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> 405a0e3ad6STejun Heo #include <linux/gfp.h> 411da177e4SLinus Torvalds #include <linux/module.h> 421da177e4SLinus Torvalds 431da177e4SLinus Torvalds /* People can turn this off for buggy TCP's found in printers etc. */ 44ab32ea5dSBrian Haley int sysctl_tcp_retrans_collapse __read_mostly = 1; 451da177e4SLinus Torvalds 4615d99e02SRick Jones /* People can turn this on to work with those rare, broken TCPs that 4715d99e02SRick Jones * interpret the window field as a signed quantity. 4815d99e02SRick Jones */ 49ab32ea5dSBrian Haley int sysctl_tcp_workaround_signed_windows __read_mostly = 0; 5015d99e02SRick Jones 511da177e4SLinus Torvalds /* This limits the percentage of the congestion window which we 521da177e4SLinus Torvalds * will allow a single TSO frame to consume. Building TSO frames 531da177e4SLinus Torvalds * which are too large can cause TCP streams to be bursty. 541da177e4SLinus Torvalds */ 55ab32ea5dSBrian Haley int sysctl_tcp_tso_win_divisor __read_mostly = 3; 561da177e4SLinus Torvalds 57ab32ea5dSBrian Haley int sysctl_tcp_mtu_probing __read_mostly = 0; 58ab32ea5dSBrian Haley int sysctl_tcp_base_mss __read_mostly = 512; 595d424d5aSJohn Heffner 6035089bb2SDavid S. Miller /* By default, RFC2861 behavior. */ 61ab32ea5dSBrian Haley int sysctl_tcp_slow_start_after_idle __read_mostly = 1; 6235089bb2SDavid S. Miller 63519855c5SWilliam Allen Simpson int sysctl_tcp_cookie_size __read_mostly = 0; /* TCP_COOKIE_MAX */ 64e6b09ccaSDavid S. Miller EXPORT_SYMBOL_GPL(sysctl_tcp_cookie_size); 65519855c5SWilliam Allen Simpson 66519855c5SWilliam Allen Simpson 6767edfef7SAndi Kleen /* Account for new data that has been sent to the network. */ 6866f5fe62SIlpo Järvinen static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb) 696ff03ac3SIlpo Järvinen { 706ff03ac3SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 7166f5fe62SIlpo Järvinen unsigned int prior_packets = tp->packets_out; 729e412ba7SIlpo Järvinen 73fe067e8aSDavid S. Miller tcp_advance_send_head(sk, skb); 741da177e4SLinus Torvalds tp->snd_nxt = TCP_SKB_CB(skb)->end_seq; 758512430eSIlpo Järvinen 768512430eSIlpo Järvinen /* Don't override Nagle indefinately with F-RTO */ 778512430eSIlpo Järvinen if (tp->frto_counter == 2) 788512430eSIlpo Järvinen tp->frto_counter = 3; 7966f5fe62SIlpo Järvinen 8066f5fe62SIlpo Järvinen tp->packets_out += tcp_skb_pcount(skb); 8166f5fe62SIlpo Järvinen if (!prior_packets) 8266f5fe62SIlpo Järvinen inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 8366f5fe62SIlpo Järvinen inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 841da177e4SLinus Torvalds } 851da177e4SLinus Torvalds 861da177e4SLinus Torvalds /* SND.NXT, if window was not shrunk. 871da177e4SLinus Torvalds * If window has been shrunk, what should we make? It is not clear at all. 881da177e4SLinus Torvalds * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-( 891da177e4SLinus Torvalds * Anything in between SND.UNA...SND.UNA+SND.WND also can be already 901da177e4SLinus Torvalds * invalid. OK, let's make this for now: 911da177e4SLinus Torvalds */ 929e412ba7SIlpo Järvinen static inline __u32 tcp_acceptable_seq(struct sock *sk) 931da177e4SLinus Torvalds { 949e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 959e412ba7SIlpo Järvinen 9690840defSIlpo Järvinen if (!before(tcp_wnd_end(tp), tp->snd_nxt)) 971da177e4SLinus Torvalds return tp->snd_nxt; 981da177e4SLinus Torvalds else 9990840defSIlpo Järvinen return tcp_wnd_end(tp); 1001da177e4SLinus Torvalds } 1011da177e4SLinus Torvalds 1021da177e4SLinus Torvalds /* Calculate mss to advertise in SYN segment. 1031da177e4SLinus Torvalds * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that: 1041da177e4SLinus Torvalds * 1051da177e4SLinus Torvalds * 1. It is independent of path mtu. 1061da177e4SLinus Torvalds * 2. Ideally, it is maximal possible segment size i.e. 65535-40. 1071da177e4SLinus Torvalds * 3. For IPv4 it is reasonable to calculate it from maximal MTU of 1081da177e4SLinus Torvalds * attached devices, because some buggy hosts are confused by 1091da177e4SLinus Torvalds * large MSS. 1101da177e4SLinus Torvalds * 4. We do not make 3, we advertise MSS, calculated from first 1111da177e4SLinus Torvalds * hop device mtu, but allow to raise it to ip_rt_min_advmss. 1121da177e4SLinus Torvalds * This may be overridden via information stored in routing table. 1131da177e4SLinus Torvalds * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible, 1141da177e4SLinus Torvalds * probably even Jumbo". 1151da177e4SLinus Torvalds */ 1161da177e4SLinus Torvalds static __u16 tcp_advertise_mss(struct sock *sk) 1171da177e4SLinus Torvalds { 1181da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1191da177e4SLinus Torvalds struct dst_entry *dst = __sk_dst_get(sk); 1201da177e4SLinus Torvalds int mss = tp->advmss; 1211da177e4SLinus Torvalds 1221da177e4SLinus Torvalds if (dst && dst_metric(dst, RTAX_ADVMSS) < mss) { 1231da177e4SLinus Torvalds mss = dst_metric(dst, RTAX_ADVMSS); 1241da177e4SLinus Torvalds tp->advmss = mss; 1251da177e4SLinus Torvalds } 1261da177e4SLinus Torvalds 1271da177e4SLinus Torvalds return (__u16)mss; 1281da177e4SLinus Torvalds } 1291da177e4SLinus Torvalds 1301da177e4SLinus Torvalds /* RFC2861. Reset CWND after idle period longer RTO to "restart window". 1311da177e4SLinus Torvalds * This is the first part of cwnd validation mechanism. */ 132463c84b9SArnaldo Carvalho de Melo static void tcp_cwnd_restart(struct sock *sk, struct dst_entry *dst) 1331da177e4SLinus Torvalds { 134463c84b9SArnaldo Carvalho de Melo struct tcp_sock *tp = tcp_sk(sk); 1351da177e4SLinus Torvalds s32 delta = tcp_time_stamp - tp->lsndtime; 1361da177e4SLinus Torvalds u32 restart_cwnd = tcp_init_cwnd(tp, dst); 1371da177e4SLinus Torvalds u32 cwnd = tp->snd_cwnd; 1381da177e4SLinus Torvalds 1396687e988SArnaldo Carvalho de Melo tcp_ca_event(sk, CA_EVENT_CWND_RESTART); 1401da177e4SLinus Torvalds 1416687e988SArnaldo Carvalho de Melo tp->snd_ssthresh = tcp_current_ssthresh(sk); 1421da177e4SLinus Torvalds restart_cwnd = min(restart_cwnd, cwnd); 1431da177e4SLinus Torvalds 144463c84b9SArnaldo Carvalho de Melo while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd) 1451da177e4SLinus Torvalds cwnd >>= 1; 1461da177e4SLinus Torvalds tp->snd_cwnd = max(cwnd, restart_cwnd); 1471da177e4SLinus Torvalds tp->snd_cwnd_stamp = tcp_time_stamp; 1481da177e4SLinus Torvalds tp->snd_cwnd_used = 0; 1491da177e4SLinus Torvalds } 1501da177e4SLinus Torvalds 15167edfef7SAndi Kleen /* Congestion state accounting after a packet has been sent. */ 15240efc6faSStephen Hemminger static void tcp_event_data_sent(struct tcp_sock *tp, 1531da177e4SLinus Torvalds struct sk_buff *skb, struct sock *sk) 1541da177e4SLinus Torvalds { 155463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 156463c84b9SArnaldo Carvalho de Melo const u32 now = tcp_time_stamp; 1571da177e4SLinus Torvalds 15835089bb2SDavid S. Miller if (sysctl_tcp_slow_start_after_idle && 15935089bb2SDavid S. Miller (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto)) 160463c84b9SArnaldo Carvalho de Melo tcp_cwnd_restart(sk, __sk_dst_get(sk)); 1611da177e4SLinus Torvalds 1621da177e4SLinus Torvalds tp->lsndtime = now; 1631da177e4SLinus Torvalds 1641da177e4SLinus Torvalds /* If it is a reply for ato after last received 1651da177e4SLinus Torvalds * packet, enter pingpong mode. 1661da177e4SLinus Torvalds */ 167463c84b9SArnaldo Carvalho de Melo if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato) 168463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pingpong = 1; 1691da177e4SLinus Torvalds } 1701da177e4SLinus Torvalds 17167edfef7SAndi Kleen /* Account for an ACK we sent. */ 17240efc6faSStephen Hemminger static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts) 1731da177e4SLinus Torvalds { 174463c84b9SArnaldo Carvalho de Melo tcp_dec_quickack_mode(sk, pkts); 175463c84b9SArnaldo Carvalho de Melo inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); 1761da177e4SLinus Torvalds } 1771da177e4SLinus Torvalds 1781da177e4SLinus Torvalds /* Determine a window scaling and initial window to offer. 1791da177e4SLinus Torvalds * Based on the assumption that the given amount of space 1801da177e4SLinus Torvalds * will be offered. Store the results in the tp structure. 1811da177e4SLinus Torvalds * NOTE: for smooth operation initial space offering should 1821da177e4SLinus Torvalds * be a multiple of mss if possible. We assume here that mss >= 1. 1831da177e4SLinus Torvalds * This MUST be enforced by all callers. 1841da177e4SLinus Torvalds */ 1851da177e4SLinus Torvalds void tcp_select_initial_window(int __space, __u32 mss, 1861da177e4SLinus Torvalds __u32 *rcv_wnd, __u32 *window_clamp, 18731d12926Slaurent chavey int wscale_ok, __u8 *rcv_wscale, 18831d12926Slaurent chavey __u32 init_rcv_wnd) 1891da177e4SLinus Torvalds { 1901da177e4SLinus Torvalds unsigned int space = (__space < 0 ? 0 : __space); 1911da177e4SLinus Torvalds 1921da177e4SLinus Torvalds /* If no clamp set the clamp to the max possible scaled window */ 1931da177e4SLinus Torvalds if (*window_clamp == 0) 1941da177e4SLinus Torvalds (*window_clamp) = (65535 << 14); 1951da177e4SLinus Torvalds space = min(*window_clamp, space); 1961da177e4SLinus Torvalds 1971da177e4SLinus Torvalds /* Quantize space offering to a multiple of mss if possible. */ 1981da177e4SLinus Torvalds if (space > mss) 1991da177e4SLinus Torvalds space = (space / mss) * mss; 2001da177e4SLinus Torvalds 2011da177e4SLinus Torvalds /* NOTE: offering an initial window larger than 32767 20215d99e02SRick Jones * will break some buggy TCP stacks. If the admin tells us 20315d99e02SRick Jones * it is likely we could be speaking with such a buggy stack 20415d99e02SRick Jones * we will truncate our initial window offering to 32K-1 20515d99e02SRick Jones * unless the remote has sent us a window scaling option, 20615d99e02SRick Jones * which we interpret as a sign the remote TCP is not 20715d99e02SRick Jones * misinterpreting the window field as a signed quantity. 2081da177e4SLinus Torvalds */ 20915d99e02SRick Jones if (sysctl_tcp_workaround_signed_windows) 2101da177e4SLinus Torvalds (*rcv_wnd) = min(space, MAX_TCP_WINDOW); 21115d99e02SRick Jones else 21215d99e02SRick Jones (*rcv_wnd) = space; 21315d99e02SRick Jones 2141da177e4SLinus Torvalds (*rcv_wscale) = 0; 2151da177e4SLinus Torvalds if (wscale_ok) { 2161da177e4SLinus Torvalds /* Set window scaling on max possible window 2171da177e4SLinus Torvalds * See RFC1323 for an explanation of the limit to 14 2181da177e4SLinus Torvalds */ 2191da177e4SLinus Torvalds space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max); 220316c1592SStephen Hemminger space = min_t(u32, space, *window_clamp); 2211da177e4SLinus Torvalds while (space > 65535 && (*rcv_wscale) < 14) { 2221da177e4SLinus Torvalds space >>= 1; 2231da177e4SLinus Torvalds (*rcv_wscale)++; 2241da177e4SLinus Torvalds } 2251da177e4SLinus Torvalds } 2261da177e4SLinus Torvalds 2270705c6f0SGerrit Renker /* Set initial window to value enough for senders, following RFC5681. */ 2281da177e4SLinus Torvalds if (mss > (1 << *rcv_wscale)) { 2290705c6f0SGerrit Renker int init_cwnd = rfc3390_bytes_to_packets(mss); 2300705c6f0SGerrit Renker 23131d12926Slaurent chavey /* when initializing use the value from init_rcv_wnd 23231d12926Slaurent chavey * rather than the default from above 23331d12926Slaurent chavey */ 234*b1afde60SNandita Dukkipati if (init_rcv_wnd) 235*b1afde60SNandita Dukkipati *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss); 236*b1afde60SNandita Dukkipati else 237*b1afde60SNandita Dukkipati *rcv_wnd = min(*rcv_wnd, init_cwnd * mss); 2381da177e4SLinus Torvalds } 2391da177e4SLinus Torvalds 2401da177e4SLinus Torvalds /* Set the clamp no higher than max representable value */ 2411da177e4SLinus Torvalds (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp); 2421da177e4SLinus Torvalds } 2434bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_select_initial_window); 2441da177e4SLinus Torvalds 2451da177e4SLinus Torvalds /* Chose a new window to advertise, update state in tcp_sock for the 2461da177e4SLinus Torvalds * socket, and return result with RFC1323 scaling applied. The return 2471da177e4SLinus Torvalds * value can be stuffed directly into th->window for an outgoing 2481da177e4SLinus Torvalds * frame. 2491da177e4SLinus Torvalds */ 25040efc6faSStephen Hemminger static u16 tcp_select_window(struct sock *sk) 2511da177e4SLinus Torvalds { 2521da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2531da177e4SLinus Torvalds u32 cur_win = tcp_receive_window(tp); 2541da177e4SLinus Torvalds u32 new_win = __tcp_select_window(sk); 2551da177e4SLinus Torvalds 2561da177e4SLinus Torvalds /* Never shrink the offered window */ 2571da177e4SLinus Torvalds if (new_win < cur_win) { 2581da177e4SLinus Torvalds /* Danger Will Robinson! 2591da177e4SLinus Torvalds * Don't update rcv_wup/rcv_wnd here or else 2601da177e4SLinus Torvalds * we will not be able to advertise a zero 2611da177e4SLinus Torvalds * window in time. --DaveM 2621da177e4SLinus Torvalds * 2631da177e4SLinus Torvalds * Relax Will Robinson. 2641da177e4SLinus Torvalds */ 265607bfbf2SPatrick McHardy new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale); 2661da177e4SLinus Torvalds } 2671da177e4SLinus Torvalds tp->rcv_wnd = new_win; 2681da177e4SLinus Torvalds tp->rcv_wup = tp->rcv_nxt; 2691da177e4SLinus Torvalds 2701da177e4SLinus Torvalds /* Make sure we do not exceed the maximum possible 2711da177e4SLinus Torvalds * scaled window. 2721da177e4SLinus Torvalds */ 27315d99e02SRick Jones if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows) 2741da177e4SLinus Torvalds new_win = min(new_win, MAX_TCP_WINDOW); 2751da177e4SLinus Torvalds else 2761da177e4SLinus Torvalds new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); 2771da177e4SLinus Torvalds 2781da177e4SLinus Torvalds /* RFC1323 scaling applied */ 2791da177e4SLinus Torvalds new_win >>= tp->rx_opt.rcv_wscale; 2801da177e4SLinus Torvalds 2811da177e4SLinus Torvalds /* If we advertise zero window, disable fast path. */ 2821da177e4SLinus Torvalds if (new_win == 0) 2831da177e4SLinus Torvalds tp->pred_flags = 0; 2841da177e4SLinus Torvalds 2851da177e4SLinus Torvalds return new_win; 2861da177e4SLinus Torvalds } 2871da177e4SLinus Torvalds 28867edfef7SAndi Kleen /* Packet ECN state for a SYN-ACK */ 289056834d9SIlpo Järvinen static inline void TCP_ECN_send_synack(struct tcp_sock *tp, struct sk_buff *skb) 290bdf1ee5dSIlpo Järvinen { 291a3433f35SChangli Gao TCP_SKB_CB(skb)->flags &= ~TCPHDR_CWR; 292bdf1ee5dSIlpo Järvinen if (!(tp->ecn_flags & TCP_ECN_OK)) 293a3433f35SChangli Gao TCP_SKB_CB(skb)->flags &= ~TCPHDR_ECE; 294bdf1ee5dSIlpo Järvinen } 295bdf1ee5dSIlpo Järvinen 29667edfef7SAndi Kleen /* Packet ECN state for a SYN. */ 297bdf1ee5dSIlpo Järvinen static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb) 298bdf1ee5dSIlpo Järvinen { 299bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 300bdf1ee5dSIlpo Järvinen 301bdf1ee5dSIlpo Järvinen tp->ecn_flags = 0; 302255cac91SIlpo Järvinen if (sysctl_tcp_ecn == 1) { 303a3433f35SChangli Gao TCP_SKB_CB(skb)->flags |= TCPHDR_ECE | TCPHDR_CWR; 304bdf1ee5dSIlpo Järvinen tp->ecn_flags = TCP_ECN_OK; 305bdf1ee5dSIlpo Järvinen } 306bdf1ee5dSIlpo Järvinen } 307bdf1ee5dSIlpo Järvinen 308bdf1ee5dSIlpo Järvinen static __inline__ void 309bdf1ee5dSIlpo Järvinen TCP_ECN_make_synack(struct request_sock *req, struct tcphdr *th) 310bdf1ee5dSIlpo Järvinen { 311bdf1ee5dSIlpo Järvinen if (inet_rsk(req)->ecn_ok) 312bdf1ee5dSIlpo Järvinen th->ece = 1; 313bdf1ee5dSIlpo Järvinen } 314bdf1ee5dSIlpo Järvinen 31567edfef7SAndi Kleen /* Set up ECN state for a packet on a ESTABLISHED socket that is about to 31667edfef7SAndi Kleen * be sent. 31767edfef7SAndi Kleen */ 318bdf1ee5dSIlpo Järvinen static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb, 319bdf1ee5dSIlpo Järvinen int tcp_header_len) 320bdf1ee5dSIlpo Järvinen { 321bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 322bdf1ee5dSIlpo Järvinen 323bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_OK) { 324bdf1ee5dSIlpo Järvinen /* Not-retransmitted data segment: set ECT and inject CWR. */ 325bdf1ee5dSIlpo Järvinen if (skb->len != tcp_header_len && 326bdf1ee5dSIlpo Järvinen !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { 327bdf1ee5dSIlpo Järvinen INET_ECN_xmit(sk); 328bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) { 329bdf1ee5dSIlpo Järvinen tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; 330bdf1ee5dSIlpo Järvinen tcp_hdr(skb)->cwr = 1; 331bdf1ee5dSIlpo Järvinen skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; 332bdf1ee5dSIlpo Järvinen } 333bdf1ee5dSIlpo Järvinen } else { 334bdf1ee5dSIlpo Järvinen /* ACK or retransmitted segment: clear ECT|CE */ 335bdf1ee5dSIlpo Järvinen INET_ECN_dontxmit(sk); 336bdf1ee5dSIlpo Järvinen } 337bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) 338bdf1ee5dSIlpo Järvinen tcp_hdr(skb)->ece = 1; 339bdf1ee5dSIlpo Järvinen } 340bdf1ee5dSIlpo Järvinen } 341bdf1ee5dSIlpo Järvinen 342e870a8efSIlpo Järvinen /* Constructs common control bits of non-data skb. If SYN/FIN is present, 343e870a8efSIlpo Järvinen * auto increment end seqno. 344e870a8efSIlpo Järvinen */ 345e870a8efSIlpo Järvinen static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags) 346e870a8efSIlpo Järvinen { 3472e8e18efSDavid S. Miller skb->ip_summed = CHECKSUM_PARTIAL; 348e870a8efSIlpo Järvinen skb->csum = 0; 349e870a8efSIlpo Järvinen 350e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->flags = flags; 351e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->sacked = 0; 352e870a8efSIlpo Järvinen 353e870a8efSIlpo Järvinen skb_shinfo(skb)->gso_segs = 1; 354e870a8efSIlpo Järvinen skb_shinfo(skb)->gso_size = 0; 355e870a8efSIlpo Järvinen skb_shinfo(skb)->gso_type = 0; 356e870a8efSIlpo Järvinen 357e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->seq = seq; 358a3433f35SChangli Gao if (flags & (TCPHDR_SYN | TCPHDR_FIN)) 359e870a8efSIlpo Järvinen seq++; 360e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->end_seq = seq; 361e870a8efSIlpo Järvinen } 362e870a8efSIlpo Järvinen 36333f5f57eSIlpo Järvinen static inline int tcp_urg_mode(const struct tcp_sock *tp) 36433f5f57eSIlpo Järvinen { 36533f5f57eSIlpo Järvinen return tp->snd_una != tp->snd_up; 36633f5f57eSIlpo Järvinen } 36733f5f57eSIlpo Järvinen 36833ad798cSAdam Langley #define OPTION_SACK_ADVERTISE (1 << 0) 36933ad798cSAdam Langley #define OPTION_TS (1 << 1) 37033ad798cSAdam Langley #define OPTION_MD5 (1 << 2) 37189e95a61SOri Finkelman #define OPTION_WSCALE (1 << 3) 372bd0388aeSWilliam Allen Simpson #define OPTION_COOKIE_EXTENSION (1 << 4) 37333ad798cSAdam Langley 37433ad798cSAdam Langley struct tcp_out_options { 37533ad798cSAdam Langley u8 options; /* bit field of OPTION_* */ 37633ad798cSAdam Langley u8 ws; /* window scale, 0 to disable */ 37733ad798cSAdam Langley u8 num_sack_blocks; /* number of SACK blocks to include */ 378bd0388aeSWilliam Allen Simpson u8 hash_size; /* bytes in hash_location */ 37933ad798cSAdam Langley u16 mss; /* 0 to disable */ 38033ad798cSAdam Langley __u32 tsval, tsecr; /* need to include OPTION_TS */ 381bd0388aeSWilliam Allen Simpson __u8 *hash_location; /* temporary pointer, overloaded */ 38233ad798cSAdam Langley }; 38333ad798cSAdam Langley 384bd0388aeSWilliam Allen Simpson /* The sysctl int routines are generic, so check consistency here. 385bd0388aeSWilliam Allen Simpson */ 386bd0388aeSWilliam Allen Simpson static u8 tcp_cookie_size_check(u8 desired) 387bd0388aeSWilliam Allen Simpson { 388bd0388aeSWilliam Allen Simpson if (desired > 0) { 389bd0388aeSWilliam Allen Simpson /* previously specified */ 390bd0388aeSWilliam Allen Simpson return desired; 391bd0388aeSWilliam Allen Simpson } 392bd0388aeSWilliam Allen Simpson if (sysctl_tcp_cookie_size <= 0) { 393bd0388aeSWilliam Allen Simpson /* no default specified */ 394bd0388aeSWilliam Allen Simpson return 0; 395bd0388aeSWilliam Allen Simpson } 396bd0388aeSWilliam Allen Simpson if (sysctl_tcp_cookie_size <= TCP_COOKIE_MIN) { 397bd0388aeSWilliam Allen Simpson /* value too small, specify minimum */ 398bd0388aeSWilliam Allen Simpson return TCP_COOKIE_MIN; 399bd0388aeSWilliam Allen Simpson } 400bd0388aeSWilliam Allen Simpson if (sysctl_tcp_cookie_size >= TCP_COOKIE_MAX) { 401bd0388aeSWilliam Allen Simpson /* value too large, specify maximum */ 402bd0388aeSWilliam Allen Simpson return TCP_COOKIE_MAX; 403bd0388aeSWilliam Allen Simpson } 404bd0388aeSWilliam Allen Simpson if (0x1 & sysctl_tcp_cookie_size) { 405bd0388aeSWilliam Allen Simpson /* 8-bit multiple, illegal, fix it */ 406bd0388aeSWilliam Allen Simpson return (u8)(sysctl_tcp_cookie_size + 0x1); 407bd0388aeSWilliam Allen Simpson } 408bd0388aeSWilliam Allen Simpson return (u8)sysctl_tcp_cookie_size; 409bd0388aeSWilliam Allen Simpson } 410bd0388aeSWilliam Allen Simpson 41167edfef7SAndi Kleen /* Write previously computed TCP options to the packet. 