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; 5217f9b838bSDaniel Lee u8 *p = (u8 *)ptr; 5227f9b838bSDaniel Lee u32 len; /* Fast Open option length */ 5232100c8d2SYuchung Cheng 5247f9b838bSDaniel Lee if (foc->exp) { 5257f9b838bSDaniel Lee len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len; 5267f9b838bSDaniel Lee *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) | 5272100c8d2SYuchung Cheng TCPOPT_FASTOPEN_MAGIC); 5287f9b838bSDaniel Lee p += TCPOLEN_EXP_FASTOPEN_BASE; 5297f9b838bSDaniel Lee } else { 5307f9b838bSDaniel Lee len = TCPOLEN_FASTOPEN_BASE + foc->len; 5317f9b838bSDaniel Lee *p++ = TCPOPT_FASTOPEN; 5327f9b838bSDaniel Lee *p++ = len; 5332100c8d2SYuchung Cheng } 5347f9b838bSDaniel Lee 5357f9b838bSDaniel Lee memcpy(p, foc->val, foc->len); 5367f9b838bSDaniel Lee if ((len & 3) == 2) { 5377f9b838bSDaniel Lee p[foc->len] = TCPOPT_NOP; 5387f9b838bSDaniel Lee p[foc->len + 1] = TCPOPT_NOP; 5397f9b838bSDaniel Lee } 5407f9b838bSDaniel Lee ptr += (len + 3) >> 2; 5412100c8d2SYuchung Cheng } 54240efc6faSStephen Hemminger } 54340efc6faSStephen Hemminger 54467edfef7SAndi Kleen /* Compute TCP options for SYN packets. This is not the final 54567edfef7SAndi Kleen * network wire format yet. 54667edfef7SAndi Kleen */ 54795c96174SEric Dumazet static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb, 54833ad798cSAdam Langley struct tcp_out_options *opts, 549cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 550cf533ea5SEric Dumazet { 55133ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 55295c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 553783237e8SYuchung Cheng struct tcp_fastopen_request *fastopen = tp->fastopen_req; 55433ad798cSAdam Langley 555cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 55633ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 55733ad798cSAdam Langley if (*md5) { 55833ad798cSAdam Langley opts->options |= OPTION_MD5; 559bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 560cfb6eeb4SYOSHIFUJI Hideaki } 56133ad798cSAdam Langley #else 56233ad798cSAdam Langley *md5 = NULL; 563cfb6eeb4SYOSHIFUJI Hideaki #endif 56433ad798cSAdam Langley 56533ad798cSAdam Langley /* We always get an MSS option. The option bytes which will be seen in 56633ad798cSAdam Langley * normal data packets should timestamps be used, must be in the MSS 56733ad798cSAdam Langley * advertised. But we subtract them from tp->mss_cache so that 56833ad798cSAdam Langley * calculations in tcp_sendmsg are simpler etc. So account for this 56933ad798cSAdam Langley * fact here if necessary. If we don't do this correctly, as a 57033ad798cSAdam Langley * receiver we won't recognize data packets as being full sized when we 57133ad798cSAdam Langley * should, and thus we won't abide by the delayed ACK rules correctly. 57233ad798cSAdam Langley * SACKs don't matter, we never delay an ACK when we have any of those 57333ad798cSAdam Langley * going out. */ 57433ad798cSAdam Langley opts->mss = tcp_advertise_mss(sk); 575bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 57633ad798cSAdam Langley 57751456b29SIan Morris if (likely(sysctl_tcp_timestamps && !*md5)) { 57833ad798cSAdam Langley opts->options |= OPTION_TS; 5797faee5c0SEric Dumazet opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset; 58033ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 581bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 58233ad798cSAdam Langley } 583bb5b7c11SDavid S. Miller if (likely(sysctl_tcp_window_scaling)) { 58433ad798cSAdam Langley opts->ws = tp->rx_opt.rcv_wscale; 58589e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 586bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 58733ad798cSAdam Langley } 588bb5b7c11SDavid S. Miller if (likely(sysctl_tcp_sack)) { 58933ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 590b32d1310SDavid S. Miller if (unlikely(!(OPTION_TS & opts->options))) 591bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 59233ad798cSAdam Langley } 59333ad798cSAdam Langley 594783237e8SYuchung Cheng if (fastopen && fastopen->cookie.len >= 0) { 595*2646c831SDaniel Lee u32 need = fastopen->cookie.len; 596*2646c831SDaniel Lee 597*2646c831SDaniel Lee need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE : 598*2646c831SDaniel Lee TCPOLEN_FASTOPEN_BASE; 599783237e8SYuchung Cheng need = (need + 3) & ~3U; /* Align to 32 bits */ 600783237e8SYuchung Cheng if (remaining >= need) { 601783237e8SYuchung Cheng opts->options |= OPTION_FAST_OPEN_COOKIE; 602783237e8SYuchung Cheng opts->fastopen_cookie = &fastopen->cookie; 603783237e8SYuchung Cheng remaining -= need; 604783237e8SYuchung Cheng tp->syn_fastopen = 1; 605*2646c831SDaniel Lee tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0; 606783237e8SYuchung Cheng } 607783237e8SYuchung Cheng } 608bd0388aeSWilliam Allen Simpson 609bd0388aeSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 61033ad798cSAdam Langley } 61133ad798cSAdam Langley 61267edfef7SAndi Kleen /* Set up TCP options for SYN-ACKs. */ 61395c96174SEric Dumazet static unsigned int tcp_synack_options(struct sock *sk, 61433ad798cSAdam Langley struct request_sock *req, 61595c96174SEric Dumazet unsigned int mss, struct sk_buff *skb, 61633ad798cSAdam Langley struct tcp_out_options *opts, 61780f03e27SEric Dumazet const struct tcp_md5sig_key *md5, 6188336886fSJerry Chu struct tcp_fastopen_cookie *foc) 6194957faadSWilliam Allen Simpson { 62033ad798cSAdam Langley struct inet_request_sock *ireq = inet_rsk(req); 62195c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 62233ad798cSAdam Langley 62333ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 62480f03e27SEric Dumazet if (md5) { 62533ad798cSAdam Langley opts->options |= OPTION_MD5; 6264957faadSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 6274957faadSWilliam Allen Simpson 6284957faadSWilliam Allen Simpson /* We can't fit any SACK blocks in a packet with MD5 + TS 6294957faadSWilliam Allen Simpson * options. There was discussion about disabling SACK 6304957faadSWilliam Allen Simpson * rather than TS in order to fit in better with old, 6314957faadSWilliam Allen Simpson * buggy kernels, but that was deemed to be unnecessary. 6324957faadSWilliam Allen Simpson */ 633de213e5eSEric Dumazet ireq->tstamp_ok &= !ireq->sack_ok; 63433ad798cSAdam Langley } 63533ad798cSAdam Langley #endif 63633ad798cSAdam Langley 6374957faadSWilliam Allen Simpson /* We always send an MSS option. */ 63833ad798cSAdam Langley opts->mss = mss; 6394957faadSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 64033ad798cSAdam Langley 64133ad798cSAdam Langley if (likely(ireq->wscale_ok)) { 64233ad798cSAdam Langley opts->ws = ireq->rcv_wscale; 64389e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 6444957faadSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 64533ad798cSAdam Langley } 646de213e5eSEric Dumazet if (likely(ireq->tstamp_ok)) { 64733ad798cSAdam Langley opts->options |= OPTION_TS; 6487faee5c0SEric Dumazet opts->tsval = tcp_skb_timestamp(skb); 64933ad798cSAdam Langley opts->tsecr = req->ts_recent; 6504957faadSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 65133ad798cSAdam Langley } 65233ad798cSAdam Langley if (likely(ireq->sack_ok)) { 65333ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 654de213e5eSEric Dumazet if (unlikely(!ireq->tstamp_ok)) 6554957faadSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 65633ad798cSAdam Langley } 6577f9b838bSDaniel Lee if (foc != NULL && foc->len >= 0) { 6587f9b838bSDaniel Lee u32 need = foc->len; 6597f9b838bSDaniel Lee 6607f9b838bSDaniel Lee need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE : 6617f9b838bSDaniel Lee TCPOLEN_FASTOPEN_BASE; 6628336886fSJerry Chu need = (need + 3) & ~3U; /* Align to 32 bits */ 6638336886fSJerry Chu if (remaining >= need) { 6648336886fSJerry Chu opts->options |= OPTION_FAST_OPEN_COOKIE; 6658336886fSJerry Chu opts->fastopen_cookie = foc; 6668336886fSJerry Chu remaining -= need; 6678336886fSJerry Chu } 6688336886fSJerry Chu } 6694957faadSWilliam Allen Simpson 6704957faadSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 67133ad798cSAdam Langley } 67233ad798cSAdam Langley 67367edfef7SAndi Kleen /* Compute TCP options for ESTABLISHED sockets. This is not the 67467edfef7SAndi Kleen * final wire format yet. 67567edfef7SAndi Kleen */ 67695c96174SEric Dumazet static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb, 67733ad798cSAdam Langley struct tcp_out_options *opts, 678cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 679cf533ea5SEric Dumazet { 68033ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 68195c96174SEric Dumazet unsigned int size = 0; 682cabeccbdSIlpo Järvinen unsigned int eff_sacks; 68333ad798cSAdam Langley 6845843ef42SAndi Kleen opts->options = 0; 6855843ef42SAndi Kleen 68633ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 68733ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 68833ad798cSAdam Langley if (unlikely(*md5)) { 68933ad798cSAdam Langley opts->options |= OPTION_MD5; 69033ad798cSAdam Langley size += TCPOLEN_MD5SIG_ALIGNED; 69133ad798cSAdam Langley } 69233ad798cSAdam Langley #else 69333ad798cSAdam Langley *md5 = NULL; 69433ad798cSAdam Langley #endif 69533ad798cSAdam Langley 69633ad798cSAdam Langley if (likely(tp->rx_opt.tstamp_ok)) { 69733ad798cSAdam Langley opts->options |= OPTION_TS; 6987faee5c0SEric Dumazet opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0; 69933ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 70033ad798cSAdam Langley size += TCPOLEN_TSTAMP_ALIGNED; 70133ad798cSAdam Langley } 70233ad798cSAdam Langley 703cabeccbdSIlpo Järvinen eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; 704cabeccbdSIlpo Järvinen if (unlikely(eff_sacks)) { 70595c96174SEric Dumazet const unsigned int remaining = MAX_TCP_OPTION_SPACE - size; 70633ad798cSAdam Langley opts->num_sack_blocks = 70795c96174SEric Dumazet min_t(unsigned int, eff_sacks, 70833ad798cSAdam Langley (remaining - TCPOLEN_SACK_BASE_ALIGNED) / 70933ad798cSAdam Langley TCPOLEN_SACK_PERBLOCK); 71033ad798cSAdam Langley size += TCPOLEN_SACK_BASE_ALIGNED + 71133ad798cSAdam Langley opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; 71233ad798cSAdam Langley } 71333ad798cSAdam Langley 71433ad798cSAdam Langley return size; 71540efc6faSStephen Hemminger } 7161da177e4SLinus Torvalds 71746d3ceabSEric Dumazet 71846d3ceabSEric Dumazet /* TCP SMALL QUEUES (TSQ) 71946d3ceabSEric Dumazet * 72046d3ceabSEric Dumazet * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev) 72146d3ceabSEric Dumazet * to reduce RTT and bufferbloat. 72246d3ceabSEric Dumazet * We do this using a special skb destructor (tcp_wfree). 72346d3ceabSEric Dumazet * 72446d3ceabSEric Dumazet * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb 72546d3ceabSEric Dumazet * needs to be reallocated in a driver. 7268e3bff96Sstephen hemminger * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc 72746d3ceabSEric Dumazet * 72846d3ceabSEric Dumazet * Since transmit from skb destructor is forbidden, we use a tasklet 72946d3ceabSEric Dumazet * to process all sockets that eventually need to send more skbs. 73046d3ceabSEric Dumazet * We use one tasklet per cpu, with its own queue of sockets. 73146d3ceabSEric Dumazet */ 73246d3ceabSEric Dumazet struct tsq_tasklet { 73346d3ceabSEric Dumazet struct tasklet_struct tasklet; 73446d3ceabSEric Dumazet struct list_head head; /* queue of tcp sockets */ 73546d3ceabSEric Dumazet }; 73646d3ceabSEric Dumazet static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet); 73746d3ceabSEric Dumazet 7386f458dfbSEric Dumazet static void tcp_tsq_handler(struct sock *sk) 7396f458dfbSEric Dumazet { 7406f458dfbSEric Dumazet if ((1 << sk->sk_state) & 7416f458dfbSEric Dumazet (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING | 7426f458dfbSEric Dumazet TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) 743bf06200eSJohn Ogness tcp_write_xmit(sk, tcp_current_mss(sk), tcp_sk(sk)->nonagle, 744bf06200eSJohn Ogness 0, GFP_ATOMIC); 7456f458dfbSEric Dumazet } 74646d3ceabSEric Dumazet /* 7478e3bff96Sstephen hemminger * One tasklet per cpu tries to send more skbs. 74846d3ceabSEric Dumazet * We run in tasklet context but need to disable irqs when 7498e3bff96Sstephen hemminger * transferring tsq->head because tcp_wfree() might 75046d3ceabSEric Dumazet * interrupt us (non NAPI drivers) 75146d3ceabSEric Dumazet */ 75246d3ceabSEric Dumazet static void tcp_tasklet_func(unsigned long data) 75346d3ceabSEric Dumazet { 75446d3ceabSEric Dumazet struct tsq_tasklet *tsq = (struct tsq_tasklet *)data; 75546d3ceabSEric Dumazet LIST_HEAD(list); 75646d3ceabSEric Dumazet unsigned long flags; 75746d3ceabSEric Dumazet struct list_head *q, *n; 75846d3ceabSEric Dumazet struct tcp_sock *tp; 75946d3ceabSEric Dumazet struct sock *sk; 76046d3ceabSEric Dumazet 76146d3ceabSEric Dumazet local_irq_save(flags); 76246d3ceabSEric Dumazet list_splice_init(&tsq->head, &list); 76346d3ceabSEric Dumazet local_irq_restore(flags); 76446d3ceabSEric Dumazet 76546d3ceabSEric Dumazet list_for_each_safe(q, n, &list) { 76646d3ceabSEric Dumazet tp = list_entry(q, struct tcp_sock, tsq_node); 76746d3ceabSEric Dumazet list_del(&tp->tsq_node); 76846d3ceabSEric Dumazet 76946d3ceabSEric Dumazet sk = (struct sock *)tp; 77046d3ceabSEric Dumazet bh_lock_sock(sk); 77146d3ceabSEric Dumazet 77246d3ceabSEric Dumazet if (!sock_owned_by_user(sk)) { 7736f458dfbSEric Dumazet tcp_tsq_handler(sk); 77446d3ceabSEric Dumazet } else { 77546d3ceabSEric Dumazet /* defer the work to tcp_release_cb() */ 7766f458dfbSEric Dumazet set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags); 77746d3ceabSEric Dumazet } 77846d3ceabSEric Dumazet bh_unlock_sock(sk); 77946d3ceabSEric Dumazet 78046d3ceabSEric Dumazet clear_bit(TSQ_QUEUED, &tp->tsq_flags); 78146d3ceabSEric Dumazet sk_free(sk); 78246d3ceabSEric Dumazet } 78346d3ceabSEric Dumazet } 78446d3ceabSEric Dumazet 7856f458dfbSEric Dumazet #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \ 7866f458dfbSEric Dumazet (1UL << TCP_WRITE_TIMER_DEFERRED) | \ 787563d34d0SEric Dumazet (1UL << TCP_DELACK_TIMER_DEFERRED) | \ 788563d34d0SEric Dumazet (1UL << TCP_MTU_REDUCED_DEFERRED)) 78946d3ceabSEric Dumazet /** 79046d3ceabSEric Dumazet * tcp_release_cb - tcp release_sock() callback 79146d3ceabSEric Dumazet * @sk: socket 79246d3ceabSEric Dumazet * 79346d3ceabSEric Dumazet * called from release_sock() to perform protocol dependent 79446d3ceabSEric Dumazet * actions before socket release. 79546d3ceabSEric Dumazet */ 79646d3ceabSEric Dumazet void tcp_release_cb(struct sock *sk) 79746d3ceabSEric Dumazet { 79846d3ceabSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 7996f458dfbSEric Dumazet unsigned long flags, nflags; 80046d3ceabSEric Dumazet 8016f458dfbSEric Dumazet /* perform an atomic operation only if at least one flag is set */ 8026f458dfbSEric Dumazet do { 8036f458dfbSEric Dumazet flags = tp->tsq_flags; 8046f458dfbSEric Dumazet if (!(flags & TCP_DEFERRED_ALL)) 8056f458dfbSEric Dumazet return; 8066f458dfbSEric Dumazet nflags = flags & ~TCP_DEFERRED_ALL; 8076f458dfbSEric Dumazet } while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags); 8086f458dfbSEric Dumazet 8096f458dfbSEric Dumazet if (flags & (1UL << TCP_TSQ_DEFERRED)) 8106f458dfbSEric Dumazet tcp_tsq_handler(sk); 8116f458dfbSEric Dumazet 812c3f9b018SEric Dumazet /* Here begins the tricky part : 813c3f9b018SEric Dumazet * We are called from release_sock() with : 814c3f9b018SEric Dumazet * 1) BH disabled 815c3f9b018SEric Dumazet * 2) sk_lock.slock spinlock held 816c3f9b018SEric Dumazet * 3) socket owned by us (sk->sk_lock.owned == 1) 817c3f9b018SEric Dumazet * 818c3f9b018SEric Dumazet * But following code is meant to be called from BH handlers, 819c3f9b018SEric Dumazet * so we should keep BH disabled, but early release socket ownership 820c3f9b018SEric Dumazet */ 821c3f9b018SEric Dumazet sock_release_ownership(sk); 822c3f9b018SEric Dumazet 823144d56e9SEric Dumazet if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) { 8246f458dfbSEric Dumazet tcp_write_timer_handler(sk); 825144d56e9SEric Dumazet __sock_put(sk); 826144d56e9SEric Dumazet } 827144d56e9SEric Dumazet if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) { 8286f458dfbSEric Dumazet tcp_delack_timer_handler(sk); 829144d56e9SEric Dumazet __sock_put(sk); 830144d56e9SEric Dumazet } 831144d56e9SEric Dumazet if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) { 8324fab9071SNeal Cardwell inet_csk(sk)->icsk_af_ops->mtu_reduced(sk); 833144d56e9SEric Dumazet __sock_put(sk); 834144d56e9SEric Dumazet } 83546d3ceabSEric Dumazet } 83646d3ceabSEric Dumazet EXPORT_SYMBOL(tcp_release_cb); 83746d3ceabSEric Dumazet 83846d3ceabSEric Dumazet void __init tcp_tasklet_init(void) 83946d3ceabSEric Dumazet { 84046d3ceabSEric Dumazet int i; 84146d3ceabSEric Dumazet 84246d3ceabSEric Dumazet for_each_possible_cpu(i) { 84346d3ceabSEric Dumazet struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i); 84446d3ceabSEric Dumazet 84546d3ceabSEric Dumazet INIT_LIST_HEAD(&tsq->head); 84646d3ceabSEric Dumazet tasklet_init(&tsq->tasklet, 84746d3ceabSEric Dumazet tcp_tasklet_func, 84846d3ceabSEric Dumazet (unsigned long)tsq); 84946d3ceabSEric Dumazet } 85046d3ceabSEric Dumazet } 85146d3ceabSEric Dumazet 85246d3ceabSEric Dumazet /* 85346d3ceabSEric Dumazet * Write buffer destructor automatically called from kfree_skb. 8548e3bff96Sstephen hemminger * We can't xmit new skbs from this context, as we might already 85546d3ceabSEric Dumazet * hold qdisc lock. 85646d3ceabSEric Dumazet */ 857d6a4a104SEric Dumazet void tcp_wfree(struct sk_buff *skb) 85846d3ceabSEric Dumazet { 85946d3ceabSEric Dumazet struct sock *sk = skb->sk; 86046d3ceabSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 8619b462d02SEric Dumazet int wmem; 8629b462d02SEric Dumazet 8639b462d02SEric Dumazet /* Keep one reference on sk_wmem_alloc. 8649b462d02SEric Dumazet * Will be released by sk_free() from here or tcp_tasklet_func() 8659b462d02SEric Dumazet */ 8669b462d02SEric Dumazet wmem = atomic_sub_return(skb->truesize - 1, &sk->sk_wmem_alloc); 8679b462d02SEric Dumazet 8689b462d02SEric Dumazet /* If this softirq is serviced by ksoftirqd, we are likely under stress. 8699b462d02SEric Dumazet * Wait until our queues (qdisc + devices) are drained. 8709b462d02SEric Dumazet * This gives : 8719b462d02SEric Dumazet * - less callbacks to tcp_write_xmit(), reducing stress (batches) 8729b462d02SEric Dumazet * - chance for incoming ACK (processed by another cpu maybe) 8739b462d02SEric Dumazet * to migrate this flow (skb->ooo_okay will be eventually set) 8749b462d02SEric Dumazet */ 8759b462d02SEric Dumazet if (wmem >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current) 8769b462d02SEric Dumazet goto out; 87746d3ceabSEric Dumazet 87846d3ceabSEric Dumazet if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) && 87946d3ceabSEric Dumazet !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) { 88046d3ceabSEric Dumazet unsigned long flags; 88146d3ceabSEric Dumazet struct tsq_tasklet *tsq; 88246d3ceabSEric Dumazet 88346d3ceabSEric Dumazet /* queue this socket to tasklet queue */ 88446d3ceabSEric Dumazet local_irq_save(flags); 885903ceff7SChristoph Lameter tsq = this_cpu_ptr(&tsq_tasklet); 88646d3ceabSEric Dumazet list_add(&tp->tsq_node, &tsq->head); 88746d3ceabSEric Dumazet tasklet_schedule(&tsq->tasklet); 88846d3ceabSEric Dumazet local_irq_restore(flags); 8899b462d02SEric Dumazet return; 89046d3ceabSEric Dumazet } 8919b462d02SEric Dumazet out: 8929b462d02SEric Dumazet sk_free(sk); 89346d3ceabSEric Dumazet } 89446d3ceabSEric Dumazet 8951da177e4SLinus Torvalds /* This routine actually transmits TCP packets queued in by 8961da177e4SLinus Torvalds * tcp_do_sendmsg(). This is used by both the initial 8971da177e4SLinus Torvalds * transmission and possible later retransmissions. 8981da177e4SLinus Torvalds * All SKB's seen here are completely headerless. It is our 8991da177e4SLinus Torvalds * job to build the TCP header, and pass the packet down to 9001da177e4SLinus Torvalds * IP so it can do the same plus pass the packet off to the 9011da177e4SLinus Torvalds * device. 