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) 3882100c8d2SYuchung Cheng #define OPTION_FAST_OPEN_COOKIE (1 << 8) 38933ad798cSAdam Langley 39033ad798cSAdam Langley struct tcp_out_options { 3912100c8d2SYuchung Cheng u16 options; /* bit field of OPTION_* */ 3922100c8d2SYuchung 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 */ 3972100c8d2SYuchung Cheng __u32 tsval, tsecr; /* need to include OPTION_TS */ 3982100c8d2SYuchung 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 { 4472100c8d2SYuchung 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 } 5692100c8d2SYuchung Cheng 5702100c8d2SYuchung Cheng if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) { 5712100c8d2SYuchung Cheng struct tcp_fastopen_cookie *foc = opts->fastopen_cookie; 5722100c8d2SYuchung Cheng 5732100c8d2SYuchung Cheng *ptr++ = htonl((TCPOPT_EXP << 24) | 5742100c8d2SYuchung Cheng ((TCPOLEN_EXP_FASTOPEN_BASE + foc->len) << 16) | 5752100c8d2SYuchung Cheng TCPOPT_FASTOPEN_MAGIC); 5762100c8d2SYuchung Cheng 5772100c8d2SYuchung Cheng memcpy(ptr, foc->val, foc->len); 5782100c8d2SYuchung Cheng if ((foc->len & 3) == 2) { 5792100c8d2SYuchung Cheng u8 *align = ((u8 *)ptr) + foc->len; 5802100c8d2SYuchung Cheng align[0] = align[1] = TCPOPT_NOP; 5812100c8d2SYuchung Cheng } 5822100c8d2SYuchung Cheng ptr += (foc->len + 3) >> 2; 5832100c8d2SYuchung 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; 599783237e8SYuchung Cheng struct tcp_fastopen_request *fastopen = tp->fastopen_req; 60033ad798cSAdam Langley 601cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 60233ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 60333ad798cSAdam Langley if (*md5) { 60433ad798cSAdam Langley opts->options |= OPTION_MD5; 605bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 606cfb6eeb4SYOSHIFUJI Hideaki } 60733ad798cSAdam Langley #else 60833ad798cSAdam Langley *md5 = NULL; 609cfb6eeb4SYOSHIFUJI Hideaki #endif 61033ad798cSAdam Langley 61133ad798cSAdam Langley /* We always get an MSS option. The option bytes which will be seen in 61233ad798cSAdam Langley * normal data packets should timestamps be used, must be in the MSS 61333ad798cSAdam Langley * advertised. But we subtract them from tp->mss_cache so that 61433ad798cSAdam Langley * calculations in tcp_sendmsg are simpler etc. So account for this 61533ad798cSAdam Langley * fact here if necessary. If we don't do this correctly, as a 61633ad798cSAdam Langley * receiver we won't recognize data packets as being full sized when we 61733ad798cSAdam Langley * should, and thus we won't abide by the delayed ACK rules correctly. 61833ad798cSAdam Langley * SACKs don't matter, we never delay an ACK when we have any of those 61933ad798cSAdam Langley * going out. */ 62033ad798cSAdam Langley opts->mss = tcp_advertise_mss(sk); 621bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 62233ad798cSAdam Langley 623bb5b7c11SDavid S. Miller if (likely(sysctl_tcp_timestamps && *md5 == NULL)) { 62433ad798cSAdam Langley opts->options |= OPTION_TS; 62533ad798cSAdam Langley opts->tsval = TCP_SKB_CB(skb)->when; 62633ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 627bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 62833ad798cSAdam Langley } 629bb5b7c11SDavid S. Miller if (likely(sysctl_tcp_window_scaling)) { 63033ad798cSAdam Langley opts->ws = tp->rx_opt.rcv_wscale; 63189e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 632bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 63333ad798cSAdam Langley } 634bb5b7c11SDavid S. Miller if (likely(sysctl_tcp_sack)) { 63533ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 636b32d1310SDavid S. Miller if (unlikely(!(OPTION_TS & opts->options))) 637bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 63833ad798cSAdam Langley } 63933ad798cSAdam Langley 640783237e8SYuchung Cheng if (fastopen && fastopen->cookie.len >= 0) { 641783237e8SYuchung Cheng u32 need = TCPOLEN_EXP_FASTOPEN_BASE + fastopen->cookie.len; 642783237e8SYuchung Cheng need = (need + 3) & ~3U; /* Align to 32 bits */ 643783237e8SYuchung Cheng if (remaining >= need) { 644783237e8SYuchung Cheng opts->options |= OPTION_FAST_OPEN_COOKIE; 645783237e8SYuchung Cheng opts->fastopen_cookie = &fastopen->cookie; 646783237e8SYuchung Cheng remaining -= need; 647783237e8SYuchung Cheng tp->syn_fastopen = 1; 648783237e8SYuchung Cheng } 649783237e8SYuchung Cheng } 650bd0388aeSWilliam Allen Simpson /* Note that timestamps are required by the specification. 651bd0388aeSWilliam Allen Simpson * 652bd0388aeSWilliam Allen Simpson * Odd numbers of bytes are prohibited by the specification, ensuring 653bd0388aeSWilliam Allen Simpson * that the cookie is 16-bit aligned, and the resulting cookie pair is 654bd0388aeSWilliam Allen Simpson * 32-bit aligned. 655bd0388aeSWilliam Allen Simpson */ 656bd0388aeSWilliam Allen Simpson if (*md5 == NULL && 657bd0388aeSWilliam Allen Simpson (OPTION_TS & opts->options) && 658bd0388aeSWilliam Allen Simpson cookie_size > 0) { 659bd0388aeSWilliam Allen Simpson int need = TCPOLEN_COOKIE_BASE + cookie_size; 660bd0388aeSWilliam Allen Simpson 661bd0388aeSWilliam Allen Simpson if (0x2 & need) { 662bd0388aeSWilliam Allen Simpson /* 32-bit multiple */ 663bd0388aeSWilliam Allen Simpson need += 2; /* NOPs */ 664bd0388aeSWilliam Allen Simpson 665bd0388aeSWilliam Allen Simpson if (need > remaining) { 666bd0388aeSWilliam Allen Simpson /* try shrinking cookie to fit */ 667bd0388aeSWilliam Allen Simpson cookie_size -= 2; 668bd0388aeSWilliam Allen Simpson need -= 4; 669bd0388aeSWilliam Allen Simpson } 670bd0388aeSWilliam Allen Simpson } 671bd0388aeSWilliam Allen Simpson while (need > remaining && TCP_COOKIE_MIN <= cookie_size) { 672bd0388aeSWilliam Allen Simpson cookie_size -= 4; 673bd0388aeSWilliam Allen Simpson need -= 4; 674bd0388aeSWilliam Allen Simpson } 675bd0388aeSWilliam Allen Simpson if (TCP_COOKIE_MIN <= cookie_size) { 676bd0388aeSWilliam Allen Simpson opts->options |= OPTION_COOKIE_EXTENSION; 677bd0388aeSWilliam Allen Simpson opts->hash_location = (__u8 *)&cvp->cookie_pair[0]; 678bd0388aeSWilliam Allen Simpson opts->hash_size = cookie_size; 679bd0388aeSWilliam Allen Simpson 680bd0388aeSWilliam Allen Simpson /* Remember for future incarnations. */ 681bd0388aeSWilliam Allen Simpson cvp->cookie_desired = cookie_size; 682bd0388aeSWilliam Allen Simpson 683bd0388aeSWilliam Allen Simpson if (cvp->cookie_desired != cvp->cookie_pair_size) { 684bd0388aeSWilliam Allen Simpson /* Currently use random bytes as a nonce, 685bd0388aeSWilliam Allen Simpson * assuming these are completely unpredictable 686bd0388aeSWilliam Allen Simpson * by hostile users of the same system. 687bd0388aeSWilliam Allen Simpson */ 688bd0388aeSWilliam Allen Simpson get_random_bytes(&cvp->cookie_pair[0], 689bd0388aeSWilliam Allen Simpson cookie_size); 690bd0388aeSWilliam Allen Simpson cvp->cookie_pair_size = cookie_size; 691bd0388aeSWilliam Allen Simpson } 692bd0388aeSWilliam Allen Simpson 693bd0388aeSWilliam Allen Simpson remaining -= need; 694bd0388aeSWilliam Allen Simpson } 695bd0388aeSWilliam Allen Simpson } 696bd0388aeSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 69733ad798cSAdam Langley } 69833ad798cSAdam Langley 69967edfef7SAndi Kleen /* Set up TCP options for SYN-ACKs. */ 70095c96174SEric Dumazet static unsigned int tcp_synack_options(struct sock *sk, 70133ad798cSAdam Langley struct request_sock *req, 70295c96174SEric Dumazet unsigned int mss, struct sk_buff *skb, 70333ad798cSAdam Langley struct tcp_out_options *opts, 7044957faadSWilliam Allen Simpson struct tcp_md5sig_key **md5, 7054957faadSWilliam Allen Simpson struct tcp_extend_values *xvp) 7064957faadSWilliam Allen Simpson { 70733ad798cSAdam Langley struct inet_request_sock *ireq = inet_rsk(req); 70895c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 7094957faadSWilliam Allen Simpson u8 cookie_plus = (xvp != NULL && !xvp->cookie_out_never) ? 7104957faadSWilliam Allen Simpson xvp->cookie_plus : 7114957faadSWilliam Allen Simpson 0; 71233ad798cSAdam Langley 71333ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 71433ad798cSAdam Langley *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req); 71533ad798cSAdam Langley if (*md5) { 71633ad798cSAdam Langley opts->options |= OPTION_MD5; 7174957faadSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 7184957faadSWilliam Allen Simpson 7194957faadSWilliam Allen Simpson /* We can't fit any SACK blocks in a packet with MD5 + TS 7204957faadSWilliam Allen Simpson * options. There was discussion about disabling SACK 7214957faadSWilliam Allen Simpson * rather than TS in order to fit in better with old, 7224957faadSWilliam Allen Simpson * buggy kernels, but that was deemed to be unnecessary. 7234957faadSWilliam Allen Simpson */ 724de213e5eSEric Dumazet ireq->tstamp_ok &= !ireq->sack_ok; 72533ad798cSAdam Langley } 72633ad798cSAdam Langley #else 72733ad798cSAdam Langley *md5 = NULL; 72833ad798cSAdam Langley #endif 72933ad798cSAdam Langley 7304957faadSWilliam Allen Simpson /* We always send an MSS option. */ 73133ad798cSAdam Langley opts->mss = mss; 7324957faadSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 73333ad798cSAdam Langley 73433ad798cSAdam Langley if (likely(ireq->wscale_ok)) { 73533ad798cSAdam Langley opts->ws = ireq->rcv_wscale; 73689e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 7374957faadSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 73833ad798cSAdam Langley } 739de213e5eSEric Dumazet if (likely(ireq->tstamp_ok)) { 74033ad798cSAdam Langley opts->options |= OPTION_TS; 74133ad798cSAdam Langley opts->tsval = TCP_SKB_CB(skb)->when; 74233ad798cSAdam Langley opts->tsecr = req->ts_recent; 7434957faadSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 74433ad798cSAdam Langley } 74533ad798cSAdam Langley if (likely(ireq->sack_ok)) { 74633ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 747de213e5eSEric Dumazet if (unlikely(!ireq->tstamp_ok)) 7484957faadSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 74933ad798cSAdam Langley } 75033ad798cSAdam Langley 7514957faadSWilliam Allen Simpson /* Similar rationale to tcp_syn_options() applies here, too. 7524957faadSWilliam Allen Simpson * If the <SYN> options fit, the same options should fit now! 7534957faadSWilliam Allen Simpson */ 7544957faadSWilliam Allen Simpson if (*md5 == NULL && 755de213e5eSEric Dumazet ireq->tstamp_ok && 7564957faadSWilliam Allen Simpson cookie_plus > TCPOLEN_COOKIE_BASE) { 7574957faadSWilliam Allen Simpson int need = cookie_plus; /* has TCPOLEN_COOKIE_BASE */ 7584957faadSWilliam Allen Simpson 7594957faadSWilliam Allen Simpson if (0x2 & need) { 7604957faadSWilliam Allen Simpson /* 32-bit multiple */ 7614957faadSWilliam Allen Simpson need += 2; /* NOPs */ 7624957faadSWilliam Allen Simpson } 7634957faadSWilliam Allen Simpson if (need <= remaining) { 7644957faadSWilliam Allen Simpson opts->options |= OPTION_COOKIE_EXTENSION; 7654957faadSWilliam Allen Simpson opts->hash_size = cookie_plus - TCPOLEN_COOKIE_BASE; 7664957faadSWilliam Allen Simpson remaining -= need; 7674957faadSWilliam Allen Simpson } else { 7684957faadSWilliam Allen Simpson /* There's no error return, so flag it. */ 7694957faadSWilliam Allen Simpson xvp->cookie_out_never = 1; /* true */ 7704957faadSWilliam Allen Simpson opts->hash_size = 0; 7714957faadSWilliam Allen Simpson } 7724957faadSWilliam Allen Simpson } 7734957faadSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 77433ad798cSAdam Langley } 77533ad798cSAdam Langley 77667edfef7SAndi Kleen /* Compute TCP options for ESTABLISHED sockets. This is not the 77767edfef7SAndi Kleen * final wire format yet. 77867edfef7SAndi Kleen */ 77995c96174SEric Dumazet static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb, 78033ad798cSAdam Langley struct tcp_out_options *opts, 781cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 782cf533ea5SEric Dumazet { 78333ad798cSAdam Langley struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL; 78433ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 78595c96174SEric Dumazet unsigned int size = 0; 786cabeccbdSIlpo Järvinen unsigned int eff_sacks; 78733ad798cSAdam Langley 78833ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 78933ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 79033ad798cSAdam Langley if (unlikely(*md5)) { 79133ad798cSAdam Langley opts->options |= OPTION_MD5; 79233ad798cSAdam Langley size += TCPOLEN_MD5SIG_ALIGNED; 79333ad798cSAdam Langley } 79433ad798cSAdam Langley #else 79533ad798cSAdam Langley *md5 = NULL; 79633ad798cSAdam Langley #endif 79733ad798cSAdam Langley 79833ad798cSAdam Langley if (likely(tp->rx_opt.tstamp_ok)) { 79933ad798cSAdam Langley opts->options |= OPTION_TS; 80033ad798cSAdam Langley opts->tsval = tcb ? tcb->when : 0; 80133ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 80233ad798cSAdam Langley size += TCPOLEN_TSTAMP_ALIGNED; 80333ad798cSAdam Langley } 80433ad798cSAdam Langley 805cabeccbdSIlpo Järvinen eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; 806cabeccbdSIlpo Järvinen if (unlikely(eff_sacks)) { 80795c96174SEric Dumazet const unsigned int remaining = MAX_TCP_OPTION_SPACE - size; 80833ad798cSAdam Langley opts->num_sack_blocks = 80995c96174SEric Dumazet min_t(unsigned int, eff_sacks, 81033ad798cSAdam Langley (remaining - TCPOLEN_SACK_BASE_ALIGNED) / 81133ad798cSAdam Langley TCPOLEN_SACK_PERBLOCK); 81233ad798cSAdam Langley size += TCPOLEN_SACK_BASE_ALIGNED + 81333ad798cSAdam Langley opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; 81433ad798cSAdam Langley } 81533ad798cSAdam Langley 81633ad798cSAdam Langley return size; 81740efc6faSStephen Hemminger } 8181da177e4SLinus Torvalds 81946d3ceabSEric Dumazet 82046d3ceabSEric Dumazet /* TCP SMALL QUEUES (TSQ) 82146d3ceabSEric Dumazet * 82246d3ceabSEric Dumazet * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev) 82346d3ceabSEric Dumazet * to reduce RTT and bufferbloat. 82446d3ceabSEric Dumazet * We do this using a special skb destructor (tcp_wfree). 82546d3ceabSEric Dumazet * 82646d3ceabSEric Dumazet * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb 82746d3ceabSEric Dumazet * needs to be reallocated in a driver. 82846d3ceabSEric Dumazet * The invariant being skb->truesize substracted from sk->sk_wmem_alloc 82946d3ceabSEric Dumazet * 83046d3ceabSEric Dumazet * Since transmit from skb destructor is forbidden, we use a tasklet 83146d3ceabSEric Dumazet * to process all sockets that eventually need to send more skbs. 83246d3ceabSEric Dumazet * We use one tasklet per cpu, with its own queue of sockets. 83346d3ceabSEric Dumazet */ 83446d3ceabSEric Dumazet struct tsq_tasklet { 83546d3ceabSEric Dumazet struct tasklet_struct tasklet; 83646d3ceabSEric Dumazet struct list_head head; /* queue of tcp sockets */ 83746d3ceabSEric Dumazet }; 83846d3ceabSEric Dumazet static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet); 83946d3ceabSEric Dumazet 84046d3ceabSEric Dumazet /* 84146d3ceabSEric Dumazet * One tasklest per cpu tries to send more skbs. 84246d3ceabSEric Dumazet * We run in tasklet context but need to disable irqs when 84346d3ceabSEric Dumazet * transfering tsq->head because tcp_wfree() might 84446d3ceabSEric Dumazet * interrupt us (non NAPI drivers) 84546d3ceabSEric Dumazet */ 84646d3ceabSEric Dumazet static void tcp_tasklet_func(unsigned long data) 84746d3ceabSEric Dumazet { 84846d3ceabSEric Dumazet struct tsq_tasklet *tsq = (struct tsq_tasklet *)data; 84946d3ceabSEric Dumazet LIST_HEAD(list); 85046d3ceabSEric Dumazet unsigned long flags; 85146d3ceabSEric Dumazet struct list_head *q, *n; 85246d3ceabSEric Dumazet struct tcp_sock *tp; 85346d3ceabSEric Dumazet struct sock *sk; 85446d3ceabSEric Dumazet 85546d3ceabSEric Dumazet local_irq_save(flags); 85646d3ceabSEric Dumazet list_splice_init(&tsq->head, &list); 85746d3ceabSEric Dumazet local_irq_restore(flags); 85846d3ceabSEric Dumazet 85946d3ceabSEric Dumazet list_for_each_safe(q, n, &list) { 86046d3ceabSEric Dumazet tp = list_entry(q, struct tcp_sock, tsq_node); 86146d3ceabSEric Dumazet list_del(&tp->tsq_node); 86246d3ceabSEric Dumazet 86346d3ceabSEric Dumazet sk = (struct sock *)tp; 86446d3ceabSEric Dumazet bh_lock_sock(sk); 86546d3ceabSEric Dumazet 86646d3ceabSEric Dumazet if (!sock_owned_by_user(sk)) { 86746d3ceabSEric Dumazet if ((1 << sk->sk_state) & 86846d3ceabSEric Dumazet (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | 869d01cb207SEric Dumazet TCPF_CLOSING | TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) 87046d3ceabSEric Dumazet tcp_write_xmit(sk, 87146d3ceabSEric Dumazet tcp_current_mss(sk), 87246d3ceabSEric Dumazet 0, 0, 87346d3ceabSEric Dumazet GFP_ATOMIC); 87446d3ceabSEric Dumazet } else { 87546d3ceabSEric Dumazet /* defer the work to tcp_release_cb() */ 87646d3ceabSEric Dumazet set_bit(TSQ_OWNED, &tp->tsq_flags); 87746d3ceabSEric Dumazet } 87846d3ceabSEric Dumazet bh_unlock_sock(sk); 87946d3ceabSEric Dumazet 88046d3ceabSEric Dumazet clear_bit(TSQ_QUEUED, &tp->tsq_flags); 88146d3ceabSEric Dumazet sk_free(sk); 88246d3ceabSEric Dumazet } 88346d3ceabSEric Dumazet } 88446d3ceabSEric Dumazet 88546d3ceabSEric Dumazet /** 88646d3ceabSEric Dumazet * tcp_release_cb - tcp release_sock() callback 88746d3ceabSEric Dumazet * @sk: socket 88846d3ceabSEric Dumazet * 88946d3ceabSEric Dumazet * called from release_sock() to perform protocol dependent 89046d3ceabSEric Dumazet * actions before socket release. 