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 { 12098f26fb1cSEric Dumazet if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) { 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 133884fa7933SPatrick McHardy if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { 13391da177e4SLinus Torvalds /* Copy and checksum data tail into the new buffer. */ 1340056834d9SIlpo Järvinen buff->csum = csum_partial_copy_nocheck(skb->data + len, 1341056834d9SIlpo Järvinen skb_put(buff, nsize), 13421da177e4SLinus Torvalds nsize, 0); 13431da177e4SLinus Torvalds 13441da177e4SLinus Torvalds skb_trim(skb, len); 13451da177e4SLinus Torvalds 13461da177e4SLinus Torvalds skb->csum = csum_block_sub(skb->csum, buff->csum, len); 13471da177e4SLinus Torvalds } else { 134884fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 13491da177e4SLinus Torvalds skb_split(skb, buff, len); 13501da177e4SLinus Torvalds } 13511da177e4SLinus Torvalds 13521da177e4SLinus Torvalds buff->ip_summed = skb->ip_summed; 13531da177e4SLinus Torvalds 1354a61bbcf2SPatrick McHardy buff->tstamp = skb->tstamp; 1355490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 13561da177e4SLinus Torvalds 13576475be16SDavid S. Miller old_factor = tcp_skb_pcount(skb); 13586475be16SDavid S. Miller 13591da177e4SLinus Torvalds /* Fix up tso_factor for both original and new SKB. */ 13605bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 13615bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 13621da177e4SLinus Torvalds 1363b9f64820SYuchung Cheng /* Update delivered info for the new segment */ 1364b9f64820SYuchung Cheng TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx; 1365b9f64820SYuchung Cheng 13666475be16SDavid S. Miller /* If this packet has been sent out already, we must 13676475be16SDavid S. Miller * adjust the various packet counters. 13686475be16SDavid S. Miller */ 1369cf0b450cSHerbert Xu if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 13706475be16SDavid S. Miller int diff = old_factor - tcp_skb_pcount(skb) - 13716475be16SDavid S. Miller tcp_skb_pcount(buff); 13721da177e4SLinus Torvalds 1373797108d1SIlpo Järvinen if (diff) 1374797108d1SIlpo Järvinen tcp_adjust_pcount(sk, skb, diff); 13751da177e4SLinus Torvalds } 13761da177e4SLinus Torvalds 13771da177e4SLinus Torvalds /* Link BUFF into the send queue. */ 1378f4a775d1SEric Dumazet __skb_header_release(buff); 137975c119afSEric Dumazet tcp_insert_write_queue_after(skb, buff, sk, tcp_queue); 1380f67971e6SEric Dumazet if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE) 1381e2080072SEric Dumazet list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor); 13821da177e4SLinus Torvalds 13831da177e4SLinus Torvalds return 0; 13841da177e4SLinus Torvalds } 13851da177e4SLinus Torvalds 1386f4d01666SEric Dumazet /* This is similar to __pskb_pull_tail(). The difference is that pulled 1387f4d01666SEric Dumazet * data is not copied, but immediately discarded. 13881da177e4SLinus Torvalds */ 13897162fb24SEric Dumazet static int __pskb_trim_head(struct sk_buff *skb, int len) 13901da177e4SLinus Torvalds { 13917b7fc97aSEric Dumazet struct skb_shared_info *shinfo; 13921da177e4SLinus Torvalds int i, k, eat; 13931da177e4SLinus Torvalds 13944fa48bf3SEric Dumazet eat = min_t(int, len, skb_headlen(skb)); 13954fa48bf3SEric Dumazet if (eat) { 13964fa48bf3SEric Dumazet __skb_pull(skb, eat); 13974fa48bf3SEric Dumazet len -= eat; 13984fa48bf3SEric Dumazet if (!len) 13997162fb24SEric Dumazet return 0; 14004fa48bf3SEric Dumazet } 14011da177e4SLinus Torvalds eat = len; 14021da177e4SLinus Torvalds k = 0; 14037b7fc97aSEric Dumazet shinfo = skb_shinfo(skb); 14047b7fc97aSEric Dumazet for (i = 0; i < shinfo->nr_frags; i++) { 14057b7fc97aSEric Dumazet int size = skb_frag_size(&shinfo->frags[i]); 14069e903e08SEric Dumazet 14079e903e08SEric Dumazet if (size <= eat) { 1408aff65da0SIan Campbell skb_frag_unref(skb, i); 14099e903e08SEric Dumazet eat -= size; 14101da177e4SLinus Torvalds } else { 14117b7fc97aSEric Dumazet shinfo->frags[k] = shinfo->frags[i]; 14121da177e4SLinus Torvalds if (eat) { 14137b7fc97aSEric Dumazet shinfo->frags[k].page_offset += eat; 14147b7fc97aSEric Dumazet skb_frag_size_sub(&shinfo->frags[k], eat); 14151da177e4SLinus Torvalds eat = 0; 14161da177e4SLinus Torvalds } 14171da177e4SLinus Torvalds k++; 14181da177e4SLinus Torvalds } 14191da177e4SLinus Torvalds } 14207b7fc97aSEric Dumazet shinfo->nr_frags = k; 14211da177e4SLinus Torvalds 14221da177e4SLinus Torvalds skb->data_len -= len; 14231da177e4SLinus Torvalds skb->len = skb->data_len; 14247162fb24SEric Dumazet return len; 14251da177e4SLinus Torvalds } 14261da177e4SLinus Torvalds 142767edfef7SAndi Kleen /* Remove acked data from a packet in the transmit queue. */ 14281da177e4SLinus Torvalds int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 14291da177e4SLinus Torvalds { 14307162fb24SEric Dumazet u32 delta_truesize; 14317162fb24SEric Dumazet 143214bbd6a5SPravin B Shelar if (skb_unclone(skb, GFP_ATOMIC)) 14331da177e4SLinus Torvalds return -ENOMEM; 14341da177e4SLinus Torvalds 14357162fb24SEric Dumazet delta_truesize = __pskb_trim_head(skb, len); 14361da177e4SLinus Torvalds 14371da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq += len; 143884fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 14391da177e4SLinus Torvalds 14407162fb24SEric Dumazet if (delta_truesize) { 14417162fb24SEric Dumazet skb->truesize -= delta_truesize; 14427162fb24SEric Dumazet sk->sk_wmem_queued -= delta_truesize; 14437162fb24SEric Dumazet sk_mem_uncharge(sk, delta_truesize); 14441da177e4SLinus Torvalds sock_set_flag(sk, SOCK_QUEUE_SHRUNK); 14457162fb24SEric Dumazet } 14461da177e4SLinus Torvalds 14475b35e1e6SNeal Cardwell /* Any change of skb->len requires recalculation of tso factor. */ 14481da177e4SLinus Torvalds if (tcp_skb_pcount(skb) > 1) 14495bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb)); 14501da177e4SLinus Torvalds 14511da177e4SLinus Torvalds return 0; 14521da177e4SLinus Torvalds } 14531da177e4SLinus Torvalds 14541b63edd6SYuchung Cheng /* Calculate MSS not accounting any TCP options. */ 14551b63edd6SYuchung Cheng static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu) 14565d424d5aSJohn Heffner { 1457cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1458cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 14595d424d5aSJohn Heffner int mss_now; 14605d424d5aSJohn Heffner 14615d424d5aSJohn Heffner /* Calculate base mss without TCP options: 14625d424d5aSJohn Heffner It is MMS_S - sizeof(tcphdr) of rfc1122 14635d424d5aSJohn Heffner */ 14645d424d5aSJohn Heffner mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 14655d424d5aSJohn Heffner 146667469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 146767469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 146867469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 146967469601SEric Dumazet 147067469601SEric Dumazet if (dst && dst_allfrag(dst)) 147167469601SEric Dumazet mss_now -= icsk->icsk_af_ops->net_frag_header_len; 147267469601SEric Dumazet } 147367469601SEric Dumazet 14745d424d5aSJohn Heffner /* Clamp it (mss_clamp does not include tcp options) */ 14755d424d5aSJohn Heffner if (mss_now > tp->rx_opt.mss_clamp) 14765d424d5aSJohn Heffner mss_now = tp->rx_opt.mss_clamp; 14775d424d5aSJohn Heffner 14785d424d5aSJohn Heffner /* Now subtract optional transport overhead */ 14795d424d5aSJohn Heffner mss_now -= icsk->icsk_ext_hdr_len; 14805d424d5aSJohn Heffner 14815d424d5aSJohn Heffner /* Then reserve room for full set of TCP options and 8 bytes of data */ 14825d424d5aSJohn Heffner if (mss_now < 48) 14835d424d5aSJohn Heffner mss_now = 48; 14845d424d5aSJohn Heffner return mss_now; 14855d424d5aSJohn Heffner } 14865d424d5aSJohn Heffner 14871b63edd6SYuchung Cheng /* Calculate MSS. Not accounting for SACKs here. */ 14881b63edd6SYuchung Cheng int tcp_mtu_to_mss(struct sock *sk, int pmtu) 14891b63edd6SYuchung Cheng { 14901b63edd6SYuchung Cheng /* Subtract TCP options size, not including SACKs */ 14911b63edd6SYuchung Cheng return __tcp_mtu_to_mss(sk, pmtu) - 14921b63edd6SYuchung Cheng (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr)); 14931b63edd6SYuchung Cheng } 14941b63edd6SYuchung Cheng 14955d424d5aSJohn Heffner /* Inverse of above */ 149667469601SEric Dumazet int tcp_mss_to_mtu(struct sock *sk, int mss) 14975d424d5aSJohn Heffner { 1498cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1499cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 15005d424d5aSJohn Heffner int mtu; 15015d424d5aSJohn Heffner 15025d424d5aSJohn Heffner mtu = mss + 15035d424d5aSJohn Heffner tp->tcp_header_len + 15045d424d5aSJohn Heffner icsk->icsk_ext_hdr_len + 15055d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 15065d424d5aSJohn Heffner 150767469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 150867469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 150967469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 151067469601SEric Dumazet 151167469601SEric Dumazet if (dst && dst_allfrag(dst)) 151267469601SEric Dumazet mtu += icsk->icsk_af_ops->net_frag_header_len; 151367469601SEric Dumazet } 15145d424d5aSJohn Heffner return mtu; 15155d424d5aSJohn Heffner } 1516556c6b46SNeal Cardwell EXPORT_SYMBOL(tcp_mss_to_mtu); 15175d424d5aSJohn Heffner 151867edfef7SAndi Kleen /* MTU probing init per socket */ 15195d424d5aSJohn Heffner void tcp_mtup_init(struct sock *sk) 15205d424d5aSJohn Heffner { 15215d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 15225d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 1523b0f9ca53SFan Du struct net *net = sock_net(sk); 15245d424d5aSJohn Heffner 1525b0f9ca53SFan Du icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1; 15265d424d5aSJohn Heffner icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 15275d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 1528b0f9ca53SFan Du icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss); 15295d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 153005cbc0dbSFan Du if (icsk->icsk_mtup.enabled) 1531c74df29aSEric Dumazet icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; 15325d424d5aSJohn Heffner } 15334bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_mtup_init); 15345d424d5aSJohn Heffner 15351da177e4SLinus Torvalds /* This function synchronize snd mss to current pmtu/exthdr set. 15361da177e4SLinus Torvalds 15371da177e4SLinus Torvalds tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 15381da177e4SLinus Torvalds for TCP options, but includes only bare TCP header. 15391da177e4SLinus Torvalds 15401da177e4SLinus Torvalds tp->rx_opt.mss_clamp is mss negotiated at connection setup. 1541caa20d9aSStephen Hemminger It is minimum of user_mss and mss received with SYN. 15421da177e4SLinus Torvalds It also does not include TCP options. 15431da177e4SLinus Torvalds 1544d83d8461SArnaldo Carvalho de Melo inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. 15451da177e4SLinus Torvalds 15461da177e4SLinus Torvalds tp->mss_cache is current effective sending mss, including 15471da177e4SLinus Torvalds all tcp options except for SACKs. It is evaluated, 15481da177e4SLinus Torvalds taking into account current pmtu, but never exceeds 15491da177e4SLinus Torvalds tp->rx_opt.mss_clamp. 15501da177e4SLinus Torvalds 15511da177e4SLinus Torvalds NOTE1. rfc1122 clearly states that advertised MSS 15521da177e4SLinus Torvalds DOES NOT include either tcp or ip options. 15531da177e4SLinus Torvalds 1554d83d8461SArnaldo Carvalho de Melo NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache 1555d83d8461SArnaldo Carvalho de Melo are READ ONLY outside this function. --ANK (980731) 15561da177e4SLinus Torvalds */ 15571da177e4SLinus Torvalds unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 15581da177e4SLinus Torvalds { 15591da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1560d83d8461SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 15615d424d5aSJohn Heffner int mss_now; 15621da177e4SLinus Torvalds 15635d424d5aSJohn Heffner if (icsk->icsk_mtup.search_high > pmtu) 15645d424d5aSJohn Heffner icsk->icsk_mtup.search_high = pmtu; 15651da177e4SLinus Torvalds 15665d424d5aSJohn Heffner mss_now = tcp_mtu_to_mss(sk, pmtu); 1567409d22b4SIlpo Järvinen mss_now = tcp_bound_to_half_wnd(tp, mss_now); 15681da177e4SLinus Torvalds 15691da177e4SLinus Torvalds /* And store cached results */ 1570d83d8461SArnaldo Carvalho de Melo icsk->icsk_pmtu_cookie = pmtu; 15715d424d5aSJohn Heffner if (icsk->icsk_mtup.enabled) 15725d424d5aSJohn Heffner mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); 1573c1b4a7e6SDavid S. Miller tp->mss_cache = mss_now; 15741da177e4SLinus Torvalds 15751da177e4SLinus Torvalds return mss_now; 15761da177e4SLinus Torvalds } 15774bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_sync_mss); 15781da177e4SLinus Torvalds 15791da177e4SLinus Torvalds /* Compute the current effective MSS, taking SACKs and IP options, 15801da177e4SLinus Torvalds * and even PMTU discovery events into account. 15811da177e4SLinus Torvalds */ 15820c54b85fSIlpo Järvinen unsigned int tcp_current_mss(struct sock *sk) 15831da177e4SLinus Torvalds { 1584cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1585cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1586c1b4a7e6SDavid S. Miller u32 mss_now; 158795c96174SEric Dumazet unsigned int header_len; 158833ad798cSAdam Langley struct tcp_out_options opts; 158933ad798cSAdam Langley struct tcp_md5sig_key *md5; 15901da177e4SLinus Torvalds 1591c1b4a7e6SDavid S. Miller mss_now = tp->mss_cache; 1592c1b4a7e6SDavid S. Miller 15931da177e4SLinus Torvalds if (dst) { 15941da177e4SLinus Torvalds u32 mtu = dst_mtu(dst); 1595d83d8461SArnaldo Carvalho de Melo if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 15961da177e4SLinus Torvalds mss_now = tcp_sync_mss(sk, mtu); 15971da177e4SLinus Torvalds } 15981da177e4SLinus Torvalds 159933ad798cSAdam Langley header_len = tcp_established_options(sk, NULL, &opts, &md5) + 160033ad798cSAdam Langley sizeof(struct tcphdr); 160133ad798cSAdam Langley /* The mss_cache is sized based on tp->tcp_header_len, which assumes 160233ad798cSAdam Langley * some common options. If this is an odd packet (because we have SACK 160333ad798cSAdam Langley * blocks etc) then our calculated header_len will be different, and 160433ad798cSAdam Langley * we have to adjust mss_now correspondingly */ 160533ad798cSAdam Langley if (header_len != tp->tcp_header_len) { 160633ad798cSAdam Langley int delta = (int) header_len - tp->tcp_header_len; 160733ad798cSAdam Langley mss_now -= delta; 160833ad798cSAdam Langley } 1609cfb6eeb4SYOSHIFUJI Hideaki 16101da177e4SLinus Torvalds return mss_now; 16111da177e4SLinus Torvalds } 16121da177e4SLinus Torvalds 161386fd14adSWeiping Pan /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. 161486fd14adSWeiping Pan * As additional protections, we do not touch cwnd in retransmission phases, 161586fd14adSWeiping Pan * and if application hit its sndbuf limit recently. 