9021da177e4SLinus Torvalds * 9031da177e4SLinus Torvalds * We are working here with either a clone of the original 9041da177e4SLinus Torvalds * SKB, or a fresh unique copy made by the retransmit engine. 9051da177e4SLinus Torvalds */ 906056834d9SIlpo Järvinen static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, 907056834d9SIlpo Järvinen gfp_t gfp_mask) 9081da177e4SLinus Torvalds { 9096687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 910dfb4b9dcSDavid S. Miller struct inet_sock *inet; 911dfb4b9dcSDavid S. Miller struct tcp_sock *tp; 912dfb4b9dcSDavid S. Miller struct tcp_skb_cb *tcb; 91333ad798cSAdam Langley struct tcp_out_options opts; 91495c96174SEric Dumazet unsigned int tcp_options_size, tcp_header_size; 915cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 9161da177e4SLinus Torvalds struct tcphdr *th; 9171da177e4SLinus Torvalds int err; 9181da177e4SLinus Torvalds 919dfb4b9dcSDavid S. Miller BUG_ON(!skb || !tcp_skb_pcount(skb)); 920dfb4b9dcSDavid S. Miller 921ccdbb6e9SEric Dumazet if (clone_it) { 922740b0f18SEric Dumazet skb_mstamp_get(&skb->skb_mstamp); 923dfb4b9dcSDavid S. Miller 924dfb4b9dcSDavid S. Miller if (unlikely(skb_cloned(skb))) 925dfb4b9dcSDavid S. Miller skb = pskb_copy(skb, gfp_mask); 926dfb4b9dcSDavid S. Miller else 927dfb4b9dcSDavid S. Miller skb = skb_clone(skb, gfp_mask); 928dfb4b9dcSDavid S. Miller if (unlikely(!skb)) 929dfb4b9dcSDavid S. Miller return -ENOBUFS; 930dfb4b9dcSDavid S. Miller } 931dfb4b9dcSDavid S. Miller 932dfb4b9dcSDavid S. Miller inet = inet_sk(sk); 933dfb4b9dcSDavid S. Miller tp = tcp_sk(sk); 934dfb4b9dcSDavid S. Miller tcb = TCP_SKB_CB(skb); 93533ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 9361da177e4SLinus Torvalds 9374de075e0SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) 93833ad798cSAdam Langley tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); 93933ad798cSAdam Langley else 94033ad798cSAdam Langley tcp_options_size = tcp_established_options(sk, skb, &opts, 94133ad798cSAdam Langley &md5); 94233ad798cSAdam Langley tcp_header_size = tcp_options_size + sizeof(struct tcphdr); 9431da177e4SLinus Torvalds 944547669d4SEric Dumazet if (tcp_packets_in_flight(tp) == 0) 9456687e988SArnaldo Carvalho de Melo tcp_ca_event(sk, CA_EVENT_TX_START); 946547669d4SEric Dumazet 947547669d4SEric Dumazet /* if no packet is in qdisc/device queue, then allow XPS to select 948b2532eb9SEric Dumazet * another queue. We can be called from tcp_tsq_handler() 949b2532eb9SEric Dumazet * which holds one reference to sk_wmem_alloc. 950b2532eb9SEric Dumazet * 951b2532eb9SEric Dumazet * TODO: Ideally, in-flight pure ACK packets should not matter here. 952b2532eb9SEric Dumazet * One way to get this would be to set skb->truesize = 2 on them. 953547669d4SEric Dumazet */ 954b2532eb9SEric Dumazet skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1); 9551da177e4SLinus Torvalds 956aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 957aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 95846d3ceabSEric Dumazet 95946d3ceabSEric Dumazet skb_orphan(skb); 96046d3ceabSEric Dumazet skb->sk = sk; 96198781965SEric Dumazet skb->destructor = skb_is_tcp_pure_ack(skb) ? sock_wfree : tcp_wfree; 962b73c3d0eSTom Herbert skb_set_hash_from_sk(skb, sk); 96346d3ceabSEric Dumazet atomic_add(skb->truesize, &sk->sk_wmem_alloc); 9641da177e4SLinus Torvalds 9651da177e4SLinus Torvalds /* Build TCP header and checksum it. */ 966aa8223c7SArnaldo Carvalho de Melo th = tcp_hdr(skb); 967c720c7e8SEric Dumazet th->source = inet->inet_sport; 968c720c7e8SEric Dumazet th->dest = inet->inet_dport; 9691da177e4SLinus Torvalds th->seq = htonl(tcb->seq); 9701da177e4SLinus Torvalds th->ack_seq = htonl(tp->rcv_nxt); 971df7a3b07SAl Viro *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | 9724de075e0SEric Dumazet tcb->tcp_flags); 973dfb4b9dcSDavid S. Miller 9744de075e0SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) { 9751da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments 9761da177e4SLinus Torvalds * is never scaled. 9771da177e4SLinus Torvalds */ 978600ff0c2SIlpo Järvinen th->window = htons(min(tp->rcv_wnd, 65535U)); 9791da177e4SLinus Torvalds } else { 9801da177e4SLinus Torvalds th->window = htons(tcp_select_window(sk)); 9811da177e4SLinus Torvalds } 9821da177e4SLinus Torvalds th->check = 0; 9831da177e4SLinus Torvalds th->urg_ptr = 0; 9841da177e4SLinus Torvalds 98533f5f57eSIlpo Järvinen /* The urg_mode check is necessary during a below snd_una win probe */ 9867691367dSHerbert Xu if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { 9877691367dSHerbert Xu if (before(tp->snd_up, tcb->seq + 0x10000)) { 9881da177e4SLinus Torvalds th->urg_ptr = htons(tp->snd_up - tcb->seq); 9891da177e4SLinus Torvalds th->urg = 1; 9907691367dSHerbert Xu } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { 9910eae88f3SEric Dumazet th->urg_ptr = htons(0xFFFF); 9927691367dSHerbert Xu th->urg = 1; 9937691367dSHerbert Xu } 9941da177e4SLinus Torvalds } 9951da177e4SLinus Torvalds 996bd0388aeSWilliam Allen Simpson tcp_options_write((__be32 *)(th + 1), tp, &opts); 9974de075e0SEric Dumazet if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0)) 998735d3831SFlorian Westphal tcp_ecn_send(sk, skb, tcp_header_size); 999dfb4b9dcSDavid S. Miller 1000cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 1001cfb6eeb4SYOSHIFUJI Hideaki /* Calculate the MD5 hash, as we have all we need now */ 1002cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 1003a465419bSEric Dumazet sk_nocaps_add(sk, NETIF_F_GSO_MASK); 1004bd0388aeSWilliam Allen Simpson tp->af_specific->calc_md5_hash(opts.hash_location, 100539f8e58eSEric Dumazet md5, sk, skb); 1006cfb6eeb4SYOSHIFUJI Hideaki } 1007cfb6eeb4SYOSHIFUJI Hideaki #endif 1008cfb6eeb4SYOSHIFUJI Hideaki 1009bb296246SHerbert Xu icsk->icsk_af_ops->send_check(sk, skb); 10101da177e4SLinus Torvalds 10114de075e0SEric Dumazet if (likely(tcb->tcp_flags & TCPHDR_ACK)) 1012fc6415bcSDavid S. Miller tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); 10131da177e4SLinus Torvalds 10141da177e4SLinus Torvalds if (skb->len != tcp_header_size) 1015cf533ea5SEric Dumazet tcp_event_data_sent(tp, sk); 10161da177e4SLinus Torvalds 1017bd37a088SWei Yongjun if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) 1018aa2ea058STom Herbert TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, 1019aa2ea058STom Herbert tcp_skb_pcount(skb)); 10201da177e4SLinus Torvalds 1021cd7d8498SEric Dumazet /* OK, its time to fill skb_shinfo(skb)->gso_segs */ 1022cd7d8498SEric Dumazet skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb); 1023cd7d8498SEric Dumazet 10247faee5c0SEric Dumazet /* Our usage of tstamp should remain private */ 10257faee5c0SEric Dumazet skb->tstamp.tv64 = 0; 1026971f10ecSEric Dumazet 1027971f10ecSEric Dumazet /* Cleanup our debris for IP stacks */ 1028971f10ecSEric Dumazet memset(skb->cb, 0, max(sizeof(struct inet_skb_parm), 1029971f10ecSEric Dumazet sizeof(struct inet6_skb_parm))); 1030971f10ecSEric Dumazet 1031b0270e91SEric Dumazet err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl); 10327faee5c0SEric Dumazet 103383de47cdSHua Zhong if (likely(err <= 0)) 10341da177e4SLinus Torvalds return err; 10351da177e4SLinus Torvalds 10365ee2c941SChristoph Paasch tcp_enter_cwr(sk); 10371da177e4SLinus Torvalds 1038b9df3cb8SGerrit Renker return net_xmit_eval(err); 10391da177e4SLinus Torvalds } 10401da177e4SLinus Torvalds 104167edfef7SAndi Kleen /* This routine just queues the buffer for sending. 10421da177e4SLinus Torvalds * 10431da177e4SLinus Torvalds * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, 10441da177e4SLinus Torvalds * otherwise socket can stall. 10451da177e4SLinus Torvalds */ 10461da177e4SLinus Torvalds static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) 10471da177e4SLinus Torvalds { 10481da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 10491da177e4SLinus Torvalds 10501da177e4SLinus Torvalds /* Advance write_seq and place onto the write_queue. */ 10511da177e4SLinus Torvalds tp->write_seq = TCP_SKB_CB(skb)->end_seq; 1052f4a775d1SEric Dumazet __skb_header_release(skb); 1053fe067e8aSDavid S. Miller tcp_add_write_queue_tail(sk, skb); 10543ab224beSHideo Aoki sk->sk_wmem_queued += skb->truesize; 10553ab224beSHideo Aoki sk_mem_charge(sk, skb->truesize); 10561da177e4SLinus Torvalds } 10571da177e4SLinus Torvalds 105867edfef7SAndi Kleen /* Initialize TSO segments for a packet. */ 1059cf533ea5SEric Dumazet static void tcp_set_skb_tso_segs(const struct sock *sk, struct sk_buff *skb, 1060056834d9SIlpo Järvinen unsigned int mss_now) 1061f6302d1dSDavid S. Miller { 10627b7fc97aSEric Dumazet struct skb_shared_info *shinfo = skb_shinfo(skb); 10637b7fc97aSEric Dumazet 1064c52e2421SEric Dumazet /* Make sure we own this skb before messing gso_size/gso_segs */ 1065c52e2421SEric Dumazet WARN_ON_ONCE(skb_cloned(skb)); 1066c52e2421SEric Dumazet 10678f26fb1cSEric Dumazet if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) { 1068f6302d1dSDavid S. Miller /* Avoid the costly divide in the normal 1069f6302d1dSDavid S. Miller * non-TSO case. 1070f6302d1dSDavid S. Miller */ 1071cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, 1); 10727b7fc97aSEric Dumazet shinfo->gso_size = 0; 10737b7fc97aSEric Dumazet shinfo->gso_type = 0; 1074f6302d1dSDavid S. Miller } else { 1075cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now)); 10767b7fc97aSEric Dumazet shinfo->gso_size = mss_now; 10777b7fc97aSEric Dumazet shinfo->gso_type = sk->sk_gso_type; 10781da177e4SLinus Torvalds } 10791da177e4SLinus Torvalds } 10801da177e4SLinus Torvalds 108191fed7a1SIlpo Järvinen /* When a modification to fackets out becomes necessary, we need to check 108268f8353bSIlpo Järvinen * skb is counted to fackets_out or not. 108391fed7a1SIlpo Järvinen */ 1084cf533ea5SEric Dumazet static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb, 108591fed7a1SIlpo Järvinen int decr) 108691fed7a1SIlpo Järvinen { 1087a47e5a98SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1088a47e5a98SIlpo Järvinen 1089dc86967bSIlpo Järvinen if (!tp->sacked_out || tcp_is_reno(tp)) 109091fed7a1SIlpo Järvinen return; 109191fed7a1SIlpo Järvinen 10926859d494SIlpo Järvinen if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq)) 109391fed7a1SIlpo Järvinen tp->fackets_out -= decr; 109491fed7a1SIlpo Järvinen } 109591fed7a1SIlpo Järvinen 1096797108d1SIlpo Järvinen /* Pcount in the middle of the write queue got changed, we need to do various 1097797108d1SIlpo Järvinen * tweaks to fix counters 1098797108d1SIlpo Järvinen */ 1099cf533ea5SEric Dumazet static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) 1100797108d1SIlpo Järvinen { 1101797108d1SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1102797108d1SIlpo Järvinen 1103797108d1SIlpo Järvinen tp->packets_out -= decr; 1104797108d1SIlpo Järvinen 1105797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 1106797108d1SIlpo Järvinen tp->sacked_out -= decr; 1107797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 1108797108d1SIlpo Järvinen tp->retrans_out -= decr; 1109797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 1110797108d1SIlpo Järvinen tp->lost_out -= decr; 1111797108d1SIlpo Järvinen 1112797108d1SIlpo Järvinen /* Reno case is special. Sigh... */ 1113797108d1SIlpo Järvinen if (tcp_is_reno(tp) && decr > 0) 1114797108d1SIlpo Järvinen tp->sacked_out -= min_t(u32, tp->sacked_out, decr); 1115797108d1SIlpo Järvinen 1116797108d1SIlpo Järvinen tcp_adjust_fackets_out(sk, skb, decr); 1117797108d1SIlpo Järvinen 1118797108d1SIlpo Järvinen if (tp->lost_skb_hint && 1119797108d1SIlpo Järvinen before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && 112052cf3cc8SIlpo Järvinen (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))) 1121797108d1SIlpo Järvinen tp->lost_cnt_hint -= decr; 1122797108d1SIlpo Järvinen 1123797108d1SIlpo Järvinen tcp_verify_left_out(tp); 1124797108d1SIlpo Järvinen } 1125797108d1SIlpo Järvinen 1126490cc7d0SWillem de Bruijn static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2) 1127490cc7d0SWillem de Bruijn { 1128490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo = skb_shinfo(skb); 1129490cc7d0SWillem de Bruijn 1130490cc7d0SWillem de Bruijn if (unlikely(shinfo->tx_flags & SKBTX_ANY_TSTAMP) && 1131490cc7d0SWillem de Bruijn !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) { 1132490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo2 = skb_shinfo(skb2); 1133490cc7d0SWillem de Bruijn u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP; 1134490cc7d0SWillem de Bruijn 1135490cc7d0SWillem de Bruijn shinfo->tx_flags &= ~tsflags; 1136490cc7d0SWillem de Bruijn shinfo2->tx_flags |= tsflags; 1137490cc7d0SWillem de Bruijn swap(shinfo->tskey, shinfo2->tskey); 1138490cc7d0SWillem de Bruijn } 1139490cc7d0SWillem de Bruijn } 1140490cc7d0SWillem de Bruijn 11411da177e4SLinus Torvalds /* Function to create two new TCP segments. Shrinks the given segment 11421da177e4SLinus Torvalds * to the specified size and appends a new segment with the rest of the 11431da177e4SLinus Torvalds * packet to the list. This won't be called frequently, I hope. 11441da177e4SLinus Torvalds * Remember, these are still headerless SKBs at this point. 11451da177e4SLinus Torvalds */ 1146056834d9SIlpo Järvinen int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, 11476cc55e09SOctavian Purdila unsigned int mss_now, gfp_t gfp) 11481da177e4SLinus Torvalds { 11491da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 11501da177e4SLinus Torvalds struct sk_buff *buff; 11516475be16SDavid S. Miller int nsize, old_factor; 1152b60b49eaSHerbert Xu int nlen; 11539ce01461SIlpo Järvinen u8 flags; 11541da177e4SLinus Torvalds 11552fceec13SIlpo Järvinen if (WARN_ON(len > skb->len)) 11562fceec13SIlpo Järvinen return -EINVAL; 11576a438bbeSStephen Hemminger 11581da177e4SLinus Torvalds nsize = skb_headlen(skb) - len; 11591da177e4SLinus Torvalds if (nsize < 0) 11601da177e4SLinus Torvalds nsize = 0; 11611da177e4SLinus Torvalds 11626cc55e09SOctavian Purdila if (skb_unclone(skb, gfp)) 11631da177e4SLinus Torvalds return -ENOMEM; 11641da177e4SLinus Torvalds 11651da177e4SLinus Torvalds /* Get a new skb... force flag on. */ 11666cc55e09SOctavian Purdila buff = sk_stream_alloc_skb(sk, nsize, gfp); 116751456b29SIan Morris if (!buff) 11681da177e4SLinus Torvalds return -ENOMEM; /* We'll just try again later. */ 1169ef5cb973SHerbert Xu 11703ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 11713ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1172b60b49eaSHerbert Xu nlen = skb->len - len - nsize; 1173b60b49eaSHerbert Xu buff->truesize += nlen; 1174b60b49eaSHerbert Xu skb->truesize -= nlen; 11751da177e4SLinus Torvalds 11761da177e4SLinus Torvalds /* Correct the sequence numbers. */ 11771da177e4SLinus Torvalds TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 11781da177e4SLinus Torvalds TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 11791da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 11801da177e4SLinus Torvalds 11811da177e4SLinus Torvalds /* PSH and FIN should only be set in the second packet. */ 11824de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 11834de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 11844de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1185e14c3cafSHerbert Xu TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; 11861da177e4SLinus Torvalds 118784fa7933SPatrick McHardy if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { 11881da177e4SLinus Torvalds /* Copy and checksum data tail into the new buffer. */ 1189056834d9SIlpo Järvinen buff->csum = csum_partial_copy_nocheck(skb->data + len, 1190056834d9SIlpo Järvinen skb_put(buff, nsize), 11911da177e4SLinus Torvalds nsize, 0); 11921da177e4SLinus Torvalds 11931da177e4SLinus Torvalds skb_trim(skb, len); 11941da177e4SLinus Torvalds 11951da177e4SLinus Torvalds skb->csum = csum_block_sub(skb->csum, buff->csum, len); 11961da177e4SLinus Torvalds } else { 119784fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 11981da177e4SLinus Torvalds skb_split(skb, buff, len); 11991da177e4SLinus Torvalds } 12001da177e4SLinus Torvalds 12011da177e4SLinus Torvalds buff->ip_summed = skb->ip_summed; 12021da177e4SLinus Torvalds 1203a61bbcf2SPatrick McHardy buff->tstamp = skb->tstamp; 1204490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 12051da177e4SLinus Torvalds 12066475be16SDavid S. Miller old_factor = tcp_skb_pcount(skb); 12076475be16SDavid S. Miller 12081da177e4SLinus Torvalds /* Fix up tso_factor for both original and new SKB. */ 1209846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1210846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, buff, mss_now); 12111da177e4SLinus Torvalds 12126475be16SDavid S. Miller /* If this packet has been sent out already, we must 12136475be16SDavid S. Miller * adjust the various packet counters. 12146475be16SDavid S. Miller */ 1215cf0b450cSHerbert Xu if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 12166475be16SDavid S. Miller int diff = old_factor - tcp_skb_pcount(skb) - 12176475be16SDavid S. Miller tcp_skb_pcount(buff); 12181da177e4SLinus Torvalds 1219797108d1SIlpo Järvinen if (diff) 1220797108d1SIlpo Järvinen tcp_adjust_pcount(sk, skb, diff); 12211da177e4SLinus Torvalds } 12221da177e4SLinus Torvalds 12231da177e4SLinus Torvalds /* Link BUFF into the send queue. */ 1224f4a775d1SEric Dumazet __skb_header_release(buff); 1225fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 12261da177e4SLinus Torvalds 12271da177e4SLinus Torvalds return 0; 12281da177e4SLinus Torvalds } 12291da177e4SLinus Torvalds 12301da177e4SLinus Torvalds /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c 12311da177e4SLinus Torvalds * eventually). The difference is that pulled data not copied, but 12321da177e4SLinus Torvalds * immediately discarded. 