1457c8996SThomas Gleixner // SPDX-License-Identifier: GPL-2.0-only 21da177e4SLinus Torvalds /* 31da177e4SLinus Torvalds * INET An implementation of the TCP/IP protocol suite for the LINUX 41da177e4SLinus Torvalds * operating system. INET is implemented using the BSD Socket 51da177e4SLinus Torvalds * interface as the means of communication with the user level. 61da177e4SLinus Torvalds * 71da177e4SLinus Torvalds * Implementation of the Transmission Control Protocol(TCP). 81da177e4SLinus Torvalds * 902c30a84SJesper Juhl * Authors: Ross Biro 101da177e4SLinus Torvalds * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> 111da177e4SLinus Torvalds * Mark Evans, <evansmp@uhura.aston.ac.uk> 121da177e4SLinus Torvalds * Corey Minyard <wf-rch!minyard@relay.EU.net> 131da177e4SLinus Torvalds * Florian La Roche, <flla@stud.uni-sb.de> 141da177e4SLinus Torvalds * Charles Hedrick, <hedrick@klinzhai.rutgers.edu> 151da177e4SLinus Torvalds * Linus Torvalds, <torvalds@cs.helsinki.fi> 161da177e4SLinus Torvalds * Alan Cox, <gw4pts@gw4pts.ampr.org> 171da177e4SLinus Torvalds * Matthew Dillon, <dillon@apollo.west.oic.com> 181da177e4SLinus Torvalds * Arnt Gulbrandsen, <agulbra@nvg.unit.no> 191da177e4SLinus Torvalds * Jorge Cwik, <jorge@laser.satlink.net> 201da177e4SLinus Torvalds */ 211da177e4SLinus Torvalds 221da177e4SLinus Torvalds /* 231da177e4SLinus Torvalds * Changes: Pedro Roque : Retransmit queue handled by TCP. 241da177e4SLinus Torvalds * : Fragmentation on mtu decrease 251da177e4SLinus Torvalds * : Segment collapse on retransmit 261da177e4SLinus Torvalds * : AF independence 271da177e4SLinus Torvalds * 281da177e4SLinus Torvalds * Linus Torvalds : send_delayed_ack 291da177e4SLinus Torvalds * David S. Miller : Charge memory using the right skb 301da177e4SLinus Torvalds * during syn/ack processing. 311da177e4SLinus Torvalds * David S. Miller : Output engine completely rewritten. 321da177e4SLinus Torvalds * Andrea Arcangeli: SYNACK carry ts_recent in tsecr. 331da177e4SLinus Torvalds * Cacophonix Gaul : draft-minshall-nagle-01 341da177e4SLinus Torvalds * J Hadi Salim : ECN support 351da177e4SLinus Torvalds * 361da177e4SLinus Torvalds */ 371da177e4SLinus Torvalds 3891df42beSJoe Perches #define pr_fmt(fmt) "TCP: " fmt 3991df42beSJoe Perches 401da177e4SLinus Torvalds #include <net/tcp.h> 41eda7acddSPeter Krystad #include <net/mptcp.h> 421da177e4SLinus Torvalds 431da177e4SLinus Torvalds #include <linux/compiler.h> 445a0e3ad6STejun Heo #include <linux/gfp.h> 451da177e4SLinus Torvalds #include <linux/module.h> 4660e2a778SUrsula Braun #include <linux/static_key.h> 47f6d827b1SMina Almasry #include <linux/skbuff_ref.h> 481da177e4SLinus Torvalds 49e086101bSCong Wang #include <trace/events/tcp.h> 5035089bb2SDavid S. Miller 519799ccb0SEric Dumazet /* Refresh clocks of a TCP socket, 529799ccb0SEric Dumazet * ensuring monotically increasing values. 539799ccb0SEric Dumazet */ 549799ccb0SEric Dumazet void tcp_mstamp_refresh(struct tcp_sock *tp) 559799ccb0SEric Dumazet { 569799ccb0SEric Dumazet u64 val = tcp_clock_ns(); 579799ccb0SEric Dumazet 585f6188a8SEric Dumazet tp->tcp_clock_cache = val; 59e6d14070SEric Dumazet tp->tcp_mstamp = div_u64(val, NSEC_PER_USEC); 609799ccb0SEric Dumazet } 619799ccb0SEric Dumazet 6246d3ceabSEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 6346d3ceabSEric Dumazet int push_one, gfp_t gfp); 64519855c5SWilliam Allen Simpson 6567edfef7SAndi Kleen /* Account for new data that has been sent to the network. */ 6675c119afSEric Dumazet static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb) 676ff03ac3SIlpo Järvinen { 686ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 696ff03ac3SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 7066f5fe62SIlpo Järvinen unsigned int prior_packets = tp->packets_out; 719e412ba7SIlpo Järvinen 72e0d694d6SEric Dumazet WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(skb)->end_seq); 738512430eSIlpo Järvinen 7475c119afSEric Dumazet __skb_unlink(skb, &sk->sk_write_queue); 7575c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, skb); 7675c119afSEric Dumazet 7785369750SCambda Zhu if (tp->highest_sack == NULL) 7885369750SCambda Zhu tp->highest_sack = skb; 7985369750SCambda Zhu 8066f5fe62SIlpo Järvinen tp->packets_out += tcp_skb_pcount(skb); 81bec41a11SYuchung Cheng if (!prior_packets || icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) 82750ea2baSYuchung Cheng tcp_rearm_rto(sk); 83f19c29e3SYuchung Cheng 84f7324acdSDavid S. Miller NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT, 85f19c29e3SYuchung Cheng tcp_skb_pcount(skb)); 864bfe744fSEric Dumazet tcp_check_space(sk); 876a5dc9e5SEric Dumazet } 881da177e4SLinus Torvalds 89a4ecb15aSCui, Cheng /* SND.NXT, if window was not shrunk or the amount of shrunk was less than one 90a4ecb15aSCui, Cheng * window scaling factor due to loss of precision. 911da177e4SLinus Torvalds * If window has been shrunk, what should we make? It is not clear at all. 921da177e4SLinus Torvalds * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-( 931da177e4SLinus Torvalds * Anything in between SND.UNA...SND.UNA+SND.WND also can be already 941da177e4SLinus Torvalds * invalid. OK, let's make this for now: 951da177e4SLinus Torvalds */ 96cf533ea5SEric Dumazet static inline __u32 tcp_acceptable_seq(const struct sock *sk) 971da177e4SLinus Torvalds { 98cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 999e412ba7SIlpo Järvinen 100a4ecb15aSCui, Cheng if (!before(tcp_wnd_end(tp), tp->snd_nxt) || 101a4ecb15aSCui, Cheng (tp->rx_opt.wscale_ok && 102a4ecb15aSCui, Cheng ((tp->snd_nxt - tcp_wnd_end(tp)) < (1 << tp->rx_opt.rcv_wscale)))) 1031da177e4SLinus Torvalds return tp->snd_nxt; 1041da177e4SLinus Torvalds else 10590840defSIlpo Järvinen return tcp_wnd_end(tp); 1061da177e4SLinus Torvalds } 1071da177e4SLinus Torvalds 1081da177e4SLinus Torvalds /* Calculate mss to advertise in SYN segment. 1091da177e4SLinus Torvalds * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that: 1101da177e4SLinus Torvalds * 1111da177e4SLinus Torvalds * 1. It is independent of path mtu. 1121da177e4SLinus Torvalds * 2. Ideally, it is maximal possible segment size i.e. 65535-40. 1131da177e4SLinus Torvalds * 3. For IPv4 it is reasonable to calculate it from maximal MTU of 1141da177e4SLinus Torvalds * attached devices, because some buggy hosts are confused by 1151da177e4SLinus Torvalds * large MSS. 1161da177e4SLinus Torvalds * 4. We do not make 3, we advertise MSS, calculated from first 1171da177e4SLinus Torvalds * hop device mtu, but allow to raise it to ip_rt_min_advmss. 1181da177e4SLinus Torvalds * This may be overridden via information stored in routing table. 1191da177e4SLinus Torvalds * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible, 1201da177e4SLinus Torvalds * probably even Jumbo". 1211da177e4SLinus Torvalds */ 1221da177e4SLinus Torvalds static __u16 tcp_advertise_mss(struct sock *sk) 1231da177e4SLinus Torvalds { 1241da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 125cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1261da177e4SLinus Torvalds int mss = tp->advmss; 1271da177e4SLinus Torvalds 1280dbaee3bSDavid S. Miller if (dst) { 1290dbaee3bSDavid S. Miller unsigned int metric = dst_metric_advmss(dst); 1300dbaee3bSDavid S. Miller 1310dbaee3bSDavid S. Miller if (metric < mss) { 1320dbaee3bSDavid S. Miller mss = metric; 1331da177e4SLinus Torvalds tp->advmss = mss; 1341da177e4SLinus Torvalds } 1350dbaee3bSDavid S. Miller } 1361da177e4SLinus Torvalds 1371da177e4SLinus Torvalds return (__u16)mss; 1381da177e4SLinus Torvalds } 1391da177e4SLinus Torvalds 1401da177e4SLinus Torvalds /* RFC2861. Reset CWND after idle period longer RTO to "restart window". 1416f021c62SEric Dumazet * This is the first part of cwnd validation mechanism. 1426f021c62SEric Dumazet */ 1436f021c62SEric Dumazet void tcp_cwnd_restart(struct sock *sk, s32 delta) 1441da177e4SLinus Torvalds { 145463c84b9SArnaldo Carvalho de Melo struct tcp_sock *tp = tcp_sk(sk); 1466f021c62SEric Dumazet u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk)); 14740570375SEric Dumazet u32 cwnd = tcp_snd_cwnd(tp); 1481da177e4SLinus Torvalds 1496687e988SArnaldo Carvalho de Melo tcp_ca_event(sk, CA_EVENT_CWND_RESTART); 1501da177e4SLinus Torvalds 1516687e988SArnaldo Carvalho de Melo tp->snd_ssthresh = tcp_current_ssthresh(sk); 1521da177e4SLinus Torvalds restart_cwnd = min(restart_cwnd, cwnd); 1531da177e4SLinus Torvalds 154463c84b9SArnaldo Carvalho de Melo while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd) 1551da177e4SLinus Torvalds cwnd >>= 1; 15640570375SEric Dumazet tcp_snd_cwnd_set(tp, max(cwnd, restart_cwnd)); 157c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 1581da177e4SLinus Torvalds tp->snd_cwnd_used = 0; 1591da177e4SLinus Torvalds } 1601da177e4SLinus Torvalds 16167edfef7SAndi Kleen /* Congestion state accounting after a packet has been sent. */ 16240efc6faSStephen Hemminger static void tcp_event_data_sent(struct tcp_sock *tp, 163cf533ea5SEric Dumazet struct sock *sk) 1641da177e4SLinus Torvalds { 165463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 166d635fbe2SEric Dumazet const u32 now = tcp_jiffies32; 1671da177e4SLinus Torvalds 16805c5a46dSNeal Cardwell if (tcp_packets_in_flight(tp) == 0) 16905c5a46dSNeal Cardwell tcp_ca_event(sk, CA_EVENT_TX_START); 17005c5a46dSNeal Cardwell 1714a41f453SWei Wang tp->lsndtime = now; 1724d8f24eeSWei Wang 1734d8f24eeSWei Wang /* If it is a reply for ato after last received 174562b1fdfSHaiyang Zhang * packet, increase pingpong count. 1754d8f24eeSWei Wang */ 1764d8f24eeSWei Wang if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato) 177562b1fdfSHaiyang Zhang inet_csk_inc_pingpong_cnt(sk); 1781da177e4SLinus Torvalds } 1791da177e4SLinus Torvalds 18067edfef7SAndi Kleen /* Account for an ACK we sent. */ 181059217c1SNeal Cardwell static inline void tcp_event_ack_sent(struct sock *sk, u32 rcv_nxt) 1821da177e4SLinus Torvalds { 1835d9f4262SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 1845d9f4262SEric Dumazet 1852b195850SEric Dumazet if (unlikely(tp->compressed_ack)) { 186200d95f4SEric Dumazet NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED, 1872b195850SEric Dumazet tp->compressed_ack); 1882b195850SEric Dumazet tp->compressed_ack = 0; 1895d9f4262SEric Dumazet if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1) 1905d9f4262SEric Dumazet __sock_put(sk); 1915d9f4262SEric Dumazet } 19227cde44aSYuchung Cheng 19327cde44aSYuchung Cheng if (unlikely(rcv_nxt != tp->rcv_nxt)) 19427cde44aSYuchung Cheng return; /* Special ACK sent by DCTCP to reflect ECN */ 195059217c1SNeal Cardwell tcp_dec_quickack_mode(sk); 196463c84b9SArnaldo Carvalho de Melo inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); 1971da177e4SLinus Torvalds } 1981da177e4SLinus Torvalds 1991da177e4SLinus Torvalds /* Determine a window scaling and initial window to offer. 2001da177e4SLinus Torvalds * Based on the assumption that the given amount of space 2011da177e4SLinus Torvalds * will be offered. Store the results in the tp structure. 2021da177e4SLinus Torvalds * NOTE: for smooth operation initial space offering should 2031da177e4SLinus Torvalds * be a multiple of mss if possible. We assume here that mss >= 1. 2041da177e4SLinus Torvalds * This MUST be enforced by all callers. 2051da177e4SLinus Torvalds */ 206ceef9ab6SEric Dumazet void tcp_select_initial_window(const struct sock *sk, int __space, __u32 mss, 207f410cbeaSEric Dumazet __u32 *rcv_wnd, __u32 *__window_clamp, 20831d12926Slaurent chavey int wscale_ok, __u8 *rcv_wscale, 20931d12926Slaurent chavey __u32 init_rcv_wnd) 2101da177e4SLinus Torvalds { 2111da177e4SLinus Torvalds unsigned int space = (__space < 0 ? 0 : __space); 212f410cbeaSEric Dumazet u32 window_clamp = READ_ONCE(*__window_clamp); 2131da177e4SLinus Torvalds 2141da177e4SLinus Torvalds /* If no clamp set the clamp to the max possible scaled window */ 215f410cbeaSEric Dumazet if (window_clamp == 0) 216f410cbeaSEric Dumazet window_clamp = (U16_MAX << TCP_MAX_WSCALE); 217f410cbeaSEric Dumazet space = min(window_clamp, space); 2181da177e4SLinus Torvalds 2191da177e4SLinus Torvalds /* Quantize space offering to a multiple of mss if possible. */ 2201da177e4SLinus Torvalds if (space > mss) 221589c49cbSGao Feng space = rounddown(space, mss); 2221da177e4SLinus Torvalds 2231da177e4SLinus Torvalds /* NOTE: offering an initial window larger than 32767 22415d99e02SRick Jones * will break some buggy TCP stacks. If the admin tells us 22515d99e02SRick Jones * it is likely we could be speaking with such a buggy stack 22615d99e02SRick Jones * we will truncate our initial window offering to 32K-1 22715d99e02SRick Jones * unless the remote has sent us a window scaling option, 22815d99e02SRick Jones * which we interpret as a sign the remote TCP is not 22915d99e02SRick Jones * misinterpreting the window field as a signed quantity. 2301da177e4SLinus Torvalds */ 2310f1e4d06SKuniyuki Iwashima if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows)) 2321da177e4SLinus Torvalds (*rcv_wnd) = min(space, MAX_TCP_WINDOW); 23315d99e02SRick Jones else 234a337531bSYuchung Cheng (*rcv_wnd) = min_t(u32, space, U16_MAX); 235a337531bSYuchung Cheng 236a337531bSYuchung Cheng if (init_rcv_wnd) 237a337531bSYuchung Cheng *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss); 23815d99e02SRick Jones 23919bf6261SEric Dumazet *rcv_wscale = 0; 2401da177e4SLinus Torvalds if (wscale_ok) { 241589c49cbSGao Feng /* Set window scaling on max possible window */ 24202739545SKuniyuki Iwashima space = max_t(u32, space, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2])); 2431227c177SKuniyuki Iwashima space = max_t(u32, space, READ_ONCE(sysctl_rmem_max)); 244f410cbeaSEric Dumazet space = min_t(u32, space, window_clamp); 24519bf6261SEric Dumazet *rcv_wscale = clamp_t(int, ilog2(space) - 15, 24619bf6261SEric Dumazet 0, TCP_MAX_WSCALE); 2471da177e4SLinus Torvalds } 2481da177e4SLinus Torvalds /* Set the clamp no higher than max representable value */ 249f410cbeaSEric Dumazet WRITE_ONCE(*__window_clamp, 250f410cbeaSEric Dumazet min_t(__u32, U16_MAX << (*rcv_wscale), window_clamp)); 2511da177e4SLinus Torvalds } 2524bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_select_initial_window); 2531da177e4SLinus Torvalds 2541da177e4SLinus Torvalds /* Chose a new window to advertise, update state in tcp_sock for the 2551da177e4SLinus Torvalds * socket, and return result with RFC1323 scaling applied. The return 2561da177e4SLinus Torvalds * value can be stuffed directly into th->window for an outgoing 2571da177e4SLinus Torvalds * frame. 2581da177e4SLinus Torvalds */ 25940efc6faSStephen Hemminger static u16 tcp_select_window(struct sock *sk) 2601da177e4SLinus Torvalds { 2611da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 262b650d953Smfreemon@cloudflare.com struct net *net = sock_net(sk); 263e2142825SMenglong Dong u32 old_win = tp->rcv_wnd; 264e2142825SMenglong Dong u32 cur_win, new_win; 2651da177e4SLinus Torvalds 266e2142825SMenglong Dong /* Make the window 0 if we failed to queue the data because we 267e2142825SMenglong Dong * are out of memory. The window is temporary, so we don't store 268e2142825SMenglong Dong * it on the socket. 269e2142825SMenglong Dong */ 270e2142825SMenglong Dong if (unlikely(inet_csk(sk)->icsk_ack.pending & ICSK_ACK_NOMEM)) 271e2142825SMenglong Dong return 0; 272e2142825SMenglong Dong 273e2142825SMenglong Dong cur_win = tcp_receive_window(tp); 274e2142825SMenglong Dong new_win = __tcp_select_window(sk); 2751da177e4SLinus Torvalds if (new_win < cur_win) { 2761da177e4SLinus Torvalds /* Danger Will Robinson! 2771da177e4SLinus Torvalds * Don't update rcv_wup/rcv_wnd here or else 2781da177e4SLinus Torvalds * we will not be able to advertise a zero 2791da177e4SLinus Torvalds * window in time. --DaveM 2801da177e4SLinus Torvalds * 2811da177e4SLinus Torvalds * Relax Will Robinson. 2821da177e4SLinus Torvalds */ 283b650d953Smfreemon@cloudflare.com if (!READ_ONCE(net->ipv4.sysctl_tcp_shrink_window) || !tp->rx_opt.rcv_wscale) { 284b650d953Smfreemon@cloudflare.com /* Never shrink the offered window */ 2858e165e20SFlorian Westphal if (new_win == 0) 286b650d953Smfreemon@cloudflare.com NET_INC_STATS(net, LINUX_MIB_TCPWANTZEROWINDOWADV); 287607bfbf2SPatrick McHardy new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale); 2881da177e4SLinus Torvalds } 289b650d953Smfreemon@cloudflare.com } 290b650d953Smfreemon@cloudflare.com 2911da177e4SLinus Torvalds tp->rcv_wnd = new_win; 2921da177e4SLinus Torvalds tp->rcv_wup = tp->rcv_nxt; 2931da177e4SLinus Torvalds 2941da177e4SLinus Torvalds /* Make sure we do not exceed the maximum possible 2951da177e4SLinus Torvalds * scaled window. 2961da177e4SLinus Torvalds */ 297ceef9ab6SEric Dumazet if (!tp->rx_opt.rcv_wscale && 298b650d953Smfreemon@cloudflare.com READ_ONCE(net->ipv4.sysctl_tcp_workaround_signed_windows)) 2991da177e4SLinus Torvalds new_win = min(new_win, MAX_TCP_WINDOW); 3001da177e4SLinus Torvalds else 3011da177e4SLinus Torvalds new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); 3021da177e4SLinus Torvalds 3031da177e4SLinus Torvalds /* RFC1323 scaling applied */ 3041da177e4SLinus Torvalds new_win >>= tp->rx_opt.rcv_wscale; 3051da177e4SLinus Torvalds 30631770e34SFlorian Westphal /* If we advertise zero window, disable fast path. */ 3078e165e20SFlorian Westphal if (new_win == 0) { 30831770e34SFlorian Westphal tp->pred_flags = 0; 3098e165e20SFlorian Westphal if (old_win) 310b650d953Smfreemon@cloudflare.com NET_INC_STATS(net, LINUX_MIB_TCPTOZEROWINDOWADV); 3118e165e20SFlorian Westphal } else if (old_win == 0) { 312b650d953Smfreemon@cloudflare.com NET_INC_STATS(net, LINUX_MIB_TCPFROMZEROWINDOWADV); 3138e165e20SFlorian Westphal } 3141da177e4SLinus Torvalds 3151da177e4SLinus Torvalds return new_win; 3161da177e4SLinus Torvalds } 3171da177e4SLinus Torvalds 31867edfef7SAndi Kleen /* Packet ECN state for a SYN-ACK */ 319735d3831SFlorian Westphal static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb) 320bdf1ee5dSIlpo Järvinen { 32130e502a3SDaniel Borkmann const struct tcp_sock *tp = tcp_sk(sk); 32230e502a3SDaniel Borkmann 3234de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR; 324bdf1ee5dSIlpo Järvinen if (!(tp->ecn_flags & TCP_ECN_OK)) 3254de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE; 32691b5b21cSLawrence Brakmo else if (tcp_ca_needs_ecn(sk) || 32791b5b21cSLawrence Brakmo tcp_bpf_ca_needs_ecn(sk)) 32830e502a3SDaniel Borkmann INET_ECN_xmit(sk); 329bdf1ee5dSIlpo Järvinen } 330bdf1ee5dSIlpo Järvinen 33167edfef7SAndi Kleen /* Packet ECN state for a SYN. */ 332735d3831SFlorian Westphal static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb) 333bdf1ee5dSIlpo Järvinen { 334bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 33591b5b21cSLawrence Brakmo bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk); 3364785a667SKuniyuki Iwashima bool use_ecn = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_ecn) == 1 || 33791b5b21cSLawrence Brakmo tcp_ca_needs_ecn(sk) || bpf_needs_ecn; 338f7b3bec6SFlorian Westphal 339f7b3bec6SFlorian Westphal if (!use_ecn) { 340f7b3bec6SFlorian Westphal const struct dst_entry *dst = __sk_dst_get(sk); 341f7b3bec6SFlorian Westphal 342f7b3bec6SFlorian Westphal if (dst && dst_feature(dst, RTAX_FEATURE_ECN)) 343f7b3bec6SFlorian Westphal use_ecn = true; 344f7b3bec6SFlorian Westphal } 345bdf1ee5dSIlpo Järvinen 346bdf1ee5dSIlpo Järvinen tp->ecn_flags = 0; 347f7b3bec6SFlorian Westphal 348f7b3bec6SFlorian Westphal if (use_ecn) { 3494de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR; 350bdf1ee5dSIlpo Järvinen tp->ecn_flags = TCP_ECN_OK; 35191b5b21cSLawrence Brakmo if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn) 35230e502a3SDaniel Borkmann INET_ECN_xmit(sk); 353bdf1ee5dSIlpo Järvinen } 354bdf1ee5dSIlpo Järvinen } 355bdf1ee5dSIlpo Järvinen 35649213555SDaniel Borkmann static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb) 35749213555SDaniel Borkmann { 35812b8d9caSKuniyuki Iwashima if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback)) 35949213555SDaniel Borkmann /* tp->ecn_flags are cleared at a later point in time when 36049213555SDaniel Borkmann * SYN ACK is ultimatively being received. 36149213555SDaniel Borkmann */ 36249213555SDaniel Borkmann TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR); 36349213555SDaniel Borkmann } 36449213555SDaniel Borkmann 365735d3831SFlorian Westphal static void 3666ac705b1SEric Dumazet tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th) 367bdf1ee5dSIlpo Järvinen { 3686ac705b1SEric Dumazet if (inet_rsk(req)->ecn_ok) 369bdf1ee5dSIlpo Järvinen th->ece = 1; 370bdf1ee5dSIlpo Järvinen } 371bdf1ee5dSIlpo Järvinen 37267edfef7SAndi Kleen /* Set up ECN state for a packet on a ESTABLISHED socket that is about to 37367edfef7SAndi Kleen * be sent. 37467edfef7SAndi Kleen */ 375735d3831SFlorian Westphal static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb, 376ea1627c2SEric Dumazet struct tcphdr *th, int tcp_header_len) 377bdf1ee5dSIlpo Järvinen { 378bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 379bdf1ee5dSIlpo Järvinen 380bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_OK) { 381bdf1ee5dSIlpo Järvinen /* Not-retransmitted data segment: set ECT and inject CWR. */ 382bdf1ee5dSIlpo Järvinen if (skb->len != tcp_header_len && 383bdf1ee5dSIlpo Järvinen !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { 384bdf1ee5dSIlpo Järvinen INET_ECN_xmit(sk); 385bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) { 386bdf1ee5dSIlpo Järvinen tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; 387ea1627c2SEric Dumazet th->cwr = 1; 388bdf1ee5dSIlpo Järvinen skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; 389bdf1ee5dSIlpo Järvinen } 39030e502a3SDaniel Borkmann } else if (!tcp_ca_needs_ecn(sk)) { 391bdf1ee5dSIlpo Järvinen /* ACK or retransmitted segment: clear ECT|CE */ 392bdf1ee5dSIlpo Järvinen INET_ECN_dontxmit(sk); 393bdf1ee5dSIlpo Järvinen } 394bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) 395ea1627c2SEric Dumazet th->ece = 1; 396bdf1ee5dSIlpo Järvinen } 397bdf1ee5dSIlpo Järvinen } 398bdf1ee5dSIlpo Järvinen 399e870a8efSIlpo Järvinen /* Constructs common control bits of non-data skb. If SYN/FIN is present, 400e870a8efSIlpo Järvinen * auto increment end seqno. 401e870a8efSIlpo Järvinen */ 402e870a8efSIlpo Järvinen static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags) 403e870a8efSIlpo Järvinen { 4042e8e18efSDavid S. Miller skb->ip_summed = CHECKSUM_PARTIAL; 405e870a8efSIlpo Järvinen 4064de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags; 407e870a8efSIlpo Järvinen 408cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, 1); 409e870a8efSIlpo Järvinen 410e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->seq = seq; 411a3433f35SChangli Gao if (flags & (TCPHDR_SYN | TCPHDR_FIN)) 412e870a8efSIlpo Järvinen seq++; 413e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->end_seq = seq; 414e870a8efSIlpo Järvinen } 415e870a8efSIlpo Järvinen 416a2a385d6SEric Dumazet static inline bool tcp_urg_mode(const struct tcp_sock *tp) 41733f5f57eSIlpo Järvinen { 41833f5f57eSIlpo Järvinen return tp->snd_una != tp->snd_up; 41933f5f57eSIlpo Järvinen } 42033f5f57eSIlpo Järvinen 4213b65abb8SLeonard Crestez #define OPTION_SACK_ADVERTISE BIT(0) 4223b65abb8SLeonard Crestez #define OPTION_TS BIT(1) 4233b65abb8SLeonard Crestez #define OPTION_MD5 BIT(2) 4243b65abb8SLeonard Crestez #define OPTION_WSCALE BIT(3) 4253b65abb8SLeonard Crestez #define OPTION_FAST_OPEN_COOKIE BIT(8) 4263b65abb8SLeonard Crestez #define OPTION_SMC BIT(9) 4273b65abb8SLeonard Crestez #define OPTION_MPTCP BIT(10) 4281e03d32bSDmitry Safonov #define OPTION_AO BIT(11) 42960e2a778SUrsula Braun 43060e2a778SUrsula Braun static void smc_options_write(__be32 *ptr, u16 *options) 43160e2a778SUrsula Braun { 43260e2a778SUrsula Braun #if IS_ENABLED(CONFIG_SMC) 43360e2a778SUrsula Braun if (static_branch_unlikely(&tcp_have_smc)) { 43460e2a778SUrsula Braun if (unlikely(OPTION_SMC & *options)) { 43560e2a778SUrsula Braun *ptr++ = htonl((TCPOPT_NOP << 24) | 43660e2a778SUrsula Braun (TCPOPT_NOP << 16) | 43760e2a778SUrsula Braun (TCPOPT_EXP << 8) | 43860e2a778SUrsula Braun (TCPOLEN_EXP_SMC_BASE)); 43960e2a778SUrsula Braun *ptr++ = htonl(TCPOPT_SMC_MAGIC); 44060e2a778SUrsula Braun } 44160e2a778SUrsula Braun } 44260e2a778SUrsula Braun #endif 44360e2a778SUrsula Braun } 44433ad798cSAdam Langley 44533ad798cSAdam Langley struct tcp_out_options { 4462100c8d2SYuchung Cheng u16 options; /* bit field of OPTION_* */ 4472100c8d2SYuchung Cheng u16 mss; /* 0 to disable */ 44833ad798cSAdam Langley u8 ws; /* window scale, 0 to disable */ 44933ad798cSAdam Langley u8 num_sack_blocks; /* number of SACK blocks to include */ 450bd0388aeSWilliam Allen Simpson u8 hash_size; /* bytes in hash_location */ 451331fca43SMartin KaFai Lau u8 bpf_opt_len; /* length of BPF hdr option */ 452bd0388aeSWilliam Allen Simpson __u8 *hash_location; /* temporary pointer, overloaded */ 4532100c8d2SYuchung Cheng __u32 tsval, tsecr; /* need to include OPTION_TS */ 4542100c8d2SYuchung Cheng struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */ 455eda7acddSPeter Krystad struct mptcp_out_options mptcp; 45633ad798cSAdam Langley }; 45733ad798cSAdam Langley 458ea66758cSPaolo Abeni static void mptcp_options_write(struct tcphdr *th, __be32 *ptr, 459ea66758cSPaolo Abeni struct tcp_sock *tp, 460fa3fe2b1SFlorian Westphal struct tcp_out_options *opts) 461eda7acddSPeter Krystad { 462eda7acddSPeter Krystad #if IS_ENABLED(CONFIG_MPTCP) 463eda7acddSPeter Krystad if (unlikely(OPTION_MPTCP & opts->options)) 464ea66758cSPaolo Abeni mptcp_write_options(th, ptr, tp, &opts->mptcp); 465eda7acddSPeter Krystad #endif 466eda7acddSPeter Krystad } 467eda7acddSPeter Krystad 468331fca43SMartin KaFai Lau #ifdef CONFIG_CGROUP_BPF 4690813a841SMartin KaFai Lau static int bpf_skops_write_hdr_opt_arg0(struct sk_buff *skb, 4700813a841SMartin KaFai Lau enum tcp_synack_type synack_type) 4710813a841SMartin KaFai Lau { 4720813a841SMartin KaFai Lau if (unlikely(!skb)) 4730813a841SMartin KaFai Lau return BPF_WRITE_HDR_TCP_CURRENT_MSS; 4740813a841SMartin KaFai Lau 4750813a841SMartin KaFai Lau if (unlikely(synack_type == TCP_SYNACK_COOKIE)) 4760813a841SMartin KaFai Lau return BPF_WRITE_HDR_TCP_SYNACK_COOKIE; 4770813a841SMartin KaFai Lau 4780813a841SMartin KaFai Lau return 0; 4790813a841SMartin KaFai Lau } 4800813a841SMartin KaFai Lau 481331fca43SMartin KaFai Lau /* req, syn_skb and synack_type are used when writing synack */ 482331fca43SMartin KaFai Lau static void bpf_skops_hdr_opt_len(struct sock *sk, struct sk_buff *skb, 483331fca43SMartin KaFai Lau struct request_sock *req, 484331fca43SMartin KaFai Lau struct sk_buff *syn_skb, 485331fca43SMartin KaFai Lau enum tcp_synack_type synack_type, 486331fca43SMartin KaFai Lau struct tcp_out_options *opts, 487331fca43SMartin KaFai Lau unsigned int *remaining) 488331fca43SMartin KaFai Lau { 4890813a841SMartin KaFai Lau struct bpf_sock_ops_kern sock_ops; 4900813a841SMartin KaFai Lau int err; 4910813a841SMartin KaFai Lau 492331fca43SMartin KaFai Lau if (likely(!BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), 493331fca43SMartin KaFai Lau BPF_SOCK_OPS_WRITE_HDR_OPT_CB_FLAG)) || 494331fca43SMartin KaFai Lau !