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; 64395a22caeSFlorian Westphal opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off; 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 | 737f9616c35SEric Dumazet TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) { 738f9616c35SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 739f9616c35SEric Dumazet 740f9616c35SEric Dumazet if (tp->lost_out > tp->retrans_out && 741f9616c35SEric Dumazet tp->snd_cwnd > tcp_packets_in_flight(tp)) 742f9616c35SEric Dumazet tcp_xmit_retransmit_queue(sk); 743f9616c35SEric Dumazet 744f9616c35SEric Dumazet tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle, 745bf06200eSJohn Ogness 0, GFP_ATOMIC); 7466f458dfbSEric Dumazet } 747f9616c35SEric Dumazet } 74846d3ceabSEric Dumazet /* 7498e3bff96Sstephen hemminger * One tasklet per cpu tries to send more skbs. 75046d3ceabSEric Dumazet * We run in tasklet context but need to disable irqs when 7518e3bff96Sstephen hemminger * transferring tsq->head because tcp_wfree() might 75246d3ceabSEric Dumazet * interrupt us (non NAPI drivers) 75346d3ceabSEric Dumazet */ 75446d3ceabSEric Dumazet static void tcp_tasklet_func(unsigned long data) 75546d3ceabSEric Dumazet { 75646d3ceabSEric Dumazet struct tsq_tasklet *tsq = (struct tsq_tasklet *)data; 75746d3ceabSEric Dumazet LIST_HEAD(list); 75846d3ceabSEric Dumazet unsigned long flags; 75946d3ceabSEric Dumazet struct list_head *q, *n; 76046d3ceabSEric Dumazet struct tcp_sock *tp; 76146d3ceabSEric Dumazet struct sock *sk; 76246d3ceabSEric Dumazet 76346d3ceabSEric Dumazet local_irq_save(flags); 76446d3ceabSEric Dumazet list_splice_init(&tsq->head, &list); 76546d3ceabSEric Dumazet local_irq_restore(flags); 76646d3ceabSEric Dumazet 76746d3ceabSEric Dumazet list_for_each_safe(q, n, &list) { 76846d3ceabSEric Dumazet tp = list_entry(q, struct tcp_sock, tsq_node); 76946d3ceabSEric Dumazet list_del(&tp->tsq_node); 77046d3ceabSEric Dumazet 77146d3ceabSEric Dumazet sk = (struct sock *)tp; 772*0a9648f1SEric Dumazet smp_mb__before_atomic(); 7737aa5470cSEric Dumazet clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags); 7747aa5470cSEric Dumazet 775b223feb9SEric Dumazet if (!sk->sk_lock.owned && 7767aa5470cSEric Dumazet test_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags)) { 77746d3ceabSEric Dumazet bh_lock_sock(sk); 77846d3ceabSEric Dumazet if (!sock_owned_by_user(sk)) { 7797aa5470cSEric Dumazet clear_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags); 7806f458dfbSEric Dumazet tcp_tsq_handler(sk); 78146d3ceabSEric Dumazet } 78246d3ceabSEric Dumazet bh_unlock_sock(sk); 783b223feb9SEric Dumazet } 78446d3ceabSEric Dumazet 78546d3ceabSEric Dumazet sk_free(sk); 78646d3ceabSEric Dumazet } 78746d3ceabSEric Dumazet } 78846d3ceabSEric Dumazet 78940fc3423SEric Dumazet #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \ 79040fc3423SEric Dumazet TCPF_WRITE_TIMER_DEFERRED | \ 79140fc3423SEric Dumazet TCPF_DELACK_TIMER_DEFERRED | \ 79240fc3423SEric Dumazet TCPF_MTU_REDUCED_DEFERRED) 79346d3ceabSEric Dumazet /** 79446d3ceabSEric Dumazet * tcp_release_cb - tcp release_sock() callback 79546d3ceabSEric Dumazet * @sk: socket 79646d3ceabSEric Dumazet * 79746d3ceabSEric Dumazet * called from release_sock() to perform protocol dependent 79846d3ceabSEric Dumazet * actions before socket release. 79946d3ceabSEric Dumazet */ 80046d3ceabSEric Dumazet void tcp_release_cb(struct sock *sk) 80146d3ceabSEric Dumazet { 8026f458dfbSEric Dumazet unsigned long flags, nflags; 80346d3ceabSEric Dumazet 8046f458dfbSEric Dumazet /* perform an atomic operation only if at least one flag is set */ 8056f458dfbSEric Dumazet do { 8067aa5470cSEric Dumazet flags = sk->sk_tsq_flags; 8076f458dfbSEric Dumazet if (!(flags & TCP_DEFERRED_ALL)) 8086f458dfbSEric Dumazet return; 8096f458dfbSEric Dumazet nflags = flags & ~TCP_DEFERRED_ALL; 8107aa5470cSEric Dumazet } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags); 8116f458dfbSEric Dumazet 81240fc3423SEric Dumazet if (flags & TCPF_TSQ_DEFERRED) 8136f458dfbSEric Dumazet tcp_tsq_handler(sk); 8146f458dfbSEric Dumazet 815c3f9b018SEric Dumazet /* Here begins the tricky part : 816c3f9b018SEric Dumazet * We are called from release_sock() with : 817c3f9b018SEric Dumazet * 1) BH disabled 818c3f9b018SEric Dumazet * 2) sk_lock.slock spinlock held 819c3f9b018SEric Dumazet * 3) socket owned by us (sk->sk_lock.owned == 1) 820c3f9b018SEric Dumazet * 821c3f9b018SEric Dumazet * But following code is meant to be called from BH handlers, 822c3f9b018SEric Dumazet * so we should keep BH disabled, but early release socket ownership 823c3f9b018SEric Dumazet */ 824c3f9b018SEric Dumazet sock_release_ownership(sk); 825c3f9b018SEric Dumazet 82640fc3423SEric Dumazet if (flags & TCPF_WRITE_TIMER_DEFERRED) { 8276f458dfbSEric Dumazet tcp_write_timer_handler(sk); 828144d56e9SEric Dumazet __sock_put(sk); 829144d56e9SEric Dumazet } 83040fc3423SEric Dumazet if (flags & TCPF_DELACK_TIMER_DEFERRED) { 8316f458dfbSEric Dumazet tcp_delack_timer_handler(sk); 832144d56e9SEric Dumazet __sock_put(sk); 833144d56e9SEric Dumazet } 83440fc3423SEric Dumazet if (flags & TCPF_MTU_REDUCED_DEFERRED) { 8354fab9071SNeal Cardwell inet_csk(sk)->icsk_af_ops->mtu_reduced(sk); 836144d56e9SEric Dumazet __sock_put(sk); 837144d56e9SEric Dumazet } 83846d3ceabSEric Dumazet } 83946d3ceabSEric Dumazet EXPORT_SYMBOL(tcp_release_cb); 84046d3ceabSEric Dumazet 84146d3ceabSEric Dumazet void __init tcp_tasklet_init(void) 84246d3ceabSEric Dumazet { 84346d3ceabSEric Dumazet int i; 84446d3ceabSEric Dumazet 84546d3ceabSEric Dumazet for_each_possible_cpu(i) { 84646d3ceabSEric Dumazet struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i); 84746d3ceabSEric Dumazet 84846d3ceabSEric Dumazet INIT_LIST_HEAD(&tsq->head); 84946d3ceabSEric Dumazet tasklet_init(&tsq->tasklet, 85046d3ceabSEric Dumazet tcp_tasklet_func, 85146d3ceabSEric Dumazet (unsigned long)tsq); 85246d3ceabSEric Dumazet } 85346d3ceabSEric Dumazet } 85446d3ceabSEric Dumazet 85546d3ceabSEric Dumazet /* 85646d3ceabSEric Dumazet * Write buffer destructor automatically called from kfree_skb. 8578e3bff96Sstephen hemminger * We can't xmit new skbs from this context, as we might already 85846d3ceabSEric Dumazet * hold qdisc lock. 85946d3ceabSEric Dumazet */ 860d6a4a104SEric Dumazet void tcp_wfree(struct sk_buff *skb) 86146d3ceabSEric Dumazet { 86246d3ceabSEric Dumazet struct sock *sk = skb->sk; 86346d3ceabSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 864408f0a6cSEric Dumazet unsigned long flags, nval, oval; 8659b462d02SEric Dumazet int wmem; 8669b462d02SEric Dumazet 8679b462d02SEric Dumazet /* Keep one reference on sk_wmem_alloc. 8689b462d02SEric Dumazet * Will be released by sk_free() from here or tcp_tasklet_func() 8699b462d02SEric Dumazet */ 8709b462d02SEric Dumazet wmem = atomic_sub_return(skb->truesize - 1, &sk->sk_wmem_alloc); 8719b462d02SEric Dumazet 8729b462d02SEric Dumazet /* If this softirq is serviced by ksoftirqd, we are likely under stress. 8739b462d02SEric Dumazet * Wait until our queues (qdisc + devices) are drained. 8749b462d02SEric Dumazet * This gives : 8759b462d02SEric Dumazet * - less callbacks to tcp_write_xmit(), reducing stress (batches) 8769b462d02SEric Dumazet * - chance for incoming ACK (processed by another cpu maybe) 8779b462d02SEric Dumazet * to migrate this flow (skb->ooo_okay will be eventually set) 8789b462d02SEric Dumazet */ 8799b462d02SEric Dumazet if (wmem >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current) 8809b462d02SEric Dumazet goto out; 88146d3ceabSEric Dumazet 8827aa5470cSEric Dumazet for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) { 88346d3ceabSEric Dumazet struct tsq_tasklet *tsq; 884a9b204d1SEric Dumazet bool empty; 88546d3ceabSEric Dumazet 886408f0a6cSEric Dumazet if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED)) 887408f0a6cSEric Dumazet goto out; 888408f0a6cSEric Dumazet 889b223feb9SEric Dumazet nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED | TCPF_TSQ_DEFERRED; 8907aa5470cSEric Dumazet nval = cmpxchg(&sk->sk_tsq_flags, oval, nval); 891408f0a6cSEric Dumazet if (nval != oval) 892408f0a6cSEric Dumazet continue; 893408f0a6cSEric Dumazet 89446d3ceabSEric Dumazet /* queue this socket to tasklet queue */ 89546d3ceabSEric Dumazet local_irq_save(flags); 896903ceff7SChristoph Lameter tsq = this_cpu_ptr(&tsq_tasklet); 897a9b204d1SEric Dumazet empty = list_empty(&tsq->head); 89846d3ceabSEric Dumazet list_add(&tp->tsq_node, &tsq->head); 899a9b204d1SEric Dumazet if (empty) 90046d3ceabSEric Dumazet tasklet_schedule(&tsq->tasklet); 90146d3ceabSEric Dumazet local_irq_restore(flags); 9029b462d02SEric Dumazet return; 90346d3ceabSEric Dumazet } 9049b462d02SEric Dumazet out: 9059b462d02SEric Dumazet sk_free(sk); 90646d3ceabSEric Dumazet } 90746d3ceabSEric Dumazet 9081da177e4SLinus Torvalds /* This routine actually transmits TCP packets queued in by 9091da177e4SLinus Torvalds * tcp_do_sendmsg(). This is used by both the initial 9101da177e4SLinus Torvalds * transmission and possible later retransmissions. 9111da177e4SLinus Torvalds * All SKB's seen here are completely headerless. It is our 9121da177e4SLinus Torvalds * job to build the TCP header, and pass the packet down to 9131da177e4SLinus Torvalds * IP so it can do the same plus pass the packet off to the 9141da177e4SLinus Torvalds * device. 9151da177e4SLinus Torvalds * 9161da177e4SLinus Torvalds * We are working here with either a clone of the original 9171da177e4SLinus Torvalds * SKB, or a fresh unique copy made by the retransmit engine. 9181da177e4SLinus Torvalds */ 919056834d9SIlpo Järvinen static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, 920056834d9SIlpo Järvinen gfp_t gfp_mask) 9211da177e4SLinus Torvalds { 9226687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 923dfb4b9dcSDavid S. Miller struct inet_sock *inet; 924dfb4b9dcSDavid S. Miller struct tcp_sock *tp; 925dfb4b9dcSDavid S. Miller struct tcp_skb_cb *tcb; 92633ad798cSAdam Langley struct tcp_out_options opts; 92795c96174SEric Dumazet unsigned int tcp_options_size, tcp_header_size; 928cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 9291da177e4SLinus Torvalds struct tcphdr *th; 9301da177e4SLinus Torvalds int err; 9311da177e4SLinus Torvalds 932dfb4b9dcSDavid S. Miller BUG_ON(!skb || !tcp_skb_pcount(skb)); 9336f094b9eSLawrence Brakmo tp = tcp_sk(sk); 934dfb4b9dcSDavid S. Miller 935ccdbb6e9SEric Dumazet if (clone_it) { 936740b0f18SEric Dumazet skb_mstamp_get(&skb->skb_mstamp); 9376f094b9eSLawrence Brakmo TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq 9386f094b9eSLawrence Brakmo - tp->snd_una; 939b9f64820SYuchung Cheng tcp_rate_skb_sent(sk, skb); 940dfb4b9dcSDavid S. Miller 941dfb4b9dcSDavid S. Miller if (unlikely(skb_cloned(skb))) 942dfb4b9dcSDavid S. Miller skb = pskb_copy(skb, gfp_mask); 943dfb4b9dcSDavid S. Miller else 944dfb4b9dcSDavid S. Miller skb = skb_clone(skb, gfp_mask); 945dfb4b9dcSDavid S. Miller if (unlikely(!skb)) 946dfb4b9dcSDavid S. Miller return -ENOBUFS; 947dfb4b9dcSDavid S. Miller } 948dfb4b9dcSDavid S. Miller 949dfb4b9dcSDavid S. Miller inet = inet_sk(sk); 950dfb4b9dcSDavid S. Miller tcb = TCP_SKB_CB(skb); 95133ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 9521da177e4SLinus Torvalds 9534de075e0SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) 95433ad798cSAdam Langley tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); 95533ad798cSAdam Langley else 95633ad798cSAdam Langley tcp_options_size = tcp_established_options(sk, skb, &opts, 95733ad798cSAdam Langley &md5); 95833ad798cSAdam Langley tcp_header_size = tcp_options_size + sizeof(struct tcphdr); 9591da177e4SLinus Torvalds 960547669d4SEric Dumazet /* if no packet is in qdisc/device queue, then allow XPS to select 961b2532eb9SEric Dumazet * another queue. We can be called from tcp_tsq_handler() 962b2532eb9SEric Dumazet * which holds one reference to sk_wmem_alloc. 963b2532eb9SEric Dumazet * 964b2532eb9SEric Dumazet * TODO: Ideally, in-flight pure ACK packets should not matter here. 965b2532eb9SEric Dumazet * One way to get this would be to set skb->truesize = 2 on them. 966547669d4SEric Dumazet */ 967b2532eb9SEric Dumazet skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1); 9681da177e4SLinus Torvalds 969aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 970aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 97146d3ceabSEric Dumazet 97246d3ceabSEric Dumazet skb_orphan(skb); 97346d3ceabSEric Dumazet skb->sk = sk; 9741d2077acSEric Dumazet skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree; 975b73c3d0eSTom Herbert skb_set_hash_from_sk(skb, sk); 97646d3ceabSEric Dumazet atomic_add(skb->truesize, &sk->sk_wmem_alloc); 9771da177e4SLinus Torvalds 9781da177e4SLinus Torvalds /* Build TCP header and checksum it. */ 979ea1627c2SEric Dumazet th = (struct tcphdr *)skb->data; 980c720c7e8SEric Dumazet th->source = inet->inet_sport; 981c720c7e8SEric Dumazet th->dest = inet->inet_dport; 9821da177e4SLinus Torvalds th->seq = htonl(tcb->seq); 9831da177e4SLinus Torvalds th->ack_seq = htonl(tp->rcv_nxt); 984df7a3b07SAl Viro *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | 9854de075e0SEric Dumazet tcb->tcp_flags); 986dfb4b9dcSDavid S. Miller 9871da177e4SLinus Torvalds th->check = 0; 9881da177e4SLinus Torvalds th->urg_ptr = 0; 9891da177e4SLinus Torvalds 99033f5f57eSIlpo Järvinen /* The urg_mode check is necessary during a below snd_una win probe */ 9917691367dSHerbert Xu if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { 9927691367dSHerbert Xu if (before(tp->snd_up, tcb->seq + 0x10000)) { 9931da177e4SLinus Torvalds th->urg_ptr = htons(tp->snd_up - tcb->seq); 9941da177e4SLinus Torvalds th->urg = 1; 9957691367dSHerbert Xu } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { 9960eae88f3SEric Dumazet th->urg_ptr = htons(0xFFFF); 9977691367dSHerbert Xu th->urg = 1; 9987691367dSHerbert Xu } 9991da177e4SLinus Torvalds } 10001da177e4SLinus Torvalds 1001bd0388aeSWilliam Allen Simpson tcp_options_write((__be32 *)(th + 1), tp, &opts); 100251466a75SEric Dumazet skb_shinfo(skb)->gso_type = sk->sk_gso_type; 1003ea1627c2SEric Dumazet if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) { 1004ea1627c2SEric Dumazet th->window = htons(tcp_select_window(sk)); 1005ea1627c2SEric Dumazet tcp_ecn_send(sk, skb, th, tcp_header_size); 1006ea1627c2SEric Dumazet } else { 1007ea1627c2SEric Dumazet /* RFC1323: The window in SYN & SYN/ACK segments 1008ea1627c2SEric Dumazet * is never scaled. 1009ea1627c2SEric Dumazet */ 1010ea1627c2SEric Dumazet th->window = htons(min(tp->rcv_wnd, 65535U)); 1011ea1627c2SEric Dumazet } 1012cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 1013cfb6eeb4SYOSHIFUJI Hideaki /* Calculate the MD5 hash, as we have all we need now */ 1014cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 1015a465419bSEric Dumazet sk_nocaps_add(sk, NETIF_F_GSO_MASK); 1016bd0388aeSWilliam Allen Simpson tp->af_specific->calc_md5_hash(opts.hash_location, 101739f8e58eSEric Dumazet md5, sk, skb); 1018cfb6eeb4SYOSHIFUJI Hideaki } 1019cfb6eeb4SYOSHIFUJI Hideaki #endif 1020cfb6eeb4SYOSHIFUJI Hideaki 1021bb296246SHerbert Xu icsk->icsk_af_ops->send_check(sk, skb); 10221da177e4SLinus Torvalds 10234de075e0SEric Dumazet if (likely(tcb->tcp_flags & TCPHDR_ACK)) 1024fc6415bcSDavid S. Miller tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); 10251da177e4SLinus Torvalds 1026a44d6eacSMartin KaFai Lau if (skb->len != tcp_header_size) { 1027cf533ea5SEric Dumazet tcp_event_data_sent(tp, sk); 1028a44d6eacSMartin KaFai Lau tp->data_segs_out += tcp_skb_pcount(skb); 1029a44d6eacSMartin KaFai Lau } 10301da177e4SLinus Torvalds 1031bd37a088SWei Yongjun if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) 1032aa2ea058STom Herbert TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, 1033aa2ea058STom Herbert tcp_skb_pcount(skb)); 10341da177e4SLinus Torvalds 10352efd055cSMarcelo Ricardo Leitner tp->segs_out += tcp_skb_pcount(skb); 1036f69ad292SEric Dumazet /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */ 1037cd7d8498SEric Dumazet skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb); 1038f69ad292SEric Dumazet skb_shinfo(skb)->gso_size = tcp_skb_mss(skb); 1039cd7d8498SEric Dumazet 10407faee5c0SEric Dumazet /* Our usage of tstamp should remain private */ 10417faee5c0SEric Dumazet skb->tstamp.tv64 = 0; 1042971f10ecSEric Dumazet 1043971f10ecSEric Dumazet /* Cleanup our debris for IP stacks */ 1044971f10ecSEric Dumazet memset(skb->cb, 0, max(sizeof(struct inet_skb_parm), 1045971f10ecSEric Dumazet sizeof(struct inet6_skb_parm))); 1046971f10ecSEric Dumazet 1047b0270e91SEric Dumazet err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl); 10487faee5c0SEric Dumazet 104983de47cdSHua Zhong if (likely(err <= 0)) 10501da177e4SLinus Torvalds return err; 10511da177e4SLinus Torvalds 10525ee2c941SChristoph Paasch tcp_enter_cwr(sk); 10531da177e4SLinus Torvalds 1054b9df3cb8SGerrit Renker return net_xmit_eval(err); 10551da177e4SLinus Torvalds } 10561da177e4SLinus Torvalds 105767edfef7SAndi Kleen /* This routine just queues the buffer for sending. 10581da177e4SLinus Torvalds * 10591da177e4SLinus Torvalds * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, 10601da177e4SLinus Torvalds * otherwise socket can stall. 10611da177e4SLinus Torvalds */ 10621da177e4SLinus Torvalds static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) 10631da177e4SLinus Torvalds { 10641da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 10651da177e4SLinus Torvalds 10661da177e4SLinus Torvalds /* Advance write_seq and place onto the write_queue. */ 10671da177e4SLinus Torvalds tp->write_seq = TCP_SKB_CB(skb)->end_seq; 1068f4a775d1SEric Dumazet __skb_header_release(skb); 1069fe067e8aSDavid S. Miller tcp_add_write_queue_tail(sk, skb); 10703ab224beSHideo Aoki sk->sk_wmem_queued += skb->truesize; 10713ab224beSHideo Aoki sk_mem_charge(sk, skb->truesize); 10721da177e4SLinus Torvalds } 10731da177e4SLinus Torvalds 107467edfef7SAndi Kleen /* Initialize TSO segments for a packet. */ 10755bbb432cSEric Dumazet static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now) 1076f6302d1dSDavid S. Miller { 10778f26fb1cSEric Dumazet if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) { 1078f6302d1dSDavid S. Miller /* Avoid the costly divide in the normal 1079f6302d1dSDavid S. Miller * non-TSO case. 1080f6302d1dSDavid S. Miller */ 1081cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, 1); 1082f69ad292SEric Dumazet TCP_SKB_CB(skb)->tcp_gso_size = 0; 1083f6302d1dSDavid S. Miller } else { 1084cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now)); 1085f69ad292SEric Dumazet TCP_SKB_CB(skb)->tcp_gso_size = mss_now; 10861da177e4SLinus Torvalds } 10871da177e4SLinus Torvalds } 10881da177e4SLinus Torvalds 108991fed7a1SIlpo Järvinen /* When a modification to fackets out becomes necessary, we need to check 109068f8353bSIlpo Järvinen * skb is counted to fackets_out or not. 109191fed7a1SIlpo Järvinen */ 1092cf533ea5SEric Dumazet static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb, 109391fed7a1SIlpo Järvinen int decr) 109491fed7a1SIlpo Järvinen { 1095a47e5a98SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1096a47e5a98SIlpo Järvinen 1097dc86967bSIlpo Järvinen if (!tp->sacked_out || tcp_is_reno(tp)) 109891fed7a1SIlpo Järvinen return; 109991fed7a1SIlpo Järvinen 11006859d494SIlpo Järvinen if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq)) 110191fed7a1SIlpo Järvinen tp->fackets_out -= decr; 110291fed7a1SIlpo Järvinen } 110391fed7a1SIlpo Järvinen 1104797108d1SIlpo Järvinen /* Pcount in the middle of the write queue got changed, we need to do various 1105797108d1SIlpo Järvinen * tweaks to fix counters 1106797108d1SIlpo Järvinen */ 1107cf533ea5SEric Dumazet static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) 1108797108d1SIlpo Järvinen { 1109797108d1SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1110797108d1SIlpo Järvinen 1111797108d1SIlpo Järvinen tp->packets_out -= decr; 1112797108d1SIlpo Järvinen 1113797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 1114797108d1SIlpo Järvinen tp->sacked_out -= decr; 1115797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 1116797108d1SIlpo Järvinen tp->retrans_out -= decr; 1117797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 1118797108d1SIlpo Järvinen tp->lost_out -= decr; 1119797108d1SIlpo Järvinen 1120797108d1SIlpo Järvinen /* Reno case is special. Sigh... */ 1121797108d1SIlpo Järvinen if (tcp_is_reno(tp) && decr > 0) 1122797108d1SIlpo Järvinen tp->sacked_out -= min_t(u32, tp->sacked_out, decr); 1123797108d1SIlpo Järvinen 1124797108d1SIlpo Järvinen tcp_adjust_fackets_out(sk, skb, decr); 1125797108d1SIlpo Järvinen 1126797108d1SIlpo Järvinen if (tp->lost_skb_hint && 1127797108d1SIlpo Järvinen before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && 112852cf3cc8SIlpo Järvinen (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))) 1129797108d1SIlpo Järvinen tp->lost_cnt_hint -= decr; 1130797108d1SIlpo Järvinen 1131797108d1SIlpo Järvinen tcp_verify_left_out(tp); 1132797108d1SIlpo Järvinen } 1133797108d1SIlpo Järvinen 11340a2cf20cSSoheil Hassas Yeganeh static bool tcp_has_tx_tstamp(const struct sk_buff *skb) 11350a2cf20cSSoheil Hassas Yeganeh { 11360a2cf20cSSoheil Hassas Yeganeh return TCP_SKB_CB(skb)->txstamp_ack || 11370a2cf20cSSoheil Hassas Yeganeh (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP); 11380a2cf20cSSoheil Hassas Yeganeh } 11390a2cf20cSSoheil Hassas Yeganeh 1140490cc7d0SWillem de Bruijn static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2) 1141490cc7d0SWillem de Bruijn { 1142490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo = skb_shinfo(skb); 1143490cc7d0SWillem de Bruijn 11440a2cf20cSSoheil Hassas Yeganeh if (unlikely(tcp_has_tx_tstamp(skb)) && 1145490cc7d0SWillem de Bruijn !