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. */ 68cf533ea5SEric Dumazet static void tcp_event_new_data_sent(struct sock *sk, const 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 7625985edcSLucas De Marchi /* Don't override Nagle indefinitely 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 */ 92cf533ea5SEric Dumazet static inline __u32 tcp_acceptable_seq(const struct sock *sk) 931da177e4SLinus Torvalds { 94cf533ea5SEric Dumazet const 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); 119cf533ea5SEric Dumazet const 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. */ 136cf533ea5SEric Dumazet static void tcp_cwnd_restart(struct sock *sk, const 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, 157cf533ea5SEric Dumazet 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 */ 298cf533ea5SEric Dumazet static inline void TCP_ECN_send_synack(const struct tcp_sock *tp, struct sk_buff *skb) 299bdf1ee5dSIlpo Järvinen { 3004de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR; 301bdf1ee5dSIlpo Järvinen if (!(tp->ecn_flags & TCP_ECN_OK)) 3024de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_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) { 3124de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR; 313bdf1ee5dSIlpo Järvinen tp->ecn_flags = TCP_ECN_OK; 314bdf1ee5dSIlpo Järvinen } 315bdf1ee5dSIlpo Järvinen } 316bdf1ee5dSIlpo Järvinen 317bdf1ee5dSIlpo Järvinen static __inline__ void 318cf533ea5SEric Dumazet TCP_ECN_make_synack(const 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 3594de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_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 */ 56695c96174SEric Dumazet static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb, 56733ad798cSAdam Langley struct tcp_out_options *opts, 568cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 569cf533ea5SEric Dumazet { 57033ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 571bd0388aeSWilliam Allen Simpson struct tcp_cookie_values *cvp = tp->cookie_values; 57295c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 573bd0388aeSWilliam Allen Simpson u8 cookie_size = (!tp->rx_opt.cookie_out_never && cvp != NULL) ? 574bd0388aeSWilliam Allen Simpson tcp_cookie_size_check(cvp->cookie_desired) : 575bd0388aeSWilliam Allen Simpson 0; 57633ad798cSAdam Langley 577cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 57833ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 57933ad798cSAdam Langley if (*md5) { 58033ad798cSAdam Langley opts->options |= OPTION_MD5; 581bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 582cfb6eeb4SYOSHIFUJI Hideaki } 58333ad798cSAdam Langley #else 58433ad798cSAdam Langley *md5 = NULL; 585cfb6eeb4SYOSHIFUJI Hideaki #endif 58633ad798cSAdam Langley 58733ad798cSAdam Langley /* We always get an MSS option. The option bytes which will be seen in 58833ad798cSAdam Langley * normal data packets should timestamps be used, must be in the MSS 58933ad798cSAdam Langley * advertised. But we subtract them from tp->mss_cache so that 59033ad798cSAdam Langley * calculations in tcp_sendmsg are simpler etc. So account for this 59133ad798cSAdam Langley * fact here if necessary. If we don't do this correctly, as a 59233ad798cSAdam Langley * receiver we won't recognize data packets as being full sized when we 59333ad798cSAdam Langley * should, and thus we won't abide by the delayed ACK rules correctly. 59433ad798cSAdam Langley * SACKs don't matter, we never delay an ACK when we have any of those 59533ad798cSAdam Langley * going out. */ 59633ad798cSAdam Langley opts->mss = tcp_advertise_mss(sk); 597bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 59833ad798cSAdam Langley 599bb5b7c11SDavid S. Miller if (likely(sysctl_tcp_timestamps && *md5 == NULL)) { 60033ad798cSAdam Langley opts->options |= OPTION_TS; 60133ad798cSAdam Langley opts->tsval = TCP_SKB_CB(skb)->when; 60233ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 603bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 60433ad798cSAdam Langley } 605bb5b7c11SDavid S. Miller if (likely(sysctl_tcp_window_scaling)) { 60633ad798cSAdam Langley opts->ws = tp->rx_opt.rcv_wscale; 60789e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 608bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 60933ad798cSAdam Langley } 610bb5b7c11SDavid S. Miller if (likely(sysctl_tcp_sack)) { 61133ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 612b32d1310SDavid S. Miller if (unlikely(!(OPTION_TS & opts->options))) 613bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 61433ad798cSAdam Langley } 61533ad798cSAdam Langley 616bd0388aeSWilliam Allen Simpson /* Note that timestamps are required by the specification. 617bd0388aeSWilliam Allen Simpson * 618bd0388aeSWilliam Allen Simpson * Odd numbers of bytes are prohibited by the specification, ensuring 619bd0388aeSWilliam Allen Simpson * that the cookie is 16-bit aligned, and the resulting cookie pair is 620bd0388aeSWilliam Allen Simpson * 32-bit aligned. 621bd0388aeSWilliam Allen Simpson */ 622bd0388aeSWilliam Allen Simpson if (*md5 == NULL && 623bd0388aeSWilliam Allen Simpson (OPTION_TS & opts->options) && 624bd0388aeSWilliam Allen Simpson cookie_size > 0) { 625bd0388aeSWilliam Allen Simpson int need = TCPOLEN_COOKIE_BASE + cookie_size; 626bd0388aeSWilliam Allen Simpson 627bd0388aeSWilliam Allen Simpson if (0x2 & need) { 628bd0388aeSWilliam Allen Simpson /* 32-bit multiple */ 629bd0388aeSWilliam Allen Simpson need += 2; /* NOPs */ 630bd0388aeSWilliam Allen Simpson 631bd0388aeSWilliam Allen Simpson if (need > remaining) { 632bd0388aeSWilliam Allen Simpson /* try shrinking cookie to fit */ 633bd0388aeSWilliam Allen Simpson cookie_size -= 2; 634bd0388aeSWilliam Allen Simpson need -= 4; 635bd0388aeSWilliam Allen Simpson } 636bd0388aeSWilliam Allen Simpson } 637bd0388aeSWilliam Allen Simpson while (need > remaining && TCP_COOKIE_MIN <= cookie_size) { 638bd0388aeSWilliam Allen Simpson cookie_size -= 4; 639bd0388aeSWilliam Allen Simpson need -= 4; 640bd0388aeSWilliam Allen Simpson } 641bd0388aeSWilliam Allen Simpson if (TCP_COOKIE_MIN <= cookie_size) { 642bd0388aeSWilliam Allen Simpson opts->options |= OPTION_COOKIE_EXTENSION; 643bd0388aeSWilliam Allen Simpson opts->hash_location = (__u8 *)&cvp->cookie_pair[0]; 644bd0388aeSWilliam Allen Simpson opts->hash_size = cookie_size; 645bd0388aeSWilliam Allen Simpson 646bd0388aeSWilliam Allen Simpson /* Remember for future incarnations. */ 647bd0388aeSWilliam Allen Simpson cvp->cookie_desired = cookie_size; 648bd0388aeSWilliam Allen Simpson 649bd0388aeSWilliam Allen Simpson if (cvp->cookie_desired != cvp->cookie_pair_size) { 650bd0388aeSWilliam Allen Simpson /* Currently use random bytes as a nonce, 651bd0388aeSWilliam Allen Simpson * assuming these are completely unpredictable 652bd0388aeSWilliam Allen Simpson * by hostile users of the same system. 653bd0388aeSWilliam Allen Simpson */ 654bd0388aeSWilliam Allen Simpson get_random_bytes(&cvp->cookie_pair[0], 655bd0388aeSWilliam Allen Simpson cookie_size); 656bd0388aeSWilliam Allen Simpson cvp->cookie_pair_size = cookie_size; 657bd0388aeSWilliam Allen Simpson } 658bd0388aeSWilliam Allen Simpson 659bd0388aeSWilliam Allen Simpson remaining -= need; 660bd0388aeSWilliam Allen Simpson } 661bd0388aeSWilliam Allen Simpson } 662bd0388aeSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 66333ad798cSAdam Langley } 66433ad798cSAdam Langley 66567edfef7SAndi Kleen /* Set up TCP options for SYN-ACKs. */ 66695c96174SEric Dumazet static unsigned int tcp_synack_options(struct sock *sk, 66733ad798cSAdam Langley struct request_sock *req, 66895c96174SEric Dumazet unsigned int mss, struct sk_buff *skb, 66933ad798cSAdam Langley struct tcp_out_options *opts, 6704957faadSWilliam Allen Simpson struct tcp_md5sig_key **md5, 6714957faadSWilliam Allen Simpson struct tcp_extend_values *xvp) 6724957faadSWilliam Allen Simpson { 67333ad798cSAdam Langley struct inet_request_sock *ireq = inet_rsk(req); 67495c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 6754957faadSWilliam Allen Simpson u8 cookie_plus = (xvp != NULL && !xvp->cookie_out_never) ? 6764957faadSWilliam Allen Simpson xvp->cookie_plus : 6774957faadSWilliam Allen Simpson 0; 67833ad798cSAdam Langley 67933ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 68033ad798cSAdam Langley *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req); 68133ad798cSAdam Langley if (*md5) { 68233ad798cSAdam Langley opts->options |= OPTION_MD5; 6834957faadSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 6844957faadSWilliam Allen Simpson 6854957faadSWilliam Allen Simpson /* We can't fit any SACK blocks in a packet with MD5 + TS 6864957faadSWilliam Allen Simpson * options. There was discussion about disabling SACK 6874957faadSWilliam Allen Simpson * rather than TS in order to fit in better with old, 6884957faadSWilliam Allen Simpson * buggy kernels, but that was deemed to be unnecessary. 6894957faadSWilliam Allen Simpson */ 690de213e5eSEric Dumazet ireq->tstamp_ok &= !ireq->sack_ok; 69133ad798cSAdam Langley } 69233ad798cSAdam Langley #else 69333ad798cSAdam Langley *md5 = NULL; 69433ad798cSAdam Langley #endif 69533ad798cSAdam Langley 6964957faadSWilliam Allen Simpson /* We always send an MSS option. */ 69733ad798cSAdam Langley opts->mss = mss; 6984957faadSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 69933ad798cSAdam Langley 70033ad798cSAdam Langley if (likely(ireq->wscale_ok)) { 70133ad798cSAdam Langley opts->ws = ireq->rcv_wscale; 70289e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 7034957faadSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 70433ad798cSAdam Langley } 705de213e5eSEric Dumazet if (likely(ireq->tstamp_ok)) { 70633ad798cSAdam Langley opts->options |= OPTION_TS; 70733ad798cSAdam Langley opts->tsval = TCP_SKB_CB(skb)->when; 70833ad798cSAdam Langley opts->tsecr = req->ts_recent; 7094957faadSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 71033ad798cSAdam Langley } 71133ad798cSAdam Langley if (likely(ireq->sack_ok)) { 71233ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 713de213e5eSEric Dumazet if (unlikely(!ireq->tstamp_ok)) 7144957faadSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 71533ad798cSAdam Langley } 71633ad798cSAdam Langley 7174957faadSWilliam Allen Simpson /* Similar rationale to tcp_syn_options() applies here, too. 7184957faadSWilliam Allen Simpson * If the <SYN> options fit, the same options should fit now! 7194957faadSWilliam Allen Simpson */ 7204957faadSWilliam Allen Simpson if (*md5 == NULL && 721de213e5eSEric Dumazet ireq->tstamp_ok && 7224957faadSWilliam Allen Simpson cookie_plus > TCPOLEN_COOKIE_BASE) { 7234957faadSWilliam Allen Simpson int need = cookie_plus; /* has TCPOLEN_COOKIE_BASE */ 7244957faadSWilliam Allen Simpson 7254957faadSWilliam Allen Simpson if (0x2 & need) { 7264957faadSWilliam Allen Simpson /* 32-bit multiple */ 7274957faadSWilliam Allen Simpson need += 2; /* NOPs */ 7284957faadSWilliam Allen Simpson } 7294957faadSWilliam Allen Simpson if (need <= remaining) { 7304957faadSWilliam Allen Simpson opts->options |= OPTION_COOKIE_EXTENSION; 7314957faadSWilliam Allen Simpson opts->hash_size = cookie_plus - TCPOLEN_COOKIE_BASE; 7324957faadSWilliam Allen Simpson remaining -= need; 7334957faadSWilliam Allen Simpson } else { 7344957faadSWilliam Allen Simpson /* There's no error return, so flag it. */ 7354957faadSWilliam Allen Simpson xvp->cookie_out_never = 1; /* true */ 7364957faadSWilliam Allen Simpson opts->hash_size = 0; 7374957faadSWilliam Allen Simpson } 7384957faadSWilliam Allen Simpson } 7394957faadSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 74033ad798cSAdam Langley } 74133ad798cSAdam Langley 74267edfef7SAndi Kleen /* Compute TCP options for ESTABLISHED sockets. This is not the 74367edfef7SAndi Kleen * final wire format yet. 74467edfef7SAndi Kleen */ 74595c96174SEric Dumazet static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb, 74633ad798cSAdam Langley struct tcp_out_options *opts, 747cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 748cf533ea5SEric Dumazet { 74933ad798cSAdam Langley struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL; 75033ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 75195c96174SEric Dumazet unsigned int size = 0; 752cabeccbdSIlpo Järvinen unsigned int eff_sacks; 75333ad798cSAdam Langley 75433ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 75533ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 75633ad798cSAdam Langley if (unlikely(*md5)) { 75733ad798cSAdam Langley opts->options |= OPTION_MD5; 75833ad798cSAdam Langley size += TCPOLEN_MD5SIG_ALIGNED; 75933ad798cSAdam Langley } 76033ad798cSAdam Langley #else 76133ad798cSAdam Langley *md5 = NULL; 76233ad798cSAdam Langley #endif 76333ad798cSAdam Langley 76433ad798cSAdam Langley if (likely(tp->rx_opt.tstamp_ok)) { 76533ad798cSAdam Langley opts->options |= OPTION_TS; 76633ad798cSAdam Langley opts->tsval = tcb ? tcb->when : 0; 76733ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 76833ad798cSAdam Langley size += TCPOLEN_TSTAMP_ALIGNED; 76933ad798cSAdam Langley } 77033ad798cSAdam Langley 771cabeccbdSIlpo Järvinen eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; 772cabeccbdSIlpo Järvinen if (unlikely(eff_sacks)) { 77395c96174SEric Dumazet const unsigned int remaining = MAX_TCP_OPTION_SPACE - size; 77433ad798cSAdam Langley opts->num_sack_blocks = 77595c96174SEric Dumazet min_t(unsigned int, eff_sacks, 77633ad798cSAdam Langley (remaining - TCPOLEN_SACK_BASE_ALIGNED) / 77733ad798cSAdam Langley TCPOLEN_SACK_PERBLOCK); 77833ad798cSAdam Langley size += TCPOLEN_SACK_BASE_ALIGNED + 77933ad798cSAdam Langley opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; 78033ad798cSAdam Langley } 78133ad798cSAdam Langley 78233ad798cSAdam Langley return size; 78340efc6faSStephen Hemminger } 7841da177e4SLinus Torvalds 7851da177e4SLinus Torvalds /* This routine actually transmits TCP packets queued in by 7861da177e4SLinus Torvalds * tcp_do_sendmsg(). This is used by both the initial 7871da177e4SLinus Torvalds * transmission and possible later retransmissions. 7881da177e4SLinus Torvalds * All SKB's seen here are completely headerless. It is our 7891da177e4SLinus Torvalds * job to build the TCP header, and pass the packet down to 7901da177e4SLinus Torvalds * IP so it can do the same plus pass the packet off to the 7911da177e4SLinus Torvalds * device. 7921da177e4SLinus Torvalds * 7931da177e4SLinus Torvalds * We are working here with either a clone of the original 7941da177e4SLinus Torvalds * SKB, or a fresh unique copy made by the retransmit engine. 