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 { 3667b7fc97aSEric Dumazet struct skb_shared_info *shinfo = skb_shinfo(skb); 3677b7fc97aSEric Dumazet 3682e8e18efSDavid S. Miller skb->ip_summed = CHECKSUM_PARTIAL; 369e870a8efSIlpo Järvinen skb->csum = 0; 370e870a8efSIlpo Järvinen 3714de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags; 372e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->sacked = 0; 373e870a8efSIlpo Järvinen 3747b7fc97aSEric Dumazet shinfo->gso_segs = 1; 3757b7fc97aSEric Dumazet shinfo->gso_size = 0; 3767b7fc97aSEric Dumazet shinfo->gso_type = 0; 377e870a8efSIlpo Järvinen 378e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->seq = seq; 379a3433f35SChangli Gao if (flags & (TCPHDR_SYN | TCPHDR_FIN)) 380e870a8efSIlpo Järvinen seq++; 381e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->end_seq = seq; 382e870a8efSIlpo Järvinen } 383e870a8efSIlpo Järvinen 384a2a385d6SEric Dumazet static inline bool tcp_urg_mode(const struct tcp_sock *tp) 38533f5f57eSIlpo Järvinen { 38633f5f57eSIlpo Järvinen return tp->snd_una != tp->snd_up; 38733f5f57eSIlpo Järvinen } 38833f5f57eSIlpo Järvinen 38933ad798cSAdam Langley #define OPTION_SACK_ADVERTISE (1 << 0) 39033ad798cSAdam Langley #define OPTION_TS (1 << 1) 39133ad798cSAdam Langley #define OPTION_MD5 (1 << 2) 39289e95a61SOri Finkelman #define OPTION_WSCALE (1 << 3) 3932100c8d2SYuchung Cheng #define OPTION_FAST_OPEN_COOKIE (1 << 8) 39433ad798cSAdam Langley 39533ad798cSAdam Langley struct tcp_out_options { 3962100c8d2SYuchung Cheng u16 options; /* bit field of OPTION_* */ 3972100c8d2SYuchung Cheng u16 mss; /* 0 to disable */ 39833ad798cSAdam Langley u8 ws; /* window scale, 0 to disable */ 39933ad798cSAdam Langley u8 num_sack_blocks; /* number of SACK blocks to include */ 400bd0388aeSWilliam Allen Simpson u8 hash_size; /* bytes in hash_location */ 401bd0388aeSWilliam Allen Simpson __u8 *hash_location; /* temporary pointer, overloaded */ 4022100c8d2SYuchung Cheng __u32 tsval, tsecr; /* need to include OPTION_TS */ 4032100c8d2SYuchung Cheng struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */ 40433ad798cSAdam Langley }; 40533ad798cSAdam Langley 40667edfef7SAndi Kleen /* Write previously computed TCP options to the packet. 40767edfef7SAndi Kleen * 40867edfef7SAndi Kleen * Beware: Something in the Internet is very sensitive to the ordering of 409fd6149d3SIlpo Järvinen * TCP options, we learned this through the hard way, so be careful here. 410fd6149d3SIlpo Järvinen * Luckily we can at least blame others for their non-compliance but from 4118e3bff96Sstephen hemminger * inter-operability perspective it seems that we're somewhat stuck with 412fd6149d3SIlpo Järvinen * the ordering which we have been using if we want to keep working with 413fd6149d3SIlpo Järvinen * those broken things (not that it currently hurts anybody as there isn't 414fd6149d3SIlpo Järvinen * particular reason why the ordering would need to be changed). 415fd6149d3SIlpo Järvinen * 416fd6149d3SIlpo Järvinen * At least SACK_PERM as the first option is known to lead to a disaster 417fd6149d3SIlpo Järvinen * (but it may well be that other scenarios fail similarly). 418fd6149d3SIlpo Järvinen */ 41933ad798cSAdam Langley static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp, 420bd0388aeSWilliam Allen Simpson struct tcp_out_options *opts) 421bd0388aeSWilliam Allen Simpson { 4222100c8d2SYuchung Cheng u16 options = opts->options; /* mungable copy */ 423bd0388aeSWilliam Allen Simpson 424bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_MD5 & options)) { 4251a2c6181SChristoph Paasch *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | 4261a2c6181SChristoph Paasch (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG); 427bd0388aeSWilliam Allen Simpson /* overload cookie hash location */ 428bd0388aeSWilliam Allen Simpson opts->hash_location = (__u8 *)ptr; 42933ad798cSAdam Langley ptr += 4; 43033ad798cSAdam Langley } 43133ad798cSAdam Langley 432fd6149d3SIlpo Järvinen if (unlikely(opts->mss)) { 433fd6149d3SIlpo Järvinen *ptr++ = htonl((TCPOPT_MSS << 24) | 434fd6149d3SIlpo Järvinen (TCPOLEN_MSS << 16) | 435fd6149d3SIlpo Järvinen opts->mss); 436fd6149d3SIlpo Järvinen } 437fd6149d3SIlpo Järvinen 438bd0388aeSWilliam Allen Simpson if (likely(OPTION_TS & options)) { 439bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 44033ad798cSAdam Langley *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | 44133ad798cSAdam Langley (TCPOLEN_SACK_PERM << 16) | 44233ad798cSAdam Langley (TCPOPT_TIMESTAMP << 8) | 44333ad798cSAdam Langley TCPOLEN_TIMESTAMP); 444bd0388aeSWilliam Allen Simpson options &= ~OPTION_SACK_ADVERTISE; 44533ad798cSAdam Langley } else { 446496c98dfSYOSHIFUJI Hideaki *ptr++ = htonl((TCPOPT_NOP << 24) | 44740efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 44840efc6faSStephen Hemminger (TCPOPT_TIMESTAMP << 8) | 44940efc6faSStephen Hemminger TCPOLEN_TIMESTAMP); 45040efc6faSStephen Hemminger } 45133ad798cSAdam Langley *ptr++ = htonl(opts->tsval); 45233ad798cSAdam Langley *ptr++ = htonl(opts->tsecr); 45333ad798cSAdam Langley } 45433ad798cSAdam Langley 455bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 45633ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 45733ad798cSAdam Langley (TCPOPT_NOP << 16) | 45833ad798cSAdam Langley (TCPOPT_SACK_PERM << 8) | 45933ad798cSAdam Langley TCPOLEN_SACK_PERM); 46033ad798cSAdam Langley } 46133ad798cSAdam Langley 462bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_WSCALE & options)) { 46333ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 46433ad798cSAdam Langley (TCPOPT_WINDOW << 16) | 46533ad798cSAdam Langley (TCPOLEN_WINDOW << 8) | 46633ad798cSAdam Langley opts->ws); 46733ad798cSAdam Langley } 46833ad798cSAdam Langley 46933ad798cSAdam Langley if (unlikely(opts->num_sack_blocks)) { 47033ad798cSAdam Langley struct tcp_sack_block *sp = tp->rx_opt.dsack ? 47133ad798cSAdam Langley tp->duplicate_sack : tp->selective_acks; 47240efc6faSStephen Hemminger int this_sack; 47340efc6faSStephen Hemminger 47440efc6faSStephen Hemminger *ptr++ = htonl((TCPOPT_NOP << 24) | 47540efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 47640efc6faSStephen Hemminger (TCPOPT_SACK << 8) | 47733ad798cSAdam Langley (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * 47840efc6faSStephen Hemminger TCPOLEN_SACK_PERBLOCK))); 4792de979bdSStephen Hemminger 48033ad798cSAdam Langley for (this_sack = 0; this_sack < opts->num_sack_blocks; 48133ad798cSAdam Langley ++this_sack) { 48240efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].start_seq); 48340efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].end_seq); 48440efc6faSStephen Hemminger } 4852de979bdSStephen Hemminger 48640efc6faSStephen Hemminger tp->rx_opt.dsack = 0; 48740efc6faSStephen Hemminger } 4882100c8d2SYuchung Cheng 4892100c8d2SYuchung Cheng if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) { 4902100c8d2SYuchung Cheng struct tcp_fastopen_cookie *foc = opts->fastopen_cookie; 4912100c8d2SYuchung Cheng 4922100c8d2SYuchung Cheng *ptr++ = htonl((TCPOPT_EXP << 24) | 4932100c8d2SYuchung Cheng ((TCPOLEN_EXP_FASTOPEN_BASE + foc->len) << 16) | 4942100c8d2SYuchung Cheng TCPOPT_FASTOPEN_MAGIC); 4952100c8d2SYuchung Cheng 4962100c8d2SYuchung Cheng memcpy(ptr, foc->val, foc->len); 4972100c8d2SYuchung Cheng if ((foc->len & 3) == 2) { 4982100c8d2SYuchung Cheng u8 *align = ((u8 *)ptr) + foc->len; 4992100c8d2SYuchung Cheng align[0] = align[1] = TCPOPT_NOP; 5002100c8d2SYuchung Cheng } 5012100c8d2SYuchung Cheng ptr += (foc->len + 3) >> 2; 5022100c8d2SYuchung Cheng } 50340efc6faSStephen Hemminger } 50440efc6faSStephen Hemminger 50567edfef7SAndi Kleen /* Compute TCP options for SYN packets. This is not the final 50667edfef7SAndi Kleen * network wire format yet. 50767edfef7SAndi Kleen */ 50895c96174SEric Dumazet static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb, 50933ad798cSAdam Langley struct tcp_out_options *opts, 510cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 511cf533ea5SEric Dumazet { 51233ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 51395c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 514783237e8SYuchung Cheng struct tcp_fastopen_request *fastopen = tp->fastopen_req; 51533ad798cSAdam Langley 516cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 51733ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 51833ad798cSAdam Langley if (*md5) { 51933ad798cSAdam Langley opts->options |= OPTION_MD5; 520bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 521cfb6eeb4SYOSHIFUJI Hideaki } 52233ad798cSAdam Langley #else 52333ad798cSAdam Langley *md5 = NULL; 524cfb6eeb4SYOSHIFUJI Hideaki #endif 52533ad798cSAdam Langley 52633ad798cSAdam Langley /* We always get an MSS option. The option bytes which will be seen in 52733ad798cSAdam Langley * normal data packets should timestamps be used, must be in the MSS 52833ad798cSAdam Langley * advertised. But we subtract them from tp->mss_cache so that 52933ad798cSAdam Langley * calculations in tcp_sendmsg are simpler etc. So account for this 53033ad798cSAdam Langley * fact here if necessary. If we don't do this correctly, as a 53133ad798cSAdam Langley * receiver we won't recognize data packets as being full sized when we 53233ad798cSAdam Langley * should, and thus we won't abide by the delayed ACK rules correctly. 53333ad798cSAdam Langley * SACKs don't matter, we never delay an ACK when we have any of those 53433ad798cSAdam Langley * going out. */ 53533ad798cSAdam Langley opts->mss = tcp_advertise_mss(sk); 536bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 53733ad798cSAdam Langley 538bb5b7c11SDavid S. Miller if (likely(sysctl_tcp_timestamps && *md5 == NULL)) { 53933ad798cSAdam Langley opts->options |= OPTION_TS; 540ee684b6fSAndrey Vagin opts->tsval = TCP_SKB_CB(skb)->when + tp->tsoffset; 54133ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 542bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 54333ad798cSAdam Langley } 544bb5b7c11SDavid S. Miller if (likely(sysctl_tcp_window_scaling)) { 54533ad798cSAdam Langley opts->ws = tp->rx_opt.rcv_wscale; 54689e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 547bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 54833ad798cSAdam Langley } 549bb5b7c11SDavid S. Miller if (likely(sysctl_tcp_sack)) { 55033ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 551b32d1310SDavid S. Miller if (unlikely(!(OPTION_TS & opts->options))) 552bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 55333ad798cSAdam Langley } 55433ad798cSAdam Langley 555783237e8SYuchung Cheng if (fastopen && fastopen->cookie.len >= 0) { 556783237e8SYuchung Cheng u32 need = TCPOLEN_EXP_FASTOPEN_BASE + fastopen->cookie.len; 557783237e8SYuchung Cheng need = (need + 3) & ~3U; /* Align to 32 bits */ 558783237e8SYuchung Cheng if (remaining >= need) { 559783237e8SYuchung Cheng opts->options |= OPTION_FAST_OPEN_COOKIE; 560783237e8SYuchung Cheng opts->fastopen_cookie = &fastopen->cookie; 561783237e8SYuchung Cheng remaining -= need; 562783237e8SYuchung Cheng tp->syn_fastopen = 1; 563783237e8SYuchung Cheng } 564783237e8SYuchung Cheng } 565bd0388aeSWilliam Allen Simpson 566bd0388aeSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 56733ad798cSAdam Langley } 56833ad798cSAdam Langley 56967edfef7SAndi Kleen /* Set up TCP options for SYN-ACKs. */ 57095c96174SEric Dumazet static unsigned int tcp_synack_options(struct sock *sk, 57133ad798cSAdam Langley struct request_sock *req, 57295c96174SEric Dumazet unsigned int mss, struct sk_buff *skb, 57333ad798cSAdam Langley struct tcp_out_options *opts, 5744957faadSWilliam Allen Simpson struct tcp_md5sig_key **md5, 5758336886fSJerry Chu struct tcp_fastopen_cookie *foc) 5764957faadSWilliam Allen Simpson { 57733ad798cSAdam Langley struct inet_request_sock *ireq = inet_rsk(req); 57895c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 57933ad798cSAdam Langley 58033ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 58133ad798cSAdam Langley *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req); 58233ad798cSAdam Langley if (*md5) { 58333ad798cSAdam Langley opts->options |= OPTION_MD5; 5844957faadSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 5854957faadSWilliam Allen Simpson 5864957faadSWilliam Allen Simpson /* We can't fit any SACK blocks in a packet with MD5 + TS 5874957faadSWilliam Allen Simpson * options. There was discussion about disabling SACK 5884957faadSWilliam Allen Simpson * rather than TS in order to fit in better with old, 5894957faadSWilliam Allen Simpson * buggy kernels, but that was deemed to be unnecessary. 5904957faadSWilliam Allen Simpson */ 591de213e5eSEric Dumazet ireq->tstamp_ok &= !ireq->sack_ok; 59233ad798cSAdam Langley } 59333ad798cSAdam Langley #else 59433ad798cSAdam Langley *md5 = NULL; 59533ad798cSAdam Langley #endif 59633ad798cSAdam Langley 5974957faadSWilliam Allen Simpson /* We always send an MSS option. */ 59833ad798cSAdam Langley opts->mss = mss; 5994957faadSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 60033ad798cSAdam Langley 60133ad798cSAdam Langley if (likely(ireq->wscale_ok)) { 60233ad798cSAdam Langley opts->ws = ireq->rcv_wscale; 60389e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 6044957faadSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 60533ad798cSAdam Langley } 606de213e5eSEric Dumazet if (likely(ireq->tstamp_ok)) { 60733ad798cSAdam Langley opts->options |= OPTION_TS; 60833ad798cSAdam Langley opts->tsval = TCP_SKB_CB(skb)->when; 60933ad798cSAdam Langley opts->tsecr = req->ts_recent; 6104957faadSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 61133ad798cSAdam Langley } 61233ad798cSAdam Langley if (likely(ireq->sack_ok)) { 61333ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 614de213e5eSEric Dumazet if (unlikely(!ireq->tstamp_ok)) 6154957faadSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 61633ad798cSAdam Langley } 6178336886fSJerry Chu if (foc != NULL) { 6188336886fSJerry Chu u32 need = TCPOLEN_EXP_FASTOPEN_BASE + foc->len; 6198336886fSJerry Chu need = (need + 3) & ~3U; /* Align to 32 bits */ 6208336886fSJerry Chu if (remaining >= need) { 6218336886fSJerry Chu opts->options |= OPTION_FAST_OPEN_COOKIE; 6228336886fSJerry Chu opts->fastopen_cookie = foc; 6238336886fSJerry Chu remaining -= need; 6248336886fSJerry Chu } 6258336886fSJerry Chu } 6264957faadSWilliam Allen Simpson 6274957faadSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 62833ad798cSAdam Langley } 62933ad798cSAdam Langley 63067edfef7SAndi Kleen /* Compute TCP options for ESTABLISHED sockets. This is not the 63167edfef7SAndi Kleen * final wire format yet. 63267edfef7SAndi Kleen */ 63395c96174SEric Dumazet static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb, 63433ad798cSAdam Langley struct tcp_out_options *opts, 635cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 636cf533ea5SEric Dumazet { 63733ad798cSAdam Langley struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL; 63833ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 63995c96174SEric Dumazet unsigned int size = 0; 640cabeccbdSIlpo Järvinen unsigned int eff_sacks; 64133ad798cSAdam Langley 6425843ef42SAndi Kleen opts->options = 0; 6435843ef42SAndi Kleen 64433ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 64533ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 64633ad798cSAdam Langley if (unlikely(*md5)) { 64733ad798cSAdam Langley opts->options |= OPTION_MD5; 64833ad798cSAdam Langley size += TCPOLEN_MD5SIG_ALIGNED; 64933ad798cSAdam Langley } 65033ad798cSAdam Langley #else 65133ad798cSAdam Langley *md5 = NULL; 65233ad798cSAdam Langley #endif 65333ad798cSAdam Langley 65433ad798cSAdam Langley if (likely(tp->rx_opt.tstamp_ok)) { 65533ad798cSAdam Langley opts->options |= OPTION_TS; 656ee684b6fSAndrey Vagin opts->tsval = tcb ? tcb->when + tp->tsoffset : 0; 65733ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 65833ad798cSAdam Langley size += TCPOLEN_TSTAMP_ALIGNED; 65933ad798cSAdam Langley } 66033ad798cSAdam Langley 661cabeccbdSIlpo Järvinen eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; 662cabeccbdSIlpo Järvinen if (unlikely(eff_sacks)) { 66395c96174SEric Dumazet const unsigned int remaining = MAX_TCP_OPTION_SPACE - size; 66433ad798cSAdam Langley opts->num_sack_blocks = 66595c96174SEric Dumazet min_t(unsigned int, eff_sacks, 66633ad798cSAdam Langley (remaining - TCPOLEN_SACK_BASE_ALIGNED) / 66733ad798cSAdam Langley TCPOLEN_SACK_PERBLOCK); 66833ad798cSAdam Langley size += TCPOLEN_SACK_BASE_ALIGNED + 66933ad798cSAdam Langley opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; 67033ad798cSAdam Langley } 67133ad798cSAdam Langley 67233ad798cSAdam Langley return size; 67340efc6faSStephen Hemminger } 6741da177e4SLinus Torvalds 67546d3ceabSEric Dumazet 67646d3ceabSEric Dumazet /* TCP SMALL QUEUES (TSQ) 67746d3ceabSEric Dumazet * 67846d3ceabSEric Dumazet * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev) 67946d3ceabSEric Dumazet * to reduce RTT and bufferbloat. 68046d3ceabSEric Dumazet * We do this using a special skb destructor (tcp_wfree). 68146d3ceabSEric Dumazet * 68246d3ceabSEric Dumazet * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb 68346d3ceabSEric Dumazet * needs to be reallocated in a driver. 6848e3bff96Sstephen hemminger * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc 68546d3ceabSEric Dumazet * 68646d3ceabSEric Dumazet * Since transmit from skb destructor is forbidden, we use a tasklet 68746d3ceabSEric Dumazet * to process all sockets that eventually need to send more skbs. 68846d3ceabSEric Dumazet * We use one tasklet per cpu, with its own queue of sockets. 68946d3ceabSEric Dumazet */ 69046d3ceabSEric Dumazet struct tsq_tasklet { 69146d3ceabSEric Dumazet struct tasklet_struct tasklet; 69246d3ceabSEric Dumazet struct list_head head; /* queue of tcp sockets */ 69346d3ceabSEric Dumazet }; 69446d3ceabSEric Dumazet static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet); 69546d3ceabSEric Dumazet 6966f458dfbSEric Dumazet static void tcp_tsq_handler(struct sock *sk) 6976f458dfbSEric Dumazet { 6986f458dfbSEric Dumazet if ((1 << sk->sk_state) & 6996f458dfbSEric Dumazet (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING | 7006f458dfbSEric Dumazet TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) 701bf06200eSJohn Ogness tcp_write_xmit(sk, tcp_current_mss(sk), tcp_sk(sk)->nonagle, 702bf06200eSJohn Ogness 0, GFP_ATOMIC); 7036f458dfbSEric Dumazet } 70446d3ceabSEric Dumazet /* 7058e3bff96Sstephen hemminger * One tasklet per cpu tries to send more skbs. 70646d3ceabSEric Dumazet * We run in tasklet context but need to disable irqs when 7078e3bff96Sstephen hemminger * transferring tsq->head because tcp_wfree() might 70846d3ceabSEric Dumazet * interrupt us (non NAPI drivers) 70946d3ceabSEric Dumazet */ 71046d3ceabSEric Dumazet static void tcp_tasklet_func(unsigned long data) 71146d3ceabSEric Dumazet { 71246d3ceabSEric Dumazet struct tsq_tasklet *tsq = (struct tsq_tasklet *)data; 71346d3ceabSEric Dumazet LIST_HEAD(list); 71446d3ceabSEric Dumazet unsigned long flags; 71546d3ceabSEric Dumazet struct list_head *q, *n; 71646d3ceabSEric Dumazet struct tcp_sock *tp; 71746d3ceabSEric Dumazet struct sock *sk; 71846d3ceabSEric Dumazet 71946d3ceabSEric Dumazet local_irq_save(flags); 72046d3ceabSEric Dumazet list_splice_init(&tsq->head, &list); 72146d3ceabSEric Dumazet local_irq_restore(flags); 72246d3ceabSEric Dumazet 72346d3ceabSEric Dumazet list_for_each_safe(q, n, &list) { 72446d3ceabSEric Dumazet tp = list_entry(q, struct tcp_sock, tsq_node); 72546d3ceabSEric Dumazet list_del(&tp->tsq_node); 72646d3ceabSEric Dumazet 72746d3ceabSEric Dumazet sk = (struct sock *)tp; 72846d3ceabSEric Dumazet bh_lock_sock(sk); 72946d3ceabSEric Dumazet 73046d3ceabSEric Dumazet if (!sock_owned_by_user(sk)) { 7316f458dfbSEric Dumazet tcp_tsq_handler(sk); 73246d3ceabSEric Dumazet } else { 73346d3ceabSEric Dumazet /* defer the work to tcp_release_cb() */ 7346f458dfbSEric Dumazet set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags); 73546d3ceabSEric Dumazet } 73646d3ceabSEric Dumazet bh_unlock_sock(sk); 73746d3ceabSEric Dumazet 73846d3ceabSEric Dumazet clear_bit(TSQ_QUEUED, &tp->tsq_flags); 73946d3ceabSEric Dumazet sk_free(sk); 74046d3ceabSEric Dumazet } 74146d3ceabSEric Dumazet } 74246d3ceabSEric Dumazet 7436f458dfbSEric Dumazet #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \ 7446f458dfbSEric Dumazet (1UL << TCP_WRITE_TIMER_DEFERRED) | \ 745563d34d0SEric Dumazet (1UL << TCP_DELACK_TIMER_DEFERRED) | \ 746563d34d0SEric Dumazet (1UL << TCP_MTU_REDUCED_DEFERRED)) 74746d3ceabSEric Dumazet /** 74846d3ceabSEric Dumazet * tcp_release_cb - tcp release_sock() callback 74946d3ceabSEric Dumazet * @sk: socket 75046d3ceabSEric Dumazet * 75146d3ceabSEric Dumazet * called from release_sock() to perform protocol dependent 75246d3ceabSEric Dumazet * actions before socket release. 