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 68c9bee3b7SEric Dumazet unsigned int sysctl_tcp_notsent_lowat __read_mostly = UINT_MAX; 69c9bee3b7SEric Dumazet EXPORT_SYMBOL(sysctl_tcp_notsent_lowat); 70c9bee3b7SEric Dumazet 7146d3ceabSEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 7246d3ceabSEric Dumazet int push_one, gfp_t gfp); 73519855c5SWilliam Allen Simpson 7467edfef7SAndi Kleen /* Account for new data that has been sent to the network. */ 75cf533ea5SEric Dumazet static void tcp_event_new_data_sent(struct sock *sk, const struct sk_buff *skb) 766ff03ac3SIlpo Järvinen { 776ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 786ff03ac3SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 7966f5fe62SIlpo Järvinen unsigned int prior_packets = tp->packets_out; 809e412ba7SIlpo Järvinen 81fe067e8aSDavid S. Miller tcp_advance_send_head(sk, skb); 821da177e4SLinus Torvalds tp->snd_nxt = TCP_SKB_CB(skb)->end_seq; 838512430eSIlpo Järvinen 8466f5fe62SIlpo Järvinen tp->packets_out += tcp_skb_pcount(skb); 856ba8a3b1SNandita Dukkipati if (!prior_packets || icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS || 866a5dc9e5SEric Dumazet icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) { 87750ea2baSYuchung Cheng tcp_rearm_rto(sk); 881da177e4SLinus Torvalds } 896a5dc9e5SEric Dumazet } 901da177e4SLinus Torvalds 911da177e4SLinus Torvalds /* SND.NXT, if window was not shrunk. 921da177e4SLinus Torvalds * If window has been shrunk, what should we make? It is not clear at all. 931da177e4SLinus Torvalds * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-( 941da177e4SLinus Torvalds * Anything in between SND.UNA...SND.UNA+SND.WND also can be already 951da177e4SLinus Torvalds * invalid. OK, let's make this for now: 961da177e4SLinus Torvalds */ 97cf533ea5SEric Dumazet static inline __u32 tcp_acceptable_seq(const struct sock *sk) 981da177e4SLinus Torvalds { 99cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1009e412ba7SIlpo Järvinen 10190840defSIlpo Järvinen if (!before(tcp_wnd_end(tp), tp->snd_nxt)) 1021da177e4SLinus Torvalds return tp->snd_nxt; 1031da177e4SLinus Torvalds else 10490840defSIlpo Järvinen return tcp_wnd_end(tp); 1051da177e4SLinus Torvalds } 1061da177e4SLinus Torvalds 1071da177e4SLinus Torvalds /* Calculate mss to advertise in SYN segment. 1081da177e4SLinus Torvalds * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that: 1091da177e4SLinus Torvalds * 1101da177e4SLinus Torvalds * 1. It is independent of path mtu. 1111da177e4SLinus Torvalds * 2. Ideally, it is maximal possible segment size i.e. 65535-40. 1121da177e4SLinus Torvalds * 3. For IPv4 it is reasonable to calculate it from maximal MTU of 1131da177e4SLinus Torvalds * attached devices, because some buggy hosts are confused by 1141da177e4SLinus Torvalds * large MSS. 1151da177e4SLinus Torvalds * 4. We do not make 3, we advertise MSS, calculated from first 1161da177e4SLinus Torvalds * hop device mtu, but allow to raise it to ip_rt_min_advmss. 1171da177e4SLinus Torvalds * This may be overridden via information stored in routing table. 1181da177e4SLinus Torvalds * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible, 1191da177e4SLinus Torvalds * probably even Jumbo". 1201da177e4SLinus Torvalds */ 1211da177e4SLinus Torvalds static __u16 tcp_advertise_mss(struct sock *sk) 1221da177e4SLinus Torvalds { 1231da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 124cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1251da177e4SLinus Torvalds int mss = tp->advmss; 1261da177e4SLinus Torvalds 1270dbaee3bSDavid S. Miller if (dst) { 1280dbaee3bSDavid S. Miller unsigned int metric = dst_metric_advmss(dst); 1290dbaee3bSDavid S. Miller 1300dbaee3bSDavid S. Miller if (metric < mss) { 1310dbaee3bSDavid S. Miller mss = metric; 1321da177e4SLinus Torvalds tp->advmss = mss; 1331da177e4SLinus Torvalds } 1340dbaee3bSDavid S. Miller } 1351da177e4SLinus Torvalds 1361da177e4SLinus Torvalds return (__u16)mss; 1371da177e4SLinus Torvalds } 1381da177e4SLinus Torvalds 1391da177e4SLinus Torvalds /* RFC2861. Reset CWND after idle period longer RTO to "restart window". 1401da177e4SLinus Torvalds * This is the first part of cwnd validation mechanism. */ 141cf533ea5SEric Dumazet static void tcp_cwnd_restart(struct sock *sk, const struct dst_entry *dst) 1421da177e4SLinus Torvalds { 143463c84b9SArnaldo Carvalho de Melo struct tcp_sock *tp = tcp_sk(sk); 1441da177e4SLinus Torvalds s32 delta = tcp_time_stamp - tp->lsndtime; 1451da177e4SLinus Torvalds u32 restart_cwnd = tcp_init_cwnd(tp, dst); 1461da177e4SLinus Torvalds u32 cwnd = tp->snd_cwnd; 1471da177e4SLinus Torvalds 1486687e988SArnaldo Carvalho de Melo tcp_ca_event(sk, CA_EVENT_CWND_RESTART); 1491da177e4SLinus Torvalds 1506687e988SArnaldo Carvalho de Melo tp->snd_ssthresh = tcp_current_ssthresh(sk); 1511da177e4SLinus Torvalds restart_cwnd = min(restart_cwnd, cwnd); 1521da177e4SLinus Torvalds 153463c84b9SArnaldo Carvalho de Melo while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd) 1541da177e4SLinus Torvalds cwnd >>= 1; 1551da177e4SLinus Torvalds tp->snd_cwnd = max(cwnd, restart_cwnd); 1561da177e4SLinus Torvalds tp->snd_cwnd_stamp = tcp_time_stamp; 1571da177e4SLinus Torvalds tp->snd_cwnd_used = 0; 1581da177e4SLinus Torvalds } 1591da177e4SLinus Torvalds 16067edfef7SAndi Kleen /* Congestion state accounting after a packet has been sent. */ 16140efc6faSStephen Hemminger static void tcp_event_data_sent(struct tcp_sock *tp, 162cf533ea5SEric Dumazet struct sock *sk) 1631da177e4SLinus Torvalds { 164463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 165463c84b9SArnaldo Carvalho de Melo const u32 now = tcp_time_stamp; 166bcefe17cSCong Wang const struct dst_entry *dst = __sk_dst_get(sk); 1671da177e4SLinus Torvalds 16835089bb2SDavid S. Miller if (sysctl_tcp_slow_start_after_idle && 16935089bb2SDavid S. Miller (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto)) 170463c84b9SArnaldo Carvalho de Melo tcp_cwnd_restart(sk, __sk_dst_get(sk)); 1711da177e4SLinus Torvalds 1721da177e4SLinus Torvalds tp->lsndtime = now; 1731da177e4SLinus Torvalds 1741da177e4SLinus Torvalds /* If it is a reply for ato after last received 1751da177e4SLinus Torvalds * packet, enter pingpong mode. 1761da177e4SLinus Torvalds */ 177bcefe17cSCong Wang if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato && 178bcefe17cSCong Wang (!dst || !dst_metric(dst, RTAX_QUICKACK))) 179463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pingpong = 1; 1801da177e4SLinus Torvalds } 1811da177e4SLinus Torvalds 18267edfef7SAndi Kleen /* Account for an ACK we sent. */ 18340efc6faSStephen Hemminger static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts) 1841da177e4SLinus Torvalds { 185463c84b9SArnaldo Carvalho de Melo tcp_dec_quickack_mode(sk, pkts); 186463c84b9SArnaldo Carvalho de Melo inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); 1871da177e4SLinus Torvalds } 1881da177e4SLinus Torvalds 18985f16525SYuchung Cheng 19085f16525SYuchung Cheng u32 tcp_default_init_rwnd(u32 mss) 19185f16525SYuchung Cheng { 19285f16525SYuchung Cheng /* Initial receive window should be twice of TCP_INIT_CWND to 1939ef71e0cSWeiping Pan * enable proper sending of new unsent data during fast recovery 19485f16525SYuchung Cheng * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a 19585f16525SYuchung Cheng * limit when mss is larger than 1460. 19685f16525SYuchung Cheng */ 19785f16525SYuchung Cheng u32 init_rwnd = TCP_INIT_CWND * 2; 19885f16525SYuchung Cheng 19985f16525SYuchung Cheng if (mss > 1460) 20085f16525SYuchung Cheng init_rwnd = max((1460 * init_rwnd) / mss, 2U); 20185f16525SYuchung Cheng return init_rwnd; 20285f16525SYuchung Cheng } 20385f16525SYuchung Cheng 2041da177e4SLinus Torvalds /* Determine a window scaling and initial window to offer. 2051da177e4SLinus Torvalds * Based on the assumption that the given amount of space 2061da177e4SLinus Torvalds * will be offered. Store the results in the tp structure. 2071da177e4SLinus Torvalds * NOTE: for smooth operation initial space offering should 2081da177e4SLinus Torvalds * be a multiple of mss if possible. We assume here that mss >= 1. 2091da177e4SLinus Torvalds * This MUST be enforced by all callers. 2101da177e4SLinus Torvalds */ 2111da177e4SLinus Torvalds void tcp_select_initial_window(int __space, __u32 mss, 2121da177e4SLinus Torvalds __u32 *rcv_wnd, __u32 *window_clamp, 21331d12926Slaurent chavey int wscale_ok, __u8 *rcv_wscale, 21431d12926Slaurent chavey __u32 init_rcv_wnd) 2151da177e4SLinus Torvalds { 2161da177e4SLinus Torvalds unsigned int space = (__space < 0 ? 0 : __space); 2171da177e4SLinus Torvalds 2181da177e4SLinus Torvalds /* If no clamp set the clamp to the max possible scaled window */ 2191da177e4SLinus Torvalds if (*window_clamp == 0) 2201da177e4SLinus Torvalds (*window_clamp) = (65535 << 14); 2211da177e4SLinus Torvalds space = min(*window_clamp, space); 2221da177e4SLinus Torvalds 2231da177e4SLinus Torvalds /* Quantize space offering to a multiple of mss if possible. */ 2241da177e4SLinus Torvalds if (space > mss) 2251da177e4SLinus Torvalds space = (space / mss) * mss; 2261da177e4SLinus Torvalds 2271da177e4SLinus Torvalds /* NOTE: offering an initial window larger than 32767 22815d99e02SRick Jones * will break some buggy TCP stacks. If the admin tells us 22915d99e02SRick Jones * it is likely we could be speaking with such a buggy stack 23015d99e02SRick Jones * we will truncate our initial window offering to 32K-1 23115d99e02SRick Jones * unless the remote has sent us a window scaling option, 23215d99e02SRick Jones * which we interpret as a sign the remote TCP is not 23315d99e02SRick Jones * misinterpreting the window field as a signed quantity. 2341da177e4SLinus Torvalds */ 23515d99e02SRick Jones if (sysctl_tcp_workaround_signed_windows) 2361da177e4SLinus Torvalds (*rcv_wnd) = min(space, MAX_TCP_WINDOW); 23715d99e02SRick Jones else 23815d99e02SRick Jones (*rcv_wnd) = space; 23915d99e02SRick Jones 2401da177e4SLinus Torvalds (*rcv_wscale) = 0; 2411da177e4SLinus Torvalds if (wscale_ok) { 2421da177e4SLinus Torvalds /* Set window scaling on max possible window 2431da177e4SLinus Torvalds * See RFC1323 for an explanation of the limit to 14 2441da177e4SLinus Torvalds */ 2451da177e4SLinus Torvalds space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max); 246316c1592SStephen Hemminger space = min_t(u32, space, *window_clamp); 2471da177e4SLinus Torvalds while (space > 65535 && (*rcv_wscale) < 14) { 2481da177e4SLinus Torvalds space >>= 1; 2491da177e4SLinus Torvalds (*rcv_wscale)++; 2501da177e4SLinus Torvalds } 2511da177e4SLinus Torvalds } 2521da177e4SLinus Torvalds 2531da177e4SLinus Torvalds if (mss > (1 << *rcv_wscale)) { 25485f16525SYuchung Cheng if (!init_rcv_wnd) /* Use default unless specified otherwise */ 25585f16525SYuchung Cheng init_rcv_wnd = tcp_default_init_rwnd(mss); 256b1afde60SNandita Dukkipati *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss); 2571da177e4SLinus Torvalds } 2581da177e4SLinus Torvalds 2591da177e4SLinus Torvalds /* Set the clamp no higher than max representable value */ 2601da177e4SLinus Torvalds (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp); 2611da177e4SLinus Torvalds } 2624bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_select_initial_window); 2631da177e4SLinus Torvalds 2641da177e4SLinus Torvalds /* Chose a new window to advertise, update state in tcp_sock for the 2651da177e4SLinus Torvalds * socket, and return result with RFC1323 scaling applied. The return 2661da177e4SLinus Torvalds * value can be stuffed directly into th->window for an outgoing 2671da177e4SLinus Torvalds * frame. 2681da177e4SLinus Torvalds */ 26940efc6faSStephen Hemminger static u16 tcp_select_window(struct sock *sk) 2701da177e4SLinus Torvalds { 2711da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2721da177e4SLinus Torvalds u32 cur_win = tcp_receive_window(tp); 2731da177e4SLinus Torvalds u32 new_win = __tcp_select_window(sk); 2741da177e4SLinus Torvalds 2751da177e4SLinus Torvalds /* Never shrink the offered window */ 2761da177e4SLinus Torvalds if (new_win < cur_win) { 2771da177e4SLinus Torvalds /* Danger Will Robinson! 2781da177e4SLinus Torvalds * Don't update rcv_wup/rcv_wnd here or else 2791da177e4SLinus Torvalds * we will not be able to advertise a zero 2801da177e4SLinus Torvalds * window in time. --DaveM 2811da177e4SLinus Torvalds * 2821da177e4SLinus Torvalds * Relax Will Robinson. 2831da177e4SLinus Torvalds */ 284607bfbf2SPatrick McHardy new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale); 2851da177e4SLinus Torvalds } 2861da177e4SLinus Torvalds tp->rcv_wnd = new_win; 2871da177e4SLinus Torvalds tp->rcv_wup = tp->rcv_nxt; 2881da177e4SLinus Torvalds 2891da177e4SLinus Torvalds /* Make sure we do not exceed the maximum possible 2901da177e4SLinus Torvalds * scaled window. 2911da177e4SLinus Torvalds */ 29215d99e02SRick Jones if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows) 2931da177e4SLinus Torvalds new_win = min(new_win, MAX_TCP_WINDOW); 2941da177e4SLinus Torvalds else 2951da177e4SLinus Torvalds new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); 2961da177e4SLinus Torvalds 2971da177e4SLinus Torvalds /* RFC1323 scaling applied */ 2981da177e4SLinus Torvalds new_win >>= tp->rx_opt.rcv_wscale; 2991da177e4SLinus Torvalds 3001da177e4SLinus Torvalds /* If we advertise zero window, disable fast path. */ 3011da177e4SLinus Torvalds if (new_win == 0) 3021da177e4SLinus Torvalds tp->pred_flags = 0; 3031da177e4SLinus Torvalds 3041da177e4SLinus Torvalds return new_win; 3051da177e4SLinus Torvalds } 3061da177e4SLinus Torvalds 30767edfef7SAndi Kleen /* Packet ECN state for a SYN-ACK */ 308cf533ea5SEric Dumazet static inline void TCP_ECN_send_synack(const struct tcp_sock *tp, struct sk_buff *skb) 309bdf1ee5dSIlpo Järvinen { 3104de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR; 311bdf1ee5dSIlpo Järvinen if (!(tp->ecn_flags & TCP_ECN_OK)) 3124de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE; 313bdf1ee5dSIlpo Järvinen } 314bdf1ee5dSIlpo Järvinen 31567edfef7SAndi Kleen /* Packet ECN state for a SYN. */ 316bdf1ee5dSIlpo Järvinen static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb) 317bdf1ee5dSIlpo Järvinen { 318bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 319bdf1ee5dSIlpo Järvinen 320bdf1ee5dSIlpo Järvinen tp->ecn_flags = 0; 3215d134f1cSHannes Frederic Sowa if (sock_net(sk)->ipv4.sysctl_tcp_ecn == 1) { 3224de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR; 323bdf1ee5dSIlpo Järvinen tp->ecn_flags = TCP_ECN_OK; 324bdf1ee5dSIlpo Järvinen } 325bdf1ee5dSIlpo Järvinen } 326bdf1ee5dSIlpo Järvinen 327bdf1ee5dSIlpo Järvinen static __inline__ void 328cf533ea5SEric Dumazet TCP_ECN_make_synack(const struct request_sock *req, struct tcphdr *th) 329bdf1ee5dSIlpo Järvinen { 330bdf1ee5dSIlpo Järvinen if (inet_rsk(req)->ecn_ok) 331bdf1ee5dSIlpo Järvinen th->ece = 1; 332bdf1ee5dSIlpo Järvinen } 333bdf1ee5dSIlpo Järvinen 33467edfef7SAndi Kleen /* Set up ECN state for a packet on a ESTABLISHED socket that is about to 33567edfef7SAndi Kleen * be sent. 33667edfef7SAndi Kleen */ 337bdf1ee5dSIlpo Järvinen static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb, 338bdf1ee5dSIlpo Järvinen int tcp_header_len) 339bdf1ee5dSIlpo Järvinen { 340bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 341bdf1ee5dSIlpo Järvinen 342bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_OK) { 343bdf1ee5dSIlpo Järvinen /* Not-retransmitted data segment: set ECT and inject CWR. */ 344bdf1ee5dSIlpo Järvinen if (skb->len != tcp_header_len && 345bdf1ee5dSIlpo Järvinen !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { 346bdf1ee5dSIlpo Järvinen INET_ECN_xmit(sk); 347bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) { 348bdf1ee5dSIlpo Järvinen tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; 349bdf1ee5dSIlpo Järvinen tcp_hdr(skb)->cwr = 1; 350bdf1ee5dSIlpo Järvinen skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; 351bdf1ee5dSIlpo Järvinen } 352bdf1ee5dSIlpo Järvinen } else { 353bdf1ee5dSIlpo Järvinen /* ACK or retransmitted segment: clear ECT|CE */ 354bdf1ee5dSIlpo Järvinen INET_ECN_dontxmit(sk); 355bdf1ee5dSIlpo Järvinen } 356bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) 357bdf1ee5dSIlpo Järvinen tcp_hdr(skb)->ece = 1; 358bdf1ee5dSIlpo Järvinen } 359bdf1ee5dSIlpo Järvinen } 360bdf1ee5dSIlpo Järvinen 361e870a8efSIlpo Järvinen /* Constructs common control bits of non-data skb. If SYN/FIN is present, 362e870a8efSIlpo Järvinen * auto increment end seqno. 363e870a8efSIlpo Järvinen */ 364e870a8efSIlpo Järvinen static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags) 365e870a8efSIlpo Järvinen { 3662e8e18efSDavid S. Miller skb->ip_summed = CHECKSUM_PARTIAL; 367e870a8efSIlpo Järvinen skb->csum = 0; 368e870a8efSIlpo Järvinen 3694de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags; 370e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->sacked = 0; 371e870a8efSIlpo Järvinen 372e870a8efSIlpo Järvinen skb_shinfo(skb)->gso_segs = 1; 373e870a8efSIlpo Järvinen skb_shinfo(skb)->gso_size = 0; 374e870a8efSIlpo Järvinen skb_shinfo(skb)->gso_type = 0; 375e870a8efSIlpo Järvinen 376e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->seq = seq; 377a3433f35SChangli Gao if (flags & (TCPHDR_SYN | TCPHDR_FIN)) 378e870a8efSIlpo Järvinen seq++; 379e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->end_seq = seq; 380e870a8efSIlpo Järvinen } 381e870a8efSIlpo Järvinen 382a2a385d6SEric Dumazet static inline bool tcp_urg_mode(const struct tcp_sock *tp) 38333f5f57eSIlpo Järvinen { 38433f5f57eSIlpo Järvinen return tp->snd_una != tp->snd_up; 38533f5f57eSIlpo Järvinen } 38633f5f57eSIlpo Järvinen 38733ad798cSAdam Langley #define OPTION_SACK_ADVERTISE (1 << 0) 38833ad798cSAdam Langley #define OPTION_TS (1 << 1) 38933ad798cSAdam Langley #define OPTION_MD5 (1 << 2) 39089e95a61SOri Finkelman #define OPTION_WSCALE (1 << 3) 3912100c8d2SYuchung Cheng #define OPTION_FAST_OPEN_COOKIE (1 << 8) 39233ad798cSAdam Langley 39333ad798cSAdam Langley struct tcp_out_options { 3942100c8d2SYuchung Cheng u16 options; /* bit field of OPTION_* */ 3952100c8d2SYuchung Cheng u16 mss; /* 0 to disable */ 39633ad798cSAdam Langley u8 ws; /* window scale, 0 to disable */ 39733ad798cSAdam Langley u8 num_sack_blocks; /* number of SACK blocks to include */ 398bd0388aeSWilliam Allen Simpson u8 hash_size; /* bytes in hash_location */ 399bd0388aeSWilliam Allen Simpson __u8 *hash_location; /* temporary pointer, overloaded */ 4002100c8d2SYuchung Cheng __u32 tsval, tsecr; /* need to include OPTION_TS */ 4012100c8d2SYuchung Cheng struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */ 40233ad798cSAdam Langley }; 40333ad798cSAdam Langley 40467edfef7SAndi Kleen /* Write previously computed TCP options to the packet. 40567edfef7SAndi Kleen * 40667edfef7SAndi Kleen * Beware: Something in the Internet is very sensitive to the ordering of 407fd6149d3SIlpo Järvinen * TCP options, we learned this through the hard way, so be careful here. 408fd6149d3SIlpo Järvinen * Luckily we can at least blame others for their non-compliance but from 409fd6149d3SIlpo Järvinen * inter-operatibility perspective it seems that we're somewhat stuck with 410fd6149d3SIlpo Järvinen * the ordering which we have been using if we want to keep working with 411fd6149d3SIlpo Järvinen * those broken things (not that it currently hurts anybody as there isn't 412fd6149d3SIlpo Järvinen * particular reason why the ordering would need to be changed). 413fd6149d3SIlpo Järvinen * 414fd6149d3SIlpo Järvinen * At least SACK_PERM as the first option is known to lead to a disaster 415fd6149d3SIlpo Järvinen * (but it may well be that other scenarios fail similarly). 416fd6149d3SIlpo Järvinen */ 41733ad798cSAdam Langley static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp, 418bd0388aeSWilliam Allen Simpson struct tcp_out_options *opts) 419bd0388aeSWilliam Allen Simpson { 4202100c8d2SYuchung Cheng u16 options = opts->options; /* mungable copy */ 421bd0388aeSWilliam Allen Simpson 422bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_MD5 & options)) { 4231a2c6181SChristoph Paasch *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | 4241a2c6181SChristoph Paasch (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG); 425bd0388aeSWilliam Allen Simpson /* overload cookie hash location */ 426bd0388aeSWilliam Allen Simpson opts->hash_location = (__u8 *)ptr; 42733ad798cSAdam Langley ptr += 4; 42833ad798cSAdam Langley } 42933ad798cSAdam Langley 430fd6149d3SIlpo Järvinen if (unlikely(opts->mss)) { 431fd6149d3SIlpo Järvinen *ptr++ = htonl((TCPOPT_MSS << 24) | 432fd6149d3SIlpo Järvinen (TCPOLEN_MSS << 16) | 433fd6149d3SIlpo Järvinen opts->mss); 434fd6149d3SIlpo Järvinen } 435fd6149d3SIlpo Järvinen 436bd0388aeSWilliam Allen Simpson if (likely(OPTION_TS & options)) { 437bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 43833ad798cSAdam Langley *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | 43933ad798cSAdam Langley (TCPOLEN_SACK_PERM << 16) | 44033ad798cSAdam Langley (TCPOPT_TIMESTAMP << 8) | 44133ad798cSAdam Langley TCPOLEN_TIMESTAMP); 442bd0388aeSWilliam Allen Simpson options &= ~OPTION_SACK_ADVERTISE; 44333ad798cSAdam Langley } else { 444496c98dfSYOSHIFUJI Hideaki *ptr++ = htonl((TCPOPT_NOP << 24) | 44540efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 44640efc6faSStephen Hemminger (TCPOPT_TIMESTAMP << 8) | 44740efc6faSStephen Hemminger TCPOLEN_TIMESTAMP); 44840efc6faSStephen Hemminger } 44933ad798cSAdam Langley *ptr++ = htonl(opts->tsval); 45033ad798cSAdam Langley *ptr++ = htonl(opts->tsecr); 45133ad798cSAdam Langley } 45233ad798cSAdam Langley 453bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 45433ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 45533ad798cSAdam Langley (TCPOPT_NOP << 16) | 45633ad798cSAdam Langley (TCPOPT_SACK_PERM << 8) | 45733ad798cSAdam Langley TCPOLEN_SACK_PERM); 45833ad798cSAdam Langley } 45933ad798cSAdam Langley 460bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_WSCALE & options)) { 46133ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 46233ad798cSAdam Langley (TCPOPT_WINDOW << 16) | 46333ad798cSAdam Langley (TCPOLEN_WINDOW << 8) | 46433ad798cSAdam Langley opts->ws); 46533ad798cSAdam Langley } 46633ad798cSAdam Langley 46733ad798cSAdam Langley if (unlikely(opts->num_sack_blocks)) { 46833ad798cSAdam Langley struct tcp_sack_block *sp = tp->rx_opt.dsack ? 