7951da177e4SLinus Torvalds */ 796056834d9SIlpo Järvinen static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, 797056834d9SIlpo Järvinen gfp_t gfp_mask) 7981da177e4SLinus Torvalds { 7996687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 800dfb4b9dcSDavid S. Miller struct inet_sock *inet; 801dfb4b9dcSDavid S. Miller struct tcp_sock *tp; 802dfb4b9dcSDavid S. Miller struct tcp_skb_cb *tcb; 80333ad798cSAdam Langley struct tcp_out_options opts; 80495c96174SEric Dumazet unsigned int tcp_options_size, tcp_header_size; 805cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 8061da177e4SLinus Torvalds struct tcphdr *th; 8071da177e4SLinus Torvalds int err; 8081da177e4SLinus Torvalds 809dfb4b9dcSDavid S. Miller BUG_ON(!skb || !tcp_skb_pcount(skb)); 810dfb4b9dcSDavid S. Miller 811dfb4b9dcSDavid S. Miller /* If congestion control is doing timestamping, we must 812dfb4b9dcSDavid S. Miller * take such a timestamp before we potentially clone/copy. 813dfb4b9dcSDavid S. Miller */ 814164891aaSStephen Hemminger if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP) 815dfb4b9dcSDavid S. Miller __net_timestamp(skb); 816dfb4b9dcSDavid S. Miller 817dfb4b9dcSDavid S. Miller if (likely(clone_it)) { 818dfb4b9dcSDavid S. Miller if (unlikely(skb_cloned(skb))) 819dfb4b9dcSDavid S. Miller skb = pskb_copy(skb, gfp_mask); 820dfb4b9dcSDavid S. Miller else 821dfb4b9dcSDavid S. Miller skb = skb_clone(skb, gfp_mask); 822dfb4b9dcSDavid S. Miller if (unlikely(!skb)) 823dfb4b9dcSDavid S. Miller return -ENOBUFS; 824dfb4b9dcSDavid S. Miller } 825dfb4b9dcSDavid S. Miller 826dfb4b9dcSDavid S. Miller inet = inet_sk(sk); 827dfb4b9dcSDavid S. Miller tp = tcp_sk(sk); 828dfb4b9dcSDavid S. Miller tcb = TCP_SKB_CB(skb); 82933ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 8301da177e4SLinus Torvalds 8314de075e0SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) 83233ad798cSAdam Langley tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); 83333ad798cSAdam Langley else 83433ad798cSAdam Langley tcp_options_size = tcp_established_options(sk, skb, &opts, 83533ad798cSAdam Langley &md5); 83633ad798cSAdam Langley tcp_header_size = tcp_options_size + sizeof(struct tcphdr); 8371da177e4SLinus Torvalds 8383853b584STom Herbert if (tcp_packets_in_flight(tp) == 0) { 8396687e988SArnaldo Carvalho de Melo tcp_ca_event(sk, CA_EVENT_TX_START); 8403853b584STom Herbert skb->ooo_okay = 1; 8413853b584STom Herbert } else 8423853b584STom Herbert skb->ooo_okay = 0; 8431da177e4SLinus Torvalds 844aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 845aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 846e89862f4SDavid S. Miller skb_set_owner_w(skb, sk); 8471da177e4SLinus Torvalds 8481da177e4SLinus Torvalds /* Build TCP header and checksum it. */ 849aa8223c7SArnaldo Carvalho de Melo th = tcp_hdr(skb); 850c720c7e8SEric Dumazet th->source = inet->inet_sport; 851c720c7e8SEric Dumazet th->dest = inet->inet_dport; 8521da177e4SLinus Torvalds th->seq = htonl(tcb->seq); 8531da177e4SLinus Torvalds th->ack_seq = htonl(tp->rcv_nxt); 854df7a3b07SAl Viro *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | 8554de075e0SEric Dumazet tcb->tcp_flags); 856dfb4b9dcSDavid S. Miller 8574de075e0SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) { 8581da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments 8591da177e4SLinus Torvalds * is never scaled. 8601da177e4SLinus Torvalds */ 861600ff0c2SIlpo Järvinen th->window = htons(min(tp->rcv_wnd, 65535U)); 8621da177e4SLinus Torvalds } else { 8631da177e4SLinus Torvalds th->window = htons(tcp_select_window(sk)); 8641da177e4SLinus Torvalds } 8651da177e4SLinus Torvalds th->check = 0; 8661da177e4SLinus Torvalds th->urg_ptr = 0; 8671da177e4SLinus Torvalds 86833f5f57eSIlpo Järvinen /* The urg_mode check is necessary during a below snd_una win probe */ 8697691367dSHerbert Xu if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { 8707691367dSHerbert Xu if (before(tp->snd_up, tcb->seq + 0x10000)) { 8711da177e4SLinus Torvalds th->urg_ptr = htons(tp->snd_up - tcb->seq); 8721da177e4SLinus Torvalds th->urg = 1; 8737691367dSHerbert Xu } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { 8740eae88f3SEric Dumazet th->urg_ptr = htons(0xFFFF); 8757691367dSHerbert Xu th->urg = 1; 8767691367dSHerbert Xu } 8771da177e4SLinus Torvalds } 8781da177e4SLinus Torvalds 879bd0388aeSWilliam Allen Simpson tcp_options_write((__be32 *)(th + 1), tp, &opts); 8804de075e0SEric Dumazet if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0)) 8819e412ba7SIlpo Järvinen TCP_ECN_send(sk, skb, tcp_header_size); 882dfb4b9dcSDavid S. Miller 883cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 884cfb6eeb4SYOSHIFUJI Hideaki /* Calculate the MD5 hash, as we have all we need now */ 885cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 886a465419bSEric Dumazet sk_nocaps_add(sk, NETIF_F_GSO_MASK); 887bd0388aeSWilliam Allen Simpson tp->af_specific->calc_md5_hash(opts.hash_location, 88849a72dfbSAdam Langley md5, sk, NULL, skb); 889cfb6eeb4SYOSHIFUJI Hideaki } 890cfb6eeb4SYOSHIFUJI Hideaki #endif 891cfb6eeb4SYOSHIFUJI Hideaki 892bb296246SHerbert Xu icsk->icsk_af_ops->send_check(sk, skb); 8931da177e4SLinus Torvalds 8944de075e0SEric Dumazet if (likely(tcb->tcp_flags & TCPHDR_ACK)) 895fc6415bcSDavid S. Miller tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); 8961da177e4SLinus Torvalds 8971da177e4SLinus Torvalds if (skb->len != tcp_header_size) 898cf533ea5SEric Dumazet tcp_event_data_sent(tp, sk); 8991da177e4SLinus Torvalds 900bd37a088SWei Yongjun if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) 901aa2ea058STom Herbert TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, 902aa2ea058STom Herbert tcp_skb_pcount(skb)); 9031da177e4SLinus Torvalds 904d9d8da80SDavid S. Miller err = icsk->icsk_af_ops->queue_xmit(skb, &inet->cork.fl); 90583de47cdSHua Zhong if (likely(err <= 0)) 9061da177e4SLinus Torvalds return err; 9071da177e4SLinus Torvalds 9083cfe3baaSIlpo Järvinen tcp_enter_cwr(sk, 1); 9091da177e4SLinus Torvalds 910b9df3cb8SGerrit Renker return net_xmit_eval(err); 9111da177e4SLinus Torvalds } 9121da177e4SLinus Torvalds 91367edfef7SAndi Kleen /* This routine just queues the buffer for sending. 9141da177e4SLinus Torvalds * 9151da177e4SLinus Torvalds * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, 9161da177e4SLinus Torvalds * otherwise socket can stall. 9171da177e4SLinus Torvalds */ 9181da177e4SLinus Torvalds static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) 9191da177e4SLinus Torvalds { 9201da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 9211da177e4SLinus Torvalds 9221da177e4SLinus Torvalds /* Advance write_seq and place onto the write_queue. */ 9231da177e4SLinus Torvalds tp->write_seq = TCP_SKB_CB(skb)->end_seq; 9241da177e4SLinus Torvalds skb_header_release(skb); 925fe067e8aSDavid S. Miller tcp_add_write_queue_tail(sk, skb); 9263ab224beSHideo Aoki sk->sk_wmem_queued += skb->truesize; 9273ab224beSHideo Aoki sk_mem_charge(sk, skb->truesize); 9281da177e4SLinus Torvalds } 9291da177e4SLinus Torvalds 93067edfef7SAndi Kleen /* Initialize TSO segments for a packet. */ 931cf533ea5SEric Dumazet static void tcp_set_skb_tso_segs(const struct sock *sk, struct sk_buff *skb, 932056834d9SIlpo Järvinen unsigned int mss_now) 933f6302d1dSDavid S. Miller { 9348e5b9ddaSHerbert Xu if (skb->len <= mss_now || !sk_can_gso(sk) || 9358e5b9ddaSHerbert Xu skb->ip_summed == CHECKSUM_NONE) { 936f6302d1dSDavid S. Miller /* Avoid the costly divide in the normal 937f6302d1dSDavid S. Miller * non-TSO case. 938f6302d1dSDavid S. Miller */ 9397967168cSHerbert Xu skb_shinfo(skb)->gso_segs = 1; 9407967168cSHerbert Xu skb_shinfo(skb)->gso_size = 0; 9417967168cSHerbert Xu skb_shinfo(skb)->gso_type = 0; 942f6302d1dSDavid S. Miller } else { 943356f89e1SIlpo Järvinen skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now); 9447967168cSHerbert Xu skb_shinfo(skb)->gso_size = mss_now; 945bcd76111SHerbert Xu skb_shinfo(skb)->gso_type = sk->sk_gso_type; 9461da177e4SLinus Torvalds } 9471da177e4SLinus Torvalds } 9481da177e4SLinus Torvalds 94991fed7a1SIlpo Järvinen /* When a modification to fackets out becomes necessary, we need to check 95068f8353bSIlpo Järvinen * skb is counted to fackets_out or not. 95191fed7a1SIlpo Järvinen */ 952cf533ea5SEric Dumazet static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb, 95391fed7a1SIlpo Järvinen int decr) 95491fed7a1SIlpo Järvinen { 955a47e5a98SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 956a47e5a98SIlpo Järvinen 957dc86967bSIlpo Järvinen if (!tp->sacked_out || tcp_is_reno(tp)) 95891fed7a1SIlpo Järvinen return; 95991fed7a1SIlpo Järvinen 9606859d494SIlpo Järvinen if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq)) 96191fed7a1SIlpo Järvinen tp->fackets_out -= decr; 96291fed7a1SIlpo Järvinen } 96391fed7a1SIlpo Järvinen 964797108d1SIlpo Järvinen /* Pcount in the middle of the write queue got changed, we need to do various 965797108d1SIlpo Järvinen * tweaks to fix counters 966797108d1SIlpo Järvinen */ 967cf533ea5SEric Dumazet static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) 968797108d1SIlpo Järvinen { 969797108d1SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 970797108d1SIlpo Järvinen 971797108d1SIlpo Järvinen tp->packets_out -= decr; 972797108d1SIlpo Järvinen 973797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 974797108d1SIlpo Järvinen tp->sacked_out -= decr; 975797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 976797108d1SIlpo Järvinen tp->retrans_out -= decr; 977797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 978797108d1SIlpo Järvinen tp->lost_out -= decr; 979797108d1SIlpo Järvinen 980797108d1SIlpo Järvinen /* Reno case is special. Sigh... */ 981797108d1SIlpo Järvinen if (tcp_is_reno(tp) && decr > 0) 982797108d1SIlpo Järvinen tp->sacked_out -= min_t(u32, tp->sacked_out, decr); 983797108d1SIlpo Järvinen 984797108d1SIlpo Järvinen tcp_adjust_fackets_out(sk, skb, decr); 985797108d1SIlpo Järvinen 986797108d1SIlpo Järvinen if (tp->lost_skb_hint && 987797108d1SIlpo Järvinen before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && 98852cf3cc8SIlpo Järvinen (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))) 989797108d1SIlpo Järvinen tp->lost_cnt_hint -= decr; 990797108d1SIlpo Järvinen 991797108d1SIlpo Järvinen tcp_verify_left_out(tp); 992797108d1SIlpo Järvinen } 993797108d1SIlpo Järvinen 9941da177e4SLinus Torvalds /* Function to create two new TCP segments. Shrinks the given segment 9951da177e4SLinus Torvalds * to the specified size and appends a new segment with the rest of the 9961da177e4SLinus Torvalds * packet to the list. This won't be called frequently, I hope. 9971da177e4SLinus Torvalds * Remember, these are still headerless SKBs at this point. 9981da177e4SLinus Torvalds */ 999056834d9SIlpo Järvinen int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, 1000056834d9SIlpo Järvinen unsigned int mss_now) 10011da177e4SLinus Torvalds { 10021da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 10031da177e4SLinus Torvalds struct sk_buff *buff; 10046475be16SDavid S. Miller int nsize, old_factor; 1005b60b49eaSHerbert Xu int nlen; 10069ce01461SIlpo Järvinen u8 flags; 10071da177e4SLinus Torvalds 10082fceec13SIlpo Järvinen if (WARN_ON(len > skb->len)) 10092fceec13SIlpo Järvinen return -EINVAL; 10106a438bbeSStephen Hemminger 10111da177e4SLinus Torvalds nsize = skb_headlen(skb) - len; 10121da177e4SLinus Torvalds if (nsize < 0) 10131da177e4SLinus Torvalds nsize = 0; 10141da177e4SLinus Torvalds 10151da177e4SLinus Torvalds if (skb_cloned(skb) && 10161da177e4SLinus Torvalds skb_is_nonlinear(skb) && 10171da177e4SLinus Torvalds pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) 10181da177e4SLinus Torvalds return -ENOMEM; 10191da177e4SLinus Torvalds 10201da177e4SLinus Torvalds /* Get a new skb... force flag on. */ 10211da177e4SLinus Torvalds buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC); 10221da177e4SLinus Torvalds if (buff == NULL) 10231da177e4SLinus Torvalds return -ENOMEM; /* We'll just try again later. */ 1024ef5cb973SHerbert Xu 10253ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 10263ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1027b60b49eaSHerbert Xu nlen = skb->len - len - nsize; 1028b60b49eaSHerbert Xu buff->truesize += nlen; 1029b60b49eaSHerbert Xu skb->truesize -= nlen; 10301da177e4SLinus Torvalds 10311da177e4SLinus Torvalds /* Correct the sequence numbers. */ 10321da177e4SLinus Torvalds TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 10331da177e4SLinus Torvalds TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 10341da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 10351da177e4SLinus Torvalds 10361da177e4SLinus Torvalds /* PSH and FIN should only be set in the second packet. */ 10374de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 10384de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 10394de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1040e14c3cafSHerbert Xu TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; 10411da177e4SLinus Torvalds 104284fa7933SPatrick McHardy if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { 10431da177e4SLinus Torvalds /* Copy and checksum data tail into the new buffer. */ 1044056834d9SIlpo Järvinen buff->csum = csum_partial_copy_nocheck(skb->data + len, 1045056834d9SIlpo Järvinen skb_put(buff, nsize), 10461da177e4SLinus Torvalds nsize, 0); 10471da177e4SLinus Torvalds 10481da177e4SLinus Torvalds skb_trim(skb, len); 10491da177e4SLinus Torvalds 10501da177e4SLinus Torvalds skb->csum = csum_block_sub(skb->csum, buff->csum, len); 10511da177e4SLinus Torvalds } else { 105284fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 10531da177e4SLinus Torvalds skb_split(skb, buff, len); 10541da177e4SLinus Torvalds } 10551da177e4SLinus Torvalds 10561da177e4SLinus Torvalds buff->ip_summed = skb->ip_summed; 10571da177e4SLinus Torvalds 10581da177e4SLinus Torvalds /* Looks stupid, but our code really uses when of 10591da177e4SLinus Torvalds * skbs, which it never sent before. --ANK 10601da177e4SLinus Torvalds */ 10611da177e4SLinus Torvalds TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when; 1062a61bbcf2SPatrick McHardy buff->tstamp = skb->tstamp; 10631da177e4SLinus Torvalds 10646475be16SDavid S. Miller old_factor = tcp_skb_pcount(skb); 10656475be16SDavid S. Miller 10661da177e4SLinus Torvalds /* Fix up tso_factor for both original and new SKB. */ 1067846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1068846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, buff, mss_now); 10691da177e4SLinus Torvalds 10706475be16SDavid S. Miller /* If this packet has been sent out already, we must 10716475be16SDavid S. Miller * adjust the various packet counters. 10726475be16SDavid S. Miller */ 1073cf0b450cSHerbert Xu if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 10746475be16SDavid S. Miller int diff = old_factor - tcp_skb_pcount(skb) - 10756475be16SDavid S. Miller tcp_skb_pcount(buff); 10761da177e4SLinus Torvalds 1077797108d1SIlpo Järvinen if (diff) 1078797108d1SIlpo Järvinen tcp_adjust_pcount(sk, skb, diff); 10791da177e4SLinus Torvalds } 10801da177e4SLinus Torvalds 10811da177e4SLinus Torvalds /* Link BUFF into the send queue. */ 1082f44b5271SDavid S. Miller skb_header_release(buff); 1083fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 10841da177e4SLinus Torvalds 10851da177e4SLinus Torvalds return 0; 10861da177e4SLinus Torvalds } 10871da177e4SLinus Torvalds 10881da177e4SLinus Torvalds /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c 10891da177e4SLinus Torvalds * eventually). The difference is that pulled data not copied, but 10901da177e4SLinus Torvalds * immediately discarded. 