12331da177e4SLinus Torvalds */ 1234f2911969SHerbert Xu ~{PmVHI~} static void __pskb_trim_head(struct sk_buff *skb, int len) 12351da177e4SLinus Torvalds { 12367b7fc97aSEric Dumazet struct skb_shared_info *shinfo; 12371da177e4SLinus Torvalds int i, k, eat; 12381da177e4SLinus Torvalds 12394fa48bf3SEric Dumazet eat = min_t(int, len, skb_headlen(skb)); 12404fa48bf3SEric Dumazet if (eat) { 12414fa48bf3SEric Dumazet __skb_pull(skb, eat); 12424fa48bf3SEric Dumazet len -= eat; 12434fa48bf3SEric Dumazet if (!len) 12444fa48bf3SEric Dumazet return; 12454fa48bf3SEric Dumazet } 12461da177e4SLinus Torvalds eat = len; 12471da177e4SLinus Torvalds k = 0; 12487b7fc97aSEric Dumazet shinfo = skb_shinfo(skb); 12497b7fc97aSEric Dumazet for (i = 0; i < shinfo->nr_frags; i++) { 12507b7fc97aSEric Dumazet int size = skb_frag_size(&shinfo->frags[i]); 12519e903e08SEric Dumazet 12529e903e08SEric Dumazet if (size <= eat) { 1253aff65da0SIan Campbell skb_frag_unref(skb, i); 12549e903e08SEric Dumazet eat -= size; 12551da177e4SLinus Torvalds } else { 12567b7fc97aSEric Dumazet shinfo->frags[k] = shinfo->frags[i]; 12571da177e4SLinus Torvalds if (eat) { 12587b7fc97aSEric Dumazet shinfo->frags[k].page_offset += eat; 12597b7fc97aSEric Dumazet skb_frag_size_sub(&shinfo->frags[k], eat); 12601da177e4SLinus Torvalds eat = 0; 12611da177e4SLinus Torvalds } 12621da177e4SLinus Torvalds k++; 12631da177e4SLinus Torvalds } 12641da177e4SLinus Torvalds } 12657b7fc97aSEric Dumazet shinfo->nr_frags = k; 12661da177e4SLinus Torvalds 126727a884dcSArnaldo Carvalho de Melo skb_reset_tail_pointer(skb); 12681da177e4SLinus Torvalds skb->data_len -= len; 12691da177e4SLinus Torvalds skb->len = skb->data_len; 12701da177e4SLinus Torvalds } 12711da177e4SLinus Torvalds 127267edfef7SAndi Kleen /* Remove acked data from a packet in the transmit queue. */ 12731da177e4SLinus Torvalds int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 12741da177e4SLinus Torvalds { 127514bbd6a5SPravin B Shelar if (skb_unclone(skb, GFP_ATOMIC)) 12761da177e4SLinus Torvalds return -ENOMEM; 12771da177e4SLinus Torvalds 12784fa48bf3SEric Dumazet __pskb_trim_head(skb, len); 12791da177e4SLinus Torvalds 12801da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq += len; 128184fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 12821da177e4SLinus Torvalds 12831da177e4SLinus Torvalds skb->truesize -= len; 12841da177e4SLinus Torvalds sk->sk_wmem_queued -= len; 12853ab224beSHideo Aoki sk_mem_uncharge(sk, len); 12861da177e4SLinus Torvalds sock_set_flag(sk, SOCK_QUEUE_SHRUNK); 12871da177e4SLinus Torvalds 12885b35e1e6SNeal Cardwell /* Any change of skb->len requires recalculation of tso factor. */ 12891da177e4SLinus Torvalds if (tcp_skb_pcount(skb) > 1) 12905b35e1e6SNeal Cardwell tcp_set_skb_tso_segs(sk, skb, tcp_skb_mss(skb)); 12911da177e4SLinus Torvalds 12921da177e4SLinus Torvalds return 0; 12931da177e4SLinus Torvalds } 12941da177e4SLinus Torvalds 12951b63edd6SYuchung Cheng /* Calculate MSS not accounting any TCP options. */ 12961b63edd6SYuchung Cheng static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu) 12975d424d5aSJohn Heffner { 1298cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1299cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 13005d424d5aSJohn Heffner int mss_now; 13015d424d5aSJohn Heffner 13025d424d5aSJohn Heffner /* Calculate base mss without TCP options: 13035d424d5aSJohn Heffner It is MMS_S - sizeof(tcphdr) of rfc1122 13045d424d5aSJohn Heffner */ 13055d424d5aSJohn Heffner mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 13065d424d5aSJohn Heffner 130767469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 130867469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 130967469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 131067469601SEric Dumazet 131167469601SEric Dumazet if (dst && dst_allfrag(dst)) 131267469601SEric Dumazet mss_now -= icsk->icsk_af_ops->net_frag_header_len; 131367469601SEric Dumazet } 131467469601SEric Dumazet 13155d424d5aSJohn Heffner /* Clamp it (mss_clamp does not include tcp options) */ 13165d424d5aSJohn Heffner if (mss_now > tp->rx_opt.mss_clamp) 13175d424d5aSJohn Heffner mss_now = tp->rx_opt.mss_clamp; 13185d424d5aSJohn Heffner 13195d424d5aSJohn Heffner /* Now subtract optional transport overhead */ 13205d424d5aSJohn Heffner mss_now -= icsk->icsk_ext_hdr_len; 13215d424d5aSJohn Heffner 13225d424d5aSJohn Heffner /* Then reserve room for full set of TCP options and 8 bytes of data */ 13235d424d5aSJohn Heffner if (mss_now < 48) 13245d424d5aSJohn Heffner mss_now = 48; 13255d424d5aSJohn Heffner return mss_now; 13265d424d5aSJohn Heffner } 13275d424d5aSJohn Heffner 13281b63edd6SYuchung Cheng /* Calculate MSS. Not accounting for SACKs here. */ 13291b63edd6SYuchung Cheng int tcp_mtu_to_mss(struct sock *sk, int pmtu) 13301b63edd6SYuchung Cheng { 13311b63edd6SYuchung Cheng /* Subtract TCP options size, not including SACKs */ 13321b63edd6SYuchung Cheng return __tcp_mtu_to_mss(sk, pmtu) - 13331b63edd6SYuchung Cheng (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr)); 13341b63edd6SYuchung Cheng } 13351b63edd6SYuchung Cheng 13365d424d5aSJohn Heffner /* Inverse of above */ 133767469601SEric Dumazet int tcp_mss_to_mtu(struct sock *sk, int mss) 13385d424d5aSJohn Heffner { 1339cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1340cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 13415d424d5aSJohn Heffner int mtu; 13425d424d5aSJohn Heffner 13435d424d5aSJohn Heffner mtu = mss + 13445d424d5aSJohn Heffner tp->tcp_header_len + 13455d424d5aSJohn Heffner icsk->icsk_ext_hdr_len + 13465d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 13475d424d5aSJohn Heffner 134867469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 134967469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 135067469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 135167469601SEric Dumazet 135267469601SEric Dumazet if (dst && dst_allfrag(dst)) 135367469601SEric Dumazet mtu += icsk->icsk_af_ops->net_frag_header_len; 135467469601SEric Dumazet } 13555d424d5aSJohn Heffner return mtu; 13565d424d5aSJohn Heffner } 13575d424d5aSJohn Heffner 135867edfef7SAndi Kleen /* MTU probing init per socket */ 13595d424d5aSJohn Heffner void tcp_mtup_init(struct sock *sk) 13605d424d5aSJohn Heffner { 13615d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 13625d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 1363b0f9ca53SFan Du struct net *net = sock_net(sk); 13645d424d5aSJohn Heffner 1365b0f9ca53SFan Du icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1; 13665d424d5aSJohn Heffner icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 13675d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 1368b0f9ca53SFan Du icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss); 13695d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 137005cbc0dbSFan Du if (icsk->icsk_mtup.enabled) 137105cbc0dbSFan Du icsk->icsk_mtup.probe_timestamp = tcp_time_stamp; 13725d424d5aSJohn Heffner } 13734bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_mtup_init); 13745d424d5aSJohn Heffner 13751da177e4SLinus Torvalds /* This function synchronize snd mss to current pmtu/exthdr set. 13761da177e4SLinus Torvalds 13771da177e4SLinus Torvalds tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 13781da177e4SLinus Torvalds for TCP options, but includes only bare TCP header. 13791da177e4SLinus Torvalds 13801da177e4SLinus Torvalds tp->rx_opt.mss_clamp is mss negotiated at connection setup. 1381caa20d9aSStephen Hemminger It is minimum of user_mss and mss received with SYN. 13821da177e4SLinus Torvalds It also does not include TCP options. 13831da177e4SLinus Torvalds 1384d83d8461SArnaldo Carvalho de Melo inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. 13851da177e4SLinus Torvalds 13861da177e4SLinus Torvalds tp->mss_cache is current effective sending mss, including 13871da177e4SLinus Torvalds all tcp options except for SACKs. It is evaluated, 13881da177e4SLinus Torvalds taking into account current pmtu, but never exceeds 13891da177e4SLinus Torvalds tp->rx_opt.mss_clamp. 13901da177e4SLinus Torvalds 13911da177e4SLinus Torvalds NOTE1. rfc1122 clearly states that advertised MSS 13921da177e4SLinus Torvalds DOES NOT include either tcp or ip options. 13931da177e4SLinus Torvalds 1394d83d8461SArnaldo Carvalho de Melo NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache 1395d83d8461SArnaldo Carvalho de Melo are READ ONLY outside this function. --ANK (980731) 13961da177e4SLinus Torvalds */ 13971da177e4SLinus Torvalds unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 13981da177e4SLinus Torvalds { 13991da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1400d83d8461SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 14015d424d5aSJohn Heffner int mss_now; 14021da177e4SLinus Torvalds 14035d424d5aSJohn Heffner if (icsk->icsk_mtup.search_high > pmtu) 14045d424d5aSJohn Heffner icsk->icsk_mtup.search_high = pmtu; 14051da177e4SLinus Torvalds 14065d424d5aSJohn Heffner mss_now = tcp_mtu_to_mss(sk, pmtu); 1407409d22b4SIlpo Järvinen mss_now = tcp_bound_to_half_wnd(tp, mss_now); 14081da177e4SLinus Torvalds 14091da177e4SLinus Torvalds /* And store cached results */ 1410d83d8461SArnaldo Carvalho de Melo icsk->icsk_pmtu_cookie = pmtu; 14115d424d5aSJohn Heffner if (icsk->icsk_mtup.enabled) 14125d424d5aSJohn Heffner mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); 1413c1b4a7e6SDavid S. Miller tp->mss_cache = mss_now; 14141da177e4SLinus Torvalds 14151da177e4SLinus Torvalds return mss_now; 14161da177e4SLinus Torvalds } 14174bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_sync_mss); 14181da177e4SLinus Torvalds 14191da177e4SLinus Torvalds /* Compute the current effective MSS, taking SACKs and IP options, 14201da177e4SLinus Torvalds * and even PMTU discovery events into account. 14211da177e4SLinus Torvalds */ 14220c54b85fSIlpo Järvinen unsigned int tcp_current_mss(struct sock *sk) 14231da177e4SLinus Torvalds { 1424cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1425cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1426c1b4a7e6SDavid S. Miller u32 mss_now; 142795c96174SEric Dumazet unsigned int header_len; 142833ad798cSAdam Langley struct tcp_out_options opts; 142933ad798cSAdam Langley struct tcp_md5sig_key *md5; 14301da177e4SLinus Torvalds 1431c1b4a7e6SDavid S. Miller mss_now = tp->mss_cache; 1432c1b4a7e6SDavid S. Miller 14331da177e4SLinus Torvalds if (dst) { 14341da177e4SLinus Torvalds u32 mtu = dst_mtu(dst); 1435d83d8461SArnaldo Carvalho de Melo if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 14361da177e4SLinus Torvalds mss_now = tcp_sync_mss(sk, mtu); 14371da177e4SLinus Torvalds } 14381da177e4SLinus Torvalds 143933ad798cSAdam Langley header_len = tcp_established_options(sk, NULL, &opts, &md5) + 144033ad798cSAdam Langley sizeof(struct tcphdr); 144133ad798cSAdam Langley /* The mss_cache is sized based on tp->tcp_header_len, which assumes 144233ad798cSAdam Langley * some common options. If this is an odd packet (because we have SACK 144333ad798cSAdam Langley * blocks etc) then our calculated header_len will be different, and 144433ad798cSAdam Langley * we have to adjust mss_now correspondingly */ 144533ad798cSAdam Langley if (header_len != tp->tcp_header_len) { 144633ad798cSAdam Langley int delta = (int) header_len - tp->tcp_header_len; 144733ad798cSAdam Langley mss_now -= delta; 144833ad798cSAdam Langley } 1449cfb6eeb4SYOSHIFUJI Hideaki 14501da177e4SLinus Torvalds return mss_now; 14511da177e4SLinus Torvalds } 14521da177e4SLinus Torvalds 145386fd14adSWeiping Pan /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. 145486fd14adSWeiping Pan * As additional protections, we do not touch cwnd in retransmission phases, 145586fd14adSWeiping Pan * and if application hit its sndbuf limit recently. 145686fd14adSWeiping Pan */ 145786fd14adSWeiping Pan static void tcp_cwnd_application_limited(struct sock *sk) 1458a762a980SDavid S. Miller { 14599e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1460a762a980SDavid S. Miller 146186fd14adSWeiping Pan if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && 146286fd14adSWeiping Pan sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 146386fd14adSWeiping Pan /* Limited by application or receiver window. */ 146486fd14adSWeiping Pan u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); 146586fd14adSWeiping Pan u32 win_used = max(tp->snd_cwnd_used, init_win); 146686fd14adSWeiping Pan if (win_used < tp->snd_cwnd) { 146786fd14adSWeiping Pan tp->snd_ssthresh = tcp_current_ssthresh(sk); 146886fd14adSWeiping Pan tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1; 146986fd14adSWeiping Pan } 147086fd14adSWeiping Pan tp->snd_cwnd_used = 0; 147186fd14adSWeiping Pan } 147286fd14adSWeiping Pan tp->snd_cwnd_stamp = tcp_time_stamp; 147386fd14adSWeiping Pan } 147486fd14adSWeiping Pan 1475ca8a2263SNeal Cardwell static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited) 1476a762a980SDavid S. Miller { 1477a762a980SDavid S. Miller struct tcp_sock *tp = tcp_sk(sk); 1478a762a980SDavid S. Miller 1479ca8a2263SNeal Cardwell /* Track the maximum number of outstanding packets in each 1480ca8a2263SNeal Cardwell * window, and remember whether we were cwnd-limited then. 1481ca8a2263SNeal Cardwell */ 1482ca8a2263SNeal Cardwell if (!before(tp->snd_una, tp->max_packets_seq) || 1483ca8a2263SNeal Cardwell tp->packets_out > tp->max_packets_out) { 1484ca8a2263SNeal Cardwell tp->max_packets_out = tp->packets_out; 1485ca8a2263SNeal Cardwell tp->max_packets_seq = tp->snd_nxt; 1486ca8a2263SNeal Cardwell tp->is_cwnd_limited = is_cwnd_limited; 1487ca8a2263SNeal Cardwell } 1488e114a710SEric Dumazet 148924901551SEric Dumazet if (tcp_is_cwnd_limited(sk)) { 1490a762a980SDavid S. Miller /* Network is feed fully. */ 1491a762a980SDavid S. Miller tp->snd_cwnd_used = 0; 1492a762a980SDavid S. Miller tp->snd_cwnd_stamp = tcp_time_stamp; 1493a762a980SDavid S. Miller } else { 1494a762a980SDavid S. Miller /* Network starves. */ 1495a762a980SDavid S. Miller if (tp->packets_out > tp->snd_cwnd_used) 1496a762a980SDavid S. Miller tp->snd_cwnd_used = tp->packets_out; 1497a762a980SDavid S. Miller 149815d33c07SDavid S. Miller if (sysctl_tcp_slow_start_after_idle && 149915d33c07SDavid S. Miller (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto) 1500a762a980SDavid S. Miller tcp_cwnd_application_limited(sk); 1501a762a980SDavid S. Miller } 1502a762a980SDavid S. Miller } 1503a762a980SDavid S. Miller 1504d4589926SEric Dumazet /* Minshall's variant of the Nagle send check. */ 1505d4589926SEric Dumazet static bool tcp_minshall_check(const struct tcp_sock *tp) 1506d4589926SEric Dumazet { 1507d4589926SEric Dumazet return after(tp->snd_sml, tp->snd_una) && 1508d4589926SEric Dumazet !after(tp->snd_sml, tp->snd_nxt); 1509d4589926SEric Dumazet } 1510d4589926SEric Dumazet 1511d4589926SEric Dumazet /* Update snd_sml if this skb is under mss 1512d4589926SEric Dumazet * Note that a TSO packet might end with a sub-mss segment 1513d4589926SEric Dumazet * The test is really : 1514d4589926SEric Dumazet * if ((skb->len % mss) != 0) 1515d4589926SEric Dumazet * tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1516d4589926SEric Dumazet * But we can avoid doing the divide again given we already have 1517d4589926SEric Dumazet * skb_pcount = skb->len / mss_now 15180e3a4803SIlpo Järvinen */ 1519d4589926SEric Dumazet static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now, 1520d4589926SEric Dumazet const struct sk_buff *skb) 1521d4589926SEric Dumazet { 1522d4589926SEric Dumazet if (skb->len < tcp_skb_pcount(skb) * mss_now) 1523d4589926SEric Dumazet tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1524d4589926SEric Dumazet } 1525d4589926SEric Dumazet 1526d4589926SEric Dumazet /* Return false, if packet can be sent now without violation Nagle's rules: 1527d4589926SEric Dumazet * 1. It is full sized. (provided by caller in %partial bool) 1528d4589926SEric Dumazet * 2. Or it contains FIN. (already checked by caller) 1529d4589926SEric Dumazet * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. 1530d4589926SEric Dumazet * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 1531d4589926SEric Dumazet * With Minshall's modification: all sent small packets are ACKed. 1532d4589926SEric Dumazet */ 1533d4589926SEric Dumazet static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp, 1534cc93fc51SPeter Pan(潘卫平) int nonagle) 1535d4589926SEric Dumazet { 1536d4589926SEric Dumazet return partial && 1537d4589926SEric Dumazet ((nonagle & TCP_NAGLE_CORK) || 1538d4589926SEric Dumazet (!nonagle && tp->packets_out && tcp_minshall_check(tp))); 1539d4589926SEric Dumazet } 1540605ad7f1SEric Dumazet 1541605ad7f1SEric Dumazet /* Return how many segs we'd like on a TSO packet, 1542605ad7f1SEric Dumazet * to send one TSO packet per ms 1543605ad7f1SEric Dumazet */ 1544605ad7f1SEric Dumazet static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now) 1545605ad7f1SEric Dumazet { 1546605ad7f1SEric Dumazet u32 bytes, segs; 1547605ad7f1SEric Dumazet 1548605ad7f1SEric Dumazet bytes = min(sk->sk_pacing_rate >> 10, 1549605ad7f1SEric Dumazet sk->sk_gso_max_size - 1 - MAX_TCP_HEADER); 1550605ad7f1SEric Dumazet 1551605ad7f1SEric Dumazet /* Goal is to send at least one packet per ms, 1552605ad7f1SEric Dumazet * not one big TSO packet every 100 ms. 1553605ad7f1SEric Dumazet * This preserves ACK clocking and is consistent 1554605ad7f1SEric Dumazet * with tcp_tso_should_defer() heuristic. 1555605ad7f1SEric Dumazet */ 1556605ad7f1SEric Dumazet segs = max_t(u32, bytes / mss_now, sysctl_tcp_min_tso_segs); 1557605ad7f1SEric Dumazet 1558605ad7f1SEric Dumazet return min_t(u32, segs, sk->sk_gso_max_segs); 1559605ad7f1SEric Dumazet } 1560605ad7f1SEric Dumazet 1561d4589926SEric Dumazet /* Returns the portion of skb which can be sent right away */ 1562d4589926SEric Dumazet static unsigned int tcp_mss_split_point(const struct sock *sk, 1563d4589926SEric Dumazet const struct sk_buff *skb, 1564d4589926SEric Dumazet unsigned int mss_now, 1565d4589926SEric Dumazet unsigned int max_segs, 1566d4589926SEric Dumazet int nonagle) 1567c1b4a7e6SDavid S. Miller { 1568cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1569d4589926SEric Dumazet u32 partial, needed, window, max_len; 1570c1b4a7e6SDavid S. Miller 157190840defSIlpo Järvinen window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 15721485348dSBen Hutchings max_len = mss_now * max_segs; 15730e3a4803SIlpo Järvinen 15741485348dSBen Hutchings if (likely(max_len <= window && skb != tcp_write_queue_tail(sk))) 15751485348dSBen Hutchings return max_len; 15760e3a4803SIlpo Järvinen 15775ea3a748SIlpo Järvinen needed = min(skb->len, window); 15785ea3a748SIlpo Järvinen 15791485348dSBen Hutchings if (max_len <= needed) 15801485348dSBen Hutchings return max_len; 15810e3a4803SIlpo Järvinen 1582d4589926SEric Dumazet partial = needed % mss_now; 1583d4589926SEric Dumazet /* If last segment is not a full MSS, check if Nagle rules allow us 1584d4589926SEric Dumazet * to include this last segment in this skb. 1585d4589926SEric Dumazet * Otherwise, we'll split the skb at last MSS boundary 1586d4589926SEric Dumazet */ 1587cc93fc51SPeter Pan(潘卫平) if (tcp_nagle_check(partial != 0, tp, nonagle)) 1588d4589926SEric Dumazet return needed - partial; 1589d4589926SEric Dumazet 1590d4589926SEric Dumazet return needed; 1591c1b4a7e6SDavid S. Miller } 1592c1b4a7e6SDavid S. Miller 1593c1b4a7e6SDavid S. Miller /* Can at least one segment of SKB be sent right now, according to the 1594c1b4a7e6SDavid S. Miller * congestion window rules? If so, return how many segments are allowed. 1595c1b4a7e6SDavid S. Miller */ 1596cf533ea5SEric Dumazet static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, 1597cf533ea5SEric Dumazet const struct sk_buff *skb) 1598c1b4a7e6SDavid S. Miller { 1599d649a7a8SEric Dumazet u32 in_flight, cwnd, halfcwnd; 1600c1b4a7e6SDavid S. Miller 1601c1b4a7e6SDavid S. Miller /* Don't be strict about the congestion window for the final FIN. */ 16024de075e0SEric Dumazet if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && 16034de075e0SEric Dumazet tcp_skb_pcount(skb) == 1) 1604c1b4a7e6SDavid S. Miller return 1; 1605c1b4a7e6SDavid S. Miller 1606c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1607c1b4a7e6SDavid S. Miller cwnd = tp->snd_cwnd; 1608d649a7a8SEric Dumazet if (in_flight >= cwnd) 1609c1b4a7e6SDavid S. Miller return 0; 1610d649a7a8SEric Dumazet 1611d649a7a8SEric Dumazet /* For better scheduling, ensure we have at least 1612d649a7a8SEric Dumazet * 2 GSO packets in flight. 1613d649a7a8SEric Dumazet */ 1614d649a7a8SEric Dumazet halfcwnd = max(cwnd >> 1, 1U); 1615d649a7a8SEric Dumazet return min(halfcwnd, cwnd - in_flight); 1616c1b4a7e6SDavid S. Miller } 1617c1b4a7e6SDavid S. Miller 1618b595076aSUwe Kleine-König /* Initialize TSO state of a skb. 161967edfef7SAndi Kleen * This must be invoked the first time we consider transmitting 1620c1b4a7e6SDavid S. Miller * SKB onto the wire. 1621c1b4a7e6SDavid S. Miller */ 1622cf533ea5SEric Dumazet static int tcp_init_tso_segs(const struct sock *sk, struct sk_buff *skb, 1623056834d9SIlpo Järvinen unsigned int mss_now) 1624c1b4a7e6SDavid S. Miller { 1625c1b4a7e6SDavid S. Miller int tso_segs = tcp_skb_pcount(skb); 1626c1b4a7e6SDavid S. Miller 1627f8269a49SIlpo Järvinen if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { 1628846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1629c1b4a7e6SDavid S. Miller tso_segs = tcp_skb_pcount(skb); 1630c1b4a7e6SDavid S. Miller } 1631c1b4a7e6SDavid S. Miller return tso_segs; 1632c1b4a7e6SDavid S. Miller } 1633c1b4a7e6SDavid S. Miller 1634c1b4a7e6SDavid S. Miller 1635a2a385d6SEric Dumazet /* Return true if the Nagle test allows this packet to be 1636c1b4a7e6SDavid S. Miller * sent now. 1637c1b4a7e6SDavid S. Miller */ 1638a2a385d6SEric Dumazet static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, 1639c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1640c1b4a7e6SDavid S. Miller { 1641c1b4a7e6SDavid S. Miller /* Nagle rule does not apply to frames, which sit in the middle of the 1642c1b4a7e6SDavid S. Miller * write_queue (they have no chances to get new data). 1643c1b4a7e6SDavid S. Miller * 1644c1b4a7e6SDavid S. Miller * This is implemented in the callers, where they modify the 'nonagle' 1645c1b4a7e6SDavid S. Miller * argument based upon the location of SKB in the send queue. 