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; 5897b1ce25SShan Wei int sysctl_tcp_base_mss __read_mostly = TCP_BASE_MSS; 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 1220dbaee3bSDavid S. Miller if (dst) { 1230dbaee3bSDavid S. Miller unsigned int metric = dst_metric_advmss(dst); 1240dbaee3bSDavid S. Miller 1250dbaee3bSDavid S. Miller if (metric < mss) { 1260dbaee3bSDavid S. Miller mss = metric; 1271da177e4SLinus Torvalds tp->advmss = mss; 1281da177e4SLinus Torvalds } 1290dbaee3bSDavid S. Miller } 1301da177e4SLinus Torvalds 1311da177e4SLinus Torvalds return (__u16)mss; 1321da177e4SLinus Torvalds } 1331da177e4SLinus Torvalds 1341da177e4SLinus Torvalds /* RFC2861. Reset CWND after idle period longer RTO to "restart window". 1351da177e4SLinus Torvalds * This is the first part of cwnd validation mechanism. */ 136463c84b9SArnaldo Carvalho de Melo static void tcp_cwnd_restart(struct sock *sk, struct dst_entry *dst) 1371da177e4SLinus Torvalds { 138463c84b9SArnaldo Carvalho de Melo struct tcp_sock *tp = tcp_sk(sk); 1391da177e4SLinus Torvalds s32 delta = tcp_time_stamp - tp->lsndtime; 1401da177e4SLinus Torvalds u32 restart_cwnd = tcp_init_cwnd(tp, dst); 1411da177e4SLinus Torvalds u32 cwnd = tp->snd_cwnd; 1421da177e4SLinus Torvalds 1436687e988SArnaldo Carvalho de Melo tcp_ca_event(sk, CA_EVENT_CWND_RESTART); 1441da177e4SLinus Torvalds 1456687e988SArnaldo Carvalho de Melo tp->snd_ssthresh = tcp_current_ssthresh(sk); 1461da177e4SLinus Torvalds restart_cwnd = min(restart_cwnd, cwnd); 1471da177e4SLinus Torvalds 148463c84b9SArnaldo Carvalho de Melo while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd) 1491da177e4SLinus Torvalds cwnd >>= 1; 1501da177e4SLinus Torvalds tp->snd_cwnd = max(cwnd, restart_cwnd); 1511da177e4SLinus Torvalds tp->snd_cwnd_stamp = tcp_time_stamp; 1521da177e4SLinus Torvalds tp->snd_cwnd_used = 0; 1531da177e4SLinus Torvalds } 1541da177e4SLinus Torvalds 15567edfef7SAndi Kleen /* Congestion state accounting after a packet has been sent. */ 15640efc6faSStephen Hemminger static void tcp_event_data_sent(struct tcp_sock *tp, 1571da177e4SLinus Torvalds struct sk_buff *skb, struct sock *sk) 1581da177e4SLinus Torvalds { 159463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 160463c84b9SArnaldo Carvalho de Melo const u32 now = tcp_time_stamp; 1611da177e4SLinus Torvalds 16235089bb2SDavid S. Miller if (sysctl_tcp_slow_start_after_idle && 16335089bb2SDavid S. Miller (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto)) 164463c84b9SArnaldo Carvalho de Melo tcp_cwnd_restart(sk, __sk_dst_get(sk)); 1651da177e4SLinus Torvalds 1661da177e4SLinus Torvalds tp->lsndtime = now; 1671da177e4SLinus Torvalds 1681da177e4SLinus Torvalds /* If it is a reply for ato after last received 1691da177e4SLinus Torvalds * packet, enter pingpong mode. 1701da177e4SLinus Torvalds */ 171463c84b9SArnaldo Carvalho de Melo if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato) 172463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pingpong = 1; 1731da177e4SLinus Torvalds } 1741da177e4SLinus Torvalds 17567edfef7SAndi Kleen /* Account for an ACK we sent. */ 17640efc6faSStephen Hemminger static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts) 1771da177e4SLinus Torvalds { 178463c84b9SArnaldo Carvalho de Melo tcp_dec_quickack_mode(sk, pkts); 179463c84b9SArnaldo Carvalho de Melo inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); 1801da177e4SLinus Torvalds } 1811da177e4SLinus Torvalds 1821da177e4SLinus Torvalds /* Determine a window scaling and initial window to offer. 1831da177e4SLinus Torvalds * Based on the assumption that the given amount of space 1841da177e4SLinus Torvalds * will be offered. Store the results in the tp structure. 1851da177e4SLinus Torvalds * NOTE: for smooth operation initial space offering should 1861da177e4SLinus Torvalds * be a multiple of mss if possible. We assume here that mss >= 1. 1871da177e4SLinus Torvalds * This MUST be enforced by all callers. 1881da177e4SLinus Torvalds */ 1891da177e4SLinus Torvalds void tcp_select_initial_window(int __space, __u32 mss, 1901da177e4SLinus Torvalds __u32 *rcv_wnd, __u32 *window_clamp, 19131d12926Slaurent chavey int wscale_ok, __u8 *rcv_wscale, 19231d12926Slaurent chavey __u32 init_rcv_wnd) 1931da177e4SLinus Torvalds { 1941da177e4SLinus Torvalds unsigned int space = (__space < 0 ? 0 : __space); 1951da177e4SLinus Torvalds 1961da177e4SLinus Torvalds /* If no clamp set the clamp to the max possible scaled window */ 1971da177e4SLinus Torvalds if (*window_clamp == 0) 1981da177e4SLinus Torvalds (*window_clamp) = (65535 << 14); 1991da177e4SLinus Torvalds space = min(*window_clamp, space); 2001da177e4SLinus Torvalds 2011da177e4SLinus Torvalds /* Quantize space offering to a multiple of mss if possible. */ 2021da177e4SLinus Torvalds if (space > mss) 2031da177e4SLinus Torvalds space = (space / mss) * mss; 2041da177e4SLinus Torvalds 2051da177e4SLinus Torvalds /* NOTE: offering an initial window larger than 32767 20615d99e02SRick Jones * will break some buggy TCP stacks. If the admin tells us 20715d99e02SRick Jones * it is likely we could be speaking with such a buggy stack 20815d99e02SRick Jones * we will truncate our initial window offering to 32K-1 20915d99e02SRick Jones * unless the remote has sent us a window scaling option, 21015d99e02SRick Jones * which we interpret as a sign the remote TCP is not 21115d99e02SRick Jones * misinterpreting the window field as a signed quantity. 2121da177e4SLinus Torvalds */ 21315d99e02SRick Jones if (sysctl_tcp_workaround_signed_windows) 2141da177e4SLinus Torvalds (*rcv_wnd) = min(space, MAX_TCP_WINDOW); 21515d99e02SRick Jones else 21615d99e02SRick Jones (*rcv_wnd) = space; 21715d99e02SRick Jones 2181da177e4SLinus Torvalds (*rcv_wscale) = 0; 2191da177e4SLinus Torvalds if (wscale_ok) { 2201da177e4SLinus Torvalds /* Set window scaling on max possible window 2211da177e4SLinus Torvalds * See RFC1323 for an explanation of the limit to 14 2221da177e4SLinus Torvalds */ 2231da177e4SLinus Torvalds space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max); 224316c1592SStephen Hemminger space = min_t(u32, space, *window_clamp); 2251da177e4SLinus Torvalds while (space > 65535 && (*rcv_wscale) < 14) { 2261da177e4SLinus Torvalds space >>= 1; 2271da177e4SLinus Torvalds (*rcv_wscale)++; 2281da177e4SLinus Torvalds } 2291da177e4SLinus Torvalds } 2301da177e4SLinus Torvalds 231356f0398SNandita Dukkipati /* Set initial window to a value enough for senders starting with 232356f0398SNandita Dukkipati * initial congestion window of TCP_DEFAULT_INIT_RCVWND. Place 233356f0398SNandita Dukkipati * a limit on the initial window when mss is larger than 1460. 234356f0398SNandita Dukkipati */ 2351da177e4SLinus Torvalds if (mss > (1 << *rcv_wscale)) { 236356f0398SNandita Dukkipati int init_cwnd = TCP_DEFAULT_INIT_RCVWND; 237356f0398SNandita Dukkipati if (mss > 1460) 238356f0398SNandita Dukkipati init_cwnd = 239356f0398SNandita Dukkipati max_t(u32, (1460 * TCP_DEFAULT_INIT_RCVWND) / mss, 2); 24031d12926Slaurent chavey /* when initializing use the value from init_rcv_wnd 24131d12926Slaurent chavey * rather than the default from above 24231d12926Slaurent chavey */ 243b1afde60SNandita Dukkipati if (init_rcv_wnd) 244b1afde60SNandita Dukkipati *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss); 245b1afde60SNandita Dukkipati else 246b1afde60SNandita Dukkipati *rcv_wnd = min(*rcv_wnd, init_cwnd * mss); 2471da177e4SLinus Torvalds } 2481da177e4SLinus Torvalds 2491da177e4SLinus Torvalds /* Set the clamp no higher than max representable value */ 2501da177e4SLinus Torvalds (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp); 2511da177e4SLinus Torvalds } 2524bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_select_initial_window); 2531da177e4SLinus Torvalds 2541da177e4SLinus Torvalds /* Chose a new window to advertise, update state in tcp_sock for the 2551da177e4SLinus Torvalds * socket, and return result with RFC1323 scaling applied. The return 2561da177e4SLinus Torvalds * value can be stuffed directly into th->window for an outgoing 2571da177e4SLinus Torvalds * frame. 2581da177e4SLinus Torvalds */ 25940efc6faSStephen Hemminger static u16 tcp_select_window(struct sock *sk) 2601da177e4SLinus Torvalds { 2611da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2621da177e4SLinus Torvalds u32 cur_win = tcp_receive_window(tp); 2631da177e4SLinus Torvalds u32 new_win = __tcp_select_window(sk); 2641da177e4SLinus Torvalds 2651da177e4SLinus Torvalds /* Never shrink the offered window */ 2661da177e4SLinus Torvalds if (new_win < cur_win) { 2671da177e4SLinus Torvalds /* Danger Will Robinson! 2681da177e4SLinus Torvalds * Don't update rcv_wup/rcv_wnd here or else 2691da177e4SLinus Torvalds * we will not be able to advertise a zero 2701da177e4SLinus Torvalds * window in time. --DaveM 2711da177e4SLinus Torvalds * 2721da177e4SLinus Torvalds * Relax Will Robinson. 2731da177e4SLinus Torvalds */ 274607bfbf2SPatrick McHardy new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale); 2751da177e4SLinus Torvalds } 2761da177e4SLinus Torvalds tp->rcv_wnd = new_win; 2771da177e4SLinus Torvalds tp->rcv_wup = tp->rcv_nxt; 2781da177e4SLinus Torvalds 2791da177e4SLinus Torvalds /* Make sure we do not exceed the maximum possible 2801da177e4SLinus Torvalds * scaled window. 2811da177e4SLinus Torvalds */ 28215d99e02SRick Jones if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows) 2831da177e4SLinus Torvalds new_win = min(new_win, MAX_TCP_WINDOW); 2841da177e4SLinus Torvalds else 2851da177e4SLinus Torvalds new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); 2861da177e4SLinus Torvalds 2871da177e4SLinus Torvalds /* RFC1323 scaling applied */ 2881da177e4SLinus Torvalds new_win >>= tp->rx_opt.rcv_wscale; 2891da177e4SLinus Torvalds 2901da177e4SLinus Torvalds /* If we advertise zero window, disable fast path. */ 2911da177e4SLinus Torvalds if (new_win == 0) 2921da177e4SLinus Torvalds tp->pred_flags = 0; 2931da177e4SLinus Torvalds 2941da177e4SLinus Torvalds return new_win; 2951da177e4SLinus Torvalds } 2961da177e4SLinus Torvalds 29767edfef7SAndi Kleen /* Packet ECN state for a SYN-ACK */ 298056834d9SIlpo Järvinen static inline void TCP_ECN_send_synack(struct tcp_sock *tp, struct sk_buff *skb) 299bdf1ee5dSIlpo Järvinen { 300a3433f35SChangli Gao TCP_SKB_CB(skb)->flags &= ~TCPHDR_CWR; 301bdf1ee5dSIlpo Järvinen if (!(tp->ecn_flags & TCP_ECN_OK)) 302a3433f35SChangli Gao TCP_SKB_CB(skb)->flags &= ~TCPHDR_ECE; 303bdf1ee5dSIlpo Järvinen } 304bdf1ee5dSIlpo Järvinen 30567edfef7SAndi Kleen /* Packet ECN state for a SYN. */ 306bdf1ee5dSIlpo Järvinen static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb) 307bdf1ee5dSIlpo Järvinen { 308bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 309bdf1ee5dSIlpo Järvinen 310bdf1ee5dSIlpo Järvinen tp->ecn_flags = 0; 311255cac91SIlpo Järvinen if (sysctl_tcp_ecn == 1) { 312a3433f35SChangli Gao TCP_SKB_CB(skb)->flags |= TCPHDR_ECE | TCPHDR_CWR; 313bdf1ee5dSIlpo Järvinen tp->ecn_flags = TCP_ECN_OK; 314bdf1ee5dSIlpo Järvinen } 315bdf1ee5dSIlpo Järvinen } 316bdf1ee5dSIlpo Järvinen 317bdf1ee5dSIlpo Järvinen static __inline__ void 318bdf1ee5dSIlpo Järvinen TCP_ECN_make_synack(struct request_sock *req, struct tcphdr *th) 319bdf1ee5dSIlpo Järvinen { 320bdf1ee5dSIlpo Järvinen if (inet_rsk(req)->ecn_ok) 321bdf1ee5dSIlpo Järvinen th->ece = 1; 322bdf1ee5dSIlpo Järvinen } 323bdf1ee5dSIlpo Järvinen 32467edfef7SAndi Kleen /* Set up ECN state for a packet on a ESTABLISHED socket that is about to 32567edfef7SAndi Kleen * be sent. 32667edfef7SAndi Kleen */ 327bdf1ee5dSIlpo Järvinen static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb, 328bdf1ee5dSIlpo Järvinen int tcp_header_len) 329bdf1ee5dSIlpo Järvinen { 330bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 331bdf1ee5dSIlpo Järvinen 332bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_OK) { 333bdf1ee5dSIlpo Järvinen /* Not-retransmitted data segment: set ECT and inject CWR. */ 334bdf1ee5dSIlpo Järvinen if (skb->len != tcp_header_len && 335bdf1ee5dSIlpo Järvinen !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { 336bdf1ee5dSIlpo Järvinen INET_ECN_xmit(sk); 337bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) { 338bdf1ee5dSIlpo Järvinen tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; 339bdf1ee5dSIlpo Järvinen tcp_hdr(skb)->cwr = 1; 340bdf1ee5dSIlpo Järvinen skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; 341bdf1ee5dSIlpo Järvinen } 342bdf1ee5dSIlpo Järvinen } else { 343bdf1ee5dSIlpo Järvinen /* ACK or retransmitted segment: clear ECT|CE */ 344bdf1ee5dSIlpo Järvinen INET_ECN_dontxmit(sk); 345bdf1ee5dSIlpo Järvinen } 346bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) 347bdf1ee5dSIlpo Järvinen tcp_hdr(skb)->ece = 1; 348bdf1ee5dSIlpo Järvinen } 349bdf1ee5dSIlpo Järvinen } 350bdf1ee5dSIlpo Järvinen 351e870a8efSIlpo Järvinen /* Constructs common control bits of non-data skb. If SYN/FIN is present, 352e870a8efSIlpo Järvinen * auto increment end seqno. 353e870a8efSIlpo Järvinen */ 354e870a8efSIlpo Järvinen static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags) 355e870a8efSIlpo Järvinen { 3562e8e18efSDavid S. Miller skb->ip_summed = CHECKSUM_PARTIAL; 357e870a8efSIlpo Järvinen skb->csum = 0; 358e870a8efSIlpo Järvinen 359e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->flags = flags; 360e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->sacked = 0; 361e870a8efSIlpo Järvinen 362e870a8efSIlpo Järvinen skb_shinfo(skb)->gso_segs = 1; 363e870a8efSIlpo Järvinen skb_shinfo(skb)->gso_size = 0; 364e870a8efSIlpo Järvinen skb_shinfo(skb)->gso_type = 0; 365e870a8efSIlpo Järvinen 366e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->seq = seq; 367a3433f35SChangli Gao if (flags & (TCPHDR_SYN | TCPHDR_FIN)) 368e870a8efSIlpo Järvinen seq++; 369e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->end_seq = seq; 370e870a8efSIlpo Järvinen } 371e870a8efSIlpo Järvinen 37233f5f57eSIlpo Järvinen static inline int tcp_urg_mode(const struct tcp_sock *tp) 37333f5f57eSIlpo Järvinen { 37433f5f57eSIlpo Järvinen return tp->snd_una != tp->snd_up; 37533f5f57eSIlpo Järvinen } 37633f5f57eSIlpo Järvinen 37733ad798cSAdam Langley #define OPTION_SACK_ADVERTISE (1 << 0) 37833ad798cSAdam Langley #define OPTION_TS (1 << 1) 37933ad798cSAdam Langley #define OPTION_MD5 (1 << 2) 38089e95a61SOri Finkelman #define OPTION_WSCALE (1 << 3) 381bd0388aeSWilliam Allen Simpson #define OPTION_COOKIE_EXTENSION (1 << 4) 38233ad798cSAdam Langley 38333ad798cSAdam Langley struct tcp_out_options { 38433ad798cSAdam Langley u8 options; /* bit field of OPTION_* */ 38533ad798cSAdam Langley u8 ws; /* window scale, 0 to disable */ 38633ad798cSAdam Langley u8 num_sack_blocks; /* number of SACK blocks to include */ 387bd0388aeSWilliam Allen Simpson u8 hash_size; /* bytes in hash_location */ 38833ad798cSAdam Langley u16 mss; /* 0 to disable */ 38933ad798cSAdam Langley __u32 tsval, tsecr; /* need to include OPTION_TS */ 390bd0388aeSWilliam Allen Simpson __u8 *hash_location; /* temporary pointer, overloaded */ 39133ad798cSAdam Langley }; 39233ad798cSAdam Langley 393bd0388aeSWilliam Allen Simpson /* The sysctl int routines are generic, so check consistency here. 394bd0388aeSWilliam Allen Simpson */ 395bd0388aeSWilliam Allen Simpson static u8 tcp_cookie_size_check(u8 desired) 396bd0388aeSWilliam Allen Simpson { 397f1987257SEric Dumazet int cookie_size; 398f1987257SEric Dumazet 399f1987257SEric Dumazet if (desired > 0) 400bd0388aeSWilliam Allen Simpson /* previously specified */ 401bd0388aeSWilliam Allen Simpson return desired; 402f1987257SEric Dumazet 403f1987257SEric Dumazet cookie_size = ACCESS_ONCE(sysctl_tcp_cookie_size); 404f1987257SEric Dumazet if (cookie_size <= 0) 405bd0388aeSWilliam Allen Simpson /* no default specified */ 406bd0388aeSWilliam Allen Simpson return 0; 407f1987257SEric Dumazet 408f1987257SEric Dumazet if (cookie_size <= TCP_COOKIE_MIN) 409bd0388aeSWilliam Allen Simpson /* value too small, specify minimum */ 410bd0388aeSWilliam Allen Simpson return TCP_COOKIE_MIN; 411f1987257SEric Dumazet 412f1987257SEric Dumazet if (cookie_size >= TCP_COOKIE_MAX) 413bd0388aeSWilliam Allen Simpson /* value too large, specify maximum */ 414bd0388aeSWilliam Allen Simpson return TCP_COOKIE_MAX; 415f1987257SEric Dumazet 416f1987257SEric Dumazet if (cookie_size & 1) 417bd0388aeSWilliam Allen Simpson /* 8-bit multiple, illegal, fix it */ 418f1987257SEric Dumazet cookie_size++; 419f1987257SEric Dumazet 420f1987257SEric Dumazet return (u8)cookie_size; 421bd0388aeSWilliam Allen Simpson } 422bd0388aeSWilliam Allen Simpson 42367edfef7SAndi Kleen /* Write previously computed TCP options to the packet. 42467edfef7SAndi Kleen * 42567edfef7SAndi Kleen * Beware: Something in the Internet is very sensitive to the ordering of 426fd6149d3SIlpo Järvinen * TCP options, we learned this through the hard way, so be careful here. 427fd6149d3SIlpo Järvinen * Luckily we can at least blame others for their non-compliance but from 428fd6149d3SIlpo Järvinen * inter-operatibility perspective it seems that we're somewhat stuck with 429fd6149d3SIlpo Järvinen * the ordering which we have been using if we want to keep working with 430fd6149d3SIlpo Järvinen * those broken things (not that it currently hurts anybody as there isn't 431fd6149d3SIlpo Järvinen * particular reason why the ordering would need to be changed). 432fd6149d3SIlpo Järvinen * 433fd6149d3SIlpo Järvinen * At least SACK_PERM as the first option is known to lead to a disaster 434fd6149d3SIlpo Järvinen * (but it may well be that other scenarios fail similarly). 435fd6149d3SIlpo Järvinen */ 43633ad798cSAdam Langley static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp, 437bd0388aeSWilliam Allen Simpson struct tcp_out_options *opts) 438bd0388aeSWilliam Allen Simpson { 439bd0388aeSWilliam Allen Simpson u8 options = opts->options; /* mungable copy */ 440bd0388aeSWilliam Allen Simpson 441bd0388aeSWilliam Allen Simpson /* Having both authentication and cookies for security is redundant, 442bd0388aeSWilliam Allen Simpson * and there's certainly not enough room. Instead, the cookie-less 443bd0388aeSWilliam Allen Simpson * extension variant is proposed. 