89146d3ceabSEric Dumazet */ 89246d3ceabSEric Dumazet void tcp_release_cb(struct sock *sk) 89346d3ceabSEric Dumazet { 89446d3ceabSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 89546d3ceabSEric Dumazet 89646d3ceabSEric Dumazet if (test_and_clear_bit(TSQ_OWNED, &tp->tsq_flags)) { 89746d3ceabSEric Dumazet if ((1 << sk->sk_state) & 89846d3ceabSEric Dumazet (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | 899d01cb207SEric Dumazet TCPF_CLOSING | TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) 90046d3ceabSEric Dumazet tcp_write_xmit(sk, 90146d3ceabSEric Dumazet tcp_current_mss(sk), 90246d3ceabSEric Dumazet 0, 0, 90346d3ceabSEric Dumazet GFP_ATOMIC); 90446d3ceabSEric Dumazet } 90546d3ceabSEric Dumazet } 90646d3ceabSEric Dumazet EXPORT_SYMBOL(tcp_release_cb); 90746d3ceabSEric Dumazet 90846d3ceabSEric Dumazet void __init tcp_tasklet_init(void) 90946d3ceabSEric Dumazet { 91046d3ceabSEric Dumazet int i; 91146d3ceabSEric Dumazet 91246d3ceabSEric Dumazet for_each_possible_cpu(i) { 91346d3ceabSEric Dumazet struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i); 91446d3ceabSEric Dumazet 91546d3ceabSEric Dumazet INIT_LIST_HEAD(&tsq->head); 91646d3ceabSEric Dumazet tasklet_init(&tsq->tasklet, 91746d3ceabSEric Dumazet tcp_tasklet_func, 91846d3ceabSEric Dumazet (unsigned long)tsq); 91946d3ceabSEric Dumazet } 92046d3ceabSEric Dumazet } 92146d3ceabSEric Dumazet 92246d3ceabSEric Dumazet /* 92346d3ceabSEric Dumazet * Write buffer destructor automatically called from kfree_skb. 92446d3ceabSEric Dumazet * We cant xmit new skbs from this context, as we might already 92546d3ceabSEric Dumazet * hold qdisc lock. 92646d3ceabSEric Dumazet */ 92746d3ceabSEric Dumazet void tcp_wfree(struct sk_buff *skb) 92846d3ceabSEric Dumazet { 92946d3ceabSEric Dumazet struct sock *sk = skb->sk; 93046d3ceabSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 93146d3ceabSEric Dumazet 93246d3ceabSEric Dumazet if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) && 93346d3ceabSEric Dumazet !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) { 93446d3ceabSEric Dumazet unsigned long flags; 93546d3ceabSEric Dumazet struct tsq_tasklet *tsq; 93646d3ceabSEric Dumazet 93746d3ceabSEric Dumazet /* Keep a ref on socket. 93846d3ceabSEric Dumazet * This last ref will be released in tcp_tasklet_func() 93946d3ceabSEric Dumazet */ 94046d3ceabSEric Dumazet atomic_sub(skb->truesize - 1, &sk->sk_wmem_alloc); 94146d3ceabSEric Dumazet 94246d3ceabSEric Dumazet /* queue this socket to tasklet queue */ 94346d3ceabSEric Dumazet local_irq_save(flags); 94446d3ceabSEric Dumazet tsq = &__get_cpu_var(tsq_tasklet); 94546d3ceabSEric Dumazet list_add(&tp->tsq_node, &tsq->head); 94646d3ceabSEric Dumazet tasklet_schedule(&tsq->tasklet); 94746d3ceabSEric Dumazet local_irq_restore(flags); 94846d3ceabSEric Dumazet } else { 94946d3ceabSEric Dumazet sock_wfree(skb); 95046d3ceabSEric Dumazet } 95146d3ceabSEric Dumazet } 95246d3ceabSEric Dumazet 9531da177e4SLinus Torvalds /* This routine actually transmits TCP packets queued in by 9541da177e4SLinus Torvalds * tcp_do_sendmsg(). This is used by both the initial 9551da177e4SLinus Torvalds * transmission and possible later retransmissions. 9561da177e4SLinus Torvalds * All SKB's seen here are completely headerless. It is our 9571da177e4SLinus Torvalds * job to build the TCP header, and pass the packet down to 9581da177e4SLinus Torvalds * IP so it can do the same plus pass the packet off to the 9591da177e4SLinus Torvalds * device. 9601da177e4SLinus Torvalds * 9611da177e4SLinus Torvalds * We are working here with either a clone of the original 9621da177e4SLinus Torvalds * SKB, or a fresh unique copy made by the retransmit engine. 9631da177e4SLinus Torvalds */ 964056834d9SIlpo Järvinen static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, 965056834d9SIlpo Järvinen gfp_t gfp_mask) 9661da177e4SLinus Torvalds { 9676687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 968dfb4b9dcSDavid S. Miller struct inet_sock *inet; 969dfb4b9dcSDavid S. Miller struct tcp_sock *tp; 970dfb4b9dcSDavid S. Miller struct tcp_skb_cb *tcb; 97133ad798cSAdam Langley struct tcp_out_options opts; 97295c96174SEric Dumazet unsigned int tcp_options_size, tcp_header_size; 973cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 9741da177e4SLinus Torvalds struct tcphdr *th; 9751da177e4SLinus Torvalds int err; 9761da177e4SLinus Torvalds 977dfb4b9dcSDavid S. Miller BUG_ON(!skb || !tcp_skb_pcount(skb)); 978dfb4b9dcSDavid S. Miller 979dfb4b9dcSDavid S. Miller /* If congestion control is doing timestamping, we must 980dfb4b9dcSDavid S. Miller * take such a timestamp before we potentially clone/copy. 981dfb4b9dcSDavid S. Miller */ 982164891aaSStephen Hemminger if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP) 983dfb4b9dcSDavid S. Miller __net_timestamp(skb); 984dfb4b9dcSDavid S. Miller 985dfb4b9dcSDavid S. Miller if (likely(clone_it)) { 986dfb4b9dcSDavid S. Miller if (unlikely(skb_cloned(skb))) 987dfb4b9dcSDavid S. Miller skb = pskb_copy(skb, gfp_mask); 988dfb4b9dcSDavid S. Miller else 989dfb4b9dcSDavid S. Miller skb = skb_clone(skb, gfp_mask); 990dfb4b9dcSDavid S. Miller if (unlikely(!skb)) 991dfb4b9dcSDavid S. Miller return -ENOBUFS; 992dfb4b9dcSDavid S. Miller } 993dfb4b9dcSDavid S. Miller 994dfb4b9dcSDavid S. Miller inet = inet_sk(sk); 995dfb4b9dcSDavid S. Miller tp = tcp_sk(sk); 996dfb4b9dcSDavid S. Miller tcb = TCP_SKB_CB(skb); 99733ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 9981da177e4SLinus Torvalds 9994de075e0SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) 100033ad798cSAdam Langley tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); 100133ad798cSAdam Langley else 100233ad798cSAdam Langley tcp_options_size = tcp_established_options(sk, skb, &opts, 100333ad798cSAdam Langley &md5); 100433ad798cSAdam Langley tcp_header_size = tcp_options_size + sizeof(struct tcphdr); 10051da177e4SLinus Torvalds 10063853b584STom Herbert if (tcp_packets_in_flight(tp) == 0) { 10076687e988SArnaldo Carvalho de Melo tcp_ca_event(sk, CA_EVENT_TX_START); 10083853b584STom Herbert skb->ooo_okay = 1; 10093853b584STom Herbert } else 10103853b584STom Herbert skb->ooo_okay = 0; 10111da177e4SLinus Torvalds 1012aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 1013aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 101446d3ceabSEric Dumazet 101546d3ceabSEric Dumazet skb_orphan(skb); 101646d3ceabSEric Dumazet skb->sk = sk; 101746d3ceabSEric Dumazet skb->destructor = (sysctl_tcp_limit_output_bytes > 0) ? 101846d3ceabSEric Dumazet tcp_wfree : sock_wfree; 101946d3ceabSEric Dumazet atomic_add(skb->truesize, &sk->sk_wmem_alloc); 10201da177e4SLinus Torvalds 10211da177e4SLinus Torvalds /* Build TCP header and checksum it. */ 1022aa8223c7SArnaldo Carvalho de Melo th = tcp_hdr(skb); 1023c720c7e8SEric Dumazet th->source = inet->inet_sport; 1024c720c7e8SEric Dumazet th->dest = inet->inet_dport; 10251da177e4SLinus Torvalds th->seq = htonl(tcb->seq); 10261da177e4SLinus Torvalds th->ack_seq = htonl(tp->rcv_nxt); 1027df7a3b07SAl Viro *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | 10284de075e0SEric Dumazet tcb->tcp_flags); 1029dfb4b9dcSDavid S. Miller 10304de075e0SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) { 10311da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments 10321da177e4SLinus Torvalds * is never scaled. 10331da177e4SLinus Torvalds */ 1034600ff0c2SIlpo Järvinen th->window = htons(min(tp->rcv_wnd, 65535U)); 10351da177e4SLinus Torvalds } else { 10361da177e4SLinus Torvalds th->window = htons(tcp_select_window(sk)); 10371da177e4SLinus Torvalds } 10381da177e4SLinus Torvalds th->check = 0; 10391da177e4SLinus Torvalds th->urg_ptr = 0; 10401da177e4SLinus Torvalds 104133f5f57eSIlpo Järvinen /* The urg_mode check is necessary during a below snd_una win probe */ 10427691367dSHerbert Xu if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { 10437691367dSHerbert Xu if (before(tp->snd_up, tcb->seq + 0x10000)) { 10441da177e4SLinus Torvalds th->urg_ptr = htons(tp->snd_up - tcb->seq); 10451da177e4SLinus Torvalds th->urg = 1; 10467691367dSHerbert Xu } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { 10470eae88f3SEric Dumazet th->urg_ptr = htons(0xFFFF); 10487691367dSHerbert Xu th->urg = 1; 10497691367dSHerbert Xu } 10501da177e4SLinus Torvalds } 10511da177e4SLinus Torvalds 1052bd0388aeSWilliam Allen Simpson tcp_options_write((__be32 *)(th + 1), tp, &opts); 10534de075e0SEric Dumazet if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0)) 10549e412ba7SIlpo Järvinen TCP_ECN_send(sk, skb, tcp_header_size); 1055dfb4b9dcSDavid S. Miller 1056cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 1057cfb6eeb4SYOSHIFUJI Hideaki /* Calculate the MD5 hash, as we have all we need now */ 1058cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 1059a465419bSEric Dumazet sk_nocaps_add(sk, NETIF_F_GSO_MASK); 1060bd0388aeSWilliam Allen Simpson tp->af_specific->calc_md5_hash(opts.hash_location, 106149a72dfbSAdam Langley md5, sk, NULL, skb); 1062cfb6eeb4SYOSHIFUJI Hideaki } 1063cfb6eeb4SYOSHIFUJI Hideaki #endif 1064cfb6eeb4SYOSHIFUJI Hideaki 1065bb296246SHerbert Xu icsk->icsk_af_ops->send_check(sk, skb); 10661da177e4SLinus Torvalds 10674de075e0SEric Dumazet if (likely(tcb->tcp_flags & TCPHDR_ACK)) 1068fc6415bcSDavid S. Miller tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); 10691da177e4SLinus Torvalds 10701da177e4SLinus Torvalds if (skb->len != tcp_header_size) 1071cf533ea5SEric Dumazet tcp_event_data_sent(tp, sk); 10721da177e4SLinus Torvalds 1073bd37a088SWei Yongjun if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) 1074aa2ea058STom Herbert TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, 1075aa2ea058STom Herbert tcp_skb_pcount(skb)); 10761da177e4SLinus Torvalds 1077d9d8da80SDavid S. Miller err = icsk->icsk_af_ops->queue_xmit(skb, &inet->cork.fl); 107883de47cdSHua Zhong if (likely(err <= 0)) 10791da177e4SLinus Torvalds return err; 10801da177e4SLinus Torvalds 10813cfe3baaSIlpo Järvinen tcp_enter_cwr(sk, 1); 10821da177e4SLinus Torvalds 1083b9df3cb8SGerrit Renker return net_xmit_eval(err); 10841da177e4SLinus Torvalds } 10851da177e4SLinus Torvalds 108667edfef7SAndi Kleen /* This routine just queues the buffer for sending. 10871da177e4SLinus Torvalds * 10881da177e4SLinus Torvalds * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, 10891da177e4SLinus Torvalds * otherwise socket can stall. 10901da177e4SLinus Torvalds */ 10911da177e4SLinus Torvalds static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) 10921da177e4SLinus Torvalds { 10931da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 10941da177e4SLinus Torvalds 10951da177e4SLinus Torvalds /* Advance write_seq and place onto the write_queue. */ 10961da177e4SLinus Torvalds tp->write_seq = TCP_SKB_CB(skb)->end_seq; 10971da177e4SLinus Torvalds skb_header_release(skb); 1098fe067e8aSDavid S. Miller tcp_add_write_queue_tail(sk, skb); 10993ab224beSHideo Aoki sk->sk_wmem_queued += skb->truesize; 11003ab224beSHideo Aoki sk_mem_charge(sk, skb->truesize); 11011da177e4SLinus Torvalds } 11021da177e4SLinus Torvalds 110367edfef7SAndi Kleen /* Initialize TSO segments for a packet. */ 1104cf533ea5SEric Dumazet static void tcp_set_skb_tso_segs(const struct sock *sk, struct sk_buff *skb, 1105056834d9SIlpo Järvinen unsigned int mss_now) 1106f6302d1dSDavid S. Miller { 11078e5b9ddaSHerbert Xu if (skb->len <= mss_now || !sk_can_gso(sk) || 11088e5b9ddaSHerbert Xu skb->ip_summed == CHECKSUM_NONE) { 1109f6302d1dSDavid S. Miller /* Avoid the costly divide in the normal 1110f6302d1dSDavid S. Miller * non-TSO case. 1111f6302d1dSDavid S. Miller */ 11127967168cSHerbert Xu skb_shinfo(skb)->gso_segs = 1; 11137967168cSHerbert Xu skb_shinfo(skb)->gso_size = 0; 11147967168cSHerbert Xu skb_shinfo(skb)->gso_type = 0; 1115f6302d1dSDavid S. Miller } else { 1116356f89e1SIlpo Järvinen skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now); 11177967168cSHerbert Xu skb_shinfo(skb)->gso_size = mss_now; 1118bcd76111SHerbert Xu skb_shinfo(skb)->gso_type = sk->sk_gso_type; 11191da177e4SLinus Torvalds } 11201da177e4SLinus Torvalds } 11211da177e4SLinus Torvalds 112291fed7a1SIlpo Järvinen /* When a modification to fackets out becomes necessary, we need to check 112368f8353bSIlpo Järvinen * skb is counted to fackets_out or not. 112491fed7a1SIlpo Järvinen */ 1125cf533ea5SEric Dumazet static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb, 112691fed7a1SIlpo Järvinen int decr) 112791fed7a1SIlpo Järvinen { 1128a47e5a98SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1129a47e5a98SIlpo Järvinen 1130dc86967bSIlpo Järvinen if (!tp->sacked_out || tcp_is_reno(tp)) 113191fed7a1SIlpo Järvinen return; 113291fed7a1SIlpo Järvinen 11336859d494SIlpo Järvinen if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq)) 113491fed7a1SIlpo Järvinen tp->fackets_out -= decr; 113591fed7a1SIlpo Järvinen } 113691fed7a1SIlpo Järvinen 1137797108d1SIlpo Järvinen /* Pcount in the middle of the write queue got changed, we need to do various 1138797108d1SIlpo Järvinen * tweaks to fix counters 1139797108d1SIlpo Järvinen */ 1140cf533ea5SEric Dumazet static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) 1141797108d1SIlpo Järvinen { 1142797108d1SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1143797108d1SIlpo Järvinen 1144797108d1SIlpo Järvinen tp->packets_out -= decr; 1145797108d1SIlpo Järvinen 1146797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 1147797108d1SIlpo Järvinen tp->sacked_out -= decr; 1148797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 1149797108d1SIlpo Järvinen tp->retrans_out -= decr; 1150797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 1151797108d1SIlpo Järvinen tp->lost_out -= decr; 1152797108d1SIlpo Järvinen 1153797108d1SIlpo Järvinen /* Reno case is special. Sigh... */ 1154797108d1SIlpo Järvinen if (tcp_is_reno(tp) && decr > 0) 1155797108d1SIlpo Järvinen tp->sacked_out -= min_t(u32, tp->sacked_out, decr); 1156797108d1SIlpo Järvinen 1157797108d1SIlpo Järvinen tcp_adjust_fackets_out(sk, skb, decr); 1158797108d1SIlpo Järvinen 1159797108d1SIlpo Järvinen if (tp->lost_skb_hint && 1160797108d1SIlpo Järvinen before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && 116152cf3cc8SIlpo Järvinen (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))) 1162797108d1SIlpo Järvinen tp->lost_cnt_hint -= decr; 1163797108d1SIlpo Järvinen 1164797108d1SIlpo Järvinen tcp_verify_left_out(tp); 1165797108d1SIlpo Järvinen } 1166797108d1SIlpo Järvinen 11671da177e4SLinus Torvalds /* Function to create two new TCP segments. Shrinks the given segment 11681da177e4SLinus Torvalds * to the specified size and appends a new segment with the rest of the 11691da177e4SLinus Torvalds * packet to the list. This won't be called frequently, I hope. 11701da177e4SLinus Torvalds * Remember, these are still headerless SKBs at this point. 11711da177e4SLinus Torvalds */ 1172056834d9SIlpo Järvinen int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, 1173056834d9SIlpo Järvinen unsigned int mss_now) 11741da177e4SLinus Torvalds { 11751da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 11761da177e4SLinus Torvalds struct sk_buff *buff; 11776475be16SDavid S. Miller int nsize, old_factor; 1178b60b49eaSHerbert Xu int nlen; 11799ce01461SIlpo Järvinen u8 flags; 11801da177e4SLinus Torvalds 11812fceec13SIlpo Järvinen if (WARN_ON(len > skb->len)) 11822fceec13SIlpo Järvinen return -EINVAL; 11836a438bbeSStephen Hemminger 11841da177e4SLinus Torvalds nsize = skb_headlen(skb) - len; 11851da177e4SLinus Torvalds if (nsize < 0) 11861da177e4SLinus Torvalds nsize = 0; 11871da177e4SLinus Torvalds 11881da177e4SLinus Torvalds if (skb_cloned(skb) && 11891da177e4SLinus Torvalds skb_is_nonlinear(skb) && 11901da177e4SLinus Torvalds pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) 11911da177e4SLinus Torvalds return -ENOMEM; 11921da177e4SLinus Torvalds 11931da177e4SLinus Torvalds /* Get a new skb... force flag on. */ 11941da177e4SLinus Torvalds buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC); 11951da177e4SLinus Torvalds if (buff == NULL) 11961da177e4SLinus Torvalds return -ENOMEM; /* We'll just try again later. */ 1197ef5cb973SHerbert Xu 11983ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 11993ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1200b60b49eaSHerbert Xu nlen = skb->len - len - nsize; 1201b60b49eaSHerbert Xu buff->truesize += nlen; 1202b60b49eaSHerbert Xu skb->truesize -= nlen; 12031da177e4SLinus Torvalds 12041da177e4SLinus Torvalds /* Correct the sequence numbers. */ 12051da177e4SLinus Torvalds TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 12061da177e4SLinus Torvalds TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 12071da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 12081da177e4SLinus Torvalds 12091da177e4SLinus Torvalds /* PSH and FIN should only be set in the second packet. */ 12104de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 12114de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 12124de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1213e14c3cafSHerbert Xu TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; 12141da177e4SLinus Torvalds 121584fa7933SPatrick McHardy if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { 12161da177e4SLinus Torvalds /* Copy and checksum data tail into the new buffer. */ 1217056834d9SIlpo Järvinen buff->csum = csum_partial_copy_nocheck(skb->data + len, 1218056834d9SIlpo Järvinen skb_put(buff, nsize), 12191da177e4SLinus Torvalds nsize, 0); 12201da177e4SLinus Torvalds 12211da177e4SLinus Torvalds skb_trim(skb, len); 12221da177e4SLinus Torvalds 12231da177e4SLinus Torvalds skb->csum = csum_block_sub(skb->csum, buff->csum, len); 12241da177e4SLinus Torvalds } else { 122584fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 12261da177e4SLinus Torvalds skb_split(skb, buff, len); 12271da177e4SLinus Torvalds } 12281da177e4SLinus Torvalds 12291da177e4SLinus Torvalds buff->ip_summed = skb->ip_summed; 12301da177e4SLinus Torvalds 12311da177e4SLinus Torvalds /* Looks stupid, but our code really uses when of 12321da177e4SLinus Torvalds * skbs, which it never sent before. --ANK 12331da177e4SLinus Torvalds */ 12341da177e4SLinus Torvalds TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when; 1235a61bbcf2SPatrick McHardy buff->tstamp = skb->tstamp; 12361da177e4SLinus Torvalds 12376475be16SDavid S. Miller old_factor = tcp_skb_pcount(skb); 12386475be16SDavid S. Miller 12391da177e4SLinus Torvalds /* Fix up tso_factor for both original and new SKB. */ 1240846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1241846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, buff, mss_now); 12421da177e4SLinus Torvalds 12436475be16SDavid S. Miller /* If this packet has been sent out already, we must 12446475be16SDavid S. Miller * adjust the various packet counters. 