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> 4460e2a778SUrsula Braun #include <linux/static_key.h> 451da177e4SLinus Torvalds 46e086101bSCong Wang #include <trace/events/tcp.h> 47e086101bSCong Wang 4846d3ceabSEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 4946d3ceabSEric Dumazet int push_one, gfp_t gfp); 50519855c5SWilliam Allen Simpson 5167edfef7SAndi Kleen /* Account for new data that has been sent to the network. */ 5275c119afSEric Dumazet static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb) 536ff03ac3SIlpo Järvinen { 546ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 556ff03ac3SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 5666f5fe62SIlpo Järvinen unsigned int prior_packets = tp->packets_out; 579e412ba7SIlpo Järvinen 581da177e4SLinus Torvalds tp->snd_nxt = TCP_SKB_CB(skb)->end_seq; 598512430eSIlpo Järvinen 6075c119afSEric Dumazet __skb_unlink(skb, &sk->sk_write_queue); 6175c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, skb); 6275c119afSEric Dumazet 6366f5fe62SIlpo Järvinen tp->packets_out += tcp_skb_pcount(skb); 64bec41a11SYuchung Cheng if (!prior_packets || icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) 65750ea2baSYuchung Cheng tcp_rearm_rto(sk); 66f19c29e3SYuchung Cheng 67f7324acdSDavid S. Miller NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT, 68f19c29e3SYuchung Cheng tcp_skb_pcount(skb)); 696a5dc9e5SEric Dumazet } 701da177e4SLinus Torvalds 71a4ecb15aSCui, Cheng /* SND.NXT, if window was not shrunk or the amount of shrunk was less than one 72a4ecb15aSCui, Cheng * window scaling factor due to loss of precision. 731da177e4SLinus Torvalds * If window has been shrunk, what should we make? It is not clear at all. 741da177e4SLinus Torvalds * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-( 751da177e4SLinus Torvalds * Anything in between SND.UNA...SND.UNA+SND.WND also can be already 761da177e4SLinus Torvalds * invalid. OK, let's make this for now: 771da177e4SLinus Torvalds */ 78cf533ea5SEric Dumazet static inline __u32 tcp_acceptable_seq(const struct sock *sk) 791da177e4SLinus Torvalds { 80cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 819e412ba7SIlpo Järvinen 82a4ecb15aSCui, Cheng if (!before(tcp_wnd_end(tp), tp->snd_nxt) || 83a4ecb15aSCui, Cheng (tp->rx_opt.wscale_ok && 84a4ecb15aSCui, Cheng ((tp->snd_nxt - tcp_wnd_end(tp)) < (1 << tp->rx_opt.rcv_wscale)))) 851da177e4SLinus Torvalds return tp->snd_nxt; 861da177e4SLinus Torvalds else 8790840defSIlpo Järvinen return tcp_wnd_end(tp); 881da177e4SLinus Torvalds } 891da177e4SLinus Torvalds 901da177e4SLinus Torvalds /* Calculate mss to advertise in SYN segment. 911da177e4SLinus Torvalds * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that: 921da177e4SLinus Torvalds * 931da177e4SLinus Torvalds * 1. It is independent of path mtu. 941da177e4SLinus Torvalds * 2. Ideally, it is maximal possible segment size i.e. 65535-40. 951da177e4SLinus Torvalds * 3. For IPv4 it is reasonable to calculate it from maximal MTU of 961da177e4SLinus Torvalds * attached devices, because some buggy hosts are confused by 971da177e4SLinus Torvalds * large MSS. 981da177e4SLinus Torvalds * 4. We do not make 3, we advertise MSS, calculated from first 991da177e4SLinus Torvalds * hop device mtu, but allow to raise it to ip_rt_min_advmss. 1001da177e4SLinus Torvalds * This may be overridden via information stored in routing table. 1011da177e4SLinus Torvalds * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible, 1021da177e4SLinus Torvalds * probably even Jumbo". 1031da177e4SLinus Torvalds */ 1041da177e4SLinus Torvalds static __u16 tcp_advertise_mss(struct sock *sk) 1051da177e4SLinus Torvalds { 1061da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 107cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1081da177e4SLinus Torvalds int mss = tp->advmss; 1091da177e4SLinus Torvalds 1100dbaee3bSDavid S. Miller if (dst) { 1110dbaee3bSDavid S. Miller unsigned int metric = dst_metric_advmss(dst); 1120dbaee3bSDavid S. Miller 1130dbaee3bSDavid S. Miller if (metric < mss) { 1140dbaee3bSDavid S. Miller mss = metric; 1151da177e4SLinus Torvalds tp->advmss = mss; 1161da177e4SLinus Torvalds } 1170dbaee3bSDavid S. Miller } 1181da177e4SLinus Torvalds 1191da177e4SLinus Torvalds return (__u16)mss; 1201da177e4SLinus Torvalds } 1211da177e4SLinus Torvalds 1221da177e4SLinus Torvalds /* RFC2861. Reset CWND after idle period longer RTO to "restart window". 1236f021c62SEric Dumazet * This is the first part of cwnd validation mechanism. 1246f021c62SEric Dumazet */ 1256f021c62SEric Dumazet void tcp_cwnd_restart(struct sock *sk, s32 delta) 1261da177e4SLinus Torvalds { 127463c84b9SArnaldo Carvalho de Melo struct tcp_sock *tp = tcp_sk(sk); 1286f021c62SEric Dumazet u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk)); 1291da177e4SLinus Torvalds u32 cwnd = tp->snd_cwnd; 1301da177e4SLinus Torvalds 1316687e988SArnaldo Carvalho de Melo tcp_ca_event(sk, CA_EVENT_CWND_RESTART); 1321da177e4SLinus Torvalds 1336687e988SArnaldo Carvalho de Melo tp->snd_ssthresh = tcp_current_ssthresh(sk); 1341da177e4SLinus Torvalds restart_cwnd = min(restart_cwnd, cwnd); 1351da177e4SLinus Torvalds 136463c84b9SArnaldo Carvalho de Melo while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd) 1371da177e4SLinus Torvalds cwnd >>= 1; 1381da177e4SLinus Torvalds tp->snd_cwnd = max(cwnd, restart_cwnd); 139c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 1401da177e4SLinus Torvalds tp->snd_cwnd_used = 0; 1411da177e4SLinus Torvalds } 1421da177e4SLinus Torvalds 14367edfef7SAndi Kleen /* Congestion state accounting after a packet has been sent. */ 14440efc6faSStephen Hemminger static void tcp_event_data_sent(struct tcp_sock *tp, 145cf533ea5SEric Dumazet struct sock *sk) 1461da177e4SLinus Torvalds { 147463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 148d635fbe2SEric Dumazet const u32 now = tcp_jiffies32; 1491da177e4SLinus Torvalds 15005c5a46dSNeal Cardwell if (tcp_packets_in_flight(tp) == 0) 15105c5a46dSNeal Cardwell tcp_ca_event(sk, CA_EVENT_TX_START); 15205c5a46dSNeal Cardwell 1531da177e4SLinus Torvalds tp->lsndtime = now; 1541da177e4SLinus Torvalds 1551da177e4SLinus Torvalds /* If it is a reply for ato after last received 1561da177e4SLinus Torvalds * packet, enter pingpong mode. 1571da177e4SLinus Torvalds */ 1582251ae46SJon Maxwell if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato) 159463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pingpong = 1; 1601da177e4SLinus Torvalds } 1611da177e4SLinus Torvalds 16267edfef7SAndi Kleen /* Account for an ACK we sent. */ 16340efc6faSStephen Hemminger static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts) 1641da177e4SLinus Torvalds { 165463c84b9SArnaldo Carvalho de Melo tcp_dec_quickack_mode(sk, pkts); 166463c84b9SArnaldo Carvalho de Melo inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); 1671da177e4SLinus Torvalds } 1681da177e4SLinus Torvalds 16985f16525SYuchung Cheng 17085f16525SYuchung Cheng u32 tcp_default_init_rwnd(u32 mss) 17185f16525SYuchung Cheng { 17285f16525SYuchung Cheng /* Initial receive window should be twice of TCP_INIT_CWND to 1739ef71e0cSWeiping Pan * enable proper sending of new unsent data during fast recovery 17485f16525SYuchung Cheng * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a 17585f16525SYuchung Cheng * limit when mss is larger than 1460. 17685f16525SYuchung Cheng */ 17785f16525SYuchung Cheng u32 init_rwnd = TCP_INIT_CWND * 2; 17885f16525SYuchung Cheng 17985f16525SYuchung Cheng if (mss > 1460) 18085f16525SYuchung Cheng init_rwnd = max((1460 * init_rwnd) / mss, 2U); 18185f16525SYuchung Cheng return init_rwnd; 18285f16525SYuchung Cheng } 18385f16525SYuchung Cheng 1841da177e4SLinus Torvalds /* Determine a window scaling and initial window to offer. 1851da177e4SLinus Torvalds * Based on the assumption that the given amount of space 1861da177e4SLinus Torvalds * will be offered. Store the results in the tp structure. 1871da177e4SLinus Torvalds * NOTE: for smooth operation initial space offering should 1881da177e4SLinus Torvalds * be a multiple of mss if possible. We assume here that mss >= 1. 1891da177e4SLinus Torvalds * This MUST be enforced by all callers. 1901da177e4SLinus Torvalds */ 191ceef9ab6SEric Dumazet void tcp_select_initial_window(const struct sock *sk, int __space, __u32 mss, 1921da177e4SLinus Torvalds __u32 *rcv_wnd, __u32 *window_clamp, 19331d12926Slaurent chavey int wscale_ok, __u8 *rcv_wscale, 19431d12926Slaurent chavey __u32 init_rcv_wnd) 1951da177e4SLinus Torvalds { 1961da177e4SLinus Torvalds unsigned int space = (__space < 0 ? 0 : __space); 1971da177e4SLinus Torvalds 1981da177e4SLinus Torvalds /* If no clamp set the clamp to the max possible scaled window */ 1991da177e4SLinus Torvalds if (*window_clamp == 0) 200589c49cbSGao Feng (*window_clamp) = (U16_MAX << TCP_MAX_WSCALE); 2011da177e4SLinus Torvalds space = min(*window_clamp, space); 2021da177e4SLinus Torvalds 2031da177e4SLinus Torvalds /* Quantize space offering to a multiple of mss if possible. */ 2041da177e4SLinus Torvalds if (space > mss) 205589c49cbSGao Feng space = rounddown(space, mss); 2061da177e4SLinus Torvalds 2071da177e4SLinus Torvalds /* NOTE: offering an initial window larger than 32767 20815d99e02SRick Jones * will break some buggy TCP stacks. If the admin tells us 20915d99e02SRick Jones * it is likely we could be speaking with such a buggy stack 21015d99e02SRick Jones * we will truncate our initial window offering to 32K-1 21115d99e02SRick Jones * unless the remote has sent us a window scaling option, 21215d99e02SRick Jones * which we interpret as a sign the remote TCP is not 21315d99e02SRick Jones * misinterpreting the window field as a signed quantity. 2141da177e4SLinus Torvalds */ 215ceef9ab6SEric Dumazet if (sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows) 2161da177e4SLinus Torvalds (*rcv_wnd) = min(space, MAX_TCP_WINDOW); 21715d99e02SRick Jones else 21815d99e02SRick Jones (*rcv_wnd) = space; 21915d99e02SRick Jones 2201da177e4SLinus Torvalds (*rcv_wscale) = 0; 2211da177e4SLinus Torvalds if (wscale_ok) { 222589c49cbSGao Feng /* Set window scaling on max possible window */ 223f626300aSSoheil Hassas Yeganeh space = max_t(u32, space, sysctl_tcp_rmem[2]); 224f626300aSSoheil Hassas Yeganeh space = max_t(u32, space, sysctl_rmem_max); 225316c1592SStephen Hemminger space = min_t(u32, space, *window_clamp); 226589c49cbSGao Feng while (space > U16_MAX && (*rcv_wscale) < TCP_MAX_WSCALE) { 2271da177e4SLinus Torvalds space >>= 1; 2281da177e4SLinus Torvalds (*rcv_wscale)++; 2291da177e4SLinus Torvalds } 2301da177e4SLinus Torvalds } 2311da177e4SLinus Torvalds 2321da177e4SLinus Torvalds if (mss > (1 << *rcv_wscale)) { 23385f16525SYuchung Cheng if (!init_rcv_wnd) /* Use default unless specified otherwise */ 23485f16525SYuchung Cheng init_rcv_wnd = tcp_default_init_rwnd(mss); 235b1afde60SNandita Dukkipati *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss); 2361da177e4SLinus Torvalds } 2371da177e4SLinus Torvalds 2381da177e4SLinus Torvalds /* Set the clamp no higher than max representable value */ 239589c49cbSGao Feng (*window_clamp) = min_t(__u32, U16_MAX << (*rcv_wscale), *window_clamp); 2401da177e4SLinus Torvalds } 2414bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_select_initial_window); 2421da177e4SLinus Torvalds 2431da177e4SLinus Torvalds /* Chose a new window to advertise, update state in tcp_sock for the 2441da177e4SLinus Torvalds * socket, and return result with RFC1323 scaling applied. The return 2451da177e4SLinus Torvalds * value can be stuffed directly into th->window for an outgoing 2461da177e4SLinus Torvalds * frame. 2471da177e4SLinus Torvalds */ 24840efc6faSStephen Hemminger static u16 tcp_select_window(struct sock *sk) 2491da177e4SLinus Torvalds { 2501da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2518e165e20SFlorian Westphal u32 old_win = tp->rcv_wnd; 2521da177e4SLinus Torvalds u32 cur_win = tcp_receive_window(tp); 2531da177e4SLinus Torvalds u32 new_win = __tcp_select_window(sk); 2541da177e4SLinus Torvalds 2551da177e4SLinus Torvalds /* Never shrink the offered window */ 2561da177e4SLinus Torvalds if (new_win < cur_win) { 2571da177e4SLinus Torvalds /* Danger Will Robinson! 2581da177e4SLinus Torvalds * Don't update rcv_wup/rcv_wnd here or else 2591da177e4SLinus Torvalds * we will not be able to advertise a zero 2601da177e4SLinus Torvalds * window in time. --DaveM 2611da177e4SLinus Torvalds * 2621da177e4SLinus Torvalds * Relax Will Robinson. 2631da177e4SLinus Torvalds */ 2648e165e20SFlorian Westphal if (new_win == 0) 2658e165e20SFlorian Westphal NET_INC_STATS(sock_net(sk), 2668e165e20SFlorian Westphal LINUX_MIB_TCPWANTZEROWINDOWADV); 267607bfbf2SPatrick McHardy new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale); 2681da177e4SLinus Torvalds } 2691da177e4SLinus Torvalds tp->rcv_wnd = new_win; 2701da177e4SLinus Torvalds tp->rcv_wup = tp->rcv_nxt; 2711da177e4SLinus Torvalds 2721da177e4SLinus Torvalds /* Make sure we do not exceed the maximum possible 2731da177e4SLinus Torvalds * scaled window. 2741da177e4SLinus Torvalds */ 275ceef9ab6SEric Dumazet if (!tp->rx_opt.rcv_wscale && 276ceef9ab6SEric Dumazet sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows) 2771da177e4SLinus Torvalds new_win = min(new_win, MAX_TCP_WINDOW); 2781da177e4SLinus Torvalds else 2791da177e4SLinus Torvalds new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); 2801da177e4SLinus Torvalds 2811da177e4SLinus Torvalds /* RFC1323 scaling applied */ 2821da177e4SLinus Torvalds new_win >>= tp->rx_opt.rcv_wscale; 2831da177e4SLinus Torvalds 28431770e34SFlorian Westphal /* If we advertise zero window, disable fast path. */ 2858e165e20SFlorian Westphal if (new_win == 0) { 28631770e34SFlorian Westphal tp->pred_flags = 0; 2878e165e20SFlorian Westphal if (old_win) 2888e165e20SFlorian Westphal NET_INC_STATS(sock_net(sk), 2898e165e20SFlorian Westphal LINUX_MIB_TCPTOZEROWINDOWADV); 2908e165e20SFlorian Westphal } else if (old_win == 0) { 2918e165e20SFlorian Westphal NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV); 2928e165e20SFlorian Westphal } 2931da177e4SLinus Torvalds 2941da177e4SLinus Torvalds return new_win; 2951da177e4SLinus Torvalds } 2961da177e4SLinus Torvalds 29767edfef7SAndi Kleen /* Packet ECN state for a SYN-ACK */ 298735d3831SFlorian Westphal static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb) 299bdf1ee5dSIlpo Järvinen { 30030e502a3SDaniel Borkmann const struct tcp_sock *tp = tcp_sk(sk); 30130e502a3SDaniel Borkmann 3024de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR; 303bdf1ee5dSIlpo Järvinen if (!(tp->ecn_flags & TCP_ECN_OK)) 3044de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE; 30591b5b21cSLawrence Brakmo else if (tcp_ca_needs_ecn(sk) || 30691b5b21cSLawrence Brakmo tcp_bpf_ca_needs_ecn(sk)) 30730e502a3SDaniel Borkmann INET_ECN_xmit(sk); 308bdf1ee5dSIlpo Järvinen } 309bdf1ee5dSIlpo Järvinen 31067edfef7SAndi Kleen /* Packet ECN state for a SYN. */ 311735d3831SFlorian Westphal static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb) 312bdf1ee5dSIlpo Järvinen { 313bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 31491b5b21cSLawrence Brakmo bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk); 315f7b3bec6SFlorian Westphal bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 || 31691b5b21cSLawrence Brakmo tcp_ca_needs_ecn(sk) || bpf_needs_ecn; 317f7b3bec6SFlorian Westphal 318f7b3bec6SFlorian Westphal if (!use_ecn) { 319f7b3bec6SFlorian Westphal const struct dst_entry *dst = __sk_dst_get(sk); 320f7b3bec6SFlorian Westphal 321f7b3bec6SFlorian Westphal if (dst && dst_feature(dst, RTAX_FEATURE_ECN)) 322f7b3bec6SFlorian Westphal use_ecn = true; 323f7b3bec6SFlorian Westphal } 324bdf1ee5dSIlpo Järvinen 325bdf1ee5dSIlpo Järvinen tp->ecn_flags = 0; 326f7b3bec6SFlorian Westphal 327f7b3bec6SFlorian Westphal if (use_ecn) { 3284de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR; 329bdf1ee5dSIlpo Järvinen tp->ecn_flags = TCP_ECN_OK; 33091b5b21cSLawrence Brakmo if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn) 33130e502a3SDaniel Borkmann INET_ECN_xmit(sk); 332bdf1ee5dSIlpo Järvinen } 333bdf1ee5dSIlpo Järvinen } 334bdf1ee5dSIlpo Järvinen 33549213555SDaniel Borkmann static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb) 33649213555SDaniel Borkmann { 33749213555SDaniel Borkmann if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback) 33849213555SDaniel Borkmann /* tp->ecn_flags are cleared at a later point in time when 33949213555SDaniel Borkmann * SYN ACK is ultimatively being received. 34049213555SDaniel Borkmann */ 34149213555SDaniel Borkmann TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR); 34249213555SDaniel Borkmann } 34349213555SDaniel Borkmann 344735d3831SFlorian Westphal static void 3456ac705b1SEric Dumazet tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th) 346bdf1ee5dSIlpo Järvinen { 3476ac705b1SEric Dumazet if (inet_rsk(req)->ecn_ok) 348bdf1ee5dSIlpo Järvinen th->ece = 1; 349bdf1ee5dSIlpo Järvinen } 350bdf1ee5dSIlpo Järvinen 35167edfef7SAndi Kleen /* Set up ECN state for a packet on a ESTABLISHED socket that is about to 35267edfef7SAndi Kleen * be sent. 35367edfef7SAndi Kleen */ 354735d3831SFlorian Westphal static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb, 355ea1627c2SEric Dumazet struct tcphdr *th, int tcp_header_len) 356bdf1ee5dSIlpo Järvinen { 357bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 358bdf1ee5dSIlpo Järvinen 359bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_OK) { 360bdf1ee5dSIlpo Järvinen /* Not-retransmitted data segment: set ECT and inject CWR. */ 361bdf1ee5dSIlpo Järvinen if (skb->len != tcp_header_len && 362bdf1ee5dSIlpo Järvinen !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { 363bdf1ee5dSIlpo Järvinen INET_ECN_xmit(sk); 364bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) { 365bdf1ee5dSIlpo Järvinen tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; 366ea1627c2SEric Dumazet th->cwr = 1; 367bdf1ee5dSIlpo Järvinen skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; 368bdf1ee5dSIlpo Järvinen } 36930e502a3SDaniel Borkmann } else if (!tcp_ca_needs_ecn(sk)) { 370bdf1ee5dSIlpo Järvinen /* ACK or retransmitted segment: clear ECT|CE */ 371bdf1ee5dSIlpo Järvinen INET_ECN_dontxmit(sk); 372bdf1ee5dSIlpo Järvinen } 373bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) 374ea1627c2SEric Dumazet th->ece = 1; 375bdf1ee5dSIlpo Järvinen } 376bdf1ee5dSIlpo Järvinen } 377bdf1ee5dSIlpo Järvinen 378e870a8efSIlpo Järvinen /* Constructs common control bits of non-data skb. If SYN/FIN is present, 379e870a8efSIlpo Järvinen * auto increment end seqno. 380e870a8efSIlpo Järvinen */ 381e870a8efSIlpo Järvinen static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags) 382e870a8efSIlpo Järvinen { 3832e8e18efSDavid S. Miller skb->ip_summed = CHECKSUM_PARTIAL; 384e870a8efSIlpo Järvinen skb->csum = 0; 385e870a8efSIlpo Järvinen 3864de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags; 387e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->sacked = 0; 388e870a8efSIlpo Järvinen 389cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, 1); 390e870a8efSIlpo Järvinen 391e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->seq = seq; 392a3433f35SChangli Gao if (flags & (TCPHDR_SYN | TCPHDR_FIN)) 393e870a8efSIlpo Järvinen seq++; 394e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->end_seq = seq; 395e870a8efSIlpo Järvinen } 396e870a8efSIlpo Järvinen 397a2a385d6SEric Dumazet static inline bool tcp_urg_mode(const struct tcp_sock *tp) 39833f5f57eSIlpo Järvinen { 39933f5f57eSIlpo Järvinen return tp->snd_una != tp->snd_up; 40033f5f57eSIlpo Järvinen } 40133f5f57eSIlpo Järvinen 40233ad798cSAdam Langley #define OPTION_SACK_ADVERTISE (1 << 0) 40333ad798cSAdam Langley #define OPTION_TS (1 << 1) 40433ad798cSAdam Langley #define OPTION_MD5 (1 << 2) 40589e95a61SOri Finkelman #define OPTION_WSCALE (1 << 3) 4062100c8d2SYuchung Cheng #define OPTION_FAST_OPEN_COOKIE (1 << 8) 40760e2a778SUrsula Braun #define OPTION_SMC (1 << 9) 40860e2a778SUrsula Braun 40960e2a778SUrsula Braun static void smc_options_write(__be32 *ptr, u16 *options) 41060e2a778SUrsula Braun { 41160e2a778SUrsula Braun #if IS_ENABLED(CONFIG_SMC) 41260e2a778SUrsula Braun if (static_branch_unlikely(&tcp_have_smc)) { 41360e2a778SUrsula Braun if (unlikely(OPTION_SMC & *options)) { 41460e2a778SUrsula Braun *ptr++ = htonl((TCPOPT_NOP << 24) | 41560e2a778SUrsula Braun (TCPOPT_NOP << 16) | 41660e2a778SUrsula Braun (TCPOPT_EXP << 8) | 41760e2a778SUrsula Braun (TCPOLEN_EXP_SMC_BASE)); 41860e2a778SUrsula Braun *ptr++ = htonl(TCPOPT_SMC_MAGIC); 41960e2a778SUrsula Braun } 42060e2a778SUrsula Braun } 42160e2a778SUrsula Braun #endif 42260e2a778SUrsula Braun } 42333ad798cSAdam Langley 42433ad798cSAdam Langley struct tcp_out_options { 4252100c8d2SYuchung Cheng u16 options; /* bit field of OPTION_* */ 4262100c8d2SYuchung Cheng u16 mss; /* 0 to disable */ 42733ad798cSAdam Langley u8 ws; /* window scale, 0 to disable */ 42833ad798cSAdam Langley u8 num_sack_blocks; /* number of SACK blocks to include */ 429bd0388aeSWilliam Allen Simpson u8 hash_size; /* bytes in hash_location */ 430bd0388aeSWilliam Allen Simpson __u8 *hash_location; /* temporary pointer, overloaded */ 4312100c8d2SYuchung Cheng __u32 tsval, tsecr; /* need to include OPTION_TS */ 4322100c8d2SYuchung Cheng struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */ 43333ad798cSAdam Langley }; 43433ad798cSAdam Langley 43567edfef7SAndi Kleen /* Write previously computed TCP options to the packet. 43667edfef7SAndi Kleen * 43767edfef7SAndi Kleen * Beware: Something in the Internet is very sensitive to the ordering of 438fd6149d3SIlpo Järvinen * TCP options, we learned this through the hard way, so be careful here. 439fd6149d3SIlpo Järvinen * Luckily we can at least blame others for their non-compliance but from 4408e3bff96Sstephen hemminger * inter-operability perspective it seems that we're somewhat stuck with 441fd6149d3SIlpo Järvinen * the ordering which we have been using if we want to keep working with 442fd6149d3SIlpo Järvinen * those broken things (not that it currently hurts anybody as there isn't 443fd6149d3SIlpo Järvinen * particular reason why the ordering would need to be changed). 444fd6149d3SIlpo Järvinen * 445fd6149d3SIlpo Järvinen * At least SACK_PERM as the first option is known to lead to a disaster 446fd6149d3SIlpo Järvinen * (but it may well be that other scenarios fail similarly). 447fd6149d3SIlpo Järvinen */ 44833ad798cSAdam Langley static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp, 449bd0388aeSWilliam Allen Simpson struct tcp_out_options *opts) 450bd0388aeSWilliam Allen Simpson { 4512100c8d2SYuchung Cheng u16 options = opts->options; /* mungable copy */ 452bd0388aeSWilliam Allen Simpson 453bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_MD5 & options)) { 4541a2c6181SChristoph Paasch *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | 4551a2c6181SChristoph Paasch (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG); 456bd0388aeSWilliam Allen Simpson /* overload cookie hash location */ 457bd0388aeSWilliam Allen Simpson opts->hash_location = (__u8 *)ptr; 45833ad798cSAdam Langley ptr += 4; 45933ad798cSAdam Langley } 46033ad798cSAdam Langley 461fd6149d3SIlpo Järvinen if (unlikely(opts->mss)) { 462fd6149d3SIlpo Järvinen *ptr++ = htonl((TCPOPT_MSS << 24) | 463fd6149d3SIlpo Järvinen (TCPOLEN_MSS << 16) | 464fd6149d3SIlpo Järvinen opts->mss); 465fd6149d3SIlpo Järvinen } 466fd6149d3SIlpo Järvinen 467bd0388aeSWilliam Allen Simpson if (likely(OPTION_TS & options)) { 468bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 46933ad798cSAdam Langley *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | 47033ad798cSAdam Langley (TCPOLEN_SACK_PERM << 16) | 47133ad798cSAdam Langley (TCPOPT_TIMESTAMP << 8) | 47233ad798cSAdam Langley TCPOLEN_TIMESTAMP); 473bd0388aeSWilliam Allen Simpson options &= ~OPTION_SACK_ADVERTISE; 47433ad798cSAdam Langley } else { 475496c98dfSYOSHIFUJI Hideaki *ptr++ = htonl((TCPOPT_NOP << 24) | 47640efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 47740efc6faSStephen Hemminger (TCPOPT_TIMESTAMP << 8) | 47840efc6faSStephen Hemminger TCPOLEN_TIMESTAMP); 47940efc6faSStephen Hemminger } 48033ad798cSAdam Langley *ptr++ = htonl(opts->tsval); 48133ad798cSAdam Langley *ptr++ = htonl(opts->tsecr); 48233ad798cSAdam Langley } 48333ad798cSAdam Langley 484bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 48533ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 48633ad798cSAdam Langley (TCPOPT_NOP << 16) | 48733ad798cSAdam Langley (TCPOPT_SACK_PERM << 8) | 48833ad798cSAdam Langley TCPOLEN_SACK_PERM); 48933ad798cSAdam Langley } 49033ad798cSAdam Langley 491bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_WSCALE & options)) { 49233ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 49333ad798cSAdam Langley (TCPOPT_WINDOW << 16) | 49433ad798cSAdam Langley (TCPOLEN_WINDOW << 8) | 49533ad798cSAdam Langley opts->ws); 49633ad798cSAdam Langley } 49733ad798cSAdam Langley 49833ad798cSAdam Langley if (unlikely(opts->num_sack_blocks)) { 49933ad798cSAdam Langley struct tcp_sack_block *sp = tp->rx_opt.dsack ? 