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 68519855c5SWilliam Allen Simpson int sysctl_tcp_cookie_size __read_mostly = 0; /* TCP_COOKIE_MAX */ 69e6b09ccaSDavid S. Miller EXPORT_SYMBOL_GPL(sysctl_tcp_cookie_size); 70519855c5SWilliam Allen Simpson 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 { 77*6ba8a3b1SNandita 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 8425985edcSLucas De Marchi /* Don't override Nagle indefinitely with F-RTO */ 858512430eSIlpo Järvinen if (tp->frto_counter == 2) 868512430eSIlpo Järvinen tp->frto_counter = 3; 8766f5fe62SIlpo Järvinen 8866f5fe62SIlpo Järvinen tp->packets_out += tcp_skb_pcount(skb); 89*6ba8a3b1SNandita Dukkipati if (!prior_packets || icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS || 90*6ba8a3b1SNandita Dukkipati icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) 91750ea2baSYuchung Cheng tcp_rearm_rto(sk); 921da177e4SLinus Torvalds } 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; 1691da177e4SLinus Torvalds 17035089bb2SDavid S. Miller if (sysctl_tcp_slow_start_after_idle && 17135089bb2SDavid S. Miller (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto)) 172463c84b9SArnaldo Carvalho de Melo tcp_cwnd_restart(sk, __sk_dst_get(sk)); 1731da177e4SLinus Torvalds 1741da177e4SLinus Torvalds tp->lsndtime = now; 1751da177e4SLinus Torvalds 1761da177e4SLinus Torvalds /* If it is a reply for ato after last received 1771da177e4SLinus Torvalds * packet, enter pingpong mode. 1781da177e4SLinus Torvalds */ 179463c84b9SArnaldo Carvalho de Melo if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato) 180463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pingpong = 1; 1811da177e4SLinus Torvalds } 1821da177e4SLinus Torvalds 18367edfef7SAndi Kleen /* Account for an ACK we sent. */ 18440efc6faSStephen Hemminger static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts) 1851da177e4SLinus Torvalds { 186463c84b9SArnaldo Carvalho de Melo tcp_dec_quickack_mode(sk, pkts); 187463c84b9SArnaldo Carvalho de Melo inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); 1881da177e4SLinus Torvalds } 1891da177e4SLinus Torvalds 1901da177e4SLinus Torvalds /* Determine a window scaling and initial window to offer. 1911da177e4SLinus Torvalds * Based on the assumption that the given amount of space 1921da177e4SLinus Torvalds * will be offered. Store the results in the tp structure. 1931da177e4SLinus Torvalds * NOTE: for smooth operation initial space offering should 1941da177e4SLinus Torvalds * be a multiple of mss if possible. We assume here that mss >= 1. 1951da177e4SLinus Torvalds * This MUST be enforced by all callers. 1961da177e4SLinus Torvalds */ 1971da177e4SLinus Torvalds void tcp_select_initial_window(int __space, __u32 mss, 1981da177e4SLinus Torvalds __u32 *rcv_wnd, __u32 *window_clamp, 19931d12926Slaurent chavey int wscale_ok, __u8 *rcv_wscale, 20031d12926Slaurent chavey __u32 init_rcv_wnd) 2011da177e4SLinus Torvalds { 2021da177e4SLinus Torvalds unsigned int space = (__space < 0 ? 0 : __space); 2031da177e4SLinus Torvalds 2041da177e4SLinus Torvalds /* If no clamp set the clamp to the max possible scaled window */ 2051da177e4SLinus Torvalds if (*window_clamp == 0) 2061da177e4SLinus Torvalds (*window_clamp) = (65535 << 14); 2071da177e4SLinus Torvalds space = min(*window_clamp, space); 2081da177e4SLinus Torvalds 2091da177e4SLinus Torvalds /* Quantize space offering to a multiple of mss if possible. */ 2101da177e4SLinus Torvalds if (space > mss) 2111da177e4SLinus Torvalds space = (space / mss) * mss; 2121da177e4SLinus Torvalds 2131da177e4SLinus Torvalds /* NOTE: offering an initial window larger than 32767 21415d99e02SRick Jones * will break some buggy TCP stacks. If the admin tells us 21515d99e02SRick Jones * it is likely we could be speaking with such a buggy stack 21615d99e02SRick Jones * we will truncate our initial window offering to 32K-1 21715d99e02SRick Jones * unless the remote has sent us a window scaling option, 21815d99e02SRick Jones * which we interpret as a sign the remote TCP is not 21915d99e02SRick Jones * misinterpreting the window field as a signed quantity. 2201da177e4SLinus Torvalds */ 22115d99e02SRick Jones if (sysctl_tcp_workaround_signed_windows) 2221da177e4SLinus Torvalds (*rcv_wnd) = min(space, MAX_TCP_WINDOW); 22315d99e02SRick Jones else 22415d99e02SRick Jones (*rcv_wnd) = space; 22515d99e02SRick Jones 2261da177e4SLinus Torvalds (*rcv_wscale) = 0; 2271da177e4SLinus Torvalds if (wscale_ok) { 2281da177e4SLinus Torvalds /* Set window scaling on max possible window 2291da177e4SLinus Torvalds * See RFC1323 for an explanation of the limit to 14 2301da177e4SLinus Torvalds */ 2311da177e4SLinus Torvalds space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max); 232316c1592SStephen Hemminger space = min_t(u32, space, *window_clamp); 2331da177e4SLinus Torvalds while (space > 65535 && (*rcv_wscale) < 14) { 2341da177e4SLinus Torvalds space >>= 1; 2351da177e4SLinus Torvalds (*rcv_wscale)++; 2361da177e4SLinus Torvalds } 2371da177e4SLinus Torvalds } 2381da177e4SLinus Torvalds 239356f0398SNandita Dukkipati /* Set initial window to a value enough for senders starting with 240356f0398SNandita Dukkipati * initial congestion window of TCP_DEFAULT_INIT_RCVWND. Place 241356f0398SNandita Dukkipati * a limit on the initial window when mss is larger than 1460. 242356f0398SNandita Dukkipati */ 2431da177e4SLinus Torvalds if (mss > (1 << *rcv_wscale)) { 244356f0398SNandita Dukkipati int init_cwnd = TCP_DEFAULT_INIT_RCVWND; 245356f0398SNandita Dukkipati if (mss > 1460) 246356f0398SNandita Dukkipati init_cwnd = 247356f0398SNandita Dukkipati max_t(u32, (1460 * TCP_DEFAULT_INIT_RCVWND) / mss, 2); 24831d12926Slaurent chavey /* when initializing use the value from init_rcv_wnd 24931d12926Slaurent chavey * rather than the default from above 25031d12926Slaurent chavey */ 251b1afde60SNandita Dukkipati if (init_rcv_wnd) 252b1afde60SNandita Dukkipati *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss); 253b1afde60SNandita Dukkipati else 254b1afde60SNandita Dukkipati *rcv_wnd = min(*rcv_wnd, init_cwnd * mss); 2551da177e4SLinus Torvalds } 2561da177e4SLinus Torvalds 2571da177e4SLinus Torvalds /* Set the clamp no higher than max representable value */ 2581da177e4SLinus Torvalds (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp); 2591da177e4SLinus Torvalds } 2604bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_select_initial_window); 2611da177e4SLinus Torvalds 2621da177e4SLinus Torvalds /* Chose a new window to advertise, update state in tcp_sock for the 2631da177e4SLinus Torvalds * socket, and return result with RFC1323 scaling applied. The return 2641da177e4SLinus Torvalds * value can be stuffed directly into th->window for an outgoing 2651da177e4SLinus Torvalds * frame. 2661da177e4SLinus Torvalds */ 26740efc6faSStephen Hemminger static u16 tcp_select_window(struct sock *sk) 2681da177e4SLinus Torvalds { 2691da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2701da177e4SLinus Torvalds u32 cur_win = tcp_receive_window(tp); 2711da177e4SLinus Torvalds u32 new_win = __tcp_select_window(sk); 2721da177e4SLinus Torvalds 2731da177e4SLinus Torvalds /* Never shrink the offered window */ 2741da177e4SLinus Torvalds if (new_win < cur_win) { 2751da177e4SLinus Torvalds /* Danger Will Robinson! 2761da177e4SLinus Torvalds * Don't update rcv_wup/rcv_wnd here or else 2771da177e4SLinus Torvalds * we will not be able to advertise a zero 2781da177e4SLinus Torvalds * window in time. --DaveM 2791da177e4SLinus Torvalds * 2801da177e4SLinus Torvalds * Relax Will Robinson. 2811da177e4SLinus Torvalds */ 282607bfbf2SPatrick McHardy new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale); 2831da177e4SLinus Torvalds } 2841da177e4SLinus Torvalds tp->rcv_wnd = new_win; 2851da177e4SLinus Torvalds tp->rcv_wup = tp->rcv_nxt; 2861da177e4SLinus Torvalds 2871da177e4SLinus Torvalds /* Make sure we do not exceed the maximum possible 2881da177e4SLinus Torvalds * scaled window. 2891da177e4SLinus Torvalds */ 29015d99e02SRick Jones if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows) 2911da177e4SLinus Torvalds new_win = min(new_win, MAX_TCP_WINDOW); 2921da177e4SLinus Torvalds else 2931da177e4SLinus Torvalds new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); 2941da177e4SLinus Torvalds 2951da177e4SLinus Torvalds /* RFC1323 scaling applied */ 2961da177e4SLinus Torvalds new_win >>= tp->rx_opt.rcv_wscale; 2971da177e4SLinus Torvalds 2981da177e4SLinus Torvalds /* If we advertise zero window, disable fast path. */ 2991da177e4SLinus Torvalds if (new_win == 0) 3001da177e4SLinus Torvalds tp->pred_flags = 0; 3011da177e4SLinus Torvalds 3021da177e4SLinus Torvalds return new_win; 3031da177e4SLinus Torvalds } 3041da177e4SLinus Torvalds 30567edfef7SAndi Kleen /* Packet ECN state for a SYN-ACK */ 306cf533ea5SEric Dumazet static inline void TCP_ECN_send_synack(const struct tcp_sock *tp, struct sk_buff *skb) 307bdf1ee5dSIlpo Järvinen { 3084de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR; 309bdf1ee5dSIlpo Järvinen if (!(tp->ecn_flags & TCP_ECN_OK)) 3104de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE; 311bdf1ee5dSIlpo Järvinen } 312bdf1ee5dSIlpo Järvinen 31367edfef7SAndi Kleen /* Packet ECN state for a SYN. */ 314bdf1ee5dSIlpo Järvinen static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb) 315bdf1ee5dSIlpo Järvinen { 316bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 317bdf1ee5dSIlpo Järvinen 318bdf1ee5dSIlpo Järvinen tp->ecn_flags = 0; 3195d134f1cSHannes Frederic Sowa if (sock_net(sk)->ipv4.sysctl_tcp_ecn == 1) { 3204de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR; 321bdf1ee5dSIlpo Järvinen tp->ecn_flags = TCP_ECN_OK; 322bdf1ee5dSIlpo Järvinen } 323bdf1ee5dSIlpo Järvinen } 324bdf1ee5dSIlpo Järvinen 325bdf1ee5dSIlpo Järvinen static __inline__ void 326cf533ea5SEric Dumazet TCP_ECN_make_synack(const struct request_sock *req, struct tcphdr *th) 327bdf1ee5dSIlpo Järvinen { 328bdf1ee5dSIlpo Järvinen if (inet_rsk(req)->ecn_ok) 329bdf1ee5dSIlpo Järvinen th->ece = 1; 330bdf1ee5dSIlpo Järvinen } 331bdf1ee5dSIlpo Järvinen 33267edfef7SAndi Kleen /* Set up ECN state for a packet on a ESTABLISHED socket that is about to 33367edfef7SAndi Kleen * be sent. 33467edfef7SAndi Kleen */ 335bdf1ee5dSIlpo Järvinen static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb, 336bdf1ee5dSIlpo Järvinen int tcp_header_len) 337bdf1ee5dSIlpo Järvinen { 338bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 339bdf1ee5dSIlpo Järvinen 340bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_OK) { 341bdf1ee5dSIlpo Järvinen /* Not-retransmitted data segment: set ECT and inject CWR. */ 342bdf1ee5dSIlpo Järvinen if (skb->len != tcp_header_len && 343bdf1ee5dSIlpo Järvinen !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { 344bdf1ee5dSIlpo Järvinen INET_ECN_xmit(sk); 345bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) { 346bdf1ee5dSIlpo Järvinen tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; 347bdf1ee5dSIlpo Järvinen tcp_hdr(skb)->cwr = 1; 348bdf1ee5dSIlpo Järvinen skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; 349bdf1ee5dSIlpo Järvinen } 350bdf1ee5dSIlpo Järvinen } else { 351bdf1ee5dSIlpo Järvinen /* ACK or retransmitted segment: clear ECT|CE */ 352bdf1ee5dSIlpo Järvinen INET_ECN_dontxmit(sk); 353bdf1ee5dSIlpo Järvinen } 354bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) 355bdf1ee5dSIlpo Järvinen tcp_hdr(skb)->ece = 1; 356bdf1ee5dSIlpo Järvinen } 357bdf1ee5dSIlpo Järvinen } 358bdf1ee5dSIlpo Järvinen 359e870a8efSIlpo Järvinen /* Constructs common control bits of non-data skb. If SYN/FIN is present, 360e870a8efSIlpo Järvinen * auto increment end seqno. 361e870a8efSIlpo Järvinen */ 362e870a8efSIlpo Järvinen static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags) 363e870a8efSIlpo Järvinen { 3642e8e18efSDavid S. Miller skb->ip_summed = CHECKSUM_PARTIAL; 365e870a8efSIlpo Järvinen skb->csum = 0; 366e870a8efSIlpo Järvinen 3674de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags; 368e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->sacked = 0; 369e870a8efSIlpo Järvinen 370e870a8efSIlpo Järvinen skb_shinfo(skb)->gso_segs = 1; 371e870a8efSIlpo Järvinen skb_shinfo(skb)->gso_size = 0; 372e870a8efSIlpo Järvinen skb_shinfo(skb)->gso_type = 0; 373e870a8efSIlpo Järvinen 374e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->seq = seq; 375a3433f35SChangli Gao if (flags & (TCPHDR_SYN | TCPHDR_FIN)) 376e870a8efSIlpo Järvinen seq++; 377e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->end_seq = seq; 378e870a8efSIlpo Järvinen } 379e870a8efSIlpo Järvinen 380a2a385d6SEric Dumazet static inline bool tcp_urg_mode(const struct tcp_sock *tp) 38133f5f57eSIlpo Järvinen { 38233f5f57eSIlpo Järvinen return tp->snd_una != tp->snd_up; 38333f5f57eSIlpo Järvinen } 38433f5f57eSIlpo Järvinen 38533ad798cSAdam Langley #define OPTION_SACK_ADVERTISE (1 << 0) 38633ad798cSAdam Langley #define OPTION_TS (1 << 1) 38733ad798cSAdam Langley #define OPTION_MD5 (1 << 2) 38889e95a61SOri Finkelman #define OPTION_WSCALE (1 << 3) 389bd0388aeSWilliam Allen Simpson #define OPTION_COOKIE_EXTENSION (1 << 4) 3902100c8d2SYuchung Cheng #define OPTION_FAST_OPEN_COOKIE (1 << 8) 39133ad798cSAdam Langley 39233ad798cSAdam Langley struct tcp_out_options { 3932100c8d2SYuchung Cheng u16 options; /* bit field of OPTION_* */ 3942100c8d2SYuchung Cheng u16 mss; /* 0 to disable */ 39533ad798cSAdam Langley u8 ws; /* window scale, 0 to disable */ 39633ad798cSAdam Langley u8 num_sack_blocks; /* number of SACK blocks to include */ 397bd0388aeSWilliam Allen Simpson u8 hash_size; /* bytes in hash_location */ 398bd0388aeSWilliam Allen Simpson __u8 *hash_location; /* temporary pointer, overloaded */ 3992100c8d2SYuchung Cheng __u32 tsval, tsecr; /* need to include OPTION_TS */ 4002100c8d2SYuchung Cheng struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */ 40133ad798cSAdam Langley }; 40233ad798cSAdam Langley 403bd0388aeSWilliam Allen Simpson /* The sysctl int routines are generic, so check consistency here. 404bd0388aeSWilliam Allen Simpson */ 405bd0388aeSWilliam Allen Simpson static u8 tcp_cookie_size_check(u8 desired) 406bd0388aeSWilliam Allen Simpson { 407f1987257SEric Dumazet int cookie_size; 408f1987257SEric Dumazet 409f1987257SEric Dumazet if (desired > 0) 410bd0388aeSWilliam Allen Simpson /* previously specified */ 411bd0388aeSWilliam Allen Simpson return desired; 412f1987257SEric Dumazet 413f1987257SEric Dumazet cookie_size = ACCESS_ONCE(sysctl_tcp_cookie_size); 414f1987257SEric Dumazet if (cookie_size <= 0) 415bd0388aeSWilliam Allen Simpson /* no default specified */ 416bd0388aeSWilliam Allen Simpson return 0; 417f1987257SEric Dumazet 418f1987257SEric Dumazet if (cookie_size <= TCP_COOKIE_MIN) 419bd0388aeSWilliam Allen Simpson /* value too small, specify minimum */ 420bd0388aeSWilliam Allen Simpson return TCP_COOKIE_MIN; 421f1987257SEric Dumazet 422f1987257SEric Dumazet if (cookie_size >= TCP_COOKIE_MAX) 423bd0388aeSWilliam Allen Simpson /* value too large, specify maximum */ 424bd0388aeSWilliam Allen Simpson return TCP_COOKIE_MAX; 425f1987257SEric Dumazet 426f1987257SEric Dumazet if (cookie_size & 1) 427bd0388aeSWilliam Allen Simpson /* 8-bit multiple, illegal, fix it */ 428f1987257SEric Dumazet cookie_size++; 429f1987257SEric Dumazet 430f1987257SEric Dumazet return (u8)cookie_size; 431bd0388aeSWilliam Allen Simpson } 432bd0388aeSWilliam Allen Simpson 43367edfef7SAndi Kleen /* Write previously computed TCP options to the packet. 43467edfef7SAndi Kleen * 43567edfef7SAndi Kleen * Beware: Something in the Internet is very sensitive to the ordering of 436fd6149d3SIlpo Järvinen * TCP options, we learned this through the hard way, so be careful here. 437fd6149d3SIlpo Järvinen * Luckily we can at least blame others for their non-compliance but from 438fd6149d3SIlpo Järvinen * inter-operatibility perspective it seems that we're somewhat stuck with 439fd6149d3SIlpo Järvinen * the ordering which we have been using if we want to keep working with 440fd6149d3SIlpo Järvinen * those broken things (not that it currently hurts anybody as there isn't 441fd6149d3SIlpo Järvinen * particular reason why the ordering would need to be changed). 442fd6149d3SIlpo Järvinen * 443fd6149d3SIlpo Järvinen * At least SACK_PERM as the first option is known to lead to a disaster 444fd6149d3SIlpo Järvinen * (but it may well be that other scenarios fail similarly). 445fd6149d3SIlpo Järvinen */ 44633ad798cSAdam Langley static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp, 447bd0388aeSWilliam Allen Simpson struct tcp_out_options *opts) 448bd0388aeSWilliam Allen Simpson { 4492100c8d2SYuchung Cheng u16 options = opts->options; /* mungable copy */ 450bd0388aeSWilliam Allen Simpson 451bd0388aeSWilliam Allen Simpson /* Having both authentication and cookies for security is redundant, 452bd0388aeSWilliam Allen Simpson * and there's certainly not enough room. Instead, the cookie-less 453bd0388aeSWilliam Allen Simpson * extension variant is proposed. 454bd0388aeSWilliam Allen Simpson * 455bd0388aeSWilliam Allen Simpson * Consider the pessimal case with authentication. The options 456bd0388aeSWilliam Allen Simpson * could look like: 457bd0388aeSWilliam Allen Simpson * COOKIE|MD5(20) + MSS(4) + SACK|TS(12) + WSCALE(4) == 40 458bd0388aeSWilliam Allen Simpson */ 459bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_MD5 & options)) { 460bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_COOKIE_EXTENSION & options)) { 461bd0388aeSWilliam Allen Simpson *ptr++ = htonl((TCPOPT_COOKIE << 24) | 462bd0388aeSWilliam Allen Simpson (TCPOLEN_COOKIE_BASE << 16) | 463bd0388aeSWilliam Allen Simpson (TCPOPT_MD5SIG << 8) | 464bd0388aeSWilliam Allen Simpson TCPOLEN_MD5SIG); 465bd0388aeSWilliam Allen Simpson } else { 46633ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 46733ad798cSAdam Langley (TCPOPT_NOP << 16) | 46833ad798cSAdam Langley (TCPOPT_MD5SIG << 8) | 46933ad798cSAdam Langley TCPOLEN_MD5SIG); 470bd0388aeSWilliam Allen Simpson } 471bd0388aeSWilliam Allen Simpson options &= ~OPTION_COOKIE_EXTENSION; 472bd0388aeSWilliam Allen Simpson /* overload cookie hash location */ 473bd0388aeSWilliam Allen Simpson opts->hash_location = (__u8 *)ptr; 47433ad798cSAdam Langley ptr += 4; 47533ad798cSAdam Langley } 47633ad798cSAdam Langley 477fd6149d3SIlpo Järvinen if (unlikely(opts->mss)) { 478fd6149d3SIlpo Järvinen *ptr++ = htonl((TCPOPT_MSS << 24) | 479fd6149d3SIlpo Järvinen (TCPOLEN_MSS << 16) | 480fd6149d3SIlpo Järvinen opts->mss); 481fd6149d3SIlpo Järvinen } 482fd6149d3SIlpo Järvinen 483bd0388aeSWilliam Allen Simpson if (likely(OPTION_TS & options)) { 484bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 48533ad798cSAdam Langley *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | 48633ad798cSAdam Langley (TCPOLEN_SACK_PERM << 16) | 48733ad798cSAdam Langley (TCPOPT_TIMESTAMP << 8) | 48833ad798cSAdam Langley TCPOLEN_TIMESTAMP); 489bd0388aeSWilliam Allen Simpson options &= ~OPTION_SACK_ADVERTISE; 49033ad798cSAdam Langley } else { 491496c98dfSYOSHIFUJI Hideaki *ptr++ = htonl((TCPOPT_NOP << 24) | 49240efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 49340efc6faSStephen Hemminger (TCPOPT_TIMESTAMP << 8) | 49440efc6faSStephen Hemminger TCPOLEN_TIMESTAMP); 49540efc6faSStephen Hemminger } 49633ad798cSAdam Langley *ptr++ = htonl(opts->tsval); 49733ad798cSAdam Langley *ptr++ = htonl(opts->tsecr); 49833ad798cSAdam Langley } 49933ad798cSAdam Langley 500bd0388aeSWilliam Allen Simpson /* Specification requires after timestamp, so do it now. 501bd0388aeSWilliam Allen Simpson * 502bd0388aeSWilliam Allen Simpson * Consider the pessimal case without authentication. The options 503bd0388aeSWilliam Allen Simpson * could look like: 504bd0388aeSWilliam Allen Simpson * MSS(4) + SACK|TS(12) + COOKIE(20) + WSCALE(4) == 40 505bd0388aeSWilliam Allen Simpson */ 506bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_COOKIE_EXTENSION & options)) { 507bd0388aeSWilliam Allen Simpson __u8 *cookie_copy = opts->hash_location; 508bd0388aeSWilliam Allen Simpson u8 cookie_size = opts->hash_size; 509bd0388aeSWilliam Allen Simpson 510bd0388aeSWilliam Allen Simpson /* 8-bit multiple handled in tcp_cookie_size_check() above, 511bd0388aeSWilliam Allen Simpson * and elsewhere. 512bd0388aeSWilliam Allen Simpson */ 513bd0388aeSWilliam Allen Simpson if (0x2 & cookie_size) { 514bd0388aeSWilliam Allen Simpson __u8 *p = (__u8 *)ptr; 515bd0388aeSWilliam Allen Simpson 516bd0388aeSWilliam Allen Simpson /* 16-bit multiple */ 517bd0388aeSWilliam Allen Simpson *p++ = TCPOPT_COOKIE; 518bd0388aeSWilliam Allen Simpson *p++ = TCPOLEN_COOKIE_BASE + cookie_size; 519bd0388aeSWilliam Allen Simpson *p++ = *cookie_copy++; 520bd0388aeSWilliam Allen Simpson *p++ = *cookie_copy++; 521bd0388aeSWilliam Allen Simpson ptr++; 522bd0388aeSWilliam Allen Simpson cookie_size -= 2; 523bd0388aeSWilliam Allen Simpson } else { 524bd0388aeSWilliam Allen Simpson /* 32-bit multiple */ 525bd0388aeSWilliam Allen Simpson *ptr++ = htonl(((TCPOPT_NOP << 24) | 526bd0388aeSWilliam Allen Simpson (TCPOPT_NOP << 16) | 527bd0388aeSWilliam Allen Simpson (TCPOPT_COOKIE << 8) | 528bd0388aeSWilliam Allen Simpson TCPOLEN_COOKIE_BASE) + 529bd0388aeSWilliam Allen Simpson cookie_size); 530bd0388aeSWilliam Allen Simpson } 531bd0388aeSWilliam Allen Simpson 532bd0388aeSWilliam Allen Simpson if (cookie_size > 0) { 533bd0388aeSWilliam Allen Simpson memcpy(ptr, cookie_copy, cookie_size); 534bd0388aeSWilliam Allen Simpson ptr += (cookie_size / 4); 535bd0388aeSWilliam Allen Simpson } 536bd0388aeSWilliam Allen Simpson } 537bd0388aeSWilliam Allen Simpson 538bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 53933ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 54033ad798cSAdam Langley (TCPOPT_NOP << 16) | 54133ad798cSAdam Langley (TCPOPT_SACK_PERM << 8) | 54233ad798cSAdam Langley TCPOLEN_SACK_PERM); 54333ad798cSAdam Langley } 54433ad798cSAdam Langley 545bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_WSCALE & options)) { 54633ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 54733ad798cSAdam Langley (TCPOPT_WINDOW << 16) | 54833ad798cSAdam Langley (TCPOLEN_WINDOW << 8) | 54933ad798cSAdam Langley opts->ws); 55033ad798cSAdam Langley } 55133ad798cSAdam Langley 55233ad798cSAdam Langley if (unlikely(opts->num_sack_blocks)) { 55333ad798cSAdam Langley struct tcp_sack_block *sp = tp->rx_opt.dsack ? 