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". 1406f021c62SEric Dumazet * This is the first part of cwnd validation mechanism. 1416f021c62SEric Dumazet */ 1426f021c62SEric Dumazet void tcp_cwnd_restart(struct sock *sk, s32 delta) 1431da177e4SLinus Torvalds { 144463c84b9SArnaldo Carvalho de Melo struct tcp_sock *tp = tcp_sk(sk); 1456f021c62SEric Dumazet u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk)); 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 16705c5a46dSNeal Cardwell if (tcp_packets_in_flight(tp) == 0) 16805c5a46dSNeal Cardwell tcp_ca_event(sk, CA_EVENT_TX_START); 16905c5a46dSNeal Cardwell 1701da177e4SLinus Torvalds tp->lsndtime = now; 1711da177e4SLinus Torvalds 1721da177e4SLinus Torvalds /* If it is a reply for ato after last received 1731da177e4SLinus Torvalds * packet, enter pingpong mode. 1741da177e4SLinus Torvalds */ 1752251ae46SJon Maxwell if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato) 176463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pingpong = 1; 1771da177e4SLinus Torvalds } 1781da177e4SLinus Torvalds 17967edfef7SAndi Kleen /* Account for an ACK we sent. */ 18040efc6faSStephen Hemminger static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts) 1811da177e4SLinus Torvalds { 182463c84b9SArnaldo Carvalho de Melo tcp_dec_quickack_mode(sk, pkts); 183463c84b9SArnaldo Carvalho de Melo inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); 1841da177e4SLinus Torvalds } 1851da177e4SLinus Torvalds 18685f16525SYuchung Cheng 18785f16525SYuchung Cheng u32 tcp_default_init_rwnd(u32 mss) 18885f16525SYuchung Cheng { 18985f16525SYuchung Cheng /* Initial receive window should be twice of TCP_INIT_CWND to 1909ef71e0cSWeiping Pan * enable proper sending of new unsent data during fast recovery 19185f16525SYuchung Cheng * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a 19285f16525SYuchung Cheng * limit when mss is larger than 1460. 19385f16525SYuchung Cheng */ 19485f16525SYuchung Cheng u32 init_rwnd = TCP_INIT_CWND * 2; 19585f16525SYuchung Cheng 19685f16525SYuchung Cheng if (mss > 1460) 19785f16525SYuchung Cheng init_rwnd = max((1460 * init_rwnd) / mss, 2U); 19885f16525SYuchung Cheng return init_rwnd; 19985f16525SYuchung Cheng } 20085f16525SYuchung Cheng 2011da177e4SLinus Torvalds /* Determine a window scaling and initial window to offer. 2021da177e4SLinus Torvalds * Based on the assumption that the given amount of space 2031da177e4SLinus Torvalds * will be offered. Store the results in the tp structure. 2041da177e4SLinus Torvalds * NOTE: for smooth operation initial space offering should 2051da177e4SLinus Torvalds * be a multiple of mss if possible. We assume here that mss >= 1. 2061da177e4SLinus Torvalds * This MUST be enforced by all callers. 2071da177e4SLinus Torvalds */ 2081da177e4SLinus Torvalds void tcp_select_initial_window(int __space, __u32 mss, 2091da177e4SLinus Torvalds __u32 *rcv_wnd, __u32 *window_clamp, 21031d12926Slaurent chavey int wscale_ok, __u8 *rcv_wscale, 21131d12926Slaurent chavey __u32 init_rcv_wnd) 2121da177e4SLinus Torvalds { 2131da177e4SLinus Torvalds unsigned int space = (__space < 0 ? 0 : __space); 2141da177e4SLinus Torvalds 2151da177e4SLinus Torvalds /* If no clamp set the clamp to the max possible scaled window */ 2161da177e4SLinus Torvalds if (*window_clamp == 0) 2171da177e4SLinus Torvalds (*window_clamp) = (65535 << 14); 2181da177e4SLinus Torvalds space = min(*window_clamp, space); 2191da177e4SLinus Torvalds 2201da177e4SLinus Torvalds /* Quantize space offering to a multiple of mss if possible. */ 2211da177e4SLinus Torvalds if (space > mss) 2221da177e4SLinus Torvalds space = (space / mss) * mss; 2231da177e4SLinus Torvalds 2241da177e4SLinus Torvalds /* NOTE: offering an initial window larger than 32767 22515d99e02SRick Jones * will break some buggy TCP stacks. If the admin tells us 22615d99e02SRick Jones * it is likely we could be speaking with such a buggy stack 22715d99e02SRick Jones * we will truncate our initial window offering to 32K-1 22815d99e02SRick Jones * unless the remote has sent us a window scaling option, 22915d99e02SRick Jones * which we interpret as a sign the remote TCP is not 23015d99e02SRick Jones * misinterpreting the window field as a signed quantity. 2311da177e4SLinus Torvalds */ 23215d99e02SRick Jones if (sysctl_tcp_workaround_signed_windows) 2331da177e4SLinus Torvalds (*rcv_wnd) = min(space, MAX_TCP_WINDOW); 23415d99e02SRick Jones else 23515d99e02SRick Jones (*rcv_wnd) = space; 23615d99e02SRick Jones 2371da177e4SLinus Torvalds (*rcv_wscale) = 0; 2381da177e4SLinus Torvalds if (wscale_ok) { 2391da177e4SLinus Torvalds /* Set window scaling on max possible window 2401da177e4SLinus Torvalds * See RFC1323 for an explanation of the limit to 14 2411da177e4SLinus Torvalds */ 2421da177e4SLinus Torvalds space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max); 243316c1592SStephen Hemminger space = min_t(u32, space, *window_clamp); 2441da177e4SLinus Torvalds while (space > 65535 && (*rcv_wscale) < 14) { 2451da177e4SLinus Torvalds space >>= 1; 2461da177e4SLinus Torvalds (*rcv_wscale)++; 2471da177e4SLinus Torvalds } 2481da177e4SLinus Torvalds } 2491da177e4SLinus Torvalds 2501da177e4SLinus Torvalds if (mss > (1 << *rcv_wscale)) { 25185f16525SYuchung Cheng if (!init_rcv_wnd) /* Use default unless specified otherwise */ 25285f16525SYuchung Cheng init_rcv_wnd = tcp_default_init_rwnd(mss); 253b1afde60SNandita Dukkipati *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss); 2541da177e4SLinus Torvalds } 2551da177e4SLinus Torvalds 2561da177e4SLinus Torvalds /* Set the clamp no higher than max representable value */ 2571da177e4SLinus Torvalds (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp); 2581da177e4SLinus Torvalds } 2594bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_select_initial_window); 2601da177e4SLinus Torvalds 2611da177e4SLinus Torvalds /* Chose a new window to advertise, update state in tcp_sock for the 2621da177e4SLinus Torvalds * socket, and return result with RFC1323 scaling applied. The return 2631da177e4SLinus Torvalds * value can be stuffed directly into th->window for an outgoing 2641da177e4SLinus Torvalds * frame. 2651da177e4SLinus Torvalds */ 26640efc6faSStephen Hemminger static u16 tcp_select_window(struct sock *sk) 2671da177e4SLinus Torvalds { 2681da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2698e165e20SFlorian Westphal u32 old_win = tp->rcv_wnd; 2701da177e4SLinus Torvalds u32 cur_win = tcp_receive_window(tp); 2711da177e4SLinus Torvalds u32 new_win = __tcp_select_window(sk); 2721da177e4SLinus Torvalds 2731da177e4SLinus Torvalds /* Never shrink the offered window */ 2741da177e4SLinus Torvalds if (new_win < cur_win) { 2751da177e4SLinus Torvalds /* Danger Will Robinson! 2761da177e4SLinus Torvalds * Don't update rcv_wup/rcv_wnd here or else 2771da177e4SLinus Torvalds * we will not be able to advertise a zero 2781da177e4SLinus Torvalds * window in time. --DaveM 2791da177e4SLinus Torvalds * 2801da177e4SLinus Torvalds * Relax Will Robinson. 2811da177e4SLinus Torvalds */ 2828e165e20SFlorian Westphal if (new_win == 0) 2838e165e20SFlorian Westphal NET_INC_STATS(sock_net(sk), 2848e165e20SFlorian Westphal LINUX_MIB_TCPWANTZEROWINDOWADV); 285607bfbf2SPatrick McHardy new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale); 2861da177e4SLinus Torvalds } 2871da177e4SLinus Torvalds tp->rcv_wnd = new_win; 2881da177e4SLinus Torvalds tp->rcv_wup = tp->rcv_nxt; 2891da177e4SLinus Torvalds 2901da177e4SLinus Torvalds /* Make sure we do not exceed the maximum possible 2911da177e4SLinus Torvalds * scaled window. 2921da177e4SLinus Torvalds */ 29315d99e02SRick Jones if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows) 2941da177e4SLinus Torvalds new_win = min(new_win, MAX_TCP_WINDOW); 2951da177e4SLinus Torvalds else 2961da177e4SLinus Torvalds new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); 2971da177e4SLinus Torvalds 2981da177e4SLinus Torvalds /* RFC1323 scaling applied */ 2991da177e4SLinus Torvalds new_win >>= tp->rx_opt.rcv_wscale; 3001da177e4SLinus Torvalds 3011da177e4SLinus Torvalds /* If we advertise zero window, disable fast path. */ 3028e165e20SFlorian Westphal if (new_win == 0) { 3031da177e4SLinus Torvalds tp->pred_flags = 0; 3048e165e20SFlorian Westphal if (old_win) 3058e165e20SFlorian Westphal NET_INC_STATS(sock_net(sk), 3068e165e20SFlorian Westphal LINUX_MIB_TCPTOZEROWINDOWADV); 3078e165e20SFlorian Westphal } else if (old_win == 0) { 3088e165e20SFlorian Westphal NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV); 3098e165e20SFlorian Westphal } 3101da177e4SLinus Torvalds 3111da177e4SLinus Torvalds return new_win; 3121da177e4SLinus Torvalds } 3131da177e4SLinus Torvalds 31467edfef7SAndi Kleen /* Packet ECN state for a SYN-ACK */ 315735d3831SFlorian Westphal static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb) 316bdf1ee5dSIlpo Järvinen { 31730e502a3SDaniel Borkmann const struct tcp_sock *tp = tcp_sk(sk); 31830e502a3SDaniel Borkmann 3194de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR; 320bdf1ee5dSIlpo Järvinen if (!(tp->ecn_flags & TCP_ECN_OK)) 3214de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE; 32230e502a3SDaniel Borkmann else if (tcp_ca_needs_ecn(sk)) 32330e502a3SDaniel Borkmann INET_ECN_xmit(sk); 324bdf1ee5dSIlpo Järvinen } 325bdf1ee5dSIlpo Järvinen 32667edfef7SAndi Kleen /* Packet ECN state for a SYN. */ 327735d3831SFlorian Westphal static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb) 328bdf1ee5dSIlpo Järvinen { 329bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 330f7b3bec6SFlorian Westphal bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 || 331f7b3bec6SFlorian Westphal tcp_ca_needs_ecn(sk); 332f7b3bec6SFlorian Westphal 333f7b3bec6SFlorian Westphal if (!use_ecn) { 334f7b3bec6SFlorian Westphal const struct dst_entry *dst = __sk_dst_get(sk); 335f7b3bec6SFlorian Westphal 336f7b3bec6SFlorian Westphal if (dst && dst_feature(dst, RTAX_FEATURE_ECN)) 337f7b3bec6SFlorian Westphal use_ecn = true; 338f7b3bec6SFlorian Westphal } 339bdf1ee5dSIlpo Järvinen 340bdf1ee5dSIlpo Järvinen tp->ecn_flags = 0; 341f7b3bec6SFlorian Westphal 342f7b3bec6SFlorian Westphal if (use_ecn) { 3434de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR; 344bdf1ee5dSIlpo Järvinen tp->ecn_flags = TCP_ECN_OK; 34530e502a3SDaniel Borkmann if (tcp_ca_needs_ecn(sk)) 34630e502a3SDaniel Borkmann INET_ECN_xmit(sk); 347bdf1ee5dSIlpo Järvinen } 348bdf1ee5dSIlpo Järvinen } 349bdf1ee5dSIlpo Järvinen 35049213555SDaniel Borkmann static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb) 35149213555SDaniel Borkmann { 35249213555SDaniel Borkmann if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback) 35349213555SDaniel Borkmann /* tp->ecn_flags are cleared at a later point in time when 35449213555SDaniel Borkmann * SYN ACK is ultimatively being received. 35549213555SDaniel Borkmann */ 35649213555SDaniel Borkmann TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR); 35749213555SDaniel Borkmann } 35849213555SDaniel Borkmann 359735d3831SFlorian Westphal static void 3606ac705b1SEric Dumazet tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th) 361bdf1ee5dSIlpo Järvinen { 3626ac705b1SEric Dumazet if (inet_rsk(req)->ecn_ok) 363bdf1ee5dSIlpo Järvinen th->ece = 1; 364bdf1ee5dSIlpo Järvinen } 365bdf1ee5dSIlpo Järvinen 36667edfef7SAndi Kleen /* Set up ECN state for a packet on a ESTABLISHED socket that is about to 36767edfef7SAndi Kleen * be sent. 36867edfef7SAndi Kleen */ 369735d3831SFlorian Westphal static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb, 370bdf1ee5dSIlpo Järvinen int tcp_header_len) 371bdf1ee5dSIlpo Järvinen { 372bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 373bdf1ee5dSIlpo Järvinen 374bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_OK) { 375bdf1ee5dSIlpo Järvinen /* Not-retransmitted data segment: set ECT and inject CWR. */ 376bdf1ee5dSIlpo Järvinen if (skb->len != tcp_header_len && 377bdf1ee5dSIlpo Järvinen !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { 378bdf1ee5dSIlpo Järvinen INET_ECN_xmit(sk); 379bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) { 380bdf1ee5dSIlpo Järvinen tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; 381bdf1ee5dSIlpo Järvinen tcp_hdr(skb)->cwr = 1; 382bdf1ee5dSIlpo Järvinen skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; 383bdf1ee5dSIlpo Järvinen } 38430e502a3SDaniel Borkmann } else if (!tcp_ca_needs_ecn(sk)) { 385bdf1ee5dSIlpo Järvinen /* ACK or retransmitted segment: clear ECT|CE */ 386bdf1ee5dSIlpo Järvinen INET_ECN_dontxmit(sk); 387bdf1ee5dSIlpo Järvinen } 388bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) 389bdf1ee5dSIlpo Järvinen tcp_hdr(skb)->ece = 1; 390bdf1ee5dSIlpo Järvinen } 391bdf1ee5dSIlpo Järvinen } 392bdf1ee5dSIlpo Järvinen 393e870a8efSIlpo Järvinen /* Constructs common control bits of non-data skb. If SYN/FIN is present, 394e870a8efSIlpo Järvinen * auto increment end seqno. 395e870a8efSIlpo Järvinen */ 396e870a8efSIlpo Järvinen static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags) 397e870a8efSIlpo Järvinen { 3982e8e18efSDavid S. Miller skb->ip_summed = CHECKSUM_PARTIAL; 399e870a8efSIlpo Järvinen skb->csum = 0; 400e870a8efSIlpo Järvinen 4014de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags; 402e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->sacked = 0; 403e870a8efSIlpo Järvinen 404cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, 1); 405e870a8efSIlpo Järvinen 406e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->seq = seq; 407a3433f35SChangli Gao if (flags & (TCPHDR_SYN | TCPHDR_FIN)) 408e870a8efSIlpo Järvinen seq++; 409e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->end_seq = seq; 410e870a8efSIlpo Järvinen } 411e870a8efSIlpo Järvinen 412a2a385d6SEric Dumazet static inline bool tcp_urg_mode(const struct tcp_sock *tp) 41333f5f57eSIlpo Järvinen { 41433f5f57eSIlpo Järvinen return tp->snd_una != tp->snd_up; 41533f5f57eSIlpo Järvinen } 41633f5f57eSIlpo Järvinen 41733ad798cSAdam Langley #define OPTION_SACK_ADVERTISE (1 << 0) 41833ad798cSAdam Langley #define OPTION_TS (1 << 1) 41933ad798cSAdam Langley #define OPTION_MD5 (1 << 2) 42089e95a61SOri Finkelman #define OPTION_WSCALE (1 << 3) 4212100c8d2SYuchung Cheng #define OPTION_FAST_OPEN_COOKIE (1 << 8) 42233ad798cSAdam Langley 42333ad798cSAdam Langley struct tcp_out_options { 4242100c8d2SYuchung Cheng u16 options; /* bit field of OPTION_* */ 4252100c8d2SYuchung Cheng u16 mss; /* 0 to disable */ 42633ad798cSAdam Langley u8 ws; /* window scale, 0 to disable */ 42733ad798cSAdam Langley u8 num_sack_blocks; /* number of SACK blocks to include */ 428bd0388aeSWilliam Allen Simpson u8 hash_size; /* bytes in hash_location */ 429bd0388aeSWilliam Allen Simpson __u8 *hash_location; /* temporary pointer, overloaded */ 4302100c8d2SYuchung Cheng __u32 tsval, tsecr; /* need to include OPTION_TS */ 4312100c8d2SYuchung Cheng struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */ 43233ad798cSAdam Langley }; 43333ad798cSAdam Langley 43467edfef7SAndi Kleen /* Write previously computed TCP options to the packet. 43567edfef7SAndi Kleen * 43667edfef7SAndi Kleen * Beware: Something in the Internet is very sensitive to the ordering of 437fd6149d3SIlpo Järvinen * TCP options, we learned this through the hard way, so be careful here. 438fd6149d3SIlpo Järvinen * Luckily we can at least blame others for their non-compliance but from 4398e3bff96Sstephen hemminger * inter-operability perspective it seems that we're somewhat stuck with 440fd6149d3SIlpo Järvinen * the ordering which we have been using if we want to keep working with 441fd6149d3SIlpo Järvinen * those broken things (not that it currently hurts anybody as there isn't 442fd6149d3SIlpo Järvinen * particular reason why the ordering would need to be changed). 443fd6149d3SIlpo Järvinen * 444fd6149d3SIlpo Järvinen * At least SACK_PERM as the first option is known to lead to a disaster 445fd6149d3SIlpo Järvinen * (but it may well be that other scenarios fail similarly). 446fd6149d3SIlpo Järvinen */ 44733ad798cSAdam Langley static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp, 448bd0388aeSWilliam Allen Simpson struct tcp_out_options *opts) 449bd0388aeSWilliam Allen Simpson { 4502100c8d2SYuchung Cheng u16 options = opts->options; /* mungable copy */ 451bd0388aeSWilliam Allen Simpson 452bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_MD5 & options)) { 4531a2c6181SChristoph Paasch *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | 4541a2c6181SChristoph Paasch (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG); 455bd0388aeSWilliam Allen Simpson /* overload cookie hash location */ 456bd0388aeSWilliam Allen Simpson opts->hash_location = (__u8 *)ptr; 45733ad798cSAdam Langley ptr += 4; 45833ad798cSAdam Langley } 45933ad798cSAdam Langley 460fd6149d3SIlpo Järvinen if (unlikely(opts->mss)) { 461fd6149d3SIlpo Järvinen *ptr++ = htonl((TCPOPT_MSS << 24) | 462fd6149d3SIlpo Järvinen (TCPOLEN_MSS << 16) | 463fd6149d3SIlpo Järvinen opts->mss); 464fd6149d3SIlpo Järvinen } 465fd6149d3SIlpo Järvinen 466bd0388aeSWilliam Allen Simpson if (likely(OPTION_TS & options)) { 467bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 46833ad798cSAdam Langley *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | 46933ad798cSAdam Langley (TCPOLEN_SACK_PERM << 16) | 47033ad798cSAdam Langley (TCPOPT_TIMESTAMP << 8) | 47133ad798cSAdam Langley TCPOLEN_TIMESTAMP); 472bd0388aeSWilliam Allen Simpson options &= ~OPTION_SACK_ADVERTISE; 47333ad798cSAdam Langley } else { 474496c98dfSYOSHIFUJI Hideaki *ptr++ = htonl((TCPOPT_NOP << 24) | 47540efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 47640efc6faSStephen Hemminger (TCPOPT_TIMESTAMP << 8) | 47740efc6faSStephen Hemminger TCPOLEN_TIMESTAMP); 47840efc6faSStephen Hemminger } 47933ad798cSAdam Langley *ptr++ = htonl(opts->tsval); 48033ad798cSAdam Langley *ptr++ = htonl(opts->tsecr); 48133ad798cSAdam Langley } 48233ad798cSAdam Langley 483bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 48433ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 48533ad798cSAdam Langley (TCPOPT_NOP << 16) | 48633ad798cSAdam Langley (TCPOPT_SACK_PERM << 8) | 48733ad798cSAdam Langley TCPOLEN_SACK_PERM); 48833ad798cSAdam Langley } 48933ad798cSAdam Langley 490bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_WSCALE & options)) { 49133ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 49233ad798cSAdam Langley (TCPOPT_WINDOW << 16) | 49333ad798cSAdam Langley (TCPOLEN_WINDOW << 8) | 49433ad798cSAdam Langley opts->ws); 49533ad798cSAdam Langley } 49633ad798cSAdam Langley 49733ad798cSAdam Langley if (unlikely(opts->num_sack_blocks)) { 49833ad798cSAdam Langley struct tcp_sack_block *sp = tp->rx_opt.dsack ? 49933ad798cSAdam Langley tp->duplicate_sack : tp->selective_acks; 50040efc6faSStephen Hemminger int this_sack; 50140efc6faSStephen Hemminger 50240efc6faSStephen Hemminger *ptr++ = htonl((TCPOPT_NOP << 24) | 50340efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 50440efc6faSStephen Hemminger (TCPOPT_SACK << 8) | 50533ad798cSAdam Langley (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * 50640efc6faSStephen Hemminger TCPOLEN_SACK_PERBLOCK))); 5072de979bdSStephen Hemminger 50833ad798cSAdam Langley for (this_sack = 0; this_sack < opts->num_sack_blocks; 50933ad798cSAdam Langley ++this_sack) { 51040efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].start_seq); 51140efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].end_seq); 51240efc6faSStephen Hemminger } 5132de979bdSStephen Hemminger 51440efc6faSStephen Hemminger tp->rx_opt.dsack = 0; 51540efc6faSStephen Hemminger } 5162100c8d2SYuchung Cheng 5172100c8d2SYuchung Cheng if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) { 5182100c8d2SYuchung Cheng struct tcp_fastopen_cookie *foc = opts->fastopen_cookie; 5197f9b838bSDaniel Lee u8 *p = (u8 *)ptr; 5207f9b838bSDaniel Lee u32 len; /* Fast Open option length */ 5212100c8d2SYuchung Cheng 5227f9b838bSDaniel Lee if (foc->exp) { 5237f9b838bSDaniel Lee len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len; 5247f9b838bSDaniel Lee *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) | 5252100c8d2SYuchung Cheng TCPOPT_FASTOPEN_MAGIC); 5267f9b838bSDaniel Lee p += TCPOLEN_EXP_FASTOPEN_BASE; 5277f9b838bSDaniel Lee } else { 5287f9b838bSDaniel Lee len = TCPOLEN_FASTOPEN_BASE + foc->len; 5297f9b838bSDaniel Lee *p++ = TCPOPT_FASTOPEN; 5307f9b838bSDaniel Lee *p++ = len; 5312100c8d2SYuchung Cheng } 5327f9b838bSDaniel Lee 5337f9b838bSDaniel Lee memcpy(p, foc->val, foc->len); 5347f9b838bSDaniel Lee if ((len & 3) == 2) { 5357f9b838bSDaniel Lee p[foc->len] = TCPOPT_NOP; 5367f9b838bSDaniel Lee p[foc->len + 1] = TCPOPT_NOP; 5377f9b838bSDaniel Lee } 5387f9b838bSDaniel Lee ptr += (len + 3) >> 2; 5392100c8d2SYuchung Cheng } 54040efc6faSStephen Hemminger } 54140efc6faSStephen Hemminger 54267edfef7SAndi Kleen /* Compute TCP options for SYN packets. This is not the final 54367edfef7SAndi Kleen * network wire format yet. 54467edfef7SAndi Kleen */ 54595c96174SEric Dumazet static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb, 54633ad798cSAdam Langley struct tcp_out_options *opts, 547cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 548cf533ea5SEric Dumazet { 54933ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 55095c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 551783237e8SYuchung Cheng struct tcp_fastopen_request *fastopen = tp->fastopen_req; 55233ad798cSAdam Langley 553cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 55433ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 55533ad798cSAdam Langley if (*md5) { 55633ad798cSAdam Langley opts->options |= OPTION_MD5; 557bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 558cfb6eeb4SYOSHIFUJI Hideaki } 55933ad798cSAdam Langley #else 56033ad798cSAdam Langley *md5 = NULL; 561cfb6eeb4SYOSHIFUJI Hideaki #endif 56233ad798cSAdam Langley 56333ad798cSAdam Langley /* We always get an MSS option. The option bytes which will be seen in 56433ad798cSAdam Langley * normal data packets should timestamps be used, must be in the MSS 56533ad798cSAdam Langley * advertised. But we subtract them from tp->mss_cache so that 56633ad798cSAdam Langley * calculations in tcp_sendmsg are simpler etc. So account for this 56733ad798cSAdam Langley * fact here if necessary. If we don't do this correctly, as a 56833ad798cSAdam Langley * receiver we won't recognize data packets as being full sized when we 56933ad798cSAdam Langley * should, and thus we won't abide by the delayed ACK rules correctly. 57033ad798cSAdam Langley * SACKs don't matter, we never delay an ACK when we have any of those 57133ad798cSAdam Langley * going out. */ 57233ad798cSAdam Langley opts->mss = tcp_advertise_mss(sk); 573bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 57433ad798cSAdam Langley 57551456b29SIan Morris if (likely(sysctl_tcp_timestamps && !