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 53c39c4c6aSWei Liu /* Default TSQ limit of four TSO segments */ 54c39c4c6aSWei Liu int sysctl_tcp_limit_output_bytes __read_mostly = 262144; 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 6235089bb2SDavid S. Miller /* By default, RFC2861 behavior. */ 63ab32ea5dSBrian Haley int sysctl_tcp_slow_start_after_idle __read_mostly = 1; 6435089bb2SDavid S. Miller 6546d3ceabSEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 6646d3ceabSEric Dumazet int push_one, gfp_t gfp); 67519855c5SWilliam Allen Simpson 6867edfef7SAndi Kleen /* Account for new data that has been sent to the network. */ 69cf533ea5SEric Dumazet static void tcp_event_new_data_sent(struct sock *sk, const struct sk_buff *skb) 706ff03ac3SIlpo Järvinen { 716ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 726ff03ac3SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 7366f5fe62SIlpo Järvinen unsigned int prior_packets = tp->packets_out; 749e412ba7SIlpo Järvinen 75fe067e8aSDavid S. Miller tcp_advance_send_head(sk, skb); 761da177e4SLinus Torvalds tp->snd_nxt = TCP_SKB_CB(skb)->end_seq; 778512430eSIlpo Järvinen 7866f5fe62SIlpo Järvinen tp->packets_out += tcp_skb_pcount(skb); 796ba8a3b1SNandita Dukkipati if (!prior_packets || icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS || 806a5dc9e5SEric Dumazet icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) { 81750ea2baSYuchung Cheng tcp_rearm_rto(sk); 821da177e4SLinus Torvalds } 83f19c29e3SYuchung Cheng 84f7324acdSDavid S. Miller NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT, 85f19c29e3SYuchung Cheng tcp_skb_pcount(skb)); 866a5dc9e5SEric Dumazet } 871da177e4SLinus Torvalds 881da177e4SLinus Torvalds /* SND.NXT, if window was not shrunk. 891da177e4SLinus Torvalds * If window has been shrunk, what should we make? It is not clear at all. 901da177e4SLinus Torvalds * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-( 911da177e4SLinus Torvalds * Anything in between SND.UNA...SND.UNA+SND.WND also can be already 921da177e4SLinus Torvalds * invalid. OK, let's make this for now: 931da177e4SLinus Torvalds */ 94cf533ea5SEric Dumazet static inline __u32 tcp_acceptable_seq(const struct sock *sk) 951da177e4SLinus Torvalds { 96cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 979e412ba7SIlpo Järvinen 9890840defSIlpo Järvinen if (!before(tcp_wnd_end(tp), tp->snd_nxt)) 991da177e4SLinus Torvalds return tp->snd_nxt; 1001da177e4SLinus Torvalds else 10190840defSIlpo Järvinen return tcp_wnd_end(tp); 1021da177e4SLinus Torvalds } 1031da177e4SLinus Torvalds 1041da177e4SLinus Torvalds /* Calculate mss to advertise in SYN segment. 1051da177e4SLinus Torvalds * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that: 1061da177e4SLinus Torvalds * 1071da177e4SLinus Torvalds * 1. It is independent of path mtu. 1081da177e4SLinus Torvalds * 2. Ideally, it is maximal possible segment size i.e. 65535-40. 1091da177e4SLinus Torvalds * 3. For IPv4 it is reasonable to calculate it from maximal MTU of 1101da177e4SLinus Torvalds * attached devices, because some buggy hosts are confused by 1111da177e4SLinus Torvalds * large MSS. 1121da177e4SLinus Torvalds * 4. We do not make 3, we advertise MSS, calculated from first 1131da177e4SLinus Torvalds * hop device mtu, but allow to raise it to ip_rt_min_advmss. 1141da177e4SLinus Torvalds * This may be overridden via information stored in routing table. 1151da177e4SLinus Torvalds * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible, 1161da177e4SLinus Torvalds * probably even Jumbo". 1171da177e4SLinus Torvalds */ 1181da177e4SLinus Torvalds static __u16 tcp_advertise_mss(struct sock *sk) 1191da177e4SLinus Torvalds { 1201da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 121cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1221da177e4SLinus Torvalds int mss = tp->advmss; 1231da177e4SLinus Torvalds 1240dbaee3bSDavid S. Miller if (dst) { 1250dbaee3bSDavid S. Miller unsigned int metric = dst_metric_advmss(dst); 1260dbaee3bSDavid S. Miller 1270dbaee3bSDavid S. Miller if (metric < mss) { 1280dbaee3bSDavid S. Miller mss = metric; 1291da177e4SLinus Torvalds tp->advmss = mss; 1301da177e4SLinus Torvalds } 1310dbaee3bSDavid S. Miller } 1321da177e4SLinus Torvalds 1331da177e4SLinus Torvalds return (__u16)mss; 1341da177e4SLinus Torvalds } 1351da177e4SLinus Torvalds 1361da177e4SLinus Torvalds /* RFC2861. Reset CWND after idle period longer RTO to "restart window". 1376f021c62SEric Dumazet * This is the first part of cwnd validation mechanism. 1386f021c62SEric Dumazet */ 1396f021c62SEric Dumazet void tcp_cwnd_restart(struct sock *sk, s32 delta) 1401da177e4SLinus Torvalds { 141463c84b9SArnaldo Carvalho de Melo struct tcp_sock *tp = tcp_sk(sk); 1426f021c62SEric Dumazet u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk)); 1431da177e4SLinus Torvalds u32 cwnd = tp->snd_cwnd; 1441da177e4SLinus Torvalds 1456687e988SArnaldo Carvalho de Melo tcp_ca_event(sk, CA_EVENT_CWND_RESTART); 1461da177e4SLinus Torvalds 1476687e988SArnaldo Carvalho de Melo tp->snd_ssthresh = tcp_current_ssthresh(sk); 1481da177e4SLinus Torvalds restart_cwnd = min(restart_cwnd, cwnd); 1491da177e4SLinus Torvalds 150463c84b9SArnaldo Carvalho de Melo while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd) 1511da177e4SLinus Torvalds cwnd >>= 1; 1521da177e4SLinus Torvalds tp->snd_cwnd = max(cwnd, restart_cwnd); 1531da177e4SLinus Torvalds tp->snd_cwnd_stamp = tcp_time_stamp; 1541da177e4SLinus Torvalds tp->snd_cwnd_used = 0; 1551da177e4SLinus Torvalds } 1561da177e4SLinus Torvalds 15767edfef7SAndi Kleen /* Congestion state accounting after a packet has been sent. */ 15840efc6faSStephen Hemminger static void tcp_event_data_sent(struct tcp_sock *tp, 159cf533ea5SEric Dumazet struct sock *sk) 1601da177e4SLinus Torvalds { 161463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 162463c84b9SArnaldo Carvalho de Melo const u32 now = tcp_time_stamp; 1631da177e4SLinus Torvalds 16405c5a46dSNeal Cardwell if (tcp_packets_in_flight(tp) == 0) 16505c5a46dSNeal Cardwell tcp_ca_event(sk, CA_EVENT_TX_START); 16605c5a46dSNeal Cardwell 1671da177e4SLinus Torvalds tp->lsndtime = now; 1681da177e4SLinus Torvalds 1691da177e4SLinus Torvalds /* If it is a reply for ato after last received 1701da177e4SLinus Torvalds * packet, enter pingpong mode. 1711da177e4SLinus Torvalds */ 1722251ae46SJon Maxwell if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato) 173463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pingpong = 1; 1741da177e4SLinus Torvalds } 1751da177e4SLinus Torvalds 17667edfef7SAndi Kleen /* Account for an ACK we sent. */ 17740efc6faSStephen Hemminger static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts) 1781da177e4SLinus Torvalds { 179463c84b9SArnaldo Carvalho de Melo tcp_dec_quickack_mode(sk, pkts); 180463c84b9SArnaldo Carvalho de Melo inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); 1811da177e4SLinus Torvalds } 1821da177e4SLinus Torvalds 18385f16525SYuchung Cheng 18485f16525SYuchung Cheng u32 tcp_default_init_rwnd(u32 mss) 18585f16525SYuchung Cheng { 18685f16525SYuchung Cheng /* Initial receive window should be twice of TCP_INIT_CWND to 1879ef71e0cSWeiping Pan * enable proper sending of new unsent data during fast recovery 18885f16525SYuchung Cheng * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a 18985f16525SYuchung Cheng * limit when mss is larger than 1460. 19085f16525SYuchung Cheng */ 19185f16525SYuchung Cheng u32 init_rwnd = TCP_INIT_CWND * 2; 19285f16525SYuchung Cheng 19385f16525SYuchung Cheng if (mss > 1460) 19485f16525SYuchung Cheng init_rwnd = max((1460 * init_rwnd) / mss, 2U); 19585f16525SYuchung Cheng return init_rwnd; 19685f16525SYuchung Cheng } 19785f16525SYuchung Cheng 1981da177e4SLinus Torvalds /* Determine a window scaling and initial window to offer. 1991da177e4SLinus Torvalds * Based on the assumption that the given amount of space 2001da177e4SLinus Torvalds * will be offered. Store the results in the tp structure. 2011da177e4SLinus Torvalds * NOTE: for smooth operation initial space offering should 2021da177e4SLinus Torvalds * be a multiple of mss if possible. We assume here that mss >= 1. 2031da177e4SLinus Torvalds * This MUST be enforced by all callers. 2041da177e4SLinus Torvalds */ 2051da177e4SLinus Torvalds void tcp_select_initial_window(int __space, __u32 mss, 2061da177e4SLinus Torvalds __u32 *rcv_wnd, __u32 *window_clamp, 20731d12926Slaurent chavey int wscale_ok, __u8 *rcv_wscale, 20831d12926Slaurent chavey __u32 init_rcv_wnd) 2091da177e4SLinus Torvalds { 2101da177e4SLinus Torvalds unsigned int space = (__space < 0 ? 0 : __space); 2111da177e4SLinus Torvalds 2121da177e4SLinus Torvalds /* If no clamp set the clamp to the max possible scaled window */ 2131da177e4SLinus Torvalds if (*window_clamp == 0) 2141da177e4SLinus Torvalds (*window_clamp) = (65535 << 14); 2151da177e4SLinus Torvalds space = min(*window_clamp, space); 2161da177e4SLinus Torvalds 2171da177e4SLinus Torvalds /* Quantize space offering to a multiple of mss if possible. */ 2181da177e4SLinus Torvalds if (space > mss) 2191da177e4SLinus Torvalds space = (space / mss) * mss; 2201da177e4SLinus Torvalds 2211da177e4SLinus Torvalds /* NOTE: offering an initial window larger than 32767 22215d99e02SRick Jones * will break some buggy TCP stacks. If the admin tells us 22315d99e02SRick Jones * it is likely we could be speaking with such a buggy stack 22415d99e02SRick Jones * we will truncate our initial window offering to 32K-1 22515d99e02SRick Jones * unless the remote has sent us a window scaling option, 22615d99e02SRick Jones * which we interpret as a sign the remote TCP is not 22715d99e02SRick Jones * misinterpreting the window field as a signed quantity. 2281da177e4SLinus Torvalds */ 22915d99e02SRick Jones if (sysctl_tcp_workaround_signed_windows) 2301da177e4SLinus Torvalds (*rcv_wnd) = min(space, MAX_TCP_WINDOW); 23115d99e02SRick Jones else 23215d99e02SRick Jones (*rcv_wnd) = space; 23315d99e02SRick Jones 2341da177e4SLinus Torvalds (*rcv_wscale) = 0; 2351da177e4SLinus Torvalds if (wscale_ok) { 2361da177e4SLinus Torvalds /* Set window scaling on max possible window 2371da177e4SLinus Torvalds * See RFC1323 for an explanation of the limit to 14 2381da177e4SLinus Torvalds */ 239f626300aSSoheil Hassas Yeganeh space = max_t(u32, space, sysctl_tcp_rmem[2]); 240f626300aSSoheil Hassas Yeganeh space = max_t(u32, space, sysctl_rmem_max); 241316c1592SStephen Hemminger space = min_t(u32, space, *window_clamp); 2421da177e4SLinus Torvalds while (space > 65535 && (*rcv_wscale) < 14) { 2431da177e4SLinus Torvalds space >>= 1; 2441da177e4SLinus Torvalds (*rcv_wscale)++; 2451da177e4SLinus Torvalds } 2461da177e4SLinus Torvalds } 2471da177e4SLinus Torvalds 2481da177e4SLinus Torvalds if (mss > (1 << *rcv_wscale)) { 24985f16525SYuchung Cheng if (!init_rcv_wnd) /* Use default unless specified otherwise */ 25085f16525SYuchung Cheng init_rcv_wnd = tcp_default_init_rwnd(mss); 251b1afde60SNandita Dukkipati *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss); 2521da177e4SLinus Torvalds } 2531da177e4SLinus Torvalds 2541da177e4SLinus Torvalds /* Set the clamp no higher than max representable value */ 2551da177e4SLinus Torvalds (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp); 2561da177e4SLinus Torvalds } 2574bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_select_initial_window); 2581da177e4SLinus Torvalds 2591da177e4SLinus Torvalds /* Chose a new window to advertise, update state in tcp_sock for the 2601da177e4SLinus Torvalds * socket, and return result with RFC1323 scaling applied. The return 2611da177e4SLinus Torvalds * value can be stuffed directly into th->window for an outgoing 2621da177e4SLinus Torvalds * frame. 2631da177e4SLinus Torvalds */ 26440efc6faSStephen Hemminger static u16 tcp_select_window(struct sock *sk) 2651da177e4SLinus Torvalds { 2661da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2678e165e20SFlorian Westphal u32 old_win = tp->rcv_wnd; 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 */ 2808e165e20SFlorian Westphal if (new_win == 0) 2818e165e20SFlorian Westphal NET_INC_STATS(sock_net(sk), 2828e165e20SFlorian Westphal LINUX_MIB_TCPWANTZEROWINDOWADV); 283607bfbf2SPatrick McHardy new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale); 2841da177e4SLinus Torvalds } 2851da177e4SLinus Torvalds tp->rcv_wnd = new_win; 2861da177e4SLinus Torvalds tp->rcv_wup = tp->rcv_nxt; 2871da177e4SLinus Torvalds 2881da177e4SLinus Torvalds /* Make sure we do not exceed the maximum possible 2891da177e4SLinus Torvalds * scaled window. 2901da177e4SLinus Torvalds */ 29115d99e02SRick Jones if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows) 2921da177e4SLinus Torvalds new_win = min(new_win, MAX_TCP_WINDOW); 2931da177e4SLinus Torvalds else 2941da177e4SLinus Torvalds new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); 2951da177e4SLinus Torvalds 2961da177e4SLinus Torvalds /* RFC1323 scaling applied */ 2971da177e4SLinus Torvalds new_win >>= tp->rx_opt.rcv_wscale; 2981da177e4SLinus Torvalds 2991da177e4SLinus Torvalds /* If we advertise zero window, disable fast path. */ 3008e165e20SFlorian Westphal if (new_win == 0) { 3011da177e4SLinus Torvalds tp->pred_flags = 0; 3028e165e20SFlorian Westphal if (old_win) 3038e165e20SFlorian Westphal NET_INC_STATS(sock_net(sk), 3048e165e20SFlorian Westphal LINUX_MIB_TCPTOZEROWINDOWADV); 3058e165e20SFlorian Westphal } else if (old_win == 0) { 3068e165e20SFlorian Westphal NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV); 3078e165e20SFlorian Westphal } 3081da177e4SLinus Torvalds 3091da177e4SLinus Torvalds return new_win; 3101da177e4SLinus Torvalds } 3111da177e4SLinus Torvalds 31267edfef7SAndi Kleen /* Packet ECN state for a SYN-ACK */ 313735d3831SFlorian Westphal static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb) 314bdf1ee5dSIlpo Järvinen { 31530e502a3SDaniel Borkmann const struct tcp_sock *tp = tcp_sk(sk); 31630e502a3SDaniel Borkmann 3174de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR; 318bdf1ee5dSIlpo Järvinen if (!(tp->ecn_flags & TCP_ECN_OK)) 3194de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE; 32030e502a3SDaniel Borkmann else if (tcp_ca_needs_ecn(sk)) 32130e502a3SDaniel Borkmann INET_ECN_xmit(sk); 322bdf1ee5dSIlpo Järvinen } 323bdf1ee5dSIlpo Järvinen 32467edfef7SAndi Kleen /* Packet ECN state for a SYN. */ 325735d3831SFlorian Westphal static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb) 326bdf1ee5dSIlpo Järvinen { 327bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 328f7b3bec6SFlorian Westphal bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 || 329f7b3bec6SFlorian Westphal tcp_ca_needs_ecn(sk); 330f7b3bec6SFlorian Westphal 331f7b3bec6SFlorian Westphal if (!use_ecn) { 332f7b3bec6SFlorian Westphal const struct dst_entry *dst = __sk_dst_get(sk); 333f7b3bec6SFlorian Westphal 334f7b3bec6SFlorian Westphal if (dst && dst_feature(dst, RTAX_FEATURE_ECN)) 335f7b3bec6SFlorian Westphal use_ecn = true; 336f7b3bec6SFlorian Westphal } 337bdf1ee5dSIlpo Järvinen 338bdf1ee5dSIlpo Järvinen tp->ecn_flags = 0; 339f7b3bec6SFlorian Westphal 340f7b3bec6SFlorian Westphal if (use_ecn) { 3414de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR; 342bdf1ee5dSIlpo Järvinen tp->ecn_flags = TCP_ECN_OK; 34330e502a3SDaniel Borkmann if (tcp_ca_needs_ecn(sk)) 34430e502a3SDaniel Borkmann INET_ECN_xmit(sk); 345bdf1ee5dSIlpo Järvinen } 346bdf1ee5dSIlpo Järvinen } 347bdf1ee5dSIlpo Järvinen 34849213555SDaniel Borkmann static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb) 34949213555SDaniel Borkmann { 35049213555SDaniel Borkmann if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback) 35149213555SDaniel Borkmann /* tp->ecn_flags are cleared at a later point in time when 35249213555SDaniel Borkmann * SYN ACK is ultimatively being received. 35349213555SDaniel Borkmann */ 35449213555SDaniel Borkmann TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR); 35549213555SDaniel Borkmann } 35649213555SDaniel Borkmann 357735d3831SFlorian Westphal static void 3586ac705b1SEric Dumazet tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th) 359bdf1ee5dSIlpo Järvinen { 3606ac705b1SEric Dumazet if (inet_rsk(req)->ecn_ok) 361bdf1ee5dSIlpo Järvinen th->ece = 1; 362bdf1ee5dSIlpo Järvinen } 363bdf1ee5dSIlpo Järvinen 36467edfef7SAndi Kleen /* Set up ECN state for a packet on a ESTABLISHED socket that is about to 36567edfef7SAndi Kleen * be sent. 36667edfef7SAndi Kleen */ 367735d3831SFlorian Westphal static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb, 368ea1627c2SEric Dumazet struct tcphdr *th, int tcp_header_len) 369bdf1ee5dSIlpo Järvinen { 370bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 371bdf1ee5dSIlpo Järvinen 372bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_OK) { 373bdf1ee5dSIlpo Järvinen /* Not-retransmitted data segment: set ECT and inject CWR. */ 374bdf1ee5dSIlpo Järvinen if (skb->len != tcp_header_len && 375bdf1ee5dSIlpo Järvinen !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { 376bdf1ee5dSIlpo Järvinen INET_ECN_xmit(sk); 377bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) { 378bdf1ee5dSIlpo Järvinen tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; 379ea1627c2SEric Dumazet th->cwr = 1; 380bdf1ee5dSIlpo Järvinen skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; 381bdf1ee5dSIlpo Järvinen } 38230e502a3SDaniel Borkmann } else if (!tcp_ca_needs_ecn(sk)) { 383bdf1ee5dSIlpo Järvinen /* ACK or retransmitted segment: clear ECT|CE */ 384bdf1ee5dSIlpo Järvinen INET_ECN_dontxmit(sk); 385bdf1ee5dSIlpo Järvinen } 386bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) 387ea1627c2SEric Dumazet th->ece = 1; 388bdf1ee5dSIlpo Järvinen } 389bdf1ee5dSIlpo Järvinen } 390bdf1ee5dSIlpo Järvinen 391e870a8efSIlpo Järvinen /* Constructs common control bits of non-data skb. If SYN/FIN is present, 392e870a8efSIlpo Järvinen * auto increment end seqno. 393e870a8efSIlpo Järvinen */ 394e870a8efSIlpo Järvinen static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags) 395e870a8efSIlpo Järvinen { 3962e8e18efSDavid S. Miller skb->ip_summed = CHECKSUM_PARTIAL; 397e870a8efSIlpo Järvinen skb->csum = 0; 398e870a8efSIlpo Järvinen 3994de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags; 400e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->sacked = 0; 401e870a8efSIlpo Järvinen 402cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, 1); 403e870a8efSIlpo Järvinen 404e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->seq = seq; 405a3433f35SChangli Gao if (flags & (TCPHDR_SYN | TCPHDR_FIN)) 406e870a8efSIlpo Järvinen seq++; 407e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->end_seq = seq; 408e870a8efSIlpo Järvinen } 409e870a8efSIlpo Järvinen 410a2a385d6SEric Dumazet static inline bool tcp_urg_mode(const struct tcp_sock *tp) 41133f5f57eSIlpo Järvinen { 41233f5f57eSIlpo Järvinen return tp->snd_una != tp->snd_up; 41333f5f57eSIlpo Järvinen } 41433f5f57eSIlpo Järvinen 41533ad798cSAdam Langley #define OPTION_SACK_ADVERTISE (1 << 0) 41633ad798cSAdam Langley #define OPTION_TS (1 << 1) 41733ad798cSAdam Langley #define OPTION_MD5 (1 << 2) 41889e95a61SOri Finkelman #define OPTION_WSCALE (1 << 3) 4192100c8d2SYuchung Cheng #define OPTION_FAST_OPEN_COOKIE (1 << 8) 42033ad798cSAdam Langley 42133ad798cSAdam Langley struct tcp_out_options { 4222100c8d2SYuchung Cheng u16 options; /* bit field of OPTION_* */ 4232100c8d2SYuchung Cheng u16 mss; /* 0 to disable */ 42433ad798cSAdam Langley u8 ws; /* window scale, 0 to disable */ 42533ad798cSAdam Langley u8 num_sack_blocks; /* number of SACK blocks to include */ 426bd0388aeSWilliam Allen Simpson u8 hash_size; /* bytes in hash_location */ 427bd0388aeSWilliam Allen Simpson __u8 *hash_location; /* temporary pointer, overloaded */ 4282100c8d2SYuchung Cheng __u32 tsval, tsecr; /* need to include OPTION_TS */ 4292100c8d2SYuchung Cheng struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */ 43033ad798cSAdam Langley }; 43133ad798cSAdam Langley 43267edfef7SAndi Kleen /* Write previously computed TCP options to the packet. 43367edfef7SAndi Kleen * 43467edfef7SAndi Kleen * Beware: Something in the Internet is very sensitive to the ordering of 435fd6149d3SIlpo Järvinen * TCP options, we learned this through the hard way, so be careful here. 436fd6149d3SIlpo Järvinen * Luckily we can at least blame others for their non-compliance but from 4378e3bff96Sstephen hemminger * inter-operability perspective it seems that we're somewhat stuck with 438fd6149d3SIlpo Järvinen * the ordering which we have been using if we want to keep working with 439fd6149d3SIlpo Järvinen * those broken things (not that it currently hurts anybody as there isn't 440fd6149d3SIlpo Järvinen * particular reason why the ordering would need to be changed). 441fd6149d3SIlpo Järvinen * 442fd6149d3SIlpo Järvinen * At least SACK_PERM as the first option is known to lead to a disaster 443fd6149d3SIlpo Järvinen * (but it may well be that other scenarios fail similarly). 444fd6149d3SIlpo Järvinen */ 44533ad798cSAdam Langley static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp, 446bd0388aeSWilliam Allen Simpson struct tcp_out_options *opts) 447bd0388aeSWilliam Allen Simpson { 4482100c8d2SYuchung Cheng u16 options = opts->options; /* mungable copy */ 449bd0388aeSWilliam Allen Simpson 450bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_MD5 & options)) { 4511a2c6181SChristoph Paasch *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | 4521a2c6181SChristoph Paasch (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG); 453bd0388aeSWilliam Allen Simpson /* overload cookie hash location */ 454bd0388aeSWilliam Allen Simpson opts->hash_location = (__u8 *)ptr; 45533ad798cSAdam Langley ptr += 4; 45633ad798cSAdam Langley } 45733ad798cSAdam Langley 458fd6149d3SIlpo Järvinen if (unlikely(opts->mss)) { 459fd6149d3SIlpo Järvinen *ptr++ = htonl((TCPOPT_MSS << 24) | 460fd6149d3SIlpo Järvinen (TCPOLEN_MSS << 16) | 461fd6149d3SIlpo Järvinen opts->mss); 462fd6149d3SIlpo Järvinen } 463fd6149d3SIlpo Järvinen 464bd0388aeSWilliam Allen Simpson if (likely(OPTION_TS & options)) { 465bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 46633ad798cSAdam Langley *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | 46733ad798cSAdam Langley (TCPOLEN_SACK_PERM << 16) | 46833ad798cSAdam Langley (TCPOPT_TIMESTAMP << 8) | 46933ad798cSAdam Langley TCPOLEN_TIMESTAMP); 470bd0388aeSWilliam Allen Simpson options &= ~OPTION_SACK_ADVERTISE; 47133ad798cSAdam Langley } else { 472496c98dfSYOSHIFUJI Hideaki *ptr++ = htonl((TCPOPT_NOP << 24) | 47340efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 47440efc6faSStephen Hemminger (TCPOPT_TIMESTAMP << 8) | 47540efc6faSStephen Hemminger TCPOLEN_TIMESTAMP); 47640efc6faSStephen Hemminger } 47733ad798cSAdam Langley *ptr++ = htonl(opts->tsval); 47833ad798cSAdam Langley *ptr++ = htonl(opts->tsecr); 47933ad798cSAdam Langley } 48033ad798cSAdam Langley 481bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 48233ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 48333ad798cSAdam Langley (TCPOPT_NOP << 16) | 48433ad798cSAdam Langley (TCPOPT_SACK_PERM << 8) | 48533ad798cSAdam Langley TCPOLEN_SACK_PERM); 48633ad798cSAdam Langley } 48733ad798cSAdam Langley 488bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_WSCALE & options)) { 48933ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 49033ad798cSAdam Langley (TCPOPT_WINDOW << 16) | 49133ad798cSAdam Langley (TCPOLEN_WINDOW << 8) | 49233ad798cSAdam Langley opts->ws); 49333ad798cSAdam Langley } 49433ad798cSAdam Langley 49533ad798cSAdam Langley if (unlikely(opts->num_sack_blocks)) { 49633ad798cSAdam Langley struct tcp_sack_block *sp = tp->rx_opt.dsack ? 