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 3791df42beSJoe Perches #define pr_fmt(fmt) "TCP: " fmt 3891df42beSJoe Perches 391da177e4SLinus Torvalds #include <net/tcp.h> 401da177e4SLinus Torvalds 411da177e4SLinus Torvalds #include <linux/compiler.h> 425a0e3ad6STejun Heo #include <linux/gfp.h> 431da177e4SLinus Torvalds #include <linux/module.h> 441da177e4SLinus Torvalds 451da177e4SLinus Torvalds /* People can turn this off for buggy TCP's found in printers etc. */ 46ab32ea5dSBrian Haley int sysctl_tcp_retrans_collapse __read_mostly = 1; 471da177e4SLinus Torvalds 4815d99e02SRick Jones /* People can turn this on to work with those rare, broken TCPs that 4915d99e02SRick Jones * interpret the window field as a signed quantity. 5015d99e02SRick Jones */ 51ab32ea5dSBrian Haley int sysctl_tcp_workaround_signed_windows __read_mostly = 0; 5215d99e02SRick Jones 531da177e4SLinus Torvalds /* This limits the percentage of the congestion window which we 541da177e4SLinus Torvalds * will allow a single TSO frame to consume. Building TSO frames 551da177e4SLinus Torvalds * which are too large can cause TCP streams to be bursty. 561da177e4SLinus Torvalds */ 57ab32ea5dSBrian Haley int sysctl_tcp_tso_win_divisor __read_mostly = 3; 581da177e4SLinus Torvalds 59ab32ea5dSBrian Haley int sysctl_tcp_mtu_probing __read_mostly = 0; 6097b1ce25SShan Wei int sysctl_tcp_base_mss __read_mostly = TCP_BASE_MSS; 615d424d5aSJohn Heffner 6235089bb2SDavid S. Miller /* By default, RFC2861 behavior. */ 63ab32ea5dSBrian Haley int sysctl_tcp_slow_start_after_idle __read_mostly = 1; 6435089bb2SDavid S. Miller 65519855c5SWilliam Allen Simpson int sysctl_tcp_cookie_size __read_mostly = 0; /* TCP_COOKIE_MAX */ 66e6b09ccaSDavid S. Miller EXPORT_SYMBOL_GPL(sysctl_tcp_cookie_size); 67519855c5SWilliam Allen Simpson 68519855c5SWilliam Allen Simpson 6967edfef7SAndi Kleen /* Account for new data that has been sent to the network. */ 70cf533ea5SEric Dumazet static void tcp_event_new_data_sent(struct sock *sk, const struct sk_buff *skb) 716ff03ac3SIlpo Järvinen { 726ff03ac3SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 7366f5fe62SIlpo Järvinen unsigned int prior_packets = tp->packets_out; 749e412ba7SIlpo Järvinen 75fe067e8aSDavid S. Miller tcp_advance_send_head(sk, skb); 761da177e4SLinus Torvalds tp->snd_nxt = TCP_SKB_CB(skb)->end_seq; 778512430eSIlpo Järvinen 7825985edcSLucas De Marchi /* Don't override Nagle indefinitely with F-RTO */ 798512430eSIlpo Järvinen if (tp->frto_counter == 2) 808512430eSIlpo Järvinen tp->frto_counter = 3; 8166f5fe62SIlpo Järvinen 8266f5fe62SIlpo Järvinen tp->packets_out += tcp_skb_pcount(skb); 83750ea2baSYuchung Cheng if (!prior_packets || tp->early_retrans_delayed) 84750ea2baSYuchung Cheng tcp_rearm_rto(sk); 851da177e4SLinus Torvalds } 861da177e4SLinus Torvalds 871da177e4SLinus Torvalds /* SND.NXT, if window was not shrunk. 881da177e4SLinus Torvalds * If window has been shrunk, what should we make? It is not clear at all. 891da177e4SLinus Torvalds * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-( 901da177e4SLinus Torvalds * Anything in between SND.UNA...SND.UNA+SND.WND also can be already 911da177e4SLinus Torvalds * invalid. OK, let's make this for now: 921da177e4SLinus Torvalds */ 93cf533ea5SEric Dumazet static inline __u32 tcp_acceptable_seq(const struct sock *sk) 941da177e4SLinus Torvalds { 95cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 969e412ba7SIlpo Järvinen 9790840defSIlpo Järvinen if (!before(tcp_wnd_end(tp), tp->snd_nxt)) 981da177e4SLinus Torvalds return tp->snd_nxt; 991da177e4SLinus Torvalds else 10090840defSIlpo Järvinen return tcp_wnd_end(tp); 1011da177e4SLinus Torvalds } 1021da177e4SLinus Torvalds 1031da177e4SLinus Torvalds /* Calculate mss to advertise in SYN segment. 1041da177e4SLinus Torvalds * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that: 1051da177e4SLinus Torvalds * 1061da177e4SLinus Torvalds * 1. It is independent of path mtu. 1071da177e4SLinus Torvalds * 2. Ideally, it is maximal possible segment size i.e. 65535-40. 1081da177e4SLinus Torvalds * 3. For IPv4 it is reasonable to calculate it from maximal MTU of 1091da177e4SLinus Torvalds * attached devices, because some buggy hosts are confused by 1101da177e4SLinus Torvalds * large MSS. 1111da177e4SLinus Torvalds * 4. We do not make 3, we advertise MSS, calculated from first 1121da177e4SLinus Torvalds * hop device mtu, but allow to raise it to ip_rt_min_advmss. 1131da177e4SLinus Torvalds * This may be overridden via information stored in routing table. 1141da177e4SLinus Torvalds * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible, 1151da177e4SLinus Torvalds * probably even Jumbo". 1161da177e4SLinus Torvalds */ 1171da177e4SLinus Torvalds static __u16 tcp_advertise_mss(struct sock *sk) 1181da177e4SLinus Torvalds { 1191da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 120cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1211da177e4SLinus Torvalds int mss = tp->advmss; 1221da177e4SLinus Torvalds 1230dbaee3bSDavid S. Miller if (dst) { 1240dbaee3bSDavid S. Miller unsigned int metric = dst_metric_advmss(dst); 1250dbaee3bSDavid S. Miller 1260dbaee3bSDavid S. Miller if (metric < mss) { 1270dbaee3bSDavid S. Miller mss = metric; 1281da177e4SLinus Torvalds tp->advmss = mss; 1291da177e4SLinus Torvalds } 1300dbaee3bSDavid S. Miller } 1311da177e4SLinus Torvalds 1321da177e4SLinus Torvalds return (__u16)mss; 1331da177e4SLinus Torvalds } 1341da177e4SLinus Torvalds 1351da177e4SLinus Torvalds /* RFC2861. Reset CWND after idle period longer RTO to "restart window". 1361da177e4SLinus Torvalds * This is the first part of cwnd validation mechanism. */ 137cf533ea5SEric Dumazet static void tcp_cwnd_restart(struct sock *sk, const struct dst_entry *dst) 1381da177e4SLinus Torvalds { 139463c84b9SArnaldo Carvalho de Melo struct tcp_sock *tp = tcp_sk(sk); 1401da177e4SLinus Torvalds s32 delta = tcp_time_stamp - tp->lsndtime; 1411da177e4SLinus Torvalds u32 restart_cwnd = tcp_init_cwnd(tp, dst); 1421da177e4SLinus Torvalds u32 cwnd = tp->snd_cwnd; 1431da177e4SLinus Torvalds 1446687e988SArnaldo Carvalho de Melo tcp_ca_event(sk, CA_EVENT_CWND_RESTART); 1451da177e4SLinus Torvalds 1466687e988SArnaldo Carvalho de Melo tp->snd_ssthresh = tcp_current_ssthresh(sk); 1471da177e4SLinus Torvalds restart_cwnd = min(restart_cwnd, cwnd); 1481da177e4SLinus Torvalds 149463c84b9SArnaldo Carvalho de Melo while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd) 1501da177e4SLinus Torvalds cwnd >>= 1; 1511da177e4SLinus Torvalds tp->snd_cwnd = max(cwnd, restart_cwnd); 1521da177e4SLinus Torvalds tp->snd_cwnd_stamp = tcp_time_stamp; 1531da177e4SLinus Torvalds tp->snd_cwnd_used = 0; 1541da177e4SLinus Torvalds } 1551da177e4SLinus Torvalds 15667edfef7SAndi Kleen /* Congestion state accounting after a packet has been sent. */ 15740efc6faSStephen Hemminger static void tcp_event_data_sent(struct tcp_sock *tp, 158cf533ea5SEric Dumazet struct sock *sk) 1591da177e4SLinus Torvalds { 160463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 161463c84b9SArnaldo Carvalho de Melo const u32 now = tcp_time_stamp; 1621da177e4SLinus Torvalds 16335089bb2SDavid S. Miller if (sysctl_tcp_slow_start_after_idle && 16435089bb2SDavid S. Miller (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto)) 165463c84b9SArnaldo Carvalho de Melo tcp_cwnd_restart(sk, __sk_dst_get(sk)); 1661da177e4SLinus Torvalds 1671da177e4SLinus Torvalds tp->lsndtime = now; 1681da177e4SLinus Torvalds 1691da177e4SLinus Torvalds /* If it is a reply for ato after last received 1701da177e4SLinus Torvalds * packet, enter pingpong mode. 1711da177e4SLinus Torvalds */ 172463c84b9SArnaldo Carvalho de Melo if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato) 173463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pingpong = 1; 1741da177e4SLinus Torvalds } 1751da177e4SLinus Torvalds 17667edfef7SAndi Kleen /* Account for an ACK we sent. */ 17740efc6faSStephen Hemminger static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts) 1781da177e4SLinus Torvalds { 179463c84b9SArnaldo Carvalho de Melo tcp_dec_quickack_mode(sk, pkts); 180463c84b9SArnaldo Carvalho de Melo inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); 1811da177e4SLinus Torvalds } 1821da177e4SLinus Torvalds 1831da177e4SLinus Torvalds /* Determine a window scaling and initial window to offer. 1841da177e4SLinus Torvalds * Based on the assumption that the given amount of space 1851da177e4SLinus Torvalds * will be offered. Store the results in the tp structure. 1861da177e4SLinus Torvalds * NOTE: for smooth operation initial space offering should 1871da177e4SLinus Torvalds * be a multiple of mss if possible. We assume here that mss >= 1. 1881da177e4SLinus Torvalds * This MUST be enforced by all callers. 1891da177e4SLinus Torvalds */ 1901da177e4SLinus Torvalds void tcp_select_initial_window(int __space, __u32 mss, 1911da177e4SLinus Torvalds __u32 *rcv_wnd, __u32 *window_clamp, 19231d12926Slaurent chavey int wscale_ok, __u8 *rcv_wscale, 19331d12926Slaurent chavey __u32 init_rcv_wnd) 1941da177e4SLinus Torvalds { 1951da177e4SLinus Torvalds unsigned int space = (__space < 0 ? 0 : __space); 1961da177e4SLinus Torvalds 1971da177e4SLinus Torvalds /* If no clamp set the clamp to the max possible scaled window */ 1981da177e4SLinus Torvalds if (*window_clamp == 0) 1991da177e4SLinus Torvalds (*window_clamp) = (65535 << 14); 2001da177e4SLinus Torvalds space = min(*window_clamp, space); 2011da177e4SLinus Torvalds 2021da177e4SLinus Torvalds /* Quantize space offering to a multiple of mss if possible. */ 2031da177e4SLinus Torvalds if (space > mss) 2041da177e4SLinus Torvalds space = (space / mss) * mss; 2051da177e4SLinus Torvalds 2061da177e4SLinus Torvalds /* NOTE: offering an initial window larger than 32767 20715d99e02SRick Jones * will break some buggy TCP stacks. If the admin tells us 20815d99e02SRick Jones * it is likely we could be speaking with such a buggy stack 20915d99e02SRick Jones * we will truncate our initial window offering to 32K-1 21015d99e02SRick Jones * unless the remote has sent us a window scaling option, 21115d99e02SRick Jones * which we interpret as a sign the remote TCP is not 21215d99e02SRick Jones * misinterpreting the window field as a signed quantity. 2131da177e4SLinus Torvalds */ 21415d99e02SRick Jones if (sysctl_tcp_workaround_signed_windows) 2151da177e4SLinus Torvalds (*rcv_wnd) = min(space, MAX_TCP_WINDOW); 21615d99e02SRick Jones else 21715d99e02SRick Jones (*rcv_wnd) = space; 21815d99e02SRick Jones 2191da177e4SLinus Torvalds (*rcv_wscale) = 0; 2201da177e4SLinus Torvalds if (wscale_ok) { 2211da177e4SLinus Torvalds /* Set window scaling on max possible window 2221da177e4SLinus Torvalds * See RFC1323 for an explanation of the limit to 14 2231da177e4SLinus Torvalds */ 2241da177e4SLinus Torvalds space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max); 225316c1592SStephen Hemminger space = min_t(u32, space, *window_clamp); 2261da177e4SLinus Torvalds while (space > 65535 && (*rcv_wscale) < 14) { 2271da177e4SLinus Torvalds space >>= 1; 2281da177e4SLinus Torvalds (*rcv_wscale)++; 2291da177e4SLinus Torvalds } 2301da177e4SLinus Torvalds } 2311da177e4SLinus Torvalds 232356f0398SNandita Dukkipati /* Set initial window to a value enough for senders starting with 233356f0398SNandita Dukkipati * initial congestion window of TCP_DEFAULT_INIT_RCVWND. Place 234356f0398SNandita Dukkipati * a limit on the initial window when mss is larger than 1460. 235356f0398SNandita Dukkipati */ 2361da177e4SLinus Torvalds if (mss > (1 << *rcv_wscale)) { 237356f0398SNandita Dukkipati int init_cwnd = TCP_DEFAULT_INIT_RCVWND; 238356f0398SNandita Dukkipati if (mss > 1460) 239356f0398SNandita Dukkipati init_cwnd = 240356f0398SNandita Dukkipati max_t(u32, (1460 * TCP_DEFAULT_INIT_RCVWND) / mss, 2); 24131d12926Slaurent chavey /* when initializing use the value from init_rcv_wnd 24231d12926Slaurent chavey * rather than the default from above 24331d12926Slaurent chavey */ 244b1afde60SNandita Dukkipati if (init_rcv_wnd) 245b1afde60SNandita Dukkipati *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss); 246b1afde60SNandita Dukkipati else 247b1afde60SNandita Dukkipati *rcv_wnd = min(*rcv_wnd, init_cwnd * mss); 2481da177e4SLinus Torvalds } 2491da177e4SLinus Torvalds 2501da177e4SLinus Torvalds /* Set the clamp no higher than max representable value */ 2511da177e4SLinus Torvalds (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp); 2521da177e4SLinus Torvalds } 2534bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_select_initial_window); 2541da177e4SLinus Torvalds 2551da177e4SLinus Torvalds /* Chose a new window to advertise, update state in tcp_sock for the 2561da177e4SLinus Torvalds * socket, and return result with RFC1323 scaling applied. The return 2571da177e4SLinus Torvalds * value can be stuffed directly into th->window for an outgoing 2581da177e4SLinus Torvalds * frame. 2591da177e4SLinus Torvalds */ 26040efc6faSStephen Hemminger static u16 tcp_select_window(struct sock *sk) 2611da177e4SLinus Torvalds { 2621da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2631da177e4SLinus Torvalds u32 cur_win = tcp_receive_window(tp); 2641da177e4SLinus Torvalds u32 new_win = __tcp_select_window(sk); 2651da177e4SLinus Torvalds 2661da177e4SLinus Torvalds /* Never shrink the offered window */ 2671da177e4SLinus Torvalds if (new_win < cur_win) { 2681da177e4SLinus Torvalds /* Danger Will Robinson! 2691da177e4SLinus Torvalds * Don't update rcv_wup/rcv_wnd here or else 2701da177e4SLinus Torvalds * we will not be able to advertise a zero 2711da177e4SLinus Torvalds * window in time. --DaveM 2721da177e4SLinus Torvalds * 2731da177e4SLinus Torvalds * Relax Will Robinson. 2741da177e4SLinus Torvalds */ 275607bfbf2SPatrick McHardy new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale); 2761da177e4SLinus Torvalds } 2771da177e4SLinus Torvalds tp->rcv_wnd = new_win; 2781da177e4SLinus Torvalds tp->rcv_wup = tp->rcv_nxt; 2791da177e4SLinus Torvalds 2801da177e4SLinus Torvalds /* Make sure we do not exceed the maximum possible 2811da177e4SLinus Torvalds * scaled window. 2821da177e4SLinus Torvalds */ 28315d99e02SRick Jones if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows) 2841da177e4SLinus Torvalds new_win = min(new_win, MAX_TCP_WINDOW); 2851da177e4SLinus Torvalds else 2861da177e4SLinus Torvalds new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); 2871da177e4SLinus Torvalds 2881da177e4SLinus Torvalds /* RFC1323 scaling applied */ 2891da177e4SLinus Torvalds new_win >>= tp->rx_opt.rcv_wscale; 2901da177e4SLinus Torvalds 2911da177e4SLinus Torvalds /* If we advertise zero window, disable fast path. */ 2921da177e4SLinus Torvalds if (new_win == 0) 2931da177e4SLinus Torvalds tp->pred_flags = 0; 2941da177e4SLinus Torvalds 2951da177e4SLinus Torvalds return new_win; 2961da177e4SLinus Torvalds } 2971da177e4SLinus Torvalds 29867edfef7SAndi Kleen /* Packet ECN state for a SYN-ACK */ 299cf533ea5SEric Dumazet static inline void TCP_ECN_send_synack(const struct tcp_sock *tp, struct sk_buff *skb) 300bdf1ee5dSIlpo Järvinen { 3014de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR; 302bdf1ee5dSIlpo Järvinen if (!(tp->ecn_flags & TCP_ECN_OK)) 3034de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE; 304bdf1ee5dSIlpo Järvinen } 305bdf1ee5dSIlpo Järvinen 30667edfef7SAndi Kleen /* Packet ECN state for a SYN. */ 307bdf1ee5dSIlpo Järvinen static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb) 308bdf1ee5dSIlpo Järvinen { 309bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 310bdf1ee5dSIlpo Järvinen 311bdf1ee5dSIlpo Järvinen tp->ecn_flags = 0; 312255cac91SIlpo Järvinen if (sysctl_tcp_ecn == 1) { 3134de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR; 314bdf1ee5dSIlpo Järvinen tp->ecn_flags = TCP_ECN_OK; 315bdf1ee5dSIlpo Järvinen } 316bdf1ee5dSIlpo Järvinen } 317bdf1ee5dSIlpo Järvinen 318bdf1ee5dSIlpo Järvinen static __inline__ void 319cf533ea5SEric Dumazet TCP_ECN_make_synack(const struct request_sock *req, struct tcphdr *th) 320bdf1ee5dSIlpo Järvinen { 321bdf1ee5dSIlpo Järvinen if (inet_rsk(req)->ecn_ok) 322bdf1ee5dSIlpo Järvinen th->ece = 1; 323bdf1ee5dSIlpo Järvinen } 324bdf1ee5dSIlpo Järvinen 32567edfef7SAndi Kleen /* Set up ECN state for a packet on a ESTABLISHED socket that is about to 32667edfef7SAndi Kleen * be sent. 32767edfef7SAndi Kleen */ 328bdf1ee5dSIlpo Järvinen static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb, 329bdf1ee5dSIlpo Järvinen int tcp_header_len) 330bdf1ee5dSIlpo Järvinen { 331bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 332bdf1ee5dSIlpo Järvinen 333bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_OK) { 334bdf1ee5dSIlpo Järvinen /* Not-retransmitted data segment: set ECT and inject CWR. */ 335bdf1ee5dSIlpo Järvinen if (skb->len != tcp_header_len && 336bdf1ee5dSIlpo Järvinen !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { 337bdf1ee5dSIlpo Järvinen INET_ECN_xmit(sk); 338bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) { 339bdf1ee5dSIlpo Järvinen tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; 340bdf1ee5dSIlpo Järvinen tcp_hdr(skb)->cwr = 1; 341bdf1ee5dSIlpo Järvinen skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; 342bdf1ee5dSIlpo Järvinen } 343bdf1ee5dSIlpo Järvinen } else { 344bdf1ee5dSIlpo Järvinen /* ACK or retransmitted segment: clear ECT|CE */ 345bdf1ee5dSIlpo Järvinen INET_ECN_dontxmit(sk); 346bdf1ee5dSIlpo Järvinen } 347bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) 348bdf1ee5dSIlpo Järvinen tcp_hdr(skb)->ece = 1; 349bdf1ee5dSIlpo Järvinen } 350bdf1ee5dSIlpo Järvinen } 351bdf1ee5dSIlpo Järvinen 352e870a8efSIlpo Järvinen /* Constructs common control bits of non-data skb. If SYN/FIN is present, 353e870a8efSIlpo Järvinen * auto increment end seqno. 354e870a8efSIlpo Järvinen */ 355e870a8efSIlpo Järvinen static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags) 356e870a8efSIlpo Järvinen { 3572e8e18efSDavid S. Miller skb->ip_summed = CHECKSUM_PARTIAL; 358e870a8efSIlpo Järvinen skb->csum = 0; 359e870a8efSIlpo Järvinen 3604de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags; 361e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->sacked = 0; 362e870a8efSIlpo Järvinen 363e870a8efSIlpo Järvinen skb_shinfo(skb)->gso_segs = 1; 364e870a8efSIlpo Järvinen skb_shinfo(skb)->gso_size = 0; 365e870a8efSIlpo Järvinen skb_shinfo(skb)->gso_type = 0; 366e870a8efSIlpo Järvinen 367e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->seq = seq; 368a3433f35SChangli Gao if (flags & (TCPHDR_SYN | TCPHDR_FIN)) 369e870a8efSIlpo Järvinen seq++; 370e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->end_seq = seq; 371e870a8efSIlpo Järvinen } 372e870a8efSIlpo Järvinen 373a2a385d6SEric Dumazet static inline bool tcp_urg_mode(const struct tcp_sock *tp) 37433f5f57eSIlpo Järvinen { 37533f5f57eSIlpo Järvinen return tp->snd_una != tp->snd_up; 37633f5f57eSIlpo Järvinen } 37733f5f57eSIlpo Järvinen 37833ad798cSAdam Langley #define OPTION_SACK_ADVERTISE (1 << 0) 37933ad798cSAdam Langley #define OPTION_TS (1 << 1) 38033ad798cSAdam Langley #define OPTION_MD5 (1 << 2) 38189e95a61SOri Finkelman #define OPTION_WSCALE (1 << 3) 382bd0388aeSWilliam Allen Simpson #define OPTION_COOKIE_EXTENSION (1 << 4) 38333ad798cSAdam Langley 38433ad798cSAdam Langley struct tcp_out_options { 38533ad798cSAdam Langley u8 options; /* bit field of OPTION_* */ 38633ad798cSAdam Langley u8 ws; /* window scale, 0 to disable */ 38733ad798cSAdam Langley u8 num_sack_blocks; /* number of SACK blocks to include */ 388bd0388aeSWilliam Allen Simpson u8 hash_size; /* bytes in hash_location */ 38933ad798cSAdam Langley u16 mss; /* 0 to disable */ 39033ad798cSAdam Langley __u32 tsval, tsecr; /* need to include OPTION_TS */ 391bd0388aeSWilliam Allen Simpson __u8 *hash_location; /* temporary pointer, overloaded */ 39233ad798cSAdam Langley }; 39333ad798cSAdam Langley 394bd0388aeSWilliam Allen Simpson /* The sysctl int routines are generic, so check consistency here. 395bd0388aeSWilliam Allen Simpson */ 396bd0388aeSWilliam Allen Simpson static u8 tcp_cookie_size_check(u8 desired) 397bd0388aeSWilliam Allen Simpson { 398f1987257SEric Dumazet int cookie_size; 399f1987257SEric Dumazet 400f1987257SEric Dumazet if (desired > 0) 401bd0388aeSWilliam Allen Simpson /* previously specified */ 402bd0388aeSWilliam Allen Simpson return desired; 403f1987257SEric Dumazet 404f1987257SEric Dumazet cookie_size = ACCESS_ONCE(sysctl_tcp_cookie_size); 405f1987257SEric Dumazet if (cookie_size <= 0) 406bd0388aeSWilliam Allen Simpson /* no default specified */ 407bd0388aeSWilliam Allen Simpson return 0; 408f1987257SEric Dumazet 409f1987257SEric Dumazet if (cookie_size <= TCP_COOKIE_MIN) 410bd0388aeSWilliam Allen Simpson /* value too small, specify minimum */ 411bd0388aeSWilliam Allen Simpson return TCP_COOKIE_MIN; 412f1987257SEric Dumazet 413f1987257SEric Dumazet if (cookie_size >= TCP_COOKIE_MAX) 414bd0388aeSWilliam Allen Simpson /* value too large, specify maximum */ 415bd0388aeSWilliam Allen Simpson return TCP_COOKIE_MAX; 416f1987257SEric Dumazet 417f1987257SEric Dumazet if (cookie_size & 1) 418bd0388aeSWilliam Allen Simpson /* 8-bit multiple, illegal, fix it */ 419f1987257SEric Dumazet cookie_size++; 420f1987257SEric Dumazet 421f1987257SEric Dumazet return (u8)cookie_size; 422bd0388aeSWilliam Allen Simpson } 423bd0388aeSWilliam Allen Simpson 42467edfef7SAndi Kleen /* Write previously computed TCP options to the packet. 