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 348df7a3b07SAl Viro static void tcp_build_and_update_options(__be32 *ptr, struct tcp_sock *tp, 349cfb6eeb4SYOSHIFUJI Hideaki __u32 tstamp, __u8 **md5_hash) 35040efc6faSStephen Hemminger { 35140efc6faSStephen Hemminger if (tp->rx_opt.tstamp_ok) { 352496c98dfSYOSHIFUJI Hideaki *ptr++ = htonl((TCPOPT_NOP << 24) | 35340efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 35440efc6faSStephen Hemminger (TCPOPT_TIMESTAMP << 8) | 35540efc6faSStephen Hemminger TCPOLEN_TIMESTAMP); 35640efc6faSStephen Hemminger *ptr++ = htonl(tstamp); 35740efc6faSStephen Hemminger *ptr++ = htonl(tp->rx_opt.ts_recent); 35840efc6faSStephen Hemminger } 35940efc6faSStephen Hemminger if (tp->rx_opt.eff_sacks) { 36040efc6faSStephen Hemminger struct tcp_sack_block *sp = tp->rx_opt.dsack ? tp->duplicate_sack : tp->selective_acks; 36140efc6faSStephen Hemminger int this_sack; 36240efc6faSStephen Hemminger 36340efc6faSStephen Hemminger *ptr++ = htonl((TCPOPT_NOP << 24) | 36440efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 36540efc6faSStephen Hemminger (TCPOPT_SACK << 8) | 36640efc6faSStephen Hemminger (TCPOLEN_SACK_BASE + (tp->rx_opt.eff_sacks * 36740efc6faSStephen Hemminger TCPOLEN_SACK_PERBLOCK))); 3682de979bdSStephen Hemminger 36940efc6faSStephen Hemminger for (this_sack = 0; this_sack < tp->rx_opt.eff_sacks; this_sack++) { 37040efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].start_seq); 37140efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].end_seq); 37240efc6faSStephen Hemminger } 3732de979bdSStephen Hemminger 37440efc6faSStephen Hemminger if (tp->rx_opt.dsack) { 37540efc6faSStephen Hemminger tp->rx_opt.dsack = 0; 37640efc6faSStephen Hemminger tp->rx_opt.eff_sacks--; 37740efc6faSStephen Hemminger } 37840efc6faSStephen Hemminger } 379cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 380cfb6eeb4SYOSHIFUJI Hideaki if (md5_hash) { 381cfb6eeb4SYOSHIFUJI Hideaki *ptr++ = htonl((TCPOPT_NOP << 24) | 382cfb6eeb4SYOSHIFUJI Hideaki (TCPOPT_NOP << 16) | 383cfb6eeb4SYOSHIFUJI Hideaki (TCPOPT_MD5SIG << 8) | 384cfb6eeb4SYOSHIFUJI Hideaki TCPOLEN_MD5SIG); 385cfb6eeb4SYOSHIFUJI Hideaki *md5_hash = (__u8 *)ptr; 386cfb6eeb4SYOSHIFUJI Hideaki } 387cfb6eeb4SYOSHIFUJI Hideaki #endif 38840efc6faSStephen Hemminger } 38940efc6faSStephen Hemminger 39040efc6faSStephen Hemminger /* Construct a tcp options header for a SYN or SYN_ACK packet. 39140efc6faSStephen Hemminger * If this is every changed make sure to change the definition of 39240efc6faSStephen Hemminger * MAX_SYN_SIZE to match the new maximum number of options that you 39340efc6faSStephen Hemminger * can generate. 394cfb6eeb4SYOSHIFUJI Hideaki * 395cfb6eeb4SYOSHIFUJI Hideaki * Note - that with the RFC2385 TCP option, we make room for the 396cfb6eeb4SYOSHIFUJI Hideaki * 16 byte MD5 hash. This will be filled in later, so the pointer for the 397cfb6eeb4SYOSHIFUJI Hideaki * location to be filled is passed back up. 39840efc6faSStephen Hemminger */ 399df7a3b07SAl Viro static void tcp_syn_build_options(__be32 *ptr, int mss, int ts, int sack, 40040efc6faSStephen Hemminger int offer_wscale, int wscale, __u32 tstamp, 401cfb6eeb4SYOSHIFUJI Hideaki __u32 ts_recent, __u8 **md5_hash) 40240efc6faSStephen Hemminger { 40340efc6faSStephen Hemminger /* We always get an MSS option. 40440efc6faSStephen Hemminger * The option bytes which will be seen in normal data 40540efc6faSStephen Hemminger * packets should timestamps be used, must be in the MSS 40640efc6faSStephen Hemminger * advertised. But we subtract them from tp->mss_cache so 40740efc6faSStephen Hemminger * that calculations in tcp_sendmsg are simpler etc. 40840efc6faSStephen Hemminger * So account for this fact here if necessary. If we 40940efc6faSStephen Hemminger * don't do this correctly, as a receiver we won't 41040efc6faSStephen Hemminger * recognize data packets as being full sized when we 41140efc6faSStephen Hemminger * should, and thus we won't abide by the delayed ACK 41240efc6faSStephen Hemminger * rules correctly. 41340efc6faSStephen Hemminger * SACKs don't matter, we never delay an ACK when we 41440efc6faSStephen Hemminger * have any of those going out. 41540efc6faSStephen Hemminger */ 41640efc6faSStephen Hemminger *ptr++ = htonl((TCPOPT_MSS << 24) | (TCPOLEN_MSS << 16) | mss); 41740efc6faSStephen Hemminger if (ts) { 41840efc6faSStephen Hemminger if (sack) 419496c98dfSYOSHIFUJI Hideaki *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | 420496c98dfSYOSHIFUJI Hideaki (TCPOLEN_SACK_PERM << 16) | 421496c98dfSYOSHIFUJI Hideaki (TCPOPT_TIMESTAMP << 8) | 422496c98dfSYOSHIFUJI Hideaki TCPOLEN_TIMESTAMP); 42340efc6faSStephen Hemminger else 424496c98dfSYOSHIFUJI Hideaki *ptr++ = htonl((TCPOPT_NOP << 24) | 425496c98dfSYOSHIFUJI Hideaki (TCPOPT_NOP << 16) | 426496c98dfSYOSHIFUJI Hideaki (TCPOPT_TIMESTAMP << 8) | 427496c98dfSYOSHIFUJI Hideaki TCPOLEN_TIMESTAMP); 42840efc6faSStephen Hemminger *ptr++ = htonl(tstamp); /* TSVAL */ 42940efc6faSStephen Hemminger *ptr++ = htonl(ts_recent); /* TSECR */ 43040efc6faSStephen Hemminger } else if (sack) 431496c98dfSYOSHIFUJI Hideaki *ptr++ = htonl((TCPOPT_NOP << 24) | 432496c98dfSYOSHIFUJI Hideaki (TCPOPT_NOP << 16) | 433496c98dfSYOSHIFUJI Hideaki (TCPOPT_SACK_PERM << 8) | 434496c98dfSYOSHIFUJI Hideaki TCPOLEN_SACK_PERM); 43540efc6faSStephen Hemminger if (offer_wscale) 436496c98dfSYOSHIFUJI Hideaki *ptr++ = htonl((TCPOPT_NOP << 24) | 437496c98dfSYOSHIFUJI Hideaki (TCPOPT_WINDOW << 16) | 438496c98dfSYOSHIFUJI Hideaki (TCPOLEN_WINDOW << 8) | 439496c98dfSYOSHIFUJI Hideaki (wscale)); 440cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 441cfb6eeb4SYOSHIFUJI Hideaki /* 442cfb6eeb4SYOSHIFUJI Hideaki * If MD5 is enabled, then we set the option, and include the size 443cfb6eeb4SYOSHIFUJI Hideaki * (always 18). The actual MD5 hash is added just before the 444cfb6eeb4SYOSHIFUJI Hideaki * packet is sent. 445cfb6eeb4SYOSHIFUJI Hideaki */ 446cfb6eeb4SYOSHIFUJI Hideaki if (md5_hash) { 447cfb6eeb4SYOSHIFUJI Hideaki *ptr++ = htonl((TCPOPT_NOP << 24) | 448cfb6eeb4SYOSHIFUJI Hideaki (TCPOPT_NOP << 16) | 449cfb6eeb4SYOSHIFUJI Hideaki (TCPOPT_MD5SIG << 8) | 450cfb6eeb4SYOSHIFUJI Hideaki TCPOLEN_MD5SIG); 451cfb6eeb4SYOSHIFUJI Hideaki *md5_hash = (__u8 *)ptr; 452cfb6eeb4SYOSHIFUJI Hideaki } 453cfb6eeb4SYOSHIFUJI Hideaki #endif 45440efc6faSStephen Hemminger } 4551da177e4SLinus Torvalds 4561da177e4SLinus Torvalds /* This routine actually transmits TCP packets queued in by 4571da177e4SLinus Torvalds * tcp_do_sendmsg(). This is used by both the initial 4581da177e4SLinus Torvalds * transmission and possible later retransmissions. 4591da177e4SLinus Torvalds * All SKB's seen here are completely headerless. It is our 4601da177e4SLinus Torvalds * job to build the TCP header, and pass the packet down to 4611da177e4SLinus Torvalds * IP so it can do the same plus pass the packet off to the 4621da177e4SLinus Torvalds * device. 4631da177e4SLinus Torvalds * 4641da177e4SLinus Torvalds * We are working here with either a clone of the original 4651da177e4SLinus Torvalds * SKB, or a fresh unique copy made by the retransmit engine. 4661da177e4SLinus Torvalds */ 467056834d9SIlpo Järvinen static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, 468056834d9SIlpo Järvinen gfp_t gfp_mask) 4691da177e4SLinus Torvalds { 4706687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 471dfb4b9dcSDavid S. Miller struct inet_sock *inet; 472dfb4b9dcSDavid S. Miller struct tcp_sock *tp; 473dfb4b9dcSDavid S. Miller struct tcp_skb_cb *tcb; 474dfb4b9dcSDavid S. Miller int tcp_header_size; 475cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 476cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 477cfb6eeb4SYOSHIFUJI Hideaki __u8 *md5_hash_location; 478cfb6eeb4SYOSHIFUJI Hideaki #endif 4791da177e4SLinus Torvalds struct tcphdr *th; 4801da177e4SLinus Torvalds int sysctl_flags; 4811da177e4SLinus Torvalds int err; 4821da177e4SLinus Torvalds 483dfb4b9dcSDavid S. Miller BUG_ON(!skb || !tcp_skb_pcount(skb)); 484dfb4b9dcSDavid S. Miller 485dfb4b9dcSDavid S. Miller /* If congestion control is doing timestamping, we must 486dfb4b9dcSDavid S. Miller * take such a timestamp before we potentially clone/copy. 487dfb4b9dcSDavid S. Miller */ 488164891aaSStephen Hemminger if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP) 489dfb4b9dcSDavid S. Miller __net_timestamp(skb); 490dfb4b9dcSDavid S. Miller 491dfb4b9dcSDavid S. Miller if (likely(clone_it)) { 492dfb4b9dcSDavid S. Miller if (unlikely(skb_cloned(skb))) 493dfb4b9dcSDavid S. Miller skb = pskb_copy(skb, gfp_mask); 494dfb4b9dcSDavid S. Miller else 495dfb4b9dcSDavid S. Miller skb = skb_clone(skb, gfp_mask); 496dfb4b9dcSDavid S. Miller if (unlikely(!skb)) 497dfb4b9dcSDavid S. Miller return -ENOBUFS; 498dfb4b9dcSDavid S. Miller } 499dfb4b9dcSDavid S. Miller 500dfb4b9dcSDavid S. Miller inet = inet_sk(sk); 501dfb4b9dcSDavid S. Miller tp = tcp_sk(sk); 502dfb4b9dcSDavid S. Miller tcb = TCP_SKB_CB(skb); 503dfb4b9dcSDavid S. Miller tcp_header_size = tp->tcp_header_len; 5041da177e4SLinus Torvalds 5051da177e4SLinus Torvalds #define SYSCTL_FLAG_TSTAMPS 0x1 5061da177e4SLinus Torvalds #define SYSCTL_FLAG_WSCALE 0x2 5071da177e4SLinus Torvalds #define SYSCTL_FLAG_SACK 0x4 5081da177e4SLinus Torvalds 5091da177e4SLinus Torvalds sysctl_flags = 0; 510dfb4b9dcSDavid S. Miller if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) { 5111da177e4SLinus Torvalds tcp_header_size = sizeof(struct tcphdr) + TCPOLEN_MSS; 5121da177e4SLinus Torvalds if (sysctl_tcp_timestamps) { 5131da177e4SLinus Torvalds tcp_header_size += TCPOLEN_TSTAMP_ALIGNED; 5141da177e4SLinus Torvalds sysctl_flags |= SYSCTL_FLAG_TSTAMPS; 5151da177e4SLinus Torvalds } 5161da177e4SLinus Torvalds if (sysctl_tcp_window_scaling) { 5171da177e4SLinus Torvalds tcp_header_size += TCPOLEN_WSCALE_ALIGNED; 5181da177e4SLinus Torvalds sysctl_flags |= SYSCTL_FLAG_WSCALE; 5191da177e4SLinus Torvalds } 5201da177e4SLinus Torvalds if (sysctl_tcp_sack) { 5211da177e4SLinus Torvalds sysctl_flags |= SYSCTL_FLAG_SACK; 5221da177e4SLinus Torvalds if (!(sysctl_flags & SYSCTL_FLAG_TSTAMPS)) 5231da177e4SLinus Torvalds tcp_header_size += TCPOLEN_SACKPERM_ALIGNED; 5241da177e4SLinus Torvalds } 525dfb4b9dcSDavid S. Miller } else if (unlikely(tp->rx_opt.eff_sacks)) { 5261da177e4SLinus Torvalds /* A SACK is 2 pad bytes, a 2 byte header, plus 5271da177e4SLinus Torvalds * 2 32-bit sequence numbers for each SACK block. 5281da177e4SLinus Torvalds */ 5291da177e4SLinus Torvalds tcp_header_size += (TCPOLEN_SACK_BASE_ALIGNED + 530dfb4b9dcSDavid S. Miller (tp->rx_opt.eff_sacks * 531dfb4b9dcSDavid S. Miller TCPOLEN_SACK_PERBLOCK)); 5321da177e4SLinus Torvalds } 5331da177e4SLinus Torvalds 534317a76f9SStephen Hemminger if (tcp_packets_in_flight(tp) == 0) 5356687e988SArnaldo Carvalho de Melo tcp_ca_event(sk, CA_EVENT_TX_START); 5361da177e4SLinus Torvalds 537cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 538cfb6eeb4SYOSHIFUJI Hideaki /* 539cfb6eeb4SYOSHIFUJI Hideaki * Are we doing MD5 on this segment? If so - make 540cfb6eeb4SYOSHIFUJI Hideaki * room for it. 541cfb6eeb4SYOSHIFUJI Hideaki */ 542cfb6eeb4SYOSHIFUJI Hideaki md5 = tp->af_specific->md5_lookup(sk, sk); 543cfb6eeb4SYOSHIFUJI Hideaki if (md5) 544cfb6eeb4SYOSHIFUJI Hideaki tcp_header_size += TCPOLEN_MD5SIG_ALIGNED; 545cfb6eeb4SYOSHIFUJI Hideaki #endif 546cfb6eeb4SYOSHIFUJI Hideaki 547aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 548aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 549e89862f4SDavid S. Miller skb_set_owner_w(skb, sk); 5501da177e4SLinus Torvalds 5511da177e4SLinus Torvalds /* Build TCP header and checksum it. */ 552aa8223c7SArnaldo Carvalho de Melo th = tcp_hdr(skb); 5531da177e4SLinus Torvalds th->source = inet->sport; 5541da177e4SLinus Torvalds th->dest = inet->dport; 5551da177e4SLinus Torvalds th->seq = htonl(tcb->seq); 5561da177e4SLinus Torvalds th->ack_seq = htonl(tp->rcv_nxt); 557df7a3b07SAl Viro *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | 558dfb4b9dcSDavid S. Miller tcb->flags); 559dfb4b9dcSDavid S. Miller 560dfb4b9dcSDavid S. Miller if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) { 5611da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments 5621da177e4SLinus Torvalds * is never scaled. 5631da177e4SLinus Torvalds */ 564600ff0c2SIlpo Järvinen th->window = htons(min(tp->rcv_wnd, 65535U)); 5651da177e4SLinus Torvalds } else { 5661da177e4SLinus Torvalds th->window = htons(tcp_select_window(sk)); 5671da177e4SLinus Torvalds } 5681da177e4SLinus Torvalds th->check = 0; 5691da177e4SLinus Torvalds th->urg_ptr = 0; 5701da177e4SLinus Torvalds 571dfb4b9dcSDavid S. Miller if (unlikely(tp->urg_mode && 572dfb4b9dcSDavid S. Miller between(tp->snd_up, tcb->seq + 1, tcb->seq + 0xFFFF))) { 5731da177e4SLinus Torvalds th->urg_ptr = htons(tp->snd_up - tcb->seq); 5741da177e4SLinus Torvalds th->urg = 1; 5751da177e4SLinus Torvalds } 5761da177e4SLinus Torvalds 577dfb4b9dcSDavid S. Miller if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) { 578df7a3b07SAl Viro tcp_syn_build_options((__be32 *)(th + 1), 5791da177e4SLinus Torvalds tcp_advertise_mss(sk), 5801da177e4SLinus Torvalds (sysctl_flags & SYSCTL_FLAG_TSTAMPS), 5811da177e4SLinus Torvalds (sysctl_flags & SYSCTL_FLAG_SACK), 5821da177e4SLinus Torvalds (sysctl_flags & SYSCTL_FLAG_WSCALE), 5831da177e4SLinus Torvalds tp->rx_opt.rcv_wscale, 5841da177e4SLinus Torvalds tcb->when, 585cfb6eeb4SYOSHIFUJI Hideaki tp->rx_opt.ts_recent, 586cfb6eeb4SYOSHIFUJI Hideaki 587cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 588cfb6eeb4SYOSHIFUJI Hideaki md5 ? &md5_hash_location : 589cfb6eeb4SYOSHIFUJI Hideaki #endif 590cfb6eeb4SYOSHIFUJI Hideaki NULL); 5911da177e4SLinus Torvalds } else { 592df7a3b07SAl Viro tcp_build_and_update_options((__be32 *)(th + 1), 593cfb6eeb4SYOSHIFUJI Hideaki tp, tcb->when, 594cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 595cfb6eeb4SYOSHIFUJI Hideaki md5 ? &md5_hash_location : 596cfb6eeb4SYOSHIFUJI Hideaki #endif 597cfb6eeb4SYOSHIFUJI Hideaki NULL); 5989e412ba7SIlpo Järvinen TCP_ECN_send(sk, skb, tcp_header_size); 5991da177e4SLinus Torvalds } 600dfb4b9dcSDavid S. Miller 601cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 602cfb6eeb4SYOSHIFUJI Hideaki /* Calculate the MD5 hash, as we have all we need now */ 603cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 604cfb6eeb4SYOSHIFUJI Hideaki tp->af_specific->calc_md5_hash(md5_hash_location, 605cfb6eeb4SYOSHIFUJI Hideaki md5, 606cfb6eeb4SYOSHIFUJI Hideaki sk, NULL, NULL, 607aa8223c7SArnaldo Carvalho de Melo tcp_hdr(skb), 608cfb6eeb4SYOSHIFUJI Hideaki skb->len); 609cfb6eeb4SYOSHIFUJI Hideaki } 610cfb6eeb4SYOSHIFUJI Hideaki #endif 611cfb6eeb4SYOSHIFUJI Hideaki 6128292a17aSArnaldo Carvalho de Melo icsk->icsk_af_ops->send_check(sk, skb->len, skb); 6131da177e4SLinus Torvalds 614dfb4b9dcSDavid S. Miller if (likely(tcb->flags & TCPCB_FLAG_ACK)) 615fc6415bcSDavid S. Miller tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); 6161da177e4SLinus Torvalds 6171da177e4SLinus Torvalds if (skb->len != tcp_header_size) 6181da177e4SLinus Torvalds tcp_event_data_sent(tp, skb, sk); 6191da177e4SLinus Torvalds 620bd37a088SWei Yongjun if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) 6211da177e4SLinus Torvalds TCP_INC_STATS(TCP_MIB_OUTSEGS); 6221da177e4SLinus Torvalds 623e89862f4SDavid S. Miller err = icsk->icsk_af_ops->queue_xmit(skb, 0); 62483de47cdSHua Zhong if (likely(err <= 0)) 6251da177e4SLinus Torvalds return err; 6261da177e4SLinus Torvalds 6273cfe3baaSIlpo Järvinen tcp_enter_cwr(sk, 1); 6281da177e4SLinus Torvalds 629b9df3cb8SGerrit Renker return net_xmit_eval(err); 630dfb4b9dcSDavid S. Miller 6311da177e4SLinus Torvalds #undef SYSCTL_FLAG_TSTAMPS 6321da177e4SLinus Torvalds #undef SYSCTL_FLAG_WSCALE 6331da177e4SLinus Torvalds #undef SYSCTL_FLAG_SACK 6341da177e4SLinus Torvalds } 6351da177e4SLinus Torvalds 6361da177e4SLinus Torvalds /* This routine just queue's the buffer 6371da177e4SLinus Torvalds * 6381da177e4SLinus Torvalds * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, 6391da177e4SLinus Torvalds * otherwise socket can stall. 