11da177e4SLinus Torvalds /* 21da177e4SLinus Torvalds * INET An implementation of the TCP/IP protocol suite for the LINUX 31da177e4SLinus Torvalds * operating system. INET is implemented using the BSD Socket 41da177e4SLinus Torvalds * interface as the means of communication with the user level. 51da177e4SLinus Torvalds * 61da177e4SLinus Torvalds * Implementation of the Transmission Control Protocol(TCP). 71da177e4SLinus Torvalds * 802c30a84SJesper Juhl * Authors: Ross Biro 91da177e4SLinus Torvalds * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> 101da177e4SLinus Torvalds * Mark Evans, <evansmp@uhura.aston.ac.uk> 111da177e4SLinus Torvalds * Corey Minyard <wf-rch!minyard@relay.EU.net> 121da177e4SLinus Torvalds * Florian La Roche, <flla@stud.uni-sb.de> 131da177e4SLinus Torvalds * Charles Hedrick, <hedrick@klinzhai.rutgers.edu> 141da177e4SLinus Torvalds * Linus Torvalds, <torvalds@cs.helsinki.fi> 151da177e4SLinus Torvalds * Alan Cox, <gw4pts@gw4pts.ampr.org> 161da177e4SLinus Torvalds * Matthew Dillon, <dillon@apollo.west.oic.com> 171da177e4SLinus Torvalds * Arnt Gulbrandsen, <agulbra@nvg.unit.no> 181da177e4SLinus Torvalds * Jorge Cwik, <jorge@laser.satlink.net> 191da177e4SLinus Torvalds */ 201da177e4SLinus Torvalds 211da177e4SLinus Torvalds /* 221da177e4SLinus Torvalds * Changes: Pedro Roque : Retransmit queue handled by TCP. 231da177e4SLinus Torvalds * : Fragmentation on mtu decrease 241da177e4SLinus Torvalds * : Segment collapse on retransmit 251da177e4SLinus Torvalds * : AF independence 261da177e4SLinus Torvalds * 271da177e4SLinus Torvalds * Linus Torvalds : send_delayed_ack 281da177e4SLinus Torvalds * David S. Miller : Charge memory using the right skb 291da177e4SLinus Torvalds * during syn/ack processing. 301da177e4SLinus Torvalds * David S. Miller : Output engine completely rewritten. 311da177e4SLinus Torvalds * Andrea Arcangeli: SYNACK carry ts_recent in tsecr. 321da177e4SLinus Torvalds * Cacophonix Gaul : draft-minshall-nagle-01 331da177e4SLinus Torvalds * J Hadi Salim : ECN support 341da177e4SLinus Torvalds * 351da177e4SLinus Torvalds */ 361da177e4SLinus Torvalds 3791df42beSJoe Perches #define pr_fmt(fmt) "TCP: " fmt 3891df42beSJoe Perches 391da177e4SLinus Torvalds #include <net/tcp.h> 401da177e4SLinus Torvalds 411da177e4SLinus Torvalds #include <linux/compiler.h> 425a0e3ad6STejun Heo #include <linux/gfp.h> 431da177e4SLinus Torvalds #include <linux/module.h> 441da177e4SLinus Torvalds 451da177e4SLinus Torvalds /* People can turn this off for buggy TCP's found in printers etc. */ 46ab32ea5dSBrian Haley int sysctl_tcp_retrans_collapse __read_mostly = 1; 471da177e4SLinus Torvalds 4815d99e02SRick Jones /* People can turn this on to work with those rare, broken TCPs that 4915d99e02SRick Jones * interpret the window field as a signed quantity. 5015d99e02SRick Jones */ 51ab32ea5dSBrian Haley int sysctl_tcp_workaround_signed_windows __read_mostly = 0; 5215d99e02SRick Jones 53c39c4c6aSWei Liu /* Default TSQ limit of four TSO segments */ 54c39c4c6aSWei Liu int sysctl_tcp_limit_output_bytes __read_mostly = 262144; 5546d3ceabSEric Dumazet 561da177e4SLinus Torvalds /* This limits the percentage of the congestion window which we 571da177e4SLinus Torvalds * will allow a single TSO frame to consume. Building TSO frames 581da177e4SLinus Torvalds * which are too large can cause TCP streams to be bursty. 591da177e4SLinus Torvalds */ 60ab32ea5dSBrian Haley int sysctl_tcp_tso_win_divisor __read_mostly = 3; 611da177e4SLinus Torvalds 6235089bb2SDavid S. Miller /* By default, RFC2861 behavior. */ 63ab32ea5dSBrian Haley int sysctl_tcp_slow_start_after_idle __read_mostly = 1; 6435089bb2SDavid S. Miller 65c9bee3b7SEric Dumazet unsigned int sysctl_tcp_notsent_lowat __read_mostly = UINT_MAX; 66c9bee3b7SEric Dumazet EXPORT_SYMBOL(sysctl_tcp_notsent_lowat); 67c9bee3b7SEric Dumazet 6846d3ceabSEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 6946d3ceabSEric Dumazet int push_one, gfp_t gfp); 70519855c5SWilliam Allen Simpson 7167edfef7SAndi Kleen /* Account for new data that has been sent to the network. */ 72cf533ea5SEric Dumazet static void tcp_event_new_data_sent(struct sock *sk, const struct sk_buff *skb) 736ff03ac3SIlpo Järvinen { 746ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 756ff03ac3SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 7666f5fe62SIlpo Järvinen unsigned int prior_packets = tp->packets_out; 779e412ba7SIlpo Järvinen 78fe067e8aSDavid S. Miller tcp_advance_send_head(sk, skb); 791da177e4SLinus Torvalds tp->snd_nxt = TCP_SKB_CB(skb)->end_seq; 808512430eSIlpo Järvinen 8166f5fe62SIlpo Järvinen tp->packets_out += tcp_skb_pcount(skb); 826ba8a3b1SNandita Dukkipati if (!prior_packets || icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS || 836a5dc9e5SEric Dumazet icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) { 84750ea2baSYuchung Cheng tcp_rearm_rto(sk); 851da177e4SLinus Torvalds } 86f19c29e3SYuchung Cheng 87f7324acdSDavid S. Miller NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT, 88f19c29e3SYuchung Cheng tcp_skb_pcount(skb)); 896a5dc9e5SEric Dumazet } 901da177e4SLinus Torvalds 911da177e4SLinus Torvalds /* SND.NXT, if window was not shrunk. 921da177e4SLinus Torvalds * If window has been shrunk, what should we make? It is not clear at all. 931da177e4SLinus Torvalds * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-( 941da177e4SLinus Torvalds * Anything in between SND.UNA...SND.UNA+SND.WND also can be already 951da177e4SLinus Torvalds * invalid. OK, let's make this for now: 961da177e4SLinus Torvalds */ 97cf533ea5SEric Dumazet static inline __u32 tcp_acceptable_seq(const struct sock *sk) 981da177e4SLinus Torvalds { 99cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1009e412ba7SIlpo Järvinen 10190840defSIlpo Järvinen if (!before(tcp_wnd_end(tp), tp->snd_nxt)) 1021da177e4SLinus Torvalds return tp->snd_nxt; 1031da177e4SLinus Torvalds else 10490840defSIlpo Järvinen return tcp_wnd_end(tp); 1051da177e4SLinus Torvalds } 1061da177e4SLinus Torvalds 1071da177e4SLinus Torvalds /* Calculate mss to advertise in SYN segment. 1081da177e4SLinus Torvalds * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that: 1091da177e4SLinus Torvalds * 1101da177e4SLinus Torvalds * 1. It is independent of path mtu. 1111da177e4SLinus Torvalds * 2. Ideally, it is maximal possible segment size i.e. 65535-40. 1121da177e4SLinus Torvalds * 3. For IPv4 it is reasonable to calculate it from maximal MTU of 1131da177e4SLinus Torvalds * attached devices, because some buggy hosts are confused by 1141da177e4SLinus Torvalds * large MSS. 1151da177e4SLinus Torvalds * 4. We do not make 3, we advertise MSS, calculated from first 1161da177e4SLinus Torvalds * hop device mtu, but allow to raise it to ip_rt_min_advmss. 1171da177e4SLinus Torvalds * This may be overridden via information stored in routing table. 1181da177e4SLinus Torvalds * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible, 1191da177e4SLinus Torvalds * probably even Jumbo". 1201da177e4SLinus Torvalds */ 1211da177e4SLinus Torvalds static __u16 tcp_advertise_mss(struct sock *sk) 1221da177e4SLinus Torvalds { 1231da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 124cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1251da177e4SLinus Torvalds int mss = tp->advmss; 1261da177e4SLinus Torvalds 1270dbaee3bSDavid S. Miller if (dst) { 1280dbaee3bSDavid S. Miller unsigned int metric = dst_metric_advmss(dst); 1290dbaee3bSDavid S. Miller 1300dbaee3bSDavid S. Miller if (metric < mss) { 1310dbaee3bSDavid S. Miller mss = metric; 1321da177e4SLinus Torvalds tp->advmss = mss; 1331da177e4SLinus Torvalds } 1340dbaee3bSDavid S. Miller } 1351da177e4SLinus Torvalds 1361da177e4SLinus Torvalds return (__u16)mss; 1371da177e4SLinus Torvalds } 1381da177e4SLinus Torvalds 1391da177e4SLinus Torvalds /* RFC2861. Reset CWND after idle period longer RTO to "restart window". 1401da177e4SLinus Torvalds * This is the first part of cwnd validation mechanism. */ 141cf533ea5SEric Dumazet static void tcp_cwnd_restart(struct sock *sk, const struct dst_entry *dst) 1421da177e4SLinus Torvalds { 143463c84b9SArnaldo Carvalho de Melo struct tcp_sock *tp = tcp_sk(sk); 1441da177e4SLinus Torvalds s32 delta = tcp_time_stamp - tp->lsndtime; 1451da177e4SLinus Torvalds u32 restart_cwnd = tcp_init_cwnd(tp, dst); 1461da177e4SLinus Torvalds u32 cwnd = tp->snd_cwnd; 1471da177e4SLinus Torvalds 1486687e988SArnaldo Carvalho de Melo tcp_ca_event(sk, CA_EVENT_CWND_RESTART); 1491da177e4SLinus Torvalds 1506687e988SArnaldo Carvalho de Melo tp->snd_ssthresh = tcp_current_ssthresh(sk); 1511da177e4SLinus Torvalds restart_cwnd = min(restart_cwnd, cwnd); 1521da177e4SLinus Torvalds 153463c84b9SArnaldo Carvalho de Melo while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd) 1541da177e4SLinus Torvalds cwnd >>= 1; 1551da177e4SLinus Torvalds tp->snd_cwnd = max(cwnd, restart_cwnd); 1561da177e4SLinus Torvalds tp->snd_cwnd_stamp = tcp_time_stamp; 1571da177e4SLinus Torvalds tp->snd_cwnd_used = 0; 1581da177e4SLinus Torvalds } 1591da177e4SLinus Torvalds 16067edfef7SAndi Kleen /* Congestion state accounting after a packet has been sent. */ 16140efc6faSStephen Hemminger static void tcp_event_data_sent(struct tcp_sock *tp, 162cf533ea5SEric Dumazet struct sock *sk) 1631da177e4SLinus Torvalds { 164463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 165463c84b9SArnaldo Carvalho de Melo const u32 now = tcp_time_stamp; 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 35349213555SDaniel Borkmann static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb) 35449213555SDaniel Borkmann { 35549213555SDaniel Borkmann if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback) 35649213555SDaniel Borkmann /* tp->ecn_flags are cleared at a later point in time when 35749213555SDaniel Borkmann * SYN ACK is ultimatively being received. 35849213555SDaniel Borkmann */ 35949213555SDaniel Borkmann TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR); 36049213555SDaniel Borkmann } 36149213555SDaniel Borkmann 362735d3831SFlorian Westphal static void 363735d3831SFlorian Westphal tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th, 36430e502a3SDaniel Borkmann struct sock *sk) 365bdf1ee5dSIlpo Järvinen { 36630e502a3SDaniel Borkmann if (inet_rsk(req)->ecn_ok) { 367bdf1ee5dSIlpo Järvinen th->ece = 1; 36830e502a3SDaniel Borkmann if (tcp_ca_needs_ecn(sk)) 36930e502a3SDaniel Borkmann INET_ECN_xmit(sk); 37030e502a3SDaniel Borkmann } 371bdf1ee5dSIlpo Järvinen } 372bdf1ee5dSIlpo Järvinen 37367edfef7SAndi Kleen /* Set up ECN state for a packet on a ESTABLISHED socket that is about to 37467edfef7SAndi Kleen * be sent. 37567edfef7SAndi Kleen */ 376735d3831SFlorian Westphal static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb, 377bdf1ee5dSIlpo Järvinen int tcp_header_len) 378bdf1ee5dSIlpo Järvinen { 379bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 380bdf1ee5dSIlpo Järvinen 381bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_OK) { 382bdf1ee5dSIlpo Järvinen /* Not-retransmitted data segment: set ECT and inject CWR. */ 383bdf1ee5dSIlpo Järvinen if (skb->len != tcp_header_len && 384bdf1ee5dSIlpo Järvinen !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { 385bdf1ee5dSIlpo Järvinen INET_ECN_xmit(sk); 386bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) { 387bdf1ee5dSIlpo Järvinen tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; 388bdf1ee5dSIlpo Järvinen tcp_hdr(skb)->cwr = 1; 389bdf1ee5dSIlpo Järvinen skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; 390bdf1ee5dSIlpo Järvinen } 39130e502a3SDaniel Borkmann } else if (!tcp_ca_needs_ecn(sk)) { 392bdf1ee5dSIlpo Järvinen /* ACK or retransmitted segment: clear ECT|CE */ 393bdf1ee5dSIlpo Järvinen INET_ECN_dontxmit(sk); 394bdf1ee5dSIlpo Järvinen } 395bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) 396bdf1ee5dSIlpo Järvinen tcp_hdr(skb)->ece = 1; 397bdf1ee5dSIlpo Järvinen } 398bdf1ee5dSIlpo Järvinen } 399bdf1ee5dSIlpo Järvinen 400e870a8efSIlpo Järvinen /* Constructs common control bits of non-data skb. If SYN/FIN is present, 401e870a8efSIlpo Järvinen * auto increment end seqno. 402e870a8efSIlpo Järvinen */ 403e870a8efSIlpo Järvinen static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags) 404e870a8efSIlpo Järvinen { 4057b7fc97aSEric Dumazet struct skb_shared_info *shinfo = skb_shinfo(skb); 4067b7fc97aSEric Dumazet 4072e8e18efSDavid S. Miller skb->ip_summed = CHECKSUM_PARTIAL; 408e870a8efSIlpo Järvinen skb->csum = 0; 409e870a8efSIlpo Järvinen 4104de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags; 411e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->sacked = 0; 412e870a8efSIlpo Järvinen 413cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, 1); 4147b7fc97aSEric Dumazet shinfo->gso_size = 0; 415e870a8efSIlpo Järvinen 416e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->seq = seq; 417a3433f35SChangli Gao if (flags & (TCPHDR_SYN | TCPHDR_FIN)) 418e870a8efSIlpo Järvinen seq++; 419e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->end_seq = seq; 420e870a8efSIlpo Järvinen } 421e870a8efSIlpo Järvinen 422a2a385d6SEric Dumazet static inline bool tcp_urg_mode(const struct tcp_sock *tp) 42333f5f57eSIlpo Järvinen { 42433f5f57eSIlpo Järvinen return tp->snd_una != tp->snd_up; 42533f5f57eSIlpo Järvinen } 42633f5f57eSIlpo Järvinen 42733ad798cSAdam Langley #define OPTION_SACK_ADVERTISE (1 << 0) 42833ad798cSAdam Langley #define OPTION_TS (1 << 1) 42933ad798cSAdam Langley #define OPTION_MD5 (1 << 2) 43089e95a61SOri Finkelman #define OPTION_WSCALE (1 << 3) 4312100c8d2SYuchung Cheng #define OPTION_FAST_OPEN_COOKIE (1 << 8) 43233ad798cSAdam Langley 43333ad798cSAdam Langley struct tcp_out_options { 4342100c8d2SYuchung Cheng u16 options; /* bit field of OPTION_* */ 4352100c8d2SYuchung Cheng u16 mss; /* 0 to disable */ 43633ad798cSAdam Langley u8 ws; /* window scale, 0 to disable */ 43733ad798cSAdam Langley u8 num_sack_blocks; /* number of SACK blocks to include */ 438bd0388aeSWilliam Allen Simpson u8 hash_size; /* bytes in hash_location */ 439bd0388aeSWilliam Allen Simpson __u8 *hash_location; /* temporary pointer, overloaded */ 4402100c8d2SYuchung Cheng __u32 tsval, tsecr; /* need to include OPTION_TS */ 4412100c8d2SYuchung Cheng struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */ 44233ad798cSAdam Langley }; 44333ad798cSAdam Langley 44467edfef7SAndi Kleen /* Write previously computed TCP options to the packet. 44567edfef7SAndi Kleen * 44667edfef7SAndi Kleen * Beware: Something in the Internet is very sensitive to the ordering of 447fd6149d3SIlpo Järvinen * TCP options, we learned this through the hard way, so be careful here. 448fd6149d3SIlpo Järvinen * Luckily we can at least blame others for their non-compliance but from 4498e3bff96Sstephen hemminger * inter-operability perspective it seems that we're somewhat stuck with 450fd6149d3SIlpo Järvinen * the ordering which we have been using if we want to keep working with 451fd6149d3SIlpo Järvinen * those broken things (not that it currently hurts anybody as there isn't 452fd6149d3SIlpo Järvinen * particular reason why the ordering would need to be changed). 453fd6149d3SIlpo Järvinen * 454fd6149d3SIlpo Järvinen * At least SACK_PERM as the first option is known to lead to a disaster 455fd6149d3SIlpo Järvinen * (but it may well be that other scenarios fail similarly). 456fd6149d3SIlpo Järvinen */ 45733ad798cSAdam Langley static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp, 458bd0388aeSWilliam Allen Simpson struct tcp_out_options *opts) 459bd0388aeSWilliam Allen Simpson { 4602100c8d2SYuchung Cheng u16 options = opts->options; /* mungable copy */ 461bd0388aeSWilliam Allen Simpson 462bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_MD5 & options)) { 4631a2c6181SChristoph Paasch *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | 4641a2c6181SChristoph Paasch (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG); 465bd0388aeSWilliam Allen Simpson /* overload cookie hash location */ 466bd0388aeSWilliam Allen Simpson opts->hash_location = (__u8 *)ptr; 46733ad798cSAdam Langley ptr += 4; 46833ad798cSAdam Langley } 46933ad798cSAdam Langley 470fd6149d3SIlpo Järvinen if (unlikely(opts->mss)) { 471fd6149d3SIlpo Järvinen *ptr++ = htonl((TCPOPT_MSS << 24) | 472fd6149d3SIlpo Järvinen (TCPOLEN_MSS << 16) | 473fd6149d3SIlpo Järvinen opts->mss); 474fd6149d3SIlpo Järvinen } 475fd6149d3SIlpo Järvinen 476bd0388aeSWilliam Allen Simpson if (likely(OPTION_TS & options)) { 477bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 47833ad798cSAdam Langley *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | 47933ad798cSAdam Langley (TCPOLEN_SACK_PERM << 16) | 48033ad798cSAdam Langley (TCPOPT_TIMESTAMP << 8) | 48133ad798cSAdam Langley TCPOLEN_TIMESTAMP); 482bd0388aeSWilliam Allen Simpson options &= ~OPTION_SACK_ADVERTISE; 48333ad798cSAdam Langley } else { 484496c98dfSYOSHIFUJI Hideaki *ptr++ = htonl((TCPOPT_NOP << 24) | 48540efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 48640efc6faSStephen Hemminger (TCPOPT_TIMESTAMP << 8) | 48740efc6faSStephen Hemminger TCPOLEN_TIMESTAMP); 48840efc6faSStephen Hemminger } 48933ad798cSAdam Langley *ptr++ = htonl(opts->tsval); 49033ad798cSAdam Langley *ptr++ = htonl(opts->tsecr); 49133ad798cSAdam Langley } 49233ad798cSAdam Langley 493bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 49433ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 49533ad798cSAdam Langley (TCPOPT_NOP << 16) | 49633ad798cSAdam Langley (TCPOPT_SACK_PERM << 8) | 49733ad798cSAdam Langley TCPOLEN_SACK_PERM); 49833ad798cSAdam Langley } 49933ad798cSAdam Langley 500bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_WSCALE & options)) { 50133ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 50233ad798cSAdam Langley (TCPOPT_WINDOW << 16) | 50333ad798cSAdam Langley (TCPOLEN_WINDOW << 8) | 50433ad798cSAdam Langley opts->ws); 50533ad798cSAdam Langley } 50633ad798cSAdam Langley 50733ad798cSAdam Langley if (unlikely(opts->num_sack_blocks)) { 50833ad798cSAdam Langley struct tcp_sack_block *sp = tp->rx_opt.dsack ? 50933ad798cSAdam Langley tp->duplicate_sack : tp->selective_acks; 51040efc6faSStephen Hemminger int this_sack; 51140efc6faSStephen Hemminger 51240efc6faSStephen Hemminger *ptr++ = htonl((TCPOPT_NOP << 24) | 51340efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 51440efc6faSStephen Hemminger (TCPOPT_SACK << 8) | 51533ad798cSAdam Langley (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * 51640efc6faSStephen Hemminger TCPOLEN_SACK_PERBLOCK))); 5172de979bdSStephen Hemminger 51833ad798cSAdam Langley for (this_sack = 0; this_sack < opts->num_sack_blocks; 51933ad798cSAdam Langley ++this_sack) { 52040efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].start_seq); 52140efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].end_seq); 52240efc6faSStephen Hemminger } 5232de979bdSStephen Hemminger 52440efc6faSStephen Hemminger tp->rx_opt.dsack = 0; 52540efc6faSStephen Hemminger } 5262100c8d2SYuchung Cheng 5272100c8d2SYuchung Cheng if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) { 5282100c8d2SYuchung Cheng struct tcp_fastopen_cookie *foc = opts->fastopen_cookie; 5297f9b838bSDaniel Lee u8 *p = (u8 *)ptr; 5307f9b838bSDaniel Lee u32 len; /* Fast Open option length */ 5312100c8d2SYuchung Cheng 5327f9b838bSDaniel Lee if (foc->exp) { 5337f9b838bSDaniel Lee len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len; 5347f9b838bSDaniel Lee *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) | 5352100c8d2SYuchung Cheng TCPOPT_FASTOPEN_MAGIC); 5367f9b838bSDaniel Lee p += TCPOLEN_EXP_FASTOPEN_BASE; 5377f9b838bSDaniel Lee } else { 5387f9b838bSDaniel Lee len = TCPOLEN_FASTOPEN_BASE + foc->len; 5397f9b838bSDaniel Lee *p++ = TCPOPT_FASTOPEN; 5407f9b838bSDaniel Lee *p++ = len; 5412100c8d2SYuchung Cheng } 5427f9b838bSDaniel Lee 5437f9b838bSDaniel Lee memcpy(p, foc->val, foc->len); 5447f9b838bSDaniel Lee if ((len & 3) == 2) { 5457f9b838bSDaniel Lee p[foc->len] = TCPOPT_NOP; 5467f9b838bSDaniel Lee p[foc->len + 1] = TCPOPT_NOP; 5477f9b838bSDaniel Lee } 5487f9b838bSDaniel Lee ptr += (len + 3) >> 2; 5492100c8d2SYuchung Cheng } 55040efc6faSStephen Hemminger } 55140efc6faSStephen Hemminger 55267edfef7SAndi Kleen /* Compute TCP options for SYN packets. This is not the final 55367edfef7SAndi Kleen * network wire format yet. 55467edfef7SAndi Kleen */ 55595c96174SEric Dumazet static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb, 55633ad798cSAdam Langley struct tcp_out_options *opts, 557cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 558cf533ea5SEric Dumazet { 55933ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 56095c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 561783237e8SYuchung Cheng struct tcp_fastopen_request *fastopen = tp->fastopen_req; 56233ad798cSAdam Langley 563cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 56433ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 56533ad798cSAdam Langley if (*md5) { 56633ad798cSAdam Langley opts->options |= OPTION_MD5; 567bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 568cfb6eeb4SYOSHIFUJI Hideaki } 56933ad798cSAdam Langley #else 57033ad798cSAdam Langley *md5 = NULL; 571cfb6eeb4SYOSHIFUJI Hideaki #endif 57233ad798cSAdam Langley 57333ad798cSAdam Langley /* We always get an MSS option. The option bytes which will be seen in 57433ad798cSAdam Langley * normal data packets should timestamps be used, must be in the MSS 57533ad798cSAdam Langley * advertised. But we subtract them from tp->mss_cache so that 57633ad798cSAdam Langley * calculations in tcp_sendmsg are simpler etc. So account for this 57733ad798cSAdam Langley * fact here if necessary. If we don't do this correctly, as a 57833ad798cSAdam Langley * receiver we won't recognize data packets as being full sized when we 57933ad798cSAdam Langley * should, and thus we won't abide by the delayed ACK rules correctly. 58033ad798cSAdam Langley * SACKs don't matter, we never delay an ACK when we have any of those 58133ad798cSAdam Langley * going out. */ 58233ad798cSAdam Langley opts->mss = tcp_advertise_mss(sk); 583bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 58433ad798cSAdam Langley 58551456b29SIan Morris if (likely(sysctl_tcp_timestamps && !*md5)) { 58633ad798cSAdam Langley opts->options |= OPTION_TS; 5877faee5c0SEric Dumazet opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset; 58833ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 589bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 59033ad798cSAdam Langley } 591bb5b7c11SDavid S. Miller if (likely(sysctl_tcp_window_scaling)) { 59233ad798cSAdam Langley opts->ws = tp->rx_opt.rcv_wscale; 59389e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 594bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 59533ad798cSAdam Langley } 596bb5b7c11SDavid S. Miller if (likely(sysctl_tcp_sack)) { 59733ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 598b32d1310SDavid S. Miller if (unlikely(!