42567edfef7SAndi Kleen * 42667edfef7SAndi Kleen * Beware: Something in the Internet is very sensitive to the ordering of 427fd6149d3SIlpo Järvinen * TCP options, we learned this through the hard way, so be careful here. 428fd6149d3SIlpo Järvinen * Luckily we can at least blame others for their non-compliance but from 429fd6149d3SIlpo Järvinen * inter-operatibility perspective it seems that we're somewhat stuck with 430fd6149d3SIlpo Järvinen * the ordering which we have been using if we want to keep working with 431fd6149d3SIlpo Järvinen * those broken things (not that it currently hurts anybody as there isn't 432fd6149d3SIlpo Järvinen * particular reason why the ordering would need to be changed). 433fd6149d3SIlpo Järvinen * 434fd6149d3SIlpo Järvinen * At least SACK_PERM as the first option is known to lead to a disaster 435fd6149d3SIlpo Järvinen * (but it may well be that other scenarios fail similarly). 436fd6149d3SIlpo Järvinen */ 43733ad798cSAdam Langley static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp, 438bd0388aeSWilliam Allen Simpson struct tcp_out_options *opts) 439bd0388aeSWilliam Allen Simpson { 440bd0388aeSWilliam Allen Simpson u8 options = opts->options; /* mungable copy */ 441bd0388aeSWilliam Allen Simpson 442bd0388aeSWilliam Allen Simpson /* Having both authentication and cookies for security is redundant, 443bd0388aeSWilliam Allen Simpson * and there's certainly not enough room. Instead, the cookie-less 444bd0388aeSWilliam Allen Simpson * extension variant is proposed. 445bd0388aeSWilliam Allen Simpson * 446bd0388aeSWilliam Allen Simpson * Consider the pessimal case with authentication. The options 447bd0388aeSWilliam Allen Simpson * could look like: 448bd0388aeSWilliam Allen Simpson * COOKIE|MD5(20) + MSS(4) + SACK|TS(12) + WSCALE(4) == 40 449bd0388aeSWilliam Allen Simpson */ 450bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_MD5 & options)) { 451bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_COOKIE_EXTENSION & options)) { 452bd0388aeSWilliam Allen Simpson *ptr++ = htonl((TCPOPT_COOKIE << 24) | 453bd0388aeSWilliam Allen Simpson (TCPOLEN_COOKIE_BASE << 16) | 454bd0388aeSWilliam Allen Simpson (TCPOPT_MD5SIG << 8) | 455bd0388aeSWilliam Allen Simpson TCPOLEN_MD5SIG); 456bd0388aeSWilliam Allen Simpson } else { 45733ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 45833ad798cSAdam Langley (TCPOPT_NOP << 16) | 45933ad798cSAdam Langley (TCPOPT_MD5SIG << 8) | 46033ad798cSAdam Langley TCPOLEN_MD5SIG); 461bd0388aeSWilliam Allen Simpson } 462bd0388aeSWilliam Allen Simpson options &= ~OPTION_COOKIE_EXTENSION; 463bd0388aeSWilliam Allen Simpson /* overload cookie hash location */ 464bd0388aeSWilliam Allen Simpson opts->hash_location = (__u8 *)ptr; 46533ad798cSAdam Langley ptr += 4; 46633ad798cSAdam Langley } 46733ad798cSAdam Langley 468fd6149d3SIlpo Järvinen if (unlikely(opts->mss)) { 469fd6149d3SIlpo Järvinen *ptr++ = htonl((TCPOPT_MSS << 24) | 470fd6149d3SIlpo Järvinen (TCPOLEN_MSS << 16) | 471fd6149d3SIlpo Järvinen opts->mss); 472fd6149d3SIlpo Järvinen } 473fd6149d3SIlpo Järvinen 474bd0388aeSWilliam Allen Simpson if (likely(OPTION_TS & options)) { 475bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 47633ad798cSAdam Langley *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | 47733ad798cSAdam Langley (TCPOLEN_SACK_PERM << 16) | 47833ad798cSAdam Langley (TCPOPT_TIMESTAMP << 8) | 47933ad798cSAdam Langley TCPOLEN_TIMESTAMP); 480bd0388aeSWilliam Allen Simpson options &= ~OPTION_SACK_ADVERTISE; 48133ad798cSAdam Langley } else { 482496c98dfSYOSHIFUJI Hideaki *ptr++ = htonl((TCPOPT_NOP << 24) | 48340efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 48440efc6faSStephen Hemminger (TCPOPT_TIMESTAMP << 8) | 48540efc6faSStephen Hemminger TCPOLEN_TIMESTAMP); 48640efc6faSStephen Hemminger } 48733ad798cSAdam Langley *ptr++ = htonl(opts->tsval); 48833ad798cSAdam Langley *ptr++ = htonl(opts->tsecr); 48933ad798cSAdam Langley } 49033ad798cSAdam Langley 491bd0388aeSWilliam Allen Simpson /* Specification requires after timestamp, so do it now. 492bd0388aeSWilliam Allen Simpson * 493bd0388aeSWilliam Allen Simpson * Consider the pessimal case without authentication. The options 494bd0388aeSWilliam Allen Simpson * could look like: 495bd0388aeSWilliam Allen Simpson * MSS(4) + SACK|TS(12) + COOKIE(20) + WSCALE(4) == 40 496bd0388aeSWilliam Allen Simpson */ 497bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_COOKIE_EXTENSION & options)) { 498bd0388aeSWilliam Allen Simpson __u8 *cookie_copy = opts->hash_location; 499bd0388aeSWilliam Allen Simpson u8 cookie_size = opts->hash_size; 500bd0388aeSWilliam Allen Simpson 501bd0388aeSWilliam Allen Simpson /* 8-bit multiple handled in tcp_cookie_size_check() above, 502bd0388aeSWilliam Allen Simpson * and elsewhere. 503bd0388aeSWilliam Allen Simpson */ 504bd0388aeSWilliam Allen Simpson if (0x2 & cookie_size) { 505bd0388aeSWilliam Allen Simpson __u8 *p = (__u8 *)ptr; 506bd0388aeSWilliam Allen Simpson 507bd0388aeSWilliam Allen Simpson /* 16-bit multiple */ 508bd0388aeSWilliam Allen Simpson *p++ = TCPOPT_COOKIE; 509bd0388aeSWilliam Allen Simpson *p++ = TCPOLEN_COOKIE_BASE + cookie_size; 510bd0388aeSWilliam Allen Simpson *p++ = *cookie_copy++; 511bd0388aeSWilliam Allen Simpson *p++ = *cookie_copy++; 512bd0388aeSWilliam Allen Simpson ptr++; 513bd0388aeSWilliam Allen Simpson cookie_size -= 2; 514bd0388aeSWilliam Allen Simpson } else { 515bd0388aeSWilliam Allen Simpson /* 32-bit multiple */ 516bd0388aeSWilliam Allen Simpson *ptr++ = htonl(((TCPOPT_NOP << 24) | 517bd0388aeSWilliam Allen Simpson (TCPOPT_NOP << 16) | 518bd0388aeSWilliam Allen Simpson (TCPOPT_COOKIE << 8) | 519bd0388aeSWilliam Allen Simpson TCPOLEN_COOKIE_BASE) + 520bd0388aeSWilliam Allen Simpson cookie_size); 521bd0388aeSWilliam Allen Simpson } 522bd0388aeSWilliam Allen Simpson 523bd0388aeSWilliam Allen Simpson if (cookie_size > 0) { 524bd0388aeSWilliam Allen Simpson memcpy(ptr, cookie_copy, cookie_size); 525bd0388aeSWilliam Allen Simpson ptr += (cookie_size / 4); 526bd0388aeSWilliam Allen Simpson } 527bd0388aeSWilliam Allen Simpson } 528bd0388aeSWilliam Allen Simpson 529bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 53033ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 53133ad798cSAdam Langley (TCPOPT_NOP << 16) | 53233ad798cSAdam Langley (TCPOPT_SACK_PERM << 8) | 53333ad798cSAdam Langley TCPOLEN_SACK_PERM); 53433ad798cSAdam Langley } 53533ad798cSAdam Langley 536bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_WSCALE & options)) { 53733ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 53833ad798cSAdam Langley (TCPOPT_WINDOW << 16) | 53933ad798cSAdam Langley (TCPOLEN_WINDOW << 8) | 54033ad798cSAdam Langley opts->ws); 54133ad798cSAdam Langley } 54233ad798cSAdam Langley 54333ad798cSAdam Langley if (unlikely(opts->num_sack_blocks)) { 54433ad798cSAdam Langley struct tcp_sack_block *sp = tp->rx_opt.dsack ? 54533ad798cSAdam Langley tp->duplicate_sack : tp->selective_acks; 54640efc6faSStephen Hemminger int this_sack; 54740efc6faSStephen Hemminger 54840efc6faSStephen Hemminger *ptr++ = htonl((TCPOPT_NOP << 24) | 54940efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 55040efc6faSStephen Hemminger (TCPOPT_SACK << 8) | 55133ad798cSAdam Langley (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * 55240efc6faSStephen Hemminger TCPOLEN_SACK_PERBLOCK))); 5532de979bdSStephen Hemminger 55433ad798cSAdam Langley for (this_sack = 0; this_sack < opts->num_sack_blocks; 55533ad798cSAdam Langley ++this_sack) { 55640efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].start_seq); 55740efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].end_seq); 55840efc6faSStephen Hemminger } 5592de979bdSStephen Hemminger 56040efc6faSStephen Hemminger tp->rx_opt.dsack = 0; 56140efc6faSStephen Hemminger } 56240efc6faSStephen Hemminger } 56340efc6faSStephen Hemminger 56467edfef7SAndi Kleen /* Compute TCP options for SYN packets. This is not the final 56567edfef7SAndi Kleen * network wire format yet. 56667edfef7SAndi Kleen */ 56795c96174SEric Dumazet static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb, 56833ad798cSAdam Langley struct tcp_out_options *opts, 569cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 570cf533ea5SEric Dumazet { 57133ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 572bd0388aeSWilliam Allen Simpson struct tcp_cookie_values *cvp = tp->cookie_values; 57395c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 574bd0388aeSWilliam Allen Simpson u8 cookie_size = (!tp->rx_opt.cookie_out_never && cvp != NULL) ? 575bd0388aeSWilliam Allen Simpson tcp_cookie_size_check(cvp->cookie_desired) : 576bd0388aeSWilliam Allen Simpson 0; 57733ad798cSAdam Langley 578cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 57933ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 58033ad798cSAdam Langley if (*md5) { 58133ad798cSAdam Langley opts->options |= OPTION_MD5; 582bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 583cfb6eeb4SYOSHIFUJI Hideaki } 58433ad798cSAdam Langley #else 58533ad798cSAdam Langley *md5 = NULL; 586cfb6eeb4SYOSHIFUJI Hideaki #endif 58733ad798cSAdam Langley 58833ad798cSAdam Langley /* We always get an MSS option. The option bytes which will be seen in 58933ad798cSAdam Langley * normal data packets should timestamps be used, must be in the MSS 59033ad798cSAdam Langley * advertised. But we subtract them from tp->mss_cache so that 59133ad798cSAdam Langley * calculations in tcp_sendmsg are simpler etc. So account for this 59233ad798cSAdam Langley * fact here if necessary. If we don't do this correctly, as a 59333ad798cSAdam Langley * receiver we won't recognize data packets as being full sized when we 59433ad798cSAdam Langley * should, and thus we won't abide by the delayed ACK rules correctly. 59533ad798cSAdam Langley * SACKs don't matter, we never delay an ACK when we have any of those 59633ad798cSAdam Langley * going out. */ 59733ad798cSAdam Langley opts->mss = tcp_advertise_mss(sk); 598bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 59933ad798cSAdam Langley 600bb5b7c11SDavid S. Miller if (likely(sysctl_tcp_timestamps && *md5 == NULL)) { 60133ad798cSAdam Langley opts->options |= OPTION_TS; 60233ad798cSAdam Langley opts->tsval = TCP_SKB_CB(skb)->when; 60333ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 604bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 60533ad798cSAdam Langley } 606bb5b7c11SDavid S. Miller if (likely(sysctl_tcp_window_scaling)) { 60733ad798cSAdam Langley opts->ws = tp->rx_opt.rcv_wscale; 60889e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 609bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 61033ad798cSAdam Langley } 611bb5b7c11SDavid S. Miller if (likely(sysctl_tcp_sack)) { 61233ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 613b32d1310SDavid S. Miller if (unlikely(!(OPTION_TS & opts->options))) 614bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 61533ad798cSAdam Langley } 61633ad798cSAdam Langley 617bd0388aeSWilliam Allen Simpson /* Note that timestamps are required by the specification. 618bd0388aeSWilliam Allen Simpson * 619bd0388aeSWilliam Allen Simpson * Odd numbers of bytes are prohibited by the specification, ensuring 620bd0388aeSWilliam Allen Simpson * that the cookie is 16-bit aligned, and the resulting cookie pair is 621bd0388aeSWilliam Allen Simpson * 32-bit aligned. 622bd0388aeSWilliam Allen Simpson */ 623bd0388aeSWilliam Allen Simpson if (*md5 == NULL && 624bd0388aeSWilliam Allen Simpson (OPTION_TS & opts->options) && 625bd0388aeSWilliam Allen Simpson cookie_size > 0) { 626bd0388aeSWilliam Allen Simpson int need = TCPOLEN_COOKIE_BASE + cookie_size; 627bd0388aeSWilliam Allen Simpson 628bd0388aeSWilliam Allen Simpson if (0x2 & need) { 629bd0388aeSWilliam Allen Simpson /* 32-bit multiple */ 630bd0388aeSWilliam Allen Simpson need += 2; /* NOPs */ 631bd0388aeSWilliam Allen Simpson 632bd0388aeSWilliam Allen Simpson if (need > remaining) { 633bd0388aeSWilliam Allen Simpson /* try shrinking cookie to fit */ 634bd0388aeSWilliam Allen Simpson cookie_size -= 2; 635bd0388aeSWilliam Allen Simpson need -= 4; 636bd0388aeSWilliam Allen Simpson } 637bd0388aeSWilliam Allen Simpson } 638bd0388aeSWilliam Allen Simpson while (need > remaining && TCP_COOKIE_MIN <= cookie_size) { 639bd0388aeSWilliam Allen Simpson cookie_size -= 4; 640bd0388aeSWilliam Allen Simpson need -= 4; 641bd0388aeSWilliam Allen Simpson } 642bd0388aeSWilliam Allen Simpson if (TCP_COOKIE_MIN <= cookie_size) { 643bd0388aeSWilliam Allen Simpson opts->options |= OPTION_COOKIE_EXTENSION; 644bd0388aeSWilliam Allen Simpson opts->hash_location = (__u8 *)&cvp->cookie_pair[0]; 645bd0388aeSWilliam Allen Simpson opts->hash_size = cookie_size; 646bd0388aeSWilliam Allen Simpson 647bd0388aeSWilliam Allen Simpson /* Remember for future incarnations. */ 648bd0388aeSWilliam Allen Simpson cvp->cookie_desired = cookie_size; 649bd0388aeSWilliam Allen Simpson 650bd0388aeSWilliam Allen Simpson if (cvp->cookie_desired != cvp->cookie_pair_size) { 651bd0388aeSWilliam Allen Simpson /* Currently use random bytes as a nonce, 652bd0388aeSWilliam Allen Simpson * assuming these are completely unpredictable 653bd0388aeSWilliam Allen Simpson * by hostile users of the same system. 654bd0388aeSWilliam Allen Simpson */ 655bd0388aeSWilliam Allen Simpson get_random_bytes(&cvp->cookie_pair[0], 656bd0388aeSWilliam Allen Simpson cookie_size); 657bd0388aeSWilliam Allen Simpson cvp->cookie_pair_size = cookie_size; 658bd0388aeSWilliam Allen Simpson } 659bd0388aeSWilliam Allen Simpson 660bd0388aeSWilliam Allen Simpson remaining -= need; 661bd0388aeSWilliam Allen Simpson } 662bd0388aeSWilliam Allen Simpson } 663bd0388aeSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 66433ad798cSAdam Langley } 66533ad798cSAdam Langley 66667edfef7SAndi Kleen /* Set up TCP options for SYN-ACKs. */ 66795c96174SEric Dumazet static unsigned int tcp_synack_options(struct sock *sk, 66833ad798cSAdam Langley struct request_sock *req, 66995c96174SEric Dumazet unsigned int mss, struct sk_buff *skb, 67033ad798cSAdam Langley struct tcp_out_options *opts, 6714957faadSWilliam Allen Simpson struct tcp_md5sig_key **md5, 6724957faadSWilliam Allen Simpson struct tcp_extend_values *xvp) 6734957faadSWilliam Allen Simpson { 67433ad798cSAdam Langley struct inet_request_sock *ireq = inet_rsk(req); 67595c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 6764957faadSWilliam Allen Simpson u8 cookie_plus = (xvp != NULL && !xvp->cookie_out_never) ? 6774957faadSWilliam Allen Simpson xvp->cookie_plus : 6784957faadSWilliam Allen Simpson 0; 67933ad798cSAdam Langley 68033ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 68133ad798cSAdam Langley *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req); 68233ad798cSAdam Langley if (*md5) { 68333ad798cSAdam Langley opts->options |= OPTION_MD5; 6844957faadSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 6854957faadSWilliam Allen Simpson 6864957faadSWilliam Allen Simpson /* We can't fit any SACK blocks in a packet with MD5 + TS 6874957faadSWilliam Allen Simpson * options. There was discussion about disabling SACK 6884957faadSWilliam Allen Simpson * rather than TS in order to fit in better with old, 6894957faadSWilliam Allen Simpson * buggy kernels, but that was deemed to be unnecessary. 6904957faadSWilliam Allen Simpson */ 691de213e5eSEric Dumazet ireq->tstamp_ok &= !ireq->sack_ok; 69233ad798cSAdam Langley } 69333ad798cSAdam Langley #else 69433ad798cSAdam Langley *md5 = NULL; 69533ad798cSAdam Langley #endif 69633ad798cSAdam Langley 6974957faadSWilliam Allen Simpson /* We always send an MSS option. */ 69833ad798cSAdam Langley opts->mss = mss; 6994957faadSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 70033ad798cSAdam Langley 70133ad798cSAdam Langley if (likely(ireq->wscale_ok)) { 70233ad798cSAdam Langley opts->ws = ireq->rcv_wscale; 70389e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 7044957faadSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 70533ad798cSAdam Langley } 706de213e5eSEric Dumazet if (likely(ireq->tstamp_ok)) { 70733ad798cSAdam Langley opts->options |= OPTION_TS; 70833ad798cSAdam Langley opts->tsval = TCP_SKB_CB(skb)->when; 70933ad798cSAdam Langley opts->tsecr = req->ts_recent; 7104957faadSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 71133ad798cSAdam Langley } 71233ad798cSAdam Langley if (likely(ireq->sack_ok)) { 71333ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 714de213e5eSEric Dumazet if (unlikely(!ireq->tstamp_ok)) 7154957faadSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 71633ad798cSAdam Langley } 71733ad798cSAdam Langley 7184957faadSWilliam Allen Simpson /* Similar rationale to tcp_syn_options() applies here, too. 7194957faadSWilliam Allen Simpson * If the <SYN> options fit, the same options should fit now! 7204957faadSWilliam Allen Simpson */ 7214957faadSWilliam Allen Simpson if (*md5 == NULL && 722de213e5eSEric Dumazet ireq->tstamp_ok && 7234957faadSWilliam Allen Simpson cookie_plus > TCPOLEN_COOKIE_BASE) { 7244957faadSWilliam Allen Simpson int need = cookie_plus; /* has TCPOLEN_COOKIE_BASE */ 7254957faadSWilliam Allen Simpson 7264957faadSWilliam Allen Simpson if (0x2 & need) { 7274957faadSWilliam Allen Simpson /* 32-bit multiple */ 7284957faadSWilliam Allen Simpson need += 2; /* NOPs */ 7294957faadSWilliam Allen Simpson } 7304957faadSWilliam Allen Simpson if (need <= remaining) { 7314957faadSWilliam Allen Simpson opts->options |= OPTION_COOKIE_EXTENSION; 7324957faadSWilliam Allen Simpson opts->hash_size = cookie_plus - TCPOLEN_COOKIE_BASE; 7334957faadSWilliam Allen Simpson remaining -= need; 7344957faadSWilliam Allen Simpson } else { 7354957faadSWilliam Allen Simpson /* There's no error return, so flag it. */ 7364957faadSWilliam Allen Simpson xvp->cookie_out_never = 1; /* true */ 7374957faadSWilliam Allen Simpson opts->hash_size = 0; 7384957faadSWilliam Allen Simpson } 7394957faadSWilliam Allen Simpson } 7404957faadSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 74133ad798cSAdam Langley } 74233ad798cSAdam Langley 74367edfef7SAndi Kleen /* Compute TCP options for ESTABLISHED sockets. This is not the 74467edfef7SAndi Kleen * final wire format yet. 74567edfef7SAndi Kleen */ 74695c96174SEric Dumazet static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb, 74733ad798cSAdam Langley struct tcp_out_options *opts, 748cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 749cf533ea5SEric Dumazet { 75033ad798cSAdam Langley struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL; 75133ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 75295c96174SEric Dumazet unsigned int size = 0; 753cabeccbdSIlpo Järvinen unsigned int eff_sacks; 75433ad798cSAdam Langley 75533ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 75633ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 75733ad798cSAdam Langley if (unlikely(*md5)) { 75833ad798cSAdam Langley opts->options |= OPTION_MD5; 75933ad798cSAdam Langley size += TCPOLEN_MD5SIG_ALIGNED; 76033ad798cSAdam Langley } 76133ad798cSAdam Langley #else 76233ad798cSAdam Langley *md5 = NULL; 76333ad798cSAdam Langley #endif 76433ad798cSAdam Langley 76533ad798cSAdam Langley if (likely(tp->rx_opt.tstamp_ok)) { 76633ad798cSAdam Langley opts->options |= OPTION_TS; 76733ad798cSAdam Langley opts->tsval = tcb ? tcb->when : 0; 76833ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 76933ad798cSAdam Langley size += TCPOLEN_TSTAMP_ALIGNED; 77033ad798cSAdam Langley } 77133ad798cSAdam Langley 772cabeccbdSIlpo Järvinen eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; 773cabeccbdSIlpo Järvinen if (unlikely(eff_sacks)) { 77495c96174SEric Dumazet const unsigned int remaining = MAX_TCP_OPTION_SPACE - size; 77533ad798cSAdam Langley opts->num_sack_blocks = 77695c96174SEric Dumazet min_t(unsigned int, eff_sacks, 77733ad798cSAdam Langley (remaining - TCPOLEN_SACK_BASE_ALIGNED) / 77833ad798cSAdam Langley TCPOLEN_SACK_PERBLOCK); 77933ad798cSAdam Langley size += TCPOLEN_SACK_BASE_ALIGNED + 78033ad798cSAdam Langley opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; 78133ad798cSAdam Langley } 78233ad798cSAdam Langley 78333ad798cSAdam Langley return size; 78440efc6faSStephen Hemminger } 7851da177e4SLinus Torvalds 7861da177e4SLinus Torvalds /* This routine actually transmits TCP packets queued in by 7871da177e4SLinus Torvalds * tcp_do_sendmsg(). This is used by both the initial 7881da177e4SLinus Torvalds * transmission and possible later retransmissions. 