11da177e4SLinus Torvalds /* 21da177e4SLinus Torvalds * INET An implementation of the TCP/IP protocol suite for the LINUX 31da177e4SLinus Torvalds * operating system. INET is implemented using the BSD Socket 41da177e4SLinus Torvalds * interface as the means of communication with the user level. 51da177e4SLinus Torvalds * 61da177e4SLinus Torvalds * Implementation of the Transmission Control Protocol(TCP). 71da177e4SLinus Torvalds * 802c30a84SJesper Juhl * Authors: Ross Biro 91da177e4SLinus Torvalds * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> 101da177e4SLinus Torvalds * Mark Evans, <evansmp@uhura.aston.ac.uk> 111da177e4SLinus Torvalds * Corey Minyard <wf-rch!minyard@relay.EU.net> 121da177e4SLinus Torvalds * Florian La Roche, <flla@stud.uni-sb.de> 131da177e4SLinus Torvalds * Charles Hedrick, <hedrick@klinzhai.rutgers.edu> 141da177e4SLinus Torvalds * Linus Torvalds, <torvalds@cs.helsinki.fi> 151da177e4SLinus Torvalds * Alan Cox, <gw4pts@gw4pts.ampr.org> 161da177e4SLinus Torvalds * Matthew Dillon, <dillon@apollo.west.oic.com> 171da177e4SLinus Torvalds * Arnt Gulbrandsen, <agulbra@nvg.unit.no> 181da177e4SLinus Torvalds * Jorge Cwik, <jorge@laser.satlink.net> 191da177e4SLinus Torvalds */ 201da177e4SLinus Torvalds 211da177e4SLinus Torvalds /* 221da177e4SLinus Torvalds * Changes: Pedro Roque : Retransmit queue handled by TCP. 231da177e4SLinus Torvalds * : Fragmentation on mtu decrease 241da177e4SLinus Torvalds * : Segment collapse on retransmit 251da177e4SLinus Torvalds * : AF independence 261da177e4SLinus Torvalds * 271da177e4SLinus Torvalds * Linus Torvalds : send_delayed_ack 281da177e4SLinus Torvalds * David S. Miller : Charge memory using the right skb 291da177e4SLinus Torvalds * during syn/ack processing. 301da177e4SLinus Torvalds * David S. Miller : Output engine completely rewritten. 311da177e4SLinus Torvalds * Andrea Arcangeli: SYNACK carry ts_recent in tsecr. 321da177e4SLinus Torvalds * Cacophonix Gaul : draft-minshall-nagle-01 331da177e4SLinus Torvalds * J Hadi Salim : ECN support 341da177e4SLinus Torvalds * 351da177e4SLinus Torvalds */ 361da177e4SLinus Torvalds 3791df42beSJoe Perches #define pr_fmt(fmt) "TCP: " fmt 3891df42beSJoe Perches 391da177e4SLinus Torvalds #include <net/tcp.h> 401da177e4SLinus Torvalds 411da177e4SLinus Torvalds #include <linux/compiler.h> 425a0e3ad6STejun Heo #include <linux/gfp.h> 431da177e4SLinus Torvalds #include <linux/module.h> 441da177e4SLinus Torvalds 451da177e4SLinus Torvalds /* People can turn this off for buggy TCP's found in printers etc. */ 46ab32ea5dSBrian Haley int sysctl_tcp_retrans_collapse __read_mostly = 1; 471da177e4SLinus Torvalds 4815d99e02SRick Jones /* People can turn this on to work with those rare, broken TCPs that 4915d99e02SRick Jones * interpret the window field as a signed quantity. 5015d99e02SRick Jones */ 51ab32ea5dSBrian Haley int sysctl_tcp_workaround_signed_windows __read_mostly = 0; 5215d99e02SRick Jones 5346d3ceabSEric Dumazet /* Default TSQ limit of two TSO segments */ 5446d3ceabSEric Dumazet int sysctl_tcp_limit_output_bytes __read_mostly = 131072; 5546d3ceabSEric Dumazet 561da177e4SLinus Torvalds /* This limits the percentage of the congestion window which we 571da177e4SLinus Torvalds * will allow a single TSO frame to consume. Building TSO frames 581da177e4SLinus Torvalds * which are too large can cause TCP streams to be bursty. 591da177e4SLinus Torvalds */ 60ab32ea5dSBrian Haley int sysctl_tcp_tso_win_divisor __read_mostly = 3; 611da177e4SLinus Torvalds 62ab32ea5dSBrian Haley int sysctl_tcp_mtu_probing __read_mostly = 0; 6397b1ce25SShan Wei int sysctl_tcp_base_mss __read_mostly = TCP_BASE_MSS; 645d424d5aSJohn Heffner 6535089bb2SDavid S. Miller /* By default, RFC2861 behavior. */ 66ab32ea5dSBrian Haley int sysctl_tcp_slow_start_after_idle __read_mostly = 1; 6735089bb2SDavid S. Miller 68519855c5SWilliam Allen Simpson int sysctl_tcp_cookie_size __read_mostly = 0; /* TCP_COOKIE_MAX */ 69e6b09ccaSDavid S. Miller EXPORT_SYMBOL_GPL(sysctl_tcp_cookie_size); 70519855c5SWilliam Allen Simpson 7146d3ceabSEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 7246d3ceabSEric Dumazet int push_one, gfp_t gfp); 73519855c5SWilliam Allen Simpson 7467edfef7SAndi Kleen /* Account for new data that has been sent to the network. */ 75cf533ea5SEric Dumazet static void tcp_event_new_data_sent(struct sock *sk, const struct sk_buff *skb) 766ff03ac3SIlpo Järvinen { 776ff03ac3SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 7866f5fe62SIlpo Järvinen unsigned int prior_packets = tp->packets_out; 799e412ba7SIlpo Järvinen 80fe067e8aSDavid S. Miller tcp_advance_send_head(sk, skb); 811da177e4SLinus Torvalds tp->snd_nxt = TCP_SKB_CB(skb)->end_seq; 828512430eSIlpo Järvinen 8325985edcSLucas De Marchi /* Don't override Nagle indefinitely with F-RTO */ 848512430eSIlpo Järvinen if (tp->frto_counter == 2) 858512430eSIlpo Järvinen tp->frto_counter = 3; 8666f5fe62SIlpo Järvinen 8766f5fe62SIlpo Järvinen tp->packets_out += tcp_skb_pcount(skb); 88750ea2baSYuchung Cheng if (!prior_packets || tp->early_retrans_delayed) 89750ea2baSYuchung Cheng tcp_rearm_rto(sk); 901da177e4SLinus Torvalds } 911da177e4SLinus Torvalds 921da177e4SLinus Torvalds /* SND.NXT, if window was not shrunk. 931da177e4SLinus Torvalds * If window has been shrunk, what should we make? It is not clear at all. 941da177e4SLinus Torvalds * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-( 951da177e4SLinus Torvalds * Anything in between SND.UNA...SND.UNA+SND.WND also can be already 961da177e4SLinus Torvalds * invalid. OK, let's make this for now: 971da177e4SLinus Torvalds */ 98cf533ea5SEric Dumazet static inline __u32 tcp_acceptable_seq(const struct sock *sk) 991da177e4SLinus Torvalds { 100cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1019e412ba7SIlpo Järvinen 10290840defSIlpo Järvinen if (!before(tcp_wnd_end(tp), tp->snd_nxt)) 1031da177e4SLinus Torvalds return tp->snd_nxt; 1041da177e4SLinus Torvalds else 10590840defSIlpo Järvinen return tcp_wnd_end(tp); 1061da177e4SLinus Torvalds } 1071da177e4SLinus Torvalds 1081da177e4SLinus Torvalds /* Calculate mss to advertise in SYN segment. 1091da177e4SLinus Torvalds * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that: 1101da177e4SLinus Torvalds * 1111da177e4SLinus Torvalds * 1. It is independent of path mtu. 1121da177e4SLinus Torvalds * 2. Ideally, it is maximal possible segment size i.e. 65535-40. 1131da177e4SLinus Torvalds * 3. For IPv4 it is reasonable to calculate it from maximal MTU of 1141da177e4SLinus Torvalds * attached devices, because some buggy hosts are confused by 1151da177e4SLinus Torvalds * large MSS. 1161da177e4SLinus Torvalds * 4. We do not make 3, we advertise MSS, calculated from first 1171da177e4SLinus Torvalds * hop device mtu, but allow to raise it to ip_rt_min_advmss. 1181da177e4SLinus Torvalds * This may be overridden via information stored in routing table. 1191da177e4SLinus Torvalds * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible, 1201da177e4SLinus Torvalds * probably even Jumbo". 1211da177e4SLinus Torvalds */ 1221da177e4SLinus Torvalds static __u16 tcp_advertise_mss(struct sock *sk) 1231da177e4SLinus Torvalds { 1241da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 125cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1261da177e4SLinus Torvalds int mss = tp->advmss; 1271da177e4SLinus Torvalds 1280dbaee3bSDavid S. Miller if (dst) { 1290dbaee3bSDavid S. Miller unsigned int metric = dst_metric_advmss(dst); 1300dbaee3bSDavid S. Miller 1310dbaee3bSDavid S. Miller if (metric < mss) { 1320dbaee3bSDavid S. Miller mss = metric; 1331da177e4SLinus Torvalds tp->advmss = mss; 1341da177e4SLinus Torvalds } 1350dbaee3bSDavid S. Miller } 1361da177e4SLinus Torvalds 1371da177e4SLinus Torvalds return (__u16)mss; 1381da177e4SLinus Torvalds } 1391da177e4SLinus Torvalds 1401da177e4SLinus Torvalds /* RFC2861. Reset CWND after idle period longer RTO to "restart window". 1411da177e4SLinus Torvalds * This is the first part of cwnd validation mechanism. */ 142cf533ea5SEric Dumazet static void tcp_cwnd_restart(struct sock *sk, const struct dst_entry *dst) 1431da177e4SLinus Torvalds { 144463c84b9SArnaldo Carvalho de Melo struct tcp_sock *tp = tcp_sk(sk); 1451da177e4SLinus Torvalds s32 delta = tcp_time_stamp - tp->lsndtime; 1461da177e4SLinus Torvalds u32 restart_cwnd = tcp_init_cwnd(tp, dst); 1471da177e4SLinus Torvalds u32 cwnd = tp->snd_cwnd; 1481da177e4SLinus Torvalds 1496687e988SArnaldo Carvalho de Melo tcp_ca_event(sk, CA_EVENT_CWND_RESTART); 1501da177e4SLinus Torvalds 1516687e988SArnaldo Carvalho de Melo tp->snd_ssthresh = tcp_current_ssthresh(sk); 1521da177e4SLinus Torvalds restart_cwnd = min(restart_cwnd, cwnd); 1531da177e4SLinus Torvalds 154463c84b9SArnaldo Carvalho de Melo while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd) 1551da177e4SLinus Torvalds cwnd >>= 1; 1561da177e4SLinus Torvalds tp->snd_cwnd = max(cwnd, restart_cwnd); 1571da177e4SLinus Torvalds tp->snd_cwnd_stamp = tcp_time_stamp; 1581da177e4SLinus Torvalds tp->snd_cwnd_used = 0; 1591da177e4SLinus Torvalds } 1601da177e4SLinus Torvalds 16167edfef7SAndi Kleen /* Congestion state accounting after a packet has been sent. */ 16240efc6faSStephen Hemminger static void tcp_event_data_sent(struct tcp_sock *tp, 163cf533ea5SEric Dumazet struct sock *sk) 1641da177e4SLinus Torvalds { 165463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 166463c84b9SArnaldo Carvalho de Melo const u32 now = tcp_time_stamp; 1671da177e4SLinus Torvalds 16835089bb2SDavid S. Miller if (sysctl_tcp_slow_start_after_idle && 16935089bb2SDavid S. Miller (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto)) 170463c84b9SArnaldo Carvalho de Melo tcp_cwnd_restart(sk, __sk_dst_get(sk)); 1711da177e4SLinus Torvalds 1721da177e4SLinus Torvalds tp->lsndtime = now; 1731da177e4SLinus Torvalds 1741da177e4SLinus Torvalds /* If it is a reply for ato after last received 1751da177e4SLinus Torvalds * packet, enter pingpong mode. 1761da177e4SLinus Torvalds */ 177463c84b9SArnaldo Carvalho de Melo if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato) 178463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pingpong = 1; 1791da177e4SLinus Torvalds } 1801da177e4SLinus Torvalds 18167edfef7SAndi Kleen /* Account for an ACK we sent. */ 18240efc6faSStephen Hemminger static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts) 1831da177e4SLinus Torvalds { 184463c84b9SArnaldo Carvalho de Melo tcp_dec_quickack_mode(sk, pkts); 185463c84b9SArnaldo Carvalho de Melo inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); 1861da177e4SLinus Torvalds } 1871da177e4SLinus Torvalds 1881da177e4SLinus Torvalds /* Determine a window scaling and initial window to offer. 1891da177e4SLinus Torvalds * Based on the assumption that the given amount of space 1901da177e4SLinus Torvalds * will be offered. Store the results in the tp structure. 1911da177e4SLinus Torvalds * NOTE: for smooth operation initial space offering should 1921da177e4SLinus Torvalds * be a multiple of mss if possible. We assume here that mss >= 1. 1931da177e4SLinus Torvalds * This MUST be enforced by all callers. 1941da177e4SLinus Torvalds */ 1951da177e4SLinus Torvalds void tcp_select_initial_window(int __space, __u32 mss, 1961da177e4SLinus Torvalds __u32 *rcv_wnd, __u32 *window_clamp, 19731d12926Slaurent chavey int wscale_ok, __u8 *rcv_wscale, 19831d12926Slaurent chavey __u32 init_rcv_wnd) 1991da177e4SLinus Torvalds { 2001da177e4SLinus Torvalds unsigned int space = (__space < 0 ? 0 : __space); 2011da177e4SLinus Torvalds 2021da177e4SLinus Torvalds /* If no clamp set the clamp to the max possible scaled window */ 2031da177e4SLinus Torvalds if (*window_clamp == 0) 2041da177e4SLinus Torvalds (*window_clamp) = (65535 << 14); 2051da177e4SLinus Torvalds space = min(*window_clamp, space); 2061da177e4SLinus Torvalds 2071da177e4SLinus Torvalds /* Quantize space offering to a multiple of mss if possible. */ 2081da177e4SLinus Torvalds if (space > mss) 2091da177e4SLinus Torvalds space = (space / mss) * mss; 2101da177e4SLinus Torvalds 2111da177e4SLinus Torvalds /* NOTE: offering an initial window larger than 32767 21215d99e02SRick Jones * will break some buggy TCP stacks. If the admin tells us 21315d99e02SRick Jones * it is likely we could be speaking with such a buggy stack 21415d99e02SRick Jones * we will truncate our initial window offering to 32K-1 21515d99e02SRick Jones * unless the remote has sent us a window scaling option, 21615d99e02SRick Jones * which we interpret as a sign the remote TCP is not 21715d99e02SRick Jones * misinterpreting the window field as a signed quantity. 2181da177e4SLinus Torvalds */ 21915d99e02SRick Jones if (sysctl_tcp_workaround_signed_windows) 2201da177e4SLinus Torvalds (*rcv_wnd) = min(space, MAX_TCP_WINDOW); 22115d99e02SRick Jones else 22215d99e02SRick Jones (*rcv_wnd) = space; 22315d99e02SRick Jones 2241da177e4SLinus Torvalds (*rcv_wscale) = 0; 2251da177e4SLinus Torvalds if (wscale_ok) { 2261da177e4SLinus Torvalds /* Set window scaling on max possible window 2271da177e4SLinus Torvalds * See RFC1323 for an explanation of the limit to 14 2281da177e4SLinus Torvalds */ 2291da177e4SLinus Torvalds space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max); 230316c1592SStephen Hemminger space = min_t(u32, space, *window_clamp); 2311da177e4SLinus Torvalds while (space > 65535 && (*rcv_wscale) < 14) { 2321da177e4SLinus Torvalds space >>= 1; 2331da177e4SLinus Torvalds (*rcv_wscale)++; 2341da177e4SLinus Torvalds } 2351da177e4SLinus Torvalds } 2361da177e4SLinus Torvalds 237356f0398SNandita Dukkipati /* Set initial window to a value enough for senders starting with 238356f0398SNandita Dukkipati * initial congestion window of TCP_DEFAULT_INIT_RCVWND. Place 239356f0398SNandita Dukkipati * a limit on the initial window when mss is larger than 1460. 240356f0398SNandita Dukkipati */ 2411da177e4SLinus Torvalds if (mss > (1 << *rcv_wscale)) { 242356f0398SNandita Dukkipati int init_cwnd = TCP_DEFAULT_INIT_RCVWND; 243356f0398SNandita Dukkipati if (mss > 1460) 244356f0398SNandita Dukkipati init_cwnd = 245356f0398SNandita Dukkipati max_t(u32, (1460 * TCP_DEFAULT_INIT_RCVWND) / mss, 2); 24631d12926Slaurent chavey /* when initializing use the value from init_rcv_wnd 24731d12926Slaurent chavey * rather than the default from above 24831d12926Slaurent chavey */ 249b1afde60SNandita Dukkipati if (init_rcv_wnd) 250b1afde60SNandita Dukkipati *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss); 251b1afde60SNandita Dukkipati else 252b1afde60SNandita Dukkipati *rcv_wnd = min(*rcv_wnd, init_cwnd * mss); 2531da177e4SLinus Torvalds } 2541da177e4SLinus Torvalds 2551da177e4SLinus Torvalds /* Set the clamp no higher than max representable value */ 2561da177e4SLinus Torvalds (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp); 2571da177e4SLinus Torvalds } 2584bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_select_initial_window); 2591da177e4SLinus Torvalds 2601da177e4SLinus Torvalds /* Chose a new window to advertise, update state in tcp_sock for the 2611da177e4SLinus Torvalds * socket, and return result with RFC1323 scaling applied. The return 2621da177e4SLinus Torvalds * value can be stuffed directly into th->window for an outgoing 2631da177e4SLinus Torvalds * frame. 2641da177e4SLinus Torvalds */ 26540efc6faSStephen Hemminger static u16 tcp_select_window(struct sock *sk) 2661da177e4SLinus Torvalds { 2671da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2681da177e4SLinus Torvalds u32 cur_win = tcp_receive_window(tp); 2691da177e4SLinus Torvalds u32 new_win = __tcp_select_window(sk); 2701da177e4SLinus Torvalds 2711da177e4SLinus Torvalds /* Never shrink the offered window */ 2721da177e4SLinus Torvalds if (new_win < cur_win) { 2731da177e4SLinus Torvalds /* Danger Will Robinson! 2741da177e4SLinus Torvalds * Don't update rcv_wup/rcv_wnd here or else 2751da177e4SLinus Torvalds * we will not be able to advertise a zero 2761da177e4SLinus Torvalds * window in time. --DaveM 2771da177e4SLinus Torvalds * 2781da177e4SLinus Torvalds * Relax Will Robinson. 2791da177e4SLinus Torvalds */ 280607bfbf2SPatrick McHardy new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale); 2811da177e4SLinus Torvalds } 2821da177e4SLinus Torvalds tp->rcv_wnd = new_win; 2831da177e4SLinus Torvalds tp->rcv_wup = tp->rcv_nxt; 2841da177e4SLinus Torvalds 2851da177e4SLinus Torvalds /* Make sure we do not exceed the maximum possible 2861da177e4SLinus Torvalds * scaled window. 2871da177e4SLinus Torvalds */ 28815d99e02SRick Jones if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows) 2891da177e4SLinus Torvalds new_win = min(new_win, MAX_TCP_WINDOW); 2901da177e4SLinus Torvalds else 2911da177e4SLinus Torvalds new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); 2921da177e4SLinus Torvalds 2931da177e4SLinus Torvalds /* RFC1323 scaling applied */ 2941da177e4SLinus Torvalds new_win >>= tp->rx_opt.rcv_wscale; 2951da177e4SLinus Torvalds 2961da177e4SLinus Torvalds /* If we advertise zero window, disable fast path. */ 2971da177e4SLinus Torvalds if (new_win == 0) 2981da177e4SLinus Torvalds tp->pred_flags = 0; 2991da177e4SLinus Torvalds 3001da177e4SLinus Torvalds return new_win; 3011da177e4SLinus Torvalds } 3021da177e4SLinus Torvalds 30367edfef7SAndi Kleen /* Packet ECN state for a SYN-ACK */ 304cf533ea5SEric Dumazet static inline void TCP_ECN_send_synack(const struct tcp_sock *tp, struct sk_buff *skb) 305bdf1ee5dSIlpo Järvinen { 3064de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR; 307bdf1ee5dSIlpo Järvinen if (!(tp->ecn_flags & TCP_ECN_OK)) 3084de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE; 309bdf1ee5dSIlpo Järvinen } 310bdf1ee5dSIlpo Järvinen 31167edfef7SAndi Kleen /* Packet ECN state for a SYN. */ 312bdf1ee5dSIlpo Järvinen static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb) 313bdf1ee5dSIlpo Järvinen { 314bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 315bdf1ee5dSIlpo Järvinen 316bdf1ee5dSIlpo Järvinen tp->ecn_flags = 0; 317255cac91SIlpo Järvinen if (sysctl_tcp_ecn == 1) { 3184de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR; 319bdf1ee5dSIlpo Järvinen tp->ecn_flags = TCP_ECN_OK; 320bdf1ee5dSIlpo Järvinen } 321bdf1ee5dSIlpo Järvinen } 322bdf1ee5dSIlpo Järvinen 323bdf1ee5dSIlpo Järvinen static __inline__ void 324cf533ea5SEric Dumazet TCP_ECN_make_synack(const struct request_sock *req, struct tcphdr *th) 325bdf1ee5dSIlpo Järvinen { 326bdf1ee5dSIlpo Järvinen if (inet_rsk(req)->ecn_ok) 327bdf1ee5dSIlpo Järvinen th->ece = 1; 328bdf1ee5dSIlpo Järvinen } 329bdf1ee5dSIlpo Järvinen 33067edfef7SAndi Kleen /* Set up ECN state for a packet on a ESTABLISHED socket that is about to 33167edfef7SAndi Kleen * be sent. 33267edfef7SAndi Kleen */ 333bdf1ee5dSIlpo Järvinen static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb, 334bdf1ee5dSIlpo Järvinen int tcp_header_len) 335bdf1ee5dSIlpo Järvinen { 336bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 337bdf1ee5dSIlpo Järvinen 338bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_OK) { 339bdf1ee5dSIlpo Järvinen /* Not-retransmitted data segment: set ECT and inject CWR. */ 340bdf1ee5dSIlpo Järvinen if (skb->len != tcp_header_len && 341bdf1ee5dSIlpo Järvinen !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { 342bdf1ee5dSIlpo Järvinen INET_ECN_xmit(sk); 343bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) { 344bdf1ee5dSIlpo Järvinen tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; 345bdf1ee5dSIlpo Järvinen tcp_hdr(skb)->cwr = 1; 346bdf1ee5dSIlpo Järvinen skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; 347bdf1ee5dSIlpo Järvinen } 348bdf1ee5dSIlpo Järvinen } else { 349bdf1ee5dSIlpo Järvinen /* ACK or retransmitted segment: clear ECT|CE */ 350bdf1ee5dSIlpo Järvinen INET_ECN_dontxmit(sk); 351bdf1ee5dSIlpo Järvinen } 352bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) 353bdf1ee5dSIlpo Järvinen tcp_hdr(skb)->ece = 1; 354bdf1ee5dSIlpo Järvinen } 355bdf1ee5dSIlpo Järvinen } 356bdf1ee5dSIlpo Järvinen 357e870a8efSIlpo Järvinen /* Constructs common control bits of non-data skb. If SYN/FIN is present, 358e870a8efSIlpo Järvinen * auto increment end seqno. 