11da177e4SLinus Torvalds /* 21da177e4SLinus Torvalds * INET An implementation of the TCP/IP protocol suite for the LINUX 31da177e4SLinus Torvalds * operating system. INET is implemented using the BSD Socket 41da177e4SLinus Torvalds * interface as the means of communication with the user level. 51da177e4SLinus Torvalds * 61da177e4SLinus Torvalds * Implementation of the Transmission Control Protocol(TCP). 71da177e4SLinus Torvalds * 802c30a84SJesper Juhl * Authors: Ross Biro 91da177e4SLinus Torvalds * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> 101da177e4SLinus Torvalds * Mark Evans, <evansmp@uhura.aston.ac.uk> 111da177e4SLinus Torvalds * Corey Minyard <wf-rch!minyard@relay.EU.net> 121da177e4SLinus Torvalds * Florian La Roche, <flla@stud.uni-sb.de> 131da177e4SLinus Torvalds * Charles Hedrick, <hedrick@klinzhai.rutgers.edu> 141da177e4SLinus Torvalds * Linus Torvalds, <torvalds@cs.helsinki.fi> 151da177e4SLinus Torvalds * Alan Cox, <gw4pts@gw4pts.ampr.org> 161da177e4SLinus Torvalds * Matthew Dillon, <dillon@apollo.west.oic.com> 171da177e4SLinus Torvalds * Arnt Gulbrandsen, <agulbra@nvg.unit.no> 181da177e4SLinus Torvalds * Jorge Cwik, <jorge@laser.satlink.net> 191da177e4SLinus Torvalds */ 201da177e4SLinus Torvalds 211da177e4SLinus Torvalds /* 221da177e4SLinus Torvalds * Changes: Pedro Roque : Retransmit queue handled by TCP. 231da177e4SLinus Torvalds * : Fragmentation on mtu decrease 241da177e4SLinus Torvalds * : Segment collapse on retransmit 251da177e4SLinus Torvalds * : AF independence 261da177e4SLinus Torvalds * 271da177e4SLinus Torvalds * Linus Torvalds : send_delayed_ack 281da177e4SLinus Torvalds * David S. Miller : Charge memory using the right skb 291da177e4SLinus Torvalds * during syn/ack processing. 301da177e4SLinus Torvalds * David S. Miller : Output engine completely rewritten. 311da177e4SLinus Torvalds * Andrea Arcangeli: SYNACK carry ts_recent in tsecr. 321da177e4SLinus Torvalds * Cacophonix Gaul : draft-minshall-nagle-01 331da177e4SLinus Torvalds * J Hadi Salim : ECN support 341da177e4SLinus Torvalds * 351da177e4SLinus Torvalds */ 361da177e4SLinus Torvalds 3791df42beSJoe Perches #define pr_fmt(fmt) "TCP: " fmt 3891df42beSJoe Perches 391da177e4SLinus Torvalds #include <net/tcp.h> 401da177e4SLinus Torvalds 411da177e4SLinus Torvalds #include <linux/compiler.h> 425a0e3ad6STejun Heo #include <linux/gfp.h> 431da177e4SLinus Torvalds #include <linux/module.h> 441da177e4SLinus Torvalds 451da177e4SLinus Torvalds /* People can turn this off for buggy TCP's found in printers etc. */ 46ab32ea5dSBrian Haley int sysctl_tcp_retrans_collapse __read_mostly = 1; 471da177e4SLinus Torvalds 4815d99e02SRick Jones /* People can turn this on to work with those rare, broken TCPs that 4915d99e02SRick Jones * interpret the window field as a signed quantity. 5015d99e02SRick Jones */ 51ab32ea5dSBrian Haley int sysctl_tcp_workaround_signed_windows __read_mostly = 0; 5215d99e02SRick Jones 5346d3ceabSEric Dumazet /* Default TSQ limit of two TSO segments */ 5446d3ceabSEric Dumazet int sysctl_tcp_limit_output_bytes __read_mostly = 131072; 5546d3ceabSEric Dumazet 561da177e4SLinus Torvalds /* This limits the percentage of the congestion window which we 571da177e4SLinus Torvalds * will allow a single TSO frame to consume. Building TSO frames 581da177e4SLinus Torvalds * which are too large can cause TCP streams to be bursty. 591da177e4SLinus Torvalds */ 60ab32ea5dSBrian Haley int sysctl_tcp_tso_win_divisor __read_mostly = 3; 611da177e4SLinus Torvalds 6235089bb2SDavid S. Miller /* By default, RFC2861 behavior. */ 63ab32ea5dSBrian Haley int sysctl_tcp_slow_start_after_idle __read_mostly = 1; 6435089bb2SDavid S. Miller 65c9bee3b7SEric Dumazet unsigned int sysctl_tcp_notsent_lowat __read_mostly = UINT_MAX; 66c9bee3b7SEric Dumazet EXPORT_SYMBOL(sysctl_tcp_notsent_lowat); 67c9bee3b7SEric Dumazet 6846d3ceabSEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 6946d3ceabSEric Dumazet int push_one, gfp_t gfp); 70519855c5SWilliam Allen Simpson 7167edfef7SAndi Kleen /* Account for new data that has been sent to the network. */ 72cf533ea5SEric Dumazet static void tcp_event_new_data_sent(struct sock *sk, const struct sk_buff *skb) 736ff03ac3SIlpo Järvinen { 746ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 756ff03ac3SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 7666f5fe62SIlpo Järvinen unsigned int prior_packets = tp->packets_out; 779e412ba7SIlpo Järvinen 78fe067e8aSDavid S. Miller tcp_advance_send_head(sk, skb); 791da177e4SLinus Torvalds tp->snd_nxt = TCP_SKB_CB(skb)->end_seq; 808512430eSIlpo Järvinen 8166f5fe62SIlpo Järvinen tp->packets_out += tcp_skb_pcount(skb); 826ba8a3b1SNandita Dukkipati if (!prior_packets || icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS || 836a5dc9e5SEric Dumazet icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) { 84750ea2baSYuchung Cheng tcp_rearm_rto(sk); 851da177e4SLinus Torvalds } 86f19c29e3SYuchung Cheng 87f7324acdSDavid S. Miller NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT, 88f19c29e3SYuchung Cheng tcp_skb_pcount(skb)); 896a5dc9e5SEric Dumazet } 901da177e4SLinus Torvalds 911da177e4SLinus Torvalds /* SND.NXT, if window was not shrunk. 921da177e4SLinus Torvalds * If window has been shrunk, what should we make? It is not clear at all. 931da177e4SLinus Torvalds * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-( 941da177e4SLinus Torvalds * Anything in between SND.UNA...SND.UNA+SND.WND also can be already 951da177e4SLinus Torvalds * invalid. OK, let's make this for now: 961da177e4SLinus Torvalds */ 97cf533ea5SEric Dumazet static inline __u32 tcp_acceptable_seq(const struct sock *sk) 981da177e4SLinus Torvalds { 99cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1009e412ba7SIlpo Järvinen 10190840defSIlpo Järvinen if (!before(tcp_wnd_end(tp), tp->snd_nxt)) 1021da177e4SLinus Torvalds return tp->snd_nxt; 1031da177e4SLinus Torvalds else 10490840defSIlpo Järvinen return tcp_wnd_end(tp); 1051da177e4SLinus Torvalds } 1061da177e4SLinus Torvalds 1071da177e4SLinus Torvalds /* Calculate mss to advertise in SYN segment. 1081da177e4SLinus Torvalds * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that: 1091da177e4SLinus Torvalds * 1101da177e4SLinus Torvalds * 1. It is independent of path mtu. 1111da177e4SLinus Torvalds * 2. Ideally, it is maximal possible segment size i.e. 65535-40. 1121da177e4SLinus Torvalds * 3. For IPv4 it is reasonable to calculate it from maximal MTU of 1131da177e4SLinus Torvalds * attached devices, because some buggy hosts are confused by 1141da177e4SLinus Torvalds * large MSS. 1151da177e4SLinus Torvalds * 4. We do not make 3, we advertise MSS, calculated from first 1161da177e4SLinus Torvalds * hop device mtu, but allow to raise it to ip_rt_min_advmss. 1171da177e4SLinus Torvalds * This may be overridden via information stored in routing table. 1181da177e4SLinus Torvalds * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible, 1191da177e4SLinus Torvalds * probably even Jumbo". 1201da177e4SLinus Torvalds */ 1211da177e4SLinus Torvalds static __u16 tcp_advertise_mss(struct sock *sk) 1221da177e4SLinus Torvalds { 1231da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 124cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1251da177e4SLinus Torvalds int mss = tp->advmss; 1261da177e4SLinus Torvalds 1270dbaee3bSDavid S. Miller if (dst) { 1280dbaee3bSDavid S. Miller unsigned int metric = dst_metric_advmss(dst); 1290dbaee3bSDavid S. Miller 1300dbaee3bSDavid S. Miller if (metric < mss) { 1310dbaee3bSDavid S. Miller mss = metric; 1321da177e4SLinus Torvalds tp->advmss = mss; 1331da177e4SLinus Torvalds } 1340dbaee3bSDavid S. Miller } 1351da177e4SLinus Torvalds 1361da177e4SLinus Torvalds return (__u16)mss; 1371da177e4SLinus Torvalds } 1381da177e4SLinus Torvalds 1391da177e4SLinus Torvalds /* RFC2861. Reset CWND after idle period longer RTO to "restart window". 1401da177e4SLinus Torvalds * This is the first part of cwnd validation mechanism. */ 141cf533ea5SEric Dumazet static void tcp_cwnd_restart(struct sock *sk, const struct dst_entry *dst) 1421da177e4SLinus Torvalds { 143463c84b9SArnaldo Carvalho de Melo struct tcp_sock *tp = tcp_sk(sk); 1441da177e4SLinus Torvalds s32 delta = tcp_time_stamp - tp->lsndtime; 1451da177e4SLinus Torvalds u32 restart_cwnd = tcp_init_cwnd(tp, dst); 1461da177e4SLinus Torvalds u32 cwnd = tp->snd_cwnd; 1471da177e4SLinus Torvalds 1486687e988SArnaldo Carvalho de Melo tcp_ca_event(sk, CA_EVENT_CWND_RESTART); 1491da177e4SLinus Torvalds 1506687e988SArnaldo Carvalho de Melo tp->snd_ssthresh = tcp_current_ssthresh(sk); 1511da177e4SLinus Torvalds restart_cwnd = min(restart_cwnd, cwnd); 1521da177e4SLinus Torvalds 153463c84b9SArnaldo Carvalho de Melo while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd) 1541da177e4SLinus Torvalds cwnd >>= 1; 1551da177e4SLinus Torvalds tp->snd_cwnd = max(cwnd, restart_cwnd); 1561da177e4SLinus Torvalds tp->snd_cwnd_stamp = tcp_time_stamp; 1571da177e4SLinus Torvalds tp->snd_cwnd_used = 0; 1581da177e4SLinus Torvalds } 1591da177e4SLinus Torvalds 16067edfef7SAndi Kleen /* Congestion state accounting after a packet has been sent. */ 16140efc6faSStephen Hemminger static void tcp_event_data_sent(struct tcp_sock *tp, 162cf533ea5SEric Dumazet struct sock *sk) 1631da177e4SLinus Torvalds { 164463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 165463c84b9SArnaldo Carvalho de Melo const u32 now = tcp_time_stamp; 166bcefe17cSCong Wang const struct dst_entry *dst = __sk_dst_get(sk); 1671da177e4SLinus Torvalds 16835089bb2SDavid S. Miller if (sysctl_tcp_slow_start_after_idle && 16935089bb2SDavid S. Miller (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto)) 170463c84b9SArnaldo Carvalho de Melo tcp_cwnd_restart(sk, __sk_dst_get(sk)); 1711da177e4SLinus Torvalds 1721da177e4SLinus Torvalds tp->lsndtime = now; 1731da177e4SLinus Torvalds 1741da177e4SLinus Torvalds /* If it is a reply for ato after last received 1751da177e4SLinus Torvalds * packet, enter pingpong mode. 1761da177e4SLinus Torvalds */ 177bcefe17cSCong Wang if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato && 178bcefe17cSCong Wang (!dst || !dst_metric(dst, RTAX_QUICKACK))) 179463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pingpong = 1; 1801da177e4SLinus Torvalds } 1811da177e4SLinus Torvalds 18267edfef7SAndi Kleen /* Account for an ACK we sent. */ 18340efc6faSStephen Hemminger static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts) 1841da177e4SLinus Torvalds { 185463c84b9SArnaldo Carvalho de Melo tcp_dec_quickack_mode(sk, pkts); 186463c84b9SArnaldo Carvalho de Melo inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); 1871da177e4SLinus Torvalds } 1881da177e4SLinus Torvalds 18985f16525SYuchung Cheng 19085f16525SYuchung Cheng u32 tcp_default_init_rwnd(u32 mss) 19185f16525SYuchung Cheng { 19285f16525SYuchung Cheng /* Initial receive window should be twice of TCP_INIT_CWND to 1939ef71e0cSWeiping Pan * enable proper sending of new unsent data during fast recovery 19485f16525SYuchung Cheng * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a 19585f16525SYuchung Cheng * limit when mss is larger than 1460. 19685f16525SYuchung Cheng */ 19785f16525SYuchung Cheng u32 init_rwnd = TCP_INIT_CWND * 2; 19885f16525SYuchung Cheng 19985f16525SYuchung Cheng if (mss > 1460) 20085f16525SYuchung Cheng init_rwnd = max((1460 * init_rwnd) / mss, 2U); 20185f16525SYuchung Cheng return init_rwnd; 20285f16525SYuchung Cheng } 20385f16525SYuchung Cheng 2041da177e4SLinus Torvalds /* Determine a window scaling and initial window to offer. 2051da177e4SLinus Torvalds * Based on the assumption that the given amount of space 2061da177e4SLinus Torvalds * will be offered. Store the results in the tp structure. 2071da177e4SLinus Torvalds * NOTE: for smooth operation initial space offering should 2081da177e4SLinus Torvalds * be a multiple of mss if possible. We assume here that mss >= 1. 2091da177e4SLinus Torvalds * This MUST be enforced by all callers. 2101da177e4SLinus Torvalds */ 2111da177e4SLinus Torvalds void tcp_select_initial_window(int __space, __u32 mss, 2121da177e4SLinus Torvalds __u32 *rcv_wnd, __u32 *window_clamp, 21331d12926Slaurent chavey int wscale_ok, __u8 *rcv_wscale, 21431d12926Slaurent chavey __u32 init_rcv_wnd) 2151da177e4SLinus Torvalds { 2161da177e4SLinus Torvalds unsigned int space = (__space < 0 ? 0 : __space); 2171da177e4SLinus Torvalds 2181da177e4SLinus Torvalds /* If no clamp set the clamp to the max possible scaled window */ 2191da177e4SLinus Torvalds if (*window_clamp == 0) 2201da177e4SLinus Torvalds (*window_clamp) = (65535 << 14); 2211da177e4SLinus Torvalds space = min(*window_clamp, space); 2221da177e4SLinus Torvalds 2231da177e4SLinus Torvalds /* Quantize space offering to a multiple of mss if possible. */ 2241da177e4SLinus Torvalds if (space > mss) 2251da177e4SLinus Torvalds space = (space / mss) * mss; 2261da177e4SLinus Torvalds 2271da177e4SLinus Torvalds /* NOTE: offering an initial window larger than 32767 22815d99e02SRick Jones * will break some buggy TCP stacks. If the admin tells us 22915d99e02SRick Jones * it is likely we could be speaking with such a buggy stack 23015d99e02SRick Jones * we will truncate our initial window offering to 32K-1 23115d99e02SRick Jones * unless the remote has sent us a window scaling option, 23215d99e02SRick Jones * which we interpret as a sign the remote TCP is not 23315d99e02SRick Jones * misinterpreting the window field as a signed quantity. 2341da177e4SLinus Torvalds */ 23515d99e02SRick Jones if (sysctl_tcp_workaround_signed_windows) 2361da177e4SLinus Torvalds (*rcv_wnd) = min(space, MAX_TCP_WINDOW); 23715d99e02SRick Jones else 23815d99e02SRick Jones (*rcv_wnd) = space; 23915d99e02SRick Jones 2401da177e4SLinus Torvalds (*rcv_wscale) = 0; 2411da177e4SLinus Torvalds if (wscale_ok) { 2421da177e4SLinus Torvalds /* Set window scaling on max possible window 2431da177e4SLinus Torvalds * See RFC1323 for an explanation of the limit to 14 2441da177e4SLinus Torvalds */ 2451da177e4SLinus Torvalds space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max); 246316c1592SStephen Hemminger space = min_t(u32, space, *window_clamp); 2471da177e4SLinus Torvalds while (space > 65535 && (*rcv_wscale) < 14) { 2481da177e4SLinus Torvalds space >>= 1; 2491da177e4SLinus Torvalds (*rcv_wscale)++; 2501da177e4SLinus Torvalds } 2511da177e4SLinus Torvalds } 2521da177e4SLinus Torvalds 2531da177e4SLinus Torvalds if (mss > (1 << *rcv_wscale)) { 25485f16525SYuchung Cheng if (!init_rcv_wnd) /* Use default unless specified otherwise */ 25585f16525SYuchung Cheng init_rcv_wnd = tcp_default_init_rwnd(mss); 256b1afde60SNandita Dukkipati *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss); 2571da177e4SLinus Torvalds } 2581da177e4SLinus Torvalds 2591da177e4SLinus Torvalds /* Set the clamp no higher than max representable value */ 2601da177e4SLinus Torvalds (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp); 2611da177e4SLinus Torvalds } 2624bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_select_initial_window); 2631da177e4SLinus Torvalds 2641da177e4SLinus Torvalds /* Chose a new window to advertise, update state in tcp_sock for the 2651da177e4SLinus Torvalds * socket, and return result with RFC1323 scaling applied. The return 2661da177e4SLinus Torvalds * value can be stuffed directly into th->window for an outgoing 2671da177e4SLinus Torvalds * frame. 2681da177e4SLinus Torvalds */ 26940efc6faSStephen Hemminger static u16 tcp_select_window(struct sock *sk) 2701da177e4SLinus Torvalds { 2711da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2728e165e20SFlorian Westphal u32 old_win = tp->rcv_wnd; 2731da177e4SLinus Torvalds u32 cur_win = tcp_receive_window(tp); 2741da177e4SLinus Torvalds u32 new_win = __tcp_select_window(sk); 2751da177e4SLinus Torvalds 2761da177e4SLinus Torvalds /* Never shrink the offered window */ 2771da177e4SLinus Torvalds if (new_win < cur_win) { 2781da177e4SLinus Torvalds /* Danger Will Robinson! 2791da177e4SLinus Torvalds * Don't update rcv_wup/rcv_wnd here or else 2801da177e4SLinus Torvalds * we will not be able to advertise a zero 2811da177e4SLinus Torvalds * window in time. --DaveM 2821da177e4SLinus Torvalds * 2831da177e4SLinus Torvalds * Relax Will Robinson. 2841da177e4SLinus Torvalds */ 2858e165e20SFlorian Westphal if (new_win == 0) 2868e165e20SFlorian Westphal NET_INC_STATS(sock_net(sk), 2878e165e20SFlorian Westphal LINUX_MIB_TCPWANTZEROWINDOWADV); 288607bfbf2SPatrick McHardy new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale); 2891da177e4SLinus Torvalds } 2901da177e4SLinus Torvalds tp->rcv_wnd = new_win; 2911da177e4SLinus Torvalds tp->rcv_wup = tp->rcv_nxt; 2921da177e4SLinus Torvalds 2931da177e4SLinus Torvalds /* Make sure we do not exceed the maximum possible 2941da177e4SLinus Torvalds * scaled window. 2951da177e4SLinus Torvalds */ 29615d99e02SRick Jones if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows) 2971da177e4SLinus Torvalds new_win = min(new_win, MAX_TCP_WINDOW); 2981da177e4SLinus Torvalds else 2991da177e4SLinus Torvalds new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); 3001da177e4SLinus Torvalds 3011da177e4SLinus Torvalds /* RFC1323 scaling applied */ 3021da177e4SLinus Torvalds new_win >>= tp->rx_opt.rcv_wscale; 3031da177e4SLinus Torvalds 3041da177e4SLinus Torvalds /* If we advertise zero window, disable fast path. */ 3058e165e20SFlorian Westphal if (new_win == 0) { 3061da177e4SLinus Torvalds tp->pred_flags = 0; 3078e165e20SFlorian Westphal if (old_win) 3088e165e20SFlorian Westphal NET_INC_STATS(sock_net(sk), 3098e165e20SFlorian Westphal LINUX_MIB_TCPTOZEROWINDOWADV); 3108e165e20SFlorian Westphal } else if (old_win == 0) { 3118e165e20SFlorian Westphal NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV); 3128e165e20SFlorian Westphal } 3131da177e4SLinus Torvalds 3141da177e4SLinus Torvalds return new_win; 3151da177e4SLinus Torvalds } 3161da177e4SLinus Torvalds 31767edfef7SAndi Kleen /* Packet ECN state for a SYN-ACK */ 318735d3831SFlorian Westphal static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb) 319bdf1ee5dSIlpo Järvinen { 32030e502a3SDaniel Borkmann const struct tcp_sock *tp = tcp_sk(sk); 32130e502a3SDaniel Borkmann 3224de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR; 323bdf1ee5dSIlpo Järvinen if (!(tp->ecn_flags & TCP_ECN_OK)) 3244de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE; 32530e502a3SDaniel Borkmann else if (tcp_ca_needs_ecn(sk)) 32630e502a3SDaniel Borkmann INET_ECN_xmit(sk); 327bdf1ee5dSIlpo Järvinen } 328bdf1ee5dSIlpo Järvinen 32967edfef7SAndi Kleen /* Packet ECN state for a SYN. */ 330735d3831SFlorian Westphal static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb) 331bdf1ee5dSIlpo Järvinen { 332bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 333f7b3bec6SFlorian Westphal bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 || 334f7b3bec6SFlorian Westphal tcp_ca_needs_ecn(sk); 335f7b3bec6SFlorian Westphal 336f7b3bec6SFlorian Westphal if (!use_ecn) { 337f7b3bec6SFlorian Westphal const struct dst_entry *dst = __sk_dst_get(sk); 338f7b3bec6SFlorian Westphal 339f7b3bec6SFlorian Westphal if (dst && dst_feature(dst, RTAX_FEATURE_ECN)) 340f7b3bec6SFlorian Westphal use_ecn = true; 341f7b3bec6SFlorian Westphal } 342bdf1ee5dSIlpo Järvinen 343bdf1ee5dSIlpo Järvinen tp->ecn_flags = 0; 344f7b3bec6SFlorian Westphal 345f7b3bec6SFlorian Westphal if (use_ecn) { 3464de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR; 347bdf1ee5dSIlpo Järvinen tp->ecn_flags = TCP_ECN_OK; 34830e502a3SDaniel Borkmann if (tcp_ca_needs_ecn(sk)) 34930e502a3SDaniel Borkmann INET_ECN_xmit(sk); 350bdf1ee5dSIlpo Järvinen } 351bdf1ee5dSIlpo Järvinen } 352bdf1ee5dSIlpo Järvinen 353735d3831SFlorian Westphal static void 354735d3831SFlorian Westphal tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th, 35530e502a3SDaniel Borkmann struct sock *sk) 356bdf1ee5dSIlpo Järvinen { 35730e502a3SDaniel Borkmann if (inet_rsk(req)->ecn_ok) { 358bdf1ee5dSIlpo Järvinen th->ece = 1; 35930e502a3SDaniel Borkmann if (tcp_ca_needs_ecn(sk)) 36030e502a3SDaniel Borkmann INET_ECN_xmit(sk); 36130e502a3SDaniel Borkmann } 362bdf1ee5dSIlpo Järvinen } 363bdf1ee5dSIlpo Järvinen 36467edfef7SAndi Kleen /* Set up ECN state for a packet on a ESTABLISHED socket that is about to 36567edfef7SAndi Kleen * be sent. 36667edfef7SAndi Kleen */ 367735d3831SFlorian Westphal static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb, 368bdf1ee5dSIlpo Järvinen int tcp_header_len) 369bdf1ee5dSIlpo Järvinen { 370bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 371bdf1ee5dSIlpo Järvinen 372bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_OK) { 373bdf1ee5dSIlpo Järvinen /* Not-retransmitted data segment: set ECT and inject CWR. */ 374bdf1ee5dSIlpo Järvinen if (skb->len != tcp_header_len && 375bdf1ee5dSIlpo Järvinen !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { 376bdf1ee5dSIlpo Järvinen INET_ECN_xmit(sk); 377bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) { 378bdf1ee5dSIlpo Järvinen tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; 379bdf1ee5dSIlpo Järvinen tcp_hdr(skb)->cwr = 1; 380bdf1ee5dSIlpo Järvinen skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; 381bdf1ee5dSIlpo Järvinen } 38230e502a3SDaniel Borkmann } else if (!tcp_ca_needs_ecn(sk)) { 383bdf1ee5dSIlpo Järvinen /* ACK or retransmitted segment: clear ECT|CE */ 384bdf1ee5dSIlpo Järvinen INET_ECN_dontxmit(sk); 385bdf1ee5dSIlpo Järvinen } 386bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) 387bdf1ee5dSIlpo Järvinen tcp_hdr(skb)->ece = 1; 388bdf1ee5dSIlpo Järvinen } 389bdf1ee5dSIlpo Järvinen } 390bdf1ee5dSIlpo Järvinen 391e870a8efSIlpo Järvinen /* Constructs common control bits of non-data skb. If SYN/FIN is present, 392e870a8efSIlpo Järvinen * auto increment end seqno. 393e870a8efSIlpo Järvinen */ 394e870a8efSIlpo Järvinen static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags) 395e870a8efSIlpo Järvinen { 3967b7fc97aSEric Dumazet struct skb_shared_info *shinfo = skb_shinfo(skb); 3977b7fc97aSEric Dumazet 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); 4057b7fc97aSEric Dumazet shinfo->gso_size = 0; 4067b7fc97aSEric Dumazet shinfo->gso_type = 0; 407e870a8efSIlpo Järvinen 408e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->seq = seq; 409a3433f35SChangli Gao if (flags & (TCPHDR_SYN | TCPHDR_FIN)) 410e870a8efSIlpo Järvinen seq++; 411e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->end_seq = seq; 412e870a8efSIlpo Järvinen } 413e870a8efSIlpo Järvinen 414a2a385d6SEric Dumazet static inline bool tcp_urg_mode(const struct tcp_sock *tp) 41533f5f57eSIlpo Järvinen { 41633f5f57eSIlpo Järvinen return tp->snd_una != tp->snd_up; 41733f5f57eSIlpo Järvinen } 41833f5f57eSIlpo Järvinen 41933ad798cSAdam Langley #define OPTION_SACK_ADVERTISE (1 << 0) 42033ad798cSAdam Langley #define OPTION_TS (1 << 1) 42133ad798cSAdam Langley #define OPTION_MD5 (1 << 2) 42289e95a61SOri Finkelman #define OPTION_WSCALE (1 << 3) 4232100c8d2SYuchung Cheng #define OPTION_FAST_OPEN_COOKIE (1 << 8) 42433ad798cSAdam Langley 42533ad798cSAdam Langley struct tcp_out_options { 4262100c8d2SYuchung Cheng u16 options; /* bit field of OPTION_* */ 4272100c8d2SYuchung Cheng u16 mss; /* 0 to disable */ 42833ad798cSAdam Langley u8 ws; /* window scale, 0 to disable */ 42933ad798cSAdam Langley u8 num_sack_blocks; /* number of SACK blocks to include */ 430bd0388aeSWilliam Allen Simpson u8 hash_size; /* bytes in hash_location */ 431bd0388aeSWilliam Allen Simpson __u8 *hash_location; /* temporary pointer, overloaded */ 4322100c8d2SYuchung Cheng __u32 tsval, tsecr; /* need to include OPTION_TS */ 4332100c8d2SYuchung Cheng struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */ 43433ad798cSAdam Langley }; 43533ad798cSAdam Langley 43667edfef7SAndi Kleen /* Write previously computed TCP options to the packet. 