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 5346d3ceabSEric Dumazet /* Default TSQ limit of two TSO segments */ 5446d3ceabSEric Dumazet int sysctl_tcp_limit_output_bytes __read_mostly = 131072; 5546d3ceabSEric Dumazet 561da177e4SLinus Torvalds /* This limits the percentage of the congestion window which we 571da177e4SLinus Torvalds * will allow a single TSO frame to consume. Building TSO frames 581da177e4SLinus Torvalds * which are too large can cause TCP streams to be bursty. 591da177e4SLinus Torvalds */ 60ab32ea5dSBrian Haley int sysctl_tcp_tso_win_divisor __read_mostly = 3; 611da177e4SLinus Torvalds 62ab32ea5dSBrian Haley int sysctl_tcp_mtu_probing __read_mostly = 0; 6397b1ce25SShan Wei int sysctl_tcp_base_mss __read_mostly = TCP_BASE_MSS; 645d424d5aSJohn Heffner 6535089bb2SDavid S. Miller /* By default, RFC2861 behavior. */ 66ab32ea5dSBrian Haley int sysctl_tcp_slow_start_after_idle __read_mostly = 1; 6735089bb2SDavid S. Miller 6846d3ceabSEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 6946d3ceabSEric Dumazet int push_one, gfp_t gfp); 70519855c5SWilliam Allen Simpson 7167edfef7SAndi Kleen /* Account for new data that has been sent to the network. */ 72cf533ea5SEric Dumazet static void tcp_event_new_data_sent(struct sock *sk, const struct sk_buff *skb) 736ff03ac3SIlpo Järvinen { 746ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 756ff03ac3SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 7666f5fe62SIlpo Järvinen unsigned int prior_packets = tp->packets_out; 779e412ba7SIlpo Järvinen 78fe067e8aSDavid S. Miller tcp_advance_send_head(sk, skb); 791da177e4SLinus Torvalds tp->snd_nxt = TCP_SKB_CB(skb)->end_seq; 808512430eSIlpo Järvinen 8166f5fe62SIlpo Järvinen tp->packets_out += tcp_skb_pcount(skb); 826ba8a3b1SNandita Dukkipati if (!prior_packets || icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS || 83*6a5dc9e5SEric Dumazet icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) { 84750ea2baSYuchung Cheng tcp_rearm_rto(sk); 851da177e4SLinus Torvalds } 86*6a5dc9e5SEric Dumazet } 871da177e4SLinus Torvalds 881da177e4SLinus Torvalds /* SND.NXT, if window was not shrunk. 891da177e4SLinus Torvalds * If window has been shrunk, what should we make? It is not clear at all. 901da177e4SLinus Torvalds * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-( 911da177e4SLinus Torvalds * Anything in between SND.UNA...SND.UNA+SND.WND also can be already 921da177e4SLinus Torvalds * invalid. OK, let's make this for now: 931da177e4SLinus Torvalds */ 94cf533ea5SEric Dumazet static inline __u32 tcp_acceptable_seq(const struct sock *sk) 951da177e4SLinus Torvalds { 96cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 979e412ba7SIlpo Järvinen 9890840defSIlpo Järvinen if (!before(tcp_wnd_end(tp), tp->snd_nxt)) 991da177e4SLinus Torvalds return tp->snd_nxt; 1001da177e4SLinus Torvalds else 10190840defSIlpo Järvinen return tcp_wnd_end(tp); 1021da177e4SLinus Torvalds } 1031da177e4SLinus Torvalds 1041da177e4SLinus Torvalds /* Calculate mss to advertise in SYN segment. 1051da177e4SLinus Torvalds * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that: 1061da177e4SLinus Torvalds * 1071da177e4SLinus Torvalds * 1. It is independent of path mtu. 1081da177e4SLinus Torvalds * 2. Ideally, it is maximal possible segment size i.e. 65535-40. 1091da177e4SLinus Torvalds * 3. For IPv4 it is reasonable to calculate it from maximal MTU of 1101da177e4SLinus Torvalds * attached devices, because some buggy hosts are confused by 1111da177e4SLinus Torvalds * large MSS. 1121da177e4SLinus Torvalds * 4. We do not make 3, we advertise MSS, calculated from first 1131da177e4SLinus Torvalds * hop device mtu, but allow to raise it to ip_rt_min_advmss. 1141da177e4SLinus Torvalds * This may be overridden via information stored in routing table. 1151da177e4SLinus Torvalds * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible, 1161da177e4SLinus Torvalds * probably even Jumbo". 1171da177e4SLinus Torvalds */ 1181da177e4SLinus Torvalds static __u16 tcp_advertise_mss(struct sock *sk) 1191da177e4SLinus Torvalds { 1201da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 121cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1221da177e4SLinus Torvalds int mss = tp->advmss; 1231da177e4SLinus Torvalds 1240dbaee3bSDavid S. Miller if (dst) { 1250dbaee3bSDavid S. Miller unsigned int metric = dst_metric_advmss(dst); 1260dbaee3bSDavid S. Miller 1270dbaee3bSDavid S. Miller if (metric < mss) { 1280dbaee3bSDavid S. Miller mss = metric; 1291da177e4SLinus Torvalds tp->advmss = mss; 1301da177e4SLinus Torvalds } 1310dbaee3bSDavid S. Miller } 1321da177e4SLinus Torvalds 1331da177e4SLinus Torvalds return (__u16)mss; 1341da177e4SLinus Torvalds } 1351da177e4SLinus Torvalds 1361da177e4SLinus Torvalds /* RFC2861. Reset CWND after idle period longer RTO to "restart window". 1371da177e4SLinus Torvalds * This is the first part of cwnd validation mechanism. */ 138cf533ea5SEric Dumazet static void tcp_cwnd_restart(struct sock *sk, const struct dst_entry *dst) 1391da177e4SLinus Torvalds { 140463c84b9SArnaldo Carvalho de Melo struct tcp_sock *tp = tcp_sk(sk); 1411da177e4SLinus Torvalds s32 delta = tcp_time_stamp - tp->lsndtime; 1421da177e4SLinus Torvalds u32 restart_cwnd = tcp_init_cwnd(tp, dst); 1431da177e4SLinus Torvalds u32 cwnd = tp->snd_cwnd; 1441da177e4SLinus Torvalds 1456687e988SArnaldo Carvalho de Melo tcp_ca_event(sk, CA_EVENT_CWND_RESTART); 1461da177e4SLinus Torvalds 1476687e988SArnaldo Carvalho de Melo tp->snd_ssthresh = tcp_current_ssthresh(sk); 1481da177e4SLinus Torvalds restart_cwnd = min(restart_cwnd, cwnd); 1491da177e4SLinus Torvalds 150463c84b9SArnaldo Carvalho de Melo while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd) 1511da177e4SLinus Torvalds cwnd >>= 1; 1521da177e4SLinus Torvalds tp->snd_cwnd = max(cwnd, restart_cwnd); 1531da177e4SLinus Torvalds tp->snd_cwnd_stamp = tcp_time_stamp; 1541da177e4SLinus Torvalds tp->snd_cwnd_used = 0; 1551da177e4SLinus Torvalds } 1561da177e4SLinus Torvalds 15767edfef7SAndi Kleen /* Congestion state accounting after a packet has been sent. */ 15840efc6faSStephen Hemminger static void tcp_event_data_sent(struct tcp_sock *tp, 159cf533ea5SEric Dumazet struct sock *sk) 1601da177e4SLinus Torvalds { 161463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 162463c84b9SArnaldo Carvalho de Melo const u32 now = tcp_time_stamp; 1631da177e4SLinus Torvalds 16435089bb2SDavid S. Miller if (sysctl_tcp_slow_start_after_idle && 16535089bb2SDavid S. Miller (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto)) 166463c84b9SArnaldo Carvalho de Melo tcp_cwnd_restart(sk, __sk_dst_get(sk)); 1671da177e4SLinus Torvalds 1681da177e4SLinus Torvalds tp->lsndtime = now; 1691da177e4SLinus Torvalds 1701da177e4SLinus Torvalds /* If it is a reply for ato after last received 1711da177e4SLinus Torvalds * packet, enter pingpong mode. 1721da177e4SLinus Torvalds */ 173463c84b9SArnaldo Carvalho de Melo if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato) 174463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pingpong = 1; 1751da177e4SLinus Torvalds } 1761da177e4SLinus Torvalds 17767edfef7SAndi Kleen /* Account for an ACK we sent. */ 17840efc6faSStephen Hemminger static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts) 1791da177e4SLinus Torvalds { 180463c84b9SArnaldo Carvalho de Melo tcp_dec_quickack_mode(sk, pkts); 181463c84b9SArnaldo Carvalho de Melo inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); 1821da177e4SLinus Torvalds } 1831da177e4SLinus Torvalds 1841da177e4SLinus Torvalds /* Determine a window scaling and initial window to offer. 1851da177e4SLinus Torvalds * Based on the assumption that the given amount of space 1861da177e4SLinus Torvalds * will be offered. Store the results in the tp structure. 1871da177e4SLinus Torvalds * NOTE: for smooth operation initial space offering should 1881da177e4SLinus Torvalds * be a multiple of mss if possible. We assume here that mss >= 1. 1891da177e4SLinus Torvalds * This MUST be enforced by all callers. 1901da177e4SLinus Torvalds */ 1911da177e4SLinus Torvalds void tcp_select_initial_window(int __space, __u32 mss, 1921da177e4SLinus Torvalds __u32 *rcv_wnd, __u32 *window_clamp, 19331d12926Slaurent chavey int wscale_ok, __u8 *rcv_wscale, 19431d12926Slaurent chavey __u32 init_rcv_wnd) 1951da177e4SLinus Torvalds { 1961da177e4SLinus Torvalds unsigned int space = (__space < 0 ? 0 : __space); 1971da177e4SLinus Torvalds 1981da177e4SLinus Torvalds /* If no clamp set the clamp to the max possible scaled window */ 1991da177e4SLinus Torvalds if (*window_clamp == 0) 2001da177e4SLinus Torvalds (*window_clamp) = (65535 << 14); 2011da177e4SLinus Torvalds space = min(*window_clamp, space); 2021da177e4SLinus Torvalds 2031da177e4SLinus Torvalds /* Quantize space offering to a multiple of mss if possible. */ 2041da177e4SLinus Torvalds if (space > mss) 2051da177e4SLinus Torvalds space = (space / mss) * mss; 2061da177e4SLinus Torvalds 2071da177e4SLinus Torvalds /* NOTE: offering an initial window larger than 32767 20815d99e02SRick Jones * will break some buggy TCP stacks. If the admin tells us 20915d99e02SRick Jones * it is likely we could be speaking with such a buggy stack 21015d99e02SRick Jones * we will truncate our initial window offering to 32K-1 21115d99e02SRick Jones * unless the remote has sent us a window scaling option, 21215d99e02SRick Jones * which we interpret as a sign the remote TCP is not 21315d99e02SRick Jones * misinterpreting the window field as a signed quantity. 2141da177e4SLinus Torvalds */ 21515d99e02SRick Jones if (sysctl_tcp_workaround_signed_windows) 2161da177e4SLinus Torvalds (*rcv_wnd) = min(space, MAX_TCP_WINDOW); 21715d99e02SRick Jones else 21815d99e02SRick Jones (*rcv_wnd) = space; 21915d99e02SRick Jones 2201da177e4SLinus Torvalds (*rcv_wscale) = 0; 2211da177e4SLinus Torvalds if (wscale_ok) { 2221da177e4SLinus Torvalds /* Set window scaling on max possible window 2231da177e4SLinus Torvalds * See RFC1323 for an explanation of the limit to 14 2241da177e4SLinus Torvalds */ 2251da177e4SLinus Torvalds space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max); 226316c1592SStephen Hemminger space = min_t(u32, space, *window_clamp); 2271da177e4SLinus Torvalds while (space > 65535 && (*rcv_wscale) < 14) { 2281da177e4SLinus Torvalds space >>= 1; 2291da177e4SLinus Torvalds (*rcv_wscale)++; 2301da177e4SLinus Torvalds } 2311da177e4SLinus Torvalds } 2321da177e4SLinus Torvalds 233356f0398SNandita Dukkipati /* Set initial window to a value enough for senders starting with 234356f0398SNandita Dukkipati * initial congestion window of TCP_DEFAULT_INIT_RCVWND. Place 235356f0398SNandita Dukkipati * a limit on the initial window when mss is larger than 1460. 236356f0398SNandita Dukkipati */ 2371da177e4SLinus Torvalds if (mss > (1 << *rcv_wscale)) { 238356f0398SNandita Dukkipati int init_cwnd = TCP_DEFAULT_INIT_RCVWND; 239356f0398SNandita Dukkipati if (mss > 1460) 240356f0398SNandita Dukkipati init_cwnd = 241356f0398SNandita Dukkipati max_t(u32, (1460 * TCP_DEFAULT_INIT_RCVWND) / mss, 2); 24231d12926Slaurent chavey /* when initializing use the value from init_rcv_wnd 24331d12926Slaurent chavey * rather than the default from above 24431d12926Slaurent chavey */ 245b1afde60SNandita Dukkipati if (init_rcv_wnd) 246b1afde60SNandita Dukkipati *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss); 247b1afde60SNandita Dukkipati else 248b1afde60SNandita Dukkipati *rcv_wnd = min(*rcv_wnd, init_cwnd * mss); 2491da177e4SLinus Torvalds } 2501da177e4SLinus Torvalds 2511da177e4SLinus Torvalds /* Set the clamp no higher than max representable value */ 2521da177e4SLinus Torvalds (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp); 2531da177e4SLinus Torvalds } 2544bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_select_initial_window); 2551da177e4SLinus Torvalds 2561da177e4SLinus Torvalds /* Chose a new window to advertise, update state in tcp_sock for the 2571da177e4SLinus Torvalds * socket, and return result with RFC1323 scaling applied. The return 2581da177e4SLinus Torvalds * value can be stuffed directly into th->window for an outgoing 2591da177e4SLinus Torvalds * frame. 2601da177e4SLinus Torvalds */ 26140efc6faSStephen Hemminger static u16 tcp_select_window(struct sock *sk) 2621da177e4SLinus Torvalds { 2631da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2641da177e4SLinus Torvalds u32 cur_win = tcp_receive_window(tp); 2651da177e4SLinus Torvalds u32 new_win = __tcp_select_window(sk); 2661da177e4SLinus Torvalds 2671da177e4SLinus Torvalds /* Never shrink the offered window */ 2681da177e4SLinus Torvalds if (new_win < cur_win) { 2691da177e4SLinus Torvalds /* Danger Will Robinson! 2701da177e4SLinus Torvalds * Don't update rcv_wup/rcv_wnd here or else 2711da177e4SLinus Torvalds * we will not be able to advertise a zero 2721da177e4SLinus Torvalds * window in time. --DaveM 2731da177e4SLinus Torvalds * 2741da177e4SLinus Torvalds * Relax Will Robinson. 2751da177e4SLinus Torvalds */ 276607bfbf2SPatrick McHardy new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale); 2771da177e4SLinus Torvalds } 2781da177e4SLinus Torvalds tp->rcv_wnd = new_win; 2791da177e4SLinus Torvalds tp->rcv_wup = tp->rcv_nxt; 2801da177e4SLinus Torvalds 2811da177e4SLinus Torvalds /* Make sure we do not exceed the maximum possible 2821da177e4SLinus Torvalds * scaled window. 2831da177e4SLinus Torvalds */ 28415d99e02SRick Jones if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows) 2851da177e4SLinus Torvalds new_win = min(new_win, MAX_TCP_WINDOW); 2861da177e4SLinus Torvalds else 2871da177e4SLinus Torvalds new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); 2881da177e4SLinus Torvalds 2891da177e4SLinus Torvalds /* RFC1323 scaling applied */ 2901da177e4SLinus Torvalds new_win >>= tp->rx_opt.rcv_wscale; 2911da177e4SLinus Torvalds 2921da177e4SLinus Torvalds /* If we advertise zero window, disable fast path. */ 2931da177e4SLinus Torvalds if (new_win == 0) 2941da177e4SLinus Torvalds tp->pred_flags = 0; 2951da177e4SLinus Torvalds 2961da177e4SLinus Torvalds return new_win; 2971da177e4SLinus Torvalds } 2981da177e4SLinus Torvalds 29967edfef7SAndi Kleen /* Packet ECN state for a SYN-ACK */ 300cf533ea5SEric Dumazet static inline void TCP_ECN_send_synack(const struct tcp_sock *tp, struct sk_buff *skb) 301bdf1ee5dSIlpo Järvinen { 3024de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR; 303bdf1ee5dSIlpo Järvinen if (!(tp->ecn_flags & TCP_ECN_OK)) 3044de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE; 305bdf1ee5dSIlpo Järvinen } 306bdf1ee5dSIlpo Järvinen 30767edfef7SAndi Kleen /* Packet ECN state for a SYN. */ 308bdf1ee5dSIlpo Järvinen static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb) 309bdf1ee5dSIlpo Järvinen { 310bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 311bdf1ee5dSIlpo Järvinen 312bdf1ee5dSIlpo Järvinen tp->ecn_flags = 0; 3135d134f1cSHannes Frederic Sowa if (sock_net(sk)->ipv4.sysctl_tcp_ecn == 1) { 3144de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR; 315bdf1ee5dSIlpo Järvinen tp->ecn_flags = TCP_ECN_OK; 316bdf1ee5dSIlpo Järvinen } 317bdf1ee5dSIlpo Järvinen } 318bdf1ee5dSIlpo Järvinen 319bdf1ee5dSIlpo Järvinen static __inline__ void 320cf533ea5SEric Dumazet TCP_ECN_make_synack(const struct request_sock *req, struct tcphdr *th) 321bdf1ee5dSIlpo Järvinen { 322bdf1ee5dSIlpo Järvinen if (inet_rsk(req)->ecn_ok) 323bdf1ee5dSIlpo Järvinen th->ece = 1; 324bdf1ee5dSIlpo Järvinen } 325bdf1ee5dSIlpo Järvinen 32667edfef7SAndi Kleen /* Set up ECN state for a packet on a ESTABLISHED socket that is about to 32767edfef7SAndi Kleen * be sent. 32867edfef7SAndi Kleen */ 329bdf1ee5dSIlpo Järvinen static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb, 330bdf1ee5dSIlpo Järvinen int tcp_header_len) 331bdf1ee5dSIlpo Järvinen { 332bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 333bdf1ee5dSIlpo Järvinen 334bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_OK) { 335bdf1ee5dSIlpo Järvinen /* Not-retransmitted data segment: set ECT and inject CWR. */ 336bdf1ee5dSIlpo Järvinen if (skb->len != tcp_header_len && 337bdf1ee5dSIlpo Järvinen !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { 338bdf1ee5dSIlpo Järvinen INET_ECN_xmit(sk); 339bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) { 340bdf1ee5dSIlpo Järvinen tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; 341bdf1ee5dSIlpo Järvinen tcp_hdr(skb)->cwr = 1; 342bdf1ee5dSIlpo Järvinen skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; 343bdf1ee5dSIlpo Järvinen } 344bdf1ee5dSIlpo Järvinen } else { 345bdf1ee5dSIlpo Järvinen /* ACK or retransmitted segment: clear ECT|CE */ 346bdf1ee5dSIlpo Järvinen INET_ECN_dontxmit(sk); 347bdf1ee5dSIlpo Järvinen } 348bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) 349bdf1ee5dSIlpo Järvinen tcp_hdr(skb)->ece = 1; 350bdf1ee5dSIlpo Järvinen } 351bdf1ee5dSIlpo Järvinen } 352bdf1ee5dSIlpo Järvinen 353e870a8efSIlpo Järvinen /* Constructs common control bits of non-data skb. If SYN/FIN is present, 354e870a8efSIlpo Järvinen * auto increment end seqno. 355e870a8efSIlpo Järvinen */ 356e870a8efSIlpo Järvinen static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags) 357e870a8efSIlpo Järvinen { 3582e8e18efSDavid S. Miller skb->ip_summed = CHECKSUM_PARTIAL; 359e870a8efSIlpo Järvinen skb->csum = 0; 360e870a8efSIlpo Järvinen 3614de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags; 362e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->sacked = 0; 363e870a8efSIlpo Järvinen 364e870a8efSIlpo Järvinen skb_shinfo(skb)->gso_segs = 1; 365e870a8efSIlpo Järvinen skb_shinfo(skb)->gso_size = 0; 366e870a8efSIlpo Järvinen skb_shinfo(skb)->gso_type = 0; 367e870a8efSIlpo Järvinen 368e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->seq = seq; 369a3433f35SChangli Gao if (flags & (TCPHDR_SYN | TCPHDR_FIN)) 370e870a8efSIlpo Järvinen seq++; 371e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->end_seq = seq; 372e870a8efSIlpo Järvinen } 373e870a8efSIlpo Järvinen 374a2a385d6SEric Dumazet static inline bool tcp_urg_mode(const struct tcp_sock *tp) 37533f5f57eSIlpo Järvinen { 37633f5f57eSIlpo Järvinen return tp->snd_una != tp->snd_up; 37733f5f57eSIlpo Järvinen } 37833f5f57eSIlpo Järvinen 37933ad798cSAdam Langley #define OPTION_SACK_ADVERTISE (1 << 0) 38033ad798cSAdam Langley #define OPTION_TS (1 << 1) 38133ad798cSAdam Langley #define OPTION_MD5 (1 << 2) 38289e95a61SOri Finkelman #define OPTION_WSCALE (1 << 3) 3832100c8d2SYuchung Cheng #define OPTION_FAST_OPEN_COOKIE (1 << 8) 38433ad798cSAdam Langley 38533ad798cSAdam Langley struct tcp_out_options { 3862100c8d2SYuchung Cheng u16 options; /* bit field of OPTION_* */ 3872100c8d2SYuchung Cheng u16 mss; /* 0 to disable */ 38833ad798cSAdam Langley u8 ws; /* window scale, 0 to disable */ 38933ad798cSAdam Langley u8 num_sack_blocks; /* number of SACK blocks to include */ 390bd0388aeSWilliam Allen Simpson u8 hash_size; /* bytes in hash_location */ 391bd0388aeSWilliam Allen Simpson __u8 *hash_location; /* temporary pointer, overloaded */ 3922100c8d2SYuchung Cheng __u32 tsval, tsecr; /* need to include OPTION_TS */ 3932100c8d2SYuchung Cheng struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */ 39433ad798cSAdam Langley }; 39533ad798cSAdam Langley 39667edfef7SAndi Kleen /* Write previously computed TCP options to the packet. 39767edfef7SAndi Kleen * 39867edfef7SAndi Kleen * Beware: Something in the Internet is very sensitive to the ordering of 399fd6149d3SIlpo Järvinen * TCP options, we learned this through the hard way, so be careful here. 400fd6149d3SIlpo Järvinen * Luckily we can at least blame others for their non-compliance but from 401fd6149d3SIlpo Järvinen * inter-operatibility perspective it seems that we're somewhat stuck with 402fd6149d3SIlpo Järvinen * the ordering which we have been using if we want to keep working with 403fd6149d3SIlpo Järvinen * those broken things (not that it currently hurts anybody as there isn't 404fd6149d3SIlpo Järvinen * particular reason why the ordering would need to be changed). 405fd6149d3SIlpo Järvinen * 406fd6149d3SIlpo Järvinen * At least SACK_PERM as the first option is known to lead to a disaster 407fd6149d3SIlpo Järvinen * (but it may well be that other scenarios fail similarly). 