41267edfef7SAndi Kleen * 41367edfef7SAndi Kleen * Beware: Something in the Internet is very sensitive to the ordering of 414fd6149d3SIlpo Järvinen * TCP options, we learned this through the hard way, so be careful here. 415fd6149d3SIlpo Järvinen * Luckily we can at least blame others for their non-compliance but from 416fd6149d3SIlpo Järvinen * inter-operatibility perspective it seems that we're somewhat stuck with 417fd6149d3SIlpo Järvinen * the ordering which we have been using if we want to keep working with 418fd6149d3SIlpo Järvinen * those broken things (not that it currently hurts anybody as there isn't 419fd6149d3SIlpo Järvinen * particular reason why the ordering would need to be changed). 420fd6149d3SIlpo Järvinen * 421fd6149d3SIlpo Järvinen * At least SACK_PERM as the first option is known to lead to a disaster 422fd6149d3SIlpo Järvinen * (but it may well be that other scenarios fail similarly). 423fd6149d3SIlpo Järvinen */ 42433ad798cSAdam Langley static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp, 425bd0388aeSWilliam Allen Simpson struct tcp_out_options *opts) 426bd0388aeSWilliam Allen Simpson { 427bd0388aeSWilliam Allen Simpson u8 options = opts->options; /* mungable copy */ 428bd0388aeSWilliam Allen Simpson 429bd0388aeSWilliam Allen Simpson /* Having both authentication and cookies for security is redundant, 430bd0388aeSWilliam Allen Simpson * and there's certainly not enough room. Instead, the cookie-less 431bd0388aeSWilliam Allen Simpson * extension variant is proposed. 432bd0388aeSWilliam Allen Simpson * 433bd0388aeSWilliam Allen Simpson * Consider the pessimal case with authentication. The options 434bd0388aeSWilliam Allen Simpson * could look like: 435bd0388aeSWilliam Allen Simpson * COOKIE|MD5(20) + MSS(4) + SACK|TS(12) + WSCALE(4) == 40 436bd0388aeSWilliam Allen Simpson */ 437bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_MD5 & options)) { 438bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_COOKIE_EXTENSION & options)) { 439bd0388aeSWilliam Allen Simpson *ptr++ = htonl((TCPOPT_COOKIE << 24) | 440bd0388aeSWilliam Allen Simpson (TCPOLEN_COOKIE_BASE << 16) | 441bd0388aeSWilliam Allen Simpson (TCPOPT_MD5SIG << 8) | 442bd0388aeSWilliam Allen Simpson TCPOLEN_MD5SIG); 443bd0388aeSWilliam Allen Simpson } else { 44433ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 44533ad798cSAdam Langley (TCPOPT_NOP << 16) | 44633ad798cSAdam Langley (TCPOPT_MD5SIG << 8) | 44733ad798cSAdam Langley TCPOLEN_MD5SIG); 448bd0388aeSWilliam Allen Simpson } 449bd0388aeSWilliam Allen Simpson options &= ~OPTION_COOKIE_EXTENSION; 450bd0388aeSWilliam Allen Simpson /* overload cookie hash location */ 451bd0388aeSWilliam Allen Simpson opts->hash_location = (__u8 *)ptr; 45233ad798cSAdam Langley ptr += 4; 45333ad798cSAdam Langley } 45433ad798cSAdam Langley 455fd6149d3SIlpo Järvinen if (unlikely(opts->mss)) { 456fd6149d3SIlpo Järvinen *ptr++ = htonl((TCPOPT_MSS << 24) | 457fd6149d3SIlpo Järvinen (TCPOLEN_MSS << 16) | 458fd6149d3SIlpo Järvinen opts->mss); 459fd6149d3SIlpo Järvinen } 460fd6149d3SIlpo Järvinen 461bd0388aeSWilliam Allen Simpson if (likely(OPTION_TS & options)) { 462bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 46333ad798cSAdam Langley *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | 46433ad798cSAdam Langley (TCPOLEN_SACK_PERM << 16) | 46533ad798cSAdam Langley (TCPOPT_TIMESTAMP << 8) | 46633ad798cSAdam Langley TCPOLEN_TIMESTAMP); 467bd0388aeSWilliam Allen Simpson options &= ~OPTION_SACK_ADVERTISE; 46833ad798cSAdam Langley } else { 469496c98dfSYOSHIFUJI Hideaki *ptr++ = htonl((TCPOPT_NOP << 24) | 47040efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 47140efc6faSStephen Hemminger (TCPOPT_TIMESTAMP << 8) | 47240efc6faSStephen Hemminger TCPOLEN_TIMESTAMP); 47340efc6faSStephen Hemminger } 47433ad798cSAdam Langley *ptr++ = htonl(opts->tsval); 47533ad798cSAdam Langley *ptr++ = htonl(opts->tsecr); 47633ad798cSAdam Langley } 47733ad798cSAdam Langley 478bd0388aeSWilliam Allen Simpson /* Specification requires after timestamp, so do it now. 479bd0388aeSWilliam Allen Simpson * 480bd0388aeSWilliam Allen Simpson * Consider the pessimal case without authentication. The options 481bd0388aeSWilliam Allen Simpson * could look like: 482bd0388aeSWilliam Allen Simpson * MSS(4) + SACK|TS(12) + COOKIE(20) + WSCALE(4) == 40 483bd0388aeSWilliam Allen Simpson */ 484bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_COOKIE_EXTENSION & options)) { 485bd0388aeSWilliam Allen Simpson __u8 *cookie_copy = opts->hash_location; 486bd0388aeSWilliam Allen Simpson u8 cookie_size = opts->hash_size; 487bd0388aeSWilliam Allen Simpson 488bd0388aeSWilliam Allen Simpson /* 8-bit multiple handled in tcp_cookie_size_check() above, 489bd0388aeSWilliam Allen Simpson * and elsewhere. 490bd0388aeSWilliam Allen Simpson */ 491bd0388aeSWilliam Allen Simpson if (0x2 & cookie_size) { 492bd0388aeSWilliam Allen Simpson __u8 *p = (__u8 *)ptr; 493bd0388aeSWilliam Allen Simpson 494bd0388aeSWilliam Allen Simpson /* 16-bit multiple */ 495bd0388aeSWilliam Allen Simpson *p++ = TCPOPT_COOKIE; 496bd0388aeSWilliam Allen Simpson *p++ = TCPOLEN_COOKIE_BASE + cookie_size; 497bd0388aeSWilliam Allen Simpson *p++ = *cookie_copy++; 498bd0388aeSWilliam Allen Simpson *p++ = *cookie_copy++; 499bd0388aeSWilliam Allen Simpson ptr++; 500bd0388aeSWilliam Allen Simpson cookie_size -= 2; 501bd0388aeSWilliam Allen Simpson } else { 502bd0388aeSWilliam Allen Simpson /* 32-bit multiple */ 503bd0388aeSWilliam Allen Simpson *ptr++ = htonl(((TCPOPT_NOP << 24) | 504bd0388aeSWilliam Allen Simpson (TCPOPT_NOP << 16) | 505bd0388aeSWilliam Allen Simpson (TCPOPT_COOKIE << 8) | 506bd0388aeSWilliam Allen Simpson TCPOLEN_COOKIE_BASE) + 507bd0388aeSWilliam Allen Simpson cookie_size); 508bd0388aeSWilliam Allen Simpson } 509bd0388aeSWilliam Allen Simpson 510bd0388aeSWilliam Allen Simpson if (cookie_size > 0) { 511bd0388aeSWilliam Allen Simpson memcpy(ptr, cookie_copy, cookie_size); 512bd0388aeSWilliam Allen Simpson ptr += (cookie_size / 4); 513bd0388aeSWilliam Allen Simpson } 514bd0388aeSWilliam Allen Simpson } 515bd0388aeSWilliam Allen Simpson 516bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 51733ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 51833ad798cSAdam Langley (TCPOPT_NOP << 16) | 51933ad798cSAdam Langley (TCPOPT_SACK_PERM << 8) | 52033ad798cSAdam Langley TCPOLEN_SACK_PERM); 52133ad798cSAdam Langley } 52233ad798cSAdam Langley 523bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_WSCALE & options)) { 52433ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 52533ad798cSAdam Langley (TCPOPT_WINDOW << 16) | 52633ad798cSAdam Langley (TCPOLEN_WINDOW << 8) | 52733ad798cSAdam Langley opts->ws); 52833ad798cSAdam Langley } 52933ad798cSAdam Langley 53033ad798cSAdam Langley if (unlikely(opts->num_sack_blocks)) { 53133ad798cSAdam Langley struct tcp_sack_block *sp = tp->rx_opt.dsack ? 53233ad798cSAdam Langley tp->duplicate_sack : tp->selective_acks; 53340efc6faSStephen Hemminger int this_sack; 53440efc6faSStephen Hemminger 53540efc6faSStephen Hemminger *ptr++ = htonl((TCPOPT_NOP << 24) | 53640efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 53740efc6faSStephen Hemminger (TCPOPT_SACK << 8) | 53833ad798cSAdam Langley (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * 53940efc6faSStephen Hemminger TCPOLEN_SACK_PERBLOCK))); 5402de979bdSStephen Hemminger 54133ad798cSAdam Langley for (this_sack = 0; this_sack < opts->num_sack_blocks; 54233ad798cSAdam Langley ++this_sack) { 54340efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].start_seq); 54440efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].end_seq); 54540efc6faSStephen Hemminger } 5462de979bdSStephen Hemminger 54740efc6faSStephen Hemminger tp->rx_opt.dsack = 0; 54840efc6faSStephen Hemminger } 54940efc6faSStephen Hemminger } 55040efc6faSStephen Hemminger 55167edfef7SAndi Kleen /* Compute TCP options for SYN packets. This is not the final 55267edfef7SAndi Kleen * network wire format yet. 55367edfef7SAndi Kleen */ 55433ad798cSAdam Langley static unsigned tcp_syn_options(struct sock *sk, struct sk_buff *skb, 55533ad798cSAdam Langley struct tcp_out_options *opts, 55633ad798cSAdam Langley struct tcp_md5sig_key **md5) { 55733ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 558bd0388aeSWilliam Allen Simpson struct tcp_cookie_values *cvp = tp->cookie_values; 559bd0388aeSWilliam Allen Simpson unsigned remaining = MAX_TCP_OPTION_SPACE; 560bd0388aeSWilliam Allen Simpson u8 cookie_size = (!tp->rx_opt.cookie_out_never && cvp != NULL) ? 561bd0388aeSWilliam Allen Simpson tcp_cookie_size_check(cvp->cookie_desired) : 562bd0388aeSWilliam Allen Simpson 0; 56333ad798cSAdam Langley 564cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 56533ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 56633ad798cSAdam Langley if (*md5) { 56733ad798cSAdam Langley opts->options |= OPTION_MD5; 568bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 569cfb6eeb4SYOSHIFUJI Hideaki } 57033ad798cSAdam Langley #else 57133ad798cSAdam Langley *md5 = NULL; 572cfb6eeb4SYOSHIFUJI Hideaki #endif 57333ad798cSAdam Langley 57433ad798cSAdam Langley /* We always get an MSS option. The option bytes which will be seen in 57533ad798cSAdam Langley * normal data packets should timestamps be used, must be in the MSS 57633ad798cSAdam Langley * advertised. But we subtract them from tp->mss_cache so that 57733ad798cSAdam Langley * calculations in tcp_sendmsg are simpler etc. So account for this 57833ad798cSAdam Langley * fact here if necessary. If we don't do this correctly, as a 57933ad798cSAdam Langley * receiver we won't recognize data packets as being full sized when we 58033ad798cSAdam Langley * should, and thus we won't abide by the delayed ACK rules correctly. 58133ad798cSAdam Langley * SACKs don't matter, we never delay an ACK when we have any of those 58233ad798cSAdam Langley * going out. */ 58333ad798cSAdam Langley opts->mss = tcp_advertise_mss(sk); 584bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 58533ad798cSAdam Langley 586bb5b7c11SDavid S. Miller if (likely(sysctl_tcp_timestamps && *md5 == NULL)) { 58733ad798cSAdam Langley opts->options |= OPTION_TS; 58833ad798cSAdam Langley opts->tsval = TCP_SKB_CB(skb)->when; 58933ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 590bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 59133ad798cSAdam Langley } 592bb5b7c11SDavid S. Miller if (likely(sysctl_tcp_window_scaling)) { 59333ad798cSAdam Langley opts->ws = tp->rx_opt.rcv_wscale; 59489e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 595bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 59633ad798cSAdam Langley } 597bb5b7c11SDavid S. Miller if (likely(sysctl_tcp_sack)) { 59833ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 599b32d1310SDavid S. Miller if (unlikely(!(OPTION_TS & opts->options))) 600bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 60133ad798cSAdam Langley } 60233ad798cSAdam Langley 603bd0388aeSWilliam Allen Simpson /* Note that timestamps are required by the specification. 604bd0388aeSWilliam Allen Simpson * 605bd0388aeSWilliam Allen Simpson * Odd numbers of bytes are prohibited by the specification, ensuring 606bd0388aeSWilliam Allen Simpson * that the cookie is 16-bit aligned, and the resulting cookie pair is 607bd0388aeSWilliam Allen Simpson * 32-bit aligned. 608bd0388aeSWilliam Allen Simpson */ 609bd0388aeSWilliam Allen Simpson if (*md5 == NULL && 610bd0388aeSWilliam Allen Simpson (OPTION_TS & opts->options) && 611bd0388aeSWilliam Allen Simpson cookie_size > 0) { 612bd0388aeSWilliam Allen Simpson int need = TCPOLEN_COOKIE_BASE + cookie_size; 613bd0388aeSWilliam Allen Simpson 614bd0388aeSWilliam Allen Simpson if (0x2 & need) { 615bd0388aeSWilliam Allen Simpson /* 32-bit multiple */ 616bd0388aeSWilliam Allen Simpson need += 2; /* NOPs */ 617bd0388aeSWilliam Allen Simpson 618bd0388aeSWilliam Allen Simpson if (need > remaining) { 619bd0388aeSWilliam Allen Simpson /* try shrinking cookie to fit */ 620bd0388aeSWilliam Allen Simpson cookie_size -= 2; 621bd0388aeSWilliam Allen Simpson need -= 4; 622bd0388aeSWilliam Allen Simpson } 623bd0388aeSWilliam Allen Simpson } 624bd0388aeSWilliam Allen Simpson while (need > remaining && TCP_COOKIE_MIN <= cookie_size) { 625bd0388aeSWilliam Allen Simpson cookie_size -= 4; 626bd0388aeSWilliam Allen Simpson need -= 4; 627bd0388aeSWilliam Allen Simpson } 628bd0388aeSWilliam Allen Simpson if (TCP_COOKIE_MIN <= cookie_size) { 629bd0388aeSWilliam Allen Simpson opts->options |= OPTION_COOKIE_EXTENSION; 630bd0388aeSWilliam Allen Simpson opts->hash_location = (__u8 *)&cvp->cookie_pair[0]; 631bd0388aeSWilliam Allen Simpson opts->hash_size = cookie_size; 632bd0388aeSWilliam Allen Simpson 633bd0388aeSWilliam Allen Simpson /* Remember for future incarnations. */ 634bd0388aeSWilliam Allen Simpson cvp->cookie_desired = cookie_size; 635bd0388aeSWilliam Allen Simpson 636bd0388aeSWilliam Allen Simpson if (cvp->cookie_desired != cvp->cookie_pair_size) { 637bd0388aeSWilliam Allen Simpson /* Currently use random bytes as a nonce, 638bd0388aeSWilliam Allen Simpson * assuming these are completely unpredictable 639bd0388aeSWilliam Allen Simpson * by hostile users of the same system. 640bd0388aeSWilliam Allen Simpson */ 641bd0388aeSWilliam Allen Simpson get_random_bytes(&cvp->cookie_pair[0], 642bd0388aeSWilliam Allen Simpson cookie_size); 643bd0388aeSWilliam Allen Simpson cvp->cookie_pair_size = cookie_size; 644bd0388aeSWilliam Allen Simpson } 645bd0388aeSWilliam Allen Simpson 646bd0388aeSWilliam Allen Simpson remaining -= need; 647bd0388aeSWilliam Allen Simpson } 648bd0388aeSWilliam Allen Simpson } 649bd0388aeSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 65033ad798cSAdam Langley } 65133ad798cSAdam Langley 65267edfef7SAndi Kleen /* Set up TCP options for SYN-ACKs. */ 65333ad798cSAdam Langley static unsigned tcp_synack_options(struct sock *sk, 65433ad798cSAdam Langley struct request_sock *req, 65533ad798cSAdam Langley unsigned mss, struct sk_buff *skb, 65633ad798cSAdam Langley struct tcp_out_options *opts, 6574957faadSWilliam Allen Simpson struct tcp_md5sig_key **md5, 6584957faadSWilliam Allen Simpson struct tcp_extend_values *xvp) 6594957faadSWilliam Allen Simpson { 66033ad798cSAdam Langley struct inet_request_sock *ireq = inet_rsk(req); 6614957faadSWilliam Allen Simpson unsigned remaining = MAX_TCP_OPTION_SPACE; 6624957faadSWilliam Allen Simpson u8 cookie_plus = (xvp != NULL && !xvp->cookie_out_never) ? 6634957faadSWilliam Allen Simpson xvp->cookie_plus : 6644957faadSWilliam Allen Simpson 0; 66533ad798cSAdam Langley 66633ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 66733ad798cSAdam Langley *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req); 66833ad798cSAdam Langley if (*md5) { 66933ad798cSAdam Langley opts->options |= OPTION_MD5; 6704957faadSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 6714957faadSWilliam Allen Simpson 6724957faadSWilliam Allen Simpson /* We can't fit any SACK blocks in a packet with MD5 + TS 6734957faadSWilliam Allen Simpson * options. There was discussion about disabling SACK 6744957faadSWilliam Allen Simpson * rather than TS in order to fit in better with old, 6754957faadSWilliam Allen Simpson * buggy kernels, but that was deemed to be unnecessary. 