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 * 81da177e4SLinus Torvalds * Version: $Id: tcp_output.c,v 1.146 2002/02/01 22:01:04 davem Exp $ 91da177e4SLinus Torvalds * 1002c30a84SJesper Juhl * Authors: Ross Biro 111da177e4SLinus Torvalds * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> 121da177e4SLinus Torvalds * Mark Evans, <evansmp@uhura.aston.ac.uk> 131da177e4SLinus Torvalds * Corey Minyard <wf-rch!minyard@relay.EU.net> 141da177e4SLinus Torvalds * Florian La Roche, <flla@stud.uni-sb.de> 151da177e4SLinus Torvalds * Charles Hedrick, <hedrick@klinzhai.rutgers.edu> 161da177e4SLinus Torvalds * Linus Torvalds, <torvalds@cs.helsinki.fi> 171da177e4SLinus Torvalds * Alan Cox, <gw4pts@gw4pts.ampr.org> 181da177e4SLinus Torvalds * Matthew Dillon, <dillon@apollo.west.oic.com> 191da177e4SLinus Torvalds * Arnt Gulbrandsen, <agulbra@nvg.unit.no> 201da177e4SLinus Torvalds * Jorge Cwik, <jorge@laser.satlink.net> 211da177e4SLinus Torvalds */ 221da177e4SLinus Torvalds 231da177e4SLinus Torvalds /* 241da177e4SLinus Torvalds * Changes: Pedro Roque : Retransmit queue handled by TCP. 251da177e4SLinus Torvalds * : Fragmentation on mtu decrease 261da177e4SLinus Torvalds * : Segment collapse on retransmit 271da177e4SLinus Torvalds * : AF independence 281da177e4SLinus Torvalds * 291da177e4SLinus Torvalds * Linus Torvalds : send_delayed_ack 301da177e4SLinus Torvalds * David S. Miller : Charge memory using the right skb 311da177e4SLinus Torvalds * during syn/ack processing. 321da177e4SLinus Torvalds * David S. Miller : Output engine completely rewritten. 331da177e4SLinus Torvalds * Andrea Arcangeli: SYNACK carry ts_recent in tsecr. 341da177e4SLinus Torvalds * Cacophonix Gaul : draft-minshall-nagle-01 351da177e4SLinus Torvalds * J Hadi Salim : ECN support 361da177e4SLinus Torvalds * 371da177e4SLinus Torvalds */ 381da177e4SLinus Torvalds 391da177e4SLinus Torvalds #include <net/tcp.h> 401da177e4SLinus Torvalds 411da177e4SLinus Torvalds #include <linux/compiler.h> 421da177e4SLinus Torvalds #include <linux/module.h> 431da177e4SLinus Torvalds #include <linux/smp_lock.h> 441da177e4SLinus Torvalds 451da177e4SLinus Torvalds /* People can turn this off for buggy TCP's found in printers etc. */ 461da177e4SLinus Torvalds int sysctl_tcp_retrans_collapse = 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 */ 5115d99e02SRick Jones int sysctl_tcp_workaround_signed_windows = 0; 5215d99e02SRick Jones 531da177e4SLinus Torvalds /* This limits the percentage of the congestion window which we 541da177e4SLinus Torvalds * will allow a single TSO frame to consume. Building TSO frames 551da177e4SLinus Torvalds * which are too large can cause TCP streams to be bursty. 561da177e4SLinus Torvalds */ 57c1b4a7e6SDavid S. Miller int sysctl_tcp_tso_win_divisor = 3; 581da177e4SLinus Torvalds 595d424d5aSJohn Heffner int sysctl_tcp_mtu_probing = 0; 605d424d5aSJohn Heffner int sysctl_tcp_base_mss = 512; 615d424d5aSJohn Heffner 6240efc6faSStephen Hemminger static void update_send_head(struct sock *sk, struct tcp_sock *tp, 631da177e4SLinus Torvalds struct sk_buff *skb) 641da177e4SLinus Torvalds { 651da177e4SLinus Torvalds sk->sk_send_head = skb->next; 661da177e4SLinus Torvalds if (sk->sk_send_head == (struct sk_buff *)&sk->sk_write_queue) 671da177e4SLinus Torvalds sk->sk_send_head = NULL; 681da177e4SLinus Torvalds tp->snd_nxt = TCP_SKB_CB(skb)->end_seq; 691da177e4SLinus Torvalds tcp_packets_out_inc(sk, tp, skb); 701da177e4SLinus Torvalds } 711da177e4SLinus Torvalds 721da177e4SLinus Torvalds /* SND.NXT, if window was not shrunk. 731da177e4SLinus Torvalds * If window has been shrunk, what should we make? It is not clear at all. 741da177e4SLinus Torvalds * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-( 751da177e4SLinus Torvalds * Anything in between SND.UNA...SND.UNA+SND.WND also can be already 761da177e4SLinus Torvalds * invalid. OK, let's make this for now: 771da177e4SLinus Torvalds */ 781da177e4SLinus Torvalds static inline __u32 tcp_acceptable_seq(struct sock *sk, struct tcp_sock *tp) 791da177e4SLinus Torvalds { 801da177e4SLinus Torvalds if (!before(tp->snd_una+tp->snd_wnd, tp->snd_nxt)) 811da177e4SLinus Torvalds return tp->snd_nxt; 821da177e4SLinus Torvalds else 831da177e4SLinus Torvalds return tp->snd_una+tp->snd_wnd; 841da177e4SLinus Torvalds } 851da177e4SLinus Torvalds 861da177e4SLinus Torvalds /* Calculate mss to advertise in SYN segment. 871da177e4SLinus Torvalds * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that: 881da177e4SLinus Torvalds * 891da177e4SLinus Torvalds * 1. It is independent of path mtu. 901da177e4SLinus Torvalds * 2. Ideally, it is maximal possible segment size i.e. 65535-40. 911da177e4SLinus Torvalds * 3. For IPv4 it is reasonable to calculate it from maximal MTU of 921da177e4SLinus Torvalds * attached devices, because some buggy hosts are confused by 931da177e4SLinus Torvalds * large MSS. 941da177e4SLinus Torvalds * 4. We do not make 3, we advertise MSS, calculated from first 951da177e4SLinus Torvalds * hop device mtu, but allow to raise it to ip_rt_min_advmss. 961da177e4SLinus Torvalds * This may be overridden via information stored in routing table. 971da177e4SLinus Torvalds * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible, 981da177e4SLinus Torvalds * probably even Jumbo". 991da177e4SLinus Torvalds */ 1001da177e4SLinus Torvalds static __u16 tcp_advertise_mss(struct sock *sk) 1011da177e4SLinus Torvalds { 1021da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1031da177e4SLinus Torvalds struct dst_entry *dst = __sk_dst_get(sk); 1041da177e4SLinus Torvalds int mss = tp->advmss; 1051da177e4SLinus Torvalds 1061da177e4SLinus Torvalds if (dst && dst_metric(dst, RTAX_ADVMSS) < mss) { 1071da177e4SLinus Torvalds mss = dst_metric(dst, RTAX_ADVMSS); 1081da177e4SLinus Torvalds tp->advmss = mss; 1091da177e4SLinus Torvalds } 1101da177e4SLinus Torvalds 1111da177e4SLinus Torvalds return (__u16)mss; 1121da177e4SLinus Torvalds } 1131da177e4SLinus Torvalds 1141da177e4SLinus Torvalds /* RFC2861. Reset CWND after idle period longer RTO to "restart window". 1151da177e4SLinus Torvalds * This is the first part of cwnd validation mechanism. */ 116463c84b9SArnaldo Carvalho de Melo static void tcp_cwnd_restart(struct sock *sk, struct dst_entry *dst) 1171da177e4SLinus Torvalds { 118463c84b9SArnaldo Carvalho de Melo struct tcp_sock *tp = tcp_sk(sk); 1191da177e4SLinus Torvalds s32 delta = tcp_time_stamp - tp->lsndtime; 1201da177e4SLinus Torvalds u32 restart_cwnd = tcp_init_cwnd(tp, dst); 1211da177e4SLinus Torvalds u32 cwnd = tp->snd_cwnd; 1221da177e4SLinus Torvalds 1236687e988SArnaldo Carvalho de Melo tcp_ca_event(sk, CA_EVENT_CWND_RESTART); 1241da177e4SLinus Torvalds 1256687e988SArnaldo Carvalho de Melo tp->snd_ssthresh = tcp_current_ssthresh(sk); 1261da177e4SLinus Torvalds restart_cwnd = min(restart_cwnd, cwnd); 1271da177e4SLinus Torvalds 128463c84b9SArnaldo Carvalho de Melo while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd) 1291da177e4SLinus Torvalds cwnd >>= 1; 1301da177e4SLinus Torvalds tp->snd_cwnd = max(cwnd, restart_cwnd); 1311da177e4SLinus Torvalds tp->snd_cwnd_stamp = tcp_time_stamp; 1321da177e4SLinus Torvalds tp->snd_cwnd_used = 0; 1331da177e4SLinus Torvalds } 1341da177e4SLinus Torvalds 13540efc6faSStephen Hemminger static void tcp_event_data_sent(struct tcp_sock *tp, 1361da177e4SLinus Torvalds struct sk_buff *skb, struct sock *sk) 1371da177e4SLinus Torvalds { 138463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 139463c84b9SArnaldo Carvalho de Melo const u32 now = tcp_time_stamp; 1401da177e4SLinus Torvalds 141463c84b9SArnaldo Carvalho de Melo if (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto) 142463c84b9SArnaldo Carvalho de Melo tcp_cwnd_restart(sk, __sk_dst_get(sk)); 1431da177e4SLinus Torvalds 1441da177e4SLinus Torvalds tp->lsndtime = now; 1451da177e4SLinus Torvalds 1461da177e4SLinus Torvalds /* If it is a reply for ato after last received 1471da177e4SLinus Torvalds * packet, enter pingpong mode. 1481da177e4SLinus Torvalds */ 149463c84b9SArnaldo Carvalho de Melo if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato) 150463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pingpong = 1; 1511da177e4SLinus Torvalds } 1521da177e4SLinus Torvalds 15340efc6faSStephen Hemminger static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts) 1541da177e4SLinus Torvalds { 155463c84b9SArnaldo Carvalho de Melo tcp_dec_quickack_mode(sk, pkts); 156463c84b9SArnaldo Carvalho de Melo inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); 1571da177e4SLinus Torvalds } 1581da177e4SLinus Torvalds 1591da177e4SLinus Torvalds /* Determine a window scaling and initial window to offer. 1601da177e4SLinus Torvalds * Based on the assumption that the given amount of space 1611da177e4SLinus Torvalds * will be offered. Store the results in the tp structure. 1621da177e4SLinus Torvalds * NOTE: for smooth operation initial space offering should 1631da177e4SLinus Torvalds * be a multiple of mss if possible. We assume here that mss >= 1. 1641da177e4SLinus Torvalds * This MUST be enforced by all callers. 1651da177e4SLinus Torvalds */ 1661da177e4SLinus Torvalds void tcp_select_initial_window(int __space, __u32 mss, 1671da177e4SLinus Torvalds __u32 *rcv_wnd, __u32 *window_clamp, 1681da177e4SLinus Torvalds int wscale_ok, __u8 *rcv_wscale) 1691da177e4SLinus Torvalds { 1701da177e4SLinus Torvalds unsigned int space = (__space < 0 ? 0 : __space); 1711da177e4SLinus Torvalds 1721da177e4SLinus Torvalds /* If no clamp set the clamp to the max possible scaled window */ 1731da177e4SLinus Torvalds if (*window_clamp == 0) 1741da177e4SLinus Torvalds (*window_clamp) = (65535 << 14); 1751da177e4SLinus Torvalds space = min(*window_clamp, space); 1761da177e4SLinus Torvalds 1771da177e4SLinus Torvalds /* Quantize space offering to a multiple of mss if possible. */ 1781da177e4SLinus Torvalds if (space > mss) 1791da177e4SLinus Torvalds space = (space / mss) * mss; 1801da177e4SLinus Torvalds 1811da177e4SLinus Torvalds /* NOTE: offering an initial window larger than 32767 18215d99e02SRick Jones * will break some buggy TCP stacks. If the admin tells us 18315d99e02SRick Jones * it is likely we could be speaking with such a buggy stack 18415d99e02SRick Jones * we will truncate our initial window offering to 32K-1 18515d99e02SRick Jones * unless the remote has sent us a window scaling option, 18615d99e02SRick Jones * which we interpret as a sign the remote TCP is not 18715d99e02SRick Jones * misinterpreting the window field as a signed quantity. 1881da177e4SLinus Torvalds */ 18915d99e02SRick Jones if (sysctl_tcp_workaround_signed_windows) 1901da177e4SLinus Torvalds (*rcv_wnd) = min(space, MAX_TCP_WINDOW); 19115d99e02SRick Jones else 19215d99e02SRick Jones (*rcv_wnd) = space; 19315d99e02SRick Jones 1941da177e4SLinus Torvalds (*rcv_wscale) = 0; 1951da177e4SLinus Torvalds if (wscale_ok) { 1961da177e4SLinus Torvalds /* Set window scaling on max possible window 1971da177e4SLinus Torvalds * See RFC1323 for an explanation of the limit to 14 1981da177e4SLinus Torvalds */ 1991da177e4SLinus Torvalds space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max); 2001da177e4SLinus Torvalds while (space > 65535 && (*rcv_wscale) < 14) { 2011da177e4SLinus Torvalds space >>= 1; 2021da177e4SLinus Torvalds (*rcv_wscale)++; 2031da177e4SLinus Torvalds } 2041da177e4SLinus Torvalds } 2051da177e4SLinus Torvalds 2061da177e4SLinus Torvalds /* Set initial window to value enough for senders, 2076b251858SDavid S. Miller * following RFC2414. Senders, not following this RFC, 2081da177e4SLinus Torvalds * will be satisfied with 2. 2091da177e4SLinus Torvalds */ 2101da177e4SLinus Torvalds if (mss > (1<<*rcv_wscale)) { 21101ff367eSDavid S. Miller int init_cwnd = 4; 21201ff367eSDavid S. Miller if (mss > 1460*3) 2131da177e4SLinus Torvalds init_cwnd = 2; 21401ff367eSDavid S. Miller else if (mss > 1460) 21501ff367eSDavid S. Miller init_cwnd = 3; 2161da177e4SLinus Torvalds if (*rcv_wnd > init_cwnd*mss) 2171da177e4SLinus Torvalds *rcv_wnd = init_cwnd*mss; 2181da177e4SLinus Torvalds } 2191da177e4SLinus Torvalds 2201da177e4SLinus Torvalds /* Set the clamp no higher than max representable value */ 2211da177e4SLinus Torvalds (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp); 2221da177e4SLinus Torvalds } 2231da177e4SLinus Torvalds 2241da177e4SLinus Torvalds /* Chose a new window to advertise, update state in tcp_sock for the 2251da177e4SLinus Torvalds * socket, and return result with RFC1323 scaling applied. The return 2261da177e4SLinus Torvalds * value can be stuffed directly into th->window for an outgoing 2271da177e4SLinus Torvalds * frame. 2281da177e4SLinus Torvalds */ 22940efc6faSStephen Hemminger static u16 tcp_select_window(struct sock *sk) 2301da177e4SLinus Torvalds { 2311da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2321da177e4SLinus Torvalds u32 cur_win = tcp_receive_window(tp); 2331da177e4SLinus Torvalds u32 new_win = __tcp_select_window(sk); 2341da177e4SLinus Torvalds 2351da177e4SLinus Torvalds /* Never shrink the offered window */ 2361da177e4SLinus Torvalds if(new_win < cur_win) { 2371da177e4SLinus Torvalds /* Danger Will Robinson! 2381da177e4SLinus Torvalds * Don't update rcv_wup/rcv_wnd here or else 2391da177e4SLinus Torvalds * we will not be able to advertise a zero 2401da177e4SLinus Torvalds * window in time. --DaveM 2411da177e4SLinus Torvalds * 2421da177e4SLinus Torvalds * Relax Will Robinson. 2431da177e4SLinus Torvalds */ 2441da177e4SLinus Torvalds new_win = cur_win; 2451da177e4SLinus Torvalds } 2461da177e4SLinus Torvalds tp->rcv_wnd = new_win; 2471da177e4SLinus Torvalds tp->rcv_wup = tp->rcv_nxt; 2481da177e4SLinus Torvalds 2491da177e4SLinus Torvalds /* Make sure we do not exceed the maximum possible 2501da177e4SLinus Torvalds * scaled window. 2511da177e4SLinus Torvalds */ 25215d99e02SRick Jones if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows) 2531da177e4SLinus Torvalds new_win = min(new_win, MAX_TCP_WINDOW); 2541da177e4SLinus Torvalds else 2551da177e4SLinus Torvalds new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); 2561da177e4SLinus Torvalds 2571da177e4SLinus Torvalds /* RFC1323 scaling applied */ 2581da177e4SLinus Torvalds new_win >>= tp->rx_opt.rcv_wscale; 2591da177e4SLinus Torvalds 2601da177e4SLinus Torvalds /* If we advertise zero window, disable fast path. */ 2611da177e4SLinus Torvalds if (new_win == 0) 2621da177e4SLinus Torvalds tp->pred_flags = 0; 2631da177e4SLinus Torvalds 2641da177e4SLinus Torvalds return new_win; 2651da177e4SLinus Torvalds } 2661da177e4SLinus Torvalds 26740efc6faSStephen Hemminger static void tcp_build_and_update_options(__u32 *ptr, struct tcp_sock *tp, 26840efc6faSStephen Hemminger __u32 tstamp) 26940efc6faSStephen Hemminger { 27040efc6faSStephen Hemminger if (tp->rx_opt.tstamp_ok) { 27140efc6faSStephen Hemminger *ptr++ = __constant_htonl((TCPOPT_NOP << 24) | 27240efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 27340efc6faSStephen Hemminger (TCPOPT_TIMESTAMP << 8) | 27440efc6faSStephen Hemminger TCPOLEN_TIMESTAMP); 27540efc6faSStephen Hemminger *ptr++ = htonl(tstamp); 27640efc6faSStephen Hemminger *ptr++ = htonl(tp->rx_opt.ts_recent); 27740efc6faSStephen Hemminger } 27840efc6faSStephen Hemminger if (tp->rx_opt.eff_sacks) { 27940efc6faSStephen Hemminger struct tcp_sack_block *sp = tp->rx_opt.dsack ? tp->duplicate_sack : tp->selective_acks; 28040efc6faSStephen Hemminger int this_sack; 28140efc6faSStephen Hemminger 28240efc6faSStephen Hemminger *ptr++ = htonl((TCPOPT_NOP << 24) | 28340efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 28440efc6faSStephen Hemminger (TCPOPT_SACK << 8) | 28540efc6faSStephen Hemminger (TCPOLEN_SACK_BASE + (tp->rx_opt.eff_sacks * 28640efc6faSStephen Hemminger TCPOLEN_SACK_PERBLOCK))); 28740efc6faSStephen Hemminger for(this_sack = 0; this_sack < tp->rx_opt.eff_sacks; this_sack++) { 28840efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].start_seq); 28940efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].end_seq); 29040efc6faSStephen Hemminger } 29140efc6faSStephen Hemminger if (tp->rx_opt.dsack) { 29240efc6faSStephen Hemminger tp->rx_opt.dsack = 0; 29340efc6faSStephen Hemminger tp->rx_opt.eff_sacks--; 29440efc6faSStephen Hemminger } 29540efc6faSStephen Hemminger } 29640efc6faSStephen Hemminger } 29740efc6faSStephen Hemminger 29840efc6faSStephen Hemminger /* Construct a tcp options header for a SYN or SYN_ACK packet. 