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) { 1146490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo2 = skb_shinfo(skb2); 1147490cc7d0SWillem de Bruijn u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP; 1148490cc7d0SWillem de Bruijn 1149490cc7d0SWillem de Bruijn shinfo->tx_flags &= ~tsflags; 1150490cc7d0SWillem de Bruijn shinfo2->tx_flags |= tsflags; 1151490cc7d0SWillem de Bruijn swap(shinfo->tskey, shinfo2->tskey); 1152b51e13faSMartin KaFai Lau TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack; 1153b51e13faSMartin KaFai Lau TCP_SKB_CB(skb)->txstamp_ack = 0; 1154490cc7d0SWillem de Bruijn } 1155490cc7d0SWillem de Bruijn } 1156490cc7d0SWillem de Bruijn 1157a166140eSMartin KaFai Lau static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2) 1158a166140eSMartin KaFai Lau { 1159a166140eSMartin KaFai Lau TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor; 1160a166140eSMartin KaFai Lau TCP_SKB_CB(skb)->eor = 0; 1161a166140eSMartin KaFai Lau } 1162a166140eSMartin KaFai Lau 11631da177e4SLinus Torvalds /* Function to create two new TCP segments. Shrinks the given segment 11641da177e4SLinus Torvalds * to the specified size and appends a new segment with the rest of the 11651da177e4SLinus Torvalds * packet to the list. This won't be called frequently, I hope. 11661da177e4SLinus Torvalds * Remember, these are still headerless SKBs at this point. 11671da177e4SLinus Torvalds */ 1168056834d9SIlpo Järvinen int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, 11696cc55e09SOctavian Purdila unsigned int mss_now, gfp_t gfp) 11701da177e4SLinus Torvalds { 11711da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 11721da177e4SLinus Torvalds struct sk_buff *buff; 11736475be16SDavid S. Miller int nsize, old_factor; 1174b60b49eaSHerbert Xu int nlen; 11759ce01461SIlpo Järvinen u8 flags; 11761da177e4SLinus Torvalds 11772fceec13SIlpo Järvinen if (WARN_ON(len > skb->len)) 11782fceec13SIlpo Järvinen return -EINVAL; 11796a438bbeSStephen Hemminger 11801da177e4SLinus Torvalds nsize = skb_headlen(skb) - len; 11811da177e4SLinus Torvalds if (nsize < 0) 11821da177e4SLinus Torvalds nsize = 0; 11831da177e4SLinus Torvalds 11846cc55e09SOctavian Purdila if (skb_unclone(skb, gfp)) 11851da177e4SLinus Torvalds return -ENOMEM; 11861da177e4SLinus Torvalds 11871da177e4SLinus Torvalds /* Get a new skb... force flag on. */ 1188eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, nsize, gfp, true); 118951456b29SIan Morris if (!buff) 11901da177e4SLinus Torvalds return -ENOMEM; /* We'll just try again later. */ 1191ef5cb973SHerbert Xu 11923ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 11933ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1194b60b49eaSHerbert Xu nlen = skb->len - len - nsize; 1195b60b49eaSHerbert Xu buff->truesize += nlen; 1196b60b49eaSHerbert Xu skb->truesize -= nlen; 11971da177e4SLinus Torvalds 11981da177e4SLinus Torvalds /* Correct the sequence numbers. */ 11991da177e4SLinus Torvalds TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 12001da177e4SLinus Torvalds TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 12011da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 12021da177e4SLinus Torvalds 12031da177e4SLinus Torvalds /* PSH and FIN should only be set in the second packet. */ 12044de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 12054de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 12064de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1207e14c3cafSHerbert Xu TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; 1208a166140eSMartin KaFai Lau tcp_skb_fragment_eor(skb, buff); 12091da177e4SLinus Torvalds 121084fa7933SPatrick McHardy if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { 12111da177e4SLinus Torvalds /* Copy and checksum data tail into the new buffer. */ 1212056834d9SIlpo Järvinen buff->csum = csum_partial_copy_nocheck(skb->data + len, 1213056834d9SIlpo Järvinen skb_put(buff, nsize), 12141da177e4SLinus Torvalds nsize, 0); 12151da177e4SLinus Torvalds 12161da177e4SLinus Torvalds skb_trim(skb, len); 12171da177e4SLinus Torvalds 12181da177e4SLinus Torvalds skb->csum = csum_block_sub(skb->csum, buff->csum, len); 12191da177e4SLinus Torvalds } else { 122084fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 12211da177e4SLinus Torvalds skb_split(skb, buff, len); 12221da177e4SLinus Torvalds } 12231da177e4SLinus Torvalds 12241da177e4SLinus Torvalds buff->ip_summed = skb->ip_summed; 12251da177e4SLinus Torvalds 1226a61bbcf2SPatrick McHardy buff->tstamp = skb->tstamp; 1227490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 12281da177e4SLinus Torvalds 12296475be16SDavid S. Miller old_factor = tcp_skb_pcount(skb); 12306475be16SDavid S. Miller 12311da177e4SLinus Torvalds /* Fix up tso_factor for both original and new SKB. */ 12325bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 12335bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 12341da177e4SLinus Torvalds 1235b9f64820SYuchung Cheng /* Update delivered info for the new segment */ 1236b9f64820SYuchung Cheng TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx; 1237b9f64820SYuchung Cheng 12386475be16SDavid S. Miller /* If this packet has been sent out already, we must 12396475be16SDavid S. Miller * adjust the various packet counters. 12406475be16SDavid S. Miller */ 1241cf0b450cSHerbert Xu if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 12426475be16SDavid S. Miller int diff = old_factor - tcp_skb_pcount(skb) - 12436475be16SDavid S. Miller tcp_skb_pcount(buff); 12441da177e4SLinus Torvalds 1245797108d1SIlpo Järvinen if (diff) 1246797108d1SIlpo Järvinen tcp_adjust_pcount(sk, skb, diff); 12471da177e4SLinus Torvalds } 12481da177e4SLinus Torvalds 12491da177e4SLinus Torvalds /* Link BUFF into the send queue. */ 1250f4a775d1SEric Dumazet __skb_header_release(buff); 1251fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 12521da177e4SLinus Torvalds 12531da177e4SLinus Torvalds return 0; 12541da177e4SLinus Torvalds } 12551da177e4SLinus Torvalds 12561da177e4SLinus Torvalds /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c 12571da177e4SLinus Torvalds * eventually). The difference is that pulled data not copied, but 12581da177e4SLinus Torvalds * immediately discarded. 12591da177e4SLinus Torvalds */ 1260f2911969SHerbert Xu ~{PmVHI~} static void __pskb_trim_head(struct sk_buff *skb, int len) 12611da177e4SLinus Torvalds { 12627b7fc97aSEric Dumazet struct skb_shared_info *shinfo; 12631da177e4SLinus Torvalds int i, k, eat; 12641da177e4SLinus Torvalds 12654fa48bf3SEric Dumazet eat = min_t(int, len, skb_headlen(skb)); 12664fa48bf3SEric Dumazet if (eat) { 12674fa48bf3SEric Dumazet __skb_pull(skb, eat); 12684fa48bf3SEric Dumazet len -= eat; 12694fa48bf3SEric Dumazet if (!len) 12704fa48bf3SEric Dumazet return; 12714fa48bf3SEric Dumazet } 12721da177e4SLinus Torvalds eat = len; 12731da177e4SLinus Torvalds k = 0; 12747b7fc97aSEric Dumazet shinfo = skb_shinfo(skb); 12757b7fc97aSEric Dumazet for (i = 0; i < shinfo->nr_frags; i++) { 12767b7fc97aSEric Dumazet int size = skb_frag_size(&shinfo->frags[i]); 12779e903e08SEric Dumazet 12789e903e08SEric Dumazet if (size <= eat) { 1279aff65da0SIan Campbell skb_frag_unref(skb, i); 12809e903e08SEric Dumazet eat -= size; 12811da177e4SLinus Torvalds } else { 12827b7fc97aSEric Dumazet shinfo->frags[k] = shinfo->frags[i]; 12831da177e4SLinus Torvalds if (eat) { 12847b7fc97aSEric Dumazet shinfo->frags[k].page_offset += eat; 12857b7fc97aSEric Dumazet skb_frag_size_sub(&shinfo->frags[k], eat); 12861da177e4SLinus Torvalds eat = 0; 12871da177e4SLinus Torvalds } 12881da177e4SLinus Torvalds k++; 12891da177e4SLinus Torvalds } 12901da177e4SLinus Torvalds } 12917b7fc97aSEric Dumazet shinfo->nr_frags = k; 12921da177e4SLinus Torvalds 129327a884dcSArnaldo Carvalho de Melo skb_reset_tail_pointer(skb); 12941da177e4SLinus Torvalds skb->data_len -= len; 12951da177e4SLinus Torvalds skb->len = skb->data_len; 12961da177e4SLinus Torvalds } 12971da177e4SLinus Torvalds 129867edfef7SAndi Kleen /* Remove acked data from a packet in the transmit queue. */ 12991da177e4SLinus Torvalds int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 13001da177e4SLinus Torvalds { 130114bbd6a5SPravin B Shelar if (skb_unclone(skb, GFP_ATOMIC)) 13021da177e4SLinus Torvalds return -ENOMEM; 13031da177e4SLinus Torvalds 13044fa48bf3SEric Dumazet __pskb_trim_head(skb, len); 13051da177e4SLinus Torvalds 13061da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq += len; 130784fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 13081da177e4SLinus Torvalds 13091da177e4SLinus Torvalds skb->truesize -= len; 13101da177e4SLinus Torvalds sk->sk_wmem_queued -= len; 13113ab224beSHideo Aoki sk_mem_uncharge(sk, len); 13121da177e4SLinus Torvalds sock_set_flag(sk, SOCK_QUEUE_SHRUNK); 13131da177e4SLinus Torvalds 13145b35e1e6SNeal Cardwell /* Any change of skb->len requires recalculation of tso factor. */ 13151da177e4SLinus Torvalds if (tcp_skb_pcount(skb) > 1) 13165bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb)); 13171da177e4SLinus Torvalds 13181da177e4SLinus Torvalds return 0; 13191da177e4SLinus Torvalds } 13201da177e4SLinus Torvalds 13211b63edd6SYuchung Cheng /* Calculate MSS not accounting any TCP options. */ 13221b63edd6SYuchung Cheng static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu) 13235d424d5aSJohn Heffner { 1324cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1325cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 13265d424d5aSJohn Heffner int mss_now; 13275d424d5aSJohn Heffner 13285d424d5aSJohn Heffner /* Calculate base mss without TCP options: 13295d424d5aSJohn Heffner It is MMS_S - sizeof(tcphdr) of rfc1122 13305d424d5aSJohn Heffner */ 13315d424d5aSJohn Heffner mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 13325d424d5aSJohn Heffner 133367469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 133467469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 133567469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 133667469601SEric Dumazet 133767469601SEric Dumazet if (dst && dst_allfrag(dst)) 133867469601SEric Dumazet mss_now -= icsk->icsk_af_ops->net_frag_header_len; 133967469601SEric Dumazet } 134067469601SEric Dumazet 13415d424d5aSJohn Heffner /* Clamp it (mss_clamp does not include tcp options) */ 13425d424d5aSJohn Heffner if (mss_now > tp->rx_opt.mss_clamp) 13435d424d5aSJohn Heffner mss_now = tp->rx_opt.mss_clamp; 13445d424d5aSJohn Heffner 13455d424d5aSJohn Heffner /* Now subtract optional transport overhead */ 13465d424d5aSJohn Heffner mss_now -= icsk->icsk_ext_hdr_len; 13475d424d5aSJohn Heffner 13485d424d5aSJohn Heffner /* Then reserve room for full set of TCP options and 8 bytes of data */ 13495d424d5aSJohn Heffner if (mss_now < 48) 13505d424d5aSJohn Heffner mss_now = 48; 13515d424d5aSJohn Heffner return mss_now; 13525d424d5aSJohn Heffner } 13535d424d5aSJohn Heffner 13541b63edd6SYuchung Cheng /* Calculate MSS. Not accounting for SACKs here. */ 13551b63edd6SYuchung Cheng int tcp_mtu_to_mss(struct sock *sk, int pmtu) 13561b63edd6SYuchung Cheng { 13571b63edd6SYuchung Cheng /* Subtract TCP options size, not including SACKs */ 13581b63edd6SYuchung Cheng return __tcp_mtu_to_mss(sk, pmtu) - 13591b63edd6SYuchung Cheng (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr)); 13601b63edd6SYuchung Cheng } 13611b63edd6SYuchung Cheng 13625d424d5aSJohn Heffner /* Inverse of above */ 136367469601SEric Dumazet int tcp_mss_to_mtu(struct sock *sk, int mss) 13645d424d5aSJohn Heffner { 1365cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1366cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 13675d424d5aSJohn Heffner int mtu; 13685d424d5aSJohn Heffner 13695d424d5aSJohn Heffner mtu = mss + 13705d424d5aSJohn Heffner tp->tcp_header_len + 13715d424d5aSJohn Heffner icsk->icsk_ext_hdr_len + 13725d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 13735d424d5aSJohn Heffner 137467469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 137567469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 137667469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 137767469601SEric Dumazet 137867469601SEric Dumazet if (dst && dst_allfrag(dst)) 137967469601SEric Dumazet mtu += icsk->icsk_af_ops->net_frag_header_len; 138067469601SEric Dumazet } 13815d424d5aSJohn Heffner return mtu; 13825d424d5aSJohn Heffner } 1383556c6b46SNeal Cardwell EXPORT_SYMBOL(tcp_mss_to_mtu); 13845d424d5aSJohn Heffner 138567edfef7SAndi Kleen /* MTU probing init per socket */ 13865d424d5aSJohn Heffner void tcp_mtup_init(struct sock *sk) 13875d424d5aSJohn Heffner { 13885d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 13895d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 1390b0f9ca53SFan Du struct net *net = sock_net(sk); 13915d424d5aSJohn Heffner 1392b0f9ca53SFan Du icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1; 13935d424d5aSJohn Heffner icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 13945d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 1395b0f9ca53SFan Du icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss); 13965d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 139705cbc0dbSFan Du if (icsk->icsk_mtup.enabled) 139805cbc0dbSFan Du icsk->icsk_mtup.probe_timestamp = tcp_time_stamp; 13995d424d5aSJohn Heffner } 14004bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_mtup_init); 14015d424d5aSJohn Heffner 14021da177e4SLinus Torvalds /* This function synchronize snd mss to current pmtu/exthdr set. 14031da177e4SLinus Torvalds 14041da177e4SLinus Torvalds tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 14051da177e4SLinus Torvalds for TCP options, but includes only bare TCP header. 14061da177e4SLinus Torvalds 14071da177e4SLinus Torvalds tp->rx_opt.mss_clamp is mss negotiated at connection setup. 1408caa20d9aSStephen Hemminger It is minimum of user_mss and mss received with SYN. 14091da177e4SLinus Torvalds It also does not include TCP options. 14101da177e4SLinus Torvalds 1411d83d8461SArnaldo Carvalho de Melo inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. 14121da177e4SLinus Torvalds 14131da177e4SLinus Torvalds tp->mss_cache is current effective sending mss, including 14141da177e4SLinus Torvalds all tcp options except for SACKs. It is evaluated, 14151da177e4SLinus Torvalds taking into account current pmtu, but never exceeds 14161da177e4SLinus Torvalds tp->rx_opt.mss_clamp. 14171da177e4SLinus Torvalds 14181da177e4SLinus Torvalds NOTE1. rfc1122 clearly states that advertised MSS 14191da177e4SLinus Torvalds DOES NOT include either tcp or ip options. 14201da177e4SLinus Torvalds 1421d83d8461SArnaldo Carvalho de Melo NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache 1422d83d8461SArnaldo Carvalho de Melo are READ ONLY outside this function. --ANK (980731) 14231da177e4SLinus Torvalds */ 14241da177e4SLinus Torvalds unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 14251da177e4SLinus Torvalds { 14261da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1427d83d8461SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 14285d424d5aSJohn Heffner int mss_now; 14291da177e4SLinus Torvalds 14305d424d5aSJohn Heffner if (icsk->icsk_mtup.search_high > pmtu) 14315d424d5aSJohn Heffner icsk->icsk_mtup.search_high = pmtu; 14321da177e4SLinus Torvalds 14335d424d5aSJohn Heffner mss_now = tcp_mtu_to_mss(sk, pmtu); 1434409d22b4SIlpo Järvinen mss_now = tcp_bound_to_half_wnd(tp, mss_now); 14351da177e4SLinus Torvalds 14361da177e4SLinus Torvalds /* And store cached results */ 1437d83d8461SArnaldo Carvalho de Melo icsk->icsk_pmtu_cookie = pmtu; 14385d424d5aSJohn Heffner if (icsk->icsk_mtup.enabled) 14395d424d5aSJohn Heffner mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); 1440c1b4a7e6SDavid S. Miller tp->mss_cache = mss_now; 14411da177e4SLinus Torvalds 14421da177e4SLinus Torvalds return mss_now; 14431da177e4SLinus Torvalds } 14444bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_sync_mss); 14451da177e4SLinus Torvalds 14461da177e4SLinus Torvalds /* Compute the current effective MSS, taking SACKs and IP options, 14471da177e4SLinus Torvalds * and even PMTU discovery events into account. 14481da177e4SLinus Torvalds */ 14490c54b85fSIlpo Järvinen unsigned int tcp_current_mss(struct sock *sk) 14501da177e4SLinus Torvalds { 1451cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1452cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1453c1b4a7e6SDavid S. Miller u32 mss_now; 145495c96174SEric Dumazet unsigned int header_len; 145533ad798cSAdam Langley struct tcp_out_options opts; 145633ad798cSAdam Langley struct tcp_md5sig_key *md5; 14571da177e4SLinus Torvalds 1458c1b4a7e6SDavid S. Miller mss_now = tp->mss_cache; 1459c1b4a7e6SDavid S. Miller 14601da177e4SLinus Torvalds if (dst) { 14611da177e4SLinus Torvalds u32 mtu = dst_mtu(dst); 1462d83d8461SArnaldo Carvalho de Melo if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 14631da177e4SLinus Torvalds mss_now = tcp_sync_mss(sk, mtu); 14641da177e4SLinus Torvalds } 14651da177e4SLinus Torvalds 146633ad798cSAdam Langley header_len = tcp_established_options(sk, NULL, &opts, &md5) + 146733ad798cSAdam Langley sizeof(struct tcphdr); 146833ad798cSAdam Langley /* The mss_cache is sized based on tp->tcp_header_len, which assumes 146933ad798cSAdam Langley * some common options. If this is an odd packet (because we have SACK 147033ad798cSAdam Langley * blocks etc) then our calculated header_len will be different, and 147133ad798cSAdam Langley * we have to adjust mss_now correspondingly */ 147233ad798cSAdam Langley if (header_len != tp->tcp_header_len) { 147333ad798cSAdam Langley int delta = (int) header_len - tp->tcp_header_len; 147433ad798cSAdam Langley mss_now -= delta; 147533ad798cSAdam Langley } 1476cfb6eeb4SYOSHIFUJI Hideaki 14771da177e4SLinus Torvalds return mss_now; 14781da177e4SLinus Torvalds } 14791da177e4SLinus Torvalds 148086fd14adSWeiping Pan /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. 148186fd14adSWeiping Pan * As additional protections, we do not touch cwnd in retransmission phases, 148286fd14adSWeiping Pan * and if application hit its sndbuf limit recently. 