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 45*e086101bSCong Wang #include <trace/events/tcp.h> 46*e086101bSCong Wang 471da177e4SLinus Torvalds /* People can turn this off for buggy TCP's found in printers etc. */ 48ab32ea5dSBrian Haley int sysctl_tcp_retrans_collapse __read_mostly = 1; 491da177e4SLinus Torvalds 5015d99e02SRick Jones /* People can turn this on to work with those rare, broken TCPs that 5115d99e02SRick Jones * interpret the window field as a signed quantity. 5215d99e02SRick Jones */ 53ab32ea5dSBrian Haley int sysctl_tcp_workaround_signed_windows __read_mostly = 0; 5415d99e02SRick Jones 55c39c4c6aSWei Liu /* Default TSQ limit of four TSO segments */ 56c39c4c6aSWei Liu int sysctl_tcp_limit_output_bytes __read_mostly = 262144; 5746d3ceabSEric Dumazet 581da177e4SLinus Torvalds /* This limits the percentage of the congestion window which we 591da177e4SLinus Torvalds * will allow a single TSO frame to consume. Building TSO frames 601da177e4SLinus Torvalds * which are too large can cause TCP streams to be bursty. 611da177e4SLinus Torvalds */ 62ab32ea5dSBrian Haley int sysctl_tcp_tso_win_divisor __read_mostly = 3; 631da177e4SLinus Torvalds 6435089bb2SDavid S. Miller /* By default, RFC2861 behavior. */ 65ab32ea5dSBrian Haley int sysctl_tcp_slow_start_after_idle __read_mostly = 1; 6635089bb2SDavid S. Miller 6746d3ceabSEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 6846d3ceabSEric Dumazet int push_one, gfp_t gfp); 69519855c5SWilliam Allen Simpson 7067edfef7SAndi Kleen /* Account for new data that has been sent to the network. */ 7175c119afSEric Dumazet static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb) 726ff03ac3SIlpo Järvinen { 736ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 746ff03ac3SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 7566f5fe62SIlpo Järvinen unsigned int prior_packets = tp->packets_out; 769e412ba7SIlpo Järvinen 771da177e4SLinus Torvalds tp->snd_nxt = TCP_SKB_CB(skb)->end_seq; 788512430eSIlpo Järvinen 7975c119afSEric Dumazet __skb_unlink(skb, &sk->sk_write_queue); 8075c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, skb); 8175c119afSEric Dumazet 8266f5fe62SIlpo Järvinen tp->packets_out += tcp_skb_pcount(skb); 83bec41a11SYuchung Cheng if (!prior_packets || icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) 84750ea2baSYuchung Cheng tcp_rearm_rto(sk); 85f19c29e3SYuchung Cheng 86f7324acdSDavid S. Miller NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT, 87f19c29e3SYuchung Cheng tcp_skb_pcount(skb)); 886a5dc9e5SEric Dumazet } 891da177e4SLinus Torvalds 90a4ecb15aSCui, Cheng /* SND.NXT, if window was not shrunk or the amount of shrunk was less than one 91a4ecb15aSCui, Cheng * window scaling factor due to loss of precision. 921da177e4SLinus Torvalds * If window has been shrunk, what should we make? It is not clear at all. 931da177e4SLinus Torvalds * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-( 941da177e4SLinus Torvalds * Anything in between SND.UNA...SND.UNA+SND.WND also can be already 951da177e4SLinus Torvalds * invalid. OK, let's make this for now: 961da177e4SLinus Torvalds */ 97cf533ea5SEric Dumazet static inline __u32 tcp_acceptable_seq(const struct sock *sk) 981da177e4SLinus Torvalds { 99cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1009e412ba7SIlpo Järvinen 101a4ecb15aSCui, Cheng if (!before(tcp_wnd_end(tp), tp->snd_nxt) || 102a4ecb15aSCui, Cheng (tp->rx_opt.wscale_ok && 103a4ecb15aSCui, Cheng ((tp->snd_nxt - tcp_wnd_end(tp)) < (1 << tp->rx_opt.rcv_wscale)))) 1041da177e4SLinus Torvalds return tp->snd_nxt; 1051da177e4SLinus Torvalds else 10690840defSIlpo Järvinen return tcp_wnd_end(tp); 1071da177e4SLinus Torvalds } 1081da177e4SLinus Torvalds 1091da177e4SLinus Torvalds /* Calculate mss to advertise in SYN segment. 1101da177e4SLinus Torvalds * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that: 1111da177e4SLinus Torvalds * 1121da177e4SLinus Torvalds * 1. It is independent of path mtu. 1131da177e4SLinus Torvalds * 2. Ideally, it is maximal possible segment size i.e. 65535-40. 1141da177e4SLinus Torvalds * 3. For IPv4 it is reasonable to calculate it from maximal MTU of 1151da177e4SLinus Torvalds * attached devices, because some buggy hosts are confused by 1161da177e4SLinus Torvalds * large MSS. 1171da177e4SLinus Torvalds * 4. We do not make 3, we advertise MSS, calculated from first 1181da177e4SLinus Torvalds * hop device mtu, but allow to raise it to ip_rt_min_advmss. 1191da177e4SLinus Torvalds * This may be overridden via information stored in routing table. 1201da177e4SLinus Torvalds * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible, 1211da177e4SLinus Torvalds * probably even Jumbo". 1221da177e4SLinus Torvalds */ 1231da177e4SLinus Torvalds static __u16 tcp_advertise_mss(struct sock *sk) 1241da177e4SLinus Torvalds { 1251da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 126cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1271da177e4SLinus Torvalds int mss = tp->advmss; 1281da177e4SLinus Torvalds 1290dbaee3bSDavid S. Miller if (dst) { 1300dbaee3bSDavid S. Miller unsigned int metric = dst_metric_advmss(dst); 1310dbaee3bSDavid S. Miller 1320dbaee3bSDavid S. Miller if (metric < mss) { 1330dbaee3bSDavid S. Miller mss = metric; 1341da177e4SLinus Torvalds tp->advmss = mss; 1351da177e4SLinus Torvalds } 1360dbaee3bSDavid S. Miller } 1371da177e4SLinus Torvalds 1381da177e4SLinus Torvalds return (__u16)mss; 1391da177e4SLinus Torvalds } 1401da177e4SLinus Torvalds 1411da177e4SLinus Torvalds /* RFC2861. Reset CWND after idle period longer RTO to "restart window". 1426f021c62SEric Dumazet * This is the first part of cwnd validation mechanism. 1436f021c62SEric Dumazet */ 1446f021c62SEric Dumazet void tcp_cwnd_restart(struct sock *sk, s32 delta) 1451da177e4SLinus Torvalds { 146463c84b9SArnaldo Carvalho de Melo struct tcp_sock *tp = tcp_sk(sk); 1476f021c62SEric Dumazet u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk)); 1481da177e4SLinus Torvalds u32 cwnd = tp->snd_cwnd; 1491da177e4SLinus Torvalds 1506687e988SArnaldo Carvalho de Melo tcp_ca_event(sk, CA_EVENT_CWND_RESTART); 1511da177e4SLinus Torvalds 1526687e988SArnaldo Carvalho de Melo tp->snd_ssthresh = tcp_current_ssthresh(sk); 1531da177e4SLinus Torvalds restart_cwnd = min(restart_cwnd, cwnd); 1541da177e4SLinus Torvalds 155463c84b9SArnaldo Carvalho de Melo while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd) 1561da177e4SLinus Torvalds cwnd >>= 1; 1571da177e4SLinus Torvalds tp->snd_cwnd = max(cwnd, restart_cwnd); 158c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 1591da177e4SLinus Torvalds tp->snd_cwnd_used = 0; 1601da177e4SLinus Torvalds } 1611da177e4SLinus Torvalds 16267edfef7SAndi Kleen /* Congestion state accounting after a packet has been sent. */ 16340efc6faSStephen Hemminger static void tcp_event_data_sent(struct tcp_sock *tp, 164cf533ea5SEric Dumazet struct sock *sk) 1651da177e4SLinus Torvalds { 166463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 167d635fbe2SEric Dumazet const u32 now = tcp_jiffies32; 1681da177e4SLinus Torvalds 16905c5a46dSNeal Cardwell if (tcp_packets_in_flight(tp) == 0) 17005c5a46dSNeal Cardwell tcp_ca_event(sk, CA_EVENT_TX_START); 17105c5a46dSNeal Cardwell 1721da177e4SLinus Torvalds tp->lsndtime = now; 1731da177e4SLinus Torvalds 1741da177e4SLinus Torvalds /* If it is a reply for ato after last received 1751da177e4SLinus Torvalds * packet, enter pingpong mode. 1761da177e4SLinus Torvalds */ 1772251ae46SJon Maxwell if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato) 178463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pingpong = 1; 1791da177e4SLinus Torvalds } 1801da177e4SLinus Torvalds 18167edfef7SAndi Kleen /* Account for an ACK we sent. */ 18240efc6faSStephen Hemminger static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts) 1831da177e4SLinus Torvalds { 184463c84b9SArnaldo Carvalho de Melo tcp_dec_quickack_mode(sk, pkts); 185463c84b9SArnaldo Carvalho de Melo inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); 1861da177e4SLinus Torvalds } 1871da177e4SLinus Torvalds 18885f16525SYuchung Cheng 18985f16525SYuchung Cheng u32 tcp_default_init_rwnd(u32 mss) 19085f16525SYuchung Cheng { 19185f16525SYuchung Cheng /* Initial receive window should be twice of TCP_INIT_CWND to 1929ef71e0cSWeiping Pan * enable proper sending of new unsent data during fast recovery 19385f16525SYuchung Cheng * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a 19485f16525SYuchung Cheng * limit when mss is larger than 1460. 19585f16525SYuchung Cheng */ 19685f16525SYuchung Cheng u32 init_rwnd = TCP_INIT_CWND * 2; 19785f16525SYuchung Cheng 19885f16525SYuchung Cheng if (mss > 1460) 19985f16525SYuchung Cheng init_rwnd = max((1460 * init_rwnd) / mss, 2U); 20085f16525SYuchung Cheng return init_rwnd; 20185f16525SYuchung Cheng } 20285f16525SYuchung Cheng 2031da177e4SLinus Torvalds /* Determine a window scaling and initial window to offer. 2041da177e4SLinus Torvalds * Based on the assumption that the given amount of space 2051da177e4SLinus Torvalds * will be offered. Store the results in the tp structure. 2061da177e4SLinus Torvalds * NOTE: for smooth operation initial space offering should 2071da177e4SLinus Torvalds * be a multiple of mss if possible. We assume here that mss >= 1. 2081da177e4SLinus Torvalds * This MUST be enforced by all callers. 2091da177e4SLinus Torvalds */ 2101da177e4SLinus Torvalds void tcp_select_initial_window(int __space, __u32 mss, 2111da177e4SLinus Torvalds __u32 *rcv_wnd, __u32 *window_clamp, 21231d12926Slaurent chavey int wscale_ok, __u8 *rcv_wscale, 21331d12926Slaurent chavey __u32 init_rcv_wnd) 2141da177e4SLinus Torvalds { 2151da177e4SLinus Torvalds unsigned int space = (__space < 0 ? 0 : __space); 2161da177e4SLinus Torvalds 2171da177e4SLinus Torvalds /* If no clamp set the clamp to the max possible scaled window */ 2181da177e4SLinus Torvalds if (*window_clamp == 0) 219589c49cbSGao Feng (*window_clamp) = (U16_MAX << TCP_MAX_WSCALE); 2201da177e4SLinus Torvalds space = min(*window_clamp, space); 2211da177e4SLinus Torvalds 2221da177e4SLinus Torvalds /* Quantize space offering to a multiple of mss if possible. */ 2231da177e4SLinus Torvalds if (space > mss) 224589c49cbSGao Feng space = rounddown(space, mss); 2251da177e4SLinus Torvalds 2261da177e4SLinus Torvalds /* NOTE: offering an initial window larger than 32767 22715d99e02SRick Jones * will break some buggy TCP stacks. If the admin tells us 22815d99e02SRick Jones * it is likely we could be speaking with such a buggy stack 22915d99e02SRick Jones * we will truncate our initial window offering to 32K-1 23015d99e02SRick Jones * unless the remote has sent us a window scaling option, 23115d99e02SRick Jones * which we interpret as a sign the remote TCP is not 23215d99e02SRick Jones * misinterpreting the window field as a signed quantity. 2331da177e4SLinus Torvalds */ 23415d99e02SRick Jones if (sysctl_tcp_workaround_signed_windows) 2351da177e4SLinus Torvalds (*rcv_wnd) = min(space, MAX_TCP_WINDOW); 23615d99e02SRick Jones else 23715d99e02SRick Jones (*rcv_wnd) = space; 23815d99e02SRick Jones 2391da177e4SLinus Torvalds (*rcv_wscale) = 0; 2401da177e4SLinus Torvalds if (wscale_ok) { 241589c49cbSGao Feng /* Set window scaling on max possible window */ 242f626300aSSoheil Hassas Yeganeh space = max_t(u32, space, sysctl_tcp_rmem[2]); 243f626300aSSoheil Hassas Yeganeh space = max_t(u32, space, sysctl_rmem_max); 244316c1592SStephen Hemminger space = min_t(u32, space, *window_clamp); 245589c49cbSGao Feng while (space > U16_MAX && (*rcv_wscale) < TCP_MAX_WSCALE) { 2461da177e4SLinus Torvalds space >>= 1; 2471da177e4SLinus Torvalds (*rcv_wscale)++; 2481da177e4SLinus Torvalds } 2491da177e4SLinus Torvalds } 2501da177e4SLinus Torvalds 2511da177e4SLinus Torvalds if (mss > (1 << *rcv_wscale)) { 25285f16525SYuchung Cheng if (!init_rcv_wnd) /* Use default unless specified otherwise */ 25385f16525SYuchung Cheng init_rcv_wnd = tcp_default_init_rwnd(mss); 254b1afde60SNandita Dukkipati *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss); 2551da177e4SLinus Torvalds } 2561da177e4SLinus Torvalds 2571da177e4SLinus Torvalds /* Set the clamp no higher than max representable value */ 258589c49cbSGao Feng (*window_clamp) = min_t(__u32, U16_MAX << (*rcv_wscale), *window_clamp); 2591da177e4SLinus Torvalds } 2604bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_select_initial_window); 2611da177e4SLinus Torvalds 2621da177e4SLinus Torvalds /* Chose a new window to advertise, update state in tcp_sock for the 2631da177e4SLinus Torvalds * socket, and return result with RFC1323 scaling applied. The return 2641da177e4SLinus Torvalds * value can be stuffed directly into th->window for an outgoing 2651da177e4SLinus Torvalds * frame. 2661da177e4SLinus Torvalds */ 26740efc6faSStephen Hemminger static u16 tcp_select_window(struct sock *sk) 2681da177e4SLinus Torvalds { 2691da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2708e165e20SFlorian Westphal u32 old_win = tp->rcv_wnd; 2711da177e4SLinus Torvalds u32 cur_win = tcp_receive_window(tp); 2721da177e4SLinus Torvalds u32 new_win = __tcp_select_window(sk); 2731da177e4SLinus Torvalds 2741da177e4SLinus Torvalds /* Never shrink the offered window */ 2751da177e4SLinus Torvalds if (new_win < cur_win) { 2761da177e4SLinus Torvalds /* Danger Will Robinson! 2771da177e4SLinus Torvalds * Don't update rcv_wup/rcv_wnd here or else 2781da177e4SLinus Torvalds * we will not be able to advertise a zero 2791da177e4SLinus Torvalds * window in time. --DaveM 2801da177e4SLinus Torvalds * 2811da177e4SLinus Torvalds * Relax Will Robinson. 2821da177e4SLinus Torvalds */ 2838e165e20SFlorian Westphal if (new_win == 0) 2848e165e20SFlorian Westphal NET_INC_STATS(sock_net(sk), 2858e165e20SFlorian Westphal LINUX_MIB_TCPWANTZEROWINDOWADV); 286607bfbf2SPatrick McHardy new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale); 2871da177e4SLinus Torvalds } 2881da177e4SLinus Torvalds tp->rcv_wnd = new_win; 2891da177e4SLinus Torvalds tp->rcv_wup = tp->rcv_nxt; 2901da177e4SLinus Torvalds 2911da177e4SLinus Torvalds /* Make sure we do not exceed the maximum possible 2921da177e4SLinus Torvalds * scaled window. 2931da177e4SLinus Torvalds */ 29415d99e02SRick Jones if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows) 2951da177e4SLinus Torvalds new_win = min(new_win, MAX_TCP_WINDOW); 2961da177e4SLinus Torvalds else 2971da177e4SLinus Torvalds new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); 2981da177e4SLinus Torvalds 2991da177e4SLinus Torvalds /* RFC1323 scaling applied */ 3001da177e4SLinus Torvalds new_win >>= tp->rx_opt.rcv_wscale; 3011da177e4SLinus Torvalds 30231770e34SFlorian Westphal /* If we advertise zero window, disable fast path. */ 3038e165e20SFlorian Westphal if (new_win == 0) { 30431770e34SFlorian Westphal tp->pred_flags = 0; 3058e165e20SFlorian Westphal if (old_win) 3068e165e20SFlorian Westphal NET_INC_STATS(sock_net(sk), 3078e165e20SFlorian Westphal LINUX_MIB_TCPTOZEROWINDOWADV); 3088e165e20SFlorian Westphal } else if (old_win == 0) { 3098e165e20SFlorian Westphal NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV); 3108e165e20SFlorian Westphal } 3111da177e4SLinus Torvalds 3121da177e4SLinus Torvalds return new_win; 3131da177e4SLinus Torvalds } 3141da177e4SLinus Torvalds 31567edfef7SAndi Kleen /* Packet ECN state for a SYN-ACK */ 316735d3831SFlorian Westphal static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb) 317bdf1ee5dSIlpo Järvinen { 31830e502a3SDaniel Borkmann const struct tcp_sock *tp = tcp_sk(sk); 31930e502a3SDaniel Borkmann 3204de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR; 321bdf1ee5dSIlpo Järvinen if (!(tp->ecn_flags & TCP_ECN_OK)) 3224de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE; 32391b5b21cSLawrence Brakmo else if (tcp_ca_needs_ecn(sk) || 32491b5b21cSLawrence Brakmo tcp_bpf_ca_needs_ecn(sk)) 32530e502a3SDaniel Borkmann INET_ECN_xmit(sk); 326bdf1ee5dSIlpo Järvinen } 327bdf1ee5dSIlpo Järvinen 32867edfef7SAndi Kleen /* Packet ECN state for a SYN. */ 329735d3831SFlorian Westphal static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb) 330bdf1ee5dSIlpo Järvinen { 331bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 33291b5b21cSLawrence Brakmo bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk); 333f7b3bec6SFlorian Westphal bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 || 33491b5b21cSLawrence Brakmo tcp_ca_needs_ecn(sk) || bpf_needs_ecn; 335f7b3bec6SFlorian Westphal 336f7b3bec6SFlorian Westphal if (!use_ecn) { 337f7b3bec6SFlorian Westphal const struct dst_entry *dst = __sk_dst_get(sk); 338f7b3bec6SFlorian Westphal 339f7b3bec6SFlorian Westphal if (dst && dst_feature(dst, RTAX_FEATURE_ECN)) 340f7b3bec6SFlorian Westphal use_ecn = true; 341f7b3bec6SFlorian Westphal } 342bdf1ee5dSIlpo Järvinen 343bdf1ee5dSIlpo Järvinen tp->ecn_flags = 0; 344f7b3bec6SFlorian Westphal 345f7b3bec6SFlorian Westphal if (use_ecn) { 3464de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR; 347bdf1ee5dSIlpo Järvinen tp->ecn_flags = TCP_ECN_OK; 34891b5b21cSLawrence Brakmo if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn) 34930e502a3SDaniel Borkmann INET_ECN_xmit(sk); 350bdf1ee5dSIlpo Järvinen } 351bdf1ee5dSIlpo Järvinen } 352bdf1ee5dSIlpo Järvinen 35349213555SDaniel Borkmann static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb) 35449213555SDaniel Borkmann { 35549213555SDaniel Borkmann if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback) 35649213555SDaniel Borkmann /* tp->ecn_flags are cleared at a later point in time when 35749213555SDaniel Borkmann * SYN ACK is ultimatively being received. 35849213555SDaniel Borkmann */ 35949213555SDaniel Borkmann TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR); 36049213555SDaniel Borkmann } 36149213555SDaniel Borkmann 362735d3831SFlorian Westphal static void 3636ac705b1SEric Dumazet tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th) 364bdf1ee5dSIlpo Järvinen { 3656ac705b1SEric Dumazet if (inet_rsk(req)->ecn_ok) 366bdf1ee5dSIlpo Järvinen th->ece = 1; 367bdf1ee5dSIlpo Järvinen } 368bdf1ee5dSIlpo Järvinen 36967edfef7SAndi Kleen /* Set up ECN state for a packet on a ESTABLISHED socket that is about to 37067edfef7SAndi Kleen * be sent. 37167edfef7SAndi Kleen */ 372735d3831SFlorian Westphal static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb, 373ea1627c2SEric Dumazet struct tcphdr *th, int tcp_header_len) 374bdf1ee5dSIlpo Järvinen { 375bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 376bdf1ee5dSIlpo Järvinen 377bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_OK) { 378bdf1ee5dSIlpo Järvinen /* Not-retransmitted data segment: set ECT and inject CWR. */ 379bdf1ee5dSIlpo Järvinen if (skb->len != tcp_header_len && 380bdf1ee5dSIlpo Järvinen !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { 381bdf1ee5dSIlpo Järvinen INET_ECN_xmit(sk); 382bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) { 383bdf1ee5dSIlpo Järvinen tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; 384ea1627c2SEric Dumazet th->cwr = 1; 385bdf1ee5dSIlpo Järvinen skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; 386bdf1ee5dSIlpo Järvinen } 38730e502a3SDaniel Borkmann } else if (!tcp_ca_needs_ecn(sk)) { 388bdf1ee5dSIlpo Järvinen /* ACK or retransmitted segment: clear ECT|CE */ 389bdf1ee5dSIlpo Järvinen INET_ECN_dontxmit(sk); 390bdf1ee5dSIlpo Järvinen } 391bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) 392ea1627c2SEric Dumazet th->ece = 1; 393bdf1ee5dSIlpo Järvinen } 394bdf1ee5dSIlpo Järvinen } 395bdf1ee5dSIlpo Järvinen 396e870a8efSIlpo Järvinen /* Constructs common control bits of non-data skb. If SYN/FIN is present, 397e870a8efSIlpo Järvinen * auto increment end seqno. 398e870a8efSIlpo Järvinen */ 399e870a8efSIlpo Järvinen static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags) 400e870a8efSIlpo Järvinen { 4012e8e18efSDavid S. Miller skb->ip_summed = CHECKSUM_PARTIAL; 402e870a8efSIlpo Järvinen skb->csum = 0; 403e870a8efSIlpo Järvinen 4044de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags; 405e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->sacked = 0; 406e870a8efSIlpo Järvinen 407cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, 1); 408e870a8efSIlpo Järvinen 409e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->seq = seq; 410a3433f35SChangli Gao if (flags & (TCPHDR_SYN | TCPHDR_FIN)) 411e870a8efSIlpo Järvinen seq++; 412e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->end_seq = seq; 413e870a8efSIlpo Järvinen } 414e870a8efSIlpo Järvinen 415a2a385d6SEric Dumazet static inline bool tcp_urg_mode(const struct tcp_sock *tp) 41633f5f57eSIlpo Järvinen { 41733f5f57eSIlpo Järvinen return tp->snd_una != tp->snd_up; 41833f5f57eSIlpo Järvinen } 41933f5f57eSIlpo Järvinen 42033ad798cSAdam Langley #define OPTION_SACK_ADVERTISE (1 << 0) 42133ad798cSAdam Langley #define OPTION_TS (1 << 1) 42233ad798cSAdam Langley #define OPTION_MD5 (1 << 2) 42389e95a61SOri Finkelman #define OPTION_WSCALE (1 << 3) 4242100c8d2SYuchung Cheng #define OPTION_FAST_OPEN_COOKIE (1 << 8) 42533ad798cSAdam Langley 42633ad798cSAdam Langley struct tcp_out_options { 4272100c8d2SYuchung Cheng u16 options; /* bit field of OPTION_* */ 4282100c8d2SYuchung Cheng u16 mss; /* 0 to disable */ 42933ad798cSAdam Langley u8 ws; /* window scale, 0 to disable */ 43033ad798cSAdam Langley u8 num_sack_blocks; /* number of SACK blocks to include */ 431bd0388aeSWilliam Allen Simpson u8 hash_size; /* bytes in hash_location */ 432bd0388aeSWilliam Allen Simpson __u8 *hash_location; /* temporary pointer, overloaded */ 4332100c8d2SYuchung Cheng __u32 tsval, tsecr; /* need to include OPTION_TS */ 4342100c8d2SYuchung Cheng struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */ 43533ad798cSAdam Langley }; 43633ad798cSAdam Langley 43767edfef7SAndi Kleen /* Write previously computed TCP options to the packet. 43867edfef7SAndi Kleen * 43967edfef7SAndi Kleen * Beware: Something in the Internet is very sensitive to the ordering of 440fd6149d3SIlpo Järvinen * TCP options, we learned this through the hard way, so be careful here. 441fd6149d3SIlpo Järvinen * Luckily we can at least blame others for their non-compliance but from 4428e3bff96Sstephen hemminger * inter-operability perspective it seems that we're somewhat stuck with 443fd6149d3SIlpo Järvinen * the ordering which we have been using if we want to keep working with 444fd6149d3SIlpo Järvinen * those broken things (not that it currently hurts anybody as there isn't 445fd6149d3SIlpo Järvinen * particular reason why the ordering would need to be changed). 446fd6149d3SIlpo Järvinen * 447fd6149d3SIlpo Järvinen * At least SACK_PERM as the first option is known to lead to a disaster 448fd6149d3SIlpo Järvinen * (but it may well be that other scenarios fail similarly). 449fd6149d3SIlpo Järvinen */ 45033ad798cSAdam Langley static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp, 451bd0388aeSWilliam Allen Simpson struct tcp_out_options *opts) 452bd0388aeSWilliam Allen Simpson { 4532100c8d2SYuchung Cheng u16 options = opts->options; /* mungable copy */ 454bd0388aeSWilliam Allen Simpson 455bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_MD5 & options)) { 4561a2c6181SChristoph Paasch *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | 4571a2c6181SChristoph Paasch (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG); 458bd0388aeSWilliam Allen Simpson /* overload cookie hash location */ 459bd0388aeSWilliam Allen Simpson opts->hash_location = (__u8 *)ptr; 46033ad798cSAdam Langley ptr += 4; 46133ad798cSAdam Langley } 46233ad798cSAdam Langley 463fd6149d3SIlpo Järvinen if (unlikely(opts->mss)) { 464fd6149d3SIlpo Järvinen *ptr++ = htonl((TCPOPT_MSS << 24) | 465fd6149d3SIlpo Järvinen (TCPOLEN_MSS << 16) | 466fd6149d3SIlpo Järvinen opts->mss); 467fd6149d3SIlpo Järvinen } 468fd6149d3SIlpo Järvinen 469bd0388aeSWilliam Allen Simpson if (likely(OPTION_TS & options)) { 470bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 47133ad798cSAdam Langley *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | 47233ad798cSAdam Langley (TCPOLEN_SACK_PERM << 16) | 47333ad798cSAdam Langley (TCPOPT_TIMESTAMP << 8) | 47433ad798cSAdam Langley TCPOLEN_TIMESTAMP); 475bd0388aeSWilliam Allen Simpson options &= ~OPTION_SACK_ADVERTISE; 47633ad798cSAdam Langley } else { 477496c98dfSYOSHIFUJI Hideaki *ptr++ = htonl((TCPOPT_NOP << 24) | 47840efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 47940efc6faSStephen Hemminger (TCPOPT_TIMESTAMP << 8) | 48040efc6faSStephen Hemminger TCPOLEN_TIMESTAMP); 48140efc6faSStephen Hemminger } 48233ad798cSAdam Langley *ptr++ = htonl(opts->tsval); 48333ad798cSAdam Langley *ptr++ = htonl(opts->tsecr); 48433ad798cSAdam Langley } 48533ad798cSAdam Langley 486bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 48733ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 48833ad798cSAdam Langley (TCPOPT_NOP << 16) | 48933ad798cSAdam Langley (TCPOPT_SACK_PERM << 8) | 49033ad798cSAdam Langley TCPOLEN_SACK_PERM); 49133ad798cSAdam Langley } 49233ad798cSAdam Langley 493bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_WSCALE & options)) { 49433ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 49533ad798cSAdam Langley (TCPOPT_WINDOW << 16) | 49633ad798cSAdam Langley (TCPOLEN_WINDOW << 8) | 49733ad798cSAdam Langley opts->ws); 49833ad798cSAdam Langley } 49933ad798cSAdam Langley 50033ad798cSAdam Langley if (unlikely(opts->num_sack_blocks)) { 50133ad798cSAdam Langley struct tcp_sack_block *sp = tp->rx_opt.dsack ? 