50033ad798cSAdam Langley tp->duplicate_sack : tp->selective_acks; 50140efc6faSStephen Hemminger int this_sack; 50240efc6faSStephen Hemminger 50340efc6faSStephen Hemminger *ptr++ = htonl((TCPOPT_NOP << 24) | 50440efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 50540efc6faSStephen Hemminger (TCPOPT_SACK << 8) | 50633ad798cSAdam Langley (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * 50740efc6faSStephen Hemminger TCPOLEN_SACK_PERBLOCK))); 5082de979bdSStephen Hemminger 50933ad798cSAdam Langley for (this_sack = 0; this_sack < opts->num_sack_blocks; 51033ad798cSAdam Langley ++this_sack) { 51140efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].start_seq); 51240efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].end_seq); 51340efc6faSStephen Hemminger } 5142de979bdSStephen Hemminger 51540efc6faSStephen Hemminger tp->rx_opt.dsack = 0; 51640efc6faSStephen Hemminger } 5172100c8d2SYuchung Cheng 5182100c8d2SYuchung Cheng if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) { 5192100c8d2SYuchung Cheng struct tcp_fastopen_cookie *foc = opts->fastopen_cookie; 5207f9b838bSDaniel Lee u8 *p = (u8 *)ptr; 5217f9b838bSDaniel Lee u32 len; /* Fast Open option length */ 5222100c8d2SYuchung Cheng 5237f9b838bSDaniel Lee if (foc->exp) { 5247f9b838bSDaniel Lee len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len; 5257f9b838bSDaniel Lee *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) | 5262100c8d2SYuchung Cheng TCPOPT_FASTOPEN_MAGIC); 5277f9b838bSDaniel Lee p += TCPOLEN_EXP_FASTOPEN_BASE; 5287f9b838bSDaniel Lee } else { 5297f9b838bSDaniel Lee len = TCPOLEN_FASTOPEN_BASE + foc->len; 5307f9b838bSDaniel Lee *p++ = TCPOPT_FASTOPEN; 5317f9b838bSDaniel Lee *p++ = len; 5322100c8d2SYuchung Cheng } 5337f9b838bSDaniel Lee 5347f9b838bSDaniel Lee memcpy(p, foc->val, foc->len); 5357f9b838bSDaniel Lee if ((len & 3) == 2) { 5367f9b838bSDaniel Lee p[foc->len] = TCPOPT_NOP; 5377f9b838bSDaniel Lee p[foc->len + 1] = TCPOPT_NOP; 5387f9b838bSDaniel Lee } 5397f9b838bSDaniel Lee ptr += (len + 3) >> 2; 5402100c8d2SYuchung Cheng } 54160e2a778SUrsula Braun 54260e2a778SUrsula Braun smc_options_write(ptr, &options); 54360e2a778SUrsula Braun } 54460e2a778SUrsula Braun 54560e2a778SUrsula Braun static void smc_set_option(const struct tcp_sock *tp, 54660e2a778SUrsula Braun struct tcp_out_options *opts, 54760e2a778SUrsula Braun unsigned int *remaining) 54860e2a778SUrsula Braun { 54960e2a778SUrsula Braun #if IS_ENABLED(CONFIG_SMC) 55060e2a778SUrsula Braun if (static_branch_unlikely(&tcp_have_smc)) { 55160e2a778SUrsula Braun if (tp->syn_smc) { 55260e2a778SUrsula Braun if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) { 55360e2a778SUrsula Braun opts->options |= OPTION_SMC; 55460e2a778SUrsula Braun *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED; 55560e2a778SUrsula Braun } 55660e2a778SUrsula Braun } 55760e2a778SUrsula Braun } 55860e2a778SUrsula Braun #endif 55960e2a778SUrsula Braun } 56060e2a778SUrsula Braun 56160e2a778SUrsula Braun static void smc_set_option_cond(const struct tcp_sock *tp, 56260e2a778SUrsula Braun const struct inet_request_sock *ireq, 56360e2a778SUrsula Braun struct tcp_out_options *opts, 56460e2a778SUrsula Braun unsigned int *remaining) 56560e2a778SUrsula Braun { 56660e2a778SUrsula Braun #if IS_ENABLED(CONFIG_SMC) 56760e2a778SUrsula Braun if (static_branch_unlikely(&tcp_have_smc)) { 56860e2a778SUrsula Braun if (tp->syn_smc && ireq->smc_ok) { 56960e2a778SUrsula Braun if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) { 57060e2a778SUrsula Braun opts->options |= OPTION_SMC; 57160e2a778SUrsula Braun *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED; 57260e2a778SUrsula Braun } 57360e2a778SUrsula Braun } 57460e2a778SUrsula Braun } 57560e2a778SUrsula Braun #endif 57640efc6faSStephen Hemminger } 57740efc6faSStephen Hemminger 57867edfef7SAndi Kleen /* Compute TCP options for SYN packets. This is not the final 57967edfef7SAndi Kleen * network wire format yet. 58067edfef7SAndi Kleen */ 58195c96174SEric Dumazet static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb, 58233ad798cSAdam Langley struct tcp_out_options *opts, 583cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 584cf533ea5SEric Dumazet { 58533ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 58695c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 587783237e8SYuchung Cheng struct tcp_fastopen_request *fastopen = tp->fastopen_req; 58833ad798cSAdam Langley 589cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 59033ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 59133ad798cSAdam Langley if (*md5) { 59233ad798cSAdam Langley opts->options |= OPTION_MD5; 593bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 594cfb6eeb4SYOSHIFUJI Hideaki } 59533ad798cSAdam Langley #else 59633ad798cSAdam Langley *md5 = NULL; 597cfb6eeb4SYOSHIFUJI Hideaki #endif 59833ad798cSAdam Langley 59933ad798cSAdam Langley /* We always get an MSS option. The option bytes which will be seen in 60033ad798cSAdam Langley * normal data packets should timestamps be used, must be in the MSS 60133ad798cSAdam Langley * advertised. But we subtract them from tp->mss_cache so that 60233ad798cSAdam Langley * calculations in tcp_sendmsg are simpler etc. So account for this 60333ad798cSAdam Langley * fact here if necessary. If we don't do this correctly, as a 60433ad798cSAdam Langley * receiver we won't recognize data packets as being full sized when we 60533ad798cSAdam Langley * should, and thus we won't abide by the delayed ACK rules correctly. 60633ad798cSAdam Langley * SACKs don't matter, we never delay an ACK when we have any of those 60733ad798cSAdam Langley * going out. */ 60833ad798cSAdam Langley opts->mss = tcp_advertise_mss(sk); 609bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 61033ad798cSAdam Langley 6115d2ed052SEric Dumazet if (likely(sock_net(sk)->ipv4.sysctl_tcp_timestamps && !*md5)) { 61233ad798cSAdam Langley opts->options |= OPTION_TS; 6137faee5c0SEric Dumazet opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset; 61433ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 615bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 61633ad798cSAdam Langley } 6179bb37ef0SEric Dumazet if (likely(sock_net(sk)->ipv4.sysctl_tcp_window_scaling)) { 61833ad798cSAdam Langley opts->ws = tp->rx_opt.rcv_wscale; 61989e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 620bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 62133ad798cSAdam Langley } 622f9301034SEric Dumazet if (likely(sock_net(sk)->ipv4.sysctl_tcp_sack)) { 62333ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 624b32d1310SDavid S. Miller if (unlikely(!(OPTION_TS & opts->options))) 625bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 62633ad798cSAdam Langley } 62733ad798cSAdam Langley 628783237e8SYuchung Cheng if (fastopen && fastopen->cookie.len >= 0) { 6292646c831SDaniel Lee u32 need = fastopen->cookie.len; 6302646c831SDaniel Lee 6312646c831SDaniel Lee need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE : 6322646c831SDaniel Lee TCPOLEN_FASTOPEN_BASE; 633783237e8SYuchung Cheng need = (need + 3) & ~3U; /* Align to 32 bits */ 634783237e8SYuchung Cheng if (remaining >= need) { 635783237e8SYuchung Cheng opts->options |= OPTION_FAST_OPEN_COOKIE; 636783237e8SYuchung Cheng opts->fastopen_cookie = &fastopen->cookie; 637783237e8SYuchung Cheng remaining -= need; 638783237e8SYuchung Cheng tp->syn_fastopen = 1; 6392646c831SDaniel Lee tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0; 640783237e8SYuchung Cheng } 641783237e8SYuchung Cheng } 642bd0388aeSWilliam Allen Simpson 64360e2a778SUrsula Braun smc_set_option(tp, opts, &remaining); 64460e2a778SUrsula Braun 645bd0388aeSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 64633ad798cSAdam Langley } 64733ad798cSAdam Langley 64867edfef7SAndi Kleen /* Set up TCP options for SYN-ACKs. */ 64960e2a778SUrsula Braun static unsigned int tcp_synack_options(const struct sock *sk, 65060e2a778SUrsula Braun struct request_sock *req, 65195c96174SEric Dumazet unsigned int mss, struct sk_buff *skb, 65233ad798cSAdam Langley struct tcp_out_options *opts, 65380f03e27SEric Dumazet const struct tcp_md5sig_key *md5, 6548336886fSJerry Chu struct tcp_fastopen_cookie *foc) 6554957faadSWilliam Allen Simpson { 65633ad798cSAdam Langley struct inet_request_sock *ireq = inet_rsk(req); 65795c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 65833ad798cSAdam Langley 65933ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 66080f03e27SEric Dumazet if (md5) { 66133ad798cSAdam Langley opts->options |= OPTION_MD5; 6624957faadSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 6634957faadSWilliam Allen Simpson 6644957faadSWilliam Allen Simpson /* We can't fit any SACK blocks in a packet with MD5 + TS 6654957faadSWilliam Allen Simpson * options. There was discussion about disabling SACK 6664957faadSWilliam Allen Simpson * rather than TS in order to fit in better with old, 6674957faadSWilliam Allen Simpson * buggy kernels, but that was deemed to be unnecessary. 6684957faadSWilliam Allen Simpson */ 669de213e5eSEric Dumazet ireq->tstamp_ok &= !ireq->sack_ok; 67033ad798cSAdam Langley } 67133ad798cSAdam Langley #endif 67233ad798cSAdam Langley 6734957faadSWilliam Allen Simpson /* We always send an MSS option. */ 67433ad798cSAdam Langley opts->mss = mss; 6754957faadSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 67633ad798cSAdam Langley 67733ad798cSAdam Langley if (likely(ireq->wscale_ok)) { 67833ad798cSAdam Langley opts->ws = ireq->rcv_wscale; 67989e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 6804957faadSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 68133ad798cSAdam Langley } 682de213e5eSEric Dumazet if (likely(ireq->tstamp_ok)) { 68333ad798cSAdam Langley opts->options |= OPTION_TS; 68495a22caeSFlorian Westphal opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off; 68533ad798cSAdam Langley opts->tsecr = req->ts_recent; 6864957faadSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 68733ad798cSAdam Langley } 68833ad798cSAdam Langley if (likely(ireq->sack_ok)) { 68933ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 690de213e5eSEric Dumazet if (unlikely(!ireq->tstamp_ok)) 6914957faadSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 69233ad798cSAdam Langley } 6937f9b838bSDaniel Lee if (foc != NULL && foc->len >= 0) { 6947f9b838bSDaniel Lee u32 need = foc->len; 6957f9b838bSDaniel Lee 6967f9b838bSDaniel Lee need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE : 6977f9b838bSDaniel Lee TCPOLEN_FASTOPEN_BASE; 6988336886fSJerry Chu need = (need + 3) & ~3U; /* Align to 32 bits */ 6998336886fSJerry Chu if (remaining >= need) { 7008336886fSJerry Chu opts->options |= OPTION_FAST_OPEN_COOKIE; 7018336886fSJerry Chu opts->fastopen_cookie = foc; 7028336886fSJerry Chu remaining -= need; 7038336886fSJerry Chu } 7048336886fSJerry Chu } 7054957faadSWilliam Allen Simpson 70660e2a778SUrsula Braun smc_set_option_cond(tcp_sk(sk), ireq, opts, &remaining); 70760e2a778SUrsula Braun 7084957faadSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 70933ad798cSAdam Langley } 71033ad798cSAdam Langley 71167edfef7SAndi Kleen /* Compute TCP options for ESTABLISHED sockets. This is not the 71267edfef7SAndi Kleen * final wire format yet. 71367edfef7SAndi Kleen */ 71495c96174SEric Dumazet static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb, 71533ad798cSAdam Langley struct tcp_out_options *opts, 716cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 717cf533ea5SEric Dumazet { 71833ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 71995c96174SEric Dumazet unsigned int size = 0; 720cabeccbdSIlpo Järvinen unsigned int eff_sacks; 72133ad798cSAdam Langley 7225843ef42SAndi Kleen opts->options = 0; 7235843ef42SAndi Kleen 72433ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 72533ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 72633ad798cSAdam Langley if (unlikely(*md5)) { 72733ad798cSAdam Langley opts->options |= OPTION_MD5; 72833ad798cSAdam Langley size += TCPOLEN_MD5SIG_ALIGNED; 72933ad798cSAdam Langley } 73033ad798cSAdam Langley #else 73133ad798cSAdam Langley *md5 = NULL; 73233ad798cSAdam Langley #endif 73333ad798cSAdam Langley 73433ad798cSAdam Langley if (likely(tp->rx_opt.tstamp_ok)) { 73533ad798cSAdam Langley opts->options |= OPTION_TS; 7367faee5c0SEric Dumazet opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0; 73733ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 73833ad798cSAdam Langley size += TCPOLEN_TSTAMP_ALIGNED; 73933ad798cSAdam Langley } 74033ad798cSAdam Langley 741cabeccbdSIlpo Järvinen eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; 742cabeccbdSIlpo Järvinen if (unlikely(eff_sacks)) { 74395c96174SEric Dumazet const unsigned int remaining = MAX_TCP_OPTION_SPACE - size; 74433ad798cSAdam Langley opts->num_sack_blocks = 74595c96174SEric Dumazet min_t(unsigned int, eff_sacks, 74633ad798cSAdam Langley (remaining - TCPOLEN_SACK_BASE_ALIGNED) / 74733ad798cSAdam Langley TCPOLEN_SACK_PERBLOCK); 74833ad798cSAdam Langley size += TCPOLEN_SACK_BASE_ALIGNED + 74933ad798cSAdam Langley opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; 75033ad798cSAdam Langley } 75133ad798cSAdam Langley 75233ad798cSAdam Langley return size; 75340efc6faSStephen Hemminger } 7541da177e4SLinus Torvalds 75546d3ceabSEric Dumazet 75646d3ceabSEric Dumazet /* TCP SMALL QUEUES (TSQ) 75746d3ceabSEric Dumazet * 75846d3ceabSEric Dumazet * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev) 75946d3ceabSEric Dumazet * to reduce RTT and bufferbloat. 76046d3ceabSEric Dumazet * We do this using a special skb destructor (tcp_wfree). 76146d3ceabSEric Dumazet * 76246d3ceabSEric Dumazet * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb 76346d3ceabSEric Dumazet * needs to be reallocated in a driver. 7648e3bff96Sstephen hemminger * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc 76546d3ceabSEric Dumazet * 76646d3ceabSEric Dumazet * Since transmit from skb destructor is forbidden, we use a tasklet 76746d3ceabSEric Dumazet * to process all sockets that eventually need to send more skbs. 76846d3ceabSEric Dumazet * We use one tasklet per cpu, with its own queue of sockets. 76946d3ceabSEric Dumazet */ 77046d3ceabSEric Dumazet struct tsq_tasklet { 77146d3ceabSEric Dumazet struct tasklet_struct tasklet; 77246d3ceabSEric Dumazet struct list_head head; /* queue of tcp sockets */ 77346d3ceabSEric Dumazet }; 77446d3ceabSEric Dumazet static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet); 77546d3ceabSEric Dumazet 7766f458dfbSEric Dumazet static void tcp_tsq_handler(struct sock *sk) 7776f458dfbSEric Dumazet { 7786f458dfbSEric Dumazet if ((1 << sk->sk_state) & 7796f458dfbSEric Dumazet (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING | 780f9616c35SEric Dumazet TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) { 781f9616c35SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 782f9616c35SEric Dumazet 783f9616c35SEric Dumazet if (tp->lost_out > tp->retrans_out && 7843a91d29fSKoichiro Den tp->snd_cwnd > tcp_packets_in_flight(tp)) { 7853a91d29fSKoichiro Den tcp_mstamp_refresh(tp); 786f9616c35SEric Dumazet tcp_xmit_retransmit_queue(sk); 7873a91d29fSKoichiro Den } 788f9616c35SEric Dumazet 789f9616c35SEric Dumazet tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle, 790bf06200eSJohn Ogness 0, GFP_ATOMIC); 7916f458dfbSEric Dumazet } 792f9616c35SEric Dumazet } 79346d3ceabSEric Dumazet /* 7948e3bff96Sstephen hemminger * One tasklet per cpu tries to send more skbs. 79546d3ceabSEric Dumazet * We run in tasklet context but need to disable irqs when 7968e3bff96Sstephen hemminger * transferring tsq->head because tcp_wfree() might 79746d3ceabSEric Dumazet * interrupt us (non NAPI drivers) 79846d3ceabSEric Dumazet */ 79946d3ceabSEric Dumazet static void tcp_tasklet_func(unsigned long data) 80046d3ceabSEric Dumazet { 80146d3ceabSEric Dumazet struct tsq_tasklet *tsq = (struct tsq_tasklet *)data; 80246d3ceabSEric Dumazet LIST_HEAD(list); 80346d3ceabSEric Dumazet unsigned long flags; 80446d3ceabSEric Dumazet struct list_head *q, *n; 80546d3ceabSEric Dumazet struct tcp_sock *tp; 80646d3ceabSEric Dumazet struct sock *sk; 80746d3ceabSEric Dumazet 80846d3ceabSEric Dumazet local_irq_save(flags); 80946d3ceabSEric Dumazet list_splice_init(&tsq->head, &list); 81046d3ceabSEric Dumazet local_irq_restore(flags); 81146d3ceabSEric Dumazet 81246d3ceabSEric Dumazet list_for_each_safe(q, n, &list) { 81346d3ceabSEric Dumazet tp = list_entry(q, struct tcp_sock, tsq_node); 81446d3ceabSEric Dumazet list_del(&tp->tsq_node); 81546d3ceabSEric Dumazet 81646d3ceabSEric Dumazet sk = (struct sock *)tp; 8170a9648f1SEric Dumazet smp_mb__before_atomic(); 8187aa5470cSEric Dumazet clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags); 8197aa5470cSEric Dumazet 820b223feb9SEric Dumazet if (!sk->sk_lock.owned && 8217aa5470cSEric Dumazet test_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags)) { 82246d3ceabSEric Dumazet bh_lock_sock(sk); 82346d3ceabSEric Dumazet if (!sock_owned_by_user(sk)) { 8247aa5470cSEric Dumazet clear_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags); 8256f458dfbSEric Dumazet tcp_tsq_handler(sk); 82646d3ceabSEric Dumazet } 82746d3ceabSEric Dumazet bh_unlock_sock(sk); 828b223feb9SEric Dumazet } 82946d3ceabSEric Dumazet 83046d3ceabSEric Dumazet sk_free(sk); 83146d3ceabSEric Dumazet } 83246d3ceabSEric Dumazet } 83346d3ceabSEric Dumazet 83440fc3423SEric Dumazet #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \ 83540fc3423SEric Dumazet TCPF_WRITE_TIMER_DEFERRED | \ 83640fc3423SEric Dumazet TCPF_DELACK_TIMER_DEFERRED | \ 83740fc3423SEric Dumazet TCPF_MTU_REDUCED_DEFERRED) 83846d3ceabSEric Dumazet /** 83946d3ceabSEric Dumazet * tcp_release_cb - tcp release_sock() callback 84046d3ceabSEric Dumazet * @sk: socket 84146d3ceabSEric Dumazet * 84246d3ceabSEric Dumazet * called from release_sock() to perform protocol dependent 84346d3ceabSEric Dumazet * actions before socket release. 84446d3ceabSEric Dumazet */ 84546d3ceabSEric Dumazet void tcp_release_cb(struct sock *sk) 84646d3ceabSEric Dumazet { 8476f458dfbSEric Dumazet unsigned long flags, nflags; 84846d3ceabSEric Dumazet 8496f458dfbSEric Dumazet /* perform an atomic operation only if at least one flag is set */ 8506f458dfbSEric Dumazet do { 8517aa5470cSEric Dumazet flags = sk->sk_tsq_flags; 8526f458dfbSEric Dumazet if (!(flags & TCP_DEFERRED_ALL)) 8536f458dfbSEric Dumazet return; 8546f458dfbSEric Dumazet nflags = flags & ~TCP_DEFERRED_ALL; 8557aa5470cSEric Dumazet } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags); 8566f458dfbSEric Dumazet 85740fc3423SEric Dumazet if (flags & TCPF_TSQ_DEFERRED) 8586f458dfbSEric Dumazet tcp_tsq_handler(sk); 8596f458dfbSEric Dumazet 860c3f9b018SEric Dumazet /* Here begins the tricky part : 861c3f9b018SEric Dumazet * We are called from release_sock() with : 862c3f9b018SEric Dumazet * 1) BH disabled 863c3f9b018SEric Dumazet * 2) sk_lock.slock spinlock held 864c3f9b018SEric Dumazet * 3) socket owned by us (sk->sk_lock.owned == 1) 865c3f9b018SEric Dumazet * 866c3f9b018SEric Dumazet * But following code is meant to be called from BH handlers, 867c3f9b018SEric Dumazet * so we should keep BH disabled, but early release socket ownership 868c3f9b018SEric Dumazet */ 869c3f9b018SEric Dumazet sock_release_ownership(sk); 870c3f9b018SEric Dumazet 87140fc3423SEric Dumazet if (flags & TCPF_WRITE_TIMER_DEFERRED) { 8726f458dfbSEric Dumazet tcp_write_timer_handler(sk); 873144d56e9SEric Dumazet __sock_put(sk); 874144d56e9SEric Dumazet } 87540fc3423SEric Dumazet if (flags & TCPF_DELACK_TIMER_DEFERRED) { 8766f458dfbSEric Dumazet tcp_delack_timer_handler(sk); 877144d56e9SEric Dumazet __sock_put(sk); 878144d56e9SEric Dumazet } 87940fc3423SEric Dumazet if (flags & TCPF_MTU_REDUCED_DEFERRED) { 8804fab9071SNeal Cardwell inet_csk(sk)->icsk_af_ops->mtu_reduced(sk); 881144d56e9SEric Dumazet __sock_put(sk); 882144d56e9SEric Dumazet } 88346d3ceabSEric Dumazet } 88446d3ceabSEric Dumazet EXPORT_SYMBOL(tcp_release_cb); 88546d3ceabSEric Dumazet 88646d3ceabSEric Dumazet void __init tcp_tasklet_init(void) 88746d3ceabSEric Dumazet { 88846d3ceabSEric Dumazet int i; 88946d3ceabSEric Dumazet 89046d3ceabSEric Dumazet for_each_possible_cpu(i) { 89146d3ceabSEric Dumazet struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i); 89246d3ceabSEric Dumazet 89346d3ceabSEric Dumazet INIT_LIST_HEAD(&tsq->head); 89446d3ceabSEric Dumazet tasklet_init(&tsq->tasklet, 89546d3ceabSEric Dumazet tcp_tasklet_func, 89646d3ceabSEric Dumazet (unsigned long)tsq); 89746d3ceabSEric Dumazet } 89846d3ceabSEric Dumazet } 89946d3ceabSEric Dumazet 90046d3ceabSEric Dumazet /* 90146d3ceabSEric Dumazet * Write buffer destructor automatically called from kfree_skb. 9028e3bff96Sstephen hemminger * We can't xmit new skbs from this context, as we might already 90346d3ceabSEric Dumazet * hold qdisc lock. 90446d3ceabSEric Dumazet */ 905d6a4a104SEric Dumazet void tcp_wfree(struct sk_buff *skb) 90646d3ceabSEric Dumazet { 90746d3ceabSEric Dumazet struct sock *sk = skb->sk; 90846d3ceabSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 909408f0a6cSEric Dumazet unsigned long flags, nval, oval; 9109b462d02SEric Dumazet 9119b462d02SEric Dumazet /* Keep one reference on sk_wmem_alloc. 9129b462d02SEric Dumazet * Will be released by sk_free() from here or tcp_tasklet_func() 9139b462d02SEric Dumazet */ 91414afee4bSReshetova, Elena WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc)); 9159b462d02SEric Dumazet 9169b462d02SEric Dumazet /* If this softirq is serviced by ksoftirqd, we are likely under stress. 9179b462d02SEric Dumazet * Wait until our queues (qdisc + devices) are drained. 9189b462d02SEric Dumazet * This gives : 9199b462d02SEric Dumazet * - less callbacks to tcp_write_xmit(), reducing stress (batches) 9209b462d02SEric Dumazet * - chance for incoming ACK (processed by another cpu maybe) 9219b462d02SEric Dumazet * to migrate this flow (skb->ooo_okay will be eventually set) 9229b462d02SEric Dumazet */ 92314afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current) 9249b462d02SEric Dumazet goto out; 92546d3ceabSEric Dumazet 9267aa5470cSEric Dumazet for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) { 92746d3ceabSEric Dumazet struct tsq_tasklet *tsq; 928a9b204d1SEric Dumazet bool empty; 92946d3ceabSEric Dumazet 930408f0a6cSEric Dumazet if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED)) 931408f0a6cSEric Dumazet goto out; 932408f0a6cSEric Dumazet 933b223feb9SEric Dumazet nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED | TCPF_TSQ_DEFERRED; 9347aa5470cSEric Dumazet nval = cmpxchg(&sk->sk_tsq_flags, oval, nval); 935408f0a6cSEric Dumazet if (nval != oval) 936408f0a6cSEric Dumazet continue; 937408f0a6cSEric Dumazet 93846d3ceabSEric Dumazet /* queue this socket to tasklet queue */ 93946d3ceabSEric Dumazet local_irq_save(flags); 940903ceff7SChristoph Lameter tsq = this_cpu_ptr(&tsq_tasklet); 941a9b204d1SEric Dumazet empty = list_empty(&tsq->head); 94246d3ceabSEric Dumazet list_add(&tp->tsq_node, &tsq->head); 943a9b204d1SEric Dumazet if (empty) 94446d3ceabSEric Dumazet tasklet_schedule(&tsq->tasklet); 94546d3ceabSEric Dumazet local_irq_restore(flags); 9469b462d02SEric Dumazet return; 94746d3ceabSEric Dumazet } 9489b462d02SEric Dumazet out: 9499b462d02SEric Dumazet sk_free(sk); 95046d3ceabSEric Dumazet } 95146d3ceabSEric Dumazet 952218af599SEric Dumazet /* Note: Called under hard irq. 953218af599SEric Dumazet * We can not call TCP stack right away. 954218af599SEric Dumazet */ 955218af599SEric Dumazet enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer) 956218af599SEric Dumazet { 957218af599SEric Dumazet struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer); 958218af599SEric Dumazet struct sock *sk = (struct sock *)tp; 959218af599SEric Dumazet unsigned long nval, oval; 960218af599SEric Dumazet 961218af599SEric Dumazet for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) { 962218af599SEric Dumazet struct tsq_tasklet *tsq; 963218af599SEric Dumazet bool empty; 964218af599SEric Dumazet 965218af599SEric Dumazet if (oval & TSQF_QUEUED) 966218af599SEric Dumazet break; 967218af599SEric Dumazet 968218af599SEric Dumazet nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED | TCPF_TSQ_DEFERRED; 969218af599SEric Dumazet nval = cmpxchg(&sk->sk_tsq_flags, oval, nval); 970218af599SEric Dumazet if (nval != oval) 971218af599SEric Dumazet continue; 972218af599SEric Dumazet 97314afee4bSReshetova, Elena if (!refcount_inc_not_zero(&sk->sk_wmem_alloc)) 974218af599SEric Dumazet break; 975218af599SEric Dumazet /* queue this socket to tasklet queue */ 976218af599SEric Dumazet tsq = this_cpu_ptr(&tsq_tasklet); 977218af599SEric Dumazet empty = list_empty(&tsq->head); 978218af599SEric Dumazet list_add(&tp->tsq_node, &tsq->head); 979218af599SEric Dumazet if (empty) 980218af599SEric Dumazet tasklet_schedule(&tsq->tasklet); 981218af599SEric Dumazet break; 982218af599SEric Dumazet } 983218af599SEric Dumazet return HRTIMER_NORESTART; 984218af599SEric Dumazet } 985218af599SEric Dumazet 986218af599SEric Dumazet /* BBR congestion control needs pacing. 987218af599SEric Dumazet * Same remark for SO_MAX_PACING_RATE. 988218af599SEric Dumazet * sch_fq packet scheduler is efficiently handling pacing, 989218af599SEric Dumazet * but is not always installed/used. 990218af599SEric Dumazet * Return true if TCP stack should pace packets itself. 