408fd6149d3SIlpo Järvinen */ 40933ad798cSAdam Langley static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp, 410bd0388aeSWilliam Allen Simpson struct tcp_out_options *opts) 411bd0388aeSWilliam Allen Simpson { 4122100c8d2SYuchung Cheng u16 options = opts->options; /* mungable copy */ 413bd0388aeSWilliam Allen Simpson 414bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_MD5 & options)) { 4151a2c6181SChristoph Paasch *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | 4161a2c6181SChristoph Paasch (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG); 417bd0388aeSWilliam Allen Simpson /* overload cookie hash location */ 418bd0388aeSWilliam Allen Simpson opts->hash_location = (__u8 *)ptr; 41933ad798cSAdam Langley ptr += 4; 42033ad798cSAdam Langley } 42133ad798cSAdam Langley 422fd6149d3SIlpo Järvinen if (unlikely(opts->mss)) { 423fd6149d3SIlpo Järvinen *ptr++ = htonl((TCPOPT_MSS << 24) | 424fd6149d3SIlpo Järvinen (TCPOLEN_MSS << 16) | 425fd6149d3SIlpo Järvinen opts->mss); 426fd6149d3SIlpo Järvinen } 427fd6149d3SIlpo Järvinen 428bd0388aeSWilliam Allen Simpson if (likely(OPTION_TS & options)) { 429bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 43033ad798cSAdam Langley *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | 43133ad798cSAdam Langley (TCPOLEN_SACK_PERM << 16) | 43233ad798cSAdam Langley (TCPOPT_TIMESTAMP << 8) | 43333ad798cSAdam Langley TCPOLEN_TIMESTAMP); 434bd0388aeSWilliam Allen Simpson options &= ~OPTION_SACK_ADVERTISE; 43533ad798cSAdam Langley } else { 436496c98dfSYOSHIFUJI Hideaki *ptr++ = htonl((TCPOPT_NOP << 24) | 43740efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 43840efc6faSStephen Hemminger (TCPOPT_TIMESTAMP << 8) | 43940efc6faSStephen Hemminger TCPOLEN_TIMESTAMP); 44040efc6faSStephen Hemminger } 44133ad798cSAdam Langley *ptr++ = htonl(opts->tsval); 44233ad798cSAdam Langley *ptr++ = htonl(opts->tsecr); 44333ad798cSAdam Langley } 44433ad798cSAdam Langley 445bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 44633ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 44733ad798cSAdam Langley (TCPOPT_NOP << 16) | 44833ad798cSAdam Langley (TCPOPT_SACK_PERM << 8) | 44933ad798cSAdam Langley TCPOLEN_SACK_PERM); 45033ad798cSAdam Langley } 45133ad798cSAdam Langley 452bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_WSCALE & options)) { 45333ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 45433ad798cSAdam Langley (TCPOPT_WINDOW << 16) | 45533ad798cSAdam Langley (TCPOLEN_WINDOW << 8) | 45633ad798cSAdam Langley opts->ws); 45733ad798cSAdam Langley } 45833ad798cSAdam Langley 45933ad798cSAdam Langley if (unlikely(opts->num_sack_blocks)) { 46033ad798cSAdam Langley struct tcp_sack_block *sp = tp->rx_opt.dsack ? 46133ad798cSAdam Langley tp->duplicate_sack : tp->selective_acks; 46240efc6faSStephen Hemminger int this_sack; 46340efc6faSStephen Hemminger 46440efc6faSStephen Hemminger *ptr++ = htonl((TCPOPT_NOP << 24) | 46540efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 46640efc6faSStephen Hemminger (TCPOPT_SACK << 8) | 46733ad798cSAdam Langley (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * 46840efc6faSStephen Hemminger TCPOLEN_SACK_PERBLOCK))); 4692de979bdSStephen Hemminger 47033ad798cSAdam Langley for (this_sack = 0; this_sack < opts->num_sack_blocks; 47133ad798cSAdam Langley ++this_sack) { 47240efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].start_seq); 47340efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].end_seq); 47440efc6faSStephen Hemminger } 4752de979bdSStephen Hemminger 47640efc6faSStephen Hemminger tp->rx_opt.dsack = 0; 47740efc6faSStephen Hemminger } 4782100c8d2SYuchung Cheng 4792100c8d2SYuchung Cheng if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) { 4802100c8d2SYuchung Cheng struct tcp_fastopen_cookie *foc = opts->fastopen_cookie; 4812100c8d2SYuchung Cheng 4822100c8d2SYuchung Cheng *ptr++ = htonl((TCPOPT_EXP << 24) | 4832100c8d2SYuchung Cheng ((TCPOLEN_EXP_FASTOPEN_BASE + foc->len) << 16) | 4842100c8d2SYuchung Cheng TCPOPT_FASTOPEN_MAGIC); 4852100c8d2SYuchung Cheng 4862100c8d2SYuchung Cheng memcpy(ptr, foc->val, foc->len); 4872100c8d2SYuchung Cheng if ((foc->len & 3) == 2) { 4882100c8d2SYuchung Cheng u8 *align = ((u8 *)ptr) + foc->len; 4892100c8d2SYuchung Cheng align[0] = align[1] = TCPOPT_NOP; 4902100c8d2SYuchung Cheng } 4912100c8d2SYuchung Cheng ptr += (foc->len + 3) >> 2; 4922100c8d2SYuchung Cheng } 49340efc6faSStephen Hemminger } 49440efc6faSStephen Hemminger 49567edfef7SAndi Kleen /* Compute TCP options for SYN packets. This is not the final 49667edfef7SAndi Kleen * network wire format yet. 49767edfef7SAndi Kleen */ 49895c96174SEric Dumazet static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb, 49933ad798cSAdam Langley struct tcp_out_options *opts, 500cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 501cf533ea5SEric Dumazet { 50233ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 50395c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 504783237e8SYuchung Cheng struct tcp_fastopen_request *fastopen = tp->fastopen_req; 50533ad798cSAdam Langley 506cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 50733ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 50833ad798cSAdam Langley if (*md5) { 50933ad798cSAdam Langley opts->options |= OPTION_MD5; 510bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 511cfb6eeb4SYOSHIFUJI Hideaki } 51233ad798cSAdam Langley #else 51333ad798cSAdam Langley *md5 = NULL; 514cfb6eeb4SYOSHIFUJI Hideaki #endif 51533ad798cSAdam Langley 51633ad798cSAdam Langley /* We always get an MSS option. The option bytes which will be seen in 51733ad798cSAdam Langley * normal data packets should timestamps be used, must be in the MSS 51833ad798cSAdam Langley * advertised. But we subtract them from tp->mss_cache so that 51933ad798cSAdam Langley * calculations in tcp_sendmsg are simpler etc. So account for this 52033ad798cSAdam Langley * fact here if necessary. If we don't do this correctly, as a 52133ad798cSAdam Langley * receiver we won't recognize data packets as being full sized when we 52233ad798cSAdam Langley * should, and thus we won't abide by the delayed ACK rules correctly. 52333ad798cSAdam Langley * SACKs don't matter, we never delay an ACK when we have any of those 52433ad798cSAdam Langley * going out. */ 52533ad798cSAdam Langley opts->mss = tcp_advertise_mss(sk); 526bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 52733ad798cSAdam Langley 528bb5b7c11SDavid S. Miller if (likely(sysctl_tcp_timestamps && *md5 == NULL)) { 52933ad798cSAdam Langley opts->options |= OPTION_TS; 530ee684b6fSAndrey Vagin opts->tsval = TCP_SKB_CB(skb)->when + tp->tsoffset; 53133ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 532bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 53333ad798cSAdam Langley } 534bb5b7c11SDavid S. Miller if (likely(sysctl_tcp_window_scaling)) { 53533ad798cSAdam Langley opts->ws = tp->rx_opt.rcv_wscale; 53689e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 537bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 53833ad798cSAdam Langley } 539bb5b7c11SDavid S. Miller if (likely(sysctl_tcp_sack)) { 54033ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 541b32d1310SDavid S. Miller if (unlikely(!(OPTION_TS & opts->options))) 542bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 54333ad798cSAdam Langley } 54433ad798cSAdam Langley 545783237e8SYuchung Cheng if (fastopen && fastopen->cookie.len >= 0) { 546783237e8SYuchung Cheng u32 need = TCPOLEN_EXP_FASTOPEN_BASE + fastopen->cookie.len; 547783237e8SYuchung Cheng need = (need + 3) & ~3U; /* Align to 32 bits */ 548783237e8SYuchung Cheng if (remaining >= need) { 549783237e8SYuchung Cheng opts->options |= OPTION_FAST_OPEN_COOKIE; 550783237e8SYuchung Cheng opts->fastopen_cookie = &fastopen->cookie; 551783237e8SYuchung Cheng remaining -= need; 552783237e8SYuchung Cheng tp->syn_fastopen = 1; 553783237e8SYuchung Cheng } 554783237e8SYuchung Cheng } 555bd0388aeSWilliam Allen Simpson 556bd0388aeSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 55733ad798cSAdam Langley } 55833ad798cSAdam Langley 55967edfef7SAndi Kleen /* Set up TCP options for SYN-ACKs. */ 56095c96174SEric Dumazet static unsigned int tcp_synack_options(struct sock *sk, 56133ad798cSAdam Langley struct request_sock *req, 56295c96174SEric Dumazet unsigned int mss, struct sk_buff *skb, 56333ad798cSAdam Langley struct tcp_out_options *opts, 5644957faadSWilliam Allen Simpson struct tcp_md5sig_key **md5, 5658336886fSJerry Chu struct tcp_fastopen_cookie *foc) 5664957faadSWilliam Allen Simpson { 56733ad798cSAdam Langley struct inet_request_sock *ireq = inet_rsk(req); 56895c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 56933ad798cSAdam Langley 57033ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 57133ad798cSAdam Langley *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req); 57233ad798cSAdam Langley if (*md5) { 57333ad798cSAdam Langley opts->options |= OPTION_MD5; 5744957faadSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 5754957faadSWilliam Allen Simpson 5764957faadSWilliam Allen Simpson /* We can't fit any SACK blocks in a packet with MD5 + TS 5774957faadSWilliam Allen Simpson * options. There was discussion about disabling SACK 5784957faadSWilliam Allen Simpson * rather than TS in order to fit in better with old, 5794957faadSWilliam Allen Simpson * buggy kernels, but that was deemed to be unnecessary. 5804957faadSWilliam Allen Simpson */ 581de213e5eSEric Dumazet ireq->tstamp_ok &= !ireq->sack_ok; 58233ad798cSAdam Langley } 58333ad798cSAdam Langley #else 58433ad798cSAdam Langley *md5 = NULL; 58533ad798cSAdam Langley #endif 58633ad798cSAdam Langley 5874957faadSWilliam Allen Simpson /* We always send an MSS option. */ 58833ad798cSAdam Langley opts->mss = mss; 5894957faadSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 59033ad798cSAdam Langley 59133ad798cSAdam Langley if (likely(ireq->wscale_ok)) { 59233ad798cSAdam Langley opts->ws = ireq->rcv_wscale; 59389e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 5944957faadSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 59533ad798cSAdam Langley } 596de213e5eSEric Dumazet if (likely(ireq->tstamp_ok)) { 59733ad798cSAdam Langley opts->options |= OPTION_TS; 59833ad798cSAdam Langley opts->tsval = TCP_SKB_CB(skb)->when; 59933ad798cSAdam Langley opts->tsecr = req->ts_recent; 6004957faadSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 60133ad798cSAdam Langley } 60233ad798cSAdam Langley if (likely(ireq->sack_ok)) { 60333ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 604de213e5eSEric Dumazet if (unlikely(!ireq->tstamp_ok)) 6054957faadSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 60633ad798cSAdam Langley } 6078336886fSJerry Chu if (foc != NULL) { 6088336886fSJerry Chu u32 need = TCPOLEN_EXP_FASTOPEN_BASE + foc->len; 6098336886fSJerry Chu need = (need + 3) & ~3U; /* Align to 32 bits */ 6108336886fSJerry Chu if (remaining >= need) { 6118336886fSJerry Chu opts->options |= OPTION_FAST_OPEN_COOKIE; 6128336886fSJerry Chu opts->fastopen_cookie = foc; 6138336886fSJerry Chu remaining -= need; 6148336886fSJerry Chu } 6158336886fSJerry Chu } 6164957faadSWilliam Allen Simpson 6174957faadSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 61833ad798cSAdam Langley } 61933ad798cSAdam Langley 62067edfef7SAndi Kleen /* Compute TCP options for ESTABLISHED sockets. This is not the 62167edfef7SAndi Kleen * final wire format yet. 62267edfef7SAndi Kleen */ 62395c96174SEric Dumazet static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb, 62433ad798cSAdam Langley struct tcp_out_options *opts, 625cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 626cf533ea5SEric Dumazet { 62733ad798cSAdam Langley struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL; 62833ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 62995c96174SEric Dumazet unsigned int size = 0; 630cabeccbdSIlpo Järvinen unsigned int eff_sacks; 63133ad798cSAdam Langley 63233ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 63333ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 63433ad798cSAdam Langley if (unlikely(*md5)) { 63533ad798cSAdam Langley opts->options |= OPTION_MD5; 63633ad798cSAdam Langley size += TCPOLEN_MD5SIG_ALIGNED; 63733ad798cSAdam Langley } 63833ad798cSAdam Langley #else 63933ad798cSAdam Langley *md5 = NULL; 64033ad798cSAdam Langley #endif 64133ad798cSAdam Langley 64233ad798cSAdam Langley if (likely(tp->rx_opt.tstamp_ok)) { 64333ad798cSAdam Langley opts->options |= OPTION_TS; 644ee684b6fSAndrey Vagin opts->tsval = tcb ? tcb->when + tp->tsoffset : 0; 64533ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 64633ad798cSAdam Langley size += TCPOLEN_TSTAMP_ALIGNED; 64733ad798cSAdam Langley } 64833ad798cSAdam Langley 649cabeccbdSIlpo Järvinen eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; 650cabeccbdSIlpo Järvinen if (unlikely(eff_sacks)) { 65195c96174SEric Dumazet const unsigned int remaining = MAX_TCP_OPTION_SPACE - size; 65233ad798cSAdam Langley opts->num_sack_blocks = 65395c96174SEric Dumazet min_t(unsigned int, eff_sacks, 65433ad798cSAdam Langley (remaining - TCPOLEN_SACK_BASE_ALIGNED) / 65533ad798cSAdam Langley TCPOLEN_SACK_PERBLOCK); 65633ad798cSAdam Langley size += TCPOLEN_SACK_BASE_ALIGNED + 65733ad798cSAdam Langley opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; 65833ad798cSAdam Langley } 65933ad798cSAdam Langley 66033ad798cSAdam Langley return size; 66140efc6faSStephen Hemminger } 6621da177e4SLinus Torvalds 66346d3ceabSEric Dumazet 66446d3ceabSEric Dumazet /* TCP SMALL QUEUES (TSQ) 66546d3ceabSEric Dumazet * 66646d3ceabSEric Dumazet * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev) 66746d3ceabSEric Dumazet * to reduce RTT and bufferbloat. 66846d3ceabSEric Dumazet * We do this using a special skb destructor (tcp_wfree). 66946d3ceabSEric Dumazet * 67046d3ceabSEric Dumazet * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb 67146d3ceabSEric Dumazet * needs to be reallocated in a driver. 67246d3ceabSEric Dumazet * The invariant being skb->truesize substracted from sk->sk_wmem_alloc 67346d3ceabSEric Dumazet * 67446d3ceabSEric Dumazet * Since transmit from skb destructor is forbidden, we use a tasklet 67546d3ceabSEric Dumazet * to process all sockets that eventually need to send more skbs. 67646d3ceabSEric Dumazet * We use one tasklet per cpu, with its own queue of sockets. 67746d3ceabSEric Dumazet */ 67846d3ceabSEric Dumazet struct tsq_tasklet { 67946d3ceabSEric Dumazet struct tasklet_struct tasklet; 68046d3ceabSEric Dumazet struct list_head head; /* queue of tcp sockets */ 68146d3ceabSEric Dumazet }; 68246d3ceabSEric Dumazet static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet); 68346d3ceabSEric Dumazet 6846f458dfbSEric Dumazet static void tcp_tsq_handler(struct sock *sk) 6856f458dfbSEric Dumazet { 6866f458dfbSEric Dumazet if ((1 << sk->sk_state) & 6876f458dfbSEric Dumazet (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING | 6886f458dfbSEric Dumazet TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) 6896f458dfbSEric Dumazet tcp_write_xmit(sk, tcp_current_mss(sk), 0, 0, GFP_ATOMIC); 6906f458dfbSEric Dumazet } 69146d3ceabSEric Dumazet /* 69246d3ceabSEric Dumazet * One tasklest per cpu tries to send more skbs. 69346d3ceabSEric Dumazet * We run in tasklet context but need to disable irqs when 69446d3ceabSEric Dumazet * transfering tsq->head because tcp_wfree() might 69546d3ceabSEric Dumazet * interrupt us (non NAPI drivers) 69646d3ceabSEric Dumazet */ 69746d3ceabSEric Dumazet static void tcp_tasklet_func(unsigned long data) 69846d3ceabSEric Dumazet { 69946d3ceabSEric Dumazet struct tsq_tasklet *tsq = (struct tsq_tasklet *)data; 70046d3ceabSEric Dumazet LIST_HEAD(list); 70146d3ceabSEric Dumazet unsigned long flags; 70246d3ceabSEric Dumazet struct list_head *q, *n; 70346d3ceabSEric Dumazet struct tcp_sock *tp; 70446d3ceabSEric Dumazet struct sock *sk; 70546d3ceabSEric Dumazet 70646d3ceabSEric Dumazet local_irq_save(flags); 70746d3ceabSEric Dumazet list_splice_init(&tsq->head, &list); 70846d3ceabSEric Dumazet local_irq_restore(flags); 70946d3ceabSEric Dumazet 71046d3ceabSEric Dumazet list_for_each_safe(q, n, &list) { 71146d3ceabSEric Dumazet tp = list_entry(q, struct tcp_sock, tsq_node); 71246d3ceabSEric Dumazet list_del(&tp->tsq_node); 71346d3ceabSEric Dumazet 71446d3ceabSEric Dumazet sk = (struct sock *)tp; 71546d3ceabSEric Dumazet bh_lock_sock(sk); 71646d3ceabSEric Dumazet 71746d3ceabSEric Dumazet if (!sock_owned_by_user(sk)) { 7186f458dfbSEric Dumazet tcp_tsq_handler(sk); 71946d3ceabSEric Dumazet } else { 72046d3ceabSEric Dumazet /* defer the work to tcp_release_cb() */ 7216f458dfbSEric Dumazet set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags); 72246d3ceabSEric Dumazet } 72346d3ceabSEric Dumazet bh_unlock_sock(sk); 72446d3ceabSEric Dumazet 72546d3ceabSEric Dumazet clear_bit(TSQ_QUEUED, &tp->tsq_flags); 72646d3ceabSEric Dumazet sk_free(sk); 72746d3ceabSEric Dumazet } 72846d3ceabSEric Dumazet } 72946d3ceabSEric Dumazet 7306f458dfbSEric Dumazet #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \ 7316f458dfbSEric Dumazet (1UL << TCP_WRITE_TIMER_DEFERRED) | \ 732563d34d0SEric Dumazet (1UL << TCP_DELACK_TIMER_DEFERRED) | \ 733563d34d0SEric Dumazet (1UL << TCP_MTU_REDUCED_DEFERRED)) 73446d3ceabSEric Dumazet /** 73546d3ceabSEric Dumazet * tcp_release_cb - tcp release_sock() callback 73646d3ceabSEric Dumazet * @sk: socket 73746d3ceabSEric Dumazet * 73846d3ceabSEric Dumazet * called from release_sock() to perform protocol dependent 73946d3ceabSEric Dumazet * actions before socket release. 74046d3ceabSEric Dumazet */ 74146d3ceabSEric Dumazet void tcp_release_cb(struct sock *sk) 74246d3ceabSEric Dumazet { 74346d3ceabSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 7446f458dfbSEric Dumazet unsigned long flags, nflags; 74546d3ceabSEric Dumazet 7466f458dfbSEric Dumazet /* perform an atomic operation only if at least one flag is set */ 7476f458dfbSEric Dumazet do { 7486f458dfbSEric Dumazet flags = tp->tsq_flags; 7496f458dfbSEric Dumazet if (!(flags & TCP_DEFERRED_ALL)) 7506f458dfbSEric Dumazet return; 7516f458dfbSEric Dumazet nflags = flags & ~TCP_DEFERRED_ALL; 7526f458dfbSEric Dumazet } while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags); 7536f458dfbSEric Dumazet 7546f458dfbSEric Dumazet if (flags & (1UL << TCP_TSQ_DEFERRED)) 7556f458dfbSEric Dumazet tcp_tsq_handler(sk); 7566f458dfbSEric Dumazet 757144d56e9SEric Dumazet if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) { 7586f458dfbSEric Dumazet tcp_write_timer_handler(sk); 759144d56e9SEric Dumazet __sock_put(sk); 760144d56e9SEric Dumazet } 761144d56e9SEric Dumazet if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) { 7626f458dfbSEric Dumazet tcp_delack_timer_handler(sk); 763144d56e9SEric Dumazet __sock_put(sk); 764144d56e9SEric Dumazet } 765144d56e9SEric Dumazet if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) { 766563d34d0SEric Dumazet sk->sk_prot->mtu_reduced(sk); 767144d56e9SEric Dumazet __sock_put(sk); 768144d56e9SEric Dumazet } 76946d3ceabSEric Dumazet } 77046d3ceabSEric Dumazet EXPORT_SYMBOL(tcp_release_cb); 77146d3ceabSEric Dumazet 77246d3ceabSEric Dumazet void __init tcp_tasklet_init(void) 77346d3ceabSEric Dumazet { 77446d3ceabSEric Dumazet int i; 77546d3ceabSEric Dumazet 77646d3ceabSEric Dumazet for_each_possible_cpu(i) { 77746d3ceabSEric Dumazet struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i); 77846d3ceabSEric Dumazet 77946d3ceabSEric Dumazet INIT_LIST_HEAD(&tsq->head); 78046d3ceabSEric Dumazet tasklet_init(&tsq->tasklet, 78146d3ceabSEric Dumazet tcp_tasklet_func, 78246d3ceabSEric Dumazet (unsigned long)tsq); 78346d3ceabSEric Dumazet } 78446d3ceabSEric Dumazet } 78546d3ceabSEric Dumazet 78646d3ceabSEric Dumazet /* 78746d3ceabSEric Dumazet * Write buffer destructor automatically called from kfree_skb. 78846d3ceabSEric Dumazet * We cant xmit new skbs from this context, as we might already 78946d3ceabSEric Dumazet * hold qdisc lock. 79046d3ceabSEric Dumazet */ 791d6a4a104SEric Dumazet void tcp_wfree(struct sk_buff *skb) 79246d3ceabSEric Dumazet { 79346d3ceabSEric Dumazet struct sock *sk = skb->sk; 79446d3ceabSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 79546d3ceabSEric Dumazet 79646d3ceabSEric Dumazet if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) && 79746d3ceabSEric Dumazet !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) { 79846d3ceabSEric Dumazet unsigned long flags; 79946d3ceabSEric Dumazet struct tsq_tasklet *tsq; 80046d3ceabSEric Dumazet 80146d3ceabSEric Dumazet /* Keep a ref on socket. 80246d3ceabSEric Dumazet * This last ref will be released in tcp_tasklet_func() 80346d3ceabSEric Dumazet */ 80446d3ceabSEric Dumazet atomic_sub(skb->truesize - 1, &sk->sk_wmem_alloc); 80546d3ceabSEric Dumazet 80646d3ceabSEric Dumazet /* queue this socket to tasklet queue */ 80746d3ceabSEric Dumazet local_irq_save(flags); 80846d3ceabSEric Dumazet tsq = &__get_cpu_var(tsq_tasklet); 80946d3ceabSEric Dumazet list_add(&tp->tsq_node, &tsq->head); 81046d3ceabSEric Dumazet tasklet_schedule(&tsq->tasklet); 81146d3ceabSEric Dumazet local_irq_restore(flags); 81246d3ceabSEric Dumazet } else { 81346d3ceabSEric Dumazet sock_wfree(skb); 81446d3ceabSEric Dumazet } 81546d3ceabSEric Dumazet } 81646d3ceabSEric Dumazet 8171da177e4SLinus Torvalds /* This routine actually transmits TCP packets queued in by 8181da177e4SLinus Torvalds * tcp_do_sendmsg(). This is used by both the initial 8191da177e4SLinus Torvalds * transmission and possible later retransmissions. 8201da177e4SLinus Torvalds * All SKB's seen here are completely headerless. It is our 8211da177e4SLinus Torvalds * job to build the TCP header, and pass the packet down to 8221da177e4SLinus Torvalds * IP so it can do the same plus pass the packet off to the 8231da177e4SLinus Torvalds * device. 8241da177e4SLinus Torvalds * 8251da177e4SLinus Torvalds * We are working here with either a clone of the original 8261da177e4SLinus Torvalds * SKB, or a fresh unique copy made by the retransmit engine. 