(OPTION_TS & opts->options))) 599bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 60033ad798cSAdam Langley } 60133ad798cSAdam Langley 602783237e8SYuchung Cheng if (fastopen && fastopen->cookie.len >= 0) { 6032646c831SDaniel Lee u32 need = fastopen->cookie.len; 6042646c831SDaniel Lee 6052646c831SDaniel Lee need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE : 6062646c831SDaniel Lee TCPOLEN_FASTOPEN_BASE; 607783237e8SYuchung Cheng need = (need + 3) & ~3U; /* Align to 32 bits */ 608783237e8SYuchung Cheng if (remaining >= need) { 609783237e8SYuchung Cheng opts->options |= OPTION_FAST_OPEN_COOKIE; 610783237e8SYuchung Cheng opts->fastopen_cookie = &fastopen->cookie; 611783237e8SYuchung Cheng remaining -= need; 612783237e8SYuchung Cheng tp->syn_fastopen = 1; 6132646c831SDaniel Lee tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0; 614783237e8SYuchung Cheng } 615783237e8SYuchung Cheng } 616bd0388aeSWilliam Allen Simpson 617bd0388aeSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 61833ad798cSAdam Langley } 61933ad798cSAdam Langley 62067edfef7SAndi Kleen /* Set up TCP options for SYN-ACKs. */ 62195c96174SEric Dumazet static unsigned int tcp_synack_options(struct sock *sk, 62233ad798cSAdam Langley struct request_sock *req, 62395c96174SEric Dumazet unsigned int mss, struct sk_buff *skb, 62433ad798cSAdam Langley struct tcp_out_options *opts, 62580f03e27SEric Dumazet const struct tcp_md5sig_key *md5, 6268336886fSJerry Chu struct tcp_fastopen_cookie *foc) 6274957faadSWilliam Allen Simpson { 62833ad798cSAdam Langley struct inet_request_sock *ireq = inet_rsk(req); 62995c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 63033ad798cSAdam Langley 63133ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 63280f03e27SEric Dumazet if (md5) { 63333ad798cSAdam Langley opts->options |= OPTION_MD5; 6344957faadSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 6354957faadSWilliam Allen Simpson 6364957faadSWilliam Allen Simpson /* We can't fit any SACK blocks in a packet with MD5 + TS 6374957faadSWilliam Allen Simpson * options. There was discussion about disabling SACK 6384957faadSWilliam Allen Simpson * rather than TS in order to fit in better with old, 6394957faadSWilliam Allen Simpson * buggy kernels, but that was deemed to be unnecessary. 6404957faadSWilliam Allen Simpson */ 641de213e5eSEric Dumazet ireq->tstamp_ok &= !ireq->sack_ok; 64233ad798cSAdam Langley } 64333ad798cSAdam Langley #endif 64433ad798cSAdam Langley 6454957faadSWilliam Allen Simpson /* We always send an MSS option. */ 64633ad798cSAdam Langley opts->mss = mss; 6474957faadSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 64833ad798cSAdam Langley 64933ad798cSAdam Langley if (likely(ireq->wscale_ok)) { 65033ad798cSAdam Langley opts->ws = ireq->rcv_wscale; 65189e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 6524957faadSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 65333ad798cSAdam Langley } 654de213e5eSEric Dumazet if (likely(ireq->tstamp_ok)) { 65533ad798cSAdam Langley opts->options |= OPTION_TS; 6567faee5c0SEric Dumazet opts->tsval = tcp_skb_timestamp(skb); 65733ad798cSAdam Langley opts->tsecr = req->ts_recent; 6584957faadSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 65933ad798cSAdam Langley } 66033ad798cSAdam Langley if (likely(ireq->sack_ok)) { 66133ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 662de213e5eSEric Dumazet if (unlikely(!ireq->tstamp_ok)) 6634957faadSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 66433ad798cSAdam Langley } 6657f9b838bSDaniel Lee if (foc != NULL && foc->len >= 0) { 6667f9b838bSDaniel Lee u32 need = foc->len; 6677f9b838bSDaniel Lee 6687f9b838bSDaniel Lee need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE : 6697f9b838bSDaniel Lee TCPOLEN_FASTOPEN_BASE; 6708336886fSJerry Chu need = (need + 3) & ~3U; /* Align to 32 bits */ 6718336886fSJerry Chu if (remaining >= need) { 6728336886fSJerry Chu opts->options |= OPTION_FAST_OPEN_COOKIE; 6738336886fSJerry Chu opts->fastopen_cookie = foc; 6748336886fSJerry Chu remaining -= need; 6758336886fSJerry Chu } 6768336886fSJerry Chu } 6774957faadSWilliam Allen Simpson 6784957faadSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 67933ad798cSAdam Langley } 68033ad798cSAdam Langley 68167edfef7SAndi Kleen /* Compute TCP options for ESTABLISHED sockets. This is not the 68267edfef7SAndi Kleen * final wire format yet. 68367edfef7SAndi Kleen */ 68495c96174SEric Dumazet static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb, 68533ad798cSAdam Langley struct tcp_out_options *opts, 686cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 687cf533ea5SEric Dumazet { 68833ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 68995c96174SEric Dumazet unsigned int size = 0; 690cabeccbdSIlpo Järvinen unsigned int eff_sacks; 69133ad798cSAdam Langley 6925843ef42SAndi Kleen opts->options = 0; 6935843ef42SAndi Kleen 69433ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 69533ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 69633ad798cSAdam Langley if (unlikely(*md5)) { 69733ad798cSAdam Langley opts->options |= OPTION_MD5; 69833ad798cSAdam Langley size += TCPOLEN_MD5SIG_ALIGNED; 69933ad798cSAdam Langley } 70033ad798cSAdam Langley #else 70133ad798cSAdam Langley *md5 = NULL; 70233ad798cSAdam Langley #endif 70333ad798cSAdam Langley 70433ad798cSAdam Langley if (likely(tp->rx_opt.tstamp_ok)) { 70533ad798cSAdam Langley opts->options |= OPTION_TS; 7067faee5c0SEric Dumazet opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0; 70733ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 70833ad798cSAdam Langley size += TCPOLEN_TSTAMP_ALIGNED; 70933ad798cSAdam Langley } 71033ad798cSAdam Langley 711cabeccbdSIlpo Järvinen eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; 712cabeccbdSIlpo Järvinen if (unlikely(eff_sacks)) { 71395c96174SEric Dumazet const unsigned int remaining = MAX_TCP_OPTION_SPACE - size; 71433ad798cSAdam Langley opts->num_sack_blocks = 71595c96174SEric Dumazet min_t(unsigned int, eff_sacks, 71633ad798cSAdam Langley (remaining - TCPOLEN_SACK_BASE_ALIGNED) / 71733ad798cSAdam Langley TCPOLEN_SACK_PERBLOCK); 71833ad798cSAdam Langley size += TCPOLEN_SACK_BASE_ALIGNED + 71933ad798cSAdam Langley opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; 72033ad798cSAdam Langley } 72133ad798cSAdam Langley 72233ad798cSAdam Langley return size; 72340efc6faSStephen Hemminger } 7241da177e4SLinus Torvalds 72546d3ceabSEric Dumazet 72646d3ceabSEric Dumazet /* TCP SMALL QUEUES (TSQ) 72746d3ceabSEric Dumazet * 72846d3ceabSEric Dumazet * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev) 72946d3ceabSEric Dumazet * to reduce RTT and bufferbloat. 73046d3ceabSEric Dumazet * We do this using a special skb destructor (tcp_wfree). 73146d3ceabSEric Dumazet * 73246d3ceabSEric Dumazet * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb 73346d3ceabSEric Dumazet * needs to be reallocated in a driver. 7348e3bff96Sstephen hemminger * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc 73546d3ceabSEric Dumazet * 73646d3ceabSEric Dumazet * Since transmit from skb destructor is forbidden, we use a tasklet 73746d3ceabSEric Dumazet * to process all sockets that eventually need to send more skbs. 73846d3ceabSEric Dumazet * We use one tasklet per cpu, with its own queue of sockets. 73946d3ceabSEric Dumazet */ 74046d3ceabSEric Dumazet struct tsq_tasklet { 74146d3ceabSEric Dumazet struct tasklet_struct tasklet; 74246d3ceabSEric Dumazet struct list_head head; /* queue of tcp sockets */ 74346d3ceabSEric Dumazet }; 74446d3ceabSEric Dumazet static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet); 74546d3ceabSEric Dumazet 7466f458dfbSEric Dumazet static void tcp_tsq_handler(struct sock *sk) 7476f458dfbSEric Dumazet { 7486f458dfbSEric Dumazet if ((1 << sk->sk_state) & 7496f458dfbSEric Dumazet (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING | 7506f458dfbSEric Dumazet TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) 751bf06200eSJohn Ogness tcp_write_xmit(sk, tcp_current_mss(sk), tcp_sk(sk)->nonagle, 752bf06200eSJohn Ogness 0, GFP_ATOMIC); 7536f458dfbSEric Dumazet } 75446d3ceabSEric Dumazet /* 7558e3bff96Sstephen hemminger * One tasklet per cpu tries to send more skbs. 75646d3ceabSEric Dumazet * We run in tasklet context but need to disable irqs when 7578e3bff96Sstephen hemminger * transferring tsq->head because tcp_wfree() might 75846d3ceabSEric Dumazet * interrupt us (non NAPI drivers) 75946d3ceabSEric Dumazet */ 76046d3ceabSEric Dumazet static void tcp_tasklet_func(unsigned long data) 76146d3ceabSEric Dumazet { 76246d3ceabSEric Dumazet struct tsq_tasklet *tsq = (struct tsq_tasklet *)data; 76346d3ceabSEric Dumazet LIST_HEAD(list); 76446d3ceabSEric Dumazet unsigned long flags; 76546d3ceabSEric Dumazet struct list_head *q, *n; 76646d3ceabSEric Dumazet struct tcp_sock *tp; 76746d3ceabSEric Dumazet struct sock *sk; 76846d3ceabSEric Dumazet 76946d3ceabSEric Dumazet local_irq_save(flags); 77046d3ceabSEric Dumazet list_splice_init(&tsq->head, &list); 77146d3ceabSEric Dumazet local_irq_restore(flags); 77246d3ceabSEric Dumazet 77346d3ceabSEric Dumazet list_for_each_safe(q, n, &list) { 77446d3ceabSEric Dumazet tp = list_entry(q, struct tcp_sock, tsq_node); 77546d3ceabSEric Dumazet list_del(&tp->tsq_node); 77646d3ceabSEric Dumazet 77746d3ceabSEric Dumazet sk = (struct sock *)tp; 77846d3ceabSEric Dumazet bh_lock_sock(sk); 77946d3ceabSEric Dumazet 78046d3ceabSEric Dumazet if (!sock_owned_by_user(sk)) { 7816f458dfbSEric Dumazet tcp_tsq_handler(sk); 78246d3ceabSEric Dumazet } else { 78346d3ceabSEric Dumazet /* defer the work to tcp_release_cb() */ 7846f458dfbSEric Dumazet set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags); 78546d3ceabSEric Dumazet } 78646d3ceabSEric Dumazet bh_unlock_sock(sk); 78746d3ceabSEric Dumazet 78846d3ceabSEric Dumazet clear_bit(TSQ_QUEUED, &tp->tsq_flags); 78946d3ceabSEric Dumazet sk_free(sk); 79046d3ceabSEric Dumazet } 79146d3ceabSEric Dumazet } 79246d3ceabSEric Dumazet 7936f458dfbSEric Dumazet #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \ 7946f458dfbSEric Dumazet (1UL << TCP_WRITE_TIMER_DEFERRED) | \ 795563d34d0SEric Dumazet (1UL << TCP_DELACK_TIMER_DEFERRED) | \ 796563d34d0SEric Dumazet (1UL << TCP_MTU_REDUCED_DEFERRED)) 79746d3ceabSEric Dumazet /** 79846d3ceabSEric Dumazet * tcp_release_cb - tcp release_sock() callback 79946d3ceabSEric Dumazet * @sk: socket 80046d3ceabSEric Dumazet * 80146d3ceabSEric Dumazet * called from release_sock() to perform protocol dependent 80246d3ceabSEric Dumazet * actions before socket release. 80346d3ceabSEric Dumazet */ 80446d3ceabSEric Dumazet void tcp_release_cb(struct sock *sk) 80546d3ceabSEric Dumazet { 80646d3ceabSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 8076f458dfbSEric Dumazet unsigned long flags, nflags; 80846d3ceabSEric Dumazet 8096f458dfbSEric Dumazet /* perform an atomic operation only if at least one flag is set */ 8106f458dfbSEric Dumazet do { 8116f458dfbSEric Dumazet flags = tp->tsq_flags; 8126f458dfbSEric Dumazet if (!(flags & TCP_DEFERRED_ALL)) 8136f458dfbSEric Dumazet return; 8146f458dfbSEric Dumazet nflags = flags & ~TCP_DEFERRED_ALL; 8156f458dfbSEric Dumazet } while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags); 8166f458dfbSEric Dumazet 8176f458dfbSEric Dumazet if (flags & (1UL << TCP_TSQ_DEFERRED)) 8186f458dfbSEric Dumazet tcp_tsq_handler(sk); 8196f458dfbSEric Dumazet 820c3f9b018SEric Dumazet /* Here begins the tricky part : 821c3f9b018SEric Dumazet * We are called from release_sock() with : 822c3f9b018SEric Dumazet * 1) BH disabled 823c3f9b018SEric Dumazet * 2) sk_lock.slock spinlock held 824c3f9b018SEric Dumazet * 3) socket owned by us (sk->sk_lock.owned == 1) 825c3f9b018SEric Dumazet * 826c3f9b018SEric Dumazet * But following code is meant to be called from BH handlers, 827c3f9b018SEric Dumazet * so we should keep BH disabled, but early release socket ownership 828c3f9b018SEric Dumazet */ 829c3f9b018SEric Dumazet sock_release_ownership(sk); 830c3f9b018SEric Dumazet 831144d56e9SEric Dumazet if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) { 8326f458dfbSEric Dumazet tcp_write_timer_handler(sk); 833144d56e9SEric Dumazet __sock_put(sk); 834144d56e9SEric Dumazet } 835144d56e9SEric Dumazet if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) { 8366f458dfbSEric Dumazet tcp_delack_timer_handler(sk); 837144d56e9SEric Dumazet __sock_put(sk); 838144d56e9SEric Dumazet } 839144d56e9SEric Dumazet if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) { 8404fab9071SNeal Cardwell inet_csk(sk)->icsk_af_ops->mtu_reduced(sk); 841144d56e9SEric Dumazet __sock_put(sk); 842144d56e9SEric Dumazet } 84346d3ceabSEric Dumazet } 84446d3ceabSEric Dumazet EXPORT_SYMBOL(tcp_release_cb); 84546d3ceabSEric Dumazet 84646d3ceabSEric Dumazet void __init tcp_tasklet_init(void) 84746d3ceabSEric Dumazet { 84846d3ceabSEric Dumazet int i; 84946d3ceabSEric Dumazet 85046d3ceabSEric Dumazet for_each_possible_cpu(i) { 85146d3ceabSEric Dumazet struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i); 85246d3ceabSEric Dumazet 85346d3ceabSEric Dumazet INIT_LIST_HEAD(&tsq->head); 85446d3ceabSEric Dumazet tasklet_init(&tsq->tasklet, 85546d3ceabSEric Dumazet tcp_tasklet_func, 85646d3ceabSEric Dumazet (unsigned long)tsq); 85746d3ceabSEric Dumazet } 85846d3ceabSEric Dumazet } 85946d3ceabSEric Dumazet 86046d3ceabSEric Dumazet /* 86146d3ceabSEric Dumazet * Write buffer destructor automatically called from kfree_skb. 8628e3bff96Sstephen hemminger * We can't xmit new skbs from this context, as we might already 86346d3ceabSEric Dumazet * hold qdisc lock. 86446d3ceabSEric Dumazet */ 865d6a4a104SEric Dumazet void tcp_wfree(struct sk_buff *skb) 86646d3ceabSEric Dumazet { 86746d3ceabSEric Dumazet struct sock *sk = skb->sk; 86846d3ceabSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 8699b462d02SEric Dumazet int wmem; 8709b462d02SEric Dumazet 8719b462d02SEric Dumazet /* Keep one reference on sk_wmem_alloc. 8729b462d02SEric Dumazet * Will be released by sk_free() from here or tcp_tasklet_func() 8739b462d02SEric Dumazet */ 8749b462d02SEric Dumazet wmem = atomic_sub_return(skb->truesize - 1, &sk->sk_wmem_alloc); 8759b462d02SEric Dumazet 8769b462d02SEric Dumazet /* If this softirq is serviced by ksoftirqd, we are likely under stress. 8779b462d02SEric Dumazet * Wait until our queues (qdisc + devices) are drained. 8789b462d02SEric Dumazet * This gives : 8799b462d02SEric Dumazet * - less callbacks to tcp_write_xmit(), reducing stress (batches) 8809b462d02SEric Dumazet * - chance for incoming ACK (processed by another cpu maybe) 8819b462d02SEric Dumazet * to migrate this flow (skb->ooo_okay will be eventually set) 8829b462d02SEric Dumazet */ 8839b462d02SEric Dumazet if (wmem >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current) 8849b462d02SEric Dumazet goto out; 88546d3ceabSEric Dumazet 88646d3ceabSEric Dumazet if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) && 88746d3ceabSEric Dumazet !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) { 88846d3ceabSEric Dumazet unsigned long flags; 88946d3ceabSEric Dumazet struct tsq_tasklet *tsq; 89046d3ceabSEric Dumazet 89146d3ceabSEric Dumazet /* queue this socket to tasklet queue */ 89246d3ceabSEric Dumazet local_irq_save(flags); 893903ceff7SChristoph Lameter tsq = this_cpu_ptr(&tsq_tasklet); 89446d3ceabSEric Dumazet list_add(&tp->tsq_node, &tsq->head); 89546d3ceabSEric Dumazet tasklet_schedule(&tsq->tasklet); 89646d3ceabSEric Dumazet local_irq_restore(flags); 8979b462d02SEric Dumazet return; 89846d3ceabSEric Dumazet } 8999b462d02SEric Dumazet out: 9009b462d02SEric Dumazet sk_free(sk); 90146d3ceabSEric Dumazet } 90246d3ceabSEric Dumazet 9031da177e4SLinus Torvalds /* This routine actually transmits TCP packets queued in by 9041da177e4SLinus Torvalds * tcp_do_sendmsg(). This is used by both the initial 9051da177e4SLinus Torvalds * transmission and possible later retransmissions. 9061da177e4SLinus Torvalds * All SKB's seen here are completely headerless. It is our 9071da177e4SLinus Torvalds * job to build the TCP header, and pass the packet down to 9081da177e4SLinus Torvalds * IP so it can do the same plus pass the packet off to the 9091da177e4SLinus Torvalds * device. 9101da177e4SLinus Torvalds * 9111da177e4SLinus Torvalds * We are working here with either a clone of the original 9121da177e4SLinus Torvalds * SKB, or a fresh unique copy made by the retransmit engine. 9131da177e4SLinus Torvalds */ 914056834d9SIlpo Järvinen static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, 915056834d9SIlpo Järvinen gfp_t gfp_mask) 9161da177e4SLinus Torvalds { 9176687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 918dfb4b9dcSDavid S. Miller struct inet_sock *inet; 919dfb4b9dcSDavid S. Miller struct tcp_sock *tp; 920dfb4b9dcSDavid S. Miller struct tcp_skb_cb *tcb; 92133ad798cSAdam Langley struct tcp_out_options opts; 92295c96174SEric Dumazet unsigned int tcp_options_size, tcp_header_size; 923cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 9241da177e4SLinus Torvalds struct tcphdr *th; 9251da177e4SLinus Torvalds int err; 9261da177e4SLinus Torvalds 927dfb4b9dcSDavid S. Miller BUG_ON(!skb || !tcp_skb_pcount(skb)); 928dfb4b9dcSDavid S. Miller 929ccdbb6e9SEric Dumazet if (clone_it) { 930740b0f18SEric Dumazet skb_mstamp_get(&skb->skb_mstamp); 931dfb4b9dcSDavid S. Miller 932dfb4b9dcSDavid S. Miller if (unlikely(skb_cloned(skb))) 933dfb4b9dcSDavid S. Miller skb = pskb_copy(skb, gfp_mask); 934dfb4b9dcSDavid S. Miller else 935dfb4b9dcSDavid S. Miller skb = skb_clone(skb, gfp_mask); 936dfb4b9dcSDavid S. Miller if (unlikely(!skb)) 937dfb4b9dcSDavid S. Miller return -ENOBUFS; 938dfb4b9dcSDavid S. Miller } 939dfb4b9dcSDavid S. Miller 940dfb4b9dcSDavid S. Miller inet = inet_sk(sk); 941dfb4b9dcSDavid S. Miller tp = tcp_sk(sk); 942dfb4b9dcSDavid S. Miller tcb = TCP_SKB_CB(skb); 94333ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 9441da177e4SLinus Torvalds 9454de075e0SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) 94633ad798cSAdam Langley tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); 94733ad798cSAdam Langley else 94833ad798cSAdam Langley tcp_options_size = tcp_established_options(sk, skb, &opts, 94933ad798cSAdam Langley &md5); 95033ad798cSAdam Langley tcp_header_size = tcp_options_size + sizeof(struct tcphdr); 9511da177e4SLinus Torvalds 952547669d4SEric Dumazet if (tcp_packets_in_flight(tp) == 0) 9536687e988SArnaldo Carvalho de Melo tcp_ca_event(sk, CA_EVENT_TX_START); 954547669d4SEric Dumazet 955547669d4SEric Dumazet /* if no packet is in qdisc/device queue, then allow XPS to select 956b2532eb9SEric Dumazet * another queue. We can be called from tcp_tsq_handler() 957b2532eb9SEric Dumazet * which holds one reference to sk_wmem_alloc. 958b2532eb9SEric Dumazet * 959b2532eb9SEric Dumazet * TODO: Ideally, in-flight pure ACK packets should not matter here. 960b2532eb9SEric Dumazet * One way to get this would be to set skb->truesize = 2 on them. 961547669d4SEric Dumazet */ 962b2532eb9SEric Dumazet skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1); 9631da177e4SLinus Torvalds 964aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 965aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 96646d3ceabSEric Dumazet 96746d3ceabSEric Dumazet skb_orphan(skb); 96846d3ceabSEric Dumazet skb->sk = sk; 96998781965SEric Dumazet skb->destructor = skb_is_tcp_pure_ack(skb) ? sock_wfree : tcp_wfree; 970b73c3d0eSTom Herbert skb_set_hash_from_sk(skb, sk); 97146d3ceabSEric Dumazet atomic_add(skb->truesize, &sk->sk_wmem_alloc); 9721da177e4SLinus Torvalds 9731da177e4SLinus Torvalds /* Build TCP header and checksum it. */ 974aa8223c7SArnaldo Carvalho de Melo th = tcp_hdr(skb); 975c720c7e8SEric Dumazet th->source = inet->inet_sport; 976c720c7e8SEric Dumazet th->dest = inet->inet_dport; 9771da177e4SLinus Torvalds th->seq = htonl(tcb->seq); 9781da177e4SLinus Torvalds th->ack_seq = htonl(tp->rcv_nxt); 979df7a3b07SAl Viro *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | 9804de075e0SEric Dumazet tcb->tcp_flags); 981dfb4b9dcSDavid S. Miller 9824de075e0SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) { 9831da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments 9841da177e4SLinus Torvalds * is never scaled. 9851da177e4SLinus Torvalds */ 986600ff0c2SIlpo Järvinen th->window = htons(min(tp->rcv_wnd, 65535U)); 9871da177e4SLinus Torvalds } else { 9881da177e4SLinus Torvalds th->window = htons(tcp_select_window(sk)); 9891da177e4SLinus Torvalds } 9901da177e4SLinus Torvalds th->check = 0; 9911da177e4SLinus Torvalds th->urg_ptr = 0; 9921da177e4SLinus Torvalds 99333f5f57eSIlpo Järvinen /* The urg_mode check is necessary during a below snd_una win probe */ 9947691367dSHerbert Xu if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { 9957691367dSHerbert Xu if (before(tp->snd_up, tcb->seq + 0x10000)) { 9961da177e4SLinus Torvalds th->urg_ptr = htons(tp->snd_up - tcb->seq); 9971da177e4SLinus Torvalds th->urg = 1; 9987691367dSHerbert Xu } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { 9990eae88f3SEric Dumazet th->urg_ptr = htons(0xFFFF); 10007691367dSHerbert Xu th->urg = 1; 10017691367dSHerbert Xu } 10021da177e4SLinus Torvalds } 10031da177e4SLinus Torvalds 1004bd0388aeSWilliam Allen Simpson tcp_options_write((__be32 *)(th + 1), tp, &opts); 100551466a75SEric Dumazet skb_shinfo(skb)->gso_type = sk->sk_gso_type; 10064de075e0SEric Dumazet if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0)) 1007735d3831SFlorian Westphal tcp_ecn_send(sk, skb, tcp_header_size); 1008dfb4b9dcSDavid S. Miller 1009cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 1010cfb6eeb4SYOSHIFUJI Hideaki /* Calculate the MD5 hash, as we have all we need now */ 1011cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 1012a465419bSEric Dumazet sk_nocaps_add(sk, NETIF_F_GSO_MASK); 1013bd0388aeSWilliam Allen Simpson tp->af_specific->calc_md5_hash(opts.hash_location, 101439f8e58eSEric Dumazet md5, sk, skb); 1015cfb6eeb4SYOSHIFUJI Hideaki } 1016cfb6eeb4SYOSHIFUJI Hideaki #endif 1017cfb6eeb4SYOSHIFUJI Hideaki 1018bb296246SHerbert Xu icsk->icsk_af_ops->send_check(sk, skb); 10191da177e4SLinus Torvalds 10204de075e0SEric Dumazet if (likely(tcb->tcp_flags & TCPHDR_ACK)) 1021fc6415bcSDavid S. Miller tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); 10221da177e4SLinus Torvalds 10231da177e4SLinus Torvalds if (skb->len != tcp_header_size) 1024cf533ea5SEric Dumazet tcp_event_data_sent(tp, sk); 10251da177e4SLinus Torvalds 1026bd37a088SWei Yongjun if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) 1027aa2ea058STom Herbert TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, 1028aa2ea058STom Herbert tcp_skb_pcount(skb)); 10291da177e4SLinus Torvalds 10302efd055cSMarcelo Ricardo Leitner tp->segs_out += tcp_skb_pcount(skb); 1031cd7d8498SEric Dumazet /* OK, its time to fill skb_shinfo(skb)->gso_segs */ 1032cd7d8498SEric Dumazet skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb); 1033cd7d8498SEric Dumazet 10347faee5c0SEric Dumazet /* Our usage of tstamp should remain private */ 10357faee5c0SEric Dumazet skb->tstamp.tv64 = 0; 1036971f10ecSEric Dumazet 1037971f10ecSEric Dumazet /* Cleanup our debris for IP stacks */ 1038971f10ecSEric Dumazet memset(skb->cb, 0, max(sizeof(struct inet_skb_parm), 1039971f10ecSEric Dumazet sizeof(struct inet6_skb_parm))); 1040971f10ecSEric Dumazet 1041b0270e91SEric Dumazet err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl); 10427faee5c0SEric Dumazet 104383de47cdSHua Zhong if (likely(err <= 0)) 10441da177e4SLinus Torvalds return err; 10451da177e4SLinus Torvalds 10465ee2c941SChristoph Paasch tcp_enter_cwr(sk); 10471da177e4SLinus Torvalds 1048b9df3cb8SGerrit Renker return net_xmit_eval(err); 10491da177e4SLinus Torvalds } 10501da177e4SLinus Torvalds 105167edfef7SAndi Kleen /* This routine just queues the buffer for sending. 