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 53c39c4c6aSWei Liu /* Default TSQ limit of four TSO segments */ 54c39c4c6aSWei Liu int sysctl_tcp_limit_output_bytes __read_mostly = 262144; 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 6235089bb2SDavid S. Miller /* By default, RFC2861 behavior. */ 63ab32ea5dSBrian Haley int sysctl_tcp_slow_start_after_idle __read_mostly = 1; 6435089bb2SDavid S. Miller 65c9bee3b7SEric Dumazet unsigned int sysctl_tcp_notsent_lowat __read_mostly = UINT_MAX; 66c9bee3b7SEric Dumazet EXPORT_SYMBOL(sysctl_tcp_notsent_lowat); 67c9bee3b7SEric Dumazet 6846d3ceabSEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 6946d3ceabSEric Dumazet int push_one, gfp_t gfp); 70519855c5SWilliam Allen Simpson 7167edfef7SAndi Kleen /* Account for new data that has been sent to the network. */ 72cf533ea5SEric Dumazet static void tcp_event_new_data_sent(struct sock *sk, const struct sk_buff *skb) 736ff03ac3SIlpo Järvinen { 746ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 756ff03ac3SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 7666f5fe62SIlpo Järvinen unsigned int prior_packets = tp->packets_out; 779e412ba7SIlpo Järvinen 78fe067e8aSDavid S. Miller tcp_advance_send_head(sk, skb); 791da177e4SLinus Torvalds tp->snd_nxt = TCP_SKB_CB(skb)->end_seq; 808512430eSIlpo Järvinen 8166f5fe62SIlpo Järvinen tp->packets_out += tcp_skb_pcount(skb); 826ba8a3b1SNandita Dukkipati if (!prior_packets || icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS || 836a5dc9e5SEric Dumazet icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) { 84750ea2baSYuchung Cheng tcp_rearm_rto(sk); 851da177e4SLinus Torvalds } 86f19c29e3SYuchung Cheng 87f7324acdSDavid S. Miller NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT, 88f19c29e3SYuchung Cheng tcp_skb_pcount(skb)); 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; 1661da177e4SLinus Torvalds 16735089bb2SDavid S. Miller if (sysctl_tcp_slow_start_after_idle && 16835089bb2SDavid S. Miller (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto)) 169463c84b9SArnaldo Carvalho de Melo tcp_cwnd_restart(sk, __sk_dst_get(sk)); 1701da177e4SLinus Torvalds 1711da177e4SLinus Torvalds tp->lsndtime = now; 1721da177e4SLinus Torvalds 1731da177e4SLinus Torvalds /* If it is a reply for ato after last received 1741da177e4SLinus Torvalds * packet, enter pingpong mode. 1751da177e4SLinus Torvalds */ 1762251ae46SJon Maxwell if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato) 177463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pingpong = 1; 1781da177e4SLinus Torvalds } 1791da177e4SLinus Torvalds 18067edfef7SAndi Kleen /* Account for an ACK we sent. */ 18140efc6faSStephen Hemminger static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts) 1821da177e4SLinus Torvalds { 183463c84b9SArnaldo Carvalho de Melo tcp_dec_quickack_mode(sk, pkts); 184463c84b9SArnaldo Carvalho de Melo inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); 1851da177e4SLinus Torvalds } 1861da177e4SLinus Torvalds 18785f16525SYuchung Cheng 18885f16525SYuchung Cheng u32 tcp_default_init_rwnd(u32 mss) 18985f16525SYuchung Cheng { 19085f16525SYuchung Cheng /* Initial receive window should be twice of TCP_INIT_CWND to 1919ef71e0cSWeiping Pan * enable proper sending of new unsent data during fast recovery 19285f16525SYuchung Cheng * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a 19385f16525SYuchung Cheng * limit when mss is larger than 1460. 19485f16525SYuchung Cheng */ 19585f16525SYuchung Cheng u32 init_rwnd = TCP_INIT_CWND * 2; 19685f16525SYuchung Cheng 19785f16525SYuchung Cheng if (mss > 1460) 19885f16525SYuchung Cheng init_rwnd = max((1460 * init_rwnd) / mss, 2U); 19985f16525SYuchung Cheng return init_rwnd; 20085f16525SYuchung Cheng } 20185f16525SYuchung Cheng 2021da177e4SLinus Torvalds /* Determine a window scaling and initial window to offer. 2031da177e4SLinus Torvalds * Based on the assumption that the given amount of space 2041da177e4SLinus Torvalds * will be offered. Store the results in the tp structure. 2051da177e4SLinus Torvalds * NOTE: for smooth operation initial space offering should 2061da177e4SLinus Torvalds * be a multiple of mss if possible. We assume here that mss >= 1. 2071da177e4SLinus Torvalds * This MUST be enforced by all callers. 2081da177e4SLinus Torvalds */ 2091da177e4SLinus Torvalds void tcp_select_initial_window(int __space, __u32 mss, 2101da177e4SLinus Torvalds __u32 *rcv_wnd, __u32 *window_clamp, 21131d12926Slaurent chavey int wscale_ok, __u8 *rcv_wscale, 21231d12926Slaurent chavey __u32 init_rcv_wnd) 2131da177e4SLinus Torvalds { 2141da177e4SLinus Torvalds unsigned int space = (__space < 0 ? 0 : __space); 2151da177e4SLinus Torvalds 2161da177e4SLinus Torvalds /* If no clamp set the clamp to the max possible scaled window */ 2171da177e4SLinus Torvalds if (*window_clamp == 0) 2181da177e4SLinus Torvalds (*window_clamp) = (65535 << 14); 2191da177e4SLinus Torvalds space = min(*window_clamp, space); 2201da177e4SLinus Torvalds 2211da177e4SLinus Torvalds /* Quantize space offering to a multiple of mss if possible. */ 2221da177e4SLinus Torvalds if (space > mss) 2231da177e4SLinus Torvalds space = (space / mss) * mss; 2241da177e4SLinus Torvalds 2251da177e4SLinus Torvalds /* NOTE: offering an initial window larger than 32767 22615d99e02SRick Jones * will break some buggy TCP stacks. If the admin tells us 22715d99e02SRick Jones * it is likely we could be speaking with such a buggy stack 22815d99e02SRick Jones * we will truncate our initial window offering to 32K-1 22915d99e02SRick Jones * unless the remote has sent us a window scaling option, 23015d99e02SRick Jones * which we interpret as a sign the remote TCP is not 23115d99e02SRick Jones * misinterpreting the window field as a signed quantity. 2321da177e4SLinus Torvalds */ 23315d99e02SRick Jones if (sysctl_tcp_workaround_signed_windows) 2341da177e4SLinus Torvalds (*rcv_wnd) = min(space, MAX_TCP_WINDOW); 23515d99e02SRick Jones else 23615d99e02SRick Jones (*rcv_wnd) = space; 23715d99e02SRick Jones 2381da177e4SLinus Torvalds (*rcv_wscale) = 0; 2391da177e4SLinus Torvalds if (wscale_ok) { 2401da177e4SLinus Torvalds /* Set window scaling on max possible window 2411da177e4SLinus Torvalds * See RFC1323 for an explanation of the limit to 14 2421da177e4SLinus Torvalds */ 2431da177e4SLinus Torvalds space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max); 244316c1592SStephen Hemminger space = min_t(u32, space, *window_clamp); 2451da177e4SLinus Torvalds while (space > 65535 && (*rcv_wscale) < 14) { 2461da177e4SLinus Torvalds space >>= 1; 2471da177e4SLinus Torvalds (*rcv_wscale)++; 2481da177e4SLinus Torvalds } 2491da177e4SLinus Torvalds } 2501da177e4SLinus Torvalds 2511da177e4SLinus Torvalds if (mss > (1 << *rcv_wscale)) { 25285f16525SYuchung Cheng if (!init_rcv_wnd) /* Use default unless specified otherwise */ 25385f16525SYuchung Cheng init_rcv_wnd = tcp_default_init_rwnd(mss); 254b1afde60SNandita Dukkipati *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss); 2551da177e4SLinus Torvalds } 2561da177e4SLinus Torvalds 2571da177e4SLinus Torvalds /* Set the clamp no higher than max representable value */ 2581da177e4SLinus Torvalds (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp); 2591da177e4SLinus Torvalds } 2604bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_select_initial_window); 2611da177e4SLinus Torvalds 2621da177e4SLinus Torvalds /* Chose a new window to advertise, update state in tcp_sock for the 2631da177e4SLinus Torvalds * socket, and return result with RFC1323 scaling applied. The return 2641da177e4SLinus Torvalds * value can be stuffed directly into th->window for an outgoing 2651da177e4SLinus Torvalds * frame. 2661da177e4SLinus Torvalds */ 26740efc6faSStephen Hemminger static u16 tcp_select_window(struct sock *sk) 2681da177e4SLinus Torvalds { 2691da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2708e165e20SFlorian Westphal u32 old_win = tp->rcv_wnd; 2711da177e4SLinus Torvalds u32 cur_win = tcp_receive_window(tp); 2721da177e4SLinus Torvalds u32 new_win = __tcp_select_window(sk); 2731da177e4SLinus Torvalds 2741da177e4SLinus Torvalds /* Never shrink the offered window */ 2751da177e4SLinus Torvalds if (new_win < cur_win) { 2761da177e4SLinus Torvalds /* Danger Will Robinson! 2771da177e4SLinus Torvalds * Don't update rcv_wup/rcv_wnd here or else 2781da177e4SLinus Torvalds * we will not be able to advertise a zero 2791da177e4SLinus Torvalds * window in time. --DaveM 2801da177e4SLinus Torvalds * 2811da177e4SLinus Torvalds * Relax Will Robinson. 2821da177e4SLinus Torvalds */ 2838e165e20SFlorian Westphal if (new_win == 0) 2848e165e20SFlorian Westphal NET_INC_STATS(sock_net(sk), 2858e165e20SFlorian Westphal LINUX_MIB_TCPWANTZEROWINDOWADV); 286607bfbf2SPatrick McHardy new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale); 2871da177e4SLinus Torvalds } 2881da177e4SLinus Torvalds tp->rcv_wnd = new_win; 2891da177e4SLinus Torvalds tp->rcv_wup = tp->rcv_nxt; 2901da177e4SLinus Torvalds 2911da177e4SLinus Torvalds /* Make sure we do not exceed the maximum possible 2921da177e4SLinus Torvalds * scaled window. 2931da177e4SLinus Torvalds */ 29415d99e02SRick Jones if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows) 2951da177e4SLinus Torvalds new_win = min(new_win, MAX_TCP_WINDOW); 2961da177e4SLinus Torvalds else 2971da177e4SLinus Torvalds new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); 2981da177e4SLinus Torvalds 2991da177e4SLinus Torvalds /* RFC1323 scaling applied */ 3001da177e4SLinus Torvalds new_win >>= tp->rx_opt.rcv_wscale; 3011da177e4SLinus Torvalds 3021da177e4SLinus Torvalds /* If we advertise zero window, disable fast path. */ 3038e165e20SFlorian Westphal if (new_win == 0) { 3041da177e4SLinus Torvalds tp->pred_flags = 0; 3058e165e20SFlorian Westphal if (old_win) 3068e165e20SFlorian Westphal NET_INC_STATS(sock_net(sk), 3078e165e20SFlorian Westphal LINUX_MIB_TCPTOZEROWINDOWADV); 3088e165e20SFlorian Westphal } else if (old_win == 0) { 3098e165e20SFlorian Westphal NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV); 3108e165e20SFlorian Westphal } 3111da177e4SLinus Torvalds 3121da177e4SLinus Torvalds return new_win; 3131da177e4SLinus Torvalds } 3141da177e4SLinus Torvalds 31567edfef7SAndi Kleen /* Packet ECN state for a SYN-ACK */ 316735d3831SFlorian Westphal static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb) 317bdf1ee5dSIlpo Järvinen { 31830e502a3SDaniel Borkmann const struct tcp_sock *tp = tcp_sk(sk); 31930e502a3SDaniel Borkmann 3204de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR; 321bdf1ee5dSIlpo Järvinen if (!(tp->ecn_flags & TCP_ECN_OK)) 3224de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE; 32330e502a3SDaniel Borkmann else if (tcp_ca_needs_ecn(sk)) 32430e502a3SDaniel Borkmann INET_ECN_xmit(sk); 325bdf1ee5dSIlpo Järvinen } 326bdf1ee5dSIlpo Järvinen 32767edfef7SAndi Kleen /* Packet ECN state for a SYN. */ 328735d3831SFlorian Westphal static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb) 329bdf1ee5dSIlpo Järvinen { 330bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 331f7b3bec6SFlorian Westphal bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 || 332f7b3bec6SFlorian Westphal tcp_ca_needs_ecn(sk); 333f7b3bec6SFlorian Westphal 334f7b3bec6SFlorian Westphal if (!use_ecn) { 335f7b3bec6SFlorian Westphal const struct dst_entry *dst = __sk_dst_get(sk); 336f7b3bec6SFlorian Westphal 337f7b3bec6SFlorian Westphal if (dst && dst_feature(dst, RTAX_FEATURE_ECN)) 338f7b3bec6SFlorian Westphal use_ecn = true; 339f7b3bec6SFlorian Westphal } 340bdf1ee5dSIlpo Järvinen 341bdf1ee5dSIlpo Järvinen tp->ecn_flags = 0; 342f7b3bec6SFlorian Westphal 343f7b3bec6SFlorian Westphal if (use_ecn) { 3444de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR; 345bdf1ee5dSIlpo Järvinen tp->ecn_flags = TCP_ECN_OK; 34630e502a3SDaniel Borkmann if (tcp_ca_needs_ecn(sk)) 34730e502a3SDaniel Borkmann INET_ECN_xmit(sk); 348bdf1ee5dSIlpo Järvinen } 349bdf1ee5dSIlpo Järvinen } 350bdf1ee5dSIlpo Järvinen 35149213555SDaniel Borkmann static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb) 35249213555SDaniel Borkmann { 35349213555SDaniel Borkmann if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback) 35449213555SDaniel Borkmann /* tp->ecn_flags are cleared at a later point in time when 35549213555SDaniel Borkmann * SYN ACK is ultimatively being received. 35649213555SDaniel Borkmann */ 35749213555SDaniel Borkmann TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR); 35849213555SDaniel Borkmann } 35949213555SDaniel Borkmann 360735d3831SFlorian Westphal static void 361735d3831SFlorian Westphal tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th, 36230e502a3SDaniel Borkmann struct sock *sk) 363bdf1ee5dSIlpo Järvinen { 36430e502a3SDaniel Borkmann if (inet_rsk(req)->ecn_ok) { 365bdf1ee5dSIlpo Järvinen th->ece = 1; 36630e502a3SDaniel Borkmann if (tcp_ca_needs_ecn(sk)) 36730e502a3SDaniel Borkmann INET_ECN_xmit(sk); 36830e502a3SDaniel Borkmann } 369bdf1ee5dSIlpo Järvinen } 370bdf1ee5dSIlpo Järvinen 37167edfef7SAndi Kleen /* Set up ECN state for a packet on a ESTABLISHED socket that is about to 37267edfef7SAndi Kleen * be sent. 37367edfef7SAndi Kleen */ 374735d3831SFlorian Westphal static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb, 375bdf1ee5dSIlpo Järvinen int tcp_header_len) 376bdf1ee5dSIlpo Järvinen { 377bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 378bdf1ee5dSIlpo Järvinen 379bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_OK) { 380bdf1ee5dSIlpo Järvinen /* Not-retransmitted data segment: set ECT and inject CWR. */ 381bdf1ee5dSIlpo Järvinen if (skb->len != tcp_header_len && 382bdf1ee5dSIlpo Järvinen !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { 383bdf1ee5dSIlpo Järvinen INET_ECN_xmit(sk); 384bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) { 385bdf1ee5dSIlpo Järvinen tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; 386bdf1ee5dSIlpo Järvinen tcp_hdr(skb)->cwr = 1; 387bdf1ee5dSIlpo Järvinen skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; 388bdf1ee5dSIlpo Järvinen } 38930e502a3SDaniel Borkmann } else if (!tcp_ca_needs_ecn(sk)) { 390bdf1ee5dSIlpo Järvinen /* ACK or retransmitted segment: clear ECT|CE */ 391bdf1ee5dSIlpo Järvinen INET_ECN_dontxmit(sk); 392bdf1ee5dSIlpo Järvinen } 393bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) 394bdf1ee5dSIlpo Järvinen tcp_hdr(skb)->ece = 1; 395bdf1ee5dSIlpo Järvinen } 396bdf1ee5dSIlpo Järvinen } 397bdf1ee5dSIlpo Järvinen 398e870a8efSIlpo Järvinen /* Constructs common control bits of non-data skb. If SYN/FIN is present, 399e870a8efSIlpo Järvinen * auto increment end seqno. 400e870a8efSIlpo Järvinen */ 401e870a8efSIlpo Järvinen static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags) 402e870a8efSIlpo Järvinen { 4032e8e18efSDavid S. Miller skb->ip_summed = CHECKSUM_PARTIAL; 404e870a8efSIlpo Järvinen skb->csum = 0; 405e870a8efSIlpo Järvinen 4064de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags; 407e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->sacked = 0; 408e870a8efSIlpo Järvinen 409cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, 1); 410e870a8efSIlpo Järvinen 411e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->seq = seq; 412a3433f35SChangli Gao if (flags & (TCPHDR_SYN | TCPHDR_FIN)) 413e870a8efSIlpo Järvinen seq++; 414e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->end_seq = seq; 415e870a8efSIlpo Järvinen } 416e870a8efSIlpo Järvinen 417a2a385d6SEric Dumazet static inline bool tcp_urg_mode(const struct tcp_sock *tp) 41833f5f57eSIlpo Järvinen { 41933f5f57eSIlpo Järvinen return tp->snd_una != tp->snd_up; 42033f5f57eSIlpo Järvinen } 42133f5f57eSIlpo Järvinen 42233ad798cSAdam Langley #define OPTION_SACK_ADVERTISE (1 << 0) 42333ad798cSAdam Langley #define OPTION_TS (1 << 1) 42433ad798cSAdam Langley #define OPTION_MD5 (1 << 2) 42589e95a61SOri Finkelman #define OPTION_WSCALE (1 << 3) 4262100c8d2SYuchung Cheng #define OPTION_FAST_OPEN_COOKIE (1 << 8) 42733ad798cSAdam Langley 42833ad798cSAdam Langley struct tcp_out_options { 4292100c8d2SYuchung Cheng u16 options; /* bit field of OPTION_* */ 4302100c8d2SYuchung Cheng u16 mss; /* 0 to disable */ 43133ad798cSAdam Langley u8 ws; /* window scale, 0 to disable */ 43233ad798cSAdam Langley u8 num_sack_blocks; /* number of SACK blocks to include */ 433bd0388aeSWilliam Allen Simpson u8 hash_size; /* bytes in hash_location */ 434bd0388aeSWilliam Allen Simpson __u8 *hash_location; /* temporary pointer, overloaded */ 4352100c8d2SYuchung Cheng __u32 tsval, tsecr; /* need to include OPTION_TS */ 4362100c8d2SYuchung Cheng struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */ 43733ad798cSAdam Langley }; 43833ad798cSAdam Langley 43967edfef7SAndi Kleen /* Write previously computed TCP options to the packet. 44067edfef7SAndi Kleen * 44167edfef7SAndi Kleen * Beware: Something in the Internet is very sensitive to the ordering of 442fd6149d3SIlpo Järvinen * TCP options, we learned this through the hard way, so be careful here. 443fd6149d3SIlpo Järvinen * Luckily we can at least blame others for their non-compliance but from 4448e3bff96Sstephen hemminger * inter-operability perspective it seems that we're somewhat stuck with 445fd6149d3SIlpo Järvinen * the ordering which we have been using if we want to keep working with 446fd6149d3SIlpo Järvinen * those broken things (not that it currently hurts anybody as there isn't 447fd6149d3SIlpo Järvinen * particular reason why the ordering would need to be changed). 448fd6149d3SIlpo Järvinen * 449fd6149d3SIlpo Järvinen * At least SACK_PERM as the first option is known to lead to a disaster 450fd6149d3SIlpo Järvinen * (but it may well be that other scenarios fail similarly). 451fd6149d3SIlpo Järvinen */ 45233ad798cSAdam Langley static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp, 453bd0388aeSWilliam Allen Simpson struct tcp_out_options *opts) 454bd0388aeSWilliam Allen Simpson { 4552100c8d2SYuchung Cheng u16 options = opts->options; /* mungable copy */ 456bd0388aeSWilliam Allen Simpson 457bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_MD5 & options)) { 4581a2c6181SChristoph Paasch *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | 4591a2c6181SChristoph Paasch (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG); 460bd0388aeSWilliam Allen Simpson /* overload cookie hash location */ 461bd0388aeSWilliam Allen Simpson opts->hash_location = (__u8 *)ptr; 46233ad798cSAdam Langley ptr += 4; 46333ad798cSAdam Langley } 46433ad798cSAdam Langley 465fd6149d3SIlpo Järvinen if (unlikely(opts->mss)) { 466fd6149d3SIlpo Järvinen *ptr++ = htonl((TCPOPT_MSS << 24) | 467fd6149d3SIlpo Järvinen (TCPOLEN_MSS << 16) | 468fd6149d3SIlpo Järvinen opts->mss); 469fd6149d3SIlpo Järvinen } 470fd6149d3SIlpo Järvinen 471bd0388aeSWilliam Allen Simpson if (likely(OPTION_TS & options)) { 472bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 47333ad798cSAdam Langley *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | 47433ad798cSAdam Langley (TCPOLEN_SACK_PERM << 16) | 47533ad798cSAdam Langley (TCPOPT_TIMESTAMP << 8) | 47633ad798cSAdam Langley TCPOLEN_TIMESTAMP); 477bd0388aeSWilliam Allen Simpson options &= ~OPTION_SACK_ADVERTISE; 47833ad798cSAdam Langley } else { 479496c98dfSYOSHIFUJI Hideaki *ptr++ = htonl((TCPOPT_NOP << 24) | 48040efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 48140efc6faSStephen Hemminger (TCPOPT_TIMESTAMP << 8) | 48240efc6faSStephen Hemminger TCPOLEN_TIMESTAMP); 48340efc6faSStephen Hemminger } 48433ad798cSAdam Langley *ptr++ = htonl(opts->tsval); 48533ad798cSAdam Langley *ptr++ = htonl(opts->tsecr); 48633ad798cSAdam Langley } 48733ad798cSAdam Langley 488bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 48933ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 49033ad798cSAdam Langley (TCPOPT_NOP << 16) | 49133ad798cSAdam Langley (TCPOPT_SACK_PERM << 8) | 49233ad798cSAdam Langley TCPOLEN_SACK_PERM); 49333ad798cSAdam Langley } 49433ad798cSAdam Langley 495bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_WSCALE & options)) { 49633ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 49733ad798cSAdam Langley (TCPOPT_WINDOW << 16) | 49833ad798cSAdam Langley (TCPOLEN_WINDOW << 8) | 49933ad798cSAdam Langley opts->ws); 50033ad798cSAdam Langley } 50133ad798cSAdam Langley 50233ad798cSAdam Langley if (unlikely(opts->num_sack_blocks)) { 50333ad798cSAdam Langley struct tcp_sack_block *sp = tp->rx_opt.dsack ? 50433ad798cSAdam Langley tp->duplicate_sack : tp->selective_acks; 50540efc6faSStephen Hemminger int this_sack; 50640efc6faSStephen Hemminger 50740efc6faSStephen Hemminger *ptr++ = htonl((TCPOPT_NOP << 24) | 50840efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 50940efc6faSStephen Hemminger (TCPOPT_SACK << 8) | 51033ad798cSAdam Langley (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * 51140efc6faSStephen Hemminger TCPOLEN_SACK_PERBLOCK))); 5122de979bdSStephen Hemminger 51333ad798cSAdam Langley for (this_sack = 0; this_sack < opts->num_sack_blocks; 51433ad798cSAdam Langley ++this_sack) { 51540efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].start_seq); 51640efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].end_seq); 51740efc6faSStephen Hemminger } 5182de979bdSStephen Hemminger 51940efc6faSStephen Hemminger tp->rx_opt.dsack = 0; 52040efc6faSStephen Hemminger } 5212100c8d2SYuchung Cheng 5222100c8d2SYuchung Cheng if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) { 5232100c8d2SYuchung Cheng struct tcp_fastopen_cookie *foc = opts->fastopen_cookie; 5247f9b838bSDaniel Lee u8 *p = (u8 *)ptr; 5257f9b838bSDaniel Lee u32 len; /* Fast Open option length */ 5262100c8d2SYuchung Cheng 5277f9b838bSDaniel Lee if (foc->exp) { 5287f9b838bSDaniel Lee len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len; 5297f9b838bSDaniel Lee *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) | 5302100c8d2SYuchung Cheng TCPOPT_FASTOPEN_MAGIC); 5317f9b838bSDaniel Lee p += TCPOLEN_EXP_FASTOPEN_BASE; 5327f9b838bSDaniel Lee } else { 5337f9b838bSDaniel Lee len = TCPOLEN_FASTOPEN_BASE + foc->len; 5347f9b838bSDaniel Lee *p++ = TCPOPT_FASTOPEN; 5357f9b838bSDaniel Lee *p++ = len; 5362100c8d2SYuchung Cheng } 5377f9b838bSDaniel Lee 5387f9b838bSDaniel Lee memcpy(p, foc->val, foc->len); 5397f9b838bSDaniel Lee if ((len & 3) == 2) { 5407f9b838bSDaniel Lee p[foc->len] = TCPOPT_NOP; 5417f9b838bSDaniel Lee p[foc->len + 1] = TCPOPT_NOP; 5427f9b838bSDaniel Lee } 5437f9b838bSDaniel Lee ptr += (len + 3) >> 2; 5442100c8d2SYuchung Cheng } 54540efc6faSStephen Hemminger } 54640efc6faSStephen Hemminger 54767edfef7SAndi Kleen /* Compute TCP options for SYN packets. This is not the final 54867edfef7SAndi Kleen * network wire format yet. 54967edfef7SAndi Kleen */ 55095c96174SEric Dumazet static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb, 55133ad798cSAdam Langley struct tcp_out_options *opts, 552cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 553cf533ea5SEric Dumazet { 55433ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 55595c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 556783237e8SYuchung Cheng struct tcp_fastopen_request *fastopen = tp->fastopen_req; 55733ad798cSAdam Langley 558cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 55933ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 56033ad798cSAdam Langley if (*md5) { 56133ad798cSAdam Langley opts->options |= OPTION_MD5; 562bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 563cfb6eeb4SYOSHIFUJI Hideaki } 56433ad798cSAdam Langley #else 56533ad798cSAdam Langley *md5 = NULL; 566cfb6eeb4SYOSHIFUJI Hideaki #endif 56733ad798cSAdam Langley 56833ad798cSAdam Langley /* We always get an MSS option. The option bytes which will be seen in 56933ad798cSAdam Langley * normal data packets should timestamps be used, must be in the MSS 57033ad798cSAdam Langley * advertised. But we subtract them from tp->mss_cache so that 57133ad798cSAdam Langley * calculations in tcp_sendmsg are simpler etc. So account for this 57233ad798cSAdam Langley * fact here if necessary. If we don't do this correctly, as a 57333ad798cSAdam Langley * receiver we won't recognize data packets as being full sized when we 57433ad798cSAdam Langley * should, and thus we won't abide by the delayed ACK rules correctly. 