43767edfef7SAndi Kleen * 43867edfef7SAndi Kleen * Beware: Something in the Internet is very sensitive to the ordering of 439fd6149d3SIlpo Järvinen * TCP options, we learned this through the hard way, so be careful here. 440fd6149d3SIlpo Järvinen * Luckily we can at least blame others for their non-compliance but from 4418e3bff96Sstephen hemminger * inter-operability perspective it seems that we're somewhat stuck with 442fd6149d3SIlpo Järvinen * the ordering which we have been using if we want to keep working with 443fd6149d3SIlpo Järvinen * those broken things (not that it currently hurts anybody as there isn't 444fd6149d3SIlpo Järvinen * particular reason why the ordering would need to be changed). 445fd6149d3SIlpo Järvinen * 446fd6149d3SIlpo Järvinen * At least SACK_PERM as the first option is known to lead to a disaster 447fd6149d3SIlpo Järvinen * (but it may well be that other scenarios fail similarly). 448fd6149d3SIlpo Järvinen */ 44933ad798cSAdam Langley static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp, 450bd0388aeSWilliam Allen Simpson struct tcp_out_options *opts) 451bd0388aeSWilliam Allen Simpson { 4522100c8d2SYuchung Cheng u16 options = opts->options; /* mungable copy */ 453bd0388aeSWilliam Allen Simpson 454bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_MD5 & options)) { 4551a2c6181SChristoph Paasch *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | 4561a2c6181SChristoph Paasch (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG); 457bd0388aeSWilliam Allen Simpson /* overload cookie hash location */ 458bd0388aeSWilliam Allen Simpson opts->hash_location = (__u8 *)ptr; 45933ad798cSAdam Langley ptr += 4; 46033ad798cSAdam Langley } 46133ad798cSAdam Langley 462fd6149d3SIlpo Järvinen if (unlikely(opts->mss)) { 463fd6149d3SIlpo Järvinen *ptr++ = htonl((TCPOPT_MSS << 24) | 464fd6149d3SIlpo Järvinen (TCPOLEN_MSS << 16) | 465fd6149d3SIlpo Järvinen opts->mss); 466fd6149d3SIlpo Järvinen } 467fd6149d3SIlpo Järvinen 468bd0388aeSWilliam Allen Simpson if (likely(OPTION_TS & options)) { 469bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 47033ad798cSAdam Langley *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | 47133ad798cSAdam Langley (TCPOLEN_SACK_PERM << 16) | 47233ad798cSAdam Langley (TCPOPT_TIMESTAMP << 8) | 47333ad798cSAdam Langley TCPOLEN_TIMESTAMP); 474bd0388aeSWilliam Allen Simpson options &= ~OPTION_SACK_ADVERTISE; 47533ad798cSAdam Langley } else { 476496c98dfSYOSHIFUJI Hideaki *ptr++ = htonl((TCPOPT_NOP << 24) | 47740efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 47840efc6faSStephen Hemminger (TCPOPT_TIMESTAMP << 8) | 47940efc6faSStephen Hemminger TCPOLEN_TIMESTAMP); 48040efc6faSStephen Hemminger } 48133ad798cSAdam Langley *ptr++ = htonl(opts->tsval); 48233ad798cSAdam Langley *ptr++ = htonl(opts->tsecr); 48333ad798cSAdam Langley } 48433ad798cSAdam Langley 485bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 48633ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 48733ad798cSAdam Langley (TCPOPT_NOP << 16) | 48833ad798cSAdam Langley (TCPOPT_SACK_PERM << 8) | 48933ad798cSAdam Langley TCPOLEN_SACK_PERM); 49033ad798cSAdam Langley } 49133ad798cSAdam Langley 492bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_WSCALE & options)) { 49333ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 49433ad798cSAdam Langley (TCPOPT_WINDOW << 16) | 49533ad798cSAdam Langley (TCPOLEN_WINDOW << 8) | 49633ad798cSAdam Langley opts->ws); 49733ad798cSAdam Langley } 49833ad798cSAdam Langley 49933ad798cSAdam Langley if (unlikely(opts->num_sack_blocks)) { 50033ad798cSAdam Langley struct tcp_sack_block *sp = tp->rx_opt.dsack ? 50133ad798cSAdam Langley tp->duplicate_sack : tp->selective_acks; 50240efc6faSStephen Hemminger int this_sack; 50340efc6faSStephen Hemminger 50440efc6faSStephen Hemminger *ptr++ = htonl((TCPOPT_NOP << 24) | 50540efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 50640efc6faSStephen Hemminger (TCPOPT_SACK << 8) | 50733ad798cSAdam Langley (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * 50840efc6faSStephen Hemminger TCPOLEN_SACK_PERBLOCK))); 5092de979bdSStephen Hemminger 51033ad798cSAdam Langley for (this_sack = 0; this_sack < opts->num_sack_blocks; 51133ad798cSAdam Langley ++this_sack) { 51240efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].start_seq); 51340efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].end_seq); 51440efc6faSStephen Hemminger } 5152de979bdSStephen Hemminger 51640efc6faSStephen Hemminger tp->rx_opt.dsack = 0; 51740efc6faSStephen Hemminger } 5182100c8d2SYuchung Cheng 5192100c8d2SYuchung Cheng if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) { 5202100c8d2SYuchung Cheng struct tcp_fastopen_cookie *foc = opts->fastopen_cookie; 5212100c8d2SYuchung Cheng 5222100c8d2SYuchung Cheng *ptr++ = htonl((TCPOPT_EXP << 24) | 5232100c8d2SYuchung Cheng ((TCPOLEN_EXP_FASTOPEN_BASE + foc->len) << 16) | 5242100c8d2SYuchung Cheng TCPOPT_FASTOPEN_MAGIC); 5252100c8d2SYuchung Cheng 5262100c8d2SYuchung Cheng memcpy(ptr, foc->val, foc->len); 5272100c8d2SYuchung Cheng if ((foc->len & 3) == 2) { 5282100c8d2SYuchung Cheng u8 *align = ((u8 *)ptr) + foc->len; 5292100c8d2SYuchung Cheng align[0] = align[1] = TCPOPT_NOP; 5302100c8d2SYuchung Cheng } 5312100c8d2SYuchung Cheng ptr += (foc->len + 3) >> 2; 5322100c8d2SYuchung Cheng } 53340efc6faSStephen Hemminger } 53440efc6faSStephen Hemminger 53567edfef7SAndi Kleen /* Compute TCP options for SYN packets. This is not the final 53667edfef7SAndi Kleen * network wire format yet. 53767edfef7SAndi Kleen */ 53895c96174SEric Dumazet static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb, 53933ad798cSAdam Langley struct tcp_out_options *opts, 540cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 541cf533ea5SEric Dumazet { 54233ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 54395c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 544783237e8SYuchung Cheng struct tcp_fastopen_request *fastopen = tp->fastopen_req; 54533ad798cSAdam Langley 546cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 54733ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 54833ad798cSAdam Langley if (*md5) { 54933ad798cSAdam Langley opts->options |= OPTION_MD5; 550bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 551cfb6eeb4SYOSHIFUJI Hideaki } 55233ad798cSAdam Langley #else 55333ad798cSAdam Langley *md5 = NULL; 554cfb6eeb4SYOSHIFUJI Hideaki #endif 55533ad798cSAdam Langley 55633ad798cSAdam Langley /* We always get an MSS option. The option bytes which will be seen in 55733ad798cSAdam Langley * normal data packets should timestamps be used, must be in the MSS 55833ad798cSAdam Langley * advertised. But we subtract them from tp->mss_cache so that 55933ad798cSAdam Langley * calculations in tcp_sendmsg are simpler etc. So account for this 56033ad798cSAdam Langley * fact here if necessary. If we don't do this correctly, as a 56133ad798cSAdam Langley * receiver we won't recognize data packets as being full sized when we 56233ad798cSAdam Langley * should, and thus we won't abide by the delayed ACK rules correctly. 56333ad798cSAdam Langley * SACKs don't matter, we never delay an ACK when we have any of those 56433ad798cSAdam Langley * going out. */ 56533ad798cSAdam Langley opts->mss = tcp_advertise_mss(sk); 566bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 56733ad798cSAdam Langley 568bb5b7c11SDavid S. Miller if (likely(sysctl_tcp_timestamps && *md5 == NULL)) { 56933ad798cSAdam Langley opts->options |= OPTION_TS; 5707faee5c0SEric Dumazet opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset; 57133ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 572bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 57333ad798cSAdam Langley } 574bb5b7c11SDavid S. Miller if (likely(sysctl_tcp_window_scaling)) { 57533ad798cSAdam Langley opts->ws = tp->rx_opt.rcv_wscale; 57689e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 577bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 57833ad798cSAdam Langley } 579bb5b7c11SDavid S. Miller if (likely(sysctl_tcp_sack)) { 58033ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 581b32d1310SDavid S. Miller if (unlikely(!(OPTION_TS & opts->options))) 582bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 58333ad798cSAdam Langley } 58433ad798cSAdam Langley 585783237e8SYuchung Cheng if (fastopen && fastopen->cookie.len >= 0) { 586783237e8SYuchung Cheng u32 need = TCPOLEN_EXP_FASTOPEN_BASE + fastopen->cookie.len; 587783237e8SYuchung Cheng need = (need + 3) & ~3U; /* Align to 32 bits */ 588783237e8SYuchung Cheng if (remaining >= need) { 589783237e8SYuchung Cheng opts->options |= OPTION_FAST_OPEN_COOKIE; 590783237e8SYuchung Cheng opts->fastopen_cookie = &fastopen->cookie; 591783237e8SYuchung Cheng remaining -= need; 592783237e8SYuchung Cheng tp->syn_fastopen = 1; 593783237e8SYuchung Cheng } 594783237e8SYuchung Cheng } 595bd0388aeSWilliam Allen Simpson 596bd0388aeSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 59733ad798cSAdam Langley } 59833ad798cSAdam Langley 59967edfef7SAndi Kleen /* Set up TCP options for SYN-ACKs. */ 60095c96174SEric Dumazet static unsigned int tcp_synack_options(struct sock *sk, 60133ad798cSAdam Langley struct request_sock *req, 60295c96174SEric Dumazet unsigned int mss, struct sk_buff *skb, 60333ad798cSAdam Langley struct tcp_out_options *opts, 6044957faadSWilliam Allen Simpson struct tcp_md5sig_key **md5, 6058336886fSJerry Chu struct tcp_fastopen_cookie *foc) 6064957faadSWilliam Allen Simpson { 60733ad798cSAdam Langley struct inet_request_sock *ireq = inet_rsk(req); 60895c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 60933ad798cSAdam Langley 61033ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 61133ad798cSAdam Langley *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req); 61233ad798cSAdam Langley if (*md5) { 61333ad798cSAdam Langley opts->options |= OPTION_MD5; 6144957faadSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 6154957faadSWilliam Allen Simpson 6164957faadSWilliam Allen Simpson /* We can't fit any SACK blocks in a packet with MD5 + TS 6174957faadSWilliam Allen Simpson * options. There was discussion about disabling SACK 6184957faadSWilliam Allen Simpson * rather than TS in order to fit in better with old, 6194957faadSWilliam Allen Simpson * buggy kernels, but that was deemed to be unnecessary. 6204957faadSWilliam Allen Simpson */ 621de213e5eSEric Dumazet ireq->tstamp_ok &= !ireq->sack_ok; 62233ad798cSAdam Langley } 62333ad798cSAdam Langley #else 62433ad798cSAdam Langley *md5 = NULL; 62533ad798cSAdam Langley #endif 62633ad798cSAdam Langley 6274957faadSWilliam Allen Simpson /* We always send an MSS option. */ 62833ad798cSAdam Langley opts->mss = mss; 6294957faadSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 63033ad798cSAdam Langley 63133ad798cSAdam Langley if (likely(ireq->wscale_ok)) { 63233ad798cSAdam Langley opts->ws = ireq->rcv_wscale; 63389e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 6344957faadSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 63533ad798cSAdam Langley } 636de213e5eSEric Dumazet if (likely(ireq->tstamp_ok)) { 63733ad798cSAdam Langley opts->options |= OPTION_TS; 6387faee5c0SEric Dumazet opts->tsval = tcp_skb_timestamp(skb); 63933ad798cSAdam Langley opts->tsecr = req->ts_recent; 6404957faadSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 64133ad798cSAdam Langley } 64233ad798cSAdam Langley if (likely(ireq->sack_ok)) { 64333ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 644de213e5eSEric Dumazet if (unlikely(!ireq->tstamp_ok)) 6454957faadSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 64633ad798cSAdam Langley } 64789278c9dSYuchung Cheng if (foc != NULL && foc->len >= 0) { 6488336886fSJerry Chu u32 need = TCPOLEN_EXP_FASTOPEN_BASE + foc->len; 6498336886fSJerry Chu need = (need + 3) & ~3U; /* Align to 32 bits */ 6508336886fSJerry Chu if (remaining >= need) { 6518336886fSJerry Chu opts->options |= OPTION_FAST_OPEN_COOKIE; 6528336886fSJerry Chu opts->fastopen_cookie = foc; 6538336886fSJerry Chu remaining -= need; 6548336886fSJerry Chu } 6558336886fSJerry Chu } 6564957faadSWilliam Allen Simpson 6574957faadSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 65833ad798cSAdam Langley } 65933ad798cSAdam Langley 66067edfef7SAndi Kleen /* Compute TCP options for ESTABLISHED sockets. This is not the 66167edfef7SAndi Kleen * final wire format yet. 66267edfef7SAndi Kleen */ 66395c96174SEric Dumazet static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb, 66433ad798cSAdam Langley struct tcp_out_options *opts, 665cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 666cf533ea5SEric Dumazet { 66733ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 66895c96174SEric Dumazet unsigned int size = 0; 669cabeccbdSIlpo Järvinen unsigned int eff_sacks; 67033ad798cSAdam Langley 6715843ef42SAndi Kleen opts->options = 0; 6725843ef42SAndi Kleen 67333ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 67433ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 67533ad798cSAdam Langley if (unlikely(*md5)) { 67633ad798cSAdam Langley opts->options |= OPTION_MD5; 67733ad798cSAdam Langley size += TCPOLEN_MD5SIG_ALIGNED; 67833ad798cSAdam Langley } 67933ad798cSAdam Langley #else 68033ad798cSAdam Langley *md5 = NULL; 68133ad798cSAdam Langley #endif 68233ad798cSAdam Langley 68333ad798cSAdam Langley if (likely(tp->rx_opt.tstamp_ok)) { 68433ad798cSAdam Langley opts->options |= OPTION_TS; 6857faee5c0SEric Dumazet opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0; 68633ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 68733ad798cSAdam Langley size += TCPOLEN_TSTAMP_ALIGNED; 68833ad798cSAdam Langley } 68933ad798cSAdam Langley 690cabeccbdSIlpo Järvinen eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; 691cabeccbdSIlpo Järvinen if (unlikely(eff_sacks)) { 69295c96174SEric Dumazet const unsigned int remaining = MAX_TCP_OPTION_SPACE - size; 69333ad798cSAdam Langley opts->num_sack_blocks = 69495c96174SEric Dumazet min_t(unsigned int, eff_sacks, 69533ad798cSAdam Langley (remaining - TCPOLEN_SACK_BASE_ALIGNED) / 69633ad798cSAdam Langley TCPOLEN_SACK_PERBLOCK); 69733ad798cSAdam Langley size += TCPOLEN_SACK_BASE_ALIGNED + 69833ad798cSAdam Langley opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; 69933ad798cSAdam Langley } 70033ad798cSAdam Langley 70133ad798cSAdam Langley return size; 70240efc6faSStephen Hemminger } 7031da177e4SLinus Torvalds 70446d3ceabSEric Dumazet 70546d3ceabSEric Dumazet /* TCP SMALL QUEUES (TSQ) 70646d3ceabSEric Dumazet * 70746d3ceabSEric Dumazet * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev) 70846d3ceabSEric Dumazet * to reduce RTT and bufferbloat. 70946d3ceabSEric Dumazet * We do this using a special skb destructor (tcp_wfree). 71046d3ceabSEric Dumazet * 71146d3ceabSEric Dumazet * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb 71246d3ceabSEric Dumazet * needs to be reallocated in a driver. 7138e3bff96Sstephen hemminger * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc 71446d3ceabSEric Dumazet * 71546d3ceabSEric Dumazet * Since transmit from skb destructor is forbidden, we use a tasklet 71646d3ceabSEric Dumazet * to process all sockets that eventually need to send more skbs. 71746d3ceabSEric Dumazet * We use one tasklet per cpu, with its own queue of sockets. 71846d3ceabSEric Dumazet */ 71946d3ceabSEric Dumazet struct tsq_tasklet { 72046d3ceabSEric Dumazet struct tasklet_struct tasklet; 72146d3ceabSEric Dumazet struct list_head head; /* queue of tcp sockets */ 72246d3ceabSEric Dumazet }; 72346d3ceabSEric Dumazet static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet); 72446d3ceabSEric Dumazet 7256f458dfbSEric Dumazet static void tcp_tsq_handler(struct sock *sk) 7266f458dfbSEric Dumazet { 7276f458dfbSEric Dumazet if ((1 << sk->sk_state) & 7286f458dfbSEric Dumazet (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING | 7296f458dfbSEric Dumazet TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) 730bf06200eSJohn Ogness tcp_write_xmit(sk, tcp_current_mss(sk), tcp_sk(sk)->nonagle, 731bf06200eSJohn Ogness 0, GFP_ATOMIC); 7326f458dfbSEric Dumazet } 73346d3ceabSEric Dumazet /* 7348e3bff96Sstephen hemminger * One tasklet per cpu tries to send more skbs. 73546d3ceabSEric Dumazet * We run in tasklet context but need to disable irqs when 7368e3bff96Sstephen hemminger * transferring tsq->head because tcp_wfree() might 73746d3ceabSEric Dumazet * interrupt us (non NAPI drivers) 73846d3ceabSEric Dumazet */ 73946d3ceabSEric Dumazet static void tcp_tasklet_func(unsigned long data) 74046d3ceabSEric Dumazet { 74146d3ceabSEric Dumazet struct tsq_tasklet *tsq = (struct tsq_tasklet *)data; 74246d3ceabSEric Dumazet LIST_HEAD(list); 74346d3ceabSEric Dumazet unsigned long flags; 74446d3ceabSEric Dumazet struct list_head *q, *n; 74546d3ceabSEric Dumazet struct tcp_sock *tp; 74646d3ceabSEric Dumazet struct sock *sk; 74746d3ceabSEric Dumazet 74846d3ceabSEric Dumazet local_irq_save(flags); 74946d3ceabSEric Dumazet list_splice_init(&tsq->head, &list); 75046d3ceabSEric Dumazet local_irq_restore(flags); 75146d3ceabSEric Dumazet 75246d3ceabSEric Dumazet list_for_each_safe(q, n, &list) { 75346d3ceabSEric Dumazet tp = list_entry(q, struct tcp_sock, tsq_node); 75446d3ceabSEric Dumazet list_del(&tp->tsq_node); 75546d3ceabSEric Dumazet 75646d3ceabSEric Dumazet sk = (struct sock *)tp; 75746d3ceabSEric Dumazet bh_lock_sock(sk); 75846d3ceabSEric Dumazet 75946d3ceabSEric Dumazet if (!sock_owned_by_user(sk)) { 7606f458dfbSEric Dumazet tcp_tsq_handler(sk); 76146d3ceabSEric Dumazet } else { 76246d3ceabSEric Dumazet /* defer the work to tcp_release_cb() */ 7636f458dfbSEric Dumazet set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags); 76446d3ceabSEric Dumazet } 76546d3ceabSEric Dumazet bh_unlock_sock(sk); 76646d3ceabSEric Dumazet 76746d3ceabSEric Dumazet clear_bit(TSQ_QUEUED, &tp->tsq_flags); 76846d3ceabSEric Dumazet sk_free(sk); 76946d3ceabSEric Dumazet } 77046d3ceabSEric Dumazet } 77146d3ceabSEric Dumazet 7726f458dfbSEric Dumazet #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \ 7736f458dfbSEric Dumazet (1UL << TCP_WRITE_TIMER_DEFERRED) | \ 774563d34d0SEric Dumazet (1UL << TCP_DELACK_TIMER_DEFERRED) | \ 775563d34d0SEric Dumazet (1UL << TCP_MTU_REDUCED_DEFERRED)) 77646d3ceabSEric Dumazet /** 77746d3ceabSEric Dumazet * tcp_release_cb - tcp release_sock() callback 77846d3ceabSEric Dumazet * @sk: socket 77946d3ceabSEric Dumazet * 78046d3ceabSEric Dumazet * called from release_sock() to perform protocol dependent 78146d3ceabSEric Dumazet * actions before socket release. 78246d3ceabSEric Dumazet */ 78346d3ceabSEric Dumazet void tcp_release_cb(struct sock *sk) 78446d3ceabSEric Dumazet { 78546d3ceabSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 7866f458dfbSEric Dumazet unsigned long flags, nflags; 78746d3ceabSEric Dumazet 7886f458dfbSEric Dumazet /* perform an atomic operation only if at least one flag is set */ 7896f458dfbSEric Dumazet do { 7906f458dfbSEric Dumazet flags = tp->tsq_flags; 7916f458dfbSEric Dumazet if (!(flags & TCP_DEFERRED_ALL)) 7926f458dfbSEric Dumazet return; 7936f458dfbSEric Dumazet nflags = flags & ~TCP_DEFERRED_ALL; 7946f458dfbSEric Dumazet } while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags); 7956f458dfbSEric Dumazet 7966f458dfbSEric Dumazet if (flags & (1UL << TCP_TSQ_DEFERRED)) 7976f458dfbSEric Dumazet tcp_tsq_handler(sk); 7986f458dfbSEric Dumazet 799c3f9b018SEric Dumazet /* Here begins the tricky part : 800c3f9b018SEric Dumazet * We are called from release_sock() with : 801c3f9b018SEric Dumazet * 1) BH disabled 802c3f9b018SEric Dumazet * 2) sk_lock.slock spinlock held 803c3f9b018SEric Dumazet * 3) socket owned by us (sk->sk_lock.owned == 1) 804c3f9b018SEric Dumazet * 805c3f9b018SEric Dumazet * But following code is meant to be called from BH handlers, 806c3f9b018SEric Dumazet * so we should keep BH disabled, but early release socket ownership 807c3f9b018SEric Dumazet */ 808c3f9b018SEric Dumazet sock_release_ownership(sk); 809c3f9b018SEric Dumazet 810144d56e9SEric Dumazet if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) { 8116f458dfbSEric Dumazet tcp_write_timer_handler(sk); 812144d56e9SEric Dumazet __sock_put(sk); 813144d56e9SEric Dumazet } 814144d56e9SEric Dumazet if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) { 8156f458dfbSEric Dumazet tcp_delack_timer_handler(sk); 816144d56e9SEric Dumazet __sock_put(sk); 817144d56e9SEric Dumazet } 818144d56e9SEric Dumazet if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) { 8194fab9071SNeal Cardwell inet_csk(sk)->icsk_af_ops->mtu_reduced(sk); 820144d56e9SEric Dumazet __sock_put(sk); 821144d56e9SEric Dumazet } 82246d3ceabSEric Dumazet } 82346d3ceabSEric Dumazet EXPORT_SYMBOL(tcp_release_cb); 82446d3ceabSEric Dumazet 82546d3ceabSEric Dumazet void __init tcp_tasklet_init(void) 82646d3ceabSEric Dumazet { 82746d3ceabSEric Dumazet int i; 82846d3ceabSEric Dumazet 82946d3ceabSEric Dumazet for_each_possible_cpu(i) { 83046d3ceabSEric Dumazet struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i); 83146d3ceabSEric Dumazet 83246d3ceabSEric Dumazet INIT_LIST_HEAD(&tsq->head); 83346d3ceabSEric Dumazet tasklet_init(&tsq->tasklet, 83446d3ceabSEric Dumazet tcp_tasklet_func, 83546d3ceabSEric Dumazet (unsigned long)tsq); 83646d3ceabSEric Dumazet } 83746d3ceabSEric Dumazet } 83846d3ceabSEric Dumazet 83946d3ceabSEric Dumazet /* 84046d3ceabSEric Dumazet * Write buffer destructor automatically called from kfree_skb. 8418e3bff96Sstephen hemminger * We can't xmit new skbs from this context, as we might already 84246d3ceabSEric Dumazet * hold qdisc lock. 84346d3ceabSEric Dumazet */ 844d6a4a104SEric Dumazet void tcp_wfree(struct sk_buff *skb) 84546d3ceabSEric Dumazet { 84646d3ceabSEric Dumazet struct sock *sk = skb->sk; 84746d3ceabSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 8489b462d02SEric Dumazet int wmem; 8499b462d02SEric Dumazet 8509b462d02SEric Dumazet /* Keep one reference on sk_wmem_alloc. 8519b462d02SEric Dumazet * Will be released by sk_free() from here or tcp_tasklet_func() 8529b462d02SEric Dumazet */ 8539b462d02SEric Dumazet wmem = atomic_sub_return(skb->truesize - 1, &sk->sk_wmem_alloc); 8549b462d02SEric Dumazet 8559b462d02SEric Dumazet /* If this softirq is serviced by ksoftirqd, we are likely under stress. 8569b462d02SEric Dumazet * Wait until our queues (qdisc + devices) are drained. 8579b462d02SEric Dumazet * This gives : 8589b462d02SEric Dumazet * - less callbacks to tcp_write_xmit(), reducing stress (batches) 8599b462d02SEric Dumazet * - chance for incoming ACK (processed by another cpu maybe) 8609b462d02SEric Dumazet * to migrate this flow (skb->ooo_okay will be eventually set) 8619b462d02SEric Dumazet */ 8629b462d02SEric Dumazet if (wmem >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current) 8639b462d02SEric Dumazet goto out; 86446d3ceabSEric Dumazet 86546d3ceabSEric Dumazet if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) && 86646d3ceabSEric Dumazet !