359e870a8efSIlpo Järvinen */ 360e870a8efSIlpo Järvinen static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags) 361e870a8efSIlpo Järvinen { 3622e8e18efSDavid S. Miller skb->ip_summed = CHECKSUM_PARTIAL; 363e870a8efSIlpo Järvinen skb->csum = 0; 364e870a8efSIlpo Järvinen 3654de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags; 366e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->sacked = 0; 367e870a8efSIlpo Järvinen 368e870a8efSIlpo Järvinen skb_shinfo(skb)->gso_segs = 1; 369e870a8efSIlpo Järvinen skb_shinfo(skb)->gso_size = 0; 370e870a8efSIlpo Järvinen skb_shinfo(skb)->gso_type = 0; 371e870a8efSIlpo Järvinen 372e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->seq = seq; 373a3433f35SChangli Gao if (flags & (TCPHDR_SYN | TCPHDR_FIN)) 374e870a8efSIlpo Järvinen seq++; 375e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->end_seq = seq; 376e870a8efSIlpo Järvinen } 377e870a8efSIlpo Järvinen 378a2a385d6SEric Dumazet static inline bool tcp_urg_mode(const struct tcp_sock *tp) 37933f5f57eSIlpo Järvinen { 38033f5f57eSIlpo Järvinen return tp->snd_una != tp->snd_up; 38133f5f57eSIlpo Järvinen } 38233f5f57eSIlpo Järvinen 38333ad798cSAdam Langley #define OPTION_SACK_ADVERTISE (1 << 0) 38433ad798cSAdam Langley #define OPTION_TS (1 << 1) 38533ad798cSAdam Langley #define OPTION_MD5 (1 << 2) 38689e95a61SOri Finkelman #define OPTION_WSCALE (1 << 3) 387bd0388aeSWilliam Allen Simpson #define OPTION_COOKIE_EXTENSION (1 << 4) 388*2100c8d2SYuchung Cheng #define OPTION_FAST_OPEN_COOKIE (1 << 8) 38933ad798cSAdam Langley 39033ad798cSAdam Langley struct tcp_out_options { 391*2100c8d2SYuchung Cheng u16 options; /* bit field of OPTION_* */ 392*2100c8d2SYuchung Cheng u16 mss; /* 0 to disable */ 39333ad798cSAdam Langley u8 ws; /* window scale, 0 to disable */ 39433ad798cSAdam Langley u8 num_sack_blocks; /* number of SACK blocks to include */ 395bd0388aeSWilliam Allen Simpson u8 hash_size; /* bytes in hash_location */ 396bd0388aeSWilliam Allen Simpson __u8 *hash_location; /* temporary pointer, overloaded */ 397*2100c8d2SYuchung Cheng __u32 tsval, tsecr; /* need to include OPTION_TS */ 398*2100c8d2SYuchung Cheng struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */ 39933ad798cSAdam Langley }; 40033ad798cSAdam Langley 401bd0388aeSWilliam Allen Simpson /* The sysctl int routines are generic, so check consistency here. 402bd0388aeSWilliam Allen Simpson */ 403bd0388aeSWilliam Allen Simpson static u8 tcp_cookie_size_check(u8 desired) 404bd0388aeSWilliam Allen Simpson { 405f1987257SEric Dumazet int cookie_size; 406f1987257SEric Dumazet 407f1987257SEric Dumazet if (desired > 0) 408bd0388aeSWilliam Allen Simpson /* previously specified */ 409bd0388aeSWilliam Allen Simpson return desired; 410f1987257SEric Dumazet 411f1987257SEric Dumazet cookie_size = ACCESS_ONCE(sysctl_tcp_cookie_size); 412f1987257SEric Dumazet if (cookie_size <= 0) 413bd0388aeSWilliam Allen Simpson /* no default specified */ 414bd0388aeSWilliam Allen Simpson return 0; 415f1987257SEric Dumazet 416f1987257SEric Dumazet if (cookie_size <= TCP_COOKIE_MIN) 417bd0388aeSWilliam Allen Simpson /* value too small, specify minimum */ 418bd0388aeSWilliam Allen Simpson return TCP_COOKIE_MIN; 419f1987257SEric Dumazet 420f1987257SEric Dumazet if (cookie_size >= TCP_COOKIE_MAX) 421bd0388aeSWilliam Allen Simpson /* value too large, specify maximum */ 422bd0388aeSWilliam Allen Simpson return TCP_COOKIE_MAX; 423f1987257SEric Dumazet 424f1987257SEric Dumazet if (cookie_size & 1) 425bd0388aeSWilliam Allen Simpson /* 8-bit multiple, illegal, fix it */ 426f1987257SEric Dumazet cookie_size++; 427f1987257SEric Dumazet 428f1987257SEric Dumazet return (u8)cookie_size; 429bd0388aeSWilliam Allen Simpson } 430bd0388aeSWilliam Allen Simpson 43167edfef7SAndi Kleen /* Write previously computed TCP options to the packet. 43267edfef7SAndi Kleen * 43367edfef7SAndi Kleen * Beware: Something in the Internet is very sensitive to the ordering of 434fd6149d3SIlpo Järvinen * TCP options, we learned this through the hard way, so be careful here. 435fd6149d3SIlpo Järvinen * Luckily we can at least blame others for their non-compliance but from 436fd6149d3SIlpo Järvinen * inter-operatibility perspective it seems that we're somewhat stuck with 437fd6149d3SIlpo Järvinen * the ordering which we have been using if we want to keep working with 438fd6149d3SIlpo Järvinen * those broken things (not that it currently hurts anybody as there isn't 439fd6149d3SIlpo Järvinen * particular reason why the ordering would need to be changed). 440fd6149d3SIlpo Järvinen * 441fd6149d3SIlpo Järvinen * At least SACK_PERM as the first option is known to lead to a disaster 442fd6149d3SIlpo Järvinen * (but it may well be that other scenarios fail similarly). 443fd6149d3SIlpo Järvinen */ 44433ad798cSAdam Langley static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp, 445bd0388aeSWilliam Allen Simpson struct tcp_out_options *opts) 446bd0388aeSWilliam Allen Simpson { 447*2100c8d2SYuchung Cheng u16 options = opts->options; /* mungable copy */ 448bd0388aeSWilliam Allen Simpson 449bd0388aeSWilliam Allen Simpson /* Having both authentication and cookies for security is redundant, 450bd0388aeSWilliam Allen Simpson * and there's certainly not enough room. Instead, the cookie-less 451bd0388aeSWilliam Allen Simpson * extension variant is proposed. 452bd0388aeSWilliam Allen Simpson * 453bd0388aeSWilliam Allen Simpson * Consider the pessimal case with authentication. The options 454bd0388aeSWilliam Allen Simpson * could look like: 455bd0388aeSWilliam Allen Simpson * COOKIE|MD5(20) + MSS(4) + SACK|TS(12) + WSCALE(4) == 40 456bd0388aeSWilliam Allen Simpson */ 457bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_MD5 & options)) { 458bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_COOKIE_EXTENSION & options)) { 459bd0388aeSWilliam Allen Simpson *ptr++ = htonl((TCPOPT_COOKIE << 24) | 460bd0388aeSWilliam Allen Simpson (TCPOLEN_COOKIE_BASE << 16) | 461bd0388aeSWilliam Allen Simpson (TCPOPT_MD5SIG << 8) | 462bd0388aeSWilliam Allen Simpson TCPOLEN_MD5SIG); 463bd0388aeSWilliam Allen Simpson } else { 46433ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 46533ad798cSAdam Langley (TCPOPT_NOP << 16) | 46633ad798cSAdam Langley (TCPOPT_MD5SIG << 8) | 46733ad798cSAdam Langley TCPOLEN_MD5SIG); 468bd0388aeSWilliam Allen Simpson } 469bd0388aeSWilliam Allen Simpson options &= ~OPTION_COOKIE_EXTENSION; 470bd0388aeSWilliam Allen Simpson /* overload cookie hash location */ 471bd0388aeSWilliam Allen Simpson opts->hash_location = (__u8 *)ptr; 47233ad798cSAdam Langley ptr += 4; 47333ad798cSAdam Langley } 47433ad798cSAdam Langley 475fd6149d3SIlpo Järvinen if (unlikely(opts->mss)) { 476fd6149d3SIlpo Järvinen *ptr++ = htonl((TCPOPT_MSS << 24) | 477fd6149d3SIlpo Järvinen (TCPOLEN_MSS << 16) | 478fd6149d3SIlpo Järvinen opts->mss); 479fd6149d3SIlpo Järvinen } 480fd6149d3SIlpo Järvinen 481bd0388aeSWilliam Allen Simpson if (likely(OPTION_TS & options)) { 482bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 48333ad798cSAdam Langley *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | 48433ad798cSAdam Langley (TCPOLEN_SACK_PERM << 16) | 48533ad798cSAdam Langley (TCPOPT_TIMESTAMP << 8) | 48633ad798cSAdam Langley TCPOLEN_TIMESTAMP); 487bd0388aeSWilliam Allen Simpson options &= ~OPTION_SACK_ADVERTISE; 48833ad798cSAdam Langley } else { 489496c98dfSYOSHIFUJI Hideaki *ptr++ = htonl((TCPOPT_NOP << 24) | 49040efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 49140efc6faSStephen Hemminger (TCPOPT_TIMESTAMP << 8) | 49240efc6faSStephen Hemminger TCPOLEN_TIMESTAMP); 49340efc6faSStephen Hemminger } 49433ad798cSAdam Langley *ptr++ = htonl(opts->tsval); 49533ad798cSAdam Langley *ptr++ = htonl(opts->tsecr); 49633ad798cSAdam Langley } 49733ad798cSAdam Langley 498bd0388aeSWilliam Allen Simpson /* Specification requires after timestamp, so do it now. 499bd0388aeSWilliam Allen Simpson * 500bd0388aeSWilliam Allen Simpson * Consider the pessimal case without authentication. The options 501bd0388aeSWilliam Allen Simpson * could look like: 502bd0388aeSWilliam Allen Simpson * MSS(4) + SACK|TS(12) + COOKIE(20) + WSCALE(4) == 40 503bd0388aeSWilliam Allen Simpson */ 504bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_COOKIE_EXTENSION & options)) { 505bd0388aeSWilliam Allen Simpson __u8 *cookie_copy = opts->hash_location; 506bd0388aeSWilliam Allen Simpson u8 cookie_size = opts->hash_size; 507bd0388aeSWilliam Allen Simpson 508bd0388aeSWilliam Allen Simpson /* 8-bit multiple handled in tcp_cookie_size_check() above, 509bd0388aeSWilliam Allen Simpson * and elsewhere. 510bd0388aeSWilliam Allen Simpson */ 511bd0388aeSWilliam Allen Simpson if (0x2 & cookie_size) { 512bd0388aeSWilliam Allen Simpson __u8 *p = (__u8 *)ptr; 513bd0388aeSWilliam Allen Simpson 514bd0388aeSWilliam Allen Simpson /* 16-bit multiple */ 515bd0388aeSWilliam Allen Simpson *p++ = TCPOPT_COOKIE; 516bd0388aeSWilliam Allen Simpson *p++ = TCPOLEN_COOKIE_BASE + cookie_size; 517bd0388aeSWilliam Allen Simpson *p++ = *cookie_copy++; 518bd0388aeSWilliam Allen Simpson *p++ = *cookie_copy++; 519bd0388aeSWilliam Allen Simpson ptr++; 520bd0388aeSWilliam Allen Simpson cookie_size -= 2; 521bd0388aeSWilliam Allen Simpson } else { 522bd0388aeSWilliam Allen Simpson /* 32-bit multiple */ 523bd0388aeSWilliam Allen Simpson *ptr++ = htonl(((TCPOPT_NOP << 24) | 524bd0388aeSWilliam Allen Simpson (TCPOPT_NOP << 16) | 525bd0388aeSWilliam Allen Simpson (TCPOPT_COOKIE << 8) | 526bd0388aeSWilliam Allen Simpson TCPOLEN_COOKIE_BASE) + 527bd0388aeSWilliam Allen Simpson cookie_size); 528bd0388aeSWilliam Allen Simpson } 529bd0388aeSWilliam Allen Simpson 530bd0388aeSWilliam Allen Simpson if (cookie_size > 0) { 531bd0388aeSWilliam Allen Simpson memcpy(ptr, cookie_copy, cookie_size); 532bd0388aeSWilliam Allen Simpson ptr += (cookie_size / 4); 533bd0388aeSWilliam Allen Simpson } 534bd0388aeSWilliam Allen Simpson } 535bd0388aeSWilliam Allen Simpson 536bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 53733ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 53833ad798cSAdam Langley (TCPOPT_NOP << 16) | 53933ad798cSAdam Langley (TCPOPT_SACK_PERM << 8) | 54033ad798cSAdam Langley TCPOLEN_SACK_PERM); 54133ad798cSAdam Langley } 54233ad798cSAdam Langley 543bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_WSCALE & options)) { 54433ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 54533ad798cSAdam Langley (TCPOPT_WINDOW << 16) | 54633ad798cSAdam Langley (TCPOLEN_WINDOW << 8) | 54733ad798cSAdam Langley opts->ws); 54833ad798cSAdam Langley } 54933ad798cSAdam Langley 55033ad798cSAdam Langley if (unlikely(opts->num_sack_blocks)) { 55133ad798cSAdam Langley struct tcp_sack_block *sp = tp->rx_opt.dsack ? 55233ad798cSAdam Langley tp->duplicate_sack : tp->selective_acks; 55340efc6faSStephen Hemminger int this_sack; 55440efc6faSStephen Hemminger 55540efc6faSStephen Hemminger *ptr++ = htonl((TCPOPT_NOP << 24) | 55640efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 55740efc6faSStephen Hemminger (TCPOPT_SACK << 8) | 55833ad798cSAdam Langley (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * 55940efc6faSStephen Hemminger TCPOLEN_SACK_PERBLOCK))); 5602de979bdSStephen Hemminger 56133ad798cSAdam Langley for (this_sack = 0; this_sack < opts->num_sack_blocks; 56233ad798cSAdam Langley ++this_sack) { 56340efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].start_seq); 56440efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].end_seq); 56540efc6faSStephen Hemminger } 5662de979bdSStephen Hemminger 56740efc6faSStephen Hemminger tp->rx_opt.dsack = 0; 56840efc6faSStephen Hemminger } 569*2100c8d2SYuchung Cheng 570*2100c8d2SYuchung Cheng if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) { 571*2100c8d2SYuchung Cheng struct tcp_fastopen_cookie *foc = opts->fastopen_cookie; 572*2100c8d2SYuchung Cheng 573*2100c8d2SYuchung Cheng *ptr++ = htonl((TCPOPT_EXP << 24) | 574*2100c8d2SYuchung Cheng ((TCPOLEN_EXP_FASTOPEN_BASE + foc->len) << 16) | 575*2100c8d2SYuchung Cheng TCPOPT_FASTOPEN_MAGIC); 576*2100c8d2SYuchung Cheng 577*2100c8d2SYuchung Cheng memcpy(ptr, foc->val, foc->len); 578*2100c8d2SYuchung Cheng if ((foc->len & 3) == 2) { 579*2100c8d2SYuchung Cheng u8 *align = ((u8 *)ptr) + foc->len; 580*2100c8d2SYuchung Cheng align[0] = align[1] = TCPOPT_NOP; 581*2100c8d2SYuchung Cheng } 582*2100c8d2SYuchung Cheng ptr += (foc->len + 3) >> 2; 583*2100c8d2SYuchung Cheng } 58440efc6faSStephen Hemminger } 58540efc6faSStephen Hemminger 58667edfef7SAndi Kleen /* Compute TCP options for SYN packets. This is not the final 58767edfef7SAndi Kleen * network wire format yet. 58867edfef7SAndi Kleen */ 58995c96174SEric Dumazet static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb, 59033ad798cSAdam Langley struct tcp_out_options *opts, 591cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 592cf533ea5SEric Dumazet { 59333ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 594bd0388aeSWilliam Allen Simpson struct tcp_cookie_values *cvp = tp->cookie_values; 59595c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 596bd0388aeSWilliam Allen Simpson u8 cookie_size = (!tp->rx_opt.cookie_out_never && cvp != NULL) ? 597bd0388aeSWilliam Allen Simpson tcp_cookie_size_check(cvp->cookie_desired) : 598bd0388aeSWilliam Allen Simpson 0; 59933ad798cSAdam Langley 600cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 60133ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 60233ad798cSAdam Langley if (*md5) { 60333ad798cSAdam Langley opts->options |= OPTION_MD5; 604bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 605cfb6eeb4SYOSHIFUJI Hideaki } 60633ad798cSAdam Langley #else 60733ad798cSAdam Langley *md5 = NULL; 608cfb6eeb4SYOSHIFUJI Hideaki #endif 60933ad798cSAdam Langley 61033ad798cSAdam Langley /* We always get an MSS option. The option bytes which will be seen in 61133ad798cSAdam Langley * normal data packets should timestamps be used, must be in the MSS 61233ad798cSAdam Langley * advertised. But we subtract them from tp->mss_cache so that 61333ad798cSAdam Langley * calculations in tcp_sendmsg are simpler etc. So account for this 61433ad798cSAdam Langley * fact here if necessary. If we don't do this correctly, as a 61533ad798cSAdam Langley * receiver we won't recognize data packets as being full sized when we 61633ad798cSAdam Langley * should, and thus we won't abide by the delayed ACK rules correctly. 61733ad798cSAdam Langley * SACKs don't matter, we never delay an ACK when we have any of those 61833ad798cSAdam Langley * going out. */ 61933ad798cSAdam Langley opts->mss = tcp_advertise_mss(sk); 620bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 62133ad798cSAdam Langley 622bb5b7c11SDavid S. Miller if (likely(sysctl_tcp_timestamps && *md5 == NULL)) { 62333ad798cSAdam Langley opts->options |= OPTION_TS; 62433ad798cSAdam Langley opts->tsval = TCP_SKB_CB(skb)->when; 62533ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 626bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 62733ad798cSAdam Langley } 628bb5b7c11SDavid S. Miller if (likely(sysctl_tcp_window_scaling)) { 62933ad798cSAdam Langley opts->ws = tp->rx_opt.rcv_wscale; 63089e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 631bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 63233ad798cSAdam Langley } 633bb5b7c11SDavid S. Miller if (likely(sysctl_tcp_sack)) { 63433ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 635b32d1310SDavid S. Miller if (unlikely(!(OPTION_TS & opts->options))) 636bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 63733ad798cSAdam Langley } 63833ad798cSAdam Langley 639bd0388aeSWilliam Allen Simpson /* Note that timestamps are required by the specification. 640bd0388aeSWilliam Allen Simpson * 641bd0388aeSWilliam Allen Simpson * Odd numbers of bytes are prohibited by the specification, ensuring 642bd0388aeSWilliam Allen Simpson * that the cookie is 16-bit aligned, and the resulting cookie pair is 643bd0388aeSWilliam Allen Simpson * 32-bit aligned. 644bd0388aeSWilliam Allen Simpson */ 645bd0388aeSWilliam Allen Simpson if (*md5 == NULL && 646bd0388aeSWilliam Allen Simpson (OPTION_TS & opts->options) && 647bd0388aeSWilliam Allen Simpson cookie_size > 0) { 648bd0388aeSWilliam Allen Simpson int need = TCPOLEN_COOKIE_BASE + cookie_size; 649bd0388aeSWilliam Allen Simpson 650bd0388aeSWilliam Allen Simpson if (0x2 & need) { 651bd0388aeSWilliam Allen Simpson /* 32-bit multiple */ 652bd0388aeSWilliam Allen Simpson need += 2; /* NOPs */ 653bd0388aeSWilliam Allen Simpson 654bd0388aeSWilliam Allen Simpson if (need > remaining) { 655bd0388aeSWilliam Allen Simpson /* try shrinking cookie to fit */ 656bd0388aeSWilliam Allen Simpson cookie_size -= 2; 657bd0388aeSWilliam Allen Simpson need -= 4; 658bd0388aeSWilliam Allen Simpson } 659bd0388aeSWilliam Allen Simpson } 660bd0388aeSWilliam Allen Simpson while (need > remaining && TCP_COOKIE_MIN <= cookie_size) { 661bd0388aeSWilliam Allen Simpson cookie_size -= 4; 662bd0388aeSWilliam Allen Simpson need -= 4; 663bd0388aeSWilliam Allen Simpson } 664bd0388aeSWilliam Allen Simpson if (TCP_COOKIE_MIN <= cookie_size) { 665bd0388aeSWilliam Allen Simpson opts->options |= OPTION_COOKIE_EXTENSION; 666bd0388aeSWilliam Allen Simpson opts->hash_location = (__u8 *)&cvp->cookie_pair[0]; 667bd0388aeSWilliam Allen Simpson opts->hash_size = cookie_size; 668bd0388aeSWilliam Allen Simpson 669bd0388aeSWilliam Allen Simpson /* Remember for future incarnations. */ 670bd0388aeSWilliam Allen Simpson cvp->cookie_desired = cookie_size; 671bd0388aeSWilliam Allen Simpson 672bd0388aeSWilliam Allen Simpson if (cvp->cookie_desired != cvp->cookie_pair_size) { 673bd0388aeSWilliam Allen Simpson /* Currently use random bytes as a nonce, 674bd0388aeSWilliam Allen Simpson * assuming these are completely unpredictable 675bd0388aeSWilliam Allen Simpson * by hostile users of the same system. 676bd0388aeSWilliam Allen Simpson */ 677bd0388aeSWilliam Allen Simpson get_random_bytes(&cvp->cookie_pair[0], 678bd0388aeSWilliam Allen Simpson cookie_size); 679bd0388aeSWilliam Allen Simpson cvp->cookie_pair_size = cookie_size; 680bd0388aeSWilliam Allen Simpson } 681bd0388aeSWilliam Allen Simpson 682bd0388aeSWilliam Allen Simpson remaining -= need; 683bd0388aeSWilliam Allen Simpson } 684bd0388aeSWilliam Allen Simpson } 685bd0388aeSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 68633ad798cSAdam Langley } 68733ad798cSAdam Langley 68867edfef7SAndi Kleen /* Set up TCP options for SYN-ACKs. */ 68995c96174SEric Dumazet static unsigned int tcp_synack_options(struct sock *sk, 69033ad798cSAdam Langley struct request_sock *req, 69195c96174SEric Dumazet unsigned int mss, struct sk_buff *skb, 69233ad798cSAdam Langley struct tcp_out_options *opts, 6934957faadSWilliam Allen Simpson struct tcp_md5sig_key **md5, 6944957faadSWilliam Allen Simpson struct tcp_extend_values *xvp) 6954957faadSWilliam Allen Simpson { 69633ad798cSAdam Langley struct inet_request_sock *ireq = inet_rsk(req); 69795c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 6984957faadSWilliam Allen Simpson u8 cookie_plus = (xvp != NULL && !xvp->cookie_out_never) ? 6994957faadSWilliam Allen Simpson xvp->cookie_plus : 7004957faadSWilliam Allen Simpson 0; 70133ad798cSAdam Langley 70233ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 70333ad798cSAdam Langley *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req); 70433ad798cSAdam Langley if (*md5) { 70533ad798cSAdam Langley opts->options |= OPTION_MD5; 7064957faadSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 7074957faadSWilliam Allen Simpson 7084957faadSWilliam Allen Simpson /* We can't fit any SACK blocks in a packet with MD5 + TS 7094957faadSWilliam Allen Simpson * options. There was discussion about disabling SACK 7104957faadSWilliam Allen Simpson * rather than TS in order to fit in better with old, 7114957faadSWilliam Allen Simpson * buggy kernels, but that was deemed to be unnecessary. 