49733ad798cSAdam Langley tp->duplicate_sack : tp->selective_acks; 49840efc6faSStephen Hemminger int this_sack; 49940efc6faSStephen Hemminger 50040efc6faSStephen Hemminger *ptr++ = htonl((TCPOPT_NOP << 24) | 50140efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 50240efc6faSStephen Hemminger (TCPOPT_SACK << 8) | 50333ad798cSAdam Langley (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * 50440efc6faSStephen Hemminger TCPOLEN_SACK_PERBLOCK))); 5052de979bdSStephen Hemminger 50633ad798cSAdam Langley for (this_sack = 0; this_sack < opts->num_sack_blocks; 50733ad798cSAdam Langley ++this_sack) { 50840efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].start_seq); 50940efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].end_seq); 51040efc6faSStephen Hemminger } 5112de979bdSStephen Hemminger 51240efc6faSStephen Hemminger tp->rx_opt.dsack = 0; 51340efc6faSStephen Hemminger } 5142100c8d2SYuchung Cheng 5152100c8d2SYuchung Cheng if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) { 5162100c8d2SYuchung Cheng struct tcp_fastopen_cookie *foc = opts->fastopen_cookie; 5177f9b838bSDaniel Lee u8 *p = (u8 *)ptr; 5187f9b838bSDaniel Lee u32 len; /* Fast Open option length */ 5192100c8d2SYuchung Cheng 5207f9b838bSDaniel Lee if (foc->exp) { 5217f9b838bSDaniel Lee len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len; 5227f9b838bSDaniel Lee *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) | 5232100c8d2SYuchung Cheng TCPOPT_FASTOPEN_MAGIC); 5247f9b838bSDaniel Lee p += TCPOLEN_EXP_FASTOPEN_BASE; 5257f9b838bSDaniel Lee } else { 5267f9b838bSDaniel Lee len = TCPOLEN_FASTOPEN_BASE + foc->len; 5277f9b838bSDaniel Lee *p++ = TCPOPT_FASTOPEN; 5287f9b838bSDaniel Lee *p++ = len; 5292100c8d2SYuchung Cheng } 5307f9b838bSDaniel Lee 5317f9b838bSDaniel Lee memcpy(p, foc->val, foc->len); 5327f9b838bSDaniel Lee if ((len & 3) == 2) { 5337f9b838bSDaniel Lee p[foc->len] = TCPOPT_NOP; 5347f9b838bSDaniel Lee p[foc->len + 1] = TCPOPT_NOP; 5357f9b838bSDaniel Lee } 5367f9b838bSDaniel Lee ptr += (len + 3) >> 2; 5372100c8d2SYuchung Cheng } 53840efc6faSStephen Hemminger } 53940efc6faSStephen Hemminger 54067edfef7SAndi Kleen /* Compute TCP options for SYN packets. This is not the final 54167edfef7SAndi Kleen * network wire format yet. 54267edfef7SAndi Kleen */ 54395c96174SEric Dumazet static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb, 54433ad798cSAdam Langley struct tcp_out_options *opts, 545cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 546cf533ea5SEric Dumazet { 54733ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 54895c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 549783237e8SYuchung Cheng struct tcp_fastopen_request *fastopen = tp->fastopen_req; 55033ad798cSAdam Langley 551cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 55233ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 55333ad798cSAdam Langley if (*md5) { 55433ad798cSAdam Langley opts->options |= OPTION_MD5; 555bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 556cfb6eeb4SYOSHIFUJI Hideaki } 55733ad798cSAdam Langley #else 55833ad798cSAdam Langley *md5 = NULL; 559cfb6eeb4SYOSHIFUJI Hideaki #endif 56033ad798cSAdam Langley 56133ad798cSAdam Langley /* We always get an MSS option. The option bytes which will be seen in 56233ad798cSAdam Langley * normal data packets should timestamps be used, must be in the MSS 56333ad798cSAdam Langley * advertised. But we subtract them from tp->mss_cache so that 56433ad798cSAdam Langley * calculations in tcp_sendmsg are simpler etc. So account for this 56533ad798cSAdam Langley * fact here if necessary. If we don't do this correctly, as a 56633ad798cSAdam Langley * receiver we won't recognize data packets as being full sized when we 56733ad798cSAdam Langley * should, and thus we won't abide by the delayed ACK rules correctly. 56833ad798cSAdam Langley * SACKs don't matter, we never delay an ACK when we have any of those 56933ad798cSAdam Langley * going out. */ 57033ad798cSAdam Langley opts->mss = tcp_advertise_mss(sk); 571bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 57233ad798cSAdam Langley 57351456b29SIan Morris if (likely(sysctl_tcp_timestamps && !*md5)) { 57433ad798cSAdam Langley opts->options |= OPTION_TS; 5757faee5c0SEric Dumazet opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset; 57633ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 577bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 57833ad798cSAdam Langley } 579bb5b7c11SDavid S. Miller if (likely(sysctl_tcp_window_scaling)) { 58033ad798cSAdam Langley opts->ws = tp->rx_opt.rcv_wscale; 58189e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 582bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 58333ad798cSAdam Langley } 584bb5b7c11SDavid S. Miller if (likely(sysctl_tcp_sack)) { 58533ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 586b32d1310SDavid S. Miller if (unlikely(!(OPTION_TS & opts->options))) 587bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 58833ad798cSAdam Langley } 58933ad798cSAdam Langley 590783237e8SYuchung Cheng if (fastopen && fastopen->cookie.len >= 0) { 5912646c831SDaniel Lee u32 need = fastopen->cookie.len; 5922646c831SDaniel Lee 5932646c831SDaniel Lee need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE : 5942646c831SDaniel Lee TCPOLEN_FASTOPEN_BASE; 595783237e8SYuchung Cheng need = (need + 3) & ~3U; /* Align to 32 bits */ 596783237e8SYuchung Cheng if (remaining >= need) { 597783237e8SYuchung Cheng opts->options |= OPTION_FAST_OPEN_COOKIE; 598783237e8SYuchung Cheng opts->fastopen_cookie = &fastopen->cookie; 599783237e8SYuchung Cheng remaining -= need; 600783237e8SYuchung Cheng tp->syn_fastopen = 1; 6012646c831SDaniel Lee tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0; 602783237e8SYuchung Cheng } 603783237e8SYuchung Cheng } 604bd0388aeSWilliam Allen Simpson 605bd0388aeSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 60633ad798cSAdam Langley } 60733ad798cSAdam Langley 60867edfef7SAndi Kleen /* Set up TCP options for SYN-ACKs. */ 60937bfbddaSEric Dumazet static unsigned int tcp_synack_options(struct request_sock *req, 61095c96174SEric Dumazet unsigned int mss, struct sk_buff *skb, 61133ad798cSAdam Langley struct tcp_out_options *opts, 61280f03e27SEric Dumazet const struct tcp_md5sig_key *md5, 6138336886fSJerry Chu struct tcp_fastopen_cookie *foc) 6144957faadSWilliam Allen Simpson { 61533ad798cSAdam Langley struct inet_request_sock *ireq = inet_rsk(req); 61695c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 61733ad798cSAdam Langley 61833ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 61980f03e27SEric Dumazet if (md5) { 62033ad798cSAdam Langley opts->options |= OPTION_MD5; 6214957faadSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 6224957faadSWilliam Allen Simpson 6234957faadSWilliam Allen Simpson /* We can't fit any SACK blocks in a packet with MD5 + TS 6244957faadSWilliam Allen Simpson * options. There was discussion about disabling SACK 6254957faadSWilliam Allen Simpson * rather than TS in order to fit in better with old, 6264957faadSWilliam Allen Simpson * buggy kernels, but that was deemed to be unnecessary. 6274957faadSWilliam Allen Simpson */ 628de213e5eSEric Dumazet ireq->tstamp_ok &= !ireq->sack_ok; 62933ad798cSAdam Langley } 63033ad798cSAdam Langley #endif 63133ad798cSAdam Langley 6324957faadSWilliam Allen Simpson /* We always send an MSS option. */ 63333ad798cSAdam Langley opts->mss = mss; 6344957faadSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 63533ad798cSAdam Langley 63633ad798cSAdam Langley if (likely(ireq->wscale_ok)) { 63733ad798cSAdam Langley opts->ws = ireq->rcv_wscale; 63889e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 6394957faadSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 64033ad798cSAdam Langley } 641de213e5eSEric Dumazet if (likely(ireq->tstamp_ok)) { 64233ad798cSAdam Langley opts->options |= OPTION_TS; 6437faee5c0SEric Dumazet opts->tsval = tcp_skb_timestamp(skb); 64433ad798cSAdam Langley opts->tsecr = req->ts_recent; 6454957faadSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 64633ad798cSAdam Langley } 64733ad798cSAdam Langley if (likely(ireq->sack_ok)) { 64833ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 649de213e5eSEric Dumazet if (unlikely(!ireq->tstamp_ok)) 6504957faadSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 65133ad798cSAdam Langley } 6527f9b838bSDaniel Lee if (foc != NULL && foc->len >= 0) { 6537f9b838bSDaniel Lee u32 need = foc->len; 6547f9b838bSDaniel Lee 6557f9b838bSDaniel Lee need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE : 6567f9b838bSDaniel Lee TCPOLEN_FASTOPEN_BASE; 6578336886fSJerry Chu need = (need + 3) & ~3U; /* Align to 32 bits */ 6588336886fSJerry Chu if (remaining >= need) { 6598336886fSJerry Chu opts->options |= OPTION_FAST_OPEN_COOKIE; 6608336886fSJerry Chu opts->fastopen_cookie = foc; 6618336886fSJerry Chu remaining -= need; 6628336886fSJerry Chu } 6638336886fSJerry Chu } 6644957faadSWilliam Allen Simpson 6654957faadSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 66633ad798cSAdam Langley } 66733ad798cSAdam Langley 66867edfef7SAndi Kleen /* Compute TCP options for ESTABLISHED sockets. This is not the 66967edfef7SAndi Kleen * final wire format yet. 67067edfef7SAndi Kleen */ 67195c96174SEric Dumazet static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb, 67233ad798cSAdam Langley struct tcp_out_options *opts, 673cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 674cf533ea5SEric Dumazet { 67533ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 67695c96174SEric Dumazet unsigned int size = 0; 677cabeccbdSIlpo Järvinen unsigned int eff_sacks; 67833ad798cSAdam Langley 6795843ef42SAndi Kleen opts->options = 0; 6805843ef42SAndi Kleen 68133ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 68233ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 68333ad798cSAdam Langley if (unlikely(*md5)) { 68433ad798cSAdam Langley opts->options |= OPTION_MD5; 68533ad798cSAdam Langley size += TCPOLEN_MD5SIG_ALIGNED; 68633ad798cSAdam Langley } 68733ad798cSAdam Langley #else 68833ad798cSAdam Langley *md5 = NULL; 68933ad798cSAdam Langley #endif 69033ad798cSAdam Langley 69133ad798cSAdam Langley if (likely(tp->rx_opt.tstamp_ok)) { 69233ad798cSAdam Langley opts->options |= OPTION_TS; 6937faee5c0SEric Dumazet opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0; 69433ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 69533ad798cSAdam Langley size += TCPOLEN_TSTAMP_ALIGNED; 69633ad798cSAdam Langley } 69733ad798cSAdam Langley 698cabeccbdSIlpo Järvinen eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; 699cabeccbdSIlpo Järvinen if (unlikely(eff_sacks)) { 70095c96174SEric Dumazet const unsigned int remaining = MAX_TCP_OPTION_SPACE - size; 70133ad798cSAdam Langley opts->num_sack_blocks = 70295c96174SEric Dumazet min_t(unsigned int, eff_sacks, 70333ad798cSAdam Langley (remaining - TCPOLEN_SACK_BASE_ALIGNED) / 70433ad798cSAdam Langley TCPOLEN_SACK_PERBLOCK); 70533ad798cSAdam Langley size += TCPOLEN_SACK_BASE_ALIGNED + 70633ad798cSAdam Langley opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; 70733ad798cSAdam Langley } 70833ad798cSAdam Langley 70933ad798cSAdam Langley return size; 71040efc6faSStephen Hemminger } 7111da177e4SLinus Torvalds 71246d3ceabSEric Dumazet 71346d3ceabSEric Dumazet /* TCP SMALL QUEUES (TSQ) 71446d3ceabSEric Dumazet * 71546d3ceabSEric Dumazet * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev) 71646d3ceabSEric Dumazet * to reduce RTT and bufferbloat. 71746d3ceabSEric Dumazet * We do this using a special skb destructor (tcp_wfree). 71846d3ceabSEric Dumazet * 71946d3ceabSEric Dumazet * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb 72046d3ceabSEric Dumazet * needs to be reallocated in a driver. 7218e3bff96Sstephen hemminger * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc 72246d3ceabSEric Dumazet * 72346d3ceabSEric Dumazet * Since transmit from skb destructor is forbidden, we use a tasklet 72446d3ceabSEric Dumazet * to process all sockets that eventually need to send more skbs. 72546d3ceabSEric Dumazet * We use one tasklet per cpu, with its own queue of sockets. 72646d3ceabSEric Dumazet */ 72746d3ceabSEric Dumazet struct tsq_tasklet { 72846d3ceabSEric Dumazet struct tasklet_struct tasklet; 72946d3ceabSEric Dumazet struct list_head head; /* queue of tcp sockets */ 73046d3ceabSEric Dumazet }; 73146d3ceabSEric Dumazet static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet); 73246d3ceabSEric Dumazet 7336f458dfbSEric Dumazet static void tcp_tsq_handler(struct sock *sk) 7346f458dfbSEric Dumazet { 7356f458dfbSEric Dumazet if ((1 << sk->sk_state) & 7366f458dfbSEric Dumazet (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING | 7376f458dfbSEric Dumazet TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) 738bf06200eSJohn Ogness tcp_write_xmit(sk, tcp_current_mss(sk), tcp_sk(sk)->nonagle, 739bf06200eSJohn Ogness 0, GFP_ATOMIC); 7406f458dfbSEric Dumazet } 74146d3ceabSEric Dumazet /* 7428e3bff96Sstephen hemminger * One tasklet per cpu tries to send more skbs. 74346d3ceabSEric Dumazet * We run in tasklet context but need to disable irqs when 7448e3bff96Sstephen hemminger * transferring tsq->head because tcp_wfree() might 74546d3ceabSEric Dumazet * interrupt us (non NAPI drivers) 74646d3ceabSEric Dumazet */ 74746d3ceabSEric Dumazet static void tcp_tasklet_func(unsigned long data) 74846d3ceabSEric Dumazet { 74946d3ceabSEric Dumazet struct tsq_tasklet *tsq = (struct tsq_tasklet *)data; 75046d3ceabSEric Dumazet LIST_HEAD(list); 75146d3ceabSEric Dumazet unsigned long flags; 75246d3ceabSEric Dumazet struct list_head *q, *n; 75346d3ceabSEric Dumazet struct tcp_sock *tp; 75446d3ceabSEric Dumazet struct sock *sk; 75546d3ceabSEric Dumazet 75646d3ceabSEric Dumazet local_irq_save(flags); 75746d3ceabSEric Dumazet list_splice_init(&tsq->head, &list); 75846d3ceabSEric Dumazet local_irq_restore(flags); 75946d3ceabSEric Dumazet 76046d3ceabSEric Dumazet list_for_each_safe(q, n, &list) { 76146d3ceabSEric Dumazet tp = list_entry(q, struct tcp_sock, tsq_node); 76246d3ceabSEric Dumazet list_del(&tp->tsq_node); 76346d3ceabSEric Dumazet 76446d3ceabSEric Dumazet sk = (struct sock *)tp; 76546d3ceabSEric Dumazet bh_lock_sock(sk); 76646d3ceabSEric Dumazet 76746d3ceabSEric Dumazet if (!sock_owned_by_user(sk)) { 7686f458dfbSEric Dumazet tcp_tsq_handler(sk); 76946d3ceabSEric Dumazet } else { 77046d3ceabSEric Dumazet /* defer the work to tcp_release_cb() */ 7716f458dfbSEric Dumazet set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags); 77246d3ceabSEric Dumazet } 77346d3ceabSEric Dumazet bh_unlock_sock(sk); 77446d3ceabSEric Dumazet 77546d3ceabSEric Dumazet clear_bit(TSQ_QUEUED, &tp->tsq_flags); 77646d3ceabSEric Dumazet sk_free(sk); 77746d3ceabSEric Dumazet } 77846d3ceabSEric Dumazet } 77946d3ceabSEric Dumazet 7806f458dfbSEric Dumazet #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \ 7816f458dfbSEric Dumazet (1UL << TCP_WRITE_TIMER_DEFERRED) | \ 782563d34d0SEric Dumazet (1UL << TCP_DELACK_TIMER_DEFERRED) | \ 783563d34d0SEric Dumazet (1UL << TCP_MTU_REDUCED_DEFERRED)) 78446d3ceabSEric Dumazet /** 78546d3ceabSEric Dumazet * tcp_release_cb - tcp release_sock() callback 78646d3ceabSEric Dumazet * @sk: socket 78746d3ceabSEric Dumazet * 78846d3ceabSEric Dumazet * called from release_sock() to perform protocol dependent 78946d3ceabSEric Dumazet * actions before socket release. 79046d3ceabSEric Dumazet */ 79146d3ceabSEric Dumazet void tcp_release_cb(struct sock *sk) 79246d3ceabSEric Dumazet { 79346d3ceabSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 7946f458dfbSEric Dumazet unsigned long flags, nflags; 79546d3ceabSEric Dumazet 7966f458dfbSEric Dumazet /* perform an atomic operation only if at least one flag is set */ 7976f458dfbSEric Dumazet do { 7986f458dfbSEric Dumazet flags = tp->tsq_flags; 7996f458dfbSEric Dumazet if (!(flags & TCP_DEFERRED_ALL)) 8006f458dfbSEric Dumazet return; 8016f458dfbSEric Dumazet nflags = flags & ~TCP_DEFERRED_ALL; 8026f458dfbSEric Dumazet } while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags); 8036f458dfbSEric Dumazet 8046f458dfbSEric Dumazet if (flags & (1UL << TCP_TSQ_DEFERRED)) 8056f458dfbSEric Dumazet tcp_tsq_handler(sk); 8066f458dfbSEric Dumazet 807c3f9b018SEric Dumazet /* Here begins the tricky part : 808c3f9b018SEric Dumazet * We are called from release_sock() with : 809c3f9b018SEric Dumazet * 1) BH disabled 810c3f9b018SEric Dumazet * 2) sk_lock.slock spinlock held 811c3f9b018SEric Dumazet * 3) socket owned by us (sk->sk_lock.owned == 1) 812c3f9b018SEric Dumazet * 813c3f9b018SEric Dumazet * But following code is meant to be called from BH handlers, 814c3f9b018SEric Dumazet * so we should keep BH disabled, but early release socket ownership 815c3f9b018SEric Dumazet */ 816c3f9b018SEric Dumazet sock_release_ownership(sk); 817c3f9b018SEric Dumazet 818144d56e9SEric Dumazet if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) { 8196f458dfbSEric Dumazet tcp_write_timer_handler(sk); 820144d56e9SEric Dumazet __sock_put(sk); 821144d56e9SEric Dumazet } 822144d56e9SEric Dumazet if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) { 8236f458dfbSEric Dumazet tcp_delack_timer_handler(sk); 824144d56e9SEric Dumazet __sock_put(sk); 825144d56e9SEric Dumazet } 826144d56e9SEric Dumazet if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) { 8274fab9071SNeal Cardwell inet_csk(sk)->icsk_af_ops->mtu_reduced(sk); 828144d56e9SEric Dumazet __sock_put(sk); 829144d56e9SEric Dumazet } 83046d3ceabSEric Dumazet } 83146d3ceabSEric Dumazet EXPORT_SYMBOL(tcp_release_cb); 83246d3ceabSEric Dumazet 83346d3ceabSEric Dumazet void __init tcp_tasklet_init(void) 83446d3ceabSEric Dumazet { 83546d3ceabSEric Dumazet int i; 83646d3ceabSEric Dumazet 83746d3ceabSEric Dumazet for_each_possible_cpu(i) { 83846d3ceabSEric Dumazet struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i); 83946d3ceabSEric Dumazet 84046d3ceabSEric Dumazet INIT_LIST_HEAD(&tsq->head); 84146d3ceabSEric Dumazet tasklet_init(&tsq->tasklet, 84246d3ceabSEric Dumazet tcp_tasklet_func, 84346d3ceabSEric Dumazet (unsigned long)tsq); 84446d3ceabSEric Dumazet } 84546d3ceabSEric Dumazet } 84646d3ceabSEric Dumazet 84746d3ceabSEric Dumazet /* 84846d3ceabSEric Dumazet * Write buffer destructor automatically called from kfree_skb. 8498e3bff96Sstephen hemminger * We can't xmit new skbs from this context, as we might already 85046d3ceabSEric Dumazet * hold qdisc lock. 85146d3ceabSEric Dumazet */ 852d6a4a104SEric Dumazet void tcp_wfree(struct sk_buff *skb) 85346d3ceabSEric Dumazet { 85446d3ceabSEric Dumazet struct sock *sk = skb->sk; 85546d3ceabSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 8569b462d02SEric Dumazet int wmem; 8579b462d02SEric Dumazet 8589b462d02SEric Dumazet /* Keep one reference on sk_wmem_alloc. 8599b462d02SEric Dumazet * Will be released by sk_free() from here or tcp_tasklet_func() 8609b462d02SEric Dumazet */ 8619b462d02SEric Dumazet wmem = atomic_sub_return(skb->truesize - 1, &sk->sk_wmem_alloc); 8629b462d02SEric Dumazet 8639b462d02SEric Dumazet /* If this softirq is serviced by ksoftirqd, we are likely under stress. 8649b462d02SEric Dumazet * Wait until our queues (qdisc + devices) are drained. 8659b462d02SEric Dumazet * This gives : 8669b462d02SEric Dumazet * - less callbacks to tcp_write_xmit(), reducing stress (batches) 8679b462d02SEric Dumazet * - chance for incoming ACK (processed by another cpu maybe) 8689b462d02SEric Dumazet * to migrate this flow (skb->ooo_okay will be eventually set) 8699b462d02SEric Dumazet */ 8709b462d02SEric Dumazet if (wmem >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current) 8719b462d02SEric Dumazet goto out; 87246d3ceabSEric Dumazet 87346d3ceabSEric Dumazet if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) && 87446d3ceabSEric Dumazet !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) { 87546d3ceabSEric Dumazet unsigned long flags; 87646d3ceabSEric Dumazet struct tsq_tasklet *tsq; 87746d3ceabSEric Dumazet 87846d3ceabSEric Dumazet /* queue this socket to tasklet queue */ 87946d3ceabSEric Dumazet local_irq_save(flags); 880903ceff7SChristoph Lameter tsq = this_cpu_ptr(&tsq_tasklet); 88146d3ceabSEric Dumazet list_add(&tp->tsq_node, &tsq->head); 88246d3ceabSEric Dumazet tasklet_schedule(&tsq->tasklet); 88346d3ceabSEric Dumazet local_irq_restore(flags); 8849b462d02SEric Dumazet return; 88546d3ceabSEric Dumazet } 8869b462d02SEric Dumazet out: 8879b462d02SEric Dumazet sk_free(sk); 88846d3ceabSEric Dumazet } 88946d3ceabSEric Dumazet 8901da177e4SLinus Torvalds /* This routine actually transmits TCP packets queued in by 8911da177e4SLinus Torvalds * tcp_do_sendmsg(). This is used by both the initial 8921da177e4SLinus Torvalds * transmission and possible later retransmissions. 8931da177e4SLinus Torvalds * All SKB's seen here are completely headerless. It is our 8941da177e4SLinus Torvalds * job to build the TCP header, and pass the packet down to 8951da177e4SLinus Torvalds * IP so it can do the same plus pass the packet off to the 8961da177e4SLinus Torvalds * device. 8971da177e4SLinus Torvalds * 8981da177e4SLinus Torvalds * We are working here with either a clone of the original 8991da177e4SLinus Torvalds * SKB, or a fresh unique copy made by the retransmit engine. 9001da177e4SLinus Torvalds */ 901056834d9SIlpo Järvinen static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, 902056834d9SIlpo Järvinen gfp_t gfp_mask) 9031da177e4SLinus Torvalds { 9046687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 905dfb4b9dcSDavid S. Miller struct inet_sock *inet; 906dfb4b9dcSDavid S. Miller struct tcp_sock *tp; 907dfb4b9dcSDavid S. Miller struct tcp_skb_cb *tcb; 90833ad798cSAdam Langley struct tcp_out_options opts; 90995c96174SEric Dumazet unsigned int tcp_options_size, tcp_header_size; 910cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 9111da177e4SLinus Torvalds struct tcphdr *th; 9121da177e4SLinus Torvalds int err; 9131da177e4SLinus Torvalds 914dfb4b9dcSDavid S. Miller BUG_ON(!skb || !tcp_skb_pcount(skb)); 9156f094b9eSLawrence Brakmo tp = tcp_sk(sk); 916dfb4b9dcSDavid S. Miller 917ccdbb6e9SEric Dumazet if (clone_it) { 918740b0f18SEric Dumazet skb_mstamp_get(&skb->skb_mstamp); 9196f094b9eSLawrence Brakmo TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq 9206f094b9eSLawrence Brakmo - tp->snd_una; 921*b9f64820SYuchung Cheng tcp_rate_skb_sent(sk, skb); 922dfb4b9dcSDavid S. Miller 923dfb4b9dcSDavid S. Miller if (unlikely(skb_cloned(skb))) 924dfb4b9dcSDavid S. Miller skb = pskb_copy(skb, gfp_mask); 925dfb4b9dcSDavid S. Miller else 926dfb4b9dcSDavid S. Miller skb = skb_clone(skb, gfp_mask); 927dfb4b9dcSDavid S. Miller if (unlikely(!skb)) 928dfb4b9dcSDavid S. Miller return -ENOBUFS; 929dfb4b9dcSDavid S. Miller } 930dfb4b9dcSDavid S. Miller 931dfb4b9dcSDavid S. Miller inet = inet_sk(sk); 932dfb4b9dcSDavid S. Miller tcb = TCP_SKB_CB(skb); 93333ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 9341da177e4SLinus Torvalds 9354de075e0SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) 93633ad798cSAdam Langley tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); 93733ad798cSAdam Langley else 93833ad798cSAdam Langley tcp_options_size = tcp_established_options(sk, skb, &opts, 93933ad798cSAdam Langley &md5); 94033ad798cSAdam Langley tcp_header_size = tcp_options_size + sizeof(struct tcphdr); 9411da177e4SLinus Torvalds 942547669d4SEric Dumazet /* if no packet is in qdisc/device queue, then allow XPS to select 943b2532eb9SEric Dumazet * another queue. We can be called from tcp_tsq_handler() 944b2532eb9SEric Dumazet * which holds one reference to sk_wmem_alloc. 