57533ad798cSAdam Langley * SACKs don't matter, we never delay an ACK when we have any of those 57633ad798cSAdam Langley * going out. */ 57733ad798cSAdam Langley opts->mss = tcp_advertise_mss(sk); 578bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 57933ad798cSAdam Langley 58051456b29SIan Morris if (likely(sysctl_tcp_timestamps && !*md5)) { 58133ad798cSAdam Langley opts->options |= OPTION_TS; 5827faee5c0SEric Dumazet opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset; 58333ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 584bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 58533ad798cSAdam Langley } 586bb5b7c11SDavid S. Miller if (likely(sysctl_tcp_window_scaling)) { 58733ad798cSAdam Langley opts->ws = tp->rx_opt.rcv_wscale; 58889e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 589bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 59033ad798cSAdam Langley } 591bb5b7c11SDavid S. Miller if (likely(sysctl_tcp_sack)) { 59233ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 593b32d1310SDavid S. Miller if (unlikely(!(OPTION_TS & opts->options))) 594bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 59533ad798cSAdam Langley } 59633ad798cSAdam Langley 597783237e8SYuchung Cheng if (fastopen && fastopen->cookie.len >= 0) { 5982646c831SDaniel Lee u32 need = fastopen->cookie.len; 5992646c831SDaniel Lee 6002646c831SDaniel Lee need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE : 6012646c831SDaniel Lee TCPOLEN_FASTOPEN_BASE; 602783237e8SYuchung Cheng need = (need + 3) & ~3U; /* Align to 32 bits */ 603783237e8SYuchung Cheng if (remaining >= need) { 604783237e8SYuchung Cheng opts->options |= OPTION_FAST_OPEN_COOKIE; 605783237e8SYuchung Cheng opts->fastopen_cookie = &fastopen->cookie; 606783237e8SYuchung Cheng remaining -= need; 607783237e8SYuchung Cheng tp->syn_fastopen = 1; 6082646c831SDaniel Lee tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0; 609783237e8SYuchung Cheng } 610783237e8SYuchung Cheng } 611bd0388aeSWilliam Allen Simpson 612bd0388aeSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 61333ad798cSAdam Langley } 61433ad798cSAdam Langley 61567edfef7SAndi Kleen /* Set up TCP options for SYN-ACKs. */ 61695c96174SEric Dumazet static unsigned int tcp_synack_options(struct sock *sk, 61733ad798cSAdam Langley struct request_sock *req, 61895c96174SEric Dumazet unsigned int mss, struct sk_buff *skb, 61933ad798cSAdam Langley struct tcp_out_options *opts, 62080f03e27SEric Dumazet const struct tcp_md5sig_key *md5, 6218336886fSJerry Chu struct tcp_fastopen_cookie *foc) 6224957faadSWilliam Allen Simpson { 62333ad798cSAdam Langley struct inet_request_sock *ireq = inet_rsk(req); 62495c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 62533ad798cSAdam Langley 62633ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 62780f03e27SEric Dumazet if (md5) { 62833ad798cSAdam Langley opts->options |= OPTION_MD5; 6294957faadSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 6304957faadSWilliam Allen Simpson 6314957faadSWilliam Allen Simpson /* We can't fit any SACK blocks in a packet with MD5 + TS 6324957faadSWilliam Allen Simpson * options. There was discussion about disabling SACK 6334957faadSWilliam Allen Simpson * rather than TS in order to fit in better with old, 6344957faadSWilliam Allen Simpson * buggy kernels, but that was deemed to be unnecessary. 6354957faadSWilliam Allen Simpson */ 636de213e5eSEric Dumazet ireq->tstamp_ok &= !ireq->sack_ok; 63733ad798cSAdam Langley } 63833ad798cSAdam Langley #endif 63933ad798cSAdam Langley 6404957faadSWilliam Allen Simpson /* We always send an MSS option. */ 64133ad798cSAdam Langley opts->mss = mss; 6424957faadSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 64333ad798cSAdam Langley 64433ad798cSAdam Langley if (likely(ireq->wscale_ok)) { 64533ad798cSAdam Langley opts->ws = ireq->rcv_wscale; 64689e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 6474957faadSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 64833ad798cSAdam Langley } 649de213e5eSEric Dumazet if (likely(ireq->tstamp_ok)) { 65033ad798cSAdam Langley opts->options |= OPTION_TS; 6517faee5c0SEric Dumazet opts->tsval = tcp_skb_timestamp(skb); 65233ad798cSAdam Langley opts->tsecr = req->ts_recent; 6534957faadSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 65433ad798cSAdam Langley } 65533ad798cSAdam Langley if (likely(ireq->sack_ok)) { 65633ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 657de213e5eSEric Dumazet if (unlikely(!ireq->tstamp_ok)) 6584957faadSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 65933ad798cSAdam Langley } 6607f9b838bSDaniel Lee if (foc != NULL && foc->len >= 0) { 6617f9b838bSDaniel Lee u32 need = foc->len; 6627f9b838bSDaniel Lee 6637f9b838bSDaniel Lee need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE : 6647f9b838bSDaniel Lee TCPOLEN_FASTOPEN_BASE; 6658336886fSJerry Chu need = (need + 3) & ~3U; /* Align to 32 bits */ 6668336886fSJerry Chu if (remaining >= need) { 6678336886fSJerry Chu opts->options |= OPTION_FAST_OPEN_COOKIE; 6688336886fSJerry Chu opts->fastopen_cookie = foc; 6698336886fSJerry Chu remaining -= need; 6708336886fSJerry Chu } 6718336886fSJerry Chu } 6724957faadSWilliam Allen Simpson 6734957faadSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 67433ad798cSAdam Langley } 67533ad798cSAdam Langley 67667edfef7SAndi Kleen /* Compute TCP options for ESTABLISHED sockets. This is not the 67767edfef7SAndi Kleen * final wire format yet. 67867edfef7SAndi Kleen */ 67995c96174SEric Dumazet static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb, 68033ad798cSAdam Langley struct tcp_out_options *opts, 681cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 682cf533ea5SEric Dumazet { 68333ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 68495c96174SEric Dumazet unsigned int size = 0; 685cabeccbdSIlpo Järvinen unsigned int eff_sacks; 68633ad798cSAdam Langley 6875843ef42SAndi Kleen opts->options = 0; 6885843ef42SAndi Kleen 68933ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 69033ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 69133ad798cSAdam Langley if (unlikely(*md5)) { 69233ad798cSAdam Langley opts->options |= OPTION_MD5; 69333ad798cSAdam Langley size += TCPOLEN_MD5SIG_ALIGNED; 69433ad798cSAdam Langley } 69533ad798cSAdam Langley #else 69633ad798cSAdam Langley *md5 = NULL; 69733ad798cSAdam Langley #endif 69833ad798cSAdam Langley 69933ad798cSAdam Langley if (likely(tp->rx_opt.tstamp_ok)) { 70033ad798cSAdam Langley opts->options |= OPTION_TS; 7017faee5c0SEric Dumazet opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0; 70233ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 70333ad798cSAdam Langley size += TCPOLEN_TSTAMP_ALIGNED; 70433ad798cSAdam Langley } 70533ad798cSAdam Langley 706cabeccbdSIlpo Järvinen eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; 707cabeccbdSIlpo Järvinen if (unlikely(eff_sacks)) { 70895c96174SEric Dumazet const unsigned int remaining = MAX_TCP_OPTION_SPACE - size; 70933ad798cSAdam Langley opts->num_sack_blocks = 71095c96174SEric Dumazet min_t(unsigned int, eff_sacks, 71133ad798cSAdam Langley (remaining - TCPOLEN_SACK_BASE_ALIGNED) / 71233ad798cSAdam Langley TCPOLEN_SACK_PERBLOCK); 71333ad798cSAdam Langley size += TCPOLEN_SACK_BASE_ALIGNED + 71433ad798cSAdam Langley opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; 71533ad798cSAdam Langley } 71633ad798cSAdam Langley 71733ad798cSAdam Langley return size; 71840efc6faSStephen Hemminger } 7191da177e4SLinus Torvalds 72046d3ceabSEric Dumazet 72146d3ceabSEric Dumazet /* TCP SMALL QUEUES (TSQ) 72246d3ceabSEric Dumazet * 72346d3ceabSEric Dumazet * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev) 72446d3ceabSEric Dumazet * to reduce RTT and bufferbloat. 72546d3ceabSEric Dumazet * We do this using a special skb destructor (tcp_wfree). 72646d3ceabSEric Dumazet * 72746d3ceabSEric Dumazet * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb 72846d3ceabSEric Dumazet * needs to be reallocated in a driver. 7298e3bff96Sstephen hemminger * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc 73046d3ceabSEric Dumazet * 73146d3ceabSEric Dumazet * Since transmit from skb destructor is forbidden, we use a tasklet 73246d3ceabSEric Dumazet * to process all sockets that eventually need to send more skbs. 73346d3ceabSEric Dumazet * We use one tasklet per cpu, with its own queue of sockets. 73446d3ceabSEric Dumazet */ 73546d3ceabSEric Dumazet struct tsq_tasklet { 73646d3ceabSEric Dumazet struct tasklet_struct tasklet; 73746d3ceabSEric Dumazet struct list_head head; /* queue of tcp sockets */ 73846d3ceabSEric Dumazet }; 73946d3ceabSEric Dumazet static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet); 74046d3ceabSEric Dumazet 7416f458dfbSEric Dumazet static void tcp_tsq_handler(struct sock *sk) 7426f458dfbSEric Dumazet { 7436f458dfbSEric Dumazet if ((1 << sk->sk_state) & 7446f458dfbSEric Dumazet (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING | 7456f458dfbSEric Dumazet TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) 746bf06200eSJohn Ogness tcp_write_xmit(sk, tcp_current_mss(sk), tcp_sk(sk)->nonagle, 747bf06200eSJohn Ogness 0, GFP_ATOMIC); 7486f458dfbSEric Dumazet } 74946d3ceabSEric Dumazet /* 7508e3bff96Sstephen hemminger * One tasklet per cpu tries to send more skbs. 75146d3ceabSEric Dumazet * We run in tasklet context but need to disable irqs when 7528e3bff96Sstephen hemminger * transferring tsq->head because tcp_wfree() might 75346d3ceabSEric Dumazet * interrupt us (non NAPI drivers) 75446d3ceabSEric Dumazet */ 75546d3ceabSEric Dumazet static void tcp_tasklet_func(unsigned long data) 75646d3ceabSEric Dumazet { 75746d3ceabSEric Dumazet struct tsq_tasklet *tsq = (struct tsq_tasklet *)data; 75846d3ceabSEric Dumazet LIST_HEAD(list); 75946d3ceabSEric Dumazet unsigned long flags; 76046d3ceabSEric Dumazet struct list_head *q, *n; 76146d3ceabSEric Dumazet struct tcp_sock *tp; 76246d3ceabSEric Dumazet struct sock *sk; 76346d3ceabSEric Dumazet 76446d3ceabSEric Dumazet local_irq_save(flags); 76546d3ceabSEric Dumazet list_splice_init(&tsq->head, &list); 76646d3ceabSEric Dumazet local_irq_restore(flags); 76746d3ceabSEric Dumazet 76846d3ceabSEric Dumazet list_for_each_safe(q, n, &list) { 76946d3ceabSEric Dumazet tp = list_entry(q, struct tcp_sock, tsq_node); 77046d3ceabSEric Dumazet list_del(&tp->tsq_node); 77146d3ceabSEric Dumazet 77246d3ceabSEric Dumazet sk = (struct sock *)tp; 77346d3ceabSEric Dumazet bh_lock_sock(sk); 77446d3ceabSEric Dumazet 77546d3ceabSEric Dumazet if (!sock_owned_by_user(sk)) { 7766f458dfbSEric Dumazet tcp_tsq_handler(sk); 77746d3ceabSEric Dumazet } else { 77846d3ceabSEric Dumazet /* defer the work to tcp_release_cb() */ 7796f458dfbSEric Dumazet set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags); 78046d3ceabSEric Dumazet } 78146d3ceabSEric Dumazet bh_unlock_sock(sk); 78246d3ceabSEric Dumazet 78346d3ceabSEric Dumazet clear_bit(TSQ_QUEUED, &tp->tsq_flags); 78446d3ceabSEric Dumazet sk_free(sk); 78546d3ceabSEric Dumazet } 78646d3ceabSEric Dumazet } 78746d3ceabSEric Dumazet 7886f458dfbSEric Dumazet #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \ 7896f458dfbSEric Dumazet (1UL << TCP_WRITE_TIMER_DEFERRED) | \ 790563d34d0SEric Dumazet (1UL << TCP_DELACK_TIMER_DEFERRED) | \ 791563d34d0SEric Dumazet (1UL << TCP_MTU_REDUCED_DEFERRED)) 79246d3ceabSEric Dumazet /** 79346d3ceabSEric Dumazet * tcp_release_cb - tcp release_sock() callback 79446d3ceabSEric Dumazet * @sk: socket 79546d3ceabSEric Dumazet * 79646d3ceabSEric Dumazet * called from release_sock() to perform protocol dependent 79746d3ceabSEric Dumazet * actions before socket release. 79846d3ceabSEric Dumazet */ 79946d3ceabSEric Dumazet void tcp_release_cb(struct sock *sk) 80046d3ceabSEric Dumazet { 80146d3ceabSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 8026f458dfbSEric Dumazet unsigned long flags, nflags; 80346d3ceabSEric Dumazet 8046f458dfbSEric Dumazet /* perform an atomic operation only if at least one flag is set */ 8056f458dfbSEric Dumazet do { 8066f458dfbSEric Dumazet flags = tp->tsq_flags; 8076f458dfbSEric Dumazet if (!(flags & TCP_DEFERRED_ALL)) 8086f458dfbSEric Dumazet return; 8096f458dfbSEric Dumazet nflags = flags & ~TCP_DEFERRED_ALL; 8106f458dfbSEric Dumazet } while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags); 8116f458dfbSEric Dumazet 8126f458dfbSEric Dumazet if (flags & (1UL << TCP_TSQ_DEFERRED)) 8136f458dfbSEric Dumazet tcp_tsq_handler(sk); 8146f458dfbSEric Dumazet 815c3f9b018SEric Dumazet /* Here begins the tricky part : 816c3f9b018SEric Dumazet * We are called from release_sock() with : 817c3f9b018SEric Dumazet * 1) BH disabled 818c3f9b018SEric Dumazet * 2) sk_lock.slock spinlock held 819c3f9b018SEric Dumazet * 3) socket owned by us (sk->sk_lock.owned == 1) 820c3f9b018SEric Dumazet * 821c3f9b018SEric Dumazet * But following code is meant to be called from BH handlers, 822c3f9b018SEric Dumazet * so we should keep BH disabled, but early release socket ownership 823c3f9b018SEric Dumazet */ 824c3f9b018SEric Dumazet sock_release_ownership(sk); 825c3f9b018SEric Dumazet 826144d56e9SEric Dumazet if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) { 8276f458dfbSEric Dumazet tcp_write_timer_handler(sk); 828144d56e9SEric Dumazet __sock_put(sk); 829144d56e9SEric Dumazet } 830144d56e9SEric Dumazet if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) { 8316f458dfbSEric Dumazet tcp_delack_timer_handler(sk); 832144d56e9SEric Dumazet __sock_put(sk); 833144d56e9SEric Dumazet } 834144d56e9SEric Dumazet if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) { 8354fab9071SNeal Cardwell inet_csk(sk)->icsk_af_ops->mtu_reduced(sk); 836144d56e9SEric Dumazet __sock_put(sk); 837144d56e9SEric Dumazet } 83846d3ceabSEric Dumazet } 83946d3ceabSEric Dumazet EXPORT_SYMBOL(tcp_release_cb); 84046d3ceabSEric Dumazet 84146d3ceabSEric Dumazet void __init tcp_tasklet_init(void) 84246d3ceabSEric Dumazet { 84346d3ceabSEric Dumazet int i; 84446d3ceabSEric Dumazet 84546d3ceabSEric Dumazet for_each_possible_cpu(i) { 84646d3ceabSEric Dumazet struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i); 84746d3ceabSEric Dumazet 84846d3ceabSEric Dumazet INIT_LIST_HEAD(&tsq->head); 84946d3ceabSEric Dumazet tasklet_init(&tsq->tasklet, 85046d3ceabSEric Dumazet tcp_tasklet_func, 85146d3ceabSEric Dumazet (unsigned long)tsq); 85246d3ceabSEric Dumazet } 85346d3ceabSEric Dumazet } 85446d3ceabSEric Dumazet 85546d3ceabSEric Dumazet /* 85646d3ceabSEric Dumazet * Write buffer destructor automatically called from kfree_skb. 8578e3bff96Sstephen hemminger * We can't xmit new skbs from this context, as we might already 85846d3ceabSEric Dumazet * hold qdisc lock. 85946d3ceabSEric Dumazet */ 860d6a4a104SEric Dumazet void tcp_wfree(struct sk_buff *skb) 86146d3ceabSEric Dumazet { 86246d3ceabSEric Dumazet struct sock *sk = skb->sk; 86346d3ceabSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 8649b462d02SEric Dumazet int wmem; 8659b462d02SEric Dumazet 8669b462d02SEric Dumazet /* Keep one reference on sk_wmem_alloc. 8679b462d02SEric Dumazet * Will be released by sk_free() from here or tcp_tasklet_func() 8689b462d02SEric Dumazet */ 8699b462d02SEric Dumazet wmem = atomic_sub_return(skb->truesize - 1, &sk->sk_wmem_alloc); 8709b462d02SEric Dumazet 8719b462d02SEric Dumazet /* If this softirq is serviced by ksoftirqd, we are likely under stress. 8729b462d02SEric Dumazet * Wait until our queues (qdisc + devices) are drained. 8739b462d02SEric Dumazet * This gives : 8749b462d02SEric Dumazet * - less callbacks to tcp_write_xmit(), reducing stress (batches) 8759b462d02SEric Dumazet * - chance for incoming ACK (processed by another cpu maybe) 8769b462d02SEric Dumazet * to migrate this flow (skb->ooo_okay will be eventually set) 8779b462d02SEric Dumazet */ 8789b462d02SEric Dumazet if (wmem >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current) 8799b462d02SEric Dumazet goto out; 88046d3ceabSEric Dumazet 88146d3ceabSEric Dumazet if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) && 88246d3ceabSEric Dumazet !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) { 88346d3ceabSEric Dumazet unsigned long flags; 88446d3ceabSEric Dumazet struct tsq_tasklet *tsq; 88546d3ceabSEric Dumazet 88646d3ceabSEric Dumazet /* queue this socket to tasklet queue */ 88746d3ceabSEric Dumazet local_irq_save(flags); 888903ceff7SChristoph Lameter tsq = this_cpu_ptr(&tsq_tasklet); 88946d3ceabSEric Dumazet list_add(&tp->tsq_node, &tsq->head); 89046d3ceabSEric Dumazet tasklet_schedule(&tsq->tasklet); 89146d3ceabSEric Dumazet local_irq_restore(flags); 8929b462d02SEric Dumazet return; 89346d3ceabSEric Dumazet } 8949b462d02SEric Dumazet out: 8959b462d02SEric Dumazet sk_free(sk); 89646d3ceabSEric Dumazet } 89746d3ceabSEric Dumazet 8981da177e4SLinus Torvalds /* This routine actually transmits TCP packets queued in by 8991da177e4SLinus Torvalds * tcp_do_sendmsg(). This is used by both the initial 9001da177e4SLinus Torvalds * transmission and possible later retransmissions. 9011da177e4SLinus Torvalds * All SKB's seen here are completely headerless. It is our 9021da177e4SLinus Torvalds * job to build the TCP header, and pass the packet down to 9031da177e4SLinus Torvalds * IP so it can do the same plus pass the packet off to the 9041da177e4SLinus Torvalds * device. 9051da177e4SLinus Torvalds * 9061da177e4SLinus Torvalds * We are working here with either a clone of the original 9071da177e4SLinus Torvalds * SKB, or a fresh unique copy made by the retransmit engine. 9081da177e4SLinus Torvalds */ 909056834d9SIlpo Järvinen static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, 910056834d9SIlpo Järvinen gfp_t gfp_mask) 9111da177e4SLinus Torvalds { 9126687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 913dfb4b9dcSDavid S. Miller struct inet_sock *inet; 914dfb4b9dcSDavid S. Miller struct tcp_sock *tp; 915dfb4b9dcSDavid S. Miller struct tcp_skb_cb *tcb; 91633ad798cSAdam Langley struct tcp_out_options opts; 91795c96174SEric Dumazet unsigned int tcp_options_size, tcp_header_size; 918cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 9191da177e4SLinus Torvalds struct tcphdr *th; 9201da177e4SLinus Torvalds int err; 9211da177e4SLinus Torvalds 922dfb4b9dcSDavid S. Miller BUG_ON(!skb || !tcp_skb_pcount(skb)); 923dfb4b9dcSDavid S. Miller 924ccdbb6e9SEric Dumazet if (clone_it) { 925740b0f18SEric Dumazet skb_mstamp_get(&skb->skb_mstamp); 926dfb4b9dcSDavid S. Miller 927dfb4b9dcSDavid S. Miller if (unlikely(skb_cloned(skb))) 928dfb4b9dcSDavid S. Miller skb = pskb_copy(skb, gfp_mask); 929dfb4b9dcSDavid S. Miller else 930dfb4b9dcSDavid S. Miller skb = skb_clone(skb, gfp_mask); 931dfb4b9dcSDavid S. Miller if (unlikely(!skb)) 932dfb4b9dcSDavid S. Miller return -ENOBUFS; 933dfb4b9dcSDavid S. Miller } 934dfb4b9dcSDavid S. Miller 935dfb4b9dcSDavid S. Miller inet = inet_sk(sk); 936dfb4b9dcSDavid S. Miller tp = tcp_sk(sk); 937dfb4b9dcSDavid S. Miller tcb = TCP_SKB_CB(skb); 93833ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 9391da177e4SLinus Torvalds 9404de075e0SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) 94133ad798cSAdam Langley tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); 94233ad798cSAdam Langley else 94333ad798cSAdam Langley tcp_options_size = tcp_established_options(sk, skb, &opts, 94433ad798cSAdam Langley &md5); 94533ad798cSAdam Langley tcp_header_size = tcp_options_size + sizeof(struct tcphdr); 9461da177e4SLinus Torvalds 947547669d4SEric Dumazet if (tcp_packets_in_flight(tp) == 0) 9486687e988SArnaldo Carvalho de Melo tcp_ca_event(sk, CA_EVENT_TX_START); 949547669d4SEric Dumazet 950547669d4SEric Dumazet /* if no packet is in qdisc/device queue, then allow XPS to select 951b2532eb9SEric Dumazet * another queue. We can be called from tcp_tsq_handler() 952b2532eb9SEric Dumazet * which holds one reference to sk_wmem_alloc. 953b2532eb9SEric Dumazet * 954b2532eb9SEric Dumazet * TODO: Ideally, in-flight pure ACK packets should not matter here. 955b2532eb9SEric Dumazet * One way to get this would be to set skb->truesize = 2 on them. 956547669d4SEric Dumazet */ 957b2532eb9SEric Dumazet skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1); 9581da177e4SLinus Torvalds 959aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 960aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 96146d3ceabSEric Dumazet 96246d3ceabSEric Dumazet skb_orphan(skb); 96346d3ceabSEric Dumazet skb->sk = sk; 96498781965SEric Dumazet skb->destructor = skb_is_tcp_pure_ack(skb) ? sock_wfree : tcp_wfree; 965b73c3d0eSTom Herbert skb_set_hash_from_sk(skb, sk); 96646d3ceabSEric Dumazet atomic_add(skb->truesize, &sk->sk_wmem_alloc); 9671da177e4SLinus Torvalds 9681da177e4SLinus Torvalds /* Build TCP header and checksum it. */ 969aa8223c7SArnaldo Carvalho de Melo th = tcp_hdr(skb); 970c720c7e8SEric Dumazet th->source = inet->inet_sport; 971c720c7e8SEric Dumazet th->dest = inet->inet_dport; 9721da177e4SLinus Torvalds th->seq = htonl(tcb->seq); 9731da177e4SLinus Torvalds th->ack_seq = htonl(tp->rcv_nxt); 974df7a3b07SAl Viro *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | 9754de075e0SEric Dumazet tcb->tcp_flags); 976dfb4b9dcSDavid S. Miller 9774de075e0SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) { 9781da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments 9791da177e4SLinus Torvalds * is never scaled. 