161686fd14adSWeiping Pan */ 161786fd14adSWeiping Pan static void tcp_cwnd_application_limited(struct sock *sk) 1618a762a980SDavid S. Miller { 16199e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1620a762a980SDavid S. Miller 162186fd14adSWeiping Pan if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && 162286fd14adSWeiping Pan sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 162386fd14adSWeiping Pan /* Limited by application or receiver window. */ 162486fd14adSWeiping Pan u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); 162586fd14adSWeiping Pan u32 win_used = max(tp->snd_cwnd_used, init_win); 162686fd14adSWeiping Pan if (win_used < tp->snd_cwnd) { 162786fd14adSWeiping Pan tp->snd_ssthresh = tcp_current_ssthresh(sk); 162886fd14adSWeiping Pan tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1; 162986fd14adSWeiping Pan } 163086fd14adSWeiping Pan tp->snd_cwnd_used = 0; 163186fd14adSWeiping Pan } 1632c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 163386fd14adSWeiping Pan } 163486fd14adSWeiping Pan 1635ca8a2263SNeal Cardwell static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited) 1636a762a980SDavid S. Miller { 16371b1fc3fdSWei Wang const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 1638a762a980SDavid S. Miller struct tcp_sock *tp = tcp_sk(sk); 1639a762a980SDavid S. Miller 1640ca8a2263SNeal Cardwell /* Track the maximum number of outstanding packets in each 1641ca8a2263SNeal Cardwell * window, and remember whether we were cwnd-limited then. 1642ca8a2263SNeal Cardwell */ 1643ca8a2263SNeal Cardwell if (!before(tp->snd_una, tp->max_packets_seq) || 1644ca8a2263SNeal Cardwell tp->packets_out > tp->max_packets_out) { 1645ca8a2263SNeal Cardwell tp->max_packets_out = tp->packets_out; 1646ca8a2263SNeal Cardwell tp->max_packets_seq = tp->snd_nxt; 1647ca8a2263SNeal Cardwell tp->is_cwnd_limited = is_cwnd_limited; 1648ca8a2263SNeal Cardwell } 1649e114a710SEric Dumazet 165024901551SEric Dumazet if (tcp_is_cwnd_limited(sk)) { 1651a762a980SDavid S. Miller /* Network is feed fully. */ 1652a762a980SDavid S. Miller tp->snd_cwnd_used = 0; 1653c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 1654a762a980SDavid S. Miller } else { 1655a762a980SDavid S. Miller /* Network starves. */ 1656a762a980SDavid S. Miller if (tp->packets_out > tp->snd_cwnd_used) 1657a762a980SDavid S. Miller tp->snd_cwnd_used = tp->packets_out; 1658a762a980SDavid S. Miller 1659b510f0d2SEric Dumazet if (sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle && 1660c2203cf7SEric Dumazet (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto && 16611b1fc3fdSWei Wang !ca_ops->cong_control) 1662a762a980SDavid S. Miller tcp_cwnd_application_limited(sk); 1663b0f71bd3SFrancis Yan 1664b0f71bd3SFrancis Yan /* The following conditions together indicate the starvation 1665b0f71bd3SFrancis Yan * is caused by insufficient sender buffer: 1666b0f71bd3SFrancis Yan * 1) just sent some data (see tcp_write_xmit) 1667b0f71bd3SFrancis Yan * 2) not cwnd limited (this else condition) 166875c119afSEric Dumazet * 3) no more data to send (tcp_write_queue_empty()) 1669b0f71bd3SFrancis Yan * 4) application is hitting buffer limit (SOCK_NOSPACE) 1670b0f71bd3SFrancis Yan */ 167175c119afSEric Dumazet if (tcp_write_queue_empty(sk) && sk->sk_socket && 1672b0f71bd3SFrancis Yan test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) && 1673b0f71bd3SFrancis Yan (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) 1674b0f71bd3SFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED); 1675a762a980SDavid S. Miller } 1676a762a980SDavid S. Miller } 1677a762a980SDavid S. Miller 1678d4589926SEric Dumazet /* Minshall's variant of the Nagle send check. */ 1679d4589926SEric Dumazet static bool tcp_minshall_check(const struct tcp_sock *tp) 1680d4589926SEric Dumazet { 1681d4589926SEric Dumazet return after(tp->snd_sml, tp->snd_una) && 1682d4589926SEric Dumazet !after(tp->snd_sml, tp->snd_nxt); 1683d4589926SEric Dumazet } 1684d4589926SEric Dumazet 1685d4589926SEric Dumazet /* Update snd_sml if this skb is under mss 1686d4589926SEric Dumazet * Note that a TSO packet might end with a sub-mss segment 1687d4589926SEric Dumazet * The test is really : 1688d4589926SEric Dumazet * if ((skb->len % mss) != 0) 1689d4589926SEric Dumazet * tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1690d4589926SEric Dumazet * But we can avoid doing the divide again given we already have 1691d4589926SEric Dumazet * skb_pcount = skb->len / mss_now 16920e3a4803SIlpo Järvinen */ 1693d4589926SEric Dumazet static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now, 1694d4589926SEric Dumazet const struct sk_buff *skb) 1695d4589926SEric Dumazet { 1696d4589926SEric Dumazet if (skb->len < tcp_skb_pcount(skb) * mss_now) 1697d4589926SEric Dumazet tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1698d4589926SEric Dumazet } 1699d4589926SEric Dumazet 1700d4589926SEric Dumazet /* Return false, if packet can be sent now without violation Nagle's rules: 1701d4589926SEric Dumazet * 1. It is full sized. (provided by caller in %partial bool) 1702d4589926SEric Dumazet * 2. Or it contains FIN. (already checked by caller) 1703d4589926SEric Dumazet * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. 1704d4589926SEric Dumazet * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 1705d4589926SEric Dumazet * With Minshall's modification: all sent small packets are ACKed. 1706d4589926SEric Dumazet */ 1707d4589926SEric Dumazet static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp, 1708cc93fc51SPeter Pan(潘卫平) int nonagle) 1709d4589926SEric Dumazet { 1710d4589926SEric Dumazet return partial && 1711d4589926SEric Dumazet ((nonagle & TCP_NAGLE_CORK) || 1712d4589926SEric Dumazet (!nonagle && tp->packets_out && tcp_minshall_check(tp))); 1713d4589926SEric Dumazet } 1714605ad7f1SEric Dumazet 1715605ad7f1SEric Dumazet /* Return how many segs we'd like on a TSO packet, 1716605ad7f1SEric Dumazet * to send one TSO packet per ms 1717605ad7f1SEric Dumazet */ 17181b3878caSNeal Cardwell u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now, 17191b3878caSNeal Cardwell int min_tso_segs) 1720605ad7f1SEric Dumazet { 1721605ad7f1SEric Dumazet u32 bytes, segs; 1722605ad7f1SEric Dumazet 17233a9b76fdSEric Dumazet bytes = min(sk->sk_pacing_rate >> sk->sk_pacing_shift, 1724605ad7f1SEric Dumazet sk->sk_gso_max_size - 1 - MAX_TCP_HEADER); 1725605ad7f1SEric Dumazet 1726605ad7f1SEric Dumazet /* Goal is to send at least one packet per ms, 1727605ad7f1SEric Dumazet * not one big TSO packet every 100 ms. 1728605ad7f1SEric Dumazet * This preserves ACK clocking and is consistent 1729605ad7f1SEric Dumazet * with tcp_tso_should_defer() heuristic. 1730605ad7f1SEric Dumazet */ 17311b3878caSNeal Cardwell segs = max_t(u32, bytes / mss_now, min_tso_segs); 1732605ad7f1SEric Dumazet 1733605ad7f1SEric Dumazet return min_t(u32, segs, sk->sk_gso_max_segs); 1734605ad7f1SEric Dumazet } 17351b3878caSNeal Cardwell EXPORT_SYMBOL(tcp_tso_autosize); 1736605ad7f1SEric Dumazet 1737ed6e7268SNeal Cardwell /* Return the number of segments we want in the skb we are transmitting. 1738ed6e7268SNeal Cardwell * See if congestion control module wants to decide; otherwise, autosize. 1739ed6e7268SNeal Cardwell */ 1740ed6e7268SNeal Cardwell static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now) 1741ed6e7268SNeal Cardwell { 1742ed6e7268SNeal Cardwell const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 1743ed6e7268SNeal Cardwell u32 tso_segs = ca_ops->tso_segs_goal ? ca_ops->tso_segs_goal(sk) : 0; 1744ed6e7268SNeal Cardwell 17451b3878caSNeal Cardwell return tso_segs ? : 174626e9596eSEric Dumazet tcp_tso_autosize(sk, mss_now, 174726e9596eSEric Dumazet sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs); 1748ed6e7268SNeal Cardwell } 1749ed6e7268SNeal Cardwell 1750d4589926SEric Dumazet /* Returns the portion of skb which can be sent right away */ 1751d4589926SEric Dumazet static unsigned int tcp_mss_split_point(const struct sock *sk, 1752d4589926SEric Dumazet const struct sk_buff *skb, 1753d4589926SEric Dumazet unsigned int mss_now, 1754d4589926SEric Dumazet unsigned int max_segs, 1755d4589926SEric Dumazet int nonagle) 1756c1b4a7e6SDavid S. Miller { 1757cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1758d4589926SEric Dumazet u32 partial, needed, window, max_len; 1759c1b4a7e6SDavid S. Miller 176090840defSIlpo Järvinen window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 17611485348dSBen Hutchings max_len = mss_now * max_segs; 17620e3a4803SIlpo Järvinen 17631485348dSBen Hutchings if (likely(max_len <= window && skb != tcp_write_queue_tail(sk))) 17641485348dSBen Hutchings return max_len; 17650e3a4803SIlpo Järvinen 17665ea3a748SIlpo Järvinen needed = min(skb->len, window); 17675ea3a748SIlpo Järvinen 17681485348dSBen Hutchings if (max_len <= needed) 17691485348dSBen Hutchings return max_len; 17700e3a4803SIlpo Järvinen 1771d4589926SEric Dumazet partial = needed % mss_now; 1772d4589926SEric Dumazet /* If last segment is not a full MSS, check if Nagle rules allow us 1773d4589926SEric Dumazet * to include this last segment in this skb. 1774d4589926SEric Dumazet * Otherwise, we'll split the skb at last MSS boundary 1775d4589926SEric Dumazet */ 1776cc93fc51SPeter Pan(潘卫平) if (tcp_nagle_check(partial != 0, tp, nonagle)) 1777d4589926SEric Dumazet return needed - partial; 1778d4589926SEric Dumazet 1779d4589926SEric Dumazet return needed; 1780c1b4a7e6SDavid S. Miller } 1781c1b4a7e6SDavid S. Miller 1782c1b4a7e6SDavid S. Miller /* Can at least one segment of SKB be sent right now, according to the 1783c1b4a7e6SDavid S. Miller * congestion window rules? If so, return how many segments are allowed. 1784c1b4a7e6SDavid S. Miller */ 1785cf533ea5SEric Dumazet static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, 1786cf533ea5SEric Dumazet const struct sk_buff *skb) 1787c1b4a7e6SDavid S. Miller { 1788d649a7a8SEric Dumazet u32 in_flight, cwnd, halfcwnd; 1789c1b4a7e6SDavid S. Miller 1790c1b4a7e6SDavid S. Miller /* Don't be strict about the congestion window for the final FIN. */ 17914de075e0SEric Dumazet if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && 17924de075e0SEric Dumazet tcp_skb_pcount(skb) == 1) 1793c1b4a7e6SDavid S. Miller return 1; 1794c1b4a7e6SDavid S. Miller 1795c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1796c1b4a7e6SDavid S. Miller cwnd = tp->snd_cwnd; 1797d649a7a8SEric Dumazet if (in_flight >= cwnd) 1798c1b4a7e6SDavid S. Miller return 0; 1799d649a7a8SEric Dumazet 1800d649a7a8SEric Dumazet /* For better scheduling, ensure we have at least 1801d649a7a8SEric Dumazet * 2 GSO packets in flight. 1802d649a7a8SEric Dumazet */ 1803d649a7a8SEric Dumazet halfcwnd = max(cwnd >> 1, 1U); 1804d649a7a8SEric Dumazet return min(halfcwnd, cwnd - in_flight); 1805c1b4a7e6SDavid S. Miller } 1806c1b4a7e6SDavid S. Miller 1807b595076aSUwe Kleine-König /* Initialize TSO state of a skb. 180867edfef7SAndi Kleen * This must be invoked the first time we consider transmitting 1809c1b4a7e6SDavid S. Miller * SKB onto the wire. 1810c1b4a7e6SDavid S. Miller */ 18115bbb432cSEric Dumazet static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now) 1812c1b4a7e6SDavid S. Miller { 1813c1b4a7e6SDavid S. Miller int tso_segs = tcp_skb_pcount(skb); 1814c1b4a7e6SDavid S. Miller 1815f8269a49SIlpo Järvinen if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { 18165bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 1817c1b4a7e6SDavid S. Miller tso_segs = tcp_skb_pcount(skb); 1818c1b4a7e6SDavid S. Miller } 1819c1b4a7e6SDavid S. Miller return tso_segs; 1820c1b4a7e6SDavid S. Miller } 1821c1b4a7e6SDavid S. Miller 1822c1b4a7e6SDavid S. Miller 1823a2a385d6SEric Dumazet /* Return true if the Nagle test allows this packet to be 1824c1b4a7e6SDavid S. Miller * sent now. 1825c1b4a7e6SDavid S. Miller */ 1826a2a385d6SEric Dumazet static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, 1827c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1828c1b4a7e6SDavid S. Miller { 1829c1b4a7e6SDavid S. Miller /* Nagle rule does not apply to frames, which sit in the middle of the 1830c1b4a7e6SDavid S. Miller * write_queue (they have no chances to get new data). 1831c1b4a7e6SDavid S. Miller * 1832c1b4a7e6SDavid S. Miller * This is implemented in the callers, where they modify the 'nonagle' 1833c1b4a7e6SDavid S. Miller * argument based upon the location of SKB in the send queue. 1834c1b4a7e6SDavid S. Miller */ 1835c1b4a7e6SDavid S. Miller if (nonagle & TCP_NAGLE_PUSH) 1836a2a385d6SEric Dumazet return true; 1837c1b4a7e6SDavid S. Miller 18389b44190dSYuchung Cheng /* Don't use the nagle rule for urgent data (or for the final FIN). */ 18399b44190dSYuchung Cheng if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) 1840a2a385d6SEric Dumazet return true; 1841c1b4a7e6SDavid S. Miller 1842cc93fc51SPeter Pan(潘卫平) if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle)) 1843a2a385d6SEric Dumazet return true; 1844c1b4a7e6SDavid S. Miller 1845a2a385d6SEric Dumazet return false; 1846c1b4a7e6SDavid S. Miller } 1847c1b4a7e6SDavid S. Miller 1848c1b4a7e6SDavid S. Miller /* Does at least the first segment of SKB fit into the send window? */ 1849a2a385d6SEric Dumazet static bool tcp_snd_wnd_test(const struct tcp_sock *tp, 1850a2a385d6SEric Dumazet const struct sk_buff *skb, 1851056834d9SIlpo Järvinen unsigned int cur_mss) 1852c1b4a7e6SDavid S. Miller { 1853c1b4a7e6SDavid S. Miller u32 end_seq = TCP_SKB_CB(skb)->end_seq; 1854c1b4a7e6SDavid S. Miller 1855c1b4a7e6SDavid S. Miller if (skb->len > cur_mss) 1856c1b4a7e6SDavid S. Miller end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 1857c1b4a7e6SDavid S. Miller 185890840defSIlpo Järvinen return !after(end_seq, tcp_wnd_end(tp)); 1859c1b4a7e6SDavid S. Miller } 1860c1b4a7e6SDavid S. Miller 1861c1b4a7e6SDavid S. Miller /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 1862c1b4a7e6SDavid S. Miller * which is put after SKB on the list. It is very much like 1863c1b4a7e6SDavid S. Miller * tcp_fragment() except that it may make several kinds of assumptions 1864c1b4a7e6SDavid S. Miller * in order to speed up the splitting operation. In particular, we 1865c1b4a7e6SDavid S. Miller * know that all the data is in scatter-gather pages, and that the 1866c1b4a7e6SDavid S. Miller * packet has never been sent out before (and thus is not cloned). 1867c1b4a7e6SDavid S. Miller */ 186875c119afSEric Dumazet static int tso_fragment(struct sock *sk, enum tcp_queue tcp_queue, 186975c119afSEric Dumazet struct sk_buff *skb, unsigned int len, 1870c4ead4c5SEric Dumazet unsigned int mss_now, gfp_t gfp) 1871c1b4a7e6SDavid S. Miller { 1872c1b4a7e6SDavid S. Miller struct sk_buff *buff; 1873c1b4a7e6SDavid S. Miller int nlen = skb->len - len; 18749ce01461SIlpo Järvinen u8 flags; 1875c1b4a7e6SDavid S. Miller 1876c1b4a7e6SDavid S. Miller /* All of a TSO frame must be composed of paged data. */ 1877c8ac3774SHerbert Xu if (skb->len != skb->data_len) 187875c119afSEric Dumazet return tcp_fragment(sk, tcp_queue, skb, len, mss_now, gfp); 1879c1b4a7e6SDavid S. Miller 1880eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, gfp, true); 188151456b29SIan Morris if (unlikely(!buff)) 1882c1b4a7e6SDavid S. Miller return -ENOMEM; 1883c1b4a7e6SDavid S. Miller 18843ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 18853ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1886b60b49eaSHerbert Xu buff->truesize += nlen; 1887c1b4a7e6SDavid S. Miller skb->truesize -= nlen; 1888c1b4a7e6SDavid S. Miller 1889c1b4a7e6SDavid S. Miller /* Correct the sequence numbers. */ 1890c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 1891c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 1892c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 1893c1b4a7e6SDavid S. Miller 1894c1b4a7e6SDavid S. Miller /* PSH and FIN should only be set in the second packet. */ 18954de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 18964de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 18974de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1898c1b4a7e6SDavid S. Miller 1899c1b4a7e6SDavid S. Miller /* This packet was never sent out yet, so no SACK bits. */ 1900c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->sacked = 0; 1901c1b4a7e6SDavid S. Miller 1902a166140eSMartin KaFai Lau tcp_skb_fragment_eor(skb, buff); 1903a166140eSMartin KaFai Lau 190484fa7933SPatrick McHardy buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; 1905c1b4a7e6SDavid S. Miller skb_split(skb, buff, len); 1906490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 1907c1b4a7e6SDavid S. Miller 1908c1b4a7e6SDavid S. Miller /* Fix up tso_factor for both original and new SKB. */ 19095bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 19105bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 1911c1b4a7e6SDavid S. Miller 1912c1b4a7e6SDavid S. Miller /* Link BUFF into the send queue. */ 1913f4a775d1SEric Dumazet __skb_header_release(buff); 191475c119afSEric Dumazet tcp_insert_write_queue_after(skb, buff, sk, tcp_queue); 1915c1b4a7e6SDavid S. Miller 1916c1b4a7e6SDavid S. Miller return 0; 1917c1b4a7e6SDavid S. Miller } 1918c1b4a7e6SDavid S. Miller 1919c1b4a7e6SDavid S. Miller /* Try to defer sending, if possible, in order to minimize the amount 1920c1b4a7e6SDavid S. Miller * of TSO splitting we do. View it as a kind of TSO Nagle test. 1921c1b4a7e6SDavid S. Miller * 1922c1b4a7e6SDavid S. Miller * This algorithm is from John Heffner. 1923c1b4a7e6SDavid S. Miller */ 1924ca8a2263SNeal Cardwell static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb, 1925605ad7f1SEric Dumazet bool *is_cwnd_limited, u32 max_segs) 1926c1b4a7e6SDavid S. Miller { 19276687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 192850c8339eSEric Dumazet u32 age, send_win, cong_win, limit, in_flight; 192950c8339eSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 193050c8339eSEric Dumazet struct sk_buff *head; 1931ad9f4f50SEric Dumazet int win_divisor; 1932c1b4a7e6SDavid S. Miller 19334de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 1934ae8064acSJohn Heffner goto send_now; 1935c1b4a7e6SDavid S. Miller 193699d7662aSEric Dumazet if (icsk->icsk_ca_state >= TCP_CA_Recovery) 1937ae8064acSJohn Heffner goto send_now; 1938ae8064acSJohn Heffner 19395f852eb5SEric Dumazet /* Avoid bursty behavior by allowing defer 19405f852eb5SEric Dumazet * only if the last write was recent. 