*remaining) 495331fca43SMartin KaFai Lau return; 496331fca43SMartin KaFai Lau 4970813a841SMartin KaFai Lau /* *remaining has already been aligned to 4 bytes, so *remaining >= 4 */ 4980813a841SMartin KaFai Lau 4990813a841SMartin KaFai Lau /* init sock_ops */ 5000813a841SMartin KaFai Lau memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp)); 5010813a841SMartin KaFai Lau 5020813a841SMartin KaFai Lau sock_ops.op = BPF_SOCK_OPS_HDR_OPT_LEN_CB; 5030813a841SMartin KaFai Lau 5040813a841SMartin KaFai Lau if (req) { 5050813a841SMartin KaFai Lau /* The listen "sk" cannot be passed here because 5060813a841SMartin KaFai Lau * it is not locked. It would not make too much 5070813a841SMartin KaFai Lau * sense to do bpf_setsockopt(listen_sk) based 5080813a841SMartin KaFai Lau * on individual connection request also. 5090813a841SMartin KaFai Lau * 5100813a841SMartin KaFai Lau * Thus, "req" is passed here and the cgroup-bpf-progs 5110813a841SMartin KaFai Lau * of the listen "sk" will be run. 5120813a841SMartin KaFai Lau * 5130813a841SMartin KaFai Lau * "req" is also used here for fastopen even the "sk" here is 5140813a841SMartin KaFai Lau * a fullsock "child" sk. It is to keep the behavior 5150813a841SMartin KaFai Lau * consistent between fastopen and non-fastopen on 5160813a841SMartin KaFai Lau * the bpf programming side. 517331fca43SMartin KaFai Lau */ 5180813a841SMartin KaFai Lau sock_ops.sk = (struct sock *)req; 5190813a841SMartin KaFai Lau sock_ops.syn_skb = syn_skb; 5200813a841SMartin KaFai Lau } else { 5210813a841SMartin KaFai Lau sock_owned_by_me(sk); 5220813a841SMartin KaFai Lau 5230813a841SMartin KaFai Lau sock_ops.is_fullsock = 1; 5240813a841SMartin KaFai Lau sock_ops.sk = sk; 5250813a841SMartin KaFai Lau } 5260813a841SMartin KaFai Lau 5270813a841SMartin KaFai Lau sock_ops.args[0] = bpf_skops_write_hdr_opt_arg0(skb, synack_type); 5280813a841SMartin KaFai Lau sock_ops.remaining_opt_len = *remaining; 5290813a841SMartin KaFai Lau /* tcp_current_mss() does not pass a skb */ 5300813a841SMartin KaFai Lau if (skb) 5310813a841SMartin KaFai Lau bpf_skops_init_skb(&sock_ops, skb, 0); 5320813a841SMartin KaFai Lau 5330813a841SMartin KaFai Lau err = BPF_CGROUP_RUN_PROG_SOCK_OPS_SK(&sock_ops, sk); 5340813a841SMartin KaFai Lau 5350813a841SMartin KaFai Lau if (err || sock_ops.remaining_opt_len == *remaining) 5360813a841SMartin KaFai Lau return; 5370813a841SMartin KaFai Lau 5380813a841SMartin KaFai Lau opts->bpf_opt_len = *remaining - sock_ops.remaining_opt_len; 5390813a841SMartin KaFai Lau /* round up to 4 bytes */ 5400813a841SMartin KaFai Lau opts->bpf_opt_len = (opts->bpf_opt_len + 3) & ~3; 5410813a841SMartin KaFai Lau 5420813a841SMartin KaFai Lau *remaining -= opts->bpf_opt_len; 543331fca43SMartin KaFai Lau } 544331fca43SMartin KaFai Lau 545331fca43SMartin KaFai Lau static void bpf_skops_write_hdr_opt(struct sock *sk, struct sk_buff *skb, 546331fca43SMartin KaFai Lau struct request_sock *req, 547331fca43SMartin KaFai Lau struct sk_buff *syn_skb, 548331fca43SMartin KaFai Lau enum tcp_synack_type synack_type, 549331fca43SMartin KaFai Lau struct tcp_out_options *opts) 550331fca43SMartin KaFai Lau { 5510813a841SMartin KaFai Lau u8 first_opt_off, nr_written, max_opt_len = opts->bpf_opt_len; 5520813a841SMartin KaFai Lau struct bpf_sock_ops_kern sock_ops; 5530813a841SMartin KaFai Lau int err; 5540813a841SMartin KaFai Lau 5550813a841SMartin KaFai Lau if (likely(!max_opt_len)) 556331fca43SMartin KaFai Lau return; 557331fca43SMartin KaFai Lau 5580813a841SMartin KaFai Lau memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp)); 5590813a841SMartin KaFai Lau 5600813a841SMartin KaFai Lau sock_ops.op = BPF_SOCK_OPS_WRITE_HDR_OPT_CB; 5610813a841SMartin KaFai Lau 5620813a841SMartin KaFai Lau if (req) { 5630813a841SMartin KaFai Lau sock_ops.sk = (struct sock *)req; 5640813a841SMartin KaFai Lau sock_ops.syn_skb = syn_skb; 5650813a841SMartin KaFai Lau } else { 5660813a841SMartin KaFai Lau sock_owned_by_me(sk); 5670813a841SMartin KaFai Lau 5680813a841SMartin KaFai Lau sock_ops.is_fullsock = 1; 5690813a841SMartin KaFai Lau sock_ops.sk = sk; 5700813a841SMartin KaFai Lau } 5710813a841SMartin KaFai Lau 5720813a841SMartin KaFai Lau sock_ops.args[0] = bpf_skops_write_hdr_opt_arg0(skb, synack_type); 5730813a841SMartin KaFai Lau sock_ops.remaining_opt_len = max_opt_len; 5740813a841SMartin KaFai Lau first_opt_off = tcp_hdrlen(skb) - max_opt_len; 5750813a841SMartin KaFai Lau bpf_skops_init_skb(&sock_ops, skb, first_opt_off); 5760813a841SMartin KaFai Lau 5770813a841SMartin KaFai Lau err = BPF_CGROUP_RUN_PROG_SOCK_OPS_SK(&sock_ops, sk); 5780813a841SMartin KaFai Lau 5790813a841SMartin KaFai Lau if (err) 5800813a841SMartin KaFai Lau nr_written = 0; 5810813a841SMartin KaFai Lau else 5820813a841SMartin KaFai Lau nr_written = max_opt_len - sock_ops.remaining_opt_len; 5830813a841SMartin KaFai Lau 5840813a841SMartin KaFai Lau if (nr_written < max_opt_len) 5850813a841SMartin KaFai Lau memset(skb->data + first_opt_off + nr_written, TCPOPT_NOP, 5860813a841SMartin KaFai Lau max_opt_len - nr_written); 587331fca43SMartin KaFai Lau } 588331fca43SMartin KaFai Lau #else 589331fca43SMartin KaFai Lau static void bpf_skops_hdr_opt_len(struct sock *sk, struct sk_buff *skb, 590331fca43SMartin KaFai Lau struct request_sock *req, 591331fca43SMartin KaFai Lau struct sk_buff *syn_skb, 592331fca43SMartin KaFai Lau enum tcp_synack_type synack_type, 593331fca43SMartin KaFai Lau struct tcp_out_options *opts, 594331fca43SMartin KaFai Lau unsigned int *remaining) 595331fca43SMartin KaFai Lau { 596331fca43SMartin KaFai Lau } 597331fca43SMartin KaFai Lau 598331fca43SMartin KaFai Lau static void bpf_skops_write_hdr_opt(struct sock *sk, struct sk_buff *skb, 599331fca43SMartin KaFai Lau struct request_sock *req, 600331fca43SMartin KaFai Lau struct sk_buff *syn_skb, 601331fca43SMartin KaFai Lau enum tcp_synack_type synack_type, 602331fca43SMartin KaFai Lau struct tcp_out_options *opts) 603331fca43SMartin KaFai Lau { 604331fca43SMartin KaFai Lau } 605331fca43SMartin KaFai Lau #endif 606331fca43SMartin KaFai Lau 6077425627bSNathan Chancellor static __be32 *process_tcp_ao_options(struct tcp_sock *tp, 6087425627bSNathan Chancellor const struct tcp_request_sock *tcprsk, 6097425627bSNathan Chancellor struct tcp_out_options *opts, 6107425627bSNathan Chancellor struct tcp_key *key, __be32 *ptr) 6117425627bSNathan Chancellor { 6127425627bSNathan Chancellor #ifdef CONFIG_TCP_AO 6137425627bSNathan Chancellor u8 maclen = tcp_ao_maclen(key->ao_key); 6147425627bSNathan Chancellor 6157425627bSNathan Chancellor if (tcprsk) { 6167425627bSNathan Chancellor u8 aolen = maclen + sizeof(struct tcp_ao_hdr); 6177425627bSNathan Chancellor 6187425627bSNathan Chancellor *ptr++ = htonl((TCPOPT_AO << 24) | (aolen << 16) | 6197425627bSNathan Chancellor (tcprsk->ao_keyid << 8) | 6207425627bSNathan Chancellor (tcprsk->ao_rcv_next)); 6217425627bSNathan Chancellor } else { 6227425627bSNathan Chancellor struct tcp_ao_key *rnext_key; 6237425627bSNathan Chancellor struct tcp_ao_info *ao_info; 6247425627bSNathan Chancellor 6257425627bSNathan Chancellor ao_info = rcu_dereference_check(tp->ao_info, 6267425627bSNathan Chancellor lockdep_sock_is_held(&tp->inet_conn.icsk_inet.sk)); 6277425627bSNathan Chancellor rnext_key = READ_ONCE(ao_info->rnext_key); 6287425627bSNathan Chancellor if (WARN_ON_ONCE(!rnext_key)) 6297425627bSNathan Chancellor return ptr; 6307425627bSNathan Chancellor *ptr++ = htonl((TCPOPT_AO << 24) | 6317425627bSNathan Chancellor (tcp_ao_len(key->ao_key) << 16) | 6327425627bSNathan Chancellor (key->ao_key->sndid << 8) | 6337425627bSNathan Chancellor (rnext_key->rcvid)); 6347425627bSNathan Chancellor } 6357425627bSNathan Chancellor opts->hash_location = (__u8 *)ptr; 6367425627bSNathan Chancellor ptr += maclen / sizeof(*ptr); 6377425627bSNathan Chancellor if (unlikely(maclen % sizeof(*ptr))) { 6387425627bSNathan Chancellor memset(ptr, TCPOPT_NOP, sizeof(*ptr)); 6397425627bSNathan Chancellor ptr++; 6407425627bSNathan Chancellor } 6417425627bSNathan Chancellor #endif 6427425627bSNathan Chancellor return ptr; 6437425627bSNathan Chancellor } 6447425627bSNathan Chancellor 64567edfef7SAndi Kleen /* Write previously computed TCP options to the packet. 64667edfef7SAndi Kleen * 64767edfef7SAndi Kleen * Beware: Something in the Internet is very sensitive to the ordering of 648fd6149d3SIlpo Järvinen * TCP options, we learned this through the hard way, so be careful here. 649fd6149d3SIlpo Järvinen * Luckily we can at least blame others for their non-compliance but from 6508e3bff96Sstephen hemminger * inter-operability perspective it seems that we're somewhat stuck with 651fd6149d3SIlpo Järvinen * the ordering which we have been using if we want to keep working with 652fd6149d3SIlpo Järvinen * those broken things (not that it currently hurts anybody as there isn't 653fd6149d3SIlpo Järvinen * particular reason why the ordering would need to be changed). 654fd6149d3SIlpo Järvinen * 655fd6149d3SIlpo Järvinen * At least SACK_PERM as the first option is known to lead to a disaster 656fd6149d3SIlpo Järvinen * (but it may well be that other scenarios fail similarly). 657fd6149d3SIlpo Järvinen */ 658ea66758cSPaolo Abeni static void tcp_options_write(struct tcphdr *th, struct tcp_sock *tp, 65906b22ef2SDmitry Safonov const struct tcp_request_sock *tcprsk, 6601e03d32bSDmitry Safonov struct tcp_out_options *opts, 6611e03d32bSDmitry Safonov struct tcp_key *key) 662bd0388aeSWilliam Allen Simpson { 663ea66758cSPaolo Abeni __be32 *ptr = (__be32 *)(th + 1); 6642100c8d2SYuchung Cheng u16 options = opts->options; /* mungable copy */ 665bd0388aeSWilliam Allen Simpson 6661e03d32bSDmitry Safonov if (tcp_key_is_md5(key)) { 6671a2c6181SChristoph Paasch *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | 6681a2c6181SChristoph Paasch (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG); 669bd0388aeSWilliam Allen Simpson /* overload cookie hash location */ 670bd0388aeSWilliam Allen Simpson opts->hash_location = (__u8 *)ptr; 67133ad798cSAdam Langley ptr += 4; 6721e03d32bSDmitry Safonov } else if (tcp_key_is_ao(key)) { 6737425627bSNathan Chancellor ptr = process_tcp_ao_options(tp, tcprsk, opts, key, ptr); 6741e03d32bSDmitry Safonov } 675fd6149d3SIlpo Järvinen if (unlikely(opts->mss)) { 676fd6149d3SIlpo Järvinen *ptr++ = htonl((TCPOPT_MSS << 24) | 677fd6149d3SIlpo Järvinen (TCPOLEN_MSS << 16) | 678fd6149d3SIlpo Järvinen opts->mss); 679fd6149d3SIlpo Järvinen } 680fd6149d3SIlpo Järvinen 681bd0388aeSWilliam Allen Simpson if (likely(OPTION_TS & options)) { 682bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 68333ad798cSAdam Langley *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | 68433ad798cSAdam Langley (TCPOLEN_SACK_PERM << 16) | 68533ad798cSAdam Langley (TCPOPT_TIMESTAMP << 8) | 68633ad798cSAdam Langley TCPOLEN_TIMESTAMP); 687bd0388aeSWilliam Allen Simpson options &= ~OPTION_SACK_ADVERTISE; 68833ad798cSAdam Langley } else { 689496c98dfSYOSHIFUJI Hideaki *ptr++ = htonl((TCPOPT_NOP << 24) | 69040efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 69140efc6faSStephen Hemminger (TCPOPT_TIMESTAMP << 8) | 69240efc6faSStephen Hemminger TCPOLEN_TIMESTAMP); 69340efc6faSStephen Hemminger } 69433ad798cSAdam Langley *ptr++ = htonl(opts->tsval); 69533ad798cSAdam Langley *ptr++ = htonl(opts->tsecr); 69633ad798cSAdam Langley } 69733ad798cSAdam Langley 698bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 69933ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 70033ad798cSAdam Langley (TCPOPT_NOP << 16) | 70133ad798cSAdam Langley (TCPOPT_SACK_PERM << 8) | 70233ad798cSAdam Langley TCPOLEN_SACK_PERM); 70333ad798cSAdam Langley } 70433ad798cSAdam Langley 705bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_WSCALE & options)) { 70633ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 70733ad798cSAdam Langley (TCPOPT_WINDOW << 16) | 70833ad798cSAdam Langley (TCPOLEN_WINDOW << 8) | 70933ad798cSAdam Langley opts->ws); 71033ad798cSAdam Langley } 71133ad798cSAdam Langley 71233ad798cSAdam Langley if (unlikely(opts->num_sack_blocks)) { 71333ad798cSAdam Langley struct tcp_sack_block *sp = tp->rx_opt.dsack ? 71433ad798cSAdam Langley tp->duplicate_sack : tp->selective_acks; 71540efc6faSStephen Hemminger int this_sack; 71640efc6faSStephen Hemminger 71740efc6faSStephen Hemminger *ptr++ = htonl((TCPOPT_NOP << 24) | 71840efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 71940efc6faSStephen Hemminger (TCPOPT_SACK << 8) | 72033ad798cSAdam Langley (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * 72140efc6faSStephen Hemminger TCPOLEN_SACK_PERBLOCK))); 7222de979bdSStephen Hemminger 72333ad798cSAdam Langley for (this_sack = 0; this_sack < opts->num_sack_blocks; 72433ad798cSAdam Langley ++this_sack) { 72540efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].start_seq); 72640efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].end_seq); 72740efc6faSStephen Hemminger } 7282de979bdSStephen Hemminger 72940efc6faSStephen Hemminger tp->rx_opt.dsack = 0; 73040efc6faSStephen Hemminger } 7312100c8d2SYuchung Cheng 7322100c8d2SYuchung Cheng if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) { 7332100c8d2SYuchung Cheng struct tcp_fastopen_cookie *foc = opts->fastopen_cookie; 7347f9b838bSDaniel Lee u8 *p = (u8 *)ptr; 7357f9b838bSDaniel Lee u32 len; /* Fast Open option length */ 7362100c8d2SYuchung Cheng 7377f9b838bSDaniel Lee if (foc->exp) { 7387f9b838bSDaniel Lee len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len; 7397f9b838bSDaniel Lee *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) | 7402100c8d2SYuchung Cheng TCPOPT_FASTOPEN_MAGIC); 7417f9b838bSDaniel Lee p += TCPOLEN_EXP_FASTOPEN_BASE; 7427f9b838bSDaniel Lee } else { 7437f9b838bSDaniel Lee len = TCPOLEN_FASTOPEN_BASE + foc->len; 7447f9b838bSDaniel Lee *p++ = TCPOPT_FASTOPEN; 7457f9b838bSDaniel Lee *p++ = len; 7462100c8d2SYuchung Cheng } 7477f9b838bSDaniel Lee 7487f9b838bSDaniel Lee memcpy(p, foc->val, foc->len); 7497f9b838bSDaniel Lee if ((len & 3) == 2) { 7507f9b838bSDaniel Lee p[foc->len] = TCPOPT_NOP; 7517f9b838bSDaniel Lee p[foc->len + 1] = TCPOPT_NOP; 7527f9b838bSDaniel Lee } 7537f9b838bSDaniel Lee ptr += (len + 3) >> 2; 7542100c8d2SYuchung Cheng } 75560e2a778SUrsula Braun 75660e2a778SUrsula Braun smc_options_write(ptr, &options); 757eda7acddSPeter Krystad 758ea66758cSPaolo Abeni mptcp_options_write(th, ptr, tp, opts); 75960e2a778SUrsula Braun } 76060e2a778SUrsula Braun 76160e2a778SUrsula Braun static void smc_set_option(const struct tcp_sock *tp, 76260e2a778SUrsula Braun struct tcp_out_options *opts, 76360e2a778SUrsula Braun unsigned int *remaining) 76460e2a778SUrsula Braun { 76560e2a778SUrsula Braun #if IS_ENABLED(CONFIG_SMC) 76660e2a778SUrsula Braun if (static_branch_unlikely(&tcp_have_smc)) { 76760e2a778SUrsula Braun if (tp->syn_smc) { 76860e2a778SUrsula Braun if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) { 76960e2a778SUrsula Braun opts->options |= OPTION_SMC; 77060e2a778SUrsula Braun *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED; 77160e2a778SUrsula Braun } 77260e2a778SUrsula Braun } 77360e2a778SUrsula Braun } 77460e2a778SUrsula Braun #endif 77560e2a778SUrsula Braun } 77660e2a778SUrsula Braun 77760e2a778SUrsula Braun static void smc_set_option_cond(const struct tcp_sock *tp, 77860e2a778SUrsula Braun const struct inet_request_sock *ireq, 77960e2a778SUrsula Braun struct tcp_out_options *opts, 78060e2a778SUrsula Braun unsigned int *remaining) 78160e2a778SUrsula Braun { 78260e2a778SUrsula Braun #if IS_ENABLED(CONFIG_SMC) 78360e2a778SUrsula Braun if (static_branch_unlikely(&tcp_have_smc)) { 78460e2a778SUrsula Braun if (tp->syn_smc && ireq->smc_ok) { 78560e2a778SUrsula Braun if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) { 78660e2a778SUrsula Braun opts->options |= OPTION_SMC; 78760e2a778SUrsula Braun *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED; 78860e2a778SUrsula Braun } 78960e2a778SUrsula Braun } 79060e2a778SUrsula Braun } 79160e2a778SUrsula Braun #endif 79240efc6faSStephen Hemminger } 79340efc6faSStephen Hemminger 794cec37a6eSPeter Krystad static void mptcp_set_option_cond(const struct request_sock *req, 795cec37a6eSPeter Krystad struct tcp_out_options *opts, 796cec37a6eSPeter Krystad unsigned int *remaining) 797cec37a6eSPeter Krystad { 798cec37a6eSPeter Krystad if (rsk_is_mptcp(req)) { 799cec37a6eSPeter Krystad unsigned int size; 800cec37a6eSPeter Krystad 801cec37a6eSPeter Krystad if (mptcp_synack_options(req, &size, &opts->mptcp)) { 802cec37a6eSPeter Krystad if (*remaining >= size) { 803cec37a6eSPeter Krystad opts->options |= OPTION_MPTCP; 804cec37a6eSPeter Krystad *remaining -= size; 805cec37a6eSPeter Krystad } 806cec37a6eSPeter Krystad } 807cec37a6eSPeter Krystad } 808cec37a6eSPeter Krystad } 809cec37a6eSPeter Krystad 81067edfef7SAndi Kleen /* Compute TCP options for SYN packets. This is not the final 81167edfef7SAndi Kleen * network wire format yet. 81267edfef7SAndi Kleen */ 81395c96174SEric Dumazet static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb, 81433ad798cSAdam Langley struct tcp_out_options *opts, 8151e03d32bSDmitry Safonov struct tcp_key *key) 816cf533ea5SEric Dumazet { 81733ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 81895c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 819783237e8SYuchung Cheng struct tcp_fastopen_request *fastopen = tp->fastopen_req; 8201e03d32bSDmitry Safonov bool timestamps; 82133ad798cSAdam Langley 8221e03d32bSDmitry Safonov /* Better than switch (key.type) as it has static branches */ 8231e03d32bSDmitry Safonov if (tcp_key_is_md5(key)) { 8241e03d32bSDmitry Safonov timestamps = false; 82533ad798cSAdam Langley opts->options |= OPTION_MD5; 826bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 8271e03d32bSDmitry Safonov } else { 8281e03d32bSDmitry Safonov timestamps = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_timestamps); 8291e03d32bSDmitry Safonov if (tcp_key_is_ao(key)) { 8301e03d32bSDmitry Safonov opts->options |= OPTION_AO; 831da7dfaa6SDmitry Safonov remaining -= tcp_ao_len_aligned(key->ao_key); 832cfb6eeb4SYOSHIFUJI Hideaki } 8338c2320e8SEric Dumazet } 83433ad798cSAdam Langley 83533ad798cSAdam Langley /* We always get an MSS option. The option bytes which will be seen in 83633ad798cSAdam Langley * normal data packets should timestamps be used, must be in the MSS 83733ad798cSAdam Langley * advertised. But we subtract them from tp->mss_cache so that 83833ad798cSAdam Langley * calculations in tcp_sendmsg are simpler etc. So account for this 83933ad798cSAdam Langley * fact here if necessary. If we don't do this correctly, as a 84033ad798cSAdam Langley * receiver we won't recognize data packets as being full sized when we 84133ad798cSAdam Langley * should, and thus we won't abide by the delayed ACK rules correctly. 84233ad798cSAdam Langley * SACKs don't matter, we never delay an ACK when we have any of those 84333ad798cSAdam Langley * going out. */ 84433ad798cSAdam Langley opts->mss = tcp_advertise_mss(sk); 845bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 84633ad798cSAdam Langley 8471e03d32bSDmitry Safonov if (likely(timestamps)) { 84833ad798cSAdam Langley opts->options |= OPTION_TS; 849614e8316SEric Dumazet opts->tsval = tcp_skb_timestamp_ts(tp->tcp_usec_ts, skb) + tp->tsoffset; 85033ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 851bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 85233ad798cSAdam Langley } 8533666f666SKuniyuki Iwashima if (likely(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_window_scaling))) { 85433ad798cSAdam Langley opts->ws = tp->rx_opt.rcv_wscale; 85589e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 856bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 85733ad798cSAdam Langley } 8583666f666SKuniyuki Iwashima if (likely(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_sack))) { 85933ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 860b32d1310SDavid S. Miller if (unlikely(!(OPTION_TS & opts->options))) 861bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 86233ad798cSAdam Langley } 86333ad798cSAdam Langley 864783237e8SYuchung Cheng if (fastopen && fastopen->cookie.len >= 0) { 8652646c831SDaniel Lee u32 need = fastopen->cookie.len; 8662646c831SDaniel Lee 8672646c831SDaniel Lee need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE : 8682646c831SDaniel Lee TCPOLEN_FASTOPEN_BASE; 869783237e8SYuchung Cheng need = (need + 3) & ~3U; /* Align to 32 bits */ 870783237e8SYuchung Cheng if (remaining >= need) { 871783237e8SYuchung Cheng opts->options |= OPTION_FAST_OPEN_COOKIE; 872783237e8SYuchung Cheng opts->fastopen_cookie = &fastopen->cookie; 873783237e8SYuchung Cheng remaining -= need; 874783237e8SYuchung Cheng tp->syn_fastopen = 1; 8752646c831SDaniel Lee tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0; 876783237e8SYuchung Cheng } 877783237e8SYuchung Cheng } 878bd0388aeSWilliam Allen Simpson 87960e2a778SUrsula Braun smc_set_option(tp, opts, &remaining); 88060e2a778SUrsula Braun 881cec37a6eSPeter Krystad if (sk_is_mptcp(sk)) { 882cec37a6eSPeter Krystad unsigned int size; 883cec37a6eSPeter Krystad 884cc7972eaSChristoph Paasch if (mptcp_syn_options(sk, skb, &size, &opts->mptcp)) { 885cec37a6eSPeter Krystad opts->options |= OPTION_MPTCP; 886cec37a6eSPeter Krystad remaining -= size; 887cec37a6eSPeter Krystad } 888cec37a6eSPeter Krystad } 889cec37a6eSPeter Krystad 890331fca43SMartin KaFai Lau bpf_skops_hdr_opt_len(sk, skb, NULL, NULL, 0, opts, &remaining); 891331fca43SMartin KaFai Lau 892bd0388aeSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 89333ad798cSAdam Langley } 89433ad798cSAdam Langley 89567edfef7SAndi Kleen /* Set up TCP options for SYN-ACKs. */ 89660e2a778SUrsula Braun static unsigned int tcp_synack_options(const struct sock *sk, 89760e2a778SUrsula Braun struct request_sock *req, 89895c96174SEric Dumazet unsigned int mss, struct sk_buff *skb, 89933ad798cSAdam Langley struct tcp_out_options *opts, 9009427c6aaSDmitry Safonov const struct tcp_key *key, 901e114e1e8SEric Dumazet struct tcp_fastopen_cookie *foc, 902331fca43SMartin KaFai Lau enum tcp_synack_type synack_type, 903331fca43SMartin KaFai Lau struct sk_buff *syn_skb) 9044957faadSWilliam Allen Simpson { 90533ad798cSAdam Langley struct inet_request_sock *ireq = inet_rsk(req); 90695c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 90733ad798cSAdam Langley 9089427c6aaSDmitry Safonov if (tcp_key_is_md5(key)) { 90933ad798cSAdam Langley opts->options |= OPTION_MD5; 9104957faadSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 9114957faadSWilliam Allen Simpson 9124957faadSWilliam Allen Simpson /* We can't fit any SACK blocks in a packet with MD5 + TS 9134957faadSWilliam Allen Simpson * options. There was discussion about disabling SACK 9144957faadSWilliam Allen Simpson * rather than TS in order to fit in better with old, 9154957faadSWilliam Allen Simpson * buggy kernels, but that was deemed to be unnecessary. 9164957faadSWilliam Allen Simpson */ 917e114e1e8SEric Dumazet if (synack_type != TCP_SYNACK_COOKIE) 918de213e5eSEric Dumazet ireq->tstamp_ok &= !ireq->sack_ok; 9199427c6aaSDmitry Safonov } else if (tcp_key_is_ao(key)) { 9209427c6aaSDmitry Safonov opts->options |= OPTION_AO; 921da7dfaa6SDmitry Safonov remaining -= tcp_ao_len_aligned(key->ao_key); 9229427c6aaSDmitry Safonov ireq->tstamp_ok &= !ireq->sack_ok; 92333ad798cSAdam Langley } 92433ad798cSAdam Langley 9254957faadSWilliam Allen Simpson /* We always send an MSS option. */ 92633ad798cSAdam Langley opts->mss = mss; 9274957faadSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 92833ad798cSAdam Langley 92933ad798cSAdam Langley if (likely(ireq->wscale_ok)) { 93033ad798cSAdam Langley opts->ws = ireq->rcv_wscale; 93189e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 9324957faadSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 93333ad798cSAdam Langley } 934de213e5eSEric Dumazet if (likely(ireq->tstamp_ok)) { 93533ad798cSAdam Langley opts->options |= OPTION_TS; 936614e8316SEric Dumazet opts->tsval = tcp_skb_timestamp_ts(tcp_rsk(req)->req_usec_ts, skb) + 937614e8316SEric Dumazet tcp_rsk(req)->ts_off; 938eba20811SEric Dumazet opts->tsecr = READ_ONCE(req->ts_recent); 9394957faadSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 94033ad798cSAdam Langley } 94133ad798cSAdam Langley if (likely(ireq->sack_ok)) { 94233ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 943de213e5eSEric Dumazet if (unlikely(!ireq->tstamp_ok)) 9444957faadSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 94533ad798cSAdam Langley } 9467f9b838bSDaniel Lee if (foc != NULL && foc->len >= 0) { 9477f9b838bSDaniel Lee u32 need = foc->len; 9487f9b838bSDaniel Lee 9497f9b838bSDaniel Lee need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE : 9507f9b838bSDaniel Lee TCPOLEN_FASTOPEN_BASE; 9518336886fSJerry Chu need = (need + 3) & ~3U; /* Align to 32 bits */ 9528336886fSJerry Chu if (remaining >= need) { 9538336886fSJerry Chu opts->options |= OPTION_FAST_OPEN_COOKIE; 9548336886fSJerry Chu opts->fastopen_cookie = foc; 9558336886fSJerry Chu remaining -= need; 9568336886fSJerry Chu } 9578336886fSJerry Chu } 9584957faadSWilliam Allen Simpson 959cec37a6eSPeter Krystad mptcp_set_option_cond(req, opts, &remaining); 960cec37a6eSPeter Krystad 96160e2a778SUrsula Braun smc_set_option_cond(tcp_sk(sk), ireq, opts, &remaining); 96260e2a778SUrsula Braun 963331fca43SMartin KaFai Lau bpf_skops_hdr_opt_len((struct sock *)sk, skb, req, syn_skb, 964331fca43SMartin KaFai Lau synack_type, opts, &remaining); 965331fca43SMartin KaFai Lau 9664957faadSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 96733ad798cSAdam Langley } 96833ad798cSAdam Langley 96967edfef7SAndi Kleen /* Compute TCP options for ESTABLISHED sockets. This is not the 97067edfef7SAndi Kleen * final wire format yet. 97167edfef7SAndi Kleen */ 97295c96174SEric Dumazet static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb, 97333ad798cSAdam Langley struct tcp_out_options *opts, 9741e03d32bSDmitry Safonov struct tcp_key *key) 975cf533ea5SEric Dumazet { 97633ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 97795c96174SEric Dumazet unsigned int size = 0; 978cabeccbdSIlpo Järvinen unsigned int eff_sacks; 97933ad798cSAdam Langley 9805843ef42SAndi Kleen opts->options = 0; 9815843ef42SAndi Kleen 9821e03d32bSDmitry Safonov /* Better than switch (key.type) as it has static branches */ 9831e03d32bSDmitry Safonov if (tcp_key_is_md5(key)) { 98433ad798cSAdam Langley opts->options |= OPTION_MD5; 98533ad798cSAdam Langley size += TCPOLEN_MD5SIG_ALIGNED; 9861e03d32bSDmitry Safonov } else if (tcp_key_is_ao(key)) { 9871e03d32bSDmitry Safonov opts->options |= OPTION_AO; 988da7dfaa6SDmitry Safonov size += tcp_ao_len_aligned(key->ao_key); 98933ad798cSAdam Langley } 99033ad798cSAdam Langley 99133ad798cSAdam Langley if (likely(tp->rx_opt.tstamp_ok)) { 99233ad798cSAdam Langley opts->options |= OPTION_TS; 993614e8316SEric Dumazet opts->tsval = skb ? tcp_skb_timestamp_ts(tp->tcp_usec_ts, skb) + 994614e8316SEric Dumazet tp->tsoffset : 0; 99533ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 99633ad798cSAdam Langley size += TCPOLEN_TSTAMP_ALIGNED; 99733ad798cSAdam Langley } 99833ad798cSAdam Langley 999cec37a6eSPeter Krystad /* MPTCP options have precedence over SACK for the limited TCP 1000cec37a6eSPeter Krystad * option space because a MPTCP connection would be forced to 1001cec37a6eSPeter Krystad * fall back to regular TCP if a required multipath option is 1002cec37a6eSPeter Krystad * missing. SACK still gets a chance to use whatever space is 1003cec37a6eSPeter Krystad * left. 1004cec37a6eSPeter Krystad */ 1005cec37a6eSPeter Krystad if (sk_is_mptcp(sk)) { 1006cec37a6eSPeter Krystad unsigned int remaining = MAX_TCP_OPTION_SPACE - size; 1007cec37a6eSPeter Krystad unsigned int opt_size = 0; 1008cec37a6eSPeter Krystad 1009cec37a6eSPeter Krystad if (mptcp_established_options(sk, skb, &opt_size, remaining, 1010cec37a6eSPeter Krystad &opts->mptcp)) { 1011cec37a6eSPeter Krystad opts->options |= OPTION_MPTCP; 1012cec37a6eSPeter Krystad size += opt_size; 1013cec37a6eSPeter Krystad } 1014cec37a6eSPeter Krystad } 1015cec37a6eSPeter Krystad 1016cabeccbdSIlpo Järvinen eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; 1017cabeccbdSIlpo Järvinen if (unlikely(eff_sacks)) { 101895c96174SEric Dumazet const unsigned int remaining = MAX_TCP_OPTION_SPACE - size; 10199cfcca23SMat Martineau if (unlikely(remaining < TCPOLEN_SACK_BASE_ALIGNED + 10209cfcca23SMat Martineau TCPOLEN_SACK_PERBLOCK)) 10219cfcca23SMat Martineau return size; 10229cfcca23SMat Martineau 102333ad798cSAdam Langley opts->num_sack_blocks = 102495c96174SEric Dumazet min_t(unsigned int, eff_sacks, 102533ad798cSAdam Langley (remaining - TCPOLEN_SACK_BASE_ALIGNED) / 102633ad798cSAdam Langley TCPOLEN_SACK_PERBLOCK); 10279cfcca23SMat Martineau 102833ad798cSAdam Langley size += TCPOLEN_SACK_BASE_ALIGNED + 102933ad798cSAdam Langley opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; 103033ad798cSAdam Langley } 103133ad798cSAdam Langley 1032331fca43SMartin KaFai Lau if (unlikely(BPF_SOCK_OPS_TEST_FLAG(tp, 1033331fca43SMartin KaFai Lau BPF_SOCK_OPS_WRITE_HDR_OPT_CB_FLAG))) { 1034331fca43SMartin KaFai Lau unsigned int remaining = MAX_TCP_OPTION_SPACE - size; 1035331fca43SMartin KaFai Lau 1036331fca43SMartin KaFai Lau bpf_skops_hdr_opt_len(sk, skb, NULL, NULL, 0, opts, &remaining); 1037331fca43SMartin KaFai Lau 1038331fca43SMartin KaFai Lau size = MAX_TCP_OPTION_SPACE - remaining; 1039331fca43SMartin KaFai Lau } 1040331fca43SMartin KaFai Lau 104133ad798cSAdam Langley return size; 104240efc6faSStephen Hemminger } 10431da177e4SLinus Torvalds 104446d3ceabSEric Dumazet 104546d3ceabSEric Dumazet /* TCP SMALL QUEUES (TSQ) 104646d3ceabSEric Dumazet * 104746d3ceabSEric Dumazet * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev) 104846d3ceabSEric Dumazet * to reduce RTT and bufferbloat. 