6764957faadSWilliam Allen Simpson */ 677de213e5eSEric Dumazet ireq->tstamp_ok &= !ireq->sack_ok; 67833ad798cSAdam Langley } 67933ad798cSAdam Langley #else 68033ad798cSAdam Langley *md5 = NULL; 68133ad798cSAdam Langley #endif 68233ad798cSAdam Langley 6834957faadSWilliam Allen Simpson /* We always send an MSS option. */ 68433ad798cSAdam Langley opts->mss = mss; 6854957faadSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 68633ad798cSAdam Langley 68733ad798cSAdam Langley if (likely(ireq->wscale_ok)) { 68833ad798cSAdam Langley opts->ws = ireq->rcv_wscale; 68989e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 6904957faadSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 69133ad798cSAdam Langley } 692de213e5eSEric Dumazet if (likely(ireq->tstamp_ok)) { 69333ad798cSAdam Langley opts->options |= OPTION_TS; 69433ad798cSAdam Langley opts->tsval = TCP_SKB_CB(skb)->when; 69533ad798cSAdam Langley opts->tsecr = req->ts_recent; 6964957faadSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 69733ad798cSAdam Langley } 69833ad798cSAdam Langley if (likely(ireq->sack_ok)) { 69933ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 700de213e5eSEric Dumazet if (unlikely(!ireq->tstamp_ok)) 7014957faadSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 70233ad798cSAdam Langley } 70333ad798cSAdam Langley 7044957faadSWilliam Allen Simpson /* Similar rationale to tcp_syn_options() applies here, too. 7054957faadSWilliam Allen Simpson * If the <SYN> options fit, the same options should fit now! 7064957faadSWilliam Allen Simpson */ 7074957faadSWilliam Allen Simpson if (*md5 == NULL && 708de213e5eSEric Dumazet ireq->tstamp_ok && 7094957faadSWilliam Allen Simpson cookie_plus > TCPOLEN_COOKIE_BASE) { 7104957faadSWilliam Allen Simpson int need = cookie_plus; /* has TCPOLEN_COOKIE_BASE */ 7114957faadSWilliam Allen Simpson 7124957faadSWilliam Allen Simpson if (0x2 & need) { 7134957faadSWilliam Allen Simpson /* 32-bit multiple */ 7144957faadSWilliam Allen Simpson need += 2; /* NOPs */ 7154957faadSWilliam Allen Simpson } 7164957faadSWilliam Allen Simpson if (need <= remaining) { 7174957faadSWilliam Allen Simpson opts->options |= OPTION_COOKIE_EXTENSION; 7184957faadSWilliam Allen Simpson opts->hash_size = cookie_plus - TCPOLEN_COOKIE_BASE; 7194957faadSWilliam Allen Simpson remaining -= need; 7204957faadSWilliam Allen Simpson } else { 7214957faadSWilliam Allen Simpson /* There's no error return, so flag it. */ 7224957faadSWilliam Allen Simpson xvp->cookie_out_never = 1; /* true */ 7234957faadSWilliam Allen Simpson opts->hash_size = 0; 7244957faadSWilliam Allen Simpson } 7254957faadSWilliam Allen Simpson } 7264957faadSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 72733ad798cSAdam Langley } 72833ad798cSAdam Langley 72967edfef7SAndi Kleen /* Compute TCP options for ESTABLISHED sockets. This is not the 73067edfef7SAndi Kleen * final wire format yet. 73167edfef7SAndi Kleen */ 73233ad798cSAdam Langley static unsigned tcp_established_options(struct sock *sk, struct sk_buff *skb, 73333ad798cSAdam Langley struct tcp_out_options *opts, 73433ad798cSAdam Langley struct tcp_md5sig_key **md5) { 73533ad798cSAdam Langley struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL; 73633ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 73733ad798cSAdam Langley unsigned size = 0; 738cabeccbdSIlpo Järvinen unsigned int eff_sacks; 73933ad798cSAdam Langley 74033ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 74133ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 74233ad798cSAdam Langley if (unlikely(*md5)) { 74333ad798cSAdam Langley opts->options |= OPTION_MD5; 74433ad798cSAdam Langley size += TCPOLEN_MD5SIG_ALIGNED; 74533ad798cSAdam Langley } 74633ad798cSAdam Langley #else 74733ad798cSAdam Langley *md5 = NULL; 74833ad798cSAdam Langley #endif 74933ad798cSAdam Langley 75033ad798cSAdam Langley if (likely(tp->rx_opt.tstamp_ok)) { 75133ad798cSAdam Langley opts->options |= OPTION_TS; 75233ad798cSAdam Langley opts->tsval = tcb ? tcb->when : 0; 75333ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 75433ad798cSAdam Langley size += TCPOLEN_TSTAMP_ALIGNED; 75533ad798cSAdam Langley } 75633ad798cSAdam Langley 757cabeccbdSIlpo Järvinen eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; 758cabeccbdSIlpo Järvinen if (unlikely(eff_sacks)) { 75933ad798cSAdam Langley const unsigned remaining = MAX_TCP_OPTION_SPACE - size; 76033ad798cSAdam Langley opts->num_sack_blocks = 761cabeccbdSIlpo Järvinen min_t(unsigned, eff_sacks, 76233ad798cSAdam Langley (remaining - TCPOLEN_SACK_BASE_ALIGNED) / 76333ad798cSAdam Langley TCPOLEN_SACK_PERBLOCK); 76433ad798cSAdam Langley size += TCPOLEN_SACK_BASE_ALIGNED + 76533ad798cSAdam Langley opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; 76633ad798cSAdam Langley } 76733ad798cSAdam Langley 76833ad798cSAdam Langley return size; 76940efc6faSStephen Hemminger } 7701da177e4SLinus Torvalds 7711da177e4SLinus Torvalds /* This routine actually transmits TCP packets queued in by 7721da177e4SLinus Torvalds * tcp_do_sendmsg(). This is used by both the initial 7731da177e4SLinus Torvalds * transmission and possible later retransmissions. 7741da177e4SLinus Torvalds * All SKB's seen here are completely headerless. It is our 7751da177e4SLinus Torvalds * job to build the TCP header, and pass the packet down to 7761da177e4SLinus Torvalds * IP so it can do the same plus pass the packet off to the 7771da177e4SLinus Torvalds * device. 7781da177e4SLinus Torvalds * 7791da177e4SLinus Torvalds * We are working here with either a clone of the original 7801da177e4SLinus Torvalds * SKB, or a fresh unique copy made by the retransmit engine. 7811da177e4SLinus Torvalds */ 782056834d9SIlpo Järvinen static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, 783056834d9SIlpo Järvinen gfp_t gfp_mask) 7841da177e4SLinus Torvalds { 7856687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 786dfb4b9dcSDavid S. Miller struct inet_sock *inet; 787dfb4b9dcSDavid S. Miller struct tcp_sock *tp; 788dfb4b9dcSDavid S. Miller struct tcp_skb_cb *tcb; 78933ad798cSAdam Langley struct tcp_out_options opts; 79033ad798cSAdam Langley unsigned tcp_options_size, tcp_header_size; 791cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 7921da177e4SLinus Torvalds struct tcphdr *th; 7931da177e4SLinus Torvalds int err; 7941da177e4SLinus Torvalds 795dfb4b9dcSDavid S. Miller BUG_ON(!skb || !tcp_skb_pcount(skb)); 796dfb4b9dcSDavid S. Miller 797dfb4b9dcSDavid S. Miller /* If congestion control is doing timestamping, we must 798dfb4b9dcSDavid S. Miller * take such a timestamp before we potentially clone/copy. 799dfb4b9dcSDavid S. Miller */ 800164891aaSStephen Hemminger if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP) 801dfb4b9dcSDavid S. Miller __net_timestamp(skb); 802dfb4b9dcSDavid S. Miller 803dfb4b9dcSDavid S. Miller if (likely(clone_it)) { 804dfb4b9dcSDavid S. Miller if (unlikely(skb_cloned(skb))) 805dfb4b9dcSDavid S. Miller skb = pskb_copy(skb, gfp_mask); 806dfb4b9dcSDavid S. Miller else 807dfb4b9dcSDavid S. Miller skb = skb_clone(skb, gfp_mask); 808dfb4b9dcSDavid S. Miller if (unlikely(!skb)) 809dfb4b9dcSDavid S. Miller return -ENOBUFS; 810dfb4b9dcSDavid S. Miller } 811dfb4b9dcSDavid S. Miller 812dfb4b9dcSDavid S. Miller inet = inet_sk(sk); 813dfb4b9dcSDavid S. Miller tp = tcp_sk(sk); 814dfb4b9dcSDavid S. Miller tcb = TCP_SKB_CB(skb); 81533ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 8161da177e4SLinus Torvalds 817a3433f35SChangli Gao if (unlikely(tcb->flags & TCPHDR_SYN)) 81833ad798cSAdam Langley tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); 81933ad798cSAdam Langley else 82033ad798cSAdam Langley tcp_options_size = tcp_established_options(sk, skb, &opts, 82133ad798cSAdam Langley &md5); 82233ad798cSAdam Langley tcp_header_size = tcp_options_size + sizeof(struct tcphdr); 8231da177e4SLinus Torvalds 824317a76f9SStephen Hemminger if (tcp_packets_in_flight(tp) == 0) 8256687e988SArnaldo Carvalho de Melo tcp_ca_event(sk, CA_EVENT_TX_START); 8261da177e4SLinus Torvalds 827aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 828aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 829e89862f4SDavid S. Miller skb_set_owner_w(skb, sk); 8301da177e4SLinus Torvalds 8311da177e4SLinus Torvalds /* Build TCP header and checksum it. */ 832aa8223c7SArnaldo Carvalho de Melo th = tcp_hdr(skb); 833c720c7e8SEric Dumazet th->source = inet->inet_sport; 834c720c7e8SEric Dumazet th->dest = inet->inet_dport; 8351da177e4SLinus Torvalds th->seq = htonl(tcb->seq); 8361da177e4SLinus Torvalds th->ack_seq = htonl(tp->rcv_nxt); 837df7a3b07SAl Viro *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | 838dfb4b9dcSDavid S. Miller tcb->flags); 839dfb4b9dcSDavid S. Miller 840a3433f35SChangli Gao if (unlikely(tcb->flags & TCPHDR_SYN)) { 8411da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments 8421da177e4SLinus Torvalds * is never scaled. 8431da177e4SLinus Torvalds */ 844600ff0c2SIlpo Järvinen th->window = htons(min(tp->rcv_wnd, 65535U)); 8451da177e4SLinus Torvalds } else { 8461da177e4SLinus Torvalds th->window = htons(tcp_select_window(sk)); 8471da177e4SLinus Torvalds } 8481da177e4SLinus Torvalds th->check = 0; 8491da177e4SLinus Torvalds th->urg_ptr = 0; 8501da177e4SLinus Torvalds 85133f5f57eSIlpo Järvinen /* The urg_mode check is necessary during a below snd_una win probe */ 8527691367dSHerbert Xu if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { 8537691367dSHerbert Xu if (before(tp->snd_up, tcb->seq + 0x10000)) { 8541da177e4SLinus Torvalds th->urg_ptr = htons(tp->snd_up - tcb->seq); 8551da177e4SLinus Torvalds th->urg = 1; 8567691367dSHerbert Xu } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { 8570eae88f3SEric Dumazet th->urg_ptr = htons(0xFFFF); 8587691367dSHerbert Xu th->urg = 1; 8597691367dSHerbert Xu } 8601da177e4SLinus Torvalds } 8611da177e4SLinus Torvalds 862bd0388aeSWilliam Allen Simpson tcp_options_write((__be32 *)(th + 1), tp, &opts); 863a3433f35SChangli Gao if (likely((tcb->flags & TCPHDR_SYN) == 0)) 8649e412ba7SIlpo Järvinen TCP_ECN_send(sk, skb, tcp_header_size); 865dfb4b9dcSDavid S. Miller 866cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 867cfb6eeb4SYOSHIFUJI Hideaki /* Calculate the MD5 hash, as we have all we need now */ 868cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 869a465419bSEric Dumazet sk_nocaps_add(sk, NETIF_F_GSO_MASK); 870bd0388aeSWilliam Allen Simpson tp->af_specific->calc_md5_hash(opts.hash_location, 87149a72dfbSAdam Langley md5, sk, NULL, skb); 872cfb6eeb4SYOSHIFUJI Hideaki } 873cfb6eeb4SYOSHIFUJI Hideaki #endif 874cfb6eeb4SYOSHIFUJI Hideaki 875bb296246SHerbert Xu icsk->icsk_af_ops->send_check(sk, skb); 8761da177e4SLinus Torvalds 877a3433f35SChangli Gao if (likely(tcb->flags & TCPHDR_ACK)) 878fc6415bcSDavid S. Miller tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); 8791da177e4SLinus Torvalds 8801da177e4SLinus Torvalds if (skb->len != tcp_header_size) 8811da177e4SLinus Torvalds tcp_event_data_sent(tp, skb, sk); 8821da177e4SLinus Torvalds 883bd37a088SWei Yongjun if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) 884aa2ea058STom Herbert TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, 885aa2ea058STom Herbert tcp_skb_pcount(skb)); 8861da177e4SLinus Torvalds 8874e15ed4dSShan Wei err = icsk->icsk_af_ops->queue_xmit(skb); 88883de47cdSHua Zhong if (likely(err <= 0)) 8891da177e4SLinus Torvalds return err; 8901da177e4SLinus Torvalds 8913cfe3baaSIlpo Järvinen tcp_enter_cwr(sk, 1); 8921da177e4SLinus Torvalds 893b9df3cb8SGerrit Renker return net_xmit_eval(err); 8941da177e4SLinus Torvalds } 8951da177e4SLinus Torvalds 89667edfef7SAndi Kleen /* This routine just queues the buffer for sending. 8971da177e4SLinus Torvalds * 8981da177e4SLinus Torvalds * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, 8991da177e4SLinus Torvalds * otherwise socket can stall. 9001da177e4SLinus Torvalds */ 9011da177e4SLinus Torvalds static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) 9021da177e4SLinus Torvalds { 9031da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 9041da177e4SLinus Torvalds 9051da177e4SLinus Torvalds /* Advance write_seq and place onto the write_queue. */ 9061da177e4SLinus Torvalds tp->write_seq = TCP_SKB_CB(skb)->end_seq; 9071da177e4SLinus Torvalds skb_header_release(skb); 908fe067e8aSDavid S. Miller tcp_add_write_queue_tail(sk, skb); 9093ab224beSHideo Aoki sk->sk_wmem_queued += skb->truesize; 9103ab224beSHideo Aoki sk_mem_charge(sk, skb->truesize); 9111da177e4SLinus Torvalds } 9121da177e4SLinus Torvalds 91367edfef7SAndi Kleen /* Initialize TSO segments for a packet. */ 914056834d9SIlpo Järvinen static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb, 915056834d9SIlpo Järvinen unsigned int mss_now) 916f6302d1dSDavid S. Miller { 9178e5b9ddaSHerbert Xu if (skb->len <= mss_now || !sk_can_gso(sk) || 9188e5b9ddaSHerbert Xu skb->ip_summed == CHECKSUM_NONE) { 919f6302d1dSDavid S. Miller /* Avoid the costly divide in the normal 920f6302d1dSDavid S. Miller * non-TSO case. 921f6302d1dSDavid S. Miller */ 9227967168cSHerbert Xu skb_shinfo(skb)->gso_segs = 1; 9237967168cSHerbert Xu skb_shinfo(skb)->gso_size = 0; 9247967168cSHerbert Xu skb_shinfo(skb)->gso_type = 0; 925f6302d1dSDavid S. Miller } else { 926356f89e1SIlpo Järvinen skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now); 9277967168cSHerbert Xu skb_shinfo(skb)->gso_size = mss_now; 928bcd76111SHerbert Xu skb_shinfo(skb)->gso_type = sk->sk_gso_type; 9291da177e4SLinus Torvalds } 9301da177e4SLinus Torvalds } 9311da177e4SLinus Torvalds 93291fed7a1SIlpo Järvinen /* When a modification to fackets out becomes necessary, we need to check 93368f8353bSIlpo Järvinen * skb is counted to fackets_out or not. 93491fed7a1SIlpo Järvinen */ 935a47e5a98SIlpo Järvinen static void tcp_adjust_fackets_out(struct sock *sk, struct sk_buff *skb, 93691fed7a1SIlpo Järvinen int decr) 93791fed7a1SIlpo Järvinen { 938a47e5a98SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 939a47e5a98SIlpo Järvinen 940dc86967bSIlpo Järvinen if (!tp->sacked_out || tcp_is_reno(tp)) 94191fed7a1SIlpo Järvinen return; 94291fed7a1SIlpo Järvinen 9436859d494SIlpo Järvinen if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq)) 94491fed7a1SIlpo Järvinen tp->fackets_out -= decr; 94591fed7a1SIlpo Järvinen } 94691fed7a1SIlpo Järvinen 947797108d1SIlpo Järvinen /* Pcount in the middle of the write queue got changed, we need to do various 948797108d1SIlpo Järvinen * tweaks to fix counters 949797108d1SIlpo Järvinen */ 950797108d1SIlpo Järvinen static void tcp_adjust_pcount(struct sock *sk, struct sk_buff *skb, int decr) 951797108d1SIlpo Järvinen { 952797108d1SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 953797108d1SIlpo Järvinen 954797108d1SIlpo Järvinen tp->packets_out -= decr; 955797108d1SIlpo Järvinen 956797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 957797108d1SIlpo Järvinen tp->sacked_out -= decr; 958797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 959797108d1SIlpo Järvinen tp->retrans_out -= decr; 960797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 961797108d1SIlpo Järvinen tp->lost_out -= decr; 962797108d1SIlpo Järvinen 963797108d1SIlpo Järvinen /* Reno case is special. Sigh... */ 964797108d1SIlpo Järvinen if (tcp_is_reno(tp) && decr > 0) 965797108d1SIlpo Järvinen tp->sacked_out -= min_t(u32, tp->sacked_out, decr); 966797108d1SIlpo Järvinen 967797108d1SIlpo Järvinen tcp_adjust_fackets_out(sk, skb, decr); 968797108d1SIlpo Järvinen 969797108d1SIlpo Järvinen if (tp->lost_skb_hint && 970797108d1SIlpo Järvinen before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && 97152cf3cc8SIlpo Järvinen (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))) 972797108d1SIlpo Järvinen tp->lost_cnt_hint -= decr; 973797108d1SIlpo Järvinen 974797108d1SIlpo Järvinen tcp_verify_left_out(tp); 975797108d1SIlpo Järvinen } 976797108d1SIlpo Järvinen 9771da177e4SLinus Torvalds /* Function to create two new TCP segments. Shrinks the given segment 9781da177e4SLinus Torvalds * to the specified size and appends a new segment with the rest of the 9791da177e4SLinus Torvalds * packet to the list. This won't be called frequently, I hope. 9801da177e4SLinus Torvalds * Remember, these are still headerless SKBs at this point. 9811da177e4SLinus Torvalds */ 982056834d9SIlpo Järvinen int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, 983056834d9SIlpo Järvinen unsigned int mss_now) 9841da177e4SLinus Torvalds { 9851da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 9861da177e4SLinus Torvalds struct sk_buff *buff; 9876475be16SDavid S. Miller int nsize, old_factor; 988b60b49eaSHerbert Xu int nlen; 9899ce01461SIlpo Järvinen u8 flags; 9901da177e4SLinus Torvalds 991b2cc99f0SHerbert Xu BUG_ON(len > skb->len); 9926a438bbeSStephen Hemminger 9931da177e4SLinus Torvalds nsize = skb_headlen(skb) - len; 9941da177e4SLinus Torvalds if (nsize < 0) 9951da177e4SLinus Torvalds nsize = 0; 9961da177e4SLinus Torvalds 9971da177e4SLinus Torvalds if (skb_cloned(skb) && 9981da177e4SLinus Torvalds skb_is_nonlinear(skb) && 9991da177e4SLinus Torvalds pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) 10001da177e4SLinus Torvalds return -ENOMEM; 10011da177e4SLinus Torvalds 10021da177e4SLinus Torvalds /* Get a new skb... force flag on. */ 10031da177e4SLinus Torvalds buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC); 10041da177e4SLinus Torvalds if (buff == NULL) 10051da177e4SLinus Torvalds return -ENOMEM; /* We'll just try again later. */ 1006ef5cb973SHerbert Xu 10073ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 10083ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1009b60b49eaSHerbert Xu nlen = skb->len - len - nsize; 1010b60b49eaSHerbert Xu buff->truesize += nlen; 1011b60b49eaSHerbert Xu skb->truesize -= nlen; 10121da177e4SLinus Torvalds 10131da177e4SLinus Torvalds /* Correct the sequence numbers. */ 10141da177e4SLinus Torvalds TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 10151da177e4SLinus Torvalds TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 10161da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 10171da177e4SLinus Torvalds 10181da177e4SLinus Torvalds /* PSH and FIN should only be set in the second packet. */ 10191da177e4SLinus Torvalds flags = TCP_SKB_CB(skb)->flags; 1020a3433f35SChangli Gao TCP_SKB_CB(skb)->flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 10211da177e4SLinus Torvalds TCP_SKB_CB(buff)->flags = flags; 1022e14c3cafSHerbert Xu TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; 10231da177e4SLinus Torvalds 102484fa7933SPatrick McHardy if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { 10251da177e4SLinus Torvalds /* Copy and checksum data tail into the new buffer. */ 1026056834d9SIlpo Järvinen buff->csum = csum_partial_copy_nocheck(skb->data + len, 1027056834d9SIlpo Järvinen skb_put(buff, nsize), 10281da177e4SLinus Torvalds nsize, 0); 10291da177e4SLinus Torvalds 10301da177e4SLinus Torvalds skb_trim(skb, len); 10311da177e4SLinus Torvalds 10321da177e4SLinus Torvalds skb->csum = csum_block_sub(skb->csum, buff->csum, len); 10331da177e4SLinus Torvalds } else { 103484fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 10351da177e4SLinus Torvalds skb_split(skb, buff, len); 10361da177e4SLinus Torvalds } 10371da177e4SLinus Torvalds 10381da177e4SLinus Torvalds buff->ip_summed = skb->ip_summed; 10391da177e4SLinus Torvalds 10401da177e4SLinus Torvalds /* Looks stupid, but our code really uses when of 10411da177e4SLinus Torvalds * skbs, which it never sent before. --ANK 10421da177e4SLinus Torvalds */ 10431da177e4SLinus Torvalds TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when; 1044a61bbcf2SPatrick McHardy buff->tstamp = skb->tstamp; 10451da177e4SLinus Torvalds 10466475be16SDavid S. Miller old_factor = tcp_skb_pcount(skb); 10476475be16SDavid S. Miller 10481da177e4SLinus Torvalds /* Fix up tso_factor for both original and new SKB. */ 1049846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1050846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, buff, mss_now); 10511da177e4SLinus Torvalds 10526475be16SDavid S. Miller /* If this packet has been sent out already, we must 10536475be16SDavid S. Miller * adjust the various packet counters. 10546475be16SDavid S. Miller */ 1055cf0b450cSHerbert Xu if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 10566475be16SDavid S. Miller int diff = old_factor - tcp_skb_pcount(skb) - 10576475be16SDavid S. Miller tcp_skb_pcount(buff); 10581da177e4SLinus Torvalds 1059797108d1SIlpo Järvinen if (diff) 1060797108d1SIlpo Järvinen tcp_adjust_pcount(sk, skb, diff); 10611da177e4SLinus Torvalds } 10621da177e4SLinus Torvalds 10631da177e4SLinus Torvalds /* Link BUFF into the send queue. */ 1064f44b5271SDavid S. Miller skb_header_release(buff); 1065fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 10661da177e4SLinus Torvalds 10671da177e4SLinus Torvalds return 0; 10681da177e4SLinus Torvalds } 10691da177e4SLinus Torvalds 10701da177e4SLinus Torvalds /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c 10711da177e4SLinus Torvalds * eventually). The difference is that pulled data not copied, but 10721da177e4SLinus Torvalds * immediately discarded. 