10911da177e4SLinus Torvalds */ 1092f2911969SHerbert Xu ~{PmVHI~} static void __pskb_trim_head(struct sk_buff *skb, int len) 10931da177e4SLinus Torvalds { 10941da177e4SLinus Torvalds int i, k, eat; 10951da177e4SLinus Torvalds 10964fa48bf3SEric Dumazet eat = min_t(int, len, skb_headlen(skb)); 10974fa48bf3SEric Dumazet if (eat) { 10984fa48bf3SEric Dumazet __skb_pull(skb, eat); 109922b4a4f2SEric Dumazet skb->avail_size -= eat; 11004fa48bf3SEric Dumazet len -= eat; 11014fa48bf3SEric Dumazet if (!len) 11024fa48bf3SEric Dumazet return; 11034fa48bf3SEric Dumazet } 11041da177e4SLinus Torvalds eat = len; 11051da177e4SLinus Torvalds k = 0; 11061da177e4SLinus Torvalds for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) { 11079e903e08SEric Dumazet int size = skb_frag_size(&skb_shinfo(skb)->frags[i]); 11089e903e08SEric Dumazet 11099e903e08SEric Dumazet if (size <= eat) { 1110aff65da0SIan Campbell skb_frag_unref(skb, i); 11119e903e08SEric Dumazet eat -= size; 11121da177e4SLinus Torvalds } else { 11131da177e4SLinus Torvalds skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i]; 11141da177e4SLinus Torvalds if (eat) { 11151da177e4SLinus Torvalds skb_shinfo(skb)->frags[k].page_offset += eat; 11169e903e08SEric Dumazet skb_frag_size_sub(&skb_shinfo(skb)->frags[k], eat); 11171da177e4SLinus Torvalds eat = 0; 11181da177e4SLinus Torvalds } 11191da177e4SLinus Torvalds k++; 11201da177e4SLinus Torvalds } 11211da177e4SLinus Torvalds } 11221da177e4SLinus Torvalds skb_shinfo(skb)->nr_frags = k; 11231da177e4SLinus Torvalds 112427a884dcSArnaldo Carvalho de Melo skb_reset_tail_pointer(skb); 11251da177e4SLinus Torvalds skb->data_len -= len; 11261da177e4SLinus Torvalds skb->len = skb->data_len; 11271da177e4SLinus Torvalds } 11281da177e4SLinus Torvalds 112967edfef7SAndi Kleen /* Remove acked data from a packet in the transmit queue. */ 11301da177e4SLinus Torvalds int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 11311da177e4SLinus Torvalds { 1132056834d9SIlpo Järvinen if (skb_cloned(skb) && pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) 11331da177e4SLinus Torvalds return -ENOMEM; 11341da177e4SLinus Torvalds 11354fa48bf3SEric Dumazet __pskb_trim_head(skb, len); 11361da177e4SLinus Torvalds 11371da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq += len; 113884fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 11391da177e4SLinus Torvalds 11401da177e4SLinus Torvalds skb->truesize -= len; 11411da177e4SLinus Torvalds sk->sk_wmem_queued -= len; 11423ab224beSHideo Aoki sk_mem_uncharge(sk, len); 11431da177e4SLinus Torvalds sock_set_flag(sk, SOCK_QUEUE_SHRUNK); 11441da177e4SLinus Torvalds 11455b35e1e6SNeal Cardwell /* Any change of skb->len requires recalculation of tso factor. */ 11461da177e4SLinus Torvalds if (tcp_skb_pcount(skb) > 1) 11475b35e1e6SNeal Cardwell tcp_set_skb_tso_segs(sk, skb, tcp_skb_mss(skb)); 11481da177e4SLinus Torvalds 11491da177e4SLinus Torvalds return 0; 11501da177e4SLinus Torvalds } 11511da177e4SLinus Torvalds 115267edfef7SAndi Kleen /* Calculate MSS. Not accounting for SACKs here. */ 1153*67469601SEric Dumazet int tcp_mtu_to_mss(struct sock *sk, int pmtu) 11545d424d5aSJohn Heffner { 1155cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1156cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 11575d424d5aSJohn Heffner int mss_now; 11585d424d5aSJohn Heffner 11595d424d5aSJohn Heffner /* Calculate base mss without TCP options: 11605d424d5aSJohn Heffner It is MMS_S - sizeof(tcphdr) of rfc1122 11615d424d5aSJohn Heffner */ 11625d424d5aSJohn Heffner mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 11635d424d5aSJohn Heffner 1164*67469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 1165*67469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 1166*67469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1167*67469601SEric Dumazet 1168*67469601SEric Dumazet if (dst && dst_allfrag(dst)) 1169*67469601SEric Dumazet mss_now -= icsk->icsk_af_ops->net_frag_header_len; 1170*67469601SEric Dumazet } 1171*67469601SEric Dumazet 11725d424d5aSJohn Heffner /* Clamp it (mss_clamp does not include tcp options) */ 11735d424d5aSJohn Heffner if (mss_now > tp->rx_opt.mss_clamp) 11745d424d5aSJohn Heffner mss_now = tp->rx_opt.mss_clamp; 11755d424d5aSJohn Heffner 11765d424d5aSJohn Heffner /* Now subtract optional transport overhead */ 11775d424d5aSJohn Heffner mss_now -= icsk->icsk_ext_hdr_len; 11785d424d5aSJohn Heffner 11795d424d5aSJohn Heffner /* Then reserve room for full set of TCP options and 8 bytes of data */ 11805d424d5aSJohn Heffner if (mss_now < 48) 11815d424d5aSJohn Heffner mss_now = 48; 11825d424d5aSJohn Heffner 11835d424d5aSJohn Heffner /* Now subtract TCP options size, not including SACKs */ 11845d424d5aSJohn Heffner mss_now -= tp->tcp_header_len - sizeof(struct tcphdr); 11855d424d5aSJohn Heffner 11865d424d5aSJohn Heffner return mss_now; 11875d424d5aSJohn Heffner } 11885d424d5aSJohn Heffner 11895d424d5aSJohn Heffner /* Inverse of above */ 1190*67469601SEric Dumazet int tcp_mss_to_mtu(struct sock *sk, int mss) 11915d424d5aSJohn Heffner { 1192cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1193cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 11945d424d5aSJohn Heffner int mtu; 11955d424d5aSJohn Heffner 11965d424d5aSJohn Heffner mtu = mss + 11975d424d5aSJohn Heffner tp->tcp_header_len + 11985d424d5aSJohn Heffner icsk->icsk_ext_hdr_len + 11995d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 12005d424d5aSJohn Heffner 1201*67469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 1202*67469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 1203*67469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1204*67469601SEric Dumazet 1205*67469601SEric Dumazet if (dst && dst_allfrag(dst)) 1206*67469601SEric Dumazet mtu += icsk->icsk_af_ops->net_frag_header_len; 1207*67469601SEric Dumazet } 12085d424d5aSJohn Heffner return mtu; 12095d424d5aSJohn Heffner } 12105d424d5aSJohn Heffner 121167edfef7SAndi Kleen /* MTU probing init per socket */ 12125d424d5aSJohn Heffner void tcp_mtup_init(struct sock *sk) 12135d424d5aSJohn Heffner { 12145d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 12155d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 12165d424d5aSJohn Heffner 12175d424d5aSJohn Heffner icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1; 12185d424d5aSJohn Heffner icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 12195d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 12205d424d5aSJohn Heffner icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss); 12215d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 12225d424d5aSJohn Heffner } 12234bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_mtup_init); 12245d424d5aSJohn Heffner 12251da177e4SLinus Torvalds /* This function synchronize snd mss to current pmtu/exthdr set. 12261da177e4SLinus Torvalds 12271da177e4SLinus Torvalds tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 12281da177e4SLinus Torvalds for TCP options, but includes only bare TCP header. 12291da177e4SLinus Torvalds 12301da177e4SLinus Torvalds tp->rx_opt.mss_clamp is mss negotiated at connection setup. 1231caa20d9aSStephen Hemminger It is minimum of user_mss and mss received with SYN. 12321da177e4SLinus Torvalds It also does not include TCP options. 12331da177e4SLinus Torvalds 1234d83d8461SArnaldo Carvalho de Melo inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. 12351da177e4SLinus Torvalds 12361da177e4SLinus Torvalds tp->mss_cache is current effective sending mss, including 12371da177e4SLinus Torvalds all tcp options except for SACKs. It is evaluated, 12381da177e4SLinus Torvalds taking into account current pmtu, but never exceeds 12391da177e4SLinus Torvalds tp->rx_opt.mss_clamp. 12401da177e4SLinus Torvalds 12411da177e4SLinus Torvalds NOTE1. rfc1122 clearly states that advertised MSS 12421da177e4SLinus Torvalds DOES NOT include either tcp or ip options. 12431da177e4SLinus Torvalds 1244d83d8461SArnaldo Carvalho de Melo NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache 1245d83d8461SArnaldo Carvalho de Melo are READ ONLY outside this function. --ANK (980731) 12461da177e4SLinus Torvalds */ 12471da177e4SLinus Torvalds unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 12481da177e4SLinus Torvalds { 12491da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1250d83d8461SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 12515d424d5aSJohn Heffner int mss_now; 12521da177e4SLinus Torvalds 12535d424d5aSJohn Heffner if (icsk->icsk_mtup.search_high > pmtu) 12545d424d5aSJohn Heffner icsk->icsk_mtup.search_high = pmtu; 12551da177e4SLinus Torvalds 12565d424d5aSJohn Heffner mss_now = tcp_mtu_to_mss(sk, pmtu); 1257409d22b4SIlpo Järvinen mss_now = tcp_bound_to_half_wnd(tp, mss_now); 12581da177e4SLinus Torvalds 12591da177e4SLinus Torvalds /* And store cached results */ 1260d83d8461SArnaldo Carvalho de Melo icsk->icsk_pmtu_cookie = pmtu; 12615d424d5aSJohn Heffner if (icsk->icsk_mtup.enabled) 12625d424d5aSJohn Heffner mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); 1263c1b4a7e6SDavid S. Miller tp->mss_cache = mss_now; 12641da177e4SLinus Torvalds 12651da177e4SLinus Torvalds return mss_now; 12661da177e4SLinus Torvalds } 12674bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_sync_mss); 12681da177e4SLinus Torvalds 12691da177e4SLinus Torvalds /* Compute the current effective MSS, taking SACKs and IP options, 12701da177e4SLinus Torvalds * and even PMTU discovery events into account. 12711da177e4SLinus Torvalds */ 12720c54b85fSIlpo Järvinen unsigned int tcp_current_mss(struct sock *sk) 12731da177e4SLinus Torvalds { 1274cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1275cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1276c1b4a7e6SDavid S. Miller u32 mss_now; 127795c96174SEric Dumazet unsigned int header_len; 127833ad798cSAdam Langley struct tcp_out_options opts; 127933ad798cSAdam Langley struct tcp_md5sig_key *md5; 12801da177e4SLinus Torvalds 1281c1b4a7e6SDavid S. Miller mss_now = tp->mss_cache; 1282c1b4a7e6SDavid S. Miller 12831da177e4SLinus Torvalds if (dst) { 12841da177e4SLinus Torvalds u32 mtu = dst_mtu(dst); 1285d83d8461SArnaldo Carvalho de Melo if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 12861da177e4SLinus Torvalds mss_now = tcp_sync_mss(sk, mtu); 12871da177e4SLinus Torvalds } 12881da177e4SLinus Torvalds 128933ad798cSAdam Langley header_len = tcp_established_options(sk, NULL, &opts, &md5) + 129033ad798cSAdam Langley sizeof(struct tcphdr); 129133ad798cSAdam Langley /* The mss_cache is sized based on tp->tcp_header_len, which assumes 129233ad798cSAdam Langley * some common options. If this is an odd packet (because we have SACK 129333ad798cSAdam Langley * blocks etc) then our calculated header_len will be different, and 129433ad798cSAdam Langley * we have to adjust mss_now correspondingly */ 129533ad798cSAdam Langley if (header_len != tp->tcp_header_len) { 129633ad798cSAdam Langley int delta = (int) header_len - tp->tcp_header_len; 129733ad798cSAdam Langley mss_now -= delta; 129833ad798cSAdam Langley } 1299cfb6eeb4SYOSHIFUJI Hideaki 13001da177e4SLinus Torvalds return mss_now; 13011da177e4SLinus Torvalds } 13021da177e4SLinus Torvalds 1303a762a980SDavid S. Miller /* Congestion window validation. (RFC2861) */ 13049e412ba7SIlpo Järvinen static void tcp_cwnd_validate(struct sock *sk) 1305a762a980SDavid S. Miller { 13069e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1307a762a980SDavid S. Miller 1308d436d686SIlpo Järvinen if (tp->packets_out >= tp->snd_cwnd) { 1309a762a980SDavid S. Miller /* Network is feed fully. */ 1310a762a980SDavid S. Miller tp->snd_cwnd_used = 0; 1311a762a980SDavid S. Miller tp->snd_cwnd_stamp = tcp_time_stamp; 1312a762a980SDavid S. Miller } else { 1313a762a980SDavid S. Miller /* Network starves. */ 1314a762a980SDavid S. Miller if (tp->packets_out > tp->snd_cwnd_used) 1315a762a980SDavid S. Miller tp->snd_cwnd_used = tp->packets_out; 1316a762a980SDavid S. Miller 131715d33c07SDavid S. Miller if (sysctl_tcp_slow_start_after_idle && 131815d33c07SDavid S. Miller (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto) 1319a762a980SDavid S. Miller tcp_cwnd_application_limited(sk); 1320a762a980SDavid S. Miller } 1321a762a980SDavid S. Miller } 1322a762a980SDavid S. Miller 13230e3a4803SIlpo Järvinen /* Returns the portion of skb which can be sent right away without 13240e3a4803SIlpo Järvinen * introducing MSS oddities to segment boundaries. In rare cases where 13250e3a4803SIlpo Järvinen * mss_now != mss_cache, we will request caller to create a small skb 13260e3a4803SIlpo Järvinen * per input skb which could be mostly avoided here (if desired). 13275ea3a748SIlpo Järvinen * 13285ea3a748SIlpo Järvinen * We explicitly want to create a request for splitting write queue tail 13295ea3a748SIlpo Järvinen * to a small skb for Nagle purposes while avoiding unnecessary modulos, 13305ea3a748SIlpo Järvinen * thus all the complexity (cwnd_len is always MSS multiple which we 13315ea3a748SIlpo Järvinen * return whenever allowed by the other factors). Basically we need the 13325ea3a748SIlpo Järvinen * modulo only when the receiver window alone is the limiting factor or 13335ea3a748SIlpo Järvinen * when we would be allowed to send the split-due-to-Nagle skb fully. 13340e3a4803SIlpo Järvinen */ 1335cf533ea5SEric Dumazet static unsigned int tcp_mss_split_point(const struct sock *sk, const struct sk_buff *skb, 1336056834d9SIlpo Järvinen unsigned int mss_now, unsigned int cwnd) 1337c1b4a7e6SDavid S. Miller { 1338cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 13390e3a4803SIlpo Järvinen u32 needed, window, cwnd_len; 1340c1b4a7e6SDavid S. Miller 134190840defSIlpo Järvinen window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1342c1b4a7e6SDavid S. Miller cwnd_len = mss_now * cwnd; 13430e3a4803SIlpo Järvinen 13440e3a4803SIlpo Järvinen if (likely(cwnd_len <= window && skb != tcp_write_queue_tail(sk))) 13450e3a4803SIlpo Järvinen return cwnd_len; 13460e3a4803SIlpo Järvinen 13475ea3a748SIlpo Järvinen needed = min(skb->len, window); 13485ea3a748SIlpo Järvinen 134917515408SIlpo Järvinen if (cwnd_len <= needed) 13500e3a4803SIlpo Järvinen return cwnd_len; 13510e3a4803SIlpo Järvinen 13520e3a4803SIlpo Järvinen return needed - needed % mss_now; 1353c1b4a7e6SDavid S. Miller } 1354c1b4a7e6SDavid S. Miller 1355c1b4a7e6SDavid S. Miller /* Can at least one segment of SKB be sent right now, according to the 1356c1b4a7e6SDavid S. Miller * congestion window rules? If so, return how many segments are allowed. 1357c1b4a7e6SDavid S. Miller */ 1358cf533ea5SEric Dumazet static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, 1359cf533ea5SEric Dumazet const struct sk_buff *skb) 1360c1b4a7e6SDavid S. Miller { 1361c1b4a7e6SDavid S. Miller u32 in_flight, cwnd; 1362c1b4a7e6SDavid S. Miller 1363c1b4a7e6SDavid S. Miller /* Don't be strict about the congestion window for the final FIN. */ 13644de075e0SEric Dumazet if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && 13654de075e0SEric Dumazet tcp_skb_pcount(skb) == 1) 1366c1b4a7e6SDavid S. Miller return 1; 1367c1b4a7e6SDavid S. Miller 1368c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1369c1b4a7e6SDavid S. Miller cwnd = tp->snd_cwnd; 1370c1b4a7e6SDavid S. Miller if (in_flight < cwnd) 1371c1b4a7e6SDavid S. Miller return (cwnd - in_flight); 1372c1b4a7e6SDavid S. Miller 1373c1b4a7e6SDavid S. Miller return 0; 1374c1b4a7e6SDavid S. Miller } 1375c1b4a7e6SDavid S. Miller 1376b595076aSUwe Kleine-König /* Initialize TSO state of a skb. 137767edfef7SAndi Kleen * This must be invoked the first time we consider transmitting 1378c1b4a7e6SDavid S. Miller * SKB onto the wire. 1379c1b4a7e6SDavid S. Miller */ 1380cf533ea5SEric Dumazet static int tcp_init_tso_segs(const struct sock *sk, struct sk_buff *skb, 1381056834d9SIlpo Järvinen unsigned int mss_now) 1382c1b4a7e6SDavid S. Miller { 1383c1b4a7e6SDavid S. Miller int tso_segs = tcp_skb_pcount(skb); 1384c1b4a7e6SDavid S. Miller 1385f8269a49SIlpo Järvinen if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { 1386846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1387c1b4a7e6SDavid S. Miller tso_segs = tcp_skb_pcount(skb); 1388c1b4a7e6SDavid S. Miller } 1389c1b4a7e6SDavid S. Miller return tso_segs; 1390c1b4a7e6SDavid S. Miller } 1391c1b4a7e6SDavid S. Miller 139267edfef7SAndi Kleen /* Minshall's variant of the Nagle send check. */ 1393c1b4a7e6SDavid S. Miller static inline int tcp_minshall_check(const struct tcp_sock *tp) 1394c1b4a7e6SDavid S. Miller { 1395c1b4a7e6SDavid S. Miller return after(tp->snd_sml, tp->snd_una) && 1396c1b4a7e6SDavid S. Miller !after(tp->snd_sml, tp->snd_nxt); 1397c1b4a7e6SDavid S. Miller } 1398c1b4a7e6SDavid S. Miller 1399c1b4a7e6SDavid S. Miller /* Return 0, if packet can be sent now without violation Nagle's rules: 1400c1b4a7e6SDavid S. Miller * 1. It is full sized. 1401c1b4a7e6SDavid S. Miller * 2. Or it contains FIN. (already checked by caller) 14026d67e9beSFeng King * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. 1403c1b4a7e6SDavid S. Miller * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 1404c1b4a7e6SDavid S. Miller * With Minshall's modification: all sent small packets are ACKed. 1405c1b4a7e6SDavid S. Miller */ 1406c1b4a7e6SDavid S. Miller static inline int tcp_nagle_check(const struct tcp_sock *tp, 1407c1b4a7e6SDavid S. Miller const struct sk_buff *skb, 140895c96174SEric Dumazet unsigned int mss_now, int nonagle) 1409c1b4a7e6SDavid S. Miller { 1410a02cec21SEric Dumazet return skb->len < mss_now && 1411c1b4a7e6SDavid S. Miller ((nonagle & TCP_NAGLE_CORK) || 1412a02cec21SEric Dumazet (!nonagle && tp->packets_out && tcp_minshall_check(tp))); 1413c1b4a7e6SDavid S. Miller } 1414c1b4a7e6SDavid S. Miller 1415c1b4a7e6SDavid S. Miller /* Return non-zero if the Nagle test allows this packet to be 1416c1b4a7e6SDavid S. Miller * sent now. 1417c1b4a7e6SDavid S. Miller */ 1418cf533ea5SEric Dumazet static inline int tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, 1419c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1420c1b4a7e6SDavid S. Miller { 1421c1b4a7e6SDavid S. Miller /* Nagle rule does not apply to frames, which sit in the middle of the 1422c1b4a7e6SDavid S. Miller * write_queue (they have no chances to get new data). 1423c1b4a7e6SDavid S. Miller * 1424c1b4a7e6SDavid S. Miller * This is implemented in the callers, where they modify the 'nonagle' 1425c1b4a7e6SDavid S. Miller * argument based upon the location of SKB in the send queue. 