104946d3ceabSEric Dumazet * We do this using a special skb destructor (tcp_wfree). 105046d3ceabSEric Dumazet * 105146d3ceabSEric Dumazet * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb 105246d3ceabSEric Dumazet * needs to be reallocated in a driver. 10538e3bff96Sstephen hemminger * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc 105446d3ceabSEric Dumazet * 105546d3ceabSEric Dumazet * Since transmit from skb destructor is forbidden, we use a tasklet 105646d3ceabSEric Dumazet * to process all sockets that eventually need to send more skbs. 105746d3ceabSEric Dumazet * We use one tasklet per cpu, with its own queue of sockets. 105846d3ceabSEric Dumazet */ 105946d3ceabSEric Dumazet struct tsq_tasklet { 106046d3ceabSEric Dumazet struct tasklet_struct tasklet; 106146d3ceabSEric Dumazet struct list_head head; /* queue of tcp sockets */ 106246d3ceabSEric Dumazet }; 106346d3ceabSEric Dumazet static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet); 106446d3ceabSEric Dumazet 106573a6bab5SEric Dumazet static void tcp_tsq_write(struct sock *sk) 10666f458dfbSEric Dumazet { 10676f458dfbSEric Dumazet if ((1 << sk->sk_state) & 10686f458dfbSEric Dumazet (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING | 1069f9616c35SEric Dumazet TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) { 1070f9616c35SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 1071f9616c35SEric Dumazet 1072f9616c35SEric Dumazet if (tp->lost_out > tp->retrans_out && 107340570375SEric Dumazet tcp_snd_cwnd(tp) > tcp_packets_in_flight(tp)) { 10743a91d29fSKoichiro Den tcp_mstamp_refresh(tp); 1075f9616c35SEric Dumazet tcp_xmit_retransmit_queue(sk); 10763a91d29fSKoichiro Den } 1077f9616c35SEric Dumazet 1078f9616c35SEric Dumazet tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle, 1079bf06200eSJohn Ogness 0, GFP_ATOMIC); 10806f458dfbSEric Dumazet } 1081f9616c35SEric Dumazet } 108273a6bab5SEric Dumazet 108373a6bab5SEric Dumazet static void tcp_tsq_handler(struct sock *sk) 108473a6bab5SEric Dumazet { 108573a6bab5SEric Dumazet bh_lock_sock(sk); 108673a6bab5SEric Dumazet if (!sock_owned_by_user(sk)) 108773a6bab5SEric Dumazet tcp_tsq_write(sk); 108873a6bab5SEric Dumazet else if (!test_and_set_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags)) 108973a6bab5SEric Dumazet sock_hold(sk); 109073a6bab5SEric Dumazet bh_unlock_sock(sk); 109173a6bab5SEric Dumazet } 109246d3ceabSEric Dumazet /* 10938e3bff96Sstephen hemminger * One tasklet per cpu tries to send more skbs. 109446d3ceabSEric Dumazet * We run in tasklet context but need to disable irqs when 10958e3bff96Sstephen hemminger * transferring tsq->head because tcp_wfree() might 109646d3ceabSEric Dumazet * interrupt us (non NAPI drivers) 109746d3ceabSEric Dumazet */ 1098c6533ca8SAllen Pais static void tcp_tasklet_func(struct tasklet_struct *t) 109946d3ceabSEric Dumazet { 1100c6533ca8SAllen Pais struct tsq_tasklet *tsq = from_tasklet(tsq, t, tasklet); 110146d3ceabSEric Dumazet LIST_HEAD(list); 110246d3ceabSEric Dumazet unsigned long flags; 110346d3ceabSEric Dumazet struct list_head *q, *n; 110446d3ceabSEric Dumazet struct tcp_sock *tp; 110546d3ceabSEric Dumazet struct sock *sk; 110646d3ceabSEric Dumazet 110746d3ceabSEric Dumazet local_irq_save(flags); 110846d3ceabSEric Dumazet list_splice_init(&tsq->head, &list); 110946d3ceabSEric Dumazet local_irq_restore(flags); 111046d3ceabSEric Dumazet 111146d3ceabSEric Dumazet list_for_each_safe(q, n, &list) { 111246d3ceabSEric Dumazet tp = list_entry(q, struct tcp_sock, tsq_node); 111346d3ceabSEric Dumazet list_del(&tp->tsq_node); 111446d3ceabSEric Dumazet 111546d3ceabSEric Dumazet sk = (struct sock *)tp; 11160a9648f1SEric Dumazet smp_mb__before_atomic(); 11177aa5470cSEric Dumazet clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags); 11187aa5470cSEric Dumazet 11196f458dfbSEric Dumazet tcp_tsq_handler(sk); 112046d3ceabSEric Dumazet sk_free(sk); 112146d3ceabSEric Dumazet } 112246d3ceabSEric Dumazet } 112346d3ceabSEric Dumazet 112440fc3423SEric Dumazet #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \ 112540fc3423SEric Dumazet TCPF_WRITE_TIMER_DEFERRED | \ 112640fc3423SEric Dumazet TCPF_DELACK_TIMER_DEFERRED | \ 1127133c4c0dSEric Dumazet TCPF_MTU_REDUCED_DEFERRED | \ 1128133c4c0dSEric Dumazet TCPF_ACK_DEFERRED) 112946d3ceabSEric Dumazet /** 113046d3ceabSEric Dumazet * tcp_release_cb - tcp release_sock() callback 113146d3ceabSEric Dumazet * @sk: socket 113246d3ceabSEric Dumazet * 113346d3ceabSEric Dumazet * called from release_sock() to perform protocol dependent 113446d3ceabSEric Dumazet * actions before socket release. 113546d3ceabSEric Dumazet */ 113646d3ceabSEric Dumazet void tcp_release_cb(struct sock *sk) 113746d3ceabSEric Dumazet { 1138fac30731SEric Dumazet unsigned long flags = smp_load_acquire(&sk->sk_tsq_flags); 1139fac30731SEric Dumazet unsigned long nflags; 114046d3ceabSEric Dumazet 11416f458dfbSEric Dumazet /* perform an atomic operation only if at least one flag is set */ 11426f458dfbSEric Dumazet do { 11436f458dfbSEric Dumazet if (!(flags & TCP_DEFERRED_ALL)) 11446f458dfbSEric Dumazet return; 11456f458dfbSEric Dumazet nflags = flags & ~TCP_DEFERRED_ALL; 1146fac30731SEric Dumazet } while (!try_cmpxchg(&sk->sk_tsq_flags, &flags, nflags)); 11476f458dfbSEric Dumazet 114873a6bab5SEric Dumazet if (flags & TCPF_TSQ_DEFERRED) { 114973a6bab5SEric Dumazet tcp_tsq_write(sk); 115073a6bab5SEric Dumazet __sock_put(sk); 115173a6bab5SEric Dumazet } 1152c3f9b018SEric Dumazet 115340fc3423SEric Dumazet if (flags & TCPF_WRITE_TIMER_DEFERRED) { 11546f458dfbSEric Dumazet tcp_write_timer_handler(sk); 1155144d56e9SEric Dumazet __sock_put(sk); 1156144d56e9SEric Dumazet } 115740fc3423SEric Dumazet if (flags & TCPF_DELACK_TIMER_DEFERRED) { 11586f458dfbSEric Dumazet tcp_delack_timer_handler(sk); 1159144d56e9SEric Dumazet __sock_put(sk); 1160144d56e9SEric Dumazet } 116140fc3423SEric Dumazet if (flags & TCPF_MTU_REDUCED_DEFERRED) { 11624fab9071SNeal Cardwell inet_csk(sk)->icsk_af_ops->mtu_reduced(sk); 1163144d56e9SEric Dumazet __sock_put(sk); 1164144d56e9SEric Dumazet } 1165133c4c0dSEric Dumazet if ((flags & TCPF_ACK_DEFERRED) && inet_csk_ack_scheduled(sk)) 1166133c4c0dSEric Dumazet tcp_send_ack(sk); 116746d3ceabSEric Dumazet } 116846d3ceabSEric Dumazet EXPORT_SYMBOL(tcp_release_cb); 116946d3ceabSEric Dumazet 117046d3ceabSEric Dumazet void __init tcp_tasklet_init(void) 117146d3ceabSEric Dumazet { 117246d3ceabSEric Dumazet int i; 117346d3ceabSEric Dumazet 117446d3ceabSEric Dumazet for_each_possible_cpu(i) { 117546d3ceabSEric Dumazet struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i); 117646d3ceabSEric Dumazet 117746d3ceabSEric Dumazet INIT_LIST_HEAD(&tsq->head); 1178c6533ca8SAllen Pais tasklet_setup(&tsq->tasklet, tcp_tasklet_func); 117946d3ceabSEric Dumazet } 118046d3ceabSEric Dumazet } 118146d3ceabSEric Dumazet 118246d3ceabSEric Dumazet /* 118346d3ceabSEric Dumazet * Write buffer destructor automatically called from kfree_skb. 11848e3bff96Sstephen hemminger * We can't xmit new skbs from this context, as we might already 118546d3ceabSEric Dumazet * hold qdisc lock. 118646d3ceabSEric Dumazet */ 1187d6a4a104SEric Dumazet void tcp_wfree(struct sk_buff *skb) 118846d3ceabSEric Dumazet { 118946d3ceabSEric Dumazet struct sock *sk = skb->sk; 119046d3ceabSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 1191408f0a6cSEric Dumazet unsigned long flags, nval, oval; 1192b548b17aSEric Dumazet struct tsq_tasklet *tsq; 1193b548b17aSEric Dumazet bool empty; 11949b462d02SEric Dumazet 11959b462d02SEric Dumazet /* Keep one reference on sk_wmem_alloc. 11969b462d02SEric Dumazet * Will be released by sk_free() from here or tcp_tasklet_func() 11979b462d02SEric Dumazet */ 119814afee4bSReshetova, Elena WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc)); 11999b462d02SEric Dumazet 12009b462d02SEric Dumazet /* If this softirq is serviced by ksoftirqd, we are likely under stress. 12019b462d02SEric Dumazet * Wait until our queues (qdisc + devices) are drained. 12029b462d02SEric Dumazet * This gives : 12039b462d02SEric Dumazet * - less callbacks to tcp_write_xmit(), reducing stress (batches) 12049b462d02SEric Dumazet * - chance for incoming ACK (processed by another cpu maybe) 12059b462d02SEric Dumazet * to migrate this flow (skb->ooo_okay will be eventually set) 12069b462d02SEric Dumazet */ 120714afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current) 12089b462d02SEric Dumazet goto out; 120946d3ceabSEric Dumazet 1210b548b17aSEric Dumazet oval = smp_load_acquire(&sk->sk_tsq_flags); 1211b548b17aSEric Dumazet do { 1212408f0a6cSEric Dumazet if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED)) 1213408f0a6cSEric Dumazet goto out; 1214408f0a6cSEric Dumazet 121573a6bab5SEric Dumazet nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED; 1216b548b17aSEric Dumazet } while (!try_cmpxchg(&sk->sk_tsq_flags, &oval, nval)); 1217408f0a6cSEric Dumazet 121846d3ceabSEric Dumazet /* queue this socket to tasklet queue */ 121946d3ceabSEric Dumazet local_irq_save(flags); 1220903ceff7SChristoph Lameter tsq = this_cpu_ptr(&tsq_tasklet); 1221a9b204d1SEric Dumazet empty = list_empty(&tsq->head); 122246d3ceabSEric Dumazet list_add(&tp->tsq_node, &tsq->head); 1223a9b204d1SEric Dumazet if (empty) 122446d3ceabSEric Dumazet tasklet_schedule(&tsq->tasklet); 122546d3ceabSEric Dumazet local_irq_restore(flags); 12269b462d02SEric Dumazet return; 12279b462d02SEric Dumazet out: 12289b462d02SEric Dumazet sk_free(sk); 122946d3ceabSEric Dumazet } 123046d3ceabSEric Dumazet 123173a6bab5SEric Dumazet /* Note: Called under soft irq. 123273a6bab5SEric Dumazet * We can call TCP stack right away, unless socket is owned by user. 1233218af599SEric Dumazet */ 1234218af599SEric Dumazet enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer) 1235218af599SEric Dumazet { 1236218af599SEric Dumazet struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer); 1237218af599SEric Dumazet struct sock *sk = (struct sock *)tp; 1238218af599SEric Dumazet 123973a6bab5SEric Dumazet tcp_tsq_handler(sk); 124073a6bab5SEric Dumazet sock_put(sk); 1241218af599SEric Dumazet 1242218af599SEric Dumazet return HRTIMER_NORESTART; 1243218af599SEric Dumazet } 1244218af599SEric Dumazet 1245a7a25630SEric Dumazet static void tcp_update_skb_after_send(struct sock *sk, struct sk_buff *skb, 1246a7a25630SEric Dumazet u64 prior_wstamp) 1247e2080072SEric Dumazet { 1248ab408b6dSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 1249ab408b6dSEric Dumazet 1250ab408b6dSEric Dumazet if (sk->sk_pacing_status != SK_PACING_NONE) { 125128b24f90SEric Dumazet unsigned long rate = READ_ONCE(sk->sk_pacing_rate); 1252ab408b6dSEric Dumazet 1253ab408b6dSEric Dumazet /* Original sch_fq does not pace first 10 MSS 1254ab408b6dSEric Dumazet * Note that tp->data_segs_out overflows after 2^32 packets, 1255ab408b6dSEric Dumazet * this is a minor annoyance. 1256ab408b6dSEric Dumazet */ 125776a9ebe8SEric Dumazet if (rate != ~0UL && rate && tp->data_segs_out >= 10) { 1258a7a25630SEric Dumazet u64 len_ns = div64_ul((u64)skb->len * NSEC_PER_SEC, rate); 1259a7a25630SEric Dumazet u64 credit = tp->tcp_wstamp_ns - prior_wstamp; 1260a7a25630SEric Dumazet 1261a7a25630SEric Dumazet /* take into account OS jitter */ 1262a7a25630SEric Dumazet len_ns -= min_t(u64, len_ns / 2, credit); 1263a7a25630SEric Dumazet tp->tcp_wstamp_ns += len_ns; 1264ab408b6dSEric Dumazet } 1265ab408b6dSEric Dumazet } 1266e2080072SEric Dumazet list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue); 1267e2080072SEric Dumazet } 1268e2080072SEric Dumazet 126905e22e83SEric Dumazet INDIRECT_CALLABLE_DECLARE(int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl)); 127005e22e83SEric Dumazet INDIRECT_CALLABLE_DECLARE(int inet6_csk_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl)); 1271dd2e0b86SEric Dumazet INDIRECT_CALLABLE_DECLARE(void tcp_v4_send_check(struct sock *sk, struct sk_buff *skb)); 127205e22e83SEric Dumazet 12731da177e4SLinus Torvalds /* This routine actually transmits TCP packets queued in by 12741da177e4SLinus Torvalds * tcp_do_sendmsg(). This is used by both the initial 12751da177e4SLinus Torvalds * transmission and possible later retransmissions. 12761da177e4SLinus Torvalds * All SKB's seen here are completely headerless. It is our 12771da177e4SLinus Torvalds * job to build the TCP header, and pass the packet down to 12781da177e4SLinus Torvalds * IP so it can do the same plus pass the packet off to the 12791da177e4SLinus Torvalds * device. 12801da177e4SLinus Torvalds * 12811da177e4SLinus Torvalds * We are working here with either a clone of the original 12821da177e4SLinus Torvalds * SKB, or a fresh unique copy made by the retransmit engine. 12831da177e4SLinus Torvalds */ 12842987babbSYuchung Cheng static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, 12852987babbSYuchung Cheng int clone_it, gfp_t gfp_mask, u32 rcv_nxt) 12861da177e4SLinus Torvalds { 12876687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 1288dfb4b9dcSDavid S. Miller struct inet_sock *inet; 1289dfb4b9dcSDavid S. Miller struct tcp_sock *tp; 1290dfb4b9dcSDavid S. Miller struct tcp_skb_cb *tcb; 129133ad798cSAdam Langley struct tcp_out_options opts; 129295c96174SEric Dumazet unsigned int tcp_options_size, tcp_header_size; 12938c72c65bSEric Dumazet struct sk_buff *oskb = NULL; 12941e03d32bSDmitry Safonov struct tcp_key key; 12951da177e4SLinus Torvalds struct tcphdr *th; 1296a7a25630SEric Dumazet u64 prior_wstamp; 12971da177e4SLinus Torvalds int err; 12981da177e4SLinus Torvalds 1299dfb4b9dcSDavid S. Miller BUG_ON(!skb || !tcp_skb_pcount(skb)); 13006f094b9eSLawrence Brakmo tp = tcp_sk(sk); 13017f12422cSYuchung Cheng prior_wstamp = tp->tcp_wstamp_ns; 13027f12422cSYuchung Cheng tp->tcp_wstamp_ns = max(tp->tcp_wstamp_ns, tp->tcp_clock_cache); 1303a1ac9c8aSMartin KaFai Lau skb_set_delivery_time(skb, tp->tcp_wstamp_ns, true); 1304ccdbb6e9SEric Dumazet if (clone_it) { 13058c72c65bSEric Dumazet oskb = skb; 1306e2080072SEric Dumazet 1307e2080072SEric Dumazet tcp_skb_tsorted_save(oskb) { 1308e2080072SEric Dumazet if (unlikely(skb_cloned(oskb))) 1309e2080072SEric Dumazet skb = pskb_copy(oskb, gfp_mask); 1310dfb4b9dcSDavid S. Miller else 1311e2080072SEric Dumazet skb = skb_clone(oskb, gfp_mask); 1312e2080072SEric Dumazet } tcp_skb_tsorted_restore(oskb); 1313e2080072SEric Dumazet 1314dfb4b9dcSDavid S. Miller if (unlikely(!skb)) 1315dfb4b9dcSDavid S. Miller return -ENOBUFS; 1316b738a185SEric Dumazet /* retransmit skbs might have a non zero value in skb->dev 1317b738a185SEric Dumazet * because skb->dev is aliased with skb->rbnode.rb_left 1318b738a185SEric Dumazet */ 1319b738a185SEric Dumazet skb->dev = NULL; 1320dfb4b9dcSDavid S. Miller } 13215f6188a8SEric Dumazet 1322dfb4b9dcSDavid S. Miller inet = inet_sk(sk); 1323dfb4b9dcSDavid S. Miller tcb = TCP_SKB_CB(skb); 132433ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 13251da177e4SLinus Torvalds 13261e03d32bSDmitry Safonov tcp_get_current_key(sk, &key); 1327051ba674SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) { 13281e03d32bSDmitry Safonov tcp_options_size = tcp_syn_options(sk, skb, &opts, &key); 1329051ba674SEric Dumazet } else { 13301e03d32bSDmitry Safonov tcp_options_size = tcp_established_options(sk, skb, &opts, &key); 1331051ba674SEric Dumazet /* Force a PSH flag on all (GSO) packets to expedite GRO flush 1332051ba674SEric Dumazet * at receiver : This slightly improve GRO performance. 1333051ba674SEric Dumazet * Note that we do not force the PSH flag for non GSO packets, 1334051ba674SEric Dumazet * because they might be sent under high congestion events, 1335051ba674SEric Dumazet * and in this case it is better to delay the delivery of 1-MSS 1336051ba674SEric Dumazet * packets and thus the corresponding ACK packet that would 1337051ba674SEric Dumazet * release the following packet. 1338051ba674SEric Dumazet */ 1339051ba674SEric Dumazet if (tcp_skb_pcount(skb) > 1) 1340051ba674SEric Dumazet tcb->tcp_flags |= TCPHDR_PSH; 1341051ba674SEric Dumazet } 134233ad798cSAdam Langley tcp_header_size = tcp_options_size + sizeof(struct tcphdr); 13431da177e4SLinus Torvalds 1344726e9e8bSEric Dumazet /* We set skb->ooo_okay to one if this packet can select 1345726e9e8bSEric Dumazet * a different TX queue than prior packets of this flow, 1346726e9e8bSEric Dumazet * to avoid self inflicted reorders. 1347726e9e8bSEric Dumazet * The 'other' queue decision is based on current cpu number 1348726e9e8bSEric Dumazet * if XPS is enabled, or sk->sk_txhash otherwise. 1349726e9e8bSEric Dumazet * We can switch to another (and better) queue if: 1350726e9e8bSEric Dumazet * 1) No packet with payload is in qdisc/device queues. 1351726e9e8bSEric Dumazet * Delays in TX completion can defeat the test 1352726e9e8bSEric Dumazet * even if packets were already sent. 1353726e9e8bSEric Dumazet * 2) Or rtx queue is empty. 1354726e9e8bSEric Dumazet * This mitigates above case if ACK packets for 1355726e9e8bSEric Dumazet * all prior packets were already processed. 1356547669d4SEric Dumazet */ 1357726e9e8bSEric Dumazet skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1) || 1358726e9e8bSEric Dumazet tcp_rtx_queue_empty(sk); 13591da177e4SLinus Torvalds 136038ab52e8SEric Dumazet /* If we had to use memory reserve to allocate this skb, 136138ab52e8SEric Dumazet * this might cause drops if packet is looped back : 136238ab52e8SEric Dumazet * Other socket might not have SOCK_MEMALLOC. 136338ab52e8SEric Dumazet * Packets not looped back do not care about pfmemalloc. 136438ab52e8SEric Dumazet */ 136538ab52e8SEric Dumazet skb->pfmemalloc = 0; 136638ab52e8SEric Dumazet 1367aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 1368aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 136946d3ceabSEric Dumazet 137046d3ceabSEric Dumazet skb_orphan(skb); 137146d3ceabSEric Dumazet skb->sk = sk; 13721d2077acSEric Dumazet skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree; 137314afee4bSReshetova, Elena refcount_add(skb->truesize, &sk->sk_wmem_alloc); 13741da177e4SLinus Torvalds 1375eb44ad4eSEric Dumazet skb_set_dst_pending_confirm(skb, READ_ONCE(sk->sk_dst_pending_confirm)); 1376c3a2e837SJulian Anastasov 13771da177e4SLinus Torvalds /* Build TCP header and checksum it. */ 1378ea1627c2SEric Dumazet th = (struct tcphdr *)skb->data; 1379c720c7e8SEric Dumazet th->source = inet->inet_sport; 1380c720c7e8SEric Dumazet th->dest = inet->inet_dport; 13811da177e4SLinus Torvalds th->seq = htonl(tcb->seq); 13822987babbSYuchung Cheng th->ack_seq = htonl(rcv_nxt); 1383df7a3b07SAl Viro *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | 13844de075e0SEric Dumazet tcb->tcp_flags); 1385dfb4b9dcSDavid S. Miller 13861da177e4SLinus Torvalds th->check = 0; 13871da177e4SLinus Torvalds th->urg_ptr = 0; 13881da177e4SLinus Torvalds 138933f5f57eSIlpo Järvinen /* The urg_mode check is necessary during a below snd_una win probe */ 13907691367dSHerbert Xu if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { 13917691367dSHerbert Xu if (before(tp->snd_up, tcb->seq + 0x10000)) { 13921da177e4SLinus Torvalds th->urg_ptr = htons(tp->snd_up - tcb->seq); 13931da177e4SLinus Torvalds th->urg = 1; 13947691367dSHerbert Xu } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { 13950eae88f3SEric Dumazet th->urg_ptr = htons(0xFFFF); 13967691367dSHerbert Xu th->urg = 1; 13977691367dSHerbert Xu } 13981da177e4SLinus Torvalds } 13991da177e4SLinus Torvalds 140051466a75SEric Dumazet skb_shinfo(skb)->gso_type = sk->sk_gso_type; 1401ea1627c2SEric Dumazet if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) { 1402ea1627c2SEric Dumazet th->window = htons(tcp_select_window(sk)); 1403ea1627c2SEric Dumazet tcp_ecn_send(sk, skb, th, tcp_header_size); 1404ea1627c2SEric Dumazet } else { 1405ea1627c2SEric Dumazet /* RFC1323: The window in SYN & SYN/ACK segments 1406ea1627c2SEric Dumazet * is never scaled. 1407ea1627c2SEric Dumazet */ 1408ea1627c2SEric Dumazet th->window = htons(min(tp->rcv_wnd, 65535U)); 1409ea1627c2SEric Dumazet } 1410fa3fe2b1SFlorian Westphal 141106b22ef2SDmitry Safonov tcp_options_write(th, tp, NULL, &opts, &key); 1412fa3fe2b1SFlorian Westphal 14131e03d32bSDmitry Safonov if (tcp_key_is_md5(&key)) { 1414cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 1415cfb6eeb4SYOSHIFUJI Hideaki /* Calculate the MD5 hash, as we have all we need now */ 1416aba54656SEric Dumazet sk_gso_disable(sk); 1417bd0388aeSWilliam Allen Simpson tp->af_specific->calc_md5_hash(opts.hash_location, 14181e03d32bSDmitry Safonov key.md5_key, sk, skb); 1419cfb6eeb4SYOSHIFUJI Hideaki #endif 14201e03d32bSDmitry Safonov } else if (tcp_key_is_ao(&key)) { 14211e03d32bSDmitry Safonov int err; 14221e03d32bSDmitry Safonov 14231e03d32bSDmitry Safonov err = tcp_ao_transmit_skb(sk, skb, key.ao_key, th, 14241e03d32bSDmitry Safonov opts.hash_location); 14251e03d32bSDmitry Safonov if (err) { 14261e03d32bSDmitry Safonov kfree_skb_reason(skb, SKB_DROP_REASON_NOT_SPECIFIED); 14271e03d32bSDmitry Safonov return -ENOMEM; 14281e03d32bSDmitry Safonov } 14291e03d32bSDmitry Safonov } 1430cfb6eeb4SYOSHIFUJI Hideaki 1431331fca43SMartin KaFai Lau /* BPF prog is the last one writing header option */ 1432331fca43SMartin KaFai Lau bpf_skops_write_hdr_opt(sk, skb, NULL, NULL, 0, &opts); 1433331fca43SMartin KaFai Lau 1434dd2e0b86SEric Dumazet INDIRECT_CALL_INET(icsk->icsk_af_ops->send_check, 1435dd2e0b86SEric Dumazet tcp_v6_send_check, tcp_v4_send_check, 1436dd2e0b86SEric Dumazet sk, skb); 14371da177e4SLinus Torvalds 14384de075e0SEric Dumazet if (likely(tcb->tcp_flags & TCPHDR_ACK)) 1439059217c1SNeal Cardwell tcp_event_ack_sent(sk, rcv_nxt); 14401da177e4SLinus Torvalds 1441a44d6eacSMartin KaFai Lau if (skb->len != tcp_header_size) { 1442cf533ea5SEric Dumazet tcp_event_data_sent(tp, sk); 1443a44d6eacSMartin KaFai Lau tp->data_segs_out += tcp_skb_pcount(skb); 1444ba113c3aSWei Wang tp->bytes_sent += skb->len - tcp_header_size; 1445a44d6eacSMartin KaFai Lau } 14461da177e4SLinus Torvalds 1447bd37a088SWei Yongjun if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) 1448aa2ea058STom Herbert TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, 1449aa2ea058STom Herbert tcp_skb_pcount(skb)); 14501da177e4SLinus Torvalds 14512efd055cSMarcelo Ricardo Leitner tp->segs_out += tcp_skb_pcount(skb); 14520ae5b43dSYuchung Cheng skb_set_hash_from_sk(skb, sk); 1453f69ad292SEric Dumazet /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */ 1454cd7d8498SEric Dumazet skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb); 1455f69ad292SEric Dumazet skb_shinfo(skb)->gso_size = tcp_skb_mss(skb); 1456cd7d8498SEric Dumazet 1457d3edd06eSEric Dumazet /* Leave earliest departure time in skb->tstamp (skb->skb_mstamp_ns) */ 1458971f10ecSEric Dumazet 1459971f10ecSEric Dumazet /* Cleanup our debris for IP stacks */ 1460971f10ecSEric Dumazet memset(skb->cb, 0, max(sizeof(struct inet_skb_parm), 1461971f10ecSEric Dumazet sizeof(struct inet6_skb_parm))); 1462971f10ecSEric Dumazet 1463a842fe14SEric Dumazet tcp_add_tx_delay(skb, tp); 1464a842fe14SEric Dumazet 146505e22e83SEric Dumazet err = INDIRECT_CALL_INET(icsk->icsk_af_ops->queue_xmit, 146605e22e83SEric Dumazet inet6_csk_xmit, ip_queue_xmit, 146705e22e83SEric Dumazet sk, skb, &inet->cork.fl); 14687faee5c0SEric Dumazet 14698c72c65bSEric Dumazet if (unlikely(err > 0)) { 14705ee2c941SChristoph Paasch tcp_enter_cwr(sk); 14718c72c65bSEric Dumazet err = net_xmit_eval(err); 14728c72c65bSEric Dumazet } 1473fc225799SEric Dumazet if (!err && oskb) { 1474a7a25630SEric Dumazet tcp_update_skb_after_send(sk, oskb, prior_wstamp); 1475fc225799SEric Dumazet tcp_rate_skb_sent(sk, oskb); 1476fc225799SEric Dumazet } 14778c72c65bSEric Dumazet return err; 14781da177e4SLinus Torvalds } 14791da177e4SLinus Torvalds 14802987babbSYuchung Cheng static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, 14812987babbSYuchung Cheng gfp_t gfp_mask) 14822987babbSYuchung Cheng { 14832987babbSYuchung Cheng return __tcp_transmit_skb(sk, skb, clone_it, gfp_mask, 14842987babbSYuchung Cheng tcp_sk(sk)->rcv_nxt); 14852987babbSYuchung Cheng } 14862987babbSYuchung Cheng 148767edfef7SAndi Kleen /* This routine just queues the buffer for sending. 14881da177e4SLinus Torvalds * 14891da177e4SLinus Torvalds * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, 14901da177e4SLinus Torvalds * otherwise socket can stall. 14911da177e4SLinus Torvalds */ 14921da177e4SLinus Torvalds static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) 14931da177e4SLinus Torvalds { 14941da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 14951da177e4SLinus Torvalds 14961da177e4SLinus Torvalds /* Advance write_seq and place onto the write_queue. */ 14970f317464SEric Dumazet WRITE_ONCE(tp->write_seq, TCP_SKB_CB(skb)->end_seq); 1498f4a775d1SEric Dumazet __skb_header_release(skb); 1499fe067e8aSDavid S. Miller tcp_add_write_queue_tail(sk, skb); 1500ab4e846aSEric Dumazet sk_wmem_queued_add(sk, skb->truesize); 15013ab224beSHideo Aoki sk_mem_charge(sk, skb->truesize); 15021da177e4SLinus Torvalds } 15031da177e4SLinus Torvalds 150467edfef7SAndi Kleen /* Initialize TSO segments for a packet. */ 1505d5b38a71SEric Dumazet static int tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now) 1506f6302d1dSDavid S. Miller { 1507d5b38a71SEric Dumazet int tso_segs; 1508d5b38a71SEric Dumazet 15094a64fd6cSEric Dumazet if (skb->len <= mss_now) { 1510f6302d1dSDavid S. Miller /* Avoid the costly divide in the normal 1511f6302d1dSDavid S. Miller * non-TSO case. 1512f6302d1dSDavid S. Miller */ 1513f69ad292SEric Dumazet TCP_SKB_CB(skb)->tcp_gso_size = 0; 1514d5b38a71SEric Dumazet tcp_skb_pcount_set(skb, 1); 1515d5b38a71SEric Dumazet return 1; 15161da177e4SLinus Torvalds } 1517d5b38a71SEric Dumazet TCP_SKB_CB(skb)->tcp_gso_size = mss_now; 1518d5b38a71SEric Dumazet tso_segs = DIV_ROUND_UP(skb->len, mss_now); 1519d5b38a71SEric Dumazet tcp_skb_pcount_set(skb, tso_segs); 1520d5b38a71SEric Dumazet return tso_segs; 15211da177e4SLinus Torvalds } 15221da177e4SLinus Torvalds 1523797108d1SIlpo Järvinen /* Pcount in the middle of the write queue got changed, we need to do various 1524797108d1SIlpo Järvinen * tweaks to fix counters 1525797108d1SIlpo Järvinen */ 1526cf533ea5SEric Dumazet static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) 1527797108d1SIlpo Järvinen { 1528797108d1SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1529797108d1SIlpo Järvinen 1530797108d1SIlpo Järvinen tp->packets_out -= decr; 1531797108d1SIlpo Järvinen 1532797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 1533797108d1SIlpo Järvinen tp->sacked_out -= decr; 1534797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 1535797108d1SIlpo Järvinen tp->retrans_out -= decr; 1536797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 1537797108d1SIlpo Järvinen tp->lost_out -= decr; 1538797108d1SIlpo Järvinen 1539797108d1SIlpo Järvinen /* Reno case is special. Sigh... */ 1540797108d1SIlpo Järvinen if (tcp_is_reno(tp) && decr > 0) 1541797108d1SIlpo Järvinen tp->sacked_out -= min_t(u32, tp->sacked_out, decr); 1542797108d1SIlpo Järvinen 1543797108d1SIlpo Järvinen if (tp->lost_skb_hint && 1544797108d1SIlpo Järvinen before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && 1545713bafeaSYuchung Cheng (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) 1546797108d1SIlpo Järvinen tp->lost_cnt_hint -= decr; 1547797108d1SIlpo Järvinen 1548797108d1SIlpo Järvinen tcp_verify_left_out(tp); 1549797108d1SIlpo Järvinen } 1550797108d1SIlpo Järvinen 15510a2cf20cSSoheil Hassas Yeganeh static bool tcp_has_tx_tstamp(const struct sk_buff *skb) 15520a2cf20cSSoheil Hassas Yeganeh { 15530a2cf20cSSoheil Hassas Yeganeh return TCP_SKB_CB(skb)->txstamp_ack || 15540a2cf20cSSoheil Hassas Yeganeh (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP); 15550a2cf20cSSoheil Hassas Yeganeh } 15560a2cf20cSSoheil Hassas Yeganeh 1557490cc7d0SWillem de Bruijn static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2) 1558490cc7d0SWillem de Bruijn { 1559490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo = skb_shinfo(skb); 1560490cc7d0SWillem de Bruijn 15610a2cf20cSSoheil Hassas Yeganeh if (unlikely(tcp_has_tx_tstamp(skb)) && 1562490cc7d0SWillem de Bruijn !