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) { 86746d3ceabSEric Dumazet unsigned long flags; 86846d3ceabSEric Dumazet struct tsq_tasklet *tsq; 86946d3ceabSEric Dumazet 87046d3ceabSEric Dumazet /* queue this socket to tasklet queue */ 87146d3ceabSEric Dumazet local_irq_save(flags); 872903ceff7SChristoph Lameter tsq = this_cpu_ptr(&tsq_tasklet); 87346d3ceabSEric Dumazet list_add(&tp->tsq_node, &tsq->head); 87446d3ceabSEric Dumazet tasklet_schedule(&tsq->tasklet); 87546d3ceabSEric Dumazet local_irq_restore(flags); 8769b462d02SEric Dumazet return; 87746d3ceabSEric Dumazet } 8789b462d02SEric Dumazet out: 8799b462d02SEric Dumazet sk_free(sk); 88046d3ceabSEric Dumazet } 88146d3ceabSEric Dumazet 8821da177e4SLinus Torvalds /* This routine actually transmits TCP packets queued in by 8831da177e4SLinus Torvalds * tcp_do_sendmsg(). This is used by both the initial 8841da177e4SLinus Torvalds * transmission and possible later retransmissions. 8851da177e4SLinus Torvalds * All SKB's seen here are completely headerless. It is our 8861da177e4SLinus Torvalds * job to build the TCP header, and pass the packet down to 8871da177e4SLinus Torvalds * IP so it can do the same plus pass the packet off to the 8881da177e4SLinus Torvalds * device. 8891da177e4SLinus Torvalds * 8901da177e4SLinus Torvalds * We are working here with either a clone of the original 8911da177e4SLinus Torvalds * SKB, or a fresh unique copy made by the retransmit engine. 8921da177e4SLinus Torvalds */ 893056834d9SIlpo Järvinen static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, 894056834d9SIlpo Järvinen gfp_t gfp_mask) 8951da177e4SLinus Torvalds { 8966687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 897dfb4b9dcSDavid S. Miller struct inet_sock *inet; 898dfb4b9dcSDavid S. Miller struct tcp_sock *tp; 899dfb4b9dcSDavid S. Miller struct tcp_skb_cb *tcb; 90033ad798cSAdam Langley struct tcp_out_options opts; 90195c96174SEric Dumazet unsigned int tcp_options_size, tcp_header_size; 902cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 9031da177e4SLinus Torvalds struct tcphdr *th; 9041da177e4SLinus Torvalds int err; 9051da177e4SLinus Torvalds 906dfb4b9dcSDavid S. Miller BUG_ON(!skb || !tcp_skb_pcount(skb)); 907dfb4b9dcSDavid S. Miller 908ccdbb6e9SEric Dumazet if (clone_it) { 909740b0f18SEric Dumazet skb_mstamp_get(&skb->skb_mstamp); 910dfb4b9dcSDavid S. Miller 911dfb4b9dcSDavid S. Miller if (unlikely(skb_cloned(skb))) 912dfb4b9dcSDavid S. Miller skb = pskb_copy(skb, gfp_mask); 913dfb4b9dcSDavid S. Miller else 914dfb4b9dcSDavid S. Miller skb = skb_clone(skb, gfp_mask); 915dfb4b9dcSDavid S. Miller if (unlikely(!skb)) 916dfb4b9dcSDavid S. Miller return -ENOBUFS; 917dfb4b9dcSDavid S. Miller } 918dfb4b9dcSDavid S. Miller 919dfb4b9dcSDavid S. Miller inet = inet_sk(sk); 920dfb4b9dcSDavid S. Miller tp = tcp_sk(sk); 921dfb4b9dcSDavid S. Miller tcb = TCP_SKB_CB(skb); 92233ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 9231da177e4SLinus Torvalds 9244de075e0SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) 92533ad798cSAdam Langley tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); 92633ad798cSAdam Langley else 92733ad798cSAdam Langley tcp_options_size = tcp_established_options(sk, skb, &opts, 92833ad798cSAdam Langley &md5); 92933ad798cSAdam Langley tcp_header_size = tcp_options_size + sizeof(struct tcphdr); 9301da177e4SLinus Torvalds 931547669d4SEric Dumazet if (tcp_packets_in_flight(tp) == 0) 9326687e988SArnaldo Carvalho de Melo tcp_ca_event(sk, CA_EVENT_TX_START); 933547669d4SEric Dumazet 934547669d4SEric Dumazet /* if no packet is in qdisc/device queue, then allow XPS to select 935b2532eb9SEric Dumazet * another queue. We can be called from tcp_tsq_handler() 936b2532eb9SEric Dumazet * which holds one reference to sk_wmem_alloc. 937b2532eb9SEric Dumazet * 938b2532eb9SEric Dumazet * TODO: Ideally, in-flight pure ACK packets should not matter here. 939b2532eb9SEric Dumazet * One way to get this would be to set skb->truesize = 2 on them. 940547669d4SEric Dumazet */ 941b2532eb9SEric Dumazet skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1); 9421da177e4SLinus Torvalds 943aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 944aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 94546d3ceabSEric Dumazet 94646d3ceabSEric Dumazet skb_orphan(skb); 94746d3ceabSEric Dumazet skb->sk = sk; 94898781965SEric Dumazet skb->destructor = skb_is_tcp_pure_ack(skb) ? sock_wfree : tcp_wfree; 949b73c3d0eSTom Herbert skb_set_hash_from_sk(skb, sk); 95046d3ceabSEric Dumazet atomic_add(skb->truesize, &sk->sk_wmem_alloc); 9511da177e4SLinus Torvalds 9521da177e4SLinus Torvalds /* Build TCP header and checksum it. */ 953aa8223c7SArnaldo Carvalho de Melo th = tcp_hdr(skb); 954c720c7e8SEric Dumazet th->source = inet->inet_sport; 955c720c7e8SEric Dumazet th->dest = inet->inet_dport; 9561da177e4SLinus Torvalds th->seq = htonl(tcb->seq); 9571da177e4SLinus Torvalds th->ack_seq = htonl(tp->rcv_nxt); 958df7a3b07SAl Viro *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | 9594de075e0SEric Dumazet tcb->tcp_flags); 960dfb4b9dcSDavid S. Miller 9614de075e0SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) { 9621da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments 9631da177e4SLinus Torvalds * is never scaled. 9641da177e4SLinus Torvalds */ 965600ff0c2SIlpo Järvinen th->window = htons(min(tp->rcv_wnd, 65535U)); 9661da177e4SLinus Torvalds } else { 9671da177e4SLinus Torvalds th->window = htons(tcp_select_window(sk)); 9681da177e4SLinus Torvalds } 9691da177e4SLinus Torvalds th->check = 0; 9701da177e4SLinus Torvalds th->urg_ptr = 0; 9711da177e4SLinus Torvalds 97233f5f57eSIlpo Järvinen /* The urg_mode check is necessary during a below snd_una win probe */ 9737691367dSHerbert Xu if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { 9747691367dSHerbert Xu if (before(tp->snd_up, tcb->seq + 0x10000)) { 9751da177e4SLinus Torvalds th->urg_ptr = htons(tp->snd_up - tcb->seq); 9761da177e4SLinus Torvalds th->urg = 1; 9777691367dSHerbert Xu } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { 9780eae88f3SEric Dumazet th->urg_ptr = htons(0xFFFF); 9797691367dSHerbert Xu th->urg = 1; 9807691367dSHerbert Xu } 9811da177e4SLinus Torvalds } 9821da177e4SLinus Torvalds 983bd0388aeSWilliam Allen Simpson tcp_options_write((__be32 *)(th + 1), tp, &opts); 9844de075e0SEric Dumazet if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0)) 985735d3831SFlorian Westphal tcp_ecn_send(sk, skb, tcp_header_size); 986dfb4b9dcSDavid S. Miller 987cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 988cfb6eeb4SYOSHIFUJI Hideaki /* Calculate the MD5 hash, as we have all we need now */ 989cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 990a465419bSEric Dumazet sk_nocaps_add(sk, NETIF_F_GSO_MASK); 991bd0388aeSWilliam Allen Simpson tp->af_specific->calc_md5_hash(opts.hash_location, 99249a72dfbSAdam Langley md5, sk, NULL, skb); 993cfb6eeb4SYOSHIFUJI Hideaki } 994cfb6eeb4SYOSHIFUJI Hideaki #endif 995cfb6eeb4SYOSHIFUJI Hideaki 996bb296246SHerbert Xu icsk->icsk_af_ops->send_check(sk, skb); 9971da177e4SLinus Torvalds 9984de075e0SEric Dumazet if (likely(tcb->tcp_flags & TCPHDR_ACK)) 999fc6415bcSDavid S. Miller tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); 10001da177e4SLinus Torvalds 10011da177e4SLinus Torvalds if (skb->len != tcp_header_size) 1002cf533ea5SEric Dumazet tcp_event_data_sent(tp, sk); 10031da177e4SLinus Torvalds 1004bd37a088SWei Yongjun if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) 1005aa2ea058STom Herbert TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, 1006aa2ea058STom Herbert tcp_skb_pcount(skb)); 10071da177e4SLinus Torvalds 1008cd7d8498SEric Dumazet /* OK, its time to fill skb_shinfo(skb)->gso_segs */ 1009cd7d8498SEric Dumazet skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb); 1010cd7d8498SEric Dumazet 10117faee5c0SEric Dumazet /* Our usage of tstamp should remain private */ 10127faee5c0SEric Dumazet skb->tstamp.tv64 = 0; 1013971f10ecSEric Dumazet 1014971f10ecSEric Dumazet /* Cleanup our debris for IP stacks */ 1015971f10ecSEric Dumazet memset(skb->cb, 0, max(sizeof(struct inet_skb_parm), 1016971f10ecSEric Dumazet sizeof(struct inet6_skb_parm))); 1017971f10ecSEric Dumazet 1018b0270e91SEric Dumazet err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl); 10197faee5c0SEric Dumazet 102083de47cdSHua Zhong if (likely(err <= 0)) 10211da177e4SLinus Torvalds return err; 10221da177e4SLinus Torvalds 10235ee2c941SChristoph Paasch tcp_enter_cwr(sk); 10241da177e4SLinus Torvalds 1025b9df3cb8SGerrit Renker return net_xmit_eval(err); 10261da177e4SLinus Torvalds } 10271da177e4SLinus Torvalds 102867edfef7SAndi Kleen /* This routine just queues the buffer for sending. 10291da177e4SLinus Torvalds * 10301da177e4SLinus Torvalds * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, 10311da177e4SLinus Torvalds * otherwise socket can stall. 10321da177e4SLinus Torvalds */ 10331da177e4SLinus Torvalds static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) 10341da177e4SLinus Torvalds { 10351da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 10361da177e4SLinus Torvalds 10371da177e4SLinus Torvalds /* Advance write_seq and place onto the write_queue. */ 10381da177e4SLinus Torvalds tp->write_seq = TCP_SKB_CB(skb)->end_seq; 1039f4a775d1SEric Dumazet __skb_header_release(skb); 1040fe067e8aSDavid S. Miller tcp_add_write_queue_tail(sk, skb); 10413ab224beSHideo Aoki sk->sk_wmem_queued += skb->truesize; 10423ab224beSHideo Aoki sk_mem_charge(sk, skb->truesize); 10431da177e4SLinus Torvalds } 10441da177e4SLinus Torvalds 104567edfef7SAndi Kleen /* Initialize TSO segments for a packet. */ 1046cf533ea5SEric Dumazet static void tcp_set_skb_tso_segs(const struct sock *sk, struct sk_buff *skb, 1047056834d9SIlpo Järvinen unsigned int mss_now) 1048f6302d1dSDavid S. Miller { 10497b7fc97aSEric Dumazet struct skb_shared_info *shinfo = skb_shinfo(skb); 10507b7fc97aSEric Dumazet 1051c52e2421SEric Dumazet /* Make sure we own this skb before messing gso_size/gso_segs */ 1052c52e2421SEric Dumazet WARN_ON_ONCE(skb_cloned(skb)); 1053c52e2421SEric Dumazet 10548f26fb1cSEric Dumazet if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) { 1055f6302d1dSDavid S. Miller /* Avoid the costly divide in the normal 1056f6302d1dSDavid S. Miller * non-TSO case. 1057f6302d1dSDavid S. Miller */ 1058cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, 1); 10597b7fc97aSEric Dumazet shinfo->gso_size = 0; 10607b7fc97aSEric Dumazet shinfo->gso_type = 0; 1061f6302d1dSDavid S. Miller } else { 1062cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now)); 10637b7fc97aSEric Dumazet shinfo->gso_size = mss_now; 10647b7fc97aSEric Dumazet shinfo->gso_type = sk->sk_gso_type; 10651da177e4SLinus Torvalds } 10661da177e4SLinus Torvalds } 10671da177e4SLinus Torvalds 106891fed7a1SIlpo Järvinen /* When a modification to fackets out becomes necessary, we need to check 106968f8353bSIlpo Järvinen * skb is counted to fackets_out or not. 107091fed7a1SIlpo Järvinen */ 1071cf533ea5SEric Dumazet static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb, 107291fed7a1SIlpo Järvinen int decr) 107391fed7a1SIlpo Järvinen { 1074a47e5a98SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1075a47e5a98SIlpo Järvinen 1076dc86967bSIlpo Järvinen if (!tp->sacked_out || tcp_is_reno(tp)) 107791fed7a1SIlpo Järvinen return; 107891fed7a1SIlpo Järvinen 10796859d494SIlpo Järvinen if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq)) 108091fed7a1SIlpo Järvinen tp->fackets_out -= decr; 108191fed7a1SIlpo Järvinen } 108291fed7a1SIlpo Järvinen 1083797108d1SIlpo Järvinen /* Pcount in the middle of the write queue got changed, we need to do various 1084797108d1SIlpo Järvinen * tweaks to fix counters 1085797108d1SIlpo Järvinen */ 1086cf533ea5SEric Dumazet static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) 1087797108d1SIlpo Järvinen { 1088797108d1SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1089797108d1SIlpo Järvinen 1090797108d1SIlpo Järvinen tp->packets_out -= decr; 1091797108d1SIlpo Järvinen 1092797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 1093797108d1SIlpo Järvinen tp->sacked_out -= decr; 1094797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 1095797108d1SIlpo Järvinen tp->retrans_out -= decr; 1096797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 1097797108d1SIlpo Järvinen tp->lost_out -= decr; 1098797108d1SIlpo Järvinen 1099797108d1SIlpo Järvinen /* Reno case is special. Sigh... */ 1100797108d1SIlpo Järvinen if (tcp_is_reno(tp) && decr > 0) 1101797108d1SIlpo Järvinen tp->sacked_out -= min_t(u32, tp->sacked_out, decr); 1102797108d1SIlpo Järvinen 1103797108d1SIlpo Järvinen tcp_adjust_fackets_out(sk, skb, decr); 1104797108d1SIlpo Järvinen 1105797108d1SIlpo Järvinen if (tp->lost_skb_hint && 1106797108d1SIlpo Järvinen before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && 110752cf3cc8SIlpo Järvinen (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))) 1108797108d1SIlpo Järvinen tp->lost_cnt_hint -= decr; 1109797108d1SIlpo Järvinen 1110797108d1SIlpo Järvinen tcp_verify_left_out(tp); 1111797108d1SIlpo Järvinen } 1112797108d1SIlpo Järvinen 1113490cc7d0SWillem de Bruijn static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2) 1114490cc7d0SWillem de Bruijn { 1115490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo = skb_shinfo(skb); 1116490cc7d0SWillem de Bruijn 1117490cc7d0SWillem de Bruijn if (unlikely(shinfo->tx_flags & SKBTX_ANY_TSTAMP) && 1118490cc7d0SWillem de Bruijn !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) { 1119490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo2 = skb_shinfo(skb2); 1120490cc7d0SWillem de Bruijn u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP; 1121490cc7d0SWillem de Bruijn 1122490cc7d0SWillem de Bruijn shinfo->tx_flags &= ~tsflags; 1123490cc7d0SWillem de Bruijn shinfo2->tx_flags |= tsflags; 1124490cc7d0SWillem de Bruijn swap(shinfo->tskey, shinfo2->tskey); 1125490cc7d0SWillem de Bruijn } 1126490cc7d0SWillem de Bruijn } 1127490cc7d0SWillem de Bruijn 11281da177e4SLinus Torvalds /* Function to create two new TCP segments. Shrinks the given segment 11291da177e4SLinus Torvalds * to the specified size and appends a new segment with the rest of the 11301da177e4SLinus Torvalds * packet to the list. This won't be called frequently, I hope. 11311da177e4SLinus Torvalds * Remember, these are still headerless SKBs at this point. 11321da177e4SLinus Torvalds */ 1133056834d9SIlpo Järvinen int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, 11346cc55e09SOctavian Purdila unsigned int mss_now, gfp_t gfp) 11351da177e4SLinus Torvalds { 11361da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 11371da177e4SLinus Torvalds struct sk_buff *buff; 11386475be16SDavid S. Miller int nsize, old_factor; 1139b60b49eaSHerbert Xu int nlen; 11409ce01461SIlpo Järvinen u8 flags; 11411da177e4SLinus Torvalds 11422fceec13SIlpo Järvinen if (WARN_ON(len > skb->len)) 11432fceec13SIlpo Järvinen return -EINVAL; 11446a438bbeSStephen Hemminger 11451da177e4SLinus Torvalds nsize = skb_headlen(skb) - len; 11461da177e4SLinus Torvalds if (nsize < 0) 11471da177e4SLinus Torvalds nsize = 0; 11481da177e4SLinus Torvalds 11496cc55e09SOctavian Purdila if (skb_unclone(skb, gfp)) 11501da177e4SLinus Torvalds return -ENOMEM; 11511da177e4SLinus Torvalds 11521da177e4SLinus Torvalds /* Get a new skb... force flag on. */ 11536cc55e09SOctavian Purdila buff = sk_stream_alloc_skb(sk, nsize, gfp); 11541da177e4SLinus Torvalds if (buff == NULL) 11551da177e4SLinus Torvalds return -ENOMEM; /* We'll just try again later. */ 1156ef5cb973SHerbert Xu 11573ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 11583ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1159b60b49eaSHerbert Xu nlen = skb->len - len - nsize; 1160b60b49eaSHerbert Xu buff->truesize += nlen; 1161b60b49eaSHerbert Xu skb->truesize -= nlen; 11621da177e4SLinus Torvalds 11631da177e4SLinus Torvalds /* Correct the sequence numbers. */ 11641da177e4SLinus Torvalds TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 11651da177e4SLinus Torvalds TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 11661da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 11671da177e4SLinus Torvalds 11681da177e4SLinus Torvalds /* PSH and FIN should only be set in the second packet. */ 11694de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 11704de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 11714de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1172e14c3cafSHerbert Xu TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; 11731da177e4SLinus Torvalds 117484fa7933SPatrick McHardy if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { 11751da177e4SLinus Torvalds /* Copy and checksum data tail into the new buffer. */ 1176056834d9SIlpo Järvinen buff->csum = csum_partial_copy_nocheck(skb->data + len, 1177056834d9SIlpo Järvinen skb_put(buff, nsize), 11781da177e4SLinus Torvalds nsize, 0); 11791da177e4SLinus Torvalds 11801da177e4SLinus Torvalds skb_trim(skb, len); 11811da177e4SLinus Torvalds 11821da177e4SLinus Torvalds skb->csum = csum_block_sub(skb->csum, buff->csum, len); 11831da177e4SLinus Torvalds } else { 118484fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 11851da177e4SLinus Torvalds skb_split(skb, buff, len); 11861da177e4SLinus Torvalds } 11871da177e4SLinus Torvalds 11881da177e4SLinus Torvalds buff->ip_summed = skb->ip_summed; 11891da177e4SLinus Torvalds 1190a61bbcf2SPatrick McHardy buff->tstamp = skb->tstamp; 1191490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 11921da177e4SLinus Torvalds 11936475be16SDavid S. Miller old_factor = tcp_skb_pcount(skb); 11946475be16SDavid S. Miller 11951da177e4SLinus Torvalds /* Fix up tso_factor for both original and new SKB. */ 1196846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1197846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, buff, mss_now); 11981da177e4SLinus Torvalds 11996475be16SDavid S. Miller /* If this packet has been sent out already, we must 12006475be16SDavid S. Miller * adjust the various packet counters. 12016475be16SDavid S. Miller */ 1202cf0b450cSHerbert Xu if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 12036475be16SDavid S. Miller int diff = old_factor - tcp_skb_pcount(skb) - 12046475be16SDavid S. Miller tcp_skb_pcount(buff); 12051da177e4SLinus Torvalds 1206797108d1SIlpo Järvinen if (diff) 1207797108d1SIlpo Järvinen tcp_adjust_pcount(sk, skb, diff); 12081da177e4SLinus Torvalds } 12091da177e4SLinus Torvalds 12101da177e4SLinus Torvalds /* Link BUFF into the send queue. */ 1211f4a775d1SEric Dumazet __skb_header_release(buff); 1212fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 12131da177e4SLinus Torvalds 12141da177e4SLinus Torvalds return 0; 12151da177e4SLinus Torvalds } 12161da177e4SLinus Torvalds 12171da177e4SLinus Torvalds /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c 12181da177e4SLinus Torvalds * eventually). The difference is that pulled data not copied, but 12191da177e4SLinus Torvalds * immediately discarded. 12201da177e4SLinus Torvalds */ 1221f2911969SHerbert Xu ~{PmVHI~} static void __pskb_trim_head(struct sk_buff *skb, int len) 12221da177e4SLinus Torvalds { 12237b7fc97aSEric Dumazet struct skb_shared_info *shinfo; 12241da177e4SLinus Torvalds int i, k, eat; 12251da177e4SLinus Torvalds 12264fa48bf3SEric Dumazet eat = min_t(int, len, skb_headlen(skb)); 12274fa48bf3SEric Dumazet if (eat) { 12284fa48bf3SEric Dumazet __skb_pull(skb, eat); 12294fa48bf3SEric Dumazet len -= eat; 12304fa48bf3SEric Dumazet if (!len) 12314fa48bf3SEric Dumazet return; 12324fa48bf3SEric Dumazet } 12331da177e4SLinus Torvalds eat = len; 12341da177e4SLinus Torvalds k = 0; 12357b7fc97aSEric Dumazet shinfo = skb_shinfo(skb); 12367b7fc97aSEric Dumazet for (i = 0; i < shinfo->nr_frags; i++) { 12377b7fc97aSEric Dumazet int size = skb_frag_size(&shinfo->frags[i]); 12389e903e08SEric Dumazet 12399e903e08SEric Dumazet if (size <= eat) { 1240aff65da0SIan Campbell skb_frag_unref(skb, i); 12419e903e08SEric Dumazet eat -= size; 12421da177e4SLinus Torvalds } else { 12437b7fc97aSEric Dumazet shinfo->frags[k] = shinfo->frags[i]; 12441da177e4SLinus Torvalds if (eat) { 12457b7fc97aSEric Dumazet shinfo->frags[k].page_offset += eat; 12467b7fc97aSEric Dumazet skb_frag_size_sub(&shinfo->frags[k], eat); 12471da177e4SLinus Torvalds eat = 0; 12481da177e4SLinus Torvalds } 12491da177e4SLinus Torvalds k++; 12501da177e4SLinus Torvalds } 12511da177e4SLinus Torvalds } 12527b7fc97aSEric Dumazet shinfo->nr_frags = k; 12531da177e4SLinus Torvalds 125427a884dcSArnaldo Carvalho de Melo skb_reset_tail_pointer(skb); 12551da177e4SLinus Torvalds skb->data_len -= len; 12561da177e4SLinus Torvalds skb->len = skb->data_len; 12571da177e4SLinus Torvalds } 12581da177e4SLinus Torvalds 125967edfef7SAndi Kleen /* Remove acked data from a packet in the transmit queue. */ 12601da177e4SLinus Torvalds int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 12611da177e4SLinus Torvalds { 126214bbd6a5SPravin B Shelar if (skb_unclone(skb, GFP_ATOMIC)) 12631da177e4SLinus Torvalds return -ENOMEM; 12641da177e4SLinus Torvalds 12654fa48bf3SEric Dumazet __pskb_trim_head(skb, len); 12661da177e4SLinus Torvalds 12671da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq += len; 126884fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 12691da177e4SLinus Torvalds 12701da177e4SLinus Torvalds skb->truesize -= len; 12711da177e4SLinus Torvalds sk->sk_wmem_queued -= len; 12723ab224beSHideo Aoki sk_mem_uncharge(sk, len); 12731da177e4SLinus Torvalds sock_set_flag(sk, SOCK_QUEUE_SHRUNK); 12741da177e4SLinus Torvalds 12755b35e1e6SNeal Cardwell /* Any change of skb->len requires recalculation of tso factor. */ 12761da177e4SLinus Torvalds if (tcp_skb_pcount(skb) > 1) 12775b35e1e6SNeal Cardwell tcp_set_skb_tso_segs(sk, skb, tcp_skb_mss(skb)); 12781da177e4SLinus Torvalds 12791da177e4SLinus Torvalds return 0; 12801da177e4SLinus Torvalds } 12811da177e4SLinus Torvalds 12821b63edd6SYuchung Cheng /* Calculate MSS not accounting any TCP options. */ 12831b63edd6SYuchung Cheng static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu) 12845d424d5aSJohn Heffner { 1285cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1286cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 12875d424d5aSJohn Heffner int mss_now; 12885d424d5aSJohn Heffner 12895d424d5aSJohn Heffner /* Calculate base mss without TCP options: 12905d424d5aSJohn Heffner It is MMS_S - sizeof(tcphdr) of rfc1122 12915d424d5aSJohn Heffner */ 12925d424d5aSJohn Heffner mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 12935d424d5aSJohn Heffner 129467469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 129567469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 129667469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 129767469601SEric Dumazet 129867469601SEric Dumazet if (dst && dst_allfrag(dst)) 129967469601SEric Dumazet mss_now -= icsk->icsk_af_ops->net_frag_header_len; 130067469601SEric Dumazet } 130167469601SEric Dumazet 13025d424d5aSJohn Heffner /* Clamp it (mss_clamp does not include tcp options) */ 13035d424d5aSJohn Heffner if (mss_now > tp->rx_opt.mss_clamp) 13045d424d5aSJohn Heffner mss_now = tp->rx_opt.mss_clamp; 13055d424d5aSJohn Heffner 13065d424d5aSJohn Heffner /* Now subtract optional transport overhead */ 13075d424d5aSJohn Heffner mss_now -= icsk->icsk_ext_hdr_len; 13085d424d5aSJohn Heffner 13095d424d5aSJohn Heffner /* Then reserve room for full set of TCP options and 8 bytes of data */ 13105d424d5aSJohn Heffner if (mss_now < 48) 13115d424d5aSJohn Heffner mss_now = 48; 13125d424d5aSJohn Heffner return mss_now; 13135d424d5aSJohn Heffner } 13145d424d5aSJohn Heffner 13151b63edd6SYuchung Cheng /* Calculate MSS. Not accounting for SACKs here. */ 13161b63edd6SYuchung Cheng int tcp_mtu_to_mss(struct sock *sk, int pmtu) 13171b63edd6SYuchung Cheng { 13181b63edd6SYuchung Cheng /* Subtract TCP options size, not including SACKs */ 13191b63edd6SYuchung Cheng return __tcp_mtu_to_mss(sk, pmtu) - 13201b63edd6SYuchung Cheng (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr)); 13211b63edd6SYuchung Cheng } 13221b63edd6SYuchung Cheng 13235d424d5aSJohn Heffner /* Inverse of above */ 132467469601SEric Dumazet int tcp_mss_to_mtu(struct sock *sk, int mss) 13255d424d5aSJohn Heffner { 1326cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1327cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 13285d424d5aSJohn Heffner int mtu; 13295d424d5aSJohn Heffner 13305d424d5aSJohn Heffner mtu = mss + 13315d424d5aSJohn Heffner tp->tcp_header_len + 13325d424d5aSJohn Heffner icsk->icsk_ext_hdr_len + 13335d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 13345d424d5aSJohn Heffner 133567469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 133667469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 133767469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 133867469601SEric Dumazet 133967469601SEric Dumazet if (dst && dst_allfrag(dst)) 134067469601SEric Dumazet mtu += icsk->icsk_af_ops->net_frag_header_len; 134167469601SEric Dumazet } 13425d424d5aSJohn Heffner return mtu; 13435d424d5aSJohn Heffner } 13445d424d5aSJohn Heffner 134567edfef7SAndi Kleen /* MTU probing init per socket */ 13465d424d5aSJohn Heffner void tcp_mtup_init(struct sock *sk) 13475d424d5aSJohn Heffner { 13485d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 13495d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 1350b0f9ca53SFan Du struct net *net = sock_net(sk); 13515d424d5aSJohn Heffner 1352b0f9ca53SFan Du icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1; 13535d424d5aSJohn Heffner icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 13545d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 1355b0f9ca53SFan Du icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss); 13565d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 13575d424d5aSJohn Heffner } 13584bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_mtup_init); 13595d424d5aSJohn Heffner 13601da177e4SLinus Torvalds /* This function synchronize snd mss to current pmtu/exthdr set. 13611da177e4SLinus Torvalds 13621da177e4SLinus Torvalds tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 13631da177e4SLinus Torvalds for TCP options, but includes only bare TCP header. 