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> 4735089bb2SDavid S. Miller 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 */ 223356d1833SEric Dumazet space = max_t(u32, space, sock_net(sk)->ipv4.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 3854de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags; 386e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->sacked = 0; 387e870a8efSIlpo Järvinen 388cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, 1); 389e870a8efSIlpo Järvinen 390e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->seq = seq; 391a3433f35SChangli Gao if (flags & (TCPHDR_SYN | TCPHDR_FIN)) 392e870a8efSIlpo Järvinen seq++; 393e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->end_seq = seq; 394e870a8efSIlpo Järvinen } 395e870a8efSIlpo Järvinen 396a2a385d6SEric Dumazet static inline bool tcp_urg_mode(const struct tcp_sock *tp) 39733f5f57eSIlpo Järvinen { 39833f5f57eSIlpo Järvinen return tp->snd_una != tp->snd_up; 39933f5f57eSIlpo Järvinen } 40033f5f57eSIlpo Järvinen 40133ad798cSAdam Langley #define OPTION_SACK_ADVERTISE (1 << 0) 40233ad798cSAdam Langley #define OPTION_TS (1 << 1) 40333ad798cSAdam Langley #define OPTION_MD5 (1 << 2) 40489e95a61SOri Finkelman #define OPTION_WSCALE (1 << 3) 4052100c8d2SYuchung Cheng #define OPTION_FAST_OPEN_COOKIE (1 << 8) 40660e2a778SUrsula Braun #define OPTION_SMC (1 << 9) 40760e2a778SUrsula Braun 40860e2a778SUrsula Braun static void smc_options_write(__be32 *ptr, u16 *options) 40960e2a778SUrsula Braun { 41060e2a778SUrsula Braun #if IS_ENABLED(CONFIG_SMC) 41160e2a778SUrsula Braun if (static_branch_unlikely(&tcp_have_smc)) { 41260e2a778SUrsula Braun if (unlikely(OPTION_SMC & *options)) { 41360e2a778SUrsula Braun *ptr++ = htonl((TCPOPT_NOP << 24) | 41460e2a778SUrsula Braun (TCPOPT_NOP << 16) | 41560e2a778SUrsula Braun (TCPOPT_EXP << 8) | 41660e2a778SUrsula Braun (TCPOLEN_EXP_SMC_BASE)); 41760e2a778SUrsula Braun *ptr++ = htonl(TCPOPT_SMC_MAGIC); 41860e2a778SUrsula Braun } 41960e2a778SUrsula Braun } 42060e2a778SUrsula Braun #endif 42160e2a778SUrsula Braun } 42233ad798cSAdam Langley 42333ad798cSAdam Langley struct tcp_out_options { 4242100c8d2SYuchung Cheng u16 options; /* bit field of OPTION_* */ 4252100c8d2SYuchung Cheng u16 mss; /* 0 to disable */ 42633ad798cSAdam Langley u8 ws; /* window scale, 0 to disable */ 42733ad798cSAdam Langley u8 num_sack_blocks; /* number of SACK blocks to include */ 428bd0388aeSWilliam Allen Simpson u8 hash_size; /* bytes in hash_location */ 429bd0388aeSWilliam Allen Simpson __u8 *hash_location; /* temporary pointer, overloaded */ 4302100c8d2SYuchung Cheng __u32 tsval, tsecr; /* need to include OPTION_TS */ 4312100c8d2SYuchung Cheng struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */ 43233ad798cSAdam Langley }; 43333ad798cSAdam Langley 43467edfef7SAndi Kleen /* Write previously computed TCP options to the packet. 43567edfef7SAndi Kleen * 43667edfef7SAndi Kleen * Beware: Something in the Internet is very sensitive to the ordering of 437fd6149d3SIlpo Järvinen * TCP options, we learned this through the hard way, so be careful here. 438fd6149d3SIlpo Järvinen * Luckily we can at least blame others for their non-compliance but from 4398e3bff96Sstephen hemminger * inter-operability perspective it seems that we're somewhat stuck with 440fd6149d3SIlpo Järvinen * the ordering which we have been using if we want to keep working with 441fd6149d3SIlpo Järvinen * those broken things (not that it currently hurts anybody as there isn't 442fd6149d3SIlpo Järvinen * particular reason why the ordering would need to be changed). 443fd6149d3SIlpo Järvinen * 444fd6149d3SIlpo Järvinen * At least SACK_PERM as the first option is known to lead to a disaster 445fd6149d3SIlpo Järvinen * (but it may well be that other scenarios fail similarly). 446fd6149d3SIlpo Järvinen */ 44733ad798cSAdam Langley static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp, 448bd0388aeSWilliam Allen Simpson struct tcp_out_options *opts) 449bd0388aeSWilliam Allen Simpson { 4502100c8d2SYuchung Cheng u16 options = opts->options; /* mungable copy */ 451bd0388aeSWilliam Allen Simpson 452bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_MD5 & options)) { 4531a2c6181SChristoph Paasch *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | 4541a2c6181SChristoph Paasch (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG); 455bd0388aeSWilliam Allen Simpson /* overload cookie hash location */ 456bd0388aeSWilliam Allen Simpson opts->hash_location = (__u8 *)ptr; 45733ad798cSAdam Langley ptr += 4; 45833ad798cSAdam Langley } 45933ad798cSAdam Langley 460fd6149d3SIlpo Järvinen if (unlikely(opts->mss)) { 461fd6149d3SIlpo Järvinen *ptr++ = htonl((TCPOPT_MSS << 24) | 462fd6149d3SIlpo Järvinen (TCPOLEN_MSS << 16) | 463fd6149d3SIlpo Järvinen opts->mss); 464fd6149d3SIlpo Järvinen } 465fd6149d3SIlpo Järvinen 466bd0388aeSWilliam Allen Simpson if (likely(OPTION_TS & options)) { 467bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 46833ad798cSAdam Langley *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | 46933ad798cSAdam Langley (TCPOLEN_SACK_PERM << 16) | 47033ad798cSAdam Langley (TCPOPT_TIMESTAMP << 8) | 47133ad798cSAdam Langley TCPOLEN_TIMESTAMP); 472bd0388aeSWilliam Allen Simpson options &= ~OPTION_SACK_ADVERTISE; 47333ad798cSAdam Langley } else { 474496c98dfSYOSHIFUJI Hideaki *ptr++ = htonl((TCPOPT_NOP << 24) | 47540efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 47640efc6faSStephen Hemminger (TCPOPT_TIMESTAMP << 8) | 47740efc6faSStephen Hemminger TCPOLEN_TIMESTAMP); 47840efc6faSStephen Hemminger } 47933ad798cSAdam Langley *ptr++ = htonl(opts->tsval); 48033ad798cSAdam Langley *ptr++ = htonl(opts->tsecr); 48133ad798cSAdam Langley } 48233ad798cSAdam Langley 483bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 48433ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 48533ad798cSAdam Langley (TCPOPT_NOP << 16) | 48633ad798cSAdam Langley (TCPOPT_SACK_PERM << 8) | 48733ad798cSAdam Langley TCPOLEN_SACK_PERM); 48833ad798cSAdam Langley } 48933ad798cSAdam Langley 490bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_WSCALE & options)) { 49133ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 49233ad798cSAdam Langley (TCPOPT_WINDOW << 16) | 49333ad798cSAdam Langley (TCPOLEN_WINDOW << 8) | 49433ad798cSAdam Langley opts->ws); 49533ad798cSAdam Langley } 49633ad798cSAdam Langley 49733ad798cSAdam Langley if (unlikely(opts->num_sack_blocks)) { 49833ad798cSAdam Langley struct tcp_sack_block *sp = tp->rx_opt.dsack ? 49933ad798cSAdam Langley tp->duplicate_sack : tp->selective_acks; 50040efc6faSStephen Hemminger int this_sack; 50140efc6faSStephen Hemminger 50240efc6faSStephen Hemminger *ptr++ = htonl((TCPOPT_NOP << 24) | 50340efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 50440efc6faSStephen Hemminger (TCPOPT_SACK << 8) | 50533ad798cSAdam Langley (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * 50640efc6faSStephen Hemminger TCPOLEN_SACK_PERBLOCK))); 5072de979bdSStephen Hemminger 50833ad798cSAdam Langley for (this_sack = 0; this_sack < opts->num_sack_blocks; 50933ad798cSAdam Langley ++this_sack) { 51040efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].start_seq); 51140efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].end_seq); 51240efc6faSStephen Hemminger } 5132de979bdSStephen Hemminger 51440efc6faSStephen Hemminger tp->rx_opt.dsack = 0; 51540efc6faSStephen Hemminger } 5162100c8d2SYuchung Cheng 5172100c8d2SYuchung Cheng if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) { 5182100c8d2SYuchung Cheng struct tcp_fastopen_cookie *foc = opts->fastopen_cookie; 5197f9b838bSDaniel Lee u8 *p = (u8 *)ptr; 5207f9b838bSDaniel Lee u32 len; /* Fast Open option length */ 5212100c8d2SYuchung Cheng 5227f9b838bSDaniel Lee if (foc->exp) { 5237f9b838bSDaniel Lee len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len; 5247f9b838bSDaniel Lee *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) | 5252100c8d2SYuchung Cheng TCPOPT_FASTOPEN_MAGIC); 5267f9b838bSDaniel Lee p += TCPOLEN_EXP_FASTOPEN_BASE; 5277f9b838bSDaniel Lee } else { 5287f9b838bSDaniel Lee len = TCPOLEN_FASTOPEN_BASE + foc->len; 5297f9b838bSDaniel Lee *p++ = TCPOPT_FASTOPEN; 5307f9b838bSDaniel Lee *p++ = len; 5312100c8d2SYuchung Cheng } 5327f9b838bSDaniel Lee 5337f9b838bSDaniel Lee memcpy(p, foc->val, foc->len); 5347f9b838bSDaniel Lee if ((len & 3) == 2) { 5357f9b838bSDaniel Lee p[foc->len] = TCPOPT_NOP; 5367f9b838bSDaniel Lee p[foc->len + 1] = TCPOPT_NOP; 5377f9b838bSDaniel Lee } 5387f9b838bSDaniel Lee ptr += (len + 3) >> 2; 5392100c8d2SYuchung Cheng } 54060e2a778SUrsula Braun 54160e2a778SUrsula Braun smc_options_write(ptr, &options); 54260e2a778SUrsula Braun } 54360e2a778SUrsula Braun 54460e2a778SUrsula Braun static void smc_set_option(const struct tcp_sock *tp, 54560e2a778SUrsula Braun struct tcp_out_options *opts, 54660e2a778SUrsula Braun unsigned int *remaining) 54760e2a778SUrsula Braun { 54860e2a778SUrsula Braun #if IS_ENABLED(CONFIG_SMC) 54960e2a778SUrsula Braun if (static_branch_unlikely(&tcp_have_smc)) { 55060e2a778SUrsula Braun if (tp->syn_smc) { 55160e2a778SUrsula Braun if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) { 55260e2a778SUrsula Braun opts->options |= OPTION_SMC; 55360e2a778SUrsula Braun *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED; 55460e2a778SUrsula Braun } 55560e2a778SUrsula Braun } 55660e2a778SUrsula Braun } 55760e2a778SUrsula Braun #endif 55860e2a778SUrsula Braun } 55960e2a778SUrsula Braun 56060e2a778SUrsula Braun static void smc_set_option_cond(const struct tcp_sock *tp, 56160e2a778SUrsula Braun const struct inet_request_sock *ireq, 56260e2a778SUrsula Braun struct tcp_out_options *opts, 56360e2a778SUrsula Braun unsigned int *remaining) 56460e2a778SUrsula Braun { 56560e2a778SUrsula Braun #if IS_ENABLED(CONFIG_SMC) 56660e2a778SUrsula Braun if (static_branch_unlikely(&tcp_have_smc)) { 56760e2a778SUrsula Braun if (tp->syn_smc && ireq->smc_ok) { 56860e2a778SUrsula Braun if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) { 56960e2a778SUrsula Braun opts->options |= OPTION_SMC; 57060e2a778SUrsula Braun *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED; 57160e2a778SUrsula Braun } 57260e2a778SUrsula Braun } 57360e2a778SUrsula Braun } 57460e2a778SUrsula Braun #endif 57540efc6faSStephen Hemminger } 57640efc6faSStephen Hemminger 57767edfef7SAndi Kleen /* Compute TCP options for SYN packets. This is not the final 57867edfef7SAndi Kleen * network wire format yet. 57967edfef7SAndi Kleen */ 58095c96174SEric Dumazet static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb, 58133ad798cSAdam Langley struct tcp_out_options *opts, 582cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 583cf533ea5SEric Dumazet { 58433ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 58595c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 586783237e8SYuchung Cheng struct tcp_fastopen_request *fastopen = tp->fastopen_req; 58733ad798cSAdam Langley 588cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 58933ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 59033ad798cSAdam Langley if (*md5) { 59133ad798cSAdam Langley opts->options |= OPTION_MD5; 592bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 593cfb6eeb4SYOSHIFUJI Hideaki } 59433ad798cSAdam Langley #else 59533ad798cSAdam Langley *md5 = NULL; 596cfb6eeb4SYOSHIFUJI Hideaki #endif 59733ad798cSAdam Langley 59833ad798cSAdam Langley /* We always get an MSS option. The option bytes which will be seen in 59933ad798cSAdam Langley * normal data packets should timestamps be used, must be in the MSS 60033ad798cSAdam Langley * advertised. But we subtract them from tp->mss_cache so that 60133ad798cSAdam Langley * calculations in tcp_sendmsg are simpler etc. So account for this 60233ad798cSAdam Langley * fact here if necessary. If we don't do this correctly, as a 60333ad798cSAdam Langley * receiver we won't recognize data packets as being full sized when we 60433ad798cSAdam Langley * should, and thus we won't abide by the delayed ACK rules correctly. 60533ad798cSAdam Langley * SACKs don't matter, we never delay an ACK when we have any of those 60633ad798cSAdam Langley * going out. */ 60733ad798cSAdam Langley opts->mss = tcp_advertise_mss(sk); 608bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 60933ad798cSAdam Langley 6105d2ed052SEric Dumazet if (likely(sock_net(sk)->ipv4.sysctl_tcp_timestamps && !*md5)) { 61133ad798cSAdam Langley opts->options |= OPTION_TS; 6127faee5c0SEric Dumazet opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset; 61333ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 614bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 61533ad798cSAdam Langley } 6169bb37ef0SEric Dumazet if (likely(sock_net(sk)->ipv4.sysctl_tcp_window_scaling)) { 61733ad798cSAdam Langley opts->ws = tp->rx_opt.rcv_wscale; 61889e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 619bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 62033ad798cSAdam Langley } 621f9301034SEric Dumazet if (likely(sock_net(sk)->ipv4.sysctl_tcp_sack)) { 62233ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 623b32d1310SDavid S. Miller if (unlikely(!(OPTION_TS & opts->options))) 624bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 62533ad798cSAdam Langley } 62633ad798cSAdam Langley 627783237e8SYuchung Cheng if (fastopen && fastopen->cookie.len >= 0) { 6282646c831SDaniel Lee u32 need = fastopen->cookie.len; 6292646c831SDaniel Lee 6302646c831SDaniel Lee need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE : 6312646c831SDaniel Lee TCPOLEN_FASTOPEN_BASE; 632783237e8SYuchung Cheng need = (need + 3) & ~3U; /* Align to 32 bits */ 633783237e8SYuchung Cheng if (remaining >= need) { 634783237e8SYuchung Cheng opts->options |= OPTION_FAST_OPEN_COOKIE; 635783237e8SYuchung Cheng opts->fastopen_cookie = &fastopen->cookie; 636783237e8SYuchung Cheng remaining -= need; 637783237e8SYuchung Cheng tp->syn_fastopen = 1; 6382646c831SDaniel Lee tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0; 639783237e8SYuchung Cheng } 640783237e8SYuchung Cheng } 641bd0388aeSWilliam Allen Simpson 64260e2a778SUrsula Braun smc_set_option(tp, opts, &remaining); 64360e2a778SUrsula Braun 644bd0388aeSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 64533ad798cSAdam Langley } 64633ad798cSAdam Langley 64767edfef7SAndi Kleen /* Set up TCP options for SYN-ACKs. */ 64860e2a778SUrsula Braun static unsigned int tcp_synack_options(const struct sock *sk, 64960e2a778SUrsula Braun struct request_sock *req, 65095c96174SEric Dumazet unsigned int mss, struct sk_buff *skb, 65133ad798cSAdam Langley struct tcp_out_options *opts, 65280f03e27SEric Dumazet const struct tcp_md5sig_key *md5, 6538336886fSJerry Chu struct tcp_fastopen_cookie *foc) 6544957faadSWilliam Allen Simpson { 65533ad798cSAdam Langley struct inet_request_sock *ireq = inet_rsk(req); 65695c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 65733ad798cSAdam Langley 65833ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 65980f03e27SEric Dumazet if (md5) { 66033ad798cSAdam Langley opts->options |= OPTION_MD5; 6614957faadSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 6624957faadSWilliam Allen Simpson 6634957faadSWilliam Allen Simpson /* We can't fit any SACK blocks in a packet with MD5 + TS 6644957faadSWilliam Allen Simpson * options. There was discussion about disabling SACK 6654957faadSWilliam Allen Simpson * rather than TS in order to fit in better with old, 6664957faadSWilliam Allen Simpson * buggy kernels, but that was deemed to be unnecessary. 6674957faadSWilliam Allen Simpson */ 668de213e5eSEric Dumazet ireq->tstamp_ok &= !ireq->sack_ok; 66933ad798cSAdam Langley } 67033ad798cSAdam Langley #endif 67133ad798cSAdam Langley 6724957faadSWilliam Allen Simpson /* We always send an MSS option. */ 67333ad798cSAdam Langley opts->mss = mss; 6744957faadSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 67533ad798cSAdam Langley 67633ad798cSAdam Langley if (likely(ireq->wscale_ok)) { 67733ad798cSAdam Langley opts->ws = ireq->rcv_wscale; 67889e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 6794957faadSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 68033ad798cSAdam Langley } 681de213e5eSEric Dumazet if (likely(ireq->tstamp_ok)) { 68233ad798cSAdam Langley opts->options |= OPTION_TS; 68395a22caeSFlorian Westphal opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off; 68433ad798cSAdam Langley opts->tsecr = req->ts_recent; 6854957faadSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 68633ad798cSAdam Langley } 68733ad798cSAdam Langley if (likely(ireq->sack_ok)) { 68833ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 689de213e5eSEric Dumazet if (unlikely(!ireq->tstamp_ok)) 6904957faadSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 69133ad798cSAdam Langley } 6927f9b838bSDaniel Lee if (foc != NULL && foc->len >= 0) { 6937f9b838bSDaniel Lee u32 need = foc->len; 6947f9b838bSDaniel Lee 6957f9b838bSDaniel Lee need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE : 6967f9b838bSDaniel Lee TCPOLEN_FASTOPEN_BASE; 6978336886fSJerry Chu need = (need + 3) & ~3U; /* Align to 32 bits */ 6988336886fSJerry Chu if (remaining >= need) { 6998336886fSJerry Chu opts->options |= OPTION_FAST_OPEN_COOKIE; 7008336886fSJerry Chu opts->fastopen_cookie = foc; 7018336886fSJerry Chu remaining -= need; 7028336886fSJerry Chu } 7038336886fSJerry Chu } 7044957faadSWilliam Allen Simpson 70560e2a778SUrsula Braun smc_set_option_cond(tcp_sk(sk), ireq, opts, &remaining); 70660e2a778SUrsula Braun 7074957faadSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 70833ad798cSAdam Langley } 70933ad798cSAdam Langley 71067edfef7SAndi Kleen /* Compute TCP options for ESTABLISHED sockets. This is not the 71167edfef7SAndi Kleen * final wire format yet. 71267edfef7SAndi Kleen */ 71395c96174SEric Dumazet static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb, 71433ad798cSAdam Langley struct tcp_out_options *opts, 715cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 716cf533ea5SEric Dumazet { 71733ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 71895c96174SEric Dumazet unsigned int size = 0; 719cabeccbdSIlpo Järvinen unsigned int eff_sacks; 72033ad798cSAdam Langley 7215843ef42SAndi Kleen opts->options = 0; 7225843ef42SAndi Kleen 72333ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 72433ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 72533ad798cSAdam Langley if (unlikely(*md5)) { 72633ad798cSAdam Langley opts->options |= OPTION_MD5; 72733ad798cSAdam Langley size += TCPOLEN_MD5SIG_ALIGNED; 72833ad798cSAdam Langley } 72933ad798cSAdam Langley #else 73033ad798cSAdam Langley *md5 = NULL; 73133ad798cSAdam Langley #endif 73233ad798cSAdam Langley 73333ad798cSAdam Langley if (likely(tp->rx_opt.tstamp_ok)) { 73433ad798cSAdam Langley opts->options |= OPTION_TS; 7357faee5c0SEric Dumazet opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0; 73633ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 73733ad798cSAdam Langley size += TCPOLEN_TSTAMP_ALIGNED; 73833ad798cSAdam Langley } 73933ad798cSAdam Langley 740cabeccbdSIlpo Järvinen eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; 741cabeccbdSIlpo Järvinen if (unlikely(eff_sacks)) { 74295c96174SEric Dumazet const unsigned int remaining = MAX_TCP_OPTION_SPACE - size; 74333ad798cSAdam Langley opts->num_sack_blocks = 74495c96174SEric Dumazet min_t(unsigned int, eff_sacks, 74533ad798cSAdam Langley (remaining - TCPOLEN_SACK_BASE_ALIGNED) / 74633ad798cSAdam Langley TCPOLEN_SACK_PERBLOCK); 74733ad798cSAdam Langley size += TCPOLEN_SACK_BASE_ALIGNED + 74833ad798cSAdam Langley opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; 74933ad798cSAdam Langley } 75033ad798cSAdam Langley 75133ad798cSAdam Langley return size; 75240efc6faSStephen Hemminger } 7531da177e4SLinus Torvalds 75446d3ceabSEric Dumazet 75546d3ceabSEric Dumazet /* TCP SMALL QUEUES (TSQ) 75646d3ceabSEric Dumazet * 75746d3ceabSEric Dumazet * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev) 75846d3ceabSEric Dumazet * to reduce RTT and bufferbloat. 75946d3ceabSEric Dumazet * We do this using a special skb destructor (tcp_wfree). 76046d3ceabSEric Dumazet * 76146d3ceabSEric Dumazet * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb 76246d3ceabSEric Dumazet * needs to be reallocated in a driver. 7638e3bff96Sstephen hemminger * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc 76446d3ceabSEric Dumazet * 76546d3ceabSEric Dumazet * Since transmit from skb destructor is forbidden, we use a tasklet 76646d3ceabSEric Dumazet * to process all sockets that eventually need to send more skbs. 76746d3ceabSEric Dumazet * We use one tasklet per cpu, with its own queue of sockets. 76846d3ceabSEric Dumazet */ 76946d3ceabSEric Dumazet struct tsq_tasklet { 77046d3ceabSEric Dumazet struct tasklet_struct tasklet; 77146d3ceabSEric Dumazet struct list_head head; /* queue of tcp sockets */ 77246d3ceabSEric Dumazet }; 77346d3ceabSEric Dumazet static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet); 77446d3ceabSEric Dumazet 7756f458dfbSEric Dumazet static void tcp_tsq_handler(struct sock *sk) 7766f458dfbSEric Dumazet { 7776f458dfbSEric Dumazet if ((1 << sk->sk_state) & 7786f458dfbSEric Dumazet (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING | 779f9616c35SEric Dumazet TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) { 780f9616c35SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 781f9616c35SEric Dumazet 782f9616c35SEric Dumazet if (tp->lost_out > tp->retrans_out && 7833a91d29fSKoichiro Den tp->snd_cwnd > tcp_packets_in_flight(tp)) { 7843a91d29fSKoichiro Den tcp_mstamp_refresh(tp); 785f9616c35SEric Dumazet tcp_xmit_retransmit_queue(sk); 7863a91d29fSKoichiro Den } 787f9616c35SEric Dumazet 788f9616c35SEric Dumazet tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle, 789bf06200eSJohn Ogness 0, GFP_ATOMIC); 7906f458dfbSEric Dumazet } 791f9616c35SEric Dumazet } 79246d3ceabSEric Dumazet /* 7938e3bff96Sstephen hemminger * One tasklet per cpu tries to send more skbs. 79446d3ceabSEric Dumazet * We run in tasklet context but need to disable irqs when 7958e3bff96Sstephen hemminger * transferring tsq->head because tcp_wfree() might 79646d3ceabSEric Dumazet * interrupt us (non NAPI drivers) 79746d3ceabSEric Dumazet */ 79846d3ceabSEric Dumazet static void tcp_tasklet_func(unsigned long data) 79946d3ceabSEric Dumazet { 80046d3ceabSEric Dumazet struct tsq_tasklet *tsq = (struct tsq_tasklet *)data; 80146d3ceabSEric Dumazet LIST_HEAD(list); 80246d3ceabSEric Dumazet unsigned long flags; 80346d3ceabSEric Dumazet struct list_head *q, *n; 80446d3ceabSEric Dumazet struct tcp_sock *tp; 80546d3ceabSEric Dumazet struct sock *sk; 80646d3ceabSEric Dumazet 80746d3ceabSEric Dumazet local_irq_save(flags); 80846d3ceabSEric Dumazet list_splice_init(&tsq->head, &list); 80946d3ceabSEric Dumazet local_irq_restore(flags); 81046d3ceabSEric Dumazet 81146d3ceabSEric Dumazet list_for_each_safe(q, n, &list) { 81246d3ceabSEric Dumazet tp = list_entry(q, struct tcp_sock, tsq_node); 81346d3ceabSEric Dumazet list_del(&tp->tsq_node); 81446d3ceabSEric Dumazet 81546d3ceabSEric Dumazet sk = (struct sock *)tp; 8160a9648f1SEric Dumazet smp_mb__before_atomic(); 8177aa5470cSEric Dumazet clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags); 8187aa5470cSEric Dumazet 819b223feb9SEric Dumazet if (!sk->sk_lock.owned && 8207aa5470cSEric Dumazet test_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags)) { 82146d3ceabSEric Dumazet bh_lock_sock(sk); 82246d3ceabSEric Dumazet if (!sock_owned_by_user(sk)) { 8237aa5470cSEric Dumazet clear_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags); 8246f458dfbSEric Dumazet tcp_tsq_handler(sk); 82546d3ceabSEric Dumazet } 82646d3ceabSEric Dumazet bh_unlock_sock(sk); 827b223feb9SEric Dumazet } 82846d3ceabSEric Dumazet 82946d3ceabSEric Dumazet sk_free(sk); 83046d3ceabSEric Dumazet } 83146d3ceabSEric Dumazet } 83246d3ceabSEric Dumazet 83340fc3423SEric Dumazet #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \ 83440fc3423SEric Dumazet TCPF_WRITE_TIMER_DEFERRED | \ 83540fc3423SEric Dumazet TCPF_DELACK_TIMER_DEFERRED | \ 83640fc3423SEric Dumazet TCPF_MTU_REDUCED_DEFERRED) 83746d3ceabSEric Dumazet /** 83846d3ceabSEric Dumazet * tcp_release_cb - tcp release_sock() callback 83946d3ceabSEric Dumazet * @sk: socket 84046d3ceabSEric Dumazet * 84146d3ceabSEric Dumazet * called from release_sock() to perform protocol dependent 84246d3ceabSEric Dumazet * actions before socket release. 