19415f852eb5SEric Dumazet */ 1942d635fbe2SEric Dumazet if ((s32)(tcp_jiffies32 - tp->lsndtime) > 0) 1943ae8064acSJohn Heffner goto send_now; 1944908a75c1SDavid S. Miller 1945c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1946c1b4a7e6SDavid S. Miller 1947c8c9aeb5SStefano Brivio BUG_ON(tcp_skb_pcount(skb) <= 1); 1948c8c9aeb5SStefano Brivio BUG_ON(tp->snd_cwnd <= in_flight); 1949c1b4a7e6SDavid S. Miller 195090840defSIlpo Järvinen send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1951c1b4a7e6SDavid S. Miller 1952c1b4a7e6SDavid S. Miller /* From in_flight test above, we know that cwnd > in_flight. */ 1953c1b4a7e6SDavid S. Miller cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; 1954c1b4a7e6SDavid S. Miller 1955c1b4a7e6SDavid S. Miller limit = min(send_win, cong_win); 1956c1b4a7e6SDavid S. Miller 1957ba244fe9SDavid S. Miller /* If a full-sized TSO skb can be sent, do it. */ 1958605ad7f1SEric Dumazet if (limit >= max_segs * tp->mss_cache) 1959ae8064acSJohn Heffner goto send_now; 1960ba244fe9SDavid S. Miller 196162ad2761SIlpo Järvinen /* Middle in queue won't get any more data, full sendable already? */ 196262ad2761SIlpo Järvinen if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) 196362ad2761SIlpo Järvinen goto send_now; 196462ad2761SIlpo Järvinen 19655bbcc0f5SLinus Torvalds win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor); 1966ad9f4f50SEric Dumazet if (win_divisor) { 1967c1b4a7e6SDavid S. Miller u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); 1968c1b4a7e6SDavid S. Miller 1969c1b4a7e6SDavid S. Miller /* If at least some fraction of a window is available, 1970c1b4a7e6SDavid S. Miller * just use it. 1971c1b4a7e6SDavid S. Miller */ 1972ad9f4f50SEric Dumazet chunk /= win_divisor; 1973c1b4a7e6SDavid S. Miller if (limit >= chunk) 1974ae8064acSJohn Heffner goto send_now; 1975c1b4a7e6SDavid S. Miller } else { 1976c1b4a7e6SDavid S. Miller /* Different approach, try not to defer past a single 1977c1b4a7e6SDavid S. Miller * ACK. Receiver should ACK every other full sized 1978c1b4a7e6SDavid S. Miller * frame, so if we have space for more than 3 frames 1979c1b4a7e6SDavid S. Miller * then send now. 1980c1b4a7e6SDavid S. Miller */ 19816b5a5c0dSNeal Cardwell if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) 1982ae8064acSJohn Heffner goto send_now; 1983c1b4a7e6SDavid S. Miller } 1984c1b4a7e6SDavid S. Miller 198575c119afSEric Dumazet /* TODO : use tsorted_sent_queue ? */ 198675c119afSEric Dumazet head = tcp_rtx_queue_head(sk); 198775c119afSEric Dumazet if (!head) 198875c119afSEric Dumazet goto send_now; 19899a568de4SEric Dumazet age = tcp_stamp_us_delta(tp->tcp_mstamp, head->skb_mstamp); 199050c8339eSEric Dumazet /* If next ACK is likely to come too late (half srtt), do not defer */ 199150c8339eSEric Dumazet if (age < (tp->srtt_us >> 4)) 199250c8339eSEric Dumazet goto send_now; 199350c8339eSEric Dumazet 19945f852eb5SEric Dumazet /* Ok, it looks like it is advisable to defer. */ 1995ae8064acSJohn Heffner 1996d2e1339fSBendik Rønning Opstad if (cong_win < send_win && cong_win <= skb->len) 1997ca8a2263SNeal Cardwell *is_cwnd_limited = true; 1998ca8a2263SNeal Cardwell 1999a2a385d6SEric Dumazet return true; 2000ae8064acSJohn Heffner 2001ae8064acSJohn Heffner send_now: 2002a2a385d6SEric Dumazet return false; 2003c1b4a7e6SDavid S. Miller } 2004c1b4a7e6SDavid S. Miller 200505cbc0dbSFan Du static inline void tcp_mtu_check_reprobe(struct sock *sk) 200605cbc0dbSFan Du { 200705cbc0dbSFan Du struct inet_connection_sock *icsk = inet_csk(sk); 200805cbc0dbSFan Du struct tcp_sock *tp = tcp_sk(sk); 200905cbc0dbSFan Du struct net *net = sock_net(sk); 201005cbc0dbSFan Du u32 interval; 201105cbc0dbSFan Du s32 delta; 201205cbc0dbSFan Du 201305cbc0dbSFan Du interval = net->ipv4.sysctl_tcp_probe_interval; 2014c74df29aSEric Dumazet delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp; 201505cbc0dbSFan Du if (unlikely(delta >= interval * HZ)) { 201605cbc0dbSFan Du int mss = tcp_current_mss(sk); 201705cbc0dbSFan Du 201805cbc0dbSFan Du /* Update current search range */ 201905cbc0dbSFan Du icsk->icsk_mtup.probe_size = 0; 202005cbc0dbSFan Du icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + 202105cbc0dbSFan Du sizeof(struct tcphdr) + 202205cbc0dbSFan Du icsk->icsk_af_ops->net_header_len; 202305cbc0dbSFan Du icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss); 202405cbc0dbSFan Du 202505cbc0dbSFan Du /* Update probe time stamp */ 2026c74df29aSEric Dumazet icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; 202705cbc0dbSFan Du } 202805cbc0dbSFan Du } 202905cbc0dbSFan Du 2030*808cf9e3SIlya Lesokhin static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len) 2031*808cf9e3SIlya Lesokhin { 2032*808cf9e3SIlya Lesokhin struct sk_buff *skb, *next; 2033*808cf9e3SIlya Lesokhin 2034*808cf9e3SIlya Lesokhin skb = tcp_send_head(sk); 2035*808cf9e3SIlya Lesokhin tcp_for_write_queue_from_safe(skb, next, sk) { 2036*808cf9e3SIlya Lesokhin if (len <= skb->len) 2037*808cf9e3SIlya Lesokhin break; 2038*808cf9e3SIlya Lesokhin 2039*808cf9e3SIlya Lesokhin if (unlikely(TCP_SKB_CB(skb)->eor)) 2040*808cf9e3SIlya Lesokhin return false; 2041*808cf9e3SIlya Lesokhin 2042*808cf9e3SIlya Lesokhin len -= skb->len; 2043*808cf9e3SIlya Lesokhin } 2044*808cf9e3SIlya Lesokhin 2045*808cf9e3SIlya Lesokhin return true; 2046*808cf9e3SIlya Lesokhin } 2047*808cf9e3SIlya Lesokhin 20485d424d5aSJohn Heffner /* Create a new MTU probe if we are ready. 204967edfef7SAndi Kleen * MTU probe is regularly attempting to increase the path MTU by 205067edfef7SAndi Kleen * deliberately sending larger packets. This discovers routing 205167edfef7SAndi Kleen * changes resulting in larger path MTUs. 205267edfef7SAndi Kleen * 20535d424d5aSJohn Heffner * Returns 0 if we should wait to probe (no cwnd available), 20545d424d5aSJohn Heffner * 1 if a probe was sent, 2055056834d9SIlpo Järvinen * -1 otherwise 2056056834d9SIlpo Järvinen */ 20575d424d5aSJohn Heffner static int tcp_mtu_probe(struct sock *sk) 20585d424d5aSJohn Heffner { 20595d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 206012a59abcSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 20615d424d5aSJohn Heffner struct sk_buff *skb, *nskb, *next; 20626b58e0a5SFan Du struct net *net = sock_net(sk); 20635d424d5aSJohn Heffner int probe_size; 206491cc17c0SIlpo Järvinen int size_needed; 206512a59abcSEric Dumazet int copy, len; 20665d424d5aSJohn Heffner int mss_now; 20676b58e0a5SFan Du int interval; 20685d424d5aSJohn Heffner 20695d424d5aSJohn Heffner /* Not currently probing/verifying, 20705d424d5aSJohn Heffner * not in recovery, 20715d424d5aSJohn Heffner * have enough cwnd, and 207212a59abcSEric Dumazet * not SACKing (the variable headers throw things off) 207312a59abcSEric Dumazet */ 207412a59abcSEric Dumazet if (likely(!icsk->icsk_mtup.enabled || 20755d424d5aSJohn Heffner icsk->icsk_mtup.probe_size || 20765d424d5aSJohn Heffner inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 20775d424d5aSJohn Heffner tp->snd_cwnd < 11 || 207812a59abcSEric Dumazet tp->rx_opt.num_sacks || tp->rx_opt.dsack)) 20795d424d5aSJohn Heffner return -1; 20805d424d5aSJohn Heffner 20816b58e0a5SFan Du /* Use binary search for probe_size between tcp_mss_base, 20826b58e0a5SFan Du * and current mss_clamp. if (search_high - search_low) 20836b58e0a5SFan Du * smaller than a threshold, backoff from probing. 20846b58e0a5SFan Du */ 20850c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 20866b58e0a5SFan Du probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high + 20876b58e0a5SFan Du icsk->icsk_mtup.search_low) >> 1); 208891cc17c0SIlpo Järvinen size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; 20896b58e0a5SFan Du interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low; 209005cbc0dbSFan Du /* When misfortune happens, we are reprobing actively, 209105cbc0dbSFan Du * and then reprobe timer has expired. We stick with current 209205cbc0dbSFan Du * probing process by not resetting search range to its orignal. 209305cbc0dbSFan Du */ 20946b58e0a5SFan Du if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) || 209505cbc0dbSFan Du interval < net->ipv4.sysctl_tcp_probe_threshold) { 209605cbc0dbSFan Du /* Check whether enough time has elaplased for 209705cbc0dbSFan Du * another round of probing. 209805cbc0dbSFan Du */ 209905cbc0dbSFan Du tcp_mtu_check_reprobe(sk); 21005d424d5aSJohn Heffner return -1; 21015d424d5aSJohn Heffner } 21025d424d5aSJohn Heffner 21035d424d5aSJohn Heffner /* Have enough data in the send queue to probe? */ 21047f9c33e5SIlpo Järvinen if (tp->write_seq - tp->snd_nxt < size_needed) 21055d424d5aSJohn Heffner return -1; 21065d424d5aSJohn Heffner 210791cc17c0SIlpo Järvinen if (tp->snd_wnd < size_needed) 21085d424d5aSJohn Heffner return -1; 210990840defSIlpo Järvinen if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) 21105d424d5aSJohn Heffner return 0; 21115d424d5aSJohn Heffner 2112d67c58e9SIlpo Järvinen /* Do we need to wait to drain cwnd? With none in flight, don't stall */ 2113d67c58e9SIlpo Järvinen if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { 2114d67c58e9SIlpo Järvinen if (!tcp_packets_in_flight(tp)) 21155d424d5aSJohn Heffner return -1; 21165d424d5aSJohn Heffner else 21175d424d5aSJohn Heffner return 0; 21185d424d5aSJohn Heffner } 21195d424d5aSJohn Heffner 2120*808cf9e3SIlya Lesokhin if (!tcp_can_coalesce_send_queue_head(sk, probe_size)) 2121*808cf9e3SIlya Lesokhin return -1; 2122*808cf9e3SIlya Lesokhin 21235d424d5aSJohn Heffner /* We're allowed to probe. Build it now. */ 2124eb934478SEric Dumazet nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false); 212551456b29SIan Morris if (!nskb) 21265d424d5aSJohn Heffner return -1; 21273ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 21283ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 21295d424d5aSJohn Heffner 2130fe067e8aSDavid S. Miller skb = tcp_send_head(sk); 21315d424d5aSJohn Heffner 21325d424d5aSJohn Heffner TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 21335d424d5aSJohn Heffner TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 21344de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; 21355d424d5aSJohn Heffner TCP_SKB_CB(nskb)->sacked = 0; 21365d424d5aSJohn Heffner nskb->csum = 0; 213784fa7933SPatrick McHardy nskb->ip_summed = skb->ip_summed; 21385d424d5aSJohn Heffner 213950c4817eSIlpo Järvinen tcp_insert_write_queue_before(nskb, skb, sk); 21402b7cda9cSEric Dumazet tcp_highest_sack_replace(sk, skb, nskb); 214150c4817eSIlpo Järvinen 21425d424d5aSJohn Heffner len = 0; 2143234b6860SIlpo Järvinen tcp_for_write_queue_from_safe(skb, next, sk) { 21445d424d5aSJohn Heffner copy = min_t(int, skb->len, probe_size - len); 21452fe664f1SDouglas Caetano dos Santos if (nskb->ip_summed) { 21465d424d5aSJohn Heffner skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); 21472fe664f1SDouglas Caetano dos Santos } else { 21482fe664f1SDouglas Caetano dos Santos __wsum csum = skb_copy_and_csum_bits(skb, 0, 2149056834d9SIlpo Järvinen skb_put(nskb, copy), 21502fe664f1SDouglas Caetano dos Santos copy, 0); 21512fe664f1SDouglas Caetano dos Santos nskb->csum = csum_block_add(nskb->csum, csum, len); 21522fe664f1SDouglas Caetano dos Santos } 21535d424d5aSJohn Heffner 21545d424d5aSJohn Heffner if (skb->len <= copy) { 21555d424d5aSJohn Heffner /* We've eaten all the data from this skb. 21565d424d5aSJohn Heffner * Throw it away. */ 21574de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 2158*808cf9e3SIlya Lesokhin /* If this is the last SKB we copy and eor is set 2159*808cf9e3SIlya Lesokhin * we need to propagate it to the new skb. 2160*808cf9e3SIlya Lesokhin */ 2161*808cf9e3SIlya Lesokhin TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor; 2162fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 21633ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 21645d424d5aSJohn Heffner } else { 21654de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & 2166a3433f35SChangli Gao ~(TCPHDR_FIN|TCPHDR_PSH); 21675d424d5aSJohn Heffner if (!skb_shinfo(skb)->nr_frags) { 21685d424d5aSJohn Heffner skb_pull(skb, copy); 216984fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 2170056834d9SIlpo Järvinen skb->csum = csum_partial(skb->data, 2171056834d9SIlpo Järvinen skb->len, 0); 21725d424d5aSJohn Heffner } else { 21735d424d5aSJohn Heffner __pskb_trim_head(skb, copy); 21745bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 21755d424d5aSJohn Heffner } 21765d424d5aSJohn Heffner TCP_SKB_CB(skb)->seq += copy; 21775d424d5aSJohn Heffner } 21785d424d5aSJohn Heffner 21795d424d5aSJohn Heffner len += copy; 2180234b6860SIlpo Järvinen 2181234b6860SIlpo Järvinen if (len >= probe_size) 2182234b6860SIlpo Järvinen break; 21835d424d5aSJohn Heffner } 21845bbb432cSEric Dumazet tcp_init_tso_segs(nskb, nskb->len); 21855d424d5aSJohn Heffner 21865d424d5aSJohn Heffner /* We're ready to send. If this fails, the probe will 21877faee5c0SEric Dumazet * be resegmented into mss-sized pieces by tcp_write_xmit(). 21887faee5c0SEric Dumazet */ 21895d424d5aSJohn Heffner if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 21905d424d5aSJohn Heffner /* Decrement cwnd here because we are sending 21915d424d5aSJohn Heffner * effectively two packets. */ 21925d424d5aSJohn Heffner tp->snd_cwnd--; 219366f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, nskb); 21945d424d5aSJohn Heffner 21955d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 21960e7b1368SJohn Heffner tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 21970e7b1368SJohn Heffner tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 21985d424d5aSJohn Heffner 21995d424d5aSJohn Heffner return 1; 22005d424d5aSJohn Heffner } 22015d424d5aSJohn Heffner 22025d424d5aSJohn Heffner return -1; 22035d424d5aSJohn Heffner } 22045d424d5aSJohn Heffner 2205218af599SEric Dumazet static bool tcp_pacing_check(const struct sock *sk) 2206218af599SEric Dumazet { 2207218af599SEric Dumazet return tcp_needs_internal_pacing(sk) && 2208218af599SEric Dumazet hrtimer_active(&tcp_sk(sk)->pacing_timer); 2209218af599SEric Dumazet } 2210218af599SEric Dumazet 2211f9616c35SEric Dumazet /* TCP Small Queues : 2212f9616c35SEric Dumazet * Control number of packets in qdisc/devices to two packets / or ~1 ms. 2213f9616c35SEric Dumazet * (These limits are doubled for retransmits) 2214f9616c35SEric Dumazet * This allows for : 2215f9616c35SEric Dumazet * - better RTT estimation and ACK scheduling 2216f9616c35SEric Dumazet * - faster recovery 2217f9616c35SEric Dumazet * - high rates 2218f9616c35SEric Dumazet * Alas, some drivers / subsystems require a fair amount 2219f9616c35SEric Dumazet * of queued bytes to ensure line rate. 2220f9616c35SEric Dumazet * One example is wifi aggregation (802.11 AMPDU) 2221f9616c35SEric Dumazet */ 2222f9616c35SEric Dumazet static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb, 2223f9616c35SEric Dumazet unsigned int factor) 2224f9616c35SEric Dumazet { 2225f9616c35SEric Dumazet unsigned int limit; 2226f9616c35SEric Dumazet 22273a9b76fdSEric Dumazet limit = max(2 * skb->truesize, sk->sk_pacing_rate >> sk->sk_pacing_shift); 22289184d8bbSEric Dumazet limit = min_t(u32, limit, 22299184d8bbSEric Dumazet sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes); 2230f9616c35SEric Dumazet limit <<= factor; 2231f9616c35SEric Dumazet 223214afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) > limit) { 223375c119afSEric Dumazet /* Always send skb if rtx queue is empty. 223475eefc6cSEric Dumazet * No need to wait for TX completion to call us back, 223575eefc6cSEric Dumazet * after softirq/tasklet schedule. 223675eefc6cSEric Dumazet * This helps when TX completions are delayed too much. 223775eefc6cSEric Dumazet */ 223875c119afSEric Dumazet if (tcp_rtx_queue_empty(sk)) 223975eefc6cSEric Dumazet return false; 224075eefc6cSEric Dumazet 22417aa5470cSEric Dumazet set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags); 2242f9616c35SEric Dumazet /* It is possible TX completion already happened 2243f9616c35SEric Dumazet * before we set TSQ_THROTTLED, so we must 2244f9616c35SEric Dumazet * test again the condition. 2245f9616c35SEric Dumazet */ 2246f9616c35SEric Dumazet smp_mb__after_atomic(); 224714afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) > limit) 2248f9616c35SEric Dumazet return true; 2249f9616c35SEric Dumazet } 2250f9616c35SEric Dumazet return false; 2251f9616c35SEric Dumazet } 2252f9616c35SEric Dumazet 225305b055e8SFrancis Yan static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new) 225405b055e8SFrancis Yan { 2255628174ccSEric Dumazet const u32 now = tcp_jiffies32; 2256efe967cdSArnd Bergmann enum tcp_chrono old = tp->chrono_type; 225705b055e8SFrancis Yan 2258efe967cdSArnd Bergmann if (old > TCP_CHRONO_UNSPEC) 2259efe967cdSArnd Bergmann tp->chrono_stat[old - 1] += now - tp->chrono_start; 226005b055e8SFrancis Yan tp->chrono_start = now; 226105b055e8SFrancis Yan tp->chrono_type = new; 226205b055e8SFrancis Yan } 226305b055e8SFrancis Yan 226405b055e8SFrancis Yan void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type) 226505b055e8SFrancis Yan { 226605b055e8SFrancis Yan struct tcp_sock *tp = tcp_sk(sk); 226705b055e8SFrancis Yan 226805b055e8SFrancis Yan /* If there are multiple conditions worthy of tracking in a 22690f87230dSFrancis Yan * chronograph then the highest priority enum takes precedence 22700f87230dSFrancis Yan * over the other conditions. So that if something "more interesting" 227105b055e8SFrancis Yan * starts happening, stop the previous chrono and start a new one. 227205b055e8SFrancis Yan */ 227305b055e8SFrancis Yan if (type > tp->chrono_type) 227405b055e8SFrancis Yan tcp_chrono_set(tp, type); 227505b055e8SFrancis Yan } 227605b055e8SFrancis Yan 227705b055e8SFrancis Yan void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type) 227805b055e8SFrancis Yan { 227905b055e8SFrancis Yan struct tcp_sock *tp = tcp_sk(sk); 228005b055e8SFrancis Yan 22810f87230dSFrancis Yan 22820f87230dSFrancis Yan /* There are multiple conditions worthy of tracking in a 22830f87230dSFrancis Yan * chronograph, so that the highest priority enum takes 22840f87230dSFrancis Yan * precedence over the other conditions (see tcp_chrono_start). 