7124957faadSWilliam Allen Simpson */ 713de213e5eSEric Dumazet ireq->tstamp_ok &= !ireq->sack_ok; 71433ad798cSAdam Langley } 71533ad798cSAdam Langley #else 71633ad798cSAdam Langley *md5 = NULL; 71733ad798cSAdam Langley #endif 71833ad798cSAdam Langley 7194957faadSWilliam Allen Simpson /* We always send an MSS option. */ 72033ad798cSAdam Langley opts->mss = mss; 7214957faadSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 72233ad798cSAdam Langley 72333ad798cSAdam Langley if (likely(ireq->wscale_ok)) { 72433ad798cSAdam Langley opts->ws = ireq->rcv_wscale; 72589e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 7264957faadSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 72733ad798cSAdam Langley } 728de213e5eSEric Dumazet if (likely(ireq->tstamp_ok)) { 72933ad798cSAdam Langley opts->options |= OPTION_TS; 73033ad798cSAdam Langley opts->tsval = TCP_SKB_CB(skb)->when; 73133ad798cSAdam Langley opts->tsecr = req->ts_recent; 7324957faadSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 73333ad798cSAdam Langley } 73433ad798cSAdam Langley if (likely(ireq->sack_ok)) { 73533ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 736de213e5eSEric Dumazet if (unlikely(!ireq->tstamp_ok)) 7374957faadSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 73833ad798cSAdam Langley } 73933ad798cSAdam Langley 7404957faadSWilliam Allen Simpson /* Similar rationale to tcp_syn_options() applies here, too. 7414957faadSWilliam Allen Simpson * If the <SYN> options fit, the same options should fit now! 7424957faadSWilliam Allen Simpson */ 7434957faadSWilliam Allen Simpson if (*md5 == NULL && 744de213e5eSEric Dumazet ireq->tstamp_ok && 7454957faadSWilliam Allen Simpson cookie_plus > TCPOLEN_COOKIE_BASE) { 7464957faadSWilliam Allen Simpson int need = cookie_plus; /* has TCPOLEN_COOKIE_BASE */ 7474957faadSWilliam Allen Simpson 7484957faadSWilliam Allen Simpson if (0x2 & need) { 7494957faadSWilliam Allen Simpson /* 32-bit multiple */ 7504957faadSWilliam Allen Simpson need += 2; /* NOPs */ 7514957faadSWilliam Allen Simpson } 7524957faadSWilliam Allen Simpson if (need <= remaining) { 7534957faadSWilliam Allen Simpson opts->options |= OPTION_COOKIE_EXTENSION; 7544957faadSWilliam Allen Simpson opts->hash_size = cookie_plus - TCPOLEN_COOKIE_BASE; 7554957faadSWilliam Allen Simpson remaining -= need; 7564957faadSWilliam Allen Simpson } else { 7574957faadSWilliam Allen Simpson /* There's no error return, so flag it. */ 7584957faadSWilliam Allen Simpson xvp->cookie_out_never = 1; /* true */ 7594957faadSWilliam Allen Simpson opts->hash_size = 0; 7604957faadSWilliam Allen Simpson } 7614957faadSWilliam Allen Simpson } 7624957faadSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 76333ad798cSAdam Langley } 76433ad798cSAdam Langley 76567edfef7SAndi Kleen /* Compute TCP options for ESTABLISHED sockets. This is not the 76667edfef7SAndi Kleen * final wire format yet. 76767edfef7SAndi Kleen */ 76895c96174SEric Dumazet static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb, 76933ad798cSAdam Langley struct tcp_out_options *opts, 770cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 771cf533ea5SEric Dumazet { 77233ad798cSAdam Langley struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL; 77333ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 77495c96174SEric Dumazet unsigned int size = 0; 775cabeccbdSIlpo Järvinen unsigned int eff_sacks; 77633ad798cSAdam Langley 77733ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 77833ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 77933ad798cSAdam Langley if (unlikely(*md5)) { 78033ad798cSAdam Langley opts->options |= OPTION_MD5; 78133ad798cSAdam Langley size += TCPOLEN_MD5SIG_ALIGNED; 78233ad798cSAdam Langley } 78333ad798cSAdam Langley #else 78433ad798cSAdam Langley *md5 = NULL; 78533ad798cSAdam Langley #endif 78633ad798cSAdam Langley 78733ad798cSAdam Langley if (likely(tp->rx_opt.tstamp_ok)) { 78833ad798cSAdam Langley opts->options |= OPTION_TS; 78933ad798cSAdam Langley opts->tsval = tcb ? tcb->when : 0; 79033ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 79133ad798cSAdam Langley size += TCPOLEN_TSTAMP_ALIGNED; 79233ad798cSAdam Langley } 79333ad798cSAdam Langley 794cabeccbdSIlpo Järvinen eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; 795cabeccbdSIlpo Järvinen if (unlikely(eff_sacks)) { 79695c96174SEric Dumazet const unsigned int remaining = MAX_TCP_OPTION_SPACE - size; 79733ad798cSAdam Langley opts->num_sack_blocks = 79895c96174SEric Dumazet min_t(unsigned int, eff_sacks, 79933ad798cSAdam Langley (remaining - TCPOLEN_SACK_BASE_ALIGNED) / 80033ad798cSAdam Langley TCPOLEN_SACK_PERBLOCK); 80133ad798cSAdam Langley size += TCPOLEN_SACK_BASE_ALIGNED + 80233ad798cSAdam Langley opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; 80333ad798cSAdam Langley } 80433ad798cSAdam Langley 80533ad798cSAdam Langley return size; 80640efc6faSStephen Hemminger } 8071da177e4SLinus Torvalds 80846d3ceabSEric Dumazet 80946d3ceabSEric Dumazet /* TCP SMALL QUEUES (TSQ) 81046d3ceabSEric Dumazet * 81146d3ceabSEric Dumazet * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev) 81246d3ceabSEric Dumazet * to reduce RTT and bufferbloat. 81346d3ceabSEric Dumazet * We do this using a special skb destructor (tcp_wfree). 81446d3ceabSEric Dumazet * 81546d3ceabSEric Dumazet * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb 81646d3ceabSEric Dumazet * needs to be reallocated in a driver. 81746d3ceabSEric Dumazet * The invariant being skb->truesize substracted from sk->sk_wmem_alloc 81846d3ceabSEric Dumazet * 81946d3ceabSEric Dumazet * Since transmit from skb destructor is forbidden, we use a tasklet 82046d3ceabSEric Dumazet * to process all sockets that eventually need to send more skbs. 82146d3ceabSEric Dumazet * We use one tasklet per cpu, with its own queue of sockets. 82246d3ceabSEric Dumazet */ 82346d3ceabSEric Dumazet struct tsq_tasklet { 82446d3ceabSEric Dumazet struct tasklet_struct tasklet; 82546d3ceabSEric Dumazet struct list_head head; /* queue of tcp sockets */ 82646d3ceabSEric Dumazet }; 82746d3ceabSEric Dumazet static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet); 82846d3ceabSEric Dumazet 82946d3ceabSEric Dumazet /* 83046d3ceabSEric Dumazet * One tasklest per cpu tries to send more skbs. 83146d3ceabSEric Dumazet * We run in tasklet context but need to disable irqs when 83246d3ceabSEric Dumazet * transfering tsq->head because tcp_wfree() might 83346d3ceabSEric Dumazet * interrupt us (non NAPI drivers) 83446d3ceabSEric Dumazet */ 83546d3ceabSEric Dumazet static void tcp_tasklet_func(unsigned long data) 83646d3ceabSEric Dumazet { 83746d3ceabSEric Dumazet struct tsq_tasklet *tsq = (struct tsq_tasklet *)data; 83846d3ceabSEric Dumazet LIST_HEAD(list); 83946d3ceabSEric Dumazet unsigned long flags; 84046d3ceabSEric Dumazet struct list_head *q, *n; 84146d3ceabSEric Dumazet struct tcp_sock *tp; 84246d3ceabSEric Dumazet struct sock *sk; 84346d3ceabSEric Dumazet 84446d3ceabSEric Dumazet local_irq_save(flags); 84546d3ceabSEric Dumazet list_splice_init(&tsq->head, &list); 84646d3ceabSEric Dumazet local_irq_restore(flags); 84746d3ceabSEric Dumazet 84846d3ceabSEric Dumazet list_for_each_safe(q, n, &list) { 84946d3ceabSEric Dumazet tp = list_entry(q, struct tcp_sock, tsq_node); 85046d3ceabSEric Dumazet list_del(&tp->tsq_node); 85146d3ceabSEric Dumazet 85246d3ceabSEric Dumazet sk = (struct sock *)tp; 85346d3ceabSEric Dumazet bh_lock_sock(sk); 85446d3ceabSEric Dumazet 85546d3ceabSEric Dumazet if (!sock_owned_by_user(sk)) { 85646d3ceabSEric Dumazet if ((1 << sk->sk_state) & 85746d3ceabSEric Dumazet (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | 858d01cb207SEric Dumazet TCPF_CLOSING | TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) 85946d3ceabSEric Dumazet tcp_write_xmit(sk, 86046d3ceabSEric Dumazet tcp_current_mss(sk), 86146d3ceabSEric Dumazet 0, 0, 86246d3ceabSEric Dumazet GFP_ATOMIC); 86346d3ceabSEric Dumazet } else { 86446d3ceabSEric Dumazet /* defer the work to tcp_release_cb() */ 86546d3ceabSEric Dumazet set_bit(TSQ_OWNED, &tp->tsq_flags); 86646d3ceabSEric Dumazet } 86746d3ceabSEric Dumazet bh_unlock_sock(sk); 86846d3ceabSEric Dumazet 86946d3ceabSEric Dumazet clear_bit(TSQ_QUEUED, &tp->tsq_flags); 87046d3ceabSEric Dumazet sk_free(sk); 87146d3ceabSEric Dumazet } 87246d3ceabSEric Dumazet } 87346d3ceabSEric Dumazet 87446d3ceabSEric Dumazet /** 87546d3ceabSEric Dumazet * tcp_release_cb - tcp release_sock() callback 87646d3ceabSEric Dumazet * @sk: socket 87746d3ceabSEric Dumazet * 87846d3ceabSEric Dumazet * called from release_sock() to perform protocol dependent 87946d3ceabSEric Dumazet * actions before socket release. 88046d3ceabSEric Dumazet */ 88146d3ceabSEric Dumazet void tcp_release_cb(struct sock *sk) 88246d3ceabSEric Dumazet { 88346d3ceabSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 88446d3ceabSEric Dumazet 88546d3ceabSEric Dumazet if (test_and_clear_bit(TSQ_OWNED, &tp->tsq_flags)) { 88646d3ceabSEric Dumazet if ((1 << sk->sk_state) & 88746d3ceabSEric Dumazet (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | 888d01cb207SEric Dumazet TCPF_CLOSING | TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) 88946d3ceabSEric Dumazet tcp_write_xmit(sk, 89046d3ceabSEric Dumazet tcp_current_mss(sk), 89146d3ceabSEric Dumazet 0, 0, 89246d3ceabSEric Dumazet GFP_ATOMIC); 89346d3ceabSEric Dumazet } 89446d3ceabSEric Dumazet } 89546d3ceabSEric Dumazet EXPORT_SYMBOL(tcp_release_cb); 89646d3ceabSEric Dumazet 89746d3ceabSEric Dumazet void __init tcp_tasklet_init(void) 89846d3ceabSEric Dumazet { 89946d3ceabSEric Dumazet int i; 90046d3ceabSEric Dumazet 90146d3ceabSEric Dumazet for_each_possible_cpu(i) { 90246d3ceabSEric Dumazet struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i); 90346d3ceabSEric Dumazet 90446d3ceabSEric Dumazet INIT_LIST_HEAD(&tsq->head); 90546d3ceabSEric Dumazet tasklet_init(&tsq->tasklet, 90646d3ceabSEric Dumazet tcp_tasklet_func, 90746d3ceabSEric Dumazet (unsigned long)tsq); 90846d3ceabSEric Dumazet } 90946d3ceabSEric Dumazet } 91046d3ceabSEric Dumazet 91146d3ceabSEric Dumazet /* 91246d3ceabSEric Dumazet * Write buffer destructor automatically called from kfree_skb. 91346d3ceabSEric Dumazet * We cant xmit new skbs from this context, as we might already 91446d3ceabSEric Dumazet * hold qdisc lock. 91546d3ceabSEric Dumazet */ 91646d3ceabSEric Dumazet void tcp_wfree(struct sk_buff *skb) 91746d3ceabSEric Dumazet { 91846d3ceabSEric Dumazet struct sock *sk = skb->sk; 91946d3ceabSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 92046d3ceabSEric Dumazet 92146d3ceabSEric Dumazet if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) && 92246d3ceabSEric Dumazet !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) { 92346d3ceabSEric Dumazet unsigned long flags; 92446d3ceabSEric Dumazet struct tsq_tasklet *tsq; 92546d3ceabSEric Dumazet 92646d3ceabSEric Dumazet /* Keep a ref on socket. 92746d3ceabSEric Dumazet * This last ref will be released in tcp_tasklet_func() 92846d3ceabSEric Dumazet */ 92946d3ceabSEric Dumazet atomic_sub(skb->truesize - 1, &sk->sk_wmem_alloc); 93046d3ceabSEric Dumazet 93146d3ceabSEric Dumazet /* queue this socket to tasklet queue */ 93246d3ceabSEric Dumazet local_irq_save(flags); 93346d3ceabSEric Dumazet tsq = &__get_cpu_var(tsq_tasklet); 93446d3ceabSEric Dumazet list_add(&tp->tsq_node, &tsq->head); 93546d3ceabSEric Dumazet tasklet_schedule(&tsq->tasklet); 93646d3ceabSEric Dumazet local_irq_restore(flags); 93746d3ceabSEric Dumazet } else { 93846d3ceabSEric Dumazet sock_wfree(skb); 93946d3ceabSEric Dumazet } 94046d3ceabSEric Dumazet } 94146d3ceabSEric Dumazet 9421da177e4SLinus Torvalds /* This routine actually transmits TCP packets queued in by 9431da177e4SLinus Torvalds * tcp_do_sendmsg(). This is used by both the initial 9441da177e4SLinus Torvalds * transmission and possible later retransmissions. 9451da177e4SLinus Torvalds * All SKB's seen here are completely headerless. It is our 9461da177e4SLinus Torvalds * job to build the TCP header, and pass the packet down to 9471da177e4SLinus Torvalds * IP so it can do the same plus pass the packet off to the 9481da177e4SLinus Torvalds * device. 9491da177e4SLinus Torvalds * 9501da177e4SLinus Torvalds * We are working here with either a clone of the original 9511da177e4SLinus Torvalds * SKB, or a fresh unique copy made by the retransmit engine. 9521da177e4SLinus Torvalds */ 953056834d9SIlpo Järvinen static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, 954056834d9SIlpo Järvinen gfp_t gfp_mask) 9551da177e4SLinus Torvalds { 9566687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 957dfb4b9dcSDavid S. Miller struct inet_sock *inet; 958dfb4b9dcSDavid S. Miller struct tcp_sock *tp; 959dfb4b9dcSDavid S. Miller struct tcp_skb_cb *tcb; 96033ad798cSAdam Langley struct tcp_out_options opts; 96195c96174SEric Dumazet unsigned int tcp_options_size, tcp_header_size; 962cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 9631da177e4SLinus Torvalds struct tcphdr *th; 9641da177e4SLinus Torvalds int err; 9651da177e4SLinus Torvalds 966dfb4b9dcSDavid S. Miller BUG_ON(!skb || !tcp_skb_pcount(skb)); 967dfb4b9dcSDavid S. Miller 968dfb4b9dcSDavid S. Miller /* If congestion control is doing timestamping, we must 969dfb4b9dcSDavid S. Miller * take such a timestamp before we potentially clone/copy. 970dfb4b9dcSDavid S. Miller */ 971164891aaSStephen Hemminger if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP) 972dfb4b9dcSDavid S. Miller __net_timestamp(skb); 973dfb4b9dcSDavid S. Miller 974dfb4b9dcSDavid S. Miller if (likely(clone_it)) { 975dfb4b9dcSDavid S. Miller if (unlikely(skb_cloned(skb))) 976dfb4b9dcSDavid S. Miller skb = pskb_copy(skb, gfp_mask); 977dfb4b9dcSDavid S. Miller else 978dfb4b9dcSDavid S. Miller skb = skb_clone(skb, gfp_mask); 979dfb4b9dcSDavid S. Miller if (unlikely(!skb)) 980dfb4b9dcSDavid S. Miller return -ENOBUFS; 981dfb4b9dcSDavid S. Miller } 982dfb4b9dcSDavid S. Miller 983dfb4b9dcSDavid S. Miller inet = inet_sk(sk); 984dfb4b9dcSDavid S. Miller tp = tcp_sk(sk); 985dfb4b9dcSDavid S. Miller tcb = TCP_SKB_CB(skb); 98633ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 9871da177e4SLinus Torvalds 9884de075e0SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) 98933ad798cSAdam Langley tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); 99033ad798cSAdam Langley else 99133ad798cSAdam Langley tcp_options_size = tcp_established_options(sk, skb, &opts, 99233ad798cSAdam Langley &md5); 99333ad798cSAdam Langley tcp_header_size = tcp_options_size + sizeof(struct tcphdr); 9941da177e4SLinus Torvalds 9953853b584STom Herbert if (tcp_packets_in_flight(tp) == 0) { 9966687e988SArnaldo Carvalho de Melo tcp_ca_event(sk, CA_EVENT_TX_START); 9973853b584STom Herbert skb->ooo_okay = 1; 9983853b584STom Herbert } else 9993853b584STom Herbert skb->ooo_okay = 0; 10001da177e4SLinus Torvalds 1001aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 1002aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 100346d3ceabSEric Dumazet 100446d3ceabSEric Dumazet skb_orphan(skb); 100546d3ceabSEric Dumazet skb->sk = sk; 100646d3ceabSEric Dumazet skb->destructor = (sysctl_tcp_limit_output_bytes > 0) ? 100746d3ceabSEric Dumazet tcp_wfree : sock_wfree; 100846d3ceabSEric Dumazet atomic_add(skb->truesize, &sk->sk_wmem_alloc); 10091da177e4SLinus Torvalds 10101da177e4SLinus Torvalds /* Build TCP header and checksum it. */ 1011aa8223c7SArnaldo Carvalho de Melo th = tcp_hdr(skb); 1012c720c7e8SEric Dumazet th->source = inet->inet_sport; 1013c720c7e8SEric Dumazet th->dest = inet->inet_dport; 10141da177e4SLinus Torvalds th->seq = htonl(tcb->seq); 10151da177e4SLinus Torvalds th->ack_seq = htonl(tp->rcv_nxt); 1016df7a3b07SAl Viro *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | 10174de075e0SEric Dumazet tcb->tcp_flags); 1018dfb4b9dcSDavid S. Miller 10194de075e0SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) { 10201da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments 10211da177e4SLinus Torvalds * is never scaled. 10221da177e4SLinus Torvalds */ 1023600ff0c2SIlpo Järvinen th->window = htons(min(tp->rcv_wnd, 65535U)); 10241da177e4SLinus Torvalds } else { 10251da177e4SLinus Torvalds th->window = htons(tcp_select_window(sk)); 10261da177e4SLinus Torvalds } 10271da177e4SLinus Torvalds th->check = 0; 10281da177e4SLinus Torvalds th->urg_ptr = 0; 10291da177e4SLinus Torvalds 103033f5f57eSIlpo Järvinen /* The urg_mode check is necessary during a below snd_una win probe */ 10317691367dSHerbert Xu if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { 10327691367dSHerbert Xu if (before(tp->snd_up, tcb->seq + 0x10000)) { 10331da177e4SLinus Torvalds th->urg_ptr = htons(tp->snd_up - tcb->seq); 10341da177e4SLinus Torvalds th->urg = 1; 10357691367dSHerbert Xu } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { 10360eae88f3SEric Dumazet th->urg_ptr = htons(0xFFFF); 10377691367dSHerbert Xu th->urg = 1; 10387691367dSHerbert Xu } 10391da177e4SLinus Torvalds } 10401da177e4SLinus Torvalds 1041bd0388aeSWilliam Allen Simpson tcp_options_write((__be32 *)(th + 1), tp, &opts); 10424de075e0SEric Dumazet if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0)) 10439e412ba7SIlpo Järvinen TCP_ECN_send(sk, skb, tcp_header_size); 1044dfb4b9dcSDavid S. Miller 1045cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 1046cfb6eeb4SYOSHIFUJI Hideaki /* Calculate the MD5 hash, as we have all we need now */ 1047cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 1048a465419bSEric Dumazet sk_nocaps_add(sk, NETIF_F_GSO_MASK); 1049bd0388aeSWilliam Allen Simpson tp->af_specific->calc_md5_hash(opts.hash_location, 105049a72dfbSAdam Langley md5, sk, NULL, skb); 1051cfb6eeb4SYOSHIFUJI Hideaki } 1052cfb6eeb4SYOSHIFUJI Hideaki #endif 1053cfb6eeb4SYOSHIFUJI Hideaki 1054bb296246SHerbert Xu icsk->icsk_af_ops->send_check(sk, skb); 10551da177e4SLinus Torvalds 10564de075e0SEric Dumazet if (likely(tcb->tcp_flags & TCPHDR_ACK)) 1057fc6415bcSDavid S. Miller tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); 10581da177e4SLinus Torvalds 10591da177e4SLinus Torvalds if (skb->len != tcp_header_size) 1060cf533ea5SEric Dumazet tcp_event_data_sent(tp, sk); 10611da177e4SLinus Torvalds 1062bd37a088SWei Yongjun if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) 1063aa2ea058STom Herbert TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, 1064aa2ea058STom Herbert tcp_skb_pcount(skb)); 10651da177e4SLinus Torvalds 1066d9d8da80SDavid S. Miller err = icsk->icsk_af_ops->queue_xmit(skb, &inet->cork.fl); 106783de47cdSHua Zhong if (likely(err <= 0)) 10681da177e4SLinus Torvalds return err; 10691da177e4SLinus Torvalds 10703cfe3baaSIlpo Järvinen tcp_enter_cwr(sk, 1); 10711da177e4SLinus Torvalds 1072b9df3cb8SGerrit Renker return net_xmit_eval(err); 10731da177e4SLinus Torvalds } 10741da177e4SLinus Torvalds 107567edfef7SAndi Kleen /* This routine just queues the buffer for sending. 10761da177e4SLinus Torvalds * 10771da177e4SLinus Torvalds * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, 10781da177e4SLinus Torvalds * otherwise socket can stall. 10791da177e4SLinus Torvalds */ 10801da177e4SLinus Torvalds static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) 10811da177e4SLinus Torvalds { 10821da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 10831da177e4SLinus Torvalds 10841da177e4SLinus Torvalds /* Advance write_seq and place onto the write_queue. */ 10851da177e4SLinus Torvalds tp->write_seq = TCP_SKB_CB(skb)->end_seq; 10861da177e4SLinus Torvalds skb_header_release(skb); 1087fe067e8aSDavid S. Miller tcp_add_write_queue_tail(sk, skb); 10883ab224beSHideo Aoki sk->sk_wmem_queued += skb->truesize; 10893ab224beSHideo Aoki sk_mem_charge(sk, skb->truesize); 10901da177e4SLinus Torvalds } 10911da177e4SLinus Torvalds 109267edfef7SAndi Kleen /* Initialize TSO segments for a packet. */ 1093cf533ea5SEric Dumazet static void tcp_set_skb_tso_segs(const struct sock *sk, struct sk_buff *skb, 1094056834d9SIlpo Järvinen unsigned int mss_now) 1095f6302d1dSDavid S. Miller { 10968e5b9ddaSHerbert Xu if (skb->len <= mss_now || !sk_can_gso(sk) || 10978e5b9ddaSHerbert Xu skb->ip_summed == CHECKSUM_NONE) { 1098f6302d1dSDavid S. Miller /* Avoid the costly divide in the normal 1099f6302d1dSDavid S. Miller * non-TSO case. 1100f6302d1dSDavid S. Miller */ 11017967168cSHerbert Xu skb_shinfo(skb)->gso_segs = 1; 11027967168cSHerbert Xu skb_shinfo(skb)->gso_size = 0; 11037967168cSHerbert Xu skb_shinfo(skb)->gso_type = 0; 1104f6302d1dSDavid S. Miller } else { 1105356f89e1SIlpo Järvinen skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now); 11067967168cSHerbert Xu skb_shinfo(skb)->gso_size = mss_now; 1107bcd76111SHerbert Xu skb_shinfo(skb)->gso_type = sk->sk_gso_type; 11081da177e4SLinus Torvalds } 11091da177e4SLinus Torvalds } 11101da177e4SLinus Torvalds 111191fed7a1SIlpo Järvinen /* When a modification to fackets out becomes necessary, we need to check 111268f8353bSIlpo Järvinen * skb is counted to fackets_out or not. 111391fed7a1SIlpo Järvinen */ 1114cf533ea5SEric Dumazet static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb, 111591fed7a1SIlpo Järvinen int decr) 111691fed7a1SIlpo Järvinen { 1117a47e5a98SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1118a47e5a98SIlpo Järvinen 1119dc86967bSIlpo Järvinen if (!tp->sacked_out || tcp_is_reno(tp)) 112091fed7a1SIlpo Järvinen return; 112191fed7a1SIlpo Järvinen 11226859d494SIlpo Järvinen if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq)) 112391fed7a1SIlpo Järvinen tp->fackets_out -= decr; 112491fed7a1SIlpo Järvinen } 112591fed7a1SIlpo Järvinen 1126797108d1SIlpo Järvinen /* Pcount in the middle of the write queue got changed, we need to do various 1127797108d1SIlpo Järvinen * tweaks to fix counters 1128797108d1SIlpo Järvinen */ 1129cf533ea5SEric Dumazet static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) 1130797108d1SIlpo Järvinen { 1131797108d1SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1132797108d1SIlpo Järvinen 1133797108d1SIlpo Järvinen tp->packets_out -= decr; 1134797108d1SIlpo Järvinen 1135797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 1136797108d1SIlpo Järvinen tp->sacked_out -= decr; 1137797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 1138797108d1SIlpo Järvinen tp->retrans_out -= decr; 1139797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 1140797108d1SIlpo Järvinen tp->lost_out -= decr; 1141797108d1SIlpo Järvinen 1142797108d1SIlpo Järvinen /* Reno case is special. Sigh... */ 1143797108d1SIlpo Järvinen if (tcp_is_reno(tp) && decr > 0) 1144797108d1SIlpo Järvinen tp->sacked_out -= min_t(u32, tp->sacked_out, decr); 1145797108d1SIlpo Järvinen 1146797108d1SIlpo Järvinen tcp_adjust_fackets_out(sk, skb, decr); 1147797108d1SIlpo Järvinen 1148797108d1SIlpo Järvinen if (tp->lost_skb_hint && 1149797108d1SIlpo Järvinen before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && 115052cf3cc8SIlpo Järvinen (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))) 1151797108d1SIlpo Järvinen tp->lost_cnt_hint -= decr; 1152797108d1SIlpo Järvinen 1153797108d1SIlpo Järvinen tcp_verify_left_out(tp); 1154797108d1SIlpo Järvinen } 1155797108d1SIlpo Järvinen 11561da177e4SLinus Torvalds /* Function to create two new TCP segments. Shrinks the given segment 11571da177e4SLinus Torvalds * to the specified size and appends a new segment with the rest of the 11581da177e4SLinus Torvalds * packet to the list. This won't be called frequently, I hope. 11591da177e4SLinus Torvalds * Remember, these are still headerless SKBs at this point. 