991218af599SEric Dumazet */ 992218af599SEric Dumazet static bool tcp_needs_internal_pacing(const struct sock *sk) 993218af599SEric Dumazet { 994218af599SEric Dumazet return smp_load_acquire(&sk->sk_pacing_status) == SK_PACING_NEEDED; 995218af599SEric Dumazet } 996218af599SEric Dumazet 997218af599SEric Dumazet static void tcp_internal_pacing(struct sock *sk, const struct sk_buff *skb) 998218af599SEric Dumazet { 999218af599SEric Dumazet u64 len_ns; 1000218af599SEric Dumazet u32 rate; 1001218af599SEric Dumazet 1002218af599SEric Dumazet if (!tcp_needs_internal_pacing(sk)) 1003218af599SEric Dumazet return; 1004218af599SEric Dumazet rate = sk->sk_pacing_rate; 1005218af599SEric Dumazet if (!rate || rate == ~0U) 1006218af599SEric Dumazet return; 1007218af599SEric Dumazet 1008218af599SEric Dumazet /* Should account for header sizes as sch_fq does, 1009218af599SEric Dumazet * but lets make things simple. 1010218af599SEric Dumazet */ 1011218af599SEric Dumazet len_ns = (u64)skb->len * NSEC_PER_SEC; 1012218af599SEric Dumazet do_div(len_ns, rate); 1013218af599SEric Dumazet hrtimer_start(&tcp_sk(sk)->pacing_timer, 1014218af599SEric Dumazet ktime_add_ns(ktime_get(), len_ns), 1015218af599SEric Dumazet HRTIMER_MODE_ABS_PINNED); 1016218af599SEric Dumazet } 1017218af599SEric Dumazet 1018e2080072SEric Dumazet static void tcp_update_skb_after_send(struct tcp_sock *tp, struct sk_buff *skb) 1019e2080072SEric Dumazet { 1020e2080072SEric Dumazet skb->skb_mstamp = tp->tcp_mstamp; 1021e2080072SEric Dumazet list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue); 1022e2080072SEric Dumazet } 1023e2080072SEric Dumazet 10241da177e4SLinus Torvalds /* This routine actually transmits TCP packets queued in by 10251da177e4SLinus Torvalds * tcp_do_sendmsg(). This is used by both the initial 10261da177e4SLinus Torvalds * transmission and possible later retransmissions. 10271da177e4SLinus Torvalds * All SKB's seen here are completely headerless. It is our 10281da177e4SLinus Torvalds * job to build the TCP header, and pass the packet down to 10291da177e4SLinus Torvalds * IP so it can do the same plus pass the packet off to the 10301da177e4SLinus Torvalds * device. 10311da177e4SLinus Torvalds * 10321da177e4SLinus Torvalds * We are working here with either a clone of the original 10331da177e4SLinus Torvalds * SKB, or a fresh unique copy made by the retransmit engine. 10341da177e4SLinus Torvalds */ 1035056834d9SIlpo Järvinen static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, 1036056834d9SIlpo Järvinen gfp_t gfp_mask) 10371da177e4SLinus Torvalds { 10386687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 1039dfb4b9dcSDavid S. Miller struct inet_sock *inet; 1040dfb4b9dcSDavid S. Miller struct tcp_sock *tp; 1041dfb4b9dcSDavid S. Miller struct tcp_skb_cb *tcb; 104233ad798cSAdam Langley struct tcp_out_options opts; 104395c96174SEric Dumazet unsigned int tcp_options_size, tcp_header_size; 10448c72c65bSEric Dumazet struct sk_buff *oskb = NULL; 1045cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 10461da177e4SLinus Torvalds struct tcphdr *th; 10471da177e4SLinus Torvalds int err; 10481da177e4SLinus Torvalds 1049dfb4b9dcSDavid S. Miller BUG_ON(!skb || !tcp_skb_pcount(skb)); 10506f094b9eSLawrence Brakmo tp = tcp_sk(sk); 1051dfb4b9dcSDavid S. Miller 1052ccdbb6e9SEric Dumazet if (clone_it) { 10536f094b9eSLawrence Brakmo TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq 10546f094b9eSLawrence Brakmo - tp->snd_una; 10558c72c65bSEric Dumazet oskb = skb; 1056e2080072SEric Dumazet 1057e2080072SEric Dumazet tcp_skb_tsorted_save(oskb) { 1058e2080072SEric Dumazet if (unlikely(skb_cloned(oskb))) 1059e2080072SEric Dumazet skb = pskb_copy(oskb, gfp_mask); 1060dfb4b9dcSDavid S. Miller else 1061e2080072SEric Dumazet skb = skb_clone(oskb, gfp_mask); 1062e2080072SEric Dumazet } tcp_skb_tsorted_restore(oskb); 1063e2080072SEric Dumazet 1064dfb4b9dcSDavid S. Miller if (unlikely(!skb)) 1065dfb4b9dcSDavid S. Miller return -ENOBUFS; 1066dfb4b9dcSDavid S. Miller } 10678c72c65bSEric Dumazet skb->skb_mstamp = tp->tcp_mstamp; 1068dfb4b9dcSDavid S. Miller 1069dfb4b9dcSDavid S. Miller inet = inet_sk(sk); 1070dfb4b9dcSDavid S. Miller tcb = TCP_SKB_CB(skb); 107133ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 10721da177e4SLinus Torvalds 10734de075e0SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) 107433ad798cSAdam Langley tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); 107533ad798cSAdam Langley else 107633ad798cSAdam Langley tcp_options_size = tcp_established_options(sk, skb, &opts, 107733ad798cSAdam Langley &md5); 107833ad798cSAdam Langley tcp_header_size = tcp_options_size + sizeof(struct tcphdr); 10791da177e4SLinus Torvalds 1080547669d4SEric Dumazet /* if no packet is in qdisc/device queue, then allow XPS to select 1081b2532eb9SEric Dumazet * another queue. We can be called from tcp_tsq_handler() 1082b2532eb9SEric Dumazet * which holds one reference to sk_wmem_alloc. 1083b2532eb9SEric Dumazet * 1084b2532eb9SEric Dumazet * TODO: Ideally, in-flight pure ACK packets should not matter here. 1085b2532eb9SEric Dumazet * One way to get this would be to set skb->truesize = 2 on them. 1086547669d4SEric Dumazet */ 1087b2532eb9SEric Dumazet skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1); 10881da177e4SLinus Torvalds 108938ab52e8SEric Dumazet /* If we had to use memory reserve to allocate this skb, 109038ab52e8SEric Dumazet * this might cause drops if packet is looped back : 109138ab52e8SEric Dumazet * Other socket might not have SOCK_MEMALLOC. 109238ab52e8SEric Dumazet * Packets not looped back do not care about pfmemalloc. 109338ab52e8SEric Dumazet */ 109438ab52e8SEric Dumazet skb->pfmemalloc = 0; 109538ab52e8SEric Dumazet 1096aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 1097aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 109846d3ceabSEric Dumazet 109946d3ceabSEric Dumazet skb_orphan(skb); 110046d3ceabSEric Dumazet skb->sk = sk; 11011d2077acSEric Dumazet skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree; 1102b73c3d0eSTom Herbert skb_set_hash_from_sk(skb, sk); 110314afee4bSReshetova, Elena refcount_add(skb->truesize, &sk->sk_wmem_alloc); 11041da177e4SLinus Torvalds 1105c3a2e837SJulian Anastasov skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm); 1106c3a2e837SJulian Anastasov 11071da177e4SLinus Torvalds /* Build TCP header and checksum it. */ 1108ea1627c2SEric Dumazet th = (struct tcphdr *)skb->data; 1109c720c7e8SEric Dumazet th->source = inet->inet_sport; 1110c720c7e8SEric Dumazet th->dest = inet->inet_dport; 11111da177e4SLinus Torvalds th->seq = htonl(tcb->seq); 11121da177e4SLinus Torvalds th->ack_seq = htonl(tp->rcv_nxt); 1113df7a3b07SAl Viro *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | 11144de075e0SEric Dumazet tcb->tcp_flags); 1115dfb4b9dcSDavid S. Miller 11161da177e4SLinus Torvalds th->check = 0; 11171da177e4SLinus Torvalds th->urg_ptr = 0; 11181da177e4SLinus Torvalds 111933f5f57eSIlpo Järvinen /* The urg_mode check is necessary during a below snd_una win probe */ 11207691367dSHerbert Xu if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { 11217691367dSHerbert Xu if (before(tp->snd_up, tcb->seq + 0x10000)) { 11221da177e4SLinus Torvalds th->urg_ptr = htons(tp->snd_up - tcb->seq); 11231da177e4SLinus Torvalds th->urg = 1; 11247691367dSHerbert Xu } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { 11250eae88f3SEric Dumazet th->urg_ptr = htons(0xFFFF); 11267691367dSHerbert Xu th->urg = 1; 11277691367dSHerbert Xu } 11281da177e4SLinus Torvalds } 11291da177e4SLinus Torvalds 1130bd0388aeSWilliam Allen Simpson tcp_options_write((__be32 *)(th + 1), tp, &opts); 113151466a75SEric Dumazet skb_shinfo(skb)->gso_type = sk->sk_gso_type; 1132ea1627c2SEric Dumazet if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) { 1133ea1627c2SEric Dumazet th->window = htons(tcp_select_window(sk)); 1134ea1627c2SEric Dumazet tcp_ecn_send(sk, skb, th, tcp_header_size); 1135ea1627c2SEric Dumazet } else { 1136ea1627c2SEric Dumazet /* RFC1323: The window in SYN & SYN/ACK segments 1137ea1627c2SEric Dumazet * is never scaled. 1138ea1627c2SEric Dumazet */ 1139ea1627c2SEric Dumazet th->window = htons(min(tp->rcv_wnd, 65535U)); 1140ea1627c2SEric Dumazet } 1141cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 1142cfb6eeb4SYOSHIFUJI Hideaki /* Calculate the MD5 hash, as we have all we need now */ 1143cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 1144a465419bSEric Dumazet sk_nocaps_add(sk, NETIF_F_GSO_MASK); 1145bd0388aeSWilliam Allen Simpson tp->af_specific->calc_md5_hash(opts.hash_location, 114639f8e58eSEric Dumazet md5, sk, skb); 1147cfb6eeb4SYOSHIFUJI Hideaki } 1148cfb6eeb4SYOSHIFUJI Hideaki #endif 1149cfb6eeb4SYOSHIFUJI Hideaki 1150bb296246SHerbert Xu icsk->icsk_af_ops->send_check(sk, skb); 11511da177e4SLinus Torvalds 11524de075e0SEric Dumazet if (likely(tcb->tcp_flags & TCPHDR_ACK)) 1153fc6415bcSDavid S. Miller tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); 11541da177e4SLinus Torvalds 1155a44d6eacSMartin KaFai Lau if (skb->len != tcp_header_size) { 1156cf533ea5SEric Dumazet tcp_event_data_sent(tp, sk); 1157a44d6eacSMartin KaFai Lau tp->data_segs_out += tcp_skb_pcount(skb); 1158218af599SEric Dumazet tcp_internal_pacing(sk, skb); 1159a44d6eacSMartin KaFai Lau } 11601da177e4SLinus Torvalds 1161bd37a088SWei Yongjun if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) 1162aa2ea058STom Herbert TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, 1163aa2ea058STom Herbert tcp_skb_pcount(skb)); 11641da177e4SLinus Torvalds 11652efd055cSMarcelo Ricardo Leitner tp->segs_out += tcp_skb_pcount(skb); 1166f69ad292SEric Dumazet /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */ 1167cd7d8498SEric Dumazet skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb); 1168f69ad292SEric Dumazet skb_shinfo(skb)->gso_size = tcp_skb_mss(skb); 1169cd7d8498SEric Dumazet 11707faee5c0SEric Dumazet /* Our usage of tstamp should remain private */ 11712456e855SThomas Gleixner skb->tstamp = 0; 1172971f10ecSEric Dumazet 1173971f10ecSEric Dumazet /* Cleanup our debris for IP stacks */ 1174971f10ecSEric Dumazet memset(skb->cb, 0, max(sizeof(struct inet_skb_parm), 1175971f10ecSEric Dumazet sizeof(struct inet6_skb_parm))); 1176971f10ecSEric Dumazet 1177b0270e91SEric Dumazet err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl); 11787faee5c0SEric Dumazet 11798c72c65bSEric Dumazet if (unlikely(err > 0)) { 11805ee2c941SChristoph Paasch tcp_enter_cwr(sk); 11818c72c65bSEric Dumazet err = net_xmit_eval(err); 11828c72c65bSEric Dumazet } 1183fc225799SEric Dumazet if (!err && oskb) { 1184e2080072SEric Dumazet tcp_update_skb_after_send(tp, oskb); 1185fc225799SEric Dumazet tcp_rate_skb_sent(sk, oskb); 1186fc225799SEric Dumazet } 11878c72c65bSEric Dumazet return err; 11881da177e4SLinus Torvalds } 11891da177e4SLinus Torvalds 119067edfef7SAndi Kleen /* This routine just queues the buffer for sending. 11911da177e4SLinus Torvalds * 11921da177e4SLinus Torvalds * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, 11931da177e4SLinus Torvalds * otherwise socket can stall. 11941da177e4SLinus Torvalds */ 11951da177e4SLinus Torvalds static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) 11961da177e4SLinus Torvalds { 11971da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 11981da177e4SLinus Torvalds 11991da177e4SLinus Torvalds /* Advance write_seq and place onto the write_queue. */ 12001da177e4SLinus Torvalds tp->write_seq = TCP_SKB_CB(skb)->end_seq; 1201f4a775d1SEric Dumazet __skb_header_release(skb); 1202fe067e8aSDavid S. Miller tcp_add_write_queue_tail(sk, skb); 12033ab224beSHideo Aoki sk->sk_wmem_queued += skb->truesize; 12043ab224beSHideo Aoki sk_mem_charge(sk, skb->truesize); 12051da177e4SLinus Torvalds } 12061da177e4SLinus Torvalds 120767edfef7SAndi Kleen /* Initialize TSO segments for a packet. */ 12085bbb432cSEric Dumazet static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now) 1209f6302d1dSDavid S. Miller { 12108f26fb1cSEric Dumazet if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) { 1211f6302d1dSDavid S. Miller /* Avoid the costly divide in the normal 1212f6302d1dSDavid S. Miller * non-TSO case. 1213f6302d1dSDavid S. Miller */ 1214cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, 1); 1215f69ad292SEric Dumazet TCP_SKB_CB(skb)->tcp_gso_size = 0; 1216f6302d1dSDavid S. Miller } else { 1217cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now)); 1218f69ad292SEric Dumazet TCP_SKB_CB(skb)->tcp_gso_size = mss_now; 12191da177e4SLinus Torvalds } 12201da177e4SLinus Torvalds } 12211da177e4SLinus Torvalds 122291fed7a1SIlpo Järvinen /* When a modification to fackets out becomes necessary, we need to check 122368f8353bSIlpo Järvinen * skb is counted to fackets_out or not. 122491fed7a1SIlpo Järvinen */ 1225cf533ea5SEric Dumazet static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb, 122691fed7a1SIlpo Järvinen int decr) 122791fed7a1SIlpo Järvinen { 1228a47e5a98SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1229a47e5a98SIlpo Järvinen 1230dc86967bSIlpo Järvinen if (!tp->sacked_out || tcp_is_reno(tp)) 123191fed7a1SIlpo Järvinen return; 123291fed7a1SIlpo Järvinen 12336859d494SIlpo Järvinen if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq)) 123491fed7a1SIlpo Järvinen tp->fackets_out -= decr; 123591fed7a1SIlpo Järvinen } 123691fed7a1SIlpo Järvinen 1237797108d1SIlpo Järvinen /* Pcount in the middle of the write queue got changed, we need to do various 1238797108d1SIlpo Järvinen * tweaks to fix counters 1239797108d1SIlpo Järvinen */ 1240cf533ea5SEric Dumazet static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) 1241797108d1SIlpo Järvinen { 1242797108d1SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1243797108d1SIlpo Järvinen 1244797108d1SIlpo Järvinen tp->packets_out -= decr; 1245797108d1SIlpo Järvinen 1246797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 1247797108d1SIlpo Järvinen tp->sacked_out -= decr; 1248797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 1249797108d1SIlpo Järvinen tp->retrans_out -= decr; 1250797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 1251797108d1SIlpo Järvinen tp->lost_out -= decr; 1252797108d1SIlpo Järvinen 1253797108d1SIlpo Järvinen /* Reno case is special. Sigh... */ 1254797108d1SIlpo Järvinen if (tcp_is_reno(tp) && decr > 0) 1255797108d1SIlpo Järvinen tp->sacked_out -= min_t(u32, tp->sacked_out, decr); 1256797108d1SIlpo Järvinen 1257797108d1SIlpo Järvinen tcp_adjust_fackets_out(sk, skb, decr); 1258797108d1SIlpo Järvinen 1259797108d1SIlpo Järvinen if (tp->lost_skb_hint && 1260797108d1SIlpo Järvinen before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && 126152cf3cc8SIlpo Järvinen (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))) 1262797108d1SIlpo Järvinen tp->lost_cnt_hint -= decr; 1263797108d1SIlpo Järvinen 1264797108d1SIlpo Järvinen tcp_verify_left_out(tp); 1265797108d1SIlpo Järvinen } 1266797108d1SIlpo Järvinen 12670a2cf20cSSoheil Hassas Yeganeh static bool tcp_has_tx_tstamp(const struct sk_buff *skb) 12680a2cf20cSSoheil Hassas Yeganeh { 12690a2cf20cSSoheil Hassas Yeganeh return TCP_SKB_CB(skb)->txstamp_ack || 12700a2cf20cSSoheil Hassas Yeganeh (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP); 12710a2cf20cSSoheil Hassas Yeganeh } 12720a2cf20cSSoheil Hassas Yeganeh 1273490cc7d0SWillem de Bruijn static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2) 1274490cc7d0SWillem de Bruijn { 1275490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo = skb_shinfo(skb); 1276490cc7d0SWillem de Bruijn 12770a2cf20cSSoheil Hassas Yeganeh if (unlikely(tcp_has_tx_tstamp(skb)) && 1278490cc7d0SWillem de Bruijn !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) { 1279490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo2 = skb_shinfo(skb2); 1280490cc7d0SWillem de Bruijn u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP; 1281490cc7d0SWillem de Bruijn 1282490cc7d0SWillem de Bruijn shinfo->tx_flags &= ~tsflags; 1283490cc7d0SWillem de Bruijn shinfo2->tx_flags |= tsflags; 1284490cc7d0SWillem de Bruijn swap(shinfo->tskey, shinfo2->tskey); 1285b51e13faSMartin KaFai Lau TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack; 1286b51e13faSMartin KaFai Lau TCP_SKB_CB(skb)->txstamp_ack = 0; 1287490cc7d0SWillem de Bruijn } 1288490cc7d0SWillem de Bruijn } 1289490cc7d0SWillem de Bruijn 1290a166140eSMartin KaFai Lau static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2) 1291a166140eSMartin KaFai Lau { 1292a166140eSMartin KaFai Lau TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor; 1293a166140eSMartin KaFai Lau TCP_SKB_CB(skb)->eor = 0; 1294a166140eSMartin KaFai Lau } 1295a166140eSMartin KaFai Lau 129675c119afSEric Dumazet /* Insert buff after skb on the write or rtx queue of sk. */ 129775c119afSEric Dumazet static void tcp_insert_write_queue_after(struct sk_buff *skb, 129875c119afSEric Dumazet struct sk_buff *buff, 129975c119afSEric Dumazet struct sock *sk, 130075c119afSEric Dumazet enum tcp_queue tcp_queue) 130175c119afSEric Dumazet { 130275c119afSEric Dumazet if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE) 130375c119afSEric Dumazet __skb_queue_after(&sk->sk_write_queue, skb, buff); 130475c119afSEric Dumazet else 130575c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, buff); 130675c119afSEric Dumazet } 130775c119afSEric Dumazet 13081da177e4SLinus Torvalds /* Function to create two new TCP segments. Shrinks the given segment 13091da177e4SLinus Torvalds * to the specified size and appends a new segment with the rest of the 13101da177e4SLinus Torvalds * packet to the list. This won't be called frequently, I hope. 13111da177e4SLinus Torvalds * Remember, these are still headerless SKBs at this point. 13121da177e4SLinus Torvalds */ 131375c119afSEric Dumazet int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue, 131475c119afSEric Dumazet struct sk_buff *skb, u32 len, 13156cc55e09SOctavian Purdila unsigned int mss_now, gfp_t gfp) 13161da177e4SLinus Torvalds { 13171da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 13181da177e4SLinus Torvalds struct sk_buff *buff; 13196475be16SDavid S. Miller int nsize, old_factor; 1320b60b49eaSHerbert Xu int nlen; 13219ce01461SIlpo Järvinen u8 flags; 13221da177e4SLinus Torvalds 13232fceec13SIlpo Järvinen if (WARN_ON(len > skb->len)) 13242fceec13SIlpo Järvinen return -EINVAL; 13256a438bbeSStephen Hemminger 13261da177e4SLinus Torvalds nsize = skb_headlen(skb) - len; 13271da177e4SLinus Torvalds if (nsize < 0) 13281da177e4SLinus Torvalds nsize = 0; 13291da177e4SLinus Torvalds 13306cc55e09SOctavian Purdila if (skb_unclone(skb, gfp)) 13311da177e4SLinus Torvalds return -ENOMEM; 13321da177e4SLinus Torvalds 13331da177e4SLinus Torvalds /* Get a new skb... force flag on. */ 1334eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, nsize, gfp, true); 133551456b29SIan Morris if (!buff) 13361da177e4SLinus Torvalds return -ENOMEM; /* We'll just try again later. */ 1337ef5cb973SHerbert Xu 13383ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 13393ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1340b60b49eaSHerbert Xu nlen = skb->len - len - nsize; 1341b60b49eaSHerbert Xu buff->truesize += nlen; 1342b60b49eaSHerbert Xu skb->truesize -= nlen; 13431da177e4SLinus Torvalds 13441da177e4SLinus Torvalds /* Correct the sequence numbers. */ 13451da177e4SLinus Torvalds TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 13461da177e4SLinus Torvalds TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 13471da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 13481da177e4SLinus Torvalds 13491da177e4SLinus Torvalds /* PSH and FIN should only be set in the second packet. */ 13504de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 13514de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 13524de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1353e14c3cafSHerbert Xu TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; 1354a166140eSMartin KaFai Lau tcp_skb_fragment_eor(skb, buff); 13551da177e4SLinus Torvalds 135684fa7933SPatrick McHardy if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { 13571da177e4SLinus Torvalds /* Copy and checksum data tail into the new buffer. */ 1358056834d9SIlpo Järvinen buff->csum = csum_partial_copy_nocheck(skb->data + len, 1359056834d9SIlpo Järvinen skb_put(buff, nsize), 13601da177e4SLinus Torvalds nsize, 0); 13611da177e4SLinus Torvalds 13621da177e4SLinus Torvalds skb_trim(skb, len); 13631da177e4SLinus Torvalds 13641da177e4SLinus Torvalds skb->csum = csum_block_sub(skb->csum, buff->csum, len); 13651da177e4SLinus Torvalds } else { 136684fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 13671da177e4SLinus Torvalds skb_split(skb, buff, len); 13681da177e4SLinus Torvalds } 13691da177e4SLinus Torvalds 13701da177e4SLinus Torvalds buff->ip_summed = skb->ip_summed; 13711da177e4SLinus Torvalds 1372a61bbcf2SPatrick McHardy buff->tstamp = skb->tstamp; 1373490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 13741da177e4SLinus Torvalds 13756475be16SDavid S. Miller old_factor = tcp_skb_pcount(skb); 13766475be16SDavid S. Miller 13771da177e4SLinus Torvalds /* Fix up tso_factor for both original and new SKB. */ 13785bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 13795bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 13801da177e4SLinus Torvalds 1381b9f64820SYuchung Cheng /* Update delivered info for the new segment */ 1382b9f64820SYuchung Cheng TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx; 1383b9f64820SYuchung Cheng 13846475be16SDavid S. Miller /* If this packet has been sent out already, we must 13856475be16SDavid S. Miller * adjust the various packet counters. 13866475be16SDavid S. Miller */ 1387cf0b450cSHerbert Xu if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 13886475be16SDavid S. Miller int diff = old_factor - tcp_skb_pcount(skb) - 13896475be16SDavid S. Miller tcp_skb_pcount(buff); 13901da177e4SLinus Torvalds 1391797108d1SIlpo Järvinen if (diff) 1392797108d1SIlpo Järvinen tcp_adjust_pcount(sk, skb, diff); 13931da177e4SLinus Torvalds } 13941da177e4SLinus Torvalds 13951da177e4SLinus Torvalds /* Link BUFF into the send queue. */ 1396f4a775d1SEric Dumazet __skb_header_release(buff); 139775c119afSEric Dumazet tcp_insert_write_queue_after(skb, buff, sk, tcp_queue); 1398*f67971e6SEric Dumazet if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE) 1399e2080072SEric Dumazet list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor); 14001da177e4SLinus Torvalds 14011da177e4SLinus Torvalds return 0; 14021da177e4SLinus Torvalds } 14031da177e4SLinus Torvalds 1404f4d01666SEric Dumazet /* This is similar to __pskb_pull_tail(). The difference is that pulled 1405f4d01666SEric Dumazet * data is not copied, but immediately discarded. 