22850f87230dSFrancis Yan * If a condition stops, we only stop chrono tracking if 22860f87230dSFrancis Yan * it's the "most interesting" or current chrono we are 22870f87230dSFrancis Yan * tracking and starts busy chrono if we have pending data. 22880f87230dSFrancis Yan */ 228975c119afSEric Dumazet if (tcp_rtx_and_write_queues_empty(sk)) 229005b055e8SFrancis Yan tcp_chrono_set(tp, TCP_CHRONO_UNSPEC); 22910f87230dSFrancis Yan else if (type == tp->chrono_type) 22920f87230dSFrancis Yan tcp_chrono_set(tp, TCP_CHRONO_BUSY); 229305b055e8SFrancis Yan } 229405b055e8SFrancis Yan 22951da177e4SLinus Torvalds /* This routine writes packets to the network. It advances the 22961da177e4SLinus Torvalds * send_head. This happens as incoming acks open up the remote 22971da177e4SLinus Torvalds * window for us. 22981da177e4SLinus Torvalds * 2299f8269a49SIlpo Järvinen * LARGESEND note: !tcp_urg_mode is overkill, only frames between 2300f8269a49SIlpo Järvinen * snd_up-64k-mss .. snd_up cannot be large. However, taking into 2301f8269a49SIlpo Järvinen * account rare use of URG, this is not a big flaw. 2302f8269a49SIlpo Järvinen * 23036ba8a3b1SNandita Dukkipati * Send at most one packet when push_one > 0. Temporarily ignore 23046ba8a3b1SNandita Dukkipati * cwnd limit to force at most one packet out when push_one == 2. 23056ba8a3b1SNandita Dukkipati 2306a2a385d6SEric Dumazet * Returns true, if no segments are in flight and we have queued segments, 2307a2a385d6SEric Dumazet * but cannot send anything now because of SWS or another problem. 23081da177e4SLinus Torvalds */ 2309a2a385d6SEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 2310d5dd9175SIlpo Järvinen int push_one, gfp_t gfp) 23111da177e4SLinus Torvalds { 23121da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 231392df7b51SDavid S. Miller struct sk_buff *skb; 2314c1b4a7e6SDavid S. Miller unsigned int tso_segs, sent_pkts; 2315c1b4a7e6SDavid S. Miller int cwnd_quota; 23165d424d5aSJohn Heffner int result; 23175615f886SFrancis Yan bool is_cwnd_limited = false, is_rwnd_limited = false; 2318605ad7f1SEric Dumazet u32 max_segs; 23191da177e4SLinus Torvalds 2320c1b4a7e6SDavid S. Miller sent_pkts = 0; 23215d424d5aSJohn Heffner 2322ee1836aeSEric Dumazet tcp_mstamp_refresh(tp); 2323d5dd9175SIlpo Järvinen if (!push_one) { 23245d424d5aSJohn Heffner /* Do MTU probing. */ 2325d5dd9175SIlpo Järvinen result = tcp_mtu_probe(sk); 2326d5dd9175SIlpo Järvinen if (!result) { 2327a2a385d6SEric Dumazet return false; 23285d424d5aSJohn Heffner } else if (result > 0) { 23295d424d5aSJohn Heffner sent_pkts = 1; 23305d424d5aSJohn Heffner } 2331d5dd9175SIlpo Järvinen } 23325d424d5aSJohn Heffner 2333ed6e7268SNeal Cardwell max_segs = tcp_tso_segs(sk, mss_now); 2334fe067e8aSDavid S. Miller while ((skb = tcp_send_head(sk))) { 2335c8ac3774SHerbert Xu unsigned int limit; 2336c8ac3774SHerbert Xu 2337218af599SEric Dumazet if (tcp_pacing_check(sk)) 2338218af599SEric Dumazet break; 2339218af599SEric Dumazet 23405bbb432cSEric Dumazet tso_segs = tcp_init_tso_segs(skb, mss_now); 2341c1b4a7e6SDavid S. Miller BUG_ON(!tso_segs); 2342c1b4a7e6SDavid S. Miller 23439d186cacSAndrey Vagin if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) { 23447faee5c0SEric Dumazet /* "skb_mstamp" is used as a start point for the retransmit timer */ 2345e2080072SEric Dumazet tcp_update_skb_after_send(tp, skb); 2346ec342325SAndrew Vagin goto repair; /* Skip network transmission */ 23479d186cacSAndrey Vagin } 2348ec342325SAndrew Vagin 2349b68e9f85SHerbert Xu cwnd_quota = tcp_cwnd_test(tp, skb); 23506ba8a3b1SNandita Dukkipati if (!cwnd_quota) { 23516ba8a3b1SNandita Dukkipati if (push_one == 2) 23526ba8a3b1SNandita Dukkipati /* Force out a loss probe pkt. */ 23536ba8a3b1SNandita Dukkipati cwnd_quota = 1; 23546ba8a3b1SNandita Dukkipati else 2355b68e9f85SHerbert Xu break; 23566ba8a3b1SNandita Dukkipati } 2357b68e9f85SHerbert Xu 23585615f886SFrancis Yan if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) { 23595615f886SFrancis Yan is_rwnd_limited = true; 2360b68e9f85SHerbert Xu break; 23615615f886SFrancis Yan } 2362b68e9f85SHerbert Xu 2363d6a4e26aSEric Dumazet if (tso_segs == 1) { 2364aa93466bSDavid S. Miller if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 2365aa93466bSDavid S. Miller (tcp_skb_is_last(sk, skb) ? 2366aa93466bSDavid S. Miller nonagle : TCP_NAGLE_PUSH)))) 2367aa93466bSDavid S. Miller break; 2368c1b4a7e6SDavid S. Miller } else { 2369ca8a2263SNeal Cardwell if (!push_one && 2370605ad7f1SEric Dumazet tcp_tso_should_defer(sk, skb, &is_cwnd_limited, 2371605ad7f1SEric Dumazet max_segs)) 2372aa93466bSDavid S. Miller break; 2373c1b4a7e6SDavid S. Miller } 2374aa93466bSDavid S. Miller 2375605ad7f1SEric Dumazet limit = mss_now; 2376d6a4e26aSEric Dumazet if (tso_segs > 1 && !tcp_urg_mode(tp)) 2377605ad7f1SEric Dumazet limit = tcp_mss_split_point(sk, skb, mss_now, 2378605ad7f1SEric Dumazet min_t(unsigned int, 2379605ad7f1SEric Dumazet cwnd_quota, 2380605ad7f1SEric Dumazet max_segs), 2381605ad7f1SEric Dumazet nonagle); 2382605ad7f1SEric Dumazet 2383605ad7f1SEric Dumazet if (skb->len > limit && 238475c119afSEric Dumazet unlikely(tso_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE, 238575c119afSEric Dumazet skb, limit, mss_now, gfp))) 2386605ad7f1SEric Dumazet break; 2387605ad7f1SEric Dumazet 23887aa5470cSEric Dumazet if (test_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags)) 23897aa5470cSEric Dumazet clear_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags); 2390f9616c35SEric Dumazet if (tcp_small_queue_check(sk, skb, 0)) 239146d3ceabSEric Dumazet break; 2392c9eeec26SEric Dumazet 2393d5dd9175SIlpo Järvinen if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) 23941da177e4SLinus Torvalds break; 23951da177e4SLinus Torvalds 2396ec342325SAndrew Vagin repair: 23971da177e4SLinus Torvalds /* Advance the send_head. This one is sent out. 23981da177e4SLinus Torvalds * This call will increment packets_out. 23991da177e4SLinus Torvalds */ 240066f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 24011da177e4SLinus Torvalds 24021da177e4SLinus Torvalds tcp_minshall_update(tp, mss_now, skb); 2403a262f0cdSNandita Dukkipati sent_pkts += tcp_skb_pcount(skb); 2404d5dd9175SIlpo Järvinen 2405d5dd9175SIlpo Järvinen if (push_one) 2406d5dd9175SIlpo Järvinen break; 24071da177e4SLinus Torvalds } 24081da177e4SLinus Torvalds 24095615f886SFrancis Yan if (is_rwnd_limited) 24105615f886SFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED); 24115615f886SFrancis Yan else 24125615f886SFrancis Yan tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED); 24135615f886SFrancis Yan 2414aa93466bSDavid S. Miller if (likely(sent_pkts)) { 2415684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 2416684bad11SYuchung Cheng tp->prr_out += sent_pkts; 24176ba8a3b1SNandita Dukkipati 24186ba8a3b1SNandita Dukkipati /* Send one loss probe per tail loss episode. */ 24196ba8a3b1SNandita Dukkipati if (push_one != 2) 2420ed66dfafSNeal Cardwell tcp_schedule_loss_probe(sk, false); 2421d2e1339fSBendik Rønning Opstad is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd); 2422ca8a2263SNeal Cardwell tcp_cwnd_validate(sk, is_cwnd_limited); 2423a2a385d6SEric Dumazet return false; 24241da177e4SLinus Torvalds } 242575c119afSEric Dumazet return !tp->packets_out && !tcp_write_queue_empty(sk); 24266ba8a3b1SNandita Dukkipati } 24276ba8a3b1SNandita Dukkipati 2428ed66dfafSNeal Cardwell bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto) 24296ba8a3b1SNandita Dukkipati { 24306ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 24316ba8a3b1SNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 2432a2815817SNeal Cardwell u32 timeout, rto_delta_us; 24332ae21cf5SEric Dumazet int early_retrans; 24346ba8a3b1SNandita Dukkipati 24356ba8a3b1SNandita Dukkipati /* Don't do any loss probe on a Fast Open connection before 3WHS 24366ba8a3b1SNandita Dukkipati * finishes. 24376ba8a3b1SNandita Dukkipati */ 2438f9b99582SYuchung Cheng if (tp->fastopen_rsk) 24396ba8a3b1SNandita Dukkipati return false; 24406ba8a3b1SNandita Dukkipati 24412ae21cf5SEric Dumazet early_retrans = sock_net(sk)->ipv4.sysctl_tcp_early_retrans; 24426ba8a3b1SNandita Dukkipati /* Schedule a loss probe in 2*RTT for SACK capable connections 2443b4f70c3dSNeal Cardwell * not in loss recovery, that are either limited by cwnd or application. 24446ba8a3b1SNandita Dukkipati */ 24452ae21cf5SEric Dumazet if ((early_retrans != 3 && early_retrans != 4) || 2446bec41a11SYuchung Cheng !tp->packets_out || !tcp_is_sack(tp) || 2447b4f70c3dSNeal Cardwell (icsk->icsk_ca_state != TCP_CA_Open && 2448b4f70c3dSNeal Cardwell icsk->icsk_ca_state != TCP_CA_CWR)) 24496ba8a3b1SNandita Dukkipati return false; 24506ba8a3b1SNandita Dukkipati 2451bb4d991aSYuchung Cheng /* Probe timeout is 2*rtt. Add minimum RTO to account 2452f9b99582SYuchung Cheng * for delayed ack when there's one outstanding packet. If no RTT 2453f9b99582SYuchung Cheng * sample is available then probe after TCP_TIMEOUT_INIT. 24546ba8a3b1SNandita Dukkipati */ 2455bb4d991aSYuchung Cheng if (tp->srtt_us) { 2456bb4d991aSYuchung Cheng timeout = usecs_to_jiffies(tp->srtt_us >> 2); 24576ba8a3b1SNandita Dukkipati if (tp->packets_out == 1) 2458bb4d991aSYuchung Cheng timeout += TCP_RTO_MIN; 2459bb4d991aSYuchung Cheng else 2460bb4d991aSYuchung Cheng timeout += TCP_TIMEOUT_MIN; 2461bb4d991aSYuchung Cheng } else { 2462bb4d991aSYuchung Cheng timeout = TCP_TIMEOUT_INIT; 2463bb4d991aSYuchung Cheng } 24646ba8a3b1SNandita Dukkipati 2465a2815817SNeal Cardwell /* If the RTO formula yields an earlier time, then use that time. */ 2466ed66dfafSNeal Cardwell rto_delta_us = advancing_rto ? 2467ed66dfafSNeal Cardwell jiffies_to_usecs(inet_csk(sk)->icsk_rto) : 2468ed66dfafSNeal Cardwell tcp_rto_delta_us(sk); /* How far in future is RTO? */ 2469a2815817SNeal Cardwell if (rto_delta_us > 0) 2470a2815817SNeal Cardwell timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us)); 24716ba8a3b1SNandita Dukkipati 24726ba8a3b1SNandita Dukkipati inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, 24736ba8a3b1SNandita Dukkipati TCP_RTO_MAX); 24746ba8a3b1SNandita Dukkipati return true; 24756ba8a3b1SNandita Dukkipati } 24766ba8a3b1SNandita Dukkipati 24771f3279aeSEric Dumazet /* Thanks to skb fast clones, we can detect if a prior transmit of 24781f3279aeSEric Dumazet * a packet is still in a qdisc or driver queue. 24791f3279aeSEric Dumazet * In this case, there is very little point doing a retransmit ! 24801f3279aeSEric Dumazet */ 24811f3279aeSEric Dumazet static bool skb_still_in_host_queue(const struct sock *sk, 24821f3279aeSEric Dumazet const struct sk_buff *skb) 24831f3279aeSEric Dumazet { 248439bb5e62SEric Dumazet if (unlikely(skb_fclone_busy(sk, skb))) { 2485c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), 24861f3279aeSEric Dumazet LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES); 24871f3279aeSEric Dumazet return true; 24881f3279aeSEric Dumazet } 24891f3279aeSEric Dumazet return false; 24901f3279aeSEric Dumazet } 24911f3279aeSEric Dumazet 2492b340b264SYuchung Cheng /* When probe timeout (PTO) fires, try send a new segment if possible, else 24936ba8a3b1SNandita Dukkipati * retransmit the last segment. 24946ba8a3b1SNandita Dukkipati */ 24956ba8a3b1SNandita Dukkipati void tcp_send_loss_probe(struct sock *sk) 24966ba8a3b1SNandita Dukkipati { 24979b717a8dSNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 24986ba8a3b1SNandita Dukkipati struct sk_buff *skb; 24996ba8a3b1SNandita Dukkipati int pcount; 25006ba8a3b1SNandita Dukkipati int mss = tcp_current_mss(sk); 25016ba8a3b1SNandita Dukkipati 2502b340b264SYuchung Cheng skb = tcp_send_head(sk); 250375c119afSEric Dumazet if (skb && tcp_snd_wnd_test(tp, skb, mss)) { 2504b340b264SYuchung Cheng pcount = tp->packets_out; 2505b340b264SYuchung Cheng tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC); 2506b340b264SYuchung Cheng if (tp->packets_out > pcount) 2507b340b264SYuchung Cheng goto probe_sent; 25086ba8a3b1SNandita Dukkipati goto rearm_timer; 25096ba8a3b1SNandita Dukkipati } 251075c119afSEric Dumazet skb = skb_rb_last(&sk->tcp_rtx_queue); 25116ba8a3b1SNandita Dukkipati 25129b717a8dSNandita Dukkipati /* At most one outstanding TLP retransmission. */ 25139b717a8dSNandita Dukkipati if (tp->tlp_high_seq) 25149b717a8dSNandita Dukkipati goto rearm_timer; 25159b717a8dSNandita Dukkipati 25166ba8a3b1SNandita Dukkipati /* Retransmit last segment. */ 25176ba8a3b1SNandita Dukkipati if (WARN_ON(!skb)) 25186ba8a3b1SNandita Dukkipati goto rearm_timer; 25196ba8a3b1SNandita Dukkipati 25201f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 25211f3279aeSEric Dumazet goto rearm_timer; 25221f3279aeSEric Dumazet 25236ba8a3b1SNandita Dukkipati pcount = tcp_skb_pcount(skb); 25246ba8a3b1SNandita Dukkipati if (WARN_ON(!pcount)) 25256ba8a3b1SNandita Dukkipati goto rearm_timer; 25266ba8a3b1SNandita Dukkipati 25276ba8a3b1SNandita Dukkipati if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) { 252875c119afSEric Dumazet if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, 252975c119afSEric Dumazet (pcount - 1) * mss, mss, 25306cc55e09SOctavian Purdila GFP_ATOMIC))) 25316ba8a3b1SNandita Dukkipati goto rearm_timer; 253275c119afSEric Dumazet skb = skb_rb_next(skb); 25336ba8a3b1SNandita Dukkipati } 25346ba8a3b1SNandita Dukkipati 25356ba8a3b1SNandita Dukkipati if (WARN_ON(!skb || !tcp_skb_pcount(skb))) 25366ba8a3b1SNandita Dukkipati goto rearm_timer; 25376ba8a3b1SNandita Dukkipati 253810d3be56SEric Dumazet if (__tcp_retransmit_skb(sk, skb, 1)) 2539b340b264SYuchung Cheng goto rearm_timer; 25406ba8a3b1SNandita Dukkipati 25419b717a8dSNandita Dukkipati /* Record snd_nxt for loss detection. */ 25429b717a8dSNandita Dukkipati tp->tlp_high_seq = tp->snd_nxt; 25439b717a8dSNandita Dukkipati 2544b340b264SYuchung Cheng probe_sent: 2545c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES); 2546fcd16c0aSYuchung Cheng /* Reset s.t. tcp_rearm_rto will restart timer from now */ 2547fcd16c0aSYuchung Cheng inet_csk(sk)->icsk_pending = 0; 2548b340b264SYuchung Cheng rearm_timer: 2549fcd16c0aSYuchung Cheng tcp_rearm_rto(sk); 25501da177e4SLinus Torvalds } 25511da177e4SLinus Torvalds 2552a762a980SDavid S. Miller /* Push out any pending frames which were held back due to 2553a762a980SDavid S. Miller * TCP_CORK or attempt at coalescing tiny packets. 2554a762a980SDavid S. Miller * The socket must be locked by the caller. 2555a762a980SDavid S. Miller */ 25569e412ba7SIlpo Järvinen void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 25579e412ba7SIlpo Järvinen int nonagle) 2558a762a980SDavid S. Miller { 2559726e07a8SIlpo Järvinen /* If we are closed, the bytes will have to remain here. 2560726e07a8SIlpo Järvinen * In time closedown will finish, we empty the write queue and 2561726e07a8SIlpo Järvinen * all will be happy. 2562726e07a8SIlpo Järvinen */ 2563726e07a8SIlpo Järvinen if (unlikely(sk->sk_state == TCP_CLOSE)) 2564726e07a8SIlpo Järvinen return; 2565726e07a8SIlpo Järvinen 256699a1dec7SMel Gorman if (tcp_write_xmit(sk, cur_mss, nonagle, 0, 25677450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC))) 25689e412ba7SIlpo Järvinen tcp_check_probe_timer(sk); 2569a762a980SDavid S. Miller } 2570a762a980SDavid S. Miller 2571c1b4a7e6SDavid S. Miller /* Send _single_ skb sitting at the send head. This function requires 2572c1b4a7e6SDavid S. Miller * true push pending frames to setup probe timer etc. 2573c1b4a7e6SDavid S. Miller */ 2574c1b4a7e6SDavid S. Miller void tcp_push_one(struct sock *sk, unsigned int mss_now) 2575c1b4a7e6SDavid S. Miller { 2576fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 2577c1b4a7e6SDavid S. Miller 2578c1b4a7e6SDavid S. Miller BUG_ON(!skb || skb->len < mss_now); 2579c1b4a7e6SDavid S. Miller 2580d5dd9175SIlpo Järvinen tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); 2581c1b4a7e6SDavid S. Miller } 2582c1b4a7e6SDavid S. Miller 25831da177e4SLinus Torvalds /* This function returns the amount that we can raise the 25841da177e4SLinus Torvalds * usable window based on the following constraints 25851da177e4SLinus Torvalds * 25861da177e4SLinus Torvalds * 1. The window can never be shrunk once it is offered (RFC 793) 25871da177e4SLinus Torvalds * 2. We limit memory per socket 25881da177e4SLinus Torvalds * 25891da177e4SLinus Torvalds * RFC 1122: 25901da177e4SLinus Torvalds * "the suggested [SWS] avoidance algorithm for the receiver is to keep 25911da177e4SLinus Torvalds * RECV.NEXT + RCV.WIN fixed until: 25921da177e4SLinus Torvalds * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 25931da177e4SLinus Torvalds * 25941da177e4SLinus Torvalds * i.e. don't raise the right edge of the window until you can raise 25951da177e4SLinus Torvalds * it at least MSS bytes. 