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) { 1563490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo2 = skb_shinfo(skb2); 1564490cc7d0SWillem de Bruijn u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP; 1565490cc7d0SWillem de Bruijn 1566490cc7d0SWillem de Bruijn shinfo->tx_flags &= ~tsflags; 1567490cc7d0SWillem de Bruijn shinfo2->tx_flags |= tsflags; 1568490cc7d0SWillem de Bruijn swap(shinfo->tskey, shinfo2->tskey); 1569b51e13faSMartin KaFai Lau TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack; 1570b51e13faSMartin KaFai Lau TCP_SKB_CB(skb)->txstamp_ack = 0; 1571490cc7d0SWillem de Bruijn } 1572490cc7d0SWillem de Bruijn } 1573490cc7d0SWillem de Bruijn 1574a166140eSMartin KaFai Lau static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2) 1575a166140eSMartin KaFai Lau { 1576a166140eSMartin KaFai Lau TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor; 1577a166140eSMartin KaFai Lau TCP_SKB_CB(skb)->eor = 0; 1578a166140eSMartin KaFai Lau } 1579a166140eSMartin KaFai Lau 158075c119afSEric Dumazet /* Insert buff after skb on the write or rtx queue of sk. */ 158175c119afSEric Dumazet static void tcp_insert_write_queue_after(struct sk_buff *skb, 158275c119afSEric Dumazet struct sk_buff *buff, 158375c119afSEric Dumazet struct sock *sk, 158475c119afSEric Dumazet enum tcp_queue tcp_queue) 158575c119afSEric Dumazet { 158675c119afSEric Dumazet if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE) 158775c119afSEric Dumazet __skb_queue_after(&sk->sk_write_queue, skb, buff); 158875c119afSEric Dumazet else 158975c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, buff); 159075c119afSEric Dumazet } 159175c119afSEric Dumazet 15921da177e4SLinus Torvalds /* Function to create two new TCP segments. Shrinks the given segment 15931da177e4SLinus Torvalds * to the specified size and appends a new segment with the rest of the 15941da177e4SLinus Torvalds * packet to the list. This won't be called frequently, I hope. 15951da177e4SLinus Torvalds * Remember, these are still headerless SKBs at this point. 15961da177e4SLinus Torvalds */ 159775c119afSEric Dumazet int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue, 159875c119afSEric Dumazet struct sk_buff *skb, u32 len, 15996cc55e09SOctavian Purdila unsigned int mss_now, gfp_t gfp) 16001da177e4SLinus Torvalds { 16011da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 16021da177e4SLinus Torvalds struct sk_buff *buff; 1603b4a24397SEric Dumazet int old_factor; 1604b617158dSEric Dumazet long limit; 1605b60b49eaSHerbert Xu int nlen; 16069ce01461SIlpo Järvinen u8 flags; 16071da177e4SLinus Torvalds 16082fceec13SIlpo Järvinen if (WARN_ON(len > skb->len)) 16092fceec13SIlpo Järvinen return -EINVAL; 16106a438bbeSStephen Hemminger 1611b4a24397SEric Dumazet DEBUG_NET_WARN_ON_ONCE(skb_headlen(skb)); 16121da177e4SLinus Torvalds 1613b617158dSEric Dumazet /* tcp_sendmsg() can overshoot sk_wmem_queued by one full size skb. 1614b617158dSEric Dumazet * We need some allowance to not penalize applications setting small 1615b617158dSEric Dumazet * SO_SNDBUF values. 1616b617158dSEric Dumazet * Also allow first and last skb in retransmit queue to be split. 1617b617158dSEric Dumazet */ 16187c4e983cSAlexander Duyck limit = sk->sk_sndbuf + 2 * SKB_TRUESIZE(GSO_LEGACY_MAX_SIZE); 1619b617158dSEric Dumazet if (unlikely((sk->sk_wmem_queued >> 1) > limit && 1620b617158dSEric Dumazet tcp_queue != TCP_FRAG_IN_WRITE_QUEUE && 1621b617158dSEric Dumazet skb != tcp_rtx_queue_head(sk) && 1622b617158dSEric Dumazet skb != tcp_rtx_queue_tail(sk))) { 1623f070ef2aSEric Dumazet NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPWQUEUETOOBIG); 1624f070ef2aSEric Dumazet return -ENOMEM; 1625f070ef2aSEric Dumazet } 1626f070ef2aSEric Dumazet 1627c4777efaSEric Dumazet if (skb_unclone_keeptruesize(skb, gfp)) 16281da177e4SLinus Torvalds return -ENOMEM; 16291da177e4SLinus Torvalds 16301da177e4SLinus Torvalds /* Get a new skb... force flag on. */ 16315882efffSEric Dumazet buff = tcp_stream_alloc_skb(sk, gfp, true); 163251456b29SIan Morris if (!buff) 16331da177e4SLinus Torvalds return -ENOMEM; /* We'll just try again later. */ 163441477662SJakub Kicinski skb_copy_decrypted(buff, skb); 16355a369ca6SPaolo Abeni mptcp_skb_ext_copy(buff, skb); 1636ef5cb973SHerbert Xu 1637ab4e846aSEric Dumazet sk_wmem_queued_add(sk, buff->truesize); 16383ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1639b4a24397SEric Dumazet nlen = skb->len - len; 1640b60b49eaSHerbert Xu buff->truesize += nlen; 1641b60b49eaSHerbert Xu skb->truesize -= nlen; 16421da177e4SLinus Torvalds 16431da177e4SLinus Torvalds /* Correct the sequence numbers. */ 16441da177e4SLinus Torvalds TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 16451da177e4SLinus Torvalds TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 16461da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 16471da177e4SLinus Torvalds 16481da177e4SLinus Torvalds /* PSH and FIN should only be set in the second packet. */ 16494de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 16504de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 16514de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1652e14c3cafSHerbert Xu TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; 1653a166140eSMartin KaFai Lau tcp_skb_fragment_eor(skb, buff); 16541da177e4SLinus Torvalds 16551da177e4SLinus Torvalds skb_split(skb, buff, len); 16561da177e4SLinus Torvalds 1657a1ac9c8aSMartin KaFai Lau skb_set_delivery_time(buff, skb->tstamp, true); 1658490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 16591da177e4SLinus Torvalds 16606475be16SDavid S. Miller old_factor = tcp_skb_pcount(skb); 16616475be16SDavid S. Miller 16621da177e4SLinus Torvalds /* Fix up tso_factor for both original and new SKB. */ 16635bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 16645bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 16651da177e4SLinus Torvalds 1666b9f64820SYuchung Cheng /* Update delivered info for the new segment */ 1667b9f64820SYuchung Cheng TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx; 1668b9f64820SYuchung Cheng 16696475be16SDavid S. Miller /* If this packet has been sent out already, we must 16706475be16SDavid S. Miller * adjust the various packet counters. 16716475be16SDavid S. Miller */ 1672cf0b450cSHerbert Xu if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 16736475be16SDavid S. Miller int diff = old_factor - tcp_skb_pcount(skb) - 16746475be16SDavid S. Miller tcp_skb_pcount(buff); 16751da177e4SLinus Torvalds 1676797108d1SIlpo Järvinen if (diff) 1677797108d1SIlpo Järvinen tcp_adjust_pcount(sk, skb, diff); 16781da177e4SLinus Torvalds } 16791da177e4SLinus Torvalds 16801da177e4SLinus Torvalds /* Link BUFF into the send queue. */ 1681f4a775d1SEric Dumazet __skb_header_release(buff); 168275c119afSEric Dumazet tcp_insert_write_queue_after(skb, buff, sk, tcp_queue); 1683f67971e6SEric Dumazet if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE) 1684e2080072SEric Dumazet list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor); 16851da177e4SLinus Torvalds 16861da177e4SLinus Torvalds return 0; 16871da177e4SLinus Torvalds } 16881da177e4SLinus Torvalds 1689f4d01666SEric Dumazet /* This is similar to __pskb_pull_tail(). The difference is that pulled 1690f4d01666SEric Dumazet * data is not copied, but immediately discarded. 16911da177e4SLinus Torvalds */ 16927162fb24SEric Dumazet static int __pskb_trim_head(struct sk_buff *skb, int len) 16931da177e4SLinus Torvalds { 16947b7fc97aSEric Dumazet struct skb_shared_info *shinfo; 16951da177e4SLinus Torvalds int i, k, eat; 16961da177e4SLinus Torvalds 1697b4a24397SEric Dumazet DEBUG_NET_WARN_ON_ONCE(skb_headlen(skb)); 16981da177e4SLinus Torvalds eat = len; 16991da177e4SLinus Torvalds k = 0; 17007b7fc97aSEric Dumazet shinfo = skb_shinfo(skb); 17017b7fc97aSEric Dumazet for (i = 0; i < shinfo->nr_frags; i++) { 17027b7fc97aSEric Dumazet int size = skb_frag_size(&shinfo->frags[i]); 17039e903e08SEric Dumazet 17049e903e08SEric Dumazet if (size <= eat) { 1705aff65da0SIan Campbell skb_frag_unref(skb, i); 17069e903e08SEric Dumazet eat -= size; 17071da177e4SLinus Torvalds } else { 17087b7fc97aSEric Dumazet shinfo->frags[k] = shinfo->frags[i]; 17091da177e4SLinus Torvalds if (eat) { 1710b54c9d5bSJonathan Lemon skb_frag_off_add(&shinfo->frags[k], eat); 17117b7fc97aSEric Dumazet skb_frag_size_sub(&shinfo->frags[k], eat); 17121da177e4SLinus Torvalds eat = 0; 17131da177e4SLinus Torvalds } 17141da177e4SLinus Torvalds k++; 17151da177e4SLinus Torvalds } 17161da177e4SLinus Torvalds } 17177b7fc97aSEric Dumazet shinfo->nr_frags = k; 17181da177e4SLinus Torvalds 17191da177e4SLinus Torvalds skb->data_len -= len; 17201da177e4SLinus Torvalds skb->len = skb->data_len; 17217162fb24SEric Dumazet return len; 17221da177e4SLinus Torvalds } 17231da177e4SLinus Torvalds 172467edfef7SAndi Kleen /* Remove acked data from a packet in the transmit queue. */ 17251da177e4SLinus Torvalds int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 17261da177e4SLinus Torvalds { 17277162fb24SEric Dumazet u32 delta_truesize; 17287162fb24SEric Dumazet 1729c4777efaSEric Dumazet if (skb_unclone_keeptruesize(skb, GFP_ATOMIC)) 17301da177e4SLinus Torvalds return -ENOMEM; 17311da177e4SLinus Torvalds 17327162fb24SEric Dumazet delta_truesize = __pskb_trim_head(skb, len); 17331da177e4SLinus Torvalds 17341da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq += len; 17351da177e4SLinus Torvalds 17367162fb24SEric Dumazet skb->truesize -= delta_truesize; 1737ab4e846aSEric Dumazet sk_wmem_queued_add(sk, -delta_truesize); 17389b65b17dSTalal Ahmad if (!skb_zcopy_pure(skb)) 17397162fb24SEric Dumazet sk_mem_uncharge(sk, delta_truesize); 17401da177e4SLinus Torvalds 17415b35e1e6SNeal Cardwell /* Any change of skb->len requires recalculation of tso factor. */ 17421da177e4SLinus Torvalds if (tcp_skb_pcount(skb) > 1) 17435bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb)); 17441da177e4SLinus Torvalds 17451da177e4SLinus Torvalds return 0; 17461da177e4SLinus Torvalds } 17471da177e4SLinus Torvalds 17481b63edd6SYuchung Cheng /* Calculate MSS not accounting any TCP options. */ 17491b63edd6SYuchung Cheng static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu) 17505d424d5aSJohn Heffner { 1751cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1752cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 17535d424d5aSJohn Heffner int mss_now; 17545d424d5aSJohn Heffner 17555d424d5aSJohn Heffner /* Calculate base mss without TCP options: 17565d424d5aSJohn Heffner It is MMS_S - sizeof(tcphdr) of rfc1122 17575d424d5aSJohn Heffner */ 17585d424d5aSJohn Heffner mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 17595d424d5aSJohn Heffner 17605d424d5aSJohn Heffner /* Clamp it (mss_clamp does not include tcp options) */ 17615d424d5aSJohn Heffner if (mss_now > tp->rx_opt.mss_clamp) 17625d424d5aSJohn Heffner mss_now = tp->rx_opt.mss_clamp; 17635d424d5aSJohn Heffner 17645d424d5aSJohn Heffner /* Now subtract optional transport overhead */ 17655d424d5aSJohn Heffner mss_now -= icsk->icsk_ext_hdr_len; 17665d424d5aSJohn Heffner 17675d424d5aSJohn Heffner /* Then reserve room for full set of TCP options and 8 bytes of data */ 176878eb166cSKuniyuki Iwashima mss_now = max(mss_now, 176978eb166cSKuniyuki Iwashima READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_snd_mss)); 17705d424d5aSJohn Heffner return mss_now; 17715d424d5aSJohn Heffner } 17725d424d5aSJohn Heffner 17731b63edd6SYuchung Cheng /* Calculate MSS. Not accounting for SACKs here. */ 17741b63edd6SYuchung Cheng int tcp_mtu_to_mss(struct sock *sk, int pmtu) 17751b63edd6SYuchung Cheng { 17761b63edd6SYuchung Cheng /* Subtract TCP options size, not including SACKs */ 17771b63edd6SYuchung Cheng return __tcp_mtu_to_mss(sk, pmtu) - 17781b63edd6SYuchung Cheng (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr)); 17791b63edd6SYuchung Cheng } 1780c7bb4b89SEric Dumazet EXPORT_SYMBOL(tcp_mtu_to_mss); 17811b63edd6SYuchung Cheng 17825d424d5aSJohn Heffner /* Inverse of above */ 178367469601SEric Dumazet int tcp_mss_to_mtu(struct sock *sk, int mss) 17845d424d5aSJohn Heffner { 1785cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1786cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 17875d424d5aSJohn Heffner 1788e57a3447SYan Zhai return mss + 17895d424d5aSJohn Heffner tp->tcp_header_len + 17905d424d5aSJohn Heffner icsk->icsk_ext_hdr_len + 17915d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 17925d424d5aSJohn Heffner } 1793556c6b46SNeal Cardwell EXPORT_SYMBOL(tcp_mss_to_mtu); 17945d424d5aSJohn Heffner 179567edfef7SAndi Kleen /* MTU probing init per socket */ 17965d424d5aSJohn Heffner void tcp_mtup_init(struct sock *sk) 17975d424d5aSJohn Heffner { 17985d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 17995d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 1800b0f9ca53SFan Du struct net *net = sock_net(sk); 18015d424d5aSJohn Heffner 1802f47d00e0SKuniyuki Iwashima icsk->icsk_mtup.enabled = READ_ONCE(net->ipv4.sysctl_tcp_mtu_probing) > 1; 18035d424d5aSJohn Heffner icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 18045d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 180588d78bc0SKuniyuki Iwashima icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, READ_ONCE(net->ipv4.sysctl_tcp_base_mss)); 18065d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 180705cbc0dbSFan Du if (icsk->icsk_mtup.enabled) 1808c74df29aSEric Dumazet icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; 18095d424d5aSJohn Heffner } 18104bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_mtup_init); 18115d424d5aSJohn Heffner 18121da177e4SLinus Torvalds /* This function synchronize snd mss to current pmtu/exthdr set. 18131da177e4SLinus Torvalds 18141da177e4SLinus Torvalds tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 18151da177e4SLinus Torvalds for TCP options, but includes only bare TCP header. 18161da177e4SLinus Torvalds 18171da177e4SLinus Torvalds tp->rx_opt.mss_clamp is mss negotiated at connection setup. 1818caa20d9aSStephen Hemminger It is minimum of user_mss and mss received with SYN. 18191da177e4SLinus Torvalds It also does not include TCP options. 18201da177e4SLinus Torvalds 1821d83d8461SArnaldo Carvalho de Melo inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. 18221da177e4SLinus Torvalds 18231da177e4SLinus Torvalds tp->mss_cache is current effective sending mss, including 18241da177e4SLinus Torvalds all tcp options except for SACKs. It is evaluated, 18251da177e4SLinus Torvalds taking into account current pmtu, but never exceeds 18261da177e4SLinus Torvalds tp->rx_opt.mss_clamp. 18271da177e4SLinus Torvalds 18281da177e4SLinus Torvalds NOTE1. rfc1122 clearly states that advertised MSS 18291da177e4SLinus Torvalds DOES NOT include either tcp or ip options. 18301da177e4SLinus Torvalds 1831d83d8461SArnaldo Carvalho de Melo NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache 1832d83d8461SArnaldo Carvalho de Melo are READ ONLY outside this function. --ANK (980731) 18331da177e4SLinus Torvalds */ 18341da177e4SLinus Torvalds unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 18351da177e4SLinus Torvalds { 18361da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1837d83d8461SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 18385d424d5aSJohn Heffner int mss_now; 18391da177e4SLinus Torvalds 18405d424d5aSJohn Heffner if (icsk->icsk_mtup.search_high > pmtu) 18415d424d5aSJohn Heffner icsk->icsk_mtup.search_high = pmtu; 18421da177e4SLinus Torvalds 18435d424d5aSJohn Heffner mss_now = tcp_mtu_to_mss(sk, pmtu); 1844409d22b4SIlpo Järvinen mss_now = tcp_bound_to_half_wnd(tp, mss_now); 18451da177e4SLinus Torvalds 18461da177e4SLinus Torvalds /* And store cached results */ 1847d83d8461SArnaldo Carvalho de Melo icsk->icsk_pmtu_cookie = pmtu; 18485d424d5aSJohn Heffner if (icsk->icsk_mtup.enabled) 18495d424d5aSJohn Heffner mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); 1850c1b4a7e6SDavid S. Miller tp->mss_cache = mss_now; 18511da177e4SLinus Torvalds 18521da177e4SLinus Torvalds return mss_now; 18531da177e4SLinus Torvalds } 18544bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_sync_mss); 18551da177e4SLinus Torvalds 18561da177e4SLinus Torvalds /* Compute the current effective MSS, taking SACKs and IP options, 18571da177e4SLinus Torvalds * and even PMTU discovery events into account. 18581da177e4SLinus Torvalds */ 18590c54b85fSIlpo Järvinen unsigned int tcp_current_mss(struct sock *sk) 18601da177e4SLinus Torvalds { 1861cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1862cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1863c1b4a7e6SDavid S. Miller u32 mss_now; 186495c96174SEric Dumazet unsigned int header_len; 186533ad798cSAdam Langley struct tcp_out_options opts; 18661e03d32bSDmitry Safonov struct tcp_key key; 18671da177e4SLinus Torvalds 1868c1b4a7e6SDavid S. Miller mss_now = tp->mss_cache; 1869c1b4a7e6SDavid S. Miller 18701da177e4SLinus Torvalds if (dst) { 18711da177e4SLinus Torvalds u32 mtu = dst_mtu(dst); 1872d83d8461SArnaldo Carvalho de Melo if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 18731da177e4SLinus Torvalds mss_now = tcp_sync_mss(sk, mtu); 18741da177e4SLinus Torvalds } 18751e03d32bSDmitry Safonov tcp_get_current_key(sk, &key); 18761e03d32bSDmitry Safonov header_len = tcp_established_options(sk, NULL, &opts, &key) + 187733ad798cSAdam Langley sizeof(struct tcphdr); 187833ad798cSAdam Langley /* The mss_cache is sized based on tp->tcp_header_len, which assumes 187933ad798cSAdam Langley * some common options. If this is an odd packet (because we have SACK 188033ad798cSAdam Langley * blocks etc) then our calculated header_len will be different, and 188133ad798cSAdam Langley * we have to adjust mss_now correspondingly */ 188233ad798cSAdam Langley if (header_len != tp->tcp_header_len) { 188333ad798cSAdam Langley int delta = (int) header_len - tp->tcp_header_len; 188433ad798cSAdam Langley mss_now -= delta; 188533ad798cSAdam Langley } 1886cfb6eeb4SYOSHIFUJI Hideaki 18871da177e4SLinus Torvalds return mss_now; 18881da177e4SLinus Torvalds } 18891da177e4SLinus Torvalds 189086fd14adSWeiping Pan /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. 189186fd14adSWeiping Pan * As additional protections, we do not touch cwnd in retransmission phases, 189286fd14adSWeiping Pan * and if application hit its sndbuf limit recently. 189386fd14adSWeiping Pan */ 189486fd14adSWeiping Pan static void tcp_cwnd_application_limited(struct sock *sk) 1895a762a980SDavid S. Miller { 18969e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1897a762a980SDavid S. Miller 189886fd14adSWeiping Pan if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && 189986fd14adSWeiping Pan sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 190086fd14adSWeiping Pan /* Limited by application or receiver window. */ 190186fd14adSWeiping Pan u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); 190286fd14adSWeiping Pan u32 win_used = max(tp->snd_cwnd_used, init_win); 190340570375SEric Dumazet if (win_used < tcp_snd_cwnd(tp)) { 190486fd14adSWeiping Pan tp->snd_ssthresh = tcp_current_ssthresh(sk); 190540570375SEric Dumazet tcp_snd_cwnd_set(tp, (tcp_snd_cwnd(tp) + win_used) >> 1); 190686fd14adSWeiping Pan } 190786fd14adSWeiping Pan tp->snd_cwnd_used = 0; 190886fd14adSWeiping Pan } 1909c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 191086fd14adSWeiping Pan } 191186fd14adSWeiping Pan 1912ca8a2263SNeal Cardwell static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited) 1913a762a980SDavid S. Miller { 19141b1fc3fdSWei Wang const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 1915a762a980SDavid S. Miller struct tcp_sock *tp = tcp_sk(sk); 1916a762a980SDavid S. Miller 1917f4ce91ceSNeal Cardwell /* Track the strongest available signal of the degree to which the cwnd 1918f4ce91ceSNeal Cardwell * is fully utilized. If cwnd-limited then remember that fact for the 1919f4ce91ceSNeal Cardwell * current window. If not cwnd-limited then track the maximum number of 1920f4ce91ceSNeal Cardwell * outstanding packets in the current window. (If cwnd-limited then we 1921f4ce91ceSNeal Cardwell * chose to not update tp->max_packets_out to avoid an extra else 1922f4ce91ceSNeal Cardwell * clause with no functional impact.) 1923ca8a2263SNeal Cardwell */ 1924f4ce91ceSNeal Cardwell if (!before(tp->snd_una, tp->cwnd_usage_seq) || 1925f4ce91ceSNeal Cardwell is_cwnd_limited || 1926f4ce91ceSNeal Cardwell (!tp->is_cwnd_limited && 1927f4ce91ceSNeal Cardwell tp->packets_out > tp->max_packets_out)) { 1928ca8a2263SNeal Cardwell tp->is_cwnd_limited = is_cwnd_limited; 1929f4ce91ceSNeal Cardwell tp->max_packets_out = tp->packets_out; 1930f4ce91ceSNeal Cardwell tp->cwnd_usage_seq = tp->snd_nxt; 1931ca8a2263SNeal Cardwell } 1932e114a710SEric Dumazet 193324901551SEric Dumazet if (tcp_is_cwnd_limited(sk)) { 1934a762a980SDavid S. Miller /* Network is feed fully. */ 1935a762a980SDavid S. Miller tp->snd_cwnd_used = 0; 1936c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 1937a762a980SDavid S. Miller } else { 1938a762a980SDavid S. Miller /* Network starves. */ 1939a762a980SDavid S. Miller if (tp->packets_out > tp->snd_cwnd_used) 1940a762a980SDavid S. Miller tp->snd_cwnd_used = tp->packets_out; 1941a762a980SDavid S. Miller 19424845b571SKuniyuki Iwashima if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle) && 1943c2203cf7SEric Dumazet (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto && 19441b1fc3fdSWei Wang !ca_ops->cong_control) 1945a762a980SDavid S. Miller tcp_cwnd_application_limited(sk); 1946b0f71bd3SFrancis Yan 1947b0f71bd3SFrancis Yan /* The following conditions together indicate the starvation 1948b0f71bd3SFrancis Yan * is caused by insufficient sender buffer: 1949b0f71bd3SFrancis Yan * 1) just sent some data (see tcp_write_xmit) 1950b0f71bd3SFrancis Yan * 2) not cwnd limited (this else condition) 195175c119afSEric Dumazet * 3) no more data to send (tcp_write_queue_empty()) 1952b0f71bd3SFrancis Yan * 4) application is hitting buffer limit (SOCK_NOSPACE) 1953b0f71bd3SFrancis Yan */ 195475c119afSEric Dumazet if (tcp_write_queue_empty(sk) && sk->sk_socket && 1955b0f71bd3SFrancis Yan test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) && 1956b0f71bd3SFrancis Yan (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) 1957b0f71bd3SFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED); 1958a762a980SDavid S. Miller } 1959a762a980SDavid S. Miller } 1960a762a980SDavid S. Miller 1961d4589926SEric Dumazet /* Minshall's variant of the Nagle send check. */ 1962d4589926SEric Dumazet static bool tcp_minshall_check(const struct tcp_sock *tp) 1963d4589926SEric Dumazet { 1964d4589926SEric Dumazet return after(tp->snd_sml, tp->snd_una) && 1965d4589926SEric Dumazet !after(tp->snd_sml, tp->snd_nxt); 1966d4589926SEric Dumazet } 1967d4589926SEric Dumazet 1968d4589926SEric Dumazet /* Update snd_sml if this skb is under mss 1969d4589926SEric Dumazet * Note that a TSO packet might end with a sub-mss segment 1970d4589926SEric Dumazet * The test is really : 1971d4589926SEric Dumazet * if ((skb->len % mss) != 0) 1972d4589926SEric Dumazet * tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1973d4589926SEric Dumazet * But we can avoid doing the divide again given we already have 1974d4589926SEric Dumazet * skb_pcount = skb->len / mss_now 19750e3a4803SIlpo Järvinen */ 1976d4589926SEric Dumazet static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now, 1977d4589926SEric Dumazet const struct sk_buff *skb) 1978d4589926SEric Dumazet { 1979d4589926SEric Dumazet if (skb->len < tcp_skb_pcount(skb) * mss_now) 1980d4589926SEric Dumazet tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1981d4589926SEric Dumazet } 1982d4589926SEric Dumazet 1983d4589926SEric Dumazet /* Return false, if packet can be sent now without violation Nagle's rules: 1984d4589926SEric Dumazet * 1. It is full sized. (provided by caller in %partial bool) 1985d4589926SEric Dumazet * 2. Or it contains FIN. (already checked by caller) 1986d4589926SEric Dumazet * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. 1987d4589926SEric Dumazet * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 1988d4589926SEric Dumazet * With Minshall's modification: all sent small packets are ACKed. 1989d4589926SEric Dumazet */ 1990d4589926SEric Dumazet static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp, 1991cc93fc51SPeter Pan(潘卫平) int nonagle) 1992d4589926SEric Dumazet { 1993d4589926SEric Dumazet return partial && 1994d4589926SEric Dumazet ((nonagle & TCP_NAGLE_CORK) || 1995d4589926SEric Dumazet (!nonagle && tp->packets_out && tcp_minshall_check(tp))); 1996d4589926SEric Dumazet } 1997605ad7f1SEric Dumazet 1998605ad7f1SEric Dumazet /* Return how many segs we'd like on a TSO packet, 199965466904SEric Dumazet * depending on current pacing rate, and how close the peer is. 200065466904SEric Dumazet * 200165466904SEric Dumazet * Rationale is: 200265466904SEric Dumazet * - For close peers, we rather send bigger packets to reduce 200365466904SEric Dumazet * cpu costs, because occasional losses will be repaired fast. 200465466904SEric Dumazet * - For long distance/rtt flows, we would like to get ACK clocking 200565466904SEric Dumazet * with 1 ACK per ms. 200665466904SEric Dumazet * 200765466904SEric Dumazet * Use min_rtt to help adapt TSO burst size, with smaller min_rtt resulting 200865466904SEric Dumazet * in bigger TSO bursts. We we cut the RTT-based allowance in half 200965466904SEric Dumazet * for every 2^9 usec (aka 512 us) of RTT, so that the RTT-based allowance 201065466904SEric Dumazet * is below 1500 bytes after 6 * ~500 usec = 3ms. 2011605ad7f1SEric Dumazet */ 2012dcb8c9b4SEric Dumazet static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now, 20131b3878caSNeal Cardwell int min_tso_segs) 2014605ad7f1SEric Dumazet { 201565466904SEric Dumazet unsigned long bytes; 201665466904SEric Dumazet u32 r; 2017605ad7f1SEric Dumazet 201828b24f90SEric Dumazet bytes = READ_ONCE(sk->sk_pacing_rate) >> READ_ONCE(sk->sk_pacing_shift); 2019605ad7f1SEric Dumazet 20202455e61bSKuniyuki Iwashima r = tcp_min_rtt(tcp_sk(sk)) >> READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_rtt_log); 202165466904SEric Dumazet if (r < BITS_PER_TYPE(sk->sk_gso_max_size)) 202265466904SEric Dumazet bytes += sk->sk_gso_max_size >> r; 2023605ad7f1SEric Dumazet 202465466904SEric Dumazet bytes = min_t(unsigned long, bytes, sk->sk_gso_max_size); 202565466904SEric Dumazet 202665466904SEric Dumazet return max_t(u32, bytes / mss_now, min_tso_segs); 2027605ad7f1SEric Dumazet } 2028605ad7f1SEric Dumazet 2029ed6e7268SNeal Cardwell /* Return the number of segments we want in the skb we are transmitting. 2030ed6e7268SNeal Cardwell * See if congestion control module wants to decide; otherwise, autosize. 2031ed6e7268SNeal Cardwell */ 2032ed6e7268SNeal Cardwell static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now) 2033ed6e7268SNeal Cardwell { 2034ed6e7268SNeal Cardwell const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 2035dcb8c9b4SEric Dumazet u32 min_tso, tso_segs; 2036ed6e7268SNeal Cardwell 2037dcb8c9b4SEric Dumazet min_tso = ca_ops->min_tso_segs ? 