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 || 836a5dc9e5SEric Dumazet icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) { 84750ea2baSYuchung Cheng tcp_rearm_rto(sk); 851da177e4SLinus Torvalds } 866a5dc9e5SEric 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; 163bcefe17cSCong Wang const struct dst_entry *dst = __sk_dst_get(sk); 1641da177e4SLinus Torvalds 16535089bb2SDavid S. Miller if (sysctl_tcp_slow_start_after_idle && 16635089bb2SDavid S. Miller (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto)) 167463c84b9SArnaldo Carvalho de Melo tcp_cwnd_restart(sk, __sk_dst_get(sk)); 1681da177e4SLinus Torvalds 1691da177e4SLinus Torvalds tp->lsndtime = now; 1701da177e4SLinus Torvalds 1711da177e4SLinus Torvalds /* If it is a reply for ato after last received 1721da177e4SLinus Torvalds * packet, enter pingpong mode. 1731da177e4SLinus Torvalds */ 174bcefe17cSCong Wang if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato && 175bcefe17cSCong Wang (!dst || !dst_metric(dst, RTAX_QUICKACK))) 176463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pingpong = 1; 1771da177e4SLinus Torvalds } 1781da177e4SLinus Torvalds 17967edfef7SAndi Kleen /* Account for an ACK we sent. */ 18040efc6faSStephen Hemminger static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts) 1811da177e4SLinus Torvalds { 182463c84b9SArnaldo Carvalho de Melo tcp_dec_quickack_mode(sk, pkts); 183463c84b9SArnaldo Carvalho de Melo inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); 1841da177e4SLinus Torvalds } 1851da177e4SLinus Torvalds 18685f16525SYuchung Cheng 18785f16525SYuchung Cheng u32 tcp_default_init_rwnd(u32 mss) 18885f16525SYuchung Cheng { 18985f16525SYuchung Cheng /* Initial receive window should be twice of TCP_INIT_CWND to 1909ef71e0cSWeiping Pan * enable proper sending of new unsent data during fast recovery 19185f16525SYuchung Cheng * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a 19285f16525SYuchung Cheng * limit when mss is larger than 1460. 19385f16525SYuchung Cheng */ 19485f16525SYuchung Cheng u32 init_rwnd = TCP_INIT_CWND * 2; 19585f16525SYuchung Cheng 19685f16525SYuchung Cheng if (mss > 1460) 19785f16525SYuchung Cheng init_rwnd = max((1460 * init_rwnd) / mss, 2U); 19885f16525SYuchung Cheng return init_rwnd; 19985f16525SYuchung Cheng } 20085f16525SYuchung Cheng 2011da177e4SLinus Torvalds /* Determine a window scaling and initial window to offer. 2021da177e4SLinus Torvalds * Based on the assumption that the given amount of space 2031da177e4SLinus Torvalds * will be offered. Store the results in the tp structure. 2041da177e4SLinus Torvalds * NOTE: for smooth operation initial space offering should 2051da177e4SLinus Torvalds * be a multiple of mss if possible. We assume here that mss >= 1. 2061da177e4SLinus Torvalds * This MUST be enforced by all callers. 2071da177e4SLinus Torvalds */ 2081da177e4SLinus Torvalds void tcp_select_initial_window(int __space, __u32 mss, 2091da177e4SLinus Torvalds __u32 *rcv_wnd, __u32 *window_clamp, 21031d12926Slaurent chavey int wscale_ok, __u8 *rcv_wscale, 21131d12926Slaurent chavey __u32 init_rcv_wnd) 2121da177e4SLinus Torvalds { 2131da177e4SLinus Torvalds unsigned int space = (__space < 0 ? 0 : __space); 2141da177e4SLinus Torvalds 2151da177e4SLinus Torvalds /* If no clamp set the clamp to the max possible scaled window */ 2161da177e4SLinus Torvalds if (*window_clamp == 0) 2171da177e4SLinus Torvalds (*window_clamp) = (65535 << 14); 2181da177e4SLinus Torvalds space = min(*window_clamp, space); 2191da177e4SLinus Torvalds 2201da177e4SLinus Torvalds /* Quantize space offering to a multiple of mss if possible. */ 2211da177e4SLinus Torvalds if (space > mss) 2221da177e4SLinus Torvalds space = (space / mss) * mss; 2231da177e4SLinus Torvalds 2241da177e4SLinus Torvalds /* NOTE: offering an initial window larger than 32767 22515d99e02SRick Jones * will break some buggy TCP stacks. If the admin tells us 22615d99e02SRick Jones * it is likely we could be speaking with such a buggy stack 22715d99e02SRick Jones * we will truncate our initial window offering to 32K-1 22815d99e02SRick Jones * unless the remote has sent us a window scaling option, 22915d99e02SRick Jones * which we interpret as a sign the remote TCP is not 23015d99e02SRick Jones * misinterpreting the window field as a signed quantity. 2311da177e4SLinus Torvalds */ 23215d99e02SRick Jones if (sysctl_tcp_workaround_signed_windows) 2331da177e4SLinus Torvalds (*rcv_wnd) = min(space, MAX_TCP_WINDOW); 23415d99e02SRick Jones else 23515d99e02SRick Jones (*rcv_wnd) = space; 23615d99e02SRick Jones 2371da177e4SLinus Torvalds (*rcv_wscale) = 0; 2381da177e4SLinus Torvalds if (wscale_ok) { 2391da177e4SLinus Torvalds /* Set window scaling on max possible window 2401da177e4SLinus Torvalds * See RFC1323 for an explanation of the limit to 14 2411da177e4SLinus Torvalds */ 2421da177e4SLinus Torvalds space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max); 243316c1592SStephen Hemminger space = min_t(u32, space, *window_clamp); 2441da177e4SLinus Torvalds while (space > 65535 && (*rcv_wscale) < 14) { 2451da177e4SLinus Torvalds space >>= 1; 2461da177e4SLinus Torvalds (*rcv_wscale)++; 2471da177e4SLinus Torvalds } 2481da177e4SLinus Torvalds } 2491da177e4SLinus Torvalds 2501da177e4SLinus Torvalds if (mss > (1 << *rcv_wscale)) { 25185f16525SYuchung Cheng if (!init_rcv_wnd) /* Use default unless specified otherwise */ 25285f16525SYuchung Cheng init_rcv_wnd = tcp_default_init_rwnd(mss); 253b1afde60SNandita Dukkipati *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss); 2541da177e4SLinus Torvalds } 2551da177e4SLinus Torvalds 2561da177e4SLinus Torvalds /* Set the clamp no higher than max representable value */ 2571da177e4SLinus Torvalds (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp); 2581da177e4SLinus Torvalds } 2594bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_select_initial_window); 2601da177e4SLinus Torvalds 2611da177e4SLinus Torvalds /* Chose a new window to advertise, update state in tcp_sock for the 2621da177e4SLinus Torvalds * socket, and return result with RFC1323 scaling applied. The return 2631da177e4SLinus Torvalds * value can be stuffed directly into th->window for an outgoing 2641da177e4SLinus Torvalds * frame. 2651da177e4SLinus Torvalds */ 26640efc6faSStephen Hemminger static u16 tcp_select_window(struct sock *sk) 2671da177e4SLinus Torvalds { 2681da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2691da177e4SLinus Torvalds u32 cur_win = tcp_receive_window(tp); 2701da177e4SLinus Torvalds u32 new_win = __tcp_select_window(sk); 2711da177e4SLinus Torvalds 2721da177e4SLinus Torvalds /* Never shrink the offered window */ 2731da177e4SLinus Torvalds if (new_win < cur_win) { 2741da177e4SLinus Torvalds /* Danger Will Robinson! 2751da177e4SLinus Torvalds * Don't update rcv_wup/rcv_wnd here or else 2761da177e4SLinus Torvalds * we will not be able to advertise a zero 2771da177e4SLinus Torvalds * window in time. --DaveM 2781da177e4SLinus Torvalds * 2791da177e4SLinus Torvalds * Relax Will Robinson. 2801da177e4SLinus Torvalds */ 281607bfbf2SPatrick McHardy new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale); 2821da177e4SLinus Torvalds } 2831da177e4SLinus Torvalds tp->rcv_wnd = new_win; 2841da177e4SLinus Torvalds tp->rcv_wup = tp->rcv_nxt; 2851da177e4SLinus Torvalds 2861da177e4SLinus Torvalds /* Make sure we do not exceed the maximum possible 2871da177e4SLinus Torvalds * scaled window. 2881da177e4SLinus Torvalds */ 28915d99e02SRick Jones if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows) 2901da177e4SLinus Torvalds new_win = min(new_win, MAX_TCP_WINDOW); 2911da177e4SLinus Torvalds else 2921da177e4SLinus Torvalds new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); 2931da177e4SLinus Torvalds 2941da177e4SLinus Torvalds /* RFC1323 scaling applied */ 2951da177e4SLinus Torvalds new_win >>= tp->rx_opt.rcv_wscale; 2961da177e4SLinus Torvalds 2971da177e4SLinus Torvalds /* If we advertise zero window, disable fast path. */ 2981da177e4SLinus Torvalds if (new_win == 0) 2991da177e4SLinus Torvalds tp->pred_flags = 0; 3001da177e4SLinus Torvalds 3011da177e4SLinus Torvalds return new_win; 3021da177e4SLinus Torvalds } 3031da177e4SLinus Torvalds 30467edfef7SAndi Kleen /* Packet ECN state for a SYN-ACK */ 305cf533ea5SEric Dumazet static inline void TCP_ECN_send_synack(const struct tcp_sock *tp, struct sk_buff *skb) 306bdf1ee5dSIlpo Järvinen { 3074de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR; 308bdf1ee5dSIlpo Järvinen if (!(tp->ecn_flags & TCP_ECN_OK)) 3094de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE; 310bdf1ee5dSIlpo Järvinen } 311bdf1ee5dSIlpo Järvinen 31267edfef7SAndi Kleen /* Packet ECN state for a SYN. */ 313bdf1ee5dSIlpo Järvinen static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb) 314bdf1ee5dSIlpo Järvinen { 315bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 316bdf1ee5dSIlpo Järvinen 317bdf1ee5dSIlpo Järvinen tp->ecn_flags = 0; 3185d134f1cSHannes Frederic Sowa if (sock_net(sk)->ipv4.sysctl_tcp_ecn == 1) { 3194de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR; 320bdf1ee5dSIlpo Järvinen tp->ecn_flags = TCP_ECN_OK; 321bdf1ee5dSIlpo Järvinen } 322bdf1ee5dSIlpo Järvinen } 323bdf1ee5dSIlpo Järvinen 324bdf1ee5dSIlpo Järvinen static __inline__ void 325cf533ea5SEric Dumazet TCP_ECN_make_synack(const struct request_sock *req, struct tcphdr *th) 326bdf1ee5dSIlpo Järvinen { 327bdf1ee5dSIlpo Järvinen if (inet_rsk(req)->ecn_ok) 328bdf1ee5dSIlpo Järvinen th->ece = 1; 329bdf1ee5dSIlpo Järvinen } 330bdf1ee5dSIlpo Järvinen 33167edfef7SAndi Kleen /* Set up ECN state for a packet on a ESTABLISHED socket that is about to 33267edfef7SAndi Kleen * be sent. 33367edfef7SAndi Kleen */ 334bdf1ee5dSIlpo Järvinen static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb, 335bdf1ee5dSIlpo Järvinen int tcp_header_len) 336bdf1ee5dSIlpo Järvinen { 337bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 338bdf1ee5dSIlpo Järvinen 339bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_OK) { 340bdf1ee5dSIlpo Järvinen /* Not-retransmitted data segment: set ECT and inject CWR. */ 341bdf1ee5dSIlpo Järvinen if (skb->len != tcp_header_len && 342bdf1ee5dSIlpo Järvinen !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { 343bdf1ee5dSIlpo Järvinen INET_ECN_xmit(sk); 344bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) { 345bdf1ee5dSIlpo Järvinen tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; 346bdf1ee5dSIlpo Järvinen tcp_hdr(skb)->cwr = 1; 347bdf1ee5dSIlpo Järvinen skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; 348bdf1ee5dSIlpo Järvinen } 349bdf1ee5dSIlpo Järvinen } else { 350bdf1ee5dSIlpo Järvinen /* ACK or retransmitted segment: clear ECT|CE */ 351bdf1ee5dSIlpo Järvinen INET_ECN_dontxmit(sk); 352bdf1ee5dSIlpo Järvinen } 353bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) 354bdf1ee5dSIlpo Järvinen tcp_hdr(skb)->ece = 1; 355bdf1ee5dSIlpo Järvinen } 356bdf1ee5dSIlpo Järvinen } 357bdf1ee5dSIlpo Järvinen 358e870a8efSIlpo Järvinen /* Constructs common control bits of non-data skb. If SYN/FIN is present, 359e870a8efSIlpo Järvinen * auto increment end seqno. 360e870a8efSIlpo Järvinen */ 361e870a8efSIlpo Järvinen static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags) 362e870a8efSIlpo Järvinen { 3632e8e18efSDavid S. Miller skb->ip_summed = CHECKSUM_PARTIAL; 364e870a8efSIlpo Järvinen skb->csum = 0; 365e870a8efSIlpo Järvinen 3664de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags; 367e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->sacked = 0; 368e870a8efSIlpo Järvinen 369e870a8efSIlpo Järvinen skb_shinfo(skb)->gso_segs = 1; 370e870a8efSIlpo Järvinen skb_shinfo(skb)->gso_size = 0; 371e870a8efSIlpo Järvinen skb_shinfo(skb)->gso_type = 0; 372e870a8efSIlpo Järvinen 373e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->seq = seq; 374a3433f35SChangli Gao if (flags & (TCPHDR_SYN | TCPHDR_FIN)) 375e870a8efSIlpo Järvinen seq++; 376e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->end_seq = seq; 377e870a8efSIlpo Järvinen } 378e870a8efSIlpo Järvinen 379a2a385d6SEric Dumazet static inline bool tcp_urg_mode(const struct tcp_sock *tp) 38033f5f57eSIlpo Järvinen { 38133f5f57eSIlpo Järvinen return tp->snd_una != tp->snd_up; 38233f5f57eSIlpo Järvinen } 38333f5f57eSIlpo Järvinen 38433ad798cSAdam Langley #define OPTION_SACK_ADVERTISE (1 << 0) 38533ad798cSAdam Langley #define OPTION_TS (1 << 1) 38633ad798cSAdam Langley #define OPTION_MD5 (1 << 2) 38789e95a61SOri Finkelman #define OPTION_WSCALE (1 << 3) 3882100c8d2SYuchung Cheng #define OPTION_FAST_OPEN_COOKIE (1 << 8) 38933ad798cSAdam Langley 39033ad798cSAdam Langley struct tcp_out_options { 3912100c8d2SYuchung Cheng u16 options; /* bit field of OPTION_* */ 3922100c8d2SYuchung Cheng u16 mss; /* 0 to disable */ 39333ad798cSAdam Langley u8 ws; /* window scale, 0 to disable */ 39433ad798cSAdam Langley u8 num_sack_blocks; /* number of SACK blocks to include */ 395bd0388aeSWilliam Allen Simpson u8 hash_size; /* bytes in hash_location */ 396bd0388aeSWilliam Allen Simpson __u8 *hash_location; /* temporary pointer, overloaded */ 3972100c8d2SYuchung Cheng __u32 tsval, tsecr; /* need to include OPTION_TS */ 3982100c8d2SYuchung Cheng struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */ 39933ad798cSAdam Langley }; 40033ad798cSAdam Langley 40167edfef7SAndi Kleen /* Write previously computed TCP options to the packet. 40267edfef7SAndi Kleen * 40367edfef7SAndi Kleen * Beware: Something in the Internet is very sensitive to the ordering of 404fd6149d3SIlpo Järvinen * TCP options, we learned this through the hard way, so be careful here. 405fd6149d3SIlpo Järvinen * Luckily we can at least blame others for their non-compliance but from 406fd6149d3SIlpo Järvinen * inter-operatibility perspective it seems that we're somewhat stuck with 407fd6149d3SIlpo Järvinen * the ordering which we have been using if we want to keep working with 408fd6149d3SIlpo Järvinen * those broken things (not that it currently hurts anybody as there isn't 409fd6149d3SIlpo Järvinen * particular reason why the ordering would need to be changed). 410fd6149d3SIlpo Järvinen * 411fd6149d3SIlpo Järvinen * At least SACK_PERM as the first option is known to lead to a disaster 412fd6149d3SIlpo Järvinen * (but it may well be that other scenarios fail similarly). 413fd6149d3SIlpo Järvinen */ 41433ad798cSAdam Langley static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp, 415bd0388aeSWilliam Allen Simpson struct tcp_out_options *opts) 416bd0388aeSWilliam Allen Simpson { 4172100c8d2SYuchung Cheng u16 options = opts->options; /* mungable copy */ 418bd0388aeSWilliam Allen Simpson 419bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_MD5 & options)) { 4201a2c6181SChristoph Paasch *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | 4211a2c6181SChristoph Paasch (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG); 422bd0388aeSWilliam Allen Simpson /* overload cookie hash location */ 423bd0388aeSWilliam Allen Simpson opts->hash_location = (__u8 *)ptr; 42433ad798cSAdam Langley ptr += 4; 42533ad798cSAdam Langley } 42633ad798cSAdam Langley 427fd6149d3SIlpo Järvinen if (unlikely(opts->mss)) { 428fd6149d3SIlpo Järvinen *ptr++ = htonl((TCPOPT_MSS << 24) | 429fd6149d3SIlpo Järvinen (TCPOLEN_MSS << 16) | 430fd6149d3SIlpo Järvinen opts->mss); 431fd6149d3SIlpo Järvinen } 432fd6149d3SIlpo Järvinen 433bd0388aeSWilliam Allen Simpson if (likely(OPTION_TS & options)) { 434bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 43533ad798cSAdam Langley *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | 43633ad798cSAdam Langley (TCPOLEN_SACK_PERM << 16) | 43733ad798cSAdam Langley (TCPOPT_TIMESTAMP << 8) | 43833ad798cSAdam Langley TCPOLEN_TIMESTAMP); 439bd0388aeSWilliam Allen Simpson options &= ~OPTION_SACK_ADVERTISE; 44033ad798cSAdam Langley } else { 441496c98dfSYOSHIFUJI Hideaki *ptr++ = htonl((TCPOPT_NOP << 24) | 44240efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 44340efc6faSStephen Hemminger (TCPOPT_TIMESTAMP << 8) | 44440efc6faSStephen Hemminger TCPOLEN_TIMESTAMP); 44540efc6faSStephen Hemminger } 44633ad798cSAdam Langley *ptr++ = htonl(opts->tsval); 44733ad798cSAdam Langley *ptr++ = htonl(opts->tsecr); 44833ad798cSAdam Langley } 44933ad798cSAdam Langley 450bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 45133ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 45233ad798cSAdam Langley (TCPOPT_NOP << 16) | 45333ad798cSAdam Langley (TCPOPT_SACK_PERM << 8) | 45433ad798cSAdam Langley TCPOLEN_SACK_PERM); 45533ad798cSAdam Langley } 45633ad798cSAdam Langley 457bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_WSCALE & options)) { 45833ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 45933ad798cSAdam Langley (TCPOPT_WINDOW << 16) | 46033ad798cSAdam Langley (TCPOLEN_WINDOW << 8) | 46133ad798cSAdam Langley opts->ws); 46233ad798cSAdam Langley } 46333ad798cSAdam Langley 46433ad798cSAdam Langley if (unlikely(opts->num_sack_blocks)) { 46533ad798cSAdam Langley struct tcp_sack_block *sp = tp->rx_opt.dsack ? 46633ad798cSAdam Langley tp->duplicate_sack : tp->selective_acks; 46740efc6faSStephen Hemminger int this_sack; 46840efc6faSStephen Hemminger 46940efc6faSStephen Hemminger *ptr++ = htonl((TCPOPT_NOP << 24) | 47040efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 47140efc6faSStephen Hemminger (TCPOPT_SACK << 8) | 47233ad798cSAdam Langley (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * 47340efc6faSStephen Hemminger TCPOLEN_SACK_PERBLOCK))); 4742de979bdSStephen Hemminger 47533ad798cSAdam Langley for (this_sack = 0; this_sack < opts->num_sack_blocks; 47633ad798cSAdam Langley ++this_sack) { 47740efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].start_seq); 47840efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].end_seq); 47940efc6faSStephen Hemminger } 4802de979bdSStephen Hemminger 48140efc6faSStephen Hemminger tp->rx_opt.dsack = 0; 48240efc6faSStephen Hemminger } 4832100c8d2SYuchung Cheng 4842100c8d2SYuchung Cheng if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) { 4852100c8d2SYuchung Cheng struct tcp_fastopen_cookie *foc = opts->fastopen_cookie; 4862100c8d2SYuchung Cheng 4872100c8d2SYuchung Cheng *ptr++ = htonl((TCPOPT_EXP << 24) | 4882100c8d2SYuchung Cheng ((TCPOLEN_EXP_FASTOPEN_BASE + foc->len) << 16) | 4892100c8d2SYuchung Cheng TCPOPT_FASTOPEN_MAGIC); 4902100c8d2SYuchung Cheng 4912100c8d2SYuchung Cheng memcpy(ptr, foc->val, foc->len); 4922100c8d2SYuchung Cheng if ((foc->len & 3) == 2) { 4932100c8d2SYuchung Cheng u8 *align = ((u8 *)ptr) + foc->len; 4942100c8d2SYuchung Cheng align[0] = align[1] = TCPOPT_NOP; 4952100c8d2SYuchung Cheng } 4962100c8d2SYuchung Cheng ptr += (foc->len + 3) >> 2; 4972100c8d2SYuchung Cheng } 49840efc6faSStephen Hemminger } 49940efc6faSStephen Hemminger 50067edfef7SAndi Kleen /* Compute TCP options for SYN packets. This is not the final 50167edfef7SAndi Kleen * network wire format yet. 50267edfef7SAndi Kleen */ 50395c96174SEric Dumazet static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb, 50433ad798cSAdam Langley struct tcp_out_options *opts, 505cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 506cf533ea5SEric Dumazet { 50733ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 50895c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 509783237e8SYuchung Cheng struct tcp_fastopen_request *fastopen = tp->fastopen_req; 51033ad798cSAdam Langley 511cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 51233ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 51333ad798cSAdam Langley if (*md5) { 51433ad798cSAdam Langley opts->options |= OPTION_MD5; 515bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 516cfb6eeb4SYOSHIFUJI Hideaki } 51733ad798cSAdam Langley #else 51833ad798cSAdam Langley *md5 = NULL; 519cfb6eeb4SYOSHIFUJI Hideaki #endif 52033ad798cSAdam Langley 52133ad798cSAdam Langley /* We always get an MSS option. The option bytes which will be seen in 52233ad798cSAdam Langley * normal data packets should timestamps be used, must be in the MSS 52333ad798cSAdam Langley * advertised. But we subtract them from tp->mss_cache so that 52433ad798cSAdam Langley * calculations in tcp_sendmsg are simpler etc. So account for this 52533ad798cSAdam Langley * fact here if necessary. If we don't do this correctly, as a 52633ad798cSAdam Langley * receiver we won't recognize data packets as being full sized when we 52733ad798cSAdam Langley * should, and thus we won't abide by the delayed ACK rules correctly. 