1646c1b4a7e6SDavid S. Miller */ 1647c1b4a7e6SDavid S. Miller if (nonagle & TCP_NAGLE_PUSH) 1648a2a385d6SEric Dumazet return true; 1649c1b4a7e6SDavid S. Miller 16509b44190dSYuchung Cheng /* Don't use the nagle rule for urgent data (or for the final FIN). */ 16519b44190dSYuchung Cheng if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) 1652a2a385d6SEric Dumazet return true; 1653c1b4a7e6SDavid S. Miller 1654cc93fc51SPeter Pan(潘卫平) if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle)) 1655a2a385d6SEric Dumazet return true; 1656c1b4a7e6SDavid S. Miller 1657a2a385d6SEric Dumazet return false; 1658c1b4a7e6SDavid S. Miller } 1659c1b4a7e6SDavid S. Miller 1660c1b4a7e6SDavid S. Miller /* Does at least the first segment of SKB fit into the send window? */ 1661a2a385d6SEric Dumazet static bool tcp_snd_wnd_test(const struct tcp_sock *tp, 1662a2a385d6SEric Dumazet const struct sk_buff *skb, 1663056834d9SIlpo Järvinen unsigned int cur_mss) 1664c1b4a7e6SDavid S. Miller { 1665c1b4a7e6SDavid S. Miller u32 end_seq = TCP_SKB_CB(skb)->end_seq; 1666c1b4a7e6SDavid S. Miller 1667c1b4a7e6SDavid S. Miller if (skb->len > cur_mss) 1668c1b4a7e6SDavid S. Miller end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 1669c1b4a7e6SDavid S. Miller 167090840defSIlpo Järvinen return !after(end_seq, tcp_wnd_end(tp)); 1671c1b4a7e6SDavid S. Miller } 1672c1b4a7e6SDavid S. Miller 1673fe067e8aSDavid S. Miller /* This checks if the data bearing packet SKB (usually tcp_send_head(sk)) 1674c1b4a7e6SDavid S. Miller * should be put on the wire right now. If so, it returns the number of 1675c1b4a7e6SDavid S. Miller * packets allowed by the congestion window. 1676c1b4a7e6SDavid S. Miller */ 1677cf533ea5SEric Dumazet static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb, 1678c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1679c1b4a7e6SDavid S. Miller { 1680cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1681c1b4a7e6SDavid S. Miller unsigned int cwnd_quota; 1682c1b4a7e6SDavid S. Miller 1683846998aeSDavid S. Miller tcp_init_tso_segs(sk, skb, cur_mss); 1684c1b4a7e6SDavid S. Miller 1685c1b4a7e6SDavid S. Miller if (!tcp_nagle_test(tp, skb, cur_mss, nonagle)) 1686c1b4a7e6SDavid S. Miller return 0; 1687c1b4a7e6SDavid S. Miller 1688c1b4a7e6SDavid S. Miller cwnd_quota = tcp_cwnd_test(tp, skb); 1689056834d9SIlpo Järvinen if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss)) 1690c1b4a7e6SDavid S. Miller cwnd_quota = 0; 1691c1b4a7e6SDavid S. Miller 1692c1b4a7e6SDavid S. Miller return cwnd_quota; 1693c1b4a7e6SDavid S. Miller } 1694c1b4a7e6SDavid S. Miller 169567edfef7SAndi Kleen /* Test if sending is allowed right now. */ 1696a2a385d6SEric Dumazet bool tcp_may_send_now(struct sock *sk) 1697c1b4a7e6SDavid S. Miller { 1698cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1699fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 1700c1b4a7e6SDavid S. Miller 1701a02cec21SEric Dumazet return skb && 17020c54b85fSIlpo Järvinen tcp_snd_test(sk, skb, tcp_current_mss(sk), 1703c1b4a7e6SDavid S. Miller (tcp_skb_is_last(sk, skb) ? 1704a02cec21SEric Dumazet tp->nonagle : TCP_NAGLE_PUSH)); 1705c1b4a7e6SDavid S. Miller } 1706c1b4a7e6SDavid S. Miller 1707c1b4a7e6SDavid S. Miller /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 1708c1b4a7e6SDavid S. Miller * which is put after SKB on the list. It is very much like 1709c1b4a7e6SDavid S. Miller * tcp_fragment() except that it may make several kinds of assumptions 1710c1b4a7e6SDavid S. Miller * in order to speed up the splitting operation. In particular, we 1711c1b4a7e6SDavid S. Miller * know that all the data is in scatter-gather pages, and that the 1712c1b4a7e6SDavid S. Miller * packet has never been sent out before (and thus is not cloned). 1713c1b4a7e6SDavid S. Miller */ 1714056834d9SIlpo Järvinen static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, 1715c4ead4c5SEric Dumazet unsigned int mss_now, gfp_t gfp) 1716c1b4a7e6SDavid S. Miller { 1717c1b4a7e6SDavid S. Miller struct sk_buff *buff; 1718c1b4a7e6SDavid S. Miller int nlen = skb->len - len; 17199ce01461SIlpo Järvinen u8 flags; 1720c1b4a7e6SDavid S. Miller 1721c1b4a7e6SDavid S. Miller /* All of a TSO frame must be composed of paged data. */ 1722c8ac3774SHerbert Xu if (skb->len != skb->data_len) 17236cc55e09SOctavian Purdila return tcp_fragment(sk, skb, len, mss_now, gfp); 1724c1b4a7e6SDavid S. Miller 1725c4ead4c5SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, gfp); 172651456b29SIan Morris if (unlikely(!buff)) 1727c1b4a7e6SDavid S. Miller return -ENOMEM; 1728c1b4a7e6SDavid S. Miller 17293ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 17303ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1731b60b49eaSHerbert Xu buff->truesize += nlen; 1732c1b4a7e6SDavid S. Miller skb->truesize -= nlen; 1733c1b4a7e6SDavid S. Miller 1734c1b4a7e6SDavid S. Miller /* Correct the sequence numbers. */ 1735c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 1736c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 1737c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 1738c1b4a7e6SDavid S. Miller 1739c1b4a7e6SDavid S. Miller /* PSH and FIN should only be set in the second packet. */ 17404de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 17414de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 17424de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1743c1b4a7e6SDavid S. Miller 1744c1b4a7e6SDavid S. Miller /* This packet was never sent out yet, so no SACK bits. */ 1745c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->sacked = 0; 1746c1b4a7e6SDavid S. Miller 174784fa7933SPatrick McHardy buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; 1748c1b4a7e6SDavid S. Miller skb_split(skb, buff, len); 1749490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 1750c1b4a7e6SDavid S. Miller 1751c1b4a7e6SDavid S. Miller /* Fix up tso_factor for both original and new SKB. */ 1752846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1753846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, buff, mss_now); 1754c1b4a7e6SDavid S. Miller 1755c1b4a7e6SDavid S. Miller /* Link BUFF into the send queue. */ 1756f4a775d1SEric Dumazet __skb_header_release(buff); 1757fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 1758c1b4a7e6SDavid S. Miller 1759c1b4a7e6SDavid S. Miller return 0; 1760c1b4a7e6SDavid S. Miller } 1761c1b4a7e6SDavid S. Miller 1762c1b4a7e6SDavid S. Miller /* Try to defer sending, if possible, in order to minimize the amount 1763c1b4a7e6SDavid S. Miller * of TSO splitting we do. View it as a kind of TSO Nagle test. 1764c1b4a7e6SDavid S. Miller * 1765c1b4a7e6SDavid S. Miller * This algorithm is from John Heffner. 1766c1b4a7e6SDavid S. Miller */ 1767ca8a2263SNeal Cardwell static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb, 1768605ad7f1SEric Dumazet bool *is_cwnd_limited, u32 max_segs) 1769c1b4a7e6SDavid S. Miller { 17706687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 177150c8339eSEric Dumazet u32 age, send_win, cong_win, limit, in_flight; 177250c8339eSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 177350c8339eSEric Dumazet struct skb_mstamp now; 177450c8339eSEric Dumazet struct sk_buff *head; 1775ad9f4f50SEric Dumazet int win_divisor; 1776c1b4a7e6SDavid S. Miller 17774de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 1778ae8064acSJohn Heffner goto send_now; 1779c1b4a7e6SDavid S. Miller 1780a0ea700eSEric Dumazet if (!((1 << icsk->icsk_ca_state) & (TCPF_CA_Open | TCPF_CA_CWR))) 1781ae8064acSJohn Heffner goto send_now; 1782ae8064acSJohn Heffner 17835f852eb5SEric Dumazet /* Avoid bursty behavior by allowing defer 17845f852eb5SEric Dumazet * only if the last write was recent. 17855f852eb5SEric Dumazet */ 17865f852eb5SEric Dumazet if ((s32)(tcp_time_stamp - tp->lsndtime) > 0) 1787ae8064acSJohn Heffner goto send_now; 1788908a75c1SDavid S. Miller 1789c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1790c1b4a7e6SDavid S. Miller 1791056834d9SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight)); 1792c1b4a7e6SDavid S. Miller 179390840defSIlpo Järvinen send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1794c1b4a7e6SDavid S. Miller 1795c1b4a7e6SDavid S. Miller /* From in_flight test above, we know that cwnd > in_flight. */ 1796c1b4a7e6SDavid S. Miller cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; 1797c1b4a7e6SDavid S. Miller 1798c1b4a7e6SDavid S. Miller limit = min(send_win, cong_win); 1799c1b4a7e6SDavid S. Miller 1800ba244fe9SDavid S. Miller /* If a full-sized TSO skb can be sent, do it. */ 1801605ad7f1SEric Dumazet if (limit >= max_segs * tp->mss_cache) 1802ae8064acSJohn Heffner goto send_now; 1803ba244fe9SDavid S. Miller 180462ad2761SIlpo Järvinen /* Middle in queue won't get any more data, full sendable already? */ 180562ad2761SIlpo Järvinen if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) 180662ad2761SIlpo Järvinen goto send_now; 180762ad2761SIlpo Järvinen 1808ad9f4f50SEric Dumazet win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor); 1809ad9f4f50SEric Dumazet if (win_divisor) { 1810c1b4a7e6SDavid S. Miller u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); 1811c1b4a7e6SDavid S. Miller 1812c1b4a7e6SDavid S. Miller /* If at least some fraction of a window is available, 1813c1b4a7e6SDavid S. Miller * just use it. 1814c1b4a7e6SDavid S. Miller */ 1815ad9f4f50SEric Dumazet chunk /= win_divisor; 1816c1b4a7e6SDavid S. Miller if (limit >= chunk) 1817ae8064acSJohn Heffner goto send_now; 1818c1b4a7e6SDavid S. Miller } else { 1819c1b4a7e6SDavid S. Miller /* Different approach, try not to defer past a single 1820c1b4a7e6SDavid S. Miller * ACK. Receiver should ACK every other full sized 1821c1b4a7e6SDavid S. Miller * frame, so if we have space for more than 3 frames 1822c1b4a7e6SDavid S. Miller * then send now. 1823c1b4a7e6SDavid S. Miller */ 18246b5a5c0dSNeal Cardwell if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) 1825ae8064acSJohn Heffner goto send_now; 1826c1b4a7e6SDavid S. Miller } 1827c1b4a7e6SDavid S. Miller 182850c8339eSEric Dumazet head = tcp_write_queue_head(sk); 182950c8339eSEric Dumazet skb_mstamp_get(&now); 183050c8339eSEric Dumazet age = skb_mstamp_us_delta(&now, &head->skb_mstamp); 183150c8339eSEric Dumazet /* If next ACK is likely to come too late (half srtt), do not defer */ 183250c8339eSEric Dumazet if (age < (tp->srtt_us >> 4)) 183350c8339eSEric Dumazet goto send_now; 183450c8339eSEric Dumazet 18355f852eb5SEric Dumazet /* Ok, it looks like it is advisable to defer. */ 1836ae8064acSJohn Heffner 1837ca8a2263SNeal Cardwell if (cong_win < send_win && cong_win < skb->len) 1838ca8a2263SNeal Cardwell *is_cwnd_limited = true; 1839ca8a2263SNeal Cardwell 1840a2a385d6SEric Dumazet return true; 1841ae8064acSJohn Heffner 1842ae8064acSJohn Heffner send_now: 1843a2a385d6SEric Dumazet return false; 1844c1b4a7e6SDavid S. Miller } 1845c1b4a7e6SDavid S. Miller 184605cbc0dbSFan Du static inline void tcp_mtu_check_reprobe(struct sock *sk) 184705cbc0dbSFan Du { 184805cbc0dbSFan Du struct inet_connection_sock *icsk = inet_csk(sk); 184905cbc0dbSFan Du struct tcp_sock *tp = tcp_sk(sk); 185005cbc0dbSFan Du struct net *net = sock_net(sk); 185105cbc0dbSFan Du u32 interval; 185205cbc0dbSFan Du s32 delta; 185305cbc0dbSFan Du 185405cbc0dbSFan Du interval = net->ipv4.sysctl_tcp_probe_interval; 185505cbc0dbSFan Du delta = tcp_time_stamp - icsk->icsk_mtup.probe_timestamp; 185605cbc0dbSFan Du if (unlikely(delta >= interval * HZ)) { 185705cbc0dbSFan Du int mss = tcp_current_mss(sk); 185805cbc0dbSFan Du 185905cbc0dbSFan Du /* Update current search range */ 186005cbc0dbSFan Du icsk->icsk_mtup.probe_size = 0; 186105cbc0dbSFan Du icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + 186205cbc0dbSFan Du sizeof(struct tcphdr) + 186305cbc0dbSFan Du icsk->icsk_af_ops->net_header_len; 186405cbc0dbSFan Du icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss); 186505cbc0dbSFan Du 186605cbc0dbSFan Du /* Update probe time stamp */ 186705cbc0dbSFan Du icsk->icsk_mtup.probe_timestamp = tcp_time_stamp; 186805cbc0dbSFan Du } 186905cbc0dbSFan Du } 187005cbc0dbSFan Du 18715d424d5aSJohn Heffner /* Create a new MTU probe if we are ready. 187267edfef7SAndi Kleen * MTU probe is regularly attempting to increase the path MTU by 187367edfef7SAndi Kleen * deliberately sending larger packets. This discovers routing 187467edfef7SAndi Kleen * changes resulting in larger path MTUs. 187567edfef7SAndi Kleen * 18765d424d5aSJohn Heffner * Returns 0 if we should wait to probe (no cwnd available), 18775d424d5aSJohn Heffner * 1 if a probe was sent, 1878056834d9SIlpo Järvinen * -1 otherwise 1879056834d9SIlpo Järvinen */ 18805d424d5aSJohn Heffner static int tcp_mtu_probe(struct sock *sk) 18815d424d5aSJohn Heffner { 18825d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 18835d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 18845d424d5aSJohn Heffner struct sk_buff *skb, *nskb, *next; 18856b58e0a5SFan Du struct net *net = sock_net(sk); 18865d424d5aSJohn Heffner int len; 18875d424d5aSJohn Heffner int probe_size; 188891cc17c0SIlpo Järvinen int size_needed; 18895d424d5aSJohn Heffner int copy; 18905d424d5aSJohn Heffner int mss_now; 18916b58e0a5SFan Du int interval; 18925d424d5aSJohn Heffner 18935d424d5aSJohn Heffner /* Not currently probing/verifying, 18945d424d5aSJohn Heffner * not in recovery, 18955d424d5aSJohn Heffner * have enough cwnd, and 18965d424d5aSJohn Heffner * not SACKing (the variable headers throw things off) */ 18975d424d5aSJohn Heffner if (!icsk->icsk_mtup.enabled || 18985d424d5aSJohn Heffner icsk->icsk_mtup.probe_size || 18995d424d5aSJohn Heffner inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 19005d424d5aSJohn Heffner tp->snd_cwnd < 11 || 1901cabeccbdSIlpo Järvinen tp->rx_opt.num_sacks || tp->rx_opt.dsack) 19025d424d5aSJohn Heffner return -1; 19035d424d5aSJohn Heffner 19046b58e0a5SFan Du /* Use binary search for probe_size between tcp_mss_base, 19056b58e0a5SFan Du * and current mss_clamp. if (search_high - search_low) 19066b58e0a5SFan Du * smaller than a threshold, backoff from probing. 19076b58e0a5SFan Du */ 19080c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 19096b58e0a5SFan Du probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high + 19106b58e0a5SFan Du icsk->icsk_mtup.search_low) >> 1); 191191cc17c0SIlpo Järvinen size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; 19126b58e0a5SFan Du interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low; 191305cbc0dbSFan Du /* When misfortune happens, we are reprobing actively, 191405cbc0dbSFan Du * and then reprobe timer has expired. We stick with current 191505cbc0dbSFan Du * probing process by not resetting search range to its orignal. 191605cbc0dbSFan Du */ 19176b58e0a5SFan Du if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) || 191805cbc0dbSFan Du interval < net->ipv4.sysctl_tcp_probe_threshold) { 191905cbc0dbSFan Du /* Check whether enough time has elaplased for 192005cbc0dbSFan Du * another round of probing. 192105cbc0dbSFan Du */ 192205cbc0dbSFan Du tcp_mtu_check_reprobe(sk); 19235d424d5aSJohn Heffner return -1; 19245d424d5aSJohn Heffner } 19255d424d5aSJohn Heffner 19265d424d5aSJohn Heffner /* Have enough data in the send queue to probe? */ 19277f9c33e5SIlpo Järvinen if (tp->write_seq - tp->snd_nxt < size_needed) 19285d424d5aSJohn Heffner return -1; 19295d424d5aSJohn Heffner 193091cc17c0SIlpo Järvinen if (tp->snd_wnd < size_needed) 19315d424d5aSJohn Heffner return -1; 193290840defSIlpo Järvinen if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) 19335d424d5aSJohn Heffner return 0; 19345d424d5aSJohn Heffner 1935d67c58e9SIlpo Järvinen /* Do we need to wait to drain cwnd? With none in flight, don't stall */ 1936d67c58e9SIlpo Järvinen if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { 1937d67c58e9SIlpo Järvinen if (!tcp_packets_in_flight(tp)) 19385d424d5aSJohn Heffner return -1; 19395d424d5aSJohn Heffner else 19405d424d5aSJohn Heffner return 0; 19415d424d5aSJohn Heffner } 19425d424d5aSJohn Heffner 19435d424d5aSJohn Heffner /* We're allowed to probe. Build it now. */ 194451456b29SIan Morris nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC); 194551456b29SIan Morris if (!nskb) 19465d424d5aSJohn Heffner return -1; 19473ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 19483ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 19495d424d5aSJohn Heffner 1950fe067e8aSDavid S. Miller skb = tcp_send_head(sk); 19515d424d5aSJohn Heffner 19525d424d5aSJohn Heffner TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 19535d424d5aSJohn Heffner TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 19544de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; 19555d424d5aSJohn Heffner TCP_SKB_CB(nskb)->sacked = 0; 19565d424d5aSJohn Heffner nskb->csum = 0; 195784fa7933SPatrick McHardy nskb->ip_summed = skb->ip_summed; 19585d424d5aSJohn Heffner 195950c4817eSIlpo Järvinen tcp_insert_write_queue_before(nskb, skb, sk); 196050c4817eSIlpo Järvinen 19615d424d5aSJohn Heffner len = 0; 1962234b6860SIlpo Järvinen tcp_for_write_queue_from_safe(skb, next, sk) { 19635d424d5aSJohn Heffner copy = min_t(int, skb->len, probe_size - len); 19645d424d5aSJohn Heffner if (nskb->ip_summed) 19655d424d5aSJohn Heffner skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); 19665d424d5aSJohn Heffner else 19675d424d5aSJohn Heffner nskb->csum = skb_copy_and_csum_bits(skb, 0, 1968056834d9SIlpo Järvinen skb_put(nskb, copy), 1969056834d9SIlpo Järvinen copy, nskb->csum); 19705d424d5aSJohn Heffner 19715d424d5aSJohn Heffner if (skb->len <= copy) { 19725d424d5aSJohn Heffner /* We've eaten all the data from this skb. 19735d424d5aSJohn Heffner * Throw it away. */ 19744de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 1975fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 19763ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 19775d424d5aSJohn Heffner } else { 19784de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & 1979a3433f35SChangli Gao ~(TCPHDR_FIN|TCPHDR_PSH); 19805d424d5aSJohn Heffner if (!skb_shinfo(skb)->nr_frags) { 19815d424d5aSJohn Heffner skb_pull(skb, copy); 198284fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 1983056834d9SIlpo Järvinen skb->csum = csum_partial(skb->data, 1984056834d9SIlpo Järvinen skb->len, 0); 19855d424d5aSJohn Heffner } else { 19865d424d5aSJohn Heffner __pskb_trim_head(skb, copy); 19875d424d5aSJohn Heffner tcp_set_skb_tso_segs(sk, skb, mss_now); 19885d424d5aSJohn Heffner } 19895d424d5aSJohn Heffner TCP_SKB_CB(skb)->seq += copy; 19905d424d5aSJohn Heffner } 19915d424d5aSJohn Heffner 19925d424d5aSJohn Heffner len += copy; 1993234b6860SIlpo Järvinen 1994234b6860SIlpo Järvinen if (len >= probe_size) 1995234b6860SIlpo Järvinen break; 19965d424d5aSJohn Heffner } 19975d424d5aSJohn Heffner tcp_init_tso_segs(sk, nskb, nskb->len); 19985d424d5aSJohn Heffner 19995d424d5aSJohn Heffner /* We're ready to send. If this fails, the probe will 20007faee5c0SEric Dumazet * be resegmented into mss-sized pieces by tcp_write_xmit(). 20017faee5c0SEric Dumazet */ 20025d424d5aSJohn Heffner if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 20035d424d5aSJohn Heffner /* Decrement cwnd here because we are sending 20045d424d5aSJohn Heffner * effectively two packets. */ 20055d424d5aSJohn Heffner tp->snd_cwnd--; 200666f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, nskb); 20075d424d5aSJohn Heffner 20085d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 20090e7b1368SJohn Heffner tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 20100e7b1368SJohn Heffner tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 20115d424d5aSJohn Heffner 20125d424d5aSJohn Heffner return 1; 20135d424d5aSJohn Heffner } 20145d424d5aSJohn Heffner 20155d424d5aSJohn Heffner return -1; 20165d424d5aSJohn Heffner } 20175d424d5aSJohn Heffner 20181da177e4SLinus Torvalds /* This routine writes packets to the network. It advances the 20191da177e4SLinus Torvalds * send_head. This happens as incoming acks open up the remote 20201da177e4SLinus Torvalds * window for us. 20211da177e4SLinus Torvalds * 2022f8269a49SIlpo Järvinen * LARGESEND note: !tcp_urg_mode is overkill, only frames between 2023f8269a49SIlpo Järvinen * snd_up-64k-mss .. snd_up cannot be large. However, taking into 2024f8269a49SIlpo Järvinen * account rare use of URG, this is not a big flaw. 2025f8269a49SIlpo Järvinen * 20266ba8a3b1SNandita Dukkipati * Send at most one packet when push_one > 0. Temporarily ignore 20276ba8a3b1SNandita Dukkipati * cwnd limit to force at most one packet out when push_one == 2. 