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, 7058336886fSJerry Chu struct tcp_extend_values *xvp, 7068336886fSJerry Chu struct tcp_fastopen_cookie *foc) 7074957faadSWilliam Allen Simpson { 70833ad798cSAdam Langley struct inet_request_sock *ireq = inet_rsk(req); 70995c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 7104957faadSWilliam Allen Simpson u8 cookie_plus = (xvp != NULL && !xvp->cookie_out_never) ? 7114957faadSWilliam Allen Simpson xvp->cookie_plus : 7124957faadSWilliam Allen Simpson 0; 71333ad798cSAdam Langley 71433ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 71533ad798cSAdam Langley *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req); 71633ad798cSAdam Langley if (*md5) { 71733ad798cSAdam Langley opts->options |= OPTION_MD5; 7184957faadSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 7194957faadSWilliam Allen Simpson 7204957faadSWilliam Allen Simpson /* We can't fit any SACK blocks in a packet with MD5 + TS 7214957faadSWilliam Allen Simpson * options. There was discussion about disabling SACK 7224957faadSWilliam Allen Simpson * rather than TS in order to fit in better with old, 7234957faadSWilliam Allen Simpson * buggy kernels, but that was deemed to be unnecessary. 7244957faadSWilliam Allen Simpson */ 725de213e5eSEric Dumazet ireq->tstamp_ok &= !ireq->sack_ok; 72633ad798cSAdam Langley } 72733ad798cSAdam Langley #else 72833ad798cSAdam Langley *md5 = NULL; 72933ad798cSAdam Langley #endif 73033ad798cSAdam Langley 7314957faadSWilliam Allen Simpson /* We always send an MSS option. */ 73233ad798cSAdam Langley opts->mss = mss; 7334957faadSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 73433ad798cSAdam Langley 73533ad798cSAdam Langley if (likely(ireq->wscale_ok)) { 73633ad798cSAdam Langley opts->ws = ireq->rcv_wscale; 73789e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 7384957faadSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 73933ad798cSAdam Langley } 740de213e5eSEric Dumazet if (likely(ireq->tstamp_ok)) { 74133ad798cSAdam Langley opts->options |= OPTION_TS; 74233ad798cSAdam Langley opts->tsval = TCP_SKB_CB(skb)->when; 74333ad798cSAdam Langley opts->tsecr = req->ts_recent; 7444957faadSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 74533ad798cSAdam Langley } 74633ad798cSAdam Langley if (likely(ireq->sack_ok)) { 74733ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 748de213e5eSEric Dumazet if (unlikely(!ireq->tstamp_ok)) 7494957faadSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 75033ad798cSAdam Langley } 7518336886fSJerry Chu if (foc != NULL) { 7528336886fSJerry Chu u32 need = TCPOLEN_EXP_FASTOPEN_BASE + foc->len; 7538336886fSJerry Chu need = (need + 3) & ~3U; /* Align to 32 bits */ 7548336886fSJerry Chu if (remaining >= need) { 7558336886fSJerry Chu opts->options |= OPTION_FAST_OPEN_COOKIE; 7568336886fSJerry Chu opts->fastopen_cookie = foc; 7578336886fSJerry Chu remaining -= need; 7588336886fSJerry Chu } 7598336886fSJerry Chu } 7604957faadSWilliam Allen Simpson /* Similar rationale to tcp_syn_options() applies here, too. 7614957faadSWilliam Allen Simpson * If the <SYN> options fit, the same options should fit now! 7624957faadSWilliam Allen Simpson */ 7634957faadSWilliam Allen Simpson if (*md5 == NULL && 764de213e5eSEric Dumazet ireq->tstamp_ok && 7654957faadSWilliam Allen Simpson cookie_plus > TCPOLEN_COOKIE_BASE) { 7664957faadSWilliam Allen Simpson int need = cookie_plus; /* has TCPOLEN_COOKIE_BASE */ 7674957faadSWilliam Allen Simpson 7684957faadSWilliam Allen Simpson if (0x2 & need) { 7694957faadSWilliam Allen Simpson /* 32-bit multiple */ 7704957faadSWilliam Allen Simpson need += 2; /* NOPs */ 7714957faadSWilliam Allen Simpson } 7724957faadSWilliam Allen Simpson if (need <= remaining) { 7734957faadSWilliam Allen Simpson opts->options |= OPTION_COOKIE_EXTENSION; 7744957faadSWilliam Allen Simpson opts->hash_size = cookie_plus - TCPOLEN_COOKIE_BASE; 7754957faadSWilliam Allen Simpson remaining -= need; 7764957faadSWilliam Allen Simpson } else { 7774957faadSWilliam Allen Simpson /* There's no error return, so flag it. */ 7784957faadSWilliam Allen Simpson xvp->cookie_out_never = 1; /* true */ 7794957faadSWilliam Allen Simpson opts->hash_size = 0; 7804957faadSWilliam Allen Simpson } 7814957faadSWilliam Allen Simpson } 7824957faadSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 78333ad798cSAdam Langley } 78433ad798cSAdam Langley 78567edfef7SAndi Kleen /* Compute TCP options for ESTABLISHED sockets. This is not the 78667edfef7SAndi Kleen * final wire format yet. 78767edfef7SAndi Kleen */ 78895c96174SEric Dumazet static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb, 78933ad798cSAdam Langley struct tcp_out_options *opts, 790cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 791cf533ea5SEric Dumazet { 79233ad798cSAdam Langley struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL; 79333ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 79495c96174SEric Dumazet unsigned int size = 0; 795cabeccbdSIlpo Järvinen unsigned int eff_sacks; 79633ad798cSAdam Langley 79733ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 79833ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 79933ad798cSAdam Langley if (unlikely(*md5)) { 80033ad798cSAdam Langley opts->options |= OPTION_MD5; 80133ad798cSAdam Langley size += TCPOLEN_MD5SIG_ALIGNED; 80233ad798cSAdam Langley } 80333ad798cSAdam Langley #else 80433ad798cSAdam Langley *md5 = NULL; 80533ad798cSAdam Langley #endif 80633ad798cSAdam Langley 80733ad798cSAdam Langley if (likely(tp->rx_opt.tstamp_ok)) { 80833ad798cSAdam Langley opts->options |= OPTION_TS; 80933ad798cSAdam Langley opts->tsval = tcb ? tcb->when : 0; 81033ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 81133ad798cSAdam Langley size += TCPOLEN_TSTAMP_ALIGNED; 81233ad798cSAdam Langley } 81333ad798cSAdam Langley 814cabeccbdSIlpo Järvinen eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; 815cabeccbdSIlpo Järvinen if (unlikely(eff_sacks)) { 81695c96174SEric Dumazet const unsigned int remaining = MAX_TCP_OPTION_SPACE - size; 81733ad798cSAdam Langley opts->num_sack_blocks = 81895c96174SEric Dumazet min_t(unsigned int, eff_sacks, 81933ad798cSAdam Langley (remaining - TCPOLEN_SACK_BASE_ALIGNED) / 82033ad798cSAdam Langley TCPOLEN_SACK_PERBLOCK); 82133ad798cSAdam Langley size += TCPOLEN_SACK_BASE_ALIGNED + 82233ad798cSAdam Langley opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; 82333ad798cSAdam Langley } 82433ad798cSAdam Langley 82533ad798cSAdam Langley return size; 82640efc6faSStephen Hemminger } 8271da177e4SLinus Torvalds 82846d3ceabSEric Dumazet 82946d3ceabSEric Dumazet /* TCP SMALL QUEUES (TSQ) 83046d3ceabSEric Dumazet * 83146d3ceabSEric Dumazet * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev) 83246d3ceabSEric Dumazet * to reduce RTT and bufferbloat. 83346d3ceabSEric Dumazet * We do this using a special skb destructor (tcp_wfree). 83446d3ceabSEric Dumazet * 83546d3ceabSEric Dumazet * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb 83646d3ceabSEric Dumazet * needs to be reallocated in a driver. 83746d3ceabSEric Dumazet * The invariant being skb->truesize substracted from sk->sk_wmem_alloc 83846d3ceabSEric Dumazet * 83946d3ceabSEric Dumazet * Since transmit from skb destructor is forbidden, we use a tasklet 84046d3ceabSEric Dumazet * to process all sockets that eventually need to send more skbs. 84146d3ceabSEric Dumazet * We use one tasklet per cpu, with its own queue of sockets. 84246d3ceabSEric Dumazet */ 84346d3ceabSEric Dumazet struct tsq_tasklet { 84446d3ceabSEric Dumazet struct tasklet_struct tasklet; 84546d3ceabSEric Dumazet struct list_head head; /* queue of tcp sockets */ 84646d3ceabSEric Dumazet }; 84746d3ceabSEric Dumazet static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet); 84846d3ceabSEric Dumazet 8496f458dfbSEric Dumazet static void tcp_tsq_handler(struct sock *sk) 8506f458dfbSEric Dumazet { 8516f458dfbSEric Dumazet if ((1 << sk->sk_state) & 8526f458dfbSEric Dumazet (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING | 8536f458dfbSEric Dumazet TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) 8546f458dfbSEric Dumazet tcp_write_xmit(sk, tcp_current_mss(sk), 0, 0, GFP_ATOMIC); 8556f458dfbSEric Dumazet } 85646d3ceabSEric Dumazet /* 85746d3ceabSEric Dumazet * One tasklest per cpu tries to send more skbs. 85846d3ceabSEric Dumazet * We run in tasklet context but need to disable irqs when 85946d3ceabSEric Dumazet * transfering tsq->head because tcp_wfree() might 86046d3ceabSEric Dumazet * interrupt us (non NAPI drivers) 86146d3ceabSEric Dumazet */ 86246d3ceabSEric Dumazet static void tcp_tasklet_func(unsigned long data) 86346d3ceabSEric Dumazet { 86446d3ceabSEric Dumazet struct tsq_tasklet *tsq = (struct tsq_tasklet *)data; 86546d3ceabSEric Dumazet LIST_HEAD(list); 86646d3ceabSEric Dumazet unsigned long flags; 86746d3ceabSEric Dumazet struct list_head *q, *n; 86846d3ceabSEric Dumazet struct tcp_sock *tp; 86946d3ceabSEric Dumazet struct sock *sk; 87046d3ceabSEric Dumazet 87146d3ceabSEric Dumazet local_irq_save(flags); 87246d3ceabSEric Dumazet list_splice_init(&tsq->head, &list); 87346d3ceabSEric Dumazet local_irq_restore(flags); 87446d3ceabSEric Dumazet 87546d3ceabSEric Dumazet list_for_each_safe(q, n, &list) { 87646d3ceabSEric Dumazet tp = list_entry(q, struct tcp_sock, tsq_node); 87746d3ceabSEric Dumazet list_del(&tp->tsq_node); 87846d3ceabSEric Dumazet 87946d3ceabSEric Dumazet sk = (struct sock *)tp; 88046d3ceabSEric Dumazet bh_lock_sock(sk); 88146d3ceabSEric Dumazet 88246d3ceabSEric Dumazet if (!sock_owned_by_user(sk)) { 8836f458dfbSEric Dumazet tcp_tsq_handler(sk); 88446d3ceabSEric Dumazet } else { 88546d3ceabSEric Dumazet /* defer the work to tcp_release_cb() */ 8866f458dfbSEric Dumazet set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags); 88746d3ceabSEric Dumazet } 88846d3ceabSEric Dumazet bh_unlock_sock(sk); 88946d3ceabSEric Dumazet 89046d3ceabSEric Dumazet clear_bit(TSQ_QUEUED, &tp->tsq_flags); 89146d3ceabSEric Dumazet sk_free(sk); 89246d3ceabSEric Dumazet } 89346d3ceabSEric Dumazet } 89446d3ceabSEric Dumazet 8956f458dfbSEric Dumazet #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \ 8966f458dfbSEric Dumazet (1UL << TCP_WRITE_TIMER_DEFERRED) | \ 897563d34d0SEric Dumazet (1UL << TCP_DELACK_TIMER_DEFERRED) | \ 898563d34d0SEric Dumazet (1UL << TCP_MTU_REDUCED_DEFERRED)) 89946d3ceabSEric Dumazet /** 90046d3ceabSEric Dumazet * tcp_release_cb - tcp release_sock() callback 90146d3ceabSEric Dumazet * @sk: socket 90246d3ceabSEric Dumazet * 90346d3ceabSEric Dumazet * called from release_sock() to perform protocol dependent 90446d3ceabSEric Dumazet * actions before socket release. 90546d3ceabSEric Dumazet */ 90646d3ceabSEric Dumazet void tcp_release_cb(struct sock *sk) 90746d3ceabSEric Dumazet { 90846d3ceabSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 9096f458dfbSEric Dumazet unsigned long flags, nflags; 91046d3ceabSEric Dumazet 9116f458dfbSEric Dumazet /* perform an atomic operation only if at least one flag is set */ 9126f458dfbSEric Dumazet do { 9136f458dfbSEric Dumazet flags = tp->tsq_flags; 9146f458dfbSEric Dumazet if (!(flags & TCP_DEFERRED_ALL)) 9156f458dfbSEric Dumazet return; 9166f458dfbSEric Dumazet nflags = flags & ~TCP_DEFERRED_ALL; 9176f458dfbSEric Dumazet } while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags); 9186f458dfbSEric Dumazet 9196f458dfbSEric Dumazet if (flags & (1UL << TCP_TSQ_DEFERRED)) 9206f458dfbSEric Dumazet tcp_tsq_handler(sk); 9216f458dfbSEric Dumazet 922144d56e9SEric Dumazet if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) { 9236f458dfbSEric Dumazet tcp_write_timer_handler(sk); 924144d56e9SEric Dumazet __sock_put(sk); 925144d56e9SEric Dumazet } 926144d56e9SEric Dumazet if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) { 9276f458dfbSEric Dumazet tcp_delack_timer_handler(sk); 928144d56e9SEric Dumazet __sock_put(sk); 929144d56e9SEric Dumazet } 930144d56e9SEric Dumazet if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) { 931563d34d0SEric Dumazet sk->sk_prot->mtu_reduced(sk); 932144d56e9SEric Dumazet __sock_put(sk); 933144d56e9SEric Dumazet } 93446d3ceabSEric Dumazet } 93546d3ceabSEric Dumazet EXPORT_SYMBOL(tcp_release_cb); 93646d3ceabSEric Dumazet 93746d3ceabSEric Dumazet void __init tcp_tasklet_init(void) 93846d3ceabSEric Dumazet { 93946d3ceabSEric Dumazet int i; 94046d3ceabSEric Dumazet 94146d3ceabSEric Dumazet for_each_possible_cpu(i) { 94246d3ceabSEric Dumazet struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i); 94346d3ceabSEric Dumazet 94446d3ceabSEric Dumazet INIT_LIST_HEAD(&tsq->head); 94546d3ceabSEric Dumazet tasklet_init(&tsq->tasklet, 94646d3ceabSEric Dumazet tcp_tasklet_func, 94746d3ceabSEric Dumazet (unsigned long)tsq); 94846d3ceabSEric Dumazet } 94946d3ceabSEric Dumazet } 95046d3ceabSEric Dumazet 95146d3ceabSEric Dumazet /* 95246d3ceabSEric Dumazet * Write buffer destructor automatically called from kfree_skb. 95346d3ceabSEric Dumazet * We cant xmit new skbs from this context, as we might already 95446d3ceabSEric Dumazet * hold qdisc lock. 95546d3ceabSEric Dumazet */ 9568e7dfbc8SSilviu-Mihai Popescu static void tcp_wfree(struct sk_buff *skb) 95746d3ceabSEric Dumazet { 95846d3ceabSEric Dumazet struct sock *sk = skb->sk; 95946d3ceabSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 96046d3ceabSEric Dumazet 96146d3ceabSEric Dumazet if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) && 96246d3ceabSEric Dumazet !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) { 96346d3ceabSEric Dumazet unsigned long flags; 96446d3ceabSEric Dumazet struct tsq_tasklet *tsq; 96546d3ceabSEric Dumazet 96646d3ceabSEric Dumazet /* Keep a ref on socket. 96746d3ceabSEric Dumazet * This last ref will be released in tcp_tasklet_func() 96846d3ceabSEric Dumazet */ 96946d3ceabSEric Dumazet atomic_sub(skb->truesize - 1, &sk->sk_wmem_alloc); 97046d3ceabSEric Dumazet 97146d3ceabSEric Dumazet /* queue this socket to tasklet queue */ 97246d3ceabSEric Dumazet local_irq_save(flags); 97346d3ceabSEric Dumazet tsq = &__get_cpu_var(tsq_tasklet); 97446d3ceabSEric Dumazet list_add(&tp->tsq_node, &tsq->head); 97546d3ceabSEric Dumazet tasklet_schedule(&tsq->tasklet); 97646d3ceabSEric Dumazet local_irq_restore(flags); 97746d3ceabSEric Dumazet } else { 97846d3ceabSEric Dumazet sock_wfree(skb); 97946d3ceabSEric Dumazet } 98046d3ceabSEric Dumazet } 98146d3ceabSEric Dumazet 9821da177e4SLinus Torvalds /* This routine actually transmits TCP packets queued in by 9831da177e4SLinus Torvalds * tcp_do_sendmsg(). This is used by both the initial 9841da177e4SLinus Torvalds * transmission and possible later retransmissions. 9851da177e4SLinus Torvalds * All SKB's seen here are completely headerless. It is our 9861da177e4SLinus Torvalds * job to build the TCP header, and pass the packet down to 9871da177e4SLinus Torvalds * IP so it can do the same plus pass the packet off to the 9881da177e4SLinus Torvalds * device. 9891da177e4SLinus Torvalds * 9901da177e4SLinus Torvalds * We are working here with either a clone of the original 9911da177e4SLinus Torvalds * SKB, or a fresh unique copy made by the retransmit engine. 9921da177e4SLinus Torvalds */ 993056834d9SIlpo Järvinen static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, 994056834d9SIlpo Järvinen gfp_t gfp_mask) 9951da177e4SLinus Torvalds { 9966687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 997dfb4b9dcSDavid S. Miller struct inet_sock *inet; 998dfb4b9dcSDavid S. Miller struct tcp_sock *tp; 999dfb4b9dcSDavid S. Miller struct tcp_skb_cb *tcb; 100033ad798cSAdam Langley struct tcp_out_options opts; 100195c96174SEric Dumazet unsigned int tcp_options_size, tcp_header_size; 1002cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 10031da177e4SLinus Torvalds struct tcphdr *th; 10041da177e4SLinus Torvalds int err; 10051da177e4SLinus Torvalds 1006dfb4b9dcSDavid S. Miller BUG_ON(!skb || !tcp_skb_pcount(skb)); 1007dfb4b9dcSDavid S. Miller 1008dfb4b9dcSDavid S. Miller /* If congestion control is doing timestamping, we must 1009dfb4b9dcSDavid S. Miller * take such a timestamp before we potentially clone/copy. 1010dfb4b9dcSDavid S. Miller */ 1011164891aaSStephen Hemminger if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP) 1012dfb4b9dcSDavid S. Miller __net_timestamp(skb); 1013dfb4b9dcSDavid S. Miller 1014dfb4b9dcSDavid S. Miller if (likely(clone_it)) { 1015dfb4b9dcSDavid S. Miller if (unlikely(skb_cloned(skb))) 1016dfb4b9dcSDavid S. Miller skb = pskb_copy(skb, gfp_mask); 1017dfb4b9dcSDavid S. Miller else 1018dfb4b9dcSDavid S. Miller skb = skb_clone(skb, gfp_mask); 1019dfb4b9dcSDavid S. Miller if (unlikely(!skb)) 1020dfb4b9dcSDavid S. Miller return -ENOBUFS; 1021dfb4b9dcSDavid S. Miller } 1022dfb4b9dcSDavid S. Miller 1023dfb4b9dcSDavid S. Miller inet = inet_sk(sk); 1024dfb4b9dcSDavid S. Miller tp = tcp_sk(sk); 1025dfb4b9dcSDavid S. Miller tcb = TCP_SKB_CB(skb); 102633ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 10271da177e4SLinus Torvalds 10284de075e0SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) 102933ad798cSAdam Langley tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); 103033ad798cSAdam Langley else 103133ad798cSAdam Langley tcp_options_size = tcp_established_options(sk, skb, &opts, 103233ad798cSAdam Langley &md5); 103333ad798cSAdam Langley tcp_header_size = tcp_options_size + sizeof(struct tcphdr); 10341da177e4SLinus Torvalds 10353853b584STom Herbert if (tcp_packets_in_flight(tp) == 0) { 10366687e988SArnaldo Carvalho de Melo tcp_ca_event(sk, CA_EVENT_TX_START); 10373853b584STom Herbert skb->ooo_okay = 1; 10383853b584STom Herbert } else 10393853b584STom Herbert skb->ooo_okay = 0; 10401da177e4SLinus Torvalds 1041aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 1042aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 104346d3ceabSEric Dumazet 104446d3ceabSEric Dumazet skb_orphan(skb); 104546d3ceabSEric Dumazet skb->sk = sk; 104646d3ceabSEric Dumazet skb->destructor = (sysctl_tcp_limit_output_bytes > 0) ? 