52833ad798cSAdam Langley * SACKs don't matter, we never delay an ACK when we have any of those 52933ad798cSAdam Langley * going out. */ 53033ad798cSAdam Langley opts->mss = tcp_advertise_mss(sk); 531bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 53233ad798cSAdam Langley 533bb5b7c11SDavid S. Miller if (likely(sysctl_tcp_timestamps && *md5 == NULL)) { 53433ad798cSAdam Langley opts->options |= OPTION_TS; 535ee684b6fSAndrey Vagin opts->tsval = TCP_SKB_CB(skb)->when + tp->tsoffset; 53633ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 537bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 53833ad798cSAdam Langley } 539bb5b7c11SDavid S. Miller if (likely(sysctl_tcp_window_scaling)) { 54033ad798cSAdam Langley opts->ws = tp->rx_opt.rcv_wscale; 54189e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 542bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 54333ad798cSAdam Langley } 544bb5b7c11SDavid S. Miller if (likely(sysctl_tcp_sack)) { 54533ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 546b32d1310SDavid S. Miller if (unlikely(!(OPTION_TS & opts->options))) 547bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 54833ad798cSAdam Langley } 54933ad798cSAdam Langley 550783237e8SYuchung Cheng if (fastopen && fastopen->cookie.len >= 0) { 551783237e8SYuchung Cheng u32 need = TCPOLEN_EXP_FASTOPEN_BASE + fastopen->cookie.len; 552783237e8SYuchung Cheng need = (need + 3) & ~3U; /* Align to 32 bits */ 553783237e8SYuchung Cheng if (remaining >= need) { 554783237e8SYuchung Cheng opts->options |= OPTION_FAST_OPEN_COOKIE; 555783237e8SYuchung Cheng opts->fastopen_cookie = &fastopen->cookie; 556783237e8SYuchung Cheng remaining -= need; 557783237e8SYuchung Cheng tp->syn_fastopen = 1; 558783237e8SYuchung Cheng } 559783237e8SYuchung Cheng } 560bd0388aeSWilliam Allen Simpson 561bd0388aeSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 56233ad798cSAdam Langley } 56333ad798cSAdam Langley 56467edfef7SAndi Kleen /* Set up TCP options for SYN-ACKs. */ 56595c96174SEric Dumazet static unsigned int tcp_synack_options(struct sock *sk, 56633ad798cSAdam Langley struct request_sock *req, 56795c96174SEric Dumazet unsigned int mss, struct sk_buff *skb, 56833ad798cSAdam Langley struct tcp_out_options *opts, 5694957faadSWilliam Allen Simpson struct tcp_md5sig_key **md5, 5708336886fSJerry Chu struct tcp_fastopen_cookie *foc) 5714957faadSWilliam Allen Simpson { 57233ad798cSAdam Langley struct inet_request_sock *ireq = inet_rsk(req); 57395c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 57433ad798cSAdam Langley 57533ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 57633ad798cSAdam Langley *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req); 57733ad798cSAdam Langley if (*md5) { 57833ad798cSAdam Langley opts->options |= OPTION_MD5; 5794957faadSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 5804957faadSWilliam Allen Simpson 5814957faadSWilliam Allen Simpson /* We can't fit any SACK blocks in a packet with MD5 + TS 5824957faadSWilliam Allen Simpson * options. There was discussion about disabling SACK 5834957faadSWilliam Allen Simpson * rather than TS in order to fit in better with old, 5844957faadSWilliam Allen Simpson * buggy kernels, but that was deemed to be unnecessary. 5854957faadSWilliam Allen Simpson */ 586de213e5eSEric Dumazet ireq->tstamp_ok &= !ireq->sack_ok; 58733ad798cSAdam Langley } 58833ad798cSAdam Langley #else 58933ad798cSAdam Langley *md5 = NULL; 59033ad798cSAdam Langley #endif 59133ad798cSAdam Langley 5924957faadSWilliam Allen Simpson /* We always send an MSS option. */ 59333ad798cSAdam Langley opts->mss = mss; 5944957faadSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 59533ad798cSAdam Langley 59633ad798cSAdam Langley if (likely(ireq->wscale_ok)) { 59733ad798cSAdam Langley opts->ws = ireq->rcv_wscale; 59889e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 5994957faadSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 60033ad798cSAdam Langley } 601de213e5eSEric Dumazet if (likely(ireq->tstamp_ok)) { 60233ad798cSAdam Langley opts->options |= OPTION_TS; 60333ad798cSAdam Langley opts->tsval = TCP_SKB_CB(skb)->when; 60433ad798cSAdam Langley opts->tsecr = req->ts_recent; 6054957faadSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 60633ad798cSAdam Langley } 60733ad798cSAdam Langley if (likely(ireq->sack_ok)) { 60833ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 609de213e5eSEric Dumazet if (unlikely(!ireq->tstamp_ok)) 6104957faadSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 61133ad798cSAdam Langley } 6128336886fSJerry Chu if (foc != NULL) { 6138336886fSJerry Chu u32 need = TCPOLEN_EXP_FASTOPEN_BASE + foc->len; 6148336886fSJerry Chu need = (need + 3) & ~3U; /* Align to 32 bits */ 6158336886fSJerry Chu if (remaining >= need) { 6168336886fSJerry Chu opts->options |= OPTION_FAST_OPEN_COOKIE; 6178336886fSJerry Chu opts->fastopen_cookie = foc; 6188336886fSJerry Chu remaining -= need; 6198336886fSJerry Chu } 6208336886fSJerry Chu } 6214957faadSWilliam Allen Simpson 6224957faadSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 62333ad798cSAdam Langley } 62433ad798cSAdam Langley 62567edfef7SAndi Kleen /* Compute TCP options for ESTABLISHED sockets. This is not the 62667edfef7SAndi Kleen * final wire format yet. 62767edfef7SAndi Kleen */ 62895c96174SEric Dumazet static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb, 62933ad798cSAdam Langley struct tcp_out_options *opts, 630cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 631cf533ea5SEric Dumazet { 63233ad798cSAdam Langley struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL; 63333ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 63495c96174SEric Dumazet unsigned int size = 0; 635cabeccbdSIlpo Järvinen unsigned int eff_sacks; 63633ad798cSAdam Langley 63733ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 63833ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 63933ad798cSAdam Langley if (unlikely(*md5)) { 64033ad798cSAdam Langley opts->options |= OPTION_MD5; 64133ad798cSAdam Langley size += TCPOLEN_MD5SIG_ALIGNED; 64233ad798cSAdam Langley } 64333ad798cSAdam Langley #else 64433ad798cSAdam Langley *md5 = NULL; 64533ad798cSAdam Langley #endif 64633ad798cSAdam Langley 64733ad798cSAdam Langley if (likely(tp->rx_opt.tstamp_ok)) { 64833ad798cSAdam Langley opts->options |= OPTION_TS; 649ee684b6fSAndrey Vagin opts->tsval = tcb ? tcb->when + tp->tsoffset : 0; 65033ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 65133ad798cSAdam Langley size += TCPOLEN_TSTAMP_ALIGNED; 65233ad798cSAdam Langley } 65333ad798cSAdam Langley 654cabeccbdSIlpo Järvinen eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; 655cabeccbdSIlpo Järvinen if (unlikely(eff_sacks)) { 65695c96174SEric Dumazet const unsigned int remaining = MAX_TCP_OPTION_SPACE - size; 65733ad798cSAdam Langley opts->num_sack_blocks = 65895c96174SEric Dumazet min_t(unsigned int, eff_sacks, 65933ad798cSAdam Langley (remaining - TCPOLEN_SACK_BASE_ALIGNED) / 66033ad798cSAdam Langley TCPOLEN_SACK_PERBLOCK); 66133ad798cSAdam Langley size += TCPOLEN_SACK_BASE_ALIGNED + 66233ad798cSAdam Langley opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; 66333ad798cSAdam Langley } 66433ad798cSAdam Langley 66533ad798cSAdam Langley return size; 66640efc6faSStephen Hemminger } 6671da177e4SLinus Torvalds 66846d3ceabSEric Dumazet 66946d3ceabSEric Dumazet /* TCP SMALL QUEUES (TSQ) 67046d3ceabSEric Dumazet * 67146d3ceabSEric Dumazet * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev) 67246d3ceabSEric Dumazet * to reduce RTT and bufferbloat. 67346d3ceabSEric Dumazet * We do this using a special skb destructor (tcp_wfree). 67446d3ceabSEric Dumazet * 67546d3ceabSEric Dumazet * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb 67646d3ceabSEric Dumazet * needs to be reallocated in a driver. 67746d3ceabSEric Dumazet * The invariant being skb->truesize substracted from sk->sk_wmem_alloc 67846d3ceabSEric Dumazet * 67946d3ceabSEric Dumazet * Since transmit from skb destructor is forbidden, we use a tasklet 68046d3ceabSEric Dumazet * to process all sockets that eventually need to send more skbs. 68146d3ceabSEric Dumazet * We use one tasklet per cpu, with its own queue of sockets. 68246d3ceabSEric Dumazet */ 68346d3ceabSEric Dumazet struct tsq_tasklet { 68446d3ceabSEric Dumazet struct tasklet_struct tasklet; 68546d3ceabSEric Dumazet struct list_head head; /* queue of tcp sockets */ 68646d3ceabSEric Dumazet }; 68746d3ceabSEric Dumazet static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet); 68846d3ceabSEric Dumazet 6896f458dfbSEric Dumazet static void tcp_tsq_handler(struct sock *sk) 6906f458dfbSEric Dumazet { 6916f458dfbSEric Dumazet if ((1 << sk->sk_state) & 6926f458dfbSEric Dumazet (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING | 6936f458dfbSEric Dumazet TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) 6946f458dfbSEric Dumazet tcp_write_xmit(sk, tcp_current_mss(sk), 0, 0, GFP_ATOMIC); 6956f458dfbSEric Dumazet } 69646d3ceabSEric Dumazet /* 69746d3ceabSEric Dumazet * One tasklest per cpu tries to send more skbs. 69846d3ceabSEric Dumazet * We run in tasklet context but need to disable irqs when 69946d3ceabSEric Dumazet * transfering tsq->head because tcp_wfree() might 70046d3ceabSEric Dumazet * interrupt us (non NAPI drivers) 70146d3ceabSEric Dumazet */ 70246d3ceabSEric Dumazet static void tcp_tasklet_func(unsigned long data) 70346d3ceabSEric Dumazet { 70446d3ceabSEric Dumazet struct tsq_tasklet *tsq = (struct tsq_tasklet *)data; 70546d3ceabSEric Dumazet LIST_HEAD(list); 70646d3ceabSEric Dumazet unsigned long flags; 70746d3ceabSEric Dumazet struct list_head *q, *n; 70846d3ceabSEric Dumazet struct tcp_sock *tp; 70946d3ceabSEric Dumazet struct sock *sk; 71046d3ceabSEric Dumazet 71146d3ceabSEric Dumazet local_irq_save(flags); 71246d3ceabSEric Dumazet list_splice_init(&tsq->head, &list); 71346d3ceabSEric Dumazet local_irq_restore(flags); 71446d3ceabSEric Dumazet 71546d3ceabSEric Dumazet list_for_each_safe(q, n, &list) { 71646d3ceabSEric Dumazet tp = list_entry(q, struct tcp_sock, tsq_node); 71746d3ceabSEric Dumazet list_del(&tp->tsq_node); 71846d3ceabSEric Dumazet 71946d3ceabSEric Dumazet sk = (struct sock *)tp; 72046d3ceabSEric Dumazet bh_lock_sock(sk); 72146d3ceabSEric Dumazet 72246d3ceabSEric Dumazet if (!sock_owned_by_user(sk)) { 7236f458dfbSEric Dumazet tcp_tsq_handler(sk); 72446d3ceabSEric Dumazet } else { 72546d3ceabSEric Dumazet /* defer the work to tcp_release_cb() */ 7266f458dfbSEric Dumazet set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags); 72746d3ceabSEric Dumazet } 72846d3ceabSEric Dumazet bh_unlock_sock(sk); 72946d3ceabSEric Dumazet 73046d3ceabSEric Dumazet clear_bit(TSQ_QUEUED, &tp->tsq_flags); 73146d3ceabSEric Dumazet sk_free(sk); 73246d3ceabSEric Dumazet } 73346d3ceabSEric Dumazet } 73446d3ceabSEric Dumazet 7356f458dfbSEric Dumazet #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \ 7366f458dfbSEric Dumazet (1UL << TCP_WRITE_TIMER_DEFERRED) | \ 737563d34d0SEric Dumazet (1UL << TCP_DELACK_TIMER_DEFERRED) | \ 738563d34d0SEric Dumazet (1UL << TCP_MTU_REDUCED_DEFERRED)) 73946d3ceabSEric Dumazet /** 74046d3ceabSEric Dumazet * tcp_release_cb - tcp release_sock() callback 74146d3ceabSEric Dumazet * @sk: socket 74246d3ceabSEric Dumazet * 74346d3ceabSEric Dumazet * called from release_sock() to perform protocol dependent 74446d3ceabSEric Dumazet * actions before socket release. 74546d3ceabSEric Dumazet */ 74646d3ceabSEric Dumazet void tcp_release_cb(struct sock *sk) 74746d3ceabSEric Dumazet { 74846d3ceabSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 7496f458dfbSEric Dumazet unsigned long flags, nflags; 75046d3ceabSEric Dumazet 7516f458dfbSEric Dumazet /* perform an atomic operation only if at least one flag is set */ 7526f458dfbSEric Dumazet do { 7536f458dfbSEric Dumazet flags = tp->tsq_flags; 7546f458dfbSEric Dumazet if (!(flags & TCP_DEFERRED_ALL)) 7556f458dfbSEric Dumazet return; 7566f458dfbSEric Dumazet nflags = flags & ~TCP_DEFERRED_ALL; 7576f458dfbSEric Dumazet } while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags); 7586f458dfbSEric Dumazet 7596f458dfbSEric Dumazet if (flags & (1UL << TCP_TSQ_DEFERRED)) 7606f458dfbSEric Dumazet tcp_tsq_handler(sk); 7616f458dfbSEric Dumazet 762144d56e9SEric Dumazet if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) { 7636f458dfbSEric Dumazet tcp_write_timer_handler(sk); 764144d56e9SEric Dumazet __sock_put(sk); 765144d56e9SEric Dumazet } 766144d56e9SEric Dumazet if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) { 7676f458dfbSEric Dumazet tcp_delack_timer_handler(sk); 768144d56e9SEric Dumazet __sock_put(sk); 769144d56e9SEric Dumazet } 770144d56e9SEric Dumazet if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) { 771563d34d0SEric Dumazet sk->sk_prot->mtu_reduced(sk); 772144d56e9SEric Dumazet __sock_put(sk); 773144d56e9SEric Dumazet } 77446d3ceabSEric Dumazet } 77546d3ceabSEric Dumazet EXPORT_SYMBOL(tcp_release_cb); 77646d3ceabSEric Dumazet 77746d3ceabSEric Dumazet void __init tcp_tasklet_init(void) 77846d3ceabSEric Dumazet { 77946d3ceabSEric Dumazet int i; 78046d3ceabSEric Dumazet 78146d3ceabSEric Dumazet for_each_possible_cpu(i) { 78246d3ceabSEric Dumazet struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i); 78346d3ceabSEric Dumazet 78446d3ceabSEric Dumazet INIT_LIST_HEAD(&tsq->head); 78546d3ceabSEric Dumazet tasklet_init(&tsq->tasklet, 78646d3ceabSEric Dumazet tcp_tasklet_func, 78746d3ceabSEric Dumazet (unsigned long)tsq); 78846d3ceabSEric Dumazet } 78946d3ceabSEric Dumazet } 79046d3ceabSEric Dumazet 79146d3ceabSEric Dumazet /* 79246d3ceabSEric Dumazet * Write buffer destructor automatically called from kfree_skb. 79346d3ceabSEric Dumazet * We cant xmit new skbs from this context, as we might already 79446d3ceabSEric Dumazet * hold qdisc lock. 79546d3ceabSEric Dumazet */ 796d6a4a104SEric Dumazet void tcp_wfree(struct sk_buff *skb) 79746d3ceabSEric Dumazet { 79846d3ceabSEric Dumazet struct sock *sk = skb->sk; 79946d3ceabSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 80046d3ceabSEric Dumazet 80146d3ceabSEric Dumazet if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) && 80246d3ceabSEric Dumazet !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) { 80346d3ceabSEric Dumazet unsigned long flags; 80446d3ceabSEric Dumazet struct tsq_tasklet *tsq; 80546d3ceabSEric Dumazet 80646d3ceabSEric Dumazet /* Keep a ref on socket. 80746d3ceabSEric Dumazet * This last ref will be released in tcp_tasklet_func() 80846d3ceabSEric Dumazet */ 80946d3ceabSEric Dumazet atomic_sub(skb->truesize - 1, &sk->sk_wmem_alloc); 81046d3ceabSEric Dumazet 81146d3ceabSEric Dumazet /* queue this socket to tasklet queue */ 81246d3ceabSEric Dumazet local_irq_save(flags); 81346d3ceabSEric Dumazet tsq = &__get_cpu_var(tsq_tasklet); 81446d3ceabSEric Dumazet list_add(&tp->tsq_node, &tsq->head); 81546d3ceabSEric Dumazet tasklet_schedule(&tsq->tasklet); 81646d3ceabSEric Dumazet local_irq_restore(flags); 81746d3ceabSEric Dumazet } else { 81846d3ceabSEric Dumazet sock_wfree(skb); 81946d3ceabSEric Dumazet } 82046d3ceabSEric Dumazet } 82146d3ceabSEric Dumazet 8221da177e4SLinus Torvalds /* This routine actually transmits TCP packets queued in by 8231da177e4SLinus Torvalds * tcp_do_sendmsg(). This is used by both the initial 8241da177e4SLinus Torvalds * transmission and possible later retransmissions. 8251da177e4SLinus Torvalds * All SKB's seen here are completely headerless. It is our 8261da177e4SLinus Torvalds * job to build the TCP header, and pass the packet down to 8271da177e4SLinus Torvalds * IP so it can do the same plus pass the packet off to the 8281da177e4SLinus Torvalds * device. 8291da177e4SLinus Torvalds * 8301da177e4SLinus Torvalds * We are working here with either a clone of the original 8311da177e4SLinus Torvalds * SKB, or a fresh unique copy made by the retransmit engine. 8321da177e4SLinus Torvalds */ 833056834d9SIlpo Järvinen static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, 834056834d9SIlpo Järvinen gfp_t gfp_mask) 8351da177e4SLinus Torvalds { 8366687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 837dfb4b9dcSDavid S. Miller struct inet_sock *inet; 838dfb4b9dcSDavid S. Miller struct tcp_sock *tp; 839dfb4b9dcSDavid S. Miller struct tcp_skb_cb *tcb; 84033ad798cSAdam Langley struct tcp_out_options opts; 84195c96174SEric Dumazet unsigned int tcp_options_size, tcp_header_size; 842cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 8431da177e4SLinus Torvalds struct tcphdr *th; 8441da177e4SLinus Torvalds int err; 8451da177e4SLinus Torvalds 846dfb4b9dcSDavid S. Miller BUG_ON(!skb || !tcp_skb_pcount(skb)); 847dfb4b9dcSDavid S. Miller 848dfb4b9dcSDavid S. Miller /* If congestion control is doing timestamping, we must 849dfb4b9dcSDavid S. Miller * take such a timestamp before we potentially clone/copy. 850dfb4b9dcSDavid S. Miller */ 851164891aaSStephen Hemminger if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP) 852dfb4b9dcSDavid S. Miller __net_timestamp(skb); 853dfb4b9dcSDavid S. Miller 854dfb4b9dcSDavid S. Miller if (likely(clone_it)) { 8550e280af0SEric Dumazet const struct sk_buff *fclone = skb + 1; 8560e280af0SEric Dumazet 8570e280af0SEric Dumazet if (unlikely(skb->fclone == SKB_FCLONE_ORIG && 8580e280af0SEric Dumazet fclone->fclone == SKB_FCLONE_CLONE)) 8590e280af0SEric Dumazet NET_INC_STATS_BH(sock_net(sk), 8600e280af0SEric Dumazet LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES); 8610e280af0SEric Dumazet 862dfb4b9dcSDavid S. Miller if (unlikely(skb_cloned(skb))) 863dfb4b9dcSDavid S. Miller skb = pskb_copy(skb, gfp_mask); 864dfb4b9dcSDavid S. Miller else 865dfb4b9dcSDavid S. Miller skb = skb_clone(skb, gfp_mask); 866dfb4b9dcSDavid S. Miller if (unlikely(!skb)) 867dfb4b9dcSDavid S. Miller return -ENOBUFS; 868dfb4b9dcSDavid S. Miller } 869dfb4b9dcSDavid S. Miller 870dfb4b9dcSDavid S. Miller inet = inet_sk(sk); 871dfb4b9dcSDavid S. Miller tp = tcp_sk(sk); 872dfb4b9dcSDavid S. Miller tcb = TCP_SKB_CB(skb); 87333ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 8741da177e4SLinus Torvalds 8754de075e0SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) 87633ad798cSAdam Langley tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); 87733ad798cSAdam Langley else 87833ad798cSAdam Langley tcp_options_size = tcp_established_options(sk, skb, &opts, 87933ad798cSAdam Langley &md5); 88033ad798cSAdam Langley tcp_header_size = tcp_options_size + sizeof(struct tcphdr); 8811da177e4SLinus Torvalds 882547669d4SEric Dumazet if (tcp_packets_in_flight(tp) == 0) 8836687e988SArnaldo Carvalho de Melo tcp_ca_event(sk, CA_EVENT_TX_START); 884547669d4SEric Dumazet 885547669d4SEric Dumazet /* if no packet is in qdisc/device queue, then allow XPS to select 886547669d4SEric Dumazet * another queue. 887547669d4SEric Dumazet */ 888547669d4SEric Dumazet skb->ooo_okay = sk_wmem_alloc_get(sk) == 0; 8891da177e4SLinus Torvalds 890aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 891aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 89246d3ceabSEric Dumazet 89346d3ceabSEric Dumazet skb_orphan(skb); 89446d3ceabSEric Dumazet skb->sk = sk; 89546d3ceabSEric Dumazet skb->destructor = (sysctl_tcp_limit_output_bytes > 0) ? 89646d3ceabSEric Dumazet tcp_wfree : sock_wfree; 89746d3ceabSEric Dumazet atomic_add(skb->truesize, &sk->sk_wmem_alloc); 8981da177e4SLinus Torvalds 8991da177e4SLinus Torvalds /* Build TCP header and checksum it. */ 900aa8223c7SArnaldo Carvalho de Melo th = tcp_hdr(skb); 901c720c7e8SEric Dumazet th->source = inet->inet_sport; 902c720c7e8SEric Dumazet th->dest = inet->inet_dport; 9031da177e4SLinus Torvalds th->seq = htonl(tcb->seq); 9041da177e4SLinus Torvalds th->ack_seq = htonl(tp->rcv_nxt); 905df7a3b07SAl Viro *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | 9064de075e0SEric Dumazet tcb->tcp_flags); 907dfb4b9dcSDavid S. Miller 9084de075e0SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) { 9091da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments 9101da177e4SLinus Torvalds * is never scaled. 