14061da177e4SLinus Torvalds */ 14077162fb24SEric Dumazet static int __pskb_trim_head(struct sk_buff *skb, int len) 14081da177e4SLinus Torvalds { 14097b7fc97aSEric Dumazet struct skb_shared_info *shinfo; 14101da177e4SLinus Torvalds int i, k, eat; 14111da177e4SLinus Torvalds 14124fa48bf3SEric Dumazet eat = min_t(int, len, skb_headlen(skb)); 14134fa48bf3SEric Dumazet if (eat) { 14144fa48bf3SEric Dumazet __skb_pull(skb, eat); 14154fa48bf3SEric Dumazet len -= eat; 14164fa48bf3SEric Dumazet if (!len) 14177162fb24SEric Dumazet return 0; 14184fa48bf3SEric Dumazet } 14191da177e4SLinus Torvalds eat = len; 14201da177e4SLinus Torvalds k = 0; 14217b7fc97aSEric Dumazet shinfo = skb_shinfo(skb); 14227b7fc97aSEric Dumazet for (i = 0; i < shinfo->nr_frags; i++) { 14237b7fc97aSEric Dumazet int size = skb_frag_size(&shinfo->frags[i]); 14249e903e08SEric Dumazet 14259e903e08SEric Dumazet if (size <= eat) { 1426aff65da0SIan Campbell skb_frag_unref(skb, i); 14279e903e08SEric Dumazet eat -= size; 14281da177e4SLinus Torvalds } else { 14297b7fc97aSEric Dumazet shinfo->frags[k] = shinfo->frags[i]; 14301da177e4SLinus Torvalds if (eat) { 14317b7fc97aSEric Dumazet shinfo->frags[k].page_offset += eat; 14327b7fc97aSEric Dumazet skb_frag_size_sub(&shinfo->frags[k], eat); 14331da177e4SLinus Torvalds eat = 0; 14341da177e4SLinus Torvalds } 14351da177e4SLinus Torvalds k++; 14361da177e4SLinus Torvalds } 14371da177e4SLinus Torvalds } 14387b7fc97aSEric Dumazet shinfo->nr_frags = k; 14391da177e4SLinus Torvalds 14401da177e4SLinus Torvalds skb->data_len -= len; 14411da177e4SLinus Torvalds skb->len = skb->data_len; 14427162fb24SEric Dumazet return len; 14431da177e4SLinus Torvalds } 14441da177e4SLinus Torvalds 144567edfef7SAndi Kleen /* Remove acked data from a packet in the transmit queue. */ 14461da177e4SLinus Torvalds int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 14471da177e4SLinus Torvalds { 14487162fb24SEric Dumazet u32 delta_truesize; 14497162fb24SEric Dumazet 145014bbd6a5SPravin B Shelar if (skb_unclone(skb, GFP_ATOMIC)) 14511da177e4SLinus Torvalds return -ENOMEM; 14521da177e4SLinus Torvalds 14537162fb24SEric Dumazet delta_truesize = __pskb_trim_head(skb, len); 14541da177e4SLinus Torvalds 14551da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq += len; 145684fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 14571da177e4SLinus Torvalds 14587162fb24SEric Dumazet if (delta_truesize) { 14597162fb24SEric Dumazet skb->truesize -= delta_truesize; 14607162fb24SEric Dumazet sk->sk_wmem_queued -= delta_truesize; 14617162fb24SEric Dumazet sk_mem_uncharge(sk, delta_truesize); 14621da177e4SLinus Torvalds sock_set_flag(sk, SOCK_QUEUE_SHRUNK); 14637162fb24SEric Dumazet } 14641da177e4SLinus Torvalds 14655b35e1e6SNeal Cardwell /* Any change of skb->len requires recalculation of tso factor. */ 14661da177e4SLinus Torvalds if (tcp_skb_pcount(skb) > 1) 14675bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb)); 14681da177e4SLinus Torvalds 14691da177e4SLinus Torvalds return 0; 14701da177e4SLinus Torvalds } 14711da177e4SLinus Torvalds 14721b63edd6SYuchung Cheng /* Calculate MSS not accounting any TCP options. */ 14731b63edd6SYuchung Cheng static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu) 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 mss_now; 14785d424d5aSJohn Heffner 14795d424d5aSJohn Heffner /* Calculate base mss without TCP options: 14805d424d5aSJohn Heffner It is MMS_S - sizeof(tcphdr) of rfc1122 14815d424d5aSJohn Heffner */ 14825d424d5aSJohn Heffner mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 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 mss_now -= icsk->icsk_af_ops->net_frag_header_len; 149067469601SEric Dumazet } 149167469601SEric Dumazet 14925d424d5aSJohn Heffner /* Clamp it (mss_clamp does not include tcp options) */ 14935d424d5aSJohn Heffner if (mss_now > tp->rx_opt.mss_clamp) 14945d424d5aSJohn Heffner mss_now = tp->rx_opt.mss_clamp; 14955d424d5aSJohn Heffner 14965d424d5aSJohn Heffner /* Now subtract optional transport overhead */ 14975d424d5aSJohn Heffner mss_now -= icsk->icsk_ext_hdr_len; 14985d424d5aSJohn Heffner 14995d424d5aSJohn Heffner /* Then reserve room for full set of TCP options and 8 bytes of data */ 15005d424d5aSJohn Heffner if (mss_now < 48) 15015d424d5aSJohn Heffner mss_now = 48; 15025d424d5aSJohn Heffner return mss_now; 15035d424d5aSJohn Heffner } 15045d424d5aSJohn Heffner 15051b63edd6SYuchung Cheng /* Calculate MSS. Not accounting for SACKs here. */ 15061b63edd6SYuchung Cheng int tcp_mtu_to_mss(struct sock *sk, int pmtu) 15071b63edd6SYuchung Cheng { 15081b63edd6SYuchung Cheng /* Subtract TCP options size, not including SACKs */ 15091b63edd6SYuchung Cheng return __tcp_mtu_to_mss(sk, pmtu) - 15101b63edd6SYuchung Cheng (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr)); 15111b63edd6SYuchung Cheng } 15121b63edd6SYuchung Cheng 15135d424d5aSJohn Heffner /* Inverse of above */ 151467469601SEric Dumazet int tcp_mss_to_mtu(struct sock *sk, int mss) 15155d424d5aSJohn Heffner { 1516cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1517cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 15185d424d5aSJohn Heffner int mtu; 15195d424d5aSJohn Heffner 15205d424d5aSJohn Heffner mtu = mss + 15215d424d5aSJohn Heffner tp->tcp_header_len + 15225d424d5aSJohn Heffner icsk->icsk_ext_hdr_len + 15235d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 15245d424d5aSJohn Heffner 152567469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 152667469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 152767469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 152867469601SEric Dumazet 152967469601SEric Dumazet if (dst && dst_allfrag(dst)) 153067469601SEric Dumazet mtu += icsk->icsk_af_ops->net_frag_header_len; 153167469601SEric Dumazet } 15325d424d5aSJohn Heffner return mtu; 15335d424d5aSJohn Heffner } 1534556c6b46SNeal Cardwell EXPORT_SYMBOL(tcp_mss_to_mtu); 15355d424d5aSJohn Heffner 153667edfef7SAndi Kleen /* MTU probing init per socket */ 15375d424d5aSJohn Heffner void tcp_mtup_init(struct sock *sk) 15385d424d5aSJohn Heffner { 15395d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 15405d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 1541b0f9ca53SFan Du struct net *net = sock_net(sk); 15425d424d5aSJohn Heffner 1543b0f9ca53SFan Du icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1; 15445d424d5aSJohn Heffner icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 15455d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 1546b0f9ca53SFan Du icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss); 15475d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 154805cbc0dbSFan Du if (icsk->icsk_mtup.enabled) 1549c74df29aSEric Dumazet icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; 15505d424d5aSJohn Heffner } 15514bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_mtup_init); 15525d424d5aSJohn Heffner 15531da177e4SLinus Torvalds /* This function synchronize snd mss to current pmtu/exthdr set. 15541da177e4SLinus Torvalds 15551da177e4SLinus Torvalds tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 15561da177e4SLinus Torvalds for TCP options, but includes only bare TCP header. 15571da177e4SLinus Torvalds 15581da177e4SLinus Torvalds tp->rx_opt.mss_clamp is mss negotiated at connection setup. 1559caa20d9aSStephen Hemminger It is minimum of user_mss and mss received with SYN. 15601da177e4SLinus Torvalds It also does not include TCP options. 15611da177e4SLinus Torvalds 1562d83d8461SArnaldo Carvalho de Melo inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. 15631da177e4SLinus Torvalds 15641da177e4SLinus Torvalds tp->mss_cache is current effective sending mss, including 15651da177e4SLinus Torvalds all tcp options except for SACKs. It is evaluated, 15661da177e4SLinus Torvalds taking into account current pmtu, but never exceeds 15671da177e4SLinus Torvalds tp->rx_opt.mss_clamp. 15681da177e4SLinus Torvalds 15691da177e4SLinus Torvalds NOTE1. rfc1122 clearly states that advertised MSS 15701da177e4SLinus Torvalds DOES NOT include either tcp or ip options. 15711da177e4SLinus Torvalds 1572d83d8461SArnaldo Carvalho de Melo NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache 1573d83d8461SArnaldo Carvalho de Melo are READ ONLY outside this function. --ANK (980731) 15741da177e4SLinus Torvalds */ 15751da177e4SLinus Torvalds unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 15761da177e4SLinus Torvalds { 15771da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1578d83d8461SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 15795d424d5aSJohn Heffner int mss_now; 15801da177e4SLinus Torvalds 15815d424d5aSJohn Heffner if (icsk->icsk_mtup.search_high > pmtu) 15825d424d5aSJohn Heffner icsk->icsk_mtup.search_high = pmtu; 15831da177e4SLinus Torvalds 15845d424d5aSJohn Heffner mss_now = tcp_mtu_to_mss(sk, pmtu); 1585409d22b4SIlpo Järvinen mss_now = tcp_bound_to_half_wnd(tp, mss_now); 15861da177e4SLinus Torvalds 15871da177e4SLinus Torvalds /* And store cached results */ 1588d83d8461SArnaldo Carvalho de Melo icsk->icsk_pmtu_cookie = pmtu; 15895d424d5aSJohn Heffner if (icsk->icsk_mtup.enabled) 15905d424d5aSJohn Heffner mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); 1591c1b4a7e6SDavid S. Miller tp->mss_cache = mss_now; 15921da177e4SLinus Torvalds 15931da177e4SLinus Torvalds return mss_now; 15941da177e4SLinus Torvalds } 15954bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_sync_mss); 15961da177e4SLinus Torvalds 15971da177e4SLinus Torvalds /* Compute the current effective MSS, taking SACKs and IP options, 15981da177e4SLinus Torvalds * and even PMTU discovery events into account. 15991da177e4SLinus Torvalds */ 16000c54b85fSIlpo Järvinen unsigned int tcp_current_mss(struct sock *sk) 16011da177e4SLinus Torvalds { 1602cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1603cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1604c1b4a7e6SDavid S. Miller u32 mss_now; 160595c96174SEric Dumazet unsigned int header_len; 160633ad798cSAdam Langley struct tcp_out_options opts; 160733ad798cSAdam Langley struct tcp_md5sig_key *md5; 16081da177e4SLinus Torvalds 1609c1b4a7e6SDavid S. Miller mss_now = tp->mss_cache; 1610c1b4a7e6SDavid S. Miller 16111da177e4SLinus Torvalds if (dst) { 16121da177e4SLinus Torvalds u32 mtu = dst_mtu(dst); 1613d83d8461SArnaldo Carvalho de Melo if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 16141da177e4SLinus Torvalds mss_now = tcp_sync_mss(sk, mtu); 16151da177e4SLinus Torvalds } 16161da177e4SLinus Torvalds 161733ad798cSAdam Langley header_len = tcp_established_options(sk, NULL, &opts, &md5) + 161833ad798cSAdam Langley sizeof(struct tcphdr); 161933ad798cSAdam Langley /* The mss_cache is sized based on tp->tcp_header_len, which assumes 162033ad798cSAdam Langley * some common options. If this is an odd packet (because we have SACK 162133ad798cSAdam Langley * blocks etc) then our calculated header_len will be different, and 162233ad798cSAdam Langley * we have to adjust mss_now correspondingly */ 162333ad798cSAdam Langley if (header_len != tp->tcp_header_len) { 162433ad798cSAdam Langley int delta = (int) header_len - tp->tcp_header_len; 162533ad798cSAdam Langley mss_now -= delta; 162633ad798cSAdam Langley } 1627cfb6eeb4SYOSHIFUJI Hideaki 16281da177e4SLinus Torvalds return mss_now; 16291da177e4SLinus Torvalds } 16301da177e4SLinus Torvalds 163186fd14adSWeiping Pan /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. 163286fd14adSWeiping Pan * As additional protections, we do not touch cwnd in retransmission phases, 163386fd14adSWeiping Pan * and if application hit its sndbuf limit recently. 163486fd14adSWeiping Pan */ 163586fd14adSWeiping Pan static void tcp_cwnd_application_limited(struct sock *sk) 1636a762a980SDavid S. Miller { 16379e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1638a762a980SDavid S. Miller 163986fd14adSWeiping Pan if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && 164086fd14adSWeiping Pan sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 164186fd14adSWeiping Pan /* Limited by application or receiver window. */ 164286fd14adSWeiping Pan u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); 164386fd14adSWeiping Pan u32 win_used = max(tp->snd_cwnd_used, init_win); 164486fd14adSWeiping Pan if (win_used < tp->snd_cwnd) { 164586fd14adSWeiping Pan tp->snd_ssthresh = tcp_current_ssthresh(sk); 164686fd14adSWeiping Pan tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1; 164786fd14adSWeiping Pan } 164886fd14adSWeiping Pan tp->snd_cwnd_used = 0; 164986fd14adSWeiping Pan } 1650c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 165186fd14adSWeiping Pan } 165286fd14adSWeiping Pan 1653ca8a2263SNeal Cardwell static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited) 1654a762a980SDavid S. Miller { 16551b1fc3fdSWei Wang const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 1656a762a980SDavid S. Miller struct tcp_sock *tp = tcp_sk(sk); 1657a762a980SDavid S. Miller 1658ca8a2263SNeal Cardwell /* Track the maximum number of outstanding packets in each 1659ca8a2263SNeal Cardwell * window, and remember whether we were cwnd-limited then. 1660ca8a2263SNeal Cardwell */ 1661ca8a2263SNeal Cardwell if (!before(tp->snd_una, tp->max_packets_seq) || 1662ca8a2263SNeal Cardwell tp->packets_out > tp->max_packets_out) { 1663ca8a2263SNeal Cardwell tp->max_packets_out = tp->packets_out; 1664ca8a2263SNeal Cardwell tp->max_packets_seq = tp->snd_nxt; 1665ca8a2263SNeal Cardwell tp->is_cwnd_limited = is_cwnd_limited; 1666ca8a2263SNeal Cardwell } 1667e114a710SEric Dumazet 166824901551SEric Dumazet if (tcp_is_cwnd_limited(sk)) { 1669a762a980SDavid S. Miller /* Network is feed fully. */ 1670a762a980SDavid S. Miller tp->snd_cwnd_used = 0; 1671c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 1672a762a980SDavid S. Miller } else { 1673a762a980SDavid S. Miller /* Network starves. */ 1674a762a980SDavid S. Miller if (tp->packets_out > tp->snd_cwnd_used) 1675a762a980SDavid S. Miller tp->snd_cwnd_used = tp->packets_out; 1676a762a980SDavid S. Miller 1677b510f0d2SEric Dumazet if (sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle && 1678c2203cf7SEric Dumazet (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto && 16791b1fc3fdSWei Wang !ca_ops->cong_control) 1680a762a980SDavid S. Miller tcp_cwnd_application_limited(sk); 1681b0f71bd3SFrancis Yan 1682b0f71bd3SFrancis Yan /* The following conditions together indicate the starvation 1683b0f71bd3SFrancis Yan * is caused by insufficient sender buffer: 1684b0f71bd3SFrancis Yan * 1) just sent some data (see tcp_write_xmit) 1685b0f71bd3SFrancis Yan * 2) not cwnd limited (this else condition) 168675c119afSEric Dumazet * 3) no more data to send (tcp_write_queue_empty()) 1687b0f71bd3SFrancis Yan * 4) application is hitting buffer limit (SOCK_NOSPACE) 1688b0f71bd3SFrancis Yan */ 168975c119afSEric Dumazet if (tcp_write_queue_empty(sk) && sk->sk_socket && 1690b0f71bd3SFrancis Yan test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) && 1691b0f71bd3SFrancis Yan (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) 1692b0f71bd3SFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED); 1693a762a980SDavid S. Miller } 1694a762a980SDavid S. Miller } 1695a762a980SDavid S. Miller 1696d4589926SEric Dumazet /* Minshall's variant of the Nagle send check. */ 1697d4589926SEric Dumazet static bool tcp_minshall_check(const struct tcp_sock *tp) 1698d4589926SEric Dumazet { 1699d4589926SEric Dumazet return after(tp->snd_sml, tp->snd_una) && 1700d4589926SEric Dumazet !after(tp->snd_sml, tp->snd_nxt); 1701d4589926SEric Dumazet } 1702d4589926SEric Dumazet 1703d4589926SEric Dumazet /* Update snd_sml if this skb is under mss 1704d4589926SEric Dumazet * Note that a TSO packet might end with a sub-mss segment 1705d4589926SEric Dumazet * The test is really : 1706d4589926SEric Dumazet * if ((skb->len % mss) != 0) 1707d4589926SEric Dumazet * tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1708d4589926SEric Dumazet * But we can avoid doing the divide again given we already have 1709d4589926SEric Dumazet * skb_pcount = skb->len / mss_now 17100e3a4803SIlpo Järvinen */ 1711d4589926SEric Dumazet static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now, 1712d4589926SEric Dumazet const struct sk_buff *skb) 1713d4589926SEric Dumazet { 1714d4589926SEric Dumazet if (skb->len < tcp_skb_pcount(skb) * mss_now) 1715d4589926SEric Dumazet tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1716d4589926SEric Dumazet } 1717d4589926SEric Dumazet 1718d4589926SEric Dumazet /* Return false, if packet can be sent now without violation Nagle's rules: 1719d4589926SEric Dumazet * 1. It is full sized. (provided by caller in %partial bool) 1720d4589926SEric Dumazet * 2. Or it contains FIN. (already checked by caller) 1721d4589926SEric Dumazet * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. 1722d4589926SEric Dumazet * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 1723d4589926SEric Dumazet * With Minshall's modification: all sent small packets are ACKed. 1724d4589926SEric Dumazet */ 1725d4589926SEric Dumazet static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp, 1726cc93fc51SPeter Pan(潘卫平) int nonagle) 1727d4589926SEric Dumazet { 1728d4589926SEric Dumazet return partial && 1729d4589926SEric Dumazet ((nonagle & TCP_NAGLE_CORK) || 1730d4589926SEric Dumazet (!nonagle && tp->packets_out && tcp_minshall_check(tp))); 1731d4589926SEric Dumazet } 1732605ad7f1SEric Dumazet 1733605ad7f1SEric Dumazet /* Return how many segs we'd like on a TSO packet, 1734605ad7f1SEric Dumazet * to send one TSO packet per ms 1735605ad7f1SEric Dumazet */ 17361b3878caSNeal Cardwell u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now, 17371b3878caSNeal Cardwell int min_tso_segs) 1738605ad7f1SEric Dumazet { 1739605ad7f1SEric Dumazet u32 bytes, segs; 1740605ad7f1SEric Dumazet 1741605ad7f1SEric Dumazet bytes = min(sk->sk_pacing_rate >> 10, 1742605ad7f1SEric Dumazet sk->sk_gso_max_size - 1 - MAX_TCP_HEADER); 1743605ad7f1SEric Dumazet 1744605ad7f1SEric Dumazet /* Goal is to send at least one packet per ms, 1745605ad7f1SEric Dumazet * not one big TSO packet every 100 ms. 1746605ad7f1SEric Dumazet * This preserves ACK clocking and is consistent 1747605ad7f1SEric Dumazet * with tcp_tso_should_defer() heuristic. 1748605ad7f1SEric Dumazet */ 17491b3878caSNeal Cardwell segs = max_t(u32, bytes / mss_now, min_tso_segs); 1750605ad7f1SEric Dumazet 1751605ad7f1SEric Dumazet return min_t(u32, segs, sk->sk_gso_max_segs); 1752605ad7f1SEric Dumazet } 17531b3878caSNeal Cardwell EXPORT_SYMBOL(tcp_tso_autosize); 1754605ad7f1SEric Dumazet 1755ed6e7268SNeal Cardwell /* Return the number of segments we want in the skb we are transmitting. 1756ed6e7268SNeal Cardwell * See if congestion control module wants to decide; otherwise, autosize. 1757ed6e7268SNeal Cardwell */ 1758ed6e7268SNeal Cardwell static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now) 1759ed6e7268SNeal Cardwell { 1760ed6e7268SNeal Cardwell const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 1761ed6e7268SNeal Cardwell u32 tso_segs = ca_ops->tso_segs_goal ? ca_ops->tso_segs_goal(sk) : 0; 1762ed6e7268SNeal Cardwell 17631b3878caSNeal Cardwell return tso_segs ? : 176426e9596eSEric Dumazet tcp_tso_autosize(sk, mss_now, 176526e9596eSEric Dumazet sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs); 1766ed6e7268SNeal Cardwell } 1767ed6e7268SNeal Cardwell 1768d4589926SEric Dumazet /* Returns the portion of skb which can be sent right away */ 1769d4589926SEric Dumazet static unsigned int tcp_mss_split_point(const struct sock *sk, 1770d4589926SEric Dumazet const struct sk_buff *skb, 1771d4589926SEric Dumazet unsigned int mss_now, 1772d4589926SEric Dumazet unsigned int max_segs, 1773d4589926SEric Dumazet int nonagle) 1774c1b4a7e6SDavid S. Miller { 1775cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1776d4589926SEric Dumazet u32 partial, needed, window, max_len; 1777c1b4a7e6SDavid S. Miller 177890840defSIlpo Järvinen window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 17791485348dSBen Hutchings max_len = mss_now * max_segs; 17800e3a4803SIlpo Järvinen 17811485348dSBen Hutchings if (likely(max_len <= window && skb != tcp_write_queue_tail(sk))) 17821485348dSBen Hutchings return max_len; 17830e3a4803SIlpo Järvinen 17845ea3a748SIlpo Järvinen needed = min(skb->len, window); 17855ea3a748SIlpo Järvinen 17861485348dSBen Hutchings if (max_len <= needed) 17871485348dSBen Hutchings return max_len; 17880e3a4803SIlpo Järvinen 1789d4589926SEric Dumazet partial = needed % mss_now; 1790d4589926SEric Dumazet /* If last segment is not a full MSS, check if Nagle rules allow us 1791d4589926SEric Dumazet * to include this last segment in this skb. 1792d4589926SEric Dumazet * Otherwise, we'll split the skb at last MSS boundary 1793d4589926SEric Dumazet */ 1794cc93fc51SPeter Pan(潘卫平) if (tcp_nagle_check(partial != 0, tp, nonagle)) 1795d4589926SEric Dumazet return needed - partial; 1796d4589926SEric Dumazet 1797d4589926SEric Dumazet return needed; 1798c1b4a7e6SDavid S. Miller } 1799c1b4a7e6SDavid S. Miller 1800c1b4a7e6SDavid S. Miller /* Can at least one segment of SKB be sent right now, according to the 1801c1b4a7e6SDavid S. Miller * congestion window rules? If so, return how many segments are allowed. 1802c1b4a7e6SDavid S. Miller */ 1803cf533ea5SEric Dumazet static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, 1804cf533ea5SEric Dumazet const struct sk_buff *skb) 1805c1b4a7e6SDavid S. Miller { 1806d649a7a8SEric Dumazet u32 in_flight, cwnd, halfcwnd; 1807c1b4a7e6SDavid S. Miller 1808c1b4a7e6SDavid S. Miller /* Don't be strict about the congestion window for the final FIN. */ 18094de075e0SEric Dumazet if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && 18104de075e0SEric Dumazet tcp_skb_pcount(skb) == 1) 1811c1b4a7e6SDavid S. Miller return 1; 1812c1b4a7e6SDavid S. Miller 1813c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1814c1b4a7e6SDavid S. Miller cwnd = tp->snd_cwnd; 1815d649a7a8SEric Dumazet if (in_flight >= cwnd) 1816c1b4a7e6SDavid S. Miller return 0; 1817d649a7a8SEric Dumazet 1818d649a7a8SEric Dumazet /* For better scheduling, ensure we have at least 1819d649a7a8SEric Dumazet * 2 GSO packets in flight. 1820d649a7a8SEric Dumazet */ 1821d649a7a8SEric Dumazet halfcwnd = max(cwnd >> 1, 1U); 1822d649a7a8SEric Dumazet return min(halfcwnd, cwnd - in_flight); 1823c1b4a7e6SDavid S. Miller } 1824c1b4a7e6SDavid S. Miller 1825b595076aSUwe Kleine-König /* Initialize TSO state of a skb. 182667edfef7SAndi Kleen * This must be invoked the first time we consider transmitting 1827c1b4a7e6SDavid S. Miller * SKB onto the wire. 1828c1b4a7e6SDavid S. Miller */ 18295bbb432cSEric Dumazet static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now) 1830c1b4a7e6SDavid S. Miller { 1831c1b4a7e6SDavid S. Miller int tso_segs = tcp_skb_pcount(skb); 1832c1b4a7e6SDavid S. Miller 1833f8269a49SIlpo Järvinen if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { 18345bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 1835c1b4a7e6SDavid S. Miller tso_segs = tcp_skb_pcount(skb); 1836c1b4a7e6SDavid S. Miller } 1837c1b4a7e6SDavid S. Miller return tso_segs; 1838c1b4a7e6SDavid S. Miller } 1839c1b4a7e6SDavid S. Miller 1840c1b4a7e6SDavid S. Miller 1841a2a385d6SEric Dumazet /* Return true if the Nagle test allows this packet to be 1842c1b4a7e6SDavid S. Miller * sent now. 1843c1b4a7e6SDavid S. Miller */ 1844a2a385d6SEric Dumazet static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, 1845c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1846c1b4a7e6SDavid S. Miller { 1847c1b4a7e6SDavid S. Miller /* Nagle rule does not apply to frames, which sit in the middle of the 1848c1b4a7e6SDavid S. Miller * write_queue (they have no chances to get new data). 1849c1b4a7e6SDavid S. Miller * 1850c1b4a7e6SDavid S. Miller * This is implemented in the callers, where they modify the 'nonagle' 1851c1b4a7e6SDavid S. Miller * argument based upon the location of SKB in the send queue. 1852c1b4a7e6SDavid S. Miller */ 1853c1b4a7e6SDavid S. Miller if (nonagle & TCP_NAGLE_PUSH) 1854a2a385d6SEric Dumazet return true; 1855c1b4a7e6SDavid S. Miller 18569b44190dSYuchung Cheng /* Don't use the nagle rule for urgent data (or for the final FIN). */ 18579b44190dSYuchung Cheng if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) 1858a2a385d6SEric Dumazet return true; 1859c1b4a7e6SDavid S. Miller 1860cc93fc51SPeter Pan(潘卫平) if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle)) 1861a2a385d6SEric Dumazet return true; 1862c1b4a7e6SDavid S. Miller 1863a2a385d6SEric Dumazet return false; 1864c1b4a7e6SDavid S. Miller } 1865c1b4a7e6SDavid S. Miller 1866c1b4a7e6SDavid S. Miller /* Does at least the first segment of SKB fit into the send window? */ 1867a2a385d6SEric Dumazet static bool tcp_snd_wnd_test(const struct tcp_sock *tp, 1868a2a385d6SEric Dumazet const struct sk_buff *skb, 1869056834d9SIlpo Järvinen unsigned int cur_mss) 1870c1b4a7e6SDavid S. Miller { 1871c1b4a7e6SDavid S. Miller u32 end_seq = TCP_SKB_CB(skb)->end_seq; 1872c1b4a7e6SDavid S. Miller 1873c1b4a7e6SDavid S. Miller if (skb->len > cur_mss) 1874c1b4a7e6SDavid S. Miller end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 1875c1b4a7e6SDavid S. Miller 187690840defSIlpo Järvinen return !after(end_seq, tcp_wnd_end(tp)); 1877c1b4a7e6SDavid S. Miller } 1878c1b4a7e6SDavid S. Miller 1879c1b4a7e6SDavid S. Miller /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 1880c1b4a7e6SDavid S. Miller * which is put after SKB on the list. It is very much like 1881c1b4a7e6SDavid S. Miller * tcp_fragment() except that it may make several kinds of assumptions 1882c1b4a7e6SDavid S. Miller * in order to speed up the splitting operation. In particular, we 1883c1b4a7e6SDavid S. Miller * know that all the data is in scatter-gather pages, and that the 1884c1b4a7e6SDavid S. Miller * packet has never been sent out before (and thus is not cloned). 1885c1b4a7e6SDavid S. Miller */ 188675c119afSEric Dumazet static int tso_fragment(struct sock *sk, enum tcp_queue tcp_queue, 188775c119afSEric Dumazet struct sk_buff *skb, unsigned int len, 1888c4ead4c5SEric Dumazet unsigned int mss_now, gfp_t gfp) 1889c1b4a7e6SDavid S. Miller { 1890c1b4a7e6SDavid S. Miller struct sk_buff *buff; 1891c1b4a7e6SDavid S. Miller int nlen = skb->len - len; 18929ce01461SIlpo Järvinen u8 flags; 1893c1b4a7e6SDavid S. Miller 1894c1b4a7e6SDavid S. Miller /* All of a TSO frame must be composed of paged data. */ 1895c8ac3774SHerbert Xu if (skb->len != skb->data_len) 189675c119afSEric Dumazet return tcp_fragment(sk, tcp_queue, skb, len, mss_now, gfp); 1897c1b4a7e6SDavid S. Miller 1898eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, gfp, true); 189951456b29SIan Morris if (unlikely(!buff)) 1900c1b4a7e6SDavid S. Miller return -ENOMEM; 1901c1b4a7e6SDavid S. Miller 19023ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 19033ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1904b60b49eaSHerbert Xu buff->truesize += nlen; 1905c1b4a7e6SDavid S. Miller skb->truesize -= nlen; 1906c1b4a7e6SDavid S. Miller 1907c1b4a7e6SDavid S. Miller /* Correct the sequence numbers. */ 1908c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 1909c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 1910c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 1911c1b4a7e6SDavid S. Miller 1912c1b4a7e6SDavid S. Miller /* PSH and FIN should only be set in the second packet. */ 19134de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 19144de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 19154de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1916c1b4a7e6SDavid S. Miller 1917c1b4a7e6SDavid S. Miller /* This packet was never sent out yet, so no SACK bits. */ 1918c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->sacked = 0; 1919c1b4a7e6SDavid S. Miller 1920a166140eSMartin KaFai Lau tcp_skb_fragment_eor(skb, buff); 1921a166140eSMartin KaFai Lau 192284fa7933SPatrick McHardy buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; 1923c1b4a7e6SDavid S. Miller skb_split(skb, buff, len); 1924490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 1925c1b4a7e6SDavid S. Miller 1926c1b4a7e6SDavid S. Miller /* Fix up tso_factor for both original and new SKB. */ 19275bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 19285bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 1929c1b4a7e6SDavid S. Miller 1930c1b4a7e6SDavid S. Miller /* Link BUFF into the send queue. */ 1931f4a775d1SEric Dumazet __skb_header_release(buff); 193275c119afSEric Dumazet tcp_insert_write_queue_after(skb, buff, sk, tcp_queue); 1933c1b4a7e6SDavid S. Miller 1934c1b4a7e6SDavid S. Miller return 0; 1935c1b4a7e6SDavid S. Miller } 1936c1b4a7e6SDavid S. Miller 1937c1b4a7e6SDavid S. Miller /* Try to defer sending, if possible, in order to minimize the amount 1938c1b4a7e6SDavid S. Miller * of TSO splitting we do. View it as a kind of TSO Nagle test. 1939c1b4a7e6SDavid S. Miller * 1940c1b4a7e6SDavid S. Miller * This algorithm is from John Heffner. 