6401da177e4SLinus Torvalds */ 6411da177e4SLinus Torvalds static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) 6421da177e4SLinus Torvalds { 6431da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 6441da177e4SLinus Torvalds 6451da177e4SLinus Torvalds /* Advance write_seq and place onto the write_queue. */ 6461da177e4SLinus Torvalds tp->write_seq = TCP_SKB_CB(skb)->end_seq; 6471da177e4SLinus Torvalds skb_header_release(skb); 648fe067e8aSDavid S. Miller tcp_add_write_queue_tail(sk, skb); 6493ab224beSHideo Aoki sk->sk_wmem_queued += skb->truesize; 6503ab224beSHideo Aoki sk_mem_charge(sk, skb->truesize); 6511da177e4SLinus Torvalds } 6521da177e4SLinus Torvalds 653056834d9SIlpo Järvinen static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb, 654056834d9SIlpo Järvinen unsigned int mss_now) 655f6302d1dSDavid S. Miller { 656bcd76111SHerbert Xu if (skb->len <= mss_now || !sk_can_gso(sk)) { 657f6302d1dSDavid S. Miller /* Avoid the costly divide in the normal 658f6302d1dSDavid S. Miller * non-TSO case. 659f6302d1dSDavid S. Miller */ 6607967168cSHerbert Xu skb_shinfo(skb)->gso_segs = 1; 6617967168cSHerbert Xu skb_shinfo(skb)->gso_size = 0; 6627967168cSHerbert Xu skb_shinfo(skb)->gso_type = 0; 663f6302d1dSDavid S. Miller } else { 664356f89e1SIlpo Järvinen skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now); 6657967168cSHerbert Xu skb_shinfo(skb)->gso_size = mss_now; 666bcd76111SHerbert Xu skb_shinfo(skb)->gso_type = sk->sk_gso_type; 6671da177e4SLinus Torvalds } 6681da177e4SLinus Torvalds } 6691da177e4SLinus Torvalds 67091fed7a1SIlpo Järvinen /* When a modification to fackets out becomes necessary, we need to check 67168f8353bSIlpo Järvinen * skb is counted to fackets_out or not. 67291fed7a1SIlpo Järvinen */ 673a47e5a98SIlpo Järvinen static void tcp_adjust_fackets_out(struct sock *sk, struct sk_buff *skb, 67491fed7a1SIlpo Järvinen int decr) 67591fed7a1SIlpo Järvinen { 676a47e5a98SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 677a47e5a98SIlpo Järvinen 678dc86967bSIlpo Järvinen if (!tp->sacked_out || tcp_is_reno(tp)) 67991fed7a1SIlpo Järvinen return; 68091fed7a1SIlpo Järvinen 6816859d494SIlpo Järvinen if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq)) 68291fed7a1SIlpo Järvinen tp->fackets_out -= decr; 68391fed7a1SIlpo Järvinen } 68491fed7a1SIlpo Järvinen 6851da177e4SLinus Torvalds /* Function to create two new TCP segments. Shrinks the given segment 6861da177e4SLinus Torvalds * to the specified size and appends a new segment with the rest of the 6871da177e4SLinus Torvalds * packet to the list. This won't be called frequently, I hope. 6881da177e4SLinus Torvalds * Remember, these are still headerless SKBs at this point. 6891da177e4SLinus Torvalds */ 690056834d9SIlpo Järvinen int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, 691056834d9SIlpo Järvinen unsigned int mss_now) 6921da177e4SLinus Torvalds { 6931da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 6941da177e4SLinus Torvalds struct sk_buff *buff; 6956475be16SDavid S. Miller int nsize, old_factor; 696b60b49eaSHerbert Xu int nlen; 6971da177e4SLinus Torvalds u16 flags; 6981da177e4SLinus Torvalds 699b2cc99f0SHerbert Xu BUG_ON(len > skb->len); 7006a438bbeSStephen Hemminger 701b7689205SIlpo Järvinen tcp_clear_retrans_hints_partial(tp); 7021da177e4SLinus Torvalds nsize = skb_headlen(skb) - len; 7031da177e4SLinus Torvalds if (nsize < 0) 7041da177e4SLinus Torvalds nsize = 0; 7051da177e4SLinus Torvalds 7061da177e4SLinus Torvalds if (skb_cloned(skb) && 7071da177e4SLinus Torvalds skb_is_nonlinear(skb) && 7081da177e4SLinus Torvalds pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) 7091da177e4SLinus Torvalds return -ENOMEM; 7101da177e4SLinus Torvalds 7111da177e4SLinus Torvalds /* Get a new skb... force flag on. */ 7121da177e4SLinus Torvalds buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC); 7131da177e4SLinus Torvalds if (buff == NULL) 7141da177e4SLinus Torvalds return -ENOMEM; /* We'll just try again later. */ 715ef5cb973SHerbert Xu 7163ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 7173ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 718b60b49eaSHerbert Xu nlen = skb->len - len - nsize; 719b60b49eaSHerbert Xu buff->truesize += nlen; 720b60b49eaSHerbert Xu skb->truesize -= nlen; 7211da177e4SLinus Torvalds 7221da177e4SLinus Torvalds /* Correct the sequence numbers. */ 7231da177e4SLinus Torvalds TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 7241da177e4SLinus Torvalds TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 7251da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 7261da177e4SLinus Torvalds 7271da177e4SLinus Torvalds /* PSH and FIN should only be set in the second packet. */ 7281da177e4SLinus Torvalds flags = TCP_SKB_CB(skb)->flags; 7291da177e4SLinus Torvalds TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN | TCPCB_FLAG_PSH); 7301da177e4SLinus Torvalds TCP_SKB_CB(buff)->flags = flags; 731e14c3cafSHerbert Xu TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; 7321da177e4SLinus Torvalds 73384fa7933SPatrick McHardy if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { 7341da177e4SLinus Torvalds /* Copy and checksum data tail into the new buffer. */ 735056834d9SIlpo Järvinen buff->csum = csum_partial_copy_nocheck(skb->data + len, 736056834d9SIlpo Järvinen skb_put(buff, nsize), 7371da177e4SLinus Torvalds nsize, 0); 7381da177e4SLinus Torvalds 7391da177e4SLinus Torvalds skb_trim(skb, len); 7401da177e4SLinus Torvalds 7411da177e4SLinus Torvalds skb->csum = csum_block_sub(skb->csum, buff->csum, len); 7421da177e4SLinus Torvalds } else { 74384fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 7441da177e4SLinus Torvalds skb_split(skb, buff, len); 7451da177e4SLinus Torvalds } 7461da177e4SLinus Torvalds 7471da177e4SLinus Torvalds buff->ip_summed = skb->ip_summed; 7481da177e4SLinus Torvalds 7491da177e4SLinus Torvalds /* Looks stupid, but our code really uses when of 7501da177e4SLinus Torvalds * skbs, which it never sent before. --ANK 7511da177e4SLinus Torvalds */ 7521da177e4SLinus Torvalds TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when; 753a61bbcf2SPatrick McHardy buff->tstamp = skb->tstamp; 7541da177e4SLinus Torvalds 7556475be16SDavid S. Miller old_factor = tcp_skb_pcount(skb); 7566475be16SDavid S. Miller 7571da177e4SLinus Torvalds /* Fix up tso_factor for both original and new SKB. */ 758846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 759846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, buff, mss_now); 7601da177e4SLinus Torvalds 7616475be16SDavid S. Miller /* If this packet has been sent out already, we must 7626475be16SDavid S. Miller * adjust the various packet counters. 7636475be16SDavid S. Miller */ 764cf0b450cSHerbert Xu if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 7656475be16SDavid S. Miller int diff = old_factor - tcp_skb_pcount(skb) - 7666475be16SDavid S. Miller tcp_skb_pcount(buff); 7671da177e4SLinus Torvalds 7686475be16SDavid S. Miller tp->packets_out -= diff; 769e14c3cafSHerbert Xu 770e14c3cafSHerbert Xu if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 771e14c3cafSHerbert Xu tp->sacked_out -= diff; 772e14c3cafSHerbert Xu if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 773e14c3cafSHerbert Xu tp->retrans_out -= diff; 774e14c3cafSHerbert Xu 775b5860bbaSIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 7766475be16SDavid S. Miller tp->lost_out -= diff; 77783ca28beSHerbert Xu 77883ca28beSHerbert Xu /* Adjust Reno SACK estimate. */ 77991fed7a1SIlpo Järvinen if (tcp_is_reno(tp) && diff > 0) { 780af610b4cSIlpo Järvinen tcp_dec_pcount_approx_int(&tp->sacked_out, diff); 781005903bcSIlpo Järvinen tcp_verify_left_out(tp); 78283ca28beSHerbert Xu } 783a47e5a98SIlpo Järvinen tcp_adjust_fackets_out(sk, skb, diff); 7841da177e4SLinus Torvalds } 7851da177e4SLinus Torvalds 7861da177e4SLinus Torvalds /* Link BUFF into the send queue. */ 787f44b5271SDavid S. Miller skb_header_release(buff); 788fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 7891da177e4SLinus Torvalds 7901da177e4SLinus Torvalds return 0; 7911da177e4SLinus Torvalds } 7921da177e4SLinus Torvalds 7931da177e4SLinus Torvalds /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c 7941da177e4SLinus Torvalds * eventually). The difference is that pulled data not copied, but 7951da177e4SLinus Torvalds * immediately discarded. 7961da177e4SLinus Torvalds */ 797f2911969SHerbert Xu ~{PmVHI~} static void __pskb_trim_head(struct sk_buff *skb, int len) 7981da177e4SLinus Torvalds { 7991da177e4SLinus Torvalds int i, k, eat; 8001da177e4SLinus Torvalds 8011da177e4SLinus Torvalds eat = len; 8021da177e4SLinus Torvalds k = 0; 8031da177e4SLinus Torvalds for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) { 8041da177e4SLinus Torvalds if (skb_shinfo(skb)->frags[i].size <= eat) { 8051da177e4SLinus Torvalds put_page(skb_shinfo(skb)->frags[i].page); 8061da177e4SLinus Torvalds eat -= skb_shinfo(skb)->frags[i].size; 8071da177e4SLinus Torvalds } else { 8081da177e4SLinus Torvalds skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i]; 8091da177e4SLinus Torvalds if (eat) { 8101da177e4SLinus Torvalds skb_shinfo(skb)->frags[k].page_offset += eat; 8111da177e4SLinus Torvalds skb_shinfo(skb)->frags[k].size -= eat; 8121da177e4SLinus Torvalds eat = 0; 8131da177e4SLinus Torvalds } 8141da177e4SLinus Torvalds k++; 8151da177e4SLinus Torvalds } 8161da177e4SLinus Torvalds } 8171da177e4SLinus Torvalds skb_shinfo(skb)->nr_frags = k; 8181da177e4SLinus Torvalds 81927a884dcSArnaldo Carvalho de Melo skb_reset_tail_pointer(skb); 8201da177e4SLinus Torvalds skb->data_len -= len; 8211da177e4SLinus Torvalds skb->len = skb->data_len; 8221da177e4SLinus Torvalds } 8231da177e4SLinus Torvalds 8241da177e4SLinus Torvalds int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 8251da177e4SLinus Torvalds { 826056834d9SIlpo Järvinen if (skb_cloned(skb) && pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) 8271da177e4SLinus Torvalds return -ENOMEM; 8281da177e4SLinus Torvalds 829f2911969SHerbert Xu ~{PmVHI~} /* If len == headlen, we avoid __skb_pull to preserve alignment. */ 830f2911969SHerbert Xu ~{PmVHI~} if (unlikely(len < skb_headlen(skb))) 8311da177e4SLinus Torvalds __skb_pull(skb, len); 832f2911969SHerbert Xu ~{PmVHI~} else 833f2911969SHerbert Xu ~{PmVHI~} __pskb_trim_head(skb, len - skb_headlen(skb)); 8341da177e4SLinus Torvalds 8351da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq += len; 83684fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 8371da177e4SLinus Torvalds 8381da177e4SLinus Torvalds skb->truesize -= len; 8391da177e4SLinus Torvalds sk->sk_wmem_queued -= len; 8403ab224beSHideo Aoki sk_mem_uncharge(sk, len); 8411da177e4SLinus Torvalds sock_set_flag(sk, SOCK_QUEUE_SHRUNK); 8421da177e4SLinus Torvalds 8431da177e4SLinus Torvalds /* Any change of skb->len requires recalculation of tso 8441da177e4SLinus Torvalds * factor and mss. 8451da177e4SLinus Torvalds */ 8461da177e4SLinus Torvalds if (tcp_skb_pcount(skb) > 1) 847846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk, 1)); 8481da177e4SLinus Torvalds 8491da177e4SLinus Torvalds return 0; 8501da177e4SLinus Torvalds } 8511da177e4SLinus Torvalds 8525d424d5aSJohn Heffner /* Not accounting for SACKs here. */ 8535d424d5aSJohn Heffner int tcp_mtu_to_mss(struct sock *sk, int pmtu) 8545d424d5aSJohn Heffner { 8555d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 8565d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 8575d424d5aSJohn Heffner int mss_now; 8585d424d5aSJohn Heffner 8595d424d5aSJohn Heffner /* Calculate base mss without TCP options: 8605d424d5aSJohn Heffner It is MMS_S - sizeof(tcphdr) of rfc1122 8615d424d5aSJohn Heffner */ 8625d424d5aSJohn Heffner mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 8635d424d5aSJohn Heffner 8645d424d5aSJohn Heffner /* Clamp it (mss_clamp does not include tcp options) */ 8655d424d5aSJohn Heffner if (mss_now > tp->rx_opt.mss_clamp) 8665d424d5aSJohn Heffner mss_now = tp->rx_opt.mss_clamp; 8675d424d5aSJohn Heffner 8685d424d5aSJohn Heffner /* Now subtract optional transport overhead */ 8695d424d5aSJohn Heffner mss_now -= icsk->icsk_ext_hdr_len; 8705d424d5aSJohn Heffner 8715d424d5aSJohn Heffner /* Then reserve room for full set of TCP options and 8 bytes of data */ 8725d424d5aSJohn Heffner if (mss_now < 48) 8735d424d5aSJohn Heffner mss_now = 48; 8745d424d5aSJohn Heffner 8755d424d5aSJohn Heffner /* Now subtract TCP options size, not including SACKs */ 8765d424d5aSJohn Heffner mss_now -= tp->tcp_header_len - sizeof(struct tcphdr); 8775d424d5aSJohn Heffner 8785d424d5aSJohn Heffner return mss_now; 8795d424d5aSJohn Heffner } 8805d424d5aSJohn Heffner 8815d424d5aSJohn Heffner /* Inverse of above */ 8825d424d5aSJohn Heffner int tcp_mss_to_mtu(struct sock *sk, int mss) 8835d424d5aSJohn Heffner { 8845d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 8855d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 8865d424d5aSJohn Heffner int mtu; 8875d424d5aSJohn Heffner 8885d424d5aSJohn Heffner mtu = mss + 8895d424d5aSJohn Heffner tp->tcp_header_len + 8905d424d5aSJohn Heffner icsk->icsk_ext_hdr_len + 8915d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 8925d424d5aSJohn Heffner 8935d424d5aSJohn Heffner return mtu; 8945d424d5aSJohn Heffner } 8955d424d5aSJohn Heffner 8965d424d5aSJohn Heffner void tcp_mtup_init(struct sock *sk) 8975d424d5aSJohn Heffner { 8985d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 8995d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 9005d424d5aSJohn Heffner 9015d424d5aSJohn Heffner icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1; 9025d424d5aSJohn Heffner icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 9035d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 9045d424d5aSJohn Heffner icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss); 9055d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 9065d424d5aSJohn Heffner } 9075d424d5aSJohn Heffner 908409d22b4SIlpo Järvinen /* Bound MSS / TSO packet size with the half of the window */ 909409d22b4SIlpo Järvinen static int tcp_bound_to_half_wnd(struct tcp_sock *tp, int pktsize) 910409d22b4SIlpo Järvinen { 911409d22b4SIlpo Järvinen if (tp->max_window && pktsize > (tp->max_window >> 1)) 912409d22b4SIlpo Järvinen return max(tp->max_window >> 1, 68U - tp->tcp_header_len); 913409d22b4SIlpo Järvinen else 914409d22b4SIlpo Järvinen return pktsize; 915409d22b4SIlpo Järvinen } 916409d22b4SIlpo Järvinen 9171da177e4SLinus Torvalds /* This function synchronize snd mss to current pmtu/exthdr set. 9181da177e4SLinus Torvalds 9191da177e4SLinus Torvalds tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 9201da177e4SLinus Torvalds for TCP options, but includes only bare TCP header. 