84346d3ceabSEric Dumazet */ 84446d3ceabSEric Dumazet void tcp_release_cb(struct sock *sk) 84546d3ceabSEric Dumazet { 8466f458dfbSEric Dumazet unsigned long flags, nflags; 84746d3ceabSEric Dumazet 8486f458dfbSEric Dumazet /* perform an atomic operation only if at least one flag is set */ 8496f458dfbSEric Dumazet do { 8507aa5470cSEric Dumazet flags = sk->sk_tsq_flags; 8516f458dfbSEric Dumazet if (!(flags & TCP_DEFERRED_ALL)) 8526f458dfbSEric Dumazet return; 8536f458dfbSEric Dumazet nflags = flags & ~TCP_DEFERRED_ALL; 8547aa5470cSEric Dumazet } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags); 8556f458dfbSEric Dumazet 85640fc3423SEric Dumazet if (flags & TCPF_TSQ_DEFERRED) 8576f458dfbSEric Dumazet tcp_tsq_handler(sk); 8586f458dfbSEric Dumazet 859c3f9b018SEric Dumazet /* Here begins the tricky part : 860c3f9b018SEric Dumazet * We are called from release_sock() with : 861c3f9b018SEric Dumazet * 1) BH disabled 862c3f9b018SEric Dumazet * 2) sk_lock.slock spinlock held 863c3f9b018SEric Dumazet * 3) socket owned by us (sk->sk_lock.owned == 1) 864c3f9b018SEric Dumazet * 865c3f9b018SEric Dumazet * But following code is meant to be called from BH handlers, 866c3f9b018SEric Dumazet * so we should keep BH disabled, but early release socket ownership 867c3f9b018SEric Dumazet */ 868c3f9b018SEric Dumazet sock_release_ownership(sk); 869c3f9b018SEric Dumazet 87040fc3423SEric Dumazet if (flags & TCPF_WRITE_TIMER_DEFERRED) { 8716f458dfbSEric Dumazet tcp_write_timer_handler(sk); 872144d56e9SEric Dumazet __sock_put(sk); 873144d56e9SEric Dumazet } 87440fc3423SEric Dumazet if (flags & TCPF_DELACK_TIMER_DEFERRED) { 8756f458dfbSEric Dumazet tcp_delack_timer_handler(sk); 876144d56e9SEric Dumazet __sock_put(sk); 877144d56e9SEric Dumazet } 87840fc3423SEric Dumazet if (flags & TCPF_MTU_REDUCED_DEFERRED) { 8794fab9071SNeal Cardwell inet_csk(sk)->icsk_af_ops->mtu_reduced(sk); 880144d56e9SEric Dumazet __sock_put(sk); 881144d56e9SEric Dumazet } 88246d3ceabSEric Dumazet } 88346d3ceabSEric Dumazet EXPORT_SYMBOL(tcp_release_cb); 88446d3ceabSEric Dumazet 88546d3ceabSEric Dumazet void __init tcp_tasklet_init(void) 88646d3ceabSEric Dumazet { 88746d3ceabSEric Dumazet int i; 88846d3ceabSEric Dumazet 88946d3ceabSEric Dumazet for_each_possible_cpu(i) { 89046d3ceabSEric Dumazet struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i); 89146d3ceabSEric Dumazet 89246d3ceabSEric Dumazet INIT_LIST_HEAD(&tsq->head); 89346d3ceabSEric Dumazet tasklet_init(&tsq->tasklet, 89446d3ceabSEric Dumazet tcp_tasklet_func, 89546d3ceabSEric Dumazet (unsigned long)tsq); 89646d3ceabSEric Dumazet } 89746d3ceabSEric Dumazet } 89846d3ceabSEric Dumazet 89946d3ceabSEric Dumazet /* 90046d3ceabSEric Dumazet * Write buffer destructor automatically called from kfree_skb. 9018e3bff96Sstephen hemminger * We can't xmit new skbs from this context, as we might already 90246d3ceabSEric Dumazet * hold qdisc lock. 90346d3ceabSEric Dumazet */ 904d6a4a104SEric Dumazet void tcp_wfree(struct sk_buff *skb) 90546d3ceabSEric Dumazet { 90646d3ceabSEric Dumazet struct sock *sk = skb->sk; 90746d3ceabSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 908408f0a6cSEric Dumazet unsigned long flags, nval, oval; 9099b462d02SEric Dumazet 9109b462d02SEric Dumazet /* Keep one reference on sk_wmem_alloc. 9119b462d02SEric Dumazet * Will be released by sk_free() from here or tcp_tasklet_func() 9129b462d02SEric Dumazet */ 91314afee4bSReshetova, Elena WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc)); 9149b462d02SEric Dumazet 9159b462d02SEric Dumazet /* If this softirq is serviced by ksoftirqd, we are likely under stress. 9169b462d02SEric Dumazet * Wait until our queues (qdisc + devices) are drained. 9179b462d02SEric Dumazet * This gives : 9189b462d02SEric Dumazet * - less callbacks to tcp_write_xmit(), reducing stress (batches) 9199b462d02SEric Dumazet * - chance for incoming ACK (processed by another cpu maybe) 9209b462d02SEric Dumazet * to migrate this flow (skb->ooo_okay will be eventually set) 9219b462d02SEric Dumazet */ 92214afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current) 9239b462d02SEric Dumazet goto out; 92446d3ceabSEric Dumazet 9257aa5470cSEric Dumazet for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) { 92646d3ceabSEric Dumazet struct tsq_tasklet *tsq; 927a9b204d1SEric Dumazet bool empty; 92846d3ceabSEric Dumazet 929408f0a6cSEric Dumazet if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED)) 930408f0a6cSEric Dumazet goto out; 931408f0a6cSEric Dumazet 932b223feb9SEric Dumazet nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED | TCPF_TSQ_DEFERRED; 9337aa5470cSEric Dumazet nval = cmpxchg(&sk->sk_tsq_flags, oval, nval); 934408f0a6cSEric Dumazet if (nval != oval) 935408f0a6cSEric Dumazet continue; 936408f0a6cSEric Dumazet 93746d3ceabSEric Dumazet /* queue this socket to tasklet queue */ 93846d3ceabSEric Dumazet local_irq_save(flags); 939903ceff7SChristoph Lameter tsq = this_cpu_ptr(&tsq_tasklet); 940a9b204d1SEric Dumazet empty = list_empty(&tsq->head); 94146d3ceabSEric Dumazet list_add(&tp->tsq_node, &tsq->head); 942a9b204d1SEric Dumazet if (empty) 94346d3ceabSEric Dumazet tasklet_schedule(&tsq->tasklet); 94446d3ceabSEric Dumazet local_irq_restore(flags); 9459b462d02SEric Dumazet return; 94646d3ceabSEric Dumazet } 9479b462d02SEric Dumazet out: 9489b462d02SEric Dumazet sk_free(sk); 94946d3ceabSEric Dumazet } 95046d3ceabSEric Dumazet 951218af599SEric Dumazet /* Note: Called under hard irq. 952218af599SEric Dumazet * We can not call TCP stack right away. 953218af599SEric Dumazet */ 954218af599SEric Dumazet enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer) 955218af599SEric Dumazet { 956218af599SEric Dumazet struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer); 957218af599SEric Dumazet struct sock *sk = (struct sock *)tp; 958218af599SEric Dumazet unsigned long nval, oval; 959218af599SEric Dumazet 960218af599SEric Dumazet for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) { 961218af599SEric Dumazet struct tsq_tasklet *tsq; 962218af599SEric Dumazet bool empty; 963218af599SEric Dumazet 964218af599SEric Dumazet if (oval & TSQF_QUEUED) 965218af599SEric Dumazet break; 966218af599SEric Dumazet 967218af599SEric Dumazet nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED | TCPF_TSQ_DEFERRED; 968218af599SEric Dumazet nval = cmpxchg(&sk->sk_tsq_flags, oval, nval); 969218af599SEric Dumazet if (nval != oval) 970218af599SEric Dumazet continue; 971218af599SEric Dumazet 97214afee4bSReshetova, Elena if (!refcount_inc_not_zero(&sk->sk_wmem_alloc)) 973218af599SEric Dumazet break; 974218af599SEric Dumazet /* queue this socket to tasklet queue */ 975218af599SEric Dumazet tsq = this_cpu_ptr(&tsq_tasklet); 976218af599SEric Dumazet empty = list_empty(&tsq->head); 977218af599SEric Dumazet list_add(&tp->tsq_node, &tsq->head); 978218af599SEric Dumazet if (empty) 979218af599SEric Dumazet tasklet_schedule(&tsq->tasklet); 980218af599SEric Dumazet break; 981218af599SEric Dumazet } 982218af599SEric Dumazet return HRTIMER_NORESTART; 983218af599SEric Dumazet } 984218af599SEric Dumazet 985218af599SEric Dumazet /* BBR congestion control needs pacing. 986218af599SEric Dumazet * Same remark for SO_MAX_PACING_RATE. 987218af599SEric Dumazet * sch_fq packet scheduler is efficiently handling pacing, 988218af599SEric Dumazet * but is not always installed/used. 989218af599SEric Dumazet * Return true if TCP stack should pace packets itself. 990218af599SEric Dumazet */ 991218af599SEric Dumazet static bool tcp_needs_internal_pacing(const struct sock *sk) 992218af599SEric Dumazet { 993218af599SEric Dumazet return smp_load_acquire(&sk->sk_pacing_status) == SK_PACING_NEEDED; 994218af599SEric Dumazet } 995218af599SEric Dumazet 996218af599SEric Dumazet static void tcp_internal_pacing(struct sock *sk, const struct sk_buff *skb) 997218af599SEric Dumazet { 998218af599SEric Dumazet u64 len_ns; 999218af599SEric Dumazet u32 rate; 1000218af599SEric Dumazet 1001218af599SEric Dumazet if (!tcp_needs_internal_pacing(sk)) 1002218af599SEric Dumazet return; 1003218af599SEric Dumazet rate = sk->sk_pacing_rate; 1004218af599SEric Dumazet if (!rate || rate == ~0U) 1005218af599SEric Dumazet return; 1006218af599SEric Dumazet 1007218af599SEric Dumazet /* Should account for header sizes as sch_fq does, 1008218af599SEric Dumazet * but lets make things simple. 1009218af599SEric Dumazet */ 1010218af599SEric Dumazet len_ns = (u64)skb->len * NSEC_PER_SEC; 1011218af599SEric Dumazet do_div(len_ns, rate); 1012218af599SEric Dumazet hrtimer_start(&tcp_sk(sk)->pacing_timer, 1013218af599SEric Dumazet ktime_add_ns(ktime_get(), len_ns), 1014218af599SEric Dumazet HRTIMER_MODE_ABS_PINNED); 1015218af599SEric Dumazet } 1016218af599SEric Dumazet 1017e2080072SEric Dumazet static void tcp_update_skb_after_send(struct tcp_sock *tp, struct sk_buff *skb) 1018e2080072SEric Dumazet { 1019e2080072SEric Dumazet skb->skb_mstamp = tp->tcp_mstamp; 1020e2080072SEric Dumazet list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue); 1021e2080072SEric Dumazet } 1022e2080072SEric Dumazet 10231da177e4SLinus Torvalds /* This routine actually transmits TCP packets queued in by 10241da177e4SLinus Torvalds * tcp_do_sendmsg(). This is used by both the initial 10251da177e4SLinus Torvalds * transmission and possible later retransmissions. 10261da177e4SLinus Torvalds * All SKB's seen here are completely headerless. It is our 10271da177e4SLinus Torvalds * job to build the TCP header, and pass the packet down to 10281da177e4SLinus Torvalds * IP so it can do the same plus pass the packet off to the 10291da177e4SLinus Torvalds * device. 10301da177e4SLinus Torvalds * 10311da177e4SLinus Torvalds * We are working here with either a clone of the original 10321da177e4SLinus Torvalds * SKB, or a fresh unique copy made by the retransmit engine. 10331da177e4SLinus Torvalds */ 1034056834d9SIlpo Järvinen static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, 1035056834d9SIlpo Järvinen gfp_t gfp_mask) 10361da177e4SLinus Torvalds { 10376687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 1038dfb4b9dcSDavid S. Miller struct inet_sock *inet; 1039dfb4b9dcSDavid S. Miller struct tcp_sock *tp; 1040dfb4b9dcSDavid S. Miller struct tcp_skb_cb *tcb; 104133ad798cSAdam Langley struct tcp_out_options opts; 104295c96174SEric Dumazet unsigned int tcp_options_size, tcp_header_size; 10438c72c65bSEric Dumazet struct sk_buff *oskb = NULL; 1044cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 10451da177e4SLinus Torvalds struct tcphdr *th; 10461da177e4SLinus Torvalds int err; 10471da177e4SLinus Torvalds 1048dfb4b9dcSDavid S. Miller BUG_ON(!skb || !tcp_skb_pcount(skb)); 10496f094b9eSLawrence Brakmo tp = tcp_sk(sk); 1050dfb4b9dcSDavid S. Miller 1051ccdbb6e9SEric Dumazet if (clone_it) { 10526f094b9eSLawrence Brakmo TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq 10536f094b9eSLawrence Brakmo - tp->snd_una; 10548c72c65bSEric Dumazet oskb = skb; 1055e2080072SEric Dumazet 1056e2080072SEric Dumazet tcp_skb_tsorted_save(oskb) { 1057e2080072SEric Dumazet if (unlikely(skb_cloned(oskb))) 1058e2080072SEric Dumazet skb = pskb_copy(oskb, gfp_mask); 1059dfb4b9dcSDavid S. Miller else 1060e2080072SEric Dumazet skb = skb_clone(oskb, gfp_mask); 1061e2080072SEric Dumazet } tcp_skb_tsorted_restore(oskb); 1062e2080072SEric Dumazet 1063dfb4b9dcSDavid S. Miller if (unlikely(!skb)) 1064dfb4b9dcSDavid S. Miller return -ENOBUFS; 1065dfb4b9dcSDavid S. Miller } 10668c72c65bSEric Dumazet skb->skb_mstamp = tp->tcp_mstamp; 1067dfb4b9dcSDavid S. Miller 1068dfb4b9dcSDavid S. Miller inet = inet_sk(sk); 1069dfb4b9dcSDavid S. Miller tcb = TCP_SKB_CB(skb); 107033ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 10711da177e4SLinus Torvalds 10724de075e0SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) 107333ad798cSAdam Langley tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); 107433ad798cSAdam Langley else 107533ad798cSAdam Langley tcp_options_size = tcp_established_options(sk, skb, &opts, 107633ad798cSAdam Langley &md5); 107733ad798cSAdam Langley tcp_header_size = tcp_options_size + sizeof(struct tcphdr); 10781da177e4SLinus Torvalds 1079547669d4SEric Dumazet /* if no packet is in qdisc/device queue, then allow XPS to select 1080b2532eb9SEric Dumazet * another queue. We can be called from tcp_tsq_handler() 1081b2532eb9SEric Dumazet * which holds one reference to sk_wmem_alloc. 1082b2532eb9SEric Dumazet * 1083b2532eb9SEric Dumazet * TODO: Ideally, in-flight pure ACK packets should not matter here. 1084b2532eb9SEric Dumazet * One way to get this would be to set skb->truesize = 2 on them. 1085547669d4SEric Dumazet */ 1086b2532eb9SEric Dumazet skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1); 10871da177e4SLinus Torvalds 108838ab52e8SEric Dumazet /* If we had to use memory reserve to allocate this skb, 108938ab52e8SEric Dumazet * this might cause drops if packet is looped back : 109038ab52e8SEric Dumazet * Other socket might not have SOCK_MEMALLOC. 109138ab52e8SEric Dumazet * Packets not looped back do not care about pfmemalloc. 109238ab52e8SEric Dumazet */ 109338ab52e8SEric Dumazet skb->pfmemalloc = 0; 109438ab52e8SEric Dumazet 1095aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 1096aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 109746d3ceabSEric Dumazet 109846d3ceabSEric Dumazet skb_orphan(skb); 109946d3ceabSEric Dumazet skb->sk = sk; 11001d2077acSEric Dumazet skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree; 1101b73c3d0eSTom Herbert skb_set_hash_from_sk(skb, sk); 110214afee4bSReshetova, Elena refcount_add(skb->truesize, &sk->sk_wmem_alloc); 11031da177e4SLinus Torvalds 1104c3a2e837SJulian Anastasov skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm); 1105c3a2e837SJulian Anastasov 11061da177e4SLinus Torvalds /* Build TCP header and checksum it. */ 1107ea1627c2SEric Dumazet th = (struct tcphdr *)skb->data; 1108c720c7e8SEric Dumazet th->source = inet->inet_sport; 1109c720c7e8SEric Dumazet th->dest = inet->inet_dport; 11101da177e4SLinus Torvalds th->seq = htonl(tcb->seq); 11111da177e4SLinus Torvalds th->ack_seq = htonl(tp->rcv_nxt); 1112df7a3b07SAl Viro *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | 11134de075e0SEric Dumazet tcb->tcp_flags); 1114dfb4b9dcSDavid S. Miller 11151da177e4SLinus Torvalds th->check = 0; 11161da177e4SLinus Torvalds th->urg_ptr = 0; 11171da177e4SLinus Torvalds 111833f5f57eSIlpo Järvinen /* The urg_mode check is necessary during a below snd_una win probe */ 11197691367dSHerbert Xu if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { 11207691367dSHerbert Xu if (before(tp->snd_up, tcb->seq + 0x10000)) { 11211da177e4SLinus Torvalds th->urg_ptr = htons(tp->snd_up - tcb->seq); 11221da177e4SLinus Torvalds th->urg = 1; 11237691367dSHerbert Xu } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { 11240eae88f3SEric Dumazet th->urg_ptr = htons(0xFFFF); 11257691367dSHerbert Xu th->urg = 1; 11267691367dSHerbert Xu } 11271da177e4SLinus Torvalds } 11281da177e4SLinus Torvalds 1129bd0388aeSWilliam Allen Simpson tcp_options_write((__be32 *)(th + 1), tp, &opts); 113051466a75SEric Dumazet skb_shinfo(skb)->gso_type = sk->sk_gso_type; 1131ea1627c2SEric Dumazet if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) { 1132ea1627c2SEric Dumazet th->window = htons(tcp_select_window(sk)); 1133ea1627c2SEric Dumazet tcp_ecn_send(sk, skb, th, tcp_header_size); 1134ea1627c2SEric Dumazet } else { 1135ea1627c2SEric Dumazet /* RFC1323: The window in SYN & SYN/ACK segments 1136ea1627c2SEric Dumazet * is never scaled. 1137ea1627c2SEric Dumazet */ 1138ea1627c2SEric Dumazet th->window = htons(min(tp->rcv_wnd, 65535U)); 1139ea1627c2SEric Dumazet } 1140cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 1141cfb6eeb4SYOSHIFUJI Hideaki /* Calculate the MD5 hash, as we have all we need now */ 1142cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 1143a465419bSEric Dumazet sk_nocaps_add(sk, NETIF_F_GSO_MASK); 1144bd0388aeSWilliam Allen Simpson tp->af_specific->calc_md5_hash(opts.hash_location, 114539f8e58eSEric Dumazet md5, sk, skb); 1146cfb6eeb4SYOSHIFUJI Hideaki } 1147cfb6eeb4SYOSHIFUJI Hideaki #endif 1148cfb6eeb4SYOSHIFUJI Hideaki 1149bb296246SHerbert Xu icsk->icsk_af_ops->send_check(sk, skb); 11501da177e4SLinus Torvalds 11514de075e0SEric Dumazet if (likely(tcb->tcp_flags & TCPHDR_ACK)) 1152fc6415bcSDavid S. Miller tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); 11531da177e4SLinus Torvalds 1154a44d6eacSMartin KaFai Lau if (skb->len != tcp_header_size) { 1155cf533ea5SEric Dumazet tcp_event_data_sent(tp, sk); 1156a44d6eacSMartin KaFai Lau tp->data_segs_out += tcp_skb_pcount(skb); 1157218af599SEric Dumazet tcp_internal_pacing(sk, skb); 1158a44d6eacSMartin KaFai Lau } 11591da177e4SLinus Torvalds 1160bd37a088SWei Yongjun if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) 1161aa2ea058STom Herbert TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, 1162aa2ea058STom Herbert tcp_skb_pcount(skb)); 11631da177e4SLinus Torvalds 11642efd055cSMarcelo Ricardo Leitner tp->segs_out += tcp_skb_pcount(skb); 1165f69ad292SEric Dumazet /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */ 1166cd7d8498SEric Dumazet skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb); 1167f69ad292SEric Dumazet skb_shinfo(skb)->gso_size = tcp_skb_mss(skb); 1168cd7d8498SEric Dumazet 11697faee5c0SEric Dumazet /* Our usage of tstamp should remain private */ 11702456e855SThomas Gleixner skb->tstamp = 0; 1171971f10ecSEric Dumazet 1172971f10ecSEric Dumazet /* Cleanup our debris for IP stacks */ 1173971f10ecSEric Dumazet memset(skb->cb, 0, max(sizeof(struct inet_skb_parm), 1174971f10ecSEric Dumazet sizeof(struct inet6_skb_parm))); 1175971f10ecSEric Dumazet 1176b0270e91SEric Dumazet err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl); 11777faee5c0SEric Dumazet 11788c72c65bSEric Dumazet if (unlikely(err > 0)) { 11795ee2c941SChristoph Paasch tcp_enter_cwr(sk); 11808c72c65bSEric Dumazet err = net_xmit_eval(err); 11818c72c65bSEric Dumazet } 1182fc225799SEric Dumazet if (!err && oskb) { 1183e2080072SEric Dumazet tcp_update_skb_after_send(tp, oskb); 1184fc225799SEric Dumazet tcp_rate_skb_sent(sk, oskb); 1185fc225799SEric Dumazet } 11868c72c65bSEric Dumazet return err; 11871da177e4SLinus Torvalds } 11881da177e4SLinus Torvalds 118967edfef7SAndi Kleen /* This routine just queues the buffer for sending. 11901da177e4SLinus Torvalds * 11911da177e4SLinus Torvalds * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, 11921da177e4SLinus Torvalds * otherwise socket can stall. 11931da177e4SLinus Torvalds */ 11941da177e4SLinus Torvalds static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) 11951da177e4SLinus Torvalds { 11961da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 11971da177e4SLinus Torvalds 11981da177e4SLinus Torvalds /* Advance write_seq and place onto the write_queue. */ 11991da177e4SLinus Torvalds tp->write_seq = TCP_SKB_CB(skb)->end_seq; 1200f4a775d1SEric Dumazet __skb_header_release(skb); 1201fe067e8aSDavid S. Miller tcp_add_write_queue_tail(sk, skb); 12023ab224beSHideo Aoki sk->sk_wmem_queued += skb->truesize; 12033ab224beSHideo Aoki sk_mem_charge(sk, skb->truesize); 12041da177e4SLinus Torvalds } 12051da177e4SLinus Torvalds 120667edfef7SAndi Kleen /* Initialize TSO segments for a packet. */ 12075bbb432cSEric Dumazet static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now) 1208f6302d1dSDavid S. Miller { 12094a64fd6cSEric Dumazet if (skb->len <= mss_now) { 1210f6302d1dSDavid S. Miller /* Avoid the costly divide in the normal 1211f6302d1dSDavid S. Miller * non-TSO case. 1212f6302d1dSDavid S. Miller */ 1213cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, 1); 1214f69ad292SEric Dumazet TCP_SKB_CB(skb)->tcp_gso_size = 0; 1215f6302d1dSDavid S. Miller } else { 1216cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now)); 1217f69ad292SEric Dumazet TCP_SKB_CB(skb)->tcp_gso_size = mss_now; 12181da177e4SLinus Torvalds } 12191da177e4SLinus Torvalds } 12201da177e4SLinus Torvalds 1221797108d1SIlpo Järvinen /* Pcount in the middle of the write queue got changed, we need to do various 1222797108d1SIlpo Järvinen * tweaks to fix counters 1223797108d1SIlpo Järvinen */ 1224cf533ea5SEric Dumazet static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) 1225797108d1SIlpo Järvinen { 1226797108d1SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1227797108d1SIlpo Järvinen 1228797108d1SIlpo Järvinen tp->packets_out -= decr; 1229797108d1SIlpo Järvinen 1230797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 1231797108d1SIlpo Järvinen tp->sacked_out -= decr; 1232797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 1233797108d1SIlpo Järvinen tp->retrans_out -= decr; 1234797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 1235797108d1SIlpo Järvinen tp->lost_out -= decr; 1236797108d1SIlpo Järvinen 1237797108d1SIlpo Järvinen /* Reno case is special. Sigh... */ 1238797108d1SIlpo Järvinen if (tcp_is_reno(tp) && decr > 0) 1239797108d1SIlpo Järvinen tp->sacked_out -= min_t(u32, tp->sacked_out, decr); 1240797108d1SIlpo Järvinen 1241797108d1SIlpo Järvinen if (tp->lost_skb_hint && 1242797108d1SIlpo Järvinen before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && 1243713bafeaSYuchung Cheng (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) 1244797108d1SIlpo Järvinen tp->lost_cnt_hint -= decr; 1245797108d1SIlpo Järvinen 1246797108d1SIlpo Järvinen tcp_verify_left_out(tp); 1247797108d1SIlpo Järvinen } 1248797108d1SIlpo Järvinen 12490a2cf20cSSoheil Hassas Yeganeh static bool tcp_has_tx_tstamp(const struct sk_buff *skb) 12500a2cf20cSSoheil Hassas Yeganeh { 12510a2cf20cSSoheil Hassas Yeganeh return TCP_SKB_CB(skb)->txstamp_ack || 12520a2cf20cSSoheil Hassas Yeganeh (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP); 12530a2cf20cSSoheil Hassas Yeganeh } 12540a2cf20cSSoheil Hassas Yeganeh 1255490cc7d0SWillem de Bruijn static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2) 1256490cc7d0SWillem de Bruijn { 1257490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo = skb_shinfo(skb); 1258490cc7d0SWillem de Bruijn 12590a2cf20cSSoheil Hassas Yeganeh if (unlikely(tcp_has_tx_tstamp(skb)) && 1260490cc7d0SWillem de Bruijn !