25961da177e4SLinus Torvalds * 25971da177e4SLinus Torvalds * Unfortunately, the recommended algorithm breaks header prediction, 25981da177e4SLinus Torvalds * since header prediction assumes th->window stays fixed. 25991da177e4SLinus Torvalds * 26001da177e4SLinus Torvalds * Strictly speaking, keeping th->window fixed violates the receiver 26011da177e4SLinus Torvalds * side SWS prevention criteria. The problem is that under this rule 26021da177e4SLinus Torvalds * a stream of single byte packets will cause the right side of the 26031da177e4SLinus Torvalds * window to always advance by a single byte. 26041da177e4SLinus Torvalds * 26051da177e4SLinus Torvalds * Of course, if the sender implements sender side SWS prevention 26061da177e4SLinus Torvalds * then this will not be a problem. 26071da177e4SLinus Torvalds * 26081da177e4SLinus Torvalds * BSD seems to make the following compromise: 26091da177e4SLinus Torvalds * 26101da177e4SLinus Torvalds * If the free space is less than the 1/4 of the maximum 26111da177e4SLinus Torvalds * space available and the free space is less than 1/2 mss, 26121da177e4SLinus Torvalds * then set the window to 0. 26131da177e4SLinus Torvalds * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 26141da177e4SLinus Torvalds * Otherwise, just prevent the window from shrinking 26151da177e4SLinus Torvalds * and from being larger than the largest representable value. 26161da177e4SLinus Torvalds * 26171da177e4SLinus Torvalds * This prevents incremental opening of the window in the regime 26181da177e4SLinus Torvalds * where TCP is limited by the speed of the reader side taking 26191da177e4SLinus Torvalds * data out of the TCP receive queue. It does nothing about 26201da177e4SLinus Torvalds * those cases where the window is constrained on the sender side 26211da177e4SLinus Torvalds * because the pipeline is full. 26221da177e4SLinus Torvalds * 26231da177e4SLinus Torvalds * BSD also seems to "accidentally" limit itself to windows that are a 26241da177e4SLinus Torvalds * multiple of MSS, at least until the free space gets quite small. 26251da177e4SLinus Torvalds * This would appear to be a side effect of the mbuf implementation. 26261da177e4SLinus Torvalds * Combining these two algorithms results in the observed behavior 26271da177e4SLinus Torvalds * of having a fixed window size at almost all times. 26281da177e4SLinus Torvalds * 26291da177e4SLinus Torvalds * Below we obtain similar behavior by forcing the offered window to 26301da177e4SLinus Torvalds * a multiple of the mss when it is feasible to do so. 26311da177e4SLinus Torvalds * 26321da177e4SLinus Torvalds * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 26331da177e4SLinus Torvalds * Regular options like TIMESTAMP are taken into account. 26341da177e4SLinus Torvalds */ 26351da177e4SLinus Torvalds u32 __tcp_select_window(struct sock *sk) 26361da177e4SLinus Torvalds { 2637463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 26381da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2639caa20d9aSStephen Hemminger /* MSS for the peer's data. Previous versions used mss_clamp 26401da177e4SLinus Torvalds * here. I don't know if the value based on our guesses 26411da177e4SLinus Torvalds * of peer's MSS is better for the performance. It's more correct 26421da177e4SLinus Torvalds * but may be worse for the performance because of rcv_mss 26431da177e4SLinus Torvalds * fluctuations. --SAW 1998/11/1 26441da177e4SLinus Torvalds */ 2645463c84b9SArnaldo Carvalho de Melo int mss = icsk->icsk_ack.rcv_mss; 26461da177e4SLinus Torvalds int free_space = tcp_space(sk); 264786c1a045SFlorian Westphal int allowed_space = tcp_full_space(sk); 264886c1a045SFlorian Westphal int full_space = min_t(int, tp->window_clamp, allowed_space); 26491da177e4SLinus Torvalds int window; 26501da177e4SLinus Torvalds 265106425c30SEric Dumazet if (unlikely(mss > full_space)) { 26521da177e4SLinus Torvalds mss = full_space; 265306425c30SEric Dumazet if (mss <= 0) 265406425c30SEric Dumazet return 0; 265506425c30SEric Dumazet } 2656b92edbe0SEric Dumazet if (free_space < (full_space >> 1)) { 2657463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.quick = 0; 26581da177e4SLinus Torvalds 2659b8da51ebSEric Dumazet if (tcp_under_memory_pressure(sk)) 2660056834d9SIlpo Järvinen tp->rcv_ssthresh = min(tp->rcv_ssthresh, 2661056834d9SIlpo Järvinen 4U * tp->advmss); 26621da177e4SLinus Torvalds 266386c1a045SFlorian Westphal /* free_space might become our new window, make sure we don't 266486c1a045SFlorian Westphal * increase it due to wscale. 266586c1a045SFlorian Westphal */ 266686c1a045SFlorian Westphal free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale); 266786c1a045SFlorian Westphal 266886c1a045SFlorian Westphal /* if free space is less than mss estimate, or is below 1/16th 266986c1a045SFlorian Westphal * of the maximum allowed, try to move to zero-window, else 267086c1a045SFlorian Westphal * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and 267186c1a045SFlorian Westphal * new incoming data is dropped due to memory limits. 267286c1a045SFlorian Westphal * With large window, mss test triggers way too late in order 267386c1a045SFlorian Westphal * to announce zero window in time before rmem limit kicks in. 267486c1a045SFlorian Westphal */ 267586c1a045SFlorian Westphal if (free_space < (allowed_space >> 4) || free_space < mss) 26761da177e4SLinus Torvalds return 0; 26771da177e4SLinus Torvalds } 26781da177e4SLinus Torvalds 26791da177e4SLinus Torvalds if (free_space > tp->rcv_ssthresh) 26801da177e4SLinus Torvalds free_space = tp->rcv_ssthresh; 26811da177e4SLinus Torvalds 26821da177e4SLinus Torvalds /* Don't do rounding if we are using window scaling, since the 26831da177e4SLinus Torvalds * scaled window will not line up with the MSS boundary anyway. 26841da177e4SLinus Torvalds */ 26851da177e4SLinus Torvalds if (tp->rx_opt.rcv_wscale) { 26861da177e4SLinus Torvalds window = free_space; 26871da177e4SLinus Torvalds 26881da177e4SLinus Torvalds /* Advertise enough space so that it won't get scaled away. 26891da177e4SLinus Torvalds * Import case: prevent zero window announcement if 26901da177e4SLinus Torvalds * 1<<rcv_wscale > mss. 26911da177e4SLinus Torvalds */ 26921935299dSGao Feng window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale)); 26931da177e4SLinus Torvalds } else { 26941935299dSGao Feng window = tp->rcv_wnd; 26951da177e4SLinus Torvalds /* Get the largest window that is a nice multiple of mss. 26961da177e4SLinus Torvalds * Window clamp already applied above. 26971da177e4SLinus Torvalds * If our current window offering is within 1 mss of the 26981da177e4SLinus Torvalds * free space we just keep it. This prevents the divide 26991da177e4SLinus Torvalds * and multiply from happening most of the time. 27001da177e4SLinus Torvalds * We also don't do any window rounding when the free space 27011da177e4SLinus Torvalds * is too small. 27021da177e4SLinus Torvalds */ 27031da177e4SLinus Torvalds if (window <= free_space - mss || window > free_space) 27041935299dSGao Feng window = rounddown(free_space, mss); 270584565070SJohn Heffner else if (mss == full_space && 2706b92edbe0SEric Dumazet free_space > window + (full_space >> 1)) 270784565070SJohn Heffner window = free_space; 27081da177e4SLinus Torvalds } 27091da177e4SLinus Torvalds 27101da177e4SLinus Torvalds return window; 27111da177e4SLinus Torvalds } 27121da177e4SLinus Torvalds 2713cfea5a68SMartin KaFai Lau void tcp_skb_collapse_tstamp(struct sk_buff *skb, 2714082ac2d5SMartin KaFai Lau const struct sk_buff *next_skb) 2715082ac2d5SMartin KaFai Lau { 27160a2cf20cSSoheil Hassas Yeganeh if (unlikely(tcp_has_tx_tstamp(next_skb))) { 27170a2cf20cSSoheil Hassas Yeganeh const struct skb_shared_info *next_shinfo = 27180a2cf20cSSoheil Hassas Yeganeh skb_shinfo(next_skb); 2719082ac2d5SMartin KaFai Lau struct skb_shared_info *shinfo = skb_shinfo(skb); 2720082ac2d5SMartin KaFai Lau 27210a2cf20cSSoheil Hassas Yeganeh shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP; 2722082ac2d5SMartin KaFai Lau shinfo->tskey = next_shinfo->tskey; 27232de8023eSMartin KaFai Lau TCP_SKB_CB(skb)->txstamp_ack |= 27242de8023eSMartin KaFai Lau TCP_SKB_CB(next_skb)->txstamp_ack; 2725082ac2d5SMartin KaFai Lau } 2726082ac2d5SMartin KaFai Lau } 2727082ac2d5SMartin KaFai Lau 27284a17fc3aSIlpo Järvinen /* Collapses two adjacent SKB's during retransmission. */ 2729f8071cdeSEric Dumazet static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) 27301da177e4SLinus Torvalds { 27311da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 273275c119afSEric Dumazet struct sk_buff *next_skb = skb_rb_next(skb); 2733058dc334SIlpo Järvinen int skb_size, next_skb_size; 27341da177e4SLinus Torvalds 2735058dc334SIlpo Järvinen skb_size = skb->len; 2736058dc334SIlpo Järvinen next_skb_size = next_skb->len; 27371da177e4SLinus Torvalds 2738058dc334SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); 27391da177e4SLinus Torvalds 2740f8071cdeSEric Dumazet if (next_skb_size) { 2741f8071cdeSEric Dumazet if (next_skb_size <= skb_availroom(skb)) 2742f8071cdeSEric Dumazet skb_copy_bits(next_skb, 0, skb_put(skb, next_skb_size), 2743f8071cdeSEric Dumazet next_skb_size); 2744f8071cdeSEric Dumazet else if (!skb_shift(skb, next_skb, next_skb_size)) 2745f8071cdeSEric Dumazet return false; 2746f8071cdeSEric Dumazet } 27472b7cda9cSEric Dumazet tcp_highest_sack_replace(sk, next_skb, skb); 2748a6963a6bSIlpo Järvinen 274952d570aaSJarek Poplawski if (next_skb->ip_summed == CHECKSUM_PARTIAL) 275052d570aaSJarek Poplawski skb->ip_summed = CHECKSUM_PARTIAL; 27511da177e4SLinus Torvalds 275284fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 27531da177e4SLinus Torvalds skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); 27541da177e4SLinus Torvalds 27551da177e4SLinus Torvalds /* Update sequence range on original skb. */ 27561da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 27571da177e4SLinus Torvalds 2758e6c7d085SIlpo Järvinen /* Merge over control information. This moves PSH/FIN etc. over */ 27594de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; 27601da177e4SLinus Torvalds 27611da177e4SLinus Torvalds /* All done, get rid of second SKB and account for it so 27621da177e4SLinus Torvalds * packet counting does not break. 27631da177e4SLinus Torvalds */ 27644828e7f4SIlpo Järvinen TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; 2765a643b5d4SMartin KaFai Lau TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor; 2766b7689205SIlpo Järvinen 2767b7689205SIlpo Järvinen /* changed transmit queue under us so clear hints */ 2768ef9da47cSIlpo Järvinen tcp_clear_retrans_hints_partial(tp); 2769ef9da47cSIlpo Järvinen if (next_skb == tp->retransmit_skb_hint) 2770ef9da47cSIlpo Järvinen tp->retransmit_skb_hint = skb; 2771b7689205SIlpo Järvinen 2772797108d1SIlpo Järvinen tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); 2773797108d1SIlpo Järvinen 2774082ac2d5SMartin KaFai Lau tcp_skb_collapse_tstamp(skb, next_skb); 2775082ac2d5SMartin KaFai Lau 277675c119afSEric Dumazet tcp_rtx_queue_unlink_and_free(next_skb, sk); 2777f8071cdeSEric Dumazet return true; 27781da177e4SLinus Torvalds } 27791da177e4SLinus Torvalds 278067edfef7SAndi Kleen /* Check if coalescing SKBs is legal. */ 2781a2a385d6SEric Dumazet static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) 27824a17fc3aSIlpo Järvinen { 27834a17fc3aSIlpo Järvinen if (tcp_skb_pcount(skb) > 1) 2784a2a385d6SEric Dumazet return false; 27854a17fc3aSIlpo Järvinen if (skb_cloned(skb)) 2786a2a385d6SEric Dumazet return false; 27872331ccc5SEric Dumazet /* Some heuristics for collapsing over SACK'd could be invented */ 27884a17fc3aSIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 2789a2a385d6SEric Dumazet return false; 27904a17fc3aSIlpo Järvinen 2791a2a385d6SEric Dumazet return true; 27924a17fc3aSIlpo Järvinen } 27934a17fc3aSIlpo Järvinen 279467edfef7SAndi Kleen /* Collapse packets in the retransmit queue to make to create 279567edfef7SAndi Kleen * less packets on the wire. This is only done on retransmission. 279667edfef7SAndi Kleen */ 27974a17fc3aSIlpo Järvinen static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, 27984a17fc3aSIlpo Järvinen int space) 27994a17fc3aSIlpo Järvinen { 28004a17fc3aSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 28014a17fc3aSIlpo Järvinen struct sk_buff *skb = to, *tmp; 2802a2a385d6SEric Dumazet bool first = true; 28034a17fc3aSIlpo Järvinen 2804e0a1e5b5SEric Dumazet if (!sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse) 28054a17fc3aSIlpo Järvinen return; 28064de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 28074a17fc3aSIlpo Järvinen return; 28084a17fc3aSIlpo Järvinen 280975c119afSEric Dumazet skb_rbtree_walk_from_safe(skb, tmp) { 28104a17fc3aSIlpo Järvinen if (!tcp_can_collapse(sk, skb)) 28114a17fc3aSIlpo Järvinen break; 28124a17fc3aSIlpo Järvinen 2813a643b5d4SMartin KaFai Lau if (!tcp_skb_can_collapse_to(to)) 2814a643b5d4SMartin KaFai Lau break; 2815a643b5d4SMartin KaFai Lau 28164a17fc3aSIlpo Järvinen space -= skb->len; 28174a17fc3aSIlpo Järvinen 28184a17fc3aSIlpo Järvinen if (first) { 2819a2a385d6SEric Dumazet first = false; 28204a17fc3aSIlpo Järvinen continue; 28214a17fc3aSIlpo Järvinen } 28224a17fc3aSIlpo Järvinen 28234a17fc3aSIlpo Järvinen if (space < 0) 28244a17fc3aSIlpo Järvinen break; 28254a17fc3aSIlpo Järvinen 28264a17fc3aSIlpo Järvinen if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) 28274a17fc3aSIlpo Järvinen break; 28284a17fc3aSIlpo Järvinen 2829f8071cdeSEric Dumazet if (!tcp_collapse_retrans(sk, to)) 2830f8071cdeSEric Dumazet break; 28314a17fc3aSIlpo Järvinen } 28324a17fc3aSIlpo Järvinen } 28334a17fc3aSIlpo Järvinen 28341da177e4SLinus Torvalds /* This retransmits one SKB. Policy decisions and retransmit queue 28351da177e4SLinus Torvalds * state updates are done by the caller. Returns non-zero if an 28361da177e4SLinus Torvalds * error occurred which prevented the send. 28371da177e4SLinus Torvalds */ 283810d3be56SEric Dumazet int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) 28391da177e4SLinus Torvalds { 28405d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 284110d3be56SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 28427d227cd2SSridhar Samudrala unsigned int cur_mss; 284310d3be56SEric Dumazet int diff, len, err; 28441da177e4SLinus Torvalds 284510d3be56SEric Dumazet 284610d3be56SEric Dumazet /* Inconclusive MTU probe */ 284710d3be56SEric Dumazet if (icsk->icsk_mtup.probe_size) 28485d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 28495d424d5aSJohn Heffner 28501da177e4SLinus Torvalds /* Do not sent more than we queued. 1/4 is reserved for possible 2851caa20d9aSStephen Hemminger * copying overhead: fragmentation, tunneling, mangling etc. 28521da177e4SLinus Torvalds */ 285314afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) > 2854ffb4d6c8SEric Dumazet min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), 2855ffb4d6c8SEric Dumazet sk->sk_sndbuf)) 28561da177e4SLinus Torvalds return -EAGAIN; 28571da177e4SLinus Torvalds 28581f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 28591f3279aeSEric Dumazet return -EBUSY; 28601f3279aeSEric Dumazet 28611da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 28621da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 28631da177e4SLinus Torvalds BUG(); 28641da177e4SLinus Torvalds if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 28651da177e4SLinus Torvalds return -ENOMEM; 28661da177e4SLinus Torvalds } 28671da177e4SLinus Torvalds 28687d227cd2SSridhar Samudrala if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 28697d227cd2SSridhar Samudrala return -EHOSTUNREACH; /* Routing failure or similar. */ 28707d227cd2SSridhar Samudrala 28710c54b85fSIlpo Järvinen cur_mss = tcp_current_mss(sk); 28727d227cd2SSridhar Samudrala 28731da177e4SLinus Torvalds /* If receiver has shrunk his window, and skb is out of 28741da177e4SLinus Torvalds * new window, do not retransmit it. The exception is the 28751da177e4SLinus Torvalds * case, when window is shrunk to zero. In this case 28761da177e4SLinus Torvalds * our retransmit serves as a zero window probe. 