2038dcb8c9b4SEric Dumazet ca_ops->min_tso_segs(sk) : 2039e0bb4ab9SKuniyuki Iwashima READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs); 2040dcb8c9b4SEric Dumazet 2041dcb8c9b4SEric Dumazet tso_segs = tcp_tso_autosize(sk, mss_now, min_tso); 2042350c9f48SEric Dumazet return min_t(u32, tso_segs, sk->sk_gso_max_segs); 2043ed6e7268SNeal Cardwell } 2044ed6e7268SNeal Cardwell 2045d4589926SEric Dumazet /* Returns the portion of skb which can be sent right away */ 2046d4589926SEric Dumazet static unsigned int tcp_mss_split_point(const struct sock *sk, 2047d4589926SEric Dumazet const struct sk_buff *skb, 2048d4589926SEric Dumazet unsigned int mss_now, 2049d4589926SEric Dumazet unsigned int max_segs, 2050d4589926SEric Dumazet int nonagle) 2051c1b4a7e6SDavid S. Miller { 2052cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 2053d4589926SEric Dumazet u32 partial, needed, window, max_len; 2054c1b4a7e6SDavid S. Miller 205590840defSIlpo Järvinen window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 20561485348dSBen Hutchings max_len = mss_now * max_segs; 20570e3a4803SIlpo Järvinen 20581485348dSBen Hutchings if (likely(max_len <= window && skb != tcp_write_queue_tail(sk))) 20591485348dSBen Hutchings return max_len; 20600e3a4803SIlpo Järvinen 20615ea3a748SIlpo Järvinen needed = min(skb->len, window); 20625ea3a748SIlpo Järvinen 20631485348dSBen Hutchings if (max_len <= needed) 20641485348dSBen Hutchings return max_len; 20650e3a4803SIlpo Järvinen 2066d4589926SEric Dumazet partial = needed % mss_now; 2067d4589926SEric Dumazet /* If last segment is not a full MSS, check if Nagle rules allow us 2068d4589926SEric Dumazet * to include this last segment in this skb. 2069d4589926SEric Dumazet * Otherwise, we'll split the skb at last MSS boundary 2070d4589926SEric Dumazet */ 2071cc93fc51SPeter Pan(潘卫平) if (tcp_nagle_check(partial != 0, tp, nonagle)) 2072d4589926SEric Dumazet return needed - partial; 2073d4589926SEric Dumazet 2074d4589926SEric Dumazet return needed; 2075c1b4a7e6SDavid S. Miller } 2076c1b4a7e6SDavid S. Miller 2077c1b4a7e6SDavid S. Miller /* Can at least one segment of SKB be sent right now, according to the 2078c1b4a7e6SDavid S. Miller * congestion window rules? If so, return how many segments are allowed. 2079c1b4a7e6SDavid S. Miller */ 208022555032SEric Dumazet static u32 tcp_cwnd_test(const struct tcp_sock *tp) 2081c1b4a7e6SDavid S. Miller { 2082d649a7a8SEric Dumazet u32 in_flight, cwnd, halfcwnd; 2083c1b4a7e6SDavid S. Miller 2084c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 208540570375SEric Dumazet cwnd = tcp_snd_cwnd(tp); 2086d649a7a8SEric Dumazet if (in_flight >= cwnd) 2087c1b4a7e6SDavid S. Miller return 0; 2088d649a7a8SEric Dumazet 2089d649a7a8SEric Dumazet /* For better scheduling, ensure we have at least 2090d649a7a8SEric Dumazet * 2 GSO packets in flight. 2091d649a7a8SEric Dumazet */ 2092d649a7a8SEric Dumazet halfcwnd = max(cwnd >> 1, 1U); 2093d649a7a8SEric Dumazet return min(halfcwnd, cwnd - in_flight); 2094c1b4a7e6SDavid S. Miller } 2095c1b4a7e6SDavid S. Miller 2096b595076aSUwe Kleine-König /* Initialize TSO state of a skb. 209767edfef7SAndi Kleen * This must be invoked the first time we consider transmitting 2098c1b4a7e6SDavid S. Miller * SKB onto the wire. 2099c1b4a7e6SDavid S. Miller */ 21005bbb432cSEric Dumazet static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now) 2101c1b4a7e6SDavid S. Miller { 2102c1b4a7e6SDavid S. Miller int tso_segs = tcp_skb_pcount(skb); 2103c1b4a7e6SDavid S. Miller 2104d5b38a71SEric Dumazet if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) 2105d5b38a71SEric Dumazet return tcp_set_skb_tso_segs(skb, mss_now); 2106d5b38a71SEric Dumazet 2107c1b4a7e6SDavid S. Miller return tso_segs; 2108c1b4a7e6SDavid S. Miller } 2109c1b4a7e6SDavid S. Miller 2110c1b4a7e6SDavid S. Miller 2111a2a385d6SEric Dumazet /* Return true if the Nagle test allows this packet to be 2112c1b4a7e6SDavid S. Miller * sent now. 2113c1b4a7e6SDavid S. Miller */ 2114a2a385d6SEric Dumazet static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, 2115c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 2116c1b4a7e6SDavid S. Miller { 2117c1b4a7e6SDavid S. Miller /* Nagle rule does not apply to frames, which sit in the middle of the 2118c1b4a7e6SDavid S. Miller * write_queue (they have no chances to get new data). 2119c1b4a7e6SDavid S. Miller * 2120c1b4a7e6SDavid S. Miller * This is implemented in the callers, where they modify the 'nonagle' 2121c1b4a7e6SDavid S. Miller * argument based upon the location of SKB in the send queue. 2122c1b4a7e6SDavid S. Miller */ 2123c1b4a7e6SDavid S. Miller if (nonagle & TCP_NAGLE_PUSH) 2124a2a385d6SEric Dumazet return true; 2125c1b4a7e6SDavid S. Miller 21269b44190dSYuchung Cheng /* Don't use the nagle rule for urgent data (or for the final FIN). */ 21279b44190dSYuchung Cheng if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) 2128a2a385d6SEric Dumazet return true; 2129c1b4a7e6SDavid S. Miller 2130cc93fc51SPeter Pan(潘卫平) if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle)) 2131a2a385d6SEric Dumazet return true; 2132c1b4a7e6SDavid S. Miller 2133a2a385d6SEric Dumazet return false; 2134c1b4a7e6SDavid S. Miller } 2135c1b4a7e6SDavid S. Miller 2136c1b4a7e6SDavid S. Miller /* Does at least the first segment of SKB fit into the send window? */ 2137a2a385d6SEric Dumazet static bool tcp_snd_wnd_test(const struct tcp_sock *tp, 2138a2a385d6SEric Dumazet const struct sk_buff *skb, 2139056834d9SIlpo Järvinen unsigned int cur_mss) 2140c1b4a7e6SDavid S. Miller { 2141c1b4a7e6SDavid S. Miller u32 end_seq = TCP_SKB_CB(skb)->end_seq; 2142c1b4a7e6SDavid S. Miller 2143c1b4a7e6SDavid S. Miller if (skb->len > cur_mss) 2144c1b4a7e6SDavid S. Miller end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 2145c1b4a7e6SDavid S. Miller 214690840defSIlpo Järvinen return !after(end_seq, tcp_wnd_end(tp)); 2147c1b4a7e6SDavid S. Miller } 2148c1b4a7e6SDavid S. Miller 2149c1b4a7e6SDavid S. Miller /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 2150c1b4a7e6SDavid S. Miller * which is put after SKB on the list. It is very much like 2151c1b4a7e6SDavid S. Miller * tcp_fragment() except that it may make several kinds of assumptions 2152c1b4a7e6SDavid S. Miller * in order to speed up the splitting operation. In particular, we 2153c1b4a7e6SDavid S. Miller * know that all the data is in scatter-gather pages, and that the 2154c1b4a7e6SDavid S. Miller * packet has never been sent out before (and thus is not cloned). 2155c1b4a7e6SDavid S. Miller */ 215656483341SEric Dumazet static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, 2157c4ead4c5SEric Dumazet unsigned int mss_now, gfp_t gfp) 2158c1b4a7e6SDavid S. Miller { 2159c1b4a7e6SDavid S. Miller int nlen = skb->len - len; 216056483341SEric Dumazet struct sk_buff *buff; 21619ce01461SIlpo Järvinen u8 flags; 2162c1b4a7e6SDavid S. Miller 2163c1b4a7e6SDavid S. Miller /* All of a TSO frame must be composed of paged data. */ 2164b4a24397SEric Dumazet DEBUG_NET_WARN_ON_ONCE(skb->len != skb->data_len); 2165c1b4a7e6SDavid S. Miller 21665882efffSEric Dumazet buff = tcp_stream_alloc_skb(sk, gfp, true); 216751456b29SIan Morris if (unlikely(!buff)) 2168c1b4a7e6SDavid S. Miller return -ENOMEM; 216941477662SJakub Kicinski skb_copy_decrypted(buff, skb); 21705a369ca6SPaolo Abeni mptcp_skb_ext_copy(buff, skb); 2171c1b4a7e6SDavid S. Miller 2172ab4e846aSEric Dumazet sk_wmem_queued_add(sk, buff->truesize); 21733ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 2174b60b49eaSHerbert Xu buff->truesize += nlen; 2175c1b4a7e6SDavid S. Miller skb->truesize -= nlen; 2176c1b4a7e6SDavid S. Miller 2177c1b4a7e6SDavid S. Miller /* Correct the sequence numbers. */ 2178c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 2179c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 2180c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 2181c1b4a7e6SDavid S. Miller 2182c1b4a7e6SDavid S. Miller /* PSH and FIN should only be set in the second packet. */ 21834de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 21844de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 21854de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 2186c1b4a7e6SDavid S. Miller 2187a166140eSMartin KaFai Lau tcp_skb_fragment_eor(skb, buff); 2188a166140eSMartin KaFai Lau 2189c1b4a7e6SDavid S. Miller skb_split(skb, buff, len); 2190490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 2191c1b4a7e6SDavid S. Miller 2192c1b4a7e6SDavid S. Miller /* Fix up tso_factor for both original and new SKB. */ 21935bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 21945bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 2195c1b4a7e6SDavid S. Miller 2196c1b4a7e6SDavid S. Miller /* Link BUFF into the send queue. */ 2197f4a775d1SEric Dumazet __skb_header_release(buff); 219856483341SEric Dumazet tcp_insert_write_queue_after(skb, buff, sk, TCP_FRAG_IN_WRITE_QUEUE); 2199c1b4a7e6SDavid S. Miller 2200c1b4a7e6SDavid S. Miller return 0; 2201c1b4a7e6SDavid S. Miller } 2202c1b4a7e6SDavid S. Miller 2203c1b4a7e6SDavid S. Miller /* Try to defer sending, if possible, in order to minimize the amount 2204c1b4a7e6SDavid S. Miller * of TSO splitting we do. View it as a kind of TSO Nagle test. 2205c1b4a7e6SDavid S. Miller * 2206c1b4a7e6SDavid S. Miller * This algorithm is from John Heffner. 2207c1b4a7e6SDavid S. Miller */ 2208ca8a2263SNeal Cardwell static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb, 2209f9bfe4e6SEric Dumazet bool *is_cwnd_limited, 2210f9bfe4e6SEric Dumazet bool *is_rwnd_limited, 2211f9bfe4e6SEric Dumazet u32 max_segs) 2212c1b4a7e6SDavid S. Miller { 22136687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 2214f1c6ea38SEric Dumazet u32 send_win, cong_win, limit, in_flight; 221550c8339eSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 221650c8339eSEric Dumazet struct sk_buff *head; 2217ad9f4f50SEric Dumazet int win_divisor; 2218f1c6ea38SEric Dumazet s64 delta; 2219c1b4a7e6SDavid S. Miller 222099d7662aSEric Dumazet if (icsk->icsk_ca_state >= TCP_CA_Recovery) 2221ae8064acSJohn Heffner goto send_now; 2222ae8064acSJohn Heffner 22235f852eb5SEric Dumazet /* Avoid bursty behavior by allowing defer 2224a682850aSEric Dumazet * only if the last write was recent (1 ms). 2225a682850aSEric Dumazet * Note that tp->tcp_wstamp_ns can be in the future if we have 2226a682850aSEric Dumazet * packets waiting in a qdisc or device for EDT delivery. 22275f852eb5SEric Dumazet */ 2228a682850aSEric Dumazet delta = tp->tcp_clock_cache - tp->tcp_wstamp_ns - NSEC_PER_MSEC; 2229a682850aSEric Dumazet if (delta > 0) 2230ae8064acSJohn Heffner goto send_now; 2231908a75c1SDavid S. Miller 2232c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 2233c1b4a7e6SDavid S. Miller 2234c8c9aeb5SStefano Brivio BUG_ON(tcp_skb_pcount(skb) <= 1); 223540570375SEric Dumazet BUG_ON(tcp_snd_cwnd(tp) <= in_flight); 2236c1b4a7e6SDavid S. Miller 223790840defSIlpo Järvinen send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 2238c1b4a7e6SDavid S. Miller 2239c1b4a7e6SDavid S. Miller /* From in_flight test above, we know that cwnd > in_flight. */ 224040570375SEric Dumazet cong_win = (tcp_snd_cwnd(tp) - in_flight) * tp->mss_cache; 2241c1b4a7e6SDavid S. Miller 2242c1b4a7e6SDavid S. Miller limit = min(send_win, cong_win); 2243c1b4a7e6SDavid S. Miller 2244ba244fe9SDavid S. Miller /* If a full-sized TSO skb can be sent, do it. */ 2245605ad7f1SEric Dumazet if (limit >= max_segs * tp->mss_cache) 2246ae8064acSJohn Heffner goto send_now; 2247ba244fe9SDavid S. Miller 224862ad2761SIlpo Järvinen /* Middle in queue won't get any more data, full sendable already? */ 224962ad2761SIlpo Järvinen if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) 225062ad2761SIlpo Järvinen goto send_now; 225162ad2761SIlpo Järvinen 22525bbcc0f5SLinus Torvalds win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor); 2253ad9f4f50SEric Dumazet if (win_divisor) { 225440570375SEric Dumazet u32 chunk = min(tp->snd_wnd, tcp_snd_cwnd(tp) * tp->mss_cache); 2255c1b4a7e6SDavid S. Miller 2256c1b4a7e6SDavid S. Miller /* If at least some fraction of a window is available, 2257c1b4a7e6SDavid S. Miller * just use it. 2258c1b4a7e6SDavid S. Miller */ 2259ad9f4f50SEric Dumazet chunk /= win_divisor; 2260c1b4a7e6SDavid S. Miller if (limit >= chunk) 2261ae8064acSJohn Heffner goto send_now; 2262c1b4a7e6SDavid S. Miller } else { 2263c1b4a7e6SDavid S. Miller /* Different approach, try not to defer past a single 2264c1b4a7e6SDavid S. Miller * ACK. Receiver should ACK every other full sized 2265c1b4a7e6SDavid S. Miller * frame, so if we have space for more than 3 frames 2266c1b4a7e6SDavid S. Miller * then send now. 2267c1b4a7e6SDavid S. Miller */ 22686b5a5c0dSNeal Cardwell if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) 2269ae8064acSJohn Heffner goto send_now; 2270c1b4a7e6SDavid S. Miller } 2271c1b4a7e6SDavid S. Miller 227275c119afSEric Dumazet /* TODO : use tsorted_sent_queue ? */ 227375c119afSEric Dumazet head = tcp_rtx_queue_head(sk); 227475c119afSEric Dumazet if (!head) 227575c119afSEric Dumazet goto send_now; 2276f1c6ea38SEric Dumazet delta = tp->tcp_clock_cache - head->tstamp; 227750c8339eSEric Dumazet /* If next ACK is likely to come too late (half srtt), do not defer */ 2278f1c6ea38SEric Dumazet if ((s64)(delta - (u64)NSEC_PER_USEC * (tp->srtt_us >> 4)) < 0) 227950c8339eSEric Dumazet goto send_now; 228050c8339eSEric Dumazet 2281f9bfe4e6SEric Dumazet /* Ok, it looks like it is advisable to defer. 2282f9bfe4e6SEric Dumazet * Three cases are tracked : 2283f9bfe4e6SEric Dumazet * 1) We are cwnd-limited 2284f9bfe4e6SEric Dumazet * 2) We are rwnd-limited 2285f9bfe4e6SEric Dumazet * 3) We are application limited. 2286f9bfe4e6SEric Dumazet */ 2287f9bfe4e6SEric Dumazet if (cong_win < send_win) { 2288f9bfe4e6SEric Dumazet if (cong_win <= skb->len) { 2289ca8a2263SNeal Cardwell *is_cwnd_limited = true; 2290f9bfe4e6SEric Dumazet return true; 2291f9bfe4e6SEric Dumazet } 2292f9bfe4e6SEric Dumazet } else { 2293f9bfe4e6SEric Dumazet if (send_win <= skb->len) { 2294f9bfe4e6SEric Dumazet *is_rwnd_limited = true; 2295f9bfe4e6SEric Dumazet return true; 2296f9bfe4e6SEric Dumazet } 2297f9bfe4e6SEric Dumazet } 2298f9bfe4e6SEric Dumazet 2299f9bfe4e6SEric Dumazet /* If this packet won't get more data, do not wait. */ 2300d8ed257fSEric Dumazet if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) || 2301d8ed257fSEric Dumazet TCP_SKB_CB(skb)->eor) 2302f9bfe4e6SEric Dumazet goto send_now; 2303ca8a2263SNeal Cardwell 2304a2a385d6SEric Dumazet return true; 2305ae8064acSJohn Heffner 2306ae8064acSJohn Heffner send_now: 2307a2a385d6SEric Dumazet return false; 2308c1b4a7e6SDavid S. Miller } 2309c1b4a7e6SDavid S. Miller 231005cbc0dbSFan Du static inline void tcp_mtu_check_reprobe(struct sock *sk) 231105cbc0dbSFan Du { 231205cbc0dbSFan Du struct inet_connection_sock *icsk = inet_csk(sk); 231305cbc0dbSFan Du struct tcp_sock *tp = tcp_sk(sk); 231405cbc0dbSFan Du struct net *net = sock_net(sk); 231505cbc0dbSFan Du u32 interval; 231605cbc0dbSFan Du s32 delta; 231705cbc0dbSFan Du 23182a85388fSKuniyuki Iwashima interval = READ_ONCE(net->ipv4.sysctl_tcp_probe_interval); 2319c74df29aSEric Dumazet delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp; 232005cbc0dbSFan Du if (unlikely(delta >= interval * HZ)) { 232105cbc0dbSFan Du int mss = tcp_current_mss(sk); 232205cbc0dbSFan Du 232305cbc0dbSFan Du /* Update current search range */ 232405cbc0dbSFan Du icsk->icsk_mtup.probe_size = 0; 232505cbc0dbSFan Du icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + 232605cbc0dbSFan Du sizeof(struct tcphdr) + 232705cbc0dbSFan Du icsk->icsk_af_ops->net_header_len; 232805cbc0dbSFan Du icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss); 232905cbc0dbSFan Du 233005cbc0dbSFan Du /* Update probe time stamp */ 2331c74df29aSEric Dumazet icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; 233205cbc0dbSFan Du } 233305cbc0dbSFan Du } 233405cbc0dbSFan Du 2335808cf9e3SIlya Lesokhin static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len) 2336808cf9e3SIlya Lesokhin { 2337808cf9e3SIlya Lesokhin struct sk_buff *skb, *next; 2338808cf9e3SIlya Lesokhin 2339808cf9e3SIlya Lesokhin skb = tcp_send_head(sk); 2340808cf9e3SIlya Lesokhin tcp_for_write_queue_from_safe(skb, next, sk) { 2341808cf9e3SIlya Lesokhin if (len <= skb->len) 2342808cf9e3SIlya Lesokhin break; 2343808cf9e3SIlya Lesokhin 23449b65b17dSTalal Ahmad if (unlikely(TCP_SKB_CB(skb)->eor) || 23459b65b17dSTalal Ahmad tcp_has_tx_tstamp(skb) || 23469b65b17dSTalal Ahmad !skb_pure_zcopy_same(skb, next)) 2347808cf9e3SIlya Lesokhin return false; 2348808cf9e3SIlya Lesokhin 2349808cf9e3SIlya Lesokhin len -= skb->len; 2350808cf9e3SIlya Lesokhin } 2351808cf9e3SIlya Lesokhin 2352808cf9e3SIlya Lesokhin return true; 2353808cf9e3SIlya Lesokhin } 2354808cf9e3SIlya Lesokhin 235573601329SEric Dumazet static int tcp_clone_payload(struct sock *sk, struct sk_buff *to, 235673601329SEric Dumazet int probe_size) 235773601329SEric Dumazet { 235873601329SEric Dumazet skb_frag_t *lastfrag = NULL, *fragto = skb_shinfo(to)->frags; 235973601329SEric Dumazet int i, todo, len = 0, nr_frags = 0; 236073601329SEric Dumazet const struct sk_buff *skb; 236173601329SEric Dumazet 236273601329SEric Dumazet if (!sk_wmem_schedule(sk, to->truesize + probe_size)) 236373601329SEric Dumazet return -ENOMEM; 236473601329SEric Dumazet 236573601329SEric Dumazet skb_queue_walk(&sk->sk_write_queue, skb) { 236673601329SEric Dumazet const skb_frag_t *fragfrom = skb_shinfo(skb)->frags; 236773601329SEric Dumazet 236873601329SEric Dumazet if (skb_headlen(skb)) 236973601329SEric Dumazet return -EINVAL; 237073601329SEric Dumazet 237173601329SEric Dumazet for (i = 0; i < skb_shinfo(skb)->nr_frags; i++, fragfrom++) { 237273601329SEric Dumazet if (len >= probe_size) 237373601329SEric Dumazet goto commit; 237473601329SEric Dumazet todo = min_t(int, skb_frag_size(fragfrom), 237573601329SEric Dumazet probe_size - len); 237673601329SEric Dumazet len += todo; 237773601329SEric Dumazet if (lastfrag && 237873601329SEric Dumazet skb_frag_page(fragfrom) == skb_frag_page(lastfrag) && 237973601329SEric Dumazet skb_frag_off(fragfrom) == skb_frag_off(lastfrag) + 238073601329SEric Dumazet skb_frag_size(lastfrag)) { 238173601329SEric Dumazet skb_frag_size_add(lastfrag, todo); 238273601329SEric Dumazet continue; 238373601329SEric Dumazet } 238473601329SEric Dumazet if (unlikely(nr_frags == MAX_SKB_FRAGS)) 238573601329SEric Dumazet return -E2BIG; 238673601329SEric Dumazet skb_frag_page_copy(fragto, fragfrom); 238773601329SEric Dumazet skb_frag_off_copy(fragto, fragfrom); 238873601329SEric Dumazet skb_frag_size_set(fragto, todo); 238973601329SEric Dumazet nr_frags++; 239073601329SEric Dumazet lastfrag = fragto++; 239173601329SEric Dumazet } 239273601329SEric Dumazet } 239373601329SEric Dumazet commit: 239473601329SEric Dumazet WARN_ON_ONCE(len != probe_size); 239573601329SEric Dumazet for (i = 0; i < nr_frags; i++) 239673601329SEric Dumazet skb_frag_ref(to, i); 239773601329SEric Dumazet 239873601329SEric Dumazet skb_shinfo(to)->nr_frags = nr_frags; 239973601329SEric Dumazet to->truesize += probe_size; 240073601329SEric Dumazet to->len += probe_size; 240173601329SEric Dumazet to->data_len += probe_size; 240273601329SEric Dumazet __skb_header_release(to); 240373601329SEric Dumazet return 0; 240473601329SEric Dumazet } 240573601329SEric Dumazet 24065d424d5aSJohn Heffner /* Create a new MTU probe if we are ready. 240767edfef7SAndi Kleen * MTU probe is regularly attempting to increase the path MTU by 240867edfef7SAndi Kleen * deliberately sending larger packets. This discovers routing 240967edfef7SAndi Kleen * changes resulting in larger path MTUs. 241067edfef7SAndi Kleen * 24115d424d5aSJohn Heffner * Returns 0 if we should wait to probe (no cwnd available), 24125d424d5aSJohn Heffner * 1 if a probe was sent, 2413056834d9SIlpo Järvinen * -1 otherwise 2414056834d9SIlpo Järvinen */ 24155d424d5aSJohn Heffner static int tcp_mtu_probe(struct sock *sk) 24165d424d5aSJohn Heffner { 24175d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 241812a59abcSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 24195d424d5aSJohn Heffner struct sk_buff *skb, *nskb, *next; 24206b58e0a5SFan Du struct net *net = sock_net(sk); 24215d424d5aSJohn Heffner int probe_size; 242291cc17c0SIlpo Järvinen int size_needed; 242312a59abcSEric Dumazet int copy, len; 24245d424d5aSJohn Heffner int mss_now; 24256b58e0a5SFan Du int interval; 24265d424d5aSJohn Heffner 24275d424d5aSJohn Heffner /* Not currently probing/verifying, 24285d424d5aSJohn Heffner * not in recovery, 24295d424d5aSJohn Heffner * have enough cwnd, and 243012a59abcSEric Dumazet * not SACKing (the variable headers throw things off) 243112a59abcSEric Dumazet */ 243212a59abcSEric Dumazet if (likely(!icsk->icsk_mtup.enabled || 24335d424d5aSJohn Heffner icsk->icsk_mtup.probe_size || 24345d424d5aSJohn Heffner inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 243540570375SEric Dumazet tcp_snd_cwnd(tp) < 11 || 243612a59abcSEric Dumazet tp->rx_opt.num_sacks || tp->rx_opt.dsack)) 24375d424d5aSJohn Heffner return -1; 24385d424d5aSJohn Heffner 24396b58e0a5SFan Du /* Use binary search for probe_size between tcp_mss_base, 24406b58e0a5SFan Du * and current mss_clamp. if (search_high - search_low) 24416b58e0a5SFan Du * smaller than a threshold, backoff from probing. 24426b58e0a5SFan Du */ 24430c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 24446b58e0a5SFan Du probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high + 24456b58e0a5SFan Du icsk->icsk_mtup.search_low) >> 1); 244691cc17c0SIlpo Järvinen size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; 24476b58e0a5SFan Du interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low; 244805cbc0dbSFan Du /* When misfortune happens, we are reprobing actively, 244905cbc0dbSFan Du * and then reprobe timer has expired. We stick with current 245005cbc0dbSFan Du * probing process by not resetting search range to its orignal. 245105cbc0dbSFan Du */ 24526b58e0a5SFan Du if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) || 245392c0aa41SKuniyuki Iwashima interval < READ_ONCE(net->ipv4.sysctl_tcp_probe_threshold)) { 245405cbc0dbSFan Du /* Check whether enough time has elaplased for 245505cbc0dbSFan Du * another round of probing. 245605cbc0dbSFan Du */ 245705cbc0dbSFan Du tcp_mtu_check_reprobe(sk); 24585d424d5aSJohn Heffner return -1; 24595d424d5aSJohn Heffner } 24605d424d5aSJohn Heffner 24615d424d5aSJohn Heffner /* Have enough data in the send queue to probe? */ 24627f9c33e5SIlpo Järvinen if (tp->write_seq - tp->snd_nxt < size_needed) 24635d424d5aSJohn Heffner return -1; 24645d424d5aSJohn Heffner 246591cc17c0SIlpo Järvinen if (tp->snd_wnd < size_needed) 24665d424d5aSJohn Heffner return -1; 246790840defSIlpo Järvinen if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) 24685d424d5aSJohn Heffner return 0; 24695d424d5aSJohn Heffner 2470d67c58e9SIlpo Järvinen /* Do we need to wait to drain cwnd? With none in flight, don't stall */ 247140570375SEric Dumazet if (tcp_packets_in_flight(tp) + 2 > tcp_snd_cwnd(tp)) { 2472d67c58e9SIlpo Järvinen if (!tcp_packets_in_flight(tp)) 24735d424d5aSJohn Heffner return -1; 24745d424d5aSJohn Heffner else 24755d424d5aSJohn Heffner return 0; 24765d424d5aSJohn Heffner } 24775d424d5aSJohn Heffner 2478808cf9e3SIlya Lesokhin if (!tcp_can_coalesce_send_queue_head(sk, probe_size)) 2479808cf9e3SIlya Lesokhin return -1; 2480808cf9e3SIlya Lesokhin 24815d424d5aSJohn Heffner /* We're allowed to probe. Build it now. */ 24825882efffSEric Dumazet nskb = tcp_stream_alloc_skb(sk, GFP_ATOMIC, false); 248351456b29SIan Morris if (!nskb) 24845d424d5aSJohn Heffner return -1; 248573601329SEric Dumazet 248673601329SEric Dumazet /* build the payload, and be prepared to abort if this fails. */ 248773601329SEric Dumazet if (tcp_clone_payload(sk, nskb, probe_size)) { 248871c299c7SJakub Kicinski tcp_skb_tsorted_anchor_cleanup(nskb); 248973601329SEric Dumazet consume_skb(nskb); 249073601329SEric Dumazet return -1; 249173601329SEric Dumazet } 2492ab4e846aSEric Dumazet sk_wmem_queued_add(sk, nskb->truesize); 24933ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 24945d424d5aSJohn Heffner 2495fe067e8aSDavid S. Miller skb = tcp_send_head(sk); 249641477662SJakub Kicinski skb_copy_decrypted(nskb, skb); 24975a369ca6SPaolo Abeni mptcp_skb_ext_copy(nskb, skb); 24985d424d5aSJohn Heffner 24995d424d5aSJohn Heffner TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 25005d424d5aSJohn Heffner TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 25014de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; 25025d424d5aSJohn Heffner 250350c4817eSIlpo Järvinen tcp_insert_write_queue_before(nskb, skb, sk); 25042b7cda9cSEric Dumazet tcp_highest_sack_replace(sk, skb, nskb); 250550c4817eSIlpo Järvinen 25065d424d5aSJohn Heffner len = 0; 2507234b6860SIlpo Järvinen tcp_for_write_queue_from_safe(skb, next, sk) { 25085d424d5aSJohn Heffner copy = min_t(int, skb->len, probe_size - len); 25095d424d5aSJohn Heffner 25105d424d5aSJohn Heffner if (skb->len <= copy) { 25115d424d5aSJohn Heffner /* We've eaten all the data from this skb. 25125d424d5aSJohn Heffner * Throw it away. */ 25134de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 2514808cf9e3SIlya Lesokhin /* If this is the last SKB we copy and eor is set 2515808cf9e3SIlya Lesokhin * we need to propagate it to the new skb. 2516808cf9e3SIlya Lesokhin */ 2517808cf9e3SIlya Lesokhin TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor; 2518888a5c53SWillem de Bruijn tcp_skb_collapse_tstamp(nskb, skb); 2519fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 252003271f3aSTalal Ahmad tcp_wmem_free_skb(sk, skb); 25215d424d5aSJohn Heffner } else { 25224de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & 2523a3433f35SChangli Gao ~(TCPHDR_FIN|TCPHDR_PSH); 25245d424d5aSJohn Heffner __pskb_trim_head(skb, copy); 25255bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 25265d424d5aSJohn Heffner TCP_SKB_CB(skb)->seq += copy; 25275d424d5aSJohn Heffner } 25285d424d5aSJohn Heffner 25295d424d5aSJohn Heffner len += copy; 2530234b6860SIlpo Järvinen 2531234b6860SIlpo Järvinen if (len >= probe_size) 2532234b6860SIlpo Järvinen break; 25335d424d5aSJohn Heffner } 25345bbb432cSEric Dumazet tcp_init_tso_segs(nskb, nskb->len); 25355d424d5aSJohn Heffner 25365d424d5aSJohn Heffner /* We're ready to send. If this fails, the probe will 25377faee5c0SEric Dumazet * be resegmented into mss-sized pieces by tcp_write_xmit(). 