29940efc6faSStephen Hemminger * If this is every changed make sure to change the definition of 30040efc6faSStephen Hemminger * MAX_SYN_SIZE to match the new maximum number of options that you 30140efc6faSStephen Hemminger * can generate. 30240efc6faSStephen Hemminger */ 30340efc6faSStephen Hemminger static void tcp_syn_build_options(__u32 *ptr, int mss, int ts, int sack, 30440efc6faSStephen Hemminger int offer_wscale, int wscale, __u32 tstamp, 30540efc6faSStephen Hemminger __u32 ts_recent) 30640efc6faSStephen Hemminger { 30740efc6faSStephen Hemminger /* We always get an MSS option. 30840efc6faSStephen Hemminger * The option bytes which will be seen in normal data 30940efc6faSStephen Hemminger * packets should timestamps be used, must be in the MSS 31040efc6faSStephen Hemminger * advertised. But we subtract them from tp->mss_cache so 31140efc6faSStephen Hemminger * that calculations in tcp_sendmsg are simpler etc. 31240efc6faSStephen Hemminger * So account for this fact here if necessary. If we 31340efc6faSStephen Hemminger * don't do this correctly, as a receiver we won't 31440efc6faSStephen Hemminger * recognize data packets as being full sized when we 31540efc6faSStephen Hemminger * should, and thus we won't abide by the delayed ACK 31640efc6faSStephen Hemminger * rules correctly. 31740efc6faSStephen Hemminger * SACKs don't matter, we never delay an ACK when we 31840efc6faSStephen Hemminger * have any of those going out. 31940efc6faSStephen Hemminger */ 32040efc6faSStephen Hemminger *ptr++ = htonl((TCPOPT_MSS << 24) | (TCPOLEN_MSS << 16) | mss); 32140efc6faSStephen Hemminger if (ts) { 32240efc6faSStephen Hemminger if(sack) 32340efc6faSStephen Hemminger *ptr++ = __constant_htonl((TCPOPT_SACK_PERM << 24) | (TCPOLEN_SACK_PERM << 16) | 32440efc6faSStephen Hemminger (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP); 32540efc6faSStephen Hemminger else 32640efc6faSStephen Hemminger *ptr++ = __constant_htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | 32740efc6faSStephen Hemminger (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP); 32840efc6faSStephen Hemminger *ptr++ = htonl(tstamp); /* TSVAL */ 32940efc6faSStephen Hemminger *ptr++ = htonl(ts_recent); /* TSECR */ 33040efc6faSStephen Hemminger } else if(sack) 33140efc6faSStephen Hemminger *ptr++ = __constant_htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | 33240efc6faSStephen Hemminger (TCPOPT_SACK_PERM << 8) | TCPOLEN_SACK_PERM); 33340efc6faSStephen Hemminger if (offer_wscale) 33440efc6faSStephen Hemminger *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_WINDOW << 16) | (TCPOLEN_WINDOW << 8) | (wscale)); 33540efc6faSStephen Hemminger } 3361da177e4SLinus Torvalds 3371da177e4SLinus Torvalds /* This routine actually transmits TCP packets queued in by 3381da177e4SLinus Torvalds * tcp_do_sendmsg(). This is used by both the initial 3391da177e4SLinus Torvalds * transmission and possible later retransmissions. 3401da177e4SLinus Torvalds * All SKB's seen here are completely headerless. It is our 3411da177e4SLinus Torvalds * job to build the TCP header, and pass the packet down to 3421da177e4SLinus Torvalds * IP so it can do the same plus pass the packet off to the 3431da177e4SLinus Torvalds * device. 3441da177e4SLinus Torvalds * 3451da177e4SLinus Torvalds * We are working here with either a clone of the original 3461da177e4SLinus Torvalds * SKB, or a fresh unique copy made by the retransmit engine. 3471da177e4SLinus Torvalds */ 348dfb4b9dcSDavid S. Miller static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, gfp_t gfp_mask) 3491da177e4SLinus Torvalds { 3506687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 351dfb4b9dcSDavid S. Miller struct inet_sock *inet; 352dfb4b9dcSDavid S. Miller struct tcp_sock *tp; 353dfb4b9dcSDavid S. Miller struct tcp_skb_cb *tcb; 354dfb4b9dcSDavid S. Miller int tcp_header_size; 3551da177e4SLinus Torvalds struct tcphdr *th; 3561da177e4SLinus Torvalds int sysctl_flags; 3571da177e4SLinus Torvalds int err; 3581da177e4SLinus Torvalds 359dfb4b9dcSDavid S. Miller BUG_ON(!skb || !tcp_skb_pcount(skb)); 360dfb4b9dcSDavid S. Miller 361dfb4b9dcSDavid S. Miller /* If congestion control is doing timestamping, we must 362dfb4b9dcSDavid S. Miller * take such a timestamp before we potentially clone/copy. 363dfb4b9dcSDavid S. Miller */ 364dfb4b9dcSDavid S. Miller if (icsk->icsk_ca_ops->rtt_sample) 365dfb4b9dcSDavid S. Miller __net_timestamp(skb); 366dfb4b9dcSDavid S. Miller 367dfb4b9dcSDavid S. Miller if (likely(clone_it)) { 368dfb4b9dcSDavid S. Miller if (unlikely(skb_cloned(skb))) 369dfb4b9dcSDavid S. Miller skb = pskb_copy(skb, gfp_mask); 370dfb4b9dcSDavid S. Miller else 371dfb4b9dcSDavid S. Miller skb = skb_clone(skb, gfp_mask); 372dfb4b9dcSDavid S. Miller if (unlikely(!skb)) 373dfb4b9dcSDavid S. Miller return -ENOBUFS; 374dfb4b9dcSDavid S. Miller } 375dfb4b9dcSDavid S. Miller 376dfb4b9dcSDavid S. Miller inet = inet_sk(sk); 377dfb4b9dcSDavid S. Miller tp = tcp_sk(sk); 378dfb4b9dcSDavid S. Miller tcb = TCP_SKB_CB(skb); 379dfb4b9dcSDavid S. Miller tcp_header_size = tp->tcp_header_len; 3801da177e4SLinus Torvalds 3811da177e4SLinus Torvalds #define SYSCTL_FLAG_TSTAMPS 0x1 3821da177e4SLinus Torvalds #define SYSCTL_FLAG_WSCALE 0x2 3831da177e4SLinus Torvalds #define SYSCTL_FLAG_SACK 0x4 3841da177e4SLinus Torvalds 3851da177e4SLinus Torvalds sysctl_flags = 0; 386dfb4b9dcSDavid S. Miller if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) { 3871da177e4SLinus Torvalds tcp_header_size = sizeof(struct tcphdr) + TCPOLEN_MSS; 3881da177e4SLinus Torvalds if(sysctl_tcp_timestamps) { 3891da177e4SLinus Torvalds tcp_header_size += TCPOLEN_TSTAMP_ALIGNED; 3901da177e4SLinus Torvalds sysctl_flags |= SYSCTL_FLAG_TSTAMPS; 3911da177e4SLinus Torvalds } 3921da177e4SLinus Torvalds if (sysctl_tcp_window_scaling) { 3931da177e4SLinus Torvalds tcp_header_size += TCPOLEN_WSCALE_ALIGNED; 3941da177e4SLinus Torvalds sysctl_flags |= SYSCTL_FLAG_WSCALE; 3951da177e4SLinus Torvalds } 3961da177e4SLinus Torvalds if (sysctl_tcp_sack) { 3971da177e4SLinus Torvalds sysctl_flags |= SYSCTL_FLAG_SACK; 3981da177e4SLinus Torvalds if (!(sysctl_flags & SYSCTL_FLAG_TSTAMPS)) 3991da177e4SLinus Torvalds tcp_header_size += TCPOLEN_SACKPERM_ALIGNED; 4001da177e4SLinus Torvalds } 401dfb4b9dcSDavid S. Miller } else if (unlikely(tp->rx_opt.eff_sacks)) { 4021da177e4SLinus Torvalds /* A SACK is 2 pad bytes, a 2 byte header, plus 4031da177e4SLinus Torvalds * 2 32-bit sequence numbers for each SACK block. 4041da177e4SLinus Torvalds */ 4051da177e4SLinus Torvalds tcp_header_size += (TCPOLEN_SACK_BASE_ALIGNED + 406dfb4b9dcSDavid S. Miller (tp->rx_opt.eff_sacks * 407dfb4b9dcSDavid S. Miller TCPOLEN_SACK_PERBLOCK)); 4081da177e4SLinus Torvalds } 4091da177e4SLinus Torvalds 410317a76f9SStephen Hemminger if (tcp_packets_in_flight(tp) == 0) 4116687e988SArnaldo Carvalho de Melo tcp_ca_event(sk, CA_EVENT_TX_START); 4121da177e4SLinus Torvalds 4131da177e4SLinus Torvalds th = (struct tcphdr *) skb_push(skb, tcp_header_size); 4141da177e4SLinus Torvalds skb->h.th = th; 4151da177e4SLinus Torvalds skb_set_owner_w(skb, sk); 4161da177e4SLinus Torvalds 4171da177e4SLinus Torvalds /* Build TCP header and checksum it. */ 4181da177e4SLinus Torvalds th->source = inet->sport; 4191da177e4SLinus Torvalds th->dest = inet->dport; 4201da177e4SLinus Torvalds th->seq = htonl(tcb->seq); 4211da177e4SLinus Torvalds th->ack_seq = htonl(tp->rcv_nxt); 422dfb4b9dcSDavid S. Miller *(((__u16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | 423dfb4b9dcSDavid S. Miller tcb->flags); 424dfb4b9dcSDavid S. Miller 425dfb4b9dcSDavid S. Miller if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) { 4261da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments 4271da177e4SLinus Torvalds * is never scaled. 4281da177e4SLinus Torvalds */ 4291da177e4SLinus Torvalds th->window = htons(tp->rcv_wnd); 4301da177e4SLinus Torvalds } else { 4311da177e4SLinus Torvalds th->window = htons(tcp_select_window(sk)); 4321da177e4SLinus Torvalds } 4331da177e4SLinus Torvalds th->check = 0; 4341da177e4SLinus Torvalds th->urg_ptr = 0; 4351da177e4SLinus Torvalds 436dfb4b9dcSDavid S. Miller if (unlikely(tp->urg_mode && 437dfb4b9dcSDavid S. Miller between(tp->snd_up, tcb->seq+1, tcb->seq+0xFFFF))) { 4381da177e4SLinus Torvalds th->urg_ptr = htons(tp->snd_up-tcb->seq); 4391da177e4SLinus Torvalds th->urg = 1; 4401da177e4SLinus Torvalds } 4411da177e4SLinus Torvalds 442dfb4b9dcSDavid S. Miller if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) { 4431da177e4SLinus Torvalds tcp_syn_build_options((__u32 *)(th + 1), 4441da177e4SLinus Torvalds tcp_advertise_mss(sk), 4451da177e4SLinus Torvalds (sysctl_flags & SYSCTL_FLAG_TSTAMPS), 4461da177e4SLinus Torvalds (sysctl_flags & SYSCTL_FLAG_SACK), 4471da177e4SLinus Torvalds (sysctl_flags & SYSCTL_FLAG_WSCALE), 4481da177e4SLinus Torvalds tp->rx_opt.rcv_wscale, 4491da177e4SLinus Torvalds tcb->when, 4501da177e4SLinus Torvalds tp->rx_opt.ts_recent); 4511da177e4SLinus Torvalds } else { 4521da177e4SLinus Torvalds tcp_build_and_update_options((__u32 *)(th + 1), 4531da177e4SLinus Torvalds tp, tcb->when); 4541da177e4SLinus Torvalds TCP_ECN_send(sk, tp, skb, tcp_header_size); 4551da177e4SLinus Torvalds } 456dfb4b9dcSDavid S. Miller 4578292a17aSArnaldo Carvalho de Melo icsk->icsk_af_ops->send_check(sk, skb->len, skb); 4581da177e4SLinus Torvalds 459dfb4b9dcSDavid S. Miller if (likely(tcb->flags & TCPCB_FLAG_ACK)) 460fc6415bcSDavid S. Miller tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); 4611da177e4SLinus Torvalds 4621da177e4SLinus Torvalds if (skb->len != tcp_header_size) 4631da177e4SLinus Torvalds tcp_event_data_sent(tp, skb, sk); 4641da177e4SLinus Torvalds 4651da177e4SLinus Torvalds TCP_INC_STATS(TCP_MIB_OUTSEGS); 4661da177e4SLinus Torvalds 4678292a17aSArnaldo Carvalho de Melo err = icsk->icsk_af_ops->queue_xmit(skb, 0); 46883de47cdSHua Zhong if (likely(err <= 0)) 4691da177e4SLinus Torvalds return err; 4701da177e4SLinus Torvalds 4716687e988SArnaldo Carvalho de Melo tcp_enter_cwr(sk); 4721da177e4SLinus Torvalds 4731da177e4SLinus Torvalds /* NET_XMIT_CN is special. It does not guarantee, 4741da177e4SLinus Torvalds * that this packet is lost. It tells that device 4751da177e4SLinus Torvalds * is about to start to drop packets or already 4761da177e4SLinus Torvalds * drops some packets of the same priority and 4771da177e4SLinus Torvalds * invokes us to send less aggressively. 4781da177e4SLinus Torvalds */ 4791da177e4SLinus Torvalds return err == NET_XMIT_CN ? 0 : err; 480dfb4b9dcSDavid S. Miller 4811da177e4SLinus Torvalds #undef SYSCTL_FLAG_TSTAMPS 4821da177e4SLinus Torvalds #undef SYSCTL_FLAG_WSCALE 4831da177e4SLinus Torvalds #undef SYSCTL_FLAG_SACK 4841da177e4SLinus Torvalds } 4851da177e4SLinus Torvalds 4861da177e4SLinus Torvalds 4871da177e4SLinus Torvalds /* This routine just queue's the buffer 4881da177e4SLinus Torvalds * 4891da177e4SLinus Torvalds * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, 4901da177e4SLinus Torvalds * otherwise socket can stall. 4911da177e4SLinus Torvalds */ 4921da177e4SLinus Torvalds static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) 4931da177e4SLinus Torvalds { 4941da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 4951da177e4SLinus Torvalds 4961da177e4SLinus Torvalds /* Advance write_seq and place onto the write_queue. */ 4971da177e4SLinus Torvalds tp->write_seq = TCP_SKB_CB(skb)->end_seq; 4981da177e4SLinus Torvalds skb_header_release(skb); 4991da177e4SLinus Torvalds __skb_queue_tail(&sk->sk_write_queue, skb); 5001da177e4SLinus Torvalds sk_charge_skb(sk, skb); 5011da177e4SLinus Torvalds 5021da177e4SLinus Torvalds /* Queue it, remembering where we must start sending. */ 5031da177e4SLinus Torvalds if (sk->sk_send_head == NULL) 5041da177e4SLinus Torvalds sk->sk_send_head = skb; 5051da177e4SLinus Torvalds } 5061da177e4SLinus Torvalds 507846998aeSDavid S. Miller static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb, unsigned int mss_now) 508f6302d1dSDavid S. Miller { 509846998aeSDavid S. Miller if (skb->len <= mss_now || 510f6302d1dSDavid S. Miller !(sk->sk_route_caps & NETIF_F_TSO)) { 511f6302d1dSDavid S. Miller /* Avoid the costly divide in the normal 512f6302d1dSDavid S. Miller * non-TSO case. 513f6302d1dSDavid S. Miller */ 514f6302d1dSDavid S. Miller skb_shinfo(skb)->tso_segs = 1; 515f6302d1dSDavid S. Miller skb_shinfo(skb)->tso_size = 0; 516f6302d1dSDavid S. Miller } else { 517f6302d1dSDavid S. Miller unsigned int factor; 518f6302d1dSDavid S. Miller 519846998aeSDavid S. Miller factor = skb->len + (mss_now - 1); 520846998aeSDavid S. Miller factor /= mss_now; 521f6302d1dSDavid S. Miller skb_shinfo(skb)->tso_segs = factor; 522846998aeSDavid S. Miller skb_shinfo(skb)->tso_size = mss_now; 5231da177e4SLinus Torvalds } 5241da177e4SLinus Torvalds } 5251da177e4SLinus Torvalds 5261da177e4SLinus Torvalds /* Function to create two new TCP segments. Shrinks the given segment 5271da177e4SLinus Torvalds * to the specified size and appends a new segment with the rest of the 5281da177e4SLinus Torvalds * packet to the list. This won't be called frequently, I hope. 5291da177e4SLinus Torvalds * Remember, these are still headerless SKBs at this point. 5301da177e4SLinus Torvalds */ 5316475be16SDavid S. Miller int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, unsigned int mss_now) 5321da177e4SLinus Torvalds { 5331da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 5341da177e4SLinus Torvalds struct sk_buff *buff; 5356475be16SDavid S. Miller int nsize, old_factor; 536b60b49eaSHerbert Xu int nlen; 5371da177e4SLinus Torvalds u16 flags; 5381da177e4SLinus Torvalds 539b2cc99f0SHerbert Xu BUG_ON(len > skb->len); 5406a438bbeSStephen Hemminger 5416a438bbeSStephen Hemminger clear_all_retrans_hints(tp); 5421da177e4SLinus Torvalds nsize = skb_headlen(skb) - len; 5431da177e4SLinus Torvalds if (nsize < 0) 5441da177e4SLinus Torvalds nsize = 0; 5451da177e4SLinus Torvalds 5461da177e4SLinus Torvalds if (skb_cloned(skb) && 5471da177e4SLinus Torvalds skb_is_nonlinear(skb) && 5481da177e4SLinus Torvalds pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) 5491da177e4SLinus Torvalds return -ENOMEM; 5501da177e4SLinus Torvalds 5511da177e4SLinus Torvalds /* Get a new skb... force flag on. */ 5521da177e4SLinus Torvalds buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC); 5531da177e4SLinus Torvalds if (buff == NULL) 5541da177e4SLinus Torvalds return -ENOMEM; /* We'll just try again later. */ 555ef5cb973SHerbert Xu 556b60b49eaSHerbert Xu sk_charge_skb(sk, buff); 557b60b49eaSHerbert Xu nlen = skb->len - len - nsize; 558b60b49eaSHerbert Xu buff->truesize += nlen; 559b60b49eaSHerbert Xu skb->truesize -= nlen; 5601da177e4SLinus Torvalds 5611da177e4SLinus Torvalds /* Correct the sequence numbers. */ 5621da177e4SLinus Torvalds TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 5631da177e4SLinus Torvalds TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 5641da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 5651da177e4SLinus Torvalds 5661da177e4SLinus Torvalds /* PSH and FIN should only be set in the second packet. */ 5671da177e4SLinus Torvalds flags = TCP_SKB_CB(skb)->flags; 5681da177e4SLinus Torvalds TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH); 5691da177e4SLinus Torvalds TCP_SKB_CB(buff)->flags = flags; 570e14c3cafSHerbert Xu TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; 5711da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked &= ~TCPCB_AT_TAIL; 5721da177e4SLinus Torvalds 5731da177e4SLinus Torvalds if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_HW) { 5741da177e4SLinus Torvalds /* Copy and checksum data tail into the new buffer. */ 5751da177e4SLinus Torvalds buff->csum = csum_partial_copy_nocheck(skb->data + len, skb_put(buff, nsize), 5761da177e4SLinus Torvalds nsize, 0); 5771da177e4SLinus Torvalds 5781da177e4SLinus Torvalds skb_trim(skb, len); 5791da177e4SLinus Torvalds 5801da177e4SLinus Torvalds skb->csum = csum_block_sub(skb->csum, buff->csum, len); 5811da177e4SLinus Torvalds } else { 5821da177e4SLinus Torvalds skb->ip_summed = CHECKSUM_HW; 5831da177e4SLinus Torvalds skb_split(skb, buff, len); 5841da177e4SLinus Torvalds } 5851da177e4SLinus Torvalds 5861da177e4SLinus Torvalds buff->ip_summed = skb->ip_summed; 5871da177e4SLinus Torvalds 5881da177e4SLinus Torvalds /* Looks stupid, but our code really uses when of 5891da177e4SLinus Torvalds * skbs, which it never sent before. --ANK 5901da177e4SLinus Torvalds */ 5911da177e4SLinus Torvalds TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when; 592a61bbcf2SPatrick McHardy buff->tstamp = skb->tstamp; 5931da177e4SLinus Torvalds 5946475be16SDavid S. Miller old_factor = tcp_skb_pcount(skb); 5956475be16SDavid S. Miller 5961da177e4SLinus Torvalds /* Fix up tso_factor for both original and new SKB. */ 597846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 598846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, buff, mss_now); 5991da177e4SLinus Torvalds 6006475be16SDavid S. Miller /* If this packet has been sent out already, we must 6016475be16SDavid S. Miller * adjust the various packet counters. 