11601da177e4SLinus Torvalds */ 1161056834d9SIlpo Järvinen int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, 1162056834d9SIlpo Järvinen unsigned int mss_now) 11631da177e4SLinus Torvalds { 11641da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 11651da177e4SLinus Torvalds struct sk_buff *buff; 11666475be16SDavid S. Miller int nsize, old_factor; 1167b60b49eaSHerbert Xu int nlen; 11689ce01461SIlpo Järvinen u8 flags; 11691da177e4SLinus Torvalds 11702fceec13SIlpo Järvinen if (WARN_ON(len > skb->len)) 11712fceec13SIlpo Järvinen return -EINVAL; 11726a438bbeSStephen Hemminger 11731da177e4SLinus Torvalds nsize = skb_headlen(skb) - len; 11741da177e4SLinus Torvalds if (nsize < 0) 11751da177e4SLinus Torvalds nsize = 0; 11761da177e4SLinus Torvalds 11771da177e4SLinus Torvalds if (skb_cloned(skb) && 11781da177e4SLinus Torvalds skb_is_nonlinear(skb) && 11791da177e4SLinus Torvalds pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) 11801da177e4SLinus Torvalds return -ENOMEM; 11811da177e4SLinus Torvalds 11821da177e4SLinus Torvalds /* Get a new skb... force flag on. */ 11831da177e4SLinus Torvalds buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC); 11841da177e4SLinus Torvalds if (buff == NULL) 11851da177e4SLinus Torvalds return -ENOMEM; /* We'll just try again later. */ 1186ef5cb973SHerbert Xu 11873ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 11883ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1189b60b49eaSHerbert Xu nlen = skb->len - len - nsize; 1190b60b49eaSHerbert Xu buff->truesize += nlen; 1191b60b49eaSHerbert Xu skb->truesize -= nlen; 11921da177e4SLinus Torvalds 11931da177e4SLinus Torvalds /* Correct the sequence numbers. */ 11941da177e4SLinus Torvalds TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 11951da177e4SLinus Torvalds TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 11961da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 11971da177e4SLinus Torvalds 11981da177e4SLinus Torvalds /* PSH and FIN should only be set in the second packet. */ 11994de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 12004de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 12014de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1202e14c3cafSHerbert Xu TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; 12031da177e4SLinus Torvalds 120484fa7933SPatrick McHardy if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { 12051da177e4SLinus Torvalds /* Copy and checksum data tail into the new buffer. */ 1206056834d9SIlpo Järvinen buff->csum = csum_partial_copy_nocheck(skb->data + len, 1207056834d9SIlpo Järvinen skb_put(buff, nsize), 12081da177e4SLinus Torvalds nsize, 0); 12091da177e4SLinus Torvalds 12101da177e4SLinus Torvalds skb_trim(skb, len); 12111da177e4SLinus Torvalds 12121da177e4SLinus Torvalds skb->csum = csum_block_sub(skb->csum, buff->csum, len); 12131da177e4SLinus Torvalds } else { 121484fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 12151da177e4SLinus Torvalds skb_split(skb, buff, len); 12161da177e4SLinus Torvalds } 12171da177e4SLinus Torvalds 12181da177e4SLinus Torvalds buff->ip_summed = skb->ip_summed; 12191da177e4SLinus Torvalds 12201da177e4SLinus Torvalds /* Looks stupid, but our code really uses when of 12211da177e4SLinus Torvalds * skbs, which it never sent before. --ANK 12221da177e4SLinus Torvalds */ 12231da177e4SLinus Torvalds TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when; 1224a61bbcf2SPatrick McHardy buff->tstamp = skb->tstamp; 12251da177e4SLinus Torvalds 12266475be16SDavid S. Miller old_factor = tcp_skb_pcount(skb); 12276475be16SDavid S. Miller 12281da177e4SLinus Torvalds /* Fix up tso_factor for both original and new SKB. */ 1229846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1230846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, buff, mss_now); 12311da177e4SLinus Torvalds 12326475be16SDavid S. Miller /* If this packet has been sent out already, we must 12336475be16SDavid S. Miller * adjust the various packet counters. 12346475be16SDavid S. Miller */ 1235cf0b450cSHerbert Xu if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 12366475be16SDavid S. Miller int diff = old_factor - tcp_skb_pcount(skb) - 12376475be16SDavid S. Miller tcp_skb_pcount(buff); 12381da177e4SLinus Torvalds 1239797108d1SIlpo Järvinen if (diff) 1240797108d1SIlpo Järvinen tcp_adjust_pcount(sk, skb, diff); 12411da177e4SLinus Torvalds } 12421da177e4SLinus Torvalds 12431da177e4SLinus Torvalds /* Link BUFF into the send queue. */ 1244f44b5271SDavid S. Miller skb_header_release(buff); 1245fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 12461da177e4SLinus Torvalds 12471da177e4SLinus Torvalds return 0; 12481da177e4SLinus Torvalds } 12491da177e4SLinus Torvalds 12501da177e4SLinus Torvalds /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c 12511da177e4SLinus Torvalds * eventually). The difference is that pulled data not copied, but 12521da177e4SLinus Torvalds * immediately discarded. 12531da177e4SLinus Torvalds */ 1254f2911969SHerbert Xu ~{PmVHI~} static void __pskb_trim_head(struct sk_buff *skb, int len) 12551da177e4SLinus Torvalds { 12561da177e4SLinus Torvalds int i, k, eat; 12571da177e4SLinus Torvalds 12584fa48bf3SEric Dumazet eat = min_t(int, len, skb_headlen(skb)); 12594fa48bf3SEric Dumazet if (eat) { 12604fa48bf3SEric Dumazet __skb_pull(skb, eat); 126122b4a4f2SEric Dumazet skb->avail_size -= eat; 12624fa48bf3SEric Dumazet len -= eat; 12634fa48bf3SEric Dumazet if (!len) 12644fa48bf3SEric Dumazet return; 12654fa48bf3SEric Dumazet } 12661da177e4SLinus Torvalds eat = len; 12671da177e4SLinus Torvalds k = 0; 12681da177e4SLinus Torvalds for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) { 12699e903e08SEric Dumazet int size = skb_frag_size(&skb_shinfo(skb)->frags[i]); 12709e903e08SEric Dumazet 12719e903e08SEric Dumazet if (size <= eat) { 1272aff65da0SIan Campbell skb_frag_unref(skb, i); 12739e903e08SEric Dumazet eat -= size; 12741da177e4SLinus Torvalds } else { 12751da177e4SLinus Torvalds skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i]; 12761da177e4SLinus Torvalds if (eat) { 12771da177e4SLinus Torvalds skb_shinfo(skb)->frags[k].page_offset += eat; 12789e903e08SEric Dumazet skb_frag_size_sub(&skb_shinfo(skb)->frags[k], eat); 12791da177e4SLinus Torvalds eat = 0; 12801da177e4SLinus Torvalds } 12811da177e4SLinus Torvalds k++; 12821da177e4SLinus Torvalds } 12831da177e4SLinus Torvalds } 12841da177e4SLinus Torvalds skb_shinfo(skb)->nr_frags = k; 12851da177e4SLinus Torvalds 128627a884dcSArnaldo Carvalho de Melo skb_reset_tail_pointer(skb); 12871da177e4SLinus Torvalds skb->data_len -= len; 12881da177e4SLinus Torvalds skb->len = skb->data_len; 12891da177e4SLinus Torvalds } 12901da177e4SLinus Torvalds 129167edfef7SAndi Kleen /* Remove acked data from a packet in the transmit queue. */ 12921da177e4SLinus Torvalds int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 12931da177e4SLinus Torvalds { 1294056834d9SIlpo Järvinen if (skb_cloned(skb) && pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) 12951da177e4SLinus Torvalds return -ENOMEM; 12961da177e4SLinus Torvalds 12974fa48bf3SEric Dumazet __pskb_trim_head(skb, len); 12981da177e4SLinus Torvalds 12991da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq += len; 130084fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 13011da177e4SLinus Torvalds 13021da177e4SLinus Torvalds skb->truesize -= len; 13031da177e4SLinus Torvalds sk->sk_wmem_queued -= len; 13043ab224beSHideo Aoki sk_mem_uncharge(sk, len); 13051da177e4SLinus Torvalds sock_set_flag(sk, SOCK_QUEUE_SHRUNK); 13061da177e4SLinus Torvalds 13075b35e1e6SNeal Cardwell /* Any change of skb->len requires recalculation of tso factor. */ 13081da177e4SLinus Torvalds if (tcp_skb_pcount(skb) > 1) 13095b35e1e6SNeal Cardwell tcp_set_skb_tso_segs(sk, skb, tcp_skb_mss(skb)); 13101da177e4SLinus Torvalds 13111da177e4SLinus Torvalds return 0; 13121da177e4SLinus Torvalds } 13131da177e4SLinus Torvalds 131467edfef7SAndi Kleen /* Calculate MSS. Not accounting for SACKs here. */ 131567469601SEric Dumazet int tcp_mtu_to_mss(struct sock *sk, int pmtu) 13165d424d5aSJohn Heffner { 1317cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1318cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 13195d424d5aSJohn Heffner int mss_now; 13205d424d5aSJohn Heffner 13215d424d5aSJohn Heffner /* Calculate base mss without TCP options: 13225d424d5aSJohn Heffner It is MMS_S - sizeof(tcphdr) of rfc1122 13235d424d5aSJohn Heffner */ 13245d424d5aSJohn Heffner mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 13255d424d5aSJohn Heffner 132667469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 132767469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 132867469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 132967469601SEric Dumazet 133067469601SEric Dumazet if (dst && dst_allfrag(dst)) 133167469601SEric Dumazet mss_now -= icsk->icsk_af_ops->net_frag_header_len; 133267469601SEric Dumazet } 133367469601SEric Dumazet 13345d424d5aSJohn Heffner /* Clamp it (mss_clamp does not include tcp options) */ 13355d424d5aSJohn Heffner if (mss_now > tp->rx_opt.mss_clamp) 13365d424d5aSJohn Heffner mss_now = tp->rx_opt.mss_clamp; 13375d424d5aSJohn Heffner 13385d424d5aSJohn Heffner /* Now subtract optional transport overhead */ 13395d424d5aSJohn Heffner mss_now -= icsk->icsk_ext_hdr_len; 13405d424d5aSJohn Heffner 13415d424d5aSJohn Heffner /* Then reserve room for full set of TCP options and 8 bytes of data */ 13425d424d5aSJohn Heffner if (mss_now < 48) 13435d424d5aSJohn Heffner mss_now = 48; 13445d424d5aSJohn Heffner 13455d424d5aSJohn Heffner /* Now subtract TCP options size, not including SACKs */ 13465d424d5aSJohn Heffner mss_now -= tp->tcp_header_len - sizeof(struct tcphdr); 13475d424d5aSJohn Heffner 13485d424d5aSJohn Heffner return mss_now; 13495d424d5aSJohn Heffner } 13505d424d5aSJohn Heffner 13515d424d5aSJohn Heffner /* Inverse of above */ 135267469601SEric Dumazet int tcp_mss_to_mtu(struct sock *sk, int mss) 13535d424d5aSJohn Heffner { 1354cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1355cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 13565d424d5aSJohn Heffner int mtu; 13575d424d5aSJohn Heffner 13585d424d5aSJohn Heffner mtu = mss + 13595d424d5aSJohn Heffner tp->tcp_header_len + 13605d424d5aSJohn Heffner icsk->icsk_ext_hdr_len + 13615d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 13625d424d5aSJohn Heffner 136367469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 136467469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 136567469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 136667469601SEric Dumazet 136767469601SEric Dumazet if (dst && dst_allfrag(dst)) 136867469601SEric Dumazet mtu += icsk->icsk_af_ops->net_frag_header_len; 136967469601SEric Dumazet } 13705d424d5aSJohn Heffner return mtu; 13715d424d5aSJohn Heffner } 13725d424d5aSJohn Heffner 137367edfef7SAndi Kleen /* MTU probing init per socket */ 13745d424d5aSJohn Heffner void tcp_mtup_init(struct sock *sk) 13755d424d5aSJohn Heffner { 13765d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 13775d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 13785d424d5aSJohn Heffner 13795d424d5aSJohn Heffner icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1; 13805d424d5aSJohn Heffner icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 13815d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 13825d424d5aSJohn Heffner icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss); 13835d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 13845d424d5aSJohn Heffner } 13854bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_mtup_init); 13865d424d5aSJohn Heffner 13871da177e4SLinus Torvalds /* This function synchronize snd mss to current pmtu/exthdr set. 13881da177e4SLinus Torvalds 13891da177e4SLinus Torvalds tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 13901da177e4SLinus Torvalds for TCP options, but includes only bare TCP header. 13911da177e4SLinus Torvalds 13921da177e4SLinus Torvalds tp->rx_opt.mss_clamp is mss negotiated at connection setup. 1393caa20d9aSStephen Hemminger It is minimum of user_mss and mss received with SYN. 13941da177e4SLinus Torvalds It also does not include TCP options. 13951da177e4SLinus Torvalds 1396d83d8461SArnaldo Carvalho de Melo inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. 13971da177e4SLinus Torvalds 13981da177e4SLinus Torvalds tp->mss_cache is current effective sending mss, including 13991da177e4SLinus Torvalds all tcp options except for SACKs. It is evaluated, 14001da177e4SLinus Torvalds taking into account current pmtu, but never exceeds 14011da177e4SLinus Torvalds tp->rx_opt.mss_clamp. 14021da177e4SLinus Torvalds 14031da177e4SLinus Torvalds NOTE1. rfc1122 clearly states that advertised MSS 14041da177e4SLinus Torvalds DOES NOT include either tcp or ip options. 14051da177e4SLinus Torvalds 1406d83d8461SArnaldo Carvalho de Melo NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache 1407d83d8461SArnaldo Carvalho de Melo are READ ONLY outside this function. --ANK (980731) 14081da177e4SLinus Torvalds */ 14091da177e4SLinus Torvalds unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 14101da177e4SLinus Torvalds { 14111da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1412d83d8461SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 14135d424d5aSJohn Heffner int mss_now; 14141da177e4SLinus Torvalds 14155d424d5aSJohn Heffner if (icsk->icsk_mtup.search_high > pmtu) 14165d424d5aSJohn Heffner icsk->icsk_mtup.search_high = pmtu; 14171da177e4SLinus Torvalds 14185d424d5aSJohn Heffner mss_now = tcp_mtu_to_mss(sk, pmtu); 1419409d22b4SIlpo Järvinen mss_now = tcp_bound_to_half_wnd(tp, mss_now); 14201da177e4SLinus Torvalds 14211da177e4SLinus Torvalds /* And store cached results */ 1422d83d8461SArnaldo Carvalho de Melo icsk->icsk_pmtu_cookie = pmtu; 14235d424d5aSJohn Heffner if (icsk->icsk_mtup.enabled) 14245d424d5aSJohn Heffner mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); 1425c1b4a7e6SDavid S. Miller tp->mss_cache = mss_now; 14261da177e4SLinus Torvalds 14271da177e4SLinus Torvalds return mss_now; 14281da177e4SLinus Torvalds } 14294bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_sync_mss); 14301da177e4SLinus Torvalds 14311da177e4SLinus Torvalds /* Compute the current effective MSS, taking SACKs and IP options, 14321da177e4SLinus Torvalds * and even PMTU discovery events into account. 14331da177e4SLinus Torvalds */ 14340c54b85fSIlpo Järvinen unsigned int tcp_current_mss(struct sock *sk) 14351da177e4SLinus Torvalds { 1436cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1437cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1438c1b4a7e6SDavid S. Miller u32 mss_now; 143995c96174SEric Dumazet unsigned int header_len; 144033ad798cSAdam Langley struct tcp_out_options opts; 144133ad798cSAdam Langley struct tcp_md5sig_key *md5; 14421da177e4SLinus Torvalds 1443c1b4a7e6SDavid S. Miller mss_now = tp->mss_cache; 1444c1b4a7e6SDavid S. Miller 14451da177e4SLinus Torvalds if (dst) { 14461da177e4SLinus Torvalds u32 mtu = dst_mtu(dst); 1447d83d8461SArnaldo Carvalho de Melo if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 14481da177e4SLinus Torvalds mss_now = tcp_sync_mss(sk, mtu); 14491da177e4SLinus Torvalds } 14501da177e4SLinus Torvalds 145133ad798cSAdam Langley header_len = tcp_established_options(sk, NULL, &opts, &md5) + 145233ad798cSAdam Langley sizeof(struct tcphdr); 145333ad798cSAdam Langley /* The mss_cache is sized based on tp->tcp_header_len, which assumes 145433ad798cSAdam Langley * some common options. If this is an odd packet (because we have SACK 145533ad798cSAdam Langley * blocks etc) then our calculated header_len will be different, and 145633ad798cSAdam Langley * we have to adjust mss_now correspondingly */ 145733ad798cSAdam Langley if (header_len != tp->tcp_header_len) { 145833ad798cSAdam Langley int delta = (int) header_len - tp->tcp_header_len; 145933ad798cSAdam Langley mss_now -= delta; 146033ad798cSAdam Langley } 1461cfb6eeb4SYOSHIFUJI Hideaki 14621da177e4SLinus Torvalds return mss_now; 14631da177e4SLinus Torvalds } 14641da177e4SLinus Torvalds 1465a762a980SDavid S. Miller /* Congestion window validation. (RFC2861) */ 14669e412ba7SIlpo Järvinen static void tcp_cwnd_validate(struct sock *sk) 1467a762a980SDavid S. Miller { 14689e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1469a762a980SDavid S. Miller 1470d436d686SIlpo Järvinen if (tp->packets_out >= tp->snd_cwnd) { 1471a762a980SDavid S. Miller /* Network is feed fully. */ 1472a762a980SDavid S. Miller tp->snd_cwnd_used = 0; 1473a762a980SDavid S. Miller tp->snd_cwnd_stamp = tcp_time_stamp; 1474a762a980SDavid S. Miller } else { 1475a762a980SDavid S. Miller /* Network starves. */ 1476a762a980SDavid S. Miller if (tp->packets_out > tp->snd_cwnd_used) 1477a762a980SDavid S. Miller tp->snd_cwnd_used = tp->packets_out; 1478a762a980SDavid S. Miller 147915d33c07SDavid S. Miller if (sysctl_tcp_slow_start_after_idle && 148015d33c07SDavid S. Miller (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto) 1481a762a980SDavid S. Miller tcp_cwnd_application_limited(sk); 1482a762a980SDavid S. Miller } 1483a762a980SDavid S. Miller } 1484a762a980SDavid S. Miller 14850e3a4803SIlpo Järvinen /* Returns the portion of skb which can be sent right away without 14860e3a4803SIlpo Järvinen * introducing MSS oddities to segment boundaries. In rare cases where 14870e3a4803SIlpo Järvinen * mss_now != mss_cache, we will request caller to create a small skb 14880e3a4803SIlpo Järvinen * per input skb which could be mostly avoided here (if desired). 14895ea3a748SIlpo Järvinen * 14905ea3a748SIlpo Järvinen * We explicitly want to create a request for splitting write queue tail 14915ea3a748SIlpo Järvinen * to a small skb for Nagle purposes while avoiding unnecessary modulos, 14925ea3a748SIlpo Järvinen * thus all the complexity (cwnd_len is always MSS multiple which we 14935ea3a748SIlpo Järvinen * return whenever allowed by the other factors). Basically we need the 14945ea3a748SIlpo Järvinen * modulo only when the receiver window alone is the limiting factor or 14955ea3a748SIlpo Järvinen * when we would be allowed to send the split-due-to-Nagle skb fully. 14960e3a4803SIlpo Järvinen */ 1497cf533ea5SEric Dumazet static unsigned int tcp_mss_split_point(const struct sock *sk, const struct sk_buff *skb, 1498056834d9SIlpo Järvinen unsigned int mss_now, unsigned int cwnd) 1499c1b4a7e6SDavid S. Miller { 1500cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 15010e3a4803SIlpo Järvinen u32 needed, window, cwnd_len; 1502c1b4a7e6SDavid S. Miller 150390840defSIlpo Järvinen window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1504c1b4a7e6SDavid S. Miller cwnd_len = mss_now * cwnd; 15050e3a4803SIlpo Järvinen 15060e3a4803SIlpo Järvinen if (likely(cwnd_len <= window && skb != tcp_write_queue_tail(sk))) 15070e3a4803SIlpo Järvinen return cwnd_len; 15080e3a4803SIlpo Järvinen 15095ea3a748SIlpo Järvinen needed = min(skb->len, window); 15105ea3a748SIlpo Järvinen 151117515408SIlpo Järvinen if (cwnd_len <= needed) 15120e3a4803SIlpo Järvinen return cwnd_len; 15130e3a4803SIlpo Järvinen 15140e3a4803SIlpo Järvinen return needed - needed % mss_now; 1515c1b4a7e6SDavid S. Miller } 1516c1b4a7e6SDavid S. Miller 1517c1b4a7e6SDavid S. Miller /* Can at least one segment of SKB be sent right now, according to the 1518c1b4a7e6SDavid S. Miller * congestion window rules? If so, return how many segments are allowed. 1519c1b4a7e6SDavid S. Miller */ 1520cf533ea5SEric Dumazet static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, 1521cf533ea5SEric Dumazet const struct sk_buff *skb) 1522c1b4a7e6SDavid S. Miller { 1523c1b4a7e6SDavid S. Miller u32 in_flight, cwnd; 1524c1b4a7e6SDavid S. Miller 1525c1b4a7e6SDavid S. Miller /* Don't be strict about the congestion window for the final FIN. */ 15264de075e0SEric Dumazet if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && 15274de075e0SEric Dumazet tcp_skb_pcount(skb) == 1) 1528c1b4a7e6SDavid S. Miller return 1; 1529c1b4a7e6SDavid S. Miller 1530c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1531c1b4a7e6SDavid S. Miller cwnd = tp->snd_cwnd; 1532c1b4a7e6SDavid S. Miller if (in_flight < cwnd) 1533c1b4a7e6SDavid S. Miller return (cwnd - in_flight); 1534c1b4a7e6SDavid S. Miller 1535c1b4a7e6SDavid S. Miller return 0; 1536c1b4a7e6SDavid S. Miller } 1537c1b4a7e6SDavid S. Miller 1538b595076aSUwe Kleine-König /* Initialize TSO state of a skb. 153967edfef7SAndi Kleen * This must be invoked the first time we consider transmitting 1540c1b4a7e6SDavid S. Miller * SKB onto the wire. 1541c1b4a7e6SDavid S. Miller */ 1542cf533ea5SEric Dumazet static int tcp_init_tso_segs(const struct sock *sk, struct sk_buff *skb, 1543056834d9SIlpo Järvinen unsigned int mss_now) 1544c1b4a7e6SDavid S. Miller { 1545c1b4a7e6SDavid S. Miller int tso_segs = tcp_skb_pcount(skb); 1546c1b4a7e6SDavid S. Miller 1547f8269a49SIlpo Järvinen if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { 1548846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1549c1b4a7e6SDavid S. Miller tso_segs = tcp_skb_pcount(skb); 1550c1b4a7e6SDavid S. Miller } 1551c1b4a7e6SDavid S. Miller return tso_segs; 1552c1b4a7e6SDavid S. Miller } 1553c1b4a7e6SDavid S. Miller 155467edfef7SAndi Kleen /* Minshall's variant of the Nagle send check. */ 1555a2a385d6SEric Dumazet static inline bool tcp_minshall_check(const struct tcp_sock *tp) 1556c1b4a7e6SDavid S. Miller { 1557c1b4a7e6SDavid S. Miller return after(tp->snd_sml, tp->snd_una) && 1558c1b4a7e6SDavid S. Miller !after(tp->snd_sml, tp->snd_nxt); 1559c1b4a7e6SDavid S. Miller } 1560c1b4a7e6SDavid S. Miller 1561a2a385d6SEric Dumazet /* Return false, if packet can be sent now without violation Nagle's rules: 1562c1b4a7e6SDavid S. Miller * 1. It is full sized. 