25387faee5c0SEric Dumazet */ 25395d424d5aSJohn Heffner if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 25405d424d5aSJohn Heffner /* Decrement cwnd here because we are sending 25415d424d5aSJohn Heffner * effectively two packets. */ 254240570375SEric Dumazet tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) - 1); 254366f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, nskb); 25445d424d5aSJohn Heffner 25455d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 25460e7b1368SJohn Heffner tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 25470e7b1368SJohn Heffner tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 25485d424d5aSJohn Heffner 25495d424d5aSJohn Heffner return 1; 25505d424d5aSJohn Heffner } 25515d424d5aSJohn Heffner 25525d424d5aSJohn Heffner return -1; 25535d424d5aSJohn Heffner } 25545d424d5aSJohn Heffner 2555864e5c09SEric Dumazet static bool tcp_pacing_check(struct sock *sk) 2556218af599SEric Dumazet { 2557864e5c09SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 2558864e5c09SEric Dumazet 2559864e5c09SEric Dumazet if (!tcp_needs_internal_pacing(sk)) 2560864e5c09SEric Dumazet return false; 2561864e5c09SEric Dumazet 2562864e5c09SEric Dumazet if (tp->tcp_wstamp_ns <= tp->tcp_clock_cache) 2563864e5c09SEric Dumazet return false; 2564864e5c09SEric Dumazet 2565864e5c09SEric Dumazet if (!hrtimer_is_queued(&tp->pacing_timer)) { 2566864e5c09SEric Dumazet hrtimer_start(&tp->pacing_timer, 2567864e5c09SEric Dumazet ns_to_ktime(tp->tcp_wstamp_ns), 2568864e5c09SEric Dumazet HRTIMER_MODE_ABS_PINNED_SOFT); 2569864e5c09SEric Dumazet sock_hold(sk); 2570864e5c09SEric Dumazet } 2571864e5c09SEric Dumazet return true; 2572218af599SEric Dumazet } 2573218af599SEric Dumazet 2574f921a4a5SEric Dumazet static bool tcp_rtx_queue_empty_or_single_skb(const struct sock *sk) 2575f921a4a5SEric Dumazet { 2576f921a4a5SEric Dumazet const struct rb_node *node = sk->tcp_rtx_queue.rb_node; 2577f921a4a5SEric Dumazet 2578f921a4a5SEric Dumazet /* No skb in the rtx queue. */ 2579f921a4a5SEric Dumazet if (!node) 2580f921a4a5SEric Dumazet return true; 2581f921a4a5SEric Dumazet 2582f921a4a5SEric Dumazet /* Only one skb in rtx queue. */ 2583f921a4a5SEric Dumazet return !node->rb_left && !node->rb_right; 2584f921a4a5SEric Dumazet } 2585f921a4a5SEric Dumazet 2586f9616c35SEric Dumazet /* TCP Small Queues : 2587f9616c35SEric Dumazet * Control number of packets in qdisc/devices to two packets / or ~1 ms. 2588f9616c35SEric Dumazet * (These limits are doubled for retransmits) 2589f9616c35SEric Dumazet * This allows for : 2590f9616c35SEric Dumazet * - better RTT estimation and ACK scheduling 2591f9616c35SEric Dumazet * - faster recovery 2592f9616c35SEric Dumazet * - high rates 2593f9616c35SEric Dumazet * Alas, some drivers / subsystems require a fair amount 2594f9616c35SEric Dumazet * of queued bytes to ensure line rate. 2595f9616c35SEric Dumazet * One example is wifi aggregation (802.11 AMPDU) 2596f9616c35SEric Dumazet */ 2597f9616c35SEric Dumazet static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb, 2598f9616c35SEric Dumazet unsigned int factor) 2599f9616c35SEric Dumazet { 260076a9ebe8SEric Dumazet unsigned long limit; 2601f9616c35SEric Dumazet 260276a9ebe8SEric Dumazet limit = max_t(unsigned long, 260376a9ebe8SEric Dumazet 2 * skb->truesize, 260428b24f90SEric Dumazet READ_ONCE(sk->sk_pacing_rate) >> READ_ONCE(sk->sk_pacing_shift)); 2605c73e5807SEric Dumazet if (sk->sk_pacing_status == SK_PACING_NONE) 260676a9ebe8SEric Dumazet limit = min_t(unsigned long, limit, 26079fb90193SKuniyuki Iwashima READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes)); 2608f9616c35SEric Dumazet limit <<= factor; 2609f9616c35SEric Dumazet 2610a842fe14SEric Dumazet if (static_branch_unlikely(&tcp_tx_delay_enabled) && 2611a842fe14SEric Dumazet tcp_sk(sk)->tcp_tx_delay) { 261228b24f90SEric Dumazet u64 extra_bytes = (u64)READ_ONCE(sk->sk_pacing_rate) * 261328b24f90SEric Dumazet tcp_sk(sk)->tcp_tx_delay; 2614a842fe14SEric Dumazet 2615a842fe14SEric Dumazet /* TSQ is based on skb truesize sum (sk_wmem_alloc), so we 2616a842fe14SEric Dumazet * approximate our needs assuming an ~100% skb->truesize overhead. 2617a842fe14SEric Dumazet * USEC_PER_SEC is approximated by 2^20. 2618a842fe14SEric Dumazet * do_div(extra_bytes, USEC_PER_SEC/2) is replaced by a right shift. 2619a842fe14SEric Dumazet */ 2620a842fe14SEric Dumazet extra_bytes >>= (20 - 1); 2621a842fe14SEric Dumazet limit += extra_bytes; 2622a842fe14SEric Dumazet } 262314afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) > limit) { 2624f921a4a5SEric Dumazet /* Always send skb if rtx queue is empty or has one skb. 262575eefc6cSEric Dumazet * No need to wait for TX completion to call us back, 262675eefc6cSEric Dumazet * after softirq/tasklet schedule. 262775eefc6cSEric Dumazet * This helps when TX completions are delayed too much. 262875eefc6cSEric Dumazet */ 2629f921a4a5SEric Dumazet if (tcp_rtx_queue_empty_or_single_skb(sk)) 263075eefc6cSEric Dumazet return false; 263175eefc6cSEric Dumazet 26327aa5470cSEric Dumazet set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags); 2633f9616c35SEric Dumazet /* It is possible TX completion already happened 2634f9616c35SEric Dumazet * before we set TSQ_THROTTLED, so we must 2635f9616c35SEric Dumazet * test again the condition. 2636f9616c35SEric Dumazet */ 2637f9616c35SEric Dumazet smp_mb__after_atomic(); 2638ce8299b6SEric Dumazet if (refcount_read(&sk->sk_wmem_alloc) > limit) 2639f9616c35SEric Dumazet return true; 2640f9616c35SEric Dumazet } 2641f9616c35SEric Dumazet return false; 2642f9616c35SEric Dumazet } 2643f9616c35SEric Dumazet 264405b055e8SFrancis Yan static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new) 264505b055e8SFrancis Yan { 2646628174ccSEric Dumazet const u32 now = tcp_jiffies32; 2647efe967cdSArnd Bergmann enum tcp_chrono old = tp->chrono_type; 264805b055e8SFrancis Yan 2649efe967cdSArnd Bergmann if (old > TCP_CHRONO_UNSPEC) 2650efe967cdSArnd Bergmann tp->chrono_stat[old - 1] += now - tp->chrono_start; 265105b055e8SFrancis Yan tp->chrono_start = now; 265205b055e8SFrancis Yan tp->chrono_type = new; 265305b055e8SFrancis Yan } 265405b055e8SFrancis Yan 265505b055e8SFrancis Yan void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type) 265605b055e8SFrancis Yan { 265705b055e8SFrancis Yan struct tcp_sock *tp = tcp_sk(sk); 265805b055e8SFrancis Yan 265905b055e8SFrancis Yan /* If there are multiple conditions worthy of tracking in a 26600f87230dSFrancis Yan * chronograph then the highest priority enum takes precedence 26610f87230dSFrancis Yan * over the other conditions. So that if something "more interesting" 266205b055e8SFrancis Yan * starts happening, stop the previous chrono and start a new one. 266305b055e8SFrancis Yan */ 266405b055e8SFrancis Yan if (type > tp->chrono_type) 266505b055e8SFrancis Yan tcp_chrono_set(tp, type); 266605b055e8SFrancis Yan } 266705b055e8SFrancis Yan 266805b055e8SFrancis Yan void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type) 266905b055e8SFrancis Yan { 267005b055e8SFrancis Yan struct tcp_sock *tp = tcp_sk(sk); 267105b055e8SFrancis Yan 26720f87230dSFrancis Yan 26730f87230dSFrancis Yan /* There are multiple conditions worthy of tracking in a 26740f87230dSFrancis Yan * chronograph, so that the highest priority enum takes 26750f87230dSFrancis Yan * precedence over the other conditions (see tcp_chrono_start). 26760f87230dSFrancis Yan * If a condition stops, we only stop chrono tracking if 26770f87230dSFrancis Yan * it's the "most interesting" or current chrono we are 26780f87230dSFrancis Yan * tracking and starts busy chrono if we have pending data. 26790f87230dSFrancis Yan */ 268075c119afSEric Dumazet if (tcp_rtx_and_write_queues_empty(sk)) 268105b055e8SFrancis Yan tcp_chrono_set(tp, TCP_CHRONO_UNSPEC); 26820f87230dSFrancis Yan else if (type == tp->chrono_type) 26830f87230dSFrancis Yan tcp_chrono_set(tp, TCP_CHRONO_BUSY); 268405b055e8SFrancis Yan } 268505b055e8SFrancis Yan 2686*8ee602c6SEric Dumazet /* First skb in the write queue is smaller than ideal packet size. 2687*8ee602c6SEric Dumazet * Check if we can move payload from the second skb in the queue. 2688*8ee602c6SEric Dumazet */ 2689*8ee602c6SEric Dumazet static void tcp_grow_skb(struct sock *sk, struct sk_buff *skb, int amount) 2690*8ee602c6SEric Dumazet { 2691*8ee602c6SEric Dumazet struct sk_buff *next_skb = skb->next; 2692*8ee602c6SEric Dumazet unsigned int nlen; 2693*8ee602c6SEric Dumazet 2694*8ee602c6SEric Dumazet if (tcp_skb_is_last(sk, skb)) 2695*8ee602c6SEric Dumazet return; 2696*8ee602c6SEric Dumazet 2697*8ee602c6SEric Dumazet if (!tcp_skb_can_collapse(skb, next_skb)) 2698*8ee602c6SEric Dumazet return; 2699*8ee602c6SEric Dumazet 2700*8ee602c6SEric Dumazet nlen = min_t(u32, amount, next_skb->len); 2701*8ee602c6SEric Dumazet if (!nlen || !skb_shift(skb, next_skb, nlen)) 2702*8ee602c6SEric Dumazet return; 2703*8ee602c6SEric Dumazet 2704*8ee602c6SEric Dumazet TCP_SKB_CB(skb)->end_seq += nlen; 2705*8ee602c6SEric Dumazet TCP_SKB_CB(next_skb)->seq += nlen; 2706*8ee602c6SEric Dumazet 2707*8ee602c6SEric Dumazet if (!next_skb->len) { 2708*8ee602c6SEric Dumazet TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 2709*8ee602c6SEric Dumazet TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor; 2710*8ee602c6SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; 2711*8ee602c6SEric Dumazet tcp_unlink_write_queue(next_skb, sk); 2712*8ee602c6SEric Dumazet tcp_wmem_free_skb(sk, next_skb); 2713*8ee602c6SEric Dumazet } 2714*8ee602c6SEric Dumazet } 2715*8ee602c6SEric Dumazet 27161da177e4SLinus Torvalds /* This routine writes packets to the network. It advances the 27171da177e4SLinus Torvalds * send_head. This happens as incoming acks open up the remote 27181da177e4SLinus Torvalds * window for us. 27191da177e4SLinus Torvalds * 2720f8269a49SIlpo Järvinen * LARGESEND note: !tcp_urg_mode is overkill, only frames between 2721f8269a49SIlpo Järvinen * snd_up-64k-mss .. snd_up cannot be large. However, taking into 2722f8269a49SIlpo Järvinen * account rare use of URG, this is not a big flaw. 2723f8269a49SIlpo Järvinen * 27246ba8a3b1SNandita Dukkipati * Send at most one packet when push_one > 0. Temporarily ignore 27256ba8a3b1SNandita Dukkipati * cwnd limit to force at most one packet out when push_one == 2. 27266ba8a3b1SNandita Dukkipati 2727a2a385d6SEric Dumazet * Returns true, if no segments are in flight and we have queued segments, 2728a2a385d6SEric Dumazet * but cannot send anything now because of SWS or another problem. 27291da177e4SLinus Torvalds */ 2730a2a385d6SEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 2731d5dd9175SIlpo Järvinen int push_one, gfp_t gfp) 27321da177e4SLinus Torvalds { 27331da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 273492df7b51SDavid S. Miller struct sk_buff *skb; 2735c1b4a7e6SDavid S. Miller unsigned int tso_segs, sent_pkts; 273622555032SEric Dumazet u32 cwnd_quota, max_segs; 27375d424d5aSJohn Heffner int result; 27385615f886SFrancis Yan bool is_cwnd_limited = false, is_rwnd_limited = false; 27391da177e4SLinus Torvalds 2740c1b4a7e6SDavid S. Miller sent_pkts = 0; 27415d424d5aSJohn Heffner 2742ee1836aeSEric Dumazet tcp_mstamp_refresh(tp); 2743d5dd9175SIlpo Järvinen if (!push_one) { 27445d424d5aSJohn Heffner /* Do MTU probing. */ 2745d5dd9175SIlpo Järvinen result = tcp_mtu_probe(sk); 2746d5dd9175SIlpo Järvinen if (!result) { 2747a2a385d6SEric Dumazet return false; 27485d424d5aSJohn Heffner } else if (result > 0) { 27495d424d5aSJohn Heffner sent_pkts = 1; 27505d424d5aSJohn Heffner } 2751d5dd9175SIlpo Järvinen } 27525d424d5aSJohn Heffner 2753ed6e7268SNeal Cardwell max_segs = tcp_tso_segs(sk, mss_now); 2754fe067e8aSDavid S. Miller while ((skb = tcp_send_head(sk))) { 2755c8ac3774SHerbert Xu unsigned int limit; 2756*8ee602c6SEric Dumazet int missing_bytes; 2757c8ac3774SHerbert Xu 275879861919SEric Dumazet if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) { 275979861919SEric Dumazet /* "skb_mstamp_ns" is used as a start point for the retransmit timer */ 2760a1ac9c8aSMartin KaFai Lau tp->tcp_wstamp_ns = tp->tcp_clock_cache; 2761a1ac9c8aSMartin KaFai Lau skb_set_delivery_time(skb, tp->tcp_wstamp_ns, true); 276279861919SEric Dumazet list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue); 2763bf50b606SEric Dumazet tcp_init_tso_segs(skb, mss_now); 276479861919SEric Dumazet goto repair; /* Skip network transmission */ 276579861919SEric Dumazet } 276679861919SEric Dumazet 2767218af599SEric Dumazet if (tcp_pacing_check(sk)) 2768218af599SEric Dumazet break; 2769218af599SEric Dumazet 277022555032SEric Dumazet cwnd_quota = tcp_cwnd_test(tp); 27716ba8a3b1SNandita Dukkipati if (!cwnd_quota) { 27726ba8a3b1SNandita Dukkipati if (push_one == 2) 27736ba8a3b1SNandita Dukkipati /* Force out a loss probe pkt. */ 27746ba8a3b1SNandita Dukkipati cwnd_quota = 1; 27756ba8a3b1SNandita Dukkipati else 2776b68e9f85SHerbert Xu break; 27776ba8a3b1SNandita Dukkipati } 2778*8ee602c6SEric Dumazet cwnd_quota = min(cwnd_quota, max_segs); 2779*8ee602c6SEric Dumazet missing_bytes = cwnd_quota * mss_now - skb->len; 2780*8ee602c6SEric Dumazet if (missing_bytes > 0) 2781*8ee602c6SEric Dumazet tcp_grow_skb(sk, skb, missing_bytes); 2782b68e9f85SHerbert Xu 2783d5b38a71SEric Dumazet tso_segs = tcp_set_skb_tso_segs(skb, mss_now); 2784d5b38a71SEric Dumazet 27855615f886SFrancis Yan if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) { 27865615f886SFrancis Yan is_rwnd_limited = true; 2787b68e9f85SHerbert Xu break; 27885615f886SFrancis Yan } 2789b68e9f85SHerbert Xu 2790d6a4e26aSEric Dumazet if (tso_segs == 1) { 2791aa93466bSDavid S. Miller if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 2792aa93466bSDavid S. Miller (tcp_skb_is_last(sk, skb) ? 2793aa93466bSDavid S. Miller nonagle : TCP_NAGLE_PUSH)))) 2794aa93466bSDavid S. Miller break; 2795c1b4a7e6SDavid S. Miller } else { 2796ca8a2263SNeal Cardwell if (!push_one && 2797605ad7f1SEric Dumazet tcp_tso_should_defer(sk, skb, &is_cwnd_limited, 2798f9bfe4e6SEric Dumazet &is_rwnd_limited, max_segs)) 2799aa93466bSDavid S. Miller break; 2800c1b4a7e6SDavid S. Miller } 2801aa93466bSDavid S. Miller 2802605ad7f1SEric Dumazet limit = mss_now; 2803d6a4e26aSEric Dumazet if (tso_segs > 1 && !tcp_urg_mode(tp)) 2804605ad7f1SEric Dumazet limit = tcp_mss_split_point(sk, skb, mss_now, 2805*8ee602c6SEric Dumazet cwnd_quota, 2806605ad7f1SEric Dumazet nonagle); 2807605ad7f1SEric Dumazet 2808605ad7f1SEric Dumazet if (skb->len > limit && 280956483341SEric Dumazet unlikely(tso_fragment(sk, skb, limit, mss_now, gfp))) 2810605ad7f1SEric Dumazet break; 2811605ad7f1SEric Dumazet 2812f9616c35SEric Dumazet if (tcp_small_queue_check(sk, skb, 0)) 281346d3ceabSEric Dumazet break; 2814c9eeec26SEric Dumazet 28151f85e626SEric Dumazet /* Argh, we hit an empty skb(), presumably a thread 28161f85e626SEric Dumazet * is sleeping in sendmsg()/sk_stream_wait_memory(). 28171f85e626SEric Dumazet * We do not want to send a pure-ack packet and have 28181f85e626SEric Dumazet * a strange looking rtx queue with empty packet(s). 28191f85e626SEric Dumazet */ 28201f85e626SEric Dumazet if (TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq) 28211f85e626SEric Dumazet break; 28221f85e626SEric Dumazet 2823d5dd9175SIlpo Järvinen if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) 28241da177e4SLinus Torvalds break; 28251da177e4SLinus Torvalds 2826ec342325SAndrew Vagin repair: 28271da177e4SLinus Torvalds /* Advance the send_head. This one is sent out. 28281da177e4SLinus Torvalds * This call will increment packets_out. 28291da177e4SLinus Torvalds */ 283066f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 28311da177e4SLinus Torvalds 28321da177e4SLinus Torvalds tcp_minshall_update(tp, mss_now, skb); 2833a262f0cdSNandita Dukkipati sent_pkts += tcp_skb_pcount(skb); 2834d5dd9175SIlpo Järvinen 2835d5dd9175SIlpo Järvinen if (push_one) 2836d5dd9175SIlpo Järvinen break; 28371da177e4SLinus Torvalds } 28381da177e4SLinus Torvalds 28395615f886SFrancis Yan if (is_rwnd_limited) 28405615f886SFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED); 28415615f886SFrancis Yan else 28425615f886SFrancis Yan tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED); 28435615f886SFrancis Yan 284440570375SEric Dumazet is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tcp_snd_cwnd(tp)); 2845299bcb55SNeal Cardwell if (likely(sent_pkts || is_cwnd_limited)) 2846299bcb55SNeal Cardwell tcp_cwnd_validate(sk, is_cwnd_limited); 2847299bcb55SNeal Cardwell 2848aa93466bSDavid S. Miller if (likely(sent_pkts)) { 2849684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 2850684bad11SYuchung Cheng tp->prr_out += sent_pkts; 28516ba8a3b1SNandita Dukkipati 28526ba8a3b1SNandita Dukkipati /* Send one loss probe per tail loss episode. */ 28536ba8a3b1SNandita Dukkipati if (push_one != 2) 2854ed66dfafSNeal Cardwell tcp_schedule_loss_probe(sk, false); 2855a2a385d6SEric Dumazet return false; 28561da177e4SLinus Torvalds } 285775c119afSEric Dumazet return !tp->packets_out && !tcp_write_queue_empty(sk); 28586ba8a3b1SNandita Dukkipati } 28596ba8a3b1SNandita Dukkipati 2860ed66dfafSNeal Cardwell bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto) 28616ba8a3b1SNandita Dukkipati { 28626ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 28636ba8a3b1SNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 28641c2709cfSNeal Cardwell u32 timeout, timeout_us, rto_delta_us; 28652ae21cf5SEric Dumazet int early_retrans; 28666ba8a3b1SNandita Dukkipati 28676ba8a3b1SNandita Dukkipati /* Don't do any loss probe on a Fast Open connection before 3WHS 28686ba8a3b1SNandita Dukkipati * finishes. 28696ba8a3b1SNandita Dukkipati */ 2870d983ea6fSEric Dumazet if (rcu_access_pointer(tp->fastopen_rsk)) 28716ba8a3b1SNandita Dukkipati return false; 28726ba8a3b1SNandita Dukkipati 287352e65865SKuniyuki Iwashima early_retrans = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_early_retrans); 28746ba8a3b1SNandita Dukkipati /* Schedule a loss probe in 2*RTT for SACK capable connections 2875b4f70c3dSNeal Cardwell * not in loss recovery, that are either limited by cwnd or application. 28766ba8a3b1SNandita Dukkipati */ 28772ae21cf5SEric Dumazet if ((early_retrans != 3 && early_retrans != 4) || 2878bec41a11SYuchung Cheng !tp->packets_out || !tcp_is_sack(tp) || 2879b4f70c3dSNeal Cardwell (icsk->icsk_ca_state != TCP_CA_Open && 2880b4f70c3dSNeal Cardwell icsk->icsk_ca_state != TCP_CA_CWR)) 28816ba8a3b1SNandita Dukkipati return false; 28826ba8a3b1SNandita Dukkipati 2883bb4d991aSYuchung Cheng /* Probe timeout is 2*rtt. Add minimum RTO to account 2884f9b99582SYuchung Cheng * for delayed ack when there's one outstanding packet. If no RTT 2885f9b99582SYuchung Cheng * sample is available then probe after TCP_TIMEOUT_INIT. 28866ba8a3b1SNandita Dukkipati */ 2887bb4d991aSYuchung Cheng if (tp->srtt_us) { 28881c2709cfSNeal Cardwell timeout_us = tp->srtt_us >> 2; 28896ba8a3b1SNandita Dukkipati if (tp->packets_out == 1) 28901c2709cfSNeal Cardwell timeout_us += tcp_rto_min_us(sk); 2891bb4d991aSYuchung Cheng else 28921c2709cfSNeal Cardwell timeout_us += TCP_TIMEOUT_MIN_US; 28931c2709cfSNeal Cardwell timeout = usecs_to_jiffies(timeout_us); 2894bb4d991aSYuchung Cheng } else { 2895bb4d991aSYuchung Cheng timeout = TCP_TIMEOUT_INIT; 2896bb4d991aSYuchung Cheng } 28976ba8a3b1SNandita Dukkipati 2898a2815817SNeal Cardwell /* If the RTO formula yields an earlier time, then use that time. */ 2899ed66dfafSNeal Cardwell rto_delta_us = advancing_rto ? 2900ed66dfafSNeal Cardwell jiffies_to_usecs(inet_csk(sk)->icsk_rto) : 2901ed66dfafSNeal Cardwell tcp_rto_delta_us(sk); /* How far in future is RTO? */ 2902a2815817SNeal Cardwell if (rto_delta_us > 0) 2903a2815817SNeal Cardwell timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us)); 29046ba8a3b1SNandita Dukkipati 29058dc242adSEric Dumazet tcp_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, TCP_RTO_MAX); 29066ba8a3b1SNandita Dukkipati return true; 29076ba8a3b1SNandita Dukkipati } 29086ba8a3b1SNandita Dukkipati 29091f3279aeSEric Dumazet /* Thanks to skb fast clones, we can detect if a prior transmit of 29101f3279aeSEric Dumazet * a packet is still in a qdisc or driver queue. 29111f3279aeSEric Dumazet * In this case, there is very little point doing a retransmit ! 29121f3279aeSEric Dumazet */ 2913f4dae54eSEric Dumazet static bool skb_still_in_host_queue(struct sock *sk, 29141f3279aeSEric Dumazet const struct sk_buff *skb) 29151f3279aeSEric Dumazet { 291639bb5e62SEric Dumazet if (unlikely(skb_fclone_busy(sk, skb))) { 2917f4dae54eSEric Dumazet set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags); 2918f4dae54eSEric Dumazet smp_mb__after_atomic(); 2919f4dae54eSEric Dumazet if (skb_fclone_busy(sk, skb)) { 2920c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), 29211f3279aeSEric Dumazet LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES); 29221f3279aeSEric Dumazet return true; 29231f3279aeSEric Dumazet } 2924f4dae54eSEric Dumazet } 29251f3279aeSEric Dumazet return false; 29261f3279aeSEric Dumazet } 29271f3279aeSEric Dumazet 2928b340b264SYuchung Cheng /* When probe timeout (PTO) fires, try send a new segment if possible, else 29296ba8a3b1SNandita Dukkipati * retransmit the last segment. 29306ba8a3b1SNandita Dukkipati */ 29316ba8a3b1SNandita Dukkipati void tcp_send_loss_probe(struct sock *sk) 29326ba8a3b1SNandita Dukkipati { 29339b717a8dSNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 29346ba8a3b1SNandita Dukkipati struct sk_buff *skb; 29356ba8a3b1SNandita Dukkipati int pcount; 29366ba8a3b1SNandita Dukkipati int mss = tcp_current_mss(sk); 29376ba8a3b1SNandita Dukkipati 293876be93fcSYuchung Cheng /* At most one outstanding TLP */ 293976be93fcSYuchung Cheng if (tp->tlp_high_seq) 294076be93fcSYuchung Cheng goto rearm_timer; 294176be93fcSYuchung Cheng 294276be93fcSYuchung Cheng tp->tlp_retrans = 0; 2943b340b264SYuchung Cheng skb = tcp_send_head(sk); 294475c119afSEric Dumazet if (skb && tcp_snd_wnd_test(tp, skb, mss)) { 2945b340b264SYuchung Cheng pcount = tp->packets_out; 2946b340b264SYuchung Cheng tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC); 2947b340b264SYuchung Cheng if (tp->packets_out > pcount) 2948b340b264SYuchung Cheng goto probe_sent; 29496ba8a3b1SNandita Dukkipati goto rearm_timer; 29506ba8a3b1SNandita Dukkipati } 295175c119afSEric Dumazet skb = skb_rb_last(&sk->tcp_rtx_queue); 2952b2b7af86SYuchung Cheng if (unlikely(!skb)) { 2953b2b7af86SYuchung Cheng WARN_ONCE(tp->packets_out, 2954b2b7af86SYuchung Cheng "invalid inflight: %u state %u cwnd %u mss %d\n", 295540570375SEric Dumazet tp->packets_out, sk->sk_state, tcp_snd_cwnd(tp), mss); 2956b2b7af86SYuchung Cheng inet_csk(sk)->icsk_pending = 0; 2957b2b7af86SYuchung Cheng return; 2958b2b7af86SYuchung Cheng } 29596ba8a3b1SNandita Dukkipati 29601f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 29611f3279aeSEric Dumazet goto rearm_timer; 29621f3279aeSEric Dumazet 29636ba8a3b1SNandita Dukkipati pcount = tcp_skb_pcount(skb); 29646ba8a3b1SNandita Dukkipati if (WARN_ON(!pcount)) 29656ba8a3b1SNandita Dukkipati goto rearm_timer; 29666ba8a3b1SNandita Dukkipati 29676ba8a3b1SNandita Dukkipati if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) { 296875c119afSEric Dumazet if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, 296975c119afSEric Dumazet (pcount - 1) * mss, mss, 29706cc55e09SOctavian Purdila GFP_ATOMIC))) 29716ba8a3b1SNandita Dukkipati goto rearm_timer; 297275c119afSEric Dumazet skb = skb_rb_next(skb); 29736ba8a3b1SNandita Dukkipati } 29746ba8a3b1SNandita Dukkipati 29756ba8a3b1SNandita Dukkipati if (WARN_ON(!skb || !tcp_skb_pcount(skb))) 29766ba8a3b1SNandita Dukkipati goto rearm_timer; 29776ba8a3b1SNandita Dukkipati 297810d3be56SEric Dumazet if (__tcp_retransmit_skb(sk, skb, 1)) 2979b340b264SYuchung Cheng goto rearm_timer; 29806ba8a3b1SNandita Dukkipati 298176be93fcSYuchung Cheng tp->tlp_retrans = 1; 298276be93fcSYuchung Cheng 298376be93fcSYuchung Cheng probe_sent: 29849b717a8dSNandita Dukkipati /* Record snd_nxt for loss detection. */ 29859b717a8dSNandita Dukkipati tp->tlp_high_seq = tp->snd_nxt; 29869b717a8dSNandita Dukkipati 2987c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES); 2988fcd16c0aSYuchung Cheng /* Reset s.t. tcp_rearm_rto will restart timer from now */ 2989fcd16c0aSYuchung Cheng inet_csk(sk)->icsk_pending = 0; 2990b340b264SYuchung Cheng rearm_timer: 2991fcd16c0aSYuchung Cheng tcp_rearm_rto(sk); 29921da177e4SLinus Torvalds } 29931da177e4SLinus Torvalds 2994a762a980SDavid S. Miller /* Push out any pending frames which were held back due to 2995a762a980SDavid S. Miller * TCP_CORK or attempt at coalescing tiny packets. 2996a762a980SDavid S. Miller * The socket must be locked by the caller. 2997a762a980SDavid S. Miller */ 29989e412ba7SIlpo Järvinen void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 29999e412ba7SIlpo Järvinen int nonagle) 3000a762a980SDavid S. Miller { 3001726e07a8SIlpo Järvinen /* If we are closed, the bytes will have to remain here. 3002726e07a8SIlpo Järvinen * In time closedown will finish, we empty the write queue and 3003726e07a8SIlpo Järvinen * all will be happy. 3004726e07a8SIlpo Järvinen */ 3005726e07a8SIlpo Järvinen if (unlikely(sk->sk_state == TCP_CLOSE)) 3006726e07a8SIlpo Järvinen return; 3007726e07a8SIlpo Järvinen 300899a1dec7SMel Gorman if (tcp_write_xmit(sk, cur_mss, nonagle, 0, 30097450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC))) 30109e412ba7SIlpo Järvinen tcp_check_probe_timer(sk); 3011a762a980SDavid S. Miller } 3012a762a980SDavid S. Miller 3013c1b4a7e6SDavid S. Miller /* Send _single_ skb sitting at the send head. This function requires 3014c1b4a7e6SDavid S. Miller * true push pending frames to setup probe timer etc. 3015c1b4a7e6SDavid S. Miller */ 3016c1b4a7e6SDavid S. Miller void tcp_push_one(struct sock *sk, unsigned int mss_now) 3017c1b4a7e6SDavid S. Miller { 3018fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 3019c1b4a7e6SDavid S. Miller 3020c1b4a7e6SDavid S. Miller BUG_ON(!skb || skb->len < mss_now); 3021c1b4a7e6SDavid S. Miller 3022d5dd9175SIlpo Järvinen tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); 3023c1b4a7e6SDavid S. Miller } 3024c1b4a7e6SDavid S. Miller 30251da177e4SLinus Torvalds /* This function returns the amount that we can raise the 30261da177e4SLinus Torvalds * usable window based on the following constraints 30271da177e4SLinus Torvalds * 30281da177e4SLinus Torvalds * 1. The window can never be shrunk once it is offered (RFC 793) 30291da177e4SLinus Torvalds * 2. We limit memory per socket 30301da177e4SLinus Torvalds * 30311da177e4SLinus Torvalds * RFC 1122: 30321da177e4SLinus Torvalds * "the suggested [SWS] avoidance algorithm for the receiver is to keep 30331da177e4SLinus Torvalds * RECV.NEXT + RCV.WIN fixed until: 30341da177e4SLinus Torvalds * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 30351da177e4SLinus Torvalds * 30361da177e4SLinus Torvalds * i.e. don't raise the right edge of the window until you can raise 30371da177e4SLinus Torvalds * it at least MSS bytes. 30381da177e4SLinus Torvalds * 30391da177e4SLinus Torvalds * Unfortunately, the recommended algorithm breaks header prediction, 30401da177e4SLinus Torvalds * since header prediction assumes th->window stays fixed. 30411da177e4SLinus Torvalds * 30421da177e4SLinus Torvalds * Strictly speaking, keeping th->window fixed violates the receiver 30431da177e4SLinus Torvalds * side SWS prevention criteria. The problem is that under this rule 30441da177e4SLinus Torvalds * a stream of single byte packets will cause the right side of the 30451da177e4SLinus Torvalds * window to always advance by a single byte. 30461da177e4SLinus Torvalds * 30471da177e4SLinus Torvalds * Of course, if the sender implements sender side SWS prevention 30481da177e4SLinus Torvalds * then this will not be a problem. 30491da177e4SLinus Torvalds * 30501da177e4SLinus Torvalds * BSD seems to make the following compromise: 30511da177e4SLinus Torvalds * 30521da177e4SLinus Torvalds * If the free space is less than the 1/4 of the maximum 30531da177e4SLinus Torvalds * space available and the free space is less than 1/2 mss, 30541da177e4SLinus Torvalds * then set the window to 0. 30551da177e4SLinus Torvalds * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 30561da177e4SLinus Torvalds * Otherwise, just prevent the window from shrinking 30571da177e4SLinus Torvalds * and from being larger than the largest representable value. 30581da177e4SLinus Torvalds * 30591da177e4SLinus Torvalds * This prevents incremental opening of the window in the regime 30601da177e4SLinus Torvalds * where TCP is limited by the speed of the reader side taking 30611da177e4SLinus Torvalds * data out of the TCP receive queue. It does nothing about 30621da177e4SLinus Torvalds * those cases where the window is constrained on the sender side 30631da177e4SLinus Torvalds * because the pipeline is full. 30641da177e4SLinus Torvalds * 30651da177e4SLinus Torvalds * BSD also seems to "accidentally" limit itself to windows that are a 30661da177e4SLinus Torvalds * multiple of MSS, at least until the free space gets quite small. 30671da177e4SLinus Torvalds * This would appear to be a side effect of the mbuf implementation. 30681da177e4SLinus Torvalds * Combining these two algorithms results in the observed behavior 30691da177e4SLinus Torvalds * of having a fixed window size at almost all times. 30701da177e4SLinus Torvalds * 30711da177e4SLinus Torvalds * Below we obtain similar behavior by forcing the offered window to 30721da177e4SLinus Torvalds * a multiple of the mss when it is feasible to do so. 30731da177e4SLinus Torvalds * 30741da177e4SLinus Torvalds * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 30751da177e4SLinus Torvalds * Regular options like TIMESTAMP are taken into account. 