6026475be16SDavid S. Miller */ 603cf0b450cSHerbert Xu if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 6046475be16SDavid S. Miller int diff = old_factor - tcp_skb_pcount(skb) - 6056475be16SDavid S. Miller tcp_skb_pcount(buff); 6061da177e4SLinus Torvalds 6076475be16SDavid S. Miller tp->packets_out -= diff; 608e14c3cafSHerbert Xu 609e14c3cafSHerbert Xu if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 610e14c3cafSHerbert Xu tp->sacked_out -= diff; 611e14c3cafSHerbert Xu if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 612e14c3cafSHerbert Xu tp->retrans_out -= diff; 613e14c3cafSHerbert Xu 6146475be16SDavid S. Miller if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) { 6156475be16SDavid S. Miller tp->lost_out -= diff; 6166475be16SDavid S. Miller tp->left_out -= diff; 6176475be16SDavid S. Miller } 61883ca28beSHerbert Xu 6196475be16SDavid S. Miller if (diff > 0) { 62083ca28beSHerbert Xu /* Adjust Reno SACK estimate. */ 62183ca28beSHerbert Xu if (!tp->rx_opt.sack_ok) { 62283ca28beSHerbert Xu tp->sacked_out -= diff; 62383ca28beSHerbert Xu if ((int)tp->sacked_out < 0) 62483ca28beSHerbert Xu tp->sacked_out = 0; 62583ca28beSHerbert Xu tcp_sync_left_out(tp); 62683ca28beSHerbert Xu } 62783ca28beSHerbert Xu 6286475be16SDavid S. Miller tp->fackets_out -= diff; 6296475be16SDavid S. Miller if ((int)tp->fackets_out < 0) 6306475be16SDavid S. Miller tp->fackets_out = 0; 6316475be16SDavid S. Miller } 6321da177e4SLinus Torvalds } 6331da177e4SLinus Torvalds 6341da177e4SLinus Torvalds /* Link BUFF into the send queue. */ 635f44b5271SDavid S. Miller skb_header_release(buff); 6368728b834SDavid S. Miller __skb_append(skb, buff, &sk->sk_write_queue); 6371da177e4SLinus Torvalds 6381da177e4SLinus Torvalds return 0; 6391da177e4SLinus Torvalds } 6401da177e4SLinus Torvalds 6411da177e4SLinus Torvalds /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c 6421da177e4SLinus Torvalds * eventually). The difference is that pulled data not copied, but 6431da177e4SLinus Torvalds * immediately discarded. 6441da177e4SLinus Torvalds */ 645*f2911969SHerbert Xu ~{PmVHI~} static void __pskb_trim_head(struct sk_buff *skb, int len) 6461da177e4SLinus Torvalds { 6471da177e4SLinus Torvalds int i, k, eat; 6481da177e4SLinus Torvalds 6491da177e4SLinus Torvalds eat = len; 6501da177e4SLinus Torvalds k = 0; 6511da177e4SLinus Torvalds for (i=0; i<skb_shinfo(skb)->nr_frags; i++) { 6521da177e4SLinus Torvalds if (skb_shinfo(skb)->frags[i].size <= eat) { 6531da177e4SLinus Torvalds put_page(skb_shinfo(skb)->frags[i].page); 6541da177e4SLinus Torvalds eat -= skb_shinfo(skb)->frags[i].size; 6551da177e4SLinus Torvalds } else { 6561da177e4SLinus Torvalds skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i]; 6571da177e4SLinus Torvalds if (eat) { 6581da177e4SLinus Torvalds skb_shinfo(skb)->frags[k].page_offset += eat; 6591da177e4SLinus Torvalds skb_shinfo(skb)->frags[k].size -= eat; 6601da177e4SLinus Torvalds eat = 0; 6611da177e4SLinus Torvalds } 6621da177e4SLinus Torvalds k++; 6631da177e4SLinus Torvalds } 6641da177e4SLinus Torvalds } 6651da177e4SLinus Torvalds skb_shinfo(skb)->nr_frags = k; 6661da177e4SLinus Torvalds 6671da177e4SLinus Torvalds skb->tail = skb->data; 6681da177e4SLinus Torvalds skb->data_len -= len; 6691da177e4SLinus Torvalds skb->len = skb->data_len; 6701da177e4SLinus Torvalds } 6711da177e4SLinus Torvalds 6721da177e4SLinus Torvalds int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 6731da177e4SLinus Torvalds { 6741da177e4SLinus Torvalds if (skb_cloned(skb) && 6751da177e4SLinus Torvalds pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) 6761da177e4SLinus Torvalds return -ENOMEM; 6771da177e4SLinus Torvalds 678*f2911969SHerbert Xu ~{PmVHI~} /* If len == headlen, we avoid __skb_pull to preserve alignment. */ 679*f2911969SHerbert Xu ~{PmVHI~} if (unlikely(len < skb_headlen(skb))) 6801da177e4SLinus Torvalds __skb_pull(skb, len); 681*f2911969SHerbert Xu ~{PmVHI~} else 682*f2911969SHerbert Xu ~{PmVHI~} __pskb_trim_head(skb, len - skb_headlen(skb)); 6831da177e4SLinus Torvalds 6841da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq += len; 6851da177e4SLinus Torvalds skb->ip_summed = CHECKSUM_HW; 6861da177e4SLinus Torvalds 6871da177e4SLinus Torvalds skb->truesize -= len; 6881da177e4SLinus Torvalds sk->sk_wmem_queued -= len; 6891da177e4SLinus Torvalds sk->sk_forward_alloc += len; 6901da177e4SLinus Torvalds sock_set_flag(sk, SOCK_QUEUE_SHRUNK); 6911da177e4SLinus Torvalds 6921da177e4SLinus Torvalds /* Any change of skb->len requires recalculation of tso 6931da177e4SLinus Torvalds * factor and mss. 6941da177e4SLinus Torvalds */ 6951da177e4SLinus Torvalds if (tcp_skb_pcount(skb) > 1) 696846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk, 1)); 6971da177e4SLinus Torvalds 6981da177e4SLinus Torvalds return 0; 6991da177e4SLinus Torvalds } 7001da177e4SLinus Torvalds 7015d424d5aSJohn Heffner /* Not accounting for SACKs here. */ 7025d424d5aSJohn Heffner int tcp_mtu_to_mss(struct sock *sk, int pmtu) 7035d424d5aSJohn Heffner { 7045d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 7055d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 7065d424d5aSJohn Heffner int mss_now; 7075d424d5aSJohn Heffner 7085d424d5aSJohn Heffner /* Calculate base mss without TCP options: 7095d424d5aSJohn Heffner It is MMS_S - sizeof(tcphdr) of rfc1122 7105d424d5aSJohn Heffner */ 7115d424d5aSJohn Heffner mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 7125d424d5aSJohn Heffner 7135d424d5aSJohn Heffner /* Clamp it (mss_clamp does not include tcp options) */ 7145d424d5aSJohn Heffner if (mss_now > tp->rx_opt.mss_clamp) 7155d424d5aSJohn Heffner mss_now = tp->rx_opt.mss_clamp; 7165d424d5aSJohn Heffner 7175d424d5aSJohn Heffner /* Now subtract optional transport overhead */ 7185d424d5aSJohn Heffner mss_now -= icsk->icsk_ext_hdr_len; 7195d424d5aSJohn Heffner 7205d424d5aSJohn Heffner /* Then reserve room for full set of TCP options and 8 bytes of data */ 7215d424d5aSJohn Heffner if (mss_now < 48) 7225d424d5aSJohn Heffner mss_now = 48; 7235d424d5aSJohn Heffner 7245d424d5aSJohn Heffner /* Now subtract TCP options size, not including SACKs */ 7255d424d5aSJohn Heffner mss_now -= tp->tcp_header_len - sizeof(struct tcphdr); 7265d424d5aSJohn Heffner 7275d424d5aSJohn Heffner return mss_now; 7285d424d5aSJohn Heffner } 7295d424d5aSJohn Heffner 7305d424d5aSJohn Heffner /* Inverse of above */ 7315d424d5aSJohn Heffner int tcp_mss_to_mtu(struct sock *sk, int mss) 7325d424d5aSJohn Heffner { 7335d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 7345d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 7355d424d5aSJohn Heffner int mtu; 7365d424d5aSJohn Heffner 7375d424d5aSJohn Heffner mtu = mss + 7385d424d5aSJohn Heffner tp->tcp_header_len + 7395d424d5aSJohn Heffner icsk->icsk_ext_hdr_len + 7405d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 7415d424d5aSJohn Heffner 7425d424d5aSJohn Heffner return mtu; 7435d424d5aSJohn Heffner } 7445d424d5aSJohn Heffner 7455d424d5aSJohn Heffner void tcp_mtup_init(struct sock *sk) 7465d424d5aSJohn Heffner { 7475d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 7485d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 7495d424d5aSJohn Heffner 7505d424d5aSJohn Heffner icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1; 7515d424d5aSJohn Heffner icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 7525d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 7535d424d5aSJohn Heffner icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss); 7545d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 7555d424d5aSJohn Heffner } 7565d424d5aSJohn Heffner 7571da177e4SLinus Torvalds /* This function synchronize snd mss to current pmtu/exthdr set. 7581da177e4SLinus Torvalds 7591da177e4SLinus Torvalds tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 7601da177e4SLinus Torvalds for TCP options, but includes only bare TCP header. 7611da177e4SLinus Torvalds 7621da177e4SLinus Torvalds tp->rx_opt.mss_clamp is mss negotiated at connection setup. 763caa20d9aSStephen Hemminger It is minimum of user_mss and mss received with SYN. 7641da177e4SLinus Torvalds It also does not include TCP options. 7651da177e4SLinus Torvalds 766d83d8461SArnaldo Carvalho de Melo inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. 7671da177e4SLinus Torvalds 7681da177e4SLinus Torvalds tp->mss_cache is current effective sending mss, including 7691da177e4SLinus Torvalds all tcp options except for SACKs. It is evaluated, 7701da177e4SLinus Torvalds taking into account current pmtu, but never exceeds 7711da177e4SLinus Torvalds tp->rx_opt.mss_clamp. 7721da177e4SLinus Torvalds 7731da177e4SLinus Torvalds NOTE1. rfc1122 clearly states that advertised MSS 7741da177e4SLinus Torvalds DOES NOT include either tcp or ip options. 7751da177e4SLinus Torvalds 776d83d8461SArnaldo Carvalho de Melo NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache 777d83d8461SArnaldo Carvalho de Melo are READ ONLY outside this function. --ANK (980731) 7781da177e4SLinus Torvalds */ 7791da177e4SLinus Torvalds 7801da177e4SLinus Torvalds unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 7811da177e4SLinus Torvalds { 7821da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 783d83d8461SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 7845d424d5aSJohn Heffner int mss_now; 7851da177e4SLinus Torvalds 7865d424d5aSJohn Heffner if (icsk->icsk_mtup.search_high > pmtu) 7875d424d5aSJohn Heffner icsk->icsk_mtup.search_high = pmtu; 7881da177e4SLinus Torvalds 7895d424d5aSJohn Heffner mss_now = tcp_mtu_to_mss(sk, pmtu); 7901da177e4SLinus Torvalds 7911da177e4SLinus Torvalds /* Bound mss with half of window */ 7921da177e4SLinus Torvalds if (tp->max_window && mss_now > (tp->max_window>>1)) 7931da177e4SLinus Torvalds mss_now = max((tp->max_window>>1), 68U - tp->tcp_header_len); 7941da177e4SLinus Torvalds 7951da177e4SLinus Torvalds /* And store cached results */ 796d83d8461SArnaldo Carvalho de Melo icsk->icsk_pmtu_cookie = pmtu; 7975d424d5aSJohn Heffner if (icsk->icsk_mtup.enabled) 7985d424d5aSJohn Heffner mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); 799c1b4a7e6SDavid S. Miller tp->mss_cache = mss_now; 8001da177e4SLinus Torvalds 8011da177e4SLinus Torvalds return mss_now; 8021da177e4SLinus Torvalds } 8031da177e4SLinus Torvalds 8041da177e4SLinus Torvalds /* Compute the current effective MSS, taking SACKs and IP options, 8051da177e4SLinus Torvalds * and even PMTU discovery events into account. 8061da177e4SLinus Torvalds * 8071da177e4SLinus Torvalds * LARGESEND note: !urg_mode is overkill, only frames up to snd_up 8081da177e4SLinus Torvalds * cannot be large. However, taking into account rare use of URG, this 8091da177e4SLinus Torvalds * is not a big flaw. 8101da177e4SLinus Torvalds */ 811c1b4a7e6SDavid S. Miller unsigned int tcp_current_mss(struct sock *sk, int large_allowed) 8121da177e4SLinus Torvalds { 8131da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 8141da177e4SLinus Torvalds struct dst_entry *dst = __sk_dst_get(sk); 815c1b4a7e6SDavid S. Miller u32 mss_now; 816c1b4a7e6SDavid S. Miller u16 xmit_size_goal; 817c1b4a7e6SDavid S. Miller int doing_tso = 0; 8181da177e4SLinus Torvalds 819c1b4a7e6SDavid S. Miller mss_now = tp->mss_cache; 820c1b4a7e6SDavid S. Miller 821c1b4a7e6SDavid S. Miller if (large_allowed && 822c1b4a7e6SDavid S. Miller (sk->sk_route_caps & NETIF_F_TSO) && 823c1b4a7e6SDavid S. Miller !tp->urg_mode) 824c1b4a7e6SDavid S. Miller doing_tso = 1; 825c1b4a7e6SDavid S. Miller 8261da177e4SLinus Torvalds if (dst) { 8271da177e4SLinus Torvalds u32 mtu = dst_mtu(dst); 828d83d8461SArnaldo Carvalho de Melo if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 8291da177e4SLinus Torvalds mss_now = tcp_sync_mss(sk, mtu); 8301da177e4SLinus Torvalds } 8311da177e4SLinus Torvalds 8321da177e4SLinus Torvalds if (tp->rx_opt.eff_sacks) 8331da177e4SLinus Torvalds mss_now -= (TCPOLEN_SACK_BASE_ALIGNED + 8341da177e4SLinus Torvalds (tp->rx_opt.eff_sacks * TCPOLEN_SACK_PERBLOCK)); 835c1b4a7e6SDavid S. Miller 836c1b4a7e6SDavid S. Miller xmit_size_goal = mss_now; 837c1b4a7e6SDavid S. Miller 838c1b4a7e6SDavid S. Miller if (doing_tso) { 8398292a17aSArnaldo Carvalho de Melo xmit_size_goal = (65535 - 8408292a17aSArnaldo Carvalho de Melo inet_csk(sk)->icsk_af_ops->net_header_len - 841d83d8461SArnaldo Carvalho de Melo inet_csk(sk)->icsk_ext_hdr_len - 842d83d8461SArnaldo Carvalho de Melo tp->tcp_header_len); 843c1b4a7e6SDavid S. Miller 844c1b4a7e6SDavid S. Miller if (tp->max_window && 845c1b4a7e6SDavid S. Miller (xmit_size_goal > (tp->max_window >> 1))) 846c1b4a7e6SDavid S. Miller xmit_size_goal = max((tp->max_window >> 1), 847c1b4a7e6SDavid S. Miller 68U - tp->tcp_header_len); 848c1b4a7e6SDavid S. Miller 849c1b4a7e6SDavid S. Miller xmit_size_goal -= (xmit_size_goal % mss_now); 850c1b4a7e6SDavid S. Miller } 851c1b4a7e6SDavid S. Miller tp->xmit_size_goal = xmit_size_goal; 852c1b4a7e6SDavid S. Miller 8531da177e4SLinus Torvalds return mss_now; 8541da177e4SLinus Torvalds } 8551da177e4SLinus Torvalds 856a762a980SDavid S. Miller /* Congestion window validation. (RFC2861) */ 857a762a980SDavid S. Miller 85840efc6faSStephen Hemminger static void tcp_cwnd_validate(struct sock *sk, struct tcp_sock *tp) 859a762a980SDavid S. Miller { 860a762a980SDavid S. Miller __u32 packets_out = tp->packets_out; 861a762a980SDavid S. Miller 862a762a980SDavid S. Miller if (packets_out >= tp->snd_cwnd) { 863a762a980SDavid S. Miller /* Network is feed fully. */ 864a762a980SDavid S. Miller tp->snd_cwnd_used = 0; 865a762a980SDavid S. Miller tp->snd_cwnd_stamp = tcp_time_stamp; 866a762a980SDavid S. Miller } else { 867a762a980SDavid S. Miller /* Network starves. */ 868a762a980SDavid S. Miller if (tp->packets_out > tp->snd_cwnd_used) 869a762a980SDavid S. Miller tp->snd_cwnd_used = tp->packets_out; 870a762a980SDavid S. Miller 871463c84b9SArnaldo Carvalho de Melo if ((s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto) 872a762a980SDavid S. Miller tcp_cwnd_application_limited(sk); 873a762a980SDavid S. Miller } 874a762a980SDavid S. Miller } 875a762a980SDavid S. Miller 876c1b4a7e6SDavid S. Miller static unsigned int tcp_window_allows(struct tcp_sock *tp, struct sk_buff *skb, unsigned int mss_now, unsigned int cwnd) 877c1b4a7e6SDavid S. Miller { 878c1b4a7e6SDavid S. Miller u32 window, cwnd_len; 879c1b4a7e6SDavid S. Miller 880c1b4a7e6SDavid S. Miller window = (tp->snd_una + tp->snd_wnd - TCP_SKB_CB(skb)->seq); 881c1b4a7e6SDavid S. Miller cwnd_len = mss_now * cwnd; 882c1b4a7e6SDavid S. Miller return min(window, cwnd_len); 883c1b4a7e6SDavid S. Miller } 884c1b4a7e6SDavid S. Miller 885c1b4a7e6SDavid S. Miller /* Can at least one segment of SKB be sent right now, according to the 886c1b4a7e6SDavid S. Miller * congestion window rules? If so, return how many segments are allowed. 887c1b4a7e6SDavid S. Miller */ 888c1b4a7e6SDavid S. Miller static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp, struct sk_buff *skb) 889c1b4a7e6SDavid S. Miller { 890c1b4a7e6SDavid S. Miller u32 in_flight, cwnd; 891c1b4a7e6SDavid S. Miller 892c1b4a7e6SDavid S. Miller /* Don't be strict about the congestion window for the final FIN. */ 893c1b4a7e6SDavid S. Miller if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) 894c1b4a7e6SDavid S. Miller return 1; 895c1b4a7e6SDavid S. Miller 896c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 897c1b4a7e6SDavid S. Miller cwnd = tp->snd_cwnd; 898c1b4a7e6SDavid S. Miller if (in_flight < cwnd) 899c1b4a7e6SDavid S. Miller return (cwnd - in_flight); 900c1b4a7e6SDavid S. Miller 901c1b4a7e6SDavid S. Miller return 0; 902c1b4a7e6SDavid S. Miller } 903c1b4a7e6SDavid S. Miller 904c1b4a7e6SDavid S. Miller /* This must be invoked the first time we consider transmitting 905c1b4a7e6SDavid S. Miller * SKB onto the wire. 906c1b4a7e6SDavid S. Miller */ 90740efc6faSStephen Hemminger static int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb, unsigned int mss_now) 908c1b4a7e6SDavid S. Miller { 909c1b4a7e6SDavid S. Miller int tso_segs = tcp_skb_pcount(skb); 910c1b4a7e6SDavid S. Miller 911846998aeSDavid S. Miller if (!tso_segs || 912846998aeSDavid S. Miller (tso_segs > 1 && 913846998aeSDavid S. Miller skb_shinfo(skb)->tso_size != mss_now)) { 914846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 915c1b4a7e6SDavid S. Miller tso_segs = tcp_skb_pcount(skb); 916c1b4a7e6SDavid S. Miller } 917c1b4a7e6SDavid S. Miller return tso_segs; 918c1b4a7e6SDavid S. Miller } 919c1b4a7e6SDavid S. Miller 920c1b4a7e6SDavid S. Miller static inline int tcp_minshall_check(const struct tcp_sock *tp) 921c1b4a7e6SDavid S. Miller { 922c1b4a7e6SDavid S. Miller return after(tp->snd_sml,tp->snd_una) && 923c1b4a7e6SDavid S. Miller !after(tp->snd_sml, tp->snd_nxt); 924c1b4a7e6SDavid S. Miller } 925c1b4a7e6SDavid S. Miller 926c1b4a7e6SDavid S. Miller /* Return 0, if packet can be sent now without violation Nagle's rules: 927c1b4a7e6SDavid S. Miller * 1. It is full sized. 