46933ad798cSAdam Langley tp->duplicate_sack : tp->selective_acks; 47040efc6faSStephen Hemminger int this_sack; 47140efc6faSStephen Hemminger 47240efc6faSStephen Hemminger *ptr++ = htonl((TCPOPT_NOP << 24) | 47340efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 47440efc6faSStephen Hemminger (TCPOPT_SACK << 8) | 47533ad798cSAdam Langley (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * 47640efc6faSStephen Hemminger TCPOLEN_SACK_PERBLOCK))); 4772de979bdSStephen Hemminger 47833ad798cSAdam Langley for (this_sack = 0; this_sack < opts->num_sack_blocks; 47933ad798cSAdam Langley ++this_sack) { 48040efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].start_seq); 48140efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].end_seq); 48240efc6faSStephen Hemminger } 4832de979bdSStephen Hemminger 48440efc6faSStephen Hemminger tp->rx_opt.dsack = 0; 48540efc6faSStephen Hemminger } 4862100c8d2SYuchung Cheng 4872100c8d2SYuchung Cheng if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) { 4882100c8d2SYuchung Cheng struct tcp_fastopen_cookie *foc = opts->fastopen_cookie; 4892100c8d2SYuchung Cheng 4902100c8d2SYuchung Cheng *ptr++ = htonl((TCPOPT_EXP << 24) | 4912100c8d2SYuchung Cheng ((TCPOLEN_EXP_FASTOPEN_BASE + foc->len) << 16) | 4922100c8d2SYuchung Cheng TCPOPT_FASTOPEN_MAGIC); 4932100c8d2SYuchung Cheng 4942100c8d2SYuchung Cheng memcpy(ptr, foc->val, foc->len); 4952100c8d2SYuchung Cheng if ((foc->len & 3) == 2) { 4962100c8d2SYuchung Cheng u8 *align = ((u8 *)ptr) + foc->len; 4972100c8d2SYuchung Cheng align[0] = align[1] = TCPOPT_NOP; 4982100c8d2SYuchung Cheng } 4992100c8d2SYuchung Cheng ptr += (foc->len + 3) >> 2; 5002100c8d2SYuchung Cheng } 50140efc6faSStephen Hemminger } 50240efc6faSStephen Hemminger 50367edfef7SAndi Kleen /* Compute TCP options for SYN packets. This is not the final 50467edfef7SAndi Kleen * network wire format yet. 50567edfef7SAndi Kleen */ 50695c96174SEric Dumazet static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb, 50733ad798cSAdam Langley struct tcp_out_options *opts, 508cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 509cf533ea5SEric Dumazet { 51033ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 51195c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 512783237e8SYuchung Cheng struct tcp_fastopen_request *fastopen = tp->fastopen_req; 51333ad798cSAdam Langley 514cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 51533ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 51633ad798cSAdam Langley if (*md5) { 51733ad798cSAdam Langley opts->options |= OPTION_MD5; 518bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 519cfb6eeb4SYOSHIFUJI Hideaki } 52033ad798cSAdam Langley #else 52133ad798cSAdam Langley *md5 = NULL; 522cfb6eeb4SYOSHIFUJI Hideaki #endif 52333ad798cSAdam Langley 52433ad798cSAdam Langley /* We always get an MSS option. The option bytes which will be seen in 52533ad798cSAdam Langley * normal data packets should timestamps be used, must be in the MSS 52633ad798cSAdam Langley * advertised. But we subtract them from tp->mss_cache so that 52733ad798cSAdam Langley * calculations in tcp_sendmsg are simpler etc. So account for this 52833ad798cSAdam Langley * fact here if necessary. If we don't do this correctly, as a 52933ad798cSAdam Langley * receiver we won't recognize data packets as being full sized when we 53033ad798cSAdam Langley * should, and thus we won't abide by the delayed ACK rules correctly. 53133ad798cSAdam Langley * SACKs don't matter, we never delay an ACK when we have any of those 53233ad798cSAdam Langley * going out. */ 53333ad798cSAdam Langley opts->mss = tcp_advertise_mss(sk); 534bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 53533ad798cSAdam Langley 536bb5b7c11SDavid S. Miller if (likely(sysctl_tcp_timestamps && *md5 == NULL)) { 53733ad798cSAdam Langley opts->options |= OPTION_TS; 538ee684b6fSAndrey Vagin opts->tsval = TCP_SKB_CB(skb)->when + tp->tsoffset; 53933ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 540bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 54133ad798cSAdam Langley } 542bb5b7c11SDavid S. Miller if (likely(sysctl_tcp_window_scaling)) { 54333ad798cSAdam Langley opts->ws = tp->rx_opt.rcv_wscale; 54489e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 545bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 54633ad798cSAdam Langley } 547bb5b7c11SDavid S. Miller if (likely(sysctl_tcp_sack)) { 54833ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 549b32d1310SDavid S. Miller if (unlikely(!(OPTION_TS & opts->options))) 550bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 55133ad798cSAdam Langley } 55233ad798cSAdam Langley 553783237e8SYuchung Cheng if (fastopen && fastopen->cookie.len >= 0) { 554783237e8SYuchung Cheng u32 need = TCPOLEN_EXP_FASTOPEN_BASE + fastopen->cookie.len; 555783237e8SYuchung Cheng need = (need + 3) & ~3U; /* Align to 32 bits */ 556783237e8SYuchung Cheng if (remaining >= need) { 557783237e8SYuchung Cheng opts->options |= OPTION_FAST_OPEN_COOKIE; 558783237e8SYuchung Cheng opts->fastopen_cookie = &fastopen->cookie; 559783237e8SYuchung Cheng remaining -= need; 560783237e8SYuchung Cheng tp->syn_fastopen = 1; 561783237e8SYuchung Cheng } 562783237e8SYuchung Cheng } 563bd0388aeSWilliam Allen Simpson 564bd0388aeSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 56533ad798cSAdam Langley } 56633ad798cSAdam Langley 56767edfef7SAndi Kleen /* Set up TCP options for SYN-ACKs. */ 56895c96174SEric Dumazet static unsigned int tcp_synack_options(struct sock *sk, 56933ad798cSAdam Langley struct request_sock *req, 57095c96174SEric Dumazet unsigned int mss, struct sk_buff *skb, 57133ad798cSAdam Langley struct tcp_out_options *opts, 5724957faadSWilliam Allen Simpson struct tcp_md5sig_key **md5, 5738336886fSJerry Chu struct tcp_fastopen_cookie *foc) 5744957faadSWilliam Allen Simpson { 57533ad798cSAdam Langley struct inet_request_sock *ireq = inet_rsk(req); 57695c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 57733ad798cSAdam Langley 57833ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 57933ad798cSAdam Langley *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req); 58033ad798cSAdam Langley if (*md5) { 58133ad798cSAdam Langley opts->options |= OPTION_MD5; 5824957faadSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 5834957faadSWilliam Allen Simpson 5844957faadSWilliam Allen Simpson /* We can't fit any SACK blocks in a packet with MD5 + TS 5854957faadSWilliam Allen Simpson * options. There was discussion about disabling SACK 5864957faadSWilliam Allen Simpson * rather than TS in order to fit in better with old, 5874957faadSWilliam Allen Simpson * buggy kernels, but that was deemed to be unnecessary. 5884957faadSWilliam Allen Simpson */ 589de213e5eSEric Dumazet ireq->tstamp_ok &= !ireq->sack_ok; 59033ad798cSAdam Langley } 59133ad798cSAdam Langley #else 59233ad798cSAdam Langley *md5 = NULL; 59333ad798cSAdam Langley #endif 59433ad798cSAdam Langley 5954957faadSWilliam Allen Simpson /* We always send an MSS option. */ 59633ad798cSAdam Langley opts->mss = mss; 5974957faadSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 59833ad798cSAdam Langley 59933ad798cSAdam Langley if (likely(ireq->wscale_ok)) { 60033ad798cSAdam Langley opts->ws = ireq->rcv_wscale; 60189e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 6024957faadSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 60333ad798cSAdam Langley } 604de213e5eSEric Dumazet if (likely(ireq->tstamp_ok)) { 60533ad798cSAdam Langley opts->options |= OPTION_TS; 60633ad798cSAdam Langley opts->tsval = TCP_SKB_CB(skb)->when; 60733ad798cSAdam Langley opts->tsecr = req->ts_recent; 6084957faadSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 60933ad798cSAdam Langley } 61033ad798cSAdam Langley if (likely(ireq->sack_ok)) { 61133ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 612de213e5eSEric Dumazet if (unlikely(!ireq->tstamp_ok)) 6134957faadSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 61433ad798cSAdam Langley } 6158336886fSJerry Chu if (foc != NULL) { 6168336886fSJerry Chu u32 need = TCPOLEN_EXP_FASTOPEN_BASE + foc->len; 6178336886fSJerry Chu need = (need + 3) & ~3U; /* Align to 32 bits */ 6188336886fSJerry Chu if (remaining >= need) { 6198336886fSJerry Chu opts->options |= OPTION_FAST_OPEN_COOKIE; 6208336886fSJerry Chu opts->fastopen_cookie = foc; 6218336886fSJerry Chu remaining -= need; 6228336886fSJerry Chu } 6238336886fSJerry Chu } 6244957faadSWilliam Allen Simpson 6254957faadSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 62633ad798cSAdam Langley } 62733ad798cSAdam Langley 62867edfef7SAndi Kleen /* Compute TCP options for ESTABLISHED sockets. This is not the 62967edfef7SAndi Kleen * final wire format yet. 63067edfef7SAndi Kleen */ 63195c96174SEric Dumazet static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb, 63233ad798cSAdam Langley struct tcp_out_options *opts, 633cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 634cf533ea5SEric Dumazet { 63533ad798cSAdam Langley struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL; 63633ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 63795c96174SEric Dumazet unsigned int size = 0; 638cabeccbdSIlpo Järvinen unsigned int eff_sacks; 63933ad798cSAdam Langley 6405843ef42SAndi Kleen opts->options = 0; 6415843ef42SAndi Kleen 64233ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 64333ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 64433ad798cSAdam Langley if (unlikely(*md5)) { 64533ad798cSAdam Langley opts->options |= OPTION_MD5; 64633ad798cSAdam Langley size += TCPOLEN_MD5SIG_ALIGNED; 64733ad798cSAdam Langley } 64833ad798cSAdam Langley #else 64933ad798cSAdam Langley *md5 = NULL; 65033ad798cSAdam Langley #endif 65133ad798cSAdam Langley 65233ad798cSAdam Langley if (likely(tp->rx_opt.tstamp_ok)) { 65333ad798cSAdam Langley opts->options |= OPTION_TS; 654ee684b6fSAndrey Vagin opts->tsval = tcb ? tcb->when + tp->tsoffset : 0; 65533ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 65633ad798cSAdam Langley size += TCPOLEN_TSTAMP_ALIGNED; 65733ad798cSAdam Langley } 65833ad798cSAdam Langley 659cabeccbdSIlpo Järvinen eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; 660cabeccbdSIlpo Järvinen if (unlikely(eff_sacks)) { 66195c96174SEric Dumazet const unsigned int remaining = MAX_TCP_OPTION_SPACE - size; 66233ad798cSAdam Langley opts->num_sack_blocks = 66395c96174SEric Dumazet min_t(unsigned int, eff_sacks, 66433ad798cSAdam Langley (remaining - TCPOLEN_SACK_BASE_ALIGNED) / 66533ad798cSAdam Langley TCPOLEN_SACK_PERBLOCK); 66633ad798cSAdam Langley size += TCPOLEN_SACK_BASE_ALIGNED + 66733ad798cSAdam Langley opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; 66833ad798cSAdam Langley } 66933ad798cSAdam Langley 67033ad798cSAdam Langley return size; 67140efc6faSStephen Hemminger } 6721da177e4SLinus Torvalds 67346d3ceabSEric Dumazet 67446d3ceabSEric Dumazet /* TCP SMALL QUEUES (TSQ) 67546d3ceabSEric Dumazet * 67646d3ceabSEric Dumazet * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev) 67746d3ceabSEric Dumazet * to reduce RTT and bufferbloat. 67846d3ceabSEric Dumazet * We do this using a special skb destructor (tcp_wfree). 67946d3ceabSEric Dumazet * 68046d3ceabSEric Dumazet * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb 68146d3ceabSEric Dumazet * needs to be reallocated in a driver. 68246d3ceabSEric Dumazet * The invariant being skb->truesize substracted from sk->sk_wmem_alloc 68346d3ceabSEric Dumazet * 68446d3ceabSEric Dumazet * Since transmit from skb destructor is forbidden, we use a tasklet 68546d3ceabSEric Dumazet * to process all sockets that eventually need to send more skbs. 68646d3ceabSEric Dumazet * We use one tasklet per cpu, with its own queue of sockets. 68746d3ceabSEric Dumazet */ 68846d3ceabSEric Dumazet struct tsq_tasklet { 68946d3ceabSEric Dumazet struct tasklet_struct tasklet; 69046d3ceabSEric Dumazet struct list_head head; /* queue of tcp sockets */ 69146d3ceabSEric Dumazet }; 69246d3ceabSEric Dumazet static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet); 69346d3ceabSEric Dumazet 6946f458dfbSEric Dumazet static void tcp_tsq_handler(struct sock *sk) 6956f458dfbSEric Dumazet { 6966f458dfbSEric Dumazet if ((1 << sk->sk_state) & 6976f458dfbSEric Dumazet (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING | 6986f458dfbSEric Dumazet TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) 6996f458dfbSEric Dumazet tcp_write_xmit(sk, tcp_current_mss(sk), 0, 0, GFP_ATOMIC); 7006f458dfbSEric Dumazet } 70146d3ceabSEric Dumazet /* 70246d3ceabSEric Dumazet * One tasklest per cpu tries to send more skbs. 70346d3ceabSEric Dumazet * We run in tasklet context but need to disable irqs when 70446d3ceabSEric Dumazet * transfering tsq->head because tcp_wfree() might 70546d3ceabSEric Dumazet * interrupt us (non NAPI drivers) 70646d3ceabSEric Dumazet */ 70746d3ceabSEric Dumazet static void tcp_tasklet_func(unsigned long data) 70846d3ceabSEric Dumazet { 70946d3ceabSEric Dumazet struct tsq_tasklet *tsq = (struct tsq_tasklet *)data; 71046d3ceabSEric Dumazet LIST_HEAD(list); 71146d3ceabSEric Dumazet unsigned long flags; 71246d3ceabSEric Dumazet struct list_head *q, *n; 71346d3ceabSEric Dumazet struct tcp_sock *tp; 71446d3ceabSEric Dumazet struct sock *sk; 71546d3ceabSEric Dumazet 71646d3ceabSEric Dumazet local_irq_save(flags); 71746d3ceabSEric Dumazet list_splice_init(&tsq->head, &list); 71846d3ceabSEric Dumazet local_irq_restore(flags); 71946d3ceabSEric Dumazet 72046d3ceabSEric Dumazet list_for_each_safe(q, n, &list) { 72146d3ceabSEric Dumazet tp = list_entry(q, struct tcp_sock, tsq_node); 72246d3ceabSEric Dumazet list_del(&tp->tsq_node); 72346d3ceabSEric Dumazet 72446d3ceabSEric Dumazet sk = (struct sock *)tp; 72546d3ceabSEric Dumazet bh_lock_sock(sk); 72646d3ceabSEric Dumazet 72746d3ceabSEric Dumazet if (!sock_owned_by_user(sk)) { 7286f458dfbSEric Dumazet tcp_tsq_handler(sk); 72946d3ceabSEric Dumazet } else { 73046d3ceabSEric Dumazet /* defer the work to tcp_release_cb() */ 7316f458dfbSEric Dumazet set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags); 73246d3ceabSEric Dumazet } 73346d3ceabSEric Dumazet bh_unlock_sock(sk); 73446d3ceabSEric Dumazet 73546d3ceabSEric Dumazet clear_bit(TSQ_QUEUED, &tp->tsq_flags); 73646d3ceabSEric Dumazet sk_free(sk); 73746d3ceabSEric Dumazet } 73846d3ceabSEric Dumazet } 73946d3ceabSEric Dumazet 7406f458dfbSEric Dumazet #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \ 7416f458dfbSEric Dumazet (1UL << TCP_WRITE_TIMER_DEFERRED) | \ 742563d34d0SEric Dumazet (1UL << TCP_DELACK_TIMER_DEFERRED) | \ 743563d34d0SEric Dumazet (1UL << TCP_MTU_REDUCED_DEFERRED)) 74446d3ceabSEric Dumazet /** 74546d3ceabSEric Dumazet * tcp_release_cb - tcp release_sock() callback 74646d3ceabSEric Dumazet * @sk: socket 74746d3ceabSEric Dumazet * 74846d3ceabSEric Dumazet * called from release_sock() to perform protocol dependent 74946d3ceabSEric Dumazet * actions before socket release. 75046d3ceabSEric Dumazet */ 75146d3ceabSEric Dumazet void tcp_release_cb(struct sock *sk) 75246d3ceabSEric Dumazet { 75346d3ceabSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 7546f458dfbSEric Dumazet unsigned long flags, nflags; 75546d3ceabSEric Dumazet 7566f458dfbSEric Dumazet /* perform an atomic operation only if at least one flag is set */ 7576f458dfbSEric Dumazet do { 7586f458dfbSEric Dumazet flags = tp->tsq_flags; 7596f458dfbSEric Dumazet if (!(flags & TCP_DEFERRED_ALL)) 7606f458dfbSEric Dumazet return; 7616f458dfbSEric Dumazet nflags = flags & ~TCP_DEFERRED_ALL; 7626f458dfbSEric Dumazet } while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags); 7636f458dfbSEric Dumazet 7646f458dfbSEric Dumazet if (flags & (1UL << TCP_TSQ_DEFERRED)) 7656f458dfbSEric Dumazet tcp_tsq_handler(sk); 7666f458dfbSEric Dumazet 767144d56e9SEric Dumazet if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) { 7686f458dfbSEric Dumazet tcp_write_timer_handler(sk); 769144d56e9SEric Dumazet __sock_put(sk); 770144d56e9SEric Dumazet } 771144d56e9SEric Dumazet if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) { 7726f458dfbSEric Dumazet tcp_delack_timer_handler(sk); 773144d56e9SEric Dumazet __sock_put(sk); 774144d56e9SEric Dumazet } 775144d56e9SEric Dumazet if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) { 776563d34d0SEric Dumazet sk->sk_prot->mtu_reduced(sk); 777144d56e9SEric Dumazet __sock_put(sk); 778144d56e9SEric Dumazet } 77946d3ceabSEric Dumazet } 78046d3ceabSEric Dumazet EXPORT_SYMBOL(tcp_release_cb); 78146d3ceabSEric Dumazet 78246d3ceabSEric Dumazet void __init tcp_tasklet_init(void) 78346d3ceabSEric Dumazet { 78446d3ceabSEric Dumazet int i; 78546d3ceabSEric Dumazet 78646d3ceabSEric Dumazet for_each_possible_cpu(i) { 78746d3ceabSEric Dumazet struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i); 78846d3ceabSEric Dumazet 78946d3ceabSEric Dumazet INIT_LIST_HEAD(&tsq->head); 79046d3ceabSEric Dumazet tasklet_init(&tsq->tasklet, 79146d3ceabSEric Dumazet tcp_tasklet_func, 79246d3ceabSEric Dumazet (unsigned long)tsq); 79346d3ceabSEric Dumazet } 79446d3ceabSEric Dumazet } 79546d3ceabSEric Dumazet 79646d3ceabSEric Dumazet /* 79746d3ceabSEric Dumazet * Write buffer destructor automatically called from kfree_skb. 79846d3ceabSEric Dumazet * We cant xmit new skbs from this context, as we might already 79946d3ceabSEric Dumazet * hold qdisc lock. 80046d3ceabSEric Dumazet */ 801d6a4a104SEric Dumazet void tcp_wfree(struct sk_buff *skb) 80246d3ceabSEric Dumazet { 80346d3ceabSEric Dumazet struct sock *sk = skb->sk; 80446d3ceabSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 80546d3ceabSEric Dumazet 80646d3ceabSEric Dumazet if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) && 80746d3ceabSEric Dumazet !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) { 80846d3ceabSEric Dumazet unsigned long flags; 80946d3ceabSEric Dumazet struct tsq_tasklet *tsq; 81046d3ceabSEric Dumazet 81146d3ceabSEric Dumazet /* Keep a ref on socket. 81246d3ceabSEric Dumazet * This last ref will be released in tcp_tasklet_func() 81346d3ceabSEric Dumazet */ 81446d3ceabSEric Dumazet atomic_sub(skb->truesize - 1, &sk->sk_wmem_alloc); 81546d3ceabSEric Dumazet 81646d3ceabSEric Dumazet /* queue this socket to tasklet queue */ 81746d3ceabSEric Dumazet local_irq_save(flags); 81846d3ceabSEric Dumazet tsq = &__get_cpu_var(tsq_tasklet); 81946d3ceabSEric Dumazet list_add(&tp->tsq_node, &tsq->head); 82046d3ceabSEric Dumazet tasklet_schedule(&tsq->tasklet); 82146d3ceabSEric Dumazet local_irq_restore(flags); 82246d3ceabSEric Dumazet } else { 82346d3ceabSEric Dumazet sock_wfree(skb); 82446d3ceabSEric Dumazet } 82546d3ceabSEric Dumazet } 82646d3ceabSEric Dumazet 8271da177e4SLinus Torvalds /* This routine actually transmits TCP packets queued in by 8281da177e4SLinus Torvalds * tcp_do_sendmsg(). This is used by both the initial 8291da177e4SLinus Torvalds * transmission and possible later retransmissions. 8301da177e4SLinus Torvalds * All SKB's seen here are completely headerless. It is our 8311da177e4SLinus Torvalds * job to build the TCP header, and pass the packet down to 8321da177e4SLinus Torvalds * IP so it can do the same plus pass the packet off to the 8331da177e4SLinus Torvalds * device. 8341da177e4SLinus Torvalds * 8351da177e4SLinus Torvalds * We are working here with either a clone of the original 8361da177e4SLinus Torvalds * SKB, or a fresh unique copy made by the retransmit engine. 