75346d3ceabSEric Dumazet */ 75446d3ceabSEric Dumazet void tcp_release_cb(struct sock *sk) 75546d3ceabSEric Dumazet { 75646d3ceabSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 7576f458dfbSEric Dumazet unsigned long flags, nflags; 75846d3ceabSEric Dumazet 7596f458dfbSEric Dumazet /* perform an atomic operation only if at least one flag is set */ 7606f458dfbSEric Dumazet do { 7616f458dfbSEric Dumazet flags = tp->tsq_flags; 7626f458dfbSEric Dumazet if (!(flags & TCP_DEFERRED_ALL)) 7636f458dfbSEric Dumazet return; 7646f458dfbSEric Dumazet nflags = flags & ~TCP_DEFERRED_ALL; 7656f458dfbSEric Dumazet } while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags); 7666f458dfbSEric Dumazet 7676f458dfbSEric Dumazet if (flags & (1UL << TCP_TSQ_DEFERRED)) 7686f458dfbSEric Dumazet tcp_tsq_handler(sk); 7696f458dfbSEric Dumazet 770144d56e9SEric Dumazet if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) { 7716f458dfbSEric Dumazet tcp_write_timer_handler(sk); 772144d56e9SEric Dumazet __sock_put(sk); 773144d56e9SEric Dumazet } 774144d56e9SEric Dumazet if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) { 7756f458dfbSEric Dumazet tcp_delack_timer_handler(sk); 776144d56e9SEric Dumazet __sock_put(sk); 777144d56e9SEric Dumazet } 778144d56e9SEric Dumazet if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) { 779563d34d0SEric Dumazet sk->sk_prot->mtu_reduced(sk); 780144d56e9SEric Dumazet __sock_put(sk); 781144d56e9SEric Dumazet } 78246d3ceabSEric Dumazet } 78346d3ceabSEric Dumazet EXPORT_SYMBOL(tcp_release_cb); 78446d3ceabSEric Dumazet 78546d3ceabSEric Dumazet void __init tcp_tasklet_init(void) 78646d3ceabSEric Dumazet { 78746d3ceabSEric Dumazet int i; 78846d3ceabSEric Dumazet 78946d3ceabSEric Dumazet for_each_possible_cpu(i) { 79046d3ceabSEric Dumazet struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i); 79146d3ceabSEric Dumazet 79246d3ceabSEric Dumazet INIT_LIST_HEAD(&tsq->head); 79346d3ceabSEric Dumazet tasklet_init(&tsq->tasklet, 79446d3ceabSEric Dumazet tcp_tasklet_func, 79546d3ceabSEric Dumazet (unsigned long)tsq); 79646d3ceabSEric Dumazet } 79746d3ceabSEric Dumazet } 79846d3ceabSEric Dumazet 79946d3ceabSEric Dumazet /* 80046d3ceabSEric Dumazet * Write buffer destructor automatically called from kfree_skb. 8018e3bff96Sstephen hemminger * We can't xmit new skbs from this context, as we might already 80246d3ceabSEric Dumazet * hold qdisc lock. 80346d3ceabSEric Dumazet */ 804d6a4a104SEric Dumazet void tcp_wfree(struct sk_buff *skb) 80546d3ceabSEric Dumazet { 80646d3ceabSEric Dumazet struct sock *sk = skb->sk; 80746d3ceabSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 80846d3ceabSEric Dumazet 80946d3ceabSEric Dumazet if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) && 81046d3ceabSEric Dumazet !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) { 81146d3ceabSEric Dumazet unsigned long flags; 81246d3ceabSEric Dumazet struct tsq_tasklet *tsq; 81346d3ceabSEric Dumazet 81446d3ceabSEric Dumazet /* Keep a ref on socket. 81546d3ceabSEric Dumazet * This last ref will be released in tcp_tasklet_func() 81646d3ceabSEric Dumazet */ 81746d3ceabSEric Dumazet atomic_sub(skb->truesize - 1, &sk->sk_wmem_alloc); 81846d3ceabSEric Dumazet 81946d3ceabSEric Dumazet /* queue this socket to tasklet queue */ 82046d3ceabSEric Dumazet local_irq_save(flags); 82146d3ceabSEric Dumazet tsq = &__get_cpu_var(tsq_tasklet); 82246d3ceabSEric Dumazet list_add(&tp->tsq_node, &tsq->head); 82346d3ceabSEric Dumazet tasklet_schedule(&tsq->tasklet); 82446d3ceabSEric Dumazet local_irq_restore(flags); 82546d3ceabSEric Dumazet } else { 82646d3ceabSEric Dumazet sock_wfree(skb); 82746d3ceabSEric Dumazet } 82846d3ceabSEric Dumazet } 82946d3ceabSEric Dumazet 8301da177e4SLinus Torvalds /* This routine actually transmits TCP packets queued in by 8311da177e4SLinus Torvalds * tcp_do_sendmsg(). This is used by both the initial 8321da177e4SLinus Torvalds * transmission and possible later retransmissions. 8331da177e4SLinus Torvalds * All SKB's seen here are completely headerless. It is our 8341da177e4SLinus Torvalds * job to build the TCP header, and pass the packet down to 8351da177e4SLinus Torvalds * IP so it can do the same plus pass the packet off to the 8361da177e4SLinus Torvalds * device. 8371da177e4SLinus Torvalds * 8381da177e4SLinus Torvalds * We are working here with either a clone of the original 8391da177e4SLinus Torvalds * SKB, or a fresh unique copy made by the retransmit engine. 8401da177e4SLinus Torvalds */ 841056834d9SIlpo Järvinen static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, 842056834d9SIlpo Järvinen gfp_t gfp_mask) 8431da177e4SLinus Torvalds { 8446687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 845dfb4b9dcSDavid S. Miller struct inet_sock *inet; 846dfb4b9dcSDavid S. Miller struct tcp_sock *tp; 847dfb4b9dcSDavid S. Miller struct tcp_skb_cb *tcb; 84833ad798cSAdam Langley struct tcp_out_options opts; 84995c96174SEric Dumazet unsigned int tcp_options_size, tcp_header_size; 850cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 8511da177e4SLinus Torvalds struct tcphdr *th; 8521da177e4SLinus Torvalds int err; 8531da177e4SLinus Torvalds 854dfb4b9dcSDavid S. Miller BUG_ON(!skb || !tcp_skb_pcount(skb)); 855dfb4b9dcSDavid S. Miller 856ccdbb6e9SEric Dumazet if (clone_it) { 857ccdbb6e9SEric Dumazet const struct sk_buff *fclone = skb + 1; 858ccdbb6e9SEric Dumazet 859dfb4b9dcSDavid S. Miller /* If congestion control is doing timestamping, we must 860dfb4b9dcSDavid S. Miller * take such a timestamp before we potentially clone/copy. 861dfb4b9dcSDavid S. Miller */ 862164891aaSStephen Hemminger if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP) 863dfb4b9dcSDavid S. Miller __net_timestamp(skb); 864dfb4b9dcSDavid S. Miller 8650e280af0SEric Dumazet if (unlikely(skb->fclone == SKB_FCLONE_ORIG && 8660e280af0SEric Dumazet fclone->fclone == SKB_FCLONE_CLONE)) 8679a9bfd03SEric Dumazet NET_INC_STATS(sock_net(sk), 8680e280af0SEric Dumazet LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES); 8690e280af0SEric Dumazet 870dfb4b9dcSDavid S. Miller if (unlikely(skb_cloned(skb))) 871dfb4b9dcSDavid S. Miller skb = pskb_copy(skb, gfp_mask); 872dfb4b9dcSDavid S. Miller else 873dfb4b9dcSDavid S. Miller skb = skb_clone(skb, gfp_mask); 874dfb4b9dcSDavid S. Miller if (unlikely(!skb)) 875dfb4b9dcSDavid S. Miller return -ENOBUFS; 876dfb4b9dcSDavid S. Miller } 877dfb4b9dcSDavid S. Miller 878dfb4b9dcSDavid S. Miller inet = inet_sk(sk); 879dfb4b9dcSDavid S. Miller tp = tcp_sk(sk); 880dfb4b9dcSDavid S. Miller tcb = TCP_SKB_CB(skb); 88133ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 8821da177e4SLinus Torvalds 8834de075e0SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) 88433ad798cSAdam Langley tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); 88533ad798cSAdam Langley else 88633ad798cSAdam Langley tcp_options_size = tcp_established_options(sk, skb, &opts, 88733ad798cSAdam Langley &md5); 88833ad798cSAdam Langley tcp_header_size = tcp_options_size + sizeof(struct tcphdr); 8891da177e4SLinus Torvalds 890547669d4SEric Dumazet if (tcp_packets_in_flight(tp) == 0) 8916687e988SArnaldo Carvalho de Melo tcp_ca_event(sk, CA_EVENT_TX_START); 892547669d4SEric Dumazet 893547669d4SEric Dumazet /* if no packet is in qdisc/device queue, then allow XPS to select 894547669d4SEric Dumazet * another queue. 895547669d4SEric Dumazet */ 896547669d4SEric Dumazet skb->ooo_okay = sk_wmem_alloc_get(sk) == 0; 8971da177e4SLinus Torvalds 898aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 899aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 90046d3ceabSEric Dumazet 90146d3ceabSEric Dumazet skb_orphan(skb); 90246d3ceabSEric Dumazet skb->sk = sk; 903c9eeec26SEric Dumazet skb->destructor = tcp_wfree; 90446d3ceabSEric Dumazet atomic_add(skb->truesize, &sk->sk_wmem_alloc); 9051da177e4SLinus Torvalds 9061da177e4SLinus Torvalds /* Build TCP header and checksum it. */ 907aa8223c7SArnaldo Carvalho de Melo th = tcp_hdr(skb); 908c720c7e8SEric Dumazet th->source = inet->inet_sport; 909c720c7e8SEric Dumazet th->dest = inet->inet_dport; 9101da177e4SLinus Torvalds th->seq = htonl(tcb->seq); 9111da177e4SLinus Torvalds th->ack_seq = htonl(tp->rcv_nxt); 912df7a3b07SAl Viro *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | 9134de075e0SEric Dumazet tcb->tcp_flags); 914dfb4b9dcSDavid S. Miller 9154de075e0SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) { 9161da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments 9171da177e4SLinus Torvalds * is never scaled. 9181da177e4SLinus Torvalds */ 919600ff0c2SIlpo Järvinen th->window = htons(min(tp->rcv_wnd, 65535U)); 9201da177e4SLinus Torvalds } else { 9211da177e4SLinus Torvalds th->window = htons(tcp_select_window(sk)); 9221da177e4SLinus Torvalds } 9231da177e4SLinus Torvalds th->check = 0; 9241da177e4SLinus Torvalds th->urg_ptr = 0; 9251da177e4SLinus Torvalds 92633f5f57eSIlpo Järvinen /* The urg_mode check is necessary during a below snd_una win probe */ 9277691367dSHerbert Xu if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { 9287691367dSHerbert Xu if (before(tp->snd_up, tcb->seq + 0x10000)) { 9291da177e4SLinus Torvalds th->urg_ptr = htons(tp->snd_up - tcb->seq); 9301da177e4SLinus Torvalds th->urg = 1; 9317691367dSHerbert Xu } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { 9320eae88f3SEric Dumazet th->urg_ptr = htons(0xFFFF); 9337691367dSHerbert Xu th->urg = 1; 9347691367dSHerbert Xu } 9351da177e4SLinus Torvalds } 9361da177e4SLinus Torvalds 937bd0388aeSWilliam Allen Simpson tcp_options_write((__be32 *)(th + 1), tp, &opts); 9384de075e0SEric Dumazet if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0)) 9399e412ba7SIlpo Järvinen TCP_ECN_send(sk, skb, tcp_header_size); 940dfb4b9dcSDavid S. Miller 941cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 942cfb6eeb4SYOSHIFUJI Hideaki /* Calculate the MD5 hash, as we have all we need now */ 943cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 944a465419bSEric Dumazet sk_nocaps_add(sk, NETIF_F_GSO_MASK); 945bd0388aeSWilliam Allen Simpson tp->af_specific->calc_md5_hash(opts.hash_location, 94649a72dfbSAdam Langley md5, sk, NULL, skb); 947cfb6eeb4SYOSHIFUJI Hideaki } 948cfb6eeb4SYOSHIFUJI Hideaki #endif 949cfb6eeb4SYOSHIFUJI Hideaki 950bb296246SHerbert Xu icsk->icsk_af_ops->send_check(sk, skb); 9511da177e4SLinus Torvalds 9524de075e0SEric Dumazet if (likely(tcb->tcp_flags & TCPHDR_ACK)) 953fc6415bcSDavid S. Miller tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); 9541da177e4SLinus Torvalds 9551da177e4SLinus Torvalds if (skb->len != tcp_header_size) 956cf533ea5SEric Dumazet tcp_event_data_sent(tp, sk); 9571da177e4SLinus Torvalds 958bd37a088SWei Yongjun if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) 959aa2ea058STom Herbert TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, 960aa2ea058STom Herbert tcp_skb_pcount(skb)); 9611da177e4SLinus Torvalds 962d9d8da80SDavid S. Miller err = icsk->icsk_af_ops->queue_xmit(skb, &inet->cork.fl); 96383de47cdSHua Zhong if (likely(err <= 0)) 9641da177e4SLinus Torvalds return err; 9651da177e4SLinus Torvalds 9663cfe3baaSIlpo Järvinen tcp_enter_cwr(sk, 1); 9671da177e4SLinus Torvalds 968b9df3cb8SGerrit Renker return net_xmit_eval(err); 9691da177e4SLinus Torvalds } 9701da177e4SLinus Torvalds 97167edfef7SAndi Kleen /* This routine just queues the buffer for sending. 9721da177e4SLinus Torvalds * 9731da177e4SLinus Torvalds * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, 9741da177e4SLinus Torvalds * otherwise socket can stall. 9751da177e4SLinus Torvalds */ 9761da177e4SLinus Torvalds static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) 9771da177e4SLinus Torvalds { 9781da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 9791da177e4SLinus Torvalds 9801da177e4SLinus Torvalds /* Advance write_seq and place onto the write_queue. */ 9811da177e4SLinus Torvalds tp->write_seq = TCP_SKB_CB(skb)->end_seq; 9821da177e4SLinus Torvalds skb_header_release(skb); 983fe067e8aSDavid S. Miller tcp_add_write_queue_tail(sk, skb); 9843ab224beSHideo Aoki sk->sk_wmem_queued += skb->truesize; 9853ab224beSHideo Aoki sk_mem_charge(sk, skb->truesize); 9861da177e4SLinus Torvalds } 9871da177e4SLinus Torvalds 98867edfef7SAndi Kleen /* Initialize TSO segments for a packet. */ 989cf533ea5SEric Dumazet static void tcp_set_skb_tso_segs(const struct sock *sk, struct sk_buff *skb, 990056834d9SIlpo Järvinen unsigned int mss_now) 991f6302d1dSDavid S. Miller { 9927b7fc97aSEric Dumazet struct skb_shared_info *shinfo = skb_shinfo(skb); 9937b7fc97aSEric Dumazet 994c52e2421SEric Dumazet /* Make sure we own this skb before messing gso_size/gso_segs */ 995c52e2421SEric Dumazet WARN_ON_ONCE(skb_cloned(skb)); 996c52e2421SEric Dumazet 9978f26fb1cSEric Dumazet if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) { 998f6302d1dSDavid S. Miller /* Avoid the costly divide in the normal 999f6302d1dSDavid S. Miller * non-TSO case. 1000f6302d1dSDavid S. Miller */ 10017b7fc97aSEric Dumazet shinfo->gso_segs = 1; 10027b7fc97aSEric Dumazet shinfo->gso_size = 0; 10037b7fc97aSEric Dumazet shinfo->gso_type = 0; 1004f6302d1dSDavid S. Miller } else { 10057b7fc97aSEric Dumazet shinfo->gso_segs = DIV_ROUND_UP(skb->len, mss_now); 10067b7fc97aSEric Dumazet shinfo->gso_size = mss_now; 10077b7fc97aSEric Dumazet shinfo->gso_type = sk->sk_gso_type; 10081da177e4SLinus Torvalds } 10091da177e4SLinus Torvalds } 10101da177e4SLinus Torvalds 101191fed7a1SIlpo Järvinen /* When a modification to fackets out becomes necessary, we need to check 101268f8353bSIlpo Järvinen * skb is counted to fackets_out or not. 101391fed7a1SIlpo Järvinen */ 1014cf533ea5SEric Dumazet static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb, 101591fed7a1SIlpo Järvinen int decr) 101691fed7a1SIlpo Järvinen { 1017a47e5a98SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1018a47e5a98SIlpo Järvinen 1019dc86967bSIlpo Järvinen if (!tp->sacked_out || tcp_is_reno(tp)) 102091fed7a1SIlpo Järvinen return; 102191fed7a1SIlpo Järvinen 10226859d494SIlpo Järvinen if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq)) 102391fed7a1SIlpo Järvinen tp->fackets_out -= decr; 102491fed7a1SIlpo Järvinen } 102591fed7a1SIlpo Järvinen 1026797108d1SIlpo Järvinen /* Pcount in the middle of the write queue got changed, we need to do various 1027797108d1SIlpo Järvinen * tweaks to fix counters 1028797108d1SIlpo Järvinen */ 1029cf533ea5SEric Dumazet static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) 1030797108d1SIlpo Järvinen { 1031797108d1SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1032797108d1SIlpo Järvinen 1033797108d1SIlpo Järvinen tp->packets_out -= decr; 1034797108d1SIlpo Järvinen 1035797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 1036797108d1SIlpo Järvinen tp->sacked_out -= decr; 1037797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 1038797108d1SIlpo Järvinen tp->retrans_out -= decr; 1039797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 1040797108d1SIlpo Järvinen tp->lost_out -= decr; 1041797108d1SIlpo Järvinen 1042797108d1SIlpo Järvinen /* Reno case is special. Sigh... */ 1043797108d1SIlpo Järvinen if (tcp_is_reno(tp) && decr > 0) 1044797108d1SIlpo Järvinen tp->sacked_out -= min_t(u32, tp->sacked_out, decr); 1045797108d1SIlpo Järvinen 1046797108d1SIlpo Järvinen tcp_adjust_fackets_out(sk, skb, decr); 1047797108d1SIlpo Järvinen 1048797108d1SIlpo Järvinen if (tp->lost_skb_hint && 1049797108d1SIlpo Järvinen before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && 105052cf3cc8SIlpo Järvinen (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))) 1051797108d1SIlpo Järvinen tp->lost_cnt_hint -= decr; 1052797108d1SIlpo Järvinen 1053797108d1SIlpo Järvinen tcp_verify_left_out(tp); 1054797108d1SIlpo Järvinen } 1055797108d1SIlpo Järvinen 10561da177e4SLinus Torvalds /* Function to create two new TCP segments. Shrinks the given segment 10571da177e4SLinus Torvalds * to the specified size and appends a new segment with the rest of the 10581da177e4SLinus Torvalds * packet to the list. This won't be called frequently, I hope. 10591da177e4SLinus Torvalds * Remember, these are still headerless SKBs at this point. 10601da177e4SLinus Torvalds */ 1061056834d9SIlpo Järvinen int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, 1062056834d9SIlpo Järvinen unsigned int mss_now) 10631da177e4SLinus Torvalds { 10641da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 10651da177e4SLinus Torvalds struct sk_buff *buff; 10666475be16SDavid S. Miller int nsize, old_factor; 1067b60b49eaSHerbert Xu int nlen; 10689ce01461SIlpo Järvinen u8 flags; 10691da177e4SLinus Torvalds 10702fceec13SIlpo Järvinen if (WARN_ON(len > skb->len)) 10712fceec13SIlpo Järvinen return -EINVAL; 10726a438bbeSStephen Hemminger 10731da177e4SLinus Torvalds nsize = skb_headlen(skb) - len; 10741da177e4SLinus Torvalds if (nsize < 0) 10751da177e4SLinus Torvalds nsize = 0; 10761da177e4SLinus Torvalds 1077c52e2421SEric Dumazet if (skb_unclone(skb, GFP_ATOMIC)) 10781da177e4SLinus Torvalds return -ENOMEM; 10791da177e4SLinus Torvalds 10801da177e4SLinus Torvalds /* Get a new skb... force flag on. */ 10811da177e4SLinus Torvalds buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC); 10821da177e4SLinus Torvalds if (buff == NULL) 10831da177e4SLinus Torvalds return -ENOMEM; /* We'll just try again later. */ 1084ef5cb973SHerbert Xu 10853ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 10863ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1087b60b49eaSHerbert Xu nlen = skb->len - len - nsize; 1088b60b49eaSHerbert Xu buff->truesize += nlen; 1089b60b49eaSHerbert Xu skb->truesize -= nlen; 10901da177e4SLinus Torvalds 10911da177e4SLinus Torvalds /* Correct the sequence numbers. */ 10921da177e4SLinus Torvalds TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 10931da177e4SLinus Torvalds TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 10941da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 10951da177e4SLinus Torvalds 10961da177e4SLinus Torvalds /* PSH and FIN should only be set in the second packet. */ 10974de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 10984de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 10994de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1100e14c3cafSHerbert Xu TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; 11011da177e4SLinus Torvalds 110284fa7933SPatrick McHardy if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { 11031da177e4SLinus Torvalds /* Copy and checksum data tail into the new buffer. */ 1104056834d9SIlpo Järvinen buff->csum = csum_partial_copy_nocheck(skb->data + len, 1105056834d9SIlpo Järvinen skb_put(buff, nsize), 11061da177e4SLinus Torvalds nsize, 0); 11071da177e4SLinus Torvalds 11081da177e4SLinus Torvalds skb_trim(skb, len); 11091da177e4SLinus Torvalds 11101da177e4SLinus Torvalds skb->csum = csum_block_sub(skb->csum, buff->csum, len); 11111da177e4SLinus Torvalds } else { 111284fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 11131da177e4SLinus Torvalds skb_split(skb, buff, len); 11141da177e4SLinus Torvalds } 11151da177e4SLinus Torvalds 11161da177e4SLinus Torvalds buff->ip_summed = skb->ip_summed; 11171da177e4SLinus Torvalds 11181da177e4SLinus Torvalds /* Looks stupid, but our code really uses when of 11191da177e4SLinus Torvalds * skbs, which it never sent before. --ANK 11201da177e4SLinus Torvalds */ 11211da177e4SLinus Torvalds TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when; 1122a61bbcf2SPatrick McHardy buff->tstamp = skb->tstamp; 11231da177e4SLinus Torvalds 11246475be16SDavid S. Miller old_factor = tcp_skb_pcount(skb); 11256475be16SDavid S. Miller 11261da177e4SLinus Torvalds /* Fix up tso_factor for both original and new SKB. */ 1127846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1128846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, buff, mss_now); 11291da177e4SLinus Torvalds 11306475be16SDavid S. Miller /* If this packet has been sent out already, we must 11316475be16SDavid S. Miller * adjust the various packet counters. 