12456475be16SDavid S. Miller */ 1246cf0b450cSHerbert Xu if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 12476475be16SDavid S. Miller int diff = old_factor - tcp_skb_pcount(skb) - 12486475be16SDavid S. Miller tcp_skb_pcount(buff); 12491da177e4SLinus Torvalds 1250797108d1SIlpo Järvinen if (diff) 1251797108d1SIlpo Järvinen tcp_adjust_pcount(sk, skb, diff); 12521da177e4SLinus Torvalds } 12531da177e4SLinus Torvalds 12541da177e4SLinus Torvalds /* Link BUFF into the send queue. */ 1255f44b5271SDavid S. Miller skb_header_release(buff); 1256fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 12571da177e4SLinus Torvalds 12581da177e4SLinus Torvalds return 0; 12591da177e4SLinus Torvalds } 12601da177e4SLinus Torvalds 12611da177e4SLinus Torvalds /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c 12621da177e4SLinus Torvalds * eventually). The difference is that pulled data not copied, but 12631da177e4SLinus Torvalds * immediately discarded. 12641da177e4SLinus Torvalds */ 1265f2911969SHerbert Xu ~{PmVHI~} static void __pskb_trim_head(struct sk_buff *skb, int len) 12661da177e4SLinus Torvalds { 12671da177e4SLinus Torvalds int i, k, eat; 12681da177e4SLinus Torvalds 12694fa48bf3SEric Dumazet eat = min_t(int, len, skb_headlen(skb)); 12704fa48bf3SEric Dumazet if (eat) { 12714fa48bf3SEric Dumazet __skb_pull(skb, eat); 127222b4a4f2SEric Dumazet skb->avail_size -= eat; 12734fa48bf3SEric Dumazet len -= eat; 12744fa48bf3SEric Dumazet if (!len) 12754fa48bf3SEric Dumazet return; 12764fa48bf3SEric Dumazet } 12771da177e4SLinus Torvalds eat = len; 12781da177e4SLinus Torvalds k = 0; 12791da177e4SLinus Torvalds for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) { 12809e903e08SEric Dumazet int size = skb_frag_size(&skb_shinfo(skb)->frags[i]); 12819e903e08SEric Dumazet 12829e903e08SEric Dumazet if (size <= eat) { 1283aff65da0SIan Campbell skb_frag_unref(skb, i); 12849e903e08SEric Dumazet eat -= size; 12851da177e4SLinus Torvalds } else { 12861da177e4SLinus Torvalds skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i]; 12871da177e4SLinus Torvalds if (eat) { 12881da177e4SLinus Torvalds skb_shinfo(skb)->frags[k].page_offset += eat; 12899e903e08SEric Dumazet skb_frag_size_sub(&skb_shinfo(skb)->frags[k], eat); 12901da177e4SLinus Torvalds eat = 0; 12911da177e4SLinus Torvalds } 12921da177e4SLinus Torvalds k++; 12931da177e4SLinus Torvalds } 12941da177e4SLinus Torvalds } 12951da177e4SLinus Torvalds skb_shinfo(skb)->nr_frags = k; 12961da177e4SLinus Torvalds 129727a884dcSArnaldo Carvalho de Melo skb_reset_tail_pointer(skb); 12981da177e4SLinus Torvalds skb->data_len -= len; 12991da177e4SLinus Torvalds skb->len = skb->data_len; 13001da177e4SLinus Torvalds } 13011da177e4SLinus Torvalds 130267edfef7SAndi Kleen /* Remove acked data from a packet in the transmit queue. */ 13031da177e4SLinus Torvalds int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 13041da177e4SLinus Torvalds { 1305056834d9SIlpo Järvinen if (skb_cloned(skb) && pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) 13061da177e4SLinus Torvalds return -ENOMEM; 13071da177e4SLinus Torvalds 13084fa48bf3SEric Dumazet __pskb_trim_head(skb, len); 13091da177e4SLinus Torvalds 13101da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq += len; 131184fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 13121da177e4SLinus Torvalds 13131da177e4SLinus Torvalds skb->truesize -= len; 13141da177e4SLinus Torvalds sk->sk_wmem_queued -= len; 13153ab224beSHideo Aoki sk_mem_uncharge(sk, len); 13161da177e4SLinus Torvalds sock_set_flag(sk, SOCK_QUEUE_SHRUNK); 13171da177e4SLinus Torvalds 13185b35e1e6SNeal Cardwell /* Any change of skb->len requires recalculation of tso factor. */ 13191da177e4SLinus Torvalds if (tcp_skb_pcount(skb) > 1) 13205b35e1e6SNeal Cardwell tcp_set_skb_tso_segs(sk, skb, tcp_skb_mss(skb)); 13211da177e4SLinus Torvalds 13221da177e4SLinus Torvalds return 0; 13231da177e4SLinus Torvalds } 13241da177e4SLinus Torvalds 132567edfef7SAndi Kleen /* Calculate MSS. Not accounting for SACKs here. */ 132667469601SEric Dumazet int tcp_mtu_to_mss(struct sock *sk, int pmtu) 13275d424d5aSJohn Heffner { 1328cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1329cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 13305d424d5aSJohn Heffner int mss_now; 13315d424d5aSJohn Heffner 13325d424d5aSJohn Heffner /* Calculate base mss without TCP options: 13335d424d5aSJohn Heffner It is MMS_S - sizeof(tcphdr) of rfc1122 13345d424d5aSJohn Heffner */ 13355d424d5aSJohn Heffner mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 13365d424d5aSJohn Heffner 133767469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 133867469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 133967469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 134067469601SEric Dumazet 134167469601SEric Dumazet if (dst && dst_allfrag(dst)) 134267469601SEric Dumazet mss_now -= icsk->icsk_af_ops->net_frag_header_len; 134367469601SEric Dumazet } 134467469601SEric Dumazet 13455d424d5aSJohn Heffner /* Clamp it (mss_clamp does not include tcp options) */ 13465d424d5aSJohn Heffner if (mss_now > tp->rx_opt.mss_clamp) 13475d424d5aSJohn Heffner mss_now = tp->rx_opt.mss_clamp; 13485d424d5aSJohn Heffner 13495d424d5aSJohn Heffner /* Now subtract optional transport overhead */ 13505d424d5aSJohn Heffner mss_now -= icsk->icsk_ext_hdr_len; 13515d424d5aSJohn Heffner 13525d424d5aSJohn Heffner /* Then reserve room for full set of TCP options and 8 bytes of data */ 13535d424d5aSJohn Heffner if (mss_now < 48) 13545d424d5aSJohn Heffner mss_now = 48; 13555d424d5aSJohn Heffner 13565d424d5aSJohn Heffner /* Now subtract TCP options size, not including SACKs */ 13575d424d5aSJohn Heffner mss_now -= tp->tcp_header_len - sizeof(struct tcphdr); 13585d424d5aSJohn Heffner 13595d424d5aSJohn Heffner return mss_now; 13605d424d5aSJohn Heffner } 13615d424d5aSJohn Heffner 13625d424d5aSJohn Heffner /* Inverse of above */ 136367469601SEric Dumazet int tcp_mss_to_mtu(struct sock *sk, int mss) 13645d424d5aSJohn Heffner { 1365cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1366cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 13675d424d5aSJohn Heffner int mtu; 13685d424d5aSJohn Heffner 13695d424d5aSJohn Heffner mtu = mss + 13705d424d5aSJohn Heffner tp->tcp_header_len + 13715d424d5aSJohn Heffner icsk->icsk_ext_hdr_len + 13725d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 13735d424d5aSJohn Heffner 137467469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 137567469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 137667469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 137767469601SEric Dumazet 137867469601SEric Dumazet if (dst && dst_allfrag(dst)) 137967469601SEric Dumazet mtu += icsk->icsk_af_ops->net_frag_header_len; 138067469601SEric Dumazet } 13815d424d5aSJohn Heffner return mtu; 13825d424d5aSJohn Heffner } 13835d424d5aSJohn Heffner 138467edfef7SAndi Kleen /* MTU probing init per socket */ 13855d424d5aSJohn Heffner void tcp_mtup_init(struct sock *sk) 13865d424d5aSJohn Heffner { 13875d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 13885d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 13895d424d5aSJohn Heffner 13905d424d5aSJohn Heffner icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1; 13915d424d5aSJohn Heffner icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 13925d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 13935d424d5aSJohn Heffner icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss); 13945d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 13955d424d5aSJohn Heffner } 13964bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_mtup_init); 13975d424d5aSJohn Heffner 13981da177e4SLinus Torvalds /* This function synchronize snd mss to current pmtu/exthdr set. 13991da177e4SLinus Torvalds 14001da177e4SLinus Torvalds tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 14011da177e4SLinus Torvalds for TCP options, but includes only bare TCP header. 14021da177e4SLinus Torvalds 14031da177e4SLinus Torvalds tp->rx_opt.mss_clamp is mss negotiated at connection setup. 1404caa20d9aSStephen Hemminger It is minimum of user_mss and mss received with SYN. 14051da177e4SLinus Torvalds It also does not include TCP options. 14061da177e4SLinus Torvalds 1407d83d8461SArnaldo Carvalho de Melo inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. 14081da177e4SLinus Torvalds 14091da177e4SLinus Torvalds tp->mss_cache is current effective sending mss, including 14101da177e4SLinus Torvalds all tcp options except for SACKs. It is evaluated, 14111da177e4SLinus Torvalds taking into account current pmtu, but never exceeds 14121da177e4SLinus Torvalds tp->rx_opt.mss_clamp. 14131da177e4SLinus Torvalds 14141da177e4SLinus Torvalds NOTE1. rfc1122 clearly states that advertised MSS 14151da177e4SLinus Torvalds DOES NOT include either tcp or ip options. 14161da177e4SLinus Torvalds 1417d83d8461SArnaldo Carvalho de Melo NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache 1418d83d8461SArnaldo Carvalho de Melo are READ ONLY outside this function. --ANK (980731) 14191da177e4SLinus Torvalds */ 14201da177e4SLinus Torvalds unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 14211da177e4SLinus Torvalds { 14221da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1423d83d8461SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 14245d424d5aSJohn Heffner int mss_now; 14251da177e4SLinus Torvalds 14265d424d5aSJohn Heffner if (icsk->icsk_mtup.search_high > pmtu) 14275d424d5aSJohn Heffner icsk->icsk_mtup.search_high = pmtu; 14281da177e4SLinus Torvalds 14295d424d5aSJohn Heffner mss_now = tcp_mtu_to_mss(sk, pmtu); 1430409d22b4SIlpo Järvinen mss_now = tcp_bound_to_half_wnd(tp, mss_now); 14311da177e4SLinus Torvalds 14321da177e4SLinus Torvalds /* And store cached results */ 1433d83d8461SArnaldo Carvalho de Melo icsk->icsk_pmtu_cookie = pmtu; 14345d424d5aSJohn Heffner if (icsk->icsk_mtup.enabled) 14355d424d5aSJohn Heffner mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); 1436c1b4a7e6SDavid S. Miller tp->mss_cache = mss_now; 14371da177e4SLinus Torvalds 14381da177e4SLinus Torvalds return mss_now; 14391da177e4SLinus Torvalds } 14404bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_sync_mss); 14411da177e4SLinus Torvalds 14421da177e4SLinus Torvalds /* Compute the current effective MSS, taking SACKs and IP options, 14431da177e4SLinus Torvalds * and even PMTU discovery events into account. 14441da177e4SLinus Torvalds */ 14450c54b85fSIlpo Järvinen unsigned int tcp_current_mss(struct sock *sk) 14461da177e4SLinus Torvalds { 1447cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1448cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1449c1b4a7e6SDavid S. Miller u32 mss_now; 145095c96174SEric Dumazet unsigned int header_len; 145133ad798cSAdam Langley struct tcp_out_options opts; 145233ad798cSAdam Langley struct tcp_md5sig_key *md5; 14531da177e4SLinus Torvalds 1454c1b4a7e6SDavid S. Miller mss_now = tp->mss_cache; 1455c1b4a7e6SDavid S. Miller 14561da177e4SLinus Torvalds if (dst) { 14571da177e4SLinus Torvalds u32 mtu = dst_mtu(dst); 1458d83d8461SArnaldo Carvalho de Melo if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 14591da177e4SLinus Torvalds mss_now = tcp_sync_mss(sk, mtu); 14601da177e4SLinus Torvalds } 14611da177e4SLinus Torvalds 146233ad798cSAdam Langley header_len = tcp_established_options(sk, NULL, &opts, &md5) + 146333ad798cSAdam Langley sizeof(struct tcphdr); 146433ad798cSAdam Langley /* The mss_cache is sized based on tp->tcp_header_len, which assumes 146533ad798cSAdam Langley * some common options. If this is an odd packet (because we have SACK 146633ad798cSAdam Langley * blocks etc) then our calculated header_len will be different, and 146733ad798cSAdam Langley * we have to adjust mss_now correspondingly */ 146833ad798cSAdam Langley if (header_len != tp->tcp_header_len) { 146933ad798cSAdam Langley int delta = (int) header_len - tp->tcp_header_len; 147033ad798cSAdam Langley mss_now -= delta; 147133ad798cSAdam Langley } 1472cfb6eeb4SYOSHIFUJI Hideaki 14731da177e4SLinus Torvalds return mss_now; 14741da177e4SLinus Torvalds } 14751da177e4SLinus Torvalds 1476a762a980SDavid S. Miller /* Congestion window validation. (RFC2861) */ 14779e412ba7SIlpo Järvinen static void tcp_cwnd_validate(struct sock *sk) 1478a762a980SDavid S. Miller { 14799e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1480a762a980SDavid S. Miller 1481d436d686SIlpo Järvinen if (tp->packets_out >= tp->snd_cwnd) { 1482a762a980SDavid S. Miller /* Network is feed fully. */ 1483a762a980SDavid S. Miller tp->snd_cwnd_used = 0; 1484a762a980SDavid S. Miller tp->snd_cwnd_stamp = tcp_time_stamp; 1485a762a980SDavid S. Miller } else { 1486a762a980SDavid S. Miller /* Network starves. */ 1487a762a980SDavid S. Miller if (tp->packets_out > tp->snd_cwnd_used) 1488a762a980SDavid S. Miller tp->snd_cwnd_used = tp->packets_out; 1489a762a980SDavid S. Miller 149015d33c07SDavid S. Miller if (sysctl_tcp_slow_start_after_idle && 149115d33c07SDavid S. Miller (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto) 1492a762a980SDavid S. Miller tcp_cwnd_application_limited(sk); 1493a762a980SDavid S. Miller } 1494a762a980SDavid S. Miller } 1495a762a980SDavid S. Miller 14960e3a4803SIlpo Järvinen /* Returns the portion of skb which can be sent right away without 14970e3a4803SIlpo Järvinen * introducing MSS oddities to segment boundaries. In rare cases where 14980e3a4803SIlpo Järvinen * mss_now != mss_cache, we will request caller to create a small skb 14990e3a4803SIlpo Järvinen * per input skb which could be mostly avoided here (if desired). 15005ea3a748SIlpo Järvinen * 15015ea3a748SIlpo Järvinen * We explicitly want to create a request for splitting write queue tail 15025ea3a748SIlpo Järvinen * to a small skb for Nagle purposes while avoiding unnecessary modulos, 15035ea3a748SIlpo Järvinen * thus all the complexity (cwnd_len is always MSS multiple which we 15045ea3a748SIlpo Järvinen * return whenever allowed by the other factors). Basically we need the 15055ea3a748SIlpo Järvinen * modulo only when the receiver window alone is the limiting factor or 15065ea3a748SIlpo Järvinen * when we would be allowed to send the split-due-to-Nagle skb fully. 15070e3a4803SIlpo Järvinen */ 1508cf533ea5SEric Dumazet static unsigned int tcp_mss_split_point(const struct sock *sk, const struct sk_buff *skb, 1509056834d9SIlpo Järvinen unsigned int mss_now, unsigned int cwnd) 1510c1b4a7e6SDavid S. Miller { 1511cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 15120e3a4803SIlpo Järvinen u32 needed, window, cwnd_len; 1513c1b4a7e6SDavid S. Miller 151490840defSIlpo Järvinen window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1515c1b4a7e6SDavid S. Miller cwnd_len = mss_now * cwnd; 15160e3a4803SIlpo Järvinen 15170e3a4803SIlpo Järvinen if (likely(cwnd_len <= window && skb != tcp_write_queue_tail(sk))) 15180e3a4803SIlpo Järvinen return cwnd_len; 15190e3a4803SIlpo Järvinen 15205ea3a748SIlpo Järvinen needed = min(skb->len, window); 15215ea3a748SIlpo Järvinen 152217515408SIlpo Järvinen if (cwnd_len <= needed) 15230e3a4803SIlpo Järvinen return cwnd_len; 15240e3a4803SIlpo Järvinen 15250e3a4803SIlpo Järvinen return needed - needed % mss_now; 1526c1b4a7e6SDavid S. Miller } 1527c1b4a7e6SDavid S. Miller 1528c1b4a7e6SDavid S. Miller /* Can at least one segment of SKB be sent right now, according to the 1529c1b4a7e6SDavid S. Miller * congestion window rules? If so, return how many segments are allowed. 1530c1b4a7e6SDavid S. Miller */ 1531cf533ea5SEric Dumazet static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, 1532cf533ea5SEric Dumazet const struct sk_buff *skb) 1533c1b4a7e6SDavid S. Miller { 1534c1b4a7e6SDavid S. Miller u32 in_flight, cwnd; 1535c1b4a7e6SDavid S. Miller 1536c1b4a7e6SDavid S. Miller /* Don't be strict about the congestion window for the final FIN. */ 15374de075e0SEric Dumazet if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && 15384de075e0SEric Dumazet tcp_skb_pcount(skb) == 1) 1539c1b4a7e6SDavid S. Miller return 1; 1540c1b4a7e6SDavid S. Miller 1541c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1542c1b4a7e6SDavid S. Miller cwnd = tp->snd_cwnd; 1543c1b4a7e6SDavid S. Miller if (in_flight < cwnd) 1544c1b4a7e6SDavid S. Miller return (cwnd - in_flight); 1545c1b4a7e6SDavid S. Miller 1546c1b4a7e6SDavid S. Miller return 0; 1547c1b4a7e6SDavid S. Miller } 1548c1b4a7e6SDavid S. Miller 1549b595076aSUwe Kleine-König /* Initialize TSO state of a skb. 155067edfef7SAndi Kleen * This must be invoked the first time we consider transmitting 1551c1b4a7e6SDavid S. Miller * SKB onto the wire. 1552c1b4a7e6SDavid S. Miller */ 1553cf533ea5SEric Dumazet static int tcp_init_tso_segs(const struct sock *sk, struct sk_buff *skb, 1554056834d9SIlpo Järvinen unsigned int mss_now) 1555c1b4a7e6SDavid S. Miller { 1556c1b4a7e6SDavid S. Miller int tso_segs = tcp_skb_pcount(skb); 1557c1b4a7e6SDavid S. Miller 1558f8269a49SIlpo Järvinen if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { 1559846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1560c1b4a7e6SDavid S. Miller tso_segs = tcp_skb_pcount(skb); 1561c1b4a7e6SDavid S. Miller } 1562c1b4a7e6SDavid S. Miller return tso_segs; 1563c1b4a7e6SDavid S. Miller } 1564c1b4a7e6SDavid S. Miller 156567edfef7SAndi Kleen /* Minshall's variant of the Nagle send check. */ 1566a2a385d6SEric Dumazet static inline bool tcp_minshall_check(const struct tcp_sock *tp) 1567c1b4a7e6SDavid S. Miller { 1568c1b4a7e6SDavid S. Miller return after(tp->snd_sml, tp->snd_una) && 1569c1b4a7e6SDavid S. Miller !after(tp->snd_sml, tp->snd_nxt); 1570c1b4a7e6SDavid S. Miller } 1571c1b4a7e6SDavid S. Miller 1572a2a385d6SEric Dumazet /* Return false, if packet can be sent now without violation Nagle's rules: 1573c1b4a7e6SDavid S. Miller * 1. It is full sized. 