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) { 1261490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo2 = skb_shinfo(skb2); 1262490cc7d0SWillem de Bruijn u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP; 1263490cc7d0SWillem de Bruijn 1264490cc7d0SWillem de Bruijn shinfo->tx_flags &= ~tsflags; 1265490cc7d0SWillem de Bruijn shinfo2->tx_flags |= tsflags; 1266490cc7d0SWillem de Bruijn swap(shinfo->tskey, shinfo2->tskey); 1267b51e13faSMartin KaFai Lau TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack; 1268b51e13faSMartin KaFai Lau TCP_SKB_CB(skb)->txstamp_ack = 0; 1269490cc7d0SWillem de Bruijn } 1270490cc7d0SWillem de Bruijn } 1271490cc7d0SWillem de Bruijn 1272a166140eSMartin KaFai Lau static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2) 1273a166140eSMartin KaFai Lau { 1274a166140eSMartin KaFai Lau TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor; 1275a166140eSMartin KaFai Lau TCP_SKB_CB(skb)->eor = 0; 1276a166140eSMartin KaFai Lau } 1277a166140eSMartin KaFai Lau 127875c119afSEric Dumazet /* Insert buff after skb on the write or rtx queue of sk. */ 127975c119afSEric Dumazet static void tcp_insert_write_queue_after(struct sk_buff *skb, 128075c119afSEric Dumazet struct sk_buff *buff, 128175c119afSEric Dumazet struct sock *sk, 128275c119afSEric Dumazet enum tcp_queue tcp_queue) 128375c119afSEric Dumazet { 128475c119afSEric Dumazet if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE) 128575c119afSEric Dumazet __skb_queue_after(&sk->sk_write_queue, skb, buff); 128675c119afSEric Dumazet else 128775c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, buff); 128875c119afSEric Dumazet } 128975c119afSEric Dumazet 12901da177e4SLinus Torvalds /* Function to create two new TCP segments. Shrinks the given segment 12911da177e4SLinus Torvalds * to the specified size and appends a new segment with the rest of the 12921da177e4SLinus Torvalds * packet to the list. This won't be called frequently, I hope. 12931da177e4SLinus Torvalds * Remember, these are still headerless SKBs at this point. 12941da177e4SLinus Torvalds */ 129575c119afSEric Dumazet int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue, 129675c119afSEric Dumazet struct sk_buff *skb, u32 len, 12976cc55e09SOctavian Purdila unsigned int mss_now, gfp_t gfp) 12981da177e4SLinus Torvalds { 12991da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 13001da177e4SLinus Torvalds struct sk_buff *buff; 13016475be16SDavid S. Miller int nsize, old_factor; 1302b60b49eaSHerbert Xu int nlen; 13039ce01461SIlpo Järvinen u8 flags; 13041da177e4SLinus Torvalds 13052fceec13SIlpo Järvinen if (WARN_ON(len > skb->len)) 13062fceec13SIlpo Järvinen return -EINVAL; 13076a438bbeSStephen Hemminger 13081da177e4SLinus Torvalds nsize = skb_headlen(skb) - len; 13091da177e4SLinus Torvalds if (nsize < 0) 13101da177e4SLinus Torvalds nsize = 0; 13111da177e4SLinus Torvalds 13126cc55e09SOctavian Purdila if (skb_unclone(skb, gfp)) 13131da177e4SLinus Torvalds return -ENOMEM; 13141da177e4SLinus Torvalds 13151da177e4SLinus Torvalds /* Get a new skb... force flag on. */ 1316eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, nsize, gfp, true); 131751456b29SIan Morris if (!buff) 13181da177e4SLinus Torvalds return -ENOMEM; /* We'll just try again later. */ 1319ef5cb973SHerbert Xu 13203ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 13213ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1322b60b49eaSHerbert Xu nlen = skb->len - len - nsize; 1323b60b49eaSHerbert Xu buff->truesize += nlen; 1324b60b49eaSHerbert Xu skb->truesize -= nlen; 13251da177e4SLinus Torvalds 13261da177e4SLinus Torvalds /* Correct the sequence numbers. */ 13271da177e4SLinus Torvalds TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 13281da177e4SLinus Torvalds TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 13291da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 13301da177e4SLinus Torvalds 13311da177e4SLinus Torvalds /* PSH and FIN should only be set in the second packet. */ 13324de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 13334de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 13344de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1335e14c3cafSHerbert Xu TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; 1336a166140eSMartin KaFai Lau tcp_skb_fragment_eor(skb, buff); 13371da177e4SLinus Torvalds 13381da177e4SLinus Torvalds skb_split(skb, buff, len); 13391da177e4SLinus Torvalds 134098be9b12SEric Dumazet buff->ip_summed = CHECKSUM_PARTIAL; 13411da177e4SLinus Torvalds 1342a61bbcf2SPatrick McHardy buff->tstamp = skb->tstamp; 1343490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 13441da177e4SLinus Torvalds 13456475be16SDavid S. Miller old_factor = tcp_skb_pcount(skb); 13466475be16SDavid S. Miller 13471da177e4SLinus Torvalds /* Fix up tso_factor for both original and new SKB. */ 13485bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 13495bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 13501da177e4SLinus Torvalds 1351b9f64820SYuchung Cheng /* Update delivered info for the new segment */ 1352b9f64820SYuchung Cheng TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx; 1353b9f64820SYuchung Cheng 13546475be16SDavid S. Miller /* If this packet has been sent out already, we must 13556475be16SDavid S. Miller * adjust the various packet counters. 13566475be16SDavid S. Miller */ 1357cf0b450cSHerbert Xu if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 13586475be16SDavid S. Miller int diff = old_factor - tcp_skb_pcount(skb) - 13596475be16SDavid S. Miller tcp_skb_pcount(buff); 13601da177e4SLinus Torvalds 1361797108d1SIlpo Järvinen if (diff) 1362797108d1SIlpo Järvinen tcp_adjust_pcount(sk, skb, diff); 13631da177e4SLinus Torvalds } 13641da177e4SLinus Torvalds 13651da177e4SLinus Torvalds /* Link BUFF into the send queue. */ 1366f4a775d1SEric Dumazet __skb_header_release(buff); 136775c119afSEric Dumazet tcp_insert_write_queue_after(skb, buff, sk, tcp_queue); 1368f67971e6SEric Dumazet if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE) 1369e2080072SEric Dumazet list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor); 13701da177e4SLinus Torvalds 13711da177e4SLinus Torvalds return 0; 13721da177e4SLinus Torvalds } 13731da177e4SLinus Torvalds 1374f4d01666SEric Dumazet /* This is similar to __pskb_pull_tail(). The difference is that pulled 1375f4d01666SEric Dumazet * data is not copied, but immediately discarded. 13761da177e4SLinus Torvalds */ 13777162fb24SEric Dumazet static int __pskb_trim_head(struct sk_buff *skb, int len) 13781da177e4SLinus Torvalds { 13797b7fc97aSEric Dumazet struct skb_shared_info *shinfo; 13801da177e4SLinus Torvalds int i, k, eat; 13811da177e4SLinus Torvalds 13824fa48bf3SEric Dumazet eat = min_t(int, len, skb_headlen(skb)); 13834fa48bf3SEric Dumazet if (eat) { 13844fa48bf3SEric Dumazet __skb_pull(skb, eat); 13854fa48bf3SEric Dumazet len -= eat; 13864fa48bf3SEric Dumazet if (!len) 13877162fb24SEric Dumazet return 0; 13884fa48bf3SEric Dumazet } 13891da177e4SLinus Torvalds eat = len; 13901da177e4SLinus Torvalds k = 0; 13917b7fc97aSEric Dumazet shinfo = skb_shinfo(skb); 13927b7fc97aSEric Dumazet for (i = 0; i < shinfo->nr_frags; i++) { 13937b7fc97aSEric Dumazet int size = skb_frag_size(&shinfo->frags[i]); 13949e903e08SEric Dumazet 13959e903e08SEric Dumazet if (size <= eat) { 1396aff65da0SIan Campbell skb_frag_unref(skb, i); 13979e903e08SEric Dumazet eat -= size; 13981da177e4SLinus Torvalds } else { 13997b7fc97aSEric Dumazet shinfo->frags[k] = shinfo->frags[i]; 14001da177e4SLinus Torvalds if (eat) { 14017b7fc97aSEric Dumazet shinfo->frags[k].page_offset += eat; 14027b7fc97aSEric Dumazet skb_frag_size_sub(&shinfo->frags[k], eat); 14031da177e4SLinus Torvalds eat = 0; 14041da177e4SLinus Torvalds } 14051da177e4SLinus Torvalds k++; 14061da177e4SLinus Torvalds } 14071da177e4SLinus Torvalds } 14087b7fc97aSEric Dumazet shinfo->nr_frags = k; 14091da177e4SLinus Torvalds 14101da177e4SLinus Torvalds skb->data_len -= len; 14111da177e4SLinus Torvalds skb->len = skb->data_len; 14127162fb24SEric Dumazet return len; 14131da177e4SLinus Torvalds } 14141da177e4SLinus Torvalds 141567edfef7SAndi Kleen /* Remove acked data from a packet in the transmit queue. */ 14161da177e4SLinus Torvalds int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 14171da177e4SLinus Torvalds { 14187162fb24SEric Dumazet u32 delta_truesize; 14197162fb24SEric Dumazet 142014bbd6a5SPravin B Shelar if (skb_unclone(skb, GFP_ATOMIC)) 14211da177e4SLinus Torvalds return -ENOMEM; 14221da177e4SLinus Torvalds 14237162fb24SEric Dumazet delta_truesize = __pskb_trim_head(skb, len); 14241da177e4SLinus Torvalds 14251da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq += len; 142684fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 14271da177e4SLinus Torvalds 14287162fb24SEric Dumazet if (delta_truesize) { 14297162fb24SEric Dumazet skb->truesize -= delta_truesize; 14307162fb24SEric Dumazet sk->sk_wmem_queued -= delta_truesize; 14317162fb24SEric Dumazet sk_mem_uncharge(sk, delta_truesize); 14321da177e4SLinus Torvalds sock_set_flag(sk, SOCK_QUEUE_SHRUNK); 14337162fb24SEric Dumazet } 14341da177e4SLinus Torvalds 14355b35e1e6SNeal Cardwell /* Any change of skb->len requires recalculation of tso factor. */ 14361da177e4SLinus Torvalds if (tcp_skb_pcount(skb) > 1) 14375bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb)); 14381da177e4SLinus Torvalds 14391da177e4SLinus Torvalds return 0; 14401da177e4SLinus Torvalds } 14411da177e4SLinus Torvalds 14421b63edd6SYuchung Cheng /* Calculate MSS not accounting any TCP options. */ 14431b63edd6SYuchung Cheng static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu) 14445d424d5aSJohn Heffner { 1445cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1446cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 14475d424d5aSJohn Heffner int mss_now; 14485d424d5aSJohn Heffner 14495d424d5aSJohn Heffner /* Calculate base mss without TCP options: 14505d424d5aSJohn Heffner It is MMS_S - sizeof(tcphdr) of rfc1122 14515d424d5aSJohn Heffner */ 14525d424d5aSJohn Heffner mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 14535d424d5aSJohn Heffner 145467469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 145567469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 145667469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 145767469601SEric Dumazet 145867469601SEric Dumazet if (dst && dst_allfrag(dst)) 145967469601SEric Dumazet mss_now -= icsk->icsk_af_ops->net_frag_header_len; 146067469601SEric Dumazet } 146167469601SEric Dumazet 14625d424d5aSJohn Heffner /* Clamp it (mss_clamp does not include tcp options) */ 14635d424d5aSJohn Heffner if (mss_now > tp->rx_opt.mss_clamp) 14645d424d5aSJohn Heffner mss_now = tp->rx_opt.mss_clamp; 14655d424d5aSJohn Heffner 14665d424d5aSJohn Heffner /* Now subtract optional transport overhead */ 14675d424d5aSJohn Heffner mss_now -= icsk->icsk_ext_hdr_len; 14685d424d5aSJohn Heffner 14695d424d5aSJohn Heffner /* Then reserve room for full set of TCP options and 8 bytes of data */ 14705d424d5aSJohn Heffner if (mss_now < 48) 14715d424d5aSJohn Heffner mss_now = 48; 14725d424d5aSJohn Heffner return mss_now; 14735d424d5aSJohn Heffner } 14745d424d5aSJohn Heffner 14751b63edd6SYuchung Cheng /* Calculate MSS. Not accounting for SACKs here. */ 14761b63edd6SYuchung Cheng int tcp_mtu_to_mss(struct sock *sk, int pmtu) 14771b63edd6SYuchung Cheng { 14781b63edd6SYuchung Cheng /* Subtract TCP options size, not including SACKs */ 14791b63edd6SYuchung Cheng return __tcp_mtu_to_mss(sk, pmtu) - 14801b63edd6SYuchung Cheng (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr)); 14811b63edd6SYuchung Cheng } 14821b63edd6SYuchung Cheng 14835d424d5aSJohn Heffner /* Inverse of above */ 148467469601SEric Dumazet int tcp_mss_to_mtu(struct sock *sk, int mss) 14855d424d5aSJohn Heffner { 1486cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1487cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 14885d424d5aSJohn Heffner int mtu; 14895d424d5aSJohn Heffner 14905d424d5aSJohn Heffner mtu = mss + 14915d424d5aSJohn Heffner tp->tcp_header_len + 14925d424d5aSJohn Heffner icsk->icsk_ext_hdr_len + 14935d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 14945d424d5aSJohn Heffner 149567469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 149667469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 149767469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 149867469601SEric Dumazet 149967469601SEric Dumazet if (dst && dst_allfrag(dst)) 150067469601SEric Dumazet mtu += icsk->icsk_af_ops->net_frag_header_len; 150167469601SEric Dumazet } 15025d424d5aSJohn Heffner return mtu; 15035d424d5aSJohn Heffner } 1504556c6b46SNeal Cardwell EXPORT_SYMBOL(tcp_mss_to_mtu); 15055d424d5aSJohn Heffner 150667edfef7SAndi Kleen /* MTU probing init per socket */ 15075d424d5aSJohn Heffner void tcp_mtup_init(struct sock *sk) 15085d424d5aSJohn Heffner { 15095d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 15105d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 1511b0f9ca53SFan Du struct net *net = sock_net(sk); 15125d424d5aSJohn Heffner 1513b0f9ca53SFan Du icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1; 15145d424d5aSJohn Heffner icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 15155d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 1516b0f9ca53SFan Du icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss); 15175d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 151805cbc0dbSFan Du if (icsk->icsk_mtup.enabled) 1519c74df29aSEric Dumazet icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; 15205d424d5aSJohn Heffner } 15214bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_mtup_init); 15225d424d5aSJohn Heffner 15231da177e4SLinus Torvalds /* This function synchronize snd mss to current pmtu/exthdr set. 15241da177e4SLinus Torvalds 15251da177e4SLinus Torvalds tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 15261da177e4SLinus Torvalds for TCP options, but includes only bare TCP header. 15271da177e4SLinus Torvalds 15281da177e4SLinus Torvalds tp->rx_opt.mss_clamp is mss negotiated at connection setup. 1529caa20d9aSStephen Hemminger It is minimum of user_mss and mss received with SYN. 15301da177e4SLinus Torvalds It also does not include TCP options. 15311da177e4SLinus Torvalds 1532d83d8461SArnaldo Carvalho de Melo inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. 15331da177e4SLinus Torvalds 15341da177e4SLinus Torvalds tp->mss_cache is current effective sending mss, including 15351da177e4SLinus Torvalds all tcp options except for SACKs. It is evaluated, 15361da177e4SLinus Torvalds taking into account current pmtu, but never exceeds 15371da177e4SLinus Torvalds tp->rx_opt.mss_clamp. 15381da177e4SLinus Torvalds 15391da177e4SLinus Torvalds NOTE1. rfc1122 clearly states that advertised MSS 15401da177e4SLinus Torvalds DOES NOT include either tcp or ip options. 15411da177e4SLinus Torvalds 1542d83d8461SArnaldo Carvalho de Melo NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache 1543d83d8461SArnaldo Carvalho de Melo are READ ONLY outside this function. --ANK (980731) 15441da177e4SLinus Torvalds */ 15451da177e4SLinus Torvalds unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 15461da177e4SLinus Torvalds { 15471da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1548d83d8461SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 15495d424d5aSJohn Heffner int mss_now; 15501da177e4SLinus Torvalds 15515d424d5aSJohn Heffner if (icsk->icsk_mtup.search_high > pmtu) 15525d424d5aSJohn Heffner icsk->icsk_mtup.search_high = pmtu; 15531da177e4SLinus Torvalds 15545d424d5aSJohn Heffner mss_now = tcp_mtu_to_mss(sk, pmtu); 1555409d22b4SIlpo Järvinen mss_now = tcp_bound_to_half_wnd(tp, mss_now); 15561da177e4SLinus Torvalds 15571da177e4SLinus Torvalds /* And store cached results */ 1558d83d8461SArnaldo Carvalho de Melo icsk->icsk_pmtu_cookie = pmtu; 15595d424d5aSJohn Heffner if (icsk->icsk_mtup.enabled) 15605d424d5aSJohn Heffner mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); 1561c1b4a7e6SDavid S. Miller tp->mss_cache = mss_now; 15621da177e4SLinus Torvalds 15631da177e4SLinus Torvalds return mss_now; 15641da177e4SLinus Torvalds } 15654bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_sync_mss); 15661da177e4SLinus Torvalds 15671da177e4SLinus Torvalds /* Compute the current effective MSS, taking SACKs and IP options, 15681da177e4SLinus Torvalds * and even PMTU discovery events into account. 15691da177e4SLinus Torvalds */ 15700c54b85fSIlpo Järvinen unsigned int tcp_current_mss(struct sock *sk) 15711da177e4SLinus Torvalds { 1572cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1573cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1574c1b4a7e6SDavid S. Miller u32 mss_now; 157595c96174SEric Dumazet unsigned int header_len; 157633ad798cSAdam Langley struct tcp_out_options opts; 157733ad798cSAdam Langley struct tcp_md5sig_key *md5; 15781da177e4SLinus Torvalds 1579c1b4a7e6SDavid S. Miller mss_now = tp->mss_cache; 1580c1b4a7e6SDavid S. Miller 15811da177e4SLinus Torvalds if (dst) { 15821da177e4SLinus Torvalds u32 mtu = dst_mtu(dst); 1583d83d8461SArnaldo Carvalho de Melo if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 15841da177e4SLinus Torvalds mss_now = tcp_sync_mss(sk, mtu); 15851da177e4SLinus Torvalds } 15861da177e4SLinus Torvalds 158733ad798cSAdam Langley header_len = tcp_established_options(sk, NULL, &opts, &md5) + 158833ad798cSAdam Langley sizeof(struct tcphdr); 158933ad798cSAdam Langley /* The mss_cache is sized based on tp->tcp_header_len, which assumes 159033ad798cSAdam Langley * some common options. If this is an odd packet (because we have SACK 159133ad798cSAdam Langley * blocks etc) then our calculated header_len will be different, and 159233ad798cSAdam Langley * we have to adjust mss_now correspondingly */ 159333ad798cSAdam Langley if (header_len != tp->tcp_header_len) { 159433ad798cSAdam Langley int delta = (int) header_len - tp->tcp_header_len; 159533ad798cSAdam Langley mss_now -= delta; 159633ad798cSAdam Langley } 1597cfb6eeb4SYOSHIFUJI Hideaki 15981da177e4SLinus Torvalds return mss_now; 15991da177e4SLinus Torvalds } 16001da177e4SLinus Torvalds 160186fd14adSWeiping Pan /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. 160286fd14adSWeiping Pan * As additional protections, we do not touch cwnd in retransmission phases, 160386fd14adSWeiping Pan * and if application hit its sndbuf limit recently. 160486fd14adSWeiping Pan */ 160586fd14adSWeiping Pan static void tcp_cwnd_application_limited(struct sock *sk) 1606a762a980SDavid S. Miller { 16079e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1608a762a980SDavid S. Miller 160986fd14adSWeiping Pan if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && 161086fd14adSWeiping Pan sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 161186fd14adSWeiping Pan /* Limited by application or receiver window. */ 161286fd14adSWeiping Pan u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); 161386fd14adSWeiping Pan u32 win_used = max(tp->snd_cwnd_used, init_win); 161486fd14adSWeiping Pan if (win_used < tp->snd_cwnd) { 161586fd14adSWeiping Pan tp->snd_ssthresh = tcp_current_ssthresh(sk); 161686fd14adSWeiping Pan tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1; 161786fd14adSWeiping Pan } 161886fd14adSWeiping Pan tp->snd_cwnd_used = 0; 161986fd14adSWeiping Pan } 1620c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 162186fd14adSWeiping Pan } 162286fd14adSWeiping Pan 1623ca8a2263SNeal Cardwell static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited) 1624a762a980SDavid S. Miller { 16251b1fc3fdSWei Wang const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 1626a762a980SDavid S. Miller struct tcp_sock *tp = tcp_sk(sk); 1627a762a980SDavid S. Miller 1628ca8a2263SNeal Cardwell /* Track the maximum number of outstanding packets in each 1629ca8a2263SNeal Cardwell * window, and remember whether we were cwnd-limited then. 1630ca8a2263SNeal Cardwell */ 1631ca8a2263SNeal Cardwell if (!before(tp->snd_una, tp->max_packets_seq) || 1632ca8a2263SNeal Cardwell tp->packets_out > tp->max_packets_out) { 1633ca8a2263SNeal Cardwell tp->max_packets_out = tp->packets_out; 1634ca8a2263SNeal Cardwell tp->max_packets_seq = tp->snd_nxt; 1635ca8a2263SNeal Cardwell tp->is_cwnd_limited = is_cwnd_limited; 1636ca8a2263SNeal Cardwell } 1637e114a710SEric Dumazet 163824901551SEric Dumazet if (tcp_is_cwnd_limited(sk)) { 1639a762a980SDavid S. Miller /* Network is feed fully. */ 1640a762a980SDavid S. Miller tp->snd_cwnd_used = 0; 1641c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 1642a762a980SDavid S. Miller } else { 1643a762a980SDavid S. Miller /* Network starves. */ 1644a762a980SDavid S. Miller if (tp->packets_out > tp->snd_cwnd_used) 1645a762a980SDavid S. Miller tp->snd_cwnd_used = tp->packets_out; 1646a762a980SDavid S. Miller 1647b510f0d2SEric Dumazet if (sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle && 1648c2203cf7SEric Dumazet (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto && 16491b1fc3fdSWei Wang !ca_ops->cong_control) 1650a762a980SDavid S. Miller tcp_cwnd_application_limited(sk); 1651b0f71bd3SFrancis Yan 1652b0f71bd3SFrancis Yan /* The following conditions together indicate the starvation 1653b0f71bd3SFrancis Yan * is caused by insufficient sender buffer: 1654b0f71bd3SFrancis Yan * 1) just sent some data (see tcp_write_xmit) 1655b0f71bd3SFrancis Yan * 2) not cwnd limited (this else condition) 165675c119afSEric Dumazet * 3) no more data to send (tcp_write_queue_empty()) 1657b0f71bd3SFrancis Yan * 4) application is hitting buffer limit (SOCK_NOSPACE) 1658b0f71bd3SFrancis Yan */ 165975c119afSEric Dumazet if (tcp_write_queue_empty(sk) && sk->sk_socket && 1660b0f71bd3SFrancis Yan test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) && 1661b0f71bd3SFrancis Yan (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) 1662b0f71bd3SFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED); 1663a762a980SDavid S. Miller } 1664a762a980SDavid S. Miller } 1665a762a980SDavid S. Miller 1666d4589926SEric Dumazet /* Minshall's variant of the Nagle send check. */ 1667d4589926SEric Dumazet static bool tcp_minshall_check(const struct tcp_sock *tp) 1668d4589926SEric Dumazet { 1669d4589926SEric Dumazet return after(tp->snd_sml, tp->snd_una) && 1670d4589926SEric Dumazet !after(tp->snd_sml, tp->snd_nxt); 1671d4589926SEric Dumazet } 1672d4589926SEric Dumazet 1673d4589926SEric Dumazet /* Update snd_sml if this skb is under mss 1674d4589926SEric Dumazet * Note that a TSO packet might end with a sub-mss segment 1675d4589926SEric Dumazet * The test is really : 1676d4589926SEric Dumazet * if ((skb->len % mss) != 0) 1677d4589926SEric Dumazet * tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1678d4589926SEric Dumazet * But we can avoid doing the divide again given we already have 1679d4589926SEric Dumazet * skb_pcount = skb->len / mss_now 16800e3a4803SIlpo Järvinen */ 1681d4589926SEric Dumazet static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now, 1682d4589926SEric Dumazet const struct sk_buff *skb) 1683d4589926SEric Dumazet { 1684d4589926SEric Dumazet if (skb->len < tcp_skb_pcount(skb) * mss_now) 1685d4589926SEric Dumazet tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1686d4589926SEric Dumazet } 1687d4589926SEric Dumazet 1688d4589926SEric Dumazet /* Return false, if packet can be sent now without violation Nagle's rules: 1689d4589926SEric Dumazet * 1. It is full sized. (provided by caller in %partial bool) 1690d4589926SEric Dumazet * 2. Or it contains FIN. (already checked by caller) 1691d4589926SEric Dumazet * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. 