11326475be16SDavid S. Miller */ 1133cf0b450cSHerbert Xu if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 11346475be16SDavid S. Miller int diff = old_factor - tcp_skb_pcount(skb) - 11356475be16SDavid S. Miller tcp_skb_pcount(buff); 11361da177e4SLinus Torvalds 1137797108d1SIlpo Järvinen if (diff) 1138797108d1SIlpo Järvinen tcp_adjust_pcount(sk, skb, diff); 11391da177e4SLinus Torvalds } 11401da177e4SLinus Torvalds 11411da177e4SLinus Torvalds /* Link BUFF into the send queue. */ 1142f44b5271SDavid S. Miller skb_header_release(buff); 1143fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 11441da177e4SLinus Torvalds 11451da177e4SLinus Torvalds return 0; 11461da177e4SLinus Torvalds } 11471da177e4SLinus Torvalds 11481da177e4SLinus Torvalds /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c 11491da177e4SLinus Torvalds * eventually). The difference is that pulled data not copied, but 11501da177e4SLinus Torvalds * immediately discarded. 11511da177e4SLinus Torvalds */ 1152f2911969SHerbert Xu ~{PmVHI~} static void __pskb_trim_head(struct sk_buff *skb, int len) 11531da177e4SLinus Torvalds { 11547b7fc97aSEric Dumazet struct skb_shared_info *shinfo; 11551da177e4SLinus Torvalds int i, k, eat; 11561da177e4SLinus Torvalds 11574fa48bf3SEric Dumazet eat = min_t(int, len, skb_headlen(skb)); 11584fa48bf3SEric Dumazet if (eat) { 11594fa48bf3SEric Dumazet __skb_pull(skb, eat); 11604fa48bf3SEric Dumazet len -= eat; 11614fa48bf3SEric Dumazet if (!len) 11624fa48bf3SEric Dumazet return; 11634fa48bf3SEric Dumazet } 11641da177e4SLinus Torvalds eat = len; 11651da177e4SLinus Torvalds k = 0; 11667b7fc97aSEric Dumazet shinfo = skb_shinfo(skb); 11677b7fc97aSEric Dumazet for (i = 0; i < shinfo->nr_frags; i++) { 11687b7fc97aSEric Dumazet int size = skb_frag_size(&shinfo->frags[i]); 11699e903e08SEric Dumazet 11709e903e08SEric Dumazet if (size <= eat) { 1171aff65da0SIan Campbell skb_frag_unref(skb, i); 11729e903e08SEric Dumazet eat -= size; 11731da177e4SLinus Torvalds } else { 11747b7fc97aSEric Dumazet shinfo->frags[k] = shinfo->frags[i]; 11751da177e4SLinus Torvalds if (eat) { 11767b7fc97aSEric Dumazet shinfo->frags[k].page_offset += eat; 11777b7fc97aSEric Dumazet skb_frag_size_sub(&shinfo->frags[k], eat); 11781da177e4SLinus Torvalds eat = 0; 11791da177e4SLinus Torvalds } 11801da177e4SLinus Torvalds k++; 11811da177e4SLinus Torvalds } 11821da177e4SLinus Torvalds } 11837b7fc97aSEric Dumazet shinfo->nr_frags = k; 11841da177e4SLinus Torvalds 118527a884dcSArnaldo Carvalho de Melo skb_reset_tail_pointer(skb); 11861da177e4SLinus Torvalds skb->data_len -= len; 11871da177e4SLinus Torvalds skb->len = skb->data_len; 11881da177e4SLinus Torvalds } 11891da177e4SLinus Torvalds 119067edfef7SAndi Kleen /* Remove acked data from a packet in the transmit queue. */ 11911da177e4SLinus Torvalds int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 11921da177e4SLinus Torvalds { 119314bbd6a5SPravin B Shelar if (skb_unclone(skb, GFP_ATOMIC)) 11941da177e4SLinus Torvalds return -ENOMEM; 11951da177e4SLinus Torvalds 11964fa48bf3SEric Dumazet __pskb_trim_head(skb, len); 11971da177e4SLinus Torvalds 11981da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq += len; 119984fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 12001da177e4SLinus Torvalds 12011da177e4SLinus Torvalds skb->truesize -= len; 12021da177e4SLinus Torvalds sk->sk_wmem_queued -= len; 12033ab224beSHideo Aoki sk_mem_uncharge(sk, len); 12041da177e4SLinus Torvalds sock_set_flag(sk, SOCK_QUEUE_SHRUNK); 12051da177e4SLinus Torvalds 12065b35e1e6SNeal Cardwell /* Any change of skb->len requires recalculation of tso factor. */ 12071da177e4SLinus Torvalds if (tcp_skb_pcount(skb) > 1) 12085b35e1e6SNeal Cardwell tcp_set_skb_tso_segs(sk, skb, tcp_skb_mss(skb)); 12091da177e4SLinus Torvalds 12101da177e4SLinus Torvalds return 0; 12111da177e4SLinus Torvalds } 12121da177e4SLinus Torvalds 12131b63edd6SYuchung Cheng /* Calculate MSS not accounting any TCP options. */ 12141b63edd6SYuchung Cheng static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu) 12155d424d5aSJohn Heffner { 1216cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1217cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 12185d424d5aSJohn Heffner int mss_now; 12195d424d5aSJohn Heffner 12205d424d5aSJohn Heffner /* Calculate base mss without TCP options: 12215d424d5aSJohn Heffner It is MMS_S - sizeof(tcphdr) of rfc1122 12225d424d5aSJohn Heffner */ 12235d424d5aSJohn Heffner mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 12245d424d5aSJohn Heffner 122567469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 122667469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 122767469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 122867469601SEric Dumazet 122967469601SEric Dumazet if (dst && dst_allfrag(dst)) 123067469601SEric Dumazet mss_now -= icsk->icsk_af_ops->net_frag_header_len; 123167469601SEric Dumazet } 123267469601SEric Dumazet 12335d424d5aSJohn Heffner /* Clamp it (mss_clamp does not include tcp options) */ 12345d424d5aSJohn Heffner if (mss_now > tp->rx_opt.mss_clamp) 12355d424d5aSJohn Heffner mss_now = tp->rx_opt.mss_clamp; 12365d424d5aSJohn Heffner 12375d424d5aSJohn Heffner /* Now subtract optional transport overhead */ 12385d424d5aSJohn Heffner mss_now -= icsk->icsk_ext_hdr_len; 12395d424d5aSJohn Heffner 12405d424d5aSJohn Heffner /* Then reserve room for full set of TCP options and 8 bytes of data */ 12415d424d5aSJohn Heffner if (mss_now < 48) 12425d424d5aSJohn Heffner mss_now = 48; 12435d424d5aSJohn Heffner return mss_now; 12445d424d5aSJohn Heffner } 12455d424d5aSJohn Heffner 12461b63edd6SYuchung Cheng /* Calculate MSS. Not accounting for SACKs here. */ 12471b63edd6SYuchung Cheng int tcp_mtu_to_mss(struct sock *sk, int pmtu) 12481b63edd6SYuchung Cheng { 12491b63edd6SYuchung Cheng /* Subtract TCP options size, not including SACKs */ 12501b63edd6SYuchung Cheng return __tcp_mtu_to_mss(sk, pmtu) - 12511b63edd6SYuchung Cheng (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr)); 12521b63edd6SYuchung Cheng } 12531b63edd6SYuchung Cheng 12545d424d5aSJohn Heffner /* Inverse of above */ 125567469601SEric Dumazet int tcp_mss_to_mtu(struct sock *sk, int mss) 12565d424d5aSJohn Heffner { 1257cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1258cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 12595d424d5aSJohn Heffner int mtu; 12605d424d5aSJohn Heffner 12615d424d5aSJohn Heffner mtu = mss + 12625d424d5aSJohn Heffner tp->tcp_header_len + 12635d424d5aSJohn Heffner icsk->icsk_ext_hdr_len + 12645d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 12655d424d5aSJohn Heffner 126667469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 126767469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 126867469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 126967469601SEric Dumazet 127067469601SEric Dumazet if (dst && dst_allfrag(dst)) 127167469601SEric Dumazet mtu += icsk->icsk_af_ops->net_frag_header_len; 127267469601SEric Dumazet } 12735d424d5aSJohn Heffner return mtu; 12745d424d5aSJohn Heffner } 12755d424d5aSJohn Heffner 127667edfef7SAndi Kleen /* MTU probing init per socket */ 12775d424d5aSJohn Heffner void tcp_mtup_init(struct sock *sk) 12785d424d5aSJohn Heffner { 12795d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 12805d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 12815d424d5aSJohn Heffner 12825d424d5aSJohn Heffner icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1; 12835d424d5aSJohn Heffner icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 12845d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 12855d424d5aSJohn Heffner icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss); 12865d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 12875d424d5aSJohn Heffner } 12884bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_mtup_init); 12895d424d5aSJohn Heffner 12901da177e4SLinus Torvalds /* This function synchronize snd mss to current pmtu/exthdr set. 12911da177e4SLinus Torvalds 12921da177e4SLinus Torvalds tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 12931da177e4SLinus Torvalds for TCP options, but includes only bare TCP header. 12941da177e4SLinus Torvalds 12951da177e4SLinus Torvalds tp->rx_opt.mss_clamp is mss negotiated at connection setup. 1296caa20d9aSStephen Hemminger It is minimum of user_mss and mss received with SYN. 12971da177e4SLinus Torvalds It also does not include TCP options. 12981da177e4SLinus Torvalds 1299d83d8461SArnaldo Carvalho de Melo inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. 13001da177e4SLinus Torvalds 13011da177e4SLinus Torvalds tp->mss_cache is current effective sending mss, including 13021da177e4SLinus Torvalds all tcp options except for SACKs. It is evaluated, 13031da177e4SLinus Torvalds taking into account current pmtu, but never exceeds 13041da177e4SLinus Torvalds tp->rx_opt.mss_clamp. 13051da177e4SLinus Torvalds 13061da177e4SLinus Torvalds NOTE1. rfc1122 clearly states that advertised MSS 13071da177e4SLinus Torvalds DOES NOT include either tcp or ip options. 13081da177e4SLinus Torvalds 1309d83d8461SArnaldo Carvalho de Melo NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache 1310d83d8461SArnaldo Carvalho de Melo are READ ONLY outside this function. --ANK (980731) 13111da177e4SLinus Torvalds */ 13121da177e4SLinus Torvalds unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 13131da177e4SLinus Torvalds { 13141da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1315d83d8461SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 13165d424d5aSJohn Heffner int mss_now; 13171da177e4SLinus Torvalds 13185d424d5aSJohn Heffner if (icsk->icsk_mtup.search_high > pmtu) 13195d424d5aSJohn Heffner icsk->icsk_mtup.search_high = pmtu; 13201da177e4SLinus Torvalds 13215d424d5aSJohn Heffner mss_now = tcp_mtu_to_mss(sk, pmtu); 1322409d22b4SIlpo Järvinen mss_now = tcp_bound_to_half_wnd(tp, mss_now); 13231da177e4SLinus Torvalds 13241da177e4SLinus Torvalds /* And store cached results */ 1325d83d8461SArnaldo Carvalho de Melo icsk->icsk_pmtu_cookie = pmtu; 13265d424d5aSJohn Heffner if (icsk->icsk_mtup.enabled) 13275d424d5aSJohn Heffner mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); 1328c1b4a7e6SDavid S. Miller tp->mss_cache = mss_now; 13291da177e4SLinus Torvalds 13301da177e4SLinus Torvalds return mss_now; 13311da177e4SLinus Torvalds } 13324bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_sync_mss); 13331da177e4SLinus Torvalds 13341da177e4SLinus Torvalds /* Compute the current effective MSS, taking SACKs and IP options, 13351da177e4SLinus Torvalds * and even PMTU discovery events into account. 13361da177e4SLinus Torvalds */ 13370c54b85fSIlpo Järvinen unsigned int tcp_current_mss(struct sock *sk) 13381da177e4SLinus Torvalds { 1339cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1340cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1341c1b4a7e6SDavid S. Miller u32 mss_now; 134295c96174SEric Dumazet unsigned int header_len; 134333ad798cSAdam Langley struct tcp_out_options opts; 134433ad798cSAdam Langley struct tcp_md5sig_key *md5; 13451da177e4SLinus Torvalds 1346c1b4a7e6SDavid S. Miller mss_now = tp->mss_cache; 1347c1b4a7e6SDavid S. Miller 13481da177e4SLinus Torvalds if (dst) { 13491da177e4SLinus Torvalds u32 mtu = dst_mtu(dst); 1350d83d8461SArnaldo Carvalho de Melo if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 13511da177e4SLinus Torvalds mss_now = tcp_sync_mss(sk, mtu); 13521da177e4SLinus Torvalds } 13531da177e4SLinus Torvalds 135433ad798cSAdam Langley header_len = tcp_established_options(sk, NULL, &opts, &md5) + 135533ad798cSAdam Langley sizeof(struct tcphdr); 135633ad798cSAdam Langley /* The mss_cache is sized based on tp->tcp_header_len, which assumes 135733ad798cSAdam Langley * some common options. If this is an odd packet (because we have SACK 135833ad798cSAdam Langley * blocks etc) then our calculated header_len will be different, and 135933ad798cSAdam Langley * we have to adjust mss_now correspondingly */ 136033ad798cSAdam Langley if (header_len != tp->tcp_header_len) { 136133ad798cSAdam Langley int delta = (int) header_len - tp->tcp_header_len; 136233ad798cSAdam Langley mss_now -= delta; 136333ad798cSAdam Langley } 1364cfb6eeb4SYOSHIFUJI Hideaki 13651da177e4SLinus Torvalds return mss_now; 13661da177e4SLinus Torvalds } 13671da177e4SLinus Torvalds 1368a762a980SDavid S. Miller /* Congestion window validation. (RFC2861) */ 13699e412ba7SIlpo Järvinen static void tcp_cwnd_validate(struct sock *sk) 1370a762a980SDavid S. Miller { 13719e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1372a762a980SDavid S. Miller 1373d436d686SIlpo Järvinen if (tp->packets_out >= tp->snd_cwnd) { 1374a762a980SDavid S. Miller /* Network is feed fully. */ 1375a762a980SDavid S. Miller tp->snd_cwnd_used = 0; 1376a762a980SDavid S. Miller tp->snd_cwnd_stamp = tcp_time_stamp; 1377a762a980SDavid S. Miller } else { 1378a762a980SDavid S. Miller /* Network starves. */ 1379a762a980SDavid S. Miller if (tp->packets_out > tp->snd_cwnd_used) 1380a762a980SDavid S. Miller tp->snd_cwnd_used = tp->packets_out; 1381a762a980SDavid S. Miller 138215d33c07SDavid S. Miller if (sysctl_tcp_slow_start_after_idle && 138315d33c07SDavid S. Miller (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto) 1384a762a980SDavid S. Miller tcp_cwnd_application_limited(sk); 1385a762a980SDavid S. Miller } 1386a762a980SDavid S. Miller } 1387a762a980SDavid S. Miller 1388d4589926SEric Dumazet /* Minshall's variant of the Nagle send check. */ 1389d4589926SEric Dumazet static bool tcp_minshall_check(const struct tcp_sock *tp) 1390d4589926SEric Dumazet { 1391d4589926SEric Dumazet return after(tp->snd_sml, tp->snd_una) && 1392d4589926SEric Dumazet !after(tp->snd_sml, tp->snd_nxt); 1393d4589926SEric Dumazet } 1394d4589926SEric Dumazet 1395d4589926SEric Dumazet /* Update snd_sml if this skb is under mss 1396d4589926SEric Dumazet * Note that a TSO packet might end with a sub-mss segment 1397d4589926SEric Dumazet * The test is really : 1398d4589926SEric Dumazet * if ((skb->len % mss) != 0) 1399d4589926SEric Dumazet * tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1400d4589926SEric Dumazet * But we can avoid doing the divide again given we already have 1401d4589926SEric Dumazet * skb_pcount = skb->len / mss_now 14020e3a4803SIlpo Järvinen */ 1403d4589926SEric Dumazet static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now, 1404d4589926SEric Dumazet const struct sk_buff *skb) 1405d4589926SEric Dumazet { 1406d4589926SEric Dumazet if (skb->len < tcp_skb_pcount(skb) * mss_now) 1407d4589926SEric Dumazet tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1408d4589926SEric Dumazet } 1409d4589926SEric Dumazet 1410d4589926SEric Dumazet /* Return false, if packet can be sent now without violation Nagle's rules: 1411d4589926SEric Dumazet * 1. It is full sized. (provided by caller in %partial bool) 1412d4589926SEric Dumazet * 2. Or it contains FIN. (already checked by caller) 1413d4589926SEric Dumazet * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. 1414d4589926SEric Dumazet * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 1415d4589926SEric Dumazet * With Minshall's modification: all sent small packets are ACKed. 1416d4589926SEric Dumazet */ 1417d4589926SEric Dumazet static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp, 1418d4589926SEric Dumazet unsigned int mss_now, int nonagle) 1419d4589926SEric Dumazet { 1420d4589926SEric Dumazet return partial && 1421d4589926SEric Dumazet ((nonagle & TCP_NAGLE_CORK) || 1422d4589926SEric Dumazet (!nonagle && tp->packets_out && tcp_minshall_check(tp))); 1423d4589926SEric Dumazet } 1424d4589926SEric Dumazet /* Returns the portion of skb which can be sent right away */ 1425d4589926SEric Dumazet static unsigned int tcp_mss_split_point(const struct sock *sk, 1426d4589926SEric Dumazet const struct sk_buff *skb, 1427d4589926SEric Dumazet unsigned int mss_now, 1428d4589926SEric Dumazet unsigned int max_segs, 1429d4589926SEric Dumazet int nonagle) 1430c1b4a7e6SDavid S. Miller { 1431cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1432d4589926SEric Dumazet u32 partial, needed, window, max_len; 1433c1b4a7e6SDavid S. Miller 143490840defSIlpo Järvinen window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 14351485348dSBen Hutchings max_len = mss_now * max_segs; 14360e3a4803SIlpo Järvinen 14371485348dSBen Hutchings if (likely(max_len <= window && skb != tcp_write_queue_tail(sk))) 14381485348dSBen Hutchings return max_len; 14390e3a4803SIlpo Järvinen 14405ea3a748SIlpo Järvinen needed = min(skb->len, window); 14415ea3a748SIlpo Järvinen 14421485348dSBen Hutchings if (max_len <= needed) 14431485348dSBen Hutchings return max_len; 14440e3a4803SIlpo Järvinen 1445d4589926SEric Dumazet partial = needed % mss_now; 1446d4589926SEric Dumazet /* If last segment is not a full MSS, check if Nagle rules allow us 1447d4589926SEric Dumazet * to include this last segment in this skb. 1448d4589926SEric Dumazet * Otherwise, we'll split the skb at last MSS boundary 1449d4589926SEric Dumazet */ 1450d4589926SEric Dumazet if (tcp_nagle_check(partial != 0, tp, mss_now, nonagle)) 1451d4589926SEric Dumazet return needed - partial; 1452d4589926SEric Dumazet 1453d4589926SEric Dumazet return needed; 1454c1b4a7e6SDavid S. Miller } 1455c1b4a7e6SDavid S. Miller 1456c1b4a7e6SDavid S. Miller /* Can at least one segment of SKB be sent right now, according to the 1457c1b4a7e6SDavid S. Miller * congestion window rules? If so, return how many segments are allowed. 