148386fd14adSWeiping Pan */ 148486fd14adSWeiping Pan static void tcp_cwnd_application_limited(struct sock *sk) 1485a762a980SDavid S. Miller { 14869e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1487a762a980SDavid S. Miller 148886fd14adSWeiping Pan if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && 148986fd14adSWeiping Pan sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 149086fd14adSWeiping Pan /* Limited by application or receiver window. */ 149186fd14adSWeiping Pan u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); 149286fd14adSWeiping Pan u32 win_used = max(tp->snd_cwnd_used, init_win); 149386fd14adSWeiping Pan if (win_used < tp->snd_cwnd) { 149486fd14adSWeiping Pan tp->snd_ssthresh = tcp_current_ssthresh(sk); 149586fd14adSWeiping Pan tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1; 149686fd14adSWeiping Pan } 149786fd14adSWeiping Pan tp->snd_cwnd_used = 0; 149886fd14adSWeiping Pan } 149986fd14adSWeiping Pan tp->snd_cwnd_stamp = tcp_time_stamp; 150086fd14adSWeiping Pan } 150186fd14adSWeiping Pan 1502ca8a2263SNeal Cardwell static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited) 1503a762a980SDavid S. Miller { 1504a762a980SDavid S. Miller struct tcp_sock *tp = tcp_sk(sk); 1505a762a980SDavid S. Miller 1506ca8a2263SNeal Cardwell /* Track the maximum number of outstanding packets in each 1507ca8a2263SNeal Cardwell * window, and remember whether we were cwnd-limited then. 1508ca8a2263SNeal Cardwell */ 1509ca8a2263SNeal Cardwell if (!before(tp->snd_una, tp->max_packets_seq) || 1510ca8a2263SNeal Cardwell tp->packets_out > tp->max_packets_out) { 1511ca8a2263SNeal Cardwell tp->max_packets_out = tp->packets_out; 1512ca8a2263SNeal Cardwell tp->max_packets_seq = tp->snd_nxt; 1513ca8a2263SNeal Cardwell tp->is_cwnd_limited = is_cwnd_limited; 1514ca8a2263SNeal Cardwell } 1515e114a710SEric Dumazet 151624901551SEric Dumazet if (tcp_is_cwnd_limited(sk)) { 1517a762a980SDavid S. Miller /* Network is feed fully. */ 1518a762a980SDavid S. Miller tp->snd_cwnd_used = 0; 1519a762a980SDavid S. Miller tp->snd_cwnd_stamp = tcp_time_stamp; 1520a762a980SDavid S. Miller } else { 1521a762a980SDavid S. Miller /* Network starves. */ 1522a762a980SDavid S. Miller if (tp->packets_out > tp->snd_cwnd_used) 1523a762a980SDavid S. Miller tp->snd_cwnd_used = tp->packets_out; 1524a762a980SDavid S. Miller 152515d33c07SDavid S. Miller if (sysctl_tcp_slow_start_after_idle && 152615d33c07SDavid S. Miller (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto) 1527a762a980SDavid S. Miller tcp_cwnd_application_limited(sk); 1528b0f71bd3SFrancis Yan 1529b0f71bd3SFrancis Yan /* The following conditions together indicate the starvation 1530b0f71bd3SFrancis Yan * is caused by insufficient sender buffer: 1531b0f71bd3SFrancis Yan * 1) just sent some data (see tcp_write_xmit) 1532b0f71bd3SFrancis Yan * 2) not cwnd limited (this else condition) 1533b0f71bd3SFrancis Yan * 3) no more data to send (null tcp_send_head ) 1534b0f71bd3SFrancis Yan * 4) application is hitting buffer limit (SOCK_NOSPACE) 1535b0f71bd3SFrancis Yan */ 1536b0f71bd3SFrancis Yan if (!tcp_send_head(sk) && sk->sk_socket && 1537b0f71bd3SFrancis Yan test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) && 1538b0f71bd3SFrancis Yan (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) 1539b0f71bd3SFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED); 1540a762a980SDavid S. Miller } 1541a762a980SDavid S. Miller } 1542a762a980SDavid S. Miller 1543d4589926SEric Dumazet /* Minshall's variant of the Nagle send check. */ 1544d4589926SEric Dumazet static bool tcp_minshall_check(const struct tcp_sock *tp) 1545d4589926SEric Dumazet { 1546d4589926SEric Dumazet return after(tp->snd_sml, tp->snd_una) && 1547d4589926SEric Dumazet !after(tp->snd_sml, tp->snd_nxt); 1548d4589926SEric Dumazet } 1549d4589926SEric Dumazet 1550d4589926SEric Dumazet /* Update snd_sml if this skb is under mss 1551d4589926SEric Dumazet * Note that a TSO packet might end with a sub-mss segment 1552d4589926SEric Dumazet * The test is really : 1553d4589926SEric Dumazet * if ((skb->len % mss) != 0) 1554d4589926SEric Dumazet * tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1555d4589926SEric Dumazet * But we can avoid doing the divide again given we already have 1556d4589926SEric Dumazet * skb_pcount = skb->len / mss_now 15570e3a4803SIlpo Järvinen */ 1558d4589926SEric Dumazet static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now, 1559d4589926SEric Dumazet const struct sk_buff *skb) 1560d4589926SEric Dumazet { 1561d4589926SEric Dumazet if (skb->len < tcp_skb_pcount(skb) * mss_now) 1562d4589926SEric Dumazet tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1563d4589926SEric Dumazet } 1564d4589926SEric Dumazet 1565d4589926SEric Dumazet /* Return false, if packet can be sent now without violation Nagle's rules: 1566d4589926SEric Dumazet * 1. It is full sized. (provided by caller in %partial bool) 1567d4589926SEric Dumazet * 2. Or it contains FIN. (already checked by caller) 1568d4589926SEric Dumazet * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. 1569d4589926SEric Dumazet * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 1570d4589926SEric Dumazet * With Minshall's modification: all sent small packets are ACKed. 1571d4589926SEric Dumazet */ 1572d4589926SEric Dumazet static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp, 1573cc93fc51SPeter Pan(潘卫平) int nonagle) 1574d4589926SEric Dumazet { 1575d4589926SEric Dumazet return partial && 1576d4589926SEric Dumazet ((nonagle & TCP_NAGLE_CORK) || 1577d4589926SEric Dumazet (!nonagle && tp->packets_out && tcp_minshall_check(tp))); 1578d4589926SEric Dumazet } 1579605ad7f1SEric Dumazet 1580605ad7f1SEric Dumazet /* Return how many segs we'd like on a TSO packet, 1581605ad7f1SEric Dumazet * to send one TSO packet per ms 1582605ad7f1SEric Dumazet */ 15831b3878caSNeal Cardwell u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now, 15841b3878caSNeal Cardwell int min_tso_segs) 1585605ad7f1SEric Dumazet { 1586605ad7f1SEric Dumazet u32 bytes, segs; 1587605ad7f1SEric Dumazet 1588605ad7f1SEric Dumazet bytes = min(sk->sk_pacing_rate >> 10, 1589605ad7f1SEric Dumazet sk->sk_gso_max_size - 1 - MAX_TCP_HEADER); 1590605ad7f1SEric Dumazet 1591605ad7f1SEric Dumazet /* Goal is to send at least one packet per ms, 1592605ad7f1SEric Dumazet * not one big TSO packet every 100 ms. 1593605ad7f1SEric Dumazet * This preserves ACK clocking and is consistent 1594605ad7f1SEric Dumazet * with tcp_tso_should_defer() heuristic. 1595605ad7f1SEric Dumazet */ 15961b3878caSNeal Cardwell segs = max_t(u32, bytes / mss_now, min_tso_segs); 1597605ad7f1SEric Dumazet 1598605ad7f1SEric Dumazet return min_t(u32, segs, sk->sk_gso_max_segs); 1599605ad7f1SEric Dumazet } 16001b3878caSNeal Cardwell EXPORT_SYMBOL(tcp_tso_autosize); 1601605ad7f1SEric Dumazet 1602ed6e7268SNeal Cardwell /* Return the number of segments we want in the skb we are transmitting. 1603ed6e7268SNeal Cardwell * See if congestion control module wants to decide; otherwise, autosize. 1604ed6e7268SNeal Cardwell */ 1605ed6e7268SNeal Cardwell static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now) 1606ed6e7268SNeal Cardwell { 1607ed6e7268SNeal Cardwell const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 1608ed6e7268SNeal Cardwell u32 tso_segs = ca_ops->tso_segs_goal ? ca_ops->tso_segs_goal(sk) : 0; 1609ed6e7268SNeal Cardwell 16101b3878caSNeal Cardwell return tso_segs ? : 16111b3878caSNeal Cardwell tcp_tso_autosize(sk, mss_now, sysctl_tcp_min_tso_segs); 1612ed6e7268SNeal Cardwell } 1613ed6e7268SNeal Cardwell 1614d4589926SEric Dumazet /* Returns the portion of skb which can be sent right away */ 1615d4589926SEric Dumazet static unsigned int tcp_mss_split_point(const struct sock *sk, 1616d4589926SEric Dumazet const struct sk_buff *skb, 1617d4589926SEric Dumazet unsigned int mss_now, 1618d4589926SEric Dumazet unsigned int max_segs, 1619d4589926SEric Dumazet int nonagle) 1620c1b4a7e6SDavid S. Miller { 1621cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1622d4589926SEric Dumazet u32 partial, needed, window, max_len; 1623c1b4a7e6SDavid S. Miller 162490840defSIlpo Järvinen window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 16251485348dSBen Hutchings max_len = mss_now * max_segs; 16260e3a4803SIlpo Järvinen 16271485348dSBen Hutchings if (likely(max_len <= window && skb != tcp_write_queue_tail(sk))) 16281485348dSBen Hutchings return max_len; 16290e3a4803SIlpo Järvinen 16305ea3a748SIlpo Järvinen needed = min(skb->len, window); 16315ea3a748SIlpo Järvinen 16321485348dSBen Hutchings if (max_len <= needed) 16331485348dSBen Hutchings return max_len; 16340e3a4803SIlpo Järvinen 1635d4589926SEric Dumazet partial = needed % mss_now; 1636d4589926SEric Dumazet /* If last segment is not a full MSS, check if Nagle rules allow us 1637d4589926SEric Dumazet * to include this last segment in this skb. 1638d4589926SEric Dumazet * Otherwise, we'll split the skb at last MSS boundary 1639d4589926SEric Dumazet */ 1640cc93fc51SPeter Pan(潘卫平) if (tcp_nagle_check(partial != 0, tp, nonagle)) 1641d4589926SEric Dumazet return needed - partial; 1642d4589926SEric Dumazet 1643d4589926SEric Dumazet return needed; 1644c1b4a7e6SDavid S. Miller } 1645c1b4a7e6SDavid S. Miller 1646c1b4a7e6SDavid S. Miller /* Can at least one segment of SKB be sent right now, according to the 1647c1b4a7e6SDavid S. Miller * congestion window rules? If so, return how many segments are allowed. 1648c1b4a7e6SDavid S. Miller */ 1649cf533ea5SEric Dumazet static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, 1650cf533ea5SEric Dumazet const struct sk_buff *skb) 1651c1b4a7e6SDavid S. Miller { 1652d649a7a8SEric Dumazet u32 in_flight, cwnd, halfcwnd; 1653c1b4a7e6SDavid S. Miller 1654c1b4a7e6SDavid S. Miller /* Don't be strict about the congestion window for the final FIN. */ 16554de075e0SEric Dumazet if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && 16564de075e0SEric Dumazet tcp_skb_pcount(skb) == 1) 1657c1b4a7e6SDavid S. Miller return 1; 1658c1b4a7e6SDavid S. Miller 1659c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1660c1b4a7e6SDavid S. Miller cwnd = tp->snd_cwnd; 1661d649a7a8SEric Dumazet if (in_flight >= cwnd) 1662c1b4a7e6SDavid S. Miller return 0; 1663d649a7a8SEric Dumazet 1664d649a7a8SEric Dumazet /* For better scheduling, ensure we have at least 1665d649a7a8SEric Dumazet * 2 GSO packets in flight. 1666d649a7a8SEric Dumazet */ 1667d649a7a8SEric Dumazet halfcwnd = max(cwnd >> 1, 1U); 1668d649a7a8SEric Dumazet return min(halfcwnd, cwnd - in_flight); 1669c1b4a7e6SDavid S. Miller } 1670c1b4a7e6SDavid S. Miller 1671b595076aSUwe Kleine-König /* Initialize TSO state of a skb. 167267edfef7SAndi Kleen * This must be invoked the first time we consider transmitting 1673c1b4a7e6SDavid S. Miller * SKB onto the wire. 1674c1b4a7e6SDavid S. Miller */ 16755bbb432cSEric Dumazet static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now) 1676c1b4a7e6SDavid S. Miller { 1677c1b4a7e6SDavid S. Miller int tso_segs = tcp_skb_pcount(skb); 1678c1b4a7e6SDavid S. Miller 1679f8269a49SIlpo Järvinen if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { 16805bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 1681c1b4a7e6SDavid S. Miller tso_segs = tcp_skb_pcount(skb); 1682c1b4a7e6SDavid S. Miller } 1683c1b4a7e6SDavid S. Miller return tso_segs; 1684c1b4a7e6SDavid S. Miller } 1685c1b4a7e6SDavid S. Miller 1686c1b4a7e6SDavid S. Miller 1687a2a385d6SEric Dumazet /* Return true if the Nagle test allows this packet to be 1688c1b4a7e6SDavid S. Miller * sent now. 1689c1b4a7e6SDavid S. Miller */ 1690a2a385d6SEric Dumazet static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, 1691c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1692c1b4a7e6SDavid S. Miller { 1693c1b4a7e6SDavid S. Miller /* Nagle rule does not apply to frames, which sit in the middle of the 1694c1b4a7e6SDavid S. Miller * write_queue (they have no chances to get new data). 1695c1b4a7e6SDavid S. Miller * 1696c1b4a7e6SDavid S. Miller * This is implemented in the callers, where they modify the 'nonagle' 1697c1b4a7e6SDavid S. Miller * argument based upon the location of SKB in the send queue. 1698c1b4a7e6SDavid S. Miller */ 1699c1b4a7e6SDavid S. Miller if (nonagle & TCP_NAGLE_PUSH) 1700a2a385d6SEric Dumazet return true; 1701c1b4a7e6SDavid S. Miller 17029b44190dSYuchung Cheng /* Don't use the nagle rule for urgent data (or for the final FIN). */ 17039b44190dSYuchung Cheng if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) 1704a2a385d6SEric Dumazet return true; 1705c1b4a7e6SDavid S. Miller 1706cc93fc51SPeter Pan(潘卫平) if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle)) 1707a2a385d6SEric Dumazet return true; 1708c1b4a7e6SDavid S. Miller 1709a2a385d6SEric Dumazet return false; 1710c1b4a7e6SDavid S. Miller } 1711c1b4a7e6SDavid S. Miller 1712c1b4a7e6SDavid S. Miller /* Does at least the first segment of SKB fit into the send window? */ 1713a2a385d6SEric Dumazet static bool tcp_snd_wnd_test(const struct tcp_sock *tp, 1714a2a385d6SEric Dumazet const struct sk_buff *skb, 1715056834d9SIlpo Järvinen unsigned int cur_mss) 1716c1b4a7e6SDavid S. Miller { 1717c1b4a7e6SDavid S. Miller u32 end_seq = TCP_SKB_CB(skb)->end_seq; 1718c1b4a7e6SDavid S. Miller 1719c1b4a7e6SDavid S. Miller if (skb->len > cur_mss) 1720c1b4a7e6SDavid S. Miller end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 1721c1b4a7e6SDavid S. Miller 172290840defSIlpo Järvinen return !after(end_seq, tcp_wnd_end(tp)); 1723c1b4a7e6SDavid S. Miller } 1724c1b4a7e6SDavid S. Miller 1725fe067e8aSDavid S. Miller /* This checks if the data bearing packet SKB (usually tcp_send_head(sk)) 1726c1b4a7e6SDavid S. Miller * should be put on the wire right now. If so, it returns the number of 1727c1b4a7e6SDavid S. Miller * packets allowed by the congestion window. 1728c1b4a7e6SDavid S. Miller */ 1729cf533ea5SEric Dumazet static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb, 1730c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1731c1b4a7e6SDavid S. Miller { 1732cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1733c1b4a7e6SDavid S. Miller unsigned int cwnd_quota; 1734c1b4a7e6SDavid S. Miller 17355bbb432cSEric Dumazet tcp_init_tso_segs(skb, cur_mss); 1736c1b4a7e6SDavid S. Miller 1737c1b4a7e6SDavid S. Miller if (!tcp_nagle_test(tp, skb, cur_mss, nonagle)) 1738c1b4a7e6SDavid S. Miller return 0; 1739c1b4a7e6SDavid S. Miller 1740c1b4a7e6SDavid S. Miller cwnd_quota = tcp_cwnd_test(tp, skb); 1741056834d9SIlpo Järvinen if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss)) 1742c1b4a7e6SDavid S. Miller cwnd_quota = 0; 1743c1b4a7e6SDavid S. Miller 1744c1b4a7e6SDavid S. Miller return cwnd_quota; 1745c1b4a7e6SDavid S. Miller } 1746c1b4a7e6SDavid S. Miller 174767edfef7SAndi Kleen /* Test if sending is allowed right now. */ 1748a2a385d6SEric Dumazet bool tcp_may_send_now(struct sock *sk) 1749c1b4a7e6SDavid S. Miller { 1750cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1751fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 1752c1b4a7e6SDavid S. Miller 1753a02cec21SEric Dumazet return skb && 17540c54b85fSIlpo Järvinen tcp_snd_test(sk, skb, tcp_current_mss(sk), 1755c1b4a7e6SDavid S. Miller (tcp_skb_is_last(sk, skb) ? 1756a02cec21SEric Dumazet tp->nonagle : TCP_NAGLE_PUSH)); 1757c1b4a7e6SDavid S. Miller } 1758c1b4a7e6SDavid S. Miller 1759c1b4a7e6SDavid S. Miller /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 1760c1b4a7e6SDavid S. Miller * which is put after SKB on the list. It is very much like 1761c1b4a7e6SDavid S. Miller * tcp_fragment() except that it may make several kinds of assumptions 1762c1b4a7e6SDavid S. Miller * in order to speed up the splitting operation. In particular, we 1763c1b4a7e6SDavid S. Miller * know that all the data is in scatter-gather pages, and that the 1764c1b4a7e6SDavid S. Miller * packet has never been sent out before (and thus is not cloned). 1765c1b4a7e6SDavid S. Miller */ 1766056834d9SIlpo Järvinen static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, 1767c4ead4c5SEric Dumazet unsigned int mss_now, gfp_t gfp) 1768c1b4a7e6SDavid S. Miller { 1769c1b4a7e6SDavid S. Miller struct sk_buff *buff; 1770c1b4a7e6SDavid S. Miller int nlen = skb->len - len; 17719ce01461SIlpo Järvinen u8 flags; 1772c1b4a7e6SDavid S. Miller 1773c1b4a7e6SDavid S. Miller /* All of a TSO frame must be composed of paged data. */ 1774c8ac3774SHerbert Xu if (skb->len != skb->data_len) 17756cc55e09SOctavian Purdila return tcp_fragment(sk, skb, len, mss_now, gfp); 1776c1b4a7e6SDavid S. Miller 1777eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, gfp, true); 177851456b29SIan Morris if (unlikely(!buff)) 1779c1b4a7e6SDavid S. Miller return -ENOMEM; 1780c1b4a7e6SDavid S. Miller 17813ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 17823ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1783b60b49eaSHerbert Xu buff->truesize += nlen; 1784c1b4a7e6SDavid S. Miller skb->truesize -= nlen; 1785c1b4a7e6SDavid S. Miller 1786c1b4a7e6SDavid S. Miller /* Correct the sequence numbers. */ 1787c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 1788c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 1789c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 1790c1b4a7e6SDavid S. Miller 1791c1b4a7e6SDavid S. Miller /* PSH and FIN should only be set in the second packet. */ 17924de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 17934de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 17944de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1795c1b4a7e6SDavid S. Miller 1796c1b4a7e6SDavid S. Miller /* This packet was never sent out yet, so no SACK bits. */ 1797c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->sacked = 0; 1798c1b4a7e6SDavid S. Miller 1799a166140eSMartin KaFai Lau tcp_skb_fragment_eor(skb, buff); 1800a166140eSMartin KaFai Lau 180184fa7933SPatrick McHardy buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; 1802c1b4a7e6SDavid S. Miller skb_split(skb, buff, len); 1803490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 1804c1b4a7e6SDavid S. Miller 1805c1b4a7e6SDavid S. Miller /* Fix up tso_factor for both original and new SKB. */ 18065bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 18075bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 1808c1b4a7e6SDavid S. Miller 1809c1b4a7e6SDavid S. Miller /* Link BUFF into the send queue. */ 1810f4a775d1SEric Dumazet __skb_header_release(buff); 1811fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 1812c1b4a7e6SDavid S. Miller 1813c1b4a7e6SDavid S. Miller return 0; 1814c1b4a7e6SDavid S. Miller } 1815c1b4a7e6SDavid S. Miller 1816c1b4a7e6SDavid S. Miller /* Try to defer sending, if possible, in order to minimize the amount 1817c1b4a7e6SDavid S. Miller * of TSO splitting we do. View it as a kind of TSO Nagle test. 1818c1b4a7e6SDavid S. Miller * 1819c1b4a7e6SDavid S. Miller * This algorithm is from John Heffner. 1820c1b4a7e6SDavid S. Miller */ 1821ca8a2263SNeal Cardwell static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb, 1822605ad7f1SEric Dumazet bool *is_cwnd_limited, u32 max_segs) 1823c1b4a7e6SDavid S. Miller { 18246687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 182550c8339eSEric Dumazet u32 age, send_win, cong_win, limit, in_flight; 182650c8339eSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 182750c8339eSEric Dumazet struct skb_mstamp now; 182850c8339eSEric Dumazet struct sk_buff *head; 1829ad9f4f50SEric Dumazet int win_divisor; 1830c1b4a7e6SDavid S. Miller 18314de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 1832ae8064acSJohn Heffner goto send_now; 1833c1b4a7e6SDavid S. Miller 183499d7662aSEric Dumazet if (icsk->icsk_ca_state >= TCP_CA_Recovery) 1835ae8064acSJohn Heffner goto send_now; 1836ae8064acSJohn Heffner 18375f852eb5SEric Dumazet /* Avoid bursty behavior by allowing defer 18385f852eb5SEric Dumazet * only if the last write was recent. 