55433ad798cSAdam Langley tp->duplicate_sack : tp->selective_acks; 55540efc6faSStephen Hemminger int this_sack; 55640efc6faSStephen Hemminger 55740efc6faSStephen Hemminger *ptr++ = htonl((TCPOPT_NOP << 24) | 55840efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 55940efc6faSStephen Hemminger (TCPOPT_SACK << 8) | 56033ad798cSAdam Langley (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * 56140efc6faSStephen Hemminger TCPOLEN_SACK_PERBLOCK))); 5622de979bdSStephen Hemminger 56333ad798cSAdam Langley for (this_sack = 0; this_sack < opts->num_sack_blocks; 56433ad798cSAdam Langley ++this_sack) { 56540efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].start_seq); 56640efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].end_seq); 56740efc6faSStephen Hemminger } 5682de979bdSStephen Hemminger 56940efc6faSStephen Hemminger tp->rx_opt.dsack = 0; 57040efc6faSStephen Hemminger } 5712100c8d2SYuchung Cheng 5722100c8d2SYuchung Cheng if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) { 5732100c8d2SYuchung Cheng struct tcp_fastopen_cookie *foc = opts->fastopen_cookie; 5742100c8d2SYuchung Cheng 5752100c8d2SYuchung Cheng *ptr++ = htonl((TCPOPT_EXP << 24) | 5762100c8d2SYuchung Cheng ((TCPOLEN_EXP_FASTOPEN_BASE + foc->len) << 16) | 5772100c8d2SYuchung Cheng TCPOPT_FASTOPEN_MAGIC); 5782100c8d2SYuchung Cheng 5792100c8d2SYuchung Cheng memcpy(ptr, foc->val, foc->len); 5802100c8d2SYuchung Cheng if ((foc->len & 3) == 2) { 5812100c8d2SYuchung Cheng u8 *align = ((u8 *)ptr) + foc->len; 5822100c8d2SYuchung Cheng align[0] = align[1] = TCPOPT_NOP; 5832100c8d2SYuchung Cheng } 5842100c8d2SYuchung Cheng ptr += (foc->len + 3) >> 2; 5852100c8d2SYuchung Cheng } 58640efc6faSStephen Hemminger } 58740efc6faSStephen Hemminger 58867edfef7SAndi Kleen /* Compute TCP options for SYN packets. This is not the final 58967edfef7SAndi Kleen * network wire format yet. 59067edfef7SAndi Kleen */ 59195c96174SEric Dumazet static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb, 59233ad798cSAdam Langley struct tcp_out_options *opts, 593cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 594cf533ea5SEric Dumazet { 59533ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 596bd0388aeSWilliam Allen Simpson struct tcp_cookie_values *cvp = tp->cookie_values; 59795c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 598bd0388aeSWilliam Allen Simpson u8 cookie_size = (!tp->rx_opt.cookie_out_never && cvp != NULL) ? 599bd0388aeSWilliam Allen Simpson tcp_cookie_size_check(cvp->cookie_desired) : 600bd0388aeSWilliam Allen Simpson 0; 601783237e8SYuchung Cheng struct tcp_fastopen_request *fastopen = tp->fastopen_req; 60233ad798cSAdam Langley 603cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 60433ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 60533ad798cSAdam Langley if (*md5) { 60633ad798cSAdam Langley opts->options |= OPTION_MD5; 607bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 608cfb6eeb4SYOSHIFUJI Hideaki } 60933ad798cSAdam Langley #else 61033ad798cSAdam Langley *md5 = NULL; 611cfb6eeb4SYOSHIFUJI Hideaki #endif 61233ad798cSAdam Langley 61333ad798cSAdam Langley /* We always get an MSS option. The option bytes which will be seen in 61433ad798cSAdam Langley * normal data packets should timestamps be used, must be in the MSS 61533ad798cSAdam Langley * advertised. But we subtract them from tp->mss_cache so that 61633ad798cSAdam Langley * calculations in tcp_sendmsg are simpler etc. So account for this 61733ad798cSAdam Langley * fact here if necessary. If we don't do this correctly, as a 61833ad798cSAdam Langley * receiver we won't recognize data packets as being full sized when we 61933ad798cSAdam Langley * should, and thus we won't abide by the delayed ACK rules correctly. 62033ad798cSAdam Langley * SACKs don't matter, we never delay an ACK when we have any of those 62133ad798cSAdam Langley * going out. */ 62233ad798cSAdam Langley opts->mss = tcp_advertise_mss(sk); 623bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 62433ad798cSAdam Langley 625bb5b7c11SDavid S. Miller if (likely(sysctl_tcp_timestamps && *md5 == NULL)) { 62633ad798cSAdam Langley opts->options |= OPTION_TS; 627ee684b6fSAndrey Vagin opts->tsval = TCP_SKB_CB(skb)->when + tp->tsoffset; 62833ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 629bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 63033ad798cSAdam Langley } 631bb5b7c11SDavid S. Miller if (likely(sysctl_tcp_window_scaling)) { 63233ad798cSAdam Langley opts->ws = tp->rx_opt.rcv_wscale; 63389e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 634bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 63533ad798cSAdam Langley } 636bb5b7c11SDavid S. Miller if (likely(sysctl_tcp_sack)) { 63733ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 638b32d1310SDavid S. Miller if (unlikely(!(OPTION_TS & opts->options))) 639bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 64033ad798cSAdam Langley } 64133ad798cSAdam Langley 642783237e8SYuchung Cheng if (fastopen && fastopen->cookie.len >= 0) { 643783237e8SYuchung Cheng u32 need = TCPOLEN_EXP_FASTOPEN_BASE + fastopen->cookie.len; 644783237e8SYuchung Cheng need = (need + 3) & ~3U; /* Align to 32 bits */ 645783237e8SYuchung Cheng if (remaining >= need) { 646783237e8SYuchung Cheng opts->options |= OPTION_FAST_OPEN_COOKIE; 647783237e8SYuchung Cheng opts->fastopen_cookie = &fastopen->cookie; 648783237e8SYuchung Cheng remaining -= need; 649783237e8SYuchung Cheng tp->syn_fastopen = 1; 650783237e8SYuchung Cheng } 651783237e8SYuchung Cheng } 652bd0388aeSWilliam Allen Simpson /* Note that timestamps are required by the specification. 653bd0388aeSWilliam Allen Simpson * 654bd0388aeSWilliam Allen Simpson * Odd numbers of bytes are prohibited by the specification, ensuring 655bd0388aeSWilliam Allen Simpson * that the cookie is 16-bit aligned, and the resulting cookie pair is 656bd0388aeSWilliam Allen Simpson * 32-bit aligned. 657bd0388aeSWilliam Allen Simpson */ 658bd0388aeSWilliam Allen Simpson if (*md5 == NULL && 659bd0388aeSWilliam Allen Simpson (OPTION_TS & opts->options) && 660bd0388aeSWilliam Allen Simpson cookie_size > 0) { 661bd0388aeSWilliam Allen Simpson int need = TCPOLEN_COOKIE_BASE + cookie_size; 662bd0388aeSWilliam Allen Simpson 663bd0388aeSWilliam Allen Simpson if (0x2 & need) { 664bd0388aeSWilliam Allen Simpson /* 32-bit multiple */ 665bd0388aeSWilliam Allen Simpson need += 2; /* NOPs */ 666bd0388aeSWilliam Allen Simpson 667bd0388aeSWilliam Allen Simpson if (need > remaining) { 668bd0388aeSWilliam Allen Simpson /* try shrinking cookie to fit */ 669bd0388aeSWilliam Allen Simpson cookie_size -= 2; 670bd0388aeSWilliam Allen Simpson need -= 4; 671bd0388aeSWilliam Allen Simpson } 672bd0388aeSWilliam Allen Simpson } 673bd0388aeSWilliam Allen Simpson while (need > remaining && TCP_COOKIE_MIN <= cookie_size) { 674bd0388aeSWilliam Allen Simpson cookie_size -= 4; 675bd0388aeSWilliam Allen Simpson need -= 4; 676bd0388aeSWilliam Allen Simpson } 677bd0388aeSWilliam Allen Simpson if (TCP_COOKIE_MIN <= cookie_size) { 678bd0388aeSWilliam Allen Simpson opts->options |= OPTION_COOKIE_EXTENSION; 679bd0388aeSWilliam Allen Simpson opts->hash_location = (__u8 *)&cvp->cookie_pair[0]; 680bd0388aeSWilliam Allen Simpson opts->hash_size = cookie_size; 681bd0388aeSWilliam Allen Simpson 682bd0388aeSWilliam Allen Simpson /* Remember for future incarnations. */ 683bd0388aeSWilliam Allen Simpson cvp->cookie_desired = cookie_size; 684bd0388aeSWilliam Allen Simpson 685bd0388aeSWilliam Allen Simpson if (cvp->cookie_desired != cvp->cookie_pair_size) { 686bd0388aeSWilliam Allen Simpson /* Currently use random bytes as a nonce, 687bd0388aeSWilliam Allen Simpson * assuming these are completely unpredictable 688bd0388aeSWilliam Allen Simpson * by hostile users of the same system. 689bd0388aeSWilliam Allen Simpson */ 690bd0388aeSWilliam Allen Simpson get_random_bytes(&cvp->cookie_pair[0], 691bd0388aeSWilliam Allen Simpson cookie_size); 692bd0388aeSWilliam Allen Simpson cvp->cookie_pair_size = cookie_size; 693bd0388aeSWilliam Allen Simpson } 694bd0388aeSWilliam Allen Simpson 695bd0388aeSWilliam Allen Simpson remaining -= need; 696bd0388aeSWilliam Allen Simpson } 697bd0388aeSWilliam Allen Simpson } 698bd0388aeSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 69933ad798cSAdam Langley } 70033ad798cSAdam Langley 70167edfef7SAndi Kleen /* Set up TCP options for SYN-ACKs. */ 70295c96174SEric Dumazet static unsigned int tcp_synack_options(struct sock *sk, 70333ad798cSAdam Langley struct request_sock *req, 70495c96174SEric Dumazet unsigned int mss, struct sk_buff *skb, 70533ad798cSAdam Langley struct tcp_out_options *opts, 7064957faadSWilliam Allen Simpson struct tcp_md5sig_key **md5, 7078336886fSJerry Chu struct tcp_extend_values *xvp, 7088336886fSJerry Chu struct tcp_fastopen_cookie *foc) 7094957faadSWilliam Allen Simpson { 71033ad798cSAdam Langley struct inet_request_sock *ireq = inet_rsk(req); 71195c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 7124957faadSWilliam Allen Simpson u8 cookie_plus = (xvp != NULL && !xvp->cookie_out_never) ? 7134957faadSWilliam Allen Simpson xvp->cookie_plus : 7144957faadSWilliam Allen Simpson 0; 71533ad798cSAdam Langley 71633ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 71733ad798cSAdam Langley *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req); 71833ad798cSAdam Langley if (*md5) { 71933ad798cSAdam Langley opts->options |= OPTION_MD5; 7204957faadSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 7214957faadSWilliam Allen Simpson 7224957faadSWilliam Allen Simpson /* We can't fit any SACK blocks in a packet with MD5 + TS 7234957faadSWilliam Allen Simpson * options. There was discussion about disabling SACK 7244957faadSWilliam Allen Simpson * rather than TS in order to fit in better with old, 7254957faadSWilliam Allen Simpson * buggy kernels, but that was deemed to be unnecessary. 7264957faadSWilliam Allen Simpson */ 727de213e5eSEric Dumazet ireq->tstamp_ok &= !ireq->sack_ok; 72833ad798cSAdam Langley } 72933ad798cSAdam Langley #else 73033ad798cSAdam Langley *md5 = NULL; 73133ad798cSAdam Langley #endif 73233ad798cSAdam Langley 7334957faadSWilliam Allen Simpson /* We always send an MSS option. */ 73433ad798cSAdam Langley opts->mss = mss; 7354957faadSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 73633ad798cSAdam Langley 73733ad798cSAdam Langley if (likely(ireq->wscale_ok)) { 73833ad798cSAdam Langley opts->ws = ireq->rcv_wscale; 73989e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 7404957faadSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 74133ad798cSAdam Langley } 742de213e5eSEric Dumazet if (likely(ireq->tstamp_ok)) { 74333ad798cSAdam Langley opts->options |= OPTION_TS; 74433ad798cSAdam Langley opts->tsval = TCP_SKB_CB(skb)->when; 74533ad798cSAdam Langley opts->tsecr = req->ts_recent; 7464957faadSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 74733ad798cSAdam Langley } 74833ad798cSAdam Langley if (likely(ireq->sack_ok)) { 74933ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 750de213e5eSEric Dumazet if (unlikely(!ireq->tstamp_ok)) 7514957faadSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 75233ad798cSAdam Langley } 7538336886fSJerry Chu if (foc != NULL) { 7548336886fSJerry Chu u32 need = TCPOLEN_EXP_FASTOPEN_BASE + foc->len; 7558336886fSJerry Chu need = (need + 3) & ~3U; /* Align to 32 bits */ 7568336886fSJerry Chu if (remaining >= need) { 7578336886fSJerry Chu opts->options |= OPTION_FAST_OPEN_COOKIE; 7588336886fSJerry Chu opts->fastopen_cookie = foc; 7598336886fSJerry Chu remaining -= need; 7608336886fSJerry Chu } 7618336886fSJerry Chu } 7624957faadSWilliam Allen Simpson /* Similar rationale to tcp_syn_options() applies here, too. 7634957faadSWilliam Allen Simpson * If the <SYN> options fit, the same options should fit now! 7644957faadSWilliam Allen Simpson */ 7654957faadSWilliam Allen Simpson if (*md5 == NULL && 766de213e5eSEric Dumazet ireq->tstamp_ok && 7674957faadSWilliam Allen Simpson cookie_plus > TCPOLEN_COOKIE_BASE) { 7684957faadSWilliam Allen Simpson int need = cookie_plus; /* has TCPOLEN_COOKIE_BASE */ 7694957faadSWilliam Allen Simpson 7704957faadSWilliam Allen Simpson if (0x2 & need) { 7714957faadSWilliam Allen Simpson /* 32-bit multiple */ 7724957faadSWilliam Allen Simpson need += 2; /* NOPs */ 7734957faadSWilliam Allen Simpson } 7744957faadSWilliam Allen Simpson if (need <= remaining) { 7754957faadSWilliam Allen Simpson opts->options |= OPTION_COOKIE_EXTENSION; 7764957faadSWilliam Allen Simpson opts->hash_size = cookie_plus - TCPOLEN_COOKIE_BASE; 7774957faadSWilliam Allen Simpson remaining -= need; 7784957faadSWilliam Allen Simpson } else { 7794957faadSWilliam Allen Simpson /* There's no error return, so flag it. */ 7804957faadSWilliam Allen Simpson xvp->cookie_out_never = 1; /* true */ 7814957faadSWilliam Allen Simpson opts->hash_size = 0; 7824957faadSWilliam Allen Simpson } 7834957faadSWilliam Allen Simpson } 7844957faadSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 78533ad798cSAdam Langley } 78633ad798cSAdam Langley 78767edfef7SAndi Kleen /* Compute TCP options for ESTABLISHED sockets. This is not the 78867edfef7SAndi Kleen * final wire format yet. 78967edfef7SAndi Kleen */ 79095c96174SEric Dumazet static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb, 79133ad798cSAdam Langley struct tcp_out_options *opts, 792cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 793cf533ea5SEric Dumazet { 79433ad798cSAdam Langley struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL; 79533ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 79695c96174SEric Dumazet unsigned int size = 0; 797cabeccbdSIlpo Järvinen unsigned int eff_sacks; 79833ad798cSAdam Langley 79933ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 80033ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 80133ad798cSAdam Langley if (unlikely(*md5)) { 80233ad798cSAdam Langley opts->options |= OPTION_MD5; 80333ad798cSAdam Langley size += TCPOLEN_MD5SIG_ALIGNED; 80433ad798cSAdam Langley } 80533ad798cSAdam Langley #else 80633ad798cSAdam Langley *md5 = NULL; 80733ad798cSAdam Langley #endif 80833ad798cSAdam Langley 80933ad798cSAdam Langley if (likely(tp->rx_opt.tstamp_ok)) { 81033ad798cSAdam Langley opts->options |= OPTION_TS; 811ee684b6fSAndrey Vagin opts->tsval = tcb ? tcb->when + tp->tsoffset : 0; 81233ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 81333ad798cSAdam Langley size += TCPOLEN_TSTAMP_ALIGNED; 81433ad798cSAdam Langley } 81533ad798cSAdam Langley 816cabeccbdSIlpo Järvinen eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; 817cabeccbdSIlpo Järvinen if (unlikely(eff_sacks)) { 81895c96174SEric Dumazet const unsigned int remaining = MAX_TCP_OPTION_SPACE - size; 81933ad798cSAdam Langley opts->num_sack_blocks = 82095c96174SEric Dumazet min_t(unsigned int, eff_sacks, 82133ad798cSAdam Langley (remaining - TCPOLEN_SACK_BASE_ALIGNED) / 82233ad798cSAdam Langley TCPOLEN_SACK_PERBLOCK); 82333ad798cSAdam Langley size += TCPOLEN_SACK_BASE_ALIGNED + 82433ad798cSAdam Langley opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; 82533ad798cSAdam Langley } 82633ad798cSAdam Langley 82733ad798cSAdam Langley return size; 82840efc6faSStephen Hemminger } 8291da177e4SLinus Torvalds 83046d3ceabSEric Dumazet 83146d3ceabSEric Dumazet /* TCP SMALL QUEUES (TSQ) 83246d3ceabSEric Dumazet * 83346d3ceabSEric Dumazet * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev) 83446d3ceabSEric Dumazet * to reduce RTT and bufferbloat. 83546d3ceabSEric Dumazet * We do this using a special skb destructor (tcp_wfree). 83646d3ceabSEric Dumazet * 83746d3ceabSEric Dumazet * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb 83846d3ceabSEric Dumazet * needs to be reallocated in a driver. 83946d3ceabSEric Dumazet * The invariant being skb->truesize substracted from sk->sk_wmem_alloc 84046d3ceabSEric Dumazet * 84146d3ceabSEric Dumazet * Since transmit from skb destructor is forbidden, we use a tasklet 84246d3ceabSEric Dumazet * to process all sockets that eventually need to send more skbs. 84346d3ceabSEric Dumazet * We use one tasklet per cpu, with its own queue of sockets. 84446d3ceabSEric Dumazet */ 84546d3ceabSEric Dumazet struct tsq_tasklet { 84646d3ceabSEric Dumazet struct tasklet_struct tasklet; 84746d3ceabSEric Dumazet struct list_head head; /* queue of tcp sockets */ 84846d3ceabSEric Dumazet }; 84946d3ceabSEric Dumazet static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet); 85046d3ceabSEric Dumazet 8516f458dfbSEric Dumazet static void tcp_tsq_handler(struct sock *sk) 8526f458dfbSEric Dumazet { 8536f458dfbSEric Dumazet if ((1 << sk->sk_state) & 8546f458dfbSEric Dumazet (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING | 8556f458dfbSEric Dumazet TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) 8566f458dfbSEric Dumazet tcp_write_xmit(sk, tcp_current_mss(sk), 0, 0, GFP_ATOMIC); 8576f458dfbSEric Dumazet } 85846d3ceabSEric Dumazet /* 85946d3ceabSEric Dumazet * One tasklest per cpu tries to send more skbs. 86046d3ceabSEric Dumazet * We run in tasklet context but need to disable irqs when 86146d3ceabSEric Dumazet * transfering tsq->head because tcp_wfree() might 86246d3ceabSEric Dumazet * interrupt us (non NAPI drivers) 86346d3ceabSEric Dumazet */ 86446d3ceabSEric Dumazet static void tcp_tasklet_func(unsigned long data) 86546d3ceabSEric Dumazet { 86646d3ceabSEric Dumazet struct tsq_tasklet *tsq = (struct tsq_tasklet *)data; 86746d3ceabSEric Dumazet LIST_HEAD(list); 86846d3ceabSEric Dumazet unsigned long flags; 86946d3ceabSEric Dumazet struct list_head *q, *n; 87046d3ceabSEric Dumazet struct tcp_sock *tp; 87146d3ceabSEric Dumazet struct sock *sk; 87246d3ceabSEric Dumazet 87346d3ceabSEric Dumazet local_irq_save(flags); 87446d3ceabSEric Dumazet list_splice_init(&tsq->head, &list); 87546d3ceabSEric Dumazet local_irq_restore(flags); 87646d3ceabSEric Dumazet 87746d3ceabSEric Dumazet list_for_each_safe(q, n, &list) { 87846d3ceabSEric Dumazet tp = list_entry(q, struct tcp_sock, tsq_node); 87946d3ceabSEric Dumazet list_del(&tp->tsq_node); 88046d3ceabSEric Dumazet 88146d3ceabSEric Dumazet sk = (struct sock *)tp; 88246d3ceabSEric Dumazet bh_lock_sock(sk); 88346d3ceabSEric Dumazet 88446d3ceabSEric Dumazet if (!sock_owned_by_user(sk)) { 8856f458dfbSEric Dumazet tcp_tsq_handler(sk); 88646d3ceabSEric Dumazet } else { 88746d3ceabSEric Dumazet /* defer the work to tcp_release_cb() */ 8886f458dfbSEric Dumazet set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags); 88946d3ceabSEric Dumazet } 89046d3ceabSEric Dumazet bh_unlock_sock(sk); 89146d3ceabSEric Dumazet 89246d3ceabSEric Dumazet clear_bit(TSQ_QUEUED, &tp->tsq_flags); 89346d3ceabSEric Dumazet sk_free(sk); 89446d3ceabSEric Dumazet } 89546d3ceabSEric Dumazet } 89646d3ceabSEric Dumazet 8976f458dfbSEric Dumazet #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \ 8986f458dfbSEric Dumazet (1UL << TCP_WRITE_TIMER_DEFERRED) | \ 899563d34d0SEric Dumazet (1UL << TCP_DELACK_TIMER_DEFERRED) | \ 900563d34d0SEric Dumazet (1UL << TCP_MTU_REDUCED_DEFERRED)) 90146d3ceabSEric Dumazet /** 90246d3ceabSEric Dumazet * tcp_release_cb - tcp release_sock() callback 90346d3ceabSEric Dumazet * @sk: socket 90446d3ceabSEric Dumazet * 90546d3ceabSEric Dumazet * called from release_sock() to perform protocol dependent 90646d3ceabSEric Dumazet * actions before socket release. 90746d3ceabSEric Dumazet */ 90846d3ceabSEric Dumazet void tcp_release_cb(struct sock *sk) 90946d3ceabSEric Dumazet { 91046d3ceabSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 9116f458dfbSEric Dumazet unsigned long flags, nflags; 91246d3ceabSEric Dumazet 9136f458dfbSEric Dumazet /* perform an atomic operation only if at least one flag is set */ 9146f458dfbSEric Dumazet do { 9156f458dfbSEric Dumazet flags = tp->tsq_flags; 9166f458dfbSEric Dumazet if (!(flags & TCP_DEFERRED_ALL)) 9176f458dfbSEric Dumazet return; 9186f458dfbSEric Dumazet nflags = flags & ~TCP_DEFERRED_ALL; 9196f458dfbSEric Dumazet } while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags); 9206f458dfbSEric Dumazet 9216f458dfbSEric Dumazet if (flags & (1UL << TCP_TSQ_DEFERRED)) 9226f458dfbSEric Dumazet tcp_tsq_handler(sk); 9236f458dfbSEric Dumazet 924144d56e9SEric Dumazet if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) { 9256f458dfbSEric Dumazet tcp_write_timer_handler(sk); 926144d56e9SEric Dumazet __sock_put(sk); 927144d56e9SEric Dumazet } 928144d56e9SEric Dumazet if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) { 9296f458dfbSEric Dumazet tcp_delack_timer_handler(sk); 930144d56e9SEric Dumazet __sock_put(sk); 931144d56e9SEric Dumazet } 932144d56e9SEric Dumazet if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) { 933563d34d0SEric Dumazet sk->sk_prot->mtu_reduced(sk); 934144d56e9SEric Dumazet __sock_put(sk); 935144d56e9SEric Dumazet } 93646d3ceabSEric Dumazet } 93746d3ceabSEric Dumazet EXPORT_SYMBOL(tcp_release_cb); 93846d3ceabSEric Dumazet 93946d3ceabSEric Dumazet void __init tcp_tasklet_init(void) 94046d3ceabSEric Dumazet { 94146d3ceabSEric Dumazet int i; 94246d3ceabSEric Dumazet 94346d3ceabSEric Dumazet for_each_possible_cpu(i) { 94446d3ceabSEric Dumazet struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i); 94546d3ceabSEric Dumazet 94646d3ceabSEric Dumazet INIT_LIST_HEAD(&tsq->head); 94746d3ceabSEric Dumazet tasklet_init(&tsq->tasklet, 94846d3ceabSEric Dumazet tcp_tasklet_func, 94946d3ceabSEric Dumazet (unsigned long)tsq); 95046d3ceabSEric Dumazet } 95146d3ceabSEric Dumazet } 95246d3ceabSEric Dumazet 95346d3ceabSEric Dumazet /* 95446d3ceabSEric Dumazet * Write buffer destructor automatically called from kfree_skb. 95546d3ceabSEric Dumazet * We cant xmit new skbs from this context, as we might already 95646d3ceabSEric Dumazet * hold qdisc lock. 95746d3ceabSEric Dumazet */ 9588e7dfbc8SSilviu-Mihai Popescu static void tcp_wfree(struct sk_buff *skb) 95946d3ceabSEric Dumazet { 96046d3ceabSEric Dumazet struct sock *sk = skb->sk; 96146d3ceabSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 96246d3ceabSEric Dumazet 96346d3ceabSEric Dumazet if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) && 96446d3ceabSEric Dumazet !