1458c1b4a7e6SDavid S. Miller */ 1459cf533ea5SEric Dumazet static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, 1460cf533ea5SEric Dumazet const struct sk_buff *skb) 1461c1b4a7e6SDavid S. Miller { 1462c1b4a7e6SDavid S. Miller u32 in_flight, cwnd; 1463c1b4a7e6SDavid S. Miller 1464c1b4a7e6SDavid S. Miller /* Don't be strict about the congestion window for the final FIN. */ 14654de075e0SEric Dumazet if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && 14664de075e0SEric Dumazet tcp_skb_pcount(skb) == 1) 1467c1b4a7e6SDavid S. Miller return 1; 1468c1b4a7e6SDavid S. Miller 1469c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1470c1b4a7e6SDavid S. Miller cwnd = tp->snd_cwnd; 1471c1b4a7e6SDavid S. Miller if (in_flight < cwnd) 1472c1b4a7e6SDavid S. Miller return (cwnd - in_flight); 1473c1b4a7e6SDavid S. Miller 1474c1b4a7e6SDavid S. Miller return 0; 1475c1b4a7e6SDavid S. Miller } 1476c1b4a7e6SDavid S. Miller 1477b595076aSUwe Kleine-König /* Initialize TSO state of a skb. 147867edfef7SAndi Kleen * This must be invoked the first time we consider transmitting 1479c1b4a7e6SDavid S. Miller * SKB onto the wire. 1480c1b4a7e6SDavid S. Miller */ 1481cf533ea5SEric Dumazet static int tcp_init_tso_segs(const struct sock *sk, struct sk_buff *skb, 1482056834d9SIlpo Järvinen unsigned int mss_now) 1483c1b4a7e6SDavid S. Miller { 1484c1b4a7e6SDavid S. Miller int tso_segs = tcp_skb_pcount(skb); 1485c1b4a7e6SDavid S. Miller 1486f8269a49SIlpo Järvinen if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { 1487846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1488c1b4a7e6SDavid S. Miller tso_segs = tcp_skb_pcount(skb); 1489c1b4a7e6SDavid S. Miller } 1490c1b4a7e6SDavid S. Miller return tso_segs; 1491c1b4a7e6SDavid S. Miller } 1492c1b4a7e6SDavid S. Miller 1493c1b4a7e6SDavid S. Miller 1494a2a385d6SEric Dumazet /* Return true if the Nagle test allows this packet to be 1495c1b4a7e6SDavid S. Miller * sent now. 1496c1b4a7e6SDavid S. Miller */ 1497a2a385d6SEric Dumazet static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, 1498c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1499c1b4a7e6SDavid S. Miller { 1500c1b4a7e6SDavid S. Miller /* Nagle rule does not apply to frames, which sit in the middle of the 1501c1b4a7e6SDavid S. Miller * write_queue (they have no chances to get new data). 1502c1b4a7e6SDavid S. Miller * 1503c1b4a7e6SDavid S. Miller * This is implemented in the callers, where they modify the 'nonagle' 1504c1b4a7e6SDavid S. Miller * argument based upon the location of SKB in the send queue. 1505c1b4a7e6SDavid S. Miller */ 1506c1b4a7e6SDavid S. Miller if (nonagle & TCP_NAGLE_PUSH) 1507a2a385d6SEric Dumazet return true; 1508c1b4a7e6SDavid S. Miller 15099b44190dSYuchung Cheng /* Don't use the nagle rule for urgent data (or for the final FIN). */ 15109b44190dSYuchung Cheng if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) 1511a2a385d6SEric Dumazet return true; 1512c1b4a7e6SDavid S. Miller 1513d4589926SEric Dumazet if (!tcp_nagle_check(skb->len < cur_mss, tp, cur_mss, nonagle)) 1514a2a385d6SEric Dumazet return true; 1515c1b4a7e6SDavid S. Miller 1516a2a385d6SEric Dumazet return false; 1517c1b4a7e6SDavid S. Miller } 1518c1b4a7e6SDavid S. Miller 1519c1b4a7e6SDavid S. Miller /* Does at least the first segment of SKB fit into the send window? */ 1520a2a385d6SEric Dumazet static bool tcp_snd_wnd_test(const struct tcp_sock *tp, 1521a2a385d6SEric Dumazet const struct sk_buff *skb, 1522056834d9SIlpo Järvinen unsigned int cur_mss) 1523c1b4a7e6SDavid S. Miller { 1524c1b4a7e6SDavid S. Miller u32 end_seq = TCP_SKB_CB(skb)->end_seq; 1525c1b4a7e6SDavid S. Miller 1526c1b4a7e6SDavid S. Miller if (skb->len > cur_mss) 1527c1b4a7e6SDavid S. Miller end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 1528c1b4a7e6SDavid S. Miller 152990840defSIlpo Järvinen return !after(end_seq, tcp_wnd_end(tp)); 1530c1b4a7e6SDavid S. Miller } 1531c1b4a7e6SDavid S. Miller 1532fe067e8aSDavid S. Miller /* This checks if the data bearing packet SKB (usually tcp_send_head(sk)) 1533c1b4a7e6SDavid S. Miller * should be put on the wire right now. If so, it returns the number of 1534c1b4a7e6SDavid S. Miller * packets allowed by the congestion window. 1535c1b4a7e6SDavid S. Miller */ 1536cf533ea5SEric Dumazet static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb, 1537c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1538c1b4a7e6SDavid S. Miller { 1539cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1540c1b4a7e6SDavid S. Miller unsigned int cwnd_quota; 1541c1b4a7e6SDavid S. Miller 1542846998aeSDavid S. Miller tcp_init_tso_segs(sk, skb, cur_mss); 1543c1b4a7e6SDavid S. Miller 1544c1b4a7e6SDavid S. Miller if (!tcp_nagle_test(tp, skb, cur_mss, nonagle)) 1545c1b4a7e6SDavid S. Miller return 0; 1546c1b4a7e6SDavid S. Miller 1547c1b4a7e6SDavid S. Miller cwnd_quota = tcp_cwnd_test(tp, skb); 1548056834d9SIlpo Järvinen if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss)) 1549c1b4a7e6SDavid S. Miller cwnd_quota = 0; 1550c1b4a7e6SDavid S. Miller 1551c1b4a7e6SDavid S. Miller return cwnd_quota; 1552c1b4a7e6SDavid S. Miller } 1553c1b4a7e6SDavid S. Miller 155467edfef7SAndi Kleen /* Test if sending is allowed right now. */ 1555a2a385d6SEric Dumazet bool tcp_may_send_now(struct sock *sk) 1556c1b4a7e6SDavid S. Miller { 1557cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1558fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 1559c1b4a7e6SDavid S. Miller 1560a02cec21SEric Dumazet return skb && 15610c54b85fSIlpo Järvinen tcp_snd_test(sk, skb, tcp_current_mss(sk), 1562c1b4a7e6SDavid S. Miller (tcp_skb_is_last(sk, skb) ? 1563a02cec21SEric Dumazet tp->nonagle : TCP_NAGLE_PUSH)); 1564c1b4a7e6SDavid S. Miller } 1565c1b4a7e6SDavid S. Miller 1566c1b4a7e6SDavid S. Miller /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 1567c1b4a7e6SDavid S. Miller * which is put after SKB on the list. It is very much like 1568c1b4a7e6SDavid S. Miller * tcp_fragment() except that it may make several kinds of assumptions 1569c1b4a7e6SDavid S. Miller * in order to speed up the splitting operation. In particular, we 1570c1b4a7e6SDavid S. Miller * know that all the data is in scatter-gather pages, and that the 1571c1b4a7e6SDavid S. Miller * packet has never been sent out before (and thus is not cloned). 1572c1b4a7e6SDavid S. Miller */ 1573056834d9SIlpo Järvinen static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, 1574c4ead4c5SEric Dumazet unsigned int mss_now, gfp_t gfp) 1575c1b4a7e6SDavid S. Miller { 1576c1b4a7e6SDavid S. Miller struct sk_buff *buff; 1577c1b4a7e6SDavid S. Miller int nlen = skb->len - len; 15789ce01461SIlpo Järvinen u8 flags; 1579c1b4a7e6SDavid S. Miller 1580c1b4a7e6SDavid S. Miller /* All of a TSO frame must be composed of paged data. */ 1581c8ac3774SHerbert Xu if (skb->len != skb->data_len) 1582c8ac3774SHerbert Xu return tcp_fragment(sk, skb, len, mss_now); 1583c1b4a7e6SDavid S. Miller 1584c4ead4c5SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, gfp); 1585c1b4a7e6SDavid S. Miller if (unlikely(buff == NULL)) 1586c1b4a7e6SDavid S. Miller return -ENOMEM; 1587c1b4a7e6SDavid S. Miller 15883ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 15893ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1590b60b49eaSHerbert Xu buff->truesize += nlen; 1591c1b4a7e6SDavid S. Miller skb->truesize -= nlen; 1592c1b4a7e6SDavid S. Miller 1593c1b4a7e6SDavid S. Miller /* Correct the sequence numbers. */ 1594c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 1595c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 1596c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 1597c1b4a7e6SDavid S. Miller 1598c1b4a7e6SDavid S. Miller /* PSH and FIN should only be set in the second packet. */ 15994de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 16004de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 16014de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1602c1b4a7e6SDavid S. Miller 1603c1b4a7e6SDavid S. Miller /* This packet was never sent out yet, so no SACK bits. */ 1604c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->sacked = 0; 1605c1b4a7e6SDavid S. Miller 160684fa7933SPatrick McHardy buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; 1607c1b4a7e6SDavid S. Miller skb_split(skb, buff, len); 1608c1b4a7e6SDavid S. Miller 1609c1b4a7e6SDavid S. Miller /* Fix up tso_factor for both original and new SKB. */ 1610846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1611846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, buff, mss_now); 1612c1b4a7e6SDavid S. Miller 1613c1b4a7e6SDavid S. Miller /* Link BUFF into the send queue. */ 1614c1b4a7e6SDavid S. Miller skb_header_release(buff); 1615fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 1616c1b4a7e6SDavid S. Miller 1617c1b4a7e6SDavid S. Miller return 0; 1618c1b4a7e6SDavid S. Miller } 1619c1b4a7e6SDavid S. Miller 1620c1b4a7e6SDavid S. Miller /* Try to defer sending, if possible, in order to minimize the amount 1621c1b4a7e6SDavid S. Miller * of TSO splitting we do. View it as a kind of TSO Nagle test. 1622c1b4a7e6SDavid S. Miller * 1623c1b4a7e6SDavid S. Miller * This algorithm is from John Heffner. 1624c1b4a7e6SDavid S. Miller */ 1625a2a385d6SEric Dumazet static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb) 1626c1b4a7e6SDavid S. Miller { 16279e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 16286687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 1629c1b4a7e6SDavid S. Miller u32 send_win, cong_win, limit, in_flight; 1630ad9f4f50SEric Dumazet int win_divisor; 1631c1b4a7e6SDavid S. Miller 16324de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 1633ae8064acSJohn Heffner goto send_now; 1634c1b4a7e6SDavid S. Miller 16356687e988SArnaldo Carvalho de Melo if (icsk->icsk_ca_state != TCP_CA_Open) 1636ae8064acSJohn Heffner goto send_now; 1637ae8064acSJohn Heffner 1638ae8064acSJohn Heffner /* Defer for less than two clock ticks. */ 1639bd515c3eSIlpo Järvinen if (tp->tso_deferred && 1640a2acde07SIlpo Järvinen (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1) 1641ae8064acSJohn Heffner goto send_now; 1642908a75c1SDavid S. Miller 1643c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1644c1b4a7e6SDavid S. Miller 1645056834d9SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight)); 1646c1b4a7e6SDavid S. Miller 164790840defSIlpo Järvinen send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1648c1b4a7e6SDavid S. Miller 1649c1b4a7e6SDavid S. Miller /* From in_flight test above, we know that cwnd > in_flight. */ 1650c1b4a7e6SDavid S. Miller cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; 1651c1b4a7e6SDavid S. Miller 1652c1b4a7e6SDavid S. Miller limit = min(send_win, cong_win); 1653c1b4a7e6SDavid S. Miller 1654ba244fe9SDavid S. Miller /* If a full-sized TSO skb can be sent, do it. */ 16551485348dSBen Hutchings if (limit >= min_t(unsigned int, sk->sk_gso_max_size, 165695bd09ebSEric Dumazet tp->xmit_size_goal_segs * tp->mss_cache)) 1657ae8064acSJohn Heffner goto send_now; 1658ba244fe9SDavid S. Miller 165962ad2761SIlpo Järvinen /* Middle in queue won't get any more data, full sendable already? */ 166062ad2761SIlpo Järvinen if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) 166162ad2761SIlpo Järvinen goto send_now; 166262ad2761SIlpo Järvinen 1663ad9f4f50SEric Dumazet win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor); 1664ad9f4f50SEric Dumazet if (win_divisor) { 1665c1b4a7e6SDavid S. Miller u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); 1666c1b4a7e6SDavid S. Miller 1667c1b4a7e6SDavid S. Miller /* If at least some fraction of a window is available, 1668c1b4a7e6SDavid S. Miller * just use it. 1669c1b4a7e6SDavid S. Miller */ 1670ad9f4f50SEric Dumazet chunk /= win_divisor; 1671c1b4a7e6SDavid S. Miller if (limit >= chunk) 1672ae8064acSJohn Heffner goto send_now; 1673c1b4a7e6SDavid S. Miller } else { 1674c1b4a7e6SDavid S. Miller /* Different approach, try not to defer past a single 1675c1b4a7e6SDavid S. Miller * ACK. Receiver should ACK every other full sized 1676c1b4a7e6SDavid S. Miller * frame, so if we have space for more than 3 frames 1677c1b4a7e6SDavid S. Miller * then send now. 1678c1b4a7e6SDavid S. Miller */ 16796b5a5c0dSNeal Cardwell if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) 1680ae8064acSJohn Heffner goto send_now; 1681c1b4a7e6SDavid S. Miller } 1682c1b4a7e6SDavid S. Miller 1683f4541d60SEric Dumazet /* Ok, it looks like it is advisable to defer. 1684f4541d60SEric Dumazet * Do not rearm the timer if already set to not break TCP ACK clocking. 1685f4541d60SEric Dumazet */ 1686f4541d60SEric Dumazet if (!tp->tso_deferred) 1687ae8064acSJohn Heffner tp->tso_deferred = 1 | (jiffies << 1); 1688ae8064acSJohn Heffner 1689a2a385d6SEric Dumazet return true; 1690ae8064acSJohn Heffner 1691ae8064acSJohn Heffner send_now: 1692ae8064acSJohn Heffner tp->tso_deferred = 0; 1693a2a385d6SEric Dumazet return false; 1694c1b4a7e6SDavid S. Miller } 1695c1b4a7e6SDavid S. Miller 16965d424d5aSJohn Heffner /* Create a new MTU probe if we are ready. 169767edfef7SAndi Kleen * MTU probe is regularly attempting to increase the path MTU by 169867edfef7SAndi Kleen * deliberately sending larger packets. This discovers routing 169967edfef7SAndi Kleen * changes resulting in larger path MTUs. 170067edfef7SAndi Kleen * 17015d424d5aSJohn Heffner * Returns 0 if we should wait to probe (no cwnd available), 17025d424d5aSJohn Heffner * 1 if a probe was sent, 1703056834d9SIlpo Järvinen * -1 otherwise 1704056834d9SIlpo Järvinen */ 17055d424d5aSJohn Heffner static int tcp_mtu_probe(struct sock *sk) 17065d424d5aSJohn Heffner { 17075d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 17085d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 17095d424d5aSJohn Heffner struct sk_buff *skb, *nskb, *next; 17105d424d5aSJohn Heffner int len; 17115d424d5aSJohn Heffner int probe_size; 171291cc17c0SIlpo Järvinen int size_needed; 17135d424d5aSJohn Heffner int copy; 17145d424d5aSJohn Heffner int mss_now; 17155d424d5aSJohn Heffner 17165d424d5aSJohn Heffner /* Not currently probing/verifying, 17175d424d5aSJohn Heffner * not in recovery, 17185d424d5aSJohn Heffner * have enough cwnd, and 17195d424d5aSJohn Heffner * not SACKing (the variable headers throw things off) */ 17205d424d5aSJohn Heffner if (!icsk->icsk_mtup.enabled || 17215d424d5aSJohn Heffner icsk->icsk_mtup.probe_size || 17225d424d5aSJohn Heffner inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 17235d424d5aSJohn Heffner tp->snd_cwnd < 11 || 1724cabeccbdSIlpo Järvinen tp->rx_opt.num_sacks || tp->rx_opt.dsack) 17255d424d5aSJohn Heffner return -1; 17265d424d5aSJohn Heffner 17275d424d5aSJohn Heffner /* Very simple search strategy: just double the MSS. */ 17280c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 17295d424d5aSJohn Heffner probe_size = 2 * tp->mss_cache; 173091cc17c0SIlpo Järvinen size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; 17315d424d5aSJohn Heffner if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) { 17325d424d5aSJohn Heffner /* TODO: set timer for probe_converge_event */ 17335d424d5aSJohn Heffner return -1; 17345d424d5aSJohn Heffner } 17355d424d5aSJohn Heffner 17365d424d5aSJohn Heffner /* Have enough data in the send queue to probe? */ 17377f9c33e5SIlpo Järvinen if (tp->write_seq - tp->snd_nxt < size_needed) 17385d424d5aSJohn Heffner return -1; 17395d424d5aSJohn Heffner 174091cc17c0SIlpo Järvinen if (tp->snd_wnd < size_needed) 17415d424d5aSJohn Heffner return -1; 174290840defSIlpo Järvinen if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) 17435d424d5aSJohn Heffner return 0; 17445d424d5aSJohn Heffner 1745d67c58e9SIlpo Järvinen /* Do we need to wait to drain cwnd? With none in flight, don't stall */ 1746d67c58e9SIlpo Järvinen if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { 1747d67c58e9SIlpo Järvinen if (!tcp_packets_in_flight(tp)) 17485d424d5aSJohn Heffner return -1; 17495d424d5aSJohn Heffner else 17505d424d5aSJohn Heffner return 0; 17515d424d5aSJohn Heffner } 17525d424d5aSJohn Heffner 17535d424d5aSJohn Heffner /* We're allowed to probe. Build it now. */ 17545d424d5aSJohn Heffner if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL) 17555d424d5aSJohn Heffner return -1; 17563ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 17573ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 17585d424d5aSJohn Heffner 1759fe067e8aSDavid S. Miller skb = tcp_send_head(sk); 17605d424d5aSJohn Heffner 17615d424d5aSJohn Heffner TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 17625d424d5aSJohn Heffner TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 17634de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; 17645d424d5aSJohn Heffner TCP_SKB_CB(nskb)->sacked = 0; 17655d424d5aSJohn Heffner nskb->csum = 0; 176684fa7933SPatrick McHardy nskb->ip_summed = skb->ip_summed; 17675d424d5aSJohn Heffner 176850c4817eSIlpo Järvinen tcp_insert_write_queue_before(nskb, skb, sk); 176950c4817eSIlpo Järvinen 17705d424d5aSJohn Heffner len = 0; 1771234b6860SIlpo Järvinen tcp_for_write_queue_from_safe(skb, next, sk) { 17725d424d5aSJohn Heffner copy = min_t(int, skb->len, probe_size - len); 17735d424d5aSJohn Heffner if (nskb->ip_summed) 17745d424d5aSJohn Heffner skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); 17755d424d5aSJohn Heffner else 17765d424d5aSJohn Heffner nskb->csum = skb_copy_and_csum_bits(skb, 0, 1777056834d9SIlpo Järvinen skb_put(nskb, copy), 1778056834d9SIlpo Järvinen copy, nskb->csum); 17795d424d5aSJohn Heffner 17805d424d5aSJohn Heffner if (skb->len <= copy) { 17815d424d5aSJohn Heffner /* We've eaten all the data from this skb. 17825d424d5aSJohn Heffner * Throw it away. */ 17834de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 1784fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 17853ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 17865d424d5aSJohn Heffner } else { 17874de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & 1788a3433f35SChangli Gao ~(TCPHDR_FIN|TCPHDR_PSH); 17895d424d5aSJohn Heffner if (!skb_shinfo(skb)->nr_frags) { 17905d424d5aSJohn Heffner skb_pull(skb, copy); 179184fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 1792056834d9SIlpo Järvinen skb->csum = csum_partial(skb->data, 1793056834d9SIlpo Järvinen skb->len, 0); 17945d424d5aSJohn Heffner } else { 17955d424d5aSJohn Heffner __pskb_trim_head(skb, copy); 17965d424d5aSJohn Heffner tcp_set_skb_tso_segs(sk, skb, mss_now); 17975d424d5aSJohn Heffner } 17985d424d5aSJohn Heffner TCP_SKB_CB(skb)->seq += copy; 17995d424d5aSJohn Heffner } 18005d424d5aSJohn Heffner 18015d424d5aSJohn Heffner len += copy; 1802234b6860SIlpo Järvinen 1803234b6860SIlpo Järvinen if (len >= probe_size) 1804234b6860SIlpo Järvinen break; 18055d424d5aSJohn Heffner } 18065d424d5aSJohn Heffner tcp_init_tso_segs(sk, nskb, nskb->len); 18075d424d5aSJohn Heffner 18085d424d5aSJohn Heffner /* We're ready to send. If this fails, the probe will 18095d424d5aSJohn Heffner * be resegmented into mss-sized pieces by tcp_write_xmit(). */ 18105d424d5aSJohn Heffner TCP_SKB_CB(nskb)->when = tcp_time_stamp; 18115d424d5aSJohn Heffner if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 18125d424d5aSJohn Heffner /* Decrement cwnd here because we are sending 18135d424d5aSJohn Heffner * effectively two packets. */ 18145d424d5aSJohn Heffner tp->snd_cwnd--; 181566f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, nskb); 18165d424d5aSJohn Heffner 18175d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 18180e7b1368SJohn Heffner tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 18190e7b1368SJohn Heffner tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 18205d424d5aSJohn Heffner 18215d424d5aSJohn Heffner return 1; 18225d424d5aSJohn Heffner } 18235d424d5aSJohn Heffner 18245d424d5aSJohn Heffner return -1; 18255d424d5aSJohn Heffner } 18265d424d5aSJohn Heffner 18271da177e4SLinus Torvalds /* This routine writes packets to the network. It advances the 18281da177e4SLinus Torvalds * send_head. This happens as incoming acks open up the remote 18291da177e4SLinus Torvalds * window for us. 18301da177e4SLinus Torvalds * 1831f8269a49SIlpo Järvinen * LARGESEND note: !tcp_urg_mode is overkill, only frames between 1832f8269a49SIlpo Järvinen * snd_up-64k-mss .. snd_up cannot be large. However, taking into 1833f8269a49SIlpo Järvinen * account rare use of URG, this is not a big flaw. 1834f8269a49SIlpo Järvinen * 18356ba8a3b1SNandita Dukkipati * Send at most one packet when push_one > 0. Temporarily ignore 18366ba8a3b1SNandita Dukkipati * cwnd limit to force at most one packet out when push_one == 2. 