11da177e4SLinus Torvalds /* 21da177e4SLinus Torvalds * INET An implementation of the TCP/IP protocol suite for the LINUX 31da177e4SLinus Torvalds * operating system. INET is implemented using the BSD Socket 41da177e4SLinus Torvalds * interface as the means of communication with the user level. 51da177e4SLinus Torvalds * 61da177e4SLinus Torvalds * Implementation of the Transmission Control Protocol(TCP). 71da177e4SLinus Torvalds * 802c30a84SJesper Juhl * Authors: Ross Biro 91da177e4SLinus Torvalds * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> 101da177e4SLinus Torvalds * Mark Evans, <evansmp@uhura.aston.ac.uk> 111da177e4SLinus Torvalds * Corey Minyard <wf-rch!minyard@relay.EU.net> 121da177e4SLinus Torvalds * Florian La Roche, <flla@stud.uni-sb.de> 131da177e4SLinus Torvalds * Charles Hedrick, <hedrick@klinzhai.rutgers.edu> 141da177e4SLinus Torvalds * Linus Torvalds, <torvalds@cs.helsinki.fi> 151da177e4SLinus Torvalds * Alan Cox, <gw4pts@gw4pts.ampr.org> 161da177e4SLinus Torvalds * Matthew Dillon, <dillon@apollo.west.oic.com> 171da177e4SLinus Torvalds * Arnt Gulbrandsen, <agulbra@nvg.unit.no> 181da177e4SLinus Torvalds * Jorge Cwik, <jorge@laser.satlink.net> 191da177e4SLinus Torvalds */ 201da177e4SLinus Torvalds 211da177e4SLinus Torvalds /* 221da177e4SLinus Torvalds * Changes: Pedro Roque : Retransmit queue handled by TCP. 231da177e4SLinus Torvalds * : Fragmentation on mtu decrease 241da177e4SLinus Torvalds * : Segment collapse on retransmit 251da177e4SLinus Torvalds * : AF independence 261da177e4SLinus Torvalds * 271da177e4SLinus Torvalds * Linus Torvalds : send_delayed_ack 281da177e4SLinus Torvalds * David S. Miller : Charge memory using the right skb 291da177e4SLinus Torvalds * during syn/ack processing. 301da177e4SLinus Torvalds * David S. Miller : Output engine completely rewritten. 311da177e4SLinus Torvalds * Andrea Arcangeli: SYNACK carry ts_recent in tsecr. 321da177e4SLinus Torvalds * Cacophonix Gaul : draft-minshall-nagle-01 331da177e4SLinus Torvalds * J Hadi Salim : ECN support 341da177e4SLinus Torvalds * 351da177e4SLinus Torvalds */ 361da177e4SLinus Torvalds 3791df42beSJoe Perches #define pr_fmt(fmt) "TCP: " fmt 3891df42beSJoe Perches 391da177e4SLinus Torvalds #include <net/tcp.h> 401da177e4SLinus Torvalds 411da177e4SLinus Torvalds #include <linux/compiler.h> 425a0e3ad6STejun Heo #include <linux/gfp.h> 431da177e4SLinus Torvalds #include <linux/module.h> 441da177e4SLinus Torvalds 451da177e4SLinus Torvalds /* People can turn this off for buggy TCP's found in printers etc. */ 46ab32ea5dSBrian Haley int sysctl_tcp_retrans_collapse __read_mostly = 1; 471da177e4SLinus Torvalds 4815d99e02SRick Jones /* People can turn this on to work with those rare, broken TCPs that 4915d99e02SRick Jones * interpret the window field as a signed quantity. 5015d99e02SRick Jones */ 51ab32ea5dSBrian Haley int sysctl_tcp_workaround_signed_windows __read_mostly = 0; 5215d99e02SRick Jones 5346d3ceabSEric Dumazet /* Default TSQ limit of two TSO segments */ 5446d3ceabSEric Dumazet int sysctl_tcp_limit_output_bytes __read_mostly = 131072; 5546d3ceabSEric Dumazet 561da177e4SLinus Torvalds /* This limits the percentage of the congestion window which we 571da177e4SLinus Torvalds * will allow a single TSO frame to consume. Building TSO frames 581da177e4SLinus Torvalds * which are too large can cause TCP streams to be bursty. 591da177e4SLinus Torvalds */ 60ab32ea5dSBrian Haley int sysctl_tcp_tso_win_divisor __read_mostly = 3; 611da177e4SLinus Torvalds 62ab32ea5dSBrian Haley int sysctl_tcp_mtu_probing __read_mostly = 0; 6397b1ce25SShan Wei int sysctl_tcp_base_mss __read_mostly = TCP_BASE_MSS; 645d424d5aSJohn Heffner 6535089bb2SDavid S. Miller /* By default, RFC2861 behavior. */ 66ab32ea5dSBrian Haley int sysctl_tcp_slow_start_after_idle __read_mostly = 1; 6735089bb2SDavid S. Miller 68c9bee3b7SEric Dumazet unsigned int sysctl_tcp_notsent_lowat __read_mostly = UINT_MAX; 69c9bee3b7SEric Dumazet EXPORT_SYMBOL(sysctl_tcp_notsent_lowat); 70c9bee3b7SEric Dumazet 7146d3ceabSEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 7246d3ceabSEric Dumazet int push_one, gfp_t gfp); 73519855c5SWilliam Allen Simpson 7467edfef7SAndi Kleen /* Account for new data that has been sent to the network. */ 75cf533ea5SEric Dumazet static void tcp_event_new_data_sent(struct sock *sk, const struct sk_buff *skb) 766ff03ac3SIlpo Järvinen { 776ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 786ff03ac3SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 7966f5fe62SIlpo Järvinen unsigned int prior_packets = tp->packets_out; 809e412ba7SIlpo Järvinen 81fe067e8aSDavid S. Miller tcp_advance_send_head(sk, skb); 821da177e4SLinus Torvalds tp->snd_nxt = TCP_SKB_CB(skb)->end_seq; 838512430eSIlpo Järvinen 8466f5fe62SIlpo Järvinen tp->packets_out += tcp_skb_pcount(skb); 856ba8a3b1SNandita Dukkipati if (!prior_packets || icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS || 866a5dc9e5SEric Dumazet icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) { 87750ea2baSYuchung Cheng tcp_rearm_rto(sk); 881da177e4SLinus Torvalds } 89f19c29e3SYuchung Cheng 90f7324acdSDavid S. Miller NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT, 91f19c29e3SYuchung Cheng tcp_skb_pcount(skb)); 926a5dc9e5SEric Dumazet } 931da177e4SLinus Torvalds 941da177e4SLinus Torvalds /* SND.NXT, if window was not shrunk. 951da177e4SLinus Torvalds * If window has been shrunk, what should we make? It is not clear at all. 961da177e4SLinus Torvalds * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-( 971da177e4SLinus Torvalds * Anything in between SND.UNA...SND.UNA+SND.WND also can be already 981da177e4SLinus Torvalds * invalid. OK, let's make this for now: 991da177e4SLinus Torvalds */ 100cf533ea5SEric Dumazet static inline __u32 tcp_acceptable_seq(const struct sock *sk) 1011da177e4SLinus Torvalds { 102cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1039e412ba7SIlpo Järvinen 10490840defSIlpo Järvinen if (!before(tcp_wnd_end(tp), tp->snd_nxt)) 1051da177e4SLinus Torvalds return tp->snd_nxt; 1061da177e4SLinus Torvalds else 10790840defSIlpo Järvinen return tcp_wnd_end(tp); 1081da177e4SLinus Torvalds } 1091da177e4SLinus Torvalds 1101da177e4SLinus Torvalds /* Calculate mss to advertise in SYN segment. 1111da177e4SLinus Torvalds * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that: 1121da177e4SLinus Torvalds * 1131da177e4SLinus Torvalds * 1. It is independent of path mtu. 1141da177e4SLinus Torvalds * 2. Ideally, it is maximal possible segment size i.e. 65535-40. 1151da177e4SLinus Torvalds * 3. For IPv4 it is reasonable to calculate it from maximal MTU of 1161da177e4SLinus Torvalds * attached devices, because some buggy hosts are confused by 1171da177e4SLinus Torvalds * large MSS. 1181da177e4SLinus Torvalds * 4. We do not make 3, we advertise MSS, calculated from first 1191da177e4SLinus Torvalds * hop device mtu, but allow to raise it to ip_rt_min_advmss. 1201da177e4SLinus Torvalds * This may be overridden via information stored in routing table. 1211da177e4SLinus Torvalds * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible, 1221da177e4SLinus Torvalds * probably even Jumbo". 1231da177e4SLinus Torvalds */ 1241da177e4SLinus Torvalds static __u16 tcp_advertise_mss(struct sock *sk) 1251da177e4SLinus Torvalds { 1261da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 127cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1281da177e4SLinus Torvalds int mss = tp->advmss; 1291da177e4SLinus Torvalds 1300dbaee3bSDavid S. Miller if (dst) { 1310dbaee3bSDavid S. Miller unsigned int metric = dst_metric_advmss(dst); 1320dbaee3bSDavid S. Miller 1330dbaee3bSDavid S. Miller if (metric < mss) { 1340dbaee3bSDavid S. Miller mss = metric; 1351da177e4SLinus Torvalds tp->advmss = mss; 1361da177e4SLinus Torvalds } 1370dbaee3bSDavid S. Miller } 1381da177e4SLinus Torvalds 1391da177e4SLinus Torvalds return (__u16)mss; 1401da177e4SLinus Torvalds } 1411da177e4SLinus Torvalds 1421da177e4SLinus Torvalds /* RFC2861. Reset CWND after idle period longer RTO to "restart window". 1431da177e4SLinus Torvalds * This is the first part of cwnd validation mechanism. */ 144cf533ea5SEric Dumazet static void tcp_cwnd_restart(struct sock *sk, const struct dst_entry *dst) 1451da177e4SLinus Torvalds { 146463c84b9SArnaldo Carvalho de Melo struct tcp_sock *tp = tcp_sk(sk); 1471da177e4SLinus Torvalds s32 delta = tcp_time_stamp - tp->lsndtime; 1481da177e4SLinus Torvalds u32 restart_cwnd = tcp_init_cwnd(tp, dst); 1491da177e4SLinus Torvalds u32 cwnd = tp->snd_cwnd; 1501da177e4SLinus Torvalds 1516687e988SArnaldo Carvalho de Melo tcp_ca_event(sk, CA_EVENT_CWND_RESTART); 1521da177e4SLinus Torvalds 1536687e988SArnaldo Carvalho de Melo tp->snd_ssthresh = tcp_current_ssthresh(sk); 1541da177e4SLinus Torvalds restart_cwnd = min(restart_cwnd, cwnd); 1551da177e4SLinus Torvalds 156463c84b9SArnaldo Carvalho de Melo while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd) 1571da177e4SLinus Torvalds cwnd >>= 1; 1581da177e4SLinus Torvalds tp->snd_cwnd = max(cwnd, restart_cwnd); 1591da177e4SLinus Torvalds tp->snd_cwnd_stamp = tcp_time_stamp; 1601da177e4SLinus Torvalds tp->snd_cwnd_used = 0; 1611da177e4SLinus Torvalds } 1621da177e4SLinus Torvalds 16367edfef7SAndi Kleen /* Congestion state accounting after a packet has been sent. */ 16440efc6faSStephen Hemminger static void tcp_event_data_sent(struct tcp_sock *tp, 165cf533ea5SEric Dumazet struct sock *sk) 1661da177e4SLinus Torvalds { 167463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 168463c84b9SArnaldo Carvalho de Melo const u32 now = tcp_time_stamp; 169bcefe17cSCong Wang const struct dst_entry *dst = __sk_dst_get(sk); 1701da177e4SLinus Torvalds 17135089bb2SDavid S. Miller if (sysctl_tcp_slow_start_after_idle && 17235089bb2SDavid S. Miller (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto)) 173463c84b9SArnaldo Carvalho de Melo tcp_cwnd_restart(sk, __sk_dst_get(sk)); 1741da177e4SLinus Torvalds 1751da177e4SLinus Torvalds tp->lsndtime = now; 1761da177e4SLinus Torvalds 1771da177e4SLinus Torvalds /* If it is a reply for ato after last received 1781da177e4SLinus Torvalds * packet, enter pingpong mode. 1791da177e4SLinus Torvalds */ 180bcefe17cSCong Wang if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato && 181bcefe17cSCong Wang (!dst || !dst_metric(dst, RTAX_QUICKACK))) 182463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pingpong = 1; 1831da177e4SLinus Torvalds } 1841da177e4SLinus Torvalds 18567edfef7SAndi Kleen /* Account for an ACK we sent. */ 18640efc6faSStephen Hemminger static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts) 1871da177e4SLinus Torvalds { 188463c84b9SArnaldo Carvalho de Melo tcp_dec_quickack_mode(sk, pkts); 189463c84b9SArnaldo Carvalho de Melo inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); 1901da177e4SLinus Torvalds } 1911da177e4SLinus Torvalds 19285f16525SYuchung Cheng 19385f16525SYuchung Cheng u32 tcp_default_init_rwnd(u32 mss) 19485f16525SYuchung Cheng { 19585f16525SYuchung Cheng /* Initial receive window should be twice of TCP_INIT_CWND to 1969ef71e0cSWeiping Pan * enable proper sending of new unsent data during fast recovery 19785f16525SYuchung Cheng * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a 19885f16525SYuchung Cheng * limit when mss is larger than 1460. 19985f16525SYuchung Cheng */ 20085f16525SYuchung Cheng u32 init_rwnd = TCP_INIT_CWND * 2; 20185f16525SYuchung Cheng 20285f16525SYuchung Cheng if (mss > 1460) 20385f16525SYuchung Cheng init_rwnd = max((1460 * init_rwnd) / mss, 2U); 20485f16525SYuchung Cheng return init_rwnd; 20585f16525SYuchung Cheng } 20685f16525SYuchung Cheng 2071da177e4SLinus Torvalds /* Determine a window scaling and initial window to offer. 2081da177e4SLinus Torvalds * Based on the assumption that the given amount of space 2091da177e4SLinus Torvalds * will be offered. Store the results in the tp structure. 2101da177e4SLinus Torvalds * NOTE: for smooth operation initial space offering should 2111da177e4SLinus Torvalds * be a multiple of mss if possible. We assume here that mss >= 1. 2121da177e4SLinus Torvalds * This MUST be enforced by all callers. 2131da177e4SLinus Torvalds */ 2141da177e4SLinus Torvalds void tcp_select_initial_window(int __space, __u32 mss, 2151da177e4SLinus Torvalds __u32 *rcv_wnd, __u32 *window_clamp, 21631d12926Slaurent chavey int wscale_ok, __u8 *rcv_wscale, 21731d12926Slaurent chavey __u32 init_rcv_wnd) 2181da177e4SLinus Torvalds { 2191da177e4SLinus Torvalds unsigned int space = (__space < 0 ? 0 : __space); 2201da177e4SLinus Torvalds 2211da177e4SLinus Torvalds /* If no clamp set the clamp to the max possible scaled window */ 2221da177e4SLinus Torvalds if (*window_clamp == 0) 2231da177e4SLinus Torvalds (*window_clamp) = (65535 << 14); 2241da177e4SLinus Torvalds space = min(*window_clamp, space); 2251da177e4SLinus Torvalds 2261da177e4SLinus Torvalds /* Quantize space offering to a multiple of mss if possible. */ 2271da177e4SLinus Torvalds if (space > mss) 2281da177e4SLinus Torvalds space = (space / mss) * mss; 2291da177e4SLinus Torvalds 2301da177e4SLinus Torvalds /* NOTE: offering an initial window larger than 32767 23115d99e02SRick Jones * will break some buggy TCP stacks. If the admin tells us 23215d99e02SRick Jones * it is likely we could be speaking with such a buggy stack 23315d99e02SRick Jones * we will truncate our initial window offering to 32K-1 23415d99e02SRick Jones * unless the remote has sent us a window scaling option, 23515d99e02SRick Jones * which we interpret as a sign the remote TCP is not 23615d99e02SRick Jones * misinterpreting the window field as a signed quantity. 2371da177e4SLinus Torvalds */ 23815d99e02SRick Jones if (sysctl_tcp_workaround_signed_windows) 2391da177e4SLinus Torvalds (*rcv_wnd) = min(space, MAX_TCP_WINDOW); 24015d99e02SRick Jones else 24115d99e02SRick Jones (*rcv_wnd) = space; 24215d99e02SRick Jones 2431da177e4SLinus Torvalds (*rcv_wscale) = 0; 2441da177e4SLinus Torvalds if (wscale_ok) { 2451da177e4SLinus Torvalds /* Set window scaling on max possible window 2461da177e4SLinus Torvalds * See RFC1323 for an explanation of the limit to 14 2471da177e4SLinus Torvalds */ 2481da177e4SLinus Torvalds space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max); 249316c1592SStephen Hemminger space = min_t(u32, space, *window_clamp); 2501da177e4SLinus Torvalds while (space > 65535 && (*rcv_wscale) < 14) { 2511da177e4SLinus Torvalds space >>= 1; 2521da177e4SLinus Torvalds (*rcv_wscale)++; 2531da177e4SLinus Torvalds } 2541da177e4SLinus Torvalds } 2551da177e4SLinus Torvalds 2561da177e4SLinus Torvalds if (mss > (1 << *rcv_wscale)) { 25785f16525SYuchung Cheng if (!init_rcv_wnd) /* Use default unless specified otherwise */ 25885f16525SYuchung Cheng init_rcv_wnd = tcp_default_init_rwnd(mss); 259b1afde60SNandita Dukkipati *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss); 2601da177e4SLinus Torvalds } 2611da177e4SLinus Torvalds 2621da177e4SLinus Torvalds /* Set the clamp no higher than max representable value */ 2631da177e4SLinus Torvalds (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp); 2641da177e4SLinus Torvalds } 2654bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_select_initial_window); 2661da177e4SLinus Torvalds 2671da177e4SLinus Torvalds /* Chose a new window to advertise, update state in tcp_sock for the 2681da177e4SLinus Torvalds * socket, and return result with RFC1323 scaling applied. The return 2691da177e4SLinus Torvalds * value can be stuffed directly into th->window for an outgoing 2701da177e4SLinus Torvalds * frame. 2711da177e4SLinus Torvalds */ 27240efc6faSStephen Hemminger static u16 tcp_select_window(struct sock *sk) 2731da177e4SLinus Torvalds { 2741da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2758e165e20SFlorian Westphal u32 old_win = tp->rcv_wnd; 2761da177e4SLinus Torvalds u32 cur_win = tcp_receive_window(tp); 2771da177e4SLinus Torvalds u32 new_win = __tcp_select_window(sk); 2781da177e4SLinus Torvalds 2791da177e4SLinus Torvalds /* Never shrink the offered window */ 2801da177e4SLinus Torvalds if (new_win < cur_win) { 2811da177e4SLinus Torvalds /* Danger Will Robinson! 2821da177e4SLinus Torvalds * Don't update rcv_wup/rcv_wnd here or else 2831da177e4SLinus Torvalds * we will not be able to advertise a zero 2841da177e4SLinus Torvalds * window in time. --DaveM 2851da177e4SLinus Torvalds * 2861da177e4SLinus Torvalds * Relax Will Robinson. 2871da177e4SLinus Torvalds */ 2888e165e20SFlorian Westphal if (new_win == 0) 2898e165e20SFlorian Westphal NET_INC_STATS(sock_net(sk), 2908e165e20SFlorian Westphal LINUX_MIB_TCPWANTZEROWINDOWADV); 291607bfbf2SPatrick McHardy new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale); 2921da177e4SLinus Torvalds } 2931da177e4SLinus Torvalds tp->rcv_wnd = new_win; 2941da177e4SLinus Torvalds tp->rcv_wup = tp->rcv_nxt; 2951da177e4SLinus Torvalds 2961da177e4SLinus Torvalds /* Make sure we do not exceed the maximum possible 2971da177e4SLinus Torvalds * scaled window. 2981da177e4SLinus Torvalds */ 29915d99e02SRick Jones if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows) 3001da177e4SLinus Torvalds new_win = min(new_win, MAX_TCP_WINDOW); 3011da177e4SLinus Torvalds else 3021da177e4SLinus Torvalds new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); 3031da177e4SLinus Torvalds 3041da177e4SLinus Torvalds /* RFC1323 scaling applied */ 3051da177e4SLinus Torvalds new_win >>= tp->rx_opt.rcv_wscale; 3061da177e4SLinus Torvalds 3071da177e4SLinus Torvalds /* If we advertise zero window, disable fast path. */ 3088e165e20SFlorian Westphal if (new_win == 0) { 3091da177e4SLinus Torvalds tp->pred_flags = 0; 3108e165e20SFlorian Westphal if (old_win) 3118e165e20SFlorian Westphal NET_INC_STATS(sock_net(sk), 3128e165e20SFlorian Westphal LINUX_MIB_TCPTOZEROWINDOWADV); 3138e165e20SFlorian Westphal } else if (old_win == 0) { 3148e165e20SFlorian Westphal NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV); 3158e165e20SFlorian Westphal } 3161da177e4SLinus Torvalds 3171da177e4SLinus Torvalds return new_win; 3181da177e4SLinus Torvalds } 3191da177e4SLinus Torvalds 32067edfef7SAndi Kleen /* Packet ECN state for a SYN-ACK */ 321cf533ea5SEric Dumazet static inline void TCP_ECN_send_synack(const struct tcp_sock *tp, struct sk_buff *skb) 322bdf1ee5dSIlpo Järvinen { 3234de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR; 324bdf1ee5dSIlpo Järvinen if (!(tp->ecn_flags & TCP_ECN_OK)) 3254de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE; 326bdf1ee5dSIlpo Järvinen } 327bdf1ee5dSIlpo Järvinen 32867edfef7SAndi Kleen /* Packet ECN state for a SYN. */ 329bdf1ee5dSIlpo Järvinen static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb) 330bdf1ee5dSIlpo Järvinen { 331bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 332bdf1ee5dSIlpo Järvinen 333bdf1ee5dSIlpo Järvinen tp->ecn_flags = 0; 3345d134f1cSHannes Frederic Sowa if (sock_net(sk)->ipv4.sysctl_tcp_ecn == 1) { 3354de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR; 336bdf1ee5dSIlpo Järvinen tp->ecn_flags = TCP_ECN_OK; 337bdf1ee5dSIlpo Järvinen } 338bdf1ee5dSIlpo Järvinen } 339bdf1ee5dSIlpo Järvinen 340bdf1ee5dSIlpo Järvinen static __inline__ void 341cf533ea5SEric Dumazet TCP_ECN_make_synack(const struct request_sock *req, struct tcphdr *th) 342bdf1ee5dSIlpo Järvinen { 343bdf1ee5dSIlpo Järvinen if (inet_rsk(req)->ecn_ok) 344bdf1ee5dSIlpo Järvinen th->ece = 1; 345bdf1ee5dSIlpo Järvinen } 346bdf1ee5dSIlpo Järvinen 34767edfef7SAndi Kleen /* Set up ECN state for a packet on a ESTABLISHED socket that is about to 34867edfef7SAndi Kleen * be sent. 34967edfef7SAndi Kleen */ 350bdf1ee5dSIlpo Järvinen static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb, 351bdf1ee5dSIlpo Järvinen int tcp_header_len) 352bdf1ee5dSIlpo Järvinen { 353bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 354bdf1ee5dSIlpo Järvinen 355bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_OK) { 356bdf1ee5dSIlpo Järvinen /* Not-retransmitted data segment: set ECT and inject CWR. */ 357bdf1ee5dSIlpo Järvinen if (skb->len != tcp_header_len && 358bdf1ee5dSIlpo Järvinen !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { 359bdf1ee5dSIlpo Järvinen INET_ECN_xmit(sk); 360bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) { 361bdf1ee5dSIlpo Järvinen tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; 362bdf1ee5dSIlpo Järvinen tcp_hdr(skb)->cwr = 1; 363bdf1ee5dSIlpo Järvinen skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; 364bdf1ee5dSIlpo Järvinen } 365bdf1ee5dSIlpo Järvinen } else { 366bdf1ee5dSIlpo Järvinen /* ACK or retransmitted segment: clear ECT|CE */ 367bdf1ee5dSIlpo Järvinen INET_ECN_dontxmit(sk); 368bdf1ee5dSIlpo Järvinen } 369bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) 370bdf1ee5dSIlpo Järvinen tcp_hdr(skb)->ece = 1; 371bdf1ee5dSIlpo Järvinen } 372bdf1ee5dSIlpo Järvinen } 373bdf1ee5dSIlpo Järvinen 374e870a8efSIlpo Järvinen /* Constructs common control bits of non-data skb. If SYN/FIN is present, 375e870a8efSIlpo Järvinen * auto increment end seqno. 376e870a8efSIlpo Järvinen */ 377e870a8efSIlpo Järvinen static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags) 378e870a8efSIlpo Järvinen { 3797b7fc97aSEric Dumazet struct skb_shared_info *shinfo = skb_shinfo(skb); 3807b7fc97aSEric Dumazet 3812e8e18efSDavid S. Miller skb->ip_summed = CHECKSUM_PARTIAL; 382e870a8efSIlpo Järvinen skb->csum = 0; 383e870a8efSIlpo Järvinen 3844de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags; 385e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->sacked = 0; 386e870a8efSIlpo Järvinen 3877b7fc97aSEric Dumazet shinfo->gso_segs = 1; 3887b7fc97aSEric Dumazet shinfo->gso_size = 0; 3897b7fc97aSEric Dumazet shinfo->gso_type = 0; 390e870a8efSIlpo Järvinen 391e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->seq = seq; 392a3433f35SChangli Gao if (flags & (TCPHDR_SYN | TCPHDR_FIN)) 393e870a8efSIlpo Järvinen seq++; 394e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->end_seq = seq; 395e870a8efSIlpo Järvinen } 396e870a8efSIlpo Järvinen 397a2a385d6SEric Dumazet static inline bool tcp_urg_mode(const struct tcp_sock *tp) 39833f5f57eSIlpo Järvinen { 39933f5f57eSIlpo Järvinen return tp->snd_una != tp->snd_up; 40033f5f57eSIlpo Järvinen } 40133f5f57eSIlpo Järvinen 40233ad798cSAdam Langley #define OPTION_SACK_ADVERTISE (1 << 0) 40333ad798cSAdam Langley #define OPTION_TS (1 << 1) 40433ad798cSAdam Langley #define OPTION_MD5 (1 << 2) 40589e95a61SOri Finkelman #define OPTION_WSCALE (1 << 3) 4062100c8d2SYuchung Cheng #define OPTION_FAST_OPEN_COOKIE (1 << 8) 40733ad798cSAdam Langley 40833ad798cSAdam Langley struct tcp_out_options { 4092100c8d2SYuchung Cheng u16 options; /* bit field of OPTION_* */ 4102100c8d2SYuchung Cheng u16 mss; /* 0 to disable */ 41133ad798cSAdam Langley u8 ws; /* window scale, 0 to disable */ 41233ad798cSAdam Langley u8 num_sack_blocks; /* number of SACK blocks to include */ 413bd0388aeSWilliam Allen Simpson u8 hash_size; /* bytes in hash_location */ 414bd0388aeSWilliam Allen Simpson __u8 *hash_location; /* temporary pointer, overloaded */ 4152100c8d2SYuchung Cheng __u32 tsval, tsecr; /* need to include OPTION_TS */ 4162100c8d2SYuchung Cheng struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */ 41733ad798cSAdam Langley }; 41833ad798cSAdam Langley 41967edfef7SAndi Kleen /* Write previously computed TCP options to the packet. 