18395f852eb5SEric Dumazet */ 18405f852eb5SEric Dumazet if ((s32)(tcp_time_stamp - tp->lsndtime) > 0) 1841ae8064acSJohn Heffner goto send_now; 1842908a75c1SDavid S. Miller 1843c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1844c1b4a7e6SDavid S. Miller 1845056834d9SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight)); 1846c1b4a7e6SDavid S. Miller 184790840defSIlpo Järvinen send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1848c1b4a7e6SDavid S. Miller 1849c1b4a7e6SDavid S. Miller /* From in_flight test above, we know that cwnd > in_flight. */ 1850c1b4a7e6SDavid S. Miller cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; 1851c1b4a7e6SDavid S. Miller 1852c1b4a7e6SDavid S. Miller limit = min(send_win, cong_win); 1853c1b4a7e6SDavid S. Miller 1854ba244fe9SDavid S. Miller /* If a full-sized TSO skb can be sent, do it. */ 1855605ad7f1SEric Dumazet if (limit >= max_segs * tp->mss_cache) 1856ae8064acSJohn Heffner goto send_now; 1857ba244fe9SDavid S. Miller 185862ad2761SIlpo Järvinen /* Middle in queue won't get any more data, full sendable already? */ 185962ad2761SIlpo Järvinen if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) 186062ad2761SIlpo Järvinen goto send_now; 186162ad2761SIlpo Järvinen 1862ad9f4f50SEric Dumazet win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor); 1863ad9f4f50SEric Dumazet if (win_divisor) { 1864c1b4a7e6SDavid S. Miller u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); 1865c1b4a7e6SDavid S. Miller 1866c1b4a7e6SDavid S. Miller /* If at least some fraction of a window is available, 1867c1b4a7e6SDavid S. Miller * just use it. 1868c1b4a7e6SDavid S. Miller */ 1869ad9f4f50SEric Dumazet chunk /= win_divisor; 1870c1b4a7e6SDavid S. Miller if (limit >= chunk) 1871ae8064acSJohn Heffner goto send_now; 1872c1b4a7e6SDavid S. Miller } else { 1873c1b4a7e6SDavid S. Miller /* Different approach, try not to defer past a single 1874c1b4a7e6SDavid S. Miller * ACK. Receiver should ACK every other full sized 1875c1b4a7e6SDavid S. Miller * frame, so if we have space for more than 3 frames 1876c1b4a7e6SDavid S. Miller * then send now. 1877c1b4a7e6SDavid S. Miller */ 18786b5a5c0dSNeal Cardwell if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) 1879ae8064acSJohn Heffner goto send_now; 1880c1b4a7e6SDavid S. Miller } 1881c1b4a7e6SDavid S. Miller 188250c8339eSEric Dumazet head = tcp_write_queue_head(sk); 188350c8339eSEric Dumazet skb_mstamp_get(&now); 188450c8339eSEric Dumazet age = skb_mstamp_us_delta(&now, &head->skb_mstamp); 188550c8339eSEric Dumazet /* If next ACK is likely to come too late (half srtt), do not defer */ 188650c8339eSEric Dumazet if (age < (tp->srtt_us >> 4)) 188750c8339eSEric Dumazet goto send_now; 188850c8339eSEric Dumazet 18895f852eb5SEric Dumazet /* Ok, it looks like it is advisable to defer. */ 1890ae8064acSJohn Heffner 1891d2e1339fSBendik Rønning Opstad if (cong_win < send_win && cong_win <= skb->len) 1892ca8a2263SNeal Cardwell *is_cwnd_limited = true; 1893ca8a2263SNeal Cardwell 1894a2a385d6SEric Dumazet return true; 1895ae8064acSJohn Heffner 1896ae8064acSJohn Heffner send_now: 1897a2a385d6SEric Dumazet return false; 1898c1b4a7e6SDavid S. Miller } 1899c1b4a7e6SDavid S. Miller 190005cbc0dbSFan Du static inline void tcp_mtu_check_reprobe(struct sock *sk) 190105cbc0dbSFan Du { 190205cbc0dbSFan Du struct inet_connection_sock *icsk = inet_csk(sk); 190305cbc0dbSFan Du struct tcp_sock *tp = tcp_sk(sk); 190405cbc0dbSFan Du struct net *net = sock_net(sk); 190505cbc0dbSFan Du u32 interval; 190605cbc0dbSFan Du s32 delta; 190705cbc0dbSFan Du 190805cbc0dbSFan Du interval = net->ipv4.sysctl_tcp_probe_interval; 190905cbc0dbSFan Du delta = tcp_time_stamp - icsk->icsk_mtup.probe_timestamp; 191005cbc0dbSFan Du if (unlikely(delta >= interval * HZ)) { 191105cbc0dbSFan Du int mss = tcp_current_mss(sk); 191205cbc0dbSFan Du 191305cbc0dbSFan Du /* Update current search range */ 191405cbc0dbSFan Du icsk->icsk_mtup.probe_size = 0; 191505cbc0dbSFan Du icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + 191605cbc0dbSFan Du sizeof(struct tcphdr) + 191705cbc0dbSFan Du icsk->icsk_af_ops->net_header_len; 191805cbc0dbSFan Du icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss); 191905cbc0dbSFan Du 192005cbc0dbSFan Du /* Update probe time stamp */ 192105cbc0dbSFan Du icsk->icsk_mtup.probe_timestamp = tcp_time_stamp; 192205cbc0dbSFan Du } 192305cbc0dbSFan Du } 192405cbc0dbSFan Du 19255d424d5aSJohn Heffner /* Create a new MTU probe if we are ready. 192667edfef7SAndi Kleen * MTU probe is regularly attempting to increase the path MTU by 192767edfef7SAndi Kleen * deliberately sending larger packets. This discovers routing 192867edfef7SAndi Kleen * changes resulting in larger path MTUs. 192967edfef7SAndi Kleen * 19305d424d5aSJohn Heffner * Returns 0 if we should wait to probe (no cwnd available), 19315d424d5aSJohn Heffner * 1 if a probe was sent, 1932056834d9SIlpo Järvinen * -1 otherwise 1933056834d9SIlpo Järvinen */ 19345d424d5aSJohn Heffner static int tcp_mtu_probe(struct sock *sk) 19355d424d5aSJohn Heffner { 19365d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 193712a59abcSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 19385d424d5aSJohn Heffner struct sk_buff *skb, *nskb, *next; 19396b58e0a5SFan Du struct net *net = sock_net(sk); 19405d424d5aSJohn Heffner int probe_size; 194191cc17c0SIlpo Järvinen int size_needed; 194212a59abcSEric Dumazet int copy, len; 19435d424d5aSJohn Heffner int mss_now; 19446b58e0a5SFan Du int interval; 19455d424d5aSJohn Heffner 19465d424d5aSJohn Heffner /* Not currently probing/verifying, 19475d424d5aSJohn Heffner * not in recovery, 19485d424d5aSJohn Heffner * have enough cwnd, and 194912a59abcSEric Dumazet * not SACKing (the variable headers throw things off) 195012a59abcSEric Dumazet */ 195112a59abcSEric Dumazet if (likely(!icsk->icsk_mtup.enabled || 19525d424d5aSJohn Heffner icsk->icsk_mtup.probe_size || 19535d424d5aSJohn Heffner inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 19545d424d5aSJohn Heffner tp->snd_cwnd < 11 || 195512a59abcSEric Dumazet tp->rx_opt.num_sacks || tp->rx_opt.dsack)) 19565d424d5aSJohn Heffner return -1; 19575d424d5aSJohn Heffner 19586b58e0a5SFan Du /* Use binary search for probe_size between tcp_mss_base, 19596b58e0a5SFan Du * and current mss_clamp. if (search_high - search_low) 19606b58e0a5SFan Du * smaller than a threshold, backoff from probing. 19616b58e0a5SFan Du */ 19620c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 19636b58e0a5SFan Du probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high + 19646b58e0a5SFan Du icsk->icsk_mtup.search_low) >> 1); 196591cc17c0SIlpo Järvinen size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; 19666b58e0a5SFan Du interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low; 196705cbc0dbSFan Du /* When misfortune happens, we are reprobing actively, 196805cbc0dbSFan Du * and then reprobe timer has expired. We stick with current 196905cbc0dbSFan Du * probing process by not resetting search range to its orignal. 197005cbc0dbSFan Du */ 19716b58e0a5SFan Du if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) || 197205cbc0dbSFan Du interval < net->ipv4.sysctl_tcp_probe_threshold) { 197305cbc0dbSFan Du /* Check whether enough time has elaplased for 197405cbc0dbSFan Du * another round of probing. 197505cbc0dbSFan Du */ 197605cbc0dbSFan Du tcp_mtu_check_reprobe(sk); 19775d424d5aSJohn Heffner return -1; 19785d424d5aSJohn Heffner } 19795d424d5aSJohn Heffner 19805d424d5aSJohn Heffner /* Have enough data in the send queue to probe? */ 19817f9c33e5SIlpo Järvinen if (tp->write_seq - tp->snd_nxt < size_needed) 19825d424d5aSJohn Heffner return -1; 19835d424d5aSJohn Heffner 198491cc17c0SIlpo Järvinen if (tp->snd_wnd < size_needed) 19855d424d5aSJohn Heffner return -1; 198690840defSIlpo Järvinen if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) 19875d424d5aSJohn Heffner return 0; 19885d424d5aSJohn Heffner 1989d67c58e9SIlpo Järvinen /* Do we need to wait to drain cwnd? With none in flight, don't stall */ 1990d67c58e9SIlpo Järvinen if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { 1991d67c58e9SIlpo Järvinen if (!tcp_packets_in_flight(tp)) 19925d424d5aSJohn Heffner return -1; 19935d424d5aSJohn Heffner else 19945d424d5aSJohn Heffner return 0; 19955d424d5aSJohn Heffner } 19965d424d5aSJohn Heffner 19975d424d5aSJohn Heffner /* We're allowed to probe. Build it now. */ 1998eb934478SEric Dumazet nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false); 199951456b29SIan Morris if (!nskb) 20005d424d5aSJohn Heffner return -1; 20013ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 20023ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 20035d424d5aSJohn Heffner 2004fe067e8aSDavid S. Miller skb = tcp_send_head(sk); 20055d424d5aSJohn Heffner 20065d424d5aSJohn Heffner TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 20075d424d5aSJohn Heffner TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 20084de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; 20095d424d5aSJohn Heffner TCP_SKB_CB(nskb)->sacked = 0; 20105d424d5aSJohn Heffner nskb->csum = 0; 201184fa7933SPatrick McHardy nskb->ip_summed = skb->ip_summed; 20125d424d5aSJohn Heffner 201350c4817eSIlpo Järvinen tcp_insert_write_queue_before(nskb, skb, sk); 201450c4817eSIlpo Järvinen 20155d424d5aSJohn Heffner len = 0; 2016234b6860SIlpo Järvinen tcp_for_write_queue_from_safe(skb, next, sk) { 20175d424d5aSJohn Heffner copy = min_t(int, skb->len, probe_size - len); 20182fe664f1SDouglas Caetano dos Santos if (nskb->ip_summed) { 20195d424d5aSJohn Heffner skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); 20202fe664f1SDouglas Caetano dos Santos } else { 20212fe664f1SDouglas Caetano dos Santos __wsum csum = skb_copy_and_csum_bits(skb, 0, 2022056834d9SIlpo Järvinen skb_put(nskb, copy), 20232fe664f1SDouglas Caetano dos Santos copy, 0); 20242fe664f1SDouglas Caetano dos Santos nskb->csum = csum_block_add(nskb->csum, csum, len); 20252fe664f1SDouglas Caetano dos Santos } 20265d424d5aSJohn Heffner 20275d424d5aSJohn Heffner if (skb->len <= copy) { 20285d424d5aSJohn Heffner /* We've eaten all the data from this skb. 20295d424d5aSJohn Heffner * Throw it away. */ 20304de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 2031fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 20323ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 20335d424d5aSJohn Heffner } else { 20344de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & 2035a3433f35SChangli Gao ~(TCPHDR_FIN|TCPHDR_PSH); 20365d424d5aSJohn Heffner if (!skb_shinfo(skb)->nr_frags) { 20375d424d5aSJohn Heffner skb_pull(skb, copy); 203884fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 2039056834d9SIlpo Järvinen skb->csum = csum_partial(skb->data, 2040056834d9SIlpo Järvinen skb->len, 0); 20415d424d5aSJohn Heffner } else { 20425d424d5aSJohn Heffner __pskb_trim_head(skb, copy); 20435bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 20445d424d5aSJohn Heffner } 20455d424d5aSJohn Heffner TCP_SKB_CB(skb)->seq += copy; 20465d424d5aSJohn Heffner } 20475d424d5aSJohn Heffner 20485d424d5aSJohn Heffner len += copy; 2049234b6860SIlpo Järvinen 2050234b6860SIlpo Järvinen if (len >= probe_size) 2051234b6860SIlpo Järvinen break; 20525d424d5aSJohn Heffner } 20535bbb432cSEric Dumazet tcp_init_tso_segs(nskb, nskb->len); 20545d424d5aSJohn Heffner 20555d424d5aSJohn Heffner /* We're ready to send. If this fails, the probe will 20567faee5c0SEric Dumazet * be resegmented into mss-sized pieces by tcp_write_xmit(). 20577faee5c0SEric Dumazet */ 20585d424d5aSJohn Heffner if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 20595d424d5aSJohn Heffner /* Decrement cwnd here because we are sending 20605d424d5aSJohn Heffner * effectively two packets. */ 20615d424d5aSJohn Heffner tp->snd_cwnd--; 206266f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, nskb); 20635d424d5aSJohn Heffner 20645d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 20650e7b1368SJohn Heffner tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 20660e7b1368SJohn Heffner tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 20675d424d5aSJohn Heffner 20685d424d5aSJohn Heffner return 1; 20695d424d5aSJohn Heffner } 20705d424d5aSJohn Heffner 20715d424d5aSJohn Heffner return -1; 20725d424d5aSJohn Heffner } 20735d424d5aSJohn Heffner 2074f9616c35SEric Dumazet /* TCP Small Queues : 2075f9616c35SEric Dumazet * Control number of packets in qdisc/devices to two packets / or ~1 ms. 2076f9616c35SEric Dumazet * (These limits are doubled for retransmits) 2077f9616c35SEric Dumazet * This allows for : 2078f9616c35SEric Dumazet * - better RTT estimation and ACK scheduling 2079f9616c35SEric Dumazet * - faster recovery 2080f9616c35SEric Dumazet * - high rates 2081f9616c35SEric Dumazet * Alas, some drivers / subsystems require a fair amount 2082f9616c35SEric Dumazet * of queued bytes to ensure line rate. 2083f9616c35SEric Dumazet * One example is wifi aggregation (802.11 AMPDU) 2084f9616c35SEric Dumazet */ 2085f9616c35SEric Dumazet static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb, 2086f9616c35SEric Dumazet unsigned int factor) 2087f9616c35SEric Dumazet { 2088f9616c35SEric Dumazet unsigned int limit; 2089f9616c35SEric Dumazet 2090f9616c35SEric Dumazet limit = max(2 * skb->truesize, sk->sk_pacing_rate >> 10); 2091f9616c35SEric Dumazet limit = min_t(u32, limit, sysctl_tcp_limit_output_bytes); 2092f9616c35SEric Dumazet limit <<= factor; 2093f9616c35SEric Dumazet 2094f9616c35SEric Dumazet if (atomic_read(&sk->sk_wmem_alloc) > limit) { 209575eefc6cSEric Dumazet /* Always send the 1st or 2nd skb in write queue. 209675eefc6cSEric Dumazet * No need to wait for TX completion to call us back, 209775eefc6cSEric Dumazet * after softirq/tasklet schedule. 209875eefc6cSEric Dumazet * This helps when TX completions are delayed too much. 209975eefc6cSEric Dumazet */ 210075eefc6cSEric Dumazet if (skb == sk->sk_write_queue.next || 210175eefc6cSEric Dumazet skb->prev == sk->sk_write_queue.next) 210275eefc6cSEric Dumazet return false; 210375eefc6cSEric Dumazet 21047aa5470cSEric Dumazet set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags); 2105f9616c35SEric Dumazet /* It is possible TX completion already happened 2106f9616c35SEric Dumazet * before we set TSQ_THROTTLED, so we must 2107f9616c35SEric Dumazet * test again the condition. 2108f9616c35SEric Dumazet */ 2109f9616c35SEric Dumazet smp_mb__after_atomic(); 2110f9616c35SEric Dumazet if (atomic_read(&sk->sk_wmem_alloc) > limit) 2111f9616c35SEric Dumazet return true; 2112f9616c35SEric Dumazet } 2113f9616c35SEric Dumazet return false; 2114f9616c35SEric Dumazet } 2115f9616c35SEric Dumazet 211605b055e8SFrancis Yan static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new) 211705b055e8SFrancis Yan { 211805b055e8SFrancis Yan const u32 now = tcp_time_stamp; 211905b055e8SFrancis Yan 212005b055e8SFrancis Yan if (tp->chrono_type > TCP_CHRONO_UNSPEC) 212105b055e8SFrancis Yan tp->chrono_stat[tp->chrono_type - 1] += now - tp->chrono_start; 212205b055e8SFrancis Yan tp->chrono_start = now; 212305b055e8SFrancis Yan tp->chrono_type = new; 212405b055e8SFrancis Yan } 212505b055e8SFrancis Yan 212605b055e8SFrancis Yan void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type) 212705b055e8SFrancis Yan { 212805b055e8SFrancis Yan struct tcp_sock *tp = tcp_sk(sk); 212905b055e8SFrancis Yan 213005b055e8SFrancis Yan /* If there are multiple conditions worthy of tracking in a 21310f87230dSFrancis Yan * chronograph then the highest priority enum takes precedence 21320f87230dSFrancis Yan * over the other conditions. So that if something "more interesting" 213305b055e8SFrancis Yan * starts happening, stop the previous chrono and start a new one. 213405b055e8SFrancis Yan */ 213505b055e8SFrancis Yan if (type > tp->chrono_type) 213605b055e8SFrancis Yan tcp_chrono_set(tp, type); 213705b055e8SFrancis Yan } 213805b055e8SFrancis Yan 213905b055e8SFrancis Yan void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type) 214005b055e8SFrancis Yan { 214105b055e8SFrancis Yan struct tcp_sock *tp = tcp_sk(sk); 214205b055e8SFrancis Yan 21430f87230dSFrancis Yan 21440f87230dSFrancis Yan /* There are multiple conditions worthy of tracking in a 21450f87230dSFrancis Yan * chronograph, so that the highest priority enum takes 21460f87230dSFrancis Yan * precedence over the other conditions (see tcp_chrono_start). 21470f87230dSFrancis Yan * If a condition stops, we only stop chrono tracking if 21480f87230dSFrancis Yan * it's the "most interesting" or current chrono we are 21490f87230dSFrancis Yan * tracking and starts busy chrono if we have pending data. 21500f87230dSFrancis Yan */ 21510f87230dSFrancis Yan if (tcp_write_queue_empty(sk)) 215205b055e8SFrancis Yan tcp_chrono_set(tp, TCP_CHRONO_UNSPEC); 21530f87230dSFrancis Yan else if (type == tp->chrono_type) 21540f87230dSFrancis Yan tcp_chrono_set(tp, TCP_CHRONO_BUSY); 215505b055e8SFrancis Yan } 215605b055e8SFrancis Yan 21571da177e4SLinus Torvalds /* This routine writes packets to the network. It advances the 21581da177e4SLinus Torvalds * send_head. This happens as incoming acks open up the remote 21591da177e4SLinus Torvalds * window for us. 21601da177e4SLinus Torvalds * 2161f8269a49SIlpo Järvinen * LARGESEND note: !tcp_urg_mode is overkill, only frames between 2162f8269a49SIlpo Järvinen * snd_up-64k-mss .. snd_up cannot be large. However, taking into 2163f8269a49SIlpo Järvinen * account rare use of URG, this is not a big flaw. 2164f8269a49SIlpo Järvinen * 21656ba8a3b1SNandita Dukkipati * Send at most one packet when push_one > 0. Temporarily ignore 21666ba8a3b1SNandita Dukkipati * cwnd limit to force at most one packet out when push_one == 2. 21676ba8a3b1SNandita Dukkipati 2168a2a385d6SEric Dumazet * Returns true, if no segments are in flight and we have queued segments, 2169a2a385d6SEric Dumazet * but cannot send anything now because of SWS or another problem. 21701da177e4SLinus Torvalds */ 2171a2a385d6SEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 2172d5dd9175SIlpo Järvinen int push_one, gfp_t gfp) 21731da177e4SLinus Torvalds { 21741da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 217592df7b51SDavid S. Miller struct sk_buff *skb; 2176c1b4a7e6SDavid S. Miller unsigned int tso_segs, sent_pkts; 2177c1b4a7e6SDavid S. Miller int cwnd_quota; 21785d424d5aSJohn Heffner int result; 21795615f886SFrancis Yan bool is_cwnd_limited = false, is_rwnd_limited = false; 2180605ad7f1SEric Dumazet u32 max_segs; 21811da177e4SLinus Torvalds 2182c1b4a7e6SDavid S. Miller sent_pkts = 0; 21835d424d5aSJohn Heffner 2184d5dd9175SIlpo Järvinen if (!push_one) { 21855d424d5aSJohn Heffner /* Do MTU probing. */ 2186d5dd9175SIlpo Järvinen result = tcp_mtu_probe(sk); 2187d5dd9175SIlpo Järvinen if (!result) { 2188a2a385d6SEric Dumazet return false; 21895d424d5aSJohn Heffner } else if (result > 0) { 21905d424d5aSJohn Heffner sent_pkts = 1; 21915d424d5aSJohn Heffner } 2192d5dd9175SIlpo Järvinen } 21935d424d5aSJohn Heffner 2194ed6e7268SNeal Cardwell max_segs = tcp_tso_segs(sk, mss_now); 2195fe067e8aSDavid S. Miller while ((skb = tcp_send_head(sk))) { 2196c8ac3774SHerbert Xu unsigned int limit; 2197c8ac3774SHerbert Xu 21985bbb432cSEric Dumazet tso_segs = tcp_init_tso_segs(skb, mss_now); 2199c1b4a7e6SDavid S. Miller BUG_ON(!tso_segs); 2200c1b4a7e6SDavid S. Miller 22019d186cacSAndrey Vagin if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) { 22027faee5c0SEric Dumazet /* "skb_mstamp" is used as a start point for the retransmit timer */ 22037faee5c0SEric Dumazet skb_mstamp_get(&skb->skb_mstamp); 2204ec342325SAndrew Vagin goto repair; /* Skip network transmission */ 22059d186cacSAndrey Vagin } 2206ec342325SAndrew Vagin 2207b68e9f85SHerbert Xu cwnd_quota = tcp_cwnd_test(tp, skb); 22086ba8a3b1SNandita Dukkipati if (!cwnd_quota) { 22096ba8a3b1SNandita Dukkipati if (push_one == 2) 22106ba8a3b1SNandita Dukkipati /* Force out a loss probe pkt. */ 22116ba8a3b1SNandita Dukkipati cwnd_quota = 1; 22126ba8a3b1SNandita Dukkipati else 2213b68e9f85SHerbert Xu break; 22146ba8a3b1SNandita Dukkipati } 2215b68e9f85SHerbert Xu 22165615f886SFrancis Yan if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) { 22175615f886SFrancis Yan is_rwnd_limited = true; 2218b68e9f85SHerbert Xu break; 22195615f886SFrancis Yan } 2220b68e9f85SHerbert Xu 2221d6a4e26aSEric Dumazet if (tso_segs == 1) { 2222aa93466bSDavid S. Miller if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 2223aa93466bSDavid S. Miller (tcp_skb_is_last(sk, skb) ? 