30761da177e4SLinus Torvalds */ 30771da177e4SLinus Torvalds u32 __tcp_select_window(struct sock *sk) 30781da177e4SLinus Torvalds { 3079463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 30801da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 3081b650d953Smfreemon@cloudflare.com struct net *net = sock_net(sk); 3082caa20d9aSStephen Hemminger /* MSS for the peer's data. Previous versions used mss_clamp 30831da177e4SLinus Torvalds * here. I don't know if the value based on our guesses 30841da177e4SLinus Torvalds * of peer's MSS is better for the performance. It's more correct 30851da177e4SLinus Torvalds * but may be worse for the performance because of rcv_mss 30861da177e4SLinus Torvalds * fluctuations. --SAW 1998/11/1 30871da177e4SLinus Torvalds */ 3088463c84b9SArnaldo Carvalho de Melo int mss = icsk->icsk_ack.rcv_mss; 30891da177e4SLinus Torvalds int free_space = tcp_space(sk); 309086c1a045SFlorian Westphal int allowed_space = tcp_full_space(sk); 3091071c8ed6SFlorian Westphal int full_space, window; 3092071c8ed6SFlorian Westphal 3093071c8ed6SFlorian Westphal if (sk_is_mptcp(sk)) 3094071c8ed6SFlorian Westphal mptcp_space(sk, &free_space, &allowed_space); 3095071c8ed6SFlorian Westphal 3096071c8ed6SFlorian Westphal full_space = min_t(int, tp->window_clamp, allowed_space); 30971da177e4SLinus Torvalds 309806425c30SEric Dumazet if (unlikely(mss > full_space)) { 30991da177e4SLinus Torvalds mss = full_space; 310006425c30SEric Dumazet if (mss <= 0) 310106425c30SEric Dumazet return 0; 310206425c30SEric Dumazet } 3103b650d953Smfreemon@cloudflare.com 3104b650d953Smfreemon@cloudflare.com /* Only allow window shrink if the sysctl is enabled and we have 3105b650d953Smfreemon@cloudflare.com * a non-zero scaling factor in effect. 3106b650d953Smfreemon@cloudflare.com */ 3107b650d953Smfreemon@cloudflare.com if (READ_ONCE(net->ipv4.sysctl_tcp_shrink_window) && tp->rx_opt.rcv_wscale) 3108b650d953Smfreemon@cloudflare.com goto shrink_window_allowed; 3109b650d953Smfreemon@cloudflare.com 3110b650d953Smfreemon@cloudflare.com /* do not allow window to shrink */ 3111b650d953Smfreemon@cloudflare.com 3112b92edbe0SEric Dumazet if (free_space < (full_space >> 1)) { 3113463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.quick = 0; 31141da177e4SLinus Torvalds 3115b8da51ebSEric Dumazet if (tcp_under_memory_pressure(sk)) 3116053f3684SWei Wang tcp_adjust_rcv_ssthresh(sk); 31171da177e4SLinus Torvalds 311886c1a045SFlorian Westphal /* free_space might become our new window, make sure we don't 311986c1a045SFlorian Westphal * increase it due to wscale. 312086c1a045SFlorian Westphal */ 312186c1a045SFlorian Westphal free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale); 312286c1a045SFlorian Westphal 312386c1a045SFlorian Westphal /* if free space is less than mss estimate, or is below 1/16th 312486c1a045SFlorian Westphal * of the maximum allowed, try to move to zero-window, else 312586c1a045SFlorian Westphal * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and 312686c1a045SFlorian Westphal * new incoming data is dropped due to memory limits. 312786c1a045SFlorian Westphal * With large window, mss test triggers way too late in order 312886c1a045SFlorian Westphal * to announce zero window in time before rmem limit kicks in. 312986c1a045SFlorian Westphal */ 313086c1a045SFlorian Westphal if (free_space < (allowed_space >> 4) || free_space < mss) 31311da177e4SLinus Torvalds return 0; 31321da177e4SLinus Torvalds } 31331da177e4SLinus Torvalds 31341da177e4SLinus Torvalds if (free_space > tp->rcv_ssthresh) 31351da177e4SLinus Torvalds free_space = tp->rcv_ssthresh; 31361da177e4SLinus Torvalds 31371da177e4SLinus Torvalds /* Don't do rounding if we are using window scaling, since the 31381da177e4SLinus Torvalds * scaled window will not line up with the MSS boundary anyway. 31391da177e4SLinus Torvalds */ 31401da177e4SLinus Torvalds if (tp->rx_opt.rcv_wscale) { 31411da177e4SLinus Torvalds window = free_space; 31421da177e4SLinus Torvalds 31431da177e4SLinus Torvalds /* Advertise enough space so that it won't get scaled away. 31441da177e4SLinus Torvalds * Import case: prevent zero window announcement if 31451da177e4SLinus Torvalds * 1<<rcv_wscale > mss. 31461da177e4SLinus Torvalds */ 31471935299dSGao Feng window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale)); 31481da177e4SLinus Torvalds } else { 31491935299dSGao Feng window = tp->rcv_wnd; 31501da177e4SLinus Torvalds /* Get the largest window that is a nice multiple of mss. 31511da177e4SLinus Torvalds * Window clamp already applied above. 31521da177e4SLinus Torvalds * If our current window offering is within 1 mss of the 31531da177e4SLinus Torvalds * free space we just keep it. This prevents the divide 31541da177e4SLinus Torvalds * and multiply from happening most of the time. 31551da177e4SLinus Torvalds * We also don't do any window rounding when the free space 31561da177e4SLinus Torvalds * is too small. 31571da177e4SLinus Torvalds */ 31581da177e4SLinus Torvalds if (window <= free_space - mss || window > free_space) 31591935299dSGao Feng window = rounddown(free_space, mss); 316084565070SJohn Heffner else if (mss == full_space && 3161b92edbe0SEric Dumazet free_space > window + (full_space >> 1)) 316284565070SJohn Heffner window = free_space; 31631da177e4SLinus Torvalds } 31641da177e4SLinus Torvalds 31651da177e4SLinus Torvalds return window; 3166b650d953Smfreemon@cloudflare.com 3167b650d953Smfreemon@cloudflare.com shrink_window_allowed: 3168b650d953Smfreemon@cloudflare.com /* new window should always be an exact multiple of scaling factor */ 3169b650d953Smfreemon@cloudflare.com free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale); 3170b650d953Smfreemon@cloudflare.com 3171b650d953Smfreemon@cloudflare.com if (free_space < (full_space >> 1)) { 3172b650d953Smfreemon@cloudflare.com icsk->icsk_ack.quick = 0; 3173b650d953Smfreemon@cloudflare.com 3174b650d953Smfreemon@cloudflare.com if (tcp_under_memory_pressure(sk)) 3175b650d953Smfreemon@cloudflare.com tcp_adjust_rcv_ssthresh(sk); 3176b650d953Smfreemon@cloudflare.com 3177b650d953Smfreemon@cloudflare.com /* if free space is too low, return a zero window */ 3178b650d953Smfreemon@cloudflare.com if (free_space < (allowed_space >> 4) || free_space < mss || 3179b650d953Smfreemon@cloudflare.com free_space < (1 << tp->rx_opt.rcv_wscale)) 3180b650d953Smfreemon@cloudflare.com return 0; 3181b650d953Smfreemon@cloudflare.com } 3182b650d953Smfreemon@cloudflare.com 3183b650d953Smfreemon@cloudflare.com if (free_space > tp->rcv_ssthresh) { 3184b650d953Smfreemon@cloudflare.com free_space = tp->rcv_ssthresh; 3185b650d953Smfreemon@cloudflare.com /* new window should always be an exact multiple of scaling factor 3186b650d953Smfreemon@cloudflare.com * 3187b650d953Smfreemon@cloudflare.com * For this case, we ALIGN "up" (increase free_space) because 3188b650d953Smfreemon@cloudflare.com * we know free_space is not zero here, it has been reduced from 3189b650d953Smfreemon@cloudflare.com * the memory-based limit, and rcv_ssthresh is not a hard limit 3190b650d953Smfreemon@cloudflare.com * (unlike sk_rcvbuf). 3191b650d953Smfreemon@cloudflare.com */ 3192b650d953Smfreemon@cloudflare.com free_space = ALIGN(free_space, (1 << tp->rx_opt.rcv_wscale)); 3193b650d953Smfreemon@cloudflare.com } 3194b650d953Smfreemon@cloudflare.com 3195b650d953Smfreemon@cloudflare.com return free_space; 31961da177e4SLinus Torvalds } 31971da177e4SLinus Torvalds 3198cfea5a68SMartin KaFai Lau void tcp_skb_collapse_tstamp(struct sk_buff *skb, 3199082ac2d5SMartin KaFai Lau const struct sk_buff *next_skb) 3200082ac2d5SMartin KaFai Lau { 32010a2cf20cSSoheil Hassas Yeganeh if (unlikely(tcp_has_tx_tstamp(next_skb))) { 32020a2cf20cSSoheil Hassas Yeganeh const struct skb_shared_info *next_shinfo = 32030a2cf20cSSoheil Hassas Yeganeh skb_shinfo(next_skb); 3204082ac2d5SMartin KaFai Lau struct skb_shared_info *shinfo = skb_shinfo(skb); 3205082ac2d5SMartin KaFai Lau 32060a2cf20cSSoheil Hassas Yeganeh shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP; 3207082ac2d5SMartin KaFai Lau shinfo->tskey = next_shinfo->tskey; 32082de8023eSMartin KaFai Lau TCP_SKB_CB(skb)->txstamp_ack |= 32092de8023eSMartin KaFai Lau TCP_SKB_CB(next_skb)->txstamp_ack; 3210082ac2d5SMartin KaFai Lau } 3211082ac2d5SMartin KaFai Lau } 3212082ac2d5SMartin KaFai Lau 32134a17fc3aSIlpo Järvinen /* Collapses two adjacent SKB's during retransmission. */ 3214f8071cdeSEric Dumazet static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) 32151da177e4SLinus Torvalds { 32161da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 321775c119afSEric Dumazet struct sk_buff *next_skb = skb_rb_next(skb); 321813dde04fSWei Yongjun int next_skb_size; 32191da177e4SLinus Torvalds 3220058dc334SIlpo Järvinen next_skb_size = next_skb->len; 32211da177e4SLinus Torvalds 3222058dc334SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); 32231da177e4SLinus Torvalds 3224bd446314SEric Dumazet if (next_skb_size && !tcp_skb_shift(skb, next_skb, 1, next_skb_size)) 3225f8071cdeSEric Dumazet return false; 3226bd446314SEric Dumazet 32272b7cda9cSEric Dumazet tcp_highest_sack_replace(sk, next_skb, skb); 3228a6963a6bSIlpo Järvinen 32291da177e4SLinus Torvalds /* Update sequence range on original skb. */ 32301da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 32311da177e4SLinus Torvalds 3232e6c7d085SIlpo Järvinen /* Merge over control information. This moves PSH/FIN etc. over */ 32334de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; 32341da177e4SLinus Torvalds 32351da177e4SLinus Torvalds /* All done, get rid of second SKB and account for it so 32361da177e4SLinus Torvalds * packet counting does not break. 32371da177e4SLinus Torvalds */ 32384828e7f4SIlpo Järvinen TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; 3239a643b5d4SMartin KaFai Lau TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor; 3240b7689205SIlpo Järvinen 3241b7689205SIlpo Järvinen /* changed transmit queue under us so clear hints */ 3242ef9da47cSIlpo Järvinen tcp_clear_retrans_hints_partial(tp); 3243ef9da47cSIlpo Järvinen if (next_skb == tp->retransmit_skb_hint) 3244ef9da47cSIlpo Järvinen tp->retransmit_skb_hint = skb; 3245b7689205SIlpo Järvinen 3246797108d1SIlpo Järvinen tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); 3247797108d1SIlpo Järvinen 3248082ac2d5SMartin KaFai Lau tcp_skb_collapse_tstamp(skb, next_skb); 3249082ac2d5SMartin KaFai Lau 325075c119afSEric Dumazet tcp_rtx_queue_unlink_and_free(next_skb, sk); 3251f8071cdeSEric Dumazet return true; 32521da177e4SLinus Torvalds } 32531da177e4SLinus Torvalds 325467edfef7SAndi Kleen /* Check if coalescing SKBs is legal. */ 3255a2a385d6SEric Dumazet static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) 32564a17fc3aSIlpo Järvinen { 32574a17fc3aSIlpo Järvinen if (tcp_skb_pcount(skb) > 1) 3258a2a385d6SEric Dumazet return false; 32594a17fc3aSIlpo Järvinen if (skb_cloned(skb)) 3260a2a385d6SEric Dumazet return false; 32612331ccc5SEric Dumazet /* Some heuristics for collapsing over SACK'd could be invented */ 32624a17fc3aSIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 3263a2a385d6SEric Dumazet return false; 32644a17fc3aSIlpo Järvinen 3265a2a385d6SEric Dumazet return true; 32664a17fc3aSIlpo Järvinen } 32674a17fc3aSIlpo Järvinen 326867edfef7SAndi Kleen /* Collapse packets in the retransmit queue to make to create 326967edfef7SAndi Kleen * less packets on the wire. This is only done on retransmission. 327067edfef7SAndi Kleen */ 32714a17fc3aSIlpo Järvinen static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, 32724a17fc3aSIlpo Järvinen int space) 32734a17fc3aSIlpo Järvinen { 32744a17fc3aSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 32754a17fc3aSIlpo Järvinen struct sk_buff *skb = to, *tmp; 3276a2a385d6SEric Dumazet bool first = true; 32774a17fc3aSIlpo Järvinen 32781a63cb91SKuniyuki Iwashima if (!READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse)) 32794a17fc3aSIlpo Järvinen return; 32804de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 32814a17fc3aSIlpo Järvinen return; 32824a17fc3aSIlpo Järvinen 328375c119afSEric Dumazet skb_rbtree_walk_from_safe(skb, tmp) { 32844a17fc3aSIlpo Järvinen if (!tcp_can_collapse(sk, skb)) 32854a17fc3aSIlpo Järvinen break; 32864a17fc3aSIlpo Järvinen 328785712484SMat Martineau if (!tcp_skb_can_collapse(to, skb)) 3288a643b5d4SMartin KaFai Lau break; 3289a643b5d4SMartin KaFai Lau 32904a17fc3aSIlpo Järvinen space -= skb->len; 32914a17fc3aSIlpo Järvinen 32924a17fc3aSIlpo Järvinen if (first) { 3293a2a385d6SEric Dumazet first = false; 32944a17fc3aSIlpo Järvinen continue; 32954a17fc3aSIlpo Järvinen } 32964a17fc3aSIlpo Järvinen 32974a17fc3aSIlpo Järvinen if (space < 0) 32984a17fc3aSIlpo Järvinen break; 32994a17fc3aSIlpo Järvinen 33004a17fc3aSIlpo Järvinen if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) 33014a17fc3aSIlpo Järvinen break; 33024a17fc3aSIlpo Järvinen 3303f8071cdeSEric Dumazet if (!tcp_collapse_retrans(sk, to)) 3304f8071cdeSEric Dumazet break; 33054a17fc3aSIlpo Järvinen } 33064a17fc3aSIlpo Järvinen } 33074a17fc3aSIlpo Järvinen 33081da177e4SLinus Torvalds /* This retransmits one SKB. Policy decisions and retransmit queue 33091da177e4SLinus Torvalds * state updates are done by the caller. Returns non-zero if an 33101da177e4SLinus Torvalds * error occurred which prevented the send. 33111da177e4SLinus Torvalds */ 331210d3be56SEric Dumazet int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) 33131da177e4SLinus Torvalds { 33145d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 331510d3be56SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 33167d227cd2SSridhar Samudrala unsigned int cur_mss; 331710d3be56SEric Dumazet int diff, len, err; 3318536a6c8eSYonglong Li int avail_wnd; 331910d3be56SEric Dumazet 332010d3be56SEric Dumazet /* Inconclusive MTU probe */ 332110d3be56SEric Dumazet if (icsk->icsk_mtup.probe_size) 33225d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 33235d424d5aSJohn Heffner 33241f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 33251f3279aeSEric Dumazet return -EBUSY; 33261f3279aeSEric Dumazet 3327f99cd562SDong Chenchen start: 33281da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 3329f99cd562SDong Chenchen if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 3330f99cd562SDong Chenchen TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_SYN; 3331f99cd562SDong Chenchen TCP_SKB_CB(skb)->seq++; 3332f99cd562SDong Chenchen goto start; 3333f99cd562SDong Chenchen } 33347f582b24SEric Dumazet if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) { 33357f582b24SEric Dumazet WARN_ON_ONCE(1); 33367f582b24SEric Dumazet return -EINVAL; 33377f582b24SEric Dumazet } 33381da177e4SLinus Torvalds if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 33391da177e4SLinus Torvalds return -ENOMEM; 33401da177e4SLinus Torvalds } 33411da177e4SLinus Torvalds 33427d227cd2SSridhar Samudrala if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 33437d227cd2SSridhar Samudrala return -EHOSTUNREACH; /* Routing failure or similar. */ 33447d227cd2SSridhar Samudrala 33450c54b85fSIlpo Järvinen cur_mss = tcp_current_mss(sk); 3346536a6c8eSYonglong Li avail_wnd = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 33477d227cd2SSridhar Samudrala 33481da177e4SLinus Torvalds /* If receiver has shrunk his window, and skb is out of 33491da177e4SLinus Torvalds * new window, do not retransmit it. The exception is the 33501da177e4SLinus Torvalds * case, when window is shrunk to zero. In this case 3351536a6c8eSYonglong Li * our retransmit of one segment serves as a zero window probe. 33521da177e4SLinus Torvalds */ 3353536a6c8eSYonglong Li if (avail_wnd <= 0) { 3354536a6c8eSYonglong Li if (TCP_SKB_CB(skb)->seq != tp->snd_una) 33551da177e4SLinus Torvalds return -EAGAIN; 3356536a6c8eSYonglong Li avail_wnd = cur_mss; 3357536a6c8eSYonglong Li } 33581da177e4SLinus Torvalds 335910d3be56SEric Dumazet len = cur_mss * segs; 3360536a6c8eSYonglong Li if (len > avail_wnd) { 3361536a6c8eSYonglong Li len = rounddown(avail_wnd, cur_mss); 3362536a6c8eSYonglong Li if (!len) 3363536a6c8eSYonglong Li len = avail_wnd; 3364536a6c8eSYonglong Li } 336510d3be56SEric Dumazet if (skb->len > len) { 336675c119afSEric Dumazet if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len, 336775c119afSEric Dumazet cur_mss, GFP_ATOMIC)) 33681da177e4SLinus Torvalds return -ENOMEM; /* We'll try again later. */ 336902276f3cSIlpo Järvinen } else { 3370c4777efaSEric Dumazet if (skb_unclone_keeptruesize(skb, GFP_ATOMIC)) 3371c52e2421SEric Dumazet return -ENOMEM; 337210d3be56SEric Dumazet 337310d3be56SEric Dumazet diff = tcp_skb_pcount(skb); 337410d3be56SEric Dumazet tcp_set_skb_tso_segs(skb, cur_mss); 337510d3be56SEric Dumazet diff -= tcp_skb_pcount(skb); 337610d3be56SEric Dumazet if (diff) 337710d3be56SEric Dumazet tcp_adjust_pcount(sk, skb, diff); 3378536a6c8eSYonglong Li avail_wnd = min_t(int, avail_wnd, cur_mss); 3379536a6c8eSYonglong Li if (skb->len < avail_wnd) 3380536a6c8eSYonglong Li tcp_retrans_try_collapse(sk, skb, avail_wnd); 33811da177e4SLinus Torvalds } 33821da177e4SLinus Torvalds 338349213555SDaniel Borkmann /* RFC3168, section 6.1.1.1. ECN fallback */ 338449213555SDaniel Borkmann if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN) 338549213555SDaniel Borkmann tcp_ecn_clear_syn(sk, skb); 338649213555SDaniel Borkmann 3387678550c6SYuchung Cheng /* Update global and local TCP statistics. */ 3388678550c6SYuchung Cheng segs = tcp_skb_pcount(skb); 3389678550c6SYuchung Cheng TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs); 3390678550c6SYuchung Cheng if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 3391678550c6SYuchung Cheng __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 3392678550c6SYuchung Cheng tp->total_retrans += segs; 3393fb31c9b9SWei Wang tp->bytes_retrans += skb->len; 3394678550c6SYuchung Cheng 339550bceae9SThomas Graf /* make sure skb->data is aligned on arches that require it 339650bceae9SThomas Graf * and check if ack-trimming & collapsing extended the headroom 339750bceae9SThomas Graf * beyond what csum_start can cover. 339850bceae9SThomas Graf */ 339950bceae9SThomas Graf if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) || 340050bceae9SThomas Graf skb_headroom(skb) >= 0xFFFF)) { 340110a81980SEric Dumazet struct sk_buff *nskb; 340210a81980SEric Dumazet 3403e2080072SEric Dumazet tcp_skb_tsorted_save(skb) { 340410a81980SEric Dumazet nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC); 340507f8e4d0SFlorian Westphal if (nskb) { 340607f8e4d0SFlorian Westphal nskb->dev = NULL; 340707f8e4d0SFlorian Westphal err = tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC); 340807f8e4d0SFlorian Westphal } else { 340907f8e4d0SFlorian Westphal err = -ENOBUFS; 341007f8e4d0SFlorian Westphal } 3411e2080072SEric Dumazet } tcp_skb_tsorted_restore(skb); 3412e2080072SEric Dumazet 34135889e2c0SYousuk Seung if (!err) { 3414a7a25630SEric Dumazet tcp_update_skb_after_send(sk, skb, tp->tcp_wstamp_ns); 34155889e2c0SYousuk Seung tcp_rate_skb_sent(sk, skb); 34165889e2c0SYousuk Seung } 3417117632e6SEric Dumazet } else { 3418c84a5711SYuchung Cheng err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 3419117632e6SEric Dumazet } 3420c84a5711SYuchung Cheng 34217f12422cSYuchung Cheng /* To avoid taking spuriously low RTT samples based on a timestamp 34227f12422cSYuchung Cheng * for a transmit that never happened, always mark EVER_RETRANS 34237f12422cSYuchung Cheng */ 34247f12422cSYuchung Cheng TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS; 34257f12422cSYuchung Cheng 3426a31ad29eSLawrence Brakmo if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RETRANS_CB_FLAG)) 3427a31ad29eSLawrence Brakmo tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RETRANS_CB, 3428a31ad29eSLawrence Brakmo TCP_SKB_CB(skb)->seq, segs, err); 3429a31ad29eSLawrence Brakmo 3430fc9f3501SEric Dumazet if (likely(!err)) { 3431e086101bSCong Wang trace_tcp_retransmit_skb(sk, skb); 3432678550c6SYuchung Cheng } else if (err != -EBUSY) { 3433ec641b39SYuchung Cheng NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL, segs); 3434fc9f3501SEric Dumazet } 3435c84a5711SYuchung Cheng return err; 343693b174adSYuchung Cheng } 343793b174adSYuchung Cheng 343810d3be56SEric Dumazet int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) 343993b174adSYuchung Cheng { 344093b174adSYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 344110d3be56SEric Dumazet int err = __tcp_retransmit_skb(sk, skb, segs); 34421da177e4SLinus Torvalds 34431da177e4SLinus Torvalds if (err == 0) { 34441da177e4SLinus Torvalds #if FASTRETRANS_DEBUG > 0 34451da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 3446e87cc472SJoe Perches net_dbg_ratelimited("retrans_out leaked\n"); 34471da177e4SLinus Torvalds } 34481da177e4SLinus Torvalds #endif 34491da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 34501da177e4SLinus Torvalds tp->retrans_out += tcp_skb_pcount(skb); 34517ae18975SYuchung Cheng } 34521da177e4SLinus Torvalds 34537ae18975SYuchung Cheng /* Save stamp of the first (attempted) retransmit. */ 34541da177e4SLinus Torvalds if (!tp->retrans_stamp) 3455614e8316SEric Dumazet tp->retrans_stamp = tcp_skb_timestamp_ts(tp->tcp_usec_ts, skb); 34561da177e4SLinus Torvalds 34576e08d5e3SYuchung Cheng if (tp->undo_retrans < 0) 34586e08d5e3SYuchung Cheng tp->undo_retrans = 0; 34596e08d5e3SYuchung Cheng tp->undo_retrans += tcp_skb_pcount(skb); 34601da177e4SLinus Torvalds return err; 34611da177e4SLinus Torvalds } 34621da177e4SLinus Torvalds 34631da177e4SLinus Torvalds /* This gets called after a retransmit timeout, and the initially 34641da177e4SLinus Torvalds * retransmitted data is acknowledged. It tries to continue 34651da177e4SLinus Torvalds * resending the rest of the retransmit queue, until either 34661da177e4SLinus Torvalds * we've sent it all or the congestion window limit is reached. 34671da177e4SLinus Torvalds */ 34681da177e4SLinus Torvalds void tcp_xmit_retransmit_queue(struct sock *sk) 34691da177e4SLinus Torvalds { 34706687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 3471b9f1f1ceSEric Dumazet struct sk_buff *skb, *rtx_head, *hole = NULL; 34721da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 3473916e6d1aSEric Dumazet bool rearm_timer = false; 3474840a3cbeSYuchung Cheng u32 max_segs; 347561eb55f4SIlpo Järvinen int mib_idx; 34766a438bbeSStephen Hemminger 347745e77d31SIlpo Järvinen if (!tp->packets_out) 347845e77d31SIlpo Järvinen return; 347945e77d31SIlpo Järvinen 348075c119afSEric Dumazet rtx_head = tcp_rtx_queue_head(sk); 3481b9f1f1ceSEric Dumazet skb = tp->retransmit_skb_hint ?: rtx_head; 3482ed6e7268SNeal Cardwell max_segs = tcp_tso_segs(sk, tcp_current_mss(sk)); 348375c119afSEric Dumazet skb_rbtree_walk_from(skb) { 3484dca0aaf8SEric Dumazet __u8 sacked; 348510d3be56SEric Dumazet int segs; 34861da177e4SLinus Torvalds 3487218af599SEric Dumazet if (tcp_pacing_check(sk)) 3488218af599SEric Dumazet break; 3489218af599SEric Dumazet 34906a438bbeSStephen Hemminger /* we could do better than to assign each time */ 349151456b29SIan Morris if (!hole) 34926a438bbeSStephen Hemminger tp->retransmit_skb_hint = skb; 34936a438bbeSStephen Hemminger 349440570375SEric Dumazet segs = tcp_snd_cwnd(tp) - tcp_packets_in_flight(tp); 349510d3be56SEric Dumazet if (segs <= 0) 3496916e6d1aSEric Dumazet break; 3497dca0aaf8SEric Dumazet sacked = TCP_SKB_CB(skb)->sacked; 3498a3d2e9f8SEric Dumazet /* In case tcp_shift_skb_data() have aggregated large skbs, 3499a3d2e9f8SEric Dumazet * we need to make sure not sending too bigs TSO packets 3500a3d2e9f8SEric Dumazet */ 3501a3d2e9f8SEric Dumazet segs = min_t(int, segs, max_segs); 35020e1c54c2SIlpo Järvinen 3503840a3cbeSYuchung Cheng if (tp->retrans_out >= tp->lost_out) { 3504006f582cSIlpo Järvinen break; 35050e1c54c2SIlpo Järvinen } else if (!(sacked & TCPCB_LOST)) { 350651456b29SIan Morris if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) 35070e1c54c2SIlpo Järvinen hole = skb; 350861eb55f4SIlpo Järvinen continue; 35091da177e4SLinus Torvalds 35100e1c54c2SIlpo Järvinen } else { 35110e1c54c2SIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Loss) 35120e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFASTRETRANS; 35130e1c54c2SIlpo Järvinen else 35140e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; 35150e1c54c2SIlpo Järvinen } 35160e1c54c2SIlpo Järvinen 35170e1c54c2SIlpo Järvinen if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) 351861eb55f4SIlpo Järvinen continue; 351940b215e5SPavel Emelyanov 3520f9616c35SEric Dumazet if (tcp_small_queue_check(sk, skb, 1)) 3521916e6d1aSEric Dumazet break; 3522f9616c35SEric Dumazet 352310d3be56SEric Dumazet if (tcp_retransmit_skb(sk, skb, segs)) 3524916e6d1aSEric Dumazet break; 352524ab6becSYuchung Cheng 3526de1d6578SYuchung Cheng NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb)); 35271da177e4SLinus Torvalds 3528684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 3529a262f0cdSNandita Dukkipati tp->prr_out += tcp_skb_pcount(skb); 3530a262f0cdSNandita Dukkipati 353175c119afSEric Dumazet if (skb == rtx_head && 353257dde7f7SYuchung Cheng icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT) 3533916e6d1aSEric Dumazet rearm_timer = true; 3534916e6d1aSEric Dumazet 3535916e6d1aSEric Dumazet } 3536916e6d1aSEric Dumazet if (rearm_timer) 35373f80e08fSEric Dumazet tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 35383f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, 35398dc242adSEric Dumazet TCP_RTO_MAX); 35401da177e4SLinus Torvalds } 35411da177e4SLinus Torvalds 3542d83769a5SEric Dumazet /* We allow to exceed memory limits for FIN packets to expedite 3543d83769a5SEric Dumazet * connection tear down and (memory) recovery. 3544845704a5SEric Dumazet * Otherwise tcp_send_fin() could be tempted to either delay FIN 3545845704a5SEric Dumazet * or even be forced to close flow without any FIN. 3546a6c5ea4cSEric Dumazet * In general, we want to allow one skb per socket to avoid hangs 3547a6c5ea4cSEric Dumazet * with edge trigger epoll() 3548d83769a5SEric Dumazet */ 3549a6c5ea4cSEric Dumazet void sk_forced_mem_schedule(struct sock *sk, int size) 3550d83769a5SEric Dumazet { 3551c4ee1185SEric Dumazet int delta, amt; 3552d83769a5SEric Dumazet 3553c4ee1185SEric Dumazet delta = size - sk->sk_forward_alloc; 3554c4ee1185SEric Dumazet if (delta <= 0) 3555d83769a5SEric Dumazet return; 3556c4ee1185SEric Dumazet amt = sk_mem_pages(delta); 35575e6300e7SEric Dumazet sk_forward_alloc_add(sk, amt << PAGE_SHIFT); 3558e805605cSJohannes Weiner sk_memory_allocated_add(sk, amt); 3559e805605cSJohannes Weiner 3560baac50bbSJohannes Weiner if (mem_cgroup_sockets_enabled && sk->sk_memcg) 35614b1327beSWei Wang mem_cgroup_charge_skmem(sk->sk_memcg, amt, 35624b1327beSWei Wang gfp_memcg_charge() | __GFP_NOFAIL); 3563d83769a5SEric Dumazet } 3564d83769a5SEric Dumazet 3565845704a5SEric Dumazet /* Send a FIN. The caller locks the socket for us. 3566845704a5SEric Dumazet * We should try to send a FIN packet really hard, but eventually give up. 35671da177e4SLinus Torvalds */ 35681da177e4SLinus Torvalds void tcp_send_fin(struct sock *sk) 35691da177e4SLinus Torvalds { 3570ee2aabd3SEric Dumazet struct sk_buff *skb, *tskb, *tail = tcp_write_queue_tail(sk); 35711da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 35721da177e4SLinus Torvalds 3573845704a5SEric Dumazet /* Optimization, tack on the FIN if we have one skb in write queue and 3574845704a5SEric Dumazet * this skb was not yet sent, or we are under memory pressure. 3575845704a5SEric Dumazet * Note: in the latter case, FIN packet will be sent after a timeout, 3576845704a5SEric Dumazet * as TCP stack thinks it has already been transmitted. 35771da177e4SLinus Torvalds */ 3578ee2aabd3SEric Dumazet tskb = tail; 357975c119afSEric Dumazet if (!tskb && tcp_under_memory_pressure(sk)) 358075c119afSEric Dumazet tskb = skb_rb_last(&sk->tcp_rtx_queue); 358175c119afSEric Dumazet 358275c119afSEric Dumazet if (tskb) { 3583845704a5SEric Dumazet TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN; 3584845704a5SEric Dumazet TCP_SKB_CB(tskb)->end_seq++; 35851da177e4SLinus Torvalds tp->write_seq++; 3586ee2aabd3SEric Dumazet if (!tail) { 3587845704a5SEric Dumazet /* This means tskb was already sent. 3588845704a5SEric Dumazet * Pretend we included the FIN on previous transmit. 