1941c1b4a7e6SDavid S. Miller */ 1942ca8a2263SNeal Cardwell static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb, 1943605ad7f1SEric Dumazet bool *is_cwnd_limited, u32 max_segs) 1944c1b4a7e6SDavid S. Miller { 19456687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 194650c8339eSEric Dumazet u32 age, send_win, cong_win, limit, in_flight; 194750c8339eSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 194850c8339eSEric Dumazet struct sk_buff *head; 1949ad9f4f50SEric Dumazet int win_divisor; 1950c1b4a7e6SDavid S. Miller 19514de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 1952ae8064acSJohn Heffner goto send_now; 1953c1b4a7e6SDavid S. Miller 195499d7662aSEric Dumazet if (icsk->icsk_ca_state >= TCP_CA_Recovery) 1955ae8064acSJohn Heffner goto send_now; 1956ae8064acSJohn Heffner 19575f852eb5SEric Dumazet /* Avoid bursty behavior by allowing defer 19585f852eb5SEric Dumazet * only if the last write was recent. 19595f852eb5SEric Dumazet */ 1960d635fbe2SEric Dumazet if ((s32)(tcp_jiffies32 - tp->lsndtime) > 0) 1961ae8064acSJohn Heffner goto send_now; 1962908a75c1SDavid S. Miller 1963c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1964c1b4a7e6SDavid S. Miller 1965056834d9SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight)); 1966c1b4a7e6SDavid S. Miller 196790840defSIlpo Järvinen send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1968c1b4a7e6SDavid S. Miller 1969c1b4a7e6SDavid S. Miller /* From in_flight test above, we know that cwnd > in_flight. */ 1970c1b4a7e6SDavid S. Miller cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; 1971c1b4a7e6SDavid S. Miller 1972c1b4a7e6SDavid S. Miller limit = min(send_win, cong_win); 1973c1b4a7e6SDavid S. Miller 1974ba244fe9SDavid S. Miller /* If a full-sized TSO skb can be sent, do it. */ 1975605ad7f1SEric Dumazet if (limit >= max_segs * tp->mss_cache) 1976ae8064acSJohn Heffner goto send_now; 1977ba244fe9SDavid S. Miller 197862ad2761SIlpo Järvinen /* Middle in queue won't get any more data, full sendable already? */ 197962ad2761SIlpo Järvinen if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) 198062ad2761SIlpo Järvinen goto send_now; 198162ad2761SIlpo Järvinen 1982d06a9904SEric Dumazet win_divisor = ACCESS_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor); 1983ad9f4f50SEric Dumazet if (win_divisor) { 1984c1b4a7e6SDavid S. Miller u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); 1985c1b4a7e6SDavid S. Miller 1986c1b4a7e6SDavid S. Miller /* If at least some fraction of a window is available, 1987c1b4a7e6SDavid S. Miller * just use it. 1988c1b4a7e6SDavid S. Miller */ 1989ad9f4f50SEric Dumazet chunk /= win_divisor; 1990c1b4a7e6SDavid S. Miller if (limit >= chunk) 1991ae8064acSJohn Heffner goto send_now; 1992c1b4a7e6SDavid S. Miller } else { 1993c1b4a7e6SDavid S. Miller /* Different approach, try not to defer past a single 1994c1b4a7e6SDavid S. Miller * ACK. Receiver should ACK every other full sized 1995c1b4a7e6SDavid S. Miller * frame, so if we have space for more than 3 frames 1996c1b4a7e6SDavid S. Miller * then send now. 1997c1b4a7e6SDavid S. Miller */ 19986b5a5c0dSNeal Cardwell if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) 1999ae8064acSJohn Heffner goto send_now; 2000c1b4a7e6SDavid S. Miller } 2001c1b4a7e6SDavid S. Miller 200275c119afSEric Dumazet /* TODO : use tsorted_sent_queue ? */ 200375c119afSEric Dumazet head = tcp_rtx_queue_head(sk); 200475c119afSEric Dumazet if (!head) 200575c119afSEric Dumazet goto send_now; 20069a568de4SEric Dumazet age = tcp_stamp_us_delta(tp->tcp_mstamp, head->skb_mstamp); 200750c8339eSEric Dumazet /* If next ACK is likely to come too late (half srtt), do not defer */ 200850c8339eSEric Dumazet if (age < (tp->srtt_us >> 4)) 200950c8339eSEric Dumazet goto send_now; 201050c8339eSEric Dumazet 20115f852eb5SEric Dumazet /* Ok, it looks like it is advisable to defer. */ 2012ae8064acSJohn Heffner 2013d2e1339fSBendik Rønning Opstad if (cong_win < send_win && cong_win <= skb->len) 2014ca8a2263SNeal Cardwell *is_cwnd_limited = true; 2015ca8a2263SNeal Cardwell 2016a2a385d6SEric Dumazet return true; 2017ae8064acSJohn Heffner 2018ae8064acSJohn Heffner send_now: 2019a2a385d6SEric Dumazet return false; 2020c1b4a7e6SDavid S. Miller } 2021c1b4a7e6SDavid S. Miller 202205cbc0dbSFan Du static inline void tcp_mtu_check_reprobe(struct sock *sk) 202305cbc0dbSFan Du { 202405cbc0dbSFan Du struct inet_connection_sock *icsk = inet_csk(sk); 202505cbc0dbSFan Du struct tcp_sock *tp = tcp_sk(sk); 202605cbc0dbSFan Du struct net *net = sock_net(sk); 202705cbc0dbSFan Du u32 interval; 202805cbc0dbSFan Du s32 delta; 202905cbc0dbSFan Du 203005cbc0dbSFan Du interval = net->ipv4.sysctl_tcp_probe_interval; 2031c74df29aSEric Dumazet delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp; 203205cbc0dbSFan Du if (unlikely(delta >= interval * HZ)) { 203305cbc0dbSFan Du int mss = tcp_current_mss(sk); 203405cbc0dbSFan Du 203505cbc0dbSFan Du /* Update current search range */ 203605cbc0dbSFan Du icsk->icsk_mtup.probe_size = 0; 203705cbc0dbSFan Du icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + 203805cbc0dbSFan Du sizeof(struct tcphdr) + 203905cbc0dbSFan Du icsk->icsk_af_ops->net_header_len; 204005cbc0dbSFan Du icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss); 204105cbc0dbSFan Du 204205cbc0dbSFan Du /* Update probe time stamp */ 2043c74df29aSEric Dumazet icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; 204405cbc0dbSFan Du } 204505cbc0dbSFan Du } 204605cbc0dbSFan Du 20475d424d5aSJohn Heffner /* Create a new MTU probe if we are ready. 204867edfef7SAndi Kleen * MTU probe is regularly attempting to increase the path MTU by 204967edfef7SAndi Kleen * deliberately sending larger packets. This discovers routing 205067edfef7SAndi Kleen * changes resulting in larger path MTUs. 205167edfef7SAndi Kleen * 20525d424d5aSJohn Heffner * Returns 0 if we should wait to probe (no cwnd available), 20535d424d5aSJohn Heffner * 1 if a probe was sent, 2054056834d9SIlpo Järvinen * -1 otherwise 2055056834d9SIlpo Järvinen */ 20565d424d5aSJohn Heffner static int tcp_mtu_probe(struct sock *sk) 20575d424d5aSJohn Heffner { 20585d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 205912a59abcSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 20605d424d5aSJohn Heffner struct sk_buff *skb, *nskb, *next; 20616b58e0a5SFan Du struct net *net = sock_net(sk); 20625d424d5aSJohn Heffner int probe_size; 206391cc17c0SIlpo Järvinen int size_needed; 206412a59abcSEric Dumazet int copy, len; 20655d424d5aSJohn Heffner int mss_now; 20666b58e0a5SFan Du int interval; 20675d424d5aSJohn Heffner 20685d424d5aSJohn Heffner /* Not currently probing/verifying, 20695d424d5aSJohn Heffner * not in recovery, 20705d424d5aSJohn Heffner * have enough cwnd, and 207112a59abcSEric Dumazet * not SACKing (the variable headers throw things off) 207212a59abcSEric Dumazet */ 207312a59abcSEric Dumazet if (likely(!icsk->icsk_mtup.enabled || 20745d424d5aSJohn Heffner icsk->icsk_mtup.probe_size || 20755d424d5aSJohn Heffner inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 20765d424d5aSJohn Heffner tp->snd_cwnd < 11 || 207712a59abcSEric Dumazet tp->rx_opt.num_sacks || tp->rx_opt.dsack)) 20785d424d5aSJohn Heffner return -1; 20795d424d5aSJohn Heffner 20806b58e0a5SFan Du /* Use binary search for probe_size between tcp_mss_base, 20816b58e0a5SFan Du * and current mss_clamp. if (search_high - search_low) 20826b58e0a5SFan Du * smaller than a threshold, backoff from probing. 20836b58e0a5SFan Du */ 20840c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 20856b58e0a5SFan Du probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high + 20866b58e0a5SFan Du icsk->icsk_mtup.search_low) >> 1); 208791cc17c0SIlpo Järvinen size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; 20886b58e0a5SFan Du interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low; 208905cbc0dbSFan Du /* When misfortune happens, we are reprobing actively, 209005cbc0dbSFan Du * and then reprobe timer has expired. We stick with current 209105cbc0dbSFan Du * probing process by not resetting search range to its orignal. 209205cbc0dbSFan Du */ 20936b58e0a5SFan Du if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) || 209405cbc0dbSFan Du interval < net->ipv4.sysctl_tcp_probe_threshold) { 209505cbc0dbSFan Du /* Check whether enough time has elaplased for 209605cbc0dbSFan Du * another round of probing. 209705cbc0dbSFan Du */ 209805cbc0dbSFan Du tcp_mtu_check_reprobe(sk); 20995d424d5aSJohn Heffner return -1; 21005d424d5aSJohn Heffner } 21015d424d5aSJohn Heffner 21025d424d5aSJohn Heffner /* Have enough data in the send queue to probe? */ 21037f9c33e5SIlpo Järvinen if (tp->write_seq - tp->snd_nxt < size_needed) 21045d424d5aSJohn Heffner return -1; 21055d424d5aSJohn Heffner 210691cc17c0SIlpo Järvinen if (tp->snd_wnd < size_needed) 21075d424d5aSJohn Heffner return -1; 210890840defSIlpo Järvinen if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) 21095d424d5aSJohn Heffner return 0; 21105d424d5aSJohn Heffner 2111d67c58e9SIlpo Järvinen /* Do we need to wait to drain cwnd? With none in flight, don't stall */ 2112d67c58e9SIlpo Järvinen if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { 2113d67c58e9SIlpo Järvinen if (!tcp_packets_in_flight(tp)) 21145d424d5aSJohn Heffner return -1; 21155d424d5aSJohn Heffner else 21165d424d5aSJohn Heffner return 0; 21175d424d5aSJohn Heffner } 21185d424d5aSJohn Heffner 21195d424d5aSJohn Heffner /* We're allowed to probe. Build it now. */ 2120eb934478SEric Dumazet nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false); 212151456b29SIan Morris if (!nskb) 21225d424d5aSJohn Heffner return -1; 21233ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 21243ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 21255d424d5aSJohn Heffner 2126fe067e8aSDavid S. Miller skb = tcp_send_head(sk); 21275d424d5aSJohn Heffner 21285d424d5aSJohn Heffner TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 21295d424d5aSJohn Heffner TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 21304de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; 21315d424d5aSJohn Heffner TCP_SKB_CB(nskb)->sacked = 0; 21325d424d5aSJohn Heffner nskb->csum = 0; 213384fa7933SPatrick McHardy nskb->ip_summed = skb->ip_summed; 21345d424d5aSJohn Heffner 213550c4817eSIlpo Järvinen tcp_insert_write_queue_before(nskb, skb, sk); 21362b7cda9cSEric Dumazet tcp_highest_sack_replace(sk, skb, nskb); 213750c4817eSIlpo Järvinen 21385d424d5aSJohn Heffner len = 0; 2139234b6860SIlpo Järvinen tcp_for_write_queue_from_safe(skb, next, sk) { 21405d424d5aSJohn Heffner copy = min_t(int, skb->len, probe_size - len); 21412fe664f1SDouglas Caetano dos Santos if (nskb->ip_summed) { 21425d424d5aSJohn Heffner skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); 21432fe664f1SDouglas Caetano dos Santos } else { 21442fe664f1SDouglas Caetano dos Santos __wsum csum = skb_copy_and_csum_bits(skb, 0, 2145056834d9SIlpo Järvinen skb_put(nskb, copy), 21462fe664f1SDouglas Caetano dos Santos copy, 0); 21472fe664f1SDouglas Caetano dos Santos nskb->csum = csum_block_add(nskb->csum, csum, len); 21482fe664f1SDouglas Caetano dos Santos } 21495d424d5aSJohn Heffner 21505d424d5aSJohn Heffner if (skb->len <= copy) { 21515d424d5aSJohn Heffner /* We've eaten all the data from this skb. 21525d424d5aSJohn Heffner * Throw it away. */ 21534de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 2154fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 21553ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 21565d424d5aSJohn Heffner } else { 21574de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & 2158a3433f35SChangli Gao ~(TCPHDR_FIN|TCPHDR_PSH); 21595d424d5aSJohn Heffner if (!skb_shinfo(skb)->nr_frags) { 21605d424d5aSJohn Heffner skb_pull(skb, copy); 216184fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 2162056834d9SIlpo Järvinen skb->csum = csum_partial(skb->data, 2163056834d9SIlpo Järvinen skb->len, 0); 21645d424d5aSJohn Heffner } else { 21655d424d5aSJohn Heffner __pskb_trim_head(skb, copy); 21665bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 21675d424d5aSJohn Heffner } 21685d424d5aSJohn Heffner TCP_SKB_CB(skb)->seq += copy; 21695d424d5aSJohn Heffner } 21705d424d5aSJohn Heffner 21715d424d5aSJohn Heffner len += copy; 2172234b6860SIlpo Järvinen 2173234b6860SIlpo Järvinen if (len >= probe_size) 2174234b6860SIlpo Järvinen break; 21755d424d5aSJohn Heffner } 21765bbb432cSEric Dumazet tcp_init_tso_segs(nskb, nskb->len); 21775d424d5aSJohn Heffner 21785d424d5aSJohn Heffner /* We're ready to send. If this fails, the probe will 21797faee5c0SEric Dumazet * be resegmented into mss-sized pieces by tcp_write_xmit(). 21807faee5c0SEric Dumazet */ 21815d424d5aSJohn Heffner if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 21825d424d5aSJohn Heffner /* Decrement cwnd here because we are sending 21835d424d5aSJohn Heffner * effectively two packets. */ 21845d424d5aSJohn Heffner tp->snd_cwnd--; 218566f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, nskb); 21865d424d5aSJohn Heffner 21875d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 21880e7b1368SJohn Heffner tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 21890e7b1368SJohn Heffner tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 21905d424d5aSJohn Heffner 21915d424d5aSJohn Heffner return 1; 21925d424d5aSJohn Heffner } 21935d424d5aSJohn Heffner 21945d424d5aSJohn Heffner return -1; 21955d424d5aSJohn Heffner } 21965d424d5aSJohn Heffner 2197218af599SEric Dumazet static bool tcp_pacing_check(const struct sock *sk) 2198218af599SEric Dumazet { 2199218af599SEric Dumazet return tcp_needs_internal_pacing(sk) && 2200218af599SEric Dumazet hrtimer_active(&tcp_sk(sk)->pacing_timer); 2201218af599SEric Dumazet } 2202218af599SEric Dumazet 2203f9616c35SEric Dumazet /* TCP Small Queues : 2204f9616c35SEric Dumazet * Control number of packets in qdisc/devices to two packets / or ~1 ms. 2205f9616c35SEric Dumazet * (These limits are doubled for retransmits) 2206f9616c35SEric Dumazet * This allows for : 2207f9616c35SEric Dumazet * - better RTT estimation and ACK scheduling 2208f9616c35SEric Dumazet * - faster recovery 2209f9616c35SEric Dumazet * - high rates 2210f9616c35SEric Dumazet * Alas, some drivers / subsystems require a fair amount 2211f9616c35SEric Dumazet * of queued bytes to ensure line rate. 2212f9616c35SEric Dumazet * One example is wifi aggregation (802.11 AMPDU) 2213f9616c35SEric Dumazet */ 2214f9616c35SEric Dumazet static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb, 2215f9616c35SEric Dumazet unsigned int factor) 2216f9616c35SEric Dumazet { 2217f9616c35SEric Dumazet unsigned int limit; 2218f9616c35SEric Dumazet 2219f9616c35SEric Dumazet limit = max(2 * skb->truesize, sk->sk_pacing_rate >> 10); 22209184d8bbSEric Dumazet limit = min_t(u32, limit, 22219184d8bbSEric Dumazet sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes); 2222f9616c35SEric Dumazet limit <<= factor; 2223f9616c35SEric Dumazet 222414afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) > limit) { 222575c119afSEric Dumazet /* Always send skb if rtx queue is empty. 222675eefc6cSEric Dumazet * No need to wait for TX completion to call us back, 222775eefc6cSEric Dumazet * after softirq/tasklet schedule. 222875eefc6cSEric Dumazet * This helps when TX completions are delayed too much. 222975eefc6cSEric Dumazet */ 223075c119afSEric Dumazet if (tcp_rtx_queue_empty(sk)) 223175eefc6cSEric Dumazet return false; 223275eefc6cSEric Dumazet 22337aa5470cSEric Dumazet set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags); 2234f9616c35SEric Dumazet /* It is possible TX completion already happened 2235f9616c35SEric Dumazet * before we set TSQ_THROTTLED, so we must 2236f9616c35SEric Dumazet * test again the condition. 2237f9616c35SEric Dumazet */ 2238f9616c35SEric Dumazet smp_mb__after_atomic(); 223914afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) > limit) 2240f9616c35SEric Dumazet return true; 2241f9616c35SEric Dumazet } 2242f9616c35SEric Dumazet return false; 2243f9616c35SEric Dumazet } 2244f9616c35SEric Dumazet 224505b055e8SFrancis Yan static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new) 224605b055e8SFrancis Yan { 2247628174ccSEric Dumazet const u32 now = tcp_jiffies32; 2248efe967cdSArnd Bergmann enum tcp_chrono old = tp->chrono_type; 224905b055e8SFrancis Yan 2250efe967cdSArnd Bergmann if (old > TCP_CHRONO_UNSPEC) 2251efe967cdSArnd Bergmann tp->chrono_stat[old - 1] += now - tp->chrono_start; 225205b055e8SFrancis Yan tp->chrono_start = now; 225305b055e8SFrancis Yan tp->chrono_type = new; 225405b055e8SFrancis Yan } 225505b055e8SFrancis Yan 225605b055e8SFrancis Yan void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type) 225705b055e8SFrancis Yan { 225805b055e8SFrancis Yan struct tcp_sock *tp = tcp_sk(sk); 225905b055e8SFrancis Yan 226005b055e8SFrancis Yan /* If there are multiple conditions worthy of tracking in a 22610f87230dSFrancis Yan * chronograph then the highest priority enum takes precedence 22620f87230dSFrancis Yan * over the other conditions. So that if something "more interesting" 226305b055e8SFrancis Yan * starts happening, stop the previous chrono and start a new one. 226405b055e8SFrancis Yan */ 226505b055e8SFrancis Yan if (type > tp->chrono_type) 226605b055e8SFrancis Yan tcp_chrono_set(tp, type); 226705b055e8SFrancis Yan } 226805b055e8SFrancis Yan 226905b055e8SFrancis Yan void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type) 227005b055e8SFrancis Yan { 227105b055e8SFrancis Yan struct tcp_sock *tp = tcp_sk(sk); 227205b055e8SFrancis Yan 22730f87230dSFrancis Yan 22740f87230dSFrancis Yan /* There are multiple conditions worthy of tracking in a 22750f87230dSFrancis Yan * chronograph, so that the highest priority enum takes 22760f87230dSFrancis Yan * precedence over the other conditions (see tcp_chrono_start). 22770f87230dSFrancis Yan * If a condition stops, we only stop chrono tracking if 22780f87230dSFrancis Yan * it's the "most interesting" or current chrono we are 22790f87230dSFrancis Yan * tracking and starts busy chrono if we have pending data. 22800f87230dSFrancis Yan */ 228175c119afSEric Dumazet if (tcp_rtx_and_write_queues_empty(sk)) 228205b055e8SFrancis Yan tcp_chrono_set(tp, TCP_CHRONO_UNSPEC); 22830f87230dSFrancis Yan else if (type == tp->chrono_type) 22840f87230dSFrancis Yan tcp_chrono_set(tp, TCP_CHRONO_BUSY); 228505b055e8SFrancis Yan } 228605b055e8SFrancis Yan 22871da177e4SLinus Torvalds /* This routine writes packets to the network. It advances the 22881da177e4SLinus Torvalds * send_head. This happens as incoming acks open up the remote 22891da177e4SLinus Torvalds * window for us. 22901da177e4SLinus Torvalds * 2291f8269a49SIlpo Järvinen * LARGESEND note: !tcp_urg_mode is overkill, only frames between 2292f8269a49SIlpo Järvinen * snd_up-64k-mss .. snd_up cannot be large. However, taking into 2293f8269a49SIlpo Järvinen * account rare use of URG, this is not a big flaw. 2294f8269a49SIlpo Järvinen * 22956ba8a3b1SNandita Dukkipati * Send at most one packet when push_one > 0. Temporarily ignore 22966ba8a3b1SNandita Dukkipati * cwnd limit to force at most one packet out when push_one == 2. 22976ba8a3b1SNandita Dukkipati 2298a2a385d6SEric Dumazet * Returns true, if no segments are in flight and we have queued segments, 2299a2a385d6SEric Dumazet * but cannot send anything now because of SWS or another problem. 23001da177e4SLinus Torvalds */ 2301a2a385d6SEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 2302d5dd9175SIlpo Järvinen int push_one, gfp_t gfp) 23031da177e4SLinus Torvalds { 23041da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 230592df7b51SDavid S. Miller struct sk_buff *skb; 2306c1b4a7e6SDavid S. Miller unsigned int tso_segs, sent_pkts; 2307c1b4a7e6SDavid S. Miller int cwnd_quota; 23085d424d5aSJohn Heffner int result; 23095615f886SFrancis Yan bool is_cwnd_limited = false, is_rwnd_limited = false; 2310605ad7f1SEric Dumazet u32 max_segs; 23111da177e4SLinus Torvalds 2312c1b4a7e6SDavid S. Miller sent_pkts = 0; 23135d424d5aSJohn Heffner 2314ee1836aeSEric Dumazet tcp_mstamp_refresh(tp); 2315d5dd9175SIlpo Järvinen if (!push_one) { 23165d424d5aSJohn Heffner /* Do MTU probing. */ 2317d5dd9175SIlpo Järvinen result = tcp_mtu_probe(sk); 2318d5dd9175SIlpo Järvinen if (!result) { 2319a2a385d6SEric Dumazet return false; 23205d424d5aSJohn Heffner } else if (result > 0) { 23215d424d5aSJohn Heffner sent_pkts = 1; 23225d424d5aSJohn Heffner } 2323d5dd9175SIlpo Järvinen } 23245d424d5aSJohn Heffner 2325ed6e7268SNeal Cardwell max_segs = tcp_tso_segs(sk, mss_now); 2326fe067e8aSDavid S. Miller while ((skb = tcp_send_head(sk))) { 2327c8ac3774SHerbert Xu unsigned int limit; 2328c8ac3774SHerbert Xu 2329218af599SEric Dumazet if (tcp_pacing_check(sk)) 2330218af599SEric Dumazet break; 2331218af599SEric Dumazet 23325bbb432cSEric Dumazet tso_segs = tcp_init_tso_segs(skb, mss_now); 2333c1b4a7e6SDavid S. Miller BUG_ON(!tso_segs); 2334c1b4a7e6SDavid S. Miller 23359d186cacSAndrey Vagin if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) { 23367faee5c0SEric Dumazet /* "skb_mstamp" is used as a start point for the retransmit timer */ 2337e2080072SEric Dumazet tcp_update_skb_after_send(tp, skb); 2338ec342325SAndrew Vagin goto repair; /* Skip network transmission */ 23399d186cacSAndrey Vagin } 2340ec342325SAndrew Vagin 2341b68e9f85SHerbert Xu cwnd_quota = tcp_cwnd_test(tp, skb); 23426ba8a3b1SNandita Dukkipati if (!cwnd_quota) { 23436ba8a3b1SNandita Dukkipati if (push_one == 2) 23446ba8a3b1SNandita Dukkipati /* Force out a loss probe pkt. */ 23456ba8a3b1SNandita Dukkipati cwnd_quota = 1; 23466ba8a3b1SNandita Dukkipati else 2347b68e9f85SHerbert Xu break; 23486ba8a3b1SNandita Dukkipati } 2349b68e9f85SHerbert Xu 23505615f886SFrancis Yan if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) { 23515615f886SFrancis Yan is_rwnd_limited = true; 2352b68e9f85SHerbert Xu break; 23535615f886SFrancis Yan } 2354b68e9f85SHerbert Xu 2355d6a4e26aSEric Dumazet if (tso_segs == 1) { 2356aa93466bSDavid S. Miller if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 2357aa93466bSDavid S. Miller (tcp_skb_is_last(sk, skb) ? 2358aa93466bSDavid S. Miller nonagle : TCP_NAGLE_PUSH)))) 2359aa93466bSDavid S. Miller break; 2360c1b4a7e6SDavid S. Miller } else { 2361ca8a2263SNeal Cardwell if (!push_one && 2362605ad7f1SEric Dumazet tcp_tso_should_defer(sk, skb, &is_cwnd_limited, 2363605ad7f1SEric Dumazet max_segs)) 2364aa93466bSDavid S. Miller break; 2365c1b4a7e6SDavid S. Miller } 2366aa93466bSDavid S. Miller 2367605ad7f1SEric Dumazet limit = mss_now; 2368d6a4e26aSEric Dumazet if (tso_segs > 1 && !tcp_urg_mode(tp)) 2369605ad7f1SEric Dumazet limit = tcp_mss_split_point(sk, skb, mss_now, 2370605ad7f1SEric Dumazet min_t(unsigned int, 2371605ad7f1SEric Dumazet cwnd_quota, 2372605ad7f1SEric Dumazet max_segs), 2373605ad7f1SEric Dumazet nonagle); 2374605ad7f1SEric Dumazet 2375605ad7f1SEric Dumazet if (skb->len > limit && 237675c119afSEric Dumazet unlikely(tso_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE, 237775c119afSEric Dumazet skb, limit, mss_now, gfp))) 2378605ad7f1SEric Dumazet break; 2379605ad7f1SEric Dumazet 23807aa5470cSEric Dumazet if (test_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags)) 23817aa5470cSEric Dumazet clear_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags); 2382f9616c35SEric Dumazet if (tcp_small_queue_check(sk, skb, 0)) 238346d3ceabSEric Dumazet break; 2384c9eeec26SEric Dumazet 2385d5dd9175SIlpo Järvinen if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) 23861da177e4SLinus Torvalds break; 23871da177e4SLinus Torvalds 2388ec342325SAndrew Vagin repair: 23891da177e4SLinus Torvalds /* Advance the send_head. This one is sent out. 23901da177e4SLinus Torvalds * This call will increment packets_out. 23911da177e4SLinus Torvalds */ 239266f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 23931da177e4SLinus Torvalds 23941da177e4SLinus Torvalds tcp_minshall_update(tp, mss_now, skb); 2395a262f0cdSNandita Dukkipati sent_pkts += tcp_skb_pcount(skb); 2396d5dd9175SIlpo Järvinen 2397d5dd9175SIlpo Järvinen if (push_one) 2398d5dd9175SIlpo Järvinen break; 23991da177e4SLinus Torvalds } 24001da177e4SLinus Torvalds 24015615f886SFrancis Yan if (is_rwnd_limited) 24025615f886SFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED); 24035615f886SFrancis Yan else 24045615f886SFrancis Yan tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED); 24055615f886SFrancis Yan 2406aa93466bSDavid S. Miller if (likely(sent_pkts)) { 2407684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 2408684bad11SYuchung Cheng tp->prr_out += sent_pkts; 24096ba8a3b1SNandita Dukkipati 24106ba8a3b1SNandita Dukkipati /* Send one loss probe per tail loss episode. */ 24116ba8a3b1SNandita Dukkipati if (push_one != 2) 24126ba8a3b1SNandita Dukkipati tcp_schedule_loss_probe(sk); 2413d2e1339fSBendik Rønning Opstad is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd); 2414ca8a2263SNeal Cardwell tcp_cwnd_validate(sk, is_cwnd_limited); 2415a2a385d6SEric Dumazet return false; 24161da177e4SLinus Torvalds } 241775c119afSEric Dumazet return !tp->packets_out && !tcp_write_queue_empty(sk); 24186ba8a3b1SNandita Dukkipati } 24196ba8a3b1SNandita Dukkipati 24206ba8a3b1SNandita Dukkipati bool tcp_schedule_loss_probe(struct sock *sk) 24216ba8a3b1SNandita Dukkipati { 24226ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 24236ba8a3b1SNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 2424a2815817SNeal Cardwell u32 timeout, rto_delta_us; 24252ae21cf5SEric Dumazet int early_retrans; 24266ba8a3b1SNandita Dukkipati 24276ba8a3b1SNandita Dukkipati /* Don't do any loss probe on a Fast Open connection before 3WHS 24286ba8a3b1SNandita Dukkipati * finishes. 