9801da177e4SLinus Torvalds */ 981600ff0c2SIlpo Järvinen th->window = htons(min(tp->rcv_wnd, 65535U)); 9821da177e4SLinus Torvalds } else { 9831da177e4SLinus Torvalds th->window = htons(tcp_select_window(sk)); 9841da177e4SLinus Torvalds } 9851da177e4SLinus Torvalds th->check = 0; 9861da177e4SLinus Torvalds th->urg_ptr = 0; 9871da177e4SLinus Torvalds 98833f5f57eSIlpo Järvinen /* The urg_mode check is necessary during a below snd_una win probe */ 9897691367dSHerbert Xu if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { 9907691367dSHerbert Xu if (before(tp->snd_up, tcb->seq + 0x10000)) { 9911da177e4SLinus Torvalds th->urg_ptr = htons(tp->snd_up - tcb->seq); 9921da177e4SLinus Torvalds th->urg = 1; 9937691367dSHerbert Xu } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { 9940eae88f3SEric Dumazet th->urg_ptr = htons(0xFFFF); 9957691367dSHerbert Xu th->urg = 1; 9967691367dSHerbert Xu } 9971da177e4SLinus Torvalds } 9981da177e4SLinus Torvalds 999bd0388aeSWilliam Allen Simpson tcp_options_write((__be32 *)(th + 1), tp, &opts); 100051466a75SEric Dumazet skb_shinfo(skb)->gso_type = sk->sk_gso_type; 10014de075e0SEric Dumazet if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0)) 1002735d3831SFlorian Westphal tcp_ecn_send(sk, skb, tcp_header_size); 1003dfb4b9dcSDavid S. Miller 1004cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 1005cfb6eeb4SYOSHIFUJI Hideaki /* Calculate the MD5 hash, as we have all we need now */ 1006cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 1007a465419bSEric Dumazet sk_nocaps_add(sk, NETIF_F_GSO_MASK); 1008bd0388aeSWilliam Allen Simpson tp->af_specific->calc_md5_hash(opts.hash_location, 100939f8e58eSEric Dumazet md5, sk, skb); 1010cfb6eeb4SYOSHIFUJI Hideaki } 1011cfb6eeb4SYOSHIFUJI Hideaki #endif 1012cfb6eeb4SYOSHIFUJI Hideaki 1013bb296246SHerbert Xu icsk->icsk_af_ops->send_check(sk, skb); 10141da177e4SLinus Torvalds 10154de075e0SEric Dumazet if (likely(tcb->tcp_flags & TCPHDR_ACK)) 1016fc6415bcSDavid S. Miller tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); 10171da177e4SLinus Torvalds 10181da177e4SLinus Torvalds if (skb->len != tcp_header_size) 1019cf533ea5SEric Dumazet tcp_event_data_sent(tp, sk); 10201da177e4SLinus Torvalds 1021bd37a088SWei Yongjun if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) 1022aa2ea058STom Herbert TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, 1023aa2ea058STom Herbert tcp_skb_pcount(skb)); 10241da177e4SLinus Torvalds 10252efd055cSMarcelo Ricardo Leitner tp->segs_out += tcp_skb_pcount(skb); 1026f69ad292SEric Dumazet /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */ 1027cd7d8498SEric Dumazet skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb); 1028f69ad292SEric Dumazet skb_shinfo(skb)->gso_size = tcp_skb_mss(skb); 1029cd7d8498SEric Dumazet 10307faee5c0SEric Dumazet /* Our usage of tstamp should remain private */ 10317faee5c0SEric Dumazet skb->tstamp.tv64 = 0; 1032971f10ecSEric Dumazet 1033971f10ecSEric Dumazet /* Cleanup our debris for IP stacks */ 1034971f10ecSEric Dumazet memset(skb->cb, 0, max(sizeof(struct inet_skb_parm), 1035971f10ecSEric Dumazet sizeof(struct inet6_skb_parm))); 1036971f10ecSEric Dumazet 1037b0270e91SEric Dumazet err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl); 10387faee5c0SEric Dumazet 103983de47cdSHua Zhong if (likely(err <= 0)) 10401da177e4SLinus Torvalds return err; 10411da177e4SLinus Torvalds 10425ee2c941SChristoph Paasch tcp_enter_cwr(sk); 10431da177e4SLinus Torvalds 1044b9df3cb8SGerrit Renker return net_xmit_eval(err); 10451da177e4SLinus Torvalds } 10461da177e4SLinus Torvalds 104767edfef7SAndi Kleen /* This routine just queues the buffer for sending. 10481da177e4SLinus Torvalds * 10491da177e4SLinus Torvalds * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, 10501da177e4SLinus Torvalds * otherwise socket can stall. 10511da177e4SLinus Torvalds */ 10521da177e4SLinus Torvalds static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) 10531da177e4SLinus Torvalds { 10541da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 10551da177e4SLinus Torvalds 10561da177e4SLinus Torvalds /* Advance write_seq and place onto the write_queue. */ 10571da177e4SLinus Torvalds tp->write_seq = TCP_SKB_CB(skb)->end_seq; 1058f4a775d1SEric Dumazet __skb_header_release(skb); 1059fe067e8aSDavid S. Miller tcp_add_write_queue_tail(sk, skb); 10603ab224beSHideo Aoki sk->sk_wmem_queued += skb->truesize; 10613ab224beSHideo Aoki sk_mem_charge(sk, skb->truesize); 10621da177e4SLinus Torvalds } 10631da177e4SLinus Torvalds 106467edfef7SAndi Kleen /* Initialize TSO segments for a packet. */ 10655bbb432cSEric Dumazet static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now) 1066f6302d1dSDavid S. Miller { 10678f26fb1cSEric Dumazet if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) { 1068f6302d1dSDavid S. Miller /* Avoid the costly divide in the normal 1069f6302d1dSDavid S. Miller * non-TSO case. 1070f6302d1dSDavid S. Miller */ 1071cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, 1); 1072f69ad292SEric Dumazet TCP_SKB_CB(skb)->tcp_gso_size = 0; 1073f6302d1dSDavid S. Miller } else { 1074cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now)); 1075f69ad292SEric Dumazet TCP_SKB_CB(skb)->tcp_gso_size = mss_now; 10761da177e4SLinus Torvalds } 10771da177e4SLinus Torvalds } 10781da177e4SLinus Torvalds 107991fed7a1SIlpo Järvinen /* When a modification to fackets out becomes necessary, we need to check 108068f8353bSIlpo Järvinen * skb is counted to fackets_out or not. 108191fed7a1SIlpo Järvinen */ 1082cf533ea5SEric Dumazet static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb, 108391fed7a1SIlpo Järvinen int decr) 108491fed7a1SIlpo Järvinen { 1085a47e5a98SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1086a47e5a98SIlpo Järvinen 1087dc86967bSIlpo Järvinen if (!tp->sacked_out || tcp_is_reno(tp)) 108891fed7a1SIlpo Järvinen return; 108991fed7a1SIlpo Järvinen 10906859d494SIlpo Järvinen if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq)) 109191fed7a1SIlpo Järvinen tp->fackets_out -= decr; 109291fed7a1SIlpo Järvinen } 109391fed7a1SIlpo Järvinen 1094797108d1SIlpo Järvinen /* Pcount in the middle of the write queue got changed, we need to do various 1095797108d1SIlpo Järvinen * tweaks to fix counters 1096797108d1SIlpo Järvinen */ 1097cf533ea5SEric Dumazet static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) 1098797108d1SIlpo Järvinen { 1099797108d1SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1100797108d1SIlpo Järvinen 1101797108d1SIlpo Järvinen tp->packets_out -= decr; 1102797108d1SIlpo Järvinen 1103797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 1104797108d1SIlpo Järvinen tp->sacked_out -= decr; 1105797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 1106797108d1SIlpo Järvinen tp->retrans_out -= decr; 1107797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 1108797108d1SIlpo Järvinen tp->lost_out -= decr; 1109797108d1SIlpo Järvinen 1110797108d1SIlpo Järvinen /* Reno case is special. Sigh... */ 1111797108d1SIlpo Järvinen if (tcp_is_reno(tp) && decr > 0) 1112797108d1SIlpo Järvinen tp->sacked_out -= min_t(u32, tp->sacked_out, decr); 1113797108d1SIlpo Järvinen 1114797108d1SIlpo Järvinen tcp_adjust_fackets_out(sk, skb, decr); 1115797108d1SIlpo Järvinen 1116797108d1SIlpo Järvinen if (tp->lost_skb_hint && 1117797108d1SIlpo Järvinen before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && 111852cf3cc8SIlpo Järvinen (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))) 1119797108d1SIlpo Järvinen tp->lost_cnt_hint -= decr; 1120797108d1SIlpo Järvinen 1121797108d1SIlpo Järvinen tcp_verify_left_out(tp); 1122797108d1SIlpo Järvinen } 1123797108d1SIlpo Järvinen 1124490cc7d0SWillem de Bruijn static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2) 1125490cc7d0SWillem de Bruijn { 1126490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo = skb_shinfo(skb); 1127490cc7d0SWillem de Bruijn 1128490cc7d0SWillem de Bruijn if (unlikely(shinfo->tx_flags & SKBTX_ANY_TSTAMP) && 1129490cc7d0SWillem de Bruijn !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) { 1130490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo2 = skb_shinfo(skb2); 1131490cc7d0SWillem de Bruijn u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP; 1132490cc7d0SWillem de Bruijn 1133490cc7d0SWillem de Bruijn shinfo->tx_flags &= ~tsflags; 1134490cc7d0SWillem de Bruijn shinfo2->tx_flags |= tsflags; 1135490cc7d0SWillem de Bruijn swap(shinfo->tskey, shinfo2->tskey); 1136490cc7d0SWillem de Bruijn } 1137490cc7d0SWillem de Bruijn } 1138490cc7d0SWillem de Bruijn 11391da177e4SLinus Torvalds /* Function to create two new TCP segments. Shrinks the given segment 11401da177e4SLinus Torvalds * to the specified size and appends a new segment with the rest of the 11411da177e4SLinus Torvalds * packet to the list. This won't be called frequently, I hope. 11421da177e4SLinus Torvalds * Remember, these are still headerless SKBs at this point. 11431da177e4SLinus Torvalds */ 1144056834d9SIlpo Järvinen int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, 11456cc55e09SOctavian Purdila unsigned int mss_now, gfp_t gfp) 11461da177e4SLinus Torvalds { 11471da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 11481da177e4SLinus Torvalds struct sk_buff *buff; 11496475be16SDavid S. Miller int nsize, old_factor; 1150b60b49eaSHerbert Xu int nlen; 11519ce01461SIlpo Järvinen u8 flags; 11521da177e4SLinus Torvalds 11532fceec13SIlpo Järvinen if (WARN_ON(len > skb->len)) 11542fceec13SIlpo Järvinen return -EINVAL; 11556a438bbeSStephen Hemminger 11561da177e4SLinus Torvalds nsize = skb_headlen(skb) - len; 11571da177e4SLinus Torvalds if (nsize < 0) 11581da177e4SLinus Torvalds nsize = 0; 11591da177e4SLinus Torvalds 11606cc55e09SOctavian Purdila if (skb_unclone(skb, gfp)) 11611da177e4SLinus Torvalds return -ENOMEM; 11621da177e4SLinus Torvalds 11631da177e4SLinus Torvalds /* Get a new skb... force flag on. */ 1164eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, nsize, gfp, true); 116551456b29SIan Morris if (!buff) 11661da177e4SLinus Torvalds return -ENOMEM; /* We'll just try again later. */ 1167ef5cb973SHerbert Xu 11683ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 11693ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1170b60b49eaSHerbert Xu nlen = skb->len - len - nsize; 1171b60b49eaSHerbert Xu buff->truesize += nlen; 1172b60b49eaSHerbert Xu skb->truesize -= nlen; 11731da177e4SLinus Torvalds 11741da177e4SLinus Torvalds /* Correct the sequence numbers. */ 11751da177e4SLinus Torvalds TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 11761da177e4SLinus Torvalds TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 11771da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 11781da177e4SLinus Torvalds 11791da177e4SLinus Torvalds /* PSH and FIN should only be set in the second packet. */ 11804de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 11814de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 11824de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1183e14c3cafSHerbert Xu TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; 11841da177e4SLinus Torvalds 118584fa7933SPatrick McHardy if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { 11861da177e4SLinus Torvalds /* Copy and checksum data tail into the new buffer. */ 1187056834d9SIlpo Järvinen buff->csum = csum_partial_copy_nocheck(skb->data + len, 1188056834d9SIlpo Järvinen skb_put(buff, nsize), 11891da177e4SLinus Torvalds nsize, 0); 11901da177e4SLinus Torvalds 11911da177e4SLinus Torvalds skb_trim(skb, len); 11921da177e4SLinus Torvalds 11931da177e4SLinus Torvalds skb->csum = csum_block_sub(skb->csum, buff->csum, len); 11941da177e4SLinus Torvalds } else { 119584fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 11961da177e4SLinus Torvalds skb_split(skb, buff, len); 11971da177e4SLinus Torvalds } 11981da177e4SLinus Torvalds 11991da177e4SLinus Torvalds buff->ip_summed = skb->ip_summed; 12001da177e4SLinus Torvalds 1201a61bbcf2SPatrick McHardy buff->tstamp = skb->tstamp; 1202490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 12031da177e4SLinus Torvalds 12046475be16SDavid S. Miller old_factor = tcp_skb_pcount(skb); 12056475be16SDavid S. Miller 12061da177e4SLinus Torvalds /* Fix up tso_factor for both original and new SKB. */ 12075bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 12085bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 12091da177e4SLinus Torvalds 12106475be16SDavid S. Miller /* If this packet has been sent out already, we must 12116475be16SDavid S. Miller * adjust the various packet counters. 12126475be16SDavid S. Miller */ 1213cf0b450cSHerbert Xu if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 12146475be16SDavid S. Miller int diff = old_factor - tcp_skb_pcount(skb) - 12156475be16SDavid S. Miller tcp_skb_pcount(buff); 12161da177e4SLinus Torvalds 1217797108d1SIlpo Järvinen if (diff) 1218797108d1SIlpo Järvinen tcp_adjust_pcount(sk, skb, diff); 12191da177e4SLinus Torvalds } 12201da177e4SLinus Torvalds 12211da177e4SLinus Torvalds /* Link BUFF into the send queue. */ 1222f4a775d1SEric Dumazet __skb_header_release(buff); 1223fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 12241da177e4SLinus Torvalds 12251da177e4SLinus Torvalds return 0; 12261da177e4SLinus Torvalds } 12271da177e4SLinus Torvalds 12281da177e4SLinus Torvalds /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c 12291da177e4SLinus Torvalds * eventually). The difference is that pulled data not copied, but 12301da177e4SLinus Torvalds * immediately discarded. 12311da177e4SLinus Torvalds */ 1232f2911969SHerbert Xu ~{PmVHI~} static void __pskb_trim_head(struct sk_buff *skb, int len) 12331da177e4SLinus Torvalds { 12347b7fc97aSEric Dumazet struct skb_shared_info *shinfo; 12351da177e4SLinus Torvalds int i, k, eat; 12361da177e4SLinus Torvalds 12374fa48bf3SEric Dumazet eat = min_t(int, len, skb_headlen(skb)); 12384fa48bf3SEric Dumazet if (eat) { 12394fa48bf3SEric Dumazet __skb_pull(skb, eat); 12404fa48bf3SEric Dumazet len -= eat; 12414fa48bf3SEric Dumazet if (!len) 12424fa48bf3SEric Dumazet return; 12434fa48bf3SEric Dumazet } 12441da177e4SLinus Torvalds eat = len; 12451da177e4SLinus Torvalds k = 0; 12467b7fc97aSEric Dumazet shinfo = skb_shinfo(skb); 12477b7fc97aSEric Dumazet for (i = 0; i < shinfo->nr_frags; i++) { 12487b7fc97aSEric Dumazet int size = skb_frag_size(&shinfo->frags[i]); 12499e903e08SEric Dumazet 12509e903e08SEric Dumazet if (size <= eat) { 1251aff65da0SIan Campbell skb_frag_unref(skb, i); 12529e903e08SEric Dumazet eat -= size; 12531da177e4SLinus Torvalds } else { 12547b7fc97aSEric Dumazet shinfo->frags[k] = shinfo->frags[i]; 12551da177e4SLinus Torvalds if (eat) { 12567b7fc97aSEric Dumazet shinfo->frags[k].page_offset += eat; 12577b7fc97aSEric Dumazet skb_frag_size_sub(&shinfo->frags[k], eat); 12581da177e4SLinus Torvalds eat = 0; 12591da177e4SLinus Torvalds } 12601da177e4SLinus Torvalds k++; 12611da177e4SLinus Torvalds } 12621da177e4SLinus Torvalds } 12637b7fc97aSEric Dumazet shinfo->nr_frags = k; 12641da177e4SLinus Torvalds 126527a884dcSArnaldo Carvalho de Melo skb_reset_tail_pointer(skb); 12661da177e4SLinus Torvalds skb->data_len -= len; 12671da177e4SLinus Torvalds skb->len = skb->data_len; 12681da177e4SLinus Torvalds } 12691da177e4SLinus Torvalds 127067edfef7SAndi Kleen /* Remove acked data from a packet in the transmit queue. */ 12711da177e4SLinus Torvalds int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 12721da177e4SLinus Torvalds { 127314bbd6a5SPravin B Shelar if (skb_unclone(skb, GFP_ATOMIC)) 12741da177e4SLinus Torvalds return -ENOMEM; 12751da177e4SLinus Torvalds 12764fa48bf3SEric Dumazet __pskb_trim_head(skb, len); 12771da177e4SLinus Torvalds 12781da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq += len; 127984fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 12801da177e4SLinus Torvalds 12811da177e4SLinus Torvalds skb->truesize -= len; 12821da177e4SLinus Torvalds sk->sk_wmem_queued -= len; 12833ab224beSHideo Aoki sk_mem_uncharge(sk, len); 12841da177e4SLinus Torvalds sock_set_flag(sk, SOCK_QUEUE_SHRUNK); 12851da177e4SLinus Torvalds 12865b35e1e6SNeal Cardwell /* Any change of skb->len requires recalculation of tso factor. */ 12871da177e4SLinus Torvalds if (tcp_skb_pcount(skb) > 1) 12885bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb)); 12891da177e4SLinus Torvalds 12901da177e4SLinus Torvalds return 0; 12911da177e4SLinus Torvalds } 12921da177e4SLinus Torvalds 12931b63edd6SYuchung Cheng /* Calculate MSS not accounting any TCP options. */ 12941b63edd6SYuchung Cheng static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu) 12955d424d5aSJohn Heffner { 1296cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1297cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 12985d424d5aSJohn Heffner int mss_now; 12995d424d5aSJohn Heffner 13005d424d5aSJohn Heffner /* Calculate base mss without TCP options: 13015d424d5aSJohn Heffner It is MMS_S - sizeof(tcphdr) of rfc1122 13025d424d5aSJohn Heffner */ 13035d424d5aSJohn Heffner mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 13045d424d5aSJohn Heffner 130567469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 130667469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 130767469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 130867469601SEric Dumazet 130967469601SEric Dumazet if (dst && dst_allfrag(dst)) 131067469601SEric Dumazet mss_now -= icsk->icsk_af_ops->net_frag_header_len; 131167469601SEric Dumazet } 131267469601SEric Dumazet 13135d424d5aSJohn Heffner /* Clamp it (mss_clamp does not include tcp options) */ 13145d424d5aSJohn Heffner if (mss_now > tp->rx_opt.mss_clamp) 13155d424d5aSJohn Heffner mss_now = tp->rx_opt.mss_clamp; 13165d424d5aSJohn Heffner 13175d424d5aSJohn Heffner /* Now subtract optional transport overhead */ 13185d424d5aSJohn Heffner mss_now -= icsk->icsk_ext_hdr_len; 13195d424d5aSJohn Heffner 13205d424d5aSJohn Heffner /* Then reserve room for full set of TCP options and 8 bytes of data */ 13215d424d5aSJohn Heffner if (mss_now < 48) 13225d424d5aSJohn Heffner mss_now = 48; 13235d424d5aSJohn Heffner return mss_now; 13245d424d5aSJohn Heffner } 13255d424d5aSJohn Heffner 13261b63edd6SYuchung Cheng /* Calculate MSS. Not accounting for SACKs here. */ 13271b63edd6SYuchung Cheng int tcp_mtu_to_mss(struct sock *sk, int pmtu) 13281b63edd6SYuchung Cheng { 13291b63edd6SYuchung Cheng /* Subtract TCP options size, not including SACKs */ 13301b63edd6SYuchung Cheng return __tcp_mtu_to_mss(sk, pmtu) - 13311b63edd6SYuchung Cheng (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr)); 13321b63edd6SYuchung Cheng } 13331b63edd6SYuchung Cheng 13345d424d5aSJohn Heffner /* Inverse of above */ 133567469601SEric Dumazet int tcp_mss_to_mtu(struct sock *sk, int mss) 13365d424d5aSJohn Heffner { 1337cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1338cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 13395d424d5aSJohn Heffner int mtu; 13405d424d5aSJohn Heffner 13415d424d5aSJohn Heffner mtu = mss + 13425d424d5aSJohn Heffner tp->tcp_header_len + 13435d424d5aSJohn Heffner icsk->icsk_ext_hdr_len + 13445d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 13455d424d5aSJohn Heffner 134667469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 134767469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 134867469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 134967469601SEric Dumazet 135067469601SEric Dumazet if (dst && dst_allfrag(dst)) 135167469601SEric Dumazet mtu += icsk->icsk_af_ops->net_frag_header_len; 135267469601SEric Dumazet } 13535d424d5aSJohn Heffner return mtu; 13545d424d5aSJohn Heffner } 13555d424d5aSJohn Heffner 135667edfef7SAndi Kleen /* MTU probing init per socket */ 13575d424d5aSJohn Heffner void tcp_mtup_init(struct sock *sk) 13585d424d5aSJohn Heffner { 13595d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 13605d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 1361b0f9ca53SFan Du struct net *net = sock_net(sk); 13625d424d5aSJohn Heffner 1363b0f9ca53SFan Du icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1; 13645d424d5aSJohn Heffner icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 13655d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 1366b0f9ca53SFan Du icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss); 13675d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 136805cbc0dbSFan Du if (icsk->icsk_mtup.enabled) 136905cbc0dbSFan Du icsk->icsk_mtup.probe_timestamp = tcp_time_stamp; 13705d424d5aSJohn Heffner } 13714bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_mtup_init); 13725d424d5aSJohn Heffner 13731da177e4SLinus Torvalds /* This function synchronize snd mss to current pmtu/exthdr set. 13741da177e4SLinus Torvalds 13751da177e4SLinus Torvalds tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 13761da177e4SLinus Torvalds for TCP options, but includes only bare TCP header. 13771da177e4SLinus Torvalds 13781da177e4SLinus Torvalds tp->rx_opt.mss_clamp is mss negotiated at connection setup. 1379caa20d9aSStephen Hemminger It is minimum of user_mss and mss received with SYN. 13801da177e4SLinus Torvalds It also does not include TCP options. 13811da177e4SLinus Torvalds 1382d83d8461SArnaldo Carvalho de Melo inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. 13831da177e4SLinus Torvalds 13841da177e4SLinus Torvalds tp->mss_cache is current effective sending mss, including 13851da177e4SLinus Torvalds all tcp options except for SACKs. It is evaluated, 13861da177e4SLinus Torvalds taking into account current pmtu, but never exceeds 13871da177e4SLinus Torvalds tp->rx_opt.mss_clamp. 13881da177e4SLinus Torvalds 13891da177e4SLinus Torvalds NOTE1. rfc1122 clearly states that advertised MSS 13901da177e4SLinus Torvalds DOES NOT include either tcp or ip options. 13911da177e4SLinus Torvalds 1392d83d8461SArnaldo Carvalho de Melo NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache 1393d83d8461SArnaldo Carvalho de Melo are READ ONLY outside this function. --ANK (980731) 13941da177e4SLinus Torvalds */ 13951da177e4SLinus Torvalds unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 13961da177e4SLinus Torvalds { 13971da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1398d83d8461SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 13995d424d5aSJohn Heffner int mss_now; 14001da177e4SLinus Torvalds 14015d424d5aSJohn Heffner if (icsk->icsk_mtup.search_high > pmtu) 14025d424d5aSJohn Heffner icsk->icsk_mtup.search_high = pmtu; 14031da177e4SLinus Torvalds 14045d424d5aSJohn Heffner mss_now = tcp_mtu_to_mss(sk, pmtu); 1405409d22b4SIlpo Järvinen mss_now = tcp_bound_to_half_wnd(tp, mss_now); 14061da177e4SLinus Torvalds 14071da177e4SLinus Torvalds /* And store cached results */ 1408d83d8461SArnaldo Carvalho de Melo icsk->icsk_pmtu_cookie = pmtu; 14095d424d5aSJohn Heffner if (icsk->icsk_mtup.enabled) 14105d424d5aSJohn Heffner mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); 1411c1b4a7e6SDavid S. Miller tp->mss_cache = mss_now; 14121da177e4SLinus Torvalds 14131da177e4SLinus Torvalds return mss_now; 14141da177e4SLinus Torvalds } 14154bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_sync_mss); 14161da177e4SLinus Torvalds 14171da177e4SLinus Torvalds /* Compute the current effective MSS, taking SACKs and IP options, 14181da177e4SLinus Torvalds * and even PMTU discovery events into account. 