9211da177e4SLinus Torvalds 9221da177e4SLinus Torvalds tp->rx_opt.mss_clamp is mss negotiated at connection setup. 923caa20d9aSStephen Hemminger It is minimum of user_mss and mss received with SYN. 9241da177e4SLinus Torvalds It also does not include TCP options. 9251da177e4SLinus Torvalds 926d83d8461SArnaldo Carvalho de Melo inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. 9271da177e4SLinus Torvalds 9281da177e4SLinus Torvalds tp->mss_cache is current effective sending mss, including 9291da177e4SLinus Torvalds all tcp options except for SACKs. It is evaluated, 9301da177e4SLinus Torvalds taking into account current pmtu, but never exceeds 9311da177e4SLinus Torvalds tp->rx_opt.mss_clamp. 9321da177e4SLinus Torvalds 9331da177e4SLinus Torvalds NOTE1. rfc1122 clearly states that advertised MSS 9341da177e4SLinus Torvalds DOES NOT include either tcp or ip options. 9351da177e4SLinus Torvalds 936d83d8461SArnaldo Carvalho de Melo NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache 937d83d8461SArnaldo Carvalho de Melo are READ ONLY outside this function. --ANK (980731) 9381da177e4SLinus Torvalds */ 9391da177e4SLinus Torvalds unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 9401da177e4SLinus Torvalds { 9411da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 942d83d8461SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 9435d424d5aSJohn Heffner int mss_now; 9441da177e4SLinus Torvalds 9455d424d5aSJohn Heffner if (icsk->icsk_mtup.search_high > pmtu) 9465d424d5aSJohn Heffner icsk->icsk_mtup.search_high = pmtu; 9471da177e4SLinus Torvalds 9485d424d5aSJohn Heffner mss_now = tcp_mtu_to_mss(sk, pmtu); 949409d22b4SIlpo Järvinen mss_now = tcp_bound_to_half_wnd(tp, mss_now); 9501da177e4SLinus Torvalds 9511da177e4SLinus Torvalds /* And store cached results */ 952d83d8461SArnaldo Carvalho de Melo icsk->icsk_pmtu_cookie = pmtu; 9535d424d5aSJohn Heffner if (icsk->icsk_mtup.enabled) 9545d424d5aSJohn Heffner mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); 955c1b4a7e6SDavid S. Miller tp->mss_cache = mss_now; 9561da177e4SLinus Torvalds 9571da177e4SLinus Torvalds return mss_now; 9581da177e4SLinus Torvalds } 9591da177e4SLinus Torvalds 9601da177e4SLinus Torvalds /* Compute the current effective MSS, taking SACKs and IP options, 9611da177e4SLinus Torvalds * and even PMTU discovery events into account. 9621da177e4SLinus Torvalds * 9631da177e4SLinus Torvalds * LARGESEND note: !urg_mode is overkill, only frames up to snd_up 9641da177e4SLinus Torvalds * cannot be large. However, taking into account rare use of URG, this 9651da177e4SLinus Torvalds * is not a big flaw. 9661da177e4SLinus Torvalds */ 967c1b4a7e6SDavid S. Miller unsigned int tcp_current_mss(struct sock *sk, int large_allowed) 9681da177e4SLinus Torvalds { 9691da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 9701da177e4SLinus Torvalds struct dst_entry *dst = __sk_dst_get(sk); 971c1b4a7e6SDavid S. Miller u32 mss_now; 972c1b4a7e6SDavid S. Miller u16 xmit_size_goal; 973c1b4a7e6SDavid S. Miller int doing_tso = 0; 9741da177e4SLinus Torvalds 975c1b4a7e6SDavid S. Miller mss_now = tp->mss_cache; 976c1b4a7e6SDavid S. Miller 977bcd76111SHerbert Xu if (large_allowed && sk_can_gso(sk) && !tp->urg_mode) 978c1b4a7e6SDavid S. Miller doing_tso = 1; 979c1b4a7e6SDavid S. Miller 9801da177e4SLinus Torvalds if (dst) { 9811da177e4SLinus Torvalds u32 mtu = dst_mtu(dst); 982d83d8461SArnaldo Carvalho de Melo if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 9831da177e4SLinus Torvalds mss_now = tcp_sync_mss(sk, mtu); 9841da177e4SLinus Torvalds } 9851da177e4SLinus Torvalds 9861da177e4SLinus Torvalds if (tp->rx_opt.eff_sacks) 9871da177e4SLinus Torvalds mss_now -= (TCPOLEN_SACK_BASE_ALIGNED + 9881da177e4SLinus Torvalds (tp->rx_opt.eff_sacks * TCPOLEN_SACK_PERBLOCK)); 989c1b4a7e6SDavid S. Miller 990cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 991cfb6eeb4SYOSHIFUJI Hideaki if (tp->af_specific->md5_lookup(sk, sk)) 992cfb6eeb4SYOSHIFUJI Hideaki mss_now -= TCPOLEN_MD5SIG_ALIGNED; 993cfb6eeb4SYOSHIFUJI Hideaki #endif 994cfb6eeb4SYOSHIFUJI Hideaki 995c1b4a7e6SDavid S. Miller xmit_size_goal = mss_now; 996c1b4a7e6SDavid S. Miller 997c1b4a7e6SDavid S. Miller if (doing_tso) { 99882cc1a7aSPeter P Waskiewicz Jr xmit_size_goal = ((sk->sk_gso_max_size - 1) - 9998292a17aSArnaldo Carvalho de Melo inet_csk(sk)->icsk_af_ops->net_header_len - 1000d83d8461SArnaldo Carvalho de Melo inet_csk(sk)->icsk_ext_hdr_len - 1001d83d8461SArnaldo Carvalho de Melo tp->tcp_header_len); 1002c1b4a7e6SDavid S. Miller 1003409d22b4SIlpo Järvinen xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal); 1004c1b4a7e6SDavid S. Miller xmit_size_goal -= (xmit_size_goal % mss_now); 1005c1b4a7e6SDavid S. Miller } 1006c1b4a7e6SDavid S. Miller tp->xmit_size_goal = xmit_size_goal; 1007c1b4a7e6SDavid S. Miller 10081da177e4SLinus Torvalds return mss_now; 10091da177e4SLinus Torvalds } 10101da177e4SLinus Torvalds 1011a762a980SDavid S. Miller /* Congestion window validation. (RFC2861) */ 10129e412ba7SIlpo Järvinen static void tcp_cwnd_validate(struct sock *sk) 1013a762a980SDavid S. Miller { 10149e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1015a762a980SDavid S. Miller 1016d436d686SIlpo Järvinen if (tp->packets_out >= tp->snd_cwnd) { 1017a762a980SDavid S. Miller /* Network is feed fully. */ 1018a762a980SDavid S. Miller tp->snd_cwnd_used = 0; 1019a762a980SDavid S. Miller tp->snd_cwnd_stamp = tcp_time_stamp; 1020a762a980SDavid S. Miller } else { 1021a762a980SDavid S. Miller /* Network starves. */ 1022a762a980SDavid S. Miller if (tp->packets_out > tp->snd_cwnd_used) 1023a762a980SDavid S. Miller tp->snd_cwnd_used = tp->packets_out; 1024a762a980SDavid S. Miller 102515d33c07SDavid S. Miller if (sysctl_tcp_slow_start_after_idle && 102615d33c07SDavid S. Miller (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto) 1027a762a980SDavid S. Miller tcp_cwnd_application_limited(sk); 1028a762a980SDavid S. Miller } 1029a762a980SDavid S. Miller } 1030a762a980SDavid S. Miller 10310e3a4803SIlpo Järvinen /* Returns the portion of skb which can be sent right away without 10320e3a4803SIlpo Järvinen * introducing MSS oddities to segment boundaries. In rare cases where 10330e3a4803SIlpo Järvinen * mss_now != mss_cache, we will request caller to create a small skb 10340e3a4803SIlpo Järvinen * per input skb which could be mostly avoided here (if desired). 10355ea3a748SIlpo Järvinen * 10365ea3a748SIlpo Järvinen * We explicitly want to create a request for splitting write queue tail 10375ea3a748SIlpo Järvinen * to a small skb for Nagle purposes while avoiding unnecessary modulos, 10385ea3a748SIlpo Järvinen * thus all the complexity (cwnd_len is always MSS multiple which we 10395ea3a748SIlpo Järvinen * return whenever allowed by the other factors). Basically we need the 10405ea3a748SIlpo Järvinen * modulo only when the receiver window alone is the limiting factor or 10415ea3a748SIlpo Järvinen * when we would be allowed to send the split-due-to-Nagle skb fully. 10420e3a4803SIlpo Järvinen */ 10430e3a4803SIlpo Järvinen static unsigned int tcp_mss_split_point(struct sock *sk, struct sk_buff *skb, 1044056834d9SIlpo Järvinen unsigned int mss_now, unsigned int cwnd) 1045c1b4a7e6SDavid S. Miller { 10460e3a4803SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 10470e3a4803SIlpo Järvinen u32 needed, window, cwnd_len; 1048c1b4a7e6SDavid S. Miller 104990840defSIlpo Järvinen window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1050c1b4a7e6SDavid S. Miller cwnd_len = mss_now * cwnd; 10510e3a4803SIlpo Järvinen 10520e3a4803SIlpo Järvinen if (likely(cwnd_len <= window && skb != tcp_write_queue_tail(sk))) 10530e3a4803SIlpo Järvinen return cwnd_len; 10540e3a4803SIlpo Järvinen 10555ea3a748SIlpo Järvinen needed = min(skb->len, window); 10565ea3a748SIlpo Järvinen 105717515408SIlpo Järvinen if (cwnd_len <= needed) 10580e3a4803SIlpo Järvinen return cwnd_len; 10590e3a4803SIlpo Järvinen 10600e3a4803SIlpo Järvinen return needed - needed % mss_now; 1061c1b4a7e6SDavid S. Miller } 1062c1b4a7e6SDavid S. Miller 1063c1b4a7e6SDavid S. Miller /* Can at least one segment of SKB be sent right now, according to the 1064c1b4a7e6SDavid S. Miller * congestion window rules? If so, return how many segments are allowed. 1065c1b4a7e6SDavid S. Miller */ 1066056834d9SIlpo Järvinen static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp, 1067056834d9SIlpo Järvinen struct sk_buff *skb) 1068c1b4a7e6SDavid S. Miller { 1069c1b4a7e6SDavid S. Miller u32 in_flight, cwnd; 1070c1b4a7e6SDavid S. Miller 1071c1b4a7e6SDavid S. Miller /* Don't be strict about the congestion window for the final FIN. */ 1072104439a8SJohn Heffner if ((TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) && 1073104439a8SJohn Heffner tcp_skb_pcount(skb) == 1) 1074c1b4a7e6SDavid S. Miller return 1; 1075c1b4a7e6SDavid S. Miller 1076c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1077c1b4a7e6SDavid S. Miller cwnd = tp->snd_cwnd; 1078c1b4a7e6SDavid S. Miller if (in_flight < cwnd) 1079c1b4a7e6SDavid S. Miller return (cwnd - in_flight); 1080c1b4a7e6SDavid S. Miller 1081c1b4a7e6SDavid S. Miller return 0; 1082c1b4a7e6SDavid S. Miller } 1083c1b4a7e6SDavid S. Miller 1084c1b4a7e6SDavid S. Miller /* This must be invoked the first time we consider transmitting 1085c1b4a7e6SDavid S. Miller * SKB onto the wire. 1086c1b4a7e6SDavid S. Miller */ 1087056834d9SIlpo Järvinen static int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb, 1088056834d9SIlpo Järvinen unsigned int mss_now) 1089c1b4a7e6SDavid S. Miller { 1090c1b4a7e6SDavid S. Miller int tso_segs = tcp_skb_pcount(skb); 1091c1b4a7e6SDavid S. Miller 1092056834d9SIlpo Järvinen if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { 1093846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1094c1b4a7e6SDavid S. Miller tso_segs = tcp_skb_pcount(skb); 1095c1b4a7e6SDavid S. Miller } 1096c1b4a7e6SDavid S. Miller return tso_segs; 1097c1b4a7e6SDavid S. Miller } 1098c1b4a7e6SDavid S. Miller 1099c1b4a7e6SDavid S. Miller static inline int tcp_minshall_check(const struct tcp_sock *tp) 1100c1b4a7e6SDavid S. Miller { 1101c1b4a7e6SDavid S. Miller return after(tp->snd_sml,tp->snd_una) && 1102c1b4a7e6SDavid S. Miller !after(tp->snd_sml, tp->snd_nxt); 1103c1b4a7e6SDavid S. Miller } 1104c1b4a7e6SDavid S. Miller 1105c1b4a7e6SDavid S. Miller /* Return 0, if packet can be sent now without violation Nagle's rules: 1106c1b4a7e6SDavid S. Miller * 1. It is full sized. 1107c1b4a7e6SDavid S. Miller * 2. Or it contains FIN. (already checked by caller) 1108c1b4a7e6SDavid S. Miller * 3. Or TCP_NODELAY was set. 1109c1b4a7e6SDavid S. Miller * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 1110c1b4a7e6SDavid S. Miller * With Minshall's modification: all sent small packets are ACKed. 1111c1b4a7e6SDavid S. Miller */ 1112c1b4a7e6SDavid S. Miller static inline int tcp_nagle_check(const struct tcp_sock *tp, 1113c1b4a7e6SDavid S. Miller const struct sk_buff *skb, 1114c1b4a7e6SDavid S. Miller unsigned mss_now, int nonagle) 1115c1b4a7e6SDavid S. Miller { 1116c1b4a7e6SDavid S. Miller return (skb->len < mss_now && 1117c1b4a7e6SDavid S. Miller ((nonagle & TCP_NAGLE_CORK) || 1118056834d9SIlpo Järvinen (!nonagle && tp->packets_out && tcp_minshall_check(tp)))); 1119c1b4a7e6SDavid S. Miller } 1120c1b4a7e6SDavid S. Miller 1121c1b4a7e6SDavid S. Miller /* Return non-zero if the Nagle test allows this packet to be 1122c1b4a7e6SDavid S. Miller * sent now. 1123c1b4a7e6SDavid S. Miller */ 1124c1b4a7e6SDavid S. Miller static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb, 1125c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1126c1b4a7e6SDavid S. Miller { 1127c1b4a7e6SDavid S. Miller /* Nagle rule does not apply to frames, which sit in the middle of the 1128c1b4a7e6SDavid S. Miller * write_queue (they have no chances to get new data). 1129c1b4a7e6SDavid S. Miller * 1130c1b4a7e6SDavid S. Miller * This is implemented in the callers, where they modify the 'nonagle' 1131c1b4a7e6SDavid S. Miller * argument based upon the location of SKB in the send queue. 1132c1b4a7e6SDavid S. Miller */ 1133c1b4a7e6SDavid S. Miller if (nonagle & TCP_NAGLE_PUSH) 1134c1b4a7e6SDavid S. Miller return 1; 1135c1b4a7e6SDavid S. Miller 1136d551e454SIlpo Järvinen /* Don't use the nagle rule for urgent data (or for the final FIN). 1137d551e454SIlpo Järvinen * Nagle can be ignored during F-RTO too (see RFC4138). 1138d551e454SIlpo Järvinen */ 1139d551e454SIlpo Järvinen if (tp->urg_mode || (tp->frto_counter == 2) || 1140c1b4a7e6SDavid S. Miller (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)) 1141c1b4a7e6SDavid S. Miller return 1; 1142c1b4a7e6SDavid S. Miller 1143c1b4a7e6SDavid S. Miller if (!tcp_nagle_check(tp, skb, cur_mss, nonagle)) 1144c1b4a7e6SDavid S. Miller return 1; 1145c1b4a7e6SDavid S. Miller 1146c1b4a7e6SDavid S. Miller return 0; 1147c1b4a7e6SDavid S. Miller } 1148c1b4a7e6SDavid S. Miller 1149c1b4a7e6SDavid S. Miller /* Does at least the first segment of SKB fit into the send window? */ 1150056834d9SIlpo Järvinen static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb, 1151056834d9SIlpo Järvinen unsigned int cur_mss) 1152c1b4a7e6SDavid S. Miller { 1153c1b4a7e6SDavid S. Miller u32 end_seq = TCP_SKB_CB(skb)->end_seq; 1154c1b4a7e6SDavid S. Miller 1155c1b4a7e6SDavid S. Miller if (skb->len > cur_mss) 1156c1b4a7e6SDavid S. Miller end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 1157c1b4a7e6SDavid S. Miller 115890840defSIlpo Järvinen return !after(end_seq, tcp_wnd_end(tp)); 1159c1b4a7e6SDavid S. Miller } 1160c1b4a7e6SDavid S. Miller 1161fe067e8aSDavid S. Miller /* This checks if the data bearing packet SKB (usually tcp_send_head(sk)) 1162c1b4a7e6SDavid S. Miller * should be put on the wire right now. If so, it returns the number of 1163c1b4a7e6SDavid S. Miller * packets allowed by the congestion window. 1164c1b4a7e6SDavid S. Miller */ 1165c1b4a7e6SDavid S. Miller static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb, 1166c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1167c1b4a7e6SDavid S. Miller { 1168c1b4a7e6SDavid S. Miller struct tcp_sock *tp = tcp_sk(sk); 1169c1b4a7e6SDavid S. Miller unsigned int cwnd_quota; 1170c1b4a7e6SDavid S. Miller 1171846998aeSDavid S. Miller tcp_init_tso_segs(sk, skb, cur_mss); 1172c1b4a7e6SDavid S. Miller 1173c1b4a7e6SDavid S. Miller if (!tcp_nagle_test(tp, skb, cur_mss, nonagle)) 1174c1b4a7e6SDavid S. Miller return 0; 1175c1b4a7e6SDavid S. Miller 1176c1b4a7e6SDavid S. Miller cwnd_quota = tcp_cwnd_test(tp, skb); 1177056834d9SIlpo Järvinen if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss)) 1178c1b4a7e6SDavid S. Miller cwnd_quota = 0; 1179c1b4a7e6SDavid S. Miller 1180c1b4a7e6SDavid S. Miller return cwnd_quota; 1181c1b4a7e6SDavid S. Miller } 1182c1b4a7e6SDavid S. Miller 11839e412ba7SIlpo Järvinen int tcp_may_send_now(struct sock *sk) 1184c1b4a7e6SDavid S. Miller { 11859e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1186fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 1187c1b4a7e6SDavid S. Miller 1188c1b4a7e6SDavid S. Miller return (skb && 1189c1b4a7e6SDavid S. Miller tcp_snd_test(sk, skb, tcp_current_mss(sk, 1), 1190c1b4a7e6SDavid S. Miller (tcp_skb_is_last(sk, skb) ? 