50233ad798cSAdam Langley tp->duplicate_sack : tp->selective_acks; 50340efc6faSStephen Hemminger int this_sack; 50440efc6faSStephen Hemminger 50540efc6faSStephen Hemminger *ptr++ = htonl((TCPOPT_NOP << 24) | 50640efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 50740efc6faSStephen Hemminger (TCPOPT_SACK << 8) | 50833ad798cSAdam Langley (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * 50940efc6faSStephen Hemminger TCPOLEN_SACK_PERBLOCK))); 5102de979bdSStephen Hemminger 51133ad798cSAdam Langley for (this_sack = 0; this_sack < opts->num_sack_blocks; 51233ad798cSAdam Langley ++this_sack) { 51340efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].start_seq); 51440efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].end_seq); 51540efc6faSStephen Hemminger } 5162de979bdSStephen Hemminger 51740efc6faSStephen Hemminger tp->rx_opt.dsack = 0; 51840efc6faSStephen Hemminger } 5192100c8d2SYuchung Cheng 5202100c8d2SYuchung Cheng if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) { 5212100c8d2SYuchung Cheng struct tcp_fastopen_cookie *foc = opts->fastopen_cookie; 5227f9b838bSDaniel Lee u8 *p = (u8 *)ptr; 5237f9b838bSDaniel Lee u32 len; /* Fast Open option length */ 5242100c8d2SYuchung Cheng 5257f9b838bSDaniel Lee if (foc->exp) { 5267f9b838bSDaniel Lee len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len; 5277f9b838bSDaniel Lee *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) | 5282100c8d2SYuchung Cheng TCPOPT_FASTOPEN_MAGIC); 5297f9b838bSDaniel Lee p += TCPOLEN_EXP_FASTOPEN_BASE; 5307f9b838bSDaniel Lee } else { 5317f9b838bSDaniel Lee len = TCPOLEN_FASTOPEN_BASE + foc->len; 5327f9b838bSDaniel Lee *p++ = TCPOPT_FASTOPEN; 5337f9b838bSDaniel Lee *p++ = len; 5342100c8d2SYuchung Cheng } 5357f9b838bSDaniel Lee 5367f9b838bSDaniel Lee memcpy(p, foc->val, foc->len); 5377f9b838bSDaniel Lee if ((len & 3) == 2) { 5387f9b838bSDaniel Lee p[foc->len] = TCPOPT_NOP; 5397f9b838bSDaniel Lee p[foc->len + 1] = TCPOPT_NOP; 5407f9b838bSDaniel Lee } 5417f9b838bSDaniel Lee ptr += (len + 3) >> 2; 5422100c8d2SYuchung Cheng } 54340efc6faSStephen Hemminger } 54440efc6faSStephen Hemminger 54567edfef7SAndi Kleen /* Compute TCP options for SYN packets. This is not the final 54667edfef7SAndi Kleen * network wire format yet. 54767edfef7SAndi Kleen */ 54895c96174SEric Dumazet static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb, 54933ad798cSAdam Langley struct tcp_out_options *opts, 550cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 551cf533ea5SEric Dumazet { 55233ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 55395c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 554783237e8SYuchung Cheng struct tcp_fastopen_request *fastopen = tp->fastopen_req; 55533ad798cSAdam Langley 556cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 55733ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 55833ad798cSAdam Langley if (*md5) { 55933ad798cSAdam Langley opts->options |= OPTION_MD5; 560bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 561cfb6eeb4SYOSHIFUJI Hideaki } 56233ad798cSAdam Langley #else 56333ad798cSAdam Langley *md5 = NULL; 564cfb6eeb4SYOSHIFUJI Hideaki #endif 56533ad798cSAdam Langley 56633ad798cSAdam Langley /* We always get an MSS option. The option bytes which will be seen in 56733ad798cSAdam Langley * normal data packets should timestamps be used, must be in the MSS 56833ad798cSAdam Langley * advertised. But we subtract them from tp->mss_cache so that 56933ad798cSAdam Langley * calculations in tcp_sendmsg are simpler etc. So account for this 57033ad798cSAdam Langley * fact here if necessary. If we don't do this correctly, as a 57133ad798cSAdam Langley * receiver we won't recognize data packets as being full sized when we 57233ad798cSAdam Langley * should, and thus we won't abide by the delayed ACK rules correctly. 57333ad798cSAdam Langley * SACKs don't matter, we never delay an ACK when we have any of those 57433ad798cSAdam Langley * going out. */ 57533ad798cSAdam Langley opts->mss = tcp_advertise_mss(sk); 576bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 57733ad798cSAdam Langley 5785d2ed052SEric Dumazet if (likely(sock_net(sk)->ipv4.sysctl_tcp_timestamps && !*md5)) { 57933ad798cSAdam Langley opts->options |= OPTION_TS; 5807faee5c0SEric Dumazet opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset; 58133ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 582bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 58333ad798cSAdam Langley } 5849bb37ef0SEric Dumazet if (likely(sock_net(sk)->ipv4.sysctl_tcp_window_scaling)) { 58533ad798cSAdam Langley opts->ws = tp->rx_opt.rcv_wscale; 58689e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 587bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 58833ad798cSAdam Langley } 589f9301034SEric Dumazet if (likely(sock_net(sk)->ipv4.sysctl_tcp_sack)) { 59033ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 591b32d1310SDavid S. Miller if (unlikely(!(OPTION_TS & opts->options))) 592bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 59333ad798cSAdam Langley } 59433ad798cSAdam Langley 595783237e8SYuchung Cheng if (fastopen && fastopen->cookie.len >= 0) { 5962646c831SDaniel Lee u32 need = fastopen->cookie.len; 5972646c831SDaniel Lee 5982646c831SDaniel Lee need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE : 5992646c831SDaniel Lee TCPOLEN_FASTOPEN_BASE; 600783237e8SYuchung Cheng need = (need + 3) & ~3U; /* Align to 32 bits */ 601783237e8SYuchung Cheng if (remaining >= need) { 602783237e8SYuchung Cheng opts->options |= OPTION_FAST_OPEN_COOKIE; 603783237e8SYuchung Cheng opts->fastopen_cookie = &fastopen->cookie; 604783237e8SYuchung Cheng remaining -= need; 605783237e8SYuchung Cheng tp->syn_fastopen = 1; 6062646c831SDaniel Lee tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0; 607783237e8SYuchung Cheng } 608783237e8SYuchung Cheng } 609bd0388aeSWilliam Allen Simpson 610bd0388aeSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 61133ad798cSAdam Langley } 61233ad798cSAdam Langley 61367edfef7SAndi Kleen /* Set up TCP options for SYN-ACKs. */ 61437bfbddaSEric Dumazet static unsigned int tcp_synack_options(struct request_sock *req, 61595c96174SEric Dumazet unsigned int mss, struct sk_buff *skb, 61633ad798cSAdam Langley struct tcp_out_options *opts, 61780f03e27SEric Dumazet const struct tcp_md5sig_key *md5, 6188336886fSJerry Chu struct tcp_fastopen_cookie *foc) 6194957faadSWilliam Allen Simpson { 62033ad798cSAdam Langley struct inet_request_sock *ireq = inet_rsk(req); 62195c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 62233ad798cSAdam Langley 62333ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 62480f03e27SEric Dumazet if (md5) { 62533ad798cSAdam Langley opts->options |= OPTION_MD5; 6264957faadSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 6274957faadSWilliam Allen Simpson 6284957faadSWilliam Allen Simpson /* We can't fit any SACK blocks in a packet with MD5 + TS 6294957faadSWilliam Allen Simpson * options. There was discussion about disabling SACK 6304957faadSWilliam Allen Simpson * rather than TS in order to fit in better with old, 6314957faadSWilliam Allen Simpson * buggy kernels, but that was deemed to be unnecessary. 6324957faadSWilliam Allen Simpson */ 633de213e5eSEric Dumazet ireq->tstamp_ok &= !ireq->sack_ok; 63433ad798cSAdam Langley } 63533ad798cSAdam Langley #endif 63633ad798cSAdam Langley 6374957faadSWilliam Allen Simpson /* We always send an MSS option. */ 63833ad798cSAdam Langley opts->mss = mss; 6394957faadSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 64033ad798cSAdam Langley 64133ad798cSAdam Langley if (likely(ireq->wscale_ok)) { 64233ad798cSAdam Langley opts->ws = ireq->rcv_wscale; 64389e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 6444957faadSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 64533ad798cSAdam Langley } 646de213e5eSEric Dumazet if (likely(ireq->tstamp_ok)) { 64733ad798cSAdam Langley opts->options |= OPTION_TS; 64895a22caeSFlorian Westphal opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off; 64933ad798cSAdam Langley opts->tsecr = req->ts_recent; 6504957faadSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 65133ad798cSAdam Langley } 65233ad798cSAdam Langley if (likely(ireq->sack_ok)) { 65333ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 654de213e5eSEric Dumazet if (unlikely(!ireq->tstamp_ok)) 6554957faadSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 65633ad798cSAdam Langley } 6577f9b838bSDaniel Lee if (foc != NULL && foc->len >= 0) { 6587f9b838bSDaniel Lee u32 need = foc->len; 6597f9b838bSDaniel Lee 6607f9b838bSDaniel Lee need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE : 6617f9b838bSDaniel Lee TCPOLEN_FASTOPEN_BASE; 6628336886fSJerry Chu need = (need + 3) & ~3U; /* Align to 32 bits */ 6638336886fSJerry Chu if (remaining >= need) { 6648336886fSJerry Chu opts->options |= OPTION_FAST_OPEN_COOKIE; 6658336886fSJerry Chu opts->fastopen_cookie = foc; 6668336886fSJerry Chu remaining -= need; 6678336886fSJerry Chu } 6688336886fSJerry Chu } 6694957faadSWilliam Allen Simpson 6704957faadSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 67133ad798cSAdam Langley } 67233ad798cSAdam Langley 67367edfef7SAndi Kleen /* Compute TCP options for ESTABLISHED sockets. This is not the 67467edfef7SAndi Kleen * final wire format yet. 67567edfef7SAndi Kleen */ 67695c96174SEric Dumazet static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb, 67733ad798cSAdam Langley struct tcp_out_options *opts, 678cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 679cf533ea5SEric Dumazet { 68033ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 68195c96174SEric Dumazet unsigned int size = 0; 682cabeccbdSIlpo Järvinen unsigned int eff_sacks; 68333ad798cSAdam Langley 6845843ef42SAndi Kleen opts->options = 0; 6855843ef42SAndi Kleen 68633ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 68733ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 68833ad798cSAdam Langley if (unlikely(*md5)) { 68933ad798cSAdam Langley opts->options |= OPTION_MD5; 69033ad798cSAdam Langley size += TCPOLEN_MD5SIG_ALIGNED; 69133ad798cSAdam Langley } 69233ad798cSAdam Langley #else 69333ad798cSAdam Langley *md5 = NULL; 69433ad798cSAdam Langley #endif 69533ad798cSAdam Langley 69633ad798cSAdam Langley if (likely(tp->rx_opt.tstamp_ok)) { 69733ad798cSAdam Langley opts->options |= OPTION_TS; 6987faee5c0SEric Dumazet opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0; 69933ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 70033ad798cSAdam Langley size += TCPOLEN_TSTAMP_ALIGNED; 70133ad798cSAdam Langley } 70233ad798cSAdam Langley 703cabeccbdSIlpo Järvinen eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; 704cabeccbdSIlpo Järvinen if (unlikely(eff_sacks)) { 70595c96174SEric Dumazet const unsigned int remaining = MAX_TCP_OPTION_SPACE - size; 70633ad798cSAdam Langley opts->num_sack_blocks = 70795c96174SEric Dumazet min_t(unsigned int, eff_sacks, 70833ad798cSAdam Langley (remaining - TCPOLEN_SACK_BASE_ALIGNED) / 70933ad798cSAdam Langley TCPOLEN_SACK_PERBLOCK); 71033ad798cSAdam Langley size += TCPOLEN_SACK_BASE_ALIGNED + 71133ad798cSAdam Langley opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; 71233ad798cSAdam Langley } 71333ad798cSAdam Langley 71433ad798cSAdam Langley return size; 71540efc6faSStephen Hemminger } 7161da177e4SLinus Torvalds 71746d3ceabSEric Dumazet 71846d3ceabSEric Dumazet /* TCP SMALL QUEUES (TSQ) 71946d3ceabSEric Dumazet * 72046d3ceabSEric Dumazet * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev) 72146d3ceabSEric Dumazet * to reduce RTT and bufferbloat. 72246d3ceabSEric Dumazet * We do this using a special skb destructor (tcp_wfree). 72346d3ceabSEric Dumazet * 72446d3ceabSEric Dumazet * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb 72546d3ceabSEric Dumazet * needs to be reallocated in a driver. 7268e3bff96Sstephen hemminger * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc 72746d3ceabSEric Dumazet * 72846d3ceabSEric Dumazet * Since transmit from skb destructor is forbidden, we use a tasklet 72946d3ceabSEric Dumazet * to process all sockets that eventually need to send more skbs. 73046d3ceabSEric Dumazet * We use one tasklet per cpu, with its own queue of sockets. 73146d3ceabSEric Dumazet */ 73246d3ceabSEric Dumazet struct tsq_tasklet { 73346d3ceabSEric Dumazet struct tasklet_struct tasklet; 73446d3ceabSEric Dumazet struct list_head head; /* queue of tcp sockets */ 73546d3ceabSEric Dumazet }; 73646d3ceabSEric Dumazet static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet); 73746d3ceabSEric Dumazet 7386f458dfbSEric Dumazet static void tcp_tsq_handler(struct sock *sk) 7396f458dfbSEric Dumazet { 7406f458dfbSEric Dumazet if ((1 << sk->sk_state) & 7416f458dfbSEric Dumazet (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING | 742f9616c35SEric Dumazet TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) { 743f9616c35SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 744f9616c35SEric Dumazet 745f9616c35SEric Dumazet if (tp->lost_out > tp->retrans_out && 746f9616c35SEric Dumazet tp->snd_cwnd > tcp_packets_in_flight(tp)) 747f9616c35SEric Dumazet tcp_xmit_retransmit_queue(sk); 748f9616c35SEric Dumazet 749f9616c35SEric Dumazet tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle, 750bf06200eSJohn Ogness 0, GFP_ATOMIC); 7516f458dfbSEric Dumazet } 752f9616c35SEric Dumazet } 75346d3ceabSEric Dumazet /* 7548e3bff96Sstephen hemminger * One tasklet per cpu tries to send more skbs. 75546d3ceabSEric Dumazet * We run in tasklet context but need to disable irqs when 7568e3bff96Sstephen hemminger * transferring tsq->head because tcp_wfree() might 75746d3ceabSEric Dumazet * interrupt us (non NAPI drivers) 75846d3ceabSEric Dumazet */ 75946d3ceabSEric Dumazet static void tcp_tasklet_func(unsigned long data) 76046d3ceabSEric Dumazet { 76146d3ceabSEric Dumazet struct tsq_tasklet *tsq = (struct tsq_tasklet *)data; 76246d3ceabSEric Dumazet LIST_HEAD(list); 76346d3ceabSEric Dumazet unsigned long flags; 76446d3ceabSEric Dumazet struct list_head *q, *n; 76546d3ceabSEric Dumazet struct tcp_sock *tp; 76646d3ceabSEric Dumazet struct sock *sk; 76746d3ceabSEric Dumazet 76846d3ceabSEric Dumazet local_irq_save(flags); 76946d3ceabSEric Dumazet list_splice_init(&tsq->head, &list); 77046d3ceabSEric Dumazet local_irq_restore(flags); 77146d3ceabSEric Dumazet 77246d3ceabSEric Dumazet list_for_each_safe(q, n, &list) { 77346d3ceabSEric Dumazet tp = list_entry(q, struct tcp_sock, tsq_node); 77446d3ceabSEric Dumazet list_del(&tp->tsq_node); 77546d3ceabSEric Dumazet 77646d3ceabSEric Dumazet sk = (struct sock *)tp; 7770a9648f1SEric Dumazet smp_mb__before_atomic(); 7787aa5470cSEric Dumazet clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags); 7797aa5470cSEric Dumazet 780b223feb9SEric Dumazet if (!sk->sk_lock.owned && 7817aa5470cSEric Dumazet test_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags)) { 78246d3ceabSEric Dumazet bh_lock_sock(sk); 78346d3ceabSEric Dumazet if (!sock_owned_by_user(sk)) { 7847aa5470cSEric Dumazet clear_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags); 7856f458dfbSEric Dumazet tcp_tsq_handler(sk); 78646d3ceabSEric Dumazet } 78746d3ceabSEric Dumazet bh_unlock_sock(sk); 788b223feb9SEric Dumazet } 78946d3ceabSEric Dumazet 79046d3ceabSEric Dumazet sk_free(sk); 79146d3ceabSEric Dumazet } 79246d3ceabSEric Dumazet } 79346d3ceabSEric Dumazet 79440fc3423SEric Dumazet #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \ 79540fc3423SEric Dumazet TCPF_WRITE_TIMER_DEFERRED | \ 79640fc3423SEric Dumazet TCPF_DELACK_TIMER_DEFERRED | \ 79740fc3423SEric Dumazet TCPF_MTU_REDUCED_DEFERRED) 79846d3ceabSEric Dumazet /** 79946d3ceabSEric Dumazet * tcp_release_cb - tcp release_sock() callback 80046d3ceabSEric Dumazet * @sk: socket 80146d3ceabSEric Dumazet * 80246d3ceabSEric Dumazet * called from release_sock() to perform protocol dependent 80346d3ceabSEric Dumazet * actions before socket release. 80446d3ceabSEric Dumazet */ 80546d3ceabSEric Dumazet void tcp_release_cb(struct sock *sk) 80646d3ceabSEric Dumazet { 8076f458dfbSEric Dumazet unsigned long flags, nflags; 80846d3ceabSEric Dumazet 8096f458dfbSEric Dumazet /* perform an atomic operation only if at least one flag is set */ 8106f458dfbSEric Dumazet do { 8117aa5470cSEric Dumazet flags = sk->sk_tsq_flags; 8126f458dfbSEric Dumazet if (!(flags & TCP_DEFERRED_ALL)) 8136f458dfbSEric Dumazet return; 8146f458dfbSEric Dumazet nflags = flags & ~TCP_DEFERRED_ALL; 8157aa5470cSEric Dumazet } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags); 8166f458dfbSEric Dumazet 81740fc3423SEric Dumazet if (flags & TCPF_TSQ_DEFERRED) 8186f458dfbSEric Dumazet tcp_tsq_handler(sk); 8196f458dfbSEric Dumazet 820c3f9b018SEric Dumazet /* Here begins the tricky part : 821c3f9b018SEric Dumazet * We are called from release_sock() with : 822c3f9b018SEric Dumazet * 1) BH disabled 823c3f9b018SEric Dumazet * 2) sk_lock.slock spinlock held 824c3f9b018SEric Dumazet * 3) socket owned by us (sk->sk_lock.owned == 1) 825c3f9b018SEric Dumazet * 826c3f9b018SEric Dumazet * But following code is meant to be called from BH handlers, 827c3f9b018SEric Dumazet * so we should keep BH disabled, but early release socket ownership 828c3f9b018SEric Dumazet */ 829c3f9b018SEric Dumazet sock_release_ownership(sk); 830c3f9b018SEric Dumazet 83140fc3423SEric Dumazet if (flags & TCPF_WRITE_TIMER_DEFERRED) { 8326f458dfbSEric Dumazet tcp_write_timer_handler(sk); 833144d56e9SEric Dumazet __sock_put(sk); 834144d56e9SEric Dumazet } 83540fc3423SEric Dumazet if (flags & TCPF_DELACK_TIMER_DEFERRED) { 8366f458dfbSEric Dumazet tcp_delack_timer_handler(sk); 837144d56e9SEric Dumazet __sock_put(sk); 838144d56e9SEric Dumazet } 83940fc3423SEric Dumazet if (flags & TCPF_MTU_REDUCED_DEFERRED) { 8404fab9071SNeal Cardwell inet_csk(sk)->icsk_af_ops->mtu_reduced(sk); 841144d56e9SEric Dumazet __sock_put(sk); 842144d56e9SEric Dumazet } 84346d3ceabSEric Dumazet } 84446d3ceabSEric Dumazet EXPORT_SYMBOL(tcp_release_cb); 84546d3ceabSEric Dumazet 84646d3ceabSEric Dumazet void __init tcp_tasklet_init(void) 84746d3ceabSEric Dumazet { 84846d3ceabSEric Dumazet int i; 84946d3ceabSEric Dumazet 85046d3ceabSEric Dumazet for_each_possible_cpu(i) { 85146d3ceabSEric Dumazet struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i); 85246d3ceabSEric Dumazet 85346d3ceabSEric Dumazet INIT_LIST_HEAD(&tsq->head); 85446d3ceabSEric Dumazet tasklet_init(&tsq->tasklet, 85546d3ceabSEric Dumazet tcp_tasklet_func, 85646d3ceabSEric Dumazet (unsigned long)tsq); 85746d3ceabSEric Dumazet } 85846d3ceabSEric Dumazet } 85946d3ceabSEric Dumazet 86046d3ceabSEric Dumazet /* 86146d3ceabSEric Dumazet * Write buffer destructor automatically called from kfree_skb. 8628e3bff96Sstephen hemminger * We can't xmit new skbs from this context, as we might already 86346d3ceabSEric Dumazet * hold qdisc lock. 86446d3ceabSEric Dumazet */ 865d6a4a104SEric Dumazet void tcp_wfree(struct sk_buff *skb) 86646d3ceabSEric Dumazet { 86746d3ceabSEric Dumazet struct sock *sk = skb->sk; 86846d3ceabSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 869408f0a6cSEric Dumazet unsigned long flags, nval, oval; 8709b462d02SEric Dumazet 8719b462d02SEric Dumazet /* Keep one reference on sk_wmem_alloc. 8729b462d02SEric Dumazet * Will be released by sk_free() from here or tcp_tasklet_func() 8739b462d02SEric Dumazet */ 87414afee4bSReshetova, Elena WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc)); 8759b462d02SEric Dumazet 8769b462d02SEric Dumazet /* If this softirq is serviced by ksoftirqd, we are likely under stress. 8779b462d02SEric Dumazet * Wait until our queues (qdisc + devices) are drained. 8789b462d02SEric Dumazet * This gives : 8799b462d02SEric Dumazet * - less callbacks to tcp_write_xmit(), reducing stress (batches) 8809b462d02SEric Dumazet * - chance for incoming ACK (processed by another cpu maybe) 8819b462d02SEric Dumazet * to migrate this flow (skb->ooo_okay will be eventually set) 8829b462d02SEric Dumazet */ 88314afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current) 8849b462d02SEric Dumazet goto out; 88546d3ceabSEric Dumazet 8867aa5470cSEric Dumazet for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) { 88746d3ceabSEric Dumazet struct tsq_tasklet *tsq; 888a9b204d1SEric Dumazet bool empty; 88946d3ceabSEric Dumazet 890408f0a6cSEric Dumazet if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED)) 891408f0a6cSEric Dumazet goto out; 892408f0a6cSEric Dumazet 893b223feb9SEric Dumazet nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED | TCPF_TSQ_DEFERRED; 8947aa5470cSEric Dumazet nval = cmpxchg(&sk->sk_tsq_flags, oval, nval); 895408f0a6cSEric Dumazet if (nval != oval) 896408f0a6cSEric Dumazet continue; 897408f0a6cSEric Dumazet 89846d3ceabSEric Dumazet /* queue this socket to tasklet queue */ 89946d3ceabSEric Dumazet local_irq_save(flags); 900903ceff7SChristoph Lameter tsq = this_cpu_ptr(&tsq_tasklet); 901a9b204d1SEric Dumazet empty = list_empty(&tsq->head); 90246d3ceabSEric Dumazet list_add(&tp->tsq_node, &tsq->head); 903a9b204d1SEric Dumazet if (empty) 90446d3ceabSEric Dumazet tasklet_schedule(&tsq->tasklet); 90546d3ceabSEric Dumazet local_irq_restore(flags); 9069b462d02SEric Dumazet return; 90746d3ceabSEric Dumazet } 9089b462d02SEric Dumazet out: 9099b462d02SEric Dumazet sk_free(sk); 91046d3ceabSEric Dumazet } 91146d3ceabSEric Dumazet 912218af599SEric Dumazet /* Note: Called under hard irq. 913218af599SEric Dumazet * We can not call TCP stack right away. 914218af599SEric Dumazet */ 915218af599SEric Dumazet enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer) 916218af599SEric Dumazet { 917218af599SEric Dumazet struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer); 918218af599SEric Dumazet struct sock *sk = (struct sock *)tp; 919218af599SEric Dumazet unsigned long nval, oval; 920218af599SEric Dumazet 921218af599SEric Dumazet for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) { 922218af599SEric Dumazet struct tsq_tasklet *tsq; 923218af599SEric Dumazet bool empty; 924218af599SEric Dumazet 925218af599SEric Dumazet if (oval & TSQF_QUEUED) 926218af599SEric Dumazet break; 927218af599SEric Dumazet 928218af599SEric Dumazet nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED | TCPF_TSQ_DEFERRED; 929218af599SEric Dumazet nval = cmpxchg(&sk->sk_tsq_flags, oval, nval); 930218af599SEric Dumazet if (nval != oval) 931218af599SEric Dumazet continue; 932218af599SEric Dumazet 93314afee4bSReshetova, Elena if (!refcount_inc_not_zero(&sk->sk_wmem_alloc)) 934218af599SEric Dumazet break; 935218af599SEric Dumazet /* queue this socket to tasklet queue */ 936218af599SEric Dumazet tsq = this_cpu_ptr(&tsq_tasklet); 937218af599SEric Dumazet empty = list_empty(&tsq->head); 938218af599SEric Dumazet list_add(&tp->tsq_node, &tsq->head); 939218af599SEric Dumazet if (empty) 940218af599SEric Dumazet tasklet_schedule(&tsq->tasklet); 941218af599SEric Dumazet break; 942218af599SEric Dumazet } 943218af599SEric Dumazet return HRTIMER_NORESTART; 944218af599SEric Dumazet } 945218af599SEric Dumazet 946218af599SEric Dumazet /* BBR congestion control needs pacing. 947218af599SEric Dumazet * Same remark for SO_MAX_PACING_RATE. 948218af599SEric Dumazet * sch_fq packet scheduler is efficiently handling pacing, 949218af599SEric Dumazet * but is not always installed/used. 950218af599SEric Dumazet * Return true if TCP stack should pace packets itself. 951218af599SEric Dumazet */ 952218af599SEric Dumazet static bool tcp_needs_internal_pacing(const struct sock *sk) 953218af599SEric Dumazet { 954218af599SEric Dumazet return smp_load_acquire(&sk->sk_pacing_status) == SK_PACING_NEEDED; 955218af599SEric Dumazet } 956218af599SEric Dumazet 957218af599SEric Dumazet static void tcp_internal_pacing(struct sock *sk, const struct sk_buff *skb) 958218af599SEric Dumazet { 959218af599SEric Dumazet u64 len_ns; 960218af599SEric Dumazet u32 rate; 961218af599SEric Dumazet 962218af599SEric Dumazet if (!tcp_needs_internal_pacing(sk)) 963218af599SEric Dumazet return; 964218af599SEric Dumazet rate = sk->sk_pacing_rate; 965218af599SEric Dumazet if (!rate || rate == ~0U) 966218af599SEric Dumazet return; 967218af599SEric Dumazet 968218af599SEric Dumazet /* Should account for header sizes as sch_fq does, 969218af599SEric Dumazet * but lets make things simple. 970218af599SEric Dumazet */ 971218af599SEric Dumazet len_ns = (u64)skb->len * NSEC_PER_SEC; 972218af599SEric Dumazet do_div(len_ns, rate); 973218af599SEric Dumazet hrtimer_start(&tcp_sk(sk)->pacing_timer, 974218af599SEric Dumazet ktime_add_ns(ktime_get(), len_ns), 975218af599SEric Dumazet HRTIMER_MODE_ABS_PINNED); 976218af599SEric Dumazet } 977218af599SEric Dumazet 978e2080072SEric Dumazet static void tcp_update_skb_after_send(struct tcp_sock *tp, struct sk_buff *skb) 979e2080072SEric Dumazet { 980e2080072SEric Dumazet skb->skb_mstamp = tp->tcp_mstamp; 981e2080072SEric Dumazet list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue); 982e2080072SEric Dumazet } 983e2080072SEric Dumazet 9841da177e4SLinus Torvalds /* This routine actually transmits TCP packets queued in by 9851da177e4SLinus Torvalds * tcp_do_sendmsg(). This is used by both the initial 9861da177e4SLinus Torvalds * transmission and possible later retransmissions. 