2224aa93466bSDavid S. Miller nonagle : TCP_NAGLE_PUSH)))) 2225aa93466bSDavid S. Miller break; 2226c1b4a7e6SDavid S. Miller } else { 2227ca8a2263SNeal Cardwell if (!push_one && 2228605ad7f1SEric Dumazet tcp_tso_should_defer(sk, skb, &is_cwnd_limited, 2229605ad7f1SEric Dumazet max_segs)) 2230aa93466bSDavid S. Miller break; 2231c1b4a7e6SDavid S. Miller } 2232aa93466bSDavid S. Miller 2233605ad7f1SEric Dumazet limit = mss_now; 2234d6a4e26aSEric Dumazet if (tso_segs > 1 && !tcp_urg_mode(tp)) 2235605ad7f1SEric Dumazet limit = tcp_mss_split_point(sk, skb, mss_now, 2236605ad7f1SEric Dumazet min_t(unsigned int, 2237605ad7f1SEric Dumazet cwnd_quota, 2238605ad7f1SEric Dumazet max_segs), 2239605ad7f1SEric Dumazet nonagle); 2240605ad7f1SEric Dumazet 2241605ad7f1SEric Dumazet if (skb->len > limit && 2242605ad7f1SEric Dumazet unlikely(tso_fragment(sk, skb, limit, mss_now, gfp))) 2243605ad7f1SEric Dumazet break; 2244605ad7f1SEric Dumazet 22457aa5470cSEric Dumazet if (test_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags)) 22467aa5470cSEric Dumazet clear_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags); 2247f9616c35SEric Dumazet if (tcp_small_queue_check(sk, skb, 0)) 224846d3ceabSEric Dumazet break; 2249c9eeec26SEric Dumazet 2250d5dd9175SIlpo Järvinen if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) 22511da177e4SLinus Torvalds break; 22521da177e4SLinus Torvalds 2253ec342325SAndrew Vagin repair: 22541da177e4SLinus Torvalds /* Advance the send_head. This one is sent out. 22551da177e4SLinus Torvalds * This call will increment packets_out. 22561da177e4SLinus Torvalds */ 225766f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 22581da177e4SLinus Torvalds 22591da177e4SLinus Torvalds tcp_minshall_update(tp, mss_now, skb); 2260a262f0cdSNandita Dukkipati sent_pkts += tcp_skb_pcount(skb); 2261d5dd9175SIlpo Järvinen 2262d5dd9175SIlpo Järvinen if (push_one) 2263d5dd9175SIlpo Järvinen break; 22641da177e4SLinus Torvalds } 22651da177e4SLinus Torvalds 22665615f886SFrancis Yan if (is_rwnd_limited) 22675615f886SFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED); 22685615f886SFrancis Yan else 22695615f886SFrancis Yan tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED); 22705615f886SFrancis Yan 2271aa93466bSDavid S. Miller if (likely(sent_pkts)) { 2272684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 2273684bad11SYuchung Cheng tp->prr_out += sent_pkts; 22746ba8a3b1SNandita Dukkipati 22756ba8a3b1SNandita Dukkipati /* Send one loss probe per tail loss episode. */ 22766ba8a3b1SNandita Dukkipati if (push_one != 2) 22776ba8a3b1SNandita Dukkipati tcp_schedule_loss_probe(sk); 2278d2e1339fSBendik Rønning Opstad is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd); 2279ca8a2263SNeal Cardwell tcp_cwnd_validate(sk, is_cwnd_limited); 2280a2a385d6SEric Dumazet return false; 22811da177e4SLinus Torvalds } 2282b340b264SYuchung Cheng return !tp->packets_out && tcp_send_head(sk); 22836ba8a3b1SNandita Dukkipati } 22846ba8a3b1SNandita Dukkipati 22856ba8a3b1SNandita Dukkipati bool tcp_schedule_loss_probe(struct sock *sk) 22866ba8a3b1SNandita Dukkipati { 22876ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 22886ba8a3b1SNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 22896ba8a3b1SNandita Dukkipati u32 timeout, tlp_time_stamp, rto_time_stamp; 2290740b0f18SEric Dumazet u32 rtt = usecs_to_jiffies(tp->srtt_us >> 3); 22916ba8a3b1SNandita Dukkipati 22926ba8a3b1SNandita Dukkipati if (WARN_ON(icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS)) 22936ba8a3b1SNandita Dukkipati return false; 22946ba8a3b1SNandita Dukkipati /* No consecutive loss probes. */ 22956ba8a3b1SNandita Dukkipati if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) { 22966ba8a3b1SNandita Dukkipati tcp_rearm_rto(sk); 22976ba8a3b1SNandita Dukkipati return false; 22986ba8a3b1SNandita Dukkipati } 22996ba8a3b1SNandita Dukkipati /* Don't do any loss probe on a Fast Open connection before 3WHS 23006ba8a3b1SNandita Dukkipati * finishes. 23016ba8a3b1SNandita Dukkipati */ 2302f9b99582SYuchung Cheng if (tp->fastopen_rsk) 23036ba8a3b1SNandita Dukkipati return false; 23046ba8a3b1SNandita Dukkipati 23056ba8a3b1SNandita Dukkipati /* TLP is only scheduled when next timer event is RTO. */ 23066ba8a3b1SNandita Dukkipati if (icsk->icsk_pending != ICSK_TIME_RETRANS) 23076ba8a3b1SNandita Dukkipati return false; 23086ba8a3b1SNandita Dukkipati 23096ba8a3b1SNandita Dukkipati /* Schedule a loss probe in 2*RTT for SACK capable connections 23106ba8a3b1SNandita Dukkipati * in Open state, that are either limited by cwnd or application. 23116ba8a3b1SNandita Dukkipati */ 2312f9b99582SYuchung Cheng if (sysctl_tcp_early_retrans < 3 || !tp->packets_out || 23136ba8a3b1SNandita Dukkipati !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open) 23146ba8a3b1SNandita Dukkipati return false; 23156ba8a3b1SNandita Dukkipati 23166ba8a3b1SNandita Dukkipati if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) && 23176ba8a3b1SNandita Dukkipati tcp_send_head(sk)) 23186ba8a3b1SNandita Dukkipati return false; 23196ba8a3b1SNandita Dukkipati 23206ba8a3b1SNandita Dukkipati /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account 2321f9b99582SYuchung Cheng * for delayed ack when there's one outstanding packet. If no RTT 2322f9b99582SYuchung Cheng * sample is available then probe after TCP_TIMEOUT_INIT. 23236ba8a3b1SNandita Dukkipati */ 2324f9b99582SYuchung Cheng timeout = rtt << 1 ? : TCP_TIMEOUT_INIT; 23256ba8a3b1SNandita Dukkipati if (tp->packets_out == 1) 23266ba8a3b1SNandita Dukkipati timeout = max_t(u32, timeout, 23276ba8a3b1SNandita Dukkipati (rtt + (rtt >> 1) + TCP_DELACK_MAX)); 23286ba8a3b1SNandita Dukkipati timeout = max_t(u32, timeout, msecs_to_jiffies(10)); 23296ba8a3b1SNandita Dukkipati 23306ba8a3b1SNandita Dukkipati /* If RTO is shorter, just schedule TLP in its place. */ 23316ba8a3b1SNandita Dukkipati tlp_time_stamp = tcp_time_stamp + timeout; 23326ba8a3b1SNandita Dukkipati rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout; 23336ba8a3b1SNandita Dukkipati if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) { 23346ba8a3b1SNandita Dukkipati s32 delta = rto_time_stamp - tcp_time_stamp; 23356ba8a3b1SNandita Dukkipati if (delta > 0) 23366ba8a3b1SNandita Dukkipati timeout = delta; 23376ba8a3b1SNandita Dukkipati } 23386ba8a3b1SNandita Dukkipati 23396ba8a3b1SNandita Dukkipati inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, 23406ba8a3b1SNandita Dukkipati TCP_RTO_MAX); 23416ba8a3b1SNandita Dukkipati return true; 23426ba8a3b1SNandita Dukkipati } 23436ba8a3b1SNandita Dukkipati 23441f3279aeSEric Dumazet /* Thanks to skb fast clones, we can detect if a prior transmit of 23451f3279aeSEric Dumazet * a packet is still in a qdisc or driver queue. 23461f3279aeSEric Dumazet * In this case, there is very little point doing a retransmit ! 23471f3279aeSEric Dumazet */ 23481f3279aeSEric Dumazet static bool skb_still_in_host_queue(const struct sock *sk, 23491f3279aeSEric Dumazet const struct sk_buff *skb) 23501f3279aeSEric Dumazet { 235139bb5e62SEric Dumazet if (unlikely(skb_fclone_busy(sk, skb))) { 2352c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), 23531f3279aeSEric Dumazet LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES); 23541f3279aeSEric Dumazet return true; 23551f3279aeSEric Dumazet } 23561f3279aeSEric Dumazet return false; 23571f3279aeSEric Dumazet } 23581f3279aeSEric Dumazet 2359b340b264SYuchung Cheng /* When probe timeout (PTO) fires, try send a new segment if possible, else 23606ba8a3b1SNandita Dukkipati * retransmit the last segment. 23616ba8a3b1SNandita Dukkipati */ 23626ba8a3b1SNandita Dukkipati void tcp_send_loss_probe(struct sock *sk) 23636ba8a3b1SNandita Dukkipati { 23649b717a8dSNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 23656ba8a3b1SNandita Dukkipati struct sk_buff *skb; 23666ba8a3b1SNandita Dukkipati int pcount; 23676ba8a3b1SNandita Dukkipati int mss = tcp_current_mss(sk); 23686ba8a3b1SNandita Dukkipati 2369b340b264SYuchung Cheng skb = tcp_send_head(sk); 2370b340b264SYuchung Cheng if (skb) { 2371b340b264SYuchung Cheng if (tcp_snd_wnd_test(tp, skb, mss)) { 2372b340b264SYuchung Cheng pcount = tp->packets_out; 2373b340b264SYuchung Cheng tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC); 2374b340b264SYuchung Cheng if (tp->packets_out > pcount) 2375b340b264SYuchung Cheng goto probe_sent; 23766ba8a3b1SNandita Dukkipati goto rearm_timer; 23776ba8a3b1SNandita Dukkipati } 2378b340b264SYuchung Cheng skb = tcp_write_queue_prev(sk, skb); 2379b340b264SYuchung Cheng } else { 2380b340b264SYuchung Cheng skb = tcp_write_queue_tail(sk); 2381b340b264SYuchung Cheng } 23826ba8a3b1SNandita Dukkipati 23839b717a8dSNandita Dukkipati /* At most one outstanding TLP retransmission. */ 23849b717a8dSNandita Dukkipati if (tp->tlp_high_seq) 23859b717a8dSNandita Dukkipati goto rearm_timer; 23869b717a8dSNandita Dukkipati 23876ba8a3b1SNandita Dukkipati /* Retransmit last segment. */ 23886ba8a3b1SNandita Dukkipati if (WARN_ON(!skb)) 23896ba8a3b1SNandita Dukkipati goto rearm_timer; 23906ba8a3b1SNandita Dukkipati 23911f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 23921f3279aeSEric Dumazet goto rearm_timer; 23931f3279aeSEric Dumazet 23946ba8a3b1SNandita Dukkipati pcount = tcp_skb_pcount(skb); 23956ba8a3b1SNandita Dukkipati if (WARN_ON(!pcount)) 23966ba8a3b1SNandita Dukkipati goto rearm_timer; 23976ba8a3b1SNandita Dukkipati 23986ba8a3b1SNandita Dukkipati if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) { 23996cc55e09SOctavian Purdila if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss, 24006cc55e09SOctavian Purdila GFP_ATOMIC))) 24016ba8a3b1SNandita Dukkipati goto rearm_timer; 2402b340b264SYuchung Cheng skb = tcp_write_queue_next(sk, skb); 24036ba8a3b1SNandita Dukkipati } 24046ba8a3b1SNandita Dukkipati 24056ba8a3b1SNandita Dukkipati if (WARN_ON(!skb || !tcp_skb_pcount(skb))) 24066ba8a3b1SNandita Dukkipati goto rearm_timer; 24076ba8a3b1SNandita Dukkipati 240810d3be56SEric Dumazet if (__tcp_retransmit_skb(sk, skb, 1)) 2409b340b264SYuchung Cheng goto rearm_timer; 24106ba8a3b1SNandita Dukkipati 24119b717a8dSNandita Dukkipati /* Record snd_nxt for loss detection. */ 24129b717a8dSNandita Dukkipati tp->tlp_high_seq = tp->snd_nxt; 24139b717a8dSNandita Dukkipati 2414b340b264SYuchung Cheng probe_sent: 2415c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES); 2416fcd16c0aSYuchung Cheng /* Reset s.t. tcp_rearm_rto will restart timer from now */ 2417fcd16c0aSYuchung Cheng inet_csk(sk)->icsk_pending = 0; 2418b340b264SYuchung Cheng rearm_timer: 2419fcd16c0aSYuchung Cheng tcp_rearm_rto(sk); 24201da177e4SLinus Torvalds } 24211da177e4SLinus Torvalds 2422a762a980SDavid S. Miller /* Push out any pending frames which were held back due to 2423a762a980SDavid S. Miller * TCP_CORK or attempt at coalescing tiny packets. 2424a762a980SDavid S. Miller * The socket must be locked by the caller. 2425a762a980SDavid S. Miller */ 24269e412ba7SIlpo Järvinen void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 24279e412ba7SIlpo Järvinen int nonagle) 2428a762a980SDavid S. Miller { 2429726e07a8SIlpo Järvinen /* If we are closed, the bytes will have to remain here. 2430726e07a8SIlpo Järvinen * In time closedown will finish, we empty the write queue and 2431726e07a8SIlpo Järvinen * all will be happy. 2432726e07a8SIlpo Järvinen */ 2433726e07a8SIlpo Järvinen if (unlikely(sk->sk_state == TCP_CLOSE)) 2434726e07a8SIlpo Järvinen return; 2435726e07a8SIlpo Järvinen 243699a1dec7SMel Gorman if (tcp_write_xmit(sk, cur_mss, nonagle, 0, 24377450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC))) 24389e412ba7SIlpo Järvinen tcp_check_probe_timer(sk); 2439a762a980SDavid S. Miller } 2440a762a980SDavid S. Miller 2441c1b4a7e6SDavid S. Miller /* Send _single_ skb sitting at the send head. This function requires 2442c1b4a7e6SDavid S. Miller * true push pending frames to setup probe timer etc. 2443c1b4a7e6SDavid S. Miller */ 2444c1b4a7e6SDavid S. Miller void tcp_push_one(struct sock *sk, unsigned int mss_now) 2445c1b4a7e6SDavid S. Miller { 2446fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 2447c1b4a7e6SDavid S. Miller 2448c1b4a7e6SDavid S. Miller BUG_ON(!skb || skb->len < mss_now); 2449c1b4a7e6SDavid S. Miller 2450d5dd9175SIlpo Järvinen tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); 2451c1b4a7e6SDavid S. Miller } 2452c1b4a7e6SDavid S. Miller 24531da177e4SLinus Torvalds /* This function returns the amount that we can raise the 24541da177e4SLinus Torvalds * usable window based on the following constraints 24551da177e4SLinus Torvalds * 24561da177e4SLinus Torvalds * 1. The window can never be shrunk once it is offered (RFC 793) 24571da177e4SLinus Torvalds * 2. We limit memory per socket 24581da177e4SLinus Torvalds * 24591da177e4SLinus Torvalds * RFC 1122: 24601da177e4SLinus Torvalds * "the suggested [SWS] avoidance algorithm for the receiver is to keep 24611da177e4SLinus Torvalds * RECV.NEXT + RCV.WIN fixed until: 24621da177e4SLinus Torvalds * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 24631da177e4SLinus Torvalds * 24641da177e4SLinus Torvalds * i.e. don't raise the right edge of the window until you can raise 24651da177e4SLinus Torvalds * it at least MSS bytes. 24661da177e4SLinus Torvalds * 24671da177e4SLinus Torvalds * Unfortunately, the recommended algorithm breaks header prediction, 24681da177e4SLinus Torvalds * since header prediction assumes th->window stays fixed. 24691da177e4SLinus Torvalds * 24701da177e4SLinus Torvalds * Strictly speaking, keeping th->window fixed violates the receiver 24711da177e4SLinus Torvalds * side SWS prevention criteria. The problem is that under this rule 24721da177e4SLinus Torvalds * a stream of single byte packets will cause the right side of the 24731da177e4SLinus Torvalds * window to always advance by a single byte. 24741da177e4SLinus Torvalds * 24751da177e4SLinus Torvalds * Of course, if the sender implements sender side SWS prevention 24761da177e4SLinus Torvalds * then this will not be a problem. 24771da177e4SLinus Torvalds * 24781da177e4SLinus Torvalds * BSD seems to make the following compromise: 24791da177e4SLinus Torvalds * 24801da177e4SLinus Torvalds * If the free space is less than the 1/4 of the maximum 24811da177e4SLinus Torvalds * space available and the free space is less than 1/2 mss, 24821da177e4SLinus Torvalds * then set the window to 0. 24831da177e4SLinus Torvalds * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 24841da177e4SLinus Torvalds * Otherwise, just prevent the window from shrinking 24851da177e4SLinus Torvalds * and from being larger than the largest representable value. 24861da177e4SLinus Torvalds * 24871da177e4SLinus Torvalds * This prevents incremental opening of the window in the regime 24881da177e4SLinus Torvalds * where TCP is limited by the speed of the reader side taking 24891da177e4SLinus Torvalds * data out of the TCP receive queue. It does nothing about 24901da177e4SLinus Torvalds * those cases where the window is constrained on the sender side 24911da177e4SLinus Torvalds * because the pipeline is full. 24921da177e4SLinus Torvalds * 24931da177e4SLinus Torvalds * BSD also seems to "accidentally" limit itself to windows that are a 24941da177e4SLinus Torvalds * multiple of MSS, at least until the free space gets quite small. 24951da177e4SLinus Torvalds * This would appear to be a side effect of the mbuf implementation. 24961da177e4SLinus Torvalds * Combining these two algorithms results in the observed behavior 24971da177e4SLinus Torvalds * of having a fixed window size at almost all times. 24981da177e4SLinus Torvalds * 24991da177e4SLinus Torvalds * Below we obtain similar behavior by forcing the offered window to 25001da177e4SLinus Torvalds * a multiple of the mss when it is feasible to do so. 25011da177e4SLinus Torvalds * 25021da177e4SLinus Torvalds * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 25031da177e4SLinus Torvalds * Regular options like TIMESTAMP are taken into account. 25041da177e4SLinus Torvalds */ 25051da177e4SLinus Torvalds u32 __tcp_select_window(struct sock *sk) 25061da177e4SLinus Torvalds { 2507463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 25081da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2509caa20d9aSStephen Hemminger /* MSS for the peer's data. Previous versions used mss_clamp 25101da177e4SLinus Torvalds * here. I don't know if the value based on our guesses 25111da177e4SLinus Torvalds * of peer's MSS is better for the performance. It's more correct 25121da177e4SLinus Torvalds * but may be worse for the performance because of rcv_mss 25131da177e4SLinus Torvalds * fluctuations. --SAW 1998/11/1 25141da177e4SLinus Torvalds */ 2515463c84b9SArnaldo Carvalho de Melo int mss = icsk->icsk_ack.rcv_mss; 25161da177e4SLinus Torvalds int free_space = tcp_space(sk); 251786c1a045SFlorian Westphal int allowed_space = tcp_full_space(sk); 251886c1a045SFlorian Westphal int full_space = min_t(int, tp->window_clamp, allowed_space); 25191da177e4SLinus Torvalds int window; 25201da177e4SLinus Torvalds 25211da177e4SLinus Torvalds if (mss > full_space) 25221da177e4SLinus Torvalds mss = full_space; 25231da177e4SLinus Torvalds 2524b92edbe0SEric Dumazet if (free_space < (full_space >> 1)) { 2525463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.quick = 0; 25261da177e4SLinus Torvalds 2527b8da51ebSEric Dumazet if (tcp_under_memory_pressure(sk)) 2528056834d9SIlpo Järvinen tp->rcv_ssthresh = min(tp->rcv_ssthresh, 2529056834d9SIlpo Järvinen 4U * tp->advmss); 25301da177e4SLinus Torvalds 253186c1a045SFlorian Westphal /* free_space might become our new window, make sure we don't 253286c1a045SFlorian Westphal * increase it due to wscale. 253386c1a045SFlorian Westphal */ 253486c1a045SFlorian Westphal free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale); 253586c1a045SFlorian Westphal 253686c1a045SFlorian Westphal /* if free space is less than mss estimate, or is below 1/16th 253786c1a045SFlorian Westphal * of the maximum allowed, try to move to zero-window, else 253886c1a045SFlorian Westphal * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and 253986c1a045SFlorian Westphal * new incoming data is dropped due to memory limits. 254086c1a045SFlorian Westphal * With large window, mss test triggers way too late in order 254186c1a045SFlorian Westphal * to announce zero window in time before rmem limit kicks in. 254286c1a045SFlorian Westphal */ 254386c1a045SFlorian Westphal if (free_space < (allowed_space >> 4) || free_space < mss) 25441da177e4SLinus Torvalds return 0; 25451da177e4SLinus Torvalds } 25461da177e4SLinus Torvalds 25471da177e4SLinus Torvalds if (free_space > tp->rcv_ssthresh) 25481da177e4SLinus Torvalds free_space = tp->rcv_ssthresh; 25491da177e4SLinus Torvalds 25501da177e4SLinus Torvalds /* Don't do rounding if we are using window scaling, since the 25511da177e4SLinus Torvalds * scaled window will not line up with the MSS boundary anyway. 25521da177e4SLinus Torvalds */ 25531da177e4SLinus Torvalds window = tp->rcv_wnd; 25541da177e4SLinus Torvalds if (tp->rx_opt.rcv_wscale) { 25551da177e4SLinus Torvalds window = free_space; 25561da177e4SLinus Torvalds 25571da177e4SLinus Torvalds /* Advertise enough space so that it won't get scaled away. 25581da177e4SLinus Torvalds * Import case: prevent zero window announcement if 25591da177e4SLinus Torvalds * 1<<rcv_wscale > mss. 25601da177e4SLinus Torvalds */ 25611da177e4SLinus Torvalds if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window) 25621da177e4SLinus Torvalds window = (((window >> tp->rx_opt.rcv_wscale) + 1) 25631da177e4SLinus Torvalds << tp->rx_opt.rcv_wscale); 25641da177e4SLinus Torvalds } else { 25651da177e4SLinus Torvalds /* Get the largest window that is a nice multiple of mss. 25661da177e4SLinus Torvalds * Window clamp already applied above. 25671da177e4SLinus Torvalds * If our current window offering is within 1 mss of the 25681da177e4SLinus Torvalds * free space we just keep it. This prevents the divide 25691da177e4SLinus Torvalds * and multiply from happening most of the time. 25701da177e4SLinus Torvalds * We also don't do any window rounding when the free space 25711da177e4SLinus Torvalds * is too small. 