1426c1b4a7e6SDavid S. Miller */ 1427c1b4a7e6SDavid S. Miller if (nonagle & TCP_NAGLE_PUSH) 1428c1b4a7e6SDavid S. Miller return 1; 1429c1b4a7e6SDavid S. Miller 1430d551e454SIlpo Järvinen /* Don't use the nagle rule for urgent data (or for the final FIN). 1431d551e454SIlpo Järvinen * Nagle can be ignored during F-RTO too (see RFC4138). 1432d551e454SIlpo Järvinen */ 143333f5f57eSIlpo Järvinen if (tcp_urg_mode(tp) || (tp->frto_counter == 2) || 14344de075e0SEric Dumazet (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) 1435c1b4a7e6SDavid S. Miller return 1; 1436c1b4a7e6SDavid S. Miller 1437c1b4a7e6SDavid S. Miller if (!tcp_nagle_check(tp, skb, cur_mss, nonagle)) 1438c1b4a7e6SDavid S. Miller return 1; 1439c1b4a7e6SDavid S. Miller 1440c1b4a7e6SDavid S. Miller return 0; 1441c1b4a7e6SDavid S. Miller } 1442c1b4a7e6SDavid S. Miller 1443c1b4a7e6SDavid S. Miller /* Does at least the first segment of SKB fit into the send window? */ 1444cf533ea5SEric Dumazet static inline int tcp_snd_wnd_test(const struct tcp_sock *tp, const struct sk_buff *skb, 1445056834d9SIlpo Järvinen unsigned int cur_mss) 1446c1b4a7e6SDavid S. Miller { 1447c1b4a7e6SDavid S. Miller u32 end_seq = TCP_SKB_CB(skb)->end_seq; 1448c1b4a7e6SDavid S. Miller 1449c1b4a7e6SDavid S. Miller if (skb->len > cur_mss) 1450c1b4a7e6SDavid S. Miller end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 1451c1b4a7e6SDavid S. Miller 145290840defSIlpo Järvinen return !after(end_seq, tcp_wnd_end(tp)); 1453c1b4a7e6SDavid S. Miller } 1454c1b4a7e6SDavid S. Miller 1455fe067e8aSDavid S. Miller /* This checks if the data bearing packet SKB (usually tcp_send_head(sk)) 1456c1b4a7e6SDavid S. Miller * should be put on the wire right now. If so, it returns the number of 1457c1b4a7e6SDavid S. Miller * packets allowed by the congestion window. 1458c1b4a7e6SDavid S. Miller */ 1459cf533ea5SEric Dumazet static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb, 1460c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1461c1b4a7e6SDavid S. Miller { 1462cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1463c1b4a7e6SDavid S. Miller unsigned int cwnd_quota; 1464c1b4a7e6SDavid S. Miller 1465846998aeSDavid S. Miller tcp_init_tso_segs(sk, skb, cur_mss); 1466c1b4a7e6SDavid S. Miller 1467c1b4a7e6SDavid S. Miller if (!tcp_nagle_test(tp, skb, cur_mss, nonagle)) 1468c1b4a7e6SDavid S. Miller return 0; 1469c1b4a7e6SDavid S. Miller 1470c1b4a7e6SDavid S. Miller cwnd_quota = tcp_cwnd_test(tp, skb); 1471056834d9SIlpo Järvinen if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss)) 1472c1b4a7e6SDavid S. Miller cwnd_quota = 0; 1473c1b4a7e6SDavid S. Miller 1474c1b4a7e6SDavid S. Miller return cwnd_quota; 1475c1b4a7e6SDavid S. Miller } 1476c1b4a7e6SDavid S. Miller 147767edfef7SAndi Kleen /* Test if sending is allowed right now. */ 14789e412ba7SIlpo Järvinen int tcp_may_send_now(struct sock *sk) 1479c1b4a7e6SDavid S. Miller { 1480cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1481fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 1482c1b4a7e6SDavid S. Miller 1483a02cec21SEric Dumazet return skb && 14840c54b85fSIlpo Järvinen tcp_snd_test(sk, skb, tcp_current_mss(sk), 1485c1b4a7e6SDavid S. Miller (tcp_skb_is_last(sk, skb) ? 1486a02cec21SEric Dumazet tp->nonagle : TCP_NAGLE_PUSH)); 1487c1b4a7e6SDavid S. Miller } 1488c1b4a7e6SDavid S. Miller 1489c1b4a7e6SDavid S. Miller /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 1490c1b4a7e6SDavid S. Miller * which is put after SKB on the list. It is very much like 1491c1b4a7e6SDavid S. Miller * tcp_fragment() except that it may make several kinds of assumptions 1492c1b4a7e6SDavid S. Miller * in order to speed up the splitting operation. In particular, we 1493c1b4a7e6SDavid S. Miller * know that all the data is in scatter-gather pages, and that the 1494c1b4a7e6SDavid S. Miller * packet has never been sent out before (and thus is not cloned). 1495c1b4a7e6SDavid S. Miller */ 1496056834d9SIlpo Järvinen static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, 1497c4ead4c5SEric Dumazet unsigned int mss_now, gfp_t gfp) 1498c1b4a7e6SDavid S. Miller { 1499c1b4a7e6SDavid S. Miller struct sk_buff *buff; 1500c1b4a7e6SDavid S. Miller int nlen = skb->len - len; 15019ce01461SIlpo Järvinen u8 flags; 1502c1b4a7e6SDavid S. Miller 1503c1b4a7e6SDavid S. Miller /* All of a TSO frame must be composed of paged data. */ 1504c8ac3774SHerbert Xu if (skb->len != skb->data_len) 1505c8ac3774SHerbert Xu return tcp_fragment(sk, skb, len, mss_now); 1506c1b4a7e6SDavid S. Miller 1507c4ead4c5SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, gfp); 1508c1b4a7e6SDavid S. Miller if (unlikely(buff == NULL)) 1509c1b4a7e6SDavid S. Miller return -ENOMEM; 1510c1b4a7e6SDavid S. Miller 15113ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 15123ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1513b60b49eaSHerbert Xu buff->truesize += nlen; 1514c1b4a7e6SDavid S. Miller skb->truesize -= nlen; 1515c1b4a7e6SDavid S. Miller 1516c1b4a7e6SDavid S. Miller /* Correct the sequence numbers. */ 1517c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 1518c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 1519c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 1520c1b4a7e6SDavid S. Miller 1521c1b4a7e6SDavid S. Miller /* PSH and FIN should only be set in the second packet. */ 15224de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 15234de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 15244de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1525c1b4a7e6SDavid S. Miller 1526c1b4a7e6SDavid S. Miller /* This packet was never sent out yet, so no SACK bits. */ 1527c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->sacked = 0; 1528c1b4a7e6SDavid S. Miller 152984fa7933SPatrick McHardy buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; 1530c1b4a7e6SDavid S. Miller skb_split(skb, buff, len); 1531c1b4a7e6SDavid S. Miller 1532c1b4a7e6SDavid S. Miller /* Fix up tso_factor for both original and new SKB. */ 1533846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1534846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, buff, mss_now); 1535c1b4a7e6SDavid S. Miller 1536c1b4a7e6SDavid S. Miller /* Link BUFF into the send queue. */ 1537c1b4a7e6SDavid S. Miller skb_header_release(buff); 1538fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 1539c1b4a7e6SDavid S. Miller 1540c1b4a7e6SDavid S. Miller return 0; 1541c1b4a7e6SDavid S. Miller } 1542c1b4a7e6SDavid S. Miller 1543c1b4a7e6SDavid S. Miller /* Try to defer sending, if possible, in order to minimize the amount 1544c1b4a7e6SDavid S. Miller * of TSO splitting we do. View it as a kind of TSO Nagle test. 1545c1b4a7e6SDavid S. Miller * 1546c1b4a7e6SDavid S. Miller * This algorithm is from John Heffner. 1547c1b4a7e6SDavid S. Miller */ 15489e412ba7SIlpo Järvinen static int tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb) 1549c1b4a7e6SDavid S. Miller { 15509e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 15516687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 1552c1b4a7e6SDavid S. Miller u32 send_win, cong_win, limit, in_flight; 1553ad9f4f50SEric Dumazet int win_divisor; 1554c1b4a7e6SDavid S. Miller 15554de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 1556ae8064acSJohn Heffner goto send_now; 1557c1b4a7e6SDavid S. Miller 15586687e988SArnaldo Carvalho de Melo if (icsk->icsk_ca_state != TCP_CA_Open) 1559ae8064acSJohn Heffner goto send_now; 1560ae8064acSJohn Heffner 1561ae8064acSJohn Heffner /* Defer for less than two clock ticks. */ 1562bd515c3eSIlpo Järvinen if (tp->tso_deferred && 1563a2acde07SIlpo Järvinen (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1) 1564ae8064acSJohn Heffner goto send_now; 1565908a75c1SDavid S. Miller 1566c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1567c1b4a7e6SDavid S. Miller 1568056834d9SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight)); 1569c1b4a7e6SDavid S. Miller 157090840defSIlpo Järvinen send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1571c1b4a7e6SDavid S. Miller 1572c1b4a7e6SDavid S. Miller /* From in_flight test above, we know that cwnd > in_flight. */ 1573c1b4a7e6SDavid S. Miller cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; 1574c1b4a7e6SDavid S. Miller 1575c1b4a7e6SDavid S. Miller limit = min(send_win, cong_win); 1576c1b4a7e6SDavid S. Miller 1577ba244fe9SDavid S. Miller /* If a full-sized TSO skb can be sent, do it. */ 157882cc1a7aSPeter P Waskiewicz Jr if (limit >= sk->sk_gso_max_size) 1579ae8064acSJohn Heffner goto send_now; 1580ba244fe9SDavid S. Miller 158162ad2761SIlpo Järvinen /* Middle in queue won't get any more data, full sendable already? */ 158262ad2761SIlpo Järvinen if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) 158362ad2761SIlpo Järvinen goto send_now; 158462ad2761SIlpo Järvinen 1585ad9f4f50SEric Dumazet win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor); 1586ad9f4f50SEric Dumazet if (win_divisor) { 1587c1b4a7e6SDavid S. Miller u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); 1588c1b4a7e6SDavid S. Miller 1589c1b4a7e6SDavid S. Miller /* If at least some fraction of a window is available, 1590c1b4a7e6SDavid S. Miller * just use it. 1591c1b4a7e6SDavid S. Miller */ 1592ad9f4f50SEric Dumazet chunk /= win_divisor; 1593c1b4a7e6SDavid S. Miller if (limit >= chunk) 1594ae8064acSJohn Heffner goto send_now; 1595c1b4a7e6SDavid S. Miller } else { 1596c1b4a7e6SDavid S. Miller /* Different approach, try not to defer past a single 1597c1b4a7e6SDavid S. Miller * ACK. Receiver should ACK every other full sized 1598c1b4a7e6SDavid S. Miller * frame, so if we have space for more than 3 frames 1599c1b4a7e6SDavid S. Miller * then send now. 1600c1b4a7e6SDavid S. Miller */ 16016b5a5c0dSNeal Cardwell if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) 1602ae8064acSJohn Heffner goto send_now; 1603c1b4a7e6SDavid S. Miller } 1604c1b4a7e6SDavid S. Miller 1605c1b4a7e6SDavid S. Miller /* Ok, it looks like it is advisable to defer. */ 1606ae8064acSJohn Heffner tp->tso_deferred = 1 | (jiffies << 1); 1607ae8064acSJohn Heffner 1608c1b4a7e6SDavid S. Miller return 1; 1609ae8064acSJohn Heffner 1610ae8064acSJohn Heffner send_now: 1611ae8064acSJohn Heffner tp->tso_deferred = 0; 1612ae8064acSJohn Heffner return 0; 1613c1b4a7e6SDavid S. Miller } 1614c1b4a7e6SDavid S. Miller 16155d424d5aSJohn Heffner /* Create a new MTU probe if we are ready. 161667edfef7SAndi Kleen * MTU probe is regularly attempting to increase the path MTU by 161767edfef7SAndi Kleen * deliberately sending larger packets. This discovers routing 161867edfef7SAndi Kleen * changes resulting in larger path MTUs. 161967edfef7SAndi Kleen * 16205d424d5aSJohn Heffner * Returns 0 if we should wait to probe (no cwnd available), 16215d424d5aSJohn Heffner * 1 if a probe was sent, 1622056834d9SIlpo Järvinen * -1 otherwise 1623056834d9SIlpo Järvinen */ 16245d424d5aSJohn Heffner static int tcp_mtu_probe(struct sock *sk) 16255d424d5aSJohn Heffner { 16265d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 16275d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 16285d424d5aSJohn Heffner struct sk_buff *skb, *nskb, *next; 16295d424d5aSJohn Heffner int len; 16305d424d5aSJohn Heffner int probe_size; 163191cc17c0SIlpo Järvinen int size_needed; 16325d424d5aSJohn Heffner int copy; 16335d424d5aSJohn Heffner int mss_now; 16345d424d5aSJohn Heffner 16355d424d5aSJohn Heffner /* Not currently probing/verifying, 16365d424d5aSJohn Heffner * not in recovery, 16375d424d5aSJohn Heffner * have enough cwnd, and 16385d424d5aSJohn Heffner * not SACKing (the variable headers throw things off) */ 16395d424d5aSJohn Heffner if (!icsk->icsk_mtup.enabled || 16405d424d5aSJohn Heffner icsk->icsk_mtup.probe_size || 16415d424d5aSJohn Heffner inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 16425d424d5aSJohn Heffner tp->snd_cwnd < 11 || 1643cabeccbdSIlpo Järvinen tp->rx_opt.num_sacks || tp->rx_opt.dsack) 16445d424d5aSJohn Heffner return -1; 16455d424d5aSJohn Heffner 16465d424d5aSJohn Heffner /* Very simple search strategy: just double the MSS. */ 16470c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 16485d424d5aSJohn Heffner probe_size = 2 * tp->mss_cache; 164991cc17c0SIlpo Järvinen size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; 16505d424d5aSJohn Heffner if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) { 16515d424d5aSJohn Heffner /* TODO: set timer for probe_converge_event */ 16525d424d5aSJohn Heffner return -1; 16535d424d5aSJohn Heffner } 16545d424d5aSJohn Heffner 16555d424d5aSJohn Heffner /* Have enough data in the send queue to probe? */ 16567f9c33e5SIlpo Järvinen if (tp->write_seq - tp->snd_nxt < size_needed) 16575d424d5aSJohn Heffner return -1; 16585d424d5aSJohn Heffner 165991cc17c0SIlpo Järvinen if (tp->snd_wnd < size_needed) 16605d424d5aSJohn Heffner return -1; 166190840defSIlpo Järvinen if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) 16625d424d5aSJohn Heffner return 0; 16635d424d5aSJohn Heffner 1664d67c58e9SIlpo Järvinen /* Do we need to wait to drain cwnd? With none in flight, don't stall */ 1665d67c58e9SIlpo Järvinen if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { 1666d67c58e9SIlpo Järvinen if (!tcp_packets_in_flight(tp)) 16675d424d5aSJohn Heffner return -1; 16685d424d5aSJohn Heffner else 16695d424d5aSJohn Heffner return 0; 16705d424d5aSJohn Heffner } 16715d424d5aSJohn Heffner 16725d424d5aSJohn Heffner /* We're allowed to probe. Build it now. */ 16735d424d5aSJohn Heffner if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL) 16745d424d5aSJohn Heffner return -1; 16753ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 16763ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 16775d424d5aSJohn Heffner 1678fe067e8aSDavid S. Miller skb = tcp_send_head(sk); 16795d424d5aSJohn Heffner 16805d424d5aSJohn Heffner TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 16815d424d5aSJohn Heffner TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 16824de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; 16835d424d5aSJohn Heffner TCP_SKB_CB(nskb)->sacked = 0; 16845d424d5aSJohn Heffner nskb->csum = 0; 168584fa7933SPatrick McHardy nskb->ip_summed = skb->ip_summed; 16865d424d5aSJohn Heffner 168750c4817eSIlpo Järvinen tcp_insert_write_queue_before(nskb, skb, sk); 168850c4817eSIlpo Järvinen 16895d424d5aSJohn Heffner len = 0; 1690234b6860SIlpo Järvinen tcp_for_write_queue_from_safe(skb, next, sk) { 16915d424d5aSJohn Heffner copy = min_t(int, skb->len, probe_size - len); 16925d424d5aSJohn Heffner if (nskb->ip_summed) 16935d424d5aSJohn Heffner skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); 16945d424d5aSJohn Heffner else 16955d424d5aSJohn Heffner nskb->csum = skb_copy_and_csum_bits(skb, 0, 1696056834d9SIlpo Järvinen skb_put(nskb, copy), 1697056834d9SIlpo Järvinen copy, nskb->csum); 16985d424d5aSJohn Heffner 16995d424d5aSJohn Heffner if (skb->len <= copy) { 17005d424d5aSJohn Heffner /* We've eaten all the data from this skb. 17015d424d5aSJohn Heffner * Throw it away. */ 17024de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 1703fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 17043ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 17055d424d5aSJohn Heffner } else { 17064de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & 1707a3433f35SChangli Gao ~(TCPHDR_FIN|TCPHDR_PSH); 17085d424d5aSJohn Heffner if (!skb_shinfo(skb)->nr_frags) { 17095d424d5aSJohn Heffner skb_pull(skb, copy); 171084fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 1711056834d9SIlpo Järvinen skb->csum = csum_partial(skb->data, 1712056834d9SIlpo Järvinen skb->len, 0); 17135d424d5aSJohn Heffner } else { 17145d424d5aSJohn Heffner __pskb_trim_head(skb, copy); 17155d424d5aSJohn Heffner tcp_set_skb_tso_segs(sk, skb, mss_now); 17165d424d5aSJohn Heffner } 17175d424d5aSJohn Heffner TCP_SKB_CB(skb)->seq += copy; 17185d424d5aSJohn Heffner } 17195d424d5aSJohn Heffner 17205d424d5aSJohn Heffner len += copy; 1721234b6860SIlpo Järvinen 1722234b6860SIlpo Järvinen if (len >= probe_size) 1723234b6860SIlpo Järvinen break; 17245d424d5aSJohn Heffner } 17255d424d5aSJohn Heffner tcp_init_tso_segs(sk, nskb, nskb->len); 17265d424d5aSJohn Heffner 17275d424d5aSJohn Heffner /* We're ready to send. If this fails, the probe will 17285d424d5aSJohn Heffner * be resegmented into mss-sized pieces by tcp_write_xmit(). */ 17295d424d5aSJohn Heffner TCP_SKB_CB(nskb)->when = tcp_time_stamp; 17305d424d5aSJohn Heffner if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 17315d424d5aSJohn Heffner /* Decrement cwnd here because we are sending 17325d424d5aSJohn Heffner * effectively two packets. */ 17335d424d5aSJohn Heffner tp->snd_cwnd--; 173466f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, nskb); 17355d424d5aSJohn Heffner 17365d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 17370e7b1368SJohn Heffner tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 17380e7b1368SJohn Heffner tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 17395d424d5aSJohn Heffner 17405d424d5aSJohn Heffner return 1; 17415d424d5aSJohn Heffner } 17425d424d5aSJohn Heffner 17435d424d5aSJohn Heffner return -1; 17445d424d5aSJohn Heffner } 17455d424d5aSJohn Heffner 17461da177e4SLinus Torvalds /* This routine writes packets to the network. It advances the 17471da177e4SLinus Torvalds * send_head. This happens as incoming acks open up the remote 17481da177e4SLinus Torvalds * window for us. 17491da177e4SLinus Torvalds * 1750f8269a49SIlpo Järvinen * LARGESEND note: !tcp_urg_mode is overkill, only frames between 1751f8269a49SIlpo Järvinen * snd_up-64k-mss .. snd_up cannot be large. However, taking into 1752f8269a49SIlpo Järvinen * account rare use of URG, this is not a big flaw. 1753f8269a49SIlpo Järvinen * 17541da177e4SLinus Torvalds * Returns 1, if no segments are in flight and we have queued segments, but 17551da177e4SLinus Torvalds * cannot send anything now because of SWS or another problem. 