18376ba8a3b1SNandita Dukkipati 1838a2a385d6SEric Dumazet * Returns true, if no segments are in flight and we have queued segments, 1839a2a385d6SEric Dumazet * but cannot send anything now because of SWS or another problem. 18401da177e4SLinus Torvalds */ 1841a2a385d6SEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 1842d5dd9175SIlpo Järvinen int push_one, gfp_t gfp) 18431da177e4SLinus Torvalds { 18441da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 184592df7b51SDavid S. Miller struct sk_buff *skb; 1846c1b4a7e6SDavid S. Miller unsigned int tso_segs, sent_pkts; 1847c1b4a7e6SDavid S. Miller int cwnd_quota; 18485d424d5aSJohn Heffner int result; 18491da177e4SLinus Torvalds 1850c1b4a7e6SDavid S. Miller sent_pkts = 0; 18515d424d5aSJohn Heffner 1852d5dd9175SIlpo Järvinen if (!push_one) { 18535d424d5aSJohn Heffner /* Do MTU probing. */ 1854d5dd9175SIlpo Järvinen result = tcp_mtu_probe(sk); 1855d5dd9175SIlpo Järvinen if (!result) { 1856a2a385d6SEric Dumazet return false; 18575d424d5aSJohn Heffner } else if (result > 0) { 18585d424d5aSJohn Heffner sent_pkts = 1; 18595d424d5aSJohn Heffner } 1860d5dd9175SIlpo Järvinen } 18615d424d5aSJohn Heffner 1862fe067e8aSDavid S. Miller while ((skb = tcp_send_head(sk))) { 1863c8ac3774SHerbert Xu unsigned int limit; 1864c8ac3774SHerbert Xu 1865b68e9f85SHerbert Xu tso_segs = tcp_init_tso_segs(sk, skb, mss_now); 1866c1b4a7e6SDavid S. Miller BUG_ON(!tso_segs); 1867c1b4a7e6SDavid S. Miller 1868ec342325SAndrew Vagin if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) 1869ec342325SAndrew Vagin goto repair; /* Skip network transmission */ 1870ec342325SAndrew Vagin 1871b68e9f85SHerbert Xu cwnd_quota = tcp_cwnd_test(tp, skb); 18726ba8a3b1SNandita Dukkipati if (!cwnd_quota) { 18736ba8a3b1SNandita Dukkipati if (push_one == 2) 18746ba8a3b1SNandita Dukkipati /* Force out a loss probe pkt. */ 18756ba8a3b1SNandita Dukkipati cwnd_quota = 1; 18766ba8a3b1SNandita Dukkipati else 1877b68e9f85SHerbert Xu break; 18786ba8a3b1SNandita Dukkipati } 1879b68e9f85SHerbert Xu 1880b68e9f85SHerbert Xu if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) 1881b68e9f85SHerbert Xu break; 1882b68e9f85SHerbert Xu 1883c1b4a7e6SDavid S. Miller if (tso_segs == 1) { 1884aa93466bSDavid S. Miller if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 1885aa93466bSDavid S. Miller (tcp_skb_is_last(sk, skb) ? 1886aa93466bSDavid S. Miller nonagle : TCP_NAGLE_PUSH)))) 1887aa93466bSDavid S. Miller break; 1888c1b4a7e6SDavid S. Miller } else { 1889d5dd9175SIlpo Järvinen if (!push_one && tcp_tso_should_defer(sk, skb)) 1890aa93466bSDavid S. Miller break; 1891c1b4a7e6SDavid S. Miller } 1892aa93466bSDavid S. Miller 1893c9eeec26SEric Dumazet /* TCP Small Queues : 1894c9eeec26SEric Dumazet * Control number of packets in qdisc/devices to two packets / or ~1 ms. 1895c9eeec26SEric Dumazet * This allows for : 1896c9eeec26SEric Dumazet * - better RTT estimation and ACK scheduling 1897c9eeec26SEric Dumazet * - faster recovery 1898c9eeec26SEric Dumazet * - high rates 189998e09386SEric Dumazet * Alas, some drivers / subsystems require a fair amount 190098e09386SEric Dumazet * of queued bytes to ensure line rate. 190198e09386SEric Dumazet * One example is wifi aggregation (802.11 AMPDU) 190246d3ceabSEric Dumazet */ 190398e09386SEric Dumazet limit = max_t(unsigned int, sysctl_tcp_limit_output_bytes, 190498e09386SEric Dumazet sk->sk_pacing_rate >> 10); 1905c9eeec26SEric Dumazet 1906c9eeec26SEric Dumazet if (atomic_read(&sk->sk_wmem_alloc) > limit) { 190746d3ceabSEric Dumazet set_bit(TSQ_THROTTLED, &tp->tsq_flags); 1908bf06200eSJohn Ogness /* It is possible TX completion already happened 1909bf06200eSJohn Ogness * before we set TSQ_THROTTLED, so we must 1910bf06200eSJohn Ogness * test again the condition. 1911bf06200eSJohn Ogness * We abuse smp_mb__after_clear_bit() because 1912bf06200eSJohn Ogness * there is no smp_mb__after_set_bit() yet 1913bf06200eSJohn Ogness */ 1914bf06200eSJohn Ogness smp_mb__after_clear_bit(); 1915bf06200eSJohn Ogness if (atomic_read(&sk->sk_wmem_alloc) > limit) 191646d3ceabSEric Dumazet break; 191746d3ceabSEric Dumazet } 1918c9eeec26SEric Dumazet 1919c8ac3774SHerbert Xu limit = mss_now; 1920f8269a49SIlpo Järvinen if (tso_segs > 1 && !tcp_urg_mode(tp)) 19210e3a4803SIlpo Järvinen limit = tcp_mss_split_point(sk, skb, mss_now, 19221485348dSBen Hutchings min_t(unsigned int, 19231485348dSBen Hutchings cwnd_quota, 1924d4589926SEric Dumazet sk->sk_gso_max_segs), 1925d4589926SEric Dumazet nonagle); 1926c8ac3774SHerbert Xu 1927c8ac3774SHerbert Xu if (skb->len > limit && 1928c4ead4c5SEric Dumazet unlikely(tso_fragment(sk, skb, limit, mss_now, gfp))) 19291da177e4SLinus Torvalds break; 19301da177e4SLinus Torvalds 19311da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 1932c1b4a7e6SDavid S. Miller 1933d5dd9175SIlpo Järvinen if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) 19341da177e4SLinus Torvalds break; 19351da177e4SLinus Torvalds 1936ec342325SAndrew Vagin repair: 19371da177e4SLinus Torvalds /* Advance the send_head. This one is sent out. 19381da177e4SLinus Torvalds * This call will increment packets_out. 19391da177e4SLinus Torvalds */ 194066f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 19411da177e4SLinus Torvalds 19421da177e4SLinus Torvalds tcp_minshall_update(tp, mss_now, skb); 1943a262f0cdSNandita Dukkipati sent_pkts += tcp_skb_pcount(skb); 1944d5dd9175SIlpo Järvinen 1945d5dd9175SIlpo Järvinen if (push_one) 1946d5dd9175SIlpo Järvinen break; 19471da177e4SLinus Torvalds } 19481da177e4SLinus Torvalds 1949aa93466bSDavid S. Miller if (likely(sent_pkts)) { 1950684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 1951684bad11SYuchung Cheng tp->prr_out += sent_pkts; 19526ba8a3b1SNandita Dukkipati 19536ba8a3b1SNandita Dukkipati /* Send one loss probe per tail loss episode. */ 19546ba8a3b1SNandita Dukkipati if (push_one != 2) 19556ba8a3b1SNandita Dukkipati tcp_schedule_loss_probe(sk); 19569e412ba7SIlpo Järvinen tcp_cwnd_validate(sk); 1957a2a385d6SEric Dumazet return false; 19581da177e4SLinus Torvalds } 19596ba8a3b1SNandita Dukkipati return (push_one == 2) || (!tp->packets_out && tcp_send_head(sk)); 19606ba8a3b1SNandita Dukkipati } 19616ba8a3b1SNandita Dukkipati 19626ba8a3b1SNandita Dukkipati bool tcp_schedule_loss_probe(struct sock *sk) 19636ba8a3b1SNandita Dukkipati { 19646ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 19656ba8a3b1SNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 19666ba8a3b1SNandita Dukkipati u32 timeout, tlp_time_stamp, rto_time_stamp; 19676ba8a3b1SNandita Dukkipati u32 rtt = tp->srtt >> 3; 19686ba8a3b1SNandita Dukkipati 19696ba8a3b1SNandita Dukkipati if (WARN_ON(icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS)) 19706ba8a3b1SNandita Dukkipati return false; 19716ba8a3b1SNandita Dukkipati /* No consecutive loss probes. */ 19726ba8a3b1SNandita Dukkipati if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) { 19736ba8a3b1SNandita Dukkipati tcp_rearm_rto(sk); 19746ba8a3b1SNandita Dukkipati return false; 19756ba8a3b1SNandita Dukkipati } 19766ba8a3b1SNandita Dukkipati /* Don't do any loss probe on a Fast Open connection before 3WHS 19776ba8a3b1SNandita Dukkipati * finishes. 19786ba8a3b1SNandita Dukkipati */ 19796ba8a3b1SNandita Dukkipati if (sk->sk_state == TCP_SYN_RECV) 19806ba8a3b1SNandita Dukkipati return false; 19816ba8a3b1SNandita Dukkipati 19826ba8a3b1SNandita Dukkipati /* TLP is only scheduled when next timer event is RTO. */ 19836ba8a3b1SNandita Dukkipati if (icsk->icsk_pending != ICSK_TIME_RETRANS) 19846ba8a3b1SNandita Dukkipati return false; 19856ba8a3b1SNandita Dukkipati 19866ba8a3b1SNandita Dukkipati /* Schedule a loss probe in 2*RTT for SACK capable connections 19876ba8a3b1SNandita Dukkipati * in Open state, that are either limited by cwnd or application. 19886ba8a3b1SNandita Dukkipati */ 19894a5ab4e2SEric Dumazet if (sysctl_tcp_early_retrans < 3 || !tp->srtt || !tp->packets_out || 19906ba8a3b1SNandita Dukkipati !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open) 19916ba8a3b1SNandita Dukkipati return false; 19926ba8a3b1SNandita Dukkipati 19936ba8a3b1SNandita Dukkipati if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) && 19946ba8a3b1SNandita Dukkipati tcp_send_head(sk)) 19956ba8a3b1SNandita Dukkipati return false; 19966ba8a3b1SNandita Dukkipati 19976ba8a3b1SNandita Dukkipati /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account 19986ba8a3b1SNandita Dukkipati * for delayed ack when there's one outstanding packet. 19996ba8a3b1SNandita Dukkipati */ 20006ba8a3b1SNandita Dukkipati timeout = rtt << 1; 20016ba8a3b1SNandita Dukkipati if (tp->packets_out == 1) 20026ba8a3b1SNandita Dukkipati timeout = max_t(u32, timeout, 20036ba8a3b1SNandita Dukkipati (rtt + (rtt >> 1) + TCP_DELACK_MAX)); 20046ba8a3b1SNandita Dukkipati timeout = max_t(u32, timeout, msecs_to_jiffies(10)); 20056ba8a3b1SNandita Dukkipati 20066ba8a3b1SNandita Dukkipati /* If RTO is shorter, just schedule TLP in its place. */ 20076ba8a3b1SNandita Dukkipati tlp_time_stamp = tcp_time_stamp + timeout; 20086ba8a3b1SNandita Dukkipati rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout; 20096ba8a3b1SNandita Dukkipati if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) { 20106ba8a3b1SNandita Dukkipati s32 delta = rto_time_stamp - tcp_time_stamp; 20116ba8a3b1SNandita Dukkipati if (delta > 0) 20126ba8a3b1SNandita Dukkipati timeout = delta; 20136ba8a3b1SNandita Dukkipati } 20146ba8a3b1SNandita Dukkipati 20156ba8a3b1SNandita Dukkipati inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, 20166ba8a3b1SNandita Dukkipati TCP_RTO_MAX); 20176ba8a3b1SNandita Dukkipati return true; 20186ba8a3b1SNandita Dukkipati } 20196ba8a3b1SNandita Dukkipati 20206ba8a3b1SNandita Dukkipati /* When probe timeout (PTO) fires, send a new segment if one exists, else 20216ba8a3b1SNandita Dukkipati * retransmit the last segment. 20226ba8a3b1SNandita Dukkipati */ 20236ba8a3b1SNandita Dukkipati void tcp_send_loss_probe(struct sock *sk) 20246ba8a3b1SNandita Dukkipati { 20259b717a8dSNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 20266ba8a3b1SNandita Dukkipati struct sk_buff *skb; 20276ba8a3b1SNandita Dukkipati int pcount; 20286ba8a3b1SNandita Dukkipati int mss = tcp_current_mss(sk); 20296ba8a3b1SNandita Dukkipati int err = -1; 20306ba8a3b1SNandita Dukkipati 20316ba8a3b1SNandita Dukkipati if (tcp_send_head(sk) != NULL) { 20326ba8a3b1SNandita Dukkipati err = tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC); 20336ba8a3b1SNandita Dukkipati goto rearm_timer; 20346ba8a3b1SNandita Dukkipati } 20356ba8a3b1SNandita Dukkipati 20369b717a8dSNandita Dukkipati /* At most one outstanding TLP retransmission. */ 20379b717a8dSNandita Dukkipati if (tp->tlp_high_seq) 20389b717a8dSNandita Dukkipati goto rearm_timer; 20399b717a8dSNandita Dukkipati 20406ba8a3b1SNandita Dukkipati /* Retransmit last segment. */ 20416ba8a3b1SNandita Dukkipati skb = tcp_write_queue_tail(sk); 20426ba8a3b1SNandita Dukkipati if (WARN_ON(!skb)) 20436ba8a3b1SNandita Dukkipati goto rearm_timer; 20446ba8a3b1SNandita Dukkipati 20456ba8a3b1SNandita Dukkipati pcount = tcp_skb_pcount(skb); 20466ba8a3b1SNandita Dukkipati if (WARN_ON(!pcount)) 20476ba8a3b1SNandita Dukkipati goto rearm_timer; 20486ba8a3b1SNandita Dukkipati 20496ba8a3b1SNandita Dukkipati if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) { 20506ba8a3b1SNandita Dukkipati if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss))) 20516ba8a3b1SNandita Dukkipati goto rearm_timer; 20526ba8a3b1SNandita Dukkipati skb = tcp_write_queue_tail(sk); 20536ba8a3b1SNandita Dukkipati } 20546ba8a3b1SNandita Dukkipati 20556ba8a3b1SNandita Dukkipati if (WARN_ON(!skb || !tcp_skb_pcount(skb))) 20566ba8a3b1SNandita Dukkipati goto rearm_timer; 20576ba8a3b1SNandita Dukkipati 20586ba8a3b1SNandita Dukkipati /* Probe with zero data doesn't trigger fast recovery. */ 20596ba8a3b1SNandita Dukkipati if (skb->len > 0) 20606ba8a3b1SNandita Dukkipati err = __tcp_retransmit_skb(sk, skb); 20616ba8a3b1SNandita Dukkipati 20629b717a8dSNandita Dukkipati /* Record snd_nxt for loss detection. */ 20639b717a8dSNandita Dukkipati if (likely(!err)) 20649b717a8dSNandita Dukkipati tp->tlp_high_seq = tp->snd_nxt; 20659b717a8dSNandita Dukkipati 20666ba8a3b1SNandita Dukkipati rearm_timer: 20676ba8a3b1SNandita Dukkipati inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 20686ba8a3b1SNandita Dukkipati inet_csk(sk)->icsk_rto, 20696ba8a3b1SNandita Dukkipati TCP_RTO_MAX); 20706ba8a3b1SNandita Dukkipati 20716ba8a3b1SNandita Dukkipati if (likely(!err)) 20726ba8a3b1SNandita Dukkipati NET_INC_STATS_BH(sock_net(sk), 20736ba8a3b1SNandita Dukkipati LINUX_MIB_TCPLOSSPROBES); 20746ba8a3b1SNandita Dukkipati return; 20751da177e4SLinus Torvalds } 20761da177e4SLinus Torvalds 2077a762a980SDavid S. Miller /* Push out any pending frames which were held back due to 2078a762a980SDavid S. Miller * TCP_CORK or attempt at coalescing tiny packets. 2079a762a980SDavid S. Miller * The socket must be locked by the caller. 2080a762a980SDavid S. Miller */ 20819e412ba7SIlpo Järvinen void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 20829e412ba7SIlpo Järvinen int nonagle) 2083a762a980SDavid S. Miller { 2084726e07a8SIlpo Järvinen /* If we are closed, the bytes will have to remain here. 2085726e07a8SIlpo Järvinen * In time closedown will finish, we empty the write queue and 2086726e07a8SIlpo Järvinen * all will be happy. 2087726e07a8SIlpo Järvinen */ 2088726e07a8SIlpo Järvinen if (unlikely(sk->sk_state == TCP_CLOSE)) 2089726e07a8SIlpo Järvinen return; 2090726e07a8SIlpo Järvinen 209199a1dec7SMel Gorman if (tcp_write_xmit(sk, cur_mss, nonagle, 0, 209299a1dec7SMel Gorman sk_gfp_atomic(sk, GFP_ATOMIC))) 20939e412ba7SIlpo Järvinen tcp_check_probe_timer(sk); 2094a762a980SDavid S. Miller } 2095a762a980SDavid S. Miller 2096c1b4a7e6SDavid S. Miller /* Send _single_ skb sitting at the send head. This function requires 2097c1b4a7e6SDavid S. Miller * true push pending frames to setup probe timer etc. 2098c1b4a7e6SDavid S. Miller */ 2099c1b4a7e6SDavid S. Miller void tcp_push_one(struct sock *sk, unsigned int mss_now) 2100c1b4a7e6SDavid S. Miller { 2101fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 2102c1b4a7e6SDavid S. Miller 2103c1b4a7e6SDavid S. Miller BUG_ON(!skb || skb->len < mss_now); 2104c1b4a7e6SDavid S. Miller 2105d5dd9175SIlpo Järvinen tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); 2106c1b4a7e6SDavid S. Miller } 2107c1b4a7e6SDavid S. Miller 21081da177e4SLinus Torvalds /* This function returns the amount that we can raise the 21091da177e4SLinus Torvalds * usable window based on the following constraints 21101da177e4SLinus Torvalds * 21111da177e4SLinus Torvalds * 1. The window can never be shrunk once it is offered (RFC 793) 21121da177e4SLinus Torvalds * 2. We limit memory per socket 21131da177e4SLinus Torvalds * 21141da177e4SLinus Torvalds * RFC 1122: 21151da177e4SLinus Torvalds * "the suggested [SWS] avoidance algorithm for the receiver is to keep 21161da177e4SLinus Torvalds * RECV.NEXT + RCV.WIN fixed until: 21171da177e4SLinus Torvalds * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 21181da177e4SLinus Torvalds * 21191da177e4SLinus Torvalds * i.e. don't raise the right edge of the window until you can raise 21201da177e4SLinus Torvalds * it at least MSS bytes. 21211da177e4SLinus Torvalds * 21221da177e4SLinus Torvalds * Unfortunately, the recommended algorithm breaks header prediction, 21231da177e4SLinus Torvalds * since header prediction assumes th->window stays fixed. 21241da177e4SLinus Torvalds * 21251da177e4SLinus Torvalds * Strictly speaking, keeping th->window fixed violates the receiver 21261da177e4SLinus Torvalds * side SWS prevention criteria. The problem is that under this rule 21271da177e4SLinus Torvalds * a stream of single byte packets will cause the right side of the 21281da177e4SLinus Torvalds * window to always advance by a single byte. 21291da177e4SLinus Torvalds * 21301da177e4SLinus Torvalds * Of course, if the sender implements sender side SWS prevention 21311da177e4SLinus Torvalds * then this will not be a problem. 21321da177e4SLinus Torvalds * 21331da177e4SLinus Torvalds * BSD seems to make the following compromise: 21341da177e4SLinus Torvalds * 21351da177e4SLinus Torvalds * If the free space is less than the 1/4 of the maximum 21361da177e4SLinus Torvalds * space available and the free space is less than 1/2 mss, 21371da177e4SLinus Torvalds * then set the window to 0. 21381da177e4SLinus Torvalds * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 21391da177e4SLinus Torvalds * Otherwise, just prevent the window from shrinking 21401da177e4SLinus Torvalds * and from being larger than the largest representable value. 21411da177e4SLinus Torvalds * 21421da177e4SLinus Torvalds * This prevents incremental opening of the window in the regime 21431da177e4SLinus Torvalds * where TCP is limited by the speed of the reader side taking 21441da177e4SLinus Torvalds * data out of the TCP receive queue. It does nothing about 21451da177e4SLinus Torvalds * those cases where the window is constrained on the sender side 21461da177e4SLinus Torvalds * because the pipeline is full. 21471da177e4SLinus Torvalds * 21481da177e4SLinus Torvalds * BSD also seems to "accidentally" limit itself to windows that are a 21491da177e4SLinus Torvalds * multiple of MSS, at least until the free space gets quite small. 21501da177e4SLinus Torvalds * This would appear to be a side effect of the mbuf implementation. 21511da177e4SLinus Torvalds * Combining these two algorithms results in the observed behavior 21521da177e4SLinus Torvalds * of having a fixed window size at almost all times. 21531da177e4SLinus Torvalds * 21541da177e4SLinus Torvalds * Below we obtain similar behavior by forcing the offered window to 21551da177e4SLinus Torvalds * a multiple of the mss when it is feasible to do so. 21561da177e4SLinus Torvalds * 21571da177e4SLinus Torvalds * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 21581da177e4SLinus Torvalds * Regular options like TIMESTAMP are taken into account. 