1563c1b4a7e6SDavid S. Miller * 2. Or it contains FIN. (already checked by caller) 15646d67e9beSFeng King * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. 1565c1b4a7e6SDavid S. Miller * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 1566c1b4a7e6SDavid S. Miller * With Minshall's modification: all sent small packets are ACKed. 1567c1b4a7e6SDavid S. Miller */ 1568a2a385d6SEric Dumazet static inline bool tcp_nagle_check(const struct tcp_sock *tp, 1569c1b4a7e6SDavid S. Miller const struct sk_buff *skb, 157095c96174SEric Dumazet unsigned int mss_now, int nonagle) 1571c1b4a7e6SDavid S. Miller { 1572a02cec21SEric Dumazet return skb->len < mss_now && 1573c1b4a7e6SDavid S. Miller ((nonagle & TCP_NAGLE_CORK) || 1574a02cec21SEric Dumazet (!nonagle && tp->packets_out && tcp_minshall_check(tp))); 1575c1b4a7e6SDavid S. Miller } 1576c1b4a7e6SDavid S. Miller 1577a2a385d6SEric Dumazet /* Return true if the Nagle test allows this packet to be 1578c1b4a7e6SDavid S. Miller * sent now. 1579c1b4a7e6SDavid S. Miller */ 1580a2a385d6SEric Dumazet static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, 1581c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1582c1b4a7e6SDavid S. Miller { 1583c1b4a7e6SDavid S. Miller /* Nagle rule does not apply to frames, which sit in the middle of the 1584c1b4a7e6SDavid S. Miller * write_queue (they have no chances to get new data). 1585c1b4a7e6SDavid S. Miller * 1586c1b4a7e6SDavid S. Miller * This is implemented in the callers, where they modify the 'nonagle' 1587c1b4a7e6SDavid S. Miller * argument based upon the location of SKB in the send queue. 1588c1b4a7e6SDavid S. Miller */ 1589c1b4a7e6SDavid S. Miller if (nonagle & TCP_NAGLE_PUSH) 1590a2a385d6SEric Dumazet return true; 1591c1b4a7e6SDavid S. Miller 1592d551e454SIlpo Järvinen /* Don't use the nagle rule for urgent data (or for the final FIN). 1593d551e454SIlpo Järvinen * Nagle can be ignored during F-RTO too (see RFC4138). 1594d551e454SIlpo Järvinen */ 159533f5f57eSIlpo Järvinen if (tcp_urg_mode(tp) || (tp->frto_counter == 2) || 15964de075e0SEric Dumazet (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) 1597a2a385d6SEric Dumazet return true; 1598c1b4a7e6SDavid S. Miller 1599c1b4a7e6SDavid S. Miller if (!tcp_nagle_check(tp, skb, cur_mss, nonagle)) 1600a2a385d6SEric Dumazet return true; 1601c1b4a7e6SDavid S. Miller 1602a2a385d6SEric Dumazet return false; 1603c1b4a7e6SDavid S. Miller } 1604c1b4a7e6SDavid S. Miller 1605c1b4a7e6SDavid S. Miller /* Does at least the first segment of SKB fit into the send window? */ 1606a2a385d6SEric Dumazet static bool tcp_snd_wnd_test(const struct tcp_sock *tp, 1607a2a385d6SEric Dumazet const struct sk_buff *skb, 1608056834d9SIlpo Järvinen unsigned int cur_mss) 1609c1b4a7e6SDavid S. Miller { 1610c1b4a7e6SDavid S. Miller u32 end_seq = TCP_SKB_CB(skb)->end_seq; 1611c1b4a7e6SDavid S. Miller 1612c1b4a7e6SDavid S. Miller if (skb->len > cur_mss) 1613c1b4a7e6SDavid S. Miller end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 1614c1b4a7e6SDavid S. Miller 161590840defSIlpo Järvinen return !after(end_seq, tcp_wnd_end(tp)); 1616c1b4a7e6SDavid S. Miller } 1617c1b4a7e6SDavid S. Miller 1618fe067e8aSDavid S. Miller /* This checks if the data bearing packet SKB (usually tcp_send_head(sk)) 1619c1b4a7e6SDavid S. Miller * should be put on the wire right now. If so, it returns the number of 1620c1b4a7e6SDavid S. Miller * packets allowed by the congestion window. 1621c1b4a7e6SDavid S. Miller */ 1622cf533ea5SEric Dumazet static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb, 1623c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1624c1b4a7e6SDavid S. Miller { 1625cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1626c1b4a7e6SDavid S. Miller unsigned int cwnd_quota; 1627c1b4a7e6SDavid S. Miller 1628846998aeSDavid S. Miller tcp_init_tso_segs(sk, skb, cur_mss); 1629c1b4a7e6SDavid S. Miller 1630c1b4a7e6SDavid S. Miller if (!tcp_nagle_test(tp, skb, cur_mss, nonagle)) 1631c1b4a7e6SDavid S. Miller return 0; 1632c1b4a7e6SDavid S. Miller 1633c1b4a7e6SDavid S. Miller cwnd_quota = tcp_cwnd_test(tp, skb); 1634056834d9SIlpo Järvinen if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss)) 1635c1b4a7e6SDavid S. Miller cwnd_quota = 0; 1636c1b4a7e6SDavid S. Miller 1637c1b4a7e6SDavid S. Miller return cwnd_quota; 1638c1b4a7e6SDavid S. Miller } 1639c1b4a7e6SDavid S. Miller 164067edfef7SAndi Kleen /* Test if sending is allowed right now. */ 1641a2a385d6SEric Dumazet bool tcp_may_send_now(struct sock *sk) 1642c1b4a7e6SDavid S. Miller { 1643cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1644fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 1645c1b4a7e6SDavid S. Miller 1646a02cec21SEric Dumazet return skb && 16470c54b85fSIlpo Järvinen tcp_snd_test(sk, skb, tcp_current_mss(sk), 1648c1b4a7e6SDavid S. Miller (tcp_skb_is_last(sk, skb) ? 1649a02cec21SEric Dumazet tp->nonagle : TCP_NAGLE_PUSH)); 1650c1b4a7e6SDavid S. Miller } 1651c1b4a7e6SDavid S. Miller 1652c1b4a7e6SDavid S. Miller /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 1653c1b4a7e6SDavid S. Miller * which is put after SKB on the list. It is very much like 1654c1b4a7e6SDavid S. Miller * tcp_fragment() except that it may make several kinds of assumptions 1655c1b4a7e6SDavid S. Miller * in order to speed up the splitting operation. In particular, we 1656c1b4a7e6SDavid S. Miller * know that all the data is in scatter-gather pages, and that the 1657c1b4a7e6SDavid S. Miller * packet has never been sent out before (and thus is not cloned). 1658c1b4a7e6SDavid S. Miller */ 1659056834d9SIlpo Järvinen static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, 1660c4ead4c5SEric Dumazet unsigned int mss_now, gfp_t gfp) 1661c1b4a7e6SDavid S. Miller { 1662c1b4a7e6SDavid S. Miller struct sk_buff *buff; 1663c1b4a7e6SDavid S. Miller int nlen = skb->len - len; 16649ce01461SIlpo Järvinen u8 flags; 1665c1b4a7e6SDavid S. Miller 1666c1b4a7e6SDavid S. Miller /* All of a TSO frame must be composed of paged data. */ 1667c8ac3774SHerbert Xu if (skb->len != skb->data_len) 1668c8ac3774SHerbert Xu return tcp_fragment(sk, skb, len, mss_now); 1669c1b4a7e6SDavid S. Miller 1670c4ead4c5SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, gfp); 1671c1b4a7e6SDavid S. Miller if (unlikely(buff == NULL)) 1672c1b4a7e6SDavid S. Miller return -ENOMEM; 1673c1b4a7e6SDavid S. Miller 16743ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 16753ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1676b60b49eaSHerbert Xu buff->truesize += nlen; 1677c1b4a7e6SDavid S. Miller skb->truesize -= nlen; 1678c1b4a7e6SDavid S. Miller 1679c1b4a7e6SDavid S. Miller /* Correct the sequence numbers. */ 1680c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 1681c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 1682c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 1683c1b4a7e6SDavid S. Miller 1684c1b4a7e6SDavid S. Miller /* PSH and FIN should only be set in the second packet. */ 16854de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 16864de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 16874de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1688c1b4a7e6SDavid S. Miller 1689c1b4a7e6SDavid S. Miller /* This packet was never sent out yet, so no SACK bits. */ 1690c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->sacked = 0; 1691c1b4a7e6SDavid S. Miller 169284fa7933SPatrick McHardy buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; 1693c1b4a7e6SDavid S. Miller skb_split(skb, buff, len); 1694c1b4a7e6SDavid S. Miller 1695c1b4a7e6SDavid S. Miller /* Fix up tso_factor for both original and new SKB. */ 1696846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1697846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, buff, mss_now); 1698c1b4a7e6SDavid S. Miller 1699c1b4a7e6SDavid S. Miller /* Link BUFF into the send queue. */ 1700c1b4a7e6SDavid S. Miller skb_header_release(buff); 1701fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 1702c1b4a7e6SDavid S. Miller 1703c1b4a7e6SDavid S. Miller return 0; 1704c1b4a7e6SDavid S. Miller } 1705c1b4a7e6SDavid S. Miller 1706c1b4a7e6SDavid S. Miller /* Try to defer sending, if possible, in order to minimize the amount 1707c1b4a7e6SDavid S. Miller * of TSO splitting we do. View it as a kind of TSO Nagle test. 1708c1b4a7e6SDavid S. Miller * 1709c1b4a7e6SDavid S. Miller * This algorithm is from John Heffner. 1710c1b4a7e6SDavid S. Miller */ 1711a2a385d6SEric Dumazet static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb) 1712c1b4a7e6SDavid S. Miller { 17139e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 17146687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 1715c1b4a7e6SDavid S. Miller u32 send_win, cong_win, limit, in_flight; 1716ad9f4f50SEric Dumazet int win_divisor; 1717c1b4a7e6SDavid S. Miller 17184de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 1719ae8064acSJohn Heffner goto send_now; 1720c1b4a7e6SDavid S. Miller 17216687e988SArnaldo Carvalho de Melo if (icsk->icsk_ca_state != TCP_CA_Open) 1722ae8064acSJohn Heffner goto send_now; 1723ae8064acSJohn Heffner 1724ae8064acSJohn Heffner /* Defer for less than two clock ticks. */ 1725bd515c3eSIlpo Järvinen if (tp->tso_deferred && 1726a2acde07SIlpo Järvinen (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1) 1727ae8064acSJohn Heffner goto send_now; 1728908a75c1SDavid S. Miller 1729c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1730c1b4a7e6SDavid S. Miller 1731056834d9SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight)); 1732c1b4a7e6SDavid S. Miller 173390840defSIlpo Järvinen send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1734c1b4a7e6SDavid S. Miller 1735c1b4a7e6SDavid S. Miller /* From in_flight test above, we know that cwnd > in_flight. */ 1736c1b4a7e6SDavid S. Miller cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; 1737c1b4a7e6SDavid S. Miller 1738c1b4a7e6SDavid S. Miller limit = min(send_win, cong_win); 1739c1b4a7e6SDavid S. Miller 1740ba244fe9SDavid S. Miller /* If a full-sized TSO skb can be sent, do it. */ 174182cc1a7aSPeter P Waskiewicz Jr if (limit >= sk->sk_gso_max_size) 1742ae8064acSJohn Heffner goto send_now; 1743ba244fe9SDavid S. Miller 174462ad2761SIlpo Järvinen /* Middle in queue won't get any more data, full sendable already? */ 174562ad2761SIlpo Järvinen if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) 174662ad2761SIlpo Järvinen goto send_now; 174762ad2761SIlpo Järvinen 1748ad9f4f50SEric Dumazet win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor); 1749ad9f4f50SEric Dumazet if (win_divisor) { 1750c1b4a7e6SDavid S. Miller u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); 1751c1b4a7e6SDavid S. Miller 1752c1b4a7e6SDavid S. Miller /* If at least some fraction of a window is available, 1753c1b4a7e6SDavid S. Miller * just use it. 1754c1b4a7e6SDavid S. Miller */ 1755ad9f4f50SEric Dumazet chunk /= win_divisor; 1756c1b4a7e6SDavid S. Miller if (limit >= chunk) 1757ae8064acSJohn Heffner goto send_now; 1758c1b4a7e6SDavid S. Miller } else { 1759c1b4a7e6SDavid S. Miller /* Different approach, try not to defer past a single 1760c1b4a7e6SDavid S. Miller * ACK. Receiver should ACK every other full sized 1761c1b4a7e6SDavid S. Miller * frame, so if we have space for more than 3 frames 1762c1b4a7e6SDavid S. Miller * then send now. 1763c1b4a7e6SDavid S. Miller */ 17646b5a5c0dSNeal Cardwell if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) 1765ae8064acSJohn Heffner goto send_now; 1766c1b4a7e6SDavid S. Miller } 1767c1b4a7e6SDavid S. Miller 1768c1b4a7e6SDavid S. Miller /* Ok, it looks like it is advisable to defer. */ 1769ae8064acSJohn Heffner tp->tso_deferred = 1 | (jiffies << 1); 1770ae8064acSJohn Heffner 1771a2a385d6SEric Dumazet return true; 1772ae8064acSJohn Heffner 1773ae8064acSJohn Heffner send_now: 1774ae8064acSJohn Heffner tp->tso_deferred = 0; 1775a2a385d6SEric Dumazet return false; 1776c1b4a7e6SDavid S. Miller } 1777c1b4a7e6SDavid S. Miller 17785d424d5aSJohn Heffner /* Create a new MTU probe if we are ready. 177967edfef7SAndi Kleen * MTU probe is regularly attempting to increase the path MTU by 178067edfef7SAndi Kleen * deliberately sending larger packets. This discovers routing 178167edfef7SAndi Kleen * changes resulting in larger path MTUs. 178267edfef7SAndi Kleen * 17835d424d5aSJohn Heffner * Returns 0 if we should wait to probe (no cwnd available), 17845d424d5aSJohn Heffner * 1 if a probe was sent, 1785056834d9SIlpo Järvinen * -1 otherwise 1786056834d9SIlpo Järvinen */ 17875d424d5aSJohn Heffner static int tcp_mtu_probe(struct sock *sk) 17885d424d5aSJohn Heffner { 17895d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 17905d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 17915d424d5aSJohn Heffner struct sk_buff *skb, *nskb, *next; 17925d424d5aSJohn Heffner int len; 17935d424d5aSJohn Heffner int probe_size; 179491cc17c0SIlpo Järvinen int size_needed; 17955d424d5aSJohn Heffner int copy; 17965d424d5aSJohn Heffner int mss_now; 17975d424d5aSJohn Heffner 17985d424d5aSJohn Heffner /* Not currently probing/verifying, 17995d424d5aSJohn Heffner * not in recovery, 18005d424d5aSJohn Heffner * have enough cwnd, and 18015d424d5aSJohn Heffner * not SACKing (the variable headers throw things off) */ 18025d424d5aSJohn Heffner if (!icsk->icsk_mtup.enabled || 18035d424d5aSJohn Heffner icsk->icsk_mtup.probe_size || 18045d424d5aSJohn Heffner inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 18055d424d5aSJohn Heffner tp->snd_cwnd < 11 || 1806cabeccbdSIlpo Järvinen tp->rx_opt.num_sacks || tp->rx_opt.dsack) 18075d424d5aSJohn Heffner return -1; 18085d424d5aSJohn Heffner 18095d424d5aSJohn Heffner /* Very simple search strategy: just double the MSS. */ 18100c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 18115d424d5aSJohn Heffner probe_size = 2 * tp->mss_cache; 181291cc17c0SIlpo Järvinen size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; 18135d424d5aSJohn Heffner if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) { 18145d424d5aSJohn Heffner /* TODO: set timer for probe_converge_event */ 18155d424d5aSJohn Heffner return -1; 18165d424d5aSJohn Heffner } 18175d424d5aSJohn Heffner 18185d424d5aSJohn Heffner /* Have enough data in the send queue to probe? */ 18197f9c33e5SIlpo Järvinen if (tp->write_seq - tp->snd_nxt < size_needed) 18205d424d5aSJohn Heffner return -1; 18215d424d5aSJohn Heffner 182291cc17c0SIlpo Järvinen if (tp->snd_wnd < size_needed) 18235d424d5aSJohn Heffner return -1; 182490840defSIlpo Järvinen if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) 18255d424d5aSJohn Heffner return 0; 18265d424d5aSJohn Heffner 1827d67c58e9SIlpo Järvinen /* Do we need to wait to drain cwnd? With none in flight, don't stall */ 1828d67c58e9SIlpo Järvinen if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { 1829d67c58e9SIlpo Järvinen if (!tcp_packets_in_flight(tp)) 18305d424d5aSJohn Heffner return -1; 18315d424d5aSJohn Heffner else 18325d424d5aSJohn Heffner return 0; 18335d424d5aSJohn Heffner } 18345d424d5aSJohn Heffner 18355d424d5aSJohn Heffner /* We're allowed to probe. Build it now. */ 18365d424d5aSJohn Heffner if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL) 18375d424d5aSJohn Heffner return -1; 18383ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 18393ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 18405d424d5aSJohn Heffner 1841fe067e8aSDavid S. Miller skb = tcp_send_head(sk); 18425d424d5aSJohn Heffner 18435d424d5aSJohn Heffner TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 18445d424d5aSJohn Heffner TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 18454de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; 18465d424d5aSJohn Heffner TCP_SKB_CB(nskb)->sacked = 0; 18475d424d5aSJohn Heffner nskb->csum = 0; 184884fa7933SPatrick McHardy nskb->ip_summed = skb->ip_summed; 18495d424d5aSJohn Heffner 185050c4817eSIlpo Järvinen tcp_insert_write_queue_before(nskb, skb, sk); 185150c4817eSIlpo Järvinen 18525d424d5aSJohn Heffner len = 0; 1853234b6860SIlpo Järvinen tcp_for_write_queue_from_safe(skb, next, sk) { 18545d424d5aSJohn Heffner copy = min_t(int, skb->len, probe_size - len); 18555d424d5aSJohn Heffner if (nskb->ip_summed) 18565d424d5aSJohn Heffner skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); 18575d424d5aSJohn Heffner else 18585d424d5aSJohn Heffner nskb->csum = skb_copy_and_csum_bits(skb, 0, 1859056834d9SIlpo Järvinen skb_put(nskb, copy), 1860056834d9SIlpo Järvinen copy, nskb->csum); 18615d424d5aSJohn Heffner 18625d424d5aSJohn Heffner if (skb->len <= copy) { 18635d424d5aSJohn Heffner /* We've eaten all the data from this skb. 18645d424d5aSJohn Heffner * Throw it away. */ 18654de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 1866fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 18673ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 18685d424d5aSJohn Heffner } else { 18694de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & 1870a3433f35SChangli Gao ~(TCPHDR_FIN|TCPHDR_PSH); 18715d424d5aSJohn Heffner if (!skb_shinfo(skb)->nr_frags) { 18725d424d5aSJohn Heffner skb_pull(skb, copy); 187384fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 1874056834d9SIlpo Järvinen skb->csum = csum_partial(skb->data, 1875056834d9SIlpo Järvinen skb->len, 0); 18765d424d5aSJohn Heffner } else { 18775d424d5aSJohn Heffner __pskb_trim_head(skb, copy); 18785d424d5aSJohn Heffner tcp_set_skb_tso_segs(sk, skb, mss_now); 18795d424d5aSJohn Heffner } 18805d424d5aSJohn Heffner TCP_SKB_CB(skb)->seq += copy; 18815d424d5aSJohn Heffner } 18825d424d5aSJohn Heffner 18835d424d5aSJohn Heffner len += copy; 1884234b6860SIlpo Järvinen 1885234b6860SIlpo Järvinen if (len >= probe_size) 1886234b6860SIlpo Järvinen break; 18875d424d5aSJohn Heffner } 18885d424d5aSJohn Heffner tcp_init_tso_segs(sk, nskb, nskb->len); 18895d424d5aSJohn Heffner 18905d424d5aSJohn Heffner /* We're ready to send. If this fails, the probe will 18915d424d5aSJohn Heffner * be resegmented into mss-sized pieces by tcp_write_xmit(). */ 18925d424d5aSJohn Heffner TCP_SKB_CB(nskb)->when = tcp_time_stamp; 18935d424d5aSJohn Heffner if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 18945d424d5aSJohn Heffner /* Decrement cwnd here because we are sending 18955d424d5aSJohn Heffner * effectively two packets. */ 18965d424d5aSJohn Heffner tp->snd_cwnd--; 189766f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, nskb); 18985d424d5aSJohn Heffner 18995d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 19000e7b1368SJohn Heffner tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 19010e7b1368SJohn Heffner tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 19025d424d5aSJohn Heffner 19035d424d5aSJohn Heffner return 1; 19045d424d5aSJohn Heffner } 19055d424d5aSJohn Heffner 19065d424d5aSJohn Heffner return -1; 19075d424d5aSJohn Heffner } 19085d424d5aSJohn Heffner 19091da177e4SLinus Torvalds /* This routine writes packets to the network. It advances the 19101da177e4SLinus Torvalds * send_head. This happens as incoming acks open up the remote 19111da177e4SLinus Torvalds * window for us. 19121da177e4SLinus Torvalds * 1913f8269a49SIlpo Järvinen * LARGESEND note: !tcp_urg_mode is overkill, only frames between 1914f8269a49SIlpo Järvinen * snd_up-64k-mss .. snd_up cannot be large. However, taking into 1915f8269a49SIlpo Järvinen * account rare use of URG, this is not a big flaw. 1916f8269a49SIlpo Järvinen * 1917a2a385d6SEric Dumazet * Returns true, if no segments are in flight and we have queued segments, 1918a2a385d6SEric Dumazet * but cannot send anything now because of SWS or another problem. 19191da177e4SLinus Torvalds */ 1920a2a385d6SEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 1921d5dd9175SIlpo Järvinen int push_one, gfp_t gfp) 19221da177e4SLinus Torvalds { 19231da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 192492df7b51SDavid S. Miller struct sk_buff *skb; 1925c1b4a7e6SDavid S. Miller unsigned int tso_segs, sent_pkts; 1926c1b4a7e6SDavid S. Miller int cwnd_quota; 19275d424d5aSJohn Heffner int result; 19281da177e4SLinus Torvalds 1929c1b4a7e6SDavid S. Miller sent_pkts = 0; 19305d424d5aSJohn Heffner 1931d5dd9175SIlpo Järvinen if (!push_one) { 19325d424d5aSJohn Heffner /* Do MTU probing. */ 1933d5dd9175SIlpo Järvinen result = tcp_mtu_probe(sk); 1934d5dd9175SIlpo Järvinen if (!result) { 1935a2a385d6SEric Dumazet return false; 19365d424d5aSJohn Heffner } else if (result > 0) { 19375d424d5aSJohn Heffner sent_pkts = 1; 19385d424d5aSJohn Heffner } 1939d5dd9175SIlpo Järvinen } 19405d424d5aSJohn Heffner 1941fe067e8aSDavid S. Miller while ((skb = tcp_send_head(sk))) { 1942c8ac3774SHerbert Xu unsigned int limit; 1943c8ac3774SHerbert Xu 194446d3ceabSEric Dumazet 1945b68e9f85SHerbert Xu tso_segs = tcp_init_tso_segs(sk, skb, mss_now); 1946c1b4a7e6SDavid S. Miller BUG_ON(!tso_segs); 1947c1b4a7e6SDavid S. Miller 1948b68e9f85SHerbert Xu cwnd_quota = tcp_cwnd_test(tp, skb); 1949b68e9f85SHerbert Xu if (!cwnd_quota) 1950b68e9f85SHerbert Xu break; 1951b68e9f85SHerbert Xu 1952b68e9f85SHerbert Xu if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) 1953b68e9f85SHerbert Xu break; 1954b68e9f85SHerbert Xu 1955c1b4a7e6SDavid S. Miller if (tso_segs == 1) { 1956aa93466bSDavid S. Miller if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 1957aa93466bSDavid S. Miller (tcp_skb_is_last(sk, skb) ? 