28771da177e4SLinus Torvalds */ 28789d4fb27dSJoe Perches if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && 28799d4fb27dSJoe Perches TCP_SKB_CB(skb)->seq != tp->snd_una) 28801da177e4SLinus Torvalds return -EAGAIN; 28811da177e4SLinus Torvalds 288210d3be56SEric Dumazet len = cur_mss * segs; 288310d3be56SEric Dumazet if (skb->len > len) { 288475c119afSEric Dumazet if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len, 288575c119afSEric Dumazet cur_mss, GFP_ATOMIC)) 28861da177e4SLinus Torvalds return -ENOMEM; /* We'll try again later. */ 288702276f3cSIlpo Järvinen } else { 2888c52e2421SEric Dumazet if (skb_unclone(skb, GFP_ATOMIC)) 2889c52e2421SEric Dumazet return -ENOMEM; 289010d3be56SEric Dumazet 289110d3be56SEric Dumazet diff = tcp_skb_pcount(skb); 289210d3be56SEric Dumazet tcp_set_skb_tso_segs(skb, cur_mss); 289310d3be56SEric Dumazet diff -= tcp_skb_pcount(skb); 289410d3be56SEric Dumazet if (diff) 289510d3be56SEric Dumazet tcp_adjust_pcount(sk, skb, diff); 289610d3be56SEric Dumazet if (skb->len < cur_mss) 289710d3be56SEric Dumazet tcp_retrans_try_collapse(sk, skb, cur_mss); 28981da177e4SLinus Torvalds } 28991da177e4SLinus Torvalds 290049213555SDaniel Borkmann /* RFC3168, section 6.1.1.1. ECN fallback */ 290149213555SDaniel Borkmann if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN) 290249213555SDaniel Borkmann tcp_ecn_clear_syn(sk, skb); 290349213555SDaniel Borkmann 2904678550c6SYuchung Cheng /* Update global and local TCP statistics. */ 2905678550c6SYuchung Cheng segs = tcp_skb_pcount(skb); 2906678550c6SYuchung Cheng TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs); 2907678550c6SYuchung Cheng if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 2908678550c6SYuchung Cheng __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 2909678550c6SYuchung Cheng tp->total_retrans += segs; 2910678550c6SYuchung Cheng 291150bceae9SThomas Graf /* make sure skb->data is aligned on arches that require it 291250bceae9SThomas Graf * and check if ack-trimming & collapsing extended the headroom 291350bceae9SThomas Graf * beyond what csum_start can cover. 291450bceae9SThomas Graf */ 291550bceae9SThomas Graf if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) || 291650bceae9SThomas Graf skb_headroom(skb) >= 0xFFFF)) { 291710a81980SEric Dumazet struct sk_buff *nskb; 291810a81980SEric Dumazet 2919e2080072SEric Dumazet tcp_skb_tsorted_save(skb) { 292010a81980SEric Dumazet nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC); 2921c84a5711SYuchung Cheng err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) : 2922117632e6SEric Dumazet -ENOBUFS; 2923e2080072SEric Dumazet } tcp_skb_tsorted_restore(skb); 2924e2080072SEric Dumazet 29255889e2c0SYousuk Seung if (!err) { 2926e2080072SEric Dumazet tcp_update_skb_after_send(tp, skb); 29275889e2c0SYousuk Seung tcp_rate_skb_sent(sk, skb); 29285889e2c0SYousuk Seung } 2929117632e6SEric Dumazet } else { 2930c84a5711SYuchung Cheng err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 2931117632e6SEric Dumazet } 2932c84a5711SYuchung Cheng 2933a31ad29eSLawrence Brakmo if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RETRANS_CB_FLAG)) 2934a31ad29eSLawrence Brakmo tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RETRANS_CB, 2935a31ad29eSLawrence Brakmo TCP_SKB_CB(skb)->seq, segs, err); 2936a31ad29eSLawrence Brakmo 2937fc9f3501SEric Dumazet if (likely(!err)) { 2938c84a5711SYuchung Cheng TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS; 2939e086101bSCong Wang trace_tcp_retransmit_skb(sk, skb); 2940678550c6SYuchung Cheng } else if (err != -EBUSY) { 2941678550c6SYuchung Cheng NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL); 2942fc9f3501SEric Dumazet } 2943c84a5711SYuchung Cheng return err; 294493b174adSYuchung Cheng } 294593b174adSYuchung Cheng 294610d3be56SEric Dumazet int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) 294793b174adSYuchung Cheng { 294893b174adSYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 294910d3be56SEric Dumazet int err = __tcp_retransmit_skb(sk, skb, segs); 29501da177e4SLinus Torvalds 29511da177e4SLinus Torvalds if (err == 0) { 29521da177e4SLinus Torvalds #if FASTRETRANS_DEBUG > 0 29531da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 2954e87cc472SJoe Perches net_dbg_ratelimited("retrans_out leaked\n"); 29551da177e4SLinus Torvalds } 29561da177e4SLinus Torvalds #endif 29571da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 29581da177e4SLinus Torvalds tp->retrans_out += tcp_skb_pcount(skb); 29591da177e4SLinus Torvalds 29601da177e4SLinus Torvalds /* Save stamp of the first retransmit. */ 29611da177e4SLinus Torvalds if (!tp->retrans_stamp) 29627faee5c0SEric Dumazet tp->retrans_stamp = tcp_skb_timestamp(skb); 29631da177e4SLinus Torvalds 29641da177e4SLinus Torvalds } 29656e08d5e3SYuchung Cheng 29666e08d5e3SYuchung Cheng if (tp->undo_retrans < 0) 29676e08d5e3SYuchung Cheng tp->undo_retrans = 0; 29686e08d5e3SYuchung Cheng tp->undo_retrans += tcp_skb_pcount(skb); 29691da177e4SLinus Torvalds return err; 29701da177e4SLinus Torvalds } 29711da177e4SLinus Torvalds 29721da177e4SLinus Torvalds /* This gets called after a retransmit timeout, and the initially 29731da177e4SLinus Torvalds * retransmitted data is acknowledged. It tries to continue 29741da177e4SLinus Torvalds * resending the rest of the retransmit queue, until either 29751da177e4SLinus Torvalds * we've sent it all or the congestion window limit is reached. 29761da177e4SLinus Torvalds */ 29771da177e4SLinus Torvalds void tcp_xmit_retransmit_queue(struct sock *sk) 29781da177e4SLinus Torvalds { 29796687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 2980b9f1f1ceSEric Dumazet struct sk_buff *skb, *rtx_head, *hole = NULL; 29811da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2982840a3cbeSYuchung Cheng u32 max_segs; 298361eb55f4SIlpo Järvinen int mib_idx; 29846a438bbeSStephen Hemminger 298545e77d31SIlpo Järvinen if (!tp->packets_out) 298645e77d31SIlpo Järvinen return; 298745e77d31SIlpo Järvinen 298875c119afSEric Dumazet rtx_head = tcp_rtx_queue_head(sk); 2989b9f1f1ceSEric Dumazet skb = tp->retransmit_skb_hint ?: rtx_head; 2990ed6e7268SNeal Cardwell max_segs = tcp_tso_segs(sk, tcp_current_mss(sk)); 299175c119afSEric Dumazet skb_rbtree_walk_from(skb) { 2992dca0aaf8SEric Dumazet __u8 sacked; 299310d3be56SEric Dumazet int segs; 29941da177e4SLinus Torvalds 2995218af599SEric Dumazet if (tcp_pacing_check(sk)) 2996218af599SEric Dumazet break; 2997218af599SEric Dumazet 29986a438bbeSStephen Hemminger /* we could do better than to assign each time */ 299951456b29SIan Morris if (!hole) 30006a438bbeSStephen Hemminger tp->retransmit_skb_hint = skb; 30016a438bbeSStephen Hemminger 300210d3be56SEric Dumazet segs = tp->snd_cwnd - tcp_packets_in_flight(tp); 300310d3be56SEric Dumazet if (segs <= 0) 30041da177e4SLinus Torvalds return; 3005dca0aaf8SEric Dumazet sacked = TCP_SKB_CB(skb)->sacked; 3006a3d2e9f8SEric Dumazet /* In case tcp_shift_skb_data() have aggregated large skbs, 3007a3d2e9f8SEric Dumazet * we need to make sure not sending too bigs TSO packets 3008a3d2e9f8SEric Dumazet */ 3009a3d2e9f8SEric Dumazet segs = min_t(int, segs, max_segs); 30100e1c54c2SIlpo Järvinen 3011840a3cbeSYuchung Cheng if (tp->retrans_out >= tp->lost_out) { 3012006f582cSIlpo Järvinen break; 30130e1c54c2SIlpo Järvinen } else if (!(sacked & TCPCB_LOST)) { 301451456b29SIan Morris if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) 30150e1c54c2SIlpo Järvinen hole = skb; 301661eb55f4SIlpo Järvinen continue; 30171da177e4SLinus Torvalds 30180e1c54c2SIlpo Järvinen } else { 30190e1c54c2SIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Loss) 30200e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFASTRETRANS; 30210e1c54c2SIlpo Järvinen else 30220e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; 30230e1c54c2SIlpo Järvinen } 30240e1c54c2SIlpo Järvinen 30250e1c54c2SIlpo Järvinen if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) 302661eb55f4SIlpo Järvinen continue; 302740b215e5SPavel Emelyanov 3028f9616c35SEric Dumazet if (tcp_small_queue_check(sk, skb, 1)) 3029f9616c35SEric Dumazet return; 3030f9616c35SEric Dumazet 303110d3be56SEric Dumazet if (tcp_retransmit_skb(sk, skb, segs)) 30321da177e4SLinus Torvalds return; 303324ab6becSYuchung Cheng 3034de1d6578SYuchung Cheng NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb)); 30351da177e4SLinus Torvalds 3036684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 3037a262f0cdSNandita Dukkipati tp->prr_out += tcp_skb_pcount(skb); 3038a262f0cdSNandita Dukkipati 303975c119afSEric Dumazet if (skb == rtx_head && 304057dde7f7SYuchung Cheng icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT) 3041463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 30423f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, 30433f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 30441da177e4SLinus Torvalds } 30451da177e4SLinus Torvalds } 30461da177e4SLinus Torvalds 3047d83769a5SEric Dumazet /* We allow to exceed memory limits for FIN packets to expedite 3048d83769a5SEric Dumazet * connection tear down and (memory) recovery. 3049845704a5SEric Dumazet * Otherwise tcp_send_fin() could be tempted to either delay FIN 3050845704a5SEric Dumazet * or even be forced to close flow without any FIN. 3051a6c5ea4cSEric Dumazet * In general, we want to allow one skb per socket to avoid hangs 3052a6c5ea4cSEric Dumazet * with edge trigger epoll() 3053d83769a5SEric Dumazet */ 3054a6c5ea4cSEric Dumazet void sk_forced_mem_schedule(struct sock *sk, int size) 3055d83769a5SEric Dumazet { 3056e805605cSJohannes Weiner int amt; 3057d83769a5SEric Dumazet 3058d83769a5SEric Dumazet if (size <= sk->sk_forward_alloc) 3059d83769a5SEric Dumazet return; 3060d83769a5SEric Dumazet amt = sk_mem_pages(size); 3061d83769a5SEric Dumazet sk->sk_forward_alloc += amt * SK_MEM_QUANTUM; 3062e805605cSJohannes Weiner sk_memory_allocated_add(sk, amt); 3063e805605cSJohannes Weiner 3064baac50bbSJohannes Weiner if (mem_cgroup_sockets_enabled && sk->sk_memcg) 3065baac50bbSJohannes Weiner mem_cgroup_charge_skmem(sk->sk_memcg, amt); 3066d83769a5SEric Dumazet } 3067d83769a5SEric Dumazet 3068845704a5SEric Dumazet /* Send a FIN. The caller locks the socket for us. 3069845704a5SEric Dumazet * We should try to send a FIN packet really hard, but eventually give up. 30701da177e4SLinus Torvalds */ 30711da177e4SLinus Torvalds void tcp_send_fin(struct sock *sk) 30721da177e4SLinus Torvalds { 3073845704a5SEric Dumazet struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk); 30741da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 30751da177e4SLinus Torvalds 3076845704a5SEric Dumazet /* Optimization, tack on the FIN if we have one skb in write queue and 3077845704a5SEric Dumazet * this skb was not yet sent, or we are under memory pressure. 3078845704a5SEric Dumazet * Note: in the latter case, FIN packet will be sent after a timeout, 3079845704a5SEric Dumazet * as TCP stack thinks it has already been transmitted. 30801da177e4SLinus Torvalds */ 308175c119afSEric Dumazet if (!tskb && tcp_under_memory_pressure(sk)) 308275c119afSEric Dumazet tskb = skb_rb_last(&sk->tcp_rtx_queue); 308375c119afSEric Dumazet 308475c119afSEric Dumazet if (tskb) { 3085845704a5SEric Dumazet coalesce: 3086845704a5SEric Dumazet TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN; 3087845704a5SEric Dumazet TCP_SKB_CB(tskb)->end_seq++; 30881da177e4SLinus Torvalds tp->write_seq++; 308975c119afSEric Dumazet if (tcp_write_queue_empty(sk)) { 3090845704a5SEric Dumazet /* This means tskb was already sent. 3091845704a5SEric Dumazet * Pretend we included the FIN on previous transmit. 3092845704a5SEric Dumazet * We need to set tp->snd_nxt to the value it would have 3093845704a5SEric Dumazet * if FIN had been sent. This is because retransmit path 3094845704a5SEric Dumazet * does not change tp->snd_nxt. 3095845704a5SEric Dumazet */ 3096845704a5SEric Dumazet tp->snd_nxt++; 3097845704a5SEric Dumazet return; 3098845704a5SEric Dumazet } 30991da177e4SLinus Torvalds } else { 3100845704a5SEric Dumazet skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation); 3101845704a5SEric Dumazet if (unlikely(!skb)) { 3102845704a5SEric Dumazet if (tskb) 3103845704a5SEric Dumazet goto coalesce; 3104845704a5SEric Dumazet return; 31051da177e4SLinus Torvalds } 3106e2080072SEric Dumazet INIT_LIST_HEAD(&skb->tcp_tsorted_anchor); 3107d83769a5SEric Dumazet skb_reserve(skb, MAX_TCP_HEADER); 3108a6c5ea4cSEric Dumazet sk_forced_mem_schedule(sk, skb->truesize); 31091da177e4SLinus Torvalds /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 3110e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tp->write_seq, 3111a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_FIN); 31121da177e4SLinus Torvalds tcp_queue_skb(sk, skb); 31131da177e4SLinus Torvalds } 3114845704a5SEric Dumazet __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF); 31151da177e4SLinus Torvalds } 31161da177e4SLinus Torvalds 31171da177e4SLinus Torvalds /* We get here when a process closes a file descriptor (either due to 31181da177e4SLinus Torvalds * an explicit close() or as a byproduct of exit()'ing) and there 31191da177e4SLinus Torvalds * was unread data in the receive queue. This behavior is recommended 312065bb723cSGerrit Renker * by RFC 2525, section 2.17. -DaveM 31211da177e4SLinus Torvalds */ 3122dd0fc66fSAl Viro void tcp_send_active_reset(struct sock *sk, gfp_t priority) 31231da177e4SLinus Torvalds { 31241da177e4SLinus Torvalds struct sk_buff *skb; 31251da177e4SLinus Torvalds 31267cc2b043SGao Feng TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); 31277cc2b043SGao Feng 31281da177e4SLinus Torvalds /* NOTE: No TCP options attached and we never retransmit this. */ 31291da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, priority); 31301da177e4SLinus Torvalds if (!skb) { 31314e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 31321da177e4SLinus Torvalds return; 31331da177e4SLinus Torvalds } 31341da177e4SLinus Torvalds 31351da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 31361da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 3137e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), 3138a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_RST); 31399a568de4SEric Dumazet tcp_mstamp_refresh(tcp_sk(sk)); 31401da177e4SLinus Torvalds /* Send it off. */ 3141dfb4b9dcSDavid S. Miller if (tcp_transmit_skb(sk, skb, 0, priority)) 31424e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 3143c24b14c4SSong Liu 3144c24b14c4SSong Liu /* skb of trace_tcp_send_reset() keeps the skb that caused RST, 3145c24b14c4SSong Liu * skb here is different to the troublesome skb, so use NULL 3146c24b14c4SSong Liu */ 3147c24b14c4SSong Liu trace_tcp_send_reset(sk, NULL); 31481da177e4SLinus Torvalds } 31491da177e4SLinus Torvalds 315067edfef7SAndi Kleen /* Send a crossed SYN-ACK during socket establishment. 315167edfef7SAndi Kleen * WARNING: This routine must only be called when we have already sent 31521da177e4SLinus Torvalds * a SYN packet that crossed the incoming SYN that caused this routine 31531da177e4SLinus Torvalds * to get called. If this assumption fails then the initial rcv_wnd 31541da177e4SLinus Torvalds * and rcv_wscale values will not be correct. 31551da177e4SLinus Torvalds */ 31561da177e4SLinus Torvalds int tcp_send_synack(struct sock *sk) 31571da177e4SLinus Torvalds { 31581da177e4SLinus Torvalds struct sk_buff *skb; 31591da177e4SLinus Torvalds 316075c119afSEric Dumazet skb = tcp_rtx_queue_head(sk); 316151456b29SIan Morris if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 316275c119afSEric Dumazet pr_err("%s: wrong queue state\n", __func__); 31631da177e4SLinus Torvalds return -EFAULT; 31641da177e4SLinus Torvalds } 31654de075e0SEric Dumazet if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { 31661da177e4SLinus Torvalds if (skb_cloned(skb)) { 3167e2080072SEric Dumazet struct sk_buff *nskb; 3168e2080072SEric Dumazet 3169e2080072SEric Dumazet tcp_skb_tsorted_save(skb) { 3170e2080072SEric Dumazet nskb = skb_copy(skb, GFP_ATOMIC); 3171e2080072SEric Dumazet } tcp_skb_tsorted_restore(skb); 317251456b29SIan Morris if (!nskb) 31731da177e4SLinus Torvalds return -ENOMEM; 3174e2080072SEric Dumazet INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor); 317575c119afSEric Dumazet tcp_rtx_queue_unlink_and_free(skb, sk); 3176f4a775d1SEric Dumazet __skb_header_release(nskb); 317775c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb); 31783ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 31793ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 31801da177e4SLinus Torvalds skb = nskb; 31811da177e4SLinus Torvalds } 31821da177e4SLinus Torvalds 31834de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; 3184735d3831SFlorian Westphal tcp_ecn_send_synack(sk, skb); 31851da177e4SLinus Torvalds } 3186dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 31871da177e4SLinus Torvalds } 31881da177e4SLinus Torvalds 31894aea39c1SEric Dumazet /** 31904aea39c1SEric Dumazet * tcp_make_synack - Prepare a SYN-ACK. 31914aea39c1SEric Dumazet * sk: listener socket 31924aea39c1SEric Dumazet * dst: dst entry attached to the SYNACK 31934aea39c1SEric Dumazet * req: request_sock pointer 31944aea39c1SEric Dumazet * 31954aea39c1SEric Dumazet * Allocate one skb and build a SYNACK packet. 