42067edfef7SAndi Kleen * 42167edfef7SAndi Kleen * Beware: Something in the Internet is very sensitive to the ordering of 422fd6149d3SIlpo Järvinen * TCP options, we learned this through the hard way, so be careful here. 423fd6149d3SIlpo Järvinen * Luckily we can at least blame others for their non-compliance but from 4248e3bff96Sstephen hemminger * inter-operability perspective it seems that we're somewhat stuck with 425fd6149d3SIlpo Järvinen * the ordering which we have been using if we want to keep working with 426fd6149d3SIlpo Järvinen * those broken things (not that it currently hurts anybody as there isn't 427fd6149d3SIlpo Järvinen * particular reason why the ordering would need to be changed). 428fd6149d3SIlpo Järvinen * 429fd6149d3SIlpo Järvinen * At least SACK_PERM as the first option is known to lead to a disaster 430fd6149d3SIlpo Järvinen * (but it may well be that other scenarios fail similarly). 431fd6149d3SIlpo Järvinen */ 43233ad798cSAdam Langley static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp, 433bd0388aeSWilliam Allen Simpson struct tcp_out_options *opts) 434bd0388aeSWilliam Allen Simpson { 4352100c8d2SYuchung Cheng u16 options = opts->options; /* mungable copy */ 436bd0388aeSWilliam Allen Simpson 437bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_MD5 & options)) { 4381a2c6181SChristoph Paasch *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | 4391a2c6181SChristoph Paasch (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG); 440bd0388aeSWilliam Allen Simpson /* overload cookie hash location */ 441bd0388aeSWilliam Allen Simpson opts->hash_location = (__u8 *)ptr; 44233ad798cSAdam Langley ptr += 4; 44333ad798cSAdam Langley } 44433ad798cSAdam Langley 445fd6149d3SIlpo Järvinen if (unlikely(opts->mss)) { 446fd6149d3SIlpo Järvinen *ptr++ = htonl((TCPOPT_MSS << 24) | 447fd6149d3SIlpo Järvinen (TCPOLEN_MSS << 16) | 448fd6149d3SIlpo Järvinen opts->mss); 449fd6149d3SIlpo Järvinen } 450fd6149d3SIlpo Järvinen 451bd0388aeSWilliam Allen Simpson if (likely(OPTION_TS & options)) { 452bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 45333ad798cSAdam Langley *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | 45433ad798cSAdam Langley (TCPOLEN_SACK_PERM << 16) | 45533ad798cSAdam Langley (TCPOPT_TIMESTAMP << 8) | 45633ad798cSAdam Langley TCPOLEN_TIMESTAMP); 457bd0388aeSWilliam Allen Simpson options &= ~OPTION_SACK_ADVERTISE; 45833ad798cSAdam Langley } else { 459496c98dfSYOSHIFUJI Hideaki *ptr++ = htonl((TCPOPT_NOP << 24) | 46040efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 46140efc6faSStephen Hemminger (TCPOPT_TIMESTAMP << 8) | 46240efc6faSStephen Hemminger TCPOLEN_TIMESTAMP); 46340efc6faSStephen Hemminger } 46433ad798cSAdam Langley *ptr++ = htonl(opts->tsval); 46533ad798cSAdam Langley *ptr++ = htonl(opts->tsecr); 46633ad798cSAdam Langley } 46733ad798cSAdam Langley 468bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 46933ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 47033ad798cSAdam Langley (TCPOPT_NOP << 16) | 47133ad798cSAdam Langley (TCPOPT_SACK_PERM << 8) | 47233ad798cSAdam Langley TCPOLEN_SACK_PERM); 47333ad798cSAdam Langley } 47433ad798cSAdam Langley 475bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_WSCALE & options)) { 47633ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 47733ad798cSAdam Langley (TCPOPT_WINDOW << 16) | 47833ad798cSAdam Langley (TCPOLEN_WINDOW << 8) | 47933ad798cSAdam Langley opts->ws); 48033ad798cSAdam Langley } 48133ad798cSAdam Langley 48233ad798cSAdam Langley if (unlikely(opts->num_sack_blocks)) { 48333ad798cSAdam Langley struct tcp_sack_block *sp = tp->rx_opt.dsack ? 48433ad798cSAdam Langley tp->duplicate_sack : tp->selective_acks; 48540efc6faSStephen Hemminger int this_sack; 48640efc6faSStephen Hemminger 48740efc6faSStephen Hemminger *ptr++ = htonl((TCPOPT_NOP << 24) | 48840efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 48940efc6faSStephen Hemminger (TCPOPT_SACK << 8) | 49033ad798cSAdam Langley (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * 49140efc6faSStephen Hemminger TCPOLEN_SACK_PERBLOCK))); 4922de979bdSStephen Hemminger 49333ad798cSAdam Langley for (this_sack = 0; this_sack < opts->num_sack_blocks; 49433ad798cSAdam Langley ++this_sack) { 49540efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].start_seq); 49640efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].end_seq); 49740efc6faSStephen Hemminger } 4982de979bdSStephen Hemminger 49940efc6faSStephen Hemminger tp->rx_opt.dsack = 0; 50040efc6faSStephen Hemminger } 5012100c8d2SYuchung Cheng 5022100c8d2SYuchung Cheng if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) { 5032100c8d2SYuchung Cheng struct tcp_fastopen_cookie *foc = opts->fastopen_cookie; 5042100c8d2SYuchung Cheng 5052100c8d2SYuchung Cheng *ptr++ = htonl((TCPOPT_EXP << 24) | 5062100c8d2SYuchung Cheng ((TCPOLEN_EXP_FASTOPEN_BASE + foc->len) << 16) | 5072100c8d2SYuchung Cheng TCPOPT_FASTOPEN_MAGIC); 5082100c8d2SYuchung Cheng 5092100c8d2SYuchung Cheng memcpy(ptr, foc->val, foc->len); 5102100c8d2SYuchung Cheng if ((foc->len & 3) == 2) { 5112100c8d2SYuchung Cheng u8 *align = ((u8 *)ptr) + foc->len; 5122100c8d2SYuchung Cheng align[0] = align[1] = TCPOPT_NOP; 5132100c8d2SYuchung Cheng } 5142100c8d2SYuchung Cheng ptr += (foc->len + 3) >> 2; 5152100c8d2SYuchung Cheng } 51640efc6faSStephen Hemminger } 51740efc6faSStephen Hemminger 51867edfef7SAndi Kleen /* Compute TCP options for SYN packets. This is not the final 51967edfef7SAndi Kleen * network wire format yet. 52067edfef7SAndi Kleen */ 52195c96174SEric Dumazet static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb, 52233ad798cSAdam Langley struct tcp_out_options *opts, 523cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 524cf533ea5SEric Dumazet { 52533ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 52695c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 527783237e8SYuchung Cheng struct tcp_fastopen_request *fastopen = tp->fastopen_req; 52833ad798cSAdam Langley 529cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 53033ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 53133ad798cSAdam Langley if (*md5) { 53233ad798cSAdam Langley opts->options |= OPTION_MD5; 533bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 534cfb6eeb4SYOSHIFUJI Hideaki } 53533ad798cSAdam Langley #else 53633ad798cSAdam Langley *md5 = NULL; 537cfb6eeb4SYOSHIFUJI Hideaki #endif 53833ad798cSAdam Langley 53933ad798cSAdam Langley /* We always get an MSS option. The option bytes which will be seen in 54033ad798cSAdam Langley * normal data packets should timestamps be used, must be in the MSS 54133ad798cSAdam Langley * advertised. But we subtract them from tp->mss_cache so that 54233ad798cSAdam Langley * calculations in tcp_sendmsg are simpler etc. So account for this 54333ad798cSAdam Langley * fact here if necessary. If we don't do this correctly, as a 54433ad798cSAdam Langley * receiver we won't recognize data packets as being full sized when we 54533ad798cSAdam Langley * should, and thus we won't abide by the delayed ACK rules correctly. 54633ad798cSAdam Langley * SACKs don't matter, we never delay an ACK when we have any of those 54733ad798cSAdam Langley * going out. */ 54833ad798cSAdam Langley opts->mss = tcp_advertise_mss(sk); 549bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 55033ad798cSAdam Langley 551bb5b7c11SDavid S. Miller if (likely(sysctl_tcp_timestamps && *md5 == NULL)) { 55233ad798cSAdam Langley opts->options |= OPTION_TS; 553ee684b6fSAndrey Vagin opts->tsval = TCP_SKB_CB(skb)->when + tp->tsoffset; 55433ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 555bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 55633ad798cSAdam Langley } 557bb5b7c11SDavid S. Miller if (likely(sysctl_tcp_window_scaling)) { 55833ad798cSAdam Langley opts->ws = tp->rx_opt.rcv_wscale; 55989e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 560bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 56133ad798cSAdam Langley } 562bb5b7c11SDavid S. Miller if (likely(sysctl_tcp_sack)) { 56333ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 564b32d1310SDavid S. Miller if (unlikely(!(OPTION_TS & opts->options))) 565bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 56633ad798cSAdam Langley } 56733ad798cSAdam Langley 568783237e8SYuchung Cheng if (fastopen && fastopen->cookie.len >= 0) { 569783237e8SYuchung Cheng u32 need = TCPOLEN_EXP_FASTOPEN_BASE + fastopen->cookie.len; 570783237e8SYuchung Cheng need = (need + 3) & ~3U; /* Align to 32 bits */ 571783237e8SYuchung Cheng if (remaining >= need) { 572783237e8SYuchung Cheng opts->options |= OPTION_FAST_OPEN_COOKIE; 573783237e8SYuchung Cheng opts->fastopen_cookie = &fastopen->cookie; 574783237e8SYuchung Cheng remaining -= need; 575783237e8SYuchung Cheng tp->syn_fastopen = 1; 576783237e8SYuchung Cheng } 577783237e8SYuchung Cheng } 578bd0388aeSWilliam Allen Simpson 579bd0388aeSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 58033ad798cSAdam Langley } 58133ad798cSAdam Langley 58267edfef7SAndi Kleen /* Set up TCP options for SYN-ACKs. */ 58395c96174SEric Dumazet static unsigned int tcp_synack_options(struct sock *sk, 58433ad798cSAdam Langley struct request_sock *req, 58595c96174SEric Dumazet unsigned int mss, struct sk_buff *skb, 58633ad798cSAdam Langley struct tcp_out_options *opts, 5874957faadSWilliam Allen Simpson struct tcp_md5sig_key **md5, 5888336886fSJerry Chu struct tcp_fastopen_cookie *foc) 5894957faadSWilliam Allen Simpson { 59033ad798cSAdam Langley struct inet_request_sock *ireq = inet_rsk(req); 59195c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 59233ad798cSAdam Langley 59333ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 59433ad798cSAdam Langley *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req); 59533ad798cSAdam Langley if (*md5) { 59633ad798cSAdam Langley opts->options |= OPTION_MD5; 5974957faadSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 5984957faadSWilliam Allen Simpson 5994957faadSWilliam Allen Simpson /* We can't fit any SACK blocks in a packet with MD5 + TS 6004957faadSWilliam Allen Simpson * options. There was discussion about disabling SACK 6014957faadSWilliam Allen Simpson * rather than TS in order to fit in better with old, 6024957faadSWilliam Allen Simpson * buggy kernels, but that was deemed to be unnecessary. 6034957faadSWilliam Allen Simpson */ 604de213e5eSEric Dumazet ireq->tstamp_ok &= !ireq->sack_ok; 60533ad798cSAdam Langley } 60633ad798cSAdam Langley #else 60733ad798cSAdam Langley *md5 = NULL; 60833ad798cSAdam Langley #endif 60933ad798cSAdam Langley 6104957faadSWilliam Allen Simpson /* We always send an MSS option. */ 61133ad798cSAdam Langley opts->mss = mss; 6124957faadSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 61333ad798cSAdam Langley 61433ad798cSAdam Langley if (likely(ireq->wscale_ok)) { 61533ad798cSAdam Langley opts->ws = ireq->rcv_wscale; 61689e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 6174957faadSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 61833ad798cSAdam Langley } 619de213e5eSEric Dumazet if (likely(ireq->tstamp_ok)) { 62033ad798cSAdam Langley opts->options |= OPTION_TS; 62133ad798cSAdam Langley opts->tsval = TCP_SKB_CB(skb)->when; 62233ad798cSAdam Langley opts->tsecr = req->ts_recent; 6234957faadSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 62433ad798cSAdam Langley } 62533ad798cSAdam Langley if (likely(ireq->sack_ok)) { 62633ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 627de213e5eSEric Dumazet if (unlikely(!ireq->tstamp_ok)) 6284957faadSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 62933ad798cSAdam Langley } 63089278c9dSYuchung Cheng if (foc != NULL && foc->len >= 0) { 6318336886fSJerry Chu u32 need = TCPOLEN_EXP_FASTOPEN_BASE + foc->len; 6328336886fSJerry Chu need = (need + 3) & ~3U; /* Align to 32 bits */ 6338336886fSJerry Chu if (remaining >= need) { 6348336886fSJerry Chu opts->options |= OPTION_FAST_OPEN_COOKIE; 6358336886fSJerry Chu opts->fastopen_cookie = foc; 6368336886fSJerry Chu remaining -= need; 6378336886fSJerry Chu } 6388336886fSJerry Chu } 6394957faadSWilliam Allen Simpson 6404957faadSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 64133ad798cSAdam Langley } 64233ad798cSAdam Langley 64367edfef7SAndi Kleen /* Compute TCP options for ESTABLISHED sockets. This is not the 64467edfef7SAndi Kleen * final wire format yet. 64567edfef7SAndi Kleen */ 64695c96174SEric Dumazet static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb, 64733ad798cSAdam Langley struct tcp_out_options *opts, 648cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 649cf533ea5SEric Dumazet { 65033ad798cSAdam Langley struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL; 65133ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 65295c96174SEric Dumazet unsigned int size = 0; 653cabeccbdSIlpo Järvinen unsigned int eff_sacks; 65433ad798cSAdam Langley 6555843ef42SAndi Kleen opts->options = 0; 6565843ef42SAndi Kleen 65733ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 65833ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 65933ad798cSAdam Langley if (unlikely(*md5)) { 66033ad798cSAdam Langley opts->options |= OPTION_MD5; 66133ad798cSAdam Langley size += TCPOLEN_MD5SIG_ALIGNED; 66233ad798cSAdam Langley } 66333ad798cSAdam Langley #else 66433ad798cSAdam Langley *md5 = NULL; 66533ad798cSAdam Langley #endif 66633ad798cSAdam Langley 66733ad798cSAdam Langley if (likely(tp->rx_opt.tstamp_ok)) { 66833ad798cSAdam Langley opts->options |= OPTION_TS; 669ee684b6fSAndrey Vagin opts->tsval = tcb ? tcb->when + tp->tsoffset : 0; 67033ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 67133ad798cSAdam Langley size += TCPOLEN_TSTAMP_ALIGNED; 67233ad798cSAdam Langley } 67333ad798cSAdam Langley 674cabeccbdSIlpo Järvinen eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; 675cabeccbdSIlpo Järvinen if (unlikely(eff_sacks)) { 67695c96174SEric Dumazet const unsigned int remaining = MAX_TCP_OPTION_SPACE - size; 67733ad798cSAdam Langley opts->num_sack_blocks = 67895c96174SEric Dumazet min_t(unsigned int, eff_sacks, 67933ad798cSAdam Langley (remaining - TCPOLEN_SACK_BASE_ALIGNED) / 68033ad798cSAdam Langley TCPOLEN_SACK_PERBLOCK); 68133ad798cSAdam Langley size += TCPOLEN_SACK_BASE_ALIGNED + 68233ad798cSAdam Langley opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; 68333ad798cSAdam Langley } 68433ad798cSAdam Langley 68533ad798cSAdam Langley return size; 68640efc6faSStephen Hemminger } 6871da177e4SLinus Torvalds 68846d3ceabSEric Dumazet 68946d3ceabSEric Dumazet /* TCP SMALL QUEUES (TSQ) 69046d3ceabSEric Dumazet * 69146d3ceabSEric Dumazet * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev) 69246d3ceabSEric Dumazet * to reduce RTT and bufferbloat. 69346d3ceabSEric Dumazet * We do this using a special skb destructor (tcp_wfree). 69446d3ceabSEric Dumazet * 69546d3ceabSEric Dumazet * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb 69646d3ceabSEric Dumazet * needs to be reallocated in a driver. 6978e3bff96Sstephen hemminger * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc 69846d3ceabSEric Dumazet * 69946d3ceabSEric Dumazet * Since transmit from skb destructor is forbidden, we use a tasklet 70046d3ceabSEric Dumazet * to process all sockets that eventually need to send more skbs. 70146d3ceabSEric Dumazet * We use one tasklet per cpu, with its own queue of sockets. 70246d3ceabSEric Dumazet */ 70346d3ceabSEric Dumazet struct tsq_tasklet { 70446d3ceabSEric Dumazet struct tasklet_struct tasklet; 70546d3ceabSEric Dumazet struct list_head head; /* queue of tcp sockets */ 70646d3ceabSEric Dumazet }; 70746d3ceabSEric Dumazet static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet); 70846d3ceabSEric Dumazet 7096f458dfbSEric Dumazet static void tcp_tsq_handler(struct sock *sk) 7106f458dfbSEric Dumazet { 7116f458dfbSEric Dumazet if ((1 << sk->sk_state) & 7126f458dfbSEric Dumazet (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING | 7136f458dfbSEric Dumazet TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) 714bf06200eSJohn Ogness tcp_write_xmit(sk, tcp_current_mss(sk), tcp_sk(sk)->nonagle, 715bf06200eSJohn Ogness 0, GFP_ATOMIC); 7166f458dfbSEric Dumazet } 71746d3ceabSEric Dumazet /* 7188e3bff96Sstephen hemminger * One tasklet per cpu tries to send more skbs. 71946d3ceabSEric Dumazet * We run in tasklet context but need to disable irqs when 7208e3bff96Sstephen hemminger * transferring tsq->head because tcp_wfree() might 72146d3ceabSEric Dumazet * interrupt us (non NAPI drivers) 72246d3ceabSEric Dumazet */ 72346d3ceabSEric Dumazet static void tcp_tasklet_func(unsigned long data) 72446d3ceabSEric Dumazet { 72546d3ceabSEric Dumazet struct tsq_tasklet *tsq = (struct tsq_tasklet *)data; 72646d3ceabSEric Dumazet LIST_HEAD(list); 72746d3ceabSEric Dumazet unsigned long flags; 72846d3ceabSEric Dumazet struct list_head *q, *n; 72946d3ceabSEric Dumazet struct tcp_sock *tp; 73046d3ceabSEric Dumazet struct sock *sk; 73146d3ceabSEric Dumazet 73246d3ceabSEric Dumazet local_irq_save(flags); 73346d3ceabSEric Dumazet list_splice_init(&tsq->head, &list); 73446d3ceabSEric Dumazet local_irq_restore(flags); 73546d3ceabSEric Dumazet 73646d3ceabSEric Dumazet list_for_each_safe(q, n, &list) { 73746d3ceabSEric Dumazet tp = list_entry(q, struct tcp_sock, tsq_node); 73846d3ceabSEric Dumazet list_del(&tp->tsq_node); 73946d3ceabSEric Dumazet 74046d3ceabSEric Dumazet sk = (struct sock *)tp; 74146d3ceabSEric Dumazet bh_lock_sock(sk); 74246d3ceabSEric Dumazet 74346d3ceabSEric Dumazet if (!sock_owned_by_user(sk)) { 7446f458dfbSEric Dumazet tcp_tsq_handler(sk); 74546d3ceabSEric Dumazet } else { 74646d3ceabSEric Dumazet /* defer the work to tcp_release_cb() */ 7476f458dfbSEric Dumazet set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags); 74846d3ceabSEric Dumazet } 74946d3ceabSEric Dumazet bh_unlock_sock(sk); 75046d3ceabSEric Dumazet 75146d3ceabSEric Dumazet clear_bit(TSQ_QUEUED, &tp->tsq_flags); 75246d3ceabSEric Dumazet sk_free(sk); 75346d3ceabSEric Dumazet } 75446d3ceabSEric Dumazet } 75546d3ceabSEric Dumazet 7566f458dfbSEric Dumazet #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \ 7576f458dfbSEric Dumazet (1UL << TCP_WRITE_TIMER_DEFERRED) | \ 758563d34d0SEric Dumazet (1UL << TCP_DELACK_TIMER_DEFERRED) | \ 759563d34d0SEric Dumazet (1UL << TCP_MTU_REDUCED_DEFERRED)) 76046d3ceabSEric Dumazet /** 76146d3ceabSEric Dumazet * tcp_release_cb - tcp release_sock() callback 76246d3ceabSEric Dumazet * @sk: socket 76346d3ceabSEric Dumazet * 76446d3ceabSEric Dumazet * called from release_sock() to perform protocol dependent 76546d3ceabSEric Dumazet * actions before socket release. 76646d3ceabSEric Dumazet */ 76746d3ceabSEric Dumazet void tcp_release_cb(struct sock *sk) 76846d3ceabSEric Dumazet { 76946d3ceabSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 7706f458dfbSEric Dumazet unsigned long flags, nflags; 77146d3ceabSEric Dumazet 7726f458dfbSEric Dumazet /* perform an atomic operation only if at least one flag is set */ 7736f458dfbSEric Dumazet do { 7746f458dfbSEric Dumazet flags = tp->tsq_flags; 7756f458dfbSEric Dumazet if (!(flags & TCP_DEFERRED_ALL)) 7766f458dfbSEric Dumazet return; 7776f458dfbSEric Dumazet nflags = flags & ~TCP_DEFERRED_ALL; 7786f458dfbSEric Dumazet } while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags); 7796f458dfbSEric Dumazet 7806f458dfbSEric Dumazet if (flags & (1UL << TCP_TSQ_DEFERRED)) 7816f458dfbSEric Dumazet tcp_tsq_handler(sk); 7826f458dfbSEric Dumazet 783c3f9b018SEric Dumazet /* Here begins the tricky part : 784c3f9b018SEric Dumazet * We are called from release_sock() with : 785c3f9b018SEric Dumazet * 1) BH disabled 786c3f9b018SEric Dumazet * 2) sk_lock.slock spinlock held 787c3f9b018SEric Dumazet * 3) socket owned by us (sk->sk_lock.owned == 1) 788c3f9b018SEric Dumazet * 789c3f9b018SEric Dumazet * But following code is meant to be called from BH handlers, 790c3f9b018SEric Dumazet * so we should keep BH disabled, but early release socket ownership 791c3f9b018SEric Dumazet */ 792c3f9b018SEric Dumazet sock_release_ownership(sk); 793c3f9b018SEric Dumazet 794144d56e9SEric Dumazet if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) { 7956f458dfbSEric Dumazet tcp_write_timer_handler(sk); 796144d56e9SEric Dumazet __sock_put(sk); 797144d56e9SEric Dumazet } 798144d56e9SEric Dumazet if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) { 7996f458dfbSEric Dumazet tcp_delack_timer_handler(sk); 800144d56e9SEric Dumazet __sock_put(sk); 801144d56e9SEric Dumazet } 802144d56e9SEric Dumazet if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) { 803563d34d0SEric Dumazet sk->sk_prot->mtu_reduced(sk); 804144d56e9SEric Dumazet __sock_put(sk); 805144d56e9SEric Dumazet } 80646d3ceabSEric Dumazet } 80746d3ceabSEric Dumazet EXPORT_SYMBOL(tcp_release_cb); 80846d3ceabSEric Dumazet 80946d3ceabSEric Dumazet void __init tcp_tasklet_init(void) 81046d3ceabSEric Dumazet { 81146d3ceabSEric Dumazet int i; 81246d3ceabSEric Dumazet 81346d3ceabSEric Dumazet for_each_possible_cpu(i) { 81446d3ceabSEric Dumazet struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i); 81546d3ceabSEric Dumazet 81646d3ceabSEric Dumazet INIT_LIST_HEAD(&tsq->head); 81746d3ceabSEric Dumazet tasklet_init(&tsq->tasklet, 81846d3ceabSEric Dumazet tcp_tasklet_func, 81946d3ceabSEric Dumazet (unsigned long)tsq); 82046d3ceabSEric Dumazet } 82146d3ceabSEric Dumazet } 82246d3ceabSEric Dumazet 82346d3ceabSEric Dumazet /* 82446d3ceabSEric Dumazet * Write buffer destructor automatically called from kfree_skb. 8258e3bff96Sstephen hemminger * We can't xmit new skbs from this context, as we might already 82646d3ceabSEric Dumazet * hold qdisc lock. 82746d3ceabSEric Dumazet */ 828d6a4a104SEric Dumazet void tcp_wfree(struct sk_buff *skb) 82946d3ceabSEric Dumazet { 83046d3ceabSEric Dumazet struct sock *sk = skb->sk; 83146d3ceabSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 83246d3ceabSEric Dumazet 83346d3ceabSEric Dumazet if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) && 83446d3ceabSEric Dumazet !