444bd0388aeSWilliam Allen Simpson * 445bd0388aeSWilliam Allen Simpson * Consider the pessimal case with authentication. The options 446bd0388aeSWilliam Allen Simpson * could look like: 447bd0388aeSWilliam Allen Simpson * COOKIE|MD5(20) + MSS(4) + SACK|TS(12) + WSCALE(4) == 40 448bd0388aeSWilliam Allen Simpson */ 449bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_MD5 & options)) { 450bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_COOKIE_EXTENSION & options)) { 451bd0388aeSWilliam Allen Simpson *ptr++ = htonl((TCPOPT_COOKIE << 24) | 452bd0388aeSWilliam Allen Simpson (TCPOLEN_COOKIE_BASE << 16) | 453bd0388aeSWilliam Allen Simpson (TCPOPT_MD5SIG << 8) | 454bd0388aeSWilliam Allen Simpson TCPOLEN_MD5SIG); 455bd0388aeSWilliam Allen Simpson } else { 45633ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 45733ad798cSAdam Langley (TCPOPT_NOP << 16) | 45833ad798cSAdam Langley (TCPOPT_MD5SIG << 8) | 45933ad798cSAdam Langley TCPOLEN_MD5SIG); 460bd0388aeSWilliam Allen Simpson } 461bd0388aeSWilliam Allen Simpson options &= ~OPTION_COOKIE_EXTENSION; 462bd0388aeSWilliam Allen Simpson /* overload cookie hash location */ 463bd0388aeSWilliam Allen Simpson opts->hash_location = (__u8 *)ptr; 46433ad798cSAdam Langley ptr += 4; 46533ad798cSAdam Langley } 46633ad798cSAdam Langley 467fd6149d3SIlpo Järvinen if (unlikely(opts->mss)) { 468fd6149d3SIlpo Järvinen *ptr++ = htonl((TCPOPT_MSS << 24) | 469fd6149d3SIlpo Järvinen (TCPOLEN_MSS << 16) | 470fd6149d3SIlpo Järvinen opts->mss); 471fd6149d3SIlpo Järvinen } 472fd6149d3SIlpo Järvinen 473bd0388aeSWilliam Allen Simpson if (likely(OPTION_TS & options)) { 474bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 47533ad798cSAdam Langley *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | 47633ad798cSAdam Langley (TCPOLEN_SACK_PERM << 16) | 47733ad798cSAdam Langley (TCPOPT_TIMESTAMP << 8) | 47833ad798cSAdam Langley TCPOLEN_TIMESTAMP); 479bd0388aeSWilliam Allen Simpson options &= ~OPTION_SACK_ADVERTISE; 48033ad798cSAdam Langley } else { 481496c98dfSYOSHIFUJI Hideaki *ptr++ = htonl((TCPOPT_NOP << 24) | 48240efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 48340efc6faSStephen Hemminger (TCPOPT_TIMESTAMP << 8) | 48440efc6faSStephen Hemminger TCPOLEN_TIMESTAMP); 48540efc6faSStephen Hemminger } 48633ad798cSAdam Langley *ptr++ = htonl(opts->tsval); 48733ad798cSAdam Langley *ptr++ = htonl(opts->tsecr); 48833ad798cSAdam Langley } 48933ad798cSAdam Langley 490bd0388aeSWilliam Allen Simpson /* Specification requires after timestamp, so do it now. 491bd0388aeSWilliam Allen Simpson * 492bd0388aeSWilliam Allen Simpson * Consider the pessimal case without authentication. The options 493bd0388aeSWilliam Allen Simpson * could look like: 494bd0388aeSWilliam Allen Simpson * MSS(4) + SACK|TS(12) + COOKIE(20) + WSCALE(4) == 40 495bd0388aeSWilliam Allen Simpson */ 496bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_COOKIE_EXTENSION & options)) { 497bd0388aeSWilliam Allen Simpson __u8 *cookie_copy = opts->hash_location; 498bd0388aeSWilliam Allen Simpson u8 cookie_size = opts->hash_size; 499bd0388aeSWilliam Allen Simpson 500bd0388aeSWilliam Allen Simpson /* 8-bit multiple handled in tcp_cookie_size_check() above, 501bd0388aeSWilliam Allen Simpson * and elsewhere. 502bd0388aeSWilliam Allen Simpson */ 503bd0388aeSWilliam Allen Simpson if (0x2 & cookie_size) { 504bd0388aeSWilliam Allen Simpson __u8 *p = (__u8 *)ptr; 505bd0388aeSWilliam Allen Simpson 506bd0388aeSWilliam Allen Simpson /* 16-bit multiple */ 507bd0388aeSWilliam Allen Simpson *p++ = TCPOPT_COOKIE; 508bd0388aeSWilliam Allen Simpson *p++ = TCPOLEN_COOKIE_BASE + cookie_size; 509bd0388aeSWilliam Allen Simpson *p++ = *cookie_copy++; 510bd0388aeSWilliam Allen Simpson *p++ = *cookie_copy++; 511bd0388aeSWilliam Allen Simpson ptr++; 512bd0388aeSWilliam Allen Simpson cookie_size -= 2; 513bd0388aeSWilliam Allen Simpson } else { 514bd0388aeSWilliam Allen Simpson /* 32-bit multiple */ 515bd0388aeSWilliam Allen Simpson *ptr++ = htonl(((TCPOPT_NOP << 24) | 516bd0388aeSWilliam Allen Simpson (TCPOPT_NOP << 16) | 517bd0388aeSWilliam Allen Simpson (TCPOPT_COOKIE << 8) | 518bd0388aeSWilliam Allen Simpson TCPOLEN_COOKIE_BASE) + 519bd0388aeSWilliam Allen Simpson cookie_size); 520bd0388aeSWilliam Allen Simpson } 521bd0388aeSWilliam Allen Simpson 522bd0388aeSWilliam Allen Simpson if (cookie_size > 0) { 523bd0388aeSWilliam Allen Simpson memcpy(ptr, cookie_copy, cookie_size); 524bd0388aeSWilliam Allen Simpson ptr += (cookie_size / 4); 525bd0388aeSWilliam Allen Simpson } 526bd0388aeSWilliam Allen Simpson } 527bd0388aeSWilliam Allen Simpson 528bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 52933ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 53033ad798cSAdam Langley (TCPOPT_NOP << 16) | 53133ad798cSAdam Langley (TCPOPT_SACK_PERM << 8) | 53233ad798cSAdam Langley TCPOLEN_SACK_PERM); 53333ad798cSAdam Langley } 53433ad798cSAdam Langley 535bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_WSCALE & options)) { 53633ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 53733ad798cSAdam Langley (TCPOPT_WINDOW << 16) | 53833ad798cSAdam Langley (TCPOLEN_WINDOW << 8) | 53933ad798cSAdam Langley opts->ws); 54033ad798cSAdam Langley } 54133ad798cSAdam Langley 54233ad798cSAdam Langley if (unlikely(opts->num_sack_blocks)) { 54333ad798cSAdam Langley struct tcp_sack_block *sp = tp->rx_opt.dsack ? 54433ad798cSAdam Langley tp->duplicate_sack : tp->selective_acks; 54540efc6faSStephen Hemminger int this_sack; 54640efc6faSStephen Hemminger 54740efc6faSStephen Hemminger *ptr++ = htonl((TCPOPT_NOP << 24) | 54840efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 54940efc6faSStephen Hemminger (TCPOPT_SACK << 8) | 55033ad798cSAdam Langley (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * 55140efc6faSStephen Hemminger TCPOLEN_SACK_PERBLOCK))); 5522de979bdSStephen Hemminger 55333ad798cSAdam Langley for (this_sack = 0; this_sack < opts->num_sack_blocks; 55433ad798cSAdam Langley ++this_sack) { 55540efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].start_seq); 55640efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].end_seq); 55740efc6faSStephen Hemminger } 5582de979bdSStephen Hemminger 55940efc6faSStephen Hemminger tp->rx_opt.dsack = 0; 56040efc6faSStephen Hemminger } 56140efc6faSStephen Hemminger } 56240efc6faSStephen Hemminger 56367edfef7SAndi Kleen /* Compute TCP options for SYN packets. This is not the final 56467edfef7SAndi Kleen * network wire format yet. 56567edfef7SAndi Kleen */ 56633ad798cSAdam Langley static unsigned tcp_syn_options(struct sock *sk, struct sk_buff *skb, 56733ad798cSAdam Langley struct tcp_out_options *opts, 56833ad798cSAdam Langley struct tcp_md5sig_key **md5) { 56933ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 570bd0388aeSWilliam Allen Simpson struct tcp_cookie_values *cvp = tp->cookie_values; 571bd0388aeSWilliam Allen Simpson unsigned remaining = MAX_TCP_OPTION_SPACE; 572bd0388aeSWilliam Allen Simpson u8 cookie_size = (!tp->rx_opt.cookie_out_never && cvp != NULL) ? 573bd0388aeSWilliam Allen Simpson tcp_cookie_size_check(cvp->cookie_desired) : 574bd0388aeSWilliam Allen Simpson 0; 57533ad798cSAdam Langley 576cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 57733ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 57833ad798cSAdam Langley if (*md5) { 57933ad798cSAdam Langley opts->options |= OPTION_MD5; 580bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 581cfb6eeb4SYOSHIFUJI Hideaki } 58233ad798cSAdam Langley #else 58333ad798cSAdam Langley *md5 = NULL; 584cfb6eeb4SYOSHIFUJI Hideaki #endif 58533ad798cSAdam Langley 58633ad798cSAdam Langley /* We always get an MSS option. The option bytes which will be seen in 58733ad798cSAdam Langley * normal data packets should timestamps be used, must be in the MSS 58833ad798cSAdam Langley * advertised. But we subtract them from tp->mss_cache so that 58933ad798cSAdam Langley * calculations in tcp_sendmsg are simpler etc. So account for this 59033ad798cSAdam Langley * fact here if necessary. If we don't do this correctly, as a 59133ad798cSAdam Langley * receiver we won't recognize data packets as being full sized when we 59233ad798cSAdam Langley * should, and thus we won't abide by the delayed ACK rules correctly. 59333ad798cSAdam Langley * SACKs don't matter, we never delay an ACK when we have any of those 59433ad798cSAdam Langley * going out. */ 59533ad798cSAdam Langley opts->mss = tcp_advertise_mss(sk); 596bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 59733ad798cSAdam Langley 598bb5b7c11SDavid S. Miller if (likely(sysctl_tcp_timestamps && *md5 == NULL)) { 59933ad798cSAdam Langley opts->options |= OPTION_TS; 60033ad798cSAdam Langley opts->tsval = TCP_SKB_CB(skb)->when; 60133ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 602bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 60333ad798cSAdam Langley } 604bb5b7c11SDavid S. Miller if (likely(sysctl_tcp_window_scaling)) { 60533ad798cSAdam Langley opts->ws = tp->rx_opt.rcv_wscale; 60689e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 607bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 60833ad798cSAdam Langley } 609bb5b7c11SDavid S. Miller if (likely(sysctl_tcp_sack)) { 61033ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 611b32d1310SDavid S. Miller if (unlikely(!(OPTION_TS & opts->options))) 612bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 61333ad798cSAdam Langley } 61433ad798cSAdam Langley 615bd0388aeSWilliam Allen Simpson /* Note that timestamps are required by the specification. 616bd0388aeSWilliam Allen Simpson * 617bd0388aeSWilliam Allen Simpson * Odd numbers of bytes are prohibited by the specification, ensuring 618bd0388aeSWilliam Allen Simpson * that the cookie is 16-bit aligned, and the resulting cookie pair is 619bd0388aeSWilliam Allen Simpson * 32-bit aligned. 620bd0388aeSWilliam Allen Simpson */ 621bd0388aeSWilliam Allen Simpson if (*md5 == NULL && 622bd0388aeSWilliam Allen Simpson (OPTION_TS & opts->options) && 623bd0388aeSWilliam Allen Simpson cookie_size > 0) { 624bd0388aeSWilliam Allen Simpson int need = TCPOLEN_COOKIE_BASE + cookie_size; 625bd0388aeSWilliam Allen Simpson 626bd0388aeSWilliam Allen Simpson if (0x2 & need) { 627bd0388aeSWilliam Allen Simpson /* 32-bit multiple */ 628bd0388aeSWilliam Allen Simpson need += 2; /* NOPs */ 629bd0388aeSWilliam Allen Simpson 630bd0388aeSWilliam Allen Simpson if (need > remaining) { 631bd0388aeSWilliam Allen Simpson /* try shrinking cookie to fit */ 632bd0388aeSWilliam Allen Simpson cookie_size -= 2; 633bd0388aeSWilliam Allen Simpson need -= 4; 634bd0388aeSWilliam Allen Simpson } 635bd0388aeSWilliam Allen Simpson } 636bd0388aeSWilliam Allen Simpson while (need > remaining && TCP_COOKIE_MIN <= cookie_size) { 637bd0388aeSWilliam Allen Simpson cookie_size -= 4; 638bd0388aeSWilliam Allen Simpson need -= 4; 639bd0388aeSWilliam Allen Simpson } 640bd0388aeSWilliam Allen Simpson if (TCP_COOKIE_MIN <= cookie_size) { 641bd0388aeSWilliam Allen Simpson opts->options |= OPTION_COOKIE_EXTENSION; 642bd0388aeSWilliam Allen Simpson opts->hash_location = (__u8 *)&cvp->cookie_pair[0]; 643bd0388aeSWilliam Allen Simpson opts->hash_size = cookie_size; 644bd0388aeSWilliam Allen Simpson 645bd0388aeSWilliam Allen Simpson /* Remember for future incarnations. */ 646bd0388aeSWilliam Allen Simpson cvp->cookie_desired = cookie_size; 647bd0388aeSWilliam Allen Simpson 648bd0388aeSWilliam Allen Simpson if (cvp->cookie_desired != cvp->cookie_pair_size) { 649bd0388aeSWilliam Allen Simpson /* Currently use random bytes as a nonce, 650bd0388aeSWilliam Allen Simpson * assuming these are completely unpredictable 651bd0388aeSWilliam Allen Simpson * by hostile users of the same system. 652bd0388aeSWilliam Allen Simpson */ 653bd0388aeSWilliam Allen Simpson get_random_bytes(&cvp->cookie_pair[0], 654bd0388aeSWilliam Allen Simpson cookie_size); 655bd0388aeSWilliam Allen Simpson cvp->cookie_pair_size = cookie_size; 656bd0388aeSWilliam Allen Simpson } 657bd0388aeSWilliam Allen Simpson 658bd0388aeSWilliam Allen Simpson remaining -= need; 659bd0388aeSWilliam Allen Simpson } 660bd0388aeSWilliam Allen Simpson } 661bd0388aeSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 66233ad798cSAdam Langley } 66333ad798cSAdam Langley 66467edfef7SAndi Kleen /* Set up TCP options for SYN-ACKs. */ 66533ad798cSAdam Langley static unsigned tcp_synack_options(struct sock *sk, 66633ad798cSAdam Langley struct request_sock *req, 66733ad798cSAdam Langley unsigned mss, struct sk_buff *skb, 66833ad798cSAdam Langley struct tcp_out_options *opts, 6694957faadSWilliam Allen Simpson struct tcp_md5sig_key **md5, 6704957faadSWilliam Allen Simpson struct tcp_extend_values *xvp) 6714957faadSWilliam Allen Simpson { 67233ad798cSAdam Langley struct inet_request_sock *ireq = inet_rsk(req); 6734957faadSWilliam Allen Simpson unsigned remaining = MAX_TCP_OPTION_SPACE; 6744957faadSWilliam Allen Simpson u8 cookie_plus = (xvp != NULL && !xvp->cookie_out_never) ? 6754957faadSWilliam Allen Simpson xvp->cookie_plus : 6764957faadSWilliam Allen Simpson 0; 67733ad798cSAdam Langley 67833ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 67933ad798cSAdam Langley *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req); 68033ad798cSAdam Langley if (*md5) { 68133ad798cSAdam Langley opts->options |= OPTION_MD5; 6824957faadSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 6834957faadSWilliam Allen Simpson 6844957faadSWilliam Allen Simpson /* We can't fit any SACK blocks in a packet with MD5 + TS 6854957faadSWilliam Allen Simpson * options. There was discussion about disabling SACK 6864957faadSWilliam Allen Simpson * rather than TS in order to fit in better with old, 6874957faadSWilliam Allen Simpson * buggy kernels, but that was deemed to be unnecessary. 6884957faadSWilliam Allen Simpson */ 689de213e5eSEric Dumazet ireq->tstamp_ok &= !ireq->sack_ok; 69033ad798cSAdam Langley } 69133ad798cSAdam Langley #else 69233ad798cSAdam Langley *md5 = NULL; 69333ad798cSAdam Langley #endif 69433ad798cSAdam Langley 6954957faadSWilliam Allen Simpson /* We always send an MSS option. */ 69633ad798cSAdam Langley opts->mss = mss; 6974957faadSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 69833ad798cSAdam Langley 69933ad798cSAdam Langley if (likely(ireq->wscale_ok)) { 70033ad798cSAdam Langley opts->ws = ireq->rcv_wscale; 70189e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 7024957faadSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 70333ad798cSAdam Langley } 704de213e5eSEric Dumazet if (likely(ireq->tstamp_ok)) { 70533ad798cSAdam Langley opts->options |= OPTION_TS; 70633ad798cSAdam Langley opts->tsval = TCP_SKB_CB(skb)->when; 70733ad798cSAdam Langley opts->tsecr = req->ts_recent; 7084957faadSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 70933ad798cSAdam Langley } 71033ad798cSAdam Langley if (likely(ireq->sack_ok)) { 71133ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 712de213e5eSEric Dumazet if (unlikely(!ireq->tstamp_ok)) 7134957faadSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 71433ad798cSAdam Langley } 71533ad798cSAdam Langley 7164957faadSWilliam Allen Simpson /* Similar rationale to tcp_syn_options() applies here, too. 7174957faadSWilliam Allen Simpson * If the <SYN> options fit, the same options should fit now! 7184957faadSWilliam Allen Simpson */ 7194957faadSWilliam Allen Simpson if (*md5 == NULL && 720de213e5eSEric Dumazet ireq->tstamp_ok && 7214957faadSWilliam Allen Simpson cookie_plus > TCPOLEN_COOKIE_BASE) { 7224957faadSWilliam Allen Simpson int need = cookie_plus; /* has TCPOLEN_COOKIE_BASE */ 7234957faadSWilliam Allen Simpson 7244957faadSWilliam Allen Simpson if (0x2 & need) { 7254957faadSWilliam Allen Simpson /* 32-bit multiple */ 7264957faadSWilliam Allen Simpson need += 2; /* NOPs */ 7274957faadSWilliam Allen Simpson } 7284957faadSWilliam Allen Simpson if (need <= remaining) { 7294957faadSWilliam Allen Simpson opts->options |= OPTION_COOKIE_EXTENSION; 7304957faadSWilliam Allen Simpson opts->hash_size = cookie_plus - TCPOLEN_COOKIE_BASE; 7314957faadSWilliam Allen Simpson remaining -= need; 7324957faadSWilliam Allen Simpson } else { 7334957faadSWilliam Allen Simpson /* There's no error return, so flag it. */ 7344957faadSWilliam Allen Simpson xvp->cookie_out_never = 1; /* true */ 7354957faadSWilliam Allen Simpson opts->hash_size = 0; 7364957faadSWilliam Allen Simpson } 7374957faadSWilliam Allen Simpson } 7384957faadSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 73933ad798cSAdam Langley } 74033ad798cSAdam Langley 74167edfef7SAndi Kleen /* Compute TCP options for ESTABLISHED sockets. This is not the 74267edfef7SAndi Kleen * final wire format yet. 74367edfef7SAndi Kleen */ 74433ad798cSAdam Langley static unsigned tcp_established_options(struct sock *sk, struct sk_buff *skb, 74533ad798cSAdam Langley struct tcp_out_options *opts, 74633ad798cSAdam Langley struct tcp_md5sig_key **md5) { 74733ad798cSAdam Langley struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL; 74833ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 74933ad798cSAdam Langley unsigned size = 0; 750cabeccbdSIlpo Järvinen unsigned int eff_sacks; 75133ad798cSAdam Langley 75233ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 75333ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 75433ad798cSAdam Langley if (unlikely(*md5)) { 75533ad798cSAdam Langley opts->options |= OPTION_MD5; 75633ad798cSAdam Langley size += TCPOLEN_MD5SIG_ALIGNED; 75733ad798cSAdam Langley } 75833ad798cSAdam Langley #else 75933ad798cSAdam Langley *md5 = NULL; 76033ad798cSAdam Langley #endif 76133ad798cSAdam Langley 76233ad798cSAdam Langley if (likely(tp->rx_opt.tstamp_ok)) { 76333ad798cSAdam Langley opts->options |= OPTION_TS; 76433ad798cSAdam Langley opts->tsval = tcb ? tcb->when : 0; 76533ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 76633ad798cSAdam Langley size += TCPOLEN_TSTAMP_ALIGNED; 76733ad798cSAdam Langley } 76833ad798cSAdam Langley 769cabeccbdSIlpo Järvinen eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; 770cabeccbdSIlpo Järvinen if (unlikely(eff_sacks)) { 77133ad798cSAdam Langley const unsigned remaining = MAX_TCP_OPTION_SPACE - size; 77233ad798cSAdam Langley opts->num_sack_blocks = 773cabeccbdSIlpo Järvinen min_t(unsigned, eff_sacks, 77433ad798cSAdam Langley (remaining - TCPOLEN_SACK_BASE_ALIGNED) / 77533ad798cSAdam Langley TCPOLEN_SACK_PERBLOCK); 77633ad798cSAdam Langley size += TCPOLEN_SACK_BASE_ALIGNED + 77733ad798cSAdam Langley opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; 77833ad798cSAdam Langley } 77933ad798cSAdam Langley 78033ad798cSAdam Langley return size; 78140efc6faSStephen Hemminger } 7821da177e4SLinus Torvalds 7831da177e4SLinus Torvalds /* This routine actually transmits TCP packets queued in by 7841da177e4SLinus Torvalds * tcp_do_sendmsg(). This is used by both the initial 7851da177e4SLinus Torvalds * transmission and possible later retransmissions. 7861da177e4SLinus Torvalds * All SKB's seen here are completely headerless. It is our 7871da177e4SLinus Torvalds * job to build the TCP header, and pass the packet down to 7881da177e4SLinus Torvalds * IP so it can do the same plus pass the packet off to the 7891da177e4SLinus Torvalds * device. 7901da177e4SLinus Torvalds * 7911da177e4SLinus Torvalds * We are working here with either a clone of the original 7921da177e4SLinus Torvalds * SKB, or a fresh unique copy made by the retransmit engine. 