9871da177e4SLinus Torvalds * All SKB's seen here are completely headerless. It is our 9881da177e4SLinus Torvalds * job to build the TCP header, and pass the packet down to 9891da177e4SLinus Torvalds * IP so it can do the same plus pass the packet off to the 9901da177e4SLinus Torvalds * device. 9911da177e4SLinus Torvalds * 9921da177e4SLinus Torvalds * We are working here with either a clone of the original 9931da177e4SLinus Torvalds * SKB, or a fresh unique copy made by the retransmit engine. 9941da177e4SLinus Torvalds */ 995056834d9SIlpo Järvinen static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, 996056834d9SIlpo Järvinen gfp_t gfp_mask) 9971da177e4SLinus Torvalds { 9986687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 999dfb4b9dcSDavid S. Miller struct inet_sock *inet; 1000dfb4b9dcSDavid S. Miller struct tcp_sock *tp; 1001dfb4b9dcSDavid S. Miller struct tcp_skb_cb *tcb; 100233ad798cSAdam Langley struct tcp_out_options opts; 100395c96174SEric Dumazet unsigned int tcp_options_size, tcp_header_size; 10048c72c65bSEric Dumazet struct sk_buff *oskb = NULL; 1005cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 10061da177e4SLinus Torvalds struct tcphdr *th; 10071da177e4SLinus Torvalds int err; 10081da177e4SLinus Torvalds 1009dfb4b9dcSDavid S. Miller BUG_ON(!skb || !tcp_skb_pcount(skb)); 10106f094b9eSLawrence Brakmo tp = tcp_sk(sk); 1011dfb4b9dcSDavid S. Miller 1012ccdbb6e9SEric Dumazet if (clone_it) { 10136f094b9eSLawrence Brakmo TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq 10146f094b9eSLawrence Brakmo - tp->snd_una; 10158c72c65bSEric Dumazet oskb = skb; 1016e2080072SEric Dumazet 1017e2080072SEric Dumazet tcp_skb_tsorted_save(oskb) { 1018e2080072SEric Dumazet if (unlikely(skb_cloned(oskb))) 1019e2080072SEric Dumazet skb = pskb_copy(oskb, gfp_mask); 1020dfb4b9dcSDavid S. Miller else 1021e2080072SEric Dumazet skb = skb_clone(oskb, gfp_mask); 1022e2080072SEric Dumazet } tcp_skb_tsorted_restore(oskb); 1023e2080072SEric Dumazet 1024dfb4b9dcSDavid S. Miller if (unlikely(!skb)) 1025dfb4b9dcSDavid S. Miller return -ENOBUFS; 1026dfb4b9dcSDavid S. Miller } 10278c72c65bSEric Dumazet skb->skb_mstamp = tp->tcp_mstamp; 1028dfb4b9dcSDavid S. Miller 1029dfb4b9dcSDavid S. Miller inet = inet_sk(sk); 1030dfb4b9dcSDavid S. Miller tcb = TCP_SKB_CB(skb); 103133ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 10321da177e4SLinus Torvalds 10334de075e0SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) 103433ad798cSAdam Langley tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); 103533ad798cSAdam Langley else 103633ad798cSAdam Langley tcp_options_size = tcp_established_options(sk, skb, &opts, 103733ad798cSAdam Langley &md5); 103833ad798cSAdam Langley tcp_header_size = tcp_options_size + sizeof(struct tcphdr); 10391da177e4SLinus Torvalds 1040547669d4SEric Dumazet /* if no packet is in qdisc/device queue, then allow XPS to select 1041b2532eb9SEric Dumazet * another queue. We can be called from tcp_tsq_handler() 1042b2532eb9SEric Dumazet * which holds one reference to sk_wmem_alloc. 1043b2532eb9SEric Dumazet * 1044b2532eb9SEric Dumazet * TODO: Ideally, in-flight pure ACK packets should not matter here. 1045b2532eb9SEric Dumazet * One way to get this would be to set skb->truesize = 2 on them. 1046547669d4SEric Dumazet */ 1047b2532eb9SEric Dumazet skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1); 10481da177e4SLinus Torvalds 104938ab52e8SEric Dumazet /* If we had to use memory reserve to allocate this skb, 105038ab52e8SEric Dumazet * this might cause drops if packet is looped back : 105138ab52e8SEric Dumazet * Other socket might not have SOCK_MEMALLOC. 105238ab52e8SEric Dumazet * Packets not looped back do not care about pfmemalloc. 105338ab52e8SEric Dumazet */ 105438ab52e8SEric Dumazet skb->pfmemalloc = 0; 105538ab52e8SEric Dumazet 1056aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 1057aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 105846d3ceabSEric Dumazet 105946d3ceabSEric Dumazet skb_orphan(skb); 106046d3ceabSEric Dumazet skb->sk = sk; 10611d2077acSEric Dumazet skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree; 1062b73c3d0eSTom Herbert skb_set_hash_from_sk(skb, sk); 106314afee4bSReshetova, Elena refcount_add(skb->truesize, &sk->sk_wmem_alloc); 10641da177e4SLinus Torvalds 1065c3a2e837SJulian Anastasov skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm); 1066c3a2e837SJulian Anastasov 10671da177e4SLinus Torvalds /* Build TCP header and checksum it. */ 1068ea1627c2SEric Dumazet th = (struct tcphdr *)skb->data; 1069c720c7e8SEric Dumazet th->source = inet->inet_sport; 1070c720c7e8SEric Dumazet th->dest = inet->inet_dport; 10711da177e4SLinus Torvalds th->seq = htonl(tcb->seq); 10721da177e4SLinus Torvalds th->ack_seq = htonl(tp->rcv_nxt); 1073df7a3b07SAl Viro *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | 10744de075e0SEric Dumazet tcb->tcp_flags); 1075dfb4b9dcSDavid S. Miller 10761da177e4SLinus Torvalds th->check = 0; 10771da177e4SLinus Torvalds th->urg_ptr = 0; 10781da177e4SLinus Torvalds 107933f5f57eSIlpo Järvinen /* The urg_mode check is necessary during a below snd_una win probe */ 10807691367dSHerbert Xu if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { 10817691367dSHerbert Xu if (before(tp->snd_up, tcb->seq + 0x10000)) { 10821da177e4SLinus Torvalds th->urg_ptr = htons(tp->snd_up - tcb->seq); 10831da177e4SLinus Torvalds th->urg = 1; 10847691367dSHerbert Xu } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { 10850eae88f3SEric Dumazet th->urg_ptr = htons(0xFFFF); 10867691367dSHerbert Xu th->urg = 1; 10877691367dSHerbert Xu } 10881da177e4SLinus Torvalds } 10891da177e4SLinus Torvalds 1090bd0388aeSWilliam Allen Simpson tcp_options_write((__be32 *)(th + 1), tp, &opts); 109151466a75SEric Dumazet skb_shinfo(skb)->gso_type = sk->sk_gso_type; 1092ea1627c2SEric Dumazet if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) { 1093ea1627c2SEric Dumazet th->window = htons(tcp_select_window(sk)); 1094ea1627c2SEric Dumazet tcp_ecn_send(sk, skb, th, tcp_header_size); 1095ea1627c2SEric Dumazet } else { 1096ea1627c2SEric Dumazet /* RFC1323: The window in SYN & SYN/ACK segments 1097ea1627c2SEric Dumazet * is never scaled. 1098ea1627c2SEric Dumazet */ 1099ea1627c2SEric Dumazet th->window = htons(min(tp->rcv_wnd, 65535U)); 1100ea1627c2SEric Dumazet } 1101cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 1102cfb6eeb4SYOSHIFUJI Hideaki /* Calculate the MD5 hash, as we have all we need now */ 1103cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 1104a465419bSEric Dumazet sk_nocaps_add(sk, NETIF_F_GSO_MASK); 1105bd0388aeSWilliam Allen Simpson tp->af_specific->calc_md5_hash(opts.hash_location, 110639f8e58eSEric Dumazet md5, sk, skb); 1107cfb6eeb4SYOSHIFUJI Hideaki } 1108cfb6eeb4SYOSHIFUJI Hideaki #endif 1109cfb6eeb4SYOSHIFUJI Hideaki 1110bb296246SHerbert Xu icsk->icsk_af_ops->send_check(sk, skb); 11111da177e4SLinus Torvalds 11124de075e0SEric Dumazet if (likely(tcb->tcp_flags & TCPHDR_ACK)) 1113fc6415bcSDavid S. Miller tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); 11141da177e4SLinus Torvalds 1115a44d6eacSMartin KaFai Lau if (skb->len != tcp_header_size) { 1116cf533ea5SEric Dumazet tcp_event_data_sent(tp, sk); 1117a44d6eacSMartin KaFai Lau tp->data_segs_out += tcp_skb_pcount(skb); 1118218af599SEric Dumazet tcp_internal_pacing(sk, skb); 1119a44d6eacSMartin KaFai Lau } 11201da177e4SLinus Torvalds 1121bd37a088SWei Yongjun if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) 1122aa2ea058STom Herbert TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, 1123aa2ea058STom Herbert tcp_skb_pcount(skb)); 11241da177e4SLinus Torvalds 11252efd055cSMarcelo Ricardo Leitner tp->segs_out += tcp_skb_pcount(skb); 1126f69ad292SEric Dumazet /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */ 1127cd7d8498SEric Dumazet skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb); 1128f69ad292SEric Dumazet skb_shinfo(skb)->gso_size = tcp_skb_mss(skb); 1129cd7d8498SEric Dumazet 11307faee5c0SEric Dumazet /* Our usage of tstamp should remain private */ 11312456e855SThomas Gleixner skb->tstamp = 0; 1132971f10ecSEric Dumazet 1133971f10ecSEric Dumazet /* Cleanup our debris for IP stacks */ 1134971f10ecSEric Dumazet memset(skb->cb, 0, max(sizeof(struct inet_skb_parm), 1135971f10ecSEric Dumazet sizeof(struct inet6_skb_parm))); 1136971f10ecSEric Dumazet 1137b0270e91SEric Dumazet err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl); 11387faee5c0SEric Dumazet 11398c72c65bSEric Dumazet if (unlikely(err > 0)) { 11405ee2c941SChristoph Paasch tcp_enter_cwr(sk); 11418c72c65bSEric Dumazet err = net_xmit_eval(err); 11428c72c65bSEric Dumazet } 1143fc225799SEric Dumazet if (!err && oskb) { 1144e2080072SEric Dumazet tcp_update_skb_after_send(tp, oskb); 1145fc225799SEric Dumazet tcp_rate_skb_sent(sk, oskb); 1146fc225799SEric Dumazet } 11478c72c65bSEric Dumazet return err; 11481da177e4SLinus Torvalds } 11491da177e4SLinus Torvalds 115067edfef7SAndi Kleen /* This routine just queues the buffer for sending. 11511da177e4SLinus Torvalds * 11521da177e4SLinus Torvalds * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, 11531da177e4SLinus Torvalds * otherwise socket can stall. 11541da177e4SLinus Torvalds */ 11551da177e4SLinus Torvalds static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) 11561da177e4SLinus Torvalds { 11571da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 11581da177e4SLinus Torvalds 11591da177e4SLinus Torvalds /* Advance write_seq and place onto the write_queue. */ 11601da177e4SLinus Torvalds tp->write_seq = TCP_SKB_CB(skb)->end_seq; 1161f4a775d1SEric Dumazet __skb_header_release(skb); 1162fe067e8aSDavid S. Miller tcp_add_write_queue_tail(sk, skb); 11633ab224beSHideo Aoki sk->sk_wmem_queued += skb->truesize; 11643ab224beSHideo Aoki sk_mem_charge(sk, skb->truesize); 11651da177e4SLinus Torvalds } 11661da177e4SLinus Torvalds 116767edfef7SAndi Kleen /* Initialize TSO segments for a packet. */ 11685bbb432cSEric Dumazet static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now) 1169f6302d1dSDavid S. Miller { 11708f26fb1cSEric Dumazet if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) { 1171f6302d1dSDavid S. Miller /* Avoid the costly divide in the normal 1172f6302d1dSDavid S. Miller * non-TSO case. 1173f6302d1dSDavid S. Miller */ 1174cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, 1); 1175f69ad292SEric Dumazet TCP_SKB_CB(skb)->tcp_gso_size = 0; 1176f6302d1dSDavid S. Miller } else { 1177cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now)); 1178f69ad292SEric Dumazet TCP_SKB_CB(skb)->tcp_gso_size = mss_now; 11791da177e4SLinus Torvalds } 11801da177e4SLinus Torvalds } 11811da177e4SLinus Torvalds 118291fed7a1SIlpo Järvinen /* When a modification to fackets out becomes necessary, we need to check 118368f8353bSIlpo Järvinen * skb is counted to fackets_out or not. 118491fed7a1SIlpo Järvinen */ 1185cf533ea5SEric Dumazet static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb, 118691fed7a1SIlpo Järvinen int decr) 118791fed7a1SIlpo Järvinen { 1188a47e5a98SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1189a47e5a98SIlpo Järvinen 1190dc86967bSIlpo Järvinen if (!tp->sacked_out || tcp_is_reno(tp)) 119191fed7a1SIlpo Järvinen return; 119291fed7a1SIlpo Järvinen 11936859d494SIlpo Järvinen if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq)) 119491fed7a1SIlpo Järvinen tp->fackets_out -= decr; 119591fed7a1SIlpo Järvinen } 119691fed7a1SIlpo Järvinen 1197797108d1SIlpo Järvinen /* Pcount in the middle of the write queue got changed, we need to do various 1198797108d1SIlpo Järvinen * tweaks to fix counters 1199797108d1SIlpo Järvinen */ 1200cf533ea5SEric Dumazet static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) 1201797108d1SIlpo Järvinen { 1202797108d1SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1203797108d1SIlpo Järvinen 1204797108d1SIlpo Järvinen tp->packets_out -= decr; 1205797108d1SIlpo Järvinen 1206797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 1207797108d1SIlpo Järvinen tp->sacked_out -= decr; 1208797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 1209797108d1SIlpo Järvinen tp->retrans_out -= decr; 1210797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 1211797108d1SIlpo Järvinen tp->lost_out -= decr; 1212797108d1SIlpo Järvinen 1213797108d1SIlpo Järvinen /* Reno case is special. Sigh... */ 1214797108d1SIlpo Järvinen if (tcp_is_reno(tp) && decr > 0) 1215797108d1SIlpo Järvinen tp->sacked_out -= min_t(u32, tp->sacked_out, decr); 1216797108d1SIlpo Järvinen 1217797108d1SIlpo Järvinen tcp_adjust_fackets_out(sk, skb, decr); 1218797108d1SIlpo Järvinen 1219797108d1SIlpo Järvinen if (tp->lost_skb_hint && 1220797108d1SIlpo Järvinen before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && 122152cf3cc8SIlpo Järvinen (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))) 1222797108d1SIlpo Järvinen tp->lost_cnt_hint -= decr; 1223797108d1SIlpo Järvinen 1224797108d1SIlpo Järvinen tcp_verify_left_out(tp); 1225797108d1SIlpo Järvinen } 1226797108d1SIlpo Järvinen 12270a2cf20cSSoheil Hassas Yeganeh static bool tcp_has_tx_tstamp(const struct sk_buff *skb) 12280a2cf20cSSoheil Hassas Yeganeh { 12290a2cf20cSSoheil Hassas Yeganeh return TCP_SKB_CB(skb)->txstamp_ack || 12300a2cf20cSSoheil Hassas Yeganeh (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP); 12310a2cf20cSSoheil Hassas Yeganeh } 12320a2cf20cSSoheil Hassas Yeganeh 1233490cc7d0SWillem de Bruijn static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2) 1234490cc7d0SWillem de Bruijn { 1235490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo = skb_shinfo(skb); 1236490cc7d0SWillem de Bruijn 12370a2cf20cSSoheil Hassas Yeganeh if (unlikely(tcp_has_tx_tstamp(skb)) && 1238490cc7d0SWillem de Bruijn !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) { 1239490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo2 = skb_shinfo(skb2); 1240490cc7d0SWillem de Bruijn u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP; 1241490cc7d0SWillem de Bruijn 1242490cc7d0SWillem de Bruijn shinfo->tx_flags &= ~tsflags; 1243490cc7d0SWillem de Bruijn shinfo2->tx_flags |= tsflags; 1244490cc7d0SWillem de Bruijn swap(shinfo->tskey, shinfo2->tskey); 1245b51e13faSMartin KaFai Lau TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack; 1246b51e13faSMartin KaFai Lau TCP_SKB_CB(skb)->txstamp_ack = 0; 1247490cc7d0SWillem de Bruijn } 1248490cc7d0SWillem de Bruijn } 1249490cc7d0SWillem de Bruijn 1250a166140eSMartin KaFai Lau static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2) 1251a166140eSMartin KaFai Lau { 1252a166140eSMartin KaFai Lau TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor; 1253a166140eSMartin KaFai Lau TCP_SKB_CB(skb)->eor = 0; 1254a166140eSMartin KaFai Lau } 1255a166140eSMartin KaFai Lau 125675c119afSEric Dumazet /* Insert buff after skb on the write or rtx queue of sk. */ 125775c119afSEric Dumazet static void tcp_insert_write_queue_after(struct sk_buff *skb, 125875c119afSEric Dumazet struct sk_buff *buff, 125975c119afSEric Dumazet struct sock *sk, 126075c119afSEric Dumazet enum tcp_queue tcp_queue) 126175c119afSEric Dumazet { 126275c119afSEric Dumazet if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE) 126375c119afSEric Dumazet __skb_queue_after(&sk->sk_write_queue, skb, buff); 126475c119afSEric Dumazet else 126575c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, buff); 126675c119afSEric Dumazet } 126775c119afSEric Dumazet 12681da177e4SLinus Torvalds /* Function to create two new TCP segments. Shrinks the given segment 12691da177e4SLinus Torvalds * to the specified size and appends a new segment with the rest of the 12701da177e4SLinus Torvalds * packet to the list. This won't be called frequently, I hope. 12711da177e4SLinus Torvalds * Remember, these are still headerless SKBs at this point. 12721da177e4SLinus Torvalds */ 127375c119afSEric Dumazet int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue, 127475c119afSEric Dumazet struct sk_buff *skb, u32 len, 12756cc55e09SOctavian Purdila unsigned int mss_now, gfp_t gfp) 12761da177e4SLinus Torvalds { 12771da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 12781da177e4SLinus Torvalds struct sk_buff *buff; 12796475be16SDavid S. Miller int nsize, old_factor; 1280b60b49eaSHerbert Xu int nlen; 12819ce01461SIlpo Järvinen u8 flags; 12821da177e4SLinus Torvalds 12832fceec13SIlpo Järvinen if (WARN_ON(len > skb->len)) 12842fceec13SIlpo Järvinen return -EINVAL; 12856a438bbeSStephen Hemminger 12861da177e4SLinus Torvalds nsize = skb_headlen(skb) - len; 12871da177e4SLinus Torvalds if (nsize < 0) 12881da177e4SLinus Torvalds nsize = 0; 12891da177e4SLinus Torvalds 12906cc55e09SOctavian Purdila if (skb_unclone(skb, gfp)) 12911da177e4SLinus Torvalds return -ENOMEM; 12921da177e4SLinus Torvalds 12931da177e4SLinus Torvalds /* Get a new skb... force flag on. */ 1294eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, nsize, gfp, true); 129551456b29SIan Morris if (!buff) 12961da177e4SLinus Torvalds return -ENOMEM; /* We'll just try again later. */ 1297ef5cb973SHerbert Xu 12983ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 12993ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1300b60b49eaSHerbert Xu nlen = skb->len - len - nsize; 1301b60b49eaSHerbert Xu buff->truesize += nlen; 1302b60b49eaSHerbert Xu skb->truesize -= nlen; 13031da177e4SLinus Torvalds 13041da177e4SLinus Torvalds /* Correct the sequence numbers. */ 13051da177e4SLinus Torvalds TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 13061da177e4SLinus Torvalds TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 13071da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 13081da177e4SLinus Torvalds 13091da177e4SLinus Torvalds /* PSH and FIN should only be set in the second packet. */ 13104de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 13114de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 13124de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1313e14c3cafSHerbert Xu TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; 1314a166140eSMartin KaFai Lau tcp_skb_fragment_eor(skb, buff); 13151da177e4SLinus Torvalds 131684fa7933SPatrick McHardy if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { 13171da177e4SLinus Torvalds /* Copy and checksum data tail into the new buffer. */ 1318056834d9SIlpo Järvinen buff->csum = csum_partial_copy_nocheck(skb->data + len, 1319056834d9SIlpo Järvinen skb_put(buff, nsize), 13201da177e4SLinus Torvalds nsize, 0); 13211da177e4SLinus Torvalds 13221da177e4SLinus Torvalds skb_trim(skb, len); 13231da177e4SLinus Torvalds 13241da177e4SLinus Torvalds skb->csum = csum_block_sub(skb->csum, buff->csum, len); 13251da177e4SLinus Torvalds } else { 132684fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 13271da177e4SLinus Torvalds skb_split(skb, buff, len); 13281da177e4SLinus Torvalds } 13291da177e4SLinus Torvalds 13301da177e4SLinus Torvalds buff->ip_summed = skb->ip_summed; 13311da177e4SLinus Torvalds 1332a61bbcf2SPatrick McHardy buff->tstamp = skb->tstamp; 1333490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 13341da177e4SLinus Torvalds 13356475be16SDavid S. Miller old_factor = tcp_skb_pcount(skb); 13366475be16SDavid S. Miller 13371da177e4SLinus Torvalds /* Fix up tso_factor for both original and new SKB. */ 13385bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 13395bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 13401da177e4SLinus Torvalds 1341b9f64820SYuchung Cheng /* Update delivered info for the new segment */ 1342b9f64820SYuchung Cheng TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx; 1343b9f64820SYuchung Cheng 13446475be16SDavid S. Miller /* If this packet has been sent out already, we must 13456475be16SDavid S. Miller * adjust the various packet counters. 13466475be16SDavid S. Miller */ 1347cf0b450cSHerbert Xu if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 13486475be16SDavid S. Miller int diff = old_factor - tcp_skb_pcount(skb) - 13496475be16SDavid S. Miller tcp_skb_pcount(buff); 13501da177e4SLinus Torvalds 1351797108d1SIlpo Järvinen if (diff) 1352797108d1SIlpo Järvinen tcp_adjust_pcount(sk, skb, diff); 13531da177e4SLinus Torvalds } 13541da177e4SLinus Torvalds 13551da177e4SLinus Torvalds /* Link BUFF into the send queue. */ 1356f4a775d1SEric Dumazet __skb_header_release(buff); 135775c119afSEric Dumazet tcp_insert_write_queue_after(skb, buff, sk, tcp_queue); 1358e2080072SEric Dumazet list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor); 13591da177e4SLinus Torvalds 13601da177e4SLinus Torvalds return 0; 13611da177e4SLinus Torvalds } 13621da177e4SLinus Torvalds 1363f4d01666SEric Dumazet /* This is similar to __pskb_pull_tail(). The difference is that pulled 1364f4d01666SEric Dumazet * data is not copied, but immediately discarded. 13651da177e4SLinus Torvalds */ 13667162fb24SEric Dumazet static int __pskb_trim_head(struct sk_buff *skb, int len) 13671da177e4SLinus Torvalds { 13687b7fc97aSEric Dumazet struct skb_shared_info *shinfo; 13691da177e4SLinus Torvalds int i, k, eat; 13701da177e4SLinus Torvalds 13714fa48bf3SEric Dumazet eat = min_t(int, len, skb_headlen(skb)); 13724fa48bf3SEric Dumazet if (eat) { 13734fa48bf3SEric Dumazet __skb_pull(skb, eat); 13744fa48bf3SEric Dumazet len -= eat; 13754fa48bf3SEric Dumazet if (!len) 13767162fb24SEric Dumazet return 0; 13774fa48bf3SEric Dumazet } 13781da177e4SLinus Torvalds eat = len; 13791da177e4SLinus Torvalds k = 0; 13807b7fc97aSEric Dumazet shinfo = skb_shinfo(skb); 13817b7fc97aSEric Dumazet for (i = 0; i < shinfo->nr_frags; i++) { 13827b7fc97aSEric Dumazet int size = skb_frag_size(&shinfo->frags[i]); 13839e903e08SEric Dumazet 13849e903e08SEric Dumazet if (size <= eat) { 1385aff65da0SIan Campbell skb_frag_unref(skb, i); 13869e903e08SEric Dumazet eat -= size; 13871da177e4SLinus Torvalds } else { 13887b7fc97aSEric Dumazet shinfo->frags[k] = shinfo->frags[i]; 13891da177e4SLinus Torvalds if (eat) { 13907b7fc97aSEric Dumazet shinfo->frags[k].page_offset += eat; 13917b7fc97aSEric Dumazet skb_frag_size_sub(&shinfo->frags[k], eat); 13921da177e4SLinus Torvalds eat = 0; 13931da177e4SLinus Torvalds } 13941da177e4SLinus Torvalds k++; 13951da177e4SLinus Torvalds } 13961da177e4SLinus Torvalds } 13977b7fc97aSEric Dumazet shinfo->nr_frags = k; 13981da177e4SLinus Torvalds 13991da177e4SLinus Torvalds skb->data_len -= len; 14001da177e4SLinus Torvalds skb->len = skb->data_len; 14017162fb24SEric Dumazet return len; 14021da177e4SLinus Torvalds } 14031da177e4SLinus Torvalds 140467edfef7SAndi Kleen /* Remove acked data from a packet in the transmit queue. */ 14051da177e4SLinus Torvalds int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 14061da177e4SLinus Torvalds { 14077162fb24SEric Dumazet u32 delta_truesize; 14087162fb24SEric Dumazet 140914bbd6a5SPravin B Shelar if (skb_unclone(skb, GFP_ATOMIC)) 14101da177e4SLinus Torvalds return -ENOMEM; 14111da177e4SLinus Torvalds 14127162fb24SEric Dumazet delta_truesize = __pskb_trim_head(skb, len); 14131da177e4SLinus Torvalds 14141da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq += len; 141584fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 14161da177e4SLinus Torvalds 14177162fb24SEric Dumazet if (delta_truesize) { 14187162fb24SEric Dumazet skb->truesize -= delta_truesize; 14197162fb24SEric Dumazet sk->sk_wmem_queued -= delta_truesize; 14207162fb24SEric Dumazet sk_mem_uncharge(sk, delta_truesize); 14211da177e4SLinus Torvalds sock_set_flag(sk, SOCK_QUEUE_SHRUNK); 14227162fb24SEric Dumazet } 14231da177e4SLinus Torvalds 14245b35e1e6SNeal Cardwell /* Any change of skb->len requires recalculation of tso factor. */ 14251da177e4SLinus Torvalds if (tcp_skb_pcount(skb) > 1) 14265bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb)); 14271da177e4SLinus Torvalds 14281da177e4SLinus Torvalds return 0; 14291da177e4SLinus Torvalds } 14301da177e4SLinus Torvalds 14311b63edd6SYuchung Cheng /* Calculate MSS not accounting any TCP options. */ 14321b63edd6SYuchung Cheng static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu) 14335d424d5aSJohn Heffner { 1434cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1435cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 14365d424d5aSJohn Heffner int mss_now; 14375d424d5aSJohn Heffner 14385d424d5aSJohn Heffner /* Calculate base mss without TCP options: 14395d424d5aSJohn Heffner It is MMS_S - sizeof(tcphdr) of rfc1122 14405d424d5aSJohn Heffner */ 14415d424d5aSJohn Heffner mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 14425d424d5aSJohn Heffner 144367469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 144467469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 144567469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 144667469601SEric Dumazet 144767469601SEric Dumazet if (dst && dst_allfrag(dst)) 144867469601SEric Dumazet mss_now -= icsk->icsk_af_ops->net_frag_header_len; 144967469601SEric Dumazet } 145067469601SEric Dumazet 14515d424d5aSJohn Heffner /* Clamp it (mss_clamp does not include tcp options) */ 14525d424d5aSJohn Heffner if (mss_now > tp->rx_opt.mss_clamp) 14535d424d5aSJohn Heffner mss_now = tp->rx_opt.mss_clamp; 14545d424d5aSJohn Heffner 14555d424d5aSJohn Heffner /* Now subtract optional transport overhead */ 14565d424d5aSJohn Heffner mss_now -= icsk->icsk_ext_hdr_len; 14575d424d5aSJohn Heffner 14585d424d5aSJohn Heffner /* Then reserve room for full set of TCP options and 8 bytes of data */ 14595d424d5aSJohn Heffner if (mss_now < 48) 14605d424d5aSJohn Heffner mss_now = 48; 14615d424d5aSJohn Heffner return mss_now; 14625d424d5aSJohn Heffner } 14635d424d5aSJohn Heffner 14641b63edd6SYuchung Cheng /* Calculate MSS. Not accounting for SACKs here. */ 14651b63edd6SYuchung Cheng int tcp_mtu_to_mss(struct sock *sk, int pmtu) 14661b63edd6SYuchung Cheng { 14671b63edd6SYuchung Cheng /* Subtract TCP options size, not including SACKs */ 14681b63edd6SYuchung Cheng return __tcp_mtu_to_mss(sk, pmtu) - 14691b63edd6SYuchung Cheng (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr)); 14701b63edd6SYuchung Cheng } 14711b63edd6SYuchung Cheng 14725d424d5aSJohn Heffner /* Inverse of above */ 147367469601SEric Dumazet int tcp_mss_to_mtu(struct sock *sk, int mss) 14745d424d5aSJohn Heffner { 1475cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1476cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 14775d424d5aSJohn Heffner int mtu; 14785d424d5aSJohn Heffner 14795d424d5aSJohn Heffner mtu = mss + 14805d424d5aSJohn Heffner tp->tcp_header_len + 14815d424d5aSJohn Heffner icsk->icsk_ext_hdr_len + 14825d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 14835d424d5aSJohn Heffner 148467469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 148567469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 148667469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 148767469601SEric Dumazet 148867469601SEric Dumazet if (dst && dst_allfrag(dst)) 148967469601SEric Dumazet mtu += icsk->icsk_af_ops->net_frag_header_len; 149067469601SEric Dumazet } 14915d424d5aSJohn Heffner return mtu; 14925d424d5aSJohn Heffner } 1493556c6b46SNeal Cardwell EXPORT_SYMBOL(tcp_mss_to_mtu); 14945d424d5aSJohn Heffner 149567edfef7SAndi Kleen /* MTU probing init per socket */ 14965d424d5aSJohn Heffner void tcp_mtup_init(struct sock *sk) 14975d424d5aSJohn Heffner { 14985d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 14995d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 1500b0f9ca53SFan Du struct net *net = sock_net(sk); 15015d424d5aSJohn Heffner 1502b0f9ca53SFan Du icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1; 15035d424d5aSJohn Heffner icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 15045d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 1505b0f9ca53SFan Du icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss); 15065d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 150705cbc0dbSFan Du if (icsk->icsk_mtup.enabled) 1508c74df29aSEric Dumazet icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; 15095d424d5aSJohn Heffner } 15104bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_mtup_init); 15115d424d5aSJohn Heffner 15121da177e4SLinus Torvalds /* This function synchronize snd mss to current pmtu/exthdr set. 15131da177e4SLinus Torvalds 15141da177e4SLinus Torvalds tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 15151da177e4SLinus Torvalds for TCP options, but includes only bare TCP header. 