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) { 83546d3ceabSEric Dumazet unsigned long flags; 83646d3ceabSEric Dumazet struct tsq_tasklet *tsq; 83746d3ceabSEric Dumazet 83846d3ceabSEric Dumazet /* Keep a ref on socket. 83946d3ceabSEric Dumazet * This last ref will be released in tcp_tasklet_func() 84046d3ceabSEric Dumazet */ 84146d3ceabSEric Dumazet atomic_sub(skb->truesize - 1, &sk->sk_wmem_alloc); 84246d3ceabSEric Dumazet 84346d3ceabSEric Dumazet /* queue this socket to tasklet queue */ 84446d3ceabSEric Dumazet local_irq_save(flags); 84546d3ceabSEric Dumazet tsq = &__get_cpu_var(tsq_tasklet); 84646d3ceabSEric Dumazet list_add(&tp->tsq_node, &tsq->head); 84746d3ceabSEric Dumazet tasklet_schedule(&tsq->tasklet); 84846d3ceabSEric Dumazet local_irq_restore(flags); 84946d3ceabSEric Dumazet } else { 85046d3ceabSEric Dumazet sock_wfree(skb); 85146d3ceabSEric Dumazet } 85246d3ceabSEric Dumazet } 85346d3ceabSEric Dumazet 8541da177e4SLinus Torvalds /* This routine actually transmits TCP packets queued in by 8551da177e4SLinus Torvalds * tcp_do_sendmsg(). This is used by both the initial 8561da177e4SLinus Torvalds * transmission and possible later retransmissions. 8571da177e4SLinus Torvalds * All SKB's seen here are completely headerless. It is our 8581da177e4SLinus Torvalds * job to build the TCP header, and pass the packet down to 8591da177e4SLinus Torvalds * IP so it can do the same plus pass the packet off to the 8601da177e4SLinus Torvalds * device. 8611da177e4SLinus Torvalds * 8621da177e4SLinus Torvalds * We are working here with either a clone of the original 8631da177e4SLinus Torvalds * SKB, or a fresh unique copy made by the retransmit engine. 8641da177e4SLinus Torvalds */ 865056834d9SIlpo Järvinen static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, 866056834d9SIlpo Järvinen gfp_t gfp_mask) 8671da177e4SLinus Torvalds { 8686687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 869dfb4b9dcSDavid S. Miller struct inet_sock *inet; 870dfb4b9dcSDavid S. Miller struct tcp_sock *tp; 871dfb4b9dcSDavid S. Miller struct tcp_skb_cb *tcb; 87233ad798cSAdam Langley struct tcp_out_options opts; 87395c96174SEric Dumazet unsigned int tcp_options_size, tcp_header_size; 874cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 8751da177e4SLinus Torvalds struct tcphdr *th; 8761da177e4SLinus Torvalds int err; 8771da177e4SLinus Torvalds 878dfb4b9dcSDavid S. Miller BUG_ON(!skb || !tcp_skb_pcount(skb)); 879dfb4b9dcSDavid S. Miller 880ccdbb6e9SEric Dumazet if (clone_it) { 881740b0f18SEric Dumazet skb_mstamp_get(&skb->skb_mstamp); 882dfb4b9dcSDavid S. Miller 883dfb4b9dcSDavid S. Miller if (unlikely(skb_cloned(skb))) 884dfb4b9dcSDavid S. Miller skb = pskb_copy(skb, gfp_mask); 885dfb4b9dcSDavid S. Miller else 886dfb4b9dcSDavid S. Miller skb = skb_clone(skb, gfp_mask); 887dfb4b9dcSDavid S. Miller if (unlikely(!skb)) 888dfb4b9dcSDavid S. Miller return -ENOBUFS; 88921962692SEric Dumazet /* Our usage of tstamp should remain private */ 89021962692SEric Dumazet skb->tstamp.tv64 = 0; 891dfb4b9dcSDavid S. Miller } 892dfb4b9dcSDavid S. Miller 893dfb4b9dcSDavid S. Miller inet = inet_sk(sk); 894dfb4b9dcSDavid S. Miller tp = tcp_sk(sk); 895dfb4b9dcSDavid S. Miller tcb = TCP_SKB_CB(skb); 89633ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 8971da177e4SLinus Torvalds 8984de075e0SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) 89933ad798cSAdam Langley tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); 90033ad798cSAdam Langley else 90133ad798cSAdam Langley tcp_options_size = tcp_established_options(sk, skb, &opts, 90233ad798cSAdam Langley &md5); 90333ad798cSAdam Langley tcp_header_size = tcp_options_size + sizeof(struct tcphdr); 9041da177e4SLinus Torvalds 905547669d4SEric Dumazet if (tcp_packets_in_flight(tp) == 0) 9066687e988SArnaldo Carvalho de Melo tcp_ca_event(sk, CA_EVENT_TX_START); 907547669d4SEric Dumazet 908547669d4SEric Dumazet /* if no packet is in qdisc/device queue, then allow XPS to select 909547669d4SEric Dumazet * another queue. 910547669d4SEric Dumazet */ 911547669d4SEric Dumazet skb->ooo_okay = sk_wmem_alloc_get(sk) == 0; 9121da177e4SLinus Torvalds 913aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 914aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 91546d3ceabSEric Dumazet 91646d3ceabSEric Dumazet skb_orphan(skb); 91746d3ceabSEric Dumazet skb->sk = sk; 918c9eeec26SEric Dumazet skb->destructor = tcp_wfree; 91946d3ceabSEric Dumazet atomic_add(skb->truesize, &sk->sk_wmem_alloc); 9201da177e4SLinus Torvalds 9211da177e4SLinus Torvalds /* Build TCP header and checksum it. */ 922aa8223c7SArnaldo Carvalho de Melo th = tcp_hdr(skb); 923c720c7e8SEric Dumazet th->source = inet->inet_sport; 924c720c7e8SEric Dumazet th->dest = inet->inet_dport; 9251da177e4SLinus Torvalds th->seq = htonl(tcb->seq); 9261da177e4SLinus Torvalds th->ack_seq = htonl(tp->rcv_nxt); 927df7a3b07SAl Viro *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | 9284de075e0SEric Dumazet tcb->tcp_flags); 929dfb4b9dcSDavid S. Miller 9304de075e0SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) { 9311da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments 9321da177e4SLinus Torvalds * is never scaled. 9331da177e4SLinus Torvalds */ 934600ff0c2SIlpo Järvinen th->window = htons(min(tp->rcv_wnd, 65535U)); 9351da177e4SLinus Torvalds } else { 9361da177e4SLinus Torvalds th->window = htons(tcp_select_window(sk)); 9371da177e4SLinus Torvalds } 9381da177e4SLinus Torvalds th->check = 0; 9391da177e4SLinus Torvalds th->urg_ptr = 0; 9401da177e4SLinus Torvalds 94133f5f57eSIlpo Järvinen /* The urg_mode check is necessary during a below snd_una win probe */ 9427691367dSHerbert Xu if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { 9437691367dSHerbert Xu if (before(tp->snd_up, tcb->seq + 0x10000)) { 9441da177e4SLinus Torvalds th->urg_ptr = htons(tp->snd_up - tcb->seq); 9451da177e4SLinus Torvalds th->urg = 1; 9467691367dSHerbert Xu } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { 9470eae88f3SEric Dumazet th->urg_ptr = htons(0xFFFF); 9487691367dSHerbert Xu th->urg = 1; 9497691367dSHerbert Xu } 9501da177e4SLinus Torvalds } 9511da177e4SLinus Torvalds 952bd0388aeSWilliam Allen Simpson tcp_options_write((__be32 *)(th + 1), tp, &opts); 9534de075e0SEric Dumazet if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0)) 9549e412ba7SIlpo Järvinen TCP_ECN_send(sk, skb, tcp_header_size); 955dfb4b9dcSDavid S. Miller 956cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 957cfb6eeb4SYOSHIFUJI Hideaki /* Calculate the MD5 hash, as we have all we need now */ 958cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 959a465419bSEric Dumazet sk_nocaps_add(sk, NETIF_F_GSO_MASK); 960bd0388aeSWilliam Allen Simpson tp->af_specific->calc_md5_hash(opts.hash_location, 96149a72dfbSAdam Langley md5, sk, NULL, skb); 962cfb6eeb4SYOSHIFUJI Hideaki } 963cfb6eeb4SYOSHIFUJI Hideaki #endif 964cfb6eeb4SYOSHIFUJI Hideaki 965bb296246SHerbert Xu icsk->icsk_af_ops->send_check(sk, skb); 9661da177e4SLinus Torvalds 9674de075e0SEric Dumazet if (likely(tcb->tcp_flags & TCPHDR_ACK)) 968fc6415bcSDavid S. Miller tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); 9691da177e4SLinus Torvalds 9701da177e4SLinus Torvalds if (skb->len != tcp_header_size) 971cf533ea5SEric Dumazet tcp_event_data_sent(tp, sk); 9721da177e4SLinus Torvalds 973bd37a088SWei Yongjun if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) 974aa2ea058STom Herbert TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, 975aa2ea058STom Herbert tcp_skb_pcount(skb)); 9761da177e4SLinus Torvalds 977b0270e91SEric Dumazet err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl); 97883de47cdSHua Zhong if (likely(err <= 0)) 9791da177e4SLinus Torvalds return err; 9801da177e4SLinus Torvalds 9813cfe3baaSIlpo Järvinen tcp_enter_cwr(sk, 1); 9821da177e4SLinus Torvalds 983b9df3cb8SGerrit Renker return net_xmit_eval(err); 9841da177e4SLinus Torvalds } 9851da177e4SLinus Torvalds 98667edfef7SAndi Kleen /* This routine just queues the buffer for sending. 9871da177e4SLinus Torvalds * 9881da177e4SLinus Torvalds * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, 9891da177e4SLinus Torvalds * otherwise socket can stall. 9901da177e4SLinus Torvalds */ 9911da177e4SLinus Torvalds static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) 9921da177e4SLinus Torvalds { 9931da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 9941da177e4SLinus Torvalds 9951da177e4SLinus Torvalds /* Advance write_seq and place onto the write_queue. */ 9961da177e4SLinus Torvalds tp->write_seq = TCP_SKB_CB(skb)->end_seq; 9971da177e4SLinus Torvalds skb_header_release(skb); 998fe067e8aSDavid S. Miller tcp_add_write_queue_tail(sk, skb); 9993ab224beSHideo Aoki sk->sk_wmem_queued += skb->truesize; 10003ab224beSHideo Aoki sk_mem_charge(sk, skb->truesize); 10011da177e4SLinus Torvalds } 10021da177e4SLinus Torvalds 100367edfef7SAndi Kleen /* Initialize TSO segments for a packet. */ 1004cf533ea5SEric Dumazet static void tcp_set_skb_tso_segs(const struct sock *sk, struct sk_buff *skb, 1005056834d9SIlpo Järvinen unsigned int mss_now) 1006f6302d1dSDavid S. Miller { 10077b7fc97aSEric Dumazet struct skb_shared_info *shinfo = skb_shinfo(skb); 10087b7fc97aSEric Dumazet 1009c52e2421SEric Dumazet /* Make sure we own this skb before messing gso_size/gso_segs */ 1010c52e2421SEric Dumazet WARN_ON_ONCE(skb_cloned(skb)); 1011c52e2421SEric Dumazet 10128f26fb1cSEric Dumazet if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) { 1013f6302d1dSDavid S. Miller /* Avoid the costly divide in the normal 1014f6302d1dSDavid S. Miller * non-TSO case. 1015f6302d1dSDavid S. Miller */ 10167b7fc97aSEric Dumazet shinfo->gso_segs = 1; 10177b7fc97aSEric Dumazet shinfo->gso_size = 0; 10187b7fc97aSEric Dumazet shinfo->gso_type = 0; 1019f6302d1dSDavid S. Miller } else { 10207b7fc97aSEric Dumazet shinfo->gso_segs = DIV_ROUND_UP(skb->len, mss_now); 10217b7fc97aSEric Dumazet shinfo->gso_size = mss_now; 10227b7fc97aSEric Dumazet shinfo->gso_type = sk->sk_gso_type; 10231da177e4SLinus Torvalds } 10241da177e4SLinus Torvalds } 10251da177e4SLinus Torvalds 102691fed7a1SIlpo Järvinen /* When a modification to fackets out becomes necessary, we need to check 102768f8353bSIlpo Järvinen * skb is counted to fackets_out or not. 102891fed7a1SIlpo Järvinen */ 1029cf533ea5SEric Dumazet static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb, 103091fed7a1SIlpo Järvinen int decr) 103191fed7a1SIlpo Järvinen { 1032a47e5a98SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1033a47e5a98SIlpo Järvinen 1034dc86967bSIlpo Järvinen if (!tp->sacked_out || tcp_is_reno(tp)) 103591fed7a1SIlpo Järvinen return; 103691fed7a1SIlpo Järvinen 10376859d494SIlpo Järvinen if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq)) 103891fed7a1SIlpo Järvinen tp->fackets_out -= decr; 103991fed7a1SIlpo Järvinen } 104091fed7a1SIlpo Järvinen 1041797108d1SIlpo Järvinen /* Pcount in the middle of the write queue got changed, we need to do various 1042797108d1SIlpo Järvinen * tweaks to fix counters 1043797108d1SIlpo Järvinen */ 1044cf533ea5SEric Dumazet static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) 1045797108d1SIlpo Järvinen { 1046797108d1SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1047797108d1SIlpo Järvinen 1048797108d1SIlpo Järvinen tp->packets_out -= decr; 1049797108d1SIlpo Järvinen 1050797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 1051797108d1SIlpo Järvinen tp->sacked_out -= decr; 1052797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 1053797108d1SIlpo Järvinen tp->retrans_out -= decr; 1054797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 1055797108d1SIlpo Järvinen tp->lost_out -= decr; 1056797108d1SIlpo Järvinen 1057797108d1SIlpo Järvinen /* Reno case is special. Sigh... */ 1058797108d1SIlpo Järvinen if (tcp_is_reno(tp) && decr > 0) 1059797108d1SIlpo Järvinen tp->sacked_out -= min_t(u32, tp->sacked_out, decr); 1060797108d1SIlpo Järvinen 1061797108d1SIlpo Järvinen tcp_adjust_fackets_out(sk, skb, decr); 1062797108d1SIlpo Järvinen 1063797108d1SIlpo Järvinen if (tp->lost_skb_hint && 1064797108d1SIlpo Järvinen before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && 106552cf3cc8SIlpo Järvinen (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))) 1066797108d1SIlpo Järvinen tp->lost_cnt_hint -= decr; 1067797108d1SIlpo Järvinen 1068797108d1SIlpo Järvinen tcp_verify_left_out(tp); 1069797108d1SIlpo Järvinen } 1070797108d1SIlpo Järvinen 10711da177e4SLinus Torvalds /* Function to create two new TCP segments. Shrinks the given segment 10721da177e4SLinus Torvalds * to the specified size and appends a new segment with the rest of the 10731da177e4SLinus Torvalds * packet to the list. This won't be called frequently, I hope. 10741da177e4SLinus Torvalds * Remember, these are still headerless SKBs at this point. 10751da177e4SLinus Torvalds */ 1076056834d9SIlpo Järvinen int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, 1077*6cc55e09SOctavian Purdila unsigned int mss_now, gfp_t gfp) 10781da177e4SLinus Torvalds { 10791da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 10801da177e4SLinus Torvalds struct sk_buff *buff; 10816475be16SDavid S. Miller int nsize, old_factor; 1082b60b49eaSHerbert Xu int nlen; 10839ce01461SIlpo Järvinen u8 flags; 10841da177e4SLinus Torvalds 10852fceec13SIlpo Järvinen if (WARN_ON(len > skb->len)) 10862fceec13SIlpo Järvinen return -EINVAL; 10876a438bbeSStephen Hemminger 10881da177e4SLinus Torvalds nsize = skb_headlen(skb) - len; 10891da177e4SLinus Torvalds if (nsize < 0) 10901da177e4SLinus Torvalds nsize = 0; 10911da177e4SLinus Torvalds 1092*6cc55e09SOctavian Purdila if (skb_unclone(skb, gfp)) 10931da177e4SLinus Torvalds return -ENOMEM; 10941da177e4SLinus Torvalds 10951da177e4SLinus Torvalds /* Get a new skb... force flag on. */ 1096*6cc55e09SOctavian Purdila buff = sk_stream_alloc_skb(sk, nsize, gfp); 10971da177e4SLinus Torvalds if (buff == NULL) 10981da177e4SLinus Torvalds return -ENOMEM; /* We'll just try again later. */ 1099ef5cb973SHerbert Xu 11003ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 11013ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1102b60b49eaSHerbert Xu nlen = skb->len - len - nsize; 1103b60b49eaSHerbert Xu buff->truesize += nlen; 1104b60b49eaSHerbert Xu skb->truesize -= nlen; 11051da177e4SLinus Torvalds 11061da177e4SLinus Torvalds /* Correct the sequence numbers. */ 11071da177e4SLinus Torvalds TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 11081da177e4SLinus Torvalds TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 11091da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 11101da177e4SLinus Torvalds 11111da177e4SLinus Torvalds /* PSH and FIN should only be set in the second packet. */ 11124de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 11134de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 11144de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1115e14c3cafSHerbert Xu TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; 11161da177e4SLinus Torvalds 111784fa7933SPatrick McHardy if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { 11181da177e4SLinus Torvalds /* Copy and checksum data tail into the new buffer. */ 1119056834d9SIlpo Järvinen buff->csum = csum_partial_copy_nocheck(skb->data + len, 1120056834d9SIlpo Järvinen skb_put(buff, nsize), 11211da177e4SLinus Torvalds nsize, 0); 11221da177e4SLinus Torvalds 11231da177e4SLinus Torvalds skb_trim(skb, len); 11241da177e4SLinus Torvalds 11251da177e4SLinus Torvalds skb->csum = csum_block_sub(skb->csum, buff->csum, len); 11261da177e4SLinus Torvalds } else { 112784fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 11281da177e4SLinus Torvalds skb_split(skb, buff, len); 11291da177e4SLinus Torvalds } 11301da177e4SLinus Torvalds 11311da177e4SLinus Torvalds buff->ip_summed = skb->ip_summed; 11321da177e4SLinus Torvalds 11331da177e4SLinus Torvalds /* Looks stupid, but our code really uses when of 11341da177e4SLinus Torvalds * skbs, which it never sent before. --ANK 11351da177e4SLinus Torvalds */ 11361da177e4SLinus Torvalds TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when; 1137a61bbcf2SPatrick McHardy buff->tstamp = skb->tstamp; 11381da177e4SLinus Torvalds 11396475be16SDavid S. Miller old_factor = tcp_skb_pcount(skb); 11406475be16SDavid S. Miller 11411da177e4SLinus Torvalds /* Fix up tso_factor for both original and new SKB. */ 1142846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1143846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, buff, mss_now); 11441da177e4SLinus Torvalds 11456475be16SDavid S. Miller /* If this packet has been sent out already, we must 11466475be16SDavid S. Miller * adjust the various packet counters. 11476475be16SDavid S. Miller */ 1148cf0b450cSHerbert Xu if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 11496475be16SDavid S. Miller int diff = old_factor - tcp_skb_pcount(skb) - 11506475be16SDavid S. Miller tcp_skb_pcount(buff); 11511da177e4SLinus Torvalds 1152797108d1SIlpo Järvinen if (diff) 1153797108d1SIlpo Järvinen tcp_adjust_pcount(sk, skb, diff); 11541da177e4SLinus Torvalds } 11551da177e4SLinus Torvalds 11561da177e4SLinus Torvalds /* Link BUFF into the send queue. */ 1157f44b5271SDavid S. Miller skb_header_release(buff); 1158fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 11591da177e4SLinus Torvalds 11601da177e4SLinus Torvalds return 0; 11611da177e4SLinus Torvalds } 11621da177e4SLinus Torvalds 11631da177e4SLinus Torvalds /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c 11641da177e4SLinus Torvalds * eventually). The difference is that pulled data not copied, but 11651da177e4SLinus Torvalds * immediately discarded. 11661da177e4SLinus Torvalds */ 1167f2911969SHerbert Xu ~{PmVHI~} static void __pskb_trim_head(struct sk_buff *skb, int len) 11681da177e4SLinus Torvalds { 11697b7fc97aSEric Dumazet struct skb_shared_info *shinfo; 11701da177e4SLinus Torvalds int i, k, eat; 11711da177e4SLinus Torvalds 11724fa48bf3SEric Dumazet eat = min_t(int, len, skb_headlen(skb)); 11734fa48bf3SEric Dumazet if (eat) { 11744fa48bf3SEric Dumazet __skb_pull(skb, eat); 11754fa48bf3SEric Dumazet len -= eat; 11764fa48bf3SEric Dumazet if (!len) 11774fa48bf3SEric Dumazet return; 11784fa48bf3SEric Dumazet } 11791da177e4SLinus Torvalds eat = len; 11801da177e4SLinus Torvalds k = 0; 11817b7fc97aSEric Dumazet shinfo = skb_shinfo(skb); 11827b7fc97aSEric Dumazet for (i = 0; i < shinfo->nr_frags; i++) { 11837b7fc97aSEric Dumazet int size = skb_frag_size(&shinfo->frags[i]); 11849e903e08SEric Dumazet 11859e903e08SEric Dumazet if (size <= eat) { 1186aff65da0SIan Campbell skb_frag_unref(skb, i); 11879e903e08SEric Dumazet eat -= size; 11881da177e4SLinus Torvalds } else { 11897b7fc97aSEric Dumazet shinfo->frags[k] = shinfo->frags[i]; 11901da177e4SLinus Torvalds if (eat) { 11917b7fc97aSEric Dumazet shinfo->frags[k].page_offset += eat; 11927b7fc97aSEric Dumazet skb_frag_size_sub(&shinfo->frags[k], eat); 11931da177e4SLinus Torvalds eat = 0; 11941da177e4SLinus Torvalds } 11951da177e4SLinus Torvalds k++; 11961da177e4SLinus Torvalds } 11971da177e4SLinus Torvalds } 11987b7fc97aSEric Dumazet shinfo->nr_frags = k; 11991da177e4SLinus Torvalds 120027a884dcSArnaldo Carvalho de Melo skb_reset_tail_pointer(skb); 12011da177e4SLinus Torvalds skb->data_len -= len; 12021da177e4SLinus Torvalds skb->len = skb->data_len; 12031da177e4SLinus Torvalds } 12041da177e4SLinus Torvalds 120567edfef7SAndi Kleen /* Remove acked data from a packet in the transmit queue. */ 12061da177e4SLinus Torvalds int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 12071da177e4SLinus Torvalds { 120814bbd6a5SPravin B Shelar if (skb_unclone(skb, GFP_ATOMIC)) 12091da177e4SLinus Torvalds return -ENOMEM; 12101da177e4SLinus Torvalds 12114fa48bf3SEric Dumazet __pskb_trim_head(skb, len); 12121da177e4SLinus Torvalds 12131da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq += len; 121484fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 12151da177e4SLinus Torvalds 12161da177e4SLinus Torvalds skb->truesize -= len; 12171da177e4SLinus Torvalds sk->sk_wmem_queued -= len; 12183ab224beSHideo Aoki sk_mem_uncharge(sk, len); 12191da177e4SLinus Torvalds sock_set_flag(sk, SOCK_QUEUE_SHRUNK); 12201da177e4SLinus Torvalds 12215b35e1e6SNeal Cardwell /* Any change of skb->len requires recalculation of tso factor. */ 12221da177e4SLinus Torvalds if (tcp_skb_pcount(skb) > 1) 12235b35e1e6SNeal Cardwell tcp_set_skb_tso_segs(sk, skb, tcp_skb_mss(skb)); 12241da177e4SLinus Torvalds 12251da177e4SLinus Torvalds return 0; 12261da177e4SLinus Torvalds } 12271da177e4SLinus Torvalds 12281b63edd6SYuchung Cheng /* Calculate MSS not accounting any TCP options. */ 12291b63edd6SYuchung Cheng static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu) 12305d424d5aSJohn Heffner { 1231cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1232cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 12335d424d5aSJohn Heffner int mss_now; 12345d424d5aSJohn Heffner 12355d424d5aSJohn Heffner /* Calculate base mss without TCP options: 12365d424d5aSJohn Heffner It is MMS_S - sizeof(tcphdr) of rfc1122 12375d424d5aSJohn Heffner */ 12385d424d5aSJohn Heffner mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 12395d424d5aSJohn Heffner 124067469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 124167469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 124267469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 124367469601SEric Dumazet 124467469601SEric Dumazet if (dst && dst_allfrag(dst)) 124567469601SEric Dumazet mss_now -= icsk->icsk_af_ops->net_frag_header_len; 124667469601SEric Dumazet } 124767469601SEric Dumazet 12485d424d5aSJohn Heffner /* Clamp it (mss_clamp does not include tcp options) */ 12495d424d5aSJohn Heffner if (mss_now > tp->rx_opt.mss_clamp) 12505d424d5aSJohn Heffner mss_now = tp->rx_opt.mss_clamp; 12515d424d5aSJohn Heffner 12525d424d5aSJohn Heffner /* Now subtract optional transport overhead */ 12535d424d5aSJohn Heffner mss_now -= icsk->icsk_ext_hdr_len; 12545d424d5aSJohn Heffner 12555d424d5aSJohn Heffner /* Then reserve room for full set of TCP options and 8 bytes of data */ 12565d424d5aSJohn Heffner if (mss_now < 48) 12575d424d5aSJohn Heffner mss_now = 48; 12585d424d5aSJohn Heffner return mss_now; 12595d424d5aSJohn Heffner } 12605d424d5aSJohn Heffner 12611b63edd6SYuchung Cheng /* Calculate MSS. Not accounting for SACKs here. */ 12621b63edd6SYuchung Cheng int tcp_mtu_to_mss(struct sock *sk, int pmtu) 12631b63edd6SYuchung Cheng { 12641b63edd6SYuchung Cheng /* Subtract TCP options size, not including SACKs */ 12651b63edd6SYuchung Cheng return __tcp_mtu_to_mss(sk, pmtu) - 12661b63edd6SYuchung Cheng (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr)); 12671b63edd6SYuchung Cheng } 12681b63edd6SYuchung Cheng 12695d424d5aSJohn Heffner /* Inverse of above */ 127067469601SEric Dumazet int tcp_mss_to_mtu(struct sock *sk, int mss) 12715d424d5aSJohn Heffner { 1272cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1273cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 12745d424d5aSJohn Heffner int mtu; 12755d424d5aSJohn Heffner 12765d424d5aSJohn Heffner mtu = mss + 12775d424d5aSJohn Heffner tp->tcp_header_len + 12785d424d5aSJohn Heffner icsk->icsk_ext_hdr_len + 12795d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 12805d424d5aSJohn Heffner 128167469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 128267469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 128367469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 128467469601SEric Dumazet 128567469601SEric Dumazet if (dst && dst_allfrag(dst)) 128667469601SEric Dumazet mtu += icsk->icsk_af_ops->net_frag_header_len; 128767469601SEric Dumazet } 12885d424d5aSJohn Heffner return mtu; 12895d424d5aSJohn Heffner } 12905d424d5aSJohn Heffner 129167edfef7SAndi Kleen /* MTU probing init per socket */ 12925d424d5aSJohn Heffner void tcp_mtup_init(struct sock *sk) 12935d424d5aSJohn Heffner { 12945d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 12955d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 12965d424d5aSJohn Heffner 12975d424d5aSJohn Heffner icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1; 12985d424d5aSJohn Heffner icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 12995d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 13005d424d5aSJohn Heffner icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss); 13015d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 13025d424d5aSJohn Heffner } 13034bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_mtup_init); 13045d424d5aSJohn Heffner 13051da177e4SLinus Torvalds /* This function synchronize snd mss to current pmtu/exthdr set. 13061da177e4SLinus Torvalds 13071da177e4SLinus Torvalds tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 13081da177e4SLinus Torvalds for TCP options, but includes only bare TCP header. 