7891da177e4SLinus Torvalds * All SKB's seen here are completely headerless. It is our 7901da177e4SLinus Torvalds * job to build the TCP header, and pass the packet down to 7911da177e4SLinus Torvalds * IP so it can do the same plus pass the packet off to the 7921da177e4SLinus Torvalds * device. 7931da177e4SLinus Torvalds * 7941da177e4SLinus Torvalds * We are working here with either a clone of the original 7951da177e4SLinus Torvalds * SKB, or a fresh unique copy made by the retransmit engine. 7961da177e4SLinus Torvalds */ 797056834d9SIlpo Järvinen static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, 798056834d9SIlpo Järvinen gfp_t gfp_mask) 7991da177e4SLinus Torvalds { 8006687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 801dfb4b9dcSDavid S. Miller struct inet_sock *inet; 802dfb4b9dcSDavid S. Miller struct tcp_sock *tp; 803dfb4b9dcSDavid S. Miller struct tcp_skb_cb *tcb; 80433ad798cSAdam Langley struct tcp_out_options opts; 80595c96174SEric Dumazet unsigned int tcp_options_size, tcp_header_size; 806cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 8071da177e4SLinus Torvalds struct tcphdr *th; 8081da177e4SLinus Torvalds int err; 8091da177e4SLinus Torvalds 810dfb4b9dcSDavid S. Miller BUG_ON(!skb || !tcp_skb_pcount(skb)); 811dfb4b9dcSDavid S. Miller 812dfb4b9dcSDavid S. Miller /* If congestion control is doing timestamping, we must 813dfb4b9dcSDavid S. Miller * take such a timestamp before we potentially clone/copy. 814dfb4b9dcSDavid S. Miller */ 815164891aaSStephen Hemminger if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP) 816dfb4b9dcSDavid S. Miller __net_timestamp(skb); 817dfb4b9dcSDavid S. Miller 818dfb4b9dcSDavid S. Miller if (likely(clone_it)) { 819dfb4b9dcSDavid S. Miller if (unlikely(skb_cloned(skb))) 820dfb4b9dcSDavid S. Miller skb = pskb_copy(skb, gfp_mask); 821dfb4b9dcSDavid S. Miller else 822dfb4b9dcSDavid S. Miller skb = skb_clone(skb, gfp_mask); 823dfb4b9dcSDavid S. Miller if (unlikely(!skb)) 824dfb4b9dcSDavid S. Miller return -ENOBUFS; 825dfb4b9dcSDavid S. Miller } 826dfb4b9dcSDavid S. Miller 827dfb4b9dcSDavid S. Miller inet = inet_sk(sk); 828dfb4b9dcSDavid S. Miller tp = tcp_sk(sk); 829dfb4b9dcSDavid S. Miller tcb = TCP_SKB_CB(skb); 83033ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 8311da177e4SLinus Torvalds 8324de075e0SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) 83333ad798cSAdam Langley tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); 83433ad798cSAdam Langley else 83533ad798cSAdam Langley tcp_options_size = tcp_established_options(sk, skb, &opts, 83633ad798cSAdam Langley &md5); 83733ad798cSAdam Langley tcp_header_size = tcp_options_size + sizeof(struct tcphdr); 8381da177e4SLinus Torvalds 8393853b584STom Herbert if (tcp_packets_in_flight(tp) == 0) { 8406687e988SArnaldo Carvalho de Melo tcp_ca_event(sk, CA_EVENT_TX_START); 8413853b584STom Herbert skb->ooo_okay = 1; 8423853b584STom Herbert } else 8433853b584STom Herbert skb->ooo_okay = 0; 8441da177e4SLinus Torvalds 845aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 846aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 847e89862f4SDavid S. Miller skb_set_owner_w(skb, sk); 8481da177e4SLinus Torvalds 8491da177e4SLinus Torvalds /* Build TCP header and checksum it. */ 850aa8223c7SArnaldo Carvalho de Melo th = tcp_hdr(skb); 851c720c7e8SEric Dumazet th->source = inet->inet_sport; 852c720c7e8SEric Dumazet th->dest = inet->inet_dport; 8531da177e4SLinus Torvalds th->seq = htonl(tcb->seq); 8541da177e4SLinus Torvalds th->ack_seq = htonl(tp->rcv_nxt); 855df7a3b07SAl Viro *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | 8564de075e0SEric Dumazet tcb->tcp_flags); 857dfb4b9dcSDavid S. Miller 8584de075e0SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) { 8591da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments 8601da177e4SLinus Torvalds * is never scaled. 8611da177e4SLinus Torvalds */ 862600ff0c2SIlpo Järvinen th->window = htons(min(tp->rcv_wnd, 65535U)); 8631da177e4SLinus Torvalds } else { 8641da177e4SLinus Torvalds th->window = htons(tcp_select_window(sk)); 8651da177e4SLinus Torvalds } 8661da177e4SLinus Torvalds th->check = 0; 8671da177e4SLinus Torvalds th->urg_ptr = 0; 8681da177e4SLinus Torvalds 86933f5f57eSIlpo Järvinen /* The urg_mode check is necessary during a below snd_una win probe */ 8707691367dSHerbert Xu if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { 8717691367dSHerbert Xu if (before(tp->snd_up, tcb->seq + 0x10000)) { 8721da177e4SLinus Torvalds th->urg_ptr = htons(tp->snd_up - tcb->seq); 8731da177e4SLinus Torvalds th->urg = 1; 8747691367dSHerbert Xu } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { 8750eae88f3SEric Dumazet th->urg_ptr = htons(0xFFFF); 8767691367dSHerbert Xu th->urg = 1; 8777691367dSHerbert Xu } 8781da177e4SLinus Torvalds } 8791da177e4SLinus Torvalds 880bd0388aeSWilliam Allen Simpson tcp_options_write((__be32 *)(th + 1), tp, &opts); 8814de075e0SEric Dumazet if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0)) 8829e412ba7SIlpo Järvinen TCP_ECN_send(sk, skb, tcp_header_size); 883dfb4b9dcSDavid S. Miller 884cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 885cfb6eeb4SYOSHIFUJI Hideaki /* Calculate the MD5 hash, as we have all we need now */ 886cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 887a465419bSEric Dumazet sk_nocaps_add(sk, NETIF_F_GSO_MASK); 888bd0388aeSWilliam Allen Simpson tp->af_specific->calc_md5_hash(opts.hash_location, 88949a72dfbSAdam Langley md5, sk, NULL, skb); 890cfb6eeb4SYOSHIFUJI Hideaki } 891cfb6eeb4SYOSHIFUJI Hideaki #endif 892cfb6eeb4SYOSHIFUJI Hideaki 893bb296246SHerbert Xu icsk->icsk_af_ops->send_check(sk, skb); 8941da177e4SLinus Torvalds 8954de075e0SEric Dumazet if (likely(tcb->tcp_flags & TCPHDR_ACK)) 896fc6415bcSDavid S. Miller tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); 8971da177e4SLinus Torvalds 8981da177e4SLinus Torvalds if (skb->len != tcp_header_size) 899cf533ea5SEric Dumazet tcp_event_data_sent(tp, sk); 9001da177e4SLinus Torvalds 901bd37a088SWei Yongjun if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) 902aa2ea058STom Herbert TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, 903aa2ea058STom Herbert tcp_skb_pcount(skb)); 9041da177e4SLinus Torvalds 905d9d8da80SDavid S. Miller err = icsk->icsk_af_ops->queue_xmit(skb, &inet->cork.fl); 90683de47cdSHua Zhong if (likely(err <= 0)) 9071da177e4SLinus Torvalds return err; 9081da177e4SLinus Torvalds 9093cfe3baaSIlpo Järvinen tcp_enter_cwr(sk, 1); 9101da177e4SLinus Torvalds 911b9df3cb8SGerrit Renker return net_xmit_eval(err); 9121da177e4SLinus Torvalds } 9131da177e4SLinus Torvalds 91467edfef7SAndi Kleen /* This routine just queues the buffer for sending. 9151da177e4SLinus Torvalds * 9161da177e4SLinus Torvalds * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, 9171da177e4SLinus Torvalds * otherwise socket can stall. 9181da177e4SLinus Torvalds */ 9191da177e4SLinus Torvalds static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) 9201da177e4SLinus Torvalds { 9211da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 9221da177e4SLinus Torvalds 9231da177e4SLinus Torvalds /* Advance write_seq and place onto the write_queue. */ 9241da177e4SLinus Torvalds tp->write_seq = TCP_SKB_CB(skb)->end_seq; 9251da177e4SLinus Torvalds skb_header_release(skb); 926fe067e8aSDavid S. Miller tcp_add_write_queue_tail(sk, skb); 9273ab224beSHideo Aoki sk->sk_wmem_queued += skb->truesize; 9283ab224beSHideo Aoki sk_mem_charge(sk, skb->truesize); 9291da177e4SLinus Torvalds } 9301da177e4SLinus Torvalds 93167edfef7SAndi Kleen /* Initialize TSO segments for a packet. */ 932cf533ea5SEric Dumazet static void tcp_set_skb_tso_segs(const struct sock *sk, struct sk_buff *skb, 933056834d9SIlpo Järvinen unsigned int mss_now) 934f6302d1dSDavid S. Miller { 9358e5b9ddaSHerbert Xu if (skb->len <= mss_now || !sk_can_gso(sk) || 9368e5b9ddaSHerbert Xu skb->ip_summed == CHECKSUM_NONE) { 937f6302d1dSDavid S. Miller /* Avoid the costly divide in the normal 938f6302d1dSDavid S. Miller * non-TSO case. 939f6302d1dSDavid S. Miller */ 9407967168cSHerbert Xu skb_shinfo(skb)->gso_segs = 1; 9417967168cSHerbert Xu skb_shinfo(skb)->gso_size = 0; 9427967168cSHerbert Xu skb_shinfo(skb)->gso_type = 0; 943f6302d1dSDavid S. Miller } else { 944356f89e1SIlpo Järvinen skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now); 9457967168cSHerbert Xu skb_shinfo(skb)->gso_size = mss_now; 946bcd76111SHerbert Xu skb_shinfo(skb)->gso_type = sk->sk_gso_type; 9471da177e4SLinus Torvalds } 9481da177e4SLinus Torvalds } 9491da177e4SLinus Torvalds 95091fed7a1SIlpo Järvinen /* When a modification to fackets out becomes necessary, we need to check 95168f8353bSIlpo Järvinen * skb is counted to fackets_out or not. 95291fed7a1SIlpo Järvinen */ 953cf533ea5SEric Dumazet static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb, 95491fed7a1SIlpo Järvinen int decr) 95591fed7a1SIlpo Järvinen { 956a47e5a98SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 957a47e5a98SIlpo Järvinen 958dc86967bSIlpo Järvinen if (!tp->sacked_out || tcp_is_reno(tp)) 95991fed7a1SIlpo Järvinen return; 96091fed7a1SIlpo Järvinen 9616859d494SIlpo Järvinen if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq)) 96291fed7a1SIlpo Järvinen tp->fackets_out -= decr; 96391fed7a1SIlpo Järvinen } 96491fed7a1SIlpo Järvinen 965797108d1SIlpo Järvinen /* Pcount in the middle of the write queue got changed, we need to do various 966797108d1SIlpo Järvinen * tweaks to fix counters 967797108d1SIlpo Järvinen */ 968cf533ea5SEric Dumazet static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) 969797108d1SIlpo Järvinen { 970797108d1SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 971797108d1SIlpo Järvinen 972797108d1SIlpo Järvinen tp->packets_out -= decr; 973797108d1SIlpo Järvinen 974797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 975797108d1SIlpo Järvinen tp->sacked_out -= decr; 976797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 977797108d1SIlpo Järvinen tp->retrans_out -= decr; 978797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 979797108d1SIlpo Järvinen tp->lost_out -= decr; 980797108d1SIlpo Järvinen 981797108d1SIlpo Järvinen /* Reno case is special. Sigh... */ 982797108d1SIlpo Järvinen if (tcp_is_reno(tp) && decr > 0) 983797108d1SIlpo Järvinen tp->sacked_out -= min_t(u32, tp->sacked_out, decr); 984797108d1SIlpo Järvinen 985797108d1SIlpo Järvinen tcp_adjust_fackets_out(sk, skb, decr); 986797108d1SIlpo Järvinen 987797108d1SIlpo Järvinen if (tp->lost_skb_hint && 988797108d1SIlpo Järvinen before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && 98952cf3cc8SIlpo Järvinen (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))) 990797108d1SIlpo Järvinen tp->lost_cnt_hint -= decr; 991797108d1SIlpo Järvinen 992797108d1SIlpo Järvinen tcp_verify_left_out(tp); 993797108d1SIlpo Järvinen } 994797108d1SIlpo Järvinen 9951da177e4SLinus Torvalds /* Function to create two new TCP segments. Shrinks the given segment 9961da177e4SLinus Torvalds * to the specified size and appends a new segment with the rest of the 9971da177e4SLinus Torvalds * packet to the list. This won't be called frequently, I hope. 9981da177e4SLinus Torvalds * Remember, these are still headerless SKBs at this point. 9991da177e4SLinus Torvalds */ 1000056834d9SIlpo Järvinen int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, 1001056834d9SIlpo Järvinen unsigned int mss_now) 10021da177e4SLinus Torvalds { 10031da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 10041da177e4SLinus Torvalds struct sk_buff *buff; 10056475be16SDavid S. Miller int nsize, old_factor; 1006b60b49eaSHerbert Xu int nlen; 10079ce01461SIlpo Järvinen u8 flags; 10081da177e4SLinus Torvalds 10092fceec13SIlpo Järvinen if (WARN_ON(len > skb->len)) 10102fceec13SIlpo Järvinen return -EINVAL; 10116a438bbeSStephen Hemminger 10121da177e4SLinus Torvalds nsize = skb_headlen(skb) - len; 10131da177e4SLinus Torvalds if (nsize < 0) 10141da177e4SLinus Torvalds nsize = 0; 10151da177e4SLinus Torvalds 10161da177e4SLinus Torvalds if (skb_cloned(skb) && 10171da177e4SLinus Torvalds skb_is_nonlinear(skb) && 10181da177e4SLinus Torvalds pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) 10191da177e4SLinus Torvalds return -ENOMEM; 10201da177e4SLinus Torvalds 10211da177e4SLinus Torvalds /* Get a new skb... force flag on. */ 10221da177e4SLinus Torvalds buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC); 10231da177e4SLinus Torvalds if (buff == NULL) 10241da177e4SLinus Torvalds return -ENOMEM; /* We'll just try again later. */ 1025ef5cb973SHerbert Xu 10263ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 10273ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1028b60b49eaSHerbert Xu nlen = skb->len - len - nsize; 1029b60b49eaSHerbert Xu buff->truesize += nlen; 1030b60b49eaSHerbert Xu skb->truesize -= nlen; 10311da177e4SLinus Torvalds 10321da177e4SLinus Torvalds /* Correct the sequence numbers. */ 10331da177e4SLinus Torvalds TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 10341da177e4SLinus Torvalds TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 10351da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 10361da177e4SLinus Torvalds 10371da177e4SLinus Torvalds /* PSH and FIN should only be set in the second packet. */ 10384de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 10394de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 10404de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1041e14c3cafSHerbert Xu TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; 10421da177e4SLinus Torvalds 104384fa7933SPatrick McHardy if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { 10441da177e4SLinus Torvalds /* Copy and checksum data tail into the new buffer. */ 1045056834d9SIlpo Järvinen buff->csum = csum_partial_copy_nocheck(skb->data + len, 1046056834d9SIlpo Järvinen skb_put(buff, nsize), 10471da177e4SLinus Torvalds nsize, 0); 10481da177e4SLinus Torvalds 10491da177e4SLinus Torvalds skb_trim(skb, len); 10501da177e4SLinus Torvalds 10511da177e4SLinus Torvalds skb->csum = csum_block_sub(skb->csum, buff->csum, len); 10521da177e4SLinus Torvalds } else { 105384fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 10541da177e4SLinus Torvalds skb_split(skb, buff, len); 10551da177e4SLinus Torvalds } 10561da177e4SLinus Torvalds 10571da177e4SLinus Torvalds buff->ip_summed = skb->ip_summed; 10581da177e4SLinus Torvalds 10591da177e4SLinus Torvalds /* Looks stupid, but our code really uses when of 10601da177e4SLinus Torvalds * skbs, which it never sent before. --ANK 10611da177e4SLinus Torvalds */ 10621da177e4SLinus Torvalds TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when; 1063a61bbcf2SPatrick McHardy buff->tstamp = skb->tstamp; 10641da177e4SLinus Torvalds 10656475be16SDavid S. Miller old_factor = tcp_skb_pcount(skb); 10666475be16SDavid S. Miller 10671da177e4SLinus Torvalds /* Fix up tso_factor for both original and new SKB. */ 1068846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1069846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, buff, mss_now); 10701da177e4SLinus Torvalds 10716475be16SDavid S. Miller /* If this packet has been sent out already, we must 10726475be16SDavid S. Miller * adjust the various packet counters. 10736475be16SDavid S. Miller */ 1074cf0b450cSHerbert Xu if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 10756475be16SDavid S. Miller int diff = old_factor - tcp_skb_pcount(skb) - 10766475be16SDavid S. Miller tcp_skb_pcount(buff); 10771da177e4SLinus Torvalds 1078797108d1SIlpo Järvinen if (diff) 1079797108d1SIlpo Järvinen tcp_adjust_pcount(sk, skb, diff); 10801da177e4SLinus Torvalds } 10811da177e4SLinus Torvalds 10821da177e4SLinus Torvalds /* Link BUFF into the send queue. */ 1083f44b5271SDavid S. Miller skb_header_release(buff); 1084fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 10851da177e4SLinus Torvalds 10861da177e4SLinus Torvalds return 0; 10871da177e4SLinus Torvalds } 10881da177e4SLinus Torvalds 10891da177e4SLinus Torvalds /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c 10901da177e4SLinus Torvalds * eventually). The difference is that pulled data not copied, but 10911da177e4SLinus Torvalds * immediately discarded. 