9111da177e4SLinus Torvalds */ 912600ff0c2SIlpo Järvinen th->window = htons(min(tp->rcv_wnd, 65535U)); 9131da177e4SLinus Torvalds } else { 9141da177e4SLinus Torvalds th->window = htons(tcp_select_window(sk)); 9151da177e4SLinus Torvalds } 9161da177e4SLinus Torvalds th->check = 0; 9171da177e4SLinus Torvalds th->urg_ptr = 0; 9181da177e4SLinus Torvalds 91933f5f57eSIlpo Järvinen /* The urg_mode check is necessary during a below snd_una win probe */ 9207691367dSHerbert Xu if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { 9217691367dSHerbert Xu if (before(tp->snd_up, tcb->seq + 0x10000)) { 9221da177e4SLinus Torvalds th->urg_ptr = htons(tp->snd_up - tcb->seq); 9231da177e4SLinus Torvalds th->urg = 1; 9247691367dSHerbert Xu } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { 9250eae88f3SEric Dumazet th->urg_ptr = htons(0xFFFF); 9267691367dSHerbert Xu th->urg = 1; 9277691367dSHerbert Xu } 9281da177e4SLinus Torvalds } 9291da177e4SLinus Torvalds 930bd0388aeSWilliam Allen Simpson tcp_options_write((__be32 *)(th + 1), tp, &opts); 9314de075e0SEric Dumazet if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0)) 9329e412ba7SIlpo Järvinen TCP_ECN_send(sk, skb, tcp_header_size); 933dfb4b9dcSDavid S. Miller 934cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 935cfb6eeb4SYOSHIFUJI Hideaki /* Calculate the MD5 hash, as we have all we need now */ 936cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 937a465419bSEric Dumazet sk_nocaps_add(sk, NETIF_F_GSO_MASK); 938bd0388aeSWilliam Allen Simpson tp->af_specific->calc_md5_hash(opts.hash_location, 93949a72dfbSAdam Langley md5, sk, NULL, skb); 940cfb6eeb4SYOSHIFUJI Hideaki } 941cfb6eeb4SYOSHIFUJI Hideaki #endif 942cfb6eeb4SYOSHIFUJI Hideaki 943bb296246SHerbert Xu icsk->icsk_af_ops->send_check(sk, skb); 9441da177e4SLinus Torvalds 9454de075e0SEric Dumazet if (likely(tcb->tcp_flags & TCPHDR_ACK)) 946fc6415bcSDavid S. Miller tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); 9471da177e4SLinus Torvalds 9481da177e4SLinus Torvalds if (skb->len != tcp_header_size) 949cf533ea5SEric Dumazet tcp_event_data_sent(tp, sk); 9501da177e4SLinus Torvalds 951bd37a088SWei Yongjun if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) 952aa2ea058STom Herbert TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, 953aa2ea058STom Herbert tcp_skb_pcount(skb)); 9541da177e4SLinus Torvalds 955d9d8da80SDavid S. Miller err = icsk->icsk_af_ops->queue_xmit(skb, &inet->cork.fl); 95683de47cdSHua Zhong if (likely(err <= 0)) 9571da177e4SLinus Torvalds return err; 9581da177e4SLinus Torvalds 9593cfe3baaSIlpo Järvinen tcp_enter_cwr(sk, 1); 9601da177e4SLinus Torvalds 961b9df3cb8SGerrit Renker return net_xmit_eval(err); 9621da177e4SLinus Torvalds } 9631da177e4SLinus Torvalds 96467edfef7SAndi Kleen /* This routine just queues the buffer for sending. 9651da177e4SLinus Torvalds * 9661da177e4SLinus Torvalds * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, 9671da177e4SLinus Torvalds * otherwise socket can stall. 9681da177e4SLinus Torvalds */ 9691da177e4SLinus Torvalds static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) 9701da177e4SLinus Torvalds { 9711da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 9721da177e4SLinus Torvalds 9731da177e4SLinus Torvalds /* Advance write_seq and place onto the write_queue. */ 9741da177e4SLinus Torvalds tp->write_seq = TCP_SKB_CB(skb)->end_seq; 9751da177e4SLinus Torvalds skb_header_release(skb); 976fe067e8aSDavid S. Miller tcp_add_write_queue_tail(sk, skb); 9773ab224beSHideo Aoki sk->sk_wmem_queued += skb->truesize; 9783ab224beSHideo Aoki sk_mem_charge(sk, skb->truesize); 9791da177e4SLinus Torvalds } 9801da177e4SLinus Torvalds 98167edfef7SAndi Kleen /* Initialize TSO segments for a packet. */ 982cf533ea5SEric Dumazet static void tcp_set_skb_tso_segs(const struct sock *sk, struct sk_buff *skb, 983056834d9SIlpo Järvinen unsigned int mss_now) 984f6302d1dSDavid S. Miller { 9858e5b9ddaSHerbert Xu if (skb->len <= mss_now || !sk_can_gso(sk) || 9868e5b9ddaSHerbert Xu skb->ip_summed == CHECKSUM_NONE) { 987f6302d1dSDavid S. Miller /* Avoid the costly divide in the normal 988f6302d1dSDavid S. Miller * non-TSO case. 989f6302d1dSDavid S. Miller */ 9907967168cSHerbert Xu skb_shinfo(skb)->gso_segs = 1; 9917967168cSHerbert Xu skb_shinfo(skb)->gso_size = 0; 992c9af6db4SPravin B Shelar skb_shinfo(skb)->gso_type = 0; 993f6302d1dSDavid S. Miller } else { 994356f89e1SIlpo Järvinen skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now); 9957967168cSHerbert Xu skb_shinfo(skb)->gso_size = mss_now; 996c9af6db4SPravin B Shelar skb_shinfo(skb)->gso_type = sk->sk_gso_type; 9971da177e4SLinus Torvalds } 9981da177e4SLinus Torvalds } 9991da177e4SLinus Torvalds 100091fed7a1SIlpo Järvinen /* When a modification to fackets out becomes necessary, we need to check 100168f8353bSIlpo Järvinen * skb is counted to fackets_out or not. 100291fed7a1SIlpo Järvinen */ 1003cf533ea5SEric Dumazet static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb, 100491fed7a1SIlpo Järvinen int decr) 100591fed7a1SIlpo Järvinen { 1006a47e5a98SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1007a47e5a98SIlpo Järvinen 1008dc86967bSIlpo Järvinen if (!tp->sacked_out || tcp_is_reno(tp)) 100991fed7a1SIlpo Järvinen return; 101091fed7a1SIlpo Järvinen 10116859d494SIlpo Järvinen if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq)) 101291fed7a1SIlpo Järvinen tp->fackets_out -= decr; 101391fed7a1SIlpo Järvinen } 101491fed7a1SIlpo Järvinen 1015797108d1SIlpo Järvinen /* Pcount in the middle of the write queue got changed, we need to do various 1016797108d1SIlpo Järvinen * tweaks to fix counters 1017797108d1SIlpo Järvinen */ 1018cf533ea5SEric Dumazet static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) 1019797108d1SIlpo Järvinen { 1020797108d1SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1021797108d1SIlpo Järvinen 1022797108d1SIlpo Järvinen tp->packets_out -= decr; 1023797108d1SIlpo Järvinen 1024797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 1025797108d1SIlpo Järvinen tp->sacked_out -= decr; 1026797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 1027797108d1SIlpo Järvinen tp->retrans_out -= decr; 1028797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 1029797108d1SIlpo Järvinen tp->lost_out -= decr; 1030797108d1SIlpo Järvinen 1031797108d1SIlpo Järvinen /* Reno case is special. Sigh... */ 1032797108d1SIlpo Järvinen if (tcp_is_reno(tp) && decr > 0) 1033797108d1SIlpo Järvinen tp->sacked_out -= min_t(u32, tp->sacked_out, decr); 1034797108d1SIlpo Järvinen 1035797108d1SIlpo Järvinen tcp_adjust_fackets_out(sk, skb, decr); 1036797108d1SIlpo Järvinen 1037797108d1SIlpo Järvinen if (tp->lost_skb_hint && 1038797108d1SIlpo Järvinen before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && 103952cf3cc8SIlpo Järvinen (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))) 1040797108d1SIlpo Järvinen tp->lost_cnt_hint -= decr; 1041797108d1SIlpo Järvinen 1042797108d1SIlpo Järvinen tcp_verify_left_out(tp); 1043797108d1SIlpo Järvinen } 1044797108d1SIlpo Järvinen 10451da177e4SLinus Torvalds /* Function to create two new TCP segments. Shrinks the given segment 10461da177e4SLinus Torvalds * to the specified size and appends a new segment with the rest of the 10471da177e4SLinus Torvalds * packet to the list. This won't be called frequently, I hope. 10481da177e4SLinus Torvalds * Remember, these are still headerless SKBs at this point. 10491da177e4SLinus Torvalds */ 1050056834d9SIlpo Järvinen int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, 1051056834d9SIlpo Järvinen unsigned int mss_now) 10521da177e4SLinus Torvalds { 10531da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 10541da177e4SLinus Torvalds struct sk_buff *buff; 10556475be16SDavid S. Miller int nsize, old_factor; 1056b60b49eaSHerbert Xu int nlen; 10579ce01461SIlpo Järvinen u8 flags; 10581da177e4SLinus Torvalds 10592fceec13SIlpo Järvinen if (WARN_ON(len > skb->len)) 10602fceec13SIlpo Järvinen return -EINVAL; 10616a438bbeSStephen Hemminger 10621da177e4SLinus Torvalds nsize = skb_headlen(skb) - len; 10631da177e4SLinus Torvalds if (nsize < 0) 10641da177e4SLinus Torvalds nsize = 0; 10651da177e4SLinus Torvalds 10661da177e4SLinus Torvalds if (skb_cloned(skb) && 10671da177e4SLinus Torvalds skb_is_nonlinear(skb) && 10681da177e4SLinus Torvalds pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) 10691da177e4SLinus Torvalds return -ENOMEM; 10701da177e4SLinus Torvalds 10711da177e4SLinus Torvalds /* Get a new skb... force flag on. */ 10721da177e4SLinus Torvalds buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC); 10731da177e4SLinus Torvalds if (buff == NULL) 10741da177e4SLinus Torvalds return -ENOMEM; /* We'll just try again later. */ 1075ef5cb973SHerbert Xu 10763ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 10773ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1078b60b49eaSHerbert Xu nlen = skb->len - len - nsize; 1079b60b49eaSHerbert Xu buff->truesize += nlen; 1080b60b49eaSHerbert Xu skb->truesize -= nlen; 10811da177e4SLinus Torvalds 10821da177e4SLinus Torvalds /* Correct the sequence numbers. */ 10831da177e4SLinus Torvalds TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 10841da177e4SLinus Torvalds TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 10851da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 10861da177e4SLinus Torvalds 10871da177e4SLinus Torvalds /* PSH and FIN should only be set in the second packet. */ 10884de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 10894de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 10904de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1091e14c3cafSHerbert Xu TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; 10921da177e4SLinus Torvalds 109384fa7933SPatrick McHardy if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { 10941da177e4SLinus Torvalds /* Copy and checksum data tail into the new buffer. */ 1095056834d9SIlpo Järvinen buff->csum = csum_partial_copy_nocheck(skb->data + len, 1096056834d9SIlpo Järvinen skb_put(buff, nsize), 10971da177e4SLinus Torvalds nsize, 0); 10981da177e4SLinus Torvalds 10991da177e4SLinus Torvalds skb_trim(skb, len); 11001da177e4SLinus Torvalds 11011da177e4SLinus Torvalds skb->csum = csum_block_sub(skb->csum, buff->csum, len); 11021da177e4SLinus Torvalds } else { 110384fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 11041da177e4SLinus Torvalds skb_split(skb, buff, len); 11051da177e4SLinus Torvalds } 11061da177e4SLinus Torvalds 11071da177e4SLinus Torvalds buff->ip_summed = skb->ip_summed; 11081da177e4SLinus Torvalds 11091da177e4SLinus Torvalds /* Looks stupid, but our code really uses when of 11101da177e4SLinus Torvalds * skbs, which it never sent before. --ANK 11111da177e4SLinus Torvalds */ 11121da177e4SLinus Torvalds TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when; 1113a61bbcf2SPatrick McHardy buff->tstamp = skb->tstamp; 11141da177e4SLinus Torvalds 11156475be16SDavid S. Miller old_factor = tcp_skb_pcount(skb); 11166475be16SDavid S. Miller 11171da177e4SLinus Torvalds /* Fix up tso_factor for both original and new SKB. */ 1118846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1119846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, buff, mss_now); 11201da177e4SLinus Torvalds 11216475be16SDavid S. Miller /* If this packet has been sent out already, we must 11226475be16SDavid S. Miller * adjust the various packet counters. 11236475be16SDavid S. Miller */ 1124cf0b450cSHerbert Xu if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 11256475be16SDavid S. Miller int diff = old_factor - tcp_skb_pcount(skb) - 11266475be16SDavid S. Miller tcp_skb_pcount(buff); 11271da177e4SLinus Torvalds 1128797108d1SIlpo Järvinen if (diff) 1129797108d1SIlpo Järvinen tcp_adjust_pcount(sk, skb, diff); 11301da177e4SLinus Torvalds } 11311da177e4SLinus Torvalds 11321da177e4SLinus Torvalds /* Link BUFF into the send queue. */ 1133f44b5271SDavid S. Miller skb_header_release(buff); 1134fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 11351da177e4SLinus Torvalds 11361da177e4SLinus Torvalds return 0; 11371da177e4SLinus Torvalds } 11381da177e4SLinus Torvalds 11391da177e4SLinus Torvalds /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c 11401da177e4SLinus Torvalds * eventually). The difference is that pulled data not copied, but 11411da177e4SLinus Torvalds * immediately discarded. 11421da177e4SLinus Torvalds */ 1143f2911969SHerbert Xu ~{PmVHI~} static void __pskb_trim_head(struct sk_buff *skb, int len) 11441da177e4SLinus Torvalds { 11451da177e4SLinus Torvalds int i, k, eat; 11461da177e4SLinus Torvalds 11474fa48bf3SEric Dumazet eat = min_t(int, len, skb_headlen(skb)); 11484fa48bf3SEric Dumazet if (eat) { 11494fa48bf3SEric Dumazet __skb_pull(skb, eat); 11504fa48bf3SEric Dumazet len -= eat; 11514fa48bf3SEric Dumazet if (!len) 11524fa48bf3SEric Dumazet return; 11534fa48bf3SEric Dumazet } 11541da177e4SLinus Torvalds eat = len; 11551da177e4SLinus Torvalds k = 0; 11561da177e4SLinus Torvalds for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) { 11579e903e08SEric Dumazet int size = skb_frag_size(&skb_shinfo(skb)->frags[i]); 11589e903e08SEric Dumazet 11599e903e08SEric Dumazet if (size <= eat) { 1160aff65da0SIan Campbell skb_frag_unref(skb, i); 11619e903e08SEric Dumazet eat -= size; 11621da177e4SLinus Torvalds } else { 11631da177e4SLinus Torvalds skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i]; 11641da177e4SLinus Torvalds if (eat) { 11651da177e4SLinus Torvalds skb_shinfo(skb)->frags[k].page_offset += eat; 11669e903e08SEric Dumazet skb_frag_size_sub(&skb_shinfo(skb)->frags[k], eat); 11671da177e4SLinus Torvalds eat = 0; 11681da177e4SLinus Torvalds } 11691da177e4SLinus Torvalds k++; 11701da177e4SLinus Torvalds } 11711da177e4SLinus Torvalds } 11721da177e4SLinus Torvalds skb_shinfo(skb)->nr_frags = k; 11731da177e4SLinus Torvalds 117427a884dcSArnaldo Carvalho de Melo skb_reset_tail_pointer(skb); 11751da177e4SLinus Torvalds skb->data_len -= len; 11761da177e4SLinus Torvalds skb->len = skb->data_len; 11771da177e4SLinus Torvalds } 11781da177e4SLinus Torvalds 117967edfef7SAndi Kleen /* Remove acked data from a packet in the transmit queue. */ 11801da177e4SLinus Torvalds int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 11811da177e4SLinus Torvalds { 118214bbd6a5SPravin B Shelar if (skb_unclone(skb, GFP_ATOMIC)) 11831da177e4SLinus Torvalds return -ENOMEM; 11841da177e4SLinus Torvalds 11854fa48bf3SEric Dumazet __pskb_trim_head(skb, len); 11861da177e4SLinus Torvalds 11871da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq += len; 118884fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 11891da177e4SLinus Torvalds 11901da177e4SLinus Torvalds skb->truesize -= len; 11911da177e4SLinus Torvalds sk->sk_wmem_queued -= len; 11923ab224beSHideo Aoki sk_mem_uncharge(sk, len); 11931da177e4SLinus Torvalds sock_set_flag(sk, SOCK_QUEUE_SHRUNK); 11941da177e4SLinus Torvalds 11955b35e1e6SNeal Cardwell /* Any change of skb->len requires recalculation of tso factor. */ 11961da177e4SLinus Torvalds if (tcp_skb_pcount(skb) > 1) 11975b35e1e6SNeal Cardwell tcp_set_skb_tso_segs(sk, skb, tcp_skb_mss(skb)); 11981da177e4SLinus Torvalds 11991da177e4SLinus Torvalds return 0; 12001da177e4SLinus Torvalds } 12011da177e4SLinus Torvalds 12021b63edd6SYuchung Cheng /* Calculate MSS not accounting any TCP options. */ 12031b63edd6SYuchung Cheng static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu) 12045d424d5aSJohn Heffner { 1205cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1206cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 12075d424d5aSJohn Heffner int mss_now; 12085d424d5aSJohn Heffner 12095d424d5aSJohn Heffner /* Calculate base mss without TCP options: 12105d424d5aSJohn Heffner It is MMS_S - sizeof(tcphdr) of rfc1122 12115d424d5aSJohn Heffner */ 12125d424d5aSJohn Heffner mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 12135d424d5aSJohn Heffner 121467469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 121567469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 121667469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 121767469601SEric Dumazet 121867469601SEric Dumazet if (dst && dst_allfrag(dst)) 121967469601SEric Dumazet mss_now -= icsk->icsk_af_ops->net_frag_header_len; 122067469601SEric Dumazet } 122167469601SEric Dumazet 12225d424d5aSJohn Heffner /* Clamp it (mss_clamp does not include tcp options) */ 12235d424d5aSJohn Heffner if (mss_now > tp->rx_opt.mss_clamp) 12245d424d5aSJohn Heffner mss_now = tp->rx_opt.mss_clamp; 12255d424d5aSJohn Heffner 12265d424d5aSJohn Heffner /* Now subtract optional transport overhead */ 12275d424d5aSJohn Heffner mss_now -= icsk->icsk_ext_hdr_len; 12285d424d5aSJohn Heffner 12295d424d5aSJohn Heffner /* Then reserve room for full set of TCP options and 8 bytes of data */ 12305d424d5aSJohn Heffner if (mss_now < 48) 12315d424d5aSJohn Heffner mss_now = 48; 12325d424d5aSJohn Heffner return mss_now; 12335d424d5aSJohn Heffner } 12345d424d5aSJohn Heffner 12351b63edd6SYuchung Cheng /* Calculate MSS. Not accounting for SACKs here. */ 12361b63edd6SYuchung Cheng int tcp_mtu_to_mss(struct sock *sk, int pmtu) 12371b63edd6SYuchung Cheng { 12381b63edd6SYuchung Cheng /* Subtract TCP options size, not including SACKs */ 12391b63edd6SYuchung Cheng return __tcp_mtu_to_mss(sk, pmtu) - 12401b63edd6SYuchung Cheng (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr)); 12411b63edd6SYuchung Cheng } 12421b63edd6SYuchung Cheng 12435d424d5aSJohn Heffner /* Inverse of above */ 124467469601SEric Dumazet int tcp_mss_to_mtu(struct sock *sk, int mss) 12455d424d5aSJohn Heffner { 1246cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1247cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 12485d424d5aSJohn Heffner int mtu; 12495d424d5aSJohn Heffner 12505d424d5aSJohn Heffner mtu = mss + 12515d424d5aSJohn Heffner tp->tcp_header_len + 12525d424d5aSJohn Heffner icsk->icsk_ext_hdr_len + 12535d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 12545d424d5aSJohn Heffner 125567469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 125667469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 125767469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 125867469601SEric Dumazet 125967469601SEric Dumazet if (dst && dst_allfrag(dst)) 126067469601SEric Dumazet mtu += icsk->icsk_af_ops->net_frag_header_len; 126167469601SEric Dumazet } 12625d424d5aSJohn Heffner return mtu; 12635d424d5aSJohn Heffner } 12645d424d5aSJohn Heffner 126567edfef7SAndi Kleen /* MTU probing init per socket */ 12665d424d5aSJohn Heffner void tcp_mtup_init(struct sock *sk) 12675d424d5aSJohn Heffner { 12685d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 12695d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 12705d424d5aSJohn Heffner 12715d424d5aSJohn Heffner icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1; 12725d424d5aSJohn Heffner icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 12735d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 12745d424d5aSJohn Heffner icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss); 12755d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 12765d424d5aSJohn Heffner } 12774bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_mtup_init); 12785d424d5aSJohn Heffner 12791da177e4SLinus Torvalds /* This function synchronize snd mss to current pmtu/exthdr set. 12801da177e4SLinus Torvalds 12811da177e4SLinus Torvalds tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 12821da177e4SLinus Torvalds for TCP options, but includes only bare TCP header. 12831da177e4SLinus Torvalds 12841da177e4SLinus Torvalds tp->rx_opt.mss_clamp is mss negotiated at connection setup. 1285caa20d9aSStephen Hemminger It is minimum of user_mss and mss received with SYN. 12861da177e4SLinus Torvalds It also does not include TCP options. 12871da177e4SLinus Torvalds 1288d83d8461SArnaldo Carvalho de Melo inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. 12891da177e4SLinus Torvalds 12901da177e4SLinus Torvalds tp->mss_cache is current effective sending mss, including 12911da177e4SLinus Torvalds all tcp options except for SACKs. It is evaluated, 12921da177e4SLinus Torvalds taking into account current pmtu, but never exceeds 12931da177e4SLinus Torvalds tp->rx_opt.mss_clamp. 12941da177e4SLinus Torvalds 12951da177e4SLinus Torvalds NOTE1. rfc1122 clearly states that advertised MSS 12961da177e4SLinus Torvalds DOES NOT include either tcp or ip options. 12971da177e4SLinus Torvalds 1298d83d8461SArnaldo Carvalho de Melo NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache 1299d83d8461SArnaldo Carvalho de Melo are READ ONLY outside this function. --ANK (980731) 13001da177e4SLinus Torvalds */ 13011da177e4SLinus Torvalds unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 13021da177e4SLinus Torvalds { 13031da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1304d83d8461SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 13055d424d5aSJohn Heffner int mss_now; 13061da177e4SLinus Torvalds 13075d424d5aSJohn Heffner if (icsk->icsk_mtup.search_high > pmtu) 13085d424d5aSJohn Heffner icsk->icsk_mtup.search_high = pmtu; 13091da177e4SLinus Torvalds 13105d424d5aSJohn Heffner mss_now = tcp_mtu_to_mss(sk, pmtu); 1311409d22b4SIlpo Järvinen mss_now = tcp_bound_to_half_wnd(tp, mss_now); 13121da177e4SLinus Torvalds 13131da177e4SLinus Torvalds /* And store cached results */ 1314d83d8461SArnaldo Carvalho de Melo icsk->icsk_pmtu_cookie = pmtu; 13155d424d5aSJohn Heffner if (icsk->icsk_mtup.enabled) 13165d424d5aSJohn Heffner mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); 1317c1b4a7e6SDavid S. Miller tp->mss_cache = mss_now; 13181da177e4SLinus Torvalds 13191da177e4SLinus Torvalds return mss_now; 13201da177e4SLinus Torvalds } 13214bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_sync_mss); 13221da177e4SLinus Torvalds 13231da177e4SLinus Torvalds /* Compute the current effective MSS, taking SACKs and IP options, 13241da177e4SLinus Torvalds * and even PMTU discovery events into account. 