13091da177e4SLinus Torvalds 13101da177e4SLinus Torvalds tp->rx_opt.mss_clamp is mss negotiated at connection setup. 1311caa20d9aSStephen Hemminger It is minimum of user_mss and mss received with SYN. 13121da177e4SLinus Torvalds It also does not include TCP options. 13131da177e4SLinus Torvalds 1314d83d8461SArnaldo Carvalho de Melo inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. 13151da177e4SLinus Torvalds 13161da177e4SLinus Torvalds tp->mss_cache is current effective sending mss, including 13171da177e4SLinus Torvalds all tcp options except for SACKs. It is evaluated, 13181da177e4SLinus Torvalds taking into account current pmtu, but never exceeds 13191da177e4SLinus Torvalds tp->rx_opt.mss_clamp. 13201da177e4SLinus Torvalds 13211da177e4SLinus Torvalds NOTE1. rfc1122 clearly states that advertised MSS 13221da177e4SLinus Torvalds DOES NOT include either tcp or ip options. 13231da177e4SLinus Torvalds 1324d83d8461SArnaldo Carvalho de Melo NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache 1325d83d8461SArnaldo Carvalho de Melo are READ ONLY outside this function. --ANK (980731) 13261da177e4SLinus Torvalds */ 13271da177e4SLinus Torvalds unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 13281da177e4SLinus Torvalds { 13291da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1330d83d8461SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 13315d424d5aSJohn Heffner int mss_now; 13321da177e4SLinus Torvalds 13335d424d5aSJohn Heffner if (icsk->icsk_mtup.search_high > pmtu) 13345d424d5aSJohn Heffner icsk->icsk_mtup.search_high = pmtu; 13351da177e4SLinus Torvalds 13365d424d5aSJohn Heffner mss_now = tcp_mtu_to_mss(sk, pmtu); 1337409d22b4SIlpo Järvinen mss_now = tcp_bound_to_half_wnd(tp, mss_now); 13381da177e4SLinus Torvalds 13391da177e4SLinus Torvalds /* And store cached results */ 1340d83d8461SArnaldo Carvalho de Melo icsk->icsk_pmtu_cookie = pmtu; 13415d424d5aSJohn Heffner if (icsk->icsk_mtup.enabled) 13425d424d5aSJohn Heffner mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); 1343c1b4a7e6SDavid S. Miller tp->mss_cache = mss_now; 13441da177e4SLinus Torvalds 13451da177e4SLinus Torvalds return mss_now; 13461da177e4SLinus Torvalds } 13474bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_sync_mss); 13481da177e4SLinus Torvalds 13491da177e4SLinus Torvalds /* Compute the current effective MSS, taking SACKs and IP options, 13501da177e4SLinus Torvalds * and even PMTU discovery events into account. 13511da177e4SLinus Torvalds */ 13520c54b85fSIlpo Järvinen unsigned int tcp_current_mss(struct sock *sk) 13531da177e4SLinus Torvalds { 1354cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1355cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1356c1b4a7e6SDavid S. Miller u32 mss_now; 135795c96174SEric Dumazet unsigned int header_len; 135833ad798cSAdam Langley struct tcp_out_options opts; 135933ad798cSAdam Langley struct tcp_md5sig_key *md5; 13601da177e4SLinus Torvalds 1361c1b4a7e6SDavid S. Miller mss_now = tp->mss_cache; 1362c1b4a7e6SDavid S. Miller 13631da177e4SLinus Torvalds if (dst) { 13641da177e4SLinus Torvalds u32 mtu = dst_mtu(dst); 1365d83d8461SArnaldo Carvalho de Melo if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 13661da177e4SLinus Torvalds mss_now = tcp_sync_mss(sk, mtu); 13671da177e4SLinus Torvalds } 13681da177e4SLinus Torvalds 136933ad798cSAdam Langley header_len = tcp_established_options(sk, NULL, &opts, &md5) + 137033ad798cSAdam Langley sizeof(struct tcphdr); 137133ad798cSAdam Langley /* The mss_cache is sized based on tp->tcp_header_len, which assumes 137233ad798cSAdam Langley * some common options. If this is an odd packet (because we have SACK 137333ad798cSAdam Langley * blocks etc) then our calculated header_len will be different, and 137433ad798cSAdam Langley * we have to adjust mss_now correspondingly */ 137533ad798cSAdam Langley if (header_len != tp->tcp_header_len) { 137633ad798cSAdam Langley int delta = (int) header_len - tp->tcp_header_len; 137733ad798cSAdam Langley mss_now -= delta; 137833ad798cSAdam Langley } 1379cfb6eeb4SYOSHIFUJI Hideaki 13801da177e4SLinus Torvalds return mss_now; 13811da177e4SLinus Torvalds } 13821da177e4SLinus Torvalds 138386fd14adSWeiping Pan /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. 138486fd14adSWeiping Pan * As additional protections, we do not touch cwnd in retransmission phases, 138586fd14adSWeiping Pan * and if application hit its sndbuf limit recently. 138686fd14adSWeiping Pan */ 138786fd14adSWeiping Pan static void tcp_cwnd_application_limited(struct sock *sk) 138886fd14adSWeiping Pan { 138986fd14adSWeiping Pan struct tcp_sock *tp = tcp_sk(sk); 139086fd14adSWeiping Pan 139186fd14adSWeiping Pan if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && 139286fd14adSWeiping Pan sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 139386fd14adSWeiping Pan /* Limited by application or receiver window. */ 139486fd14adSWeiping Pan u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); 139586fd14adSWeiping Pan u32 win_used = max(tp->snd_cwnd_used, init_win); 139686fd14adSWeiping Pan if (win_used < tp->snd_cwnd) { 139786fd14adSWeiping Pan tp->snd_ssthresh = tcp_current_ssthresh(sk); 139886fd14adSWeiping Pan tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1; 139986fd14adSWeiping Pan } 140086fd14adSWeiping Pan tp->snd_cwnd_used = 0; 140186fd14adSWeiping Pan } 140286fd14adSWeiping Pan tp->snd_cwnd_stamp = tcp_time_stamp; 140386fd14adSWeiping Pan } 140486fd14adSWeiping Pan 1405ca8a2263SNeal Cardwell static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited) 1406a762a980SDavid S. Miller { 14079e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1408a762a980SDavid S. Miller 1409ca8a2263SNeal Cardwell /* Track the maximum number of outstanding packets in each 1410ca8a2263SNeal Cardwell * window, and remember whether we were cwnd-limited then. 1411ca8a2263SNeal Cardwell */ 1412ca8a2263SNeal Cardwell if (!before(tp->snd_una, tp->max_packets_seq) || 1413ca8a2263SNeal Cardwell tp->packets_out > tp->max_packets_out) { 1414ca8a2263SNeal Cardwell tp->max_packets_out = tp->packets_out; 1415ca8a2263SNeal Cardwell tp->max_packets_seq = tp->snd_nxt; 1416ca8a2263SNeal Cardwell tp->is_cwnd_limited = is_cwnd_limited; 1417ca8a2263SNeal Cardwell } 1418e114a710SEric Dumazet 141924901551SEric Dumazet if (tcp_is_cwnd_limited(sk)) { 1420a762a980SDavid S. Miller /* Network is feed fully. */ 1421a762a980SDavid S. Miller tp->snd_cwnd_used = 0; 1422a762a980SDavid S. Miller tp->snd_cwnd_stamp = tcp_time_stamp; 1423a762a980SDavid S. Miller } else { 1424a762a980SDavid S. Miller /* Network starves. */ 1425a762a980SDavid S. Miller if (tp->packets_out > tp->snd_cwnd_used) 1426a762a980SDavid S. Miller tp->snd_cwnd_used = tp->packets_out; 1427a762a980SDavid S. Miller 142815d33c07SDavid S. Miller if (sysctl_tcp_slow_start_after_idle && 142915d33c07SDavid S. Miller (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto) 1430a762a980SDavid S. Miller tcp_cwnd_application_limited(sk); 1431a762a980SDavid S. Miller } 1432a762a980SDavid S. Miller } 1433a762a980SDavid S. Miller 1434d4589926SEric Dumazet /* Minshall's variant of the Nagle send check. */ 1435d4589926SEric Dumazet static bool tcp_minshall_check(const struct tcp_sock *tp) 1436d4589926SEric Dumazet { 1437d4589926SEric Dumazet return after(tp->snd_sml, tp->snd_una) && 1438d4589926SEric Dumazet !after(tp->snd_sml, tp->snd_nxt); 1439d4589926SEric Dumazet } 1440d4589926SEric Dumazet 1441d4589926SEric Dumazet /* Update snd_sml if this skb is under mss 1442d4589926SEric Dumazet * Note that a TSO packet might end with a sub-mss segment 1443d4589926SEric Dumazet * The test is really : 1444d4589926SEric Dumazet * if ((skb->len % mss) != 0) 1445d4589926SEric Dumazet * tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1446d4589926SEric Dumazet * But we can avoid doing the divide again given we already have 1447d4589926SEric Dumazet * skb_pcount = skb->len / mss_now 14480e3a4803SIlpo Järvinen */ 1449d4589926SEric Dumazet static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now, 1450d4589926SEric Dumazet const struct sk_buff *skb) 1451d4589926SEric Dumazet { 1452d4589926SEric Dumazet if (skb->len < tcp_skb_pcount(skb) * mss_now) 1453d4589926SEric Dumazet tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1454d4589926SEric Dumazet } 1455d4589926SEric Dumazet 1456d4589926SEric Dumazet /* Return false, if packet can be sent now without violation Nagle's rules: 1457d4589926SEric Dumazet * 1. It is full sized. (provided by caller in %partial bool) 1458d4589926SEric Dumazet * 2. Or it contains FIN. (already checked by caller) 1459d4589926SEric Dumazet * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. 1460d4589926SEric Dumazet * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 1461d4589926SEric Dumazet * With Minshall's modification: all sent small packets are ACKed. 1462d4589926SEric Dumazet */ 1463d4589926SEric Dumazet static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp, 1464cc93fc51SPeter Pan(潘卫平) int nonagle) 1465d4589926SEric Dumazet { 1466d4589926SEric Dumazet return partial && 1467d4589926SEric Dumazet ((nonagle & TCP_NAGLE_CORK) || 1468d4589926SEric Dumazet (!nonagle && tp->packets_out && tcp_minshall_check(tp))); 1469d4589926SEric Dumazet } 1470d4589926SEric Dumazet /* Returns the portion of skb which can be sent right away */ 1471d4589926SEric Dumazet static unsigned int tcp_mss_split_point(const struct sock *sk, 1472d4589926SEric Dumazet const struct sk_buff *skb, 1473d4589926SEric Dumazet unsigned int mss_now, 1474d4589926SEric Dumazet unsigned int max_segs, 1475d4589926SEric Dumazet int nonagle) 1476c1b4a7e6SDavid S. Miller { 1477cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1478d4589926SEric Dumazet u32 partial, needed, window, max_len; 1479c1b4a7e6SDavid S. Miller 148090840defSIlpo Järvinen window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 14811485348dSBen Hutchings max_len = mss_now * max_segs; 14820e3a4803SIlpo Järvinen 14831485348dSBen Hutchings if (likely(max_len <= window && skb != tcp_write_queue_tail(sk))) 14841485348dSBen Hutchings return max_len; 14850e3a4803SIlpo Järvinen 14865ea3a748SIlpo Järvinen needed = min(skb->len, window); 14875ea3a748SIlpo Järvinen 14881485348dSBen Hutchings if (max_len <= needed) 14891485348dSBen Hutchings return max_len; 14900e3a4803SIlpo Järvinen 1491d4589926SEric Dumazet partial = needed % mss_now; 1492d4589926SEric Dumazet /* If last segment is not a full MSS, check if Nagle rules allow us 1493d4589926SEric Dumazet * to include this last segment in this skb. 1494d4589926SEric Dumazet * Otherwise, we'll split the skb at last MSS boundary 1495d4589926SEric Dumazet */ 1496cc93fc51SPeter Pan(潘卫平) if (tcp_nagle_check(partial != 0, tp, nonagle)) 1497d4589926SEric Dumazet return needed - partial; 1498d4589926SEric Dumazet 1499d4589926SEric Dumazet return needed; 1500c1b4a7e6SDavid S. Miller } 1501c1b4a7e6SDavid S. Miller 1502c1b4a7e6SDavid S. Miller /* Can at least one segment of SKB be sent right now, according to the 1503c1b4a7e6SDavid S. Miller * congestion window rules? If so, return how many segments are allowed. 1504c1b4a7e6SDavid S. Miller */ 1505cf533ea5SEric Dumazet static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, 1506cf533ea5SEric Dumazet const struct sk_buff *skb) 1507c1b4a7e6SDavid S. Miller { 1508c1b4a7e6SDavid S. Miller u32 in_flight, cwnd; 1509c1b4a7e6SDavid S. Miller 1510c1b4a7e6SDavid S. Miller /* Don't be strict about the congestion window for the final FIN. */ 15114de075e0SEric Dumazet if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && 15124de075e0SEric Dumazet tcp_skb_pcount(skb) == 1) 1513c1b4a7e6SDavid S. Miller return 1; 1514c1b4a7e6SDavid S. Miller 1515c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1516c1b4a7e6SDavid S. Miller cwnd = tp->snd_cwnd; 1517c1b4a7e6SDavid S. Miller if (in_flight < cwnd) 1518c1b4a7e6SDavid S. Miller return (cwnd - in_flight); 1519c1b4a7e6SDavid S. Miller 1520c1b4a7e6SDavid S. Miller return 0; 1521c1b4a7e6SDavid S. Miller } 1522c1b4a7e6SDavid S. Miller 1523b595076aSUwe Kleine-König /* Initialize TSO state of a skb. 152467edfef7SAndi Kleen * This must be invoked the first time we consider transmitting 1525c1b4a7e6SDavid S. Miller * SKB onto the wire. 1526c1b4a7e6SDavid S. Miller */ 1527cf533ea5SEric Dumazet static int tcp_init_tso_segs(const struct sock *sk, struct sk_buff *skb, 1528056834d9SIlpo Järvinen unsigned int mss_now) 1529c1b4a7e6SDavid S. Miller { 1530c1b4a7e6SDavid S. Miller int tso_segs = tcp_skb_pcount(skb); 1531c1b4a7e6SDavid S. Miller 1532f8269a49SIlpo Järvinen if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { 1533846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1534c1b4a7e6SDavid S. Miller tso_segs = tcp_skb_pcount(skb); 1535c1b4a7e6SDavid S. Miller } 1536c1b4a7e6SDavid S. Miller return tso_segs; 1537c1b4a7e6SDavid S. Miller } 1538c1b4a7e6SDavid S. Miller 1539c1b4a7e6SDavid S. Miller 1540a2a385d6SEric Dumazet /* Return true if the Nagle test allows this packet to be 1541c1b4a7e6SDavid S. Miller * sent now. 1542c1b4a7e6SDavid S. Miller */ 1543a2a385d6SEric Dumazet static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, 1544c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1545c1b4a7e6SDavid S. Miller { 1546c1b4a7e6SDavid S. Miller /* Nagle rule does not apply to frames, which sit in the middle of the 1547c1b4a7e6SDavid S. Miller * write_queue (they have no chances to get new data). 1548c1b4a7e6SDavid S. Miller * 1549c1b4a7e6SDavid S. Miller * This is implemented in the callers, where they modify the 'nonagle' 1550c1b4a7e6SDavid S. Miller * argument based upon the location of SKB in the send queue. 1551c1b4a7e6SDavid S. Miller */ 1552c1b4a7e6SDavid S. Miller if (nonagle & TCP_NAGLE_PUSH) 1553a2a385d6SEric Dumazet return true; 1554c1b4a7e6SDavid S. Miller 15559b44190dSYuchung Cheng /* Don't use the nagle rule for urgent data (or for the final FIN). */ 15569b44190dSYuchung Cheng if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) 1557a2a385d6SEric Dumazet return true; 1558c1b4a7e6SDavid S. Miller 1559cc93fc51SPeter Pan(潘卫平) if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle)) 1560a2a385d6SEric Dumazet return true; 1561c1b4a7e6SDavid S. Miller 1562a2a385d6SEric Dumazet return false; 1563c1b4a7e6SDavid S. Miller } 1564c1b4a7e6SDavid S. Miller 1565c1b4a7e6SDavid S. Miller /* Does at least the first segment of SKB fit into the send window? */ 1566a2a385d6SEric Dumazet static bool tcp_snd_wnd_test(const struct tcp_sock *tp, 1567a2a385d6SEric Dumazet const struct sk_buff *skb, 1568056834d9SIlpo Järvinen unsigned int cur_mss) 1569c1b4a7e6SDavid S. Miller { 1570c1b4a7e6SDavid S. Miller u32 end_seq = TCP_SKB_CB(skb)->end_seq; 1571c1b4a7e6SDavid S. Miller 1572c1b4a7e6SDavid S. Miller if (skb->len > cur_mss) 1573c1b4a7e6SDavid S. Miller end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 1574c1b4a7e6SDavid S. Miller 157590840defSIlpo Järvinen return !after(end_seq, tcp_wnd_end(tp)); 1576c1b4a7e6SDavid S. Miller } 1577c1b4a7e6SDavid S. Miller 1578fe067e8aSDavid S. Miller /* This checks if the data bearing packet SKB (usually tcp_send_head(sk)) 1579c1b4a7e6SDavid S. Miller * should be put on the wire right now. If so, it returns the number of 1580c1b4a7e6SDavid S. Miller * packets allowed by the congestion window. 1581c1b4a7e6SDavid S. Miller */ 1582cf533ea5SEric Dumazet static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb, 1583c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1584c1b4a7e6SDavid S. Miller { 1585cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1586c1b4a7e6SDavid S. Miller unsigned int cwnd_quota; 1587c1b4a7e6SDavid S. Miller 1588846998aeSDavid S. Miller tcp_init_tso_segs(sk, skb, cur_mss); 1589c1b4a7e6SDavid S. Miller 1590c1b4a7e6SDavid S. Miller if (!tcp_nagle_test(tp, skb, cur_mss, nonagle)) 1591c1b4a7e6SDavid S. Miller return 0; 1592c1b4a7e6SDavid S. Miller 1593c1b4a7e6SDavid S. Miller cwnd_quota = tcp_cwnd_test(tp, skb); 1594056834d9SIlpo Järvinen if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss)) 1595c1b4a7e6SDavid S. Miller cwnd_quota = 0; 1596c1b4a7e6SDavid S. Miller 1597c1b4a7e6SDavid S. Miller return cwnd_quota; 1598c1b4a7e6SDavid S. Miller } 1599c1b4a7e6SDavid S. Miller 160067edfef7SAndi Kleen /* Test if sending is allowed right now. */ 1601a2a385d6SEric Dumazet bool tcp_may_send_now(struct sock *sk) 1602c1b4a7e6SDavid S. Miller { 1603cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1604fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 1605c1b4a7e6SDavid S. Miller 1606a02cec21SEric Dumazet return skb && 16070c54b85fSIlpo Järvinen tcp_snd_test(sk, skb, tcp_current_mss(sk), 1608c1b4a7e6SDavid S. Miller (tcp_skb_is_last(sk, skb) ? 1609a02cec21SEric Dumazet tp->nonagle : TCP_NAGLE_PUSH)); 1610c1b4a7e6SDavid S. Miller } 1611c1b4a7e6SDavid S. Miller 1612c1b4a7e6SDavid S. Miller /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 1613c1b4a7e6SDavid S. Miller * which is put after SKB on the list. It is very much like 1614c1b4a7e6SDavid S. Miller * tcp_fragment() except that it may make several kinds of assumptions 1615c1b4a7e6SDavid S. Miller * in order to speed up the splitting operation. In particular, we 1616c1b4a7e6SDavid S. Miller * know that all the data is in scatter-gather pages, and that the 1617c1b4a7e6SDavid S. Miller * packet has never been sent out before (and thus is not cloned). 1618c1b4a7e6SDavid S. Miller */ 1619056834d9SIlpo Järvinen static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, 1620c4ead4c5SEric Dumazet unsigned int mss_now, gfp_t gfp) 1621c1b4a7e6SDavid S. Miller { 1622c1b4a7e6SDavid S. Miller struct sk_buff *buff; 1623c1b4a7e6SDavid S. Miller int nlen = skb->len - len; 16249ce01461SIlpo Järvinen u8 flags; 1625c1b4a7e6SDavid S. Miller 1626c1b4a7e6SDavid S. Miller /* All of a TSO frame must be composed of paged data. */ 1627c8ac3774SHerbert Xu if (skb->len != skb->data_len) 1628*6cc55e09SOctavian Purdila return tcp_fragment(sk, skb, len, mss_now, gfp); 1629c1b4a7e6SDavid S. Miller 1630c4ead4c5SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, gfp); 1631c1b4a7e6SDavid S. Miller if (unlikely(buff == NULL)) 1632c1b4a7e6SDavid S. Miller return -ENOMEM; 1633c1b4a7e6SDavid S. Miller 16343ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 16353ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1636b60b49eaSHerbert Xu buff->truesize += nlen; 1637c1b4a7e6SDavid S. Miller skb->truesize -= nlen; 1638c1b4a7e6SDavid S. Miller 1639c1b4a7e6SDavid S. Miller /* Correct the sequence numbers. */ 1640c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 1641c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 1642c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 1643c1b4a7e6SDavid S. Miller 1644c1b4a7e6SDavid S. Miller /* PSH and FIN should only be set in the second packet. */ 16454de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 16464de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 16474de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1648c1b4a7e6SDavid S. Miller 1649c1b4a7e6SDavid S. Miller /* This packet was never sent out yet, so no SACK bits. */ 1650c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->sacked = 0; 1651c1b4a7e6SDavid S. Miller 165284fa7933SPatrick McHardy buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; 1653c1b4a7e6SDavid S. Miller skb_split(skb, buff, len); 1654c1b4a7e6SDavid S. Miller 1655c1b4a7e6SDavid S. Miller /* Fix up tso_factor for both original and new SKB. */ 1656846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1657846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, buff, mss_now); 1658c1b4a7e6SDavid S. Miller 1659c1b4a7e6SDavid S. Miller /* Link BUFF into the send queue. */ 1660c1b4a7e6SDavid S. Miller skb_header_release(buff); 1661fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 1662c1b4a7e6SDavid S. Miller 1663c1b4a7e6SDavid S. Miller return 0; 1664c1b4a7e6SDavid S. Miller } 1665c1b4a7e6SDavid S. Miller 1666c1b4a7e6SDavid S. Miller /* Try to defer sending, if possible, in order to minimize the amount 1667c1b4a7e6SDavid S. Miller * of TSO splitting we do. View it as a kind of TSO Nagle test. 1668c1b4a7e6SDavid S. Miller * 1669c1b4a7e6SDavid S. Miller * This algorithm is from John Heffner. 