11914e67d876SIlpo Järvinen tp->nonagle : TCP_NAGLE_PUSH))); 1192c1b4a7e6SDavid S. Miller } 1193c1b4a7e6SDavid S. Miller 1194c1b4a7e6SDavid S. Miller /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 1195c1b4a7e6SDavid S. Miller * which is put after SKB on the list. It is very much like 1196c1b4a7e6SDavid S. Miller * tcp_fragment() except that it may make several kinds of assumptions 1197c1b4a7e6SDavid S. Miller * in order to speed up the splitting operation. In particular, we 1198c1b4a7e6SDavid S. Miller * know that all the data is in scatter-gather pages, and that the 1199c1b4a7e6SDavid S. Miller * packet has never been sent out before (and thus is not cloned). 1200c1b4a7e6SDavid S. Miller */ 1201056834d9SIlpo Järvinen static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, 1202056834d9SIlpo Järvinen unsigned int mss_now) 1203c1b4a7e6SDavid S. Miller { 1204c1b4a7e6SDavid S. Miller struct sk_buff *buff; 1205c1b4a7e6SDavid S. Miller int nlen = skb->len - len; 1206c1b4a7e6SDavid S. Miller u16 flags; 1207c1b4a7e6SDavid S. Miller 1208c1b4a7e6SDavid S. Miller /* All of a TSO frame must be composed of paged data. */ 1209c8ac3774SHerbert Xu if (skb->len != skb->data_len) 1210c8ac3774SHerbert Xu return tcp_fragment(sk, skb, len, mss_now); 1211c1b4a7e6SDavid S. Miller 1212df97c708SPavel Emelyanov buff = sk_stream_alloc_skb(sk, 0, GFP_ATOMIC); 1213c1b4a7e6SDavid S. Miller if (unlikely(buff == NULL)) 1214c1b4a7e6SDavid S. Miller return -ENOMEM; 1215c1b4a7e6SDavid S. Miller 12163ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 12173ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1218b60b49eaSHerbert Xu buff->truesize += nlen; 1219c1b4a7e6SDavid S. Miller skb->truesize -= nlen; 1220c1b4a7e6SDavid S. Miller 1221c1b4a7e6SDavid S. Miller /* Correct the sequence numbers. */ 1222c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 1223c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 1224c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 1225c1b4a7e6SDavid S. Miller 1226c1b4a7e6SDavid S. Miller /* PSH and FIN should only be set in the second packet. */ 1227c1b4a7e6SDavid S. Miller flags = TCP_SKB_CB(skb)->flags; 1228c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN | TCPCB_FLAG_PSH); 1229c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->flags = flags; 1230c1b4a7e6SDavid S. Miller 1231c1b4a7e6SDavid S. Miller /* This packet was never sent out yet, so no SACK bits. */ 1232c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->sacked = 0; 1233c1b4a7e6SDavid S. Miller 123484fa7933SPatrick McHardy buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; 1235c1b4a7e6SDavid S. Miller skb_split(skb, buff, len); 1236c1b4a7e6SDavid S. Miller 1237c1b4a7e6SDavid S. Miller /* Fix up tso_factor for both original and new SKB. */ 1238846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1239846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, buff, mss_now); 1240c1b4a7e6SDavid S. Miller 1241c1b4a7e6SDavid S. Miller /* Link BUFF into the send queue. */ 1242c1b4a7e6SDavid S. Miller skb_header_release(buff); 1243fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 1244c1b4a7e6SDavid S. Miller 1245c1b4a7e6SDavid S. Miller return 0; 1246c1b4a7e6SDavid S. Miller } 1247c1b4a7e6SDavid S. Miller 1248c1b4a7e6SDavid S. Miller /* Try to defer sending, if possible, in order to minimize the amount 1249c1b4a7e6SDavid S. Miller * of TSO splitting we do. View it as a kind of TSO Nagle test. 1250c1b4a7e6SDavid S. Miller * 1251c1b4a7e6SDavid S. Miller * This algorithm is from John Heffner. 1252c1b4a7e6SDavid S. Miller */ 12539e412ba7SIlpo Järvinen static int tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb) 1254c1b4a7e6SDavid S. Miller { 12559e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 12566687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 1257c1b4a7e6SDavid S. Miller u32 send_win, cong_win, limit, in_flight; 1258c1b4a7e6SDavid S. Miller 1259c1b4a7e6SDavid S. Miller if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) 1260ae8064acSJohn Heffner goto send_now; 1261c1b4a7e6SDavid S. Miller 12626687e988SArnaldo Carvalho de Melo if (icsk->icsk_ca_state != TCP_CA_Open) 1263ae8064acSJohn Heffner goto send_now; 1264ae8064acSJohn Heffner 1265ae8064acSJohn Heffner /* Defer for less than two clock ticks. */ 1266bd515c3eSIlpo Järvinen if (tp->tso_deferred && 1267bd515c3eSIlpo Järvinen ((jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1) 1268ae8064acSJohn Heffner goto send_now; 1269908a75c1SDavid S. Miller 1270c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1271c1b4a7e6SDavid S. Miller 1272056834d9SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight)); 1273c1b4a7e6SDavid S. Miller 127490840defSIlpo Järvinen send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1275c1b4a7e6SDavid S. Miller 1276c1b4a7e6SDavid S. Miller /* From in_flight test above, we know that cwnd > in_flight. */ 1277c1b4a7e6SDavid S. Miller cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; 1278c1b4a7e6SDavid S. Miller 1279c1b4a7e6SDavid S. Miller limit = min(send_win, cong_win); 1280c1b4a7e6SDavid S. Miller 1281ba244fe9SDavid S. Miller /* If a full-sized TSO skb can be sent, do it. */ 128282cc1a7aSPeter P Waskiewicz Jr if (limit >= sk->sk_gso_max_size) 1283ae8064acSJohn Heffner goto send_now; 1284ba244fe9SDavid S. Miller 1285c1b4a7e6SDavid S. Miller if (sysctl_tcp_tso_win_divisor) { 1286c1b4a7e6SDavid S. Miller u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); 1287c1b4a7e6SDavid S. Miller 1288c1b4a7e6SDavid S. Miller /* If at least some fraction of a window is available, 1289c1b4a7e6SDavid S. Miller * just use it. 1290c1b4a7e6SDavid S. Miller */ 1291c1b4a7e6SDavid S. Miller chunk /= sysctl_tcp_tso_win_divisor; 1292c1b4a7e6SDavid S. Miller if (limit >= chunk) 1293ae8064acSJohn Heffner goto send_now; 1294c1b4a7e6SDavid S. Miller } else { 1295c1b4a7e6SDavid S. Miller /* Different approach, try not to defer past a single 1296c1b4a7e6SDavid S. Miller * ACK. Receiver should ACK every other full sized 1297c1b4a7e6SDavid S. Miller * frame, so if we have space for more than 3 frames 1298c1b4a7e6SDavid S. Miller * then send now. 1299c1b4a7e6SDavid S. Miller */ 1300c1b4a7e6SDavid S. Miller if (limit > tcp_max_burst(tp) * tp->mss_cache) 1301ae8064acSJohn Heffner goto send_now; 1302c1b4a7e6SDavid S. Miller } 1303c1b4a7e6SDavid S. Miller 1304c1b4a7e6SDavid S. Miller /* Ok, it looks like it is advisable to defer. */ 1305ae8064acSJohn Heffner tp->tso_deferred = 1 | (jiffies << 1); 1306ae8064acSJohn Heffner 1307c1b4a7e6SDavid S. Miller return 1; 1308ae8064acSJohn Heffner 1309ae8064acSJohn Heffner send_now: 1310ae8064acSJohn Heffner tp->tso_deferred = 0; 1311ae8064acSJohn Heffner return 0; 1312c1b4a7e6SDavid S. Miller } 1313c1b4a7e6SDavid S. Miller 13145d424d5aSJohn Heffner /* Create a new MTU probe if we are ready. 13155d424d5aSJohn Heffner * Returns 0 if we should wait to probe (no cwnd available), 13165d424d5aSJohn Heffner * 1 if a probe was sent, 1317056834d9SIlpo Järvinen * -1 otherwise 1318056834d9SIlpo Järvinen */ 13195d424d5aSJohn Heffner static int tcp_mtu_probe(struct sock *sk) 13205d424d5aSJohn Heffner { 13215d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 13225d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 13235d424d5aSJohn Heffner struct sk_buff *skb, *nskb, *next; 13245d424d5aSJohn Heffner int len; 13255d424d5aSJohn Heffner int probe_size; 132691cc17c0SIlpo Järvinen int size_needed; 13275d424d5aSJohn Heffner int copy; 13285d424d5aSJohn Heffner int mss_now; 13295d424d5aSJohn Heffner 13305d424d5aSJohn Heffner /* Not currently probing/verifying, 13315d424d5aSJohn Heffner * not in recovery, 13325d424d5aSJohn Heffner * have enough cwnd, and 13335d424d5aSJohn Heffner * not SACKing (the variable headers throw things off) */ 13345d424d5aSJohn Heffner if (!icsk->icsk_mtup.enabled || 13355d424d5aSJohn Heffner icsk->icsk_mtup.probe_size || 13365d424d5aSJohn Heffner inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 13375d424d5aSJohn Heffner tp->snd_cwnd < 11 || 13385d424d5aSJohn Heffner tp->rx_opt.eff_sacks) 13395d424d5aSJohn Heffner return -1; 13405d424d5aSJohn Heffner 13415d424d5aSJohn Heffner /* Very simple search strategy: just double the MSS. */ 13425d424d5aSJohn Heffner mss_now = tcp_current_mss(sk, 0); 13435d424d5aSJohn Heffner probe_size = 2 * tp->mss_cache; 134491cc17c0SIlpo Järvinen size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; 13455d424d5aSJohn Heffner if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) { 13465d424d5aSJohn Heffner /* TODO: set timer for probe_converge_event */ 13475d424d5aSJohn Heffner return -1; 13485d424d5aSJohn Heffner } 13495d424d5aSJohn Heffner 13505d424d5aSJohn Heffner /* Have enough data in the send queue to probe? */ 13517f9c33e5SIlpo Järvinen if (tp->write_seq - tp->snd_nxt < size_needed) 13525d424d5aSJohn Heffner return -1; 13535d424d5aSJohn Heffner 135491cc17c0SIlpo Järvinen if (tp->snd_wnd < size_needed) 13555d424d5aSJohn Heffner return -1; 135690840defSIlpo Järvinen if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) 13575d424d5aSJohn Heffner return 0; 13585d424d5aSJohn Heffner 1359d67c58e9SIlpo Järvinen /* Do we need to wait to drain cwnd? With none in flight, don't stall */ 1360d67c58e9SIlpo Järvinen if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { 1361d67c58e9SIlpo Järvinen if (!tcp_packets_in_flight(tp)) 13625d424d5aSJohn Heffner return -1; 13635d424d5aSJohn Heffner else 13645d424d5aSJohn Heffner return 0; 13655d424d5aSJohn Heffner } 13665d424d5aSJohn Heffner 13675d424d5aSJohn Heffner /* We're allowed to probe. Build it now. */ 13685d424d5aSJohn Heffner if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL) 13695d424d5aSJohn Heffner return -1; 13703ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 13713ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 13725d424d5aSJohn Heffner 1373fe067e8aSDavid S. Miller skb = tcp_send_head(sk); 13745d424d5aSJohn Heffner 13755d424d5aSJohn Heffner TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 13765d424d5aSJohn Heffner TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 13775d424d5aSJohn Heffner TCP_SKB_CB(nskb)->flags = TCPCB_FLAG_ACK; 13785d424d5aSJohn Heffner TCP_SKB_CB(nskb)->sacked = 0; 13795d424d5aSJohn Heffner nskb->csum = 0; 138084fa7933SPatrick McHardy nskb->ip_summed = skb->ip_summed; 13815d424d5aSJohn Heffner 138250c4817eSIlpo Järvinen tcp_insert_write_queue_before(nskb, skb, sk); 138350c4817eSIlpo Järvinen 13845d424d5aSJohn Heffner len = 0; 1385234b6860SIlpo Järvinen tcp_for_write_queue_from_safe(skb, next, sk) { 13865d424d5aSJohn Heffner copy = min_t(int, skb->len, probe_size - len); 13875d424d5aSJohn Heffner if (nskb->ip_summed) 13885d424d5aSJohn Heffner skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); 13895d424d5aSJohn Heffner else 13905d424d5aSJohn Heffner nskb->csum = skb_copy_and_csum_bits(skb, 0, 1391056834d9SIlpo Järvinen skb_put(nskb, copy), 1392056834d9SIlpo Järvinen copy, nskb->csum); 13935d424d5aSJohn Heffner 13945d424d5aSJohn Heffner if (skb->len <= copy) { 13955d424d5aSJohn Heffner /* We've eaten all the data from this skb. 13965d424d5aSJohn Heffner * Throw it away. */ 13975d424d5aSJohn Heffner TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags; 1398fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 13993ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 14005d424d5aSJohn Heffner } else { 14015d424d5aSJohn Heffner TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags & 14025d424d5aSJohn Heffner ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH); 14035d424d5aSJohn Heffner if (!skb_shinfo(skb)->nr_frags) { 14045d424d5aSJohn Heffner skb_pull(skb, copy); 140584fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 1406056834d9SIlpo Järvinen skb->csum = csum_partial(skb->data, 1407056834d9SIlpo Järvinen skb->len, 0); 14085d424d5aSJohn Heffner } else { 14095d424d5aSJohn Heffner __pskb_trim_head(skb, copy); 14105d424d5aSJohn Heffner tcp_set_skb_tso_segs(sk, skb, mss_now); 14115d424d5aSJohn Heffner } 14125d424d5aSJohn Heffner TCP_SKB_CB(skb)->seq += copy; 14135d424d5aSJohn Heffner } 14145d424d5aSJohn Heffner 14155d424d5aSJohn Heffner len += copy; 1416234b6860SIlpo Järvinen 1417234b6860SIlpo Järvinen if (len >= probe_size) 1418234b6860SIlpo Järvinen break; 14195d424d5aSJohn Heffner } 14205d424d5aSJohn Heffner tcp_init_tso_segs(sk, nskb, nskb->len); 14215d424d5aSJohn Heffner 14225d424d5aSJohn Heffner /* We're ready to send. If this fails, the probe will 14235d424d5aSJohn Heffner * be resegmented into mss-sized pieces by tcp_write_xmit(). */ 14245d424d5aSJohn Heffner TCP_SKB_CB(nskb)->when = tcp_time_stamp; 14255d424d5aSJohn Heffner if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 14265d424d5aSJohn Heffner /* Decrement cwnd here because we are sending 14275d424d5aSJohn Heffner * effectively two packets. */ 14285d424d5aSJohn Heffner tp->snd_cwnd--; 142966f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, nskb); 14305d424d5aSJohn Heffner 14315d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 14320e7b1368SJohn Heffner tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 14330e7b1368SJohn Heffner tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 14345d424d5aSJohn Heffner 14355d424d5aSJohn Heffner return 1; 14365d424d5aSJohn Heffner } 14375d424d5aSJohn Heffner 14385d424d5aSJohn Heffner return -1; 14395d424d5aSJohn Heffner } 14405d424d5aSJohn Heffner 14411da177e4SLinus Torvalds /* This routine writes packets to the network. It advances the 14421da177e4SLinus Torvalds * send_head. This happens as incoming acks open up the remote 14431da177e4SLinus Torvalds * window for us. 14441da177e4SLinus Torvalds * 14451da177e4SLinus Torvalds * Returns 1, if no segments are in flight and we have queued segments, but 14461da177e4SLinus Torvalds * cannot send anything now because of SWS or another problem. 14471da177e4SLinus Torvalds */ 1448a2e2a59cSDavid S. Miller static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle) 14491da177e4SLinus Torvalds { 14501da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 145192df7b51SDavid S. Miller struct sk_buff *skb; 1452c1b4a7e6SDavid S. Miller unsigned int tso_segs, sent_pkts; 1453c1b4a7e6SDavid S. Miller int cwnd_quota; 14545d424d5aSJohn Heffner int result; 14551da177e4SLinus Torvalds 14561da177e4SLinus Torvalds /* If we are closed, the bytes will have to remain here. 14571da177e4SLinus Torvalds * In time closedown will finish, we empty the write queue and all 14581da177e4SLinus Torvalds * will be happy. 14591da177e4SLinus Torvalds */ 146092df7b51SDavid S. Miller if (unlikely(sk->sk_state == TCP_CLOSE)) 146192df7b51SDavid S. Miller return 0; 146292df7b51SDavid S. Miller 1463c1b4a7e6SDavid S. Miller sent_pkts = 0; 14645d424d5aSJohn Heffner 14655d424d5aSJohn Heffner /* Do MTU probing. */ 14665d424d5aSJohn Heffner if ((result = tcp_mtu_probe(sk)) == 0) { 14675d424d5aSJohn Heffner return 0; 14685d424d5aSJohn Heffner } else if (result > 0) { 14695d424d5aSJohn Heffner sent_pkts = 1; 14705d424d5aSJohn Heffner } 14715d424d5aSJohn Heffner 1472fe067e8aSDavid S. Miller while ((skb = tcp_send_head(sk))) { 1473c8ac3774SHerbert Xu unsigned int limit; 1474c8ac3774SHerbert Xu 1475b68e9f85SHerbert Xu tso_segs = tcp_init_tso_segs(sk, skb, mss_now); 1476c1b4a7e6SDavid S. Miller BUG_ON(!tso_segs); 1477c1b4a7e6SDavid S. Miller 1478b68e9f85SHerbert Xu cwnd_quota = tcp_cwnd_test(tp, skb); 1479b68e9f85SHerbert Xu if (!cwnd_quota) 1480b68e9f85SHerbert Xu break; 1481b68e9f85SHerbert Xu 1482b68e9f85SHerbert Xu if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) 1483b68e9f85SHerbert Xu break; 1484b68e9f85SHerbert Xu 1485c1b4a7e6SDavid S. Miller if (tso_segs == 1) { 1486aa93466bSDavid S. Miller if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 1487aa93466bSDavid S. Miller (tcp_skb_is_last(sk, skb) ? 