13251da177e4SLinus Torvalds */ 13260c54b85fSIlpo Järvinen unsigned int tcp_current_mss(struct sock *sk) 13271da177e4SLinus Torvalds { 1328cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1329cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1330c1b4a7e6SDavid S. Miller u32 mss_now; 133195c96174SEric Dumazet unsigned int header_len; 133233ad798cSAdam Langley struct tcp_out_options opts; 133333ad798cSAdam Langley struct tcp_md5sig_key *md5; 13341da177e4SLinus Torvalds 1335c1b4a7e6SDavid S. Miller mss_now = tp->mss_cache; 1336c1b4a7e6SDavid S. Miller 13371da177e4SLinus Torvalds if (dst) { 13381da177e4SLinus Torvalds u32 mtu = dst_mtu(dst); 1339d83d8461SArnaldo Carvalho de Melo if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 13401da177e4SLinus Torvalds mss_now = tcp_sync_mss(sk, mtu); 13411da177e4SLinus Torvalds } 13421da177e4SLinus Torvalds 134333ad798cSAdam Langley header_len = tcp_established_options(sk, NULL, &opts, &md5) + 134433ad798cSAdam Langley sizeof(struct tcphdr); 134533ad798cSAdam Langley /* The mss_cache is sized based on tp->tcp_header_len, which assumes 134633ad798cSAdam Langley * some common options. If this is an odd packet (because we have SACK 134733ad798cSAdam Langley * blocks etc) then our calculated header_len will be different, and 134833ad798cSAdam Langley * we have to adjust mss_now correspondingly */ 134933ad798cSAdam Langley if (header_len != tp->tcp_header_len) { 135033ad798cSAdam Langley int delta = (int) header_len - tp->tcp_header_len; 135133ad798cSAdam Langley mss_now -= delta; 135233ad798cSAdam Langley } 1353cfb6eeb4SYOSHIFUJI Hideaki 13541da177e4SLinus Torvalds return mss_now; 13551da177e4SLinus Torvalds } 13561da177e4SLinus Torvalds 1357a762a980SDavid S. Miller /* Congestion window validation. (RFC2861) */ 13589e412ba7SIlpo Järvinen static void tcp_cwnd_validate(struct sock *sk) 1359a762a980SDavid S. Miller { 13609e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1361a762a980SDavid S. Miller 1362d436d686SIlpo Järvinen if (tp->packets_out >= tp->snd_cwnd) { 1363a762a980SDavid S. Miller /* Network is feed fully. */ 1364a762a980SDavid S. Miller tp->snd_cwnd_used = 0; 1365a762a980SDavid S. Miller tp->snd_cwnd_stamp = tcp_time_stamp; 1366a762a980SDavid S. Miller } else { 1367a762a980SDavid S. Miller /* Network starves. */ 1368a762a980SDavid S. Miller if (tp->packets_out > tp->snd_cwnd_used) 1369a762a980SDavid S. Miller tp->snd_cwnd_used = tp->packets_out; 1370a762a980SDavid S. Miller 137115d33c07SDavid S. Miller if (sysctl_tcp_slow_start_after_idle && 137215d33c07SDavid S. Miller (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto) 1373a762a980SDavid S. Miller tcp_cwnd_application_limited(sk); 1374a762a980SDavid S. Miller } 1375a762a980SDavid S. Miller } 1376a762a980SDavid S. Miller 13770e3a4803SIlpo Järvinen /* Returns the portion of skb which can be sent right away without 13780e3a4803SIlpo Järvinen * introducing MSS oddities to segment boundaries. In rare cases where 13790e3a4803SIlpo Järvinen * mss_now != mss_cache, we will request caller to create a small skb 13800e3a4803SIlpo Järvinen * per input skb which could be mostly avoided here (if desired). 13815ea3a748SIlpo Järvinen * 13825ea3a748SIlpo Järvinen * We explicitly want to create a request for splitting write queue tail 13835ea3a748SIlpo Järvinen * to a small skb for Nagle purposes while avoiding unnecessary modulos, 13845ea3a748SIlpo Järvinen * thus all the complexity (cwnd_len is always MSS multiple which we 13855ea3a748SIlpo Järvinen * return whenever allowed by the other factors). Basically we need the 13865ea3a748SIlpo Järvinen * modulo only when the receiver window alone is the limiting factor or 13875ea3a748SIlpo Järvinen * when we would be allowed to send the split-due-to-Nagle skb fully. 13880e3a4803SIlpo Järvinen */ 1389cf533ea5SEric Dumazet static unsigned int tcp_mss_split_point(const struct sock *sk, const struct sk_buff *skb, 13901485348dSBen Hutchings unsigned int mss_now, unsigned int max_segs) 1391c1b4a7e6SDavid S. Miller { 1392cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 13931485348dSBen Hutchings u32 needed, window, max_len; 1394c1b4a7e6SDavid S. Miller 139590840defSIlpo Järvinen window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 13961485348dSBen Hutchings max_len = mss_now * max_segs; 13970e3a4803SIlpo Järvinen 13981485348dSBen Hutchings if (likely(max_len <= window && skb != tcp_write_queue_tail(sk))) 13991485348dSBen Hutchings return max_len; 14000e3a4803SIlpo Järvinen 14015ea3a748SIlpo Järvinen needed = min(skb->len, window); 14025ea3a748SIlpo Järvinen 14031485348dSBen Hutchings if (max_len <= needed) 14041485348dSBen Hutchings return max_len; 14050e3a4803SIlpo Järvinen 14060e3a4803SIlpo Järvinen return needed - needed % mss_now; 1407c1b4a7e6SDavid S. Miller } 1408c1b4a7e6SDavid S. Miller 1409c1b4a7e6SDavid S. Miller /* Can at least one segment of SKB be sent right now, according to the 1410c1b4a7e6SDavid S. Miller * congestion window rules? If so, return how many segments are allowed. 1411c1b4a7e6SDavid S. Miller */ 1412cf533ea5SEric Dumazet static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, 1413cf533ea5SEric Dumazet const struct sk_buff *skb) 1414c1b4a7e6SDavid S. Miller { 1415c1b4a7e6SDavid S. Miller u32 in_flight, cwnd; 1416c1b4a7e6SDavid S. Miller 1417c1b4a7e6SDavid S. Miller /* Don't be strict about the congestion window for the final FIN. */ 14184de075e0SEric Dumazet if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && 14194de075e0SEric Dumazet tcp_skb_pcount(skb) == 1) 1420c1b4a7e6SDavid S. Miller return 1; 1421c1b4a7e6SDavid S. Miller 1422c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1423c1b4a7e6SDavid S. Miller cwnd = tp->snd_cwnd; 1424c1b4a7e6SDavid S. Miller if (in_flight < cwnd) 1425c1b4a7e6SDavid S. Miller return (cwnd - in_flight); 1426c1b4a7e6SDavid S. Miller 1427c1b4a7e6SDavid S. Miller return 0; 1428c1b4a7e6SDavid S. Miller } 1429c1b4a7e6SDavid S. Miller 1430b595076aSUwe Kleine-König /* Initialize TSO state of a skb. 143167edfef7SAndi Kleen * This must be invoked the first time we consider transmitting 1432c1b4a7e6SDavid S. Miller * SKB onto the wire. 1433c1b4a7e6SDavid S. Miller */ 1434cf533ea5SEric Dumazet static int tcp_init_tso_segs(const struct sock *sk, struct sk_buff *skb, 1435056834d9SIlpo Järvinen unsigned int mss_now) 1436c1b4a7e6SDavid S. Miller { 1437c1b4a7e6SDavid S. Miller int tso_segs = tcp_skb_pcount(skb); 1438c1b4a7e6SDavid S. Miller 1439f8269a49SIlpo Järvinen if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { 1440846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1441c1b4a7e6SDavid S. Miller tso_segs = tcp_skb_pcount(skb); 1442c1b4a7e6SDavid S. Miller } 1443c1b4a7e6SDavid S. Miller return tso_segs; 1444c1b4a7e6SDavid S. Miller } 1445c1b4a7e6SDavid S. Miller 144667edfef7SAndi Kleen /* Minshall's variant of the Nagle send check. */ 1447a2a385d6SEric Dumazet static inline bool tcp_minshall_check(const struct tcp_sock *tp) 1448c1b4a7e6SDavid S. Miller { 1449c1b4a7e6SDavid S. Miller return after(tp->snd_sml, tp->snd_una) && 1450c1b4a7e6SDavid S. Miller !after(tp->snd_sml, tp->snd_nxt); 1451c1b4a7e6SDavid S. Miller } 1452c1b4a7e6SDavid S. Miller 1453a2a385d6SEric Dumazet /* Return false, if packet can be sent now without violation Nagle's rules: 1454c1b4a7e6SDavid S. Miller * 1. It is full sized. 1455c1b4a7e6SDavid S. Miller * 2. Or it contains FIN. (already checked by caller) 14566d67e9beSFeng King * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. 1457c1b4a7e6SDavid S. Miller * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 1458c1b4a7e6SDavid S. Miller * With Minshall's modification: all sent small packets are ACKed. 1459c1b4a7e6SDavid S. Miller */ 1460a2a385d6SEric Dumazet static inline bool tcp_nagle_check(const struct tcp_sock *tp, 1461c1b4a7e6SDavid S. Miller const struct sk_buff *skb, 146295c96174SEric Dumazet unsigned int mss_now, int nonagle) 1463c1b4a7e6SDavid S. Miller { 1464a02cec21SEric Dumazet return skb->len < mss_now && 1465c1b4a7e6SDavid S. Miller ((nonagle & TCP_NAGLE_CORK) || 1466a02cec21SEric Dumazet (!nonagle && tp->packets_out && tcp_minshall_check(tp))); 1467c1b4a7e6SDavid S. Miller } 1468c1b4a7e6SDavid S. Miller 1469a2a385d6SEric Dumazet /* Return true if the Nagle test allows this packet to be 1470c1b4a7e6SDavid S. Miller * sent now. 1471c1b4a7e6SDavid S. Miller */ 1472a2a385d6SEric Dumazet static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, 1473c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1474c1b4a7e6SDavid S. Miller { 1475c1b4a7e6SDavid S. Miller /* Nagle rule does not apply to frames, which sit in the middle of the 1476c1b4a7e6SDavid S. Miller * write_queue (they have no chances to get new data). 1477c1b4a7e6SDavid S. Miller * 1478c1b4a7e6SDavid S. Miller * This is implemented in the callers, where they modify the 'nonagle' 1479c1b4a7e6SDavid S. Miller * argument based upon the location of SKB in the send queue. 1480c1b4a7e6SDavid S. Miller */ 1481c1b4a7e6SDavid S. Miller if (nonagle & TCP_NAGLE_PUSH) 1482a2a385d6SEric Dumazet return true; 1483c1b4a7e6SDavid S. Miller 14849b44190dSYuchung Cheng /* Don't use the nagle rule for urgent data (or for the final FIN). */ 14859b44190dSYuchung Cheng if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) 1486a2a385d6SEric Dumazet return true; 1487c1b4a7e6SDavid S. Miller 1488c1b4a7e6SDavid S. Miller if (!tcp_nagle_check(tp, skb, cur_mss, nonagle)) 1489a2a385d6SEric Dumazet return true; 1490c1b4a7e6SDavid S. Miller 1491a2a385d6SEric Dumazet return false; 1492c1b4a7e6SDavid S. Miller } 1493c1b4a7e6SDavid S. Miller 1494c1b4a7e6SDavid S. Miller /* Does at least the first segment of SKB fit into the send window? */ 1495a2a385d6SEric Dumazet static bool tcp_snd_wnd_test(const struct tcp_sock *tp, 1496a2a385d6SEric Dumazet const struct sk_buff *skb, 1497056834d9SIlpo Järvinen unsigned int cur_mss) 1498c1b4a7e6SDavid S. Miller { 1499c1b4a7e6SDavid S. Miller u32 end_seq = TCP_SKB_CB(skb)->end_seq; 1500c1b4a7e6SDavid S. Miller 1501c1b4a7e6SDavid S. Miller if (skb->len > cur_mss) 1502c1b4a7e6SDavid S. Miller end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 1503c1b4a7e6SDavid S. Miller 150490840defSIlpo Järvinen return !after(end_seq, tcp_wnd_end(tp)); 1505c1b4a7e6SDavid S. Miller } 1506c1b4a7e6SDavid S. Miller 1507fe067e8aSDavid S. Miller /* This checks if the data bearing packet SKB (usually tcp_send_head(sk)) 1508c1b4a7e6SDavid S. Miller * should be put on the wire right now. If so, it returns the number of 1509c1b4a7e6SDavid S. Miller * packets allowed by the congestion window. 1510c1b4a7e6SDavid S. Miller */ 1511cf533ea5SEric Dumazet static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb, 1512c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1513c1b4a7e6SDavid S. Miller { 1514cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1515c1b4a7e6SDavid S. Miller unsigned int cwnd_quota; 1516c1b4a7e6SDavid S. Miller 1517846998aeSDavid S. Miller tcp_init_tso_segs(sk, skb, cur_mss); 1518c1b4a7e6SDavid S. Miller 1519c1b4a7e6SDavid S. Miller if (!tcp_nagle_test(tp, skb, cur_mss, nonagle)) 1520c1b4a7e6SDavid S. Miller return 0; 1521c1b4a7e6SDavid S. Miller 1522c1b4a7e6SDavid S. Miller cwnd_quota = tcp_cwnd_test(tp, skb); 1523056834d9SIlpo Järvinen if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss)) 1524c1b4a7e6SDavid S. Miller cwnd_quota = 0; 1525c1b4a7e6SDavid S. Miller 1526c1b4a7e6SDavid S. Miller return cwnd_quota; 1527c1b4a7e6SDavid S. Miller } 1528c1b4a7e6SDavid S. Miller 152967edfef7SAndi Kleen /* Test if sending is allowed right now. */ 1530a2a385d6SEric Dumazet bool tcp_may_send_now(struct sock *sk) 1531c1b4a7e6SDavid S. Miller { 1532cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1533fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 1534c1b4a7e6SDavid S. Miller 1535a02cec21SEric Dumazet return skb && 15360c54b85fSIlpo Järvinen tcp_snd_test(sk, skb, tcp_current_mss(sk), 1537c1b4a7e6SDavid S. Miller (tcp_skb_is_last(sk, skb) ? 1538a02cec21SEric Dumazet tp->nonagle : TCP_NAGLE_PUSH)); 1539c1b4a7e6SDavid S. Miller } 1540c1b4a7e6SDavid S. Miller 1541c1b4a7e6SDavid S. Miller /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 1542c1b4a7e6SDavid S. Miller * which is put after SKB on the list. It is very much like 1543c1b4a7e6SDavid S. Miller * tcp_fragment() except that it may make several kinds of assumptions 1544c1b4a7e6SDavid S. Miller * in order to speed up the splitting operation. In particular, we 1545c1b4a7e6SDavid S. Miller * know that all the data is in scatter-gather pages, and that the 1546c1b4a7e6SDavid S. Miller * packet has never been sent out before (and thus is not cloned). 1547c1b4a7e6SDavid S. Miller */ 1548056834d9SIlpo Järvinen static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, 1549c4ead4c5SEric Dumazet unsigned int mss_now, gfp_t gfp) 1550c1b4a7e6SDavid S. Miller { 1551c1b4a7e6SDavid S. Miller struct sk_buff *buff; 1552c1b4a7e6SDavid S. Miller int nlen = skb->len - len; 15539ce01461SIlpo Järvinen u8 flags; 1554c1b4a7e6SDavid S. Miller 1555c1b4a7e6SDavid S. Miller /* All of a TSO frame must be composed of paged data. */ 1556c8ac3774SHerbert Xu if (skb->len != skb->data_len) 1557c8ac3774SHerbert Xu return tcp_fragment(sk, skb, len, mss_now); 1558c1b4a7e6SDavid S. Miller 1559c4ead4c5SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, gfp); 1560c1b4a7e6SDavid S. Miller if (unlikely(buff == NULL)) 1561c1b4a7e6SDavid S. Miller return -ENOMEM; 1562c1b4a7e6SDavid S. Miller 15633ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 15643ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1565b60b49eaSHerbert Xu buff->truesize += nlen; 1566c1b4a7e6SDavid S. Miller skb->truesize -= nlen; 1567c1b4a7e6SDavid S. Miller 1568c1b4a7e6SDavid S. Miller /* Correct the sequence numbers. */ 1569c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 1570c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 1571c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 1572c1b4a7e6SDavid S. Miller 1573c1b4a7e6SDavid S. Miller /* PSH and FIN should only be set in the second packet. */ 15744de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 15754de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 15764de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1577c1b4a7e6SDavid S. Miller 1578c1b4a7e6SDavid S. Miller /* This packet was never sent out yet, so no SACK bits. */ 1579c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->sacked = 0; 1580c1b4a7e6SDavid S. Miller 158184fa7933SPatrick McHardy buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; 1582c1b4a7e6SDavid S. Miller skb_split(skb, buff, len); 1583c1b4a7e6SDavid S. Miller 1584c1b4a7e6SDavid S. Miller /* Fix up tso_factor for both original and new SKB. */ 1585846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1586846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, buff, mss_now); 1587c1b4a7e6SDavid S. Miller 1588c1b4a7e6SDavid S. Miller /* Link BUFF into the send queue. */ 1589c1b4a7e6SDavid S. Miller skb_header_release(buff); 1590fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 1591c1b4a7e6SDavid S. Miller 1592c1b4a7e6SDavid S. Miller return 0; 1593c1b4a7e6SDavid S. Miller } 1594c1b4a7e6SDavid S. Miller 1595c1b4a7e6SDavid S. Miller /* Try to defer sending, if possible, in order to minimize the amount 1596c1b4a7e6SDavid S. Miller * of TSO splitting we do. View it as a kind of TSO Nagle test. 1597c1b4a7e6SDavid S. Miller * 1598c1b4a7e6SDavid S. Miller * This algorithm is from John Heffner. 1599c1b4a7e6SDavid S. Miller */ 1600a2a385d6SEric Dumazet static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb) 1601c1b4a7e6SDavid S. Miller { 16029e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 16036687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 1604c1b4a7e6SDavid S. Miller u32 send_win, cong_win, limit, in_flight; 1605ad9f4f50SEric Dumazet int win_divisor; 1606c1b4a7e6SDavid S. Miller 16074de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 1608ae8064acSJohn Heffner goto send_now; 1609c1b4a7e6SDavid S. Miller 16106687e988SArnaldo Carvalho de Melo if (icsk->icsk_ca_state != TCP_CA_Open) 1611ae8064acSJohn Heffner goto send_now; 1612ae8064acSJohn Heffner 1613ae8064acSJohn Heffner /* Defer for less than two clock ticks. */ 1614bd515c3eSIlpo Järvinen if (tp->tso_deferred && 1615a2acde07SIlpo Järvinen (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1) 1616ae8064acSJohn Heffner goto send_now; 1617908a75c1SDavid S. Miller 1618c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1619c1b4a7e6SDavid S. Miller 1620056834d9SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight)); 1621c1b4a7e6SDavid S. Miller 162290840defSIlpo Järvinen send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1623c1b4a7e6SDavid S. Miller 1624c1b4a7e6SDavid S. Miller /* From in_flight test above, we know that cwnd > in_flight. */ 1625c1b4a7e6SDavid S. Miller cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; 1626c1b4a7e6SDavid S. Miller 1627c1b4a7e6SDavid S. Miller limit = min(send_win, cong_win); 1628c1b4a7e6SDavid S. Miller 1629ba244fe9SDavid S. Miller /* If a full-sized TSO skb can be sent, do it. */ 16301485348dSBen Hutchings if (limit >= min_t(unsigned int, sk->sk_gso_max_size, 16311485348dSBen Hutchings sk->sk_gso_max_segs * tp->mss_cache)) 1632ae8064acSJohn Heffner goto send_now; 1633ba244fe9SDavid S. Miller 163462ad2761SIlpo Järvinen /* Middle in queue won't get any more data, full sendable already? */ 163562ad2761SIlpo Järvinen if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) 163662ad2761SIlpo Järvinen goto send_now; 163762ad2761SIlpo Järvinen 1638ad9f4f50SEric Dumazet win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor); 1639ad9f4f50SEric Dumazet if (win_divisor) { 1640c1b4a7e6SDavid S. Miller u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); 1641c1b4a7e6SDavid S. Miller 1642c1b4a7e6SDavid S. Miller /* If at least some fraction of a window is available, 1643c1b4a7e6SDavid S. Miller * just use it. 1644c1b4a7e6SDavid S. Miller */ 1645ad9f4f50SEric Dumazet chunk /= win_divisor; 1646c1b4a7e6SDavid S. Miller if (limit >= chunk) 1647ae8064acSJohn Heffner goto send_now; 1648c1b4a7e6SDavid S. Miller } else { 1649c1b4a7e6SDavid S. Miller /* Different approach, try not to defer past a single 1650c1b4a7e6SDavid S. Miller * ACK. Receiver should ACK every other full sized 1651c1b4a7e6SDavid S. Miller * frame, so if we have space for more than 3 frames 1652c1b4a7e6SDavid S. Miller * then send now. 1653c1b4a7e6SDavid S. Miller */ 16546b5a5c0dSNeal Cardwell if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) 1655ae8064acSJohn Heffner goto send_now; 1656c1b4a7e6SDavid S. Miller } 1657c1b4a7e6SDavid S. Miller 1658f4541d60SEric Dumazet /* Ok, it looks like it is advisable to defer. 1659f4541d60SEric Dumazet * Do not rearm the timer if already set to not break TCP ACK clocking. 1660f4541d60SEric Dumazet */ 1661f4541d60SEric Dumazet if (!tp->tso_deferred) 1662ae8064acSJohn Heffner tp->tso_deferred = 1 | (jiffies << 1); 1663ae8064acSJohn Heffner 1664a2a385d6SEric Dumazet return true; 1665ae8064acSJohn Heffner 1666ae8064acSJohn Heffner send_now: 1667ae8064acSJohn Heffner tp->tso_deferred = 0; 1668a2a385d6SEric Dumazet return false; 1669c1b4a7e6SDavid S. Miller } 1670c1b4a7e6SDavid S. Miller 16715d424d5aSJohn Heffner /* Create a new MTU probe if we are ready. 