10921da177e4SLinus Torvalds */ 1093f2911969SHerbert Xu ~{PmVHI~} static void __pskb_trim_head(struct sk_buff *skb, int len) 10941da177e4SLinus Torvalds { 10951da177e4SLinus Torvalds int i, k, eat; 10961da177e4SLinus Torvalds 10974fa48bf3SEric Dumazet eat = min_t(int, len, skb_headlen(skb)); 10984fa48bf3SEric Dumazet if (eat) { 10994fa48bf3SEric Dumazet __skb_pull(skb, eat); 110022b4a4f2SEric Dumazet skb->avail_size -= eat; 11014fa48bf3SEric Dumazet len -= eat; 11024fa48bf3SEric Dumazet if (!len) 11034fa48bf3SEric Dumazet return; 11044fa48bf3SEric Dumazet } 11051da177e4SLinus Torvalds eat = len; 11061da177e4SLinus Torvalds k = 0; 11071da177e4SLinus Torvalds for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) { 11089e903e08SEric Dumazet int size = skb_frag_size(&skb_shinfo(skb)->frags[i]); 11099e903e08SEric Dumazet 11109e903e08SEric Dumazet if (size <= eat) { 1111aff65da0SIan Campbell skb_frag_unref(skb, i); 11129e903e08SEric Dumazet eat -= size; 11131da177e4SLinus Torvalds } else { 11141da177e4SLinus Torvalds skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i]; 11151da177e4SLinus Torvalds if (eat) { 11161da177e4SLinus Torvalds skb_shinfo(skb)->frags[k].page_offset += eat; 11179e903e08SEric Dumazet skb_frag_size_sub(&skb_shinfo(skb)->frags[k], eat); 11181da177e4SLinus Torvalds eat = 0; 11191da177e4SLinus Torvalds } 11201da177e4SLinus Torvalds k++; 11211da177e4SLinus Torvalds } 11221da177e4SLinus Torvalds } 11231da177e4SLinus Torvalds skb_shinfo(skb)->nr_frags = k; 11241da177e4SLinus Torvalds 112527a884dcSArnaldo Carvalho de Melo skb_reset_tail_pointer(skb); 11261da177e4SLinus Torvalds skb->data_len -= len; 11271da177e4SLinus Torvalds skb->len = skb->data_len; 11281da177e4SLinus Torvalds } 11291da177e4SLinus Torvalds 113067edfef7SAndi Kleen /* Remove acked data from a packet in the transmit queue. */ 11311da177e4SLinus Torvalds int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 11321da177e4SLinus Torvalds { 1133056834d9SIlpo Järvinen if (skb_cloned(skb) && pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) 11341da177e4SLinus Torvalds return -ENOMEM; 11351da177e4SLinus Torvalds 11364fa48bf3SEric Dumazet __pskb_trim_head(skb, len); 11371da177e4SLinus Torvalds 11381da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq += len; 113984fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 11401da177e4SLinus Torvalds 11411da177e4SLinus Torvalds skb->truesize -= len; 11421da177e4SLinus Torvalds sk->sk_wmem_queued -= len; 11433ab224beSHideo Aoki sk_mem_uncharge(sk, len); 11441da177e4SLinus Torvalds sock_set_flag(sk, SOCK_QUEUE_SHRUNK); 11451da177e4SLinus Torvalds 11465b35e1e6SNeal Cardwell /* Any change of skb->len requires recalculation of tso factor. */ 11471da177e4SLinus Torvalds if (tcp_skb_pcount(skb) > 1) 11485b35e1e6SNeal Cardwell tcp_set_skb_tso_segs(sk, skb, tcp_skb_mss(skb)); 11491da177e4SLinus Torvalds 11501da177e4SLinus Torvalds return 0; 11511da177e4SLinus Torvalds } 11521da177e4SLinus Torvalds 115367edfef7SAndi Kleen /* Calculate MSS. Not accounting for SACKs here. */ 115467469601SEric Dumazet int tcp_mtu_to_mss(struct sock *sk, int pmtu) 11555d424d5aSJohn Heffner { 1156cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1157cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 11585d424d5aSJohn Heffner int mss_now; 11595d424d5aSJohn Heffner 11605d424d5aSJohn Heffner /* Calculate base mss without TCP options: 11615d424d5aSJohn Heffner It is MMS_S - sizeof(tcphdr) of rfc1122 11625d424d5aSJohn Heffner */ 11635d424d5aSJohn Heffner mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 11645d424d5aSJohn Heffner 116567469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 116667469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 116767469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 116867469601SEric Dumazet 116967469601SEric Dumazet if (dst && dst_allfrag(dst)) 117067469601SEric Dumazet mss_now -= icsk->icsk_af_ops->net_frag_header_len; 117167469601SEric Dumazet } 117267469601SEric Dumazet 11735d424d5aSJohn Heffner /* Clamp it (mss_clamp does not include tcp options) */ 11745d424d5aSJohn Heffner if (mss_now > tp->rx_opt.mss_clamp) 11755d424d5aSJohn Heffner mss_now = tp->rx_opt.mss_clamp; 11765d424d5aSJohn Heffner 11775d424d5aSJohn Heffner /* Now subtract optional transport overhead */ 11785d424d5aSJohn Heffner mss_now -= icsk->icsk_ext_hdr_len; 11795d424d5aSJohn Heffner 11805d424d5aSJohn Heffner /* Then reserve room for full set of TCP options and 8 bytes of data */ 11815d424d5aSJohn Heffner if (mss_now < 48) 11825d424d5aSJohn Heffner mss_now = 48; 11835d424d5aSJohn Heffner 11845d424d5aSJohn Heffner /* Now subtract TCP options size, not including SACKs */ 11855d424d5aSJohn Heffner mss_now -= tp->tcp_header_len - sizeof(struct tcphdr); 11865d424d5aSJohn Heffner 11875d424d5aSJohn Heffner return mss_now; 11885d424d5aSJohn Heffner } 11895d424d5aSJohn Heffner 11905d424d5aSJohn Heffner /* Inverse of above */ 119167469601SEric Dumazet int tcp_mss_to_mtu(struct sock *sk, int mss) 11925d424d5aSJohn Heffner { 1193cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1194cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 11955d424d5aSJohn Heffner int mtu; 11965d424d5aSJohn Heffner 11975d424d5aSJohn Heffner mtu = mss + 11985d424d5aSJohn Heffner tp->tcp_header_len + 11995d424d5aSJohn Heffner icsk->icsk_ext_hdr_len + 12005d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 12015d424d5aSJohn Heffner 120267469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 120367469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 120467469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 120567469601SEric Dumazet 120667469601SEric Dumazet if (dst && dst_allfrag(dst)) 120767469601SEric Dumazet mtu += icsk->icsk_af_ops->net_frag_header_len; 120867469601SEric Dumazet } 12095d424d5aSJohn Heffner return mtu; 12105d424d5aSJohn Heffner } 12115d424d5aSJohn Heffner 121267edfef7SAndi Kleen /* MTU probing init per socket */ 12135d424d5aSJohn Heffner void tcp_mtup_init(struct sock *sk) 12145d424d5aSJohn Heffner { 12155d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 12165d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 12175d424d5aSJohn Heffner 12185d424d5aSJohn Heffner icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1; 12195d424d5aSJohn Heffner icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 12205d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 12215d424d5aSJohn Heffner icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss); 12225d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 12235d424d5aSJohn Heffner } 12244bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_mtup_init); 12255d424d5aSJohn Heffner 12261da177e4SLinus Torvalds /* This function synchronize snd mss to current pmtu/exthdr set. 12271da177e4SLinus Torvalds 12281da177e4SLinus Torvalds tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 12291da177e4SLinus Torvalds for TCP options, but includes only bare TCP header. 12301da177e4SLinus Torvalds 12311da177e4SLinus Torvalds tp->rx_opt.mss_clamp is mss negotiated at connection setup. 1232caa20d9aSStephen Hemminger It is minimum of user_mss and mss received with SYN. 12331da177e4SLinus Torvalds It also does not include TCP options. 12341da177e4SLinus Torvalds 1235d83d8461SArnaldo Carvalho de Melo inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. 12361da177e4SLinus Torvalds 12371da177e4SLinus Torvalds tp->mss_cache is current effective sending mss, including 12381da177e4SLinus Torvalds all tcp options except for SACKs. It is evaluated, 12391da177e4SLinus Torvalds taking into account current pmtu, but never exceeds 12401da177e4SLinus Torvalds tp->rx_opt.mss_clamp. 12411da177e4SLinus Torvalds 12421da177e4SLinus Torvalds NOTE1. rfc1122 clearly states that advertised MSS 12431da177e4SLinus Torvalds DOES NOT include either tcp or ip options. 12441da177e4SLinus Torvalds 1245d83d8461SArnaldo Carvalho de Melo NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache 1246d83d8461SArnaldo Carvalho de Melo are READ ONLY outside this function. --ANK (980731) 12471da177e4SLinus Torvalds */ 12481da177e4SLinus Torvalds unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 12491da177e4SLinus Torvalds { 12501da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1251d83d8461SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 12525d424d5aSJohn Heffner int mss_now; 12531da177e4SLinus Torvalds 12545d424d5aSJohn Heffner if (icsk->icsk_mtup.search_high > pmtu) 12555d424d5aSJohn Heffner icsk->icsk_mtup.search_high = pmtu; 12561da177e4SLinus Torvalds 12575d424d5aSJohn Heffner mss_now = tcp_mtu_to_mss(sk, pmtu); 1258409d22b4SIlpo Järvinen mss_now = tcp_bound_to_half_wnd(tp, mss_now); 12591da177e4SLinus Torvalds 12601da177e4SLinus Torvalds /* And store cached results */ 1261d83d8461SArnaldo Carvalho de Melo icsk->icsk_pmtu_cookie = pmtu; 12625d424d5aSJohn Heffner if (icsk->icsk_mtup.enabled) 12635d424d5aSJohn Heffner mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); 1264c1b4a7e6SDavid S. Miller tp->mss_cache = mss_now; 12651da177e4SLinus Torvalds 12661da177e4SLinus Torvalds return mss_now; 12671da177e4SLinus Torvalds } 12684bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_sync_mss); 12691da177e4SLinus Torvalds 12701da177e4SLinus Torvalds /* Compute the current effective MSS, taking SACKs and IP options, 12711da177e4SLinus Torvalds * and even PMTU discovery events into account. 12721da177e4SLinus Torvalds */ 12730c54b85fSIlpo Järvinen unsigned int tcp_current_mss(struct sock *sk) 12741da177e4SLinus Torvalds { 1275cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1276cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1277c1b4a7e6SDavid S. Miller u32 mss_now; 127895c96174SEric Dumazet unsigned int header_len; 127933ad798cSAdam Langley struct tcp_out_options opts; 128033ad798cSAdam Langley struct tcp_md5sig_key *md5; 12811da177e4SLinus Torvalds 1282c1b4a7e6SDavid S. Miller mss_now = tp->mss_cache; 1283c1b4a7e6SDavid S. Miller 12841da177e4SLinus Torvalds if (dst) { 12851da177e4SLinus Torvalds u32 mtu = dst_mtu(dst); 1286d83d8461SArnaldo Carvalho de Melo if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 12871da177e4SLinus Torvalds mss_now = tcp_sync_mss(sk, mtu); 12881da177e4SLinus Torvalds } 12891da177e4SLinus Torvalds 129033ad798cSAdam Langley header_len = tcp_established_options(sk, NULL, &opts, &md5) + 129133ad798cSAdam Langley sizeof(struct tcphdr); 129233ad798cSAdam Langley /* The mss_cache is sized based on tp->tcp_header_len, which assumes 129333ad798cSAdam Langley * some common options. If this is an odd packet (because we have SACK 129433ad798cSAdam Langley * blocks etc) then our calculated header_len will be different, and 129533ad798cSAdam Langley * we have to adjust mss_now correspondingly */ 129633ad798cSAdam Langley if (header_len != tp->tcp_header_len) { 129733ad798cSAdam Langley int delta = (int) header_len - tp->tcp_header_len; 129833ad798cSAdam Langley mss_now -= delta; 129933ad798cSAdam Langley } 1300cfb6eeb4SYOSHIFUJI Hideaki 13011da177e4SLinus Torvalds return mss_now; 13021da177e4SLinus Torvalds } 13031da177e4SLinus Torvalds 1304a762a980SDavid S. Miller /* Congestion window validation. (RFC2861) */ 13059e412ba7SIlpo Järvinen static void tcp_cwnd_validate(struct sock *sk) 1306a762a980SDavid S. Miller { 13079e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1308a762a980SDavid S. Miller 1309d436d686SIlpo Järvinen if (tp->packets_out >= tp->snd_cwnd) { 1310a762a980SDavid S. Miller /* Network is feed fully. */ 1311a762a980SDavid S. Miller tp->snd_cwnd_used = 0; 1312a762a980SDavid S. Miller tp->snd_cwnd_stamp = tcp_time_stamp; 1313a762a980SDavid S. Miller } else { 1314a762a980SDavid S. Miller /* Network starves. */ 1315a762a980SDavid S. Miller if (tp->packets_out > tp->snd_cwnd_used) 1316a762a980SDavid S. Miller tp->snd_cwnd_used = tp->packets_out; 1317a762a980SDavid S. Miller 131815d33c07SDavid S. Miller if (sysctl_tcp_slow_start_after_idle && 131915d33c07SDavid S. Miller (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto) 1320a762a980SDavid S. Miller tcp_cwnd_application_limited(sk); 1321a762a980SDavid S. Miller } 1322a762a980SDavid S. Miller } 1323a762a980SDavid S. Miller 13240e3a4803SIlpo Järvinen /* Returns the portion of skb which can be sent right away without 13250e3a4803SIlpo Järvinen * introducing MSS oddities to segment boundaries. In rare cases where 13260e3a4803SIlpo Järvinen * mss_now != mss_cache, we will request caller to create a small skb 13270e3a4803SIlpo Järvinen * per input skb which could be mostly avoided here (if desired). 13285ea3a748SIlpo Järvinen * 13295ea3a748SIlpo Järvinen * We explicitly want to create a request for splitting write queue tail 13305ea3a748SIlpo Järvinen * to a small skb for Nagle purposes while avoiding unnecessary modulos, 13315ea3a748SIlpo Järvinen * thus all the complexity (cwnd_len is always MSS multiple which we 13325ea3a748SIlpo Järvinen * return whenever allowed by the other factors). Basically we need the 13335ea3a748SIlpo Järvinen * modulo only when the receiver window alone is the limiting factor or 13345ea3a748SIlpo Järvinen * when we would be allowed to send the split-due-to-Nagle skb fully. 13350e3a4803SIlpo Järvinen */ 1336cf533ea5SEric Dumazet static unsigned int tcp_mss_split_point(const struct sock *sk, const struct sk_buff *skb, 1337056834d9SIlpo Järvinen unsigned int mss_now, unsigned int cwnd) 1338c1b4a7e6SDavid S. Miller { 1339cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 13400e3a4803SIlpo Järvinen u32 needed, window, cwnd_len; 1341c1b4a7e6SDavid S. Miller 134290840defSIlpo Järvinen window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1343c1b4a7e6SDavid S. Miller cwnd_len = mss_now * cwnd; 13440e3a4803SIlpo Järvinen 13450e3a4803SIlpo Järvinen if (likely(cwnd_len <= window && skb != tcp_write_queue_tail(sk))) 13460e3a4803SIlpo Järvinen return cwnd_len; 13470e3a4803SIlpo Järvinen 13485ea3a748SIlpo Järvinen needed = min(skb->len, window); 13495ea3a748SIlpo Järvinen 135017515408SIlpo Järvinen if (cwnd_len <= needed) 13510e3a4803SIlpo Järvinen return cwnd_len; 13520e3a4803SIlpo Järvinen 13530e3a4803SIlpo Järvinen return needed - needed % mss_now; 1354c1b4a7e6SDavid S. Miller } 1355c1b4a7e6SDavid S. Miller 1356c1b4a7e6SDavid S. Miller /* Can at least one segment of SKB be sent right now, according to the 1357c1b4a7e6SDavid S. Miller * congestion window rules? If so, return how many segments are allowed. 1358c1b4a7e6SDavid S. Miller */ 1359cf533ea5SEric Dumazet static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, 1360cf533ea5SEric Dumazet const struct sk_buff *skb) 1361c1b4a7e6SDavid S. Miller { 1362c1b4a7e6SDavid S. Miller u32 in_flight, cwnd; 1363c1b4a7e6SDavid S. Miller 1364c1b4a7e6SDavid S. Miller /* Don't be strict about the congestion window for the final FIN. */ 13654de075e0SEric Dumazet if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && 13664de075e0SEric Dumazet tcp_skb_pcount(skb) == 1) 1367c1b4a7e6SDavid S. Miller return 1; 1368c1b4a7e6SDavid S. Miller 1369c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1370c1b4a7e6SDavid S. Miller cwnd = tp->snd_cwnd; 1371c1b4a7e6SDavid S. Miller if (in_flight < cwnd) 1372c1b4a7e6SDavid S. Miller return (cwnd - in_flight); 1373c1b4a7e6SDavid S. Miller 1374c1b4a7e6SDavid S. Miller return 0; 1375c1b4a7e6SDavid S. Miller } 1376c1b4a7e6SDavid S. Miller 1377b595076aSUwe Kleine-König /* Initialize TSO state of a skb. 137867edfef7SAndi Kleen * This must be invoked the first time we consider transmitting 1379c1b4a7e6SDavid S. Miller * SKB onto the wire. 1380c1b4a7e6SDavid S. Miller */ 1381cf533ea5SEric Dumazet static int tcp_init_tso_segs(const struct sock *sk, struct sk_buff *skb, 1382056834d9SIlpo Järvinen unsigned int mss_now) 1383c1b4a7e6SDavid S. Miller { 1384c1b4a7e6SDavid S. Miller int tso_segs = tcp_skb_pcount(skb); 1385c1b4a7e6SDavid S. Miller 1386f8269a49SIlpo Järvinen if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { 1387846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1388c1b4a7e6SDavid S. Miller tso_segs = tcp_skb_pcount(skb); 1389c1b4a7e6SDavid S. Miller } 1390c1b4a7e6SDavid S. Miller return tso_segs; 1391c1b4a7e6SDavid S. Miller } 1392c1b4a7e6SDavid S. Miller 139367edfef7SAndi Kleen /* Minshall's variant of the Nagle send check. */ 1394a2a385d6SEric Dumazet static inline bool tcp_minshall_check(const struct tcp_sock *tp) 1395c1b4a7e6SDavid S. Miller { 1396c1b4a7e6SDavid S. Miller return after(tp->snd_sml, tp->snd_una) && 1397c1b4a7e6SDavid S. Miller !after(tp->snd_sml, tp->snd_nxt); 1398c1b4a7e6SDavid S. Miller } 1399c1b4a7e6SDavid S. Miller 1400a2a385d6SEric Dumazet /* Return false, if packet can be sent now without violation Nagle's rules: 1401c1b4a7e6SDavid S. Miller * 1. It is full sized. 1402c1b4a7e6SDavid S. Miller * 2. Or it contains FIN. (already checked by caller) 14036d67e9beSFeng King * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. 1404c1b4a7e6SDavid S. Miller * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 1405c1b4a7e6SDavid S. Miller * With Minshall's modification: all sent small packets are ACKed. 1406c1b4a7e6SDavid S. Miller */ 1407a2a385d6SEric Dumazet static inline bool tcp_nagle_check(const struct tcp_sock *tp, 1408c1b4a7e6SDavid S. Miller const struct sk_buff *skb, 140995c96174SEric Dumazet unsigned int mss_now, int nonagle) 1410c1b4a7e6SDavid S. Miller { 1411a02cec21SEric Dumazet return skb->len < mss_now && 1412c1b4a7e6SDavid S. Miller ((nonagle & TCP_NAGLE_CORK) || 1413a02cec21SEric Dumazet (!nonagle && tp->packets_out && tcp_minshall_check(tp))); 1414c1b4a7e6SDavid S. Miller } 1415c1b4a7e6SDavid S. Miller 1416a2a385d6SEric Dumazet /* Return true if the Nagle test allows this packet to be 1417c1b4a7e6SDavid S. Miller * sent now. 1418c1b4a7e6SDavid S. Miller */ 1419a2a385d6SEric Dumazet static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, 1420c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1421c1b4a7e6SDavid S. Miller { 1422c1b4a7e6SDavid S. Miller /* Nagle rule does not apply to frames, which sit in the middle of the 1423c1b4a7e6SDavid S. Miller * write_queue (they have no chances to get new data). 1424c1b4a7e6SDavid S. Miller * 1425c1b4a7e6SDavid S. Miller * This is implemented in the callers, where they modify the 'nonagle' 1426c1b4a7e6SDavid S. Miller * argument based upon the location of SKB in the send queue. 1427c1b4a7e6SDavid S. Miller */ 1428c1b4a7e6SDavid S. Miller if (nonagle & TCP_NAGLE_PUSH) 1429a2a385d6SEric Dumazet return true; 1430c1b4a7e6SDavid S. Miller 1431d551e454SIlpo Järvinen /* Don't use the nagle rule for urgent data (or for the final FIN). 1432d551e454SIlpo Järvinen * Nagle can be ignored during F-RTO too (see RFC4138). 1433d551e454SIlpo Järvinen */ 143433f5f57eSIlpo Järvinen if (tcp_urg_mode(tp) || (tp->frto_counter == 2) || 14354de075e0SEric Dumazet (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) 1436a2a385d6SEric Dumazet return true; 1437c1b4a7e6SDavid S. Miller 1438c1b4a7e6SDavid S. Miller if (!tcp_nagle_check(tp, skb, cur_mss, nonagle)) 1439a2a385d6SEric Dumazet return true; 1440c1b4a7e6SDavid S. Miller 1441a2a385d6SEric Dumazet return false; 1442c1b4a7e6SDavid S. Miller } 1443c1b4a7e6SDavid S. Miller 1444c1b4a7e6SDavid S. Miller /* Does at least the first segment of SKB fit into the send window? */ 1445a2a385d6SEric Dumazet static bool tcp_snd_wnd_test(const struct tcp_sock *tp, 1446a2a385d6SEric Dumazet const struct sk_buff *skb, 1447056834d9SIlpo Järvinen unsigned int cur_mss) 1448c1b4a7e6SDavid S. Miller { 1449c1b4a7e6SDavid S. Miller u32 end_seq = TCP_SKB_CB(skb)->end_seq; 1450c1b4a7e6SDavid S. Miller 1451c1b4a7e6SDavid S. Miller if (skb->len > cur_mss) 1452c1b4a7e6SDavid S. Miller end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 1453c1b4a7e6SDavid S. Miller 145490840defSIlpo Järvinen return !after(end_seq, tcp_wnd_end(tp)); 1455c1b4a7e6SDavid S. Miller } 1456c1b4a7e6SDavid S. Miller 1457fe067e8aSDavid S. Miller /* This checks if the data bearing packet SKB (usually tcp_send_head(sk)) 1458c1b4a7e6SDavid S. Miller * should be put on the wire right now. If so, it returns the number of 1459c1b4a7e6SDavid S. Miller * packets allowed by the congestion window. 1460c1b4a7e6SDavid S. Miller */ 1461cf533ea5SEric Dumazet static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb, 1462c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1463c1b4a7e6SDavid S. Miller { 1464cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1465c1b4a7e6SDavid S. Miller unsigned int cwnd_quota; 1466c1b4a7e6SDavid S. Miller 1467846998aeSDavid S. Miller tcp_init_tso_segs(sk, skb, cur_mss); 1468c1b4a7e6SDavid S. Miller 1469c1b4a7e6SDavid S. Miller if (!tcp_nagle_test(tp, skb, cur_mss, nonagle)) 1470c1b4a7e6SDavid S. Miller return 0; 1471c1b4a7e6SDavid S. Miller 1472c1b4a7e6SDavid S. Miller cwnd_quota = tcp_cwnd_test(tp, skb); 1473056834d9SIlpo Järvinen if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss)) 1474c1b4a7e6SDavid S. Miller cwnd_quota = 0; 1475c1b4a7e6SDavid S. Miller 1476c1b4a7e6SDavid S. Miller return cwnd_quota; 1477c1b4a7e6SDavid S. Miller } 1478c1b4a7e6SDavid S. Miller 147967edfef7SAndi Kleen /* Test if sending is allowed right now. */ 1480a2a385d6SEric Dumazet bool tcp_may_send_now(struct sock *sk) 1481c1b4a7e6SDavid S. Miller { 1482cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1483fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 1484c1b4a7e6SDavid S. Miller 1485a02cec21SEric Dumazet return skb && 14860c54b85fSIlpo Järvinen tcp_snd_test(sk, skb, tcp_current_mss(sk), 1487c1b4a7e6SDavid S. Miller (tcp_skb_is_last(sk, skb) ? 