8271da177e4SLinus Torvalds */ 828056834d9SIlpo Järvinen static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, 829056834d9SIlpo Järvinen gfp_t gfp_mask) 8301da177e4SLinus Torvalds { 8316687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 832dfb4b9dcSDavid S. Miller struct inet_sock *inet; 833dfb4b9dcSDavid S. Miller struct tcp_sock *tp; 834dfb4b9dcSDavid S. Miller struct tcp_skb_cb *tcb; 83533ad798cSAdam Langley struct tcp_out_options opts; 83695c96174SEric Dumazet unsigned int tcp_options_size, tcp_header_size; 837cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 8381da177e4SLinus Torvalds struct tcphdr *th; 8391da177e4SLinus Torvalds int err; 8401da177e4SLinus Torvalds 841dfb4b9dcSDavid S. Miller BUG_ON(!skb || !tcp_skb_pcount(skb)); 842dfb4b9dcSDavid S. Miller 843dfb4b9dcSDavid S. Miller /* If congestion control is doing timestamping, we must 844dfb4b9dcSDavid S. Miller * take such a timestamp before we potentially clone/copy. 845dfb4b9dcSDavid S. Miller */ 846164891aaSStephen Hemminger if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP) 847dfb4b9dcSDavid S. Miller __net_timestamp(skb); 848dfb4b9dcSDavid S. Miller 849dfb4b9dcSDavid S. Miller if (likely(clone_it)) { 8500e280af0SEric Dumazet const struct sk_buff *fclone = skb + 1; 8510e280af0SEric Dumazet 8520e280af0SEric Dumazet if (unlikely(skb->fclone == SKB_FCLONE_ORIG && 8530e280af0SEric Dumazet fclone->fclone == SKB_FCLONE_CLONE)) 8540e280af0SEric Dumazet NET_INC_STATS_BH(sock_net(sk), 8550e280af0SEric Dumazet LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES); 8560e280af0SEric Dumazet 857dfb4b9dcSDavid S. Miller if (unlikely(skb_cloned(skb))) 858dfb4b9dcSDavid S. Miller skb = pskb_copy(skb, gfp_mask); 859dfb4b9dcSDavid S. Miller else 860dfb4b9dcSDavid S. Miller skb = skb_clone(skb, gfp_mask); 861dfb4b9dcSDavid S. Miller if (unlikely(!skb)) 862dfb4b9dcSDavid S. Miller return -ENOBUFS; 863dfb4b9dcSDavid S. Miller } 864dfb4b9dcSDavid S. Miller 865dfb4b9dcSDavid S. Miller inet = inet_sk(sk); 866dfb4b9dcSDavid S. Miller tp = tcp_sk(sk); 867dfb4b9dcSDavid S. Miller tcb = TCP_SKB_CB(skb); 86833ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 8691da177e4SLinus Torvalds 8704de075e0SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) 87133ad798cSAdam Langley tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); 87233ad798cSAdam Langley else 87333ad798cSAdam Langley tcp_options_size = tcp_established_options(sk, skb, &opts, 87433ad798cSAdam Langley &md5); 87533ad798cSAdam Langley tcp_header_size = tcp_options_size + sizeof(struct tcphdr); 8761da177e4SLinus Torvalds 8773853b584STom Herbert if (tcp_packets_in_flight(tp) == 0) { 8786687e988SArnaldo Carvalho de Melo tcp_ca_event(sk, CA_EVENT_TX_START); 8793853b584STom Herbert skb->ooo_okay = 1; 8803853b584STom Herbert } else 8813853b584STom Herbert skb->ooo_okay = 0; 8821da177e4SLinus Torvalds 883aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 884aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 88546d3ceabSEric Dumazet 88646d3ceabSEric Dumazet skb_orphan(skb); 88746d3ceabSEric Dumazet skb->sk = sk; 88846d3ceabSEric Dumazet skb->destructor = (sysctl_tcp_limit_output_bytes > 0) ? 88946d3ceabSEric Dumazet tcp_wfree : sock_wfree; 89046d3ceabSEric Dumazet atomic_add(skb->truesize, &sk->sk_wmem_alloc); 8911da177e4SLinus Torvalds 8921da177e4SLinus Torvalds /* Build TCP header and checksum it. */ 893aa8223c7SArnaldo Carvalho de Melo th = tcp_hdr(skb); 894c720c7e8SEric Dumazet th->source = inet->inet_sport; 895c720c7e8SEric Dumazet th->dest = inet->inet_dport; 8961da177e4SLinus Torvalds th->seq = htonl(tcb->seq); 8971da177e4SLinus Torvalds th->ack_seq = htonl(tp->rcv_nxt); 898df7a3b07SAl Viro *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | 8994de075e0SEric Dumazet tcb->tcp_flags); 900dfb4b9dcSDavid S. Miller 9014de075e0SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) { 9021da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments 9031da177e4SLinus Torvalds * is never scaled. 9041da177e4SLinus Torvalds */ 905600ff0c2SIlpo Järvinen th->window = htons(min(tp->rcv_wnd, 65535U)); 9061da177e4SLinus Torvalds } else { 9071da177e4SLinus Torvalds th->window = htons(tcp_select_window(sk)); 9081da177e4SLinus Torvalds } 9091da177e4SLinus Torvalds th->check = 0; 9101da177e4SLinus Torvalds th->urg_ptr = 0; 9111da177e4SLinus Torvalds 91233f5f57eSIlpo Järvinen /* The urg_mode check is necessary during a below snd_una win probe */ 9137691367dSHerbert Xu if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { 9147691367dSHerbert Xu if (before(tp->snd_up, tcb->seq + 0x10000)) { 9151da177e4SLinus Torvalds th->urg_ptr = htons(tp->snd_up - tcb->seq); 9161da177e4SLinus Torvalds th->urg = 1; 9177691367dSHerbert Xu } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { 9180eae88f3SEric Dumazet th->urg_ptr = htons(0xFFFF); 9197691367dSHerbert Xu th->urg = 1; 9207691367dSHerbert Xu } 9211da177e4SLinus Torvalds } 9221da177e4SLinus Torvalds 923bd0388aeSWilliam Allen Simpson tcp_options_write((__be32 *)(th + 1), tp, &opts); 9244de075e0SEric Dumazet if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0)) 9259e412ba7SIlpo Järvinen TCP_ECN_send(sk, skb, tcp_header_size); 926dfb4b9dcSDavid S. Miller 927cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 928cfb6eeb4SYOSHIFUJI Hideaki /* Calculate the MD5 hash, as we have all we need now */ 929cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 930a465419bSEric Dumazet sk_nocaps_add(sk, NETIF_F_GSO_MASK); 931bd0388aeSWilliam Allen Simpson tp->af_specific->calc_md5_hash(opts.hash_location, 93249a72dfbSAdam Langley md5, sk, NULL, skb); 933cfb6eeb4SYOSHIFUJI Hideaki } 934cfb6eeb4SYOSHIFUJI Hideaki #endif 935cfb6eeb4SYOSHIFUJI Hideaki 936bb296246SHerbert Xu icsk->icsk_af_ops->send_check(sk, skb); 9371da177e4SLinus Torvalds 9384de075e0SEric Dumazet if (likely(tcb->tcp_flags & TCPHDR_ACK)) 939fc6415bcSDavid S. Miller tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); 9401da177e4SLinus Torvalds 9411da177e4SLinus Torvalds if (skb->len != tcp_header_size) 942cf533ea5SEric Dumazet tcp_event_data_sent(tp, sk); 9431da177e4SLinus Torvalds 944bd37a088SWei Yongjun if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) 945aa2ea058STom Herbert TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, 946aa2ea058STom Herbert tcp_skb_pcount(skb)); 9471da177e4SLinus Torvalds 948d9d8da80SDavid S. Miller err = icsk->icsk_af_ops->queue_xmit(skb, &inet->cork.fl); 94983de47cdSHua Zhong if (likely(err <= 0)) 9501da177e4SLinus Torvalds return err; 9511da177e4SLinus Torvalds 9523cfe3baaSIlpo Järvinen tcp_enter_cwr(sk, 1); 9531da177e4SLinus Torvalds 954b9df3cb8SGerrit Renker return net_xmit_eval(err); 9551da177e4SLinus Torvalds } 9561da177e4SLinus Torvalds 95767edfef7SAndi Kleen /* This routine just queues the buffer for sending. 9581da177e4SLinus Torvalds * 9591da177e4SLinus Torvalds * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, 9601da177e4SLinus Torvalds * otherwise socket can stall. 9611da177e4SLinus Torvalds */ 9621da177e4SLinus Torvalds static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) 9631da177e4SLinus Torvalds { 9641da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 9651da177e4SLinus Torvalds 9661da177e4SLinus Torvalds /* Advance write_seq and place onto the write_queue. */ 9671da177e4SLinus Torvalds tp->write_seq = TCP_SKB_CB(skb)->end_seq; 9681da177e4SLinus Torvalds skb_header_release(skb); 969fe067e8aSDavid S. Miller tcp_add_write_queue_tail(sk, skb); 9703ab224beSHideo Aoki sk->sk_wmem_queued += skb->truesize; 9713ab224beSHideo Aoki sk_mem_charge(sk, skb->truesize); 9721da177e4SLinus Torvalds } 9731da177e4SLinus Torvalds 97467edfef7SAndi Kleen /* Initialize TSO segments for a packet. */ 975cf533ea5SEric Dumazet static void tcp_set_skb_tso_segs(const struct sock *sk, struct sk_buff *skb, 976056834d9SIlpo Järvinen unsigned int mss_now) 977f6302d1dSDavid S. Miller { 9788e5b9ddaSHerbert Xu if (skb->len <= mss_now || !sk_can_gso(sk) || 9798e5b9ddaSHerbert Xu skb->ip_summed == CHECKSUM_NONE) { 980f6302d1dSDavid S. Miller /* Avoid the costly divide in the normal 981f6302d1dSDavid S. Miller * non-TSO case. 982f6302d1dSDavid S. Miller */ 9837967168cSHerbert Xu skb_shinfo(skb)->gso_segs = 1; 9847967168cSHerbert Xu skb_shinfo(skb)->gso_size = 0; 985c9af6db4SPravin B Shelar skb_shinfo(skb)->gso_type = 0; 986f6302d1dSDavid S. Miller } else { 987356f89e1SIlpo Järvinen skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now); 9887967168cSHerbert Xu skb_shinfo(skb)->gso_size = mss_now; 989c9af6db4SPravin B Shelar skb_shinfo(skb)->gso_type = sk->sk_gso_type; 9901da177e4SLinus Torvalds } 9911da177e4SLinus Torvalds } 9921da177e4SLinus Torvalds 99391fed7a1SIlpo Järvinen /* When a modification to fackets out becomes necessary, we need to check 99468f8353bSIlpo Järvinen * skb is counted to fackets_out or not. 99591fed7a1SIlpo Järvinen */ 996cf533ea5SEric Dumazet static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb, 99791fed7a1SIlpo Järvinen int decr) 99891fed7a1SIlpo Järvinen { 999a47e5a98SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1000a47e5a98SIlpo Järvinen 1001dc86967bSIlpo Järvinen if (!tp->sacked_out || tcp_is_reno(tp)) 100291fed7a1SIlpo Järvinen return; 100391fed7a1SIlpo Järvinen 10046859d494SIlpo Järvinen if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq)) 100591fed7a1SIlpo Järvinen tp->fackets_out -= decr; 100691fed7a1SIlpo Järvinen } 100791fed7a1SIlpo Järvinen 1008797108d1SIlpo Järvinen /* Pcount in the middle of the write queue got changed, we need to do various 1009797108d1SIlpo Järvinen * tweaks to fix counters 1010797108d1SIlpo Järvinen */ 1011cf533ea5SEric Dumazet static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) 1012797108d1SIlpo Järvinen { 1013797108d1SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1014797108d1SIlpo Järvinen 1015797108d1SIlpo Järvinen tp->packets_out -= decr; 1016797108d1SIlpo Järvinen 1017797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 1018797108d1SIlpo Järvinen tp->sacked_out -= decr; 1019797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 1020797108d1SIlpo Järvinen tp->retrans_out -= decr; 1021797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 1022797108d1SIlpo Järvinen tp->lost_out -= decr; 1023797108d1SIlpo Järvinen 1024797108d1SIlpo Järvinen /* Reno case is special. Sigh... */ 1025797108d1SIlpo Järvinen if (tcp_is_reno(tp) && decr > 0) 1026797108d1SIlpo Järvinen tp->sacked_out -= min_t(u32, tp->sacked_out, decr); 1027797108d1SIlpo Järvinen 1028797108d1SIlpo Järvinen tcp_adjust_fackets_out(sk, skb, decr); 1029797108d1SIlpo Järvinen 1030797108d1SIlpo Järvinen if (tp->lost_skb_hint && 1031797108d1SIlpo Järvinen before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && 103252cf3cc8SIlpo Järvinen (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))) 1033797108d1SIlpo Järvinen tp->lost_cnt_hint -= decr; 1034797108d1SIlpo Järvinen 1035797108d1SIlpo Järvinen tcp_verify_left_out(tp); 1036797108d1SIlpo Järvinen } 1037797108d1SIlpo Järvinen 10381da177e4SLinus Torvalds /* Function to create two new TCP segments. Shrinks the given segment 10391da177e4SLinus Torvalds * to the specified size and appends a new segment with the rest of the 10401da177e4SLinus Torvalds * packet to the list. This won't be called frequently, I hope. 10411da177e4SLinus Torvalds * Remember, these are still headerless SKBs at this point. 10421da177e4SLinus Torvalds */ 1043056834d9SIlpo Järvinen int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, 1044056834d9SIlpo Järvinen unsigned int mss_now) 10451da177e4SLinus Torvalds { 10461da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 10471da177e4SLinus Torvalds struct sk_buff *buff; 10486475be16SDavid S. Miller int nsize, old_factor; 1049b60b49eaSHerbert Xu int nlen; 10509ce01461SIlpo Järvinen u8 flags; 10511da177e4SLinus Torvalds 10522fceec13SIlpo Järvinen if (WARN_ON(len > skb->len)) 10532fceec13SIlpo Järvinen return -EINVAL; 10546a438bbeSStephen Hemminger 10551da177e4SLinus Torvalds nsize = skb_headlen(skb) - len; 10561da177e4SLinus Torvalds if (nsize < 0) 10571da177e4SLinus Torvalds nsize = 0; 10581da177e4SLinus Torvalds 10591da177e4SLinus Torvalds if (skb_cloned(skb) && 10601da177e4SLinus Torvalds skb_is_nonlinear(skb) && 10611da177e4SLinus Torvalds pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) 10621da177e4SLinus Torvalds return -ENOMEM; 10631da177e4SLinus Torvalds 10641da177e4SLinus Torvalds /* Get a new skb... force flag on. */ 10651da177e4SLinus Torvalds buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC); 10661da177e4SLinus Torvalds if (buff == NULL) 10671da177e4SLinus Torvalds return -ENOMEM; /* We'll just try again later. */ 1068ef5cb973SHerbert Xu 10693ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 10703ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1071b60b49eaSHerbert Xu nlen = skb->len - len - nsize; 1072b60b49eaSHerbert Xu buff->truesize += nlen; 1073b60b49eaSHerbert Xu skb->truesize -= nlen; 10741da177e4SLinus Torvalds 10751da177e4SLinus Torvalds /* Correct the sequence numbers. */ 10761da177e4SLinus Torvalds TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 10771da177e4SLinus Torvalds TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 10781da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 10791da177e4SLinus Torvalds 10801da177e4SLinus Torvalds /* PSH and FIN should only be set in the second packet. */ 10814de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 10824de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 10834de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1084e14c3cafSHerbert Xu TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; 10851da177e4SLinus Torvalds 108684fa7933SPatrick McHardy if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { 10871da177e4SLinus Torvalds /* Copy and checksum data tail into the new buffer. */ 1088056834d9SIlpo Järvinen buff->csum = csum_partial_copy_nocheck(skb->data + len, 1089056834d9SIlpo Järvinen skb_put(buff, nsize), 10901da177e4SLinus Torvalds nsize, 0); 10911da177e4SLinus Torvalds 10921da177e4SLinus Torvalds skb_trim(skb, len); 10931da177e4SLinus Torvalds 10941da177e4SLinus Torvalds skb->csum = csum_block_sub(skb->csum, buff->csum, len); 10951da177e4SLinus Torvalds } else { 109684fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 10971da177e4SLinus Torvalds skb_split(skb, buff, len); 10981da177e4SLinus Torvalds } 10991da177e4SLinus Torvalds 11001da177e4SLinus Torvalds buff->ip_summed = skb->ip_summed; 11011da177e4SLinus Torvalds 11021da177e4SLinus Torvalds /* Looks stupid, but our code really uses when of 11031da177e4SLinus Torvalds * skbs, which it never sent before. --ANK 11041da177e4SLinus Torvalds */ 11051da177e4SLinus Torvalds TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when; 1106a61bbcf2SPatrick McHardy buff->tstamp = skb->tstamp; 11071da177e4SLinus Torvalds 11086475be16SDavid S. Miller old_factor = tcp_skb_pcount(skb); 11096475be16SDavid S. Miller 11101da177e4SLinus Torvalds /* Fix up tso_factor for both original and new SKB. */ 1111846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1112846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, buff, mss_now); 11131da177e4SLinus Torvalds 11146475be16SDavid S. Miller /* If this packet has been sent out already, we must 11156475be16SDavid S. Miller * adjust the various packet counters. 11166475be16SDavid S. Miller */ 1117cf0b450cSHerbert Xu if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 11186475be16SDavid S. Miller int diff = old_factor - tcp_skb_pcount(skb) - 11196475be16SDavid S. Miller tcp_skb_pcount(buff); 11201da177e4SLinus Torvalds 1121797108d1SIlpo Järvinen if (diff) 1122797108d1SIlpo Järvinen tcp_adjust_pcount(sk, skb, diff); 11231da177e4SLinus Torvalds } 11241da177e4SLinus Torvalds 11251da177e4SLinus Torvalds /* Link BUFF into the send queue. */ 1126f44b5271SDavid S. Miller skb_header_release(buff); 1127fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 11281da177e4SLinus Torvalds 11291da177e4SLinus Torvalds return 0; 11301da177e4SLinus Torvalds } 11311da177e4SLinus Torvalds 11321da177e4SLinus Torvalds /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c 11331da177e4SLinus Torvalds * eventually). The difference is that pulled data not copied, but 11341da177e4SLinus Torvalds * immediately discarded. 11351da177e4SLinus Torvalds */ 1136f2911969SHerbert Xu ~{PmVHI~} static void __pskb_trim_head(struct sk_buff *skb, int len) 11371da177e4SLinus Torvalds { 11381da177e4SLinus Torvalds int i, k, eat; 11391da177e4SLinus Torvalds 11404fa48bf3SEric Dumazet eat = min_t(int, len, skb_headlen(skb)); 11414fa48bf3SEric Dumazet if (eat) { 11424fa48bf3SEric Dumazet __skb_pull(skb, eat); 11434fa48bf3SEric Dumazet len -= eat; 11444fa48bf3SEric Dumazet if (!len) 11454fa48bf3SEric Dumazet return; 11464fa48bf3SEric Dumazet } 11471da177e4SLinus Torvalds eat = len; 11481da177e4SLinus Torvalds k = 0; 11491da177e4SLinus Torvalds for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) { 11509e903e08SEric Dumazet int size = skb_frag_size(&skb_shinfo(skb)->frags[i]); 11519e903e08SEric Dumazet 11529e903e08SEric Dumazet if (size <= eat) { 1153aff65da0SIan Campbell skb_frag_unref(skb, i); 11549e903e08SEric Dumazet eat -= size; 11551da177e4SLinus Torvalds } else { 11561da177e4SLinus Torvalds skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i]; 11571da177e4SLinus Torvalds if (eat) { 11581da177e4SLinus Torvalds skb_shinfo(skb)->frags[k].page_offset += eat; 11599e903e08SEric Dumazet skb_frag_size_sub(&skb_shinfo(skb)->frags[k], eat); 11601da177e4SLinus Torvalds eat = 0; 11611da177e4SLinus Torvalds } 11621da177e4SLinus Torvalds k++; 11631da177e4SLinus Torvalds } 11641da177e4SLinus Torvalds } 11651da177e4SLinus Torvalds skb_shinfo(skb)->nr_frags = k; 11661da177e4SLinus Torvalds 116727a884dcSArnaldo Carvalho de Melo skb_reset_tail_pointer(skb); 11681da177e4SLinus Torvalds skb->data_len -= len; 11691da177e4SLinus Torvalds skb->len = skb->data_len; 11701da177e4SLinus Torvalds } 11711da177e4SLinus Torvalds 117267edfef7SAndi Kleen /* Remove acked data from a packet in the transmit queue. */ 11731da177e4SLinus Torvalds int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 11741da177e4SLinus Torvalds { 117514bbd6a5SPravin B Shelar if (skb_unclone(skb, GFP_ATOMIC)) 11761da177e4SLinus Torvalds return -ENOMEM; 11771da177e4SLinus Torvalds 11784fa48bf3SEric Dumazet __pskb_trim_head(skb, len); 11791da177e4SLinus Torvalds 11801da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq += len; 118184fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 11821da177e4SLinus Torvalds 11831da177e4SLinus Torvalds skb->truesize -= len; 11841da177e4SLinus Torvalds sk->sk_wmem_queued -= len; 11853ab224beSHideo Aoki sk_mem_uncharge(sk, len); 11861da177e4SLinus Torvalds sock_set_flag(sk, SOCK_QUEUE_SHRUNK); 11871da177e4SLinus Torvalds 11885b35e1e6SNeal Cardwell /* Any change of skb->len requires recalculation of tso factor. */ 11891da177e4SLinus Torvalds if (tcp_skb_pcount(skb) > 1) 11905b35e1e6SNeal Cardwell tcp_set_skb_tso_segs(sk, skb, tcp_skb_mss(skb)); 11911da177e4SLinus Torvalds 11921da177e4SLinus Torvalds return 0; 11931da177e4SLinus Torvalds } 11941da177e4SLinus Torvalds 11951b63edd6SYuchung Cheng /* Calculate MSS not accounting any TCP options. */ 11961b63edd6SYuchung Cheng static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu) 11975d424d5aSJohn Heffner { 1198cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1199cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 12005d424d5aSJohn Heffner int mss_now; 12015d424d5aSJohn Heffner 12025d424d5aSJohn Heffner /* Calculate base mss without TCP options: 12035d424d5aSJohn Heffner It is MMS_S - sizeof(tcphdr) of rfc1122 12045d424d5aSJohn Heffner */ 12055d424d5aSJohn Heffner mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 12065d424d5aSJohn Heffner 120767469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 120867469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 120967469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 121067469601SEric Dumazet 121167469601SEric Dumazet if (dst && dst_allfrag(dst)) 121267469601SEric Dumazet mss_now -= icsk->icsk_af_ops->net_frag_header_len; 121367469601SEric Dumazet } 121467469601SEric Dumazet 12155d424d5aSJohn Heffner /* Clamp it (mss_clamp does not include tcp options) */ 12165d424d5aSJohn Heffner if (mss_now > tp->rx_opt.mss_clamp) 12175d424d5aSJohn Heffner mss_now = tp->rx_opt.mss_clamp; 12185d424d5aSJohn Heffner 12195d424d5aSJohn Heffner /* Now subtract optional transport overhead */ 12205d424d5aSJohn Heffner mss_now -= icsk->icsk_ext_hdr_len; 12215d424d5aSJohn Heffner 12225d424d5aSJohn Heffner /* Then reserve room for full set of TCP options and 8 bytes of data */ 12235d424d5aSJohn Heffner if (mss_now < 48) 12245d424d5aSJohn Heffner mss_now = 48; 12255d424d5aSJohn Heffner return mss_now; 12265d424d5aSJohn Heffner } 12275d424d5aSJohn Heffner 12281b63edd6SYuchung Cheng /* Calculate MSS. Not accounting for SACKs here. */ 12291b63edd6SYuchung Cheng int tcp_mtu_to_mss(struct sock *sk, int pmtu) 12301b63edd6SYuchung Cheng { 12311b63edd6SYuchung Cheng /* Subtract TCP options size, not including SACKs */ 12321b63edd6SYuchung Cheng return __tcp_mtu_to_mss(sk, pmtu) - 12331b63edd6SYuchung Cheng (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr)); 12341b63edd6SYuchung Cheng } 12351b63edd6SYuchung Cheng 12365d424d5aSJohn Heffner /* Inverse of above */ 123767469601SEric Dumazet int tcp_mss_to_mtu(struct sock *sk, int mss) 12385d424d5aSJohn Heffner { 1239cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1240cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 12415d424d5aSJohn Heffner int mtu; 12425d424d5aSJohn Heffner 12435d424d5aSJohn Heffner mtu = mss + 12445d424d5aSJohn Heffner tp->tcp_header_len + 12455d424d5aSJohn Heffner icsk->icsk_ext_hdr_len + 12465d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 12475d424d5aSJohn Heffner 124867469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 124967469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 125067469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 125167469601SEric Dumazet 125267469601SEric Dumazet if (dst && dst_allfrag(dst)) 125367469601SEric Dumazet mtu += icsk->icsk_af_ops->net_frag_header_len; 125467469601SEric Dumazet } 12555d424d5aSJohn Heffner return mtu; 12565d424d5aSJohn Heffner } 12575d424d5aSJohn Heffner 125867edfef7SAndi Kleen /* MTU probing init per socket */ 12595d424d5aSJohn Heffner void tcp_mtup_init(struct sock *sk) 12605d424d5aSJohn Heffner { 12615d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 12625d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 12635d424d5aSJohn Heffner 12645d424d5aSJohn Heffner icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1; 12655d424d5aSJohn Heffner icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 12665d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 12675d424d5aSJohn Heffner icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss); 12685d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 12695d424d5aSJohn Heffner } 12704bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_mtup_init); 12715d424d5aSJohn Heffner 12721da177e4SLinus Torvalds /* This function synchronize snd mss to current pmtu/exthdr set. 12731da177e4SLinus Torvalds 12741da177e4SLinus Torvalds tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 12751da177e4SLinus Torvalds for TCP options, but includes only bare TCP header. 