8371da177e4SLinus Torvalds */ 838056834d9SIlpo Järvinen static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, 839056834d9SIlpo Järvinen gfp_t gfp_mask) 8401da177e4SLinus Torvalds { 8416687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 842dfb4b9dcSDavid S. Miller struct inet_sock *inet; 843dfb4b9dcSDavid S. Miller struct tcp_sock *tp; 844dfb4b9dcSDavid S. Miller struct tcp_skb_cb *tcb; 84533ad798cSAdam Langley struct tcp_out_options opts; 84695c96174SEric Dumazet unsigned int tcp_options_size, tcp_header_size; 847cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 8481da177e4SLinus Torvalds struct tcphdr *th; 8491da177e4SLinus Torvalds int err; 8501da177e4SLinus Torvalds 851dfb4b9dcSDavid S. Miller BUG_ON(!skb || !tcp_skb_pcount(skb)); 852dfb4b9dcSDavid S. Miller 853dfb4b9dcSDavid S. Miller /* If congestion control is doing timestamping, we must 854dfb4b9dcSDavid S. Miller * take such a timestamp before we potentially clone/copy. 855dfb4b9dcSDavid S. Miller */ 856164891aaSStephen Hemminger if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP) 857dfb4b9dcSDavid S. Miller __net_timestamp(skb); 858dfb4b9dcSDavid S. Miller 859dfb4b9dcSDavid S. Miller if (likely(clone_it)) { 8600e280af0SEric Dumazet const struct sk_buff *fclone = skb + 1; 8610e280af0SEric Dumazet 8620e280af0SEric Dumazet if (unlikely(skb->fclone == SKB_FCLONE_ORIG && 8630e280af0SEric Dumazet fclone->fclone == SKB_FCLONE_CLONE)) 8640e280af0SEric Dumazet NET_INC_STATS_BH(sock_net(sk), 8650e280af0SEric Dumazet LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES); 8660e280af0SEric Dumazet 867dfb4b9dcSDavid S. Miller if (unlikely(skb_cloned(skb))) 868dfb4b9dcSDavid S. Miller skb = pskb_copy(skb, gfp_mask); 869dfb4b9dcSDavid S. Miller else 870dfb4b9dcSDavid S. Miller skb = skb_clone(skb, gfp_mask); 871dfb4b9dcSDavid S. Miller if (unlikely(!skb)) 872dfb4b9dcSDavid S. Miller return -ENOBUFS; 873dfb4b9dcSDavid S. Miller } 874dfb4b9dcSDavid S. Miller 875dfb4b9dcSDavid S. Miller inet = inet_sk(sk); 876dfb4b9dcSDavid S. Miller tp = tcp_sk(sk); 877dfb4b9dcSDavid S. Miller tcb = TCP_SKB_CB(skb); 87833ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 8791da177e4SLinus Torvalds 8804de075e0SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) 88133ad798cSAdam Langley tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); 88233ad798cSAdam Langley else 88333ad798cSAdam Langley tcp_options_size = tcp_established_options(sk, skb, &opts, 88433ad798cSAdam Langley &md5); 88533ad798cSAdam Langley tcp_header_size = tcp_options_size + sizeof(struct tcphdr); 8861da177e4SLinus Torvalds 887547669d4SEric Dumazet if (tcp_packets_in_flight(tp) == 0) 8886687e988SArnaldo Carvalho de Melo tcp_ca_event(sk, CA_EVENT_TX_START); 889547669d4SEric Dumazet 890547669d4SEric Dumazet /* if no packet is in qdisc/device queue, then allow XPS to select 891547669d4SEric Dumazet * another queue. 892547669d4SEric Dumazet */ 893547669d4SEric Dumazet skb->ooo_okay = sk_wmem_alloc_get(sk) == 0; 8941da177e4SLinus Torvalds 895aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 896aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 89746d3ceabSEric Dumazet 89846d3ceabSEric Dumazet skb_orphan(skb); 89946d3ceabSEric Dumazet skb->sk = sk; 900c9eeec26SEric Dumazet skb->destructor = tcp_wfree; 90146d3ceabSEric Dumazet atomic_add(skb->truesize, &sk->sk_wmem_alloc); 9021da177e4SLinus Torvalds 9031da177e4SLinus Torvalds /* Build TCP header and checksum it. */ 904aa8223c7SArnaldo Carvalho de Melo th = tcp_hdr(skb); 905c720c7e8SEric Dumazet th->source = inet->inet_sport; 906c720c7e8SEric Dumazet th->dest = inet->inet_dport; 9071da177e4SLinus Torvalds th->seq = htonl(tcb->seq); 9081da177e4SLinus Torvalds th->ack_seq = htonl(tp->rcv_nxt); 909df7a3b07SAl Viro *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | 9104de075e0SEric Dumazet tcb->tcp_flags); 911dfb4b9dcSDavid S. Miller 9124de075e0SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) { 9131da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments 9141da177e4SLinus Torvalds * is never scaled. 9151da177e4SLinus Torvalds */ 916600ff0c2SIlpo Järvinen th->window = htons(min(tp->rcv_wnd, 65535U)); 9171da177e4SLinus Torvalds } else { 9181da177e4SLinus Torvalds th->window = htons(tcp_select_window(sk)); 9191da177e4SLinus Torvalds } 9201da177e4SLinus Torvalds th->check = 0; 9211da177e4SLinus Torvalds th->urg_ptr = 0; 9221da177e4SLinus Torvalds 92333f5f57eSIlpo Järvinen /* The urg_mode check is necessary during a below snd_una win probe */ 9247691367dSHerbert Xu if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { 9257691367dSHerbert Xu if (before(tp->snd_up, tcb->seq + 0x10000)) { 9261da177e4SLinus Torvalds th->urg_ptr = htons(tp->snd_up - tcb->seq); 9271da177e4SLinus Torvalds th->urg = 1; 9287691367dSHerbert Xu } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { 9290eae88f3SEric Dumazet th->urg_ptr = htons(0xFFFF); 9307691367dSHerbert Xu th->urg = 1; 9317691367dSHerbert Xu } 9321da177e4SLinus Torvalds } 9331da177e4SLinus Torvalds 934bd0388aeSWilliam Allen Simpson tcp_options_write((__be32 *)(th + 1), tp, &opts); 9354de075e0SEric Dumazet if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0)) 9369e412ba7SIlpo Järvinen TCP_ECN_send(sk, skb, tcp_header_size); 937dfb4b9dcSDavid S. Miller 938cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 939cfb6eeb4SYOSHIFUJI Hideaki /* Calculate the MD5 hash, as we have all we need now */ 940cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 941a465419bSEric Dumazet sk_nocaps_add(sk, NETIF_F_GSO_MASK); 942bd0388aeSWilliam Allen Simpson tp->af_specific->calc_md5_hash(opts.hash_location, 94349a72dfbSAdam Langley md5, sk, NULL, skb); 944cfb6eeb4SYOSHIFUJI Hideaki } 945cfb6eeb4SYOSHIFUJI Hideaki #endif 946cfb6eeb4SYOSHIFUJI Hideaki 947bb296246SHerbert Xu icsk->icsk_af_ops->send_check(sk, skb); 9481da177e4SLinus Torvalds 9494de075e0SEric Dumazet if (likely(tcb->tcp_flags & TCPHDR_ACK)) 950fc6415bcSDavid S. Miller tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); 9511da177e4SLinus Torvalds 9521da177e4SLinus Torvalds if (skb->len != tcp_header_size) 953cf533ea5SEric Dumazet tcp_event_data_sent(tp, sk); 9541da177e4SLinus Torvalds 955bd37a088SWei Yongjun if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) 956aa2ea058STom Herbert TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, 957aa2ea058STom Herbert tcp_skb_pcount(skb)); 9581da177e4SLinus Torvalds 959d9d8da80SDavid S. Miller err = icsk->icsk_af_ops->queue_xmit(skb, &inet->cork.fl); 96083de47cdSHua Zhong if (likely(err <= 0)) 9611da177e4SLinus Torvalds return err; 9621da177e4SLinus Torvalds 9633cfe3baaSIlpo Järvinen tcp_enter_cwr(sk, 1); 9641da177e4SLinus Torvalds 965b9df3cb8SGerrit Renker return net_xmit_eval(err); 9661da177e4SLinus Torvalds } 9671da177e4SLinus Torvalds 96867edfef7SAndi Kleen /* This routine just queues the buffer for sending. 9691da177e4SLinus Torvalds * 9701da177e4SLinus Torvalds * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, 9711da177e4SLinus Torvalds * otherwise socket can stall. 9721da177e4SLinus Torvalds */ 9731da177e4SLinus Torvalds static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) 9741da177e4SLinus Torvalds { 9751da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 9761da177e4SLinus Torvalds 9771da177e4SLinus Torvalds /* Advance write_seq and place onto the write_queue. */ 9781da177e4SLinus Torvalds tp->write_seq = TCP_SKB_CB(skb)->end_seq; 9791da177e4SLinus Torvalds skb_header_release(skb); 980fe067e8aSDavid S. Miller tcp_add_write_queue_tail(sk, skb); 9813ab224beSHideo Aoki sk->sk_wmem_queued += skb->truesize; 9823ab224beSHideo Aoki sk_mem_charge(sk, skb->truesize); 9831da177e4SLinus Torvalds } 9841da177e4SLinus Torvalds 98567edfef7SAndi Kleen /* Initialize TSO segments for a packet. */ 986cf533ea5SEric Dumazet static void tcp_set_skb_tso_segs(const struct sock *sk, struct sk_buff *skb, 987056834d9SIlpo Järvinen unsigned int mss_now) 988f6302d1dSDavid S. Miller { 989*c52e2421SEric Dumazet /* Make sure we own this skb before messing gso_size/gso_segs */ 990*c52e2421SEric Dumazet WARN_ON_ONCE(skb_cloned(skb)); 991*c52e2421SEric Dumazet 9928e5b9ddaSHerbert Xu if (skb->len <= mss_now || !sk_can_gso(sk) || 9938e5b9ddaSHerbert Xu skb->ip_summed == CHECKSUM_NONE) { 994f6302d1dSDavid S. Miller /* Avoid the costly divide in the normal 995f6302d1dSDavid S. Miller * non-TSO case. 996f6302d1dSDavid S. Miller */ 9977967168cSHerbert Xu skb_shinfo(skb)->gso_segs = 1; 9987967168cSHerbert Xu skb_shinfo(skb)->gso_size = 0; 999c9af6db4SPravin B Shelar skb_shinfo(skb)->gso_type = 0; 1000f6302d1dSDavid S. Miller } else { 1001356f89e1SIlpo Järvinen skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now); 10027967168cSHerbert Xu skb_shinfo(skb)->gso_size = mss_now; 1003c9af6db4SPravin B Shelar skb_shinfo(skb)->gso_type = sk->sk_gso_type; 10041da177e4SLinus Torvalds } 10051da177e4SLinus Torvalds } 10061da177e4SLinus Torvalds 100791fed7a1SIlpo Järvinen /* When a modification to fackets out becomes necessary, we need to check 100868f8353bSIlpo Järvinen * skb is counted to fackets_out or not. 100991fed7a1SIlpo Järvinen */ 1010cf533ea5SEric Dumazet static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb, 101191fed7a1SIlpo Järvinen int decr) 101291fed7a1SIlpo Järvinen { 1013a47e5a98SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1014a47e5a98SIlpo Järvinen 1015dc86967bSIlpo Järvinen if (!tp->sacked_out || tcp_is_reno(tp)) 101691fed7a1SIlpo Järvinen return; 101791fed7a1SIlpo Järvinen 10186859d494SIlpo Järvinen if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq)) 101991fed7a1SIlpo Järvinen tp->fackets_out -= decr; 102091fed7a1SIlpo Järvinen } 102191fed7a1SIlpo Järvinen 1022797108d1SIlpo Järvinen /* Pcount in the middle of the write queue got changed, we need to do various 1023797108d1SIlpo Järvinen * tweaks to fix counters 1024797108d1SIlpo Järvinen */ 1025cf533ea5SEric Dumazet static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) 1026797108d1SIlpo Järvinen { 1027797108d1SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1028797108d1SIlpo Järvinen 1029797108d1SIlpo Järvinen tp->packets_out -= decr; 1030797108d1SIlpo Järvinen 1031797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 1032797108d1SIlpo Järvinen tp->sacked_out -= decr; 1033797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 1034797108d1SIlpo Järvinen tp->retrans_out -= decr; 1035797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 1036797108d1SIlpo Järvinen tp->lost_out -= decr; 1037797108d1SIlpo Järvinen 1038797108d1SIlpo Järvinen /* Reno case is special. Sigh... */ 1039797108d1SIlpo Järvinen if (tcp_is_reno(tp) && decr > 0) 1040797108d1SIlpo Järvinen tp->sacked_out -= min_t(u32, tp->sacked_out, decr); 1041797108d1SIlpo Järvinen 1042797108d1SIlpo Järvinen tcp_adjust_fackets_out(sk, skb, decr); 1043797108d1SIlpo Järvinen 1044797108d1SIlpo Järvinen if (tp->lost_skb_hint && 1045797108d1SIlpo Järvinen before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && 104652cf3cc8SIlpo Järvinen (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))) 1047797108d1SIlpo Järvinen tp->lost_cnt_hint -= decr; 1048797108d1SIlpo Järvinen 1049797108d1SIlpo Järvinen tcp_verify_left_out(tp); 1050797108d1SIlpo Järvinen } 1051797108d1SIlpo Järvinen 10521da177e4SLinus Torvalds /* Function to create two new TCP segments. Shrinks the given segment 10531da177e4SLinus Torvalds * to the specified size and appends a new segment with the rest of the 10541da177e4SLinus Torvalds * packet to the list. This won't be called frequently, I hope. 10551da177e4SLinus Torvalds * Remember, these are still headerless SKBs at this point. 10561da177e4SLinus Torvalds */ 1057056834d9SIlpo Järvinen int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, 1058056834d9SIlpo Järvinen unsigned int mss_now) 10591da177e4SLinus Torvalds { 10601da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 10611da177e4SLinus Torvalds struct sk_buff *buff; 10626475be16SDavid S. Miller int nsize, old_factor; 1063b60b49eaSHerbert Xu int nlen; 10649ce01461SIlpo Järvinen u8 flags; 10651da177e4SLinus Torvalds 10662fceec13SIlpo Järvinen if (WARN_ON(len > skb->len)) 10672fceec13SIlpo Järvinen return -EINVAL; 10686a438bbeSStephen Hemminger 10691da177e4SLinus Torvalds nsize = skb_headlen(skb) - len; 10701da177e4SLinus Torvalds if (nsize < 0) 10711da177e4SLinus Torvalds nsize = 0; 10721da177e4SLinus Torvalds 1073*c52e2421SEric Dumazet if (skb_unclone(skb, GFP_ATOMIC)) 10741da177e4SLinus Torvalds return -ENOMEM; 10751da177e4SLinus Torvalds 10761da177e4SLinus Torvalds /* Get a new skb... force flag on. */ 10771da177e4SLinus Torvalds buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC); 10781da177e4SLinus Torvalds if (buff == NULL) 10791da177e4SLinus Torvalds return -ENOMEM; /* We'll just try again later. */ 1080ef5cb973SHerbert Xu 10813ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 10823ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1083b60b49eaSHerbert Xu nlen = skb->len - len - nsize; 1084b60b49eaSHerbert Xu buff->truesize += nlen; 1085b60b49eaSHerbert Xu skb->truesize -= nlen; 10861da177e4SLinus Torvalds 10871da177e4SLinus Torvalds /* Correct the sequence numbers. */ 10881da177e4SLinus Torvalds TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 10891da177e4SLinus Torvalds TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 10901da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 10911da177e4SLinus Torvalds 10921da177e4SLinus Torvalds /* PSH and FIN should only be set in the second packet. */ 10934de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 10944de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 10954de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1096e14c3cafSHerbert Xu TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; 10971da177e4SLinus Torvalds 109884fa7933SPatrick McHardy if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { 10991da177e4SLinus Torvalds /* Copy and checksum data tail into the new buffer. */ 1100056834d9SIlpo Järvinen buff->csum = csum_partial_copy_nocheck(skb->data + len, 1101056834d9SIlpo Järvinen skb_put(buff, nsize), 11021da177e4SLinus Torvalds nsize, 0); 11031da177e4SLinus Torvalds 11041da177e4SLinus Torvalds skb_trim(skb, len); 11051da177e4SLinus Torvalds 11061da177e4SLinus Torvalds skb->csum = csum_block_sub(skb->csum, buff->csum, len); 11071da177e4SLinus Torvalds } else { 110884fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 11091da177e4SLinus Torvalds skb_split(skb, buff, len); 11101da177e4SLinus Torvalds } 11111da177e4SLinus Torvalds 11121da177e4SLinus Torvalds buff->ip_summed = skb->ip_summed; 11131da177e4SLinus Torvalds 11141da177e4SLinus Torvalds /* Looks stupid, but our code really uses when of 11151da177e4SLinus Torvalds * skbs, which it never sent before. --ANK 11161da177e4SLinus Torvalds */ 11171da177e4SLinus Torvalds TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when; 1118a61bbcf2SPatrick McHardy buff->tstamp = skb->tstamp; 11191da177e4SLinus Torvalds 11206475be16SDavid S. Miller old_factor = tcp_skb_pcount(skb); 11216475be16SDavid S. Miller 11221da177e4SLinus Torvalds /* Fix up tso_factor for both original and new SKB. */ 1123846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1124846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, buff, mss_now); 11251da177e4SLinus Torvalds 11266475be16SDavid S. Miller /* If this packet has been sent out already, we must 11276475be16SDavid S. Miller * adjust the various packet counters. 11286475be16SDavid S. Miller */ 1129cf0b450cSHerbert Xu if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 11306475be16SDavid S. Miller int diff = old_factor - tcp_skb_pcount(skb) - 11316475be16SDavid S. Miller tcp_skb_pcount(buff); 11321da177e4SLinus Torvalds 1133797108d1SIlpo Järvinen if (diff) 1134797108d1SIlpo Järvinen tcp_adjust_pcount(sk, skb, diff); 11351da177e4SLinus Torvalds } 11361da177e4SLinus Torvalds 11371da177e4SLinus Torvalds /* Link BUFF into the send queue. */ 1138f44b5271SDavid S. Miller skb_header_release(buff); 1139fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 11401da177e4SLinus Torvalds 11411da177e4SLinus Torvalds return 0; 11421da177e4SLinus Torvalds } 11431da177e4SLinus Torvalds 11441da177e4SLinus Torvalds /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c 11451da177e4SLinus Torvalds * eventually). The difference is that pulled data not copied, but 11461da177e4SLinus Torvalds * immediately discarded. 11471da177e4SLinus Torvalds */ 1148f2911969SHerbert Xu ~{PmVHI~} static void __pskb_trim_head(struct sk_buff *skb, int len) 11491da177e4SLinus Torvalds { 11501da177e4SLinus Torvalds int i, k, eat; 11511da177e4SLinus Torvalds 11524fa48bf3SEric Dumazet eat = min_t(int, len, skb_headlen(skb)); 11534fa48bf3SEric Dumazet if (eat) { 11544fa48bf3SEric Dumazet __skb_pull(skb, eat); 11554fa48bf3SEric Dumazet len -= eat; 11564fa48bf3SEric Dumazet if (!len) 11574fa48bf3SEric Dumazet return; 11584fa48bf3SEric Dumazet } 11591da177e4SLinus Torvalds eat = len; 11601da177e4SLinus Torvalds k = 0; 11611da177e4SLinus Torvalds for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) { 11629e903e08SEric Dumazet int size = skb_frag_size(&skb_shinfo(skb)->frags[i]); 11639e903e08SEric Dumazet 11649e903e08SEric Dumazet if (size <= eat) { 1165aff65da0SIan Campbell skb_frag_unref(skb, i); 11669e903e08SEric Dumazet eat -= size; 11671da177e4SLinus Torvalds } else { 11681da177e4SLinus Torvalds skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i]; 11691da177e4SLinus Torvalds if (eat) { 11701da177e4SLinus Torvalds skb_shinfo(skb)->frags[k].page_offset += eat; 11719e903e08SEric Dumazet skb_frag_size_sub(&skb_shinfo(skb)->frags[k], eat); 11721da177e4SLinus Torvalds eat = 0; 11731da177e4SLinus Torvalds } 11741da177e4SLinus Torvalds k++; 11751da177e4SLinus Torvalds } 11761da177e4SLinus Torvalds } 11771da177e4SLinus Torvalds skb_shinfo(skb)->nr_frags = k; 11781da177e4SLinus Torvalds 117927a884dcSArnaldo Carvalho de Melo skb_reset_tail_pointer(skb); 11801da177e4SLinus Torvalds skb->data_len -= len; 11811da177e4SLinus Torvalds skb->len = skb->data_len; 11821da177e4SLinus Torvalds } 11831da177e4SLinus Torvalds 118467edfef7SAndi Kleen /* Remove acked data from a packet in the transmit queue. */ 11851da177e4SLinus Torvalds int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 11861da177e4SLinus Torvalds { 118714bbd6a5SPravin B Shelar if (skb_unclone(skb, GFP_ATOMIC)) 11881da177e4SLinus Torvalds return -ENOMEM; 11891da177e4SLinus Torvalds 11904fa48bf3SEric Dumazet __pskb_trim_head(skb, len); 11911da177e4SLinus Torvalds 11921da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq += len; 119384fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 11941da177e4SLinus Torvalds 11951da177e4SLinus Torvalds skb->truesize -= len; 11961da177e4SLinus Torvalds sk->sk_wmem_queued -= len; 11973ab224beSHideo Aoki sk_mem_uncharge(sk, len); 11981da177e4SLinus Torvalds sock_set_flag(sk, SOCK_QUEUE_SHRUNK); 11991da177e4SLinus Torvalds 12005b35e1e6SNeal Cardwell /* Any change of skb->len requires recalculation of tso factor. */ 12011da177e4SLinus Torvalds if (tcp_skb_pcount(skb) > 1) 12025b35e1e6SNeal Cardwell tcp_set_skb_tso_segs(sk, skb, tcp_skb_mss(skb)); 12031da177e4SLinus Torvalds 12041da177e4SLinus Torvalds return 0; 12051da177e4SLinus Torvalds } 12061da177e4SLinus Torvalds 12071b63edd6SYuchung Cheng /* Calculate MSS not accounting any TCP options. */ 12081b63edd6SYuchung Cheng static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu) 12095d424d5aSJohn Heffner { 1210cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1211cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 12125d424d5aSJohn Heffner int mss_now; 12135d424d5aSJohn Heffner 12145d424d5aSJohn Heffner /* Calculate base mss without TCP options: 12155d424d5aSJohn Heffner It is MMS_S - sizeof(tcphdr) of rfc1122 12165d424d5aSJohn Heffner */ 12175d424d5aSJohn Heffner mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 12185d424d5aSJohn Heffner 121967469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 122067469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 122167469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 122267469601SEric Dumazet 122367469601SEric Dumazet if (dst && dst_allfrag(dst)) 122467469601SEric Dumazet mss_now -= icsk->icsk_af_ops->net_frag_header_len; 122567469601SEric Dumazet } 122667469601SEric Dumazet 12275d424d5aSJohn Heffner /* Clamp it (mss_clamp does not include tcp options) */ 12285d424d5aSJohn Heffner if (mss_now > tp->rx_opt.mss_clamp) 12295d424d5aSJohn Heffner mss_now = tp->rx_opt.mss_clamp; 12305d424d5aSJohn Heffner 12315d424d5aSJohn Heffner /* Now subtract optional transport overhead */ 12325d424d5aSJohn Heffner mss_now -= icsk->icsk_ext_hdr_len; 12335d424d5aSJohn Heffner 12345d424d5aSJohn Heffner /* Then reserve room for full set of TCP options and 8 bytes of data */ 12355d424d5aSJohn Heffner if (mss_now < 48) 12365d424d5aSJohn Heffner mss_now = 48; 12375d424d5aSJohn Heffner return mss_now; 12385d424d5aSJohn Heffner } 12395d424d5aSJohn Heffner 12401b63edd6SYuchung Cheng /* Calculate MSS. Not accounting for SACKs here. */ 12411b63edd6SYuchung Cheng int tcp_mtu_to_mss(struct sock *sk, int pmtu) 12421b63edd6SYuchung Cheng { 12431b63edd6SYuchung Cheng /* Subtract TCP options size, not including SACKs */ 12441b63edd6SYuchung Cheng return __tcp_mtu_to_mss(sk, pmtu) - 12451b63edd6SYuchung Cheng (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr)); 12461b63edd6SYuchung Cheng } 12471b63edd6SYuchung Cheng 12485d424d5aSJohn Heffner /* Inverse of above */ 124967469601SEric Dumazet int tcp_mss_to_mtu(struct sock *sk, int mss) 12505d424d5aSJohn Heffner { 1251cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1252cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 12535d424d5aSJohn Heffner int mtu; 12545d424d5aSJohn Heffner 12555d424d5aSJohn Heffner mtu = mss + 12565d424d5aSJohn Heffner tp->tcp_header_len + 12575d424d5aSJohn Heffner icsk->icsk_ext_hdr_len + 12585d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 12595d424d5aSJohn Heffner 126067469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 126167469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 126267469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 126367469601SEric Dumazet 126467469601SEric Dumazet if (dst && dst_allfrag(dst)) 126567469601SEric Dumazet mtu += icsk->icsk_af_ops->net_frag_header_len; 126667469601SEric Dumazet } 12675d424d5aSJohn Heffner return mtu; 12685d424d5aSJohn Heffner } 12695d424d5aSJohn Heffner 127067edfef7SAndi Kleen /* MTU probing init per socket */ 12715d424d5aSJohn Heffner void tcp_mtup_init(struct sock *sk) 12725d424d5aSJohn Heffner { 12735d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 12745d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 12755d424d5aSJohn Heffner 12765d424d5aSJohn Heffner icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1; 12775d424d5aSJohn Heffner icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 12785d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 12795d424d5aSJohn Heffner icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss); 12805d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 12815d424d5aSJohn Heffner } 12824bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_mtup_init); 12835d424d5aSJohn Heffner 12841da177e4SLinus Torvalds /* This function synchronize snd mss to current pmtu/exthdr set. 12851da177e4SLinus Torvalds 12861da177e4SLinus Torvalds tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 12871da177e4SLinus Torvalds for TCP options, but includes only bare TCP header. 