7931da177e4SLinus Torvalds */ 794056834d9SIlpo Järvinen static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, 795056834d9SIlpo Järvinen gfp_t gfp_mask) 7961da177e4SLinus Torvalds { 7976687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 798dfb4b9dcSDavid S. Miller struct inet_sock *inet; 799dfb4b9dcSDavid S. Miller struct tcp_sock *tp; 800dfb4b9dcSDavid S. Miller struct tcp_skb_cb *tcb; 80133ad798cSAdam Langley struct tcp_out_options opts; 80233ad798cSAdam Langley unsigned tcp_options_size, tcp_header_size; 803cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 8041da177e4SLinus Torvalds struct tcphdr *th; 8051da177e4SLinus Torvalds int err; 8061da177e4SLinus Torvalds 807dfb4b9dcSDavid S. Miller BUG_ON(!skb || !tcp_skb_pcount(skb)); 808dfb4b9dcSDavid S. Miller 809dfb4b9dcSDavid S. Miller /* If congestion control is doing timestamping, we must 810dfb4b9dcSDavid S. Miller * take such a timestamp before we potentially clone/copy. 811dfb4b9dcSDavid S. Miller */ 812164891aaSStephen Hemminger if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP) 813dfb4b9dcSDavid S. Miller __net_timestamp(skb); 814dfb4b9dcSDavid S. Miller 815dfb4b9dcSDavid S. Miller if (likely(clone_it)) { 816dfb4b9dcSDavid S. Miller if (unlikely(skb_cloned(skb))) 817dfb4b9dcSDavid S. Miller skb = pskb_copy(skb, gfp_mask); 818dfb4b9dcSDavid S. Miller else 819dfb4b9dcSDavid S. Miller skb = skb_clone(skb, gfp_mask); 820dfb4b9dcSDavid S. Miller if (unlikely(!skb)) 821dfb4b9dcSDavid S. Miller return -ENOBUFS; 822dfb4b9dcSDavid S. Miller } 823dfb4b9dcSDavid S. Miller 824dfb4b9dcSDavid S. Miller inet = inet_sk(sk); 825dfb4b9dcSDavid S. Miller tp = tcp_sk(sk); 826dfb4b9dcSDavid S. Miller tcb = TCP_SKB_CB(skb); 82733ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 8281da177e4SLinus Torvalds 829a3433f35SChangli Gao if (unlikely(tcb->flags & TCPHDR_SYN)) 83033ad798cSAdam Langley tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); 83133ad798cSAdam Langley else 83233ad798cSAdam Langley tcp_options_size = tcp_established_options(sk, skb, &opts, 83333ad798cSAdam Langley &md5); 83433ad798cSAdam Langley tcp_header_size = tcp_options_size + sizeof(struct tcphdr); 8351da177e4SLinus Torvalds 8363853b584STom Herbert if (tcp_packets_in_flight(tp) == 0) { 8376687e988SArnaldo Carvalho de Melo tcp_ca_event(sk, CA_EVENT_TX_START); 8383853b584STom Herbert skb->ooo_okay = 1; 8393853b584STom Herbert } else 8403853b584STom Herbert skb->ooo_okay = 0; 8411da177e4SLinus Torvalds 842aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 843aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 844e89862f4SDavid S. Miller skb_set_owner_w(skb, sk); 8451da177e4SLinus Torvalds 8461da177e4SLinus Torvalds /* Build TCP header and checksum it. */ 847aa8223c7SArnaldo Carvalho de Melo th = tcp_hdr(skb); 848c720c7e8SEric Dumazet th->source = inet->inet_sport; 849c720c7e8SEric Dumazet th->dest = inet->inet_dport; 8501da177e4SLinus Torvalds th->seq = htonl(tcb->seq); 8511da177e4SLinus Torvalds th->ack_seq = htonl(tp->rcv_nxt); 852df7a3b07SAl Viro *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | 853dfb4b9dcSDavid S. Miller tcb->flags); 854dfb4b9dcSDavid S. Miller 855a3433f35SChangli Gao if (unlikely(tcb->flags & TCPHDR_SYN)) { 8561da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments 8571da177e4SLinus Torvalds * is never scaled. 8581da177e4SLinus Torvalds */ 859600ff0c2SIlpo Järvinen th->window = htons(min(tp->rcv_wnd, 65535U)); 8601da177e4SLinus Torvalds } else { 8611da177e4SLinus Torvalds th->window = htons(tcp_select_window(sk)); 8621da177e4SLinus Torvalds } 8631da177e4SLinus Torvalds th->check = 0; 8641da177e4SLinus Torvalds th->urg_ptr = 0; 8651da177e4SLinus Torvalds 86633f5f57eSIlpo Järvinen /* The urg_mode check is necessary during a below snd_una win probe */ 8677691367dSHerbert Xu if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { 8687691367dSHerbert Xu if (before(tp->snd_up, tcb->seq + 0x10000)) { 8691da177e4SLinus Torvalds th->urg_ptr = htons(tp->snd_up - tcb->seq); 8701da177e4SLinus Torvalds th->urg = 1; 8717691367dSHerbert Xu } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { 8720eae88f3SEric Dumazet th->urg_ptr = htons(0xFFFF); 8737691367dSHerbert Xu th->urg = 1; 8747691367dSHerbert Xu } 8751da177e4SLinus Torvalds } 8761da177e4SLinus Torvalds 877bd0388aeSWilliam Allen Simpson tcp_options_write((__be32 *)(th + 1), tp, &opts); 878a3433f35SChangli Gao if (likely((tcb->flags & TCPHDR_SYN) == 0)) 8799e412ba7SIlpo Järvinen TCP_ECN_send(sk, skb, tcp_header_size); 880dfb4b9dcSDavid S. Miller 881cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 882cfb6eeb4SYOSHIFUJI Hideaki /* Calculate the MD5 hash, as we have all we need now */ 883cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 884a465419bSEric Dumazet sk_nocaps_add(sk, NETIF_F_GSO_MASK); 885bd0388aeSWilliam Allen Simpson tp->af_specific->calc_md5_hash(opts.hash_location, 88649a72dfbSAdam Langley md5, sk, NULL, skb); 887cfb6eeb4SYOSHIFUJI Hideaki } 888cfb6eeb4SYOSHIFUJI Hideaki #endif 889cfb6eeb4SYOSHIFUJI Hideaki 890bb296246SHerbert Xu icsk->icsk_af_ops->send_check(sk, skb); 8911da177e4SLinus Torvalds 892a3433f35SChangli Gao if (likely(tcb->flags & TCPHDR_ACK)) 893fc6415bcSDavid S. Miller tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); 8941da177e4SLinus Torvalds 8951da177e4SLinus Torvalds if (skb->len != tcp_header_size) 8961da177e4SLinus Torvalds tcp_event_data_sent(tp, skb, sk); 8971da177e4SLinus Torvalds 898bd37a088SWei Yongjun if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) 899aa2ea058STom Herbert TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, 900aa2ea058STom Herbert tcp_skb_pcount(skb)); 9011da177e4SLinus Torvalds 9024e15ed4dSShan Wei err = icsk->icsk_af_ops->queue_xmit(skb); 90383de47cdSHua Zhong if (likely(err <= 0)) 9041da177e4SLinus Torvalds return err; 9051da177e4SLinus Torvalds 9063cfe3baaSIlpo Järvinen tcp_enter_cwr(sk, 1); 9071da177e4SLinus Torvalds 908b9df3cb8SGerrit Renker return net_xmit_eval(err); 9091da177e4SLinus Torvalds } 9101da177e4SLinus Torvalds 91167edfef7SAndi Kleen /* This routine just queues the buffer for sending. 9121da177e4SLinus Torvalds * 9131da177e4SLinus Torvalds * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, 9141da177e4SLinus Torvalds * otherwise socket can stall. 9151da177e4SLinus Torvalds */ 9161da177e4SLinus Torvalds static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) 9171da177e4SLinus Torvalds { 9181da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 9191da177e4SLinus Torvalds 9201da177e4SLinus Torvalds /* Advance write_seq and place onto the write_queue. */ 9211da177e4SLinus Torvalds tp->write_seq = TCP_SKB_CB(skb)->end_seq; 9221da177e4SLinus Torvalds skb_header_release(skb); 923fe067e8aSDavid S. Miller tcp_add_write_queue_tail(sk, skb); 9243ab224beSHideo Aoki sk->sk_wmem_queued += skb->truesize; 9253ab224beSHideo Aoki sk_mem_charge(sk, skb->truesize); 9261da177e4SLinus Torvalds } 9271da177e4SLinus Torvalds 92867edfef7SAndi Kleen /* Initialize TSO segments for a packet. */ 929056834d9SIlpo Järvinen static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb, 930056834d9SIlpo Järvinen unsigned int mss_now) 931f6302d1dSDavid S. Miller { 9328e5b9ddaSHerbert Xu if (skb->len <= mss_now || !sk_can_gso(sk) || 9338e5b9ddaSHerbert Xu skb->ip_summed == CHECKSUM_NONE) { 934f6302d1dSDavid S. Miller /* Avoid the costly divide in the normal 935f6302d1dSDavid S. Miller * non-TSO case. 936f6302d1dSDavid S. Miller */ 9377967168cSHerbert Xu skb_shinfo(skb)->gso_segs = 1; 9387967168cSHerbert Xu skb_shinfo(skb)->gso_size = 0; 9397967168cSHerbert Xu skb_shinfo(skb)->gso_type = 0; 940f6302d1dSDavid S. Miller } else { 941356f89e1SIlpo Järvinen skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now); 9427967168cSHerbert Xu skb_shinfo(skb)->gso_size = mss_now; 943bcd76111SHerbert Xu skb_shinfo(skb)->gso_type = sk->sk_gso_type; 9441da177e4SLinus Torvalds } 9451da177e4SLinus Torvalds } 9461da177e4SLinus Torvalds 94791fed7a1SIlpo Järvinen /* When a modification to fackets out becomes necessary, we need to check 94868f8353bSIlpo Järvinen * skb is counted to fackets_out or not. 94991fed7a1SIlpo Järvinen */ 950a47e5a98SIlpo Järvinen static void tcp_adjust_fackets_out(struct sock *sk, struct sk_buff *skb, 95191fed7a1SIlpo Järvinen int decr) 95291fed7a1SIlpo Järvinen { 953a47e5a98SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 954a47e5a98SIlpo Järvinen 955dc86967bSIlpo Järvinen if (!tp->sacked_out || tcp_is_reno(tp)) 95691fed7a1SIlpo Järvinen return; 95791fed7a1SIlpo Järvinen 9586859d494SIlpo Järvinen if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq)) 95991fed7a1SIlpo Järvinen tp->fackets_out -= decr; 96091fed7a1SIlpo Järvinen } 96191fed7a1SIlpo Järvinen 962797108d1SIlpo Järvinen /* Pcount in the middle of the write queue got changed, we need to do various 963797108d1SIlpo Järvinen * tweaks to fix counters 964797108d1SIlpo Järvinen */ 965797108d1SIlpo Järvinen static void tcp_adjust_pcount(struct sock *sk, struct sk_buff *skb, int decr) 966797108d1SIlpo Järvinen { 967797108d1SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 968797108d1SIlpo Järvinen 969797108d1SIlpo Järvinen tp->packets_out -= decr; 970797108d1SIlpo Järvinen 971797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 972797108d1SIlpo Järvinen tp->sacked_out -= decr; 973797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 974797108d1SIlpo Järvinen tp->retrans_out -= decr; 975797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 976797108d1SIlpo Järvinen tp->lost_out -= decr; 977797108d1SIlpo Järvinen 978797108d1SIlpo Järvinen /* Reno case is special. Sigh... */ 979797108d1SIlpo Järvinen if (tcp_is_reno(tp) && decr > 0) 980797108d1SIlpo Järvinen tp->sacked_out -= min_t(u32, tp->sacked_out, decr); 981797108d1SIlpo Järvinen 982797108d1SIlpo Järvinen tcp_adjust_fackets_out(sk, skb, decr); 983797108d1SIlpo Järvinen 984797108d1SIlpo Järvinen if (tp->lost_skb_hint && 985797108d1SIlpo Järvinen before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && 98652cf3cc8SIlpo Järvinen (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))) 987797108d1SIlpo Järvinen tp->lost_cnt_hint -= decr; 988797108d1SIlpo Järvinen 989797108d1SIlpo Järvinen tcp_verify_left_out(tp); 990797108d1SIlpo Järvinen } 991797108d1SIlpo Järvinen 9921da177e4SLinus Torvalds /* Function to create two new TCP segments. Shrinks the given segment 9931da177e4SLinus Torvalds * to the specified size and appends a new segment with the rest of the 9941da177e4SLinus Torvalds * packet to the list. This won't be called frequently, I hope. 9951da177e4SLinus Torvalds * Remember, these are still headerless SKBs at this point. 9961da177e4SLinus Torvalds */ 997056834d9SIlpo Järvinen int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, 998056834d9SIlpo Järvinen unsigned int mss_now) 9991da177e4SLinus Torvalds { 10001da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 10011da177e4SLinus Torvalds struct sk_buff *buff; 10026475be16SDavid S. Miller int nsize, old_factor; 1003b60b49eaSHerbert Xu int nlen; 10049ce01461SIlpo Järvinen u8 flags; 10051da177e4SLinus Torvalds 1006*2fceec13SIlpo Järvinen if (WARN_ON(len > skb->len)) 1007*2fceec13SIlpo Järvinen return -EINVAL; 10086a438bbeSStephen Hemminger 10091da177e4SLinus Torvalds nsize = skb_headlen(skb) - len; 10101da177e4SLinus Torvalds if (nsize < 0) 10111da177e4SLinus Torvalds nsize = 0; 10121da177e4SLinus Torvalds 10131da177e4SLinus Torvalds if (skb_cloned(skb) && 10141da177e4SLinus Torvalds skb_is_nonlinear(skb) && 10151da177e4SLinus Torvalds pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) 10161da177e4SLinus Torvalds return -ENOMEM; 10171da177e4SLinus Torvalds 10181da177e4SLinus Torvalds /* Get a new skb... force flag on. */ 10191da177e4SLinus Torvalds buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC); 10201da177e4SLinus Torvalds if (buff == NULL) 10211da177e4SLinus Torvalds return -ENOMEM; /* We'll just try again later. */ 1022ef5cb973SHerbert Xu 10233ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 10243ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1025b60b49eaSHerbert Xu nlen = skb->len - len - nsize; 1026b60b49eaSHerbert Xu buff->truesize += nlen; 1027b60b49eaSHerbert Xu skb->truesize -= nlen; 10281da177e4SLinus Torvalds 10291da177e4SLinus Torvalds /* Correct the sequence numbers. */ 10301da177e4SLinus Torvalds TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 10311da177e4SLinus Torvalds TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 10321da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 10331da177e4SLinus Torvalds 10341da177e4SLinus Torvalds /* PSH and FIN should only be set in the second packet. */ 10351da177e4SLinus Torvalds flags = TCP_SKB_CB(skb)->flags; 1036a3433f35SChangli Gao TCP_SKB_CB(skb)->flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 10371da177e4SLinus Torvalds TCP_SKB_CB(buff)->flags = flags; 1038e14c3cafSHerbert Xu TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; 10391da177e4SLinus Torvalds 104084fa7933SPatrick McHardy if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { 10411da177e4SLinus Torvalds /* Copy and checksum data tail into the new buffer. */ 1042056834d9SIlpo Järvinen buff->csum = csum_partial_copy_nocheck(skb->data + len, 1043056834d9SIlpo Järvinen skb_put(buff, nsize), 10441da177e4SLinus Torvalds nsize, 0); 10451da177e4SLinus Torvalds 10461da177e4SLinus Torvalds skb_trim(skb, len); 10471da177e4SLinus Torvalds 10481da177e4SLinus Torvalds skb->csum = csum_block_sub(skb->csum, buff->csum, len); 10491da177e4SLinus Torvalds } else { 105084fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 10511da177e4SLinus Torvalds skb_split(skb, buff, len); 10521da177e4SLinus Torvalds } 10531da177e4SLinus Torvalds 10541da177e4SLinus Torvalds buff->ip_summed = skb->ip_summed; 10551da177e4SLinus Torvalds 10561da177e4SLinus Torvalds /* Looks stupid, but our code really uses when of 10571da177e4SLinus Torvalds * skbs, which it never sent before. --ANK 10581da177e4SLinus Torvalds */ 10591da177e4SLinus Torvalds TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when; 1060a61bbcf2SPatrick McHardy buff->tstamp = skb->tstamp; 10611da177e4SLinus Torvalds 10626475be16SDavid S. Miller old_factor = tcp_skb_pcount(skb); 10636475be16SDavid S. Miller 10641da177e4SLinus Torvalds /* Fix up tso_factor for both original and new SKB. */ 1065846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1066846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, buff, mss_now); 10671da177e4SLinus Torvalds 10686475be16SDavid S. Miller /* If this packet has been sent out already, we must 10696475be16SDavid S. Miller * adjust the various packet counters. 10706475be16SDavid S. Miller */ 1071cf0b450cSHerbert Xu if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 10726475be16SDavid S. Miller int diff = old_factor - tcp_skb_pcount(skb) - 10736475be16SDavid S. Miller tcp_skb_pcount(buff); 10741da177e4SLinus Torvalds 1075797108d1SIlpo Järvinen if (diff) 1076797108d1SIlpo Järvinen tcp_adjust_pcount(sk, skb, diff); 10771da177e4SLinus Torvalds } 10781da177e4SLinus Torvalds 10791da177e4SLinus Torvalds /* Link BUFF into the send queue. */ 1080f44b5271SDavid S. Miller skb_header_release(buff); 1081fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 10821da177e4SLinus Torvalds 10831da177e4SLinus Torvalds return 0; 10841da177e4SLinus Torvalds } 10851da177e4SLinus Torvalds 10861da177e4SLinus Torvalds /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c 10871da177e4SLinus Torvalds * eventually). The difference is that pulled data not copied, but 10881da177e4SLinus Torvalds * immediately discarded. 10891da177e4SLinus Torvalds */ 1090f2911969SHerbert Xu ~{PmVHI~} static void __pskb_trim_head(struct sk_buff *skb, int len) 10911da177e4SLinus Torvalds { 10921da177e4SLinus Torvalds int i, k, eat; 10931da177e4SLinus Torvalds 10941da177e4SLinus Torvalds eat = len; 10951da177e4SLinus Torvalds k = 0; 10961da177e4SLinus Torvalds for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) { 10971da177e4SLinus Torvalds if (skb_shinfo(skb)->frags[i].size <= eat) { 10981da177e4SLinus Torvalds put_page(skb_shinfo(skb)->frags[i].page); 10991da177e4SLinus Torvalds eat -= skb_shinfo(skb)->frags[i].size; 11001da177e4SLinus Torvalds } else { 11011da177e4SLinus Torvalds skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i]; 11021da177e4SLinus Torvalds if (eat) { 11031da177e4SLinus Torvalds skb_shinfo(skb)->frags[k].page_offset += eat; 11041da177e4SLinus Torvalds skb_shinfo(skb)->frags[k].size -= eat; 11051da177e4SLinus Torvalds eat = 0; 11061da177e4SLinus Torvalds } 11071da177e4SLinus Torvalds k++; 11081da177e4SLinus Torvalds } 11091da177e4SLinus Torvalds } 11101da177e4SLinus Torvalds skb_shinfo(skb)->nr_frags = k; 11111da177e4SLinus Torvalds 111227a884dcSArnaldo Carvalho de Melo skb_reset_tail_pointer(skb); 11131da177e4SLinus Torvalds skb->data_len -= len; 11141da177e4SLinus Torvalds skb->len = skb->data_len; 11151da177e4SLinus Torvalds } 11161da177e4SLinus Torvalds 111767edfef7SAndi Kleen /* Remove acked data from a packet in the transmit queue. */ 11181da177e4SLinus Torvalds int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 11191da177e4SLinus Torvalds { 1120056834d9SIlpo Järvinen if (skb_cloned(skb) && pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) 11211da177e4SLinus Torvalds return -ENOMEM; 11221da177e4SLinus Torvalds 1123f2911969SHerbert Xu ~{PmVHI~} /* If len == headlen, we avoid __skb_pull to preserve alignment. */ 1124f2911969SHerbert Xu ~{PmVHI~} if (unlikely(len < skb_headlen(skb))) 11251da177e4SLinus Torvalds __skb_pull(skb, len); 1126f2911969SHerbert Xu ~{PmVHI~} else 1127f2911969SHerbert Xu ~{PmVHI~} __pskb_trim_head(skb, len - skb_headlen(skb)); 11281da177e4SLinus Torvalds 11291da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq += len; 113084fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 11311da177e4SLinus Torvalds 11321da177e4SLinus Torvalds skb->truesize -= len; 11331da177e4SLinus Torvalds sk->sk_wmem_queued -= len; 11343ab224beSHideo Aoki sk_mem_uncharge(sk, len); 11351da177e4SLinus Torvalds sock_set_flag(sk, SOCK_QUEUE_SHRUNK); 11361da177e4SLinus Torvalds 11371da177e4SLinus Torvalds /* Any change of skb->len requires recalculation of tso 11381da177e4SLinus Torvalds * factor and mss. 11391da177e4SLinus Torvalds */ 11401da177e4SLinus Torvalds if (tcp_skb_pcount(skb) > 1) 11410c54b85fSIlpo Järvinen tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk)); 11421da177e4SLinus Torvalds 11431da177e4SLinus Torvalds return 0; 11441da177e4SLinus Torvalds } 11451da177e4SLinus Torvalds 114667edfef7SAndi Kleen /* Calculate MSS. Not accounting for SACKs here. */ 11475d424d5aSJohn Heffner int tcp_mtu_to_mss(struct sock *sk, int pmtu) 11485d424d5aSJohn Heffner { 11495d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 11505d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 