31964aea39c1SEric Dumazet * @dst is consumed : Caller should not use it again. 31974aea39c1SEric Dumazet */ 31985d062de7SEric Dumazet struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst, 3199e6b4d113SWilliam Allen Simpson struct request_sock *req, 3200ca6fb065SEric Dumazet struct tcp_fastopen_cookie *foc, 3201b3d05147SEric Dumazet enum tcp_synack_type synack_type) 32021da177e4SLinus Torvalds { 32032e6599cbSArnaldo Carvalho de Melo struct inet_request_sock *ireq = inet_rsk(req); 32045d062de7SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 320580f03e27SEric Dumazet struct tcp_md5sig_key *md5 = NULL; 32065d062de7SEric Dumazet struct tcp_out_options opts; 32075d062de7SEric Dumazet struct sk_buff *skb; 3208bd0388aeSWilliam Allen Simpson int tcp_header_size; 32095d062de7SEric Dumazet struct tcphdr *th; 3210f5fff5dcSTom Quetchenbach int mss; 32111da177e4SLinus Torvalds 3212ca6fb065SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 32134aea39c1SEric Dumazet if (unlikely(!skb)) { 32144aea39c1SEric Dumazet dst_release(dst); 32151da177e4SLinus Torvalds return NULL; 32164aea39c1SEric Dumazet } 32171da177e4SLinus Torvalds /* Reserve space for headers. */ 32181da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 32191da177e4SLinus Torvalds 3220b3d05147SEric Dumazet switch (synack_type) { 3221b3d05147SEric Dumazet case TCP_SYNACK_NORMAL: 32229e17f8a4SEric Dumazet skb_set_owner_w(skb, req_to_sk(req)); 3223b3d05147SEric Dumazet break; 3224b3d05147SEric Dumazet case TCP_SYNACK_COOKIE: 3225b3d05147SEric Dumazet /* Under synflood, we do not attach skb to a socket, 3226b3d05147SEric Dumazet * to avoid false sharing. 3227b3d05147SEric Dumazet */ 3228b3d05147SEric Dumazet break; 3229b3d05147SEric Dumazet case TCP_SYNACK_FASTOPEN: 3230ca6fb065SEric Dumazet /* sk is a const pointer, because we want to express multiple 3231ca6fb065SEric Dumazet * cpu might call us concurrently. 3232ca6fb065SEric Dumazet * sk->sk_wmem_alloc in an atomic, we can promote to rw. 3233ca6fb065SEric Dumazet */ 3234ca6fb065SEric Dumazet skb_set_owner_w(skb, (struct sock *)sk); 3235b3d05147SEric Dumazet break; 3236ca6fb065SEric Dumazet } 32374aea39c1SEric Dumazet skb_dst_set(skb, dst); 32381da177e4SLinus Torvalds 32393541f9e8SEric Dumazet mss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 3240f5fff5dcSTom Quetchenbach 324133ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 32428b5f12d0SFlorian Westphal #ifdef CONFIG_SYN_COOKIES 32438b5f12d0SFlorian Westphal if (unlikely(req->cookie_ts)) 32449a568de4SEric Dumazet skb->skb_mstamp = cookie_init_timestamp(req); 32458b5f12d0SFlorian Westphal else 32468b5f12d0SFlorian Westphal #endif 32479a568de4SEric Dumazet skb->skb_mstamp = tcp_clock_us(); 324880f03e27SEric Dumazet 324980f03e27SEric Dumazet #ifdef CONFIG_TCP_MD5SIG 325080f03e27SEric Dumazet rcu_read_lock(); 3251fd3a154aSEric Dumazet md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req)); 325280f03e27SEric Dumazet #endif 325358d607d3SEric Dumazet skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4); 325460e2a778SUrsula Braun tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5, 325560e2a778SUrsula Braun foc) + sizeof(*th); 325633ad798cSAdam Langley 3257aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 3258aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 32591da177e4SLinus Torvalds 3260ea1627c2SEric Dumazet th = (struct tcphdr *)skb->data; 32611da177e4SLinus Torvalds memset(th, 0, sizeof(struct tcphdr)); 32621da177e4SLinus Torvalds th->syn = 1; 32631da177e4SLinus Torvalds th->ack = 1; 32646ac705b1SEric Dumazet tcp_ecn_make_synack(req, th); 3265b44084c2SEric Dumazet th->source = htons(ireq->ir_num); 3266634fb979SEric Dumazet th->dest = ireq->ir_rmt_port; 3267e05a90ecSJamal Hadi Salim skb->mark = ireq->ir_mark; 32683b117750SEric Dumazet skb->ip_summed = CHECKSUM_PARTIAL; 32693b117750SEric Dumazet th->seq = htonl(tcp_rsk(req)->snt_isn); 32708336886fSJerry Chu /* XXX data is queued and acked as is. No buffer/window check */ 32718336886fSJerry Chu th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt); 32721da177e4SLinus Torvalds 32731da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 3274ed53d0abSEric Dumazet th->window = htons(min(req->rsk_rcv_wnd, 65535U)); 32755d062de7SEric Dumazet tcp_options_write((__be32 *)(th + 1), NULL, &opts); 32761da177e4SLinus Torvalds th->doff = (tcp_header_size >> 2); 327790bbcc60SEric Dumazet __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS); 3278cfb6eeb4SYOSHIFUJI Hideaki 3279cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 3280cfb6eeb4SYOSHIFUJI Hideaki /* Okay, we have all we need - do the md5 hash if needed */ 328180f03e27SEric Dumazet if (md5) 3282bd0388aeSWilliam Allen Simpson tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, 328339f8e58eSEric Dumazet md5, req_to_sk(req), skb); 328480f03e27SEric Dumazet rcu_read_unlock(); 3285cfb6eeb4SYOSHIFUJI Hideaki #endif 3286cfb6eeb4SYOSHIFUJI Hideaki 3287b50edd78SEric Dumazet /* Do not fool tcpdump (if any), clean our debris */ 32882456e855SThomas Gleixner skb->tstamp = 0; 32891da177e4SLinus Torvalds return skb; 32901da177e4SLinus Torvalds } 32914bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_make_synack); 32921da177e4SLinus Torvalds 329381164413SDaniel Borkmann static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst) 329481164413SDaniel Borkmann { 329581164413SDaniel Borkmann struct inet_connection_sock *icsk = inet_csk(sk); 329681164413SDaniel Borkmann const struct tcp_congestion_ops *ca; 329781164413SDaniel Borkmann u32 ca_key = dst_metric(dst, RTAX_CC_ALGO); 329881164413SDaniel Borkmann 329981164413SDaniel Borkmann if (ca_key == TCP_CA_UNSPEC) 330081164413SDaniel Borkmann return; 330181164413SDaniel Borkmann 330281164413SDaniel Borkmann rcu_read_lock(); 330381164413SDaniel Borkmann ca = tcp_ca_find_key(ca_key); 330481164413SDaniel Borkmann if (likely(ca && try_module_get(ca->owner))) { 330581164413SDaniel Borkmann module_put(icsk->icsk_ca_ops->owner); 330681164413SDaniel Borkmann icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst); 330781164413SDaniel Borkmann icsk->icsk_ca_ops = ca; 330881164413SDaniel Borkmann } 330981164413SDaniel Borkmann rcu_read_unlock(); 331081164413SDaniel Borkmann } 331181164413SDaniel Borkmann 331267edfef7SAndi Kleen /* Do all connect socket setups that can be done AF independent. */ 3313f7e56a76Sstephen hemminger static void tcp_connect_init(struct sock *sk) 33141da177e4SLinus Torvalds { 3315cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 33161da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 33171da177e4SLinus Torvalds __u8 rcv_wscale; 331813d3b1ebSLawrence Brakmo u32 rcv_wnd; 33191da177e4SLinus Torvalds 33201da177e4SLinus Torvalds /* We'll fix this up when we get a response from the other end. 33211da177e4SLinus Torvalds * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 33221da177e4SLinus Torvalds */ 33235d2ed052SEric Dumazet tp->tcp_header_len = sizeof(struct tcphdr); 33245d2ed052SEric Dumazet if (sock_net(sk)->ipv4.sysctl_tcp_timestamps) 33255d2ed052SEric Dumazet tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED; 33261da177e4SLinus Torvalds 3327cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 332800db4124SIan Morris if (tp->af_specific->md5_lookup(sk, sk)) 3329cfb6eeb4SYOSHIFUJI Hideaki tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 3330cfb6eeb4SYOSHIFUJI Hideaki #endif 3331cfb6eeb4SYOSHIFUJI Hideaki 33321da177e4SLinus Torvalds /* If user gave his TCP_MAXSEG, record it to clamp */ 33331da177e4SLinus Torvalds if (tp->rx_opt.user_mss) 33341da177e4SLinus Torvalds tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 33351da177e4SLinus Torvalds tp->max_window = 0; 33365d424d5aSJohn Heffner tcp_mtup_init(sk); 33371da177e4SLinus Torvalds tcp_sync_mss(sk, dst_mtu(dst)); 33381da177e4SLinus Torvalds 333981164413SDaniel Borkmann tcp_ca_dst_init(sk, dst); 334081164413SDaniel Borkmann 33411da177e4SLinus Torvalds if (!tp->window_clamp) 33421da177e4SLinus Torvalds tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 33433541f9e8SEric Dumazet tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 3344f5fff5dcSTom Quetchenbach 33451da177e4SLinus Torvalds tcp_initialize_rcv_mss(sk); 33461da177e4SLinus Torvalds 3347e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 3348e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 3349e88c64f0SHagen Paul Pfeifer (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) 3350e88c64f0SHagen Paul Pfeifer tp->window_clamp = tcp_full_space(sk); 3351e88c64f0SHagen Paul Pfeifer 335213d3b1ebSLawrence Brakmo rcv_wnd = tcp_rwnd_init_bpf(sk); 335313d3b1ebSLawrence Brakmo if (rcv_wnd == 0) 335413d3b1ebSLawrence Brakmo rcv_wnd = dst_metric(dst, RTAX_INITRWND); 335513d3b1ebSLawrence Brakmo 3356ceef9ab6SEric Dumazet tcp_select_initial_window(sk, tcp_full_space(sk), 33571da177e4SLinus Torvalds tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 33581da177e4SLinus Torvalds &tp->rcv_wnd, 33591da177e4SLinus Torvalds &tp->window_clamp, 33609bb37ef0SEric Dumazet sock_net(sk)->ipv4.sysctl_tcp_window_scaling, 336131d12926Slaurent chavey &rcv_wscale, 336213d3b1ebSLawrence Brakmo rcv_wnd); 33631da177e4SLinus Torvalds 33641da177e4SLinus Torvalds tp->rx_opt.rcv_wscale = rcv_wscale; 33651da177e4SLinus Torvalds tp->rcv_ssthresh = tp->rcv_wnd; 33661da177e4SLinus Torvalds 33671da177e4SLinus Torvalds sk->sk_err = 0; 33681da177e4SLinus Torvalds sock_reset_flag(sk, SOCK_DONE); 33691da177e4SLinus Torvalds tp->snd_wnd = 0; 3370ee7537b6SHantzis Fotis tcp_init_wl(tp, 0); 33711da177e4SLinus Torvalds tp->snd_una = tp->write_seq; 33721da177e4SLinus Torvalds tp->snd_sml = tp->write_seq; 337333f5f57eSIlpo Järvinen tp->snd_up = tp->write_seq; 3374370816aeSPavel Emelyanov tp->snd_nxt = tp->write_seq; 3375ee995283SPavel Emelyanov 3376ee995283SPavel Emelyanov if (likely(!tp->repair)) 33771da177e4SLinus Torvalds tp->rcv_nxt = 0; 3378c7781a6eSAndrew Vagin else 337970eabf0eSEric Dumazet tp->rcv_tstamp = tcp_jiffies32; 3380ee995283SPavel Emelyanov tp->rcv_wup = tp->rcv_nxt; 3381ee995283SPavel Emelyanov tp->copied_seq = tp->rcv_nxt; 33821da177e4SLinus Torvalds 33838550f328SLawrence Brakmo inet_csk(sk)->icsk_rto = tcp_timeout_init(sk); 3384463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_retransmits = 0; 33851da177e4SLinus Torvalds tcp_clear_retrans(tp); 33861da177e4SLinus Torvalds } 33871da177e4SLinus Torvalds 3388783237e8SYuchung Cheng static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) 3389783237e8SYuchung Cheng { 3390783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3391783237e8SYuchung Cheng struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 3392783237e8SYuchung Cheng 3393783237e8SYuchung Cheng tcb->end_seq += skb->len; 3394f4a775d1SEric Dumazet __skb_header_release(skb); 3395783237e8SYuchung Cheng sk->sk_wmem_queued += skb->truesize; 3396783237e8SYuchung Cheng sk_mem_charge(sk, skb->truesize); 3397783237e8SYuchung Cheng tp->write_seq = tcb->end_seq; 3398783237e8SYuchung Cheng tp->packets_out += tcp_skb_pcount(skb); 3399783237e8SYuchung Cheng } 3400783237e8SYuchung Cheng 3401783237e8SYuchung Cheng /* Build and send a SYN with data and (cached) Fast Open cookie. However, 3402783237e8SYuchung Cheng * queue a data-only packet after the regular SYN, such that regular SYNs 3403783237e8SYuchung Cheng * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges 3404783237e8SYuchung Cheng * only the SYN sequence, the data are retransmitted in the first ACK. 3405783237e8SYuchung Cheng * If cookie is not cached or other error occurs, falls back to send a 3406783237e8SYuchung Cheng * regular SYN with Fast Open cookie request option. 3407783237e8SYuchung Cheng */ 3408783237e8SYuchung Cheng static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) 3409783237e8SYuchung Cheng { 3410783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3411783237e8SYuchung Cheng struct tcp_fastopen_request *fo = tp->fastopen_req; 3412065263f4SWei Wang int space, err = 0; 3413355a901eSEric Dumazet struct sk_buff *syn_data; 3414783237e8SYuchung Cheng 341567da22d2SYuchung Cheng tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ 3416065263f4SWei Wang if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie)) 3417783237e8SYuchung Cheng goto fallback; 3418783237e8SYuchung Cheng 3419783237e8SYuchung Cheng /* MSS for SYN-data is based on cached MSS and bounded by PMTU and 3420783237e8SYuchung Cheng * user-MSS. Reserve maximum option space for middleboxes that add 3421783237e8SYuchung Cheng * private TCP options. The cost is reduced data space in SYN :( 3422783237e8SYuchung Cheng */ 34233541f9e8SEric Dumazet tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp); 34243541f9e8SEric Dumazet 34251b63edd6SYuchung Cheng space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) - 3426783237e8SYuchung Cheng MAX_TCP_OPTION_SPACE; 3427783237e8SYuchung Cheng 3428f5ddcbbbSEric Dumazet space = min_t(size_t, space, fo->size); 3429f5ddcbbbSEric Dumazet 3430f5ddcbbbSEric Dumazet /* limit to order-0 allocations */ 3431f5ddcbbbSEric Dumazet space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER)); 3432f5ddcbbbSEric Dumazet 3433eb934478SEric Dumazet syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false); 3434355a901eSEric Dumazet if (!syn_data) 3435783237e8SYuchung Cheng goto fallback; 3436355a901eSEric Dumazet syn_data->ip_summed = CHECKSUM_PARTIAL; 3437355a901eSEric Dumazet memcpy(syn_data->cb, syn->cb, sizeof(syn->cb)); 343807e100f9SEric Dumazet if (space) { 343907e100f9SEric Dumazet int copied = copy_from_iter(skb_put(syn_data, space), space, 344057be5bdaSAl Viro &fo->data->msg_iter); 344157be5bdaSAl Viro if (unlikely(!copied)) { 3442ba233b34SEric Dumazet tcp_skb_tsorted_anchor_cleanup(syn_data); 3443355a901eSEric Dumazet kfree_skb(syn_data); 3444783237e8SYuchung Cheng goto fallback; 3445783237e8SYuchung Cheng } 344657be5bdaSAl Viro if (copied != space) { 344757be5bdaSAl Viro skb_trim(syn_data, copied); 344857be5bdaSAl Viro space = copied; 344957be5bdaSAl Viro } 345007e100f9SEric Dumazet } 3451355a901eSEric Dumazet /* No more data pending in inet_wait_for_connect() */ 3452355a901eSEric Dumazet if (space == fo->size) 3453355a901eSEric Dumazet fo->data = NULL; 3454355a901eSEric Dumazet fo->copied = space; 3455783237e8SYuchung Cheng 3456355a901eSEric Dumazet tcp_connect_queue_skb(sk, syn_data); 34570f87230dSFrancis Yan if (syn_data->len) 34580f87230dSFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_BUSY); 3459355a901eSEric Dumazet 3460355a901eSEric Dumazet err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation); 3461355a901eSEric Dumazet 3462355a901eSEric Dumazet syn->skb_mstamp = syn_data->skb_mstamp; 3463355a901eSEric Dumazet 3464355a901eSEric Dumazet /* Now full SYN+DATA was cloned and sent (or not), 3465355a901eSEric Dumazet * remove the SYN from the original skb (syn_data) 3466355a901eSEric Dumazet * we keep in write queue in case of a retransmit, as we 3467355a901eSEric Dumazet * also have the SYN packet (with no data) in the same queue. 