1574c1b4a7e6SDavid S. Miller * 2. Or it contains FIN. (already checked by caller) 15756d67e9beSFeng King * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. 1576c1b4a7e6SDavid S. Miller * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 1577c1b4a7e6SDavid S. Miller * With Minshall's modification: all sent small packets are ACKed. 1578c1b4a7e6SDavid S. Miller */ 1579a2a385d6SEric Dumazet static inline bool tcp_nagle_check(const struct tcp_sock *tp, 1580c1b4a7e6SDavid S. Miller const struct sk_buff *skb, 158195c96174SEric Dumazet unsigned int mss_now, int nonagle) 1582c1b4a7e6SDavid S. Miller { 1583a02cec21SEric Dumazet return skb->len < mss_now && 1584c1b4a7e6SDavid S. Miller ((nonagle & TCP_NAGLE_CORK) || 1585a02cec21SEric Dumazet (!nonagle && tp->packets_out && tcp_minshall_check(tp))); 1586c1b4a7e6SDavid S. Miller } 1587c1b4a7e6SDavid S. Miller 1588a2a385d6SEric Dumazet /* Return true if the Nagle test allows this packet to be 1589c1b4a7e6SDavid S. Miller * sent now. 1590c1b4a7e6SDavid S. Miller */ 1591a2a385d6SEric Dumazet static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, 1592c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1593c1b4a7e6SDavid S. Miller { 1594c1b4a7e6SDavid S. Miller /* Nagle rule does not apply to frames, which sit in the middle of the 1595c1b4a7e6SDavid S. Miller * write_queue (they have no chances to get new data). 1596c1b4a7e6SDavid S. Miller * 1597c1b4a7e6SDavid S. Miller * This is implemented in the callers, where they modify the 'nonagle' 1598c1b4a7e6SDavid S. Miller * argument based upon the location of SKB in the send queue. 1599c1b4a7e6SDavid S. Miller */ 1600c1b4a7e6SDavid S. Miller if (nonagle & TCP_NAGLE_PUSH) 1601a2a385d6SEric Dumazet return true; 1602c1b4a7e6SDavid S. Miller 1603d551e454SIlpo Järvinen /* Don't use the nagle rule for urgent data (or for the final FIN). 1604d551e454SIlpo Järvinen * Nagle can be ignored during F-RTO too (see RFC4138). 1605d551e454SIlpo Järvinen */ 160633f5f57eSIlpo Järvinen if (tcp_urg_mode(tp) || (tp->frto_counter == 2) || 16074de075e0SEric Dumazet (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) 1608a2a385d6SEric Dumazet return true; 1609c1b4a7e6SDavid S. Miller 1610c1b4a7e6SDavid S. Miller if (!tcp_nagle_check(tp, skb, cur_mss, nonagle)) 1611a2a385d6SEric Dumazet return true; 1612c1b4a7e6SDavid S. Miller 1613a2a385d6SEric Dumazet return false; 1614c1b4a7e6SDavid S. Miller } 1615c1b4a7e6SDavid S. Miller 1616c1b4a7e6SDavid S. Miller /* Does at least the first segment of SKB fit into the send window? */ 1617a2a385d6SEric Dumazet static bool tcp_snd_wnd_test(const struct tcp_sock *tp, 1618a2a385d6SEric Dumazet const struct sk_buff *skb, 1619056834d9SIlpo Järvinen unsigned int cur_mss) 1620c1b4a7e6SDavid S. Miller { 1621c1b4a7e6SDavid S. Miller u32 end_seq = TCP_SKB_CB(skb)->end_seq; 1622c1b4a7e6SDavid S. Miller 1623c1b4a7e6SDavid S. Miller if (skb->len > cur_mss) 1624c1b4a7e6SDavid S. Miller end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 1625c1b4a7e6SDavid S. Miller 162690840defSIlpo Järvinen return !after(end_seq, tcp_wnd_end(tp)); 1627c1b4a7e6SDavid S. Miller } 1628c1b4a7e6SDavid S. Miller 1629fe067e8aSDavid S. Miller /* This checks if the data bearing packet SKB (usually tcp_send_head(sk)) 1630c1b4a7e6SDavid S. Miller * should be put on the wire right now. If so, it returns the number of 1631c1b4a7e6SDavid S. Miller * packets allowed by the congestion window. 1632c1b4a7e6SDavid S. Miller */ 1633cf533ea5SEric Dumazet static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb, 1634c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1635c1b4a7e6SDavid S. Miller { 1636cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1637c1b4a7e6SDavid S. Miller unsigned int cwnd_quota; 1638c1b4a7e6SDavid S. Miller 1639846998aeSDavid S. Miller tcp_init_tso_segs(sk, skb, cur_mss); 1640c1b4a7e6SDavid S. Miller 1641c1b4a7e6SDavid S. Miller if (!tcp_nagle_test(tp, skb, cur_mss, nonagle)) 1642c1b4a7e6SDavid S. Miller return 0; 1643c1b4a7e6SDavid S. Miller 1644c1b4a7e6SDavid S. Miller cwnd_quota = tcp_cwnd_test(tp, skb); 1645056834d9SIlpo Järvinen if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss)) 1646c1b4a7e6SDavid S. Miller cwnd_quota = 0; 1647c1b4a7e6SDavid S. Miller 1648c1b4a7e6SDavid S. Miller return cwnd_quota; 1649c1b4a7e6SDavid S. Miller } 1650c1b4a7e6SDavid S. Miller 165167edfef7SAndi Kleen /* Test if sending is allowed right now. */ 1652a2a385d6SEric Dumazet bool tcp_may_send_now(struct sock *sk) 1653c1b4a7e6SDavid S. Miller { 1654cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1655fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 1656c1b4a7e6SDavid S. Miller 1657a02cec21SEric Dumazet return skb && 16580c54b85fSIlpo Järvinen tcp_snd_test(sk, skb, tcp_current_mss(sk), 1659c1b4a7e6SDavid S. Miller (tcp_skb_is_last(sk, skb) ? 1660a02cec21SEric Dumazet tp->nonagle : TCP_NAGLE_PUSH)); 1661c1b4a7e6SDavid S. Miller } 1662c1b4a7e6SDavid S. Miller 1663c1b4a7e6SDavid S. Miller /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 1664c1b4a7e6SDavid S. Miller * which is put after SKB on the list. It is very much like 1665c1b4a7e6SDavid S. Miller * tcp_fragment() except that it may make several kinds of assumptions 1666c1b4a7e6SDavid S. Miller * in order to speed up the splitting operation. In particular, we 1667c1b4a7e6SDavid S. Miller * know that all the data is in scatter-gather pages, and that the 1668c1b4a7e6SDavid S. Miller * packet has never been sent out before (and thus is not cloned). 1669c1b4a7e6SDavid S. Miller */ 1670056834d9SIlpo Järvinen static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, 1671c4ead4c5SEric Dumazet unsigned int mss_now, gfp_t gfp) 1672c1b4a7e6SDavid S. Miller { 1673c1b4a7e6SDavid S. Miller struct sk_buff *buff; 1674c1b4a7e6SDavid S. Miller int nlen = skb->len - len; 16759ce01461SIlpo Järvinen u8 flags; 1676c1b4a7e6SDavid S. Miller 1677c1b4a7e6SDavid S. Miller /* All of a TSO frame must be composed of paged data. */ 1678c8ac3774SHerbert Xu if (skb->len != skb->data_len) 1679c8ac3774SHerbert Xu return tcp_fragment(sk, skb, len, mss_now); 1680c1b4a7e6SDavid S. Miller 1681c4ead4c5SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, gfp); 1682c1b4a7e6SDavid S. Miller if (unlikely(buff == NULL)) 1683c1b4a7e6SDavid S. Miller return -ENOMEM; 1684c1b4a7e6SDavid S. Miller 16853ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 16863ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1687b60b49eaSHerbert Xu buff->truesize += nlen; 1688c1b4a7e6SDavid S. Miller skb->truesize -= nlen; 1689c1b4a7e6SDavid S. Miller 1690c1b4a7e6SDavid S. Miller /* Correct the sequence numbers. */ 1691c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 1692c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 1693c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 1694c1b4a7e6SDavid S. Miller 1695c1b4a7e6SDavid S. Miller /* PSH and FIN should only be set in the second packet. */ 16964de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 16974de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 16984de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1699c1b4a7e6SDavid S. Miller 1700c1b4a7e6SDavid S. Miller /* This packet was never sent out yet, so no SACK bits. */ 1701c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->sacked = 0; 1702c1b4a7e6SDavid S. Miller 170384fa7933SPatrick McHardy buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; 1704c1b4a7e6SDavid S. Miller skb_split(skb, buff, len); 1705c1b4a7e6SDavid S. Miller 1706c1b4a7e6SDavid S. Miller /* Fix up tso_factor for both original and new SKB. */ 1707846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1708846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, buff, mss_now); 1709c1b4a7e6SDavid S. Miller 1710c1b4a7e6SDavid S. Miller /* Link BUFF into the send queue. */ 1711c1b4a7e6SDavid S. Miller skb_header_release(buff); 1712fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 1713c1b4a7e6SDavid S. Miller 1714c1b4a7e6SDavid S. Miller return 0; 1715c1b4a7e6SDavid S. Miller } 1716c1b4a7e6SDavid S. Miller 1717c1b4a7e6SDavid S. Miller /* Try to defer sending, if possible, in order to minimize the amount 1718c1b4a7e6SDavid S. Miller * of TSO splitting we do. View it as a kind of TSO Nagle test. 1719c1b4a7e6SDavid S. Miller * 1720c1b4a7e6SDavid S. Miller * This algorithm is from John Heffner. 1721c1b4a7e6SDavid S. Miller */ 1722a2a385d6SEric Dumazet static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb) 1723c1b4a7e6SDavid S. Miller { 17249e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 17256687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 1726c1b4a7e6SDavid S. Miller u32 send_win, cong_win, limit, in_flight; 1727ad9f4f50SEric Dumazet int win_divisor; 1728c1b4a7e6SDavid S. Miller 17294de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 1730ae8064acSJohn Heffner goto send_now; 1731c1b4a7e6SDavid S. Miller 17326687e988SArnaldo Carvalho de Melo if (icsk->icsk_ca_state != TCP_CA_Open) 1733ae8064acSJohn Heffner goto send_now; 1734ae8064acSJohn Heffner 1735ae8064acSJohn Heffner /* Defer for less than two clock ticks. */ 1736bd515c3eSIlpo Järvinen if (tp->tso_deferred && 1737a2acde07SIlpo Järvinen (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1) 1738ae8064acSJohn Heffner goto send_now; 1739908a75c1SDavid S. Miller 1740c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1741c1b4a7e6SDavid S. Miller 1742056834d9SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight)); 1743c1b4a7e6SDavid S. Miller 174490840defSIlpo Järvinen send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1745c1b4a7e6SDavid S. Miller 1746c1b4a7e6SDavid S. Miller /* From in_flight test above, we know that cwnd > in_flight. */ 1747c1b4a7e6SDavid S. Miller cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; 1748c1b4a7e6SDavid S. Miller 1749c1b4a7e6SDavid S. Miller limit = min(send_win, cong_win); 1750c1b4a7e6SDavid S. Miller 1751ba244fe9SDavid S. Miller /* If a full-sized TSO skb can be sent, do it. */ 175282cc1a7aSPeter P Waskiewicz Jr if (limit >= sk->sk_gso_max_size) 1753ae8064acSJohn Heffner goto send_now; 1754ba244fe9SDavid S. Miller 175562ad2761SIlpo Järvinen /* Middle in queue won't get any more data, full sendable already? */ 175662ad2761SIlpo Järvinen if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) 175762ad2761SIlpo Järvinen goto send_now; 175862ad2761SIlpo Järvinen 1759ad9f4f50SEric Dumazet win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor); 1760ad9f4f50SEric Dumazet if (win_divisor) { 1761c1b4a7e6SDavid S. Miller u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); 1762c1b4a7e6SDavid S. Miller 1763c1b4a7e6SDavid S. Miller /* If at least some fraction of a window is available, 1764c1b4a7e6SDavid S. Miller * just use it. 1765c1b4a7e6SDavid S. Miller */ 1766ad9f4f50SEric Dumazet chunk /= win_divisor; 1767c1b4a7e6SDavid S. Miller if (limit >= chunk) 1768ae8064acSJohn Heffner goto send_now; 1769c1b4a7e6SDavid S. Miller } else { 1770c1b4a7e6SDavid S. Miller /* Different approach, try not to defer past a single 1771c1b4a7e6SDavid S. Miller * ACK. Receiver should ACK every other full sized 1772c1b4a7e6SDavid S. Miller * frame, so if we have space for more than 3 frames 1773c1b4a7e6SDavid S. Miller * then send now. 1774c1b4a7e6SDavid S. Miller */ 17756b5a5c0dSNeal Cardwell if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) 1776ae8064acSJohn Heffner goto send_now; 1777c1b4a7e6SDavid S. Miller } 1778c1b4a7e6SDavid S. Miller 1779c1b4a7e6SDavid S. Miller /* Ok, it looks like it is advisable to defer. */ 1780ae8064acSJohn Heffner tp->tso_deferred = 1 | (jiffies << 1); 1781ae8064acSJohn Heffner 1782a2a385d6SEric Dumazet return true; 1783ae8064acSJohn Heffner 1784ae8064acSJohn Heffner send_now: 1785ae8064acSJohn Heffner tp->tso_deferred = 0; 1786a2a385d6SEric Dumazet return false; 1787c1b4a7e6SDavid S. Miller } 1788c1b4a7e6SDavid S. Miller 17895d424d5aSJohn Heffner /* Create a new MTU probe if we are ready. 179067edfef7SAndi Kleen * MTU probe is regularly attempting to increase the path MTU by 179167edfef7SAndi Kleen * deliberately sending larger packets. This discovers routing 179267edfef7SAndi Kleen * changes resulting in larger path MTUs. 179367edfef7SAndi Kleen * 17945d424d5aSJohn Heffner * Returns 0 if we should wait to probe (no cwnd available), 17955d424d5aSJohn Heffner * 1 if a probe was sent, 1796056834d9SIlpo Järvinen * -1 otherwise 1797056834d9SIlpo Järvinen */ 17985d424d5aSJohn Heffner static int tcp_mtu_probe(struct sock *sk) 17995d424d5aSJohn Heffner { 18005d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 18015d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 18025d424d5aSJohn Heffner struct sk_buff *skb, *nskb, *next; 18035d424d5aSJohn Heffner int len; 18045d424d5aSJohn Heffner int probe_size; 180591cc17c0SIlpo Järvinen int size_needed; 18065d424d5aSJohn Heffner int copy; 18075d424d5aSJohn Heffner int mss_now; 18085d424d5aSJohn Heffner 18095d424d5aSJohn Heffner /* Not currently probing/verifying, 18105d424d5aSJohn Heffner * not in recovery, 18115d424d5aSJohn Heffner * have enough cwnd, and 18125d424d5aSJohn Heffner * not SACKing (the variable headers throw things off) */ 18135d424d5aSJohn Heffner if (!icsk->icsk_mtup.enabled || 18145d424d5aSJohn Heffner icsk->icsk_mtup.probe_size || 18155d424d5aSJohn Heffner inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 18165d424d5aSJohn Heffner tp->snd_cwnd < 11 || 1817cabeccbdSIlpo Järvinen tp->rx_opt.num_sacks || tp->rx_opt.dsack) 18185d424d5aSJohn Heffner return -1; 18195d424d5aSJohn Heffner 18205d424d5aSJohn Heffner /* Very simple search strategy: just double the MSS. */ 18210c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 18225d424d5aSJohn Heffner probe_size = 2 * tp->mss_cache; 182391cc17c0SIlpo Järvinen size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; 18245d424d5aSJohn Heffner if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) { 18255d424d5aSJohn Heffner /* TODO: set timer for probe_converge_event */ 18265d424d5aSJohn Heffner return -1; 18275d424d5aSJohn Heffner } 18285d424d5aSJohn Heffner 18295d424d5aSJohn Heffner /* Have enough data in the send queue to probe? */ 18307f9c33e5SIlpo Järvinen if (tp->write_seq - tp->snd_nxt < size_needed) 18315d424d5aSJohn Heffner return -1; 18325d424d5aSJohn Heffner 183391cc17c0SIlpo Järvinen if (tp->snd_wnd < size_needed) 18345d424d5aSJohn Heffner return -1; 183590840defSIlpo Järvinen if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) 18365d424d5aSJohn Heffner return 0; 18375d424d5aSJohn Heffner 1838d67c58e9SIlpo Järvinen /* Do we need to wait to drain cwnd? With none in flight, don't stall */ 1839d67c58e9SIlpo Järvinen if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { 1840d67c58e9SIlpo Järvinen if (!tcp_packets_in_flight(tp)) 18415d424d5aSJohn Heffner return -1; 18425d424d5aSJohn Heffner else 18435d424d5aSJohn Heffner return 0; 18445d424d5aSJohn Heffner } 18455d424d5aSJohn Heffner 18465d424d5aSJohn Heffner /* We're allowed to probe. Build it now. */ 18475d424d5aSJohn Heffner if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL) 18485d424d5aSJohn Heffner return -1; 18493ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 18503ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 18515d424d5aSJohn Heffner 1852fe067e8aSDavid S. Miller skb = tcp_send_head(sk); 18535d424d5aSJohn Heffner 18545d424d5aSJohn Heffner TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 18555d424d5aSJohn Heffner TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 18564de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; 18575d424d5aSJohn Heffner TCP_SKB_CB(nskb)->sacked = 0; 18585d424d5aSJohn Heffner nskb->csum = 0; 185984fa7933SPatrick McHardy nskb->ip_summed = skb->ip_summed; 18605d424d5aSJohn Heffner 186150c4817eSIlpo Järvinen tcp_insert_write_queue_before(nskb, skb, sk); 186250c4817eSIlpo Järvinen 18635d424d5aSJohn Heffner len = 0; 1864234b6860SIlpo Järvinen tcp_for_write_queue_from_safe(skb, next, sk) { 18655d424d5aSJohn Heffner copy = min_t(int, skb->len, probe_size - len); 18665d424d5aSJohn Heffner if (nskb->ip_summed) 18675d424d5aSJohn Heffner skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); 18685d424d5aSJohn Heffner else 18695d424d5aSJohn Heffner nskb->csum = skb_copy_and_csum_bits(skb, 0, 1870056834d9SIlpo Järvinen skb_put(nskb, copy), 1871056834d9SIlpo Järvinen copy, nskb->csum); 18725d424d5aSJohn Heffner 18735d424d5aSJohn Heffner if (skb->len <= copy) { 18745d424d5aSJohn Heffner /* We've eaten all the data from this skb. 18755d424d5aSJohn Heffner * Throw it away. */ 18764de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 1877fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 18783ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 18795d424d5aSJohn Heffner } else { 18804de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & 1881a3433f35SChangli Gao ~(TCPHDR_FIN|TCPHDR_PSH); 18825d424d5aSJohn Heffner if (!skb_shinfo(skb)->nr_frags) { 18835d424d5aSJohn Heffner skb_pull(skb, copy); 188484fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 1885056834d9SIlpo Järvinen skb->csum = csum_partial(skb->data, 1886056834d9SIlpo Järvinen skb->len, 0); 18875d424d5aSJohn Heffner } else { 18885d424d5aSJohn Heffner __pskb_trim_head(skb, copy); 18895d424d5aSJohn Heffner tcp_set_skb_tso_segs(sk, skb, mss_now); 18905d424d5aSJohn Heffner } 18915d424d5aSJohn Heffner TCP_SKB_CB(skb)->seq += copy; 18925d424d5aSJohn Heffner } 18935d424d5aSJohn Heffner 18945d424d5aSJohn Heffner len += copy; 1895234b6860SIlpo Järvinen 1896234b6860SIlpo Järvinen if (len >= probe_size) 1897234b6860SIlpo Järvinen break; 18985d424d5aSJohn Heffner } 18995d424d5aSJohn Heffner tcp_init_tso_segs(sk, nskb, nskb->len); 19005d424d5aSJohn Heffner 19015d424d5aSJohn Heffner /* We're ready to send. If this fails, the probe will 19025d424d5aSJohn Heffner * be resegmented into mss-sized pieces by tcp_write_xmit(). */ 19035d424d5aSJohn Heffner TCP_SKB_CB(nskb)->when = tcp_time_stamp; 19045d424d5aSJohn Heffner if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 19055d424d5aSJohn Heffner /* Decrement cwnd here because we are sending 19065d424d5aSJohn Heffner * effectively two packets. */ 19075d424d5aSJohn Heffner tp->snd_cwnd--; 190866f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, nskb); 19095d424d5aSJohn Heffner 19105d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 19110e7b1368SJohn Heffner tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 19120e7b1368SJohn Heffner tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 19135d424d5aSJohn Heffner 19145d424d5aSJohn Heffner return 1; 19155d424d5aSJohn Heffner } 19165d424d5aSJohn Heffner 19175d424d5aSJohn Heffner return -1; 19185d424d5aSJohn Heffner } 19195d424d5aSJohn Heffner 19201da177e4SLinus Torvalds /* This routine writes packets to the network. It advances the 19211da177e4SLinus Torvalds * send_head. This happens as incoming acks open up the remote 19221da177e4SLinus Torvalds * window for us. 19231da177e4SLinus Torvalds * 1924f8269a49SIlpo Järvinen * LARGESEND note: !tcp_urg_mode is overkill, only frames between 1925f8269a49SIlpo Järvinen * snd_up-64k-mss .. snd_up cannot be large. However, taking into 1926f8269a49SIlpo Järvinen * account rare use of URG, this is not a big flaw. 1927f8269a49SIlpo Järvinen * 1928a2a385d6SEric Dumazet * Returns true, if no segments are in flight and we have queued segments, 1929a2a385d6SEric Dumazet * but cannot send anything now because of SWS or another problem. 