12881da177e4SLinus Torvalds 12891da177e4SLinus Torvalds tp->rx_opt.mss_clamp is mss negotiated at connection setup. 1290caa20d9aSStephen Hemminger It is minimum of user_mss and mss received with SYN. 12911da177e4SLinus Torvalds It also does not include TCP options. 12921da177e4SLinus Torvalds 1293d83d8461SArnaldo Carvalho de Melo inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. 12941da177e4SLinus Torvalds 12951da177e4SLinus Torvalds tp->mss_cache is current effective sending mss, including 12961da177e4SLinus Torvalds all tcp options except for SACKs. It is evaluated, 12971da177e4SLinus Torvalds taking into account current pmtu, but never exceeds 12981da177e4SLinus Torvalds tp->rx_opt.mss_clamp. 12991da177e4SLinus Torvalds 13001da177e4SLinus Torvalds NOTE1. rfc1122 clearly states that advertised MSS 13011da177e4SLinus Torvalds DOES NOT include either tcp or ip options. 13021da177e4SLinus Torvalds 1303d83d8461SArnaldo Carvalho de Melo NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache 1304d83d8461SArnaldo Carvalho de Melo are READ ONLY outside this function. --ANK (980731) 13051da177e4SLinus Torvalds */ 13061da177e4SLinus Torvalds unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 13071da177e4SLinus Torvalds { 13081da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1309d83d8461SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 13105d424d5aSJohn Heffner int mss_now; 13111da177e4SLinus Torvalds 13125d424d5aSJohn Heffner if (icsk->icsk_mtup.search_high > pmtu) 13135d424d5aSJohn Heffner icsk->icsk_mtup.search_high = pmtu; 13141da177e4SLinus Torvalds 13155d424d5aSJohn Heffner mss_now = tcp_mtu_to_mss(sk, pmtu); 1316409d22b4SIlpo Järvinen mss_now = tcp_bound_to_half_wnd(tp, mss_now); 13171da177e4SLinus Torvalds 13181da177e4SLinus Torvalds /* And store cached results */ 1319d83d8461SArnaldo Carvalho de Melo icsk->icsk_pmtu_cookie = pmtu; 13205d424d5aSJohn Heffner if (icsk->icsk_mtup.enabled) 13215d424d5aSJohn Heffner mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); 1322c1b4a7e6SDavid S. Miller tp->mss_cache = mss_now; 13231da177e4SLinus Torvalds 13241da177e4SLinus Torvalds return mss_now; 13251da177e4SLinus Torvalds } 13264bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_sync_mss); 13271da177e4SLinus Torvalds 13281da177e4SLinus Torvalds /* Compute the current effective MSS, taking SACKs and IP options, 13291da177e4SLinus Torvalds * and even PMTU discovery events into account. 13301da177e4SLinus Torvalds */ 13310c54b85fSIlpo Järvinen unsigned int tcp_current_mss(struct sock *sk) 13321da177e4SLinus Torvalds { 1333cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1334cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1335c1b4a7e6SDavid S. Miller u32 mss_now; 133695c96174SEric Dumazet unsigned int header_len; 133733ad798cSAdam Langley struct tcp_out_options opts; 133833ad798cSAdam Langley struct tcp_md5sig_key *md5; 13391da177e4SLinus Torvalds 1340c1b4a7e6SDavid S. Miller mss_now = tp->mss_cache; 1341c1b4a7e6SDavid S. Miller 13421da177e4SLinus Torvalds if (dst) { 13431da177e4SLinus Torvalds u32 mtu = dst_mtu(dst); 1344d83d8461SArnaldo Carvalho de Melo if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 13451da177e4SLinus Torvalds mss_now = tcp_sync_mss(sk, mtu); 13461da177e4SLinus Torvalds } 13471da177e4SLinus Torvalds 134833ad798cSAdam Langley header_len = tcp_established_options(sk, NULL, &opts, &md5) + 134933ad798cSAdam Langley sizeof(struct tcphdr); 135033ad798cSAdam Langley /* The mss_cache is sized based on tp->tcp_header_len, which assumes 135133ad798cSAdam Langley * some common options. If this is an odd packet (because we have SACK 135233ad798cSAdam Langley * blocks etc) then our calculated header_len will be different, and 135333ad798cSAdam Langley * we have to adjust mss_now correspondingly */ 135433ad798cSAdam Langley if (header_len != tp->tcp_header_len) { 135533ad798cSAdam Langley int delta = (int) header_len - tp->tcp_header_len; 135633ad798cSAdam Langley mss_now -= delta; 135733ad798cSAdam Langley } 1358cfb6eeb4SYOSHIFUJI Hideaki 13591da177e4SLinus Torvalds return mss_now; 13601da177e4SLinus Torvalds } 13611da177e4SLinus Torvalds 1362a762a980SDavid S. Miller /* Congestion window validation. (RFC2861) */ 13639e412ba7SIlpo Järvinen static void tcp_cwnd_validate(struct sock *sk) 1364a762a980SDavid S. Miller { 13659e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1366a762a980SDavid S. Miller 1367d436d686SIlpo Järvinen if (tp->packets_out >= tp->snd_cwnd) { 1368a762a980SDavid S. Miller /* Network is feed fully. */ 1369a762a980SDavid S. Miller tp->snd_cwnd_used = 0; 1370a762a980SDavid S. Miller tp->snd_cwnd_stamp = tcp_time_stamp; 1371a762a980SDavid S. Miller } else { 1372a762a980SDavid S. Miller /* Network starves. */ 1373a762a980SDavid S. Miller if (tp->packets_out > tp->snd_cwnd_used) 1374a762a980SDavid S. Miller tp->snd_cwnd_used = tp->packets_out; 1375a762a980SDavid S. Miller 137615d33c07SDavid S. Miller if (sysctl_tcp_slow_start_after_idle && 137715d33c07SDavid S. Miller (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto) 1378a762a980SDavid S. Miller tcp_cwnd_application_limited(sk); 1379a762a980SDavid S. Miller } 1380a762a980SDavid S. Miller } 1381a762a980SDavid S. Miller 13820e3a4803SIlpo Järvinen /* Returns the portion of skb which can be sent right away without 13830e3a4803SIlpo Järvinen * introducing MSS oddities to segment boundaries. In rare cases where 13840e3a4803SIlpo Järvinen * mss_now != mss_cache, we will request caller to create a small skb 13850e3a4803SIlpo Järvinen * per input skb which could be mostly avoided here (if desired). 13865ea3a748SIlpo Järvinen * 13875ea3a748SIlpo Järvinen * We explicitly want to create a request for splitting write queue tail 13885ea3a748SIlpo Järvinen * to a small skb for Nagle purposes while avoiding unnecessary modulos, 13895ea3a748SIlpo Järvinen * thus all the complexity (cwnd_len is always MSS multiple which we 13905ea3a748SIlpo Järvinen * return whenever allowed by the other factors). Basically we need the 13915ea3a748SIlpo Järvinen * modulo only when the receiver window alone is the limiting factor or 13925ea3a748SIlpo Järvinen * when we would be allowed to send the split-due-to-Nagle skb fully. 13930e3a4803SIlpo Järvinen */ 1394cf533ea5SEric Dumazet static unsigned int tcp_mss_split_point(const struct sock *sk, const struct sk_buff *skb, 13951485348dSBen Hutchings unsigned int mss_now, unsigned int max_segs) 1396c1b4a7e6SDavid S. Miller { 1397cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 13981485348dSBen Hutchings u32 needed, window, max_len; 1399c1b4a7e6SDavid S. Miller 140090840defSIlpo Järvinen window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 14011485348dSBen Hutchings max_len = mss_now * max_segs; 14020e3a4803SIlpo Järvinen 14031485348dSBen Hutchings if (likely(max_len <= window && skb != tcp_write_queue_tail(sk))) 14041485348dSBen Hutchings return max_len; 14050e3a4803SIlpo Järvinen 14065ea3a748SIlpo Järvinen needed = min(skb->len, window); 14075ea3a748SIlpo Järvinen 14081485348dSBen Hutchings if (max_len <= needed) 14091485348dSBen Hutchings return max_len; 14100e3a4803SIlpo Järvinen 14110e3a4803SIlpo Järvinen return needed - needed % mss_now; 1412c1b4a7e6SDavid S. Miller } 1413c1b4a7e6SDavid S. Miller 1414c1b4a7e6SDavid S. Miller /* Can at least one segment of SKB be sent right now, according to the 1415c1b4a7e6SDavid S. Miller * congestion window rules? If so, return how many segments are allowed. 1416c1b4a7e6SDavid S. Miller */ 1417cf533ea5SEric Dumazet static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, 1418cf533ea5SEric Dumazet const struct sk_buff *skb) 1419c1b4a7e6SDavid S. Miller { 1420c1b4a7e6SDavid S. Miller u32 in_flight, cwnd; 1421c1b4a7e6SDavid S. Miller 1422c1b4a7e6SDavid S. Miller /* Don't be strict about the congestion window for the final FIN. */ 14234de075e0SEric Dumazet if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && 14244de075e0SEric Dumazet tcp_skb_pcount(skb) == 1) 1425c1b4a7e6SDavid S. Miller return 1; 1426c1b4a7e6SDavid S. Miller 1427c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1428c1b4a7e6SDavid S. Miller cwnd = tp->snd_cwnd; 1429c1b4a7e6SDavid S. Miller if (in_flight < cwnd) 1430c1b4a7e6SDavid S. Miller return (cwnd - in_flight); 1431c1b4a7e6SDavid S. Miller 1432c1b4a7e6SDavid S. Miller return 0; 1433c1b4a7e6SDavid S. Miller } 1434c1b4a7e6SDavid S. Miller 1435b595076aSUwe Kleine-König /* Initialize TSO state of a skb. 143667edfef7SAndi Kleen * This must be invoked the first time we consider transmitting 1437c1b4a7e6SDavid S. Miller * SKB onto the wire. 1438c1b4a7e6SDavid S. Miller */ 1439cf533ea5SEric Dumazet static int tcp_init_tso_segs(const struct sock *sk, struct sk_buff *skb, 1440056834d9SIlpo Järvinen unsigned int mss_now) 1441c1b4a7e6SDavid S. Miller { 1442c1b4a7e6SDavid S. Miller int tso_segs = tcp_skb_pcount(skb); 1443c1b4a7e6SDavid S. Miller 1444f8269a49SIlpo Järvinen if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { 1445846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1446c1b4a7e6SDavid S. Miller tso_segs = tcp_skb_pcount(skb); 1447c1b4a7e6SDavid S. Miller } 1448c1b4a7e6SDavid S. Miller return tso_segs; 1449c1b4a7e6SDavid S. Miller } 1450c1b4a7e6SDavid S. Miller 145167edfef7SAndi Kleen /* Minshall's variant of the Nagle send check. */ 1452a2a385d6SEric Dumazet static inline bool tcp_minshall_check(const struct tcp_sock *tp) 1453c1b4a7e6SDavid S. Miller { 1454c1b4a7e6SDavid S. Miller return after(tp->snd_sml, tp->snd_una) && 1455c1b4a7e6SDavid S. Miller !after(tp->snd_sml, tp->snd_nxt); 1456c1b4a7e6SDavid S. Miller } 1457c1b4a7e6SDavid S. Miller 1458a2a385d6SEric Dumazet /* Return false, if packet can be sent now without violation Nagle's rules: 1459c1b4a7e6SDavid S. Miller * 1. It is full sized. 1460c1b4a7e6SDavid S. Miller * 2. Or it contains FIN. (already checked by caller) 14616d67e9beSFeng King * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. 1462c1b4a7e6SDavid S. Miller * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 1463c1b4a7e6SDavid S. Miller * With Minshall's modification: all sent small packets are ACKed. 1464c1b4a7e6SDavid S. Miller */ 1465a2a385d6SEric Dumazet static inline bool tcp_nagle_check(const struct tcp_sock *tp, 1466c1b4a7e6SDavid S. Miller const struct sk_buff *skb, 146795c96174SEric Dumazet unsigned int mss_now, int nonagle) 1468c1b4a7e6SDavid S. Miller { 1469a02cec21SEric Dumazet return skb->len < mss_now && 1470c1b4a7e6SDavid S. Miller ((nonagle & TCP_NAGLE_CORK) || 1471a02cec21SEric Dumazet (!nonagle && tp->packets_out && tcp_minshall_check(tp))); 1472c1b4a7e6SDavid S. Miller } 1473c1b4a7e6SDavid S. Miller 1474a2a385d6SEric Dumazet /* Return true if the Nagle test allows this packet to be 1475c1b4a7e6SDavid S. Miller * sent now. 1476c1b4a7e6SDavid S. Miller */ 1477a2a385d6SEric Dumazet static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, 1478c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1479c1b4a7e6SDavid S. Miller { 1480c1b4a7e6SDavid S. Miller /* Nagle rule does not apply to frames, which sit in the middle of the 1481c1b4a7e6SDavid S. Miller * write_queue (they have no chances to get new data). 1482c1b4a7e6SDavid S. Miller * 1483c1b4a7e6SDavid S. Miller * This is implemented in the callers, where they modify the 'nonagle' 1484c1b4a7e6SDavid S. Miller * argument based upon the location of SKB in the send queue. 1485c1b4a7e6SDavid S. Miller */ 1486c1b4a7e6SDavid S. Miller if (nonagle & TCP_NAGLE_PUSH) 1487a2a385d6SEric Dumazet return true; 1488c1b4a7e6SDavid S. Miller 14899b44190dSYuchung Cheng /* Don't use the nagle rule for urgent data (or for the final FIN). */ 14909b44190dSYuchung Cheng if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) 1491a2a385d6SEric Dumazet return true; 1492c1b4a7e6SDavid S. Miller 1493c1b4a7e6SDavid S. Miller if (!tcp_nagle_check(tp, skb, cur_mss, nonagle)) 1494a2a385d6SEric Dumazet return true; 1495c1b4a7e6SDavid S. Miller 1496a2a385d6SEric Dumazet return false; 1497c1b4a7e6SDavid S. Miller } 1498c1b4a7e6SDavid S. Miller 1499c1b4a7e6SDavid S. Miller /* Does at least the first segment of SKB fit into the send window? */ 1500a2a385d6SEric Dumazet static bool tcp_snd_wnd_test(const struct tcp_sock *tp, 1501a2a385d6SEric Dumazet const struct sk_buff *skb, 1502056834d9SIlpo Järvinen unsigned int cur_mss) 1503c1b4a7e6SDavid S. Miller { 1504c1b4a7e6SDavid S. Miller u32 end_seq = TCP_SKB_CB(skb)->end_seq; 1505c1b4a7e6SDavid S. Miller 1506c1b4a7e6SDavid S. Miller if (skb->len > cur_mss) 1507c1b4a7e6SDavid S. Miller end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 1508c1b4a7e6SDavid S. Miller 150990840defSIlpo Järvinen return !after(end_seq, tcp_wnd_end(tp)); 1510c1b4a7e6SDavid S. Miller } 1511c1b4a7e6SDavid S. Miller 1512fe067e8aSDavid S. Miller /* This checks if the data bearing packet SKB (usually tcp_send_head(sk)) 1513c1b4a7e6SDavid S. Miller * should be put on the wire right now. If so, it returns the number of 1514c1b4a7e6SDavid S. Miller * packets allowed by the congestion window. 1515c1b4a7e6SDavid S. Miller */ 1516cf533ea5SEric Dumazet static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb, 1517c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1518c1b4a7e6SDavid S. Miller { 1519cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1520c1b4a7e6SDavid S. Miller unsigned int cwnd_quota; 1521c1b4a7e6SDavid S. Miller 1522846998aeSDavid S. Miller tcp_init_tso_segs(sk, skb, cur_mss); 1523c1b4a7e6SDavid S. Miller 1524c1b4a7e6SDavid S. Miller if (!tcp_nagle_test(tp, skb, cur_mss, nonagle)) 1525c1b4a7e6SDavid S. Miller return 0; 1526c1b4a7e6SDavid S. Miller 1527c1b4a7e6SDavid S. Miller cwnd_quota = tcp_cwnd_test(tp, skb); 1528056834d9SIlpo Järvinen if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss)) 1529c1b4a7e6SDavid S. Miller cwnd_quota = 0; 1530c1b4a7e6SDavid S. Miller 1531c1b4a7e6SDavid S. Miller return cwnd_quota; 1532c1b4a7e6SDavid S. Miller } 1533c1b4a7e6SDavid S. Miller 153467edfef7SAndi Kleen /* Test if sending is allowed right now. */ 1535a2a385d6SEric Dumazet bool tcp_may_send_now(struct sock *sk) 1536c1b4a7e6SDavid S. Miller { 1537cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1538fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 1539c1b4a7e6SDavid S. Miller 1540a02cec21SEric Dumazet return skb && 15410c54b85fSIlpo Järvinen tcp_snd_test(sk, skb, tcp_current_mss(sk), 1542c1b4a7e6SDavid S. Miller (tcp_skb_is_last(sk, skb) ? 1543a02cec21SEric Dumazet tp->nonagle : TCP_NAGLE_PUSH)); 1544c1b4a7e6SDavid S. Miller } 1545c1b4a7e6SDavid S. Miller 1546c1b4a7e6SDavid S. Miller /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 1547c1b4a7e6SDavid S. Miller * which is put after SKB on the list. It is very much like 1548c1b4a7e6SDavid S. Miller * tcp_fragment() except that it may make several kinds of assumptions 1549c1b4a7e6SDavid S. Miller * in order to speed up the splitting operation. In particular, we 1550c1b4a7e6SDavid S. Miller * know that all the data is in scatter-gather pages, and that the 1551c1b4a7e6SDavid S. Miller * packet has never been sent out before (and thus is not cloned). 1552c1b4a7e6SDavid S. Miller */ 1553056834d9SIlpo Järvinen static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, 1554c4ead4c5SEric Dumazet unsigned int mss_now, gfp_t gfp) 1555c1b4a7e6SDavid S. Miller { 1556c1b4a7e6SDavid S. Miller struct sk_buff *buff; 1557c1b4a7e6SDavid S. Miller int nlen = skb->len - len; 15589ce01461SIlpo Järvinen u8 flags; 1559c1b4a7e6SDavid S. Miller 1560c1b4a7e6SDavid S. Miller /* All of a TSO frame must be composed of paged data. */ 1561c8ac3774SHerbert Xu if (skb->len != skb->data_len) 1562c8ac3774SHerbert Xu return tcp_fragment(sk, skb, len, mss_now); 1563c1b4a7e6SDavid S. Miller 1564c4ead4c5SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, gfp); 1565c1b4a7e6SDavid S. Miller if (unlikely(buff == NULL)) 1566c1b4a7e6SDavid S. Miller return -ENOMEM; 1567c1b4a7e6SDavid S. Miller 15683ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 15693ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1570b60b49eaSHerbert Xu buff->truesize += nlen; 1571c1b4a7e6SDavid S. Miller skb->truesize -= nlen; 1572c1b4a7e6SDavid S. Miller 1573c1b4a7e6SDavid S. Miller /* Correct the sequence numbers. */ 1574c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 1575c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 1576c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 1577c1b4a7e6SDavid S. Miller 1578c1b4a7e6SDavid S. Miller /* PSH and FIN should only be set in the second packet. */ 15794de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 15804de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 15814de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1582c1b4a7e6SDavid S. Miller 1583c1b4a7e6SDavid S. Miller /* This packet was never sent out yet, so no SACK bits. */ 1584c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->sacked = 0; 1585c1b4a7e6SDavid S. Miller 158684fa7933SPatrick McHardy buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; 1587c1b4a7e6SDavid S. Miller skb_split(skb, buff, len); 1588c1b4a7e6SDavid S. Miller 1589c1b4a7e6SDavid S. Miller /* Fix up tso_factor for both original and new SKB. */ 1590846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1591846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, buff, mss_now); 1592c1b4a7e6SDavid S. Miller 1593c1b4a7e6SDavid S. Miller /* Link BUFF into the send queue. */ 1594c1b4a7e6SDavid S. Miller skb_header_release(buff); 1595fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 1596c1b4a7e6SDavid S. Miller 1597c1b4a7e6SDavid S. Miller return 0; 1598c1b4a7e6SDavid S. Miller } 1599c1b4a7e6SDavid S. Miller 1600c1b4a7e6SDavid S. Miller /* Try to defer sending, if possible, in order to minimize the amount 1601c1b4a7e6SDavid S. Miller * of TSO splitting we do. View it as a kind of TSO Nagle test. 1602c1b4a7e6SDavid S. Miller * 1603c1b4a7e6SDavid S. Miller * This algorithm is from John Heffner. 1604c1b4a7e6SDavid S. Miller */ 1605a2a385d6SEric Dumazet static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb) 1606c1b4a7e6SDavid S. Miller { 16079e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 16086687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 1609c1b4a7e6SDavid S. Miller u32 send_win, cong_win, limit, in_flight; 1610ad9f4f50SEric Dumazet int win_divisor; 1611c1b4a7e6SDavid S. Miller 16124de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 1613ae8064acSJohn Heffner goto send_now; 1614c1b4a7e6SDavid S. Miller 16156687e988SArnaldo Carvalho de Melo if (icsk->icsk_ca_state != TCP_CA_Open) 1616ae8064acSJohn Heffner goto send_now; 1617ae8064acSJohn Heffner 1618ae8064acSJohn Heffner /* Defer for less than two clock ticks. */ 1619bd515c3eSIlpo Järvinen if (tp->tso_deferred && 1620a2acde07SIlpo Järvinen (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1) 1621ae8064acSJohn Heffner goto send_now; 1622908a75c1SDavid S. Miller 1623c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1624c1b4a7e6SDavid S. Miller 1625056834d9SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight)); 1626c1b4a7e6SDavid S. Miller 162790840defSIlpo Järvinen send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1628c1b4a7e6SDavid S. Miller 1629c1b4a7e6SDavid S. Miller /* From in_flight test above, we know that cwnd > in_flight. */ 1630c1b4a7e6SDavid S. Miller cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; 1631c1b4a7e6SDavid S. Miller 1632c1b4a7e6SDavid S. Miller limit = min(send_win, cong_win); 1633c1b4a7e6SDavid S. Miller 1634ba244fe9SDavid S. Miller /* If a full-sized TSO skb can be sent, do it. */ 16351485348dSBen Hutchings if (limit >= min_t(unsigned int, sk->sk_gso_max_size, 163695bd09ebSEric Dumazet tp->xmit_size_goal_segs * tp->mss_cache)) 1637ae8064acSJohn Heffner goto send_now; 1638ba244fe9SDavid S. Miller 163962ad2761SIlpo Järvinen /* Middle in queue won't get any more data, full sendable already? */ 164062ad2761SIlpo Järvinen if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) 164162ad2761SIlpo Järvinen goto send_now; 164262ad2761SIlpo Järvinen 1643ad9f4f50SEric Dumazet win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor); 1644ad9f4f50SEric Dumazet if (win_divisor) { 1645c1b4a7e6SDavid S. Miller u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); 1646c1b4a7e6SDavid S. Miller 1647c1b4a7e6SDavid S. Miller /* If at least some fraction of a window is available, 1648c1b4a7e6SDavid S. Miller * just use it. 1649c1b4a7e6SDavid S. Miller */ 1650ad9f4f50SEric Dumazet chunk /= win_divisor; 1651c1b4a7e6SDavid S. Miller if (limit >= chunk) 1652ae8064acSJohn Heffner goto send_now; 1653c1b4a7e6SDavid S. Miller } else { 1654c1b4a7e6SDavid S. Miller /* Different approach, try not to defer past a single 1655c1b4a7e6SDavid S. Miller * ACK. Receiver should ACK every other full sized 1656c1b4a7e6SDavid S. Miller * frame, so if we have space for more than 3 frames 1657c1b4a7e6SDavid S. Miller * then send now. 1658c1b4a7e6SDavid S. Miller */ 16596b5a5c0dSNeal Cardwell if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) 1660ae8064acSJohn Heffner goto send_now; 1661c1b4a7e6SDavid S. Miller } 1662c1b4a7e6SDavid S. Miller 1663f4541d60SEric Dumazet /* Ok, it looks like it is advisable to defer. 