12761da177e4SLinus Torvalds 12771da177e4SLinus Torvalds tp->rx_opt.mss_clamp is mss negotiated at connection setup. 1278caa20d9aSStephen Hemminger It is minimum of user_mss and mss received with SYN. 12791da177e4SLinus Torvalds It also does not include TCP options. 12801da177e4SLinus Torvalds 1281d83d8461SArnaldo Carvalho de Melo inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. 12821da177e4SLinus Torvalds 12831da177e4SLinus Torvalds tp->mss_cache is current effective sending mss, including 12841da177e4SLinus Torvalds all tcp options except for SACKs. It is evaluated, 12851da177e4SLinus Torvalds taking into account current pmtu, but never exceeds 12861da177e4SLinus Torvalds tp->rx_opt.mss_clamp. 12871da177e4SLinus Torvalds 12881da177e4SLinus Torvalds NOTE1. rfc1122 clearly states that advertised MSS 12891da177e4SLinus Torvalds DOES NOT include either tcp or ip options. 12901da177e4SLinus Torvalds 1291d83d8461SArnaldo Carvalho de Melo NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache 1292d83d8461SArnaldo Carvalho de Melo are READ ONLY outside this function. --ANK (980731) 12931da177e4SLinus Torvalds */ 12941da177e4SLinus Torvalds unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 12951da177e4SLinus Torvalds { 12961da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1297d83d8461SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 12985d424d5aSJohn Heffner int mss_now; 12991da177e4SLinus Torvalds 13005d424d5aSJohn Heffner if (icsk->icsk_mtup.search_high > pmtu) 13015d424d5aSJohn Heffner icsk->icsk_mtup.search_high = pmtu; 13021da177e4SLinus Torvalds 13035d424d5aSJohn Heffner mss_now = tcp_mtu_to_mss(sk, pmtu); 1304409d22b4SIlpo Järvinen mss_now = tcp_bound_to_half_wnd(tp, mss_now); 13051da177e4SLinus Torvalds 13061da177e4SLinus Torvalds /* And store cached results */ 1307d83d8461SArnaldo Carvalho de Melo icsk->icsk_pmtu_cookie = pmtu; 13085d424d5aSJohn Heffner if (icsk->icsk_mtup.enabled) 13095d424d5aSJohn Heffner mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); 1310c1b4a7e6SDavid S. Miller tp->mss_cache = mss_now; 13111da177e4SLinus Torvalds 13121da177e4SLinus Torvalds return mss_now; 13131da177e4SLinus Torvalds } 13144bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_sync_mss); 13151da177e4SLinus Torvalds 13161da177e4SLinus Torvalds /* Compute the current effective MSS, taking SACKs and IP options, 13171da177e4SLinus Torvalds * and even PMTU discovery events into account. 13181da177e4SLinus Torvalds */ 13190c54b85fSIlpo Järvinen unsigned int tcp_current_mss(struct sock *sk) 13201da177e4SLinus Torvalds { 1321cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1322cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1323c1b4a7e6SDavid S. Miller u32 mss_now; 132495c96174SEric Dumazet unsigned int header_len; 132533ad798cSAdam Langley struct tcp_out_options opts; 132633ad798cSAdam Langley struct tcp_md5sig_key *md5; 13271da177e4SLinus Torvalds 1328c1b4a7e6SDavid S. Miller mss_now = tp->mss_cache; 1329c1b4a7e6SDavid S. Miller 13301da177e4SLinus Torvalds if (dst) { 13311da177e4SLinus Torvalds u32 mtu = dst_mtu(dst); 1332d83d8461SArnaldo Carvalho de Melo if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 13331da177e4SLinus Torvalds mss_now = tcp_sync_mss(sk, mtu); 13341da177e4SLinus Torvalds } 13351da177e4SLinus Torvalds 133633ad798cSAdam Langley header_len = tcp_established_options(sk, NULL, &opts, &md5) + 133733ad798cSAdam Langley sizeof(struct tcphdr); 133833ad798cSAdam Langley /* The mss_cache is sized based on tp->tcp_header_len, which assumes 133933ad798cSAdam Langley * some common options. If this is an odd packet (because we have SACK 134033ad798cSAdam Langley * blocks etc) then our calculated header_len will be different, and 134133ad798cSAdam Langley * we have to adjust mss_now correspondingly */ 134233ad798cSAdam Langley if (header_len != tp->tcp_header_len) { 134333ad798cSAdam Langley int delta = (int) header_len - tp->tcp_header_len; 134433ad798cSAdam Langley mss_now -= delta; 134533ad798cSAdam Langley } 1346cfb6eeb4SYOSHIFUJI Hideaki 13471da177e4SLinus Torvalds return mss_now; 13481da177e4SLinus Torvalds } 13491da177e4SLinus Torvalds 1350a762a980SDavid S. Miller /* Congestion window validation. (RFC2861) */ 13519e412ba7SIlpo Järvinen static void tcp_cwnd_validate(struct sock *sk) 1352a762a980SDavid S. Miller { 13539e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1354a762a980SDavid S. Miller 1355d436d686SIlpo Järvinen if (tp->packets_out >= tp->snd_cwnd) { 1356a762a980SDavid S. Miller /* Network is feed fully. */ 1357a762a980SDavid S. Miller tp->snd_cwnd_used = 0; 1358a762a980SDavid S. Miller tp->snd_cwnd_stamp = tcp_time_stamp; 1359a762a980SDavid S. Miller } else { 1360a762a980SDavid S. Miller /* Network starves. */ 1361a762a980SDavid S. Miller if (tp->packets_out > tp->snd_cwnd_used) 1362a762a980SDavid S. Miller tp->snd_cwnd_used = tp->packets_out; 1363a762a980SDavid S. Miller 136415d33c07SDavid S. Miller if (sysctl_tcp_slow_start_after_idle && 136515d33c07SDavid S. Miller (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto) 1366a762a980SDavid S. Miller tcp_cwnd_application_limited(sk); 1367a762a980SDavid S. Miller } 1368a762a980SDavid S. Miller } 1369a762a980SDavid S. Miller 13700e3a4803SIlpo Järvinen /* Returns the portion of skb which can be sent right away without 13710e3a4803SIlpo Järvinen * introducing MSS oddities to segment boundaries. In rare cases where 13720e3a4803SIlpo Järvinen * mss_now != mss_cache, we will request caller to create a small skb 13730e3a4803SIlpo Järvinen * per input skb which could be mostly avoided here (if desired). 13745ea3a748SIlpo Järvinen * 13755ea3a748SIlpo Järvinen * We explicitly want to create a request for splitting write queue tail 13765ea3a748SIlpo Järvinen * to a small skb for Nagle purposes while avoiding unnecessary modulos, 13775ea3a748SIlpo Järvinen * thus all the complexity (cwnd_len is always MSS multiple which we 13785ea3a748SIlpo Järvinen * return whenever allowed by the other factors). Basically we need the 13795ea3a748SIlpo Järvinen * modulo only when the receiver window alone is the limiting factor or 13805ea3a748SIlpo Järvinen * when we would be allowed to send the split-due-to-Nagle skb fully. 13810e3a4803SIlpo Järvinen */ 1382cf533ea5SEric Dumazet static unsigned int tcp_mss_split_point(const struct sock *sk, const struct sk_buff *skb, 13831485348dSBen Hutchings unsigned int mss_now, unsigned int max_segs) 1384c1b4a7e6SDavid S. Miller { 1385cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 13861485348dSBen Hutchings u32 needed, window, max_len; 1387c1b4a7e6SDavid S. Miller 138890840defSIlpo Järvinen window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 13891485348dSBen Hutchings max_len = mss_now * max_segs; 13900e3a4803SIlpo Järvinen 13911485348dSBen Hutchings if (likely(max_len <= window && skb != tcp_write_queue_tail(sk))) 13921485348dSBen Hutchings return max_len; 13930e3a4803SIlpo Järvinen 13945ea3a748SIlpo Järvinen needed = min(skb->len, window); 13955ea3a748SIlpo Järvinen 13961485348dSBen Hutchings if (max_len <= needed) 13971485348dSBen Hutchings return max_len; 13980e3a4803SIlpo Järvinen 13990e3a4803SIlpo Järvinen return needed - needed % mss_now; 1400c1b4a7e6SDavid S. Miller } 1401c1b4a7e6SDavid S. Miller 1402c1b4a7e6SDavid S. Miller /* Can at least one segment of SKB be sent right now, according to the 1403c1b4a7e6SDavid S. Miller * congestion window rules? If so, return how many segments are allowed. 1404c1b4a7e6SDavid S. Miller */ 1405cf533ea5SEric Dumazet static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, 1406cf533ea5SEric Dumazet const struct sk_buff *skb) 1407c1b4a7e6SDavid S. Miller { 1408c1b4a7e6SDavid S. Miller u32 in_flight, cwnd; 1409c1b4a7e6SDavid S. Miller 1410c1b4a7e6SDavid S. Miller /* Don't be strict about the congestion window for the final FIN. */ 14114de075e0SEric Dumazet if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && 14124de075e0SEric Dumazet tcp_skb_pcount(skb) == 1) 1413c1b4a7e6SDavid S. Miller return 1; 1414c1b4a7e6SDavid S. Miller 1415c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1416c1b4a7e6SDavid S. Miller cwnd = tp->snd_cwnd; 1417c1b4a7e6SDavid S. Miller if (in_flight < cwnd) 1418c1b4a7e6SDavid S. Miller return (cwnd - in_flight); 1419c1b4a7e6SDavid S. Miller 1420c1b4a7e6SDavid S. Miller return 0; 1421c1b4a7e6SDavid S. Miller } 1422c1b4a7e6SDavid S. Miller 1423b595076aSUwe Kleine-König /* Initialize TSO state of a skb. 142467edfef7SAndi Kleen * This must be invoked the first time we consider transmitting 1425c1b4a7e6SDavid S. Miller * SKB onto the wire. 1426c1b4a7e6SDavid S. Miller */ 1427cf533ea5SEric Dumazet static int tcp_init_tso_segs(const struct sock *sk, struct sk_buff *skb, 1428056834d9SIlpo Järvinen unsigned int mss_now) 1429c1b4a7e6SDavid S. Miller { 1430c1b4a7e6SDavid S. Miller int tso_segs = tcp_skb_pcount(skb); 1431c1b4a7e6SDavid S. Miller 1432f8269a49SIlpo Järvinen if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { 1433846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1434c1b4a7e6SDavid S. Miller tso_segs = tcp_skb_pcount(skb); 1435c1b4a7e6SDavid S. Miller } 1436c1b4a7e6SDavid S. Miller return tso_segs; 1437c1b4a7e6SDavid S. Miller } 1438c1b4a7e6SDavid S. Miller 143967edfef7SAndi Kleen /* Minshall's variant of the Nagle send check. */ 1440a2a385d6SEric Dumazet static inline bool tcp_minshall_check(const struct tcp_sock *tp) 1441c1b4a7e6SDavid S. Miller { 1442c1b4a7e6SDavid S. Miller return after(tp->snd_sml, tp->snd_una) && 1443c1b4a7e6SDavid S. Miller !after(tp->snd_sml, tp->snd_nxt); 1444c1b4a7e6SDavid S. Miller } 1445c1b4a7e6SDavid S. Miller 1446a2a385d6SEric Dumazet /* Return false, if packet can be sent now without violation Nagle's rules: 1447c1b4a7e6SDavid S. Miller * 1. It is full sized. 1448c1b4a7e6SDavid S. Miller * 2. Or it contains FIN. (already checked by caller) 14496d67e9beSFeng King * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. 1450c1b4a7e6SDavid S. Miller * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 1451c1b4a7e6SDavid S. Miller * With Minshall's modification: all sent small packets are ACKed. 1452c1b4a7e6SDavid S. Miller */ 1453a2a385d6SEric Dumazet static inline bool tcp_nagle_check(const struct tcp_sock *tp, 1454c1b4a7e6SDavid S. Miller const struct sk_buff *skb, 145595c96174SEric Dumazet unsigned int mss_now, int nonagle) 1456c1b4a7e6SDavid S. Miller { 1457a02cec21SEric Dumazet return skb->len < mss_now && 1458c1b4a7e6SDavid S. Miller ((nonagle & TCP_NAGLE_CORK) || 1459a02cec21SEric Dumazet (!nonagle && tp->packets_out && tcp_minshall_check(tp))); 1460c1b4a7e6SDavid S. Miller } 1461c1b4a7e6SDavid S. Miller 1462a2a385d6SEric Dumazet /* Return true if the Nagle test allows this packet to be 1463c1b4a7e6SDavid S. Miller * sent now. 1464c1b4a7e6SDavid S. Miller */ 1465a2a385d6SEric Dumazet static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, 1466c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1467c1b4a7e6SDavid S. Miller { 1468c1b4a7e6SDavid S. Miller /* Nagle rule does not apply to frames, which sit in the middle of the 1469c1b4a7e6SDavid S. Miller * write_queue (they have no chances to get new data). 1470c1b4a7e6SDavid S. Miller * 1471c1b4a7e6SDavid S. Miller * This is implemented in the callers, where they modify the 'nonagle' 1472c1b4a7e6SDavid S. Miller * argument based upon the location of SKB in the send queue. 1473c1b4a7e6SDavid S. Miller */ 1474c1b4a7e6SDavid S. Miller if (nonagle & TCP_NAGLE_PUSH) 1475a2a385d6SEric Dumazet return true; 1476c1b4a7e6SDavid S. Miller 14779b44190dSYuchung Cheng /* Don't use the nagle rule for urgent data (or for the final FIN). */ 14789b44190dSYuchung Cheng if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) 1479a2a385d6SEric Dumazet return true; 1480c1b4a7e6SDavid S. Miller 1481c1b4a7e6SDavid S. Miller if (!tcp_nagle_check(tp, skb, cur_mss, nonagle)) 1482a2a385d6SEric Dumazet return true; 1483c1b4a7e6SDavid S. Miller 1484a2a385d6SEric Dumazet return false; 1485c1b4a7e6SDavid S. Miller } 1486c1b4a7e6SDavid S. Miller 1487c1b4a7e6SDavid S. Miller /* Does at least the first segment of SKB fit into the send window? */ 1488a2a385d6SEric Dumazet static bool tcp_snd_wnd_test(const struct tcp_sock *tp, 1489a2a385d6SEric Dumazet const struct sk_buff *skb, 1490056834d9SIlpo Järvinen unsigned int cur_mss) 1491c1b4a7e6SDavid S. Miller { 1492c1b4a7e6SDavid S. Miller u32 end_seq = TCP_SKB_CB(skb)->end_seq; 1493c1b4a7e6SDavid S. Miller 1494c1b4a7e6SDavid S. Miller if (skb->len > cur_mss) 1495c1b4a7e6SDavid S. Miller end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 1496c1b4a7e6SDavid S. Miller 149790840defSIlpo Järvinen return !after(end_seq, tcp_wnd_end(tp)); 1498c1b4a7e6SDavid S. Miller } 1499c1b4a7e6SDavid S. Miller 1500fe067e8aSDavid S. Miller /* This checks if the data bearing packet SKB (usually tcp_send_head(sk)) 1501c1b4a7e6SDavid S. Miller * should be put on the wire right now. If so, it returns the number of 1502c1b4a7e6SDavid S. Miller * packets allowed by the congestion window. 1503c1b4a7e6SDavid S. Miller */ 1504cf533ea5SEric Dumazet static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb, 1505c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1506c1b4a7e6SDavid S. Miller { 1507cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1508c1b4a7e6SDavid S. Miller unsigned int cwnd_quota; 1509c1b4a7e6SDavid S. Miller 1510846998aeSDavid S. Miller tcp_init_tso_segs(sk, skb, cur_mss); 1511c1b4a7e6SDavid S. Miller 1512c1b4a7e6SDavid S. Miller if (!tcp_nagle_test(tp, skb, cur_mss, nonagle)) 1513c1b4a7e6SDavid S. Miller return 0; 1514c1b4a7e6SDavid S. Miller 1515c1b4a7e6SDavid S. Miller cwnd_quota = tcp_cwnd_test(tp, skb); 1516056834d9SIlpo Järvinen if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss)) 1517c1b4a7e6SDavid S. Miller cwnd_quota = 0; 1518c1b4a7e6SDavid S. Miller 1519c1b4a7e6SDavid S. Miller return cwnd_quota; 1520c1b4a7e6SDavid S. Miller } 1521c1b4a7e6SDavid S. Miller 152267edfef7SAndi Kleen /* Test if sending is allowed right now. */ 1523a2a385d6SEric Dumazet bool tcp_may_send_now(struct sock *sk) 1524c1b4a7e6SDavid S. Miller { 1525cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1526fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 1527c1b4a7e6SDavid S. Miller 1528a02cec21SEric Dumazet return skb && 15290c54b85fSIlpo Järvinen tcp_snd_test(sk, skb, tcp_current_mss(sk), 1530c1b4a7e6SDavid S. Miller (tcp_skb_is_last(sk, skb) ? 1531a02cec21SEric Dumazet tp->nonagle : TCP_NAGLE_PUSH)); 1532c1b4a7e6SDavid S. Miller } 1533c1b4a7e6SDavid S. Miller 1534c1b4a7e6SDavid S. Miller /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 1535c1b4a7e6SDavid S. Miller * which is put after SKB on the list. It is very much like 1536c1b4a7e6SDavid S. Miller * tcp_fragment() except that it may make several kinds of assumptions 1537c1b4a7e6SDavid S. Miller * in order to speed up the splitting operation. In particular, we 1538c1b4a7e6SDavid S. Miller * know that all the data is in scatter-gather pages, and that the 1539c1b4a7e6SDavid S. Miller * packet has never been sent out before (and thus is not cloned). 1540c1b4a7e6SDavid S. Miller */ 1541056834d9SIlpo Järvinen static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, 1542c4ead4c5SEric Dumazet unsigned int mss_now, gfp_t gfp) 1543c1b4a7e6SDavid S. Miller { 1544c1b4a7e6SDavid S. Miller struct sk_buff *buff; 1545c1b4a7e6SDavid S. Miller int nlen = skb->len - len; 15469ce01461SIlpo Järvinen u8 flags; 1547c1b4a7e6SDavid S. Miller 1548c1b4a7e6SDavid S. Miller /* All of a TSO frame must be composed of paged data. */ 1549c8ac3774SHerbert Xu if (skb->len != skb->data_len) 1550c8ac3774SHerbert Xu return tcp_fragment(sk, skb, len, mss_now); 1551c1b4a7e6SDavid S. Miller 1552c4ead4c5SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, gfp); 1553c1b4a7e6SDavid S. Miller if (unlikely(buff == NULL)) 1554c1b4a7e6SDavid S. Miller return -ENOMEM; 1555c1b4a7e6SDavid S. Miller 15563ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 15573ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1558b60b49eaSHerbert Xu buff->truesize += nlen; 1559c1b4a7e6SDavid S. Miller skb->truesize -= nlen; 1560c1b4a7e6SDavid S. Miller 1561c1b4a7e6SDavid S. Miller /* Correct the sequence numbers. */ 1562c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 1563c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 1564c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 1565c1b4a7e6SDavid S. Miller 1566c1b4a7e6SDavid S. Miller /* PSH and FIN should only be set in the second packet. */ 15674de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 15684de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 15694de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1570c1b4a7e6SDavid S. Miller 1571c1b4a7e6SDavid S. Miller /* This packet was never sent out yet, so no SACK bits. */ 1572c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->sacked = 0; 1573c1b4a7e6SDavid S. Miller 157484fa7933SPatrick McHardy buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; 1575c1b4a7e6SDavid S. Miller skb_split(skb, buff, len); 1576c1b4a7e6SDavid S. Miller 1577c1b4a7e6SDavid S. Miller /* Fix up tso_factor for both original and new SKB. */ 1578846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1579846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, buff, mss_now); 1580c1b4a7e6SDavid S. Miller 1581c1b4a7e6SDavid S. Miller /* Link BUFF into the send queue. */ 1582c1b4a7e6SDavid S. Miller skb_header_release(buff); 1583fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 1584c1b4a7e6SDavid S. Miller 1585c1b4a7e6SDavid S. Miller return 0; 1586c1b4a7e6SDavid S. Miller } 1587c1b4a7e6SDavid S. Miller 1588c1b4a7e6SDavid S. Miller /* Try to defer sending, if possible, in order to minimize the amount 1589c1b4a7e6SDavid S. Miller * of TSO splitting we do. View it as a kind of TSO Nagle test. 1590c1b4a7e6SDavid S. Miller * 1591c1b4a7e6SDavid S. Miller * This algorithm is from John Heffner. 1592c1b4a7e6SDavid S. Miller */ 1593a2a385d6SEric Dumazet static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb) 1594c1b4a7e6SDavid S. Miller { 15959e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 15966687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 1597c1b4a7e6SDavid S. Miller u32 send_win, cong_win, limit, in_flight; 1598ad9f4f50SEric Dumazet int win_divisor; 1599c1b4a7e6SDavid S. Miller 16004de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 1601ae8064acSJohn Heffner goto send_now; 1602c1b4a7e6SDavid S. Miller 16036687e988SArnaldo Carvalho de Melo if (icsk->icsk_ca_state != TCP_CA_Open) 1604ae8064acSJohn Heffner goto send_now; 1605ae8064acSJohn Heffner 1606ae8064acSJohn Heffner /* Defer for less than two clock ticks. */ 1607bd515c3eSIlpo Järvinen if (tp->tso_deferred && 1608a2acde07SIlpo Järvinen (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1) 1609ae8064acSJohn Heffner goto send_now; 1610908a75c1SDavid S. Miller 1611c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1612c1b4a7e6SDavid S. Miller 1613056834d9SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight)); 1614c1b4a7e6SDavid S. Miller 161590840defSIlpo Järvinen send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1616c1b4a7e6SDavid S. Miller 1617c1b4a7e6SDavid S. Miller /* From in_flight test above, we know that cwnd > in_flight. */ 1618c1b4a7e6SDavid S. Miller cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; 1619c1b4a7e6SDavid S. Miller 1620c1b4a7e6SDavid S. Miller limit = min(send_win, cong_win); 1621c1b4a7e6SDavid S. Miller 1622ba244fe9SDavid S. Miller /* If a full-sized TSO skb can be sent, do it. */ 16231485348dSBen Hutchings if (limit >= min_t(unsigned int, sk->sk_gso_max_size, 16241485348dSBen Hutchings sk->sk_gso_max_segs * tp->mss_cache)) 1625ae8064acSJohn Heffner goto send_now; 1626ba244fe9SDavid S. Miller 162762ad2761SIlpo Järvinen /* Middle in queue won't get any more data, full sendable already? */ 162862ad2761SIlpo Järvinen if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) 162962ad2761SIlpo Järvinen goto send_now; 163062ad2761SIlpo Järvinen 1631ad9f4f50SEric Dumazet win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor); 1632ad9f4f50SEric Dumazet if (win_divisor) { 1633c1b4a7e6SDavid S. Miller u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); 1634c1b4a7e6SDavid S. Miller 1635c1b4a7e6SDavid S. Miller /* If at least some fraction of a window is available, 1636c1b4a7e6SDavid S. Miller * just use it. 1637c1b4a7e6SDavid S. Miller */ 1638ad9f4f50SEric Dumazet chunk /= win_divisor; 1639c1b4a7e6SDavid S. Miller if (limit >= chunk) 1640ae8064acSJohn Heffner goto send_now; 1641c1b4a7e6SDavid S. Miller } else { 1642c1b4a7e6SDavid S. Miller /* Different approach, try not to defer past a single 1643c1b4a7e6SDavid S. Miller * ACK. Receiver should ACK every other full sized 1644c1b4a7e6SDavid S. Miller * frame, so if we have space for more than 3 frames 1645c1b4a7e6SDavid S. Miller * then send now. 1646c1b4a7e6SDavid S. Miller */ 16476b5a5c0dSNeal Cardwell if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) 1648ae8064acSJohn Heffner goto send_now; 1649c1b4a7e6SDavid S. Miller } 1650c1b4a7e6SDavid S. Miller 1651f4541d60SEric Dumazet /* Ok, it looks like it is advisable to defer. 