10521da177e4SLinus Torvalds * 10531da177e4SLinus Torvalds * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, 10541da177e4SLinus Torvalds * otherwise socket can stall. 10551da177e4SLinus Torvalds */ 10561da177e4SLinus Torvalds static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) 10571da177e4SLinus Torvalds { 10581da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 10591da177e4SLinus Torvalds 10601da177e4SLinus Torvalds /* Advance write_seq and place onto the write_queue. */ 10611da177e4SLinus Torvalds tp->write_seq = TCP_SKB_CB(skb)->end_seq; 1062f4a775d1SEric Dumazet __skb_header_release(skb); 1063fe067e8aSDavid S. Miller tcp_add_write_queue_tail(sk, skb); 10643ab224beSHideo Aoki sk->sk_wmem_queued += skb->truesize; 10653ab224beSHideo Aoki sk_mem_charge(sk, skb->truesize); 10661da177e4SLinus Torvalds } 10671da177e4SLinus Torvalds 106867edfef7SAndi Kleen /* Initialize TSO segments for a packet. */ 1069*5bbb432cSEric Dumazet static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now) 1070f6302d1dSDavid S. Miller { 10717b7fc97aSEric Dumazet struct skb_shared_info *shinfo = skb_shinfo(skb); 10727b7fc97aSEric Dumazet 1073c52e2421SEric Dumazet /* Make sure we own this skb before messing gso_size/gso_segs */ 1074c52e2421SEric Dumazet WARN_ON_ONCE(skb_cloned(skb)); 1075c52e2421SEric Dumazet 10768f26fb1cSEric Dumazet if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) { 1077f6302d1dSDavid S. Miller /* Avoid the costly divide in the normal 1078f6302d1dSDavid S. Miller * non-TSO case. 1079f6302d1dSDavid S. Miller */ 1080cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, 1); 10817b7fc97aSEric Dumazet shinfo->gso_size = 0; 1082f6302d1dSDavid S. Miller } else { 1083cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now)); 10847b7fc97aSEric Dumazet shinfo->gso_size = mss_now; 10851da177e4SLinus Torvalds } 10861da177e4SLinus Torvalds } 10871da177e4SLinus Torvalds 108891fed7a1SIlpo Järvinen /* When a modification to fackets out becomes necessary, we need to check 108968f8353bSIlpo Järvinen * skb is counted to fackets_out or not. 109091fed7a1SIlpo Järvinen */ 1091cf533ea5SEric Dumazet static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb, 109291fed7a1SIlpo Järvinen int decr) 109391fed7a1SIlpo Järvinen { 1094a47e5a98SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1095a47e5a98SIlpo Järvinen 1096dc86967bSIlpo Järvinen if (!tp->sacked_out || tcp_is_reno(tp)) 109791fed7a1SIlpo Järvinen return; 109891fed7a1SIlpo Järvinen 10996859d494SIlpo Järvinen if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq)) 110091fed7a1SIlpo Järvinen tp->fackets_out -= decr; 110191fed7a1SIlpo Järvinen } 110291fed7a1SIlpo Järvinen 1103797108d1SIlpo Järvinen /* Pcount in the middle of the write queue got changed, we need to do various 1104797108d1SIlpo Järvinen * tweaks to fix counters 1105797108d1SIlpo Järvinen */ 1106cf533ea5SEric Dumazet static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) 1107797108d1SIlpo Järvinen { 1108797108d1SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1109797108d1SIlpo Järvinen 1110797108d1SIlpo Järvinen tp->packets_out -= decr; 1111797108d1SIlpo Järvinen 1112797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 1113797108d1SIlpo Järvinen tp->sacked_out -= decr; 1114797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 1115797108d1SIlpo Järvinen tp->retrans_out -= decr; 1116797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 1117797108d1SIlpo Järvinen tp->lost_out -= decr; 1118797108d1SIlpo Järvinen 1119797108d1SIlpo Järvinen /* Reno case is special. Sigh... */ 1120797108d1SIlpo Järvinen if (tcp_is_reno(tp) && decr > 0) 1121797108d1SIlpo Järvinen tp->sacked_out -= min_t(u32, tp->sacked_out, decr); 1122797108d1SIlpo Järvinen 1123797108d1SIlpo Järvinen tcp_adjust_fackets_out(sk, skb, decr); 1124797108d1SIlpo Järvinen 1125797108d1SIlpo Järvinen if (tp->lost_skb_hint && 1126797108d1SIlpo Järvinen before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && 112752cf3cc8SIlpo Järvinen (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))) 1128797108d1SIlpo Järvinen tp->lost_cnt_hint -= decr; 1129797108d1SIlpo Järvinen 1130797108d1SIlpo Järvinen tcp_verify_left_out(tp); 1131797108d1SIlpo Järvinen } 1132797108d1SIlpo Järvinen 1133490cc7d0SWillem de Bruijn static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2) 1134490cc7d0SWillem de Bruijn { 1135490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo = skb_shinfo(skb); 1136490cc7d0SWillem de Bruijn 1137490cc7d0SWillem de Bruijn if (unlikely(shinfo->tx_flags & SKBTX_ANY_TSTAMP) && 1138490cc7d0SWillem de Bruijn !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) { 1139490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo2 = skb_shinfo(skb2); 1140490cc7d0SWillem de Bruijn u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP; 1141490cc7d0SWillem de Bruijn 1142490cc7d0SWillem de Bruijn shinfo->tx_flags &= ~tsflags; 1143490cc7d0SWillem de Bruijn shinfo2->tx_flags |= tsflags; 1144490cc7d0SWillem de Bruijn swap(shinfo->tskey, shinfo2->tskey); 1145490cc7d0SWillem de Bruijn } 1146490cc7d0SWillem de Bruijn } 1147490cc7d0SWillem de Bruijn 11481da177e4SLinus Torvalds /* Function to create two new TCP segments. Shrinks the given segment 11491da177e4SLinus Torvalds * to the specified size and appends a new segment with the rest of the 11501da177e4SLinus Torvalds * packet to the list. This won't be called frequently, I hope. 11511da177e4SLinus Torvalds * Remember, these are still headerless SKBs at this point. 11521da177e4SLinus Torvalds */ 1153056834d9SIlpo Järvinen int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, 11546cc55e09SOctavian Purdila unsigned int mss_now, gfp_t gfp) 11551da177e4SLinus Torvalds { 11561da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 11571da177e4SLinus Torvalds struct sk_buff *buff; 11586475be16SDavid S. Miller int nsize, old_factor; 1159b60b49eaSHerbert Xu int nlen; 11609ce01461SIlpo Järvinen u8 flags; 11611da177e4SLinus Torvalds 11622fceec13SIlpo Järvinen if (WARN_ON(len > skb->len)) 11632fceec13SIlpo Järvinen return -EINVAL; 11646a438bbeSStephen Hemminger 11651da177e4SLinus Torvalds nsize = skb_headlen(skb) - len; 11661da177e4SLinus Torvalds if (nsize < 0) 11671da177e4SLinus Torvalds nsize = 0; 11681da177e4SLinus Torvalds 11696cc55e09SOctavian Purdila if (skb_unclone(skb, gfp)) 11701da177e4SLinus Torvalds return -ENOMEM; 11711da177e4SLinus Torvalds 11721da177e4SLinus Torvalds /* Get a new skb... force flag on. */ 1173eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, nsize, gfp, true); 117451456b29SIan Morris if (!buff) 11751da177e4SLinus Torvalds return -ENOMEM; /* We'll just try again later. */ 1176ef5cb973SHerbert Xu 11773ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 11783ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1179b60b49eaSHerbert Xu nlen = skb->len - len - nsize; 1180b60b49eaSHerbert Xu buff->truesize += nlen; 1181b60b49eaSHerbert Xu skb->truesize -= nlen; 11821da177e4SLinus Torvalds 11831da177e4SLinus Torvalds /* Correct the sequence numbers. */ 11841da177e4SLinus Torvalds TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 11851da177e4SLinus Torvalds TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 11861da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 11871da177e4SLinus Torvalds 11881da177e4SLinus Torvalds /* PSH and FIN should only be set in the second packet. */ 11894de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 11904de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 11914de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1192e14c3cafSHerbert Xu TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; 11931da177e4SLinus Torvalds 119484fa7933SPatrick McHardy if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { 11951da177e4SLinus Torvalds /* Copy and checksum data tail into the new buffer. */ 1196056834d9SIlpo Järvinen buff->csum = csum_partial_copy_nocheck(skb->data + len, 1197056834d9SIlpo Järvinen skb_put(buff, nsize), 11981da177e4SLinus Torvalds nsize, 0); 11991da177e4SLinus Torvalds 12001da177e4SLinus Torvalds skb_trim(skb, len); 12011da177e4SLinus Torvalds 12021da177e4SLinus Torvalds skb->csum = csum_block_sub(skb->csum, buff->csum, len); 12031da177e4SLinus Torvalds } else { 120484fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 12051da177e4SLinus Torvalds skb_split(skb, buff, len); 12061da177e4SLinus Torvalds } 12071da177e4SLinus Torvalds 12081da177e4SLinus Torvalds buff->ip_summed = skb->ip_summed; 12091da177e4SLinus Torvalds 1210a61bbcf2SPatrick McHardy buff->tstamp = skb->tstamp; 1211490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 12121da177e4SLinus Torvalds 12136475be16SDavid S. Miller old_factor = tcp_skb_pcount(skb); 12146475be16SDavid S. Miller 12151da177e4SLinus Torvalds /* Fix up tso_factor for both original and new SKB. */ 1216*5bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 1217*5bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 12181da177e4SLinus Torvalds 12196475be16SDavid S. Miller /* If this packet has been sent out already, we must 12206475be16SDavid S. Miller * adjust the various packet counters. 12216475be16SDavid S. Miller */ 1222cf0b450cSHerbert Xu if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 12236475be16SDavid S. Miller int diff = old_factor - tcp_skb_pcount(skb) - 12246475be16SDavid S. Miller tcp_skb_pcount(buff); 12251da177e4SLinus Torvalds 1226797108d1SIlpo Järvinen if (diff) 1227797108d1SIlpo Järvinen tcp_adjust_pcount(sk, skb, diff); 12281da177e4SLinus Torvalds } 12291da177e4SLinus Torvalds 12301da177e4SLinus Torvalds /* Link BUFF into the send queue. */ 1231f4a775d1SEric Dumazet __skb_header_release(buff); 1232fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 12331da177e4SLinus Torvalds 12341da177e4SLinus Torvalds return 0; 12351da177e4SLinus Torvalds } 12361da177e4SLinus Torvalds 12371da177e4SLinus Torvalds /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c 12381da177e4SLinus Torvalds * eventually). The difference is that pulled data not copied, but 12391da177e4SLinus Torvalds * immediately discarded. 12401da177e4SLinus Torvalds */ 1241f2911969SHerbert Xu ~{PmVHI~} static void __pskb_trim_head(struct sk_buff *skb, int len) 12421da177e4SLinus Torvalds { 12437b7fc97aSEric Dumazet struct skb_shared_info *shinfo; 12441da177e4SLinus Torvalds int i, k, eat; 12451da177e4SLinus Torvalds 12464fa48bf3SEric Dumazet eat = min_t(int, len, skb_headlen(skb)); 12474fa48bf3SEric Dumazet if (eat) { 12484fa48bf3SEric Dumazet __skb_pull(skb, eat); 12494fa48bf3SEric Dumazet len -= eat; 12504fa48bf3SEric Dumazet if (!len) 12514fa48bf3SEric Dumazet return; 12524fa48bf3SEric Dumazet } 12531da177e4SLinus Torvalds eat = len; 12541da177e4SLinus Torvalds k = 0; 12557b7fc97aSEric Dumazet shinfo = skb_shinfo(skb); 12567b7fc97aSEric Dumazet for (i = 0; i < shinfo->nr_frags; i++) { 12577b7fc97aSEric Dumazet int size = skb_frag_size(&shinfo->frags[i]); 12589e903e08SEric Dumazet 12599e903e08SEric Dumazet if (size <= eat) { 1260aff65da0SIan Campbell skb_frag_unref(skb, i); 12619e903e08SEric Dumazet eat -= size; 12621da177e4SLinus Torvalds } else { 12637b7fc97aSEric Dumazet shinfo->frags[k] = shinfo->frags[i]; 12641da177e4SLinus Torvalds if (eat) { 12657b7fc97aSEric Dumazet shinfo->frags[k].page_offset += eat; 12667b7fc97aSEric Dumazet skb_frag_size_sub(&shinfo->frags[k], eat); 12671da177e4SLinus Torvalds eat = 0; 12681da177e4SLinus Torvalds } 12691da177e4SLinus Torvalds k++; 12701da177e4SLinus Torvalds } 12711da177e4SLinus Torvalds } 12727b7fc97aSEric Dumazet shinfo->nr_frags = k; 12731da177e4SLinus Torvalds 127427a884dcSArnaldo Carvalho de Melo skb_reset_tail_pointer(skb); 12751da177e4SLinus Torvalds skb->data_len -= len; 12761da177e4SLinus Torvalds skb->len = skb->data_len; 12771da177e4SLinus Torvalds } 12781da177e4SLinus Torvalds 127967edfef7SAndi Kleen /* Remove acked data from a packet in the transmit queue. */ 12801da177e4SLinus Torvalds int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 12811da177e4SLinus Torvalds { 128214bbd6a5SPravin B Shelar if (skb_unclone(skb, GFP_ATOMIC)) 12831da177e4SLinus Torvalds return -ENOMEM; 12841da177e4SLinus Torvalds 12854fa48bf3SEric Dumazet __pskb_trim_head(skb, len); 12861da177e4SLinus Torvalds 12871da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq += len; 128884fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 12891da177e4SLinus Torvalds 12901da177e4SLinus Torvalds skb->truesize -= len; 12911da177e4SLinus Torvalds sk->sk_wmem_queued -= len; 12923ab224beSHideo Aoki sk_mem_uncharge(sk, len); 12931da177e4SLinus Torvalds sock_set_flag(sk, SOCK_QUEUE_SHRUNK); 12941da177e4SLinus Torvalds 12955b35e1e6SNeal Cardwell /* Any change of skb->len requires recalculation of tso factor. */ 12961da177e4SLinus Torvalds if (tcp_skb_pcount(skb) > 1) 1297*5bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb)); 12981da177e4SLinus Torvalds 12991da177e4SLinus Torvalds return 0; 13001da177e4SLinus Torvalds } 13011da177e4SLinus Torvalds 13021b63edd6SYuchung Cheng /* Calculate MSS not accounting any TCP options. */ 13031b63edd6SYuchung Cheng static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu) 13045d424d5aSJohn Heffner { 1305cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1306cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 13075d424d5aSJohn Heffner int mss_now; 13085d424d5aSJohn Heffner 13095d424d5aSJohn Heffner /* Calculate base mss without TCP options: 13105d424d5aSJohn Heffner It is MMS_S - sizeof(tcphdr) of rfc1122 13115d424d5aSJohn Heffner */ 13125d424d5aSJohn Heffner mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 13135d424d5aSJohn Heffner 131467469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 131567469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 131667469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 131767469601SEric Dumazet 131867469601SEric Dumazet if (dst && dst_allfrag(dst)) 131967469601SEric Dumazet mss_now -= icsk->icsk_af_ops->net_frag_header_len; 132067469601SEric Dumazet } 132167469601SEric Dumazet 13225d424d5aSJohn Heffner /* Clamp it (mss_clamp does not include tcp options) */ 13235d424d5aSJohn Heffner if (mss_now > tp->rx_opt.mss_clamp) 13245d424d5aSJohn Heffner mss_now = tp->rx_opt.mss_clamp; 13255d424d5aSJohn Heffner 13265d424d5aSJohn Heffner /* Now subtract optional transport overhead */ 13275d424d5aSJohn Heffner mss_now -= icsk->icsk_ext_hdr_len; 13285d424d5aSJohn Heffner 13295d424d5aSJohn Heffner /* Then reserve room for full set of TCP options and 8 bytes of data */ 13305d424d5aSJohn Heffner if (mss_now < 48) 13315d424d5aSJohn Heffner mss_now = 48; 13325d424d5aSJohn Heffner return mss_now; 13335d424d5aSJohn Heffner } 13345d424d5aSJohn Heffner 13351b63edd6SYuchung Cheng /* Calculate MSS. Not accounting for SACKs here. */ 13361b63edd6SYuchung Cheng int tcp_mtu_to_mss(struct sock *sk, int pmtu) 13371b63edd6SYuchung Cheng { 13381b63edd6SYuchung Cheng /* Subtract TCP options size, not including SACKs */ 13391b63edd6SYuchung Cheng return __tcp_mtu_to_mss(sk, pmtu) - 13401b63edd6SYuchung Cheng (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr)); 13411b63edd6SYuchung Cheng } 13421b63edd6SYuchung Cheng 13435d424d5aSJohn Heffner /* Inverse of above */ 134467469601SEric Dumazet int tcp_mss_to_mtu(struct sock *sk, int mss) 13455d424d5aSJohn Heffner { 1346cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1347cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 13485d424d5aSJohn Heffner int mtu; 13495d424d5aSJohn Heffner 13505d424d5aSJohn Heffner mtu = mss + 13515d424d5aSJohn Heffner tp->tcp_header_len + 13525d424d5aSJohn Heffner icsk->icsk_ext_hdr_len + 13535d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 13545d424d5aSJohn Heffner 135567469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 135667469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 135767469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 135867469601SEric Dumazet 135967469601SEric Dumazet if (dst && dst_allfrag(dst)) 136067469601SEric Dumazet mtu += icsk->icsk_af_ops->net_frag_header_len; 136167469601SEric Dumazet } 13625d424d5aSJohn Heffner return mtu; 13635d424d5aSJohn Heffner } 13645d424d5aSJohn Heffner 136567edfef7SAndi Kleen /* MTU probing init per socket */ 13665d424d5aSJohn Heffner void tcp_mtup_init(struct sock *sk) 13675d424d5aSJohn Heffner { 13685d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 13695d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 1370b0f9ca53SFan Du struct net *net = sock_net(sk); 13715d424d5aSJohn Heffner 1372b0f9ca53SFan Du icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1; 13735d424d5aSJohn Heffner icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 13745d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 1375b0f9ca53SFan Du icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss); 13765d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 137705cbc0dbSFan Du if (icsk->icsk_mtup.enabled) 137805cbc0dbSFan Du icsk->icsk_mtup.probe_timestamp = tcp_time_stamp; 13795d424d5aSJohn Heffner } 13804bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_mtup_init); 13815d424d5aSJohn Heffner 13821da177e4SLinus Torvalds /* This function synchronize snd mss to current pmtu/exthdr set. 13831da177e4SLinus Torvalds 13841da177e4SLinus Torvalds tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 13851da177e4SLinus Torvalds for TCP options, but includes only bare TCP header. 13861da177e4SLinus Torvalds 13871da177e4SLinus Torvalds tp->rx_opt.mss_clamp is mss negotiated at connection setup. 1388caa20d9aSStephen Hemminger It is minimum of user_mss and mss received with SYN. 13891da177e4SLinus Torvalds It also does not include TCP options. 13901da177e4SLinus Torvalds 1391d83d8461SArnaldo Carvalho de Melo inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. 13921da177e4SLinus Torvalds 13931da177e4SLinus Torvalds tp->mss_cache is current effective sending mss, including 13941da177e4SLinus Torvalds all tcp options except for SACKs. It is evaluated, 13951da177e4SLinus Torvalds taking into account current pmtu, but never exceeds 13961da177e4SLinus Torvalds tp->rx_opt.mss_clamp. 13971da177e4SLinus Torvalds 13981da177e4SLinus Torvalds NOTE1. rfc1122 clearly states that advertised MSS 13991da177e4SLinus Torvalds DOES NOT include either tcp or ip options. 14001da177e4SLinus Torvalds 1401d83d8461SArnaldo Carvalho de Melo NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache 1402d83d8461SArnaldo Carvalho de Melo are READ ONLY outside this function. --ANK (980731) 14031da177e4SLinus Torvalds */ 14041da177e4SLinus Torvalds unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 14051da177e4SLinus Torvalds { 14061da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1407d83d8461SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 14085d424d5aSJohn Heffner int mss_now; 14091da177e4SLinus Torvalds 14105d424d5aSJohn Heffner if (icsk->icsk_mtup.search_high > pmtu) 14115d424d5aSJohn Heffner icsk->icsk_mtup.search_high = pmtu; 14121da177e4SLinus Torvalds 14135d424d5aSJohn Heffner mss_now = tcp_mtu_to_mss(sk, pmtu); 1414409d22b4SIlpo Järvinen mss_now = tcp_bound_to_half_wnd(tp, mss_now); 14151da177e4SLinus Torvalds 14161da177e4SLinus Torvalds /* And store cached results */ 1417d83d8461SArnaldo Carvalho de Melo icsk->icsk_pmtu_cookie = pmtu; 14185d424d5aSJohn Heffner if (icsk->icsk_mtup.enabled) 14195d424d5aSJohn Heffner mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); 1420c1b4a7e6SDavid S. Miller tp->mss_cache = mss_now; 14211da177e4SLinus Torvalds 14221da177e4SLinus Torvalds return mss_now; 14231da177e4SLinus Torvalds } 14244bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_sync_mss); 14251da177e4SLinus Torvalds 14261da177e4SLinus Torvalds /* Compute the current effective MSS, taking SACKs and IP options, 14271da177e4SLinus Torvalds * and even PMTU discovery events into account. 14281da177e4SLinus Torvalds */ 14290c54b85fSIlpo Järvinen unsigned int tcp_current_mss(struct sock *sk) 14301da177e4SLinus Torvalds { 1431cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1432cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1433c1b4a7e6SDavid S. Miller u32 mss_now; 143495c96174SEric Dumazet unsigned int header_len; 143533ad798cSAdam Langley struct tcp_out_options opts; 143633ad798cSAdam Langley struct tcp_md5sig_key *md5; 14371da177e4SLinus Torvalds 1438c1b4a7e6SDavid S. Miller mss_now = tp->mss_cache; 1439c1b4a7e6SDavid S. Miller 14401da177e4SLinus Torvalds if (dst) { 14411da177e4SLinus Torvalds u32 mtu = dst_mtu(dst); 1442d83d8461SArnaldo Carvalho de Melo if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 14431da177e4SLinus Torvalds mss_now = tcp_sync_mss(sk, mtu); 14441da177e4SLinus Torvalds } 14451da177e4SLinus Torvalds 144633ad798cSAdam Langley header_len = tcp_established_options(sk, NULL, &opts, &md5) + 144733ad798cSAdam Langley sizeof(struct tcphdr); 144833ad798cSAdam Langley /* The mss_cache is sized based on tp->tcp_header_len, which assumes 144933ad798cSAdam Langley * some common options. If this is an odd packet (because we have SACK 145033ad798cSAdam Langley * blocks etc) then our calculated header_len will be different, and 145133ad798cSAdam Langley * we have to adjust mss_now correspondingly */ 145233ad798cSAdam Langley if (header_len != tp->tcp_header_len) { 145333ad798cSAdam Langley int delta = (int) header_len - tp->tcp_header_len; 145433ad798cSAdam Langley mss_now -= delta; 145533ad798cSAdam Langley } 1456cfb6eeb4SYOSHIFUJI Hideaki 14571da177e4SLinus Torvalds return mss_now; 14581da177e4SLinus Torvalds } 14591da177e4SLinus Torvalds 146086fd14adSWeiping Pan /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. 