24296ba8a3b1SNandita Dukkipati */ 2430f9b99582SYuchung Cheng if (tp->fastopen_rsk) 24316ba8a3b1SNandita Dukkipati return false; 24326ba8a3b1SNandita Dukkipati 24332ae21cf5SEric Dumazet early_retrans = sock_net(sk)->ipv4.sysctl_tcp_early_retrans; 24346ba8a3b1SNandita Dukkipati /* Schedule a loss probe in 2*RTT for SACK capable connections 24356ba8a3b1SNandita Dukkipati * in Open state, that are either limited by cwnd or application. 24366ba8a3b1SNandita Dukkipati */ 24372ae21cf5SEric Dumazet if ((early_retrans != 3 && early_retrans != 4) || 2438bec41a11SYuchung Cheng !tp->packets_out || !tcp_is_sack(tp) || 2439bec41a11SYuchung Cheng icsk->icsk_ca_state != TCP_CA_Open) 24406ba8a3b1SNandita Dukkipati return false; 24416ba8a3b1SNandita Dukkipati 24426ba8a3b1SNandita Dukkipati if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) && 244375c119afSEric Dumazet !tcp_write_queue_empty(sk)) 24446ba8a3b1SNandita Dukkipati return false; 24456ba8a3b1SNandita Dukkipati 2446bb4d991aSYuchung Cheng /* Probe timeout is 2*rtt. Add minimum RTO to account 2447f9b99582SYuchung Cheng * for delayed ack when there's one outstanding packet. If no RTT 2448f9b99582SYuchung Cheng * sample is available then probe after TCP_TIMEOUT_INIT. 24496ba8a3b1SNandita Dukkipati */ 2450bb4d991aSYuchung Cheng if (tp->srtt_us) { 2451bb4d991aSYuchung Cheng timeout = usecs_to_jiffies(tp->srtt_us >> 2); 24526ba8a3b1SNandita Dukkipati if (tp->packets_out == 1) 2453bb4d991aSYuchung Cheng timeout += TCP_RTO_MIN; 2454bb4d991aSYuchung Cheng else 2455bb4d991aSYuchung Cheng timeout += TCP_TIMEOUT_MIN; 2456bb4d991aSYuchung Cheng } else { 2457bb4d991aSYuchung Cheng timeout = TCP_TIMEOUT_INIT; 2458bb4d991aSYuchung Cheng } 24596ba8a3b1SNandita Dukkipati 2460a2815817SNeal Cardwell /* If the RTO formula yields an earlier time, then use that time. */ 2461a2815817SNeal Cardwell rto_delta_us = tcp_rto_delta_us(sk); /* How far in future is RTO? */ 2462a2815817SNeal Cardwell if (rto_delta_us > 0) 2463a2815817SNeal Cardwell timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us)); 24646ba8a3b1SNandita Dukkipati 24656ba8a3b1SNandita Dukkipati inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, 24666ba8a3b1SNandita Dukkipati TCP_RTO_MAX); 24676ba8a3b1SNandita Dukkipati return true; 24686ba8a3b1SNandita Dukkipati } 24696ba8a3b1SNandita Dukkipati 24701f3279aeSEric Dumazet /* Thanks to skb fast clones, we can detect if a prior transmit of 24711f3279aeSEric Dumazet * a packet is still in a qdisc or driver queue. 24721f3279aeSEric Dumazet * In this case, there is very little point doing a retransmit ! 24731f3279aeSEric Dumazet */ 24741f3279aeSEric Dumazet static bool skb_still_in_host_queue(const struct sock *sk, 24751f3279aeSEric Dumazet const struct sk_buff *skb) 24761f3279aeSEric Dumazet { 247739bb5e62SEric Dumazet if (unlikely(skb_fclone_busy(sk, skb))) { 2478c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), 24791f3279aeSEric Dumazet LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES); 24801f3279aeSEric Dumazet return true; 24811f3279aeSEric Dumazet } 24821f3279aeSEric Dumazet return false; 24831f3279aeSEric Dumazet } 24841f3279aeSEric Dumazet 2485b340b264SYuchung Cheng /* When probe timeout (PTO) fires, try send a new segment if possible, else 24866ba8a3b1SNandita Dukkipati * retransmit the last segment. 24876ba8a3b1SNandita Dukkipati */ 24886ba8a3b1SNandita Dukkipati void tcp_send_loss_probe(struct sock *sk) 24896ba8a3b1SNandita Dukkipati { 24909b717a8dSNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 24916ba8a3b1SNandita Dukkipati struct sk_buff *skb; 24926ba8a3b1SNandita Dukkipati int pcount; 24936ba8a3b1SNandita Dukkipati int mss = tcp_current_mss(sk); 24946ba8a3b1SNandita Dukkipati 2495b340b264SYuchung Cheng skb = tcp_send_head(sk); 249675c119afSEric Dumazet if (skb && tcp_snd_wnd_test(tp, skb, mss)) { 2497b340b264SYuchung Cheng pcount = tp->packets_out; 2498b340b264SYuchung Cheng tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC); 2499b340b264SYuchung Cheng if (tp->packets_out > pcount) 2500b340b264SYuchung Cheng goto probe_sent; 25016ba8a3b1SNandita Dukkipati goto rearm_timer; 25026ba8a3b1SNandita Dukkipati } 250375c119afSEric Dumazet skb = skb_rb_last(&sk->tcp_rtx_queue); 25046ba8a3b1SNandita Dukkipati 25059b717a8dSNandita Dukkipati /* At most one outstanding TLP retransmission. */ 25069b717a8dSNandita Dukkipati if (tp->tlp_high_seq) 25079b717a8dSNandita Dukkipati goto rearm_timer; 25089b717a8dSNandita Dukkipati 25096ba8a3b1SNandita Dukkipati /* Retransmit last segment. */ 25106ba8a3b1SNandita Dukkipati if (WARN_ON(!skb)) 25116ba8a3b1SNandita Dukkipati goto rearm_timer; 25126ba8a3b1SNandita Dukkipati 25131f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 25141f3279aeSEric Dumazet goto rearm_timer; 25151f3279aeSEric Dumazet 25166ba8a3b1SNandita Dukkipati pcount = tcp_skb_pcount(skb); 25176ba8a3b1SNandita Dukkipati if (WARN_ON(!pcount)) 25186ba8a3b1SNandita Dukkipati goto rearm_timer; 25196ba8a3b1SNandita Dukkipati 25206ba8a3b1SNandita Dukkipati if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) { 252175c119afSEric Dumazet if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, 252275c119afSEric Dumazet (pcount - 1) * mss, mss, 25236cc55e09SOctavian Purdila GFP_ATOMIC))) 25246ba8a3b1SNandita Dukkipati goto rearm_timer; 252575c119afSEric Dumazet skb = skb_rb_next(skb); 25266ba8a3b1SNandita Dukkipati } 25276ba8a3b1SNandita Dukkipati 25286ba8a3b1SNandita Dukkipati if (WARN_ON(!skb || !tcp_skb_pcount(skb))) 25296ba8a3b1SNandita Dukkipati goto rearm_timer; 25306ba8a3b1SNandita Dukkipati 253110d3be56SEric Dumazet if (__tcp_retransmit_skb(sk, skb, 1)) 2532b340b264SYuchung Cheng goto rearm_timer; 25336ba8a3b1SNandita Dukkipati 25349b717a8dSNandita Dukkipati /* Record snd_nxt for loss detection. */ 25359b717a8dSNandita Dukkipati tp->tlp_high_seq = tp->snd_nxt; 25369b717a8dSNandita Dukkipati 2537b340b264SYuchung Cheng probe_sent: 2538c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES); 2539fcd16c0aSYuchung Cheng /* Reset s.t. tcp_rearm_rto will restart timer from now */ 2540fcd16c0aSYuchung Cheng inet_csk(sk)->icsk_pending = 0; 2541b340b264SYuchung Cheng rearm_timer: 2542fcd16c0aSYuchung Cheng tcp_rearm_rto(sk); 25431da177e4SLinus Torvalds } 25441da177e4SLinus Torvalds 2545a762a980SDavid S. Miller /* Push out any pending frames which were held back due to 2546a762a980SDavid S. Miller * TCP_CORK or attempt at coalescing tiny packets. 2547a762a980SDavid S. Miller * The socket must be locked by the caller. 2548a762a980SDavid S. Miller */ 25499e412ba7SIlpo Järvinen void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 25509e412ba7SIlpo Järvinen int nonagle) 2551a762a980SDavid S. Miller { 2552726e07a8SIlpo Järvinen /* If we are closed, the bytes will have to remain here. 2553726e07a8SIlpo Järvinen * In time closedown will finish, we empty the write queue and 2554726e07a8SIlpo Järvinen * all will be happy. 2555726e07a8SIlpo Järvinen */ 2556726e07a8SIlpo Järvinen if (unlikely(sk->sk_state == TCP_CLOSE)) 2557726e07a8SIlpo Järvinen return; 2558726e07a8SIlpo Järvinen 255999a1dec7SMel Gorman if (tcp_write_xmit(sk, cur_mss, nonagle, 0, 25607450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC))) 25619e412ba7SIlpo Järvinen tcp_check_probe_timer(sk); 2562a762a980SDavid S. Miller } 2563a762a980SDavid S. Miller 2564c1b4a7e6SDavid S. Miller /* Send _single_ skb sitting at the send head. This function requires 2565c1b4a7e6SDavid S. Miller * true push pending frames to setup probe timer etc. 2566c1b4a7e6SDavid S. Miller */ 2567c1b4a7e6SDavid S. Miller void tcp_push_one(struct sock *sk, unsigned int mss_now) 2568c1b4a7e6SDavid S. Miller { 2569fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 2570c1b4a7e6SDavid S. Miller 2571c1b4a7e6SDavid S. Miller BUG_ON(!skb || skb->len < mss_now); 2572c1b4a7e6SDavid S. Miller 2573d5dd9175SIlpo Järvinen tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); 2574c1b4a7e6SDavid S. Miller } 2575c1b4a7e6SDavid S. Miller 25761da177e4SLinus Torvalds /* This function returns the amount that we can raise the 25771da177e4SLinus Torvalds * usable window based on the following constraints 25781da177e4SLinus Torvalds * 25791da177e4SLinus Torvalds * 1. The window can never be shrunk once it is offered (RFC 793) 25801da177e4SLinus Torvalds * 2. We limit memory per socket 25811da177e4SLinus Torvalds * 25821da177e4SLinus Torvalds * RFC 1122: 25831da177e4SLinus Torvalds * "the suggested [SWS] avoidance algorithm for the receiver is to keep 25841da177e4SLinus Torvalds * RECV.NEXT + RCV.WIN fixed until: 25851da177e4SLinus Torvalds * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 25861da177e4SLinus Torvalds * 25871da177e4SLinus Torvalds * i.e. don't raise the right edge of the window until you can raise 25881da177e4SLinus Torvalds * it at least MSS bytes. 25891da177e4SLinus Torvalds * 25901da177e4SLinus Torvalds * Unfortunately, the recommended algorithm breaks header prediction, 25911da177e4SLinus Torvalds * since header prediction assumes th->window stays fixed. 25921da177e4SLinus Torvalds * 25931da177e4SLinus Torvalds * Strictly speaking, keeping th->window fixed violates the receiver 25941da177e4SLinus Torvalds * side SWS prevention criteria. The problem is that under this rule 25951da177e4SLinus Torvalds * a stream of single byte packets will cause the right side of the 25961da177e4SLinus Torvalds * window to always advance by a single byte. 25971da177e4SLinus Torvalds * 25981da177e4SLinus Torvalds * Of course, if the sender implements sender side SWS prevention 25991da177e4SLinus Torvalds * then this will not be a problem. 26001da177e4SLinus Torvalds * 26011da177e4SLinus Torvalds * BSD seems to make the following compromise: 26021da177e4SLinus Torvalds * 26031da177e4SLinus Torvalds * If the free space is less than the 1/4 of the maximum 26041da177e4SLinus Torvalds * space available and the free space is less than 1/2 mss, 26051da177e4SLinus Torvalds * then set the window to 0. 26061da177e4SLinus Torvalds * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 26071da177e4SLinus Torvalds * Otherwise, just prevent the window from shrinking 26081da177e4SLinus Torvalds * and from being larger than the largest representable value. 26091da177e4SLinus Torvalds * 26101da177e4SLinus Torvalds * This prevents incremental opening of the window in the regime 26111da177e4SLinus Torvalds * where TCP is limited by the speed of the reader side taking 26121da177e4SLinus Torvalds * data out of the TCP receive queue. It does nothing about 26131da177e4SLinus Torvalds * those cases where the window is constrained on the sender side 26141da177e4SLinus Torvalds * because the pipeline is full. 26151da177e4SLinus Torvalds * 26161da177e4SLinus Torvalds * BSD also seems to "accidentally" limit itself to windows that are a 26171da177e4SLinus Torvalds * multiple of MSS, at least until the free space gets quite small. 26181da177e4SLinus Torvalds * This would appear to be a side effect of the mbuf implementation. 26191da177e4SLinus Torvalds * Combining these two algorithms results in the observed behavior 26201da177e4SLinus Torvalds * of having a fixed window size at almost all times. 26211da177e4SLinus Torvalds * 26221da177e4SLinus Torvalds * Below we obtain similar behavior by forcing the offered window to 26231da177e4SLinus Torvalds * a multiple of the mss when it is feasible to do so. 26241da177e4SLinus Torvalds * 26251da177e4SLinus Torvalds * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 26261da177e4SLinus Torvalds * Regular options like TIMESTAMP are taken into account. 26271da177e4SLinus Torvalds */ 26281da177e4SLinus Torvalds u32 __tcp_select_window(struct sock *sk) 26291da177e4SLinus Torvalds { 2630463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 26311da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2632caa20d9aSStephen Hemminger /* MSS for the peer's data. Previous versions used mss_clamp 26331da177e4SLinus Torvalds * here. I don't know if the value based on our guesses 26341da177e4SLinus Torvalds * of peer's MSS is better for the performance. It's more correct 26351da177e4SLinus Torvalds * but may be worse for the performance because of rcv_mss 26361da177e4SLinus Torvalds * fluctuations. --SAW 1998/11/1 26371da177e4SLinus Torvalds */ 2638463c84b9SArnaldo Carvalho de Melo int mss = icsk->icsk_ack.rcv_mss; 26391da177e4SLinus Torvalds int free_space = tcp_space(sk); 264086c1a045SFlorian Westphal int allowed_space = tcp_full_space(sk); 264186c1a045SFlorian Westphal int full_space = min_t(int, tp->window_clamp, allowed_space); 26421da177e4SLinus Torvalds int window; 26431da177e4SLinus Torvalds 264406425c30SEric Dumazet if (unlikely(mss > full_space)) { 26451da177e4SLinus Torvalds mss = full_space; 264606425c30SEric Dumazet if (mss <= 0) 264706425c30SEric Dumazet return 0; 264806425c30SEric Dumazet } 2649b92edbe0SEric Dumazet if (free_space < (full_space >> 1)) { 2650463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.quick = 0; 26511da177e4SLinus Torvalds 2652b8da51ebSEric Dumazet if (tcp_under_memory_pressure(sk)) 2653056834d9SIlpo Järvinen tp->rcv_ssthresh = min(tp->rcv_ssthresh, 2654056834d9SIlpo Järvinen 4U * tp->advmss); 26551da177e4SLinus Torvalds 265686c1a045SFlorian Westphal /* free_space might become our new window, make sure we don't 265786c1a045SFlorian Westphal * increase it due to wscale. 265886c1a045SFlorian Westphal */ 265986c1a045SFlorian Westphal free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale); 266086c1a045SFlorian Westphal 266186c1a045SFlorian Westphal /* if free space is less than mss estimate, or is below 1/16th 266286c1a045SFlorian Westphal * of the maximum allowed, try to move to zero-window, else 266386c1a045SFlorian Westphal * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and 266486c1a045SFlorian Westphal * new incoming data is dropped due to memory limits. 266586c1a045SFlorian Westphal * With large window, mss test triggers way too late in order 266686c1a045SFlorian Westphal * to announce zero window in time before rmem limit kicks in. 266786c1a045SFlorian Westphal */ 266886c1a045SFlorian Westphal if (free_space < (allowed_space >> 4) || free_space < mss) 26691da177e4SLinus Torvalds return 0; 26701da177e4SLinus Torvalds } 26711da177e4SLinus Torvalds 26721da177e4SLinus Torvalds if (free_space > tp->rcv_ssthresh) 26731da177e4SLinus Torvalds free_space = tp->rcv_ssthresh; 26741da177e4SLinus Torvalds 26751da177e4SLinus Torvalds /* Don't do rounding if we are using window scaling, since the 26761da177e4SLinus Torvalds * scaled window will not line up with the MSS boundary anyway. 26771da177e4SLinus Torvalds */ 26781da177e4SLinus Torvalds if (tp->rx_opt.rcv_wscale) { 26791da177e4SLinus Torvalds window = free_space; 26801da177e4SLinus Torvalds 26811da177e4SLinus Torvalds /* Advertise enough space so that it won't get scaled away. 26821da177e4SLinus Torvalds * Import case: prevent zero window announcement if 26831da177e4SLinus Torvalds * 1<<rcv_wscale > mss. 26841da177e4SLinus Torvalds */ 26851935299dSGao Feng window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale)); 26861da177e4SLinus Torvalds } else { 26871935299dSGao Feng window = tp->rcv_wnd; 26881da177e4SLinus Torvalds /* Get the largest window that is a nice multiple of mss. 26891da177e4SLinus Torvalds * Window clamp already applied above. 26901da177e4SLinus Torvalds * If our current window offering is within 1 mss of the 26911da177e4SLinus Torvalds * free space we just keep it. This prevents the divide 26921da177e4SLinus Torvalds * and multiply from happening most of the time. 26931da177e4SLinus Torvalds * We also don't do any window rounding when the free space 26941da177e4SLinus Torvalds * is too small. 26951da177e4SLinus Torvalds */ 26961da177e4SLinus Torvalds if (window <= free_space - mss || window > free_space) 26971935299dSGao Feng window = rounddown(free_space, mss); 269884565070SJohn Heffner else if (mss == full_space && 2699b92edbe0SEric Dumazet free_space > window + (full_space >> 1)) 270084565070SJohn Heffner window = free_space; 27011da177e4SLinus Torvalds } 27021da177e4SLinus Torvalds 27031da177e4SLinus Torvalds return window; 27041da177e4SLinus Torvalds } 27051da177e4SLinus Torvalds 2706cfea5a68SMartin KaFai Lau void tcp_skb_collapse_tstamp(struct sk_buff *skb, 2707082ac2d5SMartin KaFai Lau const struct sk_buff *next_skb) 2708082ac2d5SMartin KaFai Lau { 27090a2cf20cSSoheil Hassas Yeganeh if (unlikely(tcp_has_tx_tstamp(next_skb))) { 27100a2cf20cSSoheil Hassas Yeganeh const struct skb_shared_info *next_shinfo = 27110a2cf20cSSoheil Hassas Yeganeh skb_shinfo(next_skb); 2712082ac2d5SMartin KaFai Lau struct skb_shared_info *shinfo = skb_shinfo(skb); 2713082ac2d5SMartin KaFai Lau 27140a2cf20cSSoheil Hassas Yeganeh shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP; 2715082ac2d5SMartin KaFai Lau shinfo->tskey = next_shinfo->tskey; 27162de8023eSMartin KaFai Lau TCP_SKB_CB(skb)->txstamp_ack |= 27172de8023eSMartin KaFai Lau TCP_SKB_CB(next_skb)->txstamp_ack; 2718082ac2d5SMartin KaFai Lau } 2719082ac2d5SMartin KaFai Lau } 2720082ac2d5SMartin KaFai Lau 27214a17fc3aSIlpo Järvinen /* Collapses two adjacent SKB's during retransmission. */ 2722f8071cdeSEric Dumazet static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) 27231da177e4SLinus Torvalds { 27241da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 272575c119afSEric Dumazet struct sk_buff *next_skb = skb_rb_next(skb); 2726058dc334SIlpo Järvinen int skb_size, next_skb_size; 27271da177e4SLinus Torvalds 2728058dc334SIlpo Järvinen skb_size = skb->len; 2729058dc334SIlpo Järvinen next_skb_size = next_skb->len; 27301da177e4SLinus Torvalds 2731058dc334SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); 27321da177e4SLinus Torvalds 2733f8071cdeSEric Dumazet if (next_skb_size) { 2734f8071cdeSEric Dumazet if (next_skb_size <= skb_availroom(skb)) 2735f8071cdeSEric Dumazet skb_copy_bits(next_skb, 0, skb_put(skb, next_skb_size), 2736f8071cdeSEric Dumazet next_skb_size); 2737f8071cdeSEric Dumazet else if (!skb_shift(skb, next_skb, next_skb_size)) 2738f8071cdeSEric Dumazet return false; 2739f8071cdeSEric Dumazet } 27402b7cda9cSEric Dumazet tcp_highest_sack_replace(sk, next_skb, skb); 2741a6963a6bSIlpo Järvinen 274252d570aaSJarek Poplawski if (next_skb->ip_summed == CHECKSUM_PARTIAL) 274352d570aaSJarek Poplawski skb->ip_summed = CHECKSUM_PARTIAL; 27441da177e4SLinus Torvalds 274584fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 27461da177e4SLinus Torvalds skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); 27471da177e4SLinus Torvalds 27481da177e4SLinus Torvalds /* Update sequence range on original skb. */ 27491da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 27501da177e4SLinus Torvalds 2751e6c7d085SIlpo Järvinen /* Merge over control information. This moves PSH/FIN etc. over */ 27524de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; 27531da177e4SLinus Torvalds 27541da177e4SLinus Torvalds /* All done, get rid of second SKB and account for it so 27551da177e4SLinus Torvalds * packet counting does not break. 27561da177e4SLinus Torvalds */ 27574828e7f4SIlpo Järvinen TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; 2758a643b5d4SMartin KaFai Lau TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor; 2759b7689205SIlpo Järvinen 2760b7689205SIlpo Järvinen /* changed transmit queue under us so clear hints */ 2761ef9da47cSIlpo Järvinen tcp_clear_retrans_hints_partial(tp); 2762ef9da47cSIlpo Järvinen if (next_skb == tp->retransmit_skb_hint) 2763ef9da47cSIlpo Järvinen tp->retransmit_skb_hint = skb; 2764b7689205SIlpo Järvinen 2765797108d1SIlpo Järvinen tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); 2766797108d1SIlpo Järvinen 2767082ac2d5SMartin KaFai Lau tcp_skb_collapse_tstamp(skb, next_skb); 2768082ac2d5SMartin KaFai Lau 276975c119afSEric Dumazet tcp_rtx_queue_unlink_and_free(next_skb, sk); 2770f8071cdeSEric Dumazet return true; 27711da177e4SLinus Torvalds } 27721da177e4SLinus Torvalds 277367edfef7SAndi Kleen /* Check if coalescing SKBs is legal. */ 2774a2a385d6SEric Dumazet static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) 27754a17fc3aSIlpo Järvinen { 27764a17fc3aSIlpo Järvinen if (tcp_skb_pcount(skb) > 1) 2777a2a385d6SEric Dumazet return false; 27784a17fc3aSIlpo Järvinen if (skb_cloned(skb)) 2779a2a385d6SEric Dumazet return false; 27802331ccc5SEric Dumazet /* Some heuristics for collapsing over SACK'd could be invented */ 27814a17fc3aSIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 2782a2a385d6SEric Dumazet return false; 27834a17fc3aSIlpo Järvinen 2784a2a385d6SEric Dumazet return true; 27854a17fc3aSIlpo Järvinen } 27864a17fc3aSIlpo Järvinen 278767edfef7SAndi Kleen /* Collapse packets in the retransmit queue to make to create 278867edfef7SAndi Kleen * less packets on the wire. This is only done on retransmission. 278967edfef7SAndi Kleen */ 27904a17fc3aSIlpo Järvinen static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, 27914a17fc3aSIlpo Järvinen int space) 27924a17fc3aSIlpo Järvinen { 27934a17fc3aSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 27944a17fc3aSIlpo Järvinen struct sk_buff *skb = to, *tmp; 2795a2a385d6SEric Dumazet bool first = true; 27964a17fc3aSIlpo Järvinen 2797e0a1e5b5SEric Dumazet if (!sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse) 27984a17fc3aSIlpo Järvinen return; 27994de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 28004a17fc3aSIlpo Järvinen return; 28014a17fc3aSIlpo Järvinen 280275c119afSEric Dumazet skb_rbtree_walk_from_safe(skb, tmp) { 28034a17fc3aSIlpo Järvinen if (!tcp_can_collapse(sk, skb)) 28044a17fc3aSIlpo Järvinen break; 28054a17fc3aSIlpo Järvinen 2806a643b5d4SMartin KaFai Lau if (!tcp_skb_can_collapse_to(to)) 2807a643b5d4SMartin KaFai Lau break; 2808a643b5d4SMartin KaFai Lau 28094a17fc3aSIlpo Järvinen space -= skb->len; 28104a17fc3aSIlpo Järvinen 28114a17fc3aSIlpo Järvinen if (first) { 2812a2a385d6SEric Dumazet first = false; 28134a17fc3aSIlpo Järvinen continue; 28144a17fc3aSIlpo Järvinen } 28154a17fc3aSIlpo Järvinen 28164a17fc3aSIlpo Järvinen if (space < 0) 28174a17fc3aSIlpo Järvinen break; 28184a17fc3aSIlpo Järvinen 28194a17fc3aSIlpo Järvinen if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) 28204a17fc3aSIlpo Järvinen break; 28214a17fc3aSIlpo Järvinen 2822f8071cdeSEric Dumazet if (!tcp_collapse_retrans(sk, to)) 2823f8071cdeSEric Dumazet break; 28244a17fc3aSIlpo Järvinen } 28254a17fc3aSIlpo Järvinen } 28264a17fc3aSIlpo Järvinen 28271da177e4SLinus Torvalds /* This retransmits one SKB. Policy decisions and retransmit queue 28281da177e4SLinus Torvalds * state updates are done by the caller. Returns non-zero if an 28291da177e4SLinus Torvalds * error occurred which prevented the send. 28301da177e4SLinus Torvalds */ 283110d3be56SEric Dumazet int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) 28321da177e4SLinus Torvalds { 28335d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 283410d3be56SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 28357d227cd2SSridhar Samudrala unsigned int cur_mss; 283610d3be56SEric Dumazet int diff, len, err; 28371da177e4SLinus Torvalds 283810d3be56SEric Dumazet 283910d3be56SEric Dumazet /* Inconclusive MTU probe */ 284010d3be56SEric Dumazet if (icsk->icsk_mtup.probe_size) 28415d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 28425d424d5aSJohn Heffner 28431da177e4SLinus Torvalds /* Do not sent more than we queued. 1/4 is reserved for possible 2844caa20d9aSStephen Hemminger * copying overhead: fragmentation, tunneling, mangling etc. 28451da177e4SLinus Torvalds */ 284614afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) > 2847ffb4d6c8SEric Dumazet min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), 2848ffb4d6c8SEric Dumazet sk->sk_sndbuf)) 28491da177e4SLinus Torvalds return -EAGAIN; 28501da177e4SLinus Torvalds 28511f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 28521f3279aeSEric Dumazet return -EBUSY; 28531f3279aeSEric Dumazet 28541da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 28551da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 28561da177e4SLinus Torvalds BUG(); 28571da177e4SLinus Torvalds if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 28581da177e4SLinus Torvalds return -ENOMEM; 28591da177e4SLinus Torvalds } 28601da177e4SLinus Torvalds 28617d227cd2SSridhar Samudrala if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 28627d227cd2SSridhar Samudrala return -EHOSTUNREACH; /* Routing failure or similar. */ 28637d227cd2SSridhar Samudrala 28640c54b85fSIlpo Järvinen cur_mss = tcp_current_mss(sk); 28657d227cd2SSridhar Samudrala 28661da177e4SLinus Torvalds /* If receiver has shrunk his window, and skb is out of 28671da177e4SLinus Torvalds * new window, do not retransmit it. The exception is the 28681da177e4SLinus Torvalds * case, when window is shrunk to zero. In this case 28691da177e4SLinus Torvalds * our retransmit serves as a zero window probe. 