15161da177e4SLinus Torvalds 15171da177e4SLinus Torvalds tp->rx_opt.mss_clamp is mss negotiated at connection setup. 1518caa20d9aSStephen Hemminger It is minimum of user_mss and mss received with SYN. 15191da177e4SLinus Torvalds It also does not include TCP options. 15201da177e4SLinus Torvalds 1521d83d8461SArnaldo Carvalho de Melo inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. 15221da177e4SLinus Torvalds 15231da177e4SLinus Torvalds tp->mss_cache is current effective sending mss, including 15241da177e4SLinus Torvalds all tcp options except for SACKs. It is evaluated, 15251da177e4SLinus Torvalds taking into account current pmtu, but never exceeds 15261da177e4SLinus Torvalds tp->rx_opt.mss_clamp. 15271da177e4SLinus Torvalds 15281da177e4SLinus Torvalds NOTE1. rfc1122 clearly states that advertised MSS 15291da177e4SLinus Torvalds DOES NOT include either tcp or ip options. 15301da177e4SLinus Torvalds 1531d83d8461SArnaldo Carvalho de Melo NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache 1532d83d8461SArnaldo Carvalho de Melo are READ ONLY outside this function. --ANK (980731) 15331da177e4SLinus Torvalds */ 15341da177e4SLinus Torvalds unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 15351da177e4SLinus Torvalds { 15361da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1537d83d8461SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 15385d424d5aSJohn Heffner int mss_now; 15391da177e4SLinus Torvalds 15405d424d5aSJohn Heffner if (icsk->icsk_mtup.search_high > pmtu) 15415d424d5aSJohn Heffner icsk->icsk_mtup.search_high = pmtu; 15421da177e4SLinus Torvalds 15435d424d5aSJohn Heffner mss_now = tcp_mtu_to_mss(sk, pmtu); 1544409d22b4SIlpo Järvinen mss_now = tcp_bound_to_half_wnd(tp, mss_now); 15451da177e4SLinus Torvalds 15461da177e4SLinus Torvalds /* And store cached results */ 1547d83d8461SArnaldo Carvalho de Melo icsk->icsk_pmtu_cookie = pmtu; 15485d424d5aSJohn Heffner if (icsk->icsk_mtup.enabled) 15495d424d5aSJohn Heffner mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); 1550c1b4a7e6SDavid S. Miller tp->mss_cache = mss_now; 15511da177e4SLinus Torvalds 15521da177e4SLinus Torvalds return mss_now; 15531da177e4SLinus Torvalds } 15544bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_sync_mss); 15551da177e4SLinus Torvalds 15561da177e4SLinus Torvalds /* Compute the current effective MSS, taking SACKs and IP options, 15571da177e4SLinus Torvalds * and even PMTU discovery events into account. 15581da177e4SLinus Torvalds */ 15590c54b85fSIlpo Järvinen unsigned int tcp_current_mss(struct sock *sk) 15601da177e4SLinus Torvalds { 1561cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1562cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1563c1b4a7e6SDavid S. Miller u32 mss_now; 156495c96174SEric Dumazet unsigned int header_len; 156533ad798cSAdam Langley struct tcp_out_options opts; 156633ad798cSAdam Langley struct tcp_md5sig_key *md5; 15671da177e4SLinus Torvalds 1568c1b4a7e6SDavid S. Miller mss_now = tp->mss_cache; 1569c1b4a7e6SDavid S. Miller 15701da177e4SLinus Torvalds if (dst) { 15711da177e4SLinus Torvalds u32 mtu = dst_mtu(dst); 1572d83d8461SArnaldo Carvalho de Melo if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 15731da177e4SLinus Torvalds mss_now = tcp_sync_mss(sk, mtu); 15741da177e4SLinus Torvalds } 15751da177e4SLinus Torvalds 157633ad798cSAdam Langley header_len = tcp_established_options(sk, NULL, &opts, &md5) + 157733ad798cSAdam Langley sizeof(struct tcphdr); 157833ad798cSAdam Langley /* The mss_cache is sized based on tp->tcp_header_len, which assumes 157933ad798cSAdam Langley * some common options. If this is an odd packet (because we have SACK 158033ad798cSAdam Langley * blocks etc) then our calculated header_len will be different, and 158133ad798cSAdam Langley * we have to adjust mss_now correspondingly */ 158233ad798cSAdam Langley if (header_len != tp->tcp_header_len) { 158333ad798cSAdam Langley int delta = (int) header_len - tp->tcp_header_len; 158433ad798cSAdam Langley mss_now -= delta; 158533ad798cSAdam Langley } 1586cfb6eeb4SYOSHIFUJI Hideaki 15871da177e4SLinus Torvalds return mss_now; 15881da177e4SLinus Torvalds } 15891da177e4SLinus Torvalds 159086fd14adSWeiping Pan /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. 159186fd14adSWeiping Pan * As additional protections, we do not touch cwnd in retransmission phases, 159286fd14adSWeiping Pan * and if application hit its sndbuf limit recently. 159386fd14adSWeiping Pan */ 159486fd14adSWeiping Pan static void tcp_cwnd_application_limited(struct sock *sk) 1595a762a980SDavid S. Miller { 15969e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1597a762a980SDavid S. Miller 159886fd14adSWeiping Pan if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && 159986fd14adSWeiping Pan sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 160086fd14adSWeiping Pan /* Limited by application or receiver window. */ 160186fd14adSWeiping Pan u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); 160286fd14adSWeiping Pan u32 win_used = max(tp->snd_cwnd_used, init_win); 160386fd14adSWeiping Pan if (win_used < tp->snd_cwnd) { 160486fd14adSWeiping Pan tp->snd_ssthresh = tcp_current_ssthresh(sk); 160586fd14adSWeiping Pan tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1; 160686fd14adSWeiping Pan } 160786fd14adSWeiping Pan tp->snd_cwnd_used = 0; 160886fd14adSWeiping Pan } 1609c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 161086fd14adSWeiping Pan } 161186fd14adSWeiping Pan 1612ca8a2263SNeal Cardwell static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited) 1613a762a980SDavid S. Miller { 16141b1fc3fdSWei Wang const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 1615a762a980SDavid S. Miller struct tcp_sock *tp = tcp_sk(sk); 1616a762a980SDavid S. Miller 1617ca8a2263SNeal Cardwell /* Track the maximum number of outstanding packets in each 1618ca8a2263SNeal Cardwell * window, and remember whether we were cwnd-limited then. 1619ca8a2263SNeal Cardwell */ 1620ca8a2263SNeal Cardwell if (!before(tp->snd_una, tp->max_packets_seq) || 1621ca8a2263SNeal Cardwell tp->packets_out > tp->max_packets_out) { 1622ca8a2263SNeal Cardwell tp->max_packets_out = tp->packets_out; 1623ca8a2263SNeal Cardwell tp->max_packets_seq = tp->snd_nxt; 1624ca8a2263SNeal Cardwell tp->is_cwnd_limited = is_cwnd_limited; 1625ca8a2263SNeal Cardwell } 1626e114a710SEric Dumazet 162724901551SEric Dumazet if (tcp_is_cwnd_limited(sk)) { 1628a762a980SDavid S. Miller /* Network is feed fully. */ 1629a762a980SDavid S. Miller tp->snd_cwnd_used = 0; 1630c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 1631a762a980SDavid S. Miller } else { 1632a762a980SDavid S. Miller /* Network starves. */ 1633a762a980SDavid S. Miller if (tp->packets_out > tp->snd_cwnd_used) 1634a762a980SDavid S. Miller tp->snd_cwnd_used = tp->packets_out; 1635a762a980SDavid S. Miller 163615d33c07SDavid S. Miller if (sysctl_tcp_slow_start_after_idle && 1637c2203cf7SEric Dumazet (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto && 16381b1fc3fdSWei Wang !ca_ops->cong_control) 1639a762a980SDavid S. Miller tcp_cwnd_application_limited(sk); 1640b0f71bd3SFrancis Yan 1641b0f71bd3SFrancis Yan /* The following conditions together indicate the starvation 1642b0f71bd3SFrancis Yan * is caused by insufficient sender buffer: 1643b0f71bd3SFrancis Yan * 1) just sent some data (see tcp_write_xmit) 1644b0f71bd3SFrancis Yan * 2) not cwnd limited (this else condition) 164575c119afSEric Dumazet * 3) no more data to send (tcp_write_queue_empty()) 1646b0f71bd3SFrancis Yan * 4) application is hitting buffer limit (SOCK_NOSPACE) 1647b0f71bd3SFrancis Yan */ 164875c119afSEric Dumazet if (tcp_write_queue_empty(sk) && sk->sk_socket && 1649b0f71bd3SFrancis Yan test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) && 1650b0f71bd3SFrancis Yan (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) 1651b0f71bd3SFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED); 1652a762a980SDavid S. Miller } 1653a762a980SDavid S. Miller } 1654a762a980SDavid S. Miller 1655d4589926SEric Dumazet /* Minshall's variant of the Nagle send check. */ 1656d4589926SEric Dumazet static bool tcp_minshall_check(const struct tcp_sock *tp) 1657d4589926SEric Dumazet { 1658d4589926SEric Dumazet return after(tp->snd_sml, tp->snd_una) && 1659d4589926SEric Dumazet !after(tp->snd_sml, tp->snd_nxt); 1660d4589926SEric Dumazet } 1661d4589926SEric Dumazet 1662d4589926SEric Dumazet /* Update snd_sml if this skb is under mss 1663d4589926SEric Dumazet * Note that a TSO packet might end with a sub-mss segment 1664d4589926SEric Dumazet * The test is really : 1665d4589926SEric Dumazet * if ((skb->len % mss) != 0) 1666d4589926SEric Dumazet * tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1667d4589926SEric Dumazet * But we can avoid doing the divide again given we already have 1668d4589926SEric Dumazet * skb_pcount = skb->len / mss_now 16690e3a4803SIlpo Järvinen */ 1670d4589926SEric Dumazet static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now, 1671d4589926SEric Dumazet const struct sk_buff *skb) 1672d4589926SEric Dumazet { 1673d4589926SEric Dumazet if (skb->len < tcp_skb_pcount(skb) * mss_now) 1674d4589926SEric Dumazet tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1675d4589926SEric Dumazet } 1676d4589926SEric Dumazet 1677d4589926SEric Dumazet /* Return false, if packet can be sent now without violation Nagle's rules: 1678d4589926SEric Dumazet * 1. It is full sized. (provided by caller in %partial bool) 1679d4589926SEric Dumazet * 2. Or it contains FIN. (already checked by caller) 1680d4589926SEric Dumazet * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. 1681d4589926SEric Dumazet * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 1682d4589926SEric Dumazet * With Minshall's modification: all sent small packets are ACKed. 1683d4589926SEric Dumazet */ 1684d4589926SEric Dumazet static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp, 1685cc93fc51SPeter Pan(潘卫平) int nonagle) 1686d4589926SEric Dumazet { 1687d4589926SEric Dumazet return partial && 1688d4589926SEric Dumazet ((nonagle & TCP_NAGLE_CORK) || 1689d4589926SEric Dumazet (!nonagle && tp->packets_out && tcp_minshall_check(tp))); 1690d4589926SEric Dumazet } 1691605ad7f1SEric Dumazet 1692605ad7f1SEric Dumazet /* Return how many segs we'd like on a TSO packet, 1693605ad7f1SEric Dumazet * to send one TSO packet per ms 1694605ad7f1SEric Dumazet */ 16951b3878caSNeal Cardwell u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now, 16961b3878caSNeal Cardwell int min_tso_segs) 1697605ad7f1SEric Dumazet { 1698605ad7f1SEric Dumazet u32 bytes, segs; 1699605ad7f1SEric Dumazet 1700605ad7f1SEric Dumazet bytes = min(sk->sk_pacing_rate >> 10, 1701605ad7f1SEric Dumazet sk->sk_gso_max_size - 1 - MAX_TCP_HEADER); 1702605ad7f1SEric Dumazet 1703605ad7f1SEric Dumazet /* Goal is to send at least one packet per ms, 1704605ad7f1SEric Dumazet * not one big TSO packet every 100 ms. 1705605ad7f1SEric Dumazet * This preserves ACK clocking and is consistent 1706605ad7f1SEric Dumazet * with tcp_tso_should_defer() heuristic. 1707605ad7f1SEric Dumazet */ 17081b3878caSNeal Cardwell segs = max_t(u32, bytes / mss_now, min_tso_segs); 1709605ad7f1SEric Dumazet 1710605ad7f1SEric Dumazet return min_t(u32, segs, sk->sk_gso_max_segs); 1711605ad7f1SEric Dumazet } 17121b3878caSNeal Cardwell EXPORT_SYMBOL(tcp_tso_autosize); 1713605ad7f1SEric Dumazet 1714ed6e7268SNeal Cardwell /* Return the number of segments we want in the skb we are transmitting. 1715ed6e7268SNeal Cardwell * See if congestion control module wants to decide; otherwise, autosize. 1716ed6e7268SNeal Cardwell */ 1717ed6e7268SNeal Cardwell static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now) 1718ed6e7268SNeal Cardwell { 1719ed6e7268SNeal Cardwell const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 1720ed6e7268SNeal Cardwell u32 tso_segs = ca_ops->tso_segs_goal ? ca_ops->tso_segs_goal(sk) : 0; 1721ed6e7268SNeal Cardwell 17221b3878caSNeal Cardwell return tso_segs ? : 17231b3878caSNeal Cardwell tcp_tso_autosize(sk, mss_now, sysctl_tcp_min_tso_segs); 1724ed6e7268SNeal Cardwell } 1725ed6e7268SNeal Cardwell 1726d4589926SEric Dumazet /* Returns the portion of skb which can be sent right away */ 1727d4589926SEric Dumazet static unsigned int tcp_mss_split_point(const struct sock *sk, 1728d4589926SEric Dumazet const struct sk_buff *skb, 1729d4589926SEric Dumazet unsigned int mss_now, 1730d4589926SEric Dumazet unsigned int max_segs, 1731d4589926SEric Dumazet int nonagle) 1732c1b4a7e6SDavid S. Miller { 1733cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1734d4589926SEric Dumazet u32 partial, needed, window, max_len; 1735c1b4a7e6SDavid S. Miller 173690840defSIlpo Järvinen window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 17371485348dSBen Hutchings max_len = mss_now * max_segs; 17380e3a4803SIlpo Järvinen 17391485348dSBen Hutchings if (likely(max_len <= window && skb != tcp_write_queue_tail(sk))) 17401485348dSBen Hutchings return max_len; 17410e3a4803SIlpo Järvinen 17425ea3a748SIlpo Järvinen needed = min(skb->len, window); 17435ea3a748SIlpo Järvinen 17441485348dSBen Hutchings if (max_len <= needed) 17451485348dSBen Hutchings return max_len; 17460e3a4803SIlpo Järvinen 1747d4589926SEric Dumazet partial = needed % mss_now; 1748d4589926SEric Dumazet /* If last segment is not a full MSS, check if Nagle rules allow us 1749d4589926SEric Dumazet * to include this last segment in this skb. 1750d4589926SEric Dumazet * Otherwise, we'll split the skb at last MSS boundary 1751d4589926SEric Dumazet */ 1752cc93fc51SPeter Pan(潘卫平) if (tcp_nagle_check(partial != 0, tp, nonagle)) 1753d4589926SEric Dumazet return needed - partial; 1754d4589926SEric Dumazet 1755d4589926SEric Dumazet return needed; 1756c1b4a7e6SDavid S. Miller } 1757c1b4a7e6SDavid S. Miller 1758c1b4a7e6SDavid S. Miller /* Can at least one segment of SKB be sent right now, according to the 1759c1b4a7e6SDavid S. Miller * congestion window rules? If so, return how many segments are allowed. 1760c1b4a7e6SDavid S. Miller */ 1761cf533ea5SEric Dumazet static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, 1762cf533ea5SEric Dumazet const struct sk_buff *skb) 1763c1b4a7e6SDavid S. Miller { 1764d649a7a8SEric Dumazet u32 in_flight, cwnd, halfcwnd; 1765c1b4a7e6SDavid S. Miller 1766c1b4a7e6SDavid S. Miller /* Don't be strict about the congestion window for the final FIN. */ 17674de075e0SEric Dumazet if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && 17684de075e0SEric Dumazet tcp_skb_pcount(skb) == 1) 1769c1b4a7e6SDavid S. Miller return 1; 1770c1b4a7e6SDavid S. Miller 1771c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1772c1b4a7e6SDavid S. Miller cwnd = tp->snd_cwnd; 1773d649a7a8SEric Dumazet if (in_flight >= cwnd) 1774c1b4a7e6SDavid S. Miller return 0; 1775d649a7a8SEric Dumazet 1776d649a7a8SEric Dumazet /* For better scheduling, ensure we have at least 1777d649a7a8SEric Dumazet * 2 GSO packets in flight. 1778d649a7a8SEric Dumazet */ 1779d649a7a8SEric Dumazet halfcwnd = max(cwnd >> 1, 1U); 1780d649a7a8SEric Dumazet return min(halfcwnd, cwnd - in_flight); 1781c1b4a7e6SDavid S. Miller } 1782c1b4a7e6SDavid S. Miller 1783b595076aSUwe Kleine-König /* Initialize TSO state of a skb. 178467edfef7SAndi Kleen * This must be invoked the first time we consider transmitting 1785c1b4a7e6SDavid S. Miller * SKB onto the wire. 1786c1b4a7e6SDavid S. Miller */ 17875bbb432cSEric Dumazet static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now) 1788c1b4a7e6SDavid S. Miller { 1789c1b4a7e6SDavid S. Miller int tso_segs = tcp_skb_pcount(skb); 1790c1b4a7e6SDavid S. Miller 1791f8269a49SIlpo Järvinen if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { 17925bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 1793c1b4a7e6SDavid S. Miller tso_segs = tcp_skb_pcount(skb); 1794c1b4a7e6SDavid S. Miller } 1795c1b4a7e6SDavid S. Miller return tso_segs; 1796c1b4a7e6SDavid S. Miller } 1797c1b4a7e6SDavid S. Miller 1798c1b4a7e6SDavid S. Miller 1799a2a385d6SEric Dumazet /* Return true if the Nagle test allows this packet to be 1800c1b4a7e6SDavid S. Miller * sent now. 1801c1b4a7e6SDavid S. Miller */ 1802a2a385d6SEric Dumazet static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, 1803c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1804c1b4a7e6SDavid S. Miller { 1805c1b4a7e6SDavid S. Miller /* Nagle rule does not apply to frames, which sit in the middle of the 1806c1b4a7e6SDavid S. Miller * write_queue (they have no chances to get new data). 1807c1b4a7e6SDavid S. Miller * 1808c1b4a7e6SDavid S. Miller * This is implemented in the callers, where they modify the 'nonagle' 1809c1b4a7e6SDavid S. Miller * argument based upon the location of SKB in the send queue. 1810c1b4a7e6SDavid S. Miller */ 1811c1b4a7e6SDavid S. Miller if (nonagle & TCP_NAGLE_PUSH) 1812a2a385d6SEric Dumazet return true; 1813c1b4a7e6SDavid S. Miller 18149b44190dSYuchung Cheng /* Don't use the nagle rule for urgent data (or for the final FIN). */ 18159b44190dSYuchung Cheng if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) 1816a2a385d6SEric Dumazet return true; 1817c1b4a7e6SDavid S. Miller 1818cc93fc51SPeter Pan(潘卫平) if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle)) 1819a2a385d6SEric Dumazet return true; 1820c1b4a7e6SDavid S. Miller 1821a2a385d6SEric Dumazet return false; 1822c1b4a7e6SDavid S. Miller } 1823c1b4a7e6SDavid S. Miller 1824c1b4a7e6SDavid S. Miller /* Does at least the first segment of SKB fit into the send window? */ 1825a2a385d6SEric Dumazet static bool tcp_snd_wnd_test(const struct tcp_sock *tp, 1826a2a385d6SEric Dumazet const struct sk_buff *skb, 1827056834d9SIlpo Järvinen unsigned int cur_mss) 1828c1b4a7e6SDavid S. Miller { 1829c1b4a7e6SDavid S. Miller u32 end_seq = TCP_SKB_CB(skb)->end_seq; 1830c1b4a7e6SDavid S. Miller 1831c1b4a7e6SDavid S. Miller if (skb->len > cur_mss) 1832c1b4a7e6SDavid S. Miller end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 1833c1b4a7e6SDavid S. Miller 183490840defSIlpo Järvinen return !after(end_seq, tcp_wnd_end(tp)); 1835c1b4a7e6SDavid S. Miller } 1836c1b4a7e6SDavid S. Miller 1837c1b4a7e6SDavid S. Miller /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 1838c1b4a7e6SDavid S. Miller * which is put after SKB on the list. It is very much like 1839c1b4a7e6SDavid S. Miller * tcp_fragment() except that it may make several kinds of assumptions 1840c1b4a7e6SDavid S. Miller * in order to speed up the splitting operation. In particular, we 1841c1b4a7e6SDavid S. Miller * know that all the data is in scatter-gather pages, and that the 1842c1b4a7e6SDavid S. Miller * packet has never been sent out before (and thus is not cloned). 1843c1b4a7e6SDavid S. Miller */ 184475c119afSEric Dumazet static int tso_fragment(struct sock *sk, enum tcp_queue tcp_queue, 184575c119afSEric Dumazet struct sk_buff *skb, unsigned int len, 1846c4ead4c5SEric Dumazet unsigned int mss_now, gfp_t gfp) 1847c1b4a7e6SDavid S. Miller { 1848c1b4a7e6SDavid S. Miller struct sk_buff *buff; 1849c1b4a7e6SDavid S. Miller int nlen = skb->len - len; 18509ce01461SIlpo Järvinen u8 flags; 1851c1b4a7e6SDavid S. Miller 1852c1b4a7e6SDavid S. Miller /* All of a TSO frame must be composed of paged data. */ 1853c8ac3774SHerbert Xu if (skb->len != skb->data_len) 185475c119afSEric Dumazet return tcp_fragment(sk, tcp_queue, skb, len, mss_now, gfp); 1855c1b4a7e6SDavid S. Miller 1856eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, gfp, true); 185751456b29SIan Morris if (unlikely(!buff)) 1858c1b4a7e6SDavid S. Miller return -ENOMEM; 1859c1b4a7e6SDavid S. Miller 18603ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 18613ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1862b60b49eaSHerbert Xu buff->truesize += nlen; 1863c1b4a7e6SDavid S. Miller skb->truesize -= nlen; 1864c1b4a7e6SDavid S. Miller 1865c1b4a7e6SDavid S. Miller /* Correct the sequence numbers. */ 1866c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 1867c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 1868c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 1869c1b4a7e6SDavid S. Miller 1870c1b4a7e6SDavid S. Miller /* PSH and FIN should only be set in the second packet. */ 18714de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 18724de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 18734de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1874c1b4a7e6SDavid S. Miller 1875c1b4a7e6SDavid S. Miller /* This packet was never sent out yet, so no SACK bits. */ 1876c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->sacked = 0; 1877c1b4a7e6SDavid S. Miller 1878a166140eSMartin KaFai Lau tcp_skb_fragment_eor(skb, buff); 1879a166140eSMartin KaFai Lau 188084fa7933SPatrick McHardy buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; 1881c1b4a7e6SDavid S. Miller skb_split(skb, buff, len); 1882490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 1883c1b4a7e6SDavid S. Miller 1884c1b4a7e6SDavid S. Miller /* Fix up tso_factor for both original and new SKB. */ 18855bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 18865bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 1887c1b4a7e6SDavid S. Miller 1888c1b4a7e6SDavid S. Miller /* Link BUFF into the send queue. */ 1889f4a775d1SEric Dumazet __skb_header_release(buff); 189075c119afSEric Dumazet tcp_insert_write_queue_after(skb, buff, sk, tcp_queue); 1891c1b4a7e6SDavid S. Miller 1892c1b4a7e6SDavid S. Miller return 0; 1893c1b4a7e6SDavid S. Miller } 1894c1b4a7e6SDavid S. Miller 1895c1b4a7e6SDavid S. Miller /* Try to defer sending, if possible, in order to minimize the amount 1896c1b4a7e6SDavid S. Miller * of TSO splitting we do. View it as a kind of TSO Nagle test. 1897c1b4a7e6SDavid S. Miller * 1898c1b4a7e6SDavid S. Miller * This algorithm is from John Heffner. 