19301da177e4SLinus Torvalds */ 1931a2a385d6SEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 1932d5dd9175SIlpo Järvinen int push_one, gfp_t gfp) 19331da177e4SLinus Torvalds { 19341da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 193592df7b51SDavid S. Miller struct sk_buff *skb; 1936c1b4a7e6SDavid S. Miller unsigned int tso_segs, sent_pkts; 1937c1b4a7e6SDavid S. Miller int cwnd_quota; 19385d424d5aSJohn Heffner int result; 19391da177e4SLinus Torvalds 1940c1b4a7e6SDavid S. Miller sent_pkts = 0; 19415d424d5aSJohn Heffner 1942d5dd9175SIlpo Järvinen if (!push_one) { 19435d424d5aSJohn Heffner /* Do MTU probing. */ 1944d5dd9175SIlpo Järvinen result = tcp_mtu_probe(sk); 1945d5dd9175SIlpo Järvinen if (!result) { 1946a2a385d6SEric Dumazet return false; 19475d424d5aSJohn Heffner } else if (result > 0) { 19485d424d5aSJohn Heffner sent_pkts = 1; 19495d424d5aSJohn Heffner } 1950d5dd9175SIlpo Järvinen } 19515d424d5aSJohn Heffner 1952fe067e8aSDavid S. Miller while ((skb = tcp_send_head(sk))) { 1953c8ac3774SHerbert Xu unsigned int limit; 1954c8ac3774SHerbert Xu 195546d3ceabSEric Dumazet 1956b68e9f85SHerbert Xu tso_segs = tcp_init_tso_segs(sk, skb, mss_now); 1957c1b4a7e6SDavid S. Miller BUG_ON(!tso_segs); 1958c1b4a7e6SDavid S. Miller 1959b68e9f85SHerbert Xu cwnd_quota = tcp_cwnd_test(tp, skb); 1960b68e9f85SHerbert Xu if (!cwnd_quota) 1961b68e9f85SHerbert Xu break; 1962b68e9f85SHerbert Xu 1963b68e9f85SHerbert Xu if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) 1964b68e9f85SHerbert Xu break; 1965b68e9f85SHerbert Xu 1966c1b4a7e6SDavid S. Miller if (tso_segs == 1) { 1967aa93466bSDavid S. Miller if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 1968aa93466bSDavid S. Miller (tcp_skb_is_last(sk, skb) ? 1969aa93466bSDavid S. Miller nonagle : TCP_NAGLE_PUSH)))) 1970aa93466bSDavid S. Miller break; 1971c1b4a7e6SDavid S. Miller } else { 1972d5dd9175SIlpo Järvinen if (!push_one && tcp_tso_should_defer(sk, skb)) 1973aa93466bSDavid S. Miller break; 1974c1b4a7e6SDavid S. Miller } 1975aa93466bSDavid S. Miller 197646d3ceabSEric Dumazet /* TSQ : sk_wmem_alloc accounts skb truesize, 197746d3ceabSEric Dumazet * including skb overhead. But thats OK. 197846d3ceabSEric Dumazet */ 197946d3ceabSEric Dumazet if (atomic_read(&sk->sk_wmem_alloc) >= sysctl_tcp_limit_output_bytes) { 198046d3ceabSEric Dumazet set_bit(TSQ_THROTTLED, &tp->tsq_flags); 198146d3ceabSEric Dumazet break; 198246d3ceabSEric Dumazet } 1983c8ac3774SHerbert Xu limit = mss_now; 1984f8269a49SIlpo Järvinen if (tso_segs > 1 && !tcp_urg_mode(tp)) 19850e3a4803SIlpo Järvinen limit = tcp_mss_split_point(sk, skb, mss_now, 19860e3a4803SIlpo Järvinen cwnd_quota); 1987c8ac3774SHerbert Xu 1988c8ac3774SHerbert Xu if (skb->len > limit && 1989c4ead4c5SEric Dumazet unlikely(tso_fragment(sk, skb, limit, mss_now, gfp))) 19901da177e4SLinus Torvalds break; 19911da177e4SLinus Torvalds 19921da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 1993c1b4a7e6SDavid S. Miller 1994d5dd9175SIlpo Järvinen if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) 19951da177e4SLinus Torvalds break; 19961da177e4SLinus Torvalds 19971da177e4SLinus Torvalds /* Advance the send_head. This one is sent out. 19981da177e4SLinus Torvalds * This call will increment packets_out. 19991da177e4SLinus Torvalds */ 200066f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 20011da177e4SLinus Torvalds 20021da177e4SLinus Torvalds tcp_minshall_update(tp, mss_now, skb); 2003a262f0cdSNandita Dukkipati sent_pkts += tcp_skb_pcount(skb); 2004d5dd9175SIlpo Järvinen 2005d5dd9175SIlpo Järvinen if (push_one) 2006d5dd9175SIlpo Järvinen break; 20071da177e4SLinus Torvalds } 2008a262f0cdSNandita Dukkipati if (inet_csk(sk)->icsk_ca_state == TCP_CA_Recovery) 2009a262f0cdSNandita Dukkipati tp->prr_out += sent_pkts; 20101da177e4SLinus Torvalds 2011aa93466bSDavid S. Miller if (likely(sent_pkts)) { 20129e412ba7SIlpo Järvinen tcp_cwnd_validate(sk); 2013a2a385d6SEric Dumazet return false; 20141da177e4SLinus Torvalds } 2015fe067e8aSDavid S. Miller return !tp->packets_out && tcp_send_head(sk); 20161da177e4SLinus Torvalds } 20171da177e4SLinus Torvalds 2018a762a980SDavid S. Miller /* Push out any pending frames which were held back due to 2019a762a980SDavid S. Miller * TCP_CORK or attempt at coalescing tiny packets. 2020a762a980SDavid S. Miller * The socket must be locked by the caller. 2021a762a980SDavid S. Miller */ 20229e412ba7SIlpo Järvinen void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 20239e412ba7SIlpo Järvinen int nonagle) 2024a762a980SDavid S. Miller { 2025726e07a8SIlpo Järvinen /* If we are closed, the bytes will have to remain here. 2026726e07a8SIlpo Järvinen * In time closedown will finish, we empty the write queue and 2027726e07a8SIlpo Järvinen * all will be happy. 2028726e07a8SIlpo Järvinen */ 2029726e07a8SIlpo Järvinen if (unlikely(sk->sk_state == TCP_CLOSE)) 2030726e07a8SIlpo Järvinen return; 2031726e07a8SIlpo Järvinen 2032d5dd9175SIlpo Järvinen if (tcp_write_xmit(sk, cur_mss, nonagle, 0, GFP_ATOMIC)) 20339e412ba7SIlpo Järvinen tcp_check_probe_timer(sk); 2034a762a980SDavid S. Miller } 2035a762a980SDavid S. Miller 2036c1b4a7e6SDavid S. Miller /* Send _single_ skb sitting at the send head. This function requires 2037c1b4a7e6SDavid S. Miller * true push pending frames to setup probe timer etc. 2038c1b4a7e6SDavid S. Miller */ 2039c1b4a7e6SDavid S. Miller void tcp_push_one(struct sock *sk, unsigned int mss_now) 2040c1b4a7e6SDavid S. Miller { 2041fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 2042c1b4a7e6SDavid S. Miller 2043c1b4a7e6SDavid S. Miller BUG_ON(!skb || skb->len < mss_now); 2044c1b4a7e6SDavid S. Miller 2045d5dd9175SIlpo Järvinen tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); 2046c1b4a7e6SDavid S. Miller } 2047c1b4a7e6SDavid S. Miller 20481da177e4SLinus Torvalds /* This function returns the amount that we can raise the 20491da177e4SLinus Torvalds * usable window based on the following constraints 20501da177e4SLinus Torvalds * 20511da177e4SLinus Torvalds * 1. The window can never be shrunk once it is offered (RFC 793) 20521da177e4SLinus Torvalds * 2. We limit memory per socket 20531da177e4SLinus Torvalds * 20541da177e4SLinus Torvalds * RFC 1122: 20551da177e4SLinus Torvalds * "the suggested [SWS] avoidance algorithm for the receiver is to keep 20561da177e4SLinus Torvalds * RECV.NEXT + RCV.WIN fixed until: 20571da177e4SLinus Torvalds * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 20581da177e4SLinus Torvalds * 20591da177e4SLinus Torvalds * i.e. don't raise the right edge of the window until you can raise 20601da177e4SLinus Torvalds * it at least MSS bytes. 20611da177e4SLinus Torvalds * 20621da177e4SLinus Torvalds * Unfortunately, the recommended algorithm breaks header prediction, 20631da177e4SLinus Torvalds * since header prediction assumes th->window stays fixed. 20641da177e4SLinus Torvalds * 20651da177e4SLinus Torvalds * Strictly speaking, keeping th->window fixed violates the receiver 20661da177e4SLinus Torvalds * side SWS prevention criteria. The problem is that under this rule 20671da177e4SLinus Torvalds * a stream of single byte packets will cause the right side of the 20681da177e4SLinus Torvalds * window to always advance by a single byte. 20691da177e4SLinus Torvalds * 20701da177e4SLinus Torvalds * Of course, if the sender implements sender side SWS prevention 20711da177e4SLinus Torvalds * then this will not be a problem. 20721da177e4SLinus Torvalds * 20731da177e4SLinus Torvalds * BSD seems to make the following compromise: 20741da177e4SLinus Torvalds * 20751da177e4SLinus Torvalds * If the free space is less than the 1/4 of the maximum 20761da177e4SLinus Torvalds * space available and the free space is less than 1/2 mss, 20771da177e4SLinus Torvalds * then set the window to 0. 20781da177e4SLinus Torvalds * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 20791da177e4SLinus Torvalds * Otherwise, just prevent the window from shrinking 20801da177e4SLinus Torvalds * and from being larger than the largest representable value. 20811da177e4SLinus Torvalds * 20821da177e4SLinus Torvalds * This prevents incremental opening of the window in the regime 20831da177e4SLinus Torvalds * where TCP is limited by the speed of the reader side taking 20841da177e4SLinus Torvalds * data out of the TCP receive queue. It does nothing about 20851da177e4SLinus Torvalds * those cases where the window is constrained on the sender side 20861da177e4SLinus Torvalds * because the pipeline is full. 20871da177e4SLinus Torvalds * 20881da177e4SLinus Torvalds * BSD also seems to "accidentally" limit itself to windows that are a 20891da177e4SLinus Torvalds * multiple of MSS, at least until the free space gets quite small. 20901da177e4SLinus Torvalds * This would appear to be a side effect of the mbuf implementation. 20911da177e4SLinus Torvalds * Combining these two algorithms results in the observed behavior 20921da177e4SLinus Torvalds * of having a fixed window size at almost all times. 20931da177e4SLinus Torvalds * 20941da177e4SLinus Torvalds * Below we obtain similar behavior by forcing the offered window to 20951da177e4SLinus Torvalds * a multiple of the mss when it is feasible to do so. 20961da177e4SLinus Torvalds * 20971da177e4SLinus Torvalds * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 20981da177e4SLinus Torvalds * Regular options like TIMESTAMP are taken into account. 20991da177e4SLinus Torvalds */ 21001da177e4SLinus Torvalds u32 __tcp_select_window(struct sock *sk) 21011da177e4SLinus Torvalds { 2102463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 21031da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2104caa20d9aSStephen Hemminger /* MSS for the peer's data. Previous versions used mss_clamp 21051da177e4SLinus Torvalds * here. I don't know if the value based on our guesses 21061da177e4SLinus Torvalds * of peer's MSS is better for the performance. It's more correct 21071da177e4SLinus Torvalds * but may be worse for the performance because of rcv_mss 21081da177e4SLinus Torvalds * fluctuations. --SAW 1998/11/1 21091da177e4SLinus Torvalds */ 2110463c84b9SArnaldo Carvalho de Melo int mss = icsk->icsk_ack.rcv_mss; 21111da177e4SLinus Torvalds int free_space = tcp_space(sk); 21121da177e4SLinus Torvalds int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk)); 21131da177e4SLinus Torvalds int window; 21141da177e4SLinus Torvalds 21151da177e4SLinus Torvalds if (mss > full_space) 21161da177e4SLinus Torvalds mss = full_space; 21171da177e4SLinus Torvalds 2118b92edbe0SEric Dumazet if (free_space < (full_space >> 1)) { 2119463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.quick = 0; 21201da177e4SLinus Torvalds 2121180d8cd9SGlauber Costa if (sk_under_memory_pressure(sk)) 2122056834d9SIlpo Järvinen tp->rcv_ssthresh = min(tp->rcv_ssthresh, 2123056834d9SIlpo Järvinen 4U * tp->advmss); 21241da177e4SLinus Torvalds 21251da177e4SLinus Torvalds if (free_space < mss) 21261da177e4SLinus Torvalds return 0; 21271da177e4SLinus Torvalds } 21281da177e4SLinus Torvalds 21291da177e4SLinus Torvalds if (free_space > tp->rcv_ssthresh) 21301da177e4SLinus Torvalds free_space = tp->rcv_ssthresh; 21311da177e4SLinus Torvalds 21321da177e4SLinus Torvalds /* Don't do rounding if we are using window scaling, since the 21331da177e4SLinus Torvalds * scaled window will not line up with the MSS boundary anyway. 21341da177e4SLinus Torvalds */ 21351da177e4SLinus Torvalds window = tp->rcv_wnd; 21361da177e4SLinus Torvalds if (tp->rx_opt.rcv_wscale) { 21371da177e4SLinus Torvalds window = free_space; 21381da177e4SLinus Torvalds 21391da177e4SLinus Torvalds /* Advertise enough space so that it won't get scaled away. 21401da177e4SLinus Torvalds * Import case: prevent zero window announcement if 21411da177e4SLinus Torvalds * 1<<rcv_wscale > mss. 21421da177e4SLinus Torvalds */ 21431da177e4SLinus Torvalds if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window) 21441da177e4SLinus Torvalds window = (((window >> tp->rx_opt.rcv_wscale) + 1) 21451da177e4SLinus Torvalds << tp->rx_opt.rcv_wscale); 21461da177e4SLinus Torvalds } else { 21471da177e4SLinus Torvalds /* Get the largest window that is a nice multiple of mss. 21481da177e4SLinus Torvalds * Window clamp already applied above. 21491da177e4SLinus Torvalds * If our current window offering is within 1 mss of the 21501da177e4SLinus Torvalds * free space we just keep it. This prevents the divide 21511da177e4SLinus Torvalds * and multiply from happening most of the time. 21521da177e4SLinus Torvalds * We also don't do any window rounding when the free space 21531da177e4SLinus Torvalds * is too small. 21541da177e4SLinus Torvalds */ 21551da177e4SLinus Torvalds if (window <= free_space - mss || window > free_space) 21561da177e4SLinus Torvalds window = (free_space / mss) * mss; 215784565070SJohn Heffner else if (mss == full_space && 2158b92edbe0SEric Dumazet free_space > window + (full_space >> 1)) 215984565070SJohn Heffner window = free_space; 21601da177e4SLinus Torvalds } 21611da177e4SLinus Torvalds 21621da177e4SLinus Torvalds return window; 21631da177e4SLinus Torvalds } 21641da177e4SLinus Torvalds 21654a17fc3aSIlpo Järvinen /* Collapses two adjacent SKB's during retransmission. */ 21664a17fc3aSIlpo Järvinen static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) 21671da177e4SLinus Torvalds { 21681da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2169fe067e8aSDavid S. Miller struct sk_buff *next_skb = tcp_write_queue_next(sk, skb); 2170058dc334SIlpo Järvinen int skb_size, next_skb_size; 21711da177e4SLinus Torvalds 2172058dc334SIlpo Järvinen skb_size = skb->len; 2173058dc334SIlpo Järvinen next_skb_size = next_skb->len; 21741da177e4SLinus Torvalds 2175058dc334SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); 21761da177e4SLinus Torvalds 21776859d494SIlpo Järvinen tcp_highest_sack_combine(sk, next_skb, skb); 2178a6963a6bSIlpo Järvinen 2179fe067e8aSDavid S. Miller tcp_unlink_write_queue(next_skb, sk); 21801da177e4SLinus Torvalds 2181058dc334SIlpo Järvinen skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size), 21821a4e2d09SArnaldo Carvalho de Melo next_skb_size); 21831da177e4SLinus Torvalds 218452d570aaSJarek Poplawski if (next_skb->ip_summed == CHECKSUM_PARTIAL) 218552d570aaSJarek Poplawski skb->ip_summed = CHECKSUM_PARTIAL; 21861da177e4SLinus Torvalds 218784fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 21881da177e4SLinus Torvalds skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); 21891da177e4SLinus Torvalds 21901da177e4SLinus Torvalds /* Update sequence range on original skb. */ 21911da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 21921da177e4SLinus Torvalds 2193e6c7d085SIlpo Järvinen /* Merge over control information. This moves PSH/FIN etc. over */ 21944de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; 21951da177e4SLinus Torvalds 21961da177e4SLinus Torvalds /* All done, get rid of second SKB and account for it so 21971da177e4SLinus Torvalds * packet counting does not break. 21981da177e4SLinus Torvalds */ 21994828e7f4SIlpo Järvinen TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; 2200b7689205SIlpo Järvinen 2201b7689205SIlpo Järvinen /* changed transmit queue under us so clear hints */ 2202ef9da47cSIlpo Järvinen tcp_clear_retrans_hints_partial(tp); 2203ef9da47cSIlpo Järvinen if (next_skb == tp->retransmit_skb_hint) 2204ef9da47cSIlpo Järvinen tp->retransmit_skb_hint = skb; 2205b7689205SIlpo Järvinen 2206797108d1SIlpo Järvinen tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); 2207797108d1SIlpo Järvinen 22083ab224beSHideo Aoki sk_wmem_free_skb(sk, next_skb); 22091da177e4SLinus Torvalds } 22101da177e4SLinus Torvalds 221167edfef7SAndi Kleen /* Check if coalescing SKBs is legal. */ 2212a2a385d6SEric Dumazet static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) 22134a17fc3aSIlpo Järvinen { 22144a17fc3aSIlpo Järvinen if (tcp_skb_pcount(skb) > 1) 2215a2a385d6SEric Dumazet return false; 22164a17fc3aSIlpo Järvinen /* TODO: SACK collapsing could be used to remove this condition */ 22174a17fc3aSIlpo Järvinen if (skb_shinfo(skb)->nr_frags != 0) 2218a2a385d6SEric Dumazet return false; 22194a17fc3aSIlpo Järvinen if (skb_cloned(skb)) 2220a2a385d6SEric Dumazet return false; 22214a17fc3aSIlpo Järvinen if (skb == tcp_send_head(sk)) 2222a2a385d6SEric Dumazet return false; 22234a17fc3aSIlpo Järvinen /* Some heurestics for collapsing over SACK'd could be invented */ 22244a17fc3aSIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 2225a2a385d6SEric Dumazet return false; 22264a17fc3aSIlpo Järvinen 2227a2a385d6SEric Dumazet return true; 22284a17fc3aSIlpo Järvinen } 22294a17fc3aSIlpo Järvinen 223067edfef7SAndi Kleen /* Collapse packets in the retransmit queue to make to create 223167edfef7SAndi Kleen * less packets on the wire. This is only done on retransmission. 