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) { 96546d3ceabSEric Dumazet unsigned long flags; 96646d3ceabSEric Dumazet struct tsq_tasklet *tsq; 96746d3ceabSEric Dumazet 96846d3ceabSEric Dumazet /* Keep a ref on socket. 96946d3ceabSEric Dumazet * This last ref will be released in tcp_tasklet_func() 97046d3ceabSEric Dumazet */ 97146d3ceabSEric Dumazet atomic_sub(skb->truesize - 1, &sk->sk_wmem_alloc); 97246d3ceabSEric Dumazet 97346d3ceabSEric Dumazet /* queue this socket to tasklet queue */ 97446d3ceabSEric Dumazet local_irq_save(flags); 97546d3ceabSEric Dumazet tsq = &__get_cpu_var(tsq_tasklet); 97646d3ceabSEric Dumazet list_add(&tp->tsq_node, &tsq->head); 97746d3ceabSEric Dumazet tasklet_schedule(&tsq->tasklet); 97846d3ceabSEric Dumazet local_irq_restore(flags); 97946d3ceabSEric Dumazet } else { 98046d3ceabSEric Dumazet sock_wfree(skb); 98146d3ceabSEric Dumazet } 98246d3ceabSEric Dumazet } 98346d3ceabSEric Dumazet 9841da177e4SLinus Torvalds /* This routine actually transmits TCP packets queued in by 9851da177e4SLinus Torvalds * tcp_do_sendmsg(). This is used by both the initial 9861da177e4SLinus Torvalds * transmission and possible later retransmissions. 9871da177e4SLinus Torvalds * All SKB's seen here are completely headerless. It is our 9881da177e4SLinus Torvalds * job to build the TCP header, and pass the packet down to 9891da177e4SLinus Torvalds * IP so it can do the same plus pass the packet off to the 9901da177e4SLinus Torvalds * device. 9911da177e4SLinus Torvalds * 9921da177e4SLinus Torvalds * We are working here with either a clone of the original 9931da177e4SLinus Torvalds * SKB, or a fresh unique copy made by the retransmit engine. 9941da177e4SLinus Torvalds */ 995056834d9SIlpo Järvinen static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, 996056834d9SIlpo Järvinen gfp_t gfp_mask) 9971da177e4SLinus Torvalds { 9986687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 999dfb4b9dcSDavid S. Miller struct inet_sock *inet; 1000dfb4b9dcSDavid S. Miller struct tcp_sock *tp; 1001dfb4b9dcSDavid S. Miller struct tcp_skb_cb *tcb; 100233ad798cSAdam Langley struct tcp_out_options opts; 100395c96174SEric Dumazet unsigned int tcp_options_size, tcp_header_size; 1004cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 10051da177e4SLinus Torvalds struct tcphdr *th; 10061da177e4SLinus Torvalds int err; 10071da177e4SLinus Torvalds 1008dfb4b9dcSDavid S. Miller BUG_ON(!skb || !tcp_skb_pcount(skb)); 1009dfb4b9dcSDavid S. Miller 1010dfb4b9dcSDavid S. Miller /* If congestion control is doing timestamping, we must 1011dfb4b9dcSDavid S. Miller * take such a timestamp before we potentially clone/copy. 1012dfb4b9dcSDavid S. Miller */ 1013164891aaSStephen Hemminger if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP) 1014dfb4b9dcSDavid S. Miller __net_timestamp(skb); 1015dfb4b9dcSDavid S. Miller 1016dfb4b9dcSDavid S. Miller if (likely(clone_it)) { 1017dfb4b9dcSDavid S. Miller if (unlikely(skb_cloned(skb))) 1018dfb4b9dcSDavid S. Miller skb = pskb_copy(skb, gfp_mask); 1019dfb4b9dcSDavid S. Miller else 1020dfb4b9dcSDavid S. Miller skb = skb_clone(skb, gfp_mask); 1021dfb4b9dcSDavid S. Miller if (unlikely(!skb)) 1022dfb4b9dcSDavid S. Miller return -ENOBUFS; 1023dfb4b9dcSDavid S. Miller } 1024dfb4b9dcSDavid S. Miller 1025dfb4b9dcSDavid S. Miller inet = inet_sk(sk); 1026dfb4b9dcSDavid S. Miller tp = tcp_sk(sk); 1027dfb4b9dcSDavid S. Miller tcb = TCP_SKB_CB(skb); 102833ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 10291da177e4SLinus Torvalds 10304de075e0SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) 103133ad798cSAdam Langley tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); 103233ad798cSAdam Langley else 103333ad798cSAdam Langley tcp_options_size = tcp_established_options(sk, skb, &opts, 103433ad798cSAdam Langley &md5); 103533ad798cSAdam Langley tcp_header_size = tcp_options_size + sizeof(struct tcphdr); 10361da177e4SLinus Torvalds 10373853b584STom Herbert if (tcp_packets_in_flight(tp) == 0) { 10386687e988SArnaldo Carvalho de Melo tcp_ca_event(sk, CA_EVENT_TX_START); 10393853b584STom Herbert skb->ooo_okay = 1; 10403853b584STom Herbert } else 10413853b584STom Herbert skb->ooo_okay = 0; 10421da177e4SLinus Torvalds 1043aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 1044aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 104546d3ceabSEric Dumazet 104646d3ceabSEric Dumazet skb_orphan(skb); 104746d3ceabSEric Dumazet skb->sk = sk; 104846d3ceabSEric Dumazet skb->destructor = (sysctl_tcp_limit_output_bytes > 0) ? 104946d3ceabSEric Dumazet tcp_wfree : sock_wfree; 105046d3ceabSEric Dumazet atomic_add(skb->truesize, &sk->sk_wmem_alloc); 10511da177e4SLinus Torvalds 10521da177e4SLinus Torvalds /* Build TCP header and checksum it. */ 1053aa8223c7SArnaldo Carvalho de Melo th = tcp_hdr(skb); 1054c720c7e8SEric Dumazet th->source = inet->inet_sport; 1055c720c7e8SEric Dumazet th->dest = inet->inet_dport; 10561da177e4SLinus Torvalds th->seq = htonl(tcb->seq); 10571da177e4SLinus Torvalds th->ack_seq = htonl(tp->rcv_nxt); 1058df7a3b07SAl Viro *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | 10594de075e0SEric Dumazet tcb->tcp_flags); 1060dfb4b9dcSDavid S. Miller 10614de075e0SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) { 10621da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments 10631da177e4SLinus Torvalds * is never scaled. 10641da177e4SLinus Torvalds */ 1065600ff0c2SIlpo Järvinen th->window = htons(min(tp->rcv_wnd, 65535U)); 10661da177e4SLinus Torvalds } else { 10671da177e4SLinus Torvalds th->window = htons(tcp_select_window(sk)); 10681da177e4SLinus Torvalds } 10691da177e4SLinus Torvalds th->check = 0; 10701da177e4SLinus Torvalds th->urg_ptr = 0; 10711da177e4SLinus Torvalds 107233f5f57eSIlpo Järvinen /* The urg_mode check is necessary during a below snd_una win probe */ 10737691367dSHerbert Xu if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { 10747691367dSHerbert Xu if (before(tp->snd_up, tcb->seq + 0x10000)) { 10751da177e4SLinus Torvalds th->urg_ptr = htons(tp->snd_up - tcb->seq); 10761da177e4SLinus Torvalds th->urg = 1; 10777691367dSHerbert Xu } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { 10780eae88f3SEric Dumazet th->urg_ptr = htons(0xFFFF); 10797691367dSHerbert Xu th->urg = 1; 10807691367dSHerbert Xu } 10811da177e4SLinus Torvalds } 10821da177e4SLinus Torvalds 1083bd0388aeSWilliam Allen Simpson tcp_options_write((__be32 *)(th + 1), tp, &opts); 10844de075e0SEric Dumazet if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0)) 10859e412ba7SIlpo Järvinen TCP_ECN_send(sk, skb, tcp_header_size); 1086dfb4b9dcSDavid S. Miller 1087cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 1088cfb6eeb4SYOSHIFUJI Hideaki /* Calculate the MD5 hash, as we have all we need now */ 1089cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 1090a465419bSEric Dumazet sk_nocaps_add(sk, NETIF_F_GSO_MASK); 1091bd0388aeSWilliam Allen Simpson tp->af_specific->calc_md5_hash(opts.hash_location, 109249a72dfbSAdam Langley md5, sk, NULL, skb); 1093cfb6eeb4SYOSHIFUJI Hideaki } 1094cfb6eeb4SYOSHIFUJI Hideaki #endif 1095cfb6eeb4SYOSHIFUJI Hideaki 1096bb296246SHerbert Xu icsk->icsk_af_ops->send_check(sk, skb); 10971da177e4SLinus Torvalds 10984de075e0SEric Dumazet if (likely(tcb->tcp_flags & TCPHDR_ACK)) 1099fc6415bcSDavid S. Miller tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); 11001da177e4SLinus Torvalds 11011da177e4SLinus Torvalds if (skb->len != tcp_header_size) 1102cf533ea5SEric Dumazet tcp_event_data_sent(tp, sk); 11031da177e4SLinus Torvalds 1104bd37a088SWei Yongjun if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) 1105aa2ea058STom Herbert TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, 1106aa2ea058STom Herbert tcp_skb_pcount(skb)); 11071da177e4SLinus Torvalds 1108d9d8da80SDavid S. Miller err = icsk->icsk_af_ops->queue_xmit(skb, &inet->cork.fl); 110983de47cdSHua Zhong if (likely(err <= 0)) 11101da177e4SLinus Torvalds return err; 11111da177e4SLinus Torvalds 11123cfe3baaSIlpo Järvinen tcp_enter_cwr(sk, 1); 11131da177e4SLinus Torvalds 1114b9df3cb8SGerrit Renker return net_xmit_eval(err); 11151da177e4SLinus Torvalds } 11161da177e4SLinus Torvalds 111767edfef7SAndi Kleen /* This routine just queues the buffer for sending. 11181da177e4SLinus Torvalds * 11191da177e4SLinus Torvalds * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, 11201da177e4SLinus Torvalds * otherwise socket can stall. 11211da177e4SLinus Torvalds */ 11221da177e4SLinus Torvalds static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) 11231da177e4SLinus Torvalds { 11241da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 11251da177e4SLinus Torvalds 11261da177e4SLinus Torvalds /* Advance write_seq and place onto the write_queue. */ 11271da177e4SLinus Torvalds tp->write_seq = TCP_SKB_CB(skb)->end_seq; 11281da177e4SLinus Torvalds skb_header_release(skb); 1129fe067e8aSDavid S. Miller tcp_add_write_queue_tail(sk, skb); 11303ab224beSHideo Aoki sk->sk_wmem_queued += skb->truesize; 11313ab224beSHideo Aoki sk_mem_charge(sk, skb->truesize); 11321da177e4SLinus Torvalds } 11331da177e4SLinus Torvalds 113467edfef7SAndi Kleen /* Initialize TSO segments for a packet. */ 1135cf533ea5SEric Dumazet static void tcp_set_skb_tso_segs(const struct sock *sk, struct sk_buff *skb, 1136056834d9SIlpo Järvinen unsigned int mss_now) 1137f6302d1dSDavid S. Miller { 11388e5b9ddaSHerbert Xu if (skb->len <= mss_now || !sk_can_gso(sk) || 11398e5b9ddaSHerbert Xu skb->ip_summed == CHECKSUM_NONE) { 1140f6302d1dSDavid S. Miller /* Avoid the costly divide in the normal 1141f6302d1dSDavid S. Miller * non-TSO case. 1142f6302d1dSDavid S. Miller */ 11437967168cSHerbert Xu skb_shinfo(skb)->gso_segs = 1; 11447967168cSHerbert Xu skb_shinfo(skb)->gso_size = 0; 1145c9af6db4SPravin B Shelar skb_shinfo(skb)->gso_type = 0; 1146f6302d1dSDavid S. Miller } else { 1147356f89e1SIlpo Järvinen skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now); 11487967168cSHerbert Xu skb_shinfo(skb)->gso_size = mss_now; 1149c9af6db4SPravin B Shelar skb_shinfo(skb)->gso_type = sk->sk_gso_type; 11501da177e4SLinus Torvalds } 11511da177e4SLinus Torvalds } 11521da177e4SLinus Torvalds 115391fed7a1SIlpo Järvinen /* When a modification to fackets out becomes necessary, we need to check 115468f8353bSIlpo Järvinen * skb is counted to fackets_out or not. 115591fed7a1SIlpo Järvinen */ 1156cf533ea5SEric Dumazet static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb, 115791fed7a1SIlpo Järvinen int decr) 115891fed7a1SIlpo Järvinen { 1159a47e5a98SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1160a47e5a98SIlpo Järvinen 1161dc86967bSIlpo Järvinen if (!tp->sacked_out || tcp_is_reno(tp)) 116291fed7a1SIlpo Järvinen return; 116391fed7a1SIlpo Järvinen 11646859d494SIlpo Järvinen if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq)) 116591fed7a1SIlpo Järvinen tp->fackets_out -= decr; 116691fed7a1SIlpo Järvinen } 116791fed7a1SIlpo Järvinen 1168797108d1SIlpo Järvinen /* Pcount in the middle of the write queue got changed, we need to do various 1169797108d1SIlpo Järvinen * tweaks to fix counters 1170797108d1SIlpo Järvinen */ 1171cf533ea5SEric Dumazet static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) 1172797108d1SIlpo Järvinen { 1173797108d1SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1174797108d1SIlpo Järvinen 1175797108d1SIlpo Järvinen tp->packets_out -= decr; 1176797108d1SIlpo Järvinen 1177797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 1178797108d1SIlpo Järvinen tp->sacked_out -= decr; 1179797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 1180797108d1SIlpo Järvinen tp->retrans_out -= decr; 1181797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 1182797108d1SIlpo Järvinen tp->lost_out -= decr; 1183797108d1SIlpo Järvinen 1184797108d1SIlpo Järvinen /* Reno case is special. Sigh... */ 1185797108d1SIlpo Järvinen if (tcp_is_reno(tp) && decr > 0) 1186797108d1SIlpo Järvinen tp->sacked_out -= min_t(u32, tp->sacked_out, decr); 1187797108d1SIlpo Järvinen 1188797108d1SIlpo Järvinen tcp_adjust_fackets_out(sk, skb, decr); 1189797108d1SIlpo Järvinen 1190797108d1SIlpo Järvinen if (tp->lost_skb_hint && 1191797108d1SIlpo Järvinen before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && 119252cf3cc8SIlpo Järvinen (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))) 1193797108d1SIlpo Järvinen tp->lost_cnt_hint -= decr; 1194797108d1SIlpo Järvinen 1195797108d1SIlpo Järvinen tcp_verify_left_out(tp); 1196797108d1SIlpo Järvinen } 1197797108d1SIlpo Järvinen 11981da177e4SLinus Torvalds /* Function to create two new TCP segments. Shrinks the given segment 11991da177e4SLinus Torvalds * to the specified size and appends a new segment with the rest of the 12001da177e4SLinus Torvalds * packet to the list. This won't be called frequently, I hope. 12011da177e4SLinus Torvalds * Remember, these are still headerless SKBs at this point. 12021da177e4SLinus Torvalds */ 1203056834d9SIlpo Järvinen int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, 1204056834d9SIlpo Järvinen unsigned int mss_now) 12051da177e4SLinus Torvalds { 12061da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 12071da177e4SLinus Torvalds struct sk_buff *buff; 12086475be16SDavid S. Miller int nsize, old_factor; 1209b60b49eaSHerbert Xu int nlen; 12109ce01461SIlpo Järvinen u8 flags; 12111da177e4SLinus Torvalds 12122fceec13SIlpo Järvinen if (WARN_ON(len > skb->len)) 12132fceec13SIlpo Järvinen return -EINVAL; 12146a438bbeSStephen Hemminger 12151da177e4SLinus Torvalds nsize = skb_headlen(skb) - len; 12161da177e4SLinus Torvalds if (nsize < 0) 12171da177e4SLinus Torvalds nsize = 0; 12181da177e4SLinus Torvalds 12191da177e4SLinus Torvalds if (skb_cloned(skb) && 12201da177e4SLinus Torvalds skb_is_nonlinear(skb) && 12211da177e4SLinus Torvalds pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) 12221da177e4SLinus Torvalds return -ENOMEM; 12231da177e4SLinus Torvalds 12241da177e4SLinus Torvalds /* Get a new skb... force flag on. */ 12251da177e4SLinus Torvalds buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC); 12261da177e4SLinus Torvalds if (buff == NULL) 12271da177e4SLinus Torvalds return -ENOMEM; /* We'll just try again later. */ 1228ef5cb973SHerbert Xu 12293ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 12303ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1231b60b49eaSHerbert Xu nlen = skb->len - len - nsize; 1232b60b49eaSHerbert Xu buff->truesize += nlen; 1233b60b49eaSHerbert Xu skb->truesize -= nlen; 12341da177e4SLinus Torvalds 12351da177e4SLinus Torvalds /* Correct the sequence numbers. */ 12361da177e4SLinus Torvalds TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 12371da177e4SLinus Torvalds TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 12381da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 12391da177e4SLinus Torvalds 12401da177e4SLinus Torvalds /* PSH and FIN should only be set in the second packet. */ 12414de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 12424de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 12434de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1244e14c3cafSHerbert Xu TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; 12451da177e4SLinus Torvalds 124684fa7933SPatrick McHardy if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { 12471da177e4SLinus Torvalds /* Copy and checksum data tail into the new buffer. */ 1248056834d9SIlpo Järvinen buff->csum = csum_partial_copy_nocheck(skb->data + len, 1249056834d9SIlpo Järvinen skb_put(buff, nsize), 12501da177e4SLinus Torvalds nsize, 0); 12511da177e4SLinus Torvalds 12521da177e4SLinus Torvalds skb_trim(skb, len); 12531da177e4SLinus Torvalds 12541da177e4SLinus Torvalds skb->csum = csum_block_sub(skb->csum, buff->csum, len); 12551da177e4SLinus Torvalds } else { 125684fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 12571da177e4SLinus Torvalds skb_split(skb, buff, len); 12581da177e4SLinus Torvalds } 12591da177e4SLinus Torvalds 12601da177e4SLinus Torvalds buff->ip_summed = skb->ip_summed; 12611da177e4SLinus Torvalds 12621da177e4SLinus Torvalds /* Looks stupid, but our code really uses when of 12631da177e4SLinus Torvalds * skbs, which it never sent before. --ANK 12641da177e4SLinus Torvalds */ 12651da177e4SLinus Torvalds TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when; 1266a61bbcf2SPatrick McHardy buff->tstamp = skb->tstamp; 12671da177e4SLinus Torvalds 12686475be16SDavid S. Miller old_factor = tcp_skb_pcount(skb); 12696475be16SDavid S. Miller 12701da177e4SLinus Torvalds /* Fix up tso_factor for both original and new SKB. */ 1271846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1272846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, buff, mss_now); 12731da177e4SLinus Torvalds 12746475be16SDavid S. Miller /* If this packet has been sent out already, we must 12756475be16SDavid S. Miller * adjust the various packet counters. 12766475be16SDavid S. Miller */ 1277cf0b450cSHerbert Xu if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 12786475be16SDavid S. Miller int diff = old_factor - tcp_skb_pcount(skb) - 12796475be16SDavid S. Miller tcp_skb_pcount(buff); 12801da177e4SLinus Torvalds 1281797108d1SIlpo Järvinen if (diff) 1282797108d1SIlpo Järvinen tcp_adjust_pcount(sk, skb, diff); 12831da177e4SLinus Torvalds } 12841da177e4SLinus Torvalds 12851da177e4SLinus Torvalds /* Link BUFF into the send queue. */ 1286f44b5271SDavid S. Miller skb_header_release(buff); 1287fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 12881da177e4SLinus Torvalds 12891da177e4SLinus Torvalds return 0; 12901da177e4SLinus Torvalds } 12911da177e4SLinus Torvalds 12921da177e4SLinus Torvalds /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c 12931da177e4SLinus Torvalds * eventually). The difference is that pulled data not copied, but 12941da177e4SLinus Torvalds * immediately discarded. 12951da177e4SLinus Torvalds */ 1296f2911969SHerbert Xu ~{PmVHI~} static void __pskb_trim_head(struct sk_buff *skb, int len) 12971da177e4SLinus Torvalds { 12981da177e4SLinus Torvalds int i, k, eat; 12991da177e4SLinus Torvalds 13004fa48bf3SEric Dumazet eat = min_t(int, len, skb_headlen(skb)); 13014fa48bf3SEric Dumazet if (eat) { 13024fa48bf3SEric Dumazet __skb_pull(skb, eat); 130322b4a4f2SEric Dumazet skb->avail_size -= eat; 13044fa48bf3SEric Dumazet len -= eat; 13054fa48bf3SEric Dumazet if (!len) 13064fa48bf3SEric Dumazet return; 13074fa48bf3SEric Dumazet } 13081da177e4SLinus Torvalds eat = len; 13091da177e4SLinus Torvalds k = 0; 13101da177e4SLinus Torvalds for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) { 13119e903e08SEric Dumazet int size = skb_frag_size(&skb_shinfo(skb)->frags[i]); 13129e903e08SEric Dumazet 13139e903e08SEric Dumazet if (size <= eat) { 1314aff65da0SIan Campbell skb_frag_unref(skb, i); 13159e903e08SEric Dumazet eat -= size; 13161da177e4SLinus Torvalds } else { 13171da177e4SLinus Torvalds skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i]; 13181da177e4SLinus Torvalds if (eat) { 13191da177e4SLinus Torvalds skb_shinfo(skb)->frags[k].page_offset += eat; 13209e903e08SEric Dumazet skb_frag_size_sub(&skb_shinfo(skb)->frags[k], eat); 13211da177e4SLinus Torvalds eat = 0; 13221da177e4SLinus Torvalds } 13231da177e4SLinus Torvalds k++; 13241da177e4SLinus Torvalds } 13251da177e4SLinus Torvalds } 13261da177e4SLinus Torvalds skb_shinfo(skb)->nr_frags = k; 13271da177e4SLinus Torvalds 132827a884dcSArnaldo Carvalho de Melo skb_reset_tail_pointer(skb); 13291da177e4SLinus Torvalds skb->data_len -= len; 13301da177e4SLinus Torvalds skb->len = skb->data_len; 13311da177e4SLinus Torvalds } 13321da177e4SLinus Torvalds 133367edfef7SAndi Kleen /* Remove acked data from a packet in the transmit queue. */ 13341da177e4SLinus Torvalds int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 13351da177e4SLinus Torvalds { 133614bbd6a5SPravin B Shelar if (skb_unclone(skb, GFP_ATOMIC)) 13371da177e4SLinus Torvalds return -ENOMEM; 13381da177e4SLinus Torvalds 13394fa48bf3SEric Dumazet __pskb_trim_head(skb, len); 13401da177e4SLinus Torvalds 13411da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq += len; 134284fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 13431da177e4SLinus Torvalds 13441da177e4SLinus Torvalds skb->truesize -= len; 13451da177e4SLinus Torvalds sk->sk_wmem_queued -= len; 13463ab224beSHideo Aoki sk_mem_uncharge(sk, len); 13471da177e4SLinus Torvalds sock_set_flag(sk, SOCK_QUEUE_SHRUNK); 13481da177e4SLinus Torvalds 13495b35e1e6SNeal Cardwell /* Any change of skb->len requires recalculation of tso factor. */ 13501da177e4SLinus Torvalds if (tcp_skb_pcount(skb) > 1) 13515b35e1e6SNeal Cardwell tcp_set_skb_tso_segs(sk, skb, tcp_skb_mss(skb)); 13521da177e4SLinus Torvalds 13531da177e4SLinus Torvalds return 0; 13541da177e4SLinus Torvalds } 13551da177e4SLinus Torvalds 13561b63edd6SYuchung Cheng /* Calculate MSS not accounting any TCP options. */ 13571b63edd6SYuchung Cheng static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu) 13585d424d5aSJohn Heffner { 1359cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1360cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 13615d424d5aSJohn Heffner int mss_now; 13625d424d5aSJohn Heffner 13635d424d5aSJohn Heffner /* Calculate base mss without TCP options: 13645d424d5aSJohn Heffner It is MMS_S - sizeof(tcphdr) of rfc1122 13655d424d5aSJohn Heffner */ 13665d424d5aSJohn Heffner mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 13675d424d5aSJohn Heffner 136867469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 136967469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 137067469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 137167469601SEric Dumazet 137267469601SEric Dumazet if (dst && dst_allfrag(dst)) 137367469601SEric Dumazet mss_now -= icsk->icsk_af_ops->net_frag_header_len; 137467469601SEric Dumazet } 137567469601SEric Dumazet 13765d424d5aSJohn Heffner /* Clamp it (mss_clamp does not include tcp options) */ 13775d424d5aSJohn Heffner if (mss_now > tp->rx_opt.mss_clamp) 13785d424d5aSJohn Heffner mss_now = tp->rx_opt.mss_clamp; 13795d424d5aSJohn Heffner 13805d424d5aSJohn Heffner /* Now subtract optional transport overhead */ 13815d424d5aSJohn Heffner mss_now -= icsk->icsk_ext_hdr_len; 13825d424d5aSJohn Heffner 13835d424d5aSJohn Heffner /* Then reserve room for full set of TCP options and 8 bytes of data */ 13845d424d5aSJohn Heffner if (mss_now < 48) 13855d424d5aSJohn Heffner mss_now = 48; 13865d424d5aSJohn Heffner return mss_now; 13875d424d5aSJohn Heffner } 13885d424d5aSJohn Heffner 13891b63edd6SYuchung Cheng /* Calculate MSS. Not accounting for SACKs here. */ 13901b63edd6SYuchung Cheng int tcp_mtu_to_mss(struct sock *sk, int pmtu) 13911b63edd6SYuchung Cheng { 13921b63edd6SYuchung Cheng /* Subtract TCP options size, not including SACKs */ 13931b63edd6SYuchung Cheng return __tcp_mtu_to_mss(sk, pmtu) - 13941b63edd6SYuchung Cheng (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr)); 13951b63edd6SYuchung Cheng } 13961b63edd6SYuchung Cheng 13975d424d5aSJohn Heffner /* Inverse of above */ 139867469601SEric Dumazet int tcp_mss_to_mtu(struct sock *sk, int mss) 13995d424d5aSJohn Heffner { 1400cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1401cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 14025d424d5aSJohn Heffner int mtu; 14035d424d5aSJohn Heffner 14045d424d5aSJohn Heffner mtu = mss + 14055d424d5aSJohn Heffner tp->tcp_header_len + 14065d424d5aSJohn Heffner icsk->icsk_ext_hdr_len + 14075d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 14085d424d5aSJohn Heffner 140967469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 141067469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 141167469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 141267469601SEric Dumazet 141367469601SEric Dumazet if (dst && dst_allfrag(dst)) 141467469601SEric Dumazet mtu += icsk->icsk_af_ops->net_frag_header_len; 141567469601SEric Dumazet } 14165d424d5aSJohn Heffner return mtu; 14175d424d5aSJohn Heffner } 14185d424d5aSJohn Heffner 141967edfef7SAndi Kleen /* MTU probing init per socket */ 14205d424d5aSJohn Heffner void tcp_mtup_init(struct sock *sk) 14215d424d5aSJohn Heffner { 14225d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 14235d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 14245d424d5aSJohn Heffner 14255d424d5aSJohn Heffner icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1; 14265d424d5aSJohn Heffner icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 14275d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 14285d424d5aSJohn Heffner icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss); 14295d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 14305d424d5aSJohn Heffner } 14314bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_mtup_init); 14325d424d5aSJohn Heffner 14331da177e4SLinus Torvalds /* This function synchronize snd mss to current pmtu/exthdr set. 14341da177e4SLinus Torvalds 14351da177e4SLinus Torvalds tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 14361da177e4SLinus Torvalds for TCP options, but includes only bare TCP header. 