167267edfef7SAndi Kleen * MTU probe is regularly attempting to increase the path MTU by 167367edfef7SAndi Kleen * deliberately sending larger packets. This discovers routing 167467edfef7SAndi Kleen * changes resulting in larger path MTUs. 167567edfef7SAndi Kleen * 16765d424d5aSJohn Heffner * Returns 0 if we should wait to probe (no cwnd available), 16775d424d5aSJohn Heffner * 1 if a probe was sent, 1678056834d9SIlpo Järvinen * -1 otherwise 1679056834d9SIlpo Järvinen */ 16805d424d5aSJohn Heffner static int tcp_mtu_probe(struct sock *sk) 16815d424d5aSJohn Heffner { 16825d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 16835d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 16845d424d5aSJohn Heffner struct sk_buff *skb, *nskb, *next; 16855d424d5aSJohn Heffner int len; 16865d424d5aSJohn Heffner int probe_size; 168791cc17c0SIlpo Järvinen int size_needed; 16885d424d5aSJohn Heffner int copy; 16895d424d5aSJohn Heffner int mss_now; 16905d424d5aSJohn Heffner 16915d424d5aSJohn Heffner /* Not currently probing/verifying, 16925d424d5aSJohn Heffner * not in recovery, 16935d424d5aSJohn Heffner * have enough cwnd, and 16945d424d5aSJohn Heffner * not SACKing (the variable headers throw things off) */ 16955d424d5aSJohn Heffner if (!icsk->icsk_mtup.enabled || 16965d424d5aSJohn Heffner icsk->icsk_mtup.probe_size || 16975d424d5aSJohn Heffner inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 16985d424d5aSJohn Heffner tp->snd_cwnd < 11 || 1699cabeccbdSIlpo Järvinen tp->rx_opt.num_sacks || tp->rx_opt.dsack) 17005d424d5aSJohn Heffner return -1; 17015d424d5aSJohn Heffner 17025d424d5aSJohn Heffner /* Very simple search strategy: just double the MSS. */ 17030c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 17045d424d5aSJohn Heffner probe_size = 2 * tp->mss_cache; 170591cc17c0SIlpo Järvinen size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; 17065d424d5aSJohn Heffner if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) { 17075d424d5aSJohn Heffner /* TODO: set timer for probe_converge_event */ 17085d424d5aSJohn Heffner return -1; 17095d424d5aSJohn Heffner } 17105d424d5aSJohn Heffner 17115d424d5aSJohn Heffner /* Have enough data in the send queue to probe? */ 17127f9c33e5SIlpo Järvinen if (tp->write_seq - tp->snd_nxt < size_needed) 17135d424d5aSJohn Heffner return -1; 17145d424d5aSJohn Heffner 171591cc17c0SIlpo Järvinen if (tp->snd_wnd < size_needed) 17165d424d5aSJohn Heffner return -1; 171790840defSIlpo Järvinen if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) 17185d424d5aSJohn Heffner return 0; 17195d424d5aSJohn Heffner 1720d67c58e9SIlpo Järvinen /* Do we need to wait to drain cwnd? With none in flight, don't stall */ 1721d67c58e9SIlpo Järvinen if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { 1722d67c58e9SIlpo Järvinen if (!tcp_packets_in_flight(tp)) 17235d424d5aSJohn Heffner return -1; 17245d424d5aSJohn Heffner else 17255d424d5aSJohn Heffner return 0; 17265d424d5aSJohn Heffner } 17275d424d5aSJohn Heffner 17285d424d5aSJohn Heffner /* We're allowed to probe. Build it now. */ 17295d424d5aSJohn Heffner if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL) 17305d424d5aSJohn Heffner return -1; 17313ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 17323ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 17335d424d5aSJohn Heffner 1734fe067e8aSDavid S. Miller skb = tcp_send_head(sk); 17355d424d5aSJohn Heffner 17365d424d5aSJohn Heffner TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 17375d424d5aSJohn Heffner TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 17384de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; 17395d424d5aSJohn Heffner TCP_SKB_CB(nskb)->sacked = 0; 17405d424d5aSJohn Heffner nskb->csum = 0; 174184fa7933SPatrick McHardy nskb->ip_summed = skb->ip_summed; 17425d424d5aSJohn Heffner 174350c4817eSIlpo Järvinen tcp_insert_write_queue_before(nskb, skb, sk); 174450c4817eSIlpo Järvinen 17455d424d5aSJohn Heffner len = 0; 1746234b6860SIlpo Järvinen tcp_for_write_queue_from_safe(skb, next, sk) { 17475d424d5aSJohn Heffner copy = min_t(int, skb->len, probe_size - len); 17485d424d5aSJohn Heffner if (nskb->ip_summed) 17495d424d5aSJohn Heffner skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); 17505d424d5aSJohn Heffner else 17515d424d5aSJohn Heffner nskb->csum = skb_copy_and_csum_bits(skb, 0, 1752056834d9SIlpo Järvinen skb_put(nskb, copy), 1753056834d9SIlpo Järvinen copy, nskb->csum); 17545d424d5aSJohn Heffner 17555d424d5aSJohn Heffner if (skb->len <= copy) { 17565d424d5aSJohn Heffner /* We've eaten all the data from this skb. 17575d424d5aSJohn Heffner * Throw it away. */ 17584de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 1759fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 17603ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 17615d424d5aSJohn Heffner } else { 17624de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & 1763a3433f35SChangli Gao ~(TCPHDR_FIN|TCPHDR_PSH); 17645d424d5aSJohn Heffner if (!skb_shinfo(skb)->nr_frags) { 17655d424d5aSJohn Heffner skb_pull(skb, copy); 176684fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 1767056834d9SIlpo Järvinen skb->csum = csum_partial(skb->data, 1768056834d9SIlpo Järvinen skb->len, 0); 17695d424d5aSJohn Heffner } else { 17705d424d5aSJohn Heffner __pskb_trim_head(skb, copy); 17715d424d5aSJohn Heffner tcp_set_skb_tso_segs(sk, skb, mss_now); 17725d424d5aSJohn Heffner } 17735d424d5aSJohn Heffner TCP_SKB_CB(skb)->seq += copy; 17745d424d5aSJohn Heffner } 17755d424d5aSJohn Heffner 17765d424d5aSJohn Heffner len += copy; 1777234b6860SIlpo Järvinen 1778234b6860SIlpo Järvinen if (len >= probe_size) 1779234b6860SIlpo Järvinen break; 17805d424d5aSJohn Heffner } 17815d424d5aSJohn Heffner tcp_init_tso_segs(sk, nskb, nskb->len); 17825d424d5aSJohn Heffner 17835d424d5aSJohn Heffner /* We're ready to send. If this fails, the probe will 17845d424d5aSJohn Heffner * be resegmented into mss-sized pieces by tcp_write_xmit(). */ 17855d424d5aSJohn Heffner TCP_SKB_CB(nskb)->when = tcp_time_stamp; 17865d424d5aSJohn Heffner if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 17875d424d5aSJohn Heffner /* Decrement cwnd here because we are sending 17885d424d5aSJohn Heffner * effectively two packets. */ 17895d424d5aSJohn Heffner tp->snd_cwnd--; 179066f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, nskb); 17915d424d5aSJohn Heffner 17925d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 17930e7b1368SJohn Heffner tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 17940e7b1368SJohn Heffner tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 17955d424d5aSJohn Heffner 17965d424d5aSJohn Heffner return 1; 17975d424d5aSJohn Heffner } 17985d424d5aSJohn Heffner 17995d424d5aSJohn Heffner return -1; 18005d424d5aSJohn Heffner } 18015d424d5aSJohn Heffner 18021da177e4SLinus Torvalds /* This routine writes packets to the network. It advances the 18031da177e4SLinus Torvalds * send_head. This happens as incoming acks open up the remote 18041da177e4SLinus Torvalds * window for us. 18051da177e4SLinus Torvalds * 1806f8269a49SIlpo Järvinen * LARGESEND note: !tcp_urg_mode is overkill, only frames between 1807f8269a49SIlpo Järvinen * snd_up-64k-mss .. snd_up cannot be large. However, taking into 1808f8269a49SIlpo Järvinen * account rare use of URG, this is not a big flaw. 1809f8269a49SIlpo Järvinen * 18106ba8a3b1SNandita Dukkipati * Send at most one packet when push_one > 0. Temporarily ignore 18116ba8a3b1SNandita Dukkipati * cwnd limit to force at most one packet out when push_one == 2. 18126ba8a3b1SNandita Dukkipati 1813a2a385d6SEric Dumazet * Returns true, if no segments are in flight and we have queued segments, 1814a2a385d6SEric Dumazet * but cannot send anything now because of SWS or another problem. 18151da177e4SLinus Torvalds */ 1816a2a385d6SEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 1817d5dd9175SIlpo Järvinen int push_one, gfp_t gfp) 18181da177e4SLinus Torvalds { 18191da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 182092df7b51SDavid S. Miller struct sk_buff *skb; 1821c1b4a7e6SDavid S. Miller unsigned int tso_segs, sent_pkts; 1822c1b4a7e6SDavid S. Miller int cwnd_quota; 18235d424d5aSJohn Heffner int result; 18241da177e4SLinus Torvalds 1825c1b4a7e6SDavid S. Miller sent_pkts = 0; 18265d424d5aSJohn Heffner 1827d5dd9175SIlpo Järvinen if (!push_one) { 18285d424d5aSJohn Heffner /* Do MTU probing. */ 1829d5dd9175SIlpo Järvinen result = tcp_mtu_probe(sk); 1830d5dd9175SIlpo Järvinen if (!result) { 1831a2a385d6SEric Dumazet return false; 18325d424d5aSJohn Heffner } else if (result > 0) { 18335d424d5aSJohn Heffner sent_pkts = 1; 18345d424d5aSJohn Heffner } 1835d5dd9175SIlpo Järvinen } 18365d424d5aSJohn Heffner 1837fe067e8aSDavid S. Miller while ((skb = tcp_send_head(sk))) { 1838c8ac3774SHerbert Xu unsigned int limit; 1839c8ac3774SHerbert Xu 184046d3ceabSEric Dumazet 1841b68e9f85SHerbert Xu tso_segs = tcp_init_tso_segs(sk, skb, mss_now); 1842c1b4a7e6SDavid S. Miller BUG_ON(!tso_segs); 1843c1b4a7e6SDavid S. Miller 1844ec342325SAndrew Vagin if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) 1845ec342325SAndrew Vagin goto repair; /* Skip network transmission */ 1846ec342325SAndrew Vagin 1847b68e9f85SHerbert Xu cwnd_quota = tcp_cwnd_test(tp, skb); 18486ba8a3b1SNandita Dukkipati if (!cwnd_quota) { 18496ba8a3b1SNandita Dukkipati if (push_one == 2) 18506ba8a3b1SNandita Dukkipati /* Force out a loss probe pkt. */ 18516ba8a3b1SNandita Dukkipati cwnd_quota = 1; 18526ba8a3b1SNandita Dukkipati else 1853b68e9f85SHerbert Xu break; 18546ba8a3b1SNandita Dukkipati } 1855b68e9f85SHerbert Xu 1856b68e9f85SHerbert Xu if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) 1857b68e9f85SHerbert Xu break; 1858b68e9f85SHerbert Xu 1859c1b4a7e6SDavid S. Miller if (tso_segs == 1) { 1860aa93466bSDavid S. Miller if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 1861aa93466bSDavid S. Miller (tcp_skb_is_last(sk, skb) ? 1862aa93466bSDavid S. Miller nonagle : TCP_NAGLE_PUSH)))) 1863aa93466bSDavid S. Miller break; 1864c1b4a7e6SDavid S. Miller } else { 1865d5dd9175SIlpo Järvinen if (!push_one && tcp_tso_should_defer(sk, skb)) 1866aa93466bSDavid S. Miller break; 1867c1b4a7e6SDavid S. Miller } 1868aa93466bSDavid S. Miller 186946d3ceabSEric Dumazet /* TSQ : sk_wmem_alloc accounts skb truesize, 187046d3ceabSEric Dumazet * including skb overhead. But thats OK. 187146d3ceabSEric Dumazet */ 187246d3ceabSEric Dumazet if (atomic_read(&sk->sk_wmem_alloc) >= sysctl_tcp_limit_output_bytes) { 187346d3ceabSEric Dumazet set_bit(TSQ_THROTTLED, &tp->tsq_flags); 187446d3ceabSEric Dumazet break; 187546d3ceabSEric Dumazet } 1876c8ac3774SHerbert Xu limit = mss_now; 1877f8269a49SIlpo Järvinen if (tso_segs > 1 && !tcp_urg_mode(tp)) 18780e3a4803SIlpo Järvinen limit = tcp_mss_split_point(sk, skb, mss_now, 18791485348dSBen Hutchings min_t(unsigned int, 18801485348dSBen Hutchings cwnd_quota, 18811485348dSBen Hutchings sk->sk_gso_max_segs)); 1882c8ac3774SHerbert Xu 1883c8ac3774SHerbert Xu if (skb->len > limit && 1884c4ead4c5SEric Dumazet unlikely(tso_fragment(sk, skb, limit, mss_now, gfp))) 18851da177e4SLinus Torvalds break; 18861da177e4SLinus Torvalds 18871da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 1888c1b4a7e6SDavid S. Miller 1889d5dd9175SIlpo Järvinen if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) 18901da177e4SLinus Torvalds break; 18911da177e4SLinus Torvalds 1892ec342325SAndrew Vagin repair: 18931da177e4SLinus Torvalds /* Advance the send_head. This one is sent out. 18941da177e4SLinus Torvalds * This call will increment packets_out. 18951da177e4SLinus Torvalds */ 189666f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 18971da177e4SLinus Torvalds 18981da177e4SLinus Torvalds tcp_minshall_update(tp, mss_now, skb); 1899a262f0cdSNandita Dukkipati sent_pkts += tcp_skb_pcount(skb); 1900d5dd9175SIlpo Järvinen 1901d5dd9175SIlpo Järvinen if (push_one) 1902d5dd9175SIlpo Järvinen break; 19031da177e4SLinus Torvalds } 19041da177e4SLinus Torvalds 1905aa93466bSDavid S. Miller if (likely(sent_pkts)) { 1906684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 1907684bad11SYuchung Cheng tp->prr_out += sent_pkts; 19086ba8a3b1SNandita Dukkipati 19096ba8a3b1SNandita Dukkipati /* Send one loss probe per tail loss episode. */ 19106ba8a3b1SNandita Dukkipati if (push_one != 2) 19116ba8a3b1SNandita Dukkipati tcp_schedule_loss_probe(sk); 19129e412ba7SIlpo Järvinen tcp_cwnd_validate(sk); 1913a2a385d6SEric Dumazet return false; 19141da177e4SLinus Torvalds } 19156ba8a3b1SNandita Dukkipati return (push_one == 2) || (!tp->packets_out && tcp_send_head(sk)); 19166ba8a3b1SNandita Dukkipati } 19176ba8a3b1SNandita Dukkipati 19186ba8a3b1SNandita Dukkipati bool tcp_schedule_loss_probe(struct sock *sk) 19196ba8a3b1SNandita Dukkipati { 19206ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 19216ba8a3b1SNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 19226ba8a3b1SNandita Dukkipati u32 timeout, tlp_time_stamp, rto_time_stamp; 19236ba8a3b1SNandita Dukkipati u32 rtt = tp->srtt >> 3; 19246ba8a3b1SNandita Dukkipati 19256ba8a3b1SNandita Dukkipati if (WARN_ON(icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS)) 19266ba8a3b1SNandita Dukkipati return false; 19276ba8a3b1SNandita Dukkipati /* No consecutive loss probes. */ 19286ba8a3b1SNandita Dukkipati if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) { 19296ba8a3b1SNandita Dukkipati tcp_rearm_rto(sk); 19306ba8a3b1SNandita Dukkipati return false; 19316ba8a3b1SNandita Dukkipati } 19326ba8a3b1SNandita Dukkipati /* Don't do any loss probe on a Fast Open connection before 3WHS 19336ba8a3b1SNandita Dukkipati * finishes. 19346ba8a3b1SNandita Dukkipati */ 19356ba8a3b1SNandita Dukkipati if (sk->sk_state == TCP_SYN_RECV) 19366ba8a3b1SNandita Dukkipati return false; 19376ba8a3b1SNandita Dukkipati 19386ba8a3b1SNandita Dukkipati /* TLP is only scheduled when next timer event is RTO. */ 19396ba8a3b1SNandita Dukkipati if (icsk->icsk_pending != ICSK_TIME_RETRANS) 19406ba8a3b1SNandita Dukkipati return false; 19416ba8a3b1SNandita Dukkipati 19426ba8a3b1SNandita Dukkipati /* Schedule a loss probe in 2*RTT for SACK capable connections 19436ba8a3b1SNandita Dukkipati * in Open state, that are either limited by cwnd or application. 19446ba8a3b1SNandita Dukkipati */ 19456ba8a3b1SNandita Dukkipati if (sysctl_tcp_early_retrans < 3 || !rtt || !tp->packets_out || 19466ba8a3b1SNandita Dukkipati !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open) 19476ba8a3b1SNandita Dukkipati return false; 19486ba8a3b1SNandita Dukkipati 19496ba8a3b1SNandita Dukkipati if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) && 19506ba8a3b1SNandita Dukkipati tcp_send_head(sk)) 19516ba8a3b1SNandita Dukkipati return false; 19526ba8a3b1SNandita Dukkipati 19536ba8a3b1SNandita Dukkipati /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account 19546ba8a3b1SNandita Dukkipati * for delayed ack when there's one outstanding packet. 19556ba8a3b1SNandita Dukkipati */ 19566ba8a3b1SNandita Dukkipati timeout = rtt << 1; 19576ba8a3b1SNandita Dukkipati if (tp->packets_out == 1) 19586ba8a3b1SNandita Dukkipati timeout = max_t(u32, timeout, 19596ba8a3b1SNandita Dukkipati (rtt + (rtt >> 1) + TCP_DELACK_MAX)); 19606ba8a3b1SNandita Dukkipati timeout = max_t(u32, timeout, msecs_to_jiffies(10)); 19616ba8a3b1SNandita Dukkipati 19626ba8a3b1SNandita Dukkipati /* If RTO is shorter, just schedule TLP in its place. */ 19636ba8a3b1SNandita Dukkipati tlp_time_stamp = tcp_time_stamp + timeout; 19646ba8a3b1SNandita Dukkipati rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout; 19656ba8a3b1SNandita Dukkipati if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) { 19666ba8a3b1SNandita Dukkipati s32 delta = rto_time_stamp - tcp_time_stamp; 19676ba8a3b1SNandita Dukkipati if (delta > 0) 19686ba8a3b1SNandita Dukkipati timeout = delta; 19696ba8a3b1SNandita Dukkipati } 19706ba8a3b1SNandita Dukkipati 19716ba8a3b1SNandita Dukkipati inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, 19726ba8a3b1SNandita Dukkipati TCP_RTO_MAX); 19736ba8a3b1SNandita Dukkipati return true; 19746ba8a3b1SNandita Dukkipati } 19756ba8a3b1SNandita Dukkipati 19766ba8a3b1SNandita Dukkipati /* When probe timeout (PTO) fires, send a new segment if one exists, else 19776ba8a3b1SNandita Dukkipati * retransmit the last segment. 19786ba8a3b1SNandita Dukkipati */ 19796ba8a3b1SNandita Dukkipati void tcp_send_loss_probe(struct sock *sk) 19806ba8a3b1SNandita Dukkipati { 19819b717a8dSNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 19826ba8a3b1SNandita Dukkipati struct sk_buff *skb; 19836ba8a3b1SNandita Dukkipati int pcount; 19846ba8a3b1SNandita Dukkipati int mss = tcp_current_mss(sk); 19856ba8a3b1SNandita Dukkipati int err = -1; 19866ba8a3b1SNandita Dukkipati 19876ba8a3b1SNandita Dukkipati if (tcp_send_head(sk) != NULL) { 19886ba8a3b1SNandita Dukkipati err = tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC); 19896ba8a3b1SNandita Dukkipati goto rearm_timer; 19906ba8a3b1SNandita Dukkipati } 19916ba8a3b1SNandita Dukkipati 19929b717a8dSNandita Dukkipati /* At most one outstanding TLP retransmission. */ 19939b717a8dSNandita Dukkipati if (tp->tlp_high_seq) 19949b717a8dSNandita Dukkipati goto rearm_timer; 19959b717a8dSNandita Dukkipati 19966ba8a3b1SNandita Dukkipati /* Retransmit last segment. */ 19976ba8a3b1SNandita Dukkipati skb = tcp_write_queue_tail(sk); 19986ba8a3b1SNandita Dukkipati if (WARN_ON(!skb)) 19996ba8a3b1SNandita Dukkipati goto rearm_timer; 20006ba8a3b1SNandita Dukkipati 20016ba8a3b1SNandita Dukkipati pcount = tcp_skb_pcount(skb); 20026ba8a3b1SNandita Dukkipati if (WARN_ON(!pcount)) 20036ba8a3b1SNandita Dukkipati goto rearm_timer; 20046ba8a3b1SNandita Dukkipati 20056ba8a3b1SNandita Dukkipati if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) { 20066ba8a3b1SNandita Dukkipati if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss))) 20076ba8a3b1SNandita Dukkipati goto rearm_timer; 20086ba8a3b1SNandita Dukkipati skb = tcp_write_queue_tail(sk); 20096ba8a3b1SNandita Dukkipati } 20106ba8a3b1SNandita Dukkipati 20116ba8a3b1SNandita Dukkipati if (WARN_ON(!skb || !tcp_skb_pcount(skb))) 20126ba8a3b1SNandita Dukkipati goto rearm_timer; 20136ba8a3b1SNandita Dukkipati 20146ba8a3b1SNandita Dukkipati /* Probe with zero data doesn't trigger fast recovery. */ 20156ba8a3b1SNandita Dukkipati if (skb->len > 0) 20166ba8a3b1SNandita Dukkipati err = __tcp_retransmit_skb(sk, skb); 20176ba8a3b1SNandita Dukkipati 20189b717a8dSNandita Dukkipati /* Record snd_nxt for loss detection. */ 20199b717a8dSNandita Dukkipati if (likely(!err)) 20209b717a8dSNandita Dukkipati tp->tlp_high_seq = tp->snd_nxt; 20219b717a8dSNandita Dukkipati 20226ba8a3b1SNandita Dukkipati rearm_timer: 20236ba8a3b1SNandita Dukkipati inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 20246ba8a3b1SNandita Dukkipati inet_csk(sk)->icsk_rto, 20256ba8a3b1SNandita Dukkipati TCP_RTO_MAX); 20266ba8a3b1SNandita Dukkipati 20276ba8a3b1SNandita Dukkipati if (likely(!err)) 20286ba8a3b1SNandita Dukkipati NET_INC_STATS_BH(sock_net(sk), 20296ba8a3b1SNandita Dukkipati LINUX_MIB_TCPLOSSPROBES); 20306ba8a3b1SNandita Dukkipati return; 20311da177e4SLinus Torvalds } 20321da177e4SLinus Torvalds 2033a762a980SDavid S. Miller /* Push out any pending frames which were held back due to 2034a762a980SDavid S. Miller * TCP_CORK or attempt at coalescing tiny packets. 2035a762a980SDavid S. Miller * The socket must be locked by the caller. 2036a762a980SDavid S. Miller */ 20379e412ba7SIlpo Järvinen void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 20389e412ba7SIlpo Järvinen int nonagle) 2039a762a980SDavid S. Miller { 2040726e07a8SIlpo Järvinen /* If we are closed, the bytes will have to remain here. 2041726e07a8SIlpo Järvinen * In time closedown will finish, we empty the write queue and 2042726e07a8SIlpo Järvinen * all will be happy. 2043726e07a8SIlpo Järvinen */ 2044726e07a8SIlpo Järvinen if (unlikely(sk->sk_state == TCP_CLOSE)) 2045726e07a8SIlpo Järvinen return; 2046726e07a8SIlpo Järvinen 204799a1dec7SMel Gorman if (tcp_write_xmit(sk, cur_mss, nonagle, 0, 204899a1dec7SMel Gorman sk_gfp_atomic(sk, GFP_ATOMIC))) 20499e412ba7SIlpo Järvinen tcp_check_probe_timer(sk); 2050a762a980SDavid S. Miller } 2051a762a980SDavid S. Miller 2052c1b4a7e6SDavid S. Miller /* Send _single_ skb sitting at the send head. This function requires 2053c1b4a7e6SDavid S. Miller * true push pending frames to setup probe timer etc. 2054c1b4a7e6SDavid S. Miller */ 2055c1b4a7e6SDavid S. Miller void tcp_push_one(struct sock *sk, unsigned int mss_now) 2056c1b4a7e6SDavid S. Miller { 2057fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 2058c1b4a7e6SDavid S. Miller 2059c1b4a7e6SDavid S. Miller BUG_ON(!skb || skb->len < mss_now); 2060c1b4a7e6SDavid S. Miller 2061d5dd9175SIlpo Järvinen tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); 2062c1b4a7e6SDavid S. Miller } 2063c1b4a7e6SDavid S. Miller 20641da177e4SLinus Torvalds /* This function returns the amount that we can raise the 20651da177e4SLinus Torvalds * usable window based on the following constraints 20661da177e4SLinus Torvalds * 20671da177e4SLinus Torvalds * 1. The window can never be shrunk once it is offered (RFC 793) 20681da177e4SLinus Torvalds * 2. We limit memory per socket 20691da177e4SLinus Torvalds * 20701da177e4SLinus Torvalds * RFC 1122: 20711da177e4SLinus Torvalds * "the suggested [SWS] avoidance algorithm for the receiver is to keep 20721da177e4SLinus Torvalds * RECV.NEXT + RCV.WIN fixed until: 20731da177e4SLinus Torvalds * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 20741da177e4SLinus Torvalds * 20751da177e4SLinus Torvalds * i.e. don't raise the right edge of the window until you can raise 20761da177e4SLinus Torvalds * it at least MSS bytes. 