14191da177e4SLinus Torvalds */ 14200c54b85fSIlpo Järvinen unsigned int tcp_current_mss(struct sock *sk) 14211da177e4SLinus Torvalds { 1422cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1423cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1424c1b4a7e6SDavid S. Miller u32 mss_now; 142595c96174SEric Dumazet unsigned int header_len; 142633ad798cSAdam Langley struct tcp_out_options opts; 142733ad798cSAdam Langley struct tcp_md5sig_key *md5; 14281da177e4SLinus Torvalds 1429c1b4a7e6SDavid S. Miller mss_now = tp->mss_cache; 1430c1b4a7e6SDavid S. Miller 14311da177e4SLinus Torvalds if (dst) { 14321da177e4SLinus Torvalds u32 mtu = dst_mtu(dst); 1433d83d8461SArnaldo Carvalho de Melo if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 14341da177e4SLinus Torvalds mss_now = tcp_sync_mss(sk, mtu); 14351da177e4SLinus Torvalds } 14361da177e4SLinus Torvalds 143733ad798cSAdam Langley header_len = tcp_established_options(sk, NULL, &opts, &md5) + 143833ad798cSAdam Langley sizeof(struct tcphdr); 143933ad798cSAdam Langley /* The mss_cache is sized based on tp->tcp_header_len, which assumes 144033ad798cSAdam Langley * some common options. If this is an odd packet (because we have SACK 144133ad798cSAdam Langley * blocks etc) then our calculated header_len will be different, and 144233ad798cSAdam Langley * we have to adjust mss_now correspondingly */ 144333ad798cSAdam Langley if (header_len != tp->tcp_header_len) { 144433ad798cSAdam Langley int delta = (int) header_len - tp->tcp_header_len; 144533ad798cSAdam Langley mss_now -= delta; 144633ad798cSAdam Langley } 1447cfb6eeb4SYOSHIFUJI Hideaki 14481da177e4SLinus Torvalds return mss_now; 14491da177e4SLinus Torvalds } 14501da177e4SLinus Torvalds 145186fd14adSWeiping Pan /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. 145286fd14adSWeiping Pan * As additional protections, we do not touch cwnd in retransmission phases, 145386fd14adSWeiping Pan * and if application hit its sndbuf limit recently. 145486fd14adSWeiping Pan */ 145586fd14adSWeiping Pan static void tcp_cwnd_application_limited(struct sock *sk) 1456a762a980SDavid S. Miller { 14579e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1458a762a980SDavid S. Miller 145986fd14adSWeiping Pan if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && 146086fd14adSWeiping Pan sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 146186fd14adSWeiping Pan /* Limited by application or receiver window. */ 146286fd14adSWeiping Pan u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); 146386fd14adSWeiping Pan u32 win_used = max(tp->snd_cwnd_used, init_win); 146486fd14adSWeiping Pan if (win_used < tp->snd_cwnd) { 146586fd14adSWeiping Pan tp->snd_ssthresh = tcp_current_ssthresh(sk); 146686fd14adSWeiping Pan tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1; 146786fd14adSWeiping Pan } 146886fd14adSWeiping Pan tp->snd_cwnd_used = 0; 146986fd14adSWeiping Pan } 147086fd14adSWeiping Pan tp->snd_cwnd_stamp = tcp_time_stamp; 147186fd14adSWeiping Pan } 147286fd14adSWeiping Pan 1473ca8a2263SNeal Cardwell static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited) 1474a762a980SDavid S. Miller { 1475a762a980SDavid S. Miller struct tcp_sock *tp = tcp_sk(sk); 1476a762a980SDavid S. Miller 1477ca8a2263SNeal Cardwell /* Track the maximum number of outstanding packets in each 1478ca8a2263SNeal Cardwell * window, and remember whether we were cwnd-limited then. 1479ca8a2263SNeal Cardwell */ 1480ca8a2263SNeal Cardwell if (!before(tp->snd_una, tp->max_packets_seq) || 1481ca8a2263SNeal Cardwell tp->packets_out > tp->max_packets_out) { 1482ca8a2263SNeal Cardwell tp->max_packets_out = tp->packets_out; 1483ca8a2263SNeal Cardwell tp->max_packets_seq = tp->snd_nxt; 1484ca8a2263SNeal Cardwell tp->is_cwnd_limited = is_cwnd_limited; 1485ca8a2263SNeal Cardwell } 1486e114a710SEric Dumazet 148724901551SEric Dumazet if (tcp_is_cwnd_limited(sk)) { 1488a762a980SDavid S. Miller /* Network is feed fully. */ 1489a762a980SDavid S. Miller tp->snd_cwnd_used = 0; 1490a762a980SDavid S. Miller tp->snd_cwnd_stamp = tcp_time_stamp; 1491a762a980SDavid S. Miller } else { 1492a762a980SDavid S. Miller /* Network starves. */ 1493a762a980SDavid S. Miller if (tp->packets_out > tp->snd_cwnd_used) 1494a762a980SDavid S. Miller tp->snd_cwnd_used = tp->packets_out; 1495a762a980SDavid S. Miller 149615d33c07SDavid S. Miller if (sysctl_tcp_slow_start_after_idle && 149715d33c07SDavid S. Miller (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto) 1498a762a980SDavid S. Miller tcp_cwnd_application_limited(sk); 1499a762a980SDavid S. Miller } 1500a762a980SDavid S. Miller } 1501a762a980SDavid S. Miller 1502d4589926SEric Dumazet /* Minshall's variant of the Nagle send check. */ 1503d4589926SEric Dumazet static bool tcp_minshall_check(const struct tcp_sock *tp) 1504d4589926SEric Dumazet { 1505d4589926SEric Dumazet return after(tp->snd_sml, tp->snd_una) && 1506d4589926SEric Dumazet !after(tp->snd_sml, tp->snd_nxt); 1507d4589926SEric Dumazet } 1508d4589926SEric Dumazet 1509d4589926SEric Dumazet /* Update snd_sml if this skb is under mss 1510d4589926SEric Dumazet * Note that a TSO packet might end with a sub-mss segment 1511d4589926SEric Dumazet * The test is really : 1512d4589926SEric Dumazet * if ((skb->len % mss) != 0) 1513d4589926SEric Dumazet * tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1514d4589926SEric Dumazet * But we can avoid doing the divide again given we already have 1515d4589926SEric Dumazet * skb_pcount = skb->len / mss_now 15160e3a4803SIlpo Järvinen */ 1517d4589926SEric Dumazet static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now, 1518d4589926SEric Dumazet const struct sk_buff *skb) 1519d4589926SEric Dumazet { 1520d4589926SEric Dumazet if (skb->len < tcp_skb_pcount(skb) * mss_now) 1521d4589926SEric Dumazet tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1522d4589926SEric Dumazet } 1523d4589926SEric Dumazet 1524d4589926SEric Dumazet /* Return false, if packet can be sent now without violation Nagle's rules: 1525d4589926SEric Dumazet * 1. It is full sized. (provided by caller in %partial bool) 1526d4589926SEric Dumazet * 2. Or it contains FIN. (already checked by caller) 1527d4589926SEric Dumazet * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. 1528d4589926SEric Dumazet * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 1529d4589926SEric Dumazet * With Minshall's modification: all sent small packets are ACKed. 1530d4589926SEric Dumazet */ 1531d4589926SEric Dumazet static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp, 1532cc93fc51SPeter Pan(潘卫平) int nonagle) 1533d4589926SEric Dumazet { 1534d4589926SEric Dumazet return partial && 1535d4589926SEric Dumazet ((nonagle & TCP_NAGLE_CORK) || 1536d4589926SEric Dumazet (!nonagle && tp->packets_out && tcp_minshall_check(tp))); 1537d4589926SEric Dumazet } 1538605ad7f1SEric Dumazet 1539605ad7f1SEric Dumazet /* Return how many segs we'd like on a TSO packet, 1540605ad7f1SEric Dumazet * to send one TSO packet per ms 1541605ad7f1SEric Dumazet */ 1542605ad7f1SEric Dumazet static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now) 1543605ad7f1SEric Dumazet { 1544605ad7f1SEric Dumazet u32 bytes, segs; 1545605ad7f1SEric Dumazet 1546605ad7f1SEric Dumazet bytes = min(sk->sk_pacing_rate >> 10, 1547605ad7f1SEric Dumazet sk->sk_gso_max_size - 1 - MAX_TCP_HEADER); 1548605ad7f1SEric Dumazet 1549605ad7f1SEric Dumazet /* Goal is to send at least one packet per ms, 1550605ad7f1SEric Dumazet * not one big TSO packet every 100 ms. 1551605ad7f1SEric Dumazet * This preserves ACK clocking and is consistent 1552605ad7f1SEric Dumazet * with tcp_tso_should_defer() heuristic. 1553605ad7f1SEric Dumazet */ 1554605ad7f1SEric Dumazet segs = max_t(u32, bytes / mss_now, sysctl_tcp_min_tso_segs); 1555605ad7f1SEric Dumazet 1556605ad7f1SEric Dumazet return min_t(u32, segs, sk->sk_gso_max_segs); 1557605ad7f1SEric Dumazet } 1558605ad7f1SEric Dumazet 1559d4589926SEric Dumazet /* Returns the portion of skb which can be sent right away */ 1560d4589926SEric Dumazet static unsigned int tcp_mss_split_point(const struct sock *sk, 1561d4589926SEric Dumazet const struct sk_buff *skb, 1562d4589926SEric Dumazet unsigned int mss_now, 1563d4589926SEric Dumazet unsigned int max_segs, 1564d4589926SEric Dumazet int nonagle) 1565c1b4a7e6SDavid S. Miller { 1566cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1567d4589926SEric Dumazet u32 partial, needed, window, max_len; 1568c1b4a7e6SDavid S. Miller 156990840defSIlpo Järvinen window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 15701485348dSBen Hutchings max_len = mss_now * max_segs; 15710e3a4803SIlpo Järvinen 15721485348dSBen Hutchings if (likely(max_len <= window && skb != tcp_write_queue_tail(sk))) 15731485348dSBen Hutchings return max_len; 15740e3a4803SIlpo Järvinen 15755ea3a748SIlpo Järvinen needed = min(skb->len, window); 15765ea3a748SIlpo Järvinen 15771485348dSBen Hutchings if (max_len <= needed) 15781485348dSBen Hutchings return max_len; 15790e3a4803SIlpo Järvinen 1580d4589926SEric Dumazet partial = needed % mss_now; 1581d4589926SEric Dumazet /* If last segment is not a full MSS, check if Nagle rules allow us 1582d4589926SEric Dumazet * to include this last segment in this skb. 1583d4589926SEric Dumazet * Otherwise, we'll split the skb at last MSS boundary 1584d4589926SEric Dumazet */ 1585cc93fc51SPeter Pan(潘卫平) if (tcp_nagle_check(partial != 0, tp, nonagle)) 1586d4589926SEric Dumazet return needed - partial; 1587d4589926SEric Dumazet 1588d4589926SEric Dumazet return needed; 1589c1b4a7e6SDavid S. Miller } 1590c1b4a7e6SDavid S. Miller 1591c1b4a7e6SDavid S. Miller /* Can at least one segment of SKB be sent right now, according to the 1592c1b4a7e6SDavid S. Miller * congestion window rules? If so, return how many segments are allowed. 1593c1b4a7e6SDavid S. Miller */ 1594cf533ea5SEric Dumazet static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, 1595cf533ea5SEric Dumazet const struct sk_buff *skb) 1596c1b4a7e6SDavid S. Miller { 1597d649a7a8SEric Dumazet u32 in_flight, cwnd, halfcwnd; 1598c1b4a7e6SDavid S. Miller 1599c1b4a7e6SDavid S. Miller /* Don't be strict about the congestion window for the final FIN. */ 16004de075e0SEric Dumazet if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && 16014de075e0SEric Dumazet tcp_skb_pcount(skb) == 1) 1602c1b4a7e6SDavid S. Miller return 1; 1603c1b4a7e6SDavid S. Miller 1604c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1605c1b4a7e6SDavid S. Miller cwnd = tp->snd_cwnd; 1606d649a7a8SEric Dumazet if (in_flight >= cwnd) 1607c1b4a7e6SDavid S. Miller return 0; 1608d649a7a8SEric Dumazet 1609d649a7a8SEric Dumazet /* For better scheduling, ensure we have at least 1610d649a7a8SEric Dumazet * 2 GSO packets in flight. 1611d649a7a8SEric Dumazet */ 1612d649a7a8SEric Dumazet halfcwnd = max(cwnd >> 1, 1U); 1613d649a7a8SEric Dumazet return min(halfcwnd, cwnd - in_flight); 1614c1b4a7e6SDavid S. Miller } 1615c1b4a7e6SDavid S. Miller 1616b595076aSUwe Kleine-König /* Initialize TSO state of a skb. 161767edfef7SAndi Kleen * This must be invoked the first time we consider transmitting 1618c1b4a7e6SDavid S. Miller * SKB onto the wire. 1619c1b4a7e6SDavid S. Miller */ 16205bbb432cSEric Dumazet static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now) 1621c1b4a7e6SDavid S. Miller { 1622c1b4a7e6SDavid S. Miller int tso_segs = tcp_skb_pcount(skb); 1623c1b4a7e6SDavid S. Miller 1624f8269a49SIlpo Järvinen if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { 16255bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 1626c1b4a7e6SDavid S. Miller tso_segs = tcp_skb_pcount(skb); 1627c1b4a7e6SDavid S. Miller } 1628c1b4a7e6SDavid S. Miller return tso_segs; 1629c1b4a7e6SDavid S. Miller } 1630c1b4a7e6SDavid S. Miller 1631c1b4a7e6SDavid S. Miller 1632a2a385d6SEric Dumazet /* Return true if the Nagle test allows this packet to be 1633c1b4a7e6SDavid S. Miller * sent now. 1634c1b4a7e6SDavid S. Miller */ 1635a2a385d6SEric Dumazet static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, 1636c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1637c1b4a7e6SDavid S. Miller { 1638c1b4a7e6SDavid S. Miller /* Nagle rule does not apply to frames, which sit in the middle of the 1639c1b4a7e6SDavid S. Miller * write_queue (they have no chances to get new data). 1640c1b4a7e6SDavid S. Miller * 1641c1b4a7e6SDavid S. Miller * This is implemented in the callers, where they modify the 'nonagle' 1642c1b4a7e6SDavid S. Miller * argument based upon the location of SKB in the send queue. 1643c1b4a7e6SDavid S. Miller */ 1644c1b4a7e6SDavid S. Miller if (nonagle & TCP_NAGLE_PUSH) 1645a2a385d6SEric Dumazet return true; 1646c1b4a7e6SDavid S. Miller 16479b44190dSYuchung Cheng /* Don't use the nagle rule for urgent data (or for the final FIN). */ 16489b44190dSYuchung Cheng if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) 1649a2a385d6SEric Dumazet return true; 1650c1b4a7e6SDavid S. Miller 1651cc93fc51SPeter Pan(潘卫平) if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle)) 1652a2a385d6SEric Dumazet return true; 1653c1b4a7e6SDavid S. Miller 1654a2a385d6SEric Dumazet return false; 1655c1b4a7e6SDavid S. Miller } 1656c1b4a7e6SDavid S. Miller 1657c1b4a7e6SDavid S. Miller /* Does at least the first segment of SKB fit into the send window? */ 1658a2a385d6SEric Dumazet static bool tcp_snd_wnd_test(const struct tcp_sock *tp, 1659a2a385d6SEric Dumazet const struct sk_buff *skb, 1660056834d9SIlpo Järvinen unsigned int cur_mss) 1661c1b4a7e6SDavid S. Miller { 1662c1b4a7e6SDavid S. Miller u32 end_seq = TCP_SKB_CB(skb)->end_seq; 1663c1b4a7e6SDavid S. Miller 1664c1b4a7e6SDavid S. Miller if (skb->len > cur_mss) 1665c1b4a7e6SDavid S. Miller end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 1666c1b4a7e6SDavid S. Miller 166790840defSIlpo Järvinen return !after(end_seq, tcp_wnd_end(tp)); 1668c1b4a7e6SDavid S. Miller } 1669c1b4a7e6SDavid S. Miller 1670fe067e8aSDavid S. Miller /* This checks if the data bearing packet SKB (usually tcp_send_head(sk)) 1671c1b4a7e6SDavid S. Miller * should be put on the wire right now. If so, it returns the number of 1672c1b4a7e6SDavid S. Miller * packets allowed by the congestion window. 1673c1b4a7e6SDavid S. Miller */ 1674cf533ea5SEric Dumazet static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb, 1675c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1676c1b4a7e6SDavid S. Miller { 1677cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1678c1b4a7e6SDavid S. Miller unsigned int cwnd_quota; 1679c1b4a7e6SDavid S. Miller 16805bbb432cSEric Dumazet tcp_init_tso_segs(skb, cur_mss); 1681c1b4a7e6SDavid S. Miller 1682c1b4a7e6SDavid S. Miller if (!tcp_nagle_test(tp, skb, cur_mss, nonagle)) 1683c1b4a7e6SDavid S. Miller return 0; 1684c1b4a7e6SDavid S. Miller 1685c1b4a7e6SDavid S. Miller cwnd_quota = tcp_cwnd_test(tp, skb); 1686056834d9SIlpo Järvinen if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss)) 1687c1b4a7e6SDavid S. Miller cwnd_quota = 0; 1688c1b4a7e6SDavid S. Miller 1689c1b4a7e6SDavid S. Miller return cwnd_quota; 1690c1b4a7e6SDavid S. Miller } 1691c1b4a7e6SDavid S. Miller 169267edfef7SAndi Kleen /* Test if sending is allowed right now. */ 1693a2a385d6SEric Dumazet bool tcp_may_send_now(struct sock *sk) 1694c1b4a7e6SDavid S. Miller { 1695cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1696fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 1697c1b4a7e6SDavid S. Miller 1698a02cec21SEric Dumazet return skb && 16990c54b85fSIlpo Järvinen tcp_snd_test(sk, skb, tcp_current_mss(sk), 1700c1b4a7e6SDavid S. Miller (tcp_skb_is_last(sk, skb) ? 1701a02cec21SEric Dumazet tp->nonagle : TCP_NAGLE_PUSH)); 1702c1b4a7e6SDavid S. Miller } 1703c1b4a7e6SDavid S. Miller 1704c1b4a7e6SDavid S. Miller /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 1705c1b4a7e6SDavid S. Miller * which is put after SKB on the list. It is very much like 1706c1b4a7e6SDavid S. Miller * tcp_fragment() except that it may make several kinds of assumptions 1707c1b4a7e6SDavid S. Miller * in order to speed up the splitting operation. In particular, we 1708c1b4a7e6SDavid S. Miller * know that all the data is in scatter-gather pages, and that the 1709c1b4a7e6SDavid S. Miller * packet has never been sent out before (and thus is not cloned). 1710c1b4a7e6SDavid S. Miller */ 1711056834d9SIlpo Järvinen static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, 1712c4ead4c5SEric Dumazet unsigned int mss_now, gfp_t gfp) 1713c1b4a7e6SDavid S. Miller { 1714c1b4a7e6SDavid S. Miller struct sk_buff *buff; 1715c1b4a7e6SDavid S. Miller int nlen = skb->len - len; 17169ce01461SIlpo Järvinen u8 flags; 1717c1b4a7e6SDavid S. Miller 1718c1b4a7e6SDavid S. Miller /* All of a TSO frame must be composed of paged data. */ 1719c8ac3774SHerbert Xu if (skb->len != skb->data_len) 17206cc55e09SOctavian Purdila return tcp_fragment(sk, skb, len, mss_now, gfp); 1721c1b4a7e6SDavid S. Miller 1722eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, gfp, true); 172351456b29SIan Morris if (unlikely(!buff)) 1724c1b4a7e6SDavid S. Miller return -ENOMEM; 1725c1b4a7e6SDavid S. Miller 17263ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 17273ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1728b60b49eaSHerbert Xu buff->truesize += nlen; 1729c1b4a7e6SDavid S. Miller skb->truesize -= nlen; 1730c1b4a7e6SDavid S. Miller 1731c1b4a7e6SDavid S. Miller /* Correct the sequence numbers. */ 1732c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 1733c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 1734c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 1735c1b4a7e6SDavid S. Miller 1736c1b4a7e6SDavid S. Miller /* PSH and FIN should only be set in the second packet. */ 17374de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 17384de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 17394de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1740c1b4a7e6SDavid S. Miller 1741c1b4a7e6SDavid S. Miller /* This packet was never sent out yet, so no SACK bits. */ 1742c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->sacked = 0; 1743c1b4a7e6SDavid S. Miller 174484fa7933SPatrick McHardy buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; 1745c1b4a7e6SDavid S. Miller skb_split(skb, buff, len); 1746490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 1747c1b4a7e6SDavid S. Miller 1748c1b4a7e6SDavid S. Miller /* Fix up tso_factor for both original and new SKB. */ 17495bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 17505bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 1751c1b4a7e6SDavid S. Miller 1752c1b4a7e6SDavid S. Miller /* Link BUFF into the send queue. */ 1753f4a775d1SEric Dumazet __skb_header_release(buff); 1754fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 1755c1b4a7e6SDavid S. Miller 1756c1b4a7e6SDavid S. Miller return 0; 1757c1b4a7e6SDavid S. Miller } 1758c1b4a7e6SDavid S. Miller 1759c1b4a7e6SDavid S. Miller /* Try to defer sending, if possible, in order to minimize the amount 1760c1b4a7e6SDavid S. Miller * of TSO splitting we do. View it as a kind of TSO Nagle test. 1761c1b4a7e6SDavid S. Miller * 1762c1b4a7e6SDavid S. Miller * This algorithm is from John Heffner. 1763c1b4a7e6SDavid S. Miller */ 1764ca8a2263SNeal Cardwell static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb, 1765605ad7f1SEric Dumazet bool *is_cwnd_limited, u32 max_segs) 1766c1b4a7e6SDavid S. Miller { 17676687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 176850c8339eSEric Dumazet u32 age, send_win, cong_win, limit, in_flight; 176950c8339eSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 177050c8339eSEric Dumazet struct skb_mstamp now; 177150c8339eSEric Dumazet struct sk_buff *head; 1772ad9f4f50SEric Dumazet int win_divisor; 1773c1b4a7e6SDavid S. Miller 17744de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 1775ae8064acSJohn Heffner goto send_now; 1776c1b4a7e6SDavid S. Miller 177799d7662aSEric Dumazet if (icsk->icsk_ca_state >= TCP_CA_Recovery) 1778ae8064acSJohn Heffner goto send_now; 1779ae8064acSJohn Heffner 17805f852eb5SEric Dumazet /* Avoid bursty behavior by allowing defer 17815f852eb5SEric Dumazet * only if the last write was recent. 17825f852eb5SEric Dumazet */ 17835f852eb5SEric Dumazet if ((s32)(tcp_time_stamp - tp->lsndtime) > 0) 1784ae8064acSJohn Heffner goto send_now; 1785908a75c1SDavid S. Miller 1786c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1787c1b4a7e6SDavid S. Miller 1788056834d9SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight)); 1789c1b4a7e6SDavid S. Miller 179090840defSIlpo Järvinen send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1791c1b4a7e6SDavid S. Miller 1792c1b4a7e6SDavid S. Miller /* From in_flight test above, we know that cwnd > in_flight. */ 1793c1b4a7e6SDavid S. Miller cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; 1794c1b4a7e6SDavid S. Miller 1795c1b4a7e6SDavid S. Miller limit = min(send_win, cong_win); 1796c1b4a7e6SDavid S. Miller 1797ba244fe9SDavid S. Miller /* If a full-sized TSO skb can be sent, do it. */ 1798605ad7f1SEric Dumazet if (limit >= max_segs * tp->mss_cache) 1799ae8064acSJohn Heffner goto send_now; 1800ba244fe9SDavid S. Miller 180162ad2761SIlpo Järvinen /* Middle in queue won't get any more data, full sendable already? */ 180262ad2761SIlpo Järvinen if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) 180362ad2761SIlpo Järvinen goto send_now; 180462ad2761SIlpo Järvinen 1805ad9f4f50SEric Dumazet win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor); 1806ad9f4f50SEric Dumazet if (win_divisor) { 1807c1b4a7e6SDavid S. Miller u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); 1808c1b4a7e6SDavid S. Miller 1809c1b4a7e6SDavid S. Miller /* If at least some fraction of a window is available, 1810c1b4a7e6SDavid S. Miller * just use it. 1811c1b4a7e6SDavid S. Miller */ 1812ad9f4f50SEric Dumazet chunk /= win_divisor; 1813c1b4a7e6SDavid S. Miller if (limit >= chunk) 1814ae8064acSJohn Heffner goto send_now; 1815c1b4a7e6SDavid S. Miller } else { 1816c1b4a7e6SDavid S. Miller /* Different approach, try not to defer past a single 1817c1b4a7e6SDavid S. Miller * ACK. Receiver should ACK every other full sized 1818c1b4a7e6SDavid S. Miller * frame, so if we have space for more than 3 frames 1819c1b4a7e6SDavid S. Miller * then send now. 1820c1b4a7e6SDavid S. Miller */ 18216b5a5c0dSNeal Cardwell if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) 1822ae8064acSJohn Heffner goto send_now; 1823c1b4a7e6SDavid S. Miller } 1824c1b4a7e6SDavid S. Miller 182550c8339eSEric Dumazet head = tcp_write_queue_head(sk); 182650c8339eSEric Dumazet skb_mstamp_get(&now); 182750c8339eSEric Dumazet age = skb_mstamp_us_delta(&now, &head->skb_mstamp); 182850c8339eSEric Dumazet /* If next ACK is likely to come too late (half srtt), do not defer */ 182950c8339eSEric Dumazet if (age < (tp->srtt_us >> 4)) 183050c8339eSEric Dumazet goto send_now; 183150c8339eSEric Dumazet 18325f852eb5SEric Dumazet /* Ok, it looks like it is advisable to defer. */ 1833ae8064acSJohn Heffner 1834ca8a2263SNeal Cardwell if (cong_win < send_win && cong_win < skb->len) 1835ca8a2263SNeal Cardwell *is_cwnd_limited = true; 1836ca8a2263SNeal Cardwell 1837a2a385d6SEric Dumazet return true; 1838ae8064acSJohn Heffner 1839ae8064acSJohn Heffner send_now: 1840a2a385d6SEric Dumazet return false; 1841c1b4a7e6SDavid S. Miller } 1842c1b4a7e6SDavid S. Miller 184305cbc0dbSFan Du static inline void tcp_mtu_check_reprobe(struct sock *sk) 184405cbc0dbSFan Du { 184505cbc0dbSFan Du struct inet_connection_sock *icsk = inet_csk(sk); 184605cbc0dbSFan Du struct tcp_sock *tp = tcp_sk(sk); 184705cbc0dbSFan Du struct net *net = sock_net(sk); 184805cbc0dbSFan Du u32 interval; 184905cbc0dbSFan Du s32 delta; 185005cbc0dbSFan Du 185105cbc0dbSFan Du interval = net->ipv4.sysctl_tcp_probe_interval; 185205cbc0dbSFan Du delta = tcp_time_stamp - icsk->icsk_mtup.probe_timestamp; 185305cbc0dbSFan Du if (unlikely(delta >= interval * HZ)) { 185405cbc0dbSFan Du int mss = tcp_current_mss(sk); 185505cbc0dbSFan Du 185605cbc0dbSFan Du /* Update current search range */ 185705cbc0dbSFan Du icsk->icsk_mtup.probe_size = 0; 185805cbc0dbSFan Du icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + 185905cbc0dbSFan Du sizeof(struct tcphdr) + 186005cbc0dbSFan Du icsk->icsk_af_ops->net_header_len; 186105cbc0dbSFan Du icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss); 186205cbc0dbSFan Du 186305cbc0dbSFan Du /* Update probe time stamp */ 186405cbc0dbSFan Du icsk->icsk_mtup.probe_timestamp = tcp_time_stamp; 186505cbc0dbSFan Du } 186605cbc0dbSFan Du } 186705cbc0dbSFan Du 18685d424d5aSJohn Heffner /* Create a new MTU probe if we are ready. 