13641da177e4SLinus Torvalds 13651da177e4SLinus Torvalds tp->rx_opt.mss_clamp is mss negotiated at connection setup. 1366caa20d9aSStephen Hemminger It is minimum of user_mss and mss received with SYN. 13671da177e4SLinus Torvalds It also does not include TCP options. 13681da177e4SLinus Torvalds 1369d83d8461SArnaldo Carvalho de Melo inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. 13701da177e4SLinus Torvalds 13711da177e4SLinus Torvalds tp->mss_cache is current effective sending mss, including 13721da177e4SLinus Torvalds all tcp options except for SACKs. It is evaluated, 13731da177e4SLinus Torvalds taking into account current pmtu, but never exceeds 13741da177e4SLinus Torvalds tp->rx_opt.mss_clamp. 13751da177e4SLinus Torvalds 13761da177e4SLinus Torvalds NOTE1. rfc1122 clearly states that advertised MSS 13771da177e4SLinus Torvalds DOES NOT include either tcp or ip options. 13781da177e4SLinus Torvalds 1379d83d8461SArnaldo Carvalho de Melo NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache 1380d83d8461SArnaldo Carvalho de Melo are READ ONLY outside this function. --ANK (980731) 13811da177e4SLinus Torvalds */ 13821da177e4SLinus Torvalds unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 13831da177e4SLinus Torvalds { 13841da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1385d83d8461SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 13865d424d5aSJohn Heffner int mss_now; 13871da177e4SLinus Torvalds 13885d424d5aSJohn Heffner if (icsk->icsk_mtup.search_high > pmtu) 13895d424d5aSJohn Heffner icsk->icsk_mtup.search_high = pmtu; 13901da177e4SLinus Torvalds 13915d424d5aSJohn Heffner mss_now = tcp_mtu_to_mss(sk, pmtu); 1392409d22b4SIlpo Järvinen mss_now = tcp_bound_to_half_wnd(tp, mss_now); 13931da177e4SLinus Torvalds 13941da177e4SLinus Torvalds /* And store cached results */ 1395d83d8461SArnaldo Carvalho de Melo icsk->icsk_pmtu_cookie = pmtu; 13965d424d5aSJohn Heffner if (icsk->icsk_mtup.enabled) 13975d424d5aSJohn Heffner mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); 1398c1b4a7e6SDavid S. Miller tp->mss_cache = mss_now; 13991da177e4SLinus Torvalds 14001da177e4SLinus Torvalds return mss_now; 14011da177e4SLinus Torvalds } 14024bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_sync_mss); 14031da177e4SLinus Torvalds 14041da177e4SLinus Torvalds /* Compute the current effective MSS, taking SACKs and IP options, 14051da177e4SLinus Torvalds * and even PMTU discovery events into account. 14061da177e4SLinus Torvalds */ 14070c54b85fSIlpo Järvinen unsigned int tcp_current_mss(struct sock *sk) 14081da177e4SLinus Torvalds { 1409cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1410cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1411c1b4a7e6SDavid S. Miller u32 mss_now; 141295c96174SEric Dumazet unsigned int header_len; 141333ad798cSAdam Langley struct tcp_out_options opts; 141433ad798cSAdam Langley struct tcp_md5sig_key *md5; 14151da177e4SLinus Torvalds 1416c1b4a7e6SDavid S. Miller mss_now = tp->mss_cache; 1417c1b4a7e6SDavid S. Miller 14181da177e4SLinus Torvalds if (dst) { 14191da177e4SLinus Torvalds u32 mtu = dst_mtu(dst); 1420d83d8461SArnaldo Carvalho de Melo if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 14211da177e4SLinus Torvalds mss_now = tcp_sync_mss(sk, mtu); 14221da177e4SLinus Torvalds } 14231da177e4SLinus Torvalds 142433ad798cSAdam Langley header_len = tcp_established_options(sk, NULL, &opts, &md5) + 142533ad798cSAdam Langley sizeof(struct tcphdr); 142633ad798cSAdam Langley /* The mss_cache is sized based on tp->tcp_header_len, which assumes 142733ad798cSAdam Langley * some common options. If this is an odd packet (because we have SACK 142833ad798cSAdam Langley * blocks etc) then our calculated header_len will be different, and 142933ad798cSAdam Langley * we have to adjust mss_now correspondingly */ 143033ad798cSAdam Langley if (header_len != tp->tcp_header_len) { 143133ad798cSAdam Langley int delta = (int) header_len - tp->tcp_header_len; 143233ad798cSAdam Langley mss_now -= delta; 143333ad798cSAdam Langley } 1434cfb6eeb4SYOSHIFUJI Hideaki 14351da177e4SLinus Torvalds return mss_now; 14361da177e4SLinus Torvalds } 14371da177e4SLinus Torvalds 143886fd14adSWeiping Pan /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. 143986fd14adSWeiping Pan * As additional protections, we do not touch cwnd in retransmission phases, 144086fd14adSWeiping Pan * and if application hit its sndbuf limit recently. 144186fd14adSWeiping Pan */ 144286fd14adSWeiping Pan static void tcp_cwnd_application_limited(struct sock *sk) 1443a762a980SDavid S. Miller { 14449e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1445a762a980SDavid S. Miller 144686fd14adSWeiping Pan if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && 144786fd14adSWeiping Pan sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 144886fd14adSWeiping Pan /* Limited by application or receiver window. */ 144986fd14adSWeiping Pan u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); 145086fd14adSWeiping Pan u32 win_used = max(tp->snd_cwnd_used, init_win); 145186fd14adSWeiping Pan if (win_used < tp->snd_cwnd) { 145286fd14adSWeiping Pan tp->snd_ssthresh = tcp_current_ssthresh(sk); 145386fd14adSWeiping Pan tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1; 145486fd14adSWeiping Pan } 145586fd14adSWeiping Pan tp->snd_cwnd_used = 0; 145686fd14adSWeiping Pan } 145786fd14adSWeiping Pan tp->snd_cwnd_stamp = tcp_time_stamp; 145886fd14adSWeiping Pan } 145986fd14adSWeiping Pan 1460ca8a2263SNeal Cardwell static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited) 1461a762a980SDavid S. Miller { 1462a762a980SDavid S. Miller struct tcp_sock *tp = tcp_sk(sk); 1463a762a980SDavid S. Miller 1464ca8a2263SNeal Cardwell /* Track the maximum number of outstanding packets in each 1465ca8a2263SNeal Cardwell * window, and remember whether we were cwnd-limited then. 1466ca8a2263SNeal Cardwell */ 1467ca8a2263SNeal Cardwell if (!before(tp->snd_una, tp->max_packets_seq) || 1468ca8a2263SNeal Cardwell tp->packets_out > tp->max_packets_out) { 1469ca8a2263SNeal Cardwell tp->max_packets_out = tp->packets_out; 1470ca8a2263SNeal Cardwell tp->max_packets_seq = tp->snd_nxt; 1471ca8a2263SNeal Cardwell tp->is_cwnd_limited = is_cwnd_limited; 1472ca8a2263SNeal Cardwell } 1473e114a710SEric Dumazet 147424901551SEric Dumazet if (tcp_is_cwnd_limited(sk)) { 1475a762a980SDavid S. Miller /* Network is feed fully. */ 1476a762a980SDavid S. Miller tp->snd_cwnd_used = 0; 1477a762a980SDavid S. Miller tp->snd_cwnd_stamp = tcp_time_stamp; 1478a762a980SDavid S. Miller } else { 1479a762a980SDavid S. Miller /* Network starves. */ 1480a762a980SDavid S. Miller if (tp->packets_out > tp->snd_cwnd_used) 1481a762a980SDavid S. Miller tp->snd_cwnd_used = tp->packets_out; 1482a762a980SDavid S. Miller 148315d33c07SDavid S. Miller if (sysctl_tcp_slow_start_after_idle && 148415d33c07SDavid S. Miller (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto) 1485a762a980SDavid S. Miller tcp_cwnd_application_limited(sk); 1486a762a980SDavid S. Miller } 1487a762a980SDavid S. Miller } 1488a762a980SDavid S. Miller 1489d4589926SEric Dumazet /* Minshall's variant of the Nagle send check. */ 1490d4589926SEric Dumazet static bool tcp_minshall_check(const struct tcp_sock *tp) 1491d4589926SEric Dumazet { 1492d4589926SEric Dumazet return after(tp->snd_sml, tp->snd_una) && 1493d4589926SEric Dumazet !after(tp->snd_sml, tp->snd_nxt); 1494d4589926SEric Dumazet } 1495d4589926SEric Dumazet 1496d4589926SEric Dumazet /* Update snd_sml if this skb is under mss 1497d4589926SEric Dumazet * Note that a TSO packet might end with a sub-mss segment 1498d4589926SEric Dumazet * The test is really : 1499d4589926SEric Dumazet * if ((skb->len % mss) != 0) 1500d4589926SEric Dumazet * tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1501d4589926SEric Dumazet * But we can avoid doing the divide again given we already have 1502d4589926SEric Dumazet * skb_pcount = skb->len / mss_now 15030e3a4803SIlpo Järvinen */ 1504d4589926SEric Dumazet static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now, 1505d4589926SEric Dumazet const struct sk_buff *skb) 1506d4589926SEric Dumazet { 1507d4589926SEric Dumazet if (skb->len < tcp_skb_pcount(skb) * mss_now) 1508d4589926SEric Dumazet tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1509d4589926SEric Dumazet } 1510d4589926SEric Dumazet 1511d4589926SEric Dumazet /* Return false, if packet can be sent now without violation Nagle's rules: 1512d4589926SEric Dumazet * 1. It is full sized. (provided by caller in %partial bool) 1513d4589926SEric Dumazet * 2. Or it contains FIN. (already checked by caller) 1514d4589926SEric Dumazet * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. 1515d4589926SEric Dumazet * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 1516d4589926SEric Dumazet * With Minshall's modification: all sent small packets are ACKed. 1517d4589926SEric Dumazet */ 1518d4589926SEric Dumazet static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp, 1519cc93fc51SPeter Pan(潘卫平) int nonagle) 1520d4589926SEric Dumazet { 1521d4589926SEric Dumazet return partial && 1522d4589926SEric Dumazet ((nonagle & TCP_NAGLE_CORK) || 1523d4589926SEric Dumazet (!nonagle && tp->packets_out && tcp_minshall_check(tp))); 1524d4589926SEric Dumazet } 1525605ad7f1SEric Dumazet 1526605ad7f1SEric Dumazet /* Return how many segs we'd like on a TSO packet, 1527605ad7f1SEric Dumazet * to send one TSO packet per ms 1528605ad7f1SEric Dumazet */ 1529605ad7f1SEric Dumazet static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now) 1530605ad7f1SEric Dumazet { 1531605ad7f1SEric Dumazet u32 bytes, segs; 1532605ad7f1SEric Dumazet 1533605ad7f1SEric Dumazet bytes = min(sk->sk_pacing_rate >> 10, 1534605ad7f1SEric Dumazet sk->sk_gso_max_size - 1 - MAX_TCP_HEADER); 1535605ad7f1SEric Dumazet 1536605ad7f1SEric Dumazet /* Goal is to send at least one packet per ms, 1537605ad7f1SEric Dumazet * not one big TSO packet every 100 ms. 1538605ad7f1SEric Dumazet * This preserves ACK clocking and is consistent 1539605ad7f1SEric Dumazet * with tcp_tso_should_defer() heuristic. 1540605ad7f1SEric Dumazet */ 1541605ad7f1SEric Dumazet segs = max_t(u32, bytes / mss_now, sysctl_tcp_min_tso_segs); 1542605ad7f1SEric Dumazet 1543605ad7f1SEric Dumazet return min_t(u32, segs, sk->sk_gso_max_segs); 1544605ad7f1SEric Dumazet } 1545605ad7f1SEric Dumazet 1546d4589926SEric Dumazet /* Returns the portion of skb which can be sent right away */ 1547d4589926SEric Dumazet static unsigned int tcp_mss_split_point(const struct sock *sk, 1548d4589926SEric Dumazet const struct sk_buff *skb, 1549d4589926SEric Dumazet unsigned int mss_now, 1550d4589926SEric Dumazet unsigned int max_segs, 1551d4589926SEric Dumazet int nonagle) 1552c1b4a7e6SDavid S. Miller { 1553cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1554d4589926SEric Dumazet u32 partial, needed, window, max_len; 1555c1b4a7e6SDavid S. Miller 155690840defSIlpo Järvinen window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 15571485348dSBen Hutchings max_len = mss_now * max_segs; 15580e3a4803SIlpo Järvinen 15591485348dSBen Hutchings if (likely(max_len <= window && skb != tcp_write_queue_tail(sk))) 15601485348dSBen Hutchings return max_len; 15610e3a4803SIlpo Järvinen 15625ea3a748SIlpo Järvinen needed = min(skb->len, window); 15635ea3a748SIlpo Järvinen 15641485348dSBen Hutchings if (max_len <= needed) 15651485348dSBen Hutchings return max_len; 15660e3a4803SIlpo Järvinen 1567d4589926SEric Dumazet partial = needed % mss_now; 1568d4589926SEric Dumazet /* If last segment is not a full MSS, check if Nagle rules allow us 1569d4589926SEric Dumazet * to include this last segment in this skb. 1570d4589926SEric Dumazet * Otherwise, we'll split the skb at last MSS boundary 1571d4589926SEric Dumazet */ 1572cc93fc51SPeter Pan(潘卫平) if (tcp_nagle_check(partial != 0, tp, nonagle)) 1573d4589926SEric Dumazet return needed - partial; 1574d4589926SEric Dumazet 1575d4589926SEric Dumazet return needed; 1576c1b4a7e6SDavid S. Miller } 1577c1b4a7e6SDavid S. Miller 1578c1b4a7e6SDavid S. Miller /* Can at least one segment of SKB be sent right now, according to the 1579c1b4a7e6SDavid S. Miller * congestion window rules? If so, return how many segments are allowed. 1580c1b4a7e6SDavid S. Miller */ 1581cf533ea5SEric Dumazet static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, 1582cf533ea5SEric Dumazet const struct sk_buff *skb) 1583c1b4a7e6SDavid S. Miller { 1584d649a7a8SEric Dumazet u32 in_flight, cwnd, halfcwnd; 1585c1b4a7e6SDavid S. Miller 1586c1b4a7e6SDavid S. Miller /* Don't be strict about the congestion window for the final FIN. */ 15874de075e0SEric Dumazet if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && 15884de075e0SEric Dumazet tcp_skb_pcount(skb) == 1) 1589c1b4a7e6SDavid S. Miller return 1; 1590c1b4a7e6SDavid S. Miller 1591c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1592c1b4a7e6SDavid S. Miller cwnd = tp->snd_cwnd; 1593d649a7a8SEric Dumazet if (in_flight >= cwnd) 1594c1b4a7e6SDavid S. Miller return 0; 1595d649a7a8SEric Dumazet 1596d649a7a8SEric Dumazet /* For better scheduling, ensure we have at least 1597d649a7a8SEric Dumazet * 2 GSO packets in flight. 1598d649a7a8SEric Dumazet */ 1599d649a7a8SEric Dumazet halfcwnd = max(cwnd >> 1, 1U); 1600d649a7a8SEric Dumazet return min(halfcwnd, cwnd - in_flight); 1601c1b4a7e6SDavid S. Miller } 1602c1b4a7e6SDavid S. Miller 1603b595076aSUwe Kleine-König /* Initialize TSO state of a skb. 160467edfef7SAndi Kleen * This must be invoked the first time we consider transmitting 1605c1b4a7e6SDavid S. Miller * SKB onto the wire. 1606c1b4a7e6SDavid S. Miller */ 1607cf533ea5SEric Dumazet static int tcp_init_tso_segs(const struct sock *sk, struct sk_buff *skb, 1608056834d9SIlpo Järvinen unsigned int mss_now) 1609c1b4a7e6SDavid S. Miller { 1610c1b4a7e6SDavid S. Miller int tso_segs = tcp_skb_pcount(skb); 1611c1b4a7e6SDavid S. Miller 1612f8269a49SIlpo Järvinen if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { 1613846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1614c1b4a7e6SDavid S. Miller tso_segs = tcp_skb_pcount(skb); 1615c1b4a7e6SDavid S. Miller } 1616c1b4a7e6SDavid S. Miller return tso_segs; 1617c1b4a7e6SDavid S. Miller } 1618c1b4a7e6SDavid S. Miller 1619c1b4a7e6SDavid S. Miller 1620a2a385d6SEric Dumazet /* Return true if the Nagle test allows this packet to be 1621c1b4a7e6SDavid S. Miller * sent now. 1622c1b4a7e6SDavid S. Miller */ 1623a2a385d6SEric Dumazet static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, 1624c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1625c1b4a7e6SDavid S. Miller { 1626c1b4a7e6SDavid S. Miller /* Nagle rule does not apply to frames, which sit in the middle of the 1627c1b4a7e6SDavid S. Miller * write_queue (they have no chances to get new data). 1628c1b4a7e6SDavid S. Miller * 1629c1b4a7e6SDavid S. Miller * This is implemented in the callers, where they modify the 'nonagle' 1630c1b4a7e6SDavid S. Miller * argument based upon the location of SKB in the send queue. 1631c1b4a7e6SDavid S. Miller */ 1632c1b4a7e6SDavid S. Miller if (nonagle & TCP_NAGLE_PUSH) 1633a2a385d6SEric Dumazet return true; 1634c1b4a7e6SDavid S. Miller 16359b44190dSYuchung Cheng /* Don't use the nagle rule for urgent data (or for the final FIN). */ 16369b44190dSYuchung Cheng if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) 1637a2a385d6SEric Dumazet return true; 1638c1b4a7e6SDavid S. Miller 1639cc93fc51SPeter Pan(潘卫平) if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle)) 1640a2a385d6SEric Dumazet return true; 1641c1b4a7e6SDavid S. Miller 1642a2a385d6SEric Dumazet return false; 1643c1b4a7e6SDavid S. Miller } 1644c1b4a7e6SDavid S. Miller 1645c1b4a7e6SDavid S. Miller /* Does at least the first segment of SKB fit into the send window? */ 1646a2a385d6SEric Dumazet static bool tcp_snd_wnd_test(const struct tcp_sock *tp, 1647a2a385d6SEric Dumazet const struct sk_buff *skb, 1648056834d9SIlpo Järvinen unsigned int cur_mss) 1649c1b4a7e6SDavid S. Miller { 1650c1b4a7e6SDavid S. Miller u32 end_seq = TCP_SKB_CB(skb)->end_seq; 1651c1b4a7e6SDavid S. Miller 1652c1b4a7e6SDavid S. Miller if (skb->len > cur_mss) 1653c1b4a7e6SDavid S. Miller end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 1654c1b4a7e6SDavid S. Miller 165590840defSIlpo Järvinen return !after(end_seq, tcp_wnd_end(tp)); 1656c1b4a7e6SDavid S. Miller } 1657c1b4a7e6SDavid S. Miller 1658fe067e8aSDavid S. Miller /* This checks if the data bearing packet SKB (usually tcp_send_head(sk)) 1659c1b4a7e6SDavid S. Miller * should be put on the wire right now. If so, it returns the number of 1660c1b4a7e6SDavid S. Miller * packets allowed by the congestion window. 1661c1b4a7e6SDavid S. Miller */ 1662cf533ea5SEric Dumazet static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb, 1663c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1664c1b4a7e6SDavid S. Miller { 1665cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1666c1b4a7e6SDavid S. Miller unsigned int cwnd_quota; 1667c1b4a7e6SDavid S. Miller 1668846998aeSDavid S. Miller tcp_init_tso_segs(sk, skb, cur_mss); 1669c1b4a7e6SDavid S. Miller 1670c1b4a7e6SDavid S. Miller if (!tcp_nagle_test(tp, skb, cur_mss, nonagle)) 1671c1b4a7e6SDavid S. Miller return 0; 1672c1b4a7e6SDavid S. Miller 1673c1b4a7e6SDavid S. Miller cwnd_quota = tcp_cwnd_test(tp, skb); 1674056834d9SIlpo Järvinen if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss)) 1675c1b4a7e6SDavid S. Miller cwnd_quota = 0; 1676c1b4a7e6SDavid S. Miller 1677c1b4a7e6SDavid S. Miller return cwnd_quota; 1678c1b4a7e6SDavid S. Miller } 1679c1b4a7e6SDavid S. Miller 168067edfef7SAndi Kleen /* Test if sending is allowed right now. */ 1681a2a385d6SEric Dumazet bool tcp_may_send_now(struct sock *sk) 1682c1b4a7e6SDavid S. Miller { 1683cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1684fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 1685c1b4a7e6SDavid S. Miller 1686a02cec21SEric Dumazet return skb && 16870c54b85fSIlpo Järvinen tcp_snd_test(sk, skb, tcp_current_mss(sk), 1688c1b4a7e6SDavid S. Miller (tcp_skb_is_last(sk, skb) ? 1689a02cec21SEric Dumazet tp->nonagle : TCP_NAGLE_PUSH)); 1690c1b4a7e6SDavid S. Miller } 1691c1b4a7e6SDavid S. Miller 1692c1b4a7e6SDavid S. Miller /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 1693c1b4a7e6SDavid S. Miller * which is put after SKB on the list. It is very much like 1694c1b4a7e6SDavid S. Miller * tcp_fragment() except that it may make several kinds of assumptions 1695c1b4a7e6SDavid S. Miller * in order to speed up the splitting operation. In particular, we 1696c1b4a7e6SDavid S. Miller * know that all the data is in scatter-gather pages, and that the 1697c1b4a7e6SDavid S. Miller * packet has never been sent out before (and thus is not cloned). 