20286ba8a3b1SNandita Dukkipati 2029a2a385d6SEric Dumazet * Returns true, if no segments are in flight and we have queued segments, 2030a2a385d6SEric Dumazet * but cannot send anything now because of SWS or another problem. 20311da177e4SLinus Torvalds */ 2032a2a385d6SEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 2033d5dd9175SIlpo Järvinen int push_one, gfp_t gfp) 20341da177e4SLinus Torvalds { 20351da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 203692df7b51SDavid S. Miller struct sk_buff *skb; 2037c1b4a7e6SDavid S. Miller unsigned int tso_segs, sent_pkts; 2038c1b4a7e6SDavid S. Miller int cwnd_quota; 20395d424d5aSJohn Heffner int result; 2040ca8a2263SNeal Cardwell bool is_cwnd_limited = false; 2041605ad7f1SEric Dumazet u32 max_segs; 20421da177e4SLinus Torvalds 2043c1b4a7e6SDavid S. Miller sent_pkts = 0; 20445d424d5aSJohn Heffner 2045d5dd9175SIlpo Järvinen if (!push_one) { 20465d424d5aSJohn Heffner /* Do MTU probing. */ 2047d5dd9175SIlpo Järvinen result = tcp_mtu_probe(sk); 2048d5dd9175SIlpo Järvinen if (!result) { 2049a2a385d6SEric Dumazet return false; 20505d424d5aSJohn Heffner } else if (result > 0) { 20515d424d5aSJohn Heffner sent_pkts = 1; 20525d424d5aSJohn Heffner } 2053d5dd9175SIlpo Järvinen } 20545d424d5aSJohn Heffner 2055605ad7f1SEric Dumazet max_segs = tcp_tso_autosize(sk, mss_now); 2056fe067e8aSDavid S. Miller while ((skb = tcp_send_head(sk))) { 2057c8ac3774SHerbert Xu unsigned int limit; 2058c8ac3774SHerbert Xu 2059b68e9f85SHerbert Xu tso_segs = tcp_init_tso_segs(sk, skb, mss_now); 2060c1b4a7e6SDavid S. Miller BUG_ON(!tso_segs); 2061c1b4a7e6SDavid S. Miller 20629d186cacSAndrey Vagin if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) { 20637faee5c0SEric Dumazet /* "skb_mstamp" is used as a start point for the retransmit timer */ 20647faee5c0SEric Dumazet skb_mstamp_get(&skb->skb_mstamp); 2065ec342325SAndrew Vagin goto repair; /* Skip network transmission */ 20669d186cacSAndrey Vagin } 2067ec342325SAndrew Vagin 2068b68e9f85SHerbert Xu cwnd_quota = tcp_cwnd_test(tp, skb); 20696ba8a3b1SNandita Dukkipati if (!cwnd_quota) { 2070ca8a2263SNeal Cardwell is_cwnd_limited = true; 20716ba8a3b1SNandita Dukkipati if (push_one == 2) 20726ba8a3b1SNandita Dukkipati /* Force out a loss probe pkt. */ 20736ba8a3b1SNandita Dukkipati cwnd_quota = 1; 20746ba8a3b1SNandita Dukkipati else 2075b68e9f85SHerbert Xu break; 20766ba8a3b1SNandita Dukkipati } 2077b68e9f85SHerbert Xu 2078b68e9f85SHerbert Xu if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) 2079b68e9f85SHerbert Xu break; 2080b68e9f85SHerbert Xu 2081843925f3SHerbert Xu if (tso_segs == 1 || !max_segs) { 2082aa93466bSDavid S. Miller if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 2083aa93466bSDavid S. Miller (tcp_skb_is_last(sk, skb) ? 2084aa93466bSDavid S. Miller nonagle : TCP_NAGLE_PUSH)))) 2085aa93466bSDavid S. Miller break; 2086c1b4a7e6SDavid S. Miller } else { 2087ca8a2263SNeal Cardwell if (!push_one && 2088605ad7f1SEric Dumazet tcp_tso_should_defer(sk, skb, &is_cwnd_limited, 2089605ad7f1SEric Dumazet max_segs)) 2090aa93466bSDavid S. Miller break; 2091c1b4a7e6SDavid S. Miller } 2092aa93466bSDavid S. Miller 2093605ad7f1SEric Dumazet limit = mss_now; 2094843925f3SHerbert Xu if (tso_segs > 1 && max_segs && !tcp_urg_mode(tp)) 2095605ad7f1SEric Dumazet limit = tcp_mss_split_point(sk, skb, mss_now, 2096605ad7f1SEric Dumazet min_t(unsigned int, 2097605ad7f1SEric Dumazet cwnd_quota, 2098605ad7f1SEric Dumazet max_segs), 2099605ad7f1SEric Dumazet nonagle); 2100605ad7f1SEric Dumazet 2101605ad7f1SEric Dumazet if (skb->len > limit && 2102605ad7f1SEric Dumazet unlikely(tso_fragment(sk, skb, limit, mss_now, gfp))) 2103605ad7f1SEric Dumazet break; 2104605ad7f1SEric Dumazet 2105c9eeec26SEric Dumazet /* TCP Small Queues : 2106c9eeec26SEric Dumazet * Control number of packets in qdisc/devices to two packets / or ~1 ms. 2107c9eeec26SEric Dumazet * This allows for : 2108c9eeec26SEric Dumazet * - better RTT estimation and ACK scheduling 2109c9eeec26SEric Dumazet * - faster recovery 2110c9eeec26SEric Dumazet * - high rates 211198e09386SEric Dumazet * Alas, some drivers / subsystems require a fair amount 211298e09386SEric Dumazet * of queued bytes to ensure line rate. 211398e09386SEric Dumazet * One example is wifi aggregation (802.11 AMPDU) 211446d3ceabSEric Dumazet */ 2115605ad7f1SEric Dumazet limit = max(2 * skb->truesize, sk->sk_pacing_rate >> 10); 2116605ad7f1SEric Dumazet limit = min_t(u32, limit, sysctl_tcp_limit_output_bytes); 2117c9eeec26SEric Dumazet 2118c9eeec26SEric Dumazet if (atomic_read(&sk->sk_wmem_alloc) > limit) { 211946d3ceabSEric Dumazet set_bit(TSQ_THROTTLED, &tp->tsq_flags); 2120bf06200eSJohn Ogness /* It is possible TX completion already happened 2121bf06200eSJohn Ogness * before we set TSQ_THROTTLED, so we must 2122bf06200eSJohn Ogness * test again the condition. 2123bf06200eSJohn Ogness */ 21244e857c58SPeter Zijlstra smp_mb__after_atomic(); 2125bf06200eSJohn Ogness if (atomic_read(&sk->sk_wmem_alloc) > limit) 212646d3ceabSEric Dumazet break; 212746d3ceabSEric Dumazet } 2128c9eeec26SEric Dumazet 2129d5dd9175SIlpo Järvinen if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) 21301da177e4SLinus Torvalds break; 21311da177e4SLinus Torvalds 2132ec342325SAndrew Vagin repair: 21331da177e4SLinus Torvalds /* Advance the send_head. This one is sent out. 21341da177e4SLinus Torvalds * This call will increment packets_out. 21351da177e4SLinus Torvalds */ 213666f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 21371da177e4SLinus Torvalds 21381da177e4SLinus Torvalds tcp_minshall_update(tp, mss_now, skb); 2139a262f0cdSNandita Dukkipati sent_pkts += tcp_skb_pcount(skb); 2140d5dd9175SIlpo Järvinen 2141d5dd9175SIlpo Järvinen if (push_one) 2142d5dd9175SIlpo Järvinen break; 21431da177e4SLinus Torvalds } 21441da177e4SLinus Torvalds 2145aa93466bSDavid S. Miller if (likely(sent_pkts)) { 2146684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 2147684bad11SYuchung Cheng tp->prr_out += sent_pkts; 21486ba8a3b1SNandita Dukkipati 21496ba8a3b1SNandita Dukkipati /* Send one loss probe per tail loss episode. */ 21506ba8a3b1SNandita Dukkipati if (push_one != 2) 21516ba8a3b1SNandita Dukkipati tcp_schedule_loss_probe(sk); 2152ca8a2263SNeal Cardwell tcp_cwnd_validate(sk, is_cwnd_limited); 2153a2a385d6SEric Dumazet return false; 21541da177e4SLinus Torvalds } 21556ba8a3b1SNandita Dukkipati return (push_one == 2) || (!tp->packets_out && tcp_send_head(sk)); 21566ba8a3b1SNandita Dukkipati } 21576ba8a3b1SNandita Dukkipati 21586ba8a3b1SNandita Dukkipati bool tcp_schedule_loss_probe(struct sock *sk) 21596ba8a3b1SNandita Dukkipati { 21606ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 21616ba8a3b1SNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 21626ba8a3b1SNandita Dukkipati u32 timeout, tlp_time_stamp, rto_time_stamp; 2163740b0f18SEric Dumazet u32 rtt = usecs_to_jiffies(tp->srtt_us >> 3); 21646ba8a3b1SNandita Dukkipati 21656ba8a3b1SNandita Dukkipati if (WARN_ON(icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS)) 21666ba8a3b1SNandita Dukkipati return false; 21676ba8a3b1SNandita Dukkipati /* No consecutive loss probes. */ 21686ba8a3b1SNandita Dukkipati if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) { 21696ba8a3b1SNandita Dukkipati tcp_rearm_rto(sk); 21706ba8a3b1SNandita Dukkipati return false; 21716ba8a3b1SNandita Dukkipati } 21726ba8a3b1SNandita Dukkipati /* Don't do any loss probe on a Fast Open connection before 3WHS 21736ba8a3b1SNandita Dukkipati * finishes. 21746ba8a3b1SNandita Dukkipati */ 21756ba8a3b1SNandita Dukkipati if (sk->sk_state == TCP_SYN_RECV) 21766ba8a3b1SNandita Dukkipati return false; 21776ba8a3b1SNandita Dukkipati 21786ba8a3b1SNandita Dukkipati /* TLP is only scheduled when next timer event is RTO. */ 21796ba8a3b1SNandita Dukkipati if (icsk->icsk_pending != ICSK_TIME_RETRANS) 21806ba8a3b1SNandita Dukkipati return false; 21816ba8a3b1SNandita Dukkipati 21826ba8a3b1SNandita Dukkipati /* Schedule a loss probe in 2*RTT for SACK capable connections 21836ba8a3b1SNandita Dukkipati * in Open state, that are either limited by cwnd or application. 21846ba8a3b1SNandita Dukkipati */ 2185740b0f18SEric Dumazet if (sysctl_tcp_early_retrans < 3 || !tp->srtt_us || !tp->packets_out || 21866ba8a3b1SNandita Dukkipati !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open) 21876ba8a3b1SNandita Dukkipati return false; 21886ba8a3b1SNandita Dukkipati 21896ba8a3b1SNandita Dukkipati if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) && 21906ba8a3b1SNandita Dukkipati tcp_send_head(sk)) 21916ba8a3b1SNandita Dukkipati return false; 21926ba8a3b1SNandita Dukkipati 21936ba8a3b1SNandita Dukkipati /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account 21946ba8a3b1SNandita Dukkipati * for delayed ack when there's one outstanding packet. 21956ba8a3b1SNandita Dukkipati */ 21966ba8a3b1SNandita Dukkipati timeout = rtt << 1; 21976ba8a3b1SNandita Dukkipati if (tp->packets_out == 1) 21986ba8a3b1SNandita Dukkipati timeout = max_t(u32, timeout, 21996ba8a3b1SNandita Dukkipati (rtt + (rtt >> 1) + TCP_DELACK_MAX)); 22006ba8a3b1SNandita Dukkipati timeout = max_t(u32, timeout, msecs_to_jiffies(10)); 22016ba8a3b1SNandita Dukkipati 22026ba8a3b1SNandita Dukkipati /* If RTO is shorter, just schedule TLP in its place. */ 22036ba8a3b1SNandita Dukkipati tlp_time_stamp = tcp_time_stamp + timeout; 22046ba8a3b1SNandita Dukkipati rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout; 22056ba8a3b1SNandita Dukkipati if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) { 22066ba8a3b1SNandita Dukkipati s32 delta = rto_time_stamp - tcp_time_stamp; 22076ba8a3b1SNandita Dukkipati if (delta > 0) 22086ba8a3b1SNandita Dukkipati timeout = delta; 22096ba8a3b1SNandita Dukkipati } 22106ba8a3b1SNandita Dukkipati 22116ba8a3b1SNandita Dukkipati inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, 22126ba8a3b1SNandita Dukkipati TCP_RTO_MAX); 22136ba8a3b1SNandita Dukkipati return true; 22146ba8a3b1SNandita Dukkipati } 22156ba8a3b1SNandita Dukkipati 22161f3279aeSEric Dumazet /* Thanks to skb fast clones, we can detect if a prior transmit of 22171f3279aeSEric Dumazet * a packet is still in a qdisc or driver queue. 22181f3279aeSEric Dumazet * In this case, there is very little point doing a retransmit ! 22191f3279aeSEric Dumazet * Note: This is called from BH context only. 22201f3279aeSEric Dumazet */ 22211f3279aeSEric Dumazet static bool skb_still_in_host_queue(const struct sock *sk, 22221f3279aeSEric Dumazet const struct sk_buff *skb) 22231f3279aeSEric Dumazet { 222439bb5e62SEric Dumazet if (unlikely(skb_fclone_busy(sk, skb))) { 22251f3279aeSEric Dumazet NET_INC_STATS_BH(sock_net(sk), 22261f3279aeSEric Dumazet LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES); 22271f3279aeSEric Dumazet return true; 22281f3279aeSEric Dumazet } 22291f3279aeSEric Dumazet return false; 22301f3279aeSEric Dumazet } 22311f3279aeSEric Dumazet 22326ba8a3b1SNandita Dukkipati /* When probe timeout (PTO) fires, send a new segment if one exists, else 22336ba8a3b1SNandita Dukkipati * retransmit the last segment. 22346ba8a3b1SNandita Dukkipati */ 22356ba8a3b1SNandita Dukkipati void tcp_send_loss_probe(struct sock *sk) 22366ba8a3b1SNandita Dukkipati { 22379b717a8dSNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 22386ba8a3b1SNandita Dukkipati struct sk_buff *skb; 22396ba8a3b1SNandita Dukkipati int pcount; 22406ba8a3b1SNandita Dukkipati int mss = tcp_current_mss(sk); 22416ba8a3b1SNandita Dukkipati int err = -1; 22426ba8a3b1SNandita Dukkipati 224300db4124SIan Morris if (tcp_send_head(sk)) { 22446ba8a3b1SNandita Dukkipati err = tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC); 22456ba8a3b1SNandita Dukkipati goto rearm_timer; 22466ba8a3b1SNandita Dukkipati } 22476ba8a3b1SNandita Dukkipati 22489b717a8dSNandita Dukkipati /* At most one outstanding TLP retransmission. */ 22499b717a8dSNandita Dukkipati if (tp->tlp_high_seq) 22509b717a8dSNandita Dukkipati goto rearm_timer; 22519b717a8dSNandita Dukkipati 22526ba8a3b1SNandita Dukkipati /* Retransmit last segment. */ 22536ba8a3b1SNandita Dukkipati skb = tcp_write_queue_tail(sk); 22546ba8a3b1SNandita Dukkipati if (WARN_ON(!skb)) 22556ba8a3b1SNandita Dukkipati goto rearm_timer; 22566ba8a3b1SNandita Dukkipati 22571f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 22581f3279aeSEric Dumazet goto rearm_timer; 22591f3279aeSEric Dumazet 22606ba8a3b1SNandita Dukkipati pcount = tcp_skb_pcount(skb); 22616ba8a3b1SNandita Dukkipati if (WARN_ON(!pcount)) 22626ba8a3b1SNandita Dukkipati goto rearm_timer; 22636ba8a3b1SNandita Dukkipati 22646ba8a3b1SNandita Dukkipati if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) { 22656cc55e09SOctavian Purdila if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss, 22666cc55e09SOctavian Purdila GFP_ATOMIC))) 22676ba8a3b1SNandita Dukkipati goto rearm_timer; 22686ba8a3b1SNandita Dukkipati skb = tcp_write_queue_tail(sk); 22696ba8a3b1SNandita Dukkipati } 22706ba8a3b1SNandita Dukkipati 22716ba8a3b1SNandita Dukkipati if (WARN_ON(!skb || !tcp_skb_pcount(skb))) 22726ba8a3b1SNandita Dukkipati goto rearm_timer; 22736ba8a3b1SNandita Dukkipati 22746ba8a3b1SNandita Dukkipati err = __tcp_retransmit_skb(sk, skb); 22756ba8a3b1SNandita Dukkipati 22769b717a8dSNandita Dukkipati /* Record snd_nxt for loss detection. */ 22779b717a8dSNandita Dukkipati if (likely(!err)) 22789b717a8dSNandita Dukkipati tp->tlp_high_seq = tp->snd_nxt; 22799b717a8dSNandita Dukkipati 22806ba8a3b1SNandita Dukkipati rearm_timer: 22816ba8a3b1SNandita Dukkipati inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 22826ba8a3b1SNandita Dukkipati inet_csk(sk)->icsk_rto, 22836ba8a3b1SNandita Dukkipati TCP_RTO_MAX); 22846ba8a3b1SNandita Dukkipati 22856ba8a3b1SNandita Dukkipati if (likely(!err)) 22866ba8a3b1SNandita Dukkipati NET_INC_STATS_BH(sock_net(sk), 22876ba8a3b1SNandita Dukkipati LINUX_MIB_TCPLOSSPROBES); 22881da177e4SLinus Torvalds } 22891da177e4SLinus Torvalds 2290a762a980SDavid S. Miller /* Push out any pending frames which were held back due to 2291a762a980SDavid S. Miller * TCP_CORK or attempt at coalescing tiny packets. 2292a762a980SDavid S. Miller * The socket must be locked by the caller. 2293a762a980SDavid S. Miller */ 22949e412ba7SIlpo Järvinen void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 22959e412ba7SIlpo Järvinen int nonagle) 2296a762a980SDavid S. Miller { 2297726e07a8SIlpo Järvinen /* If we are closed, the bytes will have to remain here. 2298726e07a8SIlpo Järvinen * In time closedown will finish, we empty the write queue and 2299726e07a8SIlpo Järvinen * all will be happy. 2300726e07a8SIlpo Järvinen */ 2301726e07a8SIlpo Järvinen if (unlikely(sk->sk_state == TCP_CLOSE)) 2302726e07a8SIlpo Järvinen return; 2303726e07a8SIlpo Järvinen 230499a1dec7SMel Gorman if (tcp_write_xmit(sk, cur_mss, nonagle, 0, 230599a1dec7SMel Gorman sk_gfp_atomic(sk, GFP_ATOMIC))) 23069e412ba7SIlpo Järvinen tcp_check_probe_timer(sk); 2307a762a980SDavid S. Miller } 2308a762a980SDavid S. Miller 2309c1b4a7e6SDavid S. Miller /* Send _single_ skb sitting at the send head. This function requires 2310c1b4a7e6SDavid S. Miller * true push pending frames to setup probe timer etc. 2311c1b4a7e6SDavid S. Miller */ 2312c1b4a7e6SDavid S. Miller void tcp_push_one(struct sock *sk, unsigned int mss_now) 2313c1b4a7e6SDavid S. Miller { 2314fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 2315c1b4a7e6SDavid S. Miller 2316c1b4a7e6SDavid S. Miller BUG_ON(!skb || skb->len < mss_now); 2317c1b4a7e6SDavid S. Miller 2318d5dd9175SIlpo Järvinen tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); 2319c1b4a7e6SDavid S. Miller } 2320c1b4a7e6SDavid S. Miller 23211da177e4SLinus Torvalds /* This function returns the amount that we can raise the 23221da177e4SLinus Torvalds * usable window based on the following constraints 23231da177e4SLinus Torvalds * 23241da177e4SLinus Torvalds * 1. The window can never be shrunk once it is offered (RFC 793) 23251da177e4SLinus Torvalds * 2. We limit memory per socket 23261da177e4SLinus Torvalds * 23271da177e4SLinus Torvalds * RFC 1122: 23281da177e4SLinus Torvalds * "the suggested [SWS] avoidance algorithm for the receiver is to keep 23291da177e4SLinus Torvalds * RECV.NEXT + RCV.WIN fixed until: 23301da177e4SLinus Torvalds * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 23311da177e4SLinus Torvalds * 23321da177e4SLinus Torvalds * i.e. don't raise the right edge of the window until you can raise 23331da177e4SLinus Torvalds * it at least MSS bytes. 23341da177e4SLinus Torvalds * 23351da177e4SLinus Torvalds * Unfortunately, the recommended algorithm breaks header prediction, 23361da177e4SLinus Torvalds * since header prediction assumes th->window stays fixed. 23371da177e4SLinus Torvalds * 23381da177e4SLinus Torvalds * Strictly speaking, keeping th->window fixed violates the receiver 23391da177e4SLinus Torvalds * side SWS prevention criteria. The problem is that under this rule 23401da177e4SLinus Torvalds * a stream of single byte packets will cause the right side of the 23411da177e4SLinus Torvalds * window to always advance by a single byte. 23421da177e4SLinus Torvalds * 23431da177e4SLinus Torvalds * Of course, if the sender implements sender side SWS prevention 23441da177e4SLinus Torvalds * then this will not be a problem. 23451da177e4SLinus Torvalds * 23461da177e4SLinus Torvalds * BSD seems to make the following compromise: 23471da177e4SLinus Torvalds * 23481da177e4SLinus Torvalds * If the free space is less than the 1/4 of the maximum 23491da177e4SLinus Torvalds * space available and the free space is less than 1/2 mss, 23501da177e4SLinus Torvalds * then set the window to 0. 23511da177e4SLinus Torvalds * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 23521da177e4SLinus Torvalds * Otherwise, just prevent the window from shrinking 23531da177e4SLinus Torvalds * and from being larger than the largest representable value. 23541da177e4SLinus Torvalds * 23551da177e4SLinus Torvalds * This prevents incremental opening of the window in the regime 23561da177e4SLinus Torvalds * where TCP is limited by the speed of the reader side taking 23571da177e4SLinus Torvalds * data out of the TCP receive queue. It does nothing about 23581da177e4SLinus Torvalds * those cases where the window is constrained on the sender side 23591da177e4SLinus Torvalds * because the pipeline is full. 23601da177e4SLinus Torvalds * 23611da177e4SLinus Torvalds * BSD also seems to "accidentally" limit itself to windows that are a 23621da177e4SLinus Torvalds * multiple of MSS, at least until the free space gets quite small. 23631da177e4SLinus Torvalds * This would appear to be a side effect of the mbuf implementation. 23641da177e4SLinus Torvalds * Combining these two algorithms results in the observed behavior 23651da177e4SLinus Torvalds * of having a fixed window size at almost all times. 23661da177e4SLinus Torvalds * 23671da177e4SLinus Torvalds * Below we obtain similar behavior by forcing the offered window to 23681da177e4SLinus Torvalds * a multiple of the mss when it is feasible to do so. 23691da177e4SLinus Torvalds * 23701da177e4SLinus Torvalds * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 23711da177e4SLinus Torvalds * Regular options like TIMESTAMP are taken into account. 