1664f4541d60SEric Dumazet * Do not rearm the timer if already set to not break TCP ACK clocking. 1665f4541d60SEric Dumazet */ 1666f4541d60SEric Dumazet if (!tp->tso_deferred) 1667ae8064acSJohn Heffner tp->tso_deferred = 1 | (jiffies << 1); 1668ae8064acSJohn Heffner 1669a2a385d6SEric Dumazet return true; 1670ae8064acSJohn Heffner 1671ae8064acSJohn Heffner send_now: 1672ae8064acSJohn Heffner tp->tso_deferred = 0; 1673a2a385d6SEric Dumazet return false; 1674c1b4a7e6SDavid S. Miller } 1675c1b4a7e6SDavid S. Miller 16765d424d5aSJohn Heffner /* Create a new MTU probe if we are ready. 167767edfef7SAndi Kleen * MTU probe is regularly attempting to increase the path MTU by 167867edfef7SAndi Kleen * deliberately sending larger packets. This discovers routing 167967edfef7SAndi Kleen * changes resulting in larger path MTUs. 168067edfef7SAndi Kleen * 16815d424d5aSJohn Heffner * Returns 0 if we should wait to probe (no cwnd available), 16825d424d5aSJohn Heffner * 1 if a probe was sent, 1683056834d9SIlpo Järvinen * -1 otherwise 1684056834d9SIlpo Järvinen */ 16855d424d5aSJohn Heffner static int tcp_mtu_probe(struct sock *sk) 16865d424d5aSJohn Heffner { 16875d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 16885d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 16895d424d5aSJohn Heffner struct sk_buff *skb, *nskb, *next; 16905d424d5aSJohn Heffner int len; 16915d424d5aSJohn Heffner int probe_size; 169291cc17c0SIlpo Järvinen int size_needed; 16935d424d5aSJohn Heffner int copy; 16945d424d5aSJohn Heffner int mss_now; 16955d424d5aSJohn Heffner 16965d424d5aSJohn Heffner /* Not currently probing/verifying, 16975d424d5aSJohn Heffner * not in recovery, 16985d424d5aSJohn Heffner * have enough cwnd, and 16995d424d5aSJohn Heffner * not SACKing (the variable headers throw things off) */ 17005d424d5aSJohn Heffner if (!icsk->icsk_mtup.enabled || 17015d424d5aSJohn Heffner icsk->icsk_mtup.probe_size || 17025d424d5aSJohn Heffner inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 17035d424d5aSJohn Heffner tp->snd_cwnd < 11 || 1704cabeccbdSIlpo Järvinen tp->rx_opt.num_sacks || tp->rx_opt.dsack) 17055d424d5aSJohn Heffner return -1; 17065d424d5aSJohn Heffner 17075d424d5aSJohn Heffner /* Very simple search strategy: just double the MSS. */ 17080c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 17095d424d5aSJohn Heffner probe_size = 2 * tp->mss_cache; 171091cc17c0SIlpo Järvinen size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; 17115d424d5aSJohn Heffner if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) { 17125d424d5aSJohn Heffner /* TODO: set timer for probe_converge_event */ 17135d424d5aSJohn Heffner return -1; 17145d424d5aSJohn Heffner } 17155d424d5aSJohn Heffner 17165d424d5aSJohn Heffner /* Have enough data in the send queue to probe? */ 17177f9c33e5SIlpo Järvinen if (tp->write_seq - tp->snd_nxt < size_needed) 17185d424d5aSJohn Heffner return -1; 17195d424d5aSJohn Heffner 172091cc17c0SIlpo Järvinen if (tp->snd_wnd < size_needed) 17215d424d5aSJohn Heffner return -1; 172290840defSIlpo Järvinen if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) 17235d424d5aSJohn Heffner return 0; 17245d424d5aSJohn Heffner 1725d67c58e9SIlpo Järvinen /* Do we need to wait to drain cwnd? With none in flight, don't stall */ 1726d67c58e9SIlpo Järvinen if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { 1727d67c58e9SIlpo Järvinen if (!tcp_packets_in_flight(tp)) 17285d424d5aSJohn Heffner return -1; 17295d424d5aSJohn Heffner else 17305d424d5aSJohn Heffner return 0; 17315d424d5aSJohn Heffner } 17325d424d5aSJohn Heffner 17335d424d5aSJohn Heffner /* We're allowed to probe. Build it now. */ 17345d424d5aSJohn Heffner if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL) 17355d424d5aSJohn Heffner return -1; 17363ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 17373ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 17385d424d5aSJohn Heffner 1739fe067e8aSDavid S. Miller skb = tcp_send_head(sk); 17405d424d5aSJohn Heffner 17415d424d5aSJohn Heffner TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 17425d424d5aSJohn Heffner TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 17434de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; 17445d424d5aSJohn Heffner TCP_SKB_CB(nskb)->sacked = 0; 17455d424d5aSJohn Heffner nskb->csum = 0; 174684fa7933SPatrick McHardy nskb->ip_summed = skb->ip_summed; 17475d424d5aSJohn Heffner 174850c4817eSIlpo Järvinen tcp_insert_write_queue_before(nskb, skb, sk); 174950c4817eSIlpo Järvinen 17505d424d5aSJohn Heffner len = 0; 1751234b6860SIlpo Järvinen tcp_for_write_queue_from_safe(skb, next, sk) { 17525d424d5aSJohn Heffner copy = min_t(int, skb->len, probe_size - len); 17535d424d5aSJohn Heffner if (nskb->ip_summed) 17545d424d5aSJohn Heffner skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); 17555d424d5aSJohn Heffner else 17565d424d5aSJohn Heffner nskb->csum = skb_copy_and_csum_bits(skb, 0, 1757056834d9SIlpo Järvinen skb_put(nskb, copy), 1758056834d9SIlpo Järvinen copy, nskb->csum); 17595d424d5aSJohn Heffner 17605d424d5aSJohn Heffner if (skb->len <= copy) { 17615d424d5aSJohn Heffner /* We've eaten all the data from this skb. 17625d424d5aSJohn Heffner * Throw it away. */ 17634de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 1764fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 17653ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 17665d424d5aSJohn Heffner } else { 17674de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & 1768a3433f35SChangli Gao ~(TCPHDR_FIN|TCPHDR_PSH); 17695d424d5aSJohn Heffner if (!skb_shinfo(skb)->nr_frags) { 17705d424d5aSJohn Heffner skb_pull(skb, copy); 177184fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 1772056834d9SIlpo Järvinen skb->csum = csum_partial(skb->data, 1773056834d9SIlpo Järvinen skb->len, 0); 17745d424d5aSJohn Heffner } else { 17755d424d5aSJohn Heffner __pskb_trim_head(skb, copy); 17765d424d5aSJohn Heffner tcp_set_skb_tso_segs(sk, skb, mss_now); 17775d424d5aSJohn Heffner } 17785d424d5aSJohn Heffner TCP_SKB_CB(skb)->seq += copy; 17795d424d5aSJohn Heffner } 17805d424d5aSJohn Heffner 17815d424d5aSJohn Heffner len += copy; 1782234b6860SIlpo Järvinen 1783234b6860SIlpo Järvinen if (len >= probe_size) 1784234b6860SIlpo Järvinen break; 17855d424d5aSJohn Heffner } 17865d424d5aSJohn Heffner tcp_init_tso_segs(sk, nskb, nskb->len); 17875d424d5aSJohn Heffner 17885d424d5aSJohn Heffner /* We're ready to send. If this fails, the probe will 17895d424d5aSJohn Heffner * be resegmented into mss-sized pieces by tcp_write_xmit(). */ 17905d424d5aSJohn Heffner TCP_SKB_CB(nskb)->when = tcp_time_stamp; 17915d424d5aSJohn Heffner if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 17925d424d5aSJohn Heffner /* Decrement cwnd here because we are sending 17935d424d5aSJohn Heffner * effectively two packets. */ 17945d424d5aSJohn Heffner tp->snd_cwnd--; 179566f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, nskb); 17965d424d5aSJohn Heffner 17975d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 17980e7b1368SJohn Heffner tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 17990e7b1368SJohn Heffner tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 18005d424d5aSJohn Heffner 18015d424d5aSJohn Heffner return 1; 18025d424d5aSJohn Heffner } 18035d424d5aSJohn Heffner 18045d424d5aSJohn Heffner return -1; 18055d424d5aSJohn Heffner } 18065d424d5aSJohn Heffner 18071da177e4SLinus Torvalds /* This routine writes packets to the network. It advances the 18081da177e4SLinus Torvalds * send_head. This happens as incoming acks open up the remote 18091da177e4SLinus Torvalds * window for us. 18101da177e4SLinus Torvalds * 1811f8269a49SIlpo Järvinen * LARGESEND note: !tcp_urg_mode is overkill, only frames between 1812f8269a49SIlpo Järvinen * snd_up-64k-mss .. snd_up cannot be large. However, taking into 1813f8269a49SIlpo Järvinen * account rare use of URG, this is not a big flaw. 1814f8269a49SIlpo Järvinen * 18156ba8a3b1SNandita Dukkipati * Send at most one packet when push_one > 0. Temporarily ignore 18166ba8a3b1SNandita Dukkipati * cwnd limit to force at most one packet out when push_one == 2. 18176ba8a3b1SNandita Dukkipati 1818a2a385d6SEric Dumazet * Returns true, if no segments are in flight and we have queued segments, 1819a2a385d6SEric Dumazet * but cannot send anything now because of SWS or another problem. 18201da177e4SLinus Torvalds */ 1821a2a385d6SEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 1822d5dd9175SIlpo Järvinen int push_one, gfp_t gfp) 18231da177e4SLinus Torvalds { 18241da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 182592df7b51SDavid S. Miller struct sk_buff *skb; 1826c1b4a7e6SDavid S. Miller unsigned int tso_segs, sent_pkts; 1827c1b4a7e6SDavid S. Miller int cwnd_quota; 18285d424d5aSJohn Heffner int result; 18291da177e4SLinus Torvalds 1830c1b4a7e6SDavid S. Miller sent_pkts = 0; 18315d424d5aSJohn Heffner 1832d5dd9175SIlpo Järvinen if (!push_one) { 18335d424d5aSJohn Heffner /* Do MTU probing. */ 1834d5dd9175SIlpo Järvinen result = tcp_mtu_probe(sk); 1835d5dd9175SIlpo Järvinen if (!result) { 1836a2a385d6SEric Dumazet return false; 18375d424d5aSJohn Heffner } else if (result > 0) { 18385d424d5aSJohn Heffner sent_pkts = 1; 18395d424d5aSJohn Heffner } 1840d5dd9175SIlpo Järvinen } 18415d424d5aSJohn Heffner 1842fe067e8aSDavid S. Miller while ((skb = tcp_send_head(sk))) { 1843c8ac3774SHerbert Xu unsigned int limit; 1844c8ac3774SHerbert Xu 1845b68e9f85SHerbert Xu tso_segs = tcp_init_tso_segs(sk, skb, mss_now); 1846c1b4a7e6SDavid S. Miller BUG_ON(!tso_segs); 1847c1b4a7e6SDavid S. Miller 1848ec342325SAndrew Vagin if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) 1849ec342325SAndrew Vagin goto repair; /* Skip network transmission */ 1850ec342325SAndrew Vagin 1851b68e9f85SHerbert Xu cwnd_quota = tcp_cwnd_test(tp, skb); 18526ba8a3b1SNandita Dukkipati if (!cwnd_quota) { 18536ba8a3b1SNandita Dukkipati if (push_one == 2) 18546ba8a3b1SNandita Dukkipati /* Force out a loss probe pkt. */ 18556ba8a3b1SNandita Dukkipati cwnd_quota = 1; 18566ba8a3b1SNandita Dukkipati else 1857b68e9f85SHerbert Xu break; 18586ba8a3b1SNandita Dukkipati } 1859b68e9f85SHerbert Xu 1860b68e9f85SHerbert Xu if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) 1861b68e9f85SHerbert Xu break; 1862b68e9f85SHerbert Xu 1863c1b4a7e6SDavid S. Miller if (tso_segs == 1) { 1864aa93466bSDavid S. Miller if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 1865aa93466bSDavid S. Miller (tcp_skb_is_last(sk, skb) ? 1866aa93466bSDavid S. Miller nonagle : TCP_NAGLE_PUSH)))) 1867aa93466bSDavid S. Miller break; 1868c1b4a7e6SDavid S. Miller } else { 1869d5dd9175SIlpo Järvinen if (!push_one && tcp_tso_should_defer(sk, skb)) 1870aa93466bSDavid S. Miller break; 1871c1b4a7e6SDavid S. Miller } 1872aa93466bSDavid S. Miller 1873c9eeec26SEric Dumazet /* TCP Small Queues : 1874c9eeec26SEric Dumazet * Control number of packets in qdisc/devices to two packets / or ~1 ms. 1875c9eeec26SEric Dumazet * This allows for : 1876c9eeec26SEric Dumazet * - better RTT estimation and ACK scheduling 1877c9eeec26SEric Dumazet * - faster recovery 1878c9eeec26SEric Dumazet * - high rates 187946d3ceabSEric Dumazet */ 1880c9eeec26SEric Dumazet limit = max(skb->truesize, sk->sk_pacing_rate >> 10); 1881c9eeec26SEric Dumazet 1882c9eeec26SEric Dumazet if (atomic_read(&sk->sk_wmem_alloc) > limit) { 188346d3ceabSEric Dumazet set_bit(TSQ_THROTTLED, &tp->tsq_flags); 188446d3ceabSEric Dumazet break; 188546d3ceabSEric Dumazet } 1886c9eeec26SEric Dumazet 1887c8ac3774SHerbert Xu limit = mss_now; 1888f8269a49SIlpo Järvinen if (tso_segs > 1 && !tcp_urg_mode(tp)) 18890e3a4803SIlpo Järvinen limit = tcp_mss_split_point(sk, skb, mss_now, 18901485348dSBen Hutchings min_t(unsigned int, 18911485348dSBen Hutchings cwnd_quota, 18921485348dSBen Hutchings sk->sk_gso_max_segs)); 1893c8ac3774SHerbert Xu 1894c8ac3774SHerbert Xu if (skb->len > limit && 1895c4ead4c5SEric Dumazet unlikely(tso_fragment(sk, skb, limit, mss_now, gfp))) 18961da177e4SLinus Torvalds break; 18971da177e4SLinus Torvalds 18981da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 1899c1b4a7e6SDavid S. Miller 1900d5dd9175SIlpo Järvinen if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) 19011da177e4SLinus Torvalds break; 19021da177e4SLinus Torvalds 1903ec342325SAndrew Vagin repair: 19041da177e4SLinus Torvalds /* Advance the send_head. This one is sent out. 19051da177e4SLinus Torvalds * This call will increment packets_out. 19061da177e4SLinus Torvalds */ 190766f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 19081da177e4SLinus Torvalds 19091da177e4SLinus Torvalds tcp_minshall_update(tp, mss_now, skb); 1910a262f0cdSNandita Dukkipati sent_pkts += tcp_skb_pcount(skb); 1911d5dd9175SIlpo Järvinen 1912d5dd9175SIlpo Järvinen if (push_one) 1913d5dd9175SIlpo Järvinen break; 19141da177e4SLinus Torvalds } 19151da177e4SLinus Torvalds 1916aa93466bSDavid S. Miller if (likely(sent_pkts)) { 1917684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 1918684bad11SYuchung Cheng tp->prr_out += sent_pkts; 19196ba8a3b1SNandita Dukkipati 19206ba8a3b1SNandita Dukkipati /* Send one loss probe per tail loss episode. */ 19216ba8a3b1SNandita Dukkipati if (push_one != 2) 19226ba8a3b1SNandita Dukkipati tcp_schedule_loss_probe(sk); 19239e412ba7SIlpo Järvinen tcp_cwnd_validate(sk); 1924a2a385d6SEric Dumazet return false; 19251da177e4SLinus Torvalds } 19266ba8a3b1SNandita Dukkipati return (push_one == 2) || (!tp->packets_out && tcp_send_head(sk)); 19276ba8a3b1SNandita Dukkipati } 19286ba8a3b1SNandita Dukkipati 19296ba8a3b1SNandita Dukkipati bool tcp_schedule_loss_probe(struct sock *sk) 19306ba8a3b1SNandita Dukkipati { 19316ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 19326ba8a3b1SNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 19336ba8a3b1SNandita Dukkipati u32 timeout, tlp_time_stamp, rto_time_stamp; 19346ba8a3b1SNandita Dukkipati u32 rtt = tp->srtt >> 3; 19356ba8a3b1SNandita Dukkipati 19366ba8a3b1SNandita Dukkipati if (WARN_ON(icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS)) 19376ba8a3b1SNandita Dukkipati return false; 19386ba8a3b1SNandita Dukkipati /* No consecutive loss probes. */ 19396ba8a3b1SNandita Dukkipati if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) { 19406ba8a3b1SNandita Dukkipati tcp_rearm_rto(sk); 19416ba8a3b1SNandita Dukkipati return false; 19426ba8a3b1SNandita Dukkipati } 19436ba8a3b1SNandita Dukkipati /* Don't do any loss probe on a Fast Open connection before 3WHS 19446ba8a3b1SNandita Dukkipati * finishes. 19456ba8a3b1SNandita Dukkipati */ 19466ba8a3b1SNandita Dukkipati if (sk->sk_state == TCP_SYN_RECV) 19476ba8a3b1SNandita Dukkipati return false; 19486ba8a3b1SNandita Dukkipati 19496ba8a3b1SNandita Dukkipati /* TLP is only scheduled when next timer event is RTO. */ 19506ba8a3b1SNandita Dukkipati if (icsk->icsk_pending != ICSK_TIME_RETRANS) 19516ba8a3b1SNandita Dukkipati return false; 19526ba8a3b1SNandita Dukkipati 19536ba8a3b1SNandita Dukkipati /* Schedule a loss probe in 2*RTT for SACK capable connections 19546ba8a3b1SNandita Dukkipati * in Open state, that are either limited by cwnd or application. 19556ba8a3b1SNandita Dukkipati */ 19566ba8a3b1SNandita Dukkipati if (sysctl_tcp_early_retrans < 3 || !rtt || !tp->packets_out || 19576ba8a3b1SNandita Dukkipati !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open) 19586ba8a3b1SNandita Dukkipati return false; 19596ba8a3b1SNandita Dukkipati 19606ba8a3b1SNandita Dukkipati if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) && 19616ba8a3b1SNandita Dukkipati tcp_send_head(sk)) 19626ba8a3b1SNandita Dukkipati return false; 19636ba8a3b1SNandita Dukkipati 19646ba8a3b1SNandita Dukkipati /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account 19656ba8a3b1SNandita Dukkipati * for delayed ack when there's one outstanding packet. 19666ba8a3b1SNandita Dukkipati */ 19676ba8a3b1SNandita Dukkipati timeout = rtt << 1; 19686ba8a3b1SNandita Dukkipati if (tp->packets_out == 1) 19696ba8a3b1SNandita Dukkipati timeout = max_t(u32, timeout, 19706ba8a3b1SNandita Dukkipati (rtt + (rtt >> 1) + TCP_DELACK_MAX)); 19716ba8a3b1SNandita Dukkipati timeout = max_t(u32, timeout, msecs_to_jiffies(10)); 19726ba8a3b1SNandita Dukkipati 19736ba8a3b1SNandita Dukkipati /* If RTO is shorter, just schedule TLP in its place. */ 19746ba8a3b1SNandita Dukkipati tlp_time_stamp = tcp_time_stamp + timeout; 19756ba8a3b1SNandita Dukkipati rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout; 19766ba8a3b1SNandita Dukkipati if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) { 19776ba8a3b1SNandita Dukkipati s32 delta = rto_time_stamp - tcp_time_stamp; 19786ba8a3b1SNandita Dukkipati if (delta > 0) 19796ba8a3b1SNandita Dukkipati timeout = delta; 19806ba8a3b1SNandita Dukkipati } 19816ba8a3b1SNandita Dukkipati 19826ba8a3b1SNandita Dukkipati inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, 19836ba8a3b1SNandita Dukkipati TCP_RTO_MAX); 19846ba8a3b1SNandita Dukkipati return true; 19856ba8a3b1SNandita Dukkipati } 19866ba8a3b1SNandita Dukkipati 19876ba8a3b1SNandita Dukkipati /* When probe timeout (PTO) fires, send a new segment if one exists, else 19886ba8a3b1SNandita Dukkipati * retransmit the last segment. 19896ba8a3b1SNandita Dukkipati */ 19906ba8a3b1SNandita Dukkipati void tcp_send_loss_probe(struct sock *sk) 19916ba8a3b1SNandita Dukkipati { 19929b717a8dSNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 19936ba8a3b1SNandita Dukkipati struct sk_buff *skb; 19946ba8a3b1SNandita Dukkipati int pcount; 19956ba8a3b1SNandita Dukkipati int mss = tcp_current_mss(sk); 19966ba8a3b1SNandita Dukkipati int err = -1; 19976ba8a3b1SNandita Dukkipati 19986ba8a3b1SNandita Dukkipati if (tcp_send_head(sk) != NULL) { 19996ba8a3b1SNandita Dukkipati err = tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC); 20006ba8a3b1SNandita Dukkipati goto rearm_timer; 20016ba8a3b1SNandita Dukkipati } 20026ba8a3b1SNandita Dukkipati 20039b717a8dSNandita Dukkipati /* At most one outstanding TLP retransmission. */ 20049b717a8dSNandita Dukkipati if (tp->tlp_high_seq) 20059b717a8dSNandita Dukkipati goto rearm_timer; 20069b717a8dSNandita Dukkipati 20076ba8a3b1SNandita Dukkipati /* Retransmit last segment. */ 20086ba8a3b1SNandita Dukkipati skb = tcp_write_queue_tail(sk); 20096ba8a3b1SNandita Dukkipati if (WARN_ON(!skb)) 20106ba8a3b1SNandita Dukkipati goto rearm_timer; 20116ba8a3b1SNandita Dukkipati 20126ba8a3b1SNandita Dukkipati pcount = tcp_skb_pcount(skb); 20136ba8a3b1SNandita Dukkipati if (WARN_ON(!pcount)) 20146ba8a3b1SNandita Dukkipati goto rearm_timer; 20156ba8a3b1SNandita Dukkipati 20166ba8a3b1SNandita Dukkipati if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) { 20176ba8a3b1SNandita Dukkipati if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss))) 20186ba8a3b1SNandita Dukkipati goto rearm_timer; 20196ba8a3b1SNandita Dukkipati skb = tcp_write_queue_tail(sk); 20206ba8a3b1SNandita Dukkipati } 20216ba8a3b1SNandita Dukkipati 20226ba8a3b1SNandita Dukkipati if (WARN_ON(!skb || !tcp_skb_pcount(skb))) 20236ba8a3b1SNandita Dukkipati goto rearm_timer; 20246ba8a3b1SNandita Dukkipati 20256ba8a3b1SNandita Dukkipati /* Probe with zero data doesn't trigger fast recovery. */ 20266ba8a3b1SNandita Dukkipati if (skb->len > 0) 20276ba8a3b1SNandita Dukkipati err = __tcp_retransmit_skb(sk, skb); 20286ba8a3b1SNandita Dukkipati 20299b717a8dSNandita Dukkipati /* Record snd_nxt for loss detection. */ 20309b717a8dSNandita Dukkipati if (likely(!err)) 20319b717a8dSNandita Dukkipati tp->tlp_high_seq = tp->snd_nxt; 20329b717a8dSNandita Dukkipati 20336ba8a3b1SNandita Dukkipati rearm_timer: 20346ba8a3b1SNandita Dukkipati inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 20356ba8a3b1SNandita Dukkipati inet_csk(sk)->icsk_rto, 20366ba8a3b1SNandita Dukkipati TCP_RTO_MAX); 20376ba8a3b1SNandita Dukkipati 20386ba8a3b1SNandita Dukkipati if (likely(!err)) 20396ba8a3b1SNandita Dukkipati NET_INC_STATS_BH(sock_net(sk), 20406ba8a3b1SNandita Dukkipati LINUX_MIB_TCPLOSSPROBES); 20416ba8a3b1SNandita Dukkipati return; 20421da177e4SLinus Torvalds } 20431da177e4SLinus Torvalds 2044a762a980SDavid S. Miller /* Push out any pending frames which were held back due to 2045a762a980SDavid S. Miller * TCP_CORK or attempt at coalescing tiny packets. 2046a762a980SDavid S. Miller * The socket must be locked by the caller. 