1692d4589926SEric Dumazet * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 1693d4589926SEric Dumazet * With Minshall's modification: all sent small packets are ACKed. 1694d4589926SEric Dumazet */ 1695d4589926SEric Dumazet static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp, 1696cc93fc51SPeter Pan(潘卫平) int nonagle) 1697d4589926SEric Dumazet { 1698d4589926SEric Dumazet return partial && 1699d4589926SEric Dumazet ((nonagle & TCP_NAGLE_CORK) || 1700d4589926SEric Dumazet (!nonagle && tp->packets_out && tcp_minshall_check(tp))); 1701d4589926SEric Dumazet } 1702605ad7f1SEric Dumazet 1703605ad7f1SEric Dumazet /* Return how many segs we'd like on a TSO packet, 1704605ad7f1SEric Dumazet * to send one TSO packet per ms 1705605ad7f1SEric Dumazet */ 1706*dcb8c9b4SEric Dumazet static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now, 17071b3878caSNeal Cardwell int min_tso_segs) 1708605ad7f1SEric Dumazet { 1709605ad7f1SEric Dumazet u32 bytes, segs; 1710605ad7f1SEric Dumazet 17113a9b76fdSEric Dumazet bytes = min(sk->sk_pacing_rate >> sk->sk_pacing_shift, 1712605ad7f1SEric Dumazet sk->sk_gso_max_size - 1 - MAX_TCP_HEADER); 1713605ad7f1SEric Dumazet 1714605ad7f1SEric Dumazet /* Goal is to send at least one packet per ms, 1715605ad7f1SEric Dumazet * not one big TSO packet every 100 ms. 1716605ad7f1SEric Dumazet * This preserves ACK clocking and is consistent 1717605ad7f1SEric Dumazet * with tcp_tso_should_defer() heuristic. 1718605ad7f1SEric Dumazet */ 17191b3878caSNeal Cardwell segs = max_t(u32, bytes / mss_now, min_tso_segs); 1720605ad7f1SEric Dumazet 1721350c9f48SEric Dumazet return segs; 1722605ad7f1SEric Dumazet } 1723605ad7f1SEric Dumazet 1724ed6e7268SNeal Cardwell /* Return the number of segments we want in the skb we are transmitting. 1725ed6e7268SNeal Cardwell * See if congestion control module wants to decide; otherwise, autosize. 1726ed6e7268SNeal Cardwell */ 1727ed6e7268SNeal Cardwell static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now) 1728ed6e7268SNeal Cardwell { 1729ed6e7268SNeal Cardwell const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 1730*dcb8c9b4SEric Dumazet u32 min_tso, tso_segs; 1731ed6e7268SNeal Cardwell 1732*dcb8c9b4SEric Dumazet min_tso = ca_ops->min_tso_segs ? 1733*dcb8c9b4SEric Dumazet ca_ops->min_tso_segs(sk) : 1734*dcb8c9b4SEric Dumazet sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs; 1735*dcb8c9b4SEric Dumazet 1736*dcb8c9b4SEric Dumazet tso_segs = tcp_tso_autosize(sk, mss_now, min_tso); 1737350c9f48SEric Dumazet return min_t(u32, tso_segs, sk->sk_gso_max_segs); 1738ed6e7268SNeal Cardwell } 1739ed6e7268SNeal Cardwell 1740d4589926SEric Dumazet /* Returns the portion of skb which can be sent right away */ 1741d4589926SEric Dumazet static unsigned int tcp_mss_split_point(const struct sock *sk, 1742d4589926SEric Dumazet const struct sk_buff *skb, 1743d4589926SEric Dumazet unsigned int mss_now, 1744d4589926SEric Dumazet unsigned int max_segs, 1745d4589926SEric Dumazet int nonagle) 1746c1b4a7e6SDavid S. Miller { 1747cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1748d4589926SEric Dumazet u32 partial, needed, window, max_len; 1749c1b4a7e6SDavid S. Miller 175090840defSIlpo Järvinen window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 17511485348dSBen Hutchings max_len = mss_now * max_segs; 17520e3a4803SIlpo Järvinen 17531485348dSBen Hutchings if (likely(max_len <= window && skb != tcp_write_queue_tail(sk))) 17541485348dSBen Hutchings return max_len; 17550e3a4803SIlpo Järvinen 17565ea3a748SIlpo Järvinen needed = min(skb->len, window); 17575ea3a748SIlpo Järvinen 17581485348dSBen Hutchings if (max_len <= needed) 17591485348dSBen Hutchings return max_len; 17600e3a4803SIlpo Järvinen 1761d4589926SEric Dumazet partial = needed % mss_now; 1762d4589926SEric Dumazet /* If last segment is not a full MSS, check if Nagle rules allow us 1763d4589926SEric Dumazet * to include this last segment in this skb. 1764d4589926SEric Dumazet * Otherwise, we'll split the skb at last MSS boundary 1765d4589926SEric Dumazet */ 1766cc93fc51SPeter Pan(潘卫平) if (tcp_nagle_check(partial != 0, tp, nonagle)) 1767d4589926SEric Dumazet return needed - partial; 1768d4589926SEric Dumazet 1769d4589926SEric Dumazet return needed; 1770c1b4a7e6SDavid S. Miller } 1771c1b4a7e6SDavid S. Miller 1772c1b4a7e6SDavid S. Miller /* Can at least one segment of SKB be sent right now, according to the 1773c1b4a7e6SDavid S. Miller * congestion window rules? If so, return how many segments are allowed. 1774c1b4a7e6SDavid S. Miller */ 1775cf533ea5SEric Dumazet static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, 1776cf533ea5SEric Dumazet const struct sk_buff *skb) 1777c1b4a7e6SDavid S. Miller { 1778d649a7a8SEric Dumazet u32 in_flight, cwnd, halfcwnd; 1779c1b4a7e6SDavid S. Miller 1780c1b4a7e6SDavid S. Miller /* Don't be strict about the congestion window for the final FIN. */ 17814de075e0SEric Dumazet if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && 17824de075e0SEric Dumazet tcp_skb_pcount(skb) == 1) 1783c1b4a7e6SDavid S. Miller return 1; 1784c1b4a7e6SDavid S. Miller 1785c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1786c1b4a7e6SDavid S. Miller cwnd = tp->snd_cwnd; 1787d649a7a8SEric Dumazet if (in_flight >= cwnd) 1788c1b4a7e6SDavid S. Miller return 0; 1789d649a7a8SEric Dumazet 1790d649a7a8SEric Dumazet /* For better scheduling, ensure we have at least 1791d649a7a8SEric Dumazet * 2 GSO packets in flight. 1792d649a7a8SEric Dumazet */ 1793d649a7a8SEric Dumazet halfcwnd = max(cwnd >> 1, 1U); 1794d649a7a8SEric Dumazet return min(halfcwnd, cwnd - in_flight); 1795c1b4a7e6SDavid S. Miller } 1796c1b4a7e6SDavid S. Miller 1797b595076aSUwe Kleine-König /* Initialize TSO state of a skb. 179867edfef7SAndi Kleen * This must be invoked the first time we consider transmitting 1799c1b4a7e6SDavid S. Miller * SKB onto the wire. 1800c1b4a7e6SDavid S. Miller */ 18015bbb432cSEric Dumazet static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now) 1802c1b4a7e6SDavid S. Miller { 1803c1b4a7e6SDavid S. Miller int tso_segs = tcp_skb_pcount(skb); 1804c1b4a7e6SDavid S. Miller 1805f8269a49SIlpo Järvinen if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { 18065bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 1807c1b4a7e6SDavid S. Miller tso_segs = tcp_skb_pcount(skb); 1808c1b4a7e6SDavid S. Miller } 1809c1b4a7e6SDavid S. Miller return tso_segs; 1810c1b4a7e6SDavid S. Miller } 1811c1b4a7e6SDavid S. Miller 1812c1b4a7e6SDavid S. Miller 1813a2a385d6SEric Dumazet /* Return true if the Nagle test allows this packet to be 1814c1b4a7e6SDavid S. Miller * sent now. 1815c1b4a7e6SDavid S. Miller */ 1816a2a385d6SEric Dumazet static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, 1817c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1818c1b4a7e6SDavid S. Miller { 1819c1b4a7e6SDavid S. Miller /* Nagle rule does not apply to frames, which sit in the middle of the 1820c1b4a7e6SDavid S. Miller * write_queue (they have no chances to get new data). 1821c1b4a7e6SDavid S. Miller * 1822c1b4a7e6SDavid S. Miller * This is implemented in the callers, where they modify the 'nonagle' 1823c1b4a7e6SDavid S. Miller * argument based upon the location of SKB in the send queue. 1824c1b4a7e6SDavid S. Miller */ 1825c1b4a7e6SDavid S. Miller if (nonagle & TCP_NAGLE_PUSH) 1826a2a385d6SEric Dumazet return true; 1827c1b4a7e6SDavid S. Miller 18289b44190dSYuchung Cheng /* Don't use the nagle rule for urgent data (or for the final FIN). */ 18299b44190dSYuchung Cheng if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) 1830a2a385d6SEric Dumazet return true; 1831c1b4a7e6SDavid S. Miller 1832cc93fc51SPeter Pan(潘卫平) if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle)) 1833a2a385d6SEric Dumazet return true; 1834c1b4a7e6SDavid S. Miller 1835a2a385d6SEric Dumazet return false; 1836c1b4a7e6SDavid S. Miller } 1837c1b4a7e6SDavid S. Miller 1838c1b4a7e6SDavid S. Miller /* Does at least the first segment of SKB fit into the send window? */ 1839a2a385d6SEric Dumazet static bool tcp_snd_wnd_test(const struct tcp_sock *tp, 1840a2a385d6SEric Dumazet const struct sk_buff *skb, 1841056834d9SIlpo Järvinen unsigned int cur_mss) 1842c1b4a7e6SDavid S. Miller { 1843c1b4a7e6SDavid S. Miller u32 end_seq = TCP_SKB_CB(skb)->end_seq; 1844c1b4a7e6SDavid S. Miller 1845c1b4a7e6SDavid S. Miller if (skb->len > cur_mss) 1846c1b4a7e6SDavid S. Miller end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 1847c1b4a7e6SDavid S. Miller 184890840defSIlpo Järvinen return !after(end_seq, tcp_wnd_end(tp)); 1849c1b4a7e6SDavid S. Miller } 1850c1b4a7e6SDavid S. Miller 1851c1b4a7e6SDavid S. Miller /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 1852c1b4a7e6SDavid S. Miller * which is put after SKB on the list. It is very much like 1853c1b4a7e6SDavid S. Miller * tcp_fragment() except that it may make several kinds of assumptions 1854c1b4a7e6SDavid S. Miller * in order to speed up the splitting operation. In particular, we 1855c1b4a7e6SDavid S. Miller * know that all the data is in scatter-gather pages, and that the 1856c1b4a7e6SDavid S. Miller * packet has never been sent out before (and thus is not cloned). 1857c1b4a7e6SDavid S. Miller */ 185875c119afSEric Dumazet static int tso_fragment(struct sock *sk, enum tcp_queue tcp_queue, 185975c119afSEric Dumazet struct sk_buff *skb, unsigned int len, 1860c4ead4c5SEric Dumazet unsigned int mss_now, gfp_t gfp) 1861c1b4a7e6SDavid S. Miller { 1862c1b4a7e6SDavid S. Miller struct sk_buff *buff; 1863c1b4a7e6SDavid S. Miller int nlen = skb->len - len; 18649ce01461SIlpo Järvinen u8 flags; 1865c1b4a7e6SDavid S. Miller 1866c1b4a7e6SDavid S. Miller /* All of a TSO frame must be composed of paged data. */ 1867c8ac3774SHerbert Xu if (skb->len != skb->data_len) 186875c119afSEric Dumazet return tcp_fragment(sk, tcp_queue, skb, len, mss_now, gfp); 1869c1b4a7e6SDavid S. Miller 1870eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, gfp, true); 187151456b29SIan Morris if (unlikely(!buff)) 1872c1b4a7e6SDavid S. Miller return -ENOMEM; 1873c1b4a7e6SDavid S. Miller 18743ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 18753ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1876b60b49eaSHerbert Xu buff->truesize += nlen; 1877c1b4a7e6SDavid S. Miller skb->truesize -= nlen; 1878c1b4a7e6SDavid S. Miller 1879c1b4a7e6SDavid S. Miller /* Correct the sequence numbers. */ 1880c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 1881c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 1882c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 1883c1b4a7e6SDavid S. Miller 1884c1b4a7e6SDavid S. Miller /* PSH and FIN should only be set in the second packet. */ 18854de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 18864de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 18874de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1888c1b4a7e6SDavid S. Miller 1889c1b4a7e6SDavid S. Miller /* This packet was never sent out yet, so no SACK bits. */ 1890c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->sacked = 0; 1891c1b4a7e6SDavid S. Miller 1892a166140eSMartin KaFai Lau tcp_skb_fragment_eor(skb, buff); 1893a166140eSMartin KaFai Lau 189498be9b12SEric Dumazet buff->ip_summed = CHECKSUM_PARTIAL; 1895c1b4a7e6SDavid S. Miller skb_split(skb, buff, len); 1896490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 1897c1b4a7e6SDavid S. Miller 1898c1b4a7e6SDavid S. Miller /* Fix up tso_factor for both original and new SKB. */ 18995bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 19005bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 1901c1b4a7e6SDavid S. Miller 1902c1b4a7e6SDavid S. Miller /* Link BUFF into the send queue. */ 1903f4a775d1SEric Dumazet __skb_header_release(buff); 190475c119afSEric Dumazet tcp_insert_write_queue_after(skb, buff, sk, tcp_queue); 1905c1b4a7e6SDavid S. Miller 1906c1b4a7e6SDavid S. Miller return 0; 1907c1b4a7e6SDavid S. Miller } 1908c1b4a7e6SDavid S. Miller 1909c1b4a7e6SDavid S. Miller /* Try to defer sending, if possible, in order to minimize the amount 1910c1b4a7e6SDavid S. Miller * of TSO splitting we do. View it as a kind of TSO Nagle test. 1911c1b4a7e6SDavid S. Miller * 1912c1b4a7e6SDavid S. Miller * This algorithm is from John Heffner. 1913c1b4a7e6SDavid S. Miller */ 1914ca8a2263SNeal Cardwell static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb, 1915605ad7f1SEric Dumazet bool *is_cwnd_limited, u32 max_segs) 1916c1b4a7e6SDavid S. Miller { 19176687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 191850c8339eSEric Dumazet u32 age, send_win, cong_win, limit, in_flight; 191950c8339eSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 192050c8339eSEric Dumazet struct sk_buff *head; 1921ad9f4f50SEric Dumazet int win_divisor; 1922c1b4a7e6SDavid S. Miller 19234de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 1924ae8064acSJohn Heffner goto send_now; 1925c1b4a7e6SDavid S. Miller 192699d7662aSEric Dumazet if (icsk->icsk_ca_state >= TCP_CA_Recovery) 1927ae8064acSJohn Heffner goto send_now; 1928ae8064acSJohn Heffner 19295f852eb5SEric Dumazet /* Avoid bursty behavior by allowing defer 19305f852eb5SEric Dumazet * only if the last write was recent. 19315f852eb5SEric Dumazet */ 1932d635fbe2SEric Dumazet if ((s32)(tcp_jiffies32 - tp->lsndtime) > 0) 1933ae8064acSJohn Heffner goto send_now; 1934908a75c1SDavid S. Miller 1935c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1936c1b4a7e6SDavid S. Miller 1937c8c9aeb5SStefano Brivio BUG_ON(tcp_skb_pcount(skb) <= 1); 1938c8c9aeb5SStefano Brivio BUG_ON(tp->snd_cwnd <= in_flight); 1939c1b4a7e6SDavid S. Miller 194090840defSIlpo Järvinen send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1941c1b4a7e6SDavid S. Miller 1942c1b4a7e6SDavid S. Miller /* From in_flight test above, we know that cwnd > in_flight. */ 1943c1b4a7e6SDavid S. Miller cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; 1944c1b4a7e6SDavid S. Miller 1945c1b4a7e6SDavid S. Miller limit = min(send_win, cong_win); 1946c1b4a7e6SDavid S. Miller 1947ba244fe9SDavid S. Miller /* If a full-sized TSO skb can be sent, do it. */ 1948605ad7f1SEric Dumazet if (limit >= max_segs * tp->mss_cache) 1949ae8064acSJohn Heffner goto send_now; 1950ba244fe9SDavid S. Miller 195162ad2761SIlpo Järvinen /* Middle in queue won't get any more data, full sendable already? */ 195262ad2761SIlpo Järvinen if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) 195362ad2761SIlpo Järvinen goto send_now; 195462ad2761SIlpo Järvinen 19555bbcc0f5SLinus Torvalds win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor); 1956ad9f4f50SEric Dumazet if (win_divisor) { 1957c1b4a7e6SDavid S. Miller u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); 1958c1b4a7e6SDavid S. Miller 1959c1b4a7e6SDavid S. Miller /* If at least some fraction of a window is available, 1960c1b4a7e6SDavid S. Miller * just use it. 1961c1b4a7e6SDavid S. Miller */ 1962ad9f4f50SEric Dumazet chunk /= win_divisor; 1963c1b4a7e6SDavid S. Miller if (limit >= chunk) 1964ae8064acSJohn Heffner goto send_now; 1965c1b4a7e6SDavid S. Miller } else { 1966c1b4a7e6SDavid S. Miller /* Different approach, try not to defer past a single 1967c1b4a7e6SDavid S. Miller * ACK. Receiver should ACK every other full sized 1968c1b4a7e6SDavid S. Miller * frame, so if we have space for more than 3 frames 1969c1b4a7e6SDavid S. Miller * then send now. 1970c1b4a7e6SDavid S. Miller */ 19716b5a5c0dSNeal Cardwell if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) 1972ae8064acSJohn Heffner goto send_now; 1973c1b4a7e6SDavid S. Miller } 1974c1b4a7e6SDavid S. Miller 197575c119afSEric Dumazet /* TODO : use tsorted_sent_queue ? */ 197675c119afSEric Dumazet head = tcp_rtx_queue_head(sk); 197775c119afSEric Dumazet if (!head) 197875c119afSEric Dumazet goto send_now; 19799a568de4SEric Dumazet age = tcp_stamp_us_delta(tp->tcp_mstamp, head->skb_mstamp); 198050c8339eSEric Dumazet /* If next ACK is likely to come too late (half srtt), do not defer */ 198150c8339eSEric Dumazet if (age < (tp->srtt_us >> 4)) 198250c8339eSEric Dumazet goto send_now; 198350c8339eSEric Dumazet 19845f852eb5SEric Dumazet /* Ok, it looks like it is advisable to defer. */ 1985ae8064acSJohn Heffner 1986d2e1339fSBendik Rønning Opstad if (cong_win < send_win && cong_win <= skb->len) 1987ca8a2263SNeal Cardwell *is_cwnd_limited = true; 1988ca8a2263SNeal Cardwell 1989a2a385d6SEric Dumazet return true; 1990ae8064acSJohn Heffner 1991ae8064acSJohn Heffner send_now: 1992a2a385d6SEric Dumazet return false; 1993c1b4a7e6SDavid S. Miller } 1994c1b4a7e6SDavid S. Miller 199505cbc0dbSFan Du static inline void tcp_mtu_check_reprobe(struct sock *sk) 199605cbc0dbSFan Du { 199705cbc0dbSFan Du struct inet_connection_sock *icsk = inet_csk(sk); 199805cbc0dbSFan Du struct tcp_sock *tp = tcp_sk(sk); 199905cbc0dbSFan Du struct net *net = sock_net(sk); 200005cbc0dbSFan Du u32 interval; 200105cbc0dbSFan Du s32 delta; 200205cbc0dbSFan Du 200305cbc0dbSFan Du interval = net->ipv4.sysctl_tcp_probe_interval; 2004c74df29aSEric Dumazet delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp; 200505cbc0dbSFan Du if (unlikely(delta >= interval * HZ)) { 200605cbc0dbSFan Du int mss = tcp_current_mss(sk); 200705cbc0dbSFan Du 200805cbc0dbSFan Du /* Update current search range */ 200905cbc0dbSFan Du icsk->icsk_mtup.probe_size = 0; 201005cbc0dbSFan Du icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + 201105cbc0dbSFan Du sizeof(struct tcphdr) + 201205cbc0dbSFan Du icsk->icsk_af_ops->net_header_len; 201305cbc0dbSFan Du icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss); 201405cbc0dbSFan Du 201505cbc0dbSFan Du /* Update probe time stamp */ 2016c74df29aSEric Dumazet icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; 201705cbc0dbSFan Du } 201805cbc0dbSFan Du } 201905cbc0dbSFan Du 2020808cf9e3SIlya Lesokhin static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len) 2021808cf9e3SIlya Lesokhin { 2022808cf9e3SIlya Lesokhin struct sk_buff *skb, *next; 2023808cf9e3SIlya Lesokhin 2024808cf9e3SIlya Lesokhin skb = tcp_send_head(sk); 2025808cf9e3SIlya Lesokhin tcp_for_write_queue_from_safe(skb, next, sk) { 2026808cf9e3SIlya Lesokhin if (len <= skb->len) 2027808cf9e3SIlya Lesokhin break; 2028808cf9e3SIlya Lesokhin 2029808cf9e3SIlya Lesokhin if (unlikely(TCP_SKB_CB(skb)->eor)) 2030808cf9e3SIlya Lesokhin return false; 2031808cf9e3SIlya Lesokhin 2032808cf9e3SIlya Lesokhin len -= skb->len; 2033808cf9e3SIlya Lesokhin } 2034808cf9e3SIlya Lesokhin 2035808cf9e3SIlya Lesokhin return true; 2036808cf9e3SIlya Lesokhin } 2037808cf9e3SIlya Lesokhin 20385d424d5aSJohn Heffner /* Create a new MTU probe if we are ready. 203967edfef7SAndi Kleen * MTU probe is regularly attempting to increase the path MTU by 204067edfef7SAndi Kleen * deliberately sending larger packets. This discovers routing 204167edfef7SAndi Kleen * changes resulting in larger path MTUs. 204267edfef7SAndi Kleen * 20435d424d5aSJohn Heffner * Returns 0 if we should wait to probe (no cwnd available), 20445d424d5aSJohn Heffner * 1 if a probe was sent, 2045056834d9SIlpo Järvinen * -1 otherwise 2046056834d9SIlpo Järvinen */ 20475d424d5aSJohn Heffner static int tcp_mtu_probe(struct sock *sk) 20485d424d5aSJohn Heffner { 20495d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 205012a59abcSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 20515d424d5aSJohn Heffner struct sk_buff *skb, *nskb, *next; 20526b58e0a5SFan Du struct net *net = sock_net(sk); 20535d424d5aSJohn Heffner int probe_size; 205491cc17c0SIlpo Järvinen int size_needed; 205512a59abcSEric Dumazet int copy, len; 20565d424d5aSJohn Heffner int mss_now; 20576b58e0a5SFan Du int interval; 20585d424d5aSJohn Heffner 20595d424d5aSJohn Heffner /* Not currently probing/verifying, 20605d424d5aSJohn Heffner * not in recovery, 20615d424d5aSJohn Heffner * have enough cwnd, and 206212a59abcSEric Dumazet * not SACKing (the variable headers throw things off) 206312a59abcSEric Dumazet */ 206412a59abcSEric Dumazet if (likely(!icsk->icsk_mtup.enabled || 20655d424d5aSJohn Heffner icsk->icsk_mtup.probe_size || 20665d424d5aSJohn Heffner inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 20675d424d5aSJohn Heffner tp->snd_cwnd < 11 || 206812a59abcSEric Dumazet tp->rx_opt.num_sacks || tp->rx_opt.dsack)) 20695d424d5aSJohn Heffner return -1; 20705d424d5aSJohn Heffner 20716b58e0a5SFan Du /* Use binary search for probe_size between tcp_mss_base, 20726b58e0a5SFan Du * and current mss_clamp. if (search_high - search_low) 20736b58e0a5SFan Du * smaller than a threshold, backoff from probing. 20746b58e0a5SFan Du */ 20750c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 20766b58e0a5SFan Du probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high + 20776b58e0a5SFan Du icsk->icsk_mtup.search_low) >> 1); 207891cc17c0SIlpo Järvinen size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; 20796b58e0a5SFan Du interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low; 208005cbc0dbSFan Du /* When misfortune happens, we are reprobing actively, 208105cbc0dbSFan Du * and then reprobe timer has expired. We stick with current 208205cbc0dbSFan Du * probing process by not resetting search range to its orignal. 208305cbc0dbSFan Du */ 20846b58e0a5SFan Du if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) || 208505cbc0dbSFan Du interval < net->ipv4.sysctl_tcp_probe_threshold) { 208605cbc0dbSFan Du /* Check whether enough time has elaplased for 208705cbc0dbSFan Du * another round of probing. 