14371da177e4SLinus Torvalds 14381da177e4SLinus Torvalds tp->rx_opt.mss_clamp is mss negotiated at connection setup. 1439caa20d9aSStephen Hemminger It is minimum of user_mss and mss received with SYN. 14401da177e4SLinus Torvalds It also does not include TCP options. 14411da177e4SLinus Torvalds 1442d83d8461SArnaldo Carvalho de Melo inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. 14431da177e4SLinus Torvalds 14441da177e4SLinus Torvalds tp->mss_cache is current effective sending mss, including 14451da177e4SLinus Torvalds all tcp options except for SACKs. It is evaluated, 14461da177e4SLinus Torvalds taking into account current pmtu, but never exceeds 14471da177e4SLinus Torvalds tp->rx_opt.mss_clamp. 14481da177e4SLinus Torvalds 14491da177e4SLinus Torvalds NOTE1. rfc1122 clearly states that advertised MSS 14501da177e4SLinus Torvalds DOES NOT include either tcp or ip options. 14511da177e4SLinus Torvalds 1452d83d8461SArnaldo Carvalho de Melo NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache 1453d83d8461SArnaldo Carvalho de Melo are READ ONLY outside this function. --ANK (980731) 14541da177e4SLinus Torvalds */ 14551da177e4SLinus Torvalds unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 14561da177e4SLinus Torvalds { 14571da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1458d83d8461SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 14595d424d5aSJohn Heffner int mss_now; 14601da177e4SLinus Torvalds 14615d424d5aSJohn Heffner if (icsk->icsk_mtup.search_high > pmtu) 14625d424d5aSJohn Heffner icsk->icsk_mtup.search_high = pmtu; 14631da177e4SLinus Torvalds 14645d424d5aSJohn Heffner mss_now = tcp_mtu_to_mss(sk, pmtu); 1465409d22b4SIlpo Järvinen mss_now = tcp_bound_to_half_wnd(tp, mss_now); 14661da177e4SLinus Torvalds 14671da177e4SLinus Torvalds /* And store cached results */ 1468d83d8461SArnaldo Carvalho de Melo icsk->icsk_pmtu_cookie = pmtu; 14695d424d5aSJohn Heffner if (icsk->icsk_mtup.enabled) 14705d424d5aSJohn Heffner mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); 1471c1b4a7e6SDavid S. Miller tp->mss_cache = mss_now; 14721da177e4SLinus Torvalds 14731da177e4SLinus Torvalds return mss_now; 14741da177e4SLinus Torvalds } 14754bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_sync_mss); 14761da177e4SLinus Torvalds 14771da177e4SLinus Torvalds /* Compute the current effective MSS, taking SACKs and IP options, 14781da177e4SLinus Torvalds * and even PMTU discovery events into account. 14791da177e4SLinus Torvalds */ 14800c54b85fSIlpo Järvinen unsigned int tcp_current_mss(struct sock *sk) 14811da177e4SLinus Torvalds { 1482cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1483cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1484c1b4a7e6SDavid S. Miller u32 mss_now; 148595c96174SEric Dumazet unsigned int header_len; 148633ad798cSAdam Langley struct tcp_out_options opts; 148733ad798cSAdam Langley struct tcp_md5sig_key *md5; 14881da177e4SLinus Torvalds 1489c1b4a7e6SDavid S. Miller mss_now = tp->mss_cache; 1490c1b4a7e6SDavid S. Miller 14911da177e4SLinus Torvalds if (dst) { 14921da177e4SLinus Torvalds u32 mtu = dst_mtu(dst); 1493d83d8461SArnaldo Carvalho de Melo if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 14941da177e4SLinus Torvalds mss_now = tcp_sync_mss(sk, mtu); 14951da177e4SLinus Torvalds } 14961da177e4SLinus Torvalds 149733ad798cSAdam Langley header_len = tcp_established_options(sk, NULL, &opts, &md5) + 149833ad798cSAdam Langley sizeof(struct tcphdr); 149933ad798cSAdam Langley /* The mss_cache is sized based on tp->tcp_header_len, which assumes 150033ad798cSAdam Langley * some common options. If this is an odd packet (because we have SACK 150133ad798cSAdam Langley * blocks etc) then our calculated header_len will be different, and 150233ad798cSAdam Langley * we have to adjust mss_now correspondingly */ 150333ad798cSAdam Langley if (header_len != tp->tcp_header_len) { 150433ad798cSAdam Langley int delta = (int) header_len - tp->tcp_header_len; 150533ad798cSAdam Langley mss_now -= delta; 150633ad798cSAdam Langley } 1507cfb6eeb4SYOSHIFUJI Hideaki 15081da177e4SLinus Torvalds return mss_now; 15091da177e4SLinus Torvalds } 15101da177e4SLinus Torvalds 1511a762a980SDavid S. Miller /* Congestion window validation. (RFC2861) */ 15129e412ba7SIlpo Järvinen static void tcp_cwnd_validate(struct sock *sk) 1513a762a980SDavid S. Miller { 15149e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1515a762a980SDavid S. Miller 1516d436d686SIlpo Järvinen if (tp->packets_out >= tp->snd_cwnd) { 1517a762a980SDavid S. Miller /* Network is feed fully. */ 1518a762a980SDavid S. Miller tp->snd_cwnd_used = 0; 1519a762a980SDavid S. Miller tp->snd_cwnd_stamp = tcp_time_stamp; 1520a762a980SDavid S. Miller } else { 1521a762a980SDavid S. Miller /* Network starves. */ 1522a762a980SDavid S. Miller if (tp->packets_out > tp->snd_cwnd_used) 1523a762a980SDavid S. Miller tp->snd_cwnd_used = tp->packets_out; 1524a762a980SDavid S. Miller 152515d33c07SDavid S. Miller if (sysctl_tcp_slow_start_after_idle && 152615d33c07SDavid S. Miller (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto) 1527a762a980SDavid S. Miller tcp_cwnd_application_limited(sk); 1528a762a980SDavid S. Miller } 1529a762a980SDavid S. Miller } 1530a762a980SDavid S. Miller 15310e3a4803SIlpo Järvinen /* Returns the portion of skb which can be sent right away without 15320e3a4803SIlpo Järvinen * introducing MSS oddities to segment boundaries. In rare cases where 15330e3a4803SIlpo Järvinen * mss_now != mss_cache, we will request caller to create a small skb 15340e3a4803SIlpo Järvinen * per input skb which could be mostly avoided here (if desired). 15355ea3a748SIlpo Järvinen * 15365ea3a748SIlpo Järvinen * We explicitly want to create a request for splitting write queue tail 15375ea3a748SIlpo Järvinen * to a small skb for Nagle purposes while avoiding unnecessary modulos, 15385ea3a748SIlpo Järvinen * thus all the complexity (cwnd_len is always MSS multiple which we 15395ea3a748SIlpo Järvinen * return whenever allowed by the other factors). Basically we need the 15405ea3a748SIlpo Järvinen * modulo only when the receiver window alone is the limiting factor or 15415ea3a748SIlpo Järvinen * when we would be allowed to send the split-due-to-Nagle skb fully. 15420e3a4803SIlpo Järvinen */ 1543cf533ea5SEric Dumazet static unsigned int tcp_mss_split_point(const struct sock *sk, const struct sk_buff *skb, 15441485348dSBen Hutchings unsigned int mss_now, unsigned int max_segs) 1545c1b4a7e6SDavid S. Miller { 1546cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 15471485348dSBen Hutchings u32 needed, window, max_len; 1548c1b4a7e6SDavid S. Miller 154990840defSIlpo Järvinen window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 15501485348dSBen Hutchings max_len = mss_now * max_segs; 15510e3a4803SIlpo Järvinen 15521485348dSBen Hutchings if (likely(max_len <= window && skb != tcp_write_queue_tail(sk))) 15531485348dSBen Hutchings return max_len; 15540e3a4803SIlpo Järvinen 15555ea3a748SIlpo Järvinen needed = min(skb->len, window); 15565ea3a748SIlpo Järvinen 15571485348dSBen Hutchings if (max_len <= needed) 15581485348dSBen Hutchings return max_len; 15590e3a4803SIlpo Järvinen 15600e3a4803SIlpo Järvinen return needed - needed % mss_now; 1561c1b4a7e6SDavid S. Miller } 1562c1b4a7e6SDavid S. Miller 1563c1b4a7e6SDavid S. Miller /* Can at least one segment of SKB be sent right now, according to the 1564c1b4a7e6SDavid S. Miller * congestion window rules? If so, return how many segments are allowed. 1565c1b4a7e6SDavid S. Miller */ 1566cf533ea5SEric Dumazet static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, 1567cf533ea5SEric Dumazet const struct sk_buff *skb) 1568c1b4a7e6SDavid S. Miller { 1569c1b4a7e6SDavid S. Miller u32 in_flight, cwnd; 1570c1b4a7e6SDavid S. Miller 1571c1b4a7e6SDavid S. Miller /* Don't be strict about the congestion window for the final FIN. */ 15724de075e0SEric Dumazet if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && 15734de075e0SEric Dumazet tcp_skb_pcount(skb) == 1) 1574c1b4a7e6SDavid S. Miller return 1; 1575c1b4a7e6SDavid S. Miller 1576c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1577c1b4a7e6SDavid S. Miller cwnd = tp->snd_cwnd; 1578c1b4a7e6SDavid S. Miller if (in_flight < cwnd) 1579c1b4a7e6SDavid S. Miller return (cwnd - in_flight); 1580c1b4a7e6SDavid S. Miller 1581c1b4a7e6SDavid S. Miller return 0; 1582c1b4a7e6SDavid S. Miller } 1583c1b4a7e6SDavid S. Miller 1584b595076aSUwe Kleine-König /* Initialize TSO state of a skb. 158567edfef7SAndi Kleen * This must be invoked the first time we consider transmitting 1586c1b4a7e6SDavid S. Miller * SKB onto the wire. 1587c1b4a7e6SDavid S. Miller */ 1588cf533ea5SEric Dumazet static int tcp_init_tso_segs(const struct sock *sk, struct sk_buff *skb, 1589056834d9SIlpo Järvinen unsigned int mss_now) 1590c1b4a7e6SDavid S. Miller { 1591c1b4a7e6SDavid S. Miller int tso_segs = tcp_skb_pcount(skb); 1592c1b4a7e6SDavid S. Miller 1593f8269a49SIlpo Järvinen if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { 1594846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1595c1b4a7e6SDavid S. Miller tso_segs = tcp_skb_pcount(skb); 1596c1b4a7e6SDavid S. Miller } 1597c1b4a7e6SDavid S. Miller return tso_segs; 1598c1b4a7e6SDavid S. Miller } 1599c1b4a7e6SDavid S. Miller 160067edfef7SAndi Kleen /* Minshall's variant of the Nagle send check. */ 1601a2a385d6SEric Dumazet static inline bool tcp_minshall_check(const struct tcp_sock *tp) 1602c1b4a7e6SDavid S. Miller { 1603c1b4a7e6SDavid S. Miller return after(tp->snd_sml, tp->snd_una) && 1604c1b4a7e6SDavid S. Miller !after(tp->snd_sml, tp->snd_nxt); 1605c1b4a7e6SDavid S. Miller } 1606c1b4a7e6SDavid S. Miller 1607a2a385d6SEric Dumazet /* Return false, if packet can be sent now without violation Nagle's rules: 1608c1b4a7e6SDavid S. Miller * 1. It is full sized. 1609c1b4a7e6SDavid S. Miller * 2. Or it contains FIN. (already checked by caller) 16106d67e9beSFeng King * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. 1611c1b4a7e6SDavid S. Miller * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 1612c1b4a7e6SDavid S. Miller * With Minshall's modification: all sent small packets are ACKed. 1613c1b4a7e6SDavid S. Miller */ 1614a2a385d6SEric Dumazet static inline bool tcp_nagle_check(const struct tcp_sock *tp, 1615c1b4a7e6SDavid S. Miller const struct sk_buff *skb, 161695c96174SEric Dumazet unsigned int mss_now, int nonagle) 1617c1b4a7e6SDavid S. Miller { 1618a02cec21SEric Dumazet return skb->len < mss_now && 1619c1b4a7e6SDavid S. Miller ((nonagle & TCP_NAGLE_CORK) || 1620a02cec21SEric Dumazet (!nonagle && tp->packets_out && tcp_minshall_check(tp))); 1621c1b4a7e6SDavid S. Miller } 1622c1b4a7e6SDavid S. Miller 1623a2a385d6SEric Dumazet /* Return true if the Nagle test allows this packet to be 1624c1b4a7e6SDavid S. Miller * sent now. 1625c1b4a7e6SDavid S. Miller */ 1626a2a385d6SEric Dumazet static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, 1627c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1628c1b4a7e6SDavid S. Miller { 1629c1b4a7e6SDavid S. Miller /* Nagle rule does not apply to frames, which sit in the middle of the 1630c1b4a7e6SDavid S. Miller * write_queue (they have no chances to get new data). 1631c1b4a7e6SDavid S. Miller * 1632c1b4a7e6SDavid S. Miller * This is implemented in the callers, where they modify the 'nonagle' 1633c1b4a7e6SDavid S. Miller * argument based upon the location of SKB in the send queue. 1634c1b4a7e6SDavid S. Miller */ 1635c1b4a7e6SDavid S. Miller if (nonagle & TCP_NAGLE_PUSH) 1636a2a385d6SEric Dumazet return true; 1637c1b4a7e6SDavid S. Miller 1638d551e454SIlpo Järvinen /* Don't use the nagle rule for urgent data (or for the final FIN). 1639d551e454SIlpo Järvinen * Nagle can be ignored during F-RTO too (see RFC4138). 1640d551e454SIlpo Järvinen */ 164133f5f57eSIlpo Järvinen if (tcp_urg_mode(tp) || (tp->frto_counter == 2) || 16424de075e0SEric Dumazet (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) 1643a2a385d6SEric Dumazet return true; 1644c1b4a7e6SDavid S. Miller 1645c1b4a7e6SDavid S. Miller if (!tcp_nagle_check(tp, skb, cur_mss, nonagle)) 1646a2a385d6SEric Dumazet return true; 1647c1b4a7e6SDavid S. Miller 1648a2a385d6SEric Dumazet return false; 1649c1b4a7e6SDavid S. Miller } 1650c1b4a7e6SDavid S. Miller 1651c1b4a7e6SDavid S. Miller /* Does at least the first segment of SKB fit into the send window? */ 1652a2a385d6SEric Dumazet static bool tcp_snd_wnd_test(const struct tcp_sock *tp, 1653a2a385d6SEric Dumazet const struct sk_buff *skb, 1654056834d9SIlpo Järvinen unsigned int cur_mss) 1655c1b4a7e6SDavid S. Miller { 1656c1b4a7e6SDavid S. Miller u32 end_seq = TCP_SKB_CB(skb)->end_seq; 1657c1b4a7e6SDavid S. Miller 1658c1b4a7e6SDavid S. Miller if (skb->len > cur_mss) 1659c1b4a7e6SDavid S. Miller end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 1660c1b4a7e6SDavid S. Miller 166190840defSIlpo Järvinen return !after(end_seq, tcp_wnd_end(tp)); 1662c1b4a7e6SDavid S. Miller } 1663c1b4a7e6SDavid S. Miller 1664fe067e8aSDavid S. Miller /* This checks if the data bearing packet SKB (usually tcp_send_head(sk)) 1665c1b4a7e6SDavid S. Miller * should be put on the wire right now. If so, it returns the number of 1666c1b4a7e6SDavid S. Miller * packets allowed by the congestion window. 1667c1b4a7e6SDavid S. Miller */ 1668cf533ea5SEric Dumazet static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb, 1669c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1670c1b4a7e6SDavid S. Miller { 1671cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1672c1b4a7e6SDavid S. Miller unsigned int cwnd_quota; 1673c1b4a7e6SDavid S. Miller 1674846998aeSDavid S. Miller tcp_init_tso_segs(sk, skb, cur_mss); 1675c1b4a7e6SDavid S. Miller 1676c1b4a7e6SDavid S. Miller if (!tcp_nagle_test(tp, skb, cur_mss, nonagle)) 1677c1b4a7e6SDavid S. Miller return 0; 1678c1b4a7e6SDavid S. Miller 1679c1b4a7e6SDavid S. Miller cwnd_quota = tcp_cwnd_test(tp, skb); 1680056834d9SIlpo Järvinen if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss)) 1681c1b4a7e6SDavid S. Miller cwnd_quota = 0; 1682c1b4a7e6SDavid S. Miller 1683c1b4a7e6SDavid S. Miller return cwnd_quota; 1684c1b4a7e6SDavid S. Miller } 1685c1b4a7e6SDavid S. Miller 168667edfef7SAndi Kleen /* Test if sending is allowed right now. */ 1687a2a385d6SEric Dumazet bool tcp_may_send_now(struct sock *sk) 1688c1b4a7e6SDavid S. Miller { 1689cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1690fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 1691c1b4a7e6SDavid S. Miller 1692a02cec21SEric Dumazet return skb && 16930c54b85fSIlpo Järvinen tcp_snd_test(sk, skb, tcp_current_mss(sk), 1694c1b4a7e6SDavid S. Miller (tcp_skb_is_last(sk, skb) ? 1695a02cec21SEric Dumazet tp->nonagle : TCP_NAGLE_PUSH)); 1696c1b4a7e6SDavid S. Miller } 1697c1b4a7e6SDavid S. Miller 1698c1b4a7e6SDavid S. Miller /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 1699c1b4a7e6SDavid S. Miller * which is put after SKB on the list. It is very much like 1700c1b4a7e6SDavid S. Miller * tcp_fragment() except that it may make several kinds of assumptions 1701c1b4a7e6SDavid S. Miller * in order to speed up the splitting operation. In particular, we 1702c1b4a7e6SDavid S. Miller * know that all the data is in scatter-gather pages, and that the 1703c1b4a7e6SDavid S. Miller * packet has never been sent out before (and thus is not cloned). 1704c1b4a7e6SDavid S. Miller */ 1705056834d9SIlpo Järvinen static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, 1706c4ead4c5SEric Dumazet unsigned int mss_now, gfp_t gfp) 1707c1b4a7e6SDavid S. Miller { 1708c1b4a7e6SDavid S. Miller struct sk_buff *buff; 1709c1b4a7e6SDavid S. Miller int nlen = skb->len - len; 17109ce01461SIlpo Järvinen u8 flags; 1711c1b4a7e6SDavid S. Miller 1712c1b4a7e6SDavid S. Miller /* All of a TSO frame must be composed of paged data. */ 1713c8ac3774SHerbert Xu if (skb->len != skb->data_len) 1714c8ac3774SHerbert Xu return tcp_fragment(sk, skb, len, mss_now); 1715c1b4a7e6SDavid S. Miller 1716c4ead4c5SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, gfp); 1717c1b4a7e6SDavid S. Miller if (unlikely(buff == NULL)) 1718c1b4a7e6SDavid S. Miller return -ENOMEM; 1719c1b4a7e6SDavid S. Miller 17203ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 17213ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1722b60b49eaSHerbert Xu buff->truesize += nlen; 1723c1b4a7e6SDavid S. Miller skb->truesize -= nlen; 1724c1b4a7e6SDavid S. Miller 1725c1b4a7e6SDavid S. Miller /* Correct the sequence numbers. */ 1726c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 1727c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 1728c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 1729c1b4a7e6SDavid S. Miller 1730c1b4a7e6SDavid S. Miller /* PSH and FIN should only be set in the second packet. */ 17314de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 17324de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 17334de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1734c1b4a7e6SDavid S. Miller 1735c1b4a7e6SDavid S. Miller /* This packet was never sent out yet, so no SACK bits. */ 1736c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->sacked = 0; 1737c1b4a7e6SDavid S. Miller 173884fa7933SPatrick McHardy buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; 1739c1b4a7e6SDavid S. Miller skb_split(skb, buff, len); 1740c1b4a7e6SDavid S. Miller 1741c1b4a7e6SDavid S. Miller /* Fix up tso_factor for both original and new SKB. */ 1742846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1743846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, buff, mss_now); 1744c1b4a7e6SDavid S. Miller 1745c1b4a7e6SDavid S. Miller /* Link BUFF into the send queue. */ 1746c1b4a7e6SDavid S. Miller skb_header_release(buff); 1747fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 1748c1b4a7e6SDavid S. Miller 1749c1b4a7e6SDavid S. Miller return 0; 1750c1b4a7e6SDavid S. Miller } 1751c1b4a7e6SDavid S. Miller 1752c1b4a7e6SDavid S. Miller /* Try to defer sending, if possible, in order to minimize the amount 1753c1b4a7e6SDavid S. Miller * of TSO splitting we do. View it as a kind of TSO Nagle test. 1754c1b4a7e6SDavid S. Miller * 1755c1b4a7e6SDavid S. Miller * This algorithm is from John Heffner. 