1670c1b4a7e6SDavid S. Miller */ 1671ca8a2263SNeal Cardwell static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb, 1672ca8a2263SNeal Cardwell bool *is_cwnd_limited) 1673c1b4a7e6SDavid S. Miller { 16749e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 16756687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 1676c1b4a7e6SDavid S. Miller u32 send_win, cong_win, limit, in_flight; 1677ad9f4f50SEric Dumazet int win_divisor; 1678c1b4a7e6SDavid S. Miller 16794de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 1680ae8064acSJohn Heffner goto send_now; 1681c1b4a7e6SDavid S. Miller 16826687e988SArnaldo Carvalho de Melo if (icsk->icsk_ca_state != TCP_CA_Open) 1683ae8064acSJohn Heffner goto send_now; 1684ae8064acSJohn Heffner 1685ae8064acSJohn Heffner /* Defer for less than two clock ticks. */ 1686bd515c3eSIlpo Järvinen if (tp->tso_deferred && 1687a2acde07SIlpo Järvinen (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1) 1688ae8064acSJohn Heffner goto send_now; 1689908a75c1SDavid S. Miller 1690c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1691c1b4a7e6SDavid S. Miller 1692056834d9SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight)); 1693c1b4a7e6SDavid S. Miller 169490840defSIlpo Järvinen send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1695c1b4a7e6SDavid S. Miller 1696c1b4a7e6SDavid S. Miller /* From in_flight test above, we know that cwnd > in_flight. */ 1697c1b4a7e6SDavid S. Miller cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; 1698c1b4a7e6SDavid S. Miller 1699c1b4a7e6SDavid S. Miller limit = min(send_win, cong_win); 1700c1b4a7e6SDavid S. Miller 1701ba244fe9SDavid S. Miller /* If a full-sized TSO skb can be sent, do it. */ 17021485348dSBen Hutchings if (limit >= min_t(unsigned int, sk->sk_gso_max_size, 170395bd09ebSEric Dumazet tp->xmit_size_goal_segs * tp->mss_cache)) 1704ae8064acSJohn Heffner goto send_now; 1705ba244fe9SDavid S. Miller 170662ad2761SIlpo Järvinen /* Middle in queue won't get any more data, full sendable already? */ 170762ad2761SIlpo Järvinen if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) 170862ad2761SIlpo Järvinen goto send_now; 170962ad2761SIlpo Järvinen 1710ad9f4f50SEric Dumazet win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor); 1711ad9f4f50SEric Dumazet if (win_divisor) { 1712c1b4a7e6SDavid S. Miller u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); 1713c1b4a7e6SDavid S. Miller 1714c1b4a7e6SDavid S. Miller /* If at least some fraction of a window is available, 1715c1b4a7e6SDavid S. Miller * just use it. 1716c1b4a7e6SDavid S. Miller */ 1717ad9f4f50SEric Dumazet chunk /= win_divisor; 1718c1b4a7e6SDavid S. Miller if (limit >= chunk) 1719ae8064acSJohn Heffner goto send_now; 1720c1b4a7e6SDavid S. Miller } else { 1721c1b4a7e6SDavid S. Miller /* Different approach, try not to defer past a single 1722c1b4a7e6SDavid S. Miller * ACK. Receiver should ACK every other full sized 1723c1b4a7e6SDavid S. Miller * frame, so if we have space for more than 3 frames 1724c1b4a7e6SDavid S. Miller * then send now. 1725c1b4a7e6SDavid S. Miller */ 17266b5a5c0dSNeal Cardwell if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) 1727ae8064acSJohn Heffner goto send_now; 1728c1b4a7e6SDavid S. Miller } 1729c1b4a7e6SDavid S. Miller 1730f4541d60SEric Dumazet /* Ok, it looks like it is advisable to defer. 1731f4541d60SEric Dumazet * Do not rearm the timer if already set to not break TCP ACK clocking. 1732f4541d60SEric Dumazet */ 1733f4541d60SEric Dumazet if (!tp->tso_deferred) 1734ae8064acSJohn Heffner tp->tso_deferred = 1 | (jiffies << 1); 1735ae8064acSJohn Heffner 1736ca8a2263SNeal Cardwell if (cong_win < send_win && cong_win < skb->len) 1737ca8a2263SNeal Cardwell *is_cwnd_limited = true; 1738ca8a2263SNeal Cardwell 1739a2a385d6SEric Dumazet return true; 1740ae8064acSJohn Heffner 1741ae8064acSJohn Heffner send_now: 1742ae8064acSJohn Heffner tp->tso_deferred = 0; 1743a2a385d6SEric Dumazet return false; 1744c1b4a7e6SDavid S. Miller } 1745c1b4a7e6SDavid S. Miller 17465d424d5aSJohn Heffner /* Create a new MTU probe if we are ready. 174767edfef7SAndi Kleen * MTU probe is regularly attempting to increase the path MTU by 174867edfef7SAndi Kleen * deliberately sending larger packets. This discovers routing 174967edfef7SAndi Kleen * changes resulting in larger path MTUs. 175067edfef7SAndi Kleen * 17515d424d5aSJohn Heffner * Returns 0 if we should wait to probe (no cwnd available), 17525d424d5aSJohn Heffner * 1 if a probe was sent, 1753056834d9SIlpo Järvinen * -1 otherwise 1754056834d9SIlpo Järvinen */ 17555d424d5aSJohn Heffner static int tcp_mtu_probe(struct sock *sk) 17565d424d5aSJohn Heffner { 17575d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 17585d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 17595d424d5aSJohn Heffner struct sk_buff *skb, *nskb, *next; 17605d424d5aSJohn Heffner int len; 17615d424d5aSJohn Heffner int probe_size; 176291cc17c0SIlpo Järvinen int size_needed; 17635d424d5aSJohn Heffner int copy; 17645d424d5aSJohn Heffner int mss_now; 17655d424d5aSJohn Heffner 17665d424d5aSJohn Heffner /* Not currently probing/verifying, 17675d424d5aSJohn Heffner * not in recovery, 17685d424d5aSJohn Heffner * have enough cwnd, and 17695d424d5aSJohn Heffner * not SACKing (the variable headers throw things off) */ 17705d424d5aSJohn Heffner if (!icsk->icsk_mtup.enabled || 17715d424d5aSJohn Heffner icsk->icsk_mtup.probe_size || 17725d424d5aSJohn Heffner inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 17735d424d5aSJohn Heffner tp->snd_cwnd < 11 || 1774cabeccbdSIlpo Järvinen tp->rx_opt.num_sacks || tp->rx_opt.dsack) 17755d424d5aSJohn Heffner return -1; 17765d424d5aSJohn Heffner 17775d424d5aSJohn Heffner /* Very simple search strategy: just double the MSS. */ 17780c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 17795d424d5aSJohn Heffner probe_size = 2 * tp->mss_cache; 178091cc17c0SIlpo Järvinen size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; 17815d424d5aSJohn Heffner if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) { 17825d424d5aSJohn Heffner /* TODO: set timer for probe_converge_event */ 17835d424d5aSJohn Heffner return -1; 17845d424d5aSJohn Heffner } 17855d424d5aSJohn Heffner 17865d424d5aSJohn Heffner /* Have enough data in the send queue to probe? */ 17877f9c33e5SIlpo Järvinen if (tp->write_seq - tp->snd_nxt < size_needed) 17885d424d5aSJohn Heffner return -1; 17895d424d5aSJohn Heffner 179091cc17c0SIlpo Järvinen if (tp->snd_wnd < size_needed) 17915d424d5aSJohn Heffner return -1; 179290840defSIlpo Järvinen if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) 17935d424d5aSJohn Heffner return 0; 17945d424d5aSJohn Heffner 1795d67c58e9SIlpo Järvinen /* Do we need to wait to drain cwnd? With none in flight, don't stall */ 1796d67c58e9SIlpo Järvinen if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { 1797d67c58e9SIlpo Järvinen if (!tcp_packets_in_flight(tp)) 17985d424d5aSJohn Heffner return -1; 17995d424d5aSJohn Heffner else 18005d424d5aSJohn Heffner return 0; 18015d424d5aSJohn Heffner } 18025d424d5aSJohn Heffner 18035d424d5aSJohn Heffner /* We're allowed to probe. Build it now. */ 18045d424d5aSJohn Heffner if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL) 18055d424d5aSJohn Heffner return -1; 18063ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 18073ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 18085d424d5aSJohn Heffner 1809fe067e8aSDavid S. Miller skb = tcp_send_head(sk); 18105d424d5aSJohn Heffner 18115d424d5aSJohn Heffner TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 18125d424d5aSJohn Heffner TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 18134de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; 18145d424d5aSJohn Heffner TCP_SKB_CB(nskb)->sacked = 0; 18155d424d5aSJohn Heffner nskb->csum = 0; 181684fa7933SPatrick McHardy nskb->ip_summed = skb->ip_summed; 18175d424d5aSJohn Heffner 181850c4817eSIlpo Järvinen tcp_insert_write_queue_before(nskb, skb, sk); 181950c4817eSIlpo Järvinen 18205d424d5aSJohn Heffner len = 0; 1821234b6860SIlpo Järvinen tcp_for_write_queue_from_safe(skb, next, sk) { 18225d424d5aSJohn Heffner copy = min_t(int, skb->len, probe_size - len); 18235d424d5aSJohn Heffner if (nskb->ip_summed) 18245d424d5aSJohn Heffner skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); 18255d424d5aSJohn Heffner else 18265d424d5aSJohn Heffner nskb->csum = skb_copy_and_csum_bits(skb, 0, 1827056834d9SIlpo Järvinen skb_put(nskb, copy), 1828056834d9SIlpo Järvinen copy, nskb->csum); 18295d424d5aSJohn Heffner 18305d424d5aSJohn Heffner if (skb->len <= copy) { 18315d424d5aSJohn Heffner /* We've eaten all the data from this skb. 18325d424d5aSJohn Heffner * Throw it away. */ 18334de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 1834fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 18353ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 18365d424d5aSJohn Heffner } else { 18374de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & 1838a3433f35SChangli Gao ~(TCPHDR_FIN|TCPHDR_PSH); 18395d424d5aSJohn Heffner if (!skb_shinfo(skb)->nr_frags) { 18405d424d5aSJohn Heffner skb_pull(skb, copy); 184184fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 1842056834d9SIlpo Järvinen skb->csum = csum_partial(skb->data, 1843056834d9SIlpo Järvinen skb->len, 0); 18445d424d5aSJohn Heffner } else { 18455d424d5aSJohn Heffner __pskb_trim_head(skb, copy); 18465d424d5aSJohn Heffner tcp_set_skb_tso_segs(sk, skb, mss_now); 18475d424d5aSJohn Heffner } 18485d424d5aSJohn Heffner TCP_SKB_CB(skb)->seq += copy; 18495d424d5aSJohn Heffner } 18505d424d5aSJohn Heffner 18515d424d5aSJohn Heffner len += copy; 1852234b6860SIlpo Järvinen 1853234b6860SIlpo Järvinen if (len >= probe_size) 1854234b6860SIlpo Järvinen break; 18555d424d5aSJohn Heffner } 18565d424d5aSJohn Heffner tcp_init_tso_segs(sk, nskb, nskb->len); 18575d424d5aSJohn Heffner 18585d424d5aSJohn Heffner /* We're ready to send. If this fails, the probe will 18595d424d5aSJohn Heffner * be resegmented into mss-sized pieces by tcp_write_xmit(). */ 18605d424d5aSJohn Heffner TCP_SKB_CB(nskb)->when = tcp_time_stamp; 18615d424d5aSJohn Heffner if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 18625d424d5aSJohn Heffner /* Decrement cwnd here because we are sending 18635d424d5aSJohn Heffner * effectively two packets. */ 18645d424d5aSJohn Heffner tp->snd_cwnd--; 186566f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, nskb); 18665d424d5aSJohn Heffner 18675d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 18680e7b1368SJohn Heffner tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 18690e7b1368SJohn Heffner tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 18705d424d5aSJohn Heffner 18715d424d5aSJohn Heffner return 1; 18725d424d5aSJohn Heffner } 18735d424d5aSJohn Heffner 18745d424d5aSJohn Heffner return -1; 18755d424d5aSJohn Heffner } 18765d424d5aSJohn Heffner 18771da177e4SLinus Torvalds /* This routine writes packets to the network. It advances the 18781da177e4SLinus Torvalds * send_head. This happens as incoming acks open up the remote 18791da177e4SLinus Torvalds * window for us. 18801da177e4SLinus Torvalds * 1881f8269a49SIlpo Järvinen * LARGESEND note: !tcp_urg_mode is overkill, only frames between 1882f8269a49SIlpo Järvinen * snd_up-64k-mss .. snd_up cannot be large. However, taking into 1883f8269a49SIlpo Järvinen * account rare use of URG, this is not a big flaw. 1884f8269a49SIlpo Järvinen * 18856ba8a3b1SNandita Dukkipati * Send at most one packet when push_one > 0. Temporarily ignore 18866ba8a3b1SNandita Dukkipati * cwnd limit to force at most one packet out when push_one == 2. 18876ba8a3b1SNandita Dukkipati 1888a2a385d6SEric Dumazet * Returns true, if no segments are in flight and we have queued segments, 1889a2a385d6SEric Dumazet * but cannot send anything now because of SWS or another problem. 18901da177e4SLinus Torvalds */ 1891a2a385d6SEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 1892d5dd9175SIlpo Järvinen int push_one, gfp_t gfp) 18931da177e4SLinus Torvalds { 18941da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 189592df7b51SDavid S. Miller struct sk_buff *skb; 1896ca8a2263SNeal Cardwell unsigned int tso_segs, sent_pkts; 1897c1b4a7e6SDavid S. Miller int cwnd_quota; 18985d424d5aSJohn Heffner int result; 1899ca8a2263SNeal Cardwell bool is_cwnd_limited = false; 19001da177e4SLinus Torvalds 1901c1b4a7e6SDavid S. Miller sent_pkts = 0; 19025d424d5aSJohn Heffner 1903d5dd9175SIlpo Järvinen if (!push_one) { 19045d424d5aSJohn Heffner /* Do MTU probing. */ 1905d5dd9175SIlpo Järvinen result = tcp_mtu_probe(sk); 1906d5dd9175SIlpo Järvinen if (!result) { 1907a2a385d6SEric Dumazet return false; 19085d424d5aSJohn Heffner } else if (result > 0) { 19095d424d5aSJohn Heffner sent_pkts = 1; 19105d424d5aSJohn Heffner } 1911d5dd9175SIlpo Järvinen } 19125d424d5aSJohn Heffner 1913fe067e8aSDavid S. Miller while ((skb = tcp_send_head(sk))) { 1914c8ac3774SHerbert Xu unsigned int limit; 1915c8ac3774SHerbert Xu 1916b68e9f85SHerbert Xu tso_segs = tcp_init_tso_segs(sk, skb, mss_now); 1917c1b4a7e6SDavid S. Miller BUG_ON(!tso_segs); 1918c1b4a7e6SDavid S. Miller 1919ec342325SAndrew Vagin if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) 1920ec342325SAndrew Vagin goto repair; /* Skip network transmission */ 1921ec342325SAndrew Vagin 1922b68e9f85SHerbert Xu cwnd_quota = tcp_cwnd_test(tp, skb); 19236ba8a3b1SNandita Dukkipati if (!cwnd_quota) { 1924ca8a2263SNeal Cardwell is_cwnd_limited = true; 19256ba8a3b1SNandita Dukkipati if (push_one == 2) 19266ba8a3b1SNandita Dukkipati /* Force out a loss probe pkt. */ 19276ba8a3b1SNandita Dukkipati cwnd_quota = 1; 19286ba8a3b1SNandita Dukkipati else 1929b68e9f85SHerbert Xu break; 19306ba8a3b1SNandita Dukkipati } 1931b68e9f85SHerbert Xu 1932b68e9f85SHerbert Xu if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) 1933b68e9f85SHerbert Xu break; 1934b68e9f85SHerbert Xu 1935c1b4a7e6SDavid S. Miller if (tso_segs == 1) { 1936aa93466bSDavid S. Miller if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 1937aa93466bSDavid S. Miller (tcp_skb_is_last(sk, skb) ? 1938aa93466bSDavid S. Miller nonagle : TCP_NAGLE_PUSH)))) 1939aa93466bSDavid S. Miller break; 1940c1b4a7e6SDavid S. Miller } else { 1941ca8a2263SNeal Cardwell if (!push_one && 1942ca8a2263SNeal Cardwell tcp_tso_should_defer(sk, skb, &is_cwnd_limited)) 1943ca8a2263SNeal Cardwell break; 1944c1b4a7e6SDavid S. Miller } 1945aa93466bSDavid S. Miller 1946c9eeec26SEric Dumazet /* TCP Small Queues : 1947c9eeec26SEric Dumazet * Control number of packets in qdisc/devices to two packets / or ~1 ms. 1948c9eeec26SEric Dumazet * This allows for : 1949c9eeec26SEric Dumazet * - better RTT estimation and ACK scheduling 1950c9eeec26SEric Dumazet * - faster recovery 1951c9eeec26SEric Dumazet * - high rates 195298e09386SEric Dumazet * Alas, some drivers / subsystems require a fair amount 195398e09386SEric Dumazet * of queued bytes to ensure line rate. 195498e09386SEric Dumazet * One example is wifi aggregation (802.11 AMPDU) 195546d3ceabSEric Dumazet */ 195698e09386SEric Dumazet limit = max_t(unsigned int, sysctl_tcp_limit_output_bytes, 195798e09386SEric Dumazet sk->sk_pacing_rate >> 10); 1958c9eeec26SEric Dumazet 1959c9eeec26SEric Dumazet if (atomic_read(&sk->sk_wmem_alloc) > limit) { 196046d3ceabSEric Dumazet set_bit(TSQ_THROTTLED, &tp->tsq_flags); 1961bf06200eSJohn Ogness /* It is possible TX completion already happened 1962bf06200eSJohn Ogness * before we set TSQ_THROTTLED, so we must 1963bf06200eSJohn Ogness * test again the condition. 1964bf06200eSJohn Ogness * We abuse smp_mb__after_clear_bit() because 1965bf06200eSJohn Ogness * there is no smp_mb__after_set_bit() yet 1966bf06200eSJohn Ogness */ 1967bf06200eSJohn Ogness smp_mb__after_clear_bit(); 1968ca8a2263SNeal Cardwell if (atomic_read(&sk->sk_wmem_alloc) > limit) 196946d3ceabSEric Dumazet break; 197046d3ceabSEric Dumazet } 1971c9eeec26SEric Dumazet 1972c8ac3774SHerbert Xu limit = mss_now; 1973f8269a49SIlpo Järvinen if (tso_segs > 1 && !tcp_urg_mode(tp)) 19740e3a4803SIlpo Järvinen limit = tcp_mss_split_point(sk, skb, mss_now, 19751485348dSBen Hutchings min_t(unsigned int, 19761485348dSBen Hutchings cwnd_quota, 1977d4589926SEric Dumazet sk->sk_gso_max_segs), 1978d4589926SEric Dumazet nonagle); 1979c8ac3774SHerbert Xu 1980c8ac3774SHerbert Xu if (skb->len > limit && 1981c4ead4c5SEric Dumazet unlikely(tso_fragment(sk, skb, limit, mss_now, gfp))) 19821da177e4SLinus Torvalds break; 19831da177e4SLinus Torvalds 19841da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 1985c1b4a7e6SDavid S. Miller 1986d5dd9175SIlpo Järvinen if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) 19871da177e4SLinus Torvalds break; 19881da177e4SLinus Torvalds 1989ec342325SAndrew Vagin repair: 19901da177e4SLinus Torvalds /* Advance the send_head. This one is sent out. 19911da177e4SLinus Torvalds * This call will increment packets_out. 19921da177e4SLinus Torvalds */ 199366f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 19941da177e4SLinus Torvalds 19951da177e4SLinus Torvalds tcp_minshall_update(tp, mss_now, skb); 1996a262f0cdSNandita Dukkipati sent_pkts += tcp_skb_pcount(skb); 1997d5dd9175SIlpo Järvinen 1998d5dd9175SIlpo Järvinen if (push_one) 1999d5dd9175SIlpo Järvinen break; 20001da177e4SLinus Torvalds } 20011da177e4SLinus Torvalds 2002aa93466bSDavid S. Miller if (likely(sent_pkts)) { 2003684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 2004684bad11SYuchung Cheng tp->prr_out += sent_pkts; 20056ba8a3b1SNandita Dukkipati 20066ba8a3b1SNandita Dukkipati /* Send one loss probe per tail loss episode. */ 20076ba8a3b1SNandita Dukkipati if (push_one != 2) 20086ba8a3b1SNandita Dukkipati tcp_schedule_loss_probe(sk); 2009ca8a2263SNeal Cardwell tcp_cwnd_validate(sk, is_cwnd_limited); 2010a2a385d6SEric Dumazet return false; 20111da177e4SLinus Torvalds } 20126ba8a3b1SNandita Dukkipati return (push_one == 2) || (!tp->packets_out && tcp_send_head(sk)); 20136ba8a3b1SNandita Dukkipati } 20146ba8a3b1SNandita Dukkipati 20156ba8a3b1SNandita Dukkipati bool tcp_schedule_loss_probe(struct sock *sk) 20166ba8a3b1SNandita Dukkipati { 20176ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 20186ba8a3b1SNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 20196ba8a3b1SNandita Dukkipati u32 timeout, tlp_time_stamp, rto_time_stamp; 2020740b0f18SEric Dumazet u32 rtt = usecs_to_jiffies(tp->srtt_us >> 3); 20216ba8a3b1SNandita Dukkipati 20226ba8a3b1SNandita Dukkipati if (WARN_ON(icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS)) 20236ba8a3b1SNandita Dukkipati return false; 20246ba8a3b1SNandita Dukkipati /* No consecutive loss probes. */ 20256ba8a3b1SNandita Dukkipati if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) { 20266ba8a3b1SNandita Dukkipati tcp_rearm_rto(sk); 20276ba8a3b1SNandita Dukkipati return false; 20286ba8a3b1SNandita Dukkipati } 20296ba8a3b1SNandita Dukkipati /* Don't do any loss probe on a Fast Open connection before 3WHS 20306ba8a3b1SNandita Dukkipati * finishes. 20316ba8a3b1SNandita Dukkipati */ 20326ba8a3b1SNandita Dukkipati if (sk->sk_state == TCP_SYN_RECV) 20336ba8a3b1SNandita Dukkipati return false; 20346ba8a3b1SNandita Dukkipati 20356ba8a3b1SNandita Dukkipati /* TLP is only scheduled when next timer event is RTO. */ 20366ba8a3b1SNandita Dukkipati if (icsk->icsk_pending != ICSK_TIME_RETRANS) 20376ba8a3b1SNandita Dukkipati return false; 20386ba8a3b1SNandita Dukkipati 20396ba8a3b1SNandita Dukkipati /* Schedule a loss probe in 2*RTT for SACK capable connections 20406ba8a3b1SNandita Dukkipati * in Open state, that are either limited by cwnd or application. 20416ba8a3b1SNandita Dukkipati */ 2042740b0f18SEric Dumazet if (sysctl_tcp_early_retrans < 3 || !tp->srtt_us || !tp->packets_out || 20436ba8a3b1SNandita Dukkipati !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open) 20446ba8a3b1SNandita Dukkipati return false; 20456ba8a3b1SNandita Dukkipati 20466ba8a3b1SNandita Dukkipati if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) && 20476ba8a3b1SNandita Dukkipati tcp_send_head(sk)) 20486ba8a3b1SNandita Dukkipati return false; 20496ba8a3b1SNandita Dukkipati 20506ba8a3b1SNandita Dukkipati /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account 20516ba8a3b1SNandita Dukkipati * for delayed ack when there's one outstanding packet. 20526ba8a3b1SNandita Dukkipati */ 20536ba8a3b1SNandita Dukkipati timeout = rtt << 1; 20546ba8a3b1SNandita Dukkipati if (tp->packets_out == 1) 20556ba8a3b1SNandita Dukkipati timeout = max_t(u32, timeout, 20566ba8a3b1SNandita Dukkipati (rtt + (rtt >> 1) + TCP_DELACK_MAX)); 20576ba8a3b1SNandita Dukkipati timeout = max_t(u32, timeout, msecs_to_jiffies(10)); 20586ba8a3b1SNandita Dukkipati 20596ba8a3b1SNandita Dukkipati /* If RTO is shorter, just schedule TLP in its place. */ 20606ba8a3b1SNandita Dukkipati tlp_time_stamp = tcp_time_stamp + timeout; 20616ba8a3b1SNandita Dukkipati rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout; 20626ba8a3b1SNandita Dukkipati if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) { 20636ba8a3b1SNandita Dukkipati s32 delta = rto_time_stamp - tcp_time_stamp; 20646ba8a3b1SNandita Dukkipati if (delta > 0) 20656ba8a3b1SNandita Dukkipati timeout = delta; 20666ba8a3b1SNandita Dukkipati } 20676ba8a3b1SNandita Dukkipati 20686ba8a3b1SNandita Dukkipati inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, 20696ba8a3b1SNandita Dukkipati TCP_RTO_MAX); 20706ba8a3b1SNandita Dukkipati return true; 20716ba8a3b1SNandita Dukkipati } 20726ba8a3b1SNandita Dukkipati 20731f3279aeSEric Dumazet /* Thanks to skb fast clones, we can detect if a prior transmit of 20741f3279aeSEric Dumazet * a packet is still in a qdisc or driver queue. 20751f3279aeSEric Dumazet * In this case, there is very little point doing a retransmit ! 20761f3279aeSEric Dumazet * Note: This is called from BH context only. 20771f3279aeSEric Dumazet */ 20781f3279aeSEric Dumazet static bool skb_still_in_host_queue(const struct sock *sk, 20791f3279aeSEric Dumazet const struct sk_buff *skb) 20801f3279aeSEric Dumazet { 20811f3279aeSEric Dumazet const struct sk_buff *fclone = skb + 1; 20821f3279aeSEric Dumazet 20831f3279aeSEric Dumazet if (unlikely(skb->fclone == SKB_FCLONE_ORIG && 20841f3279aeSEric Dumazet fclone->fclone == SKB_FCLONE_CLONE)) { 20851f3279aeSEric Dumazet NET_INC_STATS_BH(sock_net(sk), 20861f3279aeSEric Dumazet LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES); 20871f3279aeSEric Dumazet return true; 20881f3279aeSEric Dumazet } 20891f3279aeSEric Dumazet return false; 20901f3279aeSEric Dumazet } 20911f3279aeSEric Dumazet 20926ba8a3b1SNandita Dukkipati /* When probe timeout (PTO) fires, send a new segment if one exists, else 20936ba8a3b1SNandita Dukkipati * retransmit the last segment. 