25721da177e4SLinus Torvalds */ 25731da177e4SLinus Torvalds if (window <= free_space - mss || window > free_space) 25741da177e4SLinus Torvalds window = (free_space / mss) * mss; 257584565070SJohn Heffner else if (mss == full_space && 2576b92edbe0SEric Dumazet free_space > window + (full_space >> 1)) 257784565070SJohn Heffner window = free_space; 25781da177e4SLinus Torvalds } 25791da177e4SLinus Torvalds 25801da177e4SLinus Torvalds return window; 25811da177e4SLinus Torvalds } 25821da177e4SLinus Torvalds 2583cfea5a68SMartin KaFai Lau void tcp_skb_collapse_tstamp(struct sk_buff *skb, 2584082ac2d5SMartin KaFai Lau const struct sk_buff *next_skb) 2585082ac2d5SMartin KaFai Lau { 25860a2cf20cSSoheil Hassas Yeganeh if (unlikely(tcp_has_tx_tstamp(next_skb))) { 25870a2cf20cSSoheil Hassas Yeganeh const struct skb_shared_info *next_shinfo = 25880a2cf20cSSoheil Hassas Yeganeh skb_shinfo(next_skb); 2589082ac2d5SMartin KaFai Lau struct skb_shared_info *shinfo = skb_shinfo(skb); 2590082ac2d5SMartin KaFai Lau 25910a2cf20cSSoheil Hassas Yeganeh shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP; 2592082ac2d5SMartin KaFai Lau shinfo->tskey = next_shinfo->tskey; 25932de8023eSMartin KaFai Lau TCP_SKB_CB(skb)->txstamp_ack |= 25942de8023eSMartin KaFai Lau TCP_SKB_CB(next_skb)->txstamp_ack; 2595082ac2d5SMartin KaFai Lau } 2596082ac2d5SMartin KaFai Lau } 2597082ac2d5SMartin KaFai Lau 25984a17fc3aSIlpo Järvinen /* Collapses two adjacent SKB's during retransmission. */ 2599f8071cdeSEric Dumazet static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) 26001da177e4SLinus Torvalds { 26011da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2602fe067e8aSDavid S. Miller struct sk_buff *next_skb = tcp_write_queue_next(sk, skb); 2603058dc334SIlpo Järvinen int skb_size, next_skb_size; 26041da177e4SLinus Torvalds 2605058dc334SIlpo Järvinen skb_size = skb->len; 2606058dc334SIlpo Järvinen next_skb_size = next_skb->len; 26071da177e4SLinus Torvalds 2608058dc334SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); 26091da177e4SLinus Torvalds 2610f8071cdeSEric Dumazet if (next_skb_size) { 2611f8071cdeSEric Dumazet if (next_skb_size <= skb_availroom(skb)) 2612f8071cdeSEric Dumazet skb_copy_bits(next_skb, 0, skb_put(skb, next_skb_size), 2613f8071cdeSEric Dumazet next_skb_size); 2614f8071cdeSEric Dumazet else if (!skb_shift(skb, next_skb, next_skb_size)) 2615f8071cdeSEric Dumazet return false; 2616f8071cdeSEric Dumazet } 26176859d494SIlpo Järvinen tcp_highest_sack_combine(sk, next_skb, skb); 2618a6963a6bSIlpo Järvinen 2619fe067e8aSDavid S. Miller tcp_unlink_write_queue(next_skb, sk); 26201da177e4SLinus Torvalds 262152d570aaSJarek Poplawski if (next_skb->ip_summed == CHECKSUM_PARTIAL) 262252d570aaSJarek Poplawski skb->ip_summed = CHECKSUM_PARTIAL; 26231da177e4SLinus Torvalds 262484fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 26251da177e4SLinus Torvalds skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); 26261da177e4SLinus Torvalds 26271da177e4SLinus Torvalds /* Update sequence range on original skb. */ 26281da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 26291da177e4SLinus Torvalds 2630e6c7d085SIlpo Järvinen /* Merge over control information. This moves PSH/FIN etc. over */ 26314de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; 26321da177e4SLinus Torvalds 26331da177e4SLinus Torvalds /* All done, get rid of second SKB and account for it so 26341da177e4SLinus Torvalds * packet counting does not break. 26351da177e4SLinus Torvalds */ 26364828e7f4SIlpo Järvinen TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; 2637a643b5d4SMartin KaFai Lau TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor; 2638b7689205SIlpo Järvinen 2639b7689205SIlpo Järvinen /* changed transmit queue under us so clear hints */ 2640ef9da47cSIlpo Järvinen tcp_clear_retrans_hints_partial(tp); 2641ef9da47cSIlpo Järvinen if (next_skb == tp->retransmit_skb_hint) 2642ef9da47cSIlpo Järvinen tp->retransmit_skb_hint = skb; 2643b7689205SIlpo Järvinen 2644797108d1SIlpo Järvinen tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); 2645797108d1SIlpo Järvinen 2646082ac2d5SMartin KaFai Lau tcp_skb_collapse_tstamp(skb, next_skb); 2647082ac2d5SMartin KaFai Lau 26483ab224beSHideo Aoki sk_wmem_free_skb(sk, next_skb); 2649f8071cdeSEric Dumazet return true; 26501da177e4SLinus Torvalds } 26511da177e4SLinus Torvalds 265267edfef7SAndi Kleen /* Check if coalescing SKBs is legal. */ 2653a2a385d6SEric Dumazet static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) 26544a17fc3aSIlpo Järvinen { 26554a17fc3aSIlpo Järvinen if (tcp_skb_pcount(skb) > 1) 2656a2a385d6SEric Dumazet return false; 26574a17fc3aSIlpo Järvinen if (skb_cloned(skb)) 2658a2a385d6SEric Dumazet return false; 26594a17fc3aSIlpo Järvinen if (skb == tcp_send_head(sk)) 2660a2a385d6SEric Dumazet return false; 26612331ccc5SEric Dumazet /* Some heuristics for collapsing over SACK'd could be invented */ 26624a17fc3aSIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 2663a2a385d6SEric Dumazet return false; 26644a17fc3aSIlpo Järvinen 2665a2a385d6SEric Dumazet return true; 26664a17fc3aSIlpo Järvinen } 26674a17fc3aSIlpo Järvinen 266867edfef7SAndi Kleen /* Collapse packets in the retransmit queue to make to create 266967edfef7SAndi Kleen * less packets on the wire. This is only done on retransmission. 267067edfef7SAndi Kleen */ 26714a17fc3aSIlpo Järvinen static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, 26724a17fc3aSIlpo Järvinen int space) 26734a17fc3aSIlpo Järvinen { 26744a17fc3aSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 26754a17fc3aSIlpo Järvinen struct sk_buff *skb = to, *tmp; 2676a2a385d6SEric Dumazet bool first = true; 26774a17fc3aSIlpo Järvinen 26784a17fc3aSIlpo Järvinen if (!sysctl_tcp_retrans_collapse) 26794a17fc3aSIlpo Järvinen return; 26804de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 26814a17fc3aSIlpo Järvinen return; 26824a17fc3aSIlpo Järvinen 26834a17fc3aSIlpo Järvinen tcp_for_write_queue_from_safe(skb, tmp, sk) { 26844a17fc3aSIlpo Järvinen if (!tcp_can_collapse(sk, skb)) 26854a17fc3aSIlpo Järvinen break; 26864a17fc3aSIlpo Järvinen 2687a643b5d4SMartin KaFai Lau if (!tcp_skb_can_collapse_to(to)) 2688a643b5d4SMartin KaFai Lau break; 2689a643b5d4SMartin KaFai Lau 26904a17fc3aSIlpo Järvinen space -= skb->len; 26914a17fc3aSIlpo Järvinen 26924a17fc3aSIlpo Järvinen if (first) { 2693a2a385d6SEric Dumazet first = false; 26944a17fc3aSIlpo Järvinen continue; 26954a17fc3aSIlpo Järvinen } 26964a17fc3aSIlpo Järvinen 26974a17fc3aSIlpo Järvinen if (space < 0) 26984a17fc3aSIlpo Järvinen break; 26994a17fc3aSIlpo Järvinen 27004a17fc3aSIlpo Järvinen if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) 27014a17fc3aSIlpo Järvinen break; 27024a17fc3aSIlpo Järvinen 2703f8071cdeSEric Dumazet if (!tcp_collapse_retrans(sk, to)) 2704f8071cdeSEric Dumazet break; 27054a17fc3aSIlpo Järvinen } 27064a17fc3aSIlpo Järvinen } 27074a17fc3aSIlpo Järvinen 27081da177e4SLinus Torvalds /* This retransmits one SKB. Policy decisions and retransmit queue 27091da177e4SLinus Torvalds * state updates are done by the caller. Returns non-zero if an 27101da177e4SLinus Torvalds * error occurred which prevented the send. 27111da177e4SLinus Torvalds */ 271210d3be56SEric Dumazet int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) 27131da177e4SLinus Torvalds { 27145d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 271510d3be56SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 27167d227cd2SSridhar Samudrala unsigned int cur_mss; 271710d3be56SEric Dumazet int diff, len, err; 27181da177e4SLinus Torvalds 271910d3be56SEric Dumazet 272010d3be56SEric Dumazet /* Inconclusive MTU probe */ 272110d3be56SEric Dumazet if (icsk->icsk_mtup.probe_size) 27225d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 27235d424d5aSJohn Heffner 27241da177e4SLinus Torvalds /* Do not sent more than we queued. 1/4 is reserved for possible 2725caa20d9aSStephen Hemminger * copying overhead: fragmentation, tunneling, mangling etc. 27261da177e4SLinus Torvalds */ 27271da177e4SLinus Torvalds if (atomic_read(&sk->sk_wmem_alloc) > 2728ffb4d6c8SEric Dumazet min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), 2729ffb4d6c8SEric Dumazet sk->sk_sndbuf)) 27301da177e4SLinus Torvalds return -EAGAIN; 27311da177e4SLinus Torvalds 27321f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 27331f3279aeSEric Dumazet return -EBUSY; 27341f3279aeSEric Dumazet 27351da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 27361da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 27371da177e4SLinus Torvalds BUG(); 27381da177e4SLinus Torvalds if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 27391da177e4SLinus Torvalds return -ENOMEM; 27401da177e4SLinus Torvalds } 27411da177e4SLinus Torvalds 27427d227cd2SSridhar Samudrala if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 27437d227cd2SSridhar Samudrala return -EHOSTUNREACH; /* Routing failure or similar. */ 27447d227cd2SSridhar Samudrala 27450c54b85fSIlpo Järvinen cur_mss = tcp_current_mss(sk); 27467d227cd2SSridhar Samudrala 27471da177e4SLinus Torvalds /* If receiver has shrunk his window, and skb is out of 27481da177e4SLinus Torvalds * new window, do not retransmit it. The exception is the 27491da177e4SLinus Torvalds * case, when window is shrunk to zero. In this case 27501da177e4SLinus Torvalds * our retransmit serves as a zero window probe. 27511da177e4SLinus Torvalds */ 27529d4fb27dSJoe Perches if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && 27539d4fb27dSJoe Perches TCP_SKB_CB(skb)->seq != tp->snd_una) 27541da177e4SLinus Torvalds return -EAGAIN; 27551da177e4SLinus Torvalds 275610d3be56SEric Dumazet len = cur_mss * segs; 275710d3be56SEric Dumazet if (skb->len > len) { 275810d3be56SEric Dumazet if (tcp_fragment(sk, skb, len, cur_mss, GFP_ATOMIC)) 27591da177e4SLinus Torvalds return -ENOMEM; /* We'll try again later. */ 276002276f3cSIlpo Järvinen } else { 2761c52e2421SEric Dumazet if (skb_unclone(skb, GFP_ATOMIC)) 2762c52e2421SEric Dumazet return -ENOMEM; 276310d3be56SEric Dumazet 276410d3be56SEric Dumazet diff = tcp_skb_pcount(skb); 276510d3be56SEric Dumazet tcp_set_skb_tso_segs(skb, cur_mss); 276610d3be56SEric Dumazet diff -= tcp_skb_pcount(skb); 276710d3be56SEric Dumazet if (diff) 276810d3be56SEric Dumazet tcp_adjust_pcount(sk, skb, diff); 276910d3be56SEric Dumazet if (skb->len < cur_mss) 277010d3be56SEric Dumazet tcp_retrans_try_collapse(sk, skb, cur_mss); 27711da177e4SLinus Torvalds } 27721da177e4SLinus Torvalds 277349213555SDaniel Borkmann /* RFC3168, section 6.1.1.1. ECN fallback */ 277449213555SDaniel Borkmann if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN) 277549213555SDaniel Borkmann tcp_ecn_clear_syn(sk, skb); 277649213555SDaniel Borkmann 277750bceae9SThomas Graf /* make sure skb->data is aligned on arches that require it 277850bceae9SThomas Graf * and check if ack-trimming & collapsing extended the headroom 277950bceae9SThomas Graf * beyond what csum_start can cover. 278050bceae9SThomas Graf */ 278150bceae9SThomas Graf if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) || 278250bceae9SThomas Graf skb_headroom(skb) >= 0xFFFF)) { 278310a81980SEric Dumazet struct sk_buff *nskb; 278410a81980SEric Dumazet 278510a81980SEric Dumazet skb_mstamp_get(&skb->skb_mstamp); 278610a81980SEric Dumazet nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC); 2787c84a5711SYuchung Cheng err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) : 2788117632e6SEric Dumazet -ENOBUFS; 2789117632e6SEric Dumazet } else { 2790c84a5711SYuchung Cheng err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 2791117632e6SEric Dumazet } 2792c84a5711SYuchung Cheng 2793fc9f3501SEric Dumazet if (likely(!err)) { 279410d3be56SEric Dumazet segs = tcp_skb_pcount(skb); 279510d3be56SEric Dumazet 2796c84a5711SYuchung Cheng TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS; 2797fc9f3501SEric Dumazet /* Update global TCP statistics. */ 279810d3be56SEric Dumazet TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs); 2799fc9f3501SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 280002a1d6e7SEric Dumazet __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 280110d3be56SEric Dumazet tp->total_retrans += segs; 2802fc9f3501SEric Dumazet } 2803c84a5711SYuchung Cheng return err; 280493b174adSYuchung Cheng } 280593b174adSYuchung Cheng 280610d3be56SEric Dumazet int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) 280793b174adSYuchung Cheng { 280893b174adSYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 280910d3be56SEric Dumazet int err = __tcp_retransmit_skb(sk, skb, segs); 28101da177e4SLinus Torvalds 28111da177e4SLinus Torvalds if (err == 0) { 28121da177e4SLinus Torvalds #if FASTRETRANS_DEBUG > 0 28131da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 2814e87cc472SJoe Perches net_dbg_ratelimited("retrans_out leaked\n"); 28151da177e4SLinus Torvalds } 28161da177e4SLinus Torvalds #endif 28171da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 28181da177e4SLinus Torvalds tp->retrans_out += tcp_skb_pcount(skb); 28191da177e4SLinus Torvalds 28201da177e4SLinus Torvalds /* Save stamp of the first retransmit. */ 28211da177e4SLinus Torvalds if (!tp->retrans_stamp) 28227faee5c0SEric Dumazet tp->retrans_stamp = tcp_skb_timestamp(skb); 28231da177e4SLinus Torvalds 28241f3279aeSEric Dumazet } else if (err != -EBUSY) { 2825c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL); 28261da177e4SLinus Torvalds } 28276e08d5e3SYuchung Cheng 28286e08d5e3SYuchung Cheng if (tp->undo_retrans < 0) 28296e08d5e3SYuchung Cheng tp->undo_retrans = 0; 28306e08d5e3SYuchung Cheng tp->undo_retrans += tcp_skb_pcount(skb); 28311da177e4SLinus Torvalds return err; 28321da177e4SLinus Torvalds } 28331da177e4SLinus Torvalds 283467edfef7SAndi Kleen /* Check if we forward retransmits are possible in the current 283567edfef7SAndi Kleen * window/congestion state. 283667edfef7SAndi Kleen */ 2837a2a385d6SEric Dumazet static bool tcp_can_forward_retransmit(struct sock *sk) 2838b5afe7bcSIlpo Järvinen { 2839b5afe7bcSIlpo Järvinen const struct inet_connection_sock *icsk = inet_csk(sk); 2840cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 2841b5afe7bcSIlpo Järvinen 2842b5afe7bcSIlpo Järvinen /* Forward retransmissions are possible only during Recovery. */ 2843b5afe7bcSIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Recovery) 2844a2a385d6SEric Dumazet return false; 2845b5afe7bcSIlpo Järvinen 2846b5afe7bcSIlpo Järvinen /* No forward retransmissions in Reno are possible. */ 2847b5afe7bcSIlpo Järvinen if (tcp_is_reno(tp)) 2848a2a385d6SEric Dumazet return false; 2849b5afe7bcSIlpo Järvinen 2850b5afe7bcSIlpo Järvinen /* Yeah, we have to make difficult choice between forward transmission 2851b5afe7bcSIlpo Järvinen * and retransmission... Both ways have their merits... 2852b5afe7bcSIlpo Järvinen * 2853b5afe7bcSIlpo Järvinen * For now we do not retransmit anything, while we have some new 2854b5afe7bcSIlpo Järvinen * segments to send. In the other cases, follow rule 3 for 2855b5afe7bcSIlpo Järvinen * NextSeg() specified in RFC3517. 2856b5afe7bcSIlpo Järvinen */ 2857b5afe7bcSIlpo Järvinen 2858b5afe7bcSIlpo Järvinen if (tcp_may_send_now(sk)) 2859a2a385d6SEric Dumazet return false; 2860b5afe7bcSIlpo Järvinen 2861a2a385d6SEric Dumazet return true; 2862b5afe7bcSIlpo Järvinen } 2863b5afe7bcSIlpo Järvinen 28641da177e4SLinus Torvalds /* This gets called after a retransmit timeout, and the initially 28651da177e4SLinus Torvalds * retransmitted data is acknowledged. It tries to continue 28661da177e4SLinus Torvalds * resending the rest of the retransmit queue, until either 28671da177e4SLinus Torvalds * we've sent it all or the congestion window limit is reached. 28681da177e4SLinus Torvalds * If doing SACK, the first ACK which comes back for a timeout 28691da177e4SLinus Torvalds * based retransmit packet might feed us FACK information again. 28701da177e4SLinus Torvalds * If so, we use it to avoid unnecessarily retransmissions. 28711da177e4SLinus Torvalds */ 28721da177e4SLinus Torvalds void tcp_xmit_retransmit_queue(struct sock *sk) 28731da177e4SLinus Torvalds { 28746687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 28751da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 28761da177e4SLinus Torvalds struct sk_buff *skb; 28770e1c54c2SIlpo Järvinen struct sk_buff *hole = NULL; 2878a3d2e9f8SEric Dumazet u32 max_segs, last_lost; 287961eb55f4SIlpo Järvinen int mib_idx; 28800e1c54c2SIlpo Järvinen int fwd_rexmitting = 0; 28816a438bbeSStephen Hemminger 288245e77d31SIlpo Järvinen if (!tp->packets_out) 288345e77d31SIlpo Järvinen return; 288445e77d31SIlpo Järvinen 288508ebd172SIlpo Järvinen if (!tp->lost_out) 288608ebd172SIlpo Järvinen tp->retransmit_high = tp->snd_una; 288708ebd172SIlpo Järvinen 2888618d9f25SIlpo Järvinen if (tp->retransmit_skb_hint) { 28896a438bbeSStephen Hemminger skb = tp->retransmit_skb_hint; 2890618d9f25SIlpo Järvinen last_lost = TCP_SKB_CB(skb)->end_seq; 2891618d9f25SIlpo Järvinen if (after(last_lost, tp->retransmit_high)) 2892618d9f25SIlpo Järvinen last_lost = tp->retransmit_high; 2893618d9f25SIlpo Järvinen } else { 2894fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 2895618d9f25SIlpo Järvinen last_lost = tp->snd_una; 2896618d9f25SIlpo Järvinen } 28971da177e4SLinus Torvalds 2898ed6e7268SNeal Cardwell max_segs = tcp_tso_segs(sk, tcp_current_mss(sk)); 2899fe067e8aSDavid S. Miller tcp_for_write_queue_from(skb, sk) { 2900dca0aaf8SEric Dumazet __u8 sacked; 290110d3be56SEric Dumazet int segs; 29021da177e4SLinus Torvalds 2903fe067e8aSDavid S. Miller if (skb == tcp_send_head(sk)) 2904fe067e8aSDavid S. Miller break; 29056a438bbeSStephen Hemminger /* we could do better than to assign each time */ 290651456b29SIan Morris if (!hole) 29076a438bbeSStephen Hemminger tp->retransmit_skb_hint = skb; 29086a438bbeSStephen Hemminger 290910d3be56SEric Dumazet segs = tp->snd_cwnd - tcp_packets_in_flight(tp); 291010d3be56SEric Dumazet if (segs <= 0) 29111da177e4SLinus Torvalds return; 2912dca0aaf8SEric Dumazet sacked = TCP_SKB_CB(skb)->sacked; 2913a3d2e9f8SEric Dumazet /* In case tcp_shift_skb_data() have aggregated large skbs, 2914a3d2e9f8SEric Dumazet * we need to make sure not sending too bigs TSO packets 2915a3d2e9f8SEric Dumazet */ 2916a3d2e9f8SEric Dumazet segs = min_t(int, segs, max_segs); 29170e1c54c2SIlpo Järvinen 29180e1c54c2SIlpo Järvinen if (fwd_rexmitting) { 29190e1c54c2SIlpo Järvinen begin_fwd: 29200e1c54c2SIlpo Järvinen if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp))) 2921006f582cSIlpo Järvinen break; 29220e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFORWARDRETRANS; 29230e1c54c2SIlpo Järvinen 29240e1c54c2SIlpo Järvinen } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) { 2925618d9f25SIlpo Järvinen tp->retransmit_high = last_lost; 29260e1c54c2SIlpo Järvinen if (!tcp_can_forward_retransmit(sk)) 29270e1c54c2SIlpo Järvinen break; 29280e1c54c2SIlpo Järvinen /* Backtrack if necessary to non-L'ed skb */ 292900db4124SIan Morris if (hole) { 29300e1c54c2SIlpo Järvinen skb = hole; 29310e1c54c2SIlpo Järvinen hole = NULL; 29320e1c54c2SIlpo Järvinen } 29330e1c54c2SIlpo Järvinen fwd_rexmitting = 1; 29340e1c54c2SIlpo Järvinen goto begin_fwd; 29350e1c54c2SIlpo Järvinen 29360e1c54c2SIlpo Järvinen } else if (!(sacked & TCPCB_LOST)) { 293751456b29SIan Morris if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) 29380e1c54c2SIlpo Järvinen hole = skb; 293961eb55f4SIlpo Järvinen continue; 29401da177e4SLinus Torvalds 29410e1c54c2SIlpo Järvinen } else { 2942618d9f25SIlpo Järvinen last_lost = TCP_SKB_CB(skb)->end_seq; 29430e1c54c2SIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Loss) 29440e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFASTRETRANS; 29450e1c54c2SIlpo Järvinen else 29460e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; 29470e1c54c2SIlpo Järvinen } 29480e1c54c2SIlpo Järvinen 29490e1c54c2SIlpo Järvinen if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) 295061eb55f4SIlpo Järvinen continue; 295140b215e5SPavel Emelyanov 2952f9616c35SEric Dumazet if (tcp_small_queue_check(sk, skb, 1)) 2953f9616c35SEric Dumazet return; 2954f9616c35SEric Dumazet 295510d3be56SEric Dumazet if (tcp_retransmit_skb(sk, skb, segs)) 29561da177e4SLinus Torvalds return; 295724ab6becSYuchung Cheng 2958de1d6578SYuchung Cheng NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb)); 29591da177e4SLinus Torvalds 2960684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 2961a262f0cdSNandita Dukkipati tp->prr_out += tcp_skb_pcount(skb); 2962a262f0cdSNandita Dukkipati 2963fe067e8aSDavid S. Miller if (skb == tcp_write_queue_head(sk)) 2964463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 29653f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, 29663f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 29671da177e4SLinus Torvalds } 29681da177e4SLinus Torvalds } 29691da177e4SLinus Torvalds 2970d83769a5SEric Dumazet /* We allow to exceed memory limits for FIN packets to expedite 2971d83769a5SEric Dumazet * connection tear down and (memory) recovery. 