104746d3ceabSEric Dumazet tcp_wfree : sock_wfree; 104846d3ceabSEric Dumazet atomic_add(skb->truesize, &sk->sk_wmem_alloc); 10491da177e4SLinus Torvalds 10501da177e4SLinus Torvalds /* Build TCP header and checksum it. */ 1051aa8223c7SArnaldo Carvalho de Melo th = tcp_hdr(skb); 1052c720c7e8SEric Dumazet th->source = inet->inet_sport; 1053c720c7e8SEric Dumazet th->dest = inet->inet_dport; 10541da177e4SLinus Torvalds th->seq = htonl(tcb->seq); 10551da177e4SLinus Torvalds th->ack_seq = htonl(tp->rcv_nxt); 1056df7a3b07SAl Viro *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | 10574de075e0SEric Dumazet tcb->tcp_flags); 1058dfb4b9dcSDavid S. Miller 10594de075e0SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) { 10601da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments 10611da177e4SLinus Torvalds * is never scaled. 10621da177e4SLinus Torvalds */ 1063600ff0c2SIlpo Järvinen th->window = htons(min(tp->rcv_wnd, 65535U)); 10641da177e4SLinus Torvalds } else { 10651da177e4SLinus Torvalds th->window = htons(tcp_select_window(sk)); 10661da177e4SLinus Torvalds } 10671da177e4SLinus Torvalds th->check = 0; 10681da177e4SLinus Torvalds th->urg_ptr = 0; 10691da177e4SLinus Torvalds 107033f5f57eSIlpo Järvinen /* The urg_mode check is necessary during a below snd_una win probe */ 10717691367dSHerbert Xu if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { 10727691367dSHerbert Xu if (before(tp->snd_up, tcb->seq + 0x10000)) { 10731da177e4SLinus Torvalds th->urg_ptr = htons(tp->snd_up - tcb->seq); 10741da177e4SLinus Torvalds th->urg = 1; 10757691367dSHerbert Xu } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { 10760eae88f3SEric Dumazet th->urg_ptr = htons(0xFFFF); 10777691367dSHerbert Xu th->urg = 1; 10787691367dSHerbert Xu } 10791da177e4SLinus Torvalds } 10801da177e4SLinus Torvalds 1081bd0388aeSWilliam Allen Simpson tcp_options_write((__be32 *)(th + 1), tp, &opts); 10824de075e0SEric Dumazet if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0)) 10839e412ba7SIlpo Järvinen TCP_ECN_send(sk, skb, tcp_header_size); 1084dfb4b9dcSDavid S. Miller 1085cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 1086cfb6eeb4SYOSHIFUJI Hideaki /* Calculate the MD5 hash, as we have all we need now */ 1087cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 1088a465419bSEric Dumazet sk_nocaps_add(sk, NETIF_F_GSO_MASK); 1089bd0388aeSWilliam Allen Simpson tp->af_specific->calc_md5_hash(opts.hash_location, 109049a72dfbSAdam Langley md5, sk, NULL, skb); 1091cfb6eeb4SYOSHIFUJI Hideaki } 1092cfb6eeb4SYOSHIFUJI Hideaki #endif 1093cfb6eeb4SYOSHIFUJI Hideaki 1094bb296246SHerbert Xu icsk->icsk_af_ops->send_check(sk, skb); 10951da177e4SLinus Torvalds 10964de075e0SEric Dumazet if (likely(tcb->tcp_flags & TCPHDR_ACK)) 1097fc6415bcSDavid S. Miller tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); 10981da177e4SLinus Torvalds 10991da177e4SLinus Torvalds if (skb->len != tcp_header_size) 1100cf533ea5SEric Dumazet tcp_event_data_sent(tp, sk); 11011da177e4SLinus Torvalds 1102bd37a088SWei Yongjun if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) 1103aa2ea058STom Herbert TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, 1104aa2ea058STom Herbert tcp_skb_pcount(skb)); 11051da177e4SLinus Torvalds 1106d9d8da80SDavid S. Miller err = icsk->icsk_af_ops->queue_xmit(skb, &inet->cork.fl); 110783de47cdSHua Zhong if (likely(err <= 0)) 11081da177e4SLinus Torvalds return err; 11091da177e4SLinus Torvalds 11103cfe3baaSIlpo Järvinen tcp_enter_cwr(sk, 1); 11111da177e4SLinus Torvalds 1112b9df3cb8SGerrit Renker return net_xmit_eval(err); 11131da177e4SLinus Torvalds } 11141da177e4SLinus Torvalds 111567edfef7SAndi Kleen /* This routine just queues the buffer for sending. 11161da177e4SLinus Torvalds * 11171da177e4SLinus Torvalds * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, 11181da177e4SLinus Torvalds * otherwise socket can stall. 11191da177e4SLinus Torvalds */ 11201da177e4SLinus Torvalds static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) 11211da177e4SLinus Torvalds { 11221da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 11231da177e4SLinus Torvalds 11241da177e4SLinus Torvalds /* Advance write_seq and place onto the write_queue. */ 11251da177e4SLinus Torvalds tp->write_seq = TCP_SKB_CB(skb)->end_seq; 11261da177e4SLinus Torvalds skb_header_release(skb); 1127fe067e8aSDavid S. Miller tcp_add_write_queue_tail(sk, skb); 11283ab224beSHideo Aoki sk->sk_wmem_queued += skb->truesize; 11293ab224beSHideo Aoki sk_mem_charge(sk, skb->truesize); 11301da177e4SLinus Torvalds } 11311da177e4SLinus Torvalds 113267edfef7SAndi Kleen /* Initialize TSO segments for a packet. */ 1133cf533ea5SEric Dumazet static void tcp_set_skb_tso_segs(const struct sock *sk, struct sk_buff *skb, 1134056834d9SIlpo Järvinen unsigned int mss_now) 1135f6302d1dSDavid S. Miller { 11368e5b9ddaSHerbert Xu if (skb->len <= mss_now || !sk_can_gso(sk) || 11378e5b9ddaSHerbert Xu skb->ip_summed == CHECKSUM_NONE) { 1138f6302d1dSDavid S. Miller /* Avoid the costly divide in the normal 1139f6302d1dSDavid S. Miller * non-TSO case. 1140f6302d1dSDavid S. Miller */ 11417967168cSHerbert Xu skb_shinfo(skb)->gso_segs = 1; 11427967168cSHerbert Xu skb_shinfo(skb)->gso_size = 0; 11437967168cSHerbert Xu skb_shinfo(skb)->gso_type = 0; 1144f6302d1dSDavid S. Miller } else { 1145356f89e1SIlpo Järvinen skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now); 11467967168cSHerbert Xu skb_shinfo(skb)->gso_size = mss_now; 1147bcd76111SHerbert Xu skb_shinfo(skb)->gso_type = sk->sk_gso_type; 11481da177e4SLinus Torvalds } 11491da177e4SLinus Torvalds } 11501da177e4SLinus Torvalds 115191fed7a1SIlpo Järvinen /* When a modification to fackets out becomes necessary, we need to check 115268f8353bSIlpo Järvinen * skb is counted to fackets_out or not. 115391fed7a1SIlpo Järvinen */ 1154cf533ea5SEric Dumazet static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb, 115591fed7a1SIlpo Järvinen int decr) 115691fed7a1SIlpo Järvinen { 1157a47e5a98SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1158a47e5a98SIlpo Järvinen 1159dc86967bSIlpo Järvinen if (!tp->sacked_out || tcp_is_reno(tp)) 116091fed7a1SIlpo Järvinen return; 116191fed7a1SIlpo Järvinen 11626859d494SIlpo Järvinen if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq)) 116391fed7a1SIlpo Järvinen tp->fackets_out -= decr; 116491fed7a1SIlpo Järvinen } 116591fed7a1SIlpo Järvinen 1166797108d1SIlpo Järvinen /* Pcount in the middle of the write queue got changed, we need to do various 1167797108d1SIlpo Järvinen * tweaks to fix counters 1168797108d1SIlpo Järvinen */ 1169cf533ea5SEric Dumazet static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) 1170797108d1SIlpo Järvinen { 1171797108d1SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1172797108d1SIlpo Järvinen 1173797108d1SIlpo Järvinen tp->packets_out -= decr; 1174797108d1SIlpo Järvinen 1175797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 1176797108d1SIlpo Järvinen tp->sacked_out -= decr; 1177797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 1178797108d1SIlpo Järvinen tp->retrans_out -= decr; 1179797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 1180797108d1SIlpo Järvinen tp->lost_out -= decr; 1181797108d1SIlpo Järvinen 1182797108d1SIlpo Järvinen /* Reno case is special. Sigh... */ 1183797108d1SIlpo Järvinen if (tcp_is_reno(tp) && decr > 0) 1184797108d1SIlpo Järvinen tp->sacked_out -= min_t(u32, tp->sacked_out, decr); 1185797108d1SIlpo Järvinen 1186797108d1SIlpo Järvinen tcp_adjust_fackets_out(sk, skb, decr); 1187797108d1SIlpo Järvinen 1188797108d1SIlpo Järvinen if (tp->lost_skb_hint && 1189797108d1SIlpo Järvinen before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && 119052cf3cc8SIlpo Järvinen (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))) 1191797108d1SIlpo Järvinen tp->lost_cnt_hint -= decr; 1192797108d1SIlpo Järvinen 1193797108d1SIlpo Järvinen tcp_verify_left_out(tp); 1194797108d1SIlpo Järvinen } 1195797108d1SIlpo Järvinen 11961da177e4SLinus Torvalds /* Function to create two new TCP segments. Shrinks the given segment 11971da177e4SLinus Torvalds * to the specified size and appends a new segment with the rest of the 11981da177e4SLinus Torvalds * packet to the list. This won't be called frequently, I hope. 11991da177e4SLinus Torvalds * Remember, these are still headerless SKBs at this point. 12001da177e4SLinus Torvalds */ 1201056834d9SIlpo Järvinen int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, 1202056834d9SIlpo Järvinen unsigned int mss_now) 12031da177e4SLinus Torvalds { 12041da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 12051da177e4SLinus Torvalds struct sk_buff *buff; 12066475be16SDavid S. Miller int nsize, old_factor; 1207b60b49eaSHerbert Xu int nlen; 12089ce01461SIlpo Järvinen u8 flags; 12091da177e4SLinus Torvalds 12102fceec13SIlpo Järvinen if (WARN_ON(len > skb->len)) 12112fceec13SIlpo Järvinen return -EINVAL; 12126a438bbeSStephen Hemminger 12131da177e4SLinus Torvalds nsize = skb_headlen(skb) - len; 12141da177e4SLinus Torvalds if (nsize < 0) 12151da177e4SLinus Torvalds nsize = 0; 12161da177e4SLinus Torvalds 12171da177e4SLinus Torvalds if (skb_cloned(skb) && 12181da177e4SLinus Torvalds skb_is_nonlinear(skb) && 12191da177e4SLinus Torvalds pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) 12201da177e4SLinus Torvalds return -ENOMEM; 12211da177e4SLinus Torvalds 12221da177e4SLinus Torvalds /* Get a new skb... force flag on. */ 12231da177e4SLinus Torvalds buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC); 12241da177e4SLinus Torvalds if (buff == NULL) 12251da177e4SLinus Torvalds return -ENOMEM; /* We'll just try again later. */ 1226ef5cb973SHerbert Xu 12273ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 12283ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1229b60b49eaSHerbert Xu nlen = skb->len - len - nsize; 1230b60b49eaSHerbert Xu buff->truesize += nlen; 1231b60b49eaSHerbert Xu skb->truesize -= nlen; 12321da177e4SLinus Torvalds 12331da177e4SLinus Torvalds /* Correct the sequence numbers. */ 12341da177e4SLinus Torvalds TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 12351da177e4SLinus Torvalds TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 12361da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 12371da177e4SLinus Torvalds 12381da177e4SLinus Torvalds /* PSH and FIN should only be set in the second packet. */ 12394de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 12404de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 12414de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1242e14c3cafSHerbert Xu TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; 12431da177e4SLinus Torvalds 124484fa7933SPatrick McHardy if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { 12451da177e4SLinus Torvalds /* Copy and checksum data tail into the new buffer. */ 1246056834d9SIlpo Järvinen buff->csum = csum_partial_copy_nocheck(skb->data + len, 1247056834d9SIlpo Järvinen skb_put(buff, nsize), 12481da177e4SLinus Torvalds nsize, 0); 12491da177e4SLinus Torvalds 12501da177e4SLinus Torvalds skb_trim(skb, len); 12511da177e4SLinus Torvalds 12521da177e4SLinus Torvalds skb->csum = csum_block_sub(skb->csum, buff->csum, len); 12531da177e4SLinus Torvalds } else { 125484fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 12551da177e4SLinus Torvalds skb_split(skb, buff, len); 12561da177e4SLinus Torvalds } 12571da177e4SLinus Torvalds 12581da177e4SLinus Torvalds buff->ip_summed = skb->ip_summed; 12591da177e4SLinus Torvalds 12601da177e4SLinus Torvalds /* Looks stupid, but our code really uses when of 12611da177e4SLinus Torvalds * skbs, which it never sent before. --ANK 12621da177e4SLinus Torvalds */ 12631da177e4SLinus Torvalds TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when; 1264a61bbcf2SPatrick McHardy buff->tstamp = skb->tstamp; 12651da177e4SLinus Torvalds 12666475be16SDavid S. Miller old_factor = tcp_skb_pcount(skb); 12676475be16SDavid S. Miller 12681da177e4SLinus Torvalds /* Fix up tso_factor for both original and new SKB. */ 1269846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1270846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, buff, mss_now); 12711da177e4SLinus Torvalds 12726475be16SDavid S. Miller /* If this packet has been sent out already, we must 12736475be16SDavid S. Miller * adjust the various packet counters. 12746475be16SDavid S. Miller */ 1275cf0b450cSHerbert Xu if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 12766475be16SDavid S. Miller int diff = old_factor - tcp_skb_pcount(skb) - 12776475be16SDavid S. Miller tcp_skb_pcount(buff); 12781da177e4SLinus Torvalds 1279797108d1SIlpo Järvinen if (diff) 1280797108d1SIlpo Järvinen tcp_adjust_pcount(sk, skb, diff); 12811da177e4SLinus Torvalds } 12821da177e4SLinus Torvalds 12831da177e4SLinus Torvalds /* Link BUFF into the send queue. */ 1284f44b5271SDavid S. Miller skb_header_release(buff); 1285fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 12861da177e4SLinus Torvalds 12871da177e4SLinus Torvalds return 0; 12881da177e4SLinus Torvalds } 12891da177e4SLinus Torvalds 12901da177e4SLinus Torvalds /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c 12911da177e4SLinus Torvalds * eventually). The difference is that pulled data not copied, but 12921da177e4SLinus Torvalds * immediately discarded. 12931da177e4SLinus Torvalds */ 1294f2911969SHerbert Xu ~{PmVHI~} static void __pskb_trim_head(struct sk_buff *skb, int len) 12951da177e4SLinus Torvalds { 12961da177e4SLinus Torvalds int i, k, eat; 12971da177e4SLinus Torvalds 12984fa48bf3SEric Dumazet eat = min_t(int, len, skb_headlen(skb)); 12994fa48bf3SEric Dumazet if (eat) { 13004fa48bf3SEric Dumazet __skb_pull(skb, eat); 130122b4a4f2SEric Dumazet skb->avail_size -= eat; 13024fa48bf3SEric Dumazet len -= eat; 13034fa48bf3SEric Dumazet if (!len) 13044fa48bf3SEric Dumazet return; 13054fa48bf3SEric Dumazet } 13061da177e4SLinus Torvalds eat = len; 13071da177e4SLinus Torvalds k = 0; 13081da177e4SLinus Torvalds for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) { 13099e903e08SEric Dumazet int size = skb_frag_size(&skb_shinfo(skb)->frags[i]); 13109e903e08SEric Dumazet 13119e903e08SEric Dumazet if (size <= eat) { 1312aff65da0SIan Campbell skb_frag_unref(skb, i); 13139e903e08SEric Dumazet eat -= size; 13141da177e4SLinus Torvalds } else { 13151da177e4SLinus Torvalds skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i]; 13161da177e4SLinus Torvalds if (eat) { 13171da177e4SLinus Torvalds skb_shinfo(skb)->frags[k].page_offset += eat; 13189e903e08SEric Dumazet skb_frag_size_sub(&skb_shinfo(skb)->frags[k], eat); 13191da177e4SLinus Torvalds eat = 0; 13201da177e4SLinus Torvalds } 13211da177e4SLinus Torvalds k++; 13221da177e4SLinus Torvalds } 13231da177e4SLinus Torvalds } 13241da177e4SLinus Torvalds skb_shinfo(skb)->nr_frags = k; 13251da177e4SLinus Torvalds 132627a884dcSArnaldo Carvalho de Melo skb_reset_tail_pointer(skb); 13271da177e4SLinus Torvalds skb->data_len -= len; 13281da177e4SLinus Torvalds skb->len = skb->data_len; 13291da177e4SLinus Torvalds } 13301da177e4SLinus Torvalds 133167edfef7SAndi Kleen /* Remove acked data from a packet in the transmit queue. */ 13321da177e4SLinus Torvalds int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 13331da177e4SLinus Torvalds { 1334056834d9SIlpo Järvinen if (skb_cloned(skb) && pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) 13351da177e4SLinus Torvalds return -ENOMEM; 13361da177e4SLinus Torvalds 13374fa48bf3SEric Dumazet __pskb_trim_head(skb, len); 13381da177e4SLinus Torvalds 13391da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq += len; 134084fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 13411da177e4SLinus Torvalds 13421da177e4SLinus Torvalds skb->truesize -= len; 13431da177e4SLinus Torvalds sk->sk_wmem_queued -= len; 13443ab224beSHideo Aoki sk_mem_uncharge(sk, len); 13451da177e4SLinus Torvalds sock_set_flag(sk, SOCK_QUEUE_SHRUNK); 13461da177e4SLinus Torvalds 13475b35e1e6SNeal Cardwell /* Any change of skb->len requires recalculation of tso factor. */ 13481da177e4SLinus Torvalds if (tcp_skb_pcount(skb) > 1) 13495b35e1e6SNeal Cardwell tcp_set_skb_tso_segs(sk, skb, tcp_skb_mss(skb)); 13501da177e4SLinus Torvalds 13511da177e4SLinus Torvalds return 0; 13521da177e4SLinus Torvalds } 13531da177e4SLinus Torvalds 135467edfef7SAndi Kleen /* Calculate MSS. Not accounting for SACKs here. */ 135567469601SEric Dumazet int tcp_mtu_to_mss(struct sock *sk, int pmtu) 13565d424d5aSJohn Heffner { 1357cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1358cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 13595d424d5aSJohn Heffner int mss_now; 13605d424d5aSJohn Heffner 13615d424d5aSJohn Heffner /* Calculate base mss without TCP options: 13625d424d5aSJohn Heffner It is MMS_S - sizeof(tcphdr) of rfc1122 13635d424d5aSJohn Heffner */ 13645d424d5aSJohn Heffner mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 13655d424d5aSJohn Heffner 136667469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 136767469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 136867469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 136967469601SEric Dumazet 137067469601SEric Dumazet if (dst && dst_allfrag(dst)) 137167469601SEric Dumazet mss_now -= icsk->icsk_af_ops->net_frag_header_len; 137267469601SEric Dumazet } 137367469601SEric Dumazet 13745d424d5aSJohn Heffner /* Clamp it (mss_clamp does not include tcp options) */ 13755d424d5aSJohn Heffner if (mss_now > tp->rx_opt.mss_clamp) 13765d424d5aSJohn Heffner mss_now = tp->rx_opt.mss_clamp; 13775d424d5aSJohn Heffner 13785d424d5aSJohn Heffner /* Now subtract optional transport overhead */ 13795d424d5aSJohn Heffner mss_now -= icsk->icsk_ext_hdr_len; 13805d424d5aSJohn Heffner 13815d424d5aSJohn Heffner /* Then reserve room for full set of TCP options and 8 bytes of data */ 13825d424d5aSJohn Heffner if (mss_now < 48) 13835d424d5aSJohn Heffner mss_now = 48; 13845d424d5aSJohn Heffner 13855d424d5aSJohn Heffner /* Now subtract TCP options size, not including SACKs */ 13865d424d5aSJohn Heffner mss_now -= tp->tcp_header_len - sizeof(struct tcphdr); 13875d424d5aSJohn Heffner 13885d424d5aSJohn Heffner return mss_now; 13895d424d5aSJohn Heffner } 13905d424d5aSJohn Heffner 13915d424d5aSJohn Heffner /* Inverse of above */ 139267469601SEric Dumazet int tcp_mss_to_mtu(struct sock *sk, int mss) 13935d424d5aSJohn Heffner { 1394cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1395cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 13965d424d5aSJohn Heffner int mtu; 13975d424d5aSJohn Heffner 13985d424d5aSJohn Heffner mtu = mss + 13995d424d5aSJohn Heffner tp->tcp_header_len + 14005d424d5aSJohn Heffner icsk->icsk_ext_hdr_len + 14015d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 14025d424d5aSJohn Heffner 140367469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 140467469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 140567469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 140667469601SEric Dumazet 140767469601SEric Dumazet if (dst && dst_allfrag(dst)) 140867469601SEric Dumazet mtu += icsk->icsk_af_ops->net_frag_header_len; 140967469601SEric Dumazet } 14105d424d5aSJohn Heffner return mtu; 14115d424d5aSJohn Heffner } 14125d424d5aSJohn Heffner 141367edfef7SAndi Kleen /* MTU probing init per socket */ 14145d424d5aSJohn Heffner void tcp_mtup_init(struct sock *sk) 14155d424d5aSJohn Heffner { 14165d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 14175d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 14185d424d5aSJohn Heffner 14195d424d5aSJohn Heffner icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1; 14205d424d5aSJohn Heffner icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 14215d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 14225d424d5aSJohn Heffner icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss); 14235d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 14245d424d5aSJohn Heffner } 14254bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_mtup_init); 14265d424d5aSJohn Heffner 14271da177e4SLinus Torvalds /* This function synchronize snd mss to current pmtu/exthdr set. 14281da177e4SLinus Torvalds 14291da177e4SLinus Torvalds tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 14301da177e4SLinus Torvalds for TCP options, but includes only bare TCP header. 14311da177e4SLinus Torvalds 14321da177e4SLinus Torvalds tp->rx_opt.mss_clamp is mss negotiated at connection setup. 1433caa20d9aSStephen Hemminger It is minimum of user_mss and mss received with SYN. 14341da177e4SLinus Torvalds It also does not include TCP options. 14351da177e4SLinus Torvalds 1436d83d8461SArnaldo Carvalho de Melo inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. 14371da177e4SLinus Torvalds 14381da177e4SLinus Torvalds tp->mss_cache is current effective sending mss, including 14391da177e4SLinus Torvalds all tcp options except for SACKs. It is evaluated, 14401da177e4SLinus Torvalds taking into account current pmtu, but never exceeds 14411da177e4SLinus Torvalds tp->rx_opt.mss_clamp. 14421da177e4SLinus Torvalds 14431da177e4SLinus Torvalds NOTE1. rfc1122 clearly states that advertised MSS 14441da177e4SLinus Torvalds DOES NOT include either tcp or ip options. 14451da177e4SLinus Torvalds 1446d83d8461SArnaldo Carvalho de Melo NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache 1447d83d8461SArnaldo Carvalho de Melo are READ ONLY outside this function. --ANK (980731) 14481da177e4SLinus Torvalds */ 14491da177e4SLinus Torvalds unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 14501da177e4SLinus Torvalds { 14511da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1452d83d8461SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 14535d424d5aSJohn Heffner int mss_now; 14541da177e4SLinus Torvalds 14555d424d5aSJohn Heffner if (icsk->icsk_mtup.search_high > pmtu) 14565d424d5aSJohn Heffner icsk->icsk_mtup.search_high = pmtu; 14571da177e4SLinus Torvalds 14585d424d5aSJohn Heffner mss_now = tcp_mtu_to_mss(sk, pmtu); 1459409d22b4SIlpo Järvinen mss_now = tcp_bound_to_half_wnd(tp, mss_now); 14601da177e4SLinus Torvalds 14611da177e4SLinus Torvalds /* And store cached results */ 1462d83d8461SArnaldo Carvalho de Melo icsk->icsk_pmtu_cookie = pmtu; 14635d424d5aSJohn Heffner if (icsk->icsk_mtup.enabled) 14645d424d5aSJohn Heffner mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); 1465c1b4a7e6SDavid S. Miller tp->mss_cache = mss_now; 14661da177e4SLinus Torvalds 14671da177e4SLinus Torvalds return mss_now; 14681da177e4SLinus Torvalds } 14694bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_sync_mss); 14701da177e4SLinus Torvalds 14711da177e4SLinus Torvalds /* Compute the current effective MSS, taking SACKs and IP options, 14721da177e4SLinus Torvalds * and even PMTU discovery events into account. 14731da177e4SLinus Torvalds */ 14740c54b85fSIlpo Järvinen unsigned int tcp_current_mss(struct sock *sk) 14751da177e4SLinus Torvalds { 1476cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1477cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1478c1b4a7e6SDavid S. Miller u32 mss_now; 147995c96174SEric Dumazet unsigned int header_len; 148033ad798cSAdam Langley struct tcp_out_options opts; 148133ad798cSAdam Langley struct tcp_md5sig_key *md5; 14821da177e4SLinus Torvalds 1483c1b4a7e6SDavid S. Miller mss_now = tp->mss_cache; 1484c1b4a7e6SDavid S. Miller 14851da177e4SLinus Torvalds if (dst) { 14861da177e4SLinus Torvalds u32 mtu = dst_mtu(dst); 1487d83d8461SArnaldo Carvalho de Melo if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 14881da177e4SLinus Torvalds mss_now = tcp_sync_mss(sk, mtu); 14891da177e4SLinus Torvalds } 14901da177e4SLinus Torvalds 149133ad798cSAdam Langley header_len = tcp_established_options(sk, NULL, &opts, &md5) + 149233ad798cSAdam Langley sizeof(struct tcphdr); 149333ad798cSAdam Langley /* The mss_cache is sized based on tp->tcp_header_len, which assumes 149433ad798cSAdam Langley * some common options. If this is an odd packet (because we have SACK 149533ad798cSAdam Langley * blocks etc) then our calculated header_len will be different, and 149633ad798cSAdam Langley * we have to adjust mss_now correspondingly */ 149733ad798cSAdam Langley if (header_len != tp->tcp_header_len) { 149833ad798cSAdam Langley int delta = (int) header_len - tp->tcp_header_len; 149933ad798cSAdam Langley mss_now -= delta; 150033ad798cSAdam Langley } 1501cfb6eeb4SYOSHIFUJI Hideaki 15021da177e4SLinus Torvalds return mss_now; 15031da177e4SLinus Torvalds } 15041da177e4SLinus Torvalds 1505a762a980SDavid S. Miller /* Congestion window validation. (RFC2861) */ 15069e412ba7SIlpo Järvinen static void tcp_cwnd_validate(struct sock *sk) 1507a762a980SDavid S. Miller { 15089e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1509a762a980SDavid S. Miller 1510d436d686SIlpo Järvinen if (tp->packets_out >= tp->snd_cwnd) { 1511a762a980SDavid S. Miller /* Network is feed fully. */ 1512a762a980SDavid S. Miller tp->snd_cwnd_used = 0; 1513a762a980SDavid S. Miller tp->snd_cwnd_stamp = tcp_time_stamp; 1514a762a980SDavid S. Miller } else { 1515a762a980SDavid S. Miller /* Network starves. */ 1516a762a980SDavid S. Miller if (tp->packets_out > tp->snd_cwnd_used) 1517a762a980SDavid S. Miller tp->snd_cwnd_used = tp->packets_out; 1518a762a980SDavid S. Miller 151915d33c07SDavid S. Miller if (sysctl_tcp_slow_start_after_idle && 152015d33c07SDavid S. Miller (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto) 1521a762a980SDavid S. Miller tcp_cwnd_application_limited(sk); 1522a762a980SDavid S. Miller } 1523a762a980SDavid S. Miller } 1524a762a980SDavid S. Miller 15250e3a4803SIlpo Järvinen /* Returns the portion of skb which can be sent right away without 15260e3a4803SIlpo Järvinen * introducing MSS oddities to segment boundaries. In rare cases where 15270e3a4803SIlpo Järvinen * mss_now != mss_cache, we will request caller to create a small skb 15280e3a4803SIlpo Järvinen * per input skb which could be mostly avoided here (if desired). 15295ea3a748SIlpo Järvinen * 15305ea3a748SIlpo Järvinen * We explicitly want to create a request for splitting write queue tail 15315ea3a748SIlpo Järvinen * to a small skb for Nagle purposes while avoiding unnecessary modulos, 15325ea3a748SIlpo Järvinen * thus all the complexity (cwnd_len is always MSS multiple which we 15335ea3a748SIlpo Järvinen * return whenever allowed by the other factors). Basically we need the 15345ea3a748SIlpo Järvinen * modulo only when the receiver window alone is the limiting factor or 15355ea3a748SIlpo Järvinen * when we would be allowed to send the split-due-to-Nagle skb fully. 15360e3a4803SIlpo Järvinen */ 1537cf533ea5SEric Dumazet static unsigned int tcp_mss_split_point(const struct sock *sk, const struct sk_buff *skb, 15381485348dSBen Hutchings unsigned int mss_now, unsigned int max_segs) 1539c1b4a7e6SDavid S. Miller { 1540cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 15411485348dSBen Hutchings u32 needed, window, max_len; 1542c1b4a7e6SDavid S. Miller 154390840defSIlpo Järvinen window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 15441485348dSBen Hutchings max_len = mss_now * max_segs; 15450e3a4803SIlpo Järvinen 15461485348dSBen Hutchings if (likely(max_len <= window && skb != tcp_write_queue_tail(sk))) 15471485348dSBen Hutchings return max_len; 15480e3a4803SIlpo Järvinen 15495ea3a748SIlpo Järvinen needed = min(skb->len, window); 15505ea3a748SIlpo Järvinen 15511485348dSBen Hutchings if (max_len <= needed) 15521485348dSBen Hutchings return max_len; 15530e3a4803SIlpo Järvinen 15540e3a4803SIlpo Järvinen return needed - needed % mss_now; 1555c1b4a7e6SDavid S. Miller } 1556c1b4a7e6SDavid S. Miller 1557c1b4a7e6SDavid S. Miller /* Can at least one segment of SKB be sent right now, according to the 1558c1b4a7e6SDavid S. Miller * congestion window rules? If so, return how many segments are allowed. 1559c1b4a7e6SDavid S. Miller */ 1560cf533ea5SEric Dumazet static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, 1561cf533ea5SEric Dumazet const struct sk_buff *skb) 1562c1b4a7e6SDavid S. Miller { 1563c1b4a7e6SDavid S. Miller u32 in_flight, cwnd; 1564c1b4a7e6SDavid S. Miller 1565c1b4a7e6SDavid S. Miller /* Don't be strict about the congestion window for the final FIN. */ 15664de075e0SEric Dumazet if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && 15674de075e0SEric Dumazet tcp_skb_pcount(skb) == 1) 1568c1b4a7e6SDavid S. Miller return 1; 1569c1b4a7e6SDavid S. Miller 1570c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1571c1b4a7e6SDavid S. Miller cwnd = tp->snd_cwnd; 1572c1b4a7e6SDavid S. Miller if (in_flight < cwnd) 1573c1b4a7e6SDavid S. Miller return (cwnd - in_flight); 1574c1b4a7e6SDavid S. Miller 1575c1b4a7e6SDavid S. Miller return 0; 1576c1b4a7e6SDavid S. Miller } 1577c1b4a7e6SDavid S. Miller 1578b595076aSUwe Kleine-König /* Initialize TSO state of a skb. 157967edfef7SAndi Kleen * This must be invoked the first time we consider transmitting 1580c1b4a7e6SDavid S. Miller * SKB onto the wire. 1581c1b4a7e6SDavid S. Miller */ 1582cf533ea5SEric Dumazet static int tcp_init_tso_segs(const struct sock *sk, struct sk_buff *skb, 1583056834d9SIlpo Järvinen unsigned int mss_now) 1584c1b4a7e6SDavid S. Miller { 1585c1b4a7e6SDavid S. Miller int tso_segs = tcp_skb_pcount(skb); 1586c1b4a7e6SDavid S. Miller 1587f8269a49SIlpo Järvinen if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { 1588846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1589c1b4a7e6SDavid S. Miller tso_segs = tcp_skb_pcount(skb); 1590c1b4a7e6SDavid S. Miller } 1591c1b4a7e6SDavid S. Miller return tso_segs; 1592c1b4a7e6SDavid S. Miller } 1593c1b4a7e6SDavid S. Miller 159467edfef7SAndi Kleen /* Minshall's variant of the Nagle send check. */ 1595a2a385d6SEric Dumazet static inline bool tcp_minshall_check(const struct tcp_sock *tp) 1596c1b4a7e6SDavid S. Miller { 1597c1b4a7e6SDavid S. Miller return after(tp->snd_sml, tp->snd_una) && 1598c1b4a7e6SDavid S. Miller !after(tp->snd_sml, tp->snd_nxt); 1599c1b4a7e6SDavid S. Miller } 1600c1b4a7e6SDavid S. Miller 1601a2a385d6SEric Dumazet /* Return false, if packet can be sent now without violation Nagle's rules: 1602c1b4a7e6SDavid S. Miller * 1. It is full sized. 1603c1b4a7e6SDavid S. Miller * 2. Or it contains FIN. (already checked by caller) 16046d67e9beSFeng King * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. 1605c1b4a7e6SDavid S. Miller * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 1606c1b4a7e6SDavid S. Miller * With Minshall's modification: all sent small packets are ACKed. 1607c1b4a7e6SDavid S. Miller */ 1608a2a385d6SEric Dumazet static inline bool tcp_nagle_check(const struct tcp_sock *tp, 1609c1b4a7e6SDavid S. Miller const struct sk_buff *skb, 161095c96174SEric Dumazet unsigned int mss_now, int nonagle) 1611c1b4a7e6SDavid S. Miller { 1612a02cec21SEric Dumazet return skb->len < mss_now && 1613c1b4a7e6SDavid S. Miller ((nonagle & TCP_NAGLE_CORK) || 1614a02cec21SEric Dumazet (!nonagle && tp->packets_out && tcp_minshall_check(tp))); 1615c1b4a7e6SDavid S. Miller } 1616c1b4a7e6SDavid S. Miller 1617a2a385d6SEric Dumazet /* Return true if the Nagle test allows this packet to be 1618c1b4a7e6SDavid S. Miller * sent now. 1619c1b4a7e6SDavid S. Miller */ 1620a2a385d6SEric Dumazet static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, 1621c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1622c1b4a7e6SDavid S. Miller { 1623c1b4a7e6SDavid S. Miller /* Nagle rule does not apply to frames, which sit in the middle of the 1624c1b4a7e6SDavid S. Miller * write_queue (they have no chances to get new data). 1625c1b4a7e6SDavid S. Miller * 1626c1b4a7e6SDavid S. Miller * This is implemented in the callers, where they modify the 'nonagle' 1627c1b4a7e6SDavid S. Miller * argument based upon the location of SKB in the send queue. 1628c1b4a7e6SDavid S. Miller */ 1629c1b4a7e6SDavid S. Miller if (nonagle & TCP_NAGLE_PUSH) 1630a2a385d6SEric Dumazet return true; 1631c1b4a7e6SDavid S. Miller 1632d551e454SIlpo Järvinen /* Don't use the nagle rule for urgent data (or for the final FIN). 1633d551e454SIlpo Järvinen * Nagle can be ignored during F-RTO too (see RFC4138). 1634d551e454SIlpo Järvinen */ 163533f5f57eSIlpo Järvinen if (tcp_urg_mode(tp) || (tp->frto_counter == 2) || 16364de075e0SEric Dumazet (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) 1637a2a385d6SEric Dumazet return true; 1638c1b4a7e6SDavid S. Miller 1639c1b4a7e6SDavid S. Miller if (!tcp_nagle_check(tp, skb, cur_mss, nonagle)) 1640a2a385d6SEric Dumazet return true; 1641c1b4a7e6SDavid S. Miller 1642a2a385d6SEric Dumazet return false; 1643c1b4a7e6SDavid S. Miller } 1644c1b4a7e6SDavid S. Miller 1645c1b4a7e6SDavid S. Miller /* Does at least the first segment of SKB fit into the send window? */ 1646a2a385d6SEric Dumazet static bool tcp_snd_wnd_test(const struct tcp_sock *tp, 1647a2a385d6SEric Dumazet const struct sk_buff *skb, 1648056834d9SIlpo Järvinen unsigned int cur_mss) 1649c1b4a7e6SDavid S. Miller { 1650c1b4a7e6SDavid S. Miller u32 end_seq = TCP_SKB_CB(skb)->end_seq; 1651c1b4a7e6SDavid S. Miller 1652c1b4a7e6SDavid S. Miller if (skb->len > cur_mss) 1653c1b4a7e6SDavid S. Miller end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 1654c1b4a7e6SDavid S. Miller 165590840defSIlpo Järvinen return !after(end_seq, tcp_wnd_end(tp)); 1656c1b4a7e6SDavid S. Miller } 1657c1b4a7e6SDavid S. Miller 1658fe067e8aSDavid S. Miller /* This checks if the data bearing packet SKB (usually tcp_send_head(sk)) 1659c1b4a7e6SDavid S. Miller * should be put on the wire right now. If so, it returns the number of 1660c1b4a7e6SDavid S. Miller * packets allowed by the congestion window. 