23721da177e4SLinus Torvalds */ 23731da177e4SLinus Torvalds u32 __tcp_select_window(struct sock *sk) 23741da177e4SLinus Torvalds { 2375463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 23761da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2377caa20d9aSStephen Hemminger /* MSS for the peer's data. Previous versions used mss_clamp 23781da177e4SLinus Torvalds * here. I don't know if the value based on our guesses 23791da177e4SLinus Torvalds * of peer's MSS is better for the performance. It's more correct 23801da177e4SLinus Torvalds * but may be worse for the performance because of rcv_mss 23811da177e4SLinus Torvalds * fluctuations. --SAW 1998/11/1 23821da177e4SLinus Torvalds */ 2383463c84b9SArnaldo Carvalho de Melo int mss = icsk->icsk_ack.rcv_mss; 23841da177e4SLinus Torvalds int free_space = tcp_space(sk); 238586c1a045SFlorian Westphal int allowed_space = tcp_full_space(sk); 238686c1a045SFlorian Westphal int full_space = min_t(int, tp->window_clamp, allowed_space); 23871da177e4SLinus Torvalds int window; 23881da177e4SLinus Torvalds 23891da177e4SLinus Torvalds if (mss > full_space) 23901da177e4SLinus Torvalds mss = full_space; 23911da177e4SLinus Torvalds 2392b92edbe0SEric Dumazet if (free_space < (full_space >> 1)) { 2393463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.quick = 0; 23941da177e4SLinus Torvalds 2395180d8cd9SGlauber Costa if (sk_under_memory_pressure(sk)) 2396056834d9SIlpo Järvinen tp->rcv_ssthresh = min(tp->rcv_ssthresh, 2397056834d9SIlpo Järvinen 4U * tp->advmss); 23981da177e4SLinus Torvalds 239986c1a045SFlorian Westphal /* free_space might become our new window, make sure we don't 240086c1a045SFlorian Westphal * increase it due to wscale. 240186c1a045SFlorian Westphal */ 240286c1a045SFlorian Westphal free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale); 240386c1a045SFlorian Westphal 240486c1a045SFlorian Westphal /* if free space is less than mss estimate, or is below 1/16th 240586c1a045SFlorian Westphal * of the maximum allowed, try to move to zero-window, else 240686c1a045SFlorian Westphal * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and 240786c1a045SFlorian Westphal * new incoming data is dropped due to memory limits. 240886c1a045SFlorian Westphal * With large window, mss test triggers way too late in order 240986c1a045SFlorian Westphal * to announce zero window in time before rmem limit kicks in. 241086c1a045SFlorian Westphal */ 241186c1a045SFlorian Westphal if (free_space < (allowed_space >> 4) || free_space < mss) 24121da177e4SLinus Torvalds return 0; 24131da177e4SLinus Torvalds } 24141da177e4SLinus Torvalds 24151da177e4SLinus Torvalds if (free_space > tp->rcv_ssthresh) 24161da177e4SLinus Torvalds free_space = tp->rcv_ssthresh; 24171da177e4SLinus Torvalds 24181da177e4SLinus Torvalds /* Don't do rounding if we are using window scaling, since the 24191da177e4SLinus Torvalds * scaled window will not line up with the MSS boundary anyway. 24201da177e4SLinus Torvalds */ 24211da177e4SLinus Torvalds window = tp->rcv_wnd; 24221da177e4SLinus Torvalds if (tp->rx_opt.rcv_wscale) { 24231da177e4SLinus Torvalds window = free_space; 24241da177e4SLinus Torvalds 24251da177e4SLinus Torvalds /* Advertise enough space so that it won't get scaled away. 24261da177e4SLinus Torvalds * Import case: prevent zero window announcement if 24271da177e4SLinus Torvalds * 1<<rcv_wscale > mss. 24281da177e4SLinus Torvalds */ 24291da177e4SLinus Torvalds if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window) 24301da177e4SLinus Torvalds window = (((window >> tp->rx_opt.rcv_wscale) + 1) 24311da177e4SLinus Torvalds << tp->rx_opt.rcv_wscale); 24321da177e4SLinus Torvalds } else { 24331da177e4SLinus Torvalds /* Get the largest window that is a nice multiple of mss. 24341da177e4SLinus Torvalds * Window clamp already applied above. 24351da177e4SLinus Torvalds * If our current window offering is within 1 mss of the 24361da177e4SLinus Torvalds * free space we just keep it. This prevents the divide 24371da177e4SLinus Torvalds * and multiply from happening most of the time. 24381da177e4SLinus Torvalds * We also don't do any window rounding when the free space 24391da177e4SLinus Torvalds * is too small. 24401da177e4SLinus Torvalds */ 24411da177e4SLinus Torvalds if (window <= free_space - mss || window > free_space) 24421da177e4SLinus Torvalds window = (free_space / mss) * mss; 244384565070SJohn Heffner else if (mss == full_space && 2444b92edbe0SEric Dumazet free_space > window + (full_space >> 1)) 244584565070SJohn Heffner window = free_space; 24461da177e4SLinus Torvalds } 24471da177e4SLinus Torvalds 24481da177e4SLinus Torvalds return window; 24491da177e4SLinus Torvalds } 24501da177e4SLinus Torvalds 24514a17fc3aSIlpo Järvinen /* Collapses two adjacent SKB's during retransmission. */ 24524a17fc3aSIlpo Järvinen static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) 24531da177e4SLinus Torvalds { 24541da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2455fe067e8aSDavid S. Miller struct sk_buff *next_skb = tcp_write_queue_next(sk, skb); 2456058dc334SIlpo Järvinen int skb_size, next_skb_size; 24571da177e4SLinus Torvalds 2458058dc334SIlpo Järvinen skb_size = skb->len; 2459058dc334SIlpo Järvinen next_skb_size = next_skb->len; 24601da177e4SLinus Torvalds 2461058dc334SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); 24621da177e4SLinus Torvalds 24636859d494SIlpo Järvinen tcp_highest_sack_combine(sk, next_skb, skb); 2464a6963a6bSIlpo Järvinen 2465fe067e8aSDavid S. Miller tcp_unlink_write_queue(next_skb, sk); 24661da177e4SLinus Torvalds 2467058dc334SIlpo Järvinen skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size), 24681a4e2d09SArnaldo Carvalho de Melo next_skb_size); 24691da177e4SLinus Torvalds 247052d570aaSJarek Poplawski if (next_skb->ip_summed == CHECKSUM_PARTIAL) 247152d570aaSJarek Poplawski skb->ip_summed = CHECKSUM_PARTIAL; 24721da177e4SLinus Torvalds 247384fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 24741da177e4SLinus Torvalds skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); 24751da177e4SLinus Torvalds 24761da177e4SLinus Torvalds /* Update sequence range on original skb. */ 24771da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 24781da177e4SLinus Torvalds 2479e6c7d085SIlpo Järvinen /* Merge over control information. This moves PSH/FIN etc. over */ 24804de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; 24811da177e4SLinus Torvalds 24821da177e4SLinus Torvalds /* All done, get rid of second SKB and account for it so 24831da177e4SLinus Torvalds * packet counting does not break. 24841da177e4SLinus Torvalds */ 24854828e7f4SIlpo Järvinen TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; 2486b7689205SIlpo Järvinen 2487b7689205SIlpo Järvinen /* changed transmit queue under us so clear hints */ 2488ef9da47cSIlpo Järvinen tcp_clear_retrans_hints_partial(tp); 2489ef9da47cSIlpo Järvinen if (next_skb == tp->retransmit_skb_hint) 2490ef9da47cSIlpo Järvinen tp->retransmit_skb_hint = skb; 2491b7689205SIlpo Järvinen 2492797108d1SIlpo Järvinen tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); 2493797108d1SIlpo Järvinen 24943ab224beSHideo Aoki sk_wmem_free_skb(sk, next_skb); 24951da177e4SLinus Torvalds } 24961da177e4SLinus Torvalds 249767edfef7SAndi Kleen /* Check if coalescing SKBs is legal. */ 2498a2a385d6SEric Dumazet static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) 24994a17fc3aSIlpo Järvinen { 25004a17fc3aSIlpo Järvinen if (tcp_skb_pcount(skb) > 1) 2501a2a385d6SEric Dumazet return false; 25024a17fc3aSIlpo Järvinen /* TODO: SACK collapsing could be used to remove this condition */ 25034a17fc3aSIlpo Järvinen if (skb_shinfo(skb)->nr_frags != 0) 2504a2a385d6SEric Dumazet return false; 25054a17fc3aSIlpo Järvinen if (skb_cloned(skb)) 2506a2a385d6SEric Dumazet return false; 25074a17fc3aSIlpo Järvinen if (skb == tcp_send_head(sk)) 2508a2a385d6SEric Dumazet return false; 25094a17fc3aSIlpo Järvinen /* Some heurestics for collapsing over SACK'd could be invented */ 25104a17fc3aSIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 2511a2a385d6SEric Dumazet return false; 25124a17fc3aSIlpo Järvinen 2513a2a385d6SEric Dumazet return true; 25144a17fc3aSIlpo Järvinen } 25154a17fc3aSIlpo Järvinen 251667edfef7SAndi Kleen /* Collapse packets in the retransmit queue to make to create 251767edfef7SAndi Kleen * less packets on the wire. This is only done on retransmission. 251867edfef7SAndi Kleen */ 25194a17fc3aSIlpo Järvinen static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, 25204a17fc3aSIlpo Järvinen int space) 25214a17fc3aSIlpo Järvinen { 25224a17fc3aSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 25234a17fc3aSIlpo Järvinen struct sk_buff *skb = to, *tmp; 2524a2a385d6SEric Dumazet bool first = true; 25254a17fc3aSIlpo Järvinen 25264a17fc3aSIlpo Järvinen if (!sysctl_tcp_retrans_collapse) 25274a17fc3aSIlpo Järvinen return; 25284de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 25294a17fc3aSIlpo Järvinen return; 25304a17fc3aSIlpo Järvinen 25314a17fc3aSIlpo Järvinen tcp_for_write_queue_from_safe(skb, tmp, sk) { 25324a17fc3aSIlpo Järvinen if (!tcp_can_collapse(sk, skb)) 25334a17fc3aSIlpo Järvinen break; 25344a17fc3aSIlpo Järvinen 25354a17fc3aSIlpo Järvinen space -= skb->len; 25364a17fc3aSIlpo Järvinen 25374a17fc3aSIlpo Järvinen if (first) { 2538a2a385d6SEric Dumazet first = false; 25394a17fc3aSIlpo Järvinen continue; 25404a17fc3aSIlpo Järvinen } 25414a17fc3aSIlpo Järvinen 25424a17fc3aSIlpo Järvinen if (space < 0) 25434a17fc3aSIlpo Järvinen break; 25444a17fc3aSIlpo Järvinen /* Punt if not enough space exists in the first SKB for 25454a17fc3aSIlpo Järvinen * the data in the second 25464a17fc3aSIlpo Järvinen */ 2547a21d4572SEric Dumazet if (skb->len > skb_availroom(to)) 25484a17fc3aSIlpo Järvinen break; 25494a17fc3aSIlpo Järvinen 25504a17fc3aSIlpo Järvinen if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) 25514a17fc3aSIlpo Järvinen break; 25524a17fc3aSIlpo Järvinen 25534a17fc3aSIlpo Järvinen tcp_collapse_retrans(sk, to); 25544a17fc3aSIlpo Järvinen } 25554a17fc3aSIlpo Järvinen } 25564a17fc3aSIlpo Järvinen 25571da177e4SLinus Torvalds /* This retransmits one SKB. Policy decisions and retransmit queue 25581da177e4SLinus Torvalds * state updates are done by the caller. Returns non-zero if an 25591da177e4SLinus Torvalds * error occurred which prevented the send. 25601da177e4SLinus Torvalds */ 256193b174adSYuchung Cheng int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) 25621da177e4SLinus Torvalds { 25631da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 25645d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 25657d227cd2SSridhar Samudrala unsigned int cur_mss; 2566c84a5711SYuchung Cheng int err; 25671da177e4SLinus Torvalds 25685d424d5aSJohn Heffner /* Inconslusive MTU probe */ 25695d424d5aSJohn Heffner if (icsk->icsk_mtup.probe_size) { 25705d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 25715d424d5aSJohn Heffner } 25725d424d5aSJohn Heffner 25731da177e4SLinus Torvalds /* Do not sent more than we queued. 1/4 is reserved for possible 2574caa20d9aSStephen Hemminger * copying overhead: fragmentation, tunneling, mangling etc. 25751da177e4SLinus Torvalds */ 25761da177e4SLinus Torvalds if (atomic_read(&sk->sk_wmem_alloc) > 25771da177e4SLinus Torvalds min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf)) 25781da177e4SLinus Torvalds return -EAGAIN; 25791da177e4SLinus Torvalds 25801f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 25811f3279aeSEric Dumazet return -EBUSY; 25821f3279aeSEric Dumazet 25831da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 25841da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 25851da177e4SLinus Torvalds BUG(); 25861da177e4SLinus Torvalds if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 25871da177e4SLinus Torvalds return -ENOMEM; 25881da177e4SLinus Torvalds } 25891da177e4SLinus Torvalds 25907d227cd2SSridhar Samudrala if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 25917d227cd2SSridhar Samudrala return -EHOSTUNREACH; /* Routing failure or similar. */ 25927d227cd2SSridhar Samudrala 25930c54b85fSIlpo Järvinen cur_mss = tcp_current_mss(sk); 25947d227cd2SSridhar Samudrala 25951da177e4SLinus Torvalds /* If receiver has shrunk his window, and skb is out of 25961da177e4SLinus Torvalds * new window, do not retransmit it. The exception is the 25971da177e4SLinus Torvalds * case, when window is shrunk to zero. In this case 25981da177e4SLinus Torvalds * our retransmit serves as a zero window probe. 25991da177e4SLinus Torvalds */ 26009d4fb27dSJoe Perches if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && 26019d4fb27dSJoe Perches TCP_SKB_CB(skb)->seq != tp->snd_una) 26021da177e4SLinus Torvalds return -EAGAIN; 26031da177e4SLinus Torvalds 26041da177e4SLinus Torvalds if (skb->len > cur_mss) { 26056cc55e09SOctavian Purdila if (tcp_fragment(sk, skb, cur_mss, cur_mss, GFP_ATOMIC)) 26061da177e4SLinus Torvalds return -ENOMEM; /* We'll try again later. */ 260702276f3cSIlpo Järvinen } else { 26089eb9362eSIlpo Järvinen int oldpcount = tcp_skb_pcount(skb); 26099eb9362eSIlpo Järvinen 26109eb9362eSIlpo Järvinen if (unlikely(oldpcount > 1)) { 2611c52e2421SEric Dumazet if (skb_unclone(skb, GFP_ATOMIC)) 2612c52e2421SEric Dumazet return -ENOMEM; 261302276f3cSIlpo Järvinen tcp_init_tso_segs(sk, skb, cur_mss); 26149eb9362eSIlpo Järvinen tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb)); 26159eb9362eSIlpo Järvinen } 26161da177e4SLinus Torvalds } 26171da177e4SLinus Torvalds 26181da177e4SLinus Torvalds tcp_retrans_try_collapse(sk, skb, cur_mss); 26191da177e4SLinus Torvalds 26201da177e4SLinus Torvalds /* Make a copy, if the first transmission SKB clone we made 26211da177e4SLinus Torvalds * is still in somebody's hands, else make a clone. 26221da177e4SLinus Torvalds */ 26231da177e4SLinus Torvalds 262450bceae9SThomas Graf /* make sure skb->data is aligned on arches that require it 262550bceae9SThomas Graf * and check if ack-trimming & collapsing extended the headroom 262650bceae9SThomas Graf * beyond what csum_start can cover. 262750bceae9SThomas Graf */ 262850bceae9SThomas Graf if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) || 262950bceae9SThomas Graf skb_headroom(skb) >= 0xFFFF)) { 2630117632e6SEric Dumazet struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER, 2631117632e6SEric Dumazet GFP_ATOMIC); 2632c84a5711SYuchung Cheng err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) : 2633117632e6SEric Dumazet -ENOBUFS; 2634117632e6SEric Dumazet } else { 2635c84a5711SYuchung Cheng err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 2636117632e6SEric Dumazet } 2637c84a5711SYuchung Cheng 2638fc9f3501SEric Dumazet if (likely(!err)) { 2639c84a5711SYuchung Cheng TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS; 2640fc9f3501SEric Dumazet /* Update global TCP statistics. */ 2641fc9f3501SEric Dumazet TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); 2642fc9f3501SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 2643fc9f3501SEric Dumazet NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 2644fc9f3501SEric Dumazet tp->total_retrans++; 2645fc9f3501SEric Dumazet } 2646c84a5711SYuchung Cheng return err; 264793b174adSYuchung Cheng } 264893b174adSYuchung Cheng 264993b174adSYuchung Cheng int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) 265093b174adSYuchung Cheng { 265193b174adSYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 265293b174adSYuchung Cheng int err = __tcp_retransmit_skb(sk, skb); 26531da177e4SLinus Torvalds 26541da177e4SLinus Torvalds if (err == 0) { 26551da177e4SLinus Torvalds #if FASTRETRANS_DEBUG > 0 26561da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 2657e87cc472SJoe Perches net_dbg_ratelimited("retrans_out leaked\n"); 26581da177e4SLinus Torvalds } 26591da177e4SLinus Torvalds #endif 2660b08d6cb2SIlpo Järvinen if (!tp->retrans_out) 2661b08d6cb2SIlpo Järvinen tp->lost_retrans_low = tp->snd_nxt; 26621da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 26631da177e4SLinus Torvalds tp->retrans_out += tcp_skb_pcount(skb); 26641da177e4SLinus Torvalds 26651da177e4SLinus Torvalds /* Save stamp of the first retransmit. */ 26661da177e4SLinus Torvalds if (!tp->retrans_stamp) 26677faee5c0SEric Dumazet tp->retrans_stamp = tcp_skb_timestamp(skb); 26681da177e4SLinus Torvalds 26691da177e4SLinus Torvalds /* snd_nxt is stored to detect loss of retransmitted segment, 26701da177e4SLinus Torvalds * see tcp_input.c tcp_sacktag_write_queue(). 26711da177e4SLinus Torvalds */ 26721da177e4SLinus Torvalds TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt; 26731f3279aeSEric Dumazet } else if (err != -EBUSY) { 267424ab6becSYuchung Cheng NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL); 26751da177e4SLinus Torvalds } 26766e08d5e3SYuchung Cheng 26776e08d5e3SYuchung Cheng if (tp->undo_retrans < 0) 26786e08d5e3SYuchung Cheng tp->undo_retrans = 0; 26796e08d5e3SYuchung Cheng tp->undo_retrans += tcp_skb_pcount(skb); 26801da177e4SLinus Torvalds return err; 26811da177e4SLinus Torvalds } 26821da177e4SLinus Torvalds 268367edfef7SAndi Kleen /* Check if we forward retransmits are possible in the current 268467edfef7SAndi Kleen * window/congestion state. 268567edfef7SAndi Kleen */ 2686a2a385d6SEric Dumazet static bool tcp_can_forward_retransmit(struct sock *sk) 2687b5afe7bcSIlpo Järvinen { 2688b5afe7bcSIlpo Järvinen const struct inet_connection_sock *icsk = inet_csk(sk); 2689cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 2690b5afe7bcSIlpo Järvinen 2691b5afe7bcSIlpo Järvinen /* Forward retransmissions are possible only during Recovery. */ 2692b5afe7bcSIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Recovery) 2693a2a385d6SEric Dumazet return false; 2694b5afe7bcSIlpo Järvinen 2695b5afe7bcSIlpo Järvinen /* No forward retransmissions in Reno are possible. */ 2696b5afe7bcSIlpo Järvinen if (tcp_is_reno(tp)) 2697a2a385d6SEric Dumazet return false; 2698b5afe7bcSIlpo Järvinen 2699b5afe7bcSIlpo Järvinen /* Yeah, we have to make difficult choice between forward transmission 2700b5afe7bcSIlpo Järvinen * and retransmission... Both ways have their merits... 2701b5afe7bcSIlpo Järvinen * 2702b5afe7bcSIlpo Järvinen * For now we do not retransmit anything, while we have some new 2703b5afe7bcSIlpo Järvinen * segments to send. In the other cases, follow rule 3 for 2704b5afe7bcSIlpo Järvinen * NextSeg() specified in RFC3517. 2705b5afe7bcSIlpo Järvinen */ 2706b5afe7bcSIlpo Järvinen 2707b5afe7bcSIlpo Järvinen if (tcp_may_send_now(sk)) 2708a2a385d6SEric Dumazet return false; 2709b5afe7bcSIlpo Järvinen 2710a2a385d6SEric Dumazet return true; 2711b5afe7bcSIlpo Järvinen } 2712b5afe7bcSIlpo Järvinen 27131da177e4SLinus Torvalds /* This gets called after a retransmit timeout, and the initially 27141da177e4SLinus Torvalds * retransmitted data is acknowledged. It tries to continue 27151da177e4SLinus Torvalds * resending the rest of the retransmit queue, until either 27161da177e4SLinus Torvalds * we've sent it all or the congestion window limit is reached. 27171da177e4SLinus Torvalds * If doing SACK, the first ACK which comes back for a timeout 27181da177e4SLinus Torvalds * based retransmit packet might feed us FACK information again. 27191da177e4SLinus Torvalds * If so, we use it to avoid unnecessarily retransmissions. 