28701da177e4SLinus Torvalds */ 28719d4fb27dSJoe Perches if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && 28729d4fb27dSJoe Perches TCP_SKB_CB(skb)->seq != tp->snd_una) 28731da177e4SLinus Torvalds return -EAGAIN; 28741da177e4SLinus Torvalds 287510d3be56SEric Dumazet len = cur_mss * segs; 287610d3be56SEric Dumazet if (skb->len > len) { 287775c119afSEric Dumazet if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len, 287875c119afSEric Dumazet cur_mss, GFP_ATOMIC)) 28791da177e4SLinus Torvalds return -ENOMEM; /* We'll try again later. */ 288002276f3cSIlpo Järvinen } else { 2881c52e2421SEric Dumazet if (skb_unclone(skb, GFP_ATOMIC)) 2882c52e2421SEric Dumazet return -ENOMEM; 288310d3be56SEric Dumazet 288410d3be56SEric Dumazet diff = tcp_skb_pcount(skb); 288510d3be56SEric Dumazet tcp_set_skb_tso_segs(skb, cur_mss); 288610d3be56SEric Dumazet diff -= tcp_skb_pcount(skb); 288710d3be56SEric Dumazet if (diff) 288810d3be56SEric Dumazet tcp_adjust_pcount(sk, skb, diff); 288910d3be56SEric Dumazet if (skb->len < cur_mss) 289010d3be56SEric Dumazet tcp_retrans_try_collapse(sk, skb, cur_mss); 28911da177e4SLinus Torvalds } 28921da177e4SLinus Torvalds 289349213555SDaniel Borkmann /* RFC3168, section 6.1.1.1. ECN fallback */ 289449213555SDaniel Borkmann if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN) 289549213555SDaniel Borkmann tcp_ecn_clear_syn(sk, skb); 289649213555SDaniel Borkmann 2897678550c6SYuchung Cheng /* Update global and local TCP statistics. */ 2898678550c6SYuchung Cheng segs = tcp_skb_pcount(skb); 2899678550c6SYuchung Cheng TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs); 2900678550c6SYuchung Cheng if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 2901678550c6SYuchung Cheng __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 2902678550c6SYuchung Cheng tp->total_retrans += segs; 2903678550c6SYuchung Cheng 290450bceae9SThomas Graf /* make sure skb->data is aligned on arches that require it 290550bceae9SThomas Graf * and check if ack-trimming & collapsing extended the headroom 290650bceae9SThomas Graf * beyond what csum_start can cover. 290750bceae9SThomas Graf */ 290850bceae9SThomas Graf if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) || 290950bceae9SThomas Graf skb_headroom(skb) >= 0xFFFF)) { 291010a81980SEric Dumazet struct sk_buff *nskb; 291110a81980SEric Dumazet 2912e2080072SEric Dumazet tcp_skb_tsorted_save(skb) { 291310a81980SEric Dumazet nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC); 2914c84a5711SYuchung Cheng err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) : 2915117632e6SEric Dumazet -ENOBUFS; 2916e2080072SEric Dumazet } tcp_skb_tsorted_restore(skb); 2917e2080072SEric Dumazet 29185889e2c0SYousuk Seung if (!err) { 2919e2080072SEric Dumazet tcp_update_skb_after_send(tp, skb); 29205889e2c0SYousuk Seung tcp_rate_skb_sent(sk, skb); 29215889e2c0SYousuk Seung } 2922117632e6SEric Dumazet } else { 2923c84a5711SYuchung Cheng err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 2924117632e6SEric Dumazet } 2925c84a5711SYuchung Cheng 2926fc9f3501SEric Dumazet if (likely(!err)) { 2927c84a5711SYuchung Cheng TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS; 2928e086101bSCong Wang trace_tcp_retransmit_skb(sk, skb); 2929678550c6SYuchung Cheng } else if (err != -EBUSY) { 2930678550c6SYuchung Cheng NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL); 2931fc9f3501SEric Dumazet } 2932c84a5711SYuchung Cheng return err; 293393b174adSYuchung Cheng } 293493b174adSYuchung Cheng 293510d3be56SEric Dumazet int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) 293693b174adSYuchung Cheng { 293793b174adSYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 293810d3be56SEric Dumazet int err = __tcp_retransmit_skb(sk, skb, segs); 29391da177e4SLinus Torvalds 29401da177e4SLinus Torvalds if (err == 0) { 29411da177e4SLinus Torvalds #if FASTRETRANS_DEBUG > 0 29421da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 2943e87cc472SJoe Perches net_dbg_ratelimited("retrans_out leaked\n"); 29441da177e4SLinus Torvalds } 29451da177e4SLinus Torvalds #endif 29461da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 29471da177e4SLinus Torvalds tp->retrans_out += tcp_skb_pcount(skb); 29481da177e4SLinus Torvalds 29491da177e4SLinus Torvalds /* Save stamp of the first retransmit. */ 29501da177e4SLinus Torvalds if (!tp->retrans_stamp) 29517faee5c0SEric Dumazet tp->retrans_stamp = tcp_skb_timestamp(skb); 29521da177e4SLinus Torvalds 29531da177e4SLinus Torvalds } 29546e08d5e3SYuchung Cheng 29556e08d5e3SYuchung Cheng if (tp->undo_retrans < 0) 29566e08d5e3SYuchung Cheng tp->undo_retrans = 0; 29576e08d5e3SYuchung Cheng tp->undo_retrans += tcp_skb_pcount(skb); 29581da177e4SLinus Torvalds return err; 29591da177e4SLinus Torvalds } 29601da177e4SLinus Torvalds 29611da177e4SLinus Torvalds /* This gets called after a retransmit timeout, and the initially 29621da177e4SLinus Torvalds * retransmitted data is acknowledged. It tries to continue 29631da177e4SLinus Torvalds * resending the rest of the retransmit queue, until either 29641da177e4SLinus Torvalds * we've sent it all or the congestion window limit is reached. 29651da177e4SLinus Torvalds * If doing SACK, the first ACK which comes back for a timeout 29661da177e4SLinus Torvalds * based retransmit packet might feed us FACK information again. 29671da177e4SLinus Torvalds * If so, we use it to avoid unnecessarily retransmissions. 29681da177e4SLinus Torvalds */ 29691da177e4SLinus Torvalds void tcp_xmit_retransmit_queue(struct sock *sk) 29701da177e4SLinus Torvalds { 29716687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 2972b9f1f1ceSEric Dumazet struct sk_buff *skb, *rtx_head, *hole = NULL; 29731da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2974840a3cbeSYuchung Cheng u32 max_segs; 297561eb55f4SIlpo Järvinen int mib_idx; 29766a438bbeSStephen Hemminger 297745e77d31SIlpo Järvinen if (!tp->packets_out) 297845e77d31SIlpo Järvinen return; 297945e77d31SIlpo Järvinen 298075c119afSEric Dumazet rtx_head = tcp_rtx_queue_head(sk); 2981b9f1f1ceSEric Dumazet skb = tp->retransmit_skb_hint ?: rtx_head; 2982ed6e7268SNeal Cardwell max_segs = tcp_tso_segs(sk, tcp_current_mss(sk)); 298375c119afSEric Dumazet skb_rbtree_walk_from(skb) { 2984dca0aaf8SEric Dumazet __u8 sacked; 298510d3be56SEric Dumazet int segs; 29861da177e4SLinus Torvalds 2987218af599SEric Dumazet if (tcp_pacing_check(sk)) 2988218af599SEric Dumazet break; 2989218af599SEric Dumazet 29906a438bbeSStephen Hemminger /* we could do better than to assign each time */ 299151456b29SIan Morris if (!hole) 29926a438bbeSStephen Hemminger tp->retransmit_skb_hint = skb; 29936a438bbeSStephen Hemminger 299410d3be56SEric Dumazet segs = tp->snd_cwnd - tcp_packets_in_flight(tp); 299510d3be56SEric Dumazet if (segs <= 0) 29961da177e4SLinus Torvalds return; 2997dca0aaf8SEric Dumazet sacked = TCP_SKB_CB(skb)->sacked; 2998a3d2e9f8SEric Dumazet /* In case tcp_shift_skb_data() have aggregated large skbs, 2999a3d2e9f8SEric Dumazet * we need to make sure not sending too bigs TSO packets 3000a3d2e9f8SEric Dumazet */ 3001a3d2e9f8SEric Dumazet segs = min_t(int, segs, max_segs); 30020e1c54c2SIlpo Järvinen 3003840a3cbeSYuchung Cheng if (tp->retrans_out >= tp->lost_out) { 3004006f582cSIlpo Järvinen break; 30050e1c54c2SIlpo Järvinen } else if (!(sacked & TCPCB_LOST)) { 300651456b29SIan Morris if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) 30070e1c54c2SIlpo Järvinen hole = skb; 300861eb55f4SIlpo Järvinen continue; 30091da177e4SLinus Torvalds 30100e1c54c2SIlpo Järvinen } else { 30110e1c54c2SIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Loss) 30120e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFASTRETRANS; 30130e1c54c2SIlpo Järvinen else 30140e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; 30150e1c54c2SIlpo Järvinen } 30160e1c54c2SIlpo Järvinen 30170e1c54c2SIlpo Järvinen if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) 301861eb55f4SIlpo Järvinen continue; 301940b215e5SPavel Emelyanov 3020f9616c35SEric Dumazet if (tcp_small_queue_check(sk, skb, 1)) 3021f9616c35SEric Dumazet return; 3022f9616c35SEric Dumazet 302310d3be56SEric Dumazet if (tcp_retransmit_skb(sk, skb, segs)) 30241da177e4SLinus Torvalds return; 302524ab6becSYuchung Cheng 3026de1d6578SYuchung Cheng NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb)); 30271da177e4SLinus Torvalds 3028684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 3029a262f0cdSNandita Dukkipati tp->prr_out += tcp_skb_pcount(skb); 3030a262f0cdSNandita Dukkipati 303175c119afSEric Dumazet if (skb == rtx_head && 303257dde7f7SYuchung Cheng icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT) 3033463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 30343f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, 30353f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 30361da177e4SLinus Torvalds } 30371da177e4SLinus Torvalds } 30381da177e4SLinus Torvalds 3039d83769a5SEric Dumazet /* We allow to exceed memory limits for FIN packets to expedite 3040d83769a5SEric Dumazet * connection tear down and (memory) recovery. 3041845704a5SEric Dumazet * Otherwise tcp_send_fin() could be tempted to either delay FIN 3042845704a5SEric Dumazet * or even be forced to close flow without any FIN. 3043a6c5ea4cSEric Dumazet * In general, we want to allow one skb per socket to avoid hangs 3044a6c5ea4cSEric Dumazet * with edge trigger epoll() 3045d83769a5SEric Dumazet */ 3046a6c5ea4cSEric Dumazet void sk_forced_mem_schedule(struct sock *sk, int size) 3047d83769a5SEric Dumazet { 3048e805605cSJohannes Weiner int amt; 3049d83769a5SEric Dumazet 3050d83769a5SEric Dumazet if (size <= sk->sk_forward_alloc) 3051d83769a5SEric Dumazet return; 3052d83769a5SEric Dumazet amt = sk_mem_pages(size); 3053d83769a5SEric Dumazet sk->sk_forward_alloc += amt * SK_MEM_QUANTUM; 3054e805605cSJohannes Weiner sk_memory_allocated_add(sk, amt); 3055e805605cSJohannes Weiner 3056baac50bbSJohannes Weiner if (mem_cgroup_sockets_enabled && sk->sk_memcg) 3057baac50bbSJohannes Weiner mem_cgroup_charge_skmem(sk->sk_memcg, amt); 3058d83769a5SEric Dumazet } 3059d83769a5SEric Dumazet 3060845704a5SEric Dumazet /* Send a FIN. The caller locks the socket for us. 3061845704a5SEric Dumazet * We should try to send a FIN packet really hard, but eventually give up. 30621da177e4SLinus Torvalds */ 30631da177e4SLinus Torvalds void tcp_send_fin(struct sock *sk) 30641da177e4SLinus Torvalds { 3065845704a5SEric Dumazet struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk); 30661da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 30671da177e4SLinus Torvalds 3068845704a5SEric Dumazet /* Optimization, tack on the FIN if we have one skb in write queue and 3069845704a5SEric Dumazet * this skb was not yet sent, or we are under memory pressure. 3070845704a5SEric Dumazet * Note: in the latter case, FIN packet will be sent after a timeout, 3071845704a5SEric Dumazet * as TCP stack thinks it has already been transmitted. 30721da177e4SLinus Torvalds */ 307375c119afSEric Dumazet if (!tskb && tcp_under_memory_pressure(sk)) 307475c119afSEric Dumazet tskb = skb_rb_last(&sk->tcp_rtx_queue); 307575c119afSEric Dumazet 307675c119afSEric Dumazet if (tskb) { 3077845704a5SEric Dumazet coalesce: 3078845704a5SEric Dumazet TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN; 3079845704a5SEric Dumazet TCP_SKB_CB(tskb)->end_seq++; 30801da177e4SLinus Torvalds tp->write_seq++; 308175c119afSEric Dumazet if (tcp_write_queue_empty(sk)) { 3082845704a5SEric Dumazet /* This means tskb was already sent. 3083845704a5SEric Dumazet * Pretend we included the FIN on previous transmit. 3084845704a5SEric Dumazet * We need to set tp->snd_nxt to the value it would have 3085845704a5SEric Dumazet * if FIN had been sent. This is because retransmit path 3086845704a5SEric Dumazet * does not change tp->snd_nxt. 3087845704a5SEric Dumazet */ 3088845704a5SEric Dumazet tp->snd_nxt++; 3089845704a5SEric Dumazet return; 3090845704a5SEric Dumazet } 30911da177e4SLinus Torvalds } else { 3092845704a5SEric Dumazet skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation); 3093845704a5SEric Dumazet if (unlikely(!skb)) { 3094845704a5SEric Dumazet if (tskb) 3095845704a5SEric Dumazet goto coalesce; 3096845704a5SEric Dumazet return; 30971da177e4SLinus Torvalds } 3098e2080072SEric Dumazet INIT_LIST_HEAD(&skb->tcp_tsorted_anchor); 3099d83769a5SEric Dumazet skb_reserve(skb, MAX_TCP_HEADER); 3100a6c5ea4cSEric Dumazet sk_forced_mem_schedule(sk, skb->truesize); 31011da177e4SLinus Torvalds /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 3102e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tp->write_seq, 3103a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_FIN); 31041da177e4SLinus Torvalds tcp_queue_skb(sk, skb); 31051da177e4SLinus Torvalds } 3106845704a5SEric Dumazet __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF); 31071da177e4SLinus Torvalds } 31081da177e4SLinus Torvalds 31091da177e4SLinus Torvalds /* We get here when a process closes a file descriptor (either due to 31101da177e4SLinus Torvalds * an explicit close() or as a byproduct of exit()'ing) and there 31111da177e4SLinus Torvalds * was unread data in the receive queue. This behavior is recommended 311265bb723cSGerrit Renker * by RFC 2525, section 2.17. -DaveM 31131da177e4SLinus Torvalds */ 3114dd0fc66fSAl Viro void tcp_send_active_reset(struct sock *sk, gfp_t priority) 31151da177e4SLinus Torvalds { 31161da177e4SLinus Torvalds struct sk_buff *skb; 31171da177e4SLinus Torvalds 31187cc2b043SGao Feng TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); 31197cc2b043SGao Feng 31201da177e4SLinus Torvalds /* NOTE: No TCP options attached and we never retransmit this. */ 31211da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, priority); 31221da177e4SLinus Torvalds if (!skb) { 31234e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 31241da177e4SLinus Torvalds return; 31251da177e4SLinus Torvalds } 31261da177e4SLinus Torvalds 31271da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 31281da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 3129e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), 3130a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_RST); 31319a568de4SEric Dumazet tcp_mstamp_refresh(tcp_sk(sk)); 31321da177e4SLinus Torvalds /* Send it off. */ 3133dfb4b9dcSDavid S. Miller if (tcp_transmit_skb(sk, skb, 0, priority)) 31344e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 3135c24b14c4SSong Liu 3136c24b14c4SSong Liu /* skb of trace_tcp_send_reset() keeps the skb that caused RST, 3137c24b14c4SSong Liu * skb here is different to the troublesome skb, so use NULL 3138c24b14c4SSong Liu */ 3139c24b14c4SSong Liu trace_tcp_send_reset(sk, NULL); 31401da177e4SLinus Torvalds } 31411da177e4SLinus Torvalds 314267edfef7SAndi Kleen /* Send a crossed SYN-ACK during socket establishment. 314367edfef7SAndi Kleen * WARNING: This routine must only be called when we have already sent 31441da177e4SLinus Torvalds * a SYN packet that crossed the incoming SYN that caused this routine 31451da177e4SLinus Torvalds * to get called. If this assumption fails then the initial rcv_wnd 31461da177e4SLinus Torvalds * and rcv_wscale values will not be correct. 31471da177e4SLinus Torvalds */ 31481da177e4SLinus Torvalds int tcp_send_synack(struct sock *sk) 31491da177e4SLinus Torvalds { 31501da177e4SLinus Torvalds struct sk_buff *skb; 31511da177e4SLinus Torvalds 315275c119afSEric Dumazet skb = tcp_rtx_queue_head(sk); 315351456b29SIan Morris if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 315475c119afSEric Dumazet pr_err("%s: wrong queue state\n", __func__); 31551da177e4SLinus Torvalds return -EFAULT; 31561da177e4SLinus Torvalds } 31574de075e0SEric Dumazet if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { 31581da177e4SLinus Torvalds if (skb_cloned(skb)) { 3159e2080072SEric Dumazet struct sk_buff *nskb; 3160e2080072SEric Dumazet 3161e2080072SEric Dumazet tcp_skb_tsorted_save(skb) { 3162e2080072SEric Dumazet nskb = skb_copy(skb, GFP_ATOMIC); 3163e2080072SEric Dumazet } tcp_skb_tsorted_restore(skb); 316451456b29SIan Morris if (!nskb) 31651da177e4SLinus Torvalds return -ENOMEM; 3166e2080072SEric Dumazet INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor); 316775c119afSEric Dumazet tcp_rtx_queue_unlink_and_free(skb, sk); 3168f4a775d1SEric Dumazet __skb_header_release(nskb); 316975c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb); 31703ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 31713ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 31721da177e4SLinus Torvalds skb = nskb; 31731da177e4SLinus Torvalds } 31741da177e4SLinus Torvalds 31754de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; 3176735d3831SFlorian Westphal tcp_ecn_send_synack(sk, skb); 31771da177e4SLinus Torvalds } 3178dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 31791da177e4SLinus Torvalds } 31801da177e4SLinus Torvalds 31814aea39c1SEric Dumazet /** 31824aea39c1SEric Dumazet * tcp_make_synack - Prepare a SYN-ACK. 31834aea39c1SEric Dumazet * sk: listener socket 31844aea39c1SEric Dumazet * dst: dst entry attached to the SYNACK 31854aea39c1SEric Dumazet * req: request_sock pointer 31864aea39c1SEric Dumazet * 31874aea39c1SEric Dumazet * Allocate one skb and build a SYNACK packet. 31884aea39c1SEric Dumazet * @dst is consumed : Caller should not use it again. 31894aea39c1SEric Dumazet */ 31905d062de7SEric Dumazet struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst, 3191e6b4d113SWilliam Allen Simpson struct request_sock *req, 3192ca6fb065SEric Dumazet struct tcp_fastopen_cookie *foc, 3193b3d05147SEric Dumazet enum tcp_synack_type synack_type) 31941da177e4SLinus Torvalds { 31952e6599cbSArnaldo Carvalho de Melo struct inet_request_sock *ireq = inet_rsk(req); 31965d062de7SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 319780f03e27SEric Dumazet struct tcp_md5sig_key *md5 = NULL; 31985d062de7SEric Dumazet struct tcp_out_options opts; 31995d062de7SEric Dumazet struct sk_buff *skb; 3200bd0388aeSWilliam Allen Simpson int tcp_header_size; 32015d062de7SEric Dumazet struct tcphdr *th; 3202f5fff5dcSTom Quetchenbach int mss; 32031da177e4SLinus Torvalds 3204ca6fb065SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 32054aea39c1SEric Dumazet if (unlikely(!skb)) { 32064aea39c1SEric Dumazet dst_release(dst); 32071da177e4SLinus Torvalds return NULL; 32084aea39c1SEric Dumazet } 32091da177e4SLinus Torvalds /* Reserve space for headers. */ 32101da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 32111da177e4SLinus Torvalds 3212b3d05147SEric Dumazet switch (synack_type) { 3213b3d05147SEric Dumazet case TCP_SYNACK_NORMAL: 32149e17f8a4SEric Dumazet skb_set_owner_w(skb, req_to_sk(req)); 3215b3d05147SEric Dumazet break; 3216b3d05147SEric Dumazet case TCP_SYNACK_COOKIE: 3217b3d05147SEric Dumazet /* Under synflood, we do not attach skb to a socket, 3218b3d05147SEric Dumazet * to avoid false sharing. 3219b3d05147SEric Dumazet */ 3220b3d05147SEric Dumazet break; 3221b3d05147SEric Dumazet case TCP_SYNACK_FASTOPEN: 3222ca6fb065SEric Dumazet /* sk is a const pointer, because we want to express multiple 3223ca6fb065SEric Dumazet * cpu might call us concurrently. 3224ca6fb065SEric Dumazet * sk->sk_wmem_alloc in an atomic, we can promote to rw. 3225ca6fb065SEric Dumazet */ 3226ca6fb065SEric Dumazet skb_set_owner_w(skb, (struct sock *)sk); 3227b3d05147SEric Dumazet break; 3228ca6fb065SEric Dumazet } 32294aea39c1SEric Dumazet skb_dst_set(skb, dst); 32301da177e4SLinus Torvalds 32313541f9e8SEric Dumazet mss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 3232f5fff5dcSTom Quetchenbach 323333ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 32348b5f12d0SFlorian Westphal #ifdef CONFIG_SYN_COOKIES 32358b5f12d0SFlorian Westphal if (unlikely(req->cookie_ts)) 32369a568de4SEric Dumazet skb->skb_mstamp = cookie_init_timestamp(req); 32378b5f12d0SFlorian Westphal else 32388b5f12d0SFlorian Westphal #endif 32399a568de4SEric Dumazet skb->skb_mstamp = tcp_clock_us(); 324080f03e27SEric Dumazet 324180f03e27SEric Dumazet #ifdef CONFIG_TCP_MD5SIG 324280f03e27SEric Dumazet rcu_read_lock(); 3243fd3a154aSEric Dumazet md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req)); 324480f03e27SEric Dumazet #endif 324558d607d3SEric Dumazet skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4); 324660e2a778SUrsula Braun tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5, 324760e2a778SUrsula Braun foc) + sizeof(*th); 324833ad798cSAdam Langley 3249aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 3250aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 32511da177e4SLinus Torvalds 3252ea1627c2SEric Dumazet th = (struct tcphdr *)skb->data; 32531da177e4SLinus Torvalds memset(th, 0, sizeof(struct tcphdr)); 32541da177e4SLinus Torvalds th->syn = 1; 32551da177e4SLinus Torvalds th->ack = 1; 32566ac705b1SEric Dumazet tcp_ecn_make_synack(req, th); 3257b44084c2SEric Dumazet th->source = htons(ireq->ir_num); 3258634fb979SEric Dumazet th->dest = ireq->ir_rmt_port; 3259e05a90ecSJamal Hadi Salim skb->mark = ireq->ir_mark; 3260e870a8efSIlpo Järvinen /* Setting of flags are superfluous here for callers (and ECE is 3261e870a8efSIlpo Järvinen * not even correctly set) 3262e870a8efSIlpo Järvinen */ 3263e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn, 3264a3433f35SChangli Gao TCPHDR_SYN | TCPHDR_ACK); 32654957faadSWilliam Allen Simpson 32661da177e4SLinus Torvalds th->seq = htonl(TCP_SKB_CB(skb)->seq); 32678336886fSJerry Chu /* XXX data is queued and acked as is. No buffer/window check */ 32688336886fSJerry Chu th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt); 32691da177e4SLinus Torvalds 32701da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 3271ed53d0abSEric Dumazet th->window = htons(min(req->rsk_rcv_wnd, 65535U)); 32725d062de7SEric Dumazet tcp_options_write((__be32 *)(th + 1), NULL, &opts); 32731da177e4SLinus Torvalds th->doff = (tcp_header_size >> 2); 327490bbcc60SEric Dumazet __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS); 3275cfb6eeb4SYOSHIFUJI Hideaki 3276cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 3277cfb6eeb4SYOSHIFUJI Hideaki /* Okay, we have all we need - do the md5 hash if needed */ 327880f03e27SEric Dumazet if (md5) 3279bd0388aeSWilliam Allen Simpson tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, 328039f8e58eSEric Dumazet md5, req_to_sk(req), skb); 328180f03e27SEric Dumazet rcu_read_unlock(); 3282cfb6eeb4SYOSHIFUJI Hideaki #endif 3283cfb6eeb4SYOSHIFUJI Hideaki 3284b50edd78SEric Dumazet /* Do not fool tcpdump (if any), clean our debris */ 32852456e855SThomas Gleixner skb->tstamp = 0; 32861da177e4SLinus Torvalds return skb; 32871da177e4SLinus Torvalds } 32884bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_make_synack); 32891da177e4SLinus Torvalds 329081164413SDaniel Borkmann static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst) 329181164413SDaniel Borkmann { 329281164413SDaniel Borkmann struct inet_connection_sock *icsk = inet_csk(sk); 329381164413SDaniel Borkmann const struct tcp_congestion_ops *ca; 329481164413SDaniel Borkmann u32 ca_key = dst_metric(dst, RTAX_CC_ALGO); 329581164413SDaniel Borkmann 329681164413SDaniel Borkmann if (ca_key == TCP_CA_UNSPEC) 329781164413SDaniel Borkmann return; 329881164413SDaniel Borkmann 329981164413SDaniel Borkmann rcu_read_lock(); 330081164413SDaniel Borkmann ca = tcp_ca_find_key(ca_key); 330181164413SDaniel Borkmann if (likely(ca && try_module_get(ca->owner))) { 330281164413SDaniel Borkmann module_put(icsk->icsk_ca_ops->owner); 330381164413SDaniel Borkmann icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst); 330481164413SDaniel Borkmann icsk->icsk_ca_ops = ca; 330581164413SDaniel Borkmann } 330681164413SDaniel Borkmann rcu_read_unlock(); 330781164413SDaniel Borkmann } 330881164413SDaniel Borkmann 330967edfef7SAndi Kleen /* Do all connect socket setups that can be done AF independent. */ 3310f7e56a76Sstephen hemminger static void tcp_connect_init(struct sock *sk) 33111da177e4SLinus Torvalds { 3312cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 33131da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 33141da177e4SLinus Torvalds __u8 rcv_wscale; 331513d3b1ebSLawrence Brakmo u32 rcv_wnd; 33161da177e4SLinus Torvalds 33171da177e4SLinus Torvalds /* We'll fix this up when we get a response from the other end. 33181da177e4SLinus Torvalds * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 33191da177e4SLinus Torvalds */ 33205d2ed052SEric Dumazet tp->tcp_header_len = sizeof(struct tcphdr); 33215d2ed052SEric Dumazet if (sock_net(sk)->ipv4.sysctl_tcp_timestamps) 33225d2ed052SEric Dumazet tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED; 33231da177e4SLinus Torvalds 3324cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 332500db4124SIan Morris if (tp->af_specific->md5_lookup(sk, sk)) 3326cfb6eeb4SYOSHIFUJI Hideaki tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 3327cfb6eeb4SYOSHIFUJI Hideaki #endif 3328cfb6eeb4SYOSHIFUJI Hideaki 33291da177e4SLinus Torvalds /* If user gave his