1661c1b4a7e6SDavid S. Miller */ 1662cf533ea5SEric Dumazet static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb, 1663c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1664c1b4a7e6SDavid S. Miller { 1665cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1666c1b4a7e6SDavid S. Miller unsigned int cwnd_quota; 1667c1b4a7e6SDavid S. Miller 1668846998aeSDavid S. Miller tcp_init_tso_segs(sk, skb, cur_mss); 1669c1b4a7e6SDavid S. Miller 1670c1b4a7e6SDavid S. Miller if (!tcp_nagle_test(tp, skb, cur_mss, nonagle)) 1671c1b4a7e6SDavid S. Miller return 0; 1672c1b4a7e6SDavid S. Miller 1673c1b4a7e6SDavid S. Miller cwnd_quota = tcp_cwnd_test(tp, skb); 1674056834d9SIlpo Järvinen if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss)) 1675c1b4a7e6SDavid S. Miller cwnd_quota = 0; 1676c1b4a7e6SDavid S. Miller 1677c1b4a7e6SDavid S. Miller return cwnd_quota; 1678c1b4a7e6SDavid S. Miller } 1679c1b4a7e6SDavid S. Miller 168067edfef7SAndi Kleen /* Test if sending is allowed right now. */ 1681a2a385d6SEric Dumazet bool tcp_may_send_now(struct sock *sk) 1682c1b4a7e6SDavid S. Miller { 1683cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1684fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 1685c1b4a7e6SDavid S. Miller 1686a02cec21SEric Dumazet return skb && 16870c54b85fSIlpo Järvinen tcp_snd_test(sk, skb, tcp_current_mss(sk), 1688c1b4a7e6SDavid S. Miller (tcp_skb_is_last(sk, skb) ? 1689a02cec21SEric Dumazet tp->nonagle : TCP_NAGLE_PUSH)); 1690c1b4a7e6SDavid S. Miller } 1691c1b4a7e6SDavid S. Miller 1692c1b4a7e6SDavid S. Miller /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 1693c1b4a7e6SDavid S. Miller * which is put after SKB on the list. It is very much like 1694c1b4a7e6SDavid S. Miller * tcp_fragment() except that it may make several kinds of assumptions 1695c1b4a7e6SDavid S. Miller * in order to speed up the splitting operation. In particular, we 1696c1b4a7e6SDavid S. Miller * know that all the data is in scatter-gather pages, and that the 1697c1b4a7e6SDavid S. Miller * packet has never been sent out before (and thus is not cloned). 1698c1b4a7e6SDavid S. Miller */ 1699056834d9SIlpo Järvinen static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, 1700c4ead4c5SEric Dumazet unsigned int mss_now, gfp_t gfp) 1701c1b4a7e6SDavid S. Miller { 1702c1b4a7e6SDavid S. Miller struct sk_buff *buff; 1703c1b4a7e6SDavid S. Miller int nlen = skb->len - len; 17049ce01461SIlpo Järvinen u8 flags; 1705c1b4a7e6SDavid S. Miller 1706c1b4a7e6SDavid S. Miller /* All of a TSO frame must be composed of paged data. */ 1707c8ac3774SHerbert Xu if (skb->len != skb->data_len) 1708c8ac3774SHerbert Xu return tcp_fragment(sk, skb, len, mss_now); 1709c1b4a7e6SDavid S. Miller 1710c4ead4c5SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, gfp); 1711c1b4a7e6SDavid S. Miller if (unlikely(buff == NULL)) 1712c1b4a7e6SDavid S. Miller return -ENOMEM; 1713c1b4a7e6SDavid S. Miller 17143ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 17153ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1716b60b49eaSHerbert Xu buff->truesize += nlen; 1717c1b4a7e6SDavid S. Miller skb->truesize -= nlen; 1718c1b4a7e6SDavid S. Miller 1719c1b4a7e6SDavid S. Miller /* Correct the sequence numbers. */ 1720c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 1721c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 1722c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 1723c1b4a7e6SDavid S. Miller 1724c1b4a7e6SDavid S. Miller /* PSH and FIN should only be set in the second packet. */ 17254de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 17264de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 17274de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1728c1b4a7e6SDavid S. Miller 1729c1b4a7e6SDavid S. Miller /* This packet was never sent out yet, so no SACK bits. */ 1730c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->sacked = 0; 1731c1b4a7e6SDavid S. Miller 173284fa7933SPatrick McHardy buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; 1733c1b4a7e6SDavid S. Miller skb_split(skb, buff, len); 1734c1b4a7e6SDavid S. Miller 1735c1b4a7e6SDavid S. Miller /* Fix up tso_factor for both original and new SKB. */ 1736846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1737846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, buff, mss_now); 1738c1b4a7e6SDavid S. Miller 1739c1b4a7e6SDavid S. Miller /* Link BUFF into the send queue. */ 1740c1b4a7e6SDavid S. Miller skb_header_release(buff); 1741fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 1742c1b4a7e6SDavid S. Miller 1743c1b4a7e6SDavid S. Miller return 0; 1744c1b4a7e6SDavid S. Miller } 1745c1b4a7e6SDavid S. Miller 1746c1b4a7e6SDavid S. Miller /* Try to defer sending, if possible, in order to minimize the amount 1747c1b4a7e6SDavid S. Miller * of TSO splitting we do. View it as a kind of TSO Nagle test. 1748c1b4a7e6SDavid S. Miller * 1749c1b4a7e6SDavid S. Miller * This algorithm is from John Heffner. 1750c1b4a7e6SDavid S. Miller */ 1751a2a385d6SEric Dumazet static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb) 1752c1b4a7e6SDavid S. Miller { 17539e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 17546687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 1755c1b4a7e6SDavid S. Miller u32 send_win, cong_win, limit, in_flight; 1756ad9f4f50SEric Dumazet int win_divisor; 1757c1b4a7e6SDavid S. Miller 17584de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 1759ae8064acSJohn Heffner goto send_now; 1760c1b4a7e6SDavid S. Miller 17616687e988SArnaldo Carvalho de Melo if (icsk->icsk_ca_state != TCP_CA_Open) 1762ae8064acSJohn Heffner goto send_now; 1763ae8064acSJohn Heffner 1764ae8064acSJohn Heffner /* Defer for less than two clock ticks. */ 1765bd515c3eSIlpo Järvinen if (tp->tso_deferred && 1766a2acde07SIlpo Järvinen (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1) 1767ae8064acSJohn Heffner goto send_now; 1768908a75c1SDavid S. Miller 1769c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1770c1b4a7e6SDavid S. Miller 1771056834d9SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight)); 1772c1b4a7e6SDavid S. Miller 177390840defSIlpo Järvinen send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1774c1b4a7e6SDavid S. Miller 1775c1b4a7e6SDavid S. Miller /* From in_flight test above, we know that cwnd > in_flight. */ 1776c1b4a7e6SDavid S. Miller cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; 1777c1b4a7e6SDavid S. Miller 1778c1b4a7e6SDavid S. Miller limit = min(send_win, cong_win); 1779c1b4a7e6SDavid S. Miller 1780ba244fe9SDavid S. Miller /* If a full-sized TSO skb can be sent, do it. */ 17811485348dSBen Hutchings if (limit >= min_t(unsigned int, sk->sk_gso_max_size, 17821485348dSBen Hutchings sk->sk_gso_max_segs * tp->mss_cache)) 1783ae8064acSJohn Heffner goto send_now; 1784ba244fe9SDavid S. Miller 178562ad2761SIlpo Järvinen /* Middle in queue won't get any more data, full sendable already? */ 178662ad2761SIlpo Järvinen if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) 178762ad2761SIlpo Järvinen goto send_now; 178862ad2761SIlpo Järvinen 1789ad9f4f50SEric Dumazet win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor); 1790ad9f4f50SEric Dumazet if (win_divisor) { 1791c1b4a7e6SDavid S. Miller u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); 1792c1b4a7e6SDavid S. Miller 1793c1b4a7e6SDavid S. Miller /* If at least some fraction of a window is available, 1794c1b4a7e6SDavid S. Miller * just use it. 1795c1b4a7e6SDavid S. Miller */ 1796ad9f4f50SEric Dumazet chunk /= win_divisor; 1797c1b4a7e6SDavid S. Miller if (limit >= chunk) 1798ae8064acSJohn Heffner goto send_now; 1799c1b4a7e6SDavid S. Miller } else { 1800c1b4a7e6SDavid S. Miller /* Different approach, try not to defer past a single 1801c1b4a7e6SDavid S. Miller * ACK. Receiver should ACK every other full sized 1802c1b4a7e6SDavid S. Miller * frame, so if we have space for more than 3 frames 1803c1b4a7e6SDavid S. Miller * then send now. 1804c1b4a7e6SDavid S. Miller */ 18056b5a5c0dSNeal Cardwell if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) 1806ae8064acSJohn Heffner goto send_now; 1807c1b4a7e6SDavid S. Miller } 1808c1b4a7e6SDavid S. Miller 1809c1b4a7e6SDavid S. Miller /* Ok, it looks like it is advisable to defer. */ 1810ae8064acSJohn Heffner tp->tso_deferred = 1 | (jiffies << 1); 1811ae8064acSJohn Heffner 1812a2a385d6SEric Dumazet return true; 1813ae8064acSJohn Heffner 1814ae8064acSJohn Heffner send_now: 1815ae8064acSJohn Heffner tp->tso_deferred = 0; 1816a2a385d6SEric Dumazet return false; 1817c1b4a7e6SDavid S. Miller } 1818c1b4a7e6SDavid S. Miller 18195d424d5aSJohn Heffner /* Create a new MTU probe if we are ready. 182067edfef7SAndi Kleen * MTU probe is regularly attempting to increase the path MTU by 182167edfef7SAndi Kleen * deliberately sending larger packets. This discovers routing 182267edfef7SAndi Kleen * changes resulting in larger path MTUs. 182367edfef7SAndi Kleen * 18245d424d5aSJohn Heffner * Returns 0 if we should wait to probe (no cwnd available), 18255d424d5aSJohn Heffner * 1 if a probe was sent, 1826056834d9SIlpo Järvinen * -1 otherwise 1827056834d9SIlpo Järvinen */ 18285d424d5aSJohn Heffner static int tcp_mtu_probe(struct sock *sk) 18295d424d5aSJohn Heffner { 18305d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 18315d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 18325d424d5aSJohn Heffner struct sk_buff *skb, *nskb, *next; 18335d424d5aSJohn Heffner int len; 18345d424d5aSJohn Heffner int probe_size; 183591cc17c0SIlpo Järvinen int size_needed; 18365d424d5aSJohn Heffner int copy; 18375d424d5aSJohn Heffner int mss_now; 18385d424d5aSJohn Heffner 18395d424d5aSJohn Heffner /* Not currently probing/verifying, 18405d424d5aSJohn Heffner * not in recovery, 18415d424d5aSJohn Heffner * have enough cwnd, and 18425d424d5aSJohn Heffner * not SACKing (the variable headers throw things off) */ 18435d424d5aSJohn Heffner if (!icsk->icsk_mtup.enabled || 18445d424d5aSJohn Heffner icsk->icsk_mtup.probe_size || 18455d424d5aSJohn Heffner inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 18465d424d5aSJohn Heffner tp->snd_cwnd < 11 || 1847cabeccbdSIlpo Järvinen tp->rx_opt.num_sacks || tp->rx_opt.dsack) 18485d424d5aSJohn Heffner return -1; 18495d424d5aSJohn Heffner 18505d424d5aSJohn Heffner /* Very simple search strategy: just double the MSS. */ 18510c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 18525d424d5aSJohn Heffner probe_size = 2 * tp->mss_cache; 185391cc17c0SIlpo Järvinen size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; 18545d424d5aSJohn Heffner if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) { 18555d424d5aSJohn Heffner /* TODO: set timer for probe_converge_event */ 18565d424d5aSJohn Heffner return -1; 18575d424d5aSJohn Heffner } 18585d424d5aSJohn Heffner 18595d424d5aSJohn Heffner /* Have enough data in the send queue to probe? */ 18607f9c33e5SIlpo Järvinen if (tp->write_seq - tp->snd_nxt < size_needed) 18615d424d5aSJohn Heffner return -1; 18625d424d5aSJohn Heffner 186391cc17c0SIlpo Järvinen if (tp->snd_wnd < size_needed) 18645d424d5aSJohn Heffner return -1; 186590840defSIlpo Järvinen if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) 18665d424d5aSJohn Heffner return 0; 18675d424d5aSJohn Heffner 1868d67c58e9SIlpo Järvinen /* Do we need to wait to drain cwnd? With none in flight, don't stall */ 1869d67c58e9SIlpo Järvinen if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { 1870d67c58e9SIlpo Järvinen if (!tcp_packets_in_flight(tp)) 18715d424d5aSJohn Heffner return -1; 18725d424d5aSJohn Heffner else 18735d424d5aSJohn Heffner return 0; 18745d424d5aSJohn Heffner } 18755d424d5aSJohn Heffner 18765d424d5aSJohn Heffner /* We're allowed to probe. Build it now. */ 18775d424d5aSJohn Heffner if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL) 18785d424d5aSJohn Heffner return -1; 18793ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 18803ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 18815d424d5aSJohn Heffner 1882fe067e8aSDavid S. Miller skb = tcp_send_head(sk); 18835d424d5aSJohn Heffner 18845d424d5aSJohn Heffner TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 18855d424d5aSJohn Heffner TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 18864de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; 18875d424d5aSJohn Heffner TCP_SKB_CB(nskb)->sacked = 0; 18885d424d5aSJohn Heffner nskb->csum = 0; 188984fa7933SPatrick McHardy nskb->ip_summed = skb->ip_summed; 18905d424d5aSJohn Heffner 189150c4817eSIlpo Järvinen tcp_insert_write_queue_before(nskb, skb, sk); 189250c4817eSIlpo Järvinen 18935d424d5aSJohn Heffner len = 0; 1894234b6860SIlpo Järvinen tcp_for_write_queue_from_safe(skb, next, sk) { 18955d424d5aSJohn Heffner copy = min_t(int, skb->len, probe_size - len); 18965d424d5aSJohn Heffner if (nskb->ip_summed) 18975d424d5aSJohn Heffner skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); 18985d424d5aSJohn Heffner else 18995d424d5aSJohn Heffner nskb->csum = skb_copy_and_csum_bits(skb, 0, 1900056834d9SIlpo Järvinen skb_put(nskb, copy), 1901056834d9SIlpo Järvinen copy, nskb->csum); 19025d424d5aSJohn Heffner 19035d424d5aSJohn Heffner if (skb->len <= copy) { 19045d424d5aSJohn Heffner /* We've eaten all the data from this skb. 19055d424d5aSJohn Heffner * Throw it away. */ 19064de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 1907fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 19083ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 19095d424d5aSJohn Heffner } else { 19104de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & 1911a3433f35SChangli Gao ~(TCPHDR_FIN|TCPHDR_PSH); 19125d424d5aSJohn Heffner if (!skb_shinfo(skb)->nr_frags) { 19135d424d5aSJohn Heffner skb_pull(skb, copy); 191484fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 1915056834d9SIlpo Järvinen skb->csum = csum_partial(skb->data, 1916056834d9SIlpo Järvinen skb->len, 0); 19175d424d5aSJohn Heffner } else { 19185d424d5aSJohn Heffner __pskb_trim_head(skb, copy); 19195d424d5aSJohn Heffner tcp_set_skb_tso_segs(sk, skb, mss_now); 19205d424d5aSJohn Heffner } 19215d424d5aSJohn Heffner TCP_SKB_CB(skb)->seq += copy; 19225d424d5aSJohn Heffner } 19235d424d5aSJohn Heffner 19245d424d5aSJohn Heffner len += copy; 1925234b6860SIlpo Järvinen 1926234b6860SIlpo Järvinen if (len >= probe_size) 1927234b6860SIlpo Järvinen break; 19285d424d5aSJohn Heffner } 19295d424d5aSJohn Heffner tcp_init_tso_segs(sk, nskb, nskb->len); 19305d424d5aSJohn Heffner 19315d424d5aSJohn Heffner /* We're ready to send. If this fails, the probe will 19325d424d5aSJohn Heffner * be resegmented into mss-sized pieces by tcp_write_xmit(). */ 19335d424d5aSJohn Heffner TCP_SKB_CB(nskb)->when = tcp_time_stamp; 19345d424d5aSJohn Heffner if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 19355d424d5aSJohn Heffner /* Decrement cwnd here because we are sending 19365d424d5aSJohn Heffner * effectively two packets. */ 19375d424d5aSJohn Heffner tp->snd_cwnd--; 193866f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, nskb); 19395d424d5aSJohn Heffner 19405d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 19410e7b1368SJohn Heffner tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 19420e7b1368SJohn Heffner tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 19435d424d5aSJohn Heffner 19445d424d5aSJohn Heffner return 1; 19455d424d5aSJohn Heffner } 19465d424d5aSJohn Heffner 19475d424d5aSJohn Heffner return -1; 19485d424d5aSJohn Heffner } 19495d424d5aSJohn Heffner 19501da177e4SLinus Torvalds /* This routine writes packets to the network. It advances the 19511da177e4SLinus Torvalds * send_head. This happens as incoming acks open up the remote 19521da177e4SLinus Torvalds * window for us. 19531da177e4SLinus Torvalds * 1954f8269a49SIlpo Järvinen * LARGESEND note: !tcp_urg_mode is overkill, only frames between 1955f8269a49SIlpo Järvinen * snd_up-64k-mss .. snd_up cannot be large. However, taking into 1956f8269a49SIlpo Järvinen * account rare use of URG, this is not a big flaw. 1957f8269a49SIlpo Järvinen * 1958a2a385d6SEric Dumazet * Returns true, if no segments are in flight and we have queued segments, 1959a2a385d6SEric Dumazet * but cannot send anything now because of SWS or another problem. 19601da177e4SLinus Torvalds */ 1961a2a385d6SEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 1962d5dd9175SIlpo Järvinen int push_one, gfp_t gfp) 19631da177e4SLinus Torvalds { 19641da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 196592df7b51SDavid S. Miller struct sk_buff *skb; 1966c1b4a7e6SDavid S. Miller unsigned int tso_segs, sent_pkts; 1967c1b4a7e6SDavid S. Miller int cwnd_quota; 19685d424d5aSJohn Heffner int result; 19691da177e4SLinus Torvalds 1970c1b4a7e6SDavid S. Miller sent_pkts = 0; 19715d424d5aSJohn Heffner 1972d5dd9175SIlpo Järvinen if (!push_one) { 19735d424d5aSJohn Heffner /* Do MTU probing. */ 1974d5dd9175SIlpo Järvinen result = tcp_mtu_probe(sk); 1975d5dd9175SIlpo Järvinen if (!result) { 1976a2a385d6SEric Dumazet return false; 19775d424d5aSJohn Heffner } else if (result > 0) { 19785d424d5aSJohn Heffner sent_pkts = 1; 19795d424d5aSJohn Heffner } 1980d5dd9175SIlpo Järvinen } 19815d424d5aSJohn Heffner 1982fe067e8aSDavid S. Miller while ((skb = tcp_send_head(sk))) { 1983c8ac3774SHerbert Xu unsigned int limit; 1984c8ac3774SHerbert Xu 198546d3ceabSEric Dumazet 1986b68e9f85SHerbert Xu tso_segs = tcp_init_tso_segs(sk, skb, mss_now); 1987c1b4a7e6SDavid S. Miller BUG_ON(!tso_segs); 1988c1b4a7e6SDavid S. Miller 1989ec342325SAndrew Vagin if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) 1990ec342325SAndrew Vagin goto repair; /* Skip network transmission */ 1991ec342325SAndrew Vagin 1992b68e9f85SHerbert Xu cwnd_quota = tcp_cwnd_test(tp, skb); 1993b68e9f85SHerbert Xu if (!cwnd_quota) 1994b68e9f85SHerbert Xu break; 1995b68e9f85SHerbert Xu 1996b68e9f85SHerbert Xu if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) 1997b68e9f85SHerbert Xu break; 1998b68e9f85SHerbert Xu 1999c1b4a7e6SDavid S. Miller if (tso_segs == 1) { 2000aa93466bSDavid S. Miller if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 2001aa93466bSDavid S. Miller (tcp_skb_is_last(sk, skb) ? 