21591da177e4SLinus Torvalds */ 21601da177e4SLinus Torvalds u32 __tcp_select_window(struct sock *sk) 21611da177e4SLinus Torvalds { 2162463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 21631da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2164caa20d9aSStephen Hemminger /* MSS for the peer's data. Previous versions used mss_clamp 21651da177e4SLinus Torvalds * here. I don't know if the value based on our guesses 21661da177e4SLinus Torvalds * of peer's MSS is better for the performance. It's more correct 21671da177e4SLinus Torvalds * but may be worse for the performance because of rcv_mss 21681da177e4SLinus Torvalds * fluctuations. --SAW 1998/11/1 21691da177e4SLinus Torvalds */ 2170463c84b9SArnaldo Carvalho de Melo int mss = icsk->icsk_ack.rcv_mss; 21711da177e4SLinus Torvalds int free_space = tcp_space(sk); 21721da177e4SLinus Torvalds int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk)); 21731da177e4SLinus Torvalds int window; 21741da177e4SLinus Torvalds 21751da177e4SLinus Torvalds if (mss > full_space) 21761da177e4SLinus Torvalds mss = full_space; 21771da177e4SLinus Torvalds 2178b92edbe0SEric Dumazet if (free_space < (full_space >> 1)) { 2179463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.quick = 0; 21801da177e4SLinus Torvalds 2181180d8cd9SGlauber Costa if (sk_under_memory_pressure(sk)) 2182056834d9SIlpo Järvinen tp->rcv_ssthresh = min(tp->rcv_ssthresh, 2183056834d9SIlpo Järvinen 4U * tp->advmss); 21841da177e4SLinus Torvalds 21851da177e4SLinus Torvalds if (free_space < mss) 21861da177e4SLinus Torvalds return 0; 21871da177e4SLinus Torvalds } 21881da177e4SLinus Torvalds 21891da177e4SLinus Torvalds if (free_space > tp->rcv_ssthresh) 21901da177e4SLinus Torvalds free_space = tp->rcv_ssthresh; 21911da177e4SLinus Torvalds 21921da177e4SLinus Torvalds /* Don't do rounding if we are using window scaling, since the 21931da177e4SLinus Torvalds * scaled window will not line up with the MSS boundary anyway. 21941da177e4SLinus Torvalds */ 21951da177e4SLinus Torvalds window = tp->rcv_wnd; 21961da177e4SLinus Torvalds if (tp->rx_opt.rcv_wscale) { 21971da177e4SLinus Torvalds window = free_space; 21981da177e4SLinus Torvalds 21991da177e4SLinus Torvalds /* Advertise enough space so that it won't get scaled away. 22001da177e4SLinus Torvalds * Import case: prevent zero window announcement if 22011da177e4SLinus Torvalds * 1<<rcv_wscale > mss. 22021da177e4SLinus Torvalds */ 22031da177e4SLinus Torvalds if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window) 22041da177e4SLinus Torvalds window = (((window >> tp->rx_opt.rcv_wscale) + 1) 22051da177e4SLinus Torvalds << tp->rx_opt.rcv_wscale); 22061da177e4SLinus Torvalds } else { 22071da177e4SLinus Torvalds /* Get the largest window that is a nice multiple of mss. 22081da177e4SLinus Torvalds * Window clamp already applied above. 22091da177e4SLinus Torvalds * If our current window offering is within 1 mss of the 22101da177e4SLinus Torvalds * free space we just keep it. This prevents the divide 22111da177e4SLinus Torvalds * and multiply from happening most of the time. 22121da177e4SLinus Torvalds * We also don't do any window rounding when the free space 22131da177e4SLinus Torvalds * is too small. 22141da177e4SLinus Torvalds */ 22151da177e4SLinus Torvalds if (window <= free_space - mss || window > free_space) 22161da177e4SLinus Torvalds window = (free_space / mss) * mss; 221784565070SJohn Heffner else if (mss == full_space && 2218b92edbe0SEric Dumazet free_space > window + (full_space >> 1)) 221984565070SJohn Heffner window = free_space; 22201da177e4SLinus Torvalds } 22211da177e4SLinus Torvalds 22221da177e4SLinus Torvalds return window; 22231da177e4SLinus Torvalds } 22241da177e4SLinus Torvalds 22254a17fc3aSIlpo Järvinen /* Collapses two adjacent SKB's during retransmission. */ 22264a17fc3aSIlpo Järvinen static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) 22271da177e4SLinus Torvalds { 22281da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2229fe067e8aSDavid S. Miller struct sk_buff *next_skb = tcp_write_queue_next(sk, skb); 2230058dc334SIlpo Järvinen int skb_size, next_skb_size; 22311da177e4SLinus Torvalds 2232058dc334SIlpo Järvinen skb_size = skb->len; 2233058dc334SIlpo Järvinen next_skb_size = next_skb->len; 22341da177e4SLinus Torvalds 2235058dc334SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); 22361da177e4SLinus Torvalds 22376859d494SIlpo Järvinen tcp_highest_sack_combine(sk, next_skb, skb); 2238a6963a6bSIlpo Järvinen 2239fe067e8aSDavid S. Miller tcp_unlink_write_queue(next_skb, sk); 22401da177e4SLinus Torvalds 2241058dc334SIlpo Järvinen skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size), 22421a4e2d09SArnaldo Carvalho de Melo next_skb_size); 22431da177e4SLinus Torvalds 224452d570aaSJarek Poplawski if (next_skb->ip_summed == CHECKSUM_PARTIAL) 224552d570aaSJarek Poplawski skb->ip_summed = CHECKSUM_PARTIAL; 22461da177e4SLinus Torvalds 224784fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 22481da177e4SLinus Torvalds skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); 22491da177e4SLinus Torvalds 22501da177e4SLinus Torvalds /* Update sequence range on original skb. */ 22511da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 22521da177e4SLinus Torvalds 2253e6c7d085SIlpo Järvinen /* Merge over control information. This moves PSH/FIN etc. over */ 22544de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; 22551da177e4SLinus Torvalds 22561da177e4SLinus Torvalds /* All done, get rid of second SKB and account for it so 22571da177e4SLinus Torvalds * packet counting does not break. 22581da177e4SLinus Torvalds */ 22594828e7f4SIlpo Järvinen TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; 2260b7689205SIlpo Järvinen 2261b7689205SIlpo Järvinen /* changed transmit queue under us so clear hints */ 2262ef9da47cSIlpo Järvinen tcp_clear_retrans_hints_partial(tp); 2263ef9da47cSIlpo Järvinen if (next_skb == tp->retransmit_skb_hint) 2264ef9da47cSIlpo Järvinen tp->retransmit_skb_hint = skb; 2265b7689205SIlpo Järvinen 2266797108d1SIlpo Järvinen tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); 2267797108d1SIlpo Järvinen 22683ab224beSHideo Aoki sk_wmem_free_skb(sk, next_skb); 22691da177e4SLinus Torvalds } 22701da177e4SLinus Torvalds 227167edfef7SAndi Kleen /* Check if coalescing SKBs is legal. */ 2272a2a385d6SEric Dumazet static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) 22734a17fc3aSIlpo Järvinen { 22744a17fc3aSIlpo Järvinen if (tcp_skb_pcount(skb) > 1) 2275a2a385d6SEric Dumazet return false; 22764a17fc3aSIlpo Järvinen /* TODO: SACK collapsing could be used to remove this condition */ 22774a17fc3aSIlpo Järvinen if (skb_shinfo(skb)->nr_frags != 0) 2278a2a385d6SEric Dumazet return false; 22794a17fc3aSIlpo Järvinen if (skb_cloned(skb)) 2280a2a385d6SEric Dumazet return false; 22814a17fc3aSIlpo Järvinen if (skb == tcp_send_head(sk)) 2282a2a385d6SEric Dumazet return false; 22834a17fc3aSIlpo Järvinen /* Some heurestics for collapsing over SACK'd could be invented */ 22844a17fc3aSIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 2285a2a385d6SEric Dumazet return false; 22864a17fc3aSIlpo Järvinen 2287a2a385d6SEric Dumazet return true; 22884a17fc3aSIlpo Järvinen } 22894a17fc3aSIlpo Järvinen 229067edfef7SAndi Kleen /* Collapse packets in the retransmit queue to make to create 229167edfef7SAndi Kleen * less packets on the wire. This is only done on retransmission. 229267edfef7SAndi Kleen */ 22934a17fc3aSIlpo Järvinen static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, 22944a17fc3aSIlpo Järvinen int space) 22954a17fc3aSIlpo Järvinen { 22964a17fc3aSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 22974a17fc3aSIlpo Järvinen struct sk_buff *skb = to, *tmp; 2298a2a385d6SEric Dumazet bool first = true; 22994a17fc3aSIlpo Järvinen 23004a17fc3aSIlpo Järvinen if (!sysctl_tcp_retrans_collapse) 23014a17fc3aSIlpo Järvinen return; 23024de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 23034a17fc3aSIlpo Järvinen return; 23044a17fc3aSIlpo Järvinen 23054a17fc3aSIlpo Järvinen tcp_for_write_queue_from_safe(skb, tmp, sk) { 23064a17fc3aSIlpo Järvinen if (!tcp_can_collapse(sk, skb)) 23074a17fc3aSIlpo Järvinen break; 23084a17fc3aSIlpo Järvinen 23094a17fc3aSIlpo Järvinen space -= skb->len; 23104a17fc3aSIlpo Järvinen 23114a17fc3aSIlpo Järvinen if (first) { 2312a2a385d6SEric Dumazet first = false; 23134a17fc3aSIlpo Järvinen continue; 23144a17fc3aSIlpo Järvinen } 23154a17fc3aSIlpo Järvinen 23164a17fc3aSIlpo Järvinen if (space < 0) 23174a17fc3aSIlpo Järvinen break; 23184a17fc3aSIlpo Järvinen /* Punt if not enough space exists in the first SKB for 23194a17fc3aSIlpo Järvinen * the data in the second 23204a17fc3aSIlpo Järvinen */ 2321a21d4572SEric Dumazet if (skb->len > skb_availroom(to)) 23224a17fc3aSIlpo Järvinen break; 23234a17fc3aSIlpo Järvinen 23244a17fc3aSIlpo Järvinen if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) 23254a17fc3aSIlpo Järvinen break; 23264a17fc3aSIlpo Järvinen 23274a17fc3aSIlpo Järvinen tcp_collapse_retrans(sk, to); 23284a17fc3aSIlpo Järvinen } 23294a17fc3aSIlpo Järvinen } 23304a17fc3aSIlpo Järvinen 23311da177e4SLinus Torvalds /* This retransmits one SKB. Policy decisions and retransmit queue 23321da177e4SLinus Torvalds * state updates are done by the caller. Returns non-zero if an 23331da177e4SLinus Torvalds * error occurred which prevented the send. 23341da177e4SLinus Torvalds */ 233593b174adSYuchung Cheng int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) 23361da177e4SLinus Torvalds { 23371da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 23385d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 23397d227cd2SSridhar Samudrala unsigned int cur_mss; 2340*c84a5711SYuchung Cheng int err; 23411da177e4SLinus Torvalds 23425d424d5aSJohn Heffner /* Inconslusive MTU probe */ 23435d424d5aSJohn Heffner if (icsk->icsk_mtup.probe_size) { 23445d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 23455d424d5aSJohn Heffner } 23465d424d5aSJohn Heffner 23471da177e4SLinus Torvalds /* Do not sent more than we queued. 1/4 is reserved for possible 2348caa20d9aSStephen Hemminger * copying overhead: fragmentation, tunneling, mangling etc. 23491da177e4SLinus Torvalds */ 23501da177e4SLinus Torvalds if (atomic_read(&sk->sk_wmem_alloc) > 23511da177e4SLinus Torvalds min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf)) 23521da177e4SLinus Torvalds return -EAGAIN; 23531da177e4SLinus Torvalds 23541da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 23551da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 23561da177e4SLinus Torvalds BUG(); 23571da177e4SLinus Torvalds if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 23581da177e4SLinus Torvalds return -ENOMEM; 23591da177e4SLinus Torvalds } 23601da177e4SLinus Torvalds 23617d227cd2SSridhar Samudrala if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 23627d227cd2SSridhar Samudrala return -EHOSTUNREACH; /* Routing failure or similar. */ 23637d227cd2SSridhar Samudrala 23640c54b85fSIlpo Järvinen cur_mss = tcp_current_mss(sk); 23657d227cd2SSridhar Samudrala 23661da177e4SLinus Torvalds /* If receiver has shrunk his window, and skb is out of 23671da177e4SLinus Torvalds * new window, do not retransmit it. The exception is the 23681da177e4SLinus Torvalds * case, when window is shrunk to zero. In this case 23691da177e4SLinus Torvalds * our retransmit serves as a zero window probe. 23701da177e4SLinus Torvalds */ 23719d4fb27dSJoe Perches if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && 23729d4fb27dSJoe Perches TCP_SKB_CB(skb)->seq != tp->snd_una) 23731da177e4SLinus Torvalds return -EAGAIN; 23741da177e4SLinus Torvalds 23751da177e4SLinus Torvalds if (skb->len > cur_mss) { 2376846998aeSDavid S. Miller if (tcp_fragment(sk, skb, cur_mss, cur_mss)) 23771da177e4SLinus Torvalds return -ENOMEM; /* We'll try again later. */ 237802276f3cSIlpo Järvinen } else { 23799eb9362eSIlpo Järvinen int oldpcount = tcp_skb_pcount(skb); 23809eb9362eSIlpo Järvinen 23819eb9362eSIlpo Järvinen if (unlikely(oldpcount > 1)) { 2382c52e2421SEric Dumazet if (skb_unclone(skb, GFP_ATOMIC)) 2383c52e2421SEric Dumazet return -ENOMEM; 238402276f3cSIlpo Järvinen tcp_init_tso_segs(sk, skb, cur_mss); 23859eb9362eSIlpo Järvinen tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb)); 23869eb9362eSIlpo Järvinen } 23871da177e4SLinus Torvalds } 23881da177e4SLinus Torvalds 23891da177e4SLinus Torvalds tcp_retrans_try_collapse(sk, skb, cur_mss); 23901da177e4SLinus Torvalds 23911da177e4SLinus Torvalds /* Make a copy, if the first transmission SKB clone we made 23921da177e4SLinus Torvalds * is still in somebody's hands, else make a clone. 23931da177e4SLinus Torvalds */ 23941da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 23951da177e4SLinus Torvalds 239650bceae9SThomas Graf /* make sure skb->data is aligned on arches that require it 239750bceae9SThomas Graf * and check if ack-trimming & collapsing extended the headroom 239850bceae9SThomas Graf * beyond what csum_start can cover. 239950bceae9SThomas Graf */ 240050bceae9SThomas Graf if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) || 240150bceae9SThomas Graf skb_headroom(skb) >= 0xFFFF)) { 2402117632e6SEric Dumazet struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER, 2403117632e6SEric Dumazet GFP_ATOMIC); 2404*c84a5711SYuchung Cheng err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) : 2405117632e6SEric Dumazet -ENOBUFS; 2406117632e6SEric Dumazet } else { 2407*c84a5711SYuchung Cheng err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 2408117632e6SEric Dumazet } 2409*c84a5711SYuchung Cheng 2410*c84a5711SYuchung Cheng if (likely(!err)) 2411*c84a5711SYuchung Cheng TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS; 2412*c84a5711SYuchung Cheng return err; 241393b174adSYuchung Cheng } 241493b174adSYuchung Cheng 241593b174adSYuchung Cheng int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) 241693b174adSYuchung Cheng { 241793b174adSYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 241893b174adSYuchung Cheng int err = __tcp_retransmit_skb(sk, skb); 24191da177e4SLinus Torvalds 24201da177e4SLinus Torvalds if (err == 0) { 24211da177e4SLinus Torvalds /* Update global TCP statistics. */ 242281cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); 24231da177e4SLinus Torvalds 24241da177e4SLinus Torvalds tp->total_retrans++; 24251da177e4SLinus Torvalds 24261da177e4SLinus Torvalds #if FASTRETRANS_DEBUG > 0 24271da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 2428e87cc472SJoe Perches net_dbg_ratelimited("retrans_out leaked\n"); 24291da177e4SLinus Torvalds } 24301da177e4SLinus Torvalds #endif 2431b08d6cb2SIlpo Järvinen if (!tp->retrans_out) 2432b08d6cb2SIlpo Järvinen tp->lost_retrans_low = tp->snd_nxt; 24331da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 24341da177e4SLinus Torvalds tp->retrans_out += tcp_skb_pcount(skb); 24351da177e4SLinus Torvalds 24361da177e4SLinus Torvalds /* Save stamp of the first retransmit. */ 24371da177e4SLinus Torvalds if (!tp->retrans_stamp) 24381da177e4SLinus Torvalds tp->retrans_stamp = TCP_SKB_CB(skb)->when; 24391da177e4SLinus Torvalds 2440c24f691bSYuchung Cheng tp->undo_retrans += tcp_skb_pcount(skb); 24411da177e4SLinus Torvalds 24421da177e4SLinus Torvalds /* snd_nxt is stored to detect loss of retransmitted segment, 24431da177e4SLinus Torvalds * see tcp_input.c tcp_sacktag_write_queue(). 24441da177e4SLinus Torvalds */ 24451da177e4SLinus Torvalds TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt; 244624ab6becSYuchung Cheng } else { 244724ab6becSYuchung Cheng NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL); 24481da177e4SLinus Torvalds } 24491da177e4SLinus Torvalds return err; 24501da177e4SLinus Torvalds } 24511da177e4SLinus Torvalds 245267edfef7SAndi Kleen /* Check if we forward retransmits are possible in the current 245367edfef7SAndi Kleen * window/congestion state. 245467edfef7SAndi Kleen */ 2455a2a385d6SEric Dumazet static bool tcp_can_forward_retransmit(struct sock *sk) 2456b5afe7bcSIlpo Järvinen { 2457b5afe7bcSIlpo Järvinen const struct inet_connection_sock *icsk = inet_csk(sk); 2458cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 2459b5afe7bcSIlpo Järvinen 2460b5afe7bcSIlpo Järvinen /* Forward retransmissions are possible only during Recovery. */ 2461b5afe7bcSIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Recovery) 2462a2a385d6SEric Dumazet return false; 2463b5afe7bcSIlpo Järvinen 2464b5afe7bcSIlpo Järvinen /* No forward retransmissions in Reno are possible. */ 2465b5afe7bcSIlpo Järvinen if (tcp_is_reno(tp)) 2466a2a385d6SEric Dumazet return false; 2467b5afe7bcSIlpo Järvinen 2468b5afe7bcSIlpo Järvinen /* Yeah, we have to make difficult choice between forward transmission 2469b5afe7bcSIlpo Järvinen * and retransmission... Both ways have their merits... 2470b5afe7bcSIlpo Järvinen * 2471b5afe7bcSIlpo Järvinen * For now we do not retransmit anything, while we have some new 2472b5afe7bcSIlpo Järvinen * segments to send. In the other cases, follow rule 3 for 2473b5afe7bcSIlpo Järvinen * NextSeg() specified in RFC3517. 2474b5afe7bcSIlpo Järvinen */ 2475b5afe7bcSIlpo Järvinen 2476b5afe7bcSIlpo Järvinen if (tcp_may_send_now(sk)) 2477a2a385d6SEric Dumazet return false; 2478b5afe7bcSIlpo Järvinen 2479a2a385d6SEric Dumazet return true; 2480b5afe7bcSIlpo Järvinen } 2481b5afe7bcSIlpo Järvinen 24821da177e4SLinus Torvalds /* This gets called after a retransmit timeout, and the initially 24831da177e4SLinus Torvalds * retransmitted data is acknowledged. It tries to continue 24841da177e4SLinus Torvalds * resending the rest of the retransmit queue, until either 24851da177e4SLinus Torvalds * we've sent it all or the congestion window limit is reached. 24861da177e4SLinus Torvalds * If doing SACK, the first ACK which comes back for a timeout 24871da177e4SLinus Torvalds * based retransmit packet might feed us FACK information again. 24881da177e4SLinus Torvalds * If so, we use it to avoid unnecessarily retransmissions. 24891da177e4SLinus Torvalds */ 24901da177e4SLinus Torvalds void tcp_xmit_retransmit_queue(struct sock *sk) 24911da177e4SLinus Torvalds { 24926687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 24931da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 24941da177e4SLinus Torvalds struct sk_buff *skb; 24950e1c54c2SIlpo Järvinen struct sk_buff *hole = NULL; 2496618d9f25SIlpo Järvinen u32 last_lost; 249761eb55f4SIlpo Järvinen int mib_idx; 24980e1c54c2SIlpo Järvinen int fwd_rexmitting = 0; 24996a438bbeSStephen Hemminger 250045e77d31SIlpo Järvinen if (!tp->packets_out) 250145e77d31SIlpo Järvinen return; 250245e77d31SIlpo Järvinen 250308ebd172SIlpo Järvinen if (!tp->lost_out) 250408ebd172SIlpo Järvinen tp->retransmit_high = tp->snd_una; 250508ebd172SIlpo Järvinen 2506618d9f25SIlpo Järvinen if (tp->retransmit_skb_hint) { 25076a438bbeSStephen Hemminger skb = tp->retransmit_skb_hint; 2508618d9f25SIlpo Järvinen last_lost = TCP_SKB_CB(skb)->end_seq; 2509618d9f25SIlpo Järvinen if (after(last_lost, tp->retransmit_high)) 2510618d9f25SIlpo Järvinen last_lost = tp->retransmit_high; 2511618d9f25SIlpo Järvinen } else { 2512fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 2513618d9f25SIlpo Järvinen last_lost = tp->snd_una; 2514618d9f25SIlpo Järvinen } 25151da177e4SLinus Torvalds 2516fe067e8aSDavid S. Miller tcp_for_write_queue_from(skb, sk) { 25171da177e4SLinus Torvalds __u8 sacked = TCP_SKB_CB(skb)->sacked; 25181da177e4SLinus Torvalds 2519fe067e8aSDavid S. Miller if (skb == tcp_send_head(sk)) 2520fe067e8aSDavid S. Miller break; 25216a438bbeSStephen Hemminger /* we could do better than to assign each time */ 25220e1c54c2SIlpo Järvinen if (hole == NULL) 25236a438bbeSStephen Hemminger tp->retransmit_skb_hint = skb; 25246a438bbeSStephen Hemminger 25251da177e4SLinus Torvalds /* Assume this retransmit will generate 25261da177e4SLinus Torvalds * only one packet for congestion window 25271da177e4SLinus Torvalds * calculation purposes. This works because 25281da177e4SLinus Torvalds * tcp_retransmit_skb() will chop up the 25291da177e4SLinus Torvalds * packet to be MSS sized and all the 25301da177e4SLinus Torvalds * packet counting works out. 25311da177e4SLinus Torvalds */ 25321da177e4SLinus Torvalds if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) 25331da177e4SLinus Torvalds return; 25340e1c54c2SIlpo Järvinen 25350e1c54c2SIlpo Järvinen if (fwd_rexmitting) { 25360e1c54c2SIlpo Järvinen begin_fwd: 25370e1c54c2SIlpo Järvinen if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp))) 2538006f582cSIlpo Järvinen break; 25390e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFORWARDRETRANS; 25400e1c54c2SIlpo Järvinen 25410e1c54c2SIlpo Järvinen } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) { 2542618d9f25SIlpo Järvinen tp->retransmit_high = last_lost; 25430e1c54c2SIlpo Järvinen if (!tcp_can_forward_retransmit(sk)) 25440e1c54c2SIlpo Järvinen break; 25450e1c54c2SIlpo Järvinen /* Backtrack if necessary to non-L'ed skb */ 25460e1c54c2SIlpo Järvinen if (hole != NULL) { 25470e1c54c2SIlpo Järvinen skb = hole; 25480e1c54c2SIlpo Järvinen hole = NULL; 25490e1c54c2SIlpo Järvinen } 25500e1c54c2SIlpo Järvinen fwd_rexmitting = 1; 25510e1c54c2SIlpo Järvinen goto begin_fwd; 25520e1c54c2SIlpo Järvinen 25530e1c54c2SIlpo Järvinen } else if (!