11515d424d5aSJohn Heffner int mss_now; 11525d424d5aSJohn Heffner 11535d424d5aSJohn Heffner /* Calculate base mss without TCP options: 11545d424d5aSJohn Heffner It is MMS_S - sizeof(tcphdr) of rfc1122 11555d424d5aSJohn Heffner */ 11565d424d5aSJohn Heffner mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 11575d424d5aSJohn Heffner 11585d424d5aSJohn Heffner /* Clamp it (mss_clamp does not include tcp options) */ 11595d424d5aSJohn Heffner if (mss_now > tp->rx_opt.mss_clamp) 11605d424d5aSJohn Heffner mss_now = tp->rx_opt.mss_clamp; 11615d424d5aSJohn Heffner 11625d424d5aSJohn Heffner /* Now subtract optional transport overhead */ 11635d424d5aSJohn Heffner mss_now -= icsk->icsk_ext_hdr_len; 11645d424d5aSJohn Heffner 11655d424d5aSJohn Heffner /* Then reserve room for full set of TCP options and 8 bytes of data */ 11665d424d5aSJohn Heffner if (mss_now < 48) 11675d424d5aSJohn Heffner mss_now = 48; 11685d424d5aSJohn Heffner 11695d424d5aSJohn Heffner /* Now subtract TCP options size, not including SACKs */ 11705d424d5aSJohn Heffner mss_now -= tp->tcp_header_len - sizeof(struct tcphdr); 11715d424d5aSJohn Heffner 11725d424d5aSJohn Heffner return mss_now; 11735d424d5aSJohn Heffner } 11745d424d5aSJohn Heffner 11755d424d5aSJohn Heffner /* Inverse of above */ 11765d424d5aSJohn Heffner int tcp_mss_to_mtu(struct sock *sk, int mss) 11775d424d5aSJohn Heffner { 11785d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 11795d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 11805d424d5aSJohn Heffner int mtu; 11815d424d5aSJohn Heffner 11825d424d5aSJohn Heffner mtu = mss + 11835d424d5aSJohn Heffner tp->tcp_header_len + 11845d424d5aSJohn Heffner icsk->icsk_ext_hdr_len + 11855d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 11865d424d5aSJohn Heffner 11875d424d5aSJohn Heffner return mtu; 11885d424d5aSJohn Heffner } 11895d424d5aSJohn Heffner 119067edfef7SAndi Kleen /* MTU probing init per socket */ 11915d424d5aSJohn Heffner void tcp_mtup_init(struct sock *sk) 11925d424d5aSJohn Heffner { 11935d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 11945d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 11955d424d5aSJohn Heffner 11965d424d5aSJohn Heffner icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1; 11975d424d5aSJohn Heffner icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 11985d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 11995d424d5aSJohn Heffner icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss); 12005d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 12015d424d5aSJohn Heffner } 12024bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_mtup_init); 12035d424d5aSJohn Heffner 12041da177e4SLinus Torvalds /* This function synchronize snd mss to current pmtu/exthdr set. 12051da177e4SLinus Torvalds 12061da177e4SLinus Torvalds tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 12071da177e4SLinus Torvalds for TCP options, but includes only bare TCP header. 12081da177e4SLinus Torvalds 12091da177e4SLinus Torvalds tp->rx_opt.mss_clamp is mss negotiated at connection setup. 1210caa20d9aSStephen Hemminger It is minimum of user_mss and mss received with SYN. 12111da177e4SLinus Torvalds It also does not include TCP options. 12121da177e4SLinus Torvalds 1213d83d8461SArnaldo Carvalho de Melo inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. 12141da177e4SLinus Torvalds 12151da177e4SLinus Torvalds tp->mss_cache is current effective sending mss, including 12161da177e4SLinus Torvalds all tcp options except for SACKs. It is evaluated, 12171da177e4SLinus Torvalds taking into account current pmtu, but never exceeds 12181da177e4SLinus Torvalds tp->rx_opt.mss_clamp. 12191da177e4SLinus Torvalds 12201da177e4SLinus Torvalds NOTE1. rfc1122 clearly states that advertised MSS 12211da177e4SLinus Torvalds DOES NOT include either tcp or ip options. 12221da177e4SLinus Torvalds 1223d83d8461SArnaldo Carvalho de Melo NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache 1224d83d8461SArnaldo Carvalho de Melo are READ ONLY outside this function. --ANK (980731) 12251da177e4SLinus Torvalds */ 12261da177e4SLinus Torvalds unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 12271da177e4SLinus Torvalds { 12281da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1229d83d8461SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 12305d424d5aSJohn Heffner int mss_now; 12311da177e4SLinus Torvalds 12325d424d5aSJohn Heffner if (icsk->icsk_mtup.search_high > pmtu) 12335d424d5aSJohn Heffner icsk->icsk_mtup.search_high = pmtu; 12341da177e4SLinus Torvalds 12355d424d5aSJohn Heffner mss_now = tcp_mtu_to_mss(sk, pmtu); 1236409d22b4SIlpo Järvinen mss_now = tcp_bound_to_half_wnd(tp, mss_now); 12371da177e4SLinus Torvalds 12381da177e4SLinus Torvalds /* And store cached results */ 1239d83d8461SArnaldo Carvalho de Melo icsk->icsk_pmtu_cookie = pmtu; 12405d424d5aSJohn Heffner if (icsk->icsk_mtup.enabled) 12415d424d5aSJohn Heffner mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); 1242c1b4a7e6SDavid S. Miller tp->mss_cache = mss_now; 12431da177e4SLinus Torvalds 12441da177e4SLinus Torvalds return mss_now; 12451da177e4SLinus Torvalds } 12464bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_sync_mss); 12471da177e4SLinus Torvalds 12481da177e4SLinus Torvalds /* Compute the current effective MSS, taking SACKs and IP options, 12491da177e4SLinus Torvalds * and even PMTU discovery events into account. 12501da177e4SLinus Torvalds */ 12510c54b85fSIlpo Järvinen unsigned int tcp_current_mss(struct sock *sk) 12521da177e4SLinus Torvalds { 12531da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 12541da177e4SLinus Torvalds struct dst_entry *dst = __sk_dst_get(sk); 1255c1b4a7e6SDavid S. Miller u32 mss_now; 125633ad798cSAdam Langley unsigned header_len; 125733ad798cSAdam Langley struct tcp_out_options opts; 125833ad798cSAdam Langley struct tcp_md5sig_key *md5; 12591da177e4SLinus Torvalds 1260c1b4a7e6SDavid S. Miller mss_now = tp->mss_cache; 1261c1b4a7e6SDavid S. Miller 12621da177e4SLinus Torvalds if (dst) { 12631da177e4SLinus Torvalds u32 mtu = dst_mtu(dst); 1264d83d8461SArnaldo Carvalho de Melo if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 12651da177e4SLinus Torvalds mss_now = tcp_sync_mss(sk, mtu); 12661da177e4SLinus Torvalds } 12671da177e4SLinus Torvalds 126833ad798cSAdam Langley header_len = tcp_established_options(sk, NULL, &opts, &md5) + 126933ad798cSAdam Langley sizeof(struct tcphdr); 127033ad798cSAdam Langley /* The mss_cache is sized based on tp->tcp_header_len, which assumes 127133ad798cSAdam Langley * some common options. If this is an odd packet (because we have SACK 127233ad798cSAdam Langley * blocks etc) then our calculated header_len will be different, and 127333ad798cSAdam Langley * we have to adjust mss_now correspondingly */ 127433ad798cSAdam Langley if (header_len != tp->tcp_header_len) { 127533ad798cSAdam Langley int delta = (int) header_len - tp->tcp_header_len; 127633ad798cSAdam Langley mss_now -= delta; 127733ad798cSAdam Langley } 1278cfb6eeb4SYOSHIFUJI Hideaki 12791da177e4SLinus Torvalds return mss_now; 12801da177e4SLinus Torvalds } 12811da177e4SLinus Torvalds 1282a762a980SDavid S. Miller /* Congestion window validation. (RFC2861) */ 12839e412ba7SIlpo Järvinen static void tcp_cwnd_validate(struct sock *sk) 1284a762a980SDavid S. Miller { 12859e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1286a762a980SDavid S. Miller 1287d436d686SIlpo Järvinen if (tp->packets_out >= tp->snd_cwnd) { 1288a762a980SDavid S. Miller /* Network is feed fully. */ 1289a762a980SDavid S. Miller tp->snd_cwnd_used = 0; 1290a762a980SDavid S. Miller tp->snd_cwnd_stamp = tcp_time_stamp; 1291a762a980SDavid S. Miller } else { 1292a762a980SDavid S. Miller /* Network starves. */ 1293a762a980SDavid S. Miller if (tp->packets_out > tp->snd_cwnd_used) 1294a762a980SDavid S. Miller tp->snd_cwnd_used = tp->packets_out; 1295a762a980SDavid S. Miller 129615d33c07SDavid S. Miller if (sysctl_tcp_slow_start_after_idle && 129715d33c07SDavid S. Miller (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto) 1298a762a980SDavid S. Miller tcp_cwnd_application_limited(sk); 1299a762a980SDavid S. Miller } 1300a762a980SDavid S. Miller } 1301a762a980SDavid S. Miller 13020e3a4803SIlpo Järvinen /* Returns the portion of skb which can be sent right away without 13030e3a4803SIlpo Järvinen * introducing MSS oddities to segment boundaries. In rare cases where 13040e3a4803SIlpo Järvinen * mss_now != mss_cache, we will request caller to create a small skb 13050e3a4803SIlpo Järvinen * per input skb which could be mostly avoided here (if desired). 13065ea3a748SIlpo Järvinen * 13075ea3a748SIlpo Järvinen * We explicitly want to create a request for splitting write queue tail 13085ea3a748SIlpo Järvinen * to a small skb for Nagle purposes while avoiding unnecessary modulos, 13095ea3a748SIlpo Järvinen * thus all the complexity (cwnd_len is always MSS multiple which we 13105ea3a748SIlpo Järvinen * return whenever allowed by the other factors). Basically we need the 13115ea3a748SIlpo Järvinen * modulo only when the receiver window alone is the limiting factor or 13125ea3a748SIlpo Järvinen * when we would be allowed to send the split-due-to-Nagle skb fully. 13130e3a4803SIlpo Järvinen */ 13140e3a4803SIlpo Järvinen static unsigned int tcp_mss_split_point(struct sock *sk, struct sk_buff *skb, 1315056834d9SIlpo Järvinen unsigned int mss_now, unsigned int cwnd) 1316c1b4a7e6SDavid S. Miller { 13170e3a4803SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 13180e3a4803SIlpo Järvinen u32 needed, window, cwnd_len; 1319c1b4a7e6SDavid S. Miller 132090840defSIlpo Järvinen window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1321c1b4a7e6SDavid S. Miller cwnd_len = mss_now * cwnd; 13220e3a4803SIlpo Järvinen 13230e3a4803SIlpo Järvinen if (likely(cwnd_len <= window && skb != tcp_write_queue_tail(sk))) 13240e3a4803SIlpo Järvinen return cwnd_len; 13250e3a4803SIlpo Järvinen 13265ea3a748SIlpo Järvinen needed = min(skb->len, window); 13275ea3a748SIlpo Järvinen 132817515408SIlpo Järvinen if (cwnd_len <= needed) 13290e3a4803SIlpo Järvinen return cwnd_len; 13300e3a4803SIlpo Järvinen 13310e3a4803SIlpo Järvinen return needed - needed % mss_now; 1332c1b4a7e6SDavid S. Miller } 1333c1b4a7e6SDavid S. Miller 1334c1b4a7e6SDavid S. Miller /* Can at least one segment of SKB be sent right now, according to the 1335c1b4a7e6SDavid S. Miller * congestion window rules? If so, return how many segments are allowed. 1336c1b4a7e6SDavid S. Miller */ 1337056834d9SIlpo Järvinen static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp, 1338056834d9SIlpo Järvinen struct sk_buff *skb) 1339c1b4a7e6SDavid S. Miller { 1340c1b4a7e6SDavid S. Miller u32 in_flight, cwnd; 1341c1b4a7e6SDavid S. Miller 1342c1b4a7e6SDavid S. Miller /* Don't be strict about the congestion window for the final FIN. */ 1343a3433f35SChangli Gao if ((TCP_SKB_CB(skb)->flags & TCPHDR_FIN) && tcp_skb_pcount(skb) == 1) 1344c1b4a7e6SDavid S. Miller return 1; 1345c1b4a7e6SDavid S. Miller 1346c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1347c1b4a7e6SDavid S. Miller cwnd = tp->snd_cwnd; 1348c1b4a7e6SDavid S. Miller if (in_flight < cwnd) 1349c1b4a7e6SDavid S. Miller return (cwnd - in_flight); 1350c1b4a7e6SDavid S. Miller 1351c1b4a7e6SDavid S. Miller return 0; 1352c1b4a7e6SDavid S. Miller } 1353c1b4a7e6SDavid S. Miller 1354b595076aSUwe Kleine-König /* Initialize TSO state of a skb. 135567edfef7SAndi Kleen * This must be invoked the first time we consider transmitting 1356c1b4a7e6SDavid S. Miller * SKB onto the wire. 1357c1b4a7e6SDavid S. Miller */ 1358056834d9SIlpo Järvinen static int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb, 1359056834d9SIlpo Järvinen unsigned int mss_now) 1360c1b4a7e6SDavid S. Miller { 1361c1b4a7e6SDavid S. Miller int tso_segs = tcp_skb_pcount(skb); 1362c1b4a7e6SDavid S. Miller 1363f8269a49SIlpo Järvinen if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { 1364846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1365c1b4a7e6SDavid S. Miller tso_segs = tcp_skb_pcount(skb); 1366c1b4a7e6SDavid S. Miller } 1367c1b4a7e6SDavid S. Miller return tso_segs; 1368c1b4a7e6SDavid S. Miller } 1369c1b4a7e6SDavid S. Miller 137067edfef7SAndi Kleen /* Minshall's variant of the Nagle send check. */ 1371c1b4a7e6SDavid S. Miller static inline int tcp_minshall_check(const struct tcp_sock *tp) 1372c1b4a7e6SDavid S. Miller { 1373c1b4a7e6SDavid S. Miller return after(tp->snd_sml, tp->snd_una) && 1374c1b4a7e6SDavid S. Miller !after(tp->snd_sml, tp->snd_nxt); 1375c1b4a7e6SDavid S. Miller } 1376c1b4a7e6SDavid S. Miller 1377c1b4a7e6SDavid S. Miller /* Return 0, if packet can be sent now without violation Nagle's rules: 1378c1b4a7e6SDavid S. Miller * 1. It is full sized. 1379c1b4a7e6SDavid S. Miller * 2. Or it contains FIN. (already checked by caller) 1380c1b4a7e6SDavid S. Miller * 3. Or TCP_NODELAY was set. 1381c1b4a7e6SDavid S. Miller * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 1382c1b4a7e6SDavid S. Miller * With Minshall's modification: all sent small packets are ACKed. 1383c1b4a7e6SDavid S. Miller */ 1384c1b4a7e6SDavid S. Miller static inline int tcp_nagle_check(const struct tcp_sock *tp, 1385c1b4a7e6SDavid S. Miller const struct sk_buff *skb, 1386c1b4a7e6SDavid S. Miller unsigned mss_now, int nonagle) 1387c1b4a7e6SDavid S. Miller { 1388a02cec21SEric Dumazet return skb->len < mss_now && 1389c1b4a7e6SDavid S. Miller ((nonagle & TCP_NAGLE_CORK) || 1390a02cec21SEric Dumazet (!nonagle && tp->packets_out && tcp_minshall_check(tp))); 1391c1b4a7e6SDavid S. Miller } 1392c1b4a7e6SDavid S. Miller 1393c1b4a7e6SDavid S. Miller /* Return non-zero if the Nagle test allows this packet to be 1394c1b4a7e6SDavid S. Miller * sent now. 1395c1b4a7e6SDavid S. Miller */ 1396c1b4a7e6SDavid S. Miller static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb, 1397c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1398c1b4a7e6SDavid S. Miller { 1399c1b4a7e6SDavid S. Miller /* Nagle rule does not apply to frames, which sit in the middle of the 1400c1b4a7e6SDavid S. Miller * write_queue (they have no chances to get new data). 1401c1b4a7e6SDavid S. Miller * 1402c1b4a7e6SDavid S. Miller * This is implemented in the callers, where they modify the 'nonagle' 1403c1b4a7e6SDavid S. Miller * argument based upon the location of SKB in the send queue. 1404c1b4a7e6SDavid S. Miller */ 1405c1b4a7e6SDavid S. Miller if (nonagle & TCP_NAGLE_PUSH) 1406c1b4a7e6SDavid S. Miller return 1; 1407c1b4a7e6SDavid S. Miller 1408d551e454SIlpo Järvinen /* Don't use the nagle rule for urgent data (or for the final FIN). 1409d551e454SIlpo Järvinen * Nagle can be ignored during F-RTO too (see RFC4138). 1410d551e454SIlpo Järvinen */ 141133f5f57eSIlpo Järvinen if (tcp_urg_mode(tp) || (tp->frto_counter == 2) || 1412a3433f35SChangli Gao (TCP_SKB_CB(skb)->flags & TCPHDR_FIN)) 1413c1b4a7e6SDavid S. Miller return 1; 1414c1b4a7e6SDavid S. Miller 1415c1b4a7e6SDavid S. Miller if (!tcp_nagle_check(tp, skb, cur_mss, nonagle)) 1416c1b4a7e6SDavid S. Miller return 1; 1417c1b4a7e6SDavid S. Miller 1418c1b4a7e6SDavid S. Miller return 0; 1419c1b4a7e6SDavid S. Miller } 1420c1b4a7e6SDavid S. Miller 1421c1b4a7e6SDavid S. Miller /* Does at least the first segment of SKB fit into the send window? */ 1422056834d9SIlpo Järvinen static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb, 1423056834d9SIlpo Järvinen unsigned int cur_mss) 1424c1b4a7e6SDavid S. Miller { 1425c1b4a7e6SDavid S. Miller u32 end_seq = TCP_SKB_CB(skb)->end_seq; 1426c1b4a7e6SDavid S. Miller 1427c1b4a7e6SDavid S. Miller if (skb->len > cur_mss) 1428c1b4a7e6SDavid S. Miller end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 1429c1b4a7e6SDavid S. Miller 143090840defSIlpo Järvinen return !after(end_seq, tcp_wnd_end(tp)); 1431c1b4a7e6SDavid S. Miller } 1432c1b4a7e6SDavid S. Miller 1433fe067e8aSDavid S. Miller /* This checks if the data bearing packet SKB (usually tcp_send_head(sk)) 1434c1b4a7e6SDavid S. Miller * should be put on the wire right now. If so, it returns the number of 1435c1b4a7e6SDavid S. Miller * packets allowed by the congestion window. 1436c1b4a7e6SDavid S. Miller */ 1437c1b4a7e6SDavid S. Miller static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb, 1438c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1439c1b4a7e6SDavid S. Miller { 1440c1b4a7e6SDavid S. Miller struct tcp_sock *tp = tcp_sk(sk); 1441c1b4a7e6SDavid S. Miller unsigned int cwnd_quota; 1442c1b4a7e6SDavid S. Miller 1443846998aeSDavid S. Miller tcp_init_tso_segs(sk, skb, cur_mss); 1444c1b4a7e6SDavid S. Miller 1445c1b4a7e6SDavid S. Miller if (!tcp_nagle_test(tp, skb, cur_mss, nonagle)) 1446c1b4a7e6SDavid S. Miller return 0; 1447c1b4a7e6SDavid S. Miller 1448c1b4a7e6SDavid S. Miller cwnd_quota = tcp_cwnd_test(tp, skb); 1449056834d9SIlpo Järvinen if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss)) 1450c1b4a7e6SDavid S. Miller cwnd_quota = 0; 1451c1b4a7e6SDavid S. Miller 1452c1b4a7e6SDavid S. Miller return cwnd_quota; 1453c1b4a7e6SDavid S. Miller } 1454c1b4a7e6SDavid S. Miller 145567edfef7SAndi Kleen /* Test if sending is allowed right now. */ 14569e412ba7SIlpo Järvinen int tcp_may_send_now(struct sock *sk) 1457c1b4a7e6SDavid S. Miller { 14589e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1459fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 1460c1b4a7e6SDavid S. Miller 1461a02cec21SEric Dumazet return skb && 14620c54b85fSIlpo Järvinen tcp_snd_test(sk, skb, tcp_current_mss(sk), 1463c1b4a7e6SDavid S. Miller (tcp_skb_is_last(sk, skb) ? 1464a02cec21SEric Dumazet tp->nonagle : TCP_NAGLE_PUSH)); 1465c1b4a7e6SDavid S. Miller } 1466c1b4a7e6SDavid S. Miller 1467c1b4a7e6SDavid S. Miller /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 1468c1b4a7e6SDavid S. Miller * which is put after SKB on the list. It is very much like 1469c1b4a7e6SDavid S. Miller * tcp_fragment() except that it may make several kinds of assumptions 1470c1b4a7e6SDavid S. Miller * in order to speed up the splitting operation. In particular, we 1471c1b4a7e6SDavid S. Miller * know that all the data is in scatter-gather pages, and that the 1472c1b4a7e6SDavid S. Miller * packet has never been sent out before (and thus is not cloned). 