1488aa93466bSDavid S. Miller nonagle : TCP_NAGLE_PUSH)))) 1489aa93466bSDavid S. Miller break; 1490c1b4a7e6SDavid S. Miller } else { 14919e412ba7SIlpo Järvinen if (tcp_tso_should_defer(sk, skb)) 1492aa93466bSDavid S. Miller break; 1493c1b4a7e6SDavid S. Miller } 1494aa93466bSDavid S. Miller 1495c8ac3774SHerbert Xu limit = mss_now; 14960e3a4803SIlpo Järvinen if (tso_segs > 1) 14970e3a4803SIlpo Järvinen limit = tcp_mss_split_point(sk, skb, mss_now, 14980e3a4803SIlpo Järvinen cwnd_quota); 1499c8ac3774SHerbert Xu 1500c8ac3774SHerbert Xu if (skb->len > limit && 1501c8ac3774SHerbert Xu unlikely(tso_fragment(sk, skb, limit, mss_now))) 15021da177e4SLinus Torvalds break; 15031da177e4SLinus Torvalds 15041da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 1505c1b4a7e6SDavid S. Miller 1506dfb4b9dcSDavid S. Miller if (unlikely(tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC))) 15071da177e4SLinus Torvalds break; 15081da177e4SLinus Torvalds 15091da177e4SLinus Torvalds /* Advance the send_head. This one is sent out. 15101da177e4SLinus Torvalds * This call will increment packets_out. 15111da177e4SLinus Torvalds */ 151266f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 15131da177e4SLinus Torvalds 15141da177e4SLinus Torvalds tcp_minshall_update(tp, mss_now, skb); 1515aa93466bSDavid S. Miller sent_pkts++; 15161da177e4SLinus Torvalds } 15171da177e4SLinus Torvalds 1518aa93466bSDavid S. Miller if (likely(sent_pkts)) { 15199e412ba7SIlpo Järvinen tcp_cwnd_validate(sk); 15201da177e4SLinus Torvalds return 0; 15211da177e4SLinus Torvalds } 1522fe067e8aSDavid S. Miller return !tp->packets_out && tcp_send_head(sk); 15231da177e4SLinus Torvalds } 15241da177e4SLinus Torvalds 1525a762a980SDavid S. Miller /* Push out any pending frames which were held back due to 1526a762a980SDavid S. Miller * TCP_CORK or attempt at coalescing tiny packets. 1527a762a980SDavid S. Miller * The socket must be locked by the caller. 1528a762a980SDavid S. Miller */ 15299e412ba7SIlpo Järvinen void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 15309e412ba7SIlpo Järvinen int nonagle) 1531a762a980SDavid S. Miller { 1532fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 1533a762a980SDavid S. Miller 1534a762a980SDavid S. Miller if (skb) { 153555c97f3eSDavid S. Miller if (tcp_write_xmit(sk, cur_mss, nonagle)) 15369e412ba7SIlpo Järvinen tcp_check_probe_timer(sk); 1537a762a980SDavid S. Miller } 1538a762a980SDavid S. Miller } 1539a762a980SDavid S. Miller 1540c1b4a7e6SDavid S. Miller /* Send _single_ skb sitting at the send head. This function requires 1541c1b4a7e6SDavid S. Miller * true push pending frames to setup probe timer etc. 1542c1b4a7e6SDavid S. Miller */ 1543c1b4a7e6SDavid S. Miller void tcp_push_one(struct sock *sk, unsigned int mss_now) 1544c1b4a7e6SDavid S. Miller { 1545fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 1546c1b4a7e6SDavid S. Miller unsigned int tso_segs, cwnd_quota; 1547c1b4a7e6SDavid S. Miller 1548c1b4a7e6SDavid S. Miller BUG_ON(!skb || skb->len < mss_now); 1549c1b4a7e6SDavid S. Miller 1550846998aeSDavid S. Miller tso_segs = tcp_init_tso_segs(sk, skb, mss_now); 1551c1b4a7e6SDavid S. Miller cwnd_quota = tcp_snd_test(sk, skb, mss_now, TCP_NAGLE_PUSH); 1552c1b4a7e6SDavid S. Miller 1553c1b4a7e6SDavid S. Miller if (likely(cwnd_quota)) { 1554c8ac3774SHerbert Xu unsigned int limit; 1555c8ac3774SHerbert Xu 1556c1b4a7e6SDavid S. Miller BUG_ON(!tso_segs); 1557c1b4a7e6SDavid S. Miller 1558c8ac3774SHerbert Xu limit = mss_now; 15590e3a4803SIlpo Järvinen if (tso_segs > 1) 15600e3a4803SIlpo Järvinen limit = tcp_mss_split_point(sk, skb, mss_now, 15610e3a4803SIlpo Järvinen cwnd_quota); 1562c8ac3774SHerbert Xu 1563c8ac3774SHerbert Xu if (skb->len > limit && 1564c8ac3774SHerbert Xu unlikely(tso_fragment(sk, skb, limit, mss_now))) 1565c1b4a7e6SDavid S. Miller return; 1566c1b4a7e6SDavid S. Miller 1567c1b4a7e6SDavid S. Miller /* Send it out now. */ 1568c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->when = tcp_time_stamp; 1569c1b4a7e6SDavid S. Miller 1570dfb4b9dcSDavid S. Miller if (likely(!tcp_transmit_skb(sk, skb, 1, sk->sk_allocation))) { 157166f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 15729e412ba7SIlpo Järvinen tcp_cwnd_validate(sk); 1573c1b4a7e6SDavid S. Miller return; 1574c1b4a7e6SDavid S. Miller } 1575c1b4a7e6SDavid S. Miller } 1576c1b4a7e6SDavid S. Miller } 1577c1b4a7e6SDavid S. Miller 15781da177e4SLinus Torvalds /* This function returns the amount that we can raise the 15791da177e4SLinus Torvalds * usable window based on the following constraints 15801da177e4SLinus Torvalds * 15811da177e4SLinus Torvalds * 1. The window can never be shrunk once it is offered (RFC 793) 15821da177e4SLinus Torvalds * 2. We limit memory per socket 15831da177e4SLinus Torvalds * 15841da177e4SLinus Torvalds * RFC 1122: 15851da177e4SLinus Torvalds * "the suggested [SWS] avoidance algorithm for the receiver is to keep 15861da177e4SLinus Torvalds * RECV.NEXT + RCV.WIN fixed until: 15871da177e4SLinus Torvalds * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 15881da177e4SLinus Torvalds * 15891da177e4SLinus Torvalds * i.e. don't raise the right edge of the window until you can raise 15901da177e4SLinus Torvalds * it at least MSS bytes. 15911da177e4SLinus Torvalds * 15921da177e4SLinus Torvalds * Unfortunately, the recommended algorithm breaks header prediction, 15931da177e4SLinus Torvalds * since header prediction assumes th->window stays fixed. 15941da177e4SLinus Torvalds * 15951da177e4SLinus Torvalds * Strictly speaking, keeping th->window fixed violates the receiver 15961da177e4SLinus Torvalds * side SWS prevention criteria. The problem is that under this rule 15971da177e4SLinus Torvalds * a stream of single byte packets will cause the right side of the 15981da177e4SLinus Torvalds * window to always advance by a single byte. 15991da177e4SLinus Torvalds * 16001da177e4SLinus Torvalds * Of course, if the sender implements sender side SWS prevention 16011da177e4SLinus Torvalds * then this will not be a problem. 16021da177e4SLinus Torvalds * 16031da177e4SLinus Torvalds * BSD seems to make the following compromise: 16041da177e4SLinus Torvalds * 16051da177e4SLinus Torvalds * If the free space is less than the 1/4 of the maximum 16061da177e4SLinus Torvalds * space available and the free space is less than 1/2 mss, 16071da177e4SLinus Torvalds * then set the window to 0. 16081da177e4SLinus Torvalds * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 16091da177e4SLinus Torvalds * Otherwise, just prevent the window from shrinking 16101da177e4SLinus Torvalds * and from being larger than the largest representable value. 16111da177e4SLinus Torvalds * 16121da177e4SLinus Torvalds * This prevents incremental opening of the window in the regime 16131da177e4SLinus Torvalds * where TCP is limited by the speed of the reader side taking 16141da177e4SLinus Torvalds * data out of the TCP receive queue. It does nothing about 16151da177e4SLinus Torvalds * those cases where the window is constrained on the sender side 16161da177e4SLinus Torvalds * because the pipeline is full. 16171da177e4SLinus Torvalds * 16181da177e4SLinus Torvalds * BSD also seems to "accidentally" limit itself to windows that are a 16191da177e4SLinus Torvalds * multiple of MSS, at least until the free space gets quite small. 16201da177e4SLinus Torvalds * This would appear to be a side effect of the mbuf implementation. 16211da177e4SLinus Torvalds * Combining these two algorithms results in the observed behavior 16221da177e4SLinus Torvalds * of having a fixed window size at almost all times. 16231da177e4SLinus Torvalds * 16241da177e4SLinus Torvalds * Below we obtain similar behavior by forcing the offered window to 16251da177e4SLinus Torvalds * a multiple of the mss when it is feasible to do so. 16261da177e4SLinus Torvalds * 16271da177e4SLinus Torvalds * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 16281da177e4SLinus Torvalds * Regular options like TIMESTAMP are taken into account. 16291da177e4SLinus Torvalds */ 16301da177e4SLinus Torvalds u32 __tcp_select_window(struct sock *sk) 16311da177e4SLinus Torvalds { 1632463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 16331da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1634caa20d9aSStephen Hemminger /* MSS for the peer's data. Previous versions used mss_clamp 16351da177e4SLinus Torvalds * here. I don't know if the value based on our guesses 16361da177e4SLinus Torvalds * of peer's MSS is better for the performance. It's more correct 16371da177e4SLinus Torvalds * but may be worse for the performance because of rcv_mss 16381da177e4SLinus Torvalds * fluctuations. --SAW 1998/11/1 16391da177e4SLinus Torvalds */ 1640463c84b9SArnaldo Carvalho de Melo int mss = icsk->icsk_ack.rcv_mss; 16411da177e4SLinus Torvalds int free_space = tcp_space(sk); 16421da177e4SLinus Torvalds int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk)); 16431da177e4SLinus Torvalds int window; 16441da177e4SLinus Torvalds 16451da177e4SLinus Torvalds if (mss > full_space) 16461da177e4SLinus Torvalds mss = full_space; 16471da177e4SLinus Torvalds 1648b92edbe0SEric Dumazet if (free_space < (full_space >> 1)) { 1649463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.quick = 0; 16501da177e4SLinus Torvalds 16511da177e4SLinus Torvalds if (tcp_memory_pressure) 1652056834d9SIlpo Järvinen tp->rcv_ssthresh = min(tp->rcv_ssthresh, 1653056834d9SIlpo Järvinen 4U * tp->advmss); 16541da177e4SLinus Torvalds 16551da177e4SLinus Torvalds if (free_space < mss) 16561da177e4SLinus Torvalds return 0; 16571da177e4SLinus Torvalds } 16581da177e4SLinus Torvalds 16591da177e4SLinus Torvalds if (free_space > tp->rcv_ssthresh) 16601da177e4SLinus Torvalds free_space = tp->rcv_ssthresh; 16611da177e4SLinus Torvalds 16621da177e4SLinus Torvalds /* Don't do rounding if we are using window scaling, since the 16631da177e4SLinus Torvalds * scaled window will not line up with the MSS boundary anyway. 16641da177e4SLinus Torvalds */ 16651da177e4SLinus Torvalds window = tp->rcv_wnd; 16661da177e4SLinus Torvalds if (tp->rx_opt.rcv_wscale) { 16671da177e4SLinus Torvalds window = free_space; 16681da177e4SLinus Torvalds 16691da177e4SLinus Torvalds /* Advertise enough space so that it won't get scaled away. 16701da177e4SLinus Torvalds * Import case: prevent zero window announcement if 16711da177e4SLinus Torvalds * 1<<rcv_wscale > mss. 16721da177e4SLinus Torvalds */ 16731da177e4SLinus Torvalds if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window) 16741da177e4SLinus Torvalds window = (((window >> tp->rx_opt.rcv_wscale) + 1) 16751da177e4SLinus Torvalds << tp->rx_opt.rcv_wscale); 16761da177e4SLinus Torvalds } else { 16771da177e4SLinus Torvalds /* Get the largest window that is a nice multiple of mss. 16781da177e4SLinus Torvalds * Window clamp already applied above. 16791da177e4SLinus Torvalds * If our current window offering is within 1 mss of the 16801da177e4SLinus Torvalds * free space we just keep it. This prevents the divide 16811da177e4SLinus Torvalds * and multiply from happening most of the time. 16821da177e4SLinus Torvalds * We also don't do any window rounding when the free space 16831da177e4SLinus Torvalds * is too small. 16841da177e4SLinus Torvalds */ 16851da177e4SLinus Torvalds if (window <= free_space - mss || window > free_space) 16861da177e4SLinus Torvalds window = (free_space / mss) * mss; 168784565070SJohn Heffner else if (mss == full_space && 1688b92edbe0SEric Dumazet free_space > window + (full_space >> 1)) 168984565070SJohn Heffner window = free_space; 16901da177e4SLinus Torvalds } 16911da177e4SLinus Torvalds 16921da177e4SLinus Torvalds return window; 16931da177e4SLinus Torvalds } 16941da177e4SLinus Torvalds 16951da177e4SLinus Torvalds /* Attempt to collapse two adjacent SKB's during retransmission. */ 1696056834d9SIlpo Järvinen static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *skb, 1697056834d9SIlpo Järvinen int mss_now) 16981da177e4SLinus Torvalds { 16991da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1700fe067e8aSDavid S. Miller struct sk_buff *next_skb = tcp_write_queue_next(sk, skb); 1701058dc334SIlpo Järvinen int skb_size, next_skb_size; 1702058dc334SIlpo Järvinen u16 flags; 17031da177e4SLinus Torvalds 17041da177e4SLinus Torvalds /* The first test we must make is that neither of these two 17051da177e4SLinus Torvalds * SKB's are still referenced by someone else. 17061da177e4SLinus Torvalds */ 1707058dc334SIlpo Järvinen if (skb_cloned(skb) || skb_cloned(next_skb)) 1708058dc334SIlpo Järvinen return; 1709058dc334SIlpo Järvinen 1710058dc334SIlpo Järvinen skb_size = skb->len; 1711058dc334SIlpo Järvinen next_skb_size = next_skb->len; 1712058dc334SIlpo Järvinen flags = TCP_SKB_CB(skb)->flags; 17131da177e4SLinus Torvalds 17141da177e4SLinus Torvalds /* Also punt if next skb has been SACK'd. */ 17151da177e4SLinus Torvalds if (TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_ACKED) 17161da177e4SLinus Torvalds return; 17171da177e4SLinus Torvalds 17181da177e4SLinus Torvalds /* Next skb is out of window. */ 171990840defSIlpo Järvinen if (after(TCP_SKB_CB(next_skb)->end_seq, tcp_wnd_end(tp))) 17201da177e4SLinus Torvalds return; 17211da177e4SLinus Torvalds 17221da177e4SLinus Torvalds /* Punt if not enough space exists in the first SKB for 17231da177e4SLinus Torvalds * the data in the second, or the total combined payload 17241da177e4SLinus Torvalds * would exceed the MSS. 17251da177e4SLinus Torvalds */ 17261da177e4SLinus Torvalds if ((next_skb_size > skb_tailroom(skb)) || 17271da177e4SLinus Torvalds ((skb_size + next_skb_size) > mss_now)) 17281da177e4SLinus Torvalds return; 17291da177e4SLinus Torvalds 1730058dc334SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); 17311da177e4SLinus Torvalds 17326859d494SIlpo Järvinen tcp_highest_sack_combine(sk, next_skb, skb); 1733a6963a6bSIlpo Järvinen 17341da177e4SLinus Torvalds /* Ok. We will be able to collapse the packet. */ 1735fe067e8aSDavid S. Miller tcp_unlink_write_queue(next_skb, sk); 17361da177e4SLinus Torvalds 1737058dc334SIlpo Järvinen skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size), 17381a4e2d09SArnaldo Carvalho de Melo next_skb_size); 17391da177e4SLinus Torvalds 174052d570aaSJarek Poplawski if (next_skb->ip_summed == CHECKSUM_PARTIAL) 174152d570aaSJarek Poplawski skb->ip_summed = CHECKSUM_PARTIAL; 17421da177e4SLinus Torvalds 174384fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 17441da177e4SLinus Torvalds skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); 17451da177e4SLinus Torvalds 17461da177e4SLinus Torvalds /* Update sequence range on original skb. */ 17471da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 17481da177e4SLinus Torvalds 17491da177e4SLinus Torvalds /* Merge over control information. */ 17501da177e4SLinus Torvalds flags |= TCP_SKB_CB(next_skb)->flags; /* This moves PSH/FIN etc. over */ 17511da177e4SLinus Torvalds TCP_SKB_CB(skb)->flags = flags; 17521da177e4SLinus Torvalds 17531da177e4SLinus Torvalds /* All done, get rid of second SKB and account for it so 17541da177e4SLinus Torvalds * packet counting does not break. 