1488a02cec21SEric Dumazet tp->nonagle : TCP_NAGLE_PUSH)); 1489c1b4a7e6SDavid S. Miller } 1490c1b4a7e6SDavid S. Miller 1491c1b4a7e6SDavid S. Miller /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 1492c1b4a7e6SDavid S. Miller * which is put after SKB on the list. It is very much like 1493c1b4a7e6SDavid S. Miller * tcp_fragment() except that it may make several kinds of assumptions 1494c1b4a7e6SDavid S. Miller * in order to speed up the splitting operation. In particular, we 1495c1b4a7e6SDavid S. Miller * know that all the data is in scatter-gather pages, and that the 1496c1b4a7e6SDavid S. Miller * packet has never been sent out before (and thus is not cloned). 1497c1b4a7e6SDavid S. Miller */ 1498056834d9SIlpo Järvinen static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, 1499c4ead4c5SEric Dumazet unsigned int mss_now, gfp_t gfp) 1500c1b4a7e6SDavid S. Miller { 1501c1b4a7e6SDavid S. Miller struct sk_buff *buff; 1502c1b4a7e6SDavid S. Miller int nlen = skb->len - len; 15039ce01461SIlpo Järvinen u8 flags; 1504c1b4a7e6SDavid S. Miller 1505c1b4a7e6SDavid S. Miller /* All of a TSO frame must be composed of paged data. */ 1506c8ac3774SHerbert Xu if (skb->len != skb->data_len) 1507c8ac3774SHerbert Xu return tcp_fragment(sk, skb, len, mss_now); 1508c1b4a7e6SDavid S. Miller 1509c4ead4c5SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, gfp); 1510c1b4a7e6SDavid S. Miller if (unlikely(buff == NULL)) 1511c1b4a7e6SDavid S. Miller return -ENOMEM; 1512c1b4a7e6SDavid S. Miller 15133ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 15143ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1515b60b49eaSHerbert Xu buff->truesize += nlen; 1516c1b4a7e6SDavid S. Miller skb->truesize -= nlen; 1517c1b4a7e6SDavid S. Miller 1518c1b4a7e6SDavid S. Miller /* Correct the sequence numbers. */ 1519c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 1520c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 1521c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 1522c1b4a7e6SDavid S. Miller 1523c1b4a7e6SDavid S. Miller /* PSH and FIN should only be set in the second packet. */ 15244de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 15254de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 15264de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1527c1b4a7e6SDavid S. Miller 1528c1b4a7e6SDavid S. Miller /* This packet was never sent out yet, so no SACK bits. */ 1529c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->sacked = 0; 1530c1b4a7e6SDavid S. Miller 153184fa7933SPatrick McHardy buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; 1532c1b4a7e6SDavid S. Miller skb_split(skb, buff, len); 1533c1b4a7e6SDavid S. Miller 1534c1b4a7e6SDavid S. Miller /* Fix up tso_factor for both original and new SKB. */ 1535846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1536846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, buff, mss_now); 1537c1b4a7e6SDavid S. Miller 1538c1b4a7e6SDavid S. Miller /* Link BUFF into the send queue. */ 1539c1b4a7e6SDavid S. Miller skb_header_release(buff); 1540fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 1541c1b4a7e6SDavid S. Miller 1542c1b4a7e6SDavid S. Miller return 0; 1543c1b4a7e6SDavid S. Miller } 1544c1b4a7e6SDavid S. Miller 1545c1b4a7e6SDavid S. Miller /* Try to defer sending, if possible, in order to minimize the amount 1546c1b4a7e6SDavid S. Miller * of TSO splitting we do. View it as a kind of TSO Nagle test. 1547c1b4a7e6SDavid S. Miller * 1548c1b4a7e6SDavid S. Miller * This algorithm is from John Heffner. 1549c1b4a7e6SDavid S. Miller */ 1550a2a385d6SEric Dumazet static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb) 1551c1b4a7e6SDavid S. Miller { 15529e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 15536687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 1554c1b4a7e6SDavid S. Miller u32 send_win, cong_win, limit, in_flight; 1555ad9f4f50SEric Dumazet int win_divisor; 1556c1b4a7e6SDavid S. Miller 15574de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 1558ae8064acSJohn Heffner goto send_now; 1559c1b4a7e6SDavid S. Miller 15606687e988SArnaldo Carvalho de Melo if (icsk->icsk_ca_state != TCP_CA_Open) 1561ae8064acSJohn Heffner goto send_now; 1562ae8064acSJohn Heffner 1563ae8064acSJohn Heffner /* Defer for less than two clock ticks. */ 1564bd515c3eSIlpo Järvinen if (tp->tso_deferred && 1565a2acde07SIlpo Järvinen (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1) 1566ae8064acSJohn Heffner goto send_now; 1567908a75c1SDavid S. Miller 1568c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1569c1b4a7e6SDavid S. Miller 1570056834d9SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight)); 1571c1b4a7e6SDavid S. Miller 157290840defSIlpo Järvinen send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1573c1b4a7e6SDavid S. Miller 1574c1b4a7e6SDavid S. Miller /* From in_flight test above, we know that cwnd > in_flight. */ 1575c1b4a7e6SDavid S. Miller cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; 1576c1b4a7e6SDavid S. Miller 1577c1b4a7e6SDavid S. Miller limit = min(send_win, cong_win); 1578c1b4a7e6SDavid S. Miller 1579ba244fe9SDavid S. Miller /* If a full-sized TSO skb can be sent, do it. */ 158082cc1a7aSPeter P Waskiewicz Jr if (limit >= sk->sk_gso_max_size) 1581ae8064acSJohn Heffner goto send_now; 1582ba244fe9SDavid S. Miller 158362ad2761SIlpo Järvinen /* Middle in queue won't get any more data, full sendable already? */ 158462ad2761SIlpo Järvinen if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) 158562ad2761SIlpo Järvinen goto send_now; 158662ad2761SIlpo Järvinen 1587ad9f4f50SEric Dumazet win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor); 1588ad9f4f50SEric Dumazet if (win_divisor) { 1589c1b4a7e6SDavid S. Miller u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); 1590c1b4a7e6SDavid S. Miller 1591c1b4a7e6SDavid S. Miller /* If at least some fraction of a window is available, 1592c1b4a7e6SDavid S. Miller * just use it. 1593c1b4a7e6SDavid S. Miller */ 1594ad9f4f50SEric Dumazet chunk /= win_divisor; 1595c1b4a7e6SDavid S. Miller if (limit >= chunk) 1596ae8064acSJohn Heffner goto send_now; 1597c1b4a7e6SDavid S. Miller } else { 1598c1b4a7e6SDavid S. Miller /* Different approach, try not to defer past a single 1599c1b4a7e6SDavid S. Miller * ACK. Receiver should ACK every other full sized 1600c1b4a7e6SDavid S. Miller * frame, so if we have space for more than 3 frames 1601c1b4a7e6SDavid S. Miller * then send now. 1602c1b4a7e6SDavid S. Miller */ 16036b5a5c0dSNeal Cardwell if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) 1604ae8064acSJohn Heffner goto send_now; 1605c1b4a7e6SDavid S. Miller } 1606c1b4a7e6SDavid S. Miller 1607c1b4a7e6SDavid S. Miller /* Ok, it looks like it is advisable to defer. */ 1608ae8064acSJohn Heffner tp->tso_deferred = 1 | (jiffies << 1); 1609ae8064acSJohn Heffner 1610a2a385d6SEric Dumazet return true; 1611ae8064acSJohn Heffner 1612ae8064acSJohn Heffner send_now: 1613ae8064acSJohn Heffner tp->tso_deferred = 0; 1614a2a385d6SEric Dumazet return false; 1615c1b4a7e6SDavid S. Miller } 1616c1b4a7e6SDavid S. Miller 16175d424d5aSJohn Heffner /* Create a new MTU probe if we are ready. 161867edfef7SAndi Kleen * MTU probe is regularly attempting to increase the path MTU by 161967edfef7SAndi Kleen * deliberately sending larger packets. This discovers routing 162067edfef7SAndi Kleen * changes resulting in larger path MTUs. 162167edfef7SAndi Kleen * 16225d424d5aSJohn Heffner * Returns 0 if we should wait to probe (no cwnd available), 16235d424d5aSJohn Heffner * 1 if a probe was sent, 1624056834d9SIlpo Järvinen * -1 otherwise 1625056834d9SIlpo Järvinen */ 16265d424d5aSJohn Heffner static int tcp_mtu_probe(struct sock *sk) 16275d424d5aSJohn Heffner { 16285d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 16295d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 16305d424d5aSJohn Heffner struct sk_buff *skb, *nskb, *next; 16315d424d5aSJohn Heffner int len; 16325d424d5aSJohn Heffner int probe_size; 163391cc17c0SIlpo Järvinen int size_needed; 16345d424d5aSJohn Heffner int copy; 16355d424d5aSJohn Heffner int mss_now; 16365d424d5aSJohn Heffner 16375d424d5aSJohn Heffner /* Not currently probing/verifying, 16385d424d5aSJohn Heffner * not in recovery, 16395d424d5aSJohn Heffner * have enough cwnd, and 16405d424d5aSJohn Heffner * not SACKing (the variable headers throw things off) */ 16415d424d5aSJohn Heffner if (!icsk->icsk_mtup.enabled || 16425d424d5aSJohn Heffner icsk->icsk_mtup.probe_size || 16435d424d5aSJohn Heffner inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 16445d424d5aSJohn Heffner tp->snd_cwnd < 11 || 1645cabeccbdSIlpo Järvinen tp->rx_opt.num_sacks || tp->rx_opt.dsack) 16465d424d5aSJohn Heffner return -1; 16475d424d5aSJohn Heffner 16485d424d5aSJohn Heffner /* Very simple search strategy: just double the MSS. */ 16490c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 16505d424d5aSJohn Heffner probe_size = 2 * tp->mss_cache; 165191cc17c0SIlpo Järvinen size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; 16525d424d5aSJohn Heffner if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) { 16535d424d5aSJohn Heffner /* TODO: set timer for probe_converge_event */ 16545d424d5aSJohn Heffner return -1; 16555d424d5aSJohn Heffner } 16565d424d5aSJohn Heffner 16575d424d5aSJohn Heffner /* Have enough data in the send queue to probe? */ 16587f9c33e5SIlpo Järvinen if (tp->write_seq - tp->snd_nxt < size_needed) 16595d424d5aSJohn Heffner return -1; 16605d424d5aSJohn Heffner 166191cc17c0SIlpo Järvinen if (tp->snd_wnd < size_needed) 16625d424d5aSJohn Heffner return -1; 166390840defSIlpo Järvinen if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) 16645d424d5aSJohn Heffner return 0; 16655d424d5aSJohn Heffner 1666d67c58e9SIlpo Järvinen /* Do we need to wait to drain cwnd? With none in flight, don't stall */ 1667d67c58e9SIlpo Järvinen if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { 1668d67c58e9SIlpo Järvinen if (!tcp_packets_in_flight(tp)) 16695d424d5aSJohn Heffner return -1; 16705d424d5aSJohn Heffner else 16715d424d5aSJohn Heffner return 0; 16725d424d5aSJohn Heffner } 16735d424d5aSJohn Heffner 16745d424d5aSJohn Heffner /* We're allowed to probe. Build it now. */ 16755d424d5aSJohn Heffner if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL) 16765d424d5aSJohn Heffner return -1; 16773ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 16783ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 16795d424d5aSJohn Heffner 1680fe067e8aSDavid S. Miller skb = tcp_send_head(sk); 16815d424d5aSJohn Heffner 16825d424d5aSJohn Heffner TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 16835d424d5aSJohn Heffner TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 16844de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; 16855d424d5aSJohn Heffner TCP_SKB_CB(nskb)->sacked = 0; 16865d424d5aSJohn Heffner nskb->csum = 0; 168784fa7933SPatrick McHardy nskb->ip_summed = skb->ip_summed; 16885d424d5aSJohn Heffner 168950c4817eSIlpo Järvinen tcp_insert_write_queue_before(nskb, skb, sk); 169050c4817eSIlpo Järvinen 16915d424d5aSJohn Heffner len = 0; 1692234b6860SIlpo Järvinen tcp_for_write_queue_from_safe(skb, next, sk) { 16935d424d5aSJohn Heffner copy = min_t(int, skb->len, probe_size - len); 16945d424d5aSJohn Heffner if (nskb->ip_summed) 16955d424d5aSJohn Heffner skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); 16965d424d5aSJohn Heffner else 16975d424d5aSJohn Heffner nskb->csum = skb_copy_and_csum_bits(skb, 0, 1698056834d9SIlpo Järvinen skb_put(nskb, copy), 1699056834d9SIlpo Järvinen copy, nskb->csum); 17005d424d5aSJohn Heffner 17015d424d5aSJohn Heffner if (skb->len <= copy) { 17025d424d5aSJohn Heffner /* We've eaten all the data from this skb. 17035d424d5aSJohn Heffner * Throw it away. */ 17044de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 1705fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 17063ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 17075d424d5aSJohn Heffner } else { 17084de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & 1709a3433f35SChangli Gao ~(TCPHDR_FIN|TCPHDR_PSH); 17105d424d5aSJohn Heffner if (!skb_shinfo(skb)->nr_frags) { 17115d424d5aSJohn Heffner skb_pull(skb, copy); 171284fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 1713056834d9SIlpo Järvinen skb->csum = csum_partial(skb->data, 1714056834d9SIlpo Järvinen skb->len, 0); 17155d424d5aSJohn Heffner } else { 17165d424d5aSJohn Heffner __pskb_trim_head(skb, copy); 17175d424d5aSJohn Heffner tcp_set_skb_tso_segs(sk, skb, mss_now); 17185d424d5aSJohn Heffner } 17195d424d5aSJohn Heffner TCP_SKB_CB(skb)->seq += copy; 17205d424d5aSJohn Heffner } 17215d424d5aSJohn Heffner 17225d424d5aSJohn Heffner len += copy; 1723234b6860SIlpo Järvinen 1724234b6860SIlpo Järvinen if (len >= probe_size) 1725234b6860SIlpo Järvinen break; 17265d424d5aSJohn Heffner } 17275d424d5aSJohn Heffner tcp_init_tso_segs(sk, nskb, nskb->len); 17285d424d5aSJohn Heffner 17295d424d5aSJohn Heffner /* We're ready to send. If this fails, the probe will 17305d424d5aSJohn Heffner * be resegmented into mss-sized pieces by tcp_write_xmit(). */ 17315d424d5aSJohn Heffner TCP_SKB_CB(nskb)->when = tcp_time_stamp; 17325d424d5aSJohn Heffner if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 17335d424d5aSJohn Heffner /* Decrement cwnd here because we are sending 17345d424d5aSJohn Heffner * effectively two packets. */ 17355d424d5aSJohn Heffner tp->snd_cwnd--; 173666f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, nskb); 17375d424d5aSJohn Heffner 17385d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 17390e7b1368SJohn Heffner tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 17400e7b1368SJohn Heffner tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 17415d424d5aSJohn Heffner 17425d424d5aSJohn Heffner return 1; 17435d424d5aSJohn Heffner } 17445d424d5aSJohn Heffner 17455d424d5aSJohn Heffner return -1; 17465d424d5aSJohn Heffner } 17475d424d5aSJohn Heffner 17481da177e4SLinus Torvalds /* This routine writes packets to the network. It advances the 17491da177e4SLinus Torvalds * send_head. This happens as incoming acks open up the remote 17501da177e4SLinus Torvalds * window for us. 17511da177e4SLinus Torvalds * 1752f8269a49SIlpo Järvinen * LARGESEND note: !tcp_urg_mode is overkill, only frames between 1753f8269a49SIlpo Järvinen * snd_up-64k-mss .. snd_up cannot be large. However, taking into 1754f8269a49SIlpo Järvinen * account rare use of URG, this is not a big flaw. 1755f8269a49SIlpo Järvinen * 1756a2a385d6SEric Dumazet * Returns true, if no segments are in flight and we have queued segments, 1757a2a385d6SEric Dumazet * but cannot send anything now because of SWS or another problem. 17581da177e4SLinus Torvalds */ 1759a2a385d6SEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 1760d5dd9175SIlpo Järvinen int push_one, gfp_t gfp) 17611da177e4SLinus Torvalds { 17621da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 176392df7b51SDavid S. Miller struct sk_buff *skb; 1764c1b4a7e6SDavid S. Miller unsigned int tso_segs, sent_pkts; 1765c1b4a7e6SDavid S. Miller int cwnd_quota; 17665d424d5aSJohn Heffner int result; 17671da177e4SLinus Torvalds 1768c1b4a7e6SDavid S. Miller sent_pkts = 0; 17695d424d5aSJohn Heffner 1770d5dd9175SIlpo Järvinen if (!push_one) { 17715d424d5aSJohn Heffner /* Do MTU probing. */ 1772d5dd9175SIlpo Järvinen result = tcp_mtu_probe(sk); 1773d5dd9175SIlpo Järvinen if (!result) { 1774a2a385d6SEric Dumazet return false; 17755d424d5aSJohn Heffner } else if (result > 0) { 17765d424d5aSJohn Heffner sent_pkts = 1; 17775d424d5aSJohn Heffner } 1778d5dd9175SIlpo Järvinen } 17795d424d5aSJohn Heffner 1780fe067e8aSDavid S. Miller while ((skb = tcp_send_head(sk))) { 1781c8ac3774SHerbert Xu unsigned int limit; 1782c8ac3774SHerbert Xu 1783b68e9f85SHerbert Xu tso_segs = tcp_init_tso_segs(sk, skb, mss_now); 1784c1b4a7e6SDavid S. Miller BUG_ON(!tso_segs); 1785c1b4a7e6SDavid S. Miller 1786b68e9f85SHerbert Xu cwnd_quota = tcp_cwnd_test(tp, skb); 1787b68e9f85SHerbert Xu if (!cwnd_quota) 1788b68e9f85SHerbert Xu break; 1789b68e9f85SHerbert Xu 1790b68e9f85SHerbert Xu if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) 1791b68e9f85SHerbert Xu break; 1792b68e9f85SHerbert Xu 1793c1b4a7e6SDavid S. Miller if (tso_segs == 1) { 1794aa93466bSDavid S. Miller if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 1795aa93466bSDavid S. Miller (tcp_skb_is_last(sk, skb) ? 1796aa93466bSDavid S. Miller nonagle : TCP_NAGLE_PUSH)))) 1797aa93466bSDavid S. Miller break; 1798c1b4a7e6SDavid S. Miller } else { 1799d5dd9175SIlpo Järvinen if (!push_one && tcp_tso_should_defer(sk, skb)) 1800aa93466bSDavid S. Miller break; 1801c1b4a7e6SDavid S. Miller } 1802aa93466bSDavid S. Miller 1803c8ac3774SHerbert Xu limit = mss_now; 1804f8269a49SIlpo Järvinen if (tso_segs > 1 && !tcp_urg_mode(tp)) 18050e3a4803SIlpo Järvinen limit = tcp_mss_split_point(sk, skb, mss_now, 18060e3a4803SIlpo Järvinen cwnd_quota); 1807c8ac3774SHerbert Xu 1808c8ac3774SHerbert Xu if (skb->len > limit && 1809c4ead4c5SEric Dumazet unlikely(tso_fragment(sk, skb, limit, mss_now, gfp))) 18101da177e4SLinus Torvalds break; 18111da177e4SLinus Torvalds 18121da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 1813c1b4a7e6SDavid S. Miller 1814d5dd9175SIlpo Järvinen if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) 18151da177e4SLinus Torvalds break; 18161da177e4SLinus Torvalds 18171da177e4SLinus Torvalds /* Advance the send_head. This one is sent out. 18181da177e4SLinus Torvalds * This call will increment packets_out. 18191da177e4SLinus Torvalds */ 182066f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 18211da177e4SLinus Torvalds 18221da177e4SLinus Torvalds tcp_minshall_update(tp, mss_now, skb); 1823a262f0cdSNandita Dukkipati sent_pkts += tcp_skb_pcount(skb); 1824d5dd9175SIlpo Järvinen 1825d5dd9175SIlpo Järvinen if (push_one) 1826d5dd9175SIlpo Järvinen break; 18271da177e4SLinus Torvalds } 1828a262f0cdSNandita Dukkipati if (inet_csk(sk)->icsk_ca_state == TCP_CA_Recovery) 1829a262f0cdSNandita Dukkipati tp->prr_out += sent_pkts; 18301da177e4SLinus Torvalds 1831aa93466bSDavid S. Miller if (likely(sent_pkts)) { 18329e412ba7SIlpo Järvinen tcp_cwnd_validate(sk); 1833a2a385d6SEric Dumazet return false; 18341da177e4SLinus Torvalds } 1835fe067e8aSDavid S. Miller return !tp->packets_out && tcp_send_head(sk); 18361da177e4SLinus Torvalds } 18371da177e4SLinus Torvalds 1838a762a980SDavid S. Miller /* Push out any pending frames which were held back due to 1839a762a980SDavid S. Miller * TCP_CORK or attempt at coalescing tiny packets. 1840a762a980SDavid S. Miller * The socket must be locked by the caller. 1841a762a980SDavid S. Miller */ 18429e412ba7SIlpo Järvinen void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 18439e412ba7SIlpo Järvinen int nonagle) 1844a762a980SDavid S. Miller { 1845726e07a8SIlpo Järvinen /* If we are closed, the bytes will have to remain here. 1846726e07a8SIlpo Järvinen * In time closedown will finish, we empty the write queue and 1847726e07a8SIlpo Järvinen * all will be happy. 