10731da177e4SLinus Torvalds */ 1074f2911969SHerbert Xu ~{PmVHI~} static void __pskb_trim_head(struct sk_buff *skb, int len) 10751da177e4SLinus Torvalds { 10761da177e4SLinus Torvalds int i, k, eat; 10771da177e4SLinus Torvalds 10781da177e4SLinus Torvalds eat = len; 10791da177e4SLinus Torvalds k = 0; 10801da177e4SLinus Torvalds for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) { 10811da177e4SLinus Torvalds if (skb_shinfo(skb)->frags[i].size <= eat) { 10821da177e4SLinus Torvalds put_page(skb_shinfo(skb)->frags[i].page); 10831da177e4SLinus Torvalds eat -= skb_shinfo(skb)->frags[i].size; 10841da177e4SLinus Torvalds } else { 10851da177e4SLinus Torvalds skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i]; 10861da177e4SLinus Torvalds if (eat) { 10871da177e4SLinus Torvalds skb_shinfo(skb)->frags[k].page_offset += eat; 10881da177e4SLinus Torvalds skb_shinfo(skb)->frags[k].size -= eat; 10891da177e4SLinus Torvalds eat = 0; 10901da177e4SLinus Torvalds } 10911da177e4SLinus Torvalds k++; 10921da177e4SLinus Torvalds } 10931da177e4SLinus Torvalds } 10941da177e4SLinus Torvalds skb_shinfo(skb)->nr_frags = k; 10951da177e4SLinus Torvalds 109627a884dcSArnaldo Carvalho de Melo skb_reset_tail_pointer(skb); 10971da177e4SLinus Torvalds skb->data_len -= len; 10981da177e4SLinus Torvalds skb->len = skb->data_len; 10991da177e4SLinus Torvalds } 11001da177e4SLinus Torvalds 110167edfef7SAndi Kleen /* Remove acked data from a packet in the transmit queue. */ 11021da177e4SLinus Torvalds int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 11031da177e4SLinus Torvalds { 1104056834d9SIlpo Järvinen if (skb_cloned(skb) && pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) 11051da177e4SLinus Torvalds return -ENOMEM; 11061da177e4SLinus Torvalds 1107f2911969SHerbert Xu ~{PmVHI~} /* If len == headlen, we avoid __skb_pull to preserve alignment. */ 1108f2911969SHerbert Xu ~{PmVHI~} if (unlikely(len < skb_headlen(skb))) 11091da177e4SLinus Torvalds __skb_pull(skb, len); 1110f2911969SHerbert Xu ~{PmVHI~} else 1111f2911969SHerbert Xu ~{PmVHI~} __pskb_trim_head(skb, len - skb_headlen(skb)); 11121da177e4SLinus Torvalds 11131da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq += len; 111484fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 11151da177e4SLinus Torvalds 11161da177e4SLinus Torvalds skb->truesize -= len; 11171da177e4SLinus Torvalds sk->sk_wmem_queued -= len; 11183ab224beSHideo Aoki sk_mem_uncharge(sk, len); 11191da177e4SLinus Torvalds sock_set_flag(sk, SOCK_QUEUE_SHRUNK); 11201da177e4SLinus Torvalds 11211da177e4SLinus Torvalds /* Any change of skb->len requires recalculation of tso 11221da177e4SLinus Torvalds * factor and mss. 11231da177e4SLinus Torvalds */ 11241da177e4SLinus Torvalds if (tcp_skb_pcount(skb) > 1) 11250c54b85fSIlpo Järvinen tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk)); 11261da177e4SLinus Torvalds 11271da177e4SLinus Torvalds return 0; 11281da177e4SLinus Torvalds } 11291da177e4SLinus Torvalds 113067edfef7SAndi Kleen /* Calculate MSS. Not accounting for SACKs here. */ 11315d424d5aSJohn Heffner int tcp_mtu_to_mss(struct sock *sk, int pmtu) 11325d424d5aSJohn Heffner { 11335d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 11345d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 11355d424d5aSJohn Heffner int mss_now; 11365d424d5aSJohn Heffner 11375d424d5aSJohn Heffner /* Calculate base mss without TCP options: 11385d424d5aSJohn Heffner It is MMS_S - sizeof(tcphdr) of rfc1122 11395d424d5aSJohn Heffner */ 11405d424d5aSJohn Heffner mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 11415d424d5aSJohn Heffner 11425d424d5aSJohn Heffner /* Clamp it (mss_clamp does not include tcp options) */ 11435d424d5aSJohn Heffner if (mss_now > tp->rx_opt.mss_clamp) 11445d424d5aSJohn Heffner mss_now = tp->rx_opt.mss_clamp; 11455d424d5aSJohn Heffner 11465d424d5aSJohn Heffner /* Now subtract optional transport overhead */ 11475d424d5aSJohn Heffner mss_now -= icsk->icsk_ext_hdr_len; 11485d424d5aSJohn Heffner 11495d424d5aSJohn Heffner /* Then reserve room for full set of TCP options and 8 bytes of data */ 11505d424d5aSJohn Heffner if (mss_now < 48) 11515d424d5aSJohn Heffner mss_now = 48; 11525d424d5aSJohn Heffner 11535d424d5aSJohn Heffner /* Now subtract TCP options size, not including SACKs */ 11545d424d5aSJohn Heffner mss_now -= tp->tcp_header_len - sizeof(struct tcphdr); 11555d424d5aSJohn Heffner 11565d424d5aSJohn Heffner return mss_now; 11575d424d5aSJohn Heffner } 11585d424d5aSJohn Heffner 11595d424d5aSJohn Heffner /* Inverse of above */ 11605d424d5aSJohn Heffner int tcp_mss_to_mtu(struct sock *sk, int mss) 11615d424d5aSJohn Heffner { 11625d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 11635d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 11645d424d5aSJohn Heffner int mtu; 11655d424d5aSJohn Heffner 11665d424d5aSJohn Heffner mtu = mss + 11675d424d5aSJohn Heffner tp->tcp_header_len + 11685d424d5aSJohn Heffner icsk->icsk_ext_hdr_len + 11695d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 11705d424d5aSJohn Heffner 11715d424d5aSJohn Heffner return mtu; 11725d424d5aSJohn Heffner } 11735d424d5aSJohn Heffner 117467edfef7SAndi Kleen /* MTU probing init per socket */ 11755d424d5aSJohn Heffner void tcp_mtup_init(struct sock *sk) 11765d424d5aSJohn Heffner { 11775d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 11785d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 11795d424d5aSJohn Heffner 11805d424d5aSJohn Heffner icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1; 11815d424d5aSJohn Heffner icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 11825d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 11835d424d5aSJohn Heffner icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss); 11845d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 11855d424d5aSJohn Heffner } 11864bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_mtup_init); 11875d424d5aSJohn Heffner 11881da177e4SLinus Torvalds /* This function synchronize snd mss to current pmtu/exthdr set. 11891da177e4SLinus Torvalds 11901da177e4SLinus Torvalds tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 11911da177e4SLinus Torvalds for TCP options, but includes only bare TCP header. 11921da177e4SLinus Torvalds 11931da177e4SLinus Torvalds tp->rx_opt.mss_clamp is mss negotiated at connection setup. 1194caa20d9aSStephen Hemminger It is minimum of user_mss and mss received with SYN. 11951da177e4SLinus Torvalds It also does not include TCP options. 11961da177e4SLinus Torvalds 1197d83d8461SArnaldo Carvalho de Melo inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. 11981da177e4SLinus Torvalds 11991da177e4SLinus Torvalds tp->mss_cache is current effective sending mss, including 12001da177e4SLinus Torvalds all tcp options except for SACKs. It is evaluated, 12011da177e4SLinus Torvalds taking into account current pmtu, but never exceeds 12021da177e4SLinus Torvalds tp->rx_opt.mss_clamp. 12031da177e4SLinus Torvalds 12041da177e4SLinus Torvalds NOTE1. rfc1122 clearly states that advertised MSS 12051da177e4SLinus Torvalds DOES NOT include either tcp or ip options. 12061da177e4SLinus Torvalds 1207d83d8461SArnaldo Carvalho de Melo NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache 1208d83d8461SArnaldo Carvalho de Melo are READ ONLY outside this function. --ANK (980731) 12091da177e4SLinus Torvalds */ 12101da177e4SLinus Torvalds unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 12111da177e4SLinus Torvalds { 12121da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1213d83d8461SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 12145d424d5aSJohn Heffner int mss_now; 12151da177e4SLinus Torvalds 12165d424d5aSJohn Heffner if (icsk->icsk_mtup.search_high > pmtu) 12175d424d5aSJohn Heffner icsk->icsk_mtup.search_high = pmtu; 12181da177e4SLinus Torvalds 12195d424d5aSJohn Heffner mss_now = tcp_mtu_to_mss(sk, pmtu); 1220409d22b4SIlpo Järvinen mss_now = tcp_bound_to_half_wnd(tp, mss_now); 12211da177e4SLinus Torvalds 12221da177e4SLinus Torvalds /* And store cached results */ 1223d83d8461SArnaldo Carvalho de Melo icsk->icsk_pmtu_cookie = pmtu; 12245d424d5aSJohn Heffner if (icsk->icsk_mtup.enabled) 12255d424d5aSJohn Heffner mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); 1226c1b4a7e6SDavid S. Miller tp->mss_cache = mss_now; 12271da177e4SLinus Torvalds 12281da177e4SLinus Torvalds return mss_now; 12291da177e4SLinus Torvalds } 12304bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_sync_mss); 12311da177e4SLinus Torvalds 12321da177e4SLinus Torvalds /* Compute the current effective MSS, taking SACKs and IP options, 12331da177e4SLinus Torvalds * and even PMTU discovery events into account. 12341da177e4SLinus Torvalds */ 12350c54b85fSIlpo Järvinen unsigned int tcp_current_mss(struct sock *sk) 12361da177e4SLinus Torvalds { 12371da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 12381da177e4SLinus Torvalds struct dst_entry *dst = __sk_dst_get(sk); 1239c1b4a7e6SDavid S. Miller u32 mss_now; 124033ad798cSAdam Langley unsigned header_len; 124133ad798cSAdam Langley struct tcp_out_options opts; 124233ad798cSAdam Langley struct tcp_md5sig_key *md5; 12431da177e4SLinus Torvalds 1244c1b4a7e6SDavid S. Miller mss_now = tp->mss_cache; 1245c1b4a7e6SDavid S. Miller 12461da177e4SLinus Torvalds if (dst) { 12471da177e4SLinus Torvalds u32 mtu = dst_mtu(dst); 1248d83d8461SArnaldo Carvalho de Melo if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 12491da177e4SLinus Torvalds mss_now = tcp_sync_mss(sk, mtu); 12501da177e4SLinus Torvalds } 12511da177e4SLinus Torvalds 125233ad798cSAdam Langley header_len = tcp_established_options(sk, NULL, &opts, &md5) + 125333ad798cSAdam Langley sizeof(struct tcphdr); 125433ad798cSAdam Langley /* The mss_cache is sized based on tp->tcp_header_len, which assumes 125533ad798cSAdam Langley * some common options. If this is an odd packet (because we have SACK 125633ad798cSAdam Langley * blocks etc) then our calculated header_len will be different, and 125733ad798cSAdam Langley * we have to adjust mss_now correspondingly */ 125833ad798cSAdam Langley if (header_len != tp->tcp_header_len) { 125933ad798cSAdam Langley int delta = (int) header_len - tp->tcp_header_len; 126033ad798cSAdam Langley mss_now -= delta; 126133ad798cSAdam Langley } 1262cfb6eeb4SYOSHIFUJI Hideaki 12631da177e4SLinus Torvalds return mss_now; 12641da177e4SLinus Torvalds } 12651da177e4SLinus Torvalds 1266a762a980SDavid S. Miller /* Congestion window validation. (RFC2861) */ 12679e412ba7SIlpo Järvinen static void tcp_cwnd_validate(struct sock *sk) 1268a762a980SDavid S. Miller { 12699e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1270a762a980SDavid S. Miller 1271d436d686SIlpo Järvinen if (tp->packets_out >= tp->snd_cwnd) { 1272a762a980SDavid S. Miller /* Network is feed fully. */ 1273a762a980SDavid S. Miller tp->snd_cwnd_used = 0; 1274a762a980SDavid S. Miller tp->snd_cwnd_stamp = tcp_time_stamp; 1275a762a980SDavid S. Miller } else { 1276a762a980SDavid S. Miller /* Network starves. */ 1277a762a980SDavid S. Miller if (tp->packets_out > tp->snd_cwnd_used) 1278a762a980SDavid S. Miller tp->snd_cwnd_used = tp->packets_out; 1279a762a980SDavid S. Miller 128015d33c07SDavid S. Miller if (sysctl_tcp_slow_start_after_idle && 128115d33c07SDavid S. Miller (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto) 1282a762a980SDavid S. Miller tcp_cwnd_application_limited(sk); 1283a762a980SDavid S. Miller } 1284a762a980SDavid S. Miller } 1285a762a980SDavid S. Miller 12860e3a4803SIlpo Järvinen /* Returns the portion of skb which can be sent right away without 12870e3a4803SIlpo Järvinen * introducing MSS oddities to segment boundaries. In rare cases where 12880e3a4803SIlpo Järvinen * mss_now != mss_cache, we will request caller to create a small skb 12890e3a4803SIlpo Järvinen * per input skb which could be mostly avoided here (if desired). 12905ea3a748SIlpo Järvinen * 12915ea3a748SIlpo Järvinen * We explicitly want to create a request for splitting write queue tail 12925ea3a748SIlpo Järvinen * to a small skb for Nagle purposes while avoiding unnecessary modulos, 12935ea3a748SIlpo Järvinen * thus all the complexity (cwnd_len is always MSS multiple which we 12945ea3a748SIlpo Järvinen * return whenever allowed by the other factors). Basically we need the 12955ea3a748SIlpo Järvinen * modulo only when the receiver window alone is the limiting factor or 12965ea3a748SIlpo Järvinen * when we would be allowed to send the split-due-to-Nagle skb fully. 12970e3a4803SIlpo Järvinen */ 12980e3a4803SIlpo Järvinen static unsigned int tcp_mss_split_point(struct sock *sk, struct sk_buff *skb, 1299056834d9SIlpo Järvinen unsigned int mss_now, unsigned int cwnd) 1300c1b4a7e6SDavid S. Miller { 13010e3a4803SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 13020e3a4803SIlpo Järvinen u32 needed, window, cwnd_len; 1303c1b4a7e6SDavid S. Miller 130490840defSIlpo Järvinen window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1305c1b4a7e6SDavid S. Miller cwnd_len = mss_now * cwnd; 13060e3a4803SIlpo Järvinen 13070e3a4803SIlpo Järvinen if (likely(cwnd_len <= window && skb != tcp_write_queue_tail(sk))) 13080e3a4803SIlpo Järvinen return cwnd_len; 13090e3a4803SIlpo Järvinen 13105ea3a748SIlpo Järvinen needed = min(skb->len, window); 13115ea3a748SIlpo Järvinen 131217515408SIlpo Järvinen if (cwnd_len <= needed) 13130e3a4803SIlpo Järvinen return cwnd_len; 13140e3a4803SIlpo Järvinen 13150e3a4803SIlpo Järvinen return needed - needed % mss_now; 1316c1b4a7e6SDavid S. Miller } 1317c1b4a7e6SDavid S. Miller 1318c1b4a7e6SDavid S. Miller /* Can at least one segment of SKB be sent right now, according to the 1319c1b4a7e6SDavid S. Miller * congestion window rules? If so, return how many segments are allowed. 1320c1b4a7e6SDavid S. Miller */ 1321056834d9SIlpo Järvinen static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp, 1322056834d9SIlpo Järvinen struct sk_buff *skb) 1323c1b4a7e6SDavid S. Miller { 1324c1b4a7e6SDavid S. Miller u32 in_flight, cwnd; 1325c1b4a7e6SDavid S. Miller 1326c1b4a7e6SDavid S. Miller /* Don't be strict about the congestion window for the final FIN. */ 1327a3433f35SChangli Gao if ((TCP_SKB_CB(skb)->flags & TCPHDR_FIN) && tcp_skb_pcount(skb) == 1) 1328c1b4a7e6SDavid S. Miller return 1; 1329c1b4a7e6SDavid S. Miller 1330c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1331c1b4a7e6SDavid S. Miller cwnd = tp->snd_cwnd; 1332c1b4a7e6SDavid S. Miller if (in_flight < cwnd) 1333c1b4a7e6SDavid S. Miller return (cwnd - in_flight); 1334c1b4a7e6SDavid S. Miller 1335c1b4a7e6SDavid S. Miller return 0; 1336c1b4a7e6SDavid S. Miller } 1337c1b4a7e6SDavid S. Miller 133867edfef7SAndi Kleen /* Intialize TSO state of a skb. 133967edfef7SAndi Kleen * This must be invoked the first time we consider transmitting 1340c1b4a7e6SDavid S. Miller * SKB onto the wire. 1341c1b4a7e6SDavid S. Miller */ 1342056834d9SIlpo Järvinen static int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb, 1343056834d9SIlpo Järvinen unsigned int mss_now) 1344c1b4a7e6SDavid S. Miller { 1345c1b4a7e6SDavid S. Miller int tso_segs = tcp_skb_pcount(skb); 1346c1b4a7e6SDavid S. Miller 1347f8269a49SIlpo Järvinen if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { 1348846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1349c1b4a7e6SDavid S. Miller tso_segs = tcp_skb_pcount(skb); 1350c1b4a7e6SDavid S. Miller } 1351c1b4a7e6SDavid S. Miller return tso_segs; 1352c1b4a7e6SDavid S. Miller } 1353c1b4a7e6SDavid S. Miller 135467edfef7SAndi Kleen /* Minshall's variant of the Nagle send check. */ 1355c1b4a7e6SDavid S. Miller static inline int tcp_minshall_check(const struct tcp_sock *tp) 1356c1b4a7e6SDavid S. Miller { 1357c1b4a7e6SDavid S. Miller return after(tp->snd_sml, tp->snd_una) && 1358c1b4a7e6SDavid S. Miller !after(tp->snd_sml, tp->snd_nxt); 1359c1b4a7e6SDavid S. Miller } 1360c1b4a7e6SDavid S. Miller 1361c1b4a7e6SDavid S. Miller /* Return 0, if packet can be sent now without violation Nagle's rules: 1362c1b4a7e6SDavid S. Miller * 1. It is full sized. 1363c1b4a7e6SDavid S. Miller * 2. Or it contains FIN. (already checked by caller) 1364c1b4a7e6SDavid S. Miller * 3. Or TCP_NODELAY was set. 1365c1b4a7e6SDavid S. Miller * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 1366c1b4a7e6SDavid S. Miller * With Minshall's modification: all sent small packets are ACKed. 1367c1b4a7e6SDavid S. Miller */ 1368c1b4a7e6SDavid S. Miller static inline int tcp_nagle_check(const struct tcp_sock *tp, 1369c1b4a7e6SDavid S. Miller const struct sk_buff *skb, 1370c1b4a7e6SDavid S. Miller unsigned mss_now, int nonagle) 1371c1b4a7e6SDavid S. Miller { 1372a02cec21SEric Dumazet return skb->len < mss_now && 1373c1b4a7e6SDavid S. Miller ((nonagle & TCP_NAGLE_CORK) || 1374a02cec21SEric Dumazet (!nonagle && tp->packets_out && tcp_minshall_check(tp))); 1375c1b4a7e6SDavid S. Miller } 1376c1b4a7e6SDavid S. Miller 1377c1b4a7e6SDavid S. Miller /* Return non-zero if the Nagle test allows this packet to be 1378c1b4a7e6SDavid S. Miller * sent now. 1379c1b4a7e6SDavid S. Miller */ 1380c1b4a7e6SDavid S. Miller static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb, 1381c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1382c1b4a7e6SDavid S. Miller { 1383c1b4a7e6SDavid S. Miller /* Nagle rule does not apply to frames, which sit in the middle of the 1384c1b4a7e6SDavid S. Miller * write_queue (they have no chances to get new data). 1385c1b4a7e6SDavid S. Miller * 1386c1b4a7e6SDavid S. Miller * This is implemented in the callers, where they modify the 'nonagle' 1387c1b4a7e6SDavid S. Miller * argument based upon the location of SKB in the send queue. 1388c1b4a7e6SDavid S. Miller */ 1389c1b4a7e6SDavid S. Miller if (nonagle & TCP_NAGLE_PUSH) 1390c1b4a7e6SDavid S. Miller return 1; 1391c1b4a7e6SDavid S. Miller 1392d551e454SIlpo Järvinen /* Don't use the nagle rule for urgent data (or for the final FIN). 1393d551e454SIlpo Järvinen * Nagle can be ignored during F-RTO too (see RFC4138). 1394d551e454SIlpo Järvinen */ 139533f5f57eSIlpo Järvinen if (tcp_urg_mode(tp) || (tp->frto_counter == 2) || 1396a3433f35SChangli Gao (TCP_SKB_CB(skb)->flags & TCPHDR_FIN)) 1397c1b4a7e6SDavid S. Miller return 1; 1398c1b4a7e6SDavid S. Miller 1399c1b4a7e6SDavid S. Miller if (!tcp_nagle_check(tp, skb, cur_mss, nonagle)) 1400c1b4a7e6SDavid S. Miller return 1; 1401c1b4a7e6SDavid S. Miller 1402c1b4a7e6SDavid S. Miller return 0; 1403c1b4a7e6SDavid S. Miller } 1404c1b4a7e6SDavid S. Miller 1405c1b4a7e6SDavid S. Miller /* Does at least the first segment of SKB fit into the send window? */ 1406056834d9SIlpo Järvinen static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb, 1407056834d9SIlpo Järvinen unsigned int cur_mss) 1408c1b4a7e6SDavid S. Miller { 1409c1b4a7e6SDavid S. Miller u32 end_seq = TCP_SKB_CB(skb)->end_seq; 1410c1b4a7e6SDavid S. Miller 1411c1b4a7e6SDavid S. Miller if (skb->len > cur_mss) 1412c1b4a7e6SDavid S. Miller end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 1413c1b4a7e6SDavid S. Miller 141490840defSIlpo Järvinen return !after(end_seq, tcp_wnd_end(tp)); 1415c1b4a7e6SDavid S. Miller } 1416c1b4a7e6SDavid S. Miller 1417fe067e8aSDavid S. Miller /* This checks if the data bearing packet SKB (usually tcp_send_head(sk)) 1418c1b4a7e6SDavid S. Miller * should be put on the wire right now. If so, it returns the number of 1419c1b4a7e6SDavid S. Miller * packets allowed by the congestion window. 