17551da177e4SLinus Torvalds */ 17564828e7f4SIlpo Järvinen TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; 17571da177e4SLinus Torvalds if (TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_RETRANS) 17581da177e4SLinus Torvalds tp->retrans_out -= tcp_skb_pcount(next_skb); 1759b5860bbaSIlpo Järvinen if (TCP_SKB_CB(next_skb)->sacked & TCPCB_LOST) 17601da177e4SLinus Torvalds tp->lost_out -= tcp_skb_pcount(next_skb); 17611da177e4SLinus Torvalds /* Reno case is special. Sigh... */ 1762e60402d0SIlpo Järvinen if (tcp_is_reno(tp) && tp->sacked_out) 17631da177e4SLinus Torvalds tcp_dec_pcount_approx(&tp->sacked_out, next_skb); 17641da177e4SLinus Torvalds 1765a47e5a98SIlpo Järvinen tcp_adjust_fackets_out(sk, next_skb, tcp_skb_pcount(next_skb)); 1766e9144bd8SIlpo Järvinen tp->packets_out -= tcp_skb_pcount(next_skb); 1767b7689205SIlpo Järvinen 1768b7689205SIlpo Järvinen /* changed transmit queue under us so clear hints */ 1769b7689205SIlpo Järvinen tcp_clear_retrans_hints_partial(tp); 1770b7689205SIlpo Järvinen 17713ab224beSHideo Aoki sk_wmem_free_skb(sk, next_skb); 17721da177e4SLinus Torvalds } 17731da177e4SLinus Torvalds 17741da177e4SLinus Torvalds /* Do a simple retransmit without using the backoff mechanisms in 17751da177e4SLinus Torvalds * tcp_timer. This is used for path mtu discovery. 17761da177e4SLinus Torvalds * The socket is already locked here. 17771da177e4SLinus Torvalds */ 17781da177e4SLinus Torvalds void tcp_simple_retransmit(struct sock *sk) 17791da177e4SLinus Torvalds { 17806687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 17811da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 17821da177e4SLinus Torvalds struct sk_buff *skb; 17831da177e4SLinus Torvalds unsigned int mss = tcp_current_mss(sk, 0); 17841da177e4SLinus Torvalds int lost = 0; 17851da177e4SLinus Torvalds 1786fe067e8aSDavid S. Miller tcp_for_write_queue(skb, sk) { 1787fe067e8aSDavid S. Miller if (skb == tcp_send_head(sk)) 1788fe067e8aSDavid S. Miller break; 17891da177e4SLinus Torvalds if (skb->len > mss && 17901da177e4SLinus Torvalds !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) { 17911da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 17921da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 17931da177e4SLinus Torvalds tp->retrans_out -= tcp_skb_pcount(skb); 17941da177e4SLinus Torvalds } 17951da177e4SLinus Torvalds if (!(TCP_SKB_CB(skb)->sacked & TCPCB_LOST)) { 17961da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 17971da177e4SLinus Torvalds tp->lost_out += tcp_skb_pcount(skb); 17981da177e4SLinus Torvalds lost = 1; 17991da177e4SLinus Torvalds } 18001da177e4SLinus Torvalds } 18011da177e4SLinus Torvalds } 18021da177e4SLinus Torvalds 18035af4ec23SIlpo Järvinen tcp_clear_all_retrans_hints(tp); 18046a438bbeSStephen Hemminger 18051da177e4SLinus Torvalds if (!lost) 18061da177e4SLinus Torvalds return; 18071da177e4SLinus Torvalds 1808882bebaaSIlpo Järvinen if (tcp_is_reno(tp)) 1809882bebaaSIlpo Järvinen tcp_limit_reno_sacked(tp); 1810882bebaaSIlpo Järvinen 1811005903bcSIlpo Järvinen tcp_verify_left_out(tp); 18121da177e4SLinus Torvalds 18131da177e4SLinus Torvalds /* Don't muck with the congestion window here. 18141da177e4SLinus Torvalds * Reason is that we do not increase amount of _data_ 18151da177e4SLinus Torvalds * in network, but units changed and effective 18161da177e4SLinus Torvalds * cwnd/ssthresh really reduced now. 18171da177e4SLinus Torvalds */ 18186687e988SArnaldo Carvalho de Melo if (icsk->icsk_ca_state != TCP_CA_Loss) { 18191da177e4SLinus Torvalds tp->high_seq = tp->snd_nxt; 18206687e988SArnaldo Carvalho de Melo tp->snd_ssthresh = tcp_current_ssthresh(sk); 18211da177e4SLinus Torvalds tp->prior_ssthresh = 0; 18221da177e4SLinus Torvalds tp->undo_marker = 0; 18236687e988SArnaldo Carvalho de Melo tcp_set_ca_state(sk, TCP_CA_Loss); 18241da177e4SLinus Torvalds } 18251da177e4SLinus Torvalds tcp_xmit_retransmit_queue(sk); 18261da177e4SLinus Torvalds } 18271da177e4SLinus Torvalds 18281da177e4SLinus Torvalds /* This retransmits one SKB. Policy decisions and retransmit queue 18291da177e4SLinus Torvalds * state updates are done by the caller. Returns non-zero if an 18301da177e4SLinus Torvalds * error occurred which prevented the send. 18311da177e4SLinus Torvalds */ 18321da177e4SLinus Torvalds int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) 18331da177e4SLinus Torvalds { 18341da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 18355d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 18367d227cd2SSridhar Samudrala unsigned int cur_mss; 18371da177e4SLinus Torvalds int err; 18381da177e4SLinus Torvalds 18395d424d5aSJohn Heffner /* Inconslusive MTU probe */ 18405d424d5aSJohn Heffner if (icsk->icsk_mtup.probe_size) { 18415d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 18425d424d5aSJohn Heffner } 18435d424d5aSJohn Heffner 18441da177e4SLinus Torvalds /* Do not sent more than we queued. 1/4 is reserved for possible 1845caa20d9aSStephen Hemminger * copying overhead: fragmentation, tunneling, mangling etc. 18461da177e4SLinus Torvalds */ 18471da177e4SLinus Torvalds if (atomic_read(&sk->sk_wmem_alloc) > 18481da177e4SLinus Torvalds min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf)) 18491da177e4SLinus Torvalds return -EAGAIN; 18501da177e4SLinus Torvalds 18511da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 18521da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 18531da177e4SLinus Torvalds BUG(); 18541da177e4SLinus Torvalds if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 18551da177e4SLinus Torvalds return -ENOMEM; 18561da177e4SLinus Torvalds } 18571da177e4SLinus Torvalds 18587d227cd2SSridhar Samudrala if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 18597d227cd2SSridhar Samudrala return -EHOSTUNREACH; /* Routing failure or similar. */ 18607d227cd2SSridhar Samudrala 18617d227cd2SSridhar Samudrala cur_mss = tcp_current_mss(sk, 0); 18627d227cd2SSridhar Samudrala 18631da177e4SLinus Torvalds /* If receiver has shrunk his window, and skb is out of 18641da177e4SLinus Torvalds * new window, do not retransmit it. The exception is the 18651da177e4SLinus Torvalds * case, when window is shrunk to zero. In this case 18661da177e4SLinus Torvalds * our retransmit serves as a zero window probe. 18671da177e4SLinus Torvalds */ 186890840defSIlpo Järvinen if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) 18691da177e4SLinus Torvalds && TCP_SKB_CB(skb)->seq != tp->snd_una) 18701da177e4SLinus Torvalds return -EAGAIN; 18711da177e4SLinus Torvalds 18721da177e4SLinus Torvalds if (skb->len > cur_mss) { 1873846998aeSDavid S. Miller if (tcp_fragment(sk, skb, cur_mss, cur_mss)) 18741da177e4SLinus Torvalds return -ENOMEM; /* We'll try again later. */ 18751da177e4SLinus Torvalds } 18761da177e4SLinus Torvalds 18771da177e4SLinus Torvalds /* Collapse two adjacent packets if worthwhile and we can. */ 18781da177e4SLinus Torvalds if (!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN) && 18791da177e4SLinus Torvalds (skb->len < (cur_mss >> 1)) && 1880fe067e8aSDavid S. Miller (tcp_write_queue_next(sk, skb) != tcp_send_head(sk)) && 1881fe067e8aSDavid S. Miller (!tcp_skb_is_last(sk, skb)) && 1882056834d9SIlpo Järvinen (skb_shinfo(skb)->nr_frags == 0 && 1883056834d9SIlpo Järvinen skb_shinfo(tcp_write_queue_next(sk, skb))->nr_frags == 0) && 1884056834d9SIlpo Järvinen (tcp_skb_pcount(skb) == 1 && 1885056834d9SIlpo Järvinen tcp_skb_pcount(tcp_write_queue_next(sk, skb)) == 1) && 18861da177e4SLinus Torvalds (sysctl_tcp_retrans_collapse != 0)) 18871da177e4SLinus Torvalds tcp_retrans_try_collapse(sk, skb, cur_mss); 18881da177e4SLinus Torvalds 18891da177e4SLinus Torvalds /* Some Solaris stacks overoptimize and ignore the FIN on a 18901da177e4SLinus Torvalds * retransmit when old data is attached. So strip it off 18911da177e4SLinus Torvalds * since it is cheap to do so and saves bytes on the network. 18921da177e4SLinus Torvalds */ 18931da177e4SLinus Torvalds if (skb->len > 0 && 18941da177e4SLinus Torvalds (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) && 18951da177e4SLinus Torvalds tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) { 18961da177e4SLinus Torvalds if (!pskb_trim(skb, 0)) { 1897e870a8efSIlpo Järvinen /* Reuse, even though it does some unnecessary work */ 1898e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1, 1899e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->flags); 19001da177e4SLinus Torvalds skb->ip_summed = CHECKSUM_NONE; 19011da177e4SLinus Torvalds } 19021da177e4SLinus Torvalds } 19031da177e4SLinus Torvalds 19041da177e4SLinus Torvalds /* Make a copy, if the first transmission SKB clone we made 19051da177e4SLinus Torvalds * is still in somebody's hands, else make a clone. 19061da177e4SLinus Torvalds */ 19071da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 19081da177e4SLinus Torvalds 1909dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 19101da177e4SLinus Torvalds 19111da177e4SLinus Torvalds if (err == 0) { 19121da177e4SLinus Torvalds /* Update global TCP statistics. */ 19131da177e4SLinus Torvalds TCP_INC_STATS(TCP_MIB_RETRANSSEGS); 19141da177e4SLinus Torvalds 19151da177e4SLinus Torvalds tp->total_retrans++; 19161da177e4SLinus Torvalds 19171da177e4SLinus Torvalds #if FASTRETRANS_DEBUG > 0 19181da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 19191da177e4SLinus Torvalds if (net_ratelimit()) 19201da177e4SLinus Torvalds printk(KERN_DEBUG "retrans_out leaked.\n"); 19211da177e4SLinus Torvalds } 19221da177e4SLinus Torvalds #endif 1923b08d6cb2SIlpo Järvinen if (!tp->retrans_out) 1924b08d6cb2SIlpo Järvinen tp->lost_retrans_low = tp->snd_nxt; 19251da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 19261da177e4SLinus Torvalds tp->retrans_out += tcp_skb_pcount(skb); 19271da177e4SLinus Torvalds 19281da177e4SLinus Torvalds /* Save stamp of the first retransmit. */ 19291da177e4SLinus Torvalds if (!tp->retrans_stamp) 19301da177e4SLinus Torvalds tp->retrans_stamp = TCP_SKB_CB(skb)->when; 19311da177e4SLinus Torvalds 19321da177e4SLinus Torvalds tp->undo_retrans++; 19331da177e4SLinus Torvalds 19341da177e4SLinus Torvalds /* snd_nxt is stored to detect loss of retransmitted segment, 19351da177e4SLinus Torvalds * see tcp_input.c tcp_sacktag_write_queue(). 19361da177e4SLinus Torvalds */ 19371da177e4SLinus Torvalds TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt; 19381da177e4SLinus Torvalds } 19391da177e4SLinus Torvalds return err; 19401da177e4SLinus Torvalds } 19411da177e4SLinus Torvalds 19421da177e4SLinus Torvalds /* This gets called after a retransmit timeout, and the initially 19431da177e4SLinus Torvalds * retransmitted data is acknowledged. It tries to continue 19441da177e4SLinus Torvalds * resending the rest of the retransmit queue, until either 19451da177e4SLinus Torvalds * we've sent it all or the congestion window limit is reached. 19461da177e4SLinus Torvalds * If doing SACK, the first ACK which comes back for a timeout 19471da177e4SLinus Torvalds * based retransmit packet might feed us FACK information again. 19481da177e4SLinus Torvalds * If so, we use it to avoid unnecessarily retransmissions. 19491da177e4SLinus Torvalds */ 19501da177e4SLinus Torvalds void tcp_xmit_retransmit_queue(struct sock *sk) 19511da177e4SLinus Torvalds { 19526687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 19531da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 19541da177e4SLinus Torvalds struct sk_buff *skb; 19556a438bbeSStephen Hemminger int packet_cnt; 19566a438bbeSStephen Hemminger 19576a438bbeSStephen Hemminger if (tp->retransmit_skb_hint) { 19586a438bbeSStephen Hemminger skb = tp->retransmit_skb_hint; 19596a438bbeSStephen Hemminger packet_cnt = tp->retransmit_cnt_hint; 19606a438bbeSStephen Hemminger } else { 1961fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 19626a438bbeSStephen Hemminger packet_cnt = 0; 19636a438bbeSStephen Hemminger } 19641da177e4SLinus Torvalds 19651da177e4SLinus Torvalds /* First pass: retransmit lost packets. */ 19666a438bbeSStephen Hemminger if (tp->lost_out) { 1967fe067e8aSDavid S. Miller tcp_for_write_queue_from(skb, sk) { 19681da177e4SLinus Torvalds __u8 sacked = TCP_SKB_CB(skb)->sacked; 19691da177e4SLinus Torvalds 1970fe067e8aSDavid S. Miller if (skb == tcp_send_head(sk)) 1971fe067e8aSDavid S. Miller break; 19726a438bbeSStephen Hemminger /* we could do better than to assign each time */ 19736a438bbeSStephen Hemminger tp->retransmit_skb_hint = skb; 19746a438bbeSStephen Hemminger tp->retransmit_cnt_hint = packet_cnt; 19756a438bbeSStephen Hemminger 19761da177e4SLinus Torvalds /* Assume this retransmit will generate 19771da177e4SLinus Torvalds * only one packet for congestion window 19781da177e4SLinus Torvalds * calculation purposes. This works because 19791da177e4SLinus Torvalds * tcp_retransmit_skb() will chop up the 19801da177e4SLinus Torvalds * packet to be MSS sized and all the 19811da177e4SLinus Torvalds * packet counting works out. 19821da177e4SLinus Torvalds */ 19831da177e4SLinus Torvalds if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) 19841da177e4SLinus Torvalds return; 19851da177e4SLinus Torvalds 19861da177e4SLinus Torvalds if (sacked & TCPCB_LOST) { 19871da177e4SLinus Torvalds if (!(sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))) { 1988*40b215e5SPavel Emelyanov int mib_idx; 1989*40b215e5SPavel Emelyanov 19906a438bbeSStephen Hemminger if (tcp_retransmit_skb(sk, skb)) { 19916a438bbeSStephen Hemminger tp->retransmit_skb_hint = NULL; 19921da177e4SLinus Torvalds return; 19936a438bbeSStephen Hemminger } 19946687e988SArnaldo Carvalho de Melo if (icsk->icsk_ca_state != TCP_CA_Loss) 1995*40b215e5SPavel Emelyanov mib_idx = LINUX_MIB_TCPFASTRETRANS; 19961da177e4SLinus Torvalds else 1997*40b215e5SPavel Emelyanov mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; 1998*40b215e5SPavel Emelyanov NET_INC_STATS_BH(mib_idx); 19991da177e4SLinus Torvalds 2000fe067e8aSDavid S. Miller if (skb == tcp_write_queue_head(sk)) 2001463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 20023f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, 20033f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 20041da177e4SLinus Torvalds } 20051da177e4SLinus Torvalds 20066a438bbeSStephen Hemminger packet_cnt += tcp_skb_pcount(skb); 20076a438bbeSStephen Hemminger if (packet_cnt >= tp->lost_out) 20081da177e4SLinus Torvalds break; 20091da177e4SLinus Torvalds } 20101da177e4SLinus Torvalds } 20111da177e4SLinus Torvalds } 20121da177e4SLinus Torvalds 20131da177e4SLinus Torvalds /* OK, demanded retransmission is finished. */ 20141da177e4SLinus Torvalds 20151da177e4SLinus Torvalds /* Forward retransmissions are possible only during Recovery. */ 20166687e988SArnaldo Carvalho de Melo if (icsk->icsk_ca_state != TCP_CA_Recovery) 20171da177e4SLinus Torvalds return; 20181da177e4SLinus Torvalds 20191da177e4SLinus Torvalds /* No forward retransmissions in Reno are possible. */ 2020e60402d0SIlpo Järvinen if (tcp_is_reno(tp)) 20211da177e4SLinus Torvalds return; 20221da177e4SLinus Torvalds 20231da177e4SLinus Torvalds /* Yeah, we have to make difficult choice between forward transmission 20241da177e4SLinus Torvalds * and retransmission... Both ways have their merits... 20251da177e4SLinus Torvalds * 20261da177e4SLinus Torvalds * For now we do not retransmit anything, while we have some new 2027539d243fSIlpo Järvinen * segments to send. In the other cases, follow rule 3 for 2028539d243fSIlpo Järvinen * NextSeg() specified in RFC3517. 