208805cbc0dbSFan Du */ 208905cbc0dbSFan Du tcp_mtu_check_reprobe(sk); 20905d424d5aSJohn Heffner return -1; 20915d424d5aSJohn Heffner } 20925d424d5aSJohn Heffner 20935d424d5aSJohn Heffner /* Have enough data in the send queue to probe? */ 20947f9c33e5SIlpo Järvinen if (tp->write_seq - tp->snd_nxt < size_needed) 20955d424d5aSJohn Heffner return -1; 20965d424d5aSJohn Heffner 209791cc17c0SIlpo Järvinen if (tp->snd_wnd < size_needed) 20985d424d5aSJohn Heffner return -1; 209990840defSIlpo Järvinen if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) 21005d424d5aSJohn Heffner return 0; 21015d424d5aSJohn Heffner 2102d67c58e9SIlpo Järvinen /* Do we need to wait to drain cwnd? With none in flight, don't stall */ 2103d67c58e9SIlpo Järvinen if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { 2104d67c58e9SIlpo Järvinen if (!tcp_packets_in_flight(tp)) 21055d424d5aSJohn Heffner return -1; 21065d424d5aSJohn Heffner else 21075d424d5aSJohn Heffner return 0; 21085d424d5aSJohn Heffner } 21095d424d5aSJohn Heffner 2110808cf9e3SIlya Lesokhin if (!tcp_can_coalesce_send_queue_head(sk, probe_size)) 2111808cf9e3SIlya Lesokhin return -1; 2112808cf9e3SIlya Lesokhin 21135d424d5aSJohn Heffner /* We're allowed to probe. Build it now. */ 2114eb934478SEric Dumazet nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false); 211551456b29SIan Morris if (!nskb) 21165d424d5aSJohn Heffner return -1; 21173ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 21183ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 21195d424d5aSJohn Heffner 2120fe067e8aSDavid S. Miller skb = tcp_send_head(sk); 21215d424d5aSJohn Heffner 21225d424d5aSJohn Heffner TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 21235d424d5aSJohn Heffner TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 21244de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; 21255d424d5aSJohn Heffner TCP_SKB_CB(nskb)->sacked = 0; 21265d424d5aSJohn Heffner nskb->csum = 0; 212798be9b12SEric Dumazet nskb->ip_summed = CHECKSUM_PARTIAL; 21285d424d5aSJohn Heffner 212950c4817eSIlpo Järvinen tcp_insert_write_queue_before(nskb, skb, sk); 21302b7cda9cSEric Dumazet tcp_highest_sack_replace(sk, skb, nskb); 213150c4817eSIlpo Järvinen 21325d424d5aSJohn Heffner len = 0; 2133234b6860SIlpo Järvinen tcp_for_write_queue_from_safe(skb, next, sk) { 21345d424d5aSJohn Heffner copy = min_t(int, skb->len, probe_size - len); 21355d424d5aSJohn Heffner skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); 21365d424d5aSJohn Heffner 21375d424d5aSJohn Heffner if (skb->len <= copy) { 21385d424d5aSJohn Heffner /* We've eaten all the data from this skb. 21395d424d5aSJohn Heffner * Throw it away. */ 21404de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 2141808cf9e3SIlya Lesokhin /* If this is the last SKB we copy and eor is set 2142808cf9e3SIlya Lesokhin * we need to propagate it to the new skb. 2143808cf9e3SIlya Lesokhin */ 2144808cf9e3SIlya Lesokhin TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor; 2145fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 21463ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 21475d424d5aSJohn Heffner } else { 21484de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & 2149a3433f35SChangli Gao ~(TCPHDR_FIN|TCPHDR_PSH); 21505d424d5aSJohn Heffner if (!skb_shinfo(skb)->nr_frags) { 21515d424d5aSJohn Heffner skb_pull(skb, copy); 21525d424d5aSJohn Heffner } else { 21535d424d5aSJohn Heffner __pskb_trim_head(skb, copy); 21545bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 21555d424d5aSJohn Heffner } 21565d424d5aSJohn Heffner TCP_SKB_CB(skb)->seq += copy; 21575d424d5aSJohn Heffner } 21585d424d5aSJohn Heffner 21595d424d5aSJohn Heffner len += copy; 2160234b6860SIlpo Järvinen 2161234b6860SIlpo Järvinen if (len >= probe_size) 2162234b6860SIlpo Järvinen break; 21635d424d5aSJohn Heffner } 21645bbb432cSEric Dumazet tcp_init_tso_segs(nskb, nskb->len); 21655d424d5aSJohn Heffner 21665d424d5aSJohn Heffner /* We're ready to send. If this fails, the probe will 21677faee5c0SEric Dumazet * be resegmented into mss-sized pieces by tcp_write_xmit(). 21687faee5c0SEric Dumazet */ 21695d424d5aSJohn Heffner if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 21705d424d5aSJohn Heffner /* Decrement cwnd here because we are sending 21715d424d5aSJohn Heffner * effectively two packets. */ 21725d424d5aSJohn Heffner tp->snd_cwnd--; 217366f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, nskb); 21745d424d5aSJohn Heffner 21755d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 21760e7b1368SJohn Heffner tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 21770e7b1368SJohn Heffner tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 21785d424d5aSJohn Heffner 21795d424d5aSJohn Heffner return 1; 21805d424d5aSJohn Heffner } 21815d424d5aSJohn Heffner 21825d424d5aSJohn Heffner return -1; 21835d424d5aSJohn Heffner } 21845d424d5aSJohn Heffner 2185218af599SEric Dumazet static bool tcp_pacing_check(const struct sock *sk) 2186218af599SEric Dumazet { 2187218af599SEric Dumazet return tcp_needs_internal_pacing(sk) && 2188218af599SEric Dumazet hrtimer_active(&tcp_sk(sk)->pacing_timer); 2189218af599SEric Dumazet } 2190218af599SEric Dumazet 2191f9616c35SEric Dumazet /* TCP Small Queues : 2192f9616c35SEric Dumazet * Control number of packets in qdisc/devices to two packets / or ~1 ms. 2193f9616c35SEric Dumazet * (These limits are doubled for retransmits) 2194f9616c35SEric Dumazet * This allows for : 2195f9616c35SEric Dumazet * - better RTT estimation and ACK scheduling 2196f9616c35SEric Dumazet * - faster recovery 2197f9616c35SEric Dumazet * - high rates 2198f9616c35SEric Dumazet * Alas, some drivers / subsystems require a fair amount 2199f9616c35SEric Dumazet * of queued bytes to ensure line rate. 2200f9616c35SEric Dumazet * One example is wifi aggregation (802.11 AMPDU) 2201f9616c35SEric Dumazet */ 2202f9616c35SEric Dumazet static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb, 2203f9616c35SEric Dumazet unsigned int factor) 2204f9616c35SEric Dumazet { 2205f9616c35SEric Dumazet unsigned int limit; 2206f9616c35SEric Dumazet 22073a9b76fdSEric Dumazet limit = max(2 * skb->truesize, sk->sk_pacing_rate >> sk->sk_pacing_shift); 22089184d8bbSEric Dumazet limit = min_t(u32, limit, 22099184d8bbSEric Dumazet sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes); 2210f9616c35SEric Dumazet limit <<= factor; 2211f9616c35SEric Dumazet 221214afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) > limit) { 221375c119afSEric Dumazet /* Always send skb if rtx queue is empty. 221475eefc6cSEric Dumazet * No need to wait for TX completion to call us back, 221575eefc6cSEric Dumazet * after softirq/tasklet schedule. 221675eefc6cSEric Dumazet * This helps when TX completions are delayed too much. 221775eefc6cSEric Dumazet */ 221875c119afSEric Dumazet if (tcp_rtx_queue_empty(sk)) 221975eefc6cSEric Dumazet return false; 222075eefc6cSEric Dumazet 22217aa5470cSEric Dumazet set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags); 2222f9616c35SEric Dumazet /* It is possible TX completion already happened 2223f9616c35SEric Dumazet * before we set TSQ_THROTTLED, so we must 2224f9616c35SEric Dumazet * test again the condition. 2225f9616c35SEric Dumazet */ 2226f9616c35SEric Dumazet smp_mb__after_atomic(); 222714afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) > limit) 2228f9616c35SEric Dumazet return true; 2229f9616c35SEric Dumazet } 2230f9616c35SEric Dumazet return false; 2231f9616c35SEric Dumazet } 2232f9616c35SEric Dumazet 223305b055e8SFrancis Yan static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new) 223405b055e8SFrancis Yan { 2235628174ccSEric Dumazet const u32 now = tcp_jiffies32; 2236efe967cdSArnd Bergmann enum tcp_chrono old = tp->chrono_type; 223705b055e8SFrancis Yan 2238efe967cdSArnd Bergmann if (old > TCP_CHRONO_UNSPEC) 2239efe967cdSArnd Bergmann tp->chrono_stat[old - 1] += now - tp->chrono_start; 224005b055e8SFrancis Yan tp->chrono_start = now; 224105b055e8SFrancis Yan tp->chrono_type = new; 224205b055e8SFrancis Yan } 224305b055e8SFrancis Yan 224405b055e8SFrancis Yan void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type) 224505b055e8SFrancis Yan { 224605b055e8SFrancis Yan struct tcp_sock *tp = tcp_sk(sk); 224705b055e8SFrancis Yan 224805b055e8SFrancis Yan /* If there are multiple conditions worthy of tracking in a 22490f87230dSFrancis Yan * chronograph then the highest priority enum takes precedence 22500f87230dSFrancis Yan * over the other conditions. So that if something "more interesting" 225105b055e8SFrancis Yan * starts happening, stop the previous chrono and start a new one. 225205b055e8SFrancis Yan */ 225305b055e8SFrancis Yan if (type > tp->chrono_type) 225405b055e8SFrancis Yan tcp_chrono_set(tp, type); 225505b055e8SFrancis Yan } 225605b055e8SFrancis Yan 225705b055e8SFrancis Yan void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type) 225805b055e8SFrancis Yan { 225905b055e8SFrancis Yan struct tcp_sock *tp = tcp_sk(sk); 226005b055e8SFrancis Yan 22610f87230dSFrancis Yan 22620f87230dSFrancis Yan /* There are multiple conditions worthy of tracking in a 22630f87230dSFrancis Yan * chronograph, so that the highest priority enum takes 22640f87230dSFrancis Yan * precedence over the other conditions (see tcp_chrono_start). 22650f87230dSFrancis Yan * If a condition stops, we only stop chrono tracking if 22660f87230dSFrancis Yan * it's the "most interesting" or current chrono we are 22670f87230dSFrancis Yan * tracking and starts busy chrono if we have pending data. 22680f87230dSFrancis Yan */ 226975c119afSEric Dumazet if (tcp_rtx_and_write_queues_empty(sk)) 227005b055e8SFrancis Yan tcp_chrono_set(tp, TCP_CHRONO_UNSPEC); 22710f87230dSFrancis Yan else if (type == tp->chrono_type) 22720f87230dSFrancis Yan tcp_chrono_set(tp, TCP_CHRONO_BUSY); 227305b055e8SFrancis Yan } 227405b055e8SFrancis Yan 22751da177e4SLinus Torvalds /* This routine writes packets to the network. It advances the 22761da177e4SLinus Torvalds * send_head. This happens as incoming acks open up the remote 22771da177e4SLinus Torvalds * window for us. 22781da177e4SLinus Torvalds * 2279f8269a49SIlpo Järvinen * LARGESEND note: !tcp_urg_mode is overkill, only frames between 2280f8269a49SIlpo Järvinen * snd_up-64k-mss .. snd_up cannot be large. However, taking into 2281f8269a49SIlpo Järvinen * account rare use of URG, this is not a big flaw. 2282f8269a49SIlpo Järvinen * 22836ba8a3b1SNandita Dukkipati * Send at most one packet when push_one > 0. Temporarily ignore 22846ba8a3b1SNandita Dukkipati * cwnd limit to force at most one packet out when push_one == 2. 22856ba8a3b1SNandita Dukkipati 2286a2a385d6SEric Dumazet * Returns true, if no segments are in flight and we have queued segments, 2287a2a385d6SEric Dumazet * but cannot send anything now because of SWS or another problem. 22881da177e4SLinus Torvalds */ 2289a2a385d6SEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 2290d5dd9175SIlpo Järvinen int push_one, gfp_t gfp) 22911da177e4SLinus Torvalds { 22921da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 229392df7b51SDavid S. Miller struct sk_buff *skb; 2294c1b4a7e6SDavid S. Miller unsigned int tso_segs, sent_pkts; 2295c1b4a7e6SDavid S. Miller int cwnd_quota; 22965d424d5aSJohn Heffner int result; 22975615f886SFrancis Yan bool is_cwnd_limited = false, is_rwnd_limited = false; 2298605ad7f1SEric Dumazet u32 max_segs; 22991da177e4SLinus Torvalds 2300c1b4a7e6SDavid S. Miller sent_pkts = 0; 23015d424d5aSJohn Heffner 2302ee1836aeSEric Dumazet tcp_mstamp_refresh(tp); 2303d5dd9175SIlpo Järvinen if (!push_one) { 23045d424d5aSJohn Heffner /* Do MTU probing. */ 2305d5dd9175SIlpo Järvinen result = tcp_mtu_probe(sk); 2306d5dd9175SIlpo Järvinen if (!result) { 2307a2a385d6SEric Dumazet return false; 23085d424d5aSJohn Heffner } else if (result > 0) { 23095d424d5aSJohn Heffner sent_pkts = 1; 23105d424d5aSJohn Heffner } 2311d5dd9175SIlpo Järvinen } 23125d424d5aSJohn Heffner 2313ed6e7268SNeal Cardwell max_segs = tcp_tso_segs(sk, mss_now); 2314fe067e8aSDavid S. Miller while ((skb = tcp_send_head(sk))) { 2315c8ac3774SHerbert Xu unsigned int limit; 2316c8ac3774SHerbert Xu 2317218af599SEric Dumazet if (tcp_pacing_check(sk)) 2318218af599SEric Dumazet break; 2319218af599SEric Dumazet 23205bbb432cSEric Dumazet tso_segs = tcp_init_tso_segs(skb, mss_now); 2321c1b4a7e6SDavid S. Miller BUG_ON(!tso_segs); 2322c1b4a7e6SDavid S. Miller 23239d186cacSAndrey Vagin if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) { 23247faee5c0SEric Dumazet /* "skb_mstamp" is used as a start point for the retransmit timer */ 2325e2080072SEric Dumazet tcp_update_skb_after_send(tp, skb); 2326ec342325SAndrew Vagin goto repair; /* Skip network transmission */ 23279d186cacSAndrey Vagin } 2328ec342325SAndrew Vagin 2329b68e9f85SHerbert Xu cwnd_quota = tcp_cwnd_test(tp, skb); 23306ba8a3b1SNandita Dukkipati if (!cwnd_quota) { 23316ba8a3b1SNandita Dukkipati if (push_one == 2) 23326ba8a3b1SNandita Dukkipati /* Force out a loss probe pkt. */ 23336ba8a3b1SNandita Dukkipati cwnd_quota = 1; 23346ba8a3b1SNandita Dukkipati else 2335b68e9f85SHerbert Xu break; 23366ba8a3b1SNandita Dukkipati } 2337b68e9f85SHerbert Xu 23385615f886SFrancis Yan if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) { 23395615f886SFrancis Yan is_rwnd_limited = true; 2340b68e9f85SHerbert Xu break; 23415615f886SFrancis Yan } 2342b68e9f85SHerbert Xu 2343d6a4e26aSEric Dumazet if (tso_segs == 1) { 2344aa93466bSDavid S. Miller if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 2345aa93466bSDavid S. Miller (tcp_skb_is_last(sk, skb) ? 2346aa93466bSDavid S. Miller nonagle : TCP_NAGLE_PUSH)))) 2347aa93466bSDavid S. Miller break; 2348c1b4a7e6SDavid S. Miller } else { 2349ca8a2263SNeal Cardwell if (!push_one && 2350605ad7f1SEric Dumazet tcp_tso_should_defer(sk, skb, &is_cwnd_limited, 2351605ad7f1SEric Dumazet max_segs)) 2352aa93466bSDavid S. Miller break; 2353c1b4a7e6SDavid S. Miller } 2354aa93466bSDavid S. Miller 2355605ad7f1SEric Dumazet limit = mss_now; 2356d6a4e26aSEric Dumazet if (tso_segs > 1 && !tcp_urg_mode(tp)) 2357605ad7f1SEric Dumazet limit = tcp_mss_split_point(sk, skb, mss_now, 2358605ad7f1SEric Dumazet min_t(unsigned int, 2359605ad7f1SEric Dumazet cwnd_quota, 2360605ad7f1SEric Dumazet max_segs), 2361605ad7f1SEric Dumazet nonagle); 2362605ad7f1SEric Dumazet 2363605ad7f1SEric Dumazet if (skb->len > limit && 236475c119afSEric Dumazet unlikely(tso_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE, 236575c119afSEric Dumazet skb, limit, mss_now, gfp))) 2366605ad7f1SEric Dumazet break; 2367605ad7f1SEric Dumazet 23687aa5470cSEric Dumazet if (test_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags)) 23697aa5470cSEric Dumazet clear_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags); 2370f9616c35SEric Dumazet if (tcp_small_queue_check(sk, skb, 0)) 237146d3ceabSEric Dumazet break; 2372c9eeec26SEric Dumazet 2373d5dd9175SIlpo Järvinen if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) 23741da177e4SLinus Torvalds break; 23751da177e4SLinus Torvalds 2376ec342325SAndrew Vagin repair: 23771da177e4SLinus Torvalds /* Advance the send_head. This one is sent out. 23781da177e4SLinus Torvalds * This call will increment packets_out. 23791da177e4SLinus Torvalds */ 238066f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 23811da177e4SLinus Torvalds 23821da177e4SLinus Torvalds tcp_minshall_update(tp, mss_now, skb); 2383a262f0cdSNandita Dukkipati sent_pkts += tcp_skb_pcount(skb); 2384d5dd9175SIlpo Järvinen 2385d5dd9175SIlpo Järvinen if (push_one) 2386d5dd9175SIlpo Järvinen break; 23871da177e4SLinus Torvalds } 23881da177e4SLinus Torvalds 23895615f886SFrancis Yan if (is_rwnd_limited) 23905615f886SFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED); 23915615f886SFrancis Yan else 23925615f886SFrancis Yan tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED); 23935615f886SFrancis Yan 2394aa93466bSDavid S. Miller if (likely(sent_pkts)) { 2395684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 2396684bad11SYuchung Cheng tp->prr_out += sent_pkts; 23976ba8a3b1SNandita Dukkipati 23986ba8a3b1SNandita Dukkipati /* Send one loss probe per tail loss episode. */ 23996ba8a3b1SNandita Dukkipati if (push_one != 2) 2400ed66dfafSNeal Cardwell tcp_schedule_loss_probe(sk, false); 2401d2e1339fSBendik Rønning Opstad is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd); 2402ca8a2263SNeal Cardwell tcp_cwnd_validate(sk, is_cwnd_limited); 2403a2a385d6SEric Dumazet return false; 24041da177e4SLinus Torvalds } 240575c119afSEric Dumazet return !tp->packets_out && !tcp_write_queue_empty(sk); 24066ba8a3b1SNandita Dukkipati } 24076ba8a3b1SNandita Dukkipati 2408ed66dfafSNeal Cardwell bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto) 24096ba8a3b1SNandita Dukkipati { 24106ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 24116ba8a3b1SNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 2412a2815817SNeal Cardwell u32 timeout, rto_delta_us; 24132ae21cf5SEric Dumazet int early_retrans; 24146ba8a3b1SNandita Dukkipati 24156ba8a3b1SNandita Dukkipati /* Don't do any loss probe on a Fast Open connection before 3WHS 24166ba8a3b1SNandita Dukkipati * finishes. 24176ba8a3b1SNandita Dukkipati */ 2418f9b99582SYuchung Cheng if (tp->fastopen_rsk) 24196ba8a3b1SNandita Dukkipati return false; 24206ba8a3b1SNandita Dukkipati 24212ae21cf5SEric Dumazet early_retrans = sock_net(sk)->ipv4.sysctl_tcp_early_retrans; 24226ba8a3b1SNandita Dukkipati /* Schedule a loss probe in 2*RTT for SACK capable connections 2423b4f70c3dSNeal Cardwell * not in loss recovery, that are either limited by cwnd or application. 24246ba8a3b1SNandita Dukkipati */ 24252ae21cf5SEric Dumazet if ((early_retrans != 3 && early_retrans != 4) || 2426bec41a11SYuchung Cheng !tp->packets_out || !tcp_is_sack(tp) || 2427b4f70c3dSNeal Cardwell (icsk->icsk_ca_state != TCP_CA_Open && 2428b4f70c3dSNeal Cardwell icsk->icsk_ca_state != TCP_CA_CWR)) 24296ba8a3b1SNandita Dukkipati return false; 24306ba8a3b1SNandita Dukkipati 2431bb4d991aSYuchung Cheng /* Probe timeout is 2*rtt. Add minimum RTO to account 2432f9b99582SYuchung Cheng * for delayed ack when there's one outstanding packet. If no RTT 2433f9b99582SYuchung Cheng * sample is available then probe after TCP_TIMEOUT_INIT. 24346ba8a3b1SNandita Dukkipati */ 2435bb4d991aSYuchung Cheng if (tp->srtt_us) { 2436bb4d991aSYuchung Cheng timeout = usecs_to_jiffies(tp->srtt_us >> 2); 24376ba8a3b1SNandita Dukkipati if (tp->packets_out == 1) 2438bb4d991aSYuchung Cheng timeout += TCP_RTO_MIN; 2439bb4d991aSYuchung Cheng else 2440bb4d991aSYuchung Cheng timeout += TCP_TIMEOUT_MIN; 2441bb4d991aSYuchung Cheng } else { 2442bb4d991aSYuchung Cheng timeout = TCP_TIMEOUT_INIT; 2443bb4d991aSYuchung Cheng } 24446ba8a3b1SNandita Dukkipati 2445a2815817SNeal Cardwell /* If the RTO formula yields an earlier time, then use that time. */ 2446ed66dfafSNeal Cardwell rto_delta_us = advancing_rto ? 2447ed66dfafSNeal Cardwell jiffies_to_usecs(inet_csk(sk)->icsk_rto) : 2448ed66dfafSNeal Cardwell tcp_rto_delta_us(sk); /* How far in future is RTO? */ 2449a2815817SNeal Cardwell if (rto_delta_us > 0) 2450a2815817SNeal Cardwell timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us)); 24516ba8a3b1SNandita Dukkipati 24526ba8a3b1SNandita Dukkipati inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, 24536ba8a3b1SNandita Dukkipati TCP_RTO_MAX); 24546ba8a3b1SNandita Dukkipati return true; 24556ba8a3b1SNandita Dukkipati } 24566ba8a3b1SNandita Dukkipati 24571f3279aeSEric Dumazet /* Thanks to skb fast clones, we can detect if a prior transmit of 24581f3279aeSEric Dumazet * a packet is still in a qdisc or driver queue. 24591f3279aeSEric Dumazet * In this case, there is very little point doing a retransmit ! 24601f3279aeSEric Dumazet */ 24611f3279aeSEric Dumazet static bool skb_still_in_host_queue(const struct sock *sk, 24621f3279aeSEric Dumazet const struct sk_buff *skb) 24631f3279aeSEric Dumazet { 246439bb5e62SEric Dumazet if (unlikely(skb_fclone_busy(sk, skb))) { 2465c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), 24661f3279aeSEric Dumazet LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES); 24671f3279aeSEric Dumazet return true; 24681f3279aeSEric Dumazet } 24691f3279aeSEric Dumazet return false; 24701f3279aeSEric Dumazet } 24711f3279aeSEric Dumazet 2472b340b264SYuchung Cheng /* When probe timeout (PTO) fires, try send a new segment if possible, else 24736ba8a3b1SNandita Dukkipati * retransmit the last segment. 24746ba8a3b1SNandita Dukkipati */ 24756ba8a3b1SNandita Dukkipati void tcp_send_loss_probe(struct sock *sk) 24766ba8a3b1SNandita Dukkipati { 24779b717a8dSNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 24786ba8a3b1SNandita Dukkipati struct sk_buff *skb; 24796ba8a3b1SNandita Dukkipati int pcount; 24806ba8a3b1SNandita Dukkipati int mss = tcp_current_mss(sk); 24816ba8a3b1SNandita Dukkipati 2482b340b264SYuchung Cheng skb = tcp_send_head(sk); 248375c119afSEric Dumazet if (skb && tcp_snd_wnd_test(tp, skb, mss)) { 2484b340b264SYuchung Cheng pcount = tp->packets_out; 2485b340b264SYuchung Cheng tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC); 2486b340b264SYuchung Cheng if (tp->packets_out > pcount) 2487b340b264SYuchung Cheng goto probe_sent; 24886ba8a3b1SNandita Dukkipati goto rearm_timer; 24896ba8a3b1SNandita Dukkipati } 249075c119afSEric Dumazet skb = skb_rb_last(&sk->tcp_rtx_queue); 24916ba8a3b1SNandita Dukkipati 24929b717a8dSNandita Dukkipati /* At most one outstanding TLP retransmission. */ 24939b717a8dSNandita Dukkipati if (tp->tlp_high_seq) 24949b717a8dSNandita Dukkipati goto rearm_timer; 24959b717a8dSNandita Dukkipati 24966ba8a3b1SNandita Dukkipati /* Retransmit last segment. */ 24976ba8a3b1SNandita Dukkipati if (WARN_ON(!skb)) 24986ba8a3b1SNandita Dukkipati goto rearm_timer; 24996ba8a3b1SNandita Dukkipati 25001f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 25011f3279aeSEric Dumazet goto rearm_timer; 25021f3279aeSEric Dumazet 25036ba8a3b1SNandita Dukkipati pcount = tcp_skb_pcount(skb); 25046ba8a3b1SNandita Dukkipati if (WARN_ON(!pcount)) 25056ba8a3b1SNandita Dukkipati goto rearm_timer; 25066ba8a3b1SNandita Dukkipati 25076ba8a3b1SNandita Dukkipati if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) { 250875c119afSEric Dumazet if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, 250975c119afSEric Dumazet (pcount - 1) * mss, mss, 25106cc55e09SOctavian Purdila GFP_ATOMIC))) 25116ba8a3b1SNandita Dukkipati goto rearm_timer; 251275c119afSEric Dumazet skb = skb_rb_next(skb); 25136ba8a3b1SNandita Dukkipati } 25146ba8a3b1SNandita Dukkipati 25156ba8a3b1SNandita Dukkipati if (WARN_ON(!skb || !tcp_skb_pcount(skb))) 25166ba8a3b1SNandita Dukkipati goto rearm_timer; 25176ba8a3b1SNandita Dukkipati 251810d3be56SEric Dumazet if (__tcp_retransmit_skb(sk, skb, 1)) 2519b340b264SYuchung Cheng goto rearm_timer; 25206ba8a3b1SNandita Dukkipati 25219b717a8dSNandita Dukkipati /* Record snd_nxt for loss detection. */ 25229b717a8dSNandita Dukkipati tp->tlp_high_seq = tp->snd_nxt; 25239b717a8dSNandita Dukkipati 2524b340b264SYuchung Cheng probe_sent: 2525c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES); 2526fcd16c0aSYuchung Cheng /* Reset s.t. tcp_rearm_rto will restart timer from now */ 2527fcd16c0aSYuchung Cheng inet_csk(sk)->icsk_pending = 0; 2528b340b264SYuchung Cheng rearm_timer: 2529fcd16c0aSYuchung Cheng tcp_rearm_rto(sk); 25301da177e4SLinus Torvalds } 25311da177e4SLinus Torvalds 2532a762a980SDavid S. Miller /* Push out any pending frames which were held back due to 2533a762a980SDavid S. Miller * TCP_CORK or attempt at coalescing tiny packets. 