1420c1b4a7e6SDavid S. Miller */ 1421c1b4a7e6SDavid S. Miller static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb, 1422c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1423c1b4a7e6SDavid S. Miller { 1424c1b4a7e6SDavid S. Miller struct tcp_sock *tp = tcp_sk(sk); 1425c1b4a7e6SDavid S. Miller unsigned int cwnd_quota; 1426c1b4a7e6SDavid S. Miller 1427846998aeSDavid S. Miller tcp_init_tso_segs(sk, skb, cur_mss); 1428c1b4a7e6SDavid S. Miller 1429c1b4a7e6SDavid S. Miller if (!tcp_nagle_test(tp, skb, cur_mss, nonagle)) 1430c1b4a7e6SDavid S. Miller return 0; 1431c1b4a7e6SDavid S. Miller 1432c1b4a7e6SDavid S. Miller cwnd_quota = tcp_cwnd_test(tp, skb); 1433056834d9SIlpo Järvinen if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss)) 1434c1b4a7e6SDavid S. Miller cwnd_quota = 0; 1435c1b4a7e6SDavid S. Miller 1436c1b4a7e6SDavid S. Miller return cwnd_quota; 1437c1b4a7e6SDavid S. Miller } 1438c1b4a7e6SDavid S. Miller 143967edfef7SAndi Kleen /* Test if sending is allowed right now. */ 14409e412ba7SIlpo Järvinen int tcp_may_send_now(struct sock *sk) 1441c1b4a7e6SDavid S. Miller { 14429e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1443fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 1444c1b4a7e6SDavid S. Miller 1445a02cec21SEric Dumazet return skb && 14460c54b85fSIlpo Järvinen tcp_snd_test(sk, skb, tcp_current_mss(sk), 1447c1b4a7e6SDavid S. Miller (tcp_skb_is_last(sk, skb) ? 1448a02cec21SEric Dumazet tp->nonagle : TCP_NAGLE_PUSH)); 1449c1b4a7e6SDavid S. Miller } 1450c1b4a7e6SDavid S. Miller 1451c1b4a7e6SDavid S. Miller /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 1452c1b4a7e6SDavid S. Miller * which is put after SKB on the list. It is very much like 1453c1b4a7e6SDavid S. Miller * tcp_fragment() except that it may make several kinds of assumptions 1454c1b4a7e6SDavid S. Miller * in order to speed up the splitting operation. In particular, we 1455c1b4a7e6SDavid S. Miller * know that all the data is in scatter-gather pages, and that the 1456c1b4a7e6SDavid S. Miller * packet has never been sent out before (and thus is not cloned). 1457c1b4a7e6SDavid S. Miller */ 1458056834d9SIlpo Järvinen static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, 1459c4ead4c5SEric Dumazet unsigned int mss_now, gfp_t gfp) 1460c1b4a7e6SDavid S. Miller { 1461c1b4a7e6SDavid S. Miller struct sk_buff *buff; 1462c1b4a7e6SDavid S. Miller int nlen = skb->len - len; 14639ce01461SIlpo Järvinen u8 flags; 1464c1b4a7e6SDavid S. Miller 1465c1b4a7e6SDavid S. Miller /* All of a TSO frame must be composed of paged data. */ 1466c8ac3774SHerbert Xu if (skb->len != skb->data_len) 1467c8ac3774SHerbert Xu return tcp_fragment(sk, skb, len, mss_now); 1468c1b4a7e6SDavid S. Miller 1469c4ead4c5SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, gfp); 1470c1b4a7e6SDavid S. Miller if (unlikely(buff == NULL)) 1471c1b4a7e6SDavid S. Miller return -ENOMEM; 1472c1b4a7e6SDavid S. Miller 14733ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 14743ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1475b60b49eaSHerbert Xu buff->truesize += nlen; 1476c1b4a7e6SDavid S. Miller skb->truesize -= nlen; 1477c1b4a7e6SDavid S. Miller 1478c1b4a7e6SDavid S. Miller /* Correct the sequence numbers. */ 1479c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 1480c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 1481c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 1482c1b4a7e6SDavid S. Miller 1483c1b4a7e6SDavid S. Miller /* PSH and FIN should only be set in the second packet. */ 1484c1b4a7e6SDavid S. Miller flags = TCP_SKB_CB(skb)->flags; 1485a3433f35SChangli Gao TCP_SKB_CB(skb)->flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 1486c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->flags = flags; 1487c1b4a7e6SDavid S. Miller 1488c1b4a7e6SDavid S. Miller /* This packet was never sent out yet, so no SACK bits. */ 1489c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->sacked = 0; 1490c1b4a7e6SDavid S. Miller 149184fa7933SPatrick McHardy buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; 1492c1b4a7e6SDavid S. Miller skb_split(skb, buff, len); 1493c1b4a7e6SDavid S. Miller 1494c1b4a7e6SDavid S. Miller /* Fix up tso_factor for both original and new SKB. */ 1495846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1496846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, buff, mss_now); 1497c1b4a7e6SDavid S. Miller 1498c1b4a7e6SDavid S. Miller /* Link BUFF into the send queue. */ 1499c1b4a7e6SDavid S. Miller skb_header_release(buff); 1500fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 1501c1b4a7e6SDavid S. Miller 1502c1b4a7e6SDavid S. Miller return 0; 1503c1b4a7e6SDavid S. Miller } 1504c1b4a7e6SDavid S. Miller 1505c1b4a7e6SDavid S. Miller /* Try to defer sending, if possible, in order to minimize the amount 1506c1b4a7e6SDavid S. Miller * of TSO splitting we do. View it as a kind of TSO Nagle test. 1507c1b4a7e6SDavid S. Miller * 1508c1b4a7e6SDavid S. Miller * This algorithm is from John Heffner. 1509c1b4a7e6SDavid S. Miller */ 15109e412ba7SIlpo Järvinen static int tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb) 1511c1b4a7e6SDavid S. Miller { 15129e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 15136687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 1514c1b4a7e6SDavid S. Miller u32 send_win, cong_win, limit, in_flight; 1515c1b4a7e6SDavid S. Miller 1516a3433f35SChangli Gao if (TCP_SKB_CB(skb)->flags & TCPHDR_FIN) 1517ae8064acSJohn Heffner goto send_now; 1518c1b4a7e6SDavid S. Miller 15196687e988SArnaldo Carvalho de Melo if (icsk->icsk_ca_state != TCP_CA_Open) 1520ae8064acSJohn Heffner goto send_now; 1521ae8064acSJohn Heffner 1522ae8064acSJohn Heffner /* Defer for less than two clock ticks. */ 1523bd515c3eSIlpo Järvinen if (tp->tso_deferred && 1524a2acde07SIlpo Järvinen (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1) 1525ae8064acSJohn Heffner goto send_now; 1526908a75c1SDavid S. Miller 1527c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1528c1b4a7e6SDavid S. Miller 1529056834d9SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight)); 1530c1b4a7e6SDavid S. Miller 153190840defSIlpo Järvinen send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1532c1b4a7e6SDavid S. Miller 1533c1b4a7e6SDavid S. Miller /* From in_flight test above, we know that cwnd > in_flight. */ 1534c1b4a7e6SDavid S. Miller cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; 1535c1b4a7e6SDavid S. Miller 1536c1b4a7e6SDavid S. Miller limit = min(send_win, cong_win); 1537c1b4a7e6SDavid S. Miller 1538ba244fe9SDavid S. Miller /* If a full-sized TSO skb can be sent, do it. */ 153982cc1a7aSPeter P Waskiewicz Jr if (limit >= sk->sk_gso_max_size) 1540ae8064acSJohn Heffner goto send_now; 1541ba244fe9SDavid S. Miller 154262ad2761SIlpo Järvinen /* Middle in queue won't get any more data, full sendable already? */ 154362ad2761SIlpo Järvinen if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) 154462ad2761SIlpo Järvinen goto send_now; 154562ad2761SIlpo Järvinen 1546c1b4a7e6SDavid S. Miller if (sysctl_tcp_tso_win_divisor) { 1547c1b4a7e6SDavid S. Miller u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); 1548c1b4a7e6SDavid S. Miller 1549c1b4a7e6SDavid S. Miller /* If at least some fraction of a window is available, 1550c1b4a7e6SDavid S. Miller * just use it. 1551c1b4a7e6SDavid S. Miller */ 1552c1b4a7e6SDavid S. Miller chunk /= sysctl_tcp_tso_win_divisor; 1553c1b4a7e6SDavid S. Miller if (limit >= chunk) 1554ae8064acSJohn Heffner goto send_now; 1555c1b4a7e6SDavid S. Miller } else { 1556c1b4a7e6SDavid S. Miller /* Different approach, try not to defer past a single 1557c1b4a7e6SDavid S. Miller * ACK. Receiver should ACK every other full sized 1558c1b4a7e6SDavid S. Miller * frame, so if we have space for more than 3 frames 1559c1b4a7e6SDavid S. Miller * then send now. 1560c1b4a7e6SDavid S. Miller */ 1561c1b4a7e6SDavid S. Miller if (limit > tcp_max_burst(tp) * tp->mss_cache) 1562ae8064acSJohn Heffner goto send_now; 1563c1b4a7e6SDavid S. Miller } 1564c1b4a7e6SDavid S. Miller 1565c1b4a7e6SDavid S. Miller /* Ok, it looks like it is advisable to defer. */ 1566ae8064acSJohn Heffner tp->tso_deferred = 1 | (jiffies << 1); 1567ae8064acSJohn Heffner 1568c1b4a7e6SDavid S. Miller return 1; 1569ae8064acSJohn Heffner 1570ae8064acSJohn Heffner send_now: 1571ae8064acSJohn Heffner tp->tso_deferred = 0; 1572ae8064acSJohn Heffner return 0; 1573c1b4a7e6SDavid S. Miller } 1574c1b4a7e6SDavid S. Miller 15755d424d5aSJohn Heffner /* Create a new MTU probe if we are ready. 157667edfef7SAndi Kleen * MTU probe is regularly attempting to increase the path MTU by 157767edfef7SAndi Kleen * deliberately sending larger packets. This discovers routing 157867edfef7SAndi Kleen * changes resulting in larger path MTUs. 157967edfef7SAndi Kleen * 15805d424d5aSJohn Heffner * Returns 0 if we should wait to probe (no cwnd available), 15815d424d5aSJohn Heffner * 1 if a probe was sent, 1582056834d9SIlpo Järvinen * -1 otherwise 1583056834d9SIlpo Järvinen */ 15845d424d5aSJohn Heffner static int tcp_mtu_probe(struct sock *sk) 15855d424d5aSJohn Heffner { 15865d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 15875d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 15885d424d5aSJohn Heffner struct sk_buff *skb, *nskb, *next; 15895d424d5aSJohn Heffner int len; 15905d424d5aSJohn Heffner int probe_size; 159191cc17c0SIlpo Järvinen int size_needed; 15925d424d5aSJohn Heffner int copy; 15935d424d5aSJohn Heffner int mss_now; 15945d424d5aSJohn Heffner 15955d424d5aSJohn Heffner /* Not currently probing/verifying, 15965d424d5aSJohn Heffner * not in recovery, 15975d424d5aSJohn Heffner * have enough cwnd, and 15985d424d5aSJohn Heffner * not SACKing (the variable headers throw things off) */ 15995d424d5aSJohn Heffner if (!icsk->icsk_mtup.enabled || 16005d424d5aSJohn Heffner icsk->icsk_mtup.probe_size || 16015d424d5aSJohn Heffner inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 16025d424d5aSJohn Heffner tp->snd_cwnd < 11 || 1603cabeccbdSIlpo Järvinen tp->rx_opt.num_sacks || tp->rx_opt.dsack) 16045d424d5aSJohn Heffner return -1; 16055d424d5aSJohn Heffner 16065d424d5aSJohn Heffner /* Very simple search strategy: just double the MSS. */ 16070c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 16085d424d5aSJohn Heffner probe_size = 2 * tp->mss_cache; 160991cc17c0SIlpo Järvinen size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; 16105d424d5aSJohn Heffner if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) { 16115d424d5aSJohn Heffner /* TODO: set timer for probe_converge_event */ 16125d424d5aSJohn Heffner return -1; 16135d424d5aSJohn Heffner } 16145d424d5aSJohn Heffner 16155d424d5aSJohn Heffner /* Have enough data in the send queue to probe? */ 16167f9c33e5SIlpo Järvinen if (tp->write_seq - tp->snd_nxt < size_needed) 16175d424d5aSJohn Heffner return -1; 16185d424d5aSJohn Heffner 161991cc17c0SIlpo Järvinen if (tp->snd_wnd < size_needed) 16205d424d5aSJohn Heffner return -1; 162190840defSIlpo Järvinen if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) 16225d424d5aSJohn Heffner return 0; 16235d424d5aSJohn Heffner 1624d67c58e9SIlpo Järvinen /* Do we need to wait to drain cwnd? With none in flight, don't stall */ 1625d67c58e9SIlpo Järvinen if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { 1626d67c58e9SIlpo Järvinen if (!tcp_packets_in_flight(tp)) 16275d424d5aSJohn Heffner return -1; 16285d424d5aSJohn Heffner else 16295d424d5aSJohn Heffner return 0; 16305d424d5aSJohn Heffner } 16315d424d5aSJohn Heffner 16325d424d5aSJohn Heffner /* We're allowed to probe. Build it now. */ 16335d424d5aSJohn Heffner if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL) 16345d424d5aSJohn Heffner return -1; 16353ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 16363ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 16375d424d5aSJohn Heffner 1638fe067e8aSDavid S. Miller skb = tcp_send_head(sk); 16395d424d5aSJohn Heffner 16405d424d5aSJohn Heffner TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 16415d424d5aSJohn Heffner TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 1642a3433f35SChangli Gao TCP_SKB_CB(nskb)->flags = TCPHDR_ACK; 16435d424d5aSJohn Heffner TCP_SKB_CB(nskb)->sacked = 0; 16445d424d5aSJohn Heffner nskb->csum = 0; 164584fa7933SPatrick McHardy nskb->ip_summed = skb->ip_summed; 16465d424d5aSJohn Heffner 164750c4817eSIlpo Järvinen tcp_insert_write_queue_before(nskb, skb, sk); 164850c4817eSIlpo Järvinen 16495d424d5aSJohn Heffner len = 0; 1650234b6860SIlpo Järvinen tcp_for_write_queue_from_safe(skb, next, sk) { 16515d424d5aSJohn Heffner copy = min_t(int, skb->len, probe_size - len); 16525d424d5aSJohn Heffner if (nskb->ip_summed) 16535d424d5aSJohn Heffner skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); 16545d424d5aSJohn Heffner else 16555d424d5aSJohn Heffner nskb->csum = skb_copy_and_csum_bits(skb, 0, 1656056834d9SIlpo Järvinen skb_put(nskb, copy), 1657056834d9SIlpo Järvinen copy, nskb->csum); 16585d424d5aSJohn Heffner 16595d424d5aSJohn Heffner if (skb->len <= copy) { 16605d424d5aSJohn Heffner /* We've eaten all the data from this skb. 16615d424d5aSJohn Heffner * Throw it away. */ 16625d424d5aSJohn Heffner TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags; 1663fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 16643ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 16655d424d5aSJohn Heffner } else { 16665d424d5aSJohn Heffner TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags & 1667a3433f35SChangli Gao ~(TCPHDR_FIN|TCPHDR_PSH); 16685d424d5aSJohn Heffner if (!skb_shinfo(skb)->nr_frags) { 16695d424d5aSJohn Heffner skb_pull(skb, copy); 167084fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 1671056834d9SIlpo Järvinen skb->csum = csum_partial(skb->data, 1672056834d9SIlpo Järvinen skb->len, 0); 16735d424d5aSJohn Heffner } else { 16745d424d5aSJohn Heffner __pskb_trim_head(skb, copy); 16755d424d5aSJohn Heffner tcp_set_skb_tso_segs(sk, skb, mss_now); 16765d424d5aSJohn Heffner } 16775d424d5aSJohn Heffner TCP_SKB_CB(skb)->seq += copy; 16785d424d5aSJohn Heffner } 16795d424d5aSJohn Heffner 16805d424d5aSJohn Heffner len += copy; 1681234b6860SIlpo Järvinen 1682234b6860SIlpo Järvinen if (len >= probe_size) 1683234b6860SIlpo Järvinen break; 16845d424d5aSJohn Heffner } 16855d424d5aSJohn Heffner tcp_init_tso_segs(sk, nskb, nskb->len); 16865d424d5aSJohn Heffner 16875d424d5aSJohn Heffner /* We're ready to send. If this fails, the probe will 16885d424d5aSJohn Heffner * be resegmented into mss-sized pieces by tcp_write_xmit(). */ 16895d424d5aSJohn Heffner TCP_SKB_CB(nskb)->when = tcp_time_stamp; 16905d424d5aSJohn Heffner if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 16915d424d5aSJohn Heffner /* Decrement cwnd here because we are sending 16925d424d5aSJohn Heffner * effectively two packets. */ 16935d424d5aSJohn Heffner tp->snd_cwnd--; 169466f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, nskb); 16955d424d5aSJohn Heffner 16965d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 16970e7b1368SJohn Heffner tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 16980e7b1368SJohn Heffner tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 16995d424d5aSJohn Heffner 17005d424d5aSJohn Heffner return 1; 17015d424d5aSJohn Heffner } 17025d424d5aSJohn Heffner 17035d424d5aSJohn Heffner return -1; 17045d424d5aSJohn Heffner } 17055d424d5aSJohn Heffner 17061da177e4SLinus Torvalds /* This routine writes packets to the network. It advances the 17071da177e4SLinus Torvalds * send_head. This happens as incoming acks open up the remote 17081da177e4SLinus Torvalds * window for us. 17091da177e4SLinus Torvalds * 1710f8269a49SIlpo Järvinen * LARGESEND note: !tcp_urg_mode is overkill, only frames between 1711f8269a49SIlpo Järvinen * snd_up-64k-mss .. snd_up cannot be large. However, taking into 1712f8269a49SIlpo Järvinen * account rare use of URG, this is not a big flaw. 1713f8269a49SIlpo Järvinen * 17141da177e4SLinus Torvalds * Returns 1, if no segments are in flight and we have queued segments, but 17151da177e4SLinus Torvalds * cannot send anything now because of SWS or another problem. 17161da177e4SLinus Torvalds */ 1717d5dd9175SIlpo Järvinen static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 1718d5dd9175SIlpo Järvinen int push_one, gfp_t gfp) 17191da177e4SLinus Torvalds { 17201da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 172192df7b51SDavid S. Miller struct sk_buff *skb; 1722c1b4a7e6SDavid S. Miller unsigned int tso_segs, sent_pkts; 1723c1b4a7e6SDavid S. Miller int cwnd_quota; 17245d424d5aSJohn Heffner int result; 17251da177e4SLinus Torvalds 1726c1b4a7e6SDavid S. Miller sent_pkts = 0; 17275d424d5aSJohn Heffner 1728d5dd9175SIlpo Järvinen if (!push_one) { 17295d424d5aSJohn Heffner /* Do MTU probing. */ 1730d5dd9175SIlpo Järvinen result = tcp_mtu_probe(sk); 1731d5dd9175SIlpo Järvinen if (!result) { 17325d424d5aSJohn Heffner return 0; 17335d424d5aSJohn Heffner } else if (result > 0) { 17345d424d5aSJohn Heffner sent_pkts = 1; 17355d424d5aSJohn Heffner } 1736d5dd9175SIlpo Järvinen } 17375d424d5aSJohn Heffner 1738fe067e8aSDavid S. Miller while ((skb = tcp_send_head(sk))) { 1739c8ac3774SHerbert Xu unsigned int limit; 1740c8ac3774SHerbert Xu 1741b68e9f85SHerbert Xu tso_segs = tcp_init_tso_segs(sk, skb, mss_now); 1742c1b4a7e6SDavid S. Miller BUG_ON(!tso_segs); 1743c1b4a7e6SDavid S. Miller 1744b68e9f85SHerbert Xu cwnd_quota = tcp_cwnd_test(tp, skb); 1745b68e9f85SHerbert Xu if (!cwnd_quota) 1746b68e9f85SHerbert Xu break; 1747b68e9f85SHerbert Xu 1748b68e9f85SHerbert Xu if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) 1749b68e9f85SHerbert Xu break; 1750b68e9f85SHerbert Xu 1751c1b4a7e6SDavid S. Miller if (tso_segs == 1) { 1752aa93466bSDavid S. Miller if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 1753aa93466bSDavid S. Miller (tcp_skb_is_last(sk, skb) ? 1754aa93466bSDavid S. Miller nonagle : TCP_NAGLE_PUSH)))) 1755aa93466bSDavid S. Miller break; 1756c1b4a7e6SDavid S. Miller } else { 1757d5dd9175SIlpo Järvinen if (!push_one && tcp_tso_should_defer(sk, skb)) 1758aa93466bSDavid S. Miller break; 1759c1b4a7e6SDavid S. Miller } 1760aa93466bSDavid S. Miller 1761c8ac3774SHerbert Xu limit = mss_now; 1762f8269a49SIlpo Järvinen if (tso_segs > 1 && !tcp_urg_mode(tp)) 17630e3a4803SIlpo Järvinen limit = tcp_mss_split_point(sk, skb, mss_now, 17640e3a4803SIlpo Järvinen cwnd_quota); 1765c8ac3774SHerbert Xu 1766c8ac3774SHerbert Xu if (skb->len > limit && 1767c4ead4c5SEric Dumazet unlikely(tso_fragment(sk, skb, limit, mss_now, gfp))) 17681da177e4SLinus Torvalds break; 17691da177e4SLinus Torvalds 17701da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 1771c1b4a7e6SDavid S. Miller 1772d5dd9175SIlpo Järvinen if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) 17731da177e4SLinus Torvalds break; 17741da177e4SLinus Torvalds 17751da177e4SLinus Torvalds /* Advance the send_head. This one is sent out. 17761da177e4SLinus Torvalds * This call will increment packets_out. 17771da177e4SLinus Torvalds */ 177866f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 17791da177e4SLinus Torvalds 17801da177e4SLinus Torvalds tcp_minshall_update(tp, mss_now, skb); 1781aa93466bSDavid S. Miller sent_pkts++; 1782d5dd9175SIlpo Järvinen 1783d5dd9175SIlpo Järvinen if (push_one) 1784d5dd9175SIlpo Järvinen break; 17851da177e4SLinus Torvalds } 17861da177e4SLinus Torvalds 1787aa93466bSDavid S. Miller if (likely(sent_pkts)) { 17889e412ba7SIlpo Järvinen tcp_cwnd_validate(sk); 17891da177e4SLinus Torvalds return 0; 17901da177e4SLinus Torvalds } 1791fe067e8aSDavid S. Miller return !tp->packets_out && tcp_send_head(sk); 17921da177e4SLinus Torvalds } 17931da177e4SLinus Torvalds 1794a762a980SDavid S. Miller /* Push out any pending frames which were held back due to 1795a762a980SDavid S. Miller * TCP_CORK or attempt at coalescing tiny packets. 1796a762a980SDavid S. Miller * The socket must be locked by the caller. 