20946ba8a3b1SNandita Dukkipati */ 20956ba8a3b1SNandita Dukkipati void tcp_send_loss_probe(struct sock *sk) 20966ba8a3b1SNandita Dukkipati { 20979b717a8dSNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 20986ba8a3b1SNandita Dukkipati struct sk_buff *skb; 20996ba8a3b1SNandita Dukkipati int pcount; 21006ba8a3b1SNandita Dukkipati int mss = tcp_current_mss(sk); 21016ba8a3b1SNandita Dukkipati int err = -1; 21026ba8a3b1SNandita Dukkipati 21036ba8a3b1SNandita Dukkipati if (tcp_send_head(sk) != NULL) { 21046ba8a3b1SNandita Dukkipati err = tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC); 21056ba8a3b1SNandita Dukkipati goto rearm_timer; 21066ba8a3b1SNandita Dukkipati } 21076ba8a3b1SNandita Dukkipati 21089b717a8dSNandita Dukkipati /* At most one outstanding TLP retransmission. */ 21099b717a8dSNandita Dukkipati if (tp->tlp_high_seq) 21109b717a8dSNandita Dukkipati goto rearm_timer; 21119b717a8dSNandita Dukkipati 21126ba8a3b1SNandita Dukkipati /* Retransmit last segment. */ 21136ba8a3b1SNandita Dukkipati skb = tcp_write_queue_tail(sk); 21146ba8a3b1SNandita Dukkipati if (WARN_ON(!skb)) 21156ba8a3b1SNandita Dukkipati goto rearm_timer; 21166ba8a3b1SNandita Dukkipati 21171f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 21181f3279aeSEric Dumazet goto rearm_timer; 21191f3279aeSEric Dumazet 21206ba8a3b1SNandita Dukkipati pcount = tcp_skb_pcount(skb); 21216ba8a3b1SNandita Dukkipati if (WARN_ON(!pcount)) 21226ba8a3b1SNandita Dukkipati goto rearm_timer; 21236ba8a3b1SNandita Dukkipati 21246ba8a3b1SNandita Dukkipati if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) { 2125*6cc55e09SOctavian Purdila if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss, 2126*6cc55e09SOctavian Purdila GFP_ATOMIC))) 21276ba8a3b1SNandita Dukkipati goto rearm_timer; 21286ba8a3b1SNandita Dukkipati skb = tcp_write_queue_tail(sk); 21296ba8a3b1SNandita Dukkipati } 21306ba8a3b1SNandita Dukkipati 21316ba8a3b1SNandita Dukkipati if (WARN_ON(!skb || !tcp_skb_pcount(skb))) 21326ba8a3b1SNandita Dukkipati goto rearm_timer; 21336ba8a3b1SNandita Dukkipati 21346ba8a3b1SNandita Dukkipati /* Probe with zero data doesn't trigger fast recovery. */ 21356ba8a3b1SNandita Dukkipati if (skb->len > 0) 21366ba8a3b1SNandita Dukkipati err = __tcp_retransmit_skb(sk, skb); 21376ba8a3b1SNandita Dukkipati 21389b717a8dSNandita Dukkipati /* Record snd_nxt for loss detection. */ 21399b717a8dSNandita Dukkipati if (likely(!err)) 21409b717a8dSNandita Dukkipati tp->tlp_high_seq = tp->snd_nxt; 21419b717a8dSNandita Dukkipati 21426ba8a3b1SNandita Dukkipati rearm_timer: 21436ba8a3b1SNandita Dukkipati inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 21446ba8a3b1SNandita Dukkipati inet_csk(sk)->icsk_rto, 21456ba8a3b1SNandita Dukkipati TCP_RTO_MAX); 21466ba8a3b1SNandita Dukkipati 21476ba8a3b1SNandita Dukkipati if (likely(!err)) 21486ba8a3b1SNandita Dukkipati NET_INC_STATS_BH(sock_net(sk), 21496ba8a3b1SNandita Dukkipati LINUX_MIB_TCPLOSSPROBES); 21501da177e4SLinus Torvalds } 21511da177e4SLinus Torvalds 2152a762a980SDavid S. Miller /* Push out any pending frames which were held back due to 2153a762a980SDavid S. Miller * TCP_CORK or attempt at coalescing tiny packets. 2154a762a980SDavid S. Miller * The socket must be locked by the caller. 2155a762a980SDavid S. Miller */ 21569e412ba7SIlpo Järvinen void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 21579e412ba7SIlpo Järvinen int nonagle) 2158a762a980SDavid S. Miller { 2159726e07a8SIlpo Järvinen /* If we are closed, the bytes will have to remain here. 2160726e07a8SIlpo Järvinen * In time closedown will finish, we empty the write queue and 2161726e07a8SIlpo Järvinen * all will be happy. 2162726e07a8SIlpo Järvinen */ 2163726e07a8SIlpo Järvinen if (unlikely(sk->sk_state == TCP_CLOSE)) 2164726e07a8SIlpo Järvinen return; 2165726e07a8SIlpo Järvinen 216699a1dec7SMel Gorman if (tcp_write_xmit(sk, cur_mss, nonagle, 0, 216799a1dec7SMel Gorman sk_gfp_atomic(sk, GFP_ATOMIC))) 21689e412ba7SIlpo Järvinen tcp_check_probe_timer(sk); 2169a762a980SDavid S. Miller } 2170a762a980SDavid S. Miller 2171c1b4a7e6SDavid S. Miller /* Send _single_ skb sitting at the send head. This function requires 2172c1b4a7e6SDavid S. Miller * true push pending frames to setup probe timer etc. 2173c1b4a7e6SDavid S. Miller */ 2174c1b4a7e6SDavid S. Miller void tcp_push_one(struct sock *sk, unsigned int mss_now) 2175c1b4a7e6SDavid S. Miller { 2176fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 2177c1b4a7e6SDavid S. Miller 2178c1b4a7e6SDavid S. Miller BUG_ON(!skb || skb->len < mss_now); 2179c1b4a7e6SDavid S. Miller 2180d5dd9175SIlpo Järvinen tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); 2181c1b4a7e6SDavid S. Miller } 2182c1b4a7e6SDavid S. Miller 21831da177e4SLinus Torvalds /* This function returns the amount that we can raise the 21841da177e4SLinus Torvalds * usable window based on the following constraints 21851da177e4SLinus Torvalds * 21861da177e4SLinus Torvalds * 1. The window can never be shrunk once it is offered (RFC 793) 21871da177e4SLinus Torvalds * 2. We limit memory per socket 21881da177e4SLinus Torvalds * 21891da177e4SLinus Torvalds * RFC 1122: 21901da177e4SLinus Torvalds * "the suggested [SWS] avoidance algorithm for the receiver is to keep 21911da177e4SLinus Torvalds * RECV.NEXT + RCV.WIN fixed until: 21921da177e4SLinus Torvalds * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 21931da177e4SLinus Torvalds * 21941da177e4SLinus Torvalds * i.e. don't raise the right edge of the window until you can raise 21951da177e4SLinus Torvalds * it at least MSS bytes. 21961da177e4SLinus Torvalds * 21971da177e4SLinus Torvalds * Unfortunately, the recommended algorithm breaks header prediction, 21981da177e4SLinus Torvalds * since header prediction assumes th->window stays fixed. 21991da177e4SLinus Torvalds * 22001da177e4SLinus Torvalds * Strictly speaking, keeping th->window fixed violates the receiver 22011da177e4SLinus Torvalds * side SWS prevention criteria. The problem is that under this rule 22021da177e4SLinus Torvalds * a stream of single byte packets will cause the right side of the 22031da177e4SLinus Torvalds * window to always advance by a single byte. 22041da177e4SLinus Torvalds * 22051da177e4SLinus Torvalds * Of course, if the sender implements sender side SWS prevention 22061da177e4SLinus Torvalds * then this will not be a problem. 22071da177e4SLinus Torvalds * 22081da177e4SLinus Torvalds * BSD seems to make the following compromise: 22091da177e4SLinus Torvalds * 22101da177e4SLinus Torvalds * If the free space is less than the 1/4 of the maximum 22111da177e4SLinus Torvalds * space available and the free space is less than 1/2 mss, 22121da177e4SLinus Torvalds * then set the window to 0. 22131da177e4SLinus Torvalds * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 22141da177e4SLinus Torvalds * Otherwise, just prevent the window from shrinking 22151da177e4SLinus Torvalds * and from being larger than the largest representable value. 22161da177e4SLinus Torvalds * 22171da177e4SLinus Torvalds * This prevents incremental opening of the window in the regime 22181da177e4SLinus Torvalds * where TCP is limited by the speed of the reader side taking 22191da177e4SLinus Torvalds * data out of the TCP receive queue. It does nothing about 22201da177e4SLinus Torvalds * those cases where the window is constrained on the sender side 22211da177e4SLinus Torvalds * because the pipeline is full. 22221da177e4SLinus Torvalds * 22231da177e4SLinus Torvalds * BSD also seems to "accidentally" limit itself to windows that are a 22241da177e4SLinus Torvalds * multiple of MSS, at least until the free space gets quite small. 22251da177e4SLinus Torvalds * This would appear to be a side effect of the mbuf implementation. 22261da177e4SLinus Torvalds * Combining these two algorithms results in the observed behavior 22271da177e4SLinus Torvalds * of having a fixed window size at almost all times. 22281da177e4SLinus Torvalds * 22291da177e4SLinus Torvalds * Below we obtain similar behavior by forcing the offered window to 22301da177e4SLinus Torvalds * a multiple of the mss when it is feasible to do so. 22311da177e4SLinus Torvalds * 22321da177e4SLinus Torvalds * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 22331da177e4SLinus Torvalds * Regular options like TIMESTAMP are taken into account. 22341da177e4SLinus Torvalds */ 22351da177e4SLinus Torvalds u32 __tcp_select_window(struct sock *sk) 22361da177e4SLinus Torvalds { 2237463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 22381da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2239caa20d9aSStephen Hemminger /* MSS for the peer's data. Previous versions used mss_clamp 22401da177e4SLinus Torvalds * here. I don't know if the value based on our guesses 22411da177e4SLinus Torvalds * of peer's MSS is better for the performance. It's more correct 22421da177e4SLinus Torvalds * but may be worse for the performance because of rcv_mss 22431da177e4SLinus Torvalds * fluctuations. --SAW 1998/11/1 22441da177e4SLinus Torvalds */ 2245463c84b9SArnaldo Carvalho de Melo int mss = icsk->icsk_ack.rcv_mss; 22461da177e4SLinus Torvalds int free_space = tcp_space(sk); 224786c1a045SFlorian Westphal int allowed_space = tcp_full_space(sk); 224886c1a045SFlorian Westphal int full_space = min_t(int, tp->window_clamp, allowed_space); 22491da177e4SLinus Torvalds int window; 22501da177e4SLinus Torvalds 22511da177e4SLinus Torvalds if (mss > full_space) 22521da177e4SLinus Torvalds mss = full_space; 22531da177e4SLinus Torvalds 2254b92edbe0SEric Dumazet if (free_space < (full_space >> 1)) { 2255463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.quick = 0; 22561da177e4SLinus Torvalds 2257180d8cd9SGlauber Costa if (sk_under_memory_pressure(sk)) 2258056834d9SIlpo Järvinen tp->rcv_ssthresh = min(tp->rcv_ssthresh, 2259056834d9SIlpo Järvinen 4U * tp->advmss); 22601da177e4SLinus Torvalds 226186c1a045SFlorian Westphal /* free_space might become our new window, make sure we don't 226286c1a045SFlorian Westphal * increase it due to wscale. 226386c1a045SFlorian Westphal */ 226486c1a045SFlorian Westphal free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale); 226586c1a045SFlorian Westphal 226686c1a045SFlorian Westphal /* if free space is less than mss estimate, or is below 1/16th 226786c1a045SFlorian Westphal * of the maximum allowed, try to move to zero-window, else 226886c1a045SFlorian Westphal * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and 226986c1a045SFlorian Westphal * new incoming data is dropped due to memory limits. 227086c1a045SFlorian Westphal * With large window, mss test triggers way too late in order 227186c1a045SFlorian Westphal * to announce zero window in time before rmem limit kicks in. 227286c1a045SFlorian Westphal */ 227386c1a045SFlorian Westphal if (free_space < (allowed_space >> 4) || free_space < mss) 22741da177e4SLinus Torvalds return 0; 22751da177e4SLinus Torvalds } 22761da177e4SLinus Torvalds 22771da177e4SLinus Torvalds if (free_space > tp->rcv_ssthresh) 22781da177e4SLinus Torvalds free_space = tp->rcv_ssthresh; 22791da177e4SLinus Torvalds 22801da177e4SLinus Torvalds /* Don't do rounding if we are using window scaling, since the 22811da177e4SLinus Torvalds * scaled window will not line up with the MSS boundary anyway. 22821da177e4SLinus Torvalds */ 22831da177e4SLinus Torvalds window = tp->rcv_wnd; 22841da177e4SLinus Torvalds if (tp->rx_opt.rcv_wscale) { 22851da177e4SLinus Torvalds window = free_space; 22861da177e4SLinus Torvalds 22871da177e4SLinus Torvalds /* Advertise enough space so that it won't get scaled away. 22881da177e4SLinus Torvalds * Import case: prevent zero window announcement if 22891da177e4SLinus Torvalds * 1<<rcv_wscale > mss. 22901da177e4SLinus Torvalds */ 22911da177e4SLinus Torvalds if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window) 22921da177e4SLinus Torvalds window = (((window >> tp->rx_opt.rcv_wscale) + 1) 22931da177e4SLinus Torvalds << tp->rx_opt.rcv_wscale); 22941da177e4SLinus Torvalds } else { 22951da177e4SLinus Torvalds /* Get the largest window that is a nice multiple of mss. 22961da177e4SLinus Torvalds * Window clamp already applied above. 22971da177e4SLinus Torvalds * If our current window offering is within 1 mss of the 22981da177e4SLinus Torvalds * free space we just keep it. This prevents the divide 22991da177e4SLinus Torvalds * and multiply from happening most of the time. 23001da177e4SLinus Torvalds * We also don't do any window rounding when the free space 23011da177e4SLinus Torvalds * is too small. 23021da177e4SLinus Torvalds */ 23031da177e4SLinus Torvalds if (window <= free_space - mss || window > free_space) 23041da177e4SLinus Torvalds window = (free_space / mss) * mss; 230584565070SJohn Heffner else if (mss == full_space && 2306b92edbe0SEric Dumazet free_space > window + (full_space >> 1)) 230784565070SJohn Heffner window = free_space; 23081da177e4SLinus Torvalds } 23091da177e4SLinus Torvalds 23101da177e4SLinus Torvalds return window; 23111da177e4SLinus Torvalds } 23121da177e4SLinus Torvalds 23134a17fc3aSIlpo Järvinen /* Collapses two adjacent SKB's during retransmission. */ 23144a17fc3aSIlpo Järvinen static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) 23151da177e4SLinus Torvalds { 23161da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2317fe067e8aSDavid S. Miller struct sk_buff *next_skb = tcp_write_queue_next(sk, skb); 2318058dc334SIlpo Järvinen int skb_size, next_skb_size; 23191da177e4SLinus Torvalds 2320058dc334SIlpo Järvinen skb_size = skb->len; 2321058dc334SIlpo Järvinen next_skb_size = next_skb->len; 23221da177e4SLinus Torvalds 2323058dc334SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); 23241da177e4SLinus Torvalds 23256859d494SIlpo Järvinen tcp_highest_sack_combine(sk, next_skb, skb); 2326a6963a6bSIlpo Järvinen 2327fe067e8aSDavid S. Miller tcp_unlink_write_queue(next_skb, sk); 23281da177e4SLinus Torvalds 2329058dc334SIlpo Järvinen skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size), 23301a4e2d09SArnaldo Carvalho de Melo next_skb_size); 23311da177e4SLinus Torvalds 233252d570aaSJarek Poplawski if (next_skb->ip_summed == CHECKSUM_PARTIAL) 233352d570aaSJarek Poplawski skb->ip_summed = CHECKSUM_PARTIAL; 23341da177e4SLinus Torvalds 233584fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 23361da177e4SLinus Torvalds skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); 23371da177e4SLinus Torvalds 23381da177e4SLinus Torvalds /* Update sequence range on original skb. */ 23391da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 23401da177e4SLinus Torvalds 2341e6c7d085SIlpo Järvinen /* Merge over control information. This moves PSH/FIN etc. over */ 23424de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; 23431da177e4SLinus Torvalds 23441da177e4SLinus Torvalds /* All done, get rid of second SKB and account for it so 23451da177e4SLinus Torvalds * packet counting does not break. 23461da177e4SLinus Torvalds */ 23474828e7f4SIlpo Järvinen TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; 2348b7689205SIlpo Järvinen 2349b7689205SIlpo Järvinen /* changed transmit queue under us so clear hints */ 2350ef9da47cSIlpo Järvinen tcp_clear_retrans_hints_partial(tp); 2351ef9da47cSIlpo Järvinen if (next_skb == tp->retransmit_skb_hint) 2352ef9da47cSIlpo Järvinen tp->retransmit_skb_hint = skb; 2353b7689205SIlpo Järvinen 2354797108d1SIlpo Järvinen tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); 2355797108d1SIlpo Järvinen 23563ab224beSHideo Aoki sk_wmem_free_skb(sk, next_skb); 23571da177e4SLinus Torvalds } 23581da177e4SLinus Torvalds 235967edfef7SAndi Kleen /* Check if coalescing SKBs is legal. */ 2360a2a385d6SEric Dumazet static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) 23614a17fc3aSIlpo Järvinen { 23624a17fc3aSIlpo Järvinen if (tcp_skb_pcount(skb) > 1) 2363a2a385d6SEric Dumazet return false; 23644a17fc3aSIlpo Järvinen /* TODO: SACK collapsing could be used to remove this condition */ 23654a17fc3aSIlpo Järvinen if (skb_shinfo(skb)->nr_frags != 0) 2366a2a385d6SEric Dumazet return false; 23674a17fc3aSIlpo Järvinen if (skb_cloned(skb)) 2368a2a385d6SEric Dumazet return false; 23694a17fc3aSIlpo Järvinen if (skb == tcp_send_head(sk)) 2370a2a385d6SEric Dumazet return false; 23714a17fc3aSIlpo Järvinen /* Some heurestics for collapsing over SACK'd could be invented */ 23724a17fc3aSIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 2373a2a385d6SEric Dumazet return false; 23744a17fc3aSIlpo Järvinen 2375a2a385d6SEric Dumazet return true; 23764a17fc3aSIlpo Järvinen } 23774a17fc3aSIlpo Järvinen 237867edfef7SAndi Kleen /* Collapse packets in the retransmit queue to make to create 237967edfef7SAndi Kleen * less packets on the wire. This is only done on retransmission. 238067edfef7SAndi Kleen */ 23814a17fc3aSIlpo Järvinen static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, 23824a17fc3aSIlpo Järvinen int space) 23834a17fc3aSIlpo Järvinen { 23844a17fc3aSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 23854a17fc3aSIlpo Järvinen struct sk_buff *skb = to, *tmp; 2386a2a385d6SEric Dumazet bool first = true; 23874a17fc3aSIlpo Järvinen 23884a17fc3aSIlpo Järvinen if (!sysctl_tcp_retrans_collapse) 23894a17fc3aSIlpo Järvinen return; 23904de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 23914a17fc3aSIlpo Järvinen return; 23924a17fc3aSIlpo Järvinen 23934a17fc3aSIlpo Järvinen tcp_for_write_queue_from_safe(skb, tmp, sk) { 23944a17fc3aSIlpo Järvinen if (!tcp_can_collapse(sk, skb)) 23954a17fc3aSIlpo Järvinen break; 23964a17fc3aSIlpo Järvinen 23974a17fc3aSIlpo Järvinen space -= skb->len; 23984a17fc3aSIlpo Järvinen 23994a17fc3aSIlpo Järvinen if (first) { 2400a2a385d6SEric Dumazet first = false; 24014a17fc3aSIlpo Järvinen continue; 24024a17fc3aSIlpo Järvinen } 24034a17fc3aSIlpo Järvinen 24044a17fc3aSIlpo Järvinen if (space < 0) 24054a17fc3aSIlpo Järvinen break; 24064a17fc3aSIlpo Järvinen /* Punt if not enough space exists in the first SKB for 24074a17fc3aSIlpo Järvinen * the data in the second 24084a17fc3aSIlpo Järvinen */ 2409a21d4572SEric Dumazet if (skb->len > skb_availroom(to)) 24104a17fc3aSIlpo Järvinen break; 24114a17fc3aSIlpo Järvinen 24124a17fc3aSIlpo Järvinen if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) 24134a17fc3aSIlpo Järvinen break; 24144a17fc3aSIlpo Järvinen 24154a17fc3aSIlpo Järvinen tcp_collapse_retrans(sk, to); 24164a17fc3aSIlpo Järvinen } 24174a17fc3aSIlpo Järvinen } 24184a17fc3aSIlpo Järvinen 24191da177e4SLinus Torvalds /* This retransmits one SKB. Policy decisions and retransmit queue 24201da177e4SLinus Torvalds * state updates are done by the caller. Returns non-zero if an 24211da177e4SLinus Torvalds * error occurred which prevented the send. 24221da177e4SLinus Torvalds */ 242393b174adSYuchung Cheng int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) 24241da177e4SLinus Torvalds { 24251da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 24265d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 24277d227cd2SSridhar Samudrala unsigned int cur_mss; 2428c84a5711SYuchung Cheng int err; 24291da177e4SLinus Torvalds 24305d424d5aSJohn Heffner /* Inconslusive MTU probe */ 24315d424d5aSJohn Heffner if (icsk->icsk_mtup.probe_size) { 24325d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 24335d424d5aSJohn Heffner } 24345d424d5aSJohn Heffner 24351da177e4SLinus Torvalds /* Do not sent more than we queued. 1/4 is reserved for possible 2436caa20d9aSStephen Hemminger * copying overhead: fragmentation, tunneling, mangling etc. 24371da177e4SLinus Torvalds */ 24381da177e4SLinus Torvalds if (atomic_read(&sk->sk_wmem_alloc) > 24391da177e4SLinus Torvalds min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf)) 24401da177e4SLinus Torvalds return -EAGAIN; 24411da177e4SLinus Torvalds 24421f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 24431f3279aeSEric Dumazet return -EBUSY; 24441f3279aeSEric Dumazet 24451da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 24461da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 24471da177e4SLinus Torvalds BUG(); 24481da177e4SLinus Torvalds if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 24491da177e4SLinus Torvalds return -ENOMEM; 24501da177e4SLinus Torvalds } 24511da177e4SLinus Torvalds 24527d227cd2SSridhar Samudrala if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 24537d227cd2SSridhar Samudrala return -EHOSTUNREACH; /* Routing failure or similar. */ 24547d227cd2SSridhar Samudrala 24550c54b85fSIlpo Järvinen cur_mss = tcp_current_mss(sk); 24567d227cd2SSridhar Samudrala 24571da177e4SLinus Torvalds /* If receiver has shrunk his window, and skb is out of 24581da177e4SLinus Torvalds * new window, do not retransmit it. The exception is the 24591da177e4SLinus Torvalds * case, when window is shrunk to zero. In this case 24601da177e4SLinus Torvalds * our retransmit serves as a zero window probe. 24611da177e4SLinus Torvalds */ 24629d4fb27dSJoe Perches if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && 24639d4fb27dSJoe Perches TCP_SKB_CB(skb)->seq != tp->snd_una) 24641da177e4SLinus Torvalds return -EAGAIN; 24651da177e4SLinus Torvalds 24661da177e4SLinus Torvalds if (skb->len > cur_mss) { 2467*6cc55e09SOctavian Purdila if (tcp_fragment(sk, skb, cur_mss, cur_mss, GFP_ATOMIC)) 24681da177e4SLinus Torvalds return -ENOMEM; /* We'll try again later. */ 246902276f3cSIlpo Järvinen } else { 24709eb9362eSIlpo Järvinen int oldpcount = tcp_skb_pcount(skb); 24719eb9362eSIlpo Järvinen 24729eb9362eSIlpo Järvinen if (unlikely(oldpcount > 1)) { 2473c52e2421SEric Dumazet if (skb_unclone(skb, GFP_ATOMIC)) 2474c52e2421SEric Dumazet return -ENOMEM; 247502276f3cSIlpo Järvinen tcp_init_tso_segs(sk, skb, cur_mss); 24769eb9362eSIlpo Järvinen tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb)); 24779eb9362eSIlpo Järvinen } 24781da177e4SLinus Torvalds } 24791da177e4SLinus Torvalds 24801da177e4SLinus Torvalds tcp_retrans_try_collapse(sk, skb, cur_mss); 24811da177e4SLinus Torvalds 24821da177e4SLinus Torvalds /* Make a copy, if the first transmission SKB clone we made 24831da177e4SLinus Torvalds * is still in somebody's hands, else make a clone. 24841da177e4SLinus Torvalds */ 24851da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 24861da177e4SLinus Torvalds 248750bceae9SThomas Graf /* make sure skb->data is aligned on arches that require it 248850bceae9SThomas Graf * and check if ack-trimming & collapsing extended the headroom 248950bceae9SThomas Graf * beyond what csum_start can cover. 