1652f4541d60SEric Dumazet * Do not rearm the timer if already set to not break TCP ACK clocking. 1653f4541d60SEric Dumazet */ 1654f4541d60SEric Dumazet if (!tp->tso_deferred) 1655ae8064acSJohn Heffner tp->tso_deferred = 1 | (jiffies << 1); 1656ae8064acSJohn Heffner 1657a2a385d6SEric Dumazet return true; 1658ae8064acSJohn Heffner 1659ae8064acSJohn Heffner send_now: 1660ae8064acSJohn Heffner tp->tso_deferred = 0; 1661a2a385d6SEric Dumazet return false; 1662c1b4a7e6SDavid S. Miller } 1663c1b4a7e6SDavid S. Miller 16645d424d5aSJohn Heffner /* Create a new MTU probe if we are ready. 166567edfef7SAndi Kleen * MTU probe is regularly attempting to increase the path MTU by 166667edfef7SAndi Kleen * deliberately sending larger packets. This discovers routing 166767edfef7SAndi Kleen * changes resulting in larger path MTUs. 166867edfef7SAndi Kleen * 16695d424d5aSJohn Heffner * Returns 0 if we should wait to probe (no cwnd available), 16705d424d5aSJohn Heffner * 1 if a probe was sent, 1671056834d9SIlpo Järvinen * -1 otherwise 1672056834d9SIlpo Järvinen */ 16735d424d5aSJohn Heffner static int tcp_mtu_probe(struct sock *sk) 16745d424d5aSJohn Heffner { 16755d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 16765d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 16775d424d5aSJohn Heffner struct sk_buff *skb, *nskb, *next; 16785d424d5aSJohn Heffner int len; 16795d424d5aSJohn Heffner int probe_size; 168091cc17c0SIlpo Järvinen int size_needed; 16815d424d5aSJohn Heffner int copy; 16825d424d5aSJohn Heffner int mss_now; 16835d424d5aSJohn Heffner 16845d424d5aSJohn Heffner /* Not currently probing/verifying, 16855d424d5aSJohn Heffner * not in recovery, 16865d424d5aSJohn Heffner * have enough cwnd, and 16875d424d5aSJohn Heffner * not SACKing (the variable headers throw things off) */ 16885d424d5aSJohn Heffner if (!icsk->icsk_mtup.enabled || 16895d424d5aSJohn Heffner icsk->icsk_mtup.probe_size || 16905d424d5aSJohn Heffner inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 16915d424d5aSJohn Heffner tp->snd_cwnd < 11 || 1692cabeccbdSIlpo Järvinen tp->rx_opt.num_sacks || tp->rx_opt.dsack) 16935d424d5aSJohn Heffner return -1; 16945d424d5aSJohn Heffner 16955d424d5aSJohn Heffner /* Very simple search strategy: just double the MSS. */ 16960c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 16975d424d5aSJohn Heffner probe_size = 2 * tp->mss_cache; 169891cc17c0SIlpo Järvinen size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; 16995d424d5aSJohn Heffner if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) { 17005d424d5aSJohn Heffner /* TODO: set timer for probe_converge_event */ 17015d424d5aSJohn Heffner return -1; 17025d424d5aSJohn Heffner } 17035d424d5aSJohn Heffner 17045d424d5aSJohn Heffner /* Have enough data in the send queue to probe? */ 17057f9c33e5SIlpo Järvinen if (tp->write_seq - tp->snd_nxt < size_needed) 17065d424d5aSJohn Heffner return -1; 17075d424d5aSJohn Heffner 170891cc17c0SIlpo Järvinen if (tp->snd_wnd < size_needed) 17095d424d5aSJohn Heffner return -1; 171090840defSIlpo Järvinen if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) 17115d424d5aSJohn Heffner return 0; 17125d424d5aSJohn Heffner 1713d67c58e9SIlpo Järvinen /* Do we need to wait to drain cwnd? With none in flight, don't stall */ 1714d67c58e9SIlpo Järvinen if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { 1715d67c58e9SIlpo Järvinen if (!tcp_packets_in_flight(tp)) 17165d424d5aSJohn Heffner return -1; 17175d424d5aSJohn Heffner else 17185d424d5aSJohn Heffner return 0; 17195d424d5aSJohn Heffner } 17205d424d5aSJohn Heffner 17215d424d5aSJohn Heffner /* We're allowed to probe. Build it now. */ 17225d424d5aSJohn Heffner if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL) 17235d424d5aSJohn Heffner return -1; 17243ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 17253ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 17265d424d5aSJohn Heffner 1727fe067e8aSDavid S. Miller skb = tcp_send_head(sk); 17285d424d5aSJohn Heffner 17295d424d5aSJohn Heffner TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 17305d424d5aSJohn Heffner TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 17314de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; 17325d424d5aSJohn Heffner TCP_SKB_CB(nskb)->sacked = 0; 17335d424d5aSJohn Heffner nskb->csum = 0; 173484fa7933SPatrick McHardy nskb->ip_summed = skb->ip_summed; 17355d424d5aSJohn Heffner 173650c4817eSIlpo Järvinen tcp_insert_write_queue_before(nskb, skb, sk); 173750c4817eSIlpo Järvinen 17385d424d5aSJohn Heffner len = 0; 1739234b6860SIlpo Järvinen tcp_for_write_queue_from_safe(skb, next, sk) { 17405d424d5aSJohn Heffner copy = min_t(int, skb->len, probe_size - len); 17415d424d5aSJohn Heffner if (nskb->ip_summed) 17425d424d5aSJohn Heffner skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); 17435d424d5aSJohn Heffner else 17445d424d5aSJohn Heffner nskb->csum = skb_copy_and_csum_bits(skb, 0, 1745056834d9SIlpo Järvinen skb_put(nskb, copy), 1746056834d9SIlpo Järvinen copy, nskb->csum); 17475d424d5aSJohn Heffner 17485d424d5aSJohn Heffner if (skb->len <= copy) { 17495d424d5aSJohn Heffner /* We've eaten all the data from this skb. 17505d424d5aSJohn Heffner * Throw it away. */ 17514de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 1752fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 17533ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 17545d424d5aSJohn Heffner } else { 17554de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & 1756a3433f35SChangli Gao ~(TCPHDR_FIN|TCPHDR_PSH); 17575d424d5aSJohn Heffner if (!skb_shinfo(skb)->nr_frags) { 17585d424d5aSJohn Heffner skb_pull(skb, copy); 175984fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 1760056834d9SIlpo Järvinen skb->csum = csum_partial(skb->data, 1761056834d9SIlpo Järvinen skb->len, 0); 17625d424d5aSJohn Heffner } else { 17635d424d5aSJohn Heffner __pskb_trim_head(skb, copy); 17645d424d5aSJohn Heffner tcp_set_skb_tso_segs(sk, skb, mss_now); 17655d424d5aSJohn Heffner } 17665d424d5aSJohn Heffner TCP_SKB_CB(skb)->seq += copy; 17675d424d5aSJohn Heffner } 17685d424d5aSJohn Heffner 17695d424d5aSJohn Heffner len += copy; 1770234b6860SIlpo Järvinen 1771234b6860SIlpo Järvinen if (len >= probe_size) 1772234b6860SIlpo Järvinen break; 17735d424d5aSJohn Heffner } 17745d424d5aSJohn Heffner tcp_init_tso_segs(sk, nskb, nskb->len); 17755d424d5aSJohn Heffner 17765d424d5aSJohn Heffner /* We're ready to send. If this fails, the probe will 17775d424d5aSJohn Heffner * be resegmented into mss-sized pieces by tcp_write_xmit(). */ 17785d424d5aSJohn Heffner TCP_SKB_CB(nskb)->when = tcp_time_stamp; 17795d424d5aSJohn Heffner if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 17805d424d5aSJohn Heffner /* Decrement cwnd here because we are sending 17815d424d5aSJohn Heffner * effectively two packets. */ 17825d424d5aSJohn Heffner tp->snd_cwnd--; 178366f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, nskb); 17845d424d5aSJohn Heffner 17855d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 17860e7b1368SJohn Heffner tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 17870e7b1368SJohn Heffner tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 17885d424d5aSJohn Heffner 17895d424d5aSJohn Heffner return 1; 17905d424d5aSJohn Heffner } 17915d424d5aSJohn Heffner 17925d424d5aSJohn Heffner return -1; 17935d424d5aSJohn Heffner } 17945d424d5aSJohn Heffner 17951da177e4SLinus Torvalds /* This routine writes packets to the network. It advances the 17961da177e4SLinus Torvalds * send_head. This happens as incoming acks open up the remote 17971da177e4SLinus Torvalds * window for us. 17981da177e4SLinus Torvalds * 1799f8269a49SIlpo Järvinen * LARGESEND note: !tcp_urg_mode is overkill, only frames between 1800f8269a49SIlpo Järvinen * snd_up-64k-mss .. snd_up cannot be large. However, taking into 1801f8269a49SIlpo Järvinen * account rare use of URG, this is not a big flaw. 1802f8269a49SIlpo Järvinen * 18036ba8a3b1SNandita Dukkipati * Send at most one packet when push_one > 0. Temporarily ignore 18046ba8a3b1SNandita Dukkipati * cwnd limit to force at most one packet out when push_one == 2. 18056ba8a3b1SNandita Dukkipati 1806a2a385d6SEric Dumazet * Returns true, if no segments are in flight and we have queued segments, 1807a2a385d6SEric Dumazet * but cannot send anything now because of SWS or another problem. 18081da177e4SLinus Torvalds */ 1809a2a385d6SEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 1810d5dd9175SIlpo Järvinen int push_one, gfp_t gfp) 18111da177e4SLinus Torvalds { 18121da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 181392df7b51SDavid S. Miller struct sk_buff *skb; 1814c1b4a7e6SDavid S. Miller unsigned int tso_segs, sent_pkts; 1815c1b4a7e6SDavid S. Miller int cwnd_quota; 18165d424d5aSJohn Heffner int result; 18171da177e4SLinus Torvalds 1818c1b4a7e6SDavid S. Miller sent_pkts = 0; 18195d424d5aSJohn Heffner 1820d5dd9175SIlpo Järvinen if (!push_one) { 18215d424d5aSJohn Heffner /* Do MTU probing. */ 1822d5dd9175SIlpo Järvinen result = tcp_mtu_probe(sk); 1823d5dd9175SIlpo Järvinen if (!result) { 1824a2a385d6SEric Dumazet return false; 18255d424d5aSJohn Heffner } else if (result > 0) { 18265d424d5aSJohn Heffner sent_pkts = 1; 18275d424d5aSJohn Heffner } 1828d5dd9175SIlpo Järvinen } 18295d424d5aSJohn Heffner 1830fe067e8aSDavid S. Miller while ((skb = tcp_send_head(sk))) { 1831c8ac3774SHerbert Xu unsigned int limit; 1832c8ac3774SHerbert Xu 183346d3ceabSEric Dumazet 1834b68e9f85SHerbert Xu tso_segs = tcp_init_tso_segs(sk, skb, mss_now); 1835c1b4a7e6SDavid S. Miller BUG_ON(!tso_segs); 1836c1b4a7e6SDavid S. Miller 1837ec342325SAndrew Vagin if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) 1838ec342325SAndrew Vagin goto repair; /* Skip network transmission */ 1839ec342325SAndrew Vagin 1840b68e9f85SHerbert Xu cwnd_quota = tcp_cwnd_test(tp, skb); 18416ba8a3b1SNandita Dukkipati if (!cwnd_quota) { 18426ba8a3b1SNandita Dukkipati if (push_one == 2) 18436ba8a3b1SNandita Dukkipati /* Force out a loss probe pkt. */ 18446ba8a3b1SNandita Dukkipati cwnd_quota = 1; 18456ba8a3b1SNandita Dukkipati else 1846b68e9f85SHerbert Xu break; 18476ba8a3b1SNandita Dukkipati } 1848b68e9f85SHerbert Xu 1849b68e9f85SHerbert Xu if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) 1850b68e9f85SHerbert Xu break; 1851b68e9f85SHerbert Xu 1852c1b4a7e6SDavid S. Miller if (tso_segs == 1) { 1853aa93466bSDavid S. Miller if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 1854aa93466bSDavid S. Miller (tcp_skb_is_last(sk, skb) ? 1855aa93466bSDavid S. Miller nonagle : TCP_NAGLE_PUSH)))) 1856aa93466bSDavid S. Miller break; 1857c1b4a7e6SDavid S. Miller } else { 1858d5dd9175SIlpo Järvinen if (!push_one && tcp_tso_should_defer(sk, skb)) 1859aa93466bSDavid S. Miller break; 1860c1b4a7e6SDavid S. Miller } 1861aa93466bSDavid S. Miller 186246d3ceabSEric Dumazet /* TSQ : sk_wmem_alloc accounts skb truesize, 186346d3ceabSEric Dumazet * including skb overhead. But thats OK. 186446d3ceabSEric Dumazet */ 186546d3ceabSEric Dumazet if (atomic_read(&sk->sk_wmem_alloc) >= sysctl_tcp_limit_output_bytes) { 186646d3ceabSEric Dumazet set_bit(TSQ_THROTTLED, &tp->tsq_flags); 186746d3ceabSEric Dumazet break; 186846d3ceabSEric Dumazet } 1869c8ac3774SHerbert Xu limit = mss_now; 1870f8269a49SIlpo Järvinen if (tso_segs > 1 && !tcp_urg_mode(tp)) 18710e3a4803SIlpo Järvinen limit = tcp_mss_split_point(sk, skb, mss_now, 18721485348dSBen Hutchings min_t(unsigned int, 18731485348dSBen Hutchings cwnd_quota, 18741485348dSBen Hutchings sk->sk_gso_max_segs)); 1875c8ac3774SHerbert Xu 1876c8ac3774SHerbert Xu if (skb->len > limit && 1877c4ead4c5SEric Dumazet unlikely(tso_fragment(sk, skb, limit, mss_now, gfp))) 18781da177e4SLinus Torvalds break; 18791da177e4SLinus Torvalds 18801da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 1881c1b4a7e6SDavid S. Miller 1882d5dd9175SIlpo Järvinen if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) 18831da177e4SLinus Torvalds break; 18841da177e4SLinus Torvalds 1885ec342325SAndrew Vagin repair: 18861da177e4SLinus Torvalds /* Advance the send_head. This one is sent out. 18871da177e4SLinus Torvalds * This call will increment packets_out. 18881da177e4SLinus Torvalds */ 188966f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 18901da177e4SLinus Torvalds 18911da177e4SLinus Torvalds tcp_minshall_update(tp, mss_now, skb); 1892a262f0cdSNandita Dukkipati sent_pkts += tcp_skb_pcount(skb); 1893d5dd9175SIlpo Järvinen 1894d5dd9175SIlpo Järvinen if (push_one) 1895d5dd9175SIlpo Järvinen break; 18961da177e4SLinus Torvalds } 18971da177e4SLinus Torvalds 1898aa93466bSDavid S. Miller if (likely(sent_pkts)) { 1899684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 1900684bad11SYuchung Cheng tp->prr_out += sent_pkts; 19016ba8a3b1SNandita Dukkipati 19026ba8a3b1SNandita Dukkipati /* Send one loss probe per tail loss episode. */ 19036ba8a3b1SNandita Dukkipati if (push_one != 2) 19046ba8a3b1SNandita Dukkipati tcp_schedule_loss_probe(sk); 19059e412ba7SIlpo Järvinen tcp_cwnd_validate(sk); 1906a2a385d6SEric Dumazet return false; 19071da177e4SLinus Torvalds } 19086ba8a3b1SNandita Dukkipati return (push_one == 2) || (!tp->packets_out && tcp_send_head(sk)); 19096ba8a3b1SNandita Dukkipati } 19106ba8a3b1SNandita Dukkipati 19116ba8a3b1SNandita Dukkipati bool tcp_schedule_loss_probe(struct sock *sk) 19126ba8a3b1SNandita Dukkipati { 19136ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 19146ba8a3b1SNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 19156ba8a3b1SNandita Dukkipati u32 timeout, tlp_time_stamp, rto_time_stamp; 19166ba8a3b1SNandita Dukkipati u32 rtt = tp->srtt >> 3; 19176ba8a3b1SNandita Dukkipati 19186ba8a3b1SNandita Dukkipati if (WARN_ON(icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS)) 19196ba8a3b1SNandita Dukkipati return false; 19206ba8a3b1SNandita Dukkipati /* No consecutive loss probes. */ 19216ba8a3b1SNandita Dukkipati if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) { 19226ba8a3b1SNandita Dukkipati tcp_rearm_rto(sk); 19236ba8a3b1SNandita Dukkipati return false; 19246ba8a3b1SNandita Dukkipati } 19256ba8a3b1SNandita Dukkipati /* Don't do any loss probe on a Fast Open connection before 3WHS 19266ba8a3b1SNandita Dukkipati * finishes. 19276ba8a3b1SNandita Dukkipati */ 19286ba8a3b1SNandita Dukkipati if (sk->sk_state == TCP_SYN_RECV) 19296ba8a3b1SNandita Dukkipati return false; 19306ba8a3b1SNandita Dukkipati 19316ba8a3b1SNandita Dukkipati /* TLP is only scheduled when next timer event is RTO. */ 19326ba8a3b1SNandita Dukkipati if (icsk->icsk_pending != ICSK_TIME_RETRANS) 19336ba8a3b1SNandita Dukkipati return false; 19346ba8a3b1SNandita Dukkipati 19356ba8a3b1SNandita Dukkipati /* Schedule a loss probe in 2*RTT for SACK capable connections 19366ba8a3b1SNandita Dukkipati * in Open state, that are either limited by cwnd or application. 19376ba8a3b1SNandita Dukkipati */ 19386ba8a3b1SNandita Dukkipati if (sysctl_tcp_early_retrans < 3 || !rtt || !tp->packets_out || 19396ba8a3b1SNandita Dukkipati !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open) 19406ba8a3b1SNandita Dukkipati return false; 19416ba8a3b1SNandita Dukkipati 19426ba8a3b1SNandita Dukkipati if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) && 19436ba8a3b1SNandita Dukkipati tcp_send_head(sk)) 19446ba8a3b1SNandita Dukkipati return false; 19456ba8a3b1SNandita Dukkipati 19466ba8a3b1SNandita Dukkipati /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account 19476ba8a3b1SNandita Dukkipati * for delayed ack when there's one outstanding packet. 19486ba8a3b1SNandita Dukkipati */ 19496ba8a3b1SNandita Dukkipati timeout = rtt << 1; 19506ba8a3b1SNandita Dukkipati if (tp->packets_out == 1) 19516ba8a3b1SNandita Dukkipati timeout = max_t(u32, timeout, 19526ba8a3b1SNandita Dukkipati (rtt + (rtt >> 1) + TCP_DELACK_MAX)); 19536ba8a3b1SNandita Dukkipati timeout = max_t(u32, timeout, msecs_to_jiffies(10)); 19546ba8a3b1SNandita Dukkipati 19556ba8a3b1SNandita Dukkipati /* If RTO is shorter, just schedule TLP in its place. */ 19566ba8a3b1SNandita Dukkipati tlp_time_stamp = tcp_time_stamp + timeout; 19576ba8a3b1SNandita Dukkipati rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout; 19586ba8a3b1SNandita Dukkipati if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) { 19596ba8a3b1SNandita Dukkipati s32 delta = rto_time_stamp - tcp_time_stamp; 19606ba8a3b1SNandita Dukkipati if (delta > 0) 19616ba8a3b1SNandita Dukkipati timeout = delta; 19626ba8a3b1SNandita Dukkipati } 19636ba8a3b1SNandita Dukkipati 19646ba8a3b1SNandita Dukkipati inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, 19656ba8a3b1SNandita Dukkipati TCP_RTO_MAX); 19666ba8a3b1SNandita Dukkipati return true; 19676ba8a3b1SNandita Dukkipati } 19686ba8a3b1SNandita Dukkipati 19696ba8a3b1SNandita Dukkipati /* When probe timeout (PTO) fires, send a new segment if one exists, else 19706ba8a3b1SNandita Dukkipati * retransmit the last segment. 19716ba8a3b1SNandita Dukkipati */ 19726ba8a3b1SNandita Dukkipati void tcp_send_loss_probe(struct sock *sk) 19736ba8a3b1SNandita Dukkipati { 19749b717a8dSNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 19756ba8a3b1SNandita Dukkipati struct sk_buff *skb; 19766ba8a3b1SNandita Dukkipati int pcount; 19776ba8a3b1SNandita Dukkipati int mss = tcp_current_mss(sk); 19786ba8a3b1SNandita Dukkipati int err = -1; 19796ba8a3b1SNandita Dukkipati 19806ba8a3b1SNandita Dukkipati if (tcp_send_head(sk) != NULL) { 19816ba8a3b1SNandita Dukkipati err = tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC); 19826ba8a3b1SNandita Dukkipati goto rearm_timer; 19836ba8a3b1SNandita Dukkipati } 19846ba8a3b1SNandita Dukkipati 19859b717a8dSNandita Dukkipati /* At most one outstanding TLP retransmission. */ 19869b717a8dSNandita Dukkipati if (tp->tlp_high_seq) 19879b717a8dSNandita Dukkipati goto rearm_timer; 19889b717a8dSNandita Dukkipati 19896ba8a3b1SNandita Dukkipati /* Retransmit last segment. */ 19906ba8a3b1SNandita Dukkipati skb = tcp_write_queue_tail(sk); 19916ba8a3b1SNandita Dukkipati if (WARN_ON(!skb)) 19926ba8a3b1SNandita Dukkipati goto rearm_timer; 19936ba8a3b1SNandita Dukkipati 19946ba8a3b1SNandita Dukkipati pcount = tcp_skb_pcount(skb); 19956ba8a3b1SNandita Dukkipati if (WARN_ON(!pcount)) 19966ba8a3b1SNandita Dukkipati goto rearm_timer; 19976ba8a3b1SNandita Dukkipati 19986ba8a3b1SNandita Dukkipati if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) { 19996ba8a3b1SNandita Dukkipati if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss))) 20006ba8a3b1SNandita Dukkipati goto rearm_timer; 20016ba8a3b1SNandita Dukkipati skb = tcp_write_queue_tail(sk); 20026ba8a3b1SNandita Dukkipati } 20036ba8a3b1SNandita Dukkipati 20046ba8a3b1SNandita Dukkipati if (WARN_ON(!skb || !tcp_skb_pcount(skb))) 20056ba8a3b1SNandita Dukkipati goto rearm_timer; 20066ba8a3b1SNandita Dukkipati 20076ba8a3b1SNandita Dukkipati /* Probe with zero data doesn't trigger fast recovery. */ 20086ba8a3b1SNandita Dukkipati if (skb->len > 0) 20096ba8a3b1SNandita Dukkipati err = __tcp_retransmit_skb(sk, skb); 20106ba8a3b1SNandita Dukkipati 20119b717a8dSNandita Dukkipati /* Record snd_nxt for loss detection. */ 20129b717a8dSNandita Dukkipati if (likely(!err)) 20139b717a8dSNandita Dukkipati tp->tlp_high_seq = tp->snd_nxt; 20149b717a8dSNandita Dukkipati 20156ba8a3b1SNandita Dukkipati rearm_timer: 20166ba8a3b1SNandita Dukkipati inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 20176ba8a3b1SNandita Dukkipati inet_csk(sk)->icsk_rto, 20186ba8a3b1SNandita Dukkipati TCP_RTO_MAX); 20196ba8a3b1SNandita Dukkipati 20206ba8a3b1SNandita Dukkipati if (likely(!err)) 20216ba8a3b1SNandita Dukkipati NET_INC_STATS_BH(sock_net(sk), 20226ba8a3b1SNandita Dukkipati LINUX_MIB_TCPLOSSPROBES); 20236ba8a3b1SNandita Dukkipati return; 20241da177e4SLinus Torvalds } 20251da177e4SLinus Torvalds 2026a762a980SDavid S. Miller /* Push out any pending frames which were held back due to 2027a762a980SDavid S. Miller * TCP_CORK or attempt at coalescing tiny packets. 