(sacked & TCPCB_LOST)) { 2554ac11ba75SIlpo Järvinen if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) 25550e1c54c2SIlpo Järvinen hole = skb; 255661eb55f4SIlpo Järvinen continue; 25571da177e4SLinus Torvalds 25580e1c54c2SIlpo Järvinen } else { 2559618d9f25SIlpo Järvinen last_lost = TCP_SKB_CB(skb)->end_seq; 25600e1c54c2SIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Loss) 25610e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFASTRETRANS; 25620e1c54c2SIlpo Järvinen else 25630e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; 25640e1c54c2SIlpo Järvinen } 25650e1c54c2SIlpo Järvinen 25660e1c54c2SIlpo Järvinen if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) 256761eb55f4SIlpo Järvinen continue; 256840b215e5SPavel Emelyanov 256924ab6becSYuchung Cheng if (tcp_retransmit_skb(sk, skb)) 25701da177e4SLinus Torvalds return; 257124ab6becSYuchung Cheng 2572de0744afSPavel Emelyanov NET_INC_STATS_BH(sock_net(sk), mib_idx); 25731da177e4SLinus Torvalds 2574684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 2575a262f0cdSNandita Dukkipati tp->prr_out += tcp_skb_pcount(skb); 2576a262f0cdSNandita Dukkipati 2577fe067e8aSDavid S. Miller if (skb == tcp_write_queue_head(sk)) 2578463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 25793f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, 25803f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 25811da177e4SLinus Torvalds } 25821da177e4SLinus Torvalds } 25831da177e4SLinus Torvalds 25841da177e4SLinus Torvalds /* Send a fin. The caller locks the socket for us. This cannot be 25851da177e4SLinus Torvalds * allowed to fail queueing a FIN frame under any circumstances. 25861da177e4SLinus Torvalds */ 25871da177e4SLinus Torvalds void tcp_send_fin(struct sock *sk) 25881da177e4SLinus Torvalds { 25891da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2590fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_write_queue_tail(sk); 25911da177e4SLinus Torvalds int mss_now; 25921da177e4SLinus Torvalds 25931da177e4SLinus Torvalds /* Optimization, tack on the FIN if we have a queue of 25941da177e4SLinus Torvalds * unsent frames. But be careful about outgoing SACKS 25951da177e4SLinus Torvalds * and IP options. 25961da177e4SLinus Torvalds */ 25970c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 25981da177e4SLinus Torvalds 2599fe067e8aSDavid S. Miller if (tcp_send_head(sk) != NULL) { 26004de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_FIN; 26011da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq++; 26021da177e4SLinus Torvalds tp->write_seq++; 26031da177e4SLinus Torvalds } else { 26041da177e4SLinus Torvalds /* Socket is locked, keep trying until memory is available. */ 26051da177e4SLinus Torvalds for (;;) { 2606aa133076SWu Fengguang skb = alloc_skb_fclone(MAX_TCP_HEADER, 2607aa133076SWu Fengguang sk->sk_allocation); 26081da177e4SLinus Torvalds if (skb) 26091da177e4SLinus Torvalds break; 26101da177e4SLinus Torvalds yield(); 26111da177e4SLinus Torvalds } 26121da177e4SLinus Torvalds 26131da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 26141da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 26151da177e4SLinus Torvalds /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 2616e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tp->write_seq, 2617a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_FIN); 26181da177e4SLinus Torvalds tcp_queue_skb(sk, skb); 26191da177e4SLinus Torvalds } 26209e412ba7SIlpo Järvinen __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF); 26211da177e4SLinus Torvalds } 26221da177e4SLinus Torvalds 26231da177e4SLinus Torvalds /* We get here when a process closes a file descriptor (either due to 26241da177e4SLinus Torvalds * an explicit close() or as a byproduct of exit()'ing) and there 26251da177e4SLinus Torvalds * was unread data in the receive queue. This behavior is recommended 262665bb723cSGerrit Renker * by RFC 2525, section 2.17. -DaveM 26271da177e4SLinus Torvalds */ 2628dd0fc66fSAl Viro void tcp_send_active_reset(struct sock *sk, gfp_t priority) 26291da177e4SLinus Torvalds { 26301da177e4SLinus Torvalds struct sk_buff *skb; 26311da177e4SLinus Torvalds 26321da177e4SLinus Torvalds /* NOTE: No TCP options attached and we never retransmit this. */ 26331da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, priority); 26341da177e4SLinus Torvalds if (!skb) { 26354e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 26361da177e4SLinus Torvalds return; 26371da177e4SLinus Torvalds } 26381da177e4SLinus Torvalds 26391da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 26401da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 2641e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), 2642a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_RST); 26431da177e4SLinus Torvalds /* Send it off. */ 26441da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2645dfb4b9dcSDavid S. Miller if (tcp_transmit_skb(sk, skb, 0, priority)) 26464e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 264726af65cbSSridhar Samudrala 264881cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); 26491da177e4SLinus Torvalds } 26501da177e4SLinus Torvalds 265167edfef7SAndi Kleen /* Send a crossed SYN-ACK during socket establishment. 265267edfef7SAndi Kleen * WARNING: This routine must only be called when we have already sent 26531da177e4SLinus Torvalds * a SYN packet that crossed the incoming SYN that caused this routine 26541da177e4SLinus Torvalds * to get called. If this assumption fails then the initial rcv_wnd 26551da177e4SLinus Torvalds * and rcv_wscale values will not be correct. 26561da177e4SLinus Torvalds */ 26571da177e4SLinus Torvalds int tcp_send_synack(struct sock *sk) 26581da177e4SLinus Torvalds { 26591da177e4SLinus Torvalds struct sk_buff *skb; 26601da177e4SLinus Torvalds 2661fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 26624de075e0SEric Dumazet if (skb == NULL || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 266391df42beSJoe Perches pr_debug("%s: wrong queue state\n", __func__); 26641da177e4SLinus Torvalds return -EFAULT; 26651da177e4SLinus Torvalds } 26664de075e0SEric Dumazet if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { 26671da177e4SLinus Torvalds if (skb_cloned(skb)) { 26681da177e4SLinus Torvalds struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC); 26691da177e4SLinus Torvalds if (nskb == NULL) 26701da177e4SLinus Torvalds return -ENOMEM; 2671fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 26721da177e4SLinus Torvalds skb_header_release(nskb); 2673fe067e8aSDavid S. Miller __tcp_add_write_queue_head(sk, nskb); 26743ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 26753ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 26763ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 26771da177e4SLinus Torvalds skb = nskb; 26781da177e4SLinus Torvalds } 26791da177e4SLinus Torvalds 26804de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; 26811da177e4SLinus Torvalds TCP_ECN_send_synack(tcp_sk(sk), skb); 26821da177e4SLinus Torvalds } 26831da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2684dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 26851da177e4SLinus Torvalds } 26861da177e4SLinus Torvalds 26874aea39c1SEric Dumazet /** 26884aea39c1SEric Dumazet * tcp_make_synack - Prepare a SYN-ACK. 26894aea39c1SEric Dumazet * sk: listener socket 26904aea39c1SEric Dumazet * dst: dst entry attached to the SYNACK 26914aea39c1SEric Dumazet * req: request_sock pointer 26924aea39c1SEric Dumazet * 26934aea39c1SEric Dumazet * Allocate one skb and build a SYNACK packet. 26944aea39c1SEric Dumazet * @dst is consumed : Caller should not use it again. 26954aea39c1SEric Dumazet */ 26961da177e4SLinus Torvalds struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst, 2697e6b4d113SWilliam Allen Simpson struct request_sock *req, 26988336886fSJerry Chu struct tcp_fastopen_cookie *foc) 26991da177e4SLinus Torvalds { 2700bd0388aeSWilliam Allen Simpson struct tcp_out_options opts; 27012e6599cbSArnaldo Carvalho de Melo struct inet_request_sock *ireq = inet_rsk(req); 27021da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 27031da177e4SLinus Torvalds struct tcphdr *th; 27041da177e4SLinus Torvalds struct sk_buff *skb; 2705cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 2706bd0388aeSWilliam Allen Simpson int tcp_header_size; 2707f5fff5dcSTom Quetchenbach int mss; 27081da177e4SLinus Torvalds 2709eb8895deSPhil Oester skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC); 27104aea39c1SEric Dumazet if (unlikely(!skb)) { 27114aea39c1SEric Dumazet dst_release(dst); 27121da177e4SLinus Torvalds return NULL; 27134aea39c1SEric Dumazet } 27141da177e4SLinus Torvalds /* Reserve space for headers. */ 27151da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 27161da177e4SLinus Torvalds 27174aea39c1SEric Dumazet skb_dst_set(skb, dst); 2718ca10b9e9SEric Dumazet security_skb_owned_by(skb, sk); 27191da177e4SLinus Torvalds 27200dbaee3bSDavid S. Miller mss = dst_metric_advmss(dst); 2721f5fff5dcSTom Quetchenbach if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss) 2722f5fff5dcSTom Quetchenbach mss = tp->rx_opt.user_mss; 2723f5fff5dcSTom Quetchenbach 272433ad798cSAdam Langley if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */ 272533ad798cSAdam Langley __u8 rcv_wscale; 272633ad798cSAdam Langley /* Set this up on the first call only */ 272733ad798cSAdam Langley req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW); 2728e88c64f0SHagen Paul Pfeifer 2729e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 2730e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 2731e88c64f0SHagen Paul Pfeifer (req->window_clamp > tcp_full_space(sk) || req->window_clamp == 0)) 2732e88c64f0SHagen Paul Pfeifer req->window_clamp = tcp_full_space(sk); 2733e88c64f0SHagen Paul Pfeifer 273433ad798cSAdam Langley /* tcp_full_space because it is guaranteed to be the first packet */ 273533ad798cSAdam Langley tcp_select_initial_window(tcp_full_space(sk), 2736f5fff5dcSTom Quetchenbach mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0), 273733ad798cSAdam Langley &req->rcv_wnd, 273833ad798cSAdam Langley &req->window_clamp, 273933ad798cSAdam Langley ireq->wscale_ok, 274031d12926Slaurent chavey &rcv_wscale, 274131d12926Slaurent chavey dst_metric(dst, RTAX_INITRWND)); 274233ad798cSAdam Langley ireq->rcv_wscale = rcv_wscale; 274333ad798cSAdam Langley } 2744cfb6eeb4SYOSHIFUJI Hideaki 274533ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 27468b5f12d0SFlorian Westphal #ifdef CONFIG_SYN_COOKIES 27478b5f12d0SFlorian Westphal if (unlikely(req->cookie_ts)) 27488b5f12d0SFlorian Westphal TCP_SKB_CB(skb)->when = cookie_init_timestamp(req); 27498b5f12d0SFlorian Westphal else 27508b5f12d0SFlorian Westphal #endif 275133ad798cSAdam Langley TCP_SKB_CB(skb)->when = tcp_time_stamp; 27521a2c6181SChristoph Paasch tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, &md5, 27531a2c6181SChristoph Paasch foc) + sizeof(*th); 275433ad798cSAdam Langley 2755aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 2756aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 27571da177e4SLinus Torvalds 2758aa8223c7SArnaldo Carvalho de Melo th = tcp_hdr(skb); 27591da177e4SLinus Torvalds memset(th, 0, sizeof(struct tcphdr)); 27601da177e4SLinus Torvalds th->syn = 1; 27611da177e4SLinus Torvalds th->ack = 1; 27621da177e4SLinus Torvalds TCP_ECN_make_synack(req, th); 2763b44084c2SEric Dumazet th->source = htons(ireq->ir_num); 2764634fb979SEric Dumazet th->dest = ireq->ir_rmt_port; 2765e870a8efSIlpo Järvinen /* Setting of flags are superfluous here for callers (and ECE is 2766e870a8efSIlpo Järvinen * not even correctly set) 2767e870a8efSIlpo Järvinen */ 2768e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn, 2769a3433f35SChangli Gao TCPHDR_SYN | TCPHDR_ACK); 27704957faadSWilliam Allen Simpson 27711da177e4SLinus Torvalds th->seq = htonl(TCP_SKB_CB(skb)->seq); 27728336886fSJerry Chu /* XXX data is queued and acked as is. No buffer/window check */ 27738336886fSJerry Chu th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt); 27741da177e4SLinus Torvalds 27751da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 2776600ff0c2SIlpo Järvinen th->window = htons(min(req->rcv_wnd, 65535U)); 2777bd0388aeSWilliam Allen Simpson tcp_options_write((__be32 *)(th + 1), tp, &opts); 27781da177e4SLinus Torvalds th->doff = (tcp_header_size >> 2); 2779aa2ea058STom Herbert TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, tcp_skb_pcount(skb)); 2780cfb6eeb4SYOSHIFUJI Hideaki 2781cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 2782cfb6eeb4SYOSHIFUJI Hideaki /* Okay, we have all we need - do the md5 hash if needed */ 2783cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 2784bd0388aeSWilliam Allen Simpson tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, 278549a72dfbSAdam Langley md5, NULL, req, skb); 2786cfb6eeb4SYOSHIFUJI Hideaki } 2787cfb6eeb4SYOSHIFUJI Hideaki #endif 2788cfb6eeb4SYOSHIFUJI Hideaki 27891da177e4SLinus Torvalds return skb; 27901da177e4SLinus Torvalds } 27914bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_make_synack); 27921da177e4SLinus Torvalds 279367edfef7SAndi Kleen /* Do all connect socket setups that can be done AF independent. */ 2794f7e56a76Sstephen hemminger static void tcp_connect_init(struct sock *sk) 27951da177e4SLinus Torvalds { 2796cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 27971da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 27981da177e4SLinus Torvalds __u8 rcv_wscale; 27991da177e4SLinus Torvalds 28001da177e4SLinus Torvalds /* We'll fix this up when we get a response from the other end. 28011da177e4SLinus Torvalds * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 28021da177e4SLinus Torvalds */ 28031da177e4SLinus Torvalds tp->tcp_header_len = sizeof(struct tcphdr) + 2804bb5b7c11SDavid S. Miller (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0); 28051da177e4SLinus Torvalds 2806cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 2807cfb6eeb4SYOSHIFUJI Hideaki if (tp->af_specific->md5_lookup(sk, sk) != NULL) 2808cfb6eeb4SYOSHIFUJI Hideaki tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 2809cfb6eeb4SYOSHIFUJI Hideaki #endif 2810cfb6eeb4SYOSHIFUJI Hideaki 28111da177e4SLinus Torvalds /* If user gave his TCP_MAXSEG, record it to clamp */ 28121da177e4SLinus Torvalds if (tp->rx_opt.user_mss) 28131da177e4SLinus Torvalds tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 28141da177e4SLinus Torvalds tp->max_window = 0; 28155d424d5aSJohn Heffner tcp_mtup_init(sk); 28161da177e4SLinus Torvalds tcp_sync_mss(sk, dst_mtu(dst)); 28171da177e4SLinus Torvalds 28181da177e4SLinus Torvalds if (!tp->window_clamp) 28191da177e4SLinus Torvalds tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 28200dbaee3bSDavid S. Miller tp->advmss = dst_metric_advmss(dst); 2821f5fff5dcSTom Quetchenbach if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss) 2822f5fff5dcSTom Quetchenbach tp->advmss = tp->rx_opt.user_mss; 2823f5fff5dcSTom Quetchenbach 28241da177e4SLinus Torvalds tcp_initialize_rcv_mss(sk); 28251da177e4SLinus Torvalds 2826e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 2827e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 2828e88c64f0SHagen Paul Pfeifer (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) 2829e88c64f0SHagen Paul Pfeifer tp->window_clamp = tcp_full_space(sk); 2830e88c64f0SHagen Paul Pfeifer 28311da177e4SLinus Torvalds tcp_select_initial_window(tcp_full_space(sk), 28321da177e4SLinus Torvalds tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 28331da177e4SLinus Torvalds &tp->rcv_wnd, 28341da177e4SLinus Torvalds &tp->window_clamp, 2835bb5b7c11SDavid S. Miller sysctl_tcp_window_scaling, 283631d12926Slaurent chavey &rcv_wscale, 283731d12926Slaurent chavey dst_metric(dst, RTAX_INITRWND)); 28381da177e4SLinus Torvalds 28391da177e4SLinus Torvalds tp->rx_opt.rcv_wscale = rcv_wscale; 28401da177e4SLinus Torvalds tp->rcv_ssthresh = tp->rcv_wnd; 28411da177e4SLinus Torvalds 28421da177e4SLinus Torvalds sk->sk_err = 0; 28431da177e4SLinus Torvalds sock_reset_flag(sk, SOCK_DONE); 28441da177e4SLinus Torvalds tp->snd_wnd = 0; 2845ee7537b6SHantzis Fotis tcp_init_wl(tp, 0); 28461da177e4SLinus Torvalds tp->snd_una = tp->write_seq; 28471da177e4SLinus Torvalds tp->snd_sml = tp->write_seq; 284833f5f57eSIlpo Järvinen tp->snd_up = tp->write_seq; 2849370816aeSPavel Emelyanov tp->snd_nxt = tp->write_seq; 2850ee995283SPavel Emelyanov 2851ee995283SPavel Emelyanov if (likely(!tp->repair)) 28521da177e4SLinus Torvalds tp->rcv_nxt = 0; 2853c7781a6eSAndrew Vagin else 2854c7781a6eSAndrew Vagin tp->rcv_tstamp = tcp_time_stamp; 2855ee995283SPavel Emelyanov tp->rcv_wup = tp->rcv_nxt; 2856ee995283SPavel Emelyanov tp->copied_seq = tp->rcv_nxt; 28571da177e4SLinus Torvalds 2858463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; 2859463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_retransmits = 0; 28601da177e4SLinus Torvalds tcp_clear_retrans(tp); 28611da177e4SLinus Torvalds } 28621da177e4SLinus Torvalds 2863783237e8SYuchung Cheng static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) 2864783237e8SYuchung Cheng { 2865783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 2866783237e8SYuchung Cheng struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 2867783237e8SYuchung Cheng 2868783237e8SYuchung Cheng tcb->end_seq += skb->len; 2869783237e8SYuchung Cheng skb_header_release(skb); 2870783237e8SYuchung Cheng __tcp_add_write_queue_tail(sk, skb); 2871783237e8SYuchung Cheng sk->sk_wmem_queued += skb->truesize; 2872783237e8SYuchung Cheng sk_mem_charge(sk, skb->truesize); 2873783237e8SYuchung Cheng tp->write_seq = tcb->end_seq; 2874783237e8SYuchung Cheng tp->packets_out += tcp_skb_pcount(skb); 2875783237e8SYuchung Cheng } 2876783237e8SYuchung Cheng 2877783237e8SYuchung Cheng /* Build and send a SYN with data and (cached) Fast Open cookie. However, 2878783237e8SYuchung Cheng * queue a data-only packet after the regular SYN, such that regular SYNs 2879783237e8SYuchung Cheng * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges 2880783237e8SYuchung Cheng * only the SYN sequence, the data are retransmitted in the first ACK. 2881783237e8SYuchung Cheng * If cookie is not cached or other error occurs, falls back to send a 2882783237e8SYuchung Cheng * regular SYN with Fast Open cookie request option. 