1698c1b4a7e6SDavid S. Miller */ 1699056834d9SIlpo Järvinen static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, 1700c4ead4c5SEric Dumazet unsigned int mss_now, gfp_t gfp) 1701c1b4a7e6SDavid S. Miller { 1702c1b4a7e6SDavid S. Miller struct sk_buff *buff; 1703c1b4a7e6SDavid S. Miller int nlen = skb->len - len; 17049ce01461SIlpo Järvinen u8 flags; 1705c1b4a7e6SDavid S. Miller 1706c1b4a7e6SDavid S. Miller /* All of a TSO frame must be composed of paged data. */ 1707c8ac3774SHerbert Xu if (skb->len != skb->data_len) 17086cc55e09SOctavian Purdila return tcp_fragment(sk, skb, len, mss_now, gfp); 1709c1b4a7e6SDavid S. Miller 1710c4ead4c5SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, gfp); 1711c1b4a7e6SDavid S. Miller if (unlikely(buff == NULL)) 1712c1b4a7e6SDavid S. Miller return -ENOMEM; 1713c1b4a7e6SDavid S. Miller 17143ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 17153ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1716b60b49eaSHerbert Xu buff->truesize += nlen; 1717c1b4a7e6SDavid S. Miller skb->truesize -= nlen; 1718c1b4a7e6SDavid S. Miller 1719c1b4a7e6SDavid S. Miller /* Correct the sequence numbers. */ 1720c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 1721c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 1722c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 1723c1b4a7e6SDavid S. Miller 1724c1b4a7e6SDavid S. Miller /* PSH and FIN should only be set in the second packet. */ 17254de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 17264de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 17274de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1728c1b4a7e6SDavid S. Miller 1729c1b4a7e6SDavid S. Miller /* This packet was never sent out yet, so no SACK bits. */ 1730c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->sacked = 0; 1731c1b4a7e6SDavid S. Miller 173284fa7933SPatrick McHardy buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; 1733c1b4a7e6SDavid S. Miller skb_split(skb, buff, len); 1734490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 1735c1b4a7e6SDavid S. Miller 1736c1b4a7e6SDavid S. Miller /* Fix up tso_factor for both original and new SKB. */ 1737846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1738846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, buff, mss_now); 1739c1b4a7e6SDavid S. Miller 1740c1b4a7e6SDavid S. Miller /* Link BUFF into the send queue. */ 1741f4a775d1SEric Dumazet __skb_header_release(buff); 1742fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 1743c1b4a7e6SDavid S. Miller 1744c1b4a7e6SDavid S. Miller return 0; 1745c1b4a7e6SDavid S. Miller } 1746c1b4a7e6SDavid S. Miller 1747c1b4a7e6SDavid S. Miller /* Try to defer sending, if possible, in order to minimize the amount 1748c1b4a7e6SDavid S. Miller * of TSO splitting we do. View it as a kind of TSO Nagle test. 1749c1b4a7e6SDavid S. Miller * 1750c1b4a7e6SDavid S. Miller * This algorithm is from John Heffner. 1751c1b4a7e6SDavid S. Miller */ 1752ca8a2263SNeal Cardwell static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb, 1753605ad7f1SEric Dumazet bool *is_cwnd_limited, u32 max_segs) 1754c1b4a7e6SDavid S. Miller { 17559e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 17566687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 1757c1b4a7e6SDavid S. Miller u32 send_win, cong_win, limit, in_flight; 1758ad9f4f50SEric Dumazet int win_divisor; 1759c1b4a7e6SDavid S. Miller 17604de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 1761ae8064acSJohn Heffner goto send_now; 1762c1b4a7e6SDavid S. Miller 17636687e988SArnaldo Carvalho de Melo if (icsk->icsk_ca_state != TCP_CA_Open) 1764ae8064acSJohn Heffner goto send_now; 1765ae8064acSJohn Heffner 1766ae8064acSJohn Heffner /* Defer for less than two clock ticks. */ 1767bd515c3eSIlpo Järvinen if (tp->tso_deferred && 1768a2acde07SIlpo Järvinen (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1) 1769ae8064acSJohn Heffner goto send_now; 1770908a75c1SDavid S. Miller 1771c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1772c1b4a7e6SDavid S. Miller 1773056834d9SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight)); 1774c1b4a7e6SDavid S. Miller 177590840defSIlpo Järvinen send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1776c1b4a7e6SDavid S. Miller 1777c1b4a7e6SDavid S. Miller /* From in_flight test above, we know that cwnd > in_flight. */ 1778c1b4a7e6SDavid S. Miller cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; 1779c1b4a7e6SDavid S. Miller 1780c1b4a7e6SDavid S. Miller limit = min(send_win, cong_win); 1781c1b4a7e6SDavid S. Miller 1782ba244fe9SDavid S. Miller /* If a full-sized TSO skb can be sent, do it. */ 1783605ad7f1SEric Dumazet if (limit >= max_segs * tp->mss_cache) 1784ae8064acSJohn Heffner goto send_now; 1785ba244fe9SDavid S. Miller 178662ad2761SIlpo Järvinen /* Middle in queue won't get any more data, full sendable already? */ 178762ad2761SIlpo Järvinen if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) 178862ad2761SIlpo Järvinen goto send_now; 178962ad2761SIlpo Järvinen 1790ad9f4f50SEric Dumazet win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor); 1791ad9f4f50SEric Dumazet if (win_divisor) { 1792c1b4a7e6SDavid S. Miller u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); 1793c1b4a7e6SDavid S. Miller 1794c1b4a7e6SDavid S. Miller /* If at least some fraction of a window is available, 1795c1b4a7e6SDavid S. Miller * just use it. 1796c1b4a7e6SDavid S. Miller */ 1797ad9f4f50SEric Dumazet chunk /= win_divisor; 1798c1b4a7e6SDavid S. Miller if (limit >= chunk) 1799ae8064acSJohn Heffner goto send_now; 1800c1b4a7e6SDavid S. Miller } else { 1801c1b4a7e6SDavid S. Miller /* Different approach, try not to defer past a single 1802c1b4a7e6SDavid S. Miller * ACK. Receiver should ACK every other full sized 1803c1b4a7e6SDavid S. Miller * frame, so if we have space for more than 3 frames 1804c1b4a7e6SDavid S. Miller * then send now. 1805c1b4a7e6SDavid S. Miller */ 18066b5a5c0dSNeal Cardwell if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) 1807ae8064acSJohn Heffner goto send_now; 1808c1b4a7e6SDavid S. Miller } 1809c1b4a7e6SDavid S. Miller 1810f4541d60SEric Dumazet /* Ok, it looks like it is advisable to defer. 1811f4541d60SEric Dumazet * Do not rearm the timer if already set to not break TCP ACK clocking. 1812f4541d60SEric Dumazet */ 1813f4541d60SEric Dumazet if (!tp->tso_deferred) 1814ae8064acSJohn Heffner tp->tso_deferred = 1 | (jiffies << 1); 1815ae8064acSJohn Heffner 1816ca8a2263SNeal Cardwell if (cong_win < send_win && cong_win < skb->len) 1817ca8a2263SNeal Cardwell *is_cwnd_limited = true; 1818ca8a2263SNeal Cardwell 1819a2a385d6SEric Dumazet return true; 1820ae8064acSJohn Heffner 1821ae8064acSJohn Heffner send_now: 1822ae8064acSJohn Heffner tp->tso_deferred = 0; 1823a2a385d6SEric Dumazet return false; 1824c1b4a7e6SDavid S. Miller } 1825c1b4a7e6SDavid S. Miller 18265d424d5aSJohn Heffner /* Create a new MTU probe if we are ready. 182767edfef7SAndi Kleen * MTU probe is regularly attempting to increase the path MTU by 182867edfef7SAndi Kleen * deliberately sending larger packets. This discovers routing 182967edfef7SAndi Kleen * changes resulting in larger path MTUs. 183067edfef7SAndi Kleen * 18315d424d5aSJohn Heffner * Returns 0 if we should wait to probe (no cwnd available), 18325d424d5aSJohn Heffner * 1 if a probe was sent, 1833056834d9SIlpo Järvinen * -1 otherwise 1834056834d9SIlpo Järvinen */ 18355d424d5aSJohn Heffner static int tcp_mtu_probe(struct sock *sk) 18365d424d5aSJohn Heffner { 18375d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 18385d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 18395d424d5aSJohn Heffner struct sk_buff *skb, *nskb, *next; 18405d424d5aSJohn Heffner int len; 18415d424d5aSJohn Heffner int probe_size; 184291cc17c0SIlpo Järvinen int size_needed; 18435d424d5aSJohn Heffner int copy; 18445d424d5aSJohn Heffner int mss_now; 18455d424d5aSJohn Heffner 18465d424d5aSJohn Heffner /* Not currently probing/verifying, 18475d424d5aSJohn Heffner * not in recovery, 18485d424d5aSJohn Heffner * have enough cwnd, and 18495d424d5aSJohn Heffner * not SACKing (the variable headers throw things off) */ 18505d424d5aSJohn Heffner if (!icsk->icsk_mtup.enabled || 18515d424d5aSJohn Heffner icsk->icsk_mtup.probe_size || 18525d424d5aSJohn Heffner inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 18535d424d5aSJohn Heffner tp->snd_cwnd < 11 || 1854cabeccbdSIlpo Järvinen tp->rx_opt.num_sacks || tp->rx_opt.dsack) 18555d424d5aSJohn Heffner return -1; 18565d424d5aSJohn Heffner 18575d424d5aSJohn Heffner /* Very simple search strategy: just double the MSS. */ 18580c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 18595d424d5aSJohn Heffner probe_size = 2 * tp->mss_cache; 186091cc17c0SIlpo Järvinen size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; 18615d424d5aSJohn Heffner if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) { 18625d424d5aSJohn Heffner /* TODO: set timer for probe_converge_event */ 18635d424d5aSJohn Heffner return -1; 18645d424d5aSJohn Heffner } 18655d424d5aSJohn Heffner 18665d424d5aSJohn Heffner /* Have enough data in the send queue to probe? */ 18677f9c33e5SIlpo Järvinen if (tp->write_seq - tp->snd_nxt < size_needed) 18685d424d5aSJohn Heffner return -1; 18695d424d5aSJohn Heffner 187091cc17c0SIlpo Järvinen if (tp->snd_wnd < size_needed) 18715d424d5aSJohn Heffner return -1; 187290840defSIlpo Järvinen if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) 18735d424d5aSJohn Heffner return 0; 18745d424d5aSJohn Heffner 1875d67c58e9SIlpo Järvinen /* Do we need to wait to drain cwnd? With none in flight, don't stall */ 1876d67c58e9SIlpo Järvinen if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { 1877d67c58e9SIlpo Järvinen if (!tcp_packets_in_flight(tp)) 18785d424d5aSJohn Heffner return -1; 18795d424d5aSJohn Heffner else 18805d424d5aSJohn Heffner return 0; 18815d424d5aSJohn Heffner } 18825d424d5aSJohn Heffner 18835d424d5aSJohn Heffner /* We're allowed to probe. Build it now. */ 18845d424d5aSJohn Heffner if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL) 18855d424d5aSJohn Heffner return -1; 18863ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 18873ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 18885d424d5aSJohn Heffner 1889fe067e8aSDavid S. Miller skb = tcp_send_head(sk); 18905d424d5aSJohn Heffner 18915d424d5aSJohn Heffner TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 18925d424d5aSJohn Heffner TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 18934de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; 18945d424d5aSJohn Heffner TCP_SKB_CB(nskb)->sacked = 0; 18955d424d5aSJohn Heffner nskb->csum = 0; 189684fa7933SPatrick McHardy nskb->ip_summed = skb->ip_summed; 18975d424d5aSJohn Heffner 189850c4817eSIlpo Järvinen tcp_insert_write_queue_before(nskb, skb, sk); 189950c4817eSIlpo Järvinen 19005d424d5aSJohn Heffner len = 0; 1901234b6860SIlpo Järvinen tcp_for_write_queue_from_safe(skb, next, sk) { 19025d424d5aSJohn Heffner copy = min_t(int, skb->len, probe_size - len); 19035d424d5aSJohn Heffner if (nskb->ip_summed) 19045d424d5aSJohn Heffner skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); 19055d424d5aSJohn Heffner else 19065d424d5aSJohn Heffner nskb->csum = skb_copy_and_csum_bits(skb, 0, 1907056834d9SIlpo Järvinen skb_put(nskb, copy), 1908056834d9SIlpo Järvinen copy, nskb->csum); 19095d424d5aSJohn Heffner 19105d424d5aSJohn Heffner if (skb->len <= copy) { 19115d424d5aSJohn Heffner /* We've eaten all the data from this skb. 19125d424d5aSJohn Heffner * Throw it away. */ 19134de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 1914fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 19153ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 19165d424d5aSJohn Heffner } else { 19174de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & 1918a3433f35SChangli Gao ~(TCPHDR_FIN|TCPHDR_PSH); 19195d424d5aSJohn Heffner if (!skb_shinfo(skb)->nr_frags) { 19205d424d5aSJohn Heffner skb_pull(skb, copy); 192184fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 1922056834d9SIlpo Järvinen skb->csum = csum_partial(skb->data, 1923056834d9SIlpo Järvinen skb->len, 0); 19245d424d5aSJohn Heffner } else { 19255d424d5aSJohn Heffner __pskb_trim_head(skb, copy); 19265d424d5aSJohn Heffner tcp_set_skb_tso_segs(sk, skb, mss_now); 19275d424d5aSJohn Heffner } 19285d424d5aSJohn Heffner TCP_SKB_CB(skb)->seq += copy; 19295d424d5aSJohn Heffner } 19305d424d5aSJohn Heffner 19315d424d5aSJohn Heffner len += copy; 1932234b6860SIlpo Järvinen 1933234b6860SIlpo Järvinen if (len >= probe_size) 1934234b6860SIlpo Järvinen break; 19355d424d5aSJohn Heffner } 19365d424d5aSJohn Heffner tcp_init_tso_segs(sk, nskb, nskb->len); 19375d424d5aSJohn Heffner 19385d424d5aSJohn Heffner /* We're ready to send. If this fails, the probe will 19397faee5c0SEric Dumazet * be resegmented into mss-sized pieces by tcp_write_xmit(). 19407faee5c0SEric Dumazet */ 19415d424d5aSJohn Heffner if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 19425d424d5aSJohn Heffner /* Decrement cwnd here because we are sending 19435d424d5aSJohn Heffner * effectively two packets. */ 19445d424d5aSJohn Heffner tp->snd_cwnd--; 194566f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, nskb); 19465d424d5aSJohn Heffner 19475d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 19480e7b1368SJohn Heffner tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 19490e7b1368SJohn Heffner tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 19505d424d5aSJohn Heffner 19515d424d5aSJohn Heffner return 1; 19525d424d5aSJohn Heffner } 19535d424d5aSJohn Heffner 19545d424d5aSJohn Heffner return -1; 19555d424d5aSJohn Heffner } 19565d424d5aSJohn Heffner 19571da177e4SLinus Torvalds /* This routine writes packets to the network. It advances the 19581da177e4SLinus Torvalds * send_head. This happens as incoming acks open up the remote 19591da177e4SLinus Torvalds * window for us. 19601da177e4SLinus Torvalds * 1961f8269a49SIlpo Järvinen * LARGESEND note: !tcp_urg_mode is overkill, only frames between 1962f8269a49SIlpo Järvinen * snd_up-64k-mss .. snd_up cannot be large. However, taking into 1963f8269a49SIlpo Järvinen * account rare use of URG, this is not a big flaw. 1964f8269a49SIlpo Järvinen * 19656ba8a3b1SNandita Dukkipati * Send at most one packet when push_one > 0. Temporarily ignore 19666ba8a3b1SNandita Dukkipati * cwnd limit to force at most one packet out when push_one == 2. 19676ba8a3b1SNandita Dukkipati 1968a2a385d6SEric Dumazet * Returns true, if no segments are in flight and we have queued segments, 1969a2a385d6SEric Dumazet * but cannot send anything now because of SWS or another problem. 19701da177e4SLinus Torvalds */ 1971a2a385d6SEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 1972d5dd9175SIlpo Järvinen int push_one, gfp_t gfp) 19731da177e4SLinus Torvalds { 19741da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 197592df7b51SDavid S. Miller struct sk_buff *skb; 1976c1b4a7e6SDavid S. Miller unsigned int tso_segs, sent_pkts; 1977c1b4a7e6SDavid S. Miller int cwnd_quota; 19785d424d5aSJohn Heffner int result; 1979ca8a2263SNeal Cardwell bool is_cwnd_limited = false; 1980605ad7f1SEric Dumazet u32 max_segs; 19811da177e4SLinus Torvalds 1982c1b4a7e6SDavid S. Miller sent_pkts = 0; 19835d424d5aSJohn Heffner 1984d5dd9175SIlpo Järvinen if (!push_one) { 19855d424d5aSJohn Heffner /* Do MTU probing. */ 1986d5dd9175SIlpo Järvinen result = tcp_mtu_probe(sk); 1987d5dd9175SIlpo Järvinen if (!result) { 1988a2a385d6SEric Dumazet return false; 19895d424d5aSJohn Heffner } else if (result > 0) { 19905d424d5aSJohn Heffner sent_pkts = 1; 19915d424d5aSJohn Heffner } 1992d5dd9175SIlpo Järvinen } 19935d424d5aSJohn Heffner 1994605ad7f1SEric Dumazet max_segs = tcp_tso_autosize(sk, mss_now); 1995fe067e8aSDavid S. Miller while ((skb = tcp_send_head(sk))) { 1996c8ac3774SHerbert Xu unsigned int limit; 1997c8ac3774SHerbert Xu 1998b68e9f85SHerbert Xu tso_segs = tcp_init_tso_segs(sk, skb, mss_now); 1999c1b4a7e6SDavid S. Miller BUG_ON(!tso_segs); 2000c1b4a7e6SDavid S. Miller 20019d186cacSAndrey Vagin if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) { 20027faee5c0SEric Dumazet /* "skb_mstamp" is used as a start point for the retransmit timer */ 20037faee5c0SEric Dumazet skb_mstamp_get(&skb->skb_mstamp); 2004ec342325SAndrew Vagin goto repair; /* Skip network transmission */ 20059d186cacSAndrey Vagin } 2006ec342325SAndrew Vagin 2007b68e9f85SHerbert Xu cwnd_quota = tcp_cwnd_test(tp, skb); 20086ba8a3b1SNandita Dukkipati if (!cwnd_quota) { 2009ca8a2263SNeal Cardwell is_cwnd_limited = true; 20106ba8a3b1SNandita Dukkipati if (push_one == 2) 20116ba8a3b1SNandita Dukkipati /* Force out a loss probe pkt. */ 20126ba8a3b1SNandita Dukkipati cwnd_quota = 1; 20136ba8a3b1SNandita Dukkipati else 2014b68e9f85SHerbert Xu break; 20156ba8a3b1SNandita Dukkipati } 2016b68e9f85SHerbert Xu 2017b68e9f85SHerbert Xu if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) 2018b68e9f85SHerbert Xu break; 2019b68e9f85SHerbert Xu 2020843925f3SHerbert Xu if (tso_segs == 1 || !max_segs) { 2021aa93466bSDavid S. Miller if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 2022aa93466bSDavid S. Miller (tcp_skb_is_last(sk, skb) ? 2023aa93466bSDavid S. Miller nonagle : TCP_NAGLE_PUSH)))) 2024aa93466bSDavid S. Miller break; 2025c1b4a7e6SDavid S. Miller } else { 2026ca8a2263SNeal Cardwell if (!push_one && 2027605ad7f1SEric Dumazet tcp_tso_should_defer(sk, skb, &is_cwnd_limited, 2028605ad7f1SEric Dumazet max_segs)) 2029aa93466bSDavid S. Miller break; 2030c1b4a7e6SDavid S. Miller } 2031aa93466bSDavid S. Miller 2032605ad7f1SEric Dumazet limit = mss_now; 2033843925f3SHerbert Xu if (tso_segs > 1 && max_segs && !tcp_urg_mode(tp)) 2034605ad7f1SEric Dumazet limit = tcp_mss_split_point(sk, skb, mss_now, 2035605ad7f1SEric Dumazet min_t(unsigned int, 2036605ad7f1SEric Dumazet cwnd_quota, 2037605ad7f1SEric Dumazet max_segs), 2038605ad7f1SEric Dumazet nonagle); 2039605ad7f1SEric Dumazet 2040605ad7f1SEric Dumazet if (skb->len > limit && 2041605ad7f1SEric Dumazet unlikely(tso_fragment(sk, skb, limit, mss_now, gfp))) 2042605ad7f1SEric Dumazet break; 2043605ad7f1SEric Dumazet 2044c9eeec26SEric Dumazet /* TCP Small Queues : 2045c9eeec26SEric Dumazet * Control number of packets in qdisc/devices to two packets / or ~1 ms. 2046c9eeec26SEric Dumazet * This allows for : 2047c9eeec26SEric Dumazet * - better RTT estimation and ACK scheduling 2048c9eeec26SEric Dumazet * - faster recovery 2049c9eeec26SEric Dumazet * - high rates 205098e09386SEric Dumazet * Alas, some drivers / subsystems require a fair amount 205198e09386SEric Dumazet * of queued bytes to ensure line rate. 205298e09386SEric Dumazet * One example is wifi aggregation (802.11 AMPDU) 205346d3ceabSEric Dumazet */ 2054605ad7f1SEric Dumazet limit = max(2 * skb->truesize, sk->sk_pacing_rate >> 10); 2055605ad7f1SEric Dumazet limit = min_t(u32, limit, sysctl_tcp_limit_output_bytes); 2056c9eeec26SEric Dumazet 2057c9eeec26SEric Dumazet if (atomic_read(&sk->sk_wmem_alloc) > limit) { 205846d3ceabSEric Dumazet set_bit(TSQ_THROTTLED, &tp->tsq_flags); 2059bf06200eSJohn Ogness /* It is possible TX completion already happened 2060bf06200eSJohn Ogness * before we set TSQ_THROTTLED, so we must 2061bf06200eSJohn Ogness * test again the condition. 2062bf06200eSJohn Ogness */ 20634e857c58SPeter Zijlstra smp_mb__after_atomic(); 2064bf06200eSJohn Ogness if (atomic_read(&sk->sk_wmem_alloc) > limit) 206546d3ceabSEric Dumazet break; 206646d3ceabSEric Dumazet } 2067c9eeec26SEric Dumazet 2068d5dd9175SIlpo Järvinen if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) 20691da177e4SLinus Torvalds break; 20701da177e4SLinus Torvalds 2071ec342325SAndrew Vagin repair: 20721da177e4SLinus Torvalds /* Advance the send_head. This one is sent out. 20731da177e4SLinus Torvalds * This call will increment packets_out. 20741da177e4SLinus Torvalds */ 207566f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 20761da177e4SLinus Torvalds 20771da177e4SLinus Torvalds tcp_minshall_update(tp, mss_now, skb); 2078a262f0cdSNandita Dukkipati sent_pkts += tcp_skb_pcount(skb); 2079d5dd9175SIlpo Järvinen 2080d5dd9175SIlpo Järvinen if (push_one) 2081d5dd9175SIlpo Järvinen break; 20821da177e4SLinus Torvalds } 20831da177e4SLinus Torvalds 2084aa93466bSDavid S. Miller if (likely(sent_pkts)) { 2085684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 2086684bad11SYuchung Cheng tp->prr_out += sent_pkts; 20876ba8a3b1SNandita Dukkipati 20886ba8a3b1SNandita Dukkipati /* Send one loss probe per tail loss episode. */ 20896ba8a3b1SNandita Dukkipati if (push_one != 2) 20906ba8a3b1SNandita Dukkipati tcp_schedule_loss_probe(sk); 2091ca8a2263SNeal Cardwell tcp_cwnd_validate(sk, is_cwnd_limited); 2092a2a385d6SEric Dumazet return false; 20931da177e4SLinus Torvalds } 20946ba8a3b1SNandita Dukkipati return (push_one == 2) || (!tp->packets_out && tcp_send_head(sk)); 20956ba8a3b1SNandita Dukkipati } 20966ba8a3b1SNandita Dukkipati 20976ba8a3b1SNandita Dukkipati bool tcp_schedule_loss_probe(struct sock *sk) 20986ba8a3b1SNandita Dukkipati { 20996ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 21006ba8a3b1SNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 21016ba8a3b1SNandita Dukkipati u32 timeout, tlp_time_stamp, rto_time_stamp; 2102740b0f18SEric Dumazet u32 rtt = usecs_to_jiffies(tp->srtt_us >> 3); 21036ba8a3b1SNandita Dukkipati 21046ba8a3b1SNandita Dukkipati if (WARN_ON(icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS)) 21056ba8a3b1SNandita Dukkipati return false; 21066ba8a3b1SNandita Dukkipati /* No consecutive loss probes. */ 21076ba8a3b1SNandita Dukkipati if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) { 21086ba8a3b1SNandita Dukkipati tcp_rearm_rto(sk); 21096ba8a3b1SNandita Dukkipati return false; 21106ba8a3b1SNandita Dukkipati } 21116ba8a3b1SNandita Dukkipati /* Don't do any loss probe on a Fast Open connection before 3WHS 21126ba8a3b1SNandita Dukkipati * finishes. 21136ba8a3b1SNandita Dukkipati */ 21146ba8a3b1SNandita Dukkipati if (sk->sk_state == TCP_SYN_RECV) 21156ba8a3b1SNandita Dukkipati return false; 21166ba8a3b1SNandita Dukkipati 21176ba8a3b1SNandita Dukkipati /* TLP is only scheduled when next timer event is RTO. */ 21186ba8a3b1SNandita Dukkipati if (icsk->icsk_pending != ICSK_TIME_RETRANS) 21196ba8a3b1SNandita Dukkipati return false; 21206ba8a3b1SNandita Dukkipati 21216ba8a3b1SNandita Dukkipati /* Schedule a loss probe in 2*RTT for SACK capable connections 21226ba8a3b1SNandita Dukkipati * in Open state, that are either limited by cwnd or application. 21236ba8a3b1SNandita Dukkipati */ 2124740b0f18SEric Dumazet if (sysctl_tcp_early_retrans < 3 || !tp->srtt_us || !tp->packets_out || 21256ba8a3b1SNandita Dukkipati !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open) 21266ba8a3b1SNandita Dukkipati return false; 21276ba8a3b1SNandita Dukkipati 21286ba8a3b1SNandita Dukkipati if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) && 21296ba8a3b1SNandita Dukkipati tcp_send_head(sk)) 21306ba8a3b1SNandita Dukkipati return false; 21316ba8a3b1SNandita Dukkipati 21326ba8a3b1SNandita Dukkipati /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account 21336ba8a3b1SNandita Dukkipati * for delayed ack when there's one outstanding packet. 21346ba8a3b1SNandita Dukkipati */ 21356ba8a3b1SNandita Dukkipati timeout = rtt << 1; 21366ba8a3b1SNandita Dukkipati if (tp->packets_out == 1) 21376ba8a3b1SNandita Dukkipati timeout = max_t(u32, timeout, 21386ba8a3b1SNandita Dukkipati (rtt + (rtt >> 1) + TCP_DELACK_MAX)); 21396ba8a3b1SNandita Dukkipati timeout = max_t(u32, timeout, msecs_to_jiffies(10)); 21406ba8a3b1SNandita Dukkipati 21416ba8a3b1SNandita Dukkipati /* If RTO is shorter, just schedule TLP in its place. */ 21426ba8a3b1SNandita Dukkipati tlp_time_stamp = tcp_time_stamp + timeout; 21436ba8a3b1SNandita Dukkipati rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout; 21446ba8a3b1SNandita Dukkipati if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) { 21456ba8a3b1SNandita Dukkipati s32 delta = rto_time_stamp - tcp_time_stamp; 21466ba8a3b1SNandita Dukkipati if (delta > 0) 21476ba8a3b1SNandita Dukkipati timeout = delta; 21486ba8a3b1SNandita Dukkipati } 21496ba8a3b1SNandita Dukkipati 21506ba8a3b1SNandita Dukkipati inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, 21516ba8a3b1SNandita Dukkipati TCP_RTO_MAX); 21526ba8a3b1SNandita Dukkipati return true; 21536ba8a3b1SNandita Dukkipati } 21546ba8a3b1SNandita Dukkipati 21551f3279aeSEric Dumazet /* Thanks to skb fast clones, we can detect if a prior transmit of 21561f3279aeSEric Dumazet * a packet is still in a qdisc or driver queue. 21571f3279aeSEric Dumazet * In this case, there is very little point doing a retransmit ! 