2534a762a980SDavid S. Miller * The socket must be locked by the caller. 2535a762a980SDavid S. Miller */ 25369e412ba7SIlpo Järvinen void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 25379e412ba7SIlpo Järvinen int nonagle) 2538a762a980SDavid S. Miller { 2539726e07a8SIlpo Järvinen /* If we are closed, the bytes will have to remain here. 2540726e07a8SIlpo Järvinen * In time closedown will finish, we empty the write queue and 2541726e07a8SIlpo Järvinen * all will be happy. 2542726e07a8SIlpo Järvinen */ 2543726e07a8SIlpo Järvinen if (unlikely(sk->sk_state == TCP_CLOSE)) 2544726e07a8SIlpo Järvinen return; 2545726e07a8SIlpo Järvinen 254699a1dec7SMel Gorman if (tcp_write_xmit(sk, cur_mss, nonagle, 0, 25477450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC))) 25489e412ba7SIlpo Järvinen tcp_check_probe_timer(sk); 2549a762a980SDavid S. Miller } 2550a762a980SDavid S. Miller 2551c1b4a7e6SDavid S. Miller /* Send _single_ skb sitting at the send head. This function requires 2552c1b4a7e6SDavid S. Miller * true push pending frames to setup probe timer etc. 2553c1b4a7e6SDavid S. Miller */ 2554c1b4a7e6SDavid S. Miller void tcp_push_one(struct sock *sk, unsigned int mss_now) 2555c1b4a7e6SDavid S. Miller { 2556fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 2557c1b4a7e6SDavid S. Miller 2558c1b4a7e6SDavid S. Miller BUG_ON(!skb || skb->len < mss_now); 2559c1b4a7e6SDavid S. Miller 2560d5dd9175SIlpo Järvinen tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); 2561c1b4a7e6SDavid S. Miller } 2562c1b4a7e6SDavid S. Miller 25631da177e4SLinus Torvalds /* This function returns the amount that we can raise the 25641da177e4SLinus Torvalds * usable window based on the following constraints 25651da177e4SLinus Torvalds * 25661da177e4SLinus Torvalds * 1. The window can never be shrunk once it is offered (RFC 793) 25671da177e4SLinus Torvalds * 2. We limit memory per socket 25681da177e4SLinus Torvalds * 25691da177e4SLinus Torvalds * RFC 1122: 25701da177e4SLinus Torvalds * "the suggested [SWS] avoidance algorithm for the receiver is to keep 25711da177e4SLinus Torvalds * RECV.NEXT + RCV.WIN fixed until: 25721da177e4SLinus Torvalds * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 25731da177e4SLinus Torvalds * 25741da177e4SLinus Torvalds * i.e. don't raise the right edge of the window until you can raise 25751da177e4SLinus Torvalds * it at least MSS bytes. 25761da177e4SLinus Torvalds * 25771da177e4SLinus Torvalds * Unfortunately, the recommended algorithm breaks header prediction, 25781da177e4SLinus Torvalds * since header prediction assumes th->window stays fixed. 25791da177e4SLinus Torvalds * 25801da177e4SLinus Torvalds * Strictly speaking, keeping th->window fixed violates the receiver 25811da177e4SLinus Torvalds * side SWS prevention criteria. The problem is that under this rule 25821da177e4SLinus Torvalds * a stream of single byte packets will cause the right side of the 25831da177e4SLinus Torvalds * window to always advance by a single byte. 25841da177e4SLinus Torvalds * 25851da177e4SLinus Torvalds * Of course, if the sender implements sender side SWS prevention 25861da177e4SLinus Torvalds * then this will not be a problem. 25871da177e4SLinus Torvalds * 25881da177e4SLinus Torvalds * BSD seems to make the following compromise: 25891da177e4SLinus Torvalds * 25901da177e4SLinus Torvalds * If the free space is less than the 1/4 of the maximum 25911da177e4SLinus Torvalds * space available and the free space is less than 1/2 mss, 25921da177e4SLinus Torvalds * then set the window to 0. 25931da177e4SLinus Torvalds * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 25941da177e4SLinus Torvalds * Otherwise, just prevent the window from shrinking 25951da177e4SLinus Torvalds * and from being larger than the largest representable value. 25961da177e4SLinus Torvalds * 25971da177e4SLinus Torvalds * This prevents incremental opening of the window in the regime 25981da177e4SLinus Torvalds * where TCP is limited by the speed of the reader side taking 25991da177e4SLinus Torvalds * data out of the TCP receive queue. It does nothing about 26001da177e4SLinus Torvalds * those cases where the window is constrained on the sender side 26011da177e4SLinus Torvalds * because the pipeline is full. 26021da177e4SLinus Torvalds * 26031da177e4SLinus Torvalds * BSD also seems to "accidentally" limit itself to windows that are a 26041da177e4SLinus Torvalds * multiple of MSS, at least until the free space gets quite small. 26051da177e4SLinus Torvalds * This would appear to be a side effect of the mbuf implementation. 26061da177e4SLinus Torvalds * Combining these two algorithms results in the observed behavior 26071da177e4SLinus Torvalds * of having a fixed window size at almost all times. 26081da177e4SLinus Torvalds * 26091da177e4SLinus Torvalds * Below we obtain similar behavior by forcing the offered window to 26101da177e4SLinus Torvalds * a multiple of the mss when it is feasible to do so. 26111da177e4SLinus Torvalds * 26121da177e4SLinus Torvalds * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 26131da177e4SLinus Torvalds * Regular options like TIMESTAMP are taken into account. 26141da177e4SLinus Torvalds */ 26151da177e4SLinus Torvalds u32 __tcp_select_window(struct sock *sk) 26161da177e4SLinus Torvalds { 2617463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 26181da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2619caa20d9aSStephen Hemminger /* MSS for the peer's data. Previous versions used mss_clamp 26201da177e4SLinus Torvalds * here. I don't know if the value based on our guesses 26211da177e4SLinus Torvalds * of peer's MSS is better for the performance. It's more correct 26221da177e4SLinus Torvalds * but may be worse for the performance because of rcv_mss 26231da177e4SLinus Torvalds * fluctuations. --SAW 1998/11/1 26241da177e4SLinus Torvalds */ 2625463c84b9SArnaldo Carvalho de Melo int mss = icsk->icsk_ack.rcv_mss; 26261da177e4SLinus Torvalds int free_space = tcp_space(sk); 262786c1a045SFlorian Westphal int allowed_space = tcp_full_space(sk); 262886c1a045SFlorian Westphal int full_space = min_t(int, tp->window_clamp, allowed_space); 26291da177e4SLinus Torvalds int window; 26301da177e4SLinus Torvalds 263106425c30SEric Dumazet if (unlikely(mss > full_space)) { 26321da177e4SLinus Torvalds mss = full_space; 263306425c30SEric Dumazet if (mss <= 0) 263406425c30SEric Dumazet return 0; 263506425c30SEric Dumazet } 2636b92edbe0SEric Dumazet if (free_space < (full_space >> 1)) { 2637463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.quick = 0; 26381da177e4SLinus Torvalds 2639b8da51ebSEric Dumazet if (tcp_under_memory_pressure(sk)) 2640056834d9SIlpo Järvinen tp->rcv_ssthresh = min(tp->rcv_ssthresh, 2641056834d9SIlpo Järvinen 4U * tp->advmss); 26421da177e4SLinus Torvalds 264386c1a045SFlorian Westphal /* free_space might become our new window, make sure we don't 264486c1a045SFlorian Westphal * increase it due to wscale. 264586c1a045SFlorian Westphal */ 264686c1a045SFlorian Westphal free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale); 264786c1a045SFlorian Westphal 264886c1a045SFlorian Westphal /* if free space is less than mss estimate, or is below 1/16th 264986c1a045SFlorian Westphal * of the maximum allowed, try to move to zero-window, else 265086c1a045SFlorian Westphal * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and 265186c1a045SFlorian Westphal * new incoming data is dropped due to memory limits. 265286c1a045SFlorian Westphal * With large window, mss test triggers way too late in order 265386c1a045SFlorian Westphal * to announce zero window in time before rmem limit kicks in. 265486c1a045SFlorian Westphal */ 265586c1a045SFlorian Westphal if (free_space < (allowed_space >> 4) || free_space < mss) 26561da177e4SLinus Torvalds return 0; 26571da177e4SLinus Torvalds } 26581da177e4SLinus Torvalds 26591da177e4SLinus Torvalds if (free_space > tp->rcv_ssthresh) 26601da177e4SLinus Torvalds free_space = tp->rcv_ssthresh; 26611da177e4SLinus Torvalds 26621da177e4SLinus Torvalds /* Don't do rounding if we are using window scaling, since the 26631da177e4SLinus Torvalds * scaled window will not line up with the MSS boundary anyway. 26641da177e4SLinus Torvalds */ 26651da177e4SLinus Torvalds if (tp->rx_opt.rcv_wscale) { 26661da177e4SLinus Torvalds window = free_space; 26671da177e4SLinus Torvalds 26681da177e4SLinus Torvalds /* Advertise enough space so that it won't get scaled away. 26691da177e4SLinus Torvalds * Import case: prevent zero window announcement if 26701da177e4SLinus Torvalds * 1<<rcv_wscale > mss. 26711da177e4SLinus Torvalds */ 26721935299dSGao Feng window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale)); 26731da177e4SLinus Torvalds } else { 26741935299dSGao Feng window = tp->rcv_wnd; 26751da177e4SLinus Torvalds /* Get the largest window that is a nice multiple of mss. 26761da177e4SLinus Torvalds * Window clamp already applied above. 26771da177e4SLinus Torvalds * If our current window offering is within 1 mss of the 26781da177e4SLinus Torvalds * free space we just keep it. This prevents the divide 26791da177e4SLinus Torvalds * and multiply from happening most of the time. 26801da177e4SLinus Torvalds * We also don't do any window rounding when the free space 26811da177e4SLinus Torvalds * is too small. 26821da177e4SLinus Torvalds */ 26831da177e4SLinus Torvalds if (window <= free_space - mss || window > free_space) 26841935299dSGao Feng window = rounddown(free_space, mss); 268584565070SJohn Heffner else if (mss == full_space && 2686b92edbe0SEric Dumazet free_space > window + (full_space >> 1)) 268784565070SJohn Heffner window = free_space; 26881da177e4SLinus Torvalds } 26891da177e4SLinus Torvalds 26901da177e4SLinus Torvalds return window; 26911da177e4SLinus Torvalds } 26921da177e4SLinus Torvalds 2693cfea5a68SMartin KaFai Lau void tcp_skb_collapse_tstamp(struct sk_buff *skb, 2694082ac2d5SMartin KaFai Lau const struct sk_buff *next_skb) 2695082ac2d5SMartin KaFai Lau { 26960a2cf20cSSoheil Hassas Yeganeh if (unlikely(tcp_has_tx_tstamp(next_skb))) { 26970a2cf20cSSoheil Hassas Yeganeh const struct skb_shared_info *next_shinfo = 26980a2cf20cSSoheil Hassas Yeganeh skb_shinfo(next_skb); 2699082ac2d5SMartin KaFai Lau struct skb_shared_info *shinfo = skb_shinfo(skb); 2700082ac2d5SMartin KaFai Lau 27010a2cf20cSSoheil Hassas Yeganeh shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP; 2702082ac2d5SMartin KaFai Lau shinfo->tskey = next_shinfo->tskey; 27032de8023eSMartin KaFai Lau TCP_SKB_CB(skb)->txstamp_ack |= 27042de8023eSMartin KaFai Lau TCP_SKB_CB(next_skb)->txstamp_ack; 2705082ac2d5SMartin KaFai Lau } 2706082ac2d5SMartin KaFai Lau } 2707082ac2d5SMartin KaFai Lau 27084a17fc3aSIlpo Järvinen /* Collapses two adjacent SKB's during retransmission. */ 2709f8071cdeSEric Dumazet static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) 27101da177e4SLinus Torvalds { 27111da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 271275c119afSEric Dumazet struct sk_buff *next_skb = skb_rb_next(skb); 2713058dc334SIlpo Järvinen int skb_size, next_skb_size; 27141da177e4SLinus Torvalds 2715058dc334SIlpo Järvinen skb_size = skb->len; 2716058dc334SIlpo Järvinen next_skb_size = next_skb->len; 27171da177e4SLinus Torvalds 2718058dc334SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); 27191da177e4SLinus Torvalds 2720f8071cdeSEric Dumazet if (next_skb_size) { 2721f8071cdeSEric Dumazet if (next_skb_size <= skb_availroom(skb)) 2722f8071cdeSEric Dumazet skb_copy_bits(next_skb, 0, skb_put(skb, next_skb_size), 2723f8071cdeSEric Dumazet next_skb_size); 2724f8071cdeSEric Dumazet else if (!skb_shift(skb, next_skb, next_skb_size)) 2725f8071cdeSEric Dumazet return false; 2726f8071cdeSEric Dumazet } 27272b7cda9cSEric Dumazet tcp_highest_sack_replace(sk, next_skb, skb); 2728a6963a6bSIlpo Järvinen 27291da177e4SLinus Torvalds /* Update sequence range on original skb. */ 27301da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 27311da177e4SLinus Torvalds 2732e6c7d085SIlpo Järvinen /* Merge over control information. This moves PSH/FIN etc. over */ 27334de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; 27341da177e4SLinus Torvalds 27351da177e4SLinus Torvalds /* All done, get rid of second SKB and account for it so 27361da177e4SLinus Torvalds * packet counting does not break. 27371da177e4SLinus Torvalds */ 27384828e7f4SIlpo Järvinen TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; 2739a643b5d4SMartin KaFai Lau TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor; 2740b7689205SIlpo Järvinen 2741b7689205SIlpo Järvinen /* changed transmit queue under us so clear hints */ 2742ef9da47cSIlpo Järvinen tcp_clear_retrans_hints_partial(tp); 2743ef9da47cSIlpo Järvinen if (next_skb == tp->retransmit_skb_hint) 2744ef9da47cSIlpo Järvinen tp->retransmit_skb_hint = skb; 2745b7689205SIlpo Järvinen 2746797108d1SIlpo Järvinen tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); 2747797108d1SIlpo Järvinen 2748082ac2d5SMartin KaFai Lau tcp_skb_collapse_tstamp(skb, next_skb); 2749082ac2d5SMartin KaFai Lau 275075c119afSEric Dumazet tcp_rtx_queue_unlink_and_free(next_skb, sk); 2751f8071cdeSEric Dumazet return true; 27521da177e4SLinus Torvalds } 27531da177e4SLinus Torvalds 275467edfef7SAndi Kleen /* Check if coalescing SKBs is legal. */ 2755a2a385d6SEric Dumazet static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) 27564a17fc3aSIlpo Järvinen { 27574a17fc3aSIlpo Järvinen if (tcp_skb_pcount(skb) > 1) 2758a2a385d6SEric Dumazet return false; 27594a17fc3aSIlpo Järvinen if (skb_cloned(skb)) 2760a2a385d6SEric Dumazet return false; 27612331ccc5SEric Dumazet /* Some heuristics for collapsing over SACK'd could be invented */ 27624a17fc3aSIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 2763a2a385d6SEric Dumazet return false; 27644a17fc3aSIlpo Järvinen 2765a2a385d6SEric Dumazet return true; 27664a17fc3aSIlpo Järvinen } 27674a17fc3aSIlpo Järvinen 276867edfef7SAndi Kleen /* Collapse packets in the retransmit queue to make to create 276967edfef7SAndi Kleen * less packets on the wire. This is only done on retransmission. 277067edfef7SAndi Kleen */ 27714a17fc3aSIlpo Järvinen static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, 27724a17fc3aSIlpo Järvinen int space) 27734a17fc3aSIlpo Järvinen { 27744a17fc3aSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 27754a17fc3aSIlpo Järvinen struct sk_buff *skb = to, *tmp; 2776a2a385d6SEric Dumazet bool first = true; 27774a17fc3aSIlpo Järvinen 2778e0a1e5b5SEric Dumazet if (!sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse) 27794a17fc3aSIlpo Järvinen return; 27804de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 27814a17fc3aSIlpo Järvinen return; 27824a17fc3aSIlpo Järvinen 278375c119afSEric Dumazet skb_rbtree_walk_from_safe(skb, tmp) { 27844a17fc3aSIlpo Järvinen if (!tcp_can_collapse(sk, skb)) 27854a17fc3aSIlpo Järvinen break; 27864a17fc3aSIlpo Järvinen 2787a643b5d4SMartin KaFai Lau if (!tcp_skb_can_collapse_to(to)) 2788a643b5d4SMartin KaFai Lau break; 2789a643b5d4SMartin KaFai Lau 27904a17fc3aSIlpo Järvinen space -= skb->len; 27914a17fc3aSIlpo Järvinen 27924a17fc3aSIlpo Järvinen if (first) { 2793a2a385d6SEric Dumazet first = false; 27944a17fc3aSIlpo Järvinen continue; 27954a17fc3aSIlpo Järvinen } 27964a17fc3aSIlpo Järvinen 27974a17fc3aSIlpo Järvinen if (space < 0) 27984a17fc3aSIlpo Järvinen break; 27994a17fc3aSIlpo Järvinen 28004a17fc3aSIlpo Järvinen if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) 28014a17fc3aSIlpo Järvinen break; 28024a17fc3aSIlpo Järvinen 2803f8071cdeSEric Dumazet if (!tcp_collapse_retrans(sk, to)) 2804f8071cdeSEric Dumazet break; 28054a17fc3aSIlpo Järvinen } 28064a17fc3aSIlpo Järvinen } 28074a17fc3aSIlpo Järvinen 28081da177e4SLinus Torvalds /* This retransmits one SKB. Policy decisions and retransmit queue 28091da177e4SLinus Torvalds * state updates are done by the caller. Returns non-zero if an 28101da177e4SLinus Torvalds * error occurred which prevented the send. 28111da177e4SLinus Torvalds */ 281210d3be56SEric Dumazet int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) 28131da177e4SLinus Torvalds { 28145d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 281510d3be56SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 28167d227cd2SSridhar Samudrala unsigned int cur_mss; 281710d3be56SEric Dumazet int diff, len, err; 28181da177e4SLinus Torvalds 281910d3be56SEric Dumazet 282010d3be56SEric Dumazet /* Inconclusive MTU probe */ 282110d3be56SEric Dumazet if (icsk->icsk_mtup.probe_size) 28225d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 28235d424d5aSJohn Heffner 28241da177e4SLinus Torvalds /* Do not sent more than we queued. 1/4 is reserved for possible 2825caa20d9aSStephen Hemminger * copying overhead: fragmentation, tunneling, mangling etc. 28261da177e4SLinus Torvalds */ 282714afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) > 2828ffb4d6c8SEric Dumazet min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), 2829ffb4d6c8SEric Dumazet sk->sk_sndbuf)) 28301da177e4SLinus Torvalds return -EAGAIN; 28311da177e4SLinus Torvalds 28321f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 28331f3279aeSEric Dumazet return -EBUSY; 28341f3279aeSEric Dumazet 28351da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 28361da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 28371da177e4SLinus Torvalds BUG(); 28381da177e4SLinus Torvalds if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 28391da177e4SLinus Torvalds return -ENOMEM; 28401da177e4SLinus Torvalds } 28411da177e4SLinus Torvalds 28427d227cd2SSridhar Samudrala if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 28437d227cd2SSridhar Samudrala return -EHOSTUNREACH; /* Routing failure or similar. */ 28447d227cd2SSridhar Samudrala 28450c54b85fSIlpo Järvinen cur_mss = tcp_current_mss(sk); 28467d227cd2SSridhar Samudrala 28471da177e4SLinus Torvalds /* If receiver has shrunk his window, and skb is out of 28481da177e4SLinus Torvalds * new window, do not retransmit it. The exception is the 28491da177e4SLinus Torvalds * case, when window is shrunk to zero. In this case 28501da177e4SLinus Torvalds * our retransmit serves as a zero window probe. 28511da177e4SLinus Torvalds */ 28529d4fb27dSJoe Perches if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && 28539d4fb27dSJoe Perches TCP_SKB_CB(skb)->seq != tp->snd_una) 28541da177e4SLinus Torvalds return -EAGAIN; 28551da177e4SLinus Torvalds 285610d3be56SEric Dumazet len = cur_mss * segs; 285710d3be56SEric Dumazet if (skb->len > len) { 285875c119afSEric Dumazet if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len, 285975c119afSEric Dumazet cur_mss, GFP_ATOMIC)) 28601da177e4SLinus Torvalds return -ENOMEM; /* We'll try again later. */ 286102276f3cSIlpo Järvinen } else { 2862c52e2421SEric Dumazet if (skb_unclone(skb, GFP_ATOMIC)) 2863c52e2421SEric Dumazet return -ENOMEM; 286410d3be56SEric Dumazet 286510d3be56SEric Dumazet diff = tcp_skb_pcount(skb); 286610d3be56SEric Dumazet tcp_set_skb_tso_segs(skb, cur_mss); 286710d3be56SEric Dumazet diff -= tcp_skb_pcount(skb); 286810d3be56SEric Dumazet if (diff) 286910d3be56SEric Dumazet tcp_adjust_pcount(sk, skb, diff); 287010d3be56SEric Dumazet if (skb->len < cur_mss) 287110d3be56SEric Dumazet tcp_retrans_try_collapse(sk, skb, cur_mss); 28721da177e4SLinus Torvalds } 28731da177e4SLinus Torvalds 287449213555SDaniel Borkmann /* RFC3168, section 6.1.1.1. ECN fallback */ 287549213555SDaniel Borkmann if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN) 287649213555SDaniel Borkmann tcp_ecn_clear_syn(sk, skb); 287749213555SDaniel Borkmann 2878678550c6SYuchung Cheng /* Update global and local TCP statistics. */ 2879678550c6SYuchung Cheng segs = tcp_skb_pcount(skb); 2880678550c6SYuchung Cheng TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs); 2881678550c6SYuchung Cheng if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 2882678550c6SYuchung Cheng __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 2883678550c6SYuchung Cheng tp->total_retrans += segs; 2884678550c6SYuchung Cheng 288550bceae9SThomas Graf /* make sure skb->data is aligned on arches that require it 288650bceae9SThomas Graf * and check if ack-trimming & collapsing extended the headroom 288750bceae9SThomas Graf * beyond what csum_start can cover. 288850bceae9SThomas Graf */ 288950bceae9SThomas Graf if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) || 289050bceae9SThomas Graf skb_headroom(skb) >= 0xFFFF)) { 289110a81980SEric Dumazet struct sk_buff *nskb; 289210a81980SEric Dumazet 2893e2080072SEric Dumazet tcp_skb_tsorted_save(skb) { 289410a81980SEric Dumazet nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC); 2895c84a5711SYuchung Cheng err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) : 2896117632e6SEric Dumazet -ENOBUFS; 2897e2080072SEric Dumazet } tcp_skb_tsorted_restore(skb); 2898e2080072SEric Dumazet 28995889e2c0SYousuk Seung if (!err) { 2900e2080072SEric Dumazet tcp_update_skb_after_send(tp, skb); 29015889e2c0SYousuk Seung tcp_rate_skb_sent(sk, skb); 29025889e2c0SYousuk Seung } 2903117632e6SEric Dumazet } else { 2904c84a5711SYuchung Cheng err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 2905117632e6SEric Dumazet } 2906c84a5711SYuchung Cheng 2907a31ad29eSLawrence Brakmo if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RETRANS_CB_FLAG)) 2908a31ad29eSLawrence Brakmo tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RETRANS_CB, 2909a31ad29eSLawrence Brakmo TCP_SKB_CB(skb)->seq, segs, err); 2910a31ad29eSLawrence Brakmo 2911fc9f3501SEric Dumazet if (likely(!err)) { 2912c84a5711SYuchung Cheng TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS; 2913e086101bSCong Wang trace_tcp_retransmit_skb(sk, skb); 2914678550c6SYuchung Cheng } else if (err != -EBUSY) { 2915678550c6SYuchung Cheng NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL); 2916fc9f3501SEric Dumazet } 2917c84a5711SYuchung Cheng return err; 291893b174adSYuchung Cheng } 291993b174adSYuchung Cheng 292010d3be56SEric Dumazet int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) 292193b174adSYuchung Cheng { 292293b174adSYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 292310d3be56SEric Dumazet int err = __tcp_retransmit_skb(sk, skb, segs); 29241da177e4SLinus Torvalds 29251da177e4SLinus Torvalds if (err == 0) { 29261da177e4SLinus Torvalds #if FASTRETRANS_DEBUG > 0 29271da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 2928e87cc472SJoe Perches net_dbg_ratelimited("retrans_out leaked\n"); 29291da177e4SLinus Torvalds } 29301da177e4SLinus Torvalds #endif 29311da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 29321da177e4SLinus Torvalds tp->retrans_out += tcp_skb_pcount(skb); 29331da177e4SLinus Torvalds 29341da177e4SLinus Torvalds /* Save stamp of the first retransmit. */ 29351da177e4SLinus Torvalds if (!tp->retrans_stamp) 29367faee5c0SEric Dumazet tp->retrans_stamp = tcp_skb_timestamp(skb); 29371da177e4SLinus Torvalds 29381da177e4SLinus Torvalds } 29396e08d5e3SYuchung Cheng 29406e08d5e3SYuchung Cheng if (tp->undo_retrans < 0) 29416e08d5e3SYuchung Cheng tp->undo_retrans = 0; 29426e08d5e3SYuchung Cheng tp->undo_retrans += tcp_skb_pcount(skb); 29431da177e4SLinus Torvalds return err; 29441da177e4SLinus Torvalds } 29451da177e4SLinus Torvalds 29461da177e4SLinus Torvalds /* This gets called after a retransmit timeout, and the initially 29471da177e4SLinus Torvalds * retransmitted data is acknowledged. It tries to continue 29481da177e4SLinus Torvalds * resending the rest of the retransmit queue, until either 29491da177e4SLinus Torvalds * we've sent it all or the congestion window limit is reached. 