1756c1b4a7e6SDavid S. Miller */ 1757a2a385d6SEric Dumazet static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb) 1758c1b4a7e6SDavid S. Miller { 17599e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 17606687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 1761c1b4a7e6SDavid S. Miller u32 send_win, cong_win, limit, in_flight; 1762ad9f4f50SEric Dumazet int win_divisor; 1763c1b4a7e6SDavid S. Miller 17644de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 1765ae8064acSJohn Heffner goto send_now; 1766c1b4a7e6SDavid S. Miller 17676687e988SArnaldo Carvalho de Melo if (icsk->icsk_ca_state != TCP_CA_Open) 1768ae8064acSJohn Heffner goto send_now; 1769ae8064acSJohn Heffner 1770ae8064acSJohn Heffner /* Defer for less than two clock ticks. */ 1771bd515c3eSIlpo Järvinen if (tp->tso_deferred && 1772a2acde07SIlpo Järvinen (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1) 1773ae8064acSJohn Heffner goto send_now; 1774908a75c1SDavid S. Miller 1775c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1776c1b4a7e6SDavid S. Miller 1777056834d9SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight)); 1778c1b4a7e6SDavid S. Miller 177990840defSIlpo Järvinen send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1780c1b4a7e6SDavid S. Miller 1781c1b4a7e6SDavid S. Miller /* From in_flight test above, we know that cwnd > in_flight. */ 1782c1b4a7e6SDavid S. Miller cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; 1783c1b4a7e6SDavid S. Miller 1784c1b4a7e6SDavid S. Miller limit = min(send_win, cong_win); 1785c1b4a7e6SDavid S. Miller 1786ba244fe9SDavid S. Miller /* If a full-sized TSO skb can be sent, do it. */ 17871485348dSBen Hutchings if (limit >= min_t(unsigned int, sk->sk_gso_max_size, 17881485348dSBen Hutchings sk->sk_gso_max_segs * tp->mss_cache)) 1789ae8064acSJohn Heffner goto send_now; 1790ba244fe9SDavid S. Miller 179162ad2761SIlpo Järvinen /* Middle in queue won't get any more data, full sendable already? */ 179262ad2761SIlpo Järvinen if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) 179362ad2761SIlpo Järvinen goto send_now; 179462ad2761SIlpo Järvinen 1795ad9f4f50SEric Dumazet win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor); 1796ad9f4f50SEric Dumazet if (win_divisor) { 1797c1b4a7e6SDavid S. Miller u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); 1798c1b4a7e6SDavid S. Miller 1799c1b4a7e6SDavid S. Miller /* If at least some fraction of a window is available, 1800c1b4a7e6SDavid S. Miller * just use it. 1801c1b4a7e6SDavid S. Miller */ 1802ad9f4f50SEric Dumazet chunk /= win_divisor; 1803c1b4a7e6SDavid S. Miller if (limit >= chunk) 1804ae8064acSJohn Heffner goto send_now; 1805c1b4a7e6SDavid S. Miller } else { 1806c1b4a7e6SDavid S. Miller /* Different approach, try not to defer past a single 1807c1b4a7e6SDavid S. Miller * ACK. Receiver should ACK every other full sized 1808c1b4a7e6SDavid S. Miller * frame, so if we have space for more than 3 frames 1809c1b4a7e6SDavid S. Miller * then send now. 1810c1b4a7e6SDavid S. Miller */ 18116b5a5c0dSNeal Cardwell if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) 1812ae8064acSJohn Heffner goto send_now; 1813c1b4a7e6SDavid S. Miller } 1814c1b4a7e6SDavid S. Miller 1815c1b4a7e6SDavid S. Miller /* Ok, it looks like it is advisable to defer. */ 1816ae8064acSJohn Heffner tp->tso_deferred = 1 | (jiffies << 1); 1817ae8064acSJohn Heffner 1818a2a385d6SEric Dumazet return true; 1819ae8064acSJohn Heffner 1820ae8064acSJohn Heffner send_now: 1821ae8064acSJohn Heffner tp->tso_deferred = 0; 1822a2a385d6SEric Dumazet return false; 1823c1b4a7e6SDavid S. Miller } 1824c1b4a7e6SDavid S. Miller 18255d424d5aSJohn Heffner /* Create a new MTU probe if we are ready. 182667edfef7SAndi Kleen * MTU probe is regularly attempting to increase the path MTU by 182767edfef7SAndi Kleen * deliberately sending larger packets. This discovers routing 182867edfef7SAndi Kleen * changes resulting in larger path MTUs. 182967edfef7SAndi Kleen * 18305d424d5aSJohn Heffner * Returns 0 if we should wait to probe (no cwnd available), 18315d424d5aSJohn Heffner * 1 if a probe was sent, 1832056834d9SIlpo Järvinen * -1 otherwise 1833056834d9SIlpo Järvinen */ 18345d424d5aSJohn Heffner static int tcp_mtu_probe(struct sock *sk) 18355d424d5aSJohn Heffner { 18365d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 18375d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 18385d424d5aSJohn Heffner struct sk_buff *skb, *nskb, *next; 18395d424d5aSJohn Heffner int len; 18405d424d5aSJohn Heffner int probe_size; 184191cc17c0SIlpo Järvinen int size_needed; 18425d424d5aSJohn Heffner int copy; 18435d424d5aSJohn Heffner int mss_now; 18445d424d5aSJohn Heffner 18455d424d5aSJohn Heffner /* Not currently probing/verifying, 18465d424d5aSJohn Heffner * not in recovery, 18475d424d5aSJohn Heffner * have enough cwnd, and 18485d424d5aSJohn Heffner * not SACKing (the variable headers throw things off) */ 18495d424d5aSJohn Heffner if (!icsk->icsk_mtup.enabled || 18505d424d5aSJohn Heffner icsk->icsk_mtup.probe_size || 18515d424d5aSJohn Heffner inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 18525d424d5aSJohn Heffner tp->snd_cwnd < 11 || 1853cabeccbdSIlpo Järvinen tp->rx_opt.num_sacks || tp->rx_opt.dsack) 18545d424d5aSJohn Heffner return -1; 18555d424d5aSJohn Heffner 18565d424d5aSJohn Heffner /* Very simple search strategy: just double the MSS. */ 18570c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 18585d424d5aSJohn Heffner probe_size = 2 * tp->mss_cache; 185991cc17c0SIlpo Järvinen size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; 18605d424d5aSJohn Heffner if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) { 18615d424d5aSJohn Heffner /* TODO: set timer for probe_converge_event */ 18625d424d5aSJohn Heffner return -1; 18635d424d5aSJohn Heffner } 18645d424d5aSJohn Heffner 18655d424d5aSJohn Heffner /* Have enough data in the send queue to probe? */ 18667f9c33e5SIlpo Järvinen if (tp->write_seq - tp->snd_nxt < size_needed) 18675d424d5aSJohn Heffner return -1; 18685d424d5aSJohn Heffner 186991cc17c0SIlpo Järvinen if (tp->snd_wnd < size_needed) 18705d424d5aSJohn Heffner return -1; 187190840defSIlpo Järvinen if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) 18725d424d5aSJohn Heffner return 0; 18735d424d5aSJohn Heffner 1874d67c58e9SIlpo Järvinen /* Do we need to wait to drain cwnd? With none in flight, don't stall */ 1875d67c58e9SIlpo Järvinen if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { 1876d67c58e9SIlpo Järvinen if (!tcp_packets_in_flight(tp)) 18775d424d5aSJohn Heffner return -1; 18785d424d5aSJohn Heffner else 18795d424d5aSJohn Heffner return 0; 18805d424d5aSJohn Heffner } 18815d424d5aSJohn Heffner 18825d424d5aSJohn Heffner /* We're allowed to probe. Build it now. */ 18835d424d5aSJohn Heffner if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL) 18845d424d5aSJohn Heffner return -1; 18853ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 18863ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 18875d424d5aSJohn Heffner 1888fe067e8aSDavid S. Miller skb = tcp_send_head(sk); 18895d424d5aSJohn Heffner 18905d424d5aSJohn Heffner TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 18915d424d5aSJohn Heffner TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 18924de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; 18935d424d5aSJohn Heffner TCP_SKB_CB(nskb)->sacked = 0; 18945d424d5aSJohn Heffner nskb->csum = 0; 189584fa7933SPatrick McHardy nskb->ip_summed = skb->ip_summed; 18965d424d5aSJohn Heffner 189750c4817eSIlpo Järvinen tcp_insert_write_queue_before(nskb, skb, sk); 189850c4817eSIlpo Järvinen 18995d424d5aSJohn Heffner len = 0; 1900234b6860SIlpo Järvinen tcp_for_write_queue_from_safe(skb, next, sk) { 19015d424d5aSJohn Heffner copy = min_t(int, skb->len, probe_size - len); 19025d424d5aSJohn Heffner if (nskb->ip_summed) 19035d424d5aSJohn Heffner skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); 19045d424d5aSJohn Heffner else 19055d424d5aSJohn Heffner nskb->csum = skb_copy_and_csum_bits(skb, 0, 1906056834d9SIlpo Järvinen skb_put(nskb, copy), 1907056834d9SIlpo Järvinen copy, nskb->csum); 19085d424d5aSJohn Heffner 19095d424d5aSJohn Heffner if (skb->len <= copy) { 19105d424d5aSJohn Heffner /* We've eaten all the data from this skb. 19115d424d5aSJohn Heffner * Throw it away. */ 19124de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 1913fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 19143ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 19155d424d5aSJohn Heffner } else { 19164de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & 1917a3433f35SChangli Gao ~(TCPHDR_FIN|TCPHDR_PSH); 19185d424d5aSJohn Heffner if (!skb_shinfo(skb)->nr_frags) { 19195d424d5aSJohn Heffner skb_pull(skb, copy); 192084fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 1921056834d9SIlpo Järvinen skb->csum = csum_partial(skb->data, 1922056834d9SIlpo Järvinen skb->len, 0); 19235d424d5aSJohn Heffner } else { 19245d424d5aSJohn Heffner __pskb_trim_head(skb, copy); 19255d424d5aSJohn Heffner tcp_set_skb_tso_segs(sk, skb, mss_now); 19265d424d5aSJohn Heffner } 19275d424d5aSJohn Heffner TCP_SKB_CB(skb)->seq += copy; 19285d424d5aSJohn Heffner } 19295d424d5aSJohn Heffner 19305d424d5aSJohn Heffner len += copy; 1931234b6860SIlpo Järvinen 1932234b6860SIlpo Järvinen if (len >= probe_size) 1933234b6860SIlpo Järvinen break; 19345d424d5aSJohn Heffner } 19355d424d5aSJohn Heffner tcp_init_tso_segs(sk, nskb, nskb->len); 19365d424d5aSJohn Heffner 19375d424d5aSJohn Heffner /* We're ready to send. If this fails, the probe will 19385d424d5aSJohn Heffner * be resegmented into mss-sized pieces by tcp_write_xmit(). */ 19395d424d5aSJohn Heffner TCP_SKB_CB(nskb)->when = tcp_time_stamp; 19405d424d5aSJohn Heffner if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 19415d424d5aSJohn Heffner /* Decrement cwnd here because we are sending 19425d424d5aSJohn Heffner * effectively two packets. */ 19435d424d5aSJohn Heffner tp->snd_cwnd--; 194466f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, nskb); 19455d424d5aSJohn Heffner 19465d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 19470e7b1368SJohn Heffner tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 19480e7b1368SJohn Heffner tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 19495d424d5aSJohn Heffner 19505d424d5aSJohn Heffner return 1; 19515d424d5aSJohn Heffner } 19525d424d5aSJohn Heffner 19535d424d5aSJohn Heffner return -1; 19545d424d5aSJohn Heffner } 19555d424d5aSJohn Heffner 19561da177e4SLinus Torvalds /* This routine writes packets to the network. It advances the 19571da177e4SLinus Torvalds * send_head. This happens as incoming acks open up the remote 19581da177e4SLinus Torvalds * window for us. 19591da177e4SLinus Torvalds * 1960f8269a49SIlpo Järvinen * LARGESEND note: !tcp_urg_mode is overkill, only frames between 1961f8269a49SIlpo Järvinen * snd_up-64k-mss .. snd_up cannot be large. However, taking into 1962f8269a49SIlpo Järvinen * account rare use of URG, this is not a big flaw. 1963f8269a49SIlpo Järvinen * 1964*6ba8a3b1SNandita Dukkipati * Send at most one packet when push_one > 0. Temporarily ignore 1965*6ba8a3b1SNandita Dukkipati * cwnd limit to force at most one packet out when push_one == 2. 1966*6ba8a3b1SNandita Dukkipati 1967a2a385d6SEric Dumazet * Returns true, if no segments are in flight and we have queued segments, 1968a2a385d6SEric Dumazet * but cannot send anything now because of SWS or another problem. 19691da177e4SLinus Torvalds */ 1970a2a385d6SEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 1971d5dd9175SIlpo Järvinen int push_one, gfp_t gfp) 19721da177e4SLinus Torvalds { 19731da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 197492df7b51SDavid S. Miller struct sk_buff *skb; 1975c1b4a7e6SDavid S. Miller unsigned int tso_segs, sent_pkts; 1976c1b4a7e6SDavid S. Miller int cwnd_quota; 19775d424d5aSJohn Heffner int result; 19781da177e4SLinus Torvalds 1979c1b4a7e6SDavid S. Miller sent_pkts = 0; 19805d424d5aSJohn Heffner 1981d5dd9175SIlpo Järvinen if (!push_one) { 19825d424d5aSJohn Heffner /* Do MTU probing. */ 1983d5dd9175SIlpo Järvinen result = tcp_mtu_probe(sk); 1984d5dd9175SIlpo Järvinen if (!result) { 1985a2a385d6SEric Dumazet return false; 19865d424d5aSJohn Heffner } else if (result > 0) { 19875d424d5aSJohn Heffner sent_pkts = 1; 19885d424d5aSJohn Heffner } 1989d5dd9175SIlpo Järvinen } 19905d424d5aSJohn Heffner 1991fe067e8aSDavid S. Miller while ((skb = tcp_send_head(sk))) { 1992c8ac3774SHerbert Xu unsigned int limit; 1993c8ac3774SHerbert Xu 199446d3ceabSEric Dumazet 1995b68e9f85SHerbert Xu tso_segs = tcp_init_tso_segs(sk, skb, mss_now); 1996c1b4a7e6SDavid S. Miller BUG_ON(!tso_segs); 1997c1b4a7e6SDavid S. Miller 1998ec342325SAndrew Vagin if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) 1999ec342325SAndrew Vagin goto repair; /* Skip network transmission */ 2000ec342325SAndrew Vagin 2001b68e9f85SHerbert Xu cwnd_quota = tcp_cwnd_test(tp, skb); 2002*6ba8a3b1SNandita Dukkipati if (!cwnd_quota) { 2003*6ba8a3b1SNandita Dukkipati if (push_one == 2) 2004*6ba8a3b1SNandita Dukkipati /* Force out a loss probe pkt. */ 2005*6ba8a3b1SNandita Dukkipati cwnd_quota = 1; 2006*6ba8a3b1SNandita Dukkipati else 2007b68e9f85SHerbert Xu break; 2008*6ba8a3b1SNandita Dukkipati } 2009b68e9f85SHerbert Xu 2010b68e9f85SHerbert Xu if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) 2011b68e9f85SHerbert Xu break; 2012b68e9f85SHerbert Xu 2013c1b4a7e6SDavid S. Miller if (tso_segs == 1) { 2014aa93466bSDavid S. Miller if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 2015aa93466bSDavid S. Miller (tcp_skb_is_last(sk, skb) ? 2016aa93466bSDavid S. Miller nonagle : TCP_NAGLE_PUSH)))) 2017aa93466bSDavid S. Miller break; 2018c1b4a7e6SDavid S. Miller } else { 2019d5dd9175SIlpo Järvinen if (!push_one && tcp_tso_should_defer(sk, skb)) 2020aa93466bSDavid S. Miller break; 2021c1b4a7e6SDavid S. Miller } 2022aa93466bSDavid S. Miller 202346d3ceabSEric Dumazet /* TSQ : sk_wmem_alloc accounts skb truesize, 202446d3ceabSEric Dumazet * including skb overhead. But thats OK. 202546d3ceabSEric Dumazet */ 202646d3ceabSEric Dumazet if (atomic_read(&sk->sk_wmem_alloc) >= sysctl_tcp_limit_output_bytes) { 202746d3ceabSEric Dumazet set_bit(TSQ_THROTTLED, &tp->tsq_flags); 202846d3ceabSEric Dumazet break; 202946d3ceabSEric Dumazet } 2030c8ac3774SHerbert Xu limit = mss_now; 2031f8269a49SIlpo Järvinen if (tso_segs > 1 && !tcp_urg_mode(tp)) 20320e3a4803SIlpo Järvinen limit = tcp_mss_split_point(sk, skb, mss_now, 20331485348dSBen Hutchings min_t(unsigned int, 20341485348dSBen Hutchings cwnd_quota, 20351485348dSBen Hutchings sk->sk_gso_max_segs)); 2036c8ac3774SHerbert Xu 2037c8ac3774SHerbert Xu if (skb->len > limit && 2038c4ead4c5SEric Dumazet unlikely(tso_fragment(sk, skb, limit, mss_now, gfp))) 20391da177e4SLinus Torvalds break; 20401da177e4SLinus Torvalds 20411da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2042c1b4a7e6SDavid S. Miller 2043d5dd9175SIlpo Järvinen if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) 20441da177e4SLinus Torvalds break; 20451da177e4SLinus Torvalds 2046ec342325SAndrew Vagin repair: 20471da177e4SLinus Torvalds /* Advance the send_head. This one is sent out. 20481da177e4SLinus Torvalds * This call will increment packets_out. 20491da177e4SLinus Torvalds */ 205066f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 20511da177e4SLinus Torvalds 20521da177e4SLinus Torvalds tcp_minshall_update(tp, mss_now, skb); 2053a262f0cdSNandita Dukkipati sent_pkts += tcp_skb_pcount(skb); 2054d5dd9175SIlpo Järvinen 2055d5dd9175SIlpo Järvinen if (push_one) 2056d5dd9175SIlpo Järvinen break; 20571da177e4SLinus Torvalds } 20581da177e4SLinus Torvalds 2059aa93466bSDavid S. Miller if (likely(sent_pkts)) { 2060684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 2061684bad11SYuchung Cheng tp->prr_out += sent_pkts; 2062*6ba8a3b1SNandita Dukkipati 2063*6ba8a3b1SNandita Dukkipati /* Send one loss probe per tail loss episode. */ 2064*6ba8a3b1SNandita Dukkipati if (push_one != 2) 2065*6ba8a3b1SNandita Dukkipati tcp_schedule_loss_probe(sk); 20669e412ba7SIlpo Järvinen tcp_cwnd_validate(sk); 2067a2a385d6SEric Dumazet return false; 20681da177e4SLinus Torvalds } 2069*6ba8a3b1SNandita Dukkipati return (push_one == 2) || (!tp->packets_out && tcp_send_head(sk)); 2070*6ba8a3b1SNandita Dukkipati } 2071*6ba8a3b1SNandita Dukkipati 2072*6ba8a3b1SNandita Dukkipati bool tcp_schedule_loss_probe(struct sock *sk) 2073*6ba8a3b1SNandita Dukkipati { 2074*6ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 2075*6ba8a3b1SNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 2076*6ba8a3b1SNandita Dukkipati u32 timeout, tlp_time_stamp, rto_time_stamp; 2077*6ba8a3b1SNandita Dukkipati u32 rtt = tp->srtt >> 3; 2078*6ba8a3b1SNandita Dukkipati 2079*6ba8a3b1SNandita Dukkipati if (WARN_ON(icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS)) 2080*6ba8a3b1SNandita Dukkipati return false; 2081*6ba8a3b1SNandita Dukkipati /* No consecutive loss probes. */ 2082*6ba8a3b1SNandita Dukkipati if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) { 2083*6ba8a3b1SNandita Dukkipati tcp_rearm_rto(sk); 2084*6ba8a3b1SNandita Dukkipati return false; 2085*6ba8a3b1SNandita Dukkipati } 2086*6ba8a3b1SNandita Dukkipati /* Don't do any loss probe on a Fast Open connection before 3WHS 2087*6ba8a3b1SNandita Dukkipati * finishes. 2088*6ba8a3b1SNandita Dukkipati */ 2089*6ba8a3b1SNandita Dukkipati if (sk->sk_state == TCP_SYN_RECV) 2090*6ba8a3b1SNandita Dukkipati return false; 2091*6ba8a3b1SNandita Dukkipati 2092*6ba8a3b1SNandita Dukkipati /* TLP is only scheduled when next timer event is RTO. */ 2093*6ba8a3b1SNandita Dukkipati if (icsk->icsk_pending != ICSK_TIME_RETRANS) 2094*6ba8a3b1SNandita Dukkipati return false; 2095*6ba8a3b1SNandita Dukkipati 2096*6ba8a3b1SNandita Dukkipati /* Schedule a loss probe in 2*RTT for SACK capable connections 2097*6ba8a3b1SNandita Dukkipati * in Open state, that are either limited by cwnd or application. 2098*6ba8a3b1SNandita Dukkipati */ 2099*6ba8a3b1SNandita Dukkipati if (sysctl_tcp_early_retrans < 3 || !rtt || !tp->packets_out || 2100*6ba8a3b1SNandita Dukkipati !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open) 2101*6ba8a3b1SNandita Dukkipati return false; 2102*6ba8a3b1SNandita Dukkipati 2103*6ba8a3b1SNandita Dukkipati if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) && 2104*6ba8a3b1SNandita Dukkipati tcp_send_head(sk)) 2105*6ba8a3b1SNandita Dukkipati return false; 2106*6ba8a3b1SNandita Dukkipati 2107*6ba8a3b1SNandita Dukkipati /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account 2108*6ba8a3b1SNandita Dukkipati * for delayed ack when there's one outstanding packet. 2109*6ba8a3b1SNandita Dukkipati */ 2110*6ba8a3b1SNandita Dukkipati timeout = rtt << 1; 2111*6ba8a3b1SNandita Dukkipati if (tp->packets_out == 1) 2112*6ba8a3b1SNandita Dukkipati timeout = max_t(u32, timeout, 2113*6ba8a3b1SNandita Dukkipati (rtt + (rtt >> 1) + TCP_DELACK_MAX)); 2114*6ba8a3b1SNandita Dukkipati timeout = max_t(u32, timeout, msecs_to_jiffies(10)); 2115*6ba8a3b1SNandita Dukkipati 2116*6ba8a3b1SNandita Dukkipati /* If RTO is shorter, just schedule TLP in its place. */ 2117*6ba8a3b1SNandita Dukkipati tlp_time_stamp = tcp_time_stamp + timeout; 2118*6ba8a3b1SNandita Dukkipati rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout; 2119*6ba8a3b1SNandita Dukkipati if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) { 2120*6ba8a3b1SNandita Dukkipati s32 delta = rto_time_stamp - tcp_time_stamp; 2121*6ba8a3b1SNandita Dukkipati if (delta > 0) 2122*6ba8a3b1SNandita Dukkipati timeout = delta; 2123*6ba8a3b1SNandita Dukkipati } 2124*6ba8a3b1SNandita Dukkipati 2125*6ba8a3b1SNandita Dukkipati inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, 2126*6ba8a3b1SNandita Dukkipati TCP_RTO_MAX); 2127*6ba8a3b1SNandita Dukkipati return true; 2128*6ba8a3b1SNandita Dukkipati } 2129*6ba8a3b1SNandita Dukkipati 2130*6ba8a3b1SNandita Dukkipati /* When probe timeout (PTO) fires, send a new segment if one exists, else 2131*6ba8a3b1SNandita Dukkipati * retransmit the last segment. 2132*6ba8a3b1SNandita Dukkipati */ 2133*6ba8a3b1SNandita Dukkipati void tcp_send_loss_probe(struct sock *sk) 2134*6ba8a3b1SNandita Dukkipati { 2135*6ba8a3b1SNandita Dukkipati struct sk_buff *skb; 2136*6ba8a3b1SNandita Dukkipati int pcount; 2137*6ba8a3b1SNandita Dukkipati int mss = tcp_current_mss(sk); 2138*6ba8a3b1SNandita Dukkipati int err = -1; 2139*6ba8a3b1SNandita Dukkipati 2140*6ba8a3b1SNandita Dukkipati if (tcp_send_head(sk) != NULL) { 2141*6ba8a3b1SNandita Dukkipati err = tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC); 2142*6ba8a3b1SNandita Dukkipati goto rearm_timer; 2143*6ba8a3b1SNandita Dukkipati } 2144*6ba8a3b1SNandita Dukkipati 2145*6ba8a3b1SNandita Dukkipati /* Retransmit last segment. */ 2146*6ba8a3b1SNandita Dukkipati skb = tcp_write_queue_tail(sk); 2147*6ba8a3b1SNandita Dukkipati if (WARN_ON(!skb)) 2148*6ba8a3b1SNandita Dukkipati goto rearm_timer; 2149*6ba8a3b1SNandita Dukkipati 2150*6ba8a3b1SNandita Dukkipati pcount = tcp_skb_pcount(skb); 2151*6ba8a3b1SNandita Dukkipati if (WARN_ON(!pcount)) 2152*6ba8a3b1SNandita Dukkipati goto rearm_timer; 2153*6ba8a3b1SNandita Dukkipati 2154*6ba8a3b1SNandita Dukkipati if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) { 2155*6ba8a3b1SNandita Dukkipati if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss))) 2156*6ba8a3b1SNandita Dukkipati goto rearm_timer; 2157*6ba8a3b1SNandita Dukkipati skb = tcp_write_queue_tail(sk); 2158*6ba8a3b1SNandita Dukkipati } 2159*6ba8a3b1SNandita Dukkipati 2160*6ba8a3b1SNandita Dukkipati if (WARN_ON(!skb || !tcp_skb_pcount(skb))) 2161*6ba8a3b1SNandita Dukkipati goto rearm_timer; 2162*6ba8a3b1SNandita Dukkipati 2163*6ba8a3b1SNandita Dukkipati /* Probe with zero data doesn't trigger fast recovery. */ 2164*6ba8a3b1SNandita Dukkipati if (skb->len > 0) 2165*6ba8a3b1SNandita Dukkipati err = __tcp_retransmit_skb(sk, skb); 2166*6ba8a3b1SNandita Dukkipati 2167*6ba8a3b1SNandita Dukkipati rearm_timer: 2168*6ba8a3b1SNandita Dukkipati inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 2169*6ba8a3b1SNandita Dukkipati inet_csk(sk)->icsk_rto, 2170*6ba8a3b1SNandita Dukkipati TCP_RTO_MAX); 2171*6ba8a3b1SNandita Dukkipati 2172*6ba8a3b1SNandita Dukkipati if (likely(!err)) 2173*6ba8a3b1SNandita Dukkipati NET_INC_STATS_BH(sock_net(sk), 2174*6ba8a3b1SNandita Dukkipati LINUX_MIB_TCPLOSSPROBES); 2175*6ba8a3b1SNandita Dukkipati return; 21761da177e4SLinus Torvalds } 21771da177e4SLinus Torvalds 2178a762a980SDavid S. Miller /* Push out any pending frames which were held back due to 2179a762a980SDavid S. Miller * TCP_CORK or attempt at coalescing tiny packets. 2180a762a980SDavid S. Miller * The socket must be locked by the caller. 