945b2532eb9SEric Dumazet * 946b2532eb9SEric Dumazet * TODO: Ideally, in-flight pure ACK packets should not matter here. 947b2532eb9SEric Dumazet * One way to get this would be to set skb->truesize = 2 on them. 948547669d4SEric Dumazet */ 949b2532eb9SEric Dumazet skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1); 9501da177e4SLinus Torvalds 951aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 952aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 95346d3ceabSEric Dumazet 95446d3ceabSEric Dumazet skb_orphan(skb); 95546d3ceabSEric Dumazet skb->sk = sk; 9561d2077acSEric Dumazet skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree; 957b73c3d0eSTom Herbert skb_set_hash_from_sk(skb, sk); 95846d3ceabSEric Dumazet atomic_add(skb->truesize, &sk->sk_wmem_alloc); 9591da177e4SLinus Torvalds 9601da177e4SLinus Torvalds /* Build TCP header and checksum it. */ 961ea1627c2SEric Dumazet th = (struct tcphdr *)skb->data; 962c720c7e8SEric Dumazet th->source = inet->inet_sport; 963c720c7e8SEric Dumazet th->dest = inet->inet_dport; 9641da177e4SLinus Torvalds th->seq = htonl(tcb->seq); 9651da177e4SLinus Torvalds th->ack_seq = htonl(tp->rcv_nxt); 966df7a3b07SAl Viro *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | 9674de075e0SEric Dumazet tcb->tcp_flags); 968dfb4b9dcSDavid S. Miller 9691da177e4SLinus Torvalds th->check = 0; 9701da177e4SLinus Torvalds th->urg_ptr = 0; 9711da177e4SLinus Torvalds 97233f5f57eSIlpo Järvinen /* The urg_mode check is necessary during a below snd_una win probe */ 9737691367dSHerbert Xu if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { 9747691367dSHerbert Xu if (before(tp->snd_up, tcb->seq + 0x10000)) { 9751da177e4SLinus Torvalds th->urg_ptr = htons(tp->snd_up - tcb->seq); 9761da177e4SLinus Torvalds th->urg = 1; 9777691367dSHerbert Xu } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { 9780eae88f3SEric Dumazet th->urg_ptr = htons(0xFFFF); 9797691367dSHerbert Xu th->urg = 1; 9807691367dSHerbert Xu } 9811da177e4SLinus Torvalds } 9821da177e4SLinus Torvalds 983bd0388aeSWilliam Allen Simpson tcp_options_write((__be32 *)(th + 1), tp, &opts); 98451466a75SEric Dumazet skb_shinfo(skb)->gso_type = sk->sk_gso_type; 985ea1627c2SEric Dumazet if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) { 986ea1627c2SEric Dumazet th->window = htons(tcp_select_window(sk)); 987ea1627c2SEric Dumazet tcp_ecn_send(sk, skb, th, tcp_header_size); 988ea1627c2SEric Dumazet } else { 989ea1627c2SEric Dumazet /* RFC1323: The window in SYN & SYN/ACK segments 990ea1627c2SEric Dumazet * is never scaled. 991ea1627c2SEric Dumazet */ 992ea1627c2SEric Dumazet th->window = htons(min(tp->rcv_wnd, 65535U)); 993ea1627c2SEric Dumazet } 994cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 995cfb6eeb4SYOSHIFUJI Hideaki /* Calculate the MD5 hash, as we have all we need now */ 996cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 997a465419bSEric Dumazet sk_nocaps_add(sk, NETIF_F_GSO_MASK); 998bd0388aeSWilliam Allen Simpson tp->af_specific->calc_md5_hash(opts.hash_location, 99939f8e58eSEric Dumazet md5, sk, skb); 1000cfb6eeb4SYOSHIFUJI Hideaki } 1001cfb6eeb4SYOSHIFUJI Hideaki #endif 1002cfb6eeb4SYOSHIFUJI Hideaki 1003bb296246SHerbert Xu icsk->icsk_af_ops->send_check(sk, skb); 10041da177e4SLinus Torvalds 10054de075e0SEric Dumazet if (likely(tcb->tcp_flags & TCPHDR_ACK)) 1006fc6415bcSDavid S. Miller tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); 10071da177e4SLinus Torvalds 1008a44d6eacSMartin KaFai Lau if (skb->len != tcp_header_size) { 1009cf533ea5SEric Dumazet tcp_event_data_sent(tp, sk); 1010a44d6eacSMartin KaFai Lau tp->data_segs_out += tcp_skb_pcount(skb); 1011a44d6eacSMartin KaFai Lau } 10121da177e4SLinus Torvalds 1013bd37a088SWei Yongjun if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) 1014aa2ea058STom Herbert TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, 1015aa2ea058STom Herbert tcp_skb_pcount(skb)); 10161da177e4SLinus Torvalds 10172efd055cSMarcelo Ricardo Leitner tp->segs_out += tcp_skb_pcount(skb); 1018f69ad292SEric Dumazet /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */ 1019cd7d8498SEric Dumazet skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb); 1020f69ad292SEric Dumazet skb_shinfo(skb)->gso_size = tcp_skb_mss(skb); 1021cd7d8498SEric Dumazet 10227faee5c0SEric Dumazet /* Our usage of tstamp should remain private */ 10237faee5c0SEric Dumazet skb->tstamp.tv64 = 0; 1024971f10ecSEric Dumazet 1025971f10ecSEric Dumazet /* Cleanup our debris for IP stacks */ 1026971f10ecSEric Dumazet memset(skb->cb, 0, max(sizeof(struct inet_skb_parm), 1027971f10ecSEric Dumazet sizeof(struct inet6_skb_parm))); 1028971f10ecSEric Dumazet 1029b0270e91SEric Dumazet err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl); 10307faee5c0SEric Dumazet 103183de47cdSHua Zhong if (likely(err <= 0)) 10321da177e4SLinus Torvalds return err; 10331da177e4SLinus Torvalds 10345ee2c941SChristoph Paasch tcp_enter_cwr(sk); 10351da177e4SLinus Torvalds 1036b9df3cb8SGerrit Renker return net_xmit_eval(err); 10371da177e4SLinus Torvalds } 10381da177e4SLinus Torvalds 103967edfef7SAndi Kleen /* This routine just queues the buffer for sending. 10401da177e4SLinus Torvalds * 10411da177e4SLinus Torvalds * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, 10421da177e4SLinus Torvalds * otherwise socket can stall. 10431da177e4SLinus Torvalds */ 10441da177e4SLinus Torvalds static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) 10451da177e4SLinus Torvalds { 10461da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 10471da177e4SLinus Torvalds 10481da177e4SLinus Torvalds /* Advance write_seq and place onto the write_queue. */ 10491da177e4SLinus Torvalds tp->write_seq = TCP_SKB_CB(skb)->end_seq; 1050f4a775d1SEric Dumazet __skb_header_release(skb); 1051fe067e8aSDavid S. Miller tcp_add_write_queue_tail(sk, skb); 10523ab224beSHideo Aoki sk->sk_wmem_queued += skb->truesize; 10533ab224beSHideo Aoki sk_mem_charge(sk, skb->truesize); 10541da177e4SLinus Torvalds } 10551da177e4SLinus Torvalds 105667edfef7SAndi Kleen /* Initialize TSO segments for a packet. */ 10575bbb432cSEric Dumazet static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now) 1058f6302d1dSDavid S. Miller { 10598f26fb1cSEric Dumazet if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) { 1060f6302d1dSDavid S. Miller /* Avoid the costly divide in the normal 1061f6302d1dSDavid S. Miller * non-TSO case. 1062f6302d1dSDavid S. Miller */ 1063cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, 1); 1064f69ad292SEric Dumazet TCP_SKB_CB(skb)->tcp_gso_size = 0; 1065f6302d1dSDavid S. Miller } else { 1066cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now)); 1067f69ad292SEric Dumazet TCP_SKB_CB(skb)->tcp_gso_size = mss_now; 10681da177e4SLinus Torvalds } 10691da177e4SLinus Torvalds } 10701da177e4SLinus Torvalds 107191fed7a1SIlpo Järvinen /* When a modification to fackets out becomes necessary, we need to check 107268f8353bSIlpo Järvinen * skb is counted to fackets_out or not. 107391fed7a1SIlpo Järvinen */ 1074cf533ea5SEric Dumazet static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb, 107591fed7a1SIlpo Järvinen int decr) 107691fed7a1SIlpo Järvinen { 1077a47e5a98SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1078a47e5a98SIlpo Järvinen 1079dc86967bSIlpo Järvinen if (!tp->sacked_out || tcp_is_reno(tp)) 108091fed7a1SIlpo Järvinen return; 108191fed7a1SIlpo Järvinen 10826859d494SIlpo Järvinen if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq)) 108391fed7a1SIlpo Järvinen tp->fackets_out -= decr; 108491fed7a1SIlpo Järvinen } 108591fed7a1SIlpo Järvinen 1086797108d1SIlpo Järvinen /* Pcount in the middle of the write queue got changed, we need to do various 1087797108d1SIlpo Järvinen * tweaks to fix counters 1088797108d1SIlpo Järvinen */ 1089cf533ea5SEric Dumazet static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) 1090797108d1SIlpo Järvinen { 1091797108d1SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1092797108d1SIlpo Järvinen 1093797108d1SIlpo Järvinen tp->packets_out -= decr; 1094797108d1SIlpo Järvinen 1095797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 1096797108d1SIlpo Järvinen tp->sacked_out -= decr; 1097797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 1098797108d1SIlpo Järvinen tp->retrans_out -= decr; 1099797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 1100797108d1SIlpo Järvinen tp->lost_out -= decr; 1101797108d1SIlpo Järvinen 1102797108d1SIlpo Järvinen /* Reno case is special. Sigh... */ 1103797108d1SIlpo Järvinen if (tcp_is_reno(tp) && decr > 0) 1104797108d1SIlpo Järvinen tp->sacked_out -= min_t(u32, tp->sacked_out, decr); 1105797108d1SIlpo Järvinen 1106797108d1SIlpo Järvinen tcp_adjust_fackets_out(sk, skb, decr); 1107797108d1SIlpo Järvinen 1108797108d1SIlpo Järvinen if (tp->lost_skb_hint && 1109797108d1SIlpo Järvinen before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && 111052cf3cc8SIlpo Järvinen (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))) 1111797108d1SIlpo Järvinen tp->lost_cnt_hint -= decr; 1112797108d1SIlpo Järvinen 1113797108d1SIlpo Järvinen tcp_verify_left_out(tp); 1114797108d1SIlpo Järvinen } 1115797108d1SIlpo Järvinen 11160a2cf20cSSoheil Hassas Yeganeh static bool tcp_has_tx_tstamp(const struct sk_buff *skb) 11170a2cf20cSSoheil Hassas Yeganeh { 11180a2cf20cSSoheil Hassas Yeganeh return TCP_SKB_CB(skb)->txstamp_ack || 11190a2cf20cSSoheil Hassas Yeganeh (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP); 11200a2cf20cSSoheil Hassas Yeganeh } 11210a2cf20cSSoheil Hassas Yeganeh 1122490cc7d0SWillem de Bruijn static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2) 1123490cc7d0SWillem de Bruijn { 1124490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo = skb_shinfo(skb); 1125490cc7d0SWillem de Bruijn 11260a2cf20cSSoheil Hassas Yeganeh if (unlikely(tcp_has_tx_tstamp(skb)) && 1127490cc7d0SWillem de Bruijn !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) { 1128490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo2 = skb_shinfo(skb2); 1129490cc7d0SWillem de Bruijn u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP; 1130490cc7d0SWillem de Bruijn 1131490cc7d0SWillem de Bruijn shinfo->tx_flags &= ~tsflags; 1132490cc7d0SWillem de Bruijn shinfo2->tx_flags |= tsflags; 1133490cc7d0SWillem de Bruijn swap(shinfo->tskey, shinfo2->tskey); 1134b51e13faSMartin KaFai Lau TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack; 1135b51e13faSMartin KaFai Lau TCP_SKB_CB(skb)->txstamp_ack = 0; 1136490cc7d0SWillem de Bruijn } 1137490cc7d0SWillem de Bruijn } 1138490cc7d0SWillem de Bruijn 1139a166140eSMartin KaFai Lau static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2) 1140a166140eSMartin KaFai Lau { 1141a166140eSMartin KaFai Lau TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor; 1142a166140eSMartin KaFai Lau TCP_SKB_CB(skb)->eor = 0; 1143a166140eSMartin KaFai Lau } 1144a166140eSMartin KaFai Lau 11451da177e4SLinus Torvalds /* Function to create two new TCP segments. Shrinks the given segment 11461da177e4SLinus Torvalds * to the specified size and appends a new segment with the rest of the 11471da177e4SLinus Torvalds * packet to the list. This won't be called frequently, I hope. 11481da177e4SLinus Torvalds * Remember, these are still headerless SKBs at this point. 11491da177e4SLinus Torvalds */ 1150056834d9SIlpo Järvinen int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, 11516cc55e09SOctavian Purdila unsigned int mss_now, gfp_t gfp) 11521da177e4SLinus Torvalds { 11531da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 11541da177e4SLinus Torvalds struct sk_buff *buff; 11556475be16SDavid S. Miller int nsize, old_factor; 1156b60b49eaSHerbert Xu int nlen; 11579ce01461SIlpo Järvinen u8 flags; 11581da177e4SLinus Torvalds 11592fceec13SIlpo Järvinen if (WARN_ON(len > skb->len)) 11602fceec13SIlpo Järvinen return -EINVAL; 11616a438bbeSStephen Hemminger 11621da177e4SLinus Torvalds nsize = skb_headlen(skb) - len; 11631da177e4SLinus Torvalds if (nsize < 0) 11641da177e4SLinus Torvalds nsize = 0; 11651da177e4SLinus Torvalds 11666cc55e09SOctavian Purdila if (skb_unclone(skb, gfp)) 11671da177e4SLinus Torvalds return -ENOMEM; 11681da177e4SLinus Torvalds 11691da177e4SLinus Torvalds /* Get a new skb... force flag on. */ 1170eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, nsize, gfp, true); 117151456b29SIan Morris if (!buff) 11721da177e4SLinus Torvalds return -ENOMEM; /* We'll just try again later. */ 1173ef5cb973SHerbert Xu 11743ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 11753ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1176b60b49eaSHerbert Xu nlen = skb->len - len - nsize; 1177b60b49eaSHerbert Xu buff->truesize += nlen; 1178b60b49eaSHerbert Xu skb->truesize -= nlen; 11791da177e4SLinus Torvalds 11801da177e4SLinus Torvalds /* Correct the sequence numbers. */ 11811da177e4SLinus Torvalds TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 11821da177e4SLinus Torvalds TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 11831da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 11841da177e4SLinus Torvalds 11851da177e4SLinus Torvalds /* PSH and FIN should only be set in the second packet. */ 11864de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 11874de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 11884de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1189e14c3cafSHerbert Xu TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; 1190a166140eSMartin KaFai Lau tcp_skb_fragment_eor(skb, buff); 11911da177e4SLinus Torvalds 119284fa7933SPatrick McHardy if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { 11931da177e4SLinus Torvalds /* Copy and checksum data tail into the new buffer. */ 1194056834d9SIlpo Järvinen buff->csum = csum_partial_copy_nocheck(skb->data + len, 1195056834d9SIlpo Järvinen skb_put(buff, nsize), 11961da177e4SLinus Torvalds nsize, 0); 11971da177e4SLinus Torvalds 11981da177e4SLinus Torvalds skb_trim(skb, len); 11991da177e4SLinus Torvalds 12001da177e4SLinus Torvalds skb->csum = csum_block_sub(skb->csum, buff->csum, len); 12011da177e4SLinus Torvalds } else { 120284fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 12031da177e4SLinus Torvalds skb_split(skb, buff, len); 12041da177e4SLinus Torvalds } 12051da177e4SLinus Torvalds 12061da177e4SLinus Torvalds buff->ip_summed = skb->ip_summed; 12071da177e4SLinus Torvalds 1208a61bbcf2SPatrick McHardy buff->tstamp = skb->tstamp; 1209490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 12101da177e4SLinus Torvalds 12116475be16SDavid S. Miller old_factor = tcp_skb_pcount(skb); 12126475be16SDavid S. Miller 12131da177e4SLinus Torvalds /* Fix up tso_factor for both original and new SKB. */ 12145bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 12155bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 12161da177e4SLinus Torvalds 1217*b9f64820SYuchung Cheng /* Update delivered info for the new segment */ 1218*b9f64820SYuchung Cheng TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx; 1219*b9f64820SYuchung Cheng 12206475be16SDavid S. Miller /* If this packet has been sent out already, we must 12216475be16SDavid S. Miller * adjust the various packet counters. 12226475be16SDavid S. Miller */ 1223cf0b450cSHerbert Xu if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 12246475be16SDavid S. Miller int diff = old_factor - tcp_skb_pcount(skb) - 12256475be16SDavid S. Miller tcp_skb_pcount(buff); 12261da177e4SLinus Torvalds 1227797108d1SIlpo Järvinen if (diff) 1228797108d1SIlpo Järvinen tcp_adjust_pcount(sk, skb, diff); 12291da177e4SLinus Torvalds } 12301da177e4SLinus Torvalds 12311da177e4SLinus Torvalds /* Link BUFF into the send queue. */ 1232f4a775d1SEric Dumazet __skb_header_release(buff); 1233fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 12341da177e4SLinus Torvalds 12351da177e4SLinus Torvalds return 0; 12361da177e4SLinus Torvalds } 12371da177e4SLinus Torvalds 12381da177e4SLinus Torvalds /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c 12391da177e4SLinus Torvalds * eventually). The difference is that pulled data not copied, but 12401da177e4SLinus Torvalds * immediately discarded. 12411da177e4SLinus Torvalds */ 1242f2911969SHerbert Xu ~{PmVHI~} static void __pskb_trim_head(struct sk_buff *skb, int len) 12431da177e4SLinus Torvalds { 12447b7fc97aSEric Dumazet struct skb_shared_info *shinfo; 12451da177e4SLinus Torvalds int i, k, eat; 12461da177e4SLinus Torvalds 12474fa48bf3SEric Dumazet eat = min_t(int, len, skb_headlen(skb)); 12484fa48bf3SEric Dumazet if (eat) { 12494fa48bf3SEric Dumazet __skb_pull(skb, eat); 12504fa48bf3SEric Dumazet len -= eat; 12514fa48bf3SEric Dumazet if (!len) 12524fa48bf3SEric Dumazet return; 12534fa48bf3SEric Dumazet } 12541da177e4SLinus Torvalds eat = len; 12551da177e4SLinus Torvalds k = 0; 12567b7fc97aSEric Dumazet shinfo = skb_shinfo(skb); 12577b7fc97aSEric Dumazet for (i = 0; i < shinfo->nr_frags; i++) { 12587b7fc97aSEric Dumazet int size = skb_frag_size(&shinfo->frags[i]); 12599e903e08SEric Dumazet 12609e903e08SEric Dumazet if (size <= eat) { 1261aff65da0SIan Campbell skb_frag_unref(skb, i); 12629e903e08SEric Dumazet eat -= size; 12631da177e4SLinus Torvalds } else { 12647b7fc97aSEric Dumazet shinfo->frags[k] = shinfo->frags[i]; 12651da177e4SLinus Torvalds if (eat) { 12667b7fc97aSEric Dumazet shinfo->frags[k].page_offset += eat; 12677b7fc97aSEric Dumazet skb_frag_size_sub(&shinfo->frags[k], eat); 12681da177e4SLinus Torvalds eat = 0; 12691da177e4SLinus Torvalds } 12701da177e4SLinus Torvalds k++; 12711da177e4SLinus Torvalds } 12721da177e4SLinus Torvalds } 12737b7fc97aSEric Dumazet shinfo->nr_frags = k; 12741da177e4SLinus Torvalds 127527a884dcSArnaldo Carvalho de Melo skb_reset_tail_pointer(skb); 12761da177e4SLinus Torvalds skb->data_len -= len; 12771da177e4SLinus Torvalds skb->len = skb->data_len; 12781da177e4SLinus Torvalds } 12791da177e4SLinus Torvalds 128067edfef7SAndi Kleen /* Remove acked data from a packet in the transmit queue. */ 12811da177e4SLinus Torvalds int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 12821da177e4SLinus Torvalds { 128314bbd6a5SPravin B Shelar if (skb_unclone(skb, GFP_ATOMIC)) 12841da177e4SLinus Torvalds return -ENOMEM; 12851da177e4SLinus Torvalds 12864fa48bf3SEric Dumazet __pskb_trim_head(skb, len); 12871da177e4SLinus Torvalds 12881da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq += len; 128984fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 12901da177e4SLinus Torvalds 12911da177e4SLinus Torvalds skb->truesize -= len; 12921da177e4SLinus Torvalds sk->sk_wmem_queued -= len; 12933ab224beSHideo Aoki sk_mem_uncharge(sk, len); 12941da177e4SLinus Torvalds sock_set_flag(sk, SOCK_QUEUE_SHRUNK); 12951da177e4SLinus Torvalds 12965b35e1e6SNeal Cardwell /* Any change of skb->len requires recalculation of tso factor. */ 12971da177e4SLinus Torvalds if (tcp_skb_pcount(skb) > 1) 12985bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb)); 12991da177e4SLinus Torvalds 13001da177e4SLinus Torvalds return 0; 13011da177e4SLinus Torvalds } 13021da177e4SLinus Torvalds 13031b63edd6SYuchung Cheng /* Calculate MSS not accounting any TCP options. */ 13041b63edd6SYuchung Cheng static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu) 13055d424d5aSJohn Heffner { 1306cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1307cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 13085d424d5aSJohn Heffner int mss_now; 13095d424d5aSJohn Heffner 13105d424d5aSJohn Heffner /* Calculate base mss without TCP options: 13115d424d5aSJohn Heffner It is MMS_S - sizeof(tcphdr) of rfc1122 13125d424d5aSJohn Heffner */ 13135d424d5aSJohn Heffner mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 13145d424d5aSJohn Heffner 131567469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 131667469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 131767469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 131867469601SEric Dumazet 131967469601SEric Dumazet if (dst && dst_allfrag(dst)) 132067469601SEric Dumazet mss_now -= icsk->icsk_af_ops->net_frag_header_len; 132167469601SEric Dumazet } 132267469601SEric Dumazet 13235d424d5aSJohn Heffner /* Clamp it (mss_clamp does not include tcp options) */ 13245d424d5aSJohn Heffner if (mss_now > tp->rx_opt.mss_clamp) 13255d424d5aSJohn Heffner mss_now = tp->rx_opt.mss_clamp; 13265d424d5aSJohn Heffner 13275d424d5aSJohn Heffner /* Now subtract optional transport overhead */ 13285d424d5aSJohn Heffner mss_now -= icsk->icsk_ext_hdr_len; 13295d424d5aSJohn Heffner 13305d424d5aSJohn Heffner /* Then reserve room for full set of TCP options and 8 bytes of data */ 13315d424d5aSJohn Heffner if (mss_now < 48) 13325d424d5aSJohn Heffner mss_now = 48; 13335d424d5aSJohn Heffner return mss_now; 13345d424d5aSJohn Heffner } 13355d424d5aSJohn Heffner 13361b63edd6SYuchung Cheng /* Calculate MSS. Not accounting for SACKs here. */ 13371b63edd6SYuchung Cheng int tcp_mtu_to_mss(struct sock *sk, int pmtu) 13381b63edd6SYuchung Cheng { 13391b63edd6SYuchung Cheng /* Subtract TCP options size, not including SACKs */ 13401b63edd6SYuchung Cheng return __tcp_mtu_to_mss(sk, pmtu) - 13411b63edd6SYuchung Cheng (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr)); 13421b63edd6SYuchung Cheng } 13431b63edd6SYuchung Cheng 13445d424d5aSJohn Heffner /* Inverse of above */ 134567469601SEric Dumazet int tcp_mss_to_mtu(struct sock *sk, int mss) 13465d424d5aSJohn Heffner { 1347cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1348cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 13495d424d5aSJohn Heffner int mtu; 13505d424d5aSJohn Heffner 13515d424d5aSJohn Heffner mtu = mss + 13525d424d5aSJohn Heffner tp->tcp_header_len + 13535d424d5aSJohn Heffner icsk->icsk_ext_hdr_len + 13545d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 13555d424d5aSJohn Heffner 135667469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 135767469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 135867469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 135967469601SEric Dumazet 136067469601SEric Dumazet if (dst && dst_allfrag(dst)) 136167469601SEric Dumazet mtu += icsk->icsk_af_ops->net_frag_header_len; 136267469601SEric Dumazet } 13635d424d5aSJohn Heffner return mtu; 13645d424d5aSJohn Heffner } 13655d424d5aSJohn Heffner 136667edfef7SAndi Kleen /* MTU probing init per socket */ 13675d424d5aSJohn Heffner void tcp_mtup_init(struct sock *sk) 13685d424d5aSJohn Heffner { 13695d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 13705d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 1371b0f9ca53SFan Du struct net *net = sock_net(sk); 13725d424d5aSJohn Heffner 1373b0f9ca53SFan Du icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1; 13745d424d5aSJohn Heffner icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 13755d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 1376b0f9ca53SFan Du icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss); 13775d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 137805cbc0dbSFan Du if (icsk->icsk_mtup.enabled) 137905cbc0dbSFan Du icsk->icsk_mtup.probe_timestamp = tcp_time_stamp; 13805d424d5aSJohn Heffner } 13814bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_mtup_init); 13825d424d5aSJohn Heffner 13831da177e4SLinus Torvalds /* This function synchronize snd mss to current pmtu/exthdr set. 13841da177e4SLinus Torvalds 13851da177e4SLinus Torvalds tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 13861da177e4SLinus Torvalds for TCP options, but includes only bare TCP header. 13871da177e4SLinus Torvalds 13881da177e4SLinus Torvalds tp->rx_opt.mss_clamp is mss negotiated at connection setup. 1389caa20d9aSStephen Hemminger It is minimum of user_mss and mss received with SYN. 13901da177e4SLinus Torvalds It also does not include TCP options. 13911da177e4SLinus Torvalds 1392d83d8461SArnaldo Carvalho de Melo inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. 13931da177e4SLinus Torvalds 13941da177e4SLinus Torvalds tp->mss_cache is current effective sending mss, including 13951da177e4SLinus Torvalds all tcp options except for SACKs. It is evaluated, 13961da177e4SLinus Torvalds taking into account current pmtu, but never exceeds 13971da177e4SLinus Torvalds tp->rx_opt.mss_clamp. 13981da177e4SLinus Torvalds 13991da177e4SLinus Torvalds NOTE1. rfc1122 clearly states that advertised MSS 14001da177e4SLinus Torvalds DOES NOT include either tcp or ip options. 