29501da177e4SLinus Torvalds */ 29511da177e4SLinus Torvalds void tcp_xmit_retransmit_queue(struct sock *sk) 29521da177e4SLinus Torvalds { 29536687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 2954b9f1f1ceSEric Dumazet struct sk_buff *skb, *rtx_head, *hole = NULL; 29551da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2956840a3cbeSYuchung Cheng u32 max_segs; 295761eb55f4SIlpo Järvinen int mib_idx; 29586a438bbeSStephen Hemminger 295945e77d31SIlpo Järvinen if (!tp->packets_out) 296045e77d31SIlpo Järvinen return; 296145e77d31SIlpo Järvinen 296275c119afSEric Dumazet rtx_head = tcp_rtx_queue_head(sk); 2963b9f1f1ceSEric Dumazet skb = tp->retransmit_skb_hint ?: rtx_head; 2964ed6e7268SNeal Cardwell max_segs = tcp_tso_segs(sk, tcp_current_mss(sk)); 296575c119afSEric Dumazet skb_rbtree_walk_from(skb) { 2966dca0aaf8SEric Dumazet __u8 sacked; 296710d3be56SEric Dumazet int segs; 29681da177e4SLinus Torvalds 2969218af599SEric Dumazet if (tcp_pacing_check(sk)) 2970218af599SEric Dumazet break; 2971218af599SEric Dumazet 29726a438bbeSStephen Hemminger /* we could do better than to assign each time */ 297351456b29SIan Morris if (!hole) 29746a438bbeSStephen Hemminger tp->retransmit_skb_hint = skb; 29756a438bbeSStephen Hemminger 297610d3be56SEric Dumazet segs = tp->snd_cwnd - tcp_packets_in_flight(tp); 297710d3be56SEric Dumazet if (segs <= 0) 29781da177e4SLinus Torvalds return; 2979dca0aaf8SEric Dumazet sacked = TCP_SKB_CB(skb)->sacked; 2980a3d2e9f8SEric Dumazet /* In case tcp_shift_skb_data() have aggregated large skbs, 2981a3d2e9f8SEric Dumazet * we need to make sure not sending too bigs TSO packets 2982a3d2e9f8SEric Dumazet */ 2983a3d2e9f8SEric Dumazet segs = min_t(int, segs, max_segs); 29840e1c54c2SIlpo Järvinen 2985840a3cbeSYuchung Cheng if (tp->retrans_out >= tp->lost_out) { 2986006f582cSIlpo Järvinen break; 29870e1c54c2SIlpo Järvinen } else if (!(sacked & TCPCB_LOST)) { 298851456b29SIan Morris if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) 29890e1c54c2SIlpo Järvinen hole = skb; 299061eb55f4SIlpo Järvinen continue; 29911da177e4SLinus Torvalds 29920e1c54c2SIlpo Järvinen } else { 29930e1c54c2SIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Loss) 29940e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFASTRETRANS; 29950e1c54c2SIlpo Järvinen else 29960e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; 29970e1c54c2SIlpo Järvinen } 29980e1c54c2SIlpo Järvinen 29990e1c54c2SIlpo Järvinen if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) 300061eb55f4SIlpo Järvinen continue; 300140b215e5SPavel Emelyanov 3002f9616c35SEric Dumazet if (tcp_small_queue_check(sk, skb, 1)) 3003f9616c35SEric Dumazet return; 3004f9616c35SEric Dumazet 300510d3be56SEric Dumazet if (tcp_retransmit_skb(sk, skb, segs)) 30061da177e4SLinus Torvalds return; 300724ab6becSYuchung Cheng 3008de1d6578SYuchung Cheng NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb)); 30091da177e4SLinus Torvalds 3010684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 3011a262f0cdSNandita Dukkipati tp->prr_out += tcp_skb_pcount(skb); 3012a262f0cdSNandita Dukkipati 301375c119afSEric Dumazet if (skb == rtx_head && 301457dde7f7SYuchung Cheng icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT) 3015463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 30163f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, 30173f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 30181da177e4SLinus Torvalds } 30191da177e4SLinus Torvalds } 30201da177e4SLinus Torvalds 3021d83769a5SEric Dumazet /* We allow to exceed memory limits for FIN packets to expedite 3022d83769a5SEric Dumazet * connection tear down and (memory) recovery. 3023845704a5SEric Dumazet * Otherwise tcp_send_fin() could be tempted to either delay FIN 3024845704a5SEric Dumazet * or even be forced to close flow without any FIN. 3025a6c5ea4cSEric Dumazet * In general, we want to allow one skb per socket to avoid hangs 3026a6c5ea4cSEric Dumazet * with edge trigger epoll() 3027d83769a5SEric Dumazet */ 3028a6c5ea4cSEric Dumazet void sk_forced_mem_schedule(struct sock *sk, int size) 3029d83769a5SEric Dumazet { 3030e805605cSJohannes Weiner int amt; 3031d83769a5SEric Dumazet 3032d83769a5SEric Dumazet if (size <= sk->sk_forward_alloc) 3033d83769a5SEric Dumazet return; 3034d83769a5SEric Dumazet amt = sk_mem_pages(size); 3035d83769a5SEric Dumazet sk->sk_forward_alloc += amt * SK_MEM_QUANTUM; 3036e805605cSJohannes Weiner sk_memory_allocated_add(sk, amt); 3037e805605cSJohannes Weiner 3038baac50bbSJohannes Weiner if (mem_cgroup_sockets_enabled && sk->sk_memcg) 3039baac50bbSJohannes Weiner mem_cgroup_charge_skmem(sk->sk_memcg, amt); 3040d83769a5SEric Dumazet } 3041d83769a5SEric Dumazet 3042845704a5SEric Dumazet /* Send a FIN. The caller locks the socket for us. 3043845704a5SEric Dumazet * We should try to send a FIN packet really hard, but eventually give up. 30441da177e4SLinus Torvalds */ 30451da177e4SLinus Torvalds void tcp_send_fin(struct sock *sk) 30461da177e4SLinus Torvalds { 3047845704a5SEric Dumazet struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk); 30481da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 30491da177e4SLinus Torvalds 3050845704a5SEric Dumazet /* Optimization, tack on the FIN if we have one skb in write queue and 3051845704a5SEric Dumazet * this skb was not yet sent, or we are under memory pressure. 3052845704a5SEric Dumazet * Note: in the latter case, FIN packet will be sent after a timeout, 3053845704a5SEric Dumazet * as TCP stack thinks it has already been transmitted. 30541da177e4SLinus Torvalds */ 305575c119afSEric Dumazet if (!tskb && tcp_under_memory_pressure(sk)) 305675c119afSEric Dumazet tskb = skb_rb_last(&sk->tcp_rtx_queue); 305775c119afSEric Dumazet 305875c119afSEric Dumazet if (tskb) { 3059845704a5SEric Dumazet coalesce: 3060845704a5SEric Dumazet TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN; 3061845704a5SEric Dumazet TCP_SKB_CB(tskb)->end_seq++; 30621da177e4SLinus Torvalds tp->write_seq++; 306375c119afSEric Dumazet if (tcp_write_queue_empty(sk)) { 3064845704a5SEric Dumazet /* This means tskb was already sent. 3065845704a5SEric Dumazet * Pretend we included the FIN on previous transmit. 3066845704a5SEric Dumazet * We need to set tp->snd_nxt to the value it would have 3067845704a5SEric Dumazet * if FIN had been sent. This is because retransmit path 3068845704a5SEric Dumazet * does not change tp->snd_nxt. 3069845704a5SEric Dumazet */ 3070845704a5SEric Dumazet tp->snd_nxt++; 3071845704a5SEric Dumazet return; 3072845704a5SEric Dumazet } 30731da177e4SLinus Torvalds } else { 3074845704a5SEric Dumazet skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation); 3075845704a5SEric Dumazet if (unlikely(!skb)) { 3076845704a5SEric Dumazet if (tskb) 3077845704a5SEric Dumazet goto coalesce; 3078845704a5SEric Dumazet return; 30791da177e4SLinus Torvalds } 3080e2080072SEric Dumazet INIT_LIST_HEAD(&skb->tcp_tsorted_anchor); 3081d83769a5SEric Dumazet skb_reserve(skb, MAX_TCP_HEADER); 3082a6c5ea4cSEric Dumazet sk_forced_mem_schedule(sk, skb->truesize); 30831da177e4SLinus Torvalds /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 3084e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tp->write_seq, 3085a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_FIN); 30861da177e4SLinus Torvalds tcp_queue_skb(sk, skb); 30871da177e4SLinus Torvalds } 3088845704a5SEric Dumazet __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF); 30891da177e4SLinus Torvalds } 30901da177e4SLinus Torvalds 30911da177e4SLinus Torvalds /* We get here when a process closes a file descriptor (either due to 30921da177e4SLinus Torvalds * an explicit close() or as a byproduct of exit()'ing) and there 30931da177e4SLinus Torvalds * was unread data in the receive queue. This behavior is recommended 309465bb723cSGerrit Renker * by RFC 2525, section 2.17. -DaveM 30951da177e4SLinus Torvalds */ 3096dd0fc66fSAl Viro void tcp_send_active_reset(struct sock *sk, gfp_t priority) 30971da177e4SLinus Torvalds { 30981da177e4SLinus Torvalds struct sk_buff *skb; 30991da177e4SLinus Torvalds 31007cc2b043SGao Feng TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); 31017cc2b043SGao Feng 31021da177e4SLinus Torvalds /* NOTE: No TCP options attached and we never retransmit this. */ 31031da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, priority); 31041da177e4SLinus Torvalds if (!skb) { 31054e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 31061da177e4SLinus Torvalds return; 31071da177e4SLinus Torvalds } 31081da177e4SLinus Torvalds 31091da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 31101da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 3111e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), 3112a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_RST); 31139a568de4SEric Dumazet tcp_mstamp_refresh(tcp_sk(sk)); 31141da177e4SLinus Torvalds /* Send it off. */ 3115dfb4b9dcSDavid S. Miller if (tcp_transmit_skb(sk, skb, 0, priority)) 31164e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 3117c24b14c4SSong Liu 3118c24b14c4SSong Liu /* skb of trace_tcp_send_reset() keeps the skb that caused RST, 3119c24b14c4SSong Liu * skb here is different to the troublesome skb, so use NULL 3120c24b14c4SSong Liu */ 3121c24b14c4SSong Liu trace_tcp_send_reset(sk, NULL); 31221da177e4SLinus Torvalds } 31231da177e4SLinus Torvalds 312467edfef7SAndi Kleen /* Send a crossed SYN-ACK during socket establishment. 312567edfef7SAndi Kleen * WARNING: This routine must only be called when we have already sent 31261da177e4SLinus Torvalds * a SYN packet that crossed the incoming SYN that caused this routine 31271da177e4SLinus Torvalds * to get called. If this assumption fails then the initial rcv_wnd 31281da177e4SLinus Torvalds * and rcv_wscale values will not be correct. 31291da177e4SLinus Torvalds */ 31301da177e4SLinus Torvalds int tcp_send_synack(struct sock *sk) 31311da177e4SLinus Torvalds { 31321da177e4SLinus Torvalds struct sk_buff *skb; 31331da177e4SLinus Torvalds 313475c119afSEric Dumazet skb = tcp_rtx_queue_head(sk); 313551456b29SIan Morris if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 313675c119afSEric Dumazet pr_err("%s: wrong queue state\n", __func__); 31371da177e4SLinus Torvalds return -EFAULT; 31381da177e4SLinus Torvalds } 31394de075e0SEric Dumazet if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { 31401da177e4SLinus Torvalds if (skb_cloned(skb)) { 3141e2080072SEric Dumazet struct sk_buff *nskb; 3142e2080072SEric Dumazet 3143e2080072SEric Dumazet tcp_skb_tsorted_save(skb) { 3144e2080072SEric Dumazet nskb = skb_copy(skb, GFP_ATOMIC); 3145e2080072SEric Dumazet } tcp_skb_tsorted_restore(skb); 314651456b29SIan Morris if (!nskb) 31471da177e4SLinus Torvalds return -ENOMEM; 3148e2080072SEric Dumazet INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor); 314975c119afSEric Dumazet tcp_rtx_queue_unlink_and_free(skb, sk); 3150f4a775d1SEric Dumazet __skb_header_release(nskb); 315175c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb); 31523ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 31533ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 31541da177e4SLinus Torvalds skb = nskb; 31551da177e4SLinus Torvalds } 31561da177e4SLinus Torvalds 31574de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; 3158735d3831SFlorian Westphal tcp_ecn_send_synack(sk, skb); 31591da177e4SLinus Torvalds } 3160dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 31611da177e4SLinus Torvalds } 31621da177e4SLinus Torvalds 31634aea39c1SEric Dumazet /** 31644aea39c1SEric Dumazet * tcp_make_synack - Prepare a SYN-ACK. 31654aea39c1SEric Dumazet * sk: listener socket 31664aea39c1SEric Dumazet * dst: dst entry attached to the SYNACK 31674aea39c1SEric Dumazet * req: request_sock pointer 31684aea39c1SEric Dumazet * 31694aea39c1SEric Dumazet * Allocate one skb and build a SYNACK packet. 31704aea39c1SEric Dumazet * @dst is consumed : Caller should not use it again. 31714aea39c1SEric Dumazet */ 31725d062de7SEric Dumazet struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst, 3173e6b4d113SWilliam Allen Simpson struct request_sock *req, 3174ca6fb065SEric Dumazet struct tcp_fastopen_cookie *foc, 3175b3d05147SEric Dumazet enum tcp_synack_type synack_type) 31761da177e4SLinus Torvalds { 31772e6599cbSArnaldo Carvalho de Melo struct inet_request_sock *ireq = inet_rsk(req); 31785d062de7SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 317980f03e27SEric Dumazet struct tcp_md5sig_key *md5 = NULL; 31805d062de7SEric Dumazet struct tcp_out_options opts; 31815d062de7SEric Dumazet struct sk_buff *skb; 3182bd0388aeSWilliam Allen Simpson int tcp_header_size; 31835d062de7SEric Dumazet struct tcphdr *th; 3184f5fff5dcSTom Quetchenbach int mss; 31851da177e4SLinus Torvalds 3186ca6fb065SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 31874aea39c1SEric Dumazet if (unlikely(!skb)) { 31884aea39c1SEric Dumazet dst_release(dst); 31891da177e4SLinus Torvalds return NULL; 31904aea39c1SEric Dumazet } 31911da177e4SLinus Torvalds /* Reserve space for headers. */ 31921da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 31931da177e4SLinus Torvalds 3194b3d05147SEric Dumazet switch (synack_type) { 3195b3d05147SEric Dumazet case TCP_SYNACK_NORMAL: 31969e17f8a4SEric Dumazet skb_set_owner_w(skb, req_to_sk(req)); 3197b3d05147SEric Dumazet break; 3198b3d05147SEric Dumazet case TCP_SYNACK_COOKIE: 3199b3d05147SEric Dumazet /* Under synflood, we do not attach skb to a socket, 3200b3d05147SEric Dumazet * to avoid false sharing. 3201b3d05147SEric Dumazet */ 3202b3d05147SEric Dumazet break; 3203b3d05147SEric Dumazet case TCP_SYNACK_FASTOPEN: 3204ca6fb065SEric Dumazet /* sk is a const pointer, because we want to express multiple 3205ca6fb065SEric Dumazet * cpu might call us concurrently. 3206ca6fb065SEric Dumazet * sk->sk_wmem_alloc in an atomic, we can promote to rw. 3207ca6fb065SEric Dumazet */ 3208ca6fb065SEric Dumazet skb_set_owner_w(skb, (struct sock *)sk); 3209b3d05147SEric Dumazet break; 3210ca6fb065SEric Dumazet } 32114aea39c1SEric Dumazet skb_dst_set(skb, dst); 32121da177e4SLinus Torvalds 32133541f9e8SEric Dumazet mss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 3214f5fff5dcSTom Quetchenbach 321533ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 32168b5f12d0SFlorian Westphal #ifdef CONFIG_SYN_COOKIES 32178b5f12d0SFlorian Westphal if (unlikely(req->cookie_ts)) 32189a568de4SEric Dumazet skb->skb_mstamp = cookie_init_timestamp(req); 32198b5f12d0SFlorian Westphal else 32208b5f12d0SFlorian Westphal #endif 32219a568de4SEric Dumazet skb->skb_mstamp = tcp_clock_us(); 322280f03e27SEric Dumazet 322380f03e27SEric Dumazet #ifdef CONFIG_TCP_MD5SIG 322480f03e27SEric Dumazet rcu_read_lock(); 3225fd3a154aSEric Dumazet md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req)); 322680f03e27SEric Dumazet #endif 322758d607d3SEric Dumazet skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4); 322860e2a778SUrsula Braun tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5, 322960e2a778SUrsula Braun foc) + sizeof(*th); 323033ad798cSAdam Langley 3231aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 3232aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 32331da177e4SLinus Torvalds 3234ea1627c2SEric Dumazet th = (struct tcphdr *)skb->data; 32351da177e4SLinus Torvalds memset(th, 0, sizeof(struct tcphdr)); 32361da177e4SLinus Torvalds th->syn = 1; 32371da177e4SLinus Torvalds th->ack = 1; 32386ac705b1SEric Dumazet tcp_ecn_make_synack(req, th); 3239b44084c2SEric Dumazet th->source = htons(ireq->ir_num); 3240634fb979SEric Dumazet th->dest = ireq->ir_rmt_port; 3241e05a90ecSJamal Hadi Salim skb->mark = ireq->ir_mark; 32423b117750SEric Dumazet skb->ip_summed = CHECKSUM_PARTIAL; 32433b117750SEric Dumazet th->seq = htonl(tcp_rsk(req)->snt_isn); 32448336886fSJerry Chu /* XXX data is queued and acked as is. No buffer/window check */ 32458336886fSJerry Chu th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt); 32461da177e4SLinus Torvalds 32471da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 3248ed53d0abSEric Dumazet th->window = htons(min(req->rsk_rcv_wnd, 65535U)); 32495d062de7SEric Dumazet tcp_options_write((__be32 *)(th + 1), NULL, &opts); 32501da177e4SLinus Torvalds th->doff = (tcp_header_size >> 2); 325190bbcc60SEric Dumazet __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS); 3252cfb6eeb4SYOSHIFUJI Hideaki 3253cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 3254cfb6eeb4SYOSHIFUJI Hideaki /* Okay, we have all we need - do the md5 hash if needed */ 325580f03e27SEric Dumazet if (md5) 3256bd0388aeSWilliam Allen Simpson tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, 325739f8e58eSEric Dumazet md5, req_to_sk(req), skb); 325880f03e27SEric Dumazet rcu_read_unlock(); 3259cfb6eeb4SYOSHIFUJI Hideaki #endif 3260cfb6eeb4SYOSHIFUJI Hideaki 3261b50edd78SEric Dumazet /* Do not fool tcpdump (if any), clean our debris */ 32622456e855SThomas Gleixner skb->tstamp = 0; 32631da177e4SLinus Torvalds return skb; 32641da177e4SLinus Torvalds } 32654bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_make_synack); 32661da177e4SLinus Torvalds 326781164413SDaniel Borkmann static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst) 326881164413SDaniel Borkmann { 326981164413SDaniel Borkmann struct inet_connection_sock *icsk = inet_csk(sk); 327081164413SDaniel Borkmann const struct tcp_congestion_ops *ca; 327181164413SDaniel Borkmann u32 ca_key = dst_metric(dst, RTAX_CC_ALGO); 327281164413SDaniel Borkmann 327381164413SDaniel Borkmann if (ca_key == TCP_CA_UNSPEC) 327481164413SDaniel Borkmann return; 327581164413SDaniel Borkmann 327681164413SDaniel Borkmann rcu_read_lock(); 327781164413SDaniel Borkmann ca = tcp_ca_find_key(ca_key); 327881164413SDaniel Borkmann if (likely(ca && try_module_get(ca->owner))) { 327981164413SDaniel Borkmann module_put(icsk->icsk_ca_ops->owner); 328081164413SDaniel Borkmann icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst); 328181164413SDaniel Borkmann icsk->icsk_ca_ops = ca; 328281164413SDaniel Borkmann } 328381164413SDaniel Borkmann rcu_read_unlock(); 328481164413SDaniel Borkmann } 328581164413SDaniel Borkmann 328667edfef7SAndi Kleen /* Do all connect socket setups that can be done AF independent. */ 3287f7e56a76Sstephen hemminger static void tcp_connect_init(struct sock *sk) 32881da177e4SLinus Torvalds { 3289cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 32901da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 32911da177e4SLinus Torvalds __u8 rcv_wscale; 329213d3b1ebSLawrence Brakmo u32 rcv_wnd; 32931da177e4SLinus Torvalds 32941da177e4SLinus Torvalds /* We'll fix this up when we get a response from the other end. 32951da177e4SLinus Torvalds * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 32961da177e4SLinus Torvalds */ 32975d2ed052SEric Dumazet tp->tcp_header_len = sizeof(struct tcphdr); 32985d2ed052SEric Dumazet if (sock_net(sk)->ipv4.sysctl_tcp_timestamps) 32995d2ed052SEric Dumazet tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED; 33001da177e4SLinus Torvalds 3301cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 330200db4124SIan Morris if (tp->af_specific->md5_lookup(sk, sk)) 3303cfb6eeb4SYOSHIFUJI Hideaki tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 3304cfb6eeb4SYOSHIFUJI Hideaki #endif 3305cfb6eeb4SYOSHIFUJI Hideaki 33061da177e4SLinus Torvalds /* If user gave his TCP_MAXSEG, record it to clamp */ 33071da177e4SLinus Torvalds if (tp->rx_opt.user_mss) 33081da177e4SLinus Torvalds tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 33091da177e4SLinus Torvalds tp->max_window = 0; 33105d424d5aSJohn Heffner tcp_mtup_init(sk); 33111da177e4SLinus Torvalds tcp_sync_mss(sk, dst_mtu(dst)); 33121da177e4SLinus Torvalds 331381164413SDaniel Borkmann tcp_ca_dst_init(sk, dst); 331481164413SDaniel Borkmann 33151da177e4SLinus Torvalds if (!tp->window_clamp) 33161da177e4SLinus Torvalds tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 33173541f9e8SEric Dumazet tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 3318f5fff5dcSTom Quetchenbach 33191da177e4SLinus Torvalds tcp_initialize_rcv_mss(sk); 33201da177e4SLinus Torvalds 3321e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 3322e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 3323e88c64f0SHagen Paul Pfeifer (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) 3324e88c64f0SHagen Paul Pfeifer tp->window_clamp = tcp_full_space(sk); 3325e88c64f0SHagen Paul Pfeifer 332613d3b1ebSLawrence Brakmo rcv_wnd = tcp_rwnd_init_bpf(sk); 332713d3b1ebSLawrence Brakmo if (rcv_wnd == 0) 332813d3b1ebSLawrence Brakmo rcv_wnd = dst_metric(dst, RTAX_INITRWND); 332913d3b1ebSLawrence Brakmo 3330ceef9ab6SEric Dumazet tcp_select_initial_window(sk, tcp_full_space(sk), 33311da177e4SLinus Torvalds tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 33321da177e4SLinus Torvalds &tp->rcv_wnd, 33331da177e4SLinus Torvalds &tp->window_clamp, 33349bb37ef0SEric Dumazet sock_net(sk)->ipv4.sysctl_tcp_window_scaling, 333531d12926Slaurent chavey &rcv_wscale, 333613d3b1ebSLawrence Brakmo rcv_wnd); 33371da177e4SLinus Torvalds 33381da177e4SLinus Torvalds tp->rx_opt.rcv_wscale = rcv_wscale; 33391da177e4SLinus Torvalds tp->rcv_ssthresh = tp->rcv_wnd; 33401da177e4SLinus Torvalds 33411da177e4SLinus Torvalds sk->sk_err = 0; 33421da177e4SLinus Torvalds sock_reset_flag(sk, SOCK_DONE); 33431da177e4SLinus Torvalds tp->snd_wnd = 0; 3344ee7537b6SHantzis Fotis tcp_init_wl(tp, 0); 33451da177e4SLinus Torvalds tp->snd_una = tp->write_seq; 33461da177e4SLinus Torvalds tp->snd_sml = tp->write_seq; 334733f5f57eSIlpo Järvinen tp->snd_up = tp->write_seq; 3348370816aeSPavel Emelyanov tp->snd_nxt = tp->write_seq; 3349ee995283SPavel Emelyanov 3350ee995283SPavel Emelyanov if (likely(!tp->repair)) 33511da177e4SLinus Torvalds tp->rcv_nxt = 0; 3352c7781a6eSAndrew Vagin else 335370eabf0eSEric Dumazet tp->rcv_tstamp = tcp_jiffies32; 3354ee995283SPavel Emelyanov tp->rcv_wup = tp->rcv_nxt; 3355ee995283SPavel Emelyanov tp->copied_seq = tp->rcv_nxt; 33561da177e4SLinus Torvalds 33578550f328SLawrence Brakmo inet_csk(sk)->icsk_rto = tcp_timeout_init(sk); 3358463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_retransmits = 0; 33591da177e4SLinus Torvalds tcp_clear_retrans(tp); 33601da177e4SLinus Torvalds } 33611da177e4SLinus Torvalds 3362783237e8SYuchung Cheng static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) 3363783237e8SYuchung Cheng { 3364783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3365783237e8SYuchung Cheng struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 3366783237e8SYuchung Cheng 3367783237e8SYuchung Cheng tcb->end_seq += skb->len; 3368f4a775d1SEric Dumazet __skb_header_release(skb); 3369783237e8SYuchung Cheng sk->sk_wmem_queued += skb->truesize; 3370783237e8SYuchung Cheng sk_mem_charge(sk, skb->truesize); 3371783237e8SYuchung Cheng tp->write_seq = tcb->end_seq; 3372783237e8SYuchung Cheng tp->packets_out += tcp_skb_pcount(skb); 3373783237e8SYuchung Cheng } 3374783237e8SYuchung Cheng 3375783237e8SYuchung Cheng /* Build and send a SYN with data and (cached) Fast Open cookie. However, 3376783237e8SYuchung Cheng * queue a data-only packet after the regular SYN, such that regular SYNs 3377783237e8SYuchung Cheng * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges 3378783237e8SYuchung Cheng * only the SYN sequence, the data are retransmitted in the first ACK. 3379783237e8SYuchung Cheng * If cookie is not cached or other error occurs, falls back to send a 3380783237e8SYuchung Cheng * regular SYN with Fast Open cookie request option. 