1848726e07a8SIlpo Järvinen */ 1849726e07a8SIlpo Järvinen if (unlikely(sk->sk_state == TCP_CLOSE)) 1850726e07a8SIlpo Järvinen return; 1851726e07a8SIlpo Järvinen 1852d5dd9175SIlpo Järvinen if (tcp_write_xmit(sk, cur_mss, nonagle, 0, GFP_ATOMIC)) 18539e412ba7SIlpo Järvinen tcp_check_probe_timer(sk); 1854a762a980SDavid S. Miller } 1855a762a980SDavid S. Miller 1856c1b4a7e6SDavid S. Miller /* Send _single_ skb sitting at the send head. This function requires 1857c1b4a7e6SDavid S. Miller * true push pending frames to setup probe timer etc. 1858c1b4a7e6SDavid S. Miller */ 1859c1b4a7e6SDavid S. Miller void tcp_push_one(struct sock *sk, unsigned int mss_now) 1860c1b4a7e6SDavid S. Miller { 1861fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 1862c1b4a7e6SDavid S. Miller 1863c1b4a7e6SDavid S. Miller BUG_ON(!skb || skb->len < mss_now); 1864c1b4a7e6SDavid S. Miller 1865d5dd9175SIlpo Järvinen tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); 1866c1b4a7e6SDavid S. Miller } 1867c1b4a7e6SDavid S. Miller 18681da177e4SLinus Torvalds /* This function returns the amount that we can raise the 18691da177e4SLinus Torvalds * usable window based on the following constraints 18701da177e4SLinus Torvalds * 18711da177e4SLinus Torvalds * 1. The window can never be shrunk once it is offered (RFC 793) 18721da177e4SLinus Torvalds * 2. We limit memory per socket 18731da177e4SLinus Torvalds * 18741da177e4SLinus Torvalds * RFC 1122: 18751da177e4SLinus Torvalds * "the suggested [SWS] avoidance algorithm for the receiver is to keep 18761da177e4SLinus Torvalds * RECV.NEXT + RCV.WIN fixed until: 18771da177e4SLinus Torvalds * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 18781da177e4SLinus Torvalds * 18791da177e4SLinus Torvalds * i.e. don't raise the right edge of the window until you can raise 18801da177e4SLinus Torvalds * it at least MSS bytes. 18811da177e4SLinus Torvalds * 18821da177e4SLinus Torvalds * Unfortunately, the recommended algorithm breaks header prediction, 18831da177e4SLinus Torvalds * since header prediction assumes th->window stays fixed. 18841da177e4SLinus Torvalds * 18851da177e4SLinus Torvalds * Strictly speaking, keeping th->window fixed violates the receiver 18861da177e4SLinus Torvalds * side SWS prevention criteria. The problem is that under this rule 18871da177e4SLinus Torvalds * a stream of single byte packets will cause the right side of the 18881da177e4SLinus Torvalds * window to always advance by a single byte. 18891da177e4SLinus Torvalds * 18901da177e4SLinus Torvalds * Of course, if the sender implements sender side SWS prevention 18911da177e4SLinus Torvalds * then this will not be a problem. 18921da177e4SLinus Torvalds * 18931da177e4SLinus Torvalds * BSD seems to make the following compromise: 18941da177e4SLinus Torvalds * 18951da177e4SLinus Torvalds * If the free space is less than the 1/4 of the maximum 18961da177e4SLinus Torvalds * space available and the free space is less than 1/2 mss, 18971da177e4SLinus Torvalds * then set the window to 0. 18981da177e4SLinus Torvalds * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 18991da177e4SLinus Torvalds * Otherwise, just prevent the window from shrinking 19001da177e4SLinus Torvalds * and from being larger than the largest representable value. 19011da177e4SLinus Torvalds * 19021da177e4SLinus Torvalds * This prevents incremental opening of the window in the regime 19031da177e4SLinus Torvalds * where TCP is limited by the speed of the reader side taking 19041da177e4SLinus Torvalds * data out of the TCP receive queue. It does nothing about 19051da177e4SLinus Torvalds * those cases where the window is constrained on the sender side 19061da177e4SLinus Torvalds * because the pipeline is full. 19071da177e4SLinus Torvalds * 19081da177e4SLinus Torvalds * BSD also seems to "accidentally" limit itself to windows that are a 19091da177e4SLinus Torvalds * multiple of MSS, at least until the free space gets quite small. 19101da177e4SLinus Torvalds * This would appear to be a side effect of the mbuf implementation. 19111da177e4SLinus Torvalds * Combining these two algorithms results in the observed behavior 19121da177e4SLinus Torvalds * of having a fixed window size at almost all times. 19131da177e4SLinus Torvalds * 19141da177e4SLinus Torvalds * Below we obtain similar behavior by forcing the offered window to 19151da177e4SLinus Torvalds * a multiple of the mss when it is feasible to do so. 19161da177e4SLinus Torvalds * 19171da177e4SLinus Torvalds * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 19181da177e4SLinus Torvalds * Regular options like TIMESTAMP are taken into account. 19191da177e4SLinus Torvalds */ 19201da177e4SLinus Torvalds u32 __tcp_select_window(struct sock *sk) 19211da177e4SLinus Torvalds { 1922463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 19231da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1924caa20d9aSStephen Hemminger /* MSS for the peer's data. Previous versions used mss_clamp 19251da177e4SLinus Torvalds * here. I don't know if the value based on our guesses 19261da177e4SLinus Torvalds * of peer's MSS is better for the performance. It's more correct 19271da177e4SLinus Torvalds * but may be worse for the performance because of rcv_mss 19281da177e4SLinus Torvalds * fluctuations. --SAW 1998/11/1 19291da177e4SLinus Torvalds */ 1930463c84b9SArnaldo Carvalho de Melo int mss = icsk->icsk_ack.rcv_mss; 19311da177e4SLinus Torvalds int free_space = tcp_space(sk); 19321da177e4SLinus Torvalds int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk)); 19331da177e4SLinus Torvalds int window; 19341da177e4SLinus Torvalds 19351da177e4SLinus Torvalds if (mss > full_space) 19361da177e4SLinus Torvalds mss = full_space; 19371da177e4SLinus Torvalds 1938b92edbe0SEric Dumazet if (free_space < (full_space >> 1)) { 1939463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.quick = 0; 19401da177e4SLinus Torvalds 1941180d8cd9SGlauber Costa if (sk_under_memory_pressure(sk)) 1942056834d9SIlpo Järvinen tp->rcv_ssthresh = min(tp->rcv_ssthresh, 1943056834d9SIlpo Järvinen 4U * tp->advmss); 19441da177e4SLinus Torvalds 19451da177e4SLinus Torvalds if (free_space < mss) 19461da177e4SLinus Torvalds return 0; 19471da177e4SLinus Torvalds } 19481da177e4SLinus Torvalds 19491da177e4SLinus Torvalds if (free_space > tp->rcv_ssthresh) 19501da177e4SLinus Torvalds free_space = tp->rcv_ssthresh; 19511da177e4SLinus Torvalds 19521da177e4SLinus Torvalds /* Don't do rounding if we are using window scaling, since the 19531da177e4SLinus Torvalds * scaled window will not line up with the MSS boundary anyway. 19541da177e4SLinus Torvalds */ 19551da177e4SLinus Torvalds window = tp->rcv_wnd; 19561da177e4SLinus Torvalds if (tp->rx_opt.rcv_wscale) { 19571da177e4SLinus Torvalds window = free_space; 19581da177e4SLinus Torvalds 19591da177e4SLinus Torvalds /* Advertise enough space so that it won't get scaled away. 19601da177e4SLinus Torvalds * Import case: prevent zero window announcement if 19611da177e4SLinus Torvalds * 1<<rcv_wscale > mss. 19621da177e4SLinus Torvalds */ 19631da177e4SLinus Torvalds if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window) 19641da177e4SLinus Torvalds window = (((window >> tp->rx_opt.rcv_wscale) + 1) 19651da177e4SLinus Torvalds << tp->rx_opt.rcv_wscale); 19661da177e4SLinus Torvalds } else { 19671da177e4SLinus Torvalds /* Get the largest window that is a nice multiple of mss. 19681da177e4SLinus Torvalds * Window clamp already applied above. 19691da177e4SLinus Torvalds * If our current window offering is within 1 mss of the 19701da177e4SLinus Torvalds * free space we just keep it. This prevents the divide 19711da177e4SLinus Torvalds * and multiply from happening most of the time. 19721da177e4SLinus Torvalds * We also don't do any window rounding when the free space 19731da177e4SLinus Torvalds * is too small. 19741da177e4SLinus Torvalds */ 19751da177e4SLinus Torvalds if (window <= free_space - mss || window > free_space) 19761da177e4SLinus Torvalds window = (free_space / mss) * mss; 197784565070SJohn Heffner else if (mss == full_space && 1978b92edbe0SEric Dumazet free_space > window + (full_space >> 1)) 197984565070SJohn Heffner window = free_space; 19801da177e4SLinus Torvalds } 19811da177e4SLinus Torvalds 19821da177e4SLinus Torvalds return window; 19831da177e4SLinus Torvalds } 19841da177e4SLinus Torvalds 19854a17fc3aSIlpo Järvinen /* Collapses two adjacent SKB's during retransmission. */ 19864a17fc3aSIlpo Järvinen static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) 19871da177e4SLinus Torvalds { 19881da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1989fe067e8aSDavid S. Miller struct sk_buff *next_skb = tcp_write_queue_next(sk, skb); 1990058dc334SIlpo Järvinen int skb_size, next_skb_size; 19911da177e4SLinus Torvalds 1992058dc334SIlpo Järvinen skb_size = skb->len; 1993058dc334SIlpo Järvinen next_skb_size = next_skb->len; 19941da177e4SLinus Torvalds 1995058dc334SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); 19961da177e4SLinus Torvalds 19976859d494SIlpo Järvinen tcp_highest_sack_combine(sk, next_skb, skb); 1998a6963a6bSIlpo Järvinen 1999fe067e8aSDavid S. Miller tcp_unlink_write_queue(next_skb, sk); 20001da177e4SLinus Torvalds 2001058dc334SIlpo Järvinen skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size), 20021a4e2d09SArnaldo Carvalho de Melo next_skb_size); 20031da177e4SLinus Torvalds 200452d570aaSJarek Poplawski if (next_skb->ip_summed == CHECKSUM_PARTIAL) 200552d570aaSJarek Poplawski skb->ip_summed = CHECKSUM_PARTIAL; 20061da177e4SLinus Torvalds 200784fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 20081da177e4SLinus Torvalds skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); 20091da177e4SLinus Torvalds 20101da177e4SLinus Torvalds /* Update sequence range on original skb. */ 20111da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 20121da177e4SLinus Torvalds 2013e6c7d085SIlpo Järvinen /* Merge over control information. This moves PSH/FIN etc. over */ 20144de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; 20151da177e4SLinus Torvalds 20161da177e4SLinus Torvalds /* All done, get rid of second SKB and account for it so 20171da177e4SLinus Torvalds * packet counting does not break. 20181da177e4SLinus Torvalds */ 20194828e7f4SIlpo Järvinen TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; 2020b7689205SIlpo Järvinen 2021b7689205SIlpo Järvinen /* changed transmit queue under us so clear hints */ 2022ef9da47cSIlpo Järvinen tcp_clear_retrans_hints_partial(tp); 2023ef9da47cSIlpo Järvinen if (next_skb == tp->retransmit_skb_hint) 2024ef9da47cSIlpo Järvinen tp->retransmit_skb_hint = skb; 2025b7689205SIlpo Järvinen 2026797108d1SIlpo Järvinen tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); 2027797108d1SIlpo Järvinen 20283ab224beSHideo Aoki sk_wmem_free_skb(sk, next_skb); 20291da177e4SLinus Torvalds } 20301da177e4SLinus Torvalds 203167edfef7SAndi Kleen /* Check if coalescing SKBs is legal. */ 2032a2a385d6SEric Dumazet static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) 20334a17fc3aSIlpo Järvinen { 20344a17fc3aSIlpo Järvinen if (tcp_skb_pcount(skb) > 1) 2035a2a385d6SEric Dumazet return false; 20364a17fc3aSIlpo Järvinen /* TODO: SACK collapsing could be used to remove this condition */ 20374a17fc3aSIlpo Järvinen if (skb_shinfo(skb)->nr_frags != 0) 2038a2a385d6SEric Dumazet return false; 20394a17fc3aSIlpo Järvinen if (skb_cloned(skb)) 2040a2a385d6SEric Dumazet return false; 20414a17fc3aSIlpo Järvinen if (skb == tcp_send_head(sk)) 2042a2a385d6SEric Dumazet return false; 20434a17fc3aSIlpo Järvinen /* Some heurestics for collapsing over SACK'd could be invented */ 20444a17fc3aSIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 2045a2a385d6SEric Dumazet return false; 20464a17fc3aSIlpo Järvinen 2047a2a385d6SEric Dumazet return true; 20484a17fc3aSIlpo Järvinen } 20494a17fc3aSIlpo Järvinen 205067edfef7SAndi Kleen /* Collapse packets in the retransmit queue to make to create 205167edfef7SAndi Kleen * less packets on the wire. This is only done on retransmission. 205267edfef7SAndi Kleen */ 20534a17fc3aSIlpo Järvinen static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, 20544a17fc3aSIlpo Järvinen int space) 20554a17fc3aSIlpo Järvinen { 20564a17fc3aSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 20574a17fc3aSIlpo Järvinen struct sk_buff *skb = to, *tmp; 2058a2a385d6SEric Dumazet bool first = true; 20594a17fc3aSIlpo Järvinen 20604a17fc3aSIlpo Järvinen if (!sysctl_tcp_retrans_collapse) 20614a17fc3aSIlpo Järvinen return; 20624de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 20634a17fc3aSIlpo Järvinen return; 20644a17fc3aSIlpo Järvinen 20654a17fc3aSIlpo Järvinen tcp_for_write_queue_from_safe(skb, tmp, sk) { 20664a17fc3aSIlpo Järvinen if (!tcp_can_collapse(sk, skb)) 20674a17fc3aSIlpo Järvinen break; 20684a17fc3aSIlpo Järvinen 20694a17fc3aSIlpo Järvinen space -= skb->len; 20704a17fc3aSIlpo Järvinen 20714a17fc3aSIlpo Järvinen if (first) { 2072a2a385d6SEric Dumazet first = false; 20734a17fc3aSIlpo Järvinen continue; 20744a17fc3aSIlpo Järvinen } 20754a17fc3aSIlpo Järvinen 20764a17fc3aSIlpo Järvinen if (space < 0) 20774a17fc3aSIlpo Järvinen break; 20784a17fc3aSIlpo Järvinen /* Punt if not enough space exists in the first SKB for 20794a17fc3aSIlpo Järvinen * the data in the second 20804a17fc3aSIlpo Järvinen */ 2081a21d4572SEric Dumazet if (skb->len > skb_availroom(to)) 20824a17fc3aSIlpo Järvinen break; 20834a17fc3aSIlpo Järvinen 20844a17fc3aSIlpo Järvinen if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) 20854a17fc3aSIlpo Järvinen break; 20864a17fc3aSIlpo Järvinen 20874a17fc3aSIlpo Järvinen tcp_collapse_retrans(sk, to); 20884a17fc3aSIlpo Järvinen } 20894a17fc3aSIlpo Järvinen } 20904a17fc3aSIlpo Järvinen 20911da177e4SLinus Torvalds /* This retransmits one SKB. Policy decisions and retransmit queue 20921da177e4SLinus Torvalds * state updates are done by the caller. Returns non-zero if an 20931da177e4SLinus Torvalds * error occurred which prevented the send. 20941da177e4SLinus Torvalds */ 20951da177e4SLinus Torvalds int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) 20961da177e4SLinus Torvalds { 20971da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 20985d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 20997d227cd2SSridhar Samudrala unsigned int cur_mss; 21001da177e4SLinus Torvalds int err; 21011da177e4SLinus Torvalds 21025d424d5aSJohn Heffner /* Inconslusive MTU probe */ 21035d424d5aSJohn Heffner if (icsk->icsk_mtup.probe_size) { 21045d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 21055d424d5aSJohn Heffner } 21065d424d5aSJohn Heffner 21071da177e4SLinus Torvalds /* Do not sent more than we queued. 1/4 is reserved for possible 2108caa20d9aSStephen Hemminger * copying overhead: fragmentation, tunneling, mangling etc. 21091da177e4SLinus Torvalds */ 21101da177e4SLinus Torvalds if (atomic_read(&sk->sk_wmem_alloc) > 21111da177e4SLinus Torvalds min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf)) 21121da177e4SLinus Torvalds return -EAGAIN; 21131da177e4SLinus Torvalds 21141da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 21151da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 21161da177e4SLinus Torvalds BUG(); 21171da177e4SLinus Torvalds if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 21181da177e4SLinus Torvalds return -ENOMEM; 21191da177e4SLinus Torvalds } 21201da177e4SLinus Torvalds 21217d227cd2SSridhar Samudrala if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 21227d227cd2SSridhar Samudrala return -EHOSTUNREACH; /* Routing failure or similar. */ 21237d227cd2SSridhar Samudrala 21240c54b85fSIlpo Järvinen cur_mss = tcp_current_mss(sk); 21257d227cd2SSridhar Samudrala 21261da177e4SLinus Torvalds /* If receiver has shrunk his window, and skb is out of 21271da177e4SLinus Torvalds * new window, do not retransmit it. The exception is the 21281da177e4SLinus Torvalds * case, when window is shrunk to zero. In this case 21291da177e4SLinus Torvalds * our retransmit serves as a zero window probe. 21301da177e4SLinus Torvalds */ 21319d4fb27dSJoe Perches if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && 21329d4fb27dSJoe Perches TCP_SKB_CB(skb)->seq != tp->snd_una) 21331da177e4SLinus Torvalds return -EAGAIN; 21341da177e4SLinus Torvalds 21351da177e4SLinus Torvalds if (skb->len > cur_mss) { 2136846998aeSDavid S. Miller if (tcp_fragment(sk, skb, cur_mss, cur_mss)) 21371da177e4SLinus Torvalds return -ENOMEM; /* We'll try again later. */ 213802276f3cSIlpo Järvinen } else { 21399eb9362eSIlpo Järvinen int oldpcount = tcp_skb_pcount(skb); 21409eb9362eSIlpo Järvinen 21419eb9362eSIlpo Järvinen if (unlikely(oldpcount > 1)) { 214202276f3cSIlpo Järvinen tcp_init_tso_segs(sk, skb, cur_mss); 21439eb9362eSIlpo Järvinen tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb)); 21449eb9362eSIlpo Järvinen } 21451da177e4SLinus Torvalds } 21461da177e4SLinus Torvalds 21471da177e4SLinus Torvalds tcp_retrans_try_collapse(sk, skb, cur_mss); 21481da177e4SLinus Torvalds 21491da177e4SLinus Torvalds /* Some Solaris stacks overoptimize and ignore the FIN on a 21501da177e4SLinus Torvalds * retransmit when old data is attached. So strip it off 21511da177e4SLinus Torvalds * since it is cheap to do so and saves bytes on the network. 21521da177e4SLinus Torvalds */ 21531da177e4SLinus Torvalds if (skb->len > 0 && 21544de075e0SEric Dumazet (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && 21551da177e4SLinus Torvalds tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) { 21561da177e4SLinus Torvalds if (!pskb_trim(skb, 0)) { 2157e870a8efSIlpo Järvinen /* Reuse, even though it does some unnecessary work */ 2158e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1, 21594de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags); 21601da177e4SLinus Torvalds skb->ip_summed = CHECKSUM_NONE; 21611da177e4SLinus Torvalds } 21621da177e4SLinus Torvalds } 21631da177e4SLinus Torvalds 21641da177e4SLinus Torvalds /* Make a copy, if the first transmission SKB clone we made 21651da177e4SLinus Torvalds * is still in somebody's hands, else make a clone. 21661da177e4SLinus Torvalds */ 21671da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 21681da177e4SLinus Torvalds 2169117632e6SEric Dumazet /* make sure skb->data is aligned on arches that require it */ 2170117632e6SEric Dumazet if (unlikely(NET_IP_ALIGN && ((unsigned long)skb->data & 3))) { 2171117632e6SEric Dumazet struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER, 2172117632e6SEric Dumazet GFP_ATOMIC); 2173117632e6SEric Dumazet err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) : 2174117632e6SEric Dumazet -ENOBUFS; 2175117632e6SEric Dumazet } else { 2176dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 2177117632e6SEric Dumazet } 21781da177e4SLinus Torvalds 21791da177e4SLinus Torvalds if (err == 0) { 21801da177e4SLinus Torvalds /* Update global TCP statistics. */ 218181cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); 21821da177e4SLinus Torvalds 21831da177e4SLinus Torvalds tp->total_retrans++; 21841da177e4SLinus Torvalds 21851da177e4SLinus Torvalds #if FASTRETRANS_DEBUG > 0 21861da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 2187e87cc472SJoe Perches net_dbg_ratelimited("retrans_out leaked\n"); 21881da177e4SLinus Torvalds } 21891da177e4SLinus Torvalds #endif 2190b08d6cb2SIlpo Järvinen if (!tp->retrans_out) 2191b08d6cb2SIlpo Järvinen tp->lost_retrans_low = tp->snd_nxt; 21921da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 21931da177e4SLinus Torvalds tp->retrans_out += tcp_skb_pcount(skb); 21941da177e4SLinus Torvalds 21951da177e4SLinus Torvalds /* Save stamp of the first retransmit. */ 21961da177e4SLinus Torvalds if (!tp->retrans_stamp) 21971da177e4SLinus Torvalds tp->retrans_stamp = TCP_SKB_CB(skb)->when; 21981da177e4SLinus Torvalds 2199c24f691bSYuchung Cheng tp->undo_retrans += tcp_skb_pcount(skb); 22001da177e4SLinus Torvalds 22011da177e4SLinus Torvalds /* snd_nxt is stored to detect loss of retransmitted segment, 22021da177e4SLinus Torvalds * see tcp_input.c tcp_sacktag_write_queue(). 