928c1b4a7e6SDavid S. Miller * 2. Or it contains FIN. (already checked by caller) 929c1b4a7e6SDavid S. Miller * 3. Or TCP_NODELAY was set. 930c1b4a7e6SDavid S. Miller * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 931c1b4a7e6SDavid S. Miller * With Minshall's modification: all sent small packets are ACKed. 932c1b4a7e6SDavid S. Miller */ 933c1b4a7e6SDavid S. Miller 934c1b4a7e6SDavid S. Miller static inline int tcp_nagle_check(const struct tcp_sock *tp, 935c1b4a7e6SDavid S. Miller const struct sk_buff *skb, 936c1b4a7e6SDavid S. Miller unsigned mss_now, int nonagle) 937c1b4a7e6SDavid S. Miller { 938c1b4a7e6SDavid S. Miller return (skb->len < mss_now && 939c1b4a7e6SDavid S. Miller ((nonagle&TCP_NAGLE_CORK) || 940c1b4a7e6SDavid S. Miller (!nonagle && 941c1b4a7e6SDavid S. Miller tp->packets_out && 942c1b4a7e6SDavid S. Miller tcp_minshall_check(tp)))); 943c1b4a7e6SDavid S. Miller } 944c1b4a7e6SDavid S. Miller 945c1b4a7e6SDavid S. Miller /* Return non-zero if the Nagle test allows this packet to be 946c1b4a7e6SDavid S. Miller * sent now. 947c1b4a7e6SDavid S. Miller */ 948c1b4a7e6SDavid S. Miller static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb, 949c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 950c1b4a7e6SDavid S. Miller { 951c1b4a7e6SDavid S. Miller /* Nagle rule does not apply to frames, which sit in the middle of the 952c1b4a7e6SDavid S. Miller * write_queue (they have no chances to get new data). 953c1b4a7e6SDavid S. Miller * 954c1b4a7e6SDavid S. Miller * This is implemented in the callers, where they modify the 'nonagle' 955c1b4a7e6SDavid S. Miller * argument based upon the location of SKB in the send queue. 956c1b4a7e6SDavid S. Miller */ 957c1b4a7e6SDavid S. Miller if (nonagle & TCP_NAGLE_PUSH) 958c1b4a7e6SDavid S. Miller return 1; 959c1b4a7e6SDavid S. Miller 960c1b4a7e6SDavid S. Miller /* Don't use the nagle rule for urgent data (or for the final FIN). */ 961c1b4a7e6SDavid S. Miller if (tp->urg_mode || 962c1b4a7e6SDavid S. Miller (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)) 963c1b4a7e6SDavid S. Miller return 1; 964c1b4a7e6SDavid S. Miller 965c1b4a7e6SDavid S. Miller if (!tcp_nagle_check(tp, skb, cur_mss, nonagle)) 966c1b4a7e6SDavid S. Miller return 1; 967c1b4a7e6SDavid S. Miller 968c1b4a7e6SDavid S. Miller return 0; 969c1b4a7e6SDavid S. Miller } 970c1b4a7e6SDavid S. Miller 971c1b4a7e6SDavid S. Miller /* Does at least the first segment of SKB fit into the send window? */ 972c1b4a7e6SDavid S. Miller static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb, unsigned int cur_mss) 973c1b4a7e6SDavid S. Miller { 974c1b4a7e6SDavid S. Miller u32 end_seq = TCP_SKB_CB(skb)->end_seq; 975c1b4a7e6SDavid S. Miller 976c1b4a7e6SDavid S. Miller if (skb->len > cur_mss) 977c1b4a7e6SDavid S. Miller end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 978c1b4a7e6SDavid S. Miller 979c1b4a7e6SDavid S. Miller return !after(end_seq, tp->snd_una + tp->snd_wnd); 980c1b4a7e6SDavid S. Miller } 981c1b4a7e6SDavid S. Miller 982c1b4a7e6SDavid S. Miller /* This checks if the data bearing packet SKB (usually sk->sk_send_head) 983c1b4a7e6SDavid S. Miller * should be put on the wire right now. If so, it returns the number of 984c1b4a7e6SDavid S. Miller * packets allowed by the congestion window. 985c1b4a7e6SDavid S. Miller */ 986c1b4a7e6SDavid S. Miller static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb, 987c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 988c1b4a7e6SDavid S. Miller { 989c1b4a7e6SDavid S. Miller struct tcp_sock *tp = tcp_sk(sk); 990c1b4a7e6SDavid S. Miller unsigned int cwnd_quota; 991c1b4a7e6SDavid S. Miller 992846998aeSDavid S. Miller tcp_init_tso_segs(sk, skb, cur_mss); 993c1b4a7e6SDavid S. Miller 994c1b4a7e6SDavid S. Miller if (!tcp_nagle_test(tp, skb, cur_mss, nonagle)) 995c1b4a7e6SDavid S. Miller return 0; 996c1b4a7e6SDavid S. Miller 997c1b4a7e6SDavid S. Miller cwnd_quota = tcp_cwnd_test(tp, skb); 998c1b4a7e6SDavid S. Miller if (cwnd_quota && 999c1b4a7e6SDavid S. Miller !tcp_snd_wnd_test(tp, skb, cur_mss)) 1000c1b4a7e6SDavid S. Miller cwnd_quota = 0; 1001c1b4a7e6SDavid S. Miller 1002c1b4a7e6SDavid S. Miller return cwnd_quota; 1003c1b4a7e6SDavid S. Miller } 1004c1b4a7e6SDavid S. Miller 1005c1b4a7e6SDavid S. Miller static inline int tcp_skb_is_last(const struct sock *sk, 1006c1b4a7e6SDavid S. Miller const struct sk_buff *skb) 1007c1b4a7e6SDavid S. Miller { 1008c1b4a7e6SDavid S. Miller return skb->next == (struct sk_buff *)&sk->sk_write_queue; 1009c1b4a7e6SDavid S. Miller } 1010c1b4a7e6SDavid S. Miller 1011c1b4a7e6SDavid S. Miller int tcp_may_send_now(struct sock *sk, struct tcp_sock *tp) 1012c1b4a7e6SDavid S. Miller { 1013c1b4a7e6SDavid S. Miller struct sk_buff *skb = sk->sk_send_head; 1014c1b4a7e6SDavid S. Miller 1015c1b4a7e6SDavid S. Miller return (skb && 1016c1b4a7e6SDavid S. Miller tcp_snd_test(sk, skb, tcp_current_mss(sk, 1), 1017c1b4a7e6SDavid S. Miller (tcp_skb_is_last(sk, skb) ? 1018c1b4a7e6SDavid S. Miller TCP_NAGLE_PUSH : 1019c1b4a7e6SDavid S. Miller tp->nonagle))); 1020c1b4a7e6SDavid S. Miller } 1021c1b4a7e6SDavid S. Miller 1022c1b4a7e6SDavid S. Miller /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 1023c1b4a7e6SDavid S. Miller * which is put after SKB on the list. It is very much like 1024c1b4a7e6SDavid S. Miller * tcp_fragment() except that it may make several kinds of assumptions 1025c1b4a7e6SDavid S. Miller * in order to speed up the splitting operation. In particular, we 1026c1b4a7e6SDavid S. Miller * know that all the data is in scatter-gather pages, and that the 1027c1b4a7e6SDavid S. Miller * packet has never been sent out before (and thus is not cloned). 1028c1b4a7e6SDavid S. Miller */ 1029846998aeSDavid S. Miller static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, unsigned int mss_now) 1030c1b4a7e6SDavid S. Miller { 1031c1b4a7e6SDavid S. Miller struct sk_buff *buff; 1032c1b4a7e6SDavid S. Miller int nlen = skb->len - len; 1033c1b4a7e6SDavid S. Miller u16 flags; 1034c1b4a7e6SDavid S. Miller 1035c1b4a7e6SDavid S. Miller /* All of a TSO frame must be composed of paged data. */ 1036c8ac3774SHerbert Xu if (skb->len != skb->data_len) 1037c8ac3774SHerbert Xu return tcp_fragment(sk, skb, len, mss_now); 1038c1b4a7e6SDavid S. Miller 1039c1b4a7e6SDavid S. Miller buff = sk_stream_alloc_pskb(sk, 0, 0, GFP_ATOMIC); 1040c1b4a7e6SDavid S. Miller if (unlikely(buff == NULL)) 1041c1b4a7e6SDavid S. Miller return -ENOMEM; 1042c1b4a7e6SDavid S. Miller 1043b60b49eaSHerbert Xu sk_charge_skb(sk, buff); 1044b60b49eaSHerbert Xu buff->truesize += nlen; 1045c1b4a7e6SDavid S. Miller skb->truesize -= nlen; 1046c1b4a7e6SDavid S. Miller 1047c1b4a7e6SDavid S. Miller /* Correct the sequence numbers. */ 1048c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 1049c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 1050c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 1051c1b4a7e6SDavid S. Miller 1052c1b4a7e6SDavid S. Miller /* PSH and FIN should only be set in the second packet. */ 1053c1b4a7e6SDavid S. Miller flags = TCP_SKB_CB(skb)->flags; 1054c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH); 1055c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->flags = flags; 1056c1b4a7e6SDavid S. Miller 1057c1b4a7e6SDavid S. Miller /* This packet was never sent out yet, so no SACK bits. */ 1058c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->sacked = 0; 1059c1b4a7e6SDavid S. Miller 1060c1b4a7e6SDavid S. Miller buff->ip_summed = skb->ip_summed = CHECKSUM_HW; 1061c1b4a7e6SDavid S. Miller skb_split(skb, buff, len); 1062c1b4a7e6SDavid S. Miller 1063c1b4a7e6SDavid S. Miller /* Fix up tso_factor for both original and new SKB. */ 1064846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1065846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, buff, mss_now); 1066c1b4a7e6SDavid S. Miller 1067c1b4a7e6SDavid S. Miller /* Link BUFF into the send queue. */ 1068c1b4a7e6SDavid S. Miller skb_header_release(buff); 10698728b834SDavid S. Miller __skb_append(skb, buff, &sk->sk_write_queue); 1070c1b4a7e6SDavid S. Miller 1071c1b4a7e6SDavid S. Miller return 0; 1072c1b4a7e6SDavid S. Miller } 1073c1b4a7e6SDavid S. Miller 1074c1b4a7e6SDavid S. Miller /* Try to defer sending, if possible, in order to minimize the amount 1075c1b4a7e6SDavid S. Miller * of TSO splitting we do. View it as a kind of TSO Nagle test. 1076c1b4a7e6SDavid S. Miller * 1077c1b4a7e6SDavid S. Miller * This algorithm is from John Heffner. 1078c1b4a7e6SDavid S. Miller */ 1079c1b4a7e6SDavid S. Miller static int tcp_tso_should_defer(struct sock *sk, struct tcp_sock *tp, struct sk_buff *skb) 1080c1b4a7e6SDavid S. Miller { 10816687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 1082c1b4a7e6SDavid S. Miller u32 send_win, cong_win, limit, in_flight; 1083c1b4a7e6SDavid S. Miller 1084c1b4a7e6SDavid S. Miller if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) 1085c1b4a7e6SDavid S. Miller return 0; 1086c1b4a7e6SDavid S. Miller 10876687e988SArnaldo Carvalho de Melo if (icsk->icsk_ca_state != TCP_CA_Open) 1088908a75c1SDavid S. Miller return 0; 1089908a75c1SDavid S. Miller 1090c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1091c1b4a7e6SDavid S. Miller 1092c1b4a7e6SDavid S. Miller BUG_ON(tcp_skb_pcount(skb) <= 1 || 1093c1b4a7e6SDavid S. Miller (tp->snd_cwnd <= in_flight)); 1094c1b4a7e6SDavid S. Miller 1095c1b4a7e6SDavid S. Miller send_win = (tp->snd_una + tp->snd_wnd) - TCP_SKB_CB(skb)->seq; 1096c1b4a7e6SDavid S. Miller 1097c1b4a7e6SDavid S. Miller /* From in_flight test above, we know that cwnd > in_flight. */ 1098c1b4a7e6SDavid S. Miller cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; 1099c1b4a7e6SDavid S. Miller 1100c1b4a7e6SDavid S. Miller limit = min(send_win, cong_win); 1101c1b4a7e6SDavid S. Miller 1102ba244fe9SDavid S. Miller /* If a full-sized TSO skb can be sent, do it. */ 1103ba244fe9SDavid S. Miller if (limit >= 65536) 1104ba244fe9SDavid S. Miller return 0; 1105ba244fe9SDavid S. Miller 1106c1b4a7e6SDavid S. Miller if (sysctl_tcp_tso_win_divisor) { 1107c1b4a7e6SDavid S. Miller u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); 1108c1b4a7e6SDavid S. Miller 1109c1b4a7e6SDavid S. Miller /* If at least some fraction of a window is available, 1110c1b4a7e6SDavid S. Miller * just use it. 1111c1b4a7e6SDavid S. Miller */ 1112c1b4a7e6SDavid S. Miller chunk /= sysctl_tcp_tso_win_divisor; 1113c1b4a7e6SDavid S. Miller if (limit >= chunk) 1114c1b4a7e6SDavid S. Miller return 0; 1115c1b4a7e6SDavid S. Miller } else { 1116c1b4a7e6SDavid S. Miller /* Different approach, try not to defer past a single 1117c1b4a7e6SDavid S. Miller * ACK. Receiver should ACK every other full sized 1118c1b4a7e6SDavid S. Miller * frame, so if we have space for more than 3 frames 1119c1b4a7e6SDavid S. Miller * then send now. 1120c1b4a7e6SDavid S. Miller */ 1121c1b4a7e6SDavid S. Miller if (limit > tcp_max_burst(tp) * tp->mss_cache) 1122c1b4a7e6SDavid S. Miller return 0; 1123c1b4a7e6SDavid S. Miller } 1124c1b4a7e6SDavid S. Miller 1125c1b4a7e6SDavid S. Miller /* Ok, it looks like it is advisable to defer. */ 1126c1b4a7e6SDavid S. Miller return 1; 1127c1b4a7e6SDavid S. Miller } 1128c1b4a7e6SDavid S. Miller 11295d424d5aSJohn Heffner /* Create a new MTU probe if we are ready. 11305d424d5aSJohn Heffner * Returns 0 if we should wait to probe (no cwnd available), 11315d424d5aSJohn Heffner * 1 if a probe was sent, 11325d424d5aSJohn Heffner * -1 otherwise */ 11335d424d5aSJohn Heffner static int tcp_mtu_probe(struct sock *sk) 11345d424d5aSJohn Heffner { 11355d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 11365d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 11375d424d5aSJohn Heffner struct sk_buff *skb, *nskb, *next; 11385d424d5aSJohn Heffner int len; 11395d424d5aSJohn Heffner int probe_size; 11405d424d5aSJohn Heffner unsigned int pif; 11415d424d5aSJohn Heffner int copy; 11425d424d5aSJohn Heffner int mss_now; 11435d424d5aSJohn Heffner 11445d424d5aSJohn Heffner /* Not currently probing/verifying, 11455d424d5aSJohn Heffner * not in recovery, 11465d424d5aSJohn Heffner * have enough cwnd, and 11475d424d5aSJohn Heffner * not SACKing (the variable headers throw things off) */ 11485d424d5aSJohn Heffner if (!icsk->icsk_mtup.enabled || 11495d424d5aSJohn Heffner icsk->icsk_mtup.probe_size || 11505d424d5aSJohn Heffner inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 11515d424d5aSJohn Heffner tp->snd_cwnd < 11 || 11525d424d5aSJohn Heffner tp->rx_opt.eff_sacks) 11535d424d5aSJohn Heffner return -1; 11545d424d5aSJohn Heffner 11555d424d5aSJohn Heffner /* Very simple search strategy: just double the MSS. */ 11565d424d5aSJohn Heffner mss_now = tcp_current_mss(sk, 0); 11575d424d5aSJohn Heffner probe_size = 2*tp->mss_cache; 11585d424d5aSJohn Heffner if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) { 11595d424d5aSJohn Heffner /* TODO: set timer for probe_converge_event */ 11605d424d5aSJohn Heffner return -1; 11615d424d5aSJohn Heffner } 11625d424d5aSJohn Heffner 11635d424d5aSJohn Heffner /* Have enough data in the send queue to probe? */ 11645d424d5aSJohn Heffner len = 0; 11655d424d5aSJohn Heffner if ((skb = sk->sk_send_head) == NULL) 11665d424d5aSJohn Heffner return -1; 11675d424d5aSJohn Heffner while ((len += skb->len) < probe_size && !tcp_skb_is_last(sk, skb)) 11685d424d5aSJohn Heffner skb = skb->next; 11695d424d5aSJohn Heffner if (len < probe_size) 11705d424d5aSJohn Heffner return -1; 11715d424d5aSJohn Heffner 11725d424d5aSJohn Heffner /* Receive window check. */ 11735d424d5aSJohn Heffner if (after(TCP_SKB_CB(skb)->seq + probe_size, tp->snd_una + tp->snd_wnd)) { 11745d424d5aSJohn Heffner if (tp->snd_wnd < probe_size) 11755d424d5aSJohn Heffner return -1; 11765d424d5aSJohn Heffner else 11775d424d5aSJohn Heffner return 0; 11785d424d5aSJohn Heffner } 11795d424d5aSJohn Heffner 11805d424d5aSJohn Heffner /* Do we need to wait to drain cwnd? */ 11815d424d5aSJohn Heffner pif = tcp_packets_in_flight(tp); 11825d424d5aSJohn Heffner if (pif + 2 > tp->snd_cwnd) { 11835d424d5aSJohn Heffner /* With no packets in flight, don't stall. */ 11845d424d5aSJohn Heffner if (pif == 0) 11855d424d5aSJohn Heffner return -1; 11865d424d5aSJohn Heffner else 11875d424d5aSJohn Heffner return 0; 11885d424d5aSJohn Heffner } 11895d424d5aSJohn Heffner 11905d424d5aSJohn Heffner /* We're allowed to probe. Build it now. */ 11915d424d5aSJohn Heffner if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL) 11925d424d5aSJohn Heffner return -1; 11935d424d5aSJohn Heffner sk_charge_skb(sk, nskb); 11945d424d5aSJohn Heffner 11955d424d5aSJohn Heffner skb = sk->sk_send_head; 11965d424d5aSJohn Heffner __skb_insert(nskb, skb->prev, skb, &sk->sk_write_queue); 11975d424d5aSJohn Heffner sk->sk_send_head = nskb; 11985d424d5aSJohn Heffner 11995d424d5aSJohn Heffner TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 12005d424d5aSJohn Heffner TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 12015d424d5aSJohn Heffner TCP_SKB_CB(nskb)->flags = TCPCB_FLAG_ACK; 12025d424d5aSJohn Heffner TCP_SKB_CB(nskb)->sacked = 0; 12035d424d5aSJohn Heffner nskb->csum = 0; 12045d424d5aSJohn Heffner if (skb->ip_summed == CHECKSUM_HW) 12055d424d5aSJohn Heffner nskb->ip_summed = CHECKSUM_HW; 12065d424d5aSJohn Heffner 12075d424d5aSJohn Heffner len = 0; 12085d424d5aSJohn Heffner while (len < probe_size) { 12095d424d5aSJohn Heffner next = skb->next; 12105d424d5aSJohn Heffner 12115d424d5aSJohn Heffner copy = min_t(int, skb->len, probe_size - len); 12125d424d5aSJohn Heffner if (nskb->ip_summed) 12135d424d5aSJohn Heffner skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); 12145d424d5aSJohn Heffner else 12155d424d5aSJohn Heffner nskb->csum = skb_copy_and_csum_bits(skb, 0, 12165d424d5aSJohn Heffner