3381783237e8SYuchung Cheng */ 3382783237e8SYuchung Cheng static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) 3383783237e8SYuchung Cheng { 3384783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3385783237e8SYuchung Cheng struct tcp_fastopen_request *fo = tp->fastopen_req; 3386065263f4SWei Wang int space, err = 0; 3387355a901eSEric Dumazet struct sk_buff *syn_data; 3388783237e8SYuchung Cheng 338967da22d2SYuchung Cheng tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ 3390065263f4SWei Wang if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie)) 3391783237e8SYuchung Cheng goto fallback; 3392783237e8SYuchung Cheng 3393783237e8SYuchung Cheng /* MSS for SYN-data is based on cached MSS and bounded by PMTU and 3394783237e8SYuchung Cheng * user-MSS. Reserve maximum option space for middleboxes that add 3395783237e8SYuchung Cheng * private TCP options. The cost is reduced data space in SYN :( 3396783237e8SYuchung Cheng */ 33973541f9e8SEric Dumazet tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp); 33983541f9e8SEric Dumazet 33991b63edd6SYuchung Cheng space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) - 3400783237e8SYuchung Cheng MAX_TCP_OPTION_SPACE; 3401783237e8SYuchung Cheng 3402f5ddcbbbSEric Dumazet space = min_t(size_t, space, fo->size); 3403f5ddcbbbSEric Dumazet 3404f5ddcbbbSEric Dumazet /* limit to order-0 allocations */ 3405f5ddcbbbSEric Dumazet space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER)); 3406f5ddcbbbSEric Dumazet 3407eb934478SEric Dumazet syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false); 3408355a901eSEric Dumazet if (!syn_data) 3409783237e8SYuchung Cheng goto fallback; 3410355a901eSEric Dumazet syn_data->ip_summed = CHECKSUM_PARTIAL; 3411355a901eSEric Dumazet memcpy(syn_data->cb, syn->cb, sizeof(syn->cb)); 341207e100f9SEric Dumazet if (space) { 341307e100f9SEric Dumazet int copied = copy_from_iter(skb_put(syn_data, space), space, 341457be5bdaSAl Viro &fo->data->msg_iter); 341557be5bdaSAl Viro if (unlikely(!copied)) { 3416ba233b34SEric Dumazet tcp_skb_tsorted_anchor_cleanup(syn_data); 3417355a901eSEric Dumazet kfree_skb(syn_data); 3418783237e8SYuchung Cheng goto fallback; 3419783237e8SYuchung Cheng } 342057be5bdaSAl Viro if (copied != space) { 342157be5bdaSAl Viro skb_trim(syn_data, copied); 342257be5bdaSAl Viro space = copied; 342357be5bdaSAl Viro } 342407e100f9SEric Dumazet } 3425355a901eSEric Dumazet /* No more data pending in inet_wait_for_connect() */ 3426355a901eSEric Dumazet if (space == fo->size) 3427355a901eSEric Dumazet fo->data = NULL; 3428355a901eSEric Dumazet fo->copied = space; 3429783237e8SYuchung Cheng 3430355a901eSEric Dumazet tcp_connect_queue_skb(sk, syn_data); 34310f87230dSFrancis Yan if (syn_data->len) 34320f87230dSFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_BUSY); 3433355a901eSEric Dumazet 3434355a901eSEric Dumazet err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation); 3435355a901eSEric Dumazet 3436355a901eSEric Dumazet syn->skb_mstamp = syn_data->skb_mstamp; 3437355a901eSEric Dumazet 3438355a901eSEric Dumazet /* Now full SYN+DATA was cloned and sent (or not), 3439355a901eSEric Dumazet * remove the SYN from the original skb (syn_data) 3440355a901eSEric Dumazet * we keep in write queue in case of a retransmit, as we 3441355a901eSEric Dumazet * also have the SYN packet (with no data) in the same queue. 3442431a9124SEric Dumazet */ 3443355a901eSEric Dumazet TCP_SKB_CB(syn_data)->seq++; 3444355a901eSEric Dumazet TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH; 3445355a901eSEric Dumazet if (!err) { 344667da22d2SYuchung Cheng tp->syn_data = (fo->copied > 0); 344775c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data); 3448f19c29e3SYuchung Cheng NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT); 3449783237e8SYuchung Cheng goto done; 3450783237e8SYuchung Cheng } 3451783237e8SYuchung Cheng 345275c119afSEric Dumazet /* data was not sent, put it in write_queue */ 345375c119afSEric Dumazet __skb_queue_tail(&sk->sk_write_queue, syn_data); 3454b5b7db8dSEric Dumazet tp->packets_out -= tcp_skb_pcount(syn_data); 3455b5b7db8dSEric Dumazet 3456783237e8SYuchung Cheng fallback: 3457783237e8SYuchung Cheng /* Send a regular SYN with Fast Open cookie request option */ 3458783237e8SYuchung Cheng if (fo->cookie.len > 0) 3459783237e8SYuchung Cheng fo->cookie.len = 0; 3460783237e8SYuchung Cheng err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); 3461783237e8SYuchung Cheng if (err) 3462783237e8SYuchung Cheng tp->syn_fastopen = 0; 3463783237e8SYuchung Cheng done: 3464783237e8SYuchung Cheng fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ 3465783237e8SYuchung Cheng return err; 3466783237e8SYuchung Cheng } 3467783237e8SYuchung Cheng 346867edfef7SAndi Kleen /* Build a SYN and send it off. */ 34691da177e4SLinus Torvalds int tcp_connect(struct sock *sk) 34701da177e4SLinus Torvalds { 34711da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 34721da177e4SLinus Torvalds struct sk_buff *buff; 3473ee586811SEric Paris int err; 34741da177e4SLinus Torvalds 3475de525be2SLawrence Brakmo tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB, 0, NULL); 34768ba60924SEric Dumazet 34778ba60924SEric Dumazet if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 34788ba60924SEric Dumazet return -EHOSTUNREACH; /* Routing failure or similar. */ 34798ba60924SEric Dumazet 34801da177e4SLinus Torvalds tcp_connect_init(sk); 34811da177e4SLinus Torvalds 34822b916477SAndrey Vagin if (unlikely(tp->repair)) { 34832b916477SAndrey Vagin tcp_finish_connect(sk, NULL); 34842b916477SAndrey Vagin return 0; 34852b916477SAndrey Vagin } 34862b916477SAndrey Vagin 3487eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true); 3488355a901eSEric Dumazet if (unlikely(!buff)) 34891da177e4SLinus Torvalds return -ENOBUFS; 34901da177e4SLinus Torvalds 3491a3433f35SChangli Gao tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); 34929a568de4SEric Dumazet tcp_mstamp_refresh(tp); 34939a568de4SEric Dumazet tp->retrans_stamp = tcp_time_stamp(tp); 3494783237e8SYuchung Cheng tcp_connect_queue_skb(sk, buff); 3495735d3831SFlorian Westphal tcp_ecn_send_syn(sk, buff); 349675c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, buff); 34971da177e4SLinus Torvalds 3498783237e8SYuchung Cheng /* Send off SYN; include data in Fast Open. */ 3499783237e8SYuchung Cheng err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : 3500783237e8SYuchung Cheng tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); 3501ee586811SEric Paris if (err == -ECONNREFUSED) 3502ee586811SEric Paris return err; 3503bd37a088SWei Yongjun 3504bd37a088SWei Yongjun /* We change tp->snd_nxt after the tcp_transmit_skb() call 3505bd37a088SWei Yongjun * in order to make this packet get counted in tcpOutSegs. 3506bd37a088SWei Yongjun */ 3507bd37a088SWei Yongjun tp->snd_nxt = tp->write_seq; 3508bd37a088SWei Yongjun tp->pushed_seq = tp->write_seq; 3509b5b7db8dSEric Dumazet buff = tcp_send_head(sk); 3510b5b7db8dSEric Dumazet if (unlikely(buff)) { 3511b5b7db8dSEric Dumazet tp->snd_nxt = TCP_SKB_CB(buff)->seq; 3512b5b7db8dSEric Dumazet tp->pushed_seq = TCP_SKB_CB(buff)->seq; 3513b5b7db8dSEric Dumazet } 351481cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); 35151da177e4SLinus Torvalds 35161da177e4SLinus Torvalds /* Timer for repeating the SYN until an answer. */ 35173f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 35183f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 35191da177e4SLinus Torvalds return 0; 35201da177e4SLinus Torvalds } 35214bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_connect); 35221da177e4SLinus Torvalds 35231da177e4SLinus Torvalds /* Send out a delayed ack, the caller does the policy checking 35241da177e4SLinus Torvalds * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 35251da177e4SLinus Torvalds * for details. 35261da177e4SLinus Torvalds */ 35271da177e4SLinus Torvalds void tcp_send_delayed_ack(struct sock *sk) 35281da177e4SLinus Torvalds { 3529463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 3530463c84b9SArnaldo Carvalho de Melo int ato = icsk->icsk_ack.ato; 35311da177e4SLinus Torvalds unsigned long timeout; 35321da177e4SLinus Torvalds 35339890092eSFlorian Westphal tcp_ca_event(sk, CA_EVENT_DELAYED_ACK); 35349890092eSFlorian Westphal 35351da177e4SLinus Torvalds if (ato > TCP_DELACK_MIN) { 3536463c84b9SArnaldo Carvalho de Melo const struct tcp_sock *tp = tcp_sk(sk); 35371da177e4SLinus Torvalds int max_ato = HZ / 2; 35381da177e4SLinus Torvalds 3539056834d9SIlpo Järvinen if (icsk->icsk_ack.pingpong || 3540056834d9SIlpo Järvinen (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 35411da177e4SLinus Torvalds max_ato = TCP_DELACK_MAX; 35421da177e4SLinus Torvalds 35431da177e4SLinus Torvalds /* Slow path, intersegment interval is "high". */ 35441da177e4SLinus Torvalds 35451da177e4SLinus Torvalds /* If some rtt estimate is known, use it to bound delayed ack. 3546463c84b9SArnaldo Carvalho de Melo * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 35471da177e4SLinus Torvalds * directly. 35481da177e4SLinus Torvalds */ 3549740b0f18SEric Dumazet if (tp->srtt_us) { 3550740b0f18SEric Dumazet int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3), 3551740b0f18SEric Dumazet TCP_DELACK_MIN); 35521da177e4SLinus Torvalds 35531da177e4SLinus Torvalds if (rtt < max_ato) 35541da177e4SLinus Torvalds max_ato = rtt; 35551da177e4SLinus Torvalds } 35561da177e4SLinus Torvalds 35571da177e4SLinus Torvalds ato = min(ato, max_ato); 35581da177e4SLinus Torvalds } 35591da177e4SLinus Torvalds 35601da177e4SLinus Torvalds /* Stay within the limit we were given */ 35611da177e4SLinus Torvalds timeout = jiffies + ato; 35621da177e4SLinus Torvalds 35631da177e4SLinus Torvalds /* Use new timeout only if there wasn't a older one earlier. */ 3564463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 35651da177e4SLinus Torvalds /* If delack timer was blocked or is about to expire, 35661da177e4SLinus Torvalds * send ACK now. 35671da177e4SLinus Torvalds */ 3568463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.blocked || 3569463c84b9SArnaldo Carvalho de Melo time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 35701da177e4SLinus Torvalds tcp_send_ack(sk); 35711da177e4SLinus Torvalds return; 35721da177e4SLinus Torvalds } 35731da177e4SLinus Torvalds 3574463c84b9SArnaldo Carvalho de Melo if (!time_before(timeout, icsk->icsk_ack.timeout)) 3575463c84b9SArnaldo Carvalho de Melo timeout = icsk->icsk_ack.timeout; 35761da177e4SLinus Torvalds } 3577463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 3578463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.timeout = timeout; 3579463c84b9SArnaldo Carvalho de Melo sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 35801da177e4SLinus Torvalds } 35811da177e4SLinus Torvalds 35821da177e4SLinus Torvalds /* This routine sends an ack and also updates the window. */ 35831da177e4SLinus Torvalds void tcp_send_ack(struct sock *sk) 35841da177e4SLinus Torvalds { 35851da177e4SLinus Torvalds struct sk_buff *buff; 35861da177e4SLinus Torvalds 3587058dc334SIlpo Järvinen /* If we have been reset, we may not send again. */ 3588058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3589058dc334SIlpo Järvinen return; 3590058dc334SIlpo Järvinen 35919890092eSFlorian Westphal tcp_ca_event(sk, CA_EVENT_NON_DELAYED_ACK); 35929890092eSFlorian Westphal 35931da177e4SLinus Torvalds /* We are not putting this on the write queue, so 35941da177e4SLinus Torvalds * tcp_transmit_skb() will set the ownership to this 35951da177e4SLinus Torvalds * sock. 35961da177e4SLinus Torvalds */ 35977450aaf6SEric Dumazet buff = alloc_skb(MAX_TCP_HEADER, 35987450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 35997450aaf6SEric Dumazet if (unlikely(!buff)) { 3600463c84b9SArnaldo Carvalho de Melo inet_csk_schedule_ack(sk); 3601463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; 36023f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 36033f421baaSArnaldo Carvalho de Melo TCP_DELACK_MAX, TCP_RTO_MAX); 36041da177e4SLinus Torvalds return; 36051da177e4SLinus Torvalds } 36061da177e4SLinus Torvalds 36071da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 36081da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 3609a3433f35SChangli Gao tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); 36101da177e4SLinus Torvalds 361198781965SEric Dumazet /* We do not want pure acks influencing TCP Small Queues or fq/pacing 361298781965SEric Dumazet * too much. 361398781965SEric Dumazet * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784 361498781965SEric Dumazet */ 361598781965SEric Dumazet skb_set_tcp_pure_ack(buff); 361698781965SEric Dumazet 36171da177e4SLinus Torvalds /* Send it off, this clears delayed acks for us. */ 36187450aaf6SEric Dumazet tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0); 36191da177e4SLinus Torvalds } 3620e3118e83SDaniel Borkmann EXPORT_SYMBOL_GPL(tcp_send_ack); 36211da177e4SLinus Torvalds 36221da177e4SLinus Torvalds /* This routine sends a packet with an out of date sequence 36231da177e4SLinus Torvalds * number. It assumes the other end will try to ack it. 36241da177e4SLinus Torvalds * 36251da177e4SLinus Torvalds * Question: what should we make while urgent mode? 36261da177e4SLinus Torvalds * 4.4BSD forces sending single byte of data. We cannot send 36271da177e4SLinus Torvalds * out of window data, because we have SND.NXT==SND.MAX... 36281da177e4SLinus Torvalds * 36291da177e4SLinus Torvalds * Current solution: to send TWO zero-length segments in urgent mode: 36301da177e4SLinus Torvalds * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 36311da177e4SLinus Torvalds * out-of-date with SND.UNA-1 to probe window. 36321da177e4SLinus Torvalds */ 3633e520af48SEric Dumazet static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib) 36341da177e4SLinus Torvalds { 36351da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 36361da177e4SLinus Torvalds struct sk_buff *skb; 36371da177e4SLinus Torvalds 36381da177e4SLinus Torvalds /* We don't queue it, tcp_transmit_skb() sets ownership. */ 36397450aaf6SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, 36407450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 364151456b29SIan Morris if (!skb) 36421da177e4SLinus Torvalds return -1; 36431da177e4SLinus Torvalds 36441da177e4SLinus Torvalds /* Reserve space for headers and set control bits. */ 36451da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 36461da177e4SLinus Torvalds /* Use a previous sequence. This should cause the other 36471da177e4SLinus Torvalds * end to send an ack. Don't queue or clone SKB, just 36481da177e4SLinus Torvalds * send it. 36491da177e4SLinus Torvalds */ 3650a3433f35SChangli Gao tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); 3651e2e8009fSRenato Westphal NET_INC_STATS(sock_net(sk), mib); 36527450aaf6SEric Dumazet return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0); 36531da177e4SLinus Torvalds } 36541da177e4SLinus Torvalds 3655385e2070SEric Dumazet /* Called from setsockopt( ... TCP_REPAIR ) */ 3656ee995283SPavel Emelyanov void tcp_send_window_probe(struct sock *sk) 3657ee995283SPavel Emelyanov { 3658ee995283SPavel Emelyanov if (sk->sk_state == TCP_ESTABLISHED) { 3659ee995283SPavel Emelyanov tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; 36609a568de4SEric Dumazet tcp_mstamp_refresh(tcp_sk(sk)); 3661e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE); 3662ee995283SPavel Emelyanov } 3663ee995283SPavel Emelyanov } 3664ee995283SPavel Emelyanov 366567edfef7SAndi Kleen /* Initiate keepalive or window probe from timer. */ 3666e520af48SEric Dumazet int tcp_write_wakeup(struct sock *sk, int mib) 36671da177e4SLinus Torvalds { 36681da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 36691da177e4SLinus Torvalds struct sk_buff *skb; 36701da177e4SLinus Torvalds 3671058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3672058dc334SIlpo Järvinen return -1; 3673058dc334SIlpo Järvinen 367400db4124SIan Morris skb = tcp_send_head(sk); 367500db4124SIan Morris if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { 36761da177e4SLinus Torvalds int err; 36770c54b85fSIlpo Järvinen unsigned int mss = tcp_current_mss(sk); 367890840defSIlpo Järvinen unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 36791da177e4SLinus Torvalds 36801da177e4SLinus Torvalds if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 36811da177e4SLinus Torvalds tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 36821da177e4SLinus Torvalds 36831da177e4SLinus Torvalds /* We are probing the opening of a window 36841da177e4SLinus Torvalds * but the window size is != 0 36851da177e4SLinus Torvalds * must have been a result SWS avoidance ( sender ) 36861da177e4SLinus Torvalds */ 36871da177e4SLinus Torvalds if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 36881da177e4SLinus Torvalds skb->len > mss) { 36891da177e4SLinus Torvalds seg_size = min(seg_size, mss); 36904de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 369175c119afSEric Dumazet if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE, 369275c119afSEric Dumazet skb, seg_size, mss, GFP_ATOMIC)) 36931da177e4SLinus Torvalds return -1; 36941da177e4SLinus Torvalds } else if (!tcp_skb_pcount(skb)) 36955bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss); 36961da177e4SLinus Torvalds 36974de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 3698dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 369966f5fe62SIlpo Järvinen if (!err) 370066f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 37011da177e4SLinus Torvalds return err; 37021da177e4SLinus Torvalds } else { 370333f5f57eSIlpo Järvinen if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) 3704e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 1, mib); 3705e520af48SEric Dumazet return tcp_xmit_probe_skb(sk, 0, mib); 37061da177e4SLinus Torvalds } 37071da177e4SLinus Torvalds } 37081da177e4SLinus Torvalds 37091da177e4SLinus Torvalds /* A window probe timeout has occurred. If window is not closed send 37101da177e4SLinus Torvalds * a partial packet else a zero probe. 37111da177e4SLinus Torvalds */ 37121da177e4SLinus Torvalds void tcp_send_probe0(struct sock *sk) 37131da177e4SLinus Torvalds { 3714463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 37151da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 3716c6214a97SNikolay Borisov struct net *net = sock_net(sk); 3717fcdd1cf4SEric Dumazet unsigned long probe_max; 37181da177e4SLinus Torvalds int err; 37191da177e4SLinus Torvalds 3720e520af48SEric Dumazet err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE); 37211da177e4SLinus Torvalds 372275c119afSEric Dumazet if (tp->packets_out || tcp_write_queue_empty(sk)) { 37231da177e4SLinus Torvalds /* Cancel probe timer, if it is not required. */ 37246687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 0; 3725463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff = 0; 37261da177e4SLinus Torvalds return; 37271da177e4SLinus Torvalds } 37281da177e4SLinus Torvalds 37291da177e4SLinus Torvalds if (err <= 0) { 3730c6214a97SNikolay Borisov if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2) 3731463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff++; 37326687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out++; 3733fcdd1cf4SEric Dumazet probe_max = TCP_RTO_MAX; 37341da177e4SLinus Torvalds } else { 37351da177e4SLinus Torvalds /* If packet was not sent due to local congestion, 37366687e988SArnaldo Carvalho de Melo * do not backoff and do not remember icsk_probes_out. 37371da177e4SLinus Torvalds * Let local senders to fight for local resources. 37381da177e4SLinus Torvalds * 37391da177e4SLinus Torvalds * Use accumulated backoff yet. 37401da177e4SLinus Torvalds */ 37416687e988SArnaldo Carvalho de Melo if (!icsk->icsk_probes_out) 37426687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 1; 3743fcdd1cf4SEric Dumazet probe_max = TCP_RESOURCE_PROBE_INTERVAL; 37441da177e4SLinus Torvalds } 3745fcdd1cf4SEric Dumazet inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 374621c8fe99SEric Dumazet tcp_probe0_when(sk, probe_max), 3747fcdd1cf4SEric Dumazet TCP_RTO_MAX); 37481da177e4SLinus Torvalds } 37495db92c99SOctavian Purdila 3750ea3bea3aSEric Dumazet int tcp_rtx_synack(const struct sock *sk, struct request_sock *req) 37515db92c99SOctavian Purdila { 37525db92c99SOctavian Purdila const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific; 37535db92c99SOctavian Purdila struct flowi fl; 37545db92c99SOctavian Purdila int res; 37555db92c99SOctavian Purdila 375658d607d3SEric Dumazet tcp_rsk(req)->txhash = net_tx_rndhash(); 3757b3d05147SEric Dumazet res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL); 37585db92c99SOctavian Purdila if (!res) { 375990bbcc60SEric Dumazet __TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); 376002a1d6e7SEric Dumazet __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 37617e32b443SYuchung Cheng if (unlikely(tcp_passive_fastopen(sk))) 37627e32b443SYuchung Cheng tcp_sk(sk)->total_retrans++; 3763cf34ce3dSSong Liu trace_tcp_retransmit_synack(sk, req); 37645db92c99SOctavian Purdila } 37655db92c99SOctavian Purdila return res; 37665db92c99SOctavian Purdila } 37675db92c99SOctavian Purdila EXPORT_SYMBOL(tcp_rtx_synack); 3768