2972845704a5SEric Dumazet * Otherwise tcp_send_fin() could be tempted to either delay FIN 2973845704a5SEric Dumazet * or even be forced to close flow without any FIN. 2974a6c5ea4cSEric Dumazet * In general, we want to allow one skb per socket to avoid hangs 2975a6c5ea4cSEric Dumazet * with edge trigger epoll() 2976d83769a5SEric Dumazet */ 2977a6c5ea4cSEric Dumazet void sk_forced_mem_schedule(struct sock *sk, int size) 2978d83769a5SEric Dumazet { 2979e805605cSJohannes Weiner int amt; 2980d83769a5SEric Dumazet 2981d83769a5SEric Dumazet if (size <= sk->sk_forward_alloc) 2982d83769a5SEric Dumazet return; 2983d83769a5SEric Dumazet amt = sk_mem_pages(size); 2984d83769a5SEric Dumazet sk->sk_forward_alloc += amt * SK_MEM_QUANTUM; 2985e805605cSJohannes Weiner sk_memory_allocated_add(sk, amt); 2986e805605cSJohannes Weiner 2987baac50bbSJohannes Weiner if (mem_cgroup_sockets_enabled && sk->sk_memcg) 2988baac50bbSJohannes Weiner mem_cgroup_charge_skmem(sk->sk_memcg, amt); 2989d83769a5SEric Dumazet } 2990d83769a5SEric Dumazet 2991845704a5SEric Dumazet /* Send a FIN. The caller locks the socket for us. 2992845704a5SEric Dumazet * We should try to send a FIN packet really hard, but eventually give up. 29931da177e4SLinus Torvalds */ 29941da177e4SLinus Torvalds void tcp_send_fin(struct sock *sk) 29951da177e4SLinus Torvalds { 2996845704a5SEric Dumazet struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk); 29971da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 29981da177e4SLinus Torvalds 2999845704a5SEric Dumazet /* Optimization, tack on the FIN if we have one skb in write queue and 3000845704a5SEric Dumazet * this skb was not yet sent, or we are under memory pressure. 3001845704a5SEric Dumazet * Note: in the latter case, FIN packet will be sent after a timeout, 3002845704a5SEric Dumazet * as TCP stack thinks it has already been transmitted. 30031da177e4SLinus Torvalds */ 3004b8da51ebSEric Dumazet if (tskb && (tcp_send_head(sk) || tcp_under_memory_pressure(sk))) { 3005845704a5SEric Dumazet coalesce: 3006845704a5SEric Dumazet TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN; 3007845704a5SEric Dumazet TCP_SKB_CB(tskb)->end_seq++; 30081da177e4SLinus Torvalds tp->write_seq++; 3009845704a5SEric Dumazet if (!tcp_send_head(sk)) { 3010845704a5SEric Dumazet /* This means tskb was already sent. 3011845704a5SEric Dumazet * Pretend we included the FIN on previous transmit. 3012845704a5SEric Dumazet * We need to set tp->snd_nxt to the value it would have 3013845704a5SEric Dumazet * if FIN had been sent. This is because retransmit path 3014845704a5SEric Dumazet * does not change tp->snd_nxt. 3015845704a5SEric Dumazet */ 3016845704a5SEric Dumazet tp->snd_nxt++; 3017845704a5SEric Dumazet return; 3018845704a5SEric Dumazet } 30191da177e4SLinus Torvalds } else { 3020845704a5SEric Dumazet skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation); 3021845704a5SEric Dumazet if (unlikely(!skb)) { 3022845704a5SEric Dumazet if (tskb) 3023845704a5SEric Dumazet goto coalesce; 3024845704a5SEric Dumazet return; 30251da177e4SLinus Torvalds } 3026d83769a5SEric Dumazet skb_reserve(skb, MAX_TCP_HEADER); 3027a6c5ea4cSEric Dumazet sk_forced_mem_schedule(sk, skb->truesize); 30281da177e4SLinus Torvalds /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 3029e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tp->write_seq, 3030a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_FIN); 30311da177e4SLinus Torvalds tcp_queue_skb(sk, skb); 30321da177e4SLinus Torvalds } 3033845704a5SEric Dumazet __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF); 30341da177e4SLinus Torvalds } 30351da177e4SLinus Torvalds 30361da177e4SLinus Torvalds /* We get here when a process closes a file descriptor (either due to 30371da177e4SLinus Torvalds * an explicit close() or as a byproduct of exit()'ing) and there 30381da177e4SLinus Torvalds * was unread data in the receive queue. This behavior is recommended 303965bb723cSGerrit Renker * by RFC 2525, section 2.17. -DaveM 30401da177e4SLinus Torvalds */ 3041dd0fc66fSAl Viro void tcp_send_active_reset(struct sock *sk, gfp_t priority) 30421da177e4SLinus Torvalds { 30431da177e4SLinus Torvalds struct sk_buff *skb; 30441da177e4SLinus Torvalds 30451da177e4SLinus Torvalds /* NOTE: No TCP options attached and we never retransmit this. */ 30461da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, priority); 30471da177e4SLinus Torvalds if (!skb) { 30484e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 30491da177e4SLinus Torvalds return; 30501da177e4SLinus Torvalds } 30511da177e4SLinus Torvalds 30521da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 30531da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 3054e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), 3055a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_RST); 3056675ee231SEric Dumazet skb_mstamp_get(&skb->skb_mstamp); 30571da177e4SLinus Torvalds /* Send it off. */ 3058dfb4b9dcSDavid S. Miller if (tcp_transmit_skb(sk, skb, 0, priority)) 30594e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 306026af65cbSSridhar Samudrala 306181cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); 30621da177e4SLinus Torvalds } 30631da177e4SLinus Torvalds 306467edfef7SAndi Kleen /* Send a crossed SYN-ACK during socket establishment. 306567edfef7SAndi Kleen * WARNING: This routine must only be called when we have already sent 30661da177e4SLinus Torvalds * a SYN packet that crossed the incoming SYN that caused this routine 30671da177e4SLinus Torvalds * to get called. If this assumption fails then the initial rcv_wnd 30681da177e4SLinus Torvalds * and rcv_wscale values will not be correct. 30691da177e4SLinus Torvalds */ 30701da177e4SLinus Torvalds int tcp_send_synack(struct sock *sk) 30711da177e4SLinus Torvalds { 30721da177e4SLinus Torvalds struct sk_buff *skb; 30731da177e4SLinus Torvalds 3074fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 307551456b29SIan Morris if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 307691df42beSJoe Perches pr_debug("%s: wrong queue state\n", __func__); 30771da177e4SLinus Torvalds return -EFAULT; 30781da177e4SLinus Torvalds } 30794de075e0SEric Dumazet if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { 30801da177e4SLinus Torvalds if (skb_cloned(skb)) { 30811da177e4SLinus Torvalds struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC); 308251456b29SIan Morris if (!nskb) 30831da177e4SLinus Torvalds return -ENOMEM; 3084fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 3085f4a775d1SEric Dumazet __skb_header_release(nskb); 3086fe067e8aSDavid S. Miller __tcp_add_write_queue_head(sk, nskb); 30873ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 30883ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 30893ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 30901da177e4SLinus Torvalds skb = nskb; 30911da177e4SLinus Torvalds } 30921da177e4SLinus Torvalds 30934de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; 3094735d3831SFlorian Westphal tcp_ecn_send_synack(sk, skb); 30951da177e4SLinus Torvalds } 3096dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 30971da177e4SLinus Torvalds } 30981da177e4SLinus Torvalds 30994aea39c1SEric Dumazet /** 31004aea39c1SEric Dumazet * tcp_make_synack - Prepare a SYN-ACK. 31014aea39c1SEric Dumazet * sk: listener socket 31024aea39c1SEric Dumazet * dst: dst entry attached to the SYNACK 31034aea39c1SEric Dumazet * req: request_sock pointer 31044aea39c1SEric Dumazet * 31054aea39c1SEric Dumazet * Allocate one skb and build a SYNACK packet. 31064aea39c1SEric Dumazet * @dst is consumed : Caller should not use it again. 31074aea39c1SEric Dumazet */ 31085d062de7SEric Dumazet struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst, 3109e6b4d113SWilliam Allen Simpson struct request_sock *req, 3110ca6fb065SEric Dumazet struct tcp_fastopen_cookie *foc, 3111b3d05147SEric Dumazet enum tcp_synack_type synack_type) 31121da177e4SLinus Torvalds { 31132e6599cbSArnaldo Carvalho de Melo struct inet_request_sock *ireq = inet_rsk(req); 31145d062de7SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 311580f03e27SEric Dumazet struct tcp_md5sig_key *md5 = NULL; 31165d062de7SEric Dumazet struct tcp_out_options opts; 31175d062de7SEric Dumazet struct sk_buff *skb; 3118bd0388aeSWilliam Allen Simpson int tcp_header_size; 31195d062de7SEric Dumazet struct tcphdr *th; 31205d062de7SEric Dumazet u16 user_mss; 3121f5fff5dcSTom Quetchenbach int mss; 31221da177e4SLinus Torvalds 3123ca6fb065SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 31244aea39c1SEric Dumazet if (unlikely(!skb)) { 31254aea39c1SEric Dumazet dst_release(dst); 31261da177e4SLinus Torvalds return NULL; 31274aea39c1SEric Dumazet } 31281da177e4SLinus Torvalds /* Reserve space for headers. */ 31291da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 31301da177e4SLinus Torvalds 3131b3d05147SEric Dumazet switch (synack_type) { 3132b3d05147SEric Dumazet case TCP_SYNACK_NORMAL: 31339e17f8a4SEric Dumazet skb_set_owner_w(skb, req_to_sk(req)); 3134b3d05147SEric Dumazet break; 3135b3d05147SEric Dumazet case TCP_SYNACK_COOKIE: 3136b3d05147SEric Dumazet /* Under synflood, we do not attach skb to a socket, 3137b3d05147SEric Dumazet * to avoid false sharing. 3138b3d05147SEric Dumazet */ 3139b3d05147SEric Dumazet break; 3140b3d05147SEric Dumazet case TCP_SYNACK_FASTOPEN: 3141ca6fb065SEric Dumazet /* sk is a const pointer, because we want to express multiple 3142ca6fb065SEric Dumazet * cpu might call us concurrently. 3143ca6fb065SEric Dumazet * sk->sk_wmem_alloc in an atomic, we can promote to rw. 3144ca6fb065SEric Dumazet */ 3145ca6fb065SEric Dumazet skb_set_owner_w(skb, (struct sock *)sk); 3146b3d05147SEric Dumazet break; 3147ca6fb065SEric Dumazet } 31484aea39c1SEric Dumazet skb_dst_set(skb, dst); 31491da177e4SLinus Torvalds 31500dbaee3bSDavid S. Miller mss = dst_metric_advmss(dst); 31515d062de7SEric Dumazet user_mss = READ_ONCE(tp->rx_opt.user_mss); 31525d062de7SEric Dumazet if (user_mss && user_mss < mss) 31535d062de7SEric Dumazet mss = user_mss; 3154f5fff5dcSTom Quetchenbach 315533ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 31568b5f12d0SFlorian Westphal #ifdef CONFIG_SYN_COOKIES 31578b5f12d0SFlorian Westphal if (unlikely(req->cookie_ts)) 31587faee5c0SEric Dumazet skb->skb_mstamp.stamp_jiffies = cookie_init_timestamp(req); 31598b5f12d0SFlorian Westphal else 31608b5f12d0SFlorian Westphal #endif 31617faee5c0SEric Dumazet skb_mstamp_get(&skb->skb_mstamp); 316280f03e27SEric Dumazet 316380f03e27SEric Dumazet #ifdef CONFIG_TCP_MD5SIG 316480f03e27SEric Dumazet rcu_read_lock(); 3165fd3a154aSEric Dumazet md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req)); 316680f03e27SEric Dumazet #endif 316758d607d3SEric Dumazet skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4); 316837bfbddaSEric Dumazet tcp_header_size = tcp_synack_options(req, mss, skb, &opts, md5, foc) + 316937bfbddaSEric Dumazet sizeof(*th); 317033ad798cSAdam Langley 3171aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 3172aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 31731da177e4SLinus Torvalds 3174ea1627c2SEric Dumazet th = (struct tcphdr *)skb->data; 31751da177e4SLinus Torvalds memset(th, 0, sizeof(struct tcphdr)); 31761da177e4SLinus Torvalds th->syn = 1; 31771da177e4SLinus Torvalds th->ack = 1; 31786ac705b1SEric Dumazet tcp_ecn_make_synack(req, th); 3179b44084c2SEric Dumazet th->source = htons(ireq->ir_num); 3180634fb979SEric Dumazet th->dest = ireq->ir_rmt_port; 3181e870a8efSIlpo Järvinen /* Setting of flags are superfluous here for callers (and ECE is 3182e870a8efSIlpo Järvinen * not even correctly set) 3183e870a8efSIlpo Järvinen */ 3184e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn, 3185a3433f35SChangli Gao TCPHDR_SYN | TCPHDR_ACK); 31864957faadSWilliam Allen Simpson 31871da177e4SLinus Torvalds th->seq = htonl(TCP_SKB_CB(skb)->seq); 31888336886fSJerry Chu /* XXX data is queued and acked as is. No buffer/window check */ 31898336886fSJerry Chu th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt); 31901da177e4SLinus Torvalds 31911da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 3192ed53d0abSEric Dumazet th->window = htons(min(req->rsk_rcv_wnd, 65535U)); 31935d062de7SEric Dumazet tcp_options_write((__be32 *)(th + 1), NULL, &opts); 31941da177e4SLinus Torvalds th->doff = (tcp_header_size >> 2); 319590bbcc60SEric Dumazet __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS); 3196cfb6eeb4SYOSHIFUJI Hideaki 3197cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 3198cfb6eeb4SYOSHIFUJI Hideaki /* Okay, we have all we need - do the md5 hash if needed */ 319980f03e27SEric Dumazet if (md5) 3200bd0388aeSWilliam Allen Simpson tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, 320139f8e58eSEric Dumazet md5, req_to_sk(req), skb); 320280f03e27SEric Dumazet rcu_read_unlock(); 3203cfb6eeb4SYOSHIFUJI Hideaki #endif 3204cfb6eeb4SYOSHIFUJI Hideaki 3205b50edd78SEric Dumazet /* Do not fool tcpdump (if any), clean our debris */ 3206b50edd78SEric Dumazet skb->tstamp.tv64 = 0; 32071da177e4SLinus Torvalds return skb; 32081da177e4SLinus Torvalds } 32094bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_make_synack); 32101da177e4SLinus Torvalds 321181164413SDaniel Borkmann static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst) 321281164413SDaniel Borkmann { 321381164413SDaniel Borkmann struct inet_connection_sock *icsk = inet_csk(sk); 321481164413SDaniel Borkmann const struct tcp_congestion_ops *ca; 321581164413SDaniel Borkmann u32 ca_key = dst_metric(dst, RTAX_CC_ALGO); 321681164413SDaniel Borkmann 321781164413SDaniel Borkmann if (ca_key == TCP_CA_UNSPEC) 321881164413SDaniel Borkmann return; 321981164413SDaniel Borkmann 322081164413SDaniel Borkmann rcu_read_lock(); 322181164413SDaniel Borkmann ca = tcp_ca_find_key(ca_key); 322281164413SDaniel Borkmann if (likely(ca && try_module_get(ca->owner))) { 322381164413SDaniel Borkmann module_put(icsk->icsk_ca_ops->owner); 322481164413SDaniel Borkmann icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst); 322581164413SDaniel Borkmann icsk->icsk_ca_ops = ca; 322681164413SDaniel Borkmann } 322781164413SDaniel Borkmann rcu_read_unlock(); 322881164413SDaniel Borkmann } 322981164413SDaniel Borkmann 323067edfef7SAndi Kleen /* Do all connect socket setups that can be done AF independent. */ 3231f7e56a76Sstephen hemminger static void tcp_connect_init(struct sock *sk) 32321da177e4SLinus Torvalds { 3233cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 32341da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 32351da177e4SLinus Torvalds __u8 rcv_wscale; 32361da177e4SLinus Torvalds 32371da177e4SLinus Torvalds /* We'll fix this up when we get a response from the other end. 32381da177e4SLinus Torvalds * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 32391da177e4SLinus Torvalds */ 32401da177e4SLinus Torvalds tp->tcp_header_len = sizeof(struct tcphdr) + 3241bb5b7c11SDavid S. Miller (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0); 32421da177e4SLinus Torvalds 3243cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 324400db4124SIan Morris if (tp->af_specific->md5_lookup(sk, sk)) 3245cfb6eeb4SYOSHIFUJI Hideaki tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 3246cfb6eeb4SYOSHIFUJI Hideaki #endif 3247cfb6eeb4SYOSHIFUJI Hideaki 32481da177e4SLinus Torvalds /* If user gave his TCP_MAXSEG, record it to clamp */ 32491da177e4SLinus Torvalds if (tp->rx_opt.user_mss) 32501da177e4SLinus Torvalds tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 32511da177e4SLinus Torvalds tp->max_window = 0; 32525d424d5aSJohn Heffner tcp_mtup_init(sk); 32531da177e4SLinus Torvalds tcp_sync_mss(sk, dst_mtu(dst)); 32541da177e4SLinus Torvalds 325581164413SDaniel Borkmann tcp_ca_dst_init(sk, dst); 325681164413SDaniel Borkmann 32571da177e4SLinus Torvalds if (!tp->window_clamp) 32581da177e4SLinus Torvalds tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 32590dbaee3bSDavid S. Miller tp->advmss = dst_metric_advmss(dst); 3260f5fff5dcSTom Quetchenbach if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss) 3261f5fff5dcSTom Quetchenbach tp->advmss = tp->rx_opt.user_mss; 3262f5fff5dcSTom Quetchenbach 32631da177e4SLinus Torvalds tcp_initialize_rcv_mss(sk); 32641da177e4SLinus Torvalds 3265e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 3266e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 3267e88c64f0SHagen Paul Pfeifer (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) 3268e88c64f0SHagen Paul Pfeifer tp->window_clamp = tcp_full_space(sk); 3269e88c64f0SHagen Paul Pfeifer 32701da177e4SLinus Torvalds tcp_select_initial_window(tcp_full_space(sk), 32711da177e4SLinus Torvalds tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 32721da177e4SLinus Torvalds &tp->rcv_wnd, 32731da177e4SLinus Torvalds &tp->window_clamp, 3274bb5b7c11SDavid S. Miller sysctl_tcp_window_scaling, 327531d12926Slaurent chavey &rcv_wscale, 327631d12926Slaurent chavey dst_metric(dst, RTAX_INITRWND)); 32771da177e4SLinus Torvalds 32781da177e4SLinus Torvalds tp->rx_opt.rcv_wscale = rcv_wscale; 32791da177e4SLinus Torvalds tp->rcv_ssthresh = tp->rcv_wnd; 32801da177e4SLinus Torvalds 32811da177e4SLinus Torvalds sk->sk_err = 0; 32821da177e4SLinus Torvalds sock_reset_flag(sk, SOCK_DONE); 32831da177e4SLinus Torvalds tp->snd_wnd = 0; 3284ee7537b6SHantzis Fotis tcp_init_wl(tp, 0); 32851da177e4SLinus Torvalds tp->snd_una = tp->write_seq; 32861da177e4SLinus Torvalds tp->snd_sml = tp->write_seq; 328733f5f57eSIlpo Järvinen tp->snd_up = tp->write_seq; 3288370816aeSPavel Emelyanov tp->snd_nxt = tp->write_seq; 3289ee995283SPavel Emelyanov 3290ee995283SPavel Emelyanov if (likely(!tp->repair)) 32911da177e4SLinus Torvalds tp->rcv_nxt = 0; 3292c7781a6eSAndrew Vagin else 3293c7781a6eSAndrew Vagin tp->rcv_tstamp = tcp_time_stamp; 3294ee995283SPavel Emelyanov tp->rcv_wup = tp->rcv_nxt; 3295ee995283SPavel Emelyanov tp->copied_seq = tp->rcv_nxt; 32961da177e4SLinus Torvalds 3297463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; 3298463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_retransmits = 0; 32991da177e4SLinus Torvalds tcp_clear_retrans(tp); 33001da177e4SLinus Torvalds } 33011da177e4SLinus Torvalds 3302783237e8SYuchung Cheng static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) 3303783237e8SYuchung Cheng { 3304783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3305783237e8SYuchung Cheng struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 3306783237e8SYuchung Cheng 3307783237e8SYuchung Cheng tcb->end_seq += skb->len; 3308f4a775d1SEric Dumazet __skb_header_release(skb); 3309783237e8SYuchung Cheng __tcp_add_write_queue_tail(sk, skb); 3310783237e8SYuchung Cheng sk->sk_wmem_queued += skb->truesize; 3311783237e8SYuchung Cheng sk_mem_charge(sk, skb->truesize); 3312783237e8SYuchung Cheng tp->write_seq = tcb->end_seq; 3313783237e8SYuchung Cheng tp->packets_out += tcp_skb_pcount(skb); 3314783237e8SYuchung Cheng } 3315783237e8SYuchung Cheng 3316783237e8SYuchung Cheng /* Build and send a SYN with data and (cached) Fast Open cookie. However, 3317783237e8SYuchung Cheng * queue a data-only packet after the regular SYN, such that regular SYNs 3318783237e8SYuchung Cheng * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges 3319783237e8SYuchung Cheng * only the SYN sequence, the data are retransmitted in the first ACK. 3320783237e8SYuchung Cheng * If cookie is not cached or other error occurs, falls back to send a 3321783237e8SYuchung Cheng * regular SYN with Fast Open cookie request option. 