146186fd14adSWeiping Pan * As additional protections, we do not touch cwnd in retransmission phases, 146286fd14adSWeiping Pan * and if application hit its sndbuf limit recently. 146386fd14adSWeiping Pan */ 146486fd14adSWeiping Pan static void tcp_cwnd_application_limited(struct sock *sk) 1465a762a980SDavid S. Miller { 14669e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1467a762a980SDavid S. Miller 146886fd14adSWeiping Pan if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && 146986fd14adSWeiping Pan sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 147086fd14adSWeiping Pan /* Limited by application or receiver window. */ 147186fd14adSWeiping Pan u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); 147286fd14adSWeiping Pan u32 win_used = max(tp->snd_cwnd_used, init_win); 147386fd14adSWeiping Pan if (win_used < tp->snd_cwnd) { 147486fd14adSWeiping Pan tp->snd_ssthresh = tcp_current_ssthresh(sk); 147586fd14adSWeiping Pan tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1; 147686fd14adSWeiping Pan } 147786fd14adSWeiping Pan tp->snd_cwnd_used = 0; 147886fd14adSWeiping Pan } 147986fd14adSWeiping Pan tp->snd_cwnd_stamp = tcp_time_stamp; 148086fd14adSWeiping Pan } 148186fd14adSWeiping Pan 1482ca8a2263SNeal Cardwell static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited) 1483a762a980SDavid S. Miller { 1484a762a980SDavid S. Miller struct tcp_sock *tp = tcp_sk(sk); 1485a762a980SDavid S. Miller 1486ca8a2263SNeal Cardwell /* Track the maximum number of outstanding packets in each 1487ca8a2263SNeal Cardwell * window, and remember whether we were cwnd-limited then. 1488ca8a2263SNeal Cardwell */ 1489ca8a2263SNeal Cardwell if (!before(tp->snd_una, tp->max_packets_seq) || 1490ca8a2263SNeal Cardwell tp->packets_out > tp->max_packets_out) { 1491ca8a2263SNeal Cardwell tp->max_packets_out = tp->packets_out; 1492ca8a2263SNeal Cardwell tp->max_packets_seq = tp->snd_nxt; 1493ca8a2263SNeal Cardwell tp->is_cwnd_limited = is_cwnd_limited; 1494ca8a2263SNeal Cardwell } 1495e114a710SEric Dumazet 149624901551SEric Dumazet if (tcp_is_cwnd_limited(sk)) { 1497a762a980SDavid S. Miller /* Network is feed fully. */ 1498a762a980SDavid S. Miller tp->snd_cwnd_used = 0; 1499a762a980SDavid S. Miller tp->snd_cwnd_stamp = tcp_time_stamp; 1500a762a980SDavid S. Miller } else { 1501a762a980SDavid S. Miller /* Network starves. */ 1502a762a980SDavid S. Miller if (tp->packets_out > tp->snd_cwnd_used) 1503a762a980SDavid S. Miller tp->snd_cwnd_used = tp->packets_out; 1504a762a980SDavid S. Miller 150515d33c07SDavid S. Miller if (sysctl_tcp_slow_start_after_idle && 150615d33c07SDavid S. Miller (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto) 1507a762a980SDavid S. Miller tcp_cwnd_application_limited(sk); 1508a762a980SDavid S. Miller } 1509a762a980SDavid S. Miller } 1510a762a980SDavid S. Miller 1511d4589926SEric Dumazet /* Minshall's variant of the Nagle send check. */ 1512d4589926SEric Dumazet static bool tcp_minshall_check(const struct tcp_sock *tp) 1513d4589926SEric Dumazet { 1514d4589926SEric Dumazet return after(tp->snd_sml, tp->snd_una) && 1515d4589926SEric Dumazet !after(tp->snd_sml, tp->snd_nxt); 1516d4589926SEric Dumazet } 1517d4589926SEric Dumazet 1518d4589926SEric Dumazet /* Update snd_sml if this skb is under mss 1519d4589926SEric Dumazet * Note that a TSO packet might end with a sub-mss segment 1520d4589926SEric Dumazet * The test is really : 1521d4589926SEric Dumazet * if ((skb->len % mss) != 0) 1522d4589926SEric Dumazet * tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1523d4589926SEric Dumazet * But we can avoid doing the divide again given we already have 1524d4589926SEric Dumazet * skb_pcount = skb->len / mss_now 15250e3a4803SIlpo Järvinen */ 1526d4589926SEric Dumazet static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now, 1527d4589926SEric Dumazet const struct sk_buff *skb) 1528d4589926SEric Dumazet { 1529d4589926SEric Dumazet if (skb->len < tcp_skb_pcount(skb) * mss_now) 1530d4589926SEric Dumazet tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1531d4589926SEric Dumazet } 1532d4589926SEric Dumazet 1533d4589926SEric Dumazet /* Return false, if packet can be sent now without violation Nagle's rules: 1534d4589926SEric Dumazet * 1. It is full sized. (provided by caller in %partial bool) 1535d4589926SEric Dumazet * 2. Or it contains FIN. (already checked by caller) 1536d4589926SEric Dumazet * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. 1537d4589926SEric Dumazet * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 1538d4589926SEric Dumazet * With Minshall's modification: all sent small packets are ACKed. 1539d4589926SEric Dumazet */ 1540d4589926SEric Dumazet static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp, 1541cc93fc51SPeter Pan(潘卫平) int nonagle) 1542d4589926SEric Dumazet { 1543d4589926SEric Dumazet return partial && 1544d4589926SEric Dumazet ((nonagle & TCP_NAGLE_CORK) || 1545d4589926SEric Dumazet (!nonagle && tp->packets_out && tcp_minshall_check(tp))); 1546d4589926SEric Dumazet } 1547605ad7f1SEric Dumazet 1548605ad7f1SEric Dumazet /* Return how many segs we'd like on a TSO packet, 1549605ad7f1SEric Dumazet * to send one TSO packet per ms 1550605ad7f1SEric Dumazet */ 1551605ad7f1SEric Dumazet static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now) 1552605ad7f1SEric Dumazet { 1553605ad7f1SEric Dumazet u32 bytes, segs; 1554605ad7f1SEric Dumazet 1555605ad7f1SEric Dumazet bytes = min(sk->sk_pacing_rate >> 10, 1556605ad7f1SEric Dumazet sk->sk_gso_max_size - 1 - MAX_TCP_HEADER); 1557605ad7f1SEric Dumazet 1558605ad7f1SEric Dumazet /* Goal is to send at least one packet per ms, 1559605ad7f1SEric Dumazet * not one big TSO packet every 100 ms. 1560605ad7f1SEric Dumazet * This preserves ACK clocking and is consistent 1561605ad7f1SEric Dumazet * with tcp_tso_should_defer() heuristic. 1562605ad7f1SEric Dumazet */ 1563605ad7f1SEric Dumazet segs = max_t(u32, bytes / mss_now, sysctl_tcp_min_tso_segs); 1564605ad7f1SEric Dumazet 1565605ad7f1SEric Dumazet return min_t(u32, segs, sk->sk_gso_max_segs); 1566605ad7f1SEric Dumazet } 1567605ad7f1SEric Dumazet 1568d4589926SEric Dumazet /* Returns the portion of skb which can be sent right away */ 1569d4589926SEric Dumazet static unsigned int tcp_mss_split_point(const struct sock *sk, 1570d4589926SEric Dumazet const struct sk_buff *skb, 1571d4589926SEric Dumazet unsigned int mss_now, 1572d4589926SEric Dumazet unsigned int max_segs, 1573d4589926SEric Dumazet int nonagle) 1574c1b4a7e6SDavid S. Miller { 1575cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1576d4589926SEric Dumazet u32 partial, needed, window, max_len; 1577c1b4a7e6SDavid S. Miller 157890840defSIlpo Järvinen window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 15791485348dSBen Hutchings max_len = mss_now * max_segs; 15800e3a4803SIlpo Järvinen 15811485348dSBen Hutchings if (likely(max_len <= window && skb != tcp_write_queue_tail(sk))) 15821485348dSBen Hutchings return max_len; 15830e3a4803SIlpo Järvinen 15845ea3a748SIlpo Järvinen needed = min(skb->len, window); 15855ea3a748SIlpo Järvinen 15861485348dSBen Hutchings if (max_len <= needed) 15871485348dSBen Hutchings return max_len; 15880e3a4803SIlpo Järvinen 1589d4589926SEric Dumazet partial = needed % mss_now; 1590d4589926SEric Dumazet /* If last segment is not a full MSS, check if Nagle rules allow us 1591d4589926SEric Dumazet * to include this last segment in this skb. 1592d4589926SEric Dumazet * Otherwise, we'll split the skb at last MSS boundary 1593d4589926SEric Dumazet */ 1594cc93fc51SPeter Pan(潘卫平) if (tcp_nagle_check(partial != 0, tp, nonagle)) 1595d4589926SEric Dumazet return needed - partial; 1596d4589926SEric Dumazet 1597d4589926SEric Dumazet return needed; 1598c1b4a7e6SDavid S. Miller } 1599c1b4a7e6SDavid S. Miller 1600c1b4a7e6SDavid S. Miller /* Can at least one segment of SKB be sent right now, according to the 1601c1b4a7e6SDavid S. Miller * congestion window rules? If so, return how many segments are allowed. 1602c1b4a7e6SDavid S. Miller */ 1603cf533ea5SEric Dumazet static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, 1604cf533ea5SEric Dumazet const struct sk_buff *skb) 1605c1b4a7e6SDavid S. Miller { 1606d649a7a8SEric Dumazet u32 in_flight, cwnd, halfcwnd; 1607c1b4a7e6SDavid S. Miller 1608c1b4a7e6SDavid S. Miller /* Don't be strict about the congestion window for the final FIN. */ 16094de075e0SEric Dumazet if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && 16104de075e0SEric Dumazet tcp_skb_pcount(skb) == 1) 1611c1b4a7e6SDavid S. Miller return 1; 1612c1b4a7e6SDavid S. Miller 1613c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1614c1b4a7e6SDavid S. Miller cwnd = tp->snd_cwnd; 1615d649a7a8SEric Dumazet if (in_flight >= cwnd) 1616c1b4a7e6SDavid S. Miller return 0; 1617d649a7a8SEric Dumazet 1618d649a7a8SEric Dumazet /* For better scheduling, ensure we have at least 1619d649a7a8SEric Dumazet * 2 GSO packets in flight. 1620d649a7a8SEric Dumazet */ 1621d649a7a8SEric Dumazet halfcwnd = max(cwnd >> 1, 1U); 1622d649a7a8SEric Dumazet return min(halfcwnd, cwnd - in_flight); 1623c1b4a7e6SDavid S. Miller } 1624c1b4a7e6SDavid S. Miller 1625b595076aSUwe Kleine-König /* Initialize TSO state of a skb. 162667edfef7SAndi Kleen * This must be invoked the first time we consider transmitting 1627c1b4a7e6SDavid S. Miller * SKB onto the wire. 1628c1b4a7e6SDavid S. Miller */ 1629*5bbb432cSEric Dumazet static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now) 1630c1b4a7e6SDavid S. Miller { 1631c1b4a7e6SDavid S. Miller int tso_segs = tcp_skb_pcount(skb); 1632c1b4a7e6SDavid S. Miller 1633f8269a49SIlpo Järvinen if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { 1634*5bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 1635c1b4a7e6SDavid S. Miller tso_segs = tcp_skb_pcount(skb); 1636c1b4a7e6SDavid S. Miller } 1637c1b4a7e6SDavid S. Miller return tso_segs; 1638c1b4a7e6SDavid S. Miller } 1639c1b4a7e6SDavid S. Miller 1640c1b4a7e6SDavid S. Miller 1641a2a385d6SEric Dumazet /* Return true if the Nagle test allows this packet to be 1642c1b4a7e6SDavid S. Miller * sent now. 1643c1b4a7e6SDavid S. Miller */ 1644a2a385d6SEric Dumazet static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, 1645c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1646c1b4a7e6SDavid S. Miller { 1647c1b4a7e6SDavid S. Miller /* Nagle rule does not apply to frames, which sit in the middle of the 1648c1b4a7e6SDavid S. Miller * write_queue (they have no chances to get new data). 1649c1b4a7e6SDavid S. Miller * 1650c1b4a7e6SDavid S. Miller * This is implemented in the callers, where they modify the 'nonagle' 1651c1b4a7e6SDavid S. Miller * argument based upon the location of SKB in the send queue. 1652c1b4a7e6SDavid S. Miller */ 1653c1b4a7e6SDavid S. Miller if (nonagle & TCP_NAGLE_PUSH) 1654a2a385d6SEric Dumazet return true; 1655c1b4a7e6SDavid S. Miller 16569b44190dSYuchung Cheng /* Don't use the nagle rule for urgent data (or for the final FIN). */ 16579b44190dSYuchung Cheng if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) 1658a2a385d6SEric Dumazet return true; 1659c1b4a7e6SDavid S. Miller 1660cc93fc51SPeter Pan(潘卫平) if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle)) 1661a2a385d6SEric Dumazet return true; 1662c1b4a7e6SDavid S. Miller 1663a2a385d6SEric Dumazet return false; 1664c1b4a7e6SDavid S. Miller } 1665c1b4a7e6SDavid S. Miller 1666c1b4a7e6SDavid S. Miller /* Does at least the first segment of SKB fit into the send window? */ 1667a2a385d6SEric Dumazet static bool tcp_snd_wnd_test(const struct tcp_sock *tp, 1668a2a385d6SEric Dumazet const struct sk_buff *skb, 1669056834d9SIlpo Järvinen unsigned int cur_mss) 1670c1b4a7e6SDavid S. Miller { 1671c1b4a7e6SDavid S. Miller u32 end_seq = TCP_SKB_CB(skb)->end_seq; 1672c1b4a7e6SDavid S. Miller 1673c1b4a7e6SDavid S. Miller if (skb->len > cur_mss) 1674c1b4a7e6SDavid S. Miller end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 1675c1b4a7e6SDavid S. Miller 167690840defSIlpo Järvinen return !after(end_seq, tcp_wnd_end(tp)); 1677c1b4a7e6SDavid S. Miller } 1678c1b4a7e6SDavid S. Miller 1679fe067e8aSDavid S. Miller /* This checks if the data bearing packet SKB (usually tcp_send_head(sk)) 1680c1b4a7e6SDavid S. Miller * should be put on the wire right now. If so, it returns the number of 1681c1b4a7e6SDavid S. Miller * packets allowed by the congestion window. 1682c1b4a7e6SDavid S. Miller */ 1683cf533ea5SEric Dumazet static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb, 1684c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1685c1b4a7e6SDavid S. Miller { 1686cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1687c1b4a7e6SDavid S. Miller unsigned int cwnd_quota; 1688c1b4a7e6SDavid S. Miller 1689*5bbb432cSEric Dumazet tcp_init_tso_segs(skb, cur_mss); 1690c1b4a7e6SDavid S. Miller 1691c1b4a7e6SDavid S. Miller if (!tcp_nagle_test(tp, skb, cur_mss, nonagle)) 1692c1b4a7e6SDavid S. Miller return 0; 1693c1b4a7e6SDavid S. Miller 1694c1b4a7e6SDavid S. Miller cwnd_quota = tcp_cwnd_test(tp, skb); 1695056834d9SIlpo Järvinen if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss)) 1696c1b4a7e6SDavid S. Miller cwnd_quota = 0; 1697c1b4a7e6SDavid S. Miller 1698c1b4a7e6SDavid S. Miller return cwnd_quota; 1699c1b4a7e6SDavid S. Miller } 1700c1b4a7e6SDavid S. Miller 170167edfef7SAndi Kleen /* Test if sending is allowed right now. */ 1702a2a385d6SEric Dumazet bool tcp_may_send_now(struct sock *sk) 1703c1b4a7e6SDavid S. Miller { 1704cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1705fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 1706c1b4a7e6SDavid S. Miller 1707a02cec21SEric Dumazet return skb && 17080c54b85fSIlpo Järvinen tcp_snd_test(sk, skb, tcp_current_mss(sk), 1709c1b4a7e6SDavid S. Miller (tcp_skb_is_last(sk, skb) ? 1710a02cec21SEric Dumazet tp->nonagle : TCP_NAGLE_PUSH)); 1711c1b4a7e6SDavid S. Miller } 1712c1b4a7e6SDavid S. Miller 1713c1b4a7e6SDavid S. Miller /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 1714c1b4a7e6SDavid S. Miller * which is put after SKB on the list. It is very much like 1715c1b4a7e6SDavid S. Miller * tcp_fragment() except that it may make several kinds of assumptions 1716c1b4a7e6SDavid S. Miller * in order to speed up the splitting operation. In particular, we 1717c1b4a7e6SDavid S. Miller * know that all the data is in scatter-gather pages, and that the 1718c1b4a7e6SDavid S. Miller * packet has never been sent out before (and thus is not cloned). 1719c1b4a7e6SDavid S. Miller */ 1720056834d9SIlpo Järvinen static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, 1721c4ead4c5SEric Dumazet unsigned int mss_now, gfp_t gfp) 1722c1b4a7e6SDavid S. Miller { 1723c1b4a7e6SDavid S. Miller struct sk_buff *buff; 1724c1b4a7e6SDavid S. Miller int nlen = skb->len - len; 17259ce01461SIlpo Järvinen u8 flags; 1726c1b4a7e6SDavid S. Miller 1727c1b4a7e6SDavid S. Miller /* All of a TSO frame must be composed of paged data. */ 1728c8ac3774SHerbert Xu if (skb->len != skb->data_len) 17296cc55e09SOctavian Purdila return tcp_fragment(sk, skb, len, mss_now, gfp); 1730c1b4a7e6SDavid S. Miller 1731eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, gfp, true); 173251456b29SIan Morris if (unlikely(!buff)) 1733c1b4a7e6SDavid S. Miller return -ENOMEM; 1734c1b4a7e6SDavid S. Miller 17353ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 17363ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1737b60b49eaSHerbert Xu buff->truesize += nlen; 1738c1b4a7e6SDavid S. Miller skb->truesize -= nlen; 1739c1b4a7e6SDavid S. Miller 1740c1b4a7e6SDavid S. Miller /* Correct the sequence numbers. */ 1741c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 1742c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 1743c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 1744c1b4a7e6SDavid S. Miller 1745c1b4a7e6SDavid S. Miller /* PSH and FIN should only be set in the second packet. */ 17464de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 17474de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 17484de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1749c1b4a7e6SDavid S. Miller 1750c1b4a7e6SDavid S. Miller /* This packet was never sent out yet, so no SACK bits. */ 1751c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->sacked = 0; 1752c1b4a7e6SDavid S. Miller 175384fa7933SPatrick McHardy buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; 1754c1b4a7e6SDavid S. Miller skb_split(skb, buff, len); 1755490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 1756c1b4a7e6SDavid S. Miller 1757c1b4a7e6SDavid S. Miller /* Fix up tso_factor for both original and new SKB. */ 1758*5bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 1759*5bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 1760c1b4a7e6SDavid S. Miller 1761c1b4a7e6SDavid S. Miller /* Link BUFF into the send queue. */ 1762f4a775d1SEric Dumazet __skb_header_release(buff); 1763fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 1764c1b4a7e6SDavid S. Miller 1765c1b4a7e6SDavid S. Miller return 0; 1766c1b4a7e6SDavid S. Miller } 1767c1b4a7e6SDavid S. Miller 1768c1b4a7e6SDavid S. Miller /* Try to defer sending, if possible, in order to minimize the amount 1769c1b4a7e6SDavid S. Miller * of TSO splitting we do. View it as a kind of TSO Nagle test. 1770c1b4a7e6SDavid S. Miller * 1771c1b4a7e6SDavid S. Miller * This algorithm is from John Heffner. 1772c1b4a7e6SDavid S. Miller */ 1773ca8a2263SNeal Cardwell static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb, 1774605ad7f1SEric Dumazet bool *is_cwnd_limited, u32 max_segs) 1775c1b4a7e6SDavid S. Miller { 17766687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 177750c8339eSEric Dumazet u32 age, send_win, cong_win, limit, in_flight; 177850c8339eSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 177950c8339eSEric Dumazet struct skb_mstamp now; 178050c8339eSEric Dumazet struct sk_buff *head; 1781ad9f4f50SEric Dumazet int win_divisor; 1782c1b4a7e6SDavid S. Miller 17834de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 1784ae8064acSJohn Heffner goto send_now; 1785c1b4a7e6SDavid S. Miller 1786a0ea700eSEric Dumazet if (!((1 << icsk->icsk_ca_state) & (TCPF_CA_Open | TCPF_CA_CWR))) 1787ae8064acSJohn Heffner goto send_now; 1788ae8064acSJohn Heffner 17895f852eb5SEric Dumazet /* Avoid bursty behavior by allowing defer 17905f852eb5SEric Dumazet * only if the last write was recent. 17915f852eb5SEric Dumazet */ 17925f852eb5SEric Dumazet if ((s32)(tcp_time_stamp - tp->lsndtime) > 0) 1793ae8064acSJohn Heffner goto send_now; 1794908a75c1SDavid S. Miller 1795c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1796c1b4a7e6SDavid S. Miller 1797056834d9SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight)); 1798c1b4a7e6SDavid S. Miller 179990840defSIlpo Järvinen send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1800c1b4a7e6SDavid S. Miller 1801c1b4a7e6SDavid S. Miller /* From in_flight test above, we know that cwnd > in_flight. */ 1802c1b4a7e6SDavid S. Miller cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; 1803c1b4a7e6SDavid S. Miller 1804c1b4a7e6SDavid S. Miller limit = min(send_win, cong_win); 1805c1b4a7e6SDavid S. Miller 1806ba244fe9SDavid S. Miller /* If a full-sized TSO skb can be sent, do it. */ 1807605ad7f1SEric Dumazet if (limit >= max_segs * tp->mss_cache) 1808ae8064acSJohn Heffner goto send_now; 1809ba244fe9SDavid S. Miller 181062ad2761SIlpo Järvinen /* Middle in queue won't get any more data, full sendable already? */ 181162ad2761SIlpo Järvinen if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) 181262ad2761SIlpo Järvinen goto send_now; 181362ad2761SIlpo Järvinen 1814ad9f4f50SEric Dumazet win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor); 1815ad9f4f50SEric Dumazet if (win_divisor) { 1816c1b4a7e6SDavid S. Miller u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); 1817c1b4a7e6SDavid S. Miller 1818c1b4a7e6SDavid S. Miller /* If at least some fraction of a window is available, 1819c1b4a7e6SDavid S. Miller * just use it. 1820c1b4a7e6SDavid S. Miller */ 1821ad9f4f50SEric Dumazet chunk /= win_divisor; 1822c1b4a7e6SDavid S. Miller if (limit >= chunk) 1823ae8064acSJohn Heffner goto send_now; 1824c1b4a7e6SDavid S. Miller } else { 1825c1b4a7e6SDavid S. Miller /* Different approach, try not to defer past a single 1826c1b4a7e6SDavid S. Miller * ACK. Receiver should ACK every other full sized 1827c1b4a7e6SDavid S. Miller * frame, so if we have space for more than 3 frames 1828c1b4a7e6SDavid S. Miller * then send now. 1829c1b4a7e6SDavid S. Miller */ 18306b5a5c0dSNeal Cardwell if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) 1831ae8064acSJohn Heffner goto send_now; 1832c1b4a7e6SDavid S. Miller } 1833c1b4a7e6SDavid S. Miller 183450c8339eSEric Dumazet head = tcp_write_queue_head(sk); 183550c8339eSEric Dumazet skb_mstamp_get(&now); 183650c8339eSEric Dumazet age = skb_mstamp_us_delta(&now, &head->skb_mstamp); 183750c8339eSEric Dumazet /* If next ACK is likely to come too late (half srtt), do not defer */ 183850c8339eSEric Dumazet if (age < (tp->srtt_us >> 4)) 183950c8339eSEric Dumazet goto send_now; 184050c8339eSEric Dumazet 18415f852eb5SEric Dumazet /* Ok, it looks like it is advisable to defer. */ 1842ae8064acSJohn Heffner 1843ca8a2263SNeal Cardwell if (cong_win < send_win && cong_win < skb->len) 1844ca8a2263SNeal Cardwell *is_cwnd_limited = true; 1845ca8a2263SNeal Cardwell 1846a2a385d6SEric Dumazet return true; 1847ae8064acSJohn Heffner 1848ae8064acSJohn Heffner send_now: 1849a2a385d6SEric Dumazet return false; 1850c1b4a7e6SDavid S. Miller } 1851c1b4a7e6SDavid S. Miller 185205cbc0dbSFan Du static inline void tcp_mtu_check_reprobe(struct sock *sk) 185305cbc0dbSFan Du { 185405cbc0dbSFan Du struct inet_connection_sock *icsk = inet_csk(sk); 185505cbc0dbSFan Du struct tcp_sock *tp = tcp_sk(sk); 185605cbc0dbSFan Du struct net *net = sock_net(sk); 185705cbc0dbSFan Du u32 interval; 185805cbc0dbSFan Du s32 delta; 185905cbc0dbSFan Du 186005cbc0dbSFan Du interval = net->ipv4.sysctl_tcp_probe_interval; 186105cbc0dbSFan Du delta = tcp_time_stamp - icsk->icsk_mtup.probe_timestamp; 186205cbc0dbSFan Du if (unlikely(delta >= interval * HZ)) { 186305cbc0dbSFan Du int mss = tcp_current_mss(sk); 186405cbc0dbSFan Du 186505cbc0dbSFan Du /* Update current search range */ 186605cbc0dbSFan Du icsk->icsk_mtup.probe_size = 0; 186705cbc0dbSFan Du icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + 186805cbc0dbSFan Du sizeof(struct tcphdr) + 186905cbc0dbSFan Du icsk->icsk_af_ops->net_header_len; 187005cbc0dbSFan Du icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss); 187105cbc0dbSFan Du 187205cbc0dbSFan Du /* Update probe time stamp */ 187305cbc0dbSFan Du icsk->icsk_mtup.probe_timestamp = tcp_time_stamp; 187405cbc0dbSFan Du } 187505cbc0dbSFan Du } 187605cbc0dbSFan Du 18775d424d5aSJohn Heffner /* Create a new MTU probe if we are ready. 