14011da177e4SLinus Torvalds 1402d83d8461SArnaldo Carvalho de Melo NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache 1403d83d8461SArnaldo Carvalho de Melo are READ ONLY outside this function. --ANK (980731) 14041da177e4SLinus Torvalds */ 14051da177e4SLinus Torvalds unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 14061da177e4SLinus Torvalds { 14071da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1408d83d8461SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 14095d424d5aSJohn Heffner int mss_now; 14101da177e4SLinus Torvalds 14115d424d5aSJohn Heffner if (icsk->icsk_mtup.search_high > pmtu) 14125d424d5aSJohn Heffner icsk->icsk_mtup.search_high = pmtu; 14131da177e4SLinus Torvalds 14145d424d5aSJohn Heffner mss_now = tcp_mtu_to_mss(sk, pmtu); 1415409d22b4SIlpo Järvinen mss_now = tcp_bound_to_half_wnd(tp, mss_now); 14161da177e4SLinus Torvalds 14171da177e4SLinus Torvalds /* And store cached results */ 1418d83d8461SArnaldo Carvalho de Melo icsk->icsk_pmtu_cookie = pmtu; 14195d424d5aSJohn Heffner if (icsk->icsk_mtup.enabled) 14205d424d5aSJohn Heffner mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); 1421c1b4a7e6SDavid S. Miller tp->mss_cache = mss_now; 14221da177e4SLinus Torvalds 14231da177e4SLinus Torvalds return mss_now; 14241da177e4SLinus Torvalds } 14254bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_sync_mss); 14261da177e4SLinus Torvalds 14271da177e4SLinus Torvalds /* Compute the current effective MSS, taking SACKs and IP options, 14281da177e4SLinus Torvalds * and even PMTU discovery events into account. 14291da177e4SLinus Torvalds */ 14300c54b85fSIlpo Järvinen unsigned int tcp_current_mss(struct sock *sk) 14311da177e4SLinus Torvalds { 1432cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1433cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1434c1b4a7e6SDavid S. Miller u32 mss_now; 143595c96174SEric Dumazet unsigned int header_len; 143633ad798cSAdam Langley struct tcp_out_options opts; 143733ad798cSAdam Langley struct tcp_md5sig_key *md5; 14381da177e4SLinus Torvalds 1439c1b4a7e6SDavid S. Miller mss_now = tp->mss_cache; 1440c1b4a7e6SDavid S. Miller 14411da177e4SLinus Torvalds if (dst) { 14421da177e4SLinus Torvalds u32 mtu = dst_mtu(dst); 1443d83d8461SArnaldo Carvalho de Melo if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 14441da177e4SLinus Torvalds mss_now = tcp_sync_mss(sk, mtu); 14451da177e4SLinus Torvalds } 14461da177e4SLinus Torvalds 144733ad798cSAdam Langley header_len = tcp_established_options(sk, NULL, &opts, &md5) + 144833ad798cSAdam Langley sizeof(struct tcphdr); 144933ad798cSAdam Langley /* The mss_cache is sized based on tp->tcp_header_len, which assumes 145033ad798cSAdam Langley * some common options. If this is an odd packet (because we have SACK 145133ad798cSAdam Langley * blocks etc) then our calculated header_len will be different, and 145233ad798cSAdam Langley * we have to adjust mss_now correspondingly */ 145333ad798cSAdam Langley if (header_len != tp->tcp_header_len) { 145433ad798cSAdam Langley int delta = (int) header_len - tp->tcp_header_len; 145533ad798cSAdam Langley mss_now -= delta; 145633ad798cSAdam Langley } 1457cfb6eeb4SYOSHIFUJI Hideaki 14581da177e4SLinus Torvalds return mss_now; 14591da177e4SLinus Torvalds } 14601da177e4SLinus Torvalds 146186fd14adSWeiping Pan /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. 146286fd14adSWeiping Pan * As additional protections, we do not touch cwnd in retransmission phases, 146386fd14adSWeiping Pan * and if application hit its sndbuf limit recently. 146486fd14adSWeiping Pan */ 146586fd14adSWeiping Pan static void tcp_cwnd_application_limited(struct sock *sk) 1466a762a980SDavid S. Miller { 14679e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1468a762a980SDavid S. Miller 146986fd14adSWeiping Pan if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && 147086fd14adSWeiping Pan sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 147186fd14adSWeiping Pan /* Limited by application or receiver window. */ 147286fd14adSWeiping Pan u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); 147386fd14adSWeiping Pan u32 win_used = max(tp->snd_cwnd_used, init_win); 147486fd14adSWeiping Pan if (win_used < tp->snd_cwnd) { 147586fd14adSWeiping Pan tp->snd_ssthresh = tcp_current_ssthresh(sk); 147686fd14adSWeiping Pan tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1; 147786fd14adSWeiping Pan } 147886fd14adSWeiping Pan tp->snd_cwnd_used = 0; 147986fd14adSWeiping Pan } 148086fd14adSWeiping Pan tp->snd_cwnd_stamp = tcp_time_stamp; 148186fd14adSWeiping Pan } 148286fd14adSWeiping Pan 1483ca8a2263SNeal Cardwell static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited) 1484a762a980SDavid S. Miller { 1485a762a980SDavid S. Miller struct tcp_sock *tp = tcp_sk(sk); 1486a762a980SDavid S. Miller 1487ca8a2263SNeal Cardwell /* Track the maximum number of outstanding packets in each 1488ca8a2263SNeal Cardwell * window, and remember whether we were cwnd-limited then. 1489ca8a2263SNeal Cardwell */ 1490ca8a2263SNeal Cardwell if (!before(tp->snd_una, tp->max_packets_seq) || 1491ca8a2263SNeal Cardwell tp->packets_out > tp->max_packets_out) { 1492ca8a2263SNeal Cardwell tp->max_packets_out = tp->packets_out; 1493ca8a2263SNeal Cardwell tp->max_packets_seq = tp->snd_nxt; 1494ca8a2263SNeal Cardwell tp->is_cwnd_limited = is_cwnd_limited; 1495ca8a2263SNeal Cardwell } 1496e114a710SEric Dumazet 149724901551SEric Dumazet if (tcp_is_cwnd_limited(sk)) { 1498a762a980SDavid S. Miller /* Network is feed fully. */ 1499a762a980SDavid S. Miller tp->snd_cwnd_used = 0; 1500a762a980SDavid S. Miller tp->snd_cwnd_stamp = tcp_time_stamp; 1501a762a980SDavid S. Miller } else { 1502a762a980SDavid S. Miller /* Network starves. */ 1503a762a980SDavid S. Miller if (tp->packets_out > tp->snd_cwnd_used) 1504a762a980SDavid S. Miller tp->snd_cwnd_used = tp->packets_out; 1505a762a980SDavid S. Miller 150615d33c07SDavid S. Miller if (sysctl_tcp_slow_start_after_idle && 150715d33c07SDavid S. Miller (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto) 1508a762a980SDavid S. Miller tcp_cwnd_application_limited(sk); 1509a762a980SDavid S. Miller } 1510a762a980SDavid S. Miller } 1511a762a980SDavid S. Miller 1512d4589926SEric Dumazet /* Minshall's variant of the Nagle send check. */ 1513d4589926SEric Dumazet static bool tcp_minshall_check(const struct tcp_sock *tp) 1514d4589926SEric Dumazet { 1515d4589926SEric Dumazet return after(tp->snd_sml, tp->snd_una) && 1516d4589926SEric Dumazet !after(tp->snd_sml, tp->snd_nxt); 1517d4589926SEric Dumazet } 1518d4589926SEric Dumazet 1519d4589926SEric Dumazet /* Update snd_sml if this skb is under mss 1520d4589926SEric Dumazet * Note that a TSO packet might end with a sub-mss segment 1521d4589926SEric Dumazet * The test is really : 1522d4589926SEric Dumazet * if ((skb->len % mss) != 0) 1523d4589926SEric Dumazet * tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1524d4589926SEric Dumazet * But we can avoid doing the divide again given we already have 1525d4589926SEric Dumazet * skb_pcount = skb->len / mss_now 15260e3a4803SIlpo Järvinen */ 1527d4589926SEric Dumazet static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now, 1528d4589926SEric Dumazet const struct sk_buff *skb) 1529d4589926SEric Dumazet { 1530d4589926SEric Dumazet if (skb->len < tcp_skb_pcount(skb) * mss_now) 1531d4589926SEric Dumazet tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1532d4589926SEric Dumazet } 1533d4589926SEric Dumazet 1534d4589926SEric Dumazet /* Return false, if packet can be sent now without violation Nagle's rules: 1535d4589926SEric Dumazet * 1. It is full sized. (provided by caller in %partial bool) 1536d4589926SEric Dumazet * 2. Or it contains FIN. (already checked by caller) 1537d4589926SEric Dumazet * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. 1538d4589926SEric Dumazet * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 1539d4589926SEric Dumazet * With Minshall's modification: all sent small packets are ACKed. 1540d4589926SEric Dumazet */ 1541d4589926SEric Dumazet static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp, 1542cc93fc51SPeter Pan(潘卫平) int nonagle) 1543d4589926SEric Dumazet { 1544d4589926SEric Dumazet return partial && 1545d4589926SEric Dumazet ((nonagle & TCP_NAGLE_CORK) || 1546d4589926SEric Dumazet (!nonagle && tp->packets_out && tcp_minshall_check(tp))); 1547d4589926SEric Dumazet } 1548605ad7f1SEric Dumazet 1549605ad7f1SEric Dumazet /* Return how many segs we'd like on a TSO packet, 1550605ad7f1SEric Dumazet * to send one TSO packet per ms 1551605ad7f1SEric Dumazet */ 1552605ad7f1SEric Dumazet static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now) 1553605ad7f1SEric Dumazet { 1554605ad7f1SEric Dumazet u32 bytes, segs; 1555605ad7f1SEric Dumazet 1556605ad7f1SEric Dumazet bytes = min(sk->sk_pacing_rate >> 10, 1557605ad7f1SEric Dumazet sk->sk_gso_max_size - 1 - MAX_TCP_HEADER); 1558605ad7f1SEric Dumazet 1559605ad7f1SEric Dumazet /* Goal is to send at least one packet per ms, 1560605ad7f1SEric Dumazet * not one big TSO packet every 100 ms. 1561605ad7f1SEric Dumazet * This preserves ACK clocking and is consistent 1562605ad7f1SEric Dumazet * with tcp_tso_should_defer() heuristic. 1563605ad7f1SEric Dumazet */ 1564605ad7f1SEric Dumazet segs = max_t(u32, bytes / mss_now, sysctl_tcp_min_tso_segs); 1565605ad7f1SEric Dumazet 1566605ad7f1SEric Dumazet return min_t(u32, segs, sk->sk_gso_max_segs); 1567605ad7f1SEric Dumazet } 1568605ad7f1SEric Dumazet 1569d4589926SEric Dumazet /* Returns the portion of skb which can be sent right away */ 1570d4589926SEric Dumazet static unsigned int tcp_mss_split_point(const struct sock *sk, 1571d4589926SEric Dumazet const struct sk_buff *skb, 1572d4589926SEric Dumazet unsigned int mss_now, 1573d4589926SEric Dumazet unsigned int max_segs, 1574d4589926SEric Dumazet int nonagle) 1575c1b4a7e6SDavid S. Miller { 1576cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1577d4589926SEric Dumazet u32 partial, needed, window, max_len; 1578c1b4a7e6SDavid S. Miller 157990840defSIlpo Järvinen window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 15801485348dSBen Hutchings max_len = mss_now * max_segs; 15810e3a4803SIlpo Järvinen 15821485348dSBen Hutchings if (likely(max_len <= window && skb != tcp_write_queue_tail(sk))) 15831485348dSBen Hutchings return max_len; 15840e3a4803SIlpo Järvinen 15855ea3a748SIlpo Järvinen needed = min(skb->len, window); 15865ea3a748SIlpo Järvinen 15871485348dSBen Hutchings if (max_len <= needed) 15881485348dSBen Hutchings return max_len; 15890e3a4803SIlpo Järvinen 1590d4589926SEric Dumazet partial = needed % mss_now; 1591d4589926SEric Dumazet /* If last segment is not a full MSS, check if Nagle rules allow us 1592d4589926SEric Dumazet * to include this last segment in this skb. 1593d4589926SEric Dumazet * Otherwise, we'll split the skb at last MSS boundary 1594d4589926SEric Dumazet */ 1595cc93fc51SPeter Pan(潘卫平) if (tcp_nagle_check(partial != 0, tp, nonagle)) 1596d4589926SEric Dumazet return needed - partial; 1597d4589926SEric Dumazet 1598d4589926SEric Dumazet return needed; 1599c1b4a7e6SDavid S. Miller } 1600c1b4a7e6SDavid S. Miller 1601c1b4a7e6SDavid S. Miller /* Can at least one segment of SKB be sent right now, according to the 1602c1b4a7e6SDavid S. Miller * congestion window rules? If so, return how many segments are allowed. 1603c1b4a7e6SDavid S. Miller */ 1604cf533ea5SEric Dumazet static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, 1605cf533ea5SEric Dumazet const struct sk_buff *skb) 1606c1b4a7e6SDavid S. Miller { 1607d649a7a8SEric Dumazet u32 in_flight, cwnd, halfcwnd; 1608c1b4a7e6SDavid S. Miller 1609c1b4a7e6SDavid S. Miller /* Don't be strict about the congestion window for the final FIN. */ 16104de075e0SEric Dumazet if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && 16114de075e0SEric Dumazet tcp_skb_pcount(skb) == 1) 1612c1b4a7e6SDavid S. Miller return 1; 1613c1b4a7e6SDavid S. Miller 1614c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1615c1b4a7e6SDavid S. Miller cwnd = tp->snd_cwnd; 1616d649a7a8SEric Dumazet if (in_flight >= cwnd) 1617c1b4a7e6SDavid S. Miller return 0; 1618d649a7a8SEric Dumazet 1619d649a7a8SEric Dumazet /* For better scheduling, ensure we have at least 1620d649a7a8SEric Dumazet * 2 GSO packets in flight. 1621d649a7a8SEric Dumazet */ 1622d649a7a8SEric Dumazet halfcwnd = max(cwnd >> 1, 1U); 1623d649a7a8SEric Dumazet return min(halfcwnd, cwnd - in_flight); 1624c1b4a7e6SDavid S. Miller } 1625c1b4a7e6SDavid S. Miller 1626b595076aSUwe Kleine-König /* Initialize TSO state of a skb. 162767edfef7SAndi Kleen * This must be invoked the first time we consider transmitting 1628c1b4a7e6SDavid S. Miller * SKB onto the wire. 1629c1b4a7e6SDavid S. Miller */ 16305bbb432cSEric Dumazet static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now) 1631c1b4a7e6SDavid S. Miller { 1632c1b4a7e6SDavid S. Miller int tso_segs = tcp_skb_pcount(skb); 1633c1b4a7e6SDavid S. Miller 1634f8269a49SIlpo Järvinen if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { 16355bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 1636c1b4a7e6SDavid S. Miller tso_segs = tcp_skb_pcount(skb); 1637c1b4a7e6SDavid S. Miller } 1638c1b4a7e6SDavid S. Miller return tso_segs; 1639c1b4a7e6SDavid S. Miller } 1640c1b4a7e6SDavid S. Miller 1641c1b4a7e6SDavid S. Miller 1642a2a385d6SEric Dumazet /* Return true if the Nagle test allows this packet to be 1643c1b4a7e6SDavid S. Miller * sent now. 1644c1b4a7e6SDavid S. Miller */ 1645a2a385d6SEric Dumazet static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, 1646c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1647c1b4a7e6SDavid S. Miller { 1648c1b4a7e6SDavid S. Miller /* Nagle rule does not apply to frames, which sit in the middle of the 1649c1b4a7e6SDavid S. Miller * write_queue (they have no chances to get new data). 1650c1b4a7e6SDavid S. Miller * 1651c1b4a7e6SDavid S. Miller * This is implemented in the callers, where they modify the 'nonagle' 1652c1b4a7e6SDavid S. Miller * argument based upon the location of SKB in the send queue. 1653c1b4a7e6SDavid S. Miller */ 1654c1b4a7e6SDavid S. Miller if (nonagle & TCP_NAGLE_PUSH) 1655a2a385d6SEric Dumazet return true; 1656c1b4a7e6SDavid S. Miller 16579b44190dSYuchung Cheng /* Don't use the nagle rule for urgent data (or for the final FIN). */ 16589b44190dSYuchung Cheng if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) 1659a2a385d6SEric Dumazet return true; 1660c1b4a7e6SDavid S. Miller 1661cc93fc51SPeter Pan(潘卫平) if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle)) 1662a2a385d6SEric Dumazet return true; 1663c1b4a7e6SDavid S. Miller 1664a2a385d6SEric Dumazet return false; 1665c1b4a7e6SDavid S. Miller } 1666c1b4a7e6SDavid S. Miller 1667c1b4a7e6SDavid S. Miller /* Does at least the first segment of SKB fit into the send window? */ 1668a2a385d6SEric Dumazet static bool tcp_snd_wnd_test(const struct tcp_sock *tp, 1669a2a385d6SEric Dumazet const struct sk_buff *skb, 1670056834d9SIlpo Järvinen unsigned int cur_mss) 1671c1b4a7e6SDavid S. Miller { 1672c1b4a7e6SDavid S. Miller u32 end_seq = TCP_SKB_CB(skb)->end_seq; 1673c1b4a7e6SDavid S. Miller 1674c1b4a7e6SDavid S. Miller if (skb->len > cur_mss) 1675c1b4a7e6SDavid S. Miller end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 1676c1b4a7e6SDavid S. Miller 167790840defSIlpo Järvinen return !after(end_seq, tcp_wnd_end(tp)); 1678c1b4a7e6SDavid S. Miller } 1679c1b4a7e6SDavid S. Miller 1680fe067e8aSDavid S. Miller /* This checks if the data bearing packet SKB (usually tcp_send_head(sk)) 1681c1b4a7e6SDavid S. Miller * should be put on the wire right now. If so, it returns the number of 1682c1b4a7e6SDavid S. Miller * packets allowed by the congestion window. 1683c1b4a7e6SDavid S. Miller */ 1684cf533ea5SEric Dumazet static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb, 1685c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1686c1b4a7e6SDavid S. Miller { 1687cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1688c1b4a7e6SDavid S. Miller unsigned int cwnd_quota; 1689c1b4a7e6SDavid S. Miller 16905bbb432cSEric Dumazet tcp_init_tso_segs(skb, cur_mss); 1691c1b4a7e6SDavid S. Miller 1692c1b4a7e6SDavid S. Miller if (!tcp_nagle_test(tp, skb, cur_mss, nonagle)) 1693c1b4a7e6SDavid S. Miller return 0; 1694c1b4a7e6SDavid S. Miller 1695c1b4a7e6SDavid S. Miller cwnd_quota = tcp_cwnd_test(tp, skb); 1696056834d9SIlpo Järvinen if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss)) 1697c1b4a7e6SDavid S. Miller cwnd_quota = 0; 1698c1b4a7e6SDavid S. Miller 1699c1b4a7e6SDavid S. Miller return cwnd_quota; 1700c1b4a7e6SDavid S. Miller } 1701c1b4a7e6SDavid S. Miller 170267edfef7SAndi Kleen /* Test if sending is allowed right now. */ 1703a2a385d6SEric Dumazet bool tcp_may_send_now(struct sock *sk) 1704c1b4a7e6SDavid S. Miller { 1705cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1706fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 1707c1b4a7e6SDavid S. Miller 1708a02cec21SEric Dumazet return skb && 17090c54b85fSIlpo Järvinen tcp_snd_test(sk, skb, tcp_current_mss(sk), 1710c1b4a7e6SDavid S. Miller (tcp_skb_is_last(sk, skb) ? 1711a02cec21SEric Dumazet tp->nonagle : TCP_NAGLE_PUSH)); 1712c1b4a7e6SDavid S. Miller } 1713c1b4a7e6SDavid S. Miller 1714c1b4a7e6SDavid S. Miller /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 1715c1b4a7e6SDavid S. Miller * which is put after SKB on the list. It is very much like 1716c1b4a7e6SDavid S. Miller * tcp_fragment() except that it may make several kinds of assumptions 1717c1b4a7e6SDavid S. Miller * in order to speed up the splitting operation. In particular, we 1718c1b4a7e6SDavid S. Miller * know that all the data is in scatter-gather pages, and that the 1719c1b4a7e6SDavid S. Miller * packet has never been sent out before (and thus is not cloned). 1720c1b4a7e6SDavid S. Miller */ 1721056834d9SIlpo Järvinen static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, 1722c4ead4c5SEric Dumazet unsigned int mss_now, gfp_t gfp) 1723c1b4a7e6SDavid S. Miller { 1724c1b4a7e6SDavid S. Miller struct sk_buff *buff; 1725c1b4a7e6SDavid S. Miller int nlen = skb->len - len; 17269ce01461SIlpo Järvinen u8 flags; 1727c1b4a7e6SDavid S. Miller 1728c1b4a7e6SDavid S. Miller /* All of a TSO frame must be composed of paged data. */ 1729c8ac3774SHerbert Xu if (skb->len != skb->data_len) 17306cc55e09SOctavian Purdila return tcp_fragment(sk, skb, len, mss_now, gfp); 1731c1b4a7e6SDavid S. Miller 1732eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, gfp, true); 173351456b29SIan Morris if (unlikely(!buff)) 1734c1b4a7e6SDavid S. Miller return -ENOMEM; 1735c1b4a7e6SDavid S. Miller 17363ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 17373ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1738b60b49eaSHerbert Xu buff->truesize += nlen; 1739c1b4a7e6SDavid S. Miller skb->truesize -= nlen; 1740c1b4a7e6SDavid S. Miller 1741c1b4a7e6SDavid S. Miller /* Correct the sequence numbers. */ 1742c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 1743c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 1744c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 1745c1b4a7e6SDavid S. Miller 1746c1b4a7e6SDavid S. Miller /* PSH and FIN should only be set in the second packet. */ 17474de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 17484de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 17494de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1750c1b4a7e6SDavid S. Miller 1751c1b4a7e6SDavid S. Miller /* This packet was never sent out yet, so no SACK bits. */ 1752c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->sacked = 0; 1753c1b4a7e6SDavid S. Miller 1754a166140eSMartin KaFai Lau tcp_skb_fragment_eor(skb, buff); 1755a166140eSMartin KaFai Lau 175684fa7933SPatrick McHardy buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; 1757c1b4a7e6SDavid S. Miller skb_split(skb, buff, len); 1758490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 1759c1b4a7e6SDavid S. Miller 1760c1b4a7e6SDavid S. Miller /* Fix up tso_factor for both original and new SKB. */ 17615bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 17625bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 1763c1b4a7e6SDavid S. Miller 1764c1b4a7e6SDavid S. Miller /* Link BUFF into the send queue. */ 1765f4a775d1SEric Dumazet __skb_header_release(buff); 1766fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 1767c1b4a7e6SDavid S. Miller 1768c1b4a7e6SDavid S. Miller return 0; 1769c1b4a7e6SDavid S. Miller } 1770c1b4a7e6SDavid S. Miller 1771c1b4a7e6SDavid S. Miller /* Try to defer sending, if possible, in order to minimize the amount 1772c1b4a7e6SDavid S. Miller * of TSO splitting we do. View it as a kind of TSO Nagle test. 1773c1b4a7e6SDavid S. Miller * 1774c1b4a7e6SDavid S. Miller * This algorithm is from John Heffner. 1775c1b4a7e6SDavid S. Miller */ 1776ca8a2263SNeal Cardwell static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb, 1777605ad7f1SEric Dumazet bool *is_cwnd_limited, u32 max_segs) 1778c1b4a7e6SDavid S. Miller { 17796687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 178050c8339eSEric Dumazet u32 age, send_win, cong_win, limit, in_flight; 178150c8339eSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 178250c8339eSEric Dumazet struct skb_mstamp now; 178350c8339eSEric Dumazet struct sk_buff *head; 1784ad9f4f50SEric Dumazet int win_divisor; 1785c1b4a7e6SDavid S. Miller 17864de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 1787ae8064acSJohn Heffner goto send_now; 1788c1b4a7e6SDavid S. Miller 178999d7662aSEric Dumazet if (icsk->icsk_ca_state >= TCP_CA_Recovery) 1790ae8064acSJohn Heffner goto send_now; 1791ae8064acSJohn Heffner 17925f852eb5SEric Dumazet /* Avoid bursty behavior by allowing defer 17935f852eb5SEric Dumazet * only if the last write was recent. 