186967edfef7SAndi Kleen * MTU probe is regularly attempting to increase the path MTU by 187067edfef7SAndi Kleen * deliberately sending larger packets. This discovers routing 187167edfef7SAndi Kleen * changes resulting in larger path MTUs. 187267edfef7SAndi Kleen * 18735d424d5aSJohn Heffner * Returns 0 if we should wait to probe (no cwnd available), 18745d424d5aSJohn Heffner * 1 if a probe was sent, 1875056834d9SIlpo Järvinen * -1 otherwise 1876056834d9SIlpo Järvinen */ 18775d424d5aSJohn Heffner static int tcp_mtu_probe(struct sock *sk) 18785d424d5aSJohn Heffner { 18795d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 18805d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 18815d424d5aSJohn Heffner struct sk_buff *skb, *nskb, *next; 18826b58e0a5SFan Du struct net *net = sock_net(sk); 18835d424d5aSJohn Heffner int len; 18845d424d5aSJohn Heffner int probe_size; 188591cc17c0SIlpo Järvinen int size_needed; 18865d424d5aSJohn Heffner int copy; 18875d424d5aSJohn Heffner int mss_now; 18886b58e0a5SFan Du int interval; 18895d424d5aSJohn Heffner 18905d424d5aSJohn Heffner /* Not currently probing/verifying, 18915d424d5aSJohn Heffner * not in recovery, 18925d424d5aSJohn Heffner * have enough cwnd, and 18935d424d5aSJohn Heffner * not SACKing (the variable headers throw things off) */ 18945d424d5aSJohn Heffner if (!icsk->icsk_mtup.enabled || 18955d424d5aSJohn Heffner icsk->icsk_mtup.probe_size || 18965d424d5aSJohn Heffner inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 18975d424d5aSJohn Heffner tp->snd_cwnd < 11 || 1898cabeccbdSIlpo Järvinen tp->rx_opt.num_sacks || tp->rx_opt.dsack) 18995d424d5aSJohn Heffner return -1; 19005d424d5aSJohn Heffner 19016b58e0a5SFan Du /* Use binary search for probe_size between tcp_mss_base, 19026b58e0a5SFan Du * and current mss_clamp. if (search_high - search_low) 19036b58e0a5SFan Du * smaller than a threshold, backoff from probing. 19046b58e0a5SFan Du */ 19050c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 19066b58e0a5SFan Du probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high + 19076b58e0a5SFan Du icsk->icsk_mtup.search_low) >> 1); 190891cc17c0SIlpo Järvinen size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; 19096b58e0a5SFan Du interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low; 191005cbc0dbSFan Du /* When misfortune happens, we are reprobing actively, 191105cbc0dbSFan Du * and then reprobe timer has expired. We stick with current 191205cbc0dbSFan Du * probing process by not resetting search range to its orignal. 191305cbc0dbSFan Du */ 19146b58e0a5SFan Du if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) || 191505cbc0dbSFan Du interval < net->ipv4.sysctl_tcp_probe_threshold) { 191605cbc0dbSFan Du /* Check whether enough time has elaplased for 191705cbc0dbSFan Du * another round of probing. 191805cbc0dbSFan Du */ 191905cbc0dbSFan Du tcp_mtu_check_reprobe(sk); 19205d424d5aSJohn Heffner return -1; 19215d424d5aSJohn Heffner } 19225d424d5aSJohn Heffner 19235d424d5aSJohn Heffner /* Have enough data in the send queue to probe? */ 19247f9c33e5SIlpo Järvinen if (tp->write_seq - tp->snd_nxt < size_needed) 19255d424d5aSJohn Heffner return -1; 19265d424d5aSJohn Heffner 192791cc17c0SIlpo Järvinen if (tp->snd_wnd < size_needed) 19285d424d5aSJohn Heffner return -1; 192990840defSIlpo Järvinen if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) 19305d424d5aSJohn Heffner return 0; 19315d424d5aSJohn Heffner 1932d67c58e9SIlpo Järvinen /* Do we need to wait to drain cwnd? With none in flight, don't stall */ 1933d67c58e9SIlpo Järvinen if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { 1934d67c58e9SIlpo Järvinen if (!tcp_packets_in_flight(tp)) 19355d424d5aSJohn Heffner return -1; 19365d424d5aSJohn Heffner else 19375d424d5aSJohn Heffner return 0; 19385d424d5aSJohn Heffner } 19395d424d5aSJohn Heffner 19405d424d5aSJohn Heffner /* We're allowed to probe. Build it now. */ 1941eb934478SEric Dumazet nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false); 194251456b29SIan Morris if (!nskb) 19435d424d5aSJohn Heffner return -1; 19443ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 19453ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 19465d424d5aSJohn Heffner 1947fe067e8aSDavid S. Miller skb = tcp_send_head(sk); 19485d424d5aSJohn Heffner 19495d424d5aSJohn Heffner TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 19505d424d5aSJohn Heffner TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 19514de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; 19525d424d5aSJohn Heffner TCP_SKB_CB(nskb)->sacked = 0; 19535d424d5aSJohn Heffner nskb->csum = 0; 195484fa7933SPatrick McHardy nskb->ip_summed = skb->ip_summed; 19555d424d5aSJohn Heffner 195650c4817eSIlpo Järvinen tcp_insert_write_queue_before(nskb, skb, sk); 195750c4817eSIlpo Järvinen 19585d424d5aSJohn Heffner len = 0; 1959234b6860SIlpo Järvinen tcp_for_write_queue_from_safe(skb, next, sk) { 19605d424d5aSJohn Heffner copy = min_t(int, skb->len, probe_size - len); 19615d424d5aSJohn Heffner if (nskb->ip_summed) 19625d424d5aSJohn Heffner skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); 19635d424d5aSJohn Heffner else 19645d424d5aSJohn Heffner nskb->csum = skb_copy_and_csum_bits(skb, 0, 1965056834d9SIlpo Järvinen skb_put(nskb, copy), 1966056834d9SIlpo Järvinen copy, nskb->csum); 19675d424d5aSJohn Heffner 19685d424d5aSJohn Heffner if (skb->len <= copy) { 19695d424d5aSJohn Heffner /* We've eaten all the data from this skb. 19705d424d5aSJohn Heffner * Throw it away. */ 19714de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 1972fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 19733ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 19745d424d5aSJohn Heffner } else { 19754de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & 1976a3433f35SChangli Gao ~(TCPHDR_FIN|TCPHDR_PSH); 19775d424d5aSJohn Heffner if (!skb_shinfo(skb)->nr_frags) { 19785d424d5aSJohn Heffner skb_pull(skb, copy); 197984fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 1980056834d9SIlpo Järvinen skb->csum = csum_partial(skb->data, 1981056834d9SIlpo Järvinen skb->len, 0); 19825d424d5aSJohn Heffner } else { 19835d424d5aSJohn Heffner __pskb_trim_head(skb, copy); 19845bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 19855d424d5aSJohn Heffner } 19865d424d5aSJohn Heffner TCP_SKB_CB(skb)->seq += copy; 19875d424d5aSJohn Heffner } 19885d424d5aSJohn Heffner 19895d424d5aSJohn Heffner len += copy; 1990234b6860SIlpo Järvinen 1991234b6860SIlpo Järvinen if (len >= probe_size) 1992234b6860SIlpo Järvinen break; 19935d424d5aSJohn Heffner } 19945bbb432cSEric Dumazet tcp_init_tso_segs(nskb, nskb->len); 19955d424d5aSJohn Heffner 19965d424d5aSJohn Heffner /* We're ready to send. If this fails, the probe will 19977faee5c0SEric Dumazet * be resegmented into mss-sized pieces by tcp_write_xmit(). 19987faee5c0SEric Dumazet */ 19995d424d5aSJohn Heffner if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 20005d424d5aSJohn Heffner /* Decrement cwnd here because we are sending 20015d424d5aSJohn Heffner * effectively two packets. */ 20025d424d5aSJohn Heffner tp->snd_cwnd--; 200366f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, nskb); 20045d424d5aSJohn Heffner 20055d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 20060e7b1368SJohn Heffner tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 20070e7b1368SJohn Heffner tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 20085d424d5aSJohn Heffner 20095d424d5aSJohn Heffner return 1; 20105d424d5aSJohn Heffner } 20115d424d5aSJohn Heffner 20125d424d5aSJohn Heffner return -1; 20135d424d5aSJohn Heffner } 20145d424d5aSJohn Heffner 20151da177e4SLinus Torvalds /* This routine writes packets to the network. It advances the 20161da177e4SLinus Torvalds * send_head. This happens as incoming acks open up the remote 20171da177e4SLinus Torvalds * window for us. 20181da177e4SLinus Torvalds * 2019f8269a49SIlpo Järvinen * LARGESEND note: !tcp_urg_mode is overkill, only frames between 2020f8269a49SIlpo Järvinen * snd_up-64k-mss .. snd_up cannot be large. However, taking into 2021f8269a49SIlpo Järvinen * account rare use of URG, this is not a big flaw. 2022f8269a49SIlpo Järvinen * 20236ba8a3b1SNandita Dukkipati * Send at most one packet when push_one > 0. Temporarily ignore 20246ba8a3b1SNandita Dukkipati * cwnd limit to force at most one packet out when push_one == 2. 20256ba8a3b1SNandita Dukkipati 2026a2a385d6SEric Dumazet * Returns true, if no segments are in flight and we have queued segments, 2027a2a385d6SEric Dumazet * but cannot send anything now because of SWS or another problem. 20281da177e4SLinus Torvalds */ 2029a2a385d6SEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 2030d5dd9175SIlpo Järvinen int push_one, gfp_t gfp) 20311da177e4SLinus Torvalds { 20321da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 203392df7b51SDavid S. Miller struct sk_buff *skb; 2034c1b4a7e6SDavid S. Miller unsigned int tso_segs, sent_pkts; 2035c1b4a7e6SDavid S. Miller int cwnd_quota; 20365d424d5aSJohn Heffner int result; 2037ca8a2263SNeal Cardwell bool is_cwnd_limited = false; 2038605ad7f1SEric Dumazet u32 max_segs; 20391da177e4SLinus Torvalds 2040c1b4a7e6SDavid S. Miller sent_pkts = 0; 20415d424d5aSJohn Heffner 2042d5dd9175SIlpo Järvinen if (!push_one) { 20435d424d5aSJohn Heffner /* Do MTU probing. */ 2044d5dd9175SIlpo Järvinen result = tcp_mtu_probe(sk); 2045d5dd9175SIlpo Järvinen if (!result) { 2046a2a385d6SEric Dumazet return false; 20475d424d5aSJohn Heffner } else if (result > 0) { 20485d424d5aSJohn Heffner sent_pkts = 1; 20495d424d5aSJohn Heffner } 2050d5dd9175SIlpo Järvinen } 20515d424d5aSJohn Heffner 2052605ad7f1SEric Dumazet max_segs = tcp_tso_autosize(sk, mss_now); 2053fe067e8aSDavid S. Miller while ((skb = tcp_send_head(sk))) { 2054c8ac3774SHerbert Xu unsigned int limit; 2055c8ac3774SHerbert Xu 20565bbb432cSEric Dumazet tso_segs = tcp_init_tso_segs(skb, mss_now); 2057c1b4a7e6SDavid S. Miller BUG_ON(!tso_segs); 2058c1b4a7e6SDavid S. Miller 20599d186cacSAndrey Vagin if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) { 20607faee5c0SEric Dumazet /* "skb_mstamp" is used as a start point for the retransmit timer */ 20617faee5c0SEric Dumazet skb_mstamp_get(&skb->skb_mstamp); 2062ec342325SAndrew Vagin goto repair; /* Skip network transmission */ 20639d186cacSAndrey Vagin } 2064ec342325SAndrew Vagin 2065b68e9f85SHerbert Xu cwnd_quota = tcp_cwnd_test(tp, skb); 20666ba8a3b1SNandita Dukkipati if (!cwnd_quota) { 2067ca8a2263SNeal Cardwell is_cwnd_limited = true; 20686ba8a3b1SNandita Dukkipati if (push_one == 2) 20696ba8a3b1SNandita Dukkipati /* Force out a loss probe pkt. */ 20706ba8a3b1SNandita Dukkipati cwnd_quota = 1; 20716ba8a3b1SNandita Dukkipati else 2072b68e9f85SHerbert Xu break; 20736ba8a3b1SNandita Dukkipati } 2074b68e9f85SHerbert Xu 2075b68e9f85SHerbert Xu if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) 2076b68e9f85SHerbert Xu break; 2077b68e9f85SHerbert Xu 2078d6a4e26aSEric Dumazet if (tso_segs == 1) { 2079aa93466bSDavid S. Miller if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 2080aa93466bSDavid S. Miller (tcp_skb_is_last(sk, skb) ? 2081aa93466bSDavid S. Miller nonagle : TCP_NAGLE_PUSH)))) 2082aa93466bSDavid S. Miller break; 2083c1b4a7e6SDavid S. Miller } else { 2084ca8a2263SNeal Cardwell if (!push_one && 2085605ad7f1SEric Dumazet tcp_tso_should_defer(sk, skb, &is_cwnd_limited, 2086605ad7f1SEric Dumazet max_segs)) 2087aa93466bSDavid S. Miller break; 2088c1b4a7e6SDavid S. Miller } 2089aa93466bSDavid S. Miller 2090605ad7f1SEric Dumazet limit = mss_now; 2091d6a4e26aSEric Dumazet if (tso_segs > 1 && !tcp_urg_mode(tp)) 2092605ad7f1SEric Dumazet limit = tcp_mss_split_point(sk, skb, mss_now, 2093605ad7f1SEric Dumazet min_t(unsigned int, 2094605ad7f1SEric Dumazet cwnd_quota, 2095605ad7f1SEric Dumazet max_segs), 2096605ad7f1SEric Dumazet nonagle); 2097605ad7f1SEric Dumazet 2098605ad7f1SEric Dumazet if (skb->len > limit && 2099605ad7f1SEric Dumazet unlikely(tso_fragment(sk, skb, limit, mss_now, gfp))) 2100605ad7f1SEric Dumazet break; 2101605ad7f1SEric Dumazet 2102c9eeec26SEric Dumazet /* TCP Small Queues : 2103c9eeec26SEric Dumazet * Control number of packets in qdisc/devices to two packets / or ~1 ms. 2104c9eeec26SEric Dumazet * This allows for : 2105c9eeec26SEric Dumazet * - better RTT estimation and ACK scheduling 2106c9eeec26SEric Dumazet * - faster recovery 2107c9eeec26SEric Dumazet * - high rates 210898e09386SEric Dumazet * Alas, some drivers / subsystems require a fair amount 210998e09386SEric Dumazet * of queued bytes to ensure line rate. 211098e09386SEric Dumazet * One example is wifi aggregation (802.11 AMPDU) 211146d3ceabSEric Dumazet */ 2112605ad7f1SEric Dumazet limit = max(2 * skb->truesize, sk->sk_pacing_rate >> 10); 2113605ad7f1SEric Dumazet limit = min_t(u32, limit, sysctl_tcp_limit_output_bytes); 2114c9eeec26SEric Dumazet 2115c9eeec26SEric Dumazet if (atomic_read(&sk->sk_wmem_alloc) > limit) { 211646d3ceabSEric Dumazet set_bit(TSQ_THROTTLED, &tp->tsq_flags); 2117bf06200eSJohn Ogness /* It is possible TX completion already happened 2118bf06200eSJohn Ogness * before we set TSQ_THROTTLED, so we must 2119bf06200eSJohn Ogness * test again the condition. 2120bf06200eSJohn Ogness */ 21214e857c58SPeter Zijlstra smp_mb__after_atomic(); 2122bf06200eSJohn Ogness if (atomic_read(&sk->sk_wmem_alloc) > limit) 212346d3ceabSEric Dumazet break; 212446d3ceabSEric Dumazet } 2125c9eeec26SEric Dumazet 2126d5dd9175SIlpo Järvinen if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) 21271da177e4SLinus Torvalds break; 21281da177e4SLinus Torvalds 2129ec342325SAndrew Vagin repair: 21301da177e4SLinus Torvalds /* Advance the send_head. This one is sent out. 21311da177e4SLinus Torvalds * This call will increment packets_out. 21321da177e4SLinus Torvalds */ 213366f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 21341da177e4SLinus Torvalds 21351da177e4SLinus Torvalds tcp_minshall_update(tp, mss_now, skb); 2136a262f0cdSNandita Dukkipati sent_pkts += tcp_skb_pcount(skb); 2137d5dd9175SIlpo Järvinen 2138d5dd9175SIlpo Järvinen if (push_one) 2139d5dd9175SIlpo Järvinen break; 21401da177e4SLinus Torvalds } 21411da177e4SLinus Torvalds 2142aa93466bSDavid S. Miller if (likely(sent_pkts)) { 2143684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 2144684bad11SYuchung Cheng tp->prr_out += sent_pkts; 21456ba8a3b1SNandita Dukkipati 21466ba8a3b1SNandita Dukkipati /* Send one loss probe per tail loss episode. */ 21476ba8a3b1SNandita Dukkipati if (push_one != 2) 21486ba8a3b1SNandita Dukkipati tcp_schedule_loss_probe(sk); 2149ca8a2263SNeal Cardwell tcp_cwnd_validate(sk, is_cwnd_limited); 2150a2a385d6SEric Dumazet return false; 21511da177e4SLinus Torvalds } 21526ba8a3b1SNandita Dukkipati return (push_one == 2) || (!tp->packets_out && tcp_send_head(sk)); 21536ba8a3b1SNandita Dukkipati } 21546ba8a3b1SNandita Dukkipati 21556ba8a3b1SNandita Dukkipati bool tcp_schedule_loss_probe(struct sock *sk) 21566ba8a3b1SNandita Dukkipati { 21576ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 21586ba8a3b1SNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 21596ba8a3b1SNandita Dukkipati u32 timeout, tlp_time_stamp, rto_time_stamp; 2160740b0f18SEric Dumazet u32 rtt = usecs_to_jiffies(tp->srtt_us >> 3); 21616ba8a3b1SNandita Dukkipati 21626ba8a3b1SNandita Dukkipati if (WARN_ON(icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS)) 21636ba8a3b1SNandita Dukkipati return false; 21646ba8a3b1SNandita Dukkipati /* No consecutive loss probes. */ 21656ba8a3b1SNandita Dukkipati if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) { 21666ba8a3b1SNandita Dukkipati tcp_rearm_rto(sk); 21676ba8a3b1SNandita Dukkipati return false; 21686ba8a3b1SNandita Dukkipati } 21696ba8a3b1SNandita Dukkipati /* Don't do any loss probe on a Fast Open connection before 3WHS 21706ba8a3b1SNandita Dukkipati * finishes. 21716ba8a3b1SNandita Dukkipati */ 21726ba8a3b1SNandita Dukkipati if (sk->sk_state == TCP_SYN_RECV) 21736ba8a3b1SNandita Dukkipati return false; 21746ba8a3b1SNandita Dukkipati 21756ba8a3b1SNandita Dukkipati /* TLP is only scheduled when next timer event is RTO. */ 21766ba8a3b1SNandita Dukkipati if (icsk->icsk_pending != ICSK_TIME_RETRANS) 21776ba8a3b1SNandita Dukkipati return false; 21786ba8a3b1SNandita Dukkipati 21796ba8a3b1SNandita Dukkipati /* Schedule a loss probe in 2*RTT for SACK capable connections 21806ba8a3b1SNandita Dukkipati * in Open state, that are either limited by cwnd or application. 21816ba8a3b1SNandita Dukkipati */ 2182740b0f18SEric Dumazet if (sysctl_tcp_early_retrans < 3 || !tp->srtt_us || !tp->packets_out || 21836ba8a3b1SNandita Dukkipati !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open) 21846ba8a3b1SNandita Dukkipati return false; 21856ba8a3b1SNandita Dukkipati 21866ba8a3b1SNandita Dukkipati if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) && 21876ba8a3b1SNandita Dukkipati tcp_send_head(sk)) 21886ba8a3b1SNandita Dukkipati return false; 21896ba8a3b1SNandita Dukkipati 21906ba8a3b1SNandita Dukkipati /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account 21916ba8a3b1SNandita Dukkipati * for delayed ack when there's one outstanding packet. 21926ba8a3b1SNandita Dukkipati */ 21936ba8a3b1SNandita Dukkipati timeout = rtt << 1; 21946ba8a3b1SNandita Dukkipati if (tp->packets_out == 1) 21956ba8a3b1SNandita Dukkipati timeout = max_t(u32, timeout, 21966ba8a3b1SNandita Dukkipati (rtt + (rtt >> 1) + TCP_DELACK_MAX)); 21976ba8a3b1SNandita Dukkipati timeout = max_t(u32, timeout, msecs_to_jiffies(10)); 21986ba8a3b1SNandita Dukkipati 21996ba8a3b1SNandita Dukkipati /* If RTO is shorter, just schedule TLP in its place. */ 22006ba8a3b1SNandita Dukkipati tlp_time_stamp = tcp_time_stamp + timeout; 22016ba8a3b1SNandita Dukkipati rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout; 22026ba8a3b1SNandita Dukkipati if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) { 22036ba8a3b1SNandita Dukkipati s32 delta = rto_time_stamp - tcp_time_stamp; 22046ba8a3b1SNandita Dukkipati if (delta > 0) 22056ba8a3b1SNandita Dukkipati timeout = delta; 22066ba8a3b1SNandita Dukkipati } 22076ba8a3b1SNandita Dukkipati 22086ba8a3b1SNandita Dukkipati inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, 22096ba8a3b1SNandita Dukkipati TCP_RTO_MAX); 22106ba8a3b1SNandita Dukkipati return true; 22116ba8a3b1SNandita Dukkipati } 22126ba8a3b1SNandita Dukkipati 22131f3279aeSEric Dumazet /* Thanks to skb fast clones, we can detect if a prior transmit of 22141f3279aeSEric Dumazet * a packet is still in a qdisc or driver queue. 22151f3279aeSEric Dumazet * In this case, there is very little point doing a retransmit ! 22161f3279aeSEric Dumazet * Note: This is called from BH context only. 22171f3279aeSEric Dumazet */ 22181f3279aeSEric Dumazet static bool skb_still_in_host_queue(const struct sock *sk, 22191f3279aeSEric Dumazet const struct sk_buff *skb) 22201f3279aeSEric Dumazet { 222139bb5e62SEric Dumazet if (unlikely(skb_fclone_busy(sk, skb))) { 22221f3279aeSEric Dumazet NET_INC_STATS_BH(sock_net(sk), 22231f3279aeSEric Dumazet LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES); 22241f3279aeSEric Dumazet return true; 22251f3279aeSEric Dumazet } 22261f3279aeSEric Dumazet return false; 22271f3279aeSEric Dumazet } 22281f3279aeSEric Dumazet 22296ba8a3b1SNandita Dukkipati /* When probe timeout (PTO) fires, send a new segment if one exists, else 22306ba8a3b1SNandita Dukkipati * retransmit the last segment. 22316ba8a3b1SNandita Dukkipati */ 22326ba8a3b1SNandita Dukkipati void tcp_send_loss_probe(struct sock *sk) 22336ba8a3b1SNandita Dukkipati { 22349b717a8dSNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 22356ba8a3b1SNandita Dukkipati struct sk_buff *skb; 22366ba8a3b1SNandita Dukkipati int pcount; 22376ba8a3b1SNandita Dukkipati int mss = tcp_current_mss(sk); 22386ba8a3b1SNandita Dukkipati int err = -1; 22396ba8a3b1SNandita Dukkipati 224000db4124SIan Morris if (tcp_send_head(sk)) { 22416ba8a3b1SNandita Dukkipati err = tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC); 22426ba8a3b1SNandita Dukkipati goto rearm_timer; 22436ba8a3b1SNandita Dukkipati } 22446ba8a3b1SNandita Dukkipati 22459b717a8dSNandita Dukkipati /* At most one outstanding TLP retransmission. */ 22469b717a8dSNandita Dukkipati if (tp->tlp_high_seq) 22479b717a8dSNandita Dukkipati goto rearm_timer; 22489b717a8dSNandita Dukkipati 22496ba8a3b1SNandita Dukkipati /* Retransmit last segment. */ 22506ba8a3b1SNandita Dukkipati skb = tcp_write_queue_tail(sk); 22516ba8a3b1SNandita Dukkipati if (WARN_ON(!skb)) 22526ba8a3b1SNandita Dukkipati goto rearm_timer; 22536ba8a3b1SNandita Dukkipati 22541f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 22551f3279aeSEric Dumazet goto rearm_timer; 22561f3279aeSEric Dumazet 22576ba8a3b1SNandita Dukkipati pcount = tcp_skb_pcount(skb); 22586ba8a3b1SNandita Dukkipati if (WARN_ON(!pcount)) 22596ba8a3b1SNandita Dukkipati goto rearm_timer; 22606ba8a3b1SNandita Dukkipati 22616ba8a3b1SNandita Dukkipati if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) { 22626cc55e09SOctavian Purdila if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss, 22636cc55e09SOctavian Purdila GFP_ATOMIC))) 22646ba8a3b1SNandita Dukkipati goto rearm_timer; 22656ba8a3b1SNandita Dukkipati skb = tcp_write_queue_tail(sk); 22666ba8a3b1SNandita Dukkipati } 22676ba8a3b1SNandita Dukkipati 22686ba8a3b1SNandita Dukkipati if (WARN_ON(!skb || !tcp_skb_pcount(skb))) 22696ba8a3b1SNandita Dukkipati goto rearm_timer; 22706ba8a3b1SNandita Dukkipati 22716ba8a3b1SNandita Dukkipati err = __tcp_retransmit_skb(sk, skb); 22726ba8a3b1SNandita Dukkipati 22739b717a8dSNandita Dukkipati /* Record snd_nxt for loss detection. */ 22749b717a8dSNandita Dukkipati if (likely(!err)) 22759b717a8dSNandita Dukkipati tp->tlp_high_seq = tp->snd_nxt; 22769b717a8dSNandita Dukkipati 22776ba8a3b1SNandita Dukkipati rearm_timer: 2278*fcd16c0aSYuchung Cheng if (likely(!err)) { 2279*fcd16c0aSYuchung Cheng NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSPROBES); 2280*fcd16c0aSYuchung Cheng /* Reset s.t. tcp_rearm_rto will restart timer from now */ 2281*fcd16c0aSYuchung Cheng inet_csk(sk)->icsk_pending = 0; 2282*fcd16c0aSYuchung Cheng } 2283*fcd16c0aSYuchung Cheng tcp_rearm_rto(sk); 22841da177e4SLinus Torvalds } 22851da177e4SLinus Torvalds 2286a762a980SDavid S. Miller /* Push out any pending frames which were held back due to 2287a762a980SDavid S. Miller * TCP_CORK or attempt at coalescing tiny packets. 