17561da177e4SLinus Torvalds */ 1757d5dd9175SIlpo Järvinen static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 1758d5dd9175SIlpo Järvinen int push_one, gfp_t gfp) 17591da177e4SLinus Torvalds { 17601da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 176192df7b51SDavid S. Miller struct sk_buff *skb; 1762c1b4a7e6SDavid S. Miller unsigned int tso_segs, sent_pkts; 1763c1b4a7e6SDavid S. Miller int cwnd_quota; 17645d424d5aSJohn Heffner int result; 17651da177e4SLinus Torvalds 1766c1b4a7e6SDavid S. Miller sent_pkts = 0; 17675d424d5aSJohn Heffner 1768d5dd9175SIlpo Järvinen if (!push_one) { 17695d424d5aSJohn Heffner /* Do MTU probing. */ 1770d5dd9175SIlpo Järvinen result = tcp_mtu_probe(sk); 1771d5dd9175SIlpo Järvinen if (!result) { 17725d424d5aSJohn Heffner return 0; 17735d424d5aSJohn Heffner } else if (result > 0) { 17745d424d5aSJohn Heffner sent_pkts = 1; 17755d424d5aSJohn Heffner } 1776d5dd9175SIlpo Järvinen } 17775d424d5aSJohn Heffner 1778fe067e8aSDavid S. Miller while ((skb = tcp_send_head(sk))) { 1779c8ac3774SHerbert Xu unsigned int limit; 1780c8ac3774SHerbert Xu 1781b68e9f85SHerbert Xu tso_segs = tcp_init_tso_segs(sk, skb, mss_now); 1782c1b4a7e6SDavid S. Miller BUG_ON(!tso_segs); 1783c1b4a7e6SDavid S. Miller 1784b68e9f85SHerbert Xu cwnd_quota = tcp_cwnd_test(tp, skb); 1785b68e9f85SHerbert Xu if (!cwnd_quota) 1786b68e9f85SHerbert Xu break; 1787b68e9f85SHerbert Xu 1788b68e9f85SHerbert Xu if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) 1789b68e9f85SHerbert Xu break; 1790b68e9f85SHerbert Xu 1791c1b4a7e6SDavid S. Miller if (tso_segs == 1) { 1792aa93466bSDavid S. Miller if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 1793aa93466bSDavid S. Miller (tcp_skb_is_last(sk, skb) ? 1794aa93466bSDavid S. Miller nonagle : TCP_NAGLE_PUSH)))) 1795aa93466bSDavid S. Miller break; 1796c1b4a7e6SDavid S. Miller } else { 1797d5dd9175SIlpo Järvinen if (!push_one && tcp_tso_should_defer(sk, skb)) 1798aa93466bSDavid S. Miller break; 1799c1b4a7e6SDavid S. Miller } 1800aa93466bSDavid S. Miller 1801c8ac3774SHerbert Xu limit = mss_now; 1802f8269a49SIlpo Järvinen if (tso_segs > 1 && !tcp_urg_mode(tp)) 18030e3a4803SIlpo Järvinen limit = tcp_mss_split_point(sk, skb, mss_now, 18040e3a4803SIlpo Järvinen cwnd_quota); 1805c8ac3774SHerbert Xu 1806c8ac3774SHerbert Xu if (skb->len > limit && 1807c4ead4c5SEric Dumazet unlikely(tso_fragment(sk, skb, limit, mss_now, gfp))) 18081da177e4SLinus Torvalds break; 18091da177e4SLinus Torvalds 18101da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 1811c1b4a7e6SDavid S. Miller 1812d5dd9175SIlpo Järvinen if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) 18131da177e4SLinus Torvalds break; 18141da177e4SLinus Torvalds 18151da177e4SLinus Torvalds /* Advance the send_head. This one is sent out. 18161da177e4SLinus Torvalds * This call will increment packets_out. 18171da177e4SLinus Torvalds */ 181866f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 18191da177e4SLinus Torvalds 18201da177e4SLinus Torvalds tcp_minshall_update(tp, mss_now, skb); 1821a262f0cdSNandita Dukkipati sent_pkts += tcp_skb_pcount(skb); 1822d5dd9175SIlpo Järvinen 1823d5dd9175SIlpo Järvinen if (push_one) 1824d5dd9175SIlpo Järvinen break; 18251da177e4SLinus Torvalds } 1826a262f0cdSNandita Dukkipati if (inet_csk(sk)->icsk_ca_state == TCP_CA_Recovery) 1827a262f0cdSNandita Dukkipati tp->prr_out += sent_pkts; 18281da177e4SLinus Torvalds 1829aa93466bSDavid S. Miller if (likely(sent_pkts)) { 18309e412ba7SIlpo Järvinen tcp_cwnd_validate(sk); 18311da177e4SLinus Torvalds return 0; 18321da177e4SLinus Torvalds } 1833fe067e8aSDavid S. Miller return !tp->packets_out && tcp_send_head(sk); 18341da177e4SLinus Torvalds } 18351da177e4SLinus Torvalds 1836a762a980SDavid S. Miller /* Push out any pending frames which were held back due to 1837a762a980SDavid S. Miller * TCP_CORK or attempt at coalescing tiny packets. 1838a762a980SDavid S. Miller * The socket must be locked by the caller. 1839a762a980SDavid S. Miller */ 18409e412ba7SIlpo Järvinen void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 18419e412ba7SIlpo Järvinen int nonagle) 1842a762a980SDavid S. Miller { 1843726e07a8SIlpo Järvinen /* If we are closed, the bytes will have to remain here. 1844726e07a8SIlpo Järvinen * In time closedown will finish, we empty the write queue and 1845726e07a8SIlpo Järvinen * all will be happy. 1846726e07a8SIlpo Järvinen */ 1847726e07a8SIlpo Järvinen if (unlikely(sk->sk_state == TCP_CLOSE)) 1848726e07a8SIlpo Järvinen return; 1849726e07a8SIlpo Järvinen 1850d5dd9175SIlpo Järvinen if (tcp_write_xmit(sk, cur_mss, nonagle, 0, GFP_ATOMIC)) 18519e412ba7SIlpo Järvinen tcp_check_probe_timer(sk); 1852a762a980SDavid S. Miller } 1853a762a980SDavid S. Miller 1854c1b4a7e6SDavid S. Miller /* Send _single_ skb sitting at the send head. This function requires 1855c1b4a7e6SDavid S. Miller * true push pending frames to setup probe timer etc. 1856c1b4a7e6SDavid S. Miller */ 1857c1b4a7e6SDavid S. Miller void tcp_push_one(struct sock *sk, unsigned int mss_now) 1858c1b4a7e6SDavid S. Miller { 1859fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 1860c1b4a7e6SDavid S. Miller 1861c1b4a7e6SDavid S. Miller BUG_ON(!skb || skb->len < mss_now); 1862c1b4a7e6SDavid S. Miller 1863d5dd9175SIlpo Järvinen tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); 1864c1b4a7e6SDavid S. Miller } 1865c1b4a7e6SDavid S. Miller 18661da177e4SLinus Torvalds /* This function returns the amount that we can raise the 18671da177e4SLinus Torvalds * usable window based on the following constraints 18681da177e4SLinus Torvalds * 18691da177e4SLinus Torvalds * 1. The window can never be shrunk once it is offered (RFC 793) 18701da177e4SLinus Torvalds * 2. We limit memory per socket 18711da177e4SLinus Torvalds * 18721da177e4SLinus Torvalds * RFC 1122: 18731da177e4SLinus Torvalds * "the suggested [SWS] avoidance algorithm for the receiver is to keep 18741da177e4SLinus Torvalds * RECV.NEXT + RCV.WIN fixed until: 18751da177e4SLinus Torvalds * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 18761da177e4SLinus Torvalds * 18771da177e4SLinus Torvalds * i.e. don't raise the right edge of the window until you can raise 18781da177e4SLinus Torvalds * it at least MSS bytes. 18791da177e4SLinus Torvalds * 18801da177e4SLinus Torvalds * Unfortunately, the recommended algorithm breaks header prediction, 18811da177e4SLinus Torvalds * since header prediction assumes th->window stays fixed. 18821da177e4SLinus Torvalds * 18831da177e4SLinus Torvalds * Strictly speaking, keeping th->window fixed violates the receiver 18841da177e4SLinus Torvalds * side SWS prevention criteria. The problem is that under this rule 18851da177e4SLinus Torvalds * a stream of single byte packets will cause the right side of the 18861da177e4SLinus Torvalds * window to always advance by a single byte. 18871da177e4SLinus Torvalds * 18881da177e4SLinus Torvalds * Of course, if the sender implements sender side SWS prevention 18891da177e4SLinus Torvalds * then this will not be a problem. 18901da177e4SLinus Torvalds * 18911da177e4SLinus Torvalds * BSD seems to make the following compromise: 18921da177e4SLinus Torvalds * 18931da177e4SLinus Torvalds * If the free space is less than the 1/4 of the maximum 18941da177e4SLinus Torvalds * space available and the free space is less than 1/2 mss, 18951da177e4SLinus Torvalds * then set the window to 0. 18961da177e4SLinus Torvalds * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 18971da177e4SLinus Torvalds * Otherwise, just prevent the window from shrinking 18981da177e4SLinus Torvalds * and from being larger than the largest representable value. 18991da177e4SLinus Torvalds * 19001da177e4SLinus Torvalds * This prevents incremental opening of the window in the regime 19011da177e4SLinus Torvalds * where TCP is limited by the speed of the reader side taking 19021da177e4SLinus Torvalds * data out of the TCP receive queue. It does nothing about 19031da177e4SLinus Torvalds * those cases where the window is constrained on the sender side 19041da177e4SLinus Torvalds * because the pipeline is full. 19051da177e4SLinus Torvalds * 19061da177e4SLinus Torvalds * BSD also seems to "accidentally" limit itself to windows that are a 19071da177e4SLinus Torvalds * multiple of MSS, at least until the free space gets quite small. 19081da177e4SLinus Torvalds * This would appear to be a side effect of the mbuf implementation. 19091da177e4SLinus Torvalds * Combining these two algorithms results in the observed behavior 19101da177e4SLinus Torvalds * of having a fixed window size at almost all times. 19111da177e4SLinus Torvalds * 19121da177e4SLinus Torvalds * Below we obtain similar behavior by forcing the offered window to 19131da177e4SLinus Torvalds * a multiple of the mss when it is feasible to do so. 19141da177e4SLinus Torvalds * 19151da177e4SLinus Torvalds * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 19161da177e4SLinus Torvalds * Regular options like TIMESTAMP are taken into account. 19171da177e4SLinus Torvalds */ 19181da177e4SLinus Torvalds u32 __tcp_select_window(struct sock *sk) 19191da177e4SLinus Torvalds { 1920463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 19211da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1922caa20d9aSStephen Hemminger /* MSS for the peer's data. Previous versions used mss_clamp 19231da177e4SLinus Torvalds * here. I don't know if the value based on our guesses 19241da177e4SLinus Torvalds * of peer's MSS is better for the performance. It's more correct 19251da177e4SLinus Torvalds * but may be worse for the performance because of rcv_mss 19261da177e4SLinus Torvalds * fluctuations. --SAW 1998/11/1 19271da177e4SLinus Torvalds */ 1928463c84b9SArnaldo Carvalho de Melo int mss = icsk->icsk_ack.rcv_mss; 19291da177e4SLinus Torvalds int free_space = tcp_space(sk); 19301da177e4SLinus Torvalds int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk)); 19311da177e4SLinus Torvalds int window; 19321da177e4SLinus Torvalds 19331da177e4SLinus Torvalds if (mss > full_space) 19341da177e4SLinus Torvalds mss = full_space; 19351da177e4SLinus Torvalds 1936b92edbe0SEric Dumazet if (free_space < (full_space >> 1)) { 1937463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.quick = 0; 19381da177e4SLinus Torvalds 1939180d8cd9SGlauber Costa if (sk_under_memory_pressure(sk)) 1940056834d9SIlpo Järvinen tp->rcv_ssthresh = min(tp->rcv_ssthresh, 1941056834d9SIlpo Järvinen 4U * tp->advmss); 19421da177e4SLinus Torvalds 19431da177e4SLinus Torvalds if (free_space < mss) 19441da177e4SLinus Torvalds return 0; 19451da177e4SLinus Torvalds } 19461da177e4SLinus Torvalds 19471da177e4SLinus Torvalds if (free_space > tp->rcv_ssthresh) 19481da177e4SLinus Torvalds free_space = tp->rcv_ssthresh; 19491da177e4SLinus Torvalds 19501da177e4SLinus Torvalds /* Don't do rounding if we are using window scaling, since the 19511da177e4SLinus Torvalds * scaled window will not line up with the MSS boundary anyway. 19521da177e4SLinus Torvalds */ 19531da177e4SLinus Torvalds window = tp->rcv_wnd; 19541da177e4SLinus Torvalds if (tp->rx_opt.rcv_wscale) { 19551da177e4SLinus Torvalds window = free_space; 19561da177e4SLinus Torvalds 19571da177e4SLinus Torvalds /* Advertise enough space so that it won't get scaled away. 19581da177e4SLinus Torvalds * Import case: prevent zero window announcement if 19591da177e4SLinus Torvalds * 1<<rcv_wscale > mss. 19601da177e4SLinus Torvalds */ 19611da177e4SLinus Torvalds if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window) 19621da177e4SLinus Torvalds window = (((window >> tp->rx_opt.rcv_wscale) + 1) 19631da177e4SLinus Torvalds << tp->rx_opt.rcv_wscale); 19641da177e4SLinus Torvalds } else { 19651da177e4SLinus Torvalds /* Get the largest window that is a nice multiple of mss. 19661da177e4SLinus Torvalds * Window clamp already applied above. 19671da177e4SLinus Torvalds * If our current window offering is within 1 mss of the 19681da177e4SLinus Torvalds * free space we just keep it. This prevents the divide 19691da177e4SLinus Torvalds * and multiply from happening most of the time. 19701da177e4SLinus Torvalds * We also don't do any window rounding when the free space 19711da177e4SLinus Torvalds * is too small. 19721da177e4SLinus Torvalds */ 19731da177e4SLinus Torvalds if (window <= free_space - mss || window > free_space) 19741da177e4SLinus Torvalds window = (free_space / mss) * mss; 197584565070SJohn Heffner else if (mss == full_space && 1976b92edbe0SEric Dumazet free_space > window + (full_space >> 1)) 197784565070SJohn Heffner window = free_space; 19781da177e4SLinus Torvalds } 19791da177e4SLinus Torvalds 19801da177e4SLinus Torvalds return window; 19811da177e4SLinus Torvalds } 19821da177e4SLinus Torvalds 19834a17fc3aSIlpo Järvinen /* Collapses two adjacent SKB's during retransmission. */ 19844a17fc3aSIlpo Järvinen static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) 19851da177e4SLinus Torvalds { 19861da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1987fe067e8aSDavid S. Miller struct sk_buff *next_skb = tcp_write_queue_next(sk, skb); 1988058dc334SIlpo Järvinen int skb_size, next_skb_size; 19891da177e4SLinus Torvalds 1990058dc334SIlpo Järvinen skb_size = skb->len; 1991058dc334SIlpo Järvinen next_skb_size = next_skb->len; 19921da177e4SLinus Torvalds 1993058dc334SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); 19941da177e4SLinus Torvalds 19956859d494SIlpo Järvinen tcp_highest_sack_combine(sk, next_skb, skb); 1996a6963a6bSIlpo Järvinen 1997fe067e8aSDavid S. Miller tcp_unlink_write_queue(next_skb, sk); 19981da177e4SLinus Torvalds 1999058dc334SIlpo Järvinen skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size), 20001a4e2d09SArnaldo Carvalho de Melo next_skb_size); 20011da177e4SLinus Torvalds 200252d570aaSJarek Poplawski if (next_skb->ip_summed == CHECKSUM_PARTIAL) 200352d570aaSJarek Poplawski skb->ip_summed = CHECKSUM_PARTIAL; 20041da177e4SLinus Torvalds 200584fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 20061da177e4SLinus Torvalds skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); 20071da177e4SLinus Torvalds 20081da177e4SLinus Torvalds /* Update sequence range on original skb. */ 20091da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 20101da177e4SLinus Torvalds 2011e6c7d085SIlpo Järvinen /* Merge over control information. This moves PSH/FIN etc. over */ 20124de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; 20131da177e4SLinus Torvalds 20141da177e4SLinus Torvalds /* All done, get rid of second SKB and account for it so 20151da177e4SLinus Torvalds * packet counting does not break. 