1473c1b4a7e6SDavid S. Miller */ 1474056834d9SIlpo Järvinen static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, 1475c4ead4c5SEric Dumazet unsigned int mss_now, gfp_t gfp) 1476c1b4a7e6SDavid S. Miller { 1477c1b4a7e6SDavid S. Miller struct sk_buff *buff; 1478c1b4a7e6SDavid S. Miller int nlen = skb->len - len; 14799ce01461SIlpo Järvinen u8 flags; 1480c1b4a7e6SDavid S. Miller 1481c1b4a7e6SDavid S. Miller /* All of a TSO frame must be composed of paged data. */ 1482c8ac3774SHerbert Xu if (skb->len != skb->data_len) 1483c8ac3774SHerbert Xu return tcp_fragment(sk, skb, len, mss_now); 1484c1b4a7e6SDavid S. Miller 1485c4ead4c5SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, gfp); 1486c1b4a7e6SDavid S. Miller if (unlikely(buff == NULL)) 1487c1b4a7e6SDavid S. Miller return -ENOMEM; 1488c1b4a7e6SDavid S. Miller 14893ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 14903ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1491b60b49eaSHerbert Xu buff->truesize += nlen; 1492c1b4a7e6SDavid S. Miller skb->truesize -= nlen; 1493c1b4a7e6SDavid S. Miller 1494c1b4a7e6SDavid S. Miller /* Correct the sequence numbers. */ 1495c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 1496c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 1497c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 1498c1b4a7e6SDavid S. Miller 1499c1b4a7e6SDavid S. Miller /* PSH and FIN should only be set in the second packet. */ 1500c1b4a7e6SDavid S. Miller flags = TCP_SKB_CB(skb)->flags; 1501a3433f35SChangli Gao TCP_SKB_CB(skb)->flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 1502c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->flags = flags; 1503c1b4a7e6SDavid S. Miller 1504c1b4a7e6SDavid S. Miller /* This packet was never sent out yet, so no SACK bits. */ 1505c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->sacked = 0; 1506c1b4a7e6SDavid S. Miller 150784fa7933SPatrick McHardy buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; 1508c1b4a7e6SDavid S. Miller skb_split(skb, buff, len); 1509c1b4a7e6SDavid S. Miller 1510c1b4a7e6SDavid S. Miller /* Fix up tso_factor for both original and new SKB. */ 1511846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1512846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, buff, mss_now); 1513c1b4a7e6SDavid S. Miller 1514c1b4a7e6SDavid S. Miller /* Link BUFF into the send queue. */ 1515c1b4a7e6SDavid S. Miller skb_header_release(buff); 1516fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 1517c1b4a7e6SDavid S. Miller 1518c1b4a7e6SDavid S. Miller return 0; 1519c1b4a7e6SDavid S. Miller } 1520c1b4a7e6SDavid S. Miller 1521c1b4a7e6SDavid S. Miller /* Try to defer sending, if possible, in order to minimize the amount 1522c1b4a7e6SDavid S. Miller * of TSO splitting we do. View it as a kind of TSO Nagle test. 1523c1b4a7e6SDavid S. Miller * 1524c1b4a7e6SDavid S. Miller * This algorithm is from John Heffner. 1525c1b4a7e6SDavid S. Miller */ 15269e412ba7SIlpo Järvinen static int tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb) 1527c1b4a7e6SDavid S. Miller { 15289e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 15296687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 1530c1b4a7e6SDavid S. Miller u32 send_win, cong_win, limit, in_flight; 1531ad9f4f50SEric Dumazet int win_divisor; 1532c1b4a7e6SDavid S. Miller 1533a3433f35SChangli Gao if (TCP_SKB_CB(skb)->flags & TCPHDR_FIN) 1534ae8064acSJohn Heffner goto send_now; 1535c1b4a7e6SDavid S. Miller 15366687e988SArnaldo Carvalho de Melo if (icsk->icsk_ca_state != TCP_CA_Open) 1537ae8064acSJohn Heffner goto send_now; 1538ae8064acSJohn Heffner 1539ae8064acSJohn Heffner /* Defer for less than two clock ticks. */ 1540bd515c3eSIlpo Järvinen if (tp->tso_deferred && 1541a2acde07SIlpo Järvinen (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1) 1542ae8064acSJohn Heffner goto send_now; 1543908a75c1SDavid S. Miller 1544c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1545c1b4a7e6SDavid S. Miller 1546056834d9SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight)); 1547c1b4a7e6SDavid S. Miller 154890840defSIlpo Järvinen send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1549c1b4a7e6SDavid S. Miller 1550c1b4a7e6SDavid S. Miller /* From in_flight test above, we know that cwnd > in_flight. */ 1551c1b4a7e6SDavid S. Miller cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; 1552c1b4a7e6SDavid S. Miller 1553c1b4a7e6SDavid S. Miller limit = min(send_win, cong_win); 1554c1b4a7e6SDavid S. Miller 1555ba244fe9SDavid S. Miller /* If a full-sized TSO skb can be sent, do it. */ 155682cc1a7aSPeter P Waskiewicz Jr if (limit >= sk->sk_gso_max_size) 1557ae8064acSJohn Heffner goto send_now; 1558ba244fe9SDavid S. Miller 155962ad2761SIlpo Järvinen /* Middle in queue won't get any more data, full sendable already? */ 156062ad2761SIlpo Järvinen if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) 156162ad2761SIlpo Järvinen goto send_now; 156262ad2761SIlpo Järvinen 1563ad9f4f50SEric Dumazet win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor); 1564ad9f4f50SEric Dumazet if (win_divisor) { 1565c1b4a7e6SDavid S. Miller u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); 1566c1b4a7e6SDavid S. Miller 1567c1b4a7e6SDavid S. Miller /* If at least some fraction of a window is available, 1568c1b4a7e6SDavid S. Miller * just use it. 1569c1b4a7e6SDavid S. Miller */ 1570ad9f4f50SEric Dumazet chunk /= win_divisor; 1571c1b4a7e6SDavid S. Miller if (limit >= chunk) 1572ae8064acSJohn Heffner goto send_now; 1573c1b4a7e6SDavid S. Miller } else { 1574c1b4a7e6SDavid S. Miller /* Different approach, try not to defer past a single 1575c1b4a7e6SDavid S. Miller * ACK. Receiver should ACK every other full sized 1576c1b4a7e6SDavid S. Miller * frame, so if we have space for more than 3 frames 1577c1b4a7e6SDavid S. Miller * then send now. 1578c1b4a7e6SDavid S. Miller */ 1579c1b4a7e6SDavid S. Miller if (limit > tcp_max_burst(tp) * tp->mss_cache) 1580ae8064acSJohn Heffner goto send_now; 1581c1b4a7e6SDavid S. Miller } 1582c1b4a7e6SDavid S. Miller 1583c1b4a7e6SDavid S. Miller /* Ok, it looks like it is advisable to defer. */ 1584ae8064acSJohn Heffner tp->tso_deferred = 1 | (jiffies << 1); 1585ae8064acSJohn Heffner 1586c1b4a7e6SDavid S. Miller return 1; 1587ae8064acSJohn Heffner 1588ae8064acSJohn Heffner send_now: 1589ae8064acSJohn Heffner tp->tso_deferred = 0; 1590ae8064acSJohn Heffner return 0; 1591c1b4a7e6SDavid S. Miller } 1592c1b4a7e6SDavid S. Miller 15935d424d5aSJohn Heffner /* Create a new MTU probe if we are ready. 159467edfef7SAndi Kleen * MTU probe is regularly attempting to increase the path MTU by 159567edfef7SAndi Kleen * deliberately sending larger packets. This discovers routing 159667edfef7SAndi Kleen * changes resulting in larger path MTUs. 159767edfef7SAndi Kleen * 15985d424d5aSJohn Heffner * Returns 0 if we should wait to probe (no cwnd available), 15995d424d5aSJohn Heffner * 1 if a probe was sent, 1600056834d9SIlpo Järvinen * -1 otherwise 1601056834d9SIlpo Järvinen */ 16025d424d5aSJohn Heffner static int tcp_mtu_probe(struct sock *sk) 16035d424d5aSJohn Heffner { 16045d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 16055d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 16065d424d5aSJohn Heffner struct sk_buff *skb, *nskb, *next; 16075d424d5aSJohn Heffner int len; 16085d424d5aSJohn Heffner int probe_size; 160991cc17c0SIlpo Järvinen int size_needed; 16105d424d5aSJohn Heffner int copy; 16115d424d5aSJohn Heffner int mss_now; 16125d424d5aSJohn Heffner 16135d424d5aSJohn Heffner /* Not currently probing/verifying, 16145d424d5aSJohn Heffner * not in recovery, 16155d424d5aSJohn Heffner * have enough cwnd, and 16165d424d5aSJohn Heffner * not SACKing (the variable headers throw things off) */ 16175d424d5aSJohn Heffner if (!icsk->icsk_mtup.enabled || 16185d424d5aSJohn Heffner icsk->icsk_mtup.probe_size || 16195d424d5aSJohn Heffner inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 16205d424d5aSJohn Heffner tp->snd_cwnd < 11 || 1621cabeccbdSIlpo Järvinen tp->rx_opt.num_sacks || tp->rx_opt.dsack) 16225d424d5aSJohn Heffner return -1; 16235d424d5aSJohn Heffner 16245d424d5aSJohn Heffner /* Very simple search strategy: just double the MSS. */ 16250c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 16265d424d5aSJohn Heffner probe_size = 2 * tp->mss_cache; 162791cc17c0SIlpo Järvinen size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; 16285d424d5aSJohn Heffner if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) { 16295d424d5aSJohn Heffner /* TODO: set timer for probe_converge_event */ 16305d424d5aSJohn Heffner return -1; 16315d424d5aSJohn Heffner } 16325d424d5aSJohn Heffner 16335d424d5aSJohn Heffner /* Have enough data in the send queue to probe? */ 16347f9c33e5SIlpo Järvinen if (tp->write_seq - tp->snd_nxt < size_needed) 16355d424d5aSJohn Heffner return -1; 16365d424d5aSJohn Heffner 163791cc17c0SIlpo Järvinen if (tp->snd_wnd < size_needed) 16385d424d5aSJohn Heffner return -1; 163990840defSIlpo Järvinen if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) 16405d424d5aSJohn Heffner return 0; 16415d424d5aSJohn Heffner 1642d67c58e9SIlpo Järvinen /* Do we need to wait to drain cwnd? With none in flight, don't stall */ 1643d67c58e9SIlpo Järvinen if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { 1644d67c58e9SIlpo Järvinen if (!tcp_packets_in_flight(tp)) 16455d424d5aSJohn Heffner return -1; 16465d424d5aSJohn Heffner else 16475d424d5aSJohn Heffner return 0; 16485d424d5aSJohn Heffner } 16495d424d5aSJohn Heffner 16505d424d5aSJohn Heffner /* We're allowed to probe. Build it now. */ 16515d424d5aSJohn Heffner if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL) 16525d424d5aSJohn Heffner return -1; 16533ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 16543ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 16555d424d5aSJohn Heffner 1656fe067e8aSDavid S. Miller skb = tcp_send_head(sk); 16575d424d5aSJohn Heffner 16585d424d5aSJohn Heffner TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 16595d424d5aSJohn Heffner TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 1660a3433f35SChangli Gao TCP_SKB_CB(nskb)->flags = TCPHDR_ACK; 16615d424d5aSJohn Heffner TCP_SKB_CB(nskb)->sacked = 0; 16625d424d5aSJohn Heffner nskb->csum = 0; 166384fa7933SPatrick McHardy nskb->ip_summed = skb->ip_summed; 16645d424d5aSJohn Heffner 166550c4817eSIlpo Järvinen tcp_insert_write_queue_before(nskb, skb, sk); 166650c4817eSIlpo Järvinen 16675d424d5aSJohn Heffner len = 0; 1668234b6860SIlpo Järvinen tcp_for_write_queue_from_safe(skb, next, sk) { 16695d424d5aSJohn Heffner copy = min_t(int, skb->len, probe_size - len); 16705d424d5aSJohn Heffner if (nskb->ip_summed) 16715d424d5aSJohn Heffner skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); 16725d424d5aSJohn Heffner else 16735d424d5aSJohn Heffner nskb->csum = skb_copy_and_csum_bits(skb, 0, 1674056834d9SIlpo Järvinen skb_put(nskb, copy), 1675056834d9SIlpo Järvinen copy, nskb->csum); 16765d424d5aSJohn Heffner 16775d424d5aSJohn Heffner if (skb->len <= copy) { 16785d424d5aSJohn Heffner /* We've eaten all the data from this skb. 16795d424d5aSJohn Heffner * Throw it away. */ 16805d424d5aSJohn Heffner TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags; 1681fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 16823ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 16835d424d5aSJohn Heffner } else { 16845d424d5aSJohn Heffner TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags & 1685a3433f35SChangli Gao ~(TCPHDR_FIN|TCPHDR_PSH); 16865d424d5aSJohn Heffner if (!skb_shinfo(skb)->nr_frags) { 16875d424d5aSJohn Heffner skb_pull(skb, copy); 168884fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 1689056834d9SIlpo Järvinen skb->csum = csum_partial(skb->data, 1690056834d9SIlpo Järvinen skb->len, 0); 16915d424d5aSJohn Heffner } else { 16925d424d5aSJohn Heffner __pskb_trim_head(skb, copy); 16935d424d5aSJohn Heffner tcp_set_skb_tso_segs(sk, skb, mss_now); 16945d424d5aSJohn Heffner } 16955d424d5aSJohn Heffner TCP_SKB_CB(skb)->seq += copy; 16965d424d5aSJohn Heffner } 16975d424d5aSJohn Heffner 16985d424d5aSJohn Heffner len += copy; 1699234b6860SIlpo Järvinen 1700234b6860SIlpo Järvinen if (len >= probe_size) 1701234b6860SIlpo Järvinen break; 17025d424d5aSJohn Heffner } 17035d424d5aSJohn Heffner tcp_init_tso_segs(sk, nskb, nskb->len); 17045d424d5aSJohn Heffner 17055d424d5aSJohn Heffner /* We're ready to send. If this fails, the probe will 17065d424d5aSJohn Heffner * be resegmented into mss-sized pieces by tcp_write_xmit(). */ 17075d424d5aSJohn Heffner TCP_SKB_CB(nskb)->when = tcp_time_stamp; 17085d424d5aSJohn Heffner if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 17095d424d5aSJohn Heffner /* Decrement cwnd here because we are sending 17105d424d5aSJohn Heffner * effectively two packets. */ 17115d424d5aSJohn Heffner tp->snd_cwnd--; 171266f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, nskb); 17135d424d5aSJohn Heffner 17145d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 17150e7b1368SJohn Heffner tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 17160e7b1368SJohn Heffner tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 17175d424d5aSJohn Heffner 17185d424d5aSJohn Heffner return 1; 17195d424d5aSJohn Heffner } 17205d424d5aSJohn Heffner 17215d424d5aSJohn Heffner return -1; 17225d424d5aSJohn Heffner } 17235d424d5aSJohn Heffner 17241da177e4SLinus Torvalds /* This routine writes packets to the network. It advances the 17251da177e4SLinus Torvalds * send_head. This happens as incoming acks open up the remote 17261da177e4SLinus Torvalds * window for us. 17271da177e4SLinus Torvalds * 1728f8269a49SIlpo Järvinen * LARGESEND note: !tcp_urg_mode is overkill, only frames between 1729f8269a49SIlpo Järvinen * snd_up-64k-mss .. snd_up cannot be large. However, taking into 1730f8269a49SIlpo Järvinen * account rare use of URG, this is not a big flaw. 1731f8269a49SIlpo Järvinen * 17321da177e4SLinus Torvalds * Returns 1, if no segments are in flight and we have queued segments, but 17331da177e4SLinus Torvalds * cannot send anything now because of SWS or another problem. 17341da177e4SLinus Torvalds */ 1735d5dd9175SIlpo Järvinen static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 1736d5dd9175SIlpo Järvinen int push_one, gfp_t gfp) 17371da177e4SLinus Torvalds { 17381da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 173992df7b51SDavid S. Miller struct sk_buff *skb; 1740c1b4a7e6SDavid S. Miller unsigned int tso_segs, sent_pkts; 1741c1b4a7e6SDavid S. Miller int cwnd_quota; 17425d424d5aSJohn Heffner int result; 17431da177e4SLinus Torvalds 1744c1b4a7e6SDavid S. Miller sent_pkts = 0; 17455d424d5aSJohn Heffner 1746d5dd9175SIlpo Järvinen if (!push_one) { 17475d424d5aSJohn Heffner /* Do MTU probing. */ 1748d5dd9175SIlpo Järvinen result = tcp_mtu_probe(sk); 1749d5dd9175SIlpo Järvinen if (!result) { 17505d424d5aSJohn Heffner return 0; 17515d424d5aSJohn Heffner } else if (result > 0) { 17525d424d5aSJohn Heffner sent_pkts = 1; 17535d424d5aSJohn Heffner } 1754d5dd9175SIlpo Järvinen } 17555d424d5aSJohn Heffner 1756fe067e8aSDavid S. Miller while ((skb = tcp_send_head(sk))) { 1757c8ac3774SHerbert Xu unsigned int limit; 1758c8ac3774SHerbert Xu 1759b68e9f85SHerbert Xu tso_segs = tcp_init_tso_segs(sk, skb, mss_now); 1760c1b4a7e6SDavid S. Miller BUG_ON(!tso_segs); 1761c1b4a7e6SDavid S. Miller 1762b68e9f85SHerbert Xu cwnd_quota = tcp_cwnd_test(tp, skb); 1763b68e9f85SHerbert Xu if (!cwnd_quota) 1764b68e9f85SHerbert Xu break; 1765b68e9f85SHerbert Xu 1766b68e9f85SHerbert Xu if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) 1767b68e9f85SHerbert Xu break; 1768b68e9f85SHerbert Xu 1769c1b4a7e6SDavid S. Miller if (tso_segs == 1) { 1770aa93466bSDavid S. Miller if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 1771aa93466bSDavid S. Miller (tcp_skb_is_last(sk, skb) ? 1772aa93466bSDavid S. Miller nonagle : TCP_NAGLE_PUSH)))) 1773aa93466bSDavid S. Miller break; 1774c1b4a7e6SDavid S. Miller } else { 1775d5dd9175SIlpo Järvinen if (!push_one && tcp_tso_should_defer(sk, skb)) 1776aa93466bSDavid S. Miller break; 1777c1b4a7e6SDavid S. Miller } 1778aa93466bSDavid S. Miller 1779c8ac3774SHerbert Xu limit = mss_now; 1780f8269a49SIlpo Järvinen if (tso_segs > 1 && !tcp_urg_mode(tp)) 17810e3a4803SIlpo Järvinen limit = tcp_mss_split_point(sk, skb, mss_now, 17820e3a4803SIlpo Järvinen cwnd_quota); 1783c8ac3774SHerbert Xu 1784c8ac3774SHerbert Xu if (skb->len > limit && 1785c4ead4c5SEric Dumazet unlikely(tso_fragment(sk, skb, limit, mss_now, gfp))) 17861da177e4SLinus Torvalds break; 17871da177e4SLinus Torvalds 17881da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 1789c1b4a7e6SDavid S. Miller 1790d5dd9175SIlpo Järvinen if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) 17911da177e4SLinus Torvalds break; 17921da177e4SLinus Torvalds 17931da177e4SLinus Torvalds /* Advance the send_head. This one is sent out. 17941da177e4SLinus Torvalds * This call will increment packets_out. 17951da177e4SLinus Torvalds */ 179666f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 17971da177e4SLinus Torvalds 17981da177e4SLinus Torvalds tcp_minshall_update(tp, mss_now, skb); 1799aa93466bSDavid S. Miller sent_pkts++; 1800d5dd9175SIlpo Järvinen 1801d5dd9175SIlpo Järvinen if (push_one) 1802d5dd9175SIlpo Järvinen break; 18031da177e4SLinus Torvalds } 18041da177e4SLinus Torvalds 1805aa93466bSDavid S. Miller if (likely(sent_pkts)) { 18069e412ba7SIlpo Järvinen tcp_cwnd_validate(sk); 18071da177e4SLinus Torvalds return 0; 18081da177e4SLinus Torvalds } 1809fe067e8aSDavid S. Miller return !tp->packets_out && tcp_send_head(sk); 18101da177e4SLinus Torvalds } 18111da177e4SLinus Torvalds 1812a762a980SDavid S. Miller /* Push out any pending frames which were held back due to 1813a762a980SDavid S. Miller * TCP_CORK or attempt at coalescing tiny packets. 1814a762a980SDavid S. Miller * The socket must be locked by the caller. 1815a762a980SDavid S. Miller */ 18169e412ba7SIlpo Järvinen void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 18179e412ba7SIlpo Järvinen int nonagle) 1818a762a980SDavid S. Miller { 1819726e07a8SIlpo Järvinen /* If we are closed, the bytes will have to remain here. 1820726e07a8SIlpo Järvinen * In time closedown will finish, we empty the write queue and 1821726e07a8SIlpo Järvinen * all will be happy. 1822726e07a8SIlpo Järvinen */ 1823726e07a8SIlpo Järvinen if (unlikely(sk->sk_state == TCP_CLOSE)) 1824726e07a8SIlpo Järvinen return; 1825726e07a8SIlpo Järvinen 1826d5dd9175SIlpo Järvinen if (tcp_write_xmit(sk, cur_mss, nonagle, 0, GFP_ATOMIC)) 18279e412ba7SIlpo Järvinen tcp_check_probe_timer(sk); 1828a762a980SDavid S. Miller } 1829a762a980SDavid S. Miller 1830c1b4a7e6SDavid S. Miller /* Send _single_ skb sitting at the send head. This function requires 1831c1b4a7e6SDavid S. Miller * true push pending frames to setup probe timer etc. 1832c1b4a7e6SDavid S. Miller */ 1833c1b4a7e6SDavid S. Miller void tcp_push_one(struct sock *sk, unsigned int mss_now) 1834c1b4a7e6SDavid S. Miller { 1835fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 1836c1b4a7e6SDavid S. Miller 1837c1b4a7e6SDavid S. Miller BUG_ON(!skb || skb->len < mss_now); 1838c1b4a7e6SDavid S. Miller 1839d5dd9175SIlpo Järvinen tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); 1840c1b4a7e6SDavid S. Miller } 1841c1b4a7e6SDavid S. Miller 18421da177e4SLinus Torvalds /* This function returns the amount that we can raise the 18431da177e4SLinus Torvalds * usable window based on the following constraints 18441da177e4SLinus Torvalds * 18451da177e4SLinus Torvalds * 1. The window can never be shrunk once it is offered (RFC 793) 18461da177e4SLinus Torvalds * 2. We limit memory per socket 18471da177e4SLinus Torvalds * 18481da177e4SLinus Torvalds * RFC 1122: 18491da177e4SLinus Torvalds * "the suggested [SWS] avoidance algorithm for the receiver is to keep 18501da177e4SLinus Torvalds * RECV.NEXT + RCV.WIN fixed until: 18511da177e4SLinus Torvalds * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 18521da177e4SLinus Torvalds * 18531da177e4SLinus Torvalds * i.e. don't raise the right edge of the window until you can raise 18541da177e4SLinus Torvalds * it at least MSS bytes. 