187867edfef7SAndi Kleen * MTU probe is regularly attempting to increase the path MTU by 187967edfef7SAndi Kleen * deliberately sending larger packets. This discovers routing 188067edfef7SAndi Kleen * changes resulting in larger path MTUs. 188167edfef7SAndi Kleen * 18825d424d5aSJohn Heffner * Returns 0 if we should wait to probe (no cwnd available), 18835d424d5aSJohn Heffner * 1 if a probe was sent, 1884056834d9SIlpo Järvinen * -1 otherwise 1885056834d9SIlpo Järvinen */ 18865d424d5aSJohn Heffner static int tcp_mtu_probe(struct sock *sk) 18875d424d5aSJohn Heffner { 18885d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 18895d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 18905d424d5aSJohn Heffner struct sk_buff *skb, *nskb, *next; 18916b58e0a5SFan Du struct net *net = sock_net(sk); 18925d424d5aSJohn Heffner int len; 18935d424d5aSJohn Heffner int probe_size; 189491cc17c0SIlpo Järvinen int size_needed; 18955d424d5aSJohn Heffner int copy; 18965d424d5aSJohn Heffner int mss_now; 18976b58e0a5SFan Du int interval; 18985d424d5aSJohn Heffner 18995d424d5aSJohn Heffner /* Not currently probing/verifying, 19005d424d5aSJohn Heffner * not in recovery, 19015d424d5aSJohn Heffner * have enough cwnd, and 19025d424d5aSJohn Heffner * not SACKing (the variable headers throw things off) */ 19035d424d5aSJohn Heffner if (!icsk->icsk_mtup.enabled || 19045d424d5aSJohn Heffner icsk->icsk_mtup.probe_size || 19055d424d5aSJohn Heffner inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 19065d424d5aSJohn Heffner tp->snd_cwnd < 11 || 1907cabeccbdSIlpo Järvinen tp->rx_opt.num_sacks || tp->rx_opt.dsack) 19085d424d5aSJohn Heffner return -1; 19095d424d5aSJohn Heffner 19106b58e0a5SFan Du /* Use binary search for probe_size between tcp_mss_base, 19116b58e0a5SFan Du * and current mss_clamp. if (search_high - search_low) 19126b58e0a5SFan Du * smaller than a threshold, backoff from probing. 19136b58e0a5SFan Du */ 19140c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 19156b58e0a5SFan Du probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high + 19166b58e0a5SFan Du icsk->icsk_mtup.search_low) >> 1); 191791cc17c0SIlpo Järvinen size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; 19186b58e0a5SFan Du interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low; 191905cbc0dbSFan Du /* When misfortune happens, we are reprobing actively, 192005cbc0dbSFan Du * and then reprobe timer has expired. We stick with current 192105cbc0dbSFan Du * probing process by not resetting search range to its orignal. 192205cbc0dbSFan Du */ 19236b58e0a5SFan Du if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) || 192405cbc0dbSFan Du interval < net->ipv4.sysctl_tcp_probe_threshold) { 192505cbc0dbSFan Du /* Check whether enough time has elaplased for 192605cbc0dbSFan Du * another round of probing. 192705cbc0dbSFan Du */ 192805cbc0dbSFan Du tcp_mtu_check_reprobe(sk); 19295d424d5aSJohn Heffner return -1; 19305d424d5aSJohn Heffner } 19315d424d5aSJohn Heffner 19325d424d5aSJohn Heffner /* Have enough data in the send queue to probe? */ 19337f9c33e5SIlpo Järvinen if (tp->write_seq - tp->snd_nxt < size_needed) 19345d424d5aSJohn Heffner return -1; 19355d424d5aSJohn Heffner 193691cc17c0SIlpo Järvinen if (tp->snd_wnd < size_needed) 19375d424d5aSJohn Heffner return -1; 193890840defSIlpo Järvinen if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) 19395d424d5aSJohn Heffner return 0; 19405d424d5aSJohn Heffner 1941d67c58e9SIlpo Järvinen /* Do we need to wait to drain cwnd? With none in flight, don't stall */ 1942d67c58e9SIlpo Järvinen if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { 1943d67c58e9SIlpo Järvinen if (!tcp_packets_in_flight(tp)) 19445d424d5aSJohn Heffner return -1; 19455d424d5aSJohn Heffner else 19465d424d5aSJohn Heffner return 0; 19475d424d5aSJohn Heffner } 19485d424d5aSJohn Heffner 19495d424d5aSJohn Heffner /* We're allowed to probe. Build it now. */ 1950eb934478SEric Dumazet nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false); 195151456b29SIan Morris if (!nskb) 19525d424d5aSJohn Heffner return -1; 19533ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 19543ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 19555d424d5aSJohn Heffner 1956fe067e8aSDavid S. Miller skb = tcp_send_head(sk); 19575d424d5aSJohn Heffner 19585d424d5aSJohn Heffner TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 19595d424d5aSJohn Heffner TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 19604de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; 19615d424d5aSJohn Heffner TCP_SKB_CB(nskb)->sacked = 0; 19625d424d5aSJohn Heffner nskb->csum = 0; 196384fa7933SPatrick McHardy nskb->ip_summed = skb->ip_summed; 19645d424d5aSJohn Heffner 196550c4817eSIlpo Järvinen tcp_insert_write_queue_before(nskb, skb, sk); 196650c4817eSIlpo Järvinen 19675d424d5aSJohn Heffner len = 0; 1968234b6860SIlpo Järvinen tcp_for_write_queue_from_safe(skb, next, sk) { 19695d424d5aSJohn Heffner copy = min_t(int, skb->len, probe_size - len); 19705d424d5aSJohn Heffner if (nskb->ip_summed) 19715d424d5aSJohn Heffner skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); 19725d424d5aSJohn Heffner else 19735d424d5aSJohn Heffner nskb->csum = skb_copy_and_csum_bits(skb, 0, 1974056834d9SIlpo Järvinen skb_put(nskb, copy), 1975056834d9SIlpo Järvinen copy, nskb->csum); 19765d424d5aSJohn Heffner 19775d424d5aSJohn Heffner if (skb->len <= copy) { 19785d424d5aSJohn Heffner /* We've eaten all the data from this skb. 19795d424d5aSJohn Heffner * Throw it away. */ 19804de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 1981fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 19823ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 19835d424d5aSJohn Heffner } else { 19844de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & 1985a3433f35SChangli Gao ~(TCPHDR_FIN|TCPHDR_PSH); 19865d424d5aSJohn Heffner if (!skb_shinfo(skb)->nr_frags) { 19875d424d5aSJohn Heffner skb_pull(skb, copy); 198884fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 1989056834d9SIlpo Järvinen skb->csum = csum_partial(skb->data, 1990056834d9SIlpo Järvinen skb->len, 0); 19915d424d5aSJohn Heffner } else { 19925d424d5aSJohn Heffner __pskb_trim_head(skb, copy); 1993*5bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 19945d424d5aSJohn Heffner } 19955d424d5aSJohn Heffner TCP_SKB_CB(skb)->seq += copy; 19965d424d5aSJohn Heffner } 19975d424d5aSJohn Heffner 19985d424d5aSJohn Heffner len += copy; 1999234b6860SIlpo Järvinen 2000234b6860SIlpo Järvinen if (len >= probe_size) 2001234b6860SIlpo Järvinen break; 20025d424d5aSJohn Heffner } 2003*5bbb432cSEric Dumazet tcp_init_tso_segs(nskb, nskb->len); 20045d424d5aSJohn Heffner 20055d424d5aSJohn Heffner /* We're ready to send. If this fails, the probe will 20067faee5c0SEric Dumazet * be resegmented into mss-sized pieces by tcp_write_xmit(). 20077faee5c0SEric Dumazet */ 20085d424d5aSJohn Heffner if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 20095d424d5aSJohn Heffner /* Decrement cwnd here because we are sending 20105d424d5aSJohn Heffner * effectively two packets. */ 20115d424d5aSJohn Heffner tp->snd_cwnd--; 201266f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, nskb); 20135d424d5aSJohn Heffner 20145d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 20150e7b1368SJohn Heffner tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 20160e7b1368SJohn Heffner tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 20175d424d5aSJohn Heffner 20185d424d5aSJohn Heffner return 1; 20195d424d5aSJohn Heffner } 20205d424d5aSJohn Heffner 20215d424d5aSJohn Heffner return -1; 20225d424d5aSJohn Heffner } 20235d424d5aSJohn Heffner 20241da177e4SLinus Torvalds /* This routine writes packets to the network. It advances the 20251da177e4SLinus Torvalds * send_head. This happens as incoming acks open up the remote 20261da177e4SLinus Torvalds * window for us. 20271da177e4SLinus Torvalds * 2028f8269a49SIlpo Järvinen * LARGESEND note: !tcp_urg_mode is overkill, only frames between 2029f8269a49SIlpo Järvinen * snd_up-64k-mss .. snd_up cannot be large. However, taking into 2030f8269a49SIlpo Järvinen * account rare use of URG, this is not a big flaw. 2031f8269a49SIlpo Järvinen * 20326ba8a3b1SNandita Dukkipati * Send at most one packet when push_one > 0. Temporarily ignore 20336ba8a3b1SNandita Dukkipati * cwnd limit to force at most one packet out when push_one == 2. 20346ba8a3b1SNandita Dukkipati 2035a2a385d6SEric Dumazet * Returns true, if no segments are in flight and we have queued segments, 2036a2a385d6SEric Dumazet * but cannot send anything now because of SWS or another problem. 20371da177e4SLinus Torvalds */ 2038a2a385d6SEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 2039d5dd9175SIlpo Järvinen int push_one, gfp_t gfp) 20401da177e4SLinus Torvalds { 20411da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 204292df7b51SDavid S. Miller struct sk_buff *skb; 2043c1b4a7e6SDavid S. Miller unsigned int tso_segs, sent_pkts; 2044c1b4a7e6SDavid S. Miller int cwnd_quota; 20455d424d5aSJohn Heffner int result; 2046ca8a2263SNeal Cardwell bool is_cwnd_limited = false; 2047605ad7f1SEric Dumazet u32 max_segs; 20481da177e4SLinus Torvalds 2049c1b4a7e6SDavid S. Miller sent_pkts = 0; 20505d424d5aSJohn Heffner 2051d5dd9175SIlpo Järvinen if (!push_one) { 20525d424d5aSJohn Heffner /* Do MTU probing. */ 2053d5dd9175SIlpo Järvinen result = tcp_mtu_probe(sk); 2054d5dd9175SIlpo Järvinen if (!result) { 2055a2a385d6SEric Dumazet return false; 20565d424d5aSJohn Heffner } else if (result > 0) { 20575d424d5aSJohn Heffner sent_pkts = 1; 20585d424d5aSJohn Heffner } 2059d5dd9175SIlpo Järvinen } 20605d424d5aSJohn Heffner 2061605ad7f1SEric Dumazet max_segs = tcp_tso_autosize(sk, mss_now); 2062fe067e8aSDavid S. Miller while ((skb = tcp_send_head(sk))) { 2063c8ac3774SHerbert Xu unsigned int limit; 2064c8ac3774SHerbert Xu 2065*5bbb432cSEric Dumazet tso_segs = tcp_init_tso_segs(skb, mss_now); 2066c1b4a7e6SDavid S. Miller BUG_ON(!tso_segs); 2067c1b4a7e6SDavid S. Miller 20689d186cacSAndrey Vagin if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) { 20697faee5c0SEric Dumazet /* "skb_mstamp" is used as a start point for the retransmit timer */ 20707faee5c0SEric Dumazet skb_mstamp_get(&skb->skb_mstamp); 2071ec342325SAndrew Vagin goto repair; /* Skip network transmission */ 20729d186cacSAndrey Vagin } 2073ec342325SAndrew Vagin 2074b68e9f85SHerbert Xu cwnd_quota = tcp_cwnd_test(tp, skb); 20756ba8a3b1SNandita Dukkipati if (!cwnd_quota) { 2076ca8a2263SNeal Cardwell is_cwnd_limited = true; 20776ba8a3b1SNandita Dukkipati if (push_one == 2) 20786ba8a3b1SNandita Dukkipati /* Force out a loss probe pkt. */ 20796ba8a3b1SNandita Dukkipati cwnd_quota = 1; 20806ba8a3b1SNandita Dukkipati else 2081b68e9f85SHerbert Xu break; 20826ba8a3b1SNandita Dukkipati } 2083b68e9f85SHerbert Xu 2084b68e9f85SHerbert Xu if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) 2085b68e9f85SHerbert Xu break; 2086b68e9f85SHerbert Xu 2087d6a4e26aSEric Dumazet if (tso_segs == 1) { 2088aa93466bSDavid S. Miller if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 2089aa93466bSDavid S. Miller (tcp_skb_is_last(sk, skb) ? 2090aa93466bSDavid S. Miller nonagle : TCP_NAGLE_PUSH)))) 2091aa93466bSDavid S. Miller break; 2092c1b4a7e6SDavid S. Miller } else { 2093ca8a2263SNeal Cardwell if (!push_one && 2094605ad7f1SEric Dumazet tcp_tso_should_defer(sk, skb, &is_cwnd_limited, 2095605ad7f1SEric Dumazet max_segs)) 2096aa93466bSDavid S. Miller break; 2097c1b4a7e6SDavid S. Miller } 2098aa93466bSDavid S. Miller 2099605ad7f1SEric Dumazet limit = mss_now; 2100d6a4e26aSEric Dumazet if (tso_segs > 1 && !tcp_urg_mode(tp)) 2101605ad7f1SEric Dumazet limit = tcp_mss_split_point(sk, skb, mss_now, 2102605ad7f1SEric Dumazet min_t(unsigned int, 2103605ad7f1SEric Dumazet cwnd_quota, 2104605ad7f1SEric Dumazet max_segs), 2105605ad7f1SEric Dumazet nonagle); 2106605ad7f1SEric Dumazet 2107605ad7f1SEric Dumazet if (skb->len > limit && 2108605ad7f1SEric Dumazet unlikely(tso_fragment(sk, skb, limit, mss_now, gfp))) 2109605ad7f1SEric Dumazet break; 2110605ad7f1SEric Dumazet 2111c9eeec26SEric Dumazet /* TCP Small Queues : 2112c9eeec26SEric Dumazet * Control number of packets in qdisc/devices to two packets / or ~1 ms. 2113c9eeec26SEric Dumazet * This allows for : 2114c9eeec26SEric Dumazet * - better RTT estimation and ACK scheduling 2115c9eeec26SEric Dumazet * - faster recovery 2116c9eeec26SEric Dumazet * - high rates 211798e09386SEric Dumazet * Alas, some drivers / subsystems require a fair amount 211898e09386SEric Dumazet * of queued bytes to ensure line rate. 211998e09386SEric Dumazet * One example is wifi aggregation (802.11 AMPDU) 212046d3ceabSEric Dumazet */ 2121605ad7f1SEric Dumazet limit = max(2 * skb->truesize, sk->sk_pacing_rate >> 10); 2122605ad7f1SEric Dumazet limit = min_t(u32, limit, sysctl_tcp_limit_output_bytes); 2123c9eeec26SEric Dumazet 2124c9eeec26SEric Dumazet if (atomic_read(&sk->sk_wmem_alloc) > limit) { 212546d3ceabSEric Dumazet set_bit(TSQ_THROTTLED, &tp->tsq_flags); 2126bf06200eSJohn Ogness /* It is possible TX completion already happened 2127bf06200eSJohn Ogness * before we set TSQ_THROTTLED, so we must 2128bf06200eSJohn Ogness * test again the condition. 2129bf06200eSJohn Ogness */ 21304e857c58SPeter Zijlstra smp_mb__after_atomic(); 2131bf06200eSJohn Ogness if (atomic_read(&sk->sk_wmem_alloc) > limit) 213246d3ceabSEric Dumazet break; 213346d3ceabSEric Dumazet } 2134c9eeec26SEric Dumazet 2135d5dd9175SIlpo Järvinen if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) 21361da177e4SLinus Torvalds break; 21371da177e4SLinus Torvalds 2138ec342325SAndrew Vagin repair: 21391da177e4SLinus Torvalds /* Advance the send_head. This one is sent out. 21401da177e4SLinus Torvalds * This call will increment packets_out. 21411da177e4SLinus Torvalds */ 214266f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 21431da177e4SLinus Torvalds 21441da177e4SLinus Torvalds tcp_minshall_update(tp, mss_now, skb); 2145a262f0cdSNandita Dukkipati sent_pkts += tcp_skb_pcount(skb); 2146d5dd9175SIlpo Järvinen 2147d5dd9175SIlpo Järvinen if (push_one) 2148d5dd9175SIlpo Järvinen break; 21491da177e4SLinus Torvalds } 21501da177e4SLinus Torvalds 2151aa93466bSDavid S. Miller if (likely(sent_pkts)) { 2152684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 2153684bad11SYuchung Cheng tp->prr_out += sent_pkts; 21546ba8a3b1SNandita Dukkipati 21556ba8a3b1SNandita Dukkipati /* Send one loss probe per tail loss episode. */ 21566ba8a3b1SNandita Dukkipati if (push_one != 2) 21576ba8a3b1SNandita Dukkipati tcp_schedule_loss_probe(sk); 2158ca8a2263SNeal Cardwell tcp_cwnd_validate(sk, is_cwnd_limited); 2159a2a385d6SEric Dumazet return false; 21601da177e4SLinus Torvalds } 21616ba8a3b1SNandita Dukkipati return (push_one == 2) || (!tp->packets_out && tcp_send_head(sk)); 21626ba8a3b1SNandita Dukkipati } 21636ba8a3b1SNandita Dukkipati 21646ba8a3b1SNandita Dukkipati bool tcp_schedule_loss_probe(struct sock *sk) 21656ba8a3b1SNandita Dukkipati { 21666ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 21676ba8a3b1SNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 21686ba8a3b1SNandita Dukkipati u32 timeout, tlp_time_stamp, rto_time_stamp; 2169740b0f18SEric Dumazet u32 rtt = usecs_to_jiffies(tp->srtt_us >> 3); 21706ba8a3b1SNandita Dukkipati 21716ba8a3b1SNandita Dukkipati if (WARN_ON(icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS)) 21726ba8a3b1SNandita Dukkipati return false; 21736ba8a3b1SNandita Dukkipati /* No consecutive loss probes. */ 21746ba8a3b1SNandita Dukkipati if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) { 21756ba8a3b1SNandita Dukkipati tcp_rearm_rto(sk); 21766ba8a3b1SNandita Dukkipati return false; 21776ba8a3b1SNandita Dukkipati } 21786ba8a3b1SNandita Dukkipati /* Don't do any loss probe on a Fast Open connection before 3WHS 21796ba8a3b1SNandita Dukkipati * finishes. 21806ba8a3b1SNandita Dukkipati */ 21816ba8a3b1SNandita Dukkipati if (sk->sk_state == TCP_SYN_RECV) 21826ba8a3b1SNandita Dukkipati return false; 21836ba8a3b1SNandita Dukkipati 21846ba8a3b1SNandita Dukkipati /* TLP is only scheduled when next timer event is RTO. */ 21856ba8a3b1SNandita Dukkipati if (icsk->icsk_pending != ICSK_TIME_RETRANS) 21866ba8a3b1SNandita Dukkipati return false; 21876ba8a3b1SNandita Dukkipati 21886ba8a3b1SNandita Dukkipati /* Schedule a loss probe in 2*RTT for SACK capable connections 21896ba8a3b1SNandita Dukkipati * in Open state, that are either limited by cwnd or application. 21906ba8a3b1SNandita Dukkipati */ 2191740b0f18SEric Dumazet if (sysctl_tcp_early_retrans < 3 || !tp->srtt_us || !tp->packets_out || 21926ba8a3b1SNandita Dukkipati !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open) 21936ba8a3b1SNandita Dukkipati return false; 21946ba8a3b1SNandita Dukkipati 21956ba8a3b1SNandita Dukkipati if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) && 21966ba8a3b1SNandita Dukkipati tcp_send_head(sk)) 21976ba8a3b1SNandita Dukkipati return false; 21986ba8a3b1SNandita Dukkipati 21996ba8a3b1SNandita Dukkipati /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account 22006ba8a3b1SNandita Dukkipati * for delayed ack when there's one outstanding packet. 22016ba8a3b1SNandita Dukkipati */ 22026ba8a3b1SNandita Dukkipati timeout = rtt << 1; 22036ba8a3b1SNandita Dukkipati if (tp->packets_out == 1) 22046ba8a3b1SNandita Dukkipati timeout = max_t(u32, timeout, 22056ba8a3b1SNandita Dukkipati (rtt + (rtt >> 1) + TCP_DELACK_MAX)); 22066ba8a3b1SNandita Dukkipati timeout = max_t(u32, timeout, msecs_to_jiffies(10)); 22076ba8a3b1SNandita Dukkipati 22086ba8a3b1SNandita Dukkipati /* If RTO is shorter, just schedule TLP in its place. */ 22096ba8a3b1SNandita Dukkipati tlp_time_stamp = tcp_time_stamp + timeout; 22106ba8a3b1SNandita Dukkipati rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout; 22116ba8a3b1SNandita Dukkipati if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) { 22126ba8a3b1SNandita Dukkipati s32 delta = rto_time_stamp - tcp_time_stamp; 22136ba8a3b1SNandita Dukkipati if (delta > 0) 22146ba8a3b1SNandita Dukkipati timeout = delta; 22156ba8a3b1SNandita Dukkipati } 22166ba8a3b1SNandita Dukkipati 22176ba8a3b1SNandita Dukkipati inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, 22186ba8a3b1SNandita Dukkipati TCP_RTO_MAX); 22196ba8a3b1SNandita Dukkipati return true; 22206ba8a3b1SNandita Dukkipati } 22216ba8a3b1SNandita Dukkipati 22221f3279aeSEric Dumazet /* Thanks to skb fast clones, we can detect if a prior transmit of 22231f3279aeSEric Dumazet * a packet is still in a qdisc or driver queue. 22241f3279aeSEric Dumazet * In this case, there is very little point doing a retransmit ! 22251f3279aeSEric Dumazet * Note: This is called from BH context only. 22261f3279aeSEric Dumazet */ 22271f3279aeSEric Dumazet static bool skb_still_in_host_queue(const struct sock *sk, 22281f3279aeSEric Dumazet const struct sk_buff *skb) 22291f3279aeSEric Dumazet { 223039bb5e62SEric Dumazet if (unlikely(skb_fclone_busy(sk, skb))) { 22311f3279aeSEric Dumazet NET_INC_STATS_BH(sock_net(sk), 22321f3279aeSEric Dumazet LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES); 22331f3279aeSEric Dumazet return true; 22341f3279aeSEric Dumazet } 22351f3279aeSEric Dumazet return false; 22361f3279aeSEric Dumazet } 22371f3279aeSEric Dumazet 22386ba8a3b1SNandita Dukkipati /* When probe timeout (PTO) fires, send a new segment if one exists, else 22396ba8a3b1SNandita Dukkipati * retransmit the last segment. 22406ba8a3b1SNandita Dukkipati */ 22416ba8a3b1SNandita Dukkipati void tcp_send_loss_probe(struct sock *sk) 22426ba8a3b1SNandita Dukkipati { 22439b717a8dSNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 22446ba8a3b1SNandita Dukkipati struct sk_buff *skb; 22456ba8a3b1SNandita Dukkipati int pcount; 22466ba8a3b1SNandita Dukkipati int mss = tcp_current_mss(sk); 22476ba8a3b1SNandita Dukkipati int err = -1; 22486ba8a3b1SNandita Dukkipati 224900db4124SIan Morris if (tcp_send_head(sk)) { 22506ba8a3b1SNandita Dukkipati err = tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC); 22516ba8a3b1SNandita Dukkipati goto rearm_timer; 22526ba8a3b1SNandita Dukkipati } 22536ba8a3b1SNandita Dukkipati 22549b717a8dSNandita Dukkipati /* At most one outstanding TLP retransmission. */ 22559b717a8dSNandita Dukkipati if (tp->tlp_high_seq) 22569b717a8dSNandita Dukkipati goto rearm_timer; 22579b717a8dSNandita Dukkipati 22586ba8a3b1SNandita Dukkipati /* Retransmit last segment. */ 22596ba8a3b1SNandita Dukkipati skb = tcp_write_queue_tail(sk); 22606ba8a3b1SNandita Dukkipati if (WARN_ON(!skb)) 22616ba8a3b1SNandita Dukkipati goto rearm_timer; 22626ba8a3b1SNandita Dukkipati 22631f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 22641f3279aeSEric Dumazet goto rearm_timer; 22651f3279aeSEric Dumazet 22666ba8a3b1SNandita Dukkipati pcount = tcp_skb_pcount(skb); 22676ba8a3b1SNandita Dukkipati if (WARN_ON(!pcount)) 22686ba8a3b1SNandita Dukkipati goto rearm_timer; 22696ba8a3b1SNandita Dukkipati 22706ba8a3b1SNandita Dukkipati if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) { 22716cc55e09SOctavian Purdila if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss, 22726cc55e09SOctavian Purdila GFP_ATOMIC))) 22736ba8a3b1SNandita Dukkipati goto rearm_timer; 22746ba8a3b1SNandita Dukkipati skb = tcp_write_queue_tail(sk); 22756ba8a3b1SNandita Dukkipati } 22766ba8a3b1SNandita Dukkipati 22776ba8a3b1SNandita Dukkipati if (WARN_ON(!skb || !tcp_skb_pcount(skb))) 22786ba8a3b1SNandita Dukkipati goto rearm_timer; 22796ba8a3b1SNandita Dukkipati 22806ba8a3b1SNandita Dukkipati err = __tcp_retransmit_skb(sk, skb); 22816ba8a3b1SNandita Dukkipati 22829b717a8dSNandita Dukkipati /* Record snd_nxt for loss detection. */ 22839b717a8dSNandita Dukkipati if (likely(!err)) 22849b717a8dSNandita Dukkipati tp->tlp_high_seq = tp->snd_nxt; 22859b717a8dSNandita Dukkipati 22866ba8a3b1SNandita Dukkipati rearm_timer: 22876ba8a3b1SNandita Dukkipati inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 22886ba8a3b1SNandita Dukkipati inet_csk(sk)->icsk_rto, 22896ba8a3b1SNandita Dukkipati TCP_RTO_MAX); 22906ba8a3b1SNandita Dukkipati 22916ba8a3b1SNandita Dukkipati if (likely(!err)) 22926ba8a3b1SNandita Dukkipati NET_INC_STATS_BH(sock_net(sk), 22936ba8a3b1SNandita Dukkipati LINUX_MIB_TCPLOSSPROBES); 22941da177e4SLinus Torvalds } 22951da177e4SLinus Torvalds 2296a762a980SDavid S. Miller /* Push out any pending frames which were held back due to 2297a762a980SDavid S. Miller * TCP_CORK or attempt at coalescing tiny packets. 