17945f852eb5SEric Dumazet */ 17955f852eb5SEric Dumazet if ((s32)(tcp_time_stamp - tp->lsndtime) > 0) 1796ae8064acSJohn Heffner goto send_now; 1797908a75c1SDavid S. Miller 1798c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1799c1b4a7e6SDavid S. Miller 1800056834d9SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight)); 1801c1b4a7e6SDavid S. Miller 180290840defSIlpo Järvinen send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1803c1b4a7e6SDavid S. Miller 1804c1b4a7e6SDavid S. Miller /* From in_flight test above, we know that cwnd > in_flight. */ 1805c1b4a7e6SDavid S. Miller cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; 1806c1b4a7e6SDavid S. Miller 1807c1b4a7e6SDavid S. Miller limit = min(send_win, cong_win); 1808c1b4a7e6SDavid S. Miller 1809ba244fe9SDavid S. Miller /* If a full-sized TSO skb can be sent, do it. */ 1810605ad7f1SEric Dumazet if (limit >= max_segs * tp->mss_cache) 1811ae8064acSJohn Heffner goto send_now; 1812ba244fe9SDavid S. Miller 181362ad2761SIlpo Järvinen /* Middle in queue won't get any more data, full sendable already? */ 181462ad2761SIlpo Järvinen if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) 181562ad2761SIlpo Järvinen goto send_now; 181662ad2761SIlpo Järvinen 1817ad9f4f50SEric Dumazet win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor); 1818ad9f4f50SEric Dumazet if (win_divisor) { 1819c1b4a7e6SDavid S. Miller u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); 1820c1b4a7e6SDavid S. Miller 1821c1b4a7e6SDavid S. Miller /* If at least some fraction of a window is available, 1822c1b4a7e6SDavid S. Miller * just use it. 1823c1b4a7e6SDavid S. Miller */ 1824ad9f4f50SEric Dumazet chunk /= win_divisor; 1825c1b4a7e6SDavid S. Miller if (limit >= chunk) 1826ae8064acSJohn Heffner goto send_now; 1827c1b4a7e6SDavid S. Miller } else { 1828c1b4a7e6SDavid S. Miller /* Different approach, try not to defer past a single 1829c1b4a7e6SDavid S. Miller * ACK. Receiver should ACK every other full sized 1830c1b4a7e6SDavid S. Miller * frame, so if we have space for more than 3 frames 1831c1b4a7e6SDavid S. Miller * then send now. 1832c1b4a7e6SDavid S. Miller */ 18336b5a5c0dSNeal Cardwell if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) 1834ae8064acSJohn Heffner goto send_now; 1835c1b4a7e6SDavid S. Miller } 1836c1b4a7e6SDavid S. Miller 183750c8339eSEric Dumazet head = tcp_write_queue_head(sk); 183850c8339eSEric Dumazet skb_mstamp_get(&now); 183950c8339eSEric Dumazet age = skb_mstamp_us_delta(&now, &head->skb_mstamp); 184050c8339eSEric Dumazet /* If next ACK is likely to come too late (half srtt), do not defer */ 184150c8339eSEric Dumazet if (age < (tp->srtt_us >> 4)) 184250c8339eSEric Dumazet goto send_now; 184350c8339eSEric Dumazet 18445f852eb5SEric Dumazet /* Ok, it looks like it is advisable to defer. */ 1845ae8064acSJohn Heffner 1846d2e1339fSBendik Rønning Opstad if (cong_win < send_win && cong_win <= skb->len) 1847ca8a2263SNeal Cardwell *is_cwnd_limited = true; 1848ca8a2263SNeal Cardwell 1849a2a385d6SEric Dumazet return true; 1850ae8064acSJohn Heffner 1851ae8064acSJohn Heffner send_now: 1852a2a385d6SEric Dumazet return false; 1853c1b4a7e6SDavid S. Miller } 1854c1b4a7e6SDavid S. Miller 185505cbc0dbSFan Du static inline void tcp_mtu_check_reprobe(struct sock *sk) 185605cbc0dbSFan Du { 185705cbc0dbSFan Du struct inet_connection_sock *icsk = inet_csk(sk); 185805cbc0dbSFan Du struct tcp_sock *tp = tcp_sk(sk); 185905cbc0dbSFan Du struct net *net = sock_net(sk); 186005cbc0dbSFan Du u32 interval; 186105cbc0dbSFan Du s32 delta; 186205cbc0dbSFan Du 186305cbc0dbSFan Du interval = net->ipv4.sysctl_tcp_probe_interval; 186405cbc0dbSFan Du delta = tcp_time_stamp - icsk->icsk_mtup.probe_timestamp; 186505cbc0dbSFan Du if (unlikely(delta >= interval * HZ)) { 186605cbc0dbSFan Du int mss = tcp_current_mss(sk); 186705cbc0dbSFan Du 186805cbc0dbSFan Du /* Update current search range */ 186905cbc0dbSFan Du icsk->icsk_mtup.probe_size = 0; 187005cbc0dbSFan Du icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + 187105cbc0dbSFan Du sizeof(struct tcphdr) + 187205cbc0dbSFan Du icsk->icsk_af_ops->net_header_len; 187305cbc0dbSFan Du icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss); 187405cbc0dbSFan Du 187505cbc0dbSFan Du /* Update probe time stamp */ 187605cbc0dbSFan Du icsk->icsk_mtup.probe_timestamp = tcp_time_stamp; 187705cbc0dbSFan Du } 187805cbc0dbSFan Du } 187905cbc0dbSFan Du 18805d424d5aSJohn Heffner /* Create a new MTU probe if we are ready. 188167edfef7SAndi Kleen * MTU probe is regularly attempting to increase the path MTU by 188267edfef7SAndi Kleen * deliberately sending larger packets. This discovers routing 188367edfef7SAndi Kleen * changes resulting in larger path MTUs. 188467edfef7SAndi Kleen * 18855d424d5aSJohn Heffner * Returns 0 if we should wait to probe (no cwnd available), 18865d424d5aSJohn Heffner * 1 if a probe was sent, 1887056834d9SIlpo Järvinen * -1 otherwise 1888056834d9SIlpo Järvinen */ 18895d424d5aSJohn Heffner static int tcp_mtu_probe(struct sock *sk) 18905d424d5aSJohn Heffner { 18915d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 18925d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 18935d424d5aSJohn Heffner struct sk_buff *skb, *nskb, *next; 18946b58e0a5SFan Du struct net *net = sock_net(sk); 18955d424d5aSJohn Heffner int len; 18965d424d5aSJohn Heffner int probe_size; 189791cc17c0SIlpo Järvinen int size_needed; 18985d424d5aSJohn Heffner int copy; 18995d424d5aSJohn Heffner int mss_now; 19006b58e0a5SFan Du int interval; 19015d424d5aSJohn Heffner 19025d424d5aSJohn Heffner /* Not currently probing/verifying, 19035d424d5aSJohn Heffner * not in recovery, 19045d424d5aSJohn Heffner * have enough cwnd, and 19055d424d5aSJohn Heffner * not SACKing (the variable headers throw things off) */ 19065d424d5aSJohn Heffner if (!icsk->icsk_mtup.enabled || 19075d424d5aSJohn Heffner icsk->icsk_mtup.probe_size || 19085d424d5aSJohn Heffner inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 19095d424d5aSJohn Heffner tp->snd_cwnd < 11 || 1910cabeccbdSIlpo Järvinen tp->rx_opt.num_sacks || tp->rx_opt.dsack) 19115d424d5aSJohn Heffner return -1; 19125d424d5aSJohn Heffner 19136b58e0a5SFan Du /* Use binary search for probe_size between tcp_mss_base, 19146b58e0a5SFan Du * and current mss_clamp. if (search_high - search_low) 19156b58e0a5SFan Du * smaller than a threshold, backoff from probing. 19166b58e0a5SFan Du */ 19170c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 19186b58e0a5SFan Du probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high + 19196b58e0a5SFan Du icsk->icsk_mtup.search_low) >> 1); 192091cc17c0SIlpo Järvinen size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; 19216b58e0a5SFan Du interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low; 192205cbc0dbSFan Du /* When misfortune happens, we are reprobing actively, 192305cbc0dbSFan Du * and then reprobe timer has expired. We stick with current 192405cbc0dbSFan Du * probing process by not resetting search range to its orignal. 192505cbc0dbSFan Du */ 19266b58e0a5SFan Du if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) || 192705cbc0dbSFan Du interval < net->ipv4.sysctl_tcp_probe_threshold) { 192805cbc0dbSFan Du /* Check whether enough time has elaplased for 192905cbc0dbSFan Du * another round of probing. 193005cbc0dbSFan Du */ 193105cbc0dbSFan Du tcp_mtu_check_reprobe(sk); 19325d424d5aSJohn Heffner return -1; 19335d424d5aSJohn Heffner } 19345d424d5aSJohn Heffner 19355d424d5aSJohn Heffner /* Have enough data in the send queue to probe? */ 19367f9c33e5SIlpo Järvinen if (tp->write_seq - tp->snd_nxt < size_needed) 19375d424d5aSJohn Heffner return -1; 19385d424d5aSJohn Heffner 193991cc17c0SIlpo Järvinen if (tp->snd_wnd < size_needed) 19405d424d5aSJohn Heffner return -1; 194190840defSIlpo Järvinen if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) 19425d424d5aSJohn Heffner return 0; 19435d424d5aSJohn Heffner 1944d67c58e9SIlpo Järvinen /* Do we need to wait to drain cwnd? With none in flight, don't stall */ 1945d67c58e9SIlpo Järvinen if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { 1946d67c58e9SIlpo Järvinen if (!tcp_packets_in_flight(tp)) 19475d424d5aSJohn Heffner return -1; 19485d424d5aSJohn Heffner else 19495d424d5aSJohn Heffner return 0; 19505d424d5aSJohn Heffner } 19515d424d5aSJohn Heffner 19525d424d5aSJohn Heffner /* We're allowed to probe. Build it now. */ 1953eb934478SEric Dumazet nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false); 195451456b29SIan Morris if (!nskb) 19555d424d5aSJohn Heffner return -1; 19563ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 19573ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 19585d424d5aSJohn Heffner 1959fe067e8aSDavid S. Miller skb = tcp_send_head(sk); 19605d424d5aSJohn Heffner 19615d424d5aSJohn Heffner TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 19625d424d5aSJohn Heffner TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 19634de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; 19645d424d5aSJohn Heffner TCP_SKB_CB(nskb)->sacked = 0; 19655d424d5aSJohn Heffner nskb->csum = 0; 196684fa7933SPatrick McHardy nskb->ip_summed = skb->ip_summed; 19675d424d5aSJohn Heffner 196850c4817eSIlpo Järvinen tcp_insert_write_queue_before(nskb, skb, sk); 196950c4817eSIlpo Järvinen 19705d424d5aSJohn Heffner len = 0; 1971234b6860SIlpo Järvinen tcp_for_write_queue_from_safe(skb, next, sk) { 19725d424d5aSJohn Heffner copy = min_t(int, skb->len, probe_size - len); 19735d424d5aSJohn Heffner if (nskb->ip_summed) 19745d424d5aSJohn Heffner skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); 19755d424d5aSJohn Heffner else 19765d424d5aSJohn Heffner nskb->csum = skb_copy_and_csum_bits(skb, 0, 1977056834d9SIlpo Järvinen skb_put(nskb, copy), 1978056834d9SIlpo Järvinen copy, nskb->csum); 19795d424d5aSJohn Heffner 19805d424d5aSJohn Heffner if (skb->len <= copy) { 19815d424d5aSJohn Heffner /* We've eaten all the data from this skb. 19825d424d5aSJohn Heffner * Throw it away. */ 19834de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 1984fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 19853ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 19865d424d5aSJohn Heffner } else { 19874de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & 1988a3433f35SChangli Gao ~(TCPHDR_FIN|TCPHDR_PSH); 19895d424d5aSJohn Heffner if (!skb_shinfo(skb)->nr_frags) { 19905d424d5aSJohn Heffner skb_pull(skb, copy); 199184fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 1992056834d9SIlpo Järvinen skb->csum = csum_partial(skb->data, 1993056834d9SIlpo Järvinen skb->len, 0); 19945d424d5aSJohn Heffner } else { 19955d424d5aSJohn Heffner __pskb_trim_head(skb, copy); 19965bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 19975d424d5aSJohn Heffner } 19985d424d5aSJohn Heffner TCP_SKB_CB(skb)->seq += copy; 19995d424d5aSJohn Heffner } 20005d424d5aSJohn Heffner 20015d424d5aSJohn Heffner len += copy; 2002234b6860SIlpo Järvinen 2003234b6860SIlpo Järvinen if (len >= probe_size) 2004234b6860SIlpo Järvinen break; 20055d424d5aSJohn Heffner } 20065bbb432cSEric Dumazet tcp_init_tso_segs(nskb, nskb->len); 20075d424d5aSJohn Heffner 20085d424d5aSJohn Heffner /* We're ready to send. If this fails, the probe will 20097faee5c0SEric Dumazet * be resegmented into mss-sized pieces by tcp_write_xmit(). 20107faee5c0SEric Dumazet */ 20115d424d5aSJohn Heffner if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 20125d424d5aSJohn Heffner /* Decrement cwnd here because we are sending 20135d424d5aSJohn Heffner * effectively two packets. */ 20145d424d5aSJohn Heffner tp->snd_cwnd--; 201566f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, nskb); 20165d424d5aSJohn Heffner 20175d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 20180e7b1368SJohn Heffner tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 20190e7b1368SJohn Heffner tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 20205d424d5aSJohn Heffner 20215d424d5aSJohn Heffner return 1; 20225d424d5aSJohn Heffner } 20235d424d5aSJohn Heffner 20245d424d5aSJohn Heffner return -1; 20255d424d5aSJohn Heffner } 20265d424d5aSJohn Heffner 20271da177e4SLinus Torvalds /* This routine writes packets to the network. It advances the 20281da177e4SLinus Torvalds * send_head. This happens as incoming acks open up the remote 20291da177e4SLinus Torvalds * window for us. 20301da177e4SLinus Torvalds * 2031f8269a49SIlpo Järvinen * LARGESEND note: !tcp_urg_mode is overkill, only frames between 2032f8269a49SIlpo Järvinen * snd_up-64k-mss .. snd_up cannot be large. However, taking into 2033f8269a49SIlpo Järvinen * account rare use of URG, this is not a big flaw. 2034f8269a49SIlpo Järvinen * 20356ba8a3b1SNandita Dukkipati * Send at most one packet when push_one > 0. Temporarily ignore 20366ba8a3b1SNandita Dukkipati * cwnd limit to force at most one packet out when push_one == 2. 20376ba8a3b1SNandita Dukkipati 2038a2a385d6SEric Dumazet * Returns true, if no segments are in flight and we have queued segments, 2039a2a385d6SEric Dumazet * but cannot send anything now because of SWS or another problem. 20401da177e4SLinus Torvalds */ 2041a2a385d6SEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 2042d5dd9175SIlpo Järvinen int push_one, gfp_t gfp) 20431da177e4SLinus Torvalds { 20441da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 204592df7b51SDavid S. Miller struct sk_buff *skb; 2046c1b4a7e6SDavid S. Miller unsigned int tso_segs, sent_pkts; 2047c1b4a7e6SDavid S. Miller int cwnd_quota; 20485d424d5aSJohn Heffner int result; 2049ca8a2263SNeal Cardwell bool is_cwnd_limited = false; 2050605ad7f1SEric Dumazet u32 max_segs; 20511da177e4SLinus Torvalds 2052c1b4a7e6SDavid S. Miller sent_pkts = 0; 20535d424d5aSJohn Heffner 2054d5dd9175SIlpo Järvinen if (!push_one) { 20555d424d5aSJohn Heffner /* Do MTU probing. */ 2056d5dd9175SIlpo Järvinen result = tcp_mtu_probe(sk); 2057d5dd9175SIlpo Järvinen if (!result) { 2058a2a385d6SEric Dumazet return false; 20595d424d5aSJohn Heffner } else if (result > 0) { 20605d424d5aSJohn Heffner sent_pkts = 1; 20615d424d5aSJohn Heffner } 2062d5dd9175SIlpo Järvinen } 20635d424d5aSJohn Heffner 2064605ad7f1SEric Dumazet max_segs = tcp_tso_autosize(sk, mss_now); 2065fe067e8aSDavid S. Miller while ((skb = tcp_send_head(sk))) { 2066c8ac3774SHerbert Xu unsigned int limit; 2067c8ac3774SHerbert Xu 20685bbb432cSEric Dumazet tso_segs = tcp_init_tso_segs(skb, mss_now); 2069c1b4a7e6SDavid S. Miller BUG_ON(!tso_segs); 2070c1b4a7e6SDavid S. Miller 20719d186cacSAndrey Vagin if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) { 20727faee5c0SEric Dumazet /* "skb_mstamp" is used as a start point for the retransmit timer */ 20737faee5c0SEric Dumazet skb_mstamp_get(&skb->skb_mstamp); 2074ec342325SAndrew Vagin goto repair; /* Skip network transmission */ 20759d186cacSAndrey Vagin } 2076ec342325SAndrew Vagin 2077b68e9f85SHerbert Xu cwnd_quota = tcp_cwnd_test(tp, skb); 20786ba8a3b1SNandita Dukkipati if (!cwnd_quota) { 20796ba8a3b1SNandita Dukkipati if (push_one == 2) 20806ba8a3b1SNandita Dukkipati /* Force out a loss probe pkt. */ 20816ba8a3b1SNandita Dukkipati cwnd_quota = 1; 20826ba8a3b1SNandita Dukkipati else 2083b68e9f85SHerbert Xu break; 20846ba8a3b1SNandita Dukkipati } 2085b68e9f85SHerbert Xu 2086b68e9f85SHerbert Xu if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) 2087b68e9f85SHerbert Xu break; 2088b68e9f85SHerbert Xu 2089d6a4e26aSEric Dumazet if (tso_segs == 1) { 2090aa93466bSDavid S. Miller if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 2091aa93466bSDavid S. Miller (tcp_skb_is_last(sk, skb) ? 2092aa93466bSDavid S. Miller nonagle : TCP_NAGLE_PUSH)))) 2093aa93466bSDavid S. Miller break; 2094c1b4a7e6SDavid S. Miller } else { 2095ca8a2263SNeal Cardwell if (!push_one && 2096605ad7f1SEric Dumazet tcp_tso_should_defer(sk, skb, &is_cwnd_limited, 2097605ad7f1SEric Dumazet max_segs)) 2098aa93466bSDavid S. Miller break; 2099c1b4a7e6SDavid S. Miller } 2100aa93466bSDavid S. Miller 2101605ad7f1SEric Dumazet limit = mss_now; 2102d6a4e26aSEric Dumazet if (tso_segs > 1 && !tcp_urg_mode(tp)) 2103605ad7f1SEric Dumazet limit = tcp_mss_split_point(sk, skb, mss_now, 2104605ad7f1SEric Dumazet min_t(unsigned int, 2105605ad7f1SEric Dumazet cwnd_quota, 2106605ad7f1SEric Dumazet max_segs), 2107605ad7f1SEric Dumazet nonagle); 2108605ad7f1SEric Dumazet 2109605ad7f1SEric Dumazet if (skb->len > limit && 2110605ad7f1SEric Dumazet unlikely(tso_fragment(sk, skb, limit, mss_now, gfp))) 2111605ad7f1SEric Dumazet break; 2112605ad7f1SEric Dumazet 2113c9eeec26SEric Dumazet /* TCP Small Queues : 2114c9eeec26SEric Dumazet * Control number of packets in qdisc/devices to two packets / or ~1 ms. 2115c9eeec26SEric Dumazet * This allows for : 2116c9eeec26SEric Dumazet * - better RTT estimation and ACK scheduling 2117c9eeec26SEric Dumazet * - faster recovery 2118c9eeec26SEric Dumazet * - high rates 211998e09386SEric Dumazet * Alas, some drivers / subsystems require a fair amount 212098e09386SEric Dumazet * of queued bytes to ensure line rate. 212198e09386SEric Dumazet * One example is wifi aggregation (802.11 AMPDU) 212246d3ceabSEric Dumazet */ 2123605ad7f1SEric Dumazet limit = max(2 * skb->truesize, sk->sk_pacing_rate >> 10); 2124605ad7f1SEric Dumazet limit = min_t(u32, limit, sysctl_tcp_limit_output_bytes); 2125c9eeec26SEric Dumazet 2126c9eeec26SEric Dumazet if (atomic_read(&sk->sk_wmem_alloc) > limit) { 212746d3ceabSEric Dumazet set_bit(TSQ_THROTTLED, &tp->tsq_flags); 2128bf06200eSJohn Ogness /* It is possible TX completion already happened 2129bf06200eSJohn Ogness * before we set TSQ_THROTTLED, so we must 2130bf06200eSJohn Ogness * test again the condition. 2131bf06200eSJohn Ogness */ 21324e857c58SPeter Zijlstra smp_mb__after_atomic(); 2133bf06200eSJohn Ogness if (atomic_read(&sk->sk_wmem_alloc) > limit) 213446d3ceabSEric Dumazet break; 213546d3ceabSEric Dumazet } 2136c9eeec26SEric Dumazet 2137d5dd9175SIlpo Järvinen if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) 21381da177e4SLinus Torvalds break; 21391da177e4SLinus Torvalds 2140ec342325SAndrew Vagin repair: 21411da177e4SLinus Torvalds /* Advance the send_head. This one is sent out. 21421da177e4SLinus Torvalds * This call will increment packets_out. 21431da177e4SLinus Torvalds */ 214466f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 21451da177e4SLinus Torvalds 21461da177e4SLinus Torvalds tcp_minshall_update(tp, mss_now, skb); 2147a262f0cdSNandita Dukkipati sent_pkts += tcp_skb_pcount(skb); 2148d5dd9175SIlpo Järvinen 2149d5dd9175SIlpo Järvinen if (push_one) 2150d5dd9175SIlpo Järvinen break; 21511da177e4SLinus Torvalds } 21521da177e4SLinus Torvalds 2153aa93466bSDavid S. Miller if (likely(sent_pkts)) { 2154684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 2155684bad11SYuchung Cheng tp->prr_out += sent_pkts; 21566ba8a3b1SNandita Dukkipati 21576ba8a3b1SNandita Dukkipati /* Send one loss probe per tail loss episode. */ 21586ba8a3b1SNandita Dukkipati if (push_one != 2) 21596ba8a3b1SNandita Dukkipati tcp_schedule_loss_probe(sk); 2160d2e1339fSBendik Rønning Opstad is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd); 2161ca8a2263SNeal Cardwell tcp_cwnd_validate(sk, is_cwnd_limited); 2162a2a385d6SEric Dumazet return false; 21631da177e4SLinus Torvalds } 2164b340b264SYuchung Cheng return !tp->packets_out && tcp_send_head(sk); 21656ba8a3b1SNandita Dukkipati } 21666ba8a3b1SNandita Dukkipati 21676ba8a3b1SNandita Dukkipati bool tcp_schedule_loss_probe(struct sock *sk) 21686ba8a3b1SNandita Dukkipati { 21696ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 21706ba8a3b1SNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 21716ba8a3b1SNandita Dukkipati u32 timeout, tlp_time_stamp, rto_time_stamp; 2172740b0f18SEric Dumazet u32 rtt = usecs_to_jiffies(tp->srtt_us >> 3); 21736ba8a3b1SNandita Dukkipati 21746ba8a3b1SNandita Dukkipati if (WARN_ON(icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS)) 21756ba8a3b1SNandita Dukkipati return false; 21766ba8a3b1SNandita Dukkipati /* No consecutive loss probes. */ 21776ba8a3b1SNandita Dukkipati if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) { 21786ba8a3b1SNandita Dukkipati tcp_rearm_rto(sk); 21796ba8a3b1SNandita Dukkipati return false; 21806ba8a3b1SNandita Dukkipati } 21816ba8a3b1SNandita Dukkipati /* Don't do any loss probe on a Fast Open connection before 3WHS 21826ba8a3b1SNandita Dukkipati * finishes. 21836ba8a3b1SNandita Dukkipati */ 2184f9b99582SYuchung Cheng if (tp->fastopen_rsk) 21856ba8a3b1SNandita Dukkipati return false; 21866ba8a3b1SNandita Dukkipati 21876ba8a3b1SNandita Dukkipati /* TLP is only scheduled when next timer event is RTO. */ 21886ba8a3b1SNandita Dukkipati if (icsk->icsk_pending != ICSK_TIME_RETRANS) 21896ba8a3b1SNandita Dukkipati return false; 21906ba8a3b1SNandita Dukkipati 21916ba8a3b1SNandita Dukkipati /* Schedule a loss probe in 2*RTT for SACK capable connections 21926ba8a3b1SNandita Dukkipati * in Open state, that are either limited by cwnd or application. 21936ba8a3b1SNandita Dukkipati */ 2194f9b99582SYuchung Cheng if (sysctl_tcp_early_retrans < 3 || !tp->packets_out || 21956ba8a3b1SNandita Dukkipati !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open) 21966ba8a3b1SNandita Dukkipati return false; 21976ba8a3b1SNandita Dukkipati 21986ba8a3b1SNandita Dukkipati if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) && 21996ba8a3b1SNandita Dukkipati tcp_send_head(sk)) 22006ba8a3b1SNandita Dukkipati return false; 22016ba8a3b1SNandita Dukkipati 22026ba8a3b1SNandita Dukkipati /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account 2203f9b99582SYuchung Cheng * for delayed ack when there's one outstanding packet. If no RTT 2204f9b99582SYuchung Cheng * sample is available then probe after TCP_TIMEOUT_INIT. 22056ba8a3b1SNandita Dukkipati */ 2206f9b99582SYuchung Cheng timeout = rtt << 1 ? : TCP_TIMEOUT_INIT; 22076ba8a3b1SNandita Dukkipati if (tp->packets_out == 1) 22086ba8a3b1SNandita Dukkipati timeout = max_t(u32, timeout, 22096ba8a3b1SNandita Dukkipati (rtt + (rtt >> 1) + TCP_DELACK_MAX)); 22106ba8a3b1SNandita Dukkipati timeout = max_t(u32, timeout, msecs_to_jiffies(10)); 22116ba8a3b1SNandita Dukkipati 22126ba8a3b1SNandita Dukkipati /* If RTO is shorter, just schedule TLP in its place. */ 22136ba8a3b1SNandita Dukkipati tlp_time_stamp = tcp_time_stamp + timeout; 22146ba8a3b1SNandita Dukkipati rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout; 22156ba8a3b1SNandita Dukkipati if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) { 22166ba8a3b1SNandita Dukkipati s32 delta = rto_time_stamp - tcp_time_stamp; 22176ba8a3b1SNandita Dukkipati if (delta > 0) 22186ba8a3b1SNandita Dukkipati timeout = delta; 22196ba8a3b1SNandita Dukkipati } 22206ba8a3b1SNandita Dukkipati 22216ba8a3b1SNandita Dukkipati inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, 22226ba8a3b1SNandita Dukkipati TCP_RTO_MAX); 22236ba8a3b1SNandita Dukkipati return true; 22246ba8a3b1SNandita Dukkipati } 22256ba8a3b1SNandita Dukkipati 22261f3279aeSEric Dumazet /* Thanks to skb fast clones, we can detect if a prior transmit of 22271f3279aeSEric Dumazet * a packet is still in a qdisc or driver queue. 22281f3279aeSEric Dumazet * In this case, there is very little point doing a retransmit ! 22291f3279aeSEric Dumazet */ 22301f3279aeSEric Dumazet static bool skb_still_in_host_queue(const struct sock *sk, 22311f3279aeSEric Dumazet const struct sk_buff *skb) 22321f3279aeSEric Dumazet { 223339bb5e62SEric Dumazet if (unlikely(skb_fclone_busy(sk, skb))) { 2234c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), 22351f3279aeSEric Dumazet LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES); 22361f3279aeSEric Dumazet return true; 22371f3279aeSEric Dumazet } 22381f3279aeSEric Dumazet return false; 22391f3279aeSEric Dumazet } 22401f3279aeSEric Dumazet 2241b340b264SYuchung Cheng /* When probe timeout (PTO) fires, try send a new segment if possible, else 22426ba8a3b1SNandita Dukkipati * retransmit the last segment. 22436ba8a3b1SNandita Dukkipati */ 22446ba8a3b1SNandita Dukkipati void tcp_send_loss_probe(struct sock *sk) 22456ba8a3b1SNandita Dukkipati { 22469b717a8dSNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 22476ba8a3b1SNandita Dukkipati struct sk_buff *skb; 22486ba8a3b1SNandita Dukkipati int pcount; 22496ba8a3b1SNandita Dukkipati int mss = tcp_current_mss(sk); 22506ba8a3b1SNandita Dukkipati 2251b340b264SYuchung Cheng skb = tcp_send_head(sk); 2252b340b264SYuchung Cheng if (skb) { 2253b340b264SYuchung Cheng if (tcp_snd_wnd_test(tp, skb, mss)) { 2254b340b264SYuchung Cheng pcount = tp->packets_out; 2255b340b264SYuchung Cheng tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC); 2256b340b264SYuchung Cheng if (tp->packets_out > pcount) 2257b340b264SYuchung Cheng goto probe_sent; 22586ba8a3b1SNandita Dukkipati goto rearm_timer; 22596ba8a3b1SNandita Dukkipati } 2260b340b264SYuchung Cheng skb = tcp_write_queue_prev(sk, skb); 2261b340b264SYuchung Cheng } else { 2262b340b264SYuchung Cheng skb = tcp_write_queue_tail(sk); 2263b340b264SYuchung Cheng } 22646ba8a3b1SNandita Dukkipati 22659b717a8dSNandita Dukkipati /* At most one outstanding TLP retransmission. */ 22669b717a8dSNandita Dukkipati if (tp->tlp_high_seq) 22679b717a8dSNandita Dukkipati goto rearm_timer; 22689b717a8dSNandita Dukkipati 22696ba8a3b1SNandita Dukkipati /* Retransmit last segment. */ 22706ba8a3b1SNandita Dukkipati if (WARN_ON(!skb)) 22716ba8a3b1SNandita Dukkipati goto rearm_timer; 22726ba8a3b1SNandita Dukkipati 22731f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 22741f3279aeSEric Dumazet goto rearm_timer; 22751f3279aeSEric Dumazet 22766ba8a3b1SNandita Dukkipati pcount = tcp_skb_pcount(skb); 22776ba8a3b1SNandita Dukkipati if (WARN_ON(!pcount)) 22786ba8a3b1SNandita Dukkipati goto rearm_timer; 22796ba8a3b1SNandita Dukkipati 22806ba8a3b1SNandita Dukkipati if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) { 22816cc55e09SOctavian Purdila if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss, 22826cc55e09SOctavian Purdila GFP_ATOMIC))) 22836ba8a3b1SNandita Dukkipati goto rearm_timer; 2284b340b264SYuchung Cheng skb = tcp_write_queue_next(sk, skb); 22856ba8a3b1SNandita Dukkipati } 22866ba8a3b1SNandita Dukkipati 22876ba8a3b1SNandita Dukkipati if (WARN_ON(!skb || !tcp_skb_pcount(skb))) 22886ba8a3b1SNandita Dukkipati goto rearm_timer; 22896ba8a3b1SNandita Dukkipati 229010d3be56SEric Dumazet if (__tcp_retransmit_skb(sk, skb, 1)) 2291b340b264SYuchung Cheng goto rearm_timer; 22926ba8a3b1SNandita Dukkipati 22939b717a8dSNandita Dukkipati /* Record snd_nxt for loss detection. */ 22949b717a8dSNandita Dukkipati tp->tlp_high_seq = tp->snd_nxt; 22959b717a8dSNandita Dukkipati 2296b340b264SYuchung Cheng probe_sent: 2297c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES); 2298fcd16c0aSYuchung Cheng /* Reset s.t. tcp_rearm_rto will restart timer from now */ 2299fcd16c0aSYuchung Cheng inet_csk(sk)->icsk_pending = 0; 2300b340b264SYuchung Cheng rearm_timer: 2301fcd16c0aSYuchung Cheng tcp_rearm_rto(sk); 23021da177e4SLinus Torvalds } 23031da177e4SLinus Torvalds 2304a762a980SDavid S. Miller /* Push out any pending frames which were held back due to 2305a762a980SDavid S. Miller * TCP_CORK or attempt at coalescing tiny packets. 