20161da177e4SLinus Torvalds */ 20174828e7f4SIlpo Järvinen TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; 2018b7689205SIlpo Järvinen 2019b7689205SIlpo Järvinen /* changed transmit queue under us so clear hints */ 2020ef9da47cSIlpo Järvinen tcp_clear_retrans_hints_partial(tp); 2021ef9da47cSIlpo Järvinen if (next_skb == tp->retransmit_skb_hint) 2022ef9da47cSIlpo Järvinen tp->retransmit_skb_hint = skb; 2023b7689205SIlpo Järvinen 2024797108d1SIlpo Järvinen tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); 2025797108d1SIlpo Järvinen 20263ab224beSHideo Aoki sk_wmem_free_skb(sk, next_skb); 20271da177e4SLinus Torvalds } 20281da177e4SLinus Torvalds 202967edfef7SAndi Kleen /* Check if coalescing SKBs is legal. */ 2030cf533ea5SEric Dumazet static int tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) 20314a17fc3aSIlpo Järvinen { 20324a17fc3aSIlpo Järvinen if (tcp_skb_pcount(skb) > 1) 20334a17fc3aSIlpo Järvinen return 0; 20344a17fc3aSIlpo Järvinen /* TODO: SACK collapsing could be used to remove this condition */ 20354a17fc3aSIlpo Järvinen if (skb_shinfo(skb)->nr_frags != 0) 20364a17fc3aSIlpo Järvinen return 0; 20374a17fc3aSIlpo Järvinen if (skb_cloned(skb)) 20384a17fc3aSIlpo Järvinen return 0; 20394a17fc3aSIlpo Järvinen if (skb == tcp_send_head(sk)) 20404a17fc3aSIlpo Järvinen return 0; 20414a17fc3aSIlpo Järvinen /* Some heurestics for collapsing over SACK'd could be invented */ 20424a17fc3aSIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 20434a17fc3aSIlpo Järvinen return 0; 20444a17fc3aSIlpo Järvinen 20454a17fc3aSIlpo Järvinen return 1; 20464a17fc3aSIlpo Järvinen } 20474a17fc3aSIlpo Järvinen 204867edfef7SAndi Kleen /* Collapse packets in the retransmit queue to make to create 204967edfef7SAndi Kleen * less packets on the wire. This is only done on retransmission. 205067edfef7SAndi Kleen */ 20514a17fc3aSIlpo Järvinen static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, 20524a17fc3aSIlpo Järvinen int space) 20534a17fc3aSIlpo Järvinen { 20544a17fc3aSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 20554a17fc3aSIlpo Järvinen struct sk_buff *skb = to, *tmp; 20564a17fc3aSIlpo Järvinen int first = 1; 20574a17fc3aSIlpo Järvinen 20584a17fc3aSIlpo Järvinen if (!sysctl_tcp_retrans_collapse) 20594a17fc3aSIlpo Järvinen return; 20604de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 20614a17fc3aSIlpo Järvinen return; 20624a17fc3aSIlpo Järvinen 20634a17fc3aSIlpo Järvinen tcp_for_write_queue_from_safe(skb, tmp, sk) { 20644a17fc3aSIlpo Järvinen if (!tcp_can_collapse(sk, skb)) 20654a17fc3aSIlpo Järvinen break; 20664a17fc3aSIlpo Järvinen 20674a17fc3aSIlpo Järvinen space -= skb->len; 20684a17fc3aSIlpo Järvinen 20694a17fc3aSIlpo Järvinen if (first) { 20704a17fc3aSIlpo Järvinen first = 0; 20714a17fc3aSIlpo Järvinen continue; 20724a17fc3aSIlpo Järvinen } 20734a17fc3aSIlpo Järvinen 20744a17fc3aSIlpo Järvinen if (space < 0) 20754a17fc3aSIlpo Järvinen break; 20764a17fc3aSIlpo Järvinen /* Punt if not enough space exists in the first SKB for 20774a17fc3aSIlpo Järvinen * the data in the second 20784a17fc3aSIlpo Järvinen */ 2079a21d4572SEric Dumazet if (skb->len > skb_availroom(to)) 20804a17fc3aSIlpo Järvinen break; 20814a17fc3aSIlpo Järvinen 20824a17fc3aSIlpo Järvinen if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) 20834a17fc3aSIlpo Järvinen break; 20844a17fc3aSIlpo Järvinen 20854a17fc3aSIlpo Järvinen tcp_collapse_retrans(sk, to); 20864a17fc3aSIlpo Järvinen } 20874a17fc3aSIlpo Järvinen } 20884a17fc3aSIlpo Järvinen 20891da177e4SLinus Torvalds /* This retransmits one SKB. Policy decisions and retransmit queue 20901da177e4SLinus Torvalds * state updates are done by the caller. Returns non-zero if an 20911da177e4SLinus Torvalds * error occurred which prevented the send. 20921da177e4SLinus Torvalds */ 20931da177e4SLinus Torvalds int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) 20941da177e4SLinus Torvalds { 20951da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 20965d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 20977d227cd2SSridhar Samudrala unsigned int cur_mss; 20981da177e4SLinus Torvalds int err; 20991da177e4SLinus Torvalds 21005d424d5aSJohn Heffner /* Inconslusive MTU probe */ 21015d424d5aSJohn Heffner if (icsk->icsk_mtup.probe_size) { 21025d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 21035d424d5aSJohn Heffner } 21045d424d5aSJohn Heffner 21051da177e4SLinus Torvalds /* Do not sent more than we queued. 1/4 is reserved for possible 2106caa20d9aSStephen Hemminger * copying overhead: fragmentation, tunneling, mangling etc. 21071da177e4SLinus Torvalds */ 21081da177e4SLinus Torvalds if (atomic_read(&sk->sk_wmem_alloc) > 21091da177e4SLinus Torvalds min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf)) 21101da177e4SLinus Torvalds return -EAGAIN; 21111da177e4SLinus Torvalds 21121da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 21131da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 21141da177e4SLinus Torvalds BUG(); 21151da177e4SLinus Torvalds if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 21161da177e4SLinus Torvalds return -ENOMEM; 21171da177e4SLinus Torvalds } 21181da177e4SLinus Torvalds 21197d227cd2SSridhar Samudrala if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 21207d227cd2SSridhar Samudrala return -EHOSTUNREACH; /* Routing failure or similar. */ 21217d227cd2SSridhar Samudrala 21220c54b85fSIlpo Järvinen cur_mss = tcp_current_mss(sk); 21237d227cd2SSridhar Samudrala 21241da177e4SLinus Torvalds /* If receiver has shrunk his window, and skb is out of 21251da177e4SLinus Torvalds * new window, do not retransmit it. The exception is the 21261da177e4SLinus Torvalds * case, when window is shrunk to zero. In this case 21271da177e4SLinus Torvalds * our retransmit serves as a zero window probe. 21281da177e4SLinus Torvalds */ 21299d4fb27dSJoe Perches if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && 21309d4fb27dSJoe Perches TCP_SKB_CB(skb)->seq != tp->snd_una) 21311da177e4SLinus Torvalds return -EAGAIN; 21321da177e4SLinus Torvalds 21331da177e4SLinus Torvalds if (skb->len > cur_mss) { 2134846998aeSDavid S. Miller if (tcp_fragment(sk, skb, cur_mss, cur_mss)) 21351da177e4SLinus Torvalds return -ENOMEM; /* We'll try again later. */ 213602276f3cSIlpo Järvinen } else { 21379eb9362eSIlpo Järvinen int oldpcount = tcp_skb_pcount(skb); 21389eb9362eSIlpo Järvinen 21399eb9362eSIlpo Järvinen if (unlikely(oldpcount > 1)) { 214002276f3cSIlpo Järvinen tcp_init_tso_segs(sk, skb, cur_mss); 21419eb9362eSIlpo Järvinen tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb)); 21429eb9362eSIlpo Järvinen } 21431da177e4SLinus Torvalds } 21441da177e4SLinus Torvalds 21451da177e4SLinus Torvalds tcp_retrans_try_collapse(sk, skb, cur_mss); 21461da177e4SLinus Torvalds 21471da177e4SLinus Torvalds /* Some Solaris stacks overoptimize and ignore the FIN on a 21481da177e4SLinus Torvalds * retransmit when old data is attached. So strip it off 21491da177e4SLinus Torvalds * since it is cheap to do so and saves bytes on the network. 21501da177e4SLinus Torvalds */ 21511da177e4SLinus Torvalds if (skb->len > 0 && 21524de075e0SEric Dumazet (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && 21531da177e4SLinus Torvalds tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) { 21541da177e4SLinus Torvalds if (!pskb_trim(skb, 0)) { 2155e870a8efSIlpo Järvinen /* Reuse, even though it does some unnecessary work */ 2156e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1, 21574de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags); 21581da177e4SLinus Torvalds skb->ip_summed = CHECKSUM_NONE; 21591da177e4SLinus Torvalds } 21601da177e4SLinus Torvalds } 21611da177e4SLinus Torvalds 21621da177e4SLinus Torvalds /* Make a copy, if the first transmission SKB clone we made 21631da177e4SLinus Torvalds * is still in somebody's hands, else make a clone. 21641da177e4SLinus Torvalds */ 21651da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 21661da177e4SLinus Torvalds 2167117632e6SEric Dumazet /* make sure skb->data is aligned on arches that require it */ 2168117632e6SEric Dumazet if (unlikely(NET_IP_ALIGN && ((unsigned long)skb->data & 3))) { 2169117632e6SEric Dumazet struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER, 2170117632e6SEric Dumazet GFP_ATOMIC); 2171117632e6SEric Dumazet err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) : 2172117632e6SEric Dumazet -ENOBUFS; 2173117632e6SEric Dumazet } else { 2174dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 2175117632e6SEric Dumazet } 21761da177e4SLinus Torvalds 21771da177e4SLinus Torvalds if (err == 0) { 21781da177e4SLinus Torvalds /* Update global TCP statistics. */ 217981cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); 21801da177e4SLinus Torvalds 21811da177e4SLinus Torvalds tp->total_retrans++; 21821da177e4SLinus Torvalds 21831da177e4SLinus Torvalds #if FASTRETRANS_DEBUG > 0 21841da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 21851da177e4SLinus Torvalds if (net_ratelimit()) 21861da177e4SLinus Torvalds printk(KERN_DEBUG "retrans_out leaked.\n"); 21871da177e4SLinus Torvalds } 21881da177e4SLinus Torvalds #endif 2189b08d6cb2SIlpo Järvinen if (!tp->retrans_out) 2190b08d6cb2SIlpo Järvinen tp->lost_retrans_low = tp->snd_nxt; 21911da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 21921da177e4SLinus Torvalds tp->retrans_out += tcp_skb_pcount(skb); 21931da177e4SLinus Torvalds 21941da177e4SLinus Torvalds /* Save stamp of the first retransmit. */ 21951da177e4SLinus Torvalds if (!tp->retrans_stamp) 21961da177e4SLinus Torvalds tp->retrans_stamp = TCP_SKB_CB(skb)->when; 21971da177e4SLinus Torvalds 2198c24f691bSYuchung Cheng tp->undo_retrans += tcp_skb_pcount(skb); 21991da177e4SLinus Torvalds 22001da177e4SLinus Torvalds /* snd_nxt is stored to detect loss of retransmitted segment, 22011da177e4SLinus Torvalds * see tcp_input.c tcp_sacktag_write_queue(). 22021da177e4SLinus Torvalds */ 22031da177e4SLinus Torvalds TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt; 22041da177e4SLinus Torvalds } 22051da177e4SLinus Torvalds return err; 22061da177e4SLinus Torvalds } 22071da177e4SLinus Torvalds 220867edfef7SAndi Kleen /* Check if we forward retransmits are possible in the current 220967edfef7SAndi Kleen * window/congestion state. 221067edfef7SAndi Kleen */ 2211b5afe7bcSIlpo Järvinen static int tcp_can_forward_retransmit(struct sock *sk) 2212b5afe7bcSIlpo Järvinen { 2213b5afe7bcSIlpo Järvinen const struct inet_connection_sock *icsk = inet_csk(sk); 2214cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 2215b5afe7bcSIlpo Järvinen 2216b5afe7bcSIlpo Järvinen /* Forward retransmissions are possible only during Recovery. */ 2217b5afe7bcSIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Recovery) 2218b5afe7bcSIlpo Järvinen return 0; 2219b5afe7bcSIlpo Järvinen 2220b5afe7bcSIlpo Järvinen /* No forward retransmissions in Reno are possible. */ 2221b5afe7bcSIlpo Järvinen if (tcp_is_reno(tp)) 2222b5afe7bcSIlpo Järvinen return 0; 2223b5afe7bcSIlpo Järvinen 2224b5afe7bcSIlpo Järvinen /* Yeah, we have to make difficult choice between forward transmission 2225b5afe7bcSIlpo Järvinen * and retransmission... Both ways have their merits... 2226b5afe7bcSIlpo Järvinen * 2227b5afe7bcSIlpo Järvinen * For now we do not retransmit anything, while we have some new 2228b5afe7bcSIlpo Järvinen * segments to send. In the other cases, follow rule 3 for 2229b5afe7bcSIlpo Järvinen * NextSeg() specified in RFC3517. 2230b5afe7bcSIlpo Järvinen */ 2231b5afe7bcSIlpo Järvinen 2232b5afe7bcSIlpo Järvinen if (tcp_may_send_now(sk)) 2233b5afe7bcSIlpo Järvinen return 0; 2234b5afe7bcSIlpo Järvinen 2235b5afe7bcSIlpo Järvinen return 1; 2236b5afe7bcSIlpo Järvinen } 2237b5afe7bcSIlpo Järvinen 22381da177e4SLinus Torvalds /* This gets called after a retransmit timeout, and the initially 22391da177e4SLinus Torvalds * retransmitted data is acknowledged. It tries to continue 22401da177e4SLinus Torvalds * resending the rest of the retransmit queue, until either 22411da177e4SLinus Torvalds * we've sent it all or the congestion window limit is reached. 22421da177e4SLinus Torvalds * If doing SACK, the first ACK which comes back for a timeout 22431da177e4SLinus Torvalds * based retransmit packet might feed us FACK information again. 22441da177e4SLinus Torvalds * If so, we use it to avoid unnecessarily retransmissions. 22451da177e4SLinus Torvalds */ 22461da177e4SLinus Torvalds void tcp_xmit_retransmit_queue(struct sock *sk) 22471da177e4SLinus Torvalds { 22486687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 22491da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 22501da177e4SLinus Torvalds struct sk_buff *skb; 22510e1c54c2SIlpo Järvinen struct sk_buff *hole = NULL; 2252618d9f25SIlpo Järvinen u32 last_lost; 225361eb55f4SIlpo Järvinen int mib_idx; 22540e1c54c2SIlpo Järvinen int fwd_rexmitting = 0; 22556a438bbeSStephen Hemminger 225645e77d31SIlpo Järvinen if (!tp->packets_out) 225745e77d31SIlpo Järvinen return; 225845e77d31SIlpo Järvinen 225908ebd172SIlpo Järvinen if (!tp->lost_out) 226008ebd172SIlpo Järvinen tp->retransmit_high = tp->snd_una; 226108ebd172SIlpo Järvinen 2262618d9f25SIlpo Järvinen if (tp->retransmit_skb_hint) { 22636a438bbeSStephen Hemminger skb = tp->retransmit_skb_hint; 2264618d9f25SIlpo Järvinen last_lost = TCP_SKB_CB(skb)->end_seq; 2265618d9f25SIlpo Järvinen if (after(last_lost, tp->retransmit_high)) 2266618d9f25SIlpo Järvinen last_lost = tp->retransmit_high; 2267618d9f25SIlpo Järvinen } else { 2268fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 2269618d9f25SIlpo Järvinen last_lost = tp->snd_una; 2270618d9f25SIlpo Järvinen } 22711da177e4SLinus Torvalds 2272fe067e8aSDavid S. Miller tcp_for_write_queue_from(skb, sk) { 22731da177e4SLinus Torvalds __u8 sacked = TCP_SKB_CB(skb)->sacked; 22741da177e4SLinus Torvalds 2275fe067e8aSDavid S. Miller if (skb == tcp_send_head(sk)) 2276fe067e8aSDavid S. Miller break; 22776a438bbeSStephen Hemminger /* we could do better than to assign each time */ 22780e1c54c2SIlpo Järvinen if (hole == NULL) 22796a438bbeSStephen Hemminger tp->retransmit_skb_hint = skb; 22806a438bbeSStephen Hemminger 22811da177e4SLinus Torvalds /* Assume this retransmit will generate 22821da177e4SLinus Torvalds * only one packet for congestion window 22831da177e4SLinus Torvalds * calculation purposes. This works because 22841da177e4SLinus Torvalds * tcp_retransmit_skb() will chop up the 22851da177e4SLinus Torvalds * packet to be MSS sized and all the 22861da177e4SLinus Torvalds * packet counting works out. 22871da177e4SLinus Torvalds */ 22881da177e4SLinus Torvalds if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) 22891da177e4SLinus Torvalds return; 22900e1c54c2SIlpo Järvinen 22910e1c54c2SIlpo Järvinen if (fwd_rexmitting) { 22920e1c54c2SIlpo Järvinen begin_fwd: 22930e1c54c2SIlpo Järvinen if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp))) 2294006f582cSIlpo Järvinen break; 22950e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFORWARDRETRANS; 22960e1c54c2SIlpo Järvinen 22970e1c54c2SIlpo Järvinen } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) { 2298618d9f25SIlpo Järvinen tp->retransmit_high = last_lost; 22990e1c54c2SIlpo Järvinen if (!tcp_can_forward_retransmit(sk)) 23000e1c54c2SIlpo Järvinen break; 23010e1c54c2SIlpo Järvinen /* Backtrack if necessary to non-L'ed skb */ 23020e1c54c2SIlpo Järvinen if (hole != NULL) { 23030e1c54c2SIlpo Järvinen skb = hole; 23040e1c54c2SIlpo Järvinen hole = NULL; 23050e1c54c2SIlpo Järvinen } 23060e1c54c2SIlpo Järvinen fwd_rexmitting = 1; 23070e1c54c2SIlpo Järvinen goto begin_fwd; 23080e1c54c2SIlpo Järvinen 23090e1c54c2SIlpo Järvinen } else if (!