2047a762a980SDavid S. Miller */ 20489e412ba7SIlpo Järvinen void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 20499e412ba7SIlpo Järvinen int nonagle) 2050a762a980SDavid S. Miller { 2051726e07a8SIlpo Järvinen /* If we are closed, the bytes will have to remain here. 2052726e07a8SIlpo Järvinen * In time closedown will finish, we empty the write queue and 2053726e07a8SIlpo Järvinen * all will be happy. 2054726e07a8SIlpo Järvinen */ 2055726e07a8SIlpo Järvinen if (unlikely(sk->sk_state == TCP_CLOSE)) 2056726e07a8SIlpo Järvinen return; 2057726e07a8SIlpo Järvinen 205899a1dec7SMel Gorman if (tcp_write_xmit(sk, cur_mss, nonagle, 0, 205999a1dec7SMel Gorman sk_gfp_atomic(sk, GFP_ATOMIC))) 20609e412ba7SIlpo Järvinen tcp_check_probe_timer(sk); 2061a762a980SDavid S. Miller } 2062a762a980SDavid S. Miller 2063c1b4a7e6SDavid S. Miller /* Send _single_ skb sitting at the send head. This function requires 2064c1b4a7e6SDavid S. Miller * true push pending frames to setup probe timer etc. 2065c1b4a7e6SDavid S. Miller */ 2066c1b4a7e6SDavid S. Miller void tcp_push_one(struct sock *sk, unsigned int mss_now) 2067c1b4a7e6SDavid S. Miller { 2068fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 2069c1b4a7e6SDavid S. Miller 2070c1b4a7e6SDavid S. Miller BUG_ON(!skb || skb->len < mss_now); 2071c1b4a7e6SDavid S. Miller 2072d5dd9175SIlpo Järvinen tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); 2073c1b4a7e6SDavid S. Miller } 2074c1b4a7e6SDavid S. Miller 20751da177e4SLinus Torvalds /* This function returns the amount that we can raise the 20761da177e4SLinus Torvalds * usable window based on the following constraints 20771da177e4SLinus Torvalds * 20781da177e4SLinus Torvalds * 1. The window can never be shrunk once it is offered (RFC 793) 20791da177e4SLinus Torvalds * 2. We limit memory per socket 20801da177e4SLinus Torvalds * 20811da177e4SLinus Torvalds * RFC 1122: 20821da177e4SLinus Torvalds * "the suggested [SWS] avoidance algorithm for the receiver is to keep 20831da177e4SLinus Torvalds * RECV.NEXT + RCV.WIN fixed until: 20841da177e4SLinus Torvalds * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 20851da177e4SLinus Torvalds * 20861da177e4SLinus Torvalds * i.e. don't raise the right edge of the window until you can raise 20871da177e4SLinus Torvalds * it at least MSS bytes. 20881da177e4SLinus Torvalds * 20891da177e4SLinus Torvalds * Unfortunately, the recommended algorithm breaks header prediction, 20901da177e4SLinus Torvalds * since header prediction assumes th->window stays fixed. 20911da177e4SLinus Torvalds * 20921da177e4SLinus Torvalds * Strictly speaking, keeping th->window fixed violates the receiver 20931da177e4SLinus Torvalds * side SWS prevention criteria. The problem is that under this rule 20941da177e4SLinus Torvalds * a stream of single byte packets will cause the right side of the 20951da177e4SLinus Torvalds * window to always advance by a single byte. 20961da177e4SLinus Torvalds * 20971da177e4SLinus Torvalds * Of course, if the sender implements sender side SWS prevention 20981da177e4SLinus Torvalds * then this will not be a problem. 20991da177e4SLinus Torvalds * 21001da177e4SLinus Torvalds * BSD seems to make the following compromise: 21011da177e4SLinus Torvalds * 21021da177e4SLinus Torvalds * If the free space is less than the 1/4 of the maximum 21031da177e4SLinus Torvalds * space available and the free space is less than 1/2 mss, 21041da177e4SLinus Torvalds * then set the window to 0. 21051da177e4SLinus Torvalds * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 21061da177e4SLinus Torvalds * Otherwise, just prevent the window from shrinking 21071da177e4SLinus Torvalds * and from being larger than the largest representable value. 21081da177e4SLinus Torvalds * 21091da177e4SLinus Torvalds * This prevents incremental opening of the window in the regime 21101da177e4SLinus Torvalds * where TCP is limited by the speed of the reader side taking 21111da177e4SLinus Torvalds * data out of the TCP receive queue. It does nothing about 21121da177e4SLinus Torvalds * those cases where the window is constrained on the sender side 21131da177e4SLinus Torvalds * because the pipeline is full. 21141da177e4SLinus Torvalds * 21151da177e4SLinus Torvalds * BSD also seems to "accidentally" limit itself to windows that are a 21161da177e4SLinus Torvalds * multiple of MSS, at least until the free space gets quite small. 21171da177e4SLinus Torvalds * This would appear to be a side effect of the mbuf implementation. 21181da177e4SLinus Torvalds * Combining these two algorithms results in the observed behavior 21191da177e4SLinus Torvalds * of having a fixed window size at almost all times. 21201da177e4SLinus Torvalds * 21211da177e4SLinus Torvalds * Below we obtain similar behavior by forcing the offered window to 21221da177e4SLinus Torvalds * a multiple of the mss when it is feasible to do so. 21231da177e4SLinus Torvalds * 21241da177e4SLinus Torvalds * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 21251da177e4SLinus Torvalds * Regular options like TIMESTAMP are taken into account. 21261da177e4SLinus Torvalds */ 21271da177e4SLinus Torvalds u32 __tcp_select_window(struct sock *sk) 21281da177e4SLinus Torvalds { 2129463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 21301da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2131caa20d9aSStephen Hemminger /* MSS for the peer's data. Previous versions used mss_clamp 21321da177e4SLinus Torvalds * here. I don't know if the value based on our guesses 21331da177e4SLinus Torvalds * of peer's MSS is better for the performance. It's more correct 21341da177e4SLinus Torvalds * but may be worse for the performance because of rcv_mss 21351da177e4SLinus Torvalds * fluctuations. --SAW 1998/11/1 21361da177e4SLinus Torvalds */ 2137463c84b9SArnaldo Carvalho de Melo int mss = icsk->icsk_ack.rcv_mss; 21381da177e4SLinus Torvalds int free_space = tcp_space(sk); 21391da177e4SLinus Torvalds int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk)); 21401da177e4SLinus Torvalds int window; 21411da177e4SLinus Torvalds 21421da177e4SLinus Torvalds if (mss > full_space) 21431da177e4SLinus Torvalds mss = full_space; 21441da177e4SLinus Torvalds 2145b92edbe0SEric Dumazet if (free_space < (full_space >> 1)) { 2146463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.quick = 0; 21471da177e4SLinus Torvalds 2148180d8cd9SGlauber Costa if (sk_under_memory_pressure(sk)) 2149056834d9SIlpo Järvinen tp->rcv_ssthresh = min(tp->rcv_ssthresh, 2150056834d9SIlpo Järvinen 4U * tp->advmss); 21511da177e4SLinus Torvalds 21521da177e4SLinus Torvalds if (free_space < mss) 21531da177e4SLinus Torvalds return 0; 21541da177e4SLinus Torvalds } 21551da177e4SLinus Torvalds 21561da177e4SLinus Torvalds if (free_space > tp->rcv_ssthresh) 21571da177e4SLinus Torvalds free_space = tp->rcv_ssthresh; 21581da177e4SLinus Torvalds 21591da177e4SLinus Torvalds /* Don't do rounding if we are using window scaling, since the 21601da177e4SLinus Torvalds * scaled window will not line up with the MSS boundary anyway. 21611da177e4SLinus Torvalds */ 21621da177e4SLinus Torvalds window = tp->rcv_wnd; 21631da177e4SLinus Torvalds if (tp->rx_opt.rcv_wscale) { 21641da177e4SLinus Torvalds window = free_space; 21651da177e4SLinus Torvalds 21661da177e4SLinus Torvalds /* Advertise enough space so that it won't get scaled away. 21671da177e4SLinus Torvalds * Import case: prevent zero window announcement if 21681da177e4SLinus Torvalds * 1<<rcv_wscale > mss. 21691da177e4SLinus Torvalds */ 21701da177e4SLinus Torvalds if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window) 21711da177e4SLinus Torvalds window = (((window >> tp->rx_opt.rcv_wscale) + 1) 21721da177e4SLinus Torvalds << tp->rx_opt.rcv_wscale); 21731da177e4SLinus Torvalds } else { 21741da177e4SLinus Torvalds /* Get the largest window that is a nice multiple of mss. 21751da177e4SLinus Torvalds * Window clamp already applied above. 21761da177e4SLinus Torvalds * If our current window offering is within 1 mss of the 21771da177e4SLinus Torvalds * free space we just keep it. This prevents the divide 21781da177e4SLinus Torvalds * and multiply from happening most of the time. 21791da177e4SLinus Torvalds * We also don't do any window rounding when the free space 21801da177e4SLinus Torvalds * is too small. 21811da177e4SLinus Torvalds */ 21821da177e4SLinus Torvalds if (window <= free_space - mss || window > free_space) 21831da177e4SLinus Torvalds window = (free_space / mss) * mss; 218484565070SJohn Heffner else if (mss == full_space && 2185b92edbe0SEric Dumazet free_space > window + (full_space >> 1)) 218684565070SJohn Heffner window = free_space; 21871da177e4SLinus Torvalds } 21881da177e4SLinus Torvalds 21891da177e4SLinus Torvalds return window; 21901da177e4SLinus Torvalds } 21911da177e4SLinus Torvalds 21924a17fc3aSIlpo Järvinen /* Collapses two adjacent SKB's during retransmission. */ 21934a17fc3aSIlpo Järvinen static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) 21941da177e4SLinus Torvalds { 21951da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2196fe067e8aSDavid S. Miller struct sk_buff *next_skb = tcp_write_queue_next(sk, skb); 2197058dc334SIlpo Järvinen int skb_size, next_skb_size; 21981da177e4SLinus Torvalds 2199058dc334SIlpo Järvinen skb_size = skb->len; 2200058dc334SIlpo Järvinen next_skb_size = next_skb->len; 22011da177e4SLinus Torvalds 2202058dc334SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); 22031da177e4SLinus Torvalds 22046859d494SIlpo Järvinen tcp_highest_sack_combine(sk, next_skb, skb); 2205a6963a6bSIlpo Järvinen 2206fe067e8aSDavid S. Miller tcp_unlink_write_queue(next_skb, sk); 22071da177e4SLinus Torvalds 2208058dc334SIlpo Järvinen skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size), 22091a4e2d09SArnaldo Carvalho de Melo next_skb_size); 22101da177e4SLinus Torvalds 221152d570aaSJarek Poplawski if (next_skb->ip_summed == CHECKSUM_PARTIAL) 221252d570aaSJarek Poplawski skb->ip_summed = CHECKSUM_PARTIAL; 22131da177e4SLinus Torvalds 221484fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 22151da177e4SLinus Torvalds skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); 22161da177e4SLinus Torvalds 22171da177e4SLinus Torvalds /* Update sequence range on original skb. */ 22181da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 22191da177e4SLinus Torvalds 2220e6c7d085SIlpo Järvinen /* Merge over control information. This moves PSH/FIN etc. over */ 22214de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; 22221da177e4SLinus Torvalds 22231da177e4SLinus Torvalds /* All done, get rid of second SKB and account for it so 22241da177e4SLinus Torvalds * packet counting does not break. 22251da177e4SLinus Torvalds */ 22264828e7f4SIlpo Järvinen TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; 2227b7689205SIlpo Järvinen 2228b7689205SIlpo Järvinen /* changed transmit queue under us so clear hints */ 2229ef9da47cSIlpo Järvinen tcp_clear_retrans_hints_partial(tp); 2230ef9da47cSIlpo Järvinen if (next_skb == tp->retransmit_skb_hint) 2231ef9da47cSIlpo Järvinen tp->retransmit_skb_hint = skb; 2232b7689205SIlpo Järvinen 2233797108d1SIlpo Järvinen tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); 2234797108d1SIlpo Järvinen 22353ab224beSHideo Aoki sk_wmem_free_skb(sk, next_skb); 22361da177e4SLinus Torvalds } 22371da177e4SLinus Torvalds 223867edfef7SAndi Kleen /* Check if coalescing SKBs is legal. */ 2239a2a385d6SEric Dumazet static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) 22404a17fc3aSIlpo Järvinen { 22414a17fc3aSIlpo Järvinen if (tcp_skb_pcount(skb) > 1) 2242a2a385d6SEric Dumazet return false; 22434a17fc3aSIlpo Järvinen /* TODO: SACK collapsing could be used to remove this condition */ 22444a17fc3aSIlpo Järvinen if (skb_shinfo(skb)->nr_frags != 0) 2245a2a385d6SEric Dumazet return false; 22464a17fc3aSIlpo Järvinen if (skb_cloned(skb)) 2247a2a385d6SEric Dumazet return false; 22484a17fc3aSIlpo Järvinen if (skb == tcp_send_head(sk)) 2249a2a385d6SEric Dumazet return false; 22504a17fc3aSIlpo Järvinen /* Some heurestics for collapsing over SACK'd could be invented */ 22514a17fc3aSIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 2252a2a385d6SEric Dumazet return false; 22534a17fc3aSIlpo Järvinen 2254a2a385d6SEric Dumazet return true; 22554a17fc3aSIlpo Järvinen } 22564a17fc3aSIlpo Järvinen 225767edfef7SAndi Kleen /* Collapse packets in the retransmit queue to make to create 225867edfef7SAndi Kleen * less packets on the wire. This is only done on retransmission. 225967edfef7SAndi Kleen */ 22604a17fc3aSIlpo Järvinen static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, 22614a17fc3aSIlpo Järvinen int space) 22624a17fc3aSIlpo Järvinen { 22634a17fc3aSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 22644a17fc3aSIlpo Järvinen struct sk_buff *skb = to, *tmp; 2265a2a385d6SEric Dumazet bool first = true; 22664a17fc3aSIlpo Järvinen 22674a17fc3aSIlpo Järvinen if (!sysctl_tcp_retrans_collapse) 22684a17fc3aSIlpo Järvinen return; 22694de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 22704a17fc3aSIlpo Järvinen return; 22714a17fc3aSIlpo Järvinen 22724a17fc3aSIlpo Järvinen tcp_for_write_queue_from_safe(skb, tmp, sk) { 22734a17fc3aSIlpo Järvinen if (!tcp_can_collapse(sk, skb)) 22744a17fc3aSIlpo Järvinen break; 22754a17fc3aSIlpo Järvinen 22764a17fc3aSIlpo Järvinen space -= skb->len; 22774a17fc3aSIlpo Järvinen 22784a17fc3aSIlpo Järvinen if (first) { 2279a2a385d6SEric Dumazet first = false; 22804a17fc3aSIlpo Järvinen continue; 22814a17fc3aSIlpo Järvinen } 22824a17fc3aSIlpo Järvinen 22834a17fc3aSIlpo Järvinen if (space < 0) 22844a17fc3aSIlpo Järvinen break; 22854a17fc3aSIlpo Järvinen /* Punt if not enough space exists in the first SKB for 22864a17fc3aSIlpo Järvinen * the data in the second 22874a17fc3aSIlpo Järvinen */ 2288a21d4572SEric Dumazet if (skb->len > skb_availroom(to)) 22894a17fc3aSIlpo Järvinen break; 22904a17fc3aSIlpo Järvinen 22914a17fc3aSIlpo Järvinen if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) 22924a17fc3aSIlpo Järvinen break; 22934a17fc3aSIlpo Järvinen 22944a17fc3aSIlpo Järvinen tcp_collapse_retrans(sk, to); 22954a17fc3aSIlpo Järvinen } 22964a17fc3aSIlpo Järvinen } 22974a17fc3aSIlpo Järvinen 22981da177e4SLinus Torvalds /* This retransmits one SKB. Policy decisions and retransmit queue 22991da177e4SLinus Torvalds * state updates are done by the caller. Returns non-zero if an 23001da177e4SLinus Torvalds * error occurred which prevented the send. 23011da177e4SLinus Torvalds */ 230293b174adSYuchung Cheng int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) 23031da177e4SLinus Torvalds { 23041da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 23055d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 23067d227cd2SSridhar Samudrala unsigned int cur_mss; 23071da177e4SLinus Torvalds 23085d424d5aSJohn Heffner /* Inconslusive MTU probe */ 23095d424d5aSJohn Heffner if (icsk->icsk_mtup.probe_size) { 23105d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 23115d424d5aSJohn Heffner } 23125d424d5aSJohn Heffner 23131da177e4SLinus Torvalds /* Do not sent more than we queued. 1/4 is reserved for possible 2314caa20d9aSStephen Hemminger * copying overhead: fragmentation, tunneling, mangling etc. 23151da177e4SLinus Torvalds */ 23161da177e4SLinus Torvalds if (atomic_read(&sk->sk_wmem_alloc) > 23171da177e4SLinus Torvalds min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf)) 23181da177e4SLinus Torvalds return -EAGAIN; 23191da177e4SLinus Torvalds 23201da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 23211da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 23221da177e4SLinus Torvalds BUG(); 23231da177e4SLinus Torvalds if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 23241da177e4SLinus Torvalds return -ENOMEM; 23251da177e4SLinus Torvalds } 23261da177e4SLinus Torvalds 23277d227cd2SSridhar Samudrala if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 23287d227cd2SSridhar Samudrala return -EHOSTUNREACH; /* Routing failure or similar. */ 23297d227cd2SSridhar Samudrala 23300c54b85fSIlpo Järvinen cur_mss = tcp_current_mss(sk); 23317d227cd2SSridhar Samudrala 23321da177e4SLinus Torvalds /* If receiver has shrunk his window, and skb is out of 23331da177e4SLinus Torvalds * new window, do not retransmit it. The exception is the 23341da177e4SLinus Torvalds * case, when window is shrunk to zero. In this case 23351da177e4SLinus Torvalds * our retransmit serves as a zero window probe. 23361da177e4SLinus Torvalds */ 23379d4fb27dSJoe Perches if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && 23389d4fb27dSJoe Perches TCP_SKB_CB(skb)->seq != tp->snd_una) 23391da177e4SLinus Torvalds return -EAGAIN; 23401da177e4SLinus Torvalds 23411da177e4SLinus Torvalds if (skb->len > cur_mss) { 2342846998aeSDavid S. Miller if (tcp_fragment(sk, skb, cur_mss, cur_mss)) 23431da177e4SLinus Torvalds return -ENOMEM; /* We'll try again later. */ 234402276f3cSIlpo Järvinen } else { 23459eb9362eSIlpo Järvinen int oldpcount = tcp_skb_pcount(skb); 23469eb9362eSIlpo Järvinen 23479eb9362eSIlpo Järvinen if (unlikely(oldpcount > 1)) { 2348*c52e2421SEric Dumazet if (skb_unclone(skb, GFP_ATOMIC)) 2349*c52e2421SEric Dumazet return -ENOMEM; 235002276f3cSIlpo Järvinen tcp_init_tso_segs(sk, skb, cur_mss); 23519eb9362eSIlpo Järvinen tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb)); 23529eb9362eSIlpo Järvinen } 23531da177e4SLinus Torvalds } 23541da177e4SLinus Torvalds 23551da177e4SLinus Torvalds tcp_retrans_try_collapse(sk, skb, cur_mss); 23561da177e4SLinus Torvalds 23571da177e4SLinus Torvalds /* Some Solaris stacks overoptimize and ignore the FIN on a 23581da177e4SLinus Torvalds * retransmit when old data is attached. So strip it off 23591da177e4SLinus Torvalds * since it is cheap to do so and saves bytes on the network. 23601da177e4SLinus Torvalds */ 23611da177e4SLinus Torvalds if (skb->len > 0 && 23624de075e0SEric Dumazet (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && 23631da177e4SLinus Torvalds tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) { 23641da177e4SLinus Torvalds if (!pskb_trim(skb, 0)) { 2365e870a8efSIlpo Järvinen /* Reuse, even though it does some unnecessary work */ 2366e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1, 23674de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags); 23681da177e4SLinus Torvalds skb->ip_summed = CHECKSUM_NONE; 23691da177e4SLinus Torvalds } 23701da177e4SLinus Torvalds } 23711da177e4SLinus Torvalds 23721da177e4SLinus Torvalds /* Make a copy, if the first transmission SKB clone we made 23731da177e4SLinus Torvalds * is still in somebody's hands, else make a clone. 23741da177e4SLinus Torvalds */ 23751da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 23761da177e4SLinus Torvalds 237750bceae9SThomas Graf /* make sure skb->data is aligned on arches that require it 237850bceae9SThomas Graf * and check if ack-trimming & collapsing extended the headroom 237950bceae9SThomas Graf * beyond what csum_start can cover. 238050bceae9SThomas Graf */ 238150bceae9SThomas Graf if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) || 238250bceae9SThomas Graf skb_headroom(skb) >= 0xFFFF)) { 2383117632e6SEric Dumazet struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER, 2384117632e6SEric Dumazet GFP_ATOMIC); 238593b174adSYuchung Cheng return nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) : 2386117632e6SEric Dumazet -ENOBUFS; 2387117632e6SEric Dumazet } else { 238893b174adSYuchung Cheng return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 2389117632e6SEric Dumazet } 239093b174adSYuchung Cheng } 239193b174adSYuchung Cheng 239293b174adSYuchung Cheng int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) 239393b174adSYuchung Cheng { 239493b174adSYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 239593b174adSYuchung Cheng int err = __tcp_retransmit_skb(sk, skb); 23961da177e4SLinus Torvalds 23971da177e4SLinus Torvalds if (err == 0) { 23981da177e4SLinus Torvalds /* Update global TCP statistics. */ 239981cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); 24001da177e4SLinus Torvalds 24011da177e4SLinus Torvalds tp->total_retrans++; 24021da177e4SLinus Torvalds 24031da177e4SLinus Torvalds #if FASTRETRANS_DEBUG > 0 24041da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 2405e87cc472SJoe Perches net_dbg_ratelimited("retrans_out leaked\n"); 24061da177e4SLinus Torvalds } 24071da177e4SLinus Torvalds #endif 2408b08d6cb2SIlpo Järvinen if (!tp->retrans_out) 2409b08d6cb2SIlpo Järvinen tp->lost_retrans_low = tp->snd_nxt; 24101da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 24111da177e4SLinus Torvalds tp->retrans_out += tcp_skb_pcount(skb); 24121da177e4SLinus Torvalds 24131da177e4SLinus Torvalds /* Save stamp of the first retransmit. */ 24141da177e4SLinus Torvalds if (!tp->retrans_stamp) 24151da177e4SLinus Torvalds tp->retrans_stamp = TCP_SKB_CB(skb)->when; 24161da177e4SLinus Torvalds 2417c24f691bSYuchung Cheng tp->undo_retrans += tcp_skb_pcount(skb); 24181da177e4SLinus Torvalds 24191da177e4SLinus Torvalds /* snd_nxt is stored to detect loss of retransmitted segment, 24201da177e4SLinus Torvalds * see tcp_input.c tcp_sacktag_write_queue(). 24211da177e4SLinus Torvalds */ 24221da177e4SLinus Torvalds TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt; 242324ab6becSYuchung Cheng } else { 242424ab6becSYuchung Cheng NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL); 24251da177e4SLinus Torvalds } 24261da177e4SLinus Torvalds return err; 24271da177e4SLinus Torvalds } 24281da177e4SLinus Torvalds 242967edfef7SAndi Kleen /* Check if we forward retransmits are possible in the current 243067edfef7SAndi Kleen * window/congestion state. 