skb_put(nskb, copy), copy, nskb->csum); 12175d424d5aSJohn Heffner 12185d424d5aSJohn Heffner if (skb->len <= copy) { 12195d424d5aSJohn Heffner /* We've eaten all the data from this skb. 12205d424d5aSJohn Heffner * Throw it away. */ 12215d424d5aSJohn Heffner TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags; 12225d424d5aSJohn Heffner __skb_unlink(skb, &sk->sk_write_queue); 12235d424d5aSJohn Heffner sk_stream_free_skb(sk, skb); 12245d424d5aSJohn Heffner } else { 12255d424d5aSJohn Heffner TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags & 12265d424d5aSJohn Heffner ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH); 12275d424d5aSJohn Heffner if (!skb_shinfo(skb)->nr_frags) { 12285d424d5aSJohn Heffner skb_pull(skb, copy); 12295d424d5aSJohn Heffner if (skb->ip_summed != CHECKSUM_HW) 12305d424d5aSJohn Heffner skb->csum = csum_partial(skb->data, skb->len, 0); 12315d424d5aSJohn Heffner } else { 12325d424d5aSJohn Heffner __pskb_trim_head(skb, copy); 12335d424d5aSJohn Heffner tcp_set_skb_tso_segs(sk, skb, mss_now); 12345d424d5aSJohn Heffner } 12355d424d5aSJohn Heffner TCP_SKB_CB(skb)->seq += copy; 12365d424d5aSJohn Heffner } 12375d424d5aSJohn Heffner 12385d424d5aSJohn Heffner len += copy; 12395d424d5aSJohn Heffner skb = next; 12405d424d5aSJohn Heffner } 12415d424d5aSJohn Heffner tcp_init_tso_segs(sk, nskb, nskb->len); 12425d424d5aSJohn Heffner 12435d424d5aSJohn Heffner /* We're ready to send. If this fails, the probe will 12445d424d5aSJohn Heffner * be resegmented into mss-sized pieces by tcp_write_xmit(). */ 12455d424d5aSJohn Heffner TCP_SKB_CB(nskb)->when = tcp_time_stamp; 12465d424d5aSJohn Heffner if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 12475d424d5aSJohn Heffner /* Decrement cwnd here because we are sending 12485d424d5aSJohn Heffner * effectively two packets. */ 12495d424d5aSJohn Heffner tp->snd_cwnd--; 12505d424d5aSJohn Heffner update_send_head(sk, tp, nskb); 12515d424d5aSJohn Heffner 12525d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 12530e7b1368SJohn Heffner tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 12540e7b1368SJohn Heffner tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 12555d424d5aSJohn Heffner 12565d424d5aSJohn Heffner return 1; 12575d424d5aSJohn Heffner } 12585d424d5aSJohn Heffner 12595d424d5aSJohn Heffner return -1; 12605d424d5aSJohn Heffner } 12615d424d5aSJohn Heffner 12625d424d5aSJohn Heffner 12631da177e4SLinus Torvalds /* This routine writes packets to the network. It advances the 12641da177e4SLinus Torvalds * send_head. This happens as incoming acks open up the remote 12651da177e4SLinus Torvalds * window for us. 12661da177e4SLinus Torvalds * 12671da177e4SLinus Torvalds * Returns 1, if no segments are in flight and we have queued segments, but 12681da177e4SLinus Torvalds * cannot send anything now because of SWS or another problem. 12691da177e4SLinus Torvalds */ 1270a2e2a59cSDavid S. Miller static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle) 12711da177e4SLinus Torvalds { 12721da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 127392df7b51SDavid S. Miller struct sk_buff *skb; 1274c1b4a7e6SDavid S. Miller unsigned int tso_segs, sent_pkts; 1275c1b4a7e6SDavid S. Miller int cwnd_quota; 12765d424d5aSJohn Heffner int result; 12771da177e4SLinus Torvalds 12781da177e4SLinus Torvalds /* If we are closed, the bytes will have to remain here. 12791da177e4SLinus Torvalds * In time closedown will finish, we empty the write queue and all 12801da177e4SLinus Torvalds * will be happy. 12811da177e4SLinus Torvalds */ 128292df7b51SDavid S. Miller if (unlikely(sk->sk_state == TCP_CLOSE)) 128392df7b51SDavid S. Miller return 0; 128492df7b51SDavid S. Miller 1285c1b4a7e6SDavid S. Miller sent_pkts = 0; 12865d424d5aSJohn Heffner 12875d424d5aSJohn Heffner /* Do MTU probing. */ 12885d424d5aSJohn Heffner if ((result = tcp_mtu_probe(sk)) == 0) { 12895d424d5aSJohn Heffner return 0; 12905d424d5aSJohn Heffner } else if (result > 0) { 12915d424d5aSJohn Heffner sent_pkts = 1; 12925d424d5aSJohn Heffner } 12935d424d5aSJohn Heffner 1294b68e9f85SHerbert Xu while ((skb = sk->sk_send_head)) { 1295c8ac3774SHerbert Xu unsigned int limit; 1296c8ac3774SHerbert Xu 1297b68e9f85SHerbert Xu tso_segs = tcp_init_tso_segs(sk, skb, mss_now); 1298c1b4a7e6SDavid S. Miller BUG_ON(!tso_segs); 1299c1b4a7e6SDavid S. Miller 1300b68e9f85SHerbert Xu cwnd_quota = tcp_cwnd_test(tp, skb); 1301b68e9f85SHerbert Xu if (!cwnd_quota) 1302b68e9f85SHerbert Xu break; 1303b68e9f85SHerbert Xu 1304b68e9f85SHerbert Xu if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) 1305b68e9f85SHerbert Xu break; 1306b68e9f85SHerbert Xu 1307c1b4a7e6SDavid S. Miller if (tso_segs == 1) { 1308aa93466bSDavid S. Miller if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 1309aa93466bSDavid S. Miller (tcp_skb_is_last(sk, skb) ? 1310aa93466bSDavid S. Miller nonagle : TCP_NAGLE_PUSH)))) 1311aa93466bSDavid S. Miller break; 1312c1b4a7e6SDavid S. Miller } else { 1313c1b4a7e6SDavid S. Miller if (tcp_tso_should_defer(sk, tp, skb)) 1314aa93466bSDavid S. Miller break; 1315c1b4a7e6SDavid S. Miller } 1316aa93466bSDavid S. Miller 1317c8ac3774SHerbert Xu limit = mss_now; 1318c1b4a7e6SDavid S. Miller if (tso_segs > 1) { 1319c8ac3774SHerbert Xu limit = tcp_window_allows(tp, skb, 1320c1b4a7e6SDavid S. Miller mss_now, cwnd_quota); 1321c1b4a7e6SDavid S. Miller 1322c1b4a7e6SDavid S. Miller if (skb->len < limit) { 1323c1b4a7e6SDavid S. Miller unsigned int trim = skb->len % mss_now; 1324c1b4a7e6SDavid S. Miller 1325c1b4a7e6SDavid S. Miller if (trim) 1326c1b4a7e6SDavid S. Miller limit = skb->len - trim; 1327c1b4a7e6SDavid S. Miller } 1328c1b4a7e6SDavid S. Miller } 1329c8ac3774SHerbert Xu 1330c8ac3774SHerbert Xu if (skb->len > limit && 1331c8ac3774SHerbert Xu unlikely(tso_fragment(sk, skb, limit, mss_now))) 13321da177e4SLinus Torvalds break; 13331da177e4SLinus Torvalds 13341da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 1335c1b4a7e6SDavid S. Miller 1336dfb4b9dcSDavid S. Miller if (unlikely(tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC))) 13371da177e4SLinus Torvalds break; 13381da177e4SLinus Torvalds 13391da177e4SLinus Torvalds /* Advance the send_head. This one is sent out. 13401da177e4SLinus Torvalds * This call will increment packets_out. 13411da177e4SLinus Torvalds */ 13421da177e4SLinus Torvalds update_send_head(sk, tp, skb); 13431da177e4SLinus Torvalds 13441da177e4SLinus Torvalds tcp_minshall_update(tp, mss_now, skb); 1345aa93466bSDavid S. Miller sent_pkts++; 13461da177e4SLinus Torvalds } 13471da177e4SLinus Torvalds 1348aa93466bSDavid S. Miller if (likely(sent_pkts)) { 13491da177e4SLinus Torvalds tcp_cwnd_validate(sk, tp); 13501da177e4SLinus Torvalds return 0; 13511da177e4SLinus Torvalds } 13521da177e4SLinus Torvalds return !tp->packets_out && sk->sk_send_head; 13531da177e4SLinus Torvalds } 13541da177e4SLinus Torvalds 1355a762a980SDavid S. Miller /* Push out any pending frames which were held back due to 1356a762a980SDavid S. Miller * TCP_CORK or attempt at coalescing tiny packets. 1357a762a980SDavid S. Miller * The socket must be locked by the caller. 1358a762a980SDavid S. Miller */ 1359a762a980SDavid S. Miller void __tcp_push_pending_frames(struct sock *sk, struct tcp_sock *tp, 1360a2e2a59cSDavid S. Miller unsigned int cur_mss, int nonagle) 1361a762a980SDavid S. Miller { 1362a762a980SDavid S. Miller struct sk_buff *skb = sk->sk_send_head; 1363a762a980SDavid S. Miller 1364a762a980SDavid S. Miller if (skb) { 136555c97f3eSDavid S. Miller if (tcp_write_xmit(sk, cur_mss, nonagle)) 1366a762a980SDavid S. Miller tcp_check_probe_timer(sk, tp); 1367a762a980SDavid S. Miller } 1368a762a980SDavid S. Miller } 1369a762a980SDavid S. Miller 1370c1b4a7e6SDavid S. Miller /* Send _single_ skb sitting at the send head. This function requires 1371c1b4a7e6SDavid S. Miller * true push pending frames to setup probe timer etc. 1372c1b4a7e6SDavid S. Miller */ 1373c1b4a7e6SDavid S. Miller void tcp_push_one(struct sock *sk, unsigned int mss_now) 1374c1b4a7e6SDavid S. Miller { 1375c1b4a7e6SDavid S. Miller struct tcp_sock *tp = tcp_sk(sk); 1376c1b4a7e6SDavid S. Miller struct sk_buff *skb = sk->sk_send_head; 1377c1b4a7e6SDavid S. Miller unsigned int tso_segs, cwnd_quota; 1378c1b4a7e6SDavid S. Miller 1379c1b4a7e6SDavid S. Miller BUG_ON(!skb || skb->len < mss_now); 1380c1b4a7e6SDavid S. Miller 1381846998aeSDavid S. Miller tso_segs = tcp_init_tso_segs(sk, skb, mss_now); 1382c1b4a7e6SDavid S. Miller cwnd_quota = tcp_snd_test(sk, skb, mss_now, TCP_NAGLE_PUSH); 1383c1b4a7e6SDavid S. Miller 1384c1b4a7e6SDavid S. Miller if (likely(cwnd_quota)) { 1385c8ac3774SHerbert Xu unsigned int limit; 1386c8ac3774SHerbert Xu 1387c1b4a7e6SDavid S. Miller BUG_ON(!tso_segs); 1388c1b4a7e6SDavid S. Miller 1389c8ac3774SHerbert Xu limit = mss_now; 1390c1b4a7e6SDavid S. Miller if (tso_segs > 1) { 1391c8ac3774SHerbert Xu limit = tcp_window_allows(tp, skb, 1392c1b4a7e6SDavid S. Miller mss_now, cwnd_quota); 1393c1b4a7e6SDavid S. Miller 1394c1b4a7e6SDavid S. Miller if (skb->len < limit) { 1395c1b4a7e6SDavid S. Miller unsigned int trim = skb->len % mss_now; 1396c1b4a7e6SDavid S. Miller 1397c1b4a7e6SDavid S. Miller if (trim) 1398c1b4a7e6SDavid S. Miller limit = skb->len - trim; 1399c1b4a7e6SDavid S. Miller } 1400c1b4a7e6SDavid S. Miller } 1401c8ac3774SHerbert Xu 1402c8ac3774SHerbert Xu if (skb->len > limit && 1403c8ac3774SHerbert Xu unlikely(tso_fragment(sk, skb, limit, mss_now))) 1404c1b4a7e6SDavid S. Miller return; 1405c1b4a7e6SDavid S. Miller 1406c1b4a7e6SDavid S. Miller /* Send it out now. */ 1407c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->when = tcp_time_stamp; 1408c1b4a7e6SDavid S. Miller 1409dfb4b9dcSDavid S. Miller if (likely(!tcp_transmit_skb(sk, skb, 1, sk->sk_allocation))) { 1410c1b4a7e6SDavid S. Miller update_send_head(sk, tp, skb); 1411c1b4a7e6SDavid S. Miller tcp_cwnd_validate(sk, tp); 1412c1b4a7e6SDavid S. Miller return; 1413c1b4a7e6SDavid S. Miller } 1414c1b4a7e6SDavid S. Miller } 1415c1b4a7e6SDavid S. Miller } 1416c1b4a7e6SDavid S. Miller 14171da177e4SLinus Torvalds /* This function returns the amount that we can raise the 14181da177e4SLinus Torvalds * usable window based on the following constraints 14191da177e4SLinus Torvalds * 14201da177e4SLinus Torvalds * 1. The window can never be shrunk once it is offered (RFC 793) 14211da177e4SLinus Torvalds * 2. We limit memory per socket 14221da177e4SLinus Torvalds * 14231da177e4SLinus Torvalds * RFC 1122: 14241da177e4SLinus Torvalds * "the suggested [SWS] avoidance algorithm for the receiver is to keep 14251da177e4SLinus Torvalds * RECV.NEXT + RCV.WIN fixed until: 14261da177e4SLinus Torvalds * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 14271da177e4SLinus Torvalds * 14281da177e4SLinus Torvalds * i.e. don't raise the right edge of the window until you can raise 14291da177e4SLinus Torvalds * it at least MSS bytes. 14301da177e4SLinus Torvalds * 14311da177e4SLinus Torvalds * Unfortunately, the recommended algorithm breaks header prediction, 14321da177e4SLinus Torvalds * since header prediction assumes th->window stays fixed. 14331da177e4SLinus Torvalds * 14341da177e4SLinus Torvalds * Strictly speaking, keeping th->window fixed violates the receiver 14351da177e4SLinus Torvalds * side SWS prevention criteria. The problem is that under this rule 14361da177e4SLinus Torvalds * a stream of single byte packets will cause the right side of the 14371da177e4SLinus Torvalds * window to always advance by a single byte. 14381da177e4SLinus Torvalds * 14391da177e4SLinus Torvalds * Of course, if the sender implements sender side SWS prevention 14401da177e4SLinus Torvalds * then this will not be a problem. 14411da177e4SLinus Torvalds * 14421da177e4SLinus Torvalds * BSD seems to make the following compromise: 14431da177e4SLinus Torvalds * 14441da177e4SLinus Torvalds * If the free space is less than the 1/4 of the maximum 14451da177e4SLinus Torvalds * space available and the free space is less than 1/2 mss, 14461da177e4SLinus Torvalds * then set the window to 0. 14471da177e4SLinus Torvalds * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 14481da177e4SLinus Torvalds * Otherwise, just prevent the window from shrinking 14491da177e4SLinus Torvalds * and from being larger than the largest representable value. 14501da177e4SLinus Torvalds * 14511da177e4SLinus Torvalds * This prevents incremental opening of the window in the regime 14521da177e4SLinus Torvalds * where TCP is limited by the speed of the reader side taking 14531da177e4SLinus Torvalds * data out of the TCP receive queue. It does nothing about 14541da177e4SLinus Torvalds * those cases where the window is constrained on the sender side 14551da177e4SLinus Torvalds * because the pipeline is full. 14561da177e4SLinus Torvalds * 14571da177e4SLinus Torvalds * BSD also seems to "accidentally" limit itself to windows that are a 14581da177e4SLinus Torvalds * multiple of MSS, at least until the free space gets quite small. 14591da177e4SLinus Torvalds * This would appear to be a side effect of the mbuf implementation. 14601da177e4SLinus Torvalds * Combining these two algorithms results in the observed behavior 14611da177e4SLinus Torvalds * of having a fixed window size at almost all times. 14621da177e4SLinus Torvalds * 14631da177e4SLinus Torvalds * Below we obtain similar behavior by forcing the offered window to 14641da177e4SLinus Torvalds * a multiple of the mss when it is feasible to do so. 14651da177e4SLinus Torvalds * 14661da177e4SLinus Torvalds * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 14671da177e4SLinus Torvalds * Regular options like TIMESTAMP are taken into account. 14681da177e4SLinus Torvalds */ 14691da177e4SLinus Torvalds u32 __tcp_select_window(struct sock *sk) 14701da177e4SLinus Torvalds { 1471463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 14721da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1473caa20d9aSStephen Hemminger /* MSS for the peer's data. Previous versions used mss_clamp 14741da177e4SLinus Torvalds * here. I don't know if the value based on our guesses 14751da177e4SLinus Torvalds * of peer's MSS is better for the performance. It's more correct 14761da177e4SLinus Torvalds * but may be worse for the performance because of rcv_mss 14771da177e4SLinus Torvalds * fluctuations. --SAW 1998/11/1 14781da177e4SLinus Torvalds */ 1479463c84b9SArnaldo Carvalho de Melo int mss = icsk->icsk_ack.rcv_mss; 14801da177e4SLinus Torvalds int free_space = tcp_space(sk); 14811da177e4SLinus Torvalds int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk)); 14821da177e4SLinus Torvalds int window; 14831da177e4SLinus Torvalds 14841da177e4SLinus Torvalds if (mss > full_space) 14851da177e4SLinus Torvalds mss = full_space; 14861da177e4SLinus Torvalds 14871da177e4SLinus Torvalds if (free_space < full_space/2) { 1488463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.quick = 0; 14891da177e4SLinus Torvalds 14901da177e4SLinus Torvalds if (tcp_memory_pressure) 14911da177e4SLinus Torvalds tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U*tp->advmss); 14921da177e4SLinus Torvalds 14931da177e4SLinus Torvalds if (free_space < mss) 14941da177e4SLinus Torvalds return 0; 14951da177e4SLinus Torvalds } 14961da177e4SLinus Torvalds 14971da177e4SLinus Torvalds if (free_space > tp->rcv_ssthresh) 14981da177e4SLinus Torvalds free_space = tp->rcv_ssthresh; 14991da177e4SLinus Torvalds 15001da177e4SLinus Torvalds /* Don't do rounding if we are using window scaling, since the 15011da177e4SLinus Torvalds * scaled window will not line up with the MSS boundary anyway. 15021da177e4SLinus Torvalds */ 15031da177e4SLinus Torvalds window = tp->rcv_wnd; 15041da177e4SLinus Torvalds if (tp->rx_opt.rcv_wscale) { 15051da177e4SLinus Torvalds window = free_space; 15061da177e4SLinus Torvalds 15071da177e4SLinus Torvalds /* Advertise enough space so that it won't get scaled away. 15081da177e4SLinus Torvalds * Import case: prevent zero window announcement if 15091da177e4SLinus Torvalds * 1<<rcv_wscale > mss. 15101da177e4SLinus Torvalds */ 15111da177e4SLinus Torvalds if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window) 15121da177e4SLinus Torvalds window = (((window >> tp->rx_opt.rcv_wscale) + 1) 15131da177e4SLinus Torvalds << tp->rx_opt.rcv_wscale); 15141da177e4SLinus Torvalds } else { 15151da177e4SLinus Torvalds /* Get the largest window that is a nice multiple of mss. 15161da177e4SLinus Torvalds * Window clamp already applied above. 