*md5)) { 57633ad798cSAdam Langley opts->options |= OPTION_TS; 5777faee5c0SEric Dumazet opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset; 57833ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 579bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 58033ad798cSAdam Langley } 581bb5b7c11SDavid S. Miller if (likely(sysctl_tcp_window_scaling)) { 58233ad798cSAdam Langley opts->ws = tp->rx_opt.rcv_wscale; 58389e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 584bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 58533ad798cSAdam Langley } 586bb5b7c11SDavid S. Miller if (likely(sysctl_tcp_sack)) { 58733ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 588b32d1310SDavid S. Miller if (unlikely(!(OPTION_TS & opts->options))) 589bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 59033ad798cSAdam Langley } 59133ad798cSAdam Langley 592783237e8SYuchung Cheng if (fastopen && fastopen->cookie.len >= 0) { 5932646c831SDaniel Lee u32 need = fastopen->cookie.len; 5942646c831SDaniel Lee 5952646c831SDaniel Lee need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE : 5962646c831SDaniel Lee TCPOLEN_FASTOPEN_BASE; 597783237e8SYuchung Cheng need = (need + 3) & ~3U; /* Align to 32 bits */ 598783237e8SYuchung Cheng if (remaining >= need) { 599783237e8SYuchung Cheng opts->options |= OPTION_FAST_OPEN_COOKIE; 600783237e8SYuchung Cheng opts->fastopen_cookie = &fastopen->cookie; 601783237e8SYuchung Cheng remaining -= need; 602783237e8SYuchung Cheng tp->syn_fastopen = 1; 6032646c831SDaniel Lee tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0; 604783237e8SYuchung Cheng } 605783237e8SYuchung Cheng } 606bd0388aeSWilliam Allen Simpson 607bd0388aeSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 60833ad798cSAdam Langley } 60933ad798cSAdam Langley 61067edfef7SAndi Kleen /* Set up TCP options for SYN-ACKs. */ 61137bfbddaSEric Dumazet static unsigned int tcp_synack_options(struct request_sock *req, 61295c96174SEric Dumazet unsigned int mss, struct sk_buff *skb, 61333ad798cSAdam Langley struct tcp_out_options *opts, 61480f03e27SEric Dumazet const struct tcp_md5sig_key *md5, 6158336886fSJerry Chu struct tcp_fastopen_cookie *foc) 6164957faadSWilliam Allen Simpson { 61733ad798cSAdam Langley struct inet_request_sock *ireq = inet_rsk(req); 61895c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 61933ad798cSAdam Langley 62033ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 62180f03e27SEric Dumazet if (md5) { 62233ad798cSAdam Langley opts->options |= OPTION_MD5; 6234957faadSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 6244957faadSWilliam Allen Simpson 6254957faadSWilliam Allen Simpson /* We can't fit any SACK blocks in a packet with MD5 + TS 6264957faadSWilliam Allen Simpson * options. There was discussion about disabling SACK 6274957faadSWilliam Allen Simpson * rather than TS in order to fit in better with old, 6284957faadSWilliam Allen Simpson * buggy kernels, but that was deemed to be unnecessary. 6294957faadSWilliam Allen Simpson */ 630de213e5eSEric Dumazet ireq->tstamp_ok &= !ireq->sack_ok; 63133ad798cSAdam Langley } 63233ad798cSAdam Langley #endif 63333ad798cSAdam Langley 6344957faadSWilliam Allen Simpson /* We always send an MSS option. */ 63533ad798cSAdam Langley opts->mss = mss; 6364957faadSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 63733ad798cSAdam Langley 63833ad798cSAdam Langley if (likely(ireq->wscale_ok)) { 63933ad798cSAdam Langley opts->ws = ireq->rcv_wscale; 64089e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 6414957faadSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 64233ad798cSAdam Langley } 643de213e5eSEric Dumazet if (likely(ireq->tstamp_ok)) { 64433ad798cSAdam Langley opts->options |= OPTION_TS; 6457faee5c0SEric Dumazet opts->tsval = tcp_skb_timestamp(skb); 64633ad798cSAdam Langley opts->tsecr = req->ts_recent; 6474957faadSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 64833ad798cSAdam Langley } 64933ad798cSAdam Langley if (likely(ireq->sack_ok)) { 65033ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 651de213e5eSEric Dumazet if (unlikely(!ireq->tstamp_ok)) 6524957faadSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 65333ad798cSAdam Langley } 6547f9b838bSDaniel Lee if (foc != NULL && foc->len >= 0) { 6557f9b838bSDaniel Lee u32 need = foc->len; 6567f9b838bSDaniel Lee 6577f9b838bSDaniel Lee need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE : 6587f9b838bSDaniel Lee TCPOLEN_FASTOPEN_BASE; 6598336886fSJerry Chu need = (need + 3) & ~3U; /* Align to 32 bits */ 6608336886fSJerry Chu if (remaining >= need) { 6618336886fSJerry Chu opts->options |= OPTION_FAST_OPEN_COOKIE; 6628336886fSJerry Chu opts->fastopen_cookie = foc; 6638336886fSJerry Chu remaining -= need; 6648336886fSJerry Chu } 6658336886fSJerry Chu } 6664957faadSWilliam Allen Simpson 6674957faadSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 66833ad798cSAdam Langley } 66933ad798cSAdam Langley 67067edfef7SAndi Kleen /* Compute TCP options for ESTABLISHED sockets. This is not the 67167edfef7SAndi Kleen * final wire format yet. 67267edfef7SAndi Kleen */ 67395c96174SEric Dumazet static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb, 67433ad798cSAdam Langley struct tcp_out_options *opts, 675cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 676cf533ea5SEric Dumazet { 67733ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 67895c96174SEric Dumazet unsigned int size = 0; 679cabeccbdSIlpo Järvinen unsigned int eff_sacks; 68033ad798cSAdam Langley 6815843ef42SAndi Kleen opts->options = 0; 6825843ef42SAndi Kleen 68333ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 68433ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 68533ad798cSAdam Langley if (unlikely(*md5)) { 68633ad798cSAdam Langley opts->options |= OPTION_MD5; 68733ad798cSAdam Langley size += TCPOLEN_MD5SIG_ALIGNED; 68833ad798cSAdam Langley } 68933ad798cSAdam Langley #else 69033ad798cSAdam Langley *md5 = NULL; 69133ad798cSAdam Langley #endif 69233ad798cSAdam Langley 69333ad798cSAdam Langley if (likely(tp->rx_opt.tstamp_ok)) { 69433ad798cSAdam Langley opts->options |= OPTION_TS; 6957faee5c0SEric Dumazet opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0; 69633ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 69733ad798cSAdam Langley size += TCPOLEN_TSTAMP_ALIGNED; 69833ad798cSAdam Langley } 69933ad798cSAdam Langley 700cabeccbdSIlpo Järvinen eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; 701cabeccbdSIlpo Järvinen if (unlikely(eff_sacks)) { 70295c96174SEric Dumazet const unsigned int remaining = MAX_TCP_OPTION_SPACE - size; 70333ad798cSAdam Langley opts->num_sack_blocks = 70495c96174SEric Dumazet min_t(unsigned int, eff_sacks, 70533ad798cSAdam Langley (remaining - TCPOLEN_SACK_BASE_ALIGNED) / 70633ad798cSAdam Langley TCPOLEN_SACK_PERBLOCK); 70733ad798cSAdam Langley size += TCPOLEN_SACK_BASE_ALIGNED + 70833ad798cSAdam Langley opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; 70933ad798cSAdam Langley } 71033ad798cSAdam Langley 71133ad798cSAdam Langley return size; 71240efc6faSStephen Hemminger } 7131da177e4SLinus Torvalds 71446d3ceabSEric Dumazet 71546d3ceabSEric Dumazet /* TCP SMALL QUEUES (TSQ) 71646d3ceabSEric Dumazet * 71746d3ceabSEric Dumazet * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev) 71846d3ceabSEric Dumazet * to reduce RTT and bufferbloat. 71946d3ceabSEric Dumazet * We do this using a special skb destructor (tcp_wfree). 72046d3ceabSEric Dumazet * 72146d3ceabSEric Dumazet * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb 72246d3ceabSEric Dumazet * needs to be reallocated in a driver. 7238e3bff96Sstephen hemminger * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc 72446d3ceabSEric Dumazet * 72546d3ceabSEric Dumazet * Since transmit from skb destructor is forbidden, we use a tasklet 72646d3ceabSEric Dumazet * to process all sockets that eventually need to send more skbs. 72746d3ceabSEric Dumazet * We use one tasklet per cpu, with its own queue of sockets. 72846d3ceabSEric Dumazet */ 72946d3ceabSEric Dumazet struct tsq_tasklet { 73046d3ceabSEric Dumazet struct tasklet_struct tasklet; 73146d3ceabSEric Dumazet struct list_head head; /* queue of tcp sockets */ 73246d3ceabSEric Dumazet }; 73346d3ceabSEric Dumazet static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet); 73446d3ceabSEric Dumazet 7356f458dfbSEric Dumazet static void tcp_tsq_handler(struct sock *sk) 7366f458dfbSEric Dumazet { 7376f458dfbSEric Dumazet if ((1 << sk->sk_state) & 7386f458dfbSEric Dumazet (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING | 7396f458dfbSEric Dumazet TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) 740bf06200eSJohn Ogness tcp_write_xmit(sk, tcp_current_mss(sk), tcp_sk(sk)->nonagle, 741bf06200eSJohn Ogness 0, GFP_ATOMIC); 7426f458dfbSEric Dumazet } 74346d3ceabSEric Dumazet /* 7448e3bff96Sstephen hemminger * One tasklet per cpu tries to send more skbs. 74546d3ceabSEric Dumazet * We run in tasklet context but need to disable irqs when 7468e3bff96Sstephen hemminger * transferring tsq->head because tcp_wfree() might 74746d3ceabSEric Dumazet * interrupt us (non NAPI drivers) 74846d3ceabSEric Dumazet */ 74946d3ceabSEric Dumazet static void tcp_tasklet_func(unsigned long data) 75046d3ceabSEric Dumazet { 75146d3ceabSEric Dumazet struct tsq_tasklet *tsq = (struct tsq_tasklet *)data; 75246d3ceabSEric Dumazet LIST_HEAD(list); 75346d3ceabSEric Dumazet unsigned long flags; 75446d3ceabSEric Dumazet struct list_head *q, *n; 75546d3ceabSEric Dumazet struct tcp_sock *tp; 75646d3ceabSEric Dumazet struct sock *sk; 75746d3ceabSEric Dumazet 75846d3ceabSEric Dumazet local_irq_save(flags); 75946d3ceabSEric Dumazet list_splice_init(&tsq->head, &list); 76046d3ceabSEric Dumazet local_irq_restore(flags); 76146d3ceabSEric Dumazet 76246d3ceabSEric Dumazet list_for_each_safe(q, n, &list) { 76346d3ceabSEric Dumazet tp = list_entry(q, struct tcp_sock, tsq_node); 76446d3ceabSEric Dumazet list_del(&tp->tsq_node); 76546d3ceabSEric Dumazet 76646d3ceabSEric Dumazet sk = (struct sock *)tp; 76746d3ceabSEric Dumazet bh_lock_sock(sk); 76846d3ceabSEric Dumazet 76946d3ceabSEric Dumazet if (!sock_owned_by_user(sk)) { 7706f458dfbSEric Dumazet tcp_tsq_handler(sk); 77146d3ceabSEric Dumazet } else { 77246d3ceabSEric Dumazet /* defer the work to tcp_release_cb() */ 7736f458dfbSEric Dumazet set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags); 77446d3ceabSEric Dumazet } 77546d3ceabSEric Dumazet bh_unlock_sock(sk); 77646d3ceabSEric Dumazet 77746d3ceabSEric Dumazet clear_bit(TSQ_QUEUED, &tp->tsq_flags); 77846d3ceabSEric Dumazet sk_free(sk); 77946d3ceabSEric Dumazet } 78046d3ceabSEric Dumazet } 78146d3ceabSEric Dumazet 7826f458dfbSEric Dumazet #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \ 7836f458dfbSEric Dumazet (1UL << TCP_WRITE_TIMER_DEFERRED) | \ 784563d34d0SEric Dumazet (1UL << TCP_DELACK_TIMER_DEFERRED) | \ 785563d34d0SEric Dumazet (1UL << TCP_MTU_REDUCED_DEFERRED)) 78646d3ceabSEric Dumazet /** 78746d3ceabSEric Dumazet * tcp_release_cb - tcp release_sock() callback 78846d3ceabSEric Dumazet * @sk: socket 78946d3ceabSEric Dumazet * 79046d3ceabSEric Dumazet * called from release_sock() to perform protocol dependent 79146d3ceabSEric Dumazet * actions before socket release. 79246d3ceabSEric Dumazet */ 79346d3ceabSEric Dumazet void tcp_release_cb(struct sock *sk) 79446d3ceabSEric Dumazet { 79546d3ceabSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 7966f458dfbSEric Dumazet unsigned long flags, nflags; 79746d3ceabSEric Dumazet 7986f458dfbSEric Dumazet /* perform an atomic operation only if at least one flag is set */ 7996f458dfbSEric Dumazet do { 8006f458dfbSEric Dumazet flags = tp->tsq_flags; 8016f458dfbSEric Dumazet if (!(flags & TCP_DEFERRED_ALL)) 8026f458dfbSEric Dumazet return; 8036f458dfbSEric Dumazet nflags = flags & ~TCP_DEFERRED_ALL; 8046f458dfbSEric Dumazet } while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags); 8056f458dfbSEric Dumazet 8066f458dfbSEric Dumazet if (flags & (1UL << TCP_TSQ_DEFERRED)) 8076f458dfbSEric Dumazet tcp_tsq_handler(sk); 8086f458dfbSEric Dumazet 809c3f9b018SEric Dumazet /* Here begins the tricky part : 810c3f9b018SEric Dumazet * We are called from release_sock() with : 811c3f9b018SEric Dumazet * 1) BH disabled 812c3f9b018SEric Dumazet * 2) sk_lock.slock spinlock held 813c3f9b018SEric Dumazet * 3) socket owned by us (sk->sk_lock.owned == 1) 814c3f9b018SEric Dumazet * 815c3f9b018SEric Dumazet * But following code is meant to be called from BH handlers, 816c3f9b018SEric Dumazet * so we should keep BH disabled, but early release socket ownership 817c3f9b018SEric Dumazet */ 818c3f9b018SEric Dumazet sock_release_ownership(sk); 819c3f9b018SEric Dumazet 820144d56e9SEric Dumazet if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) { 8216f458dfbSEric Dumazet tcp_write_timer_handler(sk); 822144d56e9SEric Dumazet __sock_put(sk); 823144d56e9SEric Dumazet } 824144d56e9SEric Dumazet if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) { 8256f458dfbSEric Dumazet tcp_delack_timer_handler(sk); 826144d56e9SEric Dumazet __sock_put(sk); 827144d56e9SEric Dumazet } 828144d56e9SEric Dumazet if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) { 8294fab9071SNeal Cardwell inet_csk(sk)->icsk_af_ops->mtu_reduced(sk); 830144d56e9SEric Dumazet __sock_put(sk); 831144d56e9SEric Dumazet } 83246d3ceabSEric Dumazet } 83346d3ceabSEric Dumazet EXPORT_SYMBOL(tcp_release_cb); 83446d3ceabSEric Dumazet 83546d3ceabSEric Dumazet void __init tcp_tasklet_init(void) 83646d3ceabSEric Dumazet { 83746d3ceabSEric Dumazet int i; 83846d3ceabSEric Dumazet 83946d3ceabSEric Dumazet for_each_possible_cpu(i) { 84046d3ceabSEric Dumazet struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i); 84146d3ceabSEric Dumazet 84246d3ceabSEric Dumazet INIT_LIST_HEAD(&tsq->head); 84346d3ceabSEric Dumazet tasklet_init(&tsq->tasklet, 84446d3ceabSEric Dumazet tcp_tasklet_func, 84546d3ceabSEric Dumazet (unsigned long)tsq); 84646d3ceabSEric Dumazet } 84746d3ceabSEric Dumazet } 84846d3ceabSEric Dumazet 84946d3ceabSEric Dumazet /* 85046d3ceabSEric Dumazet * Write buffer destructor automatically called from kfree_skb. 8518e3bff96Sstephen hemminger * We can't xmit new skbs from this context, as we might already 85246d3ceabSEric Dumazet * hold qdisc lock. 85346d3ceabSEric Dumazet */ 854d6a4a104SEric Dumazet void tcp_wfree(struct sk_buff *skb) 85546d3ceabSEric Dumazet { 85646d3ceabSEric Dumazet struct sock *sk = skb->sk; 85746d3ceabSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 8589b462d02SEric Dumazet int wmem; 8599b462d02SEric Dumazet 8609b462d02SEric Dumazet /* Keep one reference on sk_wmem_alloc. 8619b462d02SEric Dumazet * Will be released by sk_free() from here or tcp_tasklet_func() 8629b462d02SEric Dumazet */ 8639b462d02SEric Dumazet wmem = atomic_sub_return(skb->truesize - 1, &sk->sk_wmem_alloc); 8649b462d02SEric Dumazet 8659b462d02SEric Dumazet /* If this softirq is serviced by ksoftirqd, we are likely under stress. 8669b462d02SEric Dumazet * Wait until our queues (qdisc + devices) are drained. 8679b462d02SEric Dumazet * This gives : 8689b462d02SEric Dumazet * - less callbacks to tcp_write_xmit(), reducing stress (batches) 8699b462d02SEric Dumazet * - chance for incoming ACK (processed by another cpu maybe) 8709b462d02SEric Dumazet * to migrate this flow (skb->ooo_okay will be eventually set) 8719b462d02SEric Dumazet */ 8729b462d02SEric Dumazet if (wmem >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current) 8739b462d02SEric Dumazet goto out; 87446d3ceabSEric Dumazet 87546d3ceabSEric Dumazet if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) && 87646d3ceabSEric Dumazet !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) { 87746d3ceabSEric Dumazet unsigned long flags; 87846d3ceabSEric Dumazet struct tsq_tasklet *tsq; 87946d3ceabSEric Dumazet 88046d3ceabSEric Dumazet /* queue this socket to tasklet queue */ 88146d3ceabSEric Dumazet local_irq_save(flags); 882903ceff7SChristoph Lameter tsq = this_cpu_ptr(&tsq_tasklet); 88346d3ceabSEric Dumazet list_add(&tp->tsq_node, &tsq->head); 88446d3ceabSEric Dumazet tasklet_schedule(&tsq->tasklet); 88546d3ceabSEric Dumazet local_irq_restore(flags); 8869b462d02SEric Dumazet return; 88746d3ceabSEric Dumazet } 8889b462d02SEric Dumazet out: 8899b462d02SEric Dumazet sk_free(sk); 89046d3ceabSEric Dumazet } 89146d3ceabSEric Dumazet 8921da177e4SLinus Torvalds /* This routine actually transmits TCP packets queued in by 8931da177e4SLinus Torvalds * tcp_do_sendmsg(). This is used by both the initial 8941da177e4SLinus Torvalds * transmission and possible later retransmissions. 8951da177e4SLinus Torvalds * All SKB's seen here are completely headerless. It is our 8961da177e4SLinus Torvalds * job to build the TCP header, and pass the packet down to 8971da177e4SLinus Torvalds * IP so it can do the same plus pass the packet off to the 8981da177e4SLinus Torvalds * device. 8991da177e4SLinus Torvalds * 9001da177e4SLinus Torvalds * We are working here with either a clone of the original 9011da177e4SLinus Torvalds * SKB, or a fresh unique copy made by the retransmit engine. 9021da177e4SLinus Torvalds */ 903056834d9SIlpo Järvinen static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, 904056834d9SIlpo Järvinen gfp_t gfp_mask) 9051da177e4SLinus Torvalds { 9066687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 907dfb4b9dcSDavid S. Miller struct inet_sock *inet; 908dfb4b9dcSDavid S. Miller struct tcp_sock *tp; 909dfb4b9dcSDavid S. Miller struct tcp_skb_cb *tcb; 91033ad798cSAdam Langley struct tcp_out_options opts; 91195c96174SEric Dumazet unsigned int tcp_options_size, tcp_header_size; 912cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 9131da177e4SLinus Torvalds struct tcphdr *th; 9141da177e4SLinus Torvalds int err; 9151da177e4SLinus Torvalds 916dfb4b9dcSDavid S. Miller BUG_ON(!skb || !tcp_skb_pcount(skb)); 917dfb4b9dcSDavid S. Miller 918ccdbb6e9SEric Dumazet if (clone_it) { 919740b0f18SEric Dumazet skb_mstamp_get(&skb->skb_mstamp); 920dfb4b9dcSDavid S. Miller 921dfb4b9dcSDavid S. Miller if (unlikely(skb_cloned(skb))) 922dfb4b9dcSDavid S. Miller skb = pskb_copy(skb, gfp_mask); 923dfb4b9dcSDavid S. Miller else 924dfb4b9dcSDavid S. Miller skb = skb_clone(skb, gfp_mask); 925dfb4b9dcSDavid S. Miller if (unlikely(!skb)) 926dfb4b9dcSDavid S. Miller return -ENOBUFS; 927dfb4b9dcSDavid S. Miller } 928dfb4b9dcSDavid S. Miller 929dfb4b9dcSDavid S. Miller inet = inet_sk(sk); 930dfb4b9dcSDavid S. Miller tp = tcp_sk(sk); 931dfb4b9dcSDavid S. Miller tcb = TCP_SKB_CB(skb); 93233ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 9331da177e4SLinus Torvalds 9344de075e0SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) 93533ad798cSAdam Langley tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); 93633ad798cSAdam Langley else 93733ad798cSAdam Langley tcp_options_size = tcp_established_options(sk, skb, &opts, 93833ad798cSAdam Langley &md5); 93933ad798cSAdam Langley tcp_header_size = tcp_options_size + sizeof(struct tcphdr); 9401da177e4SLinus Torvalds 941547669d4SEric Dumazet /* if no packet is in qdisc/device queue, then allow XPS to select 942b2532eb9SEric Dumazet * another queue. We can be called from tcp_tsq_handler() 943b2532eb9SEric Dumazet * which holds one reference to sk_wmem_alloc. 944b2532eb9SEric Dumazet * 945b2532eb9SEric Dumazet * TODO: Ideally, in-flight pure ACK packets should not matter here. 946b2532eb9SEric Dumazet * One way to get this would be to set skb->truesize = 2 on them. 947547669d4SEric Dumazet */ 948b2532eb9SEric Dumazet skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1); 9491da177e4SLinus Torvalds 950aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 951aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 95246d3ceabSEric Dumazet 95346d3ceabSEric Dumazet skb_orphan(skb); 95446d3ceabSEric Dumazet skb->sk = sk; 95598781965SEric Dumazet skb->destructor = skb_is_tcp_pure_ack(skb) ? sock_wfree : tcp_wfree; 956b73c3d0eSTom Herbert skb_set_hash_from_sk(skb, sk); 95746d3ceabSEric Dumazet atomic_add(skb->truesize, &sk->sk_wmem_alloc); 9581da177e4SLinus Torvalds 9591da177e4SLinus Torvalds /* Build TCP header and checksum it. */ 960aa8223c7SArnaldo Carvalho de Melo th = tcp_hdr(skb); 961c720c7e8SEric Dumazet th->source = inet->inet_sport; 962c720c7e8SEric Dumazet th->dest = inet->inet_dport; 9631da177e4SLinus Torvalds th->seq = htonl(tcb->seq); 9641da177e4SLinus Torvalds th->ack_seq = htonl(tp->rcv_nxt); 965df7a3b07SAl Viro *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | 9664de075e0SEric Dumazet tcb->tcp_flags); 967dfb4b9dcSDavid S. Miller 9684de075e0SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) { 9691da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments 9701da177e4SLinus Torvalds * is never scaled. 9711da177e4SLinus Torvalds */ 972600ff0c2SIlpo Järvinen th->window = htons(min(tp->rcv_wnd, 65535U)); 9731da177e4SLinus Torvalds } else { 9741da177e4SLinus Torvalds th->window = htons(tcp_select_window(sk)); 9751da177e4SLinus Torvalds } 9761da177e4SLinus Torvalds th->check = 0; 9771da177e4SLinus Torvalds th->urg_ptr = 0; 9781da177e4SLinus Torvalds 97933f5f57eSIlpo Järvinen /* The urg_mode check is necessary during a below snd_una win probe */ 9807691367dSHerbert Xu if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { 9817691367dSHerbert Xu if (before(tp->snd_up, tcb->seq + 0x10000)) { 9821da177e4SLinus Torvalds th->urg_ptr = htons(tp->snd_up - tcb->seq); 9831da177e4SLinus Torvalds th->urg = 1; 9847691367dSHerbert Xu } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { 9850eae88f3SEric Dumazet th->urg_ptr = htons(0xFFFF); 9867691367dSHerbert Xu th->urg = 1; 9877691367dSHerbert Xu } 9881da177e4SLinus Torvalds } 9891da177e4SLinus Torvalds 990bd0388aeSWilliam Allen Simpson tcp_options_write((__be32 *)(th + 1), tp, &opts); 99151466a75SEric Dumazet skb_shinfo(skb)->gso_type = sk->sk_gso_type; 9924de075e0SEric Dumazet if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0)) 993735d3831SFlorian Westphal tcp_ecn_send(sk, skb, tcp_header_size); 994dfb4b9dcSDavid S. Miller 995cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 996cfb6eeb4SYOSHIFUJI Hideaki /* Calculate the MD5 hash, as we have all we need now */ 997cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 998a465419bSEric Dumazet sk_nocaps_add(sk, NETIF_F_GSO_MASK); 999bd0388aeSWilliam Allen Simpson tp->af_specific->calc_md5_hash(opts.hash_location, 100039f8e58eSEric Dumazet md5, sk, skb); 1001cfb6eeb4SYOSHIFUJI Hideaki } 1002cfb6eeb4SYOSHIFUJI Hideaki #endif 1003cfb6eeb4SYOSHIFUJI Hideaki 1004bb296246SHerbert Xu icsk->icsk_af_ops->send_check(sk, skb); 10051da177e4SLinus Torvalds 10064de075e0SEric Dumazet if (likely(tcb->tcp_flags & TCPHDR_ACK)) 1007fc6415bcSDavid S. Miller tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); 10081da177e4SLinus Torvalds 10091da177e4SLinus Torvalds if (skb->len != tcp_header_size) 1010cf533ea5SEric Dumazet tcp_event_data_sent(tp, sk); 10111da177e4SLinus Torvalds 1012bd37a088SWei Yongjun if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) 1013aa2ea058STom Herbert TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, 1014aa2ea058STom Herbert tcp_skb_pcount(skb)); 10151da177e4SLinus Torvalds 10162efd055cSMarcelo Ricardo Leitner tp->segs_out += tcp_skb_pcount(skb); 1017f69ad292SEric Dumazet /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */ 1018cd7d8498SEric Dumazet skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb); 1019f69ad292SEric Dumazet skb_shinfo(skb)->gso_size = tcp_skb_mss(skb); 1020cd7d8498SEric Dumazet 10217faee5c0SEric Dumazet /* Our usage of tstamp should remain private */ 10227faee5c0SEric Dumazet skb->tstamp.tv64 = 0; 1023971f10ecSEric Dumazet 1024971f10ecSEric Dumazet /* Cleanup our debris for IP stacks */ 1025971f10ecSEric Dumazet memset(skb->cb, 0, max(sizeof(struct inet_skb_parm), 1026971f10ecSEric Dumazet sizeof(struct inet6_skb_parm))); 1027971f10ecSEric Dumazet 1028b0270e91SEric Dumazet err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl); 10297faee5c0SEric Dumazet 103083de47cdSHua Zhong if (likely(err <= 0)) 10311da177e4SLinus Torvalds return err; 10321da177e4SLinus Torvalds 10335ee2c941SChristoph Paasch tcp_enter_cwr(sk); 10341da177e4SLinus Torvalds 1035b9df3cb8SGerrit Renker return net_xmit_eval(err); 10361da177e4SLinus Torvalds } 10371da177e4SLinus Torvalds 103867edfef7SAndi Kleen /* This routine just queues the buffer for sending. 