243167edfef7SAndi Kleen */ 2432a2a385d6SEric Dumazet static bool tcp_can_forward_retransmit(struct sock *sk) 2433b5afe7bcSIlpo Järvinen { 2434b5afe7bcSIlpo Järvinen const struct inet_connection_sock *icsk = inet_csk(sk); 2435cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 2436b5afe7bcSIlpo Järvinen 2437b5afe7bcSIlpo Järvinen /* Forward retransmissions are possible only during Recovery. */ 2438b5afe7bcSIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Recovery) 2439a2a385d6SEric Dumazet return false; 2440b5afe7bcSIlpo Järvinen 2441b5afe7bcSIlpo Järvinen /* No forward retransmissions in Reno are possible. */ 2442b5afe7bcSIlpo Järvinen if (tcp_is_reno(tp)) 2443a2a385d6SEric Dumazet return false; 2444b5afe7bcSIlpo Järvinen 2445b5afe7bcSIlpo Järvinen /* Yeah, we have to make difficult choice between forward transmission 2446b5afe7bcSIlpo Järvinen * and retransmission... Both ways have their merits... 2447b5afe7bcSIlpo Järvinen * 2448b5afe7bcSIlpo Järvinen * For now we do not retransmit anything, while we have some new 2449b5afe7bcSIlpo Järvinen * segments to send. In the other cases, follow rule 3 for 2450b5afe7bcSIlpo Järvinen * NextSeg() specified in RFC3517. 2451b5afe7bcSIlpo Järvinen */ 2452b5afe7bcSIlpo Järvinen 2453b5afe7bcSIlpo Järvinen if (tcp_may_send_now(sk)) 2454a2a385d6SEric Dumazet return false; 2455b5afe7bcSIlpo Järvinen 2456a2a385d6SEric Dumazet return true; 2457b5afe7bcSIlpo Järvinen } 2458b5afe7bcSIlpo Järvinen 24591da177e4SLinus Torvalds /* This gets called after a retransmit timeout, and the initially 24601da177e4SLinus Torvalds * retransmitted data is acknowledged. It tries to continue 24611da177e4SLinus Torvalds * resending the rest of the retransmit queue, until either 24621da177e4SLinus Torvalds * we've sent it all or the congestion window limit is reached. 24631da177e4SLinus Torvalds * If doing SACK, the first ACK which comes back for a timeout 24641da177e4SLinus Torvalds * based retransmit packet might feed us FACK information again. 24651da177e4SLinus Torvalds * If so, we use it to avoid unnecessarily retransmissions. 24661da177e4SLinus Torvalds */ 24671da177e4SLinus Torvalds void tcp_xmit_retransmit_queue(struct sock *sk) 24681da177e4SLinus Torvalds { 24696687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 24701da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 24711da177e4SLinus Torvalds struct sk_buff *skb; 24720e1c54c2SIlpo Järvinen struct sk_buff *hole = NULL; 2473618d9f25SIlpo Järvinen u32 last_lost; 247461eb55f4SIlpo Järvinen int mib_idx; 24750e1c54c2SIlpo Järvinen int fwd_rexmitting = 0; 24766a438bbeSStephen Hemminger 247745e77d31SIlpo Järvinen if (!tp->packets_out) 247845e77d31SIlpo Järvinen return; 247945e77d31SIlpo Järvinen 248008ebd172SIlpo Järvinen if (!tp->lost_out) 248108ebd172SIlpo Järvinen tp->retransmit_high = tp->snd_una; 248208ebd172SIlpo Järvinen 2483618d9f25SIlpo Järvinen if (tp->retransmit_skb_hint) { 24846a438bbeSStephen Hemminger skb = tp->retransmit_skb_hint; 2485618d9f25SIlpo Järvinen last_lost = TCP_SKB_CB(skb)->end_seq; 2486618d9f25SIlpo Järvinen if (after(last_lost, tp->retransmit_high)) 2487618d9f25SIlpo Järvinen last_lost = tp->retransmit_high; 2488618d9f25SIlpo Järvinen } else { 2489fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 2490618d9f25SIlpo Järvinen last_lost = tp->snd_una; 2491618d9f25SIlpo Järvinen } 24921da177e4SLinus Torvalds 2493fe067e8aSDavid S. Miller tcp_for_write_queue_from(skb, sk) { 24941da177e4SLinus Torvalds __u8 sacked = TCP_SKB_CB(skb)->sacked; 24951da177e4SLinus Torvalds 2496fe067e8aSDavid S. Miller if (skb == tcp_send_head(sk)) 2497fe067e8aSDavid S. Miller break; 24986a438bbeSStephen Hemminger /* we could do better than to assign each time */ 24990e1c54c2SIlpo Järvinen if (hole == NULL) 25006a438bbeSStephen Hemminger tp->retransmit_skb_hint = skb; 25016a438bbeSStephen Hemminger 25021da177e4SLinus Torvalds /* Assume this retransmit will generate 25031da177e4SLinus Torvalds * only one packet for congestion window 25041da177e4SLinus Torvalds * calculation purposes. This works because 25051da177e4SLinus Torvalds * tcp_retransmit_skb() will chop up the 25061da177e4SLinus Torvalds * packet to be MSS sized and all the 25071da177e4SLinus Torvalds * packet counting works out. 25081da177e4SLinus Torvalds */ 25091da177e4SLinus Torvalds if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) 25101da177e4SLinus Torvalds return; 25110e1c54c2SIlpo Järvinen 25120e1c54c2SIlpo Järvinen if (fwd_rexmitting) { 25130e1c54c2SIlpo Järvinen begin_fwd: 25140e1c54c2SIlpo Järvinen if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp))) 2515006f582cSIlpo Järvinen break; 25160e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFORWARDRETRANS; 25170e1c54c2SIlpo Järvinen 25180e1c54c2SIlpo Järvinen } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) { 2519618d9f25SIlpo Järvinen tp->retransmit_high = last_lost; 25200e1c54c2SIlpo Järvinen if (!tcp_can_forward_retransmit(sk)) 25210e1c54c2SIlpo Järvinen break; 25220e1c54c2SIlpo Järvinen /* Backtrack if necessary to non-L'ed skb */ 25230e1c54c2SIlpo Järvinen if (hole != NULL) { 25240e1c54c2SIlpo Järvinen skb = hole; 25250e1c54c2SIlpo Järvinen hole = NULL; 25260e1c54c2SIlpo Järvinen } 25270e1c54c2SIlpo Järvinen fwd_rexmitting = 1; 25280e1c54c2SIlpo Järvinen goto begin_fwd; 25290e1c54c2SIlpo Järvinen 25300e1c54c2SIlpo Järvinen } else if (!(sacked & TCPCB_LOST)) { 2531ac11ba75SIlpo Järvinen if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) 25320e1c54c2SIlpo Järvinen hole = skb; 253361eb55f4SIlpo Järvinen continue; 25341da177e4SLinus Torvalds 25350e1c54c2SIlpo Järvinen } else { 2536618d9f25SIlpo Järvinen last_lost = TCP_SKB_CB(skb)->end_seq; 25370e1c54c2SIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Loss) 25380e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFASTRETRANS; 25390e1c54c2SIlpo Järvinen else 25400e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; 25410e1c54c2SIlpo Järvinen } 25420e1c54c2SIlpo Järvinen 25430e1c54c2SIlpo Järvinen if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) 254461eb55f4SIlpo Järvinen continue; 254540b215e5SPavel Emelyanov 254624ab6becSYuchung Cheng if (tcp_retransmit_skb(sk, skb)) 25471da177e4SLinus Torvalds return; 254824ab6becSYuchung Cheng 2549de0744afSPavel Emelyanov NET_INC_STATS_BH(sock_net(sk), mib_idx); 25501da177e4SLinus Torvalds 2551684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 2552a262f0cdSNandita Dukkipati tp->prr_out += tcp_skb_pcount(skb); 2553a262f0cdSNandita Dukkipati 2554fe067e8aSDavid S. Miller if (skb == tcp_write_queue_head(sk)) 2555463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 25563f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, 25573f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 25581da177e4SLinus Torvalds } 25591da177e4SLinus Torvalds } 25601da177e4SLinus Torvalds 25611da177e4SLinus Torvalds /* Send a fin. The caller locks the socket for us. This cannot be 25621da177e4SLinus Torvalds * allowed to fail queueing a FIN frame under any circumstances. 25631da177e4SLinus Torvalds */ 25641da177e4SLinus Torvalds void tcp_send_fin(struct sock *sk) 25651da177e4SLinus Torvalds { 25661da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2567fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_write_queue_tail(sk); 25681da177e4SLinus Torvalds int mss_now; 25691da177e4SLinus Torvalds 25701da177e4SLinus Torvalds /* Optimization, tack on the FIN if we have a queue of 25711da177e4SLinus Torvalds * unsent frames. But be careful about outgoing SACKS 25721da177e4SLinus Torvalds * and IP options. 25731da177e4SLinus Torvalds */ 25740c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 25751da177e4SLinus Torvalds 2576fe067e8aSDavid S. Miller if (tcp_send_head(sk) != NULL) { 25774de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_FIN; 25781da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq++; 25791da177e4SLinus Torvalds tp->write_seq++; 25801da177e4SLinus Torvalds } else { 25811da177e4SLinus Torvalds /* Socket is locked, keep trying until memory is available. */ 25821da177e4SLinus Torvalds for (;;) { 2583aa133076SWu Fengguang skb = alloc_skb_fclone(MAX_TCP_HEADER, 2584aa133076SWu Fengguang sk->sk_allocation); 25851da177e4SLinus Torvalds if (skb) 25861da177e4SLinus Torvalds break; 25871da177e4SLinus Torvalds yield(); 25881da177e4SLinus Torvalds } 25891da177e4SLinus Torvalds 25901da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 25911da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 25921da177e4SLinus Torvalds /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 2593e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tp->write_seq, 2594a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_FIN); 25951da177e4SLinus Torvalds tcp_queue_skb(sk, skb); 25961da177e4SLinus Torvalds } 25979e412ba7SIlpo Järvinen __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF); 25981da177e4SLinus Torvalds } 25991da177e4SLinus Torvalds 26001da177e4SLinus Torvalds /* We get here when a process closes a file descriptor (either due to 26011da177e4SLinus Torvalds * an explicit close() or as a byproduct of exit()'ing) and there 26021da177e4SLinus Torvalds * was unread data in the receive queue. This behavior is recommended 260365bb723cSGerrit Renker * by RFC 2525, section 2.17. -DaveM 26041da177e4SLinus Torvalds */ 2605dd0fc66fSAl Viro void tcp_send_active_reset(struct sock *sk, gfp_t priority) 26061da177e4SLinus Torvalds { 26071da177e4SLinus Torvalds struct sk_buff *skb; 26081da177e4SLinus Torvalds 26091da177e4SLinus Torvalds /* NOTE: No TCP options attached and we never retransmit this. */ 26101da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, priority); 26111da177e4SLinus Torvalds if (!skb) { 26124e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 26131da177e4SLinus Torvalds return; 26141da177e4SLinus Torvalds } 26151da177e4SLinus Torvalds 26161da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 26171da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 2618e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), 2619a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_RST); 26201da177e4SLinus Torvalds /* Send it off. */ 26211da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2622dfb4b9dcSDavid S. Miller if (tcp_transmit_skb(sk, skb, 0, priority)) 26234e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 262426af65cbSSridhar Samudrala 262581cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); 26261da177e4SLinus Torvalds } 26271da177e4SLinus Torvalds 262867edfef7SAndi Kleen /* Send a crossed SYN-ACK during socket establishment. 262967edfef7SAndi Kleen * WARNING: This routine must only be called when we have already sent 26301da177e4SLinus Torvalds * a SYN packet that crossed the incoming SYN that caused this routine 26311da177e4SLinus Torvalds * to get called. If this assumption fails then the initial rcv_wnd 26321da177e4SLinus Torvalds * and rcv_wscale values will not be correct. 26331da177e4SLinus Torvalds */ 26341da177e4SLinus Torvalds int tcp_send_synack(struct sock *sk) 26351da177e4SLinus Torvalds { 26361da177e4SLinus Torvalds struct sk_buff *skb; 26371da177e4SLinus Torvalds 2638fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 26394de075e0SEric Dumazet if (skb == NULL || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 264091df42beSJoe Perches pr_debug("%s: wrong queue state\n", __func__); 26411da177e4SLinus Torvalds return -EFAULT; 26421da177e4SLinus Torvalds } 26434de075e0SEric Dumazet if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { 26441da177e4SLinus Torvalds if (skb_cloned(skb)) { 26451da177e4SLinus Torvalds struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC); 26461da177e4SLinus Torvalds if (nskb == NULL) 26471da177e4SLinus Torvalds return -ENOMEM; 2648fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 26491da177e4SLinus Torvalds skb_header_release(nskb); 2650fe067e8aSDavid S. Miller __tcp_add_write_queue_head(sk, nskb); 26513ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 26523ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 26533ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 26541da177e4SLinus Torvalds skb = nskb; 26551da177e4SLinus Torvalds } 26561da177e4SLinus Torvalds 26574de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; 26581da177e4SLinus Torvalds TCP_ECN_send_synack(tcp_sk(sk), skb); 26591da177e4SLinus Torvalds } 26601da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2661dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 26621da177e4SLinus Torvalds } 26631da177e4SLinus Torvalds 26644aea39c1SEric Dumazet /** 26654aea39c1SEric Dumazet * tcp_make_synack - Prepare a SYN-ACK. 26664aea39c1SEric Dumazet * sk: listener socket 26674aea39c1SEric Dumazet * dst: dst entry attached to the SYNACK 26684aea39c1SEric Dumazet * req: request_sock pointer 26694aea39c1SEric Dumazet * 26704aea39c1SEric Dumazet * Allocate one skb and build a SYNACK packet. 26714aea39c1SEric Dumazet * @dst is consumed : Caller should not use it again. 26724aea39c1SEric Dumazet */ 26731da177e4SLinus Torvalds struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst, 2674e6b4d113SWilliam Allen Simpson struct request_sock *req, 26758336886fSJerry Chu struct tcp_fastopen_cookie *foc) 26761da177e4SLinus Torvalds { 2677bd0388aeSWilliam Allen Simpson struct tcp_out_options opts; 26782e6599cbSArnaldo Carvalho de Melo struct inet_request_sock *ireq = inet_rsk(req); 26791da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 26801da177e4SLinus Torvalds struct tcphdr *th; 26811da177e4SLinus Torvalds struct sk_buff *skb; 2682cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 2683bd0388aeSWilliam Allen Simpson int tcp_header_size; 2684f5fff5dcSTom Quetchenbach int mss; 26851da177e4SLinus Torvalds 2686eb8895deSPhil Oester skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC); 26874aea39c1SEric Dumazet if (unlikely(!skb)) { 26884aea39c1SEric Dumazet dst_release(dst); 26891da177e4SLinus Torvalds return NULL; 26904aea39c1SEric Dumazet } 26911da177e4SLinus Torvalds /* Reserve space for headers. */ 26921da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 26931da177e4SLinus Torvalds 26944aea39c1SEric Dumazet skb_dst_set(skb, dst); 2695ca10b9e9SEric Dumazet security_skb_owned_by(skb, sk); 26961da177e4SLinus Torvalds 26970dbaee3bSDavid S. Miller mss = dst_metric_advmss(dst); 2698f5fff5dcSTom Quetchenbach if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss) 2699f5fff5dcSTom Quetchenbach mss = tp->rx_opt.user_mss; 2700f5fff5dcSTom Quetchenbach 270133ad798cSAdam Langley if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */ 270233ad798cSAdam Langley __u8 rcv_wscale; 270333ad798cSAdam Langley /* Set this up on the first call only */ 270433ad798cSAdam Langley req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW); 2705e88c64f0SHagen Paul Pfeifer 2706e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 2707e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 2708e88c64f0SHagen Paul Pfeifer (req->window_clamp > tcp_full_space(sk) || req->window_clamp == 0)) 2709e88c64f0SHagen Paul Pfeifer req->window_clamp = tcp_full_space(sk); 2710e88c64f0SHagen Paul Pfeifer 271133ad798cSAdam Langley /* tcp_full_space because it is guaranteed to be the first packet */ 271233ad798cSAdam Langley tcp_select_initial_window(tcp_full_space(sk), 2713f5fff5dcSTom Quetchenbach mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0), 271433ad798cSAdam Langley &req->rcv_wnd, 271533ad798cSAdam Langley &req->window_clamp, 271633ad798cSAdam Langley ireq->wscale_ok, 271731d12926Slaurent chavey &rcv_wscale, 271831d12926Slaurent chavey dst_metric(dst, RTAX_INITRWND)); 271933ad798cSAdam Langley ireq->rcv_wscale = rcv_wscale; 272033ad798cSAdam Langley } 2721cfb6eeb4SYOSHIFUJI Hideaki 272233ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 27238b5f12d0SFlorian Westphal #ifdef CONFIG_SYN_COOKIES 27248b5f12d0SFlorian Westphal if (unlikely(req->cookie_ts)) 27258b5f12d0SFlorian Westphal TCP_SKB_CB(skb)->when = cookie_init_timestamp(req); 27268b5f12d0SFlorian Westphal else 27278b5f12d0SFlorian Westphal #endif 272833ad798cSAdam Langley TCP_SKB_CB(skb)->when = tcp_time_stamp; 27291a2c6181SChristoph Paasch tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, &md5, 27301a2c6181SChristoph Paasch foc) + sizeof(*th); 273133ad798cSAdam Langley 2732aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 2733aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 27341da177e4SLinus Torvalds 2735aa8223c7SArnaldo Carvalho de Melo th = tcp_hdr(skb); 27361da177e4SLinus Torvalds memset(th, 0, sizeof(struct tcphdr)); 27371da177e4SLinus Torvalds th->syn = 1; 27381da177e4SLinus Torvalds th->ack = 1; 27391da177e4SLinus Torvalds TCP_ECN_make_synack(req, th); 2740a3116ac5SKOVACS Krisztian th->source = ireq->loc_port; 27412e6599cbSArnaldo Carvalho de Melo th->dest = ireq->rmt_port; 2742e870a8efSIlpo Järvinen /* Setting of flags are superfluous here for callers (and ECE is 2743e870a8efSIlpo Järvinen * not even correctly set) 2744e870a8efSIlpo Järvinen */ 2745e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn, 2746a3433f35SChangli Gao TCPHDR_SYN | TCPHDR_ACK); 27474957faadSWilliam Allen Simpson 27481da177e4SLinus Torvalds th->seq = htonl(TCP_SKB_CB(skb)->seq); 27498336886fSJerry Chu /* XXX data is queued and acked as is. No buffer/window check */ 27508336886fSJerry Chu th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt); 27511da177e4SLinus Torvalds 27521da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 2753600ff0c2SIlpo Järvinen th->window = htons(min(req->rcv_wnd, 65535U)); 2754bd0388aeSWilliam Allen Simpson tcp_options_write((__be32 *)(th + 1), tp, &opts); 27551da177e4SLinus Torvalds th->doff = (tcp_header_size >> 2); 2756aa2ea058STom Herbert TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, tcp_skb_pcount(skb)); 2757cfb6eeb4SYOSHIFUJI Hideaki 2758cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 2759cfb6eeb4SYOSHIFUJI Hideaki /* Okay, we have all we need - do the md5 hash if needed */ 2760cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 2761bd0388aeSWilliam Allen Simpson tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, 276249a72dfbSAdam Langley md5, NULL, req, skb); 2763cfb6eeb4SYOSHIFUJI Hideaki } 2764cfb6eeb4SYOSHIFUJI Hideaki #endif 2765cfb6eeb4SYOSHIFUJI Hideaki 27661da177e4SLinus Torvalds return skb; 27671da177e4SLinus Torvalds } 27684bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_make_synack); 27691da177e4SLinus Torvalds 277067edfef7SAndi Kleen /* Do all connect socket setups that can be done AF independent. */ 2771370816aeSPavel Emelyanov void tcp_connect_init(struct sock *sk) 27721da177e4SLinus Torvalds { 2773cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 27741da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 27751da177e4SLinus Torvalds __u8 rcv_wscale; 27761da177e4SLinus Torvalds 27771da177e4SLinus Torvalds /* We'll fix this up when we get a response from the other end. 27781da177e4SLinus Torvalds * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 27791da177e4SLinus Torvalds */ 27801da177e4SLinus Torvalds tp->tcp_header_len = sizeof(struct tcphdr) + 2781bb5b7c11SDavid S. Miller (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0); 27821da177e4SLinus Torvalds 2783cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 2784cfb6eeb4SYOSHIFUJI Hideaki if (tp->af_specific->md5_lookup(sk, sk) != NULL) 2785cfb6eeb4SYOSHIFUJI Hideaki tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 2786cfb6eeb4SYOSHIFUJI Hideaki #endif 2787cfb6eeb4SYOSHIFUJI Hideaki 27881da177e4SLinus Torvalds /* If user gave his TCP_MAXSEG, record it to clamp */ 27891da177e4SLinus Torvalds if (tp->rx_opt.user_mss) 27901da177e4SLinus Torvalds tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 27911da177e4SLinus Torvalds tp->max_window = 0; 27925d424d5aSJohn Heffner tcp_mtup_init(sk); 27931da177e4SLinus Torvalds tcp_sync_mss(sk, dst_mtu(dst)); 27941da177e4SLinus Torvalds 27951da177e4SLinus Torvalds if (!tp->window_clamp) 27961da177e4SLinus Torvalds tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 27970dbaee3bSDavid S. Miller tp->advmss = dst_metric_advmss(dst); 2798f5fff5dcSTom Quetchenbach if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss) 2799f5fff5dcSTom Quetchenbach tp->advmss = tp->rx_opt.user_mss; 2800f5fff5dcSTom Quetchenbach 28011da177e4SLinus Torvalds tcp_initialize_rcv_mss(sk); 28021da177e4SLinus Torvalds 2803e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 2804e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 2805e88c64f0SHagen Paul Pfeifer (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) 2806e88c64f0SHagen Paul Pfeifer tp->window_clamp = tcp_full_space(sk); 2807e88c64f0SHagen Paul Pfeifer 28081da177e4SLinus Torvalds tcp_select_initial_window(tcp_full_space(sk), 28091da177e4SLinus Torvalds tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 28101da177e4SLinus Torvalds &tp->rcv_wnd, 28111da177e4SLinus Torvalds &tp->window_clamp, 2812bb5b7c11SDavid S. Miller sysctl_tcp_window_scaling, 281331d12926Slaurent chavey &rcv_wscale, 281431d12926Slaurent chavey dst_metric(dst, RTAX_INITRWND)); 28151da177e4SLinus Torvalds 28161da177e4SLinus Torvalds tp->rx_opt.rcv_wscale = rcv_wscale; 28171da177e4SLinus Torvalds tp->rcv_ssthresh = tp->rcv_wnd; 28181da177e4SLinus Torvalds 28191da177e4SLinus Torvalds sk->sk_err = 0; 28201da177e4SLinus Torvalds sock_reset_flag(sk, SOCK_DONE); 28211da177e4SLinus Torvalds tp->snd_wnd = 0; 2822ee7537b6SHantzis Fotis tcp_init_wl(tp, 0); 28231da177e4SLinus Torvalds tp->snd_una = tp->write_seq; 28241da177e4SLinus Torvalds tp->snd_sml = tp->write_seq; 282533f5f57eSIlpo Järvinen tp->snd_up = tp->write_seq; 2826370816aeSPavel Emelyanov tp->snd_nxt = tp->write_seq; 2827ee995283SPavel Emelyanov 2828ee995283SPavel Emelyanov if (likely(!tp->repair)) 28291da177e4SLinus Torvalds tp->rcv_nxt = 0; 2830c7781a6eSAndrew Vagin else 2831c7781a6eSAndrew Vagin tp->rcv_tstamp = tcp_time_stamp; 2832ee995283SPavel Emelyanov tp->rcv_wup = tp->rcv_nxt; 2833ee995283SPavel Emelyanov tp->copied_seq = tp->rcv_nxt; 28341da177e4SLinus Torvalds 2835463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; 2836463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_retransmits = 0; 28371da177e4SLinus Torvalds tcp_clear_retrans(tp); 28381da177e4SLinus Torvalds } 28391da177e4SLinus Torvalds 2840783237e8SYuchung Cheng static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) 2841783237e8SYuchung Cheng { 2842783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 2843783237e8SYuchung Cheng struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 2844783237e8SYuchung Cheng 2845783237e8SYuchung Cheng tcb->end_seq += skb->len; 2846783237e8SYuchung Cheng skb_header_release(skb); 2847783237e8SYuchung Cheng __tcp_add_write_queue_tail(sk, skb); 2848783237e8SYuchung Cheng sk->sk_wmem_queued += skb->truesize; 2849783237e8SYuchung Cheng sk_mem_charge(sk, skb->truesize); 2850783237e8SYuchung Cheng tp->write_seq = tcb->end_seq; 2851783237e8SYuchung Cheng tp->packets_out += tcp_skb_pcount(skb); 2852783237e8SYuchung Cheng } 2853783237e8SYuchung Cheng 2854783237e8SYuchung Cheng /* Build and send a SYN with data and (cached) Fast Open cookie. However, 2855783237e8SYuchung Cheng * queue a data-only packet after the regular SYN, such that regular SYNs 2856783237e8SYuchung Cheng * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges 2857783237e8SYuchung Cheng * only the SYN sequence, the data are retransmitted in the first ACK. 2858783237e8SYuchung Cheng * If cookie is not cached or other error occurs, falls back to send a 2859783237e8SYuchung Cheng * regular SYN with Fast Open cookie request option. 