(sacked & TCPCB_LOST)) { 2310ac11ba75SIlpo Järvinen if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) 23110e1c54c2SIlpo Järvinen hole = skb; 231261eb55f4SIlpo Järvinen continue; 23131da177e4SLinus Torvalds 23140e1c54c2SIlpo Järvinen } else { 2315618d9f25SIlpo Järvinen last_lost = TCP_SKB_CB(skb)->end_seq; 23160e1c54c2SIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Loss) 23170e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFASTRETRANS; 23180e1c54c2SIlpo Järvinen else 23190e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; 23200e1c54c2SIlpo Järvinen } 23210e1c54c2SIlpo Järvinen 23220e1c54c2SIlpo Järvinen if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) 232361eb55f4SIlpo Järvinen continue; 232440b215e5SPavel Emelyanov 232509e9b813SEric Dumazet if (tcp_retransmit_skb(sk, skb)) { 232609e9b813SEric Dumazet NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL); 23271da177e4SLinus Torvalds return; 232809e9b813SEric Dumazet } 2329de0744afSPavel Emelyanov NET_INC_STATS_BH(sock_net(sk), mib_idx); 23301da177e4SLinus Torvalds 2331a262f0cdSNandita Dukkipati if (inet_csk(sk)->icsk_ca_state == TCP_CA_Recovery) 2332a262f0cdSNandita Dukkipati tp->prr_out += tcp_skb_pcount(skb); 2333a262f0cdSNandita Dukkipati 2334fe067e8aSDavid S. Miller if (skb == tcp_write_queue_head(sk)) 2335463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 23363f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, 23373f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 23381da177e4SLinus Torvalds } 23391da177e4SLinus Torvalds } 23401da177e4SLinus Torvalds 23411da177e4SLinus Torvalds /* Send a fin. The caller locks the socket for us. This cannot be 23421da177e4SLinus Torvalds * allowed to fail queueing a FIN frame under any circumstances. 23431da177e4SLinus Torvalds */ 23441da177e4SLinus Torvalds void tcp_send_fin(struct sock *sk) 23451da177e4SLinus Torvalds { 23461da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2347fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_write_queue_tail(sk); 23481da177e4SLinus Torvalds int mss_now; 23491da177e4SLinus Torvalds 23501da177e4SLinus Torvalds /* Optimization, tack on the FIN if we have a queue of 23511da177e4SLinus Torvalds * unsent frames. But be careful about outgoing SACKS 23521da177e4SLinus Torvalds * and IP options. 23531da177e4SLinus Torvalds */ 23540c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 23551da177e4SLinus Torvalds 2356fe067e8aSDavid S. Miller if (tcp_send_head(sk) != NULL) { 23574de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_FIN; 23581da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq++; 23591da177e4SLinus Torvalds tp->write_seq++; 23601da177e4SLinus Torvalds } else { 23611da177e4SLinus Torvalds /* Socket is locked, keep trying until memory is available. */ 23621da177e4SLinus Torvalds for (;;) { 2363aa133076SWu Fengguang skb = alloc_skb_fclone(MAX_TCP_HEADER, 2364aa133076SWu Fengguang sk->sk_allocation); 23651da177e4SLinus Torvalds if (skb) 23661da177e4SLinus Torvalds break; 23671da177e4SLinus Torvalds yield(); 23681da177e4SLinus Torvalds } 23691da177e4SLinus Torvalds 23701da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 23711da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 23721da177e4SLinus Torvalds /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 2373e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tp->write_seq, 2374a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_FIN); 23751da177e4SLinus Torvalds tcp_queue_skb(sk, skb); 23761da177e4SLinus Torvalds } 23779e412ba7SIlpo Järvinen __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF); 23781da177e4SLinus Torvalds } 23791da177e4SLinus Torvalds 23801da177e4SLinus Torvalds /* We get here when a process closes a file descriptor (either due to 23811da177e4SLinus Torvalds * an explicit close() or as a byproduct of exit()'ing) and there 23821da177e4SLinus Torvalds * was unread data in the receive queue. This behavior is recommended 238365bb723cSGerrit Renker * by RFC 2525, section 2.17. -DaveM 23841da177e4SLinus Torvalds */ 2385dd0fc66fSAl Viro void tcp_send_active_reset(struct sock *sk, gfp_t priority) 23861da177e4SLinus Torvalds { 23871da177e4SLinus Torvalds struct sk_buff *skb; 23881da177e4SLinus Torvalds 23891da177e4SLinus Torvalds /* NOTE: No TCP options attached and we never retransmit this. */ 23901da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, priority); 23911da177e4SLinus Torvalds if (!skb) { 23924e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 23931da177e4SLinus Torvalds return; 23941da177e4SLinus Torvalds } 23951da177e4SLinus Torvalds 23961da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 23971da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 2398e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), 2399a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_RST); 24001da177e4SLinus Torvalds /* Send it off. */ 24011da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2402dfb4b9dcSDavid S. Miller if (tcp_transmit_skb(sk, skb, 0, priority)) 24034e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 240426af65cbSSridhar Samudrala 240581cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); 24061da177e4SLinus Torvalds } 24071da177e4SLinus Torvalds 240867edfef7SAndi Kleen /* Send a crossed SYN-ACK during socket establishment. 240967edfef7SAndi Kleen * WARNING: This routine must only be called when we have already sent 24101da177e4SLinus Torvalds * a SYN packet that crossed the incoming SYN that caused this routine 24111da177e4SLinus Torvalds * to get called. If this assumption fails then the initial rcv_wnd 24121da177e4SLinus Torvalds * and rcv_wscale values will not be correct. 24131da177e4SLinus Torvalds */ 24141da177e4SLinus Torvalds int tcp_send_synack(struct sock *sk) 24151da177e4SLinus Torvalds { 24161da177e4SLinus Torvalds struct sk_buff *skb; 24171da177e4SLinus Torvalds 2418fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 24194de075e0SEric Dumazet if (skb == NULL || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 24201da177e4SLinus Torvalds printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n"); 24211da177e4SLinus Torvalds return -EFAULT; 24221da177e4SLinus Torvalds } 24234de075e0SEric Dumazet if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { 24241da177e4SLinus Torvalds if (skb_cloned(skb)) { 24251da177e4SLinus Torvalds struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC); 24261da177e4SLinus Torvalds if (nskb == NULL) 24271da177e4SLinus Torvalds return -ENOMEM; 2428fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 24291da177e4SLinus Torvalds skb_header_release(nskb); 2430fe067e8aSDavid S. Miller __tcp_add_write_queue_head(sk, nskb); 24313ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 24323ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 24333ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 24341da177e4SLinus Torvalds skb = nskb; 24351da177e4SLinus Torvalds } 24361da177e4SLinus Torvalds 24374de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; 24381da177e4SLinus Torvalds TCP_ECN_send_synack(tcp_sk(sk), skb); 24391da177e4SLinus Torvalds } 24401da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2441dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 24421da177e4SLinus Torvalds } 24431da177e4SLinus Torvalds 244467edfef7SAndi Kleen /* Prepare a SYN-ACK. */ 24451da177e4SLinus Torvalds struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst, 2446e6b4d113SWilliam Allen Simpson struct request_sock *req, 2447e6b4d113SWilliam Allen Simpson struct request_values *rvp) 24481da177e4SLinus Torvalds { 2449bd0388aeSWilliam Allen Simpson struct tcp_out_options opts; 24504957faadSWilliam Allen Simpson struct tcp_extend_values *xvp = tcp_xv(rvp); 24512e6599cbSArnaldo Carvalho de Melo struct inet_request_sock *ireq = inet_rsk(req); 24521da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 245328b2774aSEric Dumazet const struct tcp_cookie_values *cvp = tp->cookie_values; 24541da177e4SLinus Torvalds struct tcphdr *th; 24551da177e4SLinus Torvalds struct sk_buff *skb; 2456cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 2457bd0388aeSWilliam Allen Simpson int tcp_header_size; 2458f5fff5dcSTom Quetchenbach int mss; 245928b2774aSEric Dumazet int s_data_desired = 0; 24601da177e4SLinus Torvalds 246128b2774aSEric Dumazet if (cvp != NULL && cvp->s_data_constant && cvp->s_data_desired) 246228b2774aSEric Dumazet s_data_desired = cvp->s_data_desired; 246328b2774aSEric Dumazet skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15 + s_data_desired, 1, GFP_ATOMIC); 24641da177e4SLinus Torvalds if (skb == NULL) 24651da177e4SLinus Torvalds return NULL; 24661da177e4SLinus Torvalds 24671da177e4SLinus Torvalds /* Reserve space for headers. */ 24681da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 24691da177e4SLinus Torvalds 2470adf30907SEric Dumazet skb_dst_set(skb, dst_clone(dst)); 24711da177e4SLinus Torvalds 24720dbaee3bSDavid S. Miller mss = dst_metric_advmss(dst); 2473f5fff5dcSTom Quetchenbach if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss) 2474f5fff5dcSTom Quetchenbach mss = tp->rx_opt.user_mss; 2475f5fff5dcSTom Quetchenbach 247633ad798cSAdam Langley if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */ 247733ad798cSAdam Langley __u8 rcv_wscale; 247833ad798cSAdam Langley /* Set this up on the first call only */ 247933ad798cSAdam Langley req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW); 2480e88c64f0SHagen Paul Pfeifer 2481e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 2482e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 2483e88c64f0SHagen Paul Pfeifer (req->window_clamp > tcp_full_space(sk) || req->window_clamp == 0)) 2484e88c64f0SHagen Paul Pfeifer req->window_clamp = tcp_full_space(sk); 2485e88c64f0SHagen Paul Pfeifer 248633ad798cSAdam Langley /* tcp_full_space because it is guaranteed to be the first packet */ 248733ad798cSAdam Langley tcp_select_initial_window(tcp_full_space(sk), 2488f5fff5dcSTom Quetchenbach mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0), 248933ad798cSAdam Langley &req->rcv_wnd, 249033ad798cSAdam Langley &req->window_clamp, 249133ad798cSAdam Langley ireq->wscale_ok, 249231d12926Slaurent chavey &rcv_wscale, 249331d12926Slaurent chavey dst_metric(dst, RTAX_INITRWND)); 249433ad798cSAdam Langley ireq->rcv_wscale = rcv_wscale; 249533ad798cSAdam Langley } 2496cfb6eeb4SYOSHIFUJI Hideaki 249733ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 24988b5f12d0SFlorian Westphal #ifdef CONFIG_SYN_COOKIES 24998b5f12d0SFlorian Westphal if (unlikely(req->cookie_ts)) 25008b5f12d0SFlorian Westphal TCP_SKB_CB(skb)->when = cookie_init_timestamp(req); 25018b5f12d0SFlorian Westphal else 25028b5f12d0SFlorian Westphal #endif 250333ad798cSAdam Langley TCP_SKB_CB(skb)->when = tcp_time_stamp; 2504f5fff5dcSTom Quetchenbach tcp_header_size = tcp_synack_options(sk, req, mss, 25054957faadSWilliam Allen Simpson skb, &opts, &md5, xvp) 25064957faadSWilliam Allen Simpson + sizeof(*th); 250733ad798cSAdam Langley 2508aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 2509aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 25101da177e4SLinus Torvalds 2511aa8223c7SArnaldo Carvalho de Melo th = tcp_hdr(skb); 25121da177e4SLinus Torvalds memset(th, 0, sizeof(struct tcphdr)); 25131da177e4SLinus Torvalds th->syn = 1; 25141da177e4SLinus Torvalds th->ack = 1; 25151da177e4SLinus Torvalds TCP_ECN_make_synack(req, th); 2516a3116ac5SKOVACS Krisztian th->source = ireq->loc_port; 25172e6599cbSArnaldo Carvalho de Melo th->dest = ireq->rmt_port; 2518e870a8efSIlpo Järvinen /* Setting of flags are superfluous here for callers (and ECE is 2519e870a8efSIlpo Järvinen * not even correctly set) 2520e870a8efSIlpo Järvinen */ 2521e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn, 2522a3433f35SChangli Gao TCPHDR_SYN | TCPHDR_ACK); 25234957faadSWilliam Allen Simpson 25244957faadSWilliam Allen Simpson if (OPTION_COOKIE_EXTENSION & opts.options) { 252528b2774aSEric Dumazet if (s_data_desired) { 252628b2774aSEric Dumazet u8 *buf = skb_put(skb, s_data_desired); 25274957faadSWilliam Allen Simpson 25284957faadSWilliam Allen Simpson /* copy data directly from the listening socket. */ 252928b2774aSEric Dumazet memcpy(buf, cvp->s_data_payload, s_data_desired); 253028b2774aSEric Dumazet TCP_SKB_CB(skb)->end_seq += s_data_desired; 25314957faadSWilliam Allen Simpson } 25324957faadSWilliam Allen Simpson 25334957faadSWilliam Allen Simpson if (opts.hash_size > 0) { 25344957faadSWilliam Allen Simpson __u32 workspace[SHA_WORKSPACE_WORDS]; 25354957faadSWilliam Allen Simpson u32 *mess = &xvp->cookie_bakery[COOKIE_DIGEST_WORDS]; 25364957faadSWilliam Allen Simpson u32 *tail = &mess[COOKIE_MESSAGE_WORDS-1]; 25374957faadSWilliam Allen Simpson 25384957faadSWilliam Allen Simpson /* Secret recipe depends on the Timestamp, (future) 25394957faadSWilliam Allen Simpson * Sequence and Acknowledgment Numbers, Initiator 25404957faadSWilliam Allen Simpson * Cookie, and others handled by IP variant caller. 25414957faadSWilliam Allen Simpson */ 25424957faadSWilliam Allen Simpson *tail-- ^= opts.tsval; 25434957faadSWilliam Allen Simpson *tail-- ^= tcp_rsk(req)->rcv_isn + 1; 25444957faadSWilliam Allen Simpson *tail-- ^= TCP_SKB_CB(skb)->seq + 1; 25454957faadSWilliam Allen Simpson 25464957faadSWilliam Allen Simpson /* recommended */ 25470eae88f3SEric Dumazet *tail-- ^= (((__force u32)th->dest << 16) | (__force u32)th->source); 2548f9a2e69eSDavid S. Miller *tail-- ^= (u32)(unsigned long)cvp; /* per sockopt */ 25494957faadSWilliam Allen Simpson 25504957faadSWilliam Allen Simpson sha_transform((__u32 *)&xvp->cookie_bakery[0], 25514957faadSWilliam Allen Simpson (char *)mess, 25524957faadSWilliam Allen Simpson &workspace[0]); 25534957faadSWilliam Allen Simpson opts.hash_location = 25544957faadSWilliam Allen Simpson (__u8 *)&xvp->cookie_bakery[0]; 25554957faadSWilliam Allen Simpson } 25564957faadSWilliam Allen Simpson } 25574957faadSWilliam Allen Simpson 25581da177e4SLinus Torvalds th->seq = htonl(TCP_SKB_CB(skb)->seq); 25592e6599cbSArnaldo Carvalho de Melo th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1); 25601da177e4SLinus Torvalds 25611da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 2562600ff0c2SIlpo Järvinen th->window = htons(min(req->rcv_wnd, 65535U)); 2563bd0388aeSWilliam Allen Simpson tcp_options_write((__be32 *)(th + 1), tp, &opts); 25641da177e4SLinus Torvalds th->doff = (tcp_header_size >> 2); 2565aa2ea058STom Herbert TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, tcp_skb_pcount(skb)); 2566cfb6eeb4SYOSHIFUJI Hideaki 2567cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 2568cfb6eeb4SYOSHIFUJI Hideaki /* Okay, we have all we need - do the md5 hash if needed */ 2569cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 2570bd0388aeSWilliam Allen Simpson tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, 257149a72dfbSAdam Langley md5, NULL, req, skb); 2572cfb6eeb4SYOSHIFUJI Hideaki } 2573cfb6eeb4SYOSHIFUJI Hideaki #endif 2574cfb6eeb4SYOSHIFUJI Hideaki 25751da177e4SLinus Torvalds return skb; 25761da177e4SLinus Torvalds } 25774bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_make_synack); 25781da177e4SLinus Torvalds 257967edfef7SAndi Kleen /* Do all connect socket setups that can be done AF independent. */ 2580370816aeSPavel Emelyanov void tcp_connect_init(struct sock *sk) 25811da177e4SLinus Torvalds { 2582cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 25831da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 25841da177e4SLinus Torvalds __u8 rcv_wscale; 25851da177e4SLinus Torvalds 25861da177e4SLinus Torvalds /* We'll fix this up when we get a response from the other end. 25871da177e4SLinus Torvalds * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 25881da177e4SLinus Torvalds */ 25891da177e4SLinus Torvalds tp->tcp_header_len = sizeof(struct tcphdr) + 2590bb5b7c11SDavid S. Miller (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0); 25911da177e4SLinus Torvalds 2592cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 2593cfb6eeb4SYOSHIFUJI Hideaki if (tp->af_specific->md5_lookup(sk, sk) != NULL) 2594cfb6eeb4SYOSHIFUJI Hideaki tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 2595cfb6eeb4SYOSHIFUJI Hideaki #endif 2596cfb6eeb4SYOSHIFUJI Hideaki 25971da177e4SLinus Torvalds /* If user gave his TCP_MAXSEG, record it to clamp */ 25981da177e4SLinus Torvalds if (tp->rx_opt.user_mss) 25991da177e4SLinus Torvalds tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 26001da177e4SLinus Torvalds tp->max_window = 0; 26015d424d5aSJohn Heffner tcp_mtup_init(sk); 26021da177e4SLinus Torvalds tcp_sync_mss(sk, dst_mtu(dst)); 26031da177e4SLinus Torvalds 26041da177e4SLinus Torvalds if (!tp->window_clamp) 26051da177e4SLinus Torvalds tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 26060dbaee3bSDavid S. Miller tp->advmss = dst_metric_advmss(dst); 2607f5fff5dcSTom Quetchenbach if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss) 