TCP_MAXSEG, record it to clamp */ 33301da177e4SLinus Torvalds if (tp->rx_opt.user_mss) 33311da177e4SLinus Torvalds tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 33321da177e4SLinus Torvalds tp->max_window = 0; 33335d424d5aSJohn Heffner tcp_mtup_init(sk); 33341da177e4SLinus Torvalds tcp_sync_mss(sk, dst_mtu(dst)); 33351da177e4SLinus Torvalds 333681164413SDaniel Borkmann tcp_ca_dst_init(sk, dst); 333781164413SDaniel Borkmann 33381da177e4SLinus Torvalds if (!tp->window_clamp) 33391da177e4SLinus Torvalds tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 33403541f9e8SEric Dumazet tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 3341f5fff5dcSTom Quetchenbach 33421da177e4SLinus Torvalds tcp_initialize_rcv_mss(sk); 33431da177e4SLinus Torvalds 3344e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 3345e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 3346e88c64f0SHagen Paul Pfeifer (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) 3347e88c64f0SHagen Paul Pfeifer tp->window_clamp = tcp_full_space(sk); 3348e88c64f0SHagen Paul Pfeifer 334913d3b1ebSLawrence Brakmo rcv_wnd = tcp_rwnd_init_bpf(sk); 335013d3b1ebSLawrence Brakmo if (rcv_wnd == 0) 335113d3b1ebSLawrence Brakmo rcv_wnd = dst_metric(dst, RTAX_INITRWND); 335213d3b1ebSLawrence Brakmo 3353ceef9ab6SEric Dumazet tcp_select_initial_window(sk, tcp_full_space(sk), 33541da177e4SLinus Torvalds tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 33551da177e4SLinus Torvalds &tp->rcv_wnd, 33561da177e4SLinus Torvalds &tp->window_clamp, 33579bb37ef0SEric Dumazet sock_net(sk)->ipv4.sysctl_tcp_window_scaling, 335831d12926Slaurent chavey &rcv_wscale, 335913d3b1ebSLawrence Brakmo rcv_wnd); 33601da177e4SLinus Torvalds 33611da177e4SLinus Torvalds tp->rx_opt.rcv_wscale = rcv_wscale; 33621da177e4SLinus Torvalds tp->rcv_ssthresh = tp->rcv_wnd; 33631da177e4SLinus Torvalds 33641da177e4SLinus Torvalds sk->sk_err = 0; 33651da177e4SLinus Torvalds sock_reset_flag(sk, SOCK_DONE); 33661da177e4SLinus Torvalds tp->snd_wnd = 0; 3367ee7537b6SHantzis Fotis tcp_init_wl(tp, 0); 33681da177e4SLinus Torvalds tp->snd_una = tp->write_seq; 33691da177e4SLinus Torvalds tp->snd_sml = tp->write_seq; 337033f5f57eSIlpo Järvinen tp->snd_up = tp->write_seq; 3371370816aeSPavel Emelyanov tp->snd_nxt = tp->write_seq; 3372ee995283SPavel Emelyanov 3373ee995283SPavel Emelyanov if (likely(!tp->repair)) 33741da177e4SLinus Torvalds tp->rcv_nxt = 0; 3375c7781a6eSAndrew Vagin else 337670eabf0eSEric Dumazet tp->rcv_tstamp = tcp_jiffies32; 3377ee995283SPavel Emelyanov tp->rcv_wup = tp->rcv_nxt; 3378ee995283SPavel Emelyanov tp->copied_seq = tp->rcv_nxt; 33791da177e4SLinus Torvalds 33808550f328SLawrence Brakmo inet_csk(sk)->icsk_rto = tcp_timeout_init(sk); 3381463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_retransmits = 0; 33821da177e4SLinus Torvalds tcp_clear_retrans(tp); 33831da177e4SLinus Torvalds } 33841da177e4SLinus Torvalds 3385783237e8SYuchung Cheng static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) 3386783237e8SYuchung Cheng { 3387783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3388783237e8SYuchung Cheng struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 3389783237e8SYuchung Cheng 3390783237e8SYuchung Cheng tcb->end_seq += skb->len; 3391f4a775d1SEric Dumazet __skb_header_release(skb); 3392783237e8SYuchung Cheng sk->sk_wmem_queued += skb->truesize; 3393783237e8SYuchung Cheng sk_mem_charge(sk, skb->truesize); 3394783237e8SYuchung Cheng tp->write_seq = tcb->end_seq; 3395783237e8SYuchung Cheng tp->packets_out += tcp_skb_pcount(skb); 3396783237e8SYuchung Cheng } 3397783237e8SYuchung Cheng 3398783237e8SYuchung Cheng /* Build and send a SYN with data and (cached) Fast Open cookie. However, 3399783237e8SYuchung Cheng * queue a data-only packet after the regular SYN, such that regular SYNs 3400783237e8SYuchung Cheng * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges 3401783237e8SYuchung Cheng * only the SYN sequence, the data are retransmitted in the first ACK. 3402783237e8SYuchung Cheng * If cookie is not cached or other error occurs, falls back to send a 3403783237e8SYuchung Cheng * regular SYN with Fast Open cookie request option. 3404783237e8SYuchung Cheng */ 3405783237e8SYuchung Cheng static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) 3406783237e8SYuchung Cheng { 3407783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3408783237e8SYuchung Cheng struct tcp_fastopen_request *fo = tp->fastopen_req; 3409065263f4SWei Wang int space, err = 0; 3410355a901eSEric Dumazet struct sk_buff *syn_data; 3411783237e8SYuchung Cheng 341267da22d2SYuchung Cheng tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ 3413065263f4SWei Wang if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie)) 3414783237e8SYuchung Cheng goto fallback; 3415783237e8SYuchung Cheng 3416783237e8SYuchung Cheng /* MSS for SYN-data is based on cached MSS and bounded by PMTU and 3417783237e8SYuchung Cheng * user-MSS. Reserve maximum option space for middleboxes that add 3418783237e8SYuchung Cheng * private TCP options. The cost is reduced data space in SYN :( 3419783237e8SYuchung Cheng */ 34203541f9e8SEric Dumazet tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp); 34213541f9e8SEric Dumazet 34221b63edd6SYuchung Cheng space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) - 3423783237e8SYuchung Cheng MAX_TCP_OPTION_SPACE; 3424783237e8SYuchung Cheng 3425f5ddcbbbSEric Dumazet space = min_t(size_t, space, fo->size); 3426f5ddcbbbSEric Dumazet 3427f5ddcbbbSEric Dumazet /* limit to order-0 allocations */ 3428f5ddcbbbSEric Dumazet space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER)); 3429f5ddcbbbSEric Dumazet 3430eb934478SEric Dumazet syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false); 3431355a901eSEric Dumazet if (!syn_data) 3432783237e8SYuchung Cheng goto fallback; 3433355a901eSEric Dumazet syn_data->ip_summed = CHECKSUM_PARTIAL; 3434355a901eSEric Dumazet memcpy(syn_data->cb, syn->cb, sizeof(syn->cb)); 343507e100f9SEric Dumazet if (space) { 343607e100f9SEric Dumazet int copied = copy_from_iter(skb_put(syn_data, space), space, 343757be5bdaSAl Viro &fo->data->msg_iter); 343857be5bdaSAl Viro if (unlikely(!copied)) { 3439ba233b34SEric Dumazet tcp_skb_tsorted_anchor_cleanup(syn_data); 3440355a901eSEric Dumazet kfree_skb(syn_data); 3441783237e8SYuchung Cheng goto fallback; 3442783237e8SYuchung Cheng } 344357be5bdaSAl Viro if (copied != space) { 344457be5bdaSAl Viro skb_trim(syn_data, copied); 344557be5bdaSAl Viro space = copied; 344657be5bdaSAl Viro } 344707e100f9SEric Dumazet } 3448355a901eSEric Dumazet /* No more data pending in inet_wait_for_connect() */ 3449355a901eSEric Dumazet if (space == fo->size) 3450355a901eSEric Dumazet fo->data = NULL; 3451355a901eSEric Dumazet fo->copied = space; 3452783237e8SYuchung Cheng 3453355a901eSEric Dumazet tcp_connect_queue_skb(sk, syn_data); 34540f87230dSFrancis Yan if (syn_data->len) 34550f87230dSFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_BUSY); 3456355a901eSEric Dumazet 3457355a901eSEric Dumazet err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation); 3458355a901eSEric Dumazet 3459355a901eSEric Dumazet syn->skb_mstamp = syn_data->skb_mstamp; 3460355a901eSEric Dumazet 3461355a901eSEric Dumazet /* Now full SYN+DATA was cloned and sent (or not), 3462355a901eSEric Dumazet * remove the SYN from the original skb (syn_data) 3463355a901eSEric Dumazet * we keep in write queue in case of a retransmit, as we 3464355a901eSEric Dumazet * also have the SYN packet (with no data) in the same queue. 3465431a9124SEric Dumazet */ 3466355a901eSEric Dumazet TCP_SKB_CB(syn_data)->seq++; 3467355a901eSEric Dumazet TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH; 3468355a901eSEric Dumazet if (!err) { 346967da22d2SYuchung Cheng tp->syn_data = (fo->copied > 0); 347075c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data); 3471f19c29e3SYuchung Cheng NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT); 3472783237e8SYuchung Cheng goto done; 3473783237e8SYuchung Cheng } 3474783237e8SYuchung Cheng 347575c119afSEric Dumazet /* data was not sent, put it in write_queue */ 347675c119afSEric Dumazet __skb_queue_tail(&sk->sk_write_queue, syn_data); 3477b5b7db8dSEric Dumazet tp->packets_out -= tcp_skb_pcount(syn_data); 3478b5b7db8dSEric Dumazet 3479783237e8SYuchung Cheng fallback: 3480783237e8SYuchung Cheng /* Send a regular SYN with Fast Open cookie request option */ 3481783237e8SYuchung Cheng if (fo->cookie.len > 0) 3482783237e8SYuchung Cheng fo->cookie.len = 0; 3483783237e8SYuchung Cheng err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); 3484783237e8SYuchung Cheng if (err) 3485783237e8SYuchung Cheng tp->syn_fastopen = 0; 3486783237e8SYuchung Cheng done: 3487783237e8SYuchung Cheng fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ 3488783237e8SYuchung Cheng return err; 3489783237e8SYuchung Cheng } 3490783237e8SYuchung Cheng 349167edfef7SAndi Kleen /* Build a SYN and send it off. */ 34921da177e4SLinus Torvalds int tcp_connect(struct sock *sk) 34931da177e4SLinus Torvalds { 34941da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 34951da177e4SLinus Torvalds struct sk_buff *buff; 3496ee586811SEric Paris int err; 34971da177e4SLinus Torvalds 34989872a4bdSLawrence Brakmo tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB); 34998ba60924SEric Dumazet 35008ba60924SEric Dumazet if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 35018ba60924SEric Dumazet return -EHOSTUNREACH; /* Routing failure or similar. */ 35028ba60924SEric Dumazet 35031da177e4SLinus Torvalds tcp_connect_init(sk); 35041da177e4SLinus Torvalds 35052b916477SAndrey Vagin if (unlikely(tp->repair)) { 35062b916477SAndrey Vagin tcp_finish_connect(sk, NULL); 35072b916477SAndrey Vagin return 0; 35082b916477SAndrey Vagin } 35092b916477SAndrey Vagin 3510eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true); 3511355a901eSEric Dumazet if (unlikely(!buff)) 35121da177e4SLinus Torvalds return -ENOBUFS; 35131da177e4SLinus Torvalds 3514a3433f35SChangli Gao tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); 35159a568de4SEric Dumazet tcp_mstamp_refresh(tp); 35169a568de4SEric Dumazet tp->retrans_stamp = tcp_time_stamp(tp); 3517783237e8SYuchung Cheng tcp_connect_queue_skb(sk, buff); 3518735d3831SFlorian Westphal tcp_ecn_send_syn(sk, buff); 351975c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, buff); 35201da177e4SLinus Torvalds 3521783237e8SYuchung Cheng /* Send off SYN; include data in Fast Open. */ 3522783237e8SYuchung Cheng err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : 3523783237e8SYuchung Cheng tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); 3524ee586811SEric Paris if (err == -ECONNREFUSED) 3525ee586811SEric Paris return err; 3526bd37a088SWei Yongjun 3527bd37a088SWei Yongjun /* We change tp->snd_nxt after the tcp_transmit_skb() call 3528bd37a088SWei Yongjun * in order to make this packet get counted in tcpOutSegs. 3529bd37a088SWei Yongjun */ 3530bd37a088SWei Yongjun tp->snd_nxt = tp->write_seq; 3531bd37a088SWei Yongjun tp->pushed_seq = tp->write_seq; 3532b5b7db8dSEric Dumazet buff = tcp_send_head(sk); 3533b5b7db8dSEric Dumazet if (unlikely(buff)) { 3534b5b7db8dSEric Dumazet tp->snd_nxt = TCP_SKB_CB(buff)->seq; 3535b5b7db8dSEric Dumazet tp->pushed_seq = TCP_SKB_CB(buff)->seq; 3536b5b7db8dSEric Dumazet } 353781cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); 35381da177e4SLinus Torvalds 35391da177e4SLinus Torvalds /* Timer for repeating the SYN until an answer. */ 35403f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 35413f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 35421da177e4SLinus Torvalds return 0; 35431da177e4SLinus Torvalds } 35444bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_connect); 35451da177e4SLinus Torvalds 35461da177e4SLinus Torvalds /* Send out a delayed ack, the caller does the policy checking 35471da177e4SLinus Torvalds * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 35481da177e4SLinus Torvalds * for details. 35491da177e4SLinus Torvalds */ 35501da177e4SLinus Torvalds void tcp_send_delayed_ack(struct sock *sk) 35511da177e4SLinus Torvalds { 3552463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 3553463c84b9SArnaldo Carvalho de Melo int ato = icsk->icsk_ack.ato; 35541da177e4SLinus Torvalds unsigned long timeout; 35551da177e4SLinus Torvalds 35569890092eSFlorian Westphal tcp_ca_event(sk, CA_EVENT_DELAYED_ACK); 35579890092eSFlorian Westphal 35581da177e4SLinus Torvalds if (ato > TCP_DELACK_MIN) { 3559463c84b9SArnaldo Carvalho de Melo const struct tcp_sock *tp = tcp_sk(sk); 35601da177e4SLinus Torvalds int max_ato = HZ / 2; 35611da177e4SLinus Torvalds 3562056834d9SIlpo Järvinen if (icsk->icsk_ack.pingpong || 3563056834d9SIlpo Järvinen (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 35641da177e4SLinus Torvalds max_ato = TCP_DELACK_MAX; 35651da177e4SLinus Torvalds 35661da177e4SLinus Torvalds /* Slow path, intersegment interval is "high". */ 35671da177e4SLinus Torvalds 35681da177e4SLinus Torvalds /* If some rtt estimate is known, use it to bound delayed ack. 3569463c84b9SArnaldo Carvalho de Melo * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 35701da177e4SLinus Torvalds * directly. 35711da177e4SLinus Torvalds */ 3572740b0f18SEric Dumazet if (tp->srtt_us) { 3573740b0f18SEric Dumazet int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3), 3574740b0f18SEric Dumazet TCP_DELACK_MIN); 35751da177e4SLinus Torvalds 35761da177e4SLinus Torvalds if (rtt < max_ato) 35771da177e4SLinus Torvalds max_ato = rtt; 35781da177e4SLinus Torvalds } 35791da177e4SLinus Torvalds 35801da177e4SLinus Torvalds ato = min(ato, max_ato); 35811da177e4SLinus Torvalds } 35821da177e4SLinus Torvalds 35831da177e4SLinus Torvalds /* Stay within the limit we were given */ 35841da177e4SLinus Torvalds timeout = jiffies + ato; 35851da177e4SLinus Torvalds 35861da177e4SLinus Torvalds /* Use new timeout only if there wasn't a older one earlier. */ 3587463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 35881da177e4SLinus Torvalds /* If delack timer was blocked or is about to expire, 35891da177e4SLinus Torvalds * send ACK now. 35901da177e4SLinus Torvalds */ 3591463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.blocked || 3592463c84b9SArnaldo Carvalho de Melo time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 35931da177e4SLinus Torvalds tcp_send_ack(sk); 35941da177e4SLinus Torvalds return; 35951da177e4SLinus Torvalds } 35961da177e4SLinus Torvalds 3597463c84b9SArnaldo Carvalho de Melo if (!time_before(timeout, icsk->icsk_ack.timeout)) 3598463c84b9SArnaldo Carvalho de Melo timeout = icsk->icsk_ack.timeout; 35991da177e4SLinus Torvalds } 3600463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 3601463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.timeout = timeout; 3602463c84b9SArnaldo Carvalho de Melo sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 36031da177e4SLinus Torvalds } 36041da177e4SLinus Torvalds 36051da177e4SLinus Torvalds /* This routine sends an ack and also updates the window. */ 36061da177e4SLinus Torvalds void tcp_send_ack(struct sock *sk) 36071da177e4SLinus Torvalds { 36081da177e4SLinus Torvalds struct sk_buff *buff; 36091da177e4SLinus Torvalds 3610058dc334SIlpo Järvinen /* If we have been reset, we may not send again. */ 3611058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3612058dc334SIlpo Järvinen return; 3613058dc334SIlpo Järvinen 36149890092eSFlorian Westphal tcp_ca_event(sk, CA_EVENT_NON_DELAYED_ACK); 36159890092eSFlorian Westphal 36161da177e4SLinus Torvalds /* We are not putting this on the write queue, so 36171da177e4SLinus Torvalds * tcp_transmit_skb() will set the ownership to this 36181da177e4SLinus Torvalds * sock. 36191da177e4SLinus Torvalds */ 36207450aaf6SEric Dumazet buff = alloc_skb(MAX_TCP_HEADER, 36217450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 36227450aaf6SEric Dumazet if (unlikely(!buff)) { 3623463c84b9SArnaldo Carvalho de Melo inet_csk_schedule_ack(sk); 3624463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; 36253f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 36263f421baaSArnaldo Carvalho de Melo TCP_DELACK_MAX, TCP_RTO_MAX); 36271da177e4SLinus Torvalds return; 36281da177e4SLinus Torvalds } 36291da177e4SLinus Torvalds 36301da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 36311da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 3632a3433f35SChangli Gao tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); 36331da177e4SLinus Torvalds 363498781965SEric Dumazet /* We do not want pure acks influencing TCP Small Queues or fq/pacing 363598781965SEric Dumazet * too much. 363698781965SEric Dumazet * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784 363798781965SEric Dumazet */ 363898781965SEric Dumazet skb_set_tcp_pure_ack(buff); 363998781965SEric Dumazet 36401da177e4SLinus Torvalds /* Send it off, this clears delayed acks for us. */ 36417450aaf6SEric Dumazet tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0); 36421da177e4SLinus Torvalds } 3643e3118e83SDaniel Borkmann EXPORT_SYMBOL_GPL(tcp_send_ack); 36441da177e4SLinus Torvalds 36451da177e4SLinus Torvalds /* This routine sends a packet with an out of date sequence 36461da177e4SLinus Torvalds * number. It assumes the other end will try to ack it. 36471da177e4SLinus Torvalds * 36481da177e4SLinus Torvalds * Question: what should we make while urgent mode? 36491da177e4SLinus Torvalds * 4.4BSD forces sending single byte of data. We cannot send 36501da177e4SLinus Torvalds * out of window data, because we have SND.NXT==SND.MAX... 36511da177e4SLinus Torvalds * 36521da177e4SLinus Torvalds * Current solution: to send TWO zero-length segments in urgent mode: 36531da177e4SLinus Torvalds * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 36541da177e4SLinus Torvalds * out-of-date with SND.UNA-1 to probe window. 36551da177e4SLinus Torvalds */ 3656e520af48SEric Dumazet static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib) 36571da177e4SLinus Torvalds { 36581da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 36591da177e4SLinus Torvalds struct sk_buff *skb; 36601da177e4SLinus Torvalds 36611da177e4SLinus Torvalds /* We don't queue it, tcp_transmit_skb() sets ownership. */ 36627450aaf6SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, 36637450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 366451456b29SIan Morris if (!skb) 36651da177e4SLinus Torvalds return -1; 36661da177e4SLinus Torvalds 36671da177e4SLinus Torvalds /* Reserve space for headers and set control bits. */ 36681da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 36691da177e4SLinus Torvalds /* Use a previous sequence. This should cause the other 36701da177e4SLinus Torvalds * end to send an ack. Don't queue or clone SKB, just 36711da177e4SLinus Torvalds * send it. 36721da177e4SLinus Torvalds */ 3673a3433f35SChangli Gao tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); 3674e2e8009fSRenato Westphal NET_INC_STATS(sock_net(sk), mib); 36757450aaf6SEric Dumazet return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0); 36761da177e4SLinus Torvalds } 36771da177e4SLinus Torvalds 3678385e2070SEric Dumazet /* Called from setsockopt( ... TCP_REPAIR ) */ 3679ee995283SPavel Emelyanov void tcp_send_window_probe(struct sock *sk) 3680ee995283SPavel Emelyanov { 3681ee995283SPavel Emelyanov if (sk->sk_state == TCP_ESTABLISHED) { 3682ee995283SPavel Emelyanov tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; 36839a568de4SEric Dumazet tcp_mstamp_refresh(tcp_sk(sk)); 3684e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE); 3685ee995283SPavel Emelyanov } 3686ee995283SPavel Emelyanov } 3687ee995283SPavel Emelyanov 368867edfef7SAndi Kleen /* Initiate keepalive or window probe from timer. */ 3689e520af48SEric Dumazet int tcp_write_wakeup(struct sock *sk, int mib) 36901da177e4SLinus Torvalds { 36911da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 36921da177e4SLinus Torvalds struct sk_buff *skb; 36931da177e4SLinus Torvalds 3694058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3695058dc334SIlpo Järvinen return -1; 3696058dc334SIlpo Järvinen 369700db4124SIan Morris skb = tcp_send_head(sk); 369800db4124SIan Morris if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { 36991da177e4SLinus Torvalds int err; 37000c54b85fSIlpo Järvinen unsigned int mss = tcp_current_mss(sk); 370190840defSIlpo Järvinen unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 37021da177e4SLinus Torvalds 37031da177e4SLinus Torvalds if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 37041da177e4SLinus Torvalds tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 37051da177e4SLinus Torvalds 37061da177e4SLinus Torvalds /* We are probing the opening of a window 37071da177e4SLinus Torvalds * but the window size is != 0 37081da177e4SLinus Torvalds * must have been a result SWS avoidance ( sender ) 37091da177e4SLinus Torvalds */ 37101da177e4SLinus Torvalds if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 37111da177e4SLinus Torvalds skb->len > mss) { 37121da177e4SLinus Torvalds seg_size = min(seg_size, mss); 37134de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 371475c119afSEric Dumazet if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE, 371575c119afSEric Dumazet skb, seg_size, mss, GFP_ATOMIC)) 37161da177e4SLinus Torvalds return -1; 37171da177e4SLinus Torvalds } else if (!tcp_skb_pcount(skb)) 37185bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss); 37191da177e4SLinus Torvalds 37204de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 3721dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 372266f5fe62SIlpo Järvinen if (!err) 372366f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 37241da177e4SLinus Torvalds return err; 37251da177e4SLinus Torvalds } else { 372633f5f57eSIlpo Järvinen if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) 3727e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 1, mib); 3728e520af48SEric Dumazet return tcp_xmit_probe_skb(sk, 0, mib); 37291da177e4SLinus Torvalds } 37301da177e4SLinus Torvalds } 37311da177e4SLinus Torvalds 37321da177e4SLinus Torvalds /* A window probe timeout has occurred. If window is not closed send 37331da177e4SLinus Torvalds * a partial packet else a zero probe. 37341da177e4SLinus Torvalds */ 37351da177e4SLinus Torvalds void tcp_send_probe0(struct sock *sk) 37361da177e4SLinus Torvalds { 3737463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 37381da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 3739c6214a97SNikolay Borisov struct net *net = sock_net(sk); 3740fcdd1cf4SEric Dumazet unsigned long probe_max; 37411da177e4SLinus Torvalds int err; 37421da177e4SLinus Torvalds 3743e520af48SEric Dumazet err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE); 37441da177e4SLinus Torvalds 374575c119afSEric Dumazet if (tp->packets_out || tcp_write_queue_empty(sk)) { 37461da177e4SLinus Torvalds /* Cancel probe timer, if it is not required. */ 37476687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 0; 3748463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff = 0; 37491da177e4SLinus Torvalds return; 37501da177e4SLinus Torvalds } 37511da177e4SLinus Torvalds 37521da177e4SLinus Torvalds if (err <= 0) { 3753c6214a97SNikolay Borisov if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2) 3754463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff++; 37556687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out++; 3756fcdd1cf4SEric Dumazet probe_max = TCP_RTO_MAX; 37571da177e4SLinus Torvalds } else { 37581da177e4SLinus Torvalds /* If packet was not sent due to local congestion, 37596687e988SArnaldo Carvalho de Melo * do not backoff and do not remember icsk_probes_out. 37601da177e4SLinus Torvalds * Let local senders to fight for local resources. 37611da177e4SLinus Torvalds * 37621da177e4SLinus Torvalds * Use accumulated backoff yet. 37631da177e4SLinus Torvalds */ 37646687e988SArnaldo Carvalho de Melo if (!icsk->icsk_probes_out) 37656687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 1; 3766fcdd1cf4SEric Dumazet probe_max = TCP_RESOURCE_PROBE_INTERVAL; 37671da177e4SLinus Torvalds } 3768fcdd1cf4SEric Dumazet inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 376921c8fe99SEric Dumazet tcp_probe0_when(sk, probe_max), 3770fcdd1cf4SEric Dumazet TCP_RTO_MAX); 37711da177e4SLinus Torvalds } 37725db92c99SOctavian Purdila 3773ea3bea3aSEric Dumazet int tcp_rtx_synack(const struct sock *sk, struct request_sock *req) 37745db92c99SOctavian Purdila { 37755db92c99SOctavian Purdila const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific; 37765db92c99SOctavian Purdila struct flowi fl; 37775db92c99SOctavian Purdila int res; 37785db92c99SOctavian Purdila 377958d607d3SEric Dumazet tcp_rsk(req)->txhash = net_tx_rndhash(); 3780b3d05147SEric Dumazet res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL); 37815db92c99SOctavian Purdila if (!res) { 378290bbcc60SEric Dumazet __TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); 378302a1d6e7SEric Dumazet __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 37847e32b443SYuchung Cheng if (unlikely(tcp_passive_fastopen(sk))) 37857e32b443SYuchung Cheng tcp_sk(sk)->total_retrans++; 3786cf34ce3dSSong Liu trace_tcp_retransmit_synack(sk, req); 37875db92c99SOctavian Purdila } 37885db92c99SOctavian Purdila return res; 37895db92c99SOctavian Purdila } 37905db92c99SOctavian Purdila EXPORT_SYMBOL(tcp_rtx_synack); 3791