21581f3279aeSEric Dumazet * Note: This is called from BH context only. 21591f3279aeSEric Dumazet */ 21601f3279aeSEric Dumazet static bool skb_still_in_host_queue(const struct sock *sk, 21611f3279aeSEric Dumazet const struct sk_buff *skb) 21621f3279aeSEric Dumazet { 216339bb5e62SEric Dumazet if (unlikely(skb_fclone_busy(sk, skb))) { 21641f3279aeSEric Dumazet NET_INC_STATS_BH(sock_net(sk), 21651f3279aeSEric Dumazet LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES); 21661f3279aeSEric Dumazet return true; 21671f3279aeSEric Dumazet } 21681f3279aeSEric Dumazet return false; 21691f3279aeSEric Dumazet } 21701f3279aeSEric Dumazet 21716ba8a3b1SNandita Dukkipati /* When probe timeout (PTO) fires, send a new segment if one exists, else 21726ba8a3b1SNandita Dukkipati * retransmit the last segment. 21736ba8a3b1SNandita Dukkipati */ 21746ba8a3b1SNandita Dukkipati void tcp_send_loss_probe(struct sock *sk) 21756ba8a3b1SNandita Dukkipati { 21769b717a8dSNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 21776ba8a3b1SNandita Dukkipati struct sk_buff *skb; 21786ba8a3b1SNandita Dukkipati int pcount; 21796ba8a3b1SNandita Dukkipati int mss = tcp_current_mss(sk); 21806ba8a3b1SNandita Dukkipati int err = -1; 21816ba8a3b1SNandita Dukkipati 21826ba8a3b1SNandita Dukkipati if (tcp_send_head(sk) != NULL) { 21836ba8a3b1SNandita Dukkipati err = tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC); 21846ba8a3b1SNandita Dukkipati goto rearm_timer; 21856ba8a3b1SNandita Dukkipati } 21866ba8a3b1SNandita Dukkipati 21879b717a8dSNandita Dukkipati /* At most one outstanding TLP retransmission. */ 21889b717a8dSNandita Dukkipati if (tp->tlp_high_seq) 21899b717a8dSNandita Dukkipati goto rearm_timer; 21909b717a8dSNandita Dukkipati 21916ba8a3b1SNandita Dukkipati /* Retransmit last segment. */ 21926ba8a3b1SNandita Dukkipati skb = tcp_write_queue_tail(sk); 21936ba8a3b1SNandita Dukkipati if (WARN_ON(!skb)) 21946ba8a3b1SNandita Dukkipati goto rearm_timer; 21956ba8a3b1SNandita Dukkipati 21961f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 21971f3279aeSEric Dumazet goto rearm_timer; 21981f3279aeSEric Dumazet 21996ba8a3b1SNandita Dukkipati pcount = tcp_skb_pcount(skb); 22006ba8a3b1SNandita Dukkipati if (WARN_ON(!pcount)) 22016ba8a3b1SNandita Dukkipati goto rearm_timer; 22026ba8a3b1SNandita Dukkipati 22036ba8a3b1SNandita Dukkipati if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) { 22046cc55e09SOctavian Purdila if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss, 22056cc55e09SOctavian Purdila GFP_ATOMIC))) 22066ba8a3b1SNandita Dukkipati goto rearm_timer; 22076ba8a3b1SNandita Dukkipati skb = tcp_write_queue_tail(sk); 22086ba8a3b1SNandita Dukkipati } 22096ba8a3b1SNandita Dukkipati 22106ba8a3b1SNandita Dukkipati if (WARN_ON(!skb || !tcp_skb_pcount(skb))) 22116ba8a3b1SNandita Dukkipati goto rearm_timer; 22126ba8a3b1SNandita Dukkipati 22136ba8a3b1SNandita Dukkipati err = __tcp_retransmit_skb(sk, skb); 22146ba8a3b1SNandita Dukkipati 22159b717a8dSNandita Dukkipati /* Record snd_nxt for loss detection. */ 22169b717a8dSNandita Dukkipati if (likely(!err)) 22179b717a8dSNandita Dukkipati tp->tlp_high_seq = tp->snd_nxt; 22189b717a8dSNandita Dukkipati 22196ba8a3b1SNandita Dukkipati rearm_timer: 22206ba8a3b1SNandita Dukkipati inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 22216ba8a3b1SNandita Dukkipati inet_csk(sk)->icsk_rto, 22226ba8a3b1SNandita Dukkipati TCP_RTO_MAX); 22236ba8a3b1SNandita Dukkipati 22246ba8a3b1SNandita Dukkipati if (likely(!err)) 22256ba8a3b1SNandita Dukkipati NET_INC_STATS_BH(sock_net(sk), 22266ba8a3b1SNandita Dukkipati LINUX_MIB_TCPLOSSPROBES); 22271da177e4SLinus Torvalds } 22281da177e4SLinus Torvalds 2229a762a980SDavid S. Miller /* Push out any pending frames which were held back due to 2230a762a980SDavid S. Miller * TCP_CORK or attempt at coalescing tiny packets. 2231a762a980SDavid S. Miller * The socket must be locked by the caller. 2232a762a980SDavid S. Miller */ 22339e412ba7SIlpo Järvinen void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 22349e412ba7SIlpo Järvinen int nonagle) 2235a762a980SDavid S. Miller { 2236726e07a8SIlpo Järvinen /* If we are closed, the bytes will have to remain here. 2237726e07a8SIlpo Järvinen * In time closedown will finish, we empty the write queue and 2238726e07a8SIlpo Järvinen * all will be happy. 2239726e07a8SIlpo Järvinen */ 2240726e07a8SIlpo Järvinen if (unlikely(sk->sk_state == TCP_CLOSE)) 2241726e07a8SIlpo Järvinen return; 2242726e07a8SIlpo Järvinen 224399a1dec7SMel Gorman if (tcp_write_xmit(sk, cur_mss, nonagle, 0, 224499a1dec7SMel Gorman sk_gfp_atomic(sk, GFP_ATOMIC))) 22459e412ba7SIlpo Järvinen tcp_check_probe_timer(sk); 2246a762a980SDavid S. Miller } 2247a762a980SDavid S. Miller 2248c1b4a7e6SDavid S. Miller /* Send _single_ skb sitting at the send head. This function requires 2249c1b4a7e6SDavid S. Miller * true push pending frames to setup probe timer etc. 2250c1b4a7e6SDavid S. Miller */ 2251c1b4a7e6SDavid S. Miller void tcp_push_one(struct sock *sk, unsigned int mss_now) 2252c1b4a7e6SDavid S. Miller { 2253fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 2254c1b4a7e6SDavid S. Miller 2255c1b4a7e6SDavid S. Miller BUG_ON(!skb || skb->len < mss_now); 2256c1b4a7e6SDavid S. Miller 2257d5dd9175SIlpo Järvinen tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); 2258c1b4a7e6SDavid S. Miller } 2259c1b4a7e6SDavid S. Miller 22601da177e4SLinus Torvalds /* This function returns the amount that we can raise the 22611da177e4SLinus Torvalds * usable window based on the following constraints 22621da177e4SLinus Torvalds * 22631da177e4SLinus Torvalds * 1. The window can never be shrunk once it is offered (RFC 793) 22641da177e4SLinus Torvalds * 2. We limit memory per socket 22651da177e4SLinus Torvalds * 22661da177e4SLinus Torvalds * RFC 1122: 22671da177e4SLinus Torvalds * "the suggested [SWS] avoidance algorithm for the receiver is to keep 22681da177e4SLinus Torvalds * RECV.NEXT + RCV.WIN fixed until: 22691da177e4SLinus Torvalds * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 22701da177e4SLinus Torvalds * 22711da177e4SLinus Torvalds * i.e. don't raise the right edge of the window until you can raise 22721da177e4SLinus Torvalds * it at least MSS bytes. 22731da177e4SLinus Torvalds * 22741da177e4SLinus Torvalds * Unfortunately, the recommended algorithm breaks header prediction, 22751da177e4SLinus Torvalds * since header prediction assumes th->window stays fixed. 22761da177e4SLinus Torvalds * 22771da177e4SLinus Torvalds * Strictly speaking, keeping th->window fixed violates the receiver 22781da177e4SLinus Torvalds * side SWS prevention criteria. The problem is that under this rule 22791da177e4SLinus Torvalds * a stream of single byte packets will cause the right side of the 22801da177e4SLinus Torvalds * window to always advance by a single byte. 22811da177e4SLinus Torvalds * 22821da177e4SLinus Torvalds * Of course, if the sender implements sender side SWS prevention 22831da177e4SLinus Torvalds * then this will not be a problem. 22841da177e4SLinus Torvalds * 22851da177e4SLinus Torvalds * BSD seems to make the following compromise: 22861da177e4SLinus Torvalds * 22871da177e4SLinus Torvalds * If the free space is less than the 1/4 of the maximum 22881da177e4SLinus Torvalds * space available and the free space is less than 1/2 mss, 22891da177e4SLinus Torvalds * then set the window to 0. 22901da177e4SLinus Torvalds * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 22911da177e4SLinus Torvalds * Otherwise, just prevent the window from shrinking 22921da177e4SLinus Torvalds * and from being larger than the largest representable value. 22931da177e4SLinus Torvalds * 22941da177e4SLinus Torvalds * This prevents incremental opening of the window in the regime 22951da177e4SLinus Torvalds * where TCP is limited by the speed of the reader side taking 22961da177e4SLinus Torvalds * data out of the TCP receive queue. It does nothing about 22971da177e4SLinus Torvalds * those cases where the window is constrained on the sender side 22981da177e4SLinus Torvalds * because the pipeline is full. 22991da177e4SLinus Torvalds * 23001da177e4SLinus Torvalds * BSD also seems to "accidentally" limit itself to windows that are a 23011da177e4SLinus Torvalds * multiple of MSS, at least until the free space gets quite small. 23021da177e4SLinus Torvalds * This would appear to be a side effect of the mbuf implementation. 23031da177e4SLinus Torvalds * Combining these two algorithms results in the observed behavior 23041da177e4SLinus Torvalds * of having a fixed window size at almost all times. 23051da177e4SLinus Torvalds * 23061da177e4SLinus Torvalds * Below we obtain similar behavior by forcing the offered window to 23071da177e4SLinus Torvalds * a multiple of the mss when it is feasible to do so. 23081da177e4SLinus Torvalds * 23091da177e4SLinus Torvalds * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 23101da177e4SLinus Torvalds * Regular options like TIMESTAMP are taken into account. 23111da177e4SLinus Torvalds */ 23121da177e4SLinus Torvalds u32 __tcp_select_window(struct sock *sk) 23131da177e4SLinus Torvalds { 2314463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 23151da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2316caa20d9aSStephen Hemminger /* MSS for the peer's data. Previous versions used mss_clamp 23171da177e4SLinus Torvalds * here. I don't know if the value based on our guesses 23181da177e4SLinus Torvalds * of peer's MSS is better for the performance. It's more correct 23191da177e4SLinus Torvalds * but may be worse for the performance because of rcv_mss 23201da177e4SLinus Torvalds * fluctuations. --SAW 1998/11/1 23211da177e4SLinus Torvalds */ 2322463c84b9SArnaldo Carvalho de Melo int mss = icsk->icsk_ack.rcv_mss; 23231da177e4SLinus Torvalds int free_space = tcp_space(sk); 232486c1a045SFlorian Westphal int allowed_space = tcp_full_space(sk); 232586c1a045SFlorian Westphal int full_space = min_t(int, tp->window_clamp, allowed_space); 23261da177e4SLinus Torvalds int window; 23271da177e4SLinus Torvalds 23281da177e4SLinus Torvalds if (mss > full_space) 23291da177e4SLinus Torvalds mss = full_space; 23301da177e4SLinus Torvalds 2331b92edbe0SEric Dumazet if (free_space < (full_space >> 1)) { 2332463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.quick = 0; 23331da177e4SLinus Torvalds 2334180d8cd9SGlauber Costa if (sk_under_memory_pressure(sk)) 2335056834d9SIlpo Järvinen tp->rcv_ssthresh = min(tp->rcv_ssthresh, 2336056834d9SIlpo Järvinen 4U * tp->advmss); 23371da177e4SLinus Torvalds 233886c1a045SFlorian Westphal /* free_space might become our new window, make sure we don't 233986c1a045SFlorian Westphal * increase it due to wscale. 234086c1a045SFlorian Westphal */ 234186c1a045SFlorian Westphal free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale); 234286c1a045SFlorian Westphal 234386c1a045SFlorian Westphal /* if free space is less than mss estimate, or is below 1/16th 234486c1a045SFlorian Westphal * of the maximum allowed, try to move to zero-window, else 234586c1a045SFlorian Westphal * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and 234686c1a045SFlorian Westphal * new incoming data is dropped due to memory limits. 234786c1a045SFlorian Westphal * With large window, mss test triggers way too late in order 234886c1a045SFlorian Westphal * to announce zero window in time before rmem limit kicks in. 234986c1a045SFlorian Westphal */ 235086c1a045SFlorian Westphal if (free_space < (allowed_space >> 4) || free_space < mss) 23511da177e4SLinus Torvalds return 0; 23521da177e4SLinus Torvalds } 23531da177e4SLinus Torvalds 23541da177e4SLinus Torvalds if (free_space > tp->rcv_ssthresh) 23551da177e4SLinus Torvalds free_space = tp->rcv_ssthresh; 23561da177e4SLinus Torvalds 23571da177e4SLinus Torvalds /* Don't do rounding if we are using window scaling, since the 23581da177e4SLinus Torvalds * scaled window will not line up with the MSS boundary anyway. 23591da177e4SLinus Torvalds */ 23601da177e4SLinus Torvalds window = tp->rcv_wnd; 23611da177e4SLinus Torvalds if (tp->rx_opt.rcv_wscale) { 23621da177e4SLinus Torvalds window = free_space; 23631da177e4SLinus Torvalds 23641da177e4SLinus Torvalds /* Advertise enough space so that it won't get scaled away. 23651da177e4SLinus Torvalds * Import case: prevent zero window announcement if 23661da177e4SLinus Torvalds * 1<<rcv_wscale > mss. 23671da177e4SLinus Torvalds */ 23681da177e4SLinus Torvalds if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window) 23691da177e4SLinus Torvalds window = (((window >> tp->rx_opt.rcv_wscale) + 1) 23701da177e4SLinus Torvalds << tp->rx_opt.rcv_wscale); 23711da177e4SLinus Torvalds } else { 23721da177e4SLinus Torvalds /* Get the largest window that is a nice multiple of mss. 23731da177e4SLinus Torvalds * Window clamp already applied above. 23741da177e4SLinus Torvalds * If our current window offering is within 1 mss of the 23751da177e4SLinus Torvalds * free space we just keep it. This prevents the divide 23761da177e4SLinus Torvalds * and multiply from happening most of the time. 23771da177e4SLinus Torvalds * We also don't do any window rounding when the free space 23781da177e4SLinus Torvalds * is too small. 23791da177e4SLinus Torvalds */ 23801da177e4SLinus Torvalds if (window <= free_space - mss || window > free_space) 23811da177e4SLinus Torvalds window = (free_space / mss) * mss; 238284565070SJohn Heffner else if (mss == full_space && 2383b92edbe0SEric Dumazet free_space > window + (full_space >> 1)) 238484565070SJohn Heffner window = free_space; 23851da177e4SLinus Torvalds } 23861da177e4SLinus Torvalds 23871da177e4SLinus Torvalds return window; 23881da177e4SLinus Torvalds } 23891da177e4SLinus Torvalds 23904a17fc3aSIlpo Järvinen /* Collapses two adjacent SKB's during retransmission. */ 23914a17fc3aSIlpo Järvinen static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) 23921da177e4SLinus Torvalds { 23931da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2394fe067e8aSDavid S. Miller struct sk_buff *next_skb = tcp_write_queue_next(sk, skb); 2395058dc334SIlpo Järvinen int skb_size, next_skb_size; 23961da177e4SLinus Torvalds 2397058dc334SIlpo Järvinen skb_size = skb->len; 2398058dc334SIlpo Järvinen next_skb_size = next_skb->len; 23991da177e4SLinus Torvalds 2400058dc334SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); 24011da177e4SLinus Torvalds 24026859d494SIlpo Järvinen tcp_highest_sack_combine(sk, next_skb, skb); 2403a6963a6bSIlpo Järvinen 2404fe067e8aSDavid S. Miller tcp_unlink_write_queue(next_skb, sk); 24051da177e4SLinus Torvalds 2406058dc334SIlpo Järvinen skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size), 24071a4e2d09SArnaldo Carvalho de Melo next_skb_size); 24081da177e4SLinus Torvalds 240952d570aaSJarek Poplawski if (next_skb->ip_summed == CHECKSUM_PARTIAL) 241052d570aaSJarek Poplawski skb->ip_summed = CHECKSUM_PARTIAL; 24111da177e4SLinus Torvalds 241284fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 24131da177e4SLinus Torvalds skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); 24141da177e4SLinus Torvalds 24151da177e4SLinus Torvalds /* Update sequence range on original skb. */ 24161da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 24171da177e4SLinus Torvalds 2418e6c7d085SIlpo Järvinen /* Merge over control information. This moves PSH/FIN etc. over */ 24194de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; 24201da177e4SLinus Torvalds 24211da177e4SLinus Torvalds /* All done, get rid of second SKB and account for it so 24221da177e4SLinus Torvalds * packet counting does not break. 24231da177e4SLinus Torvalds */ 24244828e7f4SIlpo Järvinen TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; 2425b7689205SIlpo Järvinen 2426b7689205SIlpo Järvinen /* changed transmit queue under us so clear hints */ 2427ef9da47cSIlpo Järvinen tcp_clear_retrans_hints_partial(tp); 2428ef9da47cSIlpo Järvinen if (next_skb == tp->retransmit_skb_hint) 2429ef9da47cSIlpo Järvinen tp->retransmit_skb_hint = skb; 2430b7689205SIlpo Järvinen 2431797108d1SIlpo Järvinen tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); 2432797108d1SIlpo Järvinen 24333ab224beSHideo Aoki sk_wmem_free_skb(sk, next_skb); 24341da177e4SLinus Torvalds } 24351da177e4SLinus Torvalds 243667edfef7SAndi Kleen /* Check if coalescing SKBs is legal. */ 2437a2a385d6SEric Dumazet static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) 24384a17fc3aSIlpo Järvinen { 24394a17fc3aSIlpo Järvinen if (tcp_skb_pcount(skb) > 1) 2440a2a385d6SEric Dumazet return false; 24414a17fc3aSIlpo Järvinen /* TODO: SACK collapsing could be used to remove this condition */ 24424a17fc3aSIlpo Järvinen if (skb_shinfo(skb)->nr_frags != 0) 2443a2a385d6SEric Dumazet return false; 24444a17fc3aSIlpo Järvinen if (skb_cloned(skb)) 2445a2a385d6SEric Dumazet return false; 24464a17fc3aSIlpo Järvinen if (skb == tcp_send_head(sk)) 2447a2a385d6SEric Dumazet return false; 24484a17fc3aSIlpo Järvinen /* Some heurestics for collapsing over SACK'd could be invented */ 24494a17fc3aSIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 2450a2a385d6SEric Dumazet return false; 24514a17fc3aSIlpo Järvinen 2452a2a385d6SEric Dumazet return true; 24534a17fc3aSIlpo Järvinen } 24544a17fc3aSIlpo Järvinen 245567edfef7SAndi Kleen /* Collapse packets in the retransmit queue to make to create 245667edfef7SAndi Kleen * less packets on the wire. This is only done on retransmission. 245767edfef7SAndi Kleen */ 24584a17fc3aSIlpo Järvinen static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, 24594a17fc3aSIlpo Järvinen int space) 24604a17fc3aSIlpo Järvinen { 24614a17fc3aSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 24624a17fc3aSIlpo Järvinen struct sk_buff *skb = to, *tmp; 2463a2a385d6SEric Dumazet bool first = true; 24644a17fc3aSIlpo Järvinen 24654a17fc3aSIlpo Järvinen if (!sysctl_tcp_retrans_collapse) 24664a17fc3aSIlpo Järvinen return; 24674de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 24684a17fc3aSIlpo Järvinen return; 24694a17fc3aSIlpo Järvinen 24704a17fc3aSIlpo Järvinen tcp_for_write_queue_from_safe(skb, tmp, sk) { 24714a17fc3aSIlpo Järvinen if (!tcp_can_collapse(sk, skb)) 24724a17fc3aSIlpo Järvinen break; 24734a17fc3aSIlpo Järvinen 24744a17fc3aSIlpo Järvinen space -= skb->len; 24754a17fc3aSIlpo Järvinen 24764a17fc3aSIlpo Järvinen if (first) { 2477a2a385d6SEric Dumazet first = false; 24784a17fc3aSIlpo Järvinen continue; 24794a17fc3aSIlpo Järvinen } 24804a17fc3aSIlpo Järvinen 24814a17fc3aSIlpo Järvinen if (space < 0) 24824a17fc3aSIlpo Järvinen break; 24834a17fc3aSIlpo Järvinen /* Punt if not enough space exists in the first SKB for 24844a17fc3aSIlpo Järvinen * the data in the second 24854a17fc3aSIlpo Järvinen */ 2486a21d4572SEric Dumazet if (skb->len > skb_availroom(to)) 24874a17fc3aSIlpo Järvinen break; 24884a17fc3aSIlpo Järvinen 24894a17fc3aSIlpo Järvinen if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) 24904a17fc3aSIlpo Järvinen break; 24914a17fc3aSIlpo Järvinen 24924a17fc3aSIlpo Järvinen tcp_collapse_retrans(sk, to); 24934a17fc3aSIlpo Järvinen } 24944a17fc3aSIlpo Järvinen } 24954a17fc3aSIlpo Järvinen 24961da177e4SLinus Torvalds /* This retransmits one SKB. Policy decisions and retransmit queue 24971da177e4SLinus Torvalds * state updates are done by the caller. Returns non-zero if an 24981da177e4SLinus Torvalds * error occurred which prevented the send. 24991da177e4SLinus Torvalds */ 250093b174adSYuchung Cheng int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) 25011da177e4SLinus Torvalds { 25021da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 25035d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 25047d227cd2SSridhar Samudrala unsigned int cur_mss; 2505c84a5711SYuchung Cheng int err; 25061da177e4SLinus Torvalds 25075d424d5aSJohn Heffner /* Inconslusive MTU probe */ 25085d424d5aSJohn Heffner if (icsk->icsk_mtup.probe_size) { 25095d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 25105d424d5aSJohn Heffner } 25115d424d5aSJohn Heffner 25121da177e4SLinus Torvalds /* Do not sent more than we queued. 1/4 is reserved for possible 2513caa20d9aSStephen Hemminger * copying overhead: fragmentation, tunneling, mangling etc. 25141da177e4SLinus Torvalds */ 25151da177e4SLinus Torvalds if (atomic_read(&sk->sk_wmem_alloc) > 25161da177e4SLinus Torvalds min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf)) 25171da177e4SLinus Torvalds return -EAGAIN; 25181da177e4SLinus Torvalds 25191f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 25201f3279aeSEric Dumazet return -EBUSY; 25211f3279aeSEric Dumazet 25221da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 25231da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 25241da177e4SLinus Torvalds BUG(); 25251da177e4SLinus Torvalds if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 25261da177e4SLinus Torvalds return -ENOMEM; 25271da177e4SLinus Torvalds } 25281da177e4SLinus Torvalds 25297d227cd2SSridhar Samudrala if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 25307d227cd2SSridhar Samudrala return -EHOSTUNREACH; /* Routing failure or similar. */ 25317d227cd2SSridhar Samudrala 25320c54b85fSIlpo Järvinen cur_mss = tcp_current_mss(sk); 25337d227cd2SSridhar Samudrala 25341da177e4SLinus Torvalds /* If receiver has shrunk his window, and skb is out of 25351da177e4SLinus Torvalds * new window, do not retransmit it. The exception is the 25361da177e4SLinus Torvalds * case, when window is shrunk to zero. In this case 25371da177e4SLinus Torvalds * our retransmit serves as a zero window probe. 25381da177e4SLinus Torvalds */ 25399d4fb27dSJoe Perches if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && 25409d4fb27dSJoe Perches TCP_SKB_CB(skb)->seq != tp->snd_una) 25411da177e4SLinus Torvalds return -EAGAIN; 25421da177e4SLinus Torvalds 25431da177e4SLinus Torvalds if (skb->len > cur_mss) { 25446cc55e09SOctavian Purdila if (tcp_fragment(sk, skb, cur_mss, cur_mss, GFP_ATOMIC)) 25451da177e4SLinus Torvalds return -ENOMEM; /* We'll try again later. */ 254602276f3cSIlpo Järvinen } else { 25479eb9362eSIlpo Järvinen int oldpcount = tcp_skb_pcount(skb); 25489eb9362eSIlpo Järvinen 25499eb9362eSIlpo Järvinen if (unlikely(oldpcount > 1)) { 2550c52e2421SEric Dumazet if (skb_unclone(skb, GFP_ATOMIC)) 2551c52e2421SEric Dumazet return -ENOMEM; 255202276f3cSIlpo Järvinen tcp_init_tso_segs(sk, skb, cur_mss); 25539eb9362eSIlpo Järvinen tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb)); 25549eb9362eSIlpo Järvinen } 25551da177e4SLinus Torvalds } 25561da177e4SLinus Torvalds 25571da177e4SLinus Torvalds tcp_retrans_try_collapse(sk, skb, cur_mss); 25581da177e4SLinus Torvalds 25591da177e4SLinus Torvalds /* Make a copy, if the first transmission SKB clone we made 25601da177e4SLinus Torvalds * is still in somebody's hands, else make a clone. 25611da177e4SLinus Torvalds */ 25621da177e4SLinus Torvalds 256350bceae9SThomas Graf /* make sure skb->data is aligned on arches that require it 256450bceae9SThomas Graf * and check if ack-trimming & collapsing extended the headroom 256550bceae9SThomas Graf * beyond what csum_start can cover. 