22031da177e4SLinus Torvalds */ 22041da177e4SLinus Torvalds TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt; 22051da177e4SLinus Torvalds } 22061da177e4SLinus Torvalds return err; 22071da177e4SLinus Torvalds } 22081da177e4SLinus Torvalds 220967edfef7SAndi Kleen /* Check if we forward retransmits are possible in the current 221067edfef7SAndi Kleen * window/congestion state. 221167edfef7SAndi Kleen */ 2212a2a385d6SEric Dumazet static bool tcp_can_forward_retransmit(struct sock *sk) 2213b5afe7bcSIlpo Järvinen { 2214b5afe7bcSIlpo Järvinen const struct inet_connection_sock *icsk = inet_csk(sk); 2215cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 2216b5afe7bcSIlpo Järvinen 2217b5afe7bcSIlpo Järvinen /* Forward retransmissions are possible only during Recovery. */ 2218b5afe7bcSIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Recovery) 2219a2a385d6SEric Dumazet return false; 2220b5afe7bcSIlpo Järvinen 2221b5afe7bcSIlpo Järvinen /* No forward retransmissions in Reno are possible. */ 2222b5afe7bcSIlpo Järvinen if (tcp_is_reno(tp)) 2223a2a385d6SEric Dumazet return false; 2224b5afe7bcSIlpo Järvinen 2225b5afe7bcSIlpo Järvinen /* Yeah, we have to make difficult choice between forward transmission 2226b5afe7bcSIlpo Järvinen * and retransmission... Both ways have their merits... 2227b5afe7bcSIlpo Järvinen * 2228b5afe7bcSIlpo Järvinen * For now we do not retransmit anything, while we have some new 2229b5afe7bcSIlpo Järvinen * segments to send. In the other cases, follow rule 3 for 2230b5afe7bcSIlpo Järvinen * NextSeg() specified in RFC3517. 2231b5afe7bcSIlpo Järvinen */ 2232b5afe7bcSIlpo Järvinen 2233b5afe7bcSIlpo Järvinen if (tcp_may_send_now(sk)) 2234a2a385d6SEric Dumazet return false; 2235b5afe7bcSIlpo Järvinen 2236a2a385d6SEric Dumazet return true; 2237b5afe7bcSIlpo Järvinen } 2238b5afe7bcSIlpo Järvinen 22391da177e4SLinus Torvalds /* This gets called after a retransmit timeout, and the initially 22401da177e4SLinus Torvalds * retransmitted data is acknowledged. It tries to continue 22411da177e4SLinus Torvalds * resending the rest of the retransmit queue, until either 22421da177e4SLinus Torvalds * we've sent it all or the congestion window limit is reached. 22431da177e4SLinus Torvalds * If doing SACK, the first ACK which comes back for a timeout 22441da177e4SLinus Torvalds * based retransmit packet might feed us FACK information again. 22451da177e4SLinus Torvalds * If so, we use it to avoid unnecessarily retransmissions. 22461da177e4SLinus Torvalds */ 22471da177e4SLinus Torvalds void tcp_xmit_retransmit_queue(struct sock *sk) 22481da177e4SLinus Torvalds { 22496687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 22501da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 22511da177e4SLinus Torvalds struct sk_buff *skb; 22520e1c54c2SIlpo Järvinen struct sk_buff *hole = NULL; 2253618d9f25SIlpo Järvinen u32 last_lost; 225461eb55f4SIlpo Järvinen int mib_idx; 22550e1c54c2SIlpo Järvinen int fwd_rexmitting = 0; 22566a438bbeSStephen Hemminger 225745e77d31SIlpo Järvinen if (!tp->packets_out) 225845e77d31SIlpo Järvinen return; 225945e77d31SIlpo Järvinen 226008ebd172SIlpo Järvinen if (!tp->lost_out) 226108ebd172SIlpo Järvinen tp->retransmit_high = tp->snd_una; 226208ebd172SIlpo Järvinen 2263618d9f25SIlpo Järvinen if (tp->retransmit_skb_hint) { 22646a438bbeSStephen Hemminger skb = tp->retransmit_skb_hint; 2265618d9f25SIlpo Järvinen last_lost = TCP_SKB_CB(skb)->end_seq; 2266618d9f25SIlpo Järvinen if (after(last_lost, tp->retransmit_high)) 2267618d9f25SIlpo Järvinen last_lost = tp->retransmit_high; 2268618d9f25SIlpo Järvinen } else { 2269fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 2270618d9f25SIlpo Järvinen last_lost = tp->snd_una; 2271618d9f25SIlpo Järvinen } 22721da177e4SLinus Torvalds 2273fe067e8aSDavid S. Miller tcp_for_write_queue_from(skb, sk) { 22741da177e4SLinus Torvalds __u8 sacked = TCP_SKB_CB(skb)->sacked; 22751da177e4SLinus Torvalds 2276fe067e8aSDavid S. Miller if (skb == tcp_send_head(sk)) 2277fe067e8aSDavid S. Miller break; 22786a438bbeSStephen Hemminger /* we could do better than to assign each time */ 22790e1c54c2SIlpo Järvinen if (hole == NULL) 22806a438bbeSStephen Hemminger tp->retransmit_skb_hint = skb; 22816a438bbeSStephen Hemminger 22821da177e4SLinus Torvalds /* Assume this retransmit will generate 22831da177e4SLinus Torvalds * only one packet for congestion window 22841da177e4SLinus Torvalds * calculation purposes. This works because 22851da177e4SLinus Torvalds * tcp_retransmit_skb() will chop up the 22861da177e4SLinus Torvalds * packet to be MSS sized and all the 22871da177e4SLinus Torvalds * packet counting works out. 22881da177e4SLinus Torvalds */ 22891da177e4SLinus Torvalds if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) 22901da177e4SLinus Torvalds return; 22910e1c54c2SIlpo Järvinen 22920e1c54c2SIlpo Järvinen if (fwd_rexmitting) { 22930e1c54c2SIlpo Järvinen begin_fwd: 22940e1c54c2SIlpo Järvinen if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp))) 2295006f582cSIlpo Järvinen break; 22960e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFORWARDRETRANS; 22970e1c54c2SIlpo Järvinen 22980e1c54c2SIlpo Järvinen } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) { 2299618d9f25SIlpo Järvinen tp->retransmit_high = last_lost; 23000e1c54c2SIlpo Järvinen if (!tcp_can_forward_retransmit(sk)) 23010e1c54c2SIlpo Järvinen break; 23020e1c54c2SIlpo Järvinen /* Backtrack if necessary to non-L'ed skb */ 23030e1c54c2SIlpo Järvinen if (hole != NULL) { 23040e1c54c2SIlpo Järvinen skb = hole; 23050e1c54c2SIlpo Järvinen hole = NULL; 23060e1c54c2SIlpo Järvinen } 23070e1c54c2SIlpo Järvinen fwd_rexmitting = 1; 23080e1c54c2SIlpo Järvinen goto begin_fwd; 23090e1c54c2SIlpo Järvinen 23100e1c54c2SIlpo Järvinen } else if (!(sacked & TCPCB_LOST)) { 2311ac11ba75SIlpo Järvinen if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) 23120e1c54c2SIlpo Järvinen hole = skb; 231361eb55f4SIlpo Järvinen continue; 23141da177e4SLinus Torvalds 23150e1c54c2SIlpo Järvinen } else { 2316618d9f25SIlpo Järvinen last_lost = TCP_SKB_CB(skb)->end_seq; 23170e1c54c2SIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Loss) 23180e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFASTRETRANS; 23190e1c54c2SIlpo Järvinen else 23200e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; 23210e1c54c2SIlpo Järvinen } 23220e1c54c2SIlpo Järvinen 23230e1c54c2SIlpo Järvinen if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) 232461eb55f4SIlpo Järvinen continue; 232540b215e5SPavel Emelyanov 232609e9b813SEric Dumazet if (tcp_retransmit_skb(sk, skb)) { 232709e9b813SEric Dumazet NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL); 23281da177e4SLinus Torvalds return; 232909e9b813SEric Dumazet } 2330de0744afSPavel Emelyanov NET_INC_STATS_BH(sock_net(sk), mib_idx); 23311da177e4SLinus Torvalds 2332a262f0cdSNandita Dukkipati if (inet_csk(sk)->icsk_ca_state == TCP_CA_Recovery) 2333a262f0cdSNandita Dukkipati tp->prr_out += tcp_skb_pcount(skb); 2334a262f0cdSNandita Dukkipati 2335fe067e8aSDavid S. Miller if (skb == tcp_write_queue_head(sk)) 2336463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 23373f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, 23383f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 23391da177e4SLinus Torvalds } 23401da177e4SLinus Torvalds } 23411da177e4SLinus Torvalds 23421da177e4SLinus Torvalds /* Send a fin. The caller locks the socket for us. This cannot be 23431da177e4SLinus Torvalds * allowed to fail queueing a FIN frame under any circumstances. 23441da177e4SLinus Torvalds */ 23451da177e4SLinus Torvalds void tcp_send_fin(struct sock *sk) 23461da177e4SLinus Torvalds { 23471da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2348fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_write_queue_tail(sk); 23491da177e4SLinus Torvalds int mss_now; 23501da177e4SLinus Torvalds 23511da177e4SLinus Torvalds /* Optimization, tack on the FIN if we have a queue of 23521da177e4SLinus Torvalds * unsent frames. But be careful about outgoing SACKS 23531da177e4SLinus Torvalds * and IP options. 23541da177e4SLinus Torvalds */ 23550c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 23561da177e4SLinus Torvalds 2357fe067e8aSDavid S. Miller if (tcp_send_head(sk) != NULL) { 23584de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_FIN; 23591da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq++; 23601da177e4SLinus Torvalds tp->write_seq++; 23611da177e4SLinus Torvalds } else { 23621da177e4SLinus Torvalds /* Socket is locked, keep trying until memory is available. */ 23631da177e4SLinus Torvalds for (;;) { 2364aa133076SWu Fengguang skb = alloc_skb_fclone(MAX_TCP_HEADER, 2365aa133076SWu Fengguang sk->sk_allocation); 23661da177e4SLinus Torvalds if (skb) 23671da177e4SLinus Torvalds break; 23681da177e4SLinus Torvalds yield(); 23691da177e4SLinus Torvalds } 23701da177e4SLinus Torvalds 23711da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 23721da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 23731da177e4SLinus Torvalds /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 2374e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tp->write_seq, 2375a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_FIN); 23761da177e4SLinus Torvalds tcp_queue_skb(sk, skb); 23771da177e4SLinus Torvalds } 23789e412ba7SIlpo Järvinen __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF); 23791da177e4SLinus Torvalds } 23801da177e4SLinus Torvalds 23811da177e4SLinus Torvalds /* We get here when a process closes a file descriptor (either due to 23821da177e4SLinus Torvalds * an explicit close() or as a byproduct of exit()'ing) and there 23831da177e4SLinus Torvalds * was unread data in the receive queue. This behavior is recommended 238465bb723cSGerrit Renker * by RFC 2525, section 2.17. -DaveM 23851da177e4SLinus Torvalds */ 2386dd0fc66fSAl Viro void tcp_send_active_reset(struct sock *sk, gfp_t priority) 23871da177e4SLinus Torvalds { 23881da177e4SLinus Torvalds struct sk_buff *skb; 23891da177e4SLinus Torvalds 23901da177e4SLinus Torvalds /* NOTE: No TCP options attached and we never retransmit this. */ 23911da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, priority); 23921da177e4SLinus Torvalds if (!skb) { 23934e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 23941da177e4SLinus Torvalds return; 23951da177e4SLinus Torvalds } 23961da177e4SLinus Torvalds 23971da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 23981da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 2399e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), 2400a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_RST); 24011da177e4SLinus Torvalds /* Send it off. */ 24021da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2403dfb4b9dcSDavid S. Miller if (tcp_transmit_skb(sk, skb, 0, priority)) 24044e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 240526af65cbSSridhar Samudrala 240681cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); 24071da177e4SLinus Torvalds } 24081da177e4SLinus Torvalds 240967edfef7SAndi Kleen /* Send a crossed SYN-ACK during socket establishment. 241067edfef7SAndi Kleen * WARNING: This routine must only be called when we have already sent 24111da177e4SLinus Torvalds * a SYN packet that crossed the incoming SYN that caused this routine 24121da177e4SLinus Torvalds * to get called. If this assumption fails then the initial rcv_wnd 24131da177e4SLinus Torvalds * and rcv_wscale values will not be correct. 24141da177e4SLinus Torvalds */ 24151da177e4SLinus Torvalds int tcp_send_synack(struct sock *sk) 24161da177e4SLinus Torvalds { 24171da177e4SLinus Torvalds struct sk_buff *skb; 24181da177e4SLinus Torvalds 2419fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 24204de075e0SEric Dumazet if (skb == NULL || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 242191df42beSJoe Perches pr_debug("%s: wrong queue state\n", __func__); 24221da177e4SLinus Torvalds return -EFAULT; 24231da177e4SLinus Torvalds } 24244de075e0SEric Dumazet if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { 24251da177e4SLinus Torvalds if (skb_cloned(skb)) { 24261da177e4SLinus Torvalds struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC); 24271da177e4SLinus Torvalds if (nskb == NULL) 24281da177e4SLinus Torvalds return -ENOMEM; 2429fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 24301da177e4SLinus Torvalds skb_header_release(nskb); 2431fe067e8aSDavid S. Miller __tcp_add_write_queue_head(sk, nskb); 24323ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 24333ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 24343ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 24351da177e4SLinus Torvalds skb = nskb; 24361da177e4SLinus Torvalds } 24371da177e4SLinus Torvalds 24384de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; 24391da177e4SLinus Torvalds TCP_ECN_send_synack(tcp_sk(sk), skb); 24401da177e4SLinus Torvalds } 24411da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2442dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 24431da177e4SLinus Torvalds } 24441da177e4SLinus Torvalds 2445*4aea39c1SEric Dumazet /** 2446*4aea39c1SEric Dumazet * tcp_make_synack - Prepare a SYN-ACK. 2447*4aea39c1SEric Dumazet * sk: listener socket 2448*4aea39c1SEric Dumazet * dst: dst entry attached to the SYNACK 2449*4aea39c1SEric Dumazet * req: request_sock pointer 2450*4aea39c1SEric Dumazet * rvp: request_values pointer 2451*4aea39c1SEric Dumazet * 2452*4aea39c1SEric Dumazet * Allocate one skb and build a SYNACK packet. 2453*4aea39c1SEric Dumazet * @dst is consumed : Caller should not use it again. 2454*4aea39c1SEric Dumazet */ 24551da177e4SLinus Torvalds struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst, 2456e6b4d113SWilliam Allen Simpson struct request_sock *req, 2457e6b4d113SWilliam Allen Simpson struct request_values *rvp) 24581da177e4SLinus Torvalds { 2459bd0388aeSWilliam Allen Simpson struct tcp_out_options opts; 24604957faadSWilliam Allen Simpson struct tcp_extend_values *xvp = tcp_xv(rvp); 24612e6599cbSArnaldo Carvalho de Melo struct inet_request_sock *ireq = inet_rsk(req); 24621da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 246328b2774aSEric Dumazet const struct tcp_cookie_values *cvp = tp->cookie_values; 24641da177e4SLinus Torvalds struct tcphdr *th; 24651da177e4SLinus Torvalds struct sk_buff *skb; 2466cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 2467bd0388aeSWilliam Allen Simpson int tcp_header_size; 2468f5fff5dcSTom Quetchenbach int mss; 246928b2774aSEric Dumazet int s_data_desired = 0; 24701da177e4SLinus Torvalds 247128b2774aSEric Dumazet if (cvp != NULL && cvp->s_data_constant && cvp->s_data_desired) 247228b2774aSEric Dumazet s_data_desired = cvp->s_data_desired; 247390ba9b19SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER + 15 + s_data_desired, GFP_ATOMIC); 2474*4aea39c1SEric Dumazet if (unlikely(!skb)) { 2475*4aea39c1SEric Dumazet dst_release(dst); 24761da177e4SLinus Torvalds return NULL; 2477*4aea39c1SEric Dumazet } 24781da177e4SLinus Torvalds /* Reserve space for headers. */ 24791da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 24801da177e4SLinus Torvalds 2481*4aea39c1SEric Dumazet skb_dst_set(skb, dst); 24821da177e4SLinus Torvalds 24830dbaee3bSDavid S. Miller mss = dst_metric_advmss(dst); 2484f5fff5dcSTom Quetchenbach if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss) 2485f5fff5dcSTom Quetchenbach mss = tp->rx_opt.user_mss; 2486f5fff5dcSTom Quetchenbach 248733ad798cSAdam Langley if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */ 248833ad798cSAdam Langley __u8 rcv_wscale; 248933ad798cSAdam Langley /* Set this up on the first call only */ 249033ad798cSAdam Langley req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW); 2491e88c64f0SHagen Paul Pfeifer 2492e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 2493e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 2494e88c64f0SHagen Paul Pfeifer (req->window_clamp > tcp_full_space(sk) || req->window_clamp == 0)) 2495e88c64f0SHagen Paul Pfeifer req->window_clamp = tcp_full_space(sk); 2496e88c64f0SHagen Paul Pfeifer 249733ad798cSAdam Langley /* tcp_full_space because it is guaranteed to be the first packet */ 249833ad798cSAdam Langley tcp_select_initial_window(tcp_full_space(sk), 2499f5fff5dcSTom Quetchenbach mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0), 250033ad798cSAdam Langley &req->rcv_wnd, 250133ad798cSAdam Langley &req->window_clamp, 250233ad798cSAdam Langley ireq->wscale_ok, 250331d12926Slaurent chavey &rcv_wscale, 250431d12926Slaurent chavey dst_metric(dst, RTAX_INITRWND)); 250533ad798cSAdam Langley ireq->rcv_wscale = rcv_wscale; 250633ad798cSAdam Langley } 2507cfb6eeb4SYOSHIFUJI Hideaki 250833ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 25098b5f12d0SFlorian Westphal #ifdef CONFIG_SYN_COOKIES 25108b5f12d0SFlorian Westphal if (unlikely(req->cookie_ts)) 25118b5f12d0SFlorian Westphal TCP_SKB_CB(skb)->when = cookie_init_timestamp(req); 25128b5f12d0SFlorian Westphal else 25138b5f12d0SFlorian Westphal #endif 251433ad798cSAdam Langley TCP_SKB_CB(skb)->when = tcp_time_stamp; 2515f5fff5dcSTom Quetchenbach tcp_header_size = tcp_synack_options(sk, req, mss, 25164957faadSWilliam Allen Simpson skb, &opts, &md5, xvp) 25174957faadSWilliam Allen Simpson + sizeof(*th); 251833ad798cSAdam Langley 2519aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 2520aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 25211da177e4SLinus Torvalds 2522aa8223c7SArnaldo Carvalho de Melo th = tcp_hdr(skb); 25231da177e4SLinus Torvalds memset(th, 0, sizeof(struct tcphdr)); 25241da177e4SLinus Torvalds th->syn = 1; 25251da177e4SLinus Torvalds th->ack = 1; 25261da177e4SLinus Torvalds TCP_ECN_make_synack(req, th); 2527a3116ac5SKOVACS Krisztian th->source = ireq->loc_port; 25282e6599cbSArnaldo Carvalho de Melo th->dest = ireq->rmt_port; 2529e870a8efSIlpo Järvinen /* Setting of flags are superfluous here for callers (and ECE is 2530e870a8efSIlpo Järvinen * not even correctly set) 2531e870a8efSIlpo Järvinen */ 2532e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn, 2533a3433f35SChangli Gao TCPHDR_SYN | TCPHDR_ACK); 25344957faadSWilliam Allen Simpson 25354957faadSWilliam Allen Simpson if (OPTION_COOKIE_EXTENSION & opts.options) { 253628b2774aSEric Dumazet if (s_data_desired) { 253728b2774aSEric Dumazet u8 *buf = skb_put(skb, s_data_desired); 25384957faadSWilliam Allen Simpson 25394957faadSWilliam Allen Simpson /* copy data directly from the listening socket. */ 254028b2774aSEric Dumazet memcpy(buf, cvp->s_data_payload, s_data_desired); 254128b2774aSEric Dumazet TCP_SKB_CB(skb)->end_seq += s_data_desired; 25424957faadSWilliam Allen Simpson } 25434957faadSWilliam Allen Simpson 25444957faadSWilliam Allen Simpson if (opts.hash_size > 0) { 25454957faadSWilliam Allen Simpson __u32 workspace[SHA_WORKSPACE_WORDS]; 25464957faadSWilliam Allen Simpson u32 *mess = &xvp->cookie_bakery[COOKIE_DIGEST_WORDS]; 25474957faadSWilliam Allen Simpson u32 *tail = &mess[COOKIE_MESSAGE_WORDS-1]; 25484957faadSWilliam Allen Simpson 25494957faadSWilliam Allen Simpson /* Secret recipe depends on the Timestamp, (future) 25504957faadSWilliam Allen Simpson * Sequence and Acknowledgment Numbers, Initiator 25514957faadSWilliam Allen Simpson * Cookie, and others handled by IP variant caller. 