1899c1b4a7e6SDavid S. Miller */ 1900ca8a2263SNeal Cardwell static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb, 1901605ad7f1SEric Dumazet bool *is_cwnd_limited, u32 max_segs) 1902c1b4a7e6SDavid S. Miller { 19036687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 190450c8339eSEric Dumazet u32 age, send_win, cong_win, limit, in_flight; 190550c8339eSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 190650c8339eSEric Dumazet struct sk_buff *head; 1907ad9f4f50SEric Dumazet int win_divisor; 1908c1b4a7e6SDavid S. Miller 19094de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 1910ae8064acSJohn Heffner goto send_now; 1911c1b4a7e6SDavid S. Miller 191299d7662aSEric Dumazet if (icsk->icsk_ca_state >= TCP_CA_Recovery) 1913ae8064acSJohn Heffner goto send_now; 1914ae8064acSJohn Heffner 19155f852eb5SEric Dumazet /* Avoid bursty behavior by allowing defer 19165f852eb5SEric Dumazet * only if the last write was recent. 19175f852eb5SEric Dumazet */ 1918d635fbe2SEric Dumazet if ((s32)(tcp_jiffies32 - tp->lsndtime) > 0) 1919ae8064acSJohn Heffner goto send_now; 1920908a75c1SDavid S. Miller 1921c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1922c1b4a7e6SDavid S. Miller 1923056834d9SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight)); 1924c1b4a7e6SDavid S. Miller 192590840defSIlpo Järvinen send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1926c1b4a7e6SDavid S. Miller 1927c1b4a7e6SDavid S. Miller /* From in_flight test above, we know that cwnd > in_flight. */ 1928c1b4a7e6SDavid S. Miller cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; 1929c1b4a7e6SDavid S. Miller 1930c1b4a7e6SDavid S. Miller limit = min(send_win, cong_win); 1931c1b4a7e6SDavid S. Miller 1932ba244fe9SDavid S. Miller /* If a full-sized TSO skb can be sent, do it. */ 1933605ad7f1SEric Dumazet if (limit >= max_segs * tp->mss_cache) 1934ae8064acSJohn Heffner goto send_now; 1935ba244fe9SDavid S. Miller 193662ad2761SIlpo Järvinen /* Middle in queue won't get any more data, full sendable already? */ 193762ad2761SIlpo Järvinen if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) 193862ad2761SIlpo Järvinen goto send_now; 193962ad2761SIlpo Järvinen 1940ad9f4f50SEric Dumazet win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor); 1941ad9f4f50SEric Dumazet if (win_divisor) { 1942c1b4a7e6SDavid S. Miller u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); 1943c1b4a7e6SDavid S. Miller 1944c1b4a7e6SDavid S. Miller /* If at least some fraction of a window is available, 1945c1b4a7e6SDavid S. Miller * just use it. 1946c1b4a7e6SDavid S. Miller */ 1947ad9f4f50SEric Dumazet chunk /= win_divisor; 1948c1b4a7e6SDavid S. Miller if (limit >= chunk) 1949ae8064acSJohn Heffner goto send_now; 1950c1b4a7e6SDavid S. Miller } else { 1951c1b4a7e6SDavid S. Miller /* Different approach, try not to defer past a single 1952c1b4a7e6SDavid S. Miller * ACK. Receiver should ACK every other full sized 1953c1b4a7e6SDavid S. Miller * frame, so if we have space for more than 3 frames 1954c1b4a7e6SDavid S. Miller * then send now. 1955c1b4a7e6SDavid S. Miller */ 19566b5a5c0dSNeal Cardwell if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) 1957ae8064acSJohn Heffner goto send_now; 1958c1b4a7e6SDavid S. Miller } 1959c1b4a7e6SDavid S. Miller 196075c119afSEric Dumazet /* TODO : use tsorted_sent_queue ? */ 196175c119afSEric Dumazet head = tcp_rtx_queue_head(sk); 196275c119afSEric Dumazet if (!head) 196375c119afSEric Dumazet goto send_now; 19649a568de4SEric Dumazet age = tcp_stamp_us_delta(tp->tcp_mstamp, head->skb_mstamp); 196550c8339eSEric Dumazet /* If next ACK is likely to come too late (half srtt), do not defer */ 196650c8339eSEric Dumazet if (age < (tp->srtt_us >> 4)) 196750c8339eSEric Dumazet goto send_now; 196850c8339eSEric Dumazet 19695f852eb5SEric Dumazet /* Ok, it looks like it is advisable to defer. */ 1970ae8064acSJohn Heffner 1971d2e1339fSBendik Rønning Opstad if (cong_win < send_win && cong_win <= skb->len) 1972ca8a2263SNeal Cardwell *is_cwnd_limited = true; 1973ca8a2263SNeal Cardwell 1974a2a385d6SEric Dumazet return true; 1975ae8064acSJohn Heffner 1976ae8064acSJohn Heffner send_now: 1977a2a385d6SEric Dumazet return false; 1978c1b4a7e6SDavid S. Miller } 1979c1b4a7e6SDavid S. Miller 198005cbc0dbSFan Du static inline void tcp_mtu_check_reprobe(struct sock *sk) 198105cbc0dbSFan Du { 198205cbc0dbSFan Du struct inet_connection_sock *icsk = inet_csk(sk); 198305cbc0dbSFan Du struct tcp_sock *tp = tcp_sk(sk); 198405cbc0dbSFan Du struct net *net = sock_net(sk); 198505cbc0dbSFan Du u32 interval; 198605cbc0dbSFan Du s32 delta; 198705cbc0dbSFan Du 198805cbc0dbSFan Du interval = net->ipv4.sysctl_tcp_probe_interval; 1989c74df29aSEric Dumazet delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp; 199005cbc0dbSFan Du if (unlikely(delta >= interval * HZ)) { 199105cbc0dbSFan Du int mss = tcp_current_mss(sk); 199205cbc0dbSFan Du 199305cbc0dbSFan Du /* Update current search range */ 199405cbc0dbSFan Du icsk->icsk_mtup.probe_size = 0; 199505cbc0dbSFan Du icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + 199605cbc0dbSFan Du sizeof(struct tcphdr) + 199705cbc0dbSFan Du icsk->icsk_af_ops->net_header_len; 199805cbc0dbSFan Du icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss); 199905cbc0dbSFan Du 200005cbc0dbSFan Du /* Update probe time stamp */ 2001c74df29aSEric Dumazet icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; 200205cbc0dbSFan Du } 200305cbc0dbSFan Du } 200405cbc0dbSFan Du 20055d424d5aSJohn Heffner /* Create a new MTU probe if we are ready. 200667edfef7SAndi Kleen * MTU probe is regularly attempting to increase the path MTU by 200767edfef7SAndi Kleen * deliberately sending larger packets. This discovers routing 200867edfef7SAndi Kleen * changes resulting in larger path MTUs. 200967edfef7SAndi Kleen * 20105d424d5aSJohn Heffner * Returns 0 if we should wait to probe (no cwnd available), 20115d424d5aSJohn Heffner * 1 if a probe was sent, 2012056834d9SIlpo Järvinen * -1 otherwise 2013056834d9SIlpo Järvinen */ 20145d424d5aSJohn Heffner static int tcp_mtu_probe(struct sock *sk) 20155d424d5aSJohn Heffner { 20165d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 201712a59abcSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 20185d424d5aSJohn Heffner struct sk_buff *skb, *nskb, *next; 20196b58e0a5SFan Du struct net *net = sock_net(sk); 20205d424d5aSJohn Heffner int probe_size; 202191cc17c0SIlpo Järvinen int size_needed; 202212a59abcSEric Dumazet int copy, len; 20235d424d5aSJohn Heffner int mss_now; 20246b58e0a5SFan Du int interval; 20255d424d5aSJohn Heffner 20265d424d5aSJohn Heffner /* Not currently probing/verifying, 20275d424d5aSJohn Heffner * not in recovery, 20285d424d5aSJohn Heffner * have enough cwnd, and 202912a59abcSEric Dumazet * not SACKing (the variable headers throw things off) 203012a59abcSEric Dumazet */ 203112a59abcSEric Dumazet if (likely(!icsk->icsk_mtup.enabled || 20325d424d5aSJohn Heffner icsk->icsk_mtup.probe_size || 20335d424d5aSJohn Heffner inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 20345d424d5aSJohn Heffner tp->snd_cwnd < 11 || 203512a59abcSEric Dumazet tp->rx_opt.num_sacks || tp->rx_opt.dsack)) 20365d424d5aSJohn Heffner return -1; 20375d424d5aSJohn Heffner 20386b58e0a5SFan Du /* Use binary search for probe_size between tcp_mss_base, 20396b58e0a5SFan Du * and current mss_clamp. if (search_high - search_low) 20406b58e0a5SFan Du * smaller than a threshold, backoff from probing. 20416b58e0a5SFan Du */ 20420c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 20436b58e0a5SFan Du probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high + 20446b58e0a5SFan Du icsk->icsk_mtup.search_low) >> 1); 204591cc17c0SIlpo Järvinen size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; 20466b58e0a5SFan Du interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low; 204705cbc0dbSFan Du /* When misfortune happens, we are reprobing actively, 204805cbc0dbSFan Du * and then reprobe timer has expired. We stick with current 204905cbc0dbSFan Du * probing process by not resetting search range to its orignal. 205005cbc0dbSFan Du */ 20516b58e0a5SFan Du if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) || 205205cbc0dbSFan Du interval < net->ipv4.sysctl_tcp_probe_threshold) { 205305cbc0dbSFan Du /* Check whether enough time has elaplased for 205405cbc0dbSFan Du * another round of probing. 205505cbc0dbSFan Du */ 205605cbc0dbSFan Du tcp_mtu_check_reprobe(sk); 20575d424d5aSJohn Heffner return -1; 20585d424d5aSJohn Heffner } 20595d424d5aSJohn Heffner 20605d424d5aSJohn Heffner /* Have enough data in the send queue to probe? */ 20617f9c33e5SIlpo Järvinen if (tp->write_seq - tp->snd_nxt < size_needed) 20625d424d5aSJohn Heffner return -1; 20635d424d5aSJohn Heffner 206491cc17c0SIlpo Järvinen if (tp->snd_wnd < size_needed) 20655d424d5aSJohn Heffner return -1; 206690840defSIlpo Järvinen if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) 20675d424d5aSJohn Heffner return 0; 20685d424d5aSJohn Heffner 2069d67c58e9SIlpo Järvinen /* Do we need to wait to drain cwnd? With none in flight, don't stall */ 2070d67c58e9SIlpo Järvinen if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { 2071d67c58e9SIlpo Järvinen if (!tcp_packets_in_flight(tp)) 20725d424d5aSJohn Heffner return -1; 20735d424d5aSJohn Heffner else 20745d424d5aSJohn Heffner return 0; 20755d424d5aSJohn Heffner } 20765d424d5aSJohn Heffner 20775d424d5aSJohn Heffner /* We're allowed to probe. Build it now. */ 2078eb934478SEric Dumazet nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false); 207951456b29SIan Morris if (!nskb) 20805d424d5aSJohn Heffner return -1; 20813ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 20823ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 20835d424d5aSJohn Heffner 2084fe067e8aSDavid S. Miller skb = tcp_send_head(sk); 20855d424d5aSJohn Heffner 20865d424d5aSJohn Heffner TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 20875d424d5aSJohn Heffner TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 20884de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; 20895d424d5aSJohn Heffner TCP_SKB_CB(nskb)->sacked = 0; 20905d424d5aSJohn Heffner nskb->csum = 0; 209184fa7933SPatrick McHardy nskb->ip_summed = skb->ip_summed; 20925d424d5aSJohn Heffner 209350c4817eSIlpo Järvinen tcp_insert_write_queue_before(nskb, skb, sk); 209450c4817eSIlpo Järvinen 20955d424d5aSJohn Heffner len = 0; 2096234b6860SIlpo Järvinen tcp_for_write_queue_from_safe(skb, next, sk) { 20975d424d5aSJohn Heffner copy = min_t(int, skb->len, probe_size - len); 20982fe664f1SDouglas Caetano dos Santos if (nskb->ip_summed) { 20995d424d5aSJohn Heffner skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); 21002fe664f1SDouglas Caetano dos Santos } else { 21012fe664f1SDouglas Caetano dos Santos __wsum csum = skb_copy_and_csum_bits(skb, 0, 2102056834d9SIlpo Järvinen skb_put(nskb, copy), 21032fe664f1SDouglas Caetano dos Santos copy, 0); 21042fe664f1SDouglas Caetano dos Santos nskb->csum = csum_block_add(nskb->csum, csum, len); 21052fe664f1SDouglas Caetano dos Santos } 21065d424d5aSJohn Heffner 21075d424d5aSJohn Heffner if (skb->len <= copy) { 21085d424d5aSJohn Heffner /* We've eaten all the data from this skb. 21095d424d5aSJohn Heffner * Throw it away. */ 21104de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 2111fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 21123ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 21135d424d5aSJohn Heffner } else { 21144de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & 2115a3433f35SChangli Gao ~(TCPHDR_FIN|TCPHDR_PSH); 21165d424d5aSJohn Heffner if (!skb_shinfo(skb)->nr_frags) { 21175d424d5aSJohn Heffner skb_pull(skb, copy); 211884fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 2119056834d9SIlpo Järvinen skb->csum = csum_partial(skb->data, 2120056834d9SIlpo Järvinen skb->len, 0); 21215d424d5aSJohn Heffner } else { 21225d424d5aSJohn Heffner __pskb_trim_head(skb, copy); 21235bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 21245d424d5aSJohn Heffner } 21255d424d5aSJohn Heffner TCP_SKB_CB(skb)->seq += copy; 21265d424d5aSJohn Heffner } 21275d424d5aSJohn Heffner 21285d424d5aSJohn Heffner len += copy; 2129234b6860SIlpo Järvinen 2130234b6860SIlpo Järvinen if (len >= probe_size) 2131234b6860SIlpo Järvinen break; 21325d424d5aSJohn Heffner } 21335bbb432cSEric Dumazet tcp_init_tso_segs(nskb, nskb->len); 21345d424d5aSJohn Heffner 21355d424d5aSJohn Heffner /* We're ready to send. If this fails, the probe will 21367faee5c0SEric Dumazet * be resegmented into mss-sized pieces by tcp_write_xmit(). 21377faee5c0SEric Dumazet */ 21385d424d5aSJohn Heffner if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 21395d424d5aSJohn Heffner /* Decrement cwnd here because we are sending 21405d424d5aSJohn Heffner * effectively two packets. */ 21415d424d5aSJohn Heffner tp->snd_cwnd--; 214266f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, nskb); 21435d424d5aSJohn Heffner 21445d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 21450e7b1368SJohn Heffner tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 21460e7b1368SJohn Heffner tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 21475d424d5aSJohn Heffner 21485d424d5aSJohn Heffner return 1; 21495d424d5aSJohn Heffner } 21505d424d5aSJohn Heffner 21515d424d5aSJohn Heffner return -1; 21525d424d5aSJohn Heffner } 21535d424d5aSJohn Heffner 2154218af599SEric Dumazet static bool tcp_pacing_check(const struct sock *sk) 2155218af599SEric Dumazet { 2156218af599SEric Dumazet return tcp_needs_internal_pacing(sk) && 2157218af599SEric Dumazet hrtimer_active(&tcp_sk(sk)->pacing_timer); 2158218af599SEric Dumazet } 2159218af599SEric Dumazet 2160f9616c35SEric Dumazet /* TCP Small Queues : 2161f9616c35SEric Dumazet * Control number of packets in qdisc/devices to two packets / or ~1 ms. 2162f9616c35SEric Dumazet * (These limits are doubled for retransmits) 2163f9616c35SEric Dumazet * This allows for : 2164f9616c35SEric Dumazet * - better RTT estimation and ACK scheduling 2165f9616c35SEric Dumazet * - faster recovery 2166f9616c35SEric Dumazet * - high rates 2167f9616c35SEric Dumazet * Alas, some drivers / subsystems require a fair amount 2168f9616c35SEric Dumazet * of queued bytes to ensure line rate. 2169f9616c35SEric Dumazet * One example is wifi aggregation (802.11 AMPDU) 2170f9616c35SEric Dumazet */ 2171f9616c35SEric Dumazet static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb, 2172f9616c35SEric Dumazet unsigned int factor) 2173f9616c35SEric Dumazet { 2174f9616c35SEric Dumazet unsigned int limit; 2175f9616c35SEric Dumazet 2176f9616c35SEric Dumazet limit = max(2 * skb->truesize, sk->sk_pacing_rate >> 10); 2177f9616c35SEric Dumazet limit = min_t(u32, limit, sysctl_tcp_limit_output_bytes); 2178f9616c35SEric Dumazet limit <<= factor; 2179f9616c35SEric Dumazet 218014afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) > limit) { 218175c119afSEric Dumazet /* Always send skb if rtx queue is empty. 218275eefc6cSEric Dumazet * No need to wait for TX completion to call us back, 218375eefc6cSEric Dumazet * after softirq/tasklet schedule. 218475eefc6cSEric Dumazet * This helps when TX completions are delayed too much. 218575eefc6cSEric Dumazet */ 218675c119afSEric Dumazet if (tcp_rtx_queue_empty(sk)) 218775eefc6cSEric Dumazet return false; 218875eefc6cSEric Dumazet 21897aa5470cSEric Dumazet set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags); 2190f9616c35SEric Dumazet /* It is possible TX completion already happened 2191f9616c35SEric Dumazet * before we set TSQ_THROTTLED, so we must 2192f9616c35SEric Dumazet * test again the condition. 2193f9616c35SEric Dumazet */ 2194f9616c35SEric Dumazet smp_mb__after_atomic(); 219514afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) > limit) 2196f9616c35SEric Dumazet return true; 2197f9616c35SEric Dumazet } 2198f9616c35SEric Dumazet return false; 2199f9616c35SEric Dumazet } 2200f9616c35SEric Dumazet 220105b055e8SFrancis Yan static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new) 220205b055e8SFrancis Yan { 2203628174ccSEric Dumazet const u32 now = tcp_jiffies32; 2204efe967cdSArnd Bergmann enum tcp_chrono old = tp->chrono_type; 220505b055e8SFrancis Yan 2206efe967cdSArnd Bergmann if (old > TCP_CHRONO_UNSPEC) 2207efe967cdSArnd Bergmann tp->chrono_stat[old - 1] += now - tp->chrono_start; 220805b055e8SFrancis Yan tp->chrono_start = now; 220905b055e8SFrancis Yan tp->chrono_type = new; 221005b055e8SFrancis Yan } 221105b055e8SFrancis Yan 221205b055e8SFrancis Yan void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type) 221305b055e8SFrancis Yan { 221405b055e8SFrancis Yan struct tcp_sock *tp = tcp_sk(sk); 221505b055e8SFrancis Yan 221605b055e8SFrancis Yan /* If there are multiple conditions worthy of tracking in a 22170f87230dSFrancis Yan * chronograph then the highest priority enum takes precedence 22180f87230dSFrancis Yan * over the other conditions. So that if something "more interesting" 221905b055e8SFrancis Yan * starts happening, stop the previous chrono and start a new one. 222005b055e8SFrancis Yan */ 222105b055e8SFrancis Yan if (type > tp->chrono_type) 222205b055e8SFrancis Yan tcp_chrono_set(tp, type); 222305b055e8SFrancis Yan } 222405b055e8SFrancis Yan 222505b055e8SFrancis Yan void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type) 222605b055e8SFrancis Yan { 222705b055e8SFrancis Yan struct tcp_sock *tp = tcp_sk(sk); 222805b055e8SFrancis Yan 22290f87230dSFrancis Yan 22300f87230dSFrancis Yan /* There are multiple conditions worthy of tracking in a 22310f87230dSFrancis Yan * chronograph, so that the highest priority enum takes 22320f87230dSFrancis Yan * precedence over the other conditions (see tcp_chrono_start). 22330f87230dSFrancis Yan * If a condition stops, we only stop chrono tracking if 22340f87230dSFrancis Yan * it's the "most interesting" or current chrono we are 22350f87230dSFrancis Yan * tracking and starts busy chrono if we have pending data. 22360f87230dSFrancis Yan */ 223775c119afSEric Dumazet if (tcp_rtx_and_write_queues_empty(sk)) 223805b055e8SFrancis Yan tcp_chrono_set(tp, TCP_CHRONO_UNSPEC); 22390f87230dSFrancis Yan else if (type == tp->chrono_type) 22400f87230dSFrancis Yan tcp_chrono_set(tp, TCP_CHRONO_BUSY); 224105b055e8SFrancis Yan } 224205b055e8SFrancis Yan 22431da177e4SLinus Torvalds /* This routine writes packets to the network. It advances the 22441da177e4SLinus Torvalds * send_head. This happens as incoming acks open up the remote 22451da177e4SLinus Torvalds * window for us. 22461da177e4SLinus Torvalds * 2247f8269a49SIlpo Järvinen * LARGESEND note: !tcp_urg_mode is overkill, only frames between 2248f8269a49SIlpo Järvinen * snd_up-64k-mss .. snd_up cannot be large. However, taking into 2249f8269a49SIlpo Järvinen * account rare use of URG, this is not a big flaw. 2250f8269a49SIlpo Järvinen * 22516ba8a3b1SNandita Dukkipati * Send at most one packet when push_one > 0. Temporarily ignore 22526ba8a3b1SNandita Dukkipati * cwnd limit to force at most one packet out when push_one == 2. 22536ba8a3b1SNandita Dukkipati 2254a2a385d6SEric Dumazet * Returns true, if no segments are in flight and we have queued segments, 2255a2a385d6SEric Dumazet * but cannot send anything now because of SWS or another problem. 22561da177e4SLinus Torvalds */ 2257a2a385d6SEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 2258d5dd9175SIlpo Järvinen int push_one, gfp_t gfp) 22591da177e4SLinus Torvalds { 22601da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 226192df7b51SDavid S. Miller struct sk_buff *skb; 2262c1b4a7e6SDavid S. Miller unsigned int tso_segs, sent_pkts; 2263c1b4a7e6SDavid S. Miller int cwnd_quota; 22645d424d5aSJohn Heffner int result; 22655615f886SFrancis Yan bool is_cwnd_limited = false, is_rwnd_limited = false; 2266605ad7f1SEric Dumazet u32 max_segs; 22671da177e4SLinus Torvalds 2268c1b4a7e6SDavid S. Miller sent_pkts = 0; 22695d424d5aSJohn Heffner 2270d5dd9175SIlpo Järvinen if (!push_one) { 22715d424d5aSJohn Heffner /* Do MTU probing. */ 2272d5dd9175SIlpo Järvinen result = tcp_mtu_probe(sk); 2273d5dd9175SIlpo Järvinen if (!result) { 2274a2a385d6SEric Dumazet return false; 22755d424d5aSJohn Heffner } else if (result > 0) { 22765d424d5aSJohn Heffner sent_pkts = 1; 22775d424d5aSJohn Heffner } 2278d5dd9175SIlpo Järvinen } 22795d424d5aSJohn Heffner 2280ed6e7268SNeal Cardwell max_segs = tcp_tso_segs(sk, mss_now); 22819a568de4SEric Dumazet tcp_mstamp_refresh(tp); 2282fe067e8aSDavid S. Miller while ((skb = tcp_send_head(sk))) { 2283c8ac3774SHerbert Xu unsigned int limit; 2284c8ac3774SHerbert Xu 2285218af599SEric Dumazet if (tcp_pacing_check(sk)) 2286218af599SEric Dumazet break; 2287218af599SEric Dumazet 22885bbb432cSEric Dumazet tso_segs = tcp_init_tso_segs(skb, mss_now); 2289c1b4a7e6SDavid S. Miller BUG_ON(!tso_segs); 2290c1b4a7e6SDavid S. Miller 22919d186cacSAndrey Vagin if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) { 22927faee5c0SEric Dumazet /* "skb_mstamp" is used as a start point for the retransmit timer */ 2293e2080072SEric Dumazet tcp_update_skb_after_send(tp, skb); 2294ec342325SAndrew Vagin goto repair; /* Skip network transmission */ 22959d186cacSAndrey Vagin } 2296ec342325SAndrew Vagin 2297b68e9f85SHerbert Xu cwnd_quota = tcp_cwnd_test(tp, skb); 22986ba8a3b1SNandita Dukkipati if (!cwnd_quota) { 22996ba8a3b1SNandita Dukkipati if (push_one == 2) 23006ba8a3b1SNandita Dukkipati /* Force out a loss probe pkt. */ 23016ba8a3b1SNandita Dukkipati cwnd_quota = 1; 23026ba8a3b1SNandita Dukkipati else 2303b68e9f85SHerbert Xu break; 23046ba8a3b1SNandita Dukkipati } 2305b68e9f85SHerbert Xu 23065615f886SFrancis Yan if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) { 23075615f886SFrancis Yan is_rwnd_limited = true; 2308b68e9f85SHerbert Xu break; 23095615f886SFrancis Yan } 2310b68e9f85SHerbert Xu 2311d6a4e26aSEric Dumazet if (tso_segs == 1) { 2312aa93466bSDavid S. Miller if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 2313aa93466bSDavid S. Miller (tcp_skb_is_last(sk, skb) ? 2314aa93466bSDavid S. Miller nonagle : TCP_NAGLE_PUSH)))) 2315aa93466bSDavid S. Miller break; 2316c1b4a7e6SDavid S. Miller } else { 2317ca8a2263SNeal Cardwell if (!push_one && 2318605ad7f1SEric Dumazet tcp_tso_should_defer(sk, skb, &is_cwnd_limited, 2319605ad7f1SEric Dumazet max_segs)) 2320aa93466bSDavid S. Miller break; 2321c1b4a7e6SDavid S. Miller } 2322aa93466bSDavid S. Miller 2323605ad7f1SEric Dumazet limit = mss_now; 2324d6a4e26aSEric Dumazet if (tso_segs > 1 && !tcp_urg_mode(tp)) 2325605ad7f1SEric Dumazet limit = tcp_mss_split_point(sk, skb, mss_now, 2326605ad7f1SEric Dumazet min_t(unsigned int, 2327605ad7f1SEric Dumazet cwnd_quota, 2328605ad7f1SEric Dumazet max_segs), 2329605ad7f1SEric Dumazet nonagle); 2330605ad7f1SEric Dumazet 2331605ad7f1SEric Dumazet if (skb->len > limit && 233275c119afSEric Dumazet unlikely(tso_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE, 233375c119afSEric Dumazet skb, limit, mss_now, gfp))) 2334605ad7f1SEric Dumazet break; 2335605ad7f1SEric Dumazet 23367aa5470cSEric Dumazet if (test_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags)) 23377aa5470cSEric Dumazet clear_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags); 2338f9616c35SEric Dumazet if (tcp_small_queue_check(sk, skb, 0)) 233946d3ceabSEric Dumazet break; 2340c9eeec26SEric Dumazet 2341d5dd9175SIlpo Järvinen if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) 23421da177e4SLinus Torvalds break; 23431da177e4SLinus Torvalds 2344ec342325SAndrew Vagin repair: 23451da177e4SLinus Torvalds /* Advance the send_head. This one is sent out. 23461da177e4SLinus Torvalds * This call will increment packets_out. 23471da177e4SLinus Torvalds */ 234866f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 23491da177e4SLinus Torvalds 23501da177e4SLinus Torvalds tcp_minshall_update(tp, mss_now, skb); 2351a262f0cdSNandita Dukkipati sent_pkts += tcp_skb_pcount(skb); 2352d5dd9175SIlpo Järvinen 2353d5dd9175SIlpo Järvinen if (push_one) 2354d5dd9175SIlpo Järvinen break; 23551da177e4SLinus Torvalds } 23561da177e4SLinus Torvalds 23575615f886SFrancis Yan if (is_rwnd_limited) 23585615f886SFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED); 23595615f886SFrancis Yan else 23605615f886SFrancis Yan tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED); 23615615f886SFrancis Yan 2362aa93466bSDavid S. Miller if (likely(sent_pkts)) { 2363684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 2364684bad11SYuchung Cheng tp->prr_out += sent_pkts; 23656ba8a3b1SNandita Dukkipati 23666ba8a3b1SNandita Dukkipati /* Send one loss probe per tail loss episode. */ 23676ba8a3b1SNandita Dukkipati if (push_one != 2) 23686ba8a3b1SNandita Dukkipati tcp_schedule_loss_probe(sk); 2369d2e1339fSBendik Rønning Opstad is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd); 2370ca8a2263SNeal Cardwell tcp_cwnd_validate(sk, is_cwnd_limited); 2371a2a385d6SEric Dumazet return false; 23721da177e4SLinus Torvalds } 237375c119afSEric Dumazet return !tp->packets_out && !tcp_write_queue_empty(sk); 23746ba8a3b1SNandita Dukkipati } 23756ba8a3b1SNandita Dukkipati 23766ba8a3b1SNandita Dukkipati bool tcp_schedule_loss_probe(struct sock *sk) 23776ba8a3b1SNandita Dukkipati { 23786ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 23796ba8a3b1SNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 2380a2815817SNeal Cardwell u32 timeout, rto_delta_us; 23816ba8a3b1SNandita Dukkipati 23826ba8a3b1SNandita Dukkipati /* Don't do any loss probe on a Fast Open connection before 3WHS 23836ba8a3b1SNandita Dukkipati * finishes. 