1797a762a980SDavid S. Miller */ 17989e412ba7SIlpo Järvinen void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 17999e412ba7SIlpo Järvinen int nonagle) 1800a762a980SDavid S. Miller { 1801726e07a8SIlpo Järvinen /* If we are closed, the bytes will have to remain here. 1802726e07a8SIlpo Järvinen * In time closedown will finish, we empty the write queue and 1803726e07a8SIlpo Järvinen * all will be happy. 1804726e07a8SIlpo Järvinen */ 1805726e07a8SIlpo Järvinen if (unlikely(sk->sk_state == TCP_CLOSE)) 1806726e07a8SIlpo Järvinen return; 1807726e07a8SIlpo Järvinen 1808d5dd9175SIlpo Järvinen if (tcp_write_xmit(sk, cur_mss, nonagle, 0, GFP_ATOMIC)) 18099e412ba7SIlpo Järvinen tcp_check_probe_timer(sk); 1810a762a980SDavid S. Miller } 1811a762a980SDavid S. Miller 1812c1b4a7e6SDavid S. Miller /* Send _single_ skb sitting at the send head. This function requires 1813c1b4a7e6SDavid S. Miller * true push pending frames to setup probe timer etc. 1814c1b4a7e6SDavid S. Miller */ 1815c1b4a7e6SDavid S. Miller void tcp_push_one(struct sock *sk, unsigned int mss_now) 1816c1b4a7e6SDavid S. Miller { 1817fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 1818c1b4a7e6SDavid S. Miller 1819c1b4a7e6SDavid S. Miller BUG_ON(!skb || skb->len < mss_now); 1820c1b4a7e6SDavid S. Miller 1821d5dd9175SIlpo Järvinen tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); 1822c1b4a7e6SDavid S. Miller } 1823c1b4a7e6SDavid S. Miller 18241da177e4SLinus Torvalds /* This function returns the amount that we can raise the 18251da177e4SLinus Torvalds * usable window based on the following constraints 18261da177e4SLinus Torvalds * 18271da177e4SLinus Torvalds * 1. The window can never be shrunk once it is offered (RFC 793) 18281da177e4SLinus Torvalds * 2. We limit memory per socket 18291da177e4SLinus Torvalds * 18301da177e4SLinus Torvalds * RFC 1122: 18311da177e4SLinus Torvalds * "the suggested [SWS] avoidance algorithm for the receiver is to keep 18321da177e4SLinus Torvalds * RECV.NEXT + RCV.WIN fixed until: 18331da177e4SLinus Torvalds * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 18341da177e4SLinus Torvalds * 18351da177e4SLinus Torvalds * i.e. don't raise the right edge of the window until you can raise 18361da177e4SLinus Torvalds * it at least MSS bytes. 18371da177e4SLinus Torvalds * 18381da177e4SLinus Torvalds * Unfortunately, the recommended algorithm breaks header prediction, 18391da177e4SLinus Torvalds * since header prediction assumes th->window stays fixed. 18401da177e4SLinus Torvalds * 18411da177e4SLinus Torvalds * Strictly speaking, keeping th->window fixed violates the receiver 18421da177e4SLinus Torvalds * side SWS prevention criteria. The problem is that under this rule 18431da177e4SLinus Torvalds * a stream of single byte packets will cause the right side of the 18441da177e4SLinus Torvalds * window to always advance by a single byte. 18451da177e4SLinus Torvalds * 18461da177e4SLinus Torvalds * Of course, if the sender implements sender side SWS prevention 18471da177e4SLinus Torvalds * then this will not be a problem. 18481da177e4SLinus Torvalds * 18491da177e4SLinus Torvalds * BSD seems to make the following compromise: 18501da177e4SLinus Torvalds * 18511da177e4SLinus Torvalds * If the free space is less than the 1/4 of the maximum 18521da177e4SLinus Torvalds * space available and the free space is less than 1/2 mss, 18531da177e4SLinus Torvalds * then set the window to 0. 18541da177e4SLinus Torvalds * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 18551da177e4SLinus Torvalds * Otherwise, just prevent the window from shrinking 18561da177e4SLinus Torvalds * and from being larger than the largest representable value. 18571da177e4SLinus Torvalds * 18581da177e4SLinus Torvalds * This prevents incremental opening of the window in the regime 18591da177e4SLinus Torvalds * where TCP is limited by the speed of the reader side taking 18601da177e4SLinus Torvalds * data out of the TCP receive queue. It does nothing about 18611da177e4SLinus Torvalds * those cases where the window is constrained on the sender side 18621da177e4SLinus Torvalds * because the pipeline is full. 18631da177e4SLinus Torvalds * 18641da177e4SLinus Torvalds * BSD also seems to "accidentally" limit itself to windows that are a 18651da177e4SLinus Torvalds * multiple of MSS, at least until the free space gets quite small. 18661da177e4SLinus Torvalds * This would appear to be a side effect of the mbuf implementation. 18671da177e4SLinus Torvalds * Combining these two algorithms results in the observed behavior 18681da177e4SLinus Torvalds * of having a fixed window size at almost all times. 18691da177e4SLinus Torvalds * 18701da177e4SLinus Torvalds * Below we obtain similar behavior by forcing the offered window to 18711da177e4SLinus Torvalds * a multiple of the mss when it is feasible to do so. 18721da177e4SLinus Torvalds * 18731da177e4SLinus Torvalds * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 18741da177e4SLinus Torvalds * Regular options like TIMESTAMP are taken into account. 18751da177e4SLinus Torvalds */ 18761da177e4SLinus Torvalds u32 __tcp_select_window(struct sock *sk) 18771da177e4SLinus Torvalds { 1878463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 18791da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1880caa20d9aSStephen Hemminger /* MSS for the peer's data. Previous versions used mss_clamp 18811da177e4SLinus Torvalds * here. I don't know if the value based on our guesses 18821da177e4SLinus Torvalds * of peer's MSS is better for the performance. It's more correct 18831da177e4SLinus Torvalds * but may be worse for the performance because of rcv_mss 18841da177e4SLinus Torvalds * fluctuations. --SAW 1998/11/1 18851da177e4SLinus Torvalds */ 1886463c84b9SArnaldo Carvalho de Melo int mss = icsk->icsk_ack.rcv_mss; 18871da177e4SLinus Torvalds int free_space = tcp_space(sk); 18881da177e4SLinus Torvalds int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk)); 18891da177e4SLinus Torvalds int window; 18901da177e4SLinus Torvalds 18911da177e4SLinus Torvalds if (mss > full_space) 18921da177e4SLinus Torvalds mss = full_space; 18931da177e4SLinus Torvalds 1894b92edbe0SEric Dumazet if (free_space < (full_space >> 1)) { 1895463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.quick = 0; 18961da177e4SLinus Torvalds 18971da177e4SLinus Torvalds if (tcp_memory_pressure) 1898056834d9SIlpo Järvinen tp->rcv_ssthresh = min(tp->rcv_ssthresh, 1899056834d9SIlpo Järvinen 4U * tp->advmss); 19001da177e4SLinus Torvalds 19011da177e4SLinus Torvalds if (free_space < mss) 19021da177e4SLinus Torvalds return 0; 19031da177e4SLinus Torvalds } 19041da177e4SLinus Torvalds 19051da177e4SLinus Torvalds if (free_space > tp->rcv_ssthresh) 19061da177e4SLinus Torvalds free_space = tp->rcv_ssthresh; 19071da177e4SLinus Torvalds 19081da177e4SLinus Torvalds /* Don't do rounding if we are using window scaling, since the 19091da177e4SLinus Torvalds * scaled window will not line up with the MSS boundary anyway. 19101da177e4SLinus Torvalds */ 19111da177e4SLinus Torvalds window = tp->rcv_wnd; 19121da177e4SLinus Torvalds if (tp->rx_opt.rcv_wscale) { 19131da177e4SLinus Torvalds window = free_space; 19141da177e4SLinus Torvalds 19151da177e4SLinus Torvalds /* Advertise enough space so that it won't get scaled away. 19161da177e4SLinus Torvalds * Import case: prevent zero window announcement if 19171da177e4SLinus Torvalds * 1<<rcv_wscale > mss. 19181da177e4SLinus Torvalds */ 19191da177e4SLinus Torvalds if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window) 19201da177e4SLinus Torvalds window = (((window >> tp->rx_opt.rcv_wscale) + 1) 19211da177e4SLinus Torvalds << tp->rx_opt.rcv_wscale); 19221da177e4SLinus Torvalds } else { 19231da177e4SLinus Torvalds /* Get the largest window that is a nice multiple of mss. 19241da177e4SLinus Torvalds * Window clamp already applied above. 19251da177e4SLinus Torvalds * If our current window offering is within 1 mss of the 19261da177e4SLinus Torvalds * free space we just keep it. This prevents the divide 19271da177e4SLinus Torvalds * and multiply from happening most of the time. 19281da177e4SLinus Torvalds * We also don't do any window rounding when the free space 19291da177e4SLinus Torvalds * is too small. 19301da177e4SLinus Torvalds */ 19311da177e4SLinus Torvalds if (window <= free_space - mss || window > free_space) 19321da177e4SLinus Torvalds window = (free_space / mss) * mss; 193384565070SJohn Heffner else if (mss == full_space && 1934b92edbe0SEric Dumazet free_space > window + (full_space >> 1)) 193584565070SJohn Heffner window = free_space; 19361da177e4SLinus Torvalds } 19371da177e4SLinus Torvalds 19381da177e4SLinus Torvalds return window; 19391da177e4SLinus Torvalds } 19401da177e4SLinus Torvalds 19414a17fc3aSIlpo Järvinen /* Collapses two adjacent SKB's during retransmission. */ 19424a17fc3aSIlpo Järvinen static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) 19431da177e4SLinus Torvalds { 19441da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1945fe067e8aSDavid S. Miller struct sk_buff *next_skb = tcp_write_queue_next(sk, skb); 1946058dc334SIlpo Järvinen int skb_size, next_skb_size; 19471da177e4SLinus Torvalds 1948058dc334SIlpo Järvinen skb_size = skb->len; 1949058dc334SIlpo Järvinen next_skb_size = next_skb->len; 19501da177e4SLinus Torvalds 1951058dc334SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); 19521da177e4SLinus Torvalds 19536859d494SIlpo Järvinen tcp_highest_sack_combine(sk, next_skb, skb); 1954a6963a6bSIlpo Järvinen 1955fe067e8aSDavid S. Miller tcp_unlink_write_queue(next_skb, sk); 19561da177e4SLinus Torvalds 1957058dc334SIlpo Järvinen skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size), 19581a4e2d09SArnaldo Carvalho de Melo next_skb_size); 19591da177e4SLinus Torvalds 196052d570aaSJarek Poplawski if (next_skb->ip_summed == CHECKSUM_PARTIAL) 196152d570aaSJarek Poplawski skb->ip_summed = CHECKSUM_PARTIAL; 19621da177e4SLinus Torvalds 196384fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 19641da177e4SLinus Torvalds skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); 19651da177e4SLinus Torvalds 19661da177e4SLinus Torvalds /* Update sequence range on original skb. */ 19671da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 19681da177e4SLinus Torvalds 1969e6c7d085SIlpo Järvinen /* Merge over control information. This moves PSH/FIN etc. over */ 1970e6c7d085SIlpo Järvinen TCP_SKB_CB(skb)->flags |= TCP_SKB_CB(next_skb)->flags; 19711da177e4SLinus Torvalds 19721da177e4SLinus Torvalds /* All done, get rid of second SKB and account for it so 19731da177e4SLinus Torvalds * packet counting does not break. 19741da177e4SLinus Torvalds */ 19754828e7f4SIlpo Järvinen TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; 1976b7689205SIlpo Järvinen 1977b7689205SIlpo Järvinen /* changed transmit queue under us so clear hints */ 1978ef9da47cSIlpo Järvinen tcp_clear_retrans_hints_partial(tp); 1979ef9da47cSIlpo Järvinen if (next_skb == tp->retransmit_skb_hint) 1980ef9da47cSIlpo Järvinen tp->retransmit_skb_hint = skb; 1981b7689205SIlpo Järvinen 1982797108d1SIlpo Järvinen tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); 1983797108d1SIlpo Järvinen 19843ab224beSHideo Aoki sk_wmem_free_skb(sk, next_skb); 19851da177e4SLinus Torvalds } 19861da177e4SLinus Torvalds 198767edfef7SAndi Kleen /* Check if coalescing SKBs is legal. */ 19884a17fc3aSIlpo Järvinen static int tcp_can_collapse(struct sock *sk, struct sk_buff *skb) 19894a17fc3aSIlpo Järvinen { 19904a17fc3aSIlpo Järvinen if (tcp_skb_pcount(skb) > 1) 19914a17fc3aSIlpo Järvinen return 0; 19924a17fc3aSIlpo Järvinen /* TODO: SACK collapsing could be used to remove this condition */ 19934a17fc3aSIlpo Järvinen if (skb_shinfo(skb)->nr_frags != 0) 19944a17fc3aSIlpo Järvinen return 0; 19954a17fc3aSIlpo Järvinen if (skb_cloned(skb)) 19964a17fc3aSIlpo Järvinen return 0; 19974a17fc3aSIlpo Järvinen if (skb == tcp_send_head(sk)) 19984a17fc3aSIlpo Järvinen return 0; 19994a17fc3aSIlpo Järvinen /* Some heurestics for collapsing over SACK'd could be invented */ 20004a17fc3aSIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 20014a17fc3aSIlpo Järvinen return 0; 20024a17fc3aSIlpo Järvinen 20034a17fc3aSIlpo Järvinen return 1; 20044a17fc3aSIlpo Järvinen } 20054a17fc3aSIlpo Järvinen 200667edfef7SAndi Kleen /* Collapse packets in the retransmit queue to make to create 200767edfef7SAndi Kleen * less packets on the wire. This is only done on retransmission. 200867edfef7SAndi Kleen */ 20094a17fc3aSIlpo Järvinen static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, 20104a17fc3aSIlpo Järvinen int space) 20114a17fc3aSIlpo Järvinen { 20124a17fc3aSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 20134a17fc3aSIlpo Järvinen struct sk_buff *skb = to, *tmp; 20144a17fc3aSIlpo Järvinen int first = 1; 20154a17fc3aSIlpo Järvinen 20164a17fc3aSIlpo Järvinen if (!sysctl_tcp_retrans_collapse) 20174a17fc3aSIlpo Järvinen return; 2018a3433f35SChangli Gao if (TCP_SKB_CB(skb)->flags & TCPHDR_SYN) 20194a17fc3aSIlpo Järvinen return; 20204a17fc3aSIlpo Järvinen 20214a17fc3aSIlpo Järvinen tcp_for_write_queue_from_safe(skb, tmp, sk) { 20224a17fc3aSIlpo Järvinen if (!tcp_can_collapse(sk, skb)) 20234a17fc3aSIlpo Järvinen break; 20244a17fc3aSIlpo Järvinen 20254a17fc3aSIlpo Järvinen space -= skb->len; 20264a17fc3aSIlpo Järvinen 20274a17fc3aSIlpo Järvinen if (first) { 20284a17fc3aSIlpo Järvinen first = 0; 20294a17fc3aSIlpo Järvinen continue; 20304a17fc3aSIlpo Järvinen } 20314a17fc3aSIlpo Järvinen 20324a17fc3aSIlpo Järvinen if (space < 0) 20334a17fc3aSIlpo Järvinen break; 20344a17fc3aSIlpo Järvinen /* Punt if not enough space exists in the first SKB for 20354a17fc3aSIlpo Järvinen * the data in the second 20364a17fc3aSIlpo Järvinen */ 20374a17fc3aSIlpo Järvinen if (skb->len > skb_tailroom(to)) 20384a17fc3aSIlpo Järvinen break; 20394a17fc3aSIlpo Järvinen 20404a17fc3aSIlpo Järvinen if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) 20414a17fc3aSIlpo Järvinen break; 20424a17fc3aSIlpo Järvinen 20434a17fc3aSIlpo Järvinen tcp_collapse_retrans(sk, to); 20444a17fc3aSIlpo Järvinen } 20454a17fc3aSIlpo Järvinen } 20464a17fc3aSIlpo Järvinen 20471da177e4SLinus Torvalds /* This retransmits one SKB. Policy decisions and retransmit queue 20481da177e4SLinus Torvalds * state updates are done by the caller. Returns non-zero if an 20491da177e4SLinus Torvalds * error occurred which prevented the send. 20501da177e4SLinus Torvalds */ 20511da177e4SLinus Torvalds int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) 20521da177e4SLinus Torvalds { 20531da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 20545d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 20557d227cd2SSridhar Samudrala unsigned int cur_mss; 20561da177e4SLinus Torvalds int err; 20571da177e4SLinus Torvalds 20585d424d5aSJohn Heffner /* Inconslusive MTU probe */ 20595d424d5aSJohn Heffner if (icsk->icsk_mtup.probe_size) { 20605d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 20615d424d5aSJohn Heffner } 20625d424d5aSJohn Heffner 20631da177e4SLinus Torvalds /* Do not sent more than we queued. 1/4 is reserved for possible 2064caa20d9aSStephen Hemminger * copying overhead: fragmentation, tunneling, mangling etc. 20651da177e4SLinus Torvalds */ 20661da177e4SLinus Torvalds if (atomic_read(&sk->sk_wmem_alloc) > 20671da177e4SLinus Torvalds min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf)) 20681da177e4SLinus Torvalds return -EAGAIN; 20691da177e4SLinus Torvalds 20701da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 20711da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 20721da177e4SLinus Torvalds BUG(); 20731da177e4SLinus Torvalds if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 20741da177e4SLinus Torvalds return -ENOMEM; 20751da177e4SLinus Torvalds } 20761da177e4SLinus Torvalds 20777d227cd2SSridhar Samudrala if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 20787d227cd2SSridhar Samudrala return -EHOSTUNREACH; /* Routing failure or similar. */ 20797d227cd2SSridhar Samudrala 20800c54b85fSIlpo Järvinen cur_mss = tcp_current_mss(sk); 20817d227cd2SSridhar Samudrala 20821da177e4SLinus Torvalds /* If receiver has shrunk his window, and skb is out of 20831da177e4SLinus Torvalds * new window, do not retransmit it. The exception is the 20841da177e4SLinus Torvalds * case, when window is shrunk to zero. In this case 20851da177e4SLinus Torvalds * our retransmit serves as a zero window probe. 20861da177e4SLinus Torvalds */ 20879d4fb27dSJoe Perches if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && 20889d4fb27dSJoe Perches TCP_SKB_CB(skb)->seq != tp->snd_una) 20891da177e4SLinus Torvalds return -EAGAIN; 20901da177e4SLinus Torvalds 20911da177e4SLinus Torvalds if (skb->len > cur_mss) { 2092846998aeSDavid S. Miller if (tcp_fragment(sk, skb, cur_mss, cur_mss)) 20931da177e4SLinus Torvalds return -ENOMEM; /* We'll try again later. */ 209402276f3cSIlpo Järvinen } else { 20959eb9362eSIlpo Järvinen int oldpcount = tcp_skb_pcount(skb); 20969eb9362eSIlpo Järvinen 20979eb9362eSIlpo Järvinen if (unlikely(oldpcount > 1)) { 209802276f3cSIlpo Järvinen tcp_init_tso_segs(sk, skb, cur_mss); 20999eb9362eSIlpo Järvinen tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb)); 21009eb9362eSIlpo Järvinen } 21011da177e4SLinus Torvalds } 21021da177e4SLinus Torvalds 21031da177e4SLinus Torvalds tcp_retrans_try_collapse(sk, skb, cur_mss); 21041da177e4SLinus Torvalds 21051da177e4SLinus Torvalds /* Some Solaris stacks overoptimize and ignore the FIN on a 21061da177e4SLinus Torvalds * retransmit when old data is attached. So strip it off 21071da177e4SLinus Torvalds * since it is cheap to do so and saves bytes on the network. 21081da177e4SLinus Torvalds */ 21091da177e4SLinus Torvalds if (skb->len > 0 && 2110a3433f35SChangli Gao (TCP_SKB_CB(skb)->flags & TCPHDR_FIN) && 21111da177e4SLinus Torvalds tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) { 21121da177e4SLinus Torvalds if (!pskb_trim(skb, 0)) { 2113e870a8efSIlpo Järvinen /* Reuse, even though it does some unnecessary work */ 2114e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1, 2115e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->flags); 21161da177e4SLinus Torvalds skb->ip_summed = CHECKSUM_NONE; 21171da177e4SLinus Torvalds } 21181da177e4SLinus Torvalds } 21191da177e4SLinus Torvalds 21201da177e4SLinus Torvalds /* Make a copy, if the first transmission SKB clone we made 21211da177e4SLinus Torvalds * is still in somebody's hands, else make a clone. 21221da177e4SLinus Torvalds */ 21231da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 21241da177e4SLinus Torvalds 2125dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 21261da177e4SLinus Torvalds 21271da177e4SLinus Torvalds if (err == 0) { 21281da177e4SLinus Torvalds /* Update global TCP statistics. */ 212981cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); 21301da177e4SLinus Torvalds 21311da177e4SLinus Torvalds tp->total_retrans++; 21321da177e4SLinus Torvalds 21331da177e4SLinus Torvalds #if FASTRETRANS_DEBUG > 0 21341da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 21351da177e4SLinus Torvalds if (net_ratelimit()) 21361da177e4SLinus Torvalds printk(KERN_DEBUG "retrans_out leaked.