20771da177e4SLinus Torvalds * 20781da177e4SLinus Torvalds * Unfortunately, the recommended algorithm breaks header prediction, 20791da177e4SLinus Torvalds * since header prediction assumes th->window stays fixed. 20801da177e4SLinus Torvalds * 20811da177e4SLinus Torvalds * Strictly speaking, keeping th->window fixed violates the receiver 20821da177e4SLinus Torvalds * side SWS prevention criteria. The problem is that under this rule 20831da177e4SLinus Torvalds * a stream of single byte packets will cause the right side of the 20841da177e4SLinus Torvalds * window to always advance by a single byte. 20851da177e4SLinus Torvalds * 20861da177e4SLinus Torvalds * Of course, if the sender implements sender side SWS prevention 20871da177e4SLinus Torvalds * then this will not be a problem. 20881da177e4SLinus Torvalds * 20891da177e4SLinus Torvalds * BSD seems to make the following compromise: 20901da177e4SLinus Torvalds * 20911da177e4SLinus Torvalds * If the free space is less than the 1/4 of the maximum 20921da177e4SLinus Torvalds * space available and the free space is less than 1/2 mss, 20931da177e4SLinus Torvalds * then set the window to 0. 20941da177e4SLinus Torvalds * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 20951da177e4SLinus Torvalds * Otherwise, just prevent the window from shrinking 20961da177e4SLinus Torvalds * and from being larger than the largest representable value. 20971da177e4SLinus Torvalds * 20981da177e4SLinus Torvalds * This prevents incremental opening of the window in the regime 20991da177e4SLinus Torvalds * where TCP is limited by the speed of the reader side taking 21001da177e4SLinus Torvalds * data out of the TCP receive queue. It does nothing about 21011da177e4SLinus Torvalds * those cases where the window is constrained on the sender side 21021da177e4SLinus Torvalds * because the pipeline is full. 21031da177e4SLinus Torvalds * 21041da177e4SLinus Torvalds * BSD also seems to "accidentally" limit itself to windows that are a 21051da177e4SLinus Torvalds * multiple of MSS, at least until the free space gets quite small. 21061da177e4SLinus Torvalds * This would appear to be a side effect of the mbuf implementation. 21071da177e4SLinus Torvalds * Combining these two algorithms results in the observed behavior 21081da177e4SLinus Torvalds * of having a fixed window size at almost all times. 21091da177e4SLinus Torvalds * 21101da177e4SLinus Torvalds * Below we obtain similar behavior by forcing the offered window to 21111da177e4SLinus Torvalds * a multiple of the mss when it is feasible to do so. 21121da177e4SLinus Torvalds * 21131da177e4SLinus Torvalds * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 21141da177e4SLinus Torvalds * Regular options like TIMESTAMP are taken into account. 21151da177e4SLinus Torvalds */ 21161da177e4SLinus Torvalds u32 __tcp_select_window(struct sock *sk) 21171da177e4SLinus Torvalds { 2118463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 21191da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2120caa20d9aSStephen Hemminger /* MSS for the peer's data. Previous versions used mss_clamp 21211da177e4SLinus Torvalds * here. I don't know if the value based on our guesses 21221da177e4SLinus Torvalds * of peer's MSS is better for the performance. It's more correct 21231da177e4SLinus Torvalds * but may be worse for the performance because of rcv_mss 21241da177e4SLinus Torvalds * fluctuations. --SAW 1998/11/1 21251da177e4SLinus Torvalds */ 2126463c84b9SArnaldo Carvalho de Melo int mss = icsk->icsk_ack.rcv_mss; 21271da177e4SLinus Torvalds int free_space = tcp_space(sk); 21281da177e4SLinus Torvalds int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk)); 21291da177e4SLinus Torvalds int window; 21301da177e4SLinus Torvalds 21311da177e4SLinus Torvalds if (mss > full_space) 21321da177e4SLinus Torvalds mss = full_space; 21331da177e4SLinus Torvalds 2134b92edbe0SEric Dumazet if (free_space < (full_space >> 1)) { 2135463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.quick = 0; 21361da177e4SLinus Torvalds 2137180d8cd9SGlauber Costa if (sk_under_memory_pressure(sk)) 2138056834d9SIlpo Järvinen tp->rcv_ssthresh = min(tp->rcv_ssthresh, 2139056834d9SIlpo Järvinen 4U * tp->advmss); 21401da177e4SLinus Torvalds 21411da177e4SLinus Torvalds if (free_space < mss) 21421da177e4SLinus Torvalds return 0; 21431da177e4SLinus Torvalds } 21441da177e4SLinus Torvalds 21451da177e4SLinus Torvalds if (free_space > tp->rcv_ssthresh) 21461da177e4SLinus Torvalds free_space = tp->rcv_ssthresh; 21471da177e4SLinus Torvalds 21481da177e4SLinus Torvalds /* Don't do rounding if we are using window scaling, since the 21491da177e4SLinus Torvalds * scaled window will not line up with the MSS boundary anyway. 21501da177e4SLinus Torvalds */ 21511da177e4SLinus Torvalds window = tp->rcv_wnd; 21521da177e4SLinus Torvalds if (tp->rx_opt.rcv_wscale) { 21531da177e4SLinus Torvalds window = free_space; 21541da177e4SLinus Torvalds 21551da177e4SLinus Torvalds /* Advertise enough space so that it won't get scaled away. 21561da177e4SLinus Torvalds * Import case: prevent zero window announcement if 21571da177e4SLinus Torvalds * 1<<rcv_wscale > mss. 21581da177e4SLinus Torvalds */ 21591da177e4SLinus Torvalds if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window) 21601da177e4SLinus Torvalds window = (((window >> tp->rx_opt.rcv_wscale) + 1) 21611da177e4SLinus Torvalds << tp->rx_opt.rcv_wscale); 21621da177e4SLinus Torvalds } else { 21631da177e4SLinus Torvalds /* Get the largest window that is a nice multiple of mss. 21641da177e4SLinus Torvalds * Window clamp already applied above. 21651da177e4SLinus Torvalds * If our current window offering is within 1 mss of the 21661da177e4SLinus Torvalds * free space we just keep it. This prevents the divide 21671da177e4SLinus Torvalds * and multiply from happening most of the time. 21681da177e4SLinus Torvalds * We also don't do any window rounding when the free space 21691da177e4SLinus Torvalds * is too small. 21701da177e4SLinus Torvalds */ 21711da177e4SLinus Torvalds if (window <= free_space - mss || window > free_space) 21721da177e4SLinus Torvalds window = (free_space / mss) * mss; 217384565070SJohn Heffner else if (mss == full_space && 2174b92edbe0SEric Dumazet free_space > window + (full_space >> 1)) 217584565070SJohn Heffner window = free_space; 21761da177e4SLinus Torvalds } 21771da177e4SLinus Torvalds 21781da177e4SLinus Torvalds return window; 21791da177e4SLinus Torvalds } 21801da177e4SLinus Torvalds 21814a17fc3aSIlpo Järvinen /* Collapses two adjacent SKB's during retransmission. */ 21824a17fc3aSIlpo Järvinen static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) 21831da177e4SLinus Torvalds { 21841da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2185fe067e8aSDavid S. Miller struct sk_buff *next_skb = tcp_write_queue_next(sk, skb); 2186058dc334SIlpo Järvinen int skb_size, next_skb_size; 21871da177e4SLinus Torvalds 2188058dc334SIlpo Järvinen skb_size = skb->len; 2189058dc334SIlpo Järvinen next_skb_size = next_skb->len; 21901da177e4SLinus Torvalds 2191058dc334SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); 21921da177e4SLinus Torvalds 21936859d494SIlpo Järvinen tcp_highest_sack_combine(sk, next_skb, skb); 2194a6963a6bSIlpo Järvinen 2195fe067e8aSDavid S. Miller tcp_unlink_write_queue(next_skb, sk); 21961da177e4SLinus Torvalds 2197058dc334SIlpo Järvinen skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size), 21981a4e2d09SArnaldo Carvalho de Melo next_skb_size); 21991da177e4SLinus Torvalds 220052d570aaSJarek Poplawski if (next_skb->ip_summed == CHECKSUM_PARTIAL) 220152d570aaSJarek Poplawski skb->ip_summed = CHECKSUM_PARTIAL; 22021da177e4SLinus Torvalds 220384fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 22041da177e4SLinus Torvalds skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); 22051da177e4SLinus Torvalds 22061da177e4SLinus Torvalds /* Update sequence range on original skb. */ 22071da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 22081da177e4SLinus Torvalds 2209e6c7d085SIlpo Järvinen /* Merge over control information. This moves PSH/FIN etc. over */ 22104de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; 22111da177e4SLinus Torvalds 22121da177e4SLinus Torvalds /* All done, get rid of second SKB and account for it so 22131da177e4SLinus Torvalds * packet counting does not break. 22141da177e4SLinus Torvalds */ 22154828e7f4SIlpo Järvinen TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; 2216b7689205SIlpo Järvinen 2217b7689205SIlpo Järvinen /* changed transmit queue under us so clear hints */ 2218ef9da47cSIlpo Järvinen tcp_clear_retrans_hints_partial(tp); 2219ef9da47cSIlpo Järvinen if (next_skb == tp->retransmit_skb_hint) 2220ef9da47cSIlpo Järvinen tp->retransmit_skb_hint = skb; 2221b7689205SIlpo Järvinen 2222797108d1SIlpo Järvinen tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); 2223797108d1SIlpo Järvinen 22243ab224beSHideo Aoki sk_wmem_free_skb(sk, next_skb); 22251da177e4SLinus Torvalds } 22261da177e4SLinus Torvalds 222767edfef7SAndi Kleen /* Check if coalescing SKBs is legal. */ 2228a2a385d6SEric Dumazet static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) 22294a17fc3aSIlpo Järvinen { 22304a17fc3aSIlpo Järvinen if (tcp_skb_pcount(skb) > 1) 2231a2a385d6SEric Dumazet return false; 22324a17fc3aSIlpo Järvinen /* TODO: SACK collapsing could be used to remove this condition */ 22334a17fc3aSIlpo Järvinen if (skb_shinfo(skb)->nr_frags != 0) 2234a2a385d6SEric Dumazet return false; 22354a17fc3aSIlpo Järvinen if (skb_cloned(skb)) 2236a2a385d6SEric Dumazet return false; 22374a17fc3aSIlpo Järvinen if (skb == tcp_send_head(sk)) 2238a2a385d6SEric Dumazet return false; 22394a17fc3aSIlpo Järvinen /* Some heurestics for collapsing over SACK'd could be invented */ 22404a17fc3aSIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 2241a2a385d6SEric Dumazet return false; 22424a17fc3aSIlpo Järvinen 2243a2a385d6SEric Dumazet return true; 22444a17fc3aSIlpo Järvinen } 22454a17fc3aSIlpo Järvinen 224667edfef7SAndi Kleen /* Collapse packets in the retransmit queue to make to create 224767edfef7SAndi Kleen * less packets on the wire. This is only done on retransmission. 224867edfef7SAndi Kleen */ 22494a17fc3aSIlpo Järvinen static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, 22504a17fc3aSIlpo Järvinen int space) 22514a17fc3aSIlpo Järvinen { 22524a17fc3aSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 22534a17fc3aSIlpo Järvinen struct sk_buff *skb = to, *tmp; 2254a2a385d6SEric Dumazet bool first = true; 22554a17fc3aSIlpo Järvinen 22564a17fc3aSIlpo Järvinen if (!sysctl_tcp_retrans_collapse) 22574a17fc3aSIlpo Järvinen return; 22584de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 22594a17fc3aSIlpo Järvinen return; 22604a17fc3aSIlpo Järvinen 22614a17fc3aSIlpo Järvinen tcp_for_write_queue_from_safe(skb, tmp, sk) { 22624a17fc3aSIlpo Järvinen if (!tcp_can_collapse(sk, skb)) 22634a17fc3aSIlpo Järvinen break; 22644a17fc3aSIlpo Järvinen 22654a17fc3aSIlpo Järvinen space -= skb->len; 22664a17fc3aSIlpo Järvinen 22674a17fc3aSIlpo Järvinen if (first) { 2268a2a385d6SEric Dumazet first = false; 22694a17fc3aSIlpo Järvinen continue; 22704a17fc3aSIlpo Järvinen } 22714a17fc3aSIlpo Järvinen 22724a17fc3aSIlpo Järvinen if (space < 0) 22734a17fc3aSIlpo Järvinen break; 22744a17fc3aSIlpo Järvinen /* Punt if not enough space exists in the first SKB for 22754a17fc3aSIlpo Järvinen * the data in the second 22764a17fc3aSIlpo Järvinen */ 2277a21d4572SEric Dumazet if (skb->len > skb_availroom(to)) 22784a17fc3aSIlpo Järvinen break; 22794a17fc3aSIlpo Järvinen 22804a17fc3aSIlpo Järvinen if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) 22814a17fc3aSIlpo Järvinen break; 22824a17fc3aSIlpo Järvinen 22834a17fc3aSIlpo Järvinen tcp_collapse_retrans(sk, to); 22844a17fc3aSIlpo Järvinen } 22854a17fc3aSIlpo Järvinen } 22864a17fc3aSIlpo Järvinen 22871da177e4SLinus Torvalds /* This retransmits one SKB. Policy decisions and retransmit queue 22881da177e4SLinus Torvalds * state updates are done by the caller. Returns non-zero if an 22891da177e4SLinus Torvalds * error occurred which prevented the send. 22901da177e4SLinus Torvalds */ 229193b174adSYuchung Cheng int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) 22921da177e4SLinus Torvalds { 22931da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 22945d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 22957d227cd2SSridhar Samudrala unsigned int cur_mss; 22961da177e4SLinus Torvalds 22975d424d5aSJohn Heffner /* Inconslusive MTU probe */ 22985d424d5aSJohn Heffner if (icsk->icsk_mtup.probe_size) { 22995d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 23005d424d5aSJohn Heffner } 23015d424d5aSJohn Heffner 23021da177e4SLinus Torvalds /* Do not sent more than we queued. 1/4 is reserved for possible 2303caa20d9aSStephen Hemminger * copying overhead: fragmentation, tunneling, mangling etc. 23041da177e4SLinus Torvalds */ 23051da177e4SLinus Torvalds if (atomic_read(&sk->sk_wmem_alloc) > 23061da177e4SLinus Torvalds min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf)) 23071da177e4SLinus Torvalds return -EAGAIN; 23081da177e4SLinus Torvalds 23091da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 23101da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 23111da177e4SLinus Torvalds BUG(); 23121da177e4SLinus Torvalds if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 23131da177e4SLinus Torvalds return -ENOMEM; 23141da177e4SLinus Torvalds } 23151da177e4SLinus Torvalds 23167d227cd2SSridhar Samudrala if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 23177d227cd2SSridhar Samudrala return -EHOSTUNREACH; /* Routing failure or similar. */ 23187d227cd2SSridhar Samudrala 23190c54b85fSIlpo Järvinen cur_mss = tcp_current_mss(sk); 23207d227cd2SSridhar Samudrala 23211da177e4SLinus Torvalds /* If receiver has shrunk his window, and skb is out of 23221da177e4SLinus Torvalds * new window, do not retransmit it. The exception is the 23231da177e4SLinus Torvalds * case, when window is shrunk to zero. In this case 23241da177e4SLinus Torvalds * our retransmit serves as a zero window probe. 23251da177e4SLinus Torvalds */ 23269d4fb27dSJoe Perches if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && 23279d4fb27dSJoe Perches TCP_SKB_CB(skb)->seq != tp->snd_una) 23281da177e4SLinus Torvalds return -EAGAIN; 23291da177e4SLinus Torvalds 23301da177e4SLinus Torvalds if (skb->len > cur_mss) { 2331846998aeSDavid S. Miller if (tcp_fragment(sk, skb, cur_mss, cur_mss)) 23321da177e4SLinus Torvalds return -ENOMEM; /* We'll try again later. */ 233302276f3cSIlpo Järvinen } else { 23349eb9362eSIlpo Järvinen int oldpcount = tcp_skb_pcount(skb); 23359eb9362eSIlpo Järvinen 23369eb9362eSIlpo Järvinen if (unlikely(oldpcount > 1)) { 233702276f3cSIlpo Järvinen tcp_init_tso_segs(sk, skb, cur_mss); 23389eb9362eSIlpo Järvinen tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb)); 23399eb9362eSIlpo Järvinen } 23401da177e4SLinus Torvalds } 23411da177e4SLinus Torvalds 23421da177e4SLinus Torvalds tcp_retrans_try_collapse(sk, skb, cur_mss); 23431da177e4SLinus Torvalds 23441da177e4SLinus Torvalds /* Some Solaris stacks overoptimize and ignore the FIN on a 23451da177e4SLinus Torvalds * retransmit when old data is attached. So strip it off 23461da177e4SLinus Torvalds * since it is cheap to do so and saves bytes on the network. 23471da177e4SLinus Torvalds */ 23481da177e4SLinus Torvalds if (skb->len > 0 && 23494de075e0SEric Dumazet (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && 23501da177e4SLinus Torvalds tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) { 23511da177e4SLinus Torvalds if (!pskb_trim(skb, 0)) { 2352e870a8efSIlpo Järvinen /* Reuse, even though it does some unnecessary work */ 2353e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1, 23544de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags); 23551da177e4SLinus Torvalds skb->ip_summed = CHECKSUM_NONE; 23561da177e4SLinus Torvalds } 23571da177e4SLinus Torvalds } 23581da177e4SLinus Torvalds 23591da177e4SLinus Torvalds /* Make a copy, if the first transmission SKB clone we made 23601da177e4SLinus Torvalds * is still in somebody's hands, else make a clone. 23611da177e4SLinus Torvalds */ 23621da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 23631da177e4SLinus Torvalds 236450bceae9SThomas Graf /* make sure skb->data is aligned on arches that require it 236550bceae9SThomas Graf * and check if ack-trimming & collapsing extended the headroom 236650bceae9SThomas Graf * beyond what csum_start can cover. 236750bceae9SThomas Graf */ 236850bceae9SThomas Graf if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) || 236950bceae9SThomas Graf skb_headroom(skb) >= 0xFFFF)) { 2370117632e6SEric Dumazet struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER, 2371117632e6SEric Dumazet GFP_ATOMIC); 237293b174adSYuchung Cheng return nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) : 2373117632e6SEric Dumazet -ENOBUFS; 2374117632e6SEric Dumazet } else { 237593b174adSYuchung Cheng return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 2376117632e6SEric Dumazet } 237793b174adSYuchung Cheng } 237893b174adSYuchung Cheng 237993b174adSYuchung Cheng int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) 238093b174adSYuchung Cheng { 238193b174adSYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 238293b174adSYuchung Cheng int err = __tcp_retransmit_skb(sk, skb); 23831da177e4SLinus Torvalds 23841da177e4SLinus Torvalds if (err == 0) { 23851da177e4SLinus Torvalds /* Update global TCP statistics. */ 238681cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); 23871da177e4SLinus Torvalds 23881da177e4SLinus Torvalds tp->total_retrans++; 23891da177e4SLinus Torvalds 23901da177e4SLinus Torvalds #if FASTRETRANS_DEBUG > 0 23911da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 2392e87cc472SJoe Perches net_dbg_ratelimited("retrans_out leaked\n"); 23931da177e4SLinus Torvalds } 23941da177e4SLinus Torvalds #endif 2395b08d6cb2SIlpo Järvinen if (!tp->retrans_out) 2396b08d6cb2SIlpo Järvinen tp->lost_retrans_low = tp->snd_nxt; 23971da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 23981da177e4SLinus Torvalds tp->retrans_out += tcp_skb_pcount(skb); 23991da177e4SLinus Torvalds 24001da177e4SLinus Torvalds /* Save stamp of the first retransmit. */ 24011da177e4SLinus Torvalds if (!tp->retrans_stamp) 24021da177e4SLinus Torvalds tp->retrans_stamp = TCP_SKB_CB(skb)->when; 24031da177e4SLinus Torvalds 2404c24f691bSYuchung Cheng tp->undo_retrans += tcp_skb_pcount(skb); 24051da177e4SLinus Torvalds 24061da177e4SLinus Torvalds /* snd_nxt is stored to detect loss of retransmitted segment, 24071da177e4SLinus Torvalds * see tcp_input.c tcp_sacktag_write_queue(). 24081da177e4SLinus Torvalds */ 24091da177e4SLinus Torvalds TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt; 241024ab6becSYuchung Cheng } else { 241124ab6becSYuchung Cheng NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL); 24121da177e4SLinus Torvalds } 24131da177e4SLinus Torvalds return err; 24141da177e4SLinus Torvalds } 24151da177e4SLinus Torvalds 241667edfef7SAndi Kleen /* Check if we forward retransmits are possible in the current 241767edfef7SAndi Kleen * window/congestion state. 241867edfef7SAndi Kleen */ 2419a2a385d6SEric Dumazet static bool tcp_can_forward_retransmit(struct sock *sk) 2420b5afe7bcSIlpo Järvinen { 2421b5afe7bcSIlpo Järvinen const struct inet_connection_sock *icsk = inet_csk(sk); 2422cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 2423b5afe7bcSIlpo Järvinen 2424b5afe7bcSIlpo Järvinen /* Forward retransmissions are possible only during Recovery. */ 2425b5afe7bcSIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Recovery) 2426a2a385d6SEric Dumazet return false; 2427b5afe7bcSIlpo Järvinen 2428b5afe7bcSIlpo Järvinen /* No forward retransmissions in Reno are possible. */ 2429b5afe7bcSIlpo Järvinen if (tcp_is_reno(tp)) 2430a2a385d6SEric Dumazet return false; 2431b5afe7bcSIlpo Järvinen 2432b5afe7bcSIlpo Järvinen /* Yeah, we have to make difficult choice between forward transmission 2433b5afe7bcSIlpo Järvinen * and retransmission... Both ways have their merits... 2434b5afe7bcSIlpo Järvinen * 2435b5afe7bcSIlpo Järvinen * For now we do not retransmit anything, while we have some new 2436b5afe7bcSIlpo Järvinen * segments to send. In the other cases, follow rule 3 for 2437b5afe7bcSIlpo Järvinen * NextSeg() specified in RFC3517. 