2181a762a980SDavid S. Miller */ 21829e412ba7SIlpo Järvinen void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 21839e412ba7SIlpo Järvinen int nonagle) 2184a762a980SDavid S. Miller { 2185726e07a8SIlpo Järvinen /* If we are closed, the bytes will have to remain here. 2186726e07a8SIlpo Järvinen * In time closedown will finish, we empty the write queue and 2187726e07a8SIlpo Järvinen * all will be happy. 2188726e07a8SIlpo Järvinen */ 2189726e07a8SIlpo Järvinen if (unlikely(sk->sk_state == TCP_CLOSE)) 2190726e07a8SIlpo Järvinen return; 2191726e07a8SIlpo Järvinen 219299a1dec7SMel Gorman if (tcp_write_xmit(sk, cur_mss, nonagle, 0, 219399a1dec7SMel Gorman sk_gfp_atomic(sk, GFP_ATOMIC))) 21949e412ba7SIlpo Järvinen tcp_check_probe_timer(sk); 2195a762a980SDavid S. Miller } 2196a762a980SDavid S. Miller 2197c1b4a7e6SDavid S. Miller /* Send _single_ skb sitting at the send head. This function requires 2198c1b4a7e6SDavid S. Miller * true push pending frames to setup probe timer etc. 2199c1b4a7e6SDavid S. Miller */ 2200c1b4a7e6SDavid S. Miller void tcp_push_one(struct sock *sk, unsigned int mss_now) 2201c1b4a7e6SDavid S. Miller { 2202fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 2203c1b4a7e6SDavid S. Miller 2204c1b4a7e6SDavid S. Miller BUG_ON(!skb || skb->len < mss_now); 2205c1b4a7e6SDavid S. Miller 2206d5dd9175SIlpo Järvinen tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); 2207c1b4a7e6SDavid S. Miller } 2208c1b4a7e6SDavid S. Miller 22091da177e4SLinus Torvalds /* This function returns the amount that we can raise the 22101da177e4SLinus Torvalds * usable window based on the following constraints 22111da177e4SLinus Torvalds * 22121da177e4SLinus Torvalds * 1. The window can never be shrunk once it is offered (RFC 793) 22131da177e4SLinus Torvalds * 2. We limit memory per socket 22141da177e4SLinus Torvalds * 22151da177e4SLinus Torvalds * RFC 1122: 22161da177e4SLinus Torvalds * "the suggested [SWS] avoidance algorithm for the receiver is to keep 22171da177e4SLinus Torvalds * RECV.NEXT + RCV.WIN fixed until: 22181da177e4SLinus Torvalds * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 22191da177e4SLinus Torvalds * 22201da177e4SLinus Torvalds * i.e. don't raise the right edge of the window until you can raise 22211da177e4SLinus Torvalds * it at least MSS bytes. 22221da177e4SLinus Torvalds * 22231da177e4SLinus Torvalds * Unfortunately, the recommended algorithm breaks header prediction, 22241da177e4SLinus Torvalds * since header prediction assumes th->window stays fixed. 22251da177e4SLinus Torvalds * 22261da177e4SLinus Torvalds * Strictly speaking, keeping th->window fixed violates the receiver 22271da177e4SLinus Torvalds * side SWS prevention criteria. The problem is that under this rule 22281da177e4SLinus Torvalds * a stream of single byte packets will cause the right side of the 22291da177e4SLinus Torvalds * window to always advance by a single byte. 22301da177e4SLinus Torvalds * 22311da177e4SLinus Torvalds * Of course, if the sender implements sender side SWS prevention 22321da177e4SLinus Torvalds * then this will not be a problem. 22331da177e4SLinus Torvalds * 22341da177e4SLinus Torvalds * BSD seems to make the following compromise: 22351da177e4SLinus Torvalds * 22361da177e4SLinus Torvalds * If the free space is less than the 1/4 of the maximum 22371da177e4SLinus Torvalds * space available and the free space is less than 1/2 mss, 22381da177e4SLinus Torvalds * then set the window to 0. 22391da177e4SLinus Torvalds * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 22401da177e4SLinus Torvalds * Otherwise, just prevent the window from shrinking 22411da177e4SLinus Torvalds * and from being larger than the largest representable value. 22421da177e4SLinus Torvalds * 22431da177e4SLinus Torvalds * This prevents incremental opening of the window in the regime 22441da177e4SLinus Torvalds * where TCP is limited by the speed of the reader side taking 22451da177e4SLinus Torvalds * data out of the TCP receive queue. It does nothing about 22461da177e4SLinus Torvalds * those cases where the window is constrained on the sender side 22471da177e4SLinus Torvalds * because the pipeline is full. 22481da177e4SLinus Torvalds * 22491da177e4SLinus Torvalds * BSD also seems to "accidentally" limit itself to windows that are a 22501da177e4SLinus Torvalds * multiple of MSS, at least until the free space gets quite small. 22511da177e4SLinus Torvalds * This would appear to be a side effect of the mbuf implementation. 22521da177e4SLinus Torvalds * Combining these two algorithms results in the observed behavior 22531da177e4SLinus Torvalds * of having a fixed window size at almost all times. 22541da177e4SLinus Torvalds * 22551da177e4SLinus Torvalds * Below we obtain similar behavior by forcing the offered window to 22561da177e4SLinus Torvalds * a multiple of the mss when it is feasible to do so. 22571da177e4SLinus Torvalds * 22581da177e4SLinus Torvalds * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 22591da177e4SLinus Torvalds * Regular options like TIMESTAMP are taken into account. 22601da177e4SLinus Torvalds */ 22611da177e4SLinus Torvalds u32 __tcp_select_window(struct sock *sk) 22621da177e4SLinus Torvalds { 2263463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 22641da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2265caa20d9aSStephen Hemminger /* MSS for the peer's data. Previous versions used mss_clamp 22661da177e4SLinus Torvalds * here. I don't know if the value based on our guesses 22671da177e4SLinus Torvalds * of peer's MSS is better for the performance. It's more correct 22681da177e4SLinus Torvalds * but may be worse for the performance because of rcv_mss 22691da177e4SLinus Torvalds * fluctuations. --SAW 1998/11/1 22701da177e4SLinus Torvalds */ 2271463c84b9SArnaldo Carvalho de Melo int mss = icsk->icsk_ack.rcv_mss; 22721da177e4SLinus Torvalds int free_space = tcp_space(sk); 22731da177e4SLinus Torvalds int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk)); 22741da177e4SLinus Torvalds int window; 22751da177e4SLinus Torvalds 22761da177e4SLinus Torvalds if (mss > full_space) 22771da177e4SLinus Torvalds mss = full_space; 22781da177e4SLinus Torvalds 2279b92edbe0SEric Dumazet if (free_space < (full_space >> 1)) { 2280463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.quick = 0; 22811da177e4SLinus Torvalds 2282180d8cd9SGlauber Costa if (sk_under_memory_pressure(sk)) 2283056834d9SIlpo Järvinen tp->rcv_ssthresh = min(tp->rcv_ssthresh, 2284056834d9SIlpo Järvinen 4U * tp->advmss); 22851da177e4SLinus Torvalds 22861da177e4SLinus Torvalds if (free_space < mss) 22871da177e4SLinus Torvalds return 0; 22881da177e4SLinus Torvalds } 22891da177e4SLinus Torvalds 22901da177e4SLinus Torvalds if (free_space > tp->rcv_ssthresh) 22911da177e4SLinus Torvalds free_space = tp->rcv_ssthresh; 22921da177e4SLinus Torvalds 22931da177e4SLinus Torvalds /* Don't do rounding if we are using window scaling, since the 22941da177e4SLinus Torvalds * scaled window will not line up with the MSS boundary anyway. 22951da177e4SLinus Torvalds */ 22961da177e4SLinus Torvalds window = tp->rcv_wnd; 22971da177e4SLinus Torvalds if (tp->rx_opt.rcv_wscale) { 22981da177e4SLinus Torvalds window = free_space; 22991da177e4SLinus Torvalds 23001da177e4SLinus Torvalds /* Advertise enough space so that it won't get scaled away. 23011da177e4SLinus Torvalds * Import case: prevent zero window announcement if 23021da177e4SLinus Torvalds * 1<<rcv_wscale > mss. 23031da177e4SLinus Torvalds */ 23041da177e4SLinus Torvalds if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window) 23051da177e4SLinus Torvalds window = (((window >> tp->rx_opt.rcv_wscale) + 1) 23061da177e4SLinus Torvalds << tp->rx_opt.rcv_wscale); 23071da177e4SLinus Torvalds } else { 23081da177e4SLinus Torvalds /* Get the largest window that is a nice multiple of mss. 23091da177e4SLinus Torvalds * Window clamp already applied above. 23101da177e4SLinus Torvalds * If our current window offering is within 1 mss of the 23111da177e4SLinus Torvalds * free space we just keep it. This prevents the divide 23121da177e4SLinus Torvalds * and multiply from happening most of the time. 23131da177e4SLinus Torvalds * We also don't do any window rounding when the free space 23141da177e4SLinus Torvalds * is too small. 23151da177e4SLinus Torvalds */ 23161da177e4SLinus Torvalds if (window <= free_space - mss || window > free_space) 23171da177e4SLinus Torvalds window = (free_space / mss) * mss; 231884565070SJohn Heffner else if (mss == full_space && 2319b92edbe0SEric Dumazet free_space > window + (full_space >> 1)) 232084565070SJohn Heffner window = free_space; 23211da177e4SLinus Torvalds } 23221da177e4SLinus Torvalds 23231da177e4SLinus Torvalds return window; 23241da177e4SLinus Torvalds } 23251da177e4SLinus Torvalds 23264a17fc3aSIlpo Järvinen /* Collapses two adjacent SKB's during retransmission. */ 23274a17fc3aSIlpo Järvinen static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) 23281da177e4SLinus Torvalds { 23291da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2330fe067e8aSDavid S. Miller struct sk_buff *next_skb = tcp_write_queue_next(sk, skb); 2331058dc334SIlpo Järvinen int skb_size, next_skb_size; 23321da177e4SLinus Torvalds 2333058dc334SIlpo Järvinen skb_size = skb->len; 2334058dc334SIlpo Järvinen next_skb_size = next_skb->len; 23351da177e4SLinus Torvalds 2336058dc334SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); 23371da177e4SLinus Torvalds 23386859d494SIlpo Järvinen tcp_highest_sack_combine(sk, next_skb, skb); 2339a6963a6bSIlpo Järvinen 2340fe067e8aSDavid S. Miller tcp_unlink_write_queue(next_skb, sk); 23411da177e4SLinus Torvalds 2342058dc334SIlpo Järvinen skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size), 23431a4e2d09SArnaldo Carvalho de Melo next_skb_size); 23441da177e4SLinus Torvalds 234552d570aaSJarek Poplawski if (next_skb->ip_summed == CHECKSUM_PARTIAL) 234652d570aaSJarek Poplawski skb->ip_summed = CHECKSUM_PARTIAL; 23471da177e4SLinus Torvalds 234884fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 23491da177e4SLinus Torvalds skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); 23501da177e4SLinus Torvalds 23511da177e4SLinus Torvalds /* Update sequence range on original skb. */ 23521da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 23531da177e4SLinus Torvalds 2354e6c7d085SIlpo Järvinen /* Merge over control information. This moves PSH/FIN etc. over */ 23554de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; 23561da177e4SLinus Torvalds 23571da177e4SLinus Torvalds /* All done, get rid of second SKB and account for it so 23581da177e4SLinus Torvalds * packet counting does not break. 23591da177e4SLinus Torvalds */ 23604828e7f4SIlpo Järvinen TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; 2361b7689205SIlpo Järvinen 2362b7689205SIlpo Järvinen /* changed transmit queue under us so clear hints */ 2363ef9da47cSIlpo Järvinen tcp_clear_retrans_hints_partial(tp); 2364ef9da47cSIlpo Järvinen if (next_skb == tp->retransmit_skb_hint) 2365ef9da47cSIlpo Järvinen tp->retransmit_skb_hint = skb; 2366b7689205SIlpo Järvinen 2367797108d1SIlpo Järvinen tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); 2368797108d1SIlpo Järvinen 23693ab224beSHideo Aoki sk_wmem_free_skb(sk, next_skb); 23701da177e4SLinus Torvalds } 23711da177e4SLinus Torvalds 237267edfef7SAndi Kleen /* Check if coalescing SKBs is legal. */ 2373a2a385d6SEric Dumazet static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) 23744a17fc3aSIlpo Järvinen { 23754a17fc3aSIlpo Järvinen if (tcp_skb_pcount(skb) > 1) 2376a2a385d6SEric Dumazet return false; 23774a17fc3aSIlpo Järvinen /* TODO: SACK collapsing could be used to remove this condition */ 23784a17fc3aSIlpo Järvinen if (skb_shinfo(skb)->nr_frags != 0) 2379a2a385d6SEric Dumazet return false; 23804a17fc3aSIlpo Järvinen if (skb_cloned(skb)) 2381a2a385d6SEric Dumazet return false; 23824a17fc3aSIlpo Järvinen if (skb == tcp_send_head(sk)) 2383a2a385d6SEric Dumazet return false; 23844a17fc3aSIlpo Järvinen /* Some heurestics for collapsing over SACK'd could be invented */ 23854a17fc3aSIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 2386a2a385d6SEric Dumazet return false; 23874a17fc3aSIlpo Järvinen 2388a2a385d6SEric Dumazet return true; 23894a17fc3aSIlpo Järvinen } 23904a17fc3aSIlpo Järvinen 239167edfef7SAndi Kleen /* Collapse packets in the retransmit queue to make to create 239267edfef7SAndi Kleen * less packets on the wire. This is only done on retransmission. 239367edfef7SAndi Kleen */ 23944a17fc3aSIlpo Järvinen static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, 23954a17fc3aSIlpo Järvinen int space) 23964a17fc3aSIlpo Järvinen { 23974a17fc3aSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 23984a17fc3aSIlpo Järvinen struct sk_buff *skb = to, *tmp; 2399a2a385d6SEric Dumazet bool first = true; 24004a17fc3aSIlpo Järvinen 24014a17fc3aSIlpo Järvinen if (!sysctl_tcp_retrans_collapse) 24024a17fc3aSIlpo Järvinen return; 24034de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 24044a17fc3aSIlpo Järvinen return; 24054a17fc3aSIlpo Järvinen 24064a17fc3aSIlpo Järvinen tcp_for_write_queue_from_safe(skb, tmp, sk) { 24074a17fc3aSIlpo Järvinen if (!tcp_can_collapse(sk, skb)) 24084a17fc3aSIlpo Järvinen break; 24094a17fc3aSIlpo Järvinen 24104a17fc3aSIlpo Järvinen space -= skb->len; 24114a17fc3aSIlpo Järvinen 24124a17fc3aSIlpo Järvinen if (first) { 2413a2a385d6SEric Dumazet first = false; 24144a17fc3aSIlpo Järvinen continue; 24154a17fc3aSIlpo Järvinen } 24164a17fc3aSIlpo Järvinen 24174a17fc3aSIlpo Järvinen if (space < 0) 24184a17fc3aSIlpo Järvinen break; 24194a17fc3aSIlpo Järvinen /* Punt if not enough space exists in the first SKB for 24204a17fc3aSIlpo Järvinen * the data in the second 24214a17fc3aSIlpo Järvinen */ 2422a21d4572SEric Dumazet if (skb->len > skb_availroom(to)) 24234a17fc3aSIlpo Järvinen break; 24244a17fc3aSIlpo Järvinen 24254a17fc3aSIlpo Järvinen if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) 24264a17fc3aSIlpo Järvinen break; 24274a17fc3aSIlpo Järvinen 24284a17fc3aSIlpo Järvinen tcp_collapse_retrans(sk, to); 24294a17fc3aSIlpo Järvinen } 24304a17fc3aSIlpo Järvinen } 24314a17fc3aSIlpo Järvinen 24321da177e4SLinus Torvalds /* This retransmits one SKB. Policy decisions and retransmit queue 24331da177e4SLinus Torvalds * state updates are done by the caller. Returns non-zero if an 24341da177e4SLinus Torvalds * error occurred which prevented the send. 24351da177e4SLinus Torvalds */ 243693b174adSYuchung Cheng int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) 24371da177e4SLinus Torvalds { 24381da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 24395d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 24407d227cd2SSridhar Samudrala unsigned int cur_mss; 24411da177e4SLinus Torvalds 24425d424d5aSJohn Heffner /* Inconslusive MTU probe */ 24435d424d5aSJohn Heffner if (icsk->icsk_mtup.probe_size) { 24445d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 24455d424d5aSJohn Heffner } 24465d424d5aSJohn Heffner 24471da177e4SLinus Torvalds /* Do not sent more than we queued. 1/4 is reserved for possible 2448caa20d9aSStephen Hemminger * copying overhead: fragmentation, tunneling, mangling etc. 24491da177e4SLinus Torvalds */ 24501da177e4SLinus Torvalds if (atomic_read(&sk->sk_wmem_alloc) > 24511da177e4SLinus Torvalds min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf)) 24521da177e4SLinus Torvalds return -EAGAIN; 24531da177e4SLinus Torvalds 24541da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 24551da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 24561da177e4SLinus Torvalds BUG(); 24571da177e4SLinus Torvalds if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 24581da177e4SLinus Torvalds return -ENOMEM; 24591da177e4SLinus Torvalds } 24601da177e4SLinus Torvalds 24617d227cd2SSridhar Samudrala if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 24627d227cd2SSridhar Samudrala return -EHOSTUNREACH; /* Routing failure or similar. */ 24637d227cd2SSridhar Samudrala 24640c54b85fSIlpo Järvinen cur_mss = tcp_current_mss(sk); 24657d227cd2SSridhar Samudrala 24661da177e4SLinus Torvalds /* If receiver has shrunk his window, and skb is out of 24671da177e4SLinus Torvalds * new window, do not retransmit it. The exception is the 24681da177e4SLinus Torvalds * case, when window is shrunk to zero. In this case 24691da177e4SLinus Torvalds * our retransmit serves as a zero window probe. 24701da177e4SLinus Torvalds */ 24719d4fb27dSJoe Perches if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && 24729d4fb27dSJoe Perches TCP_SKB_CB(skb)->seq != tp->snd_una) 24731da177e4SLinus Torvalds return -EAGAIN; 24741da177e4SLinus Torvalds 24751da177e4SLinus Torvalds if (skb->len > cur_mss) { 2476846998aeSDavid S. Miller if (tcp_fragment(sk, skb, cur_mss, cur_mss)) 24771da177e4SLinus Torvalds return -ENOMEM; /* We'll try again later. */ 247802276f3cSIlpo Järvinen } else { 24799eb9362eSIlpo Järvinen int oldpcount = tcp_skb_pcount(skb); 24809eb9362eSIlpo Järvinen 24819eb9362eSIlpo Järvinen if (unlikely(oldpcount > 1)) { 248202276f3cSIlpo Järvinen tcp_init_tso_segs(sk, skb, cur_mss); 24839eb9362eSIlpo Järvinen tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb)); 24849eb9362eSIlpo Järvinen } 24851da177e4SLinus Torvalds } 24861da177e4SLinus Torvalds 24871da177e4SLinus Torvalds tcp_retrans_try_collapse(sk, skb, cur_mss); 24881da177e4SLinus Torvalds 24891da177e4SLinus Torvalds /* Some Solaris stacks overoptimize and ignore the FIN on a 24901da177e4SLinus Torvalds * retransmit when old data is attached. So strip it off 24911da177e4SLinus Torvalds * since it is cheap to do so and saves bytes on the network. 24921da177e4SLinus Torvalds */ 24931da177e4SLinus Torvalds if (skb->len > 0 && 24944de075e0SEric Dumazet (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && 24951da177e4SLinus Torvalds tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) { 24961da177e4SLinus Torvalds if (!pskb_trim(skb, 0)) { 2497e870a8efSIlpo Järvinen /* Reuse, even though it does some unnecessary work */ 2498e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1, 24994de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags); 25001da177e4SLinus Torvalds skb->ip_summed = CHECKSUM_NONE; 25011da177e4SLinus Torvalds } 25021da177e4SLinus Torvalds } 25031da177e4SLinus Torvalds 25041da177e4SLinus Torvalds /* Make a copy, if the first transmission SKB clone we made 25051da177e4SLinus Torvalds * is still in somebody's hands, else make a clone. 25061da177e4SLinus Torvalds */ 25071da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 25081da177e4SLinus Torvalds 2509117632e6SEric Dumazet /* make sure skb->data is aligned on arches that require it */ 2510117632e6SEric Dumazet if (unlikely(NET_IP_ALIGN && ((unsigned long)skb->data & 3))) { 2511117632e6SEric Dumazet struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER, 2512117632e6SEric Dumazet GFP_ATOMIC); 251393b174adSYuchung Cheng return nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) : 2514117632e6SEric Dumazet -ENOBUFS; 2515117632e6SEric Dumazet } else { 251693b174adSYuchung Cheng return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 2517117632e6SEric Dumazet } 251893b174adSYuchung Cheng } 251993b174adSYuchung Cheng 252093b174adSYuchung Cheng int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) 252193b174adSYuchung Cheng { 252293b174adSYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 252393b174adSYuchung Cheng int err = __tcp_retransmit_skb(sk, skb); 25241da177e4SLinus Torvalds 25251da177e4SLinus Torvalds if (err == 0) { 25261da177e4SLinus Torvalds /* Update global TCP statistics. */ 252781cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); 25281da177e4SLinus Torvalds 25291da177e4SLinus Torvalds tp->total_retrans++; 25301da177e4SLinus Torvalds 25311da177e4SLinus Torvalds #if FASTRETRANS_DEBUG > 0 25321da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 2533e87cc472SJoe Perches net_dbg_ratelimited("retrans_out leaked\n"); 25341da177e4SLinus Torvalds } 25351da177e4SLinus Torvalds #endif 2536b08d6cb2SIlpo Järvinen if (!tp->retrans_out) 2537b08d6cb2SIlpo Järvinen tp->lost_retrans_low = tp->snd_nxt; 25381da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 25391da177e4SLinus Torvalds tp->retrans_out += tcp_skb_pcount(skb); 25401da177e4SLinus Torvalds 25411da177e4SLinus Torvalds /* Save stamp of the first retransmit. */ 25421da177e4SLinus Torvalds if (!tp->retrans_stamp) 25431da177e4SLinus Torvalds tp->retrans_stamp = TCP_SKB_CB(skb)->when; 25441da177e4SLinus Torvalds 2545c24f691bSYuchung Cheng tp->undo_retrans += tcp_skb_pcount(skb); 25461da177e4SLinus Torvalds 25471da177e4SLinus Torvalds /* snd_nxt is stored to detect loss of retransmitted segment, 25481da177e4SLinus Torvalds * see tcp_input.c tcp_sacktag_write_queue(). 25491da177e4SLinus Torvalds */ 25501da177e4SLinus Torvalds TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt; 25511da177e4SLinus Torvalds } 25521da177e4SLinus Torvalds return err; 25531da177e4SLinus Torvalds } 25541da177e4SLinus Torvalds 255567edfef7SAndi Kleen /* Check if we forward retransmits are possible in the current 255667edfef7SAndi Kleen * window/congestion state. 255767edfef7SAndi Kleen */ 2558a2a385d6SEric Dumazet static bool tcp_can_forward_retransmit(struct sock *sk) 2559b5afe7bcSIlpo Järvinen { 2560b5afe7bcSIlpo Järvinen const struct inet_connection_sock *icsk = inet_csk(sk); 2561cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 2562b5afe7bcSIlpo Järvinen 2563b5afe7bcSIlpo Järvinen /* Forward retransmissions are possible only during Recovery. */ 2564b5afe7bcSIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Recovery) 2565a2a385d6SEric Dumazet return false; 2566b5afe7bcSIlpo Järvinen 2567b5afe7bcSIlpo Järvinen /* No forward retransmissions in Reno are possible. */ 2568b5afe7bcSIlpo Järvinen if (tcp_is_reno(tp)) 2569a2a385d6SEric Dumazet return false; 2570b5afe7bcSIlpo Järvinen 2571b5afe7bcSIlpo Järvinen /* Yeah, we have to make difficult choice between forward transmission 2572b5afe7bcSIlpo Järvinen * and retransmission... Both ways have their merits... 2573b5afe7bcSIlpo Järvinen * 2574b5afe7bcSIlpo Järvinen * For now we do not retransmit anything, while we have some new 2575b5afe7bcSIlpo Järvinen * segments to send. In the other cases, follow rule 3 for 2576b5afe7bcSIlpo Järvinen * NextSeg() specified in RFC3517. 