\n"); 21371da177e4SLinus Torvalds } 21381da177e4SLinus Torvalds #endif 2139b08d6cb2SIlpo Järvinen if (!tp->retrans_out) 2140b08d6cb2SIlpo Järvinen tp->lost_retrans_low = tp->snd_nxt; 21411da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 21421da177e4SLinus Torvalds tp->retrans_out += tcp_skb_pcount(skb); 21431da177e4SLinus Torvalds 21441da177e4SLinus Torvalds /* Save stamp of the first retransmit. */ 21451da177e4SLinus Torvalds if (!tp->retrans_stamp) 21461da177e4SLinus Torvalds tp->retrans_stamp = TCP_SKB_CB(skb)->when; 21471da177e4SLinus Torvalds 21481da177e4SLinus Torvalds tp->undo_retrans++; 21491da177e4SLinus Torvalds 21501da177e4SLinus Torvalds /* snd_nxt is stored to detect loss of retransmitted segment, 21511da177e4SLinus Torvalds * see tcp_input.c tcp_sacktag_write_queue(). 21521da177e4SLinus Torvalds */ 21531da177e4SLinus Torvalds TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt; 21541da177e4SLinus Torvalds } 21551da177e4SLinus Torvalds return err; 21561da177e4SLinus Torvalds } 21571da177e4SLinus Torvalds 215867edfef7SAndi Kleen /* Check if we forward retransmits are possible in the current 215967edfef7SAndi Kleen * window/congestion state. 216067edfef7SAndi Kleen */ 2161b5afe7bcSIlpo Järvinen static int tcp_can_forward_retransmit(struct sock *sk) 2162b5afe7bcSIlpo Järvinen { 2163b5afe7bcSIlpo Järvinen const struct inet_connection_sock *icsk = inet_csk(sk); 2164b5afe7bcSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 2165b5afe7bcSIlpo Järvinen 2166b5afe7bcSIlpo Järvinen /* Forward retransmissions are possible only during Recovery. */ 2167b5afe7bcSIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Recovery) 2168b5afe7bcSIlpo Järvinen return 0; 2169b5afe7bcSIlpo Järvinen 2170b5afe7bcSIlpo Järvinen /* No forward retransmissions in Reno are possible. */ 2171b5afe7bcSIlpo Järvinen if (tcp_is_reno(tp)) 2172b5afe7bcSIlpo Järvinen return 0; 2173b5afe7bcSIlpo Järvinen 2174b5afe7bcSIlpo Järvinen /* Yeah, we have to make difficult choice between forward transmission 2175b5afe7bcSIlpo Järvinen * and retransmission... Both ways have their merits... 2176b5afe7bcSIlpo Järvinen * 2177b5afe7bcSIlpo Järvinen * For now we do not retransmit anything, while we have some new 2178b5afe7bcSIlpo Järvinen * segments to send. In the other cases, follow rule 3 for 2179b5afe7bcSIlpo Järvinen * NextSeg() specified in RFC3517. 2180b5afe7bcSIlpo Järvinen */ 2181b5afe7bcSIlpo Järvinen 2182b5afe7bcSIlpo Järvinen if (tcp_may_send_now(sk)) 2183b5afe7bcSIlpo Järvinen return 0; 2184b5afe7bcSIlpo Järvinen 2185b5afe7bcSIlpo Järvinen return 1; 2186b5afe7bcSIlpo Järvinen } 2187b5afe7bcSIlpo Järvinen 21881da177e4SLinus Torvalds /* This gets called after a retransmit timeout, and the initially 21891da177e4SLinus Torvalds * retransmitted data is acknowledged. It tries to continue 21901da177e4SLinus Torvalds * resending the rest of the retransmit queue, until either 21911da177e4SLinus Torvalds * we've sent it all or the congestion window limit is reached. 21921da177e4SLinus Torvalds * If doing SACK, the first ACK which comes back for a timeout 21931da177e4SLinus Torvalds * based retransmit packet might feed us FACK information again. 21941da177e4SLinus Torvalds * If so, we use it to avoid unnecessarily retransmissions. 21951da177e4SLinus Torvalds */ 21961da177e4SLinus Torvalds void tcp_xmit_retransmit_queue(struct sock *sk) 21971da177e4SLinus Torvalds { 21986687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 21991da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 22001da177e4SLinus Torvalds struct sk_buff *skb; 22010e1c54c2SIlpo Järvinen struct sk_buff *hole = NULL; 2202618d9f25SIlpo Järvinen u32 last_lost; 220361eb55f4SIlpo Järvinen int mib_idx; 22040e1c54c2SIlpo Järvinen int fwd_rexmitting = 0; 22056a438bbeSStephen Hemminger 220645e77d31SIlpo Järvinen if (!tp->packets_out) 220745e77d31SIlpo Järvinen return; 220845e77d31SIlpo Järvinen 220908ebd172SIlpo Järvinen if (!tp->lost_out) 221008ebd172SIlpo Järvinen tp->retransmit_high = tp->snd_una; 221108ebd172SIlpo Järvinen 2212618d9f25SIlpo Järvinen if (tp->retransmit_skb_hint) { 22136a438bbeSStephen Hemminger skb = tp->retransmit_skb_hint; 2214618d9f25SIlpo Järvinen last_lost = TCP_SKB_CB(skb)->end_seq; 2215618d9f25SIlpo Järvinen if (after(last_lost, tp->retransmit_high)) 2216618d9f25SIlpo Järvinen last_lost = tp->retransmit_high; 2217618d9f25SIlpo Järvinen } else { 2218fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 2219618d9f25SIlpo Järvinen last_lost = tp->snd_una; 2220618d9f25SIlpo Järvinen } 22211da177e4SLinus Torvalds 2222fe067e8aSDavid S. Miller tcp_for_write_queue_from(skb, sk) { 22231da177e4SLinus Torvalds __u8 sacked = TCP_SKB_CB(skb)->sacked; 22241da177e4SLinus Torvalds 2225fe067e8aSDavid S. Miller if (skb == tcp_send_head(sk)) 2226fe067e8aSDavid S. Miller break; 22276a438bbeSStephen Hemminger /* we could do better than to assign each time */ 22280e1c54c2SIlpo Järvinen if (hole == NULL) 22296a438bbeSStephen Hemminger tp->retransmit_skb_hint = skb; 22306a438bbeSStephen Hemminger 22311da177e4SLinus Torvalds /* Assume this retransmit will generate 22321da177e4SLinus Torvalds * only one packet for congestion window 22331da177e4SLinus Torvalds * calculation purposes. This works because 22341da177e4SLinus Torvalds * tcp_retransmit_skb() will chop up the 22351da177e4SLinus Torvalds * packet to be MSS sized and all the 22361da177e4SLinus Torvalds * packet counting works out. 22371da177e4SLinus Torvalds */ 22381da177e4SLinus Torvalds if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) 22391da177e4SLinus Torvalds return; 22400e1c54c2SIlpo Järvinen 22410e1c54c2SIlpo Järvinen if (fwd_rexmitting) { 22420e1c54c2SIlpo Järvinen begin_fwd: 22430e1c54c2SIlpo Järvinen if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp))) 2244006f582cSIlpo Järvinen break; 22450e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFORWARDRETRANS; 22460e1c54c2SIlpo Järvinen 22470e1c54c2SIlpo Järvinen } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) { 2248618d9f25SIlpo Järvinen tp->retransmit_high = last_lost; 22490e1c54c2SIlpo Järvinen if (!tcp_can_forward_retransmit(sk)) 22500e1c54c2SIlpo Järvinen break; 22510e1c54c2SIlpo Järvinen /* Backtrack if necessary to non-L'ed skb */ 22520e1c54c2SIlpo Järvinen if (hole != NULL) { 22530e1c54c2SIlpo Järvinen skb = hole; 22540e1c54c2SIlpo Järvinen hole = NULL; 22550e1c54c2SIlpo Järvinen } 22560e1c54c2SIlpo Järvinen fwd_rexmitting = 1; 22570e1c54c2SIlpo Järvinen goto begin_fwd; 22580e1c54c2SIlpo Järvinen 22590e1c54c2SIlpo Järvinen } else if (!(sacked & TCPCB_LOST)) { 2260ac11ba75SIlpo Järvinen if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) 22610e1c54c2SIlpo Järvinen hole = skb; 226261eb55f4SIlpo Järvinen continue; 22631da177e4SLinus Torvalds 22640e1c54c2SIlpo Järvinen } else { 2265618d9f25SIlpo Järvinen last_lost = TCP_SKB_CB(skb)->end_seq; 22660e1c54c2SIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Loss) 22670e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFASTRETRANS; 22680e1c54c2SIlpo Järvinen else 22690e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; 22700e1c54c2SIlpo Järvinen } 22710e1c54c2SIlpo Järvinen 22720e1c54c2SIlpo Järvinen if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) 227361eb55f4SIlpo Järvinen continue; 227440b215e5SPavel Emelyanov 2275f0ceb0edSIlpo Järvinen if (tcp_retransmit_skb(sk, skb)) 22761da177e4SLinus Torvalds return; 2277de0744afSPavel Emelyanov NET_INC_STATS_BH(sock_net(sk), mib_idx); 22781da177e4SLinus Torvalds 2279fe067e8aSDavid S. Miller if (skb == tcp_write_queue_head(sk)) 2280463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 22813f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, 22823f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 22831da177e4SLinus Torvalds } 22841da177e4SLinus Torvalds } 22851da177e4SLinus Torvalds 22861da177e4SLinus Torvalds /* Send a fin. The caller locks the socket for us. This cannot be 22871da177e4SLinus Torvalds * allowed to fail queueing a FIN frame under any circumstances. 22881da177e4SLinus Torvalds */ 22891da177e4SLinus Torvalds void tcp_send_fin(struct sock *sk) 22901da177e4SLinus Torvalds { 22911da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2292fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_write_queue_tail(sk); 22931da177e4SLinus Torvalds int mss_now; 22941da177e4SLinus Torvalds 22951da177e4SLinus Torvalds /* Optimization, tack on the FIN if we have a queue of 22961da177e4SLinus Torvalds * unsent frames. But be careful about outgoing SACKS 22971da177e4SLinus Torvalds * and IP options. 22981da177e4SLinus Torvalds */ 22990c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 23001da177e4SLinus Torvalds 2301fe067e8aSDavid S. Miller if (tcp_send_head(sk) != NULL) { 2302a3433f35SChangli Gao TCP_SKB_CB(skb)->flags |= TCPHDR_FIN; 23031da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq++; 23041da177e4SLinus Torvalds tp->write_seq++; 23051da177e4SLinus Torvalds } else { 23061da177e4SLinus Torvalds /* Socket is locked, keep trying until memory is available. */ 23071da177e4SLinus Torvalds for (;;) { 2308aa133076SWu Fengguang skb = alloc_skb_fclone(MAX_TCP_HEADER, 2309aa133076SWu Fengguang sk->sk_allocation); 23101da177e4SLinus Torvalds if (skb) 23111da177e4SLinus Torvalds break; 23121da177e4SLinus Torvalds yield(); 23131da177e4SLinus Torvalds } 23141da177e4SLinus Torvalds 23151da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 23161da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 23171da177e4SLinus Torvalds /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 2318e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tp->write_seq, 2319a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_FIN); 23201da177e4SLinus Torvalds tcp_queue_skb(sk, skb); 23211da177e4SLinus Torvalds } 23229e412ba7SIlpo Järvinen __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF); 23231da177e4SLinus Torvalds } 23241da177e4SLinus Torvalds 23251da177e4SLinus Torvalds /* We get here when a process closes a file descriptor (either due to 23261da177e4SLinus Torvalds * an explicit close() or as a byproduct of exit()'ing) and there 23271da177e4SLinus Torvalds * was unread data in the receive queue. This behavior is recommended 232865bb723cSGerrit Renker * by RFC 2525, section 2.17. -DaveM 23291da177e4SLinus Torvalds */ 2330dd0fc66fSAl Viro void tcp_send_active_reset(struct sock *sk, gfp_t priority) 23311da177e4SLinus Torvalds { 23321da177e4SLinus Torvalds struct sk_buff *skb; 23331da177e4SLinus Torvalds 23341da177e4SLinus Torvalds /* NOTE: No TCP options attached and we never retransmit this. */ 23351da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, priority); 23361da177e4SLinus Torvalds if (!skb) { 23374e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 23381da177e4SLinus Torvalds return; 23391da177e4SLinus Torvalds } 23401da177e4SLinus Torvalds 23411da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 23421da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 2343e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), 2344a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_RST); 23451da177e4SLinus Torvalds /* Send it off. */ 23461da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2347dfb4b9dcSDavid S. Miller if (tcp_transmit_skb(sk, skb, 0, priority)) 23484e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 234926af65cbSSridhar Samudrala 235081cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); 23511da177e4SLinus Torvalds } 23521da177e4SLinus Torvalds 235367edfef7SAndi Kleen /* Send a crossed SYN-ACK during socket establishment. 235467edfef7SAndi Kleen * WARNING: This routine must only be called when we have already sent 23551da177e4SLinus Torvalds * a SYN packet that crossed the incoming SYN that caused this routine 23561da177e4SLinus Torvalds * to get called. If this assumption fails then the initial rcv_wnd 23571da177e4SLinus Torvalds * and rcv_wscale values will not be correct. 23581da177e4SLinus Torvalds */ 23591da177e4SLinus Torvalds int tcp_send_synack(struct sock *sk) 23601da177e4SLinus Torvalds { 23611da177e4SLinus Torvalds struct sk_buff *skb; 23621da177e4SLinus Torvalds 2363fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 2364a3433f35SChangli Gao if (skb == NULL || !(TCP_SKB_CB(skb)->flags & TCPHDR_SYN)) { 23651da177e4SLinus Torvalds printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n"); 23661da177e4SLinus Torvalds return -EFAULT; 23671da177e4SLinus Torvalds } 2368a3433f35SChangli Gao if (!(TCP_SKB_CB(skb)->flags & TCPHDR_ACK)) { 23691da177e4SLinus Torvalds if (skb_cloned(skb)) { 23701da177e4SLinus Torvalds struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC); 23711da177e4SLinus Torvalds if (nskb == NULL) 23721da177e4SLinus Torvalds return -ENOMEM; 2373fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 23741da177e4SLinus Torvalds skb_header_release(nskb); 2375fe067e8aSDavid S. Miller __tcp_add_write_queue_head(sk, nskb); 23763ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 23773ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 23783ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 23791da177e4SLinus Torvalds skb = nskb; 23801da177e4SLinus Torvalds } 23811da177e4SLinus Torvalds 2382a3433f35SChangli Gao TCP_SKB_CB(skb)->flags |= TCPHDR_ACK; 23831da177e4SLinus Torvalds TCP_ECN_send_synack(tcp_sk(sk), skb); 23841da177e4SLinus Torvalds } 23851da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2386dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 23871da177e4SLinus Torvalds } 23881da177e4SLinus Torvalds 238967edfef7SAndi Kleen /* Prepare a SYN-ACK. */ 23901da177e4SLinus Torvalds struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst, 2391e6b4d113SWilliam Allen Simpson struct request_sock *req, 2392e6b4d113SWilliam Allen Simpson struct request_values *rvp) 23931da177e4SLinus Torvalds { 2394bd0388aeSWilliam Allen Simpson struct tcp_out_options opts; 23954957faadSWilliam Allen Simpson struct tcp_extend_values *xvp = tcp_xv(rvp); 23962e6599cbSArnaldo Carvalho de Melo struct inet_request_sock *ireq = inet_rsk(req); 23971da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 239828b2774aSEric Dumazet const struct tcp_cookie_values *cvp = tp->cookie_values; 23991da177e4SLinus Torvalds struct tcphdr *th; 24001da177e4SLinus Torvalds struct sk_buff *skb; 2401cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 2402bd0388aeSWilliam Allen Simpson int tcp_header_size; 2403f5fff5dcSTom Quetchenbach int mss; 240428b2774aSEric Dumazet int s_data_desired = 0; 24051da177e4SLinus Torvalds 240628b2774aSEric Dumazet if (cvp != NULL && cvp->s_data_constant && cvp->s_data_desired) 240728b2774aSEric Dumazet s_data_desired = cvp->s_data_desired; 240828b2774aSEric Dumazet skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15 + s_data_desired, 1, GFP_ATOMIC); 24091da177e4SLinus Torvalds if (skb == NULL) 24101da177e4SLinus Torvalds return NULL; 24111da177e4SLinus Torvalds 24121da177e4SLinus Torvalds /* Reserve space for headers. */ 24131da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 24141da177e4SLinus Torvalds 2415adf30907SEric Dumazet skb_dst_set(skb, dst_clone(dst)); 24161da177e4SLinus Torvalds 2417f5fff5dcSTom Quetchenbach mss = dst_metric(dst, RTAX_ADVMSS); 2418f5fff5dcSTom Quetchenbach if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss) 2419f5fff5dcSTom Quetchenbach mss = tp->rx_opt.user_mss; 2420f5fff5dcSTom Quetchenbach 242133ad798cSAdam Langley if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */ 242233ad798cSAdam Langley __u8 rcv_wscale; 242333ad798cSAdam Langley /* Set this up on the first call only */ 242433ad798cSAdam Langley req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW); 2425e88c64f0SHagen Paul Pfeifer 2426e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 2427e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 2428e88c64f0SHagen Paul Pfeifer (req->window_clamp > tcp_full_space(sk) || req->window_clamp == 0)) 2429e88c64f0SHagen Paul Pfeifer req->window_clamp = tcp_full_space(sk); 2430e88c64f0SHagen Paul Pfeifer 243133ad798cSAdam Langley /* tcp_full_space because it is guaranteed to be the first packet */ 243233ad798cSAdam Langley tcp_select_initial_window(tcp_full_space(sk), 2433f5fff5dcSTom Quetchenbach mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0), 243433ad798cSAdam Langley &req->rcv_wnd, 243533ad798cSAdam Langley &req->window_clamp, 243633ad798cSAdam Langley ireq->wscale_ok, 243731d12926Slaurent chavey &rcv_wscale, 243831d12926Slaurent chavey dst_metric(dst, RTAX_INITRWND)); 243933ad798cSAdam Langley ireq->rcv_wscale = rcv_wscale; 244033ad798cSAdam Langley } 2441cfb6eeb4SYOSHIFUJI Hideaki 244233ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 24438b5f12d0SFlorian Westphal #ifdef CONFIG_SYN_COOKIES 24448b5f12d0SFlorian Westphal if (unlikely(req->cookie_ts)) 24458b5f12d0SFlorian Westphal TCP_SKB_CB(skb)->when = cookie_init_timestamp(req); 24468b5f12d0SFlorian Westphal else 24478b5f12d0SFlorian Westphal #endif 244833ad798cSAdam Langley TCP_SKB_CB(skb)->when = tcp_time_stamp; 2449f5fff5dcSTom Quetchenbach tcp_header_size = tcp_synack_options(sk, req, mss, 24504957faadSWilliam Allen Simpson skb, &opts, &md5, xvp) 24514957faadSWilliam Allen Simpson + sizeof(*th); 245233ad798cSAdam Langley 2453aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 2454aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 24551da177e4SLinus Torvalds 2456aa8223c7SArnaldo Carvalho de Melo th = tcp_hdr(skb); 24571da177e4SLinus Torvalds memset(th, 0, sizeof(struct tcphdr)); 24581da177e4SLinus Torvalds th->syn = 1; 24591da177e4SLinus Torvalds th->ack = 1; 24601da177e4SLinus Torvalds TCP_ECN_make_synack(req, th); 2461a3116ac5SKOVACS Krisztian th->source = ireq->loc_port; 24622e6599cbSArnaldo Carvalho de Melo th->dest = ireq->rmt_port; 2463e870a8efSIlpo Järvinen /* Setting of flags are superfluous here for callers (and ECE is 2464e870a8efSIlpo Järvinen * not even correctly set) 2465e870a8efSIlpo Järvinen */ 2466e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn, 2467a3433f35SChangli Gao TCPHDR_SYN | TCPHDR_ACK); 24684957faadSWilliam Allen Simpson 24694957faadSWilliam Allen Simpson if (OPTION_COOKIE_EXTENSION & opts.options) { 247028b2774aSEric Dumazet if (s_data_desired) { 247128b2774aSEric Dumazet u8 *buf = skb_put(skb, s_data_desired); 24724957faadSWilliam Allen Simpson 24734957faadSWilliam Allen Simpson /* copy data directly from the listening socket. */ 247428b2774aSEric Dumazet memcpy(buf, cvp->s_data_payload, s_data_desired); 247528b2774aSEric Dumazet TCP_SKB_CB(skb)->end_seq += s_data_desired; 24764957faadSWilliam Allen Simpson } 24774957faadSWilliam Allen Simpson 24784957faadSWilliam Allen Simpson if (opts.hash_size > 0) { 24794957faadSWilliam Allen Simpson __u32 workspace[SHA_WORKSPACE_WORDS]; 24804957faadSWilliam Allen Simpson u32 *mess = &xvp->cookie_bakery[COOKIE_DIGEST_WORDS]; 24814957faadSWilliam Allen Simpson u32 *tail = &mess[COOKIE_MESSAGE_WORDS-1]; 24824957faadSWilliam Allen Simpson 24834957faadSWilliam Allen Simpson /* Secret recipe depends on the Timestamp, (future) 24844957faadSWilliam Allen Simpson * Sequence and Acknowledgment Numbers, Initiator 24854957faadSWilliam Allen Simpson * Cookie, and others handled by IP variant caller. 