249050bceae9SThomas Graf */ 249150bceae9SThomas Graf if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) || 249250bceae9SThomas Graf skb_headroom(skb) >= 0xFFFF)) { 2493117632e6SEric Dumazet struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER, 2494117632e6SEric Dumazet GFP_ATOMIC); 2495c84a5711SYuchung Cheng err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) : 2496117632e6SEric Dumazet -ENOBUFS; 2497117632e6SEric Dumazet } else { 2498c84a5711SYuchung Cheng err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 2499117632e6SEric Dumazet } 2500c84a5711SYuchung Cheng 2501fc9f3501SEric Dumazet if (likely(!err)) { 2502c84a5711SYuchung Cheng TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS; 2503fc9f3501SEric Dumazet /* Update global TCP statistics. */ 2504fc9f3501SEric Dumazet TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); 2505fc9f3501SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 2506fc9f3501SEric Dumazet NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 2507fc9f3501SEric Dumazet tp->total_retrans++; 2508fc9f3501SEric Dumazet } 2509c84a5711SYuchung Cheng return err; 251093b174adSYuchung Cheng } 251193b174adSYuchung Cheng 251293b174adSYuchung Cheng int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) 251393b174adSYuchung Cheng { 251493b174adSYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 251593b174adSYuchung Cheng int err = __tcp_retransmit_skb(sk, skb); 25161da177e4SLinus Torvalds 25171da177e4SLinus Torvalds if (err == 0) { 25181da177e4SLinus Torvalds #if FASTRETRANS_DEBUG > 0 25191da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 2520e87cc472SJoe Perches net_dbg_ratelimited("retrans_out leaked\n"); 25211da177e4SLinus Torvalds } 25221da177e4SLinus Torvalds #endif 2523b08d6cb2SIlpo Järvinen if (!tp->retrans_out) 2524b08d6cb2SIlpo Järvinen tp->lost_retrans_low = tp->snd_nxt; 25251da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 25261da177e4SLinus Torvalds tp->retrans_out += tcp_skb_pcount(skb); 25271da177e4SLinus Torvalds 25281da177e4SLinus Torvalds /* Save stamp of the first retransmit. */ 25291da177e4SLinus Torvalds if (!tp->retrans_stamp) 25301da177e4SLinus Torvalds tp->retrans_stamp = TCP_SKB_CB(skb)->when; 25311da177e4SLinus Torvalds 2532c24f691bSYuchung Cheng tp->undo_retrans += tcp_skb_pcount(skb); 25331da177e4SLinus Torvalds 25341da177e4SLinus Torvalds /* snd_nxt is stored to detect loss of retransmitted segment, 25351da177e4SLinus Torvalds * see tcp_input.c tcp_sacktag_write_queue(). 25361da177e4SLinus Torvalds */ 25371da177e4SLinus Torvalds TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt; 25381f3279aeSEric Dumazet } else if (err != -EBUSY) { 253924ab6becSYuchung Cheng NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL); 25401da177e4SLinus Torvalds } 25411da177e4SLinus Torvalds return err; 25421da177e4SLinus Torvalds } 25431da177e4SLinus Torvalds 254467edfef7SAndi Kleen /* Check if we forward retransmits are possible in the current 254567edfef7SAndi Kleen * window/congestion state. 254667edfef7SAndi Kleen */ 2547a2a385d6SEric Dumazet static bool tcp_can_forward_retransmit(struct sock *sk) 2548b5afe7bcSIlpo Järvinen { 2549b5afe7bcSIlpo Järvinen const struct inet_connection_sock *icsk = inet_csk(sk); 2550cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 2551b5afe7bcSIlpo Järvinen 2552b5afe7bcSIlpo Järvinen /* Forward retransmissions are possible only during Recovery. */ 2553b5afe7bcSIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Recovery) 2554a2a385d6SEric Dumazet return false; 2555b5afe7bcSIlpo Järvinen 2556b5afe7bcSIlpo Järvinen /* No forward retransmissions in Reno are possible. */ 2557b5afe7bcSIlpo Järvinen if (tcp_is_reno(tp)) 2558a2a385d6SEric Dumazet return false; 2559b5afe7bcSIlpo Järvinen 2560b5afe7bcSIlpo Järvinen /* Yeah, we have to make difficult choice between forward transmission 2561b5afe7bcSIlpo Järvinen * and retransmission... Both ways have their merits... 2562b5afe7bcSIlpo Järvinen * 2563b5afe7bcSIlpo Järvinen * For now we do not retransmit anything, while we have some new 2564b5afe7bcSIlpo Järvinen * segments to send. In the other cases, follow rule 3 for 2565b5afe7bcSIlpo Järvinen * NextSeg() specified in RFC3517. 2566b5afe7bcSIlpo Järvinen */ 2567b5afe7bcSIlpo Järvinen 2568b5afe7bcSIlpo Järvinen if (tcp_may_send_now(sk)) 2569a2a385d6SEric Dumazet return false; 2570b5afe7bcSIlpo Järvinen 2571a2a385d6SEric Dumazet return true; 2572b5afe7bcSIlpo Järvinen } 2573b5afe7bcSIlpo Järvinen 25741da177e4SLinus Torvalds /* This gets called after a retransmit timeout, and the initially 25751da177e4SLinus Torvalds * retransmitted data is acknowledged. It tries to continue 25761da177e4SLinus Torvalds * resending the rest of the retransmit queue, until either 25771da177e4SLinus Torvalds * we've sent it all or the congestion window limit is reached. 25781da177e4SLinus Torvalds * If doing SACK, the first ACK which comes back for a timeout 25791da177e4SLinus Torvalds * based retransmit packet might feed us FACK information again. 25801da177e4SLinus Torvalds * If so, we use it to avoid unnecessarily retransmissions. 25811da177e4SLinus Torvalds */ 25821da177e4SLinus Torvalds void tcp_xmit_retransmit_queue(struct sock *sk) 25831da177e4SLinus Torvalds { 25846687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 25851da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 25861da177e4SLinus Torvalds struct sk_buff *skb; 25870e1c54c2SIlpo Järvinen struct sk_buff *hole = NULL; 2588618d9f25SIlpo Järvinen u32 last_lost; 258961eb55f4SIlpo Järvinen int mib_idx; 25900e1c54c2SIlpo Järvinen int fwd_rexmitting = 0; 25916a438bbeSStephen Hemminger 259245e77d31SIlpo Järvinen if (!tp->packets_out) 259345e77d31SIlpo Järvinen return; 259445e77d31SIlpo Järvinen 259508ebd172SIlpo Järvinen if (!tp->lost_out) 259608ebd172SIlpo Järvinen tp->retransmit_high = tp->snd_una; 259708ebd172SIlpo Järvinen 2598618d9f25SIlpo Järvinen if (tp->retransmit_skb_hint) { 25996a438bbeSStephen Hemminger skb = tp->retransmit_skb_hint; 2600618d9f25SIlpo Järvinen last_lost = TCP_SKB_CB(skb)->end_seq; 2601618d9f25SIlpo Järvinen if (after(last_lost, tp->retransmit_high)) 2602618d9f25SIlpo Järvinen last_lost = tp->retransmit_high; 2603618d9f25SIlpo Järvinen } else { 2604fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 2605618d9f25SIlpo Järvinen last_lost = tp->snd_una; 2606618d9f25SIlpo Järvinen } 26071da177e4SLinus Torvalds 2608fe067e8aSDavid S. Miller tcp_for_write_queue_from(skb, sk) { 26091da177e4SLinus Torvalds __u8 sacked = TCP_SKB_CB(skb)->sacked; 26101da177e4SLinus Torvalds 2611fe067e8aSDavid S. Miller if (skb == tcp_send_head(sk)) 2612fe067e8aSDavid S. Miller break; 26136a438bbeSStephen Hemminger /* we could do better than to assign each time */ 26140e1c54c2SIlpo Järvinen if (hole == NULL) 26156a438bbeSStephen Hemminger tp->retransmit_skb_hint = skb; 26166a438bbeSStephen Hemminger 26171da177e4SLinus Torvalds /* Assume this retransmit will generate 26181da177e4SLinus Torvalds * only one packet for congestion window 26191da177e4SLinus Torvalds * calculation purposes. This works because 26201da177e4SLinus Torvalds * tcp_retransmit_skb() will chop up the 26211da177e4SLinus Torvalds * packet to be MSS sized and all the 26221da177e4SLinus Torvalds * packet counting works out. 26231da177e4SLinus Torvalds */ 26241da177e4SLinus Torvalds if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) 26251da177e4SLinus Torvalds return; 26260e1c54c2SIlpo Järvinen 26270e1c54c2SIlpo Järvinen if (fwd_rexmitting) { 26280e1c54c2SIlpo Järvinen begin_fwd: 26290e1c54c2SIlpo Järvinen if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp))) 2630006f582cSIlpo Järvinen break; 26310e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFORWARDRETRANS; 26320e1c54c2SIlpo Järvinen 26330e1c54c2SIlpo Järvinen } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) { 2634618d9f25SIlpo Järvinen tp->retransmit_high = last_lost; 26350e1c54c2SIlpo Järvinen if (!tcp_can_forward_retransmit(sk)) 26360e1c54c2SIlpo Järvinen break; 26370e1c54c2SIlpo Järvinen /* Backtrack if necessary to non-L'ed skb */ 26380e1c54c2SIlpo Järvinen if (hole != NULL) { 26390e1c54c2SIlpo Järvinen skb = hole; 26400e1c54c2SIlpo Järvinen hole = NULL; 26410e1c54c2SIlpo Järvinen } 26420e1c54c2SIlpo Järvinen fwd_rexmitting = 1; 26430e1c54c2SIlpo Järvinen goto begin_fwd; 26440e1c54c2SIlpo Järvinen 26450e1c54c2SIlpo Järvinen } else if (!(sacked & TCPCB_LOST)) { 2646ac11ba75SIlpo Järvinen if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) 26470e1c54c2SIlpo Järvinen hole = skb; 264861eb55f4SIlpo Järvinen continue; 26491da177e4SLinus Torvalds 26500e1c54c2SIlpo Järvinen } else { 2651618d9f25SIlpo Järvinen last_lost = TCP_SKB_CB(skb)->end_seq; 26520e1c54c2SIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Loss) 26530e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFASTRETRANS; 26540e1c54c2SIlpo Järvinen else 26550e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; 26560e1c54c2SIlpo Järvinen } 26570e1c54c2SIlpo Järvinen 26580e1c54c2SIlpo Järvinen if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) 265961eb55f4SIlpo Järvinen continue; 266040b215e5SPavel Emelyanov 266124ab6becSYuchung Cheng if (tcp_retransmit_skb(sk, skb)) 26621da177e4SLinus Torvalds return; 266324ab6becSYuchung Cheng 2664de0744afSPavel Emelyanov NET_INC_STATS_BH(sock_net(sk), mib_idx); 26651da177e4SLinus Torvalds 2666684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 2667a262f0cdSNandita Dukkipati tp->prr_out += tcp_skb_pcount(skb); 2668a262f0cdSNandita Dukkipati 2669fe067e8aSDavid S. Miller if (skb == tcp_write_queue_head(sk)) 2670463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 26713f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, 26723f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 26731da177e4SLinus Torvalds } 26741da177e4SLinus Torvalds } 26751da177e4SLinus Torvalds 26761da177e4SLinus Torvalds /* Send a fin. The caller locks the socket for us. This cannot be 26771da177e4SLinus Torvalds * allowed to fail queueing a FIN frame under any circumstances. 26781da177e4SLinus Torvalds */ 26791da177e4SLinus Torvalds void tcp_send_fin(struct sock *sk) 26801da177e4SLinus Torvalds { 26811da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2682fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_write_queue_tail(sk); 26831da177e4SLinus Torvalds int mss_now; 26841da177e4SLinus Torvalds 26851da177e4SLinus Torvalds /* Optimization, tack on the FIN if we have a queue of 26861da177e4SLinus Torvalds * unsent frames. But be careful about outgoing SACKS 26871da177e4SLinus Torvalds * and IP options. 26881da177e4SLinus Torvalds */ 26890c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 26901da177e4SLinus Torvalds 2691fe067e8aSDavid S. Miller if (tcp_send_head(sk) != NULL) { 26924de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_FIN; 26931da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq++; 26941da177e4SLinus Torvalds tp->write_seq++; 26951da177e4SLinus Torvalds } else { 26961da177e4SLinus Torvalds /* Socket is locked, keep trying until memory is available. */ 26971da177e4SLinus Torvalds for (;;) { 2698aa133076SWu Fengguang skb = alloc_skb_fclone(MAX_TCP_HEADER, 2699aa133076SWu Fengguang sk->sk_allocation); 27001da177e4SLinus Torvalds if (skb) 27011da177e4SLinus Torvalds break; 27021da177e4SLinus Torvalds yield(); 27031da177e4SLinus Torvalds } 27041da177e4SLinus Torvalds 27051da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 27061da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 27071da177e4SLinus Torvalds /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 2708e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tp->write_seq, 2709a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_FIN); 27101da177e4SLinus Torvalds tcp_queue_skb(sk, skb); 27111da177e4SLinus Torvalds } 27129e412ba7SIlpo Järvinen __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF); 27131da177e4SLinus Torvalds } 27141da177e4SLinus Torvalds 27151da177e4SLinus Torvalds /* We get here when a process closes a file descriptor (either due to 27161da177e4SLinus Torvalds * an explicit close() or as a byproduct of exit()'ing) and there 27171da177e4SLinus Torvalds * was unread data in the receive queue. This behavior is recommended 271865bb723cSGerrit Renker * by RFC 2525, section 2.17. -DaveM 27191da177e4SLinus Torvalds */ 2720dd0fc66fSAl Viro void tcp_send_active_reset(struct sock *sk, gfp_t priority) 27211da177e4SLinus Torvalds { 27221da177e4SLinus Torvalds struct sk_buff *skb; 27231da177e4SLinus Torvalds 27241da177e4SLinus Torvalds /* NOTE: No TCP options attached and we never retransmit this. */ 27251da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, priority); 27261da177e4SLinus Torvalds if (!skb) { 27274e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 27281da177e4SLinus Torvalds return; 27291da177e4SLinus Torvalds } 27301da177e4SLinus Torvalds 27311da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 27321da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 2733e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), 2734a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_RST); 27351da177e4SLinus Torvalds /* Send it off. */ 27361da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2737dfb4b9dcSDavid S. Miller if (tcp_transmit_skb(sk, skb, 0, priority)) 27384e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 273926af65cbSSridhar Samudrala 274081cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); 27411da177e4SLinus Torvalds } 27421da177e4SLinus Torvalds 274367edfef7SAndi Kleen /* Send a crossed SYN-ACK during socket establishment. 274467edfef7SAndi Kleen * WARNING: This routine must only be called when we have already sent 27451da177e4SLinus Torvalds * a SYN packet that crossed the incoming SYN that caused this routine 27461da177e4SLinus Torvalds * to get called. If this assumption fails then the initial rcv_wnd 27471da177e4SLinus Torvalds * and rcv_wscale values will not be correct. 27481da177e4SLinus Torvalds */ 27491da177e4SLinus Torvalds int tcp_send_synack(struct sock *sk) 27501da177e4SLinus Torvalds { 27511da177e4SLinus Torvalds struct sk_buff *skb; 27521da177e4SLinus Torvalds 2753fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 27544de075e0SEric Dumazet if (skb == NULL || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 275591df42beSJoe Perches pr_debug("%s: wrong queue state\n", __func__); 27561da177e4SLinus Torvalds return -EFAULT; 27571da177e4SLinus Torvalds } 27584de075e0SEric Dumazet if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { 27591da177e4SLinus Torvalds if (skb_cloned(skb)) { 27601da177e4SLinus Torvalds struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC); 27611da177e4SLinus Torvalds if (nskb == NULL) 27621da177e4SLinus Torvalds return -ENOMEM; 2763fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 27641da177e4SLinus Torvalds skb_header_release(nskb); 2765fe067e8aSDavid S. Miller __tcp_add_write_queue_head(sk, nskb); 27663ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 27673ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 27683ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 27691da177e4SLinus Torvalds skb = nskb; 27701da177e4SLinus Torvalds } 27711da177e4SLinus Torvalds 27724de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; 27731da177e4SLinus Torvalds TCP_ECN_send_synack(tcp_sk(sk), skb); 27741da177e4SLinus Torvalds } 27751da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2776dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 27771da177e4SLinus Torvalds } 27781da177e4SLinus Torvalds 27794aea39c1SEric Dumazet /** 27804aea39c1SEric Dumazet * tcp_make_synack - Prepare a SYN-ACK. 27814aea39c1SEric Dumazet * sk: listener socket 27824aea39c1SEric Dumazet * dst: dst entry attached to the SYNACK 27834aea39c1SEric Dumazet * req: request_sock pointer 27844aea39c1SEric Dumazet * 27854aea39c1SEric Dumazet * Allocate one skb and build a SYNACK packet. 27864aea39c1SEric Dumazet * @dst is consumed : Caller should not use it again. 27874aea39c1SEric Dumazet */ 27881da177e4SLinus Torvalds struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst, 2789e6b4d113SWilliam Allen Simpson struct request_sock *req, 27908336886fSJerry Chu struct tcp_fastopen_cookie *foc) 27911da177e4SLinus Torvalds { 2792bd0388aeSWilliam Allen Simpson struct tcp_out_options opts; 27932e6599cbSArnaldo Carvalho de Melo struct inet_request_sock *ireq = inet_rsk(req); 27941da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 27951da177e4SLinus Torvalds struct tcphdr *th; 27961da177e4SLinus Torvalds struct sk_buff *skb; 2797cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 2798bd0388aeSWilliam Allen Simpson int tcp_header_size; 2799f5fff5dcSTom Quetchenbach int mss; 28001da177e4SLinus Torvalds 2801a0b8486cSEric Dumazet skb = sock_wmalloc(sk, MAX_TCP_HEADER, 1, GFP_ATOMIC); 28024aea39c1SEric Dumazet if (unlikely(!skb)) { 28034aea39c1SEric Dumazet dst_release(dst); 28041da177e4SLinus Torvalds return NULL; 28054aea39c1SEric Dumazet } 28061da177e4SLinus Torvalds /* Reserve space for headers. */ 28071da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 28081da177e4SLinus Torvalds 28094aea39c1SEric Dumazet skb_dst_set(skb, dst); 2810ca10b9e9SEric Dumazet security_skb_owned_by(skb, sk); 28111da177e4SLinus Torvalds 28120dbaee3bSDavid S. Miller mss = dst_metric_advmss(dst); 2813f5fff5dcSTom Quetchenbach if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss) 2814f5fff5dcSTom Quetchenbach mss = tp->rx_opt.user_mss; 2815f5fff5dcSTom Quetchenbach 281633ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 28178b5f12d0SFlorian Westphal #ifdef CONFIG_SYN_COOKIES 28188b5f12d0SFlorian Westphal if (unlikely(req->cookie_ts)) 28198b5f12d0SFlorian Westphal TCP_SKB_CB(skb)->when = cookie_init_timestamp(req); 28208b5f12d0SFlorian Westphal else 28218b5f12d0SFlorian Westphal #endif 282233ad798cSAdam Langley TCP_SKB_CB(skb)->when = tcp_time_stamp; 28231a2c6181SChristoph Paasch tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, &md5, 28241a2c6181SChristoph Paasch foc) + sizeof(*th); 282533ad798cSAdam Langley 2826aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 2827aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 28281da177e4SLinus Torvalds 2829aa8223c7SArnaldo Carvalho de Melo th = tcp_hdr(skb); 28301da177e4SLinus Torvalds memset(th, 0, sizeof(struct tcphdr)); 28311da177e4SLinus Torvalds th->syn = 1; 28321da177e4SLinus Torvalds th->ack = 1; 28331da177e4SLinus Torvalds TCP_ECN_make_synack(req, th); 2834b44084c2SEric Dumazet th->source = htons(ireq->ir_num); 2835634fb979SEric Dumazet th->dest = ireq->ir_rmt_port; 2836e870a8efSIlpo Järvinen /* Setting of flags are superfluous here for callers (and ECE is 2837e870a8efSIlpo Järvinen * not even correctly set) 2838e870a8efSIlpo Järvinen */ 2839e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn, 2840a3433f35SChangli Gao TCPHDR_SYN | TCPHDR_ACK); 28414957faadSWilliam Allen Simpson 28421da177e4SLinus Torvalds th->seq = htonl(TCP_SKB_CB(skb)->seq); 28438336886fSJerry Chu /* XXX data is queued and acked as is. No buffer/window check */ 28448336886fSJerry Chu th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt); 28451da177e4SLinus Torvalds 28461da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 2847600ff0c2SIlpo Järvinen th->window = htons(min(req->rcv_wnd, 65535U)); 2848bd0388aeSWilliam Allen Simpson tcp_options_write((__be32 *)(th + 1), tp, &opts); 28491da177e4SLinus Torvalds th->doff = (tcp_header_size >> 2); 2850a0b8486cSEric Dumazet TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_OUTSEGS); 2851cfb6eeb4SYOSHIFUJI Hideaki 2852cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 2853cfb6eeb4SYOSHIFUJI Hideaki /* Okay, we have all we need - do the md5 hash if needed */ 2854cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 2855bd0388aeSWilliam Allen Simpson tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, 285649a72dfbSAdam Langley md5, NULL, req, skb); 2857cfb6eeb4SYOSHIFUJI Hideaki } 2858cfb6eeb4SYOSHIFUJI Hideaki #endif 2859cfb6eeb4SYOSHIFUJI Hideaki 28601da177e4SLinus Torvalds return skb; 28611da177e4SLinus Torvalds } 28624bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_make_synack); 28631da177e4SLinus Torvalds 286467edfef7SAndi Kleen /* Do all connect socket setups that can be done AF independent. */ 2865f7e56a76Sstephen hemminger static void tcp_connect_init(struct sock *sk) 28661da177e4SLinus Torvalds { 2867cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 28681da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 28691da177e4SLinus Torvalds __u8 rcv_wscale; 28701da177e4SLinus Torvalds 28711da177e4SLinus Torvalds /* We'll fix this up when we get a response from the other end. 28721da177e4SLinus Torvalds * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 28731da177e4SLinus Torvalds */ 28741da177e4SLinus Torvalds tp->tcp_header_len = sizeof(struct tcphdr) + 2875bb5b7c11SDavid S. Miller (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0); 28761da177e4SLinus Torvalds 2877cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 2878cfb6eeb4SYOSHIFUJI Hideaki if (tp->af_specific->md5_lookup(sk, sk) != NULL) 2879cfb6eeb4SYOSHIFUJI Hideaki tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 2880cfb6eeb4SYOSHIFUJI Hideaki #endif 2881cfb6eeb4SYOSHIFUJI Hideaki 28821da177e4SLinus Torvalds /* If user gave his TCP_MAXSEG, record it to clamp */ 28831da177e4SLinus Torvalds if (tp->rx_opt.user_mss) 28841da177e4SLinus Torvalds tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 28851da177e4SLinus Torvalds tp->max_window = 0; 28865d424d5aSJohn Heffner tcp_mtup_init(sk); 28871da177e4SLinus Torvalds tcp_sync_mss(sk, dst_mtu(dst)); 28881da177e4SLinus Torvalds 28891da177e4SLinus Torvalds if (!tp->window_clamp) 28901da177e4SLinus Torvalds tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 28910dbaee3bSDavid S. Miller tp->advmss = dst_metric_advmss(dst); 2892f5fff5dcSTom Quetchenbach if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss) 2893f5fff5dcSTom Quetchenbach tp->advmss = tp->rx_opt.user_mss; 2894f5fff5dcSTom Quetchenbach 28951da177e4SLinus Torvalds tcp_initialize_rcv_mss(sk); 28961da177e4SLinus Torvalds 2897e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 2898e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 2899e88c64f0SHagen Paul Pfeifer (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) 2900e88c64f0SHagen Paul Pfeifer tp->window_clamp = tcp_full_space(sk); 2901e88c64f0SHagen Paul Pfeifer 29021da177e4SLinus Torvalds tcp_select_initial_window(tcp_full_space(sk), 29031da177e4SLinus Torvalds tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 29041da177e4SLinus Torvalds &tp->rcv_wnd, 29051da177e4SLinus Torvalds &tp->window_clamp, 2906bb5b7c11SDavid S. Miller sysctl_tcp_window_scaling, 290731d12926Slaurent chavey &rcv_wscale, 290831d12926Slaurent chavey dst_metric(dst, RTAX_INITRWND)); 29091da177e4SLinus Torvalds 29101da177e4SLinus Torvalds tp->rx_opt.rcv_wscale = rcv_wscale; 29111da177e4SLinus Torvalds tp->rcv_ssthresh = tp->rcv_wnd; 29121da177e4SLinus Torvalds 29131da177e4SLinus Torvalds sk->sk_err = 0; 29141da177e4SLinus Torvalds sock_reset_flag(sk, SOCK_DONE); 29151da177e4SLinus Torvalds tp->snd_wnd = 0; 2916ee7537b6SHantzis Fotis tcp_init_wl(tp, 0); 29171da177e4SLinus Torvalds tp->snd_una = tp->write_seq; 29181da177e4SLinus Torvalds tp->snd_sml = tp->write_seq; 291933f5f57eSIlpo Järvinen tp->snd_up = tp->write_seq; 2920370816aeSPavel Emelyanov tp->snd_nxt = tp->write_seq; 2921ee995283SPavel Emelyanov 2922ee995283SPavel Emelyanov if (likely(!tp->repair)) 29231da177e4SLinus Torvalds tp->rcv_nxt = 0; 2924c7781a6eSAndrew Vagin else 2925c7781a6eSAndrew Vagin tp->rcv_tstamp = tcp_time_stamp; 2926ee995283SPavel Emelyanov tp->rcv_wup = tp->rcv_nxt; 2927ee995283SPavel Emelyanov tp->copied_seq = tp->rcv_nxt; 29281da177e4SLinus Torvalds 2929463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; 2930463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_retransmits = 0; 29311da177e4SLinus Torvalds tcp_clear_retrans(tp); 29321da177e4SLinus Torvalds } 29331da177e4SLinus Torvalds 2934783237e8SYuchung Cheng static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) 2935783237e8SYuchung Cheng { 2936783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 2937783237e8SYuchung Cheng struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 2938783237e8SYuchung Cheng 2939783237e8SYuchung Cheng tcb->end_seq += skb->len; 2940783237e8SYuchung Cheng skb_header_release(skb); 2941783237e8SYuchung Cheng __tcp_add_write_queue_tail(sk, skb); 2942783237e8SYuchung Cheng sk->sk_wmem_queued += skb->truesize; 2943783237e8SYuchung Cheng sk_mem_charge(sk, skb->truesize); 2944783237e8SYuchung Cheng tp->write_seq = tcb->end_seq; 2945783237e8SYuchung Cheng tp->packets_out += tcp_skb_pcount(skb); 2946783237e8SYuchung Cheng } 2947783237e8SYuchung Cheng 2948783237e8SYuchung Cheng /* Build and send a SYN with data and (cached) Fast Open cookie. However, 2949783237e8SYuchung Cheng * queue a data-only packet after the regular SYN, such that regular SYNs 2950783237e8SYuchung Cheng * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges 2951783237e8SYuchung Cheng * only the SYN sequence, the data are retransmitted in the first ACK. 2952783237e8SYuchung Cheng * If cookie is not cached or other error occurs, falls back to send a 2953783237e8SYuchung Cheng * regular SYN with Fast Open cookie request option. 