2288a762a980SDavid S. Miller * The socket must be locked by the caller. 2289a762a980SDavid S. Miller */ 22909e412ba7SIlpo Järvinen void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 22919e412ba7SIlpo Järvinen int nonagle) 2292a762a980SDavid S. Miller { 2293726e07a8SIlpo Järvinen /* If we are closed, the bytes will have to remain here. 2294726e07a8SIlpo Järvinen * In time closedown will finish, we empty the write queue and 2295726e07a8SIlpo Järvinen * all will be happy. 2296726e07a8SIlpo Järvinen */ 2297726e07a8SIlpo Järvinen if (unlikely(sk->sk_state == TCP_CLOSE)) 2298726e07a8SIlpo Järvinen return; 2299726e07a8SIlpo Järvinen 230099a1dec7SMel Gorman if (tcp_write_xmit(sk, cur_mss, nonagle, 0, 230199a1dec7SMel Gorman sk_gfp_atomic(sk, GFP_ATOMIC))) 23029e412ba7SIlpo Järvinen tcp_check_probe_timer(sk); 2303a762a980SDavid S. Miller } 2304a762a980SDavid S. Miller 2305c1b4a7e6SDavid S. Miller /* Send _single_ skb sitting at the send head. This function requires 2306c1b4a7e6SDavid S. Miller * true push pending frames to setup probe timer etc. 2307c1b4a7e6SDavid S. Miller */ 2308c1b4a7e6SDavid S. Miller void tcp_push_one(struct sock *sk, unsigned int mss_now) 2309c1b4a7e6SDavid S. Miller { 2310fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 2311c1b4a7e6SDavid S. Miller 2312c1b4a7e6SDavid S. Miller BUG_ON(!skb || skb->len < mss_now); 2313c1b4a7e6SDavid S. Miller 2314d5dd9175SIlpo Järvinen tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); 2315c1b4a7e6SDavid S. Miller } 2316c1b4a7e6SDavid S. Miller 23171da177e4SLinus Torvalds /* This function returns the amount that we can raise the 23181da177e4SLinus Torvalds * usable window based on the following constraints 23191da177e4SLinus Torvalds * 23201da177e4SLinus Torvalds * 1. The window can never be shrunk once it is offered (RFC 793) 23211da177e4SLinus Torvalds * 2. We limit memory per socket 23221da177e4SLinus Torvalds * 23231da177e4SLinus Torvalds * RFC 1122: 23241da177e4SLinus Torvalds * "the suggested [SWS] avoidance algorithm for the receiver is to keep 23251da177e4SLinus Torvalds * RECV.NEXT + RCV.WIN fixed until: 23261da177e4SLinus Torvalds * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 23271da177e4SLinus Torvalds * 23281da177e4SLinus Torvalds * i.e. don't raise the right edge of the window until you can raise 23291da177e4SLinus Torvalds * it at least MSS bytes. 23301da177e4SLinus Torvalds * 23311da177e4SLinus Torvalds * Unfortunately, the recommended algorithm breaks header prediction, 23321da177e4SLinus Torvalds * since header prediction assumes th->window stays fixed. 23331da177e4SLinus Torvalds * 23341da177e4SLinus Torvalds * Strictly speaking, keeping th->window fixed violates the receiver 23351da177e4SLinus Torvalds * side SWS prevention criteria. The problem is that under this rule 23361da177e4SLinus Torvalds * a stream of single byte packets will cause the right side of the 23371da177e4SLinus Torvalds * window to always advance by a single byte. 23381da177e4SLinus Torvalds * 23391da177e4SLinus Torvalds * Of course, if the sender implements sender side SWS prevention 23401da177e4SLinus Torvalds * then this will not be a problem. 23411da177e4SLinus Torvalds * 23421da177e4SLinus Torvalds * BSD seems to make the following compromise: 23431da177e4SLinus Torvalds * 23441da177e4SLinus Torvalds * If the free space is less than the 1/4 of the maximum 23451da177e4SLinus Torvalds * space available and the free space is less than 1/2 mss, 23461da177e4SLinus Torvalds * then set the window to 0. 23471da177e4SLinus Torvalds * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 23481da177e4SLinus Torvalds * Otherwise, just prevent the window from shrinking 23491da177e4SLinus Torvalds * and from being larger than the largest representable value. 23501da177e4SLinus Torvalds * 23511da177e4SLinus Torvalds * This prevents incremental opening of the window in the regime 23521da177e4SLinus Torvalds * where TCP is limited by the speed of the reader side taking 23531da177e4SLinus Torvalds * data out of the TCP receive queue. It does nothing about 23541da177e4SLinus Torvalds * those cases where the window is constrained on the sender side 23551da177e4SLinus Torvalds * because the pipeline is full. 23561da177e4SLinus Torvalds * 23571da177e4SLinus Torvalds * BSD also seems to "accidentally" limit itself to windows that are a 23581da177e4SLinus Torvalds * multiple of MSS, at least until the free space gets quite small. 23591da177e4SLinus Torvalds * This would appear to be a side effect of the mbuf implementation. 23601da177e4SLinus Torvalds * Combining these two algorithms results in the observed behavior 23611da177e4SLinus Torvalds * of having a fixed window size at almost all times. 23621da177e4SLinus Torvalds * 23631da177e4SLinus Torvalds * Below we obtain similar behavior by forcing the offered window to 23641da177e4SLinus Torvalds * a multiple of the mss when it is feasible to do so. 23651da177e4SLinus Torvalds * 23661da177e4SLinus Torvalds * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 23671da177e4SLinus Torvalds * Regular options like TIMESTAMP are taken into account. 23681da177e4SLinus Torvalds */ 23691da177e4SLinus Torvalds u32 __tcp_select_window(struct sock *sk) 23701da177e4SLinus Torvalds { 2371463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 23721da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2373caa20d9aSStephen Hemminger /* MSS for the peer's data. Previous versions used mss_clamp 23741da177e4SLinus Torvalds * here. I don't know if the value based on our guesses 23751da177e4SLinus Torvalds * of peer's MSS is better for the performance. It's more correct 23761da177e4SLinus Torvalds * but may be worse for the performance because of rcv_mss 23771da177e4SLinus Torvalds * fluctuations. --SAW 1998/11/1 23781da177e4SLinus Torvalds */ 2379463c84b9SArnaldo Carvalho de Melo int mss = icsk->icsk_ack.rcv_mss; 23801da177e4SLinus Torvalds int free_space = tcp_space(sk); 238186c1a045SFlorian Westphal int allowed_space = tcp_full_space(sk); 238286c1a045SFlorian Westphal int full_space = min_t(int, tp->window_clamp, allowed_space); 23831da177e4SLinus Torvalds int window; 23841da177e4SLinus Torvalds 23851da177e4SLinus Torvalds if (mss > full_space) 23861da177e4SLinus Torvalds mss = full_space; 23871da177e4SLinus Torvalds 2388b92edbe0SEric Dumazet if (free_space < (full_space >> 1)) { 2389463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.quick = 0; 23901da177e4SLinus Torvalds 2391b8da51ebSEric Dumazet if (tcp_under_memory_pressure(sk)) 2392056834d9SIlpo Järvinen tp->rcv_ssthresh = min(tp->rcv_ssthresh, 2393056834d9SIlpo Järvinen 4U * tp->advmss); 23941da177e4SLinus Torvalds 239586c1a045SFlorian Westphal /* free_space might become our new window, make sure we don't 239686c1a045SFlorian Westphal * increase it due to wscale. 239786c1a045SFlorian Westphal */ 239886c1a045SFlorian Westphal free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale); 239986c1a045SFlorian Westphal 240086c1a045SFlorian Westphal /* if free space is less than mss estimate, or is below 1/16th 240186c1a045SFlorian Westphal * of the maximum allowed, try to move to zero-window, else 240286c1a045SFlorian Westphal * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and 240386c1a045SFlorian Westphal * new incoming data is dropped due to memory limits. 240486c1a045SFlorian Westphal * With large window, mss test triggers way too late in order 240586c1a045SFlorian Westphal * to announce zero window in time before rmem limit kicks in. 240686c1a045SFlorian Westphal */ 240786c1a045SFlorian Westphal if (free_space < (allowed_space >> 4) || free_space < mss) 24081da177e4SLinus Torvalds return 0; 24091da177e4SLinus Torvalds } 24101da177e4SLinus Torvalds 24111da177e4SLinus Torvalds if (free_space > tp->rcv_ssthresh) 24121da177e4SLinus Torvalds free_space = tp->rcv_ssthresh; 24131da177e4SLinus Torvalds 24141da177e4SLinus Torvalds /* Don't do rounding if we are using window scaling, since the 24151da177e4SLinus Torvalds * scaled window will not line up with the MSS boundary anyway. 24161da177e4SLinus Torvalds */ 24171da177e4SLinus Torvalds window = tp->rcv_wnd; 24181da177e4SLinus Torvalds if (tp->rx_opt.rcv_wscale) { 24191da177e4SLinus Torvalds window = free_space; 24201da177e4SLinus Torvalds 24211da177e4SLinus Torvalds /* Advertise enough space so that it won't get scaled away. 24221da177e4SLinus Torvalds * Import case: prevent zero window announcement if 24231da177e4SLinus Torvalds * 1<<rcv_wscale > mss. 24241da177e4SLinus Torvalds */ 24251da177e4SLinus Torvalds if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window) 24261da177e4SLinus Torvalds window = (((window >> tp->rx_opt.rcv_wscale) + 1) 24271da177e4SLinus Torvalds << tp->rx_opt.rcv_wscale); 24281da177e4SLinus Torvalds } else { 24291da177e4SLinus Torvalds /* Get the largest window that is a nice multiple of mss. 24301da177e4SLinus Torvalds * Window clamp already applied above. 24311da177e4SLinus Torvalds * If our current window offering is within 1 mss of the 24321da177e4SLinus Torvalds * free space we just keep it. This prevents the divide 24331da177e4SLinus Torvalds * and multiply from happening most of the time. 24341da177e4SLinus Torvalds * We also don't do any window rounding when the free space 24351da177e4SLinus Torvalds * is too small. 24361da177e4SLinus Torvalds */ 24371da177e4SLinus Torvalds if (window <= free_space - mss || window > free_space) 24381da177e4SLinus Torvalds window = (free_space / mss) * mss; 243984565070SJohn Heffner else if (mss == full_space && 2440b92edbe0SEric Dumazet free_space > window + (full_space >> 1)) 244184565070SJohn Heffner window = free_space; 24421da177e4SLinus Torvalds } 24431da177e4SLinus Torvalds 24441da177e4SLinus Torvalds return window; 24451da177e4SLinus Torvalds } 24461da177e4SLinus Torvalds 24474a17fc3aSIlpo Järvinen /* Collapses two adjacent SKB's during retransmission. */ 24484a17fc3aSIlpo Järvinen static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) 24491da177e4SLinus Torvalds { 24501da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2451fe067e8aSDavid S. Miller struct sk_buff *next_skb = tcp_write_queue_next(sk, skb); 2452058dc334SIlpo Järvinen int skb_size, next_skb_size; 24531da177e4SLinus Torvalds 2454058dc334SIlpo Järvinen skb_size = skb->len; 2455058dc334SIlpo Järvinen next_skb_size = next_skb->len; 24561da177e4SLinus Torvalds 2457058dc334SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); 24581da177e4SLinus Torvalds 24596859d494SIlpo Järvinen tcp_highest_sack_combine(sk, next_skb, skb); 2460a6963a6bSIlpo Järvinen 2461fe067e8aSDavid S. Miller tcp_unlink_write_queue(next_skb, sk); 24621da177e4SLinus Torvalds 2463058dc334SIlpo Järvinen skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size), 24641a4e2d09SArnaldo Carvalho de Melo next_skb_size); 24651da177e4SLinus Torvalds 246652d570aaSJarek Poplawski if (next_skb->ip_summed == CHECKSUM_PARTIAL) 246752d570aaSJarek Poplawski skb->ip_summed = CHECKSUM_PARTIAL; 24681da177e4SLinus Torvalds 246984fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 24701da177e4SLinus Torvalds skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); 24711da177e4SLinus Torvalds 24721da177e4SLinus Torvalds /* Update sequence range on original skb. */ 24731da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 24741da177e4SLinus Torvalds 2475e6c7d085SIlpo Järvinen /* Merge over control information. This moves PSH/FIN etc. over */ 24764de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; 24771da177e4SLinus Torvalds 24781da177e4SLinus Torvalds /* All done, get rid of second SKB and account for it so 24791da177e4SLinus Torvalds * packet counting does not break. 24801da177e4SLinus Torvalds */ 24814828e7f4SIlpo Järvinen TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; 2482b7689205SIlpo Järvinen 2483b7689205SIlpo Järvinen /* changed transmit queue under us so clear hints */ 2484ef9da47cSIlpo Järvinen tcp_clear_retrans_hints_partial(tp); 2485ef9da47cSIlpo Järvinen if (next_skb == tp->retransmit_skb_hint) 2486ef9da47cSIlpo Järvinen tp->retransmit_skb_hint = skb; 2487b7689205SIlpo Järvinen 2488797108d1SIlpo Järvinen tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); 2489797108d1SIlpo Järvinen 24903ab224beSHideo Aoki sk_wmem_free_skb(sk, next_skb); 24911da177e4SLinus Torvalds } 24921da177e4SLinus Torvalds 249367edfef7SAndi Kleen /* Check if coalescing SKBs is legal. */ 2494a2a385d6SEric Dumazet static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) 24954a17fc3aSIlpo Järvinen { 24964a17fc3aSIlpo Järvinen if (tcp_skb_pcount(skb) > 1) 2497a2a385d6SEric Dumazet return false; 24984a17fc3aSIlpo Järvinen /* TODO: SACK collapsing could be used to remove this condition */ 24994a17fc3aSIlpo Järvinen if (skb_shinfo(skb)->nr_frags != 0) 2500a2a385d6SEric Dumazet return false; 25014a17fc3aSIlpo Järvinen if (skb_cloned(skb)) 2502a2a385d6SEric Dumazet return false; 25034a17fc3aSIlpo Järvinen if (skb == tcp_send_head(sk)) 2504a2a385d6SEric Dumazet return false; 25054a17fc3aSIlpo Järvinen /* Some heurestics for collapsing over SACK'd could be invented */ 25064a17fc3aSIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 2507a2a385d6SEric Dumazet return false; 25084a17fc3aSIlpo Järvinen 2509a2a385d6SEric Dumazet return true; 25104a17fc3aSIlpo Järvinen } 25114a17fc3aSIlpo Järvinen 251267edfef7SAndi Kleen /* Collapse packets in the retransmit queue to make to create 251367edfef7SAndi Kleen * less packets on the wire. This is only done on retransmission. 251467edfef7SAndi Kleen */ 25154a17fc3aSIlpo Järvinen static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, 25164a17fc3aSIlpo Järvinen int space) 25174a17fc3aSIlpo Järvinen { 25184a17fc3aSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 25194a17fc3aSIlpo Järvinen struct sk_buff *skb = to, *tmp; 2520a2a385d6SEric Dumazet bool first = true; 25214a17fc3aSIlpo Järvinen 25224a17fc3aSIlpo Järvinen if (!sysctl_tcp_retrans_collapse) 25234a17fc3aSIlpo Järvinen return; 25244de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 25254a17fc3aSIlpo Järvinen return; 25264a17fc3aSIlpo Järvinen 25274a17fc3aSIlpo Järvinen tcp_for_write_queue_from_safe(skb, tmp, sk) { 25284a17fc3aSIlpo Järvinen if (!tcp_can_collapse(sk, skb)) 25294a17fc3aSIlpo Järvinen break; 25304a17fc3aSIlpo Järvinen 25314a17fc3aSIlpo Järvinen space -= skb->len; 25324a17fc3aSIlpo Järvinen 25334a17fc3aSIlpo Järvinen if (first) { 2534a2a385d6SEric Dumazet first = false; 25354a17fc3aSIlpo Järvinen continue; 25364a17fc3aSIlpo Järvinen } 25374a17fc3aSIlpo Järvinen 25384a17fc3aSIlpo Järvinen if (space < 0) 25394a17fc3aSIlpo Järvinen break; 25404a17fc3aSIlpo Järvinen /* Punt if not enough space exists in the first SKB for 25414a17fc3aSIlpo Järvinen * the data in the second 25424a17fc3aSIlpo Järvinen */ 2543a21d4572SEric Dumazet if (skb->len > skb_availroom(to)) 25444a17fc3aSIlpo Järvinen break; 25454a17fc3aSIlpo Järvinen 25464a17fc3aSIlpo Järvinen if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) 25474a17fc3aSIlpo Järvinen break; 25484a17fc3aSIlpo Järvinen 25494a17fc3aSIlpo Järvinen tcp_collapse_retrans(sk, to); 25504a17fc3aSIlpo Järvinen } 25514a17fc3aSIlpo Järvinen } 25524a17fc3aSIlpo Järvinen 25531da177e4SLinus Torvalds /* This retransmits one SKB. Policy decisions and retransmit queue 25541da177e4SLinus Torvalds * state updates are done by the caller. Returns non-zero if an 25551da177e4SLinus Torvalds * error occurred which prevented the send. 25561da177e4SLinus Torvalds */ 255793b174adSYuchung Cheng int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) 25581da177e4SLinus Torvalds { 25591da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 25605d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 25617d227cd2SSridhar Samudrala unsigned int cur_mss; 2562c84a5711SYuchung Cheng int err; 25631da177e4SLinus Torvalds 25645d424d5aSJohn Heffner /* Inconslusive MTU probe */ 25655d424d5aSJohn Heffner if (icsk->icsk_mtup.probe_size) { 25665d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 25675d424d5aSJohn Heffner } 25685d424d5aSJohn Heffner 25691da177e4SLinus Torvalds /* Do not sent more than we queued. 1/4 is reserved for possible 2570caa20d9aSStephen Hemminger * copying overhead: fragmentation, tunneling, mangling etc. 25711da177e4SLinus Torvalds */ 25721da177e4SLinus Torvalds if (atomic_read(&sk->sk_wmem_alloc) > 25731da177e4SLinus Torvalds min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf)) 25741da177e4SLinus Torvalds return -EAGAIN; 25751da177e4SLinus Torvalds 25761f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 25771f3279aeSEric Dumazet return -EBUSY; 25781f3279aeSEric Dumazet 25791da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 25801da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 25811da177e4SLinus Torvalds BUG(); 25821da177e4SLinus Torvalds if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 25831da177e4SLinus Torvalds return -ENOMEM; 25841da177e4SLinus Torvalds } 25851da177e4SLinus Torvalds 25867d227cd2SSridhar Samudrala if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 25877d227cd2SSridhar Samudrala return -EHOSTUNREACH; /* Routing failure or similar. */ 25887d227cd2SSridhar Samudrala 25890c54b85fSIlpo Järvinen cur_mss = tcp_current_mss(sk); 25907d227cd2SSridhar Samudrala 25911da177e4SLinus Torvalds /* If receiver has shrunk his window, and skb is out of 25921da177e4SLinus Torvalds * new window, do not retransmit it. The exception is the 25931da177e4SLinus Torvalds * case, when window is shrunk to zero. In this case 25941da177e4SLinus Torvalds * our retransmit serves as a zero window probe. 25951da177e4SLinus Torvalds */ 25969d4fb27dSJoe Perches if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && 25979d4fb27dSJoe Perches TCP_SKB_CB(skb)->seq != tp->snd_una) 25981da177e4SLinus Torvalds return -EAGAIN; 25991da177e4SLinus Torvalds 26001da177e4SLinus Torvalds if (skb->len > cur_mss) { 26016cc55e09SOctavian Purdila if (tcp_fragment(sk, skb, cur_mss, cur_mss, GFP_ATOMIC)) 26021da177e4SLinus Torvalds return -ENOMEM; /* We'll try again later. */ 260302276f3cSIlpo Järvinen } else { 26049eb9362eSIlpo Järvinen int oldpcount = tcp_skb_pcount(skb); 26059eb9362eSIlpo Järvinen 26069eb9362eSIlpo Järvinen if (unlikely(oldpcount > 1)) { 2607c52e2421SEric Dumazet if (skb_unclone(skb, GFP_ATOMIC)) 2608c52e2421SEric Dumazet return -ENOMEM; 26095bbb432cSEric Dumazet tcp_init_tso_segs(skb, cur_mss); 26109eb9362eSIlpo Järvinen tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb)); 26119eb9362eSIlpo Järvinen } 26121da177e4SLinus Torvalds } 26131da177e4SLinus Torvalds 261449213555SDaniel Borkmann /* RFC3168, section 6.1.1.1. ECN fallback */ 261549213555SDaniel Borkmann if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN) 261649213555SDaniel Borkmann tcp_ecn_clear_syn(sk, skb); 261749213555SDaniel Borkmann 26181da177e4SLinus Torvalds tcp_retrans_try_collapse(sk, skb, cur_mss); 26191da177e4SLinus Torvalds 26201da177e4SLinus Torvalds /* Make a copy, if the first transmission SKB clone we made 26211da177e4SLinus Torvalds * is still in somebody's hands, else make a clone. 26221da177e4SLinus Torvalds */ 26231da177e4SLinus Torvalds 262450bceae9SThomas Graf /* make sure skb->data is aligned on arches that require it 262550bceae9SThomas Graf * and check if ack-trimming & collapsing extended the headroom 262650bceae9SThomas Graf * beyond what csum_start can cover. 262750bceae9SThomas Graf */ 262850bceae9SThomas Graf if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) || 262950bceae9SThomas Graf skb_headroom(skb) >= 0xFFFF)) { 2630117632e6SEric Dumazet struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER, 2631117632e6SEric Dumazet GFP_ATOMIC); 2632c84a5711SYuchung Cheng err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) : 2633117632e6SEric Dumazet -ENOBUFS; 2634117632e6SEric Dumazet } else { 2635c84a5711SYuchung Cheng err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 2636117632e6SEric Dumazet } 2637c84a5711SYuchung Cheng 2638fc9f3501SEric Dumazet if (likely(!err)) { 2639c84a5711SYuchung Cheng TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS; 2640fc9f3501SEric Dumazet /* Update global TCP statistics. */ 2641fc9f3501SEric Dumazet TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); 2642fc9f3501SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 2643fc9f3501SEric Dumazet NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 2644fc9f3501SEric Dumazet tp->total_retrans++; 2645fc9f3501SEric Dumazet } 2646c84a5711SYuchung Cheng return err; 264793b174adSYuchung Cheng } 264893b174adSYuchung Cheng 264993b174adSYuchung Cheng int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) 265093b174adSYuchung Cheng { 265193b174adSYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 265293b174adSYuchung Cheng int err = __tcp_retransmit_skb(sk, skb); 26531da177e4SLinus Torvalds 26541da177e4SLinus Torvalds if (err == 0) { 26551da177e4SLinus Torvalds #if FASTRETRANS_DEBUG > 0 26561da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 2657e87cc472SJoe Perches net_dbg_ratelimited("retrans_out leaked\n"); 26581da177e4SLinus Torvalds } 26591da177e4SLinus Torvalds #endif 2660b08d6cb2SIlpo Järvinen if (!tp->retrans_out) 2661b08d6cb2SIlpo Järvinen tp->lost_retrans_low = tp->snd_nxt; 26621da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 26631da177e4SLinus Torvalds tp->retrans_out += tcp_skb_pcount(skb); 26641da177e4SLinus Torvalds 26651da177e4SLinus Torvalds /* Save stamp of the first retransmit. */ 26661da177e4SLinus Torvalds if (!tp->retrans_stamp) 26677faee5c0SEric Dumazet tp->retrans_stamp = tcp_skb_timestamp(skb); 26681da177e4SLinus Torvalds 26691da177e4SLinus Torvalds /* snd_nxt is stored to detect loss of retransmitted segment, 26701da177e4SLinus Torvalds * see tcp_input.c tcp_sacktag_write_queue(). 26711da177e4SLinus Torvalds */ 26721da177e4SLinus Torvalds TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt; 26731f3279aeSEric Dumazet } else if (err != -EBUSY) { 267424ab6becSYuchung Cheng NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL); 26751da177e4SLinus Torvalds } 26766e08d5e3SYuchung Cheng 26776e08d5e3SYuchung Cheng if (tp->undo_retrans < 0) 26786e08d5e3SYuchung Cheng tp->undo_retrans = 0; 26796e08d5e3SYuchung Cheng tp->undo_retrans += tcp_skb_pcount(skb); 26801da177e4SLinus Torvalds return err; 26811da177e4SLinus Torvalds } 26821da177e4SLinus Torvalds 268367edfef7SAndi Kleen /* Check if we forward retransmits are possible in the current 268467edfef7SAndi Kleen * window/congestion state. 