2028a762a980SDavid S. Miller * The socket must be locked by the caller. 2029a762a980SDavid S. Miller */ 20309e412ba7SIlpo Järvinen void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 20319e412ba7SIlpo Järvinen int nonagle) 2032a762a980SDavid S. Miller { 2033726e07a8SIlpo Järvinen /* If we are closed, the bytes will have to remain here. 2034726e07a8SIlpo Järvinen * In time closedown will finish, we empty the write queue and 2035726e07a8SIlpo Järvinen * all will be happy. 2036726e07a8SIlpo Järvinen */ 2037726e07a8SIlpo Järvinen if (unlikely(sk->sk_state == TCP_CLOSE)) 2038726e07a8SIlpo Järvinen return; 2039726e07a8SIlpo Järvinen 204099a1dec7SMel Gorman if (tcp_write_xmit(sk, cur_mss, nonagle, 0, 204199a1dec7SMel Gorman sk_gfp_atomic(sk, GFP_ATOMIC))) 20429e412ba7SIlpo Järvinen tcp_check_probe_timer(sk); 2043a762a980SDavid S. Miller } 2044a762a980SDavid S. Miller 2045c1b4a7e6SDavid S. Miller /* Send _single_ skb sitting at the send head. This function requires 2046c1b4a7e6SDavid S. Miller * true push pending frames to setup probe timer etc. 2047c1b4a7e6SDavid S. Miller */ 2048c1b4a7e6SDavid S. Miller void tcp_push_one(struct sock *sk, unsigned int mss_now) 2049c1b4a7e6SDavid S. Miller { 2050fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 2051c1b4a7e6SDavid S. Miller 2052c1b4a7e6SDavid S. Miller BUG_ON(!skb || skb->len < mss_now); 2053c1b4a7e6SDavid S. Miller 2054d5dd9175SIlpo Järvinen tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); 2055c1b4a7e6SDavid S. Miller } 2056c1b4a7e6SDavid S. Miller 20571da177e4SLinus Torvalds /* This function returns the amount that we can raise the 20581da177e4SLinus Torvalds * usable window based on the following constraints 20591da177e4SLinus Torvalds * 20601da177e4SLinus Torvalds * 1. The window can never be shrunk once it is offered (RFC 793) 20611da177e4SLinus Torvalds * 2. We limit memory per socket 20621da177e4SLinus Torvalds * 20631da177e4SLinus Torvalds * RFC 1122: 20641da177e4SLinus Torvalds * "the suggested [SWS] avoidance algorithm for the receiver is to keep 20651da177e4SLinus Torvalds * RECV.NEXT + RCV.WIN fixed until: 20661da177e4SLinus Torvalds * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 20671da177e4SLinus Torvalds * 20681da177e4SLinus Torvalds * i.e. don't raise the right edge of the window until you can raise 20691da177e4SLinus Torvalds * it at least MSS bytes. 20701da177e4SLinus Torvalds * 20711da177e4SLinus Torvalds * Unfortunately, the recommended algorithm breaks header prediction, 20721da177e4SLinus Torvalds * since header prediction assumes th->window stays fixed. 20731da177e4SLinus Torvalds * 20741da177e4SLinus Torvalds * Strictly speaking, keeping th->window fixed violates the receiver 20751da177e4SLinus Torvalds * side SWS prevention criteria. The problem is that under this rule 20761da177e4SLinus Torvalds * a stream of single byte packets will cause the right side of the 20771da177e4SLinus Torvalds * window to always advance by a single byte. 20781da177e4SLinus Torvalds * 20791da177e4SLinus Torvalds * Of course, if the sender implements sender side SWS prevention 20801da177e4SLinus Torvalds * then this will not be a problem. 20811da177e4SLinus Torvalds * 20821da177e4SLinus Torvalds * BSD seems to make the following compromise: 20831da177e4SLinus Torvalds * 20841da177e4SLinus Torvalds * If the free space is less than the 1/4 of the maximum 20851da177e4SLinus Torvalds * space available and the free space is less than 1/2 mss, 20861da177e4SLinus Torvalds * then set the window to 0. 20871da177e4SLinus Torvalds * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 20881da177e4SLinus Torvalds * Otherwise, just prevent the window from shrinking 20891da177e4SLinus Torvalds * and from being larger than the largest representable value. 20901da177e4SLinus Torvalds * 20911da177e4SLinus Torvalds * This prevents incremental opening of the window in the regime 20921da177e4SLinus Torvalds * where TCP is limited by the speed of the reader side taking 20931da177e4SLinus Torvalds * data out of the TCP receive queue. It does nothing about 20941da177e4SLinus Torvalds * those cases where the window is constrained on the sender side 20951da177e4SLinus Torvalds * because the pipeline is full. 20961da177e4SLinus Torvalds * 20971da177e4SLinus Torvalds * BSD also seems to "accidentally" limit itself to windows that are a 20981da177e4SLinus Torvalds * multiple of MSS, at least until the free space gets quite small. 20991da177e4SLinus Torvalds * This would appear to be a side effect of the mbuf implementation. 21001da177e4SLinus Torvalds * Combining these two algorithms results in the observed behavior 21011da177e4SLinus Torvalds * of having a fixed window size at almost all times. 21021da177e4SLinus Torvalds * 21031da177e4SLinus Torvalds * Below we obtain similar behavior by forcing the offered window to 21041da177e4SLinus Torvalds * a multiple of the mss when it is feasible to do so. 21051da177e4SLinus Torvalds * 21061da177e4SLinus Torvalds * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 21071da177e4SLinus Torvalds * Regular options like TIMESTAMP are taken into account. 21081da177e4SLinus Torvalds */ 21091da177e4SLinus Torvalds u32 __tcp_select_window(struct sock *sk) 21101da177e4SLinus Torvalds { 2111463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 21121da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2113caa20d9aSStephen Hemminger /* MSS for the peer's data. Previous versions used mss_clamp 21141da177e4SLinus Torvalds * here. I don't know if the value based on our guesses 21151da177e4SLinus Torvalds * of peer's MSS is better for the performance. It's more correct 21161da177e4SLinus Torvalds * but may be worse for the performance because of rcv_mss 21171da177e4SLinus Torvalds * fluctuations. --SAW 1998/11/1 21181da177e4SLinus Torvalds */ 2119463c84b9SArnaldo Carvalho de Melo int mss = icsk->icsk_ack.rcv_mss; 21201da177e4SLinus Torvalds int free_space = tcp_space(sk); 21211da177e4SLinus Torvalds int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk)); 21221da177e4SLinus Torvalds int window; 21231da177e4SLinus Torvalds 21241da177e4SLinus Torvalds if (mss > full_space) 21251da177e4SLinus Torvalds mss = full_space; 21261da177e4SLinus Torvalds 2127b92edbe0SEric Dumazet if (free_space < (full_space >> 1)) { 2128463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.quick = 0; 21291da177e4SLinus Torvalds 2130180d8cd9SGlauber Costa if (sk_under_memory_pressure(sk)) 2131056834d9SIlpo Järvinen tp->rcv_ssthresh = min(tp->rcv_ssthresh, 2132056834d9SIlpo Järvinen 4U * tp->advmss); 21331da177e4SLinus Torvalds 21341da177e4SLinus Torvalds if (free_space < mss) 21351da177e4SLinus Torvalds return 0; 21361da177e4SLinus Torvalds } 21371da177e4SLinus Torvalds 21381da177e4SLinus Torvalds if (free_space > tp->rcv_ssthresh) 21391da177e4SLinus Torvalds free_space = tp->rcv_ssthresh; 21401da177e4SLinus Torvalds 21411da177e4SLinus Torvalds /* Don't do rounding if we are using window scaling, since the 21421da177e4SLinus Torvalds * scaled window will not line up with the MSS boundary anyway. 21431da177e4SLinus Torvalds */ 21441da177e4SLinus Torvalds window = tp->rcv_wnd; 21451da177e4SLinus Torvalds if (tp->rx_opt.rcv_wscale) { 21461da177e4SLinus Torvalds window = free_space; 21471da177e4SLinus Torvalds 21481da177e4SLinus Torvalds /* Advertise enough space so that it won't get scaled away. 21491da177e4SLinus Torvalds * Import case: prevent zero window announcement if 21501da177e4SLinus Torvalds * 1<<rcv_wscale > mss. 21511da177e4SLinus Torvalds */ 21521da177e4SLinus Torvalds if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window) 21531da177e4SLinus Torvalds window = (((window >> tp->rx_opt.rcv_wscale) + 1) 21541da177e4SLinus Torvalds << tp->rx_opt.rcv_wscale); 21551da177e4SLinus Torvalds } else { 21561da177e4SLinus Torvalds /* Get the largest window that is a nice multiple of mss. 21571da177e4SLinus Torvalds * Window clamp already applied above. 21581da177e4SLinus Torvalds * If our current window offering is within 1 mss of the 21591da177e4SLinus Torvalds * free space we just keep it. This prevents the divide 21601da177e4SLinus Torvalds * and multiply from happening most of the time. 21611da177e4SLinus Torvalds * We also don't do any window rounding when the free space 21621da177e4SLinus Torvalds * is too small. 21631da177e4SLinus Torvalds */ 21641da177e4SLinus Torvalds if (window <= free_space - mss || window > free_space) 21651da177e4SLinus Torvalds window = (free_space / mss) * mss; 216684565070SJohn Heffner else if (mss == full_space && 2167b92edbe0SEric Dumazet free_space > window + (full_space >> 1)) 216884565070SJohn Heffner window = free_space; 21691da177e4SLinus Torvalds } 21701da177e4SLinus Torvalds 21711da177e4SLinus Torvalds return window; 21721da177e4SLinus Torvalds } 21731da177e4SLinus Torvalds 21744a17fc3aSIlpo Järvinen /* Collapses two adjacent SKB's during retransmission. */ 21754a17fc3aSIlpo Järvinen static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) 21761da177e4SLinus Torvalds { 21771da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2178fe067e8aSDavid S. Miller struct sk_buff *next_skb = tcp_write_queue_next(sk, skb); 2179058dc334SIlpo Järvinen int skb_size, next_skb_size; 21801da177e4SLinus Torvalds 2181058dc334SIlpo Järvinen skb_size = skb->len; 2182058dc334SIlpo Järvinen next_skb_size = next_skb->len; 21831da177e4SLinus Torvalds 2184058dc334SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); 21851da177e4SLinus Torvalds 21866859d494SIlpo Järvinen tcp_highest_sack_combine(sk, next_skb, skb); 2187a6963a6bSIlpo Järvinen 2188fe067e8aSDavid S. Miller tcp_unlink_write_queue(next_skb, sk); 21891da177e4SLinus Torvalds 2190058dc334SIlpo Järvinen skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size), 21911a4e2d09SArnaldo Carvalho de Melo next_skb_size); 21921da177e4SLinus Torvalds 219352d570aaSJarek Poplawski if (next_skb->ip_summed == CHECKSUM_PARTIAL) 219452d570aaSJarek Poplawski skb->ip_summed = CHECKSUM_PARTIAL; 21951da177e4SLinus Torvalds 219684fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 21971da177e4SLinus Torvalds skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); 21981da177e4SLinus Torvalds 21991da177e4SLinus Torvalds /* Update sequence range on original skb. */ 22001da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 22011da177e4SLinus Torvalds 2202e6c7d085SIlpo Järvinen /* Merge over control information. This moves PSH/FIN etc. over */ 22034de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; 22041da177e4SLinus Torvalds 22051da177e4SLinus Torvalds /* All done, get rid of second SKB and account for it so 22061da177e4SLinus Torvalds * packet counting does not break. 22071da177e4SLinus Torvalds */ 22084828e7f4SIlpo Järvinen TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; 2209b7689205SIlpo Järvinen 2210b7689205SIlpo Järvinen /* changed transmit queue under us so clear hints */ 2211ef9da47cSIlpo Järvinen tcp_clear_retrans_hints_partial(tp); 2212ef9da47cSIlpo Järvinen if (next_skb == tp->retransmit_skb_hint) 2213ef9da47cSIlpo Järvinen tp->retransmit_skb_hint = skb; 2214b7689205SIlpo Järvinen 2215797108d1SIlpo Järvinen tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); 2216797108d1SIlpo Järvinen 22173ab224beSHideo Aoki sk_wmem_free_skb(sk, next_skb); 22181da177e4SLinus Torvalds } 22191da177e4SLinus Torvalds 222067edfef7SAndi Kleen /* Check if coalescing SKBs is legal. */ 2221a2a385d6SEric Dumazet static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) 22224a17fc3aSIlpo Järvinen { 22234a17fc3aSIlpo Järvinen if (tcp_skb_pcount(skb) > 1) 2224a2a385d6SEric Dumazet return false; 22254a17fc3aSIlpo Järvinen /* TODO: SACK collapsing could be used to remove this condition */ 22264a17fc3aSIlpo Järvinen if (skb_shinfo(skb)->nr_frags != 0) 2227a2a385d6SEric Dumazet return false; 22284a17fc3aSIlpo Järvinen if (skb_cloned(skb)) 2229a2a385d6SEric Dumazet return false; 22304a17fc3aSIlpo Järvinen if (skb == tcp_send_head(sk)) 2231a2a385d6SEric Dumazet return false; 22324a17fc3aSIlpo Järvinen /* Some heurestics for collapsing over SACK'd could be invented */ 22334a17fc3aSIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 2234a2a385d6SEric Dumazet return false; 22354a17fc3aSIlpo Järvinen 2236a2a385d6SEric Dumazet return true; 22374a17fc3aSIlpo Järvinen } 22384a17fc3aSIlpo Järvinen 223967edfef7SAndi Kleen /* Collapse packets in the retransmit queue to make to create 224067edfef7SAndi Kleen * less packets on the wire. This is only done on retransmission. 224167edfef7SAndi Kleen */ 22424a17fc3aSIlpo Järvinen static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, 22434a17fc3aSIlpo Järvinen int space) 22444a17fc3aSIlpo Järvinen { 22454a17fc3aSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 22464a17fc3aSIlpo Järvinen struct sk_buff *skb = to, *tmp; 2247a2a385d6SEric Dumazet bool first = true; 22484a17fc3aSIlpo Järvinen 22494a17fc3aSIlpo Järvinen if (!sysctl_tcp_retrans_collapse) 22504a17fc3aSIlpo Järvinen return; 22514de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 22524a17fc3aSIlpo Järvinen return; 22534a17fc3aSIlpo Järvinen 22544a17fc3aSIlpo Järvinen tcp_for_write_queue_from_safe(skb, tmp, sk) { 22554a17fc3aSIlpo Järvinen if (!tcp_can_collapse(sk, skb)) 22564a17fc3aSIlpo Järvinen break; 22574a17fc3aSIlpo Järvinen 22584a17fc3aSIlpo Järvinen space -= skb->len; 22594a17fc3aSIlpo Järvinen 22604a17fc3aSIlpo Järvinen if (first) { 2261a2a385d6SEric Dumazet first = false; 22624a17fc3aSIlpo Järvinen continue; 22634a17fc3aSIlpo Järvinen } 22644a17fc3aSIlpo Järvinen 22654a17fc3aSIlpo Järvinen if (space < 0) 22664a17fc3aSIlpo Järvinen break; 22674a17fc3aSIlpo Järvinen /* Punt if not enough space exists in the first SKB for 22684a17fc3aSIlpo Järvinen * the data in the second 22694a17fc3aSIlpo Järvinen */ 2270a21d4572SEric Dumazet if (skb->len > skb_availroom(to)) 22714a17fc3aSIlpo Järvinen break; 22724a17fc3aSIlpo Järvinen 22734a17fc3aSIlpo Järvinen if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) 22744a17fc3aSIlpo Järvinen break; 22754a17fc3aSIlpo Järvinen 22764a17fc3aSIlpo Järvinen tcp_collapse_retrans(sk, to); 22774a17fc3aSIlpo Järvinen } 22784a17fc3aSIlpo Järvinen } 22794a17fc3aSIlpo Järvinen 22801da177e4SLinus Torvalds /* This retransmits one SKB. Policy decisions and retransmit queue 22811da177e4SLinus Torvalds * state updates are done by the caller. Returns non-zero if an 22821da177e4SLinus Torvalds * error occurred which prevented the send. 22831da177e4SLinus Torvalds */ 228493b174adSYuchung Cheng int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) 22851da177e4SLinus Torvalds { 22861da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 22875d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 22887d227cd2SSridhar Samudrala unsigned int cur_mss; 22891da177e4SLinus Torvalds 22905d424d5aSJohn Heffner /* Inconslusive MTU probe */ 22915d424d5aSJohn Heffner if (icsk->icsk_mtup.probe_size) { 22925d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 22935d424d5aSJohn Heffner } 22945d424d5aSJohn Heffner 22951da177e4SLinus Torvalds /* Do not sent more than we queued. 1/4 is reserved for possible 2296caa20d9aSStephen Hemminger * copying overhead: fragmentation, tunneling, mangling etc. 22971da177e4SLinus Torvalds */ 22981da177e4SLinus Torvalds if (atomic_read(&sk->sk_wmem_alloc) > 22991da177e4SLinus Torvalds min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf)) 23001da177e4SLinus Torvalds return -EAGAIN; 23011da177e4SLinus Torvalds 23021da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 23031da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 23041da177e4SLinus Torvalds BUG(); 23051da177e4SLinus Torvalds if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 23061da177e4SLinus Torvalds return -ENOMEM; 23071da177e4SLinus Torvalds } 23081da177e4SLinus Torvalds 23097d227cd2SSridhar Samudrala if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 23107d227cd2SSridhar Samudrala return -EHOSTUNREACH; /* Routing failure or similar. */ 23117d227cd2SSridhar Samudrala 23120c54b85fSIlpo Järvinen cur_mss = tcp_current_mss(sk); 23137d227cd2SSridhar Samudrala 23141da177e4SLinus Torvalds /* If receiver has shrunk his window, and skb is out of 23151da177e4SLinus Torvalds * new window, do not retransmit it. The exception is the 23161da177e4SLinus Torvalds * case, when window is shrunk to zero. In this case 23171da177e4SLinus Torvalds * our retransmit serves as a zero window probe. 23181da177e4SLinus Torvalds */ 23199d4fb27dSJoe Perches if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && 23209d4fb27dSJoe Perches TCP_SKB_CB(skb)->seq != tp->snd_una) 23211da177e4SLinus Torvalds return -EAGAIN; 23221da177e4SLinus Torvalds 23231da177e4SLinus Torvalds if (skb->len > cur_mss) { 2324846998aeSDavid S. Miller if (tcp_fragment(sk, skb, cur_mss, cur_mss)) 23251da177e4SLinus Torvalds return -ENOMEM; /* We'll try again later. */ 232602276f3cSIlpo Järvinen } else { 23279eb9362eSIlpo Järvinen int oldpcount = tcp_skb_pcount(skb); 23289eb9362eSIlpo Järvinen 23299eb9362eSIlpo Järvinen if (unlikely(oldpcount > 1)) { 233002276f3cSIlpo Järvinen tcp_init_tso_segs(sk, skb, cur_mss); 23319eb9362eSIlpo Järvinen tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb)); 23329eb9362eSIlpo Järvinen } 23331da177e4SLinus Torvalds } 23341da177e4SLinus Torvalds 23351da177e4SLinus Torvalds tcp_retrans_try_collapse(sk, skb, cur_mss); 23361da177e4SLinus Torvalds 23371da177e4SLinus Torvalds /* Some Solaris stacks overoptimize and ignore the FIN on a 23381da177e4SLinus Torvalds * retransmit when old data is attached. So strip it off 23391da177e4SLinus Torvalds * since it is cheap to do so and saves bytes on the network. 23401da177e4SLinus Torvalds */ 23411da177e4SLinus Torvalds if (skb->len > 0 && 23424de075e0SEric Dumazet (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && 23431da177e4SLinus Torvalds tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) { 23441da177e4SLinus Torvalds if (!pskb_trim(skb, 0)) { 2345e870a8efSIlpo Järvinen /* Reuse, even though it does some unnecessary work */ 2346e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1, 23474de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags); 23481da177e4SLinus Torvalds skb->ip_summed = CHECKSUM_NONE; 23491da177e4SLinus Torvalds } 23501da177e4SLinus Torvalds } 23511da177e4SLinus Torvalds 23521da177e4SLinus Torvalds /* Make a copy, if the first transmission SKB clone we made 23531da177e4SLinus Torvalds * is still in somebody's hands, else make a clone. 23541da177e4SLinus Torvalds */ 23551da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 23561da177e4SLinus Torvalds 235750bceae9SThomas Graf /* make sure skb->data is aligned on arches that require it 235850bceae9SThomas Graf * and check if ack-trimming & collapsing extended the headroom 235950bceae9SThomas Graf * beyond what csum_start can cover. 236050bceae9SThomas Graf */ 236150bceae9SThomas Graf if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) || 236250bceae9SThomas Graf skb_headroom(skb) >= 0xFFFF)) { 2363117632e6SEric Dumazet struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER, 2364117632e6SEric Dumazet GFP_ATOMIC); 236593b174adSYuchung Cheng return nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) : 2366117632e6SEric Dumazet -ENOBUFS; 2367117632e6SEric Dumazet } else { 236893b174adSYuchung Cheng return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 2369117632e6SEric Dumazet } 237093b174adSYuchung Cheng } 237193b174adSYuchung Cheng 237293b174adSYuchung Cheng int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) 237393b174adSYuchung Cheng { 237493b174adSYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 237593b174adSYuchung Cheng int err = __tcp_retransmit_skb(sk, skb); 23761da177e4SLinus Torvalds 23771da177e4SLinus Torvalds if (err == 0) { 23781da177e4SLinus Torvalds /* Update global TCP statistics. */ 237981cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); 23801da177e4SLinus Torvalds 23811da177e4SLinus Torvalds tp->total_retrans++; 23821da177e4SLinus Torvalds 23831da177e4SLinus Torvalds #if FASTRETRANS_DEBUG > 0 23841da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 2385e87cc472SJoe Perches net_dbg_ratelimited("retrans_out leaked\n"); 23861da177e4SLinus Torvalds } 23871da177e4SLinus Torvalds #endif 2388b08d6cb2SIlpo Järvinen if (!tp->retrans_out) 2389b08d6cb2SIlpo Järvinen tp->lost_retrans_low = tp->snd_nxt; 23901da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 23911da177e4SLinus Torvalds tp->retrans_out += tcp_skb_pcount(skb); 23921da177e4SLinus Torvalds 23931da177e4SLinus Torvalds /* Save stamp of the first retransmit. */ 23941da177e4SLinus Torvalds if (!tp->retrans_stamp) 23951da177e4SLinus Torvalds tp->retrans_stamp = TCP_SKB_CB(skb)->when; 23961da177e4SLinus Torvalds 2397c24f691bSYuchung Cheng tp->undo_retrans += tcp_skb_pcount(skb); 23981da177e4SLinus Torvalds 23991da177e4SLinus Torvalds /* snd_nxt is stored to detect loss of retransmitted segment, 24001da177e4SLinus Torvalds * see tcp_input.c tcp_sacktag_write_queue(). 24011da177e4SLinus Torvalds */ 24021da177e4SLinus Torvalds TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt; 24031da177e4SLinus Torvalds } 24041da177e4SLinus Torvalds return err; 24051da177e4SLinus Torvalds } 24061da177e4SLinus Torvalds 240767edfef7SAndi Kleen /* Check if we forward retransmits are possible in the current 240867edfef7SAndi Kleen * window/congestion state. 