10391da177e4SLinus Torvalds * 10401da177e4SLinus Torvalds * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, 10411da177e4SLinus Torvalds * otherwise socket can stall. 10421da177e4SLinus Torvalds */ 10431da177e4SLinus Torvalds static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) 10441da177e4SLinus Torvalds { 10451da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 10461da177e4SLinus Torvalds 10471da177e4SLinus Torvalds /* Advance write_seq and place onto the write_queue. */ 10481da177e4SLinus Torvalds tp->write_seq = TCP_SKB_CB(skb)->end_seq; 1049f4a775d1SEric Dumazet __skb_header_release(skb); 1050fe067e8aSDavid S. Miller tcp_add_write_queue_tail(sk, skb); 10513ab224beSHideo Aoki sk->sk_wmem_queued += skb->truesize; 10523ab224beSHideo Aoki sk_mem_charge(sk, skb->truesize); 10531da177e4SLinus Torvalds } 10541da177e4SLinus Torvalds 105567edfef7SAndi Kleen /* Initialize TSO segments for a packet. */ 10565bbb432cSEric Dumazet static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now) 1057f6302d1dSDavid S. Miller { 10588f26fb1cSEric Dumazet if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) { 1059f6302d1dSDavid S. Miller /* Avoid the costly divide in the normal 1060f6302d1dSDavid S. Miller * non-TSO case. 1061f6302d1dSDavid S. Miller */ 1062cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, 1); 1063f69ad292SEric Dumazet TCP_SKB_CB(skb)->tcp_gso_size = 0; 1064f6302d1dSDavid S. Miller } else { 1065cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now)); 1066f69ad292SEric Dumazet TCP_SKB_CB(skb)->tcp_gso_size = mss_now; 10671da177e4SLinus Torvalds } 10681da177e4SLinus Torvalds } 10691da177e4SLinus Torvalds 107091fed7a1SIlpo Järvinen /* When a modification to fackets out becomes necessary, we need to check 107168f8353bSIlpo Järvinen * skb is counted to fackets_out or not. 107291fed7a1SIlpo Järvinen */ 1073cf533ea5SEric Dumazet static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb, 107491fed7a1SIlpo Järvinen int decr) 107591fed7a1SIlpo Järvinen { 1076a47e5a98SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1077a47e5a98SIlpo Järvinen 1078dc86967bSIlpo Järvinen if (!tp->sacked_out || tcp_is_reno(tp)) 107991fed7a1SIlpo Järvinen return; 108091fed7a1SIlpo Järvinen 10816859d494SIlpo Järvinen if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq)) 108291fed7a1SIlpo Järvinen tp->fackets_out -= decr; 108391fed7a1SIlpo Järvinen } 108491fed7a1SIlpo Järvinen 1085797108d1SIlpo Järvinen /* Pcount in the middle of the write queue got changed, we need to do various 1086797108d1SIlpo Järvinen * tweaks to fix counters 1087797108d1SIlpo Järvinen */ 1088cf533ea5SEric Dumazet static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) 1089797108d1SIlpo Järvinen { 1090797108d1SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1091797108d1SIlpo Järvinen 1092797108d1SIlpo Järvinen tp->packets_out -= decr; 1093797108d1SIlpo Järvinen 1094797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 1095797108d1SIlpo Järvinen tp->sacked_out -= decr; 1096797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 1097797108d1SIlpo Järvinen tp->retrans_out -= decr; 1098797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 1099797108d1SIlpo Järvinen tp->lost_out -= decr; 1100797108d1SIlpo Järvinen 1101797108d1SIlpo Järvinen /* Reno case is special. Sigh... */ 1102797108d1SIlpo Järvinen if (tcp_is_reno(tp) && decr > 0) 1103797108d1SIlpo Järvinen tp->sacked_out -= min_t(u32, tp->sacked_out, decr); 1104797108d1SIlpo Järvinen 1105797108d1SIlpo Järvinen tcp_adjust_fackets_out(sk, skb, decr); 1106797108d1SIlpo Järvinen 1107797108d1SIlpo Järvinen if (tp->lost_skb_hint && 1108797108d1SIlpo Järvinen before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && 110952cf3cc8SIlpo Järvinen (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))) 1110797108d1SIlpo Järvinen tp->lost_cnt_hint -= decr; 1111797108d1SIlpo Järvinen 1112797108d1SIlpo Järvinen tcp_verify_left_out(tp); 1113797108d1SIlpo Järvinen } 1114797108d1SIlpo Järvinen 1115490cc7d0SWillem de Bruijn static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2) 1116490cc7d0SWillem de Bruijn { 1117490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo = skb_shinfo(skb); 1118490cc7d0SWillem de Bruijn 1119490cc7d0SWillem de Bruijn if (unlikely(shinfo->tx_flags & SKBTX_ANY_TSTAMP) && 1120490cc7d0SWillem de Bruijn !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) { 1121490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo2 = skb_shinfo(skb2); 1122490cc7d0SWillem de Bruijn u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP; 1123490cc7d0SWillem de Bruijn 1124490cc7d0SWillem de Bruijn shinfo->tx_flags &= ~tsflags; 1125490cc7d0SWillem de Bruijn shinfo2->tx_flags |= tsflags; 1126490cc7d0SWillem de Bruijn swap(shinfo->tskey, shinfo2->tskey); 1127490cc7d0SWillem de Bruijn } 1128490cc7d0SWillem de Bruijn } 1129490cc7d0SWillem de Bruijn 11301da177e4SLinus Torvalds /* Function to create two new TCP segments. Shrinks the given segment 11311da177e4SLinus Torvalds * to the specified size and appends a new segment with the rest of the 11321da177e4SLinus Torvalds * packet to the list. This won't be called frequently, I hope. 11331da177e4SLinus Torvalds * Remember, these are still headerless SKBs at this point. 11341da177e4SLinus Torvalds */ 1135056834d9SIlpo Järvinen int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, 11366cc55e09SOctavian Purdila unsigned int mss_now, gfp_t gfp) 11371da177e4SLinus Torvalds { 11381da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 11391da177e4SLinus Torvalds struct sk_buff *buff; 11406475be16SDavid S. Miller int nsize, old_factor; 1141b60b49eaSHerbert Xu int nlen; 11429ce01461SIlpo Järvinen u8 flags; 11431da177e4SLinus Torvalds 11442fceec13SIlpo Järvinen if (WARN_ON(len > skb->len)) 11452fceec13SIlpo Järvinen return -EINVAL; 11466a438bbeSStephen Hemminger 11471da177e4SLinus Torvalds nsize = skb_headlen(skb) - len; 11481da177e4SLinus Torvalds if (nsize < 0) 11491da177e4SLinus Torvalds nsize = 0; 11501da177e4SLinus Torvalds 11516cc55e09SOctavian Purdila if (skb_unclone(skb, gfp)) 11521da177e4SLinus Torvalds return -ENOMEM; 11531da177e4SLinus Torvalds 11541da177e4SLinus Torvalds /* Get a new skb... force flag on. */ 1155eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, nsize, gfp, true); 115651456b29SIan Morris if (!buff) 11571da177e4SLinus Torvalds return -ENOMEM; /* We'll just try again later. */ 1158ef5cb973SHerbert Xu 11593ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 11603ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1161b60b49eaSHerbert Xu nlen = skb->len - len - nsize; 1162b60b49eaSHerbert Xu buff->truesize += nlen; 1163b60b49eaSHerbert Xu skb->truesize -= nlen; 11641da177e4SLinus Torvalds 11651da177e4SLinus Torvalds /* Correct the sequence numbers. */ 11661da177e4SLinus Torvalds TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 11671da177e4SLinus Torvalds TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 11681da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 11691da177e4SLinus Torvalds 11701da177e4SLinus Torvalds /* PSH and FIN should only be set in the second packet. */ 11714de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 11724de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 11734de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1174e14c3cafSHerbert Xu TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; 11751da177e4SLinus Torvalds 117684fa7933SPatrick McHardy if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { 11771da177e4SLinus Torvalds /* Copy and checksum data tail into the new buffer. */ 1178056834d9SIlpo Järvinen buff->csum = csum_partial_copy_nocheck(skb->data + len, 1179056834d9SIlpo Järvinen skb_put(buff, nsize), 11801da177e4SLinus Torvalds nsize, 0); 11811da177e4SLinus Torvalds 11821da177e4SLinus Torvalds skb_trim(skb, len); 11831da177e4SLinus Torvalds 11841da177e4SLinus Torvalds skb->csum = csum_block_sub(skb->csum, buff->csum, len); 11851da177e4SLinus Torvalds } else { 118684fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 11871da177e4SLinus Torvalds skb_split(skb, buff, len); 11881da177e4SLinus Torvalds } 11891da177e4SLinus Torvalds 11901da177e4SLinus Torvalds buff->ip_summed = skb->ip_summed; 11911da177e4SLinus Torvalds 1192a61bbcf2SPatrick McHardy buff->tstamp = skb->tstamp; 1193490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 11941da177e4SLinus Torvalds 11956475be16SDavid S. Miller old_factor = tcp_skb_pcount(skb); 11966475be16SDavid S. Miller 11971da177e4SLinus Torvalds /* Fix up tso_factor for both original and new SKB. */ 11985bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 11995bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 12001da177e4SLinus Torvalds 12016475be16SDavid S. Miller /* If this packet has been sent out already, we must 12026475be16SDavid S. Miller * adjust the various packet counters. 12036475be16SDavid S. Miller */ 1204cf0b450cSHerbert Xu if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 12056475be16SDavid S. Miller int diff = old_factor - tcp_skb_pcount(skb) - 12066475be16SDavid S. Miller tcp_skb_pcount(buff); 12071da177e4SLinus Torvalds 1208797108d1SIlpo Järvinen if (diff) 1209797108d1SIlpo Järvinen tcp_adjust_pcount(sk, skb, diff); 12101da177e4SLinus Torvalds } 12111da177e4SLinus Torvalds 12121da177e4SLinus Torvalds /* Link BUFF into the send queue. */ 1213f4a775d1SEric Dumazet __skb_header_release(buff); 1214fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 12151da177e4SLinus Torvalds 12161da177e4SLinus Torvalds return 0; 12171da177e4SLinus Torvalds } 12181da177e4SLinus Torvalds 12191da177e4SLinus Torvalds /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c 12201da177e4SLinus Torvalds * eventually). The difference is that pulled data not copied, but 12211da177e4SLinus Torvalds * immediately discarded. 12221da177e4SLinus Torvalds */ 1223f2911969SHerbert Xu ~{PmVHI~} static void __pskb_trim_head(struct sk_buff *skb, int len) 12241da177e4SLinus Torvalds { 12257b7fc97aSEric Dumazet struct skb_shared_info *shinfo; 12261da177e4SLinus Torvalds int i, k, eat; 12271da177e4SLinus Torvalds 12284fa48bf3SEric Dumazet eat = min_t(int, len, skb_headlen(skb)); 12294fa48bf3SEric Dumazet if (eat) { 12304fa48bf3SEric Dumazet __skb_pull(skb, eat); 12314fa48bf3SEric Dumazet len -= eat; 12324fa48bf3SEric Dumazet if (!len) 12334fa48bf3SEric Dumazet return; 12344fa48bf3SEric Dumazet } 12351da177e4SLinus Torvalds eat = len; 12361da177e4SLinus Torvalds k = 0; 12377b7fc97aSEric Dumazet shinfo = skb_shinfo(skb); 12387b7fc97aSEric Dumazet for (i = 0; i < shinfo->nr_frags; i++) { 12397b7fc97aSEric Dumazet int size = skb_frag_size(&shinfo->frags[i]); 12409e903e08SEric Dumazet 12419e903e08SEric Dumazet if (size <= eat) { 1242aff65da0SIan Campbell skb_frag_unref(skb, i); 12439e903e08SEric Dumazet eat -= size; 12441da177e4SLinus Torvalds } else { 12457b7fc97aSEric Dumazet shinfo->frags[k] = shinfo->frags[i]; 12461da177e4SLinus Torvalds if (eat) { 12477b7fc97aSEric Dumazet shinfo->frags[k].page_offset += eat; 12487b7fc97aSEric Dumazet skb_frag_size_sub(&shinfo->frags[k], eat); 12491da177e4SLinus Torvalds eat = 0; 12501da177e4SLinus Torvalds } 12511da177e4SLinus Torvalds k++; 12521da177e4SLinus Torvalds } 12531da177e4SLinus Torvalds } 12547b7fc97aSEric Dumazet shinfo->nr_frags = k; 12551da177e4SLinus Torvalds 125627a884dcSArnaldo Carvalho de Melo skb_reset_tail_pointer(skb); 12571da177e4SLinus Torvalds skb->data_len -= len; 12581da177e4SLinus Torvalds skb->len = skb->data_len; 12591da177e4SLinus Torvalds } 12601da177e4SLinus Torvalds 126167edfef7SAndi Kleen /* Remove acked data from a packet in the transmit queue. */ 12621da177e4SLinus Torvalds int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 12631da177e4SLinus Torvalds { 126414bbd6a5SPravin B Shelar if (skb_unclone(skb, GFP_ATOMIC)) 12651da177e4SLinus Torvalds return -ENOMEM; 12661da177e4SLinus Torvalds 12674fa48bf3SEric Dumazet __pskb_trim_head(skb, len); 12681da177e4SLinus Torvalds 12691da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq += len; 127084fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 12711da177e4SLinus Torvalds 12721da177e4SLinus Torvalds skb->truesize -= len; 12731da177e4SLinus Torvalds sk->sk_wmem_queued -= len; 12743ab224beSHideo Aoki sk_mem_uncharge(sk, len); 12751da177e4SLinus Torvalds sock_set_flag(sk, SOCK_QUEUE_SHRUNK); 12761da177e4SLinus Torvalds 12775b35e1e6SNeal Cardwell /* Any change of skb->len requires recalculation of tso factor. */ 12781da177e4SLinus Torvalds if (tcp_skb_pcount(skb) > 1) 12795bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb)); 12801da177e4SLinus Torvalds 12811da177e4SLinus Torvalds return 0; 12821da177e4SLinus Torvalds } 12831da177e4SLinus Torvalds 12841b63edd6SYuchung Cheng /* Calculate MSS not accounting any TCP options. */ 12851b63edd6SYuchung Cheng static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu) 12865d424d5aSJohn Heffner { 1287cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1288cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 12895d424d5aSJohn Heffner int mss_now; 12905d424d5aSJohn Heffner 12915d424d5aSJohn Heffner /* Calculate base mss without TCP options: 12925d424d5aSJohn Heffner It is MMS_S - sizeof(tcphdr) of rfc1122 12935d424d5aSJohn Heffner */ 12945d424d5aSJohn Heffner mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 12955d424d5aSJohn Heffner 129667469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 129767469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 129867469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 129967469601SEric Dumazet 130067469601SEric Dumazet if (dst && dst_allfrag(dst)) 130167469601SEric Dumazet mss_now -= icsk->icsk_af_ops->net_frag_header_len; 130267469601SEric Dumazet } 130367469601SEric Dumazet 13045d424d5aSJohn Heffner /* Clamp it (mss_clamp does not include tcp options) */ 13055d424d5aSJohn Heffner if (mss_now > tp->rx_opt.mss_clamp) 13065d424d5aSJohn Heffner mss_now = tp->rx_opt.mss_clamp; 13075d424d5aSJohn Heffner 13085d424d5aSJohn Heffner /* Now subtract optional transport overhead */ 13095d424d5aSJohn Heffner mss_now -= icsk->icsk_ext_hdr_len; 13105d424d5aSJohn Heffner 13115d424d5aSJohn Heffner /* Then reserve room for full set of TCP options and 8 bytes of data */ 13125d424d5aSJohn Heffner if (mss_now < 48) 13135d424d5aSJohn Heffner mss_now = 48; 13145d424d5aSJohn Heffner return mss_now; 13155d424d5aSJohn Heffner } 13165d424d5aSJohn Heffner 13171b63edd6SYuchung Cheng /* Calculate MSS. Not accounting for SACKs here. */ 13181b63edd6SYuchung Cheng int tcp_mtu_to_mss(struct sock *sk, int pmtu) 13191b63edd6SYuchung Cheng { 13201b63edd6SYuchung Cheng /* Subtract TCP options size, not including SACKs */ 13211b63edd6SYuchung Cheng return __tcp_mtu_to_mss(sk, pmtu) - 13221b63edd6SYuchung Cheng (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr)); 13231b63edd6SYuchung Cheng } 13241b63edd6SYuchung Cheng 13255d424d5aSJohn Heffner /* Inverse of above */ 132667469601SEric Dumazet int tcp_mss_to_mtu(struct sock *sk, int mss) 13275d424d5aSJohn Heffner { 1328cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1329cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 13305d424d5aSJohn Heffner int mtu; 13315d424d5aSJohn Heffner 13325d424d5aSJohn Heffner mtu = mss + 13335d424d5aSJohn Heffner tp->tcp_header_len + 13345d424d5aSJohn Heffner icsk->icsk_ext_hdr_len + 13355d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 13365d424d5aSJohn Heffner 133767469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 133867469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 133967469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 134067469601SEric Dumazet 134167469601SEric Dumazet if (dst && dst_allfrag(dst)) 134267469601SEric Dumazet mtu += icsk->icsk_af_ops->net_frag_header_len; 134367469601SEric Dumazet } 13445d424d5aSJohn Heffner return mtu; 13455d424d5aSJohn Heffner } 13465d424d5aSJohn Heffner 134767edfef7SAndi Kleen /* MTU probing init per socket */ 13485d424d5aSJohn Heffner void tcp_mtup_init(struct sock *sk) 13495d424d5aSJohn Heffner { 13505d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 13515d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 1352b0f9ca53SFan Du struct net *net = sock_net(sk); 13535d424d5aSJohn Heffner 1354b0f9ca53SFan Du icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1; 13555d424d5aSJohn Heffner icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 13565d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 1357b0f9ca53SFan Du icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss); 13585d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 135905cbc0dbSFan Du if (icsk->icsk_mtup.enabled) 136005cbc0dbSFan Du icsk->icsk_mtup.probe_timestamp = tcp_time_stamp; 13615d424d5aSJohn Heffner } 13624bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_mtup_init); 13635d424d5aSJohn Heffner 13641da177e4SLinus Torvalds /* This function synchronize snd mss to current pmtu/exthdr set. 13651da177e4SLinus Torvalds 13661da177e4SLinus Torvalds tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 13671da177e4SLinus Torvalds for TCP options, but includes only bare TCP header. 13681da177e4SLinus Torvalds 13691da177e4SLinus Torvalds tp->rx_opt.mss_clamp is mss negotiated at connection setup. 1370caa20d9aSStephen Hemminger It is minimum of user_mss and mss received with SYN. 13711da177e4SLinus Torvalds It also does not include TCP options. 13721da177e4SLinus Torvalds 1373d83d8461SArnaldo Carvalho de Melo inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. 13741da177e4SLinus Torvalds 13751da177e4SLinus Torvalds tp->mss_cache is current effective sending mss, including 13761da177e4SLinus Torvalds all tcp options except for SACKs. It is evaluated, 13771da177e4SLinus Torvalds taking into account current pmtu, but never exceeds 13781da177e4SLinus Torvalds tp->rx_opt.mss_clamp. 13791da177e4SLinus Torvalds 13801da177e4SLinus Torvalds NOTE1. rfc1122 clearly states that advertised MSS 13811da177e4SLinus Torvalds DOES NOT include either tcp or ip options. 13821da177e4SLinus Torvalds 1383d83d8461SArnaldo Carvalho de Melo NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache 1384d83d8461SArnaldo Carvalho de Melo are READ ONLY outside this function. --ANK (980731) 13851da177e4SLinus Torvalds */ 13861da177e4SLinus Torvalds unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 13871da177e4SLinus Torvalds { 13881da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1389d83d8461SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 13905d424d5aSJohn Heffner int mss_now; 13911da177e4SLinus Torvalds 13925d424d5aSJohn Heffner if (icsk->icsk_mtup.search_high > pmtu) 13935d424d5aSJohn Heffner icsk->icsk_mtup.search_high = pmtu; 13941da177e4SLinus Torvalds 13955d424d5aSJohn Heffner mss_now = tcp_mtu_to_mss(sk, pmtu); 1396409d22b4SIlpo Järvinen mss_now = tcp_bound_to_half_wnd(tp, mss_now); 13971da177e4SLinus Torvalds 13981da177e4SLinus Torvalds /* And store cached results */ 1399d83d8461SArnaldo Carvalho de Melo icsk->icsk_pmtu_cookie = pmtu; 14005d424d5aSJohn Heffner if (icsk->icsk_mtup.enabled) 14015d424d5aSJohn Heffner mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); 1402c1b4a7e6SDavid S. Miller tp->mss_cache = mss_now; 14031da177e4SLinus Torvalds 14041da177e4SLinus Torvalds return mss_now; 14051da177e4SLinus Torvalds } 14064bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_sync_mss); 14071da177e4SLinus Torvalds 14081da177e4SLinus Torvalds /* Compute the current effective MSS, taking SACKs and IP options, 14091da177e4SLinus Torvalds * and even PMTU discovery events into account. 14101da177e4SLinus Torvalds */ 14110c54b85fSIlpo Järvinen unsigned int tcp_current_mss(struct sock *sk) 14121da177e4SLinus Torvalds { 1413cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1414cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1415c1b4a7e6SDavid S. Miller u32 mss_now; 141695c96174SEric Dumazet unsigned int header_len; 141733ad798cSAdam Langley struct tcp_out_options opts; 141833ad798cSAdam Langley struct tcp_md5sig_key *md5; 14191da177e4SLinus Torvalds 1420c1b4a7e6SDavid S. Miller mss_now = tp->mss_cache; 1421c1b4a7e6SDavid S. Miller 14221da177e4SLinus Torvalds if (dst) { 14231da177e4SLinus Torvalds u32 mtu = dst_mtu(dst); 1424d83d8461SArnaldo Carvalho de Melo if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 14251da177e4SLinus Torvalds mss_now = tcp_sync_mss(sk, mtu); 14261da177e4SLinus Torvalds } 14271da177e4SLinus Torvalds 142833ad798cSAdam Langley header_len = tcp_established_options(sk, NULL, &opts, &md5) + 142933ad798cSAdam Langley sizeof(struct tcphdr); 143033ad798cSAdam Langley /* The mss_cache is sized based on tp->tcp_header_len, which assumes 143133ad798cSAdam Langley * some common options. If this is an odd packet (because we have SACK 143233ad798cSAdam Langley * blocks etc) then our calculated header_len will be different, and 143333ad798cSAdam Langley * we have to adjust mss_now correspondingly */ 143433ad798cSAdam Langley if (header_len != tp->tcp_header_len) { 143533ad798cSAdam Langley int delta = (int) header_len - tp->tcp_header_len; 143633ad798cSAdam Langley mss_now -= delta; 143733ad798cSAdam Langley } 1438cfb6eeb4SYOSHIFUJI Hideaki 14391da177e4SLinus Torvalds return mss_now; 14401da177e4SLinus Torvalds } 14411da177e4SLinus Torvalds 144286fd14adSWeiping Pan /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. 