2860783237e8SYuchung Cheng */ 2861783237e8SYuchung Cheng static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) 2862783237e8SYuchung Cheng { 2863783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 2864783237e8SYuchung Cheng struct tcp_fastopen_request *fo = tp->fastopen_req; 2865aab48743SYuchung Cheng int syn_loss = 0, space, i, err = 0, iovlen = fo->data->msg_iovlen; 2866783237e8SYuchung Cheng struct sk_buff *syn_data = NULL, *data; 2867aab48743SYuchung Cheng unsigned long last_syn_loss = 0; 2868783237e8SYuchung Cheng 286967da22d2SYuchung Cheng tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ 2870aab48743SYuchung Cheng tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie, 2871aab48743SYuchung Cheng &syn_loss, &last_syn_loss); 2872aab48743SYuchung Cheng /* Recurring FO SYN losses: revert to regular handshake temporarily */ 2873aab48743SYuchung Cheng if (syn_loss > 1 && 2874aab48743SYuchung Cheng time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) { 2875aab48743SYuchung Cheng fo->cookie.len = -1; 2876aab48743SYuchung Cheng goto fallback; 2877aab48743SYuchung Cheng } 2878aab48743SYuchung Cheng 287967da22d2SYuchung Cheng if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE) 288067da22d2SYuchung Cheng fo->cookie.len = -1; 288167da22d2SYuchung Cheng else if (fo->cookie.len <= 0) 2882783237e8SYuchung Cheng goto fallback; 2883783237e8SYuchung Cheng 2884783237e8SYuchung Cheng /* MSS for SYN-data is based on cached MSS and bounded by PMTU and 2885783237e8SYuchung Cheng * user-MSS. Reserve maximum option space for middleboxes that add 2886783237e8SYuchung Cheng * private TCP options. The cost is reduced data space in SYN :( 2887783237e8SYuchung Cheng */ 2888783237e8SYuchung Cheng if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp) 2889783237e8SYuchung Cheng tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 28901b63edd6SYuchung Cheng space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) - 2891783237e8SYuchung Cheng MAX_TCP_OPTION_SPACE; 2892783237e8SYuchung Cheng 2893783237e8SYuchung Cheng syn_data = skb_copy_expand(syn, skb_headroom(syn), space, 2894783237e8SYuchung Cheng sk->sk_allocation); 2895783237e8SYuchung Cheng if (syn_data == NULL) 2896783237e8SYuchung Cheng goto fallback; 2897783237e8SYuchung Cheng 2898783237e8SYuchung Cheng for (i = 0; i < iovlen && syn_data->len < space; ++i) { 2899783237e8SYuchung Cheng struct iovec *iov = &fo->data->msg_iov[i]; 2900783237e8SYuchung Cheng unsigned char __user *from = iov->iov_base; 2901783237e8SYuchung Cheng int len = iov->iov_len; 2902783237e8SYuchung Cheng 2903783237e8SYuchung Cheng if (syn_data->len + len > space) 2904783237e8SYuchung Cheng len = space - syn_data->len; 2905783237e8SYuchung Cheng else if (i + 1 == iovlen) 2906783237e8SYuchung Cheng /* No more data pending in inet_wait_for_connect() */ 2907783237e8SYuchung Cheng fo->data = NULL; 2908783237e8SYuchung Cheng 2909783237e8SYuchung Cheng if (skb_add_data(syn_data, from, len)) 2910783237e8SYuchung Cheng goto fallback; 2911783237e8SYuchung Cheng } 2912783237e8SYuchung Cheng 2913783237e8SYuchung Cheng /* Queue a data-only packet after the regular SYN for retransmission */ 2914783237e8SYuchung Cheng data = pskb_copy(syn_data, sk->sk_allocation); 2915783237e8SYuchung Cheng if (data == NULL) 2916783237e8SYuchung Cheng goto fallback; 2917783237e8SYuchung Cheng TCP_SKB_CB(data)->seq++; 2918783237e8SYuchung Cheng TCP_SKB_CB(data)->tcp_flags &= ~TCPHDR_SYN; 2919783237e8SYuchung Cheng TCP_SKB_CB(data)->tcp_flags = (TCPHDR_ACK|TCPHDR_PSH); 2920783237e8SYuchung Cheng tcp_connect_queue_skb(sk, data); 2921783237e8SYuchung Cheng fo->copied = data->len; 2922783237e8SYuchung Cheng 2923783237e8SYuchung Cheng if (tcp_transmit_skb(sk, syn_data, 0, sk->sk_allocation) == 0) { 292467da22d2SYuchung Cheng tp->syn_data = (fo->copied > 0); 2925783237e8SYuchung Cheng NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE); 2926783237e8SYuchung Cheng goto done; 2927783237e8SYuchung Cheng } 2928783237e8SYuchung Cheng syn_data = NULL; 2929783237e8SYuchung Cheng 2930783237e8SYuchung Cheng fallback: 2931783237e8SYuchung Cheng /* Send a regular SYN with Fast Open cookie request option */ 2932783237e8SYuchung Cheng if (fo->cookie.len > 0) 2933783237e8SYuchung Cheng fo->cookie.len = 0; 2934783237e8SYuchung Cheng err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); 2935783237e8SYuchung Cheng if (err) 2936783237e8SYuchung Cheng tp->syn_fastopen = 0; 2937783237e8SYuchung Cheng kfree_skb(syn_data); 2938783237e8SYuchung Cheng done: 2939783237e8SYuchung Cheng fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ 2940783237e8SYuchung Cheng return err; 2941783237e8SYuchung Cheng } 2942783237e8SYuchung Cheng 294367edfef7SAndi Kleen /* Build a SYN and send it off. */ 29441da177e4SLinus Torvalds int tcp_connect(struct sock *sk) 29451da177e4SLinus Torvalds { 29461da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 29471da177e4SLinus Torvalds struct sk_buff *buff; 2948ee586811SEric Paris int err; 29491da177e4SLinus Torvalds 29501da177e4SLinus Torvalds tcp_connect_init(sk); 29511da177e4SLinus Torvalds 29522b916477SAndrey Vagin if (unlikely(tp->repair)) { 29532b916477SAndrey Vagin tcp_finish_connect(sk, NULL); 29542b916477SAndrey Vagin return 0; 29552b916477SAndrey Vagin } 29562b916477SAndrey Vagin 2957d179cd12SDavid S. Miller buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation); 29581da177e4SLinus Torvalds if (unlikely(buff == NULL)) 29591da177e4SLinus Torvalds return -ENOBUFS; 29601da177e4SLinus Torvalds 29611da177e4SLinus Torvalds /* Reserve space for headers. */ 29621da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 29631da177e4SLinus Torvalds 2964a3433f35SChangli Gao tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); 2965783237e8SYuchung Cheng tp->retrans_stamp = TCP_SKB_CB(buff)->when = tcp_time_stamp; 2966783237e8SYuchung Cheng tcp_connect_queue_skb(sk, buff); 2967e870a8efSIlpo Järvinen TCP_ECN_send_syn(sk, buff); 29681da177e4SLinus Torvalds 2969783237e8SYuchung Cheng /* Send off SYN; include data in Fast Open. */ 2970783237e8SYuchung Cheng err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : 2971783237e8SYuchung Cheng tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); 2972ee586811SEric Paris if (err == -ECONNREFUSED) 2973ee586811SEric Paris return err; 2974bd37a088SWei Yongjun 2975bd37a088SWei Yongjun /* We change tp->snd_nxt after the tcp_transmit_skb() call 2976bd37a088SWei Yongjun * in order to make this packet get counted in tcpOutSegs. 2977bd37a088SWei Yongjun */ 2978bd37a088SWei Yongjun tp->snd_nxt = tp->write_seq; 2979bd37a088SWei Yongjun tp->pushed_seq = tp->write_seq; 298081cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); 29811da177e4SLinus Torvalds 29821da177e4SLinus Torvalds /* Timer for repeating the SYN until an answer. */ 29833f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 29843f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 29851da177e4SLinus Torvalds return 0; 29861da177e4SLinus Torvalds } 29874bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_connect); 29881da177e4SLinus Torvalds 29891da177e4SLinus Torvalds /* Send out a delayed ack, the caller does the policy checking 29901da177e4SLinus Torvalds * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 29911da177e4SLinus Torvalds * for details. 29921da177e4SLinus Torvalds */ 29931da177e4SLinus Torvalds void tcp_send_delayed_ack(struct sock *sk) 29941da177e4SLinus Torvalds { 2995463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 2996463c84b9SArnaldo Carvalho de Melo int ato = icsk->icsk_ack.ato; 29971da177e4SLinus Torvalds unsigned long timeout; 29981da177e4SLinus Torvalds 29991da177e4SLinus Torvalds if (ato > TCP_DELACK_MIN) { 3000463c84b9SArnaldo Carvalho de Melo const struct tcp_sock *tp = tcp_sk(sk); 30011da177e4SLinus Torvalds int max_ato = HZ / 2; 30021da177e4SLinus Torvalds 3003056834d9SIlpo Järvinen if (icsk->icsk_ack.pingpong || 3004056834d9SIlpo Järvinen (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 30051da177e4SLinus Torvalds max_ato = TCP_DELACK_MAX; 30061da177e4SLinus Torvalds 30071da177e4SLinus Torvalds /* Slow path, intersegment interval is "high". */ 30081da177e4SLinus Torvalds 30091da177e4SLinus Torvalds /* If some rtt estimate is known, use it to bound delayed ack. 3010463c84b9SArnaldo Carvalho de Melo * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 30111da177e4SLinus Torvalds * directly. 30121da177e4SLinus Torvalds */ 30131da177e4SLinus Torvalds if (tp->srtt) { 30141da177e4SLinus Torvalds int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN); 30151da177e4SLinus Torvalds 30161da177e4SLinus Torvalds if (rtt < max_ato) 30171da177e4SLinus Torvalds max_ato = rtt; 30181da177e4SLinus Torvalds } 30191da177e4SLinus Torvalds 30201da177e4SLinus Torvalds ato = min(ato, max_ato); 30211da177e4SLinus Torvalds } 30221da177e4SLinus Torvalds 30231da177e4SLinus Torvalds /* Stay within the limit we were given */ 30241da177e4SLinus Torvalds timeout = jiffies + ato; 30251da177e4SLinus Torvalds 30261da177e4SLinus Torvalds /* Use new timeout only if there wasn't a older one earlier. */ 3027463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 30281da177e4SLinus Torvalds /* If delack timer was blocked or is about to expire, 30291da177e4SLinus Torvalds * send ACK now. 30301da177e4SLinus Torvalds */ 3031463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.blocked || 3032463c84b9SArnaldo Carvalho de Melo time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 30331da177e4SLinus Torvalds tcp_send_ack(sk); 30341da177e4SLinus Torvalds return; 30351da177e4SLinus Torvalds } 30361da177e4SLinus Torvalds 3037463c84b9SArnaldo Carvalho de Melo if (!time_before(timeout, icsk->icsk_ack.timeout)) 3038463c84b9SArnaldo Carvalho de Melo timeout = icsk->icsk_ack.timeout; 30391da177e4SLinus Torvalds } 3040463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 3041463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.timeout = timeout; 3042463c84b9SArnaldo Carvalho de Melo sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 30431da177e4SLinus Torvalds } 30441da177e4SLinus Torvalds 30451da177e4SLinus Torvalds /* This routine sends an ack and also updates the window. */ 30461da177e4SLinus Torvalds void tcp_send_ack(struct sock *sk) 30471da177e4SLinus Torvalds { 30481da177e4SLinus Torvalds struct sk_buff *buff; 30491da177e4SLinus Torvalds 3050058dc334SIlpo Järvinen /* If we have been reset, we may not send again. */ 3051058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3052058dc334SIlpo Järvinen return; 3053058dc334SIlpo Järvinen 30541da177e4SLinus Torvalds /* We are not putting this on the write queue, so 30551da177e4SLinus Torvalds * tcp_transmit_skb() will set the ownership to this 30561da177e4SLinus Torvalds * sock. 30571da177e4SLinus Torvalds */ 305899a1dec7SMel Gorman buff = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC)); 30591da177e4SLinus Torvalds if (buff == NULL) { 3060463c84b9SArnaldo Carvalho de Melo inet_csk_schedule_ack(sk); 3061463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; 30623f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 30633f421baaSArnaldo Carvalho de Melo TCP_DELACK_MAX, TCP_RTO_MAX); 30641da177e4SLinus Torvalds return; 30651da177e4SLinus Torvalds } 30661da177e4SLinus Torvalds 30671da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 30681da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 3069a3433f35SChangli Gao tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); 30701da177e4SLinus Torvalds 30711da177e4SLinus Torvalds /* Send it off, this clears delayed acks for us. */ 30721da177e4SLinus Torvalds TCP_SKB_CB(buff)->when = tcp_time_stamp; 307399a1dec7SMel Gorman tcp_transmit_skb(sk, buff, 0, sk_gfp_atomic(sk, GFP_ATOMIC)); 30741da177e4SLinus Torvalds } 30751da177e4SLinus Torvalds 30761da177e4SLinus Torvalds /* This routine sends a packet with an out of date sequence 30771da177e4SLinus Torvalds * number. It assumes the other end will try to ack it. 30781da177e4SLinus Torvalds * 30791da177e4SLinus Torvalds * Question: what should we make while urgent mode? 30801da177e4SLinus Torvalds * 4.4BSD forces sending single byte of data. We cannot send 30811da177e4SLinus Torvalds * out of window data, because we have SND.NXT==SND.MAX... 30821da177e4SLinus Torvalds * 30831da177e4SLinus Torvalds * Current solution: to send TWO zero-length segments in urgent mode: 30841da177e4SLinus Torvalds * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 30851da177e4SLinus Torvalds * out-of-date with SND.UNA-1 to probe window. 30861da177e4SLinus Torvalds */ 30871da177e4SLinus Torvalds static int tcp_xmit_probe_skb(struct sock *sk, int urgent) 30881da177e4SLinus Torvalds { 30891da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 30901da177e4SLinus Torvalds struct sk_buff *skb; 30911da177e4SLinus Torvalds 30921da177e4SLinus Torvalds /* We don't queue it, tcp_transmit_skb() sets ownership. */ 309399a1dec7SMel Gorman skb = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC)); 30941da177e4SLinus Torvalds if (skb == NULL) 30951da177e4SLinus Torvalds return -1; 30961da177e4SLinus Torvalds 30971da177e4SLinus Torvalds /* Reserve space for headers and set control bits. */ 30981da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 30991da177e4SLinus Torvalds /* Use a previous sequence. This should cause the other 31001da177e4SLinus Torvalds * end to send an ack. Don't queue or clone SKB, just 31011da177e4SLinus Torvalds * send it. 31021da177e4SLinus Torvalds */ 3103a3433f35SChangli Gao tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); 31041da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 3105dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC); 31061da177e4SLinus Torvalds } 31071da177e4SLinus Torvalds 3108ee995283SPavel Emelyanov void tcp_send_window_probe(struct sock *sk) 3109ee995283SPavel Emelyanov { 3110ee995283SPavel Emelyanov if (sk->sk_state == TCP_ESTABLISHED) { 3111ee995283SPavel Emelyanov tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; 3112c0e88ff0SPavel Emelyanov tcp_sk(sk)->snd_nxt = tcp_sk(sk)->write_seq; 3113ee995283SPavel Emelyanov tcp_xmit_probe_skb(sk, 0); 3114ee995283SPavel Emelyanov } 3115ee995283SPavel Emelyanov } 3116ee995283SPavel Emelyanov 311767edfef7SAndi Kleen /* Initiate keepalive or window probe from timer. */ 31181da177e4SLinus Torvalds int tcp_write_wakeup(struct sock *sk) 31191da177e4SLinus Torvalds { 31201da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 31211da177e4SLinus Torvalds struct sk_buff *skb; 31221da177e4SLinus Torvalds 3123058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3124058dc334SIlpo Järvinen return -1; 3125058dc334SIlpo Järvinen 3126fe067e8aSDavid S. Miller if ((skb = tcp_send_head(sk)) != NULL && 312790840defSIlpo Järvinen before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { 31281da177e4SLinus Torvalds int err; 31290c54b85fSIlpo Järvinen unsigned int mss = tcp_current_mss(sk); 313090840defSIlpo Järvinen unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 31311da177e4SLinus Torvalds 31321da177e4SLinus Torvalds if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 31331da177e4SLinus Torvalds tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 31341da177e4SLinus Torvalds 31351da177e4SLinus Torvalds /* We are probing the opening of a window 31361da177e4SLinus Torvalds * but the window size is != 0 31371da177e4SLinus Torvalds * must have been a result SWS avoidance ( sender ) 31381da177e4SLinus Torvalds */ 31391da177e4SLinus Torvalds if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 31401da177e4SLinus Torvalds skb->len > mss) { 31411da177e4SLinus Torvalds seg_size = min(seg_size, mss); 31424de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 3143846998aeSDavid S. Miller if (tcp_fragment(sk, skb, seg_size, mss)) 31441da177e4SLinus Torvalds return -1; 31451da177e4SLinus Torvalds } else if (!tcp_skb_pcount(skb)) 3146846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss); 31471da177e4SLinus Torvalds 31484de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 31491da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 3150dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 315166f5fe62SIlpo Järvinen if (!err) 315266f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 31531da177e4SLinus Torvalds return err; 31541da177e4SLinus Torvalds } else { 315533f5f57eSIlpo Järvinen if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) 31564828e7f4SIlpo Järvinen tcp_xmit_probe_skb(sk, 1); 31571da177e4SLinus Torvalds return tcp_xmit_probe_skb(sk, 0); 31581da177e4SLinus Torvalds } 31591da177e4SLinus Torvalds } 31601da177e4SLinus Torvalds 31611da177e4SLinus Torvalds /* A window probe timeout has occurred. If window is not closed send 31621da177e4SLinus Torvalds * a partial packet else a zero probe. 31631da177e4SLinus Torvalds */ 31641da177e4SLinus Torvalds void tcp_send_probe0(struct sock *sk) 31651da177e4SLinus Torvalds { 3166463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 31671da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 31681da177e4SLinus Torvalds int err; 31691da177e4SLinus Torvalds 31701da177e4SLinus Torvalds err = tcp_write_wakeup(sk); 31711da177e4SLinus Torvalds 3172fe067e8aSDavid S. Miller if (tp->packets_out || !tcp_send_head(sk)) { 31731da177e4SLinus Torvalds /* Cancel probe timer, if it is not required. */ 31746687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 0; 3175463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff = 0; 31761da177e4SLinus Torvalds return; 31771da177e4SLinus Torvalds } 31781da177e4SLinus Torvalds 31791da177e4SLinus Torvalds if (err <= 0) { 3180463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_backoff < sysctl_tcp_retries2) 3181463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff++; 31826687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out++; 3183463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 31843f421baaSArnaldo Carvalho de Melo min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), 31853f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 31861da177e4SLinus Torvalds } else { 31871da177e4SLinus Torvalds /* If packet was not sent due to local congestion, 31886687e988SArnaldo Carvalho de Melo * do not backoff and do not remember icsk_probes_out. 31891da177e4SLinus Torvalds * Let local senders to fight for local resources. 31901da177e4SLinus Torvalds * 31911da177e4SLinus Torvalds * Use accumulated backoff yet. 31921da177e4SLinus Torvalds */ 31936687e988SArnaldo Carvalho de Melo if (!icsk->icsk_probes_out) 31946687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 1; 3195463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 3196463c84b9SArnaldo Carvalho de Melo min(icsk->icsk_rto << icsk->icsk_backoff, 31973f421baaSArnaldo Carvalho de Melo TCP_RESOURCE_PROBE_INTERVAL), 31983f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 31991da177e4SLinus Torvalds } 32001da177e4SLinus Torvalds } 3201