223267edfef7SAndi Kleen */ 22334a17fc3aSIlpo Järvinen static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, 22344a17fc3aSIlpo Järvinen int space) 22354a17fc3aSIlpo Järvinen { 22364a17fc3aSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 22374a17fc3aSIlpo Järvinen struct sk_buff *skb = to, *tmp; 2238a2a385d6SEric Dumazet bool first = true; 22394a17fc3aSIlpo Järvinen 22404a17fc3aSIlpo Järvinen if (!sysctl_tcp_retrans_collapse) 22414a17fc3aSIlpo Järvinen return; 22424de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 22434a17fc3aSIlpo Järvinen return; 22444a17fc3aSIlpo Järvinen 22454a17fc3aSIlpo Järvinen tcp_for_write_queue_from_safe(skb, tmp, sk) { 22464a17fc3aSIlpo Järvinen if (!tcp_can_collapse(sk, skb)) 22474a17fc3aSIlpo Järvinen break; 22484a17fc3aSIlpo Järvinen 22494a17fc3aSIlpo Järvinen space -= skb->len; 22504a17fc3aSIlpo Järvinen 22514a17fc3aSIlpo Järvinen if (first) { 2252a2a385d6SEric Dumazet first = false; 22534a17fc3aSIlpo Järvinen continue; 22544a17fc3aSIlpo Järvinen } 22554a17fc3aSIlpo Järvinen 22564a17fc3aSIlpo Järvinen if (space < 0) 22574a17fc3aSIlpo Järvinen break; 22584a17fc3aSIlpo Järvinen /* Punt if not enough space exists in the first SKB for 22594a17fc3aSIlpo Järvinen * the data in the second 22604a17fc3aSIlpo Järvinen */ 2261a21d4572SEric Dumazet if (skb->len > skb_availroom(to)) 22624a17fc3aSIlpo Järvinen break; 22634a17fc3aSIlpo Järvinen 22644a17fc3aSIlpo Järvinen if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) 22654a17fc3aSIlpo Järvinen break; 22664a17fc3aSIlpo Järvinen 22674a17fc3aSIlpo Järvinen tcp_collapse_retrans(sk, to); 22684a17fc3aSIlpo Järvinen } 22694a17fc3aSIlpo Järvinen } 22704a17fc3aSIlpo Järvinen 22711da177e4SLinus Torvalds /* This retransmits one SKB. Policy decisions and retransmit queue 22721da177e4SLinus Torvalds * state updates are done by the caller. Returns non-zero if an 22731da177e4SLinus Torvalds * error occurred which prevented the send. 22741da177e4SLinus Torvalds */ 22751da177e4SLinus Torvalds int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) 22761da177e4SLinus Torvalds { 22771da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 22785d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 22797d227cd2SSridhar Samudrala unsigned int cur_mss; 22801da177e4SLinus Torvalds int err; 22811da177e4SLinus Torvalds 22825d424d5aSJohn Heffner /* Inconslusive MTU probe */ 22835d424d5aSJohn Heffner if (icsk->icsk_mtup.probe_size) { 22845d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 22855d424d5aSJohn Heffner } 22865d424d5aSJohn Heffner 22871da177e4SLinus Torvalds /* Do not sent more than we queued. 1/4 is reserved for possible 2288caa20d9aSStephen Hemminger * copying overhead: fragmentation, tunneling, mangling etc. 22891da177e4SLinus Torvalds */ 22901da177e4SLinus Torvalds if (atomic_read(&sk->sk_wmem_alloc) > 22911da177e4SLinus Torvalds min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf)) 22921da177e4SLinus Torvalds return -EAGAIN; 22931da177e4SLinus Torvalds 22941da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 22951da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 22961da177e4SLinus Torvalds BUG(); 22971da177e4SLinus Torvalds if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 22981da177e4SLinus Torvalds return -ENOMEM; 22991da177e4SLinus Torvalds } 23001da177e4SLinus Torvalds 23017d227cd2SSridhar Samudrala if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 23027d227cd2SSridhar Samudrala return -EHOSTUNREACH; /* Routing failure or similar. */ 23037d227cd2SSridhar Samudrala 23040c54b85fSIlpo Järvinen cur_mss = tcp_current_mss(sk); 23057d227cd2SSridhar Samudrala 23061da177e4SLinus Torvalds /* If receiver has shrunk his window, and skb is out of 23071da177e4SLinus Torvalds * new window, do not retransmit it. The exception is the 23081da177e4SLinus Torvalds * case, when window is shrunk to zero. In this case 23091da177e4SLinus Torvalds * our retransmit serves as a zero window probe. 23101da177e4SLinus Torvalds */ 23119d4fb27dSJoe Perches if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && 23129d4fb27dSJoe Perches TCP_SKB_CB(skb)->seq != tp->snd_una) 23131da177e4SLinus Torvalds return -EAGAIN; 23141da177e4SLinus Torvalds 23151da177e4SLinus Torvalds if (skb->len > cur_mss) { 2316846998aeSDavid S. Miller if (tcp_fragment(sk, skb, cur_mss, cur_mss)) 23171da177e4SLinus Torvalds return -ENOMEM; /* We'll try again later. */ 231802276f3cSIlpo Järvinen } else { 23199eb9362eSIlpo Järvinen int oldpcount = tcp_skb_pcount(skb); 23209eb9362eSIlpo Järvinen 23219eb9362eSIlpo Järvinen if (unlikely(oldpcount > 1)) { 232202276f3cSIlpo Järvinen tcp_init_tso_segs(sk, skb, cur_mss); 23239eb9362eSIlpo Järvinen tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb)); 23249eb9362eSIlpo Järvinen } 23251da177e4SLinus Torvalds } 23261da177e4SLinus Torvalds 23271da177e4SLinus Torvalds tcp_retrans_try_collapse(sk, skb, cur_mss); 23281da177e4SLinus Torvalds 23291da177e4SLinus Torvalds /* Some Solaris stacks overoptimize and ignore the FIN on a 23301da177e4SLinus Torvalds * retransmit when old data is attached. So strip it off 23311da177e4SLinus Torvalds * since it is cheap to do so and saves bytes on the network. 23321da177e4SLinus Torvalds */ 23331da177e4SLinus Torvalds if (skb->len > 0 && 23344de075e0SEric Dumazet (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && 23351da177e4SLinus Torvalds tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) { 23361da177e4SLinus Torvalds if (!pskb_trim(skb, 0)) { 2337e870a8efSIlpo Järvinen /* Reuse, even though it does some unnecessary work */ 2338e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1, 23394de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags); 23401da177e4SLinus Torvalds skb->ip_summed = CHECKSUM_NONE; 23411da177e4SLinus Torvalds } 23421da177e4SLinus Torvalds } 23431da177e4SLinus Torvalds 23441da177e4SLinus Torvalds /* Make a copy, if the first transmission SKB clone we made 23451da177e4SLinus Torvalds * is still in somebody's hands, else make a clone. 23461da177e4SLinus Torvalds */ 23471da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 23481da177e4SLinus Torvalds 2349117632e6SEric Dumazet /* make sure skb->data is aligned on arches that require it */ 2350117632e6SEric Dumazet if (unlikely(NET_IP_ALIGN && ((unsigned long)skb->data & 3))) { 2351117632e6SEric Dumazet struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER, 2352117632e6SEric Dumazet GFP_ATOMIC); 2353117632e6SEric Dumazet err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) : 2354117632e6SEric Dumazet -ENOBUFS; 2355117632e6SEric Dumazet } else { 2356dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 2357117632e6SEric Dumazet } 23581da177e4SLinus Torvalds 23591da177e4SLinus Torvalds if (err == 0) { 23601da177e4SLinus Torvalds /* Update global TCP statistics. */ 236181cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); 23621da177e4SLinus Torvalds 23631da177e4SLinus Torvalds tp->total_retrans++; 23641da177e4SLinus Torvalds 23651da177e4SLinus Torvalds #if FASTRETRANS_DEBUG > 0 23661da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 2367e87cc472SJoe Perches net_dbg_ratelimited("retrans_out leaked\n"); 23681da177e4SLinus Torvalds } 23691da177e4SLinus Torvalds #endif 2370b08d6cb2SIlpo Järvinen if (!tp->retrans_out) 2371b08d6cb2SIlpo Järvinen tp->lost_retrans_low = tp->snd_nxt; 23721da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 23731da177e4SLinus Torvalds tp->retrans_out += tcp_skb_pcount(skb); 23741da177e4SLinus Torvalds 23751da177e4SLinus Torvalds /* Save stamp of the first retransmit. */ 23761da177e4SLinus Torvalds if (!tp->retrans_stamp) 23771da177e4SLinus Torvalds tp->retrans_stamp = TCP_SKB_CB(skb)->when; 23781da177e4SLinus Torvalds 2379c24f691bSYuchung Cheng tp->undo_retrans += tcp_skb_pcount(skb); 23801da177e4SLinus Torvalds 23811da177e4SLinus Torvalds /* snd_nxt is stored to detect loss of retransmitted segment, 23821da177e4SLinus Torvalds * see tcp_input.c tcp_sacktag_write_queue(). 23831da177e4SLinus Torvalds */ 23841da177e4SLinus Torvalds TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt; 23851da177e4SLinus Torvalds } 23861da177e4SLinus Torvalds return err; 23871da177e4SLinus Torvalds } 23881da177e4SLinus Torvalds 238967edfef7SAndi Kleen /* Check if we forward retransmits are possible in the current 239067edfef7SAndi Kleen * window/congestion state. 239167edfef7SAndi Kleen */ 2392a2a385d6SEric Dumazet static bool tcp_can_forward_retransmit(struct sock *sk) 2393b5afe7bcSIlpo Järvinen { 2394b5afe7bcSIlpo Järvinen const struct inet_connection_sock *icsk = inet_csk(sk); 2395cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 2396b5afe7bcSIlpo Järvinen 2397b5afe7bcSIlpo Järvinen /* Forward retransmissions are possible only during Recovery. */ 2398b5afe7bcSIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Recovery) 2399a2a385d6SEric Dumazet return false; 2400b5afe7bcSIlpo Järvinen 2401b5afe7bcSIlpo Järvinen /* No forward retransmissions in Reno are possible. */ 2402b5afe7bcSIlpo Järvinen if (tcp_is_reno(tp)) 2403a2a385d6SEric Dumazet return false; 2404b5afe7bcSIlpo Järvinen 2405b5afe7bcSIlpo Järvinen /* Yeah, we have to make difficult choice between forward transmission 2406b5afe7bcSIlpo Järvinen * and retransmission... Both ways have their merits... 2407b5afe7bcSIlpo Järvinen * 2408b5afe7bcSIlpo Järvinen * For now we do not retransmit anything, while we have some new 2409b5afe7bcSIlpo Järvinen * segments to send. In the other cases, follow rule 3 for 2410b5afe7bcSIlpo Järvinen * NextSeg() specified in RFC3517. 2411b5afe7bcSIlpo Järvinen */ 2412b5afe7bcSIlpo Järvinen 2413b5afe7bcSIlpo Järvinen if (tcp_may_send_now(sk)) 2414a2a385d6SEric Dumazet return false; 2415b5afe7bcSIlpo Järvinen 2416a2a385d6SEric Dumazet return true; 2417b5afe7bcSIlpo Järvinen } 2418b5afe7bcSIlpo Järvinen 24191da177e4SLinus Torvalds /* This gets called after a retransmit timeout, and the initially 24201da177e4SLinus Torvalds * retransmitted data is acknowledged. It tries to continue 24211da177e4SLinus Torvalds * resending the rest of the retransmit queue, until either 24221da177e4SLinus Torvalds * we've sent it all or the congestion window limit is reached. 24231da177e4SLinus Torvalds * If doing SACK, the first ACK which comes back for a timeout 24241da177e4SLinus Torvalds * based retransmit packet might feed us FACK information again. 24251da177e4SLinus Torvalds * If so, we use it to avoid unnecessarily retransmissions. 24261da177e4SLinus Torvalds */ 24271da177e4SLinus Torvalds void tcp_xmit_retransmit_queue(struct sock *sk) 24281da177e4SLinus Torvalds { 24296687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 24301da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 24311da177e4SLinus Torvalds struct sk_buff *skb; 24320e1c54c2SIlpo Järvinen struct sk_buff *hole = NULL; 2433618d9f25SIlpo Järvinen u32 last_lost; 243461eb55f4SIlpo Järvinen int mib_idx; 24350e1c54c2SIlpo Järvinen int fwd_rexmitting = 0; 24366a438bbeSStephen Hemminger 243745e77d31SIlpo Järvinen if (!tp->packets_out) 243845e77d31SIlpo Järvinen return; 243945e77d31SIlpo Järvinen 244008ebd172SIlpo Järvinen if (!tp->lost_out) 244108ebd172SIlpo Järvinen tp->retransmit_high = tp->snd_una; 244208ebd172SIlpo Järvinen 2443618d9f25SIlpo Järvinen if (tp->retransmit_skb_hint) { 24446a438bbeSStephen Hemminger skb = tp->retransmit_skb_hint; 2445618d9f25SIlpo Järvinen last_lost = TCP_SKB_CB(skb)->end_seq; 2446618d9f25SIlpo Järvinen if (after(last_lost, tp->retransmit_high)) 2447618d9f25SIlpo Järvinen last_lost = tp->retransmit_high; 2448618d9f25SIlpo Järvinen } else { 2449fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 2450618d9f25SIlpo Järvinen last_lost = tp->snd_una; 2451618d9f25SIlpo Järvinen } 24521da177e4SLinus Torvalds 2453fe067e8aSDavid S. Miller tcp_for_write_queue_from(skb, sk) { 24541da177e4SLinus Torvalds __u8 sacked = TCP_SKB_CB(skb)->sacked; 24551da177e4SLinus Torvalds 2456fe067e8aSDavid S. Miller if (skb == tcp_send_head(sk)) 2457fe067e8aSDavid S. Miller break; 24586a438bbeSStephen Hemminger /* we could do better than to assign each time */ 24590e1c54c2SIlpo Järvinen if (hole == NULL) 24606a438bbeSStephen Hemminger tp->retransmit_skb_hint = skb; 24616a438bbeSStephen Hemminger 24621da177e4SLinus Torvalds /* Assume this retransmit will generate 24631da177e4SLinus Torvalds * only one packet for congestion window 24641da177e4SLinus Torvalds * calculation purposes. This works because 24651da177e4SLinus Torvalds * tcp_retransmit_skb() will chop up the 24661da177e4SLinus Torvalds * packet to be MSS sized and all the 24671da177e4SLinus Torvalds * packet counting works out. 24681da177e4SLinus Torvalds */ 24691da177e4SLinus Torvalds if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) 24701da177e4SLinus Torvalds return; 24710e1c54c2SIlpo Järvinen 24720e1c54c2SIlpo Järvinen if (fwd_rexmitting) { 24730e1c54c2SIlpo Järvinen begin_fwd: 24740e1c54c2SIlpo Järvinen if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp))) 2475006f582cSIlpo Järvinen break; 24760e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFORWARDRETRANS; 24770e1c54c2SIlpo Järvinen 24780e1c54c2SIlpo Järvinen } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) { 2479618d9f25SIlpo Järvinen tp->retransmit_high = last_lost; 24800e1c54c2SIlpo Järvinen if (!tcp_can_forward_retransmit(sk)) 24810e1c54c2SIlpo Järvinen break; 24820e1c54c2SIlpo Järvinen /* Backtrack if necessary to non-L'ed skb */ 24830e1c54c2SIlpo Järvinen if (hole != NULL) { 24840e1c54c2SIlpo Järvinen skb = hole; 24850e1c54c2SIlpo Järvinen hole = NULL; 24860e1c54c2SIlpo Järvinen } 24870e1c54c2SIlpo Järvinen fwd_rexmitting = 1; 24880e1c54c2SIlpo Järvinen goto begin_fwd; 24890e1c54c2SIlpo Järvinen 24900e1c54c2SIlpo Järvinen } else if (!(sacked & TCPCB_LOST)) { 2491ac11ba75SIlpo Järvinen if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) 24920e1c54c2SIlpo Järvinen hole = skb; 249361eb55f4SIlpo Järvinen continue; 24941da177e4SLinus Torvalds 24950e1c54c2SIlpo Järvinen } else { 2496618d9f25SIlpo Järvinen last_lost = TCP_SKB_CB(skb)->end_seq; 24970e1c54c2SIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Loss) 24980e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFASTRETRANS; 24990e1c54c2SIlpo Järvinen else 25000e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; 25010e1c54c2SIlpo Järvinen } 25020e1c54c2SIlpo Järvinen 25030e1c54c2SIlpo Järvinen if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) 250461eb55f4SIlpo Järvinen continue; 250540b215e5SPavel Emelyanov 250609e9b813SEric Dumazet if (tcp_retransmit_skb(sk, skb)) { 250709e9b813SEric Dumazet NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL); 25081da177e4SLinus Torvalds return; 250909e9b813SEric Dumazet } 2510de0744afSPavel Emelyanov NET_INC_STATS_BH(sock_net(sk), mib_idx); 25111da177e4SLinus Torvalds 2512a262f0cdSNandita Dukkipati if (inet_csk(sk)->icsk_ca_state == TCP_CA_Recovery) 2513a262f0cdSNandita Dukkipati tp->prr_out += tcp_skb_pcount(skb); 2514a262f0cdSNandita Dukkipati 2515fe067e8aSDavid S. Miller if (skb == tcp_write_queue_head(sk)) 2516463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 25173f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, 25183f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 25191da177e4SLinus Torvalds } 25201da177e4SLinus Torvalds } 25211da177e4SLinus Torvalds 25221da177e4SLinus Torvalds /* Send a fin. The caller locks the socket for us. This cannot be 25231da177e4SLinus Torvalds * allowed to fail queueing a FIN frame under any circumstances. 25241da177e4SLinus Torvalds */ 25251da177e4SLinus Torvalds void tcp_send_fin(struct sock *sk) 25261da177e4SLinus Torvalds { 25271da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2528fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_write_queue_tail(sk); 25291da177e4SLinus Torvalds int mss_now; 25301da177e4SLinus Torvalds 25311da177e4SLinus Torvalds /* Optimization, tack on the FIN if we have a queue of 25321da177e4SLinus Torvalds * unsent frames. But be careful about outgoing SACKS 25331da177e4SLinus Torvalds * and IP options. 