268567edfef7SAndi Kleen */ 2686a2a385d6SEric Dumazet static bool tcp_can_forward_retransmit(struct sock *sk) 2687b5afe7bcSIlpo Järvinen { 2688b5afe7bcSIlpo Järvinen const struct inet_connection_sock *icsk = inet_csk(sk); 2689cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 2690b5afe7bcSIlpo Järvinen 2691b5afe7bcSIlpo Järvinen /* Forward retransmissions are possible only during Recovery. */ 2692b5afe7bcSIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Recovery) 2693a2a385d6SEric Dumazet return false; 2694b5afe7bcSIlpo Järvinen 2695b5afe7bcSIlpo Järvinen /* No forward retransmissions in Reno are possible. */ 2696b5afe7bcSIlpo Järvinen if (tcp_is_reno(tp)) 2697a2a385d6SEric Dumazet return false; 2698b5afe7bcSIlpo Järvinen 2699b5afe7bcSIlpo Järvinen /* Yeah, we have to make difficult choice between forward transmission 2700b5afe7bcSIlpo Järvinen * and retransmission... Both ways have their merits... 2701b5afe7bcSIlpo Järvinen * 2702b5afe7bcSIlpo Järvinen * For now we do not retransmit anything, while we have some new 2703b5afe7bcSIlpo Järvinen * segments to send. In the other cases, follow rule 3 for 2704b5afe7bcSIlpo Järvinen * NextSeg() specified in RFC3517. 2705b5afe7bcSIlpo Järvinen */ 2706b5afe7bcSIlpo Järvinen 2707b5afe7bcSIlpo Järvinen if (tcp_may_send_now(sk)) 2708a2a385d6SEric Dumazet return false; 2709b5afe7bcSIlpo Järvinen 2710a2a385d6SEric Dumazet return true; 2711b5afe7bcSIlpo Järvinen } 2712b5afe7bcSIlpo Järvinen 27131da177e4SLinus Torvalds /* This gets called after a retransmit timeout, and the initially 27141da177e4SLinus Torvalds * retransmitted data is acknowledged. It tries to continue 27151da177e4SLinus Torvalds * resending the rest of the retransmit queue, until either 27161da177e4SLinus Torvalds * we've sent it all or the congestion window limit is reached. 27171da177e4SLinus Torvalds * If doing SACK, the first ACK which comes back for a timeout 27181da177e4SLinus Torvalds * based retransmit packet might feed us FACK information again. 27191da177e4SLinus Torvalds * If so, we use it to avoid unnecessarily retransmissions. 27201da177e4SLinus Torvalds */ 27211da177e4SLinus Torvalds void tcp_xmit_retransmit_queue(struct sock *sk) 27221da177e4SLinus Torvalds { 27236687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 27241da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 27251da177e4SLinus Torvalds struct sk_buff *skb; 27260e1c54c2SIlpo Järvinen struct sk_buff *hole = NULL; 2727618d9f25SIlpo Järvinen u32 last_lost; 272861eb55f4SIlpo Järvinen int mib_idx; 27290e1c54c2SIlpo Järvinen int fwd_rexmitting = 0; 27306a438bbeSStephen Hemminger 273145e77d31SIlpo Järvinen if (!tp->packets_out) 273245e77d31SIlpo Järvinen return; 273345e77d31SIlpo Järvinen 273408ebd172SIlpo Järvinen if (!tp->lost_out) 273508ebd172SIlpo Järvinen tp->retransmit_high = tp->snd_una; 273608ebd172SIlpo Järvinen 2737618d9f25SIlpo Järvinen if (tp->retransmit_skb_hint) { 27386a438bbeSStephen Hemminger skb = tp->retransmit_skb_hint; 2739618d9f25SIlpo Järvinen last_lost = TCP_SKB_CB(skb)->end_seq; 2740618d9f25SIlpo Järvinen if (after(last_lost, tp->retransmit_high)) 2741618d9f25SIlpo Järvinen last_lost = tp->retransmit_high; 2742618d9f25SIlpo Järvinen } else { 2743fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 2744618d9f25SIlpo Järvinen last_lost = tp->snd_una; 2745618d9f25SIlpo Järvinen } 27461da177e4SLinus Torvalds 2747fe067e8aSDavid S. Miller tcp_for_write_queue_from(skb, sk) { 27481da177e4SLinus Torvalds __u8 sacked = TCP_SKB_CB(skb)->sacked; 27491da177e4SLinus Torvalds 2750fe067e8aSDavid S. Miller if (skb == tcp_send_head(sk)) 2751fe067e8aSDavid S. Miller break; 27526a438bbeSStephen Hemminger /* we could do better than to assign each time */ 275351456b29SIan Morris if (!hole) 27546a438bbeSStephen Hemminger tp->retransmit_skb_hint = skb; 27556a438bbeSStephen Hemminger 27561da177e4SLinus Torvalds /* Assume this retransmit will generate 27571da177e4SLinus Torvalds * only one packet for congestion window 27581da177e4SLinus Torvalds * calculation purposes. This works because 27591da177e4SLinus Torvalds * tcp_retransmit_skb() will chop up the 27601da177e4SLinus Torvalds * packet to be MSS sized and all the 27611da177e4SLinus Torvalds * packet counting works out. 27621da177e4SLinus Torvalds */ 27631da177e4SLinus Torvalds if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) 27641da177e4SLinus Torvalds return; 27650e1c54c2SIlpo Järvinen 27660e1c54c2SIlpo Järvinen if (fwd_rexmitting) { 27670e1c54c2SIlpo Järvinen begin_fwd: 27680e1c54c2SIlpo Järvinen if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp))) 2769006f582cSIlpo Järvinen break; 27700e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFORWARDRETRANS; 27710e1c54c2SIlpo Järvinen 27720e1c54c2SIlpo Järvinen } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) { 2773618d9f25SIlpo Järvinen tp->retransmit_high = last_lost; 27740e1c54c2SIlpo Järvinen if (!tcp_can_forward_retransmit(sk)) 27750e1c54c2SIlpo Järvinen break; 27760e1c54c2SIlpo Järvinen /* Backtrack if necessary to non-L'ed skb */ 277700db4124SIan Morris if (hole) { 27780e1c54c2SIlpo Järvinen skb = hole; 27790e1c54c2SIlpo Järvinen hole = NULL; 27800e1c54c2SIlpo Järvinen } 27810e1c54c2SIlpo Järvinen fwd_rexmitting = 1; 27820e1c54c2SIlpo Järvinen goto begin_fwd; 27830e1c54c2SIlpo Järvinen 27840e1c54c2SIlpo Järvinen } else if (!(sacked & TCPCB_LOST)) { 278551456b29SIan Morris if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) 27860e1c54c2SIlpo Järvinen hole = skb; 278761eb55f4SIlpo Järvinen continue; 27881da177e4SLinus Torvalds 27890e1c54c2SIlpo Järvinen } else { 2790618d9f25SIlpo Järvinen last_lost = TCP_SKB_CB(skb)->end_seq; 27910e1c54c2SIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Loss) 27920e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFASTRETRANS; 27930e1c54c2SIlpo Järvinen else 27940e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; 27950e1c54c2SIlpo Järvinen } 27960e1c54c2SIlpo Järvinen 27970e1c54c2SIlpo Järvinen if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) 279861eb55f4SIlpo Järvinen continue; 279940b215e5SPavel Emelyanov 280024ab6becSYuchung Cheng if (tcp_retransmit_skb(sk, skb)) 28011da177e4SLinus Torvalds return; 280224ab6becSYuchung Cheng 2803de0744afSPavel Emelyanov NET_INC_STATS_BH(sock_net(sk), mib_idx); 28041da177e4SLinus Torvalds 2805684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 2806a262f0cdSNandita Dukkipati tp->prr_out += tcp_skb_pcount(skb); 2807a262f0cdSNandita Dukkipati 2808fe067e8aSDavid S. Miller if (skb == tcp_write_queue_head(sk)) 2809463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 28103f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, 28113f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 28121da177e4SLinus Torvalds } 28131da177e4SLinus Torvalds } 28141da177e4SLinus Torvalds 2815d83769a5SEric Dumazet /* We allow to exceed memory limits for FIN packets to expedite 2816d83769a5SEric Dumazet * connection tear down and (memory) recovery. 2817845704a5SEric Dumazet * Otherwise tcp_send_fin() could be tempted to either delay FIN 2818845704a5SEric Dumazet * or even be forced to close flow without any FIN. 2819a6c5ea4cSEric Dumazet * In general, we want to allow one skb per socket to avoid hangs 2820a6c5ea4cSEric Dumazet * with edge trigger epoll() 2821d83769a5SEric Dumazet */ 2822a6c5ea4cSEric Dumazet void sk_forced_mem_schedule(struct sock *sk, int size) 2823d83769a5SEric Dumazet { 2824d83769a5SEric Dumazet int amt, status; 2825d83769a5SEric Dumazet 2826d83769a5SEric Dumazet if (size <= sk->sk_forward_alloc) 2827d83769a5SEric Dumazet return; 2828d83769a5SEric Dumazet amt = sk_mem_pages(size); 2829d83769a5SEric Dumazet sk->sk_forward_alloc += amt * SK_MEM_QUANTUM; 2830d83769a5SEric Dumazet sk_memory_allocated_add(sk, amt, &status); 2831d83769a5SEric Dumazet } 2832d83769a5SEric Dumazet 2833845704a5SEric Dumazet /* Send a FIN. The caller locks the socket for us. 2834845704a5SEric Dumazet * We should try to send a FIN packet really hard, but eventually give up. 28351da177e4SLinus Torvalds */ 28361da177e4SLinus Torvalds void tcp_send_fin(struct sock *sk) 28371da177e4SLinus Torvalds { 2838845704a5SEric Dumazet struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk); 28391da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 28401da177e4SLinus Torvalds 2841845704a5SEric Dumazet /* Optimization, tack on the FIN if we have one skb in write queue and 2842845704a5SEric Dumazet * this skb was not yet sent, or we are under memory pressure. 2843845704a5SEric Dumazet * Note: in the latter case, FIN packet will be sent after a timeout, 2844845704a5SEric Dumazet * as TCP stack thinks it has already been transmitted. 28451da177e4SLinus Torvalds */ 2846b8da51ebSEric Dumazet if (tskb && (tcp_send_head(sk) || tcp_under_memory_pressure(sk))) { 2847845704a5SEric Dumazet coalesce: 2848845704a5SEric Dumazet TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN; 2849845704a5SEric Dumazet TCP_SKB_CB(tskb)->end_seq++; 28501da177e4SLinus Torvalds tp->write_seq++; 2851845704a5SEric Dumazet if (!tcp_send_head(sk)) { 2852845704a5SEric Dumazet /* This means tskb was already sent. 2853845704a5SEric Dumazet * Pretend we included the FIN on previous transmit. 2854845704a5SEric Dumazet * We need to set tp->snd_nxt to the value it would have 2855845704a5SEric Dumazet * if FIN had been sent. This is because retransmit path 2856845704a5SEric Dumazet * does not change tp->snd_nxt. 2857845704a5SEric Dumazet */ 2858845704a5SEric Dumazet tp->snd_nxt++; 2859845704a5SEric Dumazet return; 2860845704a5SEric Dumazet } 28611da177e4SLinus Torvalds } else { 2862845704a5SEric Dumazet skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation); 2863845704a5SEric Dumazet if (unlikely(!skb)) { 2864845704a5SEric Dumazet if (tskb) 2865845704a5SEric Dumazet goto coalesce; 2866845704a5SEric Dumazet return; 28671da177e4SLinus Torvalds } 2868d83769a5SEric Dumazet skb_reserve(skb, MAX_TCP_HEADER); 2869a6c5ea4cSEric Dumazet sk_forced_mem_schedule(sk, skb->truesize); 28701da177e4SLinus Torvalds /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 2871e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tp->write_seq, 2872a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_FIN); 28731da177e4SLinus Torvalds tcp_queue_skb(sk, skb); 28741da177e4SLinus Torvalds } 2875845704a5SEric Dumazet __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF); 28761da177e4SLinus Torvalds } 28771da177e4SLinus Torvalds 28781da177e4SLinus Torvalds /* We get here when a process closes a file descriptor (either due to 28791da177e4SLinus Torvalds * an explicit close() or as a byproduct of exit()'ing) and there 28801da177e4SLinus Torvalds * was unread data in the receive queue. This behavior is recommended 288165bb723cSGerrit Renker * by RFC 2525, section 2.17. -DaveM 28821da177e4SLinus Torvalds */ 2883dd0fc66fSAl Viro void tcp_send_active_reset(struct sock *sk, gfp_t priority) 28841da177e4SLinus Torvalds { 28851da177e4SLinus Torvalds struct sk_buff *skb; 28861da177e4SLinus Torvalds 28871da177e4SLinus Torvalds /* NOTE: No TCP options attached and we never retransmit this. */ 28881da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, priority); 28891da177e4SLinus Torvalds if (!skb) { 28904e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 28911da177e4SLinus Torvalds return; 28921da177e4SLinus Torvalds } 28931da177e4SLinus Torvalds 28941da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 28951da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 2896e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), 2897a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_RST); 28981da177e4SLinus Torvalds /* Send it off. */ 2899dfb4b9dcSDavid S. Miller if (tcp_transmit_skb(sk, skb, 0, priority)) 29004e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 290126af65cbSSridhar Samudrala 290281cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); 29031da177e4SLinus Torvalds } 29041da177e4SLinus Torvalds 290567edfef7SAndi Kleen /* Send a crossed SYN-ACK during socket establishment. 290667edfef7SAndi Kleen * WARNING: This routine must only be called when we have already sent 29071da177e4SLinus Torvalds * a SYN packet that crossed the incoming SYN that caused this routine 29081da177e4SLinus Torvalds * to get called. If this assumption fails then the initial rcv_wnd 29091da177e4SLinus Torvalds * and rcv_wscale values will not be correct. 29101da177e4SLinus Torvalds */ 29111da177e4SLinus Torvalds int tcp_send_synack(struct sock *sk) 29121da177e4SLinus Torvalds { 29131da177e4SLinus Torvalds struct sk_buff *skb; 29141da177e4SLinus Torvalds 2915fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 291651456b29SIan Morris if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 291791df42beSJoe Perches pr_debug("%s: wrong queue state\n", __func__); 29181da177e4SLinus Torvalds return -EFAULT; 29191da177e4SLinus Torvalds } 29204de075e0SEric Dumazet if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { 29211da177e4SLinus Torvalds if (skb_cloned(skb)) { 29221da177e4SLinus Torvalds struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC); 292351456b29SIan Morris if (!nskb) 29241da177e4SLinus Torvalds return -ENOMEM; 2925fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 2926f4a775d1SEric Dumazet __skb_header_release(nskb); 2927fe067e8aSDavid S. Miller __tcp_add_write_queue_head(sk, nskb); 29283ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 29293ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 29303ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 29311da177e4SLinus Torvalds skb = nskb; 29321da177e4SLinus Torvalds } 29331da177e4SLinus Torvalds 29344de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; 2935735d3831SFlorian Westphal tcp_ecn_send_synack(sk, skb); 29361da177e4SLinus Torvalds } 2937dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 29381da177e4SLinus Torvalds } 29391da177e4SLinus Torvalds 29404aea39c1SEric Dumazet /** 29414aea39c1SEric Dumazet * tcp_make_synack - Prepare a SYN-ACK. 29424aea39c1SEric Dumazet * sk: listener socket 29434aea39c1SEric Dumazet * dst: dst entry attached to the SYNACK 29444aea39c1SEric Dumazet * req: request_sock pointer 29454aea39c1SEric Dumazet * 29464aea39c1SEric Dumazet * Allocate one skb and build a SYNACK packet. 29474aea39c1SEric Dumazet * @dst is consumed : Caller should not use it again. 29484aea39c1SEric Dumazet */ 29491da177e4SLinus Torvalds struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst, 2950e6b4d113SWilliam Allen Simpson struct request_sock *req, 29518336886fSJerry Chu struct tcp_fastopen_cookie *foc) 29521da177e4SLinus Torvalds { 2953bd0388aeSWilliam Allen Simpson struct tcp_out_options opts; 29542e6599cbSArnaldo Carvalho de Melo struct inet_request_sock *ireq = inet_rsk(req); 29551da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 29561da177e4SLinus Torvalds struct tcphdr *th; 29571da177e4SLinus Torvalds struct sk_buff *skb; 295880f03e27SEric Dumazet struct tcp_md5sig_key *md5 = NULL; 2959bd0388aeSWilliam Allen Simpson int tcp_header_size; 2960f5fff5dcSTom Quetchenbach int mss; 29611da177e4SLinus Torvalds 2962a0b8486cSEric Dumazet skb = sock_wmalloc(sk, MAX_TCP_HEADER, 1, GFP_ATOMIC); 29634aea39c1SEric Dumazet if (unlikely(!skb)) { 29644aea39c1SEric Dumazet dst_release(dst); 29651da177e4SLinus Torvalds return NULL; 29664aea39c1SEric Dumazet } 29671da177e4SLinus Torvalds /* Reserve space for headers. */ 29681da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 29691da177e4SLinus Torvalds 29704aea39c1SEric Dumazet skb_dst_set(skb, dst); 29711da177e4SLinus Torvalds 29720dbaee3bSDavid S. Miller mss = dst_metric_advmss(dst); 2973f5fff5dcSTom Quetchenbach if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss) 2974f5fff5dcSTom Quetchenbach mss = tp->rx_opt.user_mss; 2975f5fff5dcSTom Quetchenbach 297633ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 29778b5f12d0SFlorian Westphal #ifdef CONFIG_SYN_COOKIES 29788b5f12d0SFlorian Westphal if (unlikely(req->cookie_ts)) 29797faee5c0SEric Dumazet skb->skb_mstamp.stamp_jiffies = cookie_init_timestamp(req); 29808b5f12d0SFlorian Westphal else 29818b5f12d0SFlorian Westphal #endif 29827faee5c0SEric Dumazet skb_mstamp_get(&skb->skb_mstamp); 298380f03e27SEric Dumazet 298480f03e27SEric Dumazet #ifdef CONFIG_TCP_MD5SIG 298580f03e27SEric Dumazet rcu_read_lock(); 2986fd3a154aSEric Dumazet md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req)); 298780f03e27SEric Dumazet #endif 298880f03e27SEric Dumazet tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5, 29891a2c6181SChristoph Paasch foc) + sizeof(*th); 299033ad798cSAdam Langley 2991aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 2992aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 29931da177e4SLinus Torvalds 2994aa8223c7SArnaldo Carvalho de Melo th = tcp_hdr(skb); 29951da177e4SLinus Torvalds memset(th, 0, sizeof(struct tcphdr)); 29961da177e4SLinus Torvalds th->syn = 1; 29971da177e4SLinus Torvalds th->ack = 1; 2998735d3831SFlorian Westphal tcp_ecn_make_synack(req, th, sk); 2999b44084c2SEric Dumazet th->source = htons(ireq->ir_num); 3000634fb979SEric Dumazet th->dest = ireq->ir_rmt_port; 3001e870a8efSIlpo Järvinen /* Setting of flags are superfluous here for callers (and ECE is 3002e870a8efSIlpo Järvinen * not even correctly set) 3003e870a8efSIlpo Järvinen */ 3004e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn, 3005a3433f35SChangli Gao TCPHDR_SYN | TCPHDR_ACK); 30064957faadSWilliam Allen Simpson 30071da177e4SLinus Torvalds th->seq = htonl(TCP_SKB_CB(skb)->seq); 30088336886fSJerry Chu /* XXX data is queued and acked as is. No buffer/window check */ 30098336886fSJerry Chu th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt); 30101da177e4SLinus Torvalds 30111da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 3012600ff0c2SIlpo Järvinen th->window = htons(min(req->rcv_wnd, 65535U)); 3013bd0388aeSWilliam Allen Simpson tcp_options_write((__be32 *)(th + 1), tp, &opts); 30141da177e4SLinus Torvalds th->doff = (tcp_header_size >> 2); 3015a0b8486cSEric Dumazet TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_OUTSEGS); 3016cfb6eeb4SYOSHIFUJI Hideaki 3017cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 3018cfb6eeb4SYOSHIFUJI Hideaki /* Okay, we have all we need - do the md5 hash if needed */ 301980f03e27SEric Dumazet if (md5) 3020bd0388aeSWilliam Allen Simpson tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, 302139f8e58eSEric Dumazet md5, req_to_sk(req), skb); 302280f03e27SEric Dumazet rcu_read_unlock(); 3023cfb6eeb4SYOSHIFUJI Hideaki #endif 3024cfb6eeb4SYOSHIFUJI Hideaki 3025b50edd78SEric Dumazet /* Do not fool tcpdump (if any), clean our debris */ 3026b50edd78SEric Dumazet skb->tstamp.tv64 = 0; 30271da177e4SLinus Torvalds return skb; 30281da177e4SLinus Torvalds } 30294bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_make_synack); 30301da177e4SLinus Torvalds 303181164413SDaniel Borkmann static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst) 303281164413SDaniel Borkmann { 303381164413SDaniel Borkmann struct inet_connection_sock *icsk = inet_csk(sk); 303481164413SDaniel Borkmann const struct tcp_congestion_ops *ca; 303581164413SDaniel Borkmann u32 ca_key = dst_metric(dst, RTAX_CC_ALGO); 303681164413SDaniel Borkmann 303781164413SDaniel Borkmann if (ca_key == TCP_CA_UNSPEC) 303881164413SDaniel Borkmann return; 303981164413SDaniel Borkmann 304081164413SDaniel Borkmann rcu_read_lock(); 304181164413SDaniel Borkmann ca = tcp_ca_find_key(ca_key); 304281164413SDaniel Borkmann if (likely(ca && try_module_get(ca->owner))) { 304381164413SDaniel Borkmann module_put(icsk->icsk_ca_ops->owner); 304481164413SDaniel Borkmann icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst); 304581164413SDaniel Borkmann icsk->icsk_ca_ops = ca; 304681164413SDaniel Borkmann } 304781164413SDaniel Borkmann rcu_read_unlock(); 304881164413SDaniel Borkmann } 304981164413SDaniel Borkmann 305067edfef7SAndi Kleen /* Do all connect socket setups that can be done AF independent. */ 3051f7e56a76Sstephen hemminger static void tcp_connect_init(struct sock *sk) 30521da177e4SLinus Torvalds { 3053cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 30541da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 30551da177e4SLinus Torvalds __u8 rcv_wscale; 30561da177e4SLinus Torvalds 30571da177e4SLinus Torvalds /* We'll fix this up when we get a response from the other end. 30581da177e4SLinus Torvalds * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 30591da177e4SLinus Torvalds */ 30601da177e4SLinus Torvalds tp->tcp_header_len = sizeof(struct tcphdr) + 3061bb5b7c11SDavid S. Miller (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0); 30621da177e4SLinus Torvalds 3063cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 306400db4124SIan Morris if (tp->af_specific->md5_lookup(sk, sk)) 3065cfb6eeb4SYOSHIFUJI Hideaki tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 3066cfb6eeb4SYOSHIFUJI Hideaki #endif 3067cfb6eeb4SYOSHIFUJI Hideaki 30681da177e4SLinus Torvalds /* If user gave his TCP_MAXSEG, record it to clamp */ 30691da177e4SLinus Torvalds if (tp->rx_opt.user_mss) 30701da177e4SLinus Torvalds tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 30711da177e4SLinus Torvalds tp->max_window = 0; 30725d424d5aSJohn Heffner tcp_mtup_init(sk); 30731da177e4SLinus Torvalds tcp_sync_mss(sk, dst_mtu(dst)); 30741da177e4SLinus Torvalds 307581164413SDaniel Borkmann tcp_ca_dst_init(sk, dst); 307681164413SDaniel Borkmann 30771da177e4SLinus Torvalds if (!tp->window_clamp) 30781da177e4SLinus Torvalds tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 30790dbaee3bSDavid S. Miller tp->advmss = dst_metric_advmss(dst); 3080f5fff5dcSTom Quetchenbach if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss) 3081f5fff5dcSTom Quetchenbach tp->advmss = tp->rx_opt.user_mss; 3082f5fff5dcSTom Quetchenbach 30831da177e4SLinus Torvalds tcp_initialize_rcv_mss(sk); 30841da177e4SLinus Torvalds 3085e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 3086e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 3087e88c64f0SHagen Paul Pfeifer (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) 3088e88c64f0SHagen Paul Pfeifer tp->window_clamp = tcp_full_space(sk); 3089e88c64f0SHagen Paul Pfeifer 30901da177e4SLinus Torvalds tcp_select_initial_window(tcp_full_space(sk), 30911da177e4SLinus Torvalds tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 30921da177e4SLinus Torvalds &tp->rcv_wnd, 30931da177e4SLinus Torvalds &tp->window_clamp, 3094bb5b7c11SDavid S. Miller sysctl_tcp_window_scaling, 309531d12926Slaurent chavey &rcv_wscale, 309631d12926Slaurent chavey dst_metric(dst, RTAX_INITRWND)); 30971da177e4SLinus Torvalds 30981da177e4SLinus Torvalds tp->rx_opt.rcv_wscale = rcv_wscale; 30991da177e4SLinus Torvalds tp->rcv_ssthresh = tp->rcv_wnd; 31001da177e4SLinus Torvalds 31011da177e4SLinus Torvalds sk->sk_err = 0; 31021da177e4SLinus Torvalds sock_reset_flag(sk, SOCK_DONE); 31031da177e4SLinus Torvalds tp->snd_wnd = 0; 3104ee7537b6SHantzis Fotis tcp_init_wl(tp, 0); 31051da177e4SLinus Torvalds tp->snd_una = tp->write_seq; 31061da177e4SLinus Torvalds tp->snd_sml = tp->write_seq; 310733f5f57eSIlpo Järvinen tp->snd_up = tp->write_seq; 3108370816aeSPavel Emelyanov tp->snd_nxt = tp->write_seq; 3109ee995283SPavel Emelyanov 3110ee995283SPavel Emelyanov if (likely(!tp->repair)) 31111da177e4SLinus Torvalds tp->rcv_nxt = 0; 3112c7781a6eSAndrew Vagin else 3113c7781a6eSAndrew Vagin tp->rcv_tstamp = tcp_time_stamp; 3114ee995283SPavel Emelyanov tp->rcv_wup = tp->rcv_nxt; 3115ee995283SPavel Emelyanov tp->copied_seq = tp->rcv_nxt; 31161da177e4SLinus Torvalds 3117463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; 3118463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_retransmits = 0; 31191da177e4SLinus Torvalds tcp_clear_retrans(tp); 31201da177e4SLinus Torvalds } 31211da177e4SLinus Torvalds 3122783237e8SYuchung Cheng static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) 3123783237e8SYuchung Cheng { 3124783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3125783237e8SYuchung Cheng struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 3126783237e8SYuchung Cheng 3127783237e8SYuchung Cheng tcb->end_seq += skb->len; 3128f4a775d1SEric Dumazet __skb_header_release(skb); 3129783237e8SYuchung Cheng __tcp_add_write_queue_tail(sk, skb); 3130783237e8SYuchung Cheng sk->sk_wmem_queued += skb->truesize; 3131783237e8SYuchung Cheng sk_mem_charge(sk, skb->truesize); 3132783237e8SYuchung Cheng tp->write_seq = tcb->end_seq; 3133783237e8SYuchung Cheng tp->packets_out += tcp_skb_pcount(skb); 3134783237e8SYuchung Cheng } 3135783237e8SYuchung Cheng 3136783237e8SYuchung Cheng /* Build and send a SYN with data and (cached) Fast Open cookie. However, 3137783237e8SYuchung Cheng * queue a data-only packet after the regular SYN, such that regular SYNs 3138783237e8SYuchung Cheng * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges 3139783237e8SYuchung Cheng * only the SYN sequence, the data are retransmitted in the first ACK. 3140783237e8SYuchung Cheng * If cookie is not cached or other error occurs, falls back to send a 3141783237e8SYuchung Cheng * regular SYN with Fast Open cookie request option. 