2438b5afe7bcSIlpo Järvinen */ 2439b5afe7bcSIlpo Järvinen 2440b5afe7bcSIlpo Järvinen if (tcp_may_send_now(sk)) 2441a2a385d6SEric Dumazet return false; 2442b5afe7bcSIlpo Järvinen 2443a2a385d6SEric Dumazet return true; 2444b5afe7bcSIlpo Järvinen } 2445b5afe7bcSIlpo Järvinen 24461da177e4SLinus Torvalds /* This gets called after a retransmit timeout, and the initially 24471da177e4SLinus Torvalds * retransmitted data is acknowledged. It tries to continue 24481da177e4SLinus Torvalds * resending the rest of the retransmit queue, until either 24491da177e4SLinus Torvalds * we've sent it all or the congestion window limit is reached. 24501da177e4SLinus Torvalds * If doing SACK, the first ACK which comes back for a timeout 24511da177e4SLinus Torvalds * based retransmit packet might feed us FACK information again. 24521da177e4SLinus Torvalds * If so, we use it to avoid unnecessarily retransmissions. 24531da177e4SLinus Torvalds */ 24541da177e4SLinus Torvalds void tcp_xmit_retransmit_queue(struct sock *sk) 24551da177e4SLinus Torvalds { 24566687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 24571da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 24581da177e4SLinus Torvalds struct sk_buff *skb; 24590e1c54c2SIlpo Järvinen struct sk_buff *hole = NULL; 2460618d9f25SIlpo Järvinen u32 last_lost; 246161eb55f4SIlpo Järvinen int mib_idx; 24620e1c54c2SIlpo Järvinen int fwd_rexmitting = 0; 24636a438bbeSStephen Hemminger 246445e77d31SIlpo Järvinen if (!tp->packets_out) 246545e77d31SIlpo Järvinen return; 246645e77d31SIlpo Järvinen 246708ebd172SIlpo Järvinen if (!tp->lost_out) 246808ebd172SIlpo Järvinen tp->retransmit_high = tp->snd_una; 246908ebd172SIlpo Järvinen 2470618d9f25SIlpo Järvinen if (tp->retransmit_skb_hint) { 24716a438bbeSStephen Hemminger skb = tp->retransmit_skb_hint; 2472618d9f25SIlpo Järvinen last_lost = TCP_SKB_CB(skb)->end_seq; 2473618d9f25SIlpo Järvinen if (after(last_lost, tp->retransmit_high)) 2474618d9f25SIlpo Järvinen last_lost = tp->retransmit_high; 2475618d9f25SIlpo Järvinen } else { 2476fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 2477618d9f25SIlpo Järvinen last_lost = tp->snd_una; 2478618d9f25SIlpo Järvinen } 24791da177e4SLinus Torvalds 2480fe067e8aSDavid S. Miller tcp_for_write_queue_from(skb, sk) { 24811da177e4SLinus Torvalds __u8 sacked = TCP_SKB_CB(skb)->sacked; 24821da177e4SLinus Torvalds 2483fe067e8aSDavid S. Miller if (skb == tcp_send_head(sk)) 2484fe067e8aSDavid S. Miller break; 24856a438bbeSStephen Hemminger /* we could do better than to assign each time */ 24860e1c54c2SIlpo Järvinen if (hole == NULL) 24876a438bbeSStephen Hemminger tp->retransmit_skb_hint = skb; 24886a438bbeSStephen Hemminger 24891da177e4SLinus Torvalds /* Assume this retransmit will generate 24901da177e4SLinus Torvalds * only one packet for congestion window 24911da177e4SLinus Torvalds * calculation purposes. This works because 24921da177e4SLinus Torvalds * tcp_retransmit_skb() will chop up the 24931da177e4SLinus Torvalds * packet to be MSS sized and all the 24941da177e4SLinus Torvalds * packet counting works out. 24951da177e4SLinus Torvalds */ 24961da177e4SLinus Torvalds if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) 24971da177e4SLinus Torvalds return; 24980e1c54c2SIlpo Järvinen 24990e1c54c2SIlpo Järvinen if (fwd_rexmitting) { 25000e1c54c2SIlpo Järvinen begin_fwd: 25010e1c54c2SIlpo Järvinen if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp))) 2502006f582cSIlpo Järvinen break; 25030e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFORWARDRETRANS; 25040e1c54c2SIlpo Järvinen 25050e1c54c2SIlpo Järvinen } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) { 2506618d9f25SIlpo Järvinen tp->retransmit_high = last_lost; 25070e1c54c2SIlpo Järvinen if (!tcp_can_forward_retransmit(sk)) 25080e1c54c2SIlpo Järvinen break; 25090e1c54c2SIlpo Järvinen /* Backtrack if necessary to non-L'ed skb */ 25100e1c54c2SIlpo Järvinen if (hole != NULL) { 25110e1c54c2SIlpo Järvinen skb = hole; 25120e1c54c2SIlpo Järvinen hole = NULL; 25130e1c54c2SIlpo Järvinen } 25140e1c54c2SIlpo Järvinen fwd_rexmitting = 1; 25150e1c54c2SIlpo Järvinen goto begin_fwd; 25160e1c54c2SIlpo Järvinen 25170e1c54c2SIlpo Järvinen } else if (!(sacked & TCPCB_LOST)) { 2518ac11ba75SIlpo Järvinen if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) 25190e1c54c2SIlpo Järvinen hole = skb; 252061eb55f4SIlpo Järvinen continue; 25211da177e4SLinus Torvalds 25220e1c54c2SIlpo Järvinen } else { 2523618d9f25SIlpo Järvinen last_lost = TCP_SKB_CB(skb)->end_seq; 25240e1c54c2SIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Loss) 25250e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFASTRETRANS; 25260e1c54c2SIlpo Järvinen else 25270e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; 25280e1c54c2SIlpo Järvinen } 25290e1c54c2SIlpo Järvinen 25300e1c54c2SIlpo Järvinen if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) 253161eb55f4SIlpo Järvinen continue; 253240b215e5SPavel Emelyanov 253324ab6becSYuchung Cheng if (tcp_retransmit_skb(sk, skb)) 25341da177e4SLinus Torvalds return; 253524ab6becSYuchung Cheng 2536de0744afSPavel Emelyanov NET_INC_STATS_BH(sock_net(sk), mib_idx); 25371da177e4SLinus Torvalds 2538684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 2539a262f0cdSNandita Dukkipati tp->prr_out += tcp_skb_pcount(skb); 2540a262f0cdSNandita Dukkipati 2541fe067e8aSDavid S. Miller if (skb == tcp_write_queue_head(sk)) 2542463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 25433f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, 25443f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 25451da177e4SLinus Torvalds } 25461da177e4SLinus Torvalds } 25471da177e4SLinus Torvalds 25481da177e4SLinus Torvalds /* Send a fin. The caller locks the socket for us. This cannot be 25491da177e4SLinus Torvalds * allowed to fail queueing a FIN frame under any circumstances. 25501da177e4SLinus Torvalds */ 25511da177e4SLinus Torvalds void tcp_send_fin(struct sock *sk) 25521da177e4SLinus Torvalds { 25531da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2554fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_write_queue_tail(sk); 25551da177e4SLinus Torvalds int mss_now; 25561da177e4SLinus Torvalds 25571da177e4SLinus Torvalds /* Optimization, tack on the FIN if we have a queue of 25581da177e4SLinus Torvalds * unsent frames. But be careful about outgoing SACKS 25591da177e4SLinus Torvalds * and IP options. 25601da177e4SLinus Torvalds */ 25610c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 25621da177e4SLinus Torvalds 2563fe067e8aSDavid S. Miller if (tcp_send_head(sk) != NULL) { 25644de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_FIN; 25651da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq++; 25661da177e4SLinus Torvalds tp->write_seq++; 25671da177e4SLinus Torvalds } else { 25681da177e4SLinus Torvalds /* Socket is locked, keep trying until memory is available. */ 25691da177e4SLinus Torvalds for (;;) { 2570aa133076SWu Fengguang skb = alloc_skb_fclone(MAX_TCP_HEADER, 2571aa133076SWu Fengguang sk->sk_allocation); 25721da177e4SLinus Torvalds if (skb) 25731da177e4SLinus Torvalds break; 25741da177e4SLinus Torvalds yield(); 25751da177e4SLinus Torvalds } 25761da177e4SLinus Torvalds 25771da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 25781da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 25791da177e4SLinus Torvalds /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 2580e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tp->write_seq, 2581a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_FIN); 25821da177e4SLinus Torvalds tcp_queue_skb(sk, skb); 25831da177e4SLinus Torvalds } 25849e412ba7SIlpo Järvinen __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF); 25851da177e4SLinus Torvalds } 25861da177e4SLinus Torvalds 25871da177e4SLinus Torvalds /* We get here when a process closes a file descriptor (either due to 25881da177e4SLinus Torvalds * an explicit close() or as a byproduct of exit()'ing) and there 25891da177e4SLinus Torvalds * was unread data in the receive queue. This behavior is recommended 259065bb723cSGerrit Renker * by RFC 2525, section 2.17. -DaveM 25911da177e4SLinus Torvalds */ 2592dd0fc66fSAl Viro void tcp_send_active_reset(struct sock *sk, gfp_t priority) 25931da177e4SLinus Torvalds { 25941da177e4SLinus Torvalds struct sk_buff *skb; 25951da177e4SLinus Torvalds 25961da177e4SLinus Torvalds /* NOTE: No TCP options attached and we never retransmit this. */ 25971da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, priority); 25981da177e4SLinus Torvalds if (!skb) { 25994e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 26001da177e4SLinus Torvalds return; 26011da177e4SLinus Torvalds } 26021da177e4SLinus Torvalds 26031da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 26041da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 2605e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), 2606a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_RST); 26071da177e4SLinus Torvalds /* Send it off. */ 26081da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2609dfb4b9dcSDavid S. Miller if (tcp_transmit_skb(sk, skb, 0, priority)) 26104e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 261126af65cbSSridhar Samudrala 261281cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); 26131da177e4SLinus Torvalds } 26141da177e4SLinus Torvalds 261567edfef7SAndi Kleen /* Send a crossed SYN-ACK during socket establishment. 261667edfef7SAndi Kleen * WARNING: This routine must only be called when we have already sent 26171da177e4SLinus Torvalds * a SYN packet that crossed the incoming SYN that caused this routine 26181da177e4SLinus Torvalds * to get called. If this assumption fails then the initial rcv_wnd 26191da177e4SLinus Torvalds * and rcv_wscale values will not be correct. 26201da177e4SLinus Torvalds */ 26211da177e4SLinus Torvalds int tcp_send_synack(struct sock *sk) 26221da177e4SLinus Torvalds { 26231da177e4SLinus Torvalds struct sk_buff *skb; 26241da177e4SLinus Torvalds 2625fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 26264de075e0SEric Dumazet if (skb == NULL || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 262791df42beSJoe Perches pr_debug("%s: wrong queue state\n", __func__); 26281da177e4SLinus Torvalds return -EFAULT; 26291da177e4SLinus Torvalds } 26304de075e0SEric Dumazet if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { 26311da177e4SLinus Torvalds if (skb_cloned(skb)) { 26321da177e4SLinus Torvalds struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC); 26331da177e4SLinus Torvalds if (nskb == NULL) 26341da177e4SLinus Torvalds return -ENOMEM; 2635fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 26361da177e4SLinus Torvalds skb_header_release(nskb); 2637fe067e8aSDavid S. Miller __tcp_add_write_queue_head(sk, nskb); 26383ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 26393ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 26403ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 26411da177e4SLinus Torvalds skb = nskb; 26421da177e4SLinus Torvalds } 26431da177e4SLinus Torvalds 26444de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; 26451da177e4SLinus Torvalds TCP_ECN_send_synack(tcp_sk(sk), skb); 26461da177e4SLinus Torvalds } 26471da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2648dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 26491da177e4SLinus Torvalds } 26501da177e4SLinus Torvalds 26514aea39c1SEric Dumazet /** 26524aea39c1SEric Dumazet * tcp_make_synack - Prepare a SYN-ACK. 26534aea39c1SEric Dumazet * sk: listener socket 26544aea39c1SEric Dumazet * dst: dst entry attached to the SYNACK 26554aea39c1SEric Dumazet * req: request_sock pointer 26564aea39c1SEric Dumazet * 26574aea39c1SEric Dumazet * Allocate one skb and build a SYNACK packet. 26584aea39c1SEric Dumazet * @dst is consumed : Caller should not use it again. 26594aea39c1SEric Dumazet */ 26601da177e4SLinus Torvalds struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst, 2661e6b4d113SWilliam Allen Simpson struct request_sock *req, 26628336886fSJerry Chu struct tcp_fastopen_cookie *foc) 26631da177e4SLinus Torvalds { 2664bd0388aeSWilliam Allen Simpson struct tcp_out_options opts; 26652e6599cbSArnaldo Carvalho de Melo struct inet_request_sock *ireq = inet_rsk(req); 26661da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 26671da177e4SLinus Torvalds struct tcphdr *th; 26681da177e4SLinus Torvalds struct sk_buff *skb; 2669cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 2670bd0388aeSWilliam Allen Simpson int tcp_header_size; 2671f5fff5dcSTom Quetchenbach int mss; 26721da177e4SLinus Torvalds 2673*eb8895deSPhil Oester skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC); 26744aea39c1SEric Dumazet if (unlikely(!skb)) { 26754aea39c1SEric Dumazet dst_release(dst); 26761da177e4SLinus Torvalds return NULL; 26774aea39c1SEric Dumazet } 26781da177e4SLinus Torvalds /* Reserve space for headers. */ 26791da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 26801da177e4SLinus Torvalds 26814aea39c1SEric Dumazet skb_dst_set(skb, dst); 2682ca10b9e9SEric Dumazet security_skb_owned_by(skb, sk); 26831da177e4SLinus Torvalds 26840dbaee3bSDavid S. Miller mss = dst_metric_advmss(dst); 2685f5fff5dcSTom Quetchenbach if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss) 2686f5fff5dcSTom Quetchenbach mss = tp->rx_opt.user_mss; 2687f5fff5dcSTom Quetchenbach 268833ad798cSAdam Langley if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */ 268933ad798cSAdam Langley __u8 rcv_wscale; 269033ad798cSAdam Langley /* Set this up on the first call only */ 269133ad798cSAdam Langley req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW); 2692e88c64f0SHagen Paul Pfeifer 2693e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 2694e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 2695e88c64f0SHagen Paul Pfeifer (req->window_clamp > tcp_full_space(sk) || req->window_clamp == 0)) 2696e88c64f0SHagen Paul Pfeifer req->window_clamp = tcp_full_space(sk); 2697e88c64f0SHagen Paul Pfeifer 269833ad798cSAdam Langley /* tcp_full_space because it is guaranteed to be the first packet */ 269933ad798cSAdam Langley tcp_select_initial_window(tcp_full_space(sk), 2700f5fff5dcSTom Quetchenbach mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0), 270133ad798cSAdam Langley &req->rcv_wnd, 270233ad798cSAdam Langley &req->window_clamp, 270333ad798cSAdam Langley ireq->wscale_ok, 270431d12926Slaurent chavey &rcv_wscale, 270531d12926Slaurent chavey dst_metric(dst, RTAX_INITRWND)); 270633ad798cSAdam Langley ireq->rcv_wscale = rcv_wscale; 270733ad798cSAdam Langley } 2708cfb6eeb4SYOSHIFUJI Hideaki 270933ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 27108b5f12d0SFlorian Westphal #ifdef CONFIG_SYN_COOKIES 27118b5f12d0SFlorian Westphal if (unlikely(req->cookie_ts)) 27128b5f12d0SFlorian Westphal TCP_SKB_CB(skb)->when = cookie_init_timestamp(req); 27138b5f12d0SFlorian Westphal else 27148b5f12d0SFlorian Westphal #endif 271533ad798cSAdam Langley TCP_SKB_CB(skb)->when = tcp_time_stamp; 27161a2c6181SChristoph Paasch tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, &md5, 27171a2c6181SChristoph Paasch foc) + sizeof(*th); 271833ad798cSAdam Langley 2719aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 2720aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 27211da177e4SLinus Torvalds 2722aa8223c7SArnaldo Carvalho de Melo th = tcp_hdr(skb); 27231da177e4SLinus Torvalds memset(th, 0, sizeof(struct tcphdr)); 27241da177e4SLinus Torvalds th->syn = 1; 27251da177e4SLinus Torvalds th->ack = 1; 27261da177e4SLinus Torvalds TCP_ECN_make_synack(req, th); 2727a3116ac5SKOVACS Krisztian th->source = ireq->loc_port; 27282e6599cbSArnaldo Carvalho de Melo th->dest = ireq->rmt_port; 2729e870a8efSIlpo Järvinen /* Setting of flags are superfluous here for callers (and ECE is 2730e870a8efSIlpo Järvinen * not even correctly set) 2731e870a8efSIlpo Järvinen */ 2732e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn, 2733a3433f35SChangli Gao TCPHDR_SYN | TCPHDR_ACK); 27344957faadSWilliam Allen Simpson 27351da177e4SLinus Torvalds th->seq = htonl(TCP_SKB_CB(skb)->seq); 27368336886fSJerry Chu /* XXX data is queued and acked as is. No buffer/window check */ 27378336886fSJerry Chu th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt); 27381da177e4SLinus Torvalds 27391da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 2740600ff0c2SIlpo Järvinen th->window = htons(min(req->rcv_wnd, 65535U)); 2741bd0388aeSWilliam Allen Simpson tcp_options_write((__be32 *)(th + 1), tp, &opts); 27421da177e4SLinus Torvalds th->doff = (tcp_header_size >> 2); 2743aa2ea058STom Herbert TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, tcp_skb_pcount(skb)); 2744cfb6eeb4SYOSHIFUJI Hideaki 2745cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 2746cfb6eeb4SYOSHIFUJI Hideaki /* Okay, we have all we need - do the md5 hash if needed */ 2747cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 2748bd0388aeSWilliam Allen Simpson tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, 274949a72dfbSAdam Langley md5, NULL, req, skb); 2750cfb6eeb4SYOSHIFUJI Hideaki } 2751cfb6eeb4SYOSHIFUJI Hideaki #endif 2752cfb6eeb4SYOSHIFUJI Hideaki 27531da177e4SLinus Torvalds return skb; 27541da177e4SLinus Torvalds } 27554bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_make_synack); 27561da177e4SLinus Torvalds 275767edfef7SAndi Kleen /* Do all connect socket setups that can be done AF independent. */ 2758370816aeSPavel Emelyanov void tcp_connect_init(struct sock *sk) 27591da177e4SLinus Torvalds { 2760cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 27611da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 27621da177e4SLinus Torvalds __u8 rcv_wscale; 27631da177e4SLinus Torvalds 27641da177e4SLinus Torvalds /* We'll fix this up when we get a response from the other end. 27651da177e4SLinus Torvalds * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 27661da177e4SLinus Torvalds */ 27671da177e4SLinus Torvalds tp->tcp_header_len = sizeof(struct tcphdr) + 2768bb5b7c11SDavid S. Miller (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0); 27691da177e4SLinus Torvalds 2770cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 2771cfb6eeb4SYOSHIFUJI Hideaki if (tp->af_specific->md5_lookup(sk, sk) != NULL) 2772cfb6eeb4SYOSHIFUJI Hideaki tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 2773cfb6eeb4SYOSHIFUJI Hideaki #endif 2774cfb6eeb4SYOSHIFUJI Hideaki 27751da177e4SLinus Torvalds /* If user gave his TCP_MAXSEG, record it to clamp */ 27761da177e4SLinus Torvalds if (tp->rx_opt.user_mss) 27771da177e4SLinus Torvalds tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 27781da177e4SLinus Torvalds tp->max_window = 0; 27795d424d5aSJohn Heffner tcp_mtup_init(sk); 27801da177e4SLinus Torvalds tcp_sync_mss(sk, dst_mtu(dst)); 27811da177e4SLinus Torvalds 27821da177e4SLinus Torvalds if (!tp->window_clamp) 27831da177e4SLinus Torvalds tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 27840dbaee3bSDavid S. Miller tp->advmss = dst_metric_advmss(dst); 2785f5fff5dcSTom Quetchenbach if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss) 2786f5fff5dcSTom Quetchenbach tp->advmss = tp->rx_opt.user_mss; 2787f5fff5dcSTom Quetchenbach 27881da177e4SLinus Torvalds tcp_initialize_rcv_mss(sk); 27891da177e4SLinus Torvalds 2790e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 2791e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 2792e88c64f0SHagen Paul Pfeifer (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) 2793e88c64f0SHagen Paul Pfeifer tp->window_clamp = tcp_full_space(sk); 2794e88c64f0SHagen Paul Pfeifer 27951da177e4SLinus Torvalds tcp_select_initial_window(tcp_full_space(sk), 27961da177e4SLinus Torvalds tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 27971da177e4SLinus Torvalds &tp->rcv_wnd, 27981da177e4SLinus Torvalds &tp->window_clamp, 2799bb5b7c11SDavid S. Miller sysctl_tcp_window_scaling, 280031d12926Slaurent chavey &rcv_wscale, 280131d12926Slaurent chavey dst_metric(dst, RTAX_INITRWND)); 28021da177e4SLinus Torvalds 28031da177e4SLinus Torvalds tp->rx_opt.rcv_wscale = rcv_wscale; 28041da177e4SLinus Torvalds tp->rcv_ssthresh = tp->rcv_wnd; 28051da177e4SLinus Torvalds 28061da177e4SLinus Torvalds sk->sk_err = 0; 28071da177e4SLinus Torvalds sock_reset_flag(sk, SOCK_DONE); 28081da177e4SLinus Torvalds tp->snd_wnd = 0; 2809ee7537b6SHantzis Fotis tcp_init_wl(tp, 0); 28101da177e4SLinus Torvalds tp->snd_una = tp->write_seq; 28111da177e4SLinus Torvalds tp->snd_sml = tp->write_seq; 281233f5f57eSIlpo Järvinen tp->snd_up = tp->write_seq; 2813370816aeSPavel Emelyanov tp->snd_nxt = tp->write_seq; 2814ee995283SPavel Emelyanov 2815ee995283SPavel Emelyanov if (likely(!tp->repair)) 28161da177e4SLinus Torvalds tp->rcv_nxt = 0; 2817c7781a6eSAndrew Vagin else 2818c7781a6eSAndrew Vagin tp->rcv_tstamp = tcp_time_stamp; 2819ee995283SPavel Emelyanov tp->rcv_wup = tp->rcv_nxt; 2820ee995283SPavel Emelyanov tp->copied_seq = tp->rcv_nxt; 28211da177e4SLinus Torvalds 2822463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; 2823463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_retransmits = 0; 28241da177e4SLinus Torvalds tcp_clear_retrans(tp); 28251da177e4SLinus Torvalds } 28261da177e4SLinus Torvalds 2827783237e8SYuchung Cheng static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) 2828783237e8SYuchung Cheng { 2829783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 2830783237e8SYuchung Cheng struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 2831783237e8SYuchung Cheng 2832783237e8SYuchung Cheng tcb->end_seq += skb->len; 2833783237e8SYuchung Cheng skb_header_release(skb); 2834783237e8SYuchung Cheng __tcp_add_write_queue_tail(sk, skb); 2835783237e8SYuchung Cheng sk->sk_wmem_queued += skb->truesize; 2836783237e8SYuchung Cheng sk_mem_charge(sk, skb->truesize); 2837783237e8SYuchung Cheng tp->write_seq = tcb->end_seq; 2838783237e8SYuchung Cheng tp->packets_out += tcp_skb_pcount(skb); 2839783237e8SYuchung Cheng } 2840783237e8SYuchung Cheng 2841783237e8SYuchung Cheng /* Build and send a SYN with data and (cached) Fast Open cookie. However, 2842783237e8SYuchung Cheng * queue a data-only packet after the regular SYN, such that regular SYNs 2843783237e8SYuchung Cheng * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges 2844783237e8SYuchung Cheng * only the SYN sequence, the data are retransmitted in the first ACK. 2845783237e8SYuchung Cheng * If cookie is not cached or other error occurs, falls back to send a 2846783237e8SYuchung Cheng * regular SYN with Fast Open cookie request option. 