18551da177e4SLinus Torvalds * 18561da177e4SLinus Torvalds * Unfortunately, the recommended algorithm breaks header prediction, 18571da177e4SLinus Torvalds * since header prediction assumes th->window stays fixed. 18581da177e4SLinus Torvalds * 18591da177e4SLinus Torvalds * Strictly speaking, keeping th->window fixed violates the receiver 18601da177e4SLinus Torvalds * side SWS prevention criteria. The problem is that under this rule 18611da177e4SLinus Torvalds * a stream of single byte packets will cause the right side of the 18621da177e4SLinus Torvalds * window to always advance by a single byte. 18631da177e4SLinus Torvalds * 18641da177e4SLinus Torvalds * Of course, if the sender implements sender side SWS prevention 18651da177e4SLinus Torvalds * then this will not be a problem. 18661da177e4SLinus Torvalds * 18671da177e4SLinus Torvalds * BSD seems to make the following compromise: 18681da177e4SLinus Torvalds * 18691da177e4SLinus Torvalds * If the free space is less than the 1/4 of the maximum 18701da177e4SLinus Torvalds * space available and the free space is less than 1/2 mss, 18711da177e4SLinus Torvalds * then set the window to 0. 18721da177e4SLinus Torvalds * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 18731da177e4SLinus Torvalds * Otherwise, just prevent the window from shrinking 18741da177e4SLinus Torvalds * and from being larger than the largest representable value. 18751da177e4SLinus Torvalds * 18761da177e4SLinus Torvalds * This prevents incremental opening of the window in the regime 18771da177e4SLinus Torvalds * where TCP is limited by the speed of the reader side taking 18781da177e4SLinus Torvalds * data out of the TCP receive queue. It does nothing about 18791da177e4SLinus Torvalds * those cases where the window is constrained on the sender side 18801da177e4SLinus Torvalds * because the pipeline is full. 18811da177e4SLinus Torvalds * 18821da177e4SLinus Torvalds * BSD also seems to "accidentally" limit itself to windows that are a 18831da177e4SLinus Torvalds * multiple of MSS, at least until the free space gets quite small. 18841da177e4SLinus Torvalds * This would appear to be a side effect of the mbuf implementation. 18851da177e4SLinus Torvalds * Combining these two algorithms results in the observed behavior 18861da177e4SLinus Torvalds * of having a fixed window size at almost all times. 18871da177e4SLinus Torvalds * 18881da177e4SLinus Torvalds * Below we obtain similar behavior by forcing the offered window to 18891da177e4SLinus Torvalds * a multiple of the mss when it is feasible to do so. 18901da177e4SLinus Torvalds * 18911da177e4SLinus Torvalds * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 18921da177e4SLinus Torvalds * Regular options like TIMESTAMP are taken into account. 18931da177e4SLinus Torvalds */ 18941da177e4SLinus Torvalds u32 __tcp_select_window(struct sock *sk) 18951da177e4SLinus Torvalds { 1896463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 18971da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1898caa20d9aSStephen Hemminger /* MSS for the peer's data. Previous versions used mss_clamp 18991da177e4SLinus Torvalds * here. I don't know if the value based on our guesses 19001da177e4SLinus Torvalds * of peer's MSS is better for the performance. It's more correct 19011da177e4SLinus Torvalds * but may be worse for the performance because of rcv_mss 19021da177e4SLinus Torvalds * fluctuations. --SAW 1998/11/1 19031da177e4SLinus Torvalds */ 1904463c84b9SArnaldo Carvalho de Melo int mss = icsk->icsk_ack.rcv_mss; 19051da177e4SLinus Torvalds int free_space = tcp_space(sk); 19061da177e4SLinus Torvalds int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk)); 19071da177e4SLinus Torvalds int window; 19081da177e4SLinus Torvalds 19091da177e4SLinus Torvalds if (mss > full_space) 19101da177e4SLinus Torvalds mss = full_space; 19111da177e4SLinus Torvalds 1912b92edbe0SEric Dumazet if (free_space < (full_space >> 1)) { 1913463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.quick = 0; 19141da177e4SLinus Torvalds 19151da177e4SLinus Torvalds if (tcp_memory_pressure) 1916056834d9SIlpo Järvinen tp->rcv_ssthresh = min(tp->rcv_ssthresh, 1917056834d9SIlpo Järvinen 4U * tp->advmss); 19181da177e4SLinus Torvalds 19191da177e4SLinus Torvalds if (free_space < mss) 19201da177e4SLinus Torvalds return 0; 19211da177e4SLinus Torvalds } 19221da177e4SLinus Torvalds 19231da177e4SLinus Torvalds if (free_space > tp->rcv_ssthresh) 19241da177e4SLinus Torvalds free_space = tp->rcv_ssthresh; 19251da177e4SLinus Torvalds 19261da177e4SLinus Torvalds /* Don't do rounding if we are using window scaling, since the 19271da177e4SLinus Torvalds * scaled window will not line up with the MSS boundary anyway. 19281da177e4SLinus Torvalds */ 19291da177e4SLinus Torvalds window = tp->rcv_wnd; 19301da177e4SLinus Torvalds if (tp->rx_opt.rcv_wscale) { 19311da177e4SLinus Torvalds window = free_space; 19321da177e4SLinus Torvalds 19331da177e4SLinus Torvalds /* Advertise enough space so that it won't get scaled away. 19341da177e4SLinus Torvalds * Import case: prevent zero window announcement if 19351da177e4SLinus Torvalds * 1<<rcv_wscale > mss. 19361da177e4SLinus Torvalds */ 19371da177e4SLinus Torvalds if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window) 19381da177e4SLinus Torvalds window = (((window >> tp->rx_opt.rcv_wscale) + 1) 19391da177e4SLinus Torvalds << tp->rx_opt.rcv_wscale); 19401da177e4SLinus Torvalds } else { 19411da177e4SLinus Torvalds /* Get the largest window that is a nice multiple of mss. 19421da177e4SLinus Torvalds * Window clamp already applied above. 19431da177e4SLinus Torvalds * If our current window offering is within 1 mss of the 19441da177e4SLinus Torvalds * free space we just keep it. This prevents the divide 19451da177e4SLinus Torvalds * and multiply from happening most of the time. 19461da177e4SLinus Torvalds * We also don't do any window rounding when the free space 19471da177e4SLinus Torvalds * is too small. 19481da177e4SLinus Torvalds */ 19491da177e4SLinus Torvalds if (window <= free_space - mss || window > free_space) 19501da177e4SLinus Torvalds window = (free_space / mss) * mss; 195184565070SJohn Heffner else if (mss == full_space && 1952b92edbe0SEric Dumazet free_space > window + (full_space >> 1)) 195384565070SJohn Heffner window = free_space; 19541da177e4SLinus Torvalds } 19551da177e4SLinus Torvalds 19561da177e4SLinus Torvalds return window; 19571da177e4SLinus Torvalds } 19581da177e4SLinus Torvalds 19594a17fc3aSIlpo Järvinen /* Collapses two adjacent SKB's during retransmission. */ 19604a17fc3aSIlpo Järvinen static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) 19611da177e4SLinus Torvalds { 19621da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1963fe067e8aSDavid S. Miller struct sk_buff *next_skb = tcp_write_queue_next(sk, skb); 1964058dc334SIlpo Järvinen int skb_size, next_skb_size; 19651da177e4SLinus Torvalds 1966058dc334SIlpo Järvinen skb_size = skb->len; 1967058dc334SIlpo Järvinen next_skb_size = next_skb->len; 19681da177e4SLinus Torvalds 1969058dc334SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); 19701da177e4SLinus Torvalds 19716859d494SIlpo Järvinen tcp_highest_sack_combine(sk, next_skb, skb); 1972a6963a6bSIlpo Järvinen 1973fe067e8aSDavid S. Miller tcp_unlink_write_queue(next_skb, sk); 19741da177e4SLinus Torvalds 1975058dc334SIlpo Järvinen skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size), 19761a4e2d09SArnaldo Carvalho de Melo next_skb_size); 19771da177e4SLinus Torvalds 197852d570aaSJarek Poplawski if (next_skb->ip_summed == CHECKSUM_PARTIAL) 197952d570aaSJarek Poplawski skb->ip_summed = CHECKSUM_PARTIAL; 19801da177e4SLinus Torvalds 198184fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 19821da177e4SLinus Torvalds skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); 19831da177e4SLinus Torvalds 19841da177e4SLinus Torvalds /* Update sequence range on original skb. */ 19851da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 19861da177e4SLinus Torvalds 1987e6c7d085SIlpo Järvinen /* Merge over control information. This moves PSH/FIN etc. over */ 1988e6c7d085SIlpo Järvinen TCP_SKB_CB(skb)->flags |= TCP_SKB_CB(next_skb)->flags; 19891da177e4SLinus Torvalds 19901da177e4SLinus Torvalds /* All done, get rid of second SKB and account for it so 19911da177e4SLinus Torvalds * packet counting does not break. 19921da177e4SLinus Torvalds */ 19934828e7f4SIlpo Järvinen TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; 1994b7689205SIlpo Järvinen 1995b7689205SIlpo Järvinen /* changed transmit queue under us so clear hints */ 1996ef9da47cSIlpo Järvinen tcp_clear_retrans_hints_partial(tp); 1997ef9da47cSIlpo Järvinen if (next_skb == tp->retransmit_skb_hint) 1998ef9da47cSIlpo Järvinen tp->retransmit_skb_hint = skb; 1999b7689205SIlpo Järvinen 2000797108d1SIlpo Järvinen tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); 2001797108d1SIlpo Järvinen 20023ab224beSHideo Aoki sk_wmem_free_skb(sk, next_skb); 20031da177e4SLinus Torvalds } 20041da177e4SLinus Torvalds 200567edfef7SAndi Kleen /* Check if coalescing SKBs is legal. */ 20064a17fc3aSIlpo Järvinen static int tcp_can_collapse(struct sock *sk, struct sk_buff *skb) 20074a17fc3aSIlpo Järvinen { 20084a17fc3aSIlpo Järvinen if (tcp_skb_pcount(skb) > 1) 20094a17fc3aSIlpo Järvinen return 0; 20104a17fc3aSIlpo Järvinen /* TODO: SACK collapsing could be used to remove this condition */ 20114a17fc3aSIlpo Järvinen if (skb_shinfo(skb)->nr_frags != 0) 20124a17fc3aSIlpo Järvinen return 0; 20134a17fc3aSIlpo Järvinen if (skb_cloned(skb)) 20144a17fc3aSIlpo Järvinen return 0; 20154a17fc3aSIlpo Järvinen if (skb == tcp_send_head(sk)) 20164a17fc3aSIlpo Järvinen return 0; 20174a17fc3aSIlpo Järvinen /* Some heurestics for collapsing over SACK'd could be invented */ 20184a17fc3aSIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 20194a17fc3aSIlpo Järvinen return 0; 20204a17fc3aSIlpo Järvinen 20214a17fc3aSIlpo Järvinen return 1; 20224a17fc3aSIlpo Järvinen } 20234a17fc3aSIlpo Järvinen 202467edfef7SAndi Kleen /* Collapse packets in the retransmit queue to make to create 202567edfef7SAndi Kleen * less packets on the wire. This is only done on retransmission. 202667edfef7SAndi Kleen */ 20274a17fc3aSIlpo Järvinen static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, 20284a17fc3aSIlpo Järvinen int space) 20294a17fc3aSIlpo Järvinen { 20304a17fc3aSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 20314a17fc3aSIlpo Järvinen struct sk_buff *skb = to, *tmp; 20324a17fc3aSIlpo Järvinen int first = 1; 20334a17fc3aSIlpo Järvinen 20344a17fc3aSIlpo Järvinen if (!sysctl_tcp_retrans_collapse) 20354a17fc3aSIlpo Järvinen return; 2036a3433f35SChangli Gao if (TCP_SKB_CB(skb)->flags & TCPHDR_SYN) 20374a17fc3aSIlpo Järvinen return; 20384a17fc3aSIlpo Järvinen 20394a17fc3aSIlpo Järvinen tcp_for_write_queue_from_safe(skb, tmp, sk) { 20404a17fc3aSIlpo Järvinen if (!tcp_can_collapse(sk, skb)) 20414a17fc3aSIlpo Järvinen break; 20424a17fc3aSIlpo Järvinen 20434a17fc3aSIlpo Järvinen space -= skb->len; 20444a17fc3aSIlpo Järvinen 20454a17fc3aSIlpo Järvinen if (first) { 20464a17fc3aSIlpo Järvinen first = 0; 20474a17fc3aSIlpo Järvinen continue; 20484a17fc3aSIlpo Järvinen } 20494a17fc3aSIlpo Järvinen 20504a17fc3aSIlpo Järvinen if (space < 0) 20514a17fc3aSIlpo Järvinen break; 20524a17fc3aSIlpo Järvinen /* Punt if not enough space exists in the first SKB for 20534a17fc3aSIlpo Järvinen * the data in the second 20544a17fc3aSIlpo Järvinen */ 20554a17fc3aSIlpo Järvinen if (skb->len > skb_tailroom(to)) 20564a17fc3aSIlpo Järvinen break; 20574a17fc3aSIlpo Järvinen 20584a17fc3aSIlpo Järvinen if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) 20594a17fc3aSIlpo Järvinen break; 20604a17fc3aSIlpo Järvinen 20614a17fc3aSIlpo Järvinen tcp_collapse_retrans(sk, to); 20624a17fc3aSIlpo Järvinen } 20634a17fc3aSIlpo Järvinen } 20644a17fc3aSIlpo Järvinen 20651da177e4SLinus Torvalds /* This retransmits one SKB. Policy decisions and retransmit queue 20661da177e4SLinus Torvalds * state updates are done by the caller. Returns non-zero if an 20671da177e4SLinus Torvalds * error occurred which prevented the send. 20681da177e4SLinus Torvalds */ 20691da177e4SLinus Torvalds int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) 20701da177e4SLinus Torvalds { 20711da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 20725d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 20737d227cd2SSridhar Samudrala unsigned int cur_mss; 20741da177e4SLinus Torvalds int err; 20751da177e4SLinus Torvalds 20765d424d5aSJohn Heffner /* Inconslusive MTU probe */ 20775d424d5aSJohn Heffner if (icsk->icsk_mtup.probe_size) { 20785d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 20795d424d5aSJohn Heffner } 20805d424d5aSJohn Heffner 20811da177e4SLinus Torvalds /* Do not sent more than we queued. 1/4 is reserved for possible 2082caa20d9aSStephen Hemminger * copying overhead: fragmentation, tunneling, mangling etc. 20831da177e4SLinus Torvalds */ 20841da177e4SLinus Torvalds if (atomic_read(&sk->sk_wmem_alloc) > 20851da177e4SLinus Torvalds min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf)) 20861da177e4SLinus Torvalds return -EAGAIN; 20871da177e4SLinus Torvalds 20881da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 20891da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 20901da177e4SLinus Torvalds BUG(); 20911da177e4SLinus Torvalds if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 20921da177e4SLinus Torvalds return -ENOMEM; 20931da177e4SLinus Torvalds } 20941da177e4SLinus Torvalds 20957d227cd2SSridhar Samudrala if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 20967d227cd2SSridhar Samudrala return -EHOSTUNREACH; /* Routing failure or similar. */ 20977d227cd2SSridhar Samudrala 20980c54b85fSIlpo Järvinen cur_mss = tcp_current_mss(sk); 20997d227cd2SSridhar Samudrala 21001da177e4SLinus Torvalds /* If receiver has shrunk his window, and skb is out of 21011da177e4SLinus Torvalds * new window, do not retransmit it. The exception is the 21021da177e4SLinus Torvalds * case, when window is shrunk to zero. In this case 21031da177e4SLinus Torvalds * our retransmit serves as a zero window probe. 21041da177e4SLinus Torvalds */ 21059d4fb27dSJoe Perches if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && 21069d4fb27dSJoe Perches TCP_SKB_CB(skb)->seq != tp->snd_una) 21071da177e4SLinus Torvalds return -EAGAIN; 21081da177e4SLinus Torvalds 21091da177e4SLinus Torvalds if (skb->len > cur_mss) { 2110846998aeSDavid S. Miller if (tcp_fragment(sk, skb, cur_mss, cur_mss)) 21111da177e4SLinus Torvalds return -ENOMEM; /* We'll try again later. */ 211202276f3cSIlpo Järvinen } else { 21139eb9362eSIlpo Järvinen int oldpcount = tcp_skb_pcount(skb); 21149eb9362eSIlpo Järvinen 21159eb9362eSIlpo Järvinen if (unlikely(oldpcount > 1)) { 211602276f3cSIlpo Järvinen tcp_init_tso_segs(sk, skb, cur_mss); 21179eb9362eSIlpo Järvinen tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb)); 21189eb9362eSIlpo Järvinen } 21191da177e4SLinus Torvalds } 21201da177e4SLinus Torvalds 21211da177e4SLinus Torvalds tcp_retrans_try_collapse(sk, skb, cur_mss); 21221da177e4SLinus Torvalds 21231da177e4SLinus Torvalds /* Some Solaris stacks overoptimize and ignore the FIN on a 21241da177e4SLinus Torvalds * retransmit when old data is attached. So strip it off 21251da177e4SLinus Torvalds * since it is cheap to do so and saves bytes on the network. 21261da177e4SLinus Torvalds */ 21271da177e4SLinus Torvalds if (skb->len > 0 && 2128a3433f35SChangli Gao (TCP_SKB_CB(skb)->flags & TCPHDR_FIN) && 21291da177e4SLinus Torvalds tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) { 21301da177e4SLinus Torvalds if (!pskb_trim(skb, 0)) { 2131e870a8efSIlpo Järvinen /* Reuse, even though it does some unnecessary work */ 2132e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1, 2133e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->flags); 21341da177e4SLinus Torvalds skb->ip_summed = CHECKSUM_NONE; 21351da177e4SLinus Torvalds } 21361da177e4SLinus Torvalds } 21371da177e4SLinus Torvalds 21381da177e4SLinus Torvalds /* Make a copy, if the first transmission SKB clone we made 21391da177e4SLinus Torvalds * is still in somebody's hands, else make a clone. 21401da177e4SLinus Torvalds */ 21411da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 21421da177e4SLinus Torvalds 2143dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 21441da177e4SLinus Torvalds 21451da177e4SLinus Torvalds if (err == 0) { 21461da177e4SLinus Torvalds /* Update global TCP statistics. */ 214781cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); 21481da177e4SLinus Torvalds 21491da177e4SLinus Torvalds tp->total_retrans++; 21501da177e4SLinus Torvalds 21511da177e4SLinus Torvalds #if FASTRETRANS_DEBUG > 0 21521da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 21531da177e4SLinus Torvalds if (net_ratelimit()) 21541da177e4SLinus Torvalds printk(KERN_DEBUG "retrans_out leaked.\n"); 21551da177e4SLinus Torvalds } 21561da177e4SLinus Torvalds #endif 2157b08d6cb2SIlpo Järvinen if (!tp->retrans_out) 2158b08d6cb2SIlpo Järvinen tp->lost_retrans_low = tp->snd_nxt; 21591da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 21601da177e4SLinus Torvalds tp->retrans_out += tcp_skb_pcount(skb); 21611da177e4SLinus Torvalds 21621da177e4SLinus Torvalds /* Save stamp of the first retransmit. */ 21631da177e4SLinus Torvalds if (!tp->retrans_stamp) 21641da177e4SLinus Torvalds tp->retrans_stamp = TCP_SKB_CB(skb)->when; 21651da177e4SLinus Torvalds 2166c24f691bSYuchung Cheng tp->undo_retrans += tcp_skb_pcount(skb); 21671da177e4SLinus Torvalds 21681da177e4SLinus Torvalds /* snd_nxt is stored to detect loss of retransmitted segment, 21691da177e4SLinus Torvalds * see tcp_input.c tcp_sacktag_write_queue(). 21701da177e4SLinus Torvalds */ 21711da177e4SLinus Torvalds TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt; 21721da177e4SLinus Torvalds } 21731da177e4SLinus Torvalds return err; 21741da177e4SLinus Torvalds } 21751da177e4SLinus Torvalds 217667edfef7SAndi Kleen /* Check if we forward retransmits are possible in the current 217767edfef7SAndi Kleen * window/congestion state. 