27201da177e4SLinus Torvalds */ 27211da177e4SLinus Torvalds void tcp_xmit_retransmit_queue(struct sock *sk) 27221da177e4SLinus Torvalds { 27236687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 27241da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 27251da177e4SLinus Torvalds struct sk_buff *skb; 27260e1c54c2SIlpo Järvinen struct sk_buff *hole = NULL; 2727618d9f25SIlpo Järvinen u32 last_lost; 272861eb55f4SIlpo Järvinen int mib_idx; 27290e1c54c2SIlpo Järvinen int fwd_rexmitting = 0; 27306a438bbeSStephen Hemminger 273145e77d31SIlpo Järvinen if (!tp->packets_out) 273245e77d31SIlpo Järvinen return; 273345e77d31SIlpo Järvinen 273408ebd172SIlpo Järvinen if (!tp->lost_out) 273508ebd172SIlpo Järvinen tp->retransmit_high = tp->snd_una; 273608ebd172SIlpo Järvinen 2737618d9f25SIlpo Järvinen if (tp->retransmit_skb_hint) { 27386a438bbeSStephen Hemminger skb = tp->retransmit_skb_hint; 2739618d9f25SIlpo Järvinen last_lost = TCP_SKB_CB(skb)->end_seq; 2740618d9f25SIlpo Järvinen if (after(last_lost, tp->retransmit_high)) 2741618d9f25SIlpo Järvinen last_lost = tp->retransmit_high; 2742618d9f25SIlpo Järvinen } else { 2743fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 2744618d9f25SIlpo Järvinen last_lost = tp->snd_una; 2745618d9f25SIlpo Järvinen } 27461da177e4SLinus Torvalds 2747fe067e8aSDavid S. Miller tcp_for_write_queue_from(skb, sk) { 27481da177e4SLinus Torvalds __u8 sacked = TCP_SKB_CB(skb)->sacked; 27491da177e4SLinus Torvalds 2750fe067e8aSDavid S. Miller if (skb == tcp_send_head(sk)) 2751fe067e8aSDavid S. Miller break; 27526a438bbeSStephen Hemminger /* we could do better than to assign each time */ 275351456b29SIan Morris if (!hole) 27546a438bbeSStephen Hemminger tp->retransmit_skb_hint = skb; 27556a438bbeSStephen Hemminger 27561da177e4SLinus Torvalds /* Assume this retransmit will generate 27571da177e4SLinus Torvalds * only one packet for congestion window 27581da177e4SLinus Torvalds * calculation purposes. This works because 27591da177e4SLinus Torvalds * tcp_retransmit_skb() will chop up the 27601da177e4SLinus Torvalds * packet to be MSS sized and all the 27611da177e4SLinus Torvalds * packet counting works out. 27621da177e4SLinus Torvalds */ 27631da177e4SLinus Torvalds if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) 27641da177e4SLinus Torvalds return; 27650e1c54c2SIlpo Järvinen 27660e1c54c2SIlpo Järvinen if (fwd_rexmitting) { 27670e1c54c2SIlpo Järvinen begin_fwd: 27680e1c54c2SIlpo Järvinen if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp))) 2769006f582cSIlpo Järvinen break; 27700e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFORWARDRETRANS; 27710e1c54c2SIlpo Järvinen 27720e1c54c2SIlpo Järvinen } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) { 2773618d9f25SIlpo Järvinen tp->retransmit_high = last_lost; 27740e1c54c2SIlpo Järvinen if (!tcp_can_forward_retransmit(sk)) 27750e1c54c2SIlpo Järvinen break; 27760e1c54c2SIlpo Järvinen /* Backtrack if necessary to non-L'ed skb */ 277700db4124SIan Morris if (hole) { 27780e1c54c2SIlpo Järvinen skb = hole; 27790e1c54c2SIlpo Järvinen hole = NULL; 27800e1c54c2SIlpo Järvinen } 27810e1c54c2SIlpo Järvinen fwd_rexmitting = 1; 27820e1c54c2SIlpo Järvinen goto begin_fwd; 27830e1c54c2SIlpo Järvinen 27840e1c54c2SIlpo Järvinen } else if (!(sacked & TCPCB_LOST)) { 278551456b29SIan Morris if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) 27860e1c54c2SIlpo Järvinen hole = skb; 278761eb55f4SIlpo Järvinen continue; 27881da177e4SLinus Torvalds 27890e1c54c2SIlpo Järvinen } else { 2790618d9f25SIlpo Järvinen last_lost = TCP_SKB_CB(skb)->end_seq; 27910e1c54c2SIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Loss) 27920e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFASTRETRANS; 27930e1c54c2SIlpo Järvinen else 27940e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; 27950e1c54c2SIlpo Järvinen } 27960e1c54c2SIlpo Järvinen 27970e1c54c2SIlpo Järvinen if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) 279861eb55f4SIlpo Järvinen continue; 279940b215e5SPavel Emelyanov 280024ab6becSYuchung Cheng if (tcp_retransmit_skb(sk, skb)) 28011da177e4SLinus Torvalds return; 280224ab6becSYuchung Cheng 2803de0744afSPavel Emelyanov NET_INC_STATS_BH(sock_net(sk), mib_idx); 28041da177e4SLinus Torvalds 2805684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 2806a262f0cdSNandita Dukkipati tp->prr_out += tcp_skb_pcount(skb); 2807a262f0cdSNandita Dukkipati 2808fe067e8aSDavid S. Miller if (skb == tcp_write_queue_head(sk)) 2809463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 28103f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, 28113f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 28121da177e4SLinus Torvalds } 28131da177e4SLinus Torvalds } 28141da177e4SLinus Torvalds 28151da177e4SLinus Torvalds /* Send a fin. The caller locks the socket for us. This cannot be 28161da177e4SLinus Torvalds * allowed to fail queueing a FIN frame under any circumstances. 28171da177e4SLinus Torvalds */ 28181da177e4SLinus Torvalds void tcp_send_fin(struct sock *sk) 28191da177e4SLinus Torvalds { 28201da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2821fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_write_queue_tail(sk); 28221da177e4SLinus Torvalds int mss_now; 28231da177e4SLinus Torvalds 28241da177e4SLinus Torvalds /* Optimization, tack on the FIN if we have a queue of 28251da177e4SLinus Torvalds * unsent frames. But be careful about outgoing SACKS 28261da177e4SLinus Torvalds * and IP options. 28271da177e4SLinus Torvalds */ 28280c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 28291da177e4SLinus Torvalds 283000db4124SIan Morris if (tcp_send_head(sk)) { 28314de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_FIN; 28321da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq++; 28331da177e4SLinus Torvalds tp->write_seq++; 28341da177e4SLinus Torvalds } else { 28351da177e4SLinus Torvalds /* Socket is locked, keep trying until memory is available. */ 28361da177e4SLinus Torvalds for (;;) { 2837d22e1537SJosh Hunt skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation); 28381da177e4SLinus Torvalds if (skb) 28391da177e4SLinus Torvalds break; 28401da177e4SLinus Torvalds yield(); 28411da177e4SLinus Torvalds } 28421da177e4SLinus Torvalds /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 2843e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tp->write_seq, 2844a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_FIN); 28451da177e4SLinus Torvalds tcp_queue_skb(sk, skb); 28461da177e4SLinus Torvalds } 28479e412ba7SIlpo Järvinen __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF); 28481da177e4SLinus Torvalds } 28491da177e4SLinus Torvalds 28501da177e4SLinus Torvalds /* We get here when a process closes a file descriptor (either due to 28511da177e4SLinus Torvalds * an explicit close() or as a byproduct of exit()'ing) and there 28521da177e4SLinus Torvalds * was unread data in the receive queue. This behavior is recommended 285365bb723cSGerrit Renker * by RFC 2525, section 2.17. -DaveM 28541da177e4SLinus Torvalds */ 2855dd0fc66fSAl Viro void tcp_send_active_reset(struct sock *sk, gfp_t priority) 28561da177e4SLinus Torvalds { 28571da177e4SLinus Torvalds struct sk_buff *skb; 28581da177e4SLinus Torvalds 28591da177e4SLinus Torvalds /* NOTE: No TCP options attached and we never retransmit this. */ 28601da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, priority); 28611da177e4SLinus Torvalds if (!skb) { 28624e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 28631da177e4SLinus Torvalds return; 28641da177e4SLinus Torvalds } 28651da177e4SLinus Torvalds 28661da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 28671da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 2868e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), 2869a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_RST); 28701da177e4SLinus Torvalds /* Send it off. */ 2871dfb4b9dcSDavid S. Miller if (tcp_transmit_skb(sk, skb, 0, priority)) 28724e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 287326af65cbSSridhar Samudrala 287481cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); 28751da177e4SLinus Torvalds } 28761da177e4SLinus Torvalds 287767edfef7SAndi Kleen /* Send a crossed SYN-ACK during socket establishment. 287867edfef7SAndi Kleen * WARNING: This routine must only be called when we have already sent 28791da177e4SLinus Torvalds * a SYN packet that crossed the incoming SYN that caused this routine 28801da177e4SLinus Torvalds * to get called. If this assumption fails then the initial rcv_wnd 28811da177e4SLinus Torvalds * and rcv_wscale values will not be correct. 28821da177e4SLinus Torvalds */ 28831da177e4SLinus Torvalds int tcp_send_synack(struct sock *sk) 28841da177e4SLinus Torvalds { 28851da177e4SLinus Torvalds struct sk_buff *skb; 28861da177e4SLinus Torvalds 2887fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 288851456b29SIan Morris if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 288991df42beSJoe Perches pr_debug("%s: wrong queue state\n", __func__); 28901da177e4SLinus Torvalds return -EFAULT; 28911da177e4SLinus Torvalds } 28924de075e0SEric Dumazet if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { 28931da177e4SLinus Torvalds if (skb_cloned(skb)) { 28941da177e4SLinus Torvalds struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC); 289551456b29SIan Morris if (!nskb) 28961da177e4SLinus Torvalds return -ENOMEM; 2897fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 2898f4a775d1SEric Dumazet __skb_header_release(nskb); 2899fe067e8aSDavid S. Miller __tcp_add_write_queue_head(sk, nskb); 29003ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 29013ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 29023ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 29031da177e4SLinus Torvalds skb = nskb; 29041da177e4SLinus Torvalds } 29051da177e4SLinus Torvalds 29064de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; 2907735d3831SFlorian Westphal tcp_ecn_send_synack(sk, skb); 29081da177e4SLinus Torvalds } 2909dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 29101da177e4SLinus Torvalds } 29111da177e4SLinus Torvalds 29124aea39c1SEric Dumazet /** 29134aea39c1SEric Dumazet * tcp_make_synack - Prepare a SYN-ACK. 29144aea39c1SEric Dumazet * sk: listener socket 29154aea39c1SEric Dumazet * dst: dst entry attached to the SYNACK 29164aea39c1SEric Dumazet * req: request_sock pointer 29174aea39c1SEric Dumazet * 29184aea39c1SEric Dumazet * Allocate one skb and build a SYNACK packet. 29194aea39c1SEric Dumazet * @dst is consumed : Caller should not use it again. 29204aea39c1SEric Dumazet */ 29211da177e4SLinus Torvalds struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst, 2922e6b4d113SWilliam Allen Simpson struct request_sock *req, 29238336886fSJerry Chu struct tcp_fastopen_cookie *foc) 29241da177e4SLinus Torvalds { 2925bd0388aeSWilliam Allen Simpson struct tcp_out_options opts; 29262e6599cbSArnaldo Carvalho de Melo struct inet_request_sock *ireq = inet_rsk(req); 29271da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 29281da177e4SLinus Torvalds struct tcphdr *th; 29291da177e4SLinus Torvalds struct sk_buff *skb; 293080f03e27SEric Dumazet struct tcp_md5sig_key *md5 = NULL; 2931bd0388aeSWilliam Allen Simpson int tcp_header_size; 2932f5fff5dcSTom Quetchenbach int mss; 29331da177e4SLinus Torvalds 2934a0b8486cSEric Dumazet skb = sock_wmalloc(sk, MAX_TCP_HEADER, 1, GFP_ATOMIC); 29354aea39c1SEric Dumazet if (unlikely(!skb)) { 29364aea39c1SEric Dumazet dst_release(dst); 29371da177e4SLinus Torvalds return NULL; 29384aea39c1SEric Dumazet } 29391da177e4SLinus Torvalds /* Reserve space for headers. */ 29401da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 29411da177e4SLinus Torvalds 29424aea39c1SEric Dumazet skb_dst_set(skb, dst); 29431da177e4SLinus Torvalds 29440dbaee3bSDavid S. Miller mss = dst_metric_advmss(dst); 2945f5fff5dcSTom Quetchenbach if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss) 2946f5fff5dcSTom Quetchenbach mss = tp->rx_opt.user_mss; 2947f5fff5dcSTom Quetchenbach 294833ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 29498b5f12d0SFlorian Westphal #ifdef CONFIG_SYN_COOKIES 29508b5f12d0SFlorian Westphal if (unlikely(req->cookie_ts)) 29517faee5c0SEric Dumazet skb->skb_mstamp.stamp_jiffies = cookie_init_timestamp(req); 29528b5f12d0SFlorian Westphal else 29538b5f12d0SFlorian Westphal #endif 29547faee5c0SEric Dumazet skb_mstamp_get(&skb->skb_mstamp); 295580f03e27SEric Dumazet 295680f03e27SEric Dumazet #ifdef CONFIG_TCP_MD5SIG 295780f03e27SEric Dumazet rcu_read_lock(); 2958fd3a154aSEric Dumazet md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req)); 295980f03e27SEric Dumazet #endif 296080f03e27SEric Dumazet tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5, 29611a2c6181SChristoph Paasch foc) + sizeof(*th); 296233ad798cSAdam Langley 2963aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 2964aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 29651da177e4SLinus Torvalds 2966aa8223c7SArnaldo Carvalho de Melo th = tcp_hdr(skb); 29671da177e4SLinus Torvalds memset(th, 0, sizeof(struct tcphdr)); 29681da177e4SLinus Torvalds th->syn = 1; 29691da177e4SLinus Torvalds th->ack = 1; 2970735d3831SFlorian Westphal tcp_ecn_make_synack(req, th, sk); 2971b44084c2SEric Dumazet th->source = htons(ireq->ir_num); 2972634fb979SEric Dumazet th->dest = ireq->ir_rmt_port; 2973e870a8efSIlpo Järvinen /* Setting of flags are superfluous here for callers (and ECE is 2974e870a8efSIlpo Järvinen * not even correctly set) 2975e870a8efSIlpo Järvinen */ 2976e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn, 2977a3433f35SChangli Gao TCPHDR_SYN | TCPHDR_ACK); 29784957faadSWilliam Allen Simpson 29791da177e4SLinus Torvalds th->seq = htonl(TCP_SKB_CB(skb)->seq); 29808336886fSJerry Chu /* XXX data is queued and acked as is. No buffer/window check */ 29818336886fSJerry Chu th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt); 29821da177e4SLinus Torvalds 29831da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 2984600ff0c2SIlpo Järvinen th->window = htons(min(req->rcv_wnd, 65535U)); 2985bd0388aeSWilliam Allen Simpson tcp_options_write((__be32 *)(th + 1), tp, &opts); 29861da177e4SLinus Torvalds th->doff = (tcp_header_size >> 2); 2987a0b8486cSEric Dumazet TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_OUTSEGS); 2988cfb6eeb4SYOSHIFUJI Hideaki 2989cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 2990cfb6eeb4SYOSHIFUJI Hideaki /* Okay, we have all we need - do the md5 hash if needed */ 299180f03e27SEric Dumazet if (md5) 2992bd0388aeSWilliam Allen Simpson tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, 299339f8e58eSEric Dumazet md5, req_to_sk(req), skb); 299480f03e27SEric Dumazet rcu_read_unlock(); 2995cfb6eeb4SYOSHIFUJI Hideaki #endif 2996cfb6eeb4SYOSHIFUJI Hideaki 29971da177e4SLinus Torvalds return skb; 29981da177e4SLinus Torvalds } 29994bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_make_synack); 30001da177e4SLinus Torvalds 300181164413SDaniel Borkmann static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst) 300281164413SDaniel Borkmann { 300381164413SDaniel Borkmann struct inet_connection_sock *icsk = inet_csk(sk); 300481164413SDaniel Borkmann const struct tcp_congestion_ops *ca; 300581164413SDaniel Borkmann u32 ca_key = dst_metric(dst, RTAX_CC_ALGO); 300681164413SDaniel Borkmann 300781164413SDaniel Borkmann if (ca_key == TCP_CA_UNSPEC) 300881164413SDaniel Borkmann return; 300981164413SDaniel Borkmann 301081164413SDaniel Borkmann rcu_read_lock(); 301181164413SDaniel Borkmann ca = tcp_ca_find_key(ca_key); 301281164413SDaniel Borkmann if (likely(ca && try_module_get(ca->owner))) { 301381164413SDaniel Borkmann module_put(icsk->icsk_ca_ops->owner); 301481164413SDaniel Borkmann icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst); 301581164413SDaniel Borkmann icsk->icsk_ca_ops = ca; 301681164413SDaniel Borkmann } 301781164413SDaniel Borkmann rcu_read_unlock(); 301881164413SDaniel Borkmann } 301981164413SDaniel Borkmann 302067edfef7SAndi Kleen /* Do all connect socket setups that can be done AF independent. */ 3021f7e56a76Sstephen hemminger static void tcp_connect_init(struct sock *sk) 30221da177e4SLinus Torvalds { 3023cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 30241da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 30251da177e4SLinus Torvalds __u8 rcv_wscale; 30261da177e4SLinus Torvalds 30271da177e4SLinus Torvalds /* We'll fix this up when we get a response from the other end. 30281da177e4SLinus Torvalds * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 30291da177e4SLinus Torvalds */ 30301da177e4SLinus Torvalds tp->tcp_header_len = sizeof(struct tcphdr) + 3031bb5b7c11SDavid S. Miller (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0); 30321da177e4SLinus Torvalds 3033cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 303400db4124SIan Morris if (tp->af_specific->md5_lookup(sk, sk)) 3035cfb6eeb4SYOSHIFUJI Hideaki tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 3036cfb6eeb4SYOSHIFUJI Hideaki #endif 3037cfb6eeb4SYOSHIFUJI Hideaki 30381da177e4SLinus Torvalds /* If user gave his TCP_MAXSEG, record it to clamp */ 30391da177e4SLinus Torvalds if (tp->rx_opt.user_mss) 30401da177e4SLinus Torvalds tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 30411da177e4SLinus Torvalds tp->max_window = 0; 30425d424d5aSJohn Heffner tcp_mtup_init(sk); 30431da177e4SLinus Torvalds tcp_sync_mss(sk, dst_mtu(dst)); 30441da177e4SLinus Torvalds 304581164413SDaniel Borkmann tcp_ca_dst_init(sk, dst); 304681164413SDaniel Borkmann 30471da177e4SLinus Torvalds if (!tp->window_clamp) 30481da177e4SLinus Torvalds tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 30490dbaee3bSDavid S. Miller tp->advmss = dst_metric_advmss(dst); 3050f5fff5dcSTom Quetchenbach if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss) 3051f5fff5dcSTom Quetchenbach tp->advmss = tp->rx_opt.user_mss; 3052f5fff5dcSTom Quetchenbach 30531da177e4SLinus Torvalds tcp_initialize_rcv_mss(sk); 30541da177e4SLinus Torvalds 3055e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 3056e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 3057e88c64f0SHagen Paul Pfeifer (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) 3058e88c64f0SHagen Paul Pfeifer tp->window_clamp = tcp_full_space(sk); 3059e88c64f0SHagen Paul Pfeifer 30601da177e4SLinus Torvalds tcp_select_initial_window(tcp_full_space(sk), 30611da177e4SLinus Torvalds tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 30621da177e4SLinus Torvalds &tp->rcv_wnd, 30631da177e4SLinus Torvalds &tp->window_clamp, 3064bb5b7c11SDavid S. Miller sysctl_tcp_window_scaling, 306531d12926Slaurent chavey &rcv_wscale, 306631d12926Slaurent chavey dst_metric(dst, RTAX_INITRWND)); 30671da177e4SLinus Torvalds 30681da177e4SLinus Torvalds tp->rx_opt.rcv_wscale = rcv_wscale; 30691da177e4SLinus Torvalds tp->rcv_ssthresh = tp->rcv_wnd; 30701da177e4SLinus Torvalds 30711da177e4SLinus Torvalds sk->sk_err = 0; 30721da177e4SLinus Torvalds sock_reset_flag(sk, SOCK_DONE); 30731da177e4SLinus Torvalds tp->snd_wnd = 0; 3074ee7537b6SHantzis Fotis tcp_init_wl(tp, 0); 30751da177e4SLinus Torvalds tp->snd_una = tp->write_seq; 30761da177e4SLinus Torvalds tp->snd_sml = tp->write_seq; 307733f5f57eSIlpo Järvinen tp->snd_up = tp->write_seq; 3078370816aeSPavel Emelyanov tp->snd_nxt = tp->write_seq; 3079ee995283SPavel Emelyanov 3080ee995283SPavel Emelyanov if (likely(!tp->repair)) 30811da177e4SLinus Torvalds tp->rcv_nxt = 0; 3082c7781a6eSAndrew Vagin else 3083c7781a6eSAndrew Vagin tp->rcv_tstamp = tcp_time_stamp; 3084ee995283SPavel Emelyanov tp->rcv_wup = tp->rcv_nxt; 3085ee995283SPavel Emelyanov tp->copied_seq = tp->rcv_nxt; 30861da177e4SLinus Torvalds 3087463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; 3088463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_retransmits = 0; 30891da177e4SLinus Torvalds tcp_clear_retrans(tp); 30901da177e4SLinus Torvalds } 30911da177e4SLinus Torvalds 3092783237e8SYuchung Cheng static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) 3093783237e8SYuchung Cheng { 3094783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3095783237e8SYuchung Cheng struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 3096783237e8SYuchung Cheng 3097783237e8SYuchung Cheng tcb->end_seq += skb->len; 3098f4a775d1SEric Dumazet __skb_header_release(skb); 3099783237e8SYuchung Cheng __tcp_add_write_queue_tail(sk, skb); 3100783237e8SYuchung Cheng sk->sk_wmem_queued += skb->truesize; 3101783237e8SYuchung Cheng sk_mem_charge(sk, skb->truesize); 3102783237e8SYuchung Cheng tp->write_seq = tcb->end_seq; 3103783237e8SYuchung Cheng tp->packets_out += tcp_skb_pcount(skb); 3104783237e8SYuchung Cheng } 3105783237e8SYuchung Cheng 3106783237e8SYuchung Cheng /* Build and send a SYN with data and (cached) Fast Open cookie. However, 3107783237e8SYuchung Cheng * queue a data-only packet after the regular SYN, such that regular SYNs 3108783237e8SYuchung Cheng * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges 3109783237e8SYuchung Cheng * only the SYN sequence, the data are retransmitted in the first ACK. 3110783237e8SYuchung Cheng * If cookie is not cached or other error occurs, falls back to send a 3111783237e8SYuchung Cheng * regular SYN with Fast Open cookie request option. 