2306a762a980SDavid S. Miller * The socket must be locked by the caller. 2307a762a980SDavid S. Miller */ 23089e412ba7SIlpo Järvinen void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 23099e412ba7SIlpo Järvinen int nonagle) 2310a762a980SDavid S. Miller { 2311726e07a8SIlpo Järvinen /* If we are closed, the bytes will have to remain here. 2312726e07a8SIlpo Järvinen * In time closedown will finish, we empty the write queue and 2313726e07a8SIlpo Järvinen * all will be happy. 2314726e07a8SIlpo Järvinen */ 2315726e07a8SIlpo Järvinen if (unlikely(sk->sk_state == TCP_CLOSE)) 2316726e07a8SIlpo Järvinen return; 2317726e07a8SIlpo Järvinen 231899a1dec7SMel Gorman if (tcp_write_xmit(sk, cur_mss, nonagle, 0, 23197450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC))) 23209e412ba7SIlpo Järvinen tcp_check_probe_timer(sk); 2321a762a980SDavid S. Miller } 2322a762a980SDavid S. Miller 2323c1b4a7e6SDavid S. Miller /* Send _single_ skb sitting at the send head. This function requires 2324c1b4a7e6SDavid S. Miller * true push pending frames to setup probe timer etc. 2325c1b4a7e6SDavid S. Miller */ 2326c1b4a7e6SDavid S. Miller void tcp_push_one(struct sock *sk, unsigned int mss_now) 2327c1b4a7e6SDavid S. Miller { 2328fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 2329c1b4a7e6SDavid S. Miller 2330c1b4a7e6SDavid S. Miller BUG_ON(!skb || skb->len < mss_now); 2331c1b4a7e6SDavid S. Miller 2332d5dd9175SIlpo Järvinen tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); 2333c1b4a7e6SDavid S. Miller } 2334c1b4a7e6SDavid S. Miller 23351da177e4SLinus Torvalds /* This function returns the amount that we can raise the 23361da177e4SLinus Torvalds * usable window based on the following constraints 23371da177e4SLinus Torvalds * 23381da177e4SLinus Torvalds * 1. The window can never be shrunk once it is offered (RFC 793) 23391da177e4SLinus Torvalds * 2. We limit memory per socket 23401da177e4SLinus Torvalds * 23411da177e4SLinus Torvalds * RFC 1122: 23421da177e4SLinus Torvalds * "the suggested [SWS] avoidance algorithm for the receiver is to keep 23431da177e4SLinus Torvalds * RECV.NEXT + RCV.WIN fixed until: 23441da177e4SLinus Torvalds * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 23451da177e4SLinus Torvalds * 23461da177e4SLinus Torvalds * i.e. don't raise the right edge of the window until you can raise 23471da177e4SLinus Torvalds * it at least MSS bytes. 23481da177e4SLinus Torvalds * 23491da177e4SLinus Torvalds * Unfortunately, the recommended algorithm breaks header prediction, 23501da177e4SLinus Torvalds * since header prediction assumes th->window stays fixed. 23511da177e4SLinus Torvalds * 23521da177e4SLinus Torvalds * Strictly speaking, keeping th->window fixed violates the receiver 23531da177e4SLinus Torvalds * side SWS prevention criteria. The problem is that under this rule 23541da177e4SLinus Torvalds * a stream of single byte packets will cause the right side of the 23551da177e4SLinus Torvalds * window to always advance by a single byte. 23561da177e4SLinus Torvalds * 23571da177e4SLinus Torvalds * Of course, if the sender implements sender side SWS prevention 23581da177e4SLinus Torvalds * then this will not be a problem. 23591da177e4SLinus Torvalds * 23601da177e4SLinus Torvalds * BSD seems to make the following compromise: 23611da177e4SLinus Torvalds * 23621da177e4SLinus Torvalds * If the free space is less than the 1/4 of the maximum 23631da177e4SLinus Torvalds * space available and the free space is less than 1/2 mss, 23641da177e4SLinus Torvalds * then set the window to 0. 23651da177e4SLinus Torvalds * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 23661da177e4SLinus Torvalds * Otherwise, just prevent the window from shrinking 23671da177e4SLinus Torvalds * and from being larger than the largest representable value. 23681da177e4SLinus Torvalds * 23691da177e4SLinus Torvalds * This prevents incremental opening of the window in the regime 23701da177e4SLinus Torvalds * where TCP is limited by the speed of the reader side taking 23711da177e4SLinus Torvalds * data out of the TCP receive queue. It does nothing about 23721da177e4SLinus Torvalds * those cases where the window is constrained on the sender side 23731da177e4SLinus Torvalds * because the pipeline is full. 23741da177e4SLinus Torvalds * 23751da177e4SLinus Torvalds * BSD also seems to "accidentally" limit itself to windows that are a 23761da177e4SLinus Torvalds * multiple of MSS, at least until the free space gets quite small. 23771da177e4SLinus Torvalds * This would appear to be a side effect of the mbuf implementation. 23781da177e4SLinus Torvalds * Combining these two algorithms results in the observed behavior 23791da177e4SLinus Torvalds * of having a fixed window size at almost all times. 23801da177e4SLinus Torvalds * 23811da177e4SLinus Torvalds * Below we obtain similar behavior by forcing the offered window to 23821da177e4SLinus Torvalds * a multiple of the mss when it is feasible to do so. 23831da177e4SLinus Torvalds * 23841da177e4SLinus Torvalds * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 23851da177e4SLinus Torvalds * Regular options like TIMESTAMP are taken into account. 23861da177e4SLinus Torvalds */ 23871da177e4SLinus Torvalds u32 __tcp_select_window(struct sock *sk) 23881da177e4SLinus Torvalds { 2389463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 23901da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2391caa20d9aSStephen Hemminger /* MSS for the peer's data. Previous versions used mss_clamp 23921da177e4SLinus Torvalds * here. I don't know if the value based on our guesses 23931da177e4SLinus Torvalds * of peer's MSS is better for the performance. It's more correct 23941da177e4SLinus Torvalds * but may be worse for the performance because of rcv_mss 23951da177e4SLinus Torvalds * fluctuations. --SAW 1998/11/1 23961da177e4SLinus Torvalds */ 2397463c84b9SArnaldo Carvalho de Melo int mss = icsk->icsk_ack.rcv_mss; 23981da177e4SLinus Torvalds int free_space = tcp_space(sk); 239986c1a045SFlorian Westphal int allowed_space = tcp_full_space(sk); 240086c1a045SFlorian Westphal int full_space = min_t(int, tp->window_clamp, allowed_space); 24011da177e4SLinus Torvalds int window; 24021da177e4SLinus Torvalds 24031da177e4SLinus Torvalds if (mss > full_space) 24041da177e4SLinus Torvalds mss = full_space; 24051da177e4SLinus Torvalds 2406b92edbe0SEric Dumazet if (free_space < (full_space >> 1)) { 2407463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.quick = 0; 24081da177e4SLinus Torvalds 2409b8da51ebSEric Dumazet if (tcp_under_memory_pressure(sk)) 2410056834d9SIlpo Järvinen tp->rcv_ssthresh = min(tp->rcv_ssthresh, 2411056834d9SIlpo Järvinen 4U * tp->advmss); 24121da177e4SLinus Torvalds 241386c1a045SFlorian Westphal /* free_space might become our new window, make sure we don't 241486c1a045SFlorian Westphal * increase it due to wscale. 241586c1a045SFlorian Westphal */ 241686c1a045SFlorian Westphal free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale); 241786c1a045SFlorian Westphal 241886c1a045SFlorian Westphal /* if free space is less than mss estimate, or is below 1/16th 241986c1a045SFlorian Westphal * of the maximum allowed, try to move to zero-window, else 242086c1a045SFlorian Westphal * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and 242186c1a045SFlorian Westphal * new incoming data is dropped due to memory limits. 242286c1a045SFlorian Westphal * With large window, mss test triggers way too late in order 242386c1a045SFlorian Westphal * to announce zero window in time before rmem limit kicks in. 242486c1a045SFlorian Westphal */ 242586c1a045SFlorian Westphal if (free_space < (allowed_space >> 4) || free_space < mss) 24261da177e4SLinus Torvalds return 0; 24271da177e4SLinus Torvalds } 24281da177e4SLinus Torvalds 24291da177e4SLinus Torvalds if (free_space > tp->rcv_ssthresh) 24301da177e4SLinus Torvalds free_space = tp->rcv_ssthresh; 24311da177e4SLinus Torvalds 24321da177e4SLinus Torvalds /* Don't do rounding if we are using window scaling, since the 24331da177e4SLinus Torvalds * scaled window will not line up with the MSS boundary anyway. 24341da177e4SLinus Torvalds */ 24351da177e4SLinus Torvalds window = tp->rcv_wnd; 24361da177e4SLinus Torvalds if (tp->rx_opt.rcv_wscale) { 24371da177e4SLinus Torvalds window = free_space; 24381da177e4SLinus Torvalds 24391da177e4SLinus Torvalds /* Advertise enough space so that it won't get scaled away. 24401da177e4SLinus Torvalds * Import case: prevent zero window announcement if 24411da177e4SLinus Torvalds * 1<<rcv_wscale > mss. 24421da177e4SLinus Torvalds */ 24431da177e4SLinus Torvalds if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window) 24441da177e4SLinus Torvalds window = (((window >> tp->rx_opt.rcv_wscale) + 1) 24451da177e4SLinus Torvalds << tp->rx_opt.rcv_wscale); 24461da177e4SLinus Torvalds } else { 24471da177e4SLinus Torvalds /* Get the largest window that is a nice multiple of mss. 24481da177e4SLinus Torvalds * Window clamp already applied above. 24491da177e4SLinus Torvalds * If our current window offering is within 1 mss of the 24501da177e4SLinus Torvalds * free space we just keep it. This prevents the divide 24511da177e4SLinus Torvalds * and multiply from happening most of the time. 24521da177e4SLinus Torvalds * We also don't do any window rounding when the free space 24531da177e4SLinus Torvalds * is too small. 24541da177e4SLinus Torvalds */ 24551da177e4SLinus Torvalds if (window <= free_space - mss || window > free_space) 24561da177e4SLinus Torvalds window = (free_space / mss) * mss; 245784565070SJohn Heffner else if (mss == full_space && 2458b92edbe0SEric Dumazet free_space > window + (full_space >> 1)) 245984565070SJohn Heffner window = free_space; 24601da177e4SLinus Torvalds } 24611da177e4SLinus Torvalds 24621da177e4SLinus Torvalds return window; 24631da177e4SLinus Torvalds } 24641da177e4SLinus Torvalds 2465cfea5a68SMartin KaFai Lau void tcp_skb_collapse_tstamp(struct sk_buff *skb, 2466082ac2d5SMartin KaFai Lau const struct sk_buff *next_skb) 2467082ac2d5SMartin KaFai Lau { 24680a2cf20cSSoheil Hassas Yeganeh if (unlikely(tcp_has_tx_tstamp(next_skb))) { 24690a2cf20cSSoheil Hassas Yeganeh const struct skb_shared_info *next_shinfo = 24700a2cf20cSSoheil Hassas Yeganeh skb_shinfo(next_skb); 2471082ac2d5SMartin KaFai Lau struct skb_shared_info *shinfo = skb_shinfo(skb); 2472082ac2d5SMartin KaFai Lau 24730a2cf20cSSoheil Hassas Yeganeh shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP; 2474082ac2d5SMartin KaFai Lau shinfo->tskey = next_shinfo->tskey; 24752de8023eSMartin KaFai Lau TCP_SKB_CB(skb)->txstamp_ack |= 24762de8023eSMartin KaFai Lau TCP_SKB_CB(next_skb)->txstamp_ack; 2477082ac2d5SMartin KaFai Lau } 2478082ac2d5SMartin KaFai Lau } 2479082ac2d5SMartin KaFai Lau 24804a17fc3aSIlpo Järvinen /* Collapses two adjacent SKB's during retransmission. */ 24814a17fc3aSIlpo Järvinen static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) 24821da177e4SLinus Torvalds { 24831da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2484fe067e8aSDavid S. Miller struct sk_buff *next_skb = tcp_write_queue_next(sk, skb); 2485058dc334SIlpo Järvinen int skb_size, next_skb_size; 24861da177e4SLinus Torvalds 2487058dc334SIlpo Järvinen skb_size = skb->len; 2488058dc334SIlpo Järvinen next_skb_size = next_skb->len; 24891da177e4SLinus Torvalds 2490058dc334SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); 24911da177e4SLinus Torvalds 24926859d494SIlpo Järvinen tcp_highest_sack_combine(sk, next_skb, skb); 2493a6963a6bSIlpo Järvinen 2494fe067e8aSDavid S. Miller tcp_unlink_write_queue(next_skb, sk); 24951da177e4SLinus Torvalds 2496058dc334SIlpo Järvinen skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size), 24971a4e2d09SArnaldo Carvalho de Melo next_skb_size); 24981da177e4SLinus Torvalds 249952d570aaSJarek Poplawski if (next_skb->ip_summed == CHECKSUM_PARTIAL) 250052d570aaSJarek Poplawski skb->ip_summed = CHECKSUM_PARTIAL; 25011da177e4SLinus Torvalds 250284fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 25031da177e4SLinus Torvalds skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); 25041da177e4SLinus Torvalds 25051da177e4SLinus Torvalds /* Update sequence range on original skb. */ 25061da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 25071da177e4SLinus Torvalds 2508e6c7d085SIlpo Järvinen /* Merge over control information. This moves PSH/FIN etc. over */ 25094de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; 25101da177e4SLinus Torvalds 25111da177e4SLinus Torvalds /* All done, get rid of second SKB and account for it so 25121da177e4SLinus Torvalds * packet counting does not break. 25131da177e4SLinus Torvalds */ 25144828e7f4SIlpo Järvinen TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; 2515a643b5d4SMartin KaFai Lau TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor; 2516b7689205SIlpo Järvinen 2517b7689205SIlpo Järvinen /* changed transmit queue under us so clear hints */ 2518ef9da47cSIlpo Järvinen tcp_clear_retrans_hints_partial(tp); 2519ef9da47cSIlpo Järvinen if (next_skb == tp->retransmit_skb_hint) 2520ef9da47cSIlpo Järvinen tp->retransmit_skb_hint = skb; 2521b7689205SIlpo Järvinen 2522797108d1SIlpo Järvinen tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); 2523797108d1SIlpo Järvinen 2524082ac2d5SMartin KaFai Lau tcp_skb_collapse_tstamp(skb, next_skb); 2525082ac2d5SMartin KaFai Lau 25263ab224beSHideo Aoki sk_wmem_free_skb(sk, next_skb); 25271da177e4SLinus Torvalds } 25281da177e4SLinus Torvalds 252967edfef7SAndi Kleen /* Check if coalescing SKBs is legal. */ 2530a2a385d6SEric Dumazet static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) 25314a17fc3aSIlpo Järvinen { 25324a17fc3aSIlpo Järvinen if (tcp_skb_pcount(skb) > 1) 2533a2a385d6SEric Dumazet return false; 25344a17fc3aSIlpo Järvinen /* TODO: SACK collapsing could be used to remove this condition */ 25354a17fc3aSIlpo Järvinen if (skb_shinfo(skb)->nr_frags != 0) 2536a2a385d6SEric Dumazet return false; 25374a17fc3aSIlpo Järvinen if (skb_cloned(skb)) 2538a2a385d6SEric Dumazet return false; 25394a17fc3aSIlpo Järvinen if (skb == tcp_send_head(sk)) 2540a2a385d6SEric Dumazet return false; 25414a17fc3aSIlpo Järvinen /* Some heurestics for collapsing over SACK'd could be invented */ 25424a17fc3aSIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 2543a2a385d6SEric Dumazet return false; 25444a17fc3aSIlpo Järvinen 2545a2a385d6SEric Dumazet return true; 25464a17fc3aSIlpo Järvinen } 25474a17fc3aSIlpo Järvinen 254867edfef7SAndi Kleen /* Collapse packets in the retransmit queue to make to create 254967edfef7SAndi Kleen * less packets on the wire. This is only done on retransmission. 255067edfef7SAndi Kleen */ 25514a17fc3aSIlpo Järvinen static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, 25524a17fc3aSIlpo Järvinen int space) 25534a17fc3aSIlpo Järvinen { 25544a17fc3aSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 25554a17fc3aSIlpo Järvinen struct sk_buff *skb = to, *tmp; 2556a2a385d6SEric Dumazet bool first = true; 25574a17fc3aSIlpo Järvinen 25584a17fc3aSIlpo Järvinen if (!sysctl_tcp_retrans_collapse) 25594a17fc3aSIlpo Järvinen return; 25604de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 25614a17fc3aSIlpo Järvinen return; 25624a17fc3aSIlpo Järvinen 25634a17fc3aSIlpo Järvinen tcp_for_write_queue_from_safe(skb, tmp, sk) { 25644a17fc3aSIlpo Järvinen if (!tcp_can_collapse(sk, skb)) 25654a17fc3aSIlpo Järvinen break; 25664a17fc3aSIlpo Järvinen 2567a643b5d4SMartin KaFai Lau if (!tcp_skb_can_collapse_to(to)) 2568a643b5d4SMartin KaFai Lau break; 2569a643b5d4SMartin KaFai Lau 25704a17fc3aSIlpo Järvinen space -= skb->len; 25714a17fc3aSIlpo Järvinen 25724a17fc3aSIlpo Järvinen if (first) { 2573a2a385d6SEric Dumazet first = false; 25744a17fc3aSIlpo Järvinen continue; 25754a17fc3aSIlpo Järvinen } 25764a17fc3aSIlpo Järvinen 25774a17fc3aSIlpo Järvinen if (space < 0) 25784a17fc3aSIlpo Järvinen break; 25794a17fc3aSIlpo Järvinen /* Punt if not enough space exists in the first SKB for 25804a17fc3aSIlpo Järvinen * the data in the second 25814a17fc3aSIlpo Järvinen */ 2582a21d4572SEric Dumazet if (skb->len > skb_availroom(to)) 25834a17fc3aSIlpo Järvinen break; 25844a17fc3aSIlpo Järvinen 25854a17fc3aSIlpo Järvinen if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) 25864a17fc3aSIlpo Järvinen break; 25874a17fc3aSIlpo Järvinen 25884a17fc3aSIlpo Järvinen tcp_collapse_retrans(sk, to); 25894a17fc3aSIlpo Järvinen } 25904a17fc3aSIlpo Järvinen } 25914a17fc3aSIlpo Järvinen 25921da177e4SLinus Torvalds /* This retransmits one SKB. Policy decisions and retransmit queue 25931da177e4SLinus Torvalds * state updates are done by the caller. Returns non-zero if an 25941da177e4SLinus Torvalds * error occurred which prevented the send. 25951da177e4SLinus Torvalds */ 259610d3be56SEric Dumazet int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) 25971da177e4SLinus Torvalds { 25985d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 259910d3be56SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 26007d227cd2SSridhar Samudrala unsigned int cur_mss; 260110d3be56SEric Dumazet int diff, len, err; 26021da177e4SLinus Torvalds 260310d3be56SEric Dumazet 260410d3be56SEric Dumazet /* Inconclusive MTU probe */ 260510d3be56SEric Dumazet if (icsk->icsk_mtup.probe_size) 26065d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 26075d424d5aSJohn Heffner 26081da177e4SLinus Torvalds /* Do not sent more than we queued. 1/4 is reserved for possible 2609caa20d9aSStephen Hemminger * copying overhead: fragmentation, tunneling, mangling etc. 26101da177e4SLinus Torvalds */ 26111da177e4SLinus Torvalds if (atomic_read(&sk->sk_wmem_alloc) > 26121da177e4SLinus Torvalds min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf)) 26131da177e4SLinus Torvalds return -EAGAIN; 26141da177e4SLinus Torvalds 26151f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 26161f3279aeSEric Dumazet return -EBUSY; 26171f3279aeSEric Dumazet 26181da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 26191da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 26201da177e4SLinus Torvalds BUG(); 26211da177e4SLinus Torvalds if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 26221da177e4SLinus Torvalds return -ENOMEM; 26231da177e4SLinus Torvalds } 26241da177e4SLinus Torvalds 26257d227cd2SSridhar Samudrala if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 26267d227cd2SSridhar Samudrala return -EHOSTUNREACH; /* Routing failure or similar. */ 26277d227cd2SSridhar Samudrala 26280c54b85fSIlpo Järvinen cur_mss = tcp_current_mss(sk); 26297d227cd2SSridhar Samudrala 26301da177e4SLinus Torvalds /* If receiver has shrunk his window, and skb is out of 26311da177e4SLinus Torvalds * new window, do not retransmit it. The exception is the 26321da177e4SLinus Torvalds * case, when window is shrunk to zero. In this case 26331da177e4SLinus Torvalds * our retransmit serves as a zero window probe. 26341da177e4SLinus Torvalds */ 26359d4fb27dSJoe Perches if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && 26369d4fb27dSJoe Perches TCP_SKB_CB(skb)->seq != tp->snd_una) 26371da177e4SLinus Torvalds return -EAGAIN; 26381da177e4SLinus Torvalds 263910d3be56SEric Dumazet len = cur_mss * segs; 264010d3be56SEric Dumazet if (skb->len > len) { 264110d3be56SEric Dumazet if (tcp_fragment(sk, skb, len, cur_mss, GFP_ATOMIC)) 26421da177e4SLinus Torvalds return -ENOMEM; /* We'll try again later. */ 264302276f3cSIlpo Järvinen } else { 2644c52e2421SEric Dumazet if (skb_unclone(skb, GFP_ATOMIC)) 2645c52e2421SEric Dumazet return -ENOMEM; 264610d3be56SEric Dumazet 264710d3be56SEric Dumazet diff = tcp_skb_pcount(skb); 264810d3be56SEric Dumazet tcp_set_skb_tso_segs(skb, cur_mss); 264910d3be56SEric Dumazet diff -= tcp_skb_pcount(skb); 265010d3be56SEric Dumazet if (diff) 265110d3be56SEric Dumazet tcp_adjust_pcount(sk, skb, diff); 265210d3be56SEric Dumazet if (skb->len < cur_mss) 265310d3be56SEric Dumazet tcp_retrans_try_collapse(sk, skb, cur_mss); 26541da177e4SLinus Torvalds } 26551da177e4SLinus Torvalds 265649213555SDaniel Borkmann /* RFC3168, section 6.1.1.1. ECN fallback */ 265749213555SDaniel Borkmann if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN) 265849213555SDaniel Borkmann tcp_ecn_clear_syn(sk, skb); 265949213555SDaniel Borkmann 266050bceae9SThomas Graf /* make sure skb->data is aligned on arches that require it 266150bceae9SThomas Graf * and check if ack-trimming & collapsing extended the headroom 266250bceae9SThomas Graf * beyond what csum_start can cover. 266350bceae9SThomas Graf */ 266450bceae9SThomas Graf if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) || 266550bceae9SThomas Graf skb_headroom(skb) >= 0xFFFF)) { 266610a81980SEric Dumazet struct sk_buff *nskb; 266710a81980SEric Dumazet 266810a81980SEric Dumazet skb_mstamp_get(&skb->skb_mstamp); 266910a81980SEric Dumazet nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC); 2670c84a5711SYuchung Cheng err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) : 2671117632e6SEric Dumazet -ENOBUFS; 2672117632e6SEric Dumazet } else { 2673c84a5711SYuchung Cheng err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 2674117632e6SEric Dumazet } 2675c84a5711SYuchung Cheng 2676fc9f3501SEric Dumazet if (likely(!err)) { 267710d3be56SEric Dumazet segs = tcp_skb_pcount(skb); 267810d3be56SEric Dumazet 2679c84a5711SYuchung Cheng TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS; 2680fc9f3501SEric Dumazet /* Update global TCP statistics. */ 268110d3be56SEric Dumazet TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs); 2682fc9f3501SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 268302a1d6e7SEric Dumazet __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 268410d3be56SEric Dumazet tp->total_retrans += segs; 2685fc9f3501SEric Dumazet } 2686c84a5711SYuchung Cheng return err; 268793b174adSYuchung Cheng } 268893b174adSYuchung Cheng 268910d3be56SEric Dumazet int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) 269093b174adSYuchung Cheng { 269193b174adSYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 269210d3be56SEric Dumazet int err = __tcp_retransmit_skb(sk, skb, segs); 26931da177e4SLinus Torvalds 26941da177e4SLinus Torvalds if (err == 0) { 26951da177e4SLinus Torvalds #if FASTRETRANS_DEBUG > 0 26961da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 2697e87cc472SJoe Perches net_dbg_ratelimited("retrans_out leaked\n"); 26981da177e4SLinus Torvalds } 26991da177e4SLinus Torvalds #endif 27001da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 27011da177e4SLinus Torvalds tp->retrans_out += tcp_skb_pcount(skb); 27021da177e4SLinus Torvalds 27031da177e4SLinus Torvalds /* Save stamp of the first retransmit. */ 27041da177e4SLinus Torvalds if (!tp->retrans_stamp) 27057faee5c0SEric Dumazet tp->retrans_stamp = tcp_skb_timestamp(skb); 27061da177e4SLinus Torvalds 27071f3279aeSEric Dumazet } else if (err != -EBUSY) { 2708c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL); 27091da177e4SLinus Torvalds } 27106e08d5e3SYuchung Cheng 27116e08d5e3SYuchung Cheng if (tp->undo_retrans < 0) 27126e08d5e3SYuchung Cheng tp->undo_retrans = 0; 27136e08d5e3SYuchung Cheng tp->undo_retrans += tcp_skb_pcount(skb); 27141da177e4SLinus Torvalds return err; 27151da177e4SLinus Torvalds } 27161da177e4SLinus Torvalds 271767edfef7SAndi Kleen /* Check if we forward retransmits are possible in the current 271867edfef7SAndi Kleen * window/congestion state. 