217867edfef7SAndi Kleen */ 2179b5afe7bcSIlpo Järvinen static int tcp_can_forward_retransmit(struct sock *sk) 2180b5afe7bcSIlpo Järvinen { 2181b5afe7bcSIlpo Järvinen const struct inet_connection_sock *icsk = inet_csk(sk); 2182b5afe7bcSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 2183b5afe7bcSIlpo Järvinen 2184b5afe7bcSIlpo Järvinen /* Forward retransmissions are possible only during Recovery. */ 2185b5afe7bcSIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Recovery) 2186b5afe7bcSIlpo Järvinen return 0; 2187b5afe7bcSIlpo Järvinen 2188b5afe7bcSIlpo Järvinen /* No forward retransmissions in Reno are possible. */ 2189b5afe7bcSIlpo Järvinen if (tcp_is_reno(tp)) 2190b5afe7bcSIlpo Järvinen return 0; 2191b5afe7bcSIlpo Järvinen 2192b5afe7bcSIlpo Järvinen /* Yeah, we have to make difficult choice between forward transmission 2193b5afe7bcSIlpo Järvinen * and retransmission... Both ways have their merits... 2194b5afe7bcSIlpo Järvinen * 2195b5afe7bcSIlpo Järvinen * For now we do not retransmit anything, while we have some new 2196b5afe7bcSIlpo Järvinen * segments to send. In the other cases, follow rule 3 for 2197b5afe7bcSIlpo Järvinen * NextSeg() specified in RFC3517. 2198b5afe7bcSIlpo Järvinen */ 2199b5afe7bcSIlpo Järvinen 2200b5afe7bcSIlpo Järvinen if (tcp_may_send_now(sk)) 2201b5afe7bcSIlpo Järvinen return 0; 2202b5afe7bcSIlpo Järvinen 2203b5afe7bcSIlpo Järvinen return 1; 2204b5afe7bcSIlpo Järvinen } 2205b5afe7bcSIlpo Järvinen 22061da177e4SLinus Torvalds /* This gets called after a retransmit timeout, and the initially 22071da177e4SLinus Torvalds * retransmitted data is acknowledged. It tries to continue 22081da177e4SLinus Torvalds * resending the rest of the retransmit queue, until either 22091da177e4SLinus Torvalds * we've sent it all or the congestion window limit is reached. 22101da177e4SLinus Torvalds * If doing SACK, the first ACK which comes back for a timeout 22111da177e4SLinus Torvalds * based retransmit packet might feed us FACK information again. 22121da177e4SLinus Torvalds * If so, we use it to avoid unnecessarily retransmissions. 22131da177e4SLinus Torvalds */ 22141da177e4SLinus Torvalds void tcp_xmit_retransmit_queue(struct sock *sk) 22151da177e4SLinus Torvalds { 22166687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 22171da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 22181da177e4SLinus Torvalds struct sk_buff *skb; 22190e1c54c2SIlpo Järvinen struct sk_buff *hole = NULL; 2220618d9f25SIlpo Järvinen u32 last_lost; 222161eb55f4SIlpo Järvinen int mib_idx; 22220e1c54c2SIlpo Järvinen int fwd_rexmitting = 0; 22236a438bbeSStephen Hemminger 222445e77d31SIlpo Järvinen if (!tp->packets_out) 222545e77d31SIlpo Järvinen return; 222645e77d31SIlpo Järvinen 222708ebd172SIlpo Järvinen if (!tp->lost_out) 222808ebd172SIlpo Järvinen tp->retransmit_high = tp->snd_una; 222908ebd172SIlpo Järvinen 2230618d9f25SIlpo Järvinen if (tp->retransmit_skb_hint) { 22316a438bbeSStephen Hemminger skb = tp->retransmit_skb_hint; 2232618d9f25SIlpo Järvinen last_lost = TCP_SKB_CB(skb)->end_seq; 2233618d9f25SIlpo Järvinen if (after(last_lost, tp->retransmit_high)) 2234618d9f25SIlpo Järvinen last_lost = tp->retransmit_high; 2235618d9f25SIlpo Järvinen } else { 2236fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 2237618d9f25SIlpo Järvinen last_lost = tp->snd_una; 2238618d9f25SIlpo Järvinen } 22391da177e4SLinus Torvalds 2240fe067e8aSDavid S. Miller tcp_for_write_queue_from(skb, sk) { 22411da177e4SLinus Torvalds __u8 sacked = TCP_SKB_CB(skb)->sacked; 22421da177e4SLinus Torvalds 2243fe067e8aSDavid S. Miller if (skb == tcp_send_head(sk)) 2244fe067e8aSDavid S. Miller break; 22456a438bbeSStephen Hemminger /* we could do better than to assign each time */ 22460e1c54c2SIlpo Järvinen if (hole == NULL) 22476a438bbeSStephen Hemminger tp->retransmit_skb_hint = skb; 22486a438bbeSStephen Hemminger 22491da177e4SLinus Torvalds /* Assume this retransmit will generate 22501da177e4SLinus Torvalds * only one packet for congestion window 22511da177e4SLinus Torvalds * calculation purposes. This works because 22521da177e4SLinus Torvalds * tcp_retransmit_skb() will chop up the 22531da177e4SLinus Torvalds * packet to be MSS sized and all the 22541da177e4SLinus Torvalds * packet counting works out. 22551da177e4SLinus Torvalds */ 22561da177e4SLinus Torvalds if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) 22571da177e4SLinus Torvalds return; 22580e1c54c2SIlpo Järvinen 22590e1c54c2SIlpo Järvinen if (fwd_rexmitting) { 22600e1c54c2SIlpo Järvinen begin_fwd: 22610e1c54c2SIlpo Järvinen if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp))) 2262006f582cSIlpo Järvinen break; 22630e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFORWARDRETRANS; 22640e1c54c2SIlpo Järvinen 22650e1c54c2SIlpo Järvinen } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) { 2266618d9f25SIlpo Järvinen tp->retransmit_high = last_lost; 22670e1c54c2SIlpo Järvinen if (!tcp_can_forward_retransmit(sk)) 22680e1c54c2SIlpo Järvinen break; 22690e1c54c2SIlpo Järvinen /* Backtrack if necessary to non-L'ed skb */ 22700e1c54c2SIlpo Järvinen if (hole != NULL) { 22710e1c54c2SIlpo Järvinen skb = hole; 22720e1c54c2SIlpo Järvinen hole = NULL; 22730e1c54c2SIlpo Järvinen } 22740e1c54c2SIlpo Järvinen fwd_rexmitting = 1; 22750e1c54c2SIlpo Järvinen goto begin_fwd; 22760e1c54c2SIlpo Järvinen 22770e1c54c2SIlpo Järvinen } else if (!(sacked & TCPCB_LOST)) { 2278ac11ba75SIlpo Järvinen if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) 22790e1c54c2SIlpo Järvinen hole = skb; 228061eb55f4SIlpo Järvinen continue; 22811da177e4SLinus Torvalds 22820e1c54c2SIlpo Järvinen } else { 2283618d9f25SIlpo Järvinen last_lost = TCP_SKB_CB(skb)->end_seq; 22840e1c54c2SIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Loss) 22850e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFASTRETRANS; 22860e1c54c2SIlpo Järvinen else 22870e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; 22880e1c54c2SIlpo Järvinen } 22890e1c54c2SIlpo Järvinen 22900e1c54c2SIlpo Järvinen if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) 229161eb55f4SIlpo Järvinen continue; 229240b215e5SPavel Emelyanov 2293f0ceb0edSIlpo Järvinen if (tcp_retransmit_skb(sk, skb)) 22941da177e4SLinus Torvalds return; 2295de0744afSPavel Emelyanov NET_INC_STATS_BH(sock_net(sk), mib_idx); 22961da177e4SLinus Torvalds 2297fe067e8aSDavid S. Miller if (skb == tcp_write_queue_head(sk)) 2298463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 22993f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, 23003f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 23011da177e4SLinus Torvalds } 23021da177e4SLinus Torvalds } 23031da177e4SLinus Torvalds 23041da177e4SLinus Torvalds /* Send a fin. The caller locks the socket for us. This cannot be 23051da177e4SLinus Torvalds * allowed to fail queueing a FIN frame under any circumstances. 23061da177e4SLinus Torvalds */ 23071da177e4SLinus Torvalds void tcp_send_fin(struct sock *sk) 23081da177e4SLinus Torvalds { 23091da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2310fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_write_queue_tail(sk); 23111da177e4SLinus Torvalds int mss_now; 23121da177e4SLinus Torvalds 23131da177e4SLinus Torvalds /* Optimization, tack on the FIN if we have a queue of 23141da177e4SLinus Torvalds * unsent frames. But be careful about outgoing SACKS 23151da177e4SLinus Torvalds * and IP options. 23161da177e4SLinus Torvalds */ 23170c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 23181da177e4SLinus Torvalds 2319fe067e8aSDavid S. Miller if (tcp_send_head(sk) != NULL) { 2320a3433f35SChangli Gao TCP_SKB_CB(skb)->flags |= TCPHDR_FIN; 23211da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq++; 23221da177e4SLinus Torvalds tp->write_seq++; 23231da177e4SLinus Torvalds } else { 23241da177e4SLinus Torvalds /* Socket is locked, keep trying until memory is available. */ 23251da177e4SLinus Torvalds for (;;) { 2326aa133076SWu Fengguang skb = alloc_skb_fclone(MAX_TCP_HEADER, 2327aa133076SWu Fengguang sk->sk_allocation); 23281da177e4SLinus Torvalds if (skb) 23291da177e4SLinus Torvalds break; 23301da177e4SLinus Torvalds yield(); 23311da177e4SLinus Torvalds } 23321da177e4SLinus Torvalds 23331da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 23341da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 23351da177e4SLinus Torvalds /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 2336e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tp->write_seq, 2337a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_FIN); 23381da177e4SLinus Torvalds tcp_queue_skb(sk, skb); 23391da177e4SLinus Torvalds } 23409e412ba7SIlpo Järvinen __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF); 23411da177e4SLinus Torvalds } 23421da177e4SLinus Torvalds 23431da177e4SLinus Torvalds /* We get here when a process closes a file descriptor (either due to 23441da177e4SLinus Torvalds * an explicit close() or as a byproduct of exit()'ing) and there 23451da177e4SLinus Torvalds * was unread data in the receive queue. This behavior is recommended 234665bb723cSGerrit Renker * by RFC 2525, section 2.17. -DaveM 23471da177e4SLinus Torvalds */ 2348dd0fc66fSAl Viro void tcp_send_active_reset(struct sock *sk, gfp_t priority) 23491da177e4SLinus Torvalds { 23501da177e4SLinus Torvalds struct sk_buff *skb; 23511da177e4SLinus Torvalds 23521da177e4SLinus Torvalds /* NOTE: No TCP options attached and we never retransmit this. */ 23531da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, priority); 23541da177e4SLinus Torvalds if (!skb) { 23554e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 23561da177e4SLinus Torvalds return; 23571da177e4SLinus Torvalds } 23581da177e4SLinus Torvalds 23591da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 23601da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 2361e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), 2362a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_RST); 23631da177e4SLinus Torvalds /* Send it off. */ 23641da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2365dfb4b9dcSDavid S. Miller if (tcp_transmit_skb(sk, skb, 0, priority)) 23664e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 236726af65cbSSridhar Samudrala 236881cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); 23691da177e4SLinus Torvalds } 23701da177e4SLinus Torvalds 237167edfef7SAndi Kleen /* Send a crossed SYN-ACK during socket establishment. 237267edfef7SAndi Kleen * WARNING: This routine must only be called when we have already sent 23731da177e4SLinus Torvalds * a SYN packet that crossed the incoming SYN that caused this routine 23741da177e4SLinus Torvalds * to get called. If this assumption fails then the initial rcv_wnd 23751da177e4SLinus Torvalds * and rcv_wscale values will not be correct. 23761da177e4SLinus Torvalds */ 23771da177e4SLinus Torvalds int tcp_send_synack(struct sock *sk) 23781da177e4SLinus Torvalds { 23791da177e4SLinus Torvalds struct sk_buff *skb; 23801da177e4SLinus Torvalds 2381fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 2382a3433f35SChangli Gao if (skb == NULL || !(TCP_SKB_CB(skb)->flags & TCPHDR_SYN)) { 23831da177e4SLinus Torvalds printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n"); 23841da177e4SLinus Torvalds return -EFAULT; 23851da177e4SLinus Torvalds } 2386a3433f35SChangli Gao if (!(TCP_SKB_CB(skb)->flags & TCPHDR_ACK)) { 23871da177e4SLinus Torvalds if (skb_cloned(skb)) { 23881da177e4SLinus Torvalds struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC); 23891da177e4SLinus Torvalds if (nskb == NULL) 23901da177e4SLinus Torvalds return -ENOMEM; 2391fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 23921da177e4SLinus Torvalds skb_header_release(nskb); 2393fe067e8aSDavid S. Miller __tcp_add_write_queue_head(sk, nskb); 23943ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 23953ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 23963ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 23971da177e4SLinus Torvalds skb = nskb; 23981da177e4SLinus Torvalds } 23991da177e4SLinus Torvalds 2400a3433f35SChangli Gao TCP_SKB_CB(skb)->flags |= TCPHDR_ACK; 24011da177e4SLinus Torvalds TCP_ECN_send_synack(tcp_sk(sk), skb); 24021da177e4SLinus Torvalds } 24031da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2404dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 24051da177e4SLinus Torvalds } 24061da177e4SLinus Torvalds 240767edfef7SAndi Kleen /* Prepare a SYN-ACK. */ 24081da177e4SLinus Torvalds struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst, 2409e6b4d113SWilliam Allen Simpson struct request_sock *req, 2410e6b4d113SWilliam Allen Simpson struct request_values *rvp) 24111da177e4SLinus Torvalds { 2412bd0388aeSWilliam Allen Simpson struct tcp_out_options opts; 24134957faadSWilliam Allen Simpson struct tcp_extend_values *xvp = tcp_xv(rvp); 24142e6599cbSArnaldo Carvalho de Melo struct inet_request_sock *ireq = inet_rsk(req); 24151da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 241628b2774aSEric Dumazet const struct tcp_cookie_values *cvp = tp->cookie_values; 24171da177e4SLinus Torvalds struct tcphdr *th; 24181da177e4SLinus Torvalds struct sk_buff *skb; 2419cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 2420bd0388aeSWilliam Allen Simpson int tcp_header_size; 2421f5fff5dcSTom Quetchenbach int mss; 242228b2774aSEric Dumazet int s_data_desired = 0; 24231da177e4SLinus Torvalds 242428b2774aSEric Dumazet if (cvp != NULL && cvp->s_data_constant && cvp->s_data_desired) 242528b2774aSEric Dumazet s_data_desired = cvp->s_data_desired; 242628b2774aSEric Dumazet skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15 + s_data_desired, 1, GFP_ATOMIC); 24271da177e4SLinus Torvalds if (skb == NULL) 24281da177e4SLinus Torvalds return NULL; 24291da177e4SLinus Torvalds 24301da177e4SLinus Torvalds /* Reserve space for headers. */ 24311da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 24321da177e4SLinus Torvalds 2433adf30907SEric Dumazet skb_dst_set(skb, dst_clone(dst)); 24341da177e4SLinus Torvalds 24350dbaee3bSDavid S. Miller mss = dst_metric_advmss(dst); 2436f5fff5dcSTom Quetchenbach if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss) 2437f5fff5dcSTom Quetchenbach mss = tp->rx_opt.user_mss; 2438f5fff5dcSTom Quetchenbach 243933ad798cSAdam Langley if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */ 244033ad798cSAdam Langley __u8 rcv_wscale; 244133ad798cSAdam Langley /* Set this up on the first call only */ 244233ad798cSAdam Langley req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW); 2443e88c64f0SHagen Paul Pfeifer 2444e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 2445e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 2446e88c64f0SHagen Paul Pfeifer (req->window_clamp > tcp_full_space(sk) || req->window_clamp == 0)) 2447e88c64f0SHagen Paul Pfeifer req->window_clamp = tcp_full_space(sk); 2448e88c64f0SHagen Paul Pfeifer 244933ad798cSAdam Langley /* tcp_full_space because it is guaranteed to be the first packet */ 245033ad798cSAdam Langley tcp_select_initial_window(tcp_full_space(sk), 2451f5fff5dcSTom Quetchenbach mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0), 245233ad798cSAdam Langley &req->rcv_wnd, 245333ad798cSAdam Langley &req->window_clamp, 245433ad798cSAdam Langley ireq->wscale_ok, 245531d12926Slaurent chavey &rcv_wscale, 245631d12926Slaurent chavey dst_metric(dst, RTAX_INITRWND)); 245733ad798cSAdam Langley ireq->rcv_wscale = rcv_wscale; 245833ad798cSAdam Langley } 2459cfb6eeb4SYOSHIFUJI Hideaki 246033ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 24618b5f12d0SFlorian Westphal #ifdef CONFIG_SYN_COOKIES 24628b5f12d0SFlorian Westphal if (unlikely(req->cookie_ts)) 24638b5f12d0SFlorian Westphal TCP_SKB_CB(skb)->when = cookie_init_timestamp(req); 24648b5f12d0SFlorian Westphal else 24658b5f12d0SFlorian Westphal #endif 246633ad798cSAdam Langley TCP_SKB_CB(skb)->when = tcp_time_stamp; 2467f5fff5dcSTom Quetchenbach tcp_header_size = tcp_synack_options(sk, req, mss, 24684957faadSWilliam Allen Simpson skb, &opts, &md5, xvp) 24694957faadSWilliam Allen Simpson + sizeof(*th); 247033ad798cSAdam Langley 2471aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 2472aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 24731da177e4SLinus Torvalds 2474aa8223c7SArnaldo Carvalho de Melo th = tcp_hdr(skb); 24751da177e4SLinus Torvalds memset(th, 0, sizeof(struct tcphdr)); 24761da177e4SLinus Torvalds th->syn = 1; 24771da177e4SLinus Torvalds th->ack = 1; 24781da177e4SLinus Torvalds TCP_ECN_make_synack(req, th); 2479a3116ac5SKOVACS Krisztian th->source = ireq->loc_port; 24802e6599cbSArnaldo Carvalho de Melo th->dest = ireq->rmt_port; 2481e870a8efSIlpo Järvinen /* Setting of flags are superfluous here for callers (and ECE is 2482e870a8efSIlpo Järvinen * not even correctly set) 2483e870a8efSIlpo Järvinen */ 2484e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn, 2485a3433f35SChangli Gao TCPHDR_SYN | TCPHDR_ACK); 24864957faadSWilliam Allen Simpson 24874957faadSWilliam Allen Simpson if (OPTION_COOKIE_EXTENSION & opts.options) { 248828b2774aSEric Dumazet if (s_data_desired) { 248928b2774aSEric Dumazet u8 *buf = skb_put(skb, s_data_desired); 24904957faadSWilliam Allen Simpson 24914957faadSWilliam Allen Simpson /* copy data directly from the listening socket. */ 249228b2774aSEric Dumazet memcpy(buf, cvp->s_data_payload, s_data_desired); 249328b2774aSEric Dumazet TCP_SKB_CB(skb)->end_seq += s_data_desired; 24944957faadSWilliam Allen Simpson } 24954957faadSWilliam Allen Simpson 24964957faadSWilliam Allen Simpson if (opts.hash_size > 0) { 24974957faadSWilliam Allen Simpson __u32 workspace[SHA_WORKSPACE_WORDS]; 24984957faadSWilliam Allen Simpson u32 *mess = &xvp->cookie_bakery[COOKIE_DIGEST_WORDS]; 24994957faadSWilliam Allen Simpson u32 *tail = &mess[COOKIE_MESSAGE_WORDS-1]; 25004957faadSWilliam Allen Simpson 25014957faadSWilliam Allen Simpson /* Secret recipe depends on the Timestamp, (future) 25024957faadSWilliam Allen Simpson * Sequence and Acknowledgment Numbers, Initiator 25034957faadSWilliam Allen Simpson * Cookie, and others handled by IP variant caller. 