1958aa93466bSDavid S. Miller nonagle : TCP_NAGLE_PUSH)))) 1959aa93466bSDavid S. Miller break; 1960c1b4a7e6SDavid S. Miller } else { 1961d5dd9175SIlpo Järvinen if (!push_one && tcp_tso_should_defer(sk, skb)) 1962aa93466bSDavid S. Miller break; 1963c1b4a7e6SDavid S. Miller } 1964aa93466bSDavid S. Miller 196546d3ceabSEric Dumazet /* TSQ : sk_wmem_alloc accounts skb truesize, 196646d3ceabSEric Dumazet * including skb overhead. But thats OK. 196746d3ceabSEric Dumazet */ 196846d3ceabSEric Dumazet if (atomic_read(&sk->sk_wmem_alloc) >= sysctl_tcp_limit_output_bytes) { 196946d3ceabSEric Dumazet set_bit(TSQ_THROTTLED, &tp->tsq_flags); 197046d3ceabSEric Dumazet break; 197146d3ceabSEric Dumazet } 1972c8ac3774SHerbert Xu limit = mss_now; 1973f8269a49SIlpo Järvinen if (tso_segs > 1 && !tcp_urg_mode(tp)) 19740e3a4803SIlpo Järvinen limit = tcp_mss_split_point(sk, skb, mss_now, 19750e3a4803SIlpo Järvinen cwnd_quota); 1976c8ac3774SHerbert Xu 1977c8ac3774SHerbert Xu if (skb->len > limit && 1978c4ead4c5SEric Dumazet unlikely(tso_fragment(sk, skb, limit, mss_now, gfp))) 19791da177e4SLinus Torvalds break; 19801da177e4SLinus Torvalds 19811da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 1982c1b4a7e6SDavid S. Miller 1983d5dd9175SIlpo Järvinen if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) 19841da177e4SLinus Torvalds break; 19851da177e4SLinus Torvalds 19861da177e4SLinus Torvalds /* Advance the send_head. This one is sent out. 19871da177e4SLinus Torvalds * This call will increment packets_out. 19881da177e4SLinus Torvalds */ 198966f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 19901da177e4SLinus Torvalds 19911da177e4SLinus Torvalds tcp_minshall_update(tp, mss_now, skb); 1992a262f0cdSNandita Dukkipati sent_pkts += tcp_skb_pcount(skb); 1993d5dd9175SIlpo Järvinen 1994d5dd9175SIlpo Järvinen if (push_one) 1995d5dd9175SIlpo Järvinen break; 19961da177e4SLinus Torvalds } 1997a262f0cdSNandita Dukkipati if (inet_csk(sk)->icsk_ca_state == TCP_CA_Recovery) 1998a262f0cdSNandita Dukkipati tp->prr_out += sent_pkts; 19991da177e4SLinus Torvalds 2000aa93466bSDavid S. Miller if (likely(sent_pkts)) { 20019e412ba7SIlpo Järvinen tcp_cwnd_validate(sk); 2002a2a385d6SEric Dumazet return false; 20031da177e4SLinus Torvalds } 2004fe067e8aSDavid S. Miller return !tp->packets_out && tcp_send_head(sk); 20051da177e4SLinus Torvalds } 20061da177e4SLinus Torvalds 2007a762a980SDavid S. Miller /* Push out any pending frames which were held back due to 2008a762a980SDavid S. Miller * TCP_CORK or attempt at coalescing tiny packets. 2009a762a980SDavid S. Miller * The socket must be locked by the caller. 2010a762a980SDavid S. Miller */ 20119e412ba7SIlpo Järvinen void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 20129e412ba7SIlpo Järvinen int nonagle) 2013a762a980SDavid S. Miller { 2014726e07a8SIlpo Järvinen /* If we are closed, the bytes will have to remain here. 2015726e07a8SIlpo Järvinen * In time closedown will finish, we empty the write queue and 2016726e07a8SIlpo Järvinen * all will be happy. 2017726e07a8SIlpo Järvinen */ 2018726e07a8SIlpo Järvinen if (unlikely(sk->sk_state == TCP_CLOSE)) 2019726e07a8SIlpo Järvinen return; 2020726e07a8SIlpo Järvinen 2021d5dd9175SIlpo Järvinen if (tcp_write_xmit(sk, cur_mss, nonagle, 0, GFP_ATOMIC)) 20229e412ba7SIlpo Järvinen tcp_check_probe_timer(sk); 2023a762a980SDavid S. Miller } 2024a762a980SDavid S. Miller 2025c1b4a7e6SDavid S. Miller /* Send _single_ skb sitting at the send head. This function requires 2026c1b4a7e6SDavid S. Miller * true push pending frames to setup probe timer etc. 2027c1b4a7e6SDavid S. Miller */ 2028c1b4a7e6SDavid S. Miller void tcp_push_one(struct sock *sk, unsigned int mss_now) 2029c1b4a7e6SDavid S. Miller { 2030fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 2031c1b4a7e6SDavid S. Miller 2032c1b4a7e6SDavid S. Miller BUG_ON(!skb || skb->len < mss_now); 2033c1b4a7e6SDavid S. Miller 2034d5dd9175SIlpo Järvinen tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); 2035c1b4a7e6SDavid S. Miller } 2036c1b4a7e6SDavid S. Miller 20371da177e4SLinus Torvalds /* This function returns the amount that we can raise the 20381da177e4SLinus Torvalds * usable window based on the following constraints 20391da177e4SLinus Torvalds * 20401da177e4SLinus Torvalds * 1. The window can never be shrunk once it is offered (RFC 793) 20411da177e4SLinus Torvalds * 2. We limit memory per socket 20421da177e4SLinus Torvalds * 20431da177e4SLinus Torvalds * RFC 1122: 20441da177e4SLinus Torvalds * "the suggested [SWS] avoidance algorithm for the receiver is to keep 20451da177e4SLinus Torvalds * RECV.NEXT + RCV.WIN fixed until: 20461da177e4SLinus Torvalds * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 20471da177e4SLinus Torvalds * 20481da177e4SLinus Torvalds * i.e. don't raise the right edge of the window until you can raise 20491da177e4SLinus Torvalds * it at least MSS bytes. 20501da177e4SLinus Torvalds * 20511da177e4SLinus Torvalds * Unfortunately, the recommended algorithm breaks header prediction, 20521da177e4SLinus Torvalds * since header prediction assumes th->window stays fixed. 20531da177e4SLinus Torvalds * 20541da177e4SLinus Torvalds * Strictly speaking, keeping th->window fixed violates the receiver 20551da177e4SLinus Torvalds * side SWS prevention criteria. The problem is that under this rule 20561da177e4SLinus Torvalds * a stream of single byte packets will cause the right side of the 20571da177e4SLinus Torvalds * window to always advance by a single byte. 20581da177e4SLinus Torvalds * 20591da177e4SLinus Torvalds * Of course, if the sender implements sender side SWS prevention 20601da177e4SLinus Torvalds * then this will not be a problem. 20611da177e4SLinus Torvalds * 20621da177e4SLinus Torvalds * BSD seems to make the following compromise: 20631da177e4SLinus Torvalds * 20641da177e4SLinus Torvalds * If the free space is less than the 1/4 of the maximum 20651da177e4SLinus Torvalds * space available and the free space is less than 1/2 mss, 20661da177e4SLinus Torvalds * then set the window to 0. 20671da177e4SLinus Torvalds * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 20681da177e4SLinus Torvalds * Otherwise, just prevent the window from shrinking 20691da177e4SLinus Torvalds * and from being larger than the largest representable value. 20701da177e4SLinus Torvalds * 20711da177e4SLinus Torvalds * This prevents incremental opening of the window in the regime 20721da177e4SLinus Torvalds * where TCP is limited by the speed of the reader side taking 20731da177e4SLinus Torvalds * data out of the TCP receive queue. It does nothing about 20741da177e4SLinus Torvalds * those cases where the window is constrained on the sender side 20751da177e4SLinus Torvalds * because the pipeline is full. 20761da177e4SLinus Torvalds * 20771da177e4SLinus Torvalds * BSD also seems to "accidentally" limit itself to windows that are a 20781da177e4SLinus Torvalds * multiple of MSS, at least until the free space gets quite small. 20791da177e4SLinus Torvalds * This would appear to be a side effect of the mbuf implementation. 20801da177e4SLinus Torvalds * Combining these two algorithms results in the observed behavior 20811da177e4SLinus Torvalds * of having a fixed window size at almost all times. 20821da177e4SLinus Torvalds * 20831da177e4SLinus Torvalds * Below we obtain similar behavior by forcing the offered window to 20841da177e4SLinus Torvalds * a multiple of the mss when it is feasible to do so. 20851da177e4SLinus Torvalds * 20861da177e4SLinus Torvalds * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 20871da177e4SLinus Torvalds * Regular options like TIMESTAMP are taken into account. 20881da177e4SLinus Torvalds */ 20891da177e4SLinus Torvalds u32 __tcp_select_window(struct sock *sk) 20901da177e4SLinus Torvalds { 2091463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 20921da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2093caa20d9aSStephen Hemminger /* MSS for the peer's data. Previous versions used mss_clamp 20941da177e4SLinus Torvalds * here. I don't know if the value based on our guesses 20951da177e4SLinus Torvalds * of peer's MSS is better for the performance. It's more correct 20961da177e4SLinus Torvalds * but may be worse for the performance because of rcv_mss 20971da177e4SLinus Torvalds * fluctuations. --SAW 1998/11/1 20981da177e4SLinus Torvalds */ 2099463c84b9SArnaldo Carvalho de Melo int mss = icsk->icsk_ack.rcv_mss; 21001da177e4SLinus Torvalds int free_space = tcp_space(sk); 21011da177e4SLinus Torvalds int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk)); 21021da177e4SLinus Torvalds int window; 21031da177e4SLinus Torvalds 21041da177e4SLinus Torvalds if (mss > full_space) 21051da177e4SLinus Torvalds mss = full_space; 21061da177e4SLinus Torvalds 2107b92edbe0SEric Dumazet if (free_space < (full_space >> 1)) { 2108463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.quick = 0; 21091da177e4SLinus Torvalds 2110180d8cd9SGlauber Costa if (sk_under_memory_pressure(sk)) 2111056834d9SIlpo Järvinen tp->rcv_ssthresh = min(tp->rcv_ssthresh, 2112056834d9SIlpo Järvinen 4U * tp->advmss); 21131da177e4SLinus Torvalds 21141da177e4SLinus Torvalds if (free_space < mss) 21151da177e4SLinus Torvalds return 0; 21161da177e4SLinus Torvalds } 21171da177e4SLinus Torvalds 21181da177e4SLinus Torvalds if (free_space > tp->rcv_ssthresh) 21191da177e4SLinus Torvalds free_space = tp->rcv_ssthresh; 21201da177e4SLinus Torvalds 21211da177e4SLinus Torvalds /* Don't do rounding if we are using window scaling, since the 21221da177e4SLinus Torvalds * scaled window will not line up with the MSS boundary anyway. 21231da177e4SLinus Torvalds */ 21241da177e4SLinus Torvalds window = tp->rcv_wnd; 21251da177e4SLinus Torvalds if (tp->rx_opt.rcv_wscale) { 21261da177e4SLinus Torvalds window = free_space; 21271da177e4SLinus Torvalds 21281da177e4SLinus Torvalds /* Advertise enough space so that it won't get scaled away. 21291da177e4SLinus Torvalds * Import case: prevent zero window announcement if 21301da177e4SLinus Torvalds * 1<<rcv_wscale > mss. 21311da177e4SLinus Torvalds */ 21321da177e4SLinus Torvalds if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window) 21331da177e4SLinus Torvalds window = (((window >> tp->rx_opt.rcv_wscale) + 1) 21341da177e4SLinus Torvalds << tp->rx_opt.rcv_wscale); 21351da177e4SLinus Torvalds } else { 21361da177e4SLinus Torvalds /* Get the largest window that is a nice multiple of mss. 21371da177e4SLinus Torvalds * Window clamp already applied above. 21381da177e4SLinus Torvalds * If our current window offering is within 1 mss of the 21391da177e4SLinus Torvalds * free space we just keep it. This prevents the divide 21401da177e4SLinus Torvalds * and multiply from happening most of the time. 21411da177e4SLinus Torvalds * We also don't do any window rounding when the free space 21421da177e4SLinus Torvalds * is too small. 21431da177e4SLinus Torvalds */ 21441da177e4SLinus Torvalds if (window <= free_space - mss || window > free_space) 21451da177e4SLinus Torvalds window = (free_space / mss) * mss; 214684565070SJohn Heffner else if (mss == full_space && 2147b92edbe0SEric Dumazet free_space > window + (full_space >> 1)) 214884565070SJohn Heffner window = free_space; 21491da177e4SLinus Torvalds } 21501da177e4SLinus Torvalds 21511da177e4SLinus Torvalds return window; 21521da177e4SLinus Torvalds } 21531da177e4SLinus Torvalds 21544a17fc3aSIlpo Järvinen /* Collapses two adjacent SKB's during retransmission. */ 21554a17fc3aSIlpo Järvinen static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) 21561da177e4SLinus Torvalds { 21571da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2158fe067e8aSDavid S. Miller struct sk_buff *next_skb = tcp_write_queue_next(sk, skb); 2159058dc334SIlpo Järvinen int skb_size, next_skb_size; 21601da177e4SLinus Torvalds 2161058dc334SIlpo Järvinen skb_size = skb->len; 2162058dc334SIlpo Järvinen next_skb_size = next_skb->len; 21631da177e4SLinus Torvalds 2164058dc334SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); 21651da177e4SLinus Torvalds 21666859d494SIlpo Järvinen tcp_highest_sack_combine(sk, next_skb, skb); 2167a6963a6bSIlpo Järvinen 2168fe067e8aSDavid S. Miller tcp_unlink_write_queue(next_skb, sk); 21691da177e4SLinus Torvalds 2170058dc334SIlpo Järvinen skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size), 21711a4e2d09SArnaldo Carvalho de Melo next_skb_size); 21721da177e4SLinus Torvalds 217352d570aaSJarek Poplawski if (next_skb->ip_summed == CHECKSUM_PARTIAL) 217452d570aaSJarek Poplawski skb->ip_summed = CHECKSUM_PARTIAL; 21751da177e4SLinus Torvalds 217684fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 21771da177e4SLinus Torvalds skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); 21781da177e4SLinus Torvalds 21791da177e4SLinus Torvalds /* Update sequence range on original skb. */ 21801da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 21811da177e4SLinus Torvalds 2182e6c7d085SIlpo Järvinen /* Merge over control information. This moves PSH/FIN etc. over */ 21834de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; 21841da177e4SLinus Torvalds 21851da177e4SLinus Torvalds /* All done, get rid of second SKB and account for it so 21861da177e4SLinus Torvalds * packet counting does not break. 21871da177e4SLinus Torvalds */ 21884828e7f4SIlpo Järvinen TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; 2189b7689205SIlpo Järvinen 2190b7689205SIlpo Järvinen /* changed transmit queue under us so clear hints */ 2191ef9da47cSIlpo Järvinen tcp_clear_retrans_hints_partial(tp); 2192ef9da47cSIlpo Järvinen if (next_skb == tp->retransmit_skb_hint) 2193ef9da47cSIlpo Järvinen tp->retransmit_skb_hint = skb; 2194b7689205SIlpo Järvinen 2195797108d1SIlpo Järvinen tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); 2196797108d1SIlpo Järvinen 21973ab224beSHideo Aoki sk_wmem_free_skb(sk, next_skb); 21981da177e4SLinus Torvalds } 21991da177e4SLinus Torvalds 220067edfef7SAndi Kleen /* Check if coalescing SKBs is legal. */ 2201a2a385d6SEric Dumazet static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) 22024a17fc3aSIlpo Järvinen { 22034a17fc3aSIlpo Järvinen if (tcp_skb_pcount(skb) > 1) 2204a2a385d6SEric Dumazet return false; 22054a17fc3aSIlpo Järvinen /* TODO: SACK collapsing could be used to remove this condition */ 22064a17fc3aSIlpo Järvinen if (skb_shinfo(skb)->nr_frags != 0) 2207a2a385d6SEric Dumazet return false; 22084a17fc3aSIlpo Järvinen if (skb_cloned(skb)) 2209a2a385d6SEric Dumazet return false; 22104a17fc3aSIlpo Järvinen if (skb == tcp_send_head(sk)) 2211a2a385d6SEric Dumazet return false; 22124a17fc3aSIlpo Järvinen /* Some heurestics for collapsing over SACK'd could be invented */ 22134a17fc3aSIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 2214a2a385d6SEric Dumazet return false; 22154a17fc3aSIlpo Järvinen 2216a2a385d6SEric Dumazet return true; 22174a17fc3aSIlpo Järvinen } 22184a17fc3aSIlpo Järvinen 221967edfef7SAndi Kleen /* Collapse packets in the retransmit queue to make to create 222067edfef7SAndi Kleen * less packets on the wire. This is only done on retransmission. 222167edfef7SAndi Kleen */ 22224a17fc3aSIlpo Järvinen static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, 22234a17fc3aSIlpo Järvinen int space) 22244a17fc3aSIlpo Järvinen { 22254a17fc3aSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 22264a17fc3aSIlpo Järvinen struct sk_buff *skb = to, *tmp; 2227a2a385d6SEric Dumazet bool first = true; 22284a17fc3aSIlpo Järvinen 22294a17fc3aSIlpo Järvinen if (!sysctl_tcp_retrans_collapse) 22304a17fc3aSIlpo Järvinen return; 22314de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 22324a17fc3aSIlpo Järvinen return; 22334a17fc3aSIlpo Järvinen 22344a17fc3aSIlpo Järvinen tcp_for_write_queue_from_safe(skb, tmp, sk) { 22354a17fc3aSIlpo Järvinen if (!tcp_can_collapse(sk, skb)) 22364a17fc3aSIlpo Järvinen break; 22374a17fc3aSIlpo Järvinen 22384a17fc3aSIlpo Järvinen space -= skb->len; 22394a17fc3aSIlpo Järvinen 22404a17fc3aSIlpo Järvinen if (first) { 2241a2a385d6SEric Dumazet first = false; 22424a17fc3aSIlpo Järvinen continue; 22434a17fc3aSIlpo Järvinen } 22444a17fc3aSIlpo Järvinen 22454a17fc3aSIlpo Järvinen if (space < 0) 22464a17fc3aSIlpo Järvinen break; 22474a17fc3aSIlpo Järvinen /* Punt if not enough space exists in the first SKB for 22484a17fc3aSIlpo Järvinen * the data in the second 22494a17fc3aSIlpo Järvinen */ 2250a21d4572SEric Dumazet if (skb->len > skb_availroom(to)) 22514a17fc3aSIlpo Järvinen break; 22524a17fc3aSIlpo Järvinen 22534a17fc3aSIlpo Järvinen if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) 22544a17fc3aSIlpo Järvinen break; 22554a17fc3aSIlpo Järvinen 22564a17fc3aSIlpo Järvinen tcp_collapse_retrans(sk, to); 22574a17fc3aSIlpo Järvinen } 22584a17fc3aSIlpo Järvinen } 22594a17fc3aSIlpo Järvinen 22601da177e4SLinus Torvalds /* This retransmits one SKB. Policy decisions and retransmit queue 22611da177e4SLinus Torvalds * state updates are done by the caller. Returns non-zero if an 22621da177e4SLinus Torvalds * error occurred which prevented the send. 22631da177e4SLinus Torvalds */ 22641da177e4SLinus Torvalds int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) 22651da177e4SLinus Torvalds { 22661da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 22675d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 22687d227cd2SSridhar Samudrala unsigned int cur_mss; 22691da177e4SLinus Torvalds int err; 22701da177e4SLinus Torvalds 22715d424d5aSJohn Heffner /* Inconslusive MTU probe */ 22725d424d5aSJohn Heffner if (icsk->icsk_mtup.probe_size) { 22735d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 22745d424d5aSJohn Heffner } 22755d424d5aSJohn Heffner 22761da177e4SLinus Torvalds /* Do not sent more than we queued. 1/4 is reserved for possible 2277caa20d9aSStephen Hemminger * copying overhead: fragmentation, tunneling, mangling etc. 22781da177e4SLinus Torvalds */ 22791da177e4SLinus Torvalds if (atomic_read(&sk->sk_wmem_alloc) > 22801da177e4SLinus Torvalds min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf)) 22811da177e4SLinus Torvalds return -EAGAIN; 22821da177e4SLinus Torvalds 22831da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 22841da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 22851da177e4SLinus Torvalds BUG(); 22861da177e4SLinus Torvalds if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 22871da177e4SLinus Torvalds return -ENOMEM; 22881da177e4SLinus Torvalds } 22891da177e4SLinus Torvalds 22907d227cd2SSridhar Samudrala if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 22917d227cd2SSridhar Samudrala return -EHOSTUNREACH; /* Routing failure or similar. */ 22927d227cd2SSridhar Samudrala 22930c54b85fSIlpo Järvinen cur_mss = tcp_current_mss(sk); 22947d227cd2SSridhar Samudrala 22951da177e4SLinus Torvalds /* If receiver has shrunk his window, and skb is out of 22961da177e4SLinus Torvalds * new window, do not retransmit it. The exception is the 22971da177e4SLinus Torvalds * case, when window is shrunk to zero. In this case 22981da177e4SLinus Torvalds * our retransmit serves as a zero window probe. 22991da177e4SLinus Torvalds */ 23009d4fb27dSJoe Perches if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && 23019d4fb27dSJoe Perches TCP_SKB_CB(skb)->seq != tp->snd_una) 23021da177e4SLinus Torvalds return -EAGAIN; 23031da177e4SLinus Torvalds 23041da177e4SLinus Torvalds if (skb->len > cur_mss) { 2305846998aeSDavid S. Miller if (tcp_fragment(sk, skb, cur_mss, cur_mss)) 23061da177e4SLinus Torvalds return -ENOMEM; /* We'll try again later. */ 230702276f3cSIlpo Järvinen } else { 23089eb9362eSIlpo Järvinen int oldpcount = tcp_skb_pcount(skb); 23099eb9362eSIlpo Järvinen 23109eb9362eSIlpo Järvinen if (unlikely(oldpcount > 1)) { 231102276f3cSIlpo Järvinen tcp_init_tso_segs(sk, skb, cur_mss); 23129eb9362eSIlpo Järvinen tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb)); 23139eb9362eSIlpo Järvinen } 23141da177e4SLinus Torvalds } 23151da177e4SLinus Torvalds 23161da177e4SLinus Torvalds tcp_retrans_try_collapse(sk, skb, cur_mss); 23171da177e4SLinus Torvalds 23181da177e4SLinus Torvalds /* Some Solaris stacks overoptimize and ignore the FIN on a 23191da177e4SLinus Torvalds * retransmit when old data is attached. So strip it off 23201da177e4SLinus Torvalds * since it is cheap to do so and saves bytes on the network. 