2577b5afe7bcSIlpo Järvinen */ 2578b5afe7bcSIlpo Järvinen 2579b5afe7bcSIlpo Järvinen if (tcp_may_send_now(sk)) 2580a2a385d6SEric Dumazet return false; 2581b5afe7bcSIlpo Järvinen 2582a2a385d6SEric Dumazet return true; 2583b5afe7bcSIlpo Järvinen } 2584b5afe7bcSIlpo Järvinen 25851da177e4SLinus Torvalds /* This gets called after a retransmit timeout, and the initially 25861da177e4SLinus Torvalds * retransmitted data is acknowledged. It tries to continue 25871da177e4SLinus Torvalds * resending the rest of the retransmit queue, until either 25881da177e4SLinus Torvalds * we've sent it all or the congestion window limit is reached. 25891da177e4SLinus Torvalds * If doing SACK, the first ACK which comes back for a timeout 25901da177e4SLinus Torvalds * based retransmit packet might feed us FACK information again. 25911da177e4SLinus Torvalds * If so, we use it to avoid unnecessarily retransmissions. 25921da177e4SLinus Torvalds */ 25931da177e4SLinus Torvalds void tcp_xmit_retransmit_queue(struct sock *sk) 25941da177e4SLinus Torvalds { 25956687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 25961da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 25971da177e4SLinus Torvalds struct sk_buff *skb; 25980e1c54c2SIlpo Järvinen struct sk_buff *hole = NULL; 2599618d9f25SIlpo Järvinen u32 last_lost; 260061eb55f4SIlpo Järvinen int mib_idx; 26010e1c54c2SIlpo Järvinen int fwd_rexmitting = 0; 26026a438bbeSStephen Hemminger 260345e77d31SIlpo Järvinen if (!tp->packets_out) 260445e77d31SIlpo Järvinen return; 260545e77d31SIlpo Järvinen 260608ebd172SIlpo Järvinen if (!tp->lost_out) 260708ebd172SIlpo Järvinen tp->retransmit_high = tp->snd_una; 260808ebd172SIlpo Järvinen 2609618d9f25SIlpo Järvinen if (tp->retransmit_skb_hint) { 26106a438bbeSStephen Hemminger skb = tp->retransmit_skb_hint; 2611618d9f25SIlpo Järvinen last_lost = TCP_SKB_CB(skb)->end_seq; 2612618d9f25SIlpo Järvinen if (after(last_lost, tp->retransmit_high)) 2613618d9f25SIlpo Järvinen last_lost = tp->retransmit_high; 2614618d9f25SIlpo Järvinen } else { 2615fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 2616618d9f25SIlpo Järvinen last_lost = tp->snd_una; 2617618d9f25SIlpo Järvinen } 26181da177e4SLinus Torvalds 2619fe067e8aSDavid S. Miller tcp_for_write_queue_from(skb, sk) { 26201da177e4SLinus Torvalds __u8 sacked = TCP_SKB_CB(skb)->sacked; 26211da177e4SLinus Torvalds 2622fe067e8aSDavid S. Miller if (skb == tcp_send_head(sk)) 2623fe067e8aSDavid S. Miller break; 26246a438bbeSStephen Hemminger /* we could do better than to assign each time */ 26250e1c54c2SIlpo Järvinen if (hole == NULL) 26266a438bbeSStephen Hemminger tp->retransmit_skb_hint = skb; 26276a438bbeSStephen Hemminger 26281da177e4SLinus Torvalds /* Assume this retransmit will generate 26291da177e4SLinus Torvalds * only one packet for congestion window 26301da177e4SLinus Torvalds * calculation purposes. This works because 26311da177e4SLinus Torvalds * tcp_retransmit_skb() will chop up the 26321da177e4SLinus Torvalds * packet to be MSS sized and all the 26331da177e4SLinus Torvalds * packet counting works out. 26341da177e4SLinus Torvalds */ 26351da177e4SLinus Torvalds if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) 26361da177e4SLinus Torvalds return; 26370e1c54c2SIlpo Järvinen 26380e1c54c2SIlpo Järvinen if (fwd_rexmitting) { 26390e1c54c2SIlpo Järvinen begin_fwd: 26400e1c54c2SIlpo Järvinen if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp))) 2641006f582cSIlpo Järvinen break; 26420e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFORWARDRETRANS; 26430e1c54c2SIlpo Järvinen 26440e1c54c2SIlpo Järvinen } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) { 2645618d9f25SIlpo Järvinen tp->retransmit_high = last_lost; 26460e1c54c2SIlpo Järvinen if (!tcp_can_forward_retransmit(sk)) 26470e1c54c2SIlpo Järvinen break; 26480e1c54c2SIlpo Järvinen /* Backtrack if necessary to non-L'ed skb */ 26490e1c54c2SIlpo Järvinen if (hole != NULL) { 26500e1c54c2SIlpo Järvinen skb = hole; 26510e1c54c2SIlpo Järvinen hole = NULL; 26520e1c54c2SIlpo Järvinen } 26530e1c54c2SIlpo Järvinen fwd_rexmitting = 1; 26540e1c54c2SIlpo Järvinen goto begin_fwd; 26550e1c54c2SIlpo Järvinen 26560e1c54c2SIlpo Järvinen } else if (!(sacked & TCPCB_LOST)) { 2657ac11ba75SIlpo Järvinen if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) 26580e1c54c2SIlpo Järvinen hole = skb; 265961eb55f4SIlpo Järvinen continue; 26601da177e4SLinus Torvalds 26610e1c54c2SIlpo Järvinen } else { 2662618d9f25SIlpo Järvinen last_lost = TCP_SKB_CB(skb)->end_seq; 26630e1c54c2SIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Loss) 26640e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFASTRETRANS; 26650e1c54c2SIlpo Järvinen else 26660e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; 26670e1c54c2SIlpo Järvinen } 26680e1c54c2SIlpo Järvinen 26690e1c54c2SIlpo Järvinen if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) 267061eb55f4SIlpo Järvinen continue; 267140b215e5SPavel Emelyanov 267209e9b813SEric Dumazet if (tcp_retransmit_skb(sk, skb)) { 267309e9b813SEric Dumazet NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL); 26741da177e4SLinus Torvalds return; 267509e9b813SEric Dumazet } 2676de0744afSPavel Emelyanov NET_INC_STATS_BH(sock_net(sk), mib_idx); 26771da177e4SLinus Torvalds 2678684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 2679a262f0cdSNandita Dukkipati tp->prr_out += tcp_skb_pcount(skb); 2680a262f0cdSNandita Dukkipati 2681fe067e8aSDavid S. Miller if (skb == tcp_write_queue_head(sk)) 2682463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 26833f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, 26843f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 26851da177e4SLinus Torvalds } 26861da177e4SLinus Torvalds } 26871da177e4SLinus Torvalds 26881da177e4SLinus Torvalds /* Send a fin. The caller locks the socket for us. This cannot be 26891da177e4SLinus Torvalds * allowed to fail queueing a FIN frame under any circumstances. 26901da177e4SLinus Torvalds */ 26911da177e4SLinus Torvalds void tcp_send_fin(struct sock *sk) 26921da177e4SLinus Torvalds { 26931da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2694fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_write_queue_tail(sk); 26951da177e4SLinus Torvalds int mss_now; 26961da177e4SLinus Torvalds 26971da177e4SLinus Torvalds /* Optimization, tack on the FIN if we have a queue of 26981da177e4SLinus Torvalds * unsent frames. But be careful about outgoing SACKS 26991da177e4SLinus Torvalds * and IP options. 27001da177e4SLinus Torvalds */ 27010c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 27021da177e4SLinus Torvalds 2703fe067e8aSDavid S. Miller if (tcp_send_head(sk) != NULL) { 27044de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_FIN; 27051da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq++; 27061da177e4SLinus Torvalds tp->write_seq++; 27071da177e4SLinus Torvalds } else { 27081da177e4SLinus Torvalds /* Socket is locked, keep trying until memory is available. */ 27091da177e4SLinus Torvalds for (;;) { 2710aa133076SWu Fengguang skb = alloc_skb_fclone(MAX_TCP_HEADER, 2711aa133076SWu Fengguang sk->sk_allocation); 27121da177e4SLinus Torvalds if (skb) 27131da177e4SLinus Torvalds break; 27141da177e4SLinus Torvalds yield(); 27151da177e4SLinus Torvalds } 27161da177e4SLinus Torvalds 27171da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 27181da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 27191da177e4SLinus Torvalds /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 2720e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tp->write_seq, 2721a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_FIN); 27221da177e4SLinus Torvalds tcp_queue_skb(sk, skb); 27231da177e4SLinus Torvalds } 27249e412ba7SIlpo Järvinen __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF); 27251da177e4SLinus Torvalds } 27261da177e4SLinus Torvalds 27271da177e4SLinus Torvalds /* We get here when a process closes a file descriptor (either due to 27281da177e4SLinus Torvalds * an explicit close() or as a byproduct of exit()'ing) and there 27291da177e4SLinus Torvalds * was unread data in the receive queue. This behavior is recommended 273065bb723cSGerrit Renker * by RFC 2525, section 2.17. -DaveM 27311da177e4SLinus Torvalds */ 2732dd0fc66fSAl Viro void tcp_send_active_reset(struct sock *sk, gfp_t priority) 27331da177e4SLinus Torvalds { 27341da177e4SLinus Torvalds struct sk_buff *skb; 27351da177e4SLinus Torvalds 27361da177e4SLinus Torvalds /* NOTE: No TCP options attached and we never retransmit this. */ 27371da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, priority); 27381da177e4SLinus Torvalds if (!skb) { 27394e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 27401da177e4SLinus Torvalds return; 27411da177e4SLinus Torvalds } 27421da177e4SLinus Torvalds 27431da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 27441da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 2745e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), 2746a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_RST); 27471da177e4SLinus Torvalds /* Send it off. */ 27481da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2749dfb4b9dcSDavid S. Miller if (tcp_transmit_skb(sk, skb, 0, priority)) 27504e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 275126af65cbSSridhar Samudrala 275281cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); 27531da177e4SLinus Torvalds } 27541da177e4SLinus Torvalds 275567edfef7SAndi Kleen /* Send a crossed SYN-ACK during socket establishment. 275667edfef7SAndi Kleen * WARNING: This routine must only be called when we have already sent 27571da177e4SLinus Torvalds * a SYN packet that crossed the incoming SYN that caused this routine 27581da177e4SLinus Torvalds * to get called. If this assumption fails then the initial rcv_wnd 27591da177e4SLinus Torvalds * and rcv_wscale values will not be correct. 27601da177e4SLinus Torvalds */ 27611da177e4SLinus Torvalds int tcp_send_synack(struct sock *sk) 27621da177e4SLinus Torvalds { 27631da177e4SLinus Torvalds struct sk_buff *skb; 27641da177e4SLinus Torvalds 2765fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 27664de075e0SEric Dumazet if (skb == NULL || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 276791df42beSJoe Perches pr_debug("%s: wrong queue state\n", __func__); 27681da177e4SLinus Torvalds return -EFAULT; 27691da177e4SLinus Torvalds } 27704de075e0SEric Dumazet if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { 27711da177e4SLinus Torvalds if (skb_cloned(skb)) { 27721da177e4SLinus Torvalds struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC); 27731da177e4SLinus Torvalds if (nskb == NULL) 27741da177e4SLinus Torvalds return -ENOMEM; 2775fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 27761da177e4SLinus Torvalds skb_header_release(nskb); 2777fe067e8aSDavid S. Miller __tcp_add_write_queue_head(sk, nskb); 27783ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 27793ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 27803ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 27811da177e4SLinus Torvalds skb = nskb; 27821da177e4SLinus Torvalds } 27831da177e4SLinus Torvalds 27844de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; 27851da177e4SLinus Torvalds TCP_ECN_send_synack(tcp_sk(sk), skb); 27861da177e4SLinus Torvalds } 27871da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2788dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 27891da177e4SLinus Torvalds } 27901da177e4SLinus Torvalds 27914aea39c1SEric Dumazet /** 27924aea39c1SEric Dumazet * tcp_make_synack - Prepare a SYN-ACK. 27934aea39c1SEric Dumazet * sk: listener socket 27944aea39c1SEric Dumazet * dst: dst entry attached to the SYNACK 27954aea39c1SEric Dumazet * req: request_sock pointer 27964aea39c1SEric Dumazet * rvp: request_values pointer 27974aea39c1SEric Dumazet * 27984aea39c1SEric Dumazet * Allocate one skb and build a SYNACK packet. 27994aea39c1SEric Dumazet * @dst is consumed : Caller should not use it again. 28004aea39c1SEric Dumazet */ 28011da177e4SLinus Torvalds struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst, 2802e6b4d113SWilliam Allen Simpson struct request_sock *req, 28038336886fSJerry Chu struct request_values *rvp, 28048336886fSJerry Chu struct tcp_fastopen_cookie *foc) 28051da177e4SLinus Torvalds { 2806bd0388aeSWilliam Allen Simpson struct tcp_out_options opts; 28074957faadSWilliam Allen Simpson struct tcp_extend_values *xvp = tcp_xv(rvp); 28082e6599cbSArnaldo Carvalho de Melo struct inet_request_sock *ireq = inet_rsk(req); 28091da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 281028b2774aSEric Dumazet const struct tcp_cookie_values *cvp = tp->cookie_values; 28111da177e4SLinus Torvalds struct tcphdr *th; 28121da177e4SLinus Torvalds struct sk_buff *skb; 2813cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 2814bd0388aeSWilliam Allen Simpson int tcp_header_size; 2815f5fff5dcSTom Quetchenbach int mss; 281628b2774aSEric Dumazet int s_data_desired = 0; 28171da177e4SLinus Torvalds 281828b2774aSEric Dumazet if (cvp != NULL && cvp->s_data_constant && cvp->s_data_desired) 281928b2774aSEric Dumazet s_data_desired = cvp->s_data_desired; 282099a1dec7SMel Gorman skb = alloc_skb(MAX_TCP_HEADER + 15 + s_data_desired, 282199a1dec7SMel Gorman sk_gfp_atomic(sk, GFP_ATOMIC)); 28224aea39c1SEric Dumazet if (unlikely(!skb)) { 28234aea39c1SEric Dumazet dst_release(dst); 28241da177e4SLinus Torvalds return NULL; 28254aea39c1SEric Dumazet } 28261da177e4SLinus Torvalds /* Reserve space for headers. */ 28271da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 28281da177e4SLinus Torvalds 28294aea39c1SEric Dumazet skb_dst_set(skb, dst); 28301da177e4SLinus Torvalds 28310dbaee3bSDavid S. Miller mss = dst_metric_advmss(dst); 2832f5fff5dcSTom Quetchenbach if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss) 2833f5fff5dcSTom Quetchenbach mss = tp->rx_opt.user_mss; 2834f5fff5dcSTom Quetchenbach 283533ad798cSAdam Langley if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */ 283633ad798cSAdam Langley __u8 rcv_wscale; 283733ad798cSAdam Langley /* Set this up on the first call only */ 283833ad798cSAdam Langley req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW); 2839e88c64f0SHagen Paul Pfeifer 2840e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 2841e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 2842e88c64f0SHagen Paul Pfeifer (req->window_clamp > tcp_full_space(sk) || req->window_clamp == 0)) 2843e88c64f0SHagen Paul Pfeifer req->window_clamp = tcp_full_space(sk); 2844e88c64f0SHagen Paul Pfeifer 284533ad798cSAdam Langley /* tcp_full_space because it is guaranteed to be the first packet */ 284633ad798cSAdam Langley tcp_select_initial_window(tcp_full_space(sk), 2847f5fff5dcSTom Quetchenbach mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0), 284833ad798cSAdam Langley &req->rcv_wnd, 284933ad798cSAdam Langley &req->window_clamp, 285033ad798cSAdam Langley ireq->wscale_ok, 285131d12926Slaurent chavey &rcv_wscale, 285231d12926Slaurent chavey dst_metric(dst, RTAX_INITRWND)); 285333ad798cSAdam Langley ireq->rcv_wscale = rcv_wscale; 285433ad798cSAdam Langley } 2855cfb6eeb4SYOSHIFUJI Hideaki 285633ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 28578b5f12d0SFlorian Westphal #ifdef CONFIG_SYN_COOKIES 28588b5f12d0SFlorian Westphal if (unlikely(req->cookie_ts)) 28598b5f12d0SFlorian Westphal TCP_SKB_CB(skb)->when = cookie_init_timestamp(req); 28608b5f12d0SFlorian Westphal else 28618b5f12d0SFlorian Westphal #endif 286233ad798cSAdam Langley TCP_SKB_CB(skb)->when = tcp_time_stamp; 2863f5fff5dcSTom Quetchenbach tcp_header_size = tcp_synack_options(sk, req, mss, 28648336886fSJerry Chu skb, &opts, &md5, xvp, foc) 28654957faadSWilliam Allen Simpson + sizeof(*th); 286633ad798cSAdam Langley 2867aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 2868aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 28691da177e4SLinus Torvalds 2870aa8223c7SArnaldo Carvalho de Melo th = tcp_hdr(skb); 28711da177e4SLinus Torvalds memset(th, 0, sizeof(struct tcphdr)); 28721da177e4SLinus Torvalds th->syn = 1; 28731da177e4SLinus Torvalds th->ack = 1; 28741da177e4SLinus Torvalds TCP_ECN_make_synack(req, th); 2875a3116ac5SKOVACS Krisztian th->source = ireq->loc_port; 28762e6599cbSArnaldo Carvalho de Melo th->dest = ireq->rmt_port; 2877e870a8efSIlpo Järvinen /* Setting of flags are superfluous here for callers (and ECE is 2878e870a8efSIlpo Järvinen * not even correctly set) 2879e870a8efSIlpo Järvinen */ 2880e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn, 2881a3433f35SChangli Gao TCPHDR_SYN | TCPHDR_ACK); 28824957faadSWilliam Allen Simpson 28834957faadSWilliam Allen Simpson if (OPTION_COOKIE_EXTENSION & opts.options) { 288428b2774aSEric Dumazet if (s_data_desired) { 288528b2774aSEric Dumazet u8 *buf = skb_put(skb, s_data_desired); 28864957faadSWilliam Allen Simpson 28874957faadSWilliam Allen Simpson /* copy data directly from the listening socket. */ 288828b2774aSEric Dumazet memcpy(buf, cvp->s_data_payload, s_data_desired); 288928b2774aSEric Dumazet TCP_SKB_CB(skb)->end_seq += s_data_desired; 28904957faadSWilliam Allen Simpson } 28914957faadSWilliam Allen Simpson 28924957faadSWilliam Allen Simpson if (opts.hash_size > 0) { 28934957faadSWilliam Allen Simpson __u32 workspace[SHA_WORKSPACE_WORDS]; 28944957faadSWilliam Allen Simpson u32 *mess = &xvp->cookie_bakery[COOKIE_DIGEST_WORDS]; 28954957faadSWilliam Allen Simpson u32 *tail = &mess[COOKIE_MESSAGE_WORDS-1]; 28964957faadSWilliam Allen Simpson 28974957faadSWilliam Allen Simpson /* Secret recipe depends on the Timestamp, (future) 28984957faadSWilliam Allen Simpson * Sequence and Acknowledgment Numbers, Initiator 28994957faadSWilliam Allen Simpson * Cookie, and others handled by IP variant caller. 29004957faadSWilliam Allen Simpson */ 29014957faadSWilliam Allen Simpson *tail-- ^= opts.tsval; 29024957faadSWilliam Allen Simpson *tail-- ^= tcp_rsk(req)->rcv_isn + 1; 29034957faadSWilliam Allen Simpson *tail-- ^= TCP_SKB_CB(skb)->seq + 1; 29044957faadSWilliam Allen Simpson 29054957faadSWilliam Allen Simpson /* recommended */ 29060eae88f3SEric Dumazet *tail-- ^= (((__force u32)th->dest << 16) | (__force u32)th->source); 2907f9a2e69eSDavid S. Miller *tail-- ^= (u32)(unsigned long)cvp; /* per sockopt */ 29084957faadSWilliam Allen Simpson 29094957faadSWilliam Allen Simpson sha_transform((__u32 *)&xvp->cookie_bakery[0], 29104957faadSWilliam Allen Simpson (char *)mess, 29114957faadSWilliam Allen Simpson &workspace[0]); 29124957faadSWilliam Allen Simpson opts.hash_location = 29134957faadSWilliam Allen Simpson (__u8 *)&xvp->cookie_bakery[0]; 29144957faadSWilliam Allen Simpson } 29154957faadSWilliam Allen Simpson } 29164957faadSWilliam Allen Simpson 29171da177e4SLinus Torvalds th->seq = htonl(TCP_SKB_CB(skb)->seq); 29188336886fSJerry Chu /* XXX data is queued and acked as is. No buffer/window check */ 29198336886fSJerry Chu th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt); 29201da177e4SLinus Torvalds 29211da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 2922600ff0c2SIlpo Järvinen th->window = htons(min(req->rcv_wnd, 65535U)); 2923bd0388aeSWilliam Allen Simpson tcp_options_write((__be32 *)(th + 1), tp, &opts); 29241da177e4SLinus Torvalds th->doff = (tcp_header_size >> 2); 2925aa2ea058STom Herbert TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, tcp_skb_pcount(skb)); 2926cfb6eeb4SYOSHIFUJI Hideaki 2927cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 2928cfb6eeb4SYOSHIFUJI Hideaki /* Okay, we have all we need - do the md5 hash if needed */ 2929cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 2930bd0388aeSWilliam Allen Simpson tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, 293149a72dfbSAdam Langley md5, NULL, req, skb); 2932cfb6eeb4SYOSHIFUJI Hideaki } 2933cfb6eeb4SYOSHIFUJI Hideaki #endif 2934cfb6eeb4SYOSHIFUJI Hideaki 29351da177e4SLinus Torvalds return skb; 29361da177e4SLinus Torvalds } 29374bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_make_synack); 29381da177e4SLinus Torvalds 293967edfef7SAndi Kleen /* Do all connect socket setups that can be done AF independent. */ 2940370816aeSPavel Emelyanov void tcp_connect_init(struct sock *sk) 29411da177e4SLinus Torvalds { 2942cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 29431da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 29441da177e4SLinus Torvalds __u8 rcv_wscale; 29451da177e4SLinus Torvalds 29461da177e4SLinus Torvalds /* We'll fix this up when we get a response from the other end. 29471da177e4SLinus Torvalds * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 29481da177e4SLinus Torvalds */ 29491da177e4SLinus Torvalds tp->tcp_header_len = sizeof(struct tcphdr) + 2950bb5b7c11SDavid S. Miller (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0); 29511da177e4SLinus Torvalds 2952cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 2953cfb6eeb4SYOSHIFUJI Hideaki if (tp->af_specific->md5_lookup(sk, sk) != NULL) 2954cfb6eeb4SYOSHIFUJI Hideaki tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 2955cfb6eeb4SYOSHIFUJI Hideaki #endif 2956cfb6eeb4SYOSHIFUJI Hideaki 29571da177e4SLinus Torvalds /* If user gave his TCP_MAXSEG, record it to clamp */ 29581da177e4SLinus Torvalds if (tp->rx_opt.user_mss) 29591da177e4SLinus Torvalds tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 29601da177e4SLinus Torvalds tp->max_window = 0; 29615d424d5aSJohn Heffner tcp_mtup_init(sk); 29621da177e4SLinus Torvalds tcp_sync_mss(sk, dst_mtu(dst)); 29631da177e4SLinus Torvalds 29641da177e4SLinus Torvalds if (!tp->window_clamp) 29651da177e4SLinus Torvalds tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 29660dbaee3bSDavid S. Miller tp->advmss = dst_metric_advmss(dst); 2967f5fff5dcSTom Quetchenbach if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss) 2968f5fff5dcSTom Quetchenbach tp->advmss = tp->rx_opt.user_mss; 2969f5fff5dcSTom Quetchenbach 29701da177e4SLinus Torvalds tcp_initialize_rcv_mss(sk); 29711da177e4SLinus Torvalds 2972e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 2973e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 2974e88c64f0SHagen Paul Pfeifer (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) 2975e88c64f0SHagen Paul Pfeifer tp->window_clamp = tcp_full_space(sk); 2976e88c64f0SHagen Paul Pfeifer 29771da177e4SLinus Torvalds tcp_select_initial_window(tcp_full_space(sk), 29781da177e4SLinus Torvalds tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 29791da177e4SLinus Torvalds &tp->rcv_wnd, 29801da177e4SLinus Torvalds &tp->window_clamp, 2981bb5b7c11SDavid S. Miller sysctl_tcp_window_scaling, 298231d12926Slaurent chavey &rcv_wscale, 298331d12926Slaurent chavey dst_metric(dst, RTAX_INITRWND)); 29841da177e4SLinus Torvalds 29851da177e4SLinus Torvalds tp->rx_opt.rcv_wscale = rcv_wscale; 29861da177e4SLinus Torvalds tp->rcv_ssthresh = tp->rcv_wnd; 29871da177e4SLinus Torvalds 29881da177e4SLinus Torvalds sk->sk_err = 0; 29891da177e4SLinus Torvalds sock_reset_flag(sk, SOCK_DONE); 29901da177e4SLinus Torvalds tp->snd_wnd = 0; 2991ee7537b6SHantzis Fotis tcp_init_wl(tp, 0); 29921da177e4SLinus Torvalds tp->snd_una = tp->write_seq; 29931da177e4SLinus Torvalds tp->snd_sml = tp->write_seq; 299433f5f57eSIlpo Järvinen tp->snd_up = tp->write_seq; 2995370816aeSPavel Emelyanov tp->snd_nxt = tp->write_seq; 2996ee995283SPavel Emelyanov 2997ee995283SPavel Emelyanov if (likely(!tp->repair)) 29981da177e4SLinus Torvalds tp->rcv_nxt = 0; 2999ee995283SPavel Emelyanov tp->rcv_wup = tp->rcv_nxt; 3000ee995283SPavel Emelyanov tp->copied_seq = tp->rcv_nxt; 30011da177e4SLinus Torvalds 3002463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; 3003463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_retransmits = 0; 30041da177e4SLinus Torvalds tcp_clear_retrans(tp); 30051da177e4SLinus Torvalds } 30061da177e4SLinus Torvalds 3007783237e8SYuchung Cheng static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) 3008783237e8SYuchung Cheng { 3009783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3010783237e8SYuchung Cheng struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 3011783237e8SYuchung Cheng 3012783237e8SYuchung Cheng tcb->end_seq += skb->len; 3013783237e8SYuchung Cheng skb_header_release(skb); 3014783237e8SYuchung Cheng __tcp_add_write_queue_tail(sk, skb); 3015783237e8SYuchung Cheng sk->sk_wmem_queued += skb->truesize; 3016783237e8SYuchung Cheng sk_mem_charge(sk, skb->truesize); 3017783237e8SYuchung Cheng tp->write_seq = tcb->end_seq; 3018783237e8SYuchung Cheng tp->packets_out += tcp_skb_pcount(skb); 3019783237e8SYuchung Cheng } 3020783237e8SYuchung Cheng 3021783237e8SYuchung Cheng /* Build and send a SYN with data and (cached) Fast Open cookie. However, 3022783237e8SYuchung Cheng * queue a data-only packet after the regular SYN, such that regular SYNs 3023783237e8SYuchung Cheng * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges 3024783237e8SYuchung Cheng * only the SYN sequence, the data are retransmitted in the first ACK. 3025783237e8SYuchung Cheng * If cookie is not cached or other error occurs, falls back to send a 3026783237e8SYuchung Cheng * regular SYN with Fast Open cookie request option. 