2883783237e8SYuchung Cheng */ 2884783237e8SYuchung Cheng static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) 2885783237e8SYuchung Cheng { 2886783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 2887783237e8SYuchung Cheng struct tcp_fastopen_request *fo = tp->fastopen_req; 2888aab48743SYuchung Cheng int syn_loss = 0, space, i, err = 0, iovlen = fo->data->msg_iovlen; 2889783237e8SYuchung Cheng struct sk_buff *syn_data = NULL, *data; 2890aab48743SYuchung Cheng unsigned long last_syn_loss = 0; 2891783237e8SYuchung Cheng 289267da22d2SYuchung Cheng tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ 2893aab48743SYuchung Cheng tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie, 2894aab48743SYuchung Cheng &syn_loss, &last_syn_loss); 2895aab48743SYuchung Cheng /* Recurring FO SYN losses: revert to regular handshake temporarily */ 2896aab48743SYuchung Cheng if (syn_loss > 1 && 2897aab48743SYuchung Cheng time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) { 2898aab48743SYuchung Cheng fo->cookie.len = -1; 2899aab48743SYuchung Cheng goto fallback; 2900aab48743SYuchung Cheng } 2901aab48743SYuchung Cheng 290267da22d2SYuchung Cheng if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE) 290367da22d2SYuchung Cheng fo->cookie.len = -1; 290467da22d2SYuchung Cheng else if (fo->cookie.len <= 0) 2905783237e8SYuchung Cheng goto fallback; 2906783237e8SYuchung Cheng 2907783237e8SYuchung Cheng /* MSS for SYN-data is based on cached MSS and bounded by PMTU and 2908783237e8SYuchung Cheng * user-MSS. Reserve maximum option space for middleboxes that add 2909783237e8SYuchung Cheng * private TCP options. The cost is reduced data space in SYN :( 2910783237e8SYuchung Cheng */ 2911783237e8SYuchung Cheng if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp) 2912783237e8SYuchung Cheng tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 29131b63edd6SYuchung Cheng space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) - 2914783237e8SYuchung Cheng MAX_TCP_OPTION_SPACE; 2915783237e8SYuchung Cheng 2916f5ddcbbbSEric Dumazet space = min_t(size_t, space, fo->size); 2917f5ddcbbbSEric Dumazet 2918f5ddcbbbSEric Dumazet /* limit to order-0 allocations */ 2919f5ddcbbbSEric Dumazet space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER)); 2920f5ddcbbbSEric Dumazet 2921f5ddcbbbSEric Dumazet syn_data = skb_copy_expand(syn, MAX_TCP_HEADER, space, 2922783237e8SYuchung Cheng sk->sk_allocation); 2923783237e8SYuchung Cheng if (syn_data == NULL) 2924783237e8SYuchung Cheng goto fallback; 2925783237e8SYuchung Cheng 2926783237e8SYuchung Cheng for (i = 0; i < iovlen && syn_data->len < space; ++i) { 2927783237e8SYuchung Cheng struct iovec *iov = &fo->data->msg_iov[i]; 2928783237e8SYuchung Cheng unsigned char __user *from = iov->iov_base; 2929783237e8SYuchung Cheng int len = iov->iov_len; 2930783237e8SYuchung Cheng 2931783237e8SYuchung Cheng if (syn_data->len + len > space) 2932783237e8SYuchung Cheng len = space - syn_data->len; 2933783237e8SYuchung Cheng else if (i + 1 == iovlen) 2934783237e8SYuchung Cheng /* No more data pending in inet_wait_for_connect() */ 2935783237e8SYuchung Cheng fo->data = NULL; 2936783237e8SYuchung Cheng 2937783237e8SYuchung Cheng if (skb_add_data(syn_data, from, len)) 2938783237e8SYuchung Cheng goto fallback; 2939783237e8SYuchung Cheng } 2940783237e8SYuchung Cheng 2941783237e8SYuchung Cheng /* Queue a data-only packet after the regular SYN for retransmission */ 2942783237e8SYuchung Cheng data = pskb_copy(syn_data, sk->sk_allocation); 2943783237e8SYuchung Cheng if (data == NULL) 2944783237e8SYuchung Cheng goto fallback; 2945783237e8SYuchung Cheng TCP_SKB_CB(data)->seq++; 2946783237e8SYuchung Cheng TCP_SKB_CB(data)->tcp_flags &= ~TCPHDR_SYN; 2947783237e8SYuchung Cheng TCP_SKB_CB(data)->tcp_flags = (TCPHDR_ACK|TCPHDR_PSH); 2948783237e8SYuchung Cheng tcp_connect_queue_skb(sk, data); 2949783237e8SYuchung Cheng fo->copied = data->len; 2950783237e8SYuchung Cheng 2951783237e8SYuchung Cheng if (tcp_transmit_skb(sk, syn_data, 0, sk->sk_allocation) == 0) { 295267da22d2SYuchung Cheng tp->syn_data = (fo->copied > 0); 2953783237e8SYuchung Cheng NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE); 2954783237e8SYuchung Cheng goto done; 2955783237e8SYuchung Cheng } 2956783237e8SYuchung Cheng syn_data = NULL; 2957783237e8SYuchung Cheng 2958783237e8SYuchung Cheng fallback: 2959783237e8SYuchung Cheng /* Send a regular SYN with Fast Open cookie request option */ 2960783237e8SYuchung Cheng if (fo->cookie.len > 0) 2961783237e8SYuchung Cheng fo->cookie.len = 0; 2962783237e8SYuchung Cheng err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); 2963783237e8SYuchung Cheng if (err) 2964783237e8SYuchung Cheng tp->syn_fastopen = 0; 2965783237e8SYuchung Cheng kfree_skb(syn_data); 2966783237e8SYuchung Cheng done: 2967783237e8SYuchung Cheng fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ 2968783237e8SYuchung Cheng return err; 2969783237e8SYuchung Cheng } 2970783237e8SYuchung Cheng 297167edfef7SAndi Kleen /* Build a SYN and send it off. */ 29721da177e4SLinus Torvalds int tcp_connect(struct sock *sk) 29731da177e4SLinus Torvalds { 29741da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 29751da177e4SLinus Torvalds struct sk_buff *buff; 2976ee586811SEric Paris int err; 29771da177e4SLinus Torvalds 29781da177e4SLinus Torvalds tcp_connect_init(sk); 29791da177e4SLinus Torvalds 29802b916477SAndrey Vagin if (unlikely(tp->repair)) { 29812b916477SAndrey Vagin tcp_finish_connect(sk, NULL); 29822b916477SAndrey Vagin return 0; 29832b916477SAndrey Vagin } 29842b916477SAndrey Vagin 2985d179cd12SDavid S. Miller buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation); 29861da177e4SLinus Torvalds if (unlikely(buff == NULL)) 29871da177e4SLinus Torvalds return -ENOBUFS; 29881da177e4SLinus Torvalds 29891da177e4SLinus Torvalds /* Reserve space for headers. */ 29901da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 29911da177e4SLinus Torvalds 2992a3433f35SChangli Gao tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); 2993783237e8SYuchung Cheng tp->retrans_stamp = TCP_SKB_CB(buff)->when = tcp_time_stamp; 2994783237e8SYuchung Cheng tcp_connect_queue_skb(sk, buff); 2995e870a8efSIlpo Järvinen TCP_ECN_send_syn(sk, buff); 29961da177e4SLinus Torvalds 2997783237e8SYuchung Cheng /* Send off SYN; include data in Fast Open. */ 2998783237e8SYuchung Cheng err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : 2999783237e8SYuchung Cheng tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); 3000ee586811SEric Paris if (err == -ECONNREFUSED) 3001ee586811SEric Paris return err; 3002bd37a088SWei Yongjun 3003bd37a088SWei Yongjun /* We change tp->snd_nxt after the tcp_transmit_skb() call 3004bd37a088SWei Yongjun * in order to make this packet get counted in tcpOutSegs. 3005bd37a088SWei Yongjun */ 3006bd37a088SWei Yongjun tp->snd_nxt = tp->write_seq; 3007bd37a088SWei Yongjun tp->pushed_seq = tp->write_seq; 300881cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); 30091da177e4SLinus Torvalds 30101da177e4SLinus Torvalds /* Timer for repeating the SYN until an answer. */ 30113f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 30123f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 30131da177e4SLinus Torvalds return 0; 30141da177e4SLinus Torvalds } 30154bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_connect); 30161da177e4SLinus Torvalds 30171da177e4SLinus Torvalds /* Send out a delayed ack, the caller does the policy checking 30181da177e4SLinus Torvalds * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 30191da177e4SLinus Torvalds * for details. 30201da177e4SLinus Torvalds */ 30211da177e4SLinus Torvalds void tcp_send_delayed_ack(struct sock *sk) 30221da177e4SLinus Torvalds { 3023463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 3024463c84b9SArnaldo Carvalho de Melo int ato = icsk->icsk_ack.ato; 30251da177e4SLinus Torvalds unsigned long timeout; 30261da177e4SLinus Torvalds 30271da177e4SLinus Torvalds if (ato > TCP_DELACK_MIN) { 3028463c84b9SArnaldo Carvalho de Melo const struct tcp_sock *tp = tcp_sk(sk); 30291da177e4SLinus Torvalds int max_ato = HZ / 2; 30301da177e4SLinus Torvalds 3031056834d9SIlpo Järvinen if (icsk->icsk_ack.pingpong || 3032056834d9SIlpo Järvinen (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 30331da177e4SLinus Torvalds max_ato = TCP_DELACK_MAX; 30341da177e4SLinus Torvalds 30351da177e4SLinus Torvalds /* Slow path, intersegment interval is "high". */ 30361da177e4SLinus Torvalds 30371da177e4SLinus Torvalds /* If some rtt estimate is known, use it to bound delayed ack. 3038463c84b9SArnaldo Carvalho de Melo * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 30391da177e4SLinus Torvalds * directly. 30401da177e4SLinus Torvalds */ 30411da177e4SLinus Torvalds if (tp->srtt) { 30421da177e4SLinus Torvalds int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN); 30431da177e4SLinus Torvalds 30441da177e4SLinus Torvalds if (rtt < max_ato) 30451da177e4SLinus Torvalds max_ato = rtt; 30461da177e4SLinus Torvalds } 30471da177e4SLinus Torvalds 30481da177e4SLinus Torvalds ato = min(ato, max_ato); 30491da177e4SLinus Torvalds } 30501da177e4SLinus Torvalds 30511da177e4SLinus Torvalds /* Stay within the limit we were given */ 30521da177e4SLinus Torvalds timeout = jiffies + ato; 30531da177e4SLinus Torvalds 30541da177e4SLinus Torvalds /* Use new timeout only if there wasn't a older one earlier. */ 3055463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 30561da177e4SLinus Torvalds /* If delack timer was blocked or is about to expire, 30571da177e4SLinus Torvalds * send ACK now. 30581da177e4SLinus Torvalds */ 3059463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.blocked || 3060463c84b9SArnaldo Carvalho de Melo time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 30611da177e4SLinus Torvalds tcp_send_ack(sk); 30621da177e4SLinus Torvalds return; 30631da177e4SLinus Torvalds } 30641da177e4SLinus Torvalds 3065463c84b9SArnaldo Carvalho de Melo if (!time_before(timeout, icsk->icsk_ack.timeout)) 3066463c84b9SArnaldo Carvalho de Melo timeout = icsk->icsk_ack.timeout; 30671da177e4SLinus Torvalds } 3068463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 3069463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.timeout = timeout; 3070463c84b9SArnaldo Carvalho de Melo sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 30711da177e4SLinus Torvalds } 30721da177e4SLinus Torvalds 30731da177e4SLinus Torvalds /* This routine sends an ack and also updates the window. */ 30741da177e4SLinus Torvalds void tcp_send_ack(struct sock *sk) 30751da177e4SLinus Torvalds { 30761da177e4SLinus Torvalds struct sk_buff *buff; 30771da177e4SLinus Torvalds 3078058dc334SIlpo Järvinen /* If we have been reset, we may not send again. */ 3079058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3080058dc334SIlpo Järvinen return; 3081058dc334SIlpo Järvinen 30821da177e4SLinus Torvalds /* We are not putting this on the write queue, so 30831da177e4SLinus Torvalds * tcp_transmit_skb() will set the ownership to this 30841da177e4SLinus Torvalds * sock. 30851da177e4SLinus Torvalds */ 308699a1dec7SMel Gorman buff = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC)); 30871da177e4SLinus Torvalds if (buff == NULL) { 3088463c84b9SArnaldo Carvalho de Melo inet_csk_schedule_ack(sk); 3089463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; 30903f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 30913f421baaSArnaldo Carvalho de Melo TCP_DELACK_MAX, TCP_RTO_MAX); 30921da177e4SLinus Torvalds return; 30931da177e4SLinus Torvalds } 30941da177e4SLinus Torvalds 30951da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 30961da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 3097a3433f35SChangli Gao tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); 30981da177e4SLinus Torvalds 30991da177e4SLinus Torvalds /* Send it off, this clears delayed acks for us. */ 31001da177e4SLinus Torvalds TCP_SKB_CB(buff)->when = tcp_time_stamp; 310199a1dec7SMel Gorman tcp_transmit_skb(sk, buff, 0, sk_gfp_atomic(sk, GFP_ATOMIC)); 31021da177e4SLinus Torvalds } 31031da177e4SLinus Torvalds 31041da177e4SLinus Torvalds /* This routine sends a packet with an out of date sequence 31051da177e4SLinus Torvalds * number. It assumes the other end will try to ack it. 31061da177e4SLinus Torvalds * 31071da177e4SLinus Torvalds * Question: what should we make while urgent mode? 31081da177e4SLinus Torvalds * 4.4BSD forces sending single byte of data. We cannot send 31091da177e4SLinus Torvalds * out of window data, because we have SND.NXT==SND.MAX... 31101da177e4SLinus Torvalds * 31111da177e4SLinus Torvalds * Current solution: to send TWO zero-length segments in urgent mode: 31121da177e4SLinus Torvalds * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 31131da177e4SLinus Torvalds * out-of-date with SND.UNA-1 to probe window. 31141da177e4SLinus Torvalds */ 31151da177e4SLinus Torvalds static int tcp_xmit_probe_skb(struct sock *sk, int urgent) 31161da177e4SLinus Torvalds { 31171da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 31181da177e4SLinus Torvalds struct sk_buff *skb; 31191da177e4SLinus Torvalds 31201da177e4SLinus Torvalds /* We don't queue it, tcp_transmit_skb() sets ownership. */ 312199a1dec7SMel Gorman skb = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC)); 31221da177e4SLinus Torvalds if (skb == NULL) 31231da177e4SLinus Torvalds return -1; 31241da177e4SLinus Torvalds 31251da177e4SLinus Torvalds /* Reserve space for headers and set control bits. */ 31261da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 31271da177e4SLinus Torvalds /* Use a previous sequence. This should cause the other 31281da177e4SLinus Torvalds * end to send an ack. Don't queue or clone SKB, just 31291da177e4SLinus Torvalds * send it. 31301da177e4SLinus Torvalds */ 3131a3433f35SChangli Gao tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); 31321da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 3133dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC); 31341da177e4SLinus Torvalds } 31351da177e4SLinus Torvalds 3136ee995283SPavel Emelyanov void tcp_send_window_probe(struct sock *sk) 3137ee995283SPavel Emelyanov { 3138ee995283SPavel Emelyanov if (sk->sk_state == TCP_ESTABLISHED) { 3139ee995283SPavel Emelyanov tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; 3140ee995283SPavel Emelyanov tcp_xmit_probe_skb(sk, 0); 3141ee995283SPavel Emelyanov } 3142ee995283SPavel Emelyanov } 3143ee995283SPavel Emelyanov 314467edfef7SAndi Kleen /* Initiate keepalive or window probe from timer. */ 31451da177e4SLinus Torvalds int tcp_write_wakeup(struct sock *sk) 31461da177e4SLinus Torvalds { 31471da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 31481da177e4SLinus Torvalds struct sk_buff *skb; 31491da177e4SLinus Torvalds 3150058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3151058dc334SIlpo Järvinen return -1; 3152058dc334SIlpo Järvinen 3153fe067e8aSDavid S. Miller if ((skb = tcp_send_head(sk)) != NULL && 315490840defSIlpo Järvinen before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { 31551da177e4SLinus Torvalds int err; 31560c54b85fSIlpo Järvinen unsigned int mss = tcp_current_mss(sk); 315790840defSIlpo Järvinen unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 31581da177e4SLinus Torvalds 31591da177e4SLinus Torvalds if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 31601da177e4SLinus Torvalds tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 31611da177e4SLinus Torvalds 31621da177e4SLinus Torvalds /* We are probing the opening of a window 31631da177e4SLinus Torvalds * but the window size is != 0 31641da177e4SLinus Torvalds * must have been a result SWS avoidance ( sender ) 31651da177e4SLinus Torvalds */ 31661da177e4SLinus Torvalds if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 31671da177e4SLinus Torvalds skb->len > mss) { 31681da177e4SLinus Torvalds seg_size = min(seg_size, mss); 31694de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 3170846998aeSDavid S. Miller if (tcp_fragment(sk, skb, seg_size, mss)) 31711da177e4SLinus Torvalds return -1; 31721da177e4SLinus Torvalds } else if (!tcp_skb_pcount(skb)) 3173846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss); 31741da177e4SLinus Torvalds 31754de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 31761da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 3177dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 317866f5fe62SIlpo Järvinen if (!err) 317966f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 31801da177e4SLinus Torvalds return err; 31811da177e4SLinus Torvalds } else { 318233f5f57eSIlpo Järvinen if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) 31834828e7f4SIlpo Järvinen tcp_xmit_probe_skb(sk, 1); 31841da177e4SLinus Torvalds return tcp_xmit_probe_skb(sk, 0); 31851da177e4SLinus Torvalds } 31861da177e4SLinus Torvalds } 31871da177e4SLinus Torvalds 31881da177e4SLinus Torvalds /* A window probe timeout has occurred. If window is not closed send 31891da177e4SLinus Torvalds * a partial packet else a zero probe. 31901da177e4SLinus Torvalds */ 31911da177e4SLinus Torvalds void tcp_send_probe0(struct sock *sk) 31921da177e4SLinus Torvalds { 3193463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 31941da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 31951da177e4SLinus Torvalds int err; 31961da177e4SLinus Torvalds 31971da177e4SLinus Torvalds err = tcp_write_wakeup(sk); 31981da177e4SLinus Torvalds 3199fe067e8aSDavid S. Miller if (tp->packets_out || !tcp_send_head(sk)) { 32001da177e4SLinus Torvalds /* Cancel probe timer, if it is not required. */ 32016687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 0; 3202463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff = 0; 32031da177e4SLinus Torvalds return; 32041da177e4SLinus Torvalds } 32051da177e4SLinus Torvalds 32061da177e4SLinus Torvalds if (err <= 0) { 3207463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_backoff < sysctl_tcp_retries2) 3208463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff++; 32096687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out++; 3210463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 32113f421baaSArnaldo Carvalho de Melo min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), 32123f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 32131da177e4SLinus Torvalds } else { 32141da177e4SLinus Torvalds /* If packet was not sent due to local congestion, 32156687e988SArnaldo Carvalho de Melo * do not backoff and do not remember icsk_probes_out. 32161da177e4SLinus Torvalds * Let local senders to fight for local resources. 32171da177e4SLinus Torvalds * 32181da177e4SLinus Torvalds * Use accumulated backoff yet. 32191da177e4SLinus Torvalds */ 32206687e988SArnaldo Carvalho de Melo if (!icsk->icsk_probes_out) 32216687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 1; 3222463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 3223463c84b9SArnaldo Carvalho de Melo min(icsk->icsk_rto << icsk->icsk_backoff, 32243f421baaSArnaldo Carvalho de Melo TCP_RESOURCE_PROBE_INTERVAL), 32253f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 32261da177e4SLinus Torvalds } 32271da177e4SLinus Torvalds } 3228