23211da177e4SLinus Torvalds */ 23221da177e4SLinus Torvalds if (skb->len > 0 && 23234de075e0SEric Dumazet (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && 23241da177e4SLinus Torvalds tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) { 23251da177e4SLinus Torvalds if (!pskb_trim(skb, 0)) { 2326e870a8efSIlpo Järvinen /* Reuse, even though it does some unnecessary work */ 2327e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1, 23284de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags); 23291da177e4SLinus Torvalds skb->ip_summed = CHECKSUM_NONE; 23301da177e4SLinus Torvalds } 23311da177e4SLinus Torvalds } 23321da177e4SLinus Torvalds 23331da177e4SLinus Torvalds /* Make a copy, if the first transmission SKB clone we made 23341da177e4SLinus Torvalds * is still in somebody's hands, else make a clone. 23351da177e4SLinus Torvalds */ 23361da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 23371da177e4SLinus Torvalds 2338117632e6SEric Dumazet /* make sure skb->data is aligned on arches that require it */ 2339117632e6SEric Dumazet if (unlikely(NET_IP_ALIGN && ((unsigned long)skb->data & 3))) { 2340117632e6SEric Dumazet struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER, 2341117632e6SEric Dumazet GFP_ATOMIC); 2342117632e6SEric Dumazet err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) : 2343117632e6SEric Dumazet -ENOBUFS; 2344117632e6SEric Dumazet } else { 2345dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 2346117632e6SEric Dumazet } 23471da177e4SLinus Torvalds 23481da177e4SLinus Torvalds if (err == 0) { 23491da177e4SLinus Torvalds /* Update global TCP statistics. */ 235081cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); 23511da177e4SLinus Torvalds 23521da177e4SLinus Torvalds tp->total_retrans++; 23531da177e4SLinus Torvalds 23541da177e4SLinus Torvalds #if FASTRETRANS_DEBUG > 0 23551da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 2356e87cc472SJoe Perches net_dbg_ratelimited("retrans_out leaked\n"); 23571da177e4SLinus Torvalds } 23581da177e4SLinus Torvalds #endif 2359b08d6cb2SIlpo Järvinen if (!tp->retrans_out) 2360b08d6cb2SIlpo Järvinen tp->lost_retrans_low = tp->snd_nxt; 23611da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 23621da177e4SLinus Torvalds tp->retrans_out += tcp_skb_pcount(skb); 23631da177e4SLinus Torvalds 23641da177e4SLinus Torvalds /* Save stamp of the first retransmit. */ 23651da177e4SLinus Torvalds if (!tp->retrans_stamp) 23661da177e4SLinus Torvalds tp->retrans_stamp = TCP_SKB_CB(skb)->when; 23671da177e4SLinus Torvalds 2368c24f691bSYuchung Cheng tp->undo_retrans += tcp_skb_pcount(skb); 23691da177e4SLinus Torvalds 23701da177e4SLinus Torvalds /* snd_nxt is stored to detect loss of retransmitted segment, 23711da177e4SLinus Torvalds * see tcp_input.c tcp_sacktag_write_queue(). 23721da177e4SLinus Torvalds */ 23731da177e4SLinus Torvalds TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt; 23741da177e4SLinus Torvalds } 23751da177e4SLinus Torvalds return err; 23761da177e4SLinus Torvalds } 23771da177e4SLinus Torvalds 237867edfef7SAndi Kleen /* Check if we forward retransmits are possible in the current 237967edfef7SAndi Kleen * window/congestion state. 238067edfef7SAndi Kleen */ 2381a2a385d6SEric Dumazet static bool tcp_can_forward_retransmit(struct sock *sk) 2382b5afe7bcSIlpo Järvinen { 2383b5afe7bcSIlpo Järvinen const struct inet_connection_sock *icsk = inet_csk(sk); 2384cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 2385b5afe7bcSIlpo Järvinen 2386b5afe7bcSIlpo Järvinen /* Forward retransmissions are possible only during Recovery. */ 2387b5afe7bcSIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Recovery) 2388a2a385d6SEric Dumazet return false; 2389b5afe7bcSIlpo Järvinen 2390b5afe7bcSIlpo Järvinen /* No forward retransmissions in Reno are possible. */ 2391b5afe7bcSIlpo Järvinen if (tcp_is_reno(tp)) 2392a2a385d6SEric Dumazet return false; 2393b5afe7bcSIlpo Järvinen 2394b5afe7bcSIlpo Järvinen /* Yeah, we have to make difficult choice between forward transmission 2395b5afe7bcSIlpo Järvinen * and retransmission... Both ways have their merits... 2396b5afe7bcSIlpo Järvinen * 2397b5afe7bcSIlpo Järvinen * For now we do not retransmit anything, while we have some new 2398b5afe7bcSIlpo Järvinen * segments to send. In the other cases, follow rule 3 for 2399b5afe7bcSIlpo Järvinen * NextSeg() specified in RFC3517. 2400b5afe7bcSIlpo Järvinen */ 2401b5afe7bcSIlpo Järvinen 2402b5afe7bcSIlpo Järvinen if (tcp_may_send_now(sk)) 2403a2a385d6SEric Dumazet return false; 2404b5afe7bcSIlpo Järvinen 2405a2a385d6SEric Dumazet return true; 2406b5afe7bcSIlpo Järvinen } 2407b5afe7bcSIlpo Järvinen 24081da177e4SLinus Torvalds /* This gets called after a retransmit timeout, and the initially 24091da177e4SLinus Torvalds * retransmitted data is acknowledged. It tries to continue 24101da177e4SLinus Torvalds * resending the rest of the retransmit queue, until either 24111da177e4SLinus Torvalds * we've sent it all or the congestion window limit is reached. 24121da177e4SLinus Torvalds * If doing SACK, the first ACK which comes back for a timeout 24131da177e4SLinus Torvalds * based retransmit packet might feed us FACK information again. 24141da177e4SLinus Torvalds * If so, we use it to avoid unnecessarily retransmissions. 24151da177e4SLinus Torvalds */ 24161da177e4SLinus Torvalds void tcp_xmit_retransmit_queue(struct sock *sk) 24171da177e4SLinus Torvalds { 24186687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 24191da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 24201da177e4SLinus Torvalds struct sk_buff *skb; 24210e1c54c2SIlpo Järvinen struct sk_buff *hole = NULL; 2422618d9f25SIlpo Järvinen u32 last_lost; 242361eb55f4SIlpo Järvinen int mib_idx; 24240e1c54c2SIlpo Järvinen int fwd_rexmitting = 0; 24256a438bbeSStephen Hemminger 242645e77d31SIlpo Järvinen if (!tp->packets_out) 242745e77d31SIlpo Järvinen return; 242845e77d31SIlpo Järvinen 242908ebd172SIlpo Järvinen if (!tp->lost_out) 243008ebd172SIlpo Järvinen tp->retransmit_high = tp->snd_una; 243108ebd172SIlpo Järvinen 2432618d9f25SIlpo Järvinen if (tp->retransmit_skb_hint) { 24336a438bbeSStephen Hemminger skb = tp->retransmit_skb_hint; 2434618d9f25SIlpo Järvinen last_lost = TCP_SKB_CB(skb)->end_seq; 2435618d9f25SIlpo Järvinen if (after(last_lost, tp->retransmit_high)) 2436618d9f25SIlpo Järvinen last_lost = tp->retransmit_high; 2437618d9f25SIlpo Järvinen } else { 2438fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 2439618d9f25SIlpo Järvinen last_lost = tp->snd_una; 2440618d9f25SIlpo Järvinen } 24411da177e4SLinus Torvalds 2442fe067e8aSDavid S. Miller tcp_for_write_queue_from(skb, sk) { 24431da177e4SLinus Torvalds __u8 sacked = TCP_SKB_CB(skb)->sacked; 24441da177e4SLinus Torvalds 2445fe067e8aSDavid S. Miller if (skb == tcp_send_head(sk)) 2446fe067e8aSDavid S. Miller break; 24476a438bbeSStephen Hemminger /* we could do better than to assign each time */ 24480e1c54c2SIlpo Järvinen if (hole == NULL) 24496a438bbeSStephen Hemminger tp->retransmit_skb_hint = skb; 24506a438bbeSStephen Hemminger 24511da177e4SLinus Torvalds /* Assume this retransmit will generate 24521da177e4SLinus Torvalds * only one packet for congestion window 24531da177e4SLinus Torvalds * calculation purposes. This works because 24541da177e4SLinus Torvalds * tcp_retransmit_skb() will chop up the 24551da177e4SLinus Torvalds * packet to be MSS sized and all the 24561da177e4SLinus Torvalds * packet counting works out. 24571da177e4SLinus Torvalds */ 24581da177e4SLinus Torvalds if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) 24591da177e4SLinus Torvalds return; 24600e1c54c2SIlpo Järvinen 24610e1c54c2SIlpo Järvinen if (fwd_rexmitting) { 24620e1c54c2SIlpo Järvinen begin_fwd: 24630e1c54c2SIlpo Järvinen if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp))) 2464006f582cSIlpo Järvinen break; 24650e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFORWARDRETRANS; 24660e1c54c2SIlpo Järvinen 24670e1c54c2SIlpo Järvinen } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) { 2468618d9f25SIlpo Järvinen tp->retransmit_high = last_lost; 24690e1c54c2SIlpo Järvinen if (!tcp_can_forward_retransmit(sk)) 24700e1c54c2SIlpo Järvinen break; 24710e1c54c2SIlpo Järvinen /* Backtrack if necessary to non-L'ed skb */ 24720e1c54c2SIlpo Järvinen if (hole != NULL) { 24730e1c54c2SIlpo Järvinen skb = hole; 24740e1c54c2SIlpo Järvinen hole = NULL; 24750e1c54c2SIlpo Järvinen } 24760e1c54c2SIlpo Järvinen fwd_rexmitting = 1; 24770e1c54c2SIlpo Järvinen goto begin_fwd; 24780e1c54c2SIlpo Järvinen 24790e1c54c2SIlpo Järvinen } else if (!(sacked & TCPCB_LOST)) { 2480ac11ba75SIlpo Järvinen if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) 24810e1c54c2SIlpo Järvinen hole = skb; 248261eb55f4SIlpo Järvinen continue; 24831da177e4SLinus Torvalds 24840e1c54c2SIlpo Järvinen } else { 2485618d9f25SIlpo Järvinen last_lost = TCP_SKB_CB(skb)->end_seq; 24860e1c54c2SIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Loss) 24870e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFASTRETRANS; 24880e1c54c2SIlpo Järvinen else 24890e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; 24900e1c54c2SIlpo Järvinen } 24910e1c54c2SIlpo Järvinen 24920e1c54c2SIlpo Järvinen if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) 249361eb55f4SIlpo Järvinen continue; 249440b215e5SPavel Emelyanov 249509e9b813SEric Dumazet if (tcp_retransmit_skb(sk, skb)) { 249609e9b813SEric Dumazet NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL); 24971da177e4SLinus Torvalds return; 249809e9b813SEric Dumazet } 2499de0744afSPavel Emelyanov NET_INC_STATS_BH(sock_net(sk), mib_idx); 25001da177e4SLinus Torvalds 2501a262f0cdSNandita Dukkipati if (inet_csk(sk)->icsk_ca_state == TCP_CA_Recovery) 2502a262f0cdSNandita Dukkipati tp->prr_out += tcp_skb_pcount(skb); 2503a262f0cdSNandita Dukkipati 2504fe067e8aSDavid S. Miller if (skb == tcp_write_queue_head(sk)) 2505463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 25063f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, 25073f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 25081da177e4SLinus Torvalds } 25091da177e4SLinus Torvalds } 25101da177e4SLinus Torvalds 25111da177e4SLinus Torvalds /* Send a fin. The caller locks the socket for us. This cannot be 25121da177e4SLinus Torvalds * allowed to fail queueing a FIN frame under any circumstances. 25131da177e4SLinus Torvalds */ 25141da177e4SLinus Torvalds void tcp_send_fin(struct sock *sk) 25151da177e4SLinus Torvalds { 25161da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2517fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_write_queue_tail(sk); 25181da177e4SLinus Torvalds int mss_now; 25191da177e4SLinus Torvalds 25201da177e4SLinus Torvalds /* Optimization, tack on the FIN if we have a queue of 25211da177e4SLinus Torvalds * unsent frames. But be careful about outgoing SACKS 25221da177e4SLinus Torvalds * and IP options. 25231da177e4SLinus Torvalds */ 25240c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 25251da177e4SLinus Torvalds 2526fe067e8aSDavid S. Miller if (tcp_send_head(sk) != NULL) { 25274de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_FIN; 25281da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq++; 25291da177e4SLinus Torvalds tp->write_seq++; 25301da177e4SLinus Torvalds } else { 25311da177e4SLinus Torvalds /* Socket is locked, keep trying until memory is available. */ 25321da177e4SLinus Torvalds for (;;) { 2533aa133076SWu Fengguang skb = alloc_skb_fclone(MAX_TCP_HEADER, 2534aa133076SWu Fengguang sk->sk_allocation); 25351da177e4SLinus Torvalds if (skb) 25361da177e4SLinus Torvalds break; 25371da177e4SLinus Torvalds yield(); 25381da177e4SLinus Torvalds } 25391da177e4SLinus Torvalds 25401da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 25411da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 25421da177e4SLinus Torvalds /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 2543e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tp->write_seq, 2544a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_FIN); 25451da177e4SLinus Torvalds tcp_queue_skb(sk, skb); 25461da177e4SLinus Torvalds } 25479e412ba7SIlpo Järvinen __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF); 25481da177e4SLinus Torvalds } 25491da177e4SLinus Torvalds 25501da177e4SLinus Torvalds /* We get here when a process closes a file descriptor (either due to 25511da177e4SLinus Torvalds * an explicit close() or as a byproduct of exit()'ing) and there 25521da177e4SLinus Torvalds * was unread data in the receive queue. This behavior is recommended 255365bb723cSGerrit Renker * by RFC 2525, section 2.17. -DaveM 25541da177e4SLinus Torvalds */ 2555dd0fc66fSAl Viro void tcp_send_active_reset(struct sock *sk, gfp_t priority) 25561da177e4SLinus Torvalds { 25571da177e4SLinus Torvalds struct sk_buff *skb; 25581da177e4SLinus Torvalds 25591da177e4SLinus Torvalds /* NOTE: No TCP options attached and we never retransmit this. */ 25601da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, priority); 25611da177e4SLinus Torvalds if (!skb) { 25624e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 25631da177e4SLinus Torvalds return; 25641da177e4SLinus Torvalds } 25651da177e4SLinus Torvalds 25661da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 25671da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 2568e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), 2569a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_RST); 25701da177e4SLinus Torvalds /* Send it off. */ 25711da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2572dfb4b9dcSDavid S. Miller if (tcp_transmit_skb(sk, skb, 0, priority)) 25734e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 257426af65cbSSridhar Samudrala 257581cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); 25761da177e4SLinus Torvalds } 25771da177e4SLinus Torvalds 257867edfef7SAndi Kleen /* Send a crossed SYN-ACK during socket establishment. 257967edfef7SAndi Kleen * WARNING: This routine must only be called when we have already sent 25801da177e4SLinus Torvalds * a SYN packet that crossed the incoming SYN that caused this routine 25811da177e4SLinus Torvalds * to get called. If this assumption fails then the initial rcv_wnd 25821da177e4SLinus Torvalds * and rcv_wscale values will not be correct. 25831da177e4SLinus Torvalds */ 25841da177e4SLinus Torvalds int tcp_send_synack(struct sock *sk) 25851da177e4SLinus Torvalds { 25861da177e4SLinus Torvalds struct sk_buff *skb; 25871da177e4SLinus Torvalds 2588fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 25894de075e0SEric Dumazet if (skb == NULL || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 259091df42beSJoe Perches pr_debug("%s: wrong queue state\n", __func__); 25911da177e4SLinus Torvalds return -EFAULT; 25921da177e4SLinus Torvalds } 25934de075e0SEric Dumazet if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { 25941da177e4SLinus Torvalds if (skb_cloned(skb)) { 25951da177e4SLinus Torvalds struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC); 25961da177e4SLinus Torvalds if (nskb == NULL) 25971da177e4SLinus Torvalds return -ENOMEM; 2598fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 25991da177e4SLinus Torvalds skb_header_release(nskb); 2600fe067e8aSDavid S. Miller __tcp_add_write_queue_head(sk, nskb); 26013ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 26023ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 26033ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 26041da177e4SLinus Torvalds skb = nskb; 26051da177e4SLinus Torvalds } 26061da177e4SLinus Torvalds 26074de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; 26081da177e4SLinus Torvalds TCP_ECN_send_synack(tcp_sk(sk), skb); 26091da177e4SLinus Torvalds } 26101da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2611dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 26121da177e4SLinus Torvalds } 26131da177e4SLinus Torvalds 26144aea39c1SEric Dumazet /** 26154aea39c1SEric Dumazet * tcp_make_synack - Prepare a SYN-ACK. 26164aea39c1SEric Dumazet * sk: listener socket 26174aea39c1SEric Dumazet * dst: dst entry attached to the SYNACK 26184aea39c1SEric Dumazet * req: request_sock pointer 26194aea39c1SEric Dumazet * rvp: request_values pointer 26204aea39c1SEric Dumazet * 26214aea39c1SEric Dumazet * Allocate one skb and build a SYNACK packet. 26224aea39c1SEric Dumazet * @dst is consumed : Caller should not use it again. 26234aea39c1SEric Dumazet */ 26241da177e4SLinus Torvalds struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst, 2625e6b4d113SWilliam Allen Simpson struct request_sock *req, 2626e6b4d113SWilliam Allen Simpson struct request_values *rvp) 26271da177e4SLinus Torvalds { 2628bd0388aeSWilliam Allen Simpson struct tcp_out_options opts; 26294957faadSWilliam Allen Simpson struct tcp_extend_values *xvp = tcp_xv(rvp); 26302e6599cbSArnaldo Carvalho de Melo struct inet_request_sock *ireq = inet_rsk(req); 26311da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 263228b2774aSEric Dumazet const struct tcp_cookie_values *cvp = tp->cookie_values; 26331da177e4SLinus Torvalds struct tcphdr *th; 26341da177e4SLinus Torvalds struct sk_buff *skb; 2635cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 2636bd0388aeSWilliam Allen Simpson int tcp_header_size; 2637f5fff5dcSTom Quetchenbach int mss; 263828b2774aSEric Dumazet int s_data_desired = 0; 26391da177e4SLinus Torvalds 264028b2774aSEric Dumazet if (cvp != NULL && cvp->s_data_constant && cvp->s_data_desired) 264128b2774aSEric Dumazet s_data_desired = cvp->s_data_desired; 264290ba9b19SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER + 15 + s_data_desired, GFP_ATOMIC); 26434aea39c1SEric Dumazet if (unlikely(!skb)) { 26444aea39c1SEric Dumazet dst_release(dst); 26451da177e4SLinus Torvalds return NULL; 26464aea39c1SEric Dumazet } 26471da177e4SLinus Torvalds /* Reserve space for headers. */ 26481da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 26491da177e4SLinus Torvalds 26504aea39c1SEric Dumazet skb_dst_set(skb, dst); 26511da177e4SLinus Torvalds 26520dbaee3bSDavid S. Miller mss = dst_metric_advmss(dst); 2653f5fff5dcSTom Quetchenbach if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss) 2654f5fff5dcSTom Quetchenbach mss = tp->rx_opt.user_mss; 2655f5fff5dcSTom Quetchenbach 265633ad798cSAdam Langley if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */ 265733ad798cSAdam Langley __u8 rcv_wscale; 265833ad798cSAdam Langley /* Set this up on the first call only */ 265933ad798cSAdam Langley req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW); 2660e88c64f0SHagen Paul Pfeifer 2661e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 2662e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 2663e88c64f0SHagen Paul Pfeifer (req->window_clamp > tcp_full_space(sk) || req->window_clamp == 0)) 2664e88c64f0SHagen Paul Pfeifer req->window_clamp = tcp_full_space(sk); 2665e88c64f0SHagen Paul Pfeifer 266633ad798cSAdam Langley /* tcp_full_space because it is guaranteed to be the first packet */ 266733ad798cSAdam Langley tcp_select_initial_window(tcp_full_space(sk), 2668f5fff5dcSTom Quetchenbach mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0), 266933ad798cSAdam Langley &req->rcv_wnd, 267033ad798cSAdam Langley &req->window_clamp, 267133ad798cSAdam Langley ireq->wscale_ok, 267231d12926Slaurent chavey &rcv_wscale, 267331d12926Slaurent chavey dst_metric(dst, RTAX_INITRWND)); 267433ad798cSAdam Langley ireq->rcv_wscale = rcv_wscale; 267533ad798cSAdam Langley } 2676cfb6eeb4SYOSHIFUJI Hideaki 267733ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 26788b5f12d0SFlorian Westphal #ifdef CONFIG_SYN_COOKIES 26798b5f12d0SFlorian Westphal if (unlikely(req->cookie_ts)) 26808b5f12d0SFlorian Westphal TCP_SKB_CB(skb)->when = cookie_init_timestamp(req); 26818b5f12d0SFlorian Westphal else 26828b5f12d0SFlorian Westphal #endif 268333ad798cSAdam Langley TCP_SKB_CB(skb)->when = tcp_time_stamp; 2684f5fff5dcSTom Quetchenbach tcp_header_size = tcp_synack_options(sk, req, mss, 26854957faadSWilliam Allen Simpson skb, &opts, &md5, xvp) 26864957faadSWilliam Allen Simpson + sizeof(*th); 268733ad798cSAdam Langley 2688aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 2689aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 26901da177e4SLinus Torvalds 2691aa8223c7SArnaldo Carvalho de Melo th = tcp_hdr(skb); 26921da177e4SLinus Torvalds memset(th, 0, sizeof(struct tcphdr)); 26931da177e4SLinus Torvalds th->syn = 1; 26941da177e4SLinus Torvalds th->ack = 1; 26951da177e4SLinus Torvalds TCP_ECN_make_synack(req, th); 2696a3116ac5SKOVACS Krisztian th->source = ireq->loc_port; 26972e6599cbSArnaldo Carvalho de Melo th->dest = ireq->rmt_port; 2698e870a8efSIlpo Järvinen /* Setting of flags are superfluous here for callers (and ECE is 2699e870a8efSIlpo Järvinen * not even correctly set) 2700e870a8efSIlpo Järvinen */ 2701e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn, 2702a3433f35SChangli Gao TCPHDR_SYN | TCPHDR_ACK); 27034957faadSWilliam Allen Simpson 27044957faadSWilliam Allen Simpson if (OPTION_COOKIE_EXTENSION & opts.options) { 270528b2774aSEric Dumazet if (s_data_desired) { 270628b2774aSEric Dumazet u8 *buf = skb_put(skb, s_data_desired); 27074957faadSWilliam Allen Simpson 27084957faadSWilliam Allen Simpson /* copy data directly from the listening socket. */ 270928b2774aSEric Dumazet memcpy(buf, cvp->s_data_payload, s_data_desired); 271028b2774aSEric Dumazet TCP_SKB_CB(skb)->end_seq += s_data_desired; 27114957faadSWilliam Allen Simpson } 27124957faadSWilliam Allen Simpson 27134957faadSWilliam Allen Simpson if (opts.hash_size > 0) { 27144957faadSWilliam Allen Simpson __u32 workspace[SHA_WORKSPACE_WORDS]; 27154957faadSWilliam Allen Simpson u32 *mess = &xvp->cookie_bakery[COOKIE_DIGEST_WORDS]; 27164957faadSWilliam Allen Simpson u32 *tail = &mess[COOKIE_MESSAGE_WORDS-1]; 27174957faadSWilliam Allen Simpson 27184957faadSWilliam Allen Simpson /* Secret recipe depends on the Timestamp, (future) 27194957faadSWilliam Allen Simpson * Sequence and Acknowledgment Numbers, Initiator 27204957faadSWilliam Allen Simpson * Cookie, and others handled by IP variant caller. 