240967edfef7SAndi Kleen */ 2410a2a385d6SEric Dumazet static bool tcp_can_forward_retransmit(struct sock *sk) 2411b5afe7bcSIlpo Järvinen { 2412b5afe7bcSIlpo Järvinen const struct inet_connection_sock *icsk = inet_csk(sk); 2413cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 2414b5afe7bcSIlpo Järvinen 2415b5afe7bcSIlpo Järvinen /* Forward retransmissions are possible only during Recovery. */ 2416b5afe7bcSIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Recovery) 2417a2a385d6SEric Dumazet return false; 2418b5afe7bcSIlpo Järvinen 2419b5afe7bcSIlpo Järvinen /* No forward retransmissions in Reno are possible. */ 2420b5afe7bcSIlpo Järvinen if (tcp_is_reno(tp)) 2421a2a385d6SEric Dumazet return false; 2422b5afe7bcSIlpo Järvinen 2423b5afe7bcSIlpo Järvinen /* Yeah, we have to make difficult choice between forward transmission 2424b5afe7bcSIlpo Järvinen * and retransmission... Both ways have their merits... 2425b5afe7bcSIlpo Järvinen * 2426b5afe7bcSIlpo Järvinen * For now we do not retransmit anything, while we have some new 2427b5afe7bcSIlpo Järvinen * segments to send. In the other cases, follow rule 3 for 2428b5afe7bcSIlpo Järvinen * NextSeg() specified in RFC3517. 2429b5afe7bcSIlpo Järvinen */ 2430b5afe7bcSIlpo Järvinen 2431b5afe7bcSIlpo Järvinen if (tcp_may_send_now(sk)) 2432a2a385d6SEric Dumazet return false; 2433b5afe7bcSIlpo Järvinen 2434a2a385d6SEric Dumazet return true; 2435b5afe7bcSIlpo Järvinen } 2436b5afe7bcSIlpo Järvinen 24371da177e4SLinus Torvalds /* This gets called after a retransmit timeout, and the initially 24381da177e4SLinus Torvalds * retransmitted data is acknowledged. It tries to continue 24391da177e4SLinus Torvalds * resending the rest of the retransmit queue, until either 24401da177e4SLinus Torvalds * we've sent it all or the congestion window limit is reached. 24411da177e4SLinus Torvalds * If doing SACK, the first ACK which comes back for a timeout 24421da177e4SLinus Torvalds * based retransmit packet might feed us FACK information again. 24431da177e4SLinus Torvalds * If so, we use it to avoid unnecessarily retransmissions. 24441da177e4SLinus Torvalds */ 24451da177e4SLinus Torvalds void tcp_xmit_retransmit_queue(struct sock *sk) 24461da177e4SLinus Torvalds { 24476687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 24481da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 24491da177e4SLinus Torvalds struct sk_buff *skb; 24500e1c54c2SIlpo Järvinen struct sk_buff *hole = NULL; 2451618d9f25SIlpo Järvinen u32 last_lost; 245261eb55f4SIlpo Järvinen int mib_idx; 24530e1c54c2SIlpo Järvinen int fwd_rexmitting = 0; 24546a438bbeSStephen Hemminger 245545e77d31SIlpo Järvinen if (!tp->packets_out) 245645e77d31SIlpo Järvinen return; 245745e77d31SIlpo Järvinen 245808ebd172SIlpo Järvinen if (!tp->lost_out) 245908ebd172SIlpo Järvinen tp->retransmit_high = tp->snd_una; 246008ebd172SIlpo Järvinen 2461618d9f25SIlpo Järvinen if (tp->retransmit_skb_hint) { 24626a438bbeSStephen Hemminger skb = tp->retransmit_skb_hint; 2463618d9f25SIlpo Järvinen last_lost = TCP_SKB_CB(skb)->end_seq; 2464618d9f25SIlpo Järvinen if (after(last_lost, tp->retransmit_high)) 2465618d9f25SIlpo Järvinen last_lost = tp->retransmit_high; 2466618d9f25SIlpo Järvinen } else { 2467fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 2468618d9f25SIlpo Järvinen last_lost = tp->snd_una; 2469618d9f25SIlpo Järvinen } 24701da177e4SLinus Torvalds 2471fe067e8aSDavid S. Miller tcp_for_write_queue_from(skb, sk) { 24721da177e4SLinus Torvalds __u8 sacked = TCP_SKB_CB(skb)->sacked; 24731da177e4SLinus Torvalds 2474fe067e8aSDavid S. Miller if (skb == tcp_send_head(sk)) 2475fe067e8aSDavid S. Miller break; 24766a438bbeSStephen Hemminger /* we could do better than to assign each time */ 24770e1c54c2SIlpo Järvinen if (hole == NULL) 24786a438bbeSStephen Hemminger tp->retransmit_skb_hint = skb; 24796a438bbeSStephen Hemminger 24801da177e4SLinus Torvalds /* Assume this retransmit will generate 24811da177e4SLinus Torvalds * only one packet for congestion window 24821da177e4SLinus Torvalds * calculation purposes. This works because 24831da177e4SLinus Torvalds * tcp_retransmit_skb() will chop up the 24841da177e4SLinus Torvalds * packet to be MSS sized and all the 24851da177e4SLinus Torvalds * packet counting works out. 24861da177e4SLinus Torvalds */ 24871da177e4SLinus Torvalds if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) 24881da177e4SLinus Torvalds return; 24890e1c54c2SIlpo Järvinen 24900e1c54c2SIlpo Järvinen if (fwd_rexmitting) { 24910e1c54c2SIlpo Järvinen begin_fwd: 24920e1c54c2SIlpo Järvinen if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp))) 2493006f582cSIlpo Järvinen break; 24940e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFORWARDRETRANS; 24950e1c54c2SIlpo Järvinen 24960e1c54c2SIlpo Järvinen } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) { 2497618d9f25SIlpo Järvinen tp->retransmit_high = last_lost; 24980e1c54c2SIlpo Järvinen if (!tcp_can_forward_retransmit(sk)) 24990e1c54c2SIlpo Järvinen break; 25000e1c54c2SIlpo Järvinen /* Backtrack if necessary to non-L'ed skb */ 25010e1c54c2SIlpo Järvinen if (hole != NULL) { 25020e1c54c2SIlpo Järvinen skb = hole; 25030e1c54c2SIlpo Järvinen hole = NULL; 25040e1c54c2SIlpo Järvinen } 25050e1c54c2SIlpo Järvinen fwd_rexmitting = 1; 25060e1c54c2SIlpo Järvinen goto begin_fwd; 25070e1c54c2SIlpo Järvinen 25080e1c54c2SIlpo Järvinen } else if (!(sacked & TCPCB_LOST)) { 2509ac11ba75SIlpo Järvinen if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) 25100e1c54c2SIlpo Järvinen hole = skb; 251161eb55f4SIlpo Järvinen continue; 25121da177e4SLinus Torvalds 25130e1c54c2SIlpo Järvinen } else { 2514618d9f25SIlpo Järvinen last_lost = TCP_SKB_CB(skb)->end_seq; 25150e1c54c2SIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Loss) 25160e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFASTRETRANS; 25170e1c54c2SIlpo Järvinen else 25180e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; 25190e1c54c2SIlpo Järvinen } 25200e1c54c2SIlpo Järvinen 25210e1c54c2SIlpo Järvinen if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) 252261eb55f4SIlpo Järvinen continue; 252340b215e5SPavel Emelyanov 252409e9b813SEric Dumazet if (tcp_retransmit_skb(sk, skb)) { 252509e9b813SEric Dumazet NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL); 25261da177e4SLinus Torvalds return; 252709e9b813SEric Dumazet } 2528de0744afSPavel Emelyanov NET_INC_STATS_BH(sock_net(sk), mib_idx); 25291da177e4SLinus Torvalds 2530684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 2531a262f0cdSNandita Dukkipati tp->prr_out += tcp_skb_pcount(skb); 2532a262f0cdSNandita Dukkipati 2533fe067e8aSDavid S. Miller if (skb == tcp_write_queue_head(sk)) 2534463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 25353f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, 25363f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 25371da177e4SLinus Torvalds } 25381da177e4SLinus Torvalds } 25391da177e4SLinus Torvalds 25401da177e4SLinus Torvalds /* Send a fin. The caller locks the socket for us. This cannot be 25411da177e4SLinus Torvalds * allowed to fail queueing a FIN frame under any circumstances. 25421da177e4SLinus Torvalds */ 25431da177e4SLinus Torvalds void tcp_send_fin(struct sock *sk) 25441da177e4SLinus Torvalds { 25451da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2546fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_write_queue_tail(sk); 25471da177e4SLinus Torvalds int mss_now; 25481da177e4SLinus Torvalds 25491da177e4SLinus Torvalds /* Optimization, tack on the FIN if we have a queue of 25501da177e4SLinus Torvalds * unsent frames. But be careful about outgoing SACKS 25511da177e4SLinus Torvalds * and IP options. 25521da177e4SLinus Torvalds */ 25530c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 25541da177e4SLinus Torvalds 2555fe067e8aSDavid S. Miller if (tcp_send_head(sk) != NULL) { 25564de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_FIN; 25571da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq++; 25581da177e4SLinus Torvalds tp->write_seq++; 25591da177e4SLinus Torvalds } else { 25601da177e4SLinus Torvalds /* Socket is locked, keep trying until memory is available. */ 25611da177e4SLinus Torvalds for (;;) { 2562aa133076SWu Fengguang skb = alloc_skb_fclone(MAX_TCP_HEADER, 2563aa133076SWu Fengguang sk->sk_allocation); 25641da177e4SLinus Torvalds if (skb) 25651da177e4SLinus Torvalds break; 25661da177e4SLinus Torvalds yield(); 25671da177e4SLinus Torvalds } 25681da177e4SLinus Torvalds 25691da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 25701da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 25711da177e4SLinus Torvalds /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 2572e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tp->write_seq, 2573a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_FIN); 25741da177e4SLinus Torvalds tcp_queue_skb(sk, skb); 25751da177e4SLinus Torvalds } 25769e412ba7SIlpo Järvinen __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF); 25771da177e4SLinus Torvalds } 25781da177e4SLinus Torvalds 25791da177e4SLinus Torvalds /* We get here when a process closes a file descriptor (either due to 25801da177e4SLinus Torvalds * an explicit close() or as a byproduct of exit()'ing) and there 25811da177e4SLinus Torvalds * was unread data in the receive queue. This behavior is recommended 258265bb723cSGerrit Renker * by RFC 2525, section 2.17. -DaveM 25831da177e4SLinus Torvalds */ 2584dd0fc66fSAl Viro void tcp_send_active_reset(struct sock *sk, gfp_t priority) 25851da177e4SLinus Torvalds { 25861da177e4SLinus Torvalds struct sk_buff *skb; 25871da177e4SLinus Torvalds 25881da177e4SLinus Torvalds /* NOTE: No TCP options attached and we never retransmit this. */ 25891da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, priority); 25901da177e4SLinus Torvalds if (!skb) { 25914e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 25921da177e4SLinus Torvalds return; 25931da177e4SLinus Torvalds } 25941da177e4SLinus Torvalds 25951da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 25961da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 2597e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), 2598a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_RST); 25991da177e4SLinus Torvalds /* Send it off. */ 26001da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2601dfb4b9dcSDavid S. Miller if (tcp_transmit_skb(sk, skb, 0, priority)) 26024e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 260326af65cbSSridhar Samudrala 260481cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); 26051da177e4SLinus Torvalds } 26061da177e4SLinus Torvalds 260767edfef7SAndi Kleen /* Send a crossed SYN-ACK during socket establishment. 260867edfef7SAndi Kleen * WARNING: This routine must only be called when we have already sent 26091da177e4SLinus Torvalds * a SYN packet that crossed the incoming SYN that caused this routine 26101da177e4SLinus Torvalds * to get called. If this assumption fails then the initial rcv_wnd 26111da177e4SLinus Torvalds * and rcv_wscale values will not be correct. 26121da177e4SLinus Torvalds */ 26131da177e4SLinus Torvalds int tcp_send_synack(struct sock *sk) 26141da177e4SLinus Torvalds { 26151da177e4SLinus Torvalds struct sk_buff *skb; 26161da177e4SLinus Torvalds 2617fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 26184de075e0SEric Dumazet if (skb == NULL || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 261991df42beSJoe Perches pr_debug("%s: wrong queue state\n", __func__); 26201da177e4SLinus Torvalds return -EFAULT; 26211da177e4SLinus Torvalds } 26224de075e0SEric Dumazet if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { 26231da177e4SLinus Torvalds if (skb_cloned(skb)) { 26241da177e4SLinus Torvalds struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC); 26251da177e4SLinus Torvalds if (nskb == NULL) 26261da177e4SLinus Torvalds return -ENOMEM; 2627fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 26281da177e4SLinus Torvalds skb_header_release(nskb); 2629fe067e8aSDavid S. Miller __tcp_add_write_queue_head(sk, nskb); 26303ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 26313ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 26323ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 26331da177e4SLinus Torvalds skb = nskb; 26341da177e4SLinus Torvalds } 26351da177e4SLinus Torvalds 26364de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; 26371da177e4SLinus Torvalds TCP_ECN_send_synack(tcp_sk(sk), skb); 26381da177e4SLinus Torvalds } 26391da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2640dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 26411da177e4SLinus Torvalds } 26421da177e4SLinus Torvalds 26434aea39c1SEric Dumazet /** 26444aea39c1SEric Dumazet * tcp_make_synack - Prepare a SYN-ACK. 26454aea39c1SEric Dumazet * sk: listener socket 26464aea39c1SEric Dumazet * dst: dst entry attached to the SYNACK 26474aea39c1SEric Dumazet * req: request_sock pointer 26484aea39c1SEric Dumazet * 26494aea39c1SEric Dumazet * Allocate one skb and build a SYNACK packet. 26504aea39c1SEric Dumazet * @dst is consumed : Caller should not use it again. 26514aea39c1SEric Dumazet */ 26521da177e4SLinus Torvalds struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst, 2653e6b4d113SWilliam Allen Simpson struct request_sock *req, 26548336886fSJerry Chu struct tcp_fastopen_cookie *foc) 26551da177e4SLinus Torvalds { 2656bd0388aeSWilliam Allen Simpson struct tcp_out_options opts; 26572e6599cbSArnaldo Carvalho de Melo struct inet_request_sock *ireq = inet_rsk(req); 26581da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 26591da177e4SLinus Torvalds struct tcphdr *th; 26601da177e4SLinus Torvalds struct sk_buff *skb; 2661cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 2662bd0388aeSWilliam Allen Simpson int tcp_header_size; 2663f5fff5dcSTom Quetchenbach int mss; 26641da177e4SLinus Torvalds 26651a2c6181SChristoph Paasch skb = alloc_skb(MAX_TCP_HEADER + 15, sk_gfp_atomic(sk, GFP_ATOMIC)); 26664aea39c1SEric Dumazet if (unlikely(!skb)) { 26674aea39c1SEric Dumazet dst_release(dst); 26681da177e4SLinus Torvalds return NULL; 26694aea39c1SEric Dumazet } 26701da177e4SLinus Torvalds /* Reserve space for headers. */ 26711da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 26721da177e4SLinus Torvalds 26734aea39c1SEric Dumazet skb_dst_set(skb, dst); 2674ca10b9e9SEric Dumazet security_skb_owned_by(skb, sk); 26751da177e4SLinus Torvalds 26760dbaee3bSDavid S. Miller mss = dst_metric_advmss(dst); 2677f5fff5dcSTom Quetchenbach if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss) 2678f5fff5dcSTom Quetchenbach mss = tp->rx_opt.user_mss; 2679f5fff5dcSTom Quetchenbach 268033ad798cSAdam Langley if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */ 268133ad798cSAdam Langley __u8 rcv_wscale; 268233ad798cSAdam Langley /* Set this up on the first call only */ 268333ad798cSAdam Langley req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW); 2684e88c64f0SHagen Paul Pfeifer 2685e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 2686e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 2687e88c64f0SHagen Paul Pfeifer (req->window_clamp > tcp_full_space(sk) || req->window_clamp == 0)) 2688e88c64f0SHagen Paul Pfeifer req->window_clamp = tcp_full_space(sk); 2689e88c64f0SHagen Paul Pfeifer 269033ad798cSAdam Langley /* tcp_full_space because it is guaranteed to be the first packet */ 269133ad798cSAdam Langley tcp_select_initial_window(tcp_full_space(sk), 2692f5fff5dcSTom Quetchenbach mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0), 269333ad798cSAdam Langley &req->rcv_wnd, 269433ad798cSAdam Langley &req->window_clamp, 269533ad798cSAdam Langley ireq->wscale_ok, 269631d12926Slaurent chavey &rcv_wscale, 269731d12926Slaurent chavey dst_metric(dst, RTAX_INITRWND)); 269833ad798cSAdam Langley ireq->rcv_wscale = rcv_wscale; 269933ad798cSAdam Langley } 2700cfb6eeb4SYOSHIFUJI Hideaki 270133ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 27028b5f12d0SFlorian Westphal #ifdef CONFIG_SYN_COOKIES 27038b5f12d0SFlorian Westphal if (unlikely(req->cookie_ts)) 27048b5f12d0SFlorian Westphal TCP_SKB_CB(skb)->when = cookie_init_timestamp(req); 27058b5f12d0SFlorian Westphal else 27068b5f12d0SFlorian Westphal #endif 270733ad798cSAdam Langley TCP_SKB_CB(skb)->when = tcp_time_stamp; 27081a2c6181SChristoph Paasch tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, &md5, 27091a2c6181SChristoph Paasch foc) + sizeof(*th); 271033ad798cSAdam Langley 2711aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 2712aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 27131da177e4SLinus Torvalds 2714aa8223c7SArnaldo Carvalho de Melo th = tcp_hdr(skb); 27151da177e4SLinus Torvalds memset(th, 0, sizeof(struct tcphdr)); 27161da177e4SLinus Torvalds th->syn = 1; 27171da177e4SLinus Torvalds th->ack = 1; 27181da177e4SLinus Torvalds TCP_ECN_make_synack(req, th); 2719a3116ac5SKOVACS Krisztian th->source = ireq->loc_port; 27202e6599cbSArnaldo Carvalho de Melo th->dest = ireq->rmt_port; 2721e870a8efSIlpo Järvinen /* Setting of flags are superfluous here for callers (and ECE is 2722e870a8efSIlpo Järvinen * not even correctly set) 2723e870a8efSIlpo Järvinen */ 2724e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn, 2725a3433f35SChangli Gao TCPHDR_SYN | TCPHDR_ACK); 27264957faadSWilliam Allen Simpson 27271da177e4SLinus Torvalds th->seq = htonl(TCP_SKB_CB(skb)->seq); 27288336886fSJerry Chu /* XXX data is queued and acked as is. No buffer/window check */ 27298336886fSJerry Chu th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt); 27301da177e4SLinus Torvalds 27311da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 2732600ff0c2SIlpo Järvinen th->window = htons(min(req->rcv_wnd, 65535U)); 2733bd0388aeSWilliam Allen Simpson tcp_options_write((__be32 *)(th + 1), tp, &opts); 27341da177e4SLinus Torvalds th->doff = (tcp_header_size >> 2); 2735aa2ea058STom Herbert TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, tcp_skb_pcount(skb)); 2736cfb6eeb4SYOSHIFUJI Hideaki 2737cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 2738cfb6eeb4SYOSHIFUJI Hideaki /* Okay, we have all we need - do the md5 hash if needed */ 2739cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 2740bd0388aeSWilliam Allen Simpson tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, 274149a72dfbSAdam Langley md5, NULL, req, skb); 2742cfb6eeb4SYOSHIFUJI Hideaki } 2743cfb6eeb4SYOSHIFUJI Hideaki #endif 2744cfb6eeb4SYOSHIFUJI Hideaki 27451da177e4SLinus Torvalds return skb; 27461da177e4SLinus Torvalds } 27474bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_make_synack); 27481da177e4SLinus Torvalds 274967edfef7SAndi Kleen /* Do all connect socket setups that can be done AF independent. */ 2750370816aeSPavel Emelyanov void tcp_connect_init(struct sock *sk) 27511da177e4SLinus Torvalds { 2752cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 27531da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 27541da177e4SLinus Torvalds __u8 rcv_wscale; 27551da177e4SLinus Torvalds 27561da177e4SLinus Torvalds /* We'll fix this up when we get a response from the other end. 27571da177e4SLinus Torvalds * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 27581da177e4SLinus Torvalds */ 27591da177e4SLinus Torvalds tp->tcp_header_len = sizeof(struct tcphdr) + 2760bb5b7c11SDavid S. Miller (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0); 27611da177e4SLinus Torvalds 2762cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 2763cfb6eeb4SYOSHIFUJI Hideaki if (tp->af_specific->md5_lookup(sk, sk) != NULL) 2764cfb6eeb4SYOSHIFUJI Hideaki tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 2765cfb6eeb4SYOSHIFUJI Hideaki #endif 2766cfb6eeb4SYOSHIFUJI Hideaki 27671da177e4SLinus Torvalds /* If user gave his TCP_MAXSEG, record it to clamp */ 27681da177e4SLinus Torvalds if (tp->rx_opt.user_mss) 27691da177e4SLinus Torvalds tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 27701da177e4SLinus Torvalds tp->max_window = 0; 27715d424d5aSJohn Heffner tcp_mtup_init(sk); 27721da177e4SLinus Torvalds tcp_sync_mss(sk, dst_mtu(dst)); 27731da177e4SLinus Torvalds 27741da177e4SLinus Torvalds if (!tp->window_clamp) 27751da177e4SLinus Torvalds tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 27760dbaee3bSDavid S. Miller tp->advmss = dst_metric_advmss(dst); 2777f5fff5dcSTom Quetchenbach if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss) 2778f5fff5dcSTom Quetchenbach tp->advmss = tp->rx_opt.user_mss; 2779f5fff5dcSTom Quetchenbach 27801da177e4SLinus Torvalds tcp_initialize_rcv_mss(sk); 27811da177e4SLinus Torvalds 2782e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 2783e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 2784e88c64f0SHagen Paul Pfeifer (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) 2785e88c64f0SHagen Paul Pfeifer tp->window_clamp = tcp_full_space(sk); 2786e88c64f0SHagen Paul Pfeifer 27871da177e4SLinus Torvalds tcp_select_initial_window(tcp_full_space(sk), 27881da177e4SLinus Torvalds tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 27891da177e4SLinus Torvalds &tp->rcv_wnd, 27901da177e4SLinus Torvalds &tp->window_clamp, 2791bb5b7c11SDavid S. Miller sysctl_tcp_window_scaling, 279231d12926Slaurent chavey &rcv_wscale, 279331d12926Slaurent chavey dst_metric(dst, RTAX_INITRWND)); 27941da177e4SLinus Torvalds 27951da177e4SLinus Torvalds tp->rx_opt.rcv_wscale = rcv_wscale; 27961da177e4SLinus Torvalds tp->rcv_ssthresh = tp->rcv_wnd; 27971da177e4SLinus Torvalds 27981da177e4SLinus Torvalds sk->sk_err = 0; 27991da177e4SLinus Torvalds sock_reset_flag(sk, SOCK_DONE); 28001da177e4SLinus Torvalds tp->snd_wnd = 0; 2801ee7537b6SHantzis Fotis tcp_init_wl(tp, 0); 28021da177e4SLinus Torvalds tp->snd_una = tp->write_seq; 28031da177e4SLinus Torvalds tp->snd_sml = tp->write_seq; 280433f5f57eSIlpo Järvinen tp->snd_up = tp->write_seq; 2805370816aeSPavel Emelyanov tp->snd_nxt = tp->write_seq; 2806ee995283SPavel Emelyanov 2807ee995283SPavel Emelyanov if (likely(!tp->repair)) 28081da177e4SLinus Torvalds tp->rcv_nxt = 0; 2809ee995283SPavel Emelyanov tp->rcv_wup = tp->rcv_nxt; 2810ee995283SPavel Emelyanov tp->copied_seq = tp->rcv_nxt; 28111da177e4SLinus Torvalds 2812463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; 2813463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_retransmits = 0; 28141da177e4SLinus Torvalds tcp_clear_retrans(tp); 28151da177e4SLinus Torvalds } 28161da177e4SLinus Torvalds 2817783237e8SYuchung Cheng static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) 2818783237e8SYuchung Cheng { 2819783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 2820783237e8SYuchung Cheng struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 2821783237e8SYuchung Cheng 2822783237e8SYuchung Cheng tcb->end_seq += skb->len; 2823783237e8SYuchung Cheng skb_header_release(skb); 2824783237e8SYuchung Cheng __tcp_add_write_queue_tail(sk, skb); 2825783237e8SYuchung Cheng sk->sk_wmem_queued += skb->truesize; 2826783237e8SYuchung Cheng sk_mem_charge(sk, skb->truesize); 2827783237e8SYuchung Cheng tp->write_seq = tcb->end_seq; 2828783237e8SYuchung Cheng tp->packets_out += tcp_skb_pcount(skb); 2829783237e8SYuchung Cheng } 2830783237e8SYuchung Cheng 2831783237e8SYuchung Cheng /* Build and send a SYN with data and (cached) Fast Open cookie. However, 2832783237e8SYuchung Cheng * queue a data-only packet after the regular SYN, such that regular SYNs 2833783237e8SYuchung Cheng * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges 2834783237e8SYuchung Cheng * only the SYN sequence, the data are retransmitted in the first ACK. 2835783237e8SYuchung Cheng * If cookie is not cached or other error occurs, falls back to send a 2836783237e8SYuchung Cheng * regular SYN with Fast Open cookie request option. 