23846ba8a3b1SNandita Dukkipati */ 2385f9b99582SYuchung Cheng if (tp->fastopen_rsk) 23866ba8a3b1SNandita Dukkipati return false; 23876ba8a3b1SNandita Dukkipati 23886ba8a3b1SNandita Dukkipati /* Schedule a loss probe in 2*RTT for SACK capable connections 23896ba8a3b1SNandita Dukkipati * in Open state, that are either limited by cwnd or application. 23906ba8a3b1SNandita Dukkipati */ 2391bec41a11SYuchung Cheng if ((sysctl_tcp_early_retrans != 3 && sysctl_tcp_early_retrans != 4) || 2392bec41a11SYuchung Cheng !tp->packets_out || !tcp_is_sack(tp) || 2393bec41a11SYuchung Cheng icsk->icsk_ca_state != TCP_CA_Open) 23946ba8a3b1SNandita Dukkipati return false; 23956ba8a3b1SNandita Dukkipati 23966ba8a3b1SNandita Dukkipati if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) && 239775c119afSEric Dumazet !tcp_write_queue_empty(sk)) 23986ba8a3b1SNandita Dukkipati return false; 23996ba8a3b1SNandita Dukkipati 2400bb4d991aSYuchung Cheng /* Probe timeout is 2*rtt. Add minimum RTO to account 2401f9b99582SYuchung Cheng * for delayed ack when there's one outstanding packet. If no RTT 2402f9b99582SYuchung Cheng * sample is available then probe after TCP_TIMEOUT_INIT. 24036ba8a3b1SNandita Dukkipati */ 2404bb4d991aSYuchung Cheng if (tp->srtt_us) { 2405bb4d991aSYuchung Cheng timeout = usecs_to_jiffies(tp->srtt_us >> 2); 24066ba8a3b1SNandita Dukkipati if (tp->packets_out == 1) 2407bb4d991aSYuchung Cheng timeout += TCP_RTO_MIN; 2408bb4d991aSYuchung Cheng else 2409bb4d991aSYuchung Cheng timeout += TCP_TIMEOUT_MIN; 2410bb4d991aSYuchung Cheng } else { 2411bb4d991aSYuchung Cheng timeout = TCP_TIMEOUT_INIT; 2412bb4d991aSYuchung Cheng } 24136ba8a3b1SNandita Dukkipati 2414a2815817SNeal Cardwell /* If the RTO formula yields an earlier time, then use that time. */ 2415a2815817SNeal Cardwell rto_delta_us = tcp_rto_delta_us(sk); /* How far in future is RTO? */ 2416a2815817SNeal Cardwell if (rto_delta_us > 0) 2417a2815817SNeal Cardwell timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us)); 24186ba8a3b1SNandita Dukkipati 24196ba8a3b1SNandita Dukkipati inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, 24206ba8a3b1SNandita Dukkipati TCP_RTO_MAX); 24216ba8a3b1SNandita Dukkipati return true; 24226ba8a3b1SNandita Dukkipati } 24236ba8a3b1SNandita Dukkipati 24241f3279aeSEric Dumazet /* Thanks to skb fast clones, we can detect if a prior transmit of 24251f3279aeSEric Dumazet * a packet is still in a qdisc or driver queue. 24261f3279aeSEric Dumazet * In this case, there is very little point doing a retransmit ! 24271f3279aeSEric Dumazet */ 24281f3279aeSEric Dumazet static bool skb_still_in_host_queue(const struct sock *sk, 24291f3279aeSEric Dumazet const struct sk_buff *skb) 24301f3279aeSEric Dumazet { 243139bb5e62SEric Dumazet if (unlikely(skb_fclone_busy(sk, skb))) { 2432c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), 24331f3279aeSEric Dumazet LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES); 24341f3279aeSEric Dumazet return true; 24351f3279aeSEric Dumazet } 24361f3279aeSEric Dumazet return false; 24371f3279aeSEric Dumazet } 24381f3279aeSEric Dumazet 2439b340b264SYuchung Cheng /* When probe timeout (PTO) fires, try send a new segment if possible, else 24406ba8a3b1SNandita Dukkipati * retransmit the last segment. 24416ba8a3b1SNandita Dukkipati */ 24426ba8a3b1SNandita Dukkipati void tcp_send_loss_probe(struct sock *sk) 24436ba8a3b1SNandita Dukkipati { 24449b717a8dSNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 24456ba8a3b1SNandita Dukkipati struct sk_buff *skb; 24466ba8a3b1SNandita Dukkipati int pcount; 24476ba8a3b1SNandita Dukkipati int mss = tcp_current_mss(sk); 24486ba8a3b1SNandita Dukkipati 2449b340b264SYuchung Cheng skb = tcp_send_head(sk); 245075c119afSEric Dumazet if (skb && tcp_snd_wnd_test(tp, skb, mss)) { 2451b340b264SYuchung Cheng pcount = tp->packets_out; 2452b340b264SYuchung Cheng tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC); 2453b340b264SYuchung Cheng if (tp->packets_out > pcount) 2454b340b264SYuchung Cheng goto probe_sent; 24556ba8a3b1SNandita Dukkipati goto rearm_timer; 24566ba8a3b1SNandita Dukkipati } 245775c119afSEric Dumazet skb = skb_rb_last(&sk->tcp_rtx_queue); 24586ba8a3b1SNandita Dukkipati 24599b717a8dSNandita Dukkipati /* At most one outstanding TLP retransmission. */ 24609b717a8dSNandita Dukkipati if (tp->tlp_high_seq) 24619b717a8dSNandita Dukkipati goto rearm_timer; 24629b717a8dSNandita Dukkipati 24636ba8a3b1SNandita Dukkipati /* Retransmit last segment. */ 24646ba8a3b1SNandita Dukkipati if (WARN_ON(!skb)) 24656ba8a3b1SNandita Dukkipati goto rearm_timer; 24666ba8a3b1SNandita Dukkipati 24671f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 24681f3279aeSEric Dumazet goto rearm_timer; 24691f3279aeSEric Dumazet 24706ba8a3b1SNandita Dukkipati pcount = tcp_skb_pcount(skb); 24716ba8a3b1SNandita Dukkipati if (WARN_ON(!pcount)) 24726ba8a3b1SNandita Dukkipati goto rearm_timer; 24736ba8a3b1SNandita Dukkipati 24746ba8a3b1SNandita Dukkipati if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) { 247575c119afSEric Dumazet if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, 247675c119afSEric Dumazet (pcount - 1) * mss, mss, 24776cc55e09SOctavian Purdila GFP_ATOMIC))) 24786ba8a3b1SNandita Dukkipati goto rearm_timer; 247975c119afSEric Dumazet skb = skb_rb_next(skb); 24806ba8a3b1SNandita Dukkipati } 24816ba8a3b1SNandita Dukkipati 24826ba8a3b1SNandita Dukkipati if (WARN_ON(!skb || !tcp_skb_pcount(skb))) 24836ba8a3b1SNandita Dukkipati goto rearm_timer; 24846ba8a3b1SNandita Dukkipati 248510d3be56SEric Dumazet if (__tcp_retransmit_skb(sk, skb, 1)) 2486b340b264SYuchung Cheng goto rearm_timer; 24876ba8a3b1SNandita Dukkipati 24889b717a8dSNandita Dukkipati /* Record snd_nxt for loss detection. */ 24899b717a8dSNandita Dukkipati tp->tlp_high_seq = tp->snd_nxt; 24909b717a8dSNandita Dukkipati 2491b340b264SYuchung Cheng probe_sent: 2492c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES); 2493fcd16c0aSYuchung Cheng /* Reset s.t. tcp_rearm_rto will restart timer from now */ 2494fcd16c0aSYuchung Cheng inet_csk(sk)->icsk_pending = 0; 2495b340b264SYuchung Cheng rearm_timer: 2496fcd16c0aSYuchung Cheng tcp_rearm_rto(sk); 24971da177e4SLinus Torvalds } 24981da177e4SLinus Torvalds 2499a762a980SDavid S. Miller /* Push out any pending frames which were held back due to 2500a762a980SDavid S. Miller * TCP_CORK or attempt at coalescing tiny packets. 2501a762a980SDavid S. Miller * The socket must be locked by the caller. 2502a762a980SDavid S. Miller */ 25039e412ba7SIlpo Järvinen void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 25049e412ba7SIlpo Järvinen int nonagle) 2505a762a980SDavid S. Miller { 2506726e07a8SIlpo Järvinen /* If we are closed, the bytes will have to remain here. 2507726e07a8SIlpo Järvinen * In time closedown will finish, we empty the write queue and 2508726e07a8SIlpo Järvinen * all will be happy. 2509726e07a8SIlpo Järvinen */ 2510726e07a8SIlpo Järvinen if (unlikely(sk->sk_state == TCP_CLOSE)) 2511726e07a8SIlpo Järvinen return; 2512726e07a8SIlpo Järvinen 251399a1dec7SMel Gorman if (tcp_write_xmit(sk, cur_mss, nonagle, 0, 25147450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC))) 25159e412ba7SIlpo Järvinen tcp_check_probe_timer(sk); 2516a762a980SDavid S. Miller } 2517a762a980SDavid S. Miller 2518c1b4a7e6SDavid S. Miller /* Send _single_ skb sitting at the send head. This function requires 2519c1b4a7e6SDavid S. Miller * true push pending frames to setup probe timer etc. 2520c1b4a7e6SDavid S. Miller */ 2521c1b4a7e6SDavid S. Miller void tcp_push_one(struct sock *sk, unsigned int mss_now) 2522c1b4a7e6SDavid S. Miller { 2523fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 2524c1b4a7e6SDavid S. Miller 2525c1b4a7e6SDavid S. Miller BUG_ON(!skb || skb->len < mss_now); 2526c1b4a7e6SDavid S. Miller 2527d5dd9175SIlpo Järvinen tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); 2528c1b4a7e6SDavid S. Miller } 2529c1b4a7e6SDavid S. Miller 25301da177e4SLinus Torvalds /* This function returns the amount that we can raise the 25311da177e4SLinus Torvalds * usable window based on the following constraints 25321da177e4SLinus Torvalds * 25331da177e4SLinus Torvalds * 1. The window can never be shrunk once it is offered (RFC 793) 25341da177e4SLinus Torvalds * 2. We limit memory per socket 25351da177e4SLinus Torvalds * 25361da177e4SLinus Torvalds * RFC 1122: 25371da177e4SLinus Torvalds * "the suggested [SWS] avoidance algorithm for the receiver is to keep 25381da177e4SLinus Torvalds * RECV.NEXT + RCV.WIN fixed until: 25391da177e4SLinus Torvalds * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 25401da177e4SLinus Torvalds * 25411da177e4SLinus Torvalds * i.e. don't raise the right edge of the window until you can raise 25421da177e4SLinus Torvalds * it at least MSS bytes. 25431da177e4SLinus Torvalds * 25441da177e4SLinus Torvalds * Unfortunately, the recommended algorithm breaks header prediction, 25451da177e4SLinus Torvalds * since header prediction assumes th->window stays fixed. 25461da177e4SLinus Torvalds * 25471da177e4SLinus Torvalds * Strictly speaking, keeping th->window fixed violates the receiver 25481da177e4SLinus Torvalds * side SWS prevention criteria. The problem is that under this rule 25491da177e4SLinus Torvalds * a stream of single byte packets will cause the right side of the 25501da177e4SLinus Torvalds * window to always advance by a single byte. 25511da177e4SLinus Torvalds * 25521da177e4SLinus Torvalds * Of course, if the sender implements sender side SWS prevention 25531da177e4SLinus Torvalds * then this will not be a problem. 25541da177e4SLinus Torvalds * 25551da177e4SLinus Torvalds * BSD seems to make the following compromise: 25561da177e4SLinus Torvalds * 25571da177e4SLinus Torvalds * If the free space is less than the 1/4 of the maximum 25581da177e4SLinus Torvalds * space available and the free space is less than 1/2 mss, 25591da177e4SLinus Torvalds * then set the window to 0. 25601da177e4SLinus Torvalds * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 25611da177e4SLinus Torvalds * Otherwise, just prevent the window from shrinking 25621da177e4SLinus Torvalds * and from being larger than the largest representable value. 25631da177e4SLinus Torvalds * 25641da177e4SLinus Torvalds * This prevents incremental opening of the window in the regime 25651da177e4SLinus Torvalds * where TCP is limited by the speed of the reader side taking 25661da177e4SLinus Torvalds * data out of the TCP receive queue. It does nothing about 25671da177e4SLinus Torvalds * those cases where the window is constrained on the sender side 25681da177e4SLinus Torvalds * because the pipeline is full. 25691da177e4SLinus Torvalds * 25701da177e4SLinus Torvalds * BSD also seems to "accidentally" limit itself to windows that are a 25711da177e4SLinus Torvalds * multiple of MSS, at least until the free space gets quite small. 25721da177e4SLinus Torvalds * This would appear to be a side effect of the mbuf implementation. 25731da177e4SLinus Torvalds * Combining these two algorithms results in the observed behavior 25741da177e4SLinus Torvalds * of having a fixed window size at almost all times. 25751da177e4SLinus Torvalds * 25761da177e4SLinus Torvalds * Below we obtain similar behavior by forcing the offered window to 25771da177e4SLinus Torvalds * a multiple of the mss when it is feasible to do so. 25781da177e4SLinus Torvalds * 25791da177e4SLinus Torvalds * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 25801da177e4SLinus Torvalds * Regular options like TIMESTAMP are taken into account. 25811da177e4SLinus Torvalds */ 25821da177e4SLinus Torvalds u32 __tcp_select_window(struct sock *sk) 25831da177e4SLinus Torvalds { 2584463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 25851da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2586caa20d9aSStephen Hemminger /* MSS for the peer's data. Previous versions used mss_clamp 25871da177e4SLinus Torvalds * here. I don't know if the value based on our guesses 25881da177e4SLinus Torvalds * of peer's MSS is better for the performance. It's more correct 25891da177e4SLinus Torvalds * but may be worse for the performance because of rcv_mss 25901da177e4SLinus Torvalds * fluctuations. --SAW 1998/11/1 25911da177e4SLinus Torvalds */ 2592463c84b9SArnaldo Carvalho de Melo int mss = icsk->icsk_ack.rcv_mss; 25931da177e4SLinus Torvalds int free_space = tcp_space(sk); 259486c1a045SFlorian Westphal int allowed_space = tcp_full_space(sk); 259586c1a045SFlorian Westphal int full_space = min_t(int, tp->window_clamp, allowed_space); 25961da177e4SLinus Torvalds int window; 25971da177e4SLinus Torvalds 259806425c30SEric Dumazet if (unlikely(mss > full_space)) { 25991da177e4SLinus Torvalds mss = full_space; 260006425c30SEric Dumazet if (mss <= 0) 260106425c30SEric Dumazet return 0; 260206425c30SEric Dumazet } 2603b92edbe0SEric Dumazet if (free_space < (full_space >> 1)) { 2604463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.quick = 0; 26051da177e4SLinus Torvalds 2606b8da51ebSEric Dumazet if (tcp_under_memory_pressure(sk)) 2607056834d9SIlpo Järvinen tp->rcv_ssthresh = min(tp->rcv_ssthresh, 2608056834d9SIlpo Järvinen 4U * tp->advmss); 26091da177e4SLinus Torvalds 261086c1a045SFlorian Westphal /* free_space might become our new window, make sure we don't 261186c1a045SFlorian Westphal * increase it due to wscale. 261286c1a045SFlorian Westphal */ 261386c1a045SFlorian Westphal free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale); 261486c1a045SFlorian Westphal 261586c1a045SFlorian Westphal /* if free space is less than mss estimate, or is below 1/16th 261686c1a045SFlorian Westphal * of the maximum allowed, try to move to zero-window, else 261786c1a045SFlorian Westphal * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and 261886c1a045SFlorian Westphal * new incoming data is dropped due to memory limits. 261986c1a045SFlorian Westphal * With large window, mss test triggers way too late in order 262086c1a045SFlorian Westphal * to announce zero window in time before rmem limit kicks in. 262186c1a045SFlorian Westphal */ 262286c1a045SFlorian Westphal if (free_space < (allowed_space >> 4) || free_space < mss) 26231da177e4SLinus Torvalds return 0; 26241da177e4SLinus Torvalds } 26251da177e4SLinus Torvalds 26261da177e4SLinus Torvalds if (free_space > tp->rcv_ssthresh) 26271da177e4SLinus Torvalds free_space = tp->rcv_ssthresh; 26281da177e4SLinus Torvalds 26291da177e4SLinus Torvalds /* Don't do rounding if we are using window scaling, since the 26301da177e4SLinus Torvalds * scaled window will not line up with the MSS boundary anyway. 26311da177e4SLinus Torvalds */ 26321da177e4SLinus Torvalds if (tp->rx_opt.rcv_wscale) { 26331da177e4SLinus Torvalds window = free_space; 26341da177e4SLinus Torvalds 26351da177e4SLinus Torvalds /* Advertise enough space so that it won't get scaled away. 26361da177e4SLinus Torvalds * Import case: prevent zero window announcement if 26371da177e4SLinus Torvalds * 1<<rcv_wscale > mss. 26381da177e4SLinus Torvalds */ 26391935299dSGao Feng window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale)); 26401da177e4SLinus Torvalds } else { 26411935299dSGao Feng window = tp->rcv_wnd; 26421da177e4SLinus Torvalds /* Get the largest window that is a nice multiple of mss. 26431da177e4SLinus Torvalds * Window clamp already applied above. 26441da177e4SLinus Torvalds * If our current window offering is within 1 mss of the 26451da177e4SLinus Torvalds * free space we just keep it. This prevents the divide 26461da177e4SLinus Torvalds * and multiply from happening most of the time. 26471da177e4SLinus Torvalds * We also don't do any window rounding when the free space 26481da177e4SLinus Torvalds * is too small. 26491da177e4SLinus Torvalds */ 26501da177e4SLinus Torvalds if (window <= free_space - mss || window > free_space) 26511935299dSGao Feng window = rounddown(free_space, mss); 265284565070SJohn Heffner else if (mss == full_space && 2653b92edbe0SEric Dumazet free_space > window + (full_space >> 1)) 265484565070SJohn Heffner window = free_space; 26551da177e4SLinus Torvalds } 26561da177e4SLinus Torvalds 26571da177e4SLinus Torvalds return window; 26581da177e4SLinus Torvalds } 26591da177e4SLinus Torvalds 2660cfea5a68SMartin KaFai Lau void tcp_skb_collapse_tstamp(struct sk_buff *skb, 2661082ac2d5SMartin KaFai Lau const struct sk_buff *next_skb) 2662082ac2d5SMartin KaFai Lau { 26630a2cf20cSSoheil Hassas Yeganeh if (unlikely(tcp_has_tx_tstamp(next_skb))) { 26640a2cf20cSSoheil Hassas Yeganeh const struct skb_shared_info *next_shinfo = 26650a2cf20cSSoheil Hassas Yeganeh skb_shinfo(next_skb); 2666082ac2d5SMartin KaFai Lau struct skb_shared_info *shinfo = skb_shinfo(skb); 2667082ac2d5SMartin KaFai Lau 26680a2cf20cSSoheil Hassas Yeganeh shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP; 2669082ac2d5SMartin KaFai Lau shinfo->tskey = next_shinfo->tskey; 26702de8023eSMartin KaFai Lau TCP_SKB_CB(skb)->txstamp_ack |= 26712de8023eSMartin KaFai Lau TCP_SKB_CB(next_skb)->txstamp_ack; 2672082ac2d5SMartin KaFai Lau } 2673082ac2d5SMartin KaFai Lau } 2674082ac2d5SMartin KaFai Lau 26754a17fc3aSIlpo Järvinen /* Collapses two adjacent SKB's during retransmission. */ 2676f8071cdeSEric Dumazet static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) 26771da177e4SLinus Torvalds { 26781da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 267975c119afSEric Dumazet struct sk_buff *next_skb = skb_rb_next(skb); 2680058dc334SIlpo Järvinen int skb_size, next_skb_size; 26811da177e4SLinus Torvalds 2682058dc334SIlpo Järvinen skb_size = skb->len; 2683058dc334SIlpo Järvinen next_skb_size = next_skb->len; 26841da177e4SLinus Torvalds 2685058dc334SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); 26861da177e4SLinus Torvalds 2687f8071cdeSEric Dumazet if (next_skb_size) { 2688f8071cdeSEric Dumazet if (next_skb_size <= skb_availroom(skb)) 2689f8071cdeSEric Dumazet skb_copy_bits(next_skb, 0, skb_put(skb, next_skb_size), 2690f8071cdeSEric Dumazet next_skb_size); 2691f8071cdeSEric Dumazet else if (!skb_shift(skb, next_skb, next_skb_size)) 2692f8071cdeSEric Dumazet return false; 2693f8071cdeSEric Dumazet } 26946859d494SIlpo Järvinen tcp_highest_sack_combine(sk, next_skb, skb); 2695a6963a6bSIlpo Järvinen 269652d570aaSJarek Poplawski if (next_skb->ip_summed == CHECKSUM_PARTIAL) 269752d570aaSJarek Poplawski skb->ip_summed = CHECKSUM_PARTIAL; 26981da177e4SLinus Torvalds 269984fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 27001da177e4SLinus Torvalds skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); 27011da177e4SLinus Torvalds 27021da177e4SLinus Torvalds /* Update sequence range on original skb. */ 27031da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 27041da177e4SLinus Torvalds 2705e6c7d085SIlpo Järvinen /* Merge over control information. This moves PSH/FIN etc. over */ 27064de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; 27071da177e4SLinus Torvalds 27081da177e4SLinus Torvalds /* All done, get rid of second SKB and account for it so 27091da177e4SLinus Torvalds * packet counting does not break. 27101da177e4SLinus Torvalds */ 27114828e7f4SIlpo Järvinen TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; 2712a643b5d4SMartin KaFai Lau TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor; 2713b7689205SIlpo Järvinen 2714b7689205SIlpo Järvinen /* changed transmit queue under us so clear hints */ 2715ef9da47cSIlpo Järvinen tcp_clear_retrans_hints_partial(tp); 2716ef9da47cSIlpo Järvinen if (next_skb == tp->retransmit_skb_hint) 2717ef9da47cSIlpo Järvinen tp->retransmit_skb_hint = skb; 2718b7689205SIlpo Järvinen 2719797108d1SIlpo Järvinen tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); 2720797108d1SIlpo Järvinen 2721082ac2d5SMartin KaFai Lau tcp_skb_collapse_tstamp(skb, next_skb); 2722082ac2d5SMartin KaFai Lau 272375c119afSEric Dumazet tcp_rtx_queue_unlink_and_free(next_skb, sk); 2724f8071cdeSEric Dumazet return true; 27251da177e4SLinus Torvalds } 27261da177e4SLinus Torvalds 272767edfef7SAndi Kleen /* Check if coalescing SKBs is legal. */ 2728a2a385d6SEric Dumazet static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) 27294a17fc3aSIlpo Järvinen { 27304a17fc3aSIlpo Järvinen if (tcp_skb_pcount(skb) > 1) 2731a2a385d6SEric Dumazet return false; 27324a17fc3aSIlpo Järvinen if (skb_cloned(skb)) 2733a2a385d6SEric Dumazet return false; 27342331ccc5SEric Dumazet /* Some heuristics for collapsing over SACK'd could be invented */ 27354a17fc3aSIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 2736a2a385d6SEric Dumazet return false; 27374a17fc3aSIlpo Järvinen 2738a2a385d6SEric Dumazet return true; 27394a17fc3aSIlpo Järvinen } 27404a17fc3aSIlpo Järvinen 274167edfef7SAndi Kleen /* Collapse packets in the retransmit queue to make to create 274267edfef7SAndi Kleen * less packets on the wire. This is only done on retransmission. 274367edfef7SAndi Kleen */ 27444a17fc3aSIlpo Järvinen static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, 27454a17fc3aSIlpo Järvinen int space) 27464a17fc3aSIlpo Järvinen { 27474a17fc3aSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 27484a17fc3aSIlpo Järvinen struct sk_buff *skb = to, *tmp; 2749a2a385d6SEric Dumazet bool first = true; 27504a17fc3aSIlpo Järvinen 27514a17fc3aSIlpo Järvinen if (!sysctl_tcp_retrans_collapse) 27524a17fc3aSIlpo Järvinen return; 27534de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 27544a17fc3aSIlpo Järvinen return; 27554a17fc3aSIlpo Järvinen 275675c119afSEric Dumazet skb_rbtree_walk_from_safe(skb, tmp) { 27574a17fc3aSIlpo Järvinen if (!tcp_can_collapse(sk, skb)) 27584a17fc3aSIlpo Järvinen break; 27594a17fc3aSIlpo Järvinen 2760a643b5d4SMartin KaFai Lau if (!tcp_skb_can_collapse_to(to)) 2761a643b5d4SMartin KaFai Lau break; 2762a643b5d4SMartin KaFai Lau 27634a17fc3aSIlpo Järvinen space -= skb->len; 27644a17fc3aSIlpo Järvinen 27654a17fc3aSIlpo Järvinen if (first) { 2766a2a385d6SEric Dumazet first = false; 27674a17fc3aSIlpo Järvinen continue; 27684a17fc3aSIlpo Järvinen } 27694a17fc3aSIlpo Järvinen 27704a17fc3aSIlpo Järvinen if (space < 0) 27714a17fc3aSIlpo Järvinen break; 27724a17fc3aSIlpo Järvinen 27734a17fc3aSIlpo Järvinen if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) 27744a17fc3aSIlpo Järvinen break; 27754a17fc3aSIlpo Järvinen 2776f8071cdeSEric Dumazet if (!tcp_collapse_retrans(sk, to)) 2777f8071cdeSEric Dumazet break; 27784a17fc3aSIlpo Järvinen } 27794a17fc3aSIlpo Järvinen } 27804a17fc3aSIlpo Järvinen 27811da177e4SLinus Torvalds /* This retransmits one SKB. Policy decisions and retransmit queue 27821da177e4SLinus Torvalds * state updates are done by the caller. Returns non-zero if an 27831da177e4SLinus Torvalds * error occurred which prevented the send. 27841da177e4SLinus Torvalds */ 278510d3be56SEric Dumazet int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) 27861da177e4SLinus Torvalds { 27875d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 278810d3be56SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 27897d227cd2SSridhar Samudrala unsigned int cur_mss; 279010d3be56SEric Dumazet int diff, len, err; 27911da177e4SLinus Torvalds 279210d3be56SEric Dumazet 279310d3be56SEric Dumazet /* Inconclusive MTU probe */ 279410d3be56SEric Dumazet if (icsk->icsk_mtup.probe_size) 27955d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 27965d424d5aSJohn Heffner 27971da177e4SLinus Torvalds /* Do not sent more than we queued. 1/4 is reserved for possible 2798caa20d9aSStephen Hemminger * copying overhead: fragmentation, tunneling, mangling etc. 27991da177e4SLinus Torvalds */ 280014afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) > 2801ffb4d6c8SEric Dumazet min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), 2802ffb4d6c8SEric Dumazet sk->sk_sndbuf)) 28031da177e4SLinus Torvalds return -EAGAIN; 28041da177e4SLinus Torvalds 28051f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 28061f3279aeSEric Dumazet return -EBUSY; 28071f3279aeSEric Dumazet 28081da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 28091da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 28101da177e4SLinus Torvalds BUG(); 28111da177e4SLinus Torvalds if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 28121da177e4SLinus Torvalds return -ENOMEM; 28131da177e4SLinus Torvalds } 28141da177e4SLinus Torvalds 28157d227cd2SSridhar Samudrala if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 28167d227cd2SSridhar Samudrala return -EHOSTUNREACH; /* Routing failure or similar. */ 28177d227cd2SSridhar Samudrala 28180c54b85fSIlpo Järvinen cur_mss = tcp_current_mss(sk); 28197d227cd2SSridhar Samudrala 28201da177e4SLinus Torvalds /* If receiver has shrunk his window, and skb is out of 28211da177e4SLinus Torvalds * new window, do not retransmit it. The exception is the 28221da177e4SLinus Torvalds * case, when window is shrunk to zero. In this case 28231da177e4SLinus Torvalds * our retransmit serves as a zero window probe. 