3027783237e8SYuchung Cheng */ 3028783237e8SYuchung Cheng static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) 3029783237e8SYuchung Cheng { 3030783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3031783237e8SYuchung Cheng struct tcp_fastopen_request *fo = tp->fastopen_req; 3032aab48743SYuchung Cheng int syn_loss = 0, space, i, err = 0, iovlen = fo->data->msg_iovlen; 3033783237e8SYuchung Cheng struct sk_buff *syn_data = NULL, *data; 3034aab48743SYuchung Cheng unsigned long last_syn_loss = 0; 3035783237e8SYuchung Cheng 303667da22d2SYuchung Cheng tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ 3037aab48743SYuchung Cheng tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie, 3038aab48743SYuchung Cheng &syn_loss, &last_syn_loss); 3039aab48743SYuchung Cheng /* Recurring FO SYN losses: revert to regular handshake temporarily */ 3040aab48743SYuchung Cheng if (syn_loss > 1 && 3041aab48743SYuchung Cheng time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) { 3042aab48743SYuchung Cheng fo->cookie.len = -1; 3043aab48743SYuchung Cheng goto fallback; 3044aab48743SYuchung Cheng } 3045aab48743SYuchung Cheng 304667da22d2SYuchung Cheng if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE) 304767da22d2SYuchung Cheng fo->cookie.len = -1; 304867da22d2SYuchung Cheng else if (fo->cookie.len <= 0) 3049783237e8SYuchung Cheng goto fallback; 3050783237e8SYuchung Cheng 3051783237e8SYuchung Cheng /* MSS for SYN-data is based on cached MSS and bounded by PMTU and 3052783237e8SYuchung Cheng * user-MSS. Reserve maximum option space for middleboxes that add 3053783237e8SYuchung Cheng * private TCP options. The cost is reduced data space in SYN :( 3054783237e8SYuchung Cheng */ 3055783237e8SYuchung Cheng if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp) 3056783237e8SYuchung Cheng tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 30571b63edd6SYuchung Cheng space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) - 3058783237e8SYuchung Cheng MAX_TCP_OPTION_SPACE; 3059783237e8SYuchung Cheng 3060783237e8SYuchung Cheng syn_data = skb_copy_expand(syn, skb_headroom(syn), space, 3061783237e8SYuchung Cheng sk->sk_allocation); 3062783237e8SYuchung Cheng if (syn_data == NULL) 3063783237e8SYuchung Cheng goto fallback; 3064783237e8SYuchung Cheng 3065783237e8SYuchung Cheng for (i = 0; i < iovlen && syn_data->len < space; ++i) { 3066783237e8SYuchung Cheng struct iovec *iov = &fo->data->msg_iov[i]; 3067783237e8SYuchung Cheng unsigned char __user *from = iov->iov_base; 3068783237e8SYuchung Cheng int len = iov->iov_len; 3069783237e8SYuchung Cheng 3070783237e8SYuchung Cheng if (syn_data->len + len > space) 3071783237e8SYuchung Cheng len = space - syn_data->len; 3072783237e8SYuchung Cheng else if (i + 1 == iovlen) 3073783237e8SYuchung Cheng /* No more data pending in inet_wait_for_connect() */ 3074783237e8SYuchung Cheng fo->data = NULL; 3075783237e8SYuchung Cheng 3076783237e8SYuchung Cheng if (skb_add_data(syn_data, from, len)) 3077783237e8SYuchung Cheng goto fallback; 3078783237e8SYuchung Cheng } 3079783237e8SYuchung Cheng 3080783237e8SYuchung Cheng /* Queue a data-only packet after the regular SYN for retransmission */ 3081783237e8SYuchung Cheng data = pskb_copy(syn_data, sk->sk_allocation); 3082783237e8SYuchung Cheng if (data == NULL) 3083783237e8SYuchung Cheng goto fallback; 3084783237e8SYuchung Cheng TCP_SKB_CB(data)->seq++; 3085783237e8SYuchung Cheng TCP_SKB_CB(data)->tcp_flags &= ~TCPHDR_SYN; 3086783237e8SYuchung Cheng TCP_SKB_CB(data)->tcp_flags = (TCPHDR_ACK|TCPHDR_PSH); 3087783237e8SYuchung Cheng tcp_connect_queue_skb(sk, data); 3088783237e8SYuchung Cheng fo->copied = data->len; 3089783237e8SYuchung Cheng 3090783237e8SYuchung Cheng if (tcp_transmit_skb(sk, syn_data, 0, sk->sk_allocation) == 0) { 309167da22d2SYuchung Cheng tp->syn_data = (fo->copied > 0); 3092783237e8SYuchung Cheng NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE); 3093783237e8SYuchung Cheng goto done; 3094783237e8SYuchung Cheng } 3095783237e8SYuchung Cheng syn_data = NULL; 3096783237e8SYuchung Cheng 3097783237e8SYuchung Cheng fallback: 3098783237e8SYuchung Cheng /* Send a regular SYN with Fast Open cookie request option */ 3099783237e8SYuchung Cheng if (fo->cookie.len > 0) 3100783237e8SYuchung Cheng fo->cookie.len = 0; 3101783237e8SYuchung Cheng err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); 3102783237e8SYuchung Cheng if (err) 3103783237e8SYuchung Cheng tp->syn_fastopen = 0; 3104783237e8SYuchung Cheng kfree_skb(syn_data); 3105783237e8SYuchung Cheng done: 3106783237e8SYuchung Cheng fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ 3107783237e8SYuchung Cheng return err; 3108783237e8SYuchung Cheng } 3109783237e8SYuchung Cheng 311067edfef7SAndi Kleen /* Build a SYN and send it off. */ 31111da177e4SLinus Torvalds int tcp_connect(struct sock *sk) 31121da177e4SLinus Torvalds { 31131da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 31141da177e4SLinus Torvalds struct sk_buff *buff; 3115ee586811SEric Paris int err; 31161da177e4SLinus Torvalds 31171da177e4SLinus Torvalds tcp_connect_init(sk); 31181da177e4SLinus Torvalds 31192b916477SAndrey Vagin if (unlikely(tp->repair)) { 31202b916477SAndrey Vagin tcp_finish_connect(sk, NULL); 31212b916477SAndrey Vagin return 0; 31222b916477SAndrey Vagin } 31232b916477SAndrey Vagin 3124d179cd12SDavid S. Miller buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation); 31251da177e4SLinus Torvalds if (unlikely(buff == NULL)) 31261da177e4SLinus Torvalds return -ENOBUFS; 31271da177e4SLinus Torvalds 31281da177e4SLinus Torvalds /* Reserve space for headers. */ 31291da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 31301da177e4SLinus Torvalds 3131a3433f35SChangli Gao tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); 3132783237e8SYuchung Cheng tp->retrans_stamp = TCP_SKB_CB(buff)->when = tcp_time_stamp; 3133783237e8SYuchung Cheng tcp_connect_queue_skb(sk, buff); 3134e870a8efSIlpo Järvinen TCP_ECN_send_syn(sk, buff); 31351da177e4SLinus Torvalds 3136783237e8SYuchung Cheng /* Send off SYN; include data in Fast Open. */ 3137783237e8SYuchung Cheng err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : 3138783237e8SYuchung Cheng tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); 3139ee586811SEric Paris if (err == -ECONNREFUSED) 3140ee586811SEric Paris return err; 3141bd37a088SWei Yongjun 3142bd37a088SWei Yongjun /* We change tp->snd_nxt after the tcp_transmit_skb() call 3143bd37a088SWei Yongjun * in order to make this packet get counted in tcpOutSegs. 3144bd37a088SWei Yongjun */ 3145bd37a088SWei Yongjun tp->snd_nxt = tp->write_seq; 3146bd37a088SWei Yongjun tp->pushed_seq = tp->write_seq; 314781cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); 31481da177e4SLinus Torvalds 31491da177e4SLinus Torvalds /* Timer for repeating the SYN until an answer. */ 31503f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 31513f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 31521da177e4SLinus Torvalds return 0; 31531da177e4SLinus Torvalds } 31544bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_connect); 31551da177e4SLinus Torvalds 31561da177e4SLinus Torvalds /* Send out a delayed ack, the caller does the policy checking 31571da177e4SLinus Torvalds * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 31581da177e4SLinus Torvalds * for details. 31591da177e4SLinus Torvalds */ 31601da177e4SLinus Torvalds void tcp_send_delayed_ack(struct sock *sk) 31611da177e4SLinus Torvalds { 3162463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 3163463c84b9SArnaldo Carvalho de Melo int ato = icsk->icsk_ack.ato; 31641da177e4SLinus Torvalds unsigned long timeout; 31651da177e4SLinus Torvalds 31661da177e4SLinus Torvalds if (ato > TCP_DELACK_MIN) { 3167463c84b9SArnaldo Carvalho de Melo const struct tcp_sock *tp = tcp_sk(sk); 31681da177e4SLinus Torvalds int max_ato = HZ / 2; 31691da177e4SLinus Torvalds 3170056834d9SIlpo Järvinen if (icsk->icsk_ack.pingpong || 3171056834d9SIlpo Järvinen (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 31721da177e4SLinus Torvalds max_ato = TCP_DELACK_MAX; 31731da177e4SLinus Torvalds 31741da177e4SLinus Torvalds /* Slow path, intersegment interval is "high". */ 31751da177e4SLinus Torvalds 31761da177e4SLinus Torvalds /* If some rtt estimate is known, use it to bound delayed ack. 3177463c84b9SArnaldo Carvalho de Melo * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 31781da177e4SLinus Torvalds * directly. 31791da177e4SLinus Torvalds */ 31801da177e4SLinus Torvalds if (tp->srtt) { 31811da177e4SLinus Torvalds int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN); 31821da177e4SLinus Torvalds 31831da177e4SLinus Torvalds if (rtt < max_ato) 31841da177e4SLinus Torvalds max_ato = rtt; 31851da177e4SLinus Torvalds } 31861da177e4SLinus Torvalds 31871da177e4SLinus Torvalds ato = min(ato, max_ato); 31881da177e4SLinus Torvalds } 31891da177e4SLinus Torvalds 31901da177e4SLinus Torvalds /* Stay within the limit we were given */ 31911da177e4SLinus Torvalds timeout = jiffies + ato; 31921da177e4SLinus Torvalds 31931da177e4SLinus Torvalds /* Use new timeout only if there wasn't a older one earlier. */ 3194463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 31951da177e4SLinus Torvalds /* If delack timer was blocked or is about to expire, 31961da177e4SLinus Torvalds * send ACK now. 31971da177e4SLinus Torvalds */ 3198463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.blocked || 3199463c84b9SArnaldo Carvalho de Melo time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 32001da177e4SLinus Torvalds tcp_send_ack(sk); 32011da177e4SLinus Torvalds return; 32021da177e4SLinus Torvalds } 32031da177e4SLinus Torvalds 3204463c84b9SArnaldo Carvalho de Melo if (!time_before(timeout, icsk->icsk_ack.timeout)) 3205463c84b9SArnaldo Carvalho de Melo timeout = icsk->icsk_ack.timeout; 32061da177e4SLinus Torvalds } 3207463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 3208463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.timeout = timeout; 3209463c84b9SArnaldo Carvalho de Melo sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 32101da177e4SLinus Torvalds } 32111da177e4SLinus Torvalds 32121da177e4SLinus Torvalds /* This routine sends an ack and also updates the window. */ 32131da177e4SLinus Torvalds void tcp_send_ack(struct sock *sk) 32141da177e4SLinus Torvalds { 32151da177e4SLinus Torvalds struct sk_buff *buff; 32161da177e4SLinus Torvalds 3217058dc334SIlpo Järvinen /* If we have been reset, we may not send again. */ 3218058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3219058dc334SIlpo Järvinen return; 3220058dc334SIlpo Järvinen 32211da177e4SLinus Torvalds /* We are not putting this on the write queue, so 32221da177e4SLinus Torvalds * tcp_transmit_skb() will set the ownership to this 32231da177e4SLinus Torvalds * sock. 32241da177e4SLinus Torvalds */ 322599a1dec7SMel Gorman buff = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC)); 32261da177e4SLinus Torvalds if (buff == NULL) { 3227463c84b9SArnaldo Carvalho de Melo inet_csk_schedule_ack(sk); 3228463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; 32293f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 32303f421baaSArnaldo Carvalho de Melo TCP_DELACK_MAX, TCP_RTO_MAX); 32311da177e4SLinus Torvalds return; 32321da177e4SLinus Torvalds } 32331da177e4SLinus Torvalds 32341da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 32351da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 3236a3433f35SChangli Gao tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); 32371da177e4SLinus Torvalds 32381da177e4SLinus Torvalds /* Send it off, this clears delayed acks for us. */ 32391da177e4SLinus Torvalds TCP_SKB_CB(buff)->when = tcp_time_stamp; 324099a1dec7SMel Gorman tcp_transmit_skb(sk, buff, 0, sk_gfp_atomic(sk, GFP_ATOMIC)); 32411da177e4SLinus Torvalds } 32421da177e4SLinus Torvalds 32431da177e4SLinus Torvalds /* This routine sends a packet with an out of date sequence 32441da177e4SLinus Torvalds * number. It assumes the other end will try to ack it. 32451da177e4SLinus Torvalds * 32461da177e4SLinus Torvalds * Question: what should we make while urgent mode? 32471da177e4SLinus Torvalds * 4.4BSD forces sending single byte of data. We cannot send 32481da177e4SLinus Torvalds * out of window data, because we have SND.NXT==SND.MAX... 32491da177e4SLinus Torvalds * 32501da177e4SLinus Torvalds * Current solution: to send TWO zero-length segments in urgent mode: 32511da177e4SLinus Torvalds * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 32521da177e4SLinus Torvalds * out-of-date with SND.UNA-1 to probe window. 32531da177e4SLinus Torvalds */ 32541da177e4SLinus Torvalds static int tcp_xmit_probe_skb(struct sock *sk, int urgent) 32551da177e4SLinus Torvalds { 32561da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 32571da177e4SLinus Torvalds struct sk_buff *skb; 32581da177e4SLinus Torvalds 32591da177e4SLinus Torvalds /* We don't queue it, tcp_transmit_skb() sets ownership. */ 326099a1dec7SMel Gorman skb = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC)); 32611da177e4SLinus Torvalds if (skb == NULL) 32621da177e4SLinus Torvalds return -1; 32631da177e4SLinus Torvalds 32641da177e4SLinus Torvalds /* Reserve space for headers and set control bits. */ 32651da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 32661da177e4SLinus Torvalds /* Use a previous sequence. This should cause the other 32671da177e4SLinus Torvalds * end to send an ack. Don't queue or clone SKB, just 32681da177e4SLinus Torvalds * send it. 32691da177e4SLinus Torvalds */ 3270a3433f35SChangli Gao tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); 32711da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 3272dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC); 32731da177e4SLinus Torvalds } 32741da177e4SLinus Torvalds 3275ee995283SPavel Emelyanov void tcp_send_window_probe(struct sock *sk) 3276ee995283SPavel Emelyanov { 3277ee995283SPavel Emelyanov if (sk->sk_state == TCP_ESTABLISHED) { 3278ee995283SPavel Emelyanov tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; 3279c0e88ff0SPavel Emelyanov tcp_sk(sk)->snd_nxt = tcp_sk(sk)->write_seq; 3280ee995283SPavel Emelyanov tcp_xmit_probe_skb(sk, 0); 3281ee995283SPavel Emelyanov } 3282ee995283SPavel Emelyanov } 3283ee995283SPavel Emelyanov 328467edfef7SAndi Kleen /* Initiate keepalive or window probe from timer. */ 32851da177e4SLinus Torvalds int tcp_write_wakeup(struct sock *sk) 32861da177e4SLinus Torvalds { 32871da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 32881da177e4SLinus Torvalds struct sk_buff *skb; 32891da177e4SLinus Torvalds 3290058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3291058dc334SIlpo Järvinen return -1; 3292058dc334SIlpo Järvinen 3293fe067e8aSDavid S. Miller if ((skb = tcp_send_head(sk)) != NULL && 329490840defSIlpo Järvinen before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { 32951da177e4SLinus Torvalds int err; 32960c54b85fSIlpo Järvinen unsigned int mss = tcp_current_mss(sk); 329790840defSIlpo Järvinen unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 32981da177e4SLinus Torvalds 32991da177e4SLinus Torvalds if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 33001da177e4SLinus Torvalds tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 33011da177e4SLinus Torvalds 33021da177e4SLinus Torvalds /* We are probing the opening of a window 33031da177e4SLinus Torvalds * but the window size is != 0 33041da177e4SLinus Torvalds * must have been a result SWS avoidance ( sender ) 33051da177e4SLinus Torvalds */ 33061da177e4SLinus Torvalds if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 33071da177e4SLinus Torvalds skb->len > mss) { 33081da177e4SLinus Torvalds seg_size = min(seg_size, mss); 33094de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 3310846998aeSDavid S. Miller if (tcp_fragment(sk, skb, seg_size, mss)) 33111da177e4SLinus Torvalds return -1; 33121da177e4SLinus Torvalds } else if (!tcp_skb_pcount(skb)) 3313846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss); 33141da177e4SLinus Torvalds 33154de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 33161da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 3317dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 331866f5fe62SIlpo Järvinen if (!err) 331966f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 33201da177e4SLinus Torvalds return err; 33211da177e4SLinus Torvalds } else { 332233f5f57eSIlpo Järvinen if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) 33234828e7f4SIlpo Järvinen tcp_xmit_probe_skb(sk, 1); 33241da177e4SLinus Torvalds return tcp_xmit_probe_skb(sk, 0); 33251da177e4SLinus Torvalds } 33261da177e4SLinus Torvalds } 33271da177e4SLinus Torvalds 33281da177e4SLinus Torvalds /* A window probe timeout has occurred. If window is not closed send 33291da177e4SLinus Torvalds * a partial packet else a zero probe. 33301da177e4SLinus Torvalds */ 33311da177e4SLinus Torvalds void tcp_send_probe0(struct sock *sk) 33321da177e4SLinus Torvalds { 3333463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 33341da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 33351da177e4SLinus Torvalds int err; 33361da177e4SLinus Torvalds 33371da177e4SLinus Torvalds err = tcp_write_wakeup(sk); 33381da177e4SLinus Torvalds 3339fe067e8aSDavid S. Miller if (tp->packets_out || !tcp_send_head(sk)) { 33401da177e4SLinus Torvalds /* Cancel probe timer, if it is not required. */ 33416687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 0; 3342463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff = 0; 33431da177e4SLinus Torvalds return; 33441da177e4SLinus Torvalds } 33451da177e4SLinus Torvalds 33461da177e4SLinus Torvalds if (err <= 0) { 3347463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_backoff < sysctl_tcp_retries2) 3348463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff++; 33496687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out++; 3350463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 33513f421baaSArnaldo Carvalho de Melo min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), 33523f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 33531da177e4SLinus Torvalds } else { 33541da177e4SLinus Torvalds /* If packet was not sent due to local congestion, 33556687e988SArnaldo Carvalho de Melo * do not backoff and do not remember icsk_probes_out. 33561da177e4SLinus Torvalds * Let local senders to fight for local resources. 33571da177e4SLinus Torvalds * 33581da177e4SLinus Torvalds * Use accumulated backoff yet. 33591da177e4SLinus Torvalds */ 33606687e988SArnaldo Carvalho de Melo if (!icsk->icsk_probes_out) 33616687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 1; 3362463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 3363463c84b9SArnaldo Carvalho de Melo min(icsk->icsk_rto << icsk->icsk_backoff, 33643f421baaSArnaldo Carvalho de Melo TCP_RESOURCE_PROBE_INTERVAL), 33653f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 33661da177e4SLinus Torvalds } 33671da177e4SLinus Torvalds } 3368