15171da177e4SLinus Torvalds * If our current window offering is within 1 mss of the 15181da177e4SLinus Torvalds * free space we just keep it. This prevents the divide 15191da177e4SLinus Torvalds * and multiply from happening most of the time. 15201da177e4SLinus Torvalds * We also don't do any window rounding when the free space 15211da177e4SLinus Torvalds * is too small. 15221da177e4SLinus Torvalds */ 15231da177e4SLinus Torvalds if (window <= free_space - mss || window > free_space) 15241da177e4SLinus Torvalds window = (free_space/mss)*mss; 15251da177e4SLinus Torvalds } 15261da177e4SLinus Torvalds 15271da177e4SLinus Torvalds return window; 15281da177e4SLinus Torvalds } 15291da177e4SLinus Torvalds 15301da177e4SLinus Torvalds /* Attempt to collapse two adjacent SKB's during retransmission. */ 15311da177e4SLinus Torvalds static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *skb, int mss_now) 15321da177e4SLinus Torvalds { 15331da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 15341da177e4SLinus Torvalds struct sk_buff *next_skb = skb->next; 15351da177e4SLinus Torvalds 15361da177e4SLinus Torvalds /* The first test we must make is that neither of these two 15371da177e4SLinus Torvalds * SKB's are still referenced by someone else. 15381da177e4SLinus Torvalds */ 15391da177e4SLinus Torvalds if (!skb_cloned(skb) && !skb_cloned(next_skb)) { 15401da177e4SLinus Torvalds int skb_size = skb->len, next_skb_size = next_skb->len; 15411da177e4SLinus Torvalds u16 flags = TCP_SKB_CB(skb)->flags; 15421da177e4SLinus Torvalds 15431da177e4SLinus Torvalds /* Also punt if next skb has been SACK'd. */ 15441da177e4SLinus Torvalds if(TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_ACKED) 15451da177e4SLinus Torvalds return; 15461da177e4SLinus Torvalds 15471da177e4SLinus Torvalds /* Next skb is out of window. */ 15481da177e4SLinus Torvalds if (after(TCP_SKB_CB(next_skb)->end_seq, tp->snd_una+tp->snd_wnd)) 15491da177e4SLinus Torvalds return; 15501da177e4SLinus Torvalds 15511da177e4SLinus Torvalds /* Punt if not enough space exists in the first SKB for 15521da177e4SLinus Torvalds * the data in the second, or the total combined payload 15531da177e4SLinus Torvalds * would exceed the MSS. 15541da177e4SLinus Torvalds */ 15551da177e4SLinus Torvalds if ((next_skb_size > skb_tailroom(skb)) || 15561da177e4SLinus Torvalds ((skb_size + next_skb_size) > mss_now)) 15571da177e4SLinus Torvalds return; 15581da177e4SLinus Torvalds 15591da177e4SLinus Torvalds BUG_ON(tcp_skb_pcount(skb) != 1 || 15601da177e4SLinus Torvalds tcp_skb_pcount(next_skb) != 1); 15611da177e4SLinus Torvalds 15626a438bbeSStephen Hemminger /* changing transmit queue under us so clear hints */ 15636a438bbeSStephen Hemminger clear_all_retrans_hints(tp); 15646a438bbeSStephen Hemminger 15651da177e4SLinus Torvalds /* Ok. We will be able to collapse the packet. */ 15668728b834SDavid S. Miller __skb_unlink(next_skb, &sk->sk_write_queue); 15671da177e4SLinus Torvalds 15681da177e4SLinus Torvalds memcpy(skb_put(skb, next_skb_size), next_skb->data, next_skb_size); 15691da177e4SLinus Torvalds 15701da177e4SLinus Torvalds if (next_skb->ip_summed == CHECKSUM_HW) 15711da177e4SLinus Torvalds skb->ip_summed = CHECKSUM_HW; 15721da177e4SLinus Torvalds 15731da177e4SLinus Torvalds if (skb->ip_summed != CHECKSUM_HW) 15741da177e4SLinus Torvalds skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); 15751da177e4SLinus Torvalds 15761da177e4SLinus Torvalds /* Update sequence range on original skb. */ 15771da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 15781da177e4SLinus Torvalds 15791da177e4SLinus Torvalds /* Merge over control information. */ 15801da177e4SLinus Torvalds flags |= TCP_SKB_CB(next_skb)->flags; /* This moves PSH/FIN etc. over */ 15811da177e4SLinus Torvalds TCP_SKB_CB(skb)->flags = flags; 15821da177e4SLinus Torvalds 15831da177e4SLinus Torvalds /* All done, get rid of second SKB and account for it so 15841da177e4SLinus Torvalds * packet counting does not break. 15851da177e4SLinus Torvalds */ 15861da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked&(TCPCB_EVER_RETRANS|TCPCB_AT_TAIL); 15871da177e4SLinus Torvalds if (TCP_SKB_CB(next_skb)->sacked&TCPCB_SACKED_RETRANS) 15881da177e4SLinus Torvalds tp->retrans_out -= tcp_skb_pcount(next_skb); 15891da177e4SLinus Torvalds if (TCP_SKB_CB(next_skb)->sacked&TCPCB_LOST) { 15901da177e4SLinus Torvalds tp->lost_out -= tcp_skb_pcount(next_skb); 15911da177e4SLinus Torvalds tp->left_out -= tcp_skb_pcount(next_skb); 15921da177e4SLinus Torvalds } 15931da177e4SLinus Torvalds /* Reno case is special. Sigh... */ 15941da177e4SLinus Torvalds if (!tp->rx_opt.sack_ok && tp->sacked_out) { 15951da177e4SLinus Torvalds tcp_dec_pcount_approx(&tp->sacked_out, next_skb); 15961da177e4SLinus Torvalds tp->left_out -= tcp_skb_pcount(next_skb); 15971da177e4SLinus Torvalds } 15981da177e4SLinus Torvalds 15991da177e4SLinus Torvalds /* Not quite right: it can be > snd.fack, but 16001da177e4SLinus Torvalds * it is better to underestimate fackets. 16011da177e4SLinus Torvalds */ 16021da177e4SLinus Torvalds tcp_dec_pcount_approx(&tp->fackets_out, next_skb); 16031da177e4SLinus Torvalds tcp_packets_out_dec(tp, next_skb); 16041da177e4SLinus Torvalds sk_stream_free_skb(sk, next_skb); 16051da177e4SLinus Torvalds } 16061da177e4SLinus Torvalds } 16071da177e4SLinus Torvalds 16081da177e4SLinus Torvalds /* Do a simple retransmit without using the backoff mechanisms in 16091da177e4SLinus Torvalds * tcp_timer. This is used for path mtu discovery. 16101da177e4SLinus Torvalds * The socket is already locked here. 16111da177e4SLinus Torvalds */ 16121da177e4SLinus Torvalds void tcp_simple_retransmit(struct sock *sk) 16131da177e4SLinus Torvalds { 16146687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 16151da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 16161da177e4SLinus Torvalds struct sk_buff *skb; 16171da177e4SLinus Torvalds unsigned int mss = tcp_current_mss(sk, 0); 16181da177e4SLinus Torvalds int lost = 0; 16191da177e4SLinus Torvalds 16201da177e4SLinus Torvalds sk_stream_for_retrans_queue(skb, sk) { 16211da177e4SLinus Torvalds if (skb->len > mss && 16221da177e4SLinus Torvalds !(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) { 16231da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) { 16241da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 16251da177e4SLinus Torvalds tp->retrans_out -= tcp_skb_pcount(skb); 16261da177e4SLinus Torvalds } 16271da177e4SLinus Torvalds if (!(TCP_SKB_CB(skb)->sacked&TCPCB_LOST)) { 16281da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 16291da177e4SLinus Torvalds tp->lost_out += tcp_skb_pcount(skb); 16301da177e4SLinus Torvalds lost = 1; 16311da177e4SLinus Torvalds } 16321da177e4SLinus Torvalds } 16331da177e4SLinus Torvalds } 16341da177e4SLinus Torvalds 16356a438bbeSStephen Hemminger clear_all_retrans_hints(tp); 16366a438bbeSStephen Hemminger 16371da177e4SLinus Torvalds if (!lost) 16381da177e4SLinus Torvalds return; 16391da177e4SLinus Torvalds 16401da177e4SLinus Torvalds tcp_sync_left_out(tp); 16411da177e4SLinus Torvalds 16421da177e4SLinus Torvalds /* Don't muck with the congestion window here. 16431da177e4SLinus Torvalds * Reason is that we do not increase amount of _data_ 16441da177e4SLinus Torvalds * in network, but units changed and effective 16451da177e4SLinus Torvalds * cwnd/ssthresh really reduced now. 16461da177e4SLinus Torvalds */ 16476687e988SArnaldo Carvalho de Melo if (icsk->icsk_ca_state != TCP_CA_Loss) { 16481da177e4SLinus Torvalds tp->high_seq = tp->snd_nxt; 16496687e988SArnaldo Carvalho de Melo tp->snd_ssthresh = tcp_current_ssthresh(sk); 16501da177e4SLinus Torvalds tp->prior_ssthresh = 0; 16511da177e4SLinus Torvalds tp->undo_marker = 0; 16526687e988SArnaldo Carvalho de Melo tcp_set_ca_state(sk, TCP_CA_Loss); 16531da177e4SLinus Torvalds } 16541da177e4SLinus Torvalds tcp_xmit_retransmit_queue(sk); 16551da177e4SLinus Torvalds } 16561da177e4SLinus Torvalds 16571da177e4SLinus Torvalds /* This retransmits one SKB. Policy decisions and retransmit queue 16581da177e4SLinus Torvalds * state updates are done by the caller. Returns non-zero if an 16591da177e4SLinus Torvalds * error occurred which prevented the send. 16601da177e4SLinus Torvalds */ 16611da177e4SLinus Torvalds int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) 16621da177e4SLinus Torvalds { 16631da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 16645d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 16651da177e4SLinus Torvalds unsigned int cur_mss = tcp_current_mss(sk, 0); 16661da177e4SLinus Torvalds int err; 16671da177e4SLinus Torvalds 16685d424d5aSJohn Heffner /* Inconslusive MTU probe */ 16695d424d5aSJohn Heffner if (icsk->icsk_mtup.probe_size) { 16705d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 16715d424d5aSJohn Heffner } 16725d424d5aSJohn Heffner 16731da177e4SLinus Torvalds /* Do not sent more than we queued. 1/4 is reserved for possible 1674caa20d9aSStephen Hemminger * copying overhead: fragmentation, tunneling, mangling etc. 16751da177e4SLinus Torvalds */ 16761da177e4SLinus Torvalds if (atomic_read(&sk->sk_wmem_alloc) > 16771da177e4SLinus Torvalds min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf)) 16781da177e4SLinus Torvalds return -EAGAIN; 16791da177e4SLinus Torvalds 16801da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 16811da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 16821da177e4SLinus Torvalds BUG(); 16831da177e4SLinus Torvalds if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 16841da177e4SLinus Torvalds return -ENOMEM; 16851da177e4SLinus Torvalds } 16861da177e4SLinus Torvalds 16871da177e4SLinus Torvalds /* If receiver has shrunk his window, and skb is out of 16881da177e4SLinus Torvalds * new window, do not retransmit it. The exception is the 16891da177e4SLinus Torvalds * case, when window is shrunk to zero. In this case 16901da177e4SLinus Torvalds * our retransmit serves as a zero window probe. 16911da177e4SLinus Torvalds */ 16921da177e4SLinus Torvalds if (!before(TCP_SKB_CB(skb)->seq, tp->snd_una+tp->snd_wnd) 16931da177e4SLinus Torvalds && TCP_SKB_CB(skb)->seq != tp->snd_una) 16941da177e4SLinus Torvalds return -EAGAIN; 16951da177e4SLinus Torvalds 16961da177e4SLinus Torvalds if (skb->len > cur_mss) { 1697846998aeSDavid S. Miller if (tcp_fragment(sk, skb, cur_mss, cur_mss)) 16981da177e4SLinus Torvalds return -ENOMEM; /* We'll try again later. */ 16991da177e4SLinus Torvalds } 17001da177e4SLinus Torvalds 17011da177e4SLinus Torvalds /* Collapse two adjacent packets if worthwhile and we can. */ 17021da177e4SLinus Torvalds if(!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN) && 17031da177e4SLinus Torvalds (skb->len < (cur_mss >> 1)) && 17041da177e4SLinus Torvalds (skb->next != sk->sk_send_head) && 17051da177e4SLinus Torvalds (skb->next != (struct sk_buff *)&sk->sk_write_queue) && 17061da177e4SLinus Torvalds (skb_shinfo(skb)->nr_frags == 0 && skb_shinfo(skb->next)->nr_frags == 0) && 17071da177e4SLinus Torvalds (tcp_skb_pcount(skb) == 1 && tcp_skb_pcount(skb->next) == 1) && 17081da177e4SLinus Torvalds (sysctl_tcp_retrans_collapse != 0)) 17091da177e4SLinus Torvalds tcp_retrans_try_collapse(sk, skb, cur_mss); 17101da177e4SLinus Torvalds 17118292a17aSArnaldo Carvalho de Melo if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 17121da177e4SLinus Torvalds return -EHOSTUNREACH; /* Routing failure or similar. */ 17131da177e4SLinus Torvalds 17141da177e4SLinus Torvalds /* Some Solaris stacks overoptimize and ignore the FIN on a 17151da177e4SLinus Torvalds * retransmit when old data is attached. So strip it off 17161da177e4SLinus Torvalds * since it is cheap to do so and saves bytes on the network. 17171da177e4SLinus Torvalds */ 17181da177e4SLinus Torvalds if(skb->len > 0 && 17191da177e4SLinus Torvalds (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) && 17201da177e4SLinus Torvalds tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) { 17211da177e4SLinus Torvalds if (!pskb_trim(skb, 0)) { 17221da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq = TCP_SKB_CB(skb)->end_seq - 1; 17231da177e4SLinus Torvalds skb_shinfo(skb)->tso_segs = 1; 17241da177e4SLinus Torvalds skb_shinfo(skb)->tso_size = 0; 17251da177e4SLinus Torvalds skb->ip_summed = CHECKSUM_NONE; 17261da177e4SLinus Torvalds skb->csum = 0; 17271da177e4SLinus Torvalds } 17281da177e4SLinus Torvalds } 17291da177e4SLinus Torvalds 17301da177e4SLinus Torvalds /* Make a copy, if the first transmission SKB clone we made 17311da177e4SLinus Torvalds * is still in somebody's hands, else make a clone. 17321da177e4SLinus Torvalds */ 17331da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 17341da177e4SLinus Torvalds 1735dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 17361da177e4SLinus Torvalds 17371da177e4SLinus Torvalds if (err == 0) { 17381da177e4SLinus Torvalds /* Update global TCP statistics. */ 17391da177e4SLinus Torvalds TCP_INC_STATS(TCP_MIB_RETRANSSEGS); 17401da177e4SLinus Torvalds 17411da177e4SLinus Torvalds tp->total_retrans++; 17421da177e4SLinus Torvalds 17431da177e4SLinus Torvalds #if FASTRETRANS_DEBUG > 0 17441da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) { 17451da177e4SLinus Torvalds if (net_ratelimit()) 17461da177e4SLinus Torvalds printk(KERN_DEBUG "retrans_out leaked.\n"); 17471da177e4SLinus Torvalds } 17481da177e4SLinus Torvalds #endif 17491da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 17501da177e4SLinus Torvalds tp->retrans_out += tcp_skb_pcount(skb); 17511da177e4SLinus Torvalds 17521da177e4SLinus Torvalds /* Save stamp of the first retransmit. */ 17531da177e4SLinus Torvalds if (!tp->retrans_stamp) 17541da177e4SLinus Torvalds tp->retrans_stamp = TCP_SKB_CB(skb)->when; 17551da177e4SLinus Torvalds 17561da177e4SLinus Torvalds tp->undo_retrans++; 17571da177e4SLinus Torvalds 17581da177e4SLinus Torvalds /* snd_nxt is stored to detect loss of retransmitted segment, 17591da177e4SLinus Torvalds * see tcp_input.c tcp_sacktag_write_queue(). 17601da177e4SLinus Torvalds */ 17611da177e4SLinus Torvalds TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt; 17621da177e4SLinus Torvalds } 17631da177e4SLinus Torvalds return err; 17641da177e4SLinus Torvalds } 17651da177e4SLinus Torvalds 17661da177e4SLinus Torvalds /* This gets called after a retransmit timeout, and the initially 17671da177e4SLinus Torvalds * retransmitted data is acknowledged. It tries to continue 17681da177e4SLinus Torvalds * resending the rest of the retransmit queue, until either 17691da177e4SLinus Torvalds * we've sent it all or the congestion window limit is reached. 17701da177e4SLinus Torvalds * If doing SACK, the first ACK which comes back for a timeout 17711da177e4SLinus Torvalds * based retransmit packet might feed us FACK information again. 17721da177e4SLinus Torvalds * If so, we use it to avoid unnecessarily retransmissions. 17731da177e4SLinus Torvalds */ 17741da177e4SLinus Torvalds void tcp_xmit_retransmit_queue(struct sock *sk) 17751da177e4SLinus Torvalds { 17766687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 17771da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 17781da177e4SLinus Torvalds struct sk_buff *skb; 17796a438bbeSStephen Hemminger int packet_cnt; 17806a438bbeSStephen Hemminger 17816a438bbeSStephen Hemminger if (tp->retransmit_skb_hint) { 17826a438bbeSStephen Hemminger skb = tp->retransmit_skb_hint; 17836a438bbeSStephen Hemminger packet_cnt = tp->retransmit_cnt_hint; 17846a438bbeSStephen Hemminger }else{ 17856a438bbeSStephen Hemminger skb = sk->sk_write_queue.next; 17866a438bbeSStephen Hemminger packet_cnt = 0; 17876a438bbeSStephen Hemminger } 17881da177e4SLinus Torvalds 17891da177e4SLinus Torvalds /* First pass: retransmit lost packets. */ 17906a438bbeSStephen Hemminger if (tp->lost_out) { 17916a438bbeSStephen Hemminger sk_stream_for_retrans_queue_from(skb, sk) { 17921da177e4SLinus Torvalds __u8 sacked = TCP_SKB_CB(skb)->sacked; 17931da177e4SLinus Torvalds 17946a438bbeSStephen Hemminger /* we could do better than to assign each time */ 17956a438bbeSStephen Hemminger tp->retransmit_skb_hint = skb; 17966a438bbeSStephen Hemminger tp->retransmit_cnt_hint = packet_cnt; 17976a438bbeSStephen Hemminger 17981da177e4SLinus Torvalds /* Assume this retransmit will generate 17991da177e4SLinus Torvalds * only one packet for congestion window 18001da177e4SLinus Torvalds * calculation purposes. This works because 18011da177e4SLinus Torvalds * tcp_retransmit_skb() will chop up the 18021da177e4SLinus Torvalds * packet to be MSS sized and all the 18031da177e4SLinus Torvalds * packet counting works out. 18041da177e4SLinus Torvalds */ 18051da177e4SLinus Torvalds if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) 18061da177e4SLinus Torvalds return; 18071da177e4SLinus Torvalds 18081da177e4SLinus Torvalds if (sacked & TCPCB_LOST) { 18091da177e4SLinus Torvalds if (!