144386fd14adSWeiping Pan * As additional protections, we do not touch cwnd in retransmission phases, 144486fd14adSWeiping Pan * and if application hit its sndbuf limit recently. 144586fd14adSWeiping Pan */ 144686fd14adSWeiping Pan static void tcp_cwnd_application_limited(struct sock *sk) 1447a762a980SDavid S. Miller { 14489e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1449a762a980SDavid S. Miller 145086fd14adSWeiping Pan if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && 145186fd14adSWeiping Pan sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 145286fd14adSWeiping Pan /* Limited by application or receiver window. */ 145386fd14adSWeiping Pan u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); 145486fd14adSWeiping Pan u32 win_used = max(tp->snd_cwnd_used, init_win); 145586fd14adSWeiping Pan if (win_used < tp->snd_cwnd) { 145686fd14adSWeiping Pan tp->snd_ssthresh = tcp_current_ssthresh(sk); 145786fd14adSWeiping Pan tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1; 145886fd14adSWeiping Pan } 145986fd14adSWeiping Pan tp->snd_cwnd_used = 0; 146086fd14adSWeiping Pan } 146186fd14adSWeiping Pan tp->snd_cwnd_stamp = tcp_time_stamp; 146286fd14adSWeiping Pan } 146386fd14adSWeiping Pan 1464ca8a2263SNeal Cardwell static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited) 1465a762a980SDavid S. Miller { 1466a762a980SDavid S. Miller struct tcp_sock *tp = tcp_sk(sk); 1467a762a980SDavid S. Miller 1468ca8a2263SNeal Cardwell /* Track the maximum number of outstanding packets in each 1469ca8a2263SNeal Cardwell * window, and remember whether we were cwnd-limited then. 1470ca8a2263SNeal Cardwell */ 1471ca8a2263SNeal Cardwell if (!before(tp->snd_una, tp->max_packets_seq) || 1472ca8a2263SNeal Cardwell tp->packets_out > tp->max_packets_out) { 1473ca8a2263SNeal Cardwell tp->max_packets_out = tp->packets_out; 1474ca8a2263SNeal Cardwell tp->max_packets_seq = tp->snd_nxt; 1475ca8a2263SNeal Cardwell tp->is_cwnd_limited = is_cwnd_limited; 1476ca8a2263SNeal Cardwell } 1477e114a710SEric Dumazet 147824901551SEric Dumazet if (tcp_is_cwnd_limited(sk)) { 1479a762a980SDavid S. Miller /* Network is feed fully. */ 1480a762a980SDavid S. Miller tp->snd_cwnd_used = 0; 1481a762a980SDavid S. Miller tp->snd_cwnd_stamp = tcp_time_stamp; 1482a762a980SDavid S. Miller } else { 1483a762a980SDavid S. Miller /* Network starves. */ 1484a762a980SDavid S. Miller if (tp->packets_out > tp->snd_cwnd_used) 1485a762a980SDavid S. Miller tp->snd_cwnd_used = tp->packets_out; 1486a762a980SDavid S. Miller 148715d33c07SDavid S. Miller if (sysctl_tcp_slow_start_after_idle && 148815d33c07SDavid S. Miller (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto) 1489a762a980SDavid S. Miller tcp_cwnd_application_limited(sk); 1490a762a980SDavid S. Miller } 1491a762a980SDavid S. Miller } 1492a762a980SDavid S. Miller 1493d4589926SEric Dumazet /* Minshall's variant of the Nagle send check. */ 1494d4589926SEric Dumazet static bool tcp_minshall_check(const struct tcp_sock *tp) 1495d4589926SEric Dumazet { 1496d4589926SEric Dumazet return after(tp->snd_sml, tp->snd_una) && 1497d4589926SEric Dumazet !after(tp->snd_sml, tp->snd_nxt); 1498d4589926SEric Dumazet } 1499d4589926SEric Dumazet 1500d4589926SEric Dumazet /* Update snd_sml if this skb is under mss 1501d4589926SEric Dumazet * Note that a TSO packet might end with a sub-mss segment 1502d4589926SEric Dumazet * The test is really : 1503d4589926SEric Dumazet * if ((skb->len % mss) != 0) 1504d4589926SEric Dumazet * tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1505d4589926SEric Dumazet * But we can avoid doing the divide again given we already have 1506d4589926SEric Dumazet * skb_pcount = skb->len / mss_now 15070e3a4803SIlpo Järvinen */ 1508d4589926SEric Dumazet static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now, 1509d4589926SEric Dumazet const struct sk_buff *skb) 1510d4589926SEric Dumazet { 1511d4589926SEric Dumazet if (skb->len < tcp_skb_pcount(skb) * mss_now) 1512d4589926SEric Dumazet tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1513d4589926SEric Dumazet } 1514d4589926SEric Dumazet 1515d4589926SEric Dumazet /* Return false, if packet can be sent now without violation Nagle's rules: 1516d4589926SEric Dumazet * 1. It is full sized. (provided by caller in %partial bool) 1517d4589926SEric Dumazet * 2. Or it contains FIN. (already checked by caller) 1518d4589926SEric Dumazet * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. 1519d4589926SEric Dumazet * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 1520d4589926SEric Dumazet * With Minshall's modification: all sent small packets are ACKed. 1521d4589926SEric Dumazet */ 1522d4589926SEric Dumazet static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp, 1523cc93fc51SPeter Pan(潘卫平) int nonagle) 1524d4589926SEric Dumazet { 1525d4589926SEric Dumazet return partial && 1526d4589926SEric Dumazet ((nonagle & TCP_NAGLE_CORK) || 1527d4589926SEric Dumazet (!nonagle && tp->packets_out && tcp_minshall_check(tp))); 1528d4589926SEric Dumazet } 1529605ad7f1SEric Dumazet 1530605ad7f1SEric Dumazet /* Return how many segs we'd like on a TSO packet, 1531605ad7f1SEric Dumazet * to send one TSO packet per ms 1532605ad7f1SEric Dumazet */ 1533605ad7f1SEric Dumazet static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now) 1534605ad7f1SEric Dumazet { 1535605ad7f1SEric Dumazet u32 bytes, segs; 1536605ad7f1SEric Dumazet 1537605ad7f1SEric Dumazet bytes = min(sk->sk_pacing_rate >> 10, 1538605ad7f1SEric Dumazet sk->sk_gso_max_size - 1 - MAX_TCP_HEADER); 1539605ad7f1SEric Dumazet 1540605ad7f1SEric Dumazet /* Goal is to send at least one packet per ms, 1541605ad7f1SEric Dumazet * not one big TSO packet every 100 ms. 1542605ad7f1SEric Dumazet * This preserves ACK clocking and is consistent 1543605ad7f1SEric Dumazet * with tcp_tso_should_defer() heuristic. 1544605ad7f1SEric Dumazet */ 1545605ad7f1SEric Dumazet segs = max_t(u32, bytes / mss_now, sysctl_tcp_min_tso_segs); 1546605ad7f1SEric Dumazet 1547605ad7f1SEric Dumazet return min_t(u32, segs, sk->sk_gso_max_segs); 1548605ad7f1SEric Dumazet } 1549605ad7f1SEric Dumazet 1550d4589926SEric Dumazet /* Returns the portion of skb which can be sent right away */ 1551d4589926SEric Dumazet static unsigned int tcp_mss_split_point(const struct sock *sk, 1552d4589926SEric Dumazet const struct sk_buff *skb, 1553d4589926SEric Dumazet unsigned int mss_now, 1554d4589926SEric Dumazet unsigned int max_segs, 1555d4589926SEric Dumazet int nonagle) 1556c1b4a7e6SDavid S. Miller { 1557cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1558d4589926SEric Dumazet u32 partial, needed, window, max_len; 1559c1b4a7e6SDavid S. Miller 156090840defSIlpo Järvinen window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 15611485348dSBen Hutchings max_len = mss_now * max_segs; 15620e3a4803SIlpo Järvinen 15631485348dSBen Hutchings if (likely(max_len <= window && skb != tcp_write_queue_tail(sk))) 15641485348dSBen Hutchings return max_len; 15650e3a4803SIlpo Järvinen 15665ea3a748SIlpo Järvinen needed = min(skb->len, window); 15675ea3a748SIlpo Järvinen 15681485348dSBen Hutchings if (max_len <= needed) 15691485348dSBen Hutchings return max_len; 15700e3a4803SIlpo Järvinen 1571d4589926SEric Dumazet partial = needed % mss_now; 1572d4589926SEric Dumazet /* If last segment is not a full MSS, check if Nagle rules allow us 1573d4589926SEric Dumazet * to include this last segment in this skb. 1574d4589926SEric Dumazet * Otherwise, we'll split the skb at last MSS boundary 1575d4589926SEric Dumazet */ 1576cc93fc51SPeter Pan(潘卫平) if (tcp_nagle_check(partial != 0, tp, nonagle)) 1577d4589926SEric Dumazet return needed - partial; 1578d4589926SEric Dumazet 1579d4589926SEric Dumazet return needed; 1580c1b4a7e6SDavid S. Miller } 1581c1b4a7e6SDavid S. Miller 1582c1b4a7e6SDavid S. Miller /* Can at least one segment of SKB be sent right now, according to the 1583c1b4a7e6SDavid S. Miller * congestion window rules? If so, return how many segments are allowed. 1584c1b4a7e6SDavid S. Miller */ 1585cf533ea5SEric Dumazet static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, 1586cf533ea5SEric Dumazet const struct sk_buff *skb) 1587c1b4a7e6SDavid S. Miller { 1588d649a7a8SEric Dumazet u32 in_flight, cwnd, halfcwnd; 1589c1b4a7e6SDavid S. Miller 1590c1b4a7e6SDavid S. Miller /* Don't be strict about the congestion window for the final FIN. */ 15914de075e0SEric Dumazet if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && 15924de075e0SEric Dumazet tcp_skb_pcount(skb) == 1) 1593c1b4a7e6SDavid S. Miller return 1; 1594c1b4a7e6SDavid S. Miller 1595c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1596c1b4a7e6SDavid S. Miller cwnd = tp->snd_cwnd; 1597d649a7a8SEric Dumazet if (in_flight >= cwnd) 1598c1b4a7e6SDavid S. Miller return 0; 1599d649a7a8SEric Dumazet 1600d649a7a8SEric Dumazet /* For better scheduling, ensure we have at least 1601d649a7a8SEric Dumazet * 2 GSO packets in flight. 1602d649a7a8SEric Dumazet */ 1603d649a7a8SEric Dumazet halfcwnd = max(cwnd >> 1, 1U); 1604d649a7a8SEric Dumazet return min(halfcwnd, cwnd - in_flight); 1605c1b4a7e6SDavid S. Miller } 1606c1b4a7e6SDavid S. Miller 1607b595076aSUwe Kleine-König /* Initialize TSO state of a skb. 160867edfef7SAndi Kleen * This must be invoked the first time we consider transmitting 1609c1b4a7e6SDavid S. Miller * SKB onto the wire. 1610c1b4a7e6SDavid S. Miller */ 16115bbb432cSEric Dumazet static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now) 1612c1b4a7e6SDavid S. Miller { 1613c1b4a7e6SDavid S. Miller int tso_segs = tcp_skb_pcount(skb); 1614c1b4a7e6SDavid S. Miller 1615f8269a49SIlpo Järvinen if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { 16165bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 1617c1b4a7e6SDavid S. Miller tso_segs = tcp_skb_pcount(skb); 1618c1b4a7e6SDavid S. Miller } 1619c1b4a7e6SDavid S. Miller return tso_segs; 1620c1b4a7e6SDavid S. Miller } 1621c1b4a7e6SDavid S. Miller 1622c1b4a7e6SDavid S. Miller 1623a2a385d6SEric Dumazet /* Return true if the Nagle test allows this packet to be 1624c1b4a7e6SDavid S. Miller * sent now. 1625c1b4a7e6SDavid S. Miller */ 1626a2a385d6SEric Dumazet static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, 1627c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1628c1b4a7e6SDavid S. Miller { 1629c1b4a7e6SDavid S. Miller /* Nagle rule does not apply to frames, which sit in the middle of the 1630c1b4a7e6SDavid S. Miller * write_queue (they have no chances to get new data). 1631c1b4a7e6SDavid S. Miller * 1632c1b4a7e6SDavid S. Miller * This is implemented in the callers, where they modify the 'nonagle' 1633c1b4a7e6SDavid S. Miller * argument based upon the location of SKB in the send queue. 1634c1b4a7e6SDavid S. Miller */ 1635c1b4a7e6SDavid S. Miller if (nonagle & TCP_NAGLE_PUSH) 1636a2a385d6SEric Dumazet return true; 1637c1b4a7e6SDavid S. Miller 16389b44190dSYuchung Cheng /* Don't use the nagle rule for urgent data (or for the final FIN). */ 16399b44190dSYuchung Cheng if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) 1640a2a385d6SEric Dumazet return true; 1641c1b4a7e6SDavid S. Miller 1642cc93fc51SPeter Pan(潘卫平) if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle)) 1643a2a385d6SEric Dumazet return true; 1644c1b4a7e6SDavid S. Miller 1645a2a385d6SEric Dumazet return false; 1646c1b4a7e6SDavid S. Miller } 1647c1b4a7e6SDavid S. Miller 1648c1b4a7e6SDavid S. Miller /* Does at least the first segment of SKB fit into the send window? */ 1649a2a385d6SEric Dumazet static bool tcp_snd_wnd_test(const struct tcp_sock *tp, 1650a2a385d6SEric Dumazet const struct sk_buff *skb, 1651056834d9SIlpo Järvinen unsigned int cur_mss) 1652c1b4a7e6SDavid S. Miller { 1653c1b4a7e6SDavid S. Miller u32 end_seq = TCP_SKB_CB(skb)->end_seq; 1654c1b4a7e6SDavid S. Miller 1655c1b4a7e6SDavid S. Miller if (skb->len > cur_mss) 1656c1b4a7e6SDavid S. Miller end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 1657c1b4a7e6SDavid S. Miller 165890840defSIlpo Järvinen return !after(end_seq, tcp_wnd_end(tp)); 1659c1b4a7e6SDavid S. Miller } 1660c1b4a7e6SDavid S. Miller 1661fe067e8aSDavid S. Miller /* This checks if the data bearing packet SKB (usually tcp_send_head(sk)) 1662c1b4a7e6SDavid S. Miller * should be put on the wire right now. If so, it returns the number of 1663c1b4a7e6SDavid S. Miller * packets allowed by the congestion window. 1664c1b4a7e6SDavid S. Miller */ 1665cf533ea5SEric Dumazet static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb, 1666c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1667c1b4a7e6SDavid S. Miller { 1668cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1669c1b4a7e6SDavid S. Miller unsigned int cwnd_quota; 1670c1b4a7e6SDavid S. Miller 16715bbb432cSEric Dumazet tcp_init_tso_segs(skb, cur_mss); 1672c1b4a7e6SDavid S. Miller 1673c1b4a7e6SDavid S. Miller if (!tcp_nagle_test(tp, skb, cur_mss, nonagle)) 1674c1b4a7e6SDavid S. Miller return 0; 1675c1b4a7e6SDavid S. Miller 1676c1b4a7e6SDavid S. Miller cwnd_quota = tcp_cwnd_test(tp, skb); 1677056834d9SIlpo Järvinen if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss)) 1678c1b4a7e6SDavid S. Miller cwnd_quota = 0; 1679c1b4a7e6SDavid S. Miller 1680c1b4a7e6SDavid S. Miller return cwnd_quota; 1681c1b4a7e6SDavid S. Miller } 1682c1b4a7e6SDavid S. Miller 168367edfef7SAndi Kleen /* Test if sending is allowed right now. */ 1684a2a385d6SEric Dumazet bool tcp_may_send_now(struct sock *sk) 1685c1b4a7e6SDavid S. Miller { 1686cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1687fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 1688c1b4a7e6SDavid S. Miller 1689a02cec21SEric Dumazet return skb && 16900c54b85fSIlpo Järvinen tcp_snd_test(sk, skb, tcp_current_mss(sk), 1691c1b4a7e6SDavid S. Miller (tcp_skb_is_last(sk, skb) ? 1692a02cec21SEric Dumazet tp->nonagle : TCP_NAGLE_PUSH)); 1693c1b4a7e6SDavid S. Miller } 1694c1b4a7e6SDavid S. Miller 1695c1b4a7e6SDavid S. Miller /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 1696c1b4a7e6SDavid S. Miller * which is put after SKB on the list. It is very much like 1697c1b4a7e6SDavid S. Miller * tcp_fragment() except that it may make several kinds of assumptions 1698c1b4a7e6SDavid S. Miller * in order to speed up the splitting operation. In particular, we 1699c1b4a7e6SDavid S. Miller * know that all the data is in scatter-gather pages, and that the 1700c1b4a7e6SDavid S. Miller * packet has never been sent out before (and thus is not cloned). 1701c1b4a7e6SDavid S. Miller */ 1702056834d9SIlpo Järvinen static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, 1703c4ead4c5SEric Dumazet unsigned int mss_now, gfp_t gfp) 1704c1b4a7e6SDavid S. Miller { 1705c1b4a7e6SDavid S. Miller struct sk_buff *buff; 1706c1b4a7e6SDavid S. Miller int nlen = skb->len - len; 17079ce01461SIlpo Järvinen u8 flags; 1708c1b4a7e6SDavid S. Miller 1709c1b4a7e6SDavid S. Miller /* All of a TSO frame must be composed of paged data. */ 1710c8ac3774SHerbert Xu if (skb->len != skb->data_len) 17116cc55e09SOctavian Purdila return tcp_fragment(sk, skb, len, mss_now, gfp); 1712c1b4a7e6SDavid S. Miller 1713eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, gfp, true); 171451456b29SIan Morris if (unlikely(!buff)) 1715c1b4a7e6SDavid S. Miller return -ENOMEM; 1716c1b4a7e6SDavid S. Miller 17173ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 17183ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1719b60b49eaSHerbert Xu buff->truesize += nlen; 1720c1b4a7e6SDavid S. Miller skb->truesize -= nlen; 1721c1b4a7e6SDavid S. Miller 1722c1b4a7e6SDavid S. Miller /* Correct the sequence numbers. */ 1723c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 1724c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 1725c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 1726c1b4a7e6SDavid S. Miller 1727c1b4a7e6SDavid S. Miller /* PSH and FIN should only be set in the second packet. */ 17284de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 17294de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 17304de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1731c1b4a7e6SDavid S. Miller 1732c1b4a7e6SDavid S. Miller /* This packet was never sent out yet, so no SACK bits. */ 1733c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->sacked = 0; 1734c1b4a7e6SDavid S. Miller 173584fa7933SPatrick McHardy buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; 1736c1b4a7e6SDavid S. Miller skb_split(skb, buff, len); 1737490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 1738c1b4a7e6SDavid S. Miller 1739c1b4a7e6SDavid S. Miller /* Fix up tso_factor for both original and new SKB. */ 17405bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 17415bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 1742c1b4a7e6SDavid S. Miller 1743c1b4a7e6SDavid S. Miller /* Link BUFF into the send queue. */ 1744f4a775d1SEric Dumazet __skb_header_release(buff); 1745fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 1746c1b4a7e6SDavid S. Miller 1747c1b4a7e6SDavid S. Miller return 0; 1748c1b4a7e6SDavid S. Miller } 1749c1b4a7e6SDavid S. Miller 1750c1b4a7e6SDavid S. Miller /* Try to defer sending, if possible, in order to minimize the amount 1751c1b4a7e6SDavid S. Miller * of TSO splitting we do. View it as a kind of TSO Nagle test. 1752c1b4a7e6SDavid S. Miller * 1753c1b4a7e6SDavid S. Miller * This algorithm is from John Heffner. 1754c1b4a7e6SDavid S. Miller */ 1755ca8a2263SNeal Cardwell static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb, 1756605ad7f1SEric Dumazet bool *is_cwnd_limited, u32 max_segs) 1757c1b4a7e6SDavid S. Miller { 17586687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 175950c8339eSEric Dumazet u32 age, send_win, cong_win, limit, in_flight; 176050c8339eSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 176150c8339eSEric Dumazet struct skb_mstamp now; 176250c8339eSEric Dumazet struct sk_buff *head; 1763ad9f4f50SEric Dumazet int win_divisor; 1764c1b4a7e6SDavid S. Miller 17654de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 1766ae8064acSJohn Heffner goto send_now; 1767c1b4a7e6SDavid S. Miller 176899d7662aSEric Dumazet if (icsk->icsk_ca_state >= TCP_CA_Recovery) 1769ae8064acSJohn Heffner goto send_now; 1770ae8064acSJohn Heffner 17715f852eb5SEric Dumazet /* Avoid bursty behavior by allowing defer 17725f852eb5SEric Dumazet * only if the last write was recent. 17735f852eb5SEric Dumazet */ 17745f852eb5SEric Dumazet if ((s32)(tcp_time_stamp - tp->lsndtime) > 0) 1775ae8064acSJohn Heffner goto send_now; 1776908a75c1SDavid S. Miller 1777c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1778c1b4a7e6SDavid S. Miller 1779056834d9SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight)); 1780c1b4a7e6SDavid S. Miller 178190840defSIlpo Järvinen send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1782c1b4a7e6SDavid S. Miller 1783c1b4a7e6SDavid S. Miller /* From in_flight test above, we know that cwnd > in_flight. */ 1784c1b4a7e6SDavid S. Miller cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; 1785c1b4a7e6SDavid S. Miller 1786c1b4a7e6SDavid S. Miller limit = min(send_win, cong_win); 1787c1b4a7e6SDavid S. Miller 1788ba244fe9SDavid S. Miller /* If a full-sized TSO skb can be sent, do it. */ 1789605ad7f1SEric Dumazet if (limit >= max_segs * tp->mss_cache) 1790ae8064acSJohn Heffner goto send_now; 1791ba244fe9SDavid S. Miller 179262ad2761SIlpo Järvinen /* Middle in queue won't get any more data, full sendable already? */ 179362ad2761SIlpo Järvinen if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) 179462ad2761SIlpo Järvinen goto send_now; 179562ad2761SIlpo Järvinen 1796ad9f4f50SEric Dumazet win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor); 1797ad9f4f50SEric Dumazet if (win_divisor) { 1798c1b4a7e6SDavid S. Miller u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); 1799c1b4a7e6SDavid S. Miller 1800c1b4a7e6SDavid S. Miller /* If at least some fraction of a window is available, 1801c1b4a7e6SDavid S. Miller * just use it. 1802c1b4a7e6SDavid S. Miller */ 1803ad9f4f50SEric Dumazet chunk /= win_divisor; 1804c1b4a7e6SDavid S. Miller if (limit >= chunk) 1805ae8064acSJohn Heffner goto send_now; 1806c1b4a7e6SDavid S. Miller } else { 1807c1b4a7e6SDavid S. Miller /* Different approach, try not to defer past a single 1808c1b4a7e6SDavid S. Miller * ACK. Receiver should ACK every other full sized 1809c1b4a7e6SDavid S. Miller * frame, so if we have space for more than 3 frames 1810c1b4a7e6SDavid S. Miller * then send now. 1811c1b4a7e6SDavid S. Miller */ 18126b5a5c0dSNeal Cardwell if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) 1813ae8064acSJohn Heffner goto send_now; 1814c1b4a7e6SDavid S. Miller } 1815c1b4a7e6SDavid S. Miller 181650c8339eSEric Dumazet head = tcp_write_queue_head(sk); 181750c8339eSEric Dumazet skb_mstamp_get(&now); 181850c8339eSEric Dumazet age = skb_mstamp_us_delta(&now, &head->skb_mstamp); 181950c8339eSEric Dumazet /* If next ACK is likely to come too late (half srtt), do not defer */ 182050c8339eSEric Dumazet if (age < (tp->srtt_us >> 4)) 182150c8339eSEric Dumazet goto send_now; 182250c8339eSEric Dumazet 18235f852eb5SEric Dumazet /* Ok, it looks like it is advisable to defer. */ 1824ae8064acSJohn Heffner 1825d2e1339fSBendik Rønning Opstad if (cong_win < send_win && cong_win <= skb->len) 1826ca8a2263SNeal Cardwell *is_cwnd_limited = true; 1827ca8a2263SNeal Cardwell 1828a2a385d6SEric Dumazet return true; 1829ae8064acSJohn Heffner 1830ae8064acSJohn Heffner send_now: 1831a2a385d6SEric Dumazet return false; 1832c1b4a7e6SDavid S. Miller } 1833c1b4a7e6SDavid S. Miller 183405cbc0dbSFan Du static inline void tcp_mtu_check_reprobe(struct sock *sk) 183505cbc0dbSFan Du { 183605cbc0dbSFan Du struct inet_connection_sock *icsk = inet_csk(sk); 183705cbc0dbSFan Du struct tcp_sock *tp = tcp_sk(sk); 183805cbc0dbSFan Du struct net *net = sock_net(sk); 183905cbc0dbSFan Du u32 interval; 184005cbc0dbSFan Du s32 delta; 184105cbc0dbSFan Du 184205cbc0dbSFan Du interval = net->ipv4.sysctl_tcp_probe_interval; 184305cbc0dbSFan Du delta = tcp_time_stamp - icsk->icsk_mtup.probe_timestamp; 184405cbc0dbSFan Du if (unlikely(delta >= interval * HZ)) { 184505cbc0dbSFan Du int mss = tcp_current_mss(sk); 184605cbc0dbSFan Du 184705cbc0dbSFan Du /* Update current search range */ 184805cbc0dbSFan Du icsk->icsk_mtup.probe_size = 0; 184905cbc0dbSFan Du icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + 185005cbc0dbSFan Du sizeof(struct tcphdr) + 185105cbc0dbSFan Du icsk->icsk_af_ops->net_header_len; 185205cbc0dbSFan Du icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss); 185305cbc0dbSFan Du 185405cbc0dbSFan Du /* Update probe time stamp */ 185505cbc0dbSFan Du icsk->icsk_mtup.probe_timestamp = tcp_time_stamp; 185605cbc0dbSFan Du } 185705cbc0dbSFan Du } 185805cbc0dbSFan Du 18595d424d5aSJohn Heffner /* Create a new MTU probe if we are ready. 