24864957faadSWilliam Allen Simpson */ 24874957faadSWilliam Allen Simpson *tail-- ^= opts.tsval; 24884957faadSWilliam Allen Simpson *tail-- ^= tcp_rsk(req)->rcv_isn + 1; 24894957faadSWilliam Allen Simpson *tail-- ^= TCP_SKB_CB(skb)->seq + 1; 24904957faadSWilliam Allen Simpson 24914957faadSWilliam Allen Simpson /* recommended */ 24920eae88f3SEric Dumazet *tail-- ^= (((__force u32)th->dest << 16) | (__force u32)th->source); 2493f9a2e69eSDavid S. Miller *tail-- ^= (u32)(unsigned long)cvp; /* per sockopt */ 24944957faadSWilliam Allen Simpson 24954957faadSWilliam Allen Simpson sha_transform((__u32 *)&xvp->cookie_bakery[0], 24964957faadSWilliam Allen Simpson (char *)mess, 24974957faadSWilliam Allen Simpson &workspace[0]); 24984957faadSWilliam Allen Simpson opts.hash_location = 24994957faadSWilliam Allen Simpson (__u8 *)&xvp->cookie_bakery[0]; 25004957faadSWilliam Allen Simpson } 25014957faadSWilliam Allen Simpson } 25024957faadSWilliam Allen Simpson 25031da177e4SLinus Torvalds th->seq = htonl(TCP_SKB_CB(skb)->seq); 25042e6599cbSArnaldo Carvalho de Melo th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1); 25051da177e4SLinus Torvalds 25061da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 2507600ff0c2SIlpo Järvinen th->window = htons(min(req->rcv_wnd, 65535U)); 2508bd0388aeSWilliam Allen Simpson tcp_options_write((__be32 *)(th + 1), tp, &opts); 25091da177e4SLinus Torvalds th->doff = (tcp_header_size >> 2); 2510aa2ea058STom Herbert TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, tcp_skb_pcount(skb)); 2511cfb6eeb4SYOSHIFUJI Hideaki 2512cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 2513cfb6eeb4SYOSHIFUJI Hideaki /* Okay, we have all we need - do the md5 hash if needed */ 2514cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 2515bd0388aeSWilliam Allen Simpson tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, 251649a72dfbSAdam Langley md5, NULL, req, skb); 2517cfb6eeb4SYOSHIFUJI Hideaki } 2518cfb6eeb4SYOSHIFUJI Hideaki #endif 2519cfb6eeb4SYOSHIFUJI Hideaki 25201da177e4SLinus Torvalds return skb; 25211da177e4SLinus Torvalds } 25224bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_make_synack); 25231da177e4SLinus Torvalds 252467edfef7SAndi Kleen /* Do all connect socket setups that can be done AF independent. */ 252540efc6faSStephen Hemminger static void tcp_connect_init(struct sock *sk) 25261da177e4SLinus Torvalds { 25271da177e4SLinus Torvalds struct dst_entry *dst = __sk_dst_get(sk); 25281da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 25291da177e4SLinus Torvalds __u8 rcv_wscale; 25301da177e4SLinus Torvalds 25311da177e4SLinus Torvalds /* We'll fix this up when we get a response from the other end. 25321da177e4SLinus Torvalds * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 25331da177e4SLinus Torvalds */ 25341da177e4SLinus Torvalds tp->tcp_header_len = sizeof(struct tcphdr) + 2535bb5b7c11SDavid S. Miller (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0); 25361da177e4SLinus Torvalds 2537cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 2538cfb6eeb4SYOSHIFUJI Hideaki if (tp->af_specific->md5_lookup(sk, sk) != NULL) 2539cfb6eeb4SYOSHIFUJI Hideaki tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 2540cfb6eeb4SYOSHIFUJI Hideaki #endif 2541cfb6eeb4SYOSHIFUJI Hideaki 25421da177e4SLinus Torvalds /* If user gave his TCP_MAXSEG, record it to clamp */ 25431da177e4SLinus Torvalds if (tp->rx_opt.user_mss) 25441da177e4SLinus Torvalds tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 25451da177e4SLinus Torvalds tp->max_window = 0; 25465d424d5aSJohn Heffner tcp_mtup_init(sk); 25471da177e4SLinus Torvalds tcp_sync_mss(sk, dst_mtu(dst)); 25481da177e4SLinus Torvalds 25491da177e4SLinus Torvalds if (!tp->window_clamp) 25501da177e4SLinus Torvalds tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 25511da177e4SLinus Torvalds tp->advmss = dst_metric(dst, RTAX_ADVMSS); 2552f5fff5dcSTom Quetchenbach if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss) 2553f5fff5dcSTom Quetchenbach tp->advmss = tp->rx_opt.user_mss; 2554f5fff5dcSTom Quetchenbach 25551da177e4SLinus Torvalds tcp_initialize_rcv_mss(sk); 25561da177e4SLinus Torvalds 2557e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 2558e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 2559e88c64f0SHagen Paul Pfeifer (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) 2560e88c64f0SHagen Paul Pfeifer tp->window_clamp = tcp_full_space(sk); 2561e88c64f0SHagen Paul Pfeifer 25621da177e4SLinus Torvalds tcp_select_initial_window(tcp_full_space(sk), 25631da177e4SLinus Torvalds tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 25641da177e4SLinus Torvalds &tp->rcv_wnd, 25651da177e4SLinus Torvalds &tp->window_clamp, 2566bb5b7c11SDavid S. Miller sysctl_tcp_window_scaling, 256731d12926Slaurent chavey &rcv_wscale, 256831d12926Slaurent chavey dst_metric(dst, RTAX_INITRWND)); 25691da177e4SLinus Torvalds 25701da177e4SLinus Torvalds tp->rx_opt.rcv_wscale = rcv_wscale; 25711da177e4SLinus Torvalds tp->rcv_ssthresh = tp->rcv_wnd; 25721da177e4SLinus Torvalds 25731da177e4SLinus Torvalds sk->sk_err = 0; 25741da177e4SLinus Torvalds sock_reset_flag(sk, SOCK_DONE); 25751da177e4SLinus Torvalds tp->snd_wnd = 0; 2576ee7537b6SHantzis Fotis tcp_init_wl(tp, 0); 25771da177e4SLinus Torvalds tp->snd_una = tp->write_seq; 25781da177e4SLinus Torvalds tp->snd_sml = tp->write_seq; 257933f5f57eSIlpo Järvinen tp->snd_up = tp->write_seq; 25801da177e4SLinus Torvalds tp->rcv_nxt = 0; 25811da177e4SLinus Torvalds tp->rcv_wup = 0; 25821da177e4SLinus Torvalds tp->copied_seq = 0; 25831da177e4SLinus Torvalds 2584463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; 2585463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_retransmits = 0; 25861da177e4SLinus Torvalds tcp_clear_retrans(tp); 25871da177e4SLinus Torvalds } 25881da177e4SLinus Torvalds 258967edfef7SAndi Kleen /* Build a SYN and send it off. */ 25901da177e4SLinus Torvalds int tcp_connect(struct sock *sk) 25911da177e4SLinus Torvalds { 25921da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 25931da177e4SLinus Torvalds struct sk_buff *buff; 25941da177e4SLinus Torvalds 25951da177e4SLinus Torvalds tcp_connect_init(sk); 25961da177e4SLinus Torvalds 2597d179cd12SDavid S. Miller buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation); 25981da177e4SLinus Torvalds if (unlikely(buff == NULL)) 25991da177e4SLinus Torvalds return -ENOBUFS; 26001da177e4SLinus Torvalds 26011da177e4SLinus Torvalds /* Reserve space for headers. */ 26021da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 26031da177e4SLinus Torvalds 2604bd37a088SWei Yongjun tp->snd_nxt = tp->write_seq; 2605a3433f35SChangli Gao tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); 2606e870a8efSIlpo Järvinen TCP_ECN_send_syn(sk, buff); 26071da177e4SLinus Torvalds 26081da177e4SLinus Torvalds /* Send it off. */ 26091da177e4SLinus Torvalds TCP_SKB_CB(buff)->when = tcp_time_stamp; 26101da177e4SLinus Torvalds tp->retrans_stamp = TCP_SKB_CB(buff)->when; 26111da177e4SLinus Torvalds skb_header_release(buff); 2612fe067e8aSDavid S. Miller __tcp_add_write_queue_tail(sk, buff); 26133ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 26143ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 26151da177e4SLinus Torvalds tp->packets_out += tcp_skb_pcount(buff); 2616aa133076SWu Fengguang tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); 2617bd37a088SWei Yongjun 2618bd37a088SWei Yongjun /* We change tp->snd_nxt after the tcp_transmit_skb() call 2619bd37a088SWei Yongjun * in order to make this packet get counted in tcpOutSegs. 2620bd37a088SWei Yongjun */ 2621bd37a088SWei Yongjun tp->snd_nxt = tp->write_seq; 2622bd37a088SWei Yongjun tp->pushed_seq = tp->write_seq; 262381cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); 26241da177e4SLinus Torvalds 26251da177e4SLinus Torvalds /* Timer for repeating the SYN until an answer. */ 26263f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 26273f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 26281da177e4SLinus Torvalds return 0; 26291da177e4SLinus Torvalds } 26304bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_connect); 26311da177e4SLinus Torvalds 26321da177e4SLinus Torvalds /* Send out a delayed ack, the caller does the policy checking 26331da177e4SLinus Torvalds * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 26341da177e4SLinus Torvalds * for details. 26351da177e4SLinus Torvalds */ 26361da177e4SLinus Torvalds void tcp_send_delayed_ack(struct sock *sk) 26371da177e4SLinus Torvalds { 2638463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 2639463c84b9SArnaldo Carvalho de Melo int ato = icsk->icsk_ack.ato; 26401da177e4SLinus Torvalds unsigned long timeout; 26411da177e4SLinus Torvalds 26421da177e4SLinus Torvalds if (ato > TCP_DELACK_MIN) { 2643463c84b9SArnaldo Carvalho de Melo const struct tcp_sock *tp = tcp_sk(sk); 26441da177e4SLinus Torvalds int max_ato = HZ / 2; 26451da177e4SLinus Torvalds 2646056834d9SIlpo Järvinen if (icsk->icsk_ack.pingpong || 2647056834d9SIlpo Järvinen (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 26481da177e4SLinus Torvalds max_ato = TCP_DELACK_MAX; 26491da177e4SLinus Torvalds 26501da177e4SLinus Torvalds /* Slow path, intersegment interval is "high". */ 26511da177e4SLinus Torvalds 26521da177e4SLinus Torvalds /* If some rtt estimate is known, use it to bound delayed ack. 2653463c84b9SArnaldo Carvalho de Melo * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 26541da177e4SLinus Torvalds * directly. 26551da177e4SLinus Torvalds */ 26561da177e4SLinus Torvalds if (tp->srtt) { 26571da177e4SLinus Torvalds int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN); 26581da177e4SLinus Torvalds 26591da177e4SLinus Torvalds if (rtt < max_ato) 26601da177e4SLinus Torvalds max_ato = rtt; 26611da177e4SLinus Torvalds } 26621da177e4SLinus Torvalds 26631da177e4SLinus Torvalds ato = min(ato, max_ato); 26641da177e4SLinus Torvalds } 26651da177e4SLinus Torvalds 26661da177e4SLinus Torvalds /* Stay within the limit we were given */ 26671da177e4SLinus Torvalds timeout = jiffies + ato; 26681da177e4SLinus Torvalds 26691da177e4SLinus Torvalds /* Use new timeout only if there wasn't a older one earlier. */ 2670463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 26711da177e4SLinus Torvalds /* If delack timer was blocked or is about to expire, 26721da177e4SLinus Torvalds * send ACK now. 26731da177e4SLinus Torvalds */ 2674463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.blocked || 2675463c84b9SArnaldo Carvalho de Melo time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 26761da177e4SLinus Torvalds tcp_send_ack(sk); 26771da177e4SLinus Torvalds return; 26781da177e4SLinus Torvalds } 26791da177e4SLinus Torvalds 2680463c84b9SArnaldo Carvalho de Melo if (!time_before(timeout, icsk->icsk_ack.timeout)) 2681463c84b9SArnaldo Carvalho de Melo timeout = icsk->icsk_ack.timeout; 26821da177e4SLinus Torvalds } 2683463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 2684463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.timeout = timeout; 2685463c84b9SArnaldo Carvalho de Melo sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 26861da177e4SLinus Torvalds } 26871da177e4SLinus Torvalds 26881da177e4SLinus Torvalds /* This routine sends an ack and also updates the window. */ 26891da177e4SLinus Torvalds void tcp_send_ack(struct sock *sk) 26901da177e4SLinus Torvalds { 26911da177e4SLinus Torvalds struct sk_buff *buff; 26921da177e4SLinus Torvalds 2693058dc334SIlpo Järvinen /* If we have been reset, we may not send again. */ 2694058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 2695058dc334SIlpo Järvinen return; 2696058dc334SIlpo Järvinen 26971da177e4SLinus Torvalds /* We are not putting this on the write queue, so 26981da177e4SLinus Torvalds * tcp_transmit_skb() will set the ownership to this 26991da177e4SLinus Torvalds * sock. 27001da177e4SLinus Torvalds */ 27011da177e4SLinus Torvalds buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 27021da177e4SLinus Torvalds if (buff == NULL) { 2703463c84b9SArnaldo Carvalho de Melo inet_csk_schedule_ack(sk); 2704463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; 27053f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 27063f421baaSArnaldo Carvalho de Melo TCP_DELACK_MAX, TCP_RTO_MAX); 27071da177e4SLinus Torvalds return; 27081da177e4SLinus Torvalds } 27091da177e4SLinus Torvalds 27101da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 27111da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 2712a3433f35SChangli Gao tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); 27131da177e4SLinus Torvalds 27141da177e4SLinus Torvalds /* Send it off, this clears delayed acks for us. */ 27151da177e4SLinus Torvalds TCP_SKB_CB(buff)->when = tcp_time_stamp; 2716dfb4b9dcSDavid S. Miller tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC); 27171da177e4SLinus Torvalds } 27181da177e4SLinus Torvalds 27191da177e4SLinus Torvalds /* This routine sends a packet with an out of date sequence 27201da177e4SLinus Torvalds * number. It assumes the other end will try to ack it. 27211da177e4SLinus Torvalds * 27221da177e4SLinus Torvalds * Question: what should we make while urgent mode? 27231da177e4SLinus Torvalds * 4.4BSD forces sending single byte of data. We cannot send 27241da177e4SLinus Torvalds * out of window data, because we have SND.NXT==SND.MAX... 27251da177e4SLinus Torvalds * 27261da177e4SLinus Torvalds * Current solution: to send TWO zero-length segments in urgent mode: 27271da177e4SLinus Torvalds * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 27281da177e4SLinus Torvalds * out-of-date with SND.UNA-1 to probe window. 27291da177e4SLinus Torvalds */ 27301da177e4SLinus Torvalds static int tcp_xmit_probe_skb(struct sock *sk, int urgent) 27311da177e4SLinus Torvalds { 27321da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 27331da177e4SLinus Torvalds struct sk_buff *skb; 27341da177e4SLinus Torvalds 27351da177e4SLinus Torvalds /* We don't queue it, tcp_transmit_skb() sets ownership. */ 27361da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 27371da177e4SLinus Torvalds if (skb == NULL) 27381da177e4SLinus Torvalds return -1; 27391da177e4SLinus Torvalds 27401da177e4SLinus Torvalds /* Reserve space for headers and set control bits. */ 27411da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 27421da177e4SLinus Torvalds /* Use a previous sequence. This should cause the other 27431da177e4SLinus Torvalds * end to send an ack. Don't queue or clone SKB, just 27441da177e4SLinus Torvalds * send it. 27451da177e4SLinus Torvalds */ 2746a3433f35SChangli Gao tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); 27471da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2748dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC); 27491da177e4SLinus Torvalds } 27501da177e4SLinus Torvalds 275167edfef7SAndi Kleen /* Initiate keepalive or window probe from timer. */ 27521da177e4SLinus Torvalds int tcp_write_wakeup(struct sock *sk) 27531da177e4SLinus Torvalds { 27541da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 27551da177e4SLinus Torvalds struct sk_buff *skb; 27561da177e4SLinus Torvalds 2757058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 2758058dc334SIlpo Järvinen return -1; 2759058dc334SIlpo Järvinen 2760fe067e8aSDavid S. Miller if ((skb = tcp_send_head(sk)) != NULL && 276190840defSIlpo Järvinen before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { 27621da177e4SLinus Torvalds int err; 27630c54b85fSIlpo Järvinen unsigned int mss = tcp_current_mss(sk); 276490840defSIlpo Järvinen unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 27651da177e4SLinus Torvalds 27661da177e4SLinus Torvalds if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 27671da177e4SLinus Torvalds tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 27681da177e4SLinus Torvalds 27691da177e4SLinus Torvalds /* We are probing the opening of a window 27701da177e4SLinus Torvalds * but the window size is != 0 27711da177e4SLinus Torvalds * must have been a result SWS avoidance ( sender ) 27721da177e4SLinus Torvalds */ 27731da177e4SLinus Torvalds if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 27741da177e4SLinus Torvalds skb->len > mss) { 27751da177e4SLinus Torvalds seg_size = min(seg_size, mss); 2776a3433f35SChangli Gao TCP_SKB_CB(skb)->flags |= TCPHDR_PSH; 2777846998aeSDavid S. Miller if (tcp_fragment(sk, skb, seg_size, mss)) 27781da177e4SLinus Torvalds return -1; 27791da177e4SLinus Torvalds } else if (!tcp_skb_pcount(skb)) 2780846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss); 27811da177e4SLinus Torvalds 2782a3433f35SChangli Gao TCP_SKB_CB(skb)->flags |= TCPHDR_PSH; 27831da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2784dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 278566f5fe62SIlpo Järvinen if (!err) 278666f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 27871da177e4SLinus Torvalds return err; 27881da177e4SLinus Torvalds } else { 278933f5f57eSIlpo Järvinen if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) 27904828e7f4SIlpo Järvinen tcp_xmit_probe_skb(sk, 1); 27911da177e4SLinus Torvalds return tcp_xmit_probe_skb(sk, 0); 27921da177e4SLinus Torvalds } 27931da177e4SLinus Torvalds } 27941da177e4SLinus Torvalds 27951da177e4SLinus Torvalds /* A window probe timeout has occurred. If window is not closed send 27961da177e4SLinus Torvalds * a partial packet else a zero probe. 27971da177e4SLinus Torvalds */ 27981da177e4SLinus Torvalds void tcp_send_probe0(struct sock *sk) 27991da177e4SLinus Torvalds { 2800463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 28011da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 28021da177e4SLinus Torvalds int err; 28031da177e4SLinus Torvalds 28041da177e4SLinus Torvalds err = tcp_write_wakeup(sk); 28051da177e4SLinus Torvalds 2806fe067e8aSDavid S. Miller if (tp->packets_out || !tcp_send_head(sk)) { 28071da177e4SLinus Torvalds /* Cancel probe timer, if it is not required. */ 28086687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 0; 2809463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff = 0; 28101da177e4SLinus Torvalds return; 28111da177e4SLinus Torvalds } 28121da177e4SLinus Torvalds 28131da177e4SLinus Torvalds if (err <= 0) { 2814463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_backoff < sysctl_tcp_retries2) 2815463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff++; 28166687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out++; 2817463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 28183f421baaSArnaldo Carvalho de Melo min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), 28193f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 28201da177e4SLinus Torvalds } else { 28211da177e4SLinus Torvalds /* If packet was not sent due to local congestion, 28226687e988SArnaldo Carvalho de Melo * do not backoff and do not remember icsk_probes_out. 28231da177e4SLinus Torvalds * Let local senders to fight for local resources. 28241da177e4SLinus Torvalds * 28251da177e4SLinus Torvalds * Use accumulated backoff yet. 28261da177e4SLinus Torvalds */ 28276687e988SArnaldo Carvalho de Melo if (!icsk->icsk_probes_out) 28286687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 1; 2829463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 2830463c84b9SArnaldo Carvalho de Melo min(icsk->icsk_rto << icsk->icsk_backoff, 28313f421baaSArnaldo Carvalho de Melo TCP_RESOURCE_PROBE_INTERVAL), 28323f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 28331da177e4SLinus Torvalds } 28341da177e4SLinus Torvalds } 2835