(sacked&(TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))) { 18106a438bbeSStephen Hemminger if (tcp_retransmit_skb(sk, skb)) { 18116a438bbeSStephen Hemminger tp->retransmit_skb_hint = NULL; 18121da177e4SLinus Torvalds return; 18136a438bbeSStephen Hemminger } 18146687e988SArnaldo Carvalho de Melo if (icsk->icsk_ca_state != TCP_CA_Loss) 18151da177e4SLinus Torvalds NET_INC_STATS_BH(LINUX_MIB_TCPFASTRETRANS); 18161da177e4SLinus Torvalds else 18171da177e4SLinus Torvalds NET_INC_STATS_BH(LINUX_MIB_TCPSLOWSTARTRETRANS); 18181da177e4SLinus Torvalds 18191da177e4SLinus Torvalds if (skb == 18201da177e4SLinus Torvalds skb_peek(&sk->sk_write_queue)) 1821463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 18223f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, 18233f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 18241da177e4SLinus Torvalds } 18251da177e4SLinus Torvalds 18266a438bbeSStephen Hemminger packet_cnt += tcp_skb_pcount(skb); 18276a438bbeSStephen Hemminger if (packet_cnt >= tp->lost_out) 18281da177e4SLinus Torvalds break; 18291da177e4SLinus Torvalds } 18301da177e4SLinus Torvalds } 18311da177e4SLinus Torvalds } 18321da177e4SLinus Torvalds 18331da177e4SLinus Torvalds /* OK, demanded retransmission is finished. */ 18341da177e4SLinus Torvalds 18351da177e4SLinus Torvalds /* Forward retransmissions are possible only during Recovery. */ 18366687e988SArnaldo Carvalho de Melo if (icsk->icsk_ca_state != TCP_CA_Recovery) 18371da177e4SLinus Torvalds return; 18381da177e4SLinus Torvalds 18391da177e4SLinus Torvalds /* No forward retransmissions in Reno are possible. */ 18401da177e4SLinus Torvalds if (!tp->rx_opt.sack_ok) 18411da177e4SLinus Torvalds return; 18421da177e4SLinus Torvalds 18431da177e4SLinus Torvalds /* Yeah, we have to make difficult choice between forward transmission 18441da177e4SLinus Torvalds * and retransmission... Both ways have their merits... 18451da177e4SLinus Torvalds * 18461da177e4SLinus Torvalds * For now we do not retransmit anything, while we have some new 18471da177e4SLinus Torvalds * segments to send. 18481da177e4SLinus Torvalds */ 18491da177e4SLinus Torvalds 18501da177e4SLinus Torvalds if (tcp_may_send_now(sk, tp)) 18511da177e4SLinus Torvalds return; 18521da177e4SLinus Torvalds 18536a438bbeSStephen Hemminger if (tp->forward_skb_hint) { 18546a438bbeSStephen Hemminger skb = tp->forward_skb_hint; 18556a438bbeSStephen Hemminger packet_cnt = tp->forward_cnt_hint; 18566a438bbeSStephen Hemminger } else{ 18576a438bbeSStephen Hemminger skb = sk->sk_write_queue.next; 18581da177e4SLinus Torvalds packet_cnt = 0; 18596a438bbeSStephen Hemminger } 18601da177e4SLinus Torvalds 18616a438bbeSStephen Hemminger sk_stream_for_retrans_queue_from(skb, sk) { 18626a438bbeSStephen Hemminger tp->forward_cnt_hint = packet_cnt; 18636a438bbeSStephen Hemminger tp->forward_skb_hint = skb; 18646a438bbeSStephen Hemminger 18651da177e4SLinus Torvalds /* Similar to the retransmit loop above we 18661da177e4SLinus Torvalds * can pretend that the retransmitted SKB 18671da177e4SLinus Torvalds * we send out here will be composed of one 18681da177e4SLinus Torvalds * real MSS sized packet because tcp_retransmit_skb() 18691da177e4SLinus Torvalds * will fragment it if necessary. 18701da177e4SLinus Torvalds */ 18711da177e4SLinus Torvalds if (++packet_cnt > tp->fackets_out) 18721da177e4SLinus Torvalds break; 18731da177e4SLinus Torvalds 18741da177e4SLinus Torvalds if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) 18751da177e4SLinus Torvalds break; 18761da177e4SLinus Torvalds 18771da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) 18781da177e4SLinus Torvalds continue; 18791da177e4SLinus Torvalds 18801da177e4SLinus Torvalds /* Ok, retransmit it. */ 18816a438bbeSStephen Hemminger if (tcp_retransmit_skb(sk, skb)) { 18826a438bbeSStephen Hemminger tp->forward_skb_hint = NULL; 18831da177e4SLinus Torvalds break; 18846a438bbeSStephen Hemminger } 18851da177e4SLinus Torvalds 18861da177e4SLinus Torvalds if (skb == skb_peek(&sk->sk_write_queue)) 18873f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 18883f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, 18893f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 18901da177e4SLinus Torvalds 18911da177e4SLinus Torvalds NET_INC_STATS_BH(LINUX_MIB_TCPFORWARDRETRANS); 18921da177e4SLinus Torvalds } 18931da177e4SLinus Torvalds } 18941da177e4SLinus Torvalds 18951da177e4SLinus Torvalds 18961da177e4SLinus Torvalds /* Send a fin. The caller locks the socket for us. This cannot be 18971da177e4SLinus Torvalds * allowed to fail queueing a FIN frame under any circumstances. 18981da177e4SLinus Torvalds */ 18991da177e4SLinus Torvalds void tcp_send_fin(struct sock *sk) 19001da177e4SLinus Torvalds { 19011da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 19021da177e4SLinus Torvalds struct sk_buff *skb = skb_peek_tail(&sk->sk_write_queue); 19031da177e4SLinus Torvalds int mss_now; 19041da177e4SLinus Torvalds 19051da177e4SLinus Torvalds /* Optimization, tack on the FIN if we have a queue of 19061da177e4SLinus Torvalds * unsent frames. But be careful about outgoing SACKS 19071da177e4SLinus Torvalds * and IP options. 19081da177e4SLinus Torvalds */ 19091da177e4SLinus Torvalds mss_now = tcp_current_mss(sk, 1); 19101da177e4SLinus Torvalds 19111da177e4SLinus Torvalds if (sk->sk_send_head != NULL) { 19121da177e4SLinus Torvalds TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_FIN; 19131da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq++; 19141da177e4SLinus Torvalds tp->write_seq++; 19151da177e4SLinus Torvalds } else { 19161da177e4SLinus Torvalds /* Socket is locked, keep trying until memory is available. */ 19171da177e4SLinus Torvalds for (;;) { 1918d179cd12SDavid S. Miller skb = alloc_skb_fclone(MAX_TCP_HEADER, GFP_KERNEL); 19191da177e4SLinus Torvalds if (skb) 19201da177e4SLinus Torvalds break; 19211da177e4SLinus Torvalds yield(); 19221da177e4SLinus Torvalds } 19231da177e4SLinus Torvalds 19241da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 19251da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 19261da177e4SLinus Torvalds skb->csum = 0; 19271da177e4SLinus Torvalds TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_FIN); 19281da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked = 0; 19291da177e4SLinus Torvalds skb_shinfo(skb)->tso_segs = 1; 19301da177e4SLinus Torvalds skb_shinfo(skb)->tso_size = 0; 19311da177e4SLinus Torvalds 19321da177e4SLinus Torvalds /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 19331da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq = tp->write_seq; 19341da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1; 19351da177e4SLinus Torvalds tcp_queue_skb(sk, skb); 19361da177e4SLinus Torvalds } 19371da177e4SLinus Torvalds __tcp_push_pending_frames(sk, tp, mss_now, TCP_NAGLE_OFF); 19381da177e4SLinus Torvalds } 19391da177e4SLinus Torvalds 19401da177e4SLinus Torvalds /* We get here when a process closes a file descriptor (either due to 19411da177e4SLinus Torvalds * an explicit close() or as a byproduct of exit()'ing) and there 19421da177e4SLinus Torvalds * was unread data in the receive queue. This behavior is recommended 19431da177e4SLinus Torvalds * by draft-ietf-tcpimpl-prob-03.txt section 3.10. -DaveM 19441da177e4SLinus Torvalds */ 1945dd0fc66fSAl Viro void tcp_send_active_reset(struct sock *sk, gfp_t priority) 19461da177e4SLinus Torvalds { 19471da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 19481da177e4SLinus Torvalds struct sk_buff *skb; 19491da177e4SLinus Torvalds 19501da177e4SLinus Torvalds /* NOTE: No TCP options attached and we never retransmit this. */ 19511da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, priority); 19521da177e4SLinus Torvalds if (!skb) { 19531da177e4SLinus Torvalds NET_INC_STATS(LINUX_MIB_TCPABORTFAILED); 19541da177e4SLinus Torvalds return; 19551da177e4SLinus Torvalds } 19561da177e4SLinus Torvalds 19571da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 19581da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 19591da177e4SLinus Torvalds skb->csum = 0; 19601da177e4SLinus Torvalds TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_RST); 19611da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked = 0; 19621da177e4SLinus Torvalds skb_shinfo(skb)->tso_segs = 1; 19631da177e4SLinus Torvalds skb_shinfo(skb)->tso_size = 0; 19641da177e4SLinus Torvalds 19651da177e4SLinus Torvalds /* Send it off. */ 19661da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq = tcp_acceptable_seq(sk, tp); 19671da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq; 19681da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 1969dfb4b9dcSDavid S. Miller if (tcp_transmit_skb(sk, skb, 0, priority)) 19701da177e4SLinus Torvalds NET_INC_STATS(LINUX_MIB_TCPABORTFAILED); 19711da177e4SLinus Torvalds } 19721da177e4SLinus Torvalds 19731da177e4SLinus Torvalds /* WARNING: This routine must only be called when we have already sent 19741da177e4SLinus Torvalds * a SYN packet that crossed the incoming SYN that caused this routine 19751da177e4SLinus Torvalds * to get called. If this assumption fails then the initial rcv_wnd 19761da177e4SLinus Torvalds * and rcv_wscale values will not be correct. 19771da177e4SLinus Torvalds */ 19781da177e4SLinus Torvalds int tcp_send_synack(struct sock *sk) 19791da177e4SLinus Torvalds { 19801da177e4SLinus Torvalds struct sk_buff* skb; 19811da177e4SLinus Torvalds 19821da177e4SLinus Torvalds skb = skb_peek(&sk->sk_write_queue); 19831da177e4SLinus Torvalds if (skb == NULL || !(TCP_SKB_CB(skb)->flags&TCPCB_FLAG_SYN)) { 19841da177e4SLinus Torvalds printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n"); 19851da177e4SLinus Torvalds return -EFAULT; 19861da177e4SLinus Torvalds } 19871da177e4SLinus Torvalds if (!(TCP_SKB_CB(skb)->flags&TCPCB_FLAG_ACK)) { 19881da177e4SLinus Torvalds if (skb_cloned(skb)) { 19891da177e4SLinus Torvalds struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC); 19901da177e4SLinus Torvalds if (nskb == NULL) 19911da177e4SLinus Torvalds return -ENOMEM; 19921da177e4SLinus Torvalds __skb_unlink(skb, &sk->sk_write_queue); 19931da177e4SLinus Torvalds skb_header_release(nskb); 19941da177e4SLinus Torvalds __skb_queue_head(&sk->sk_write_queue, nskb); 19951da177e4SLinus Torvalds sk_stream_free_skb(sk, skb); 19961da177e4SLinus Torvalds sk_charge_skb(sk, nskb); 19971da177e4SLinus Torvalds skb = nskb; 19981da177e4SLinus Torvalds } 19991da177e4SLinus Torvalds 20001da177e4SLinus Torvalds TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ACK; 20011da177e4SLinus Torvalds TCP_ECN_send_synack(tcp_sk(sk), skb); 20021da177e4SLinus Torvalds } 20031da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2004dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 20051da177e4SLinus Torvalds } 20061da177e4SLinus Torvalds 20071da177e4SLinus Torvalds /* 20081da177e4SLinus Torvalds * Prepare a SYN-ACK. 20091da177e4SLinus Torvalds */ 20101da177e4SLinus Torvalds struct sk_buff * tcp_make_synack(struct sock *sk, struct dst_entry *dst, 201160236fddSArnaldo Carvalho de Melo struct request_sock *req) 20121da177e4SLinus Torvalds { 20132e6599cbSArnaldo Carvalho de Melo struct inet_request_sock *ireq = inet_rsk(req); 20141da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 20151da177e4SLinus Torvalds struct tcphdr *th; 20161da177e4SLinus Torvalds int tcp_header_size; 20171da177e4SLinus Torvalds struct sk_buff *skb; 20181da177e4SLinus Torvalds 20191da177e4SLinus Torvalds skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC); 20201da177e4SLinus Torvalds if (skb == NULL) 20211da177e4SLinus Torvalds return NULL; 20221da177e4SLinus Torvalds 20231da177e4SLinus Torvalds /* Reserve space for headers. */ 20241da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 20251da177e4SLinus Torvalds 20261da177e4SLinus Torvalds skb->dst = dst_clone(dst); 20271da177e4SLinus Torvalds 20281da177e4SLinus Torvalds tcp_header_size = (sizeof(struct tcphdr) + TCPOLEN_MSS + 20292e6599cbSArnaldo Carvalho de Melo (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0) + 20302e6599cbSArnaldo Carvalho de Melo (ireq->wscale_ok ? TCPOLEN_WSCALE_ALIGNED : 0) + 20311da177e4SLinus Torvalds /* SACK_PERM is in the place of NOP NOP of TS */ 20322e6599cbSArnaldo Carvalho de Melo ((ireq->sack_ok && !ireq->tstamp_ok) ? TCPOLEN_SACKPERM_ALIGNED : 0)); 20331da177e4SLinus Torvalds skb->h.th = th = (struct tcphdr *) skb_push(skb, tcp_header_size); 20341da177e4SLinus Torvalds 20351da177e4SLinus Torvalds memset(th, 0, sizeof(struct tcphdr)); 20361da177e4SLinus Torvalds th->syn = 1; 20371da177e4SLinus Torvalds th->ack = 1; 20381da177e4SLinus Torvalds if (dst->dev->features&NETIF_F_TSO) 20392e6599cbSArnaldo Carvalho de Melo ireq->ecn_ok = 0; 20401da177e4SLinus Torvalds TCP_ECN_make_synack(req, th); 20411da177e4SLinus Torvalds th->source = inet_sk(sk)->sport; 20422e6599cbSArnaldo Carvalho de Melo th->dest = ireq->rmt_port; 20432e6599cbSArnaldo Carvalho de Melo TCP_SKB_CB(skb)->seq = tcp_rsk(req)->snt_isn; 20441da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1; 20451da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked = 0; 20461da177e4SLinus Torvalds skb_shinfo(skb)->tso_segs = 1; 20471da177e4SLinus Torvalds skb_shinfo(skb)->tso_size = 0; 20481da177e4SLinus Torvalds th->seq = htonl(TCP_SKB_CB(skb)->seq); 20492e6599cbSArnaldo Carvalho de Melo th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1); 20501da177e4SLinus Torvalds if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */ 20511da177e4SLinus Torvalds __u8 rcv_wscale; 20521da177e4SLinus Torvalds /* Set this up on the first call only */ 20531da177e4SLinus Torvalds req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW); 20541da177e4SLinus Torvalds /* tcp_full_space because it is guaranteed to be the first packet */ 20551da177e4SLinus Torvalds tcp_select_initial_window(tcp_full_space(sk), 20562e6599cbSArnaldo Carvalho de Melo dst_metric(dst, RTAX_ADVMSS) - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0), 20571da177e4SLinus Torvalds &req->rcv_wnd, 20581da177e4SLinus Torvalds &req->window_clamp, 20592e6599cbSArnaldo Carvalho de Melo ireq->wscale_ok, 20601da177e4SLinus Torvalds &rcv_wscale); 20612e6599cbSArnaldo Carvalho de Melo ireq->rcv_wscale = rcv_wscale; 20621da177e4SLinus Torvalds } 20631da177e4SLinus Torvalds 20641da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 20651da177e4SLinus Torvalds th->window = htons(req->rcv_wnd); 20661da177e4SLinus Torvalds 20671da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 20682e6599cbSArnaldo Carvalho de Melo tcp_syn_build_options((__u32 *)(th + 1), dst_metric(dst, RTAX_ADVMSS), ireq->tstamp_ok, 20692e6599cbSArnaldo Carvalho de Melo ireq->sack_ok, ireq->wscale_ok, ireq->rcv_wscale, 20701da177e4SLinus Torvalds TCP_SKB_CB(skb)->when, 20711da177e4SLinus Torvalds req->ts_recent); 20721da177e4SLinus Torvalds 20731da177e4SLinus Torvalds skb->csum = 0; 20741da177e4SLinus Torvalds th->doff = (tcp_header_size >> 2); 20751da177e4SLinus Torvalds TCP_INC_STATS(TCP_MIB_OUTSEGS); 20761da177e4SLinus Torvalds return skb; 20771da177e4SLinus Torvalds } 20781da177e4SLinus Torvalds 20791da177e4SLinus Torvalds /* 20801da177e4SLinus Torvalds * Do all connect socket setups that can be done AF independent. 