2002aa93466bSDavid S. Miller nonagle : TCP_NAGLE_PUSH)))) 2003aa93466bSDavid S. Miller break; 2004c1b4a7e6SDavid S. Miller } else { 2005d5dd9175SIlpo Järvinen if (!push_one && tcp_tso_should_defer(sk, skb)) 2006aa93466bSDavid S. Miller break; 2007c1b4a7e6SDavid S. Miller } 2008aa93466bSDavid S. Miller 200946d3ceabSEric Dumazet /* TSQ : sk_wmem_alloc accounts skb truesize, 201046d3ceabSEric Dumazet * including skb overhead. But thats OK. 201146d3ceabSEric Dumazet */ 201246d3ceabSEric Dumazet if (atomic_read(&sk->sk_wmem_alloc) >= sysctl_tcp_limit_output_bytes) { 201346d3ceabSEric Dumazet set_bit(TSQ_THROTTLED, &tp->tsq_flags); 201446d3ceabSEric Dumazet break; 201546d3ceabSEric Dumazet } 2016c8ac3774SHerbert Xu limit = mss_now; 2017f8269a49SIlpo Järvinen if (tso_segs > 1 && !tcp_urg_mode(tp)) 20180e3a4803SIlpo Järvinen limit = tcp_mss_split_point(sk, skb, mss_now, 20191485348dSBen Hutchings min_t(unsigned int, 20201485348dSBen Hutchings cwnd_quota, 20211485348dSBen Hutchings sk->sk_gso_max_segs)); 2022c8ac3774SHerbert Xu 2023c8ac3774SHerbert Xu if (skb->len > limit && 2024c4ead4c5SEric Dumazet unlikely(tso_fragment(sk, skb, limit, mss_now, gfp))) 20251da177e4SLinus Torvalds break; 20261da177e4SLinus Torvalds 20271da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2028c1b4a7e6SDavid S. Miller 2029d5dd9175SIlpo Järvinen if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) 20301da177e4SLinus Torvalds break; 20311da177e4SLinus Torvalds 2032ec342325SAndrew Vagin repair: 20331da177e4SLinus Torvalds /* Advance the send_head. This one is sent out. 20341da177e4SLinus Torvalds * This call will increment packets_out. 20351da177e4SLinus Torvalds */ 203666f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 20371da177e4SLinus Torvalds 20381da177e4SLinus Torvalds tcp_minshall_update(tp, mss_now, skb); 2039a262f0cdSNandita Dukkipati sent_pkts += tcp_skb_pcount(skb); 2040d5dd9175SIlpo Järvinen 2041d5dd9175SIlpo Järvinen if (push_one) 2042d5dd9175SIlpo Järvinen break; 20431da177e4SLinus Torvalds } 20441da177e4SLinus Torvalds 2045aa93466bSDavid S. Miller if (likely(sent_pkts)) { 2046684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 2047684bad11SYuchung Cheng tp->prr_out += sent_pkts; 20489e412ba7SIlpo Järvinen tcp_cwnd_validate(sk); 2049a2a385d6SEric Dumazet return false; 20501da177e4SLinus Torvalds } 2051fe067e8aSDavid S. Miller return !tp->packets_out && tcp_send_head(sk); 20521da177e4SLinus Torvalds } 20531da177e4SLinus Torvalds 2054a762a980SDavid S. Miller /* Push out any pending frames which were held back due to 2055a762a980SDavid S. Miller * TCP_CORK or attempt at coalescing tiny packets. 2056a762a980SDavid S. Miller * The socket must be locked by the caller. 2057a762a980SDavid S. Miller */ 20589e412ba7SIlpo Järvinen void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 20599e412ba7SIlpo Järvinen int nonagle) 2060a762a980SDavid S. Miller { 2061726e07a8SIlpo Järvinen /* If we are closed, the bytes will have to remain here. 2062726e07a8SIlpo Järvinen * In time closedown will finish, we empty the write queue and 2063726e07a8SIlpo Järvinen * all will be happy. 2064726e07a8SIlpo Järvinen */ 2065726e07a8SIlpo Järvinen if (unlikely(sk->sk_state == TCP_CLOSE)) 2066726e07a8SIlpo Järvinen return; 2067726e07a8SIlpo Järvinen 206899a1dec7SMel Gorman if (tcp_write_xmit(sk, cur_mss, nonagle, 0, 206999a1dec7SMel Gorman sk_gfp_atomic(sk, GFP_ATOMIC))) 20709e412ba7SIlpo Järvinen tcp_check_probe_timer(sk); 2071a762a980SDavid S. Miller } 2072a762a980SDavid S. Miller 2073c1b4a7e6SDavid S. Miller /* Send _single_ skb sitting at the send head. This function requires 2074c1b4a7e6SDavid S. Miller * true push pending frames to setup probe timer etc. 2075c1b4a7e6SDavid S. Miller */ 2076c1b4a7e6SDavid S. Miller void tcp_push_one(struct sock *sk, unsigned int mss_now) 2077c1b4a7e6SDavid S. Miller { 2078fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 2079c1b4a7e6SDavid S. Miller 2080c1b4a7e6SDavid S. Miller BUG_ON(!skb || skb->len < mss_now); 2081c1b4a7e6SDavid S. Miller 2082d5dd9175SIlpo Järvinen tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); 2083c1b4a7e6SDavid S. Miller } 2084c1b4a7e6SDavid S. Miller 20851da177e4SLinus Torvalds /* This function returns the amount that we can raise the 20861da177e4SLinus Torvalds * usable window based on the following constraints 20871da177e4SLinus Torvalds * 20881da177e4SLinus Torvalds * 1. The window can never be shrunk once it is offered (RFC 793) 20891da177e4SLinus Torvalds * 2. We limit memory per socket 20901da177e4SLinus Torvalds * 20911da177e4SLinus Torvalds * RFC 1122: 20921da177e4SLinus Torvalds * "the suggested [SWS] avoidance algorithm for the receiver is to keep 20931da177e4SLinus Torvalds * RECV.NEXT + RCV.WIN fixed until: 20941da177e4SLinus Torvalds * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 20951da177e4SLinus Torvalds * 20961da177e4SLinus Torvalds * i.e. don't raise the right edge of the window until you can raise 20971da177e4SLinus Torvalds * it at least MSS bytes. 20981da177e4SLinus Torvalds * 20991da177e4SLinus Torvalds * Unfortunately, the recommended algorithm breaks header prediction, 21001da177e4SLinus Torvalds * since header prediction assumes th->window stays fixed. 21011da177e4SLinus Torvalds * 21021da177e4SLinus Torvalds * Strictly speaking, keeping th->window fixed violates the receiver 21031da177e4SLinus Torvalds * side SWS prevention criteria. The problem is that under this rule 21041da177e4SLinus Torvalds * a stream of single byte packets will cause the right side of the 21051da177e4SLinus Torvalds * window to always advance by a single byte. 21061da177e4SLinus Torvalds * 21071da177e4SLinus Torvalds * Of course, if the sender implements sender side SWS prevention 21081da177e4SLinus Torvalds * then this will not be a problem. 21091da177e4SLinus Torvalds * 21101da177e4SLinus Torvalds * BSD seems to make the following compromise: 21111da177e4SLinus Torvalds * 21121da177e4SLinus Torvalds * If the free space is less than the 1/4 of the maximum 21131da177e4SLinus Torvalds * space available and the free space is less than 1/2 mss, 21141da177e4SLinus Torvalds * then set the window to 0. 21151da177e4SLinus Torvalds * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 21161da177e4SLinus Torvalds * Otherwise, just prevent the window from shrinking 21171da177e4SLinus Torvalds * and from being larger than the largest representable value. 21181da177e4SLinus Torvalds * 21191da177e4SLinus Torvalds * This prevents incremental opening of the window in the regime 21201da177e4SLinus Torvalds * where TCP is limited by the speed of the reader side taking 21211da177e4SLinus Torvalds * data out of the TCP receive queue. It does nothing about 21221da177e4SLinus Torvalds * those cases where the window is constrained on the sender side 21231da177e4SLinus Torvalds * because the pipeline is full. 21241da177e4SLinus Torvalds * 21251da177e4SLinus Torvalds * BSD also seems to "accidentally" limit itself to windows that are a 21261da177e4SLinus Torvalds * multiple of MSS, at least until the free space gets quite small. 21271da177e4SLinus Torvalds * This would appear to be a side effect of the mbuf implementation. 21281da177e4SLinus Torvalds * Combining these two algorithms results in the observed behavior 21291da177e4SLinus Torvalds * of having a fixed window size at almost all times. 21301da177e4SLinus Torvalds * 21311da177e4SLinus Torvalds * Below we obtain similar behavior by forcing the offered window to 21321da177e4SLinus Torvalds * a multiple of the mss when it is feasible to do so. 21331da177e4SLinus Torvalds * 21341da177e4SLinus Torvalds * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 21351da177e4SLinus Torvalds * Regular options like TIMESTAMP are taken into account. 21361da177e4SLinus Torvalds */ 21371da177e4SLinus Torvalds u32 __tcp_select_window(struct sock *sk) 21381da177e4SLinus Torvalds { 2139463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 21401da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2141caa20d9aSStephen Hemminger /* MSS for the peer's data. Previous versions used mss_clamp 21421da177e4SLinus Torvalds * here. I don't know if the value based on our guesses 21431da177e4SLinus Torvalds * of peer's MSS is better for the performance. It's more correct 21441da177e4SLinus Torvalds * but may be worse for the performance because of rcv_mss 21451da177e4SLinus Torvalds * fluctuations. --SAW 1998/11/1 21461da177e4SLinus Torvalds */ 2147463c84b9SArnaldo Carvalho de Melo int mss = icsk->icsk_ack.rcv_mss; 21481da177e4SLinus Torvalds int free_space = tcp_space(sk); 21491da177e4SLinus Torvalds int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk)); 21501da177e4SLinus Torvalds int window; 21511da177e4SLinus Torvalds 21521da177e4SLinus Torvalds if (mss > full_space) 21531da177e4SLinus Torvalds mss = full_space; 21541da177e4SLinus Torvalds 2155b92edbe0SEric Dumazet if (free_space < (full_space >> 1)) { 2156463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.quick = 0; 21571da177e4SLinus Torvalds 2158180d8cd9SGlauber Costa if (sk_under_memory_pressure(sk)) 2159056834d9SIlpo Järvinen tp->rcv_ssthresh = min(tp->rcv_ssthresh, 2160056834d9SIlpo Järvinen 4U * tp->advmss); 21611da177e4SLinus Torvalds 21621da177e4SLinus Torvalds if (free_space < mss) 21631da177e4SLinus Torvalds return 0; 21641da177e4SLinus Torvalds } 21651da177e4SLinus Torvalds 21661da177e4SLinus Torvalds if (free_space > tp->rcv_ssthresh) 21671da177e4SLinus Torvalds free_space = tp->rcv_ssthresh; 21681da177e4SLinus Torvalds 21691da177e4SLinus Torvalds /* Don't do rounding if we are using window scaling, since the 21701da177e4SLinus Torvalds * scaled window will not line up with the MSS boundary anyway. 21711da177e4SLinus Torvalds */ 21721da177e4SLinus Torvalds window = tp->rcv_wnd; 21731da177e4SLinus Torvalds if (tp->rx_opt.rcv_wscale) { 21741da177e4SLinus Torvalds window = free_space; 21751da177e4SLinus Torvalds 21761da177e4SLinus Torvalds /* Advertise enough space so that it won't get scaled away. 21771da177e4SLinus Torvalds * Import case: prevent zero window announcement if 21781da177e4SLinus Torvalds * 1<<rcv_wscale > mss. 21791da177e4SLinus Torvalds */ 21801da177e4SLinus Torvalds if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window) 21811da177e4SLinus Torvalds window = (((window >> tp->rx_opt.rcv_wscale) + 1) 21821da177e4SLinus Torvalds << tp->rx_opt.rcv_wscale); 21831da177e4SLinus Torvalds } else { 21841da177e4SLinus Torvalds /* Get the largest window that is a nice multiple of mss. 21851da177e4SLinus Torvalds * Window clamp already applied above. 21861da177e4SLinus Torvalds * If our current window offering is within 1 mss of the 21871da177e4SLinus Torvalds * free space we just keep it. This prevents the divide 21881da177e4SLinus Torvalds * and multiply from happening most of the time. 21891da177e4SLinus Torvalds * We also don't do any window rounding when the free space 21901da177e4SLinus Torvalds * is too small. 21911da177e4SLinus Torvalds */ 21921da177e4SLinus Torvalds if (window <= free_space - mss || window > free_space) 21931da177e4SLinus Torvalds window = (free_space / mss) * mss; 219484565070SJohn Heffner else if (mss == full_space && 2195b92edbe0SEric Dumazet free_space > window + (full_space >> 1)) 219684565070SJohn Heffner window = free_space; 21971da177e4SLinus Torvalds } 21981da177e4SLinus Torvalds 21991da177e4SLinus Torvalds return window; 22001da177e4SLinus Torvalds } 22011da177e4SLinus Torvalds 22024a17fc3aSIlpo Järvinen /* Collapses two adjacent SKB's during retransmission. */ 22034a17fc3aSIlpo Järvinen static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) 22041da177e4SLinus Torvalds { 22051da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2206fe067e8aSDavid S. Miller struct sk_buff *next_skb = tcp_write_queue_next(sk, skb); 2207058dc334SIlpo Järvinen int skb_size, next_skb_size; 22081da177e4SLinus Torvalds 2209058dc334SIlpo Järvinen skb_size = skb->len; 2210058dc334SIlpo Järvinen next_skb_size = next_skb->len; 22111da177e4SLinus Torvalds 2212058dc334SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); 22131da177e4SLinus Torvalds 22146859d494SIlpo Järvinen tcp_highest_sack_combine(sk, next_skb, skb); 2215a6963a6bSIlpo Järvinen 2216fe067e8aSDavid S. Miller tcp_unlink_write_queue(next_skb, sk); 22171da177e4SLinus Torvalds 2218058dc334SIlpo Järvinen skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size), 22191a4e2d09SArnaldo Carvalho de Melo next_skb_size); 22201da177e4SLinus Torvalds 222152d570aaSJarek Poplawski if (next_skb->ip_summed == CHECKSUM_PARTIAL) 222252d570aaSJarek Poplawski skb->ip_summed = CHECKSUM_PARTIAL; 22231da177e4SLinus Torvalds 222484fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 22251da177e4SLinus Torvalds skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); 22261da177e4SLinus Torvalds 22271da177e4SLinus Torvalds /* Update sequence range on original skb. */ 22281da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 22291da177e4SLinus Torvalds 2230e6c7d085SIlpo Järvinen /* Merge over control information. This moves PSH/FIN etc. over */ 22314de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; 22321da177e4SLinus Torvalds 22331da177e4SLinus Torvalds /* All done, get rid of second SKB and account for it so 22341da177e4SLinus Torvalds * packet counting does not break. 22351da177e4SLinus Torvalds */ 22364828e7f4SIlpo Järvinen TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; 2237b7689205SIlpo Järvinen 2238b7689205SIlpo Järvinen /* changed transmit queue under us so clear hints */ 2239ef9da47cSIlpo Järvinen tcp_clear_retrans_hints_partial(tp); 2240ef9da47cSIlpo Järvinen if (next_skb == tp->retransmit_skb_hint) 2241ef9da47cSIlpo Järvinen tp->retransmit_skb_hint = skb; 2242b7689205SIlpo Järvinen 2243797108d1SIlpo Järvinen tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); 2244797108d1SIlpo Järvinen 22453ab224beSHideo Aoki sk_wmem_free_skb(sk, next_skb); 22461da177e4SLinus Torvalds } 22471da177e4SLinus Torvalds 224867edfef7SAndi Kleen /* Check if coalescing SKBs is legal. */ 2249a2a385d6SEric Dumazet static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) 22504a17fc3aSIlpo Järvinen { 22514a17fc3aSIlpo Järvinen if (tcp_skb_pcount(skb) > 1) 2252a2a385d6SEric Dumazet return false; 22534a17fc3aSIlpo Järvinen /* TODO: SACK collapsing could be used to remove this condition */ 22544a17fc3aSIlpo Järvinen if (skb_shinfo(skb)->nr_frags != 0) 2255a2a385d6SEric Dumazet return false; 22564a17fc3aSIlpo Järvinen if (skb_cloned(skb)) 2257a2a385d6SEric Dumazet return false; 22584a17fc3aSIlpo Järvinen if (skb == tcp_send_head(sk)) 2259a2a385d6SEric Dumazet return false; 22604a17fc3aSIlpo Järvinen /* Some heurestics for collapsing over SACK'd could be invented */ 22614a17fc3aSIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 2262a2a385d6SEric Dumazet return false; 22634a17fc3aSIlpo Järvinen 2264a2a385d6SEric Dumazet return true; 22654a17fc3aSIlpo Järvinen } 22664a17fc3aSIlpo Järvinen 226767edfef7SAndi Kleen /* Collapse packets in the retransmit queue to make to create 226867edfef7SAndi Kleen * less packets on the wire. This is only done on retransmission. 226967edfef7SAndi Kleen */ 22704a17fc3aSIlpo Järvinen static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, 22714a17fc3aSIlpo Järvinen int space) 22724a17fc3aSIlpo Järvinen { 22734a17fc3aSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 22744a17fc3aSIlpo Järvinen struct sk_buff *skb = to, *tmp; 2275a2a385d6SEric Dumazet bool first = true; 22764a17fc3aSIlpo Järvinen 22774a17fc3aSIlpo Järvinen if (!sysctl_tcp_retrans_collapse) 22784a17fc3aSIlpo Järvinen return; 22794de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 22804a17fc3aSIlpo Järvinen return; 22814a17fc3aSIlpo Järvinen 22824a17fc3aSIlpo Järvinen tcp_for_write_queue_from_safe(skb, tmp, sk) { 22834a17fc3aSIlpo Järvinen if (!tcp_can_collapse(sk, skb)) 22844a17fc3aSIlpo Järvinen break; 22854a17fc3aSIlpo Järvinen 22864a17fc3aSIlpo Järvinen space -= skb->len; 22874a17fc3aSIlpo Järvinen 22884a17fc3aSIlpo Järvinen if (first) { 2289a2a385d6SEric Dumazet first = false; 22904a17fc3aSIlpo Järvinen continue; 22914a17fc3aSIlpo Järvinen } 22924a17fc3aSIlpo Järvinen 22934a17fc3aSIlpo Järvinen if (space < 0) 22944a17fc3aSIlpo Järvinen break; 22954a17fc3aSIlpo Järvinen /* Punt if not enough space exists in the first SKB for 22964a17fc3aSIlpo Järvinen * the data in the second 22974a17fc3aSIlpo Järvinen */ 2298a21d4572SEric Dumazet if (skb->len > skb_availroom(to)) 22994a17fc3aSIlpo Järvinen break; 23004a17fc3aSIlpo Järvinen 23014a17fc3aSIlpo Järvinen if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) 23024a17fc3aSIlpo Järvinen break; 23034a17fc3aSIlpo Järvinen 23044a17fc3aSIlpo Järvinen tcp_collapse_retrans(sk, to); 23054a17fc3aSIlpo Järvinen } 23064a17fc3aSIlpo Järvinen } 23074a17fc3aSIlpo Järvinen 23081da177e4SLinus Torvalds /* This retransmits one SKB. Policy decisions and retransmit queue 23091da177e4SLinus Torvalds * state updates are done by the caller. Returns non-zero if an 23101da177e4SLinus Torvalds * error occurred which prevented the send. 23111da177e4SLinus Torvalds */ 23121da177e4SLinus Torvalds int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) 23131da177e4SLinus Torvalds { 23141da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 23155d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 23167d227cd2SSridhar Samudrala unsigned int cur_mss; 23171da177e4SLinus Torvalds int err; 23181da177e4SLinus Torvalds 23195d424d5aSJohn Heffner /* Inconslusive MTU probe */ 23205d424d5aSJohn Heffner if (icsk->icsk_mtup.probe_size) { 23215d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 23225d424d5aSJohn Heffner } 23235d424d5aSJohn Heffner 23241da177e4SLinus Torvalds /* Do not sent more than we queued. 