3468431a9124SEric Dumazet */ 3469355a901eSEric Dumazet TCP_SKB_CB(syn_data)->seq++; 3470355a901eSEric Dumazet TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH; 3471355a901eSEric Dumazet if (!err) { 347267da22d2SYuchung Cheng tp->syn_data = (fo->copied > 0); 347375c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data); 3474f19c29e3SYuchung Cheng NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT); 3475783237e8SYuchung Cheng goto done; 3476783237e8SYuchung Cheng } 3477783237e8SYuchung Cheng 347875c119afSEric Dumazet /* data was not sent, put it in write_queue */ 347975c119afSEric Dumazet __skb_queue_tail(&sk->sk_write_queue, syn_data); 3480b5b7db8dSEric Dumazet tp->packets_out -= tcp_skb_pcount(syn_data); 3481b5b7db8dSEric Dumazet 3482783237e8SYuchung Cheng fallback: 3483783237e8SYuchung Cheng /* Send a regular SYN with Fast Open cookie request option */ 3484783237e8SYuchung Cheng if (fo->cookie.len > 0) 3485783237e8SYuchung Cheng fo->cookie.len = 0; 3486783237e8SYuchung Cheng err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); 3487783237e8SYuchung Cheng if (err) 3488783237e8SYuchung Cheng tp->syn_fastopen = 0; 3489783237e8SYuchung Cheng done: 3490783237e8SYuchung Cheng fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ 3491783237e8SYuchung Cheng return err; 3492783237e8SYuchung Cheng } 3493783237e8SYuchung Cheng 349467edfef7SAndi Kleen /* Build a SYN and send it off. */ 34951da177e4SLinus Torvalds int tcp_connect(struct sock *sk) 34961da177e4SLinus Torvalds { 34971da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 34981da177e4SLinus Torvalds struct sk_buff *buff; 3499ee586811SEric Paris int err; 35001da177e4SLinus Torvalds 3501de525be2SLawrence Brakmo tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB, 0, NULL); 35028ba60924SEric Dumazet 35038ba60924SEric Dumazet if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 35048ba60924SEric Dumazet return -EHOSTUNREACH; /* Routing failure or similar. */ 35058ba60924SEric Dumazet 35061da177e4SLinus Torvalds tcp_connect_init(sk); 35071da177e4SLinus Torvalds 35082b916477SAndrey Vagin if (unlikely(tp->repair)) { 35092b916477SAndrey Vagin tcp_finish_connect(sk, NULL); 35102b916477SAndrey Vagin return 0; 35112b916477SAndrey Vagin } 35122b916477SAndrey Vagin 3513eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true); 3514355a901eSEric Dumazet if (unlikely(!buff)) 35151da177e4SLinus Torvalds return -ENOBUFS; 35161da177e4SLinus Torvalds 3517a3433f35SChangli Gao tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); 35189a568de4SEric Dumazet tcp_mstamp_refresh(tp); 35199a568de4SEric Dumazet tp->retrans_stamp = tcp_time_stamp(tp); 3520783237e8SYuchung Cheng tcp_connect_queue_skb(sk, buff); 3521735d3831SFlorian Westphal tcp_ecn_send_syn(sk, buff); 352275c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, buff); 35231da177e4SLinus Torvalds 3524783237e8SYuchung Cheng /* Send off SYN; include data in Fast Open. */ 3525783237e8SYuchung Cheng err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : 3526783237e8SYuchung Cheng tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); 3527ee586811SEric Paris if (err == -ECONNREFUSED) 3528ee586811SEric Paris return err; 3529bd37a088SWei Yongjun 3530bd37a088SWei Yongjun /* We change tp->snd_nxt after the tcp_transmit_skb() call 3531bd37a088SWei Yongjun * in order to make this packet get counted in tcpOutSegs. 3532bd37a088SWei Yongjun */ 3533bd37a088SWei Yongjun tp->snd_nxt = tp->write_seq; 3534bd37a088SWei Yongjun tp->pushed_seq = tp->write_seq; 3535b5b7db8dSEric Dumazet buff = tcp_send_head(sk); 3536b5b7db8dSEric Dumazet if (unlikely(buff)) { 3537b5b7db8dSEric Dumazet tp->snd_nxt = TCP_SKB_CB(buff)->seq; 3538b5b7db8dSEric Dumazet tp->pushed_seq = TCP_SKB_CB(buff)->seq; 3539b5b7db8dSEric Dumazet } 354081cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); 35411da177e4SLinus Torvalds 35421da177e4SLinus Torvalds /* Timer for repeating the SYN until an answer. */ 35433f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 35443f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 35451da177e4SLinus Torvalds return 0; 35461da177e4SLinus Torvalds } 35474bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_connect); 35481da177e4SLinus Torvalds 35491da177e4SLinus Torvalds /* Send out a delayed ack, the caller does the policy checking 35501da177e4SLinus Torvalds * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 35511da177e4SLinus Torvalds * for details. 35521da177e4SLinus Torvalds */ 35531da177e4SLinus Torvalds void tcp_send_delayed_ack(struct sock *sk) 35541da177e4SLinus Torvalds { 3555463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 3556463c84b9SArnaldo Carvalho de Melo int ato = icsk->icsk_ack.ato; 35571da177e4SLinus Torvalds unsigned long timeout; 35581da177e4SLinus Torvalds 35599890092eSFlorian Westphal tcp_ca_event(sk, CA_EVENT_DELAYED_ACK); 35609890092eSFlorian Westphal 35611da177e4SLinus Torvalds if (ato > TCP_DELACK_MIN) { 3562463c84b9SArnaldo Carvalho de Melo const struct tcp_sock *tp = tcp_sk(sk); 35631da177e4SLinus Torvalds int max_ato = HZ / 2; 35641da177e4SLinus Torvalds 3565056834d9SIlpo Järvinen if (icsk->icsk_ack.pingpong || 3566056834d9SIlpo Järvinen (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 35671da177e4SLinus Torvalds max_ato = TCP_DELACK_MAX; 35681da177e4SLinus Torvalds 35691da177e4SLinus Torvalds /* Slow path, intersegment interval is "high". */ 35701da177e4SLinus Torvalds 35711da177e4SLinus Torvalds /* If some rtt estimate is known, use it to bound delayed ack. 3572463c84b9SArnaldo Carvalho de Melo * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 35731da177e4SLinus Torvalds * directly. 35741da177e4SLinus Torvalds */ 3575740b0f18SEric Dumazet if (tp->srtt_us) { 3576740b0f18SEric Dumazet int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3), 3577740b0f18SEric Dumazet TCP_DELACK_MIN); 35781da177e4SLinus Torvalds 35791da177e4SLinus Torvalds if (rtt < max_ato) 35801da177e4SLinus Torvalds max_ato = rtt; 35811da177e4SLinus Torvalds } 35821da177e4SLinus Torvalds 35831da177e4SLinus Torvalds ato = min(ato, max_ato); 35841da177e4SLinus Torvalds } 35851da177e4SLinus Torvalds 35861da177e4SLinus Torvalds /* Stay within the limit we were given */ 35871da177e4SLinus Torvalds timeout = jiffies + ato; 35881da177e4SLinus Torvalds 35891da177e4SLinus Torvalds /* Use new timeout only if there wasn't a older one earlier. */ 3590463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 35911da177e4SLinus Torvalds /* If delack timer was blocked or is about to expire, 35921da177e4SLinus Torvalds * send ACK now. 35931da177e4SLinus Torvalds */ 3594463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.blocked || 3595463c84b9SArnaldo Carvalho de Melo time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 35961da177e4SLinus Torvalds tcp_send_ack(sk); 35971da177e4SLinus Torvalds return; 35981da177e4SLinus Torvalds } 35991da177e4SLinus Torvalds 3600463c84b9SArnaldo Carvalho de Melo if (!time_before(timeout, icsk->icsk_ack.timeout)) 3601463c84b9SArnaldo Carvalho de Melo timeout = icsk->icsk_ack.timeout; 36021da177e4SLinus Torvalds } 3603463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 3604463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.timeout = timeout; 3605463c84b9SArnaldo Carvalho de Melo sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 36061da177e4SLinus Torvalds } 36071da177e4SLinus Torvalds 36081da177e4SLinus Torvalds /* This routine sends an ack and also updates the window. */ 36091da177e4SLinus Torvalds void tcp_send_ack(struct sock *sk) 36101da177e4SLinus Torvalds { 36111da177e4SLinus Torvalds struct sk_buff *buff; 36121da177e4SLinus Torvalds 3613058dc334SIlpo Järvinen /* If we have been reset, we may not send again. */ 3614058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3615058dc334SIlpo Järvinen return; 3616058dc334SIlpo Järvinen 36179890092eSFlorian Westphal tcp_ca_event(sk, CA_EVENT_NON_DELAYED_ACK); 36189890092eSFlorian Westphal 36191da177e4SLinus Torvalds /* We are not putting this on the write queue, so 36201da177e4SLinus Torvalds * tcp_transmit_skb() will set the ownership to this 36211da177e4SLinus Torvalds * sock. 36221da177e4SLinus Torvalds */ 36237450aaf6SEric Dumazet buff = alloc_skb(MAX_TCP_HEADER, 36247450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 36257450aaf6SEric Dumazet if (unlikely(!buff)) { 3626463c84b9SArnaldo Carvalho de Melo inet_csk_schedule_ack(sk); 3627463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; 36283f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 36293f421baaSArnaldo Carvalho de Melo TCP_DELACK_MAX, TCP_RTO_MAX); 36301da177e4SLinus Torvalds return; 36311da177e4SLinus Torvalds } 36321da177e4SLinus Torvalds 36331da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 36341da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 3635a3433f35SChangli Gao tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); 36361da177e4SLinus Torvalds 363798781965SEric Dumazet /* We do not want pure acks influencing TCP Small Queues or fq/pacing 363898781965SEric Dumazet * too much. 363998781965SEric Dumazet * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784 364098781965SEric Dumazet */ 364198781965SEric Dumazet skb_set_tcp_pure_ack(buff); 364298781965SEric Dumazet 36431da177e4SLinus Torvalds /* Send it off, this clears delayed acks for us. */ 36447450aaf6SEric Dumazet tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0); 36451da177e4SLinus Torvalds } 3646e3118e83SDaniel Borkmann EXPORT_SYMBOL_GPL(tcp_send_ack); 36471da177e4SLinus Torvalds 36481da177e4SLinus Torvalds /* This routine sends a packet with an out of date sequence 36491da177e4SLinus Torvalds * number. It assumes the other end will try to ack it. 36501da177e4SLinus Torvalds * 36511da177e4SLinus Torvalds * Question: what should we make while urgent mode? 36521da177e4SLinus Torvalds * 4.4BSD forces sending single byte of data. We cannot send 36531da177e4SLinus Torvalds * out of window data, because we have SND.NXT==SND.MAX... 36541da177e4SLinus Torvalds * 36551da177e4SLinus Torvalds * Current solution: to send TWO zero-length segments in urgent mode: 36561da177e4SLinus Torvalds * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 36571da177e4SLinus Torvalds * out-of-date with SND.UNA-1 to probe window. 36581da177e4SLinus Torvalds */ 3659e520af48SEric Dumazet static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib) 36601da177e4SLinus Torvalds { 36611da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 36621da177e4SLinus Torvalds struct sk_buff *skb; 36631da177e4SLinus Torvalds 36641da177e4SLinus Torvalds /* We don't queue it, tcp_transmit_skb() sets ownership. */ 36657450aaf6SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, 36667450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 366751456b29SIan Morris if (!skb) 36681da177e4SLinus Torvalds return -1; 36691da177e4SLinus Torvalds 36701da177e4SLinus Torvalds /* Reserve space for headers and set control bits. */ 36711da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 36721da177e4SLinus Torvalds /* Use a previous sequence. This should cause the other 36731da177e4SLinus Torvalds * end to send an ack. Don't queue or clone SKB, just 36741da177e4SLinus Torvalds * send it. 36751da177e4SLinus Torvalds */ 3676a3433f35SChangli Gao tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); 3677e2e8009fSRenato Westphal NET_INC_STATS(sock_net(sk), mib); 36787450aaf6SEric Dumazet return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0); 36791da177e4SLinus Torvalds } 36801da177e4SLinus Torvalds 3681385e2070SEric Dumazet /* Called from setsockopt( ... TCP_REPAIR ) */ 3682ee995283SPavel Emelyanov void tcp_send_window_probe(struct sock *sk) 3683ee995283SPavel Emelyanov { 3684ee995283SPavel Emelyanov if (sk->sk_state == TCP_ESTABLISHED) { 3685ee995283SPavel Emelyanov tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; 36869a568de4SEric Dumazet tcp_mstamp_refresh(tcp_sk(sk)); 3687e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE); 3688ee995283SPavel Emelyanov } 3689ee995283SPavel Emelyanov } 3690ee995283SPavel Emelyanov 369167edfef7SAndi Kleen /* Initiate keepalive or window probe from timer. */ 3692e520af48SEric Dumazet int tcp_write_wakeup(struct sock *sk, int mib) 36931da177e4SLinus Torvalds { 36941da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 36951da177e4SLinus Torvalds struct sk_buff *skb; 36961da177e4SLinus Torvalds 3697058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3698058dc334SIlpo Järvinen return -1; 3699058dc334SIlpo Järvinen 370000db4124SIan Morris skb = tcp_send_head(sk); 370100db4124SIan Morris if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { 37021da177e4SLinus Torvalds int err; 37030c54b85fSIlpo Järvinen unsigned int mss = tcp_current_mss(sk); 370490840defSIlpo Järvinen unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 37051da177e4SLinus Torvalds 37061da177e4SLinus Torvalds if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 37071da177e4SLinus Torvalds tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 37081da177e4SLinus Torvalds 37091da177e4SLinus Torvalds /* We are probing the opening of a window 37101da177e4SLinus Torvalds * but the window size is != 0 37111da177e4SLinus Torvalds * must have been a result SWS avoidance ( sender ) 37121da177e4SLinus Torvalds */ 37131da177e4SLinus Torvalds if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 37141da177e4SLinus Torvalds skb->len > mss) { 37151da177e4SLinus Torvalds seg_size = min(seg_size, mss); 37164de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 371775c119afSEric Dumazet if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE, 371875c119afSEric Dumazet skb, seg_size, mss, GFP_ATOMIC)) 37191da177e4SLinus Torvalds return -1; 37201da177e4SLinus Torvalds } else if (!tcp_skb_pcount(skb)) 37215bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss); 37221da177e4SLinus Torvalds 37234de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 3724dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 372566f5fe62SIlpo Järvinen if (!err) 372666f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 37271da177e4SLinus Torvalds return err; 37281da177e4SLinus Torvalds } else { 372933f5f57eSIlpo Järvinen if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) 3730e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 1, mib); 3731e520af48SEric Dumazet return tcp_xmit_probe_skb(sk, 0, mib); 37321da177e4SLinus Torvalds } 37331da177e4SLinus Torvalds } 37341da177e4SLinus Torvalds 37351da177e4SLinus Torvalds /* A window probe timeout has occurred. If window is not closed send 37361da177e4SLinus Torvalds * a partial packet else a zero probe. 37371da177e4SLinus Torvalds */ 37381da177e4SLinus Torvalds void tcp_send_probe0(struct sock *sk) 37391da177e4SLinus Torvalds { 3740463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 37411da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 3742c6214a97SNikolay Borisov struct net *net = sock_net(sk); 3743fcdd1cf4SEric Dumazet unsigned long probe_max; 37441da177e4SLinus Torvalds int err; 37451da177e4SLinus Torvalds 3746e520af48SEric Dumazet err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE); 37471da177e4SLinus Torvalds 374875c119afSEric Dumazet if (tp->packets_out || tcp_write_queue_empty(sk)) { 37491da177e4SLinus Torvalds /* Cancel probe timer, if it is not required. */ 37506687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 0; 3751463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff = 0; 37521da177e4SLinus Torvalds return; 37531da177e4SLinus Torvalds } 37541da177e4SLinus Torvalds 37551da177e4SLinus Torvalds if (err <= 0) { 3756c6214a97SNikolay Borisov if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2) 3757463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff++; 37586687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out++; 3759fcdd1cf4SEric Dumazet probe_max = TCP_RTO_MAX; 37601da177e4SLinus Torvalds } else { 37611da177e4SLinus Torvalds /* If packet was not sent due to local congestion, 37626687e988SArnaldo Carvalho de Melo * do not backoff and do not remember icsk_probes_out. 37631da177e4SLinus Torvalds * Let local senders to fight for local resources. 37641da177e4SLinus Torvalds * 37651da177e4SLinus Torvalds * Use accumulated backoff yet. 37661da177e4SLinus Torvalds */ 37676687e988SArnaldo Carvalho de Melo if (!icsk->icsk_probes_out) 37686687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 1; 3769fcdd1cf4SEric Dumazet probe_max = TCP_RESOURCE_PROBE_INTERVAL; 37701da177e4SLinus Torvalds } 3771fcdd1cf4SEric Dumazet inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 377221c8fe99SEric Dumazet tcp_probe0_when(sk, probe_max), 3773fcdd1cf4SEric Dumazet TCP_RTO_MAX); 37741da177e4SLinus Torvalds } 37755db92c99SOctavian Purdila 3776ea3bea3aSEric Dumazet int tcp_rtx_synack(const struct sock *sk, struct request_sock *req) 37775db92c99SOctavian Purdila { 37785db92c99SOctavian Purdila const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific; 37795db92c99SOctavian Purdila struct flowi fl; 37805db92c99SOctavian Purdila int res; 37815db92c99SOctavian Purdila 378258d607d3SEric Dumazet tcp_rsk(req)->txhash = net_tx_rndhash(); 3783b3d05147SEric Dumazet res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL); 37845db92c99SOctavian Purdila if (!res) { 378590bbcc60SEric Dumazet __TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); 378602a1d6e7SEric Dumazet __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 37877e32b443SYuchung Cheng if (unlikely(tcp_passive_fastopen(sk))) 37887e32b443SYuchung Cheng tcp_sk(sk)->total_retrans++; 3789cf34ce3dSSong Liu trace_tcp_retransmit_synack(sk, req); 37905db92c99SOctavian Purdila } 37915db92c99SOctavian Purdila return res; 37925db92c99SOctavian Purdila } 37935db92c99SOctavian Purdila EXPORT_SYMBOL(tcp_rtx_synack); 3794