3112783237e8SYuchung Cheng */ 3113783237e8SYuchung Cheng static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) 3114783237e8SYuchung Cheng { 3115783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3116783237e8SYuchung Cheng struct tcp_fastopen_request *fo = tp->fastopen_req; 311757be5bdaSAl Viro int syn_loss = 0, space, err = 0, copied; 3118aab48743SYuchung Cheng unsigned long last_syn_loss = 0; 3119355a901eSEric Dumazet struct sk_buff *syn_data; 3120783237e8SYuchung Cheng 312167da22d2SYuchung Cheng tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ 3122aab48743SYuchung Cheng tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie, 3123aab48743SYuchung Cheng &syn_loss, &last_syn_loss); 3124aab48743SYuchung Cheng /* Recurring FO SYN losses: revert to regular handshake temporarily */ 3125aab48743SYuchung Cheng if (syn_loss > 1 && 3126aab48743SYuchung Cheng time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) { 3127aab48743SYuchung Cheng fo->cookie.len = -1; 3128aab48743SYuchung Cheng goto fallback; 3129aab48743SYuchung Cheng } 3130aab48743SYuchung Cheng 313167da22d2SYuchung Cheng if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE) 313267da22d2SYuchung Cheng fo->cookie.len = -1; 313367da22d2SYuchung Cheng else if (fo->cookie.len <= 0) 3134783237e8SYuchung Cheng goto fallback; 3135783237e8SYuchung Cheng 3136783237e8SYuchung Cheng /* MSS for SYN-data is based on cached MSS and bounded by PMTU and 3137783237e8SYuchung Cheng * user-MSS. Reserve maximum option space for middleboxes that add 3138783237e8SYuchung Cheng * private TCP options. The cost is reduced data space in SYN :( 3139783237e8SYuchung Cheng */ 3140783237e8SYuchung Cheng if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp) 3141783237e8SYuchung Cheng tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 31421b63edd6SYuchung Cheng space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) - 3143783237e8SYuchung Cheng MAX_TCP_OPTION_SPACE; 3144783237e8SYuchung Cheng 3145f5ddcbbbSEric Dumazet space = min_t(size_t, space, fo->size); 3146f5ddcbbbSEric Dumazet 3147f5ddcbbbSEric Dumazet /* limit to order-0 allocations */ 3148f5ddcbbbSEric Dumazet space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER)); 3149f5ddcbbbSEric Dumazet 3150355a901eSEric Dumazet syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation); 3151355a901eSEric Dumazet if (!syn_data) 3152783237e8SYuchung Cheng goto fallback; 3153355a901eSEric Dumazet syn_data->ip_summed = CHECKSUM_PARTIAL; 3154355a901eSEric Dumazet memcpy(syn_data->cb, syn->cb, sizeof(syn->cb)); 315557be5bdaSAl Viro copied = copy_from_iter(skb_put(syn_data, space), space, 315657be5bdaSAl Viro &fo->data->msg_iter); 315757be5bdaSAl Viro if (unlikely(!copied)) { 3158355a901eSEric Dumazet kfree_skb(syn_data); 3159783237e8SYuchung Cheng goto fallback; 3160783237e8SYuchung Cheng } 316157be5bdaSAl Viro if (copied != space) { 316257be5bdaSAl Viro skb_trim(syn_data, copied); 316357be5bdaSAl Viro space = copied; 316457be5bdaSAl Viro } 3165783237e8SYuchung Cheng 3166355a901eSEric Dumazet /* No more data pending in inet_wait_for_connect() */ 3167355a901eSEric Dumazet if (space == fo->size) 3168355a901eSEric Dumazet fo->data = NULL; 3169355a901eSEric Dumazet fo->copied = space; 3170783237e8SYuchung Cheng 3171355a901eSEric Dumazet tcp_connect_queue_skb(sk, syn_data); 3172355a901eSEric Dumazet 3173355a901eSEric Dumazet err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation); 3174355a901eSEric Dumazet 3175355a901eSEric Dumazet syn->skb_mstamp = syn_data->skb_mstamp; 3176355a901eSEric Dumazet 3177355a901eSEric Dumazet /* Now full SYN+DATA was cloned and sent (or not), 3178355a901eSEric Dumazet * remove the SYN from the original skb (syn_data) 3179355a901eSEric Dumazet * we keep in write queue in case of a retransmit, as we 3180355a901eSEric Dumazet * also have the SYN packet (with no data) in the same queue. 3181431a9124SEric Dumazet */ 3182355a901eSEric Dumazet TCP_SKB_CB(syn_data)->seq++; 3183355a901eSEric Dumazet TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH; 3184355a901eSEric Dumazet if (!err) { 318567da22d2SYuchung Cheng tp->syn_data = (fo->copied > 0); 3186f19c29e3SYuchung Cheng NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT); 3187783237e8SYuchung Cheng goto done; 3188783237e8SYuchung Cheng } 3189783237e8SYuchung Cheng 3190783237e8SYuchung Cheng fallback: 3191783237e8SYuchung Cheng /* Send a regular SYN with Fast Open cookie request option */ 3192783237e8SYuchung Cheng if (fo->cookie.len > 0) 3193783237e8SYuchung Cheng fo->cookie.len = 0; 3194783237e8SYuchung Cheng err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); 3195783237e8SYuchung Cheng if (err) 3196783237e8SYuchung Cheng tp->syn_fastopen = 0; 3197783237e8SYuchung Cheng done: 3198783237e8SYuchung Cheng fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ 3199783237e8SYuchung Cheng return err; 3200783237e8SYuchung Cheng } 3201783237e8SYuchung Cheng 320267edfef7SAndi Kleen /* Build a SYN and send it off. */ 32031da177e4SLinus Torvalds int tcp_connect(struct sock *sk) 32041da177e4SLinus Torvalds { 32051da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 32061da177e4SLinus Torvalds struct sk_buff *buff; 3207ee586811SEric Paris int err; 32081da177e4SLinus Torvalds 32091da177e4SLinus Torvalds tcp_connect_init(sk); 32101da177e4SLinus Torvalds 32112b916477SAndrey Vagin if (unlikely(tp->repair)) { 32122b916477SAndrey Vagin tcp_finish_connect(sk, NULL); 32132b916477SAndrey Vagin return 0; 32142b916477SAndrey Vagin } 32152b916477SAndrey Vagin 3216355a901eSEric Dumazet buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation); 3217355a901eSEric Dumazet if (unlikely(!buff)) 32181da177e4SLinus Torvalds return -ENOBUFS; 32191da177e4SLinus Torvalds 3220a3433f35SChangli Gao tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); 32217faee5c0SEric Dumazet tp->retrans_stamp = tcp_time_stamp; 3222783237e8SYuchung Cheng tcp_connect_queue_skb(sk, buff); 3223735d3831SFlorian Westphal tcp_ecn_send_syn(sk, buff); 32241da177e4SLinus Torvalds 3225783237e8SYuchung Cheng /* Send off SYN; include data in Fast Open. */ 3226783237e8SYuchung Cheng err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : 3227783237e8SYuchung Cheng tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); 3228ee586811SEric Paris if (err == -ECONNREFUSED) 3229ee586811SEric Paris return err; 3230bd37a088SWei Yongjun 3231bd37a088SWei Yongjun /* We change tp->snd_nxt after the tcp_transmit_skb() call 3232bd37a088SWei Yongjun * in order to make this packet get counted in tcpOutSegs. 3233bd37a088SWei Yongjun */ 3234bd37a088SWei Yongjun tp->snd_nxt = tp->write_seq; 3235bd37a088SWei Yongjun tp->pushed_seq = tp->write_seq; 323681cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); 32371da177e4SLinus Torvalds 32381da177e4SLinus Torvalds /* Timer for repeating the SYN until an answer. */ 32393f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 32403f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 32411da177e4SLinus Torvalds return 0; 32421da177e4SLinus Torvalds } 32434bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_connect); 32441da177e4SLinus Torvalds 32451da177e4SLinus Torvalds /* Send out a delayed ack, the caller does the policy checking 32461da177e4SLinus Torvalds * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 32471da177e4SLinus Torvalds * for details. 32481da177e4SLinus Torvalds */ 32491da177e4SLinus Torvalds void tcp_send_delayed_ack(struct sock *sk) 32501da177e4SLinus Torvalds { 3251463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 3252463c84b9SArnaldo Carvalho de Melo int ato = icsk->icsk_ack.ato; 32531da177e4SLinus Torvalds unsigned long timeout; 32541da177e4SLinus Torvalds 32559890092eSFlorian Westphal tcp_ca_event(sk, CA_EVENT_DELAYED_ACK); 32569890092eSFlorian Westphal 32571da177e4SLinus Torvalds if (ato > TCP_DELACK_MIN) { 3258463c84b9SArnaldo Carvalho de Melo const struct tcp_sock *tp = tcp_sk(sk); 32591da177e4SLinus Torvalds int max_ato = HZ / 2; 32601da177e4SLinus Torvalds 3261056834d9SIlpo Järvinen if (icsk->icsk_ack.pingpong || 3262056834d9SIlpo Järvinen (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 32631da177e4SLinus Torvalds max_ato = TCP_DELACK_MAX; 32641da177e4SLinus Torvalds 32651da177e4SLinus Torvalds /* Slow path, intersegment interval is "high". */ 32661da177e4SLinus Torvalds 32671da177e4SLinus Torvalds /* If some rtt estimate is known, use it to bound delayed ack. 3268463c84b9SArnaldo Carvalho de Melo * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 32691da177e4SLinus Torvalds * directly. 32701da177e4SLinus Torvalds */ 3271740b0f18SEric Dumazet if (tp->srtt_us) { 3272740b0f18SEric Dumazet int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3), 3273740b0f18SEric Dumazet TCP_DELACK_MIN); 32741da177e4SLinus Torvalds 32751da177e4SLinus Torvalds if (rtt < max_ato) 32761da177e4SLinus Torvalds max_ato = rtt; 32771da177e4SLinus Torvalds } 32781da177e4SLinus Torvalds 32791da177e4SLinus Torvalds ato = min(ato, max_ato); 32801da177e4SLinus Torvalds } 32811da177e4SLinus Torvalds 32821da177e4SLinus Torvalds /* Stay within the limit we were given */ 32831da177e4SLinus Torvalds timeout = jiffies + ato; 32841da177e4SLinus Torvalds 32851da177e4SLinus Torvalds /* Use new timeout only if there wasn't a older one earlier. */ 3286463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 32871da177e4SLinus Torvalds /* If delack timer was blocked or is about to expire, 32881da177e4SLinus Torvalds * send ACK now. 32891da177e4SLinus Torvalds */ 3290463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.blocked || 3291463c84b9SArnaldo Carvalho de Melo time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 32921da177e4SLinus Torvalds tcp_send_ack(sk); 32931da177e4SLinus Torvalds return; 32941da177e4SLinus Torvalds } 32951da177e4SLinus Torvalds 3296463c84b9SArnaldo Carvalho de Melo if (!time_before(timeout, icsk->icsk_ack.timeout)) 3297463c84b9SArnaldo Carvalho de Melo timeout = icsk->icsk_ack.timeout; 32981da177e4SLinus Torvalds } 3299463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 3300463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.timeout = timeout; 3301463c84b9SArnaldo Carvalho de Melo sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 33021da177e4SLinus Torvalds } 33031da177e4SLinus Torvalds 33041da177e4SLinus Torvalds /* This routine sends an ack and also updates the window. */ 33051da177e4SLinus Torvalds void tcp_send_ack(struct sock *sk) 33061da177e4SLinus Torvalds { 33071da177e4SLinus Torvalds struct sk_buff *buff; 33081da177e4SLinus Torvalds 3309058dc334SIlpo Järvinen /* If we have been reset, we may not send again. */ 3310058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3311058dc334SIlpo Järvinen return; 3312058dc334SIlpo Järvinen 33139890092eSFlorian Westphal tcp_ca_event(sk, CA_EVENT_NON_DELAYED_ACK); 33149890092eSFlorian Westphal 33151da177e4SLinus Torvalds /* We are not putting this on the write queue, so 33161da177e4SLinus Torvalds * tcp_transmit_skb() will set the ownership to this 33171da177e4SLinus Torvalds * sock. 33181da177e4SLinus Torvalds */ 331999a1dec7SMel Gorman buff = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC)); 332051456b29SIan Morris if (!buff) { 3321463c84b9SArnaldo Carvalho de Melo inet_csk_schedule_ack(sk); 3322463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; 33233f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 33243f421baaSArnaldo Carvalho de Melo TCP_DELACK_MAX, TCP_RTO_MAX); 33251da177e4SLinus Torvalds return; 33261da177e4SLinus Torvalds } 33271da177e4SLinus Torvalds 33281da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 33291da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 3330a3433f35SChangli Gao tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); 33311da177e4SLinus Torvalds 333298781965SEric Dumazet /* We do not want pure acks influencing TCP Small Queues or fq/pacing 333398781965SEric Dumazet * too much. 333498781965SEric Dumazet * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784 333598781965SEric Dumazet * We also avoid tcp_wfree() overhead (cache line miss accessing 333698781965SEric Dumazet * tp->tsq_flags) by using regular sock_wfree() 333798781965SEric Dumazet */ 333898781965SEric Dumazet skb_set_tcp_pure_ack(buff); 333998781965SEric Dumazet 33401da177e4SLinus Torvalds /* Send it off, this clears delayed acks for us. */ 33417faee5c0SEric Dumazet skb_mstamp_get(&buff->skb_mstamp); 334299a1dec7SMel Gorman tcp_transmit_skb(sk, buff, 0, sk_gfp_atomic(sk, GFP_ATOMIC)); 33431da177e4SLinus Torvalds } 3344e3118e83SDaniel Borkmann EXPORT_SYMBOL_GPL(tcp_send_ack); 33451da177e4SLinus Torvalds 33461da177e4SLinus Torvalds /* This routine sends a packet with an out of date sequence 33471da177e4SLinus Torvalds * number. It assumes the other end will try to ack it. 33481da177e4SLinus Torvalds * 33491da177e4SLinus Torvalds * Question: what should we make while urgent mode? 33501da177e4SLinus Torvalds * 4.4BSD forces sending single byte of data. We cannot send 33511da177e4SLinus Torvalds * out of window data, because we have SND.NXT==SND.MAX... 33521da177e4SLinus Torvalds * 33531da177e4SLinus Torvalds * Current solution: to send TWO zero-length segments in urgent mode: 33541da177e4SLinus Torvalds * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 33551da177e4SLinus Torvalds * out-of-date with SND.UNA-1 to probe window. 33561da177e4SLinus Torvalds */ 33571da177e4SLinus Torvalds static int tcp_xmit_probe_skb(struct sock *sk, int urgent) 33581da177e4SLinus Torvalds { 33591da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 33601da177e4SLinus Torvalds struct sk_buff *skb; 33611da177e4SLinus Torvalds 33621da177e4SLinus Torvalds /* We don't queue it, tcp_transmit_skb() sets ownership. */ 336399a1dec7SMel Gorman skb = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC)); 336451456b29SIan Morris if (!skb) 33651da177e4SLinus Torvalds return -1; 33661da177e4SLinus Torvalds 33671da177e4SLinus Torvalds /* Reserve space for headers and set control bits. */ 33681da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 33691da177e4SLinus Torvalds /* Use a previous sequence. This should cause the other 33701da177e4SLinus Torvalds * end to send an ack. Don't queue or clone SKB, just 33711da177e4SLinus Torvalds * send it. 33721da177e4SLinus Torvalds */ 3373a3433f35SChangli Gao tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); 33747faee5c0SEric Dumazet skb_mstamp_get(&skb->skb_mstamp); 3375dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC); 33761da177e4SLinus Torvalds } 33771da177e4SLinus Torvalds 3378ee995283SPavel Emelyanov void tcp_send_window_probe(struct sock *sk) 3379ee995283SPavel Emelyanov { 3380ee995283SPavel Emelyanov if (sk->sk_state == TCP_ESTABLISHED) { 3381ee995283SPavel Emelyanov tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; 3382ee995283SPavel Emelyanov tcp_xmit_probe_skb(sk, 0); 3383ee995283SPavel Emelyanov } 3384ee995283SPavel Emelyanov } 3385ee995283SPavel Emelyanov 338667edfef7SAndi Kleen /* Initiate keepalive or window probe from timer. */ 33871da177e4SLinus Torvalds int tcp_write_wakeup(struct sock *sk) 33881da177e4SLinus Torvalds { 33891da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 33901da177e4SLinus Torvalds struct sk_buff *skb; 33911da177e4SLinus Torvalds 3392058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3393058dc334SIlpo Järvinen return -1; 3394058dc334SIlpo Järvinen 339500db4124SIan Morris skb = tcp_send_head(sk); 339600db4124SIan Morris if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { 33971da177e4SLinus Torvalds int err; 33980c54b85fSIlpo Järvinen unsigned int mss = tcp_current_mss(sk); 339990840defSIlpo Järvinen unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 34001da177e4SLinus Torvalds 34011da177e4SLinus Torvalds if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 34021da177e4SLinus Torvalds tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 34031da177e4SLinus Torvalds 34041da177e4SLinus Torvalds /* We are probing the opening of a window 34051da177e4SLinus Torvalds * but the window size is != 0 34061da177e4SLinus Torvalds * must have been a result SWS avoidance ( sender ) 34071da177e4SLinus Torvalds */ 34081da177e4SLinus Torvalds if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 34091da177e4SLinus Torvalds skb->len > mss) { 34101da177e4SLinus Torvalds seg_size = min(seg_size, mss); 34114de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 34126cc55e09SOctavian Purdila if (tcp_fragment(sk, skb, seg_size, mss, GFP_ATOMIC)) 34131da177e4SLinus Torvalds return -1; 34141da177e4SLinus Torvalds } else if (!tcp_skb_pcount(skb)) 3415846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss); 34161da177e4SLinus Torvalds 34174de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 3418dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 341966f5fe62SIlpo Järvinen if (!err) 342066f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 34211da177e4SLinus Torvalds return err; 34221da177e4SLinus Torvalds } else { 342333f5f57eSIlpo Järvinen if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) 34244828e7f4SIlpo Järvinen tcp_xmit_probe_skb(sk, 1); 34251da177e4SLinus Torvalds return tcp_xmit_probe_skb(sk, 0); 34261da177e4SLinus Torvalds } 34271da177e4SLinus Torvalds } 34281da177e4SLinus Torvalds 34291da177e4SLinus Torvalds /* A window probe timeout has occurred. If window is not closed send 34301da177e4SLinus Torvalds * a partial packet else a zero probe. 34311da177e4SLinus Torvalds */ 34321da177e4SLinus Torvalds void tcp_send_probe0(struct sock *sk) 34331da177e4SLinus Torvalds { 3434463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 34351da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 3436fcdd1cf4SEric Dumazet unsigned long probe_max; 34371da177e4SLinus Torvalds int err; 34381da177e4SLinus Torvalds 34391da177e4SLinus Torvalds err = tcp_write_wakeup(sk); 34401da177e4SLinus Torvalds 3441fe067e8aSDavid S. Miller if (tp->packets_out || !tcp_send_head(sk)) { 34421da177e4SLinus Torvalds /* Cancel probe timer, if it is not required. */ 34436687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 0; 3444463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff = 0; 34451da177e4SLinus Torvalds return; 34461da177e4SLinus Torvalds } 34471da177e4SLinus Torvalds 34481da177e4SLinus Torvalds if (err <= 0) { 3449463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_backoff < sysctl_tcp_retries2) 3450463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff++; 34516687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out++; 3452fcdd1cf4SEric Dumazet probe_max = TCP_RTO_MAX; 34531da177e4SLinus Torvalds } else { 34541da177e4SLinus Torvalds /* If packet was not sent due to local congestion, 34556687e988SArnaldo Carvalho de Melo * do not backoff and do not remember icsk_probes_out. 34561da177e4SLinus Torvalds * Let local senders to fight for local resources. 34571da177e4SLinus Torvalds * 34581da177e4SLinus Torvalds * Use accumulated backoff yet. 34591da177e4SLinus Torvalds */ 34606687e988SArnaldo Carvalho de Melo if (!icsk->icsk_probes_out) 34616687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 1; 3462fcdd1cf4SEric Dumazet probe_max = TCP_RESOURCE_PROBE_INTERVAL; 34631da177e4SLinus Torvalds } 3464fcdd1cf4SEric Dumazet inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 3465fcdd1cf4SEric Dumazet inet_csk_rto_backoff(icsk, probe_max), 3466fcdd1cf4SEric Dumazet TCP_RTO_MAX); 34671da177e4SLinus Torvalds } 34685db92c99SOctavian Purdila 34695db92c99SOctavian Purdila int tcp_rtx_synack(struct sock *sk, struct request_sock *req) 34705db92c99SOctavian Purdila { 34715db92c99SOctavian Purdila const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific; 34725db92c99SOctavian Purdila struct flowi fl; 34735db92c99SOctavian Purdila int res; 34745db92c99SOctavian Purdila 34755db92c99SOctavian Purdila res = af_ops->send_synack(sk, NULL, &fl, req, 0, NULL); 34765db92c99SOctavian Purdila if (!res) { 34775db92c99SOctavian Purdila TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_RETRANSSEGS); 34785db92c99SOctavian Purdila NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 34795db92c99SOctavian Purdila } 34805db92c99SOctavian Purdila return res; 34815db92c99SOctavian Purdila } 34825db92c99SOctavian Purdila EXPORT_SYMBOL(tcp_rtx_synack); 3483