1/4 is reserved for possible 2325caa20d9aSStephen Hemminger * copying overhead: fragmentation, tunneling, mangling etc. 23261da177e4SLinus Torvalds */ 23271da177e4SLinus Torvalds if (atomic_read(&sk->sk_wmem_alloc) > 23281da177e4SLinus Torvalds min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf)) 23291da177e4SLinus Torvalds return -EAGAIN; 23301da177e4SLinus Torvalds 23311da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 23321da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 23331da177e4SLinus Torvalds BUG(); 23341da177e4SLinus Torvalds if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 23351da177e4SLinus Torvalds return -ENOMEM; 23361da177e4SLinus Torvalds } 23371da177e4SLinus Torvalds 23387d227cd2SSridhar Samudrala if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 23397d227cd2SSridhar Samudrala return -EHOSTUNREACH; /* Routing failure or similar. */ 23407d227cd2SSridhar Samudrala 23410c54b85fSIlpo Järvinen cur_mss = tcp_current_mss(sk); 23427d227cd2SSridhar Samudrala 23431da177e4SLinus Torvalds /* If receiver has shrunk his window, and skb is out of 23441da177e4SLinus Torvalds * new window, do not retransmit it. The exception is the 23451da177e4SLinus Torvalds * case, when window is shrunk to zero. In this case 23461da177e4SLinus Torvalds * our retransmit serves as a zero window probe. 23471da177e4SLinus Torvalds */ 23489d4fb27dSJoe Perches if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && 23499d4fb27dSJoe Perches TCP_SKB_CB(skb)->seq != tp->snd_una) 23501da177e4SLinus Torvalds return -EAGAIN; 23511da177e4SLinus Torvalds 23521da177e4SLinus Torvalds if (skb->len > cur_mss) { 2353846998aeSDavid S. Miller if (tcp_fragment(sk, skb, cur_mss, cur_mss)) 23541da177e4SLinus Torvalds return -ENOMEM; /* We'll try again later. */ 235502276f3cSIlpo Järvinen } else { 23569eb9362eSIlpo Järvinen int oldpcount = tcp_skb_pcount(skb); 23579eb9362eSIlpo Järvinen 23589eb9362eSIlpo Järvinen if (unlikely(oldpcount > 1)) { 235902276f3cSIlpo Järvinen tcp_init_tso_segs(sk, skb, cur_mss); 23609eb9362eSIlpo Järvinen tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb)); 23619eb9362eSIlpo Järvinen } 23621da177e4SLinus Torvalds } 23631da177e4SLinus Torvalds 23641da177e4SLinus Torvalds tcp_retrans_try_collapse(sk, skb, cur_mss); 23651da177e4SLinus Torvalds 23661da177e4SLinus Torvalds /* Some Solaris stacks overoptimize and ignore the FIN on a 23671da177e4SLinus Torvalds * retransmit when old data is attached. So strip it off 23681da177e4SLinus Torvalds * since it is cheap to do so and saves bytes on the network. 23691da177e4SLinus Torvalds */ 23701da177e4SLinus Torvalds if (skb->len > 0 && 23714de075e0SEric Dumazet (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && 23721da177e4SLinus Torvalds tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) { 23731da177e4SLinus Torvalds if (!pskb_trim(skb, 0)) { 2374e870a8efSIlpo Järvinen /* Reuse, even though it does some unnecessary work */ 2375e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1, 23764de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags); 23771da177e4SLinus Torvalds skb->ip_summed = CHECKSUM_NONE; 23781da177e4SLinus Torvalds } 23791da177e4SLinus Torvalds } 23801da177e4SLinus Torvalds 23811da177e4SLinus Torvalds /* Make a copy, if the first transmission SKB clone we made 23821da177e4SLinus Torvalds * is still in somebody's hands, else make a clone. 23831da177e4SLinus Torvalds */ 23841da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 23851da177e4SLinus Torvalds 2386117632e6SEric Dumazet /* make sure skb->data is aligned on arches that require it */ 2387117632e6SEric Dumazet if (unlikely(NET_IP_ALIGN && ((unsigned long)skb->data & 3))) { 2388117632e6SEric Dumazet struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER, 2389117632e6SEric Dumazet GFP_ATOMIC); 2390117632e6SEric Dumazet err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) : 2391117632e6SEric Dumazet -ENOBUFS; 2392117632e6SEric Dumazet } else { 2393dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 2394117632e6SEric Dumazet } 23951da177e4SLinus Torvalds 23961da177e4SLinus Torvalds if (err == 0) { 23971da177e4SLinus Torvalds /* Update global TCP statistics. */ 239881cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); 23991da177e4SLinus Torvalds 24001da177e4SLinus Torvalds tp->total_retrans++; 24011da177e4SLinus Torvalds 24021da177e4SLinus Torvalds #if FASTRETRANS_DEBUG > 0 24031da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 2404e87cc472SJoe Perches net_dbg_ratelimited("retrans_out leaked\n"); 24051da177e4SLinus Torvalds } 24061da177e4SLinus Torvalds #endif 2407b08d6cb2SIlpo Järvinen if (!tp->retrans_out) 2408b08d6cb2SIlpo Järvinen tp->lost_retrans_low = tp->snd_nxt; 24091da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 24101da177e4SLinus Torvalds tp->retrans_out += tcp_skb_pcount(skb); 24111da177e4SLinus Torvalds 24121da177e4SLinus Torvalds /* Save stamp of the first retransmit. */ 24131da177e4SLinus Torvalds if (!tp->retrans_stamp) 24141da177e4SLinus Torvalds tp->retrans_stamp = TCP_SKB_CB(skb)->when; 24151da177e4SLinus Torvalds 2416c24f691bSYuchung Cheng tp->undo_retrans += tcp_skb_pcount(skb); 24171da177e4SLinus Torvalds 24181da177e4SLinus Torvalds /* snd_nxt is stored to detect loss of retransmitted segment, 24191da177e4SLinus Torvalds * see tcp_input.c tcp_sacktag_write_queue(). 24201da177e4SLinus Torvalds */ 24211da177e4SLinus Torvalds TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt; 24221da177e4SLinus Torvalds } 24231da177e4SLinus Torvalds return err; 24241da177e4SLinus Torvalds } 24251da177e4SLinus Torvalds 242667edfef7SAndi Kleen /* Check if we forward retransmits are possible in the current 242767edfef7SAndi Kleen * window/congestion state. 242867edfef7SAndi Kleen */ 2429a2a385d6SEric Dumazet static bool tcp_can_forward_retransmit(struct sock *sk) 2430b5afe7bcSIlpo Järvinen { 2431b5afe7bcSIlpo Järvinen const struct inet_connection_sock *icsk = inet_csk(sk); 2432cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 2433b5afe7bcSIlpo Järvinen 2434b5afe7bcSIlpo Järvinen /* Forward retransmissions are possible only during Recovery. */ 2435b5afe7bcSIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Recovery) 2436a2a385d6SEric Dumazet return false; 2437b5afe7bcSIlpo Järvinen 2438b5afe7bcSIlpo Järvinen /* No forward retransmissions in Reno are possible. */ 2439b5afe7bcSIlpo Järvinen if (tcp_is_reno(tp)) 2440a2a385d6SEric Dumazet return false; 2441b5afe7bcSIlpo Järvinen 2442b5afe7bcSIlpo Järvinen /* Yeah, we have to make difficult choice between forward transmission 2443b5afe7bcSIlpo Järvinen * and retransmission... Both ways have their merits... 2444b5afe7bcSIlpo Järvinen * 2445b5afe7bcSIlpo Järvinen * For now we do not retransmit anything, while we have some new 2446b5afe7bcSIlpo Järvinen * segments to send. In the other cases, follow rule 3 for 2447b5afe7bcSIlpo Järvinen * NextSeg() specified in RFC3517. 2448b5afe7bcSIlpo Järvinen */ 2449b5afe7bcSIlpo Järvinen 2450b5afe7bcSIlpo Järvinen if (tcp_may_send_now(sk)) 2451a2a385d6SEric Dumazet return false; 2452b5afe7bcSIlpo Järvinen 2453a2a385d6SEric Dumazet return true; 2454b5afe7bcSIlpo Järvinen } 2455b5afe7bcSIlpo Järvinen 24561da177e4SLinus Torvalds /* This gets called after a retransmit timeout, and the initially 24571da177e4SLinus Torvalds * retransmitted data is acknowledged. It tries to continue 24581da177e4SLinus Torvalds * resending the rest of the retransmit queue, until either 24591da177e4SLinus Torvalds * we've sent it all or the congestion window limit is reached. 24601da177e4SLinus Torvalds * If doing SACK, the first ACK which comes back for a timeout 24611da177e4SLinus Torvalds * based retransmit packet might feed us FACK information again. 24621da177e4SLinus Torvalds * If so, we use it to avoid unnecessarily retransmissions. 24631da177e4SLinus Torvalds */ 24641da177e4SLinus Torvalds void tcp_xmit_retransmit_queue(struct sock *sk) 24651da177e4SLinus Torvalds { 24666687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 24671da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 24681da177e4SLinus Torvalds struct sk_buff *skb; 24690e1c54c2SIlpo Järvinen struct sk_buff *hole = NULL; 2470618d9f25SIlpo Järvinen u32 last_lost; 247161eb55f4SIlpo Järvinen int mib_idx; 24720e1c54c2SIlpo Järvinen int fwd_rexmitting = 0; 24736a438bbeSStephen Hemminger 247445e77d31SIlpo Järvinen if (!tp->packets_out) 247545e77d31SIlpo Järvinen return; 247645e77d31SIlpo Järvinen 247708ebd172SIlpo Järvinen if (!tp->lost_out) 247808ebd172SIlpo Järvinen tp->retransmit_high = tp->snd_una; 247908ebd172SIlpo Järvinen 2480618d9f25SIlpo Järvinen if (tp->retransmit_skb_hint) { 24816a438bbeSStephen Hemminger skb = tp->retransmit_skb_hint; 2482618d9f25SIlpo Järvinen last_lost = TCP_SKB_CB(skb)->end_seq; 2483618d9f25SIlpo Järvinen if (after(last_lost, tp->retransmit_high)) 2484618d9f25SIlpo Järvinen last_lost = tp->retransmit_high; 2485618d9f25SIlpo Järvinen } else { 2486fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 2487618d9f25SIlpo Järvinen last_lost = tp->snd_una; 2488618d9f25SIlpo Järvinen } 24891da177e4SLinus Torvalds 2490fe067e8aSDavid S. Miller tcp_for_write_queue_from(skb, sk) { 24911da177e4SLinus Torvalds __u8 sacked = TCP_SKB_CB(skb)->sacked; 24921da177e4SLinus Torvalds 2493fe067e8aSDavid S. Miller if (skb == tcp_send_head(sk)) 2494fe067e8aSDavid S. Miller break; 24956a438bbeSStephen Hemminger /* we could do better than to assign each time */ 24960e1c54c2SIlpo Järvinen if (hole == NULL) 24976a438bbeSStephen Hemminger tp->retransmit_skb_hint = skb; 24986a438bbeSStephen Hemminger 24991da177e4SLinus Torvalds /* Assume this retransmit will generate 25001da177e4SLinus Torvalds * only one packet for congestion window 25011da177e4SLinus Torvalds * calculation purposes. This works because 25021da177e4SLinus Torvalds * tcp_retransmit_skb() will chop up the 25031da177e4SLinus Torvalds * packet to be MSS sized and all the 25041da177e4SLinus Torvalds * packet counting works out. 25051da177e4SLinus Torvalds */ 25061da177e4SLinus Torvalds if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) 25071da177e4SLinus Torvalds return; 25080e1c54c2SIlpo Järvinen 25090e1c54c2SIlpo Järvinen if (fwd_rexmitting) { 25100e1c54c2SIlpo Järvinen begin_fwd: 25110e1c54c2SIlpo Järvinen if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp))) 2512006f582cSIlpo Järvinen break; 25130e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFORWARDRETRANS; 25140e1c54c2SIlpo Järvinen 25150e1c54c2SIlpo Järvinen } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) { 2516618d9f25SIlpo Järvinen tp->retransmit_high = last_lost; 25170e1c54c2SIlpo Järvinen if (!tcp_can_forward_retransmit(sk)) 25180e1c54c2SIlpo Järvinen break; 25190e1c54c2SIlpo Järvinen /* Backtrack if necessary to non-L'ed skb */ 25200e1c54c2SIlpo Järvinen if (hole != NULL) { 25210e1c54c2SIlpo Järvinen skb = hole; 25220e1c54c2SIlpo Järvinen hole = NULL; 25230e1c54c2SIlpo Järvinen } 25240e1c54c2SIlpo Järvinen fwd_rexmitting = 1; 25250e1c54c2SIlpo Järvinen goto begin_fwd; 25260e1c54c2SIlpo Järvinen 25270e1c54c2SIlpo Järvinen } else if (!(sacked & TCPCB_LOST)) { 2528ac11ba75SIlpo Järvinen if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) 25290e1c54c2SIlpo Järvinen hole = skb; 253061eb55f4SIlpo Järvinen continue; 25311da177e4SLinus Torvalds 25320e1c54c2SIlpo Järvinen } else { 2533618d9f25SIlpo Järvinen last_lost = TCP_SKB_CB(skb)->end_seq; 25340e1c54c2SIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Loss) 25350e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFASTRETRANS; 25360e1c54c2SIlpo Järvinen else 25370e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; 25380e1c54c2SIlpo Järvinen } 25390e1c54c2SIlpo Järvinen 25400e1c54c2SIlpo Järvinen if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) 254161eb55f4SIlpo Järvinen continue; 254240b215e5SPavel Emelyanov 254309e9b813SEric Dumazet if (tcp_retransmit_skb(sk, skb)) { 254409e9b813SEric Dumazet NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL); 25451da177e4SLinus Torvalds return; 254609e9b813SEric Dumazet } 2547de0744afSPavel Emelyanov NET_INC_STATS_BH(sock_net(sk), mib_idx); 25481da177e4SLinus Torvalds 2549684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 2550a262f0cdSNandita Dukkipati tp->prr_out += tcp_skb_pcount(skb); 2551a262f0cdSNandita Dukkipati 2552fe067e8aSDavid S. Miller if (skb == tcp_write_queue_head(sk)) 2553463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 25543f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, 25553f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 25561da177e4SLinus Torvalds } 25571da177e4SLinus Torvalds } 25581da177e4SLinus Torvalds 25591da177e4SLinus Torvalds /* Send a fin. The caller locks the socket for us. This cannot be 25601da177e4SLinus Torvalds * allowed to fail queueing a FIN frame under any circumstances. 25611da177e4SLinus Torvalds */ 25621da177e4SLinus Torvalds void tcp_send_fin(struct sock *sk) 25631da177e4SLinus Torvalds { 25641da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2565fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_write_queue_tail(sk); 25661da177e4SLinus Torvalds int mss_now; 25671da177e4SLinus Torvalds 25681da177e4SLinus Torvalds /* Optimization, tack on the FIN if we have a queue of 25691da177e4SLinus Torvalds * unsent frames. But be careful about outgoing SACKS 25701da177e4SLinus Torvalds * and IP options. 25711da177e4SLinus Torvalds */ 25720c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 25731da177e4SLinus Torvalds 2574fe067e8aSDavid S. Miller if (tcp_send_head(sk) != NULL) { 25754de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_FIN; 25761da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq++; 25771da177e4SLinus Torvalds tp->write_seq++; 25781da177e4SLinus Torvalds } else { 25791da177e4SLinus Torvalds /* Socket is locked, keep trying until memory is available. */ 25801da177e4SLinus Torvalds for (;;) { 2581aa133076SWu Fengguang skb = alloc_skb_fclone(MAX_TCP_HEADER, 2582aa133076SWu Fengguang sk->sk_allocation); 25831da177e4SLinus Torvalds if (skb) 25841da177e4SLinus Torvalds break; 25851da177e4SLinus Torvalds yield(); 25861da177e4SLinus Torvalds } 25871da177e4SLinus Torvalds 25881da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 25891da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 25901da177e4SLinus Torvalds /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 2591e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tp->write_seq, 2592a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_FIN); 25931da177e4SLinus Torvalds tcp_queue_skb(sk, skb); 25941da177e4SLinus Torvalds } 25959e412ba7SIlpo Järvinen __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF); 25961da177e4SLinus Torvalds } 25971da177e4SLinus Torvalds 25981da177e4SLinus Torvalds /* We get here when a process closes a file descriptor (either due to 25991da177e4SLinus Torvalds * an explicit close() or as a byproduct of exit()'ing) and there 26001da177e4SLinus Torvalds * was unread data in the receive queue. This behavior is recommended 260165bb723cSGerrit Renker * by RFC 2525, section 2.17. -DaveM 26021da177e4SLinus Torvalds */ 2603dd0fc66fSAl Viro void tcp_send_active_reset(struct sock *sk, gfp_t priority) 26041da177e4SLinus Torvalds { 26051da177e4SLinus Torvalds struct sk_buff *skb; 26061da177e4SLinus Torvalds 26071da177e4SLinus Torvalds /* NOTE: No TCP options attached and we never retransmit this. */ 26081da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, priority); 26091da177e4SLinus Torvalds if (!skb) { 26104e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 26111da177e4SLinus Torvalds return; 26121da177e4SLinus Torvalds } 26131da177e4SLinus Torvalds 26141da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 26151da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 2616e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), 2617a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_RST); 26181da177e4SLinus Torvalds /* Send it off. */ 26191da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2620dfb4b9dcSDavid S. Miller if (tcp_transmit_skb(sk, skb, 0, priority)) 26214e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 262226af65cbSSridhar Samudrala 262381cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); 26241da177e4SLinus Torvalds } 26251da177e4SLinus Torvalds 262667edfef7SAndi Kleen /* Send a crossed SYN-ACK during socket establishment. 