3142783237e8SYuchung Cheng */ 3143783237e8SYuchung Cheng static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) 3144783237e8SYuchung Cheng { 3145783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3146783237e8SYuchung Cheng struct tcp_fastopen_request *fo = tp->fastopen_req; 314757be5bdaSAl Viro int syn_loss = 0, space, err = 0, copied; 3148aab48743SYuchung Cheng unsigned long last_syn_loss = 0; 3149355a901eSEric Dumazet struct sk_buff *syn_data; 3150783237e8SYuchung Cheng 315167da22d2SYuchung Cheng tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ 3152aab48743SYuchung Cheng tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie, 3153aab48743SYuchung Cheng &syn_loss, &last_syn_loss); 3154aab48743SYuchung Cheng /* Recurring FO SYN losses: revert to regular handshake temporarily */ 3155aab48743SYuchung Cheng if (syn_loss > 1 && 3156aab48743SYuchung Cheng time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) { 3157aab48743SYuchung Cheng fo->cookie.len = -1; 3158aab48743SYuchung Cheng goto fallback; 3159aab48743SYuchung Cheng } 3160aab48743SYuchung Cheng 316167da22d2SYuchung Cheng if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE) 316267da22d2SYuchung Cheng fo->cookie.len = -1; 316367da22d2SYuchung Cheng else if (fo->cookie.len <= 0) 3164783237e8SYuchung Cheng goto fallback; 3165783237e8SYuchung Cheng 3166783237e8SYuchung Cheng /* MSS for SYN-data is based on cached MSS and bounded by PMTU and 3167783237e8SYuchung Cheng * user-MSS. Reserve maximum option space for middleboxes that add 3168783237e8SYuchung Cheng * private TCP options. The cost is reduced data space in SYN :( 3169783237e8SYuchung Cheng */ 3170783237e8SYuchung Cheng if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp) 3171783237e8SYuchung Cheng tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 31721b63edd6SYuchung Cheng space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) - 3173783237e8SYuchung Cheng MAX_TCP_OPTION_SPACE; 3174783237e8SYuchung Cheng 3175f5ddcbbbSEric Dumazet space = min_t(size_t, space, fo->size); 3176f5ddcbbbSEric Dumazet 3177f5ddcbbbSEric Dumazet /* limit to order-0 allocations */ 3178f5ddcbbbSEric Dumazet space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER)); 3179f5ddcbbbSEric Dumazet 3180eb934478SEric Dumazet syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false); 3181355a901eSEric Dumazet if (!syn_data) 3182783237e8SYuchung Cheng goto fallback; 3183355a901eSEric Dumazet syn_data->ip_summed = CHECKSUM_PARTIAL; 3184355a901eSEric Dumazet memcpy(syn_data->cb, syn->cb, sizeof(syn->cb)); 318557be5bdaSAl Viro copied = copy_from_iter(skb_put(syn_data, space), space, 318657be5bdaSAl Viro &fo->data->msg_iter); 318757be5bdaSAl Viro if (unlikely(!copied)) { 3188355a901eSEric Dumazet kfree_skb(syn_data); 3189783237e8SYuchung Cheng goto fallback; 3190783237e8SYuchung Cheng } 319157be5bdaSAl Viro if (copied != space) { 319257be5bdaSAl Viro skb_trim(syn_data, copied); 319357be5bdaSAl Viro space = copied; 319457be5bdaSAl Viro } 3195783237e8SYuchung Cheng 3196355a901eSEric Dumazet /* No more data pending in inet_wait_for_connect() */ 3197355a901eSEric Dumazet if (space == fo->size) 3198355a901eSEric Dumazet fo->data = NULL; 3199355a901eSEric Dumazet fo->copied = space; 3200783237e8SYuchung Cheng 3201355a901eSEric Dumazet tcp_connect_queue_skb(sk, syn_data); 3202355a901eSEric Dumazet 3203355a901eSEric Dumazet err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation); 3204355a901eSEric Dumazet 3205355a901eSEric Dumazet syn->skb_mstamp = syn_data->skb_mstamp; 3206355a901eSEric Dumazet 3207355a901eSEric Dumazet /* Now full SYN+DATA was cloned and sent (or not), 3208355a901eSEric Dumazet * remove the SYN from the original skb (syn_data) 3209355a901eSEric Dumazet * we keep in write queue in case of a retransmit, as we 3210355a901eSEric Dumazet * also have the SYN packet (with no data) in the same queue. 3211431a9124SEric Dumazet */ 3212355a901eSEric Dumazet TCP_SKB_CB(syn_data)->seq++; 3213355a901eSEric Dumazet TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH; 3214355a901eSEric Dumazet if (!err) { 321567da22d2SYuchung Cheng tp->syn_data = (fo->copied > 0); 3216f19c29e3SYuchung Cheng NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT); 3217783237e8SYuchung Cheng goto done; 3218783237e8SYuchung Cheng } 3219783237e8SYuchung Cheng 3220783237e8SYuchung Cheng fallback: 3221783237e8SYuchung Cheng /* Send a regular SYN with Fast Open cookie request option */ 3222783237e8SYuchung Cheng if (fo->cookie.len > 0) 3223783237e8SYuchung Cheng fo->cookie.len = 0; 3224783237e8SYuchung Cheng err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); 3225783237e8SYuchung Cheng if (err) 3226783237e8SYuchung Cheng tp->syn_fastopen = 0; 3227783237e8SYuchung Cheng done: 3228783237e8SYuchung Cheng fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ 3229783237e8SYuchung Cheng return err; 3230783237e8SYuchung Cheng } 3231783237e8SYuchung Cheng 323267edfef7SAndi Kleen /* Build a SYN and send it off. */ 32331da177e4SLinus Torvalds int tcp_connect(struct sock *sk) 32341da177e4SLinus Torvalds { 32351da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 32361da177e4SLinus Torvalds struct sk_buff *buff; 3237ee586811SEric Paris int err; 32381da177e4SLinus Torvalds 32391da177e4SLinus Torvalds tcp_connect_init(sk); 32401da177e4SLinus Torvalds 32412b916477SAndrey Vagin if (unlikely(tp->repair)) { 32422b916477SAndrey Vagin tcp_finish_connect(sk, NULL); 32432b916477SAndrey Vagin return 0; 32442b916477SAndrey Vagin } 32452b916477SAndrey Vagin 3246eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true); 3247355a901eSEric Dumazet if (unlikely(!buff)) 32481da177e4SLinus Torvalds return -ENOBUFS; 32491da177e4SLinus Torvalds 3250a3433f35SChangli Gao tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); 32517faee5c0SEric Dumazet tp->retrans_stamp = tcp_time_stamp; 3252783237e8SYuchung Cheng tcp_connect_queue_skb(sk, buff); 3253735d3831SFlorian Westphal tcp_ecn_send_syn(sk, buff); 32541da177e4SLinus Torvalds 3255783237e8SYuchung Cheng /* Send off SYN; include data in Fast Open. */ 3256783237e8SYuchung Cheng err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : 3257783237e8SYuchung Cheng tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); 3258ee586811SEric Paris if (err == -ECONNREFUSED) 3259ee586811SEric Paris return err; 3260bd37a088SWei Yongjun 3261bd37a088SWei Yongjun /* We change tp->snd_nxt after the tcp_transmit_skb() call 3262bd37a088SWei Yongjun * in order to make this packet get counted in tcpOutSegs. 3263bd37a088SWei Yongjun */ 3264bd37a088SWei Yongjun tp->snd_nxt = tp->write_seq; 3265bd37a088SWei Yongjun tp->pushed_seq = tp->write_seq; 326681cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); 32671da177e4SLinus Torvalds 32681da177e4SLinus Torvalds /* Timer for repeating the SYN until an answer. */ 32693f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 32703f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 32711da177e4SLinus Torvalds return 0; 32721da177e4SLinus Torvalds } 32734bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_connect); 32741da177e4SLinus Torvalds 32751da177e4SLinus Torvalds /* Send out a delayed ack, the caller does the policy checking 32761da177e4SLinus Torvalds * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 32771da177e4SLinus Torvalds * for details. 32781da177e4SLinus Torvalds */ 32791da177e4SLinus Torvalds void tcp_send_delayed_ack(struct sock *sk) 32801da177e4SLinus Torvalds { 3281463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 3282463c84b9SArnaldo Carvalho de Melo int ato = icsk->icsk_ack.ato; 32831da177e4SLinus Torvalds unsigned long timeout; 32841da177e4SLinus Torvalds 32859890092eSFlorian Westphal tcp_ca_event(sk, CA_EVENT_DELAYED_ACK); 32869890092eSFlorian Westphal 32871da177e4SLinus Torvalds if (ato > TCP_DELACK_MIN) { 3288463c84b9SArnaldo Carvalho de Melo const struct tcp_sock *tp = tcp_sk(sk); 32891da177e4SLinus Torvalds int max_ato = HZ / 2; 32901da177e4SLinus Torvalds 3291056834d9SIlpo Järvinen if (icsk->icsk_ack.pingpong || 3292056834d9SIlpo Järvinen (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 32931da177e4SLinus Torvalds max_ato = TCP_DELACK_MAX; 32941da177e4SLinus Torvalds 32951da177e4SLinus Torvalds /* Slow path, intersegment interval is "high". */ 32961da177e4SLinus Torvalds 32971da177e4SLinus Torvalds /* If some rtt estimate is known, use it to bound delayed ack. 3298463c84b9SArnaldo Carvalho de Melo * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 32991da177e4SLinus Torvalds * directly. 33001da177e4SLinus Torvalds */ 3301740b0f18SEric Dumazet if (tp->srtt_us) { 3302740b0f18SEric Dumazet int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3), 3303740b0f18SEric Dumazet TCP_DELACK_MIN); 33041da177e4SLinus Torvalds 33051da177e4SLinus Torvalds if (rtt < max_ato) 33061da177e4SLinus Torvalds max_ato = rtt; 33071da177e4SLinus Torvalds } 33081da177e4SLinus Torvalds 33091da177e4SLinus Torvalds ato = min(ato, max_ato); 33101da177e4SLinus Torvalds } 33111da177e4SLinus Torvalds 33121da177e4SLinus Torvalds /* Stay within the limit we were given */ 33131da177e4SLinus Torvalds timeout = jiffies + ato; 33141da177e4SLinus Torvalds 33151da177e4SLinus Torvalds /* Use new timeout only if there wasn't a older one earlier. */ 3316463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 33171da177e4SLinus Torvalds /* If delack timer was blocked or is about to expire, 33181da177e4SLinus Torvalds * send ACK now. 33191da177e4SLinus Torvalds */ 3320463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.blocked || 3321463c84b9SArnaldo Carvalho de Melo time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 33221da177e4SLinus Torvalds tcp_send_ack(sk); 33231da177e4SLinus Torvalds return; 33241da177e4SLinus Torvalds } 33251da177e4SLinus Torvalds 3326463c84b9SArnaldo Carvalho de Melo if (!time_before(timeout, icsk->icsk_ack.timeout)) 3327463c84b9SArnaldo Carvalho de Melo timeout = icsk->icsk_ack.timeout; 33281da177e4SLinus Torvalds } 3329463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 3330463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.timeout = timeout; 3331463c84b9SArnaldo Carvalho de Melo sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 33321da177e4SLinus Torvalds } 33331da177e4SLinus Torvalds 33341da177e4SLinus Torvalds /* This routine sends an ack and also updates the window. */ 33351da177e4SLinus Torvalds void tcp_send_ack(struct sock *sk) 33361da177e4SLinus Torvalds { 33371da177e4SLinus Torvalds struct sk_buff *buff; 33381da177e4SLinus Torvalds 3339058dc334SIlpo Järvinen /* If we have been reset, we may not send again. */ 3340058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3341058dc334SIlpo Järvinen return; 3342058dc334SIlpo Järvinen 33439890092eSFlorian Westphal tcp_ca_event(sk, CA_EVENT_NON_DELAYED_ACK); 33449890092eSFlorian Westphal 33451da177e4SLinus Torvalds /* We are not putting this on the write queue, so 33461da177e4SLinus Torvalds * tcp_transmit_skb() will set the ownership to this 33471da177e4SLinus Torvalds * sock. 33481da177e4SLinus Torvalds */ 334999a1dec7SMel Gorman buff = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC)); 335051456b29SIan Morris if (!buff) { 3351463c84b9SArnaldo Carvalho de Melo inet_csk_schedule_ack(sk); 3352463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; 33533f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 33543f421baaSArnaldo Carvalho de Melo TCP_DELACK_MAX, TCP_RTO_MAX); 33551da177e4SLinus Torvalds return; 33561da177e4SLinus Torvalds } 33571da177e4SLinus Torvalds 33581da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 33591da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 3360a3433f35SChangli Gao tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); 33611da177e4SLinus Torvalds 336298781965SEric Dumazet /* We do not want pure acks influencing TCP Small Queues or fq/pacing 336398781965SEric Dumazet * too much. 336498781965SEric Dumazet * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784 336598781965SEric Dumazet * We also avoid tcp_wfree() overhead (cache line miss accessing 336698781965SEric Dumazet * tp->tsq_flags) by using regular sock_wfree() 336798781965SEric Dumazet */ 336898781965SEric Dumazet skb_set_tcp_pure_ack(buff); 336998781965SEric Dumazet 33701da177e4SLinus Torvalds /* Send it off, this clears delayed acks for us. */ 33717faee5c0SEric Dumazet skb_mstamp_get(&buff->skb_mstamp); 337299a1dec7SMel Gorman tcp_transmit_skb(sk, buff, 0, sk_gfp_atomic(sk, GFP_ATOMIC)); 33731da177e4SLinus Torvalds } 3374e3118e83SDaniel Borkmann EXPORT_SYMBOL_GPL(tcp_send_ack); 33751da177e4SLinus Torvalds 33761da177e4SLinus Torvalds /* This routine sends a packet with an out of date sequence 33771da177e4SLinus Torvalds * number. It assumes the other end will try to ack it. 33781da177e4SLinus Torvalds * 33791da177e4SLinus Torvalds * Question: what should we make while urgent mode? 33801da177e4SLinus Torvalds * 4.4BSD forces sending single byte of data. We cannot send 33811da177e4SLinus Torvalds * out of window data, because we have SND.NXT==SND.MAX... 33821da177e4SLinus Torvalds * 33831da177e4SLinus Torvalds * Current solution: to send TWO zero-length segments in urgent mode: 33841da177e4SLinus Torvalds * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 33851da177e4SLinus Torvalds * out-of-date with SND.UNA-1 to probe window. 33861da177e4SLinus Torvalds */ 3387e520af48SEric Dumazet static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib) 33881da177e4SLinus Torvalds { 33891da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 33901da177e4SLinus Torvalds struct sk_buff *skb; 33911da177e4SLinus Torvalds 33921da177e4SLinus Torvalds /* We don't queue it, tcp_transmit_skb() sets ownership. */ 339399a1dec7SMel Gorman skb = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC)); 339451456b29SIan Morris if (!skb) 33951da177e4SLinus Torvalds return -1; 33961da177e4SLinus Torvalds 33971da177e4SLinus Torvalds /* Reserve space for headers and set control bits. */ 33981da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 33991da177e4SLinus Torvalds /* Use a previous sequence. This should cause the other 34001da177e4SLinus Torvalds * end to send an ack. Don't queue or clone SKB, just 34011da177e4SLinus Torvalds * send it. 34021da177e4SLinus Torvalds */ 3403a3433f35SChangli Gao tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); 34047faee5c0SEric Dumazet skb_mstamp_get(&skb->skb_mstamp); 3405e520af48SEric Dumazet NET_INC_STATS_BH(sock_net(sk), mib); 3406dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC); 34071da177e4SLinus Torvalds } 34081da177e4SLinus Torvalds 3409ee995283SPavel Emelyanov void tcp_send_window_probe(struct sock *sk) 3410ee995283SPavel Emelyanov { 3411ee995283SPavel Emelyanov if (sk->sk_state == TCP_ESTABLISHED) { 3412ee995283SPavel Emelyanov tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; 3413e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE); 3414ee995283SPavel Emelyanov } 3415ee995283SPavel Emelyanov } 3416ee995283SPavel Emelyanov 341767edfef7SAndi Kleen /* Initiate keepalive or window probe from timer. */ 3418e520af48SEric Dumazet int tcp_write_wakeup(struct sock *sk, int mib) 34191da177e4SLinus Torvalds { 34201da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 34211da177e4SLinus Torvalds struct sk_buff *skb; 34221da177e4SLinus Torvalds 3423058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3424058dc334SIlpo Järvinen return -1; 3425058dc334SIlpo Järvinen 342600db4124SIan Morris skb = tcp_send_head(sk); 342700db4124SIan Morris if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { 34281da177e4SLinus Torvalds int err; 34290c54b85fSIlpo Järvinen unsigned int mss = tcp_current_mss(sk); 343090840defSIlpo Järvinen unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 34311da177e4SLinus Torvalds 34321da177e4SLinus Torvalds if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 34331da177e4SLinus Torvalds tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 34341da177e4SLinus Torvalds 34351da177e4SLinus Torvalds /* We are probing the opening of a window 34361da177e4SLinus Torvalds * but the window size is != 0 34371da177e4SLinus Torvalds * must have been a result SWS avoidance ( sender ) 34381da177e4SLinus Torvalds */ 34391da177e4SLinus Torvalds if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 34401da177e4SLinus Torvalds skb->len > mss) { 34411da177e4SLinus Torvalds seg_size = min(seg_size, mss); 34424de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 34436cc55e09SOctavian Purdila if (tcp_fragment(sk, skb, seg_size, mss, GFP_ATOMIC)) 34441da177e4SLinus Torvalds return -1; 34451da177e4SLinus Torvalds } else if (!tcp_skb_pcount(skb)) 34465bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss); 34471da177e4SLinus Torvalds 34484de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 3449dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 345066f5fe62SIlpo Järvinen if (!err) 345166f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 34521da177e4SLinus Torvalds return err; 34531da177e4SLinus Torvalds } else { 345433f5f57eSIlpo Järvinen if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) 3455e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 1, mib); 3456e520af48SEric Dumazet return tcp_xmit_probe_skb(sk, 0, mib); 34571da177e4SLinus Torvalds } 34581da177e4SLinus Torvalds } 34591da177e4SLinus Torvalds 34601da177e4SLinus Torvalds /* A window probe timeout has occurred. If window is not closed send 34611da177e4SLinus Torvalds * a partial packet else a zero probe. 34621da177e4SLinus Torvalds */ 34631da177e4SLinus Torvalds void tcp_send_probe0(struct sock *sk) 34641da177e4SLinus Torvalds { 3465463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 34661da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 3467fcdd1cf4SEric Dumazet unsigned long probe_max; 34681da177e4SLinus Torvalds int err; 34691da177e4SLinus Torvalds 3470e520af48SEric Dumazet err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE); 34711da177e4SLinus Torvalds 3472fe067e8aSDavid S. Miller if (tp->packets_out || !tcp_send_head(sk)) { 34731da177e4SLinus Torvalds /* Cancel probe timer, if it is not required. */ 34746687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 0; 3475463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff = 0; 34761da177e4SLinus Torvalds return; 34771da177e4SLinus Torvalds } 34781da177e4SLinus Torvalds 34791da177e4SLinus Torvalds if (err <= 0) { 3480463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_backoff < sysctl_tcp_retries2) 3481463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff++; 34826687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out++; 3483fcdd1cf4SEric Dumazet probe_max = TCP_RTO_MAX; 34841da177e4SLinus Torvalds } else { 34851da177e4SLinus Torvalds /* If packet was not sent due to local congestion, 34866687e988SArnaldo Carvalho de Melo * do not backoff and do not remember icsk_probes_out. 34871da177e4SLinus Torvalds * Let local senders to fight for local resources. 34881da177e4SLinus Torvalds * 34891da177e4SLinus Torvalds * Use accumulated backoff yet. 34901da177e4SLinus Torvalds */ 34916687e988SArnaldo Carvalho de Melo if (!icsk->icsk_probes_out) 34926687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 1; 3493fcdd1cf4SEric Dumazet probe_max = TCP_RESOURCE_PROBE_INTERVAL; 34941da177e4SLinus Torvalds } 3495fcdd1cf4SEric Dumazet inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 349621c8fe99SEric Dumazet tcp_probe0_when(sk, probe_max), 3497fcdd1cf4SEric Dumazet TCP_RTO_MAX); 34981da177e4SLinus Torvalds } 34995db92c99SOctavian Purdila 35005db92c99SOctavian Purdila int tcp_rtx_synack(struct sock *sk, struct request_sock *req) 35015db92c99SOctavian Purdila { 35025db92c99SOctavian Purdila const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific; 35035db92c99SOctavian Purdila struct flowi fl; 35045db92c99SOctavian Purdila int res; 35055db92c99SOctavian Purdila 35065db92c99SOctavian Purdila res = af_ops->send_synack(sk, NULL, &fl, req, 0, NULL); 35075db92c99SOctavian Purdila if (!res) { 35085db92c99SOctavian Purdila TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_RETRANSSEGS); 35095db92c99SOctavian Purdila NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 35105db92c99SOctavian Purdila } 35115db92c99SOctavian Purdila return res; 35125db92c99SOctavian Purdila } 35135db92c99SOctavian Purdila EXPORT_SYMBOL(tcp_rtx_synack); 3514