25524957faadSWilliam Allen Simpson */ 25534957faadSWilliam Allen Simpson *tail-- ^= opts.tsval; 25544957faadSWilliam Allen Simpson *tail-- ^= tcp_rsk(req)->rcv_isn + 1; 25554957faadSWilliam Allen Simpson *tail-- ^= TCP_SKB_CB(skb)->seq + 1; 25564957faadSWilliam Allen Simpson 25574957faadSWilliam Allen Simpson /* recommended */ 25580eae88f3SEric Dumazet *tail-- ^= (((__force u32)th->dest << 16) | (__force u32)th->source); 2559f9a2e69eSDavid S. Miller *tail-- ^= (u32)(unsigned long)cvp; /* per sockopt */ 25604957faadSWilliam Allen Simpson 25614957faadSWilliam Allen Simpson sha_transform((__u32 *)&xvp->cookie_bakery[0], 25624957faadSWilliam Allen Simpson (char *)mess, 25634957faadSWilliam Allen Simpson &workspace[0]); 25644957faadSWilliam Allen Simpson opts.hash_location = 25654957faadSWilliam Allen Simpson (__u8 *)&xvp->cookie_bakery[0]; 25664957faadSWilliam Allen Simpson } 25674957faadSWilliam Allen Simpson } 25684957faadSWilliam Allen Simpson 25691da177e4SLinus Torvalds th->seq = htonl(TCP_SKB_CB(skb)->seq); 25702e6599cbSArnaldo Carvalho de Melo th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1); 25711da177e4SLinus Torvalds 25721da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 2573600ff0c2SIlpo Järvinen th->window = htons(min(req->rcv_wnd, 65535U)); 2574bd0388aeSWilliam Allen Simpson tcp_options_write((__be32 *)(th + 1), tp, &opts); 25751da177e4SLinus Torvalds th->doff = (tcp_header_size >> 2); 2576aa2ea058STom Herbert TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, tcp_skb_pcount(skb)); 2577cfb6eeb4SYOSHIFUJI Hideaki 2578cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 2579cfb6eeb4SYOSHIFUJI Hideaki /* Okay, we have all we need - do the md5 hash if needed */ 2580cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 2581bd0388aeSWilliam Allen Simpson tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, 258249a72dfbSAdam Langley md5, NULL, req, skb); 2583cfb6eeb4SYOSHIFUJI Hideaki } 2584cfb6eeb4SYOSHIFUJI Hideaki #endif 2585cfb6eeb4SYOSHIFUJI Hideaki 25861da177e4SLinus Torvalds return skb; 25871da177e4SLinus Torvalds } 25884bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_make_synack); 25891da177e4SLinus Torvalds 259067edfef7SAndi Kleen /* Do all connect socket setups that can be done AF independent. */ 2591370816aeSPavel Emelyanov void tcp_connect_init(struct sock *sk) 25921da177e4SLinus Torvalds { 2593cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 25941da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 25951da177e4SLinus Torvalds __u8 rcv_wscale; 25961da177e4SLinus Torvalds 25971da177e4SLinus Torvalds /* We'll fix this up when we get a response from the other end. 25981da177e4SLinus Torvalds * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 25991da177e4SLinus Torvalds */ 26001da177e4SLinus Torvalds tp->tcp_header_len = sizeof(struct tcphdr) + 2601bb5b7c11SDavid S. Miller (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0); 26021da177e4SLinus Torvalds 2603cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 2604cfb6eeb4SYOSHIFUJI Hideaki if (tp->af_specific->md5_lookup(sk, sk) != NULL) 2605cfb6eeb4SYOSHIFUJI Hideaki tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 2606cfb6eeb4SYOSHIFUJI Hideaki #endif 2607cfb6eeb4SYOSHIFUJI Hideaki 26081da177e4SLinus Torvalds /* If user gave his TCP_MAXSEG, record it to clamp */ 26091da177e4SLinus Torvalds if (tp->rx_opt.user_mss) 26101da177e4SLinus Torvalds tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 26111da177e4SLinus Torvalds tp->max_window = 0; 26125d424d5aSJohn Heffner tcp_mtup_init(sk); 26131da177e4SLinus Torvalds tcp_sync_mss(sk, dst_mtu(dst)); 26141da177e4SLinus Torvalds 26151da177e4SLinus Torvalds if (!tp->window_clamp) 26161da177e4SLinus Torvalds tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 26170dbaee3bSDavid S. Miller tp->advmss = dst_metric_advmss(dst); 2618f5fff5dcSTom Quetchenbach if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss) 2619f5fff5dcSTom Quetchenbach tp->advmss = tp->rx_opt.user_mss; 2620f5fff5dcSTom Quetchenbach 26211da177e4SLinus Torvalds tcp_initialize_rcv_mss(sk); 26221da177e4SLinus Torvalds 2623e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 2624e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 2625e88c64f0SHagen Paul Pfeifer (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) 2626e88c64f0SHagen Paul Pfeifer tp->window_clamp = tcp_full_space(sk); 2627e88c64f0SHagen Paul Pfeifer 26281da177e4SLinus Torvalds tcp_select_initial_window(tcp_full_space(sk), 26291da177e4SLinus Torvalds tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 26301da177e4SLinus Torvalds &tp->rcv_wnd, 26311da177e4SLinus Torvalds &tp->window_clamp, 2632bb5b7c11SDavid S. Miller sysctl_tcp_window_scaling, 263331d12926Slaurent chavey &rcv_wscale, 263431d12926Slaurent chavey dst_metric(dst, RTAX_INITRWND)); 26351da177e4SLinus Torvalds 26361da177e4SLinus Torvalds tp->rx_opt.rcv_wscale = rcv_wscale; 26371da177e4SLinus Torvalds tp->rcv_ssthresh = tp->rcv_wnd; 26381da177e4SLinus Torvalds 26391da177e4SLinus Torvalds sk->sk_err = 0; 26401da177e4SLinus Torvalds sock_reset_flag(sk, SOCK_DONE); 26411da177e4SLinus Torvalds tp->snd_wnd = 0; 2642ee7537b6SHantzis Fotis tcp_init_wl(tp, 0); 26431da177e4SLinus Torvalds tp->snd_una = tp->write_seq; 26441da177e4SLinus Torvalds tp->snd_sml = tp->write_seq; 264533f5f57eSIlpo Järvinen tp->snd_up = tp->write_seq; 2646370816aeSPavel Emelyanov tp->snd_nxt = tp->write_seq; 2647ee995283SPavel Emelyanov 2648ee995283SPavel Emelyanov if (likely(!tp->repair)) 26491da177e4SLinus Torvalds tp->rcv_nxt = 0; 2650ee995283SPavel Emelyanov tp->rcv_wup = tp->rcv_nxt; 2651ee995283SPavel Emelyanov tp->copied_seq = tp->rcv_nxt; 26521da177e4SLinus Torvalds 2653463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; 2654463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_retransmits = 0; 26551da177e4SLinus Torvalds tcp_clear_retrans(tp); 26561da177e4SLinus Torvalds } 26571da177e4SLinus Torvalds 265867edfef7SAndi Kleen /* Build a SYN and send it off. */ 26591da177e4SLinus Torvalds int tcp_connect(struct sock *sk) 26601da177e4SLinus Torvalds { 26611da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 26621da177e4SLinus Torvalds struct sk_buff *buff; 2663ee586811SEric Paris int err; 26641da177e4SLinus Torvalds 26651da177e4SLinus Torvalds tcp_connect_init(sk); 26661da177e4SLinus Torvalds 2667d179cd12SDavid S. Miller buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation); 26681da177e4SLinus Torvalds if (unlikely(buff == NULL)) 26691da177e4SLinus Torvalds return -ENOBUFS; 26701da177e4SLinus Torvalds 26711da177e4SLinus Torvalds /* Reserve space for headers. */ 26721da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 26731da177e4SLinus Torvalds 2674a3433f35SChangli Gao tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); 2675e870a8efSIlpo Järvinen TCP_ECN_send_syn(sk, buff); 26761da177e4SLinus Torvalds 26771da177e4SLinus Torvalds /* Send it off. */ 26781da177e4SLinus Torvalds TCP_SKB_CB(buff)->when = tcp_time_stamp; 26791da177e4SLinus Torvalds tp->retrans_stamp = TCP_SKB_CB(buff)->when; 26801da177e4SLinus Torvalds skb_header_release(buff); 2681fe067e8aSDavid S. Miller __tcp_add_write_queue_tail(sk, buff); 26823ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 26833ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 26841da177e4SLinus Torvalds tp->packets_out += tcp_skb_pcount(buff); 2685ee586811SEric Paris err = tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); 2686ee586811SEric Paris if (err == -ECONNREFUSED) 2687ee586811SEric Paris return err; 2688bd37a088SWei Yongjun 2689bd37a088SWei Yongjun /* We change tp->snd_nxt after the tcp_transmit_skb() call 2690bd37a088SWei Yongjun * in order to make this packet get counted in tcpOutSegs. 2691bd37a088SWei Yongjun */ 2692bd37a088SWei Yongjun tp->snd_nxt = tp->write_seq; 2693bd37a088SWei Yongjun tp->pushed_seq = tp->write_seq; 269481cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); 26951da177e4SLinus Torvalds 26961da177e4SLinus Torvalds /* Timer for repeating the SYN until an answer. */ 26973f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 26983f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 26991da177e4SLinus Torvalds return 0; 27001da177e4SLinus Torvalds } 27014bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_connect); 27021da177e4SLinus Torvalds 27031da177e4SLinus Torvalds /* Send out a delayed ack, the caller does the policy checking 27041da177e4SLinus Torvalds * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 27051da177e4SLinus Torvalds * for details. 27061da177e4SLinus Torvalds */ 27071da177e4SLinus Torvalds void tcp_send_delayed_ack(struct sock *sk) 27081da177e4SLinus Torvalds { 2709463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 2710463c84b9SArnaldo Carvalho de Melo int ato = icsk->icsk_ack.ato; 27111da177e4SLinus Torvalds unsigned long timeout; 27121da177e4SLinus Torvalds 27131da177e4SLinus Torvalds if (ato > TCP_DELACK_MIN) { 2714463c84b9SArnaldo Carvalho de Melo const struct tcp_sock *tp = tcp_sk(sk); 27151da177e4SLinus Torvalds int max_ato = HZ / 2; 27161da177e4SLinus Torvalds 2717056834d9SIlpo Järvinen if (icsk->icsk_ack.pingpong || 2718056834d9SIlpo Järvinen (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 27191da177e4SLinus Torvalds max_ato = TCP_DELACK_MAX; 27201da177e4SLinus Torvalds 27211da177e4SLinus Torvalds /* Slow path, intersegment interval is "high". */ 27221da177e4SLinus Torvalds 27231da177e4SLinus Torvalds /* If some rtt estimate is known, use it to bound delayed ack. 2724463c84b9SArnaldo Carvalho de Melo * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 27251da177e4SLinus Torvalds * directly. 27261da177e4SLinus Torvalds */ 27271da177e4SLinus Torvalds if (tp->srtt) { 27281da177e4SLinus Torvalds int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN); 27291da177e4SLinus Torvalds 27301da177e4SLinus Torvalds if (rtt < max_ato) 27311da177e4SLinus Torvalds max_ato = rtt; 27321da177e4SLinus Torvalds } 27331da177e4SLinus Torvalds 27341da177e4SLinus Torvalds ato = min(ato, max_ato); 27351da177e4SLinus Torvalds } 27361da177e4SLinus Torvalds 27371da177e4SLinus Torvalds /* Stay within the limit we were given */ 27381da177e4SLinus Torvalds timeout = jiffies + ato; 27391da177e4SLinus Torvalds 27401da177e4SLinus Torvalds /* Use new timeout only if there wasn't a older one earlier. */ 2741463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 27421da177e4SLinus Torvalds /* If delack timer was blocked or is about to expire, 27431da177e4SLinus Torvalds * send ACK now. 27441da177e4SLinus Torvalds */ 2745463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.blocked || 2746463c84b9SArnaldo Carvalho de Melo time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 27471da177e4SLinus Torvalds tcp_send_ack(sk); 27481da177e4SLinus Torvalds return; 27491da177e4SLinus Torvalds } 27501da177e4SLinus Torvalds 2751463c84b9SArnaldo Carvalho de Melo if (!time_before(timeout, icsk->icsk_ack.timeout)) 2752463c84b9SArnaldo Carvalho de Melo timeout = icsk->icsk_ack.timeout; 27531da177e4SLinus Torvalds } 2754463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 2755463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.timeout = timeout; 2756463c84b9SArnaldo Carvalho de Melo sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 27571da177e4SLinus Torvalds } 27581da177e4SLinus Torvalds 27591da177e4SLinus Torvalds /* This routine sends an ack and also updates the window. */ 27601da177e4SLinus Torvalds void tcp_send_ack(struct sock *sk) 27611da177e4SLinus Torvalds { 27621da177e4SLinus Torvalds struct sk_buff *buff; 27631da177e4SLinus Torvalds 2764058dc334SIlpo Järvinen /* If we have been reset, we may not send again. */ 2765058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 2766058dc334SIlpo Järvinen return; 2767058dc334SIlpo Järvinen 27681da177e4SLinus Torvalds /* We are not putting this on the write queue, so 27691da177e4SLinus Torvalds * tcp_transmit_skb() will set the ownership to this 27701da177e4SLinus Torvalds * sock. 27711da177e4SLinus Torvalds */ 27721da177e4SLinus Torvalds buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 27731da177e4SLinus Torvalds if (buff == NULL) { 2774463c84b9SArnaldo Carvalho de Melo inet_csk_schedule_ack(sk); 2775463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; 27763f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 27773f421baaSArnaldo Carvalho de Melo TCP_DELACK_MAX, TCP_RTO_MAX); 27781da177e4SLinus Torvalds return; 27791da177e4SLinus Torvalds } 27801da177e4SLinus Torvalds 27811da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 27821da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 2783a3433f35SChangli Gao tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); 27841da177e4SLinus Torvalds 27851da177e4SLinus Torvalds /* Send it off, this clears delayed acks for us. */ 27861da177e4SLinus Torvalds TCP_SKB_CB(buff)->when = tcp_time_stamp; 2787dfb4b9dcSDavid S. Miller tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC); 27881da177e4SLinus Torvalds } 27891da177e4SLinus Torvalds 27901da177e4SLinus Torvalds /* This routine sends a packet with an out of date sequence 27911da177e4SLinus Torvalds * number. It assumes the other end will try to ack it. 27921da177e4SLinus Torvalds * 27931da177e4SLinus Torvalds * Question: what should we make while urgent mode? 27941da177e4SLinus Torvalds * 4.4BSD forces sending single byte of data. We cannot send 27951da177e4SLinus Torvalds * out of window data, because we have SND.NXT==SND.MAX... 27961da177e4SLinus Torvalds * 27971da177e4SLinus Torvalds * Current solution: to send TWO zero-length segments in urgent mode: 27981da177e4SLinus Torvalds * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 27991da177e4SLinus Torvalds * out-of-date with SND.UNA-1 to probe window. 28001da177e4SLinus Torvalds */ 28011da177e4SLinus Torvalds static int tcp_xmit_probe_skb(struct sock *sk, int urgent) 28021da177e4SLinus Torvalds { 28031da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 28041da177e4SLinus Torvalds struct sk_buff *skb; 28051da177e4SLinus Torvalds 28061da177e4SLinus Torvalds /* We don't queue it, tcp_transmit_skb() sets ownership. */ 28071da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 28081da177e4SLinus Torvalds if (skb == NULL) 28091da177e4SLinus Torvalds return -1; 28101da177e4SLinus Torvalds 28111da177e4SLinus Torvalds /* Reserve space for headers and set control bits. */ 28121da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 28131da177e4SLinus Torvalds /* Use a previous sequence. This should cause the other 28141da177e4SLinus Torvalds * end to send an ack. Don't queue or clone SKB, just 28151da177e4SLinus Torvalds * send it. 28161da177e4SLinus Torvalds */ 2817a3433f35SChangli Gao tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); 28181da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2819dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC); 28201da177e4SLinus Torvalds } 28211da177e4SLinus Torvalds 2822ee995283SPavel Emelyanov void tcp_send_window_probe(struct sock *sk) 2823ee995283SPavel Emelyanov { 2824ee995283SPavel Emelyanov if (sk->sk_state == TCP_ESTABLISHED) { 2825ee995283SPavel Emelyanov tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; 2826c0e88ff0SPavel Emelyanov tcp_sk(sk)->snd_nxt = tcp_sk(sk)->write_seq; 2827ee995283SPavel Emelyanov tcp_xmit_probe_skb(sk, 0); 2828ee995283SPavel Emelyanov } 2829ee995283SPavel Emelyanov } 2830ee995283SPavel Emelyanov 283167edfef7SAndi Kleen /* Initiate keepalive or window probe from timer. */ 28321da177e4SLinus Torvalds int tcp_write_wakeup(struct sock *sk) 28331da177e4SLinus Torvalds { 28341da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 28351da177e4SLinus Torvalds struct sk_buff *skb; 28361da177e4SLinus Torvalds 2837058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 2838058dc334SIlpo Järvinen return -1; 2839058dc334SIlpo Järvinen 2840fe067e8aSDavid S. Miller if ((skb = tcp_send_head(sk)) != NULL && 284190840defSIlpo Järvinen before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { 28421da177e4SLinus Torvalds int err; 28430c54b85fSIlpo Järvinen unsigned int mss = tcp_current_mss(sk); 284490840defSIlpo Järvinen unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 28451da177e4SLinus Torvalds 28461da177e4SLinus Torvalds if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 28471da177e4SLinus Torvalds tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 28481da177e4SLinus Torvalds 28491da177e4SLinus Torvalds /* We are probing the opening of a window 28501da177e4SLinus Torvalds * but the window size is != 0 28511da177e4SLinus Torvalds * must have been a result SWS avoidance ( sender ) 28521da177e4SLinus Torvalds */ 28531da177e4SLinus Torvalds if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 28541da177e4SLinus Torvalds skb->len > mss) { 28551da177e4SLinus Torvalds seg_size = min(seg_size, mss); 28564de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 2857846998aeSDavid S. Miller if (tcp_fragment(sk, skb, seg_size, mss)) 28581da177e4SLinus Torvalds return -1; 28591da177e4SLinus Torvalds } else if (!tcp_skb_pcount(skb)) 2860846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss); 28611da177e4SLinus Torvalds 28624de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 28631da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2864dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 286566f5fe62SIlpo Järvinen if (!err) 286666f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 28671da177e4SLinus Torvalds return err; 28681da177e4SLinus Torvalds } else { 286933f5f57eSIlpo Järvinen if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) 28704828e7f4SIlpo Järvinen tcp_xmit_probe_skb(sk, 1); 28711da177e4SLinus Torvalds return tcp_xmit_probe_skb(sk, 0); 28721da177e4SLinus Torvalds } 28731da177e4SLinus Torvalds } 28741da177e4SLinus Torvalds 28751da177e4SLinus Torvalds /* A window probe timeout has occurred. If window is not closed send 28761da177e4SLinus Torvalds * a partial packet else a zero probe. 28771da177e4SLinus Torvalds */ 28781da177e4SLinus Torvalds void tcp_send_probe0(struct sock *sk) 28791da177e4SLinus Torvalds { 2880463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 28811da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 28821da177e4SLinus Torvalds int err; 28831da177e4SLinus Torvalds 28841da177e4SLinus Torvalds err = tcp_write_wakeup(sk); 28851da177e4SLinus Torvalds 2886fe067e8aSDavid S. Miller if (tp->packets_out || !tcp_send_head(sk)) { 28871da177e4SLinus Torvalds /* Cancel probe timer, if it is not required. */ 28886687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 0; 2889463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff = 0; 28901da177e4SLinus Torvalds return; 28911da177e4SLinus Torvalds } 28921da177e4SLinus Torvalds 28931da177e4SLinus Torvalds if (err <= 0) { 2894463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_backoff < sysctl_tcp_retries2) 2895463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff++; 28966687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out++; 2897463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 28983f421baaSArnaldo Carvalho de Melo min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), 28993f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 29001da177e4SLinus Torvalds } else { 29011da177e4SLinus Torvalds /* If packet was not sent due to local congestion, 29026687e988SArnaldo Carvalho de Melo * do not backoff and do not remember icsk_probes_out. 29031da177e4SLinus Torvalds * Let local senders to fight for local resources. 29041da177e4SLinus Torvalds * 29051da177e4SLinus Torvalds * Use accumulated backoff yet. 29061da177e4SLinus Torvalds */ 29076687e988SArnaldo Carvalho de Melo if (!icsk->icsk_probes_out) 29086687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 1; 2909463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 2910463c84b9SArnaldo Carvalho de Melo min(icsk->icsk_rto << icsk->icsk_backoff, 29113f421baaSArnaldo Carvalho de Melo TCP_RESOURCE_PROBE_INTERVAL), 29123f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 29131da177e4SLinus Torvalds } 29141da177e4SLinus Torvalds } 2915