262767edfef7SAndi Kleen * WARNING: This routine must only be called when we have already sent 26281da177e4SLinus Torvalds * a SYN packet that crossed the incoming SYN that caused this routine 26291da177e4SLinus Torvalds * to get called. If this assumption fails then the initial rcv_wnd 26301da177e4SLinus Torvalds * and rcv_wscale values will not be correct. 26311da177e4SLinus Torvalds */ 26321da177e4SLinus Torvalds int tcp_send_synack(struct sock *sk) 26331da177e4SLinus Torvalds { 26341da177e4SLinus Torvalds struct sk_buff *skb; 26351da177e4SLinus Torvalds 2636fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 26374de075e0SEric Dumazet if (skb == NULL || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 263891df42beSJoe Perches pr_debug("%s: wrong queue state\n", __func__); 26391da177e4SLinus Torvalds return -EFAULT; 26401da177e4SLinus Torvalds } 26414de075e0SEric Dumazet if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { 26421da177e4SLinus Torvalds if (skb_cloned(skb)) { 26431da177e4SLinus Torvalds struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC); 26441da177e4SLinus Torvalds if (nskb == NULL) 26451da177e4SLinus Torvalds return -ENOMEM; 2646fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 26471da177e4SLinus Torvalds skb_header_release(nskb); 2648fe067e8aSDavid S. Miller __tcp_add_write_queue_head(sk, nskb); 26493ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 26503ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 26513ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 26521da177e4SLinus Torvalds skb = nskb; 26531da177e4SLinus Torvalds } 26541da177e4SLinus Torvalds 26554de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; 26561da177e4SLinus Torvalds TCP_ECN_send_synack(tcp_sk(sk), skb); 26571da177e4SLinus Torvalds } 26581da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2659dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 26601da177e4SLinus Torvalds } 26611da177e4SLinus Torvalds 26624aea39c1SEric Dumazet /** 26634aea39c1SEric Dumazet * tcp_make_synack - Prepare a SYN-ACK. 26644aea39c1SEric Dumazet * sk: listener socket 26654aea39c1SEric Dumazet * dst: dst entry attached to the SYNACK 26664aea39c1SEric Dumazet * req: request_sock pointer 26674aea39c1SEric Dumazet * rvp: request_values pointer 26684aea39c1SEric Dumazet * 26694aea39c1SEric Dumazet * Allocate one skb and build a SYNACK packet. 26704aea39c1SEric Dumazet * @dst is consumed : Caller should not use it again. 26714aea39c1SEric Dumazet */ 26721da177e4SLinus Torvalds struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst, 2673e6b4d113SWilliam Allen Simpson struct request_sock *req, 26748336886fSJerry Chu struct request_values *rvp, 26758336886fSJerry Chu struct tcp_fastopen_cookie *foc) 26761da177e4SLinus Torvalds { 2677bd0388aeSWilliam Allen Simpson struct tcp_out_options opts; 26784957faadSWilliam Allen Simpson struct tcp_extend_values *xvp = tcp_xv(rvp); 26792e6599cbSArnaldo Carvalho de Melo struct inet_request_sock *ireq = inet_rsk(req); 26801da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 268128b2774aSEric Dumazet const struct tcp_cookie_values *cvp = tp->cookie_values; 26821da177e4SLinus Torvalds struct tcphdr *th; 26831da177e4SLinus Torvalds struct sk_buff *skb; 2684cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 2685bd0388aeSWilliam Allen Simpson int tcp_header_size; 2686f5fff5dcSTom Quetchenbach int mss; 268728b2774aSEric Dumazet int s_data_desired = 0; 26881da177e4SLinus Torvalds 268928b2774aSEric Dumazet if (cvp != NULL && cvp->s_data_constant && cvp->s_data_desired) 269028b2774aSEric Dumazet s_data_desired = cvp->s_data_desired; 269199a1dec7SMel Gorman skb = alloc_skb(MAX_TCP_HEADER + 15 + s_data_desired, 269299a1dec7SMel Gorman sk_gfp_atomic(sk, GFP_ATOMIC)); 26934aea39c1SEric Dumazet if (unlikely(!skb)) { 26944aea39c1SEric Dumazet dst_release(dst); 26951da177e4SLinus Torvalds return NULL; 26964aea39c1SEric Dumazet } 26971da177e4SLinus Torvalds /* Reserve space for headers. */ 26981da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 26991da177e4SLinus Torvalds 27004aea39c1SEric Dumazet skb_dst_set(skb, dst); 27011da177e4SLinus Torvalds 27020dbaee3bSDavid S. Miller mss = dst_metric_advmss(dst); 2703f5fff5dcSTom Quetchenbach if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss) 2704f5fff5dcSTom Quetchenbach mss = tp->rx_opt.user_mss; 2705f5fff5dcSTom Quetchenbach 270633ad798cSAdam Langley if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */ 270733ad798cSAdam Langley __u8 rcv_wscale; 270833ad798cSAdam Langley /* Set this up on the first call only */ 270933ad798cSAdam Langley req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW); 2710e88c64f0SHagen Paul Pfeifer 2711e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 2712e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 2713e88c64f0SHagen Paul Pfeifer (req->window_clamp > tcp_full_space(sk) || req->window_clamp == 0)) 2714e88c64f0SHagen Paul Pfeifer req->window_clamp = tcp_full_space(sk); 2715e88c64f0SHagen Paul Pfeifer 271633ad798cSAdam Langley /* tcp_full_space because it is guaranteed to be the first packet */ 271733ad798cSAdam Langley tcp_select_initial_window(tcp_full_space(sk), 2718f5fff5dcSTom Quetchenbach mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0), 271933ad798cSAdam Langley &req->rcv_wnd, 272033ad798cSAdam Langley &req->window_clamp, 272133ad798cSAdam Langley ireq->wscale_ok, 272231d12926Slaurent chavey &rcv_wscale, 272331d12926Slaurent chavey dst_metric(dst, RTAX_INITRWND)); 272433ad798cSAdam Langley ireq->rcv_wscale = rcv_wscale; 272533ad798cSAdam Langley } 2726cfb6eeb4SYOSHIFUJI Hideaki 272733ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 27288b5f12d0SFlorian Westphal #ifdef CONFIG_SYN_COOKIES 27298b5f12d0SFlorian Westphal if (unlikely(req->cookie_ts)) 27308b5f12d0SFlorian Westphal TCP_SKB_CB(skb)->when = cookie_init_timestamp(req); 27318b5f12d0SFlorian Westphal else 27328b5f12d0SFlorian Westphal #endif 273333ad798cSAdam Langley TCP_SKB_CB(skb)->when = tcp_time_stamp; 2734f5fff5dcSTom Quetchenbach tcp_header_size = tcp_synack_options(sk, req, mss, 27358336886fSJerry Chu skb, &opts, &md5, xvp, foc) 27364957faadSWilliam Allen Simpson + sizeof(*th); 273733ad798cSAdam Langley 2738aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 2739aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 27401da177e4SLinus Torvalds 2741aa8223c7SArnaldo Carvalho de Melo th = tcp_hdr(skb); 27421da177e4SLinus Torvalds memset(th, 0, sizeof(struct tcphdr)); 27431da177e4SLinus Torvalds th->syn = 1; 27441da177e4SLinus Torvalds th->ack = 1; 27451da177e4SLinus Torvalds TCP_ECN_make_synack(req, th); 2746a3116ac5SKOVACS Krisztian th->source = ireq->loc_port; 27472e6599cbSArnaldo Carvalho de Melo th->dest = ireq->rmt_port; 2748e870a8efSIlpo Järvinen /* Setting of flags are superfluous here for callers (and ECE is 2749e870a8efSIlpo Järvinen * not even correctly set) 2750e870a8efSIlpo Järvinen */ 2751e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn, 2752a3433f35SChangli Gao TCPHDR_SYN | TCPHDR_ACK); 27534957faadSWilliam Allen Simpson 27544957faadSWilliam Allen Simpson if (OPTION_COOKIE_EXTENSION & opts.options) { 275528b2774aSEric Dumazet if (s_data_desired) { 275628b2774aSEric Dumazet u8 *buf = skb_put(skb, s_data_desired); 27574957faadSWilliam Allen Simpson 27584957faadSWilliam Allen Simpson /* copy data directly from the listening socket. */ 275928b2774aSEric Dumazet memcpy(buf, cvp->s_data_payload, s_data_desired); 276028b2774aSEric Dumazet TCP_SKB_CB(skb)->end_seq += s_data_desired; 27614957faadSWilliam Allen Simpson } 27624957faadSWilliam Allen Simpson 27634957faadSWilliam Allen Simpson if (opts.hash_size > 0) { 27644957faadSWilliam Allen Simpson __u32 workspace[SHA_WORKSPACE_WORDS]; 27654957faadSWilliam Allen Simpson u32 *mess = &xvp->cookie_bakery[COOKIE_DIGEST_WORDS]; 27664957faadSWilliam Allen Simpson u32 *tail = &mess[COOKIE_MESSAGE_WORDS-1]; 27674957faadSWilliam Allen Simpson 27684957faadSWilliam Allen Simpson /* Secret recipe depends on the Timestamp, (future) 27694957faadSWilliam Allen Simpson * Sequence and Acknowledgment Numbers, Initiator 27704957faadSWilliam Allen Simpson * Cookie, and others handled by IP variant caller. 27714957faadSWilliam Allen Simpson */ 27724957faadSWilliam Allen Simpson *tail-- ^= opts.tsval; 27734957faadSWilliam Allen Simpson *tail-- ^= tcp_rsk(req)->rcv_isn + 1; 27744957faadSWilliam Allen Simpson *tail-- ^= TCP_SKB_CB(skb)->seq + 1; 27754957faadSWilliam Allen Simpson 27764957faadSWilliam Allen Simpson /* recommended */ 27770eae88f3SEric Dumazet *tail-- ^= (((__force u32)th->dest << 16) | (__force u32)th->source); 2778f9a2e69eSDavid S. Miller *tail-- ^= (u32)(unsigned long)cvp; /* per sockopt */ 27794957faadSWilliam Allen Simpson 27804957faadSWilliam Allen Simpson sha_transform((__u32 *)&xvp->cookie_bakery[0], 27814957faadSWilliam Allen Simpson (char *)mess, 27824957faadSWilliam Allen Simpson &workspace[0]); 27834957faadSWilliam Allen Simpson opts.hash_location = 27844957faadSWilliam Allen Simpson (__u8 *)&xvp->cookie_bakery[0]; 27854957faadSWilliam Allen Simpson } 27864957faadSWilliam Allen Simpson } 27874957faadSWilliam Allen Simpson 27881da177e4SLinus Torvalds th->seq = htonl(TCP_SKB_CB(skb)->seq); 27898336886fSJerry Chu /* XXX data is queued and acked as is. No buffer/window check */ 27908336886fSJerry Chu th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt); 27911da177e4SLinus Torvalds 27921da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 2793600ff0c2SIlpo Järvinen th->window = htons(min(req->rcv_wnd, 65535U)); 2794bd0388aeSWilliam Allen Simpson tcp_options_write((__be32 *)(th + 1), tp, &opts); 27951da177e4SLinus Torvalds th->doff = (tcp_header_size >> 2); 2796aa2ea058STom Herbert TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, tcp_skb_pcount(skb)); 2797cfb6eeb4SYOSHIFUJI Hideaki 2798cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 2799cfb6eeb4SYOSHIFUJI Hideaki /* Okay, we have all we need - do the md5 hash if needed */ 2800cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 2801bd0388aeSWilliam Allen Simpson tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, 280249a72dfbSAdam Langley md5, NULL, req, skb); 2803cfb6eeb4SYOSHIFUJI Hideaki } 2804cfb6eeb4SYOSHIFUJI Hideaki #endif 2805cfb6eeb4SYOSHIFUJI Hideaki 28061da177e4SLinus Torvalds return skb; 28071da177e4SLinus Torvalds } 28084bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_make_synack); 28091da177e4SLinus Torvalds 281067edfef7SAndi Kleen /* Do all connect socket setups that can be done AF independent. */ 2811370816aeSPavel Emelyanov void tcp_connect_init(struct sock *sk) 28121da177e4SLinus Torvalds { 2813cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 28141da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 28151da177e4SLinus Torvalds __u8 rcv_wscale; 28161da177e4SLinus Torvalds 28171da177e4SLinus Torvalds /* We'll fix this up when we get a response from the other end. 28181da177e4SLinus Torvalds * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 28191da177e4SLinus Torvalds */ 28201da177e4SLinus Torvalds tp->tcp_header_len = sizeof(struct tcphdr) + 2821bb5b7c11SDavid S. Miller (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0); 28221da177e4SLinus Torvalds 2823cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 2824cfb6eeb4SYOSHIFUJI Hideaki if (tp->af_specific->md5_lookup(sk, sk) != NULL) 2825cfb6eeb4SYOSHIFUJI Hideaki tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 2826cfb6eeb4SYOSHIFUJI Hideaki #endif 2827cfb6eeb4SYOSHIFUJI Hideaki 28281da177e4SLinus Torvalds /* If user gave his TCP_MAXSEG, record it to clamp */ 28291da177e4SLinus Torvalds if (tp->rx_opt.user_mss) 28301da177e4SLinus Torvalds tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 28311da177e4SLinus Torvalds tp->max_window = 0; 28325d424d5aSJohn Heffner tcp_mtup_init(sk); 28331da177e4SLinus Torvalds tcp_sync_mss(sk, dst_mtu(dst)); 28341da177e4SLinus Torvalds 28351da177e4SLinus Torvalds if (!tp->window_clamp) 28361da177e4SLinus Torvalds tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 28370dbaee3bSDavid S. Miller tp->advmss = dst_metric_advmss(dst); 2838f5fff5dcSTom Quetchenbach if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss) 2839f5fff5dcSTom Quetchenbach tp->advmss = tp->rx_opt.user_mss; 2840f5fff5dcSTom Quetchenbach 28411da177e4SLinus Torvalds tcp_initialize_rcv_mss(sk); 28421da177e4SLinus Torvalds 2843e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 2844e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 2845e88c64f0SHagen Paul Pfeifer (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) 2846e88c64f0SHagen Paul Pfeifer tp->window_clamp = tcp_full_space(sk); 2847e88c64f0SHagen Paul Pfeifer 28481da177e4SLinus Torvalds tcp_select_initial_window(tcp_full_space(sk), 28491da177e4SLinus Torvalds tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 28501da177e4SLinus Torvalds &tp->rcv_wnd, 28511da177e4SLinus Torvalds &tp->window_clamp, 2852bb5b7c11SDavid S. Miller sysctl_tcp_window_scaling, 285331d12926Slaurent chavey &rcv_wscale, 285431d12926Slaurent chavey dst_metric(dst, RTAX_INITRWND)); 28551da177e4SLinus Torvalds 28561da177e4SLinus Torvalds tp->rx_opt.rcv_wscale = rcv_wscale; 28571da177e4SLinus Torvalds tp->rcv_ssthresh = tp->rcv_wnd; 28581da177e4SLinus Torvalds 28591da177e4SLinus Torvalds sk->sk_err = 0; 28601da177e4SLinus Torvalds sock_reset_flag(sk, SOCK_DONE); 28611da177e4SLinus Torvalds tp->snd_wnd = 0; 2862ee7537b6SHantzis Fotis tcp_init_wl(tp, 0); 28631da177e4SLinus Torvalds tp->snd_una = tp->write_seq; 28641da177e4SLinus Torvalds tp->snd_sml = tp->write_seq; 286533f5f57eSIlpo Järvinen tp->snd_up = tp->write_seq; 2866370816aeSPavel Emelyanov tp->snd_nxt = tp->write_seq; 2867ee995283SPavel Emelyanov 2868ee995283SPavel Emelyanov if (likely(!tp->repair)) 28691da177e4SLinus Torvalds tp->rcv_nxt = 0; 2870ee995283SPavel Emelyanov tp->rcv_wup = tp->rcv_nxt; 2871ee995283SPavel Emelyanov tp->copied_seq = tp->rcv_nxt; 28721da177e4SLinus Torvalds 2873463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; 2874463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_retransmits = 0; 28751da177e4SLinus Torvalds tcp_clear_retrans(tp); 28761da177e4SLinus Torvalds } 28771da177e4SLinus Torvalds 2878783237e8SYuchung Cheng static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) 2879783237e8SYuchung Cheng { 2880783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 2881783237e8SYuchung Cheng struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 2882783237e8SYuchung Cheng 2883783237e8SYuchung Cheng tcb->end_seq += skb->len; 2884783237e8SYuchung Cheng skb_header_release(skb); 2885783237e8SYuchung Cheng __tcp_add_write_queue_tail(sk, skb); 2886783237e8SYuchung Cheng sk->sk_wmem_queued += skb->truesize; 2887783237e8SYuchung Cheng sk_mem_charge(sk, skb->truesize); 2888783237e8SYuchung Cheng tp->write_seq = tcb->end_seq; 2889783237e8SYuchung Cheng tp->packets_out += tcp_skb_pcount(skb); 2890783237e8SYuchung Cheng } 2891783237e8SYuchung Cheng 2892783237e8SYuchung Cheng /* Build and send a SYN with data and (cached) Fast Open cookie. However, 2893783237e8SYuchung Cheng * queue a data-only packet after the regular SYN, such that regular SYNs 2894783237e8SYuchung Cheng * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges 2895783237e8SYuchung Cheng * only the SYN sequence, the data are retransmitted in the first ACK. 2896783237e8SYuchung Cheng * If cookie is not cached or other error occurs, falls back to send a 2897783237e8SYuchung Cheng * regular SYN with Fast Open cookie request option. 