2954783237e8SYuchung Cheng */ 2955783237e8SYuchung Cheng static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) 2956783237e8SYuchung Cheng { 2957783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 2958783237e8SYuchung Cheng struct tcp_fastopen_request *fo = tp->fastopen_req; 2959aab48743SYuchung Cheng int syn_loss = 0, space, i, err = 0, iovlen = fo->data->msg_iovlen; 2960783237e8SYuchung Cheng struct sk_buff *syn_data = NULL, *data; 2961aab48743SYuchung Cheng unsigned long last_syn_loss = 0; 2962783237e8SYuchung Cheng 296367da22d2SYuchung Cheng tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ 2964aab48743SYuchung Cheng tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie, 2965aab48743SYuchung Cheng &syn_loss, &last_syn_loss); 2966aab48743SYuchung Cheng /* Recurring FO SYN losses: revert to regular handshake temporarily */ 2967aab48743SYuchung Cheng if (syn_loss > 1 && 2968aab48743SYuchung Cheng time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) { 2969aab48743SYuchung Cheng fo->cookie.len = -1; 2970aab48743SYuchung Cheng goto fallback; 2971aab48743SYuchung Cheng } 2972aab48743SYuchung Cheng 297367da22d2SYuchung Cheng if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE) 297467da22d2SYuchung Cheng fo->cookie.len = -1; 297567da22d2SYuchung Cheng else if (fo->cookie.len <= 0) 2976783237e8SYuchung Cheng goto fallback; 2977783237e8SYuchung Cheng 2978783237e8SYuchung Cheng /* MSS for SYN-data is based on cached MSS and bounded by PMTU and 2979783237e8SYuchung Cheng * user-MSS. Reserve maximum option space for middleboxes that add 2980783237e8SYuchung Cheng * private TCP options. The cost is reduced data space in SYN :( 2981783237e8SYuchung Cheng */ 2982783237e8SYuchung Cheng if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp) 2983783237e8SYuchung Cheng tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 29841b63edd6SYuchung Cheng space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) - 2985783237e8SYuchung Cheng MAX_TCP_OPTION_SPACE; 2986783237e8SYuchung Cheng 2987f5ddcbbbSEric Dumazet space = min_t(size_t, space, fo->size); 2988f5ddcbbbSEric Dumazet 2989f5ddcbbbSEric Dumazet /* limit to order-0 allocations */ 2990f5ddcbbbSEric Dumazet space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER)); 2991f5ddcbbbSEric Dumazet 2992f5ddcbbbSEric Dumazet syn_data = skb_copy_expand(syn, MAX_TCP_HEADER, space, 2993783237e8SYuchung Cheng sk->sk_allocation); 2994783237e8SYuchung Cheng if (syn_data == NULL) 2995783237e8SYuchung Cheng goto fallback; 2996783237e8SYuchung Cheng 2997783237e8SYuchung Cheng for (i = 0; i < iovlen && syn_data->len < space; ++i) { 2998783237e8SYuchung Cheng struct iovec *iov = &fo->data->msg_iov[i]; 2999783237e8SYuchung Cheng unsigned char __user *from = iov->iov_base; 3000783237e8SYuchung Cheng int len = iov->iov_len; 3001783237e8SYuchung Cheng 3002783237e8SYuchung Cheng if (syn_data->len + len > space) 3003783237e8SYuchung Cheng len = space - syn_data->len; 3004783237e8SYuchung Cheng else if (i + 1 == iovlen) 3005783237e8SYuchung Cheng /* No more data pending in inet_wait_for_connect() */ 3006783237e8SYuchung Cheng fo->data = NULL; 3007783237e8SYuchung Cheng 3008783237e8SYuchung Cheng if (skb_add_data(syn_data, from, len)) 3009783237e8SYuchung Cheng goto fallback; 3010783237e8SYuchung Cheng } 3011783237e8SYuchung Cheng 3012783237e8SYuchung Cheng /* Queue a data-only packet after the regular SYN for retransmission */ 3013783237e8SYuchung Cheng data = pskb_copy(syn_data, sk->sk_allocation); 3014783237e8SYuchung Cheng if (data == NULL) 3015783237e8SYuchung Cheng goto fallback; 3016783237e8SYuchung Cheng TCP_SKB_CB(data)->seq++; 3017783237e8SYuchung Cheng TCP_SKB_CB(data)->tcp_flags &= ~TCPHDR_SYN; 3018783237e8SYuchung Cheng TCP_SKB_CB(data)->tcp_flags = (TCPHDR_ACK|TCPHDR_PSH); 3019783237e8SYuchung Cheng tcp_connect_queue_skb(sk, data); 3020783237e8SYuchung Cheng fo->copied = data->len; 3021783237e8SYuchung Cheng 3022431a9124SEric Dumazet /* syn_data is about to be sent, we need to take current time stamps 3023431a9124SEric Dumazet * for the packets that are in write queue : SYN packet and DATA 3024431a9124SEric Dumazet */ 3025431a9124SEric Dumazet skb_mstamp_get(&syn->skb_mstamp); 3026431a9124SEric Dumazet data->skb_mstamp = syn->skb_mstamp; 3027431a9124SEric Dumazet 3028783237e8SYuchung Cheng if (tcp_transmit_skb(sk, syn_data, 0, sk->sk_allocation) == 0) { 302967da22d2SYuchung Cheng tp->syn_data = (fo->copied > 0); 3030f19c29e3SYuchung Cheng NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT); 3031783237e8SYuchung Cheng goto done; 3032783237e8SYuchung Cheng } 3033783237e8SYuchung Cheng syn_data = NULL; 3034783237e8SYuchung Cheng 3035783237e8SYuchung Cheng fallback: 3036783237e8SYuchung Cheng /* Send a regular SYN with Fast Open cookie request option */ 3037783237e8SYuchung Cheng if (fo->cookie.len > 0) 3038783237e8SYuchung Cheng fo->cookie.len = 0; 3039783237e8SYuchung Cheng err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); 3040783237e8SYuchung Cheng if (err) 3041783237e8SYuchung Cheng tp->syn_fastopen = 0; 3042783237e8SYuchung Cheng kfree_skb(syn_data); 3043783237e8SYuchung Cheng done: 3044783237e8SYuchung Cheng fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ 3045783237e8SYuchung Cheng return err; 3046783237e8SYuchung Cheng } 3047783237e8SYuchung Cheng 304867edfef7SAndi Kleen /* Build a SYN and send it off. */ 30491da177e4SLinus Torvalds int tcp_connect(struct sock *sk) 30501da177e4SLinus Torvalds { 30511da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 30521da177e4SLinus Torvalds struct sk_buff *buff; 3053ee586811SEric Paris int err; 30541da177e4SLinus Torvalds 30551da177e4SLinus Torvalds tcp_connect_init(sk); 30561da177e4SLinus Torvalds 30572b916477SAndrey Vagin if (unlikely(tp->repair)) { 30582b916477SAndrey Vagin tcp_finish_connect(sk, NULL); 30592b916477SAndrey Vagin return 0; 30602b916477SAndrey Vagin } 30612b916477SAndrey Vagin 3062d179cd12SDavid S. Miller buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation); 30631da177e4SLinus Torvalds if (unlikely(buff == NULL)) 30641da177e4SLinus Torvalds return -ENOBUFS; 30651da177e4SLinus Torvalds 30661da177e4SLinus Torvalds /* Reserve space for headers. */ 30671da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 30681da177e4SLinus Torvalds 3069a3433f35SChangli Gao tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); 3070783237e8SYuchung Cheng tp->retrans_stamp = TCP_SKB_CB(buff)->when = tcp_time_stamp; 3071783237e8SYuchung Cheng tcp_connect_queue_skb(sk, buff); 3072e870a8efSIlpo Järvinen TCP_ECN_send_syn(sk, buff); 30731da177e4SLinus Torvalds 3074783237e8SYuchung Cheng /* Send off SYN; include data in Fast Open. */ 3075783237e8SYuchung Cheng err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : 3076783237e8SYuchung Cheng tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); 3077ee586811SEric Paris if (err == -ECONNREFUSED) 3078ee586811SEric Paris return err; 3079bd37a088SWei Yongjun 3080bd37a088SWei Yongjun /* We change tp->snd_nxt after the tcp_transmit_skb() call 3081bd37a088SWei Yongjun * in order to make this packet get counted in tcpOutSegs. 3082bd37a088SWei Yongjun */ 3083bd37a088SWei Yongjun tp->snd_nxt = tp->write_seq; 3084bd37a088SWei Yongjun tp->pushed_seq = tp->write_seq; 308581cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); 30861da177e4SLinus Torvalds 30871da177e4SLinus Torvalds /* Timer for repeating the SYN until an answer. */ 30883f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 30893f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 30901da177e4SLinus Torvalds return 0; 30911da177e4SLinus Torvalds } 30924bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_connect); 30931da177e4SLinus Torvalds 30941da177e4SLinus Torvalds /* Send out a delayed ack, the caller does the policy checking 30951da177e4SLinus Torvalds * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 30961da177e4SLinus Torvalds * for details. 30971da177e4SLinus Torvalds */ 30981da177e4SLinus Torvalds void tcp_send_delayed_ack(struct sock *sk) 30991da177e4SLinus Torvalds { 3100463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 3101463c84b9SArnaldo Carvalho de Melo int ato = icsk->icsk_ack.ato; 31021da177e4SLinus Torvalds unsigned long timeout; 31031da177e4SLinus Torvalds 31041da177e4SLinus Torvalds if (ato > TCP_DELACK_MIN) { 3105463c84b9SArnaldo Carvalho de Melo const struct tcp_sock *tp = tcp_sk(sk); 31061da177e4SLinus Torvalds int max_ato = HZ / 2; 31071da177e4SLinus Torvalds 3108056834d9SIlpo Järvinen if (icsk->icsk_ack.pingpong || 3109056834d9SIlpo Järvinen (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 31101da177e4SLinus Torvalds max_ato = TCP_DELACK_MAX; 31111da177e4SLinus Torvalds 31121da177e4SLinus Torvalds /* Slow path, intersegment interval is "high". */ 31131da177e4SLinus Torvalds 31141da177e4SLinus Torvalds /* If some rtt estimate is known, use it to bound delayed ack. 3115463c84b9SArnaldo Carvalho de Melo * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 31161da177e4SLinus Torvalds * directly. 31171da177e4SLinus Torvalds */ 3118740b0f18SEric Dumazet if (tp->srtt_us) { 3119740b0f18SEric Dumazet int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3), 3120740b0f18SEric Dumazet TCP_DELACK_MIN); 31211da177e4SLinus Torvalds 31221da177e4SLinus Torvalds if (rtt < max_ato) 31231da177e4SLinus Torvalds max_ato = rtt; 31241da177e4SLinus Torvalds } 31251da177e4SLinus Torvalds 31261da177e4SLinus Torvalds ato = min(ato, max_ato); 31271da177e4SLinus Torvalds } 31281da177e4SLinus Torvalds 31291da177e4SLinus Torvalds /* Stay within the limit we were given */ 31301da177e4SLinus Torvalds timeout = jiffies + ato; 31311da177e4SLinus Torvalds 31321da177e4SLinus Torvalds /* Use new timeout only if there wasn't a older one earlier. */ 3133463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 31341da177e4SLinus Torvalds /* If delack timer was blocked or is about to expire, 31351da177e4SLinus Torvalds * send ACK now. 31361da177e4SLinus Torvalds */ 3137463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.blocked || 3138463c84b9SArnaldo Carvalho de Melo time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 31391da177e4SLinus Torvalds tcp_send_ack(sk); 31401da177e4SLinus Torvalds return; 31411da177e4SLinus Torvalds } 31421da177e4SLinus Torvalds 3143463c84b9SArnaldo Carvalho de Melo if (!time_before(timeout, icsk->icsk_ack.timeout)) 3144463c84b9SArnaldo Carvalho de Melo timeout = icsk->icsk_ack.timeout; 31451da177e4SLinus Torvalds } 3146463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 3147463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.timeout = timeout; 3148463c84b9SArnaldo Carvalho de Melo sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 31491da177e4SLinus Torvalds } 31501da177e4SLinus Torvalds 31511da177e4SLinus Torvalds /* This routine sends an ack and also updates the window. */ 31521da177e4SLinus Torvalds void tcp_send_ack(struct sock *sk) 31531da177e4SLinus Torvalds { 31541da177e4SLinus Torvalds struct sk_buff *buff; 31551da177e4SLinus Torvalds 3156058dc334SIlpo Järvinen /* If we have been reset, we may not send again. */ 3157058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3158058dc334SIlpo Järvinen return; 3159058dc334SIlpo Järvinen 31601da177e4SLinus Torvalds /* We are not putting this on the write queue, so 31611da177e4SLinus Torvalds * tcp_transmit_skb() will set the ownership to this 31621da177e4SLinus Torvalds * sock. 31631da177e4SLinus Torvalds */ 316499a1dec7SMel Gorman buff = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC)); 31651da177e4SLinus Torvalds if (buff == NULL) { 3166463c84b9SArnaldo Carvalho de Melo inet_csk_schedule_ack(sk); 3167463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; 31683f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 31693f421baaSArnaldo Carvalho de Melo TCP_DELACK_MAX, TCP_RTO_MAX); 31701da177e4SLinus Torvalds return; 31711da177e4SLinus Torvalds } 31721da177e4SLinus Torvalds 31731da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 31741da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 3175a3433f35SChangli Gao tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); 31761da177e4SLinus Torvalds 31771da177e4SLinus Torvalds /* Send it off, this clears delayed acks for us. */ 31781da177e4SLinus Torvalds TCP_SKB_CB(buff)->when = tcp_time_stamp; 317999a1dec7SMel Gorman tcp_transmit_skb(sk, buff, 0, sk_gfp_atomic(sk, GFP_ATOMIC)); 31801da177e4SLinus Torvalds } 31811da177e4SLinus Torvalds 31821da177e4SLinus Torvalds /* This routine sends a packet with an out of date sequence 31831da177e4SLinus Torvalds * number. It assumes the other end will try to ack it. 31841da177e4SLinus Torvalds * 31851da177e4SLinus Torvalds * Question: what should we make while urgent mode? 31861da177e4SLinus Torvalds * 4.4BSD forces sending single byte of data. We cannot send 31871da177e4SLinus Torvalds * out of window data, because we have SND.NXT==SND.MAX... 31881da177e4SLinus Torvalds * 31891da177e4SLinus Torvalds * Current solution: to send TWO zero-length segments in urgent mode: 31901da177e4SLinus Torvalds * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 31911da177e4SLinus Torvalds * out-of-date with SND.UNA-1 to probe window. 31921da177e4SLinus Torvalds */ 31931da177e4SLinus Torvalds static int tcp_xmit_probe_skb(struct sock *sk, int urgent) 31941da177e4SLinus Torvalds { 31951da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 31961da177e4SLinus Torvalds struct sk_buff *skb; 31971da177e4SLinus Torvalds 31981da177e4SLinus Torvalds /* We don't queue it, tcp_transmit_skb() sets ownership. */ 319999a1dec7SMel Gorman skb = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC)); 32001da177e4SLinus Torvalds if (skb == NULL) 32011da177e4SLinus Torvalds return -1; 32021da177e4SLinus Torvalds 32031da177e4SLinus Torvalds /* Reserve space for headers and set control bits. */ 32041da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 32051da177e4SLinus Torvalds /* Use a previous sequence. This should cause the other 32061da177e4SLinus Torvalds * end to send an ack. Don't queue or clone SKB, just 32071da177e4SLinus Torvalds * send it. 32081da177e4SLinus Torvalds */ 3209a3433f35SChangli Gao tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); 32101da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 3211dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC); 32121da177e4SLinus Torvalds } 32131da177e4SLinus Torvalds 3214ee995283SPavel Emelyanov void tcp_send_window_probe(struct sock *sk) 3215ee995283SPavel Emelyanov { 3216ee995283SPavel Emelyanov if (sk->sk_state == TCP_ESTABLISHED) { 3217ee995283SPavel Emelyanov tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; 3218ee995283SPavel Emelyanov tcp_xmit_probe_skb(sk, 0); 3219ee995283SPavel Emelyanov } 3220ee995283SPavel Emelyanov } 3221ee995283SPavel Emelyanov 322267edfef7SAndi Kleen /* Initiate keepalive or window probe from timer. */ 32231da177e4SLinus Torvalds int tcp_write_wakeup(struct sock *sk) 32241da177e4SLinus Torvalds { 32251da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 32261da177e4SLinus Torvalds struct sk_buff *skb; 32271da177e4SLinus Torvalds 3228058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3229058dc334SIlpo Järvinen return -1; 3230058dc334SIlpo Järvinen 3231fe067e8aSDavid S. Miller if ((skb = tcp_send_head(sk)) != NULL && 323290840defSIlpo Järvinen before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { 32331da177e4SLinus Torvalds int err; 32340c54b85fSIlpo Järvinen unsigned int mss = tcp_current_mss(sk); 323590840defSIlpo Järvinen unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 32361da177e4SLinus Torvalds 32371da177e4SLinus Torvalds if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 32381da177e4SLinus Torvalds tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 32391da177e4SLinus Torvalds 32401da177e4SLinus Torvalds /* We are probing the opening of a window 32411da177e4SLinus Torvalds * but the window size is != 0 32421da177e4SLinus Torvalds * must have been a result SWS avoidance ( sender ) 32431da177e4SLinus Torvalds */ 32441da177e4SLinus Torvalds if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 32451da177e4SLinus Torvalds skb->len > mss) { 32461da177e4SLinus Torvalds seg_size = min(seg_size, mss); 32474de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 3248*6cc55e09SOctavian Purdila if (tcp_fragment(sk, skb, seg_size, mss, GFP_ATOMIC)) 32491da177e4SLinus Torvalds return -1; 32501da177e4SLinus Torvalds } else if (!tcp_skb_pcount(skb)) 3251846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss); 32521da177e4SLinus Torvalds 32534de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 32541da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 3255dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 325666f5fe62SIlpo Järvinen if (!err) 325766f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 32581da177e4SLinus Torvalds return err; 32591da177e4SLinus Torvalds } else { 326033f5f57eSIlpo Järvinen if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) 32614828e7f4SIlpo Järvinen tcp_xmit_probe_skb(sk, 1); 32621da177e4SLinus Torvalds return tcp_xmit_probe_skb(sk, 0); 32631da177e4SLinus Torvalds } 32641da177e4SLinus Torvalds } 32651da177e4SLinus Torvalds 32661da177e4SLinus Torvalds /* A window probe timeout has occurred. If window is not closed send 32671da177e4SLinus Torvalds * a partial packet else a zero probe. 32681da177e4SLinus Torvalds */ 32691da177e4SLinus Torvalds void tcp_send_probe0(struct sock *sk) 32701da177e4SLinus Torvalds { 3271463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 32721da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 32731da177e4SLinus Torvalds int err; 32741da177e4SLinus Torvalds 32751da177e4SLinus Torvalds err = tcp_write_wakeup(sk); 32761da177e4SLinus Torvalds 3277fe067e8aSDavid S. Miller if (tp->packets_out || !tcp_send_head(sk)) { 32781da177e4SLinus Torvalds /* Cancel probe timer, if it is not required. */ 32796687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 0; 3280463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff = 0; 32811da177e4SLinus Torvalds return; 32821da177e4SLinus Torvalds } 32831da177e4SLinus Torvalds 32841da177e4SLinus Torvalds if (err <= 0) { 3285463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_backoff < sysctl_tcp_retries2) 3286463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff++; 32876687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out++; 3288463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 32893f421baaSArnaldo Carvalho de Melo min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), 32903f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 32911da177e4SLinus Torvalds } else { 32921da177e4SLinus Torvalds /* If packet was not sent due to local congestion, 32936687e988SArnaldo Carvalho de Melo * do not backoff and do not remember icsk_probes_out. 32941da177e4SLinus Torvalds * Let local senders to fight for local resources. 32951da177e4SLinus Torvalds * 32961da177e4SLinus Torvalds * Use accumulated backoff yet. 32971da177e4SLinus Torvalds */ 32986687e988SArnaldo Carvalho de Melo if (!icsk->icsk_probes_out) 32996687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 1; 3300463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 3301463c84b9SArnaldo Carvalho de Melo min(icsk->icsk_rto << icsk->icsk_backoff, 33023f421baaSArnaldo Carvalho de Melo TCP_RESOURCE_PROBE_INTERVAL), 33033f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 33041da177e4SLinus Torvalds } 33051da177e4SLinus Torvalds } 3306