27214957faadSWilliam Allen Simpson */ 27224957faadSWilliam Allen Simpson *tail-- ^= opts.tsval; 27234957faadSWilliam Allen Simpson *tail-- ^= tcp_rsk(req)->rcv_isn + 1; 27244957faadSWilliam Allen Simpson *tail-- ^= TCP_SKB_CB(skb)->seq + 1; 27254957faadSWilliam Allen Simpson 27264957faadSWilliam Allen Simpson /* recommended */ 27270eae88f3SEric Dumazet *tail-- ^= (((__force u32)th->dest << 16) | (__force u32)th->source); 2728f9a2e69eSDavid S. Miller *tail-- ^= (u32)(unsigned long)cvp; /* per sockopt */ 27294957faadSWilliam Allen Simpson 27304957faadSWilliam Allen Simpson sha_transform((__u32 *)&xvp->cookie_bakery[0], 27314957faadSWilliam Allen Simpson (char *)mess, 27324957faadSWilliam Allen Simpson &workspace[0]); 27334957faadSWilliam Allen Simpson opts.hash_location = 27344957faadSWilliam Allen Simpson (__u8 *)&xvp->cookie_bakery[0]; 27354957faadSWilliam Allen Simpson } 27364957faadSWilliam Allen Simpson } 27374957faadSWilliam Allen Simpson 27381da177e4SLinus Torvalds th->seq = htonl(TCP_SKB_CB(skb)->seq); 27392e6599cbSArnaldo Carvalho de Melo th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1); 27401da177e4SLinus Torvalds 27411da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 2742600ff0c2SIlpo Järvinen th->window = htons(min(req->rcv_wnd, 65535U)); 2743bd0388aeSWilliam Allen Simpson tcp_options_write((__be32 *)(th + 1), tp, &opts); 27441da177e4SLinus Torvalds th->doff = (tcp_header_size >> 2); 2745aa2ea058STom Herbert TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, tcp_skb_pcount(skb)); 2746cfb6eeb4SYOSHIFUJI Hideaki 2747cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 2748cfb6eeb4SYOSHIFUJI Hideaki /* Okay, we have all we need - do the md5 hash if needed */ 2749cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 2750bd0388aeSWilliam Allen Simpson tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, 275149a72dfbSAdam Langley md5, NULL, req, skb); 2752cfb6eeb4SYOSHIFUJI Hideaki } 2753cfb6eeb4SYOSHIFUJI Hideaki #endif 2754cfb6eeb4SYOSHIFUJI Hideaki 27551da177e4SLinus Torvalds return skb; 27561da177e4SLinus Torvalds } 27574bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_make_synack); 27581da177e4SLinus Torvalds 275967edfef7SAndi Kleen /* Do all connect socket setups that can be done AF independent. */ 2760370816aeSPavel Emelyanov void tcp_connect_init(struct sock *sk) 27611da177e4SLinus Torvalds { 2762cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 27631da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 27641da177e4SLinus Torvalds __u8 rcv_wscale; 27651da177e4SLinus Torvalds 27661da177e4SLinus Torvalds /* We'll fix this up when we get a response from the other end. 27671da177e4SLinus Torvalds * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 27681da177e4SLinus Torvalds */ 27691da177e4SLinus Torvalds tp->tcp_header_len = sizeof(struct tcphdr) + 2770bb5b7c11SDavid S. Miller (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0); 27711da177e4SLinus Torvalds 2772cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 2773cfb6eeb4SYOSHIFUJI Hideaki if (tp->af_specific->md5_lookup(sk, sk) != NULL) 2774cfb6eeb4SYOSHIFUJI Hideaki tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 2775cfb6eeb4SYOSHIFUJI Hideaki #endif 2776cfb6eeb4SYOSHIFUJI Hideaki 27771da177e4SLinus Torvalds /* If user gave his TCP_MAXSEG, record it to clamp */ 27781da177e4SLinus Torvalds if (tp->rx_opt.user_mss) 27791da177e4SLinus Torvalds tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 27801da177e4SLinus Torvalds tp->max_window = 0; 27815d424d5aSJohn Heffner tcp_mtup_init(sk); 27821da177e4SLinus Torvalds tcp_sync_mss(sk, dst_mtu(dst)); 27831da177e4SLinus Torvalds 27841da177e4SLinus Torvalds if (!tp->window_clamp) 27851da177e4SLinus Torvalds tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 27860dbaee3bSDavid S. Miller tp->advmss = dst_metric_advmss(dst); 2787f5fff5dcSTom Quetchenbach if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss) 2788f5fff5dcSTom Quetchenbach tp->advmss = tp->rx_opt.user_mss; 2789f5fff5dcSTom Quetchenbach 27901da177e4SLinus Torvalds tcp_initialize_rcv_mss(sk); 27911da177e4SLinus Torvalds 2792e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 2793e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 2794e88c64f0SHagen Paul Pfeifer (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) 2795e88c64f0SHagen Paul Pfeifer tp->window_clamp = tcp_full_space(sk); 2796e88c64f0SHagen Paul Pfeifer 27971da177e4SLinus Torvalds tcp_select_initial_window(tcp_full_space(sk), 27981da177e4SLinus Torvalds tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 27991da177e4SLinus Torvalds &tp->rcv_wnd, 28001da177e4SLinus Torvalds &tp->window_clamp, 2801bb5b7c11SDavid S. Miller sysctl_tcp_window_scaling, 280231d12926Slaurent chavey &rcv_wscale, 280331d12926Slaurent chavey dst_metric(dst, RTAX_INITRWND)); 28041da177e4SLinus Torvalds 28051da177e4SLinus Torvalds tp->rx_opt.rcv_wscale = rcv_wscale; 28061da177e4SLinus Torvalds tp->rcv_ssthresh = tp->rcv_wnd; 28071da177e4SLinus Torvalds 28081da177e4SLinus Torvalds sk->sk_err = 0; 28091da177e4SLinus Torvalds sock_reset_flag(sk, SOCK_DONE); 28101da177e4SLinus Torvalds tp->snd_wnd = 0; 2811ee7537b6SHantzis Fotis tcp_init_wl(tp, 0); 28121da177e4SLinus Torvalds tp->snd_una = tp->write_seq; 28131da177e4SLinus Torvalds tp->snd_sml = tp->write_seq; 281433f5f57eSIlpo Järvinen tp->snd_up = tp->write_seq; 2815370816aeSPavel Emelyanov tp->snd_nxt = tp->write_seq; 2816ee995283SPavel Emelyanov 2817ee995283SPavel Emelyanov if (likely(!tp->repair)) 28181da177e4SLinus Torvalds tp->rcv_nxt = 0; 2819ee995283SPavel Emelyanov tp->rcv_wup = tp->rcv_nxt; 2820ee995283SPavel Emelyanov tp->copied_seq = tp->rcv_nxt; 28211da177e4SLinus Torvalds 2822463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; 2823463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_retransmits = 0; 28241da177e4SLinus Torvalds tcp_clear_retrans(tp); 28251da177e4SLinus Torvalds } 28261da177e4SLinus Torvalds 282767edfef7SAndi Kleen /* Build a SYN and send it off. */ 28281da177e4SLinus Torvalds int tcp_connect(struct sock *sk) 28291da177e4SLinus Torvalds { 28301da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 28311da177e4SLinus Torvalds struct sk_buff *buff; 2832ee586811SEric Paris int err; 28331da177e4SLinus Torvalds 28341da177e4SLinus Torvalds tcp_connect_init(sk); 28351da177e4SLinus Torvalds 2836d179cd12SDavid S. Miller buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation); 28371da177e4SLinus Torvalds if (unlikely(buff == NULL)) 28381da177e4SLinus Torvalds return -ENOBUFS; 28391da177e4SLinus Torvalds 28401da177e4SLinus Torvalds /* Reserve space for headers. */ 28411da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 28421da177e4SLinus Torvalds 2843a3433f35SChangli Gao tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); 2844e870a8efSIlpo Järvinen TCP_ECN_send_syn(sk, buff); 28451da177e4SLinus Torvalds 28461da177e4SLinus Torvalds /* Send it off. */ 28471da177e4SLinus Torvalds TCP_SKB_CB(buff)->when = tcp_time_stamp; 28481da177e4SLinus Torvalds tp->retrans_stamp = TCP_SKB_CB(buff)->when; 28491da177e4SLinus Torvalds skb_header_release(buff); 2850fe067e8aSDavid S. Miller __tcp_add_write_queue_tail(sk, buff); 28513ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 28523ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 28531da177e4SLinus Torvalds tp->packets_out += tcp_skb_pcount(buff); 2854ee586811SEric Paris err = tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); 2855ee586811SEric Paris if (err == -ECONNREFUSED) 2856ee586811SEric Paris return err; 2857bd37a088SWei Yongjun 2858bd37a088SWei Yongjun /* We change tp->snd_nxt after the tcp_transmit_skb() call 2859bd37a088SWei Yongjun * in order to make this packet get counted in tcpOutSegs. 2860bd37a088SWei Yongjun */ 2861bd37a088SWei Yongjun tp->snd_nxt = tp->write_seq; 2862bd37a088SWei Yongjun tp->pushed_seq = tp->write_seq; 286381cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); 28641da177e4SLinus Torvalds 28651da177e4SLinus Torvalds /* Timer for repeating the SYN until an answer. */ 28663f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 28673f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 28681da177e4SLinus Torvalds return 0; 28691da177e4SLinus Torvalds } 28704bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_connect); 28711da177e4SLinus Torvalds 28721da177e4SLinus Torvalds /* Send out a delayed ack, the caller does the policy checking 28731da177e4SLinus Torvalds * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 28741da177e4SLinus Torvalds * for details. 28751da177e4SLinus Torvalds */ 28761da177e4SLinus Torvalds void tcp_send_delayed_ack(struct sock *sk) 28771da177e4SLinus Torvalds { 2878463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 2879463c84b9SArnaldo Carvalho de Melo int ato = icsk->icsk_ack.ato; 28801da177e4SLinus Torvalds unsigned long timeout; 28811da177e4SLinus Torvalds 28821da177e4SLinus Torvalds if (ato > TCP_DELACK_MIN) { 2883463c84b9SArnaldo Carvalho de Melo const struct tcp_sock *tp = tcp_sk(sk); 28841da177e4SLinus Torvalds int max_ato = HZ / 2; 28851da177e4SLinus Torvalds 2886056834d9SIlpo Järvinen if (icsk->icsk_ack.pingpong || 2887056834d9SIlpo Järvinen (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 28881da177e4SLinus Torvalds max_ato = TCP_DELACK_MAX; 28891da177e4SLinus Torvalds 28901da177e4SLinus Torvalds /* Slow path, intersegment interval is "high". */ 28911da177e4SLinus Torvalds 28921da177e4SLinus Torvalds /* If some rtt estimate is known, use it to bound delayed ack. 2893463c84b9SArnaldo Carvalho de Melo * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 28941da177e4SLinus Torvalds * directly. 28951da177e4SLinus Torvalds */ 28961da177e4SLinus Torvalds if (tp->srtt) { 28971da177e4SLinus Torvalds int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN); 28981da177e4SLinus Torvalds 28991da177e4SLinus Torvalds if (rtt < max_ato) 29001da177e4SLinus Torvalds max_ato = rtt; 29011da177e4SLinus Torvalds } 29021da177e4SLinus Torvalds 29031da177e4SLinus Torvalds ato = min(ato, max_ato); 29041da177e4SLinus Torvalds } 29051da177e4SLinus Torvalds 29061da177e4SLinus Torvalds /* Stay within the limit we were given */ 29071da177e4SLinus Torvalds timeout = jiffies + ato; 29081da177e4SLinus Torvalds 29091da177e4SLinus Torvalds /* Use new timeout only if there wasn't a older one earlier. */ 2910463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 29111da177e4SLinus Torvalds /* If delack timer was blocked or is about to expire, 29121da177e4SLinus Torvalds * send ACK now. 29131da177e4SLinus Torvalds */ 2914463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.blocked || 2915463c84b9SArnaldo Carvalho de Melo time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 29161da177e4SLinus Torvalds tcp_send_ack(sk); 29171da177e4SLinus Torvalds return; 29181da177e4SLinus Torvalds } 29191da177e4SLinus Torvalds 2920463c84b9SArnaldo Carvalho de Melo if (!time_before(timeout, icsk->icsk_ack.timeout)) 2921463c84b9SArnaldo Carvalho de Melo timeout = icsk->icsk_ack.timeout; 29221da177e4SLinus Torvalds } 2923463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 2924463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.timeout = timeout; 2925463c84b9SArnaldo Carvalho de Melo sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 29261da177e4SLinus Torvalds } 29271da177e4SLinus Torvalds 29281da177e4SLinus Torvalds /* This routine sends an ack and also updates the window. */ 29291da177e4SLinus Torvalds void tcp_send_ack(struct sock *sk) 29301da177e4SLinus Torvalds { 29311da177e4SLinus Torvalds struct sk_buff *buff; 29321da177e4SLinus Torvalds 2933058dc334SIlpo Järvinen /* If we have been reset, we may not send again. */ 2934058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 2935058dc334SIlpo Järvinen return; 2936058dc334SIlpo Järvinen 29371da177e4SLinus Torvalds /* We are not putting this on the write queue, so 29381da177e4SLinus Torvalds * tcp_transmit_skb() will set the ownership to this 29391da177e4SLinus Torvalds * sock. 29401da177e4SLinus Torvalds */ 29411da177e4SLinus Torvalds buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 29421da177e4SLinus Torvalds if (buff == NULL) { 2943463c84b9SArnaldo Carvalho de Melo inet_csk_schedule_ack(sk); 2944463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; 29453f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 29463f421baaSArnaldo Carvalho de Melo TCP_DELACK_MAX, TCP_RTO_MAX); 29471da177e4SLinus Torvalds return; 29481da177e4SLinus Torvalds } 29491da177e4SLinus Torvalds 29501da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 29511da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 2952a3433f35SChangli Gao tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); 29531da177e4SLinus Torvalds 29541da177e4SLinus Torvalds /* Send it off, this clears delayed acks for us. */ 29551da177e4SLinus Torvalds TCP_SKB_CB(buff)->when = tcp_time_stamp; 2956dfb4b9dcSDavid S. Miller tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC); 29571da177e4SLinus Torvalds } 29581da177e4SLinus Torvalds 29591da177e4SLinus Torvalds /* This routine sends a packet with an out of date sequence 29601da177e4SLinus Torvalds * number. It assumes the other end will try to ack it. 29611da177e4SLinus Torvalds * 29621da177e4SLinus Torvalds * Question: what should we make while urgent mode? 29631da177e4SLinus Torvalds * 4.4BSD forces sending single byte of data. We cannot send 29641da177e4SLinus Torvalds * out of window data, because we have SND.NXT==SND.MAX... 29651da177e4SLinus Torvalds * 29661da177e4SLinus Torvalds * Current solution: to send TWO zero-length segments in urgent mode: 29671da177e4SLinus Torvalds * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 29681da177e4SLinus Torvalds * out-of-date with SND.UNA-1 to probe window. 29691da177e4SLinus Torvalds */ 29701da177e4SLinus Torvalds static int tcp_xmit_probe_skb(struct sock *sk, int urgent) 29711da177e4SLinus Torvalds { 29721da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 29731da177e4SLinus Torvalds struct sk_buff *skb; 29741da177e4SLinus Torvalds 29751da177e4SLinus Torvalds /* We don't queue it, tcp_transmit_skb() sets ownership. */ 29761da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 29771da177e4SLinus Torvalds if (skb == NULL) 29781da177e4SLinus Torvalds return -1; 29791da177e4SLinus Torvalds 29801da177e4SLinus Torvalds /* Reserve space for headers and set control bits. */ 29811da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 29821da177e4SLinus Torvalds /* Use a previous sequence. This should cause the other 29831da177e4SLinus Torvalds * end to send an ack. Don't queue or clone SKB, just 29841da177e4SLinus Torvalds * send it. 29851da177e4SLinus Torvalds */ 2986a3433f35SChangli Gao tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); 29871da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2988dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC); 29891da177e4SLinus Torvalds } 29901da177e4SLinus Torvalds 2991ee995283SPavel Emelyanov void tcp_send_window_probe(struct sock *sk) 2992ee995283SPavel Emelyanov { 2993ee995283SPavel Emelyanov if (sk->sk_state == TCP_ESTABLISHED) { 2994ee995283SPavel Emelyanov tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; 2995c0e88ff0SPavel Emelyanov tcp_sk(sk)->snd_nxt = tcp_sk(sk)->write_seq; 2996ee995283SPavel Emelyanov tcp_xmit_probe_skb(sk, 0); 2997ee995283SPavel Emelyanov } 2998ee995283SPavel Emelyanov } 2999ee995283SPavel Emelyanov 300067edfef7SAndi Kleen /* Initiate keepalive or window probe from timer. */ 30011da177e4SLinus Torvalds int tcp_write_wakeup(struct sock *sk) 30021da177e4SLinus Torvalds { 30031da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 30041da177e4SLinus Torvalds struct sk_buff *skb; 30051da177e4SLinus Torvalds 3006058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3007058dc334SIlpo Järvinen return -1; 3008058dc334SIlpo Järvinen 3009fe067e8aSDavid S. Miller if ((skb = tcp_send_head(sk)) != NULL && 301090840defSIlpo Järvinen before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { 30111da177e4SLinus Torvalds int err; 30120c54b85fSIlpo Järvinen unsigned int mss = tcp_current_mss(sk); 301390840defSIlpo Järvinen unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 30141da177e4SLinus Torvalds 30151da177e4SLinus Torvalds if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 30161da177e4SLinus Torvalds tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 30171da177e4SLinus Torvalds 30181da177e4SLinus Torvalds /* We are probing the opening of a window 30191da177e4SLinus Torvalds * but the window size is != 0 30201da177e4SLinus Torvalds * must have been a result SWS avoidance ( sender ) 30211da177e4SLinus Torvalds */ 30221da177e4SLinus Torvalds if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 30231da177e4SLinus Torvalds skb->len > mss) { 30241da177e4SLinus Torvalds seg_size = min(seg_size, mss); 30254de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 3026846998aeSDavid S. Miller if (tcp_fragment(sk, skb, seg_size, mss)) 30271da177e4SLinus Torvalds return -1; 30281da177e4SLinus Torvalds } else if (!tcp_skb_pcount(skb)) 3029846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss); 30301da177e4SLinus Torvalds 30314de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 30321da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 3033dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 303466f5fe62SIlpo Järvinen if (!err) 303566f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 30361da177e4SLinus Torvalds return err; 30371da177e4SLinus Torvalds } else { 303833f5f57eSIlpo Järvinen if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) 30394828e7f4SIlpo Järvinen tcp_xmit_probe_skb(sk, 1); 30401da177e4SLinus Torvalds return tcp_xmit_probe_skb(sk, 0); 30411da177e4SLinus Torvalds } 30421da177e4SLinus Torvalds } 30431da177e4SLinus Torvalds 30441da177e4SLinus Torvalds /* A window probe timeout has occurred. If window is not closed send 30451da177e4SLinus Torvalds * a partial packet else a zero probe. 30461da177e4SLinus Torvalds */ 30471da177e4SLinus Torvalds void tcp_send_probe0(struct sock *sk) 30481da177e4SLinus Torvalds { 3049463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 30501da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 30511da177e4SLinus Torvalds int err; 30521da177e4SLinus Torvalds 30531da177e4SLinus Torvalds err = tcp_write_wakeup(sk); 30541da177e4SLinus Torvalds 3055fe067e8aSDavid S. Miller if (tp->packets_out || !tcp_send_head(sk)) { 30561da177e4SLinus Torvalds /* Cancel probe timer, if it is not required. */ 30576687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 0; 3058463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff = 0; 30591da177e4SLinus Torvalds return; 30601da177e4SLinus Torvalds } 30611da177e4SLinus Torvalds 30621da177e4SLinus Torvalds if (err <= 0) { 3063463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_backoff < sysctl_tcp_retries2) 3064463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff++; 30656687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out++; 3066463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 30673f421baaSArnaldo Carvalho de Melo min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), 30683f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 30691da177e4SLinus Torvalds } else { 30701da177e4SLinus Torvalds /* If packet was not sent due to local congestion, 30716687e988SArnaldo Carvalho de Melo * do not backoff and do not remember icsk_probes_out. 30721da177e4SLinus Torvalds * Let local senders to fight for local resources. 30731da177e4SLinus Torvalds * 30741da177e4SLinus Torvalds * Use accumulated backoff yet. 30751da177e4SLinus Torvalds */ 30766687e988SArnaldo Carvalho de Melo if (!icsk->icsk_probes_out) 30776687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 1; 3078463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 3079463c84b9SArnaldo Carvalho de Melo min(icsk->icsk_rto << icsk->icsk_backoff, 30803f421baaSArnaldo Carvalho de Melo TCP_RESOURCE_PROBE_INTERVAL), 30813f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 30821da177e4SLinus Torvalds } 30831da177e4SLinus Torvalds } 3084