2898783237e8SYuchung Cheng */ 2899783237e8SYuchung Cheng static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) 2900783237e8SYuchung Cheng { 2901783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 2902783237e8SYuchung Cheng struct tcp_fastopen_request *fo = tp->fastopen_req; 2903aab48743SYuchung Cheng int syn_loss = 0, space, i, err = 0, iovlen = fo->data->msg_iovlen; 2904783237e8SYuchung Cheng struct sk_buff *syn_data = NULL, *data; 2905aab48743SYuchung Cheng unsigned long last_syn_loss = 0; 2906783237e8SYuchung Cheng 290767da22d2SYuchung Cheng tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ 2908aab48743SYuchung Cheng tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie, 2909aab48743SYuchung Cheng &syn_loss, &last_syn_loss); 2910aab48743SYuchung Cheng /* Recurring FO SYN losses: revert to regular handshake temporarily */ 2911aab48743SYuchung Cheng if (syn_loss > 1 && 2912aab48743SYuchung Cheng time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) { 2913aab48743SYuchung Cheng fo->cookie.len = -1; 2914aab48743SYuchung Cheng goto fallback; 2915aab48743SYuchung Cheng } 2916aab48743SYuchung Cheng 291767da22d2SYuchung Cheng if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE) 291867da22d2SYuchung Cheng fo->cookie.len = -1; 291967da22d2SYuchung Cheng else if (fo->cookie.len <= 0) 2920783237e8SYuchung Cheng goto fallback; 2921783237e8SYuchung Cheng 2922783237e8SYuchung Cheng /* MSS for SYN-data is based on cached MSS and bounded by PMTU and 2923783237e8SYuchung Cheng * user-MSS. Reserve maximum option space for middleboxes that add 2924783237e8SYuchung Cheng * private TCP options. The cost is reduced data space in SYN :( 2925783237e8SYuchung Cheng */ 2926783237e8SYuchung Cheng if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp) 2927783237e8SYuchung Cheng tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 2928783237e8SYuchung Cheng space = tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) - 2929783237e8SYuchung Cheng MAX_TCP_OPTION_SPACE; 2930783237e8SYuchung Cheng 2931783237e8SYuchung Cheng syn_data = skb_copy_expand(syn, skb_headroom(syn), space, 2932783237e8SYuchung Cheng sk->sk_allocation); 2933783237e8SYuchung Cheng if (syn_data == NULL) 2934783237e8SYuchung Cheng goto fallback; 2935783237e8SYuchung Cheng 2936783237e8SYuchung Cheng for (i = 0; i < iovlen && syn_data->len < space; ++i) { 2937783237e8SYuchung Cheng struct iovec *iov = &fo->data->msg_iov[i]; 2938783237e8SYuchung Cheng unsigned char __user *from = iov->iov_base; 2939783237e8SYuchung Cheng int len = iov->iov_len; 2940783237e8SYuchung Cheng 2941783237e8SYuchung Cheng if (syn_data->len + len > space) 2942783237e8SYuchung Cheng len = space - syn_data->len; 2943783237e8SYuchung Cheng else if (i + 1 == iovlen) 2944783237e8SYuchung Cheng /* No more data pending in inet_wait_for_connect() */ 2945783237e8SYuchung Cheng fo->data = NULL; 2946783237e8SYuchung Cheng 2947783237e8SYuchung Cheng if (skb_add_data(syn_data, from, len)) 2948783237e8SYuchung Cheng goto fallback; 2949783237e8SYuchung Cheng } 2950783237e8SYuchung Cheng 2951783237e8SYuchung Cheng /* Queue a data-only packet after the regular SYN for retransmission */ 2952783237e8SYuchung Cheng data = pskb_copy(syn_data, sk->sk_allocation); 2953783237e8SYuchung Cheng if (data == NULL) 2954783237e8SYuchung Cheng goto fallback; 2955783237e8SYuchung Cheng TCP_SKB_CB(data)->seq++; 2956783237e8SYuchung Cheng TCP_SKB_CB(data)->tcp_flags &= ~TCPHDR_SYN; 2957783237e8SYuchung Cheng TCP_SKB_CB(data)->tcp_flags = (TCPHDR_ACK|TCPHDR_PSH); 2958783237e8SYuchung Cheng tcp_connect_queue_skb(sk, data); 2959783237e8SYuchung Cheng fo->copied = data->len; 2960783237e8SYuchung Cheng 2961783237e8SYuchung Cheng if (tcp_transmit_skb(sk, syn_data, 0, sk->sk_allocation) == 0) { 296267da22d2SYuchung Cheng tp->syn_data = (fo->copied > 0); 2963783237e8SYuchung Cheng NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE); 2964783237e8SYuchung Cheng goto done; 2965783237e8SYuchung Cheng } 2966783237e8SYuchung Cheng syn_data = NULL; 2967783237e8SYuchung Cheng 2968783237e8SYuchung Cheng fallback: 2969783237e8SYuchung Cheng /* Send a regular SYN with Fast Open cookie request option */ 2970783237e8SYuchung Cheng if (fo->cookie.len > 0) 2971783237e8SYuchung Cheng fo->cookie.len = 0; 2972783237e8SYuchung Cheng err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); 2973783237e8SYuchung Cheng if (err) 2974783237e8SYuchung Cheng tp->syn_fastopen = 0; 2975783237e8SYuchung Cheng kfree_skb(syn_data); 2976783237e8SYuchung Cheng done: 2977783237e8SYuchung Cheng fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ 2978783237e8SYuchung Cheng return err; 2979783237e8SYuchung Cheng } 2980783237e8SYuchung Cheng 298167edfef7SAndi Kleen /* Build a SYN and send it off. */ 29821da177e4SLinus Torvalds int tcp_connect(struct sock *sk) 29831da177e4SLinus Torvalds { 29841da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 29851da177e4SLinus Torvalds struct sk_buff *buff; 2986ee586811SEric Paris int err; 29871da177e4SLinus Torvalds 29881da177e4SLinus Torvalds tcp_connect_init(sk); 29891da177e4SLinus Torvalds 2990*2b916477SAndrey Vagin if (unlikely(tp->repair)) { 2991*2b916477SAndrey Vagin tcp_finish_connect(sk, NULL); 2992*2b916477SAndrey Vagin return 0; 2993*2b916477SAndrey Vagin } 2994*2b916477SAndrey Vagin 2995d179cd12SDavid S. Miller buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation); 29961da177e4SLinus Torvalds if (unlikely(buff == NULL)) 29971da177e4SLinus Torvalds return -ENOBUFS; 29981da177e4SLinus Torvalds 29991da177e4SLinus Torvalds /* Reserve space for headers. */ 30001da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 30011da177e4SLinus Torvalds 3002a3433f35SChangli Gao tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); 3003783237e8SYuchung Cheng tp->retrans_stamp = TCP_SKB_CB(buff)->when = tcp_time_stamp; 3004783237e8SYuchung Cheng tcp_connect_queue_skb(sk, buff); 3005e870a8efSIlpo Järvinen TCP_ECN_send_syn(sk, buff); 30061da177e4SLinus Torvalds 3007783237e8SYuchung Cheng /* Send off SYN; include data in Fast Open. */ 3008783237e8SYuchung Cheng err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : 3009783237e8SYuchung Cheng tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); 3010ee586811SEric Paris if (err == -ECONNREFUSED) 3011ee586811SEric Paris return err; 3012bd37a088SWei Yongjun 3013bd37a088SWei Yongjun /* We change tp->snd_nxt after the tcp_transmit_skb() call 3014bd37a088SWei Yongjun * in order to make this packet get counted in tcpOutSegs. 3015bd37a088SWei Yongjun */ 3016bd37a088SWei Yongjun tp->snd_nxt = tp->write_seq; 3017bd37a088SWei Yongjun tp->pushed_seq = tp->write_seq; 301881cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); 30191da177e4SLinus Torvalds 30201da177e4SLinus Torvalds /* Timer for repeating the SYN until an answer. */ 30213f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 30223f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 30231da177e4SLinus Torvalds return 0; 30241da177e4SLinus Torvalds } 30254bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_connect); 30261da177e4SLinus Torvalds 30271da177e4SLinus Torvalds /* Send out a delayed ack, the caller does the policy checking 30281da177e4SLinus Torvalds * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 30291da177e4SLinus Torvalds * for details. 30301da177e4SLinus Torvalds */ 30311da177e4SLinus Torvalds void tcp_send_delayed_ack(struct sock *sk) 30321da177e4SLinus Torvalds { 3033463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 3034463c84b9SArnaldo Carvalho de Melo int ato = icsk->icsk_ack.ato; 30351da177e4SLinus Torvalds unsigned long timeout; 30361da177e4SLinus Torvalds 30371da177e4SLinus Torvalds if (ato > TCP_DELACK_MIN) { 3038463c84b9SArnaldo Carvalho de Melo const struct tcp_sock *tp = tcp_sk(sk); 30391da177e4SLinus Torvalds int max_ato = HZ / 2; 30401da177e4SLinus Torvalds 3041056834d9SIlpo Järvinen if (icsk->icsk_ack.pingpong || 3042056834d9SIlpo Järvinen (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 30431da177e4SLinus Torvalds max_ato = TCP_DELACK_MAX; 30441da177e4SLinus Torvalds 30451da177e4SLinus Torvalds /* Slow path, intersegment interval is "high". */ 30461da177e4SLinus Torvalds 30471da177e4SLinus Torvalds /* If some rtt estimate is known, use it to bound delayed ack. 3048463c84b9SArnaldo Carvalho de Melo * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 30491da177e4SLinus Torvalds * directly. 30501da177e4SLinus Torvalds */ 30511da177e4SLinus Torvalds if (tp->srtt) { 30521da177e4SLinus Torvalds int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN); 30531da177e4SLinus Torvalds 30541da177e4SLinus Torvalds if (rtt < max_ato) 30551da177e4SLinus Torvalds max_ato = rtt; 30561da177e4SLinus Torvalds } 30571da177e4SLinus Torvalds 30581da177e4SLinus Torvalds ato = min(ato, max_ato); 30591da177e4SLinus Torvalds } 30601da177e4SLinus Torvalds 30611da177e4SLinus Torvalds /* Stay within the limit we were given */ 30621da177e4SLinus Torvalds timeout = jiffies + ato; 30631da177e4SLinus Torvalds 30641da177e4SLinus Torvalds /* Use new timeout only if there wasn't a older one earlier. */ 3065463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 30661da177e4SLinus Torvalds /* If delack timer was blocked or is about to expire, 30671da177e4SLinus Torvalds * send ACK now. 30681da177e4SLinus Torvalds */ 3069463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.blocked || 3070463c84b9SArnaldo Carvalho de Melo time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 30711da177e4SLinus Torvalds tcp_send_ack(sk); 30721da177e4SLinus Torvalds return; 30731da177e4SLinus Torvalds } 30741da177e4SLinus Torvalds 3075463c84b9SArnaldo Carvalho de Melo if (!time_before(timeout, icsk->icsk_ack.timeout)) 3076463c84b9SArnaldo Carvalho de Melo timeout = icsk->icsk_ack.timeout; 30771da177e4SLinus Torvalds } 3078463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 3079463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.timeout = timeout; 3080463c84b9SArnaldo Carvalho de Melo sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 30811da177e4SLinus Torvalds } 30821da177e4SLinus Torvalds 30831da177e4SLinus Torvalds /* This routine sends an ack and also updates the window. */ 30841da177e4SLinus Torvalds void tcp_send_ack(struct sock *sk) 30851da177e4SLinus Torvalds { 30861da177e4SLinus Torvalds struct sk_buff *buff; 30871da177e4SLinus Torvalds 3088058dc334SIlpo Järvinen /* If we have been reset, we may not send again. */ 3089058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3090058dc334SIlpo Järvinen return; 3091058dc334SIlpo Järvinen 30921da177e4SLinus Torvalds /* We are not putting this on the write queue, so 30931da177e4SLinus Torvalds * tcp_transmit_skb() will set the ownership to this 30941da177e4SLinus Torvalds * sock. 30951da177e4SLinus Torvalds */ 309699a1dec7SMel Gorman buff = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC)); 30971da177e4SLinus Torvalds if (buff == NULL) { 3098463c84b9SArnaldo Carvalho de Melo inet_csk_schedule_ack(sk); 3099463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; 31003f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 31013f421baaSArnaldo Carvalho de Melo TCP_DELACK_MAX, TCP_RTO_MAX); 31021da177e4SLinus Torvalds return; 31031da177e4SLinus Torvalds } 31041da177e4SLinus Torvalds 31051da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 31061da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 3107a3433f35SChangli Gao tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); 31081da177e4SLinus Torvalds 31091da177e4SLinus Torvalds /* Send it off, this clears delayed acks for us. */ 31101da177e4SLinus Torvalds TCP_SKB_CB(buff)->when = tcp_time_stamp; 311199a1dec7SMel Gorman tcp_transmit_skb(sk, buff, 0, sk_gfp_atomic(sk, GFP_ATOMIC)); 31121da177e4SLinus Torvalds } 31131da177e4SLinus Torvalds 31141da177e4SLinus Torvalds /* This routine sends a packet with an out of date sequence 31151da177e4SLinus Torvalds * number. It assumes the other end will try to ack it. 31161da177e4SLinus Torvalds * 31171da177e4SLinus Torvalds * Question: what should we make while urgent mode? 31181da177e4SLinus Torvalds * 4.4BSD forces sending single byte of data. We cannot send 31191da177e4SLinus Torvalds * out of window data, because we have SND.NXT==SND.MAX... 31201da177e4SLinus Torvalds * 31211da177e4SLinus Torvalds * Current solution: to send TWO zero-length segments in urgent mode: 31221da177e4SLinus Torvalds * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 31231da177e4SLinus Torvalds * out-of-date with SND.UNA-1 to probe window. 31241da177e4SLinus Torvalds */ 31251da177e4SLinus Torvalds static int tcp_xmit_probe_skb(struct sock *sk, int urgent) 31261da177e4SLinus Torvalds { 31271da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 31281da177e4SLinus Torvalds struct sk_buff *skb; 31291da177e4SLinus Torvalds 31301da177e4SLinus Torvalds /* We don't queue it, tcp_transmit_skb() sets ownership. */ 313199a1dec7SMel Gorman skb = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC)); 31321da177e4SLinus Torvalds if (skb == NULL) 31331da177e4SLinus Torvalds return -1; 31341da177e4SLinus Torvalds 31351da177e4SLinus Torvalds /* Reserve space for headers and set control bits. */ 31361da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 31371da177e4SLinus Torvalds /* Use a previous sequence. This should cause the other 31381da177e4SLinus Torvalds * end to send an ack. Don't queue or clone SKB, just 31391da177e4SLinus Torvalds * send it. 31401da177e4SLinus Torvalds */ 3141a3433f35SChangli Gao tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); 31421da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 3143dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC); 31441da177e4SLinus Torvalds } 31451da177e4SLinus Torvalds 3146ee995283SPavel Emelyanov void tcp_send_window_probe(struct sock *sk) 3147ee995283SPavel Emelyanov { 3148ee995283SPavel Emelyanov if (sk->sk_state == TCP_ESTABLISHED) { 3149ee995283SPavel Emelyanov tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; 3150c0e88ff0SPavel Emelyanov tcp_sk(sk)->snd_nxt = tcp_sk(sk)->write_seq; 3151ee995283SPavel Emelyanov tcp_xmit_probe_skb(sk, 0); 3152ee995283SPavel Emelyanov } 3153ee995283SPavel Emelyanov } 3154ee995283SPavel Emelyanov 315567edfef7SAndi Kleen /* Initiate keepalive or window probe from timer. */ 31561da177e4SLinus Torvalds int tcp_write_wakeup(struct sock *sk) 31571da177e4SLinus Torvalds { 31581da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 31591da177e4SLinus Torvalds struct sk_buff *skb; 31601da177e4SLinus Torvalds 3161058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3162058dc334SIlpo Järvinen return -1; 3163058dc334SIlpo Järvinen 3164fe067e8aSDavid S. Miller if ((skb = tcp_send_head(sk)) != NULL && 316590840defSIlpo Järvinen before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { 31661da177e4SLinus Torvalds int err; 31670c54b85fSIlpo Järvinen unsigned int mss = tcp_current_mss(sk); 316890840defSIlpo Järvinen unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 31691da177e4SLinus Torvalds 31701da177e4SLinus Torvalds if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 31711da177e4SLinus Torvalds tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 31721da177e4SLinus Torvalds 31731da177e4SLinus Torvalds /* We are probing the opening of a window 31741da177e4SLinus Torvalds * but the window size is != 0 31751da177e4SLinus Torvalds * must have been a result SWS avoidance ( sender ) 31761da177e4SLinus Torvalds */ 31771da177e4SLinus Torvalds if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 31781da177e4SLinus Torvalds skb->len > mss) { 31791da177e4SLinus Torvalds seg_size = min(seg_size, mss); 31804de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 3181846998aeSDavid S. Miller if (tcp_fragment(sk, skb, seg_size, mss)) 31821da177e4SLinus Torvalds return -1; 31831da177e4SLinus Torvalds } else if (!tcp_skb_pcount(skb)) 3184846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss); 31851da177e4SLinus Torvalds 31864de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 31871da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 3188dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 318966f5fe62SIlpo Järvinen if (!err) 319066f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 31911da177e4SLinus Torvalds return err; 31921da177e4SLinus Torvalds } else { 319333f5f57eSIlpo Järvinen if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) 31944828e7f4SIlpo Järvinen tcp_xmit_probe_skb(sk, 1); 31951da177e4SLinus Torvalds return tcp_xmit_probe_skb(sk, 0); 31961da177e4SLinus Torvalds } 31971da177e4SLinus Torvalds } 31981da177e4SLinus Torvalds 31991da177e4SLinus Torvalds /* A window probe timeout has occurred. If window is not closed send 32001da177e4SLinus Torvalds * a partial packet else a zero probe. 32011da177e4SLinus Torvalds */ 32021da177e4SLinus Torvalds void tcp_send_probe0(struct sock *sk) 32031da177e4SLinus Torvalds { 3204463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 32051da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 32061da177e4SLinus Torvalds int err; 32071da177e4SLinus Torvalds 32081da177e4SLinus Torvalds err = tcp_write_wakeup(sk); 32091da177e4SLinus Torvalds 3210fe067e8aSDavid S. Miller if (tp->packets_out || !tcp_send_head(sk)) { 32111da177e4SLinus Torvalds /* Cancel probe timer, if it is not required. */ 32126687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 0; 3213463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff = 0; 32141da177e4SLinus Torvalds return; 32151da177e4SLinus Torvalds } 32161da177e4SLinus Torvalds 32171da177e4SLinus Torvalds if (err <= 0) { 3218463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_backoff < sysctl_tcp_retries2) 3219463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff++; 32206687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out++; 3221463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 32223f421baaSArnaldo Carvalho de Melo min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), 32233f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 32241da177e4SLinus Torvalds } else { 32251da177e4SLinus Torvalds /* If packet was not sent due to local congestion, 32266687e988SArnaldo Carvalho de Melo * do not backoff and do not remember icsk_probes_out. 32271da177e4SLinus Torvalds * Let local senders to fight for local resources. 32281da177e4SLinus Torvalds * 32291da177e4SLinus Torvalds * Use accumulated backoff yet. 32301da177e4SLinus Torvalds */ 32316687e988SArnaldo Carvalho de Melo if (!icsk->icsk_probes_out) 32326687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 1; 3233463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 3234463c84b9SArnaldo Carvalho de Melo min(icsk->icsk_rto << icsk->icsk_backoff, 32353f421baaSArnaldo Carvalho de Melo TCP_RESOURCE_PROBE_INTERVAL), 32363f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 32371da177e4SLinus Torvalds } 32381da177e4SLinus Torvalds } 3239