2298a762a980SDavid S. Miller * The socket must be locked by the caller. 2299a762a980SDavid S. Miller */ 23009e412ba7SIlpo Järvinen void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 23019e412ba7SIlpo Järvinen int nonagle) 2302a762a980SDavid S. Miller { 2303726e07a8SIlpo Järvinen /* If we are closed, the bytes will have to remain here. 2304726e07a8SIlpo Järvinen * In time closedown will finish, we empty the write queue and 2305726e07a8SIlpo Järvinen * all will be happy. 2306726e07a8SIlpo Järvinen */ 2307726e07a8SIlpo Järvinen if (unlikely(sk->sk_state == TCP_CLOSE)) 2308726e07a8SIlpo Järvinen return; 2309726e07a8SIlpo Järvinen 231099a1dec7SMel Gorman if (tcp_write_xmit(sk, cur_mss, nonagle, 0, 231199a1dec7SMel Gorman sk_gfp_atomic(sk, GFP_ATOMIC))) 23129e412ba7SIlpo Järvinen tcp_check_probe_timer(sk); 2313a762a980SDavid S. Miller } 2314a762a980SDavid S. Miller 2315c1b4a7e6SDavid S. Miller /* Send _single_ skb sitting at the send head. This function requires 2316c1b4a7e6SDavid S. Miller * true push pending frames to setup probe timer etc. 2317c1b4a7e6SDavid S. Miller */ 2318c1b4a7e6SDavid S. Miller void tcp_push_one(struct sock *sk, unsigned int mss_now) 2319c1b4a7e6SDavid S. Miller { 2320fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 2321c1b4a7e6SDavid S. Miller 2322c1b4a7e6SDavid S. Miller BUG_ON(!skb || skb->len < mss_now); 2323c1b4a7e6SDavid S. Miller 2324d5dd9175SIlpo Järvinen tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); 2325c1b4a7e6SDavid S. Miller } 2326c1b4a7e6SDavid S. Miller 23271da177e4SLinus Torvalds /* This function returns the amount that we can raise the 23281da177e4SLinus Torvalds * usable window based on the following constraints 23291da177e4SLinus Torvalds * 23301da177e4SLinus Torvalds * 1. The window can never be shrunk once it is offered (RFC 793) 23311da177e4SLinus Torvalds * 2. We limit memory per socket 23321da177e4SLinus Torvalds * 23331da177e4SLinus Torvalds * RFC 1122: 23341da177e4SLinus Torvalds * "the suggested [SWS] avoidance algorithm for the receiver is to keep 23351da177e4SLinus Torvalds * RECV.NEXT + RCV.WIN fixed until: 23361da177e4SLinus Torvalds * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 23371da177e4SLinus Torvalds * 23381da177e4SLinus Torvalds * i.e. don't raise the right edge of the window until you can raise 23391da177e4SLinus Torvalds * it at least MSS bytes. 23401da177e4SLinus Torvalds * 23411da177e4SLinus Torvalds * Unfortunately, the recommended algorithm breaks header prediction, 23421da177e4SLinus Torvalds * since header prediction assumes th->window stays fixed. 23431da177e4SLinus Torvalds * 23441da177e4SLinus Torvalds * Strictly speaking, keeping th->window fixed violates the receiver 23451da177e4SLinus Torvalds * side SWS prevention criteria. The problem is that under this rule 23461da177e4SLinus Torvalds * a stream of single byte packets will cause the right side of the 23471da177e4SLinus Torvalds * window to always advance by a single byte. 23481da177e4SLinus Torvalds * 23491da177e4SLinus Torvalds * Of course, if the sender implements sender side SWS prevention 23501da177e4SLinus Torvalds * then this will not be a problem. 23511da177e4SLinus Torvalds * 23521da177e4SLinus Torvalds * BSD seems to make the following compromise: 23531da177e4SLinus Torvalds * 23541da177e4SLinus Torvalds * If the free space is less than the 1/4 of the maximum 23551da177e4SLinus Torvalds * space available and the free space is less than 1/2 mss, 23561da177e4SLinus Torvalds * then set the window to 0. 23571da177e4SLinus Torvalds * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 23581da177e4SLinus Torvalds * Otherwise, just prevent the window from shrinking 23591da177e4SLinus Torvalds * and from being larger than the largest representable value. 23601da177e4SLinus Torvalds * 23611da177e4SLinus Torvalds * This prevents incremental opening of the window in the regime 23621da177e4SLinus Torvalds * where TCP is limited by the speed of the reader side taking 23631da177e4SLinus Torvalds * data out of the TCP receive queue. It does nothing about 23641da177e4SLinus Torvalds * those cases where the window is constrained on the sender side 23651da177e4SLinus Torvalds * because the pipeline is full. 23661da177e4SLinus Torvalds * 23671da177e4SLinus Torvalds * BSD also seems to "accidentally" limit itself to windows that are a 23681da177e4SLinus Torvalds * multiple of MSS, at least until the free space gets quite small. 23691da177e4SLinus Torvalds * This would appear to be a side effect of the mbuf implementation. 23701da177e4SLinus Torvalds * Combining these two algorithms results in the observed behavior 23711da177e4SLinus Torvalds * of having a fixed window size at almost all times. 23721da177e4SLinus Torvalds * 23731da177e4SLinus Torvalds * Below we obtain similar behavior by forcing the offered window to 23741da177e4SLinus Torvalds * a multiple of the mss when it is feasible to do so. 23751da177e4SLinus Torvalds * 23761da177e4SLinus Torvalds * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 23771da177e4SLinus Torvalds * Regular options like TIMESTAMP are taken into account. 23781da177e4SLinus Torvalds */ 23791da177e4SLinus Torvalds u32 __tcp_select_window(struct sock *sk) 23801da177e4SLinus Torvalds { 2381463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 23821da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2383caa20d9aSStephen Hemminger /* MSS for the peer's data. Previous versions used mss_clamp 23841da177e4SLinus Torvalds * here. I don't know if the value based on our guesses 23851da177e4SLinus Torvalds * of peer's MSS is better for the performance. It's more correct 23861da177e4SLinus Torvalds * but may be worse for the performance because of rcv_mss 23871da177e4SLinus Torvalds * fluctuations. --SAW 1998/11/1 23881da177e4SLinus Torvalds */ 2389463c84b9SArnaldo Carvalho de Melo int mss = icsk->icsk_ack.rcv_mss; 23901da177e4SLinus Torvalds int free_space = tcp_space(sk); 239186c1a045SFlorian Westphal int allowed_space = tcp_full_space(sk); 239286c1a045SFlorian Westphal int full_space = min_t(int, tp->window_clamp, allowed_space); 23931da177e4SLinus Torvalds int window; 23941da177e4SLinus Torvalds 23951da177e4SLinus Torvalds if (mss > full_space) 23961da177e4SLinus Torvalds mss = full_space; 23971da177e4SLinus Torvalds 2398b92edbe0SEric Dumazet if (free_space < (full_space >> 1)) { 2399463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.quick = 0; 24001da177e4SLinus Torvalds 2401b8da51ebSEric Dumazet if (tcp_under_memory_pressure(sk)) 2402056834d9SIlpo Järvinen tp->rcv_ssthresh = min(tp->rcv_ssthresh, 2403056834d9SIlpo Järvinen 4U * tp->advmss); 24041da177e4SLinus Torvalds 240586c1a045SFlorian Westphal /* free_space might become our new window, make sure we don't 240686c1a045SFlorian Westphal * increase it due to wscale. 240786c1a045SFlorian Westphal */ 240886c1a045SFlorian Westphal free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale); 240986c1a045SFlorian Westphal 241086c1a045SFlorian Westphal /* if free space is less than mss estimate, or is below 1/16th 241186c1a045SFlorian Westphal * of the maximum allowed, try to move to zero-window, else 241286c1a045SFlorian Westphal * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and 241386c1a045SFlorian Westphal * new incoming data is dropped due to memory limits. 241486c1a045SFlorian Westphal * With large window, mss test triggers way too late in order 241586c1a045SFlorian Westphal * to announce zero window in time before rmem limit kicks in. 241686c1a045SFlorian Westphal */ 241786c1a045SFlorian Westphal if (free_space < (allowed_space >> 4) || free_space < mss) 24181da177e4SLinus Torvalds return 0; 24191da177e4SLinus Torvalds } 24201da177e4SLinus Torvalds 24211da177e4SLinus Torvalds if (free_space > tp->rcv_ssthresh) 24221da177e4SLinus Torvalds free_space = tp->rcv_ssthresh; 24231da177e4SLinus Torvalds 24241da177e4SLinus Torvalds /* Don't do rounding if we are using window scaling, since the 24251da177e4SLinus Torvalds * scaled window will not line up with the MSS boundary anyway. 24261da177e4SLinus Torvalds */ 24271da177e4SLinus Torvalds window = tp->rcv_wnd; 24281da177e4SLinus Torvalds if (tp->rx_opt.rcv_wscale) { 24291da177e4SLinus Torvalds window = free_space; 24301da177e4SLinus Torvalds 24311da177e4SLinus Torvalds /* Advertise enough space so that it won't get scaled away. 24321da177e4SLinus Torvalds * Import case: prevent zero window announcement if 24331da177e4SLinus Torvalds * 1<<rcv_wscale > mss. 24341da177e4SLinus Torvalds */ 24351da177e4SLinus Torvalds if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window) 24361da177e4SLinus Torvalds window = (((window >> tp->rx_opt.rcv_wscale) + 1) 24371da177e4SLinus Torvalds << tp->rx_opt.rcv_wscale); 24381da177e4SLinus Torvalds } else { 24391da177e4SLinus Torvalds /* Get the largest window that is a nice multiple of mss. 24401da177e4SLinus Torvalds * Window clamp already applied above. 24411da177e4SLinus Torvalds * If our current window offering is within 1 mss of the 24421da177e4SLinus Torvalds * free space we just keep it. This prevents the divide 24431da177e4SLinus Torvalds * and multiply from happening most of the time. 24441da177e4SLinus Torvalds * We also don't do any window rounding when the free space 24451da177e4SLinus Torvalds * is too small. 24461da177e4SLinus Torvalds */ 24471da177e4SLinus Torvalds if (window <= free_space - mss || window > free_space) 24481da177e4SLinus Torvalds window = (free_space / mss) * mss; 244984565070SJohn Heffner else if (mss == full_space && 2450b92edbe0SEric Dumazet free_space > window + (full_space >> 1)) 245184565070SJohn Heffner window = free_space; 24521da177e4SLinus Torvalds } 24531da177e4SLinus Torvalds 24541da177e4SLinus Torvalds return window; 24551da177e4SLinus Torvalds } 24561da177e4SLinus Torvalds 24574a17fc3aSIlpo Järvinen /* Collapses two adjacent SKB's during retransmission. */ 24584a17fc3aSIlpo Järvinen static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) 24591da177e4SLinus Torvalds { 24601da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2461fe067e8aSDavid S. Miller struct sk_buff *next_skb = tcp_write_queue_next(sk, skb); 2462058dc334SIlpo Järvinen int skb_size, next_skb_size; 24631da177e4SLinus Torvalds 2464058dc334SIlpo Järvinen skb_size = skb->len; 2465058dc334SIlpo Järvinen next_skb_size = next_skb->len; 24661da177e4SLinus Torvalds 2467058dc334SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); 24681da177e4SLinus Torvalds 24696859d494SIlpo Järvinen tcp_highest_sack_combine(sk, next_skb, skb); 2470a6963a6bSIlpo Järvinen 2471fe067e8aSDavid S. Miller tcp_unlink_write_queue(next_skb, sk); 24721da177e4SLinus Torvalds 2473058dc334SIlpo Järvinen skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size), 24741a4e2d09SArnaldo Carvalho de Melo next_skb_size); 24751da177e4SLinus Torvalds 247652d570aaSJarek Poplawski if (next_skb->ip_summed == CHECKSUM_PARTIAL) 247752d570aaSJarek Poplawski skb->ip_summed = CHECKSUM_PARTIAL; 24781da177e4SLinus Torvalds 247984fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 24801da177e4SLinus Torvalds skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); 24811da177e4SLinus Torvalds 24821da177e4SLinus Torvalds /* Update sequence range on original skb. */ 24831da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 24841da177e4SLinus Torvalds 2485e6c7d085SIlpo Järvinen /* Merge over control information. This moves PSH/FIN etc. over */ 24864de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; 24871da177e4SLinus Torvalds 24881da177e4SLinus Torvalds /* All done, get rid of second SKB and account for it so 24891da177e4SLinus Torvalds * packet counting does not break. 24901da177e4SLinus Torvalds */ 24914828e7f4SIlpo Järvinen TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; 2492b7689205SIlpo Järvinen 2493b7689205SIlpo Järvinen /* changed transmit queue under us so clear hints */ 2494ef9da47cSIlpo Järvinen tcp_clear_retrans_hints_partial(tp); 2495ef9da47cSIlpo Järvinen if (next_skb == tp->retransmit_skb_hint) 2496ef9da47cSIlpo Järvinen tp->retransmit_skb_hint = skb; 2497b7689205SIlpo Järvinen 2498797108d1SIlpo Järvinen tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); 2499797108d1SIlpo Järvinen 25003ab224beSHideo Aoki sk_wmem_free_skb(sk, next_skb); 25011da177e4SLinus Torvalds } 25021da177e4SLinus Torvalds 250367edfef7SAndi Kleen /* Check if coalescing SKBs is legal. */ 2504a2a385d6SEric Dumazet static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) 25054a17fc3aSIlpo Järvinen { 25064a17fc3aSIlpo Järvinen if (tcp_skb_pcount(skb) > 1) 2507a2a385d6SEric Dumazet return false; 25084a17fc3aSIlpo Järvinen /* TODO: SACK collapsing could be used to remove this condition */ 25094a17fc3aSIlpo Järvinen if (skb_shinfo(skb)->nr_frags != 0) 2510a2a385d6SEric Dumazet return false; 25114a17fc3aSIlpo Järvinen if (skb_cloned(skb)) 2512a2a385d6SEric Dumazet return false; 25134a17fc3aSIlpo Järvinen if (skb == tcp_send_head(sk)) 2514a2a385d6SEric Dumazet return false; 25154a17fc3aSIlpo Järvinen /* Some heurestics for collapsing over SACK'd could be invented */ 25164a17fc3aSIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 2517a2a385d6SEric Dumazet return false; 25184a17fc3aSIlpo Järvinen 2519a2a385d6SEric Dumazet return true; 25204a17fc3aSIlpo Järvinen } 25214a17fc3aSIlpo Järvinen 252267edfef7SAndi Kleen /* Collapse packets in the retransmit queue to make to create 252367edfef7SAndi Kleen * less packets on the wire. This is only done on retransmission. 252467edfef7SAndi Kleen */ 25254a17fc3aSIlpo Järvinen static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, 25264a17fc3aSIlpo Järvinen int space) 25274a17fc3aSIlpo Järvinen { 25284a17fc3aSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 25294a17fc3aSIlpo Järvinen struct sk_buff *skb = to, *tmp; 2530a2a385d6SEric Dumazet bool first = true; 25314a17fc3aSIlpo Järvinen 25324a17fc3aSIlpo Järvinen if (!sysctl_tcp_retrans_collapse) 25334a17fc3aSIlpo Järvinen return; 25344de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 25354a17fc3aSIlpo Järvinen return; 25364a17fc3aSIlpo Järvinen 25374a17fc3aSIlpo Järvinen tcp_for_write_queue_from_safe(skb, tmp, sk) { 25384a17fc3aSIlpo Järvinen if (!tcp_can_collapse(sk, skb)) 25394a17fc3aSIlpo Järvinen break; 25404a17fc3aSIlpo Järvinen 25414a17fc3aSIlpo Järvinen space -= skb->len; 25424a17fc3aSIlpo Järvinen 25434a17fc3aSIlpo Järvinen if (first) { 2544a2a385d6SEric Dumazet first = false; 25454a17fc3aSIlpo Järvinen continue; 25464a17fc3aSIlpo Järvinen } 25474a17fc3aSIlpo Järvinen 25484a17fc3aSIlpo Järvinen if (space < 0) 25494a17fc3aSIlpo Järvinen break; 25504a17fc3aSIlpo Järvinen /* Punt if not enough space exists in the first SKB for 25514a17fc3aSIlpo Järvinen * the data in the second 25524a17fc3aSIlpo Järvinen */ 2553a21d4572SEric Dumazet if (skb->len > skb_availroom(to)) 25544a17fc3aSIlpo Järvinen break; 25554a17fc3aSIlpo Järvinen 25564a17fc3aSIlpo Järvinen if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) 25574a17fc3aSIlpo Järvinen break; 25584a17fc3aSIlpo Järvinen 25594a17fc3aSIlpo Järvinen tcp_collapse_retrans(sk, to); 25604a17fc3aSIlpo Järvinen } 25614a17fc3aSIlpo Järvinen } 25624a17fc3aSIlpo Järvinen 25631da177e4SLinus Torvalds /* This retransmits one SKB. Policy decisions and retransmit queue 25641da177e4SLinus Torvalds * state updates are done by the caller. Returns non-zero if an 25651da177e4SLinus Torvalds * error occurred which prevented the send. 25661da177e4SLinus Torvalds */ 256793b174adSYuchung Cheng int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) 25681da177e4SLinus Torvalds { 25691da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 25705d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 25717d227cd2SSridhar Samudrala unsigned int cur_mss; 2572c84a5711SYuchung Cheng int err; 25731da177e4SLinus Torvalds 25745d424d5aSJohn Heffner /* Inconslusive MTU probe */ 25755d424d5aSJohn Heffner if (icsk->icsk_mtup.probe_size) { 25765d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 25775d424d5aSJohn Heffner } 25785d424d5aSJohn Heffner 25791da177e4SLinus Torvalds /* Do not sent more than we queued. 1/4 is reserved for possible 2580caa20d9aSStephen Hemminger * copying overhead: fragmentation, tunneling, mangling etc. 25811da177e4SLinus Torvalds */ 25821da177e4SLinus Torvalds if (atomic_read(&sk->sk_wmem_alloc) > 25831da177e4SLinus Torvalds min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf)) 25841da177e4SLinus Torvalds return -EAGAIN; 25851da177e4SLinus Torvalds 25861f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 25871f3279aeSEric Dumazet return -EBUSY; 25881f3279aeSEric Dumazet 25891da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 25901da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 25911da177e4SLinus Torvalds BUG(); 25921da177e4SLinus Torvalds if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 25931da177e4SLinus Torvalds return -ENOMEM; 25941da177e4SLinus Torvalds } 25951da177e4SLinus Torvalds 25967d227cd2SSridhar Samudrala if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 25977d227cd2SSridhar Samudrala return -EHOSTUNREACH; /* Routing failure or similar. */ 25987d227cd2SSridhar Samudrala 25990c54b85fSIlpo Järvinen cur_mss = tcp_current_mss(sk); 26007d227cd2SSridhar Samudrala 26011da177e4SLinus Torvalds /* If receiver has shrunk his window, and skb is out of 26021da177e4SLinus Torvalds * new window, do not retransmit it. The exception is the 26031da177e4SLinus Torvalds * case, when window is shrunk to zero. In this case 26041da177e4SLinus Torvalds * our retransmit serves as a zero window probe. 26051da177e4SLinus Torvalds */ 26069d4fb27dSJoe Perches if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && 26079d4fb27dSJoe Perches TCP_SKB_CB(skb)->seq != tp->snd_una) 26081da177e4SLinus Torvalds return -EAGAIN; 26091da177e4SLinus Torvalds 26101da177e4SLinus Torvalds if (skb->len > cur_mss) { 26116cc55e09SOctavian Purdila if (tcp_fragment(sk, skb, cur_mss, cur_mss, GFP_ATOMIC)) 26121da177e4SLinus Torvalds return -ENOMEM; /* We'll try again later. */ 261302276f3cSIlpo Järvinen } else { 26149eb9362eSIlpo Järvinen int oldpcount = tcp_skb_pcount(skb); 26159eb9362eSIlpo Järvinen 26169eb9362eSIlpo Järvinen if (unlikely(oldpcount > 1)) { 2617c52e2421SEric Dumazet if (skb_unclone(skb, GFP_ATOMIC)) 2618c52e2421SEric Dumazet return -ENOMEM; 2619*5bbb432cSEric Dumazet tcp_init_tso_segs(skb, cur_mss); 26209eb9362eSIlpo Järvinen tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb)); 26219eb9362eSIlpo Järvinen } 26221da177e4SLinus Torvalds } 26231da177e4SLinus Torvalds 262449213555SDaniel Borkmann /* RFC3168, section 6.1.1.1. ECN fallback */ 262549213555SDaniel Borkmann if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN) 262649213555SDaniel Borkmann tcp_ecn_clear_syn(sk, skb); 262749213555SDaniel Borkmann 26281da177e4SLinus Torvalds tcp_retrans_try_collapse(sk, skb, cur_mss); 26291da177e4SLinus Torvalds 26301da177e4SLinus Torvalds /* Make a copy, if the first transmission SKB clone we made 26311da177e4SLinus Torvalds * is still in somebody's hands, else make a clone. 26321da177e4SLinus Torvalds */ 26331da177e4SLinus Torvalds 263450bceae9SThomas Graf /* make sure skb->data is aligned on arches that require it 263550bceae9SThomas Graf * and check if ack-trimming & collapsing extended the headroom 263650bceae9SThomas Graf * beyond what csum_start can cover. 263750bceae9SThomas Graf */ 263850bceae9SThomas Graf if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) || 263950bceae9SThomas Graf skb_headroom(skb) >= 0xFFFF)) { 2640117632e6SEric Dumazet struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER, 2641117632e6SEric Dumazet GFP_ATOMIC); 2642c84a5711SYuchung Cheng err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) : 2643117632e6SEric Dumazet -ENOBUFS; 2644117632e6SEric Dumazet } else { 2645c84a5711SYuchung Cheng err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 2646117632e6SEric Dumazet } 2647c84a5711SYuchung Cheng 2648fc9f3501SEric Dumazet if (likely(!err)) { 2649c84a5711SYuchung Cheng TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS; 2650fc9f3501SEric Dumazet /* Update global TCP statistics. */ 2651fc9f3501SEric Dumazet TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); 2652fc9f3501SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 2653fc9f3501SEric Dumazet NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 2654fc9f3501SEric Dumazet tp->total_retrans++; 2655fc9f3501SEric Dumazet } 2656c84a5711SYuchung Cheng return err; 265793b174adSYuchung Cheng } 265893b174adSYuchung Cheng 265993b174adSYuchung Cheng int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) 266093b174adSYuchung Cheng { 266193b174adSYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 266293b174adSYuchung Cheng int err = __tcp_retransmit_skb(sk, skb); 26631da177e4SLinus Torvalds 26641da177e4SLinus Torvalds if (err == 0) { 26651da177e4SLinus Torvalds #if FASTRETRANS_DEBUG > 0 26661da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 2667e87cc472SJoe Perches net_dbg_ratelimited("retrans_out leaked\n"); 26681da177e4SLinus Torvalds } 26691da177e4SLinus Torvalds #endif 2670b08d6cb2SIlpo Järvinen if (!tp->retrans_out) 2671b08d6cb2SIlpo Järvinen tp->lost_retrans_low = tp->snd_nxt; 26721da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 26731da177e4SLinus Torvalds tp->retrans_out += tcp_skb_pcount(skb); 26741da177e4SLinus Torvalds 26751da177e4SLinus Torvalds /* Save stamp of the first retransmit. */ 26761da177e4SLinus Torvalds if (!tp->retrans_stamp) 26777faee5c0SEric Dumazet tp->retrans_stamp = tcp_skb_timestamp(skb); 26781da177e4SLinus Torvalds 26791da177e4SLinus Torvalds /* snd_nxt is stored to detect loss of retransmitted segment, 26801da177e4SLinus Torvalds * see tcp_input.c tcp_sacktag_write_queue(). 26811da177e4SLinus Torvalds */ 26821da177e4SLinus Torvalds TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt; 26831f3279aeSEric Dumazet } else if (err != -EBUSY) { 268424ab6becSYuchung Cheng NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL); 26851da177e4SLinus Torvalds } 26866e08d5e3SYuchung Cheng 26876e08d5e3SYuchung Cheng if (tp->undo_retrans < 0) 26886e08d5e3SYuchung Cheng tp->undo_retrans = 0; 26896e08d5e3SYuchung Cheng tp->undo_retrans += tcp_skb_pcount(skb); 26901da177e4SLinus Torvalds return err; 26911da177e4SLinus Torvalds } 26921da177e4SLinus Torvalds 269367edfef7SAndi Kleen /* Check if we forward retransmits are possible in the current 269467edfef7SAndi Kleen * window/congestion state. 