20291da177e4SLinus Torvalds */ 20301da177e4SLinus Torvalds 20319e412ba7SIlpo Järvinen if (tcp_may_send_now(sk)) 20321da177e4SLinus Torvalds return; 20331da177e4SLinus Torvalds 2034539d243fSIlpo Järvinen /* If nothing is SACKed, highest_sack in the loop won't be valid */ 2035539d243fSIlpo Järvinen if (!tp->sacked_out) 2036539d243fSIlpo Järvinen return; 2037539d243fSIlpo Järvinen 2038539d243fSIlpo Järvinen if (tp->forward_skb_hint) 20396a438bbeSStephen Hemminger skb = tp->forward_skb_hint; 2040539d243fSIlpo Järvinen else 2041fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 20421da177e4SLinus Torvalds 2043fe067e8aSDavid S. Miller tcp_for_write_queue_from(skb, sk) { 2044fe067e8aSDavid S. Miller if (skb == tcp_send_head(sk)) 2045fe067e8aSDavid S. Miller break; 20466a438bbeSStephen Hemminger tp->forward_skb_hint = skb; 20476a438bbeSStephen Hemminger 20486859d494SIlpo Järvinen if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp))) 20491da177e4SLinus Torvalds break; 20501da177e4SLinus Torvalds 20511da177e4SLinus Torvalds if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) 20521da177e4SLinus Torvalds break; 20531da177e4SLinus Torvalds 20541da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) 20551da177e4SLinus Torvalds continue; 20561da177e4SLinus Torvalds 20571da177e4SLinus Torvalds /* Ok, retransmit it. */ 20586a438bbeSStephen Hemminger if (tcp_retransmit_skb(sk, skb)) { 20596a438bbeSStephen Hemminger tp->forward_skb_hint = NULL; 20601da177e4SLinus Torvalds break; 20616a438bbeSStephen Hemminger } 20621da177e4SLinus Torvalds 2063fe067e8aSDavid S. Miller if (skb == tcp_write_queue_head(sk)) 20643f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 20653f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, 20663f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 20671da177e4SLinus Torvalds 20681da177e4SLinus Torvalds NET_INC_STATS_BH(LINUX_MIB_TCPFORWARDRETRANS); 20691da177e4SLinus Torvalds } 20701da177e4SLinus Torvalds } 20711da177e4SLinus Torvalds 20721da177e4SLinus Torvalds /* Send a fin. The caller locks the socket for us. This cannot be 20731da177e4SLinus Torvalds * allowed to fail queueing a FIN frame under any circumstances. 20741da177e4SLinus Torvalds */ 20751da177e4SLinus Torvalds void tcp_send_fin(struct sock *sk) 20761da177e4SLinus Torvalds { 20771da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2078fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_write_queue_tail(sk); 20791da177e4SLinus Torvalds int mss_now; 20801da177e4SLinus Torvalds 20811da177e4SLinus Torvalds /* Optimization, tack on the FIN if we have a queue of 20821da177e4SLinus Torvalds * unsent frames. But be careful about outgoing SACKS 20831da177e4SLinus Torvalds * and IP options. 20841da177e4SLinus Torvalds */ 20851da177e4SLinus Torvalds mss_now = tcp_current_mss(sk, 1); 20861da177e4SLinus Torvalds 2087fe067e8aSDavid S. Miller if (tcp_send_head(sk) != NULL) { 20881da177e4SLinus Torvalds TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_FIN; 20891da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq++; 20901da177e4SLinus Torvalds tp->write_seq++; 20911da177e4SLinus Torvalds } else { 20921da177e4SLinus Torvalds /* Socket is locked, keep trying until memory is available. */ 20931da177e4SLinus Torvalds for (;;) { 2094d179cd12SDavid S. Miller skb = alloc_skb_fclone(MAX_TCP_HEADER, GFP_KERNEL); 20951da177e4SLinus Torvalds if (skb) 20961da177e4SLinus Torvalds break; 20971da177e4SLinus Torvalds yield(); 20981da177e4SLinus Torvalds } 20991da177e4SLinus Torvalds 21001da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 21011da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 21021da177e4SLinus Torvalds /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 2103e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tp->write_seq, 2104e870a8efSIlpo Järvinen TCPCB_FLAG_ACK | TCPCB_FLAG_FIN); 21051da177e4SLinus Torvalds tcp_queue_skb(sk, skb); 21061da177e4SLinus Torvalds } 21079e412ba7SIlpo Järvinen __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF); 21081da177e4SLinus Torvalds } 21091da177e4SLinus Torvalds 21101da177e4SLinus Torvalds /* We get here when a process closes a file descriptor (either due to 21111da177e4SLinus Torvalds * an explicit close() or as a byproduct of exit()'ing) and there 21121da177e4SLinus Torvalds * was unread data in the receive queue. This behavior is recommended 211365bb723cSGerrit Renker * by RFC 2525, section 2.17. -DaveM 21141da177e4SLinus Torvalds */ 2115dd0fc66fSAl Viro void tcp_send_active_reset(struct sock *sk, gfp_t priority) 21161da177e4SLinus Torvalds { 21171da177e4SLinus Torvalds struct sk_buff *skb; 21181da177e4SLinus Torvalds 21191da177e4SLinus Torvalds /* NOTE: No TCP options attached and we never retransmit this. */ 21201da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, priority); 21211da177e4SLinus Torvalds if (!skb) { 21221da177e4SLinus Torvalds NET_INC_STATS(LINUX_MIB_TCPABORTFAILED); 21231da177e4SLinus Torvalds return; 21241da177e4SLinus Torvalds } 21251da177e4SLinus Torvalds 21261da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 21271da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 2128e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), 2129e870a8efSIlpo Järvinen TCPCB_FLAG_ACK | TCPCB_FLAG_RST); 21301da177e4SLinus Torvalds /* Send it off. */ 21311da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2132dfb4b9dcSDavid S. Miller if (tcp_transmit_skb(sk, skb, 0, priority)) 21331da177e4SLinus Torvalds NET_INC_STATS(LINUX_MIB_TCPABORTFAILED); 213426af65cbSSridhar Samudrala 213526af65cbSSridhar Samudrala TCP_INC_STATS(TCP_MIB_OUTRSTS); 21361da177e4SLinus Torvalds } 21371da177e4SLinus Torvalds 21381da177e4SLinus Torvalds /* WARNING: This routine must only be called when we have already sent 21391da177e4SLinus Torvalds * a SYN packet that crossed the incoming SYN that caused this routine 21401da177e4SLinus Torvalds * to get called. If this assumption fails then the initial rcv_wnd 21411da177e4SLinus Torvalds * and rcv_wscale values will not be correct. 21421da177e4SLinus Torvalds */ 21431da177e4SLinus Torvalds int tcp_send_synack(struct sock *sk) 21441da177e4SLinus Torvalds { 21451da177e4SLinus Torvalds struct sk_buff *skb; 21461da177e4SLinus Torvalds 2147fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 21481da177e4SLinus Torvalds if (skb == NULL || !(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN)) { 21491da177e4SLinus Torvalds printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n"); 21501da177e4SLinus Torvalds return -EFAULT; 21511da177e4SLinus Torvalds } 21521da177e4SLinus Torvalds if (!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_ACK)) { 21531da177e4SLinus Torvalds if (skb_cloned(skb)) { 21541da177e4SLinus Torvalds struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC); 21551da177e4SLinus Torvalds if (nskb == NULL) 21561da177e4SLinus Torvalds return -ENOMEM; 2157fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 21581da177e4SLinus Torvalds skb_header_release(nskb); 2159fe067e8aSDavid S. Miller __tcp_add_write_queue_head(sk, nskb); 21603ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 21613ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 21623ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 21631da177e4SLinus Torvalds skb = nskb; 21641da177e4SLinus Torvalds } 21651da177e4SLinus Torvalds 21661da177e4SLinus Torvalds TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ACK; 21671da177e4SLinus Torvalds TCP_ECN_send_synack(tcp_sk(sk), skb); 21681da177e4SLinus Torvalds } 21691da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2170dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 21711da177e4SLinus Torvalds } 21721da177e4SLinus Torvalds 21731da177e4SLinus Torvalds /* 21741da177e4SLinus Torvalds * Prepare a SYN-ACK. 21751da177e4SLinus Torvalds */ 21761da177e4SLinus Torvalds struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst, 217760236fddSArnaldo Carvalho de Melo struct request_sock *req) 21781da177e4SLinus Torvalds { 21792e6599cbSArnaldo Carvalho de Melo struct inet_request_sock *ireq = inet_rsk(req); 21801da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 21811da177e4SLinus Torvalds struct tcphdr *th; 21821da177e4SLinus Torvalds int tcp_header_size; 21831da177e4SLinus Torvalds struct sk_buff *skb; 2184cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 2185cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 2186cfb6eeb4SYOSHIFUJI Hideaki __u8 *md5_hash_location; 2187cfb6eeb4SYOSHIFUJI Hideaki #endif 21881da177e4SLinus Torvalds 21891da177e4SLinus Torvalds skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC); 21901da177e4SLinus Torvalds if (skb == NULL) 21911da177e4SLinus Torvalds return NULL; 21921da177e4SLinus Torvalds 21931da177e4SLinus Torvalds /* Reserve space for headers. */ 21941da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 21951da177e4SLinus Torvalds 21961da177e4SLinus Torvalds skb->dst = dst_clone(dst); 21971da177e4SLinus Torvalds 21981da177e4SLinus Torvalds tcp_header_size = (sizeof(struct tcphdr) + TCPOLEN_MSS + 21992e6599cbSArnaldo Carvalho de Melo (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0) + 22002e6599cbSArnaldo Carvalho de Melo (ireq->wscale_ok ? TCPOLEN_WSCALE_ALIGNED : 0) + 22011da177e4SLinus Torvalds /* SACK_PERM is in the place of NOP NOP of TS */ 22022e6599cbSArnaldo Carvalho de Melo ((ireq->sack_ok && !ireq->tstamp_ok) ? TCPOLEN_SACKPERM_ALIGNED : 0)); 2203cfb6eeb4SYOSHIFUJI Hideaki 2204cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 2205cfb6eeb4SYOSHIFUJI Hideaki /* Are we doing MD5 on this segment? If so - make room for it */ 2206cfb6eeb4SYOSHIFUJI Hideaki md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req); 2207cfb6eeb4SYOSHIFUJI Hideaki if (md5) 2208cfb6eeb4SYOSHIFUJI Hideaki tcp_header_size += TCPOLEN_MD5SIG_ALIGNED; 2209cfb6eeb4SYOSHIFUJI Hideaki #endif 2210aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 2211aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 22121da177e4SLinus Torvalds 2213aa8223c7SArnaldo Carvalho de Melo th = tcp_hdr(skb); 22141da177e4SLinus Torvalds memset(th, 0, sizeof(struct tcphdr)); 22151da177e4SLinus Torvalds th->syn = 1; 22161da177e4SLinus Torvalds th->ack = 1; 22171da177e4SLinus Torvalds TCP_ECN_make_synack(req, th); 22181da177e4SLinus Torvalds th->source = inet_sk(sk)->sport; 22192e6599cbSArnaldo Carvalho de Melo th->dest = ireq->rmt_port; 2220e870a8efSIlpo Järvinen /* Setting of flags are superfluous here for callers (and ECE is 2221e870a8efSIlpo Järvinen * not even correctly set) 2222e870a8efSIlpo Järvinen */ 2223e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn, 2224e870a8efSIlpo Järvinen TCPCB_FLAG_SYN | TCPCB_FLAG_ACK); 22251da177e4SLinus Torvalds th->seq = htonl(TCP_SKB_CB(skb)->seq); 22262e6599cbSArnaldo Carvalho de Melo th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1); 22271da177e4SLinus Torvalds if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */ 22281da177e4SLinus Torvalds __u8 rcv_wscale; 22291da177e4SLinus Torvalds /* Set this up on the first call only */ 22301da177e4SLinus Torvalds req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW); 22311da177e4SLinus Torvalds /* tcp_full_space because it is guaranteed to be the first packet */ 22321da177e4SLinus Torvalds tcp_select_initial_window(tcp_full_space(sk), 22332e6599cbSArnaldo Carvalho de Melo dst_metric(dst, RTAX_ADVMSS) - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0), 22341da177e4SLinus Torvalds &req->rcv_wnd, 22351da177e4SLinus Torvalds &req->window_clamp, 22362e6599cbSArnaldo Carvalho de Melo ireq->wscale_ok, 22371da177e4SLinus Torvalds &rcv_wscale); 22382e6599cbSArnaldo Carvalho de Melo ireq->rcv_wscale = rcv_wscale; 22391da177e4SLinus Torvalds } 22401da177e4SLinus Torvalds 22411da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 2242600ff0c2SIlpo Järvinen th->window = htons(min(req->rcv_wnd, 65535U)); 22434dfc2817SFlorian Westphal #ifdef CONFIG_SYN_COOKIES 22444dfc2817SFlorian Westphal if (unlikely(req->cookie_ts)) 22454dfc2817SFlorian Westphal TCP_SKB_CB(skb)->when = cookie_init_timestamp(req); 22464dfc2817SFlorian Westphal else 22474dfc2817SFlorian Westphal #endif 22481da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2249df7a3b07SAl Viro tcp_syn_build_options((__be32 *)(th + 1), dst_metric(dst, RTAX_ADVMSS), ireq->tstamp_ok, 22502e6599cbSArnaldo Carvalho de Melo ireq->sack_ok, ireq->wscale_ok, ireq->rcv_wscale, 22511da177e4SLinus Torvalds TCP_SKB_CB(skb)->when, 2252cfb6eeb4SYOSHIFUJI Hideaki req->ts_recent, 2253cfb6eeb4SYOSHIFUJI Hideaki ( 2254cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 2255cfb6eeb4SYOSHIFUJI Hideaki md5 ? &md5_hash_location : 2256cfb6eeb4SYOSHIFUJI Hideaki #endif 2257cfb6eeb4SYOSHIFUJI Hideaki NULL) 2258cfb6eeb4SYOSHIFUJI Hideaki ); 22591da177e4SLinus Torvalds 22601da177e4SLinus Torvalds th->doff = (tcp_header_size >> 2); 22611da177e4SLinus Torvalds TCP_INC_STATS(TCP_MIB_OUTSEGS); 2262cfb6eeb4SYOSHIFUJI Hideaki 2263cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 2264cfb6eeb4SYOSHIFUJI Hideaki /* Okay, we have all we need - do the md5 hash if needed */ 2265cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 2266cfb6eeb4SYOSHIFUJI Hideaki tp->af_specific->calc_md5_hash(md5_hash_location, 2267cfb6eeb4SYOSHIFUJI Hideaki md5, 2268cfb6eeb4SYOSHIFUJI Hideaki NULL, dst, req, 2269076fb722SYOSHIFUJI Hideaki tcp_hdr(skb), 2270cfb6eeb4SYOSHIFUJI Hideaki skb->len); 2271cfb6eeb4SYOSHIFUJI Hideaki } 2272cfb6eeb4SYOSHIFUJI Hideaki #endif 2273cfb6eeb4SYOSHIFUJI Hideaki 22741da177e4SLinus Torvalds return skb; 22751da177e4SLinus Torvalds } 22761da177e4SLinus Torvalds 22771da177e4SLinus Torvalds /* 22781da177e4SLinus Torvalds * Do all connect socket setups that can be done AF independent. 