186067edfef7SAndi Kleen * MTU probe is regularly attempting to increase the path MTU by 186167edfef7SAndi Kleen * deliberately sending larger packets. This discovers routing 186267edfef7SAndi Kleen * changes resulting in larger path MTUs. 186367edfef7SAndi Kleen * 18645d424d5aSJohn Heffner * Returns 0 if we should wait to probe (no cwnd available), 18655d424d5aSJohn Heffner * 1 if a probe was sent, 1866056834d9SIlpo Järvinen * -1 otherwise 1867056834d9SIlpo Järvinen */ 18685d424d5aSJohn Heffner static int tcp_mtu_probe(struct sock *sk) 18695d424d5aSJohn Heffner { 18705d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 18715d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 18725d424d5aSJohn Heffner struct sk_buff *skb, *nskb, *next; 18736b58e0a5SFan Du struct net *net = sock_net(sk); 18745d424d5aSJohn Heffner int len; 18755d424d5aSJohn Heffner int probe_size; 187691cc17c0SIlpo Järvinen int size_needed; 18775d424d5aSJohn Heffner int copy; 18785d424d5aSJohn Heffner int mss_now; 18796b58e0a5SFan Du int interval; 18805d424d5aSJohn Heffner 18815d424d5aSJohn Heffner /* Not currently probing/verifying, 18825d424d5aSJohn Heffner * not in recovery, 18835d424d5aSJohn Heffner * have enough cwnd, and 18845d424d5aSJohn Heffner * not SACKing (the variable headers throw things off) */ 18855d424d5aSJohn Heffner if (!icsk->icsk_mtup.enabled || 18865d424d5aSJohn Heffner icsk->icsk_mtup.probe_size || 18875d424d5aSJohn Heffner inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 18885d424d5aSJohn Heffner tp->snd_cwnd < 11 || 1889cabeccbdSIlpo Järvinen tp->rx_opt.num_sacks || tp->rx_opt.dsack) 18905d424d5aSJohn Heffner return -1; 18915d424d5aSJohn Heffner 18926b58e0a5SFan Du /* Use binary search for probe_size between tcp_mss_base, 18936b58e0a5SFan Du * and current mss_clamp. if (search_high - search_low) 18946b58e0a5SFan Du * smaller than a threshold, backoff from probing. 18956b58e0a5SFan Du */ 18960c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 18976b58e0a5SFan Du probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high + 18986b58e0a5SFan Du icsk->icsk_mtup.search_low) >> 1); 189991cc17c0SIlpo Järvinen size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; 19006b58e0a5SFan Du interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low; 190105cbc0dbSFan Du /* When misfortune happens, we are reprobing actively, 190205cbc0dbSFan Du * and then reprobe timer has expired. We stick with current 190305cbc0dbSFan Du * probing process by not resetting search range to its orignal. 190405cbc0dbSFan Du */ 19056b58e0a5SFan Du if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) || 190605cbc0dbSFan Du interval < net->ipv4.sysctl_tcp_probe_threshold) { 190705cbc0dbSFan Du /* Check whether enough time has elaplased for 190805cbc0dbSFan Du * another round of probing. 190905cbc0dbSFan Du */ 191005cbc0dbSFan Du tcp_mtu_check_reprobe(sk); 19115d424d5aSJohn Heffner return -1; 19125d424d5aSJohn Heffner } 19135d424d5aSJohn Heffner 19145d424d5aSJohn Heffner /* Have enough data in the send queue to probe? */ 19157f9c33e5SIlpo Järvinen if (tp->write_seq - tp->snd_nxt < size_needed) 19165d424d5aSJohn Heffner return -1; 19175d424d5aSJohn Heffner 191891cc17c0SIlpo Järvinen if (tp->snd_wnd < size_needed) 19195d424d5aSJohn Heffner return -1; 192090840defSIlpo Järvinen if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) 19215d424d5aSJohn Heffner return 0; 19225d424d5aSJohn Heffner 1923d67c58e9SIlpo Järvinen /* Do we need to wait to drain cwnd? With none in flight, don't stall */ 1924d67c58e9SIlpo Järvinen if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { 1925d67c58e9SIlpo Järvinen if (!tcp_packets_in_flight(tp)) 19265d424d5aSJohn Heffner return -1; 19275d424d5aSJohn Heffner else 19285d424d5aSJohn Heffner return 0; 19295d424d5aSJohn Heffner } 19305d424d5aSJohn Heffner 19315d424d5aSJohn Heffner /* We're allowed to probe. Build it now. */ 1932eb934478SEric Dumazet nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false); 193351456b29SIan Morris if (!nskb) 19345d424d5aSJohn Heffner return -1; 19353ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 19363ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 19375d424d5aSJohn Heffner 1938fe067e8aSDavid S. Miller skb = tcp_send_head(sk); 19395d424d5aSJohn Heffner 19405d424d5aSJohn Heffner TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 19415d424d5aSJohn Heffner TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 19424de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; 19435d424d5aSJohn Heffner TCP_SKB_CB(nskb)->sacked = 0; 19445d424d5aSJohn Heffner nskb->csum = 0; 194584fa7933SPatrick McHardy nskb->ip_summed = skb->ip_summed; 19465d424d5aSJohn Heffner 194750c4817eSIlpo Järvinen tcp_insert_write_queue_before(nskb, skb, sk); 194850c4817eSIlpo Järvinen 19495d424d5aSJohn Heffner len = 0; 1950234b6860SIlpo Järvinen tcp_for_write_queue_from_safe(skb, next, sk) { 19515d424d5aSJohn Heffner copy = min_t(int, skb->len, probe_size - len); 19525d424d5aSJohn Heffner if (nskb->ip_summed) 19535d424d5aSJohn Heffner skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); 19545d424d5aSJohn Heffner else 19555d424d5aSJohn Heffner nskb->csum = skb_copy_and_csum_bits(skb, 0, 1956056834d9SIlpo Järvinen skb_put(nskb, copy), 1957056834d9SIlpo Järvinen copy, nskb->csum); 19585d424d5aSJohn Heffner 19595d424d5aSJohn Heffner if (skb->len <= copy) { 19605d424d5aSJohn Heffner /* We've eaten all the data from this skb. 19615d424d5aSJohn Heffner * Throw it away. */ 19624de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 1963fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 19643ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 19655d424d5aSJohn Heffner } else { 19664de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & 1967a3433f35SChangli Gao ~(TCPHDR_FIN|TCPHDR_PSH); 19685d424d5aSJohn Heffner if (!skb_shinfo(skb)->nr_frags) { 19695d424d5aSJohn Heffner skb_pull(skb, copy); 197084fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 1971056834d9SIlpo Järvinen skb->csum = csum_partial(skb->data, 1972056834d9SIlpo Järvinen skb->len, 0); 19735d424d5aSJohn Heffner } else { 19745d424d5aSJohn Heffner __pskb_trim_head(skb, copy); 19755bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 19765d424d5aSJohn Heffner } 19775d424d5aSJohn Heffner TCP_SKB_CB(skb)->seq += copy; 19785d424d5aSJohn Heffner } 19795d424d5aSJohn Heffner 19805d424d5aSJohn Heffner len += copy; 1981234b6860SIlpo Järvinen 1982234b6860SIlpo Järvinen if (len >= probe_size) 1983234b6860SIlpo Järvinen break; 19845d424d5aSJohn Heffner } 19855bbb432cSEric Dumazet tcp_init_tso_segs(nskb, nskb->len); 19865d424d5aSJohn Heffner 19875d424d5aSJohn Heffner /* We're ready to send. If this fails, the probe will 19887faee5c0SEric Dumazet * be resegmented into mss-sized pieces by tcp_write_xmit(). 19897faee5c0SEric Dumazet */ 19905d424d5aSJohn Heffner if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 19915d424d5aSJohn Heffner /* Decrement cwnd here because we are sending 19925d424d5aSJohn Heffner * effectively two packets. */ 19935d424d5aSJohn Heffner tp->snd_cwnd--; 199466f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, nskb); 19955d424d5aSJohn Heffner 19965d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 19970e7b1368SJohn Heffner tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 19980e7b1368SJohn Heffner tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 19995d424d5aSJohn Heffner 20005d424d5aSJohn Heffner return 1; 20015d424d5aSJohn Heffner } 20025d424d5aSJohn Heffner 20035d424d5aSJohn Heffner return -1; 20045d424d5aSJohn Heffner } 20055d424d5aSJohn Heffner 20061da177e4SLinus Torvalds /* This routine writes packets to the network. It advances the 20071da177e4SLinus Torvalds * send_head. This happens as incoming acks open up the remote 20081da177e4SLinus Torvalds * window for us. 20091da177e4SLinus Torvalds * 2010f8269a49SIlpo Järvinen * LARGESEND note: !tcp_urg_mode is overkill, only frames between 2011f8269a49SIlpo Järvinen * snd_up-64k-mss .. snd_up cannot be large. However, taking into 2012f8269a49SIlpo Järvinen * account rare use of URG, this is not a big flaw. 2013f8269a49SIlpo Järvinen * 20146ba8a3b1SNandita Dukkipati * Send at most one packet when push_one > 0. Temporarily ignore 20156ba8a3b1SNandita Dukkipati * cwnd limit to force at most one packet out when push_one == 2. 20166ba8a3b1SNandita Dukkipati 2017a2a385d6SEric Dumazet * Returns true, if no segments are in flight and we have queued segments, 2018a2a385d6SEric Dumazet * but cannot send anything now because of SWS or another problem. 20191da177e4SLinus Torvalds */ 2020a2a385d6SEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 2021d5dd9175SIlpo Järvinen int push_one, gfp_t gfp) 20221da177e4SLinus Torvalds { 20231da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 202492df7b51SDavid S. Miller struct sk_buff *skb; 2025c1b4a7e6SDavid S. Miller unsigned int tso_segs, sent_pkts; 2026c1b4a7e6SDavid S. Miller int cwnd_quota; 20275d424d5aSJohn Heffner int result; 2028ca8a2263SNeal Cardwell bool is_cwnd_limited = false; 2029605ad7f1SEric Dumazet u32 max_segs; 20301da177e4SLinus Torvalds 2031c1b4a7e6SDavid S. Miller sent_pkts = 0; 20325d424d5aSJohn Heffner 2033d5dd9175SIlpo Järvinen if (!push_one) { 20345d424d5aSJohn Heffner /* Do MTU probing. */ 2035d5dd9175SIlpo Järvinen result = tcp_mtu_probe(sk); 2036d5dd9175SIlpo Järvinen if (!result) { 2037a2a385d6SEric Dumazet return false; 20385d424d5aSJohn Heffner } else if (result > 0) { 20395d424d5aSJohn Heffner sent_pkts = 1; 20405d424d5aSJohn Heffner } 2041d5dd9175SIlpo Järvinen } 20425d424d5aSJohn Heffner 2043605ad7f1SEric Dumazet max_segs = tcp_tso_autosize(sk, mss_now); 2044fe067e8aSDavid S. Miller while ((skb = tcp_send_head(sk))) { 2045c8ac3774SHerbert Xu unsigned int limit; 2046c8ac3774SHerbert Xu 20475bbb432cSEric Dumazet tso_segs = tcp_init_tso_segs(skb, mss_now); 2048c1b4a7e6SDavid S. Miller BUG_ON(!tso_segs); 2049c1b4a7e6SDavid S. Miller 20509d186cacSAndrey Vagin if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) { 20517faee5c0SEric Dumazet /* "skb_mstamp" is used as a start point for the retransmit timer */ 20527faee5c0SEric Dumazet skb_mstamp_get(&skb->skb_mstamp); 2053ec342325SAndrew Vagin goto repair; /* Skip network transmission */ 20549d186cacSAndrey Vagin } 2055ec342325SAndrew Vagin 2056b68e9f85SHerbert Xu cwnd_quota = tcp_cwnd_test(tp, skb); 20576ba8a3b1SNandita Dukkipati if (!cwnd_quota) { 20586ba8a3b1SNandita Dukkipati if (push_one == 2) 20596ba8a3b1SNandita Dukkipati /* Force out a loss probe pkt. */ 20606ba8a3b1SNandita Dukkipati cwnd_quota = 1; 20616ba8a3b1SNandita Dukkipati else 2062b68e9f85SHerbert Xu break; 20636ba8a3b1SNandita Dukkipati } 2064b68e9f85SHerbert Xu 2065b68e9f85SHerbert Xu if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) 2066b68e9f85SHerbert Xu break; 2067b68e9f85SHerbert Xu 2068d6a4e26aSEric Dumazet if (tso_segs == 1) { 2069aa93466bSDavid S. Miller if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 2070aa93466bSDavid S. Miller (tcp_skb_is_last(sk, skb) ? 2071aa93466bSDavid S. Miller nonagle : TCP_NAGLE_PUSH)))) 2072aa93466bSDavid S. Miller break; 2073c1b4a7e6SDavid S. Miller } else { 2074ca8a2263SNeal Cardwell if (!push_one && 2075605ad7f1SEric Dumazet tcp_tso_should_defer(sk, skb, &is_cwnd_limited, 2076605ad7f1SEric Dumazet max_segs)) 2077aa93466bSDavid S. Miller break; 2078c1b4a7e6SDavid S. Miller } 2079aa93466bSDavid S. Miller 2080605ad7f1SEric Dumazet limit = mss_now; 2081d6a4e26aSEric Dumazet if (tso_segs > 1 && !tcp_urg_mode(tp)) 2082605ad7f1SEric Dumazet limit = tcp_mss_split_point(sk, skb, mss_now, 2083605ad7f1SEric Dumazet min_t(unsigned int, 2084605ad7f1SEric Dumazet cwnd_quota, 2085605ad7f1SEric Dumazet max_segs), 2086605ad7f1SEric Dumazet nonagle); 2087605ad7f1SEric Dumazet 2088605ad7f1SEric Dumazet if (skb->len > limit && 2089605ad7f1SEric Dumazet unlikely(tso_fragment(sk, skb, limit, mss_now, gfp))) 2090605ad7f1SEric Dumazet break; 2091605ad7f1SEric Dumazet 2092c9eeec26SEric Dumazet /* TCP Small Queues : 2093c9eeec26SEric Dumazet * Control number of packets in qdisc/devices to two packets / or ~1 ms. 2094c9eeec26SEric Dumazet * This allows for : 2095c9eeec26SEric Dumazet * - better RTT estimation and ACK scheduling 2096c9eeec26SEric Dumazet * - faster recovery 2097c9eeec26SEric Dumazet * - high rates 209898e09386SEric Dumazet * Alas, some drivers / subsystems require a fair amount 209998e09386SEric Dumazet * of queued bytes to ensure line rate. 210098e09386SEric Dumazet * One example is wifi aggregation (802.11 AMPDU) 210146d3ceabSEric Dumazet */ 2102605ad7f1SEric Dumazet limit = max(2 * skb->truesize, sk->sk_pacing_rate >> 10); 2103605ad7f1SEric Dumazet limit = min_t(u32, limit, sysctl_tcp_limit_output_bytes); 2104c9eeec26SEric Dumazet 2105c9eeec26SEric Dumazet if (atomic_read(&sk->sk_wmem_alloc) > limit) { 210646d3ceabSEric Dumazet set_bit(TSQ_THROTTLED, &tp->tsq_flags); 2107bf06200eSJohn Ogness /* It is possible TX completion already happened 2108bf06200eSJohn Ogness * before we set TSQ_THROTTLED, so we must 2109bf06200eSJohn Ogness * test again the condition. 2110bf06200eSJohn Ogness */ 21114e857c58SPeter Zijlstra smp_mb__after_atomic(); 2112bf06200eSJohn Ogness if (atomic_read(&sk->sk_wmem_alloc) > limit) 211346d3ceabSEric Dumazet break; 211446d3ceabSEric Dumazet } 2115c9eeec26SEric Dumazet 2116d5dd9175SIlpo Järvinen if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) 21171da177e4SLinus Torvalds break; 21181da177e4SLinus Torvalds 2119ec342325SAndrew Vagin repair: 21201da177e4SLinus Torvalds /* Advance the send_head. This one is sent out. 21211da177e4SLinus Torvalds * This call will increment packets_out. 21221da177e4SLinus Torvalds */ 212366f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 21241da177e4SLinus Torvalds 21251da177e4SLinus Torvalds tcp_minshall_update(tp, mss_now, skb); 2126a262f0cdSNandita Dukkipati sent_pkts += tcp_skb_pcount(skb); 2127d5dd9175SIlpo Järvinen 2128d5dd9175SIlpo Järvinen if (push_one) 2129d5dd9175SIlpo Järvinen break; 21301da177e4SLinus Torvalds } 21311da177e4SLinus Torvalds 2132aa93466bSDavid S. Miller if (likely(sent_pkts)) { 2133684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 2134684bad11SYuchung Cheng tp->prr_out += sent_pkts; 21356ba8a3b1SNandita Dukkipati 21366ba8a3b1SNandita Dukkipati /* Send one loss probe per tail loss episode. */ 21376ba8a3b1SNandita Dukkipati if (push_one != 2) 21386ba8a3b1SNandita Dukkipati tcp_schedule_loss_probe(sk); 2139d2e1339fSBendik Rønning Opstad is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd); 2140ca8a2263SNeal Cardwell tcp_cwnd_validate(sk, is_cwnd_limited); 2141a2a385d6SEric Dumazet return false; 21421da177e4SLinus Torvalds } 2143b340b264SYuchung Cheng return !tp->packets_out && tcp_send_head(sk); 21446ba8a3b1SNandita Dukkipati } 21456ba8a3b1SNandita Dukkipati 21466ba8a3b1SNandita Dukkipati bool tcp_schedule_loss_probe(struct sock *sk) 21476ba8a3b1SNandita Dukkipati { 21486ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 21496ba8a3b1SNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 21506ba8a3b1SNandita Dukkipati u32 timeout, tlp_time_stamp, rto_time_stamp; 2151740b0f18SEric Dumazet u32 rtt = usecs_to_jiffies(tp->srtt_us >> 3); 21526ba8a3b1SNandita Dukkipati 21536ba8a3b1SNandita Dukkipati if (WARN_ON(icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS)) 21546ba8a3b1SNandita Dukkipati return false; 21556ba8a3b1SNandita Dukkipati /* No consecutive loss probes. */ 21566ba8a3b1SNandita Dukkipati if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) { 21576ba8a3b1SNandita Dukkipati tcp_rearm_rto(sk); 21586ba8a3b1SNandita Dukkipati return false; 21596ba8a3b1SNandita Dukkipati } 21606ba8a3b1SNandita Dukkipati /* Don't do any loss probe on a Fast Open connection before 3WHS 21616ba8a3b1SNandita Dukkipati * finishes. 21626ba8a3b1SNandita Dukkipati */ 2163f9b99582SYuchung Cheng if (tp->fastopen_rsk) 21646ba8a3b1SNandita Dukkipati return false; 21656ba8a3b1SNandita Dukkipati 21666ba8a3b1SNandita Dukkipati /* TLP is only scheduled when next timer event is RTO. */ 21676ba8a3b1SNandita Dukkipati if (icsk->icsk_pending != ICSK_TIME_RETRANS) 21686ba8a3b1SNandita Dukkipati return false; 21696ba8a3b1SNandita Dukkipati 21706ba8a3b1SNandita Dukkipati /* Schedule a loss probe in 2*RTT for SACK capable connections 21716ba8a3b1SNandita Dukkipati * in Open state, that are either limited by cwnd or application. 21726ba8a3b1SNandita Dukkipati */ 2173f9b99582SYuchung Cheng if (sysctl_tcp_early_retrans < 3 || !tp->packets_out || 21746ba8a3b1SNandita Dukkipati !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open) 21756ba8a3b1SNandita Dukkipati return false; 21766ba8a3b1SNandita Dukkipati 21776ba8a3b1SNandita Dukkipati if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) && 21786ba8a3b1SNandita Dukkipati tcp_send_head(sk)) 21796ba8a3b1SNandita Dukkipati return false; 21806ba8a3b1SNandita Dukkipati 21816ba8a3b1SNandita Dukkipati /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account 2182f9b99582SYuchung Cheng * for delayed ack when there's one outstanding packet. If no RTT 2183f9b99582SYuchung Cheng * sample is available then probe after TCP_TIMEOUT_INIT. 21846ba8a3b1SNandita Dukkipati */ 2185f9b99582SYuchung Cheng timeout = rtt << 1 ? : TCP_TIMEOUT_INIT; 21866ba8a3b1SNandita Dukkipati if (tp->packets_out == 1) 21876ba8a3b1SNandita Dukkipati timeout = max_t(u32, timeout, 21886ba8a3b1SNandita Dukkipati (rtt + (rtt >> 1) + TCP_DELACK_MAX)); 21896ba8a3b1SNandita Dukkipati timeout = max_t(u32, timeout, msecs_to_jiffies(10)); 21906ba8a3b1SNandita Dukkipati 21916ba8a3b1SNandita Dukkipati /* If RTO is shorter, just schedule TLP in its place. */ 21926ba8a3b1SNandita Dukkipati tlp_time_stamp = tcp_time_stamp + timeout; 21936ba8a3b1SNandita Dukkipati rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout; 21946ba8a3b1SNandita Dukkipati if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) { 21956ba8a3b1SNandita Dukkipati s32 delta = rto_time_stamp - tcp_time_stamp; 21966ba8a3b1SNandita Dukkipati if (delta > 0) 21976ba8a3b1SNandita Dukkipati timeout = delta; 21986ba8a3b1SNandita Dukkipati } 21996ba8a3b1SNandita Dukkipati 22006ba8a3b1SNandita Dukkipati inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, 22016ba8a3b1SNandita Dukkipati TCP_RTO_MAX); 22026ba8a3b1SNandita Dukkipati return true; 22036ba8a3b1SNandita Dukkipati } 22046ba8a3b1SNandita Dukkipati 22051f3279aeSEric Dumazet /* Thanks to skb fast clones, we can detect if a prior transmit of 22061f3279aeSEric Dumazet * a packet is still in a qdisc or driver queue. 22071f3279aeSEric Dumazet * In this case, there is very little point doing a retransmit ! 22081f3279aeSEric Dumazet * Note: This is called from BH context only. 22091f3279aeSEric Dumazet */ 22101f3279aeSEric Dumazet static bool skb_still_in_host_queue(const struct sock *sk, 22111f3279aeSEric Dumazet const struct sk_buff *skb) 22121f3279aeSEric Dumazet { 221339bb5e62SEric Dumazet if (unlikely(skb_fclone_busy(sk, skb))) { 22141f3279aeSEric Dumazet NET_INC_STATS_BH(sock_net(sk), 22151f3279aeSEric Dumazet LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES); 22161f3279aeSEric Dumazet return true; 22171f3279aeSEric Dumazet } 22181f3279aeSEric Dumazet return false; 22191f3279aeSEric Dumazet } 22201f3279aeSEric Dumazet 2221b340b264SYuchung Cheng /* When probe timeout (PTO) fires, try send a new segment if possible, else 22226ba8a3b1SNandita Dukkipati * retransmit the last segment. 22236ba8a3b1SNandita Dukkipati */ 22246ba8a3b1SNandita Dukkipati void tcp_send_loss_probe(struct sock *sk) 22256ba8a3b1SNandita Dukkipati { 22269b717a8dSNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 22276ba8a3b1SNandita Dukkipati struct sk_buff *skb; 22286ba8a3b1SNandita Dukkipati int pcount; 22296ba8a3b1SNandita Dukkipati int mss = tcp_current_mss(sk); 22306ba8a3b1SNandita Dukkipati 2231b340b264SYuchung Cheng skb = tcp_send_head(sk); 2232b340b264SYuchung Cheng if (skb) { 2233b340b264SYuchung Cheng if (tcp_snd_wnd_test(tp, skb, mss)) { 2234b340b264SYuchung Cheng pcount = tp->packets_out; 2235b340b264SYuchung Cheng tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC); 2236b340b264SYuchung Cheng if (tp->packets_out > pcount) 2237b340b264SYuchung Cheng goto probe_sent; 22386ba8a3b1SNandita Dukkipati goto rearm_timer; 22396ba8a3b1SNandita Dukkipati } 2240b340b264SYuchung Cheng skb = tcp_write_queue_prev(sk, skb); 2241b340b264SYuchung Cheng } else { 2242b340b264SYuchung Cheng skb = tcp_write_queue_tail(sk); 2243b340b264SYuchung Cheng } 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 if (WARN_ON(!skb)) 22516ba8a3b1SNandita Dukkipati goto rearm_timer; 22526ba8a3b1SNandita Dukkipati 22531f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 22541f3279aeSEric Dumazet goto rearm_timer; 22551f3279aeSEric Dumazet 22566ba8a3b1SNandita Dukkipati pcount = tcp_skb_pcount(skb); 22576ba8a3b1SNandita Dukkipati if (WARN_ON(!pcount)) 22586ba8a3b1SNandita Dukkipati goto rearm_timer; 22596ba8a3b1SNandita Dukkipati 22606ba8a3b1SNandita Dukkipati if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) { 22616cc55e09SOctavian Purdila if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss, 22626cc55e09SOctavian Purdila GFP_ATOMIC))) 22636ba8a3b1SNandita Dukkipati goto rearm_timer; 2264b340b264SYuchung Cheng skb = tcp_write_queue_next(sk, skb); 22656ba8a3b1SNandita Dukkipati } 22666ba8a3b1SNandita Dukkipati 22676ba8a3b1SNandita Dukkipati if (WARN_ON(!skb || !tcp_skb_pcount(skb))) 22686ba8a3b1SNandita Dukkipati goto rearm_timer; 22696ba8a3b1SNandita Dukkipati 2270b340b264SYuchung Cheng if (__tcp_retransmit_skb(sk, skb)) 2271b340b264SYuchung Cheng goto rearm_timer; 22726ba8a3b1SNandita Dukkipati 22739b717a8dSNandita Dukkipati /* Record snd_nxt for loss detection. */ 22749b717a8dSNandita Dukkipati tp->tlp_high_seq = tp->snd_nxt; 22759b717a8dSNandita Dukkipati 2276b340b264SYuchung Cheng probe_sent: 2277fcd16c0aSYuchung Cheng NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSPROBES); 2278fcd16c0aSYuchung Cheng /* Reset s.t. tcp_rearm_rto will restart timer from now */ 2279fcd16c0aSYuchung Cheng inet_csk(sk)->icsk_pending = 0; 2280b340b264SYuchung Cheng rearm_timer: 2281fcd16c0aSYuchung Cheng tcp_rearm_rto(sk); 22821da177e4SLinus Torvalds } 22831da177e4SLinus Torvalds 2284a762a980SDavid S. Miller /* Push out any pending frames which were held back due to 2285a762a980SDavid S. Miller * TCP_CORK or attempt at coalescing tiny packets. 2286a762a980SDavid S. Miller * The socket must be locked by the caller. 2287a762a980SDavid S. Miller */ 22889e412ba7SIlpo Järvinen void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 22899e412ba7SIlpo Järvinen int nonagle) 2290a762a980SDavid S. Miller { 2291726e07a8SIlpo Järvinen /* If we are closed, the bytes will have to remain here. 2292726e07a8SIlpo Järvinen * In time closedown will finish, we empty the write queue and 2293726e07a8SIlpo Järvinen * all will be happy. 