28241da177e4SLinus Torvalds */ 28259d4fb27dSJoe Perches if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && 28269d4fb27dSJoe Perches TCP_SKB_CB(skb)->seq != tp->snd_una) 28271da177e4SLinus Torvalds return -EAGAIN; 28281da177e4SLinus Torvalds 282910d3be56SEric Dumazet len = cur_mss * segs; 283010d3be56SEric Dumazet if (skb->len > len) { 283175c119afSEric Dumazet if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len, 283275c119afSEric Dumazet cur_mss, GFP_ATOMIC)) 28331da177e4SLinus Torvalds return -ENOMEM; /* We'll try again later. */ 283402276f3cSIlpo Järvinen } else { 2835c52e2421SEric Dumazet if (skb_unclone(skb, GFP_ATOMIC)) 2836c52e2421SEric Dumazet return -ENOMEM; 283710d3be56SEric Dumazet 283810d3be56SEric Dumazet diff = tcp_skb_pcount(skb); 283910d3be56SEric Dumazet tcp_set_skb_tso_segs(skb, cur_mss); 284010d3be56SEric Dumazet diff -= tcp_skb_pcount(skb); 284110d3be56SEric Dumazet if (diff) 284210d3be56SEric Dumazet tcp_adjust_pcount(sk, skb, diff); 284310d3be56SEric Dumazet if (skb->len < cur_mss) 284410d3be56SEric Dumazet tcp_retrans_try_collapse(sk, skb, cur_mss); 28451da177e4SLinus Torvalds } 28461da177e4SLinus Torvalds 284749213555SDaniel Borkmann /* RFC3168, section 6.1.1.1. ECN fallback */ 284849213555SDaniel Borkmann if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN) 284949213555SDaniel Borkmann tcp_ecn_clear_syn(sk, skb); 285049213555SDaniel Borkmann 2851678550c6SYuchung Cheng /* Update global and local TCP statistics. */ 2852678550c6SYuchung Cheng segs = tcp_skb_pcount(skb); 2853678550c6SYuchung Cheng TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs); 2854678550c6SYuchung Cheng if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 2855678550c6SYuchung Cheng __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 2856678550c6SYuchung Cheng tp->total_retrans += segs; 2857678550c6SYuchung Cheng 285850bceae9SThomas Graf /* make sure skb->data is aligned on arches that require it 285950bceae9SThomas Graf * and check if ack-trimming & collapsing extended the headroom 286050bceae9SThomas Graf * beyond what csum_start can cover. 286150bceae9SThomas Graf */ 286250bceae9SThomas Graf if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) || 286350bceae9SThomas Graf skb_headroom(skb) >= 0xFFFF)) { 286410a81980SEric Dumazet struct sk_buff *nskb; 286510a81980SEric Dumazet 2866e2080072SEric Dumazet tcp_skb_tsorted_save(skb) { 286710a81980SEric Dumazet nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC); 2868c84a5711SYuchung Cheng err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) : 2869117632e6SEric Dumazet -ENOBUFS; 2870e2080072SEric Dumazet } tcp_skb_tsorted_restore(skb); 2871e2080072SEric Dumazet 28728c72c65bSEric Dumazet if (!err) 2873e2080072SEric Dumazet tcp_update_skb_after_send(tp, skb); 2874117632e6SEric Dumazet } else { 2875c84a5711SYuchung Cheng err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 2876117632e6SEric Dumazet } 2877c84a5711SYuchung Cheng 2878fc9f3501SEric Dumazet if (likely(!err)) { 2879c84a5711SYuchung Cheng TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS; 2880*e086101bSCong Wang trace_tcp_retransmit_skb(sk, skb); 2881678550c6SYuchung Cheng } else if (err != -EBUSY) { 2882678550c6SYuchung Cheng NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL); 2883fc9f3501SEric Dumazet } 2884c84a5711SYuchung Cheng return err; 288593b174adSYuchung Cheng } 288693b174adSYuchung Cheng 288710d3be56SEric Dumazet int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) 288893b174adSYuchung Cheng { 288993b174adSYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 289010d3be56SEric Dumazet int err = __tcp_retransmit_skb(sk, skb, segs); 28911da177e4SLinus Torvalds 28921da177e4SLinus Torvalds if (err == 0) { 28931da177e4SLinus Torvalds #if FASTRETRANS_DEBUG > 0 28941da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 2895e87cc472SJoe Perches net_dbg_ratelimited("retrans_out leaked\n"); 28961da177e4SLinus Torvalds } 28971da177e4SLinus Torvalds #endif 28981da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 28991da177e4SLinus Torvalds tp->retrans_out += tcp_skb_pcount(skb); 29001da177e4SLinus Torvalds 29011da177e4SLinus Torvalds /* Save stamp of the first retransmit. */ 29021da177e4SLinus Torvalds if (!tp->retrans_stamp) 29037faee5c0SEric Dumazet tp->retrans_stamp = tcp_skb_timestamp(skb); 29041da177e4SLinus Torvalds 29051da177e4SLinus Torvalds } 29066e08d5e3SYuchung Cheng 29076e08d5e3SYuchung Cheng if (tp->undo_retrans < 0) 29086e08d5e3SYuchung Cheng tp->undo_retrans = 0; 29096e08d5e3SYuchung Cheng tp->undo_retrans += tcp_skb_pcount(skb); 29101da177e4SLinus Torvalds return err; 29111da177e4SLinus Torvalds } 29121da177e4SLinus Torvalds 29131da177e4SLinus Torvalds /* This gets called after a retransmit timeout, and the initially 29141da177e4SLinus Torvalds * retransmitted data is acknowledged. It tries to continue 29151da177e4SLinus Torvalds * resending the rest of the retransmit queue, until either 29161da177e4SLinus Torvalds * we've sent it all or the congestion window limit is reached. 29171da177e4SLinus Torvalds * If doing SACK, the first ACK which comes back for a timeout 29181da177e4SLinus Torvalds * based retransmit packet might feed us FACK information again. 29191da177e4SLinus Torvalds * If so, we use it to avoid unnecessarily retransmissions. 29201da177e4SLinus Torvalds */ 29211da177e4SLinus Torvalds void tcp_xmit_retransmit_queue(struct sock *sk) 29221da177e4SLinus Torvalds { 29236687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 292475c119afSEric Dumazet struct sk_buff *skb, *rtx_head = NULL, *hole = NULL; 29251da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2926840a3cbeSYuchung Cheng u32 max_segs; 292761eb55f4SIlpo Järvinen int mib_idx; 29286a438bbeSStephen Hemminger 292945e77d31SIlpo Järvinen if (!tp->packets_out) 293045e77d31SIlpo Järvinen return; 293145e77d31SIlpo Järvinen 29326a438bbeSStephen Hemminger skb = tp->retransmit_skb_hint; 293375c119afSEric Dumazet if (!skb) { 293475c119afSEric Dumazet rtx_head = tcp_rtx_queue_head(sk); 293575c119afSEric Dumazet skb = rtx_head; 2936618d9f25SIlpo Järvinen } 2937ed6e7268SNeal Cardwell max_segs = tcp_tso_segs(sk, tcp_current_mss(sk)); 293875c119afSEric Dumazet skb_rbtree_walk_from(skb) { 2939dca0aaf8SEric Dumazet __u8 sacked; 294010d3be56SEric Dumazet int segs; 29411da177e4SLinus Torvalds 2942218af599SEric Dumazet if (tcp_pacing_check(sk)) 2943218af599SEric Dumazet break; 2944218af599SEric Dumazet 29456a438bbeSStephen Hemminger /* we could do better than to assign each time */ 294651456b29SIan Morris if (!hole) 29476a438bbeSStephen Hemminger tp->retransmit_skb_hint = skb; 29486a438bbeSStephen Hemminger 294910d3be56SEric Dumazet segs = tp->snd_cwnd - tcp_packets_in_flight(tp); 295010d3be56SEric Dumazet if (segs <= 0) 29511da177e4SLinus Torvalds return; 2952dca0aaf8SEric Dumazet sacked = TCP_SKB_CB(skb)->sacked; 2953a3d2e9f8SEric Dumazet /* In case tcp_shift_skb_data() have aggregated large skbs, 2954a3d2e9f8SEric Dumazet * we need to make sure not sending too bigs TSO packets 2955a3d2e9f8SEric Dumazet */ 2956a3d2e9f8SEric Dumazet segs = min_t(int, segs, max_segs); 29570e1c54c2SIlpo Järvinen 2958840a3cbeSYuchung Cheng if (tp->retrans_out >= tp->lost_out) { 2959006f582cSIlpo Järvinen break; 29600e1c54c2SIlpo Järvinen } else if (!(sacked & TCPCB_LOST)) { 296151456b29SIan Morris if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) 29620e1c54c2SIlpo Järvinen hole = skb; 296361eb55f4SIlpo Järvinen continue; 29641da177e4SLinus Torvalds 29650e1c54c2SIlpo Järvinen } else { 29660e1c54c2SIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Loss) 29670e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFASTRETRANS; 29680e1c54c2SIlpo Järvinen else 29690e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; 29700e1c54c2SIlpo Järvinen } 29710e1c54c2SIlpo Järvinen 29720e1c54c2SIlpo Järvinen if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) 297361eb55f4SIlpo Järvinen continue; 297440b215e5SPavel Emelyanov 2975f9616c35SEric Dumazet if (tcp_small_queue_check(sk, skb, 1)) 2976f9616c35SEric Dumazet return; 2977f9616c35SEric Dumazet 297810d3be56SEric Dumazet if (tcp_retransmit_skb(sk, skb, segs)) 29791da177e4SLinus Torvalds return; 298024ab6becSYuchung Cheng 2981de1d6578SYuchung Cheng NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb)); 29821da177e4SLinus Torvalds 2983684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 2984a262f0cdSNandita Dukkipati tp->prr_out += tcp_skb_pcount(skb); 2985a262f0cdSNandita Dukkipati 298675c119afSEric Dumazet if (skb == rtx_head && 298757dde7f7SYuchung Cheng icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT) 2988463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 29893f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, 29903f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 29911da177e4SLinus Torvalds } 29921da177e4SLinus Torvalds } 29931da177e4SLinus Torvalds 2994d83769a5SEric Dumazet /* We allow to exceed memory limits for FIN packets to expedite 2995d83769a5SEric Dumazet * connection tear down and (memory) recovery. 2996845704a5SEric Dumazet * Otherwise tcp_send_fin() could be tempted to either delay FIN 2997845704a5SEric Dumazet * or even be forced to close flow without any FIN. 2998a6c5ea4cSEric Dumazet * In general, we want to allow one skb per socket to avoid hangs 2999a6c5ea4cSEric Dumazet * with edge trigger epoll() 3000d83769a5SEric Dumazet */ 3001a6c5ea4cSEric Dumazet void sk_forced_mem_schedule(struct sock *sk, int size) 3002d83769a5SEric Dumazet { 3003e805605cSJohannes Weiner int amt; 3004d83769a5SEric Dumazet 3005d83769a5SEric Dumazet if (size <= sk->sk_forward_alloc) 3006d83769a5SEric Dumazet return; 3007d83769a5SEric Dumazet amt = sk_mem_pages(size); 3008d83769a5SEric Dumazet sk->sk_forward_alloc += amt * SK_MEM_QUANTUM; 3009e805605cSJohannes Weiner sk_memory_allocated_add(sk, amt); 3010e805605cSJohannes Weiner 3011baac50bbSJohannes Weiner if (mem_cgroup_sockets_enabled && sk->sk_memcg) 3012baac50bbSJohannes Weiner mem_cgroup_charge_skmem(sk->sk_memcg, amt); 3013d83769a5SEric Dumazet } 3014d83769a5SEric Dumazet 3015845704a5SEric Dumazet /* Send a FIN. The caller locks the socket for us. 3016845704a5SEric Dumazet * We should try to send a FIN packet really hard, but eventually give up. 30171da177e4SLinus Torvalds */ 30181da177e4SLinus Torvalds void tcp_send_fin(struct sock *sk) 30191da177e4SLinus Torvalds { 3020845704a5SEric Dumazet struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk); 30211da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 30221da177e4SLinus Torvalds 3023845704a5SEric Dumazet /* Optimization, tack on the FIN if we have one skb in write queue and 3024845704a5SEric Dumazet * this skb was not yet sent, or we are under memory pressure. 3025845704a5SEric Dumazet * Note: in the latter case, FIN packet will be sent after a timeout, 3026845704a5SEric Dumazet * as TCP stack thinks it has already been transmitted. 30271da177e4SLinus Torvalds */ 302875c119afSEric Dumazet if (!tskb && tcp_under_memory_pressure(sk)) 302975c119afSEric Dumazet tskb = skb_rb_last(&sk->tcp_rtx_queue); 303075c119afSEric Dumazet 303175c119afSEric Dumazet if (tskb) { 3032845704a5SEric Dumazet coalesce: 3033845704a5SEric Dumazet TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN; 3034845704a5SEric Dumazet TCP_SKB_CB(tskb)->end_seq++; 30351da177e4SLinus Torvalds tp->write_seq++; 303675c119afSEric Dumazet if (tcp_write_queue_empty(sk)) { 3037845704a5SEric Dumazet /* This means tskb was already sent. 3038845704a5SEric Dumazet * Pretend we included the FIN on previous transmit. 3039845704a5SEric Dumazet * We need to set tp->snd_nxt to the value it would have 3040845704a5SEric Dumazet * if FIN had been sent. This is because retransmit path 3041845704a5SEric Dumazet * does not change tp->snd_nxt. 3042845704a5SEric Dumazet */ 3043845704a5SEric Dumazet tp->snd_nxt++; 3044845704a5SEric Dumazet return; 3045845704a5SEric Dumazet } 30461da177e4SLinus Torvalds } else { 3047845704a5SEric Dumazet skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation); 3048845704a5SEric Dumazet if (unlikely(!skb)) { 3049845704a5SEric Dumazet if (tskb) 3050845704a5SEric Dumazet goto coalesce; 3051845704a5SEric Dumazet return; 30521da177e4SLinus Torvalds } 3053e2080072SEric Dumazet INIT_LIST_HEAD(&skb->tcp_tsorted_anchor); 3054d83769a5SEric Dumazet skb_reserve(skb, MAX_TCP_HEADER); 3055a6c5ea4cSEric Dumazet sk_forced_mem_schedule(sk, skb->truesize); 30561da177e4SLinus Torvalds /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 3057e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tp->write_seq, 3058a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_FIN); 30591da177e4SLinus Torvalds tcp_queue_skb(sk, skb); 30601da177e4SLinus Torvalds } 3061845704a5SEric Dumazet __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF); 30621da177e4SLinus Torvalds } 30631da177e4SLinus Torvalds 30641da177e4SLinus Torvalds /* We get here when a process closes a file descriptor (either due to 30651da177e4SLinus Torvalds * an explicit close() or as a byproduct of exit()'ing) and there 30661da177e4SLinus Torvalds * was unread data in the receive queue. This behavior is recommended 306765bb723cSGerrit Renker * by RFC 2525, section 2.17. -DaveM 30681da177e4SLinus Torvalds */ 3069dd0fc66fSAl Viro void tcp_send_active_reset(struct sock *sk, gfp_t priority) 30701da177e4SLinus Torvalds { 30711da177e4SLinus Torvalds struct sk_buff *skb; 30721da177e4SLinus Torvalds 30737cc2b043SGao Feng TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); 30747cc2b043SGao Feng 30751da177e4SLinus Torvalds /* NOTE: No TCP options attached and we never retransmit this. */ 30761da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, priority); 30771da177e4SLinus Torvalds if (!skb) { 30784e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 30791da177e4SLinus Torvalds return; 30801da177e4SLinus Torvalds } 30811da177e4SLinus Torvalds 30821da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 30831da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 3084e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), 3085a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_RST); 30869a568de4SEric Dumazet tcp_mstamp_refresh(tcp_sk(sk)); 30871da177e4SLinus Torvalds /* Send it off. */ 3088dfb4b9dcSDavid S. Miller if (tcp_transmit_skb(sk, skb, 0, priority)) 30894e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 30901da177e4SLinus Torvalds } 30911da177e4SLinus Torvalds 309267edfef7SAndi Kleen /* Send a crossed SYN-ACK during socket establishment. 309367edfef7SAndi Kleen * WARNING: This routine must only be called when we have already sent 30941da177e4SLinus Torvalds * a SYN packet that crossed the incoming SYN that caused this routine 30951da177e4SLinus Torvalds * to get called. If this assumption fails then the initial rcv_wnd 30961da177e4SLinus Torvalds * and rcv_wscale values will not be correct. 30971da177e4SLinus Torvalds */ 30981da177e4SLinus Torvalds int tcp_send_synack(struct sock *sk) 30991da177e4SLinus Torvalds { 31001da177e4SLinus Torvalds struct sk_buff *skb; 31011da177e4SLinus Torvalds 310275c119afSEric Dumazet skb = tcp_rtx_queue_head(sk); 310351456b29SIan Morris if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 310475c119afSEric Dumazet pr_err("%s: wrong queue state\n", __func__); 31051da177e4SLinus Torvalds return -EFAULT; 31061da177e4SLinus Torvalds } 31074de075e0SEric Dumazet if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { 31081da177e4SLinus Torvalds if (skb_cloned(skb)) { 3109e2080072SEric Dumazet struct sk_buff *nskb; 3110e2080072SEric Dumazet 3111e2080072SEric Dumazet tcp_skb_tsorted_save(skb) { 3112e2080072SEric Dumazet nskb = skb_copy(skb, GFP_ATOMIC); 3113e2080072SEric Dumazet } tcp_skb_tsorted_restore(skb); 311451456b29SIan Morris if (!nskb) 31151da177e4SLinus Torvalds return -ENOMEM; 3116e2080072SEric Dumazet INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor); 311775c119afSEric Dumazet tcp_rtx_queue_unlink_and_free(skb, sk); 3118f4a775d1SEric Dumazet __skb_header_release(nskb); 311975c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb); 31203ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 31213ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 31221da177e4SLinus Torvalds skb = nskb; 31231da177e4SLinus Torvalds } 31241da177e4SLinus Torvalds 31254de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; 3126735d3831SFlorian Westphal tcp_ecn_send_synack(sk, skb); 31271da177e4SLinus Torvalds } 3128dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 31291da177e4SLinus Torvalds } 31301da177e4SLinus Torvalds 31314aea39c1SEric Dumazet /** 31324aea39c1SEric Dumazet * tcp_make_synack - Prepare a SYN-ACK. 31334aea39c1SEric Dumazet * sk: listener socket 31344aea39c1SEric Dumazet * dst: dst entry attached to the SYNACK 31354aea39c1SEric Dumazet * req: request_sock pointer 31364aea39c1SEric Dumazet * 31374aea39c1SEric Dumazet * Allocate one skb and build a SYNACK packet. 31384aea39c1SEric Dumazet * @dst is consumed : Caller should not use it again. 31394aea39c1SEric Dumazet */ 31405d062de7SEric Dumazet struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst, 3141e6b4d113SWilliam Allen Simpson struct request_sock *req, 3142ca6fb065SEric Dumazet struct tcp_fastopen_cookie *foc, 3143b3d05147SEric Dumazet enum tcp_synack_type synack_type) 31441da177e4SLinus Torvalds { 31452e6599cbSArnaldo Carvalho de Melo struct inet_request_sock *ireq = inet_rsk(req); 31465d062de7SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 314780f03e27SEric Dumazet struct tcp_md5sig_key *md5 = NULL; 31485d062de7SEric Dumazet struct tcp_out_options opts; 31495d062de7SEric Dumazet struct sk_buff *skb; 3150bd0388aeSWilliam Allen Simpson int tcp_header_size; 31515d062de7SEric Dumazet struct tcphdr *th; 3152f5fff5dcSTom Quetchenbach int mss; 31531da177e4SLinus Torvalds 3154ca6fb065SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 31554aea39c1SEric Dumazet if (unlikely(!skb)) { 31564aea39c1SEric Dumazet dst_release(dst); 31571da177e4SLinus Torvalds return NULL; 31584aea39c1SEric Dumazet } 31591da177e4SLinus Torvalds /* Reserve space for headers. */ 31601da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 31611da177e4SLinus Torvalds 3162b3d05147SEric Dumazet switch (synack_type) { 3163b3d05147SEric Dumazet case TCP_SYNACK_NORMAL: 31649e17f8a4SEric Dumazet skb_set_owner_w(skb, req_to_sk(req)); 3165b3d05147SEric Dumazet break; 3166b3d05147SEric Dumazet case TCP_SYNACK_COOKIE: 3167b3d05147SEric Dumazet /* Under synflood, we do not attach skb to a socket, 3168b3d05147SEric Dumazet * to avoid false sharing. 3169b3d05147SEric Dumazet */ 3170b3d05147SEric Dumazet break; 3171b3d05147SEric Dumazet case TCP_SYNACK_FASTOPEN: 3172ca6fb065SEric Dumazet /* sk is a const pointer, because we want to express multiple 3173ca6fb065SEric Dumazet * cpu might call us concurrently. 3174ca6fb065SEric Dumazet * sk->sk_wmem_alloc in an atomic, we can promote to rw. 3175ca6fb065SEric Dumazet */ 3176ca6fb065SEric Dumazet skb_set_owner_w(skb, (struct sock *)sk); 3177b3d05147SEric Dumazet break; 3178ca6fb065SEric Dumazet } 31794aea39c1SEric Dumazet skb_dst_set(skb, dst); 31801da177e4SLinus Torvalds 31813541f9e8SEric Dumazet mss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 3182f5fff5dcSTom Quetchenbach 318333ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 31848b5f12d0SFlorian Westphal #ifdef CONFIG_SYN_COOKIES 31858b5f12d0SFlorian Westphal if (unlikely(req->cookie_ts)) 31869a568de4SEric Dumazet skb->skb_mstamp = cookie_init_timestamp(req); 31878b5f12d0SFlorian Westphal else 31888b5f12d0SFlorian Westphal #endif 31899a568de4SEric Dumazet skb->skb_mstamp = tcp_clock_us(); 319080f03e27SEric Dumazet 319180f03e27SEric Dumazet #ifdef CONFIG_TCP_MD5SIG 319280f03e27SEric Dumazet rcu_read_lock(); 3193fd3a154aSEric Dumazet md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req)); 319480f03e27SEric Dumazet #endif 319558d607d3SEric Dumazet skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4); 319637bfbddaSEric Dumazet tcp_header_size = tcp_synack_options(req, mss, skb, &opts, md5, foc) + 319737bfbddaSEric Dumazet sizeof(*th); 319833ad798cSAdam Langley 3199aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 3200aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 32011da177e4SLinus Torvalds 3202ea1627c2SEric Dumazet th = (struct tcphdr *)skb->data; 32031da177e4SLinus Torvalds memset(th, 0, sizeof(struct tcphdr)); 32041da177e4SLinus Torvalds th->syn = 1; 32051da177e4SLinus Torvalds th->ack = 1; 32066ac705b1SEric Dumazet tcp_ecn_make_synack(req, th); 3207b44084c2SEric Dumazet th->source = htons(ireq->ir_num); 3208634fb979SEric Dumazet th->dest = ireq->ir_rmt_port; 3209e05a90ecSJamal Hadi Salim skb->mark = ireq->ir_mark; 3210e870a8efSIlpo Järvinen /* Setting of flags are superfluous here for callers (and ECE is 3211e870a8efSIlpo Järvinen * not even correctly set) 3212e870a8efSIlpo Järvinen */ 3213e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn, 3214a3433f35SChangli Gao TCPHDR_SYN | TCPHDR_ACK); 32154957faadSWilliam Allen Simpson 32161da177e4SLinus Torvalds th->seq = htonl(TCP_SKB_CB(skb)->seq); 32178336886fSJerry Chu /* XXX data is queued and acked as is. No buffer/window check */ 32188336886fSJerry Chu th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt); 32191da177e4SLinus Torvalds 32201da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 3221ed53d0abSEric Dumazet th->window = htons(min(req->rsk_rcv_wnd, 65535U)); 32225d062de7SEric Dumazet tcp_options_write((__be32 *)(th + 1), NULL, &opts); 32231da177e4SLinus Torvalds th->doff = (tcp_header_size >> 2); 322490bbcc60SEric Dumazet __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS); 3225cfb6eeb4SYOSHIFUJI Hideaki 3226cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 3227cfb6eeb4SYOSHIFUJI Hideaki /* Okay, we have all we need - do the md5 hash if needed */ 322880f03e27SEric Dumazet if (md5) 3229bd0388aeSWilliam Allen Simpson tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, 323039f8e58eSEric Dumazet md5, req_to_sk(req), skb); 323180f03e27SEric Dumazet rcu_read_unlock(); 3232cfb6eeb4SYOSHIFUJI Hideaki #endif 3233cfb6eeb4SYOSHIFUJI Hideaki 3234b50edd78SEric Dumazet /* Do not fool tcpdump (if any), clean our debris */ 32352456e855SThomas Gleixner skb->tstamp = 0; 32361da177e4SLinus Torvalds return skb; 32371da177e4SLinus Torvalds } 32384bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_make_synack); 32391da177e4SLinus Torvalds 324081164413SDaniel Borkmann static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst) 324181164413SDaniel Borkmann { 324281164413SDaniel Borkmann struct inet_connection_sock *icsk = inet_csk(sk); 324381164413SDaniel Borkmann const struct tcp_congestion_ops *ca; 324481164413SDaniel Borkmann u32 ca_key = dst_metric(dst, RTAX_CC_ALGO); 324581164413SDaniel Borkmann 324681164413SDaniel Borkmann if (ca_key == TCP_CA_UNSPEC) 324781164413SDaniel Borkmann return; 324881164413SDaniel Borkmann 324981164413SDaniel Borkmann rcu_read_lock(); 325081164413SDaniel Borkmann ca = tcp_ca_find_key(ca_key); 325181164413SDaniel Borkmann if (likely(ca && try_module_get(ca->owner))) { 325281164413SDaniel Borkmann module_put(icsk->icsk_ca_ops->owner); 325381164413SDaniel Borkmann icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst); 325481164413SDaniel Borkmann icsk->icsk_ca_ops = ca; 325581164413SDaniel Borkmann } 325681164413SDaniel Borkmann rcu_read_unlock(); 325781164413SDaniel Borkmann } 325881164413SDaniel Borkmann 325967edfef7SAndi Kleen /* Do all connect socket setups that can be done AF independent. */ 3260f7e56a76Sstephen hemminger static void tcp_connect_init(struct sock *sk) 32611da177e4SLinus Torvalds { 3262cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 32631da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 32641da177e4SLinus Torvalds __u8 rcv_wscale; 326513d3b1ebSLawrence Brakmo u32 rcv_wnd; 32661da177e4SLinus Torvalds 32671da177e4SLinus Torvalds /* We'll fix this up when we get a response from the other end. 32681da177e4SLinus Torvalds * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 32691da177e4SLinus Torvalds */ 32705d2ed052SEric Dumazet tp->tcp_header_len = sizeof(struct tcphdr); 32715d2ed052SEric Dumazet if (sock_net(sk)->ipv4.sysctl_tcp_timestamps) 32725d2ed052SEric Dumazet tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED; 32731da177e4SLinus Torvalds 3274cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 327500db4124SIan Morris if (tp->af_specific->md5_lookup(sk, sk)) 3276cfb6eeb4SYOSHIFUJI Hideaki tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 3277cfb6eeb4SYOSHIFUJI Hideaki #endif 3278cfb6eeb4SYOSHIFUJI Hideaki 32791da177e4SLinus Torvalds /* If user gave his TCP_MAXSEG, record it to clamp */ 32801da177e4SLinus Torvalds if (tp->rx_opt.user_mss) 32811da177e4SLinus Torvalds tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 