2608f5fff5dcSTom Quetchenbach tp->advmss = tp->rx_opt.user_mss; 2609f5fff5dcSTom Quetchenbach 26101da177e4SLinus Torvalds tcp_initialize_rcv_mss(sk); 26111da177e4SLinus Torvalds 2612e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 2613e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 2614e88c64f0SHagen Paul Pfeifer (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) 2615e88c64f0SHagen Paul Pfeifer tp->window_clamp = tcp_full_space(sk); 2616e88c64f0SHagen Paul Pfeifer 26171da177e4SLinus Torvalds tcp_select_initial_window(tcp_full_space(sk), 26181da177e4SLinus Torvalds tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 26191da177e4SLinus Torvalds &tp->rcv_wnd, 26201da177e4SLinus Torvalds &tp->window_clamp, 2621bb5b7c11SDavid S. Miller sysctl_tcp_window_scaling, 262231d12926Slaurent chavey &rcv_wscale, 262331d12926Slaurent chavey dst_metric(dst, RTAX_INITRWND)); 26241da177e4SLinus Torvalds 26251da177e4SLinus Torvalds tp->rx_opt.rcv_wscale = rcv_wscale; 26261da177e4SLinus Torvalds tp->rcv_ssthresh = tp->rcv_wnd; 26271da177e4SLinus Torvalds 26281da177e4SLinus Torvalds sk->sk_err = 0; 26291da177e4SLinus Torvalds sock_reset_flag(sk, SOCK_DONE); 26301da177e4SLinus Torvalds tp->snd_wnd = 0; 2631ee7537b6SHantzis Fotis tcp_init_wl(tp, 0); 26321da177e4SLinus Torvalds tp->snd_una = tp->write_seq; 26331da177e4SLinus Torvalds tp->snd_sml = tp->write_seq; 263433f5f57eSIlpo Järvinen tp->snd_up = tp->write_seq; 2635370816aeSPavel Emelyanov tp->snd_nxt = tp->write_seq; 2636ee995283SPavel Emelyanov 2637ee995283SPavel Emelyanov if (likely(!tp->repair)) 26381da177e4SLinus Torvalds tp->rcv_nxt = 0; 2639ee995283SPavel Emelyanov tp->rcv_wup = tp->rcv_nxt; 2640ee995283SPavel Emelyanov tp->copied_seq = tp->rcv_nxt; 26411da177e4SLinus Torvalds 2642463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; 2643463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_retransmits = 0; 26441da177e4SLinus Torvalds tcp_clear_retrans(tp); 26451da177e4SLinus Torvalds } 26461da177e4SLinus Torvalds 264767edfef7SAndi Kleen /* Build a SYN and send it off. */ 26481da177e4SLinus Torvalds int tcp_connect(struct sock *sk) 26491da177e4SLinus Torvalds { 26501da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 26511da177e4SLinus Torvalds struct sk_buff *buff; 2652ee586811SEric Paris int err; 26531da177e4SLinus Torvalds 26541da177e4SLinus Torvalds tcp_connect_init(sk); 26551da177e4SLinus Torvalds 2656d179cd12SDavid S. Miller buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation); 26571da177e4SLinus Torvalds if (unlikely(buff == NULL)) 26581da177e4SLinus Torvalds return -ENOBUFS; 26591da177e4SLinus Torvalds 26601da177e4SLinus Torvalds /* Reserve space for headers. */ 26611da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 26621da177e4SLinus Torvalds 2663a3433f35SChangli Gao tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); 2664e870a8efSIlpo Järvinen TCP_ECN_send_syn(sk, buff); 26651da177e4SLinus Torvalds 26661da177e4SLinus Torvalds /* Send it off. */ 26671da177e4SLinus Torvalds TCP_SKB_CB(buff)->when = tcp_time_stamp; 26681da177e4SLinus Torvalds tp->retrans_stamp = TCP_SKB_CB(buff)->when; 26691da177e4SLinus Torvalds skb_header_release(buff); 2670fe067e8aSDavid S. Miller __tcp_add_write_queue_tail(sk, buff); 26713ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 26723ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 26731da177e4SLinus Torvalds tp->packets_out += tcp_skb_pcount(buff); 2674ee586811SEric Paris err = tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); 2675ee586811SEric Paris if (err == -ECONNREFUSED) 2676ee586811SEric Paris return err; 2677bd37a088SWei Yongjun 2678bd37a088SWei Yongjun /* We change tp->snd_nxt after the tcp_transmit_skb() call 2679bd37a088SWei Yongjun * in order to make this packet get counted in tcpOutSegs. 2680bd37a088SWei Yongjun */ 2681bd37a088SWei Yongjun tp->snd_nxt = tp->write_seq; 2682bd37a088SWei Yongjun tp->pushed_seq = tp->write_seq; 268381cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); 26841da177e4SLinus Torvalds 26851da177e4SLinus Torvalds /* Timer for repeating the SYN until an answer. */ 26863f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 26873f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 26881da177e4SLinus Torvalds return 0; 26891da177e4SLinus Torvalds } 26904bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_connect); 26911da177e4SLinus Torvalds 26921da177e4SLinus Torvalds /* Send out a delayed ack, the caller does the policy checking 26931da177e4SLinus Torvalds * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 26941da177e4SLinus Torvalds * for details. 26951da177e4SLinus Torvalds */ 26961da177e4SLinus Torvalds void tcp_send_delayed_ack(struct sock *sk) 26971da177e4SLinus Torvalds { 2698463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 2699463c84b9SArnaldo Carvalho de Melo int ato = icsk->icsk_ack.ato; 27001da177e4SLinus Torvalds unsigned long timeout; 27011da177e4SLinus Torvalds 27021da177e4SLinus Torvalds if (ato > TCP_DELACK_MIN) { 2703463c84b9SArnaldo Carvalho de Melo const struct tcp_sock *tp = tcp_sk(sk); 27041da177e4SLinus Torvalds int max_ato = HZ / 2; 27051da177e4SLinus Torvalds 2706056834d9SIlpo Järvinen if (icsk->icsk_ack.pingpong || 2707056834d9SIlpo Järvinen (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 27081da177e4SLinus Torvalds max_ato = TCP_DELACK_MAX; 27091da177e4SLinus Torvalds 27101da177e4SLinus Torvalds /* Slow path, intersegment interval is "high". */ 27111da177e4SLinus Torvalds 27121da177e4SLinus Torvalds /* If some rtt estimate is known, use it to bound delayed ack. 2713463c84b9SArnaldo Carvalho de Melo * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 27141da177e4SLinus Torvalds * directly. 27151da177e4SLinus Torvalds */ 27161da177e4SLinus Torvalds if (tp->srtt) { 27171da177e4SLinus Torvalds int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN); 27181da177e4SLinus Torvalds 27191da177e4SLinus Torvalds if (rtt < max_ato) 27201da177e4SLinus Torvalds max_ato = rtt; 27211da177e4SLinus Torvalds } 27221da177e4SLinus Torvalds 27231da177e4SLinus Torvalds ato = min(ato, max_ato); 27241da177e4SLinus Torvalds } 27251da177e4SLinus Torvalds 27261da177e4SLinus Torvalds /* Stay within the limit we were given */ 27271da177e4SLinus Torvalds timeout = jiffies + ato; 27281da177e4SLinus Torvalds 27291da177e4SLinus Torvalds /* Use new timeout only if there wasn't a older one earlier. */ 2730463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 27311da177e4SLinus Torvalds /* If delack timer was blocked or is about to expire, 27321da177e4SLinus Torvalds * send ACK now. 27331da177e4SLinus Torvalds */ 2734463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.blocked || 2735463c84b9SArnaldo Carvalho de Melo time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 27361da177e4SLinus Torvalds tcp_send_ack(sk); 27371da177e4SLinus Torvalds return; 27381da177e4SLinus Torvalds } 27391da177e4SLinus Torvalds 2740463c84b9SArnaldo Carvalho de Melo if (!time_before(timeout, icsk->icsk_ack.timeout)) 2741463c84b9SArnaldo Carvalho de Melo timeout = icsk->icsk_ack.timeout; 27421da177e4SLinus Torvalds } 2743463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 2744463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.timeout = timeout; 2745463c84b9SArnaldo Carvalho de Melo sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 27461da177e4SLinus Torvalds } 27471da177e4SLinus Torvalds 27481da177e4SLinus Torvalds /* This routine sends an ack and also updates the window. */ 27491da177e4SLinus Torvalds void tcp_send_ack(struct sock *sk) 27501da177e4SLinus Torvalds { 27511da177e4SLinus Torvalds struct sk_buff *buff; 27521da177e4SLinus Torvalds 2753058dc334SIlpo Järvinen /* If we have been reset, we may not send again. */ 2754058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 2755058dc334SIlpo Järvinen return; 2756058dc334SIlpo Järvinen 27571da177e4SLinus Torvalds /* We are not putting this on the write queue, so 27581da177e4SLinus Torvalds * tcp_transmit_skb() will set the ownership to this 27591da177e4SLinus Torvalds * sock. 27601da177e4SLinus Torvalds */ 27611da177e4SLinus Torvalds buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 27621da177e4SLinus Torvalds if (buff == NULL) { 2763463c84b9SArnaldo Carvalho de Melo inet_csk_schedule_ack(sk); 2764463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; 27653f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 27663f421baaSArnaldo Carvalho de Melo TCP_DELACK_MAX, TCP_RTO_MAX); 27671da177e4SLinus Torvalds return; 27681da177e4SLinus Torvalds } 27691da177e4SLinus Torvalds 27701da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 27711da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 2772a3433f35SChangli Gao tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); 27731da177e4SLinus Torvalds 27741da177e4SLinus Torvalds /* Send it off, this clears delayed acks for us. */ 27751da177e4SLinus Torvalds TCP_SKB_CB(buff)->when = tcp_time_stamp; 2776dfb4b9dcSDavid S. Miller tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC); 27771da177e4SLinus Torvalds } 27781da177e4SLinus Torvalds 27791da177e4SLinus Torvalds /* This routine sends a packet with an out of date sequence 27801da177e4SLinus Torvalds * number. It assumes the other end will try to ack it. 27811da177e4SLinus Torvalds * 27821da177e4SLinus Torvalds * Question: what should we make while urgent mode? 27831da177e4SLinus Torvalds * 4.4BSD forces sending single byte of data. We cannot send 27841da177e4SLinus Torvalds * out of window data, because we have SND.NXT==SND.MAX... 27851da177e4SLinus Torvalds * 27861da177e4SLinus Torvalds * Current solution: to send TWO zero-length segments in urgent mode: 27871da177e4SLinus Torvalds * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 27881da177e4SLinus Torvalds * out-of-date with SND.UNA-1 to probe window. 27891da177e4SLinus Torvalds */ 27901da177e4SLinus Torvalds static int tcp_xmit_probe_skb(struct sock *sk, int urgent) 27911da177e4SLinus Torvalds { 27921da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 27931da177e4SLinus Torvalds struct sk_buff *skb; 27941da177e4SLinus Torvalds 27951da177e4SLinus Torvalds /* We don't queue it, tcp_transmit_skb() sets ownership. */ 27961da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 27971da177e4SLinus Torvalds if (skb == NULL) 27981da177e4SLinus Torvalds return -1; 27991da177e4SLinus Torvalds 28001da177e4SLinus Torvalds /* Reserve space for headers and set control bits. */ 28011da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 28021da177e4SLinus Torvalds /* Use a previous sequence. This should cause the other 28031da177e4SLinus Torvalds * end to send an ack. Don't queue or clone SKB, just 28041da177e4SLinus Torvalds * send it. 28051da177e4SLinus Torvalds */ 2806a3433f35SChangli Gao tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); 28071da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2808dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC); 28091da177e4SLinus Torvalds } 28101da177e4SLinus Torvalds 2811ee995283SPavel Emelyanov void tcp_send_window_probe(struct sock *sk) 2812ee995283SPavel Emelyanov { 2813ee995283SPavel Emelyanov if (sk->sk_state == TCP_ESTABLISHED) { 2814ee995283SPavel Emelyanov tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; 2815c0e88ff0SPavel Emelyanov tcp_sk(sk)->snd_nxt = tcp_sk(sk)->write_seq; 2816ee995283SPavel Emelyanov tcp_xmit_probe_skb(sk, 0); 2817ee995283SPavel Emelyanov } 2818ee995283SPavel Emelyanov } 2819ee995283SPavel Emelyanov 282067edfef7SAndi Kleen /* Initiate keepalive or window probe from timer. */ 28211da177e4SLinus Torvalds int tcp_write_wakeup(struct sock *sk) 28221da177e4SLinus Torvalds { 28231da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 28241da177e4SLinus Torvalds struct sk_buff *skb; 28251da177e4SLinus Torvalds 2826058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 2827058dc334SIlpo Järvinen return -1; 2828058dc334SIlpo Järvinen 2829fe067e8aSDavid S. Miller if ((skb = tcp_send_head(sk)) != NULL && 283090840defSIlpo Järvinen before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { 28311da177e4SLinus Torvalds int err; 28320c54b85fSIlpo Järvinen unsigned int mss = tcp_current_mss(sk); 283390840defSIlpo Järvinen unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 28341da177e4SLinus Torvalds 28351da177e4SLinus Torvalds if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 28361da177e4SLinus Torvalds tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 28371da177e4SLinus Torvalds 28381da177e4SLinus Torvalds /* We are probing the opening of a window 28391da177e4SLinus Torvalds * but the window size is != 0 28401da177e4SLinus Torvalds * must have been a result SWS avoidance ( sender ) 28411da177e4SLinus Torvalds */ 28421da177e4SLinus Torvalds if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 28431da177e4SLinus Torvalds skb->len > mss) { 28441da177e4SLinus Torvalds seg_size = min(seg_size, mss); 28454de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 2846846998aeSDavid S. Miller if (tcp_fragment(sk, skb, seg_size, mss)) 28471da177e4SLinus Torvalds return -1; 28481da177e4SLinus Torvalds } else if (!tcp_skb_pcount(skb)) 2849846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss); 28501da177e4SLinus Torvalds 28514de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 28521da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2853dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 285466f5fe62SIlpo Järvinen if (!err) 285566f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 28561da177e4SLinus Torvalds return err; 28571da177e4SLinus Torvalds } else { 285833f5f57eSIlpo Järvinen if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) 28594828e7f4SIlpo Järvinen tcp_xmit_probe_skb(sk, 1); 28601da177e4SLinus Torvalds return tcp_xmit_probe_skb(sk, 0); 28611da177e4SLinus Torvalds } 28621da177e4SLinus Torvalds } 28631da177e4SLinus Torvalds 28641da177e4SLinus Torvalds /* A window probe timeout has occurred. If window is not closed send 28651da177e4SLinus Torvalds * a partial packet else a zero probe. 28661da177e4SLinus Torvalds */ 28671da177e4SLinus Torvalds void tcp_send_probe0(struct sock *sk) 28681da177e4SLinus Torvalds { 2869463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 28701da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 28711da177e4SLinus Torvalds int err; 28721da177e4SLinus Torvalds 28731da177e4SLinus Torvalds err = tcp_write_wakeup(sk); 28741da177e4SLinus Torvalds 2875fe067e8aSDavid S. Miller if (tp->packets_out || !tcp_send_head(sk)) { 28761da177e4SLinus Torvalds /* Cancel probe timer, if it is not required. */ 28776687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 0; 2878463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff = 0; 28791da177e4SLinus Torvalds return; 28801da177e4SLinus Torvalds } 28811da177e4SLinus Torvalds 28821da177e4SLinus Torvalds if (err <= 0) { 2883463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_backoff < sysctl_tcp_retries2) 2884463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff++; 28856687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out++; 2886463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 28873f421baaSArnaldo Carvalho de Melo min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), 28883f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 28891da177e4SLinus Torvalds } else { 28901da177e4SLinus Torvalds /* If packet was not sent due to local congestion, 28916687e988SArnaldo Carvalho de Melo * do not backoff and do not remember icsk_probes_out. 28921da177e4SLinus Torvalds * Let local senders to fight for local resources. 28931da177e4SLinus Torvalds * 28941da177e4SLinus Torvalds * Use accumulated backoff yet. 28951da177e4SLinus Torvalds */ 28966687e988SArnaldo Carvalho de Melo if (!icsk->icsk_probes_out) 28976687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 1; 2898463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 2899463c84b9SArnaldo Carvalho de Melo min(icsk->icsk_rto << icsk->icsk_backoff, 29003f421baaSArnaldo Carvalho de Melo TCP_RESOURCE_PROBE_INTERVAL), 29013f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 29021da177e4SLinus Torvalds } 29031da177e4SLinus Torvalds } 2904