2294726e07a8SIlpo Järvinen */ 2295726e07a8SIlpo Järvinen if (unlikely(sk->sk_state == TCP_CLOSE)) 2296726e07a8SIlpo Järvinen return; 2297726e07a8SIlpo Järvinen 229899a1dec7SMel Gorman if (tcp_write_xmit(sk, cur_mss, nonagle, 0, 22997450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC))) 23009e412ba7SIlpo Järvinen tcp_check_probe_timer(sk); 2301a762a980SDavid S. Miller } 2302a762a980SDavid S. Miller 2303c1b4a7e6SDavid S. Miller /* Send _single_ skb sitting at the send head. This function requires 2304c1b4a7e6SDavid S. Miller * true push pending frames to setup probe timer etc. 2305c1b4a7e6SDavid S. Miller */ 2306c1b4a7e6SDavid S. Miller void tcp_push_one(struct sock *sk, unsigned int mss_now) 2307c1b4a7e6SDavid S. Miller { 2308fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 2309c1b4a7e6SDavid S. Miller 2310c1b4a7e6SDavid S. Miller BUG_ON(!skb || skb->len < mss_now); 2311c1b4a7e6SDavid S. Miller 2312d5dd9175SIlpo Järvinen tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); 2313c1b4a7e6SDavid S. Miller } 2314c1b4a7e6SDavid S. Miller 23151da177e4SLinus Torvalds /* This function returns the amount that we can raise the 23161da177e4SLinus Torvalds * usable window based on the following constraints 23171da177e4SLinus Torvalds * 23181da177e4SLinus Torvalds * 1. The window can never be shrunk once it is offered (RFC 793) 23191da177e4SLinus Torvalds * 2. We limit memory per socket 23201da177e4SLinus Torvalds * 23211da177e4SLinus Torvalds * RFC 1122: 23221da177e4SLinus Torvalds * "the suggested [SWS] avoidance algorithm for the receiver is to keep 23231da177e4SLinus Torvalds * RECV.NEXT + RCV.WIN fixed until: 23241da177e4SLinus Torvalds * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 23251da177e4SLinus Torvalds * 23261da177e4SLinus Torvalds * i.e. don't raise the right edge of the window until you can raise 23271da177e4SLinus Torvalds * it at least MSS bytes. 23281da177e4SLinus Torvalds * 23291da177e4SLinus Torvalds * Unfortunately, the recommended algorithm breaks header prediction, 23301da177e4SLinus Torvalds * since header prediction assumes th->window stays fixed. 23311da177e4SLinus Torvalds * 23321da177e4SLinus Torvalds * Strictly speaking, keeping th->window fixed violates the receiver 23331da177e4SLinus Torvalds * side SWS prevention criteria. The problem is that under this rule 23341da177e4SLinus Torvalds * a stream of single byte packets will cause the right side of the 23351da177e4SLinus Torvalds * window to always advance by a single byte. 23361da177e4SLinus Torvalds * 23371da177e4SLinus Torvalds * Of course, if the sender implements sender side SWS prevention 23381da177e4SLinus Torvalds * then this will not be a problem. 23391da177e4SLinus Torvalds * 23401da177e4SLinus Torvalds * BSD seems to make the following compromise: 23411da177e4SLinus Torvalds * 23421da177e4SLinus Torvalds * If the free space is less than the 1/4 of the maximum 23431da177e4SLinus Torvalds * space available and the free space is less than 1/2 mss, 23441da177e4SLinus Torvalds * then set the window to 0. 23451da177e4SLinus Torvalds * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 23461da177e4SLinus Torvalds * Otherwise, just prevent the window from shrinking 23471da177e4SLinus Torvalds * and from being larger than the largest representable value. 23481da177e4SLinus Torvalds * 23491da177e4SLinus Torvalds * This prevents incremental opening of the window in the regime 23501da177e4SLinus Torvalds * where TCP is limited by the speed of the reader side taking 23511da177e4SLinus Torvalds * data out of the TCP receive queue. It does nothing about 23521da177e4SLinus Torvalds * those cases where the window is constrained on the sender side 23531da177e4SLinus Torvalds * because the pipeline is full. 23541da177e4SLinus Torvalds * 23551da177e4SLinus Torvalds * BSD also seems to "accidentally" limit itself to windows that are a 23561da177e4SLinus Torvalds * multiple of MSS, at least until the free space gets quite small. 23571da177e4SLinus Torvalds * This would appear to be a side effect of the mbuf implementation. 23581da177e4SLinus Torvalds * Combining these two algorithms results in the observed behavior 23591da177e4SLinus Torvalds * of having a fixed window size at almost all times. 23601da177e4SLinus Torvalds * 23611da177e4SLinus Torvalds * Below we obtain similar behavior by forcing the offered window to 23621da177e4SLinus Torvalds * a multiple of the mss when it is feasible to do so. 23631da177e4SLinus Torvalds * 23641da177e4SLinus Torvalds * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 23651da177e4SLinus Torvalds * Regular options like TIMESTAMP are taken into account. 23661da177e4SLinus Torvalds */ 23671da177e4SLinus Torvalds u32 __tcp_select_window(struct sock *sk) 23681da177e4SLinus Torvalds { 2369463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 23701da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2371caa20d9aSStephen Hemminger /* MSS for the peer's data. Previous versions used mss_clamp 23721da177e4SLinus Torvalds * here. I don't know if the value based on our guesses 23731da177e4SLinus Torvalds * of peer's MSS is better for the performance. It's more correct 23741da177e4SLinus Torvalds * but may be worse for the performance because of rcv_mss 23751da177e4SLinus Torvalds * fluctuations. --SAW 1998/11/1 23761da177e4SLinus Torvalds */ 2377463c84b9SArnaldo Carvalho de Melo int mss = icsk->icsk_ack.rcv_mss; 23781da177e4SLinus Torvalds int free_space = tcp_space(sk); 237986c1a045SFlorian Westphal int allowed_space = tcp_full_space(sk); 238086c1a045SFlorian Westphal int full_space = min_t(int, tp->window_clamp, allowed_space); 23811da177e4SLinus Torvalds int window; 23821da177e4SLinus Torvalds 23831da177e4SLinus Torvalds if (mss > full_space) 23841da177e4SLinus Torvalds mss = full_space; 23851da177e4SLinus Torvalds 2386b92edbe0SEric Dumazet if (free_space < (full_space >> 1)) { 2387463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.quick = 0; 23881da177e4SLinus Torvalds 2389b8da51ebSEric Dumazet if (tcp_under_memory_pressure(sk)) 2390056834d9SIlpo Järvinen tp->rcv_ssthresh = min(tp->rcv_ssthresh, 2391056834d9SIlpo Järvinen 4U * tp->advmss); 23921da177e4SLinus Torvalds 239386c1a045SFlorian Westphal /* free_space might become our new window, make sure we don't 239486c1a045SFlorian Westphal * increase it due to wscale. 239586c1a045SFlorian Westphal */ 239686c1a045SFlorian Westphal free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale); 239786c1a045SFlorian Westphal 239886c1a045SFlorian Westphal /* if free space is less than mss estimate, or is below 1/16th 239986c1a045SFlorian Westphal * of the maximum allowed, try to move to zero-window, else 240086c1a045SFlorian Westphal * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and 240186c1a045SFlorian Westphal * new incoming data is dropped due to memory limits. 240286c1a045SFlorian Westphal * With large window, mss test triggers way too late in order 240386c1a045SFlorian Westphal * to announce zero window in time before rmem limit kicks in. 240486c1a045SFlorian Westphal */ 240586c1a045SFlorian Westphal if (free_space < (allowed_space >> 4) || free_space < mss) 24061da177e4SLinus Torvalds return 0; 24071da177e4SLinus Torvalds } 24081da177e4SLinus Torvalds 24091da177e4SLinus Torvalds if (free_space > tp->rcv_ssthresh) 24101da177e4SLinus Torvalds free_space = tp->rcv_ssthresh; 24111da177e4SLinus Torvalds 24121da177e4SLinus Torvalds /* Don't do rounding if we are using window scaling, since the 24131da177e4SLinus Torvalds * scaled window will not line up with the MSS boundary anyway. 24141da177e4SLinus Torvalds */ 24151da177e4SLinus Torvalds window = tp->rcv_wnd; 24161da177e4SLinus Torvalds if (tp->rx_opt.rcv_wscale) { 24171da177e4SLinus Torvalds window = free_space; 24181da177e4SLinus Torvalds 24191da177e4SLinus Torvalds /* Advertise enough space so that it won't get scaled away. 24201da177e4SLinus Torvalds * Import case: prevent zero window announcement if 24211da177e4SLinus Torvalds * 1<<rcv_wscale > mss. 24221da177e4SLinus Torvalds */ 24231da177e4SLinus Torvalds if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window) 24241da177e4SLinus Torvalds window = (((window >> tp->rx_opt.rcv_wscale) + 1) 24251da177e4SLinus Torvalds << tp->rx_opt.rcv_wscale); 24261da177e4SLinus Torvalds } else { 24271da177e4SLinus Torvalds /* Get the largest window that is a nice multiple of mss. 24281da177e4SLinus Torvalds * Window clamp already applied above. 24291da177e4SLinus Torvalds * If our current window offering is within 1 mss of the 24301da177e4SLinus Torvalds * free space we just keep it. This prevents the divide 24311da177e4SLinus Torvalds * and multiply from happening most of the time. 24321da177e4SLinus Torvalds * We also don't do any window rounding when the free space 24331da177e4SLinus Torvalds * is too small. 24341da177e4SLinus Torvalds */ 24351da177e4SLinus Torvalds if (window <= free_space - mss || window > free_space) 24361da177e4SLinus Torvalds window = (free_space / mss) * mss; 243784565070SJohn Heffner else if (mss == full_space && 2438b92edbe0SEric Dumazet free_space > window + (full_space >> 1)) 243984565070SJohn Heffner window = free_space; 24401da177e4SLinus Torvalds } 24411da177e4SLinus Torvalds 24421da177e4SLinus Torvalds return window; 24431da177e4SLinus Torvalds } 24441da177e4SLinus Torvalds 24454a17fc3aSIlpo Järvinen /* Collapses two adjacent SKB's during retransmission. */ 24464a17fc3aSIlpo Järvinen static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) 24471da177e4SLinus Torvalds { 24481da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2449fe067e8aSDavid S. Miller struct sk_buff *next_skb = tcp_write_queue_next(sk, skb); 2450058dc334SIlpo Järvinen int skb_size, next_skb_size; 24511da177e4SLinus Torvalds 2452058dc334SIlpo Järvinen skb_size = skb->len; 2453058dc334SIlpo Järvinen next_skb_size = next_skb->len; 24541da177e4SLinus Torvalds 2455058dc334SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); 24561da177e4SLinus Torvalds 24576859d494SIlpo Järvinen tcp_highest_sack_combine(sk, next_skb, skb); 2458a6963a6bSIlpo Järvinen 2459fe067e8aSDavid S. Miller tcp_unlink_write_queue(next_skb, sk); 24601da177e4SLinus Torvalds 2461058dc334SIlpo Järvinen skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size), 24621a4e2d09SArnaldo Carvalho de Melo next_skb_size); 24631da177e4SLinus Torvalds 246452d570aaSJarek Poplawski if (next_skb->ip_summed == CHECKSUM_PARTIAL) 246552d570aaSJarek Poplawski skb->ip_summed = CHECKSUM_PARTIAL; 24661da177e4SLinus Torvalds 246784fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 24681da177e4SLinus Torvalds skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); 24691da177e4SLinus Torvalds 24701da177e4SLinus Torvalds /* Update sequence range on original skb. */ 24711da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 24721da177e4SLinus Torvalds 2473e6c7d085SIlpo Järvinen /* Merge over control information. This moves PSH/FIN etc. over */ 24744de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; 24751da177e4SLinus Torvalds 24761da177e4SLinus Torvalds /* All done, get rid of second SKB and account for it so 24771da177e4SLinus Torvalds * packet counting does not break. 24781da177e4SLinus Torvalds */ 24794828e7f4SIlpo Järvinen TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; 2480b7689205SIlpo Järvinen 2481b7689205SIlpo Järvinen /* changed transmit queue under us so clear hints */ 2482ef9da47cSIlpo Järvinen tcp_clear_retrans_hints_partial(tp); 2483ef9da47cSIlpo Järvinen if (next_skb == tp->retransmit_skb_hint) 2484ef9da47cSIlpo Järvinen tp->retransmit_skb_hint = skb; 2485b7689205SIlpo Järvinen 2486797108d1SIlpo Järvinen tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); 2487797108d1SIlpo Järvinen 24883ab224beSHideo Aoki sk_wmem_free_skb(sk, next_skb); 24891da177e4SLinus Torvalds } 24901da177e4SLinus Torvalds 249167edfef7SAndi Kleen /* Check if coalescing SKBs is legal. */ 2492a2a385d6SEric Dumazet static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) 24934a17fc3aSIlpo Järvinen { 24944a17fc3aSIlpo Järvinen if (tcp_skb_pcount(skb) > 1) 2495a2a385d6SEric Dumazet return false; 24964a17fc3aSIlpo Järvinen /* TODO: SACK collapsing could be used to remove this condition */ 24974a17fc3aSIlpo Järvinen if (skb_shinfo(skb)->nr_frags != 0) 2498a2a385d6SEric Dumazet return false; 24994a17fc3aSIlpo Järvinen if (skb_cloned(skb)) 2500a2a385d6SEric Dumazet return false; 25014a17fc3aSIlpo Järvinen if (skb == tcp_send_head(sk)) 2502a2a385d6SEric Dumazet return false; 25034a17fc3aSIlpo Järvinen /* Some heurestics for collapsing over SACK'd could be invented */ 25044a17fc3aSIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 2505a2a385d6SEric Dumazet return false; 25064a17fc3aSIlpo Järvinen 2507a2a385d6SEric Dumazet return true; 25084a17fc3aSIlpo Järvinen } 25094a17fc3aSIlpo Järvinen 251067edfef7SAndi Kleen /* Collapse packets in the retransmit queue to make to create 251167edfef7SAndi Kleen * less packets on the wire. This is only done on retransmission. 251267edfef7SAndi Kleen */ 25134a17fc3aSIlpo Järvinen static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, 25144a17fc3aSIlpo Järvinen int space) 25154a17fc3aSIlpo Järvinen { 25164a17fc3aSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 25174a17fc3aSIlpo Järvinen struct sk_buff *skb = to, *tmp; 2518a2a385d6SEric Dumazet bool first = true; 25194a17fc3aSIlpo Järvinen 25204a17fc3aSIlpo Järvinen if (!sysctl_tcp_retrans_collapse) 25214a17fc3aSIlpo Järvinen return; 25224de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 25234a17fc3aSIlpo Järvinen return; 25244a17fc3aSIlpo Järvinen 25254a17fc3aSIlpo Järvinen tcp_for_write_queue_from_safe(skb, tmp, sk) { 25264a17fc3aSIlpo Järvinen if (!tcp_can_collapse(sk, skb)) 25274a17fc3aSIlpo Järvinen break; 25284a17fc3aSIlpo Järvinen 25294a17fc3aSIlpo Järvinen space -= skb->len; 25304a17fc3aSIlpo Järvinen 25314a17fc3aSIlpo Järvinen if (first) { 2532a2a385d6SEric Dumazet first = false; 25334a17fc3aSIlpo Järvinen continue; 25344a17fc3aSIlpo Järvinen } 25354a17fc3aSIlpo Järvinen 25364a17fc3aSIlpo Järvinen if (space < 0) 25374a17fc3aSIlpo Järvinen break; 25384a17fc3aSIlpo Järvinen /* Punt if not enough space exists in the first SKB for 25394a17fc3aSIlpo Järvinen * the data in the second 25404a17fc3aSIlpo Järvinen */ 2541a21d4572SEric Dumazet if (skb->len > skb_availroom(to)) 25424a17fc3aSIlpo Järvinen break; 25434a17fc3aSIlpo Järvinen 25444a17fc3aSIlpo Järvinen if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) 25454a17fc3aSIlpo Järvinen break; 25464a17fc3aSIlpo Järvinen 25474a17fc3aSIlpo Järvinen tcp_collapse_retrans(sk, to); 25484a17fc3aSIlpo Järvinen } 25494a17fc3aSIlpo Järvinen } 25504a17fc3aSIlpo Järvinen 25511da177e4SLinus Torvalds /* This retransmits one SKB. Policy decisions and retransmit queue 25521da177e4SLinus Torvalds * state updates are done by the caller. Returns non-zero if an 25531da177e4SLinus Torvalds * error occurred which prevented the send. 25541da177e4SLinus Torvalds */ 255593b174adSYuchung Cheng int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) 25561da177e4SLinus Torvalds { 25571da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 25585d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 25597d227cd2SSridhar Samudrala unsigned int cur_mss; 2560c84a5711SYuchung Cheng int err; 25611da177e4SLinus Torvalds 25625d424d5aSJohn Heffner /* Inconslusive MTU probe */ 25635d424d5aSJohn Heffner if (icsk->icsk_mtup.probe_size) { 25645d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 25655d424d5aSJohn Heffner } 25665d424d5aSJohn Heffner 25671da177e4SLinus Torvalds /* Do not sent more than we queued. 1/4 is reserved for possible 2568caa20d9aSStephen Hemminger * copying overhead: fragmentation, tunneling, mangling etc. 25691da177e4SLinus Torvalds */ 25701da177e4SLinus Torvalds if (atomic_read(&sk->sk_wmem_alloc) > 25711da177e4SLinus Torvalds min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf)) 25721da177e4SLinus Torvalds return -EAGAIN; 25731da177e4SLinus Torvalds 25741f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 25751f3279aeSEric Dumazet return -EBUSY; 25761f3279aeSEric Dumazet 25771da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 25781da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 25791da177e4SLinus Torvalds BUG(); 25801da177e4SLinus Torvalds if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 25811da177e4SLinus Torvalds return -ENOMEM; 25821da177e4SLinus Torvalds } 25831da177e4SLinus Torvalds 25847d227cd2SSridhar Samudrala if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 25857d227cd2SSridhar Samudrala return -EHOSTUNREACH; /* Routing failure or similar. */ 25867d227cd2SSridhar Samudrala 25870c54b85fSIlpo Järvinen cur_mss = tcp_current_mss(sk); 25887d227cd2SSridhar Samudrala 25891da177e4SLinus Torvalds /* If receiver has shrunk his window, and skb is out of 25901da177e4SLinus Torvalds * new window, do not retransmit it. The exception is the 25911da177e4SLinus Torvalds * case, when window is shrunk to zero. In this case 25921da177e4SLinus Torvalds * our retransmit serves as a zero window probe. 25931da177e4SLinus Torvalds */ 25949d4fb27dSJoe Perches if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && 25959d4fb27dSJoe Perches TCP_SKB_CB(skb)->seq != tp->snd_una) 25961da177e4SLinus Torvalds return -EAGAIN; 25971da177e4SLinus Torvalds 25981da177e4SLinus Torvalds if (skb->len > cur_mss) { 25996cc55e09SOctavian Purdila if (tcp_fragment(sk, skb, cur_mss, cur_mss, GFP_ATOMIC)) 26001da177e4SLinus Torvalds return -ENOMEM; /* We'll try again later. */ 260102276f3cSIlpo Järvinen } else { 26029eb9362eSIlpo Järvinen int oldpcount = tcp_skb_pcount(skb); 26039eb9362eSIlpo Järvinen 26049eb9362eSIlpo Järvinen if (unlikely(oldpcount > 1)) { 2605c52e2421SEric Dumazet if (skb_unclone(skb, GFP_ATOMIC)) 2606c52e2421SEric Dumazet return -ENOMEM; 26075bbb432cSEric Dumazet tcp_init_tso_segs(skb, cur_mss); 26089eb9362eSIlpo Järvinen tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb)); 26099eb9362eSIlpo Järvinen } 26101da177e4SLinus Torvalds } 26111da177e4SLinus Torvalds 261249213555SDaniel Borkmann /* RFC3168, section 6.1.1.1. ECN fallback */ 261349213555SDaniel Borkmann if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN) 261449213555SDaniel Borkmann tcp_ecn_clear_syn(sk, skb); 261549213555SDaniel Borkmann 26161da177e4SLinus Torvalds tcp_retrans_try_collapse(sk, skb, cur_mss); 26171da177e4SLinus Torvalds 26181da177e4SLinus Torvalds /* Make a copy, if the first transmission SKB clone we made 26191da177e4SLinus Torvalds * is still in somebody's hands, else make a clone. 26201da177e4SLinus Torvalds */ 26211da177e4SLinus Torvalds 262250bceae9SThomas Graf /* make sure skb->data is aligned on arches that require it 262350bceae9SThomas Graf * and check if ack-trimming & collapsing extended the headroom 262450bceae9SThomas Graf * beyond what csum_start can cover. 262550bceae9SThomas Graf */ 262650bceae9SThomas Graf if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) || 262750bceae9SThomas Graf skb_headroom(skb) >= 0xFFFF)) { 2628117632e6SEric Dumazet struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER, 2629117632e6SEric Dumazet GFP_ATOMIC); 2630c84a5711SYuchung Cheng err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) : 2631117632e6SEric Dumazet -ENOBUFS; 2632117632e6SEric Dumazet } else { 2633c84a5711SYuchung Cheng err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 2634117632e6SEric Dumazet } 2635c84a5711SYuchung Cheng 2636fc9f3501SEric Dumazet if (likely(!err)) { 2637c84a5711SYuchung Cheng TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS; 2638fc9f3501SEric Dumazet /* Update global TCP statistics. */ 2639fc9f3501SEric Dumazet TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); 2640fc9f3501SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 2641fc9f3501SEric Dumazet NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 2642fc9f3501SEric Dumazet tp->total_retrans++; 2643fc9f3501SEric Dumazet } 2644c84a5711SYuchung Cheng return err; 264593b174adSYuchung Cheng } 264693b174adSYuchung Cheng 264793b174adSYuchung Cheng int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) 264893b174adSYuchung Cheng { 264993b174adSYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 265093b174adSYuchung Cheng int err = __tcp_retransmit_skb(sk, skb); 26511da177e4SLinus Torvalds 26521da177e4SLinus Torvalds if (err == 0) { 26531da177e4SLinus Torvalds #if FASTRETRANS_DEBUG > 0 26541da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 2655e87cc472SJoe Perches net_dbg_ratelimited("retrans_out leaked\n"); 26561da177e4SLinus Torvalds } 26571da177e4SLinus Torvalds #endif 26581da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 26591da177e4SLinus Torvalds tp->retrans_out += tcp_skb_pcount(skb); 26601da177e4SLinus Torvalds 26611da177e4SLinus Torvalds /* Save stamp of the first retransmit. */ 26621da177e4SLinus Torvalds if (!tp->retrans_stamp) 26637faee5c0SEric Dumazet tp->retrans_stamp = tcp_skb_timestamp(skb); 26641da177e4SLinus Torvalds 26651f3279aeSEric Dumazet } else if (err != -EBUSY) { 266624ab6becSYuchung Cheng NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL); 26671da177e4SLinus Torvalds } 26686e08d5e3SYuchung Cheng 26696e08d5e3SYuchung Cheng if (tp->undo_retrans < 0) 26706e08d5e3SYuchung Cheng tp->undo_retrans = 0; 26716e08d5e3SYuchung Cheng tp->undo_retrans += tcp_skb_pcount(skb); 26721da177e4SLinus Torvalds return err; 26731da177e4SLinus Torvalds } 26741da177e4SLinus Torvalds 267567edfef7SAndi Kleen /* Check if we forward retransmits are possible in the current 267667edfef7SAndi Kleen * window/congestion state. 267767edfef7SAndi Kleen */ 2678a2a385d6SEric Dumazet static bool tcp_can_forward_retransmit(struct sock *sk) 2679b5afe7bcSIlpo Järvinen { 2680b5afe7bcSIlpo Järvinen const struct inet_connection_sock *icsk = inet_csk(sk); 2681cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 2682b5afe7bcSIlpo Järvinen 2683b5afe7bcSIlpo Järvinen /* Forward retransmissions are possible only during Recovery. */ 2684b5afe7bcSIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Recovery) 2685a2a385d6SEric Dumazet return false; 2686b5afe7bcSIlpo Järvinen 2687b5afe7bcSIlpo Järvinen /* No forward retransmissions in Reno are possible. */ 2688b5afe7bcSIlpo Järvinen if (tcp_is_reno(tp)) 2689a2a385d6SEric Dumazet return false; 2690b5afe7bcSIlpo Järvinen 2691b5afe7bcSIlpo Järvinen /* Yeah, we have to make difficult choice between forward transmission 2692b5afe7bcSIlpo Järvinen * and retransmission... Both ways have their merits... 2693b5afe7bcSIlpo Järvinen * 2694b5afe7bcSIlpo Järvinen * For now we do not retransmit anything, while we have some new 2695b5afe7bcSIlpo Järvinen * segments to send. In the other cases, follow rule 3 for 2696b5afe7bcSIlpo Järvinen * NextSeg() specified in RFC3517. 