256650bceae9SThomas Graf */ 256750bceae9SThomas Graf if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) || 256850bceae9SThomas Graf skb_headroom(skb) >= 0xFFFF)) { 2569117632e6SEric Dumazet struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER, 2570117632e6SEric Dumazet GFP_ATOMIC); 2571c84a5711SYuchung Cheng err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) : 2572117632e6SEric Dumazet -ENOBUFS; 2573117632e6SEric Dumazet } else { 2574c84a5711SYuchung Cheng err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 2575117632e6SEric Dumazet } 2576c84a5711SYuchung Cheng 2577fc9f3501SEric Dumazet if (likely(!err)) { 2578c84a5711SYuchung Cheng TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS; 2579fc9f3501SEric Dumazet /* Update global TCP statistics. */ 2580fc9f3501SEric Dumazet TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); 2581fc9f3501SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 2582fc9f3501SEric Dumazet NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 2583fc9f3501SEric Dumazet tp->total_retrans++; 2584fc9f3501SEric Dumazet } 2585c84a5711SYuchung Cheng return err; 258693b174adSYuchung Cheng } 258793b174adSYuchung Cheng 258893b174adSYuchung Cheng int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) 258993b174adSYuchung Cheng { 259093b174adSYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 259193b174adSYuchung Cheng int err = __tcp_retransmit_skb(sk, skb); 25921da177e4SLinus Torvalds 25931da177e4SLinus Torvalds if (err == 0) { 25941da177e4SLinus Torvalds #if FASTRETRANS_DEBUG > 0 25951da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 2596e87cc472SJoe Perches net_dbg_ratelimited("retrans_out leaked\n"); 25971da177e4SLinus Torvalds } 25981da177e4SLinus Torvalds #endif 2599b08d6cb2SIlpo Järvinen if (!tp->retrans_out) 2600b08d6cb2SIlpo Järvinen tp->lost_retrans_low = tp->snd_nxt; 26011da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 26021da177e4SLinus Torvalds tp->retrans_out += tcp_skb_pcount(skb); 26031da177e4SLinus Torvalds 26041da177e4SLinus Torvalds /* Save stamp of the first retransmit. */ 26051da177e4SLinus Torvalds if (!tp->retrans_stamp) 26067faee5c0SEric Dumazet tp->retrans_stamp = tcp_skb_timestamp(skb); 26071da177e4SLinus Torvalds 26081da177e4SLinus Torvalds /* snd_nxt is stored to detect loss of retransmitted segment, 26091da177e4SLinus Torvalds * see tcp_input.c tcp_sacktag_write_queue(). 26101da177e4SLinus Torvalds */ 26111da177e4SLinus Torvalds TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt; 26121f3279aeSEric Dumazet } else if (err != -EBUSY) { 261324ab6becSYuchung Cheng NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL); 26141da177e4SLinus Torvalds } 26156e08d5e3SYuchung Cheng 26166e08d5e3SYuchung Cheng if (tp->undo_retrans < 0) 26176e08d5e3SYuchung Cheng tp->undo_retrans = 0; 26186e08d5e3SYuchung Cheng tp->undo_retrans += tcp_skb_pcount(skb); 26191da177e4SLinus Torvalds return err; 26201da177e4SLinus Torvalds } 26211da177e4SLinus Torvalds 262267edfef7SAndi Kleen /* Check if we forward retransmits are possible in the current 262367edfef7SAndi Kleen * window/congestion state. 262467edfef7SAndi Kleen */ 2625a2a385d6SEric Dumazet static bool tcp_can_forward_retransmit(struct sock *sk) 2626b5afe7bcSIlpo Järvinen { 2627b5afe7bcSIlpo Järvinen const struct inet_connection_sock *icsk = inet_csk(sk); 2628cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 2629b5afe7bcSIlpo Järvinen 2630b5afe7bcSIlpo Järvinen /* Forward retransmissions are possible only during Recovery. */ 2631b5afe7bcSIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Recovery) 2632a2a385d6SEric Dumazet return false; 2633b5afe7bcSIlpo Järvinen 2634b5afe7bcSIlpo Järvinen /* No forward retransmissions in Reno are possible. */ 2635b5afe7bcSIlpo Järvinen if (tcp_is_reno(tp)) 2636a2a385d6SEric Dumazet return false; 2637b5afe7bcSIlpo Järvinen 2638b5afe7bcSIlpo Järvinen /* Yeah, we have to make difficult choice between forward transmission 2639b5afe7bcSIlpo Järvinen * and retransmission... Both ways have their merits... 2640b5afe7bcSIlpo Järvinen * 2641b5afe7bcSIlpo Järvinen * For now we do not retransmit anything, while we have some new 2642b5afe7bcSIlpo Järvinen * segments to send. In the other cases, follow rule 3 for 2643b5afe7bcSIlpo Järvinen * NextSeg() specified in RFC3517. 2644b5afe7bcSIlpo Järvinen */ 2645b5afe7bcSIlpo Järvinen 2646b5afe7bcSIlpo Järvinen if (tcp_may_send_now(sk)) 2647a2a385d6SEric Dumazet return false; 2648b5afe7bcSIlpo Järvinen 2649a2a385d6SEric Dumazet return true; 2650b5afe7bcSIlpo Järvinen } 2651b5afe7bcSIlpo Järvinen 26521da177e4SLinus Torvalds /* This gets called after a retransmit timeout, and the initially 26531da177e4SLinus Torvalds * retransmitted data is acknowledged. It tries to continue 26541da177e4SLinus Torvalds * resending the rest of the retransmit queue, until either 26551da177e4SLinus Torvalds * we've sent it all or the congestion window limit is reached. 26561da177e4SLinus Torvalds * If doing SACK, the first ACK which comes back for a timeout 26571da177e4SLinus Torvalds * based retransmit packet might feed us FACK information again. 26581da177e4SLinus Torvalds * If so, we use it to avoid unnecessarily retransmissions. 26591da177e4SLinus Torvalds */ 26601da177e4SLinus Torvalds void tcp_xmit_retransmit_queue(struct sock *sk) 26611da177e4SLinus Torvalds { 26626687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 26631da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 26641da177e4SLinus Torvalds struct sk_buff *skb; 26650e1c54c2SIlpo Järvinen struct sk_buff *hole = NULL; 2666618d9f25SIlpo Järvinen u32 last_lost; 266761eb55f4SIlpo Järvinen int mib_idx; 26680e1c54c2SIlpo Järvinen int fwd_rexmitting = 0; 26696a438bbeSStephen Hemminger 267045e77d31SIlpo Järvinen if (!tp->packets_out) 267145e77d31SIlpo Järvinen return; 267245e77d31SIlpo Järvinen 267308ebd172SIlpo Järvinen if (!tp->lost_out) 267408ebd172SIlpo Järvinen tp->retransmit_high = tp->snd_una; 267508ebd172SIlpo Järvinen 2676618d9f25SIlpo Järvinen if (tp->retransmit_skb_hint) { 26776a438bbeSStephen Hemminger skb = tp->retransmit_skb_hint; 2678618d9f25SIlpo Järvinen last_lost = TCP_SKB_CB(skb)->end_seq; 2679618d9f25SIlpo Järvinen if (after(last_lost, tp->retransmit_high)) 2680618d9f25SIlpo Järvinen last_lost = tp->retransmit_high; 2681618d9f25SIlpo Järvinen } else { 2682fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 2683618d9f25SIlpo Järvinen last_lost = tp->snd_una; 2684618d9f25SIlpo Järvinen } 26851da177e4SLinus Torvalds 2686fe067e8aSDavid S. Miller tcp_for_write_queue_from(skb, sk) { 26871da177e4SLinus Torvalds __u8 sacked = TCP_SKB_CB(skb)->sacked; 26881da177e4SLinus Torvalds 2689fe067e8aSDavid S. Miller if (skb == tcp_send_head(sk)) 2690fe067e8aSDavid S. Miller break; 26916a438bbeSStephen Hemminger /* we could do better than to assign each time */ 26920e1c54c2SIlpo Järvinen if (hole == NULL) 26936a438bbeSStephen Hemminger tp->retransmit_skb_hint = skb; 26946a438bbeSStephen Hemminger 26951da177e4SLinus Torvalds /* Assume this retransmit will generate 26961da177e4SLinus Torvalds * only one packet for congestion window 26971da177e4SLinus Torvalds * calculation purposes. This works because 26981da177e4SLinus Torvalds * tcp_retransmit_skb() will chop up the 26991da177e4SLinus Torvalds * packet to be MSS sized and all the 27001da177e4SLinus Torvalds * packet counting works out. 27011da177e4SLinus Torvalds */ 27021da177e4SLinus Torvalds if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) 27031da177e4SLinus Torvalds return; 27040e1c54c2SIlpo Järvinen 27050e1c54c2SIlpo Järvinen if (fwd_rexmitting) { 27060e1c54c2SIlpo Järvinen begin_fwd: 27070e1c54c2SIlpo Järvinen if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp))) 2708006f582cSIlpo Järvinen break; 27090e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFORWARDRETRANS; 27100e1c54c2SIlpo Järvinen 27110e1c54c2SIlpo Järvinen } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) { 2712618d9f25SIlpo Järvinen tp->retransmit_high = last_lost; 27130e1c54c2SIlpo Järvinen if (!tcp_can_forward_retransmit(sk)) 27140e1c54c2SIlpo Järvinen break; 27150e1c54c2SIlpo Järvinen /* Backtrack if necessary to non-L'ed skb */ 27160e1c54c2SIlpo Järvinen if (hole != NULL) { 27170e1c54c2SIlpo Järvinen skb = hole; 27180e1c54c2SIlpo Järvinen hole = NULL; 27190e1c54c2SIlpo Järvinen } 27200e1c54c2SIlpo Järvinen fwd_rexmitting = 1; 27210e1c54c2SIlpo Järvinen goto begin_fwd; 27220e1c54c2SIlpo Järvinen 27230e1c54c2SIlpo Järvinen } else if (!(sacked & TCPCB_LOST)) { 2724ac11ba75SIlpo Järvinen if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) 27250e1c54c2SIlpo Järvinen hole = skb; 272661eb55f4SIlpo Järvinen continue; 27271da177e4SLinus Torvalds 27280e1c54c2SIlpo Järvinen } else { 2729618d9f25SIlpo Järvinen last_lost = TCP_SKB_CB(skb)->end_seq; 27300e1c54c2SIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Loss) 27310e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFASTRETRANS; 27320e1c54c2SIlpo Järvinen else 27330e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; 27340e1c54c2SIlpo Järvinen } 27350e1c54c2SIlpo Järvinen 27360e1c54c2SIlpo Järvinen if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) 273761eb55f4SIlpo Järvinen continue; 273840b215e5SPavel Emelyanov 273924ab6becSYuchung Cheng if (tcp_retransmit_skb(sk, skb)) 27401da177e4SLinus Torvalds return; 274124ab6becSYuchung Cheng 2742de0744afSPavel Emelyanov NET_INC_STATS_BH(sock_net(sk), mib_idx); 27431da177e4SLinus Torvalds 2744684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 2745a262f0cdSNandita Dukkipati tp->prr_out += tcp_skb_pcount(skb); 2746a262f0cdSNandita Dukkipati 2747fe067e8aSDavid S. Miller if (skb == tcp_write_queue_head(sk)) 2748463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 27493f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, 27503f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 27511da177e4SLinus Torvalds } 27521da177e4SLinus Torvalds } 27531da177e4SLinus Torvalds 27541da177e4SLinus Torvalds /* Send a fin. The caller locks the socket for us. This cannot be 27551da177e4SLinus Torvalds * allowed to fail queueing a FIN frame under any circumstances. 27561da177e4SLinus Torvalds */ 27571da177e4SLinus Torvalds void tcp_send_fin(struct sock *sk) 27581da177e4SLinus Torvalds { 27591da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2760fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_write_queue_tail(sk); 27611da177e4SLinus Torvalds int mss_now; 27621da177e4SLinus Torvalds 27631da177e4SLinus Torvalds /* Optimization, tack on the FIN if we have a queue of 27641da177e4SLinus Torvalds * unsent frames. But be careful about outgoing SACKS 27651da177e4SLinus Torvalds * and IP options. 27661da177e4SLinus Torvalds */ 27670c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 27681da177e4SLinus Torvalds 2769fe067e8aSDavid S. Miller if (tcp_send_head(sk) != NULL) { 27704de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_FIN; 27711da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq++; 27721da177e4SLinus Torvalds tp->write_seq++; 27731da177e4SLinus Torvalds } else { 27741da177e4SLinus Torvalds /* Socket is locked, keep trying until memory is available. */ 27751da177e4SLinus Torvalds for (;;) { 2776d22e1537SJosh Hunt skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation); 27771da177e4SLinus Torvalds if (skb) 27781da177e4SLinus Torvalds break; 27791da177e4SLinus Torvalds yield(); 27801da177e4SLinus Torvalds } 27811da177e4SLinus Torvalds /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 2782e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tp->write_seq, 2783a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_FIN); 27841da177e4SLinus Torvalds tcp_queue_skb(sk, skb); 27851da177e4SLinus Torvalds } 27869e412ba7SIlpo Järvinen __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF); 27871da177e4SLinus Torvalds } 27881da177e4SLinus Torvalds 27891da177e4SLinus Torvalds /* We get here when a process closes a file descriptor (either due to 27901da177e4SLinus Torvalds * an explicit close() or as a byproduct of exit()'ing) and there 27911da177e4SLinus Torvalds * was unread data in the receive queue. This behavior is recommended 279265bb723cSGerrit Renker * by RFC 2525, section 2.17. -DaveM 27931da177e4SLinus Torvalds */ 2794dd0fc66fSAl Viro void tcp_send_active_reset(struct sock *sk, gfp_t priority) 27951da177e4SLinus Torvalds { 27961da177e4SLinus Torvalds struct sk_buff *skb; 27971da177e4SLinus Torvalds 27981da177e4SLinus Torvalds /* NOTE: No TCP options attached and we never retransmit this. */ 27991da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, priority); 28001da177e4SLinus Torvalds if (!skb) { 28014e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 28021da177e4SLinus Torvalds return; 28031da177e4SLinus Torvalds } 28041da177e4SLinus Torvalds 28051da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 28061da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 2807e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), 2808a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_RST); 28091da177e4SLinus Torvalds /* Send it off. */ 2810dfb4b9dcSDavid S. Miller if (tcp_transmit_skb(sk, skb, 0, priority)) 28114e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 281226af65cbSSridhar Samudrala 281381cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); 28141da177e4SLinus Torvalds } 28151da177e4SLinus Torvalds 281667edfef7SAndi Kleen /* Send a crossed SYN-ACK during socket establishment. 281767edfef7SAndi Kleen * WARNING: This routine must only be called when we have already sent 28181da177e4SLinus Torvalds * a SYN packet that crossed the incoming SYN that caused this routine 28191da177e4SLinus Torvalds * to get called. If this assumption fails then the initial rcv_wnd 28201da177e4SLinus Torvalds * and rcv_wscale values will not be correct. 28211da177e4SLinus Torvalds */ 28221da177e4SLinus Torvalds int tcp_send_synack(struct sock *sk) 28231da177e4SLinus Torvalds { 28241da177e4SLinus Torvalds struct sk_buff *skb; 28251da177e4SLinus Torvalds 2826fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 28274de075e0SEric Dumazet if (skb == NULL || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 282891df42beSJoe Perches pr_debug("%s: wrong queue state\n", __func__); 28291da177e4SLinus Torvalds return -EFAULT; 28301da177e4SLinus Torvalds } 28314de075e0SEric Dumazet if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { 28321da177e4SLinus Torvalds if (skb_cloned(skb)) { 28331da177e4SLinus Torvalds struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC); 28341da177e4SLinus Torvalds if (nskb == NULL) 28351da177e4SLinus Torvalds return -ENOMEM; 2836fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 2837f4a775d1SEric Dumazet __skb_header_release(nskb); 2838fe067e8aSDavid S. Miller __tcp_add_write_queue_head(sk, nskb); 28393ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 28403ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 28413ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 28421da177e4SLinus Torvalds skb = nskb; 28431da177e4SLinus Torvalds } 28441da177e4SLinus Torvalds 28454de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; 2846735d3831SFlorian Westphal tcp_ecn_send_synack(sk, skb); 28471da177e4SLinus Torvalds } 2848dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 28491da177e4SLinus Torvalds } 28501da177e4SLinus Torvalds 28514aea39c1SEric Dumazet /** 28524aea39c1SEric Dumazet * tcp_make_synack - Prepare a SYN-ACK. 28534aea39c1SEric Dumazet * sk: listener socket 28544aea39c1SEric Dumazet * dst: dst entry attached to the SYNACK 28554aea39c1SEric Dumazet * req: request_sock pointer 28564aea39c1SEric Dumazet * 28574aea39c1SEric Dumazet * Allocate one skb and build a SYNACK packet. 28584aea39c1SEric Dumazet * @dst is consumed : Caller should not use it again. 28594aea39c1SEric Dumazet */ 28601da177e4SLinus Torvalds struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst, 2861e6b4d113SWilliam Allen Simpson struct request_sock *req, 28628336886fSJerry Chu struct tcp_fastopen_cookie *foc) 28631da177e4SLinus Torvalds { 2864bd0388aeSWilliam Allen Simpson struct tcp_out_options opts; 28652e6599cbSArnaldo Carvalho de Melo struct inet_request_sock *ireq = inet_rsk(req); 28661da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 28671da177e4SLinus Torvalds struct tcphdr *th; 28681da177e4SLinus Torvalds struct sk_buff *skb; 2869cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 2870bd0388aeSWilliam Allen Simpson int tcp_header_size; 2871f5fff5dcSTom Quetchenbach int mss; 28721da177e4SLinus Torvalds 2873a0b8486cSEric Dumazet skb = sock_wmalloc(sk, MAX_TCP_HEADER, 1, GFP_ATOMIC); 28744aea39c1SEric Dumazet if (unlikely(!skb)) { 28754aea39c1SEric Dumazet dst_release(dst); 28761da177e4SLinus Torvalds return NULL; 28774aea39c1SEric Dumazet } 28781da177e4SLinus Torvalds /* Reserve space for headers. */ 28791da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 28801da177e4SLinus Torvalds 28814aea39c1SEric Dumazet skb_dst_set(skb, dst); 2882ca10b9e9SEric Dumazet security_skb_owned_by(skb, sk); 28831da177e4SLinus Torvalds 28840dbaee3bSDavid S. Miller mss = dst_metric_advmss(dst); 2885f5fff5dcSTom Quetchenbach if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss) 2886f5fff5dcSTom Quetchenbach mss = tp->rx_opt.user_mss; 2887f5fff5dcSTom Quetchenbach 288833ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 28898b5f12d0SFlorian Westphal #ifdef CONFIG_SYN_COOKIES 28908b5f12d0SFlorian Westphal if (unlikely(req->cookie_ts)) 28917faee5c0SEric Dumazet skb->skb_mstamp.stamp_jiffies = cookie_init_timestamp(req); 28928b5f12d0SFlorian Westphal else 28938b5f12d0SFlorian Westphal #endif 28947faee5c0SEric Dumazet skb_mstamp_get(&skb->skb_mstamp); 28951a2c6181SChristoph Paasch tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, &md5, 28961a2c6181SChristoph Paasch foc) + sizeof(*th); 289733ad798cSAdam Langley 2898aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 2899aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 29001da177e4SLinus Torvalds 2901aa8223c7SArnaldo Carvalho de Melo th = tcp_hdr(skb); 29021da177e4SLinus Torvalds memset(th, 0, sizeof(struct tcphdr)); 29031da177e4SLinus Torvalds th->syn = 1; 29041da177e4SLinus Torvalds th->ack = 1; 2905735d3831SFlorian Westphal tcp_ecn_make_synack(req, th, sk); 2906b44084c2SEric Dumazet th->source = htons(ireq->ir_num); 2907634fb979SEric Dumazet th->dest = ireq->ir_rmt_port; 2908e870a8efSIlpo Järvinen /* Setting of flags are superfluous here for callers (and ECE is 2909e870a8efSIlpo Järvinen * not even correctly set) 2910e870a8efSIlpo Järvinen */ 2911e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn, 2912a3433f35SChangli Gao TCPHDR_SYN | TCPHDR_ACK); 29134957faadSWilliam Allen Simpson 29141da177e4SLinus Torvalds th->seq = htonl(TCP_SKB_CB(skb)->seq); 29158336886fSJerry Chu /* XXX data is queued and acked as is. No buffer/window check */ 29168336886fSJerry Chu th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt); 29171da177e4SLinus Torvalds 29181da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 2919600ff0c2SIlpo Järvinen th->window = htons(min(req->rcv_wnd, 65535U)); 2920bd0388aeSWilliam Allen Simpson tcp_options_write((__be32 *)(th + 1), tp, &opts); 29211da177e4SLinus Torvalds th->doff = (tcp_header_size >> 2); 2922a0b8486cSEric Dumazet TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_OUTSEGS); 2923cfb6eeb4SYOSHIFUJI Hideaki 2924cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 2925cfb6eeb4SYOSHIFUJI Hideaki /* Okay, we have all we need - do the md5 hash if needed */ 2926cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 2927bd0388aeSWilliam Allen Simpson tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, 292849a72dfbSAdam Langley md5, NULL, req, skb); 2929cfb6eeb4SYOSHIFUJI Hideaki } 2930cfb6eeb4SYOSHIFUJI Hideaki #endif 2931cfb6eeb4SYOSHIFUJI Hideaki 2932*b50edd78SEric Dumazet /* Do not fool tcpdump (if any), clean our debris */ 2933*b50edd78SEric Dumazet skb->tstamp.tv64 = 0; 29341da177e4SLinus Torvalds return skb; 29351da177e4SLinus Torvalds } 29364bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_make_synack); 29371da177e4SLinus Torvalds 293881164413SDaniel Borkmann static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst) 293981164413SDaniel Borkmann { 294081164413SDaniel Borkmann struct inet_connection_sock *icsk = inet_csk(sk); 294181164413SDaniel Borkmann const struct tcp_congestion_ops *ca; 294281164413SDaniel Borkmann u32 ca_key = dst_metric(dst, RTAX_CC_ALGO); 294381164413SDaniel Borkmann 294481164413SDaniel Borkmann if (ca_key == TCP_CA_UNSPEC) 294581164413SDaniel Borkmann return; 294681164413SDaniel Borkmann 294781164413SDaniel Borkmann rcu_read_lock(); 294881164413SDaniel Borkmann ca = tcp_ca_find_key(ca_key); 294981164413SDaniel Borkmann if (likely(ca && try_module_get(ca->owner))) { 295081164413SDaniel Borkmann module_put(icsk->icsk_ca_ops->owner); 295181164413SDaniel Borkmann icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst); 295281164413SDaniel Borkmann icsk->icsk_ca_ops = ca; 295381164413SDaniel Borkmann } 295481164413SDaniel Borkmann rcu_read_unlock(); 295581164413SDaniel Borkmann } 295681164413SDaniel Borkmann 295767edfef7SAndi Kleen /* Do all connect socket setups that can be done AF independent. */ 2958f7e56a76Sstephen hemminger static void tcp_connect_init(struct sock *sk) 29591da177e4SLinus Torvalds { 2960cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 29611da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 29621da177e4SLinus Torvalds __u8 rcv_wscale; 29631da177e4SLinus Torvalds 29641da177e4SLinus Torvalds /* We'll fix this up when we get a response from the other end. 29651da177e4SLinus Torvalds * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 29661da177e4SLinus Torvalds */ 29671da177e4SLinus Torvalds tp->tcp_header_len = sizeof(struct tcphdr) + 2968bb5b7c11SDavid S. Miller (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0); 29691da177e4SLinus Torvalds 2970cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 2971cfb6eeb4SYOSHIFUJI Hideaki if (tp->af_specific->md5_lookup(sk, sk) != NULL) 2972cfb6eeb4SYOSHIFUJI Hideaki tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 2973cfb6eeb4SYOSHIFUJI Hideaki #endif 2974cfb6eeb4SYOSHIFUJI Hideaki 29751da177e4SLinus Torvalds /* If user gave his TCP_MAXSEG, record it to clamp */ 29761da177e4SLinus Torvalds if (tp->rx_opt.user_mss) 29771da177e4SLinus Torvalds tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 29781da177e4SLinus Torvalds tp->max_window = 0; 29795d424d5aSJohn Heffner tcp_mtup_init(sk); 29801da177e4SLinus Torvalds tcp_sync_mss(sk, dst_mtu(dst)); 29811da177e4SLinus Torvalds 298281164413SDaniel Borkmann tcp_ca_dst_init(sk, dst); 298381164413SDaniel Borkmann 29841da177e4SLinus Torvalds if (!tp->window_clamp) 29851da177e4SLinus Torvalds tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 29860dbaee3bSDavid S. Miller tp->advmss = dst_metric_advmss(dst); 2987f5fff5dcSTom Quetchenbach if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss) 2988f5fff5dcSTom Quetchenbach tp->advmss = tp->rx_opt.user_mss; 2989f5fff5dcSTom Quetchenbach 29901da177e4SLinus Torvalds tcp_initialize_rcv_mss(sk); 29911da177e4SLinus Torvalds 2992e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 2993e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 2994e88c64f0SHagen Paul Pfeifer (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) 2995e88c64f0SHagen Paul Pfeifer tp->window_clamp = tcp_full_space(sk); 2996e88c64f0SHagen Paul Pfeifer 29971da177e4SLinus Torvalds tcp_select_initial_window(tcp_full_space(sk), 29981da177e4SLinus Torvalds tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 29991da177e4SLinus Torvalds &tp->rcv_wnd, 30001da177e4SLinus Torvalds &tp->window_clamp, 3001bb5b7c11SDavid S. Miller sysctl_tcp_window_scaling, 300231d12926Slaurent chavey &rcv_wscale, 300331d12926Slaurent chavey dst_metric(dst, RTAX_INITRWND)); 30041da177e4SLinus Torvalds 30051da177e4SLinus Torvalds tp->rx_opt.rcv_wscale = rcv_wscale; 30061da177e4SLinus Torvalds tp->rcv_ssthresh = tp->rcv_wnd; 30071da177e4SLinus Torvalds 30081da177e4SLinus Torvalds sk->sk_err = 0; 30091da177e4SLinus Torvalds sock_reset_flag(sk, SOCK_DONE); 30101da177e4SLinus Torvalds tp->snd_wnd = 0; 3011ee7537b6SHantzis Fotis tcp_init_wl(tp, 0); 30121da177e4SLinus Torvalds tp->snd_una = tp->write_seq; 30131da177e4SLinus Torvalds tp->snd_sml = tp->write_seq; 301433f5f57eSIlpo Järvinen tp->snd_up = tp->write_seq; 3015370816aeSPavel Emelyanov tp->snd_nxt = tp->write_seq; 3016ee995283SPavel Emelyanov 3017ee995283SPavel Emelyanov if (likely(!tp->repair)) 30181da177e4SLinus Torvalds tp->rcv_nxt = 0; 3019c7781a6eSAndrew Vagin else 3020c7781a6eSAndrew Vagin tp->rcv_tstamp = tcp_time_stamp; 3021ee995283SPavel Emelyanov tp->rcv_wup = tp->rcv_nxt; 3022ee995283SPavel Emelyanov tp->copied_seq = tp->rcv_nxt; 30231da177e4SLinus Torvalds 3024463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; 3025463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_retransmits = 0; 30261da177e4SLinus Torvalds tcp_clear_retrans(tp); 30271da177e4SLinus Torvalds } 30281da177e4SLinus Torvalds 3029783237e8SYuchung Cheng static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) 3030783237e8SYuchung Cheng { 3031783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3032783237e8SYuchung Cheng struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 3033783237e8SYuchung Cheng 3034783237e8SYuchung Cheng tcb->end_seq += skb->len; 3035f4a775d1SEric Dumazet __skb_header_release(skb); 3036783237e8SYuchung Cheng __tcp_add_write_queue_tail(sk, skb); 3037783237e8SYuchung Cheng sk->sk_wmem_queued += skb->truesize; 3038783237e8SYuchung Cheng sk_mem_charge(sk, skb->truesize); 3039783237e8SYuchung Cheng tp->write_seq = tcb->end_seq; 3040783237e8SYuchung Cheng tp->packets_out += tcp_skb_pcount(skb); 3041783237e8SYuchung Cheng } 3042783237e8SYuchung Cheng 3043783237e8SYuchung Cheng /* Build and send a SYN with data and (cached) Fast Open cookie. However, 3044783237e8SYuchung Cheng * queue a data-only packet after the regular SYN, such that regular SYNs 3045783237e8SYuchung Cheng * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges 3046783237e8SYuchung Cheng * only the SYN sequence, the data are retransmitted in the first ACK. 3047783237e8SYuchung Cheng * If cookie is not cached or other error occurs, falls back to send a 3048783237e8SYuchung Cheng * regular SYN with Fast Open cookie request option. 