269567edfef7SAndi Kleen */ 2696a2a385d6SEric Dumazet static bool tcp_can_forward_retransmit(struct sock *sk) 2697b5afe7bcSIlpo Järvinen { 2698b5afe7bcSIlpo Järvinen const struct inet_connection_sock *icsk = inet_csk(sk); 2699cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 2700b5afe7bcSIlpo Järvinen 2701b5afe7bcSIlpo Järvinen /* Forward retransmissions are possible only during Recovery. */ 2702b5afe7bcSIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Recovery) 2703a2a385d6SEric Dumazet return false; 2704b5afe7bcSIlpo Järvinen 2705b5afe7bcSIlpo Järvinen /* No forward retransmissions in Reno are possible. */ 2706b5afe7bcSIlpo Järvinen if (tcp_is_reno(tp)) 2707a2a385d6SEric Dumazet return false; 2708b5afe7bcSIlpo Järvinen 2709b5afe7bcSIlpo Järvinen /* Yeah, we have to make difficult choice between forward transmission 2710b5afe7bcSIlpo Järvinen * and retransmission... Both ways have their merits... 2711b5afe7bcSIlpo Järvinen * 2712b5afe7bcSIlpo Järvinen * For now we do not retransmit anything, while we have some new 2713b5afe7bcSIlpo Järvinen * segments to send. In the other cases, follow rule 3 for 2714b5afe7bcSIlpo Järvinen * NextSeg() specified in RFC3517. 2715b5afe7bcSIlpo Järvinen */ 2716b5afe7bcSIlpo Järvinen 2717b5afe7bcSIlpo Järvinen if (tcp_may_send_now(sk)) 2718a2a385d6SEric Dumazet return false; 2719b5afe7bcSIlpo Järvinen 2720a2a385d6SEric Dumazet return true; 2721b5afe7bcSIlpo Järvinen } 2722b5afe7bcSIlpo Järvinen 27231da177e4SLinus Torvalds /* This gets called after a retransmit timeout, and the initially 27241da177e4SLinus Torvalds * retransmitted data is acknowledged. It tries to continue 27251da177e4SLinus Torvalds * resending the rest of the retransmit queue, until either 27261da177e4SLinus Torvalds * we've sent it all or the congestion window limit is reached. 27271da177e4SLinus Torvalds * If doing SACK, the first ACK which comes back for a timeout 27281da177e4SLinus Torvalds * based retransmit packet might feed us FACK information again. 27291da177e4SLinus Torvalds * If so, we use it to avoid unnecessarily retransmissions. 27301da177e4SLinus Torvalds */ 27311da177e4SLinus Torvalds void tcp_xmit_retransmit_queue(struct sock *sk) 27321da177e4SLinus Torvalds { 27336687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 27341da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 27351da177e4SLinus Torvalds struct sk_buff *skb; 27360e1c54c2SIlpo Järvinen struct sk_buff *hole = NULL; 2737618d9f25SIlpo Järvinen u32 last_lost; 273861eb55f4SIlpo Järvinen int mib_idx; 27390e1c54c2SIlpo Järvinen int fwd_rexmitting = 0; 27406a438bbeSStephen Hemminger 274145e77d31SIlpo Järvinen if (!tp->packets_out) 274245e77d31SIlpo Järvinen return; 274345e77d31SIlpo Järvinen 274408ebd172SIlpo Järvinen if (!tp->lost_out) 274508ebd172SIlpo Järvinen tp->retransmit_high = tp->snd_una; 274608ebd172SIlpo Järvinen 2747618d9f25SIlpo Järvinen if (tp->retransmit_skb_hint) { 27486a438bbeSStephen Hemminger skb = tp->retransmit_skb_hint; 2749618d9f25SIlpo Järvinen last_lost = TCP_SKB_CB(skb)->end_seq; 2750618d9f25SIlpo Järvinen if (after(last_lost, tp->retransmit_high)) 2751618d9f25SIlpo Järvinen last_lost = tp->retransmit_high; 2752618d9f25SIlpo Järvinen } else { 2753fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 2754618d9f25SIlpo Järvinen last_lost = tp->snd_una; 2755618d9f25SIlpo Järvinen } 27561da177e4SLinus Torvalds 2757fe067e8aSDavid S. Miller tcp_for_write_queue_from(skb, sk) { 27581da177e4SLinus Torvalds __u8 sacked = TCP_SKB_CB(skb)->sacked; 27591da177e4SLinus Torvalds 2760fe067e8aSDavid S. Miller if (skb == tcp_send_head(sk)) 2761fe067e8aSDavid S. Miller break; 27626a438bbeSStephen Hemminger /* we could do better than to assign each time */ 276351456b29SIan Morris if (!hole) 27646a438bbeSStephen Hemminger tp->retransmit_skb_hint = skb; 27656a438bbeSStephen Hemminger 27661da177e4SLinus Torvalds /* Assume this retransmit will generate 27671da177e4SLinus Torvalds * only one packet for congestion window 27681da177e4SLinus Torvalds * calculation purposes. This works because 27691da177e4SLinus Torvalds * tcp_retransmit_skb() will chop up the 27701da177e4SLinus Torvalds * packet to be MSS sized and all the 27711da177e4SLinus Torvalds * packet counting works out. 27721da177e4SLinus Torvalds */ 27731da177e4SLinus Torvalds if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) 27741da177e4SLinus Torvalds return; 27750e1c54c2SIlpo Järvinen 27760e1c54c2SIlpo Järvinen if (fwd_rexmitting) { 27770e1c54c2SIlpo Järvinen begin_fwd: 27780e1c54c2SIlpo Järvinen if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp))) 2779006f582cSIlpo Järvinen break; 27800e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFORWARDRETRANS; 27810e1c54c2SIlpo Järvinen 27820e1c54c2SIlpo Järvinen } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) { 2783618d9f25SIlpo Järvinen tp->retransmit_high = last_lost; 27840e1c54c2SIlpo Järvinen if (!tcp_can_forward_retransmit(sk)) 27850e1c54c2SIlpo Järvinen break; 27860e1c54c2SIlpo Järvinen /* Backtrack if necessary to non-L'ed skb */ 278700db4124SIan Morris if (hole) { 27880e1c54c2SIlpo Järvinen skb = hole; 27890e1c54c2SIlpo Järvinen hole = NULL; 27900e1c54c2SIlpo Järvinen } 27910e1c54c2SIlpo Järvinen fwd_rexmitting = 1; 27920e1c54c2SIlpo Järvinen goto begin_fwd; 27930e1c54c2SIlpo Järvinen 27940e1c54c2SIlpo Järvinen } else if (!(sacked & TCPCB_LOST)) { 279551456b29SIan Morris if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) 27960e1c54c2SIlpo Järvinen hole = skb; 279761eb55f4SIlpo Järvinen continue; 27981da177e4SLinus Torvalds 27990e1c54c2SIlpo Järvinen } else { 2800618d9f25SIlpo Järvinen last_lost = TCP_SKB_CB(skb)->end_seq; 28010e1c54c2SIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Loss) 28020e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFASTRETRANS; 28030e1c54c2SIlpo Järvinen else 28040e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; 28050e1c54c2SIlpo Järvinen } 28060e1c54c2SIlpo Järvinen 28070e1c54c2SIlpo Järvinen if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) 280861eb55f4SIlpo Järvinen continue; 280940b215e5SPavel Emelyanov 281024ab6becSYuchung Cheng if (tcp_retransmit_skb(sk, skb)) 28111da177e4SLinus Torvalds return; 281224ab6becSYuchung Cheng 2813de0744afSPavel Emelyanov NET_INC_STATS_BH(sock_net(sk), mib_idx); 28141da177e4SLinus Torvalds 2815684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 2816a262f0cdSNandita Dukkipati tp->prr_out += tcp_skb_pcount(skb); 2817a262f0cdSNandita Dukkipati 2818fe067e8aSDavid S. Miller if (skb == tcp_write_queue_head(sk)) 2819463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 28203f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, 28213f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 28221da177e4SLinus Torvalds } 28231da177e4SLinus Torvalds } 28241da177e4SLinus Torvalds 2825d83769a5SEric Dumazet /* We allow to exceed memory limits for FIN packets to expedite 2826d83769a5SEric Dumazet * connection tear down and (memory) recovery. 2827845704a5SEric Dumazet * Otherwise tcp_send_fin() could be tempted to either delay FIN 2828845704a5SEric Dumazet * or even be forced to close flow without any FIN. 2829a6c5ea4cSEric Dumazet * In general, we want to allow one skb per socket to avoid hangs 2830a6c5ea4cSEric Dumazet * with edge trigger epoll() 2831d83769a5SEric Dumazet */ 2832a6c5ea4cSEric Dumazet void sk_forced_mem_schedule(struct sock *sk, int size) 2833d83769a5SEric Dumazet { 2834d83769a5SEric Dumazet int amt, status; 2835d83769a5SEric Dumazet 2836d83769a5SEric Dumazet if (size <= sk->sk_forward_alloc) 2837d83769a5SEric Dumazet return; 2838d83769a5SEric Dumazet amt = sk_mem_pages(size); 2839d83769a5SEric Dumazet sk->sk_forward_alloc += amt * SK_MEM_QUANTUM; 2840d83769a5SEric Dumazet sk_memory_allocated_add(sk, amt, &status); 2841d83769a5SEric Dumazet } 2842d83769a5SEric Dumazet 2843845704a5SEric Dumazet /* Send a FIN. The caller locks the socket for us. 2844845704a5SEric Dumazet * We should try to send a FIN packet really hard, but eventually give up. 28451da177e4SLinus Torvalds */ 28461da177e4SLinus Torvalds void tcp_send_fin(struct sock *sk) 28471da177e4SLinus Torvalds { 2848845704a5SEric Dumazet struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk); 28491da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 28501da177e4SLinus Torvalds 2851845704a5SEric Dumazet /* Optimization, tack on the FIN if we have one skb in write queue and 2852845704a5SEric Dumazet * this skb was not yet sent, or we are under memory pressure. 2853845704a5SEric Dumazet * Note: in the latter case, FIN packet will be sent after a timeout, 2854845704a5SEric Dumazet * as TCP stack thinks it has already been transmitted. 28551da177e4SLinus Torvalds */ 2856b8da51ebSEric Dumazet if (tskb && (tcp_send_head(sk) || tcp_under_memory_pressure(sk))) { 2857845704a5SEric Dumazet coalesce: 2858845704a5SEric Dumazet TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN; 2859845704a5SEric Dumazet TCP_SKB_CB(tskb)->end_seq++; 28601da177e4SLinus Torvalds tp->write_seq++; 2861845704a5SEric Dumazet if (!tcp_send_head(sk)) { 2862845704a5SEric Dumazet /* This means tskb was already sent. 2863845704a5SEric Dumazet * Pretend we included the FIN on previous transmit. 2864845704a5SEric Dumazet * We need to set tp->snd_nxt to the value it would have 2865845704a5SEric Dumazet * if FIN had been sent. This is because retransmit path 2866845704a5SEric Dumazet * does not change tp->snd_nxt. 2867845704a5SEric Dumazet */ 2868845704a5SEric Dumazet tp->snd_nxt++; 2869845704a5SEric Dumazet return; 2870845704a5SEric Dumazet } 28711da177e4SLinus Torvalds } else { 2872845704a5SEric Dumazet skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation); 2873845704a5SEric Dumazet if (unlikely(!skb)) { 2874845704a5SEric Dumazet if (tskb) 2875845704a5SEric Dumazet goto coalesce; 2876845704a5SEric Dumazet return; 28771da177e4SLinus Torvalds } 2878d83769a5SEric Dumazet skb_reserve(skb, MAX_TCP_HEADER); 2879a6c5ea4cSEric Dumazet sk_forced_mem_schedule(sk, skb->truesize); 28801da177e4SLinus Torvalds /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 2881e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tp->write_seq, 2882a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_FIN); 28831da177e4SLinus Torvalds tcp_queue_skb(sk, skb); 28841da177e4SLinus Torvalds } 2885845704a5SEric Dumazet __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF); 28861da177e4SLinus Torvalds } 28871da177e4SLinus Torvalds 28881da177e4SLinus Torvalds /* We get here when a process closes a file descriptor (either due to 28891da177e4SLinus Torvalds * an explicit close() or as a byproduct of exit()'ing) and there 28901da177e4SLinus Torvalds * was unread data in the receive queue. This behavior is recommended 289165bb723cSGerrit Renker * by RFC 2525, section 2.17. -DaveM 28921da177e4SLinus Torvalds */ 2893dd0fc66fSAl Viro void tcp_send_active_reset(struct sock *sk, gfp_t priority) 28941da177e4SLinus Torvalds { 28951da177e4SLinus Torvalds struct sk_buff *skb; 28961da177e4SLinus Torvalds 28971da177e4SLinus Torvalds /* NOTE: No TCP options attached and we never retransmit this. */ 28981da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, priority); 28991da177e4SLinus Torvalds if (!skb) { 29004e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 29011da177e4SLinus Torvalds return; 29021da177e4SLinus Torvalds } 29031da177e4SLinus Torvalds 29041da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 29051da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 2906e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), 2907a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_RST); 29081da177e4SLinus Torvalds /* Send it off. */ 2909dfb4b9dcSDavid S. Miller if (tcp_transmit_skb(sk, skb, 0, priority)) 29104e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 291126af65cbSSridhar Samudrala 291281cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); 29131da177e4SLinus Torvalds } 29141da177e4SLinus Torvalds 291567edfef7SAndi Kleen /* Send a crossed SYN-ACK during socket establishment. 291667edfef7SAndi Kleen * WARNING: This routine must only be called when we have already sent 29171da177e4SLinus Torvalds * a SYN packet that crossed the incoming SYN that caused this routine 29181da177e4SLinus Torvalds * to get called. If this assumption fails then the initial rcv_wnd 29191da177e4SLinus Torvalds * and rcv_wscale values will not be correct. 29201da177e4SLinus Torvalds */ 29211da177e4SLinus Torvalds int tcp_send_synack(struct sock *sk) 29221da177e4SLinus Torvalds { 29231da177e4SLinus Torvalds struct sk_buff *skb; 29241da177e4SLinus Torvalds 2925fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 292651456b29SIan Morris if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 292791df42beSJoe Perches pr_debug("%s: wrong queue state\n", __func__); 29281da177e4SLinus Torvalds return -EFAULT; 29291da177e4SLinus Torvalds } 29304de075e0SEric Dumazet if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { 29311da177e4SLinus Torvalds if (skb_cloned(skb)) { 29321da177e4SLinus Torvalds struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC); 293351456b29SIan Morris if (!nskb) 29341da177e4SLinus Torvalds return -ENOMEM; 2935fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 2936f4a775d1SEric Dumazet __skb_header_release(nskb); 2937fe067e8aSDavid S. Miller __tcp_add_write_queue_head(sk, nskb); 29383ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 29393ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 29403ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 29411da177e4SLinus Torvalds skb = nskb; 29421da177e4SLinus Torvalds } 29431da177e4SLinus Torvalds 29444de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; 2945735d3831SFlorian Westphal tcp_ecn_send_synack(sk, skb); 29461da177e4SLinus Torvalds } 2947dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 29481da177e4SLinus Torvalds } 29491da177e4SLinus Torvalds 29504aea39c1SEric Dumazet /** 29514aea39c1SEric Dumazet * tcp_make_synack - Prepare a SYN-ACK. 29524aea39c1SEric Dumazet * sk: listener socket 29534aea39c1SEric Dumazet * dst: dst entry attached to the SYNACK 29544aea39c1SEric Dumazet * req: request_sock pointer 29554aea39c1SEric Dumazet * 29564aea39c1SEric Dumazet * Allocate one skb and build a SYNACK packet. 29574aea39c1SEric Dumazet * @dst is consumed : Caller should not use it again. 29584aea39c1SEric Dumazet */ 29591da177e4SLinus Torvalds struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst, 2960e6b4d113SWilliam Allen Simpson struct request_sock *req, 29618336886fSJerry Chu struct tcp_fastopen_cookie *foc) 29621da177e4SLinus Torvalds { 2963bd0388aeSWilliam Allen Simpson struct tcp_out_options opts; 29642e6599cbSArnaldo Carvalho de Melo struct inet_request_sock *ireq = inet_rsk(req); 29651da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 29661da177e4SLinus Torvalds struct tcphdr *th; 29671da177e4SLinus Torvalds struct sk_buff *skb; 296880f03e27SEric Dumazet struct tcp_md5sig_key *md5 = NULL; 2969bd0388aeSWilliam Allen Simpson int tcp_header_size; 2970f5fff5dcSTom Quetchenbach int mss; 29711da177e4SLinus Torvalds 2972a0b8486cSEric Dumazet skb = sock_wmalloc(sk, MAX_TCP_HEADER, 1, GFP_ATOMIC); 29734aea39c1SEric Dumazet if (unlikely(!skb)) { 29744aea39c1SEric Dumazet dst_release(dst); 29751da177e4SLinus Torvalds return NULL; 29764aea39c1SEric Dumazet } 29771da177e4SLinus Torvalds /* Reserve space for headers. */ 29781da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 29791da177e4SLinus Torvalds 29804aea39c1SEric Dumazet skb_dst_set(skb, dst); 29811da177e4SLinus Torvalds 29820dbaee3bSDavid S. Miller mss = dst_metric_advmss(dst); 2983f5fff5dcSTom Quetchenbach if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss) 2984f5fff5dcSTom Quetchenbach mss = tp->rx_opt.user_mss; 2985f5fff5dcSTom Quetchenbach 298633ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 29878b5f12d0SFlorian Westphal #ifdef CONFIG_SYN_COOKIES 29888b5f12d0SFlorian Westphal if (unlikely(req->cookie_ts)) 29897faee5c0SEric Dumazet skb->skb_mstamp.stamp_jiffies = cookie_init_timestamp(req); 29908b5f12d0SFlorian Westphal else 29918b5f12d0SFlorian Westphal #endif 29927faee5c0SEric Dumazet skb_mstamp_get(&skb->skb_mstamp); 299380f03e27SEric Dumazet 299480f03e27SEric Dumazet #ifdef CONFIG_TCP_MD5SIG 299580f03e27SEric Dumazet rcu_read_lock(); 2996fd3a154aSEric Dumazet md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req)); 299780f03e27SEric Dumazet #endif 299880f03e27SEric Dumazet tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5, 29991a2c6181SChristoph Paasch foc) + sizeof(*th); 300033ad798cSAdam Langley 3001aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 3002aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 30031da177e4SLinus Torvalds 3004aa8223c7SArnaldo Carvalho de Melo th = tcp_hdr(skb); 30051da177e4SLinus Torvalds memset(th, 0, sizeof(struct tcphdr)); 30061da177e4SLinus Torvalds th->syn = 1; 30071da177e4SLinus Torvalds th->ack = 1; 3008735d3831SFlorian Westphal tcp_ecn_make_synack(req, th, sk); 3009b44084c2SEric Dumazet th->source = htons(ireq->ir_num); 3010634fb979SEric Dumazet th->dest = ireq->ir_rmt_port; 3011e870a8efSIlpo Järvinen /* Setting of flags are superfluous here for callers (and ECE is 3012e870a8efSIlpo Järvinen * not even correctly set) 3013e870a8efSIlpo Järvinen */ 3014e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn, 3015a3433f35SChangli Gao TCPHDR_SYN | TCPHDR_ACK); 30164957faadSWilliam Allen Simpson 30171da177e4SLinus Torvalds th->seq = htonl(TCP_SKB_CB(skb)->seq); 30188336886fSJerry Chu /* XXX data is queued and acked as is. No buffer/window check */ 30198336886fSJerry Chu th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt); 30201da177e4SLinus Torvalds 30211da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 3022600ff0c2SIlpo Järvinen th->window = htons(min(req->rcv_wnd, 65535U)); 3023bd0388aeSWilliam Allen Simpson tcp_options_write((__be32 *)(th + 1), tp, &opts); 30241da177e4SLinus Torvalds th->doff = (tcp_header_size >> 2); 3025a0b8486cSEric Dumazet TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_OUTSEGS); 3026cfb6eeb4SYOSHIFUJI Hideaki 3027cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 3028cfb6eeb4SYOSHIFUJI Hideaki /* Okay, we have all we need - do the md5 hash if needed */ 302980f03e27SEric Dumazet if (md5) 3030bd0388aeSWilliam Allen Simpson tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, 303139f8e58eSEric Dumazet md5, req_to_sk(req), skb); 303280f03e27SEric Dumazet rcu_read_unlock(); 3033cfb6eeb4SYOSHIFUJI Hideaki #endif 3034cfb6eeb4SYOSHIFUJI Hideaki 3035b50edd78SEric Dumazet /* Do not fool tcpdump (if any), clean our debris */ 3036b50edd78SEric Dumazet skb->tstamp.tv64 = 0; 30371da177e4SLinus Torvalds return skb; 30381da177e4SLinus Torvalds } 30394bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_make_synack); 30401da177e4SLinus Torvalds 304181164413SDaniel Borkmann static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst) 304281164413SDaniel Borkmann { 304381164413SDaniel Borkmann struct inet_connection_sock *icsk = inet_csk(sk); 304481164413SDaniel Borkmann const struct tcp_congestion_ops *ca; 304581164413SDaniel Borkmann u32 ca_key = dst_metric(dst, RTAX_CC_ALGO); 304681164413SDaniel Borkmann 304781164413SDaniel Borkmann if (ca_key == TCP_CA_UNSPEC) 304881164413SDaniel Borkmann return; 304981164413SDaniel Borkmann 305081164413SDaniel Borkmann rcu_read_lock(); 305181164413SDaniel Borkmann ca = tcp_ca_find_key(ca_key); 305281164413SDaniel Borkmann if (likely(ca && try_module_get(ca->owner))) { 305381164413SDaniel Borkmann module_put(icsk->icsk_ca_ops->owner); 305481164413SDaniel Borkmann icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst); 305581164413SDaniel Borkmann icsk->icsk_ca_ops = ca; 305681164413SDaniel Borkmann } 305781164413SDaniel Borkmann rcu_read_unlock(); 305881164413SDaniel Borkmann } 305981164413SDaniel Borkmann 306067edfef7SAndi Kleen /* Do all connect socket setups that can be done AF independent. */ 3061f7e56a76Sstephen hemminger static void tcp_connect_init(struct sock *sk) 30621da177e4SLinus Torvalds { 3063cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 30641da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 30651da177e4SLinus Torvalds __u8 rcv_wscale; 30661da177e4SLinus Torvalds 30671da177e4SLinus Torvalds /* We'll fix this up when we get a response from the other end. 30681da177e4SLinus Torvalds * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 30691da177e4SLinus Torvalds */ 30701da177e4SLinus Torvalds tp->tcp_header_len = sizeof(struct tcphdr) + 3071bb5b7c11SDavid S. Miller (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0); 30721da177e4SLinus Torvalds 3073cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 307400db4124SIan Morris if (tp->af_specific->md5_lookup(sk, sk)) 3075cfb6eeb4SYOSHIFUJI Hideaki tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 3076cfb6eeb4SYOSHIFUJI Hideaki #endif 3077cfb6eeb4SYOSHIFUJI Hideaki 30781da177e4SLinus Torvalds /* If user gave his TCP_MAXSEG, record it to clamp */ 30791da177e4SLinus Torvalds if (tp->rx_opt.user_mss) 30801da177e4SLinus Torvalds tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 30811da177e4SLinus Torvalds tp->max_window = 0; 30825d424d5aSJohn Heffner tcp_mtup_init(sk); 30831da177e4SLinus Torvalds tcp_sync_mss(sk, dst_mtu(dst)); 30841da177e4SLinus Torvalds 308581164413SDaniel Borkmann tcp_ca_dst_init(sk, dst); 308681164413SDaniel Borkmann 30871da177e4SLinus Torvalds if (!tp->window_clamp) 30881da177e4SLinus Torvalds tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 30890dbaee3bSDavid S. Miller tp->advmss = dst_metric_advmss(dst); 3090f5fff5dcSTom Quetchenbach if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss) 3091f5fff5dcSTom Quetchenbach tp->advmss = tp->rx_opt.user_mss; 3092f5fff5dcSTom Quetchenbach 30931da177e4SLinus Torvalds tcp_initialize_rcv_mss(sk); 30941da177e4SLinus Torvalds 3095e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 3096e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 3097e88c64f0SHagen Paul Pfeifer (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) 3098e88c64f0SHagen Paul Pfeifer tp->window_clamp = tcp_full_space(sk); 3099e88c64f0SHagen Paul Pfeifer 31001da177e4SLinus Torvalds tcp_select_initial_window(tcp_full_space(sk), 31011da177e4SLinus Torvalds tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 31021da177e4SLinus Torvalds &tp->rcv_wnd, 31031da177e4SLinus Torvalds &tp->window_clamp, 3104bb5b7c11SDavid S. Miller sysctl_tcp_window_scaling, 310531d12926Slaurent chavey &rcv_wscale, 310631d12926Slaurent chavey dst_metric(dst, RTAX_INITRWND)); 31071da177e4SLinus Torvalds 31081da177e4SLinus Torvalds tp->rx_opt.rcv_wscale = rcv_wscale; 31091da177e4SLinus Torvalds tp->rcv_ssthresh = tp->rcv_wnd; 31101da177e4SLinus Torvalds 31111da177e4SLinus Torvalds sk->sk_err = 0; 31121da177e4SLinus Torvalds sock_reset_flag(sk, SOCK_DONE); 31131da177e4SLinus Torvalds tp->snd_wnd = 0; 3114ee7537b6SHantzis Fotis tcp_init_wl(tp, 0); 31151da177e4SLinus Torvalds tp->snd_una = tp->write_seq; 31161da177e4SLinus Torvalds tp->snd_sml = tp->write_seq; 311733f5f57eSIlpo Järvinen tp->snd_up = tp->write_seq; 3118370816aeSPavel Emelyanov tp->snd_nxt = tp->write_seq; 3119ee995283SPavel Emelyanov 3120ee995283SPavel Emelyanov if (likely(!tp->repair)) 31211da177e4SLinus Torvalds tp->rcv_nxt = 0; 3122c7781a6eSAndrew Vagin else 3123c7781a6eSAndrew Vagin tp->rcv_tstamp = tcp_time_stamp; 3124ee995283SPavel Emelyanov tp->rcv_wup = tp->rcv_nxt; 3125ee995283SPavel Emelyanov tp->copied_seq = tp->rcv_nxt; 31261da177e4SLinus Torvalds 3127463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; 3128463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_retransmits = 0; 31291da177e4SLinus Torvalds tcp_clear_retrans(tp); 31301da177e4SLinus Torvalds } 31311da177e4SLinus Torvalds 3132783237e8SYuchung Cheng static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) 3133783237e8SYuchung Cheng { 3134783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3135783237e8SYuchung Cheng struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 3136783237e8SYuchung Cheng 3137783237e8SYuchung Cheng tcb->end_seq += skb->len; 3138f4a775d1SEric Dumazet __skb_header_release(skb); 3139783237e8SYuchung Cheng __tcp_add_write_queue_tail(sk, skb); 3140783237e8SYuchung Cheng sk->sk_wmem_queued += skb->truesize; 3141783237e8SYuchung Cheng sk_mem_charge(sk, skb->truesize); 3142783237e8SYuchung Cheng tp->write_seq = tcb->end_seq; 3143783237e8SYuchung Cheng tp->packets_out += tcp_skb_pcount(skb); 3144783237e8SYuchung Cheng } 3145783237e8SYuchung Cheng 3146783237e8SYuchung Cheng /* Build and send a SYN with data and (cached) Fast Open cookie. However, 3147783237e8SYuchung Cheng * queue a data-only packet after the regular SYN, such that regular SYNs 3148783237e8SYuchung Cheng * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges 3149783237e8SYuchung Cheng * only the SYN sequence, the data are retransmitted in the first ACK. 3150783237e8SYuchung Cheng * If cookie is not cached or other error occurs, falls back to send a 3151783237e8SYuchung Cheng * regular SYN with Fast Open cookie request option. 