2697b5afe7bcSIlpo Järvinen */ 2698b5afe7bcSIlpo Järvinen 2699b5afe7bcSIlpo Järvinen if (tcp_may_send_now(sk)) 2700a2a385d6SEric Dumazet return false; 2701b5afe7bcSIlpo Järvinen 2702a2a385d6SEric Dumazet return true; 2703b5afe7bcSIlpo Järvinen } 2704b5afe7bcSIlpo Järvinen 27051da177e4SLinus Torvalds /* This gets called after a retransmit timeout, and the initially 27061da177e4SLinus Torvalds * retransmitted data is acknowledged. It tries to continue 27071da177e4SLinus Torvalds * resending the rest of the retransmit queue, until either 27081da177e4SLinus Torvalds * we've sent it all or the congestion window limit is reached. 27091da177e4SLinus Torvalds * If doing SACK, the first ACK which comes back for a timeout 27101da177e4SLinus Torvalds * based retransmit packet might feed us FACK information again. 27111da177e4SLinus Torvalds * If so, we use it to avoid unnecessarily retransmissions. 27121da177e4SLinus Torvalds */ 27131da177e4SLinus Torvalds void tcp_xmit_retransmit_queue(struct sock *sk) 27141da177e4SLinus Torvalds { 27156687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 27161da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 27171da177e4SLinus Torvalds struct sk_buff *skb; 27180e1c54c2SIlpo Järvinen struct sk_buff *hole = NULL; 2719618d9f25SIlpo Järvinen u32 last_lost; 272061eb55f4SIlpo Järvinen int mib_idx; 27210e1c54c2SIlpo Järvinen int fwd_rexmitting = 0; 27226a438bbeSStephen Hemminger 272345e77d31SIlpo Järvinen if (!tp->packets_out) 272445e77d31SIlpo Järvinen return; 272545e77d31SIlpo Järvinen 272608ebd172SIlpo Järvinen if (!tp->lost_out) 272708ebd172SIlpo Järvinen tp->retransmit_high = tp->snd_una; 272808ebd172SIlpo Järvinen 2729618d9f25SIlpo Järvinen if (tp->retransmit_skb_hint) { 27306a438bbeSStephen Hemminger skb = tp->retransmit_skb_hint; 2731618d9f25SIlpo Järvinen last_lost = TCP_SKB_CB(skb)->end_seq; 2732618d9f25SIlpo Järvinen if (after(last_lost, tp->retransmit_high)) 2733618d9f25SIlpo Järvinen last_lost = tp->retransmit_high; 2734618d9f25SIlpo Järvinen } else { 2735fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 2736618d9f25SIlpo Järvinen last_lost = tp->snd_una; 2737618d9f25SIlpo Järvinen } 27381da177e4SLinus Torvalds 2739fe067e8aSDavid S. Miller tcp_for_write_queue_from(skb, sk) { 27401da177e4SLinus Torvalds __u8 sacked = TCP_SKB_CB(skb)->sacked; 27411da177e4SLinus Torvalds 2742fe067e8aSDavid S. Miller if (skb == tcp_send_head(sk)) 2743fe067e8aSDavid S. Miller break; 27446a438bbeSStephen Hemminger /* we could do better than to assign each time */ 274551456b29SIan Morris if (!hole) 27466a438bbeSStephen Hemminger tp->retransmit_skb_hint = skb; 27476a438bbeSStephen Hemminger 27481da177e4SLinus Torvalds /* Assume this retransmit will generate 27491da177e4SLinus Torvalds * only one packet for congestion window 27501da177e4SLinus Torvalds * calculation purposes. This works because 27511da177e4SLinus Torvalds * tcp_retransmit_skb() will chop up the 27521da177e4SLinus Torvalds * packet to be MSS sized and all the 27531da177e4SLinus Torvalds * packet counting works out. 27541da177e4SLinus Torvalds */ 27551da177e4SLinus Torvalds if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) 27561da177e4SLinus Torvalds return; 27570e1c54c2SIlpo Järvinen 27580e1c54c2SIlpo Järvinen if (fwd_rexmitting) { 27590e1c54c2SIlpo Järvinen begin_fwd: 27600e1c54c2SIlpo Järvinen if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp))) 2761006f582cSIlpo Järvinen break; 27620e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFORWARDRETRANS; 27630e1c54c2SIlpo Järvinen 27640e1c54c2SIlpo Järvinen } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) { 2765618d9f25SIlpo Järvinen tp->retransmit_high = last_lost; 27660e1c54c2SIlpo Järvinen if (!tcp_can_forward_retransmit(sk)) 27670e1c54c2SIlpo Järvinen break; 27680e1c54c2SIlpo Järvinen /* Backtrack if necessary to non-L'ed skb */ 276900db4124SIan Morris if (hole) { 27700e1c54c2SIlpo Järvinen skb = hole; 27710e1c54c2SIlpo Järvinen hole = NULL; 27720e1c54c2SIlpo Järvinen } 27730e1c54c2SIlpo Järvinen fwd_rexmitting = 1; 27740e1c54c2SIlpo Järvinen goto begin_fwd; 27750e1c54c2SIlpo Järvinen 27760e1c54c2SIlpo Järvinen } else if (!(sacked & TCPCB_LOST)) { 277751456b29SIan Morris if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) 27780e1c54c2SIlpo Järvinen hole = skb; 277961eb55f4SIlpo Järvinen continue; 27801da177e4SLinus Torvalds 27810e1c54c2SIlpo Järvinen } else { 2782618d9f25SIlpo Järvinen last_lost = TCP_SKB_CB(skb)->end_seq; 27830e1c54c2SIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Loss) 27840e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFASTRETRANS; 27850e1c54c2SIlpo Järvinen else 27860e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; 27870e1c54c2SIlpo Järvinen } 27880e1c54c2SIlpo Järvinen 27890e1c54c2SIlpo Järvinen if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) 279061eb55f4SIlpo Järvinen continue; 279140b215e5SPavel Emelyanov 279224ab6becSYuchung Cheng if (tcp_retransmit_skb(sk, skb)) 27931da177e4SLinus Torvalds return; 279424ab6becSYuchung Cheng 2795de0744afSPavel Emelyanov NET_INC_STATS_BH(sock_net(sk), mib_idx); 27961da177e4SLinus Torvalds 2797684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 2798a262f0cdSNandita Dukkipati tp->prr_out += tcp_skb_pcount(skb); 2799a262f0cdSNandita Dukkipati 2800fe067e8aSDavid S. Miller if (skb == tcp_write_queue_head(sk)) 2801463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 28023f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, 28033f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 28041da177e4SLinus Torvalds } 28051da177e4SLinus Torvalds } 28061da177e4SLinus Torvalds 2807d83769a5SEric Dumazet /* We allow to exceed memory limits for FIN packets to expedite 2808d83769a5SEric Dumazet * connection tear down and (memory) recovery. 2809845704a5SEric Dumazet * Otherwise tcp_send_fin() could be tempted to either delay FIN 2810845704a5SEric Dumazet * or even be forced to close flow without any FIN. 2811a6c5ea4cSEric Dumazet * In general, we want to allow one skb per socket to avoid hangs 2812a6c5ea4cSEric Dumazet * with edge trigger epoll() 2813d83769a5SEric Dumazet */ 2814a6c5ea4cSEric Dumazet void sk_forced_mem_schedule(struct sock *sk, int size) 2815d83769a5SEric Dumazet { 2816e805605cSJohannes Weiner int amt; 2817d83769a5SEric Dumazet 2818d83769a5SEric Dumazet if (size <= sk->sk_forward_alloc) 2819d83769a5SEric Dumazet return; 2820d83769a5SEric Dumazet amt = sk_mem_pages(size); 2821d83769a5SEric Dumazet sk->sk_forward_alloc += amt * SK_MEM_QUANTUM; 2822e805605cSJohannes Weiner sk_memory_allocated_add(sk, amt); 2823e805605cSJohannes Weiner 2824baac50bbSJohannes Weiner if (mem_cgroup_sockets_enabled && sk->sk_memcg) 2825baac50bbSJohannes Weiner mem_cgroup_charge_skmem(sk->sk_memcg, amt); 2826d83769a5SEric Dumazet } 2827d83769a5SEric Dumazet 2828845704a5SEric Dumazet /* Send a FIN. The caller locks the socket for us. 2829845704a5SEric Dumazet * We should try to send a FIN packet really hard, but eventually give up. 28301da177e4SLinus Torvalds */ 28311da177e4SLinus Torvalds void tcp_send_fin(struct sock *sk) 28321da177e4SLinus Torvalds { 2833845704a5SEric Dumazet struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk); 28341da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 28351da177e4SLinus Torvalds 2836845704a5SEric Dumazet /* Optimization, tack on the FIN if we have one skb in write queue and 2837845704a5SEric Dumazet * this skb was not yet sent, or we are under memory pressure. 2838845704a5SEric Dumazet * Note: in the latter case, FIN packet will be sent after a timeout, 2839845704a5SEric Dumazet * as TCP stack thinks it has already been transmitted. 28401da177e4SLinus Torvalds */ 2841b8da51ebSEric Dumazet if (tskb && (tcp_send_head(sk) || tcp_under_memory_pressure(sk))) { 2842845704a5SEric Dumazet coalesce: 2843845704a5SEric Dumazet TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN; 2844845704a5SEric Dumazet TCP_SKB_CB(tskb)->end_seq++; 28451da177e4SLinus Torvalds tp->write_seq++; 2846845704a5SEric Dumazet if (!tcp_send_head(sk)) { 2847845704a5SEric Dumazet /* This means tskb was already sent. 2848845704a5SEric Dumazet * Pretend we included the FIN on previous transmit. 2849845704a5SEric Dumazet * We need to set tp->snd_nxt to the value it would have 2850845704a5SEric Dumazet * if FIN had been sent. This is because retransmit path 2851845704a5SEric Dumazet * does not change tp->snd_nxt. 2852845704a5SEric Dumazet */ 2853845704a5SEric Dumazet tp->snd_nxt++; 2854845704a5SEric Dumazet return; 2855845704a5SEric Dumazet } 28561da177e4SLinus Torvalds } else { 2857845704a5SEric Dumazet skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation); 2858845704a5SEric Dumazet if (unlikely(!skb)) { 2859845704a5SEric Dumazet if (tskb) 2860845704a5SEric Dumazet goto coalesce; 2861845704a5SEric Dumazet return; 28621da177e4SLinus Torvalds } 2863d83769a5SEric Dumazet skb_reserve(skb, MAX_TCP_HEADER); 2864a6c5ea4cSEric Dumazet sk_forced_mem_schedule(sk, skb->truesize); 28651da177e4SLinus Torvalds /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 2866e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tp->write_seq, 2867a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_FIN); 28681da177e4SLinus Torvalds tcp_queue_skb(sk, skb); 28691da177e4SLinus Torvalds } 2870845704a5SEric Dumazet __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF); 28711da177e4SLinus Torvalds } 28721da177e4SLinus Torvalds 28731da177e4SLinus Torvalds /* We get here when a process closes a file descriptor (either due to 28741da177e4SLinus Torvalds * an explicit close() or as a byproduct of exit()'ing) and there 28751da177e4SLinus Torvalds * was unread data in the receive queue. This behavior is recommended 287665bb723cSGerrit Renker * by RFC 2525, section 2.17. -DaveM 28771da177e4SLinus Torvalds */ 2878dd0fc66fSAl Viro void tcp_send_active_reset(struct sock *sk, gfp_t priority) 28791da177e4SLinus Torvalds { 28801da177e4SLinus Torvalds struct sk_buff *skb; 28811da177e4SLinus Torvalds 28821da177e4SLinus Torvalds /* NOTE: No TCP options attached and we never retransmit this. */ 28831da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, priority); 28841da177e4SLinus Torvalds if (!skb) { 28854e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 28861da177e4SLinus Torvalds return; 28871da177e4SLinus Torvalds } 28881da177e4SLinus Torvalds 28891da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 28901da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 2891e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), 2892a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_RST); 2893675ee231SEric Dumazet skb_mstamp_get(&skb->skb_mstamp); 28941da177e4SLinus Torvalds /* Send it off. */ 2895dfb4b9dcSDavid S. Miller if (tcp_transmit_skb(sk, skb, 0, priority)) 28964e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 289726af65cbSSridhar Samudrala 289881cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); 28991da177e4SLinus Torvalds } 29001da177e4SLinus Torvalds 290167edfef7SAndi Kleen /* Send a crossed SYN-ACK during socket establishment. 290267edfef7SAndi Kleen * WARNING: This routine must only be called when we have already sent 29031da177e4SLinus Torvalds * a SYN packet that crossed the incoming SYN that caused this routine 29041da177e4SLinus Torvalds * to get called. If this assumption fails then the initial rcv_wnd 29051da177e4SLinus Torvalds * and rcv_wscale values will not be correct. 29061da177e4SLinus Torvalds */ 29071da177e4SLinus Torvalds int tcp_send_synack(struct sock *sk) 29081da177e4SLinus Torvalds { 29091da177e4SLinus Torvalds struct sk_buff *skb; 29101da177e4SLinus Torvalds 2911fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 291251456b29SIan Morris if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 291391df42beSJoe Perches pr_debug("%s: wrong queue state\n", __func__); 29141da177e4SLinus Torvalds return -EFAULT; 29151da177e4SLinus Torvalds } 29164de075e0SEric Dumazet if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { 29171da177e4SLinus Torvalds if (skb_cloned(skb)) { 29181da177e4SLinus Torvalds struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC); 291951456b29SIan Morris if (!nskb) 29201da177e4SLinus Torvalds return -ENOMEM; 2921fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 2922f4a775d1SEric Dumazet __skb_header_release(nskb); 2923fe067e8aSDavid S. Miller __tcp_add_write_queue_head(sk, nskb); 29243ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 29253ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 29263ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 29271da177e4SLinus Torvalds skb = nskb; 29281da177e4SLinus Torvalds } 29291da177e4SLinus Torvalds 29304de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; 2931735d3831SFlorian Westphal tcp_ecn_send_synack(sk, skb); 29321da177e4SLinus Torvalds } 2933dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 29341da177e4SLinus Torvalds } 29351da177e4SLinus Torvalds 29364aea39c1SEric Dumazet /** 29374aea39c1SEric Dumazet * tcp_make_synack - Prepare a SYN-ACK. 29384aea39c1SEric Dumazet * sk: listener socket 29394aea39c1SEric Dumazet * dst: dst entry attached to the SYNACK 29404aea39c1SEric Dumazet * req: request_sock pointer 29414aea39c1SEric Dumazet * 29424aea39c1SEric Dumazet * Allocate one skb and build a SYNACK packet. 29434aea39c1SEric Dumazet * @dst is consumed : Caller should not use it again. 29444aea39c1SEric Dumazet */ 29455d062de7SEric Dumazet struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst, 2946e6b4d113SWilliam Allen Simpson struct request_sock *req, 2947ca6fb065SEric Dumazet struct tcp_fastopen_cookie *foc, 2948ca6fb065SEric Dumazet bool attach_req) 29491da177e4SLinus Torvalds { 29502e6599cbSArnaldo Carvalho de Melo struct inet_request_sock *ireq = inet_rsk(req); 29515d062de7SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 295280f03e27SEric Dumazet struct tcp_md5sig_key *md5 = NULL; 29535d062de7SEric Dumazet struct tcp_out_options opts; 29545d062de7SEric Dumazet struct sk_buff *skb; 2955bd0388aeSWilliam Allen Simpson int tcp_header_size; 29565d062de7SEric Dumazet struct tcphdr *th; 29575d062de7SEric Dumazet u16 user_mss; 2958f5fff5dcSTom Quetchenbach int mss; 29591da177e4SLinus Torvalds 2960ca6fb065SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 29614aea39c1SEric Dumazet if (unlikely(!skb)) { 29624aea39c1SEric Dumazet dst_release(dst); 29631da177e4SLinus Torvalds return NULL; 29644aea39c1SEric Dumazet } 29651da177e4SLinus Torvalds /* Reserve space for headers. */ 29661da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 29671da177e4SLinus Torvalds 2968ca6fb065SEric Dumazet if (attach_req) { 29699e17f8a4SEric Dumazet skb_set_owner_w(skb, req_to_sk(req)); 2970ca6fb065SEric Dumazet } else { 2971ca6fb065SEric Dumazet /* sk is a const pointer, because we want to express multiple 2972ca6fb065SEric Dumazet * cpu might call us concurrently. 2973ca6fb065SEric Dumazet * sk->sk_wmem_alloc in an atomic, we can promote to rw. 2974ca6fb065SEric Dumazet */ 2975ca6fb065SEric Dumazet skb_set_owner_w(skb, (struct sock *)sk); 2976ca6fb065SEric Dumazet } 29774aea39c1SEric Dumazet skb_dst_set(skb, dst); 29781da177e4SLinus Torvalds 29790dbaee3bSDavid S. Miller mss = dst_metric_advmss(dst); 29805d062de7SEric Dumazet user_mss = READ_ONCE(tp->rx_opt.user_mss); 29815d062de7SEric Dumazet if (user_mss && user_mss < mss) 29825d062de7SEric Dumazet mss = user_mss; 2983f5fff5dcSTom Quetchenbach 298433ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 29858b5f12d0SFlorian Westphal #ifdef CONFIG_SYN_COOKIES 29868b5f12d0SFlorian Westphal if (unlikely(req->cookie_ts)) 29877faee5c0SEric Dumazet skb->skb_mstamp.stamp_jiffies = cookie_init_timestamp(req); 29888b5f12d0SFlorian Westphal else 29898b5f12d0SFlorian Westphal #endif 29907faee5c0SEric Dumazet skb_mstamp_get(&skb->skb_mstamp); 299180f03e27SEric Dumazet 299280f03e27SEric Dumazet #ifdef CONFIG_TCP_MD5SIG 299380f03e27SEric Dumazet rcu_read_lock(); 2994fd3a154aSEric Dumazet md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req)); 299580f03e27SEric Dumazet #endif 299658d607d3SEric Dumazet skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4); 299737bfbddaSEric Dumazet tcp_header_size = tcp_synack_options(req, mss, skb, &opts, md5, foc) + 299837bfbddaSEric Dumazet sizeof(*th); 299933ad798cSAdam Langley 3000aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 3001aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 30021da177e4SLinus Torvalds 3003aa8223c7SArnaldo Carvalho de Melo th = tcp_hdr(skb); 30041da177e4SLinus Torvalds memset(th, 0, sizeof(struct tcphdr)); 30051da177e4SLinus Torvalds th->syn = 1; 30061da177e4SLinus Torvalds th->ack = 1; 30076ac705b1SEric Dumazet tcp_ecn_make_synack(req, th); 3008b44084c2SEric Dumazet th->source = htons(ireq->ir_num); 3009634fb979SEric Dumazet th->dest = ireq->ir_rmt_port; 3010e870a8efSIlpo Järvinen /* Setting of flags are superfluous here for callers (and ECE is 3011e870a8efSIlpo Järvinen * not even correctly set) 3012e870a8efSIlpo Järvinen */ 3013e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn, 3014a3433f35SChangli Gao TCPHDR_SYN | TCPHDR_ACK); 30154957faadSWilliam Allen Simpson 30161da177e4SLinus Torvalds th->seq = htonl(TCP_SKB_CB(skb)->seq); 30178336886fSJerry Chu /* XXX data is queued and acked as is. No buffer/window check */ 30188336886fSJerry Chu th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt); 30191da177e4SLinus Torvalds 30201da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 3021ed53d0abSEric Dumazet th->window = htons(min(req->rsk_rcv_wnd, 65535U)); 30225d062de7SEric Dumazet tcp_options_write((__be32 *)(th + 1), NULL, &opts); 30231da177e4SLinus Torvalds th->doff = (tcp_header_size >> 2); 3024a0b8486cSEric Dumazet TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_OUTSEGS); 3025cfb6eeb4SYOSHIFUJI Hideaki 3026cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 3027cfb6eeb4SYOSHIFUJI Hideaki /* Okay, we have all we need - do the md5 hash if needed */ 302880f03e27SEric Dumazet if (md5) 3029bd0388aeSWilliam Allen Simpson tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, 303039f8e58eSEric Dumazet md5, req_to_sk(req), skb); 303180f03e27SEric Dumazet rcu_read_unlock(); 3032cfb6eeb4SYOSHIFUJI Hideaki #endif 3033cfb6eeb4SYOSHIFUJI Hideaki 3034b50edd78SEric Dumazet /* Do not fool tcpdump (if any), clean our debris */ 3035b50edd78SEric Dumazet skb->tstamp.tv64 = 0; 30361da177e4SLinus Torvalds return skb; 30371da177e4SLinus Torvalds } 30384bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_make_synack); 30391da177e4SLinus Torvalds 304081164413SDaniel Borkmann static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst) 304181164413SDaniel Borkmann { 304281164413SDaniel Borkmann struct inet_connection_sock *icsk = inet_csk(sk); 304381164413SDaniel Borkmann const struct tcp_congestion_ops *ca; 304481164413SDaniel Borkmann u32 ca_key = dst_metric(dst, RTAX_CC_ALGO); 304581164413SDaniel Borkmann 304681164413SDaniel Borkmann if (ca_key == TCP_CA_UNSPEC) 304781164413SDaniel Borkmann return; 304881164413SDaniel Borkmann 304981164413SDaniel Borkmann rcu_read_lock(); 305081164413SDaniel Borkmann ca = tcp_ca_find_key(ca_key); 305181164413SDaniel Borkmann if (likely(ca && try_module_get(ca->owner))) { 305281164413SDaniel Borkmann module_put(icsk->icsk_ca_ops->owner); 305381164413SDaniel Borkmann icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst); 305481164413SDaniel Borkmann icsk->icsk_ca_ops = ca; 305581164413SDaniel Borkmann } 305681164413SDaniel Borkmann rcu_read_unlock(); 305781164413SDaniel Borkmann } 305881164413SDaniel Borkmann 305967edfef7SAndi Kleen /* Do all connect socket setups that can be done AF independent. */ 3060f7e56a76Sstephen hemminger static void tcp_connect_init(struct sock *sk) 30611da177e4SLinus Torvalds { 3062cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 30631da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 30641da177e4SLinus Torvalds __u8 rcv_wscale; 30651da177e4SLinus Torvalds 30661da177e4SLinus Torvalds /* We'll fix this up when we get a response from the other end. 30671da177e4SLinus Torvalds * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 30681da177e4SLinus Torvalds */ 30691da177e4SLinus Torvalds tp->tcp_header_len = sizeof(struct tcphdr) + 3070bb5b7c11SDavid S. Miller (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0); 30711da177e4SLinus Torvalds 3072cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 307300db4124SIan Morris if (tp->af_specific->md5_lookup(sk, sk)) 3074cfb6eeb4SYOSHIFUJI Hideaki tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 3075cfb6eeb4SYOSHIFUJI Hideaki #endif 3076cfb6eeb4SYOSHIFUJI Hideaki 30771da177e4SLinus Torvalds /* If user gave his TCP_MAXSEG, record it to clamp */ 30781da177e4SLinus Torvalds if (tp->rx_opt.user_mss) 30791da177e4SLinus Torvalds tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 30801da177e4SLinus Torvalds tp->max_window = 0; 30815d424d5aSJohn Heffner tcp_mtup_init(sk); 30821da177e4SLinus Torvalds tcp_sync_mss(sk, dst_mtu(dst)); 30831da177e4SLinus Torvalds 308481164413SDaniel Borkmann tcp_ca_dst_init(sk, dst); 308581164413SDaniel Borkmann 30861da177e4SLinus Torvalds if (!tp->window_clamp) 30871da177e4SLinus Torvalds tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 30880dbaee3bSDavid S. Miller tp->advmss = dst_metric_advmss(dst); 3089f5fff5dcSTom Quetchenbach if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss) 3090f5fff5dcSTom Quetchenbach tp->advmss = tp->rx_opt.user_mss; 3091f5fff5dcSTom Quetchenbach 30921da177e4SLinus Torvalds tcp_initialize_rcv_mss(sk); 30931da177e4SLinus Torvalds 3094e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 3095e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 3096e88c64f0SHagen Paul Pfeifer (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) 3097e88c64f0SHagen Paul Pfeifer tp->window_clamp = tcp_full_space(sk); 3098e88c64f0SHagen Paul Pfeifer 30991da177e4SLinus Torvalds tcp_select_initial_window(tcp_full_space(sk), 31001da177e4SLinus Torvalds tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 31011da177e4SLinus Torvalds &tp->rcv_wnd, 31021da177e4SLinus Torvalds &tp->window_clamp, 3103bb5b7c11SDavid S. Miller sysctl_tcp_window_scaling, 310431d12926Slaurent chavey &rcv_wscale, 310531d12926Slaurent chavey dst_metric(dst, RTAX_INITRWND)); 31061da177e4SLinus Torvalds 31071da177e4SLinus Torvalds tp->rx_opt.rcv_wscale = rcv_wscale; 31081da177e4SLinus Torvalds tp->rcv_ssthresh = tp->rcv_wnd; 31091da177e4SLinus Torvalds 31101da177e4SLinus Torvalds sk->sk_err = 0; 31111da177e4SLinus Torvalds sock_reset_flag(sk, SOCK_DONE); 31121da177e4SLinus Torvalds tp->snd_wnd = 0; 3113ee7537b6SHantzis Fotis tcp_init_wl(tp, 0); 31141da177e4SLinus Torvalds tp->snd_una = tp->write_seq; 31151da177e4SLinus Torvalds tp->snd_sml = tp->write_seq; 311633f5f57eSIlpo Järvinen tp->snd_up = tp->write_seq; 3117370816aeSPavel Emelyanov tp->snd_nxt = tp->write_seq; 3118ee995283SPavel Emelyanov 3119ee995283SPavel Emelyanov if (likely(!tp->repair)) 31201da177e4SLinus Torvalds tp->rcv_nxt = 0; 3121c7781a6eSAndrew Vagin else 3122c7781a6eSAndrew Vagin tp->rcv_tstamp = tcp_time_stamp; 3123ee995283SPavel Emelyanov tp->rcv_wup = tp->rcv_nxt; 3124ee995283SPavel Emelyanov tp->copied_seq = tp->rcv_nxt; 31251da177e4SLinus Torvalds 3126463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; 3127463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_retransmits = 0; 31281da177e4SLinus Torvalds tcp_clear_retrans(tp); 31291da177e4SLinus Torvalds } 31301da177e4SLinus Torvalds 3131783237e8SYuchung Cheng static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) 3132783237e8SYuchung Cheng { 3133783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3134783237e8SYuchung Cheng struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 3135783237e8SYuchung Cheng 3136783237e8SYuchung Cheng tcb->end_seq += skb->len; 3137f4a775d1SEric Dumazet __skb_header_release(skb); 3138783237e8SYuchung Cheng __tcp_add_write_queue_tail(sk, skb); 3139783237e8SYuchung Cheng sk->sk_wmem_queued += skb->truesize; 3140783237e8SYuchung Cheng sk_mem_charge(sk, skb->truesize); 3141783237e8SYuchung Cheng tp->write_seq = tcb->end_seq; 3142783237e8SYuchung Cheng tp->packets_out += tcp_skb_pcount(skb); 3143783237e8SYuchung Cheng } 3144783237e8SYuchung Cheng 3145783237e8SYuchung Cheng /* Build and send a SYN with data and (cached) Fast Open cookie. However, 3146783237e8SYuchung Cheng * queue a data-only packet after the regular SYN, such that regular SYNs 3147783237e8SYuchung Cheng * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges 3148783237e8SYuchung Cheng * only the SYN sequence, the data are retransmitted in the first ACK. 3149783237e8SYuchung Cheng * If cookie is not cached or other error occurs, falls back to send a 3150783237e8SYuchung Cheng * regular SYN with Fast Open cookie request option. 