20811da177e4SLinus Torvalds */ 208240efc6faSStephen Hemminger static void tcp_connect_init(struct sock *sk) 20831da177e4SLinus Torvalds { 20841da177e4SLinus Torvalds struct dst_entry *dst = __sk_dst_get(sk); 20851da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 20861da177e4SLinus Torvalds __u8 rcv_wscale; 20871da177e4SLinus Torvalds 20881da177e4SLinus Torvalds /* We'll fix this up when we get a response from the other end. 20891da177e4SLinus Torvalds * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 20901da177e4SLinus Torvalds */ 20911da177e4SLinus Torvalds tp->tcp_header_len = sizeof(struct tcphdr) + 20921da177e4SLinus Torvalds (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0); 20931da177e4SLinus Torvalds 20941da177e4SLinus Torvalds /* If user gave his TCP_MAXSEG, record it to clamp */ 20951da177e4SLinus Torvalds if (tp->rx_opt.user_mss) 20961da177e4SLinus Torvalds tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 20971da177e4SLinus Torvalds tp->max_window = 0; 20985d424d5aSJohn Heffner tcp_mtup_init(sk); 20991da177e4SLinus Torvalds tcp_sync_mss(sk, dst_mtu(dst)); 21001da177e4SLinus Torvalds 21011da177e4SLinus Torvalds if (!tp->window_clamp) 21021da177e4SLinus Torvalds tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 21031da177e4SLinus Torvalds tp->advmss = dst_metric(dst, RTAX_ADVMSS); 21041da177e4SLinus Torvalds tcp_initialize_rcv_mss(sk); 21051da177e4SLinus Torvalds 21061da177e4SLinus Torvalds tcp_select_initial_window(tcp_full_space(sk), 21071da177e4SLinus Torvalds tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 21081da177e4SLinus Torvalds &tp->rcv_wnd, 21091da177e4SLinus Torvalds &tp->window_clamp, 21101da177e4SLinus Torvalds sysctl_tcp_window_scaling, 21111da177e4SLinus Torvalds &rcv_wscale); 21121da177e4SLinus Torvalds 21131da177e4SLinus Torvalds tp->rx_opt.rcv_wscale = rcv_wscale; 21141da177e4SLinus Torvalds tp->rcv_ssthresh = tp->rcv_wnd; 21151da177e4SLinus Torvalds 21161da177e4SLinus Torvalds sk->sk_err = 0; 21171da177e4SLinus Torvalds sock_reset_flag(sk, SOCK_DONE); 21181da177e4SLinus Torvalds tp->snd_wnd = 0; 21191da177e4SLinus Torvalds tcp_init_wl(tp, tp->write_seq, 0); 21201da177e4SLinus Torvalds tp->snd_una = tp->write_seq; 21211da177e4SLinus Torvalds tp->snd_sml = tp->write_seq; 21221da177e4SLinus Torvalds tp->rcv_nxt = 0; 21231da177e4SLinus Torvalds tp->rcv_wup = 0; 21241da177e4SLinus Torvalds tp->copied_seq = 0; 21251da177e4SLinus Torvalds 2126463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; 2127463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_retransmits = 0; 21281da177e4SLinus Torvalds tcp_clear_retrans(tp); 21291da177e4SLinus Torvalds } 21301da177e4SLinus Torvalds 21311da177e4SLinus Torvalds /* 21321da177e4SLinus Torvalds * Build a SYN and send it off. 21331da177e4SLinus Torvalds */ 21341da177e4SLinus Torvalds int tcp_connect(struct sock *sk) 21351da177e4SLinus Torvalds { 21361da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 21371da177e4SLinus Torvalds struct sk_buff *buff; 21381da177e4SLinus Torvalds 21391da177e4SLinus Torvalds tcp_connect_init(sk); 21401da177e4SLinus Torvalds 2141d179cd12SDavid S. Miller buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation); 21421da177e4SLinus Torvalds if (unlikely(buff == NULL)) 21431da177e4SLinus Torvalds return -ENOBUFS; 21441da177e4SLinus Torvalds 21451da177e4SLinus Torvalds /* Reserve space for headers. */ 21461da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 21471da177e4SLinus Torvalds 21481da177e4SLinus Torvalds TCP_SKB_CB(buff)->flags = TCPCB_FLAG_SYN; 21491da177e4SLinus Torvalds TCP_ECN_send_syn(sk, tp, buff); 21501da177e4SLinus Torvalds TCP_SKB_CB(buff)->sacked = 0; 21511da177e4SLinus Torvalds skb_shinfo(buff)->tso_segs = 1; 21521da177e4SLinus Torvalds skb_shinfo(buff)->tso_size = 0; 21531da177e4SLinus Torvalds buff->csum = 0; 21541da177e4SLinus Torvalds TCP_SKB_CB(buff)->seq = tp->write_seq++; 21551da177e4SLinus Torvalds TCP_SKB_CB(buff)->end_seq = tp->write_seq; 21561da177e4SLinus Torvalds tp->snd_nxt = tp->write_seq; 21571da177e4SLinus Torvalds tp->pushed_seq = tp->write_seq; 21581da177e4SLinus Torvalds 21591da177e4SLinus Torvalds /* Send it off. */ 21601da177e4SLinus Torvalds TCP_SKB_CB(buff)->when = tcp_time_stamp; 21611da177e4SLinus Torvalds tp->retrans_stamp = TCP_SKB_CB(buff)->when; 21621da177e4SLinus Torvalds skb_header_release(buff); 21631da177e4SLinus Torvalds __skb_queue_tail(&sk->sk_write_queue, buff); 21641da177e4SLinus Torvalds sk_charge_skb(sk, buff); 21651da177e4SLinus Torvalds tp->packets_out += tcp_skb_pcount(buff); 2166dfb4b9dcSDavid S. Miller tcp_transmit_skb(sk, buff, 1, GFP_KERNEL); 21671da177e4SLinus Torvalds TCP_INC_STATS(TCP_MIB_ACTIVEOPENS); 21681da177e4SLinus Torvalds 21691da177e4SLinus Torvalds /* Timer for repeating the SYN until an answer. */ 21703f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 21713f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 21721da177e4SLinus Torvalds return 0; 21731da177e4SLinus Torvalds } 21741da177e4SLinus Torvalds 21751da177e4SLinus Torvalds /* Send out a delayed ack, the caller does the policy checking 21761da177e4SLinus Torvalds * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 21771da177e4SLinus Torvalds * for details. 21781da177e4SLinus Torvalds */ 21791da177e4SLinus Torvalds void tcp_send_delayed_ack(struct sock *sk) 21801da177e4SLinus Torvalds { 2181463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 2182463c84b9SArnaldo Carvalho de Melo int ato = icsk->icsk_ack.ato; 21831da177e4SLinus Torvalds unsigned long timeout; 21841da177e4SLinus Torvalds 21851da177e4SLinus Torvalds if (ato > TCP_DELACK_MIN) { 2186463c84b9SArnaldo Carvalho de Melo const struct tcp_sock *tp = tcp_sk(sk); 21871da177e4SLinus Torvalds int max_ato = HZ/2; 21881da177e4SLinus Torvalds 2189463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.pingpong || (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 21901da177e4SLinus Torvalds max_ato = TCP_DELACK_MAX; 21911da177e4SLinus Torvalds 21921da177e4SLinus Torvalds /* Slow path, intersegment interval is "high". */ 21931da177e4SLinus Torvalds 21941da177e4SLinus Torvalds /* If some rtt estimate is known, use it to bound delayed ack. 2195463c84b9SArnaldo Carvalho de Melo * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 21961da177e4SLinus Torvalds * directly. 21971da177e4SLinus Torvalds */ 21981da177e4SLinus Torvalds if (tp->srtt) { 21991da177e4SLinus Torvalds int rtt = max(tp->srtt>>3, TCP_DELACK_MIN); 22001da177e4SLinus Torvalds 22011da177e4SLinus Torvalds if (rtt < max_ato) 22021da177e4SLinus Torvalds max_ato = rtt; 22031da177e4SLinus Torvalds } 22041da177e4SLinus Torvalds 22051da177e4SLinus Torvalds ato = min(ato, max_ato); 22061da177e4SLinus Torvalds } 22071da177e4SLinus Torvalds 22081da177e4SLinus Torvalds /* Stay within the limit we were given */ 22091da177e4SLinus Torvalds timeout = jiffies + ato; 22101da177e4SLinus Torvalds 22111da177e4SLinus Torvalds /* Use new timeout only if there wasn't a older one earlier. */ 2212463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 22131da177e4SLinus Torvalds /* If delack timer was blocked or is about to expire, 22141da177e4SLinus Torvalds * send ACK now. 22151da177e4SLinus Torvalds */ 2216463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.blocked || 2217463c84b9SArnaldo Carvalho de Melo time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 22181da177e4SLinus Torvalds tcp_send_ack(sk); 22191da177e4SLinus Torvalds return; 22201da177e4SLinus Torvalds } 22211da177e4SLinus Torvalds 2222463c84b9SArnaldo Carvalho de Melo if (!time_before(timeout, icsk->icsk_ack.timeout)) 2223463c84b9SArnaldo Carvalho de Melo timeout = icsk->icsk_ack.timeout; 22241da177e4SLinus Torvalds } 2225463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 2226463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.timeout = timeout; 2227463c84b9SArnaldo Carvalho de Melo sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 22281da177e4SLinus Torvalds } 22291da177e4SLinus Torvalds 22301da177e4SLinus Torvalds /* This routine sends an ack and also updates the window. */ 22311da177e4SLinus Torvalds void tcp_send_ack(struct sock *sk) 22321da177e4SLinus Torvalds { 22331da177e4SLinus Torvalds /* If we have been reset, we may not send again. */ 22341da177e4SLinus Torvalds if (sk->sk_state != TCP_CLOSE) { 22351da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 22361da177e4SLinus Torvalds struct sk_buff *buff; 22371da177e4SLinus Torvalds 22381da177e4SLinus Torvalds /* We are not putting this on the write queue, so 22391da177e4SLinus Torvalds * tcp_transmit_skb() will set the ownership to this 22401da177e4SLinus Torvalds * sock. 22411da177e4SLinus Torvalds */ 22421da177e4SLinus Torvalds buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 22431da177e4SLinus Torvalds if (buff == NULL) { 2244463c84b9SArnaldo Carvalho de Melo inet_csk_schedule_ack(sk); 2245463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; 22463f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 22473f421baaSArnaldo Carvalho de Melo TCP_DELACK_MAX, TCP_RTO_MAX); 22481da177e4SLinus Torvalds return; 22491da177e4SLinus Torvalds } 22501da177e4SLinus Torvalds 22511da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 22521da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 22531da177e4SLinus Torvalds buff->csum = 0; 22541da177e4SLinus Torvalds TCP_SKB_CB(buff)->flags = TCPCB_FLAG_ACK; 22551da177e4SLinus Torvalds TCP_SKB_CB(buff)->sacked = 0; 22561da177e4SLinus Torvalds skb_shinfo(buff)->tso_segs = 1; 22571da177e4SLinus Torvalds skb_shinfo(buff)->tso_size = 0; 22581da177e4SLinus Torvalds 22591da177e4SLinus Torvalds /* Send it off, this clears delayed acks for us. */ 22601da177e4SLinus Torvalds TCP_SKB_CB(buff)->seq = TCP_SKB_CB(buff)->end_seq = tcp_acceptable_seq(sk, tp); 22611da177e4SLinus Torvalds TCP_SKB_CB(buff)->when = tcp_time_stamp; 2262dfb4b9dcSDavid S. Miller tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC); 22631da177e4SLinus Torvalds } 22641da177e4SLinus Torvalds } 22651da177e4SLinus Torvalds 22661da177e4SLinus Torvalds /* This routine sends a packet with an out of date sequence 22671da177e4SLinus Torvalds * number. It assumes the other end will try to ack it. 22681da177e4SLinus Torvalds * 22691da177e4SLinus Torvalds * Question: what should we make while urgent mode? 22701da177e4SLinus Torvalds * 4.4BSD forces sending single byte of data. We cannot send 22711da177e4SLinus Torvalds * out of window data, because we have SND.NXT==SND.MAX... 22721da177e4SLinus Torvalds * 22731da177e4SLinus Torvalds * Current solution: to send TWO zero-length segments in urgent mode: 22741da177e4SLinus Torvalds * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 22751da177e4SLinus Torvalds * out-of-date with SND.UNA-1 to probe window. 22761da177e4SLinus Torvalds */ 22771da177e4SLinus Torvalds static int tcp_xmit_probe_skb(struct sock *sk, int urgent) 22781da177e4SLinus Torvalds { 22791da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 22801da177e4SLinus Torvalds struct sk_buff *skb; 22811da177e4SLinus Torvalds 22821da177e4SLinus Torvalds /* We don't queue it, tcp_transmit_skb() sets ownership. */ 22831da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 22841da177e4SLinus Torvalds if (skb == NULL) 22851da177e4SLinus Torvalds return -1; 22861da177e4SLinus Torvalds 22871da177e4SLinus Torvalds /* Reserve space for headers and set control bits. */ 22881da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 22891da177e4SLinus Torvalds skb->csum = 0; 22901da177e4SLinus Torvalds TCP_SKB_CB(skb)->flags = TCPCB_FLAG_ACK; 22911da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked = urgent; 22921da177e4SLinus Torvalds skb_shinfo(skb)->tso_segs = 1; 22931da177e4SLinus Torvalds skb_shinfo(skb)->tso_size = 0; 22941da177e4SLinus Torvalds 22951da177e4SLinus Torvalds /* Use a previous sequence. This should cause the other 22961da177e4SLinus Torvalds * end to send an ack. Don't queue or clone SKB, just 22971da177e4SLinus Torvalds * send it. 22981da177e4SLinus Torvalds */ 22991da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq = urgent ? tp->snd_una : tp->snd_una - 1; 23001da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq; 23011da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2302dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC); 23031da177e4SLinus Torvalds } 23041da177e4SLinus Torvalds 23051da177e4SLinus Torvalds int tcp_write_wakeup(struct sock *sk) 23061da177e4SLinus Torvalds { 23071da177e4SLinus Torvalds if (sk->sk_state != TCP_CLOSE) { 23081da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 23091da177e4SLinus Torvalds struct sk_buff *skb; 23101da177e4SLinus Torvalds 23111da177e4SLinus Torvalds if ((skb = sk->sk_send_head) != NULL && 23121da177e4SLinus Torvalds before(TCP_SKB_CB(skb)->seq, tp->snd_una+tp->snd_wnd)) { 23131da177e4SLinus Torvalds int err; 23141da177e4SLinus Torvalds unsigned int mss = tcp_current_mss(sk, 0); 23151da177e4SLinus Torvalds unsigned int seg_size = tp->snd_una+tp->snd_wnd-TCP_SKB_CB(skb)->seq; 23161da177e4SLinus Torvalds 23171da177e4SLinus Torvalds if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 23181da177e4SLinus Torvalds tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 23191da177e4SLinus Torvalds 23201da177e4SLinus Torvalds /* We are probing the opening of a window 23211da177e4SLinus Torvalds * but the window size is != 0 23221da177e4SLinus Torvalds * must have been a result SWS avoidance ( sender ) 23231da177e4SLinus Torvalds */ 23241da177e4SLinus Torvalds if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 23251da177e4SLinus Torvalds skb->len > mss) { 23261da177e4SLinus Torvalds seg_size = min(seg_size, mss); 23271da177e4SLinus Torvalds TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH; 2328846998aeSDavid S. Miller if (tcp_fragment(sk, skb, seg_size, mss)) 23291da177e4SLinus Torvalds return -1; 23301da177e4SLinus Torvalds } else if (!tcp_skb_pcount(skb)) 2331846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss); 23321da177e4SLinus Torvalds 23331da177e4SLinus Torvalds TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH; 23341da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2335dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 23361da177e4SLinus Torvalds if (!err) { 23371da177e4SLinus Torvalds update_send_head(sk, tp, skb); 23381da177e4SLinus Torvalds } 23391da177e4SLinus Torvalds return err; 23401da177e4SLinus Torvalds } else { 23411da177e4SLinus Torvalds if (tp->urg_mode && 23421da177e4SLinus Torvalds between(tp->snd_up, tp->snd_una+1, tp->snd_una+0xFFFF)) 23431da177e4SLinus Torvalds tcp_xmit_probe_skb(sk, TCPCB_URG); 23441da177e4SLinus Torvalds return tcp_xmit_probe_skb(sk, 0); 23451da177e4SLinus Torvalds } 23461da177e4SLinus Torvalds } 23471da177e4SLinus Torvalds return -1; 23481da177e4SLinus Torvalds } 23491da177e4SLinus Torvalds 23501da177e4SLinus Torvalds /* A window probe timeout has occurred. If window is not closed send 23511da177e4SLinus Torvalds * a partial packet else a zero probe. 23521da177e4SLinus Torvalds */ 23531da177e4SLinus Torvalds void tcp_send_probe0(struct sock *sk) 23541da177e4SLinus Torvalds { 2355463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 23561da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 23571da177e4SLinus Torvalds int err; 23581da177e4SLinus Torvalds 23591da177e4SLinus Torvalds err = tcp_write_wakeup(sk); 23601da177e4SLinus Torvalds 23611da177e4SLinus Torvalds if (tp->packets_out || !sk->sk_send_head) { 23621da177e4SLinus Torvalds /* Cancel probe timer, if it is not required. */ 23636687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 0; 2364463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff = 0; 23651da177e4SLinus Torvalds return; 23661da177e4SLinus Torvalds } 23671da177e4SLinus Torvalds 23681da177e4SLinus Torvalds if (err <= 0) { 2369463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_backoff < sysctl_tcp_retries2) 2370463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff++; 23716687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out++; 2372463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 23733f421baaSArnaldo Carvalho de Melo min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), 23743f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 23751da177e4SLinus Torvalds } else { 23761da177e4SLinus Torvalds /* If packet was not sent due to local congestion, 23776687e988SArnaldo Carvalho de Melo * do not backoff and do not remember icsk_probes_out. 23781da177e4SLinus Torvalds * Let local senders to fight for local resources. 23791da177e4SLinus Torvalds * 23801da177e4SLinus Torvalds * Use accumulated backoff yet. 23811da177e4SLinus Torvalds */ 23826687e988SArnaldo Carvalho de Melo if (!icsk->icsk_probes_out) 23836687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 1; 2384463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 2385463c84b9SArnaldo Carvalho de Melo min(icsk->icsk_rto << icsk->icsk_backoff, 23863f421baaSArnaldo Carvalho de Melo TCP_RESOURCE_PROBE_INTERVAL), 23873f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 23881da177e4SLinus Torvalds } 23891da177e4SLinus Torvalds } 23901da177e4SLinus Torvalds 23911da177e4SLinus Torvalds EXPORT_SYMBOL(tcp_connect); 23921da177e4SLinus Torvalds EXPORT_SYMBOL(tcp_make_synack); 23931da177e4SLinus Torvalds EXPORT_SYMBOL(tcp_simple_retransmit); 23941da177e4SLinus Torvalds EXPORT_SYMBOL(tcp_sync_mss); 2395f4805edeSStephen Hemminger EXPORT_SYMBOL(sysctl_tcp_tso_win_divisor); 23965d424d5aSJohn Heffner EXPORT_SYMBOL(tcp_mtup_init); 2397