3152783237e8SYuchung Cheng */ 3153783237e8SYuchung Cheng static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) 3154783237e8SYuchung Cheng { 3155783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3156783237e8SYuchung Cheng struct tcp_fastopen_request *fo = tp->fastopen_req; 315757be5bdaSAl Viro int syn_loss = 0, space, err = 0, copied; 3158aab48743SYuchung Cheng unsigned long last_syn_loss = 0; 3159355a901eSEric Dumazet struct sk_buff *syn_data; 3160783237e8SYuchung Cheng 316167da22d2SYuchung Cheng tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ 3162aab48743SYuchung Cheng tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie, 3163aab48743SYuchung Cheng &syn_loss, &last_syn_loss); 3164aab48743SYuchung Cheng /* Recurring FO SYN losses: revert to regular handshake temporarily */ 3165aab48743SYuchung Cheng if (syn_loss > 1 && 3166aab48743SYuchung Cheng time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) { 3167aab48743SYuchung Cheng fo->cookie.len = -1; 3168aab48743SYuchung Cheng goto fallback; 3169aab48743SYuchung Cheng } 3170aab48743SYuchung Cheng 317167da22d2SYuchung Cheng if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE) 317267da22d2SYuchung Cheng fo->cookie.len = -1; 317367da22d2SYuchung Cheng else if (fo->cookie.len <= 0) 3174783237e8SYuchung Cheng goto fallback; 3175783237e8SYuchung Cheng 3176783237e8SYuchung Cheng /* MSS for SYN-data is based on cached MSS and bounded by PMTU and 3177783237e8SYuchung Cheng * user-MSS. Reserve maximum option space for middleboxes that add 3178783237e8SYuchung Cheng * private TCP options. The cost is reduced data space in SYN :( 3179783237e8SYuchung Cheng */ 3180783237e8SYuchung Cheng if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp) 3181783237e8SYuchung Cheng tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 31821b63edd6SYuchung Cheng space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) - 3183783237e8SYuchung Cheng MAX_TCP_OPTION_SPACE; 3184783237e8SYuchung Cheng 3185f5ddcbbbSEric Dumazet space = min_t(size_t, space, fo->size); 3186f5ddcbbbSEric Dumazet 3187f5ddcbbbSEric Dumazet /* limit to order-0 allocations */ 3188f5ddcbbbSEric Dumazet space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER)); 3189f5ddcbbbSEric Dumazet 3190eb934478SEric Dumazet syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false); 3191355a901eSEric Dumazet if (!syn_data) 3192783237e8SYuchung Cheng goto fallback; 3193355a901eSEric Dumazet syn_data->ip_summed = CHECKSUM_PARTIAL; 3194355a901eSEric Dumazet memcpy(syn_data->cb, syn->cb, sizeof(syn->cb)); 319557be5bdaSAl Viro copied = copy_from_iter(skb_put(syn_data, space), space, 319657be5bdaSAl Viro &fo->data->msg_iter); 319757be5bdaSAl Viro if (unlikely(!copied)) { 3198355a901eSEric Dumazet kfree_skb(syn_data); 3199783237e8SYuchung Cheng goto fallback; 3200783237e8SYuchung Cheng } 320157be5bdaSAl Viro if (copied != space) { 320257be5bdaSAl Viro skb_trim(syn_data, copied); 320357be5bdaSAl Viro space = copied; 320457be5bdaSAl Viro } 3205783237e8SYuchung Cheng 3206355a901eSEric Dumazet /* No more data pending in inet_wait_for_connect() */ 3207355a901eSEric Dumazet if (space == fo->size) 3208355a901eSEric Dumazet fo->data = NULL; 3209355a901eSEric Dumazet fo->copied = space; 3210783237e8SYuchung Cheng 3211355a901eSEric Dumazet tcp_connect_queue_skb(sk, syn_data); 3212355a901eSEric Dumazet 3213355a901eSEric Dumazet err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation); 3214355a901eSEric Dumazet 3215355a901eSEric Dumazet syn->skb_mstamp = syn_data->skb_mstamp; 3216355a901eSEric Dumazet 3217355a901eSEric Dumazet /* Now full SYN+DATA was cloned and sent (or not), 3218355a901eSEric Dumazet * remove the SYN from the original skb (syn_data) 3219355a901eSEric Dumazet * we keep in write queue in case of a retransmit, as we 3220355a901eSEric Dumazet * also have the SYN packet (with no data) in the same queue. 3221431a9124SEric Dumazet */ 3222355a901eSEric Dumazet TCP_SKB_CB(syn_data)->seq++; 3223355a901eSEric Dumazet TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH; 3224355a901eSEric Dumazet if (!err) { 322567da22d2SYuchung Cheng tp->syn_data = (fo->copied > 0); 3226f19c29e3SYuchung Cheng NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT); 3227783237e8SYuchung Cheng goto done; 3228783237e8SYuchung Cheng } 3229783237e8SYuchung Cheng 3230783237e8SYuchung Cheng fallback: 3231783237e8SYuchung Cheng /* Send a regular SYN with Fast Open cookie request option */ 3232783237e8SYuchung Cheng if (fo->cookie.len > 0) 3233783237e8SYuchung Cheng fo->cookie.len = 0; 3234783237e8SYuchung Cheng err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); 3235783237e8SYuchung Cheng if (err) 3236783237e8SYuchung Cheng tp->syn_fastopen = 0; 3237783237e8SYuchung Cheng done: 3238783237e8SYuchung Cheng fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ 3239783237e8SYuchung Cheng return err; 3240783237e8SYuchung Cheng } 3241783237e8SYuchung Cheng 324267edfef7SAndi Kleen /* Build a SYN and send it off. */ 32431da177e4SLinus Torvalds int tcp_connect(struct sock *sk) 32441da177e4SLinus Torvalds { 32451da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 32461da177e4SLinus Torvalds struct sk_buff *buff; 3247ee586811SEric Paris int err; 32481da177e4SLinus Torvalds 32491da177e4SLinus Torvalds tcp_connect_init(sk); 32501da177e4SLinus Torvalds 32512b916477SAndrey Vagin if (unlikely(tp->repair)) { 32522b916477SAndrey Vagin tcp_finish_connect(sk, NULL); 32532b916477SAndrey Vagin return 0; 32542b916477SAndrey Vagin } 32552b916477SAndrey Vagin 3256eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true); 3257355a901eSEric Dumazet if (unlikely(!buff)) 32581da177e4SLinus Torvalds return -ENOBUFS; 32591da177e4SLinus Torvalds 3260a3433f35SChangli Gao tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); 32617faee5c0SEric Dumazet tp->retrans_stamp = tcp_time_stamp; 3262783237e8SYuchung Cheng tcp_connect_queue_skb(sk, buff); 3263735d3831SFlorian Westphal tcp_ecn_send_syn(sk, buff); 32641da177e4SLinus Torvalds 3265783237e8SYuchung Cheng /* Send off SYN; include data in Fast Open. */ 3266783237e8SYuchung Cheng err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : 3267783237e8SYuchung Cheng tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); 3268ee586811SEric Paris if (err == -ECONNREFUSED) 3269ee586811SEric Paris return err; 3270bd37a088SWei Yongjun 3271bd37a088SWei Yongjun /* We change tp->snd_nxt after the tcp_transmit_skb() call 3272bd37a088SWei Yongjun * in order to make this packet get counted in tcpOutSegs. 3273bd37a088SWei Yongjun */ 3274bd37a088SWei Yongjun tp->snd_nxt = tp->write_seq; 3275bd37a088SWei Yongjun tp->pushed_seq = tp->write_seq; 327681cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); 32771da177e4SLinus Torvalds 32781da177e4SLinus Torvalds /* Timer for repeating the SYN until an answer. */ 32793f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 32803f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 32811da177e4SLinus Torvalds return 0; 32821da177e4SLinus Torvalds } 32834bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_connect); 32841da177e4SLinus Torvalds 32851da177e4SLinus Torvalds /* Send out a delayed ack, the caller does the policy checking 32861da177e4SLinus Torvalds * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 32871da177e4SLinus Torvalds * for details. 32881da177e4SLinus Torvalds */ 32891da177e4SLinus Torvalds void tcp_send_delayed_ack(struct sock *sk) 32901da177e4SLinus Torvalds { 3291463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 3292463c84b9SArnaldo Carvalho de Melo int ato = icsk->icsk_ack.ato; 32931da177e4SLinus Torvalds unsigned long timeout; 32941da177e4SLinus Torvalds 32959890092eSFlorian Westphal tcp_ca_event(sk, CA_EVENT_DELAYED_ACK); 32969890092eSFlorian Westphal 32971da177e4SLinus Torvalds if (ato > TCP_DELACK_MIN) { 3298463c84b9SArnaldo Carvalho de Melo const struct tcp_sock *tp = tcp_sk(sk); 32991da177e4SLinus Torvalds int max_ato = HZ / 2; 33001da177e4SLinus Torvalds 3301056834d9SIlpo Järvinen if (icsk->icsk_ack.pingpong || 3302056834d9SIlpo Järvinen (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 33031da177e4SLinus Torvalds max_ato = TCP_DELACK_MAX; 33041da177e4SLinus Torvalds 33051da177e4SLinus Torvalds /* Slow path, intersegment interval is "high". */ 33061da177e4SLinus Torvalds 33071da177e4SLinus Torvalds /* If some rtt estimate is known, use it to bound delayed ack. 3308463c84b9SArnaldo Carvalho de Melo * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 33091da177e4SLinus Torvalds * directly. 33101da177e4SLinus Torvalds */ 3311740b0f18SEric Dumazet if (tp->srtt_us) { 3312740b0f18SEric Dumazet int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3), 3313740b0f18SEric Dumazet TCP_DELACK_MIN); 33141da177e4SLinus Torvalds 33151da177e4SLinus Torvalds if (rtt < max_ato) 33161da177e4SLinus Torvalds max_ato = rtt; 33171da177e4SLinus Torvalds } 33181da177e4SLinus Torvalds 33191da177e4SLinus Torvalds ato = min(ato, max_ato); 33201da177e4SLinus Torvalds } 33211da177e4SLinus Torvalds 33221da177e4SLinus Torvalds /* Stay within the limit we were given */ 33231da177e4SLinus Torvalds timeout = jiffies + ato; 33241da177e4SLinus Torvalds 33251da177e4SLinus Torvalds /* Use new timeout only if there wasn't a older one earlier. */ 3326463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 33271da177e4SLinus Torvalds /* If delack timer was blocked or is about to expire, 33281da177e4SLinus Torvalds * send ACK now. 33291da177e4SLinus Torvalds */ 3330463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.blocked || 3331463c84b9SArnaldo Carvalho de Melo time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 33321da177e4SLinus Torvalds tcp_send_ack(sk); 33331da177e4SLinus Torvalds return; 33341da177e4SLinus Torvalds } 33351da177e4SLinus Torvalds 3336463c84b9SArnaldo Carvalho de Melo if (!time_before(timeout, icsk->icsk_ack.timeout)) 3337463c84b9SArnaldo Carvalho de Melo timeout = icsk->icsk_ack.timeout; 33381da177e4SLinus Torvalds } 3339463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 3340463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.timeout = timeout; 3341463c84b9SArnaldo Carvalho de Melo sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 33421da177e4SLinus Torvalds } 33431da177e4SLinus Torvalds 33441da177e4SLinus Torvalds /* This routine sends an ack and also updates the window. */ 33451da177e4SLinus Torvalds void tcp_send_ack(struct sock *sk) 33461da177e4SLinus Torvalds { 33471da177e4SLinus Torvalds struct sk_buff *buff; 33481da177e4SLinus Torvalds 3349058dc334SIlpo Järvinen /* If we have been reset, we may not send again. */ 3350058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3351058dc334SIlpo Järvinen return; 3352058dc334SIlpo Järvinen 33539890092eSFlorian Westphal tcp_ca_event(sk, CA_EVENT_NON_DELAYED_ACK); 33549890092eSFlorian Westphal 33551da177e4SLinus Torvalds /* We are not putting this on the write queue, so 33561da177e4SLinus Torvalds * tcp_transmit_skb() will set the ownership to this 33571da177e4SLinus Torvalds * sock. 33581da177e4SLinus Torvalds */ 335999a1dec7SMel Gorman buff = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC)); 336051456b29SIan Morris if (!buff) { 3361463c84b9SArnaldo Carvalho de Melo inet_csk_schedule_ack(sk); 3362463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; 33633f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 33643f421baaSArnaldo Carvalho de Melo TCP_DELACK_MAX, TCP_RTO_MAX); 33651da177e4SLinus Torvalds return; 33661da177e4SLinus Torvalds } 33671da177e4SLinus Torvalds 33681da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 33691da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 3370a3433f35SChangli Gao tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); 33711da177e4SLinus Torvalds 337298781965SEric Dumazet /* We do not want pure acks influencing TCP Small Queues or fq/pacing 337398781965SEric Dumazet * too much. 337498781965SEric Dumazet * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784 337598781965SEric Dumazet * We also avoid tcp_wfree() overhead (cache line miss accessing 337698781965SEric Dumazet * tp->tsq_flags) by using regular sock_wfree() 337798781965SEric Dumazet */ 337898781965SEric Dumazet skb_set_tcp_pure_ack(buff); 337998781965SEric Dumazet 33801da177e4SLinus Torvalds /* Send it off, this clears delayed acks for us. */ 33817faee5c0SEric Dumazet skb_mstamp_get(&buff->skb_mstamp); 338299a1dec7SMel Gorman tcp_transmit_skb(sk, buff, 0, sk_gfp_atomic(sk, GFP_ATOMIC)); 33831da177e4SLinus Torvalds } 3384e3118e83SDaniel Borkmann EXPORT_SYMBOL_GPL(tcp_send_ack); 33851da177e4SLinus Torvalds 33861da177e4SLinus Torvalds /* This routine sends a packet with an out of date sequence 33871da177e4SLinus Torvalds * number. It assumes the other end will try to ack it. 33881da177e4SLinus Torvalds * 33891da177e4SLinus Torvalds * Question: what should we make while urgent mode? 33901da177e4SLinus Torvalds * 4.4BSD forces sending single byte of data. We cannot send 33911da177e4SLinus Torvalds * out of window data, because we have SND.NXT==SND.MAX... 33921da177e4SLinus Torvalds * 33931da177e4SLinus Torvalds * Current solution: to send TWO zero-length segments in urgent mode: 33941da177e4SLinus Torvalds * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 33951da177e4SLinus Torvalds * out-of-date with SND.UNA-1 to probe window. 33961da177e4SLinus Torvalds */ 3397e520af48SEric Dumazet static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib) 33981da177e4SLinus Torvalds { 33991da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 34001da177e4SLinus Torvalds struct sk_buff *skb; 34011da177e4SLinus Torvalds 34021da177e4SLinus Torvalds /* We don't queue it, tcp_transmit_skb() sets ownership. */ 340399a1dec7SMel Gorman skb = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC)); 340451456b29SIan Morris if (!skb) 34051da177e4SLinus Torvalds return -1; 34061da177e4SLinus Torvalds 34071da177e4SLinus Torvalds /* Reserve space for headers and set control bits. */ 34081da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 34091da177e4SLinus Torvalds /* Use a previous sequence. This should cause the other 34101da177e4SLinus Torvalds * end to send an ack. Don't queue or clone SKB, just 34111da177e4SLinus Torvalds * send it. 34121da177e4SLinus Torvalds */ 3413a3433f35SChangli Gao tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); 34147faee5c0SEric Dumazet skb_mstamp_get(&skb->skb_mstamp); 3415e520af48SEric Dumazet NET_INC_STATS_BH(sock_net(sk), mib); 3416dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC); 34171da177e4SLinus Torvalds } 34181da177e4SLinus Torvalds 3419ee995283SPavel Emelyanov void tcp_send_window_probe(struct sock *sk) 3420ee995283SPavel Emelyanov { 3421ee995283SPavel Emelyanov if (sk->sk_state == TCP_ESTABLISHED) { 3422ee995283SPavel Emelyanov tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; 3423e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE); 3424ee995283SPavel Emelyanov } 3425ee995283SPavel Emelyanov } 3426ee995283SPavel Emelyanov 342767edfef7SAndi Kleen /* Initiate keepalive or window probe from timer. */ 3428e520af48SEric Dumazet int tcp_write_wakeup(struct sock *sk, int mib) 34291da177e4SLinus Torvalds { 34301da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 34311da177e4SLinus Torvalds struct sk_buff *skb; 34321da177e4SLinus Torvalds 3433058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3434058dc334SIlpo Järvinen return -1; 3435058dc334SIlpo Järvinen 343600db4124SIan Morris skb = tcp_send_head(sk); 343700db4124SIan Morris if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { 34381da177e4SLinus Torvalds int err; 34390c54b85fSIlpo Järvinen unsigned int mss = tcp_current_mss(sk); 344090840defSIlpo Järvinen unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 34411da177e4SLinus Torvalds 34421da177e4SLinus Torvalds if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 34431da177e4SLinus Torvalds tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 34441da177e4SLinus Torvalds 34451da177e4SLinus Torvalds /* We are probing the opening of a window 34461da177e4SLinus Torvalds * but the window size is != 0 34471da177e4SLinus Torvalds * must have been a result SWS avoidance ( sender ) 34481da177e4SLinus Torvalds */ 34491da177e4SLinus Torvalds if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 34501da177e4SLinus Torvalds skb->len > mss) { 34511da177e4SLinus Torvalds seg_size = min(seg_size, mss); 34524de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 34536cc55e09SOctavian Purdila if (tcp_fragment(sk, skb, seg_size, mss, GFP_ATOMIC)) 34541da177e4SLinus Torvalds return -1; 34551da177e4SLinus Torvalds } else if (!tcp_skb_pcount(skb)) 3456*5bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss); 34571da177e4SLinus Torvalds 34584de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 3459dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 346066f5fe62SIlpo Järvinen if (!err) 346166f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 34621da177e4SLinus Torvalds return err; 34631da177e4SLinus Torvalds } else { 346433f5f57eSIlpo Järvinen if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) 3465e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 1, mib); 3466e520af48SEric Dumazet return tcp_xmit_probe_skb(sk, 0, mib); 34671da177e4SLinus Torvalds } 34681da177e4SLinus Torvalds } 34691da177e4SLinus Torvalds 34701da177e4SLinus Torvalds /* A window probe timeout has occurred. If window is not closed send 34711da177e4SLinus Torvalds * a partial packet else a zero probe. 34721da177e4SLinus Torvalds */ 34731da177e4SLinus Torvalds void tcp_send_probe0(struct sock *sk) 34741da177e4SLinus Torvalds { 3475463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 34761da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 3477fcdd1cf4SEric Dumazet unsigned long probe_max; 34781da177e4SLinus Torvalds int err; 34791da177e4SLinus Torvalds 3480e520af48SEric Dumazet err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE); 34811da177e4SLinus Torvalds 3482fe067e8aSDavid S. Miller if (tp->packets_out || !tcp_send_head(sk)) { 34831da177e4SLinus Torvalds /* Cancel probe timer, if it is not required. */ 34846687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 0; 3485463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff = 0; 34861da177e4SLinus Torvalds return; 34871da177e4SLinus Torvalds } 34881da177e4SLinus Torvalds 34891da177e4SLinus Torvalds if (err <= 0) { 3490463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_backoff < sysctl_tcp_retries2) 3491463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff++; 34926687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out++; 3493fcdd1cf4SEric Dumazet probe_max = TCP_RTO_MAX; 34941da177e4SLinus Torvalds } else { 34951da177e4SLinus Torvalds /* If packet was not sent due to local congestion, 34966687e988SArnaldo Carvalho de Melo * do not backoff and do not remember icsk_probes_out. 34971da177e4SLinus Torvalds * Let local senders to fight for local resources. 34981da177e4SLinus Torvalds * 34991da177e4SLinus Torvalds * Use accumulated backoff yet. 35001da177e4SLinus Torvalds */ 35016687e988SArnaldo Carvalho de Melo if (!icsk->icsk_probes_out) 35026687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 1; 3503fcdd1cf4SEric Dumazet probe_max = TCP_RESOURCE_PROBE_INTERVAL; 35041da177e4SLinus Torvalds } 3505fcdd1cf4SEric Dumazet inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 350621c8fe99SEric Dumazet tcp_probe0_when(sk, probe_max), 3507fcdd1cf4SEric Dumazet TCP_RTO_MAX); 35081da177e4SLinus Torvalds } 35095db92c99SOctavian Purdila 35105db92c99SOctavian Purdila int tcp_rtx_synack(struct sock *sk, struct request_sock *req) 35115db92c99SOctavian Purdila { 35125db92c99SOctavian Purdila const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific; 35135db92c99SOctavian Purdila struct flowi fl; 35145db92c99SOctavian Purdila int res; 35155db92c99SOctavian Purdila 35165db92c99SOctavian Purdila res = af_ops->send_synack(sk, NULL, &fl, req, 0, NULL); 35175db92c99SOctavian Purdila if (!res) { 35185db92c99SOctavian Purdila TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_RETRANSSEGS); 35195db92c99SOctavian Purdila NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 35205db92c99SOctavian Purdila } 35215db92c99SOctavian Purdila return res; 35225db92c99SOctavian Purdila } 35235db92c99SOctavian Purdila EXPORT_SYMBOL(tcp_rtx_synack); 3524