2847783237e8SYuchung Cheng */ 2848783237e8SYuchung Cheng static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) 2849783237e8SYuchung Cheng { 2850783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 2851783237e8SYuchung Cheng struct tcp_fastopen_request *fo = tp->fastopen_req; 2852aab48743SYuchung Cheng int syn_loss = 0, space, i, err = 0, iovlen = fo->data->msg_iovlen; 2853783237e8SYuchung Cheng struct sk_buff *syn_data = NULL, *data; 2854aab48743SYuchung Cheng unsigned long last_syn_loss = 0; 2855783237e8SYuchung Cheng 285667da22d2SYuchung Cheng tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ 2857aab48743SYuchung Cheng tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie, 2858aab48743SYuchung Cheng &syn_loss, &last_syn_loss); 2859aab48743SYuchung Cheng /* Recurring FO SYN losses: revert to regular handshake temporarily */ 2860aab48743SYuchung Cheng if (syn_loss > 1 && 2861aab48743SYuchung Cheng time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) { 2862aab48743SYuchung Cheng fo->cookie.len = -1; 2863aab48743SYuchung Cheng goto fallback; 2864aab48743SYuchung Cheng } 2865aab48743SYuchung Cheng 286667da22d2SYuchung Cheng if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE) 286767da22d2SYuchung Cheng fo->cookie.len = -1; 286867da22d2SYuchung Cheng else if (fo->cookie.len <= 0) 2869783237e8SYuchung Cheng goto fallback; 2870783237e8SYuchung Cheng 2871783237e8SYuchung Cheng /* MSS for SYN-data is based on cached MSS and bounded by PMTU and 2872783237e8SYuchung Cheng * user-MSS. Reserve maximum option space for middleboxes that add 2873783237e8SYuchung Cheng * private TCP options. The cost is reduced data space in SYN :( 2874783237e8SYuchung Cheng */ 2875783237e8SYuchung Cheng if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp) 2876783237e8SYuchung Cheng tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 28771b63edd6SYuchung Cheng space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) - 2878783237e8SYuchung Cheng MAX_TCP_OPTION_SPACE; 2879783237e8SYuchung Cheng 2880783237e8SYuchung Cheng syn_data = skb_copy_expand(syn, skb_headroom(syn), space, 2881783237e8SYuchung Cheng sk->sk_allocation); 2882783237e8SYuchung Cheng if (syn_data == NULL) 2883783237e8SYuchung Cheng goto fallback; 2884783237e8SYuchung Cheng 2885783237e8SYuchung Cheng for (i = 0; i < iovlen && syn_data->len < space; ++i) { 2886783237e8SYuchung Cheng struct iovec *iov = &fo->data->msg_iov[i]; 2887783237e8SYuchung Cheng unsigned char __user *from = iov->iov_base; 2888783237e8SYuchung Cheng int len = iov->iov_len; 2889783237e8SYuchung Cheng 2890783237e8SYuchung Cheng if (syn_data->len + len > space) 2891783237e8SYuchung Cheng len = space - syn_data->len; 2892783237e8SYuchung Cheng else if (i + 1 == iovlen) 2893783237e8SYuchung Cheng /* No more data pending in inet_wait_for_connect() */ 2894783237e8SYuchung Cheng fo->data = NULL; 2895783237e8SYuchung Cheng 2896783237e8SYuchung Cheng if (skb_add_data(syn_data, from, len)) 2897783237e8SYuchung Cheng goto fallback; 2898783237e8SYuchung Cheng } 2899783237e8SYuchung Cheng 2900783237e8SYuchung Cheng /* Queue a data-only packet after the regular SYN for retransmission */ 2901783237e8SYuchung Cheng data = pskb_copy(syn_data, sk->sk_allocation); 2902783237e8SYuchung Cheng if (data == NULL) 2903783237e8SYuchung Cheng goto fallback; 2904783237e8SYuchung Cheng TCP_SKB_CB(data)->seq++; 2905783237e8SYuchung Cheng TCP_SKB_CB(data)->tcp_flags &= ~TCPHDR_SYN; 2906783237e8SYuchung Cheng TCP_SKB_CB(data)->tcp_flags = (TCPHDR_ACK|TCPHDR_PSH); 2907783237e8SYuchung Cheng tcp_connect_queue_skb(sk, data); 2908783237e8SYuchung Cheng fo->copied = data->len; 2909783237e8SYuchung Cheng 2910783237e8SYuchung Cheng if (tcp_transmit_skb(sk, syn_data, 0, sk->sk_allocation) == 0) { 291167da22d2SYuchung Cheng tp->syn_data = (fo->copied > 0); 2912783237e8SYuchung Cheng NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE); 2913783237e8SYuchung Cheng goto done; 2914783237e8SYuchung Cheng } 2915783237e8SYuchung Cheng syn_data = NULL; 2916783237e8SYuchung Cheng 2917783237e8SYuchung Cheng fallback: 2918783237e8SYuchung Cheng /* Send a regular SYN with Fast Open cookie request option */ 2919783237e8SYuchung Cheng if (fo->cookie.len > 0) 2920783237e8SYuchung Cheng fo->cookie.len = 0; 2921783237e8SYuchung Cheng err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); 2922783237e8SYuchung Cheng if (err) 2923783237e8SYuchung Cheng tp->syn_fastopen = 0; 2924783237e8SYuchung Cheng kfree_skb(syn_data); 2925783237e8SYuchung Cheng done: 2926783237e8SYuchung Cheng fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ 2927783237e8SYuchung Cheng return err; 2928783237e8SYuchung Cheng } 2929783237e8SYuchung Cheng 293067edfef7SAndi Kleen /* Build a SYN and send it off. */ 29311da177e4SLinus Torvalds int tcp_connect(struct sock *sk) 29321da177e4SLinus Torvalds { 29331da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 29341da177e4SLinus Torvalds struct sk_buff *buff; 2935ee586811SEric Paris int err; 29361da177e4SLinus Torvalds 29371da177e4SLinus Torvalds tcp_connect_init(sk); 29381da177e4SLinus Torvalds 29392b916477SAndrey Vagin if (unlikely(tp->repair)) { 29402b916477SAndrey Vagin tcp_finish_connect(sk, NULL); 29412b916477SAndrey Vagin return 0; 29422b916477SAndrey Vagin } 29432b916477SAndrey Vagin 2944d179cd12SDavid S. Miller buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation); 29451da177e4SLinus Torvalds if (unlikely(buff == NULL)) 29461da177e4SLinus Torvalds return -ENOBUFS; 29471da177e4SLinus Torvalds 29481da177e4SLinus Torvalds /* Reserve space for headers. */ 29491da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 29501da177e4SLinus Torvalds 2951a3433f35SChangli Gao tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); 2952783237e8SYuchung Cheng tp->retrans_stamp = TCP_SKB_CB(buff)->when = tcp_time_stamp; 2953783237e8SYuchung Cheng tcp_connect_queue_skb(sk, buff); 2954e870a8efSIlpo Järvinen TCP_ECN_send_syn(sk, buff); 29551da177e4SLinus Torvalds 2956783237e8SYuchung Cheng /* Send off SYN; include data in Fast Open. */ 2957783237e8SYuchung Cheng err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : 2958783237e8SYuchung Cheng tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); 2959ee586811SEric Paris if (err == -ECONNREFUSED) 2960ee586811SEric Paris return err; 2961bd37a088SWei Yongjun 2962bd37a088SWei Yongjun /* We change tp->snd_nxt after the tcp_transmit_skb() call 2963bd37a088SWei Yongjun * in order to make this packet get counted in tcpOutSegs. 2964bd37a088SWei Yongjun */ 2965bd37a088SWei Yongjun tp->snd_nxt = tp->write_seq; 2966bd37a088SWei Yongjun tp->pushed_seq = tp->write_seq; 296781cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); 29681da177e4SLinus Torvalds 29691da177e4SLinus Torvalds /* Timer for repeating the SYN until an answer. */ 29703f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 29713f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 29721da177e4SLinus Torvalds return 0; 29731da177e4SLinus Torvalds } 29744bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_connect); 29751da177e4SLinus Torvalds 29761da177e4SLinus Torvalds /* Send out a delayed ack, the caller does the policy checking 29771da177e4SLinus Torvalds * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 29781da177e4SLinus Torvalds * for details. 29791da177e4SLinus Torvalds */ 29801da177e4SLinus Torvalds void tcp_send_delayed_ack(struct sock *sk) 29811da177e4SLinus Torvalds { 2982463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 2983463c84b9SArnaldo Carvalho de Melo int ato = icsk->icsk_ack.ato; 29841da177e4SLinus Torvalds unsigned long timeout; 29851da177e4SLinus Torvalds 29861da177e4SLinus Torvalds if (ato > TCP_DELACK_MIN) { 2987463c84b9SArnaldo Carvalho de Melo const struct tcp_sock *tp = tcp_sk(sk); 29881da177e4SLinus Torvalds int max_ato = HZ / 2; 29891da177e4SLinus Torvalds 2990056834d9SIlpo Järvinen if (icsk->icsk_ack.pingpong || 2991056834d9SIlpo Järvinen (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 29921da177e4SLinus Torvalds max_ato = TCP_DELACK_MAX; 29931da177e4SLinus Torvalds 29941da177e4SLinus Torvalds /* Slow path, intersegment interval is "high". */ 29951da177e4SLinus Torvalds 29961da177e4SLinus Torvalds /* If some rtt estimate is known, use it to bound delayed ack. 2997463c84b9SArnaldo Carvalho de Melo * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 29981da177e4SLinus Torvalds * directly. 29991da177e4SLinus Torvalds */ 30001da177e4SLinus Torvalds if (tp->srtt) { 30011da177e4SLinus Torvalds int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN); 30021da177e4SLinus Torvalds 30031da177e4SLinus Torvalds if (rtt < max_ato) 30041da177e4SLinus Torvalds max_ato = rtt; 30051da177e4SLinus Torvalds } 30061da177e4SLinus Torvalds 30071da177e4SLinus Torvalds ato = min(ato, max_ato); 30081da177e4SLinus Torvalds } 30091da177e4SLinus Torvalds 30101da177e4SLinus Torvalds /* Stay within the limit we were given */ 30111da177e4SLinus Torvalds timeout = jiffies + ato; 30121da177e4SLinus Torvalds 30131da177e4SLinus Torvalds /* Use new timeout only if there wasn't a older one earlier. */ 3014463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 30151da177e4SLinus Torvalds /* If delack timer was blocked or is about to expire, 30161da177e4SLinus Torvalds * send ACK now. 30171da177e4SLinus Torvalds */ 3018463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.blocked || 3019463c84b9SArnaldo Carvalho de Melo time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 30201da177e4SLinus Torvalds tcp_send_ack(sk); 30211da177e4SLinus Torvalds return; 30221da177e4SLinus Torvalds } 30231da177e4SLinus Torvalds 3024463c84b9SArnaldo Carvalho de Melo if (!time_before(timeout, icsk->icsk_ack.timeout)) 3025463c84b9SArnaldo Carvalho de Melo timeout = icsk->icsk_ack.timeout; 30261da177e4SLinus Torvalds } 3027463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 3028463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.timeout = timeout; 3029463c84b9SArnaldo Carvalho de Melo sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 30301da177e4SLinus Torvalds } 30311da177e4SLinus Torvalds 30321da177e4SLinus Torvalds /* This routine sends an ack and also updates the window. */ 30331da177e4SLinus Torvalds void tcp_send_ack(struct sock *sk) 30341da177e4SLinus Torvalds { 30351da177e4SLinus Torvalds struct sk_buff *buff; 30361da177e4SLinus Torvalds 3037058dc334SIlpo Järvinen /* If we have been reset, we may not send again. */ 3038058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3039058dc334SIlpo Järvinen return; 3040058dc334SIlpo Järvinen 30411da177e4SLinus Torvalds /* We are not putting this on the write queue, so 30421da177e4SLinus Torvalds * tcp_transmit_skb() will set the ownership to this 30431da177e4SLinus Torvalds * sock. 30441da177e4SLinus Torvalds */ 304599a1dec7SMel Gorman buff = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC)); 30461da177e4SLinus Torvalds if (buff == NULL) { 3047463c84b9SArnaldo Carvalho de Melo inet_csk_schedule_ack(sk); 3048463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; 30493f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 30503f421baaSArnaldo Carvalho de Melo TCP_DELACK_MAX, TCP_RTO_MAX); 30511da177e4SLinus Torvalds return; 30521da177e4SLinus Torvalds } 30531da177e4SLinus Torvalds 30541da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 30551da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 3056a3433f35SChangli Gao tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); 30571da177e4SLinus Torvalds 30581da177e4SLinus Torvalds /* Send it off, this clears delayed acks for us. */ 30591da177e4SLinus Torvalds TCP_SKB_CB(buff)->when = tcp_time_stamp; 306099a1dec7SMel Gorman tcp_transmit_skb(sk, buff, 0, sk_gfp_atomic(sk, GFP_ATOMIC)); 30611da177e4SLinus Torvalds } 30621da177e4SLinus Torvalds 30631da177e4SLinus Torvalds /* This routine sends a packet with an out of date sequence 30641da177e4SLinus Torvalds * number. It assumes the other end will try to ack it. 30651da177e4SLinus Torvalds * 30661da177e4SLinus Torvalds * Question: what should we make while urgent mode? 30671da177e4SLinus Torvalds * 4.4BSD forces sending single byte of data. We cannot send 30681da177e4SLinus Torvalds * out of window data, because we have SND.NXT==SND.MAX... 30691da177e4SLinus Torvalds * 30701da177e4SLinus Torvalds * Current solution: to send TWO zero-length segments in urgent mode: 30711da177e4SLinus Torvalds * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 30721da177e4SLinus Torvalds * out-of-date with SND.UNA-1 to probe window. 30731da177e4SLinus Torvalds */ 30741da177e4SLinus Torvalds static int tcp_xmit_probe_skb(struct sock *sk, int urgent) 30751da177e4SLinus Torvalds { 30761da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 30771da177e4SLinus Torvalds struct sk_buff *skb; 30781da177e4SLinus Torvalds 30791da177e4SLinus Torvalds /* We don't queue it, tcp_transmit_skb() sets ownership. */ 308099a1dec7SMel Gorman skb = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC)); 30811da177e4SLinus Torvalds if (skb == NULL) 30821da177e4SLinus Torvalds return -1; 30831da177e4SLinus Torvalds 30841da177e4SLinus Torvalds /* Reserve space for headers and set control bits. */ 30851da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 30861da177e4SLinus Torvalds /* Use a previous sequence. This should cause the other 30871da177e4SLinus Torvalds * end to send an ack. Don't queue or clone SKB, just 30881da177e4SLinus Torvalds * send it. 30891da177e4SLinus Torvalds */ 3090a3433f35SChangli Gao tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); 30911da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 3092dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC); 30931da177e4SLinus Torvalds } 30941da177e4SLinus Torvalds 3095ee995283SPavel Emelyanov void tcp_send_window_probe(struct sock *sk) 3096ee995283SPavel Emelyanov { 3097ee995283SPavel Emelyanov if (sk->sk_state == TCP_ESTABLISHED) { 3098ee995283SPavel Emelyanov tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; 3099c0e88ff0SPavel Emelyanov tcp_sk(sk)->snd_nxt = tcp_sk(sk)->write_seq; 3100ee995283SPavel Emelyanov tcp_xmit_probe_skb(sk, 0); 3101ee995283SPavel Emelyanov } 3102ee995283SPavel Emelyanov } 3103ee995283SPavel Emelyanov 310467edfef7SAndi Kleen /* Initiate keepalive or window probe from timer. */ 31051da177e4SLinus Torvalds int tcp_write_wakeup(struct sock *sk) 31061da177e4SLinus Torvalds { 31071da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 31081da177e4SLinus Torvalds struct sk_buff *skb; 31091da177e4SLinus Torvalds 3110058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3111058dc334SIlpo Järvinen return -1; 3112058dc334SIlpo Järvinen 3113fe067e8aSDavid S. Miller if ((skb = tcp_send_head(sk)) != NULL && 311490840defSIlpo Järvinen before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { 31151da177e4SLinus Torvalds int err; 31160c54b85fSIlpo Järvinen unsigned int mss = tcp_current_mss(sk); 311790840defSIlpo Järvinen unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 31181da177e4SLinus Torvalds 31191da177e4SLinus Torvalds if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 31201da177e4SLinus Torvalds tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 31211da177e4SLinus Torvalds 31221da177e4SLinus Torvalds /* We are probing the opening of a window 31231da177e4SLinus Torvalds * but the window size is != 0 31241da177e4SLinus Torvalds * must have been a result SWS avoidance ( sender ) 31251da177e4SLinus Torvalds */ 31261da177e4SLinus Torvalds if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 31271da177e4SLinus Torvalds skb->len > mss) { 31281da177e4SLinus Torvalds seg_size = min(seg_size, mss); 31294de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 3130846998aeSDavid S. Miller if (tcp_fragment(sk, skb, seg_size, mss)) 31311da177e4SLinus Torvalds return -1; 31321da177e4SLinus Torvalds } else if (!tcp_skb_pcount(skb)) 3133846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss); 31341da177e4SLinus Torvalds 31354de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 31361da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 3137dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 313866f5fe62SIlpo Järvinen if (!err) 313966f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 31401da177e4SLinus Torvalds return err; 31411da177e4SLinus Torvalds } else { 314233f5f57eSIlpo Järvinen if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) 31434828e7f4SIlpo Järvinen tcp_xmit_probe_skb(sk, 1); 31441da177e4SLinus Torvalds return tcp_xmit_probe_skb(sk, 0); 31451da177e4SLinus Torvalds } 31461da177e4SLinus Torvalds } 31471da177e4SLinus Torvalds 31481da177e4SLinus Torvalds /* A window probe timeout has occurred. If window is not closed send 31491da177e4SLinus Torvalds * a partial packet else a zero probe. 31501da177e4SLinus Torvalds */ 31511da177e4SLinus Torvalds void tcp_send_probe0(struct sock *sk) 31521da177e4SLinus Torvalds { 3153463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 31541da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 31551da177e4SLinus Torvalds int err; 31561da177e4SLinus Torvalds 31571da177e4SLinus Torvalds err = tcp_write_wakeup(sk); 31581da177e4SLinus Torvalds 3159fe067e8aSDavid S. Miller if (tp->packets_out || !tcp_send_head(sk)) { 31601da177e4SLinus Torvalds /* Cancel probe timer, if it is not required. */ 31616687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 0; 3162463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff = 0; 31631da177e4SLinus Torvalds return; 31641da177e4SLinus Torvalds } 31651da177e4SLinus Torvalds 31661da177e4SLinus Torvalds if (err <= 0) { 3167463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_backoff < sysctl_tcp_retries2) 3168463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff++; 31696687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out++; 3170463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 31713f421baaSArnaldo Carvalho de Melo min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), 31723f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 31731da177e4SLinus Torvalds } else { 31741da177e4SLinus Torvalds /* If packet was not sent due to local congestion, 31756687e988SArnaldo Carvalho de Melo * do not backoff and do not remember icsk_probes_out. 31761da177e4SLinus Torvalds * Let local senders to fight for local resources. 31771da177e4SLinus Torvalds * 31781da177e4SLinus Torvalds * Use accumulated backoff yet. 31791da177e4SLinus Torvalds */ 31806687e988SArnaldo Carvalho de Melo if (!icsk->icsk_probes_out) 31816687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 1; 3182463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 3183463c84b9SArnaldo Carvalho de Melo min(icsk->icsk_rto << icsk->icsk_backoff, 31843f421baaSArnaldo Carvalho de Melo TCP_RESOURCE_PROBE_INTERVAL), 31853f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 31861da177e4SLinus Torvalds } 31871da177e4SLinus Torvalds } 3188