25044957faadSWilliam Allen Simpson */ 25054957faadSWilliam Allen Simpson *tail-- ^= opts.tsval; 25064957faadSWilliam Allen Simpson *tail-- ^= tcp_rsk(req)->rcv_isn + 1; 25074957faadSWilliam Allen Simpson *tail-- ^= TCP_SKB_CB(skb)->seq + 1; 25084957faadSWilliam Allen Simpson 25094957faadSWilliam Allen Simpson /* recommended */ 25100eae88f3SEric Dumazet *tail-- ^= (((__force u32)th->dest << 16) | (__force u32)th->source); 2511f9a2e69eSDavid S. Miller *tail-- ^= (u32)(unsigned long)cvp; /* per sockopt */ 25124957faadSWilliam Allen Simpson 25134957faadSWilliam Allen Simpson sha_transform((__u32 *)&xvp->cookie_bakery[0], 25144957faadSWilliam Allen Simpson (char *)mess, 25154957faadSWilliam Allen Simpson &workspace[0]); 25164957faadSWilliam Allen Simpson opts.hash_location = 25174957faadSWilliam Allen Simpson (__u8 *)&xvp->cookie_bakery[0]; 25184957faadSWilliam Allen Simpson } 25194957faadSWilliam Allen Simpson } 25204957faadSWilliam Allen Simpson 25211da177e4SLinus Torvalds th->seq = htonl(TCP_SKB_CB(skb)->seq); 25222e6599cbSArnaldo Carvalho de Melo th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1); 25231da177e4SLinus Torvalds 25241da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 2525600ff0c2SIlpo Järvinen th->window = htons(min(req->rcv_wnd, 65535U)); 2526bd0388aeSWilliam Allen Simpson tcp_options_write((__be32 *)(th + 1), tp, &opts); 25271da177e4SLinus Torvalds th->doff = (tcp_header_size >> 2); 2528aa2ea058STom Herbert TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, tcp_skb_pcount(skb)); 2529cfb6eeb4SYOSHIFUJI Hideaki 2530cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 2531cfb6eeb4SYOSHIFUJI Hideaki /* Okay, we have all we need - do the md5 hash if needed */ 2532cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 2533bd0388aeSWilliam Allen Simpson tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, 253449a72dfbSAdam Langley md5, NULL, req, skb); 2535cfb6eeb4SYOSHIFUJI Hideaki } 2536cfb6eeb4SYOSHIFUJI Hideaki #endif 2537cfb6eeb4SYOSHIFUJI Hideaki 25381da177e4SLinus Torvalds return skb; 25391da177e4SLinus Torvalds } 25404bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_make_synack); 25411da177e4SLinus Torvalds 254267edfef7SAndi Kleen /* Do all connect socket setups that can be done AF independent. */ 254340efc6faSStephen Hemminger static void tcp_connect_init(struct sock *sk) 25441da177e4SLinus Torvalds { 25451da177e4SLinus Torvalds struct dst_entry *dst = __sk_dst_get(sk); 25461da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 25471da177e4SLinus Torvalds __u8 rcv_wscale; 25481da177e4SLinus Torvalds 25491da177e4SLinus Torvalds /* We'll fix this up when we get a response from the other end. 25501da177e4SLinus Torvalds * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 25511da177e4SLinus Torvalds */ 25521da177e4SLinus Torvalds tp->tcp_header_len = sizeof(struct tcphdr) + 2553bb5b7c11SDavid S. Miller (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0); 25541da177e4SLinus Torvalds 2555cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 2556cfb6eeb4SYOSHIFUJI Hideaki if (tp->af_specific->md5_lookup(sk, sk) != NULL) 2557cfb6eeb4SYOSHIFUJI Hideaki tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 2558cfb6eeb4SYOSHIFUJI Hideaki #endif 2559cfb6eeb4SYOSHIFUJI Hideaki 25601da177e4SLinus Torvalds /* If user gave his TCP_MAXSEG, record it to clamp */ 25611da177e4SLinus Torvalds if (tp->rx_opt.user_mss) 25621da177e4SLinus Torvalds tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 25631da177e4SLinus Torvalds tp->max_window = 0; 25645d424d5aSJohn Heffner tcp_mtup_init(sk); 25651da177e4SLinus Torvalds tcp_sync_mss(sk, dst_mtu(dst)); 25661da177e4SLinus Torvalds 25671da177e4SLinus Torvalds if (!tp->window_clamp) 25681da177e4SLinus Torvalds tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 25690dbaee3bSDavid S. Miller tp->advmss = dst_metric_advmss(dst); 2570f5fff5dcSTom Quetchenbach if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss) 2571f5fff5dcSTom Quetchenbach tp->advmss = tp->rx_opt.user_mss; 2572f5fff5dcSTom Quetchenbach 25731da177e4SLinus Torvalds tcp_initialize_rcv_mss(sk); 25741da177e4SLinus Torvalds 2575e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 2576e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 2577e88c64f0SHagen Paul Pfeifer (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) 2578e88c64f0SHagen Paul Pfeifer tp->window_clamp = tcp_full_space(sk); 2579e88c64f0SHagen Paul Pfeifer 25801da177e4SLinus Torvalds tcp_select_initial_window(tcp_full_space(sk), 25811da177e4SLinus Torvalds tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 25821da177e4SLinus Torvalds &tp->rcv_wnd, 25831da177e4SLinus Torvalds &tp->window_clamp, 2584bb5b7c11SDavid S. Miller sysctl_tcp_window_scaling, 258531d12926Slaurent chavey &rcv_wscale, 258631d12926Slaurent chavey dst_metric(dst, RTAX_INITRWND)); 25871da177e4SLinus Torvalds 25881da177e4SLinus Torvalds tp->rx_opt.rcv_wscale = rcv_wscale; 25891da177e4SLinus Torvalds tp->rcv_ssthresh = tp->rcv_wnd; 25901da177e4SLinus Torvalds 25911da177e4SLinus Torvalds sk->sk_err = 0; 25921da177e4SLinus Torvalds sock_reset_flag(sk, SOCK_DONE); 25931da177e4SLinus Torvalds tp->snd_wnd = 0; 2594ee7537b6SHantzis Fotis tcp_init_wl(tp, 0); 25951da177e4SLinus Torvalds tp->snd_una = tp->write_seq; 25961da177e4SLinus Torvalds tp->snd_sml = tp->write_seq; 259733f5f57eSIlpo Järvinen tp->snd_up = tp->write_seq; 25981da177e4SLinus Torvalds tp->rcv_nxt = 0; 25991da177e4SLinus Torvalds tp->rcv_wup = 0; 26001da177e4SLinus Torvalds tp->copied_seq = 0; 26011da177e4SLinus Torvalds 2602463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; 2603463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_retransmits = 0; 26041da177e4SLinus Torvalds tcp_clear_retrans(tp); 26051da177e4SLinus Torvalds } 26061da177e4SLinus Torvalds 260767edfef7SAndi Kleen /* Build a SYN and send it off. */ 26081da177e4SLinus Torvalds int tcp_connect(struct sock *sk) 26091da177e4SLinus Torvalds { 26101da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 26111da177e4SLinus Torvalds struct sk_buff *buff; 2612ee586811SEric Paris int err; 26131da177e4SLinus Torvalds 26141da177e4SLinus Torvalds tcp_connect_init(sk); 26151da177e4SLinus Torvalds 2616d179cd12SDavid S. Miller buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation); 26171da177e4SLinus Torvalds if (unlikely(buff == NULL)) 26181da177e4SLinus Torvalds return -ENOBUFS; 26191da177e4SLinus Torvalds 26201da177e4SLinus Torvalds /* Reserve space for headers. */ 26211da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 26221da177e4SLinus Torvalds 2623bd37a088SWei Yongjun tp->snd_nxt = tp->write_seq; 2624a3433f35SChangli Gao tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); 2625e870a8efSIlpo Järvinen TCP_ECN_send_syn(sk, buff); 26261da177e4SLinus Torvalds 26271da177e4SLinus Torvalds /* Send it off. */ 26281da177e4SLinus Torvalds TCP_SKB_CB(buff)->when = tcp_time_stamp; 26291da177e4SLinus Torvalds tp->retrans_stamp = TCP_SKB_CB(buff)->when; 26301da177e4SLinus Torvalds skb_header_release(buff); 2631fe067e8aSDavid S. Miller __tcp_add_write_queue_tail(sk, buff); 26323ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 26333ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 26341da177e4SLinus Torvalds tp->packets_out += tcp_skb_pcount(buff); 2635ee586811SEric Paris err = tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); 2636ee586811SEric Paris if (err == -ECONNREFUSED) 2637ee586811SEric Paris return err; 2638bd37a088SWei Yongjun 2639bd37a088SWei Yongjun /* We change tp->snd_nxt after the tcp_transmit_skb() call 2640bd37a088SWei Yongjun * in order to make this packet get counted in tcpOutSegs. 2641bd37a088SWei Yongjun */ 2642bd37a088SWei Yongjun tp->snd_nxt = tp->write_seq; 2643bd37a088SWei Yongjun tp->pushed_seq = tp->write_seq; 264481cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); 26451da177e4SLinus Torvalds 26461da177e4SLinus Torvalds /* Timer for repeating the SYN until an answer. */ 26473f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 26483f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 26491da177e4SLinus Torvalds return 0; 26501da177e4SLinus Torvalds } 26514bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_connect); 26521da177e4SLinus Torvalds 26531da177e4SLinus Torvalds /* Send out a delayed ack, the caller does the policy checking 26541da177e4SLinus Torvalds * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 26551da177e4SLinus Torvalds * for details. 26561da177e4SLinus Torvalds */ 26571da177e4SLinus Torvalds void tcp_send_delayed_ack(struct sock *sk) 26581da177e4SLinus Torvalds { 2659463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 2660463c84b9SArnaldo Carvalho de Melo int ato = icsk->icsk_ack.ato; 26611da177e4SLinus Torvalds unsigned long timeout; 26621da177e4SLinus Torvalds 26631da177e4SLinus Torvalds if (ato > TCP_DELACK_MIN) { 2664463c84b9SArnaldo Carvalho de Melo const struct tcp_sock *tp = tcp_sk(sk); 26651da177e4SLinus Torvalds int max_ato = HZ / 2; 26661da177e4SLinus Torvalds 2667056834d9SIlpo Järvinen if (icsk->icsk_ack.pingpong || 2668056834d9SIlpo Järvinen (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 26691da177e4SLinus Torvalds max_ato = TCP_DELACK_MAX; 26701da177e4SLinus Torvalds 26711da177e4SLinus Torvalds /* Slow path, intersegment interval is "high". */ 26721da177e4SLinus Torvalds 26731da177e4SLinus Torvalds /* If some rtt estimate is known, use it to bound delayed ack. 2674463c84b9SArnaldo Carvalho de Melo * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 26751da177e4SLinus Torvalds * directly. 26761da177e4SLinus Torvalds */ 26771da177e4SLinus Torvalds if (tp->srtt) { 26781da177e4SLinus Torvalds int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN); 26791da177e4SLinus Torvalds 26801da177e4SLinus Torvalds if (rtt < max_ato) 26811da177e4SLinus Torvalds max_ato = rtt; 26821da177e4SLinus Torvalds } 26831da177e4SLinus Torvalds 26841da177e4SLinus Torvalds ato = min(ato, max_ato); 26851da177e4SLinus Torvalds } 26861da177e4SLinus Torvalds 26871da177e4SLinus Torvalds /* Stay within the limit we were given */ 26881da177e4SLinus Torvalds timeout = jiffies + ato; 26891da177e4SLinus Torvalds 26901da177e4SLinus Torvalds /* Use new timeout only if there wasn't a older one earlier. */ 2691463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 26921da177e4SLinus Torvalds /* If delack timer was blocked or is about to expire, 26931da177e4SLinus Torvalds * send ACK now. 26941da177e4SLinus Torvalds */ 2695463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.blocked || 2696463c84b9SArnaldo Carvalho de Melo time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 26971da177e4SLinus Torvalds tcp_send_ack(sk); 26981da177e4SLinus Torvalds return; 26991da177e4SLinus Torvalds } 27001da177e4SLinus Torvalds 2701463c84b9SArnaldo Carvalho de Melo if (!time_before(timeout, icsk->icsk_ack.timeout)) 2702463c84b9SArnaldo Carvalho de Melo timeout = icsk->icsk_ack.timeout; 27031da177e4SLinus Torvalds } 2704463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 2705463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.timeout = timeout; 2706463c84b9SArnaldo Carvalho de Melo sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 27071da177e4SLinus Torvalds } 27081da177e4SLinus Torvalds 27091da177e4SLinus Torvalds /* This routine sends an ack and also updates the window. */ 27101da177e4SLinus Torvalds void tcp_send_ack(struct sock *sk) 27111da177e4SLinus Torvalds { 27121da177e4SLinus Torvalds struct sk_buff *buff; 27131da177e4SLinus Torvalds 2714058dc334SIlpo Järvinen /* If we have been reset, we may not send again. */ 2715058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 2716058dc334SIlpo Järvinen return; 2717058dc334SIlpo Järvinen 27181da177e4SLinus Torvalds /* We are not putting this on the write queue, so 27191da177e4SLinus Torvalds * tcp_transmit_skb() will set the ownership to this 27201da177e4SLinus Torvalds * sock. 27211da177e4SLinus Torvalds */ 27221da177e4SLinus Torvalds buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 27231da177e4SLinus Torvalds if (buff == NULL) { 2724463c84b9SArnaldo Carvalho de Melo inet_csk_schedule_ack(sk); 2725463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; 27263f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 27273f421baaSArnaldo Carvalho de Melo TCP_DELACK_MAX, TCP_RTO_MAX); 27281da177e4SLinus Torvalds return; 27291da177e4SLinus Torvalds } 27301da177e4SLinus Torvalds 27311da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 27321da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 2733a3433f35SChangli Gao tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); 27341da177e4SLinus Torvalds 27351da177e4SLinus Torvalds /* Send it off, this clears delayed acks for us. */ 27361da177e4SLinus Torvalds TCP_SKB_CB(buff)->when = tcp_time_stamp; 2737dfb4b9dcSDavid S. Miller tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC); 27381da177e4SLinus Torvalds } 27391da177e4SLinus Torvalds 27401da177e4SLinus Torvalds /* This routine sends a packet with an out of date sequence 27411da177e4SLinus Torvalds * number. It assumes the other end will try to ack it. 27421da177e4SLinus Torvalds * 27431da177e4SLinus Torvalds * Question: what should we make while urgent mode? 27441da177e4SLinus Torvalds * 4.4BSD forces sending single byte of data. We cannot send 27451da177e4SLinus Torvalds * out of window data, because we have SND.NXT==SND.MAX... 27461da177e4SLinus Torvalds * 27471da177e4SLinus Torvalds * Current solution: to send TWO zero-length segments in urgent mode: 27481da177e4SLinus Torvalds * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 27491da177e4SLinus Torvalds * out-of-date with SND.UNA-1 to probe window. 27501da177e4SLinus Torvalds */ 27511da177e4SLinus Torvalds static int tcp_xmit_probe_skb(struct sock *sk, int urgent) 27521da177e4SLinus Torvalds { 27531da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 27541da177e4SLinus Torvalds struct sk_buff *skb; 27551da177e4SLinus Torvalds 27561da177e4SLinus Torvalds /* We don't queue it, tcp_transmit_skb() sets ownership. */ 27571da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 27581da177e4SLinus Torvalds if (skb == NULL) 27591da177e4SLinus Torvalds return -1; 27601da177e4SLinus Torvalds 27611da177e4SLinus Torvalds /* Reserve space for headers and set control bits. */ 27621da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 27631da177e4SLinus Torvalds /* Use a previous sequence. This should cause the other 27641da177e4SLinus Torvalds * end to send an ack. Don't queue or clone SKB, just 27651da177e4SLinus Torvalds * send it. 27661da177e4SLinus Torvalds */ 2767a3433f35SChangli Gao tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); 27681da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2769dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC); 27701da177e4SLinus Torvalds } 27711da177e4SLinus Torvalds 277267edfef7SAndi Kleen /* Initiate keepalive or window probe from timer. */ 27731da177e4SLinus Torvalds int tcp_write_wakeup(struct sock *sk) 27741da177e4SLinus Torvalds { 27751da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 27761da177e4SLinus Torvalds struct sk_buff *skb; 27771da177e4SLinus Torvalds 2778058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 2779058dc334SIlpo Järvinen return -1; 2780058dc334SIlpo Järvinen 2781fe067e8aSDavid S. Miller if ((skb = tcp_send_head(sk)) != NULL && 278290840defSIlpo Järvinen before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { 27831da177e4SLinus Torvalds int err; 27840c54b85fSIlpo Järvinen unsigned int mss = tcp_current_mss(sk); 278590840defSIlpo Järvinen unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 27861da177e4SLinus Torvalds 27871da177e4SLinus Torvalds if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 27881da177e4SLinus Torvalds tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 27891da177e4SLinus Torvalds 27901da177e4SLinus Torvalds /* We are probing the opening of a window 27911da177e4SLinus Torvalds * but the window size is != 0 27921da177e4SLinus Torvalds * must have been a result SWS avoidance ( sender ) 27931da177e4SLinus Torvalds */ 27941da177e4SLinus Torvalds if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 27951da177e4SLinus Torvalds skb->len > mss) { 27961da177e4SLinus Torvalds seg_size = min(seg_size, mss); 2797a3433f35SChangli Gao TCP_SKB_CB(skb)->flags |= TCPHDR_PSH; 2798846998aeSDavid S. Miller if (tcp_fragment(sk, skb, seg_size, mss)) 27991da177e4SLinus Torvalds return -1; 28001da177e4SLinus Torvalds } else if (!tcp_skb_pcount(skb)) 2801846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss); 28021da177e4SLinus Torvalds 2803a3433f35SChangli Gao TCP_SKB_CB(skb)->flags |= TCPHDR_PSH; 28041da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2805dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 280666f5fe62SIlpo Järvinen if (!err) 280766f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 28081da177e4SLinus Torvalds return err; 28091da177e4SLinus Torvalds } else { 281033f5f57eSIlpo Järvinen if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) 28114828e7f4SIlpo Järvinen tcp_xmit_probe_skb(sk, 1); 28121da177e4SLinus Torvalds return tcp_xmit_probe_skb(sk, 0); 28131da177e4SLinus Torvalds } 28141da177e4SLinus Torvalds } 28151da177e4SLinus Torvalds 28161da177e4SLinus Torvalds /* A window probe timeout has occurred. If window is not closed send 28171da177e4SLinus Torvalds * a partial packet else a zero probe. 28181da177e4SLinus Torvalds */ 28191da177e4SLinus Torvalds void tcp_send_probe0(struct sock *sk) 28201da177e4SLinus Torvalds { 2821463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 28221da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 28231da177e4SLinus Torvalds int err; 28241da177e4SLinus Torvalds 28251da177e4SLinus Torvalds err = tcp_write_wakeup(sk); 28261da177e4SLinus Torvalds 2827fe067e8aSDavid S. Miller if (tp->packets_out || !tcp_send_head(sk)) { 28281da177e4SLinus Torvalds /* Cancel probe timer, if it is not required. */ 28296687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 0; 2830463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff = 0; 28311da177e4SLinus Torvalds return; 28321da177e4SLinus Torvalds } 28331da177e4SLinus Torvalds 28341da177e4SLinus Torvalds if (err <= 0) { 2835463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_backoff < sysctl_tcp_retries2) 2836463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff++; 28376687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out++; 2838463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 28393f421baaSArnaldo Carvalho de Melo min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), 28403f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 28411da177e4SLinus Torvalds } else { 28421da177e4SLinus Torvalds /* If packet was not sent due to local congestion, 28436687e988SArnaldo Carvalho de Melo * do not backoff and do not remember icsk_probes_out. 28441da177e4SLinus Torvalds * Let local senders to fight for local resources. 28451da177e4SLinus Torvalds * 28461da177e4SLinus Torvalds * Use accumulated backoff yet. 28471da177e4SLinus Torvalds */ 28486687e988SArnaldo Carvalho de Melo if (!icsk->icsk_probes_out) 28496687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 1; 2850463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 2851463c84b9SArnaldo Carvalho de Melo min(icsk->icsk_rto << icsk->icsk_backoff, 28523f421baaSArnaldo Carvalho de Melo TCP_RESOURCE_PROBE_INTERVAL), 28533f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 28541da177e4SLinus Torvalds } 28551da177e4SLinus Torvalds } 2856