3322783237e8SYuchung Cheng */ 3323783237e8SYuchung Cheng static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) 3324783237e8SYuchung Cheng { 3325783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3326783237e8SYuchung Cheng struct tcp_fastopen_request *fo = tp->fastopen_req; 332707e100f9SEric Dumazet int syn_loss = 0, space, err = 0; 3328aab48743SYuchung Cheng unsigned long last_syn_loss = 0; 3329355a901eSEric Dumazet struct sk_buff *syn_data; 3330783237e8SYuchung Cheng 333167da22d2SYuchung Cheng tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ 3332aab48743SYuchung Cheng tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie, 3333aab48743SYuchung Cheng &syn_loss, &last_syn_loss); 3334aab48743SYuchung Cheng /* Recurring FO SYN losses: revert to regular handshake temporarily */ 3335aab48743SYuchung Cheng if (syn_loss > 1 && 3336aab48743SYuchung Cheng time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) { 3337aab48743SYuchung Cheng fo->cookie.len = -1; 3338aab48743SYuchung Cheng goto fallback; 3339aab48743SYuchung Cheng } 3340aab48743SYuchung Cheng 334167da22d2SYuchung Cheng if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE) 334267da22d2SYuchung Cheng fo->cookie.len = -1; 334367da22d2SYuchung Cheng else if (fo->cookie.len <= 0) 3344783237e8SYuchung Cheng goto fallback; 3345783237e8SYuchung Cheng 3346783237e8SYuchung Cheng /* MSS for SYN-data is based on cached MSS and bounded by PMTU and 3347783237e8SYuchung Cheng * user-MSS. Reserve maximum option space for middleboxes that add 3348783237e8SYuchung Cheng * private TCP options. The cost is reduced data space in SYN :( 3349783237e8SYuchung Cheng */ 3350783237e8SYuchung Cheng if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp) 3351783237e8SYuchung Cheng tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 33521b63edd6SYuchung Cheng space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) - 3353783237e8SYuchung Cheng MAX_TCP_OPTION_SPACE; 3354783237e8SYuchung Cheng 3355f5ddcbbbSEric Dumazet space = min_t(size_t, space, fo->size); 3356f5ddcbbbSEric Dumazet 3357f5ddcbbbSEric Dumazet /* limit to order-0 allocations */ 3358f5ddcbbbSEric Dumazet space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER)); 3359f5ddcbbbSEric Dumazet 3360eb934478SEric Dumazet syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false); 3361355a901eSEric Dumazet if (!syn_data) 3362783237e8SYuchung Cheng goto fallback; 3363355a901eSEric Dumazet syn_data->ip_summed = CHECKSUM_PARTIAL; 3364355a901eSEric Dumazet memcpy(syn_data->cb, syn->cb, sizeof(syn->cb)); 336507e100f9SEric Dumazet if (space) { 336607e100f9SEric Dumazet int copied = copy_from_iter(skb_put(syn_data, space), space, 336757be5bdaSAl Viro &fo->data->msg_iter); 336857be5bdaSAl Viro if (unlikely(!copied)) { 3369355a901eSEric Dumazet kfree_skb(syn_data); 3370783237e8SYuchung Cheng goto fallback; 3371783237e8SYuchung Cheng } 337257be5bdaSAl Viro if (copied != space) { 337357be5bdaSAl Viro skb_trim(syn_data, copied); 337457be5bdaSAl Viro space = copied; 337557be5bdaSAl Viro } 337607e100f9SEric Dumazet } 3377355a901eSEric Dumazet /* No more data pending in inet_wait_for_connect() */ 3378355a901eSEric Dumazet if (space == fo->size) 3379355a901eSEric Dumazet fo->data = NULL; 3380355a901eSEric Dumazet fo->copied = space; 3381783237e8SYuchung Cheng 3382355a901eSEric Dumazet tcp_connect_queue_skb(sk, syn_data); 33830f87230dSFrancis Yan if (syn_data->len) 33840f87230dSFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_BUSY); 3385355a901eSEric Dumazet 3386355a901eSEric Dumazet err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation); 3387355a901eSEric Dumazet 3388355a901eSEric Dumazet syn->skb_mstamp = syn_data->skb_mstamp; 3389355a901eSEric Dumazet 3390355a901eSEric Dumazet /* Now full SYN+DATA was cloned and sent (or not), 3391355a901eSEric Dumazet * remove the SYN from the original skb (syn_data) 3392355a901eSEric Dumazet * we keep in write queue in case of a retransmit, as we 3393355a901eSEric Dumazet * also have the SYN packet (with no data) in the same queue. 3394431a9124SEric Dumazet */ 3395355a901eSEric Dumazet TCP_SKB_CB(syn_data)->seq++; 3396355a901eSEric Dumazet TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH; 3397355a901eSEric Dumazet if (!err) { 339867da22d2SYuchung Cheng tp->syn_data = (fo->copied > 0); 3399f19c29e3SYuchung Cheng NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT); 3400783237e8SYuchung Cheng goto done; 3401783237e8SYuchung Cheng } 3402783237e8SYuchung Cheng 3403783237e8SYuchung Cheng fallback: 3404783237e8SYuchung Cheng /* Send a regular SYN with Fast Open cookie request option */ 3405783237e8SYuchung Cheng if (fo->cookie.len > 0) 3406783237e8SYuchung Cheng fo->cookie.len = 0; 3407783237e8SYuchung Cheng err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); 3408783237e8SYuchung Cheng if (err) 3409783237e8SYuchung Cheng tp->syn_fastopen = 0; 3410783237e8SYuchung Cheng done: 3411783237e8SYuchung Cheng fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ 3412783237e8SYuchung Cheng return err; 3413783237e8SYuchung Cheng } 3414783237e8SYuchung Cheng 341567edfef7SAndi Kleen /* Build a SYN and send it off. */ 34161da177e4SLinus Torvalds int tcp_connect(struct sock *sk) 34171da177e4SLinus Torvalds { 34181da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 34191da177e4SLinus Torvalds struct sk_buff *buff; 3420ee586811SEric Paris int err; 34211da177e4SLinus Torvalds 34221da177e4SLinus Torvalds tcp_connect_init(sk); 34231da177e4SLinus Torvalds 34242b916477SAndrey Vagin if (unlikely(tp->repair)) { 34252b916477SAndrey Vagin tcp_finish_connect(sk, NULL); 34262b916477SAndrey Vagin return 0; 34272b916477SAndrey Vagin } 34282b916477SAndrey Vagin 3429eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true); 3430355a901eSEric Dumazet if (unlikely(!buff)) 34311da177e4SLinus Torvalds return -ENOBUFS; 34321da177e4SLinus Torvalds 3433a3433f35SChangli Gao tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); 34347faee5c0SEric Dumazet tp->retrans_stamp = tcp_time_stamp; 3435783237e8SYuchung Cheng tcp_connect_queue_skb(sk, buff); 3436735d3831SFlorian Westphal tcp_ecn_send_syn(sk, buff); 34371da177e4SLinus Torvalds 3438783237e8SYuchung Cheng /* Send off SYN; include data in Fast Open. */ 3439783237e8SYuchung Cheng err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : 3440783237e8SYuchung Cheng tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); 3441ee586811SEric Paris if (err == -ECONNREFUSED) 3442ee586811SEric Paris return err; 3443bd37a088SWei Yongjun 3444bd37a088SWei Yongjun /* We change tp->snd_nxt after the tcp_transmit_skb() call 3445bd37a088SWei Yongjun * in order to make this packet get counted in tcpOutSegs. 3446bd37a088SWei Yongjun */ 3447bd37a088SWei Yongjun tp->snd_nxt = tp->write_seq; 3448bd37a088SWei Yongjun tp->pushed_seq = tp->write_seq; 344981cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); 34501da177e4SLinus Torvalds 34511da177e4SLinus Torvalds /* Timer for repeating the SYN until an answer. */ 34523f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 34533f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 34541da177e4SLinus Torvalds return 0; 34551da177e4SLinus Torvalds } 34564bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_connect); 34571da177e4SLinus Torvalds 34581da177e4SLinus Torvalds /* Send out a delayed ack, the caller does the policy checking 34591da177e4SLinus Torvalds * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 34601da177e4SLinus Torvalds * for details. 34611da177e4SLinus Torvalds */ 34621da177e4SLinus Torvalds void tcp_send_delayed_ack(struct sock *sk) 34631da177e4SLinus Torvalds { 3464463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 3465463c84b9SArnaldo Carvalho de Melo int ato = icsk->icsk_ack.ato; 34661da177e4SLinus Torvalds unsigned long timeout; 34671da177e4SLinus Torvalds 34689890092eSFlorian Westphal tcp_ca_event(sk, CA_EVENT_DELAYED_ACK); 34699890092eSFlorian Westphal 34701da177e4SLinus Torvalds if (ato > TCP_DELACK_MIN) { 3471463c84b9SArnaldo Carvalho de Melo const struct tcp_sock *tp = tcp_sk(sk); 34721da177e4SLinus Torvalds int max_ato = HZ / 2; 34731da177e4SLinus Torvalds 3474056834d9SIlpo Järvinen if (icsk->icsk_ack.pingpong || 3475056834d9SIlpo Järvinen (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 34761da177e4SLinus Torvalds max_ato = TCP_DELACK_MAX; 34771da177e4SLinus Torvalds 34781da177e4SLinus Torvalds /* Slow path, intersegment interval is "high". */ 34791da177e4SLinus Torvalds 34801da177e4SLinus Torvalds /* If some rtt estimate is known, use it to bound delayed ack. 3481463c84b9SArnaldo Carvalho de Melo * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 34821da177e4SLinus Torvalds * directly. 34831da177e4SLinus Torvalds */ 3484740b0f18SEric Dumazet if (tp->srtt_us) { 3485740b0f18SEric Dumazet int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3), 3486740b0f18SEric Dumazet TCP_DELACK_MIN); 34871da177e4SLinus Torvalds 34881da177e4SLinus Torvalds if (rtt < max_ato) 34891da177e4SLinus Torvalds max_ato = rtt; 34901da177e4SLinus Torvalds } 34911da177e4SLinus Torvalds 34921da177e4SLinus Torvalds ato = min(ato, max_ato); 34931da177e4SLinus Torvalds } 34941da177e4SLinus Torvalds 34951da177e4SLinus Torvalds /* Stay within the limit we were given */ 34961da177e4SLinus Torvalds timeout = jiffies + ato; 34971da177e4SLinus Torvalds 34981da177e4SLinus Torvalds /* Use new timeout only if there wasn't a older one earlier. */ 3499463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 35001da177e4SLinus Torvalds /* If delack timer was blocked or is about to expire, 35011da177e4SLinus Torvalds * send ACK now. 35021da177e4SLinus Torvalds */ 3503463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.blocked || 3504463c84b9SArnaldo Carvalho de Melo time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 35051da177e4SLinus Torvalds tcp_send_ack(sk); 35061da177e4SLinus Torvalds return; 35071da177e4SLinus Torvalds } 35081da177e4SLinus Torvalds 3509463c84b9SArnaldo Carvalho de Melo if (!time_before(timeout, icsk->icsk_ack.timeout)) 3510463c84b9SArnaldo Carvalho de Melo timeout = icsk->icsk_ack.timeout; 35111da177e4SLinus Torvalds } 3512463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 3513463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.timeout = timeout; 3514463c84b9SArnaldo Carvalho de Melo sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 35151da177e4SLinus Torvalds } 35161da177e4SLinus Torvalds 35171da177e4SLinus Torvalds /* This routine sends an ack and also updates the window. */ 35181da177e4SLinus Torvalds void tcp_send_ack(struct sock *sk) 35191da177e4SLinus Torvalds { 35201da177e4SLinus Torvalds struct sk_buff *buff; 35211da177e4SLinus Torvalds 3522058dc334SIlpo Järvinen /* If we have been reset, we may not send again. */ 3523058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3524058dc334SIlpo Järvinen return; 3525058dc334SIlpo Järvinen 35269890092eSFlorian Westphal tcp_ca_event(sk, CA_EVENT_NON_DELAYED_ACK); 35279890092eSFlorian Westphal 35281da177e4SLinus Torvalds /* We are not putting this on the write queue, so 35291da177e4SLinus Torvalds * tcp_transmit_skb() will set the ownership to this 35301da177e4SLinus Torvalds * sock. 35311da177e4SLinus Torvalds */ 35327450aaf6SEric Dumazet buff = alloc_skb(MAX_TCP_HEADER, 35337450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 35347450aaf6SEric Dumazet if (unlikely(!buff)) { 3535463c84b9SArnaldo Carvalho de Melo inet_csk_schedule_ack(sk); 3536463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; 35373f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 35383f421baaSArnaldo Carvalho de Melo TCP_DELACK_MAX, TCP_RTO_MAX); 35391da177e4SLinus Torvalds return; 35401da177e4SLinus Torvalds } 35411da177e4SLinus Torvalds 35421da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 35431da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 3544a3433f35SChangli Gao tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); 35451da177e4SLinus Torvalds 354698781965SEric Dumazet /* We do not want pure acks influencing TCP Small Queues or fq/pacing 354798781965SEric Dumazet * too much. 354898781965SEric Dumazet * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784 354998781965SEric Dumazet */ 355098781965SEric Dumazet skb_set_tcp_pure_ack(buff); 355198781965SEric Dumazet 35521da177e4SLinus Torvalds /* Send it off, this clears delayed acks for us. */ 35537faee5c0SEric Dumazet skb_mstamp_get(&buff->skb_mstamp); 35547450aaf6SEric Dumazet tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0); 35551da177e4SLinus Torvalds } 3556e3118e83SDaniel Borkmann EXPORT_SYMBOL_GPL(tcp_send_ack); 35571da177e4SLinus Torvalds 35581da177e4SLinus Torvalds /* This routine sends a packet with an out of date sequence 35591da177e4SLinus Torvalds * number. It assumes the other end will try to ack it. 35601da177e4SLinus Torvalds * 35611da177e4SLinus Torvalds * Question: what should we make while urgent mode? 35621da177e4SLinus Torvalds * 4.4BSD forces sending single byte of data. We cannot send 35631da177e4SLinus Torvalds * out of window data, because we have SND.NXT==SND.MAX... 35641da177e4SLinus Torvalds * 35651da177e4SLinus Torvalds * Current solution: to send TWO zero-length segments in urgent mode: 35661da177e4SLinus Torvalds * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 35671da177e4SLinus Torvalds * out-of-date with SND.UNA-1 to probe window. 35681da177e4SLinus Torvalds */ 3569e520af48SEric Dumazet static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib) 35701da177e4SLinus Torvalds { 35711da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 35721da177e4SLinus Torvalds struct sk_buff *skb; 35731da177e4SLinus Torvalds 35741da177e4SLinus Torvalds /* We don't queue it, tcp_transmit_skb() sets ownership. */ 35757450aaf6SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, 35767450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 357751456b29SIan Morris if (!skb) 35781da177e4SLinus Torvalds return -1; 35791da177e4SLinus Torvalds 35801da177e4SLinus Torvalds /* Reserve space for headers and set control bits. */ 35811da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 35821da177e4SLinus Torvalds /* Use a previous sequence. This should cause the other 35831da177e4SLinus Torvalds * end to send an ack. Don't queue or clone SKB, just 35841da177e4SLinus Torvalds * send it. 35851da177e4SLinus Torvalds */ 3586a3433f35SChangli Gao tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); 35877faee5c0SEric Dumazet skb_mstamp_get(&skb->skb_mstamp); 3588e2e8009fSRenato Westphal NET_INC_STATS(sock_net(sk), mib); 35897450aaf6SEric Dumazet return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0); 35901da177e4SLinus Torvalds } 35911da177e4SLinus Torvalds 3592ee995283SPavel Emelyanov void tcp_send_window_probe(struct sock *sk) 3593ee995283SPavel Emelyanov { 3594ee995283SPavel Emelyanov if (sk->sk_state == TCP_ESTABLISHED) { 3595ee995283SPavel Emelyanov tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; 3596e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE); 3597ee995283SPavel Emelyanov } 3598ee995283SPavel Emelyanov } 3599ee995283SPavel Emelyanov 360067edfef7SAndi Kleen /* Initiate keepalive or window probe from timer. */ 3601e520af48SEric Dumazet int tcp_write_wakeup(struct sock *sk, int mib) 36021da177e4SLinus Torvalds { 36031da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 36041da177e4SLinus Torvalds struct sk_buff *skb; 36051da177e4SLinus Torvalds 3606058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3607058dc334SIlpo Järvinen return -1; 3608058dc334SIlpo Järvinen 360900db4124SIan Morris skb = tcp_send_head(sk); 361000db4124SIan Morris if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { 36111da177e4SLinus Torvalds int err; 36120c54b85fSIlpo Järvinen unsigned int mss = tcp_current_mss(sk); 361390840defSIlpo Järvinen unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 36141da177e4SLinus Torvalds 36151da177e4SLinus Torvalds if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 36161da177e4SLinus Torvalds tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 36171da177e4SLinus Torvalds 36181da177e4SLinus Torvalds /* We are probing the opening of a window 36191da177e4SLinus Torvalds * but the window size is != 0 36201da177e4SLinus Torvalds * must have been a result SWS avoidance ( sender ) 36211da177e4SLinus Torvalds */ 36221da177e4SLinus Torvalds if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 36231da177e4SLinus Torvalds skb->len > mss) { 36241da177e4SLinus Torvalds seg_size = min(seg_size, mss); 36254de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 36266cc55e09SOctavian Purdila if (tcp_fragment(sk, skb, seg_size, mss, GFP_ATOMIC)) 36271da177e4SLinus Torvalds return -1; 36281da177e4SLinus Torvalds } else if (!tcp_skb_pcount(skb)) 36295bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss); 36301da177e4SLinus Torvalds 36314de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 3632dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 363366f5fe62SIlpo Järvinen if (!err) 363466f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 36351da177e4SLinus Torvalds return err; 36361da177e4SLinus Torvalds } else { 363733f5f57eSIlpo Järvinen if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) 3638e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 1, mib); 3639e520af48SEric Dumazet return tcp_xmit_probe_skb(sk, 0, mib); 36401da177e4SLinus Torvalds } 36411da177e4SLinus Torvalds } 36421da177e4SLinus Torvalds 36431da177e4SLinus Torvalds /* A window probe timeout has occurred. If window is not closed send 36441da177e4SLinus Torvalds * a partial packet else a zero probe. 36451da177e4SLinus Torvalds */ 36461da177e4SLinus Torvalds void tcp_send_probe0(struct sock *sk) 36471da177e4SLinus Torvalds { 3648463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 36491da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 3650c6214a97SNikolay Borisov struct net *net = sock_net(sk); 3651fcdd1cf4SEric Dumazet unsigned long probe_max; 36521da177e4SLinus Torvalds int err; 36531da177e4SLinus Torvalds 3654e520af48SEric Dumazet err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE); 36551da177e4SLinus Torvalds 3656fe067e8aSDavid S. Miller if (tp->packets_out || !tcp_send_head(sk)) { 36571da177e4SLinus Torvalds /* Cancel probe timer, if it is not required. */ 36586687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 0; 3659463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff = 0; 36601da177e4SLinus Torvalds return; 36611da177e4SLinus Torvalds } 36621da177e4SLinus Torvalds 36631da177e4SLinus Torvalds if (err <= 0) { 3664c6214a97SNikolay Borisov if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2) 3665463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff++; 36666687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out++; 3667fcdd1cf4SEric Dumazet probe_max = TCP_RTO_MAX; 36681da177e4SLinus Torvalds } else { 36691da177e4SLinus Torvalds /* If packet was not sent due to local congestion, 36706687e988SArnaldo Carvalho de Melo * do not backoff and do not remember icsk_probes_out. 36711da177e4SLinus Torvalds * Let local senders to fight for local resources. 36721da177e4SLinus Torvalds * 36731da177e4SLinus Torvalds * Use accumulated backoff yet. 36741da177e4SLinus Torvalds */ 36756687e988SArnaldo Carvalho de Melo if (!icsk->icsk_probes_out) 36766687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 1; 3677fcdd1cf4SEric Dumazet probe_max = TCP_RESOURCE_PROBE_INTERVAL; 36781da177e4SLinus Torvalds } 3679fcdd1cf4SEric Dumazet inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 368021c8fe99SEric Dumazet tcp_probe0_when(sk, probe_max), 3681fcdd1cf4SEric Dumazet TCP_RTO_MAX); 36821da177e4SLinus Torvalds } 36835db92c99SOctavian Purdila 3684ea3bea3aSEric Dumazet int tcp_rtx_synack(const struct sock *sk, struct request_sock *req) 36855db92c99SOctavian Purdila { 36865db92c99SOctavian Purdila const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific; 36875db92c99SOctavian Purdila struct flowi fl; 36885db92c99SOctavian Purdila int res; 36895db92c99SOctavian Purdila 369058d607d3SEric Dumazet tcp_rsk(req)->txhash = net_tx_rndhash(); 3691b3d05147SEric Dumazet res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL); 36925db92c99SOctavian Purdila if (!res) { 369390bbcc60SEric Dumazet __TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); 369402a1d6e7SEric Dumazet __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 36957e32b443SYuchung Cheng if (unlikely(tcp_passive_fastopen(sk))) 36967e32b443SYuchung Cheng tcp_sk(sk)->total_retrans++; 36975db92c99SOctavian Purdila } 36985db92c99SOctavian Purdila return res; 36995db92c99SOctavian Purdila } 37005db92c99SOctavian Purdila EXPORT_SYMBOL(tcp_rtx_synack); 3701