3049783237e8SYuchung Cheng */ 3050783237e8SYuchung Cheng static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) 3051783237e8SYuchung Cheng { 3052783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3053783237e8SYuchung Cheng struct tcp_fastopen_request *fo = tp->fastopen_req; 305457be5bdaSAl Viro int syn_loss = 0, space, err = 0, copied; 3055aab48743SYuchung Cheng unsigned long last_syn_loss = 0; 3056355a901eSEric Dumazet struct sk_buff *syn_data; 3057783237e8SYuchung Cheng 305867da22d2SYuchung Cheng tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ 3059aab48743SYuchung Cheng tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie, 3060aab48743SYuchung Cheng &syn_loss, &last_syn_loss); 3061aab48743SYuchung Cheng /* Recurring FO SYN losses: revert to regular handshake temporarily */ 3062aab48743SYuchung Cheng if (syn_loss > 1 && 3063aab48743SYuchung Cheng time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) { 3064aab48743SYuchung Cheng fo->cookie.len = -1; 3065aab48743SYuchung Cheng goto fallback; 3066aab48743SYuchung Cheng } 3067aab48743SYuchung Cheng 306867da22d2SYuchung Cheng if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE) 306967da22d2SYuchung Cheng fo->cookie.len = -1; 307067da22d2SYuchung Cheng else if (fo->cookie.len <= 0) 3071783237e8SYuchung Cheng goto fallback; 3072783237e8SYuchung Cheng 3073783237e8SYuchung Cheng /* MSS for SYN-data is based on cached MSS and bounded by PMTU and 3074783237e8SYuchung Cheng * user-MSS. Reserve maximum option space for middleboxes that add 3075783237e8SYuchung Cheng * private TCP options. The cost is reduced data space in SYN :( 3076783237e8SYuchung Cheng */ 3077783237e8SYuchung Cheng if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp) 3078783237e8SYuchung Cheng tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 30791b63edd6SYuchung Cheng space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) - 3080783237e8SYuchung Cheng MAX_TCP_OPTION_SPACE; 3081783237e8SYuchung Cheng 3082f5ddcbbbSEric Dumazet space = min_t(size_t, space, fo->size); 3083f5ddcbbbSEric Dumazet 3084f5ddcbbbSEric Dumazet /* limit to order-0 allocations */ 3085f5ddcbbbSEric Dumazet space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER)); 3086f5ddcbbbSEric Dumazet 3087355a901eSEric Dumazet syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation); 3088355a901eSEric Dumazet if (!syn_data) 3089783237e8SYuchung Cheng goto fallback; 3090355a901eSEric Dumazet syn_data->ip_summed = CHECKSUM_PARTIAL; 3091355a901eSEric Dumazet memcpy(syn_data->cb, syn->cb, sizeof(syn->cb)); 309257be5bdaSAl Viro copied = copy_from_iter(skb_put(syn_data, space), space, 309357be5bdaSAl Viro &fo->data->msg_iter); 309457be5bdaSAl Viro if (unlikely(!copied)) { 3095355a901eSEric Dumazet kfree_skb(syn_data); 3096783237e8SYuchung Cheng goto fallback; 3097783237e8SYuchung Cheng } 309857be5bdaSAl Viro if (copied != space) { 309957be5bdaSAl Viro skb_trim(syn_data, copied); 310057be5bdaSAl Viro space = copied; 310157be5bdaSAl Viro } 3102783237e8SYuchung Cheng 3103355a901eSEric Dumazet /* No more data pending in inet_wait_for_connect() */ 3104355a901eSEric Dumazet if (space == fo->size) 3105355a901eSEric Dumazet fo->data = NULL; 3106355a901eSEric Dumazet fo->copied = space; 3107783237e8SYuchung Cheng 3108355a901eSEric Dumazet tcp_connect_queue_skb(sk, syn_data); 3109355a901eSEric Dumazet 3110355a901eSEric Dumazet err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation); 3111355a901eSEric Dumazet 3112355a901eSEric Dumazet syn->skb_mstamp = syn_data->skb_mstamp; 3113355a901eSEric Dumazet 3114355a901eSEric Dumazet /* Now full SYN+DATA was cloned and sent (or not), 3115355a901eSEric Dumazet * remove the SYN from the original skb (syn_data) 3116355a901eSEric Dumazet * we keep in write queue in case of a retransmit, as we 3117355a901eSEric Dumazet * also have the SYN packet (with no data) in the same queue. 3118431a9124SEric Dumazet */ 3119355a901eSEric Dumazet TCP_SKB_CB(syn_data)->seq++; 3120355a901eSEric Dumazet TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH; 3121355a901eSEric Dumazet if (!err) { 312267da22d2SYuchung Cheng tp->syn_data = (fo->copied > 0); 3123f19c29e3SYuchung Cheng NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT); 3124783237e8SYuchung Cheng goto done; 3125783237e8SYuchung Cheng } 3126783237e8SYuchung Cheng 3127783237e8SYuchung Cheng fallback: 3128783237e8SYuchung Cheng /* Send a regular SYN with Fast Open cookie request option */ 3129783237e8SYuchung Cheng if (fo->cookie.len > 0) 3130783237e8SYuchung Cheng fo->cookie.len = 0; 3131783237e8SYuchung Cheng err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); 3132783237e8SYuchung Cheng if (err) 3133783237e8SYuchung Cheng tp->syn_fastopen = 0; 3134783237e8SYuchung Cheng done: 3135783237e8SYuchung Cheng fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ 3136783237e8SYuchung Cheng return err; 3137783237e8SYuchung Cheng } 3138783237e8SYuchung Cheng 313967edfef7SAndi Kleen /* Build a SYN and send it off. */ 31401da177e4SLinus Torvalds int tcp_connect(struct sock *sk) 31411da177e4SLinus Torvalds { 31421da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 31431da177e4SLinus Torvalds struct sk_buff *buff; 3144ee586811SEric Paris int err; 31451da177e4SLinus Torvalds 31461da177e4SLinus Torvalds tcp_connect_init(sk); 31471da177e4SLinus Torvalds 31482b916477SAndrey Vagin if (unlikely(tp->repair)) { 31492b916477SAndrey Vagin tcp_finish_connect(sk, NULL); 31502b916477SAndrey Vagin return 0; 31512b916477SAndrey Vagin } 31522b916477SAndrey Vagin 3153355a901eSEric Dumazet buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation); 3154355a901eSEric Dumazet if (unlikely(!buff)) 31551da177e4SLinus Torvalds return -ENOBUFS; 31561da177e4SLinus Torvalds 3157a3433f35SChangli Gao tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); 31587faee5c0SEric Dumazet tp->retrans_stamp = tcp_time_stamp; 3159783237e8SYuchung Cheng tcp_connect_queue_skb(sk, buff); 3160735d3831SFlorian Westphal tcp_ecn_send_syn(sk, buff); 31611da177e4SLinus Torvalds 3162783237e8SYuchung Cheng /* Send off SYN; include data in Fast Open. */ 3163783237e8SYuchung Cheng err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : 3164783237e8SYuchung Cheng tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); 3165ee586811SEric Paris if (err == -ECONNREFUSED) 3166ee586811SEric Paris return err; 3167bd37a088SWei Yongjun 3168bd37a088SWei Yongjun /* We change tp->snd_nxt after the tcp_transmit_skb() call 3169bd37a088SWei Yongjun * in order to make this packet get counted in tcpOutSegs. 3170bd37a088SWei Yongjun */ 3171bd37a088SWei Yongjun tp->snd_nxt = tp->write_seq; 3172bd37a088SWei Yongjun tp->pushed_seq = tp->write_seq; 317381cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); 31741da177e4SLinus Torvalds 31751da177e4SLinus Torvalds /* Timer for repeating the SYN until an answer. */ 31763f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 31773f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 31781da177e4SLinus Torvalds return 0; 31791da177e4SLinus Torvalds } 31804bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_connect); 31811da177e4SLinus Torvalds 31821da177e4SLinus Torvalds /* Send out a delayed ack, the caller does the policy checking 31831da177e4SLinus Torvalds * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 31841da177e4SLinus Torvalds * for details. 31851da177e4SLinus Torvalds */ 31861da177e4SLinus Torvalds void tcp_send_delayed_ack(struct sock *sk) 31871da177e4SLinus Torvalds { 3188463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 3189463c84b9SArnaldo Carvalho de Melo int ato = icsk->icsk_ack.ato; 31901da177e4SLinus Torvalds unsigned long timeout; 31911da177e4SLinus Torvalds 31929890092eSFlorian Westphal tcp_ca_event(sk, CA_EVENT_DELAYED_ACK); 31939890092eSFlorian Westphal 31941da177e4SLinus Torvalds if (ato > TCP_DELACK_MIN) { 3195463c84b9SArnaldo Carvalho de Melo const struct tcp_sock *tp = tcp_sk(sk); 31961da177e4SLinus Torvalds int max_ato = HZ / 2; 31971da177e4SLinus Torvalds 3198056834d9SIlpo Järvinen if (icsk->icsk_ack.pingpong || 3199056834d9SIlpo Järvinen (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 32001da177e4SLinus Torvalds max_ato = TCP_DELACK_MAX; 32011da177e4SLinus Torvalds 32021da177e4SLinus Torvalds /* Slow path, intersegment interval is "high". */ 32031da177e4SLinus Torvalds 32041da177e4SLinus Torvalds /* If some rtt estimate is known, use it to bound delayed ack. 3205463c84b9SArnaldo Carvalho de Melo * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 32061da177e4SLinus Torvalds * directly. 32071da177e4SLinus Torvalds */ 3208740b0f18SEric Dumazet if (tp->srtt_us) { 3209740b0f18SEric Dumazet int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3), 3210740b0f18SEric Dumazet TCP_DELACK_MIN); 32111da177e4SLinus Torvalds 32121da177e4SLinus Torvalds if (rtt < max_ato) 32131da177e4SLinus Torvalds max_ato = rtt; 32141da177e4SLinus Torvalds } 32151da177e4SLinus Torvalds 32161da177e4SLinus Torvalds ato = min(ato, max_ato); 32171da177e4SLinus Torvalds } 32181da177e4SLinus Torvalds 32191da177e4SLinus Torvalds /* Stay within the limit we were given */ 32201da177e4SLinus Torvalds timeout = jiffies + ato; 32211da177e4SLinus Torvalds 32221da177e4SLinus Torvalds /* Use new timeout only if there wasn't a older one earlier. */ 3223463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 32241da177e4SLinus Torvalds /* If delack timer was blocked or is about to expire, 32251da177e4SLinus Torvalds * send ACK now. 32261da177e4SLinus Torvalds */ 3227463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.blocked || 3228463c84b9SArnaldo Carvalho de Melo time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 32291da177e4SLinus Torvalds tcp_send_ack(sk); 32301da177e4SLinus Torvalds return; 32311da177e4SLinus Torvalds } 32321da177e4SLinus Torvalds 3233463c84b9SArnaldo Carvalho de Melo if (!time_before(timeout, icsk->icsk_ack.timeout)) 3234463c84b9SArnaldo Carvalho de Melo timeout = icsk->icsk_ack.timeout; 32351da177e4SLinus Torvalds } 3236463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 3237463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.timeout = timeout; 3238463c84b9SArnaldo Carvalho de Melo sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 32391da177e4SLinus Torvalds } 32401da177e4SLinus Torvalds 32411da177e4SLinus Torvalds /* This routine sends an ack and also updates the window. */ 32421da177e4SLinus Torvalds void tcp_send_ack(struct sock *sk) 32431da177e4SLinus Torvalds { 32441da177e4SLinus Torvalds struct sk_buff *buff; 32451da177e4SLinus Torvalds 3246058dc334SIlpo Järvinen /* If we have been reset, we may not send again. */ 3247058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3248058dc334SIlpo Järvinen return; 3249058dc334SIlpo Järvinen 32509890092eSFlorian Westphal tcp_ca_event(sk, CA_EVENT_NON_DELAYED_ACK); 32519890092eSFlorian Westphal 32521da177e4SLinus Torvalds /* We are not putting this on the write queue, so 32531da177e4SLinus Torvalds * tcp_transmit_skb() will set the ownership to this 32541da177e4SLinus Torvalds * sock. 32551da177e4SLinus Torvalds */ 325699a1dec7SMel Gorman buff = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC)); 32571da177e4SLinus Torvalds if (buff == NULL) { 3258463c84b9SArnaldo Carvalho de Melo inet_csk_schedule_ack(sk); 3259463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; 32603f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 32613f421baaSArnaldo Carvalho de Melo TCP_DELACK_MAX, TCP_RTO_MAX); 32621da177e4SLinus Torvalds return; 32631da177e4SLinus Torvalds } 32641da177e4SLinus Torvalds 32651da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 32661da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 3267a3433f35SChangli Gao tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); 32681da177e4SLinus Torvalds 326998781965SEric Dumazet /* We do not want pure acks influencing TCP Small Queues or fq/pacing 327098781965SEric Dumazet * too much. 327198781965SEric Dumazet * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784 327298781965SEric Dumazet * We also avoid tcp_wfree() overhead (cache line miss accessing 327398781965SEric Dumazet * tp->tsq_flags) by using regular sock_wfree() 327498781965SEric Dumazet */ 327598781965SEric Dumazet skb_set_tcp_pure_ack(buff); 327698781965SEric Dumazet 32771da177e4SLinus Torvalds /* Send it off, this clears delayed acks for us. */ 32787faee5c0SEric Dumazet skb_mstamp_get(&buff->skb_mstamp); 327999a1dec7SMel Gorman tcp_transmit_skb(sk, buff, 0, sk_gfp_atomic(sk, GFP_ATOMIC)); 32801da177e4SLinus Torvalds } 3281e3118e83SDaniel Borkmann EXPORT_SYMBOL_GPL(tcp_send_ack); 32821da177e4SLinus Torvalds 32831da177e4SLinus Torvalds /* This routine sends a packet with an out of date sequence 32841da177e4SLinus Torvalds * number. It assumes the other end will try to ack it. 32851da177e4SLinus Torvalds * 32861da177e4SLinus Torvalds * Question: what should we make while urgent mode? 32871da177e4SLinus Torvalds * 4.4BSD forces sending single byte of data. We cannot send 32881da177e4SLinus Torvalds * out of window data, because we have SND.NXT==SND.MAX... 32891da177e4SLinus Torvalds * 32901da177e4SLinus Torvalds * Current solution: to send TWO zero-length segments in urgent mode: 32911da177e4SLinus Torvalds * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 32921da177e4SLinus Torvalds * out-of-date with SND.UNA-1 to probe window. 32931da177e4SLinus Torvalds */ 32941da177e4SLinus Torvalds static int tcp_xmit_probe_skb(struct sock *sk, int urgent) 32951da177e4SLinus Torvalds { 32961da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 32971da177e4SLinus Torvalds struct sk_buff *skb; 32981da177e4SLinus Torvalds 32991da177e4SLinus Torvalds /* We don't queue it, tcp_transmit_skb() sets ownership. */ 330099a1dec7SMel Gorman skb = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC)); 33011da177e4SLinus Torvalds if (skb == NULL) 33021da177e4SLinus Torvalds return -1; 33031da177e4SLinus Torvalds 33041da177e4SLinus Torvalds /* Reserve space for headers and set control bits. */ 33051da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 33061da177e4SLinus Torvalds /* Use a previous sequence. This should cause the other 33071da177e4SLinus Torvalds * end to send an ack. Don't queue or clone SKB, just 33081da177e4SLinus Torvalds * send it. 33091da177e4SLinus Torvalds */ 3310a3433f35SChangli Gao tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); 33117faee5c0SEric Dumazet skb_mstamp_get(&skb->skb_mstamp); 3312dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC); 33131da177e4SLinus Torvalds } 33141da177e4SLinus Torvalds 3315ee995283SPavel Emelyanov void tcp_send_window_probe(struct sock *sk) 3316ee995283SPavel Emelyanov { 3317ee995283SPavel Emelyanov if (sk->sk_state == TCP_ESTABLISHED) { 3318ee995283SPavel Emelyanov tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; 3319ee995283SPavel Emelyanov tcp_xmit_probe_skb(sk, 0); 3320ee995283SPavel Emelyanov } 3321ee995283SPavel Emelyanov } 3322ee995283SPavel Emelyanov 332367edfef7SAndi Kleen /* Initiate keepalive or window probe from timer. */ 33241da177e4SLinus Torvalds int tcp_write_wakeup(struct sock *sk) 33251da177e4SLinus Torvalds { 33261da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 33271da177e4SLinus Torvalds struct sk_buff *skb; 33281da177e4SLinus Torvalds 3329058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3330058dc334SIlpo Järvinen return -1; 3331058dc334SIlpo Järvinen 3332fe067e8aSDavid S. Miller if ((skb = tcp_send_head(sk)) != NULL && 333390840defSIlpo Järvinen before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { 33341da177e4SLinus Torvalds int err; 33350c54b85fSIlpo Järvinen unsigned int mss = tcp_current_mss(sk); 333690840defSIlpo Järvinen unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 33371da177e4SLinus Torvalds 33381da177e4SLinus Torvalds if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 33391da177e4SLinus Torvalds tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 33401da177e4SLinus Torvalds 33411da177e4SLinus Torvalds /* We are probing the opening of a window 33421da177e4SLinus Torvalds * but the window size is != 0 33431da177e4SLinus Torvalds * must have been a result SWS avoidance ( sender ) 33441da177e4SLinus Torvalds */ 33451da177e4SLinus Torvalds if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 33461da177e4SLinus Torvalds skb->len > mss) { 33471da177e4SLinus Torvalds seg_size = min(seg_size, mss); 33484de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 33496cc55e09SOctavian Purdila if (tcp_fragment(sk, skb, seg_size, mss, GFP_ATOMIC)) 33501da177e4SLinus Torvalds return -1; 33511da177e4SLinus Torvalds } else if (!tcp_skb_pcount(skb)) 3352846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss); 33531da177e4SLinus Torvalds 33544de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 3355dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 335666f5fe62SIlpo Järvinen if (!err) 335766f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 33581da177e4SLinus Torvalds return err; 33591da177e4SLinus Torvalds } else { 336033f5f57eSIlpo Järvinen if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) 33614828e7f4SIlpo Järvinen tcp_xmit_probe_skb(sk, 1); 33621da177e4SLinus Torvalds return tcp_xmit_probe_skb(sk, 0); 33631da177e4SLinus Torvalds } 33641da177e4SLinus Torvalds } 33651da177e4SLinus Torvalds 33661da177e4SLinus Torvalds /* A window probe timeout has occurred. If window is not closed send 33671da177e4SLinus Torvalds * a partial packet else a zero probe. 33681da177e4SLinus Torvalds */ 33691da177e4SLinus Torvalds void tcp_send_probe0(struct sock *sk) 33701da177e4SLinus Torvalds { 3371463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 33721da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 3373fcdd1cf4SEric Dumazet unsigned long probe_max; 33741da177e4SLinus Torvalds int err; 33751da177e4SLinus Torvalds 33761da177e4SLinus Torvalds err = tcp_write_wakeup(sk); 33771da177e4SLinus Torvalds 3378fe067e8aSDavid S. Miller if (tp->packets_out || !tcp_send_head(sk)) { 33791da177e4SLinus Torvalds /* Cancel probe timer, if it is not required. */ 33806687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 0; 3381463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff = 0; 33821da177e4SLinus Torvalds return; 33831da177e4SLinus Torvalds } 33841da177e4SLinus Torvalds 33851da177e4SLinus Torvalds if (err <= 0) { 3386463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_backoff < sysctl_tcp_retries2) 3387463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff++; 33886687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out++; 3389fcdd1cf4SEric Dumazet probe_max = TCP_RTO_MAX; 33901da177e4SLinus Torvalds } else { 33911da177e4SLinus Torvalds /* If packet was not sent due to local congestion, 33926687e988SArnaldo Carvalho de Melo * do not backoff and do not remember icsk_probes_out. 33931da177e4SLinus Torvalds * Let local senders to fight for local resources. 33941da177e4SLinus Torvalds * 33951da177e4SLinus Torvalds * Use accumulated backoff yet. 33961da177e4SLinus Torvalds */ 33976687e988SArnaldo Carvalho de Melo if (!icsk->icsk_probes_out) 33986687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 1; 3399fcdd1cf4SEric Dumazet probe_max = TCP_RESOURCE_PROBE_INTERVAL; 34001da177e4SLinus Torvalds } 3401fcdd1cf4SEric Dumazet inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 3402fcdd1cf4SEric Dumazet inet_csk_rto_backoff(icsk, probe_max), 3403fcdd1cf4SEric Dumazet TCP_RTO_MAX); 34041da177e4SLinus Torvalds } 34055db92c99SOctavian Purdila 34065db92c99SOctavian Purdila int tcp_rtx_synack(struct sock *sk, struct request_sock *req) 34075db92c99SOctavian Purdila { 34085db92c99SOctavian Purdila const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific; 34095db92c99SOctavian Purdila struct flowi fl; 34105db92c99SOctavian Purdila int res; 34115db92c99SOctavian Purdila 34125db92c99SOctavian Purdila res = af_ops->send_synack(sk, NULL, &fl, req, 0, NULL); 34135db92c99SOctavian Purdila if (!res) { 34145db92c99SOctavian Purdila TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_RETRANSSEGS); 34155db92c99SOctavian Purdila NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 34165db92c99SOctavian Purdila } 34175db92c99SOctavian Purdila return res; 34185db92c99SOctavian Purdila } 34195db92c99SOctavian Purdila EXPORT_SYMBOL(tcp_rtx_synack); 3420