271967edfef7SAndi Kleen */ 2720a2a385d6SEric Dumazet static bool tcp_can_forward_retransmit(struct sock *sk) 2721b5afe7bcSIlpo Järvinen { 2722b5afe7bcSIlpo Järvinen const struct inet_connection_sock *icsk = inet_csk(sk); 2723cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 2724b5afe7bcSIlpo Järvinen 2725b5afe7bcSIlpo Järvinen /* Forward retransmissions are possible only during Recovery. */ 2726b5afe7bcSIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Recovery) 2727a2a385d6SEric Dumazet return false; 2728b5afe7bcSIlpo Järvinen 2729b5afe7bcSIlpo Järvinen /* No forward retransmissions in Reno are possible. */ 2730b5afe7bcSIlpo Järvinen if (tcp_is_reno(tp)) 2731a2a385d6SEric Dumazet return false; 2732b5afe7bcSIlpo Järvinen 2733b5afe7bcSIlpo Järvinen /* Yeah, we have to make difficult choice between forward transmission 2734b5afe7bcSIlpo Järvinen * and retransmission... Both ways have their merits... 2735b5afe7bcSIlpo Järvinen * 2736b5afe7bcSIlpo Järvinen * For now we do not retransmit anything, while we have some new 2737b5afe7bcSIlpo Järvinen * segments to send. In the other cases, follow rule 3 for 2738b5afe7bcSIlpo Järvinen * NextSeg() specified in RFC3517. 2739b5afe7bcSIlpo Järvinen */ 2740b5afe7bcSIlpo Järvinen 2741b5afe7bcSIlpo Järvinen if (tcp_may_send_now(sk)) 2742a2a385d6SEric Dumazet return false; 2743b5afe7bcSIlpo Järvinen 2744a2a385d6SEric Dumazet return true; 2745b5afe7bcSIlpo Järvinen } 2746b5afe7bcSIlpo Järvinen 27471da177e4SLinus Torvalds /* This gets called after a retransmit timeout, and the initially 27481da177e4SLinus Torvalds * retransmitted data is acknowledged. It tries to continue 27491da177e4SLinus Torvalds * resending the rest of the retransmit queue, until either 27501da177e4SLinus Torvalds * we've sent it all or the congestion window limit is reached. 27511da177e4SLinus Torvalds * If doing SACK, the first ACK which comes back for a timeout 27521da177e4SLinus Torvalds * based retransmit packet might feed us FACK information again. 27531da177e4SLinus Torvalds * If so, we use it to avoid unnecessarily retransmissions. 27541da177e4SLinus Torvalds */ 27551da177e4SLinus Torvalds void tcp_xmit_retransmit_queue(struct sock *sk) 27561da177e4SLinus Torvalds { 27576687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 27581da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 27591da177e4SLinus Torvalds struct sk_buff *skb; 27600e1c54c2SIlpo Järvinen struct sk_buff *hole = NULL; 2761a3d2e9f8SEric Dumazet u32 max_segs, last_lost; 276261eb55f4SIlpo Järvinen int mib_idx; 27630e1c54c2SIlpo Järvinen int fwd_rexmitting = 0; 27646a438bbeSStephen Hemminger 276545e77d31SIlpo Järvinen if (!tp->packets_out) 276645e77d31SIlpo Järvinen return; 276745e77d31SIlpo Järvinen 276808ebd172SIlpo Järvinen if (!tp->lost_out) 276908ebd172SIlpo Järvinen tp->retransmit_high = tp->snd_una; 277008ebd172SIlpo Järvinen 2771618d9f25SIlpo Järvinen if (tp->retransmit_skb_hint) { 27726a438bbeSStephen Hemminger skb = tp->retransmit_skb_hint; 2773618d9f25SIlpo Järvinen last_lost = TCP_SKB_CB(skb)->end_seq; 2774618d9f25SIlpo Järvinen if (after(last_lost, tp->retransmit_high)) 2775618d9f25SIlpo Järvinen last_lost = tp->retransmit_high; 2776618d9f25SIlpo Järvinen } else { 2777fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 2778618d9f25SIlpo Järvinen last_lost = tp->snd_una; 2779618d9f25SIlpo Järvinen } 27801da177e4SLinus Torvalds 2781a3d2e9f8SEric Dumazet max_segs = tcp_tso_autosize(sk, tcp_current_mss(sk)); 2782fe067e8aSDavid S. Miller tcp_for_write_queue_from(skb, sk) { 2783dca0aaf8SEric Dumazet __u8 sacked; 278410d3be56SEric Dumazet int segs; 27851da177e4SLinus Torvalds 2786fe067e8aSDavid S. Miller if (skb == tcp_send_head(sk)) 2787fe067e8aSDavid S. Miller break; 27886a438bbeSStephen Hemminger /* we could do better than to assign each time */ 278951456b29SIan Morris if (!hole) 27906a438bbeSStephen Hemminger tp->retransmit_skb_hint = skb; 27916a438bbeSStephen Hemminger 279210d3be56SEric Dumazet segs = tp->snd_cwnd - tcp_packets_in_flight(tp); 279310d3be56SEric Dumazet if (segs <= 0) 27941da177e4SLinus Torvalds return; 2795dca0aaf8SEric Dumazet sacked = TCP_SKB_CB(skb)->sacked; 2796a3d2e9f8SEric Dumazet /* In case tcp_shift_skb_data() have aggregated large skbs, 2797a3d2e9f8SEric Dumazet * we need to make sure not sending too bigs TSO packets 2798a3d2e9f8SEric Dumazet */ 2799a3d2e9f8SEric Dumazet segs = min_t(int, segs, max_segs); 28000e1c54c2SIlpo Järvinen 28010e1c54c2SIlpo Järvinen if (fwd_rexmitting) { 28020e1c54c2SIlpo Järvinen begin_fwd: 28030e1c54c2SIlpo Järvinen if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp))) 2804006f582cSIlpo Järvinen break; 28050e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFORWARDRETRANS; 28060e1c54c2SIlpo Järvinen 28070e1c54c2SIlpo Järvinen } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) { 2808618d9f25SIlpo Järvinen tp->retransmit_high = last_lost; 28090e1c54c2SIlpo Järvinen if (!tcp_can_forward_retransmit(sk)) 28100e1c54c2SIlpo Järvinen break; 28110e1c54c2SIlpo Järvinen /* Backtrack if necessary to non-L'ed skb */ 281200db4124SIan Morris if (hole) { 28130e1c54c2SIlpo Järvinen skb = hole; 28140e1c54c2SIlpo Järvinen hole = NULL; 28150e1c54c2SIlpo Järvinen } 28160e1c54c2SIlpo Järvinen fwd_rexmitting = 1; 28170e1c54c2SIlpo Järvinen goto begin_fwd; 28180e1c54c2SIlpo Järvinen 28190e1c54c2SIlpo Järvinen } else if (!(sacked & TCPCB_LOST)) { 282051456b29SIan Morris if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) 28210e1c54c2SIlpo Järvinen hole = skb; 282261eb55f4SIlpo Järvinen continue; 28231da177e4SLinus Torvalds 28240e1c54c2SIlpo Järvinen } else { 2825618d9f25SIlpo Järvinen last_lost = TCP_SKB_CB(skb)->end_seq; 28260e1c54c2SIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Loss) 28270e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFASTRETRANS; 28280e1c54c2SIlpo Järvinen else 28290e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; 28300e1c54c2SIlpo Järvinen } 28310e1c54c2SIlpo Järvinen 28320e1c54c2SIlpo Järvinen if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) 283361eb55f4SIlpo Järvinen continue; 283440b215e5SPavel Emelyanov 283510d3be56SEric Dumazet if (tcp_retransmit_skb(sk, skb, segs)) 28361da177e4SLinus Torvalds return; 283724ab6becSYuchung Cheng 2838c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), mib_idx); 28391da177e4SLinus Torvalds 2840684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 2841a262f0cdSNandita Dukkipati tp->prr_out += tcp_skb_pcount(skb); 2842a262f0cdSNandita Dukkipati 2843fe067e8aSDavid S. Miller if (skb == tcp_write_queue_head(sk)) 2844463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 28453f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, 28463f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 28471da177e4SLinus Torvalds } 28481da177e4SLinus Torvalds } 28491da177e4SLinus Torvalds 2850d83769a5SEric Dumazet /* We allow to exceed memory limits for FIN packets to expedite 2851d83769a5SEric Dumazet * connection tear down and (memory) recovery. 2852845704a5SEric Dumazet * Otherwise tcp_send_fin() could be tempted to either delay FIN 2853845704a5SEric Dumazet * or even be forced to close flow without any FIN. 2854a6c5ea4cSEric Dumazet * In general, we want to allow one skb per socket to avoid hangs 2855a6c5ea4cSEric Dumazet * with edge trigger epoll() 2856d83769a5SEric Dumazet */ 2857a6c5ea4cSEric Dumazet void sk_forced_mem_schedule(struct sock *sk, int size) 2858d83769a5SEric Dumazet { 2859e805605cSJohannes Weiner int amt; 2860d83769a5SEric Dumazet 2861d83769a5SEric Dumazet if (size <= sk->sk_forward_alloc) 2862d83769a5SEric Dumazet return; 2863d83769a5SEric Dumazet amt = sk_mem_pages(size); 2864d83769a5SEric Dumazet sk->sk_forward_alloc += amt * SK_MEM_QUANTUM; 2865e805605cSJohannes Weiner sk_memory_allocated_add(sk, amt); 2866e805605cSJohannes Weiner 2867baac50bbSJohannes Weiner if (mem_cgroup_sockets_enabled && sk->sk_memcg) 2868baac50bbSJohannes Weiner mem_cgroup_charge_skmem(sk->sk_memcg, amt); 2869d83769a5SEric Dumazet } 2870d83769a5SEric Dumazet 2871845704a5SEric Dumazet /* Send a FIN. The caller locks the socket for us. 2872845704a5SEric Dumazet * We should try to send a FIN packet really hard, but eventually give up. 28731da177e4SLinus Torvalds */ 28741da177e4SLinus Torvalds void tcp_send_fin(struct sock *sk) 28751da177e4SLinus Torvalds { 2876845704a5SEric Dumazet struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk); 28771da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 28781da177e4SLinus Torvalds 2879845704a5SEric Dumazet /* Optimization, tack on the FIN if we have one skb in write queue and 2880845704a5SEric Dumazet * this skb was not yet sent, or we are under memory pressure. 2881845704a5SEric Dumazet * Note: in the latter case, FIN packet will be sent after a timeout, 2882845704a5SEric Dumazet * as TCP stack thinks it has already been transmitted. 28831da177e4SLinus Torvalds */ 2884b8da51ebSEric Dumazet if (tskb && (tcp_send_head(sk) || tcp_under_memory_pressure(sk))) { 2885845704a5SEric Dumazet coalesce: 2886845704a5SEric Dumazet TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN; 2887845704a5SEric Dumazet TCP_SKB_CB(tskb)->end_seq++; 28881da177e4SLinus Torvalds tp->write_seq++; 2889845704a5SEric Dumazet if (!tcp_send_head(sk)) { 2890845704a5SEric Dumazet /* This means tskb was already sent. 2891845704a5SEric Dumazet * Pretend we included the FIN on previous transmit. 2892845704a5SEric Dumazet * We need to set tp->snd_nxt to the value it would have 2893845704a5SEric Dumazet * if FIN had been sent. This is because retransmit path 2894845704a5SEric Dumazet * does not change tp->snd_nxt. 2895845704a5SEric Dumazet */ 2896845704a5SEric Dumazet tp->snd_nxt++; 2897845704a5SEric Dumazet return; 2898845704a5SEric Dumazet } 28991da177e4SLinus Torvalds } else { 2900845704a5SEric Dumazet skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation); 2901845704a5SEric Dumazet if (unlikely(!skb)) { 2902845704a5SEric Dumazet if (tskb) 2903845704a5SEric Dumazet goto coalesce; 2904845704a5SEric Dumazet return; 29051da177e4SLinus Torvalds } 2906d83769a5SEric Dumazet skb_reserve(skb, MAX_TCP_HEADER); 2907a6c5ea4cSEric Dumazet sk_forced_mem_schedule(sk, skb->truesize); 29081da177e4SLinus Torvalds /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 2909e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tp->write_seq, 2910a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_FIN); 29111da177e4SLinus Torvalds tcp_queue_skb(sk, skb); 29121da177e4SLinus Torvalds } 2913845704a5SEric Dumazet __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF); 29141da177e4SLinus Torvalds } 29151da177e4SLinus Torvalds 29161da177e4SLinus Torvalds /* We get here when a process closes a file descriptor (either due to 29171da177e4SLinus Torvalds * an explicit close() or as a byproduct of exit()'ing) and there 29181da177e4SLinus Torvalds * was unread data in the receive queue. This behavior is recommended 291965bb723cSGerrit Renker * by RFC 2525, section 2.17. -DaveM 29201da177e4SLinus Torvalds */ 2921dd0fc66fSAl Viro void tcp_send_active_reset(struct sock *sk, gfp_t priority) 29221da177e4SLinus Torvalds { 29231da177e4SLinus Torvalds struct sk_buff *skb; 29241da177e4SLinus Torvalds 29251da177e4SLinus Torvalds /* NOTE: No TCP options attached and we never retransmit this. */ 29261da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, priority); 29271da177e4SLinus Torvalds if (!skb) { 29284e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 29291da177e4SLinus Torvalds return; 29301da177e4SLinus Torvalds } 29311da177e4SLinus Torvalds 29321da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 29331da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 2934e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), 2935a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_RST); 2936675ee231SEric Dumazet skb_mstamp_get(&skb->skb_mstamp); 29371da177e4SLinus Torvalds /* Send it off. */ 2938dfb4b9dcSDavid S. Miller if (tcp_transmit_skb(sk, skb, 0, priority)) 29394e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 294026af65cbSSridhar Samudrala 294181cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); 29421da177e4SLinus Torvalds } 29431da177e4SLinus Torvalds 294467edfef7SAndi Kleen /* Send a crossed SYN-ACK during socket establishment. 294567edfef7SAndi Kleen * WARNING: This routine must only be called when we have already sent 29461da177e4SLinus Torvalds * a SYN packet that crossed the incoming SYN that caused this routine 29471da177e4SLinus Torvalds * to get called. If this assumption fails then the initial rcv_wnd 29481da177e4SLinus Torvalds * and rcv_wscale values will not be correct. 29491da177e4SLinus Torvalds */ 29501da177e4SLinus Torvalds int tcp_send_synack(struct sock *sk) 29511da177e4SLinus Torvalds { 29521da177e4SLinus Torvalds struct sk_buff *skb; 29531da177e4SLinus Torvalds 2954fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 295551456b29SIan Morris if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 295691df42beSJoe Perches pr_debug("%s: wrong queue state\n", __func__); 29571da177e4SLinus Torvalds return -EFAULT; 29581da177e4SLinus Torvalds } 29594de075e0SEric Dumazet if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { 29601da177e4SLinus Torvalds if (skb_cloned(skb)) { 29611da177e4SLinus Torvalds struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC); 296251456b29SIan Morris if (!nskb) 29631da177e4SLinus Torvalds return -ENOMEM; 2964fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 2965f4a775d1SEric Dumazet __skb_header_release(nskb); 2966fe067e8aSDavid S. Miller __tcp_add_write_queue_head(sk, nskb); 29673ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 29683ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 29693ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 29701da177e4SLinus Torvalds skb = nskb; 29711da177e4SLinus Torvalds } 29721da177e4SLinus Torvalds 29734de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; 2974735d3831SFlorian Westphal tcp_ecn_send_synack(sk, skb); 29751da177e4SLinus Torvalds } 2976dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 29771da177e4SLinus Torvalds } 29781da177e4SLinus Torvalds 29794aea39c1SEric Dumazet /** 29804aea39c1SEric Dumazet * tcp_make_synack - Prepare a SYN-ACK. 29814aea39c1SEric Dumazet * sk: listener socket 29824aea39c1SEric Dumazet * dst: dst entry attached to the SYNACK 29834aea39c1SEric Dumazet * req: request_sock pointer 29844aea39c1SEric Dumazet * 29854aea39c1SEric Dumazet * Allocate one skb and build a SYNACK packet. 29864aea39c1SEric Dumazet * @dst is consumed : Caller should not use it again. 29874aea39c1SEric Dumazet */ 29885d062de7SEric Dumazet struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst, 2989e6b4d113SWilliam Allen Simpson struct request_sock *req, 2990ca6fb065SEric Dumazet struct tcp_fastopen_cookie *foc, 2991b3d05147SEric Dumazet enum tcp_synack_type synack_type) 29921da177e4SLinus Torvalds { 29932e6599cbSArnaldo Carvalho de Melo struct inet_request_sock *ireq = inet_rsk(req); 29945d062de7SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 299580f03e27SEric Dumazet struct tcp_md5sig_key *md5 = NULL; 29965d062de7SEric Dumazet struct tcp_out_options opts; 29975d062de7SEric Dumazet struct sk_buff *skb; 2998bd0388aeSWilliam Allen Simpson int tcp_header_size; 29995d062de7SEric Dumazet struct tcphdr *th; 30005d062de7SEric Dumazet u16 user_mss; 3001f5fff5dcSTom Quetchenbach int mss; 30021da177e4SLinus Torvalds 3003ca6fb065SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 30044aea39c1SEric Dumazet if (unlikely(!skb)) { 30054aea39c1SEric Dumazet dst_release(dst); 30061da177e4SLinus Torvalds return NULL; 30074aea39c1SEric Dumazet } 30081da177e4SLinus Torvalds /* Reserve space for headers. */ 30091da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 30101da177e4SLinus Torvalds 3011b3d05147SEric Dumazet switch (synack_type) { 3012b3d05147SEric Dumazet case TCP_SYNACK_NORMAL: 30139e17f8a4SEric Dumazet skb_set_owner_w(skb, req_to_sk(req)); 3014b3d05147SEric Dumazet break; 3015b3d05147SEric Dumazet case TCP_SYNACK_COOKIE: 3016b3d05147SEric Dumazet /* Under synflood, we do not attach skb to a socket, 3017b3d05147SEric Dumazet * to avoid false sharing. 3018b3d05147SEric Dumazet */ 3019b3d05147SEric Dumazet break; 3020b3d05147SEric Dumazet case TCP_SYNACK_FASTOPEN: 3021ca6fb065SEric Dumazet /* sk is a const pointer, because we want to express multiple 3022ca6fb065SEric Dumazet * cpu might call us concurrently. 3023ca6fb065SEric Dumazet * sk->sk_wmem_alloc in an atomic, we can promote to rw. 3024ca6fb065SEric Dumazet */ 3025ca6fb065SEric Dumazet skb_set_owner_w(skb, (struct sock *)sk); 3026b3d05147SEric Dumazet break; 3027ca6fb065SEric Dumazet } 30284aea39c1SEric Dumazet skb_dst_set(skb, dst); 30291da177e4SLinus Torvalds 30300dbaee3bSDavid S. Miller mss = dst_metric_advmss(dst); 30315d062de7SEric Dumazet user_mss = READ_ONCE(tp->rx_opt.user_mss); 30325d062de7SEric Dumazet if (user_mss && user_mss < mss) 30335d062de7SEric Dumazet mss = user_mss; 3034f5fff5dcSTom Quetchenbach 303533ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 30368b5f12d0SFlorian Westphal #ifdef CONFIG_SYN_COOKIES 30378b5f12d0SFlorian Westphal if (unlikely(req->cookie_ts)) 30387faee5c0SEric Dumazet skb->skb_mstamp.stamp_jiffies = cookie_init_timestamp(req); 30398b5f12d0SFlorian Westphal else 30408b5f12d0SFlorian Westphal #endif 30417faee5c0SEric Dumazet skb_mstamp_get(&skb->skb_mstamp); 304280f03e27SEric Dumazet 304380f03e27SEric Dumazet #ifdef CONFIG_TCP_MD5SIG 304480f03e27SEric Dumazet rcu_read_lock(); 3045fd3a154aSEric Dumazet md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req)); 304680f03e27SEric Dumazet #endif 304758d607d3SEric Dumazet skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4); 304837bfbddaSEric Dumazet tcp_header_size = tcp_synack_options(req, mss, skb, &opts, md5, foc) + 304937bfbddaSEric Dumazet sizeof(*th); 305033ad798cSAdam Langley 3051aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 3052aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 30531da177e4SLinus Torvalds 3054ea1627c2SEric Dumazet th = (struct tcphdr *)skb->data; 30551da177e4SLinus Torvalds memset(th, 0, sizeof(struct tcphdr)); 30561da177e4SLinus Torvalds th->syn = 1; 30571da177e4SLinus Torvalds th->ack = 1; 30586ac705b1SEric Dumazet tcp_ecn_make_synack(req, th); 3059b44084c2SEric Dumazet th->source = htons(ireq->ir_num); 3060634fb979SEric Dumazet th->dest = ireq->ir_rmt_port; 3061e870a8efSIlpo Järvinen /* Setting of flags are superfluous here for callers (and ECE is 3062e870a8efSIlpo Järvinen * not even correctly set) 3063e870a8efSIlpo Järvinen */ 3064e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn, 3065a3433f35SChangli Gao TCPHDR_SYN | TCPHDR_ACK); 30664957faadSWilliam Allen Simpson 30671da177e4SLinus Torvalds th->seq = htonl(TCP_SKB_CB(skb)->seq); 30688336886fSJerry Chu /* XXX data is queued and acked as is. No buffer/window check */ 30698336886fSJerry Chu th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt); 30701da177e4SLinus Torvalds 30711da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 3072ed53d0abSEric Dumazet th->window = htons(min(req->rsk_rcv_wnd, 65535U)); 30735d062de7SEric Dumazet tcp_options_write((__be32 *)(th + 1), NULL, &opts); 30741da177e4SLinus Torvalds th->doff = (tcp_header_size >> 2); 307590bbcc60SEric Dumazet __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS); 3076cfb6eeb4SYOSHIFUJI Hideaki 3077cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 3078cfb6eeb4SYOSHIFUJI Hideaki /* Okay, we have all we need - do the md5 hash if needed */ 307980f03e27SEric Dumazet if (md5) 3080bd0388aeSWilliam Allen Simpson tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, 308139f8e58eSEric Dumazet md5, req_to_sk(req), skb); 308280f03e27SEric Dumazet rcu_read_unlock(); 3083cfb6eeb4SYOSHIFUJI Hideaki #endif 3084cfb6eeb4SYOSHIFUJI Hideaki 3085b50edd78SEric Dumazet /* Do not fool tcpdump (if any), clean our debris */ 3086b50edd78SEric Dumazet skb->tstamp.tv64 = 0; 30871da177e4SLinus Torvalds return skb; 30881da177e4SLinus Torvalds } 30894bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_make_synack); 30901da177e4SLinus Torvalds 309181164413SDaniel Borkmann static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst) 309281164413SDaniel Borkmann { 309381164413SDaniel Borkmann struct inet_connection_sock *icsk = inet_csk(sk); 309481164413SDaniel Borkmann const struct tcp_congestion_ops *ca; 309581164413SDaniel Borkmann u32 ca_key = dst_metric(dst, RTAX_CC_ALGO); 309681164413SDaniel Borkmann 309781164413SDaniel Borkmann if (ca_key == TCP_CA_UNSPEC) 309881164413SDaniel Borkmann return; 309981164413SDaniel Borkmann 310081164413SDaniel Borkmann rcu_read_lock(); 310181164413SDaniel Borkmann ca = tcp_ca_find_key(ca_key); 310281164413SDaniel Borkmann if (likely(ca && try_module_get(ca->owner))) { 310381164413SDaniel Borkmann module_put(icsk->icsk_ca_ops->owner); 310481164413SDaniel Borkmann icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst); 310581164413SDaniel Borkmann icsk->icsk_ca_ops = ca; 310681164413SDaniel Borkmann } 310781164413SDaniel Borkmann rcu_read_unlock(); 310881164413SDaniel Borkmann } 310981164413SDaniel Borkmann 311067edfef7SAndi Kleen /* Do all connect socket setups that can be done AF independent. */ 3111f7e56a76Sstephen hemminger static void tcp_connect_init(struct sock *sk) 31121da177e4SLinus Torvalds { 3113cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 31141da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 31151da177e4SLinus Torvalds __u8 rcv_wscale; 31161da177e4SLinus Torvalds 31171da177e4SLinus Torvalds /* We'll fix this up when we get a response from the other end. 31181da177e4SLinus Torvalds * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 31191da177e4SLinus Torvalds */ 31201da177e4SLinus Torvalds tp->tcp_header_len = sizeof(struct tcphdr) + 3121bb5b7c11SDavid S. Miller (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0); 31221da177e4SLinus Torvalds 3123cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 312400db4124SIan Morris if (tp->af_specific->md5_lookup(sk, sk)) 3125cfb6eeb4SYOSHIFUJI Hideaki tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 3126cfb6eeb4SYOSHIFUJI Hideaki #endif 3127cfb6eeb4SYOSHIFUJI Hideaki 31281da177e4SLinus Torvalds /* If user gave his TCP_MAXSEG, record it to clamp */ 31291da177e4SLinus Torvalds if (tp->rx_opt.user_mss) 31301da177e4SLinus Torvalds tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 31311da177e4SLinus Torvalds tp->max_window = 0; 31325d424d5aSJohn Heffner tcp_mtup_init(sk); 31331da177e4SLinus Torvalds tcp_sync_mss(sk, dst_mtu(dst)); 31341da177e4SLinus Torvalds 313581164413SDaniel Borkmann tcp_ca_dst_init(sk, dst); 313681164413SDaniel Borkmann 31371da177e4SLinus Torvalds if (!tp->window_clamp) 31381da177e4SLinus Torvalds tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 31390dbaee3bSDavid S. Miller tp->advmss = dst_metric_advmss(dst); 3140f5fff5dcSTom Quetchenbach if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss) 3141f5fff5dcSTom Quetchenbach tp->advmss = tp->rx_opt.user_mss; 3142f5fff5dcSTom Quetchenbach 31431da177e4SLinus Torvalds tcp_initialize_rcv_mss(sk); 31441da177e4SLinus Torvalds 3145e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 3146e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 3147e88c64f0SHagen Paul Pfeifer (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) 3148e88c64f0SHagen Paul Pfeifer tp->window_clamp = tcp_full_space(sk); 3149e88c64f0SHagen Paul Pfeifer 31501da177e4SLinus Torvalds tcp_select_initial_window(tcp_full_space(sk), 31511da177e4SLinus Torvalds tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 31521da177e4SLinus Torvalds &tp->rcv_wnd, 31531da177e4SLinus Torvalds &tp->window_clamp, 3154bb5b7c11SDavid S. Miller sysctl_tcp_window_scaling, 315531d12926Slaurent chavey &rcv_wscale, 315631d12926Slaurent chavey dst_metric(dst, RTAX_INITRWND)); 31571da177e4SLinus Torvalds 31581da177e4SLinus Torvalds tp->rx_opt.rcv_wscale = rcv_wscale; 31591da177e4SLinus Torvalds tp->rcv_ssthresh = tp->rcv_wnd; 31601da177e4SLinus Torvalds 31611da177e4SLinus Torvalds sk->sk_err = 0; 31621da177e4SLinus Torvalds sock_reset_flag(sk, SOCK_DONE); 31631da177e4SLinus Torvalds tp->snd_wnd = 0; 3164ee7537b6SHantzis Fotis tcp_init_wl(tp, 0); 31651da177e4SLinus Torvalds tp->snd_una = tp->write_seq; 31661da177e4SLinus Torvalds tp->snd_sml = tp->write_seq; 316733f5f57eSIlpo Järvinen tp->snd_up = tp->write_seq; 3168370816aeSPavel Emelyanov tp->snd_nxt = tp->write_seq; 3169ee995283SPavel Emelyanov 3170ee995283SPavel Emelyanov if (likely(!tp->repair)) 31711da177e4SLinus Torvalds tp->rcv_nxt = 0; 3172c7781a6eSAndrew Vagin else 3173c7781a6eSAndrew Vagin tp->rcv_tstamp = tcp_time_stamp; 3174ee995283SPavel Emelyanov tp->rcv_wup = tp->rcv_nxt; 3175ee995283SPavel Emelyanov tp->copied_seq = tp->rcv_nxt; 31761da177e4SLinus Torvalds 3177463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; 3178463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_retransmits = 0; 31791da177e4SLinus Torvalds tcp_clear_retrans(tp); 31801da177e4SLinus Torvalds } 31811da177e4SLinus Torvalds 3182783237e8SYuchung Cheng static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) 3183783237e8SYuchung Cheng { 3184783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3185783237e8SYuchung Cheng struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 3186783237e8SYuchung Cheng 3187783237e8SYuchung Cheng tcb->end_seq += skb->len; 3188f4a775d1SEric Dumazet __skb_header_release(skb); 3189783237e8SYuchung Cheng __tcp_add_write_queue_tail(sk, skb); 3190783237e8SYuchung Cheng sk->sk_wmem_queued += skb->truesize; 3191783237e8SYuchung Cheng sk_mem_charge(sk, skb->truesize); 3192783237e8SYuchung Cheng tp->write_seq = tcb->end_seq; 3193783237e8SYuchung Cheng tp->packets_out += tcp_skb_pcount(skb); 3194783237e8SYuchung Cheng } 3195783237e8SYuchung Cheng 3196783237e8SYuchung Cheng /* Build and send a SYN with data and (cached) Fast Open cookie. However, 3197783237e8SYuchung Cheng * queue a data-only packet after the regular SYN, such that regular SYNs 3198783237e8SYuchung Cheng * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges 3199783237e8SYuchung Cheng * only the SYN sequence, the data are retransmitted in the first ACK. 3200783237e8SYuchung Cheng * If cookie is not cached or other error occurs, falls back to send a 3201783237e8SYuchung Cheng * regular SYN with Fast Open cookie request option. 