25341da177e4SLinus Torvalds */ 25350c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 25361da177e4SLinus Torvalds 2537fe067e8aSDavid S. Miller if (tcp_send_head(sk) != NULL) { 25384de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_FIN; 25391da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq++; 25401da177e4SLinus Torvalds tp->write_seq++; 25411da177e4SLinus Torvalds } else { 25421da177e4SLinus Torvalds /* Socket is locked, keep trying until memory is available. */ 25431da177e4SLinus Torvalds for (;;) { 2544aa133076SWu Fengguang skb = alloc_skb_fclone(MAX_TCP_HEADER, 2545aa133076SWu Fengguang sk->sk_allocation); 25461da177e4SLinus Torvalds if (skb) 25471da177e4SLinus Torvalds break; 25481da177e4SLinus Torvalds yield(); 25491da177e4SLinus Torvalds } 25501da177e4SLinus Torvalds 25511da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 25521da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 25531da177e4SLinus Torvalds /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 2554e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tp->write_seq, 2555a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_FIN); 25561da177e4SLinus Torvalds tcp_queue_skb(sk, skb); 25571da177e4SLinus Torvalds } 25589e412ba7SIlpo Järvinen __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF); 25591da177e4SLinus Torvalds } 25601da177e4SLinus Torvalds 25611da177e4SLinus Torvalds /* We get here when a process closes a file descriptor (either due to 25621da177e4SLinus Torvalds * an explicit close() or as a byproduct of exit()'ing) and there 25631da177e4SLinus Torvalds * was unread data in the receive queue. This behavior is recommended 256465bb723cSGerrit Renker * by RFC 2525, section 2.17. -DaveM 25651da177e4SLinus Torvalds */ 2566dd0fc66fSAl Viro void tcp_send_active_reset(struct sock *sk, gfp_t priority) 25671da177e4SLinus Torvalds { 25681da177e4SLinus Torvalds struct sk_buff *skb; 25691da177e4SLinus Torvalds 25701da177e4SLinus Torvalds /* NOTE: No TCP options attached and we never retransmit this. */ 25711da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, priority); 25721da177e4SLinus Torvalds if (!skb) { 25734e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 25741da177e4SLinus Torvalds return; 25751da177e4SLinus Torvalds } 25761da177e4SLinus Torvalds 25771da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 25781da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 2579e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), 2580a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_RST); 25811da177e4SLinus Torvalds /* Send it off. */ 25821da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2583dfb4b9dcSDavid S. Miller if (tcp_transmit_skb(sk, skb, 0, priority)) 25844e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 258526af65cbSSridhar Samudrala 258681cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); 25871da177e4SLinus Torvalds } 25881da177e4SLinus Torvalds 258967edfef7SAndi Kleen /* Send a crossed SYN-ACK during socket establishment. 259067edfef7SAndi Kleen * WARNING: This routine must only be called when we have already sent 25911da177e4SLinus Torvalds * a SYN packet that crossed the incoming SYN that caused this routine 25921da177e4SLinus Torvalds * to get called. If this assumption fails then the initial rcv_wnd 25931da177e4SLinus Torvalds * and rcv_wscale values will not be correct. 25941da177e4SLinus Torvalds */ 25951da177e4SLinus Torvalds int tcp_send_synack(struct sock *sk) 25961da177e4SLinus Torvalds { 25971da177e4SLinus Torvalds struct sk_buff *skb; 25981da177e4SLinus Torvalds 2599fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 26004de075e0SEric Dumazet if (skb == NULL || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 260191df42beSJoe Perches pr_debug("%s: wrong queue state\n", __func__); 26021da177e4SLinus Torvalds return -EFAULT; 26031da177e4SLinus Torvalds } 26044de075e0SEric Dumazet if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { 26051da177e4SLinus Torvalds if (skb_cloned(skb)) { 26061da177e4SLinus Torvalds struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC); 26071da177e4SLinus Torvalds if (nskb == NULL) 26081da177e4SLinus Torvalds return -ENOMEM; 2609fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 26101da177e4SLinus Torvalds skb_header_release(nskb); 2611fe067e8aSDavid S. Miller __tcp_add_write_queue_head(sk, nskb); 26123ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 26133ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 26143ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 26151da177e4SLinus Torvalds skb = nskb; 26161da177e4SLinus Torvalds } 26171da177e4SLinus Torvalds 26184de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; 26191da177e4SLinus Torvalds TCP_ECN_send_synack(tcp_sk(sk), skb); 26201da177e4SLinus Torvalds } 26211da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2622dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 26231da177e4SLinus Torvalds } 26241da177e4SLinus Torvalds 26254aea39c1SEric Dumazet /** 26264aea39c1SEric Dumazet * tcp_make_synack - Prepare a SYN-ACK. 26274aea39c1SEric Dumazet * sk: listener socket 26284aea39c1SEric Dumazet * dst: dst entry attached to the SYNACK 26294aea39c1SEric Dumazet * req: request_sock pointer 26304aea39c1SEric Dumazet * rvp: request_values pointer 26314aea39c1SEric Dumazet * 26324aea39c1SEric Dumazet * Allocate one skb and build a SYNACK packet. 26334aea39c1SEric Dumazet * @dst is consumed : Caller should not use it again. 26344aea39c1SEric Dumazet */ 26351da177e4SLinus Torvalds struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst, 2636e6b4d113SWilliam Allen Simpson struct request_sock *req, 2637e6b4d113SWilliam Allen Simpson struct request_values *rvp) 26381da177e4SLinus Torvalds { 2639bd0388aeSWilliam Allen Simpson struct tcp_out_options opts; 26404957faadSWilliam Allen Simpson struct tcp_extend_values *xvp = tcp_xv(rvp); 26412e6599cbSArnaldo Carvalho de Melo struct inet_request_sock *ireq = inet_rsk(req); 26421da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 264328b2774aSEric Dumazet const struct tcp_cookie_values *cvp = tp->cookie_values; 26441da177e4SLinus Torvalds struct tcphdr *th; 26451da177e4SLinus Torvalds struct sk_buff *skb; 2646cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 2647bd0388aeSWilliam Allen Simpson int tcp_header_size; 2648f5fff5dcSTom Quetchenbach int mss; 264928b2774aSEric Dumazet int s_data_desired = 0; 26501da177e4SLinus Torvalds 265128b2774aSEric Dumazet if (cvp != NULL && cvp->s_data_constant && cvp->s_data_desired) 265228b2774aSEric Dumazet s_data_desired = cvp->s_data_desired; 265390ba9b19SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER + 15 + s_data_desired, GFP_ATOMIC); 26544aea39c1SEric Dumazet if (unlikely(!skb)) { 26554aea39c1SEric Dumazet dst_release(dst); 26561da177e4SLinus Torvalds return NULL; 26574aea39c1SEric Dumazet } 26581da177e4SLinus Torvalds /* Reserve space for headers. */ 26591da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 26601da177e4SLinus Torvalds 26614aea39c1SEric Dumazet skb_dst_set(skb, dst); 26621da177e4SLinus Torvalds 26630dbaee3bSDavid S. Miller mss = dst_metric_advmss(dst); 2664f5fff5dcSTom Quetchenbach if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss) 2665f5fff5dcSTom Quetchenbach mss = tp->rx_opt.user_mss; 2666f5fff5dcSTom Quetchenbach 266733ad798cSAdam Langley if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */ 266833ad798cSAdam Langley __u8 rcv_wscale; 266933ad798cSAdam Langley /* Set this up on the first call only */ 267033ad798cSAdam Langley req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW); 2671e88c64f0SHagen Paul Pfeifer 2672e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 2673e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 2674e88c64f0SHagen Paul Pfeifer (req->window_clamp > tcp_full_space(sk) || req->window_clamp == 0)) 2675e88c64f0SHagen Paul Pfeifer req->window_clamp = tcp_full_space(sk); 2676e88c64f0SHagen Paul Pfeifer 267733ad798cSAdam Langley /* tcp_full_space because it is guaranteed to be the first packet */ 267833ad798cSAdam Langley tcp_select_initial_window(tcp_full_space(sk), 2679f5fff5dcSTom Quetchenbach mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0), 268033ad798cSAdam Langley &req->rcv_wnd, 268133ad798cSAdam Langley &req->window_clamp, 268233ad798cSAdam Langley ireq->wscale_ok, 268331d12926Slaurent chavey &rcv_wscale, 268431d12926Slaurent chavey dst_metric(dst, RTAX_INITRWND)); 268533ad798cSAdam Langley ireq->rcv_wscale = rcv_wscale; 268633ad798cSAdam Langley } 2687cfb6eeb4SYOSHIFUJI Hideaki 268833ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 26898b5f12d0SFlorian Westphal #ifdef CONFIG_SYN_COOKIES 26908b5f12d0SFlorian Westphal if (unlikely(req->cookie_ts)) 26918b5f12d0SFlorian Westphal TCP_SKB_CB(skb)->when = cookie_init_timestamp(req); 26928b5f12d0SFlorian Westphal else 26938b5f12d0SFlorian Westphal #endif 269433ad798cSAdam Langley TCP_SKB_CB(skb)->when = tcp_time_stamp; 2695f5fff5dcSTom Quetchenbach tcp_header_size = tcp_synack_options(sk, req, mss, 26964957faadSWilliam Allen Simpson skb, &opts, &md5, xvp) 26974957faadSWilliam Allen Simpson + sizeof(*th); 269833ad798cSAdam Langley 2699aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 2700aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 27011da177e4SLinus Torvalds 2702aa8223c7SArnaldo Carvalho de Melo th = tcp_hdr(skb); 27031da177e4SLinus Torvalds memset(th, 0, sizeof(struct tcphdr)); 27041da177e4SLinus Torvalds th->syn = 1; 27051da177e4SLinus Torvalds th->ack = 1; 27061da177e4SLinus Torvalds TCP_ECN_make_synack(req, th); 2707a3116ac5SKOVACS Krisztian th->source = ireq->loc_port; 27082e6599cbSArnaldo Carvalho de Melo th->dest = ireq->rmt_port; 2709e870a8efSIlpo Järvinen /* Setting of flags are superfluous here for callers (and ECE is 2710e870a8efSIlpo Järvinen * not even correctly set) 2711e870a8efSIlpo Järvinen */ 2712e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn, 2713a3433f35SChangli Gao TCPHDR_SYN | TCPHDR_ACK); 27144957faadSWilliam Allen Simpson 27154957faadSWilliam Allen Simpson if (OPTION_COOKIE_EXTENSION & opts.options) { 271628b2774aSEric Dumazet if (s_data_desired) { 271728b2774aSEric Dumazet u8 *buf = skb_put(skb, s_data_desired); 27184957faadSWilliam Allen Simpson 27194957faadSWilliam Allen Simpson /* copy data directly from the listening socket. */ 272028b2774aSEric Dumazet memcpy(buf, cvp->s_data_payload, s_data_desired); 272128b2774aSEric Dumazet TCP_SKB_CB(skb)->end_seq += s_data_desired; 27224957faadSWilliam Allen Simpson } 27234957faadSWilliam Allen Simpson 27244957faadSWilliam Allen Simpson if (opts.hash_size > 0) { 27254957faadSWilliam Allen Simpson __u32 workspace[SHA_WORKSPACE_WORDS]; 27264957faadSWilliam Allen Simpson u32 *mess = &xvp->cookie_bakery[COOKIE_DIGEST_WORDS]; 27274957faadSWilliam Allen Simpson u32 *tail = &mess[COOKIE_MESSAGE_WORDS-1]; 27284957faadSWilliam Allen Simpson 27294957faadSWilliam Allen Simpson /* Secret recipe depends on the Timestamp, (future) 27304957faadSWilliam Allen Simpson * Sequence and Acknowledgment Numbers, Initiator 27314957faadSWilliam Allen Simpson * Cookie, and others handled by IP variant caller. 27324957faadSWilliam Allen Simpson */ 27334957faadSWilliam Allen Simpson *tail-- ^= opts.tsval; 27344957faadSWilliam Allen Simpson *tail-- ^= tcp_rsk(req)->rcv_isn + 1; 27354957faadSWilliam Allen Simpson *tail-- ^= TCP_SKB_CB(skb)->seq + 1; 27364957faadSWilliam Allen Simpson 27374957faadSWilliam Allen Simpson /* recommended */ 27380eae88f3SEric Dumazet *tail-- ^= (((__force u32)th->dest << 16) | (__force u32)th->source); 2739f9a2e69eSDavid S. Miller *tail-- ^= (u32)(unsigned long)cvp; /* per sockopt */ 27404957faadSWilliam Allen Simpson 27414957faadSWilliam Allen Simpson sha_transform((__u32 *)&xvp->cookie_bakery[0], 27424957faadSWilliam Allen Simpson (char *)mess, 27434957faadSWilliam Allen Simpson &workspace[0]); 27444957faadSWilliam Allen Simpson opts.hash_location = 27454957faadSWilliam Allen Simpson (__u8 *)&xvp->cookie_bakery[0]; 27464957faadSWilliam Allen Simpson } 27474957faadSWilliam Allen Simpson } 27484957faadSWilliam Allen Simpson 27491da177e4SLinus Torvalds th->seq = htonl(TCP_SKB_CB(skb)->seq); 27502e6599cbSArnaldo Carvalho de Melo th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1); 27511da177e4SLinus Torvalds 27521da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 2753600ff0c2SIlpo Järvinen th->window = htons(min(req->rcv_wnd, 65535U)); 2754bd0388aeSWilliam Allen Simpson tcp_options_write((__be32 *)(th + 1), tp, &opts); 27551da177e4SLinus Torvalds th->doff = (tcp_header_size >> 2); 2756aa2ea058STom Herbert TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, tcp_skb_pcount(skb)); 2757cfb6eeb4SYOSHIFUJI Hideaki 2758cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 2759cfb6eeb4SYOSHIFUJI Hideaki /* Okay, we have all we need - do the md5 hash if needed */ 2760cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 2761bd0388aeSWilliam Allen Simpson tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, 276249a72dfbSAdam Langley md5, NULL, req, skb); 2763cfb6eeb4SYOSHIFUJI Hideaki } 2764cfb6eeb4SYOSHIFUJI Hideaki #endif 2765cfb6eeb4SYOSHIFUJI Hideaki 27661da177e4SLinus Torvalds return skb; 27671da177e4SLinus Torvalds } 27684bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_make_synack); 27691da177e4SLinus Torvalds 277067edfef7SAndi Kleen /* Do all connect socket setups that can be done AF independent. */ 2771370816aeSPavel Emelyanov void tcp_connect_init(struct sock *sk) 27721da177e4SLinus Torvalds { 2773cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 27741da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 27751da177e4SLinus Torvalds __u8 rcv_wscale; 27761da177e4SLinus Torvalds 27771da177e4SLinus Torvalds /* We'll fix this up when we get a response from the other end. 27781da177e4SLinus Torvalds * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 27791da177e4SLinus Torvalds */ 27801da177e4SLinus Torvalds tp->tcp_header_len = sizeof(struct tcphdr) + 2781bb5b7c11SDavid S. Miller (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0); 27821da177e4SLinus Torvalds 2783cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 2784cfb6eeb4SYOSHIFUJI Hideaki if (tp->af_specific->md5_lookup(sk, sk) != NULL) 2785cfb6eeb4SYOSHIFUJI Hideaki tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 2786cfb6eeb4SYOSHIFUJI Hideaki #endif 2787cfb6eeb4SYOSHIFUJI Hideaki 27881da177e4SLinus Torvalds /* If user gave his TCP_MAXSEG, record it to clamp */ 27891da177e4SLinus Torvalds if (tp->rx_opt.user_mss) 27901da177e4SLinus Torvalds tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 27911da177e4SLinus Torvalds tp->max_window = 0; 27925d424d5aSJohn Heffner tcp_mtup_init(sk); 27931da177e4SLinus Torvalds tcp_sync_mss(sk, dst_mtu(dst)); 27941da177e4SLinus Torvalds 27951da177e4SLinus Torvalds if (!tp->window_clamp) 27961da177e4SLinus Torvalds tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 27970dbaee3bSDavid S. Miller tp->advmss = dst_metric_advmss(dst); 2798f5fff5dcSTom Quetchenbach if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss) 2799f5fff5dcSTom Quetchenbach tp->advmss = tp->rx_opt.user_mss; 2800f5fff5dcSTom Quetchenbach 28011da177e4SLinus Torvalds tcp_initialize_rcv_mss(sk); 28021da177e4SLinus Torvalds 2803e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 2804e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 2805e88c64f0SHagen Paul Pfeifer (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) 2806e88c64f0SHagen Paul Pfeifer tp->window_clamp = tcp_full_space(sk); 2807e88c64f0SHagen Paul Pfeifer 28081da177e4SLinus Torvalds tcp_select_initial_window(tcp_full_space(sk), 28091da177e4SLinus Torvalds tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 28101da177e4SLinus Torvalds &tp->rcv_wnd, 28111da177e4SLinus Torvalds &tp->window_clamp, 2812bb5b7c11SDavid S. Miller sysctl_tcp_window_scaling, 281331d12926Slaurent chavey &rcv_wscale, 281431d12926Slaurent chavey dst_metric(dst, RTAX_INITRWND)); 28151da177e4SLinus Torvalds 28161da177e4SLinus Torvalds tp->rx_opt.rcv_wscale = rcv_wscale; 28171da177e4SLinus Torvalds tp->rcv_ssthresh = tp->rcv_wnd; 28181da177e4SLinus Torvalds 28191da177e4SLinus Torvalds sk->sk_err = 0; 28201da177e4SLinus Torvalds sock_reset_flag(sk, SOCK_DONE); 28211da177e4SLinus Torvalds tp->snd_wnd = 0; 2822ee7537b6SHantzis Fotis tcp_init_wl(tp, 0); 28231da177e4SLinus Torvalds tp->snd_una = tp->write_seq; 28241da177e4SLinus Torvalds tp->snd_sml = tp->write_seq; 282533f5f57eSIlpo Järvinen tp->snd_up = tp->write_seq; 2826370816aeSPavel Emelyanov tp->snd_nxt = tp->write_seq; 2827ee995283SPavel Emelyanov 2828ee995283SPavel Emelyanov if (likely(!tp->repair)) 28291da177e4SLinus Torvalds tp->rcv_nxt = 0; 2830ee995283SPavel Emelyanov tp->rcv_wup = tp->rcv_nxt; 2831ee995283SPavel Emelyanov tp->copied_seq = tp->rcv_nxt; 28321da177e4SLinus Torvalds 2833463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; 2834463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_retransmits = 0; 28351da177e4SLinus Torvalds tcp_clear_retrans(tp); 28361da177e4SLinus Torvalds } 28371da177e4SLinus Torvalds 2838783237e8SYuchung Cheng static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) 2839783237e8SYuchung Cheng { 2840783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 2841783237e8SYuchung Cheng struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 2842783237e8SYuchung Cheng 2843783237e8SYuchung Cheng tcb->end_seq += skb->len; 2844783237e8SYuchung Cheng skb_header_release(skb); 2845783237e8SYuchung Cheng __tcp_add_write_queue_tail(sk, skb); 2846783237e8SYuchung Cheng sk->sk_wmem_queued += skb->truesize; 2847783237e8SYuchung Cheng sk_mem_charge(sk, skb->truesize); 2848783237e8SYuchung Cheng tp->write_seq = tcb->end_seq; 2849783237e8SYuchung Cheng tp->packets_out += tcp_skb_pcount(skb); 2850783237e8SYuchung Cheng } 2851783237e8SYuchung Cheng 2852783237e8SYuchung Cheng /* Build and send a SYN with data and (cached) Fast Open cookie. However, 2853783237e8SYuchung Cheng * queue a data-only packet after the regular SYN, such that regular SYNs 2854783237e8SYuchung Cheng * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges 2855783237e8SYuchung Cheng * only the SYN sequence, the data are retransmitted in the first ACK. 2856783237e8SYuchung Cheng * If cookie is not cached or other error occurs, falls back to send a 2857783237e8SYuchung Cheng * regular SYN with Fast Open cookie request option. 