2837783237e8SYuchung Cheng */ 2838783237e8SYuchung Cheng static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) 2839783237e8SYuchung Cheng { 2840783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 2841783237e8SYuchung Cheng struct tcp_fastopen_request *fo = tp->fastopen_req; 2842aab48743SYuchung Cheng int syn_loss = 0, space, i, err = 0, iovlen = fo->data->msg_iovlen; 2843783237e8SYuchung Cheng struct sk_buff *syn_data = NULL, *data; 2844aab48743SYuchung Cheng unsigned long last_syn_loss = 0; 2845783237e8SYuchung Cheng 284667da22d2SYuchung Cheng tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ 2847aab48743SYuchung Cheng tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie, 2848aab48743SYuchung Cheng &syn_loss, &last_syn_loss); 2849aab48743SYuchung Cheng /* Recurring FO SYN losses: revert to regular handshake temporarily */ 2850aab48743SYuchung Cheng if (syn_loss > 1 && 2851aab48743SYuchung Cheng time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) { 2852aab48743SYuchung Cheng fo->cookie.len = -1; 2853aab48743SYuchung Cheng goto fallback; 2854aab48743SYuchung Cheng } 2855aab48743SYuchung Cheng 285667da22d2SYuchung Cheng if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE) 285767da22d2SYuchung Cheng fo->cookie.len = -1; 285867da22d2SYuchung Cheng else if (fo->cookie.len <= 0) 2859783237e8SYuchung Cheng goto fallback; 2860783237e8SYuchung Cheng 2861783237e8SYuchung Cheng /* MSS for SYN-data is based on cached MSS and bounded by PMTU and 2862783237e8SYuchung Cheng * user-MSS. Reserve maximum option space for middleboxes that add 2863783237e8SYuchung Cheng * private TCP options. The cost is reduced data space in SYN :( 2864783237e8SYuchung Cheng */ 2865783237e8SYuchung Cheng if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp) 2866783237e8SYuchung Cheng tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 28671b63edd6SYuchung Cheng space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) - 2868783237e8SYuchung Cheng MAX_TCP_OPTION_SPACE; 2869783237e8SYuchung Cheng 2870783237e8SYuchung Cheng syn_data = skb_copy_expand(syn, skb_headroom(syn), space, 2871783237e8SYuchung Cheng sk->sk_allocation); 2872783237e8SYuchung Cheng if (syn_data == NULL) 2873783237e8SYuchung Cheng goto fallback; 2874783237e8SYuchung Cheng 2875783237e8SYuchung Cheng for (i = 0; i < iovlen && syn_data->len < space; ++i) { 2876783237e8SYuchung Cheng struct iovec *iov = &fo->data->msg_iov[i]; 2877783237e8SYuchung Cheng unsigned char __user *from = iov->iov_base; 2878783237e8SYuchung Cheng int len = iov->iov_len; 2879783237e8SYuchung Cheng 2880783237e8SYuchung Cheng if (syn_data->len + len > space) 2881783237e8SYuchung Cheng len = space - syn_data->len; 2882783237e8SYuchung Cheng else if (i + 1 == iovlen) 2883783237e8SYuchung Cheng /* No more data pending in inet_wait_for_connect() */ 2884783237e8SYuchung Cheng fo->data = NULL; 2885783237e8SYuchung Cheng 2886783237e8SYuchung Cheng if (skb_add_data(syn_data, from, len)) 2887783237e8SYuchung Cheng goto fallback; 2888783237e8SYuchung Cheng } 2889783237e8SYuchung Cheng 2890783237e8SYuchung Cheng /* Queue a data-only packet after the regular SYN for retransmission */ 2891783237e8SYuchung Cheng data = pskb_copy(syn_data, sk->sk_allocation); 2892783237e8SYuchung Cheng if (data == NULL) 2893783237e8SYuchung Cheng goto fallback; 2894783237e8SYuchung Cheng TCP_SKB_CB(data)->seq++; 2895783237e8SYuchung Cheng TCP_SKB_CB(data)->tcp_flags &= ~TCPHDR_SYN; 2896783237e8SYuchung Cheng TCP_SKB_CB(data)->tcp_flags = (TCPHDR_ACK|TCPHDR_PSH); 2897783237e8SYuchung Cheng tcp_connect_queue_skb(sk, data); 2898783237e8SYuchung Cheng fo->copied = data->len; 2899783237e8SYuchung Cheng 2900783237e8SYuchung Cheng if (tcp_transmit_skb(sk, syn_data, 0, sk->sk_allocation) == 0) { 290167da22d2SYuchung Cheng tp->syn_data = (fo->copied > 0); 2902783237e8SYuchung Cheng NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE); 2903783237e8SYuchung Cheng goto done; 2904783237e8SYuchung Cheng } 2905783237e8SYuchung Cheng syn_data = NULL; 2906783237e8SYuchung Cheng 2907783237e8SYuchung Cheng fallback: 2908783237e8SYuchung Cheng /* Send a regular SYN with Fast Open cookie request option */ 2909783237e8SYuchung Cheng if (fo->cookie.len > 0) 2910783237e8SYuchung Cheng fo->cookie.len = 0; 2911783237e8SYuchung Cheng err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); 2912783237e8SYuchung Cheng if (err) 2913783237e8SYuchung Cheng tp->syn_fastopen = 0; 2914783237e8SYuchung Cheng kfree_skb(syn_data); 2915783237e8SYuchung Cheng done: 2916783237e8SYuchung Cheng fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ 2917783237e8SYuchung Cheng return err; 2918783237e8SYuchung Cheng } 2919783237e8SYuchung Cheng 292067edfef7SAndi Kleen /* Build a SYN and send it off. */ 29211da177e4SLinus Torvalds int tcp_connect(struct sock *sk) 29221da177e4SLinus Torvalds { 29231da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 29241da177e4SLinus Torvalds struct sk_buff *buff; 2925ee586811SEric Paris int err; 29261da177e4SLinus Torvalds 29271da177e4SLinus Torvalds tcp_connect_init(sk); 29281da177e4SLinus Torvalds 29292b916477SAndrey Vagin if (unlikely(tp->repair)) { 29302b916477SAndrey Vagin tcp_finish_connect(sk, NULL); 29312b916477SAndrey Vagin return 0; 29322b916477SAndrey Vagin } 29332b916477SAndrey Vagin 2934d179cd12SDavid S. Miller buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation); 29351da177e4SLinus Torvalds if (unlikely(buff == NULL)) 29361da177e4SLinus Torvalds return -ENOBUFS; 29371da177e4SLinus Torvalds 29381da177e4SLinus Torvalds /* Reserve space for headers. */ 29391da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 29401da177e4SLinus Torvalds 2941a3433f35SChangli Gao tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); 2942783237e8SYuchung Cheng tp->retrans_stamp = TCP_SKB_CB(buff)->when = tcp_time_stamp; 2943783237e8SYuchung Cheng tcp_connect_queue_skb(sk, buff); 2944e870a8efSIlpo Järvinen TCP_ECN_send_syn(sk, buff); 29451da177e4SLinus Torvalds 2946783237e8SYuchung Cheng /* Send off SYN; include data in Fast Open. */ 2947783237e8SYuchung Cheng err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : 2948783237e8SYuchung Cheng tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); 2949ee586811SEric Paris if (err == -ECONNREFUSED) 2950ee586811SEric Paris return err; 2951bd37a088SWei Yongjun 2952bd37a088SWei Yongjun /* We change tp->snd_nxt after the tcp_transmit_skb() call 2953bd37a088SWei Yongjun * in order to make this packet get counted in tcpOutSegs. 2954bd37a088SWei Yongjun */ 2955bd37a088SWei Yongjun tp->snd_nxt = tp->write_seq; 2956bd37a088SWei Yongjun tp->pushed_seq = tp->write_seq; 295781cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); 29581da177e4SLinus Torvalds 29591da177e4SLinus Torvalds /* Timer for repeating the SYN until an answer. */ 29603f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 29613f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 29621da177e4SLinus Torvalds return 0; 29631da177e4SLinus Torvalds } 29644bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_connect); 29651da177e4SLinus Torvalds 29661da177e4SLinus Torvalds /* Send out a delayed ack, the caller does the policy checking 29671da177e4SLinus Torvalds * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 29681da177e4SLinus Torvalds * for details. 29691da177e4SLinus Torvalds */ 29701da177e4SLinus Torvalds void tcp_send_delayed_ack(struct sock *sk) 29711da177e4SLinus Torvalds { 2972463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 2973463c84b9SArnaldo Carvalho de Melo int ato = icsk->icsk_ack.ato; 29741da177e4SLinus Torvalds unsigned long timeout; 29751da177e4SLinus Torvalds 29761da177e4SLinus Torvalds if (ato > TCP_DELACK_MIN) { 2977463c84b9SArnaldo Carvalho de Melo const struct tcp_sock *tp = tcp_sk(sk); 29781da177e4SLinus Torvalds int max_ato = HZ / 2; 29791da177e4SLinus Torvalds 2980056834d9SIlpo Järvinen if (icsk->icsk_ack.pingpong || 2981056834d9SIlpo Järvinen (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 29821da177e4SLinus Torvalds max_ato = TCP_DELACK_MAX; 29831da177e4SLinus Torvalds 29841da177e4SLinus Torvalds /* Slow path, intersegment interval is "high". */ 29851da177e4SLinus Torvalds 29861da177e4SLinus Torvalds /* If some rtt estimate is known, use it to bound delayed ack. 2987463c84b9SArnaldo Carvalho de Melo * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 29881da177e4SLinus Torvalds * directly. 29891da177e4SLinus Torvalds */ 29901da177e4SLinus Torvalds if (tp->srtt) { 29911da177e4SLinus Torvalds int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN); 29921da177e4SLinus Torvalds 29931da177e4SLinus Torvalds if (rtt < max_ato) 29941da177e4SLinus Torvalds max_ato = rtt; 29951da177e4SLinus Torvalds } 29961da177e4SLinus Torvalds 29971da177e4SLinus Torvalds ato = min(ato, max_ato); 29981da177e4SLinus Torvalds } 29991da177e4SLinus Torvalds 30001da177e4SLinus Torvalds /* Stay within the limit we were given */ 30011da177e4SLinus Torvalds timeout = jiffies + ato; 30021da177e4SLinus Torvalds 30031da177e4SLinus Torvalds /* Use new timeout only if there wasn't a older one earlier. */ 3004463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 30051da177e4SLinus Torvalds /* If delack timer was blocked or is about to expire, 30061da177e4SLinus Torvalds * send ACK now. 30071da177e4SLinus Torvalds */ 3008463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.blocked || 3009463c84b9SArnaldo Carvalho de Melo time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 30101da177e4SLinus Torvalds tcp_send_ack(sk); 30111da177e4SLinus Torvalds return; 30121da177e4SLinus Torvalds } 30131da177e4SLinus Torvalds 3014463c84b9SArnaldo Carvalho de Melo if (!time_before(timeout, icsk->icsk_ack.timeout)) 3015463c84b9SArnaldo Carvalho de Melo timeout = icsk->icsk_ack.timeout; 30161da177e4SLinus Torvalds } 3017463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 3018463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.timeout = timeout; 3019463c84b9SArnaldo Carvalho de Melo sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 30201da177e4SLinus Torvalds } 30211da177e4SLinus Torvalds 30221da177e4SLinus Torvalds /* This routine sends an ack and also updates the window. */ 30231da177e4SLinus Torvalds void tcp_send_ack(struct sock *sk) 30241da177e4SLinus Torvalds { 30251da177e4SLinus Torvalds struct sk_buff *buff; 30261da177e4SLinus Torvalds 3027058dc334SIlpo Järvinen /* If we have been reset, we may not send again. */ 3028058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3029058dc334SIlpo Järvinen return; 3030058dc334SIlpo Järvinen 30311da177e4SLinus Torvalds /* We are not putting this on the write queue, so 30321da177e4SLinus Torvalds * tcp_transmit_skb() will set the ownership to this 30331da177e4SLinus Torvalds * sock. 30341da177e4SLinus Torvalds */ 303599a1dec7SMel Gorman buff = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC)); 30361da177e4SLinus Torvalds if (buff == NULL) { 3037463c84b9SArnaldo Carvalho de Melo inet_csk_schedule_ack(sk); 3038463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; 30393f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 30403f421baaSArnaldo Carvalho de Melo TCP_DELACK_MAX, TCP_RTO_MAX); 30411da177e4SLinus Torvalds return; 30421da177e4SLinus Torvalds } 30431da177e4SLinus Torvalds 30441da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 30451da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 3046a3433f35SChangli Gao tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); 30471da177e4SLinus Torvalds 30481da177e4SLinus Torvalds /* Send it off, this clears delayed acks for us. */ 30491da177e4SLinus Torvalds TCP_SKB_CB(buff)->when = tcp_time_stamp; 305099a1dec7SMel Gorman tcp_transmit_skb(sk, buff, 0, sk_gfp_atomic(sk, GFP_ATOMIC)); 30511da177e4SLinus Torvalds } 30521da177e4SLinus Torvalds 30531da177e4SLinus Torvalds /* This routine sends a packet with an out of date sequence 30541da177e4SLinus Torvalds * number. It assumes the other end will try to ack it. 30551da177e4SLinus Torvalds * 30561da177e4SLinus Torvalds * Question: what should we make while urgent mode? 30571da177e4SLinus Torvalds * 4.4BSD forces sending single byte of data. We cannot send 30581da177e4SLinus Torvalds * out of window data, because we have SND.NXT==SND.MAX... 30591da177e4SLinus Torvalds * 30601da177e4SLinus Torvalds * Current solution: to send TWO zero-length segments in urgent mode: 30611da177e4SLinus Torvalds * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 30621da177e4SLinus Torvalds * out-of-date with SND.UNA-1 to probe window. 30631da177e4SLinus Torvalds */ 30641da177e4SLinus Torvalds static int tcp_xmit_probe_skb(struct sock *sk, int urgent) 30651da177e4SLinus Torvalds { 30661da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 30671da177e4SLinus Torvalds struct sk_buff *skb; 30681da177e4SLinus Torvalds 30691da177e4SLinus Torvalds /* We don't queue it, tcp_transmit_skb() sets ownership. */ 307099a1dec7SMel Gorman skb = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC)); 30711da177e4SLinus Torvalds if (skb == NULL) 30721da177e4SLinus Torvalds return -1; 30731da177e4SLinus Torvalds 30741da177e4SLinus Torvalds /* Reserve space for headers and set control bits. */ 30751da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 30761da177e4SLinus Torvalds /* Use a previous sequence. This should cause the other 30771da177e4SLinus Torvalds * end to send an ack. Don't queue or clone SKB, just 30781da177e4SLinus Torvalds * send it. 30791da177e4SLinus Torvalds */ 3080a3433f35SChangli Gao tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); 30811da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 3082dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC); 30831da177e4SLinus Torvalds } 30841da177e4SLinus Torvalds 3085ee995283SPavel Emelyanov void tcp_send_window_probe(struct sock *sk) 3086ee995283SPavel Emelyanov { 3087ee995283SPavel Emelyanov if (sk->sk_state == TCP_ESTABLISHED) { 3088ee995283SPavel Emelyanov tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; 3089c0e88ff0SPavel Emelyanov tcp_sk(sk)->snd_nxt = tcp_sk(sk)->write_seq; 3090ee995283SPavel Emelyanov tcp_xmit_probe_skb(sk, 0); 3091ee995283SPavel Emelyanov } 3092ee995283SPavel Emelyanov } 3093ee995283SPavel Emelyanov 309467edfef7SAndi Kleen /* Initiate keepalive or window probe from timer. */ 30951da177e4SLinus Torvalds int tcp_write_wakeup(struct sock *sk) 30961da177e4SLinus Torvalds { 30971da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 30981da177e4SLinus Torvalds struct sk_buff *skb; 30991da177e4SLinus Torvalds 3100058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3101058dc334SIlpo Järvinen return -1; 3102058dc334SIlpo Järvinen 3103fe067e8aSDavid S. Miller if ((skb = tcp_send_head(sk)) != NULL && 310490840defSIlpo Järvinen before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { 31051da177e4SLinus Torvalds int err; 31060c54b85fSIlpo Järvinen unsigned int mss = tcp_current_mss(sk); 310790840defSIlpo Järvinen unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 31081da177e4SLinus Torvalds 31091da177e4SLinus Torvalds if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 31101da177e4SLinus Torvalds tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 31111da177e4SLinus Torvalds 31121da177e4SLinus Torvalds /* We are probing the opening of a window 31131da177e4SLinus Torvalds * but the window size is != 0 31141da177e4SLinus Torvalds * must have been a result SWS avoidance ( sender ) 31151da177e4SLinus Torvalds */ 31161da177e4SLinus Torvalds if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 31171da177e4SLinus Torvalds skb->len > mss) { 31181da177e4SLinus Torvalds seg_size = min(seg_size, mss); 31194de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 3120846998aeSDavid S. Miller if (tcp_fragment(sk, skb, seg_size, mss)) 31211da177e4SLinus Torvalds return -1; 31221da177e4SLinus Torvalds } else if (!tcp_skb_pcount(skb)) 3123846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss); 31241da177e4SLinus Torvalds 31254de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 31261da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 3127dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 312866f5fe62SIlpo Järvinen if (!err) 312966f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 31301da177e4SLinus Torvalds return err; 31311da177e4SLinus Torvalds } else { 313233f5f57eSIlpo Järvinen if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) 31334828e7f4SIlpo Järvinen tcp_xmit_probe_skb(sk, 1); 31341da177e4SLinus Torvalds return tcp_xmit_probe_skb(sk, 0); 31351da177e4SLinus Torvalds } 31361da177e4SLinus Torvalds } 31371da177e4SLinus Torvalds 31381da177e4SLinus Torvalds /* A window probe timeout has occurred. If window is not closed send 31391da177e4SLinus Torvalds * a partial packet else a zero probe. 31401da177e4SLinus Torvalds */ 31411da177e4SLinus Torvalds void tcp_send_probe0(struct sock *sk) 31421da177e4SLinus Torvalds { 3143463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 31441da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 31451da177e4SLinus Torvalds int err; 31461da177e4SLinus Torvalds 31471da177e4SLinus Torvalds err = tcp_write_wakeup(sk); 31481da177e4SLinus Torvalds 3149fe067e8aSDavid S. Miller if (tp->packets_out || !tcp_send_head(sk)) { 31501da177e4SLinus Torvalds /* Cancel probe timer, if it is not required. */ 31516687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 0; 3152463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff = 0; 31531da177e4SLinus Torvalds return; 31541da177e4SLinus Torvalds } 31551da177e4SLinus Torvalds 31561da177e4SLinus Torvalds if (err <= 0) { 3157463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_backoff < sysctl_tcp_retries2) 3158463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff++; 31596687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out++; 3160463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 31613f421baaSArnaldo Carvalho de Melo min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), 31623f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 31631da177e4SLinus Torvalds } else { 31641da177e4SLinus Torvalds /* If packet was not sent due to local congestion, 31656687e988SArnaldo Carvalho de Melo * do not backoff and do not remember icsk_probes_out. 31661da177e4SLinus Torvalds * Let local senders to fight for local resources. 31671da177e4SLinus Torvalds * 31681da177e4SLinus Torvalds * Use accumulated backoff yet. 31691da177e4SLinus Torvalds */ 31706687e988SArnaldo Carvalho de Melo if (!icsk->icsk_probes_out) 31716687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 1; 3172463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 3173463c84b9SArnaldo Carvalho de Melo min(icsk->icsk_rto << icsk->icsk_backoff, 31743f421baaSArnaldo Carvalho de Melo TCP_RESOURCE_PROBE_INTERVAL), 31753f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 31761da177e4SLinus Torvalds } 31771da177e4SLinus Torvalds } 3178