22791da177e4SLinus Torvalds */ 228040efc6faSStephen Hemminger static void tcp_connect_init(struct sock *sk) 22811da177e4SLinus Torvalds { 22821da177e4SLinus Torvalds struct dst_entry *dst = __sk_dst_get(sk); 22831da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 22841da177e4SLinus Torvalds __u8 rcv_wscale; 22851da177e4SLinus Torvalds 22861da177e4SLinus Torvalds /* We'll fix this up when we get a response from the other end. 22871da177e4SLinus Torvalds * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 22881da177e4SLinus Torvalds */ 22891da177e4SLinus Torvalds tp->tcp_header_len = sizeof(struct tcphdr) + 22901da177e4SLinus Torvalds (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0); 22911da177e4SLinus Torvalds 2292cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 2293cfb6eeb4SYOSHIFUJI Hideaki if (tp->af_specific->md5_lookup(sk, sk) != NULL) 2294cfb6eeb4SYOSHIFUJI Hideaki tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 2295cfb6eeb4SYOSHIFUJI Hideaki #endif 2296cfb6eeb4SYOSHIFUJI Hideaki 22971da177e4SLinus Torvalds /* If user gave his TCP_MAXSEG, record it to clamp */ 22981da177e4SLinus Torvalds if (tp->rx_opt.user_mss) 22991da177e4SLinus Torvalds tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 23001da177e4SLinus Torvalds tp->max_window = 0; 23015d424d5aSJohn Heffner tcp_mtup_init(sk); 23021da177e4SLinus Torvalds tcp_sync_mss(sk, dst_mtu(dst)); 23031da177e4SLinus Torvalds 23041da177e4SLinus Torvalds if (!tp->window_clamp) 23051da177e4SLinus Torvalds tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 23061da177e4SLinus Torvalds tp->advmss = dst_metric(dst, RTAX_ADVMSS); 23071da177e4SLinus Torvalds tcp_initialize_rcv_mss(sk); 23081da177e4SLinus Torvalds 23091da177e4SLinus Torvalds tcp_select_initial_window(tcp_full_space(sk), 23101da177e4SLinus Torvalds tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 23111da177e4SLinus Torvalds &tp->rcv_wnd, 23121da177e4SLinus Torvalds &tp->window_clamp, 23131da177e4SLinus Torvalds sysctl_tcp_window_scaling, 23141da177e4SLinus Torvalds &rcv_wscale); 23151da177e4SLinus Torvalds 23161da177e4SLinus Torvalds tp->rx_opt.rcv_wscale = rcv_wscale; 23171da177e4SLinus Torvalds tp->rcv_ssthresh = tp->rcv_wnd; 23181da177e4SLinus Torvalds 23191da177e4SLinus Torvalds sk->sk_err = 0; 23201da177e4SLinus Torvalds sock_reset_flag(sk, SOCK_DONE); 23211da177e4SLinus Torvalds tp->snd_wnd = 0; 23221da177e4SLinus Torvalds tcp_init_wl(tp, tp->write_seq, 0); 23231da177e4SLinus Torvalds tp->snd_una = tp->write_seq; 23241da177e4SLinus Torvalds tp->snd_sml = tp->write_seq; 23251da177e4SLinus Torvalds tp->rcv_nxt = 0; 23261da177e4SLinus Torvalds tp->rcv_wup = 0; 23271da177e4SLinus Torvalds tp->copied_seq = 0; 23281da177e4SLinus Torvalds 2329463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; 2330463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_retransmits = 0; 23311da177e4SLinus Torvalds tcp_clear_retrans(tp); 23321da177e4SLinus Torvalds } 23331da177e4SLinus Torvalds 23341da177e4SLinus Torvalds /* 23351da177e4SLinus Torvalds * Build a SYN and send it off. 23361da177e4SLinus Torvalds */ 23371da177e4SLinus Torvalds int tcp_connect(struct sock *sk) 23381da177e4SLinus Torvalds { 23391da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 23401da177e4SLinus Torvalds struct sk_buff *buff; 23411da177e4SLinus Torvalds 23421da177e4SLinus Torvalds tcp_connect_init(sk); 23431da177e4SLinus Torvalds 2344d179cd12SDavid S. Miller buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation); 23451da177e4SLinus Torvalds if (unlikely(buff == NULL)) 23461da177e4SLinus Torvalds return -ENOBUFS; 23471da177e4SLinus Torvalds 23481da177e4SLinus Torvalds /* Reserve space for headers. */ 23491da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 23501da177e4SLinus Torvalds 2351bd37a088SWei Yongjun tp->snd_nxt = tp->write_seq; 2352e870a8efSIlpo Järvinen tcp_init_nondata_skb(buff, tp->write_seq++, TCPCB_FLAG_SYN); 2353e870a8efSIlpo Järvinen TCP_ECN_send_syn(sk, buff); 23541da177e4SLinus Torvalds 23551da177e4SLinus Torvalds /* Send it off. */ 23561da177e4SLinus Torvalds TCP_SKB_CB(buff)->when = tcp_time_stamp; 23571da177e4SLinus Torvalds tp->retrans_stamp = TCP_SKB_CB(buff)->when; 23581da177e4SLinus Torvalds skb_header_release(buff); 2359fe067e8aSDavid S. Miller __tcp_add_write_queue_tail(sk, buff); 23603ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 23613ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 23621da177e4SLinus Torvalds tp->packets_out += tcp_skb_pcount(buff); 2363dfb4b9dcSDavid S. Miller tcp_transmit_skb(sk, buff, 1, GFP_KERNEL); 2364bd37a088SWei Yongjun 2365bd37a088SWei Yongjun /* We change tp->snd_nxt after the tcp_transmit_skb() call 2366bd37a088SWei Yongjun * in order to make this packet get counted in tcpOutSegs. 2367bd37a088SWei Yongjun */ 2368bd37a088SWei Yongjun tp->snd_nxt = tp->write_seq; 2369bd37a088SWei Yongjun tp->pushed_seq = tp->write_seq; 23701da177e4SLinus Torvalds TCP_INC_STATS(TCP_MIB_ACTIVEOPENS); 23711da177e4SLinus Torvalds 23721da177e4SLinus Torvalds /* Timer for repeating the SYN until an answer. */ 23733f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 23743f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 23751da177e4SLinus Torvalds return 0; 23761da177e4SLinus Torvalds } 23771da177e4SLinus Torvalds 23781da177e4SLinus Torvalds /* Send out a delayed ack, the caller does the policy checking 23791da177e4SLinus Torvalds * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 23801da177e4SLinus Torvalds * for details. 23811da177e4SLinus Torvalds */ 23821da177e4SLinus Torvalds void tcp_send_delayed_ack(struct sock *sk) 23831da177e4SLinus Torvalds { 2384463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 2385463c84b9SArnaldo Carvalho de Melo int ato = icsk->icsk_ack.ato; 23861da177e4SLinus Torvalds unsigned long timeout; 23871da177e4SLinus Torvalds 23881da177e4SLinus Torvalds if (ato > TCP_DELACK_MIN) { 2389463c84b9SArnaldo Carvalho de Melo const struct tcp_sock *tp = tcp_sk(sk); 23901da177e4SLinus Torvalds int max_ato = HZ / 2; 23911da177e4SLinus Torvalds 2392056834d9SIlpo Järvinen if (icsk->icsk_ack.pingpong || 2393056834d9SIlpo Järvinen (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 23941da177e4SLinus Torvalds max_ato = TCP_DELACK_MAX; 23951da177e4SLinus Torvalds 23961da177e4SLinus Torvalds /* Slow path, intersegment interval is "high". */ 23971da177e4SLinus Torvalds 23981da177e4SLinus Torvalds /* If some rtt estimate is known, use it to bound delayed ack. 2399463c84b9SArnaldo Carvalho de Melo * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 24001da177e4SLinus Torvalds * directly. 24011da177e4SLinus Torvalds */ 24021da177e4SLinus Torvalds if (tp->srtt) { 24031da177e4SLinus Torvalds int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN); 24041da177e4SLinus Torvalds 24051da177e4SLinus Torvalds if (rtt < max_ato) 24061da177e4SLinus Torvalds max_ato = rtt; 24071da177e4SLinus Torvalds } 24081da177e4SLinus Torvalds 24091da177e4SLinus Torvalds ato = min(ato, max_ato); 24101da177e4SLinus Torvalds } 24111da177e4SLinus Torvalds 24121da177e4SLinus Torvalds /* Stay within the limit we were given */ 24131da177e4SLinus Torvalds timeout = jiffies + ato; 24141da177e4SLinus Torvalds 24151da177e4SLinus Torvalds /* Use new timeout only if there wasn't a older one earlier. */ 2416463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 24171da177e4SLinus Torvalds /* If delack timer was blocked or is about to expire, 24181da177e4SLinus Torvalds * send ACK now. 24191da177e4SLinus Torvalds */ 2420463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.blocked || 2421463c84b9SArnaldo Carvalho de Melo time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 24221da177e4SLinus Torvalds tcp_send_ack(sk); 24231da177e4SLinus Torvalds return; 24241da177e4SLinus Torvalds } 24251da177e4SLinus Torvalds 2426463c84b9SArnaldo Carvalho de Melo if (!time_before(timeout, icsk->icsk_ack.timeout)) 2427463c84b9SArnaldo Carvalho de Melo timeout = icsk->icsk_ack.timeout; 24281da177e4SLinus Torvalds } 2429463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 2430463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.timeout = timeout; 2431463c84b9SArnaldo Carvalho de Melo sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 24321da177e4SLinus Torvalds } 24331da177e4SLinus Torvalds 24341da177e4SLinus Torvalds /* This routine sends an ack and also updates the window. */ 24351da177e4SLinus Torvalds void tcp_send_ack(struct sock *sk) 24361da177e4SLinus Torvalds { 24371da177e4SLinus Torvalds struct sk_buff *buff; 24381da177e4SLinus Torvalds 2439058dc334SIlpo Järvinen /* If we have been reset, we may not send again. */ 2440058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 2441058dc334SIlpo Järvinen return; 2442058dc334SIlpo Järvinen 24431da177e4SLinus Torvalds /* We are not putting this on the write queue, so 24441da177e4SLinus Torvalds * tcp_transmit_skb() will set the ownership to this 24451da177e4SLinus Torvalds * sock. 24461da177e4SLinus Torvalds */ 24471da177e4SLinus Torvalds buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 24481da177e4SLinus Torvalds if (buff == NULL) { 2449463c84b9SArnaldo Carvalho de Melo inet_csk_schedule_ack(sk); 2450463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; 24513f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 24523f421baaSArnaldo Carvalho de Melo TCP_DELACK_MAX, TCP_RTO_MAX); 24531da177e4SLinus Torvalds return; 24541da177e4SLinus Torvalds } 24551da177e4SLinus Torvalds 24561da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 24571da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 2458e870a8efSIlpo Järvinen tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPCB_FLAG_ACK); 24591da177e4SLinus Torvalds 24601da177e4SLinus Torvalds /* Send it off, this clears delayed acks for us. */ 24611da177e4SLinus Torvalds TCP_SKB_CB(buff)->when = tcp_time_stamp; 2462dfb4b9dcSDavid S. Miller tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC); 24631da177e4SLinus Torvalds } 24641da177e4SLinus Torvalds 24651da177e4SLinus Torvalds /* This routine sends a packet with an out of date sequence 24661da177e4SLinus Torvalds * number. It assumes the other end will try to ack it. 24671da177e4SLinus Torvalds * 24681da177e4SLinus Torvalds * Question: what should we make while urgent mode? 24691da177e4SLinus Torvalds * 4.4BSD forces sending single byte of data. We cannot send 24701da177e4SLinus Torvalds * out of window data, because we have SND.NXT==SND.MAX... 24711da177e4SLinus Torvalds * 24721da177e4SLinus Torvalds * Current solution: to send TWO zero-length segments in urgent mode: 24731da177e4SLinus Torvalds * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 24741da177e4SLinus Torvalds * out-of-date with SND.UNA-1 to probe window. 24751da177e4SLinus Torvalds */ 24761da177e4SLinus Torvalds static int tcp_xmit_probe_skb(struct sock *sk, int urgent) 24771da177e4SLinus Torvalds { 24781da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 24791da177e4SLinus Torvalds struct sk_buff *skb; 24801da177e4SLinus Torvalds 24811da177e4SLinus Torvalds /* We don't queue it, tcp_transmit_skb() sets ownership. */ 24821da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 24831da177e4SLinus Torvalds if (skb == NULL) 24841da177e4SLinus Torvalds return -1; 24851da177e4SLinus Torvalds 24861da177e4SLinus Torvalds /* Reserve space for headers and set control bits. */ 24871da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 24881da177e4SLinus Torvalds /* Use a previous sequence. This should cause the other 24891da177e4SLinus Torvalds * end to send an ack. Don't queue or clone SKB, just 24901da177e4SLinus Torvalds * send it. 24911da177e4SLinus Torvalds */ 2492e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPCB_FLAG_ACK); 24931da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2494dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC); 24951da177e4SLinus Torvalds } 24961da177e4SLinus Torvalds 24971da177e4SLinus Torvalds int tcp_write_wakeup(struct sock *sk) 24981da177e4SLinus Torvalds { 24991da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 25001da177e4SLinus Torvalds struct sk_buff *skb; 25011da177e4SLinus Torvalds 2502058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 2503058dc334SIlpo Järvinen return -1; 2504058dc334SIlpo Järvinen 2505fe067e8aSDavid S. Miller if ((skb = tcp_send_head(sk)) != NULL && 250690840defSIlpo Järvinen before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { 25071da177e4SLinus Torvalds int err; 25081da177e4SLinus Torvalds unsigned int mss = tcp_current_mss(sk, 0); 250990840defSIlpo Järvinen unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 25101da177e4SLinus Torvalds 25111da177e4SLinus Torvalds if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 25121da177e4SLinus Torvalds tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 25131da177e4SLinus Torvalds 25141da177e4SLinus Torvalds /* We are probing the opening of a window 25151da177e4SLinus Torvalds * but the window size is != 0 25161da177e4SLinus Torvalds * must have been a result SWS avoidance ( sender ) 25171da177e4SLinus Torvalds */ 25181da177e4SLinus Torvalds if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 25191da177e4SLinus Torvalds skb->len > mss) { 25201da177e4SLinus Torvalds seg_size = min(seg_size, mss); 25211da177e4SLinus Torvalds TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH; 2522846998aeSDavid S. Miller if (tcp_fragment(sk, skb, seg_size, mss)) 25231da177e4SLinus Torvalds return -1; 25241da177e4SLinus Torvalds } else if (!tcp_skb_pcount(skb)) 2525846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss); 25261da177e4SLinus Torvalds 25271da177e4SLinus Torvalds TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH; 25281da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2529dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 253066f5fe62SIlpo Järvinen if (!err) 253166f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 25321da177e4SLinus Torvalds return err; 25331da177e4SLinus Torvalds } else { 25341da177e4SLinus Torvalds if (tp->urg_mode && 25351da177e4SLinus Torvalds between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) 25364828e7f4SIlpo Järvinen tcp_xmit_probe_skb(sk, 1); 25371da177e4SLinus Torvalds return tcp_xmit_probe_skb(sk, 0); 25381da177e4SLinus Torvalds } 25391da177e4SLinus Torvalds } 25401da177e4SLinus Torvalds 25411da177e4SLinus Torvalds /* A window probe timeout has occurred. If window is not closed send 25421da177e4SLinus Torvalds * a partial packet else a zero probe. 25431da177e4SLinus Torvalds */ 25441da177e4SLinus Torvalds void tcp_send_probe0(struct sock *sk) 25451da177e4SLinus Torvalds { 2546463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 25471da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 25481da177e4SLinus Torvalds int err; 25491da177e4SLinus Torvalds 25501da177e4SLinus Torvalds err = tcp_write_wakeup(sk); 25511da177e4SLinus Torvalds 2552fe067e8aSDavid S. Miller if (tp->packets_out || !tcp_send_head(sk)) { 25531da177e4SLinus Torvalds /* Cancel probe timer, if it is not required. */ 25546687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 0; 2555463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff = 0; 25561da177e4SLinus Torvalds return; 25571da177e4SLinus Torvalds } 25581da177e4SLinus Torvalds 25591da177e4SLinus Torvalds if (err <= 0) { 2560463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_backoff < sysctl_tcp_retries2) 2561463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff++; 25626687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out++; 2563463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 25643f421baaSArnaldo Carvalho de Melo min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), 25653f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 25661da177e4SLinus Torvalds } else { 25671da177e4SLinus Torvalds /* If packet was not sent due to local congestion, 25686687e988SArnaldo Carvalho de Melo * do not backoff and do not remember icsk_probes_out. 25691da177e4SLinus Torvalds * Let local senders to fight for local resources. 25701da177e4SLinus Torvalds * 25711da177e4SLinus Torvalds * Use accumulated backoff yet. 25721da177e4SLinus Torvalds */ 25736687e988SArnaldo Carvalho de Melo if (!icsk->icsk_probes_out) 25746687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 1; 2575463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 2576463c84b9SArnaldo Carvalho de Melo min(icsk->icsk_rto << icsk->icsk_backoff, 25773f421baaSArnaldo Carvalho de Melo TCP_RESOURCE_PROBE_INTERVAL), 25783f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 25791da177e4SLinus Torvalds } 25801da177e4SLinus Torvalds } 25811da177e4SLinus Torvalds 2582c6aefafbSGlenn Griffin EXPORT_SYMBOL(tcp_select_initial_window); 25831da177e4SLinus Torvalds EXPORT_SYMBOL(tcp_connect); 25841da177e4SLinus Torvalds EXPORT_SYMBOL(tcp_make_synack); 25851da177e4SLinus Torvalds EXPORT_SYMBOL(tcp_simple_retransmit); 25861da177e4SLinus Torvalds EXPORT_SYMBOL(tcp_sync_mss); 25875d424d5aSJohn Heffner EXPORT_SYMBOL(tcp_mtup_init); 2588