32821da177e4SLinus Torvalds tp->max_window = 0; 32835d424d5aSJohn Heffner tcp_mtup_init(sk); 32841da177e4SLinus Torvalds tcp_sync_mss(sk, dst_mtu(dst)); 32851da177e4SLinus Torvalds 328681164413SDaniel Borkmann tcp_ca_dst_init(sk, dst); 328781164413SDaniel Borkmann 32881da177e4SLinus Torvalds if (!tp->window_clamp) 32891da177e4SLinus Torvalds tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 32903541f9e8SEric Dumazet tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 3291f5fff5dcSTom Quetchenbach 32921da177e4SLinus Torvalds tcp_initialize_rcv_mss(sk); 32931da177e4SLinus Torvalds 3294e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 3295e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 3296e88c64f0SHagen Paul Pfeifer (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) 3297e88c64f0SHagen Paul Pfeifer tp->window_clamp = tcp_full_space(sk); 3298e88c64f0SHagen Paul Pfeifer 329913d3b1ebSLawrence Brakmo rcv_wnd = tcp_rwnd_init_bpf(sk); 330013d3b1ebSLawrence Brakmo if (rcv_wnd == 0) 330113d3b1ebSLawrence Brakmo rcv_wnd = dst_metric(dst, RTAX_INITRWND); 330213d3b1ebSLawrence Brakmo 33031da177e4SLinus Torvalds tcp_select_initial_window(tcp_full_space(sk), 33041da177e4SLinus Torvalds tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 33051da177e4SLinus Torvalds &tp->rcv_wnd, 33061da177e4SLinus Torvalds &tp->window_clamp, 33079bb37ef0SEric Dumazet sock_net(sk)->ipv4.sysctl_tcp_window_scaling, 330831d12926Slaurent chavey &rcv_wscale, 330913d3b1ebSLawrence Brakmo rcv_wnd); 33101da177e4SLinus Torvalds 33111da177e4SLinus Torvalds tp->rx_opt.rcv_wscale = rcv_wscale; 33121da177e4SLinus Torvalds tp->rcv_ssthresh = tp->rcv_wnd; 33131da177e4SLinus Torvalds 33141da177e4SLinus Torvalds sk->sk_err = 0; 33151da177e4SLinus Torvalds sock_reset_flag(sk, SOCK_DONE); 33161da177e4SLinus Torvalds tp->snd_wnd = 0; 3317ee7537b6SHantzis Fotis tcp_init_wl(tp, 0); 33181da177e4SLinus Torvalds tp->snd_una = tp->write_seq; 33191da177e4SLinus Torvalds tp->snd_sml = tp->write_seq; 332033f5f57eSIlpo Järvinen tp->snd_up = tp->write_seq; 3321370816aeSPavel Emelyanov tp->snd_nxt = tp->write_seq; 3322ee995283SPavel Emelyanov 3323ee995283SPavel Emelyanov if (likely(!tp->repair)) 33241da177e4SLinus Torvalds tp->rcv_nxt = 0; 3325c7781a6eSAndrew Vagin else 332670eabf0eSEric Dumazet tp->rcv_tstamp = tcp_jiffies32; 3327ee995283SPavel Emelyanov tp->rcv_wup = tp->rcv_nxt; 3328ee995283SPavel Emelyanov tp->copied_seq = tp->rcv_nxt; 33291da177e4SLinus Torvalds 33308550f328SLawrence Brakmo inet_csk(sk)->icsk_rto = tcp_timeout_init(sk); 3331463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_retransmits = 0; 33321da177e4SLinus Torvalds tcp_clear_retrans(tp); 33331da177e4SLinus Torvalds } 33341da177e4SLinus Torvalds 3335783237e8SYuchung Cheng static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) 3336783237e8SYuchung Cheng { 3337783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3338783237e8SYuchung Cheng struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 3339783237e8SYuchung Cheng 3340783237e8SYuchung Cheng tcb->end_seq += skb->len; 3341f4a775d1SEric Dumazet __skb_header_release(skb); 3342783237e8SYuchung Cheng sk->sk_wmem_queued += skb->truesize; 3343783237e8SYuchung Cheng sk_mem_charge(sk, skb->truesize); 3344783237e8SYuchung Cheng tp->write_seq = tcb->end_seq; 3345783237e8SYuchung Cheng tp->packets_out += tcp_skb_pcount(skb); 3346783237e8SYuchung Cheng } 3347783237e8SYuchung Cheng 3348783237e8SYuchung Cheng /* Build and send a SYN with data and (cached) Fast Open cookie. However, 3349783237e8SYuchung Cheng * queue a data-only packet after the regular SYN, such that regular SYNs 3350783237e8SYuchung Cheng * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges 3351783237e8SYuchung Cheng * only the SYN sequence, the data are retransmitted in the first ACK. 3352783237e8SYuchung Cheng * If cookie is not cached or other error occurs, falls back to send a 3353783237e8SYuchung Cheng * regular SYN with Fast Open cookie request option. 3354783237e8SYuchung Cheng */ 3355783237e8SYuchung Cheng static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) 3356783237e8SYuchung Cheng { 3357783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3358783237e8SYuchung Cheng struct tcp_fastopen_request *fo = tp->fastopen_req; 3359065263f4SWei Wang int space, err = 0; 3360355a901eSEric Dumazet struct sk_buff *syn_data; 3361783237e8SYuchung Cheng 336267da22d2SYuchung Cheng tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ 3363065263f4SWei Wang if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie)) 3364783237e8SYuchung Cheng goto fallback; 3365783237e8SYuchung Cheng 3366783237e8SYuchung Cheng /* MSS for SYN-data is based on cached MSS and bounded by PMTU and 3367783237e8SYuchung Cheng * user-MSS. Reserve maximum option space for middleboxes that add 3368783237e8SYuchung Cheng * private TCP options. The cost is reduced data space in SYN :( 3369783237e8SYuchung Cheng */ 33703541f9e8SEric Dumazet tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp); 33713541f9e8SEric Dumazet 33721b63edd6SYuchung Cheng space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) - 3373783237e8SYuchung Cheng MAX_TCP_OPTION_SPACE; 3374783237e8SYuchung Cheng 3375f5ddcbbbSEric Dumazet space = min_t(size_t, space, fo->size); 3376f5ddcbbbSEric Dumazet 3377f5ddcbbbSEric Dumazet /* limit to order-0 allocations */ 3378f5ddcbbbSEric Dumazet space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER)); 3379f5ddcbbbSEric Dumazet 3380eb934478SEric Dumazet syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false); 3381355a901eSEric Dumazet if (!syn_data) 3382783237e8SYuchung Cheng goto fallback; 3383355a901eSEric Dumazet syn_data->ip_summed = CHECKSUM_PARTIAL; 3384355a901eSEric Dumazet memcpy(syn_data->cb, syn->cb, sizeof(syn->cb)); 338507e100f9SEric Dumazet if (space) { 338607e100f9SEric Dumazet int copied = copy_from_iter(skb_put(syn_data, space), space, 338757be5bdaSAl Viro &fo->data->msg_iter); 338857be5bdaSAl Viro if (unlikely(!copied)) { 3389355a901eSEric Dumazet kfree_skb(syn_data); 3390783237e8SYuchung Cheng goto fallback; 3391783237e8SYuchung Cheng } 339257be5bdaSAl Viro if (copied != space) { 339357be5bdaSAl Viro skb_trim(syn_data, copied); 339457be5bdaSAl Viro space = copied; 339557be5bdaSAl Viro } 339607e100f9SEric Dumazet } 3397355a901eSEric Dumazet /* No more data pending in inet_wait_for_connect() */ 3398355a901eSEric Dumazet if (space == fo->size) 3399355a901eSEric Dumazet fo->data = NULL; 3400355a901eSEric Dumazet fo->copied = space; 3401783237e8SYuchung Cheng 3402355a901eSEric Dumazet tcp_connect_queue_skb(sk, syn_data); 34030f87230dSFrancis Yan if (syn_data->len) 34040f87230dSFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_BUSY); 3405355a901eSEric Dumazet 3406355a901eSEric Dumazet err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation); 3407355a901eSEric Dumazet 3408355a901eSEric Dumazet syn->skb_mstamp = syn_data->skb_mstamp; 3409355a901eSEric Dumazet 3410355a901eSEric Dumazet /* Now full SYN+DATA was cloned and sent (or not), 3411355a901eSEric Dumazet * remove the SYN from the original skb (syn_data) 3412355a901eSEric Dumazet * we keep in write queue in case of a retransmit, as we 3413355a901eSEric Dumazet * also have the SYN packet (with no data) in the same queue. 3414431a9124SEric Dumazet */ 3415355a901eSEric Dumazet TCP_SKB_CB(syn_data)->seq++; 3416355a901eSEric Dumazet TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH; 3417355a901eSEric Dumazet if (!err) { 341867da22d2SYuchung Cheng tp->syn_data = (fo->copied > 0); 341975c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data); 3420f19c29e3SYuchung Cheng NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT); 3421783237e8SYuchung Cheng goto done; 3422783237e8SYuchung Cheng } 3423783237e8SYuchung Cheng 342475c119afSEric Dumazet /* data was not sent, put it in write_queue */ 342575c119afSEric Dumazet __skb_queue_tail(&sk->sk_write_queue, syn_data); 3426b5b7db8dSEric Dumazet tp->packets_out -= tcp_skb_pcount(syn_data); 3427b5b7db8dSEric Dumazet 3428783237e8SYuchung Cheng fallback: 3429783237e8SYuchung Cheng /* Send a regular SYN with Fast Open cookie request option */ 3430783237e8SYuchung Cheng if (fo->cookie.len > 0) 3431783237e8SYuchung Cheng fo->cookie.len = 0; 3432783237e8SYuchung Cheng err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); 3433783237e8SYuchung Cheng if (err) 3434783237e8SYuchung Cheng tp->syn_fastopen = 0; 3435783237e8SYuchung Cheng done: 3436783237e8SYuchung Cheng fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ 3437783237e8SYuchung Cheng return err; 3438783237e8SYuchung Cheng } 3439783237e8SYuchung Cheng 344067edfef7SAndi Kleen /* Build a SYN and send it off. */ 34411da177e4SLinus Torvalds int tcp_connect(struct sock *sk) 34421da177e4SLinus Torvalds { 34431da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 34441da177e4SLinus Torvalds struct sk_buff *buff; 3445ee586811SEric Paris int err; 34461da177e4SLinus Torvalds 34479872a4bdSLawrence Brakmo tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB); 34488ba60924SEric Dumazet 34498ba60924SEric Dumazet if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 34508ba60924SEric Dumazet return -EHOSTUNREACH; /* Routing failure or similar. */ 34518ba60924SEric Dumazet 34521da177e4SLinus Torvalds tcp_connect_init(sk); 34531da177e4SLinus Torvalds 34542b916477SAndrey Vagin if (unlikely(tp->repair)) { 34552b916477SAndrey Vagin tcp_finish_connect(sk, NULL); 34562b916477SAndrey Vagin return 0; 34572b916477SAndrey Vagin } 34582b916477SAndrey Vagin 3459eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true); 3460355a901eSEric Dumazet if (unlikely(!buff)) 34611da177e4SLinus Torvalds return -ENOBUFS; 34621da177e4SLinus Torvalds 3463a3433f35SChangli Gao tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); 34649a568de4SEric Dumazet tcp_mstamp_refresh(tp); 34659a568de4SEric Dumazet tp->retrans_stamp = tcp_time_stamp(tp); 3466783237e8SYuchung Cheng tcp_connect_queue_skb(sk, buff); 3467735d3831SFlorian Westphal tcp_ecn_send_syn(sk, buff); 346875c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, buff); 34691da177e4SLinus Torvalds 3470783237e8SYuchung Cheng /* Send off SYN; include data in Fast Open. */ 3471783237e8SYuchung Cheng err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : 3472783237e8SYuchung Cheng tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); 3473ee586811SEric Paris if (err == -ECONNREFUSED) 3474ee586811SEric Paris return err; 3475bd37a088SWei Yongjun 3476bd37a088SWei Yongjun /* We change tp->snd_nxt after the tcp_transmit_skb() call 3477bd37a088SWei Yongjun * in order to make this packet get counted in tcpOutSegs. 3478bd37a088SWei Yongjun */ 3479bd37a088SWei Yongjun tp->snd_nxt = tp->write_seq; 3480bd37a088SWei Yongjun tp->pushed_seq = tp->write_seq; 3481b5b7db8dSEric Dumazet buff = tcp_send_head(sk); 3482b5b7db8dSEric Dumazet if (unlikely(buff)) { 3483b5b7db8dSEric Dumazet tp->snd_nxt = TCP_SKB_CB(buff)->seq; 3484b5b7db8dSEric Dumazet tp->pushed_seq = TCP_SKB_CB(buff)->seq; 3485b5b7db8dSEric Dumazet } 348681cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); 34871da177e4SLinus Torvalds 34881da177e4SLinus Torvalds /* Timer for repeating the SYN until an answer. */ 34893f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 34903f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 34911da177e4SLinus Torvalds return 0; 34921da177e4SLinus Torvalds } 34934bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_connect); 34941da177e4SLinus Torvalds 34951da177e4SLinus Torvalds /* Send out a delayed ack, the caller does the policy checking 34961da177e4SLinus Torvalds * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 34971da177e4SLinus Torvalds * for details. 34981da177e4SLinus Torvalds */ 34991da177e4SLinus Torvalds void tcp_send_delayed_ack(struct sock *sk) 35001da177e4SLinus Torvalds { 3501463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 3502463c84b9SArnaldo Carvalho de Melo int ato = icsk->icsk_ack.ato; 35031da177e4SLinus Torvalds unsigned long timeout; 35041da177e4SLinus Torvalds 35059890092eSFlorian Westphal tcp_ca_event(sk, CA_EVENT_DELAYED_ACK); 35069890092eSFlorian Westphal 35071da177e4SLinus Torvalds if (ato > TCP_DELACK_MIN) { 3508463c84b9SArnaldo Carvalho de Melo const struct tcp_sock *tp = tcp_sk(sk); 35091da177e4SLinus Torvalds int max_ato = HZ / 2; 35101da177e4SLinus Torvalds 3511056834d9SIlpo Järvinen if (icsk->icsk_ack.pingpong || 3512056834d9SIlpo Järvinen (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 35131da177e4SLinus Torvalds max_ato = TCP_DELACK_MAX; 35141da177e4SLinus Torvalds 35151da177e4SLinus Torvalds /* Slow path, intersegment interval is "high". */ 35161da177e4SLinus Torvalds 35171da177e4SLinus Torvalds /* If some rtt estimate is known, use it to bound delayed ack. 3518463c84b9SArnaldo Carvalho de Melo * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 35191da177e4SLinus Torvalds * directly. 35201da177e4SLinus Torvalds */ 3521740b0f18SEric Dumazet if (tp->srtt_us) { 3522740b0f18SEric Dumazet int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3), 3523740b0f18SEric Dumazet TCP_DELACK_MIN); 35241da177e4SLinus Torvalds 35251da177e4SLinus Torvalds if (rtt < max_ato) 35261da177e4SLinus Torvalds max_ato = rtt; 35271da177e4SLinus Torvalds } 35281da177e4SLinus Torvalds 35291da177e4SLinus Torvalds ato = min(ato, max_ato); 35301da177e4SLinus Torvalds } 35311da177e4SLinus Torvalds 35321da177e4SLinus Torvalds /* Stay within the limit we were given */ 35331da177e4SLinus Torvalds timeout = jiffies + ato; 35341da177e4SLinus Torvalds 35351da177e4SLinus Torvalds /* Use new timeout only if there wasn't a older one earlier. */ 3536463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 35371da177e4SLinus Torvalds /* If delack timer was blocked or is about to expire, 35381da177e4SLinus Torvalds * send ACK now. 35391da177e4SLinus Torvalds */ 3540463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.blocked || 3541463c84b9SArnaldo Carvalho de Melo time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 35421da177e4SLinus Torvalds tcp_send_ack(sk); 35431da177e4SLinus Torvalds return; 35441da177e4SLinus Torvalds } 35451da177e4SLinus Torvalds 3546463c84b9SArnaldo Carvalho de Melo if (!time_before(timeout, icsk->icsk_ack.timeout)) 3547463c84b9SArnaldo Carvalho de Melo timeout = icsk->icsk_ack.timeout; 35481da177e4SLinus Torvalds } 3549463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 3550463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.timeout = timeout; 3551463c84b9SArnaldo Carvalho de Melo sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 35521da177e4SLinus Torvalds } 35531da177e4SLinus Torvalds 35541da177e4SLinus Torvalds /* This routine sends an ack and also updates the window. */ 35551da177e4SLinus Torvalds void tcp_send_ack(struct sock *sk) 35561da177e4SLinus Torvalds { 35571da177e4SLinus Torvalds struct sk_buff *buff; 35581da177e4SLinus Torvalds 3559058dc334SIlpo Järvinen /* If we have been reset, we may not send again. */ 3560058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3561058dc334SIlpo Järvinen return; 3562058dc334SIlpo Järvinen 35639890092eSFlorian Westphal tcp_ca_event(sk, CA_EVENT_NON_DELAYED_ACK); 35649890092eSFlorian Westphal 35651da177e4SLinus Torvalds /* We are not putting this on the write queue, so 35661da177e4SLinus Torvalds * tcp_transmit_skb() will set the ownership to this 35671da177e4SLinus Torvalds * sock. 35681da177e4SLinus Torvalds */ 35697450aaf6SEric Dumazet buff = alloc_skb(MAX_TCP_HEADER, 35707450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 35717450aaf6SEric Dumazet if (unlikely(!buff)) { 3572463c84b9SArnaldo Carvalho de Melo inet_csk_schedule_ack(sk); 3573463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; 35743f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 35753f421baaSArnaldo Carvalho de Melo TCP_DELACK_MAX, TCP_RTO_MAX); 35761da177e4SLinus Torvalds return; 35771da177e4SLinus Torvalds } 35781da177e4SLinus Torvalds 35791da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 35801da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 3581a3433f35SChangli Gao tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); 35821da177e4SLinus Torvalds 358398781965SEric Dumazet /* We do not want pure acks influencing TCP Small Queues or fq/pacing 358498781965SEric Dumazet * too much. 358598781965SEric Dumazet * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784 358698781965SEric Dumazet */ 358798781965SEric Dumazet skb_set_tcp_pure_ack(buff); 358898781965SEric Dumazet 35891da177e4SLinus Torvalds /* Send it off, this clears delayed acks for us. */ 35907450aaf6SEric Dumazet tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0); 35911da177e4SLinus Torvalds } 3592e3118e83SDaniel Borkmann EXPORT_SYMBOL_GPL(tcp_send_ack); 35931da177e4SLinus Torvalds 35941da177e4SLinus Torvalds /* This routine sends a packet with an out of date sequence 35951da177e4SLinus Torvalds * number. It assumes the other end will try to ack it. 35961da177e4SLinus Torvalds * 35971da177e4SLinus Torvalds * Question: what should we make while urgent mode? 35981da177e4SLinus Torvalds * 4.4BSD forces sending single byte of data. We cannot send 35991da177e4SLinus Torvalds * out of window data, because we have SND.NXT==SND.MAX... 36001da177e4SLinus Torvalds * 36011da177e4SLinus Torvalds * Current solution: to send TWO zero-length segments in urgent mode: 36021da177e4SLinus Torvalds * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 36031da177e4SLinus Torvalds * out-of-date with SND.UNA-1 to probe window. 36041da177e4SLinus Torvalds */ 3605e520af48SEric Dumazet static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib) 36061da177e4SLinus Torvalds { 36071da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 36081da177e4SLinus Torvalds struct sk_buff *skb; 36091da177e4SLinus Torvalds 36101da177e4SLinus Torvalds /* We don't queue it, tcp_transmit_skb() sets ownership. */ 36117450aaf6SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, 36127450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 361351456b29SIan Morris if (!skb) 36141da177e4SLinus Torvalds return -1; 36151da177e4SLinus Torvalds 36161da177e4SLinus Torvalds /* Reserve space for headers and set control bits. */ 36171da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 36181da177e4SLinus Torvalds /* Use a previous sequence. This should cause the other 36191da177e4SLinus Torvalds * end to send an ack. Don't queue or clone SKB, just 36201da177e4SLinus Torvalds * send it. 36211da177e4SLinus Torvalds */ 3622a3433f35SChangli Gao tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); 3623e2e8009fSRenato Westphal NET_INC_STATS(sock_net(sk), mib); 36247450aaf6SEric Dumazet return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0); 36251da177e4SLinus Torvalds } 36261da177e4SLinus Torvalds 3627385e2070SEric Dumazet /* Called from setsockopt( ... TCP_REPAIR ) */ 3628ee995283SPavel Emelyanov void tcp_send_window_probe(struct sock *sk) 3629ee995283SPavel Emelyanov { 3630ee995283SPavel Emelyanov if (sk->sk_state == TCP_ESTABLISHED) { 3631ee995283SPavel Emelyanov tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; 36329a568de4SEric Dumazet tcp_mstamp_refresh(tcp_sk(sk)); 3633e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE); 3634ee995283SPavel Emelyanov } 3635ee995283SPavel Emelyanov } 3636ee995283SPavel Emelyanov 363767edfef7SAndi Kleen /* Initiate keepalive or window probe from timer. */ 3638e520af48SEric Dumazet int tcp_write_wakeup(struct sock *sk, int mib) 36391da177e4SLinus Torvalds { 36401da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 36411da177e4SLinus Torvalds struct sk_buff *skb; 36421da177e4SLinus Torvalds 3643058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3644058dc334SIlpo Järvinen return -1; 3645058dc334SIlpo Järvinen 364600db4124SIan Morris skb = tcp_send_head(sk); 364700db4124SIan Morris if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { 36481da177e4SLinus Torvalds int err; 36490c54b85fSIlpo Järvinen unsigned int mss = tcp_current_mss(sk); 365090840defSIlpo Järvinen unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 36511da177e4SLinus Torvalds 36521da177e4SLinus Torvalds if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 36531da177e4SLinus Torvalds tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 36541da177e4SLinus Torvalds 36551da177e4SLinus Torvalds /* We are probing the opening of a window 36561da177e4SLinus Torvalds * but the window size is != 0 36571da177e4SLinus Torvalds * must have been a result SWS avoidance ( sender ) 36581da177e4SLinus Torvalds */ 36591da177e4SLinus Torvalds if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 36601da177e4SLinus Torvalds skb->len > mss) { 36611da177e4SLinus Torvalds seg_size = min(seg_size, mss); 36624de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 366375c119afSEric Dumazet if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE, 366475c119afSEric Dumazet skb, seg_size, mss, GFP_ATOMIC)) 36651da177e4SLinus Torvalds return -1; 36661da177e4SLinus Torvalds } else if (!tcp_skb_pcount(skb)) 36675bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss); 36681da177e4SLinus Torvalds 36694de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 3670dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 367166f5fe62SIlpo Järvinen if (!err) 367266f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 36731da177e4SLinus Torvalds return err; 36741da177e4SLinus Torvalds } else { 367533f5f57eSIlpo Järvinen if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) 3676e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 1, mib); 3677e520af48SEric Dumazet return tcp_xmit_probe_skb(sk, 0, mib); 36781da177e4SLinus Torvalds } 36791da177e4SLinus Torvalds } 36801da177e4SLinus Torvalds 36811da177e4SLinus Torvalds /* A window probe timeout has occurred. If window is not closed send 36821da177e4SLinus Torvalds * a partial packet else a zero probe. 36831da177e4SLinus Torvalds */ 36841da177e4SLinus Torvalds void tcp_send_probe0(struct sock *sk) 36851da177e4SLinus Torvalds { 3686463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 36871da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 3688c6214a97SNikolay Borisov struct net *net = sock_net(sk); 3689fcdd1cf4SEric Dumazet unsigned long probe_max; 36901da177e4SLinus Torvalds int err; 36911da177e4SLinus Torvalds 3692e520af48SEric Dumazet err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE); 36931da177e4SLinus Torvalds 369475c119afSEric Dumazet if (tp->packets_out || tcp_write_queue_empty(sk)) { 36951da177e4SLinus Torvalds /* Cancel probe timer, if it is not required. */ 36966687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 0; 3697463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff = 0; 36981da177e4SLinus Torvalds return; 36991da177e4SLinus Torvalds } 37001da177e4SLinus Torvalds 37011da177e4SLinus Torvalds if (err <= 0) { 3702c6214a97SNikolay Borisov if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2) 3703463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff++; 37046687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out++; 3705fcdd1cf4SEric Dumazet probe_max = TCP_RTO_MAX; 37061da177e4SLinus Torvalds } else { 37071da177e4SLinus Torvalds /* If packet was not sent due to local congestion, 37086687e988SArnaldo Carvalho de Melo * do not backoff and do not remember icsk_probes_out. 37091da177e4SLinus Torvalds * Let local senders to fight for local resources. 37101da177e4SLinus Torvalds * 37111da177e4SLinus Torvalds * Use accumulated backoff yet. 37121da177e4SLinus Torvalds */ 37136687e988SArnaldo Carvalho de Melo if (!icsk->icsk_probes_out) 37146687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 1; 3715fcdd1cf4SEric Dumazet probe_max = TCP_RESOURCE_PROBE_INTERVAL; 37161da177e4SLinus Torvalds } 3717fcdd1cf4SEric Dumazet inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 371821c8fe99SEric Dumazet tcp_probe0_when(sk, probe_max), 3719fcdd1cf4SEric Dumazet TCP_RTO_MAX); 37201da177e4SLinus Torvalds } 37215db92c99SOctavian Purdila 3722ea3bea3aSEric Dumazet int tcp_rtx_synack(const struct sock *sk, struct request_sock *req) 37235db92c99SOctavian Purdila { 37245db92c99SOctavian Purdila const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific; 37255db92c99SOctavian Purdila struct flowi fl; 37265db92c99SOctavian Purdila int res; 37275db92c99SOctavian Purdila 372858d607d3SEric Dumazet tcp_rsk(req)->txhash = net_tx_rndhash(); 3729b3d05147SEric Dumazet res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL); 37305db92c99SOctavian Purdila if (!res) { 373190bbcc60SEric Dumazet __TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); 373202a1d6e7SEric Dumazet __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 37337e32b443SYuchung Cheng if (unlikely(tcp_passive_fastopen(sk))) 37347e32b443SYuchung Cheng tcp_sk(sk)->total_retrans++; 37355db92c99SOctavian Purdila } 37365db92c99SOctavian Purdila return res; 37375db92c99SOctavian Purdila } 37385db92c99SOctavian Purdila EXPORT_SYMBOL(tcp_rtx_synack); 3739