3151783237e8SYuchung Cheng */ 3152783237e8SYuchung Cheng static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) 3153783237e8SYuchung Cheng { 3154783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3155783237e8SYuchung Cheng struct tcp_fastopen_request *fo = tp->fastopen_req; 315607e100f9SEric Dumazet int syn_loss = 0, space, err = 0; 3157aab48743SYuchung Cheng unsigned long last_syn_loss = 0; 3158355a901eSEric Dumazet struct sk_buff *syn_data; 3159783237e8SYuchung Cheng 316067da22d2SYuchung Cheng tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ 3161aab48743SYuchung Cheng tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie, 3162aab48743SYuchung Cheng &syn_loss, &last_syn_loss); 3163aab48743SYuchung Cheng /* Recurring FO SYN losses: revert to regular handshake temporarily */ 3164aab48743SYuchung Cheng if (syn_loss > 1 && 3165aab48743SYuchung Cheng time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) { 3166aab48743SYuchung Cheng fo->cookie.len = -1; 3167aab48743SYuchung Cheng goto fallback; 3168aab48743SYuchung Cheng } 3169aab48743SYuchung Cheng 317067da22d2SYuchung Cheng if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE) 317167da22d2SYuchung Cheng fo->cookie.len = -1; 317267da22d2SYuchung Cheng else if (fo->cookie.len <= 0) 3173783237e8SYuchung Cheng goto fallback; 3174783237e8SYuchung Cheng 3175783237e8SYuchung Cheng /* MSS for SYN-data is based on cached MSS and bounded by PMTU and 3176783237e8SYuchung Cheng * user-MSS. Reserve maximum option space for middleboxes that add 3177783237e8SYuchung Cheng * private TCP options. The cost is reduced data space in SYN :( 3178783237e8SYuchung Cheng */ 3179783237e8SYuchung Cheng if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp) 3180783237e8SYuchung Cheng tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 31811b63edd6SYuchung Cheng space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) - 3182783237e8SYuchung Cheng MAX_TCP_OPTION_SPACE; 3183783237e8SYuchung Cheng 3184f5ddcbbbSEric Dumazet space = min_t(size_t, space, fo->size); 3185f5ddcbbbSEric Dumazet 3186f5ddcbbbSEric Dumazet /* limit to order-0 allocations */ 3187f5ddcbbbSEric Dumazet space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER)); 3188f5ddcbbbSEric Dumazet 3189eb934478SEric Dumazet syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false); 3190355a901eSEric Dumazet if (!syn_data) 3191783237e8SYuchung Cheng goto fallback; 3192355a901eSEric Dumazet syn_data->ip_summed = CHECKSUM_PARTIAL; 3193355a901eSEric Dumazet memcpy(syn_data->cb, syn->cb, sizeof(syn->cb)); 319407e100f9SEric Dumazet if (space) { 319507e100f9SEric Dumazet int copied = copy_from_iter(skb_put(syn_data, space), space, 319657be5bdaSAl Viro &fo->data->msg_iter); 319757be5bdaSAl Viro if (unlikely(!copied)) { 3198355a901eSEric Dumazet kfree_skb(syn_data); 3199783237e8SYuchung Cheng goto fallback; 3200783237e8SYuchung Cheng } 320157be5bdaSAl Viro if (copied != space) { 320257be5bdaSAl Viro skb_trim(syn_data, copied); 320357be5bdaSAl Viro space = copied; 320457be5bdaSAl Viro } 320507e100f9SEric Dumazet } 3206355a901eSEric Dumazet /* No more data pending in inet_wait_for_connect() */ 3207355a901eSEric Dumazet if (space == fo->size) 3208355a901eSEric Dumazet fo->data = NULL; 3209355a901eSEric Dumazet fo->copied = space; 3210783237e8SYuchung Cheng 3211355a901eSEric Dumazet tcp_connect_queue_skb(sk, syn_data); 3212355a901eSEric Dumazet 3213355a901eSEric Dumazet err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation); 3214355a901eSEric Dumazet 3215355a901eSEric Dumazet syn->skb_mstamp = syn_data->skb_mstamp; 3216355a901eSEric Dumazet 3217355a901eSEric Dumazet /* Now full SYN+DATA was cloned and sent (or not), 3218355a901eSEric Dumazet * remove the SYN from the original skb (syn_data) 3219355a901eSEric Dumazet * we keep in write queue in case of a retransmit, as we 3220355a901eSEric Dumazet * also have the SYN packet (with no data) in the same queue. 3221431a9124SEric Dumazet */ 3222355a901eSEric Dumazet TCP_SKB_CB(syn_data)->seq++; 3223355a901eSEric Dumazet TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH; 3224355a901eSEric Dumazet if (!err) { 322567da22d2SYuchung Cheng tp->syn_data = (fo->copied > 0); 3226f19c29e3SYuchung Cheng NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT); 3227783237e8SYuchung Cheng goto done; 3228783237e8SYuchung Cheng } 3229783237e8SYuchung Cheng 3230783237e8SYuchung Cheng fallback: 3231783237e8SYuchung Cheng /* Send a regular SYN with Fast Open cookie request option */ 3232783237e8SYuchung Cheng if (fo->cookie.len > 0) 3233783237e8SYuchung Cheng fo->cookie.len = 0; 3234783237e8SYuchung Cheng err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); 3235783237e8SYuchung Cheng if (err) 3236783237e8SYuchung Cheng tp->syn_fastopen = 0; 3237783237e8SYuchung Cheng done: 3238783237e8SYuchung Cheng fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ 3239783237e8SYuchung Cheng return err; 3240783237e8SYuchung Cheng } 3241783237e8SYuchung Cheng 324267edfef7SAndi Kleen /* Build a SYN and send it off. */ 32431da177e4SLinus Torvalds int tcp_connect(struct sock *sk) 32441da177e4SLinus Torvalds { 32451da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 32461da177e4SLinus Torvalds struct sk_buff *buff; 3247ee586811SEric Paris int err; 32481da177e4SLinus Torvalds 32491da177e4SLinus Torvalds tcp_connect_init(sk); 32501da177e4SLinus Torvalds 32512b916477SAndrey Vagin if (unlikely(tp->repair)) { 32522b916477SAndrey Vagin tcp_finish_connect(sk, NULL); 32532b916477SAndrey Vagin return 0; 32542b916477SAndrey Vagin } 32552b916477SAndrey Vagin 3256eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true); 3257355a901eSEric Dumazet if (unlikely(!buff)) 32581da177e4SLinus Torvalds return -ENOBUFS; 32591da177e4SLinus Torvalds 3260a3433f35SChangli Gao tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); 32617faee5c0SEric Dumazet tp->retrans_stamp = tcp_time_stamp; 3262783237e8SYuchung Cheng tcp_connect_queue_skb(sk, buff); 3263735d3831SFlorian Westphal tcp_ecn_send_syn(sk, buff); 32641da177e4SLinus Torvalds 3265783237e8SYuchung Cheng /* Send off SYN; include data in Fast Open. */ 3266783237e8SYuchung Cheng err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : 3267783237e8SYuchung Cheng tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); 3268ee586811SEric Paris if (err == -ECONNREFUSED) 3269ee586811SEric Paris return err; 3270bd37a088SWei Yongjun 3271bd37a088SWei Yongjun /* We change tp->snd_nxt after the tcp_transmit_skb() call 3272bd37a088SWei Yongjun * in order to make this packet get counted in tcpOutSegs. 3273bd37a088SWei Yongjun */ 3274bd37a088SWei Yongjun tp->snd_nxt = tp->write_seq; 3275bd37a088SWei Yongjun tp->pushed_seq = tp->write_seq; 327681cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); 32771da177e4SLinus Torvalds 32781da177e4SLinus Torvalds /* Timer for repeating the SYN until an answer. */ 32793f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 32803f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 32811da177e4SLinus Torvalds return 0; 32821da177e4SLinus Torvalds } 32834bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_connect); 32841da177e4SLinus Torvalds 32851da177e4SLinus Torvalds /* Send out a delayed ack, the caller does the policy checking 32861da177e4SLinus Torvalds * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 32871da177e4SLinus Torvalds * for details. 32881da177e4SLinus Torvalds */ 32891da177e4SLinus Torvalds void tcp_send_delayed_ack(struct sock *sk) 32901da177e4SLinus Torvalds { 3291463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 3292463c84b9SArnaldo Carvalho de Melo int ato = icsk->icsk_ack.ato; 32931da177e4SLinus Torvalds unsigned long timeout; 32941da177e4SLinus Torvalds 32959890092eSFlorian Westphal tcp_ca_event(sk, CA_EVENT_DELAYED_ACK); 32969890092eSFlorian Westphal 32971da177e4SLinus Torvalds if (ato > TCP_DELACK_MIN) { 3298463c84b9SArnaldo Carvalho de Melo const struct tcp_sock *tp = tcp_sk(sk); 32991da177e4SLinus Torvalds int max_ato = HZ / 2; 33001da177e4SLinus Torvalds 3301056834d9SIlpo Järvinen if (icsk->icsk_ack.pingpong || 3302056834d9SIlpo Järvinen (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 33031da177e4SLinus Torvalds max_ato = TCP_DELACK_MAX; 33041da177e4SLinus Torvalds 33051da177e4SLinus Torvalds /* Slow path, intersegment interval is "high". */ 33061da177e4SLinus Torvalds 33071da177e4SLinus Torvalds /* If some rtt estimate is known, use it to bound delayed ack. 3308463c84b9SArnaldo Carvalho de Melo * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 33091da177e4SLinus Torvalds * directly. 33101da177e4SLinus Torvalds */ 3311740b0f18SEric Dumazet if (tp->srtt_us) { 3312740b0f18SEric Dumazet int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3), 3313740b0f18SEric Dumazet TCP_DELACK_MIN); 33141da177e4SLinus Torvalds 33151da177e4SLinus Torvalds if (rtt < max_ato) 33161da177e4SLinus Torvalds max_ato = rtt; 33171da177e4SLinus Torvalds } 33181da177e4SLinus Torvalds 33191da177e4SLinus Torvalds ato = min(ato, max_ato); 33201da177e4SLinus Torvalds } 33211da177e4SLinus Torvalds 33221da177e4SLinus Torvalds /* Stay within the limit we were given */ 33231da177e4SLinus Torvalds timeout = jiffies + ato; 33241da177e4SLinus Torvalds 33251da177e4SLinus Torvalds /* Use new timeout only if there wasn't a older one earlier. */ 3326463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 33271da177e4SLinus Torvalds /* If delack timer was blocked or is about to expire, 33281da177e4SLinus Torvalds * send ACK now. 33291da177e4SLinus Torvalds */ 3330463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.blocked || 3331463c84b9SArnaldo Carvalho de Melo time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 33321da177e4SLinus Torvalds tcp_send_ack(sk); 33331da177e4SLinus Torvalds return; 33341da177e4SLinus Torvalds } 33351da177e4SLinus Torvalds 3336463c84b9SArnaldo Carvalho de Melo if (!time_before(timeout, icsk->icsk_ack.timeout)) 3337463c84b9SArnaldo Carvalho de Melo timeout = icsk->icsk_ack.timeout; 33381da177e4SLinus Torvalds } 3339463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 3340463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.timeout = timeout; 3341463c84b9SArnaldo Carvalho de Melo sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 33421da177e4SLinus Torvalds } 33431da177e4SLinus Torvalds 33441da177e4SLinus Torvalds /* This routine sends an ack and also updates the window. */ 33451da177e4SLinus Torvalds void tcp_send_ack(struct sock *sk) 33461da177e4SLinus Torvalds { 33471da177e4SLinus Torvalds struct sk_buff *buff; 33481da177e4SLinus Torvalds 3349058dc334SIlpo Järvinen /* If we have been reset, we may not send again. */ 3350058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3351058dc334SIlpo Järvinen return; 3352058dc334SIlpo Järvinen 33539890092eSFlorian Westphal tcp_ca_event(sk, CA_EVENT_NON_DELAYED_ACK); 33549890092eSFlorian Westphal 33551da177e4SLinus Torvalds /* We are not putting this on the write queue, so 33561da177e4SLinus Torvalds * tcp_transmit_skb() will set the ownership to this 33571da177e4SLinus Torvalds * sock. 33581da177e4SLinus Torvalds */ 33597450aaf6SEric Dumazet buff = alloc_skb(MAX_TCP_HEADER, 33607450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 33617450aaf6SEric Dumazet if (unlikely(!buff)) { 3362463c84b9SArnaldo Carvalho de Melo inet_csk_schedule_ack(sk); 3363463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; 33643f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 33653f421baaSArnaldo Carvalho de Melo TCP_DELACK_MAX, TCP_RTO_MAX); 33661da177e4SLinus Torvalds return; 33671da177e4SLinus Torvalds } 33681da177e4SLinus Torvalds 33691da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 33701da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 3371a3433f35SChangli Gao tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); 33721da177e4SLinus Torvalds 337398781965SEric Dumazet /* We do not want pure acks influencing TCP Small Queues or fq/pacing 337498781965SEric Dumazet * too much. 337598781965SEric Dumazet * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784 337698781965SEric Dumazet * We also avoid tcp_wfree() overhead (cache line miss accessing 337798781965SEric Dumazet * tp->tsq_flags) by using regular sock_wfree() 337898781965SEric Dumazet */ 337998781965SEric Dumazet skb_set_tcp_pure_ack(buff); 338098781965SEric Dumazet 33811da177e4SLinus Torvalds /* Send it off, this clears delayed acks for us. */ 33827faee5c0SEric Dumazet skb_mstamp_get(&buff->skb_mstamp); 33837450aaf6SEric Dumazet tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0); 33841da177e4SLinus Torvalds } 3385e3118e83SDaniel Borkmann EXPORT_SYMBOL_GPL(tcp_send_ack); 33861da177e4SLinus Torvalds 33871da177e4SLinus Torvalds /* This routine sends a packet with an out of date sequence 33881da177e4SLinus Torvalds * number. It assumes the other end will try to ack it. 33891da177e4SLinus Torvalds * 33901da177e4SLinus Torvalds * Question: what should we make while urgent mode? 33911da177e4SLinus Torvalds * 4.4BSD forces sending single byte of data. We cannot send 33921da177e4SLinus Torvalds * out of window data, because we have SND.NXT==SND.MAX... 33931da177e4SLinus Torvalds * 33941da177e4SLinus Torvalds * Current solution: to send TWO zero-length segments in urgent mode: 33951da177e4SLinus Torvalds * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 33961da177e4SLinus Torvalds * out-of-date with SND.UNA-1 to probe window. 33971da177e4SLinus Torvalds */ 3398e520af48SEric Dumazet static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib) 33991da177e4SLinus Torvalds { 34001da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 34011da177e4SLinus Torvalds struct sk_buff *skb; 34021da177e4SLinus Torvalds 34031da177e4SLinus Torvalds /* We don't queue it, tcp_transmit_skb() sets ownership. */ 34047450aaf6SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, 34057450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 340651456b29SIan Morris if (!skb) 34071da177e4SLinus Torvalds return -1; 34081da177e4SLinus Torvalds 34091da177e4SLinus Torvalds /* Reserve space for headers and set control bits. */ 34101da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 34111da177e4SLinus Torvalds /* Use a previous sequence. This should cause the other 34121da177e4SLinus Torvalds * end to send an ack. Don't queue or clone SKB, just 34131da177e4SLinus Torvalds * send it. 34141da177e4SLinus Torvalds */ 3415a3433f35SChangli Gao tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); 34167faee5c0SEric Dumazet skb_mstamp_get(&skb->skb_mstamp); 3417e2e8009fSRenato Westphal NET_INC_STATS(sock_net(sk), mib); 34187450aaf6SEric Dumazet return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0); 34191da177e4SLinus Torvalds } 34201da177e4SLinus Torvalds 3421ee995283SPavel Emelyanov void tcp_send_window_probe(struct sock *sk) 3422ee995283SPavel Emelyanov { 3423ee995283SPavel Emelyanov if (sk->sk_state == TCP_ESTABLISHED) { 3424ee995283SPavel Emelyanov tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; 3425e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE); 3426ee995283SPavel Emelyanov } 3427ee995283SPavel Emelyanov } 3428ee995283SPavel Emelyanov 342967edfef7SAndi Kleen /* Initiate keepalive or window probe from timer. */ 3430e520af48SEric Dumazet int tcp_write_wakeup(struct sock *sk, int mib) 34311da177e4SLinus Torvalds { 34321da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 34331da177e4SLinus Torvalds struct sk_buff *skb; 34341da177e4SLinus Torvalds 3435058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3436058dc334SIlpo Järvinen return -1; 3437058dc334SIlpo Järvinen 343800db4124SIan Morris skb = tcp_send_head(sk); 343900db4124SIan Morris if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { 34401da177e4SLinus Torvalds int err; 34410c54b85fSIlpo Järvinen unsigned int mss = tcp_current_mss(sk); 344290840defSIlpo Järvinen unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 34431da177e4SLinus Torvalds 34441da177e4SLinus Torvalds if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 34451da177e4SLinus Torvalds tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 34461da177e4SLinus Torvalds 34471da177e4SLinus Torvalds /* We are probing the opening of a window 34481da177e4SLinus Torvalds * but the window size is != 0 34491da177e4SLinus Torvalds * must have been a result SWS avoidance ( sender ) 34501da177e4SLinus Torvalds */ 34511da177e4SLinus Torvalds if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 34521da177e4SLinus Torvalds skb->len > mss) { 34531da177e4SLinus Torvalds seg_size = min(seg_size, mss); 34544de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 34556cc55e09SOctavian Purdila if (tcp_fragment(sk, skb, seg_size, mss, GFP_ATOMIC)) 34561da177e4SLinus Torvalds return -1; 34571da177e4SLinus Torvalds } else if (!tcp_skb_pcount(skb)) 34585bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss); 34591da177e4SLinus Torvalds 34604de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 3461dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 346266f5fe62SIlpo Järvinen if (!err) 346366f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 34641da177e4SLinus Torvalds return err; 34651da177e4SLinus Torvalds } else { 346633f5f57eSIlpo Järvinen if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) 3467e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 1, mib); 3468e520af48SEric Dumazet return tcp_xmit_probe_skb(sk, 0, mib); 34691da177e4SLinus Torvalds } 34701da177e4SLinus Torvalds } 34711da177e4SLinus Torvalds 34721da177e4SLinus Torvalds /* A window probe timeout has occurred. If window is not closed send 34731da177e4SLinus Torvalds * a partial packet else a zero probe. 34741da177e4SLinus Torvalds */ 34751da177e4SLinus Torvalds void tcp_send_probe0(struct sock *sk) 34761da177e4SLinus Torvalds { 3477463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 34781da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 3479*c6214a97SNikolay Borisov struct net *net = sock_net(sk); 3480fcdd1cf4SEric Dumazet unsigned long probe_max; 34811da177e4SLinus Torvalds int err; 34821da177e4SLinus Torvalds 3483e520af48SEric Dumazet err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE); 34841da177e4SLinus Torvalds 3485fe067e8aSDavid S. Miller if (tp->packets_out || !tcp_send_head(sk)) { 34861da177e4SLinus Torvalds /* Cancel probe timer, if it is not required. */ 34876687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 0; 3488463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff = 0; 34891da177e4SLinus Torvalds return; 34901da177e4SLinus Torvalds } 34911da177e4SLinus Torvalds 34921da177e4SLinus Torvalds if (err <= 0) { 3493*c6214a97SNikolay Borisov if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2) 3494463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff++; 34956687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out++; 3496fcdd1cf4SEric Dumazet probe_max = TCP_RTO_MAX; 34971da177e4SLinus Torvalds } else { 34981da177e4SLinus Torvalds /* If packet was not sent due to local congestion, 34996687e988SArnaldo Carvalho de Melo * do not backoff and do not remember icsk_probes_out. 35001da177e4SLinus Torvalds * Let local senders to fight for local resources. 35011da177e4SLinus Torvalds * 35021da177e4SLinus Torvalds * Use accumulated backoff yet. 35031da177e4SLinus Torvalds */ 35046687e988SArnaldo Carvalho de Melo if (!icsk->icsk_probes_out) 35056687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 1; 3506fcdd1cf4SEric Dumazet probe_max = TCP_RESOURCE_PROBE_INTERVAL; 35071da177e4SLinus Torvalds } 3508fcdd1cf4SEric Dumazet inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 350921c8fe99SEric Dumazet tcp_probe0_when(sk, probe_max), 3510fcdd1cf4SEric Dumazet TCP_RTO_MAX); 35111da177e4SLinus Torvalds } 35125db92c99SOctavian Purdila 3513ea3bea3aSEric Dumazet int tcp_rtx_synack(const struct sock *sk, struct request_sock *req) 35145db92c99SOctavian Purdila { 35155db92c99SOctavian Purdila const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific; 35165db92c99SOctavian Purdila struct flowi fl; 35175db92c99SOctavian Purdila int res; 35185db92c99SOctavian Purdila 351958d607d3SEric Dumazet tcp_rsk(req)->txhash = net_tx_rndhash(); 3520dc6ef6beSEric Dumazet res = af_ops->send_synack(sk, NULL, &fl, req, NULL, true); 35215db92c99SOctavian Purdila if (!res) { 35225db92c99SOctavian Purdila TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_RETRANSSEGS); 35235db92c99SOctavian Purdila NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 35245db92c99SOctavian Purdila } 35255db92c99SOctavian Purdila return res; 35265db92c99SOctavian Purdila } 35275db92c99SOctavian Purdila EXPORT_SYMBOL(tcp_rtx_synack); 3528