3202783237e8SYuchung Cheng */ 3203783237e8SYuchung Cheng static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) 3204783237e8SYuchung Cheng { 3205783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3206783237e8SYuchung Cheng struct tcp_fastopen_request *fo = tp->fastopen_req; 320707e100f9SEric Dumazet int syn_loss = 0, space, err = 0; 3208aab48743SYuchung Cheng unsigned long last_syn_loss = 0; 3209355a901eSEric Dumazet struct sk_buff *syn_data; 3210783237e8SYuchung Cheng 321167da22d2SYuchung Cheng tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ 3212aab48743SYuchung Cheng tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie, 3213aab48743SYuchung Cheng &syn_loss, &last_syn_loss); 3214aab48743SYuchung Cheng /* Recurring FO SYN losses: revert to regular handshake temporarily */ 3215aab48743SYuchung Cheng if (syn_loss > 1 && 3216aab48743SYuchung Cheng time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) { 3217aab48743SYuchung Cheng fo->cookie.len = -1; 3218aab48743SYuchung Cheng goto fallback; 3219aab48743SYuchung Cheng } 3220aab48743SYuchung Cheng 322167da22d2SYuchung Cheng if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE) 322267da22d2SYuchung Cheng fo->cookie.len = -1; 322367da22d2SYuchung Cheng else if (fo->cookie.len <= 0) 3224783237e8SYuchung Cheng goto fallback; 3225783237e8SYuchung Cheng 3226783237e8SYuchung Cheng /* MSS for SYN-data is based on cached MSS and bounded by PMTU and 3227783237e8SYuchung Cheng * user-MSS. Reserve maximum option space for middleboxes that add 3228783237e8SYuchung Cheng * private TCP options. The cost is reduced data space in SYN :( 3229783237e8SYuchung Cheng */ 3230783237e8SYuchung Cheng if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp) 3231783237e8SYuchung Cheng tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 32321b63edd6SYuchung Cheng space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) - 3233783237e8SYuchung Cheng MAX_TCP_OPTION_SPACE; 3234783237e8SYuchung Cheng 3235f5ddcbbbSEric Dumazet space = min_t(size_t, space, fo->size); 3236f5ddcbbbSEric Dumazet 3237f5ddcbbbSEric Dumazet /* limit to order-0 allocations */ 3238f5ddcbbbSEric Dumazet space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER)); 3239f5ddcbbbSEric Dumazet 3240eb934478SEric Dumazet syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false); 3241355a901eSEric Dumazet if (!syn_data) 3242783237e8SYuchung Cheng goto fallback; 3243355a901eSEric Dumazet syn_data->ip_summed = CHECKSUM_PARTIAL; 3244355a901eSEric Dumazet memcpy(syn_data->cb, syn->cb, sizeof(syn->cb)); 324507e100f9SEric Dumazet if (space) { 324607e100f9SEric Dumazet int copied = copy_from_iter(skb_put(syn_data, space), space, 324757be5bdaSAl Viro &fo->data->msg_iter); 324857be5bdaSAl Viro if (unlikely(!copied)) { 3249355a901eSEric Dumazet kfree_skb(syn_data); 3250783237e8SYuchung Cheng goto fallback; 3251783237e8SYuchung Cheng } 325257be5bdaSAl Viro if (copied != space) { 325357be5bdaSAl Viro skb_trim(syn_data, copied); 325457be5bdaSAl Viro space = copied; 325557be5bdaSAl Viro } 325607e100f9SEric Dumazet } 3257355a901eSEric Dumazet /* No more data pending in inet_wait_for_connect() */ 3258355a901eSEric Dumazet if (space == fo->size) 3259355a901eSEric Dumazet fo->data = NULL; 3260355a901eSEric Dumazet fo->copied = space; 3261783237e8SYuchung Cheng 3262355a901eSEric Dumazet tcp_connect_queue_skb(sk, syn_data); 3263355a901eSEric Dumazet 3264355a901eSEric Dumazet err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation); 3265355a901eSEric Dumazet 3266355a901eSEric Dumazet syn->skb_mstamp = syn_data->skb_mstamp; 3267355a901eSEric Dumazet 3268355a901eSEric Dumazet /* Now full SYN+DATA was cloned and sent (or not), 3269355a901eSEric Dumazet * remove the SYN from the original skb (syn_data) 3270355a901eSEric Dumazet * we keep in write queue in case of a retransmit, as we 3271355a901eSEric Dumazet * also have the SYN packet (with no data) in the same queue. 3272431a9124SEric Dumazet */ 3273355a901eSEric Dumazet TCP_SKB_CB(syn_data)->seq++; 3274355a901eSEric Dumazet TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH; 3275355a901eSEric Dumazet if (!err) { 327667da22d2SYuchung Cheng tp->syn_data = (fo->copied > 0); 3277f19c29e3SYuchung Cheng NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT); 3278783237e8SYuchung Cheng goto done; 3279783237e8SYuchung Cheng } 3280783237e8SYuchung Cheng 3281783237e8SYuchung Cheng fallback: 3282783237e8SYuchung Cheng /* Send a regular SYN with Fast Open cookie request option */ 3283783237e8SYuchung Cheng if (fo->cookie.len > 0) 3284783237e8SYuchung Cheng fo->cookie.len = 0; 3285783237e8SYuchung Cheng err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); 3286783237e8SYuchung Cheng if (err) 3287783237e8SYuchung Cheng tp->syn_fastopen = 0; 3288783237e8SYuchung Cheng done: 3289783237e8SYuchung Cheng fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ 3290783237e8SYuchung Cheng return err; 3291783237e8SYuchung Cheng } 3292783237e8SYuchung Cheng 329367edfef7SAndi Kleen /* Build a SYN and send it off. */ 32941da177e4SLinus Torvalds int tcp_connect(struct sock *sk) 32951da177e4SLinus Torvalds { 32961da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 32971da177e4SLinus Torvalds struct sk_buff *buff; 3298ee586811SEric Paris int err; 32991da177e4SLinus Torvalds 33001da177e4SLinus Torvalds tcp_connect_init(sk); 33011da177e4SLinus Torvalds 33022b916477SAndrey Vagin if (unlikely(tp->repair)) { 33032b916477SAndrey Vagin tcp_finish_connect(sk, NULL); 33042b916477SAndrey Vagin return 0; 33052b916477SAndrey Vagin } 33062b916477SAndrey Vagin 3307eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true); 3308355a901eSEric Dumazet if (unlikely(!buff)) 33091da177e4SLinus Torvalds return -ENOBUFS; 33101da177e4SLinus Torvalds 3311a3433f35SChangli Gao tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); 33127faee5c0SEric Dumazet tp->retrans_stamp = tcp_time_stamp; 3313783237e8SYuchung Cheng tcp_connect_queue_skb(sk, buff); 3314735d3831SFlorian Westphal tcp_ecn_send_syn(sk, buff); 33151da177e4SLinus Torvalds 3316783237e8SYuchung Cheng /* Send off SYN; include data in Fast Open. */ 3317783237e8SYuchung Cheng err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : 3318783237e8SYuchung Cheng tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); 3319ee586811SEric Paris if (err == -ECONNREFUSED) 3320ee586811SEric Paris return err; 3321bd37a088SWei Yongjun 3322bd37a088SWei Yongjun /* We change tp->snd_nxt after the tcp_transmit_skb() call 3323bd37a088SWei Yongjun * in order to make this packet get counted in tcpOutSegs. 3324bd37a088SWei Yongjun */ 3325bd37a088SWei Yongjun tp->snd_nxt = tp->write_seq; 3326bd37a088SWei Yongjun tp->pushed_seq = tp->write_seq; 332781cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); 33281da177e4SLinus Torvalds 33291da177e4SLinus Torvalds /* Timer for repeating the SYN until an answer. */ 33303f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 33313f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 33321da177e4SLinus Torvalds return 0; 33331da177e4SLinus Torvalds } 33344bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_connect); 33351da177e4SLinus Torvalds 33361da177e4SLinus Torvalds /* Send out a delayed ack, the caller does the policy checking 33371da177e4SLinus Torvalds * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 33381da177e4SLinus Torvalds * for details. 33391da177e4SLinus Torvalds */ 33401da177e4SLinus Torvalds void tcp_send_delayed_ack(struct sock *sk) 33411da177e4SLinus Torvalds { 3342463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 3343463c84b9SArnaldo Carvalho de Melo int ato = icsk->icsk_ack.ato; 33441da177e4SLinus Torvalds unsigned long timeout; 33451da177e4SLinus Torvalds 33469890092eSFlorian Westphal tcp_ca_event(sk, CA_EVENT_DELAYED_ACK); 33479890092eSFlorian Westphal 33481da177e4SLinus Torvalds if (ato > TCP_DELACK_MIN) { 3349463c84b9SArnaldo Carvalho de Melo const struct tcp_sock *tp = tcp_sk(sk); 33501da177e4SLinus Torvalds int max_ato = HZ / 2; 33511da177e4SLinus Torvalds 3352056834d9SIlpo Järvinen if (icsk->icsk_ack.pingpong || 3353056834d9SIlpo Järvinen (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 33541da177e4SLinus Torvalds max_ato = TCP_DELACK_MAX; 33551da177e4SLinus Torvalds 33561da177e4SLinus Torvalds /* Slow path, intersegment interval is "high". */ 33571da177e4SLinus Torvalds 33581da177e4SLinus Torvalds /* If some rtt estimate is known, use it to bound delayed ack. 3359463c84b9SArnaldo Carvalho de Melo * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 33601da177e4SLinus Torvalds * directly. 33611da177e4SLinus Torvalds */ 3362740b0f18SEric Dumazet if (tp->srtt_us) { 3363740b0f18SEric Dumazet int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3), 3364740b0f18SEric Dumazet TCP_DELACK_MIN); 33651da177e4SLinus Torvalds 33661da177e4SLinus Torvalds if (rtt < max_ato) 33671da177e4SLinus Torvalds max_ato = rtt; 33681da177e4SLinus Torvalds } 33691da177e4SLinus Torvalds 33701da177e4SLinus Torvalds ato = min(ato, max_ato); 33711da177e4SLinus Torvalds } 33721da177e4SLinus Torvalds 33731da177e4SLinus Torvalds /* Stay within the limit we were given */ 33741da177e4SLinus Torvalds timeout = jiffies + ato; 33751da177e4SLinus Torvalds 33761da177e4SLinus Torvalds /* Use new timeout only if there wasn't a older one earlier. */ 3377463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 33781da177e4SLinus Torvalds /* If delack timer was blocked or is about to expire, 33791da177e4SLinus Torvalds * send ACK now. 33801da177e4SLinus Torvalds */ 3381463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.blocked || 3382463c84b9SArnaldo Carvalho de Melo time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 33831da177e4SLinus Torvalds tcp_send_ack(sk); 33841da177e4SLinus Torvalds return; 33851da177e4SLinus Torvalds } 33861da177e4SLinus Torvalds 3387463c84b9SArnaldo Carvalho de Melo if (!time_before(timeout, icsk->icsk_ack.timeout)) 3388463c84b9SArnaldo Carvalho de Melo timeout = icsk->icsk_ack.timeout; 33891da177e4SLinus Torvalds } 3390463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 3391463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.timeout = timeout; 3392463c84b9SArnaldo Carvalho de Melo sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 33931da177e4SLinus Torvalds } 33941da177e4SLinus Torvalds 33951da177e4SLinus Torvalds /* This routine sends an ack and also updates the window. */ 33961da177e4SLinus Torvalds void tcp_send_ack(struct sock *sk) 33971da177e4SLinus Torvalds { 33981da177e4SLinus Torvalds struct sk_buff *buff; 33991da177e4SLinus Torvalds 3400058dc334SIlpo Järvinen /* If we have been reset, we may not send again. */ 3401058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3402058dc334SIlpo Järvinen return; 3403058dc334SIlpo Järvinen 34049890092eSFlorian Westphal tcp_ca_event(sk, CA_EVENT_NON_DELAYED_ACK); 34059890092eSFlorian Westphal 34061da177e4SLinus Torvalds /* We are not putting this on the write queue, so 34071da177e4SLinus Torvalds * tcp_transmit_skb() will set the ownership to this 34081da177e4SLinus Torvalds * sock. 34091da177e4SLinus Torvalds */ 34107450aaf6SEric Dumazet buff = alloc_skb(MAX_TCP_HEADER, 34117450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 34127450aaf6SEric Dumazet if (unlikely(!buff)) { 3413463c84b9SArnaldo Carvalho de Melo inet_csk_schedule_ack(sk); 3414463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; 34153f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 34163f421baaSArnaldo Carvalho de Melo TCP_DELACK_MAX, TCP_RTO_MAX); 34171da177e4SLinus Torvalds return; 34181da177e4SLinus Torvalds } 34191da177e4SLinus Torvalds 34201da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 34211da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 3422a3433f35SChangli Gao tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); 34231da177e4SLinus Torvalds 342498781965SEric Dumazet /* We do not want pure acks influencing TCP Small Queues or fq/pacing 342598781965SEric Dumazet * too much. 342698781965SEric Dumazet * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784 342798781965SEric Dumazet * We also avoid tcp_wfree() overhead (cache line miss accessing 342898781965SEric Dumazet * tp->tsq_flags) by using regular sock_wfree() 342998781965SEric Dumazet */ 343098781965SEric Dumazet skb_set_tcp_pure_ack(buff); 343198781965SEric Dumazet 34321da177e4SLinus Torvalds /* Send it off, this clears delayed acks for us. */ 34337faee5c0SEric Dumazet skb_mstamp_get(&buff->skb_mstamp); 34347450aaf6SEric Dumazet tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0); 34351da177e4SLinus Torvalds } 3436e3118e83SDaniel Borkmann EXPORT_SYMBOL_GPL(tcp_send_ack); 34371da177e4SLinus Torvalds 34381da177e4SLinus Torvalds /* This routine sends a packet with an out of date sequence 34391da177e4SLinus Torvalds * number. It assumes the other end will try to ack it. 34401da177e4SLinus Torvalds * 34411da177e4SLinus Torvalds * Question: what should we make while urgent mode? 34421da177e4SLinus Torvalds * 4.4BSD forces sending single byte of data. We cannot send 34431da177e4SLinus Torvalds * out of window data, because we have SND.NXT==SND.MAX... 34441da177e4SLinus Torvalds * 34451da177e4SLinus Torvalds * Current solution: to send TWO zero-length segments in urgent mode: 34461da177e4SLinus Torvalds * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 34471da177e4SLinus Torvalds * out-of-date with SND.UNA-1 to probe window. 34481da177e4SLinus Torvalds */ 3449e520af48SEric Dumazet static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib) 34501da177e4SLinus Torvalds { 34511da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 34521da177e4SLinus Torvalds struct sk_buff *skb; 34531da177e4SLinus Torvalds 34541da177e4SLinus Torvalds /* We don't queue it, tcp_transmit_skb() sets ownership. */ 34557450aaf6SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, 34567450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 345751456b29SIan Morris if (!skb) 34581da177e4SLinus Torvalds return -1; 34591da177e4SLinus Torvalds 34601da177e4SLinus Torvalds /* Reserve space for headers and set control bits. */ 34611da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 34621da177e4SLinus Torvalds /* Use a previous sequence. This should cause the other 34631da177e4SLinus Torvalds * end to send an ack. Don't queue or clone SKB, just 34641da177e4SLinus Torvalds * send it. 34651da177e4SLinus Torvalds */ 3466a3433f35SChangli Gao tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); 34677faee5c0SEric Dumazet skb_mstamp_get(&skb->skb_mstamp); 3468e2e8009fSRenato Westphal NET_INC_STATS(sock_net(sk), mib); 34697450aaf6SEric Dumazet return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0); 34701da177e4SLinus Torvalds } 34711da177e4SLinus Torvalds 3472ee995283SPavel Emelyanov void tcp_send_window_probe(struct sock *sk) 3473ee995283SPavel Emelyanov { 3474ee995283SPavel Emelyanov if (sk->sk_state == TCP_ESTABLISHED) { 3475ee995283SPavel Emelyanov tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; 3476e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE); 3477ee995283SPavel Emelyanov } 3478ee995283SPavel Emelyanov } 3479ee995283SPavel Emelyanov 348067edfef7SAndi Kleen /* Initiate keepalive or window probe from timer. */ 3481e520af48SEric Dumazet int tcp_write_wakeup(struct sock *sk, int mib) 34821da177e4SLinus Torvalds { 34831da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 34841da177e4SLinus Torvalds struct sk_buff *skb; 34851da177e4SLinus Torvalds 3486058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3487058dc334SIlpo Järvinen return -1; 3488058dc334SIlpo Järvinen 348900db4124SIan Morris skb = tcp_send_head(sk); 349000db4124SIan Morris if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { 34911da177e4SLinus Torvalds int err; 34920c54b85fSIlpo Järvinen unsigned int mss = tcp_current_mss(sk); 349390840defSIlpo Järvinen unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 34941da177e4SLinus Torvalds 34951da177e4SLinus Torvalds if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 34961da177e4SLinus Torvalds tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 34971da177e4SLinus Torvalds 34981da177e4SLinus Torvalds /* We are probing the opening of a window 34991da177e4SLinus Torvalds * but the window size is != 0 35001da177e4SLinus Torvalds * must have been a result SWS avoidance ( sender ) 35011da177e4SLinus Torvalds */ 35021da177e4SLinus Torvalds if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 35031da177e4SLinus Torvalds skb->len > mss) { 35041da177e4SLinus Torvalds seg_size = min(seg_size, mss); 35054de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 35066cc55e09SOctavian Purdila if (tcp_fragment(sk, skb, seg_size, mss, GFP_ATOMIC)) 35071da177e4SLinus Torvalds return -1; 35081da177e4SLinus Torvalds } else if (!tcp_skb_pcount(skb)) 35095bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss); 35101da177e4SLinus Torvalds 35114de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 3512dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 351366f5fe62SIlpo Järvinen if (!err) 351466f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 35151da177e4SLinus Torvalds return err; 35161da177e4SLinus Torvalds } else { 351733f5f57eSIlpo Järvinen if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) 3518e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 1, mib); 3519e520af48SEric Dumazet return tcp_xmit_probe_skb(sk, 0, mib); 35201da177e4SLinus Torvalds } 35211da177e4SLinus Torvalds } 35221da177e4SLinus Torvalds 35231da177e4SLinus Torvalds /* A window probe timeout has occurred. If window is not closed send 35241da177e4SLinus Torvalds * a partial packet else a zero probe. 35251da177e4SLinus Torvalds */ 35261da177e4SLinus Torvalds void tcp_send_probe0(struct sock *sk) 35271da177e4SLinus Torvalds { 3528463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 35291da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 3530c6214a97SNikolay Borisov struct net *net = sock_net(sk); 3531fcdd1cf4SEric Dumazet unsigned long probe_max; 35321da177e4SLinus Torvalds int err; 35331da177e4SLinus Torvalds 3534e520af48SEric Dumazet err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE); 35351da177e4SLinus Torvalds 3536fe067e8aSDavid S. Miller if (tp->packets_out || !tcp_send_head(sk)) { 35371da177e4SLinus Torvalds /* Cancel probe timer, if it is not required. */ 35386687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 0; 3539463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff = 0; 35401da177e4SLinus Torvalds return; 35411da177e4SLinus Torvalds } 35421da177e4SLinus Torvalds 35431da177e4SLinus Torvalds if (err <= 0) { 3544c6214a97SNikolay Borisov if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2) 3545463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff++; 35466687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out++; 3547fcdd1cf4SEric Dumazet probe_max = TCP_RTO_MAX; 35481da177e4SLinus Torvalds } else { 35491da177e4SLinus Torvalds /* If packet was not sent due to local congestion, 35506687e988SArnaldo Carvalho de Melo * do not backoff and do not remember icsk_probes_out. 35511da177e4SLinus Torvalds * Let local senders to fight for local resources. 35521da177e4SLinus Torvalds * 35531da177e4SLinus Torvalds * Use accumulated backoff yet. 35541da177e4SLinus Torvalds */ 35556687e988SArnaldo Carvalho de Melo if (!icsk->icsk_probes_out) 35566687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 1; 3557fcdd1cf4SEric Dumazet probe_max = TCP_RESOURCE_PROBE_INTERVAL; 35581da177e4SLinus Torvalds } 3559fcdd1cf4SEric Dumazet inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 356021c8fe99SEric Dumazet tcp_probe0_when(sk, probe_max), 3561fcdd1cf4SEric Dumazet TCP_RTO_MAX); 35621da177e4SLinus Torvalds } 35635db92c99SOctavian Purdila 3564ea3bea3aSEric Dumazet int tcp_rtx_synack(const struct sock *sk, struct request_sock *req) 35655db92c99SOctavian Purdila { 35665db92c99SOctavian Purdila const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific; 35675db92c99SOctavian Purdila struct flowi fl; 35685db92c99SOctavian Purdila int res; 35695db92c99SOctavian Purdila 357058d607d3SEric Dumazet tcp_rsk(req)->txhash = net_tx_rndhash(); 3571b3d05147SEric Dumazet res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL); 35725db92c99SOctavian Purdila if (!res) { 357390bbcc60SEric Dumazet __TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); 357402a1d6e7SEric Dumazet __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 35755db92c99SOctavian Purdila } 35765db92c99SOctavian Purdila return res; 35775db92c99SOctavian Purdila } 35785db92c99SOctavian Purdila EXPORT_SYMBOL(tcp_rtx_synack); 3579