2858783237e8SYuchung Cheng */ 2859783237e8SYuchung Cheng static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) 2860783237e8SYuchung Cheng { 2861783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 2862783237e8SYuchung Cheng struct tcp_fastopen_request *fo = tp->fastopen_req; 2863aab48743SYuchung Cheng int syn_loss = 0, space, i, err = 0, iovlen = fo->data->msg_iovlen; 2864783237e8SYuchung Cheng struct sk_buff *syn_data = NULL, *data; 2865aab48743SYuchung Cheng unsigned long last_syn_loss = 0; 2866783237e8SYuchung Cheng 2867*67da22d2SYuchung Cheng tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ 2868aab48743SYuchung Cheng tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie, 2869aab48743SYuchung Cheng &syn_loss, &last_syn_loss); 2870aab48743SYuchung Cheng /* Recurring FO SYN losses: revert to regular handshake temporarily */ 2871aab48743SYuchung Cheng if (syn_loss > 1 && 2872aab48743SYuchung Cheng time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) { 2873aab48743SYuchung Cheng fo->cookie.len = -1; 2874aab48743SYuchung Cheng goto fallback; 2875aab48743SYuchung Cheng } 2876aab48743SYuchung Cheng 2877*67da22d2SYuchung Cheng if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE) 2878*67da22d2SYuchung Cheng fo->cookie.len = -1; 2879*67da22d2SYuchung Cheng else if (fo->cookie.len <= 0) 2880783237e8SYuchung Cheng goto fallback; 2881783237e8SYuchung Cheng 2882783237e8SYuchung Cheng /* MSS for SYN-data is based on cached MSS and bounded by PMTU and 2883783237e8SYuchung Cheng * user-MSS. Reserve maximum option space for middleboxes that add 2884783237e8SYuchung Cheng * private TCP options. The cost is reduced data space in SYN :( 2885783237e8SYuchung Cheng */ 2886783237e8SYuchung Cheng if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp) 2887783237e8SYuchung Cheng tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 2888783237e8SYuchung Cheng space = tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) - 2889783237e8SYuchung Cheng MAX_TCP_OPTION_SPACE; 2890783237e8SYuchung Cheng 2891783237e8SYuchung Cheng syn_data = skb_copy_expand(syn, skb_headroom(syn), space, 2892783237e8SYuchung Cheng sk->sk_allocation); 2893783237e8SYuchung Cheng if (syn_data == NULL) 2894783237e8SYuchung Cheng goto fallback; 2895783237e8SYuchung Cheng 2896783237e8SYuchung Cheng for (i = 0; i < iovlen && syn_data->len < space; ++i) { 2897783237e8SYuchung Cheng struct iovec *iov = &fo->data->msg_iov[i]; 2898783237e8SYuchung Cheng unsigned char __user *from = iov->iov_base; 2899783237e8SYuchung Cheng int len = iov->iov_len; 2900783237e8SYuchung Cheng 2901783237e8SYuchung Cheng if (syn_data->len + len > space) 2902783237e8SYuchung Cheng len = space - syn_data->len; 2903783237e8SYuchung Cheng else if (i + 1 == iovlen) 2904783237e8SYuchung Cheng /* No more data pending in inet_wait_for_connect() */ 2905783237e8SYuchung Cheng fo->data = NULL; 2906783237e8SYuchung Cheng 2907783237e8SYuchung Cheng if (skb_add_data(syn_data, from, len)) 2908783237e8SYuchung Cheng goto fallback; 2909783237e8SYuchung Cheng } 2910783237e8SYuchung Cheng 2911783237e8SYuchung Cheng /* Queue a data-only packet after the regular SYN for retransmission */ 2912783237e8SYuchung Cheng data = pskb_copy(syn_data, sk->sk_allocation); 2913783237e8SYuchung Cheng if (data == NULL) 2914783237e8SYuchung Cheng goto fallback; 2915783237e8SYuchung Cheng TCP_SKB_CB(data)->seq++; 2916783237e8SYuchung Cheng TCP_SKB_CB(data)->tcp_flags &= ~TCPHDR_SYN; 2917783237e8SYuchung Cheng TCP_SKB_CB(data)->tcp_flags = (TCPHDR_ACK|TCPHDR_PSH); 2918783237e8SYuchung Cheng tcp_connect_queue_skb(sk, data); 2919783237e8SYuchung Cheng fo->copied = data->len; 2920783237e8SYuchung Cheng 2921783237e8SYuchung Cheng if (tcp_transmit_skb(sk, syn_data, 0, sk->sk_allocation) == 0) { 2922*67da22d2SYuchung Cheng tp->syn_data = (fo->copied > 0); 2923783237e8SYuchung Cheng NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE); 2924783237e8SYuchung Cheng goto done; 2925783237e8SYuchung Cheng } 2926783237e8SYuchung Cheng syn_data = NULL; 2927783237e8SYuchung Cheng 2928783237e8SYuchung Cheng fallback: 2929783237e8SYuchung Cheng /* Send a regular SYN with Fast Open cookie request option */ 2930783237e8SYuchung Cheng if (fo->cookie.len > 0) 2931783237e8SYuchung Cheng fo->cookie.len = 0; 2932783237e8SYuchung Cheng err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); 2933783237e8SYuchung Cheng if (err) 2934783237e8SYuchung Cheng tp->syn_fastopen = 0; 2935783237e8SYuchung Cheng kfree_skb(syn_data); 2936783237e8SYuchung Cheng done: 2937783237e8SYuchung Cheng fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ 2938783237e8SYuchung Cheng return err; 2939783237e8SYuchung Cheng } 2940783237e8SYuchung Cheng 294167edfef7SAndi Kleen /* Build a SYN and send it off. */ 29421da177e4SLinus Torvalds int tcp_connect(struct sock *sk) 29431da177e4SLinus Torvalds { 29441da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 29451da177e4SLinus Torvalds struct sk_buff *buff; 2946ee586811SEric Paris int err; 29471da177e4SLinus Torvalds 29481da177e4SLinus Torvalds tcp_connect_init(sk); 29491da177e4SLinus Torvalds 2950d179cd12SDavid S. Miller buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation); 29511da177e4SLinus Torvalds if (unlikely(buff == NULL)) 29521da177e4SLinus Torvalds return -ENOBUFS; 29531da177e4SLinus Torvalds 29541da177e4SLinus Torvalds /* Reserve space for headers. */ 29551da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 29561da177e4SLinus Torvalds 2957a3433f35SChangli Gao tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); 2958783237e8SYuchung Cheng tp->retrans_stamp = TCP_SKB_CB(buff)->when = tcp_time_stamp; 2959783237e8SYuchung Cheng tcp_connect_queue_skb(sk, buff); 2960e870a8efSIlpo Järvinen TCP_ECN_send_syn(sk, buff); 29611da177e4SLinus Torvalds 2962783237e8SYuchung Cheng /* Send off SYN; include data in Fast Open. */ 2963783237e8SYuchung Cheng err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : 2964783237e8SYuchung Cheng tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); 2965ee586811SEric Paris if (err == -ECONNREFUSED) 2966ee586811SEric Paris return err; 2967bd37a088SWei Yongjun 2968bd37a088SWei Yongjun /* We change tp->snd_nxt after the tcp_transmit_skb() call 2969bd37a088SWei Yongjun * in order to make this packet get counted in tcpOutSegs. 2970bd37a088SWei Yongjun */ 2971bd37a088SWei Yongjun tp->snd_nxt = tp->write_seq; 2972bd37a088SWei Yongjun tp->pushed_seq = tp->write_seq; 297381cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); 29741da177e4SLinus Torvalds 29751da177e4SLinus Torvalds /* Timer for repeating the SYN until an answer. */ 29763f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 29773f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 29781da177e4SLinus Torvalds return 0; 29791da177e4SLinus Torvalds } 29804bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_connect); 29811da177e4SLinus Torvalds 29821da177e4SLinus Torvalds /* Send out a delayed ack, the caller does the policy checking 29831da177e4SLinus Torvalds * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 29841da177e4SLinus Torvalds * for details. 29851da177e4SLinus Torvalds */ 29861da177e4SLinus Torvalds void tcp_send_delayed_ack(struct sock *sk) 29871da177e4SLinus Torvalds { 2988463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 2989463c84b9SArnaldo Carvalho de Melo int ato = icsk->icsk_ack.ato; 29901da177e4SLinus Torvalds unsigned long timeout; 29911da177e4SLinus Torvalds 29921da177e4SLinus Torvalds if (ato > TCP_DELACK_MIN) { 2993463c84b9SArnaldo Carvalho de Melo const struct tcp_sock *tp = tcp_sk(sk); 29941da177e4SLinus Torvalds int max_ato = HZ / 2; 29951da177e4SLinus Torvalds 2996056834d9SIlpo Järvinen if (icsk->icsk_ack.pingpong || 2997056834d9SIlpo Järvinen (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 29981da177e4SLinus Torvalds max_ato = TCP_DELACK_MAX; 29991da177e4SLinus Torvalds 30001da177e4SLinus Torvalds /* Slow path, intersegment interval is "high". */ 30011da177e4SLinus Torvalds 30021da177e4SLinus Torvalds /* If some rtt estimate is known, use it to bound delayed ack. 3003463c84b9SArnaldo Carvalho de Melo * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 30041da177e4SLinus Torvalds * directly. 30051da177e4SLinus Torvalds */ 30061da177e4SLinus Torvalds if (tp->srtt) { 30071da177e4SLinus Torvalds int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN); 30081da177e4SLinus Torvalds 30091da177e4SLinus Torvalds if (rtt < max_ato) 30101da177e4SLinus Torvalds max_ato = rtt; 30111da177e4SLinus Torvalds } 30121da177e4SLinus Torvalds 30131da177e4SLinus Torvalds ato = min(ato, max_ato); 30141da177e4SLinus Torvalds } 30151da177e4SLinus Torvalds 30161da177e4SLinus Torvalds /* Stay within the limit we were given */ 30171da177e4SLinus Torvalds timeout = jiffies + ato; 30181da177e4SLinus Torvalds 30191da177e4SLinus Torvalds /* Use new timeout only if there wasn't a older one earlier. */ 3020463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 30211da177e4SLinus Torvalds /* If delack timer was blocked or is about to expire, 30221da177e4SLinus Torvalds * send ACK now. 30231da177e4SLinus Torvalds */ 3024463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.blocked || 3025463c84b9SArnaldo Carvalho de Melo time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 30261da177e4SLinus Torvalds tcp_send_ack(sk); 30271da177e4SLinus Torvalds return; 30281da177e4SLinus Torvalds } 30291da177e4SLinus Torvalds 3030463c84b9SArnaldo Carvalho de Melo if (!time_before(timeout, icsk->icsk_ack.timeout)) 3031463c84b9SArnaldo Carvalho de Melo timeout = icsk->icsk_ack.timeout; 30321da177e4SLinus Torvalds } 3033463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 3034463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.timeout = timeout; 3035463c84b9SArnaldo Carvalho de Melo sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 30361da177e4SLinus Torvalds } 30371da177e4SLinus Torvalds 30381da177e4SLinus Torvalds /* This routine sends an ack and also updates the window. */ 30391da177e4SLinus Torvalds void tcp_send_ack(struct sock *sk) 30401da177e4SLinus Torvalds { 30411da177e4SLinus Torvalds struct sk_buff *buff; 30421da177e4SLinus Torvalds 3043058dc334SIlpo Järvinen /* If we have been reset, we may not send again. */ 3044058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3045058dc334SIlpo Järvinen return; 3046058dc334SIlpo Järvinen 30471da177e4SLinus Torvalds /* We are not putting this on the write queue, so 30481da177e4SLinus Torvalds * tcp_transmit_skb() will set the ownership to this 30491da177e4SLinus Torvalds * sock. 30501da177e4SLinus Torvalds */ 30511da177e4SLinus Torvalds buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 30521da177e4SLinus Torvalds if (buff == NULL) { 3053463c84b9SArnaldo Carvalho de Melo inet_csk_schedule_ack(sk); 3054463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; 30553f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 30563f421baaSArnaldo Carvalho de Melo TCP_DELACK_MAX, TCP_RTO_MAX); 30571da177e4SLinus Torvalds return; 30581da177e4SLinus Torvalds } 30591da177e4SLinus Torvalds 30601da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 30611da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 3062a3433f35SChangli Gao tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); 30631da177e4SLinus Torvalds 30641da177e4SLinus Torvalds /* Send it off, this clears delayed acks for us. */ 30651da177e4SLinus Torvalds TCP_SKB_CB(buff)->when = tcp_time_stamp; 3066dfb4b9dcSDavid S. Miller tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC); 30671da177e4SLinus Torvalds } 30681da177e4SLinus Torvalds 30691da177e4SLinus Torvalds /* This routine sends a packet with an out of date sequence 30701da177e4SLinus Torvalds * number. It assumes the other end will try to ack it. 30711da177e4SLinus Torvalds * 30721da177e4SLinus Torvalds * Question: what should we make while urgent mode? 30731da177e4SLinus Torvalds * 4.4BSD forces sending single byte of data. We cannot send 30741da177e4SLinus Torvalds * out of window data, because we have SND.NXT==SND.MAX... 30751da177e4SLinus Torvalds * 30761da177e4SLinus Torvalds * Current solution: to send TWO zero-length segments in urgent mode: 30771da177e4SLinus Torvalds * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 30781da177e4SLinus Torvalds * out-of-date with SND.UNA-1 to probe window. 30791da177e4SLinus Torvalds */ 30801da177e4SLinus Torvalds static int tcp_xmit_probe_skb(struct sock *sk, int urgent) 30811da177e4SLinus Torvalds { 30821da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 30831da177e4SLinus Torvalds struct sk_buff *skb; 30841da177e4SLinus Torvalds 30851da177e4SLinus Torvalds /* We don't queue it, tcp_transmit_skb() sets ownership. */ 30861da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 30871da177e4SLinus Torvalds if (skb == NULL) 30881da177e4SLinus Torvalds return -1; 30891da177e4SLinus Torvalds 30901da177e4SLinus Torvalds /* Reserve space for headers and set control bits. */ 30911da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 30921da177e4SLinus Torvalds /* Use a previous sequence. This should cause the other 30931da177e4SLinus Torvalds * end to send an ack. Don't queue or clone SKB, just 30941da177e4SLinus Torvalds * send it. 30951da177e4SLinus Torvalds */ 3096a3433f35SChangli Gao tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); 30971da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 3098dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC); 30991da177e4SLinus Torvalds } 31001da177e4SLinus Torvalds 3101ee995283SPavel Emelyanov void tcp_send_window_probe(struct sock *sk) 3102ee995283SPavel Emelyanov { 3103ee995283SPavel Emelyanov if (sk->sk_state == TCP_ESTABLISHED) { 3104ee995283SPavel Emelyanov tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; 3105c0e88ff0SPavel Emelyanov tcp_sk(sk)->snd_nxt = tcp_sk(sk)->write_seq; 3106ee995283SPavel Emelyanov tcp_xmit_probe_skb(sk, 0); 3107ee995283SPavel Emelyanov } 3108ee995283SPavel Emelyanov } 3109ee995283SPavel Emelyanov 311067edfef7SAndi Kleen /* Initiate keepalive or window probe from timer. */ 31111da177e4SLinus Torvalds int tcp_write_wakeup(struct sock *sk) 31121da177e4SLinus Torvalds { 31131da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 31141da177e4SLinus Torvalds struct sk_buff *skb; 31151da177e4SLinus Torvalds 3116058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3117058dc334SIlpo Järvinen return -1; 3118058dc334SIlpo Järvinen 3119fe067e8aSDavid S. Miller if ((skb = tcp_send_head(sk)) != NULL && 312090840defSIlpo Järvinen before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { 31211da177e4SLinus Torvalds int err; 31220c54b85fSIlpo Järvinen unsigned int mss = tcp_current_mss(sk); 312390840defSIlpo Järvinen unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 31241da177e4SLinus Torvalds 31251da177e4SLinus Torvalds if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 31261da177e4SLinus Torvalds tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 31271da177e4SLinus Torvalds 31281da177e4SLinus Torvalds /* We are probing the opening of a window 31291da177e4SLinus Torvalds * but the window size is != 0 31301da177e4SLinus Torvalds * must have been a result SWS avoidance ( sender ) 31311da177e4SLinus Torvalds */ 31321da177e4SLinus Torvalds if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 31331da177e4SLinus Torvalds skb->len > mss) { 31341da177e4SLinus Torvalds seg_size = min(seg_size, mss); 31354de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 3136846998aeSDavid S. Miller if (tcp_fragment(sk, skb, seg_size, mss)) 31371da177e4SLinus Torvalds return -1; 31381da177e4SLinus Torvalds } else if (!tcp_skb_pcount(skb)) 3139846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss); 31401da177e4SLinus Torvalds 31414de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 31421da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 3143dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 314466f5fe62SIlpo Järvinen if (!err) 314566f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 31461da177e4SLinus Torvalds return err; 31471da177e4SLinus Torvalds } else { 314833f5f57eSIlpo Järvinen if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) 31494828e7f4SIlpo Järvinen tcp_xmit_probe_skb(sk, 1); 31501da177e4SLinus Torvalds return tcp_xmit_probe_skb(sk, 0); 31511da177e4SLinus Torvalds } 31521da177e4SLinus Torvalds } 31531da177e4SLinus Torvalds 31541da177e4SLinus Torvalds /* A window probe timeout has occurred. If window is not closed send 31551da177e4SLinus Torvalds * a partial packet else a zero probe. 31561da177e4SLinus Torvalds */ 31571da177e4SLinus Torvalds void tcp_send_probe0(struct sock *sk) 31581da177e4SLinus Torvalds { 3159463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 31601da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 31611da177e4SLinus Torvalds int err; 31621da177e4SLinus Torvalds 31631da177e4SLinus Torvalds err = tcp_write_wakeup(sk); 31641da177e4SLinus Torvalds 3165fe067e8aSDavid S. Miller if (tp->packets_out || !tcp_send_head(sk)) { 31661da177e4SLinus Torvalds /* Cancel probe timer, if it is not required. */ 31676687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 0; 3168463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff = 0; 31691da177e4SLinus Torvalds return; 31701da177e4SLinus Torvalds } 31711da177e4SLinus Torvalds 31721da177e4SLinus Torvalds if (err <= 0) { 3173463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_backoff < sysctl_tcp_retries2) 3174463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff++; 31756687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out++; 3176463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 31773f421baaSArnaldo Carvalho de Melo min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), 31783f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 31791da177e4SLinus Torvalds } else { 31801da177e4SLinus Torvalds /* If packet was not sent due to local congestion, 31816687e988SArnaldo Carvalho de Melo * do not backoff and do not remember icsk_probes_out. 31821da177e4SLinus Torvalds * Let local senders to fight for local resources. 31831da177e4SLinus Torvalds * 31841da177e4SLinus Torvalds * Use accumulated backoff yet. 31851da177e4SLinus Torvalds */ 31866687e988SArnaldo Carvalho de Melo if (!icsk->icsk_probes_out) 31876687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 1; 3188463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 3189463c84b9SArnaldo Carvalho de Melo min(icsk->icsk_rto << icsk->icsk_backoff, 31903f421baaSArnaldo Carvalho de Melo TCP_RESOURCE_PROBE_INTERVAL), 31913f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 31921da177e4SLinus Torvalds } 31931da177e4SLinus Torvalds } 3194