3589845704a5SEric Dumazet * We need to set tp->snd_nxt to the value it would have 3590845704a5SEric Dumazet * if FIN had been sent. This is because retransmit path 3591845704a5SEric Dumazet * does not change tp->snd_nxt. 3592845704a5SEric Dumazet */ 3593e0d694d6SEric Dumazet WRITE_ONCE(tp->snd_nxt, tp->snd_nxt + 1); 3594845704a5SEric Dumazet return; 3595845704a5SEric Dumazet } 35961da177e4SLinus Torvalds } else { 3597845704a5SEric Dumazet skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation); 3598d1edc085SColin Ian King if (unlikely(!skb)) 3599845704a5SEric Dumazet return; 3600d1edc085SColin Ian King 3601e2080072SEric Dumazet INIT_LIST_HEAD(&skb->tcp_tsorted_anchor); 3602d83769a5SEric Dumazet skb_reserve(skb, MAX_TCP_HEADER); 3603a6c5ea4cSEric Dumazet sk_forced_mem_schedule(sk, skb->truesize); 36041da177e4SLinus Torvalds /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 3605e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tp->write_seq, 3606a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_FIN); 36071da177e4SLinus Torvalds tcp_queue_skb(sk, skb); 36081da177e4SLinus Torvalds } 3609845704a5SEric Dumazet __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF); 36101da177e4SLinus Torvalds } 36111da177e4SLinus Torvalds 36121da177e4SLinus Torvalds /* We get here when a process closes a file descriptor (either due to 36131da177e4SLinus Torvalds * an explicit close() or as a byproduct of exit()'ing) and there 36141da177e4SLinus Torvalds * was unread data in the receive queue. This behavior is recommended 361565bb723cSGerrit Renker * by RFC 2525, section 2.17. -DaveM 36161da177e4SLinus Torvalds */ 3617dd0fc66fSAl Viro void tcp_send_active_reset(struct sock *sk, gfp_t priority) 36181da177e4SLinus Torvalds { 36191da177e4SLinus Torvalds struct sk_buff *skb; 36201da177e4SLinus Torvalds 36217cc2b043SGao Feng TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); 36227cc2b043SGao Feng 36231da177e4SLinus Torvalds /* NOTE: No TCP options attached and we never retransmit this. */ 36241da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, priority); 36251da177e4SLinus Torvalds if (!skb) { 36264e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 36271da177e4SLinus Torvalds return; 36281da177e4SLinus Torvalds } 36291da177e4SLinus Torvalds 36301da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 36311da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 3632e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), 3633a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_RST); 36349a568de4SEric Dumazet tcp_mstamp_refresh(tcp_sk(sk)); 36351da177e4SLinus Torvalds /* Send it off. */ 3636dfb4b9dcSDavid S. Miller if (tcp_transmit_skb(sk, skb, 0, priority)) 36374e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 3638c24b14c4SSong Liu 3639c24b14c4SSong Liu /* skb of trace_tcp_send_reset() keeps the skb that caused RST, 3640c24b14c4SSong Liu * skb here is different to the troublesome skb, so use NULL 3641c24b14c4SSong Liu */ 3642c24b14c4SSong Liu trace_tcp_send_reset(sk, NULL); 36431da177e4SLinus Torvalds } 36441da177e4SLinus Torvalds 364567edfef7SAndi Kleen /* Send a crossed SYN-ACK during socket establishment. 364667edfef7SAndi Kleen * WARNING: This routine must only be called when we have already sent 36471da177e4SLinus Torvalds * a SYN packet that crossed the incoming SYN that caused this routine 36481da177e4SLinus Torvalds * to get called. If this assumption fails then the initial rcv_wnd 36491da177e4SLinus Torvalds * and rcv_wscale values will not be correct. 36501da177e4SLinus Torvalds */ 36511da177e4SLinus Torvalds int tcp_send_synack(struct sock *sk) 36521da177e4SLinus Torvalds { 36531da177e4SLinus Torvalds struct sk_buff *skb; 36541da177e4SLinus Torvalds 365575c119afSEric Dumazet skb = tcp_rtx_queue_head(sk); 365651456b29SIan Morris if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 365775c119afSEric Dumazet pr_err("%s: wrong queue state\n", __func__); 36581da177e4SLinus Torvalds return -EFAULT; 36591da177e4SLinus Torvalds } 36604de075e0SEric Dumazet if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { 36611da177e4SLinus Torvalds if (skb_cloned(skb)) { 3662e2080072SEric Dumazet struct sk_buff *nskb; 3663e2080072SEric Dumazet 3664e2080072SEric Dumazet tcp_skb_tsorted_save(skb) { 3665e2080072SEric Dumazet nskb = skb_copy(skb, GFP_ATOMIC); 3666e2080072SEric Dumazet } tcp_skb_tsorted_restore(skb); 366751456b29SIan Morris if (!nskb) 36681da177e4SLinus Torvalds return -ENOMEM; 3669e2080072SEric Dumazet INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor); 36702bec445fSEric Dumazet tcp_highest_sack_replace(sk, skb, nskb); 367175c119afSEric Dumazet tcp_rtx_queue_unlink_and_free(skb, sk); 3672f4a775d1SEric Dumazet __skb_header_release(nskb); 367375c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb); 3674ab4e846aSEric Dumazet sk_wmem_queued_add(sk, nskb->truesize); 36753ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 36761da177e4SLinus Torvalds skb = nskb; 36771da177e4SLinus Torvalds } 36781da177e4SLinus Torvalds 36794de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; 3680735d3831SFlorian Westphal tcp_ecn_send_synack(sk, skb); 36811da177e4SLinus Torvalds } 3682dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 36831da177e4SLinus Torvalds } 36841da177e4SLinus Torvalds 36854aea39c1SEric Dumazet /** 3686331fca43SMartin KaFai Lau * tcp_make_synack - Allocate one skb and build a SYNACK packet. 3687331fca43SMartin KaFai Lau * @sk: listener socket 3688331fca43SMartin KaFai Lau * @dst: dst entry attached to the SYNACK. It is consumed and caller 3689331fca43SMartin KaFai Lau * should not use it again. 3690331fca43SMartin KaFai Lau * @req: request_sock pointer 3691331fca43SMartin KaFai Lau * @foc: cookie for tcp fast open 3692331fca43SMartin KaFai Lau * @synack_type: Type of synack to prepare 3693331fca43SMartin KaFai Lau * @syn_skb: SYN packet just received. It could be NULL for rtx case. 36944aea39c1SEric Dumazet */ 36955d062de7SEric Dumazet struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst, 3696e6b4d113SWilliam Allen Simpson struct request_sock *req, 3697ca6fb065SEric Dumazet struct tcp_fastopen_cookie *foc, 3698331fca43SMartin KaFai Lau enum tcp_synack_type synack_type, 3699331fca43SMartin KaFai Lau struct sk_buff *syn_skb) 37001da177e4SLinus Torvalds { 37012e6599cbSArnaldo Carvalho de Melo struct inet_request_sock *ireq = inet_rsk(req); 37025d062de7SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 37035d062de7SEric Dumazet struct tcp_out_options opts; 37041e03d32bSDmitry Safonov struct tcp_key key = {}; 37055d062de7SEric Dumazet struct sk_buff *skb; 3706bd0388aeSWilliam Allen Simpson int tcp_header_size; 37075d062de7SEric Dumazet struct tcphdr *th; 3708f5fff5dcSTom Quetchenbach int mss; 3709a842fe14SEric Dumazet u64 now; 37101da177e4SLinus Torvalds 3711ca6fb065SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 37124aea39c1SEric Dumazet if (unlikely(!skb)) { 37134aea39c1SEric Dumazet dst_release(dst); 37141da177e4SLinus Torvalds return NULL; 37154aea39c1SEric Dumazet } 37161da177e4SLinus Torvalds /* Reserve space for headers. */ 37171da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 37181da177e4SLinus Torvalds 3719b3d05147SEric Dumazet switch (synack_type) { 3720b3d05147SEric Dumazet case TCP_SYNACK_NORMAL: 37219e17f8a4SEric Dumazet skb_set_owner_w(skb, req_to_sk(req)); 3722b3d05147SEric Dumazet break; 3723b3d05147SEric Dumazet case TCP_SYNACK_COOKIE: 3724b3d05147SEric Dumazet /* Under synflood, we do not attach skb to a socket, 3725b3d05147SEric Dumazet * to avoid false sharing. 3726b3d05147SEric Dumazet */ 3727b3d05147SEric Dumazet break; 3728b3d05147SEric Dumazet case TCP_SYNACK_FASTOPEN: 3729ca6fb065SEric Dumazet /* sk is a const pointer, because we want to express multiple 3730ca6fb065SEric Dumazet * cpu might call us concurrently. 3731ca6fb065SEric Dumazet * sk->sk_wmem_alloc in an atomic, we can promote to rw. 3732ca6fb065SEric Dumazet */ 3733ca6fb065SEric Dumazet skb_set_owner_w(skb, (struct sock *)sk); 3734b3d05147SEric Dumazet break; 3735ca6fb065SEric Dumazet } 37364aea39c1SEric Dumazet skb_dst_set(skb, dst); 37371da177e4SLinus Torvalds 37383541f9e8SEric Dumazet mss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 3739f5fff5dcSTom Quetchenbach 374033ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 3741a842fe14SEric Dumazet now = tcp_clock_ns(); 37428b5f12d0SFlorian Westphal #ifdef CONFIG_SYN_COOKIES 3743f8ace8d9SFlorian Westphal if (unlikely(synack_type == TCP_SYNACK_COOKIE && ireq->tstamp_ok)) 3744a1ac9c8aSMartin KaFai Lau skb_set_delivery_time(skb, cookie_init_timestamp(req, now), 3745a1ac9c8aSMartin KaFai Lau true); 37468b5f12d0SFlorian Westphal else 37478b5f12d0SFlorian Westphal #endif 37489e450c1eSYuchung Cheng { 3749a1ac9c8aSMartin KaFai Lau skb_set_delivery_time(skb, now, true); 37509e450c1eSYuchung Cheng if (!tcp_rsk(req)->snt_synack) /* Timestamp first SYNACK */ 37519e450c1eSYuchung Cheng tcp_rsk(req)->snt_synack = tcp_skb_timestamp_us(skb); 37529e450c1eSYuchung Cheng } 375380f03e27SEric Dumazet 37549427c6aaSDmitry Safonov #if defined(CONFIG_TCP_MD5SIG) || defined(CONFIG_TCP_AO) 375580f03e27SEric Dumazet rcu_read_lock(); 37569427c6aaSDmitry Safonov #endif 37579427c6aaSDmitry Safonov if (tcp_rsk_used_ao(req)) { 37589427c6aaSDmitry Safonov #ifdef CONFIG_TCP_AO 37599427c6aaSDmitry Safonov struct tcp_ao_key *ao_key = NULL; 37609427c6aaSDmitry Safonov u8 keyid = tcp_rsk(req)->ao_keyid; 37619427c6aaSDmitry Safonov 37629427c6aaSDmitry Safonov ao_key = tcp_sk(sk)->af_specific->ao_lookup(sk, req_to_sk(req), 37639427c6aaSDmitry Safonov keyid, -1); 37649427c6aaSDmitry Safonov /* If there is no matching key - avoid sending anything, 37659427c6aaSDmitry Safonov * especially usigned segments. It could try harder and lookup 37669427c6aaSDmitry Safonov * for another peer-matching key, but the peer has requested 37679427c6aaSDmitry Safonov * ao_keyid (RFC5925 RNextKeyID), so let's keep it simple here. 37689427c6aaSDmitry Safonov */ 37699396c4eeSDmitry Safonov if (unlikely(!ao_key)) { 37709427c6aaSDmitry Safonov rcu_read_unlock(); 37719427c6aaSDmitry Safonov kfree_skb(skb); 37729396c4eeSDmitry Safonov net_warn_ratelimited("TCP-AO: the keyid %u from SYN packet is not present - not sending SYNACK\n", 37739396c4eeSDmitry Safonov keyid); 37749427c6aaSDmitry Safonov return NULL; 37759427c6aaSDmitry Safonov } 37769427c6aaSDmitry Safonov key.ao_key = ao_key; 37779427c6aaSDmitry Safonov key.type = TCP_KEY_AO; 37789427c6aaSDmitry Safonov #endif 37799427c6aaSDmitry Safonov } else { 37809427c6aaSDmitry Safonov #ifdef CONFIG_TCP_MD5SIG 37819427c6aaSDmitry Safonov key.md5_key = tcp_rsk(req)->af_specific->req_md5_lookup(sk, 37829427c6aaSDmitry Safonov req_to_sk(req)); 37839427c6aaSDmitry Safonov if (key.md5_key) 37841e03d32bSDmitry Safonov key.type = TCP_KEY_MD5; 378580f03e27SEric Dumazet #endif 37869427c6aaSDmitry Safonov } 37875e526552SEric Dumazet skb_set_hash(skb, READ_ONCE(tcp_rsk(req)->txhash), PKT_HASH_TYPE_L4); 3788331fca43SMartin KaFai Lau /* bpf program will be interested in the tcp_flags */ 3789331fca43SMartin KaFai Lau TCP_SKB_CB(skb)->tcp_flags = TCPHDR_SYN | TCPHDR_ACK; 37909427c6aaSDmitry Safonov tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, 37919427c6aaSDmitry Safonov &key, foc, synack_type, syn_skb) 37929427c6aaSDmitry Safonov + sizeof(*th); 379333ad798cSAdam Langley 3794aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 3795aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 37961da177e4SLinus Torvalds 3797ea1627c2SEric Dumazet th = (struct tcphdr *)skb->data; 37981da177e4SLinus Torvalds memset(th, 0, sizeof(struct tcphdr)); 37991da177e4SLinus Torvalds th->syn = 1; 38001da177e4SLinus Torvalds th->ack = 1; 38016ac705b1SEric Dumazet tcp_ecn_make_synack(req, th); 3802b44084c2SEric Dumazet th->source = htons(ireq->ir_num); 3803634fb979SEric Dumazet th->dest = ireq->ir_rmt_port; 3804e05a90ecSJamal Hadi Salim skb->mark = ireq->ir_mark; 38053b117750SEric Dumazet skb->ip_summed = CHECKSUM_PARTIAL; 38063b117750SEric Dumazet th->seq = htonl(tcp_rsk(req)->snt_isn); 38078336886fSJerry Chu /* XXX data is queued and acked as is. No buffer/window check */ 38088336886fSJerry Chu th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt); 38091da177e4SLinus Torvalds 38101da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 3811ed53d0abSEric Dumazet th->window = htons(min(req->rsk_rcv_wnd, 65535U)); 38129427c6aaSDmitry Safonov tcp_options_write(th, NULL, tcp_rsk(req), &opts, &key); 38131da177e4SLinus Torvalds th->doff = (tcp_header_size >> 2); 3814bced3f7dSBreno Leitao TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS); 3815cfb6eeb4SYOSHIFUJI Hideaki 3816cfb6eeb4SYOSHIFUJI Hideaki /* Okay, we have all we need - do the md5 hash if needed */ 38179427c6aaSDmitry Safonov if (tcp_key_is_md5(&key)) { 38189427c6aaSDmitry Safonov #ifdef CONFIG_TCP_MD5SIG 3819bd0388aeSWilliam Allen Simpson tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, 38209427c6aaSDmitry Safonov key.md5_key, req_to_sk(req), skb); 38219427c6aaSDmitry Safonov #endif 38229427c6aaSDmitry Safonov } else if (tcp_key_is_ao(&key)) { 38239427c6aaSDmitry Safonov #ifdef CONFIG_TCP_AO 38249427c6aaSDmitry Safonov tcp_rsk(req)->af_specific->ao_synack_hash(opts.hash_location, 38259427c6aaSDmitry Safonov key.ao_key, req, skb, 38269427c6aaSDmitry Safonov opts.hash_location - (u8 *)th, 0); 38279427c6aaSDmitry Safonov #endif 38289427c6aaSDmitry Safonov } 38299427c6aaSDmitry Safonov #if defined(CONFIG_TCP_MD5SIG) || defined(CONFIG_TCP_AO) 383080f03e27SEric Dumazet rcu_read_unlock(); 3831cfb6eeb4SYOSHIFUJI Hideaki #endif 3832cfb6eeb4SYOSHIFUJI Hideaki 3833331fca43SMartin KaFai Lau bpf_skops_write_hdr_opt((struct sock *)sk, skb, req, syn_skb, 3834331fca43SMartin KaFai Lau synack_type, &opts); 3835331fca43SMartin KaFai Lau 3836a1ac9c8aSMartin KaFai Lau skb_set_delivery_time(skb, now, true); 3837a842fe14SEric Dumazet tcp_add_tx_delay(skb, tp); 3838a842fe14SEric Dumazet 38391da177e4SLinus Torvalds return skb; 38401da177e4SLinus Torvalds } 38414bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_make_synack); 38421da177e4SLinus Torvalds 384381164413SDaniel Borkmann static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst) 384481164413SDaniel Borkmann { 384581164413SDaniel Borkmann struct inet_connection_sock *icsk = inet_csk(sk); 384681164413SDaniel Borkmann const struct tcp_congestion_ops *ca; 384781164413SDaniel Borkmann u32 ca_key = dst_metric(dst, RTAX_CC_ALGO); 384881164413SDaniel Borkmann 384981164413SDaniel Borkmann if (ca_key == TCP_CA_UNSPEC) 385081164413SDaniel Borkmann return; 385181164413SDaniel Borkmann 385281164413SDaniel Borkmann rcu_read_lock(); 385381164413SDaniel Borkmann ca = tcp_ca_find_key(ca_key); 38540baf26b0SMartin KaFai Lau if (likely(ca && bpf_try_module_get(ca, ca->owner))) { 38550baf26b0SMartin KaFai Lau bpf_module_put(icsk->icsk_ca_ops, icsk->icsk_ca_ops->owner); 385681164413SDaniel Borkmann icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst); 385781164413SDaniel Borkmann icsk->icsk_ca_ops = ca; 385881164413SDaniel Borkmann } 385981164413SDaniel Borkmann rcu_read_unlock(); 386081164413SDaniel Borkmann } 386181164413SDaniel Borkmann 386267edfef7SAndi Kleen /* Do all connect socket setups that can be done AF independent. */ 3863f7e56a76Sstephen hemminger static void tcp_connect_init(struct sock *sk) 38641da177e4SLinus Torvalds { 3865cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 38661da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 38671da177e4SLinus Torvalds __u8 rcv_wscale; 386813d3b1ebSLawrence Brakmo u32 rcv_wnd; 38691da177e4SLinus Torvalds 38701da177e4SLinus Torvalds /* We'll fix this up when we get a response from the other end. 38711da177e4SLinus Torvalds * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 38721da177e4SLinus Torvalds */ 38735d2ed052SEric Dumazet tp->tcp_header_len = sizeof(struct tcphdr); 38743666f666SKuniyuki Iwashima if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_timestamps)) 38755d2ed052SEric Dumazet tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED; 38761da177e4SLinus Torvalds 38777c2ffaf2SDmitry Safonov tcp_ao_connect_init(sk); 38787c2ffaf2SDmitry Safonov 38791da177e4SLinus Torvalds /* If user gave his TCP_MAXSEG, record it to clamp */ 38801da177e4SLinus Torvalds if (tp->rx_opt.user_mss) 38811da177e4SLinus Torvalds tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 38821da177e4SLinus Torvalds tp->max_window = 0; 38835d424d5aSJohn Heffner tcp_mtup_init(sk); 38841da177e4SLinus Torvalds tcp_sync_mss(sk, dst_mtu(dst)); 38851da177e4SLinus Torvalds 388681164413SDaniel Borkmann tcp_ca_dst_init(sk, dst); 388781164413SDaniel Borkmann 38881da177e4SLinus Torvalds if (!tp->window_clamp) 3889f410cbeaSEric Dumazet WRITE_ONCE(tp->window_clamp, dst_metric(dst, RTAX_WINDOW)); 38903541f9e8SEric Dumazet tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 3891f5fff5dcSTom Quetchenbach 38921da177e4SLinus Torvalds tcp_initialize_rcv_mss(sk); 38931da177e4SLinus Torvalds 3894e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 3895e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 3896e88c64f0SHagen Paul Pfeifer (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) 3897f410cbeaSEric Dumazet WRITE_ONCE(tp->window_clamp, tcp_full_space(sk)); 3898e88c64f0SHagen Paul Pfeifer 389913d3b1ebSLawrence Brakmo rcv_wnd = tcp_rwnd_init_bpf(sk); 390013d3b1ebSLawrence Brakmo if (rcv_wnd == 0) 390113d3b1ebSLawrence Brakmo rcv_wnd = dst_metric(dst, RTAX_INITRWND); 390213d3b1ebSLawrence Brakmo 3903ceef9ab6SEric Dumazet tcp_select_initial_window(sk, tcp_full_space(sk), 39041da177e4SLinus Torvalds tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 39051da177e4SLinus Torvalds &tp->rcv_wnd, 39061da177e4SLinus Torvalds &tp->window_clamp, 39073666f666SKuniyuki Iwashima READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_window_scaling), 390831d12926Slaurent chavey &rcv_wscale, 390913d3b1ebSLawrence Brakmo rcv_wnd); 39101da177e4SLinus Torvalds 39111da177e4SLinus Torvalds tp->rx_opt.rcv_wscale = rcv_wscale; 39121da177e4SLinus Torvalds tp->rcv_ssthresh = tp->rcv_wnd; 39131da177e4SLinus Torvalds 3914e13ec3daSEric Dumazet WRITE_ONCE(sk->sk_err, 0); 39151da177e4SLinus Torvalds sock_reset_flag(sk, SOCK_DONE); 39161da177e4SLinus Torvalds tp->snd_wnd = 0; 3917ee7537b6SHantzis Fotis tcp_init_wl(tp, 0); 39187f582b24SEric Dumazet tcp_write_queue_purge(sk); 39191da177e4SLinus Torvalds tp->snd_una = tp->write_seq; 39201da177e4SLinus Torvalds tp->snd_sml = tp->write_seq; 392133f5f57eSIlpo Järvinen tp->snd_up = tp->write_seq; 3922e0d694d6SEric Dumazet WRITE_ONCE(tp->snd_nxt, tp->write_seq); 3923ee995283SPavel Emelyanov 3924ee995283SPavel Emelyanov if (likely(!tp->repair)) 39251da177e4SLinus Torvalds tp->rcv_nxt = 0; 3926c7781a6eSAndrew Vagin else 392770eabf0eSEric Dumazet tp->rcv_tstamp = tcp_jiffies32; 3928ee995283SPavel Emelyanov tp->rcv_wup = tp->rcv_nxt; 39297db48e98SEric Dumazet WRITE_ONCE(tp->copied_seq, tp->rcv_nxt); 39301da177e4SLinus Torvalds 39318550f328SLawrence Brakmo inet_csk(sk)->icsk_rto = tcp_timeout_init(sk); 3932463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_retransmits = 0; 39331da177e4SLinus Torvalds tcp_clear_retrans(tp); 39341da177e4SLinus Torvalds } 39351da177e4SLinus Torvalds 3936783237e8SYuchung Cheng static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) 3937783237e8SYuchung Cheng { 3938783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3939783237e8SYuchung Cheng struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 3940783237e8SYuchung Cheng 3941783237e8SYuchung Cheng tcb->end_seq += skb->len; 3942f4a775d1SEric Dumazet __skb_header_release(skb); 3943ab4e846aSEric Dumazet sk_wmem_queued_add(sk, skb->truesize); 3944783237e8SYuchung Cheng sk_mem_charge(sk, skb->truesize); 39450f317464SEric Dumazet WRITE_ONCE(tp->write_seq, tcb->end_seq); 3946783237e8SYuchung Cheng tp->packets_out += tcp_skb_pcount(skb); 3947783237e8SYuchung Cheng } 3948783237e8SYuchung Cheng 3949783237e8SYuchung Cheng /* Build and send a SYN with data and (cached) Fast Open cookie. However, 3950783237e8SYuchung Cheng * queue a data-only packet after the regular SYN, such that regular SYNs 3951783237e8SYuchung Cheng * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges 3952783237e8SYuchung Cheng * only the SYN sequence, the data are retransmitted in the first ACK. 3953783237e8SYuchung Cheng * If cookie is not cached or other error occurs, falls back to send a 3954783237e8SYuchung Cheng * regular SYN with Fast Open cookie request option. 3955783237e8SYuchung Cheng */ 3956783237e8SYuchung Cheng static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) 3957783237e8SYuchung Cheng { 3958ed0c99dcSJakub Kicinski struct inet_connection_sock *icsk = inet_csk(sk); 3959783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3960783237e8SYuchung Cheng struct tcp_fastopen_request *fo = tp->fastopen_req; 3961fbf93406SEric Dumazet struct page_frag *pfrag = sk_page_frag(sk); 3962355a901eSEric Dumazet struct sk_buff *syn_data; 3963fbf93406SEric Dumazet int space, err = 0; 3964783237e8SYuchung Cheng 396567da22d2SYuchung Cheng tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ 3966065263f4SWei Wang if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie)) 3967783237e8SYuchung Cheng goto fallback; 3968783237e8SYuchung Cheng 3969783237e8SYuchung Cheng /* MSS for SYN-data is based on cached MSS and bounded by PMTU and 3970783237e8SYuchung Cheng * user-MSS. Reserve maximum option space for middleboxes that add 3971783237e8SYuchung Cheng * private TCP options. The cost is reduced data space in SYN :( 3972783237e8SYuchung Cheng */ 39733541f9e8SEric Dumazet tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp); 3974ed0c99dcSJakub Kicinski /* Sync mss_cache after updating the mss_clamp */ 3975ed0c99dcSJakub Kicinski tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 39763541f9e8SEric Dumazet 3977ed0c99dcSJakub Kicinski space = __tcp_mtu_to_mss(sk, icsk->icsk_pmtu_cookie) - 3978783237e8SYuchung Cheng MAX_TCP_OPTION_SPACE; 3979783237e8SYuchung Cheng 3980f5ddcbbbSEric Dumazet space = min_t(size_t, space, fo->size); 3981f5ddcbbbSEric Dumazet 3982fbf93406SEric Dumazet if (space && 3983fbf93406SEric Dumazet !skb_page_frag_refill(min_t(size_t, space, PAGE_SIZE), 3984fbf93406SEric Dumazet pfrag, sk->sk_allocation)) 3985fbf93406SEric Dumazet goto fallback; 39865882efffSEric Dumazet syn_data = tcp_stream_alloc_skb(sk, sk->sk_allocation, false); 3987355a901eSEric Dumazet if (!syn_data) 3988783237e8SYuchung Cheng goto fallback; 3989355a901eSEric Dumazet memcpy(syn_data->cb, syn->cb, sizeof(syn->cb)); 399007e100f9SEric Dumazet if (space) { 3991fbf93406SEric Dumazet space = min_t(size_t, space, pfrag->size - pfrag->offset); 3992fbf93406SEric Dumazet space = tcp_wmem_schedule(sk, space); 3993fbf93406SEric Dumazet } 3994fbf93406SEric Dumazet if (space) { 3995fbf93406SEric Dumazet space = copy_page_from_iter(pfrag->page, pfrag->offset, 3996fbf93406SEric Dumazet space, &fo->data->msg_iter); 3997fbf93406SEric Dumazet if (unlikely(!space)) { 3998ba233b34SEric Dumazet tcp_skb_tsorted_anchor_cleanup(syn_data); 3999355a901eSEric Dumazet kfree_skb(syn_data); 4000783237e8SYuchung Cheng goto fallback; 4001783237e8SYuchung Cheng } 4002fbf93406SEric Dumazet skb_fill_page_desc(syn_data, 0, pfrag->page, 4003fbf93406SEric Dumazet pfrag->offset, space); 4004fbf93406SEric Dumazet page_ref_inc(pfrag->page); 4005fbf93406SEric Dumazet pfrag->offset += space; 4006fbf93406SEric Dumazet skb_len_add(syn_data, space); 4007f859a448SWillem de Bruijn skb_zcopy_set(syn_data, fo->uarg, NULL); 400807e100f9SEric Dumazet } 4009355a901eSEric Dumazet /* No more data pending in inet_wait_for_connect() */ 4010355a901eSEric Dumazet if (space == fo->size) 4011355a901eSEric Dumazet fo->data = NULL; 4012355a901eSEric Dumazet fo->copied = space; 4013783237e8SYuchung Cheng 4014355a901eSEric Dumazet tcp_connect_queue_skb(sk, syn_data); 40150f87230dSFrancis Yan if (syn_data->len) 40160f87230dSFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_BUSY); 4017355a901eSEric Dumazet 4018355a901eSEric Dumazet err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation); 4019355a901eSEric Dumazet 4020a1ac9c8aSMartin KaFai Lau skb_set_delivery_time(syn, syn_data->skb_mstamp_ns, true); 4021355a901eSEric Dumazet 4022355a901eSEric Dumazet /* Now full SYN+DATA was cloned and sent (or not), 4023355a901eSEric Dumazet * remove the SYN from the original skb (syn_data) 4024355a901eSEric Dumazet * we keep in write queue in case of a retransmit, as we 4025355a901eSEric Dumazet * also have the SYN packet (with no data) in the same queue. 4026431a9124SEric Dumazet */ 4027355a901eSEric Dumazet TCP_SKB_CB(syn_data)->seq++; 4028355a901eSEric Dumazet TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH; 4029355a901eSEric Dumazet if (!err) { 403067da22d2SYuchung Cheng tp->syn_data = (fo->copied > 0); 403175c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data); 4032f19c29e3SYuchung Cheng NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT); 4033783237e8SYuchung Cheng goto done; 4034783237e8SYuchung Cheng } 4035783237e8SYuchung Cheng 403675c119afSEric Dumazet /* data was not sent, put it in write_queue */ 403775c119afSEric Dumazet __skb_queue_tail(&sk->sk_write_queue, syn_data); 4038b5b7db8dSEric Dumazet tp->packets_out -= tcp_skb_pcount(syn_data); 4039b5b7db8dSEric Dumazet 4040783237e8SYuchung Cheng fallback: 4041783237e8SYuchung Cheng /* Send a regular SYN with Fast Open cookie request option */ 4042783237e8SYuchung Cheng if (fo->cookie.len > 0) 4043783237e8SYuchung Cheng fo->cookie.len = 0; 4044783237e8SYuchung Cheng err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); 4045783237e8SYuchung Cheng if (err) 4046783237e8SYuchung Cheng tp->syn_fastopen = 0; 4047783237e8SYuchung Cheng done: 4048783237e8SYuchung Cheng fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ 4049783237e8SYuchung Cheng return err; 4050783237e8SYuchung Cheng } 4051783237e8SYuchung Cheng 405267edfef7SAndi Kleen /* Build a SYN and send it off. */ 40531da177e4SLinus Torvalds int tcp_connect(struct sock *sk) 40541da177e4SLinus Torvalds { 40551da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 40561da177e4SLinus Torvalds struct sk_buff *buff; 4057ee586811SEric Paris int err; 40581da177e4SLinus Torvalds 4059de525be2SLawrence Brakmo tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB, 0, NULL); 40608ba60924SEric Dumazet 40610aadc739SDmitry Safonov #if defined(CONFIG_TCP_MD5SIG) && defined(CONFIG_TCP_AO) 40620aadc739SDmitry Safonov /* Has to be checked late, after setting daddr/saddr/ops. 40630aadc739SDmitry Safonov * Return error if the peer has both a md5 and a tcp-ao key 40640aadc739SDmitry Safonov * configured as this is ambiguous. 40650aadc739SDmitry Safonov */ 40660aadc739SDmitry Safonov if (unlikely(rcu_dereference_protected(tp->md5sig_info, 40670aadc739SDmitry Safonov lockdep_sock_is_held(sk)))) { 40680aadc739SDmitry Safonov bool needs_ao = !!tp->af_specific->ao_lookup(sk, sk, -1, -1); 40690aadc739SDmitry Safonov bool needs_md5 = !!tp->af_specific->md5_lookup(sk, sk); 40700aadc739SDmitry Safonov struct tcp_ao_info *ao_info; 40710aadc739SDmitry Safonov 40720aadc739SDmitry Safonov ao_info = rcu_dereference_check(tp->ao_info, 40730aadc739SDmitry Safonov lockdep_sock_is_held(sk)); 40740aadc739SDmitry Safonov if (ao_info) { 40750aadc739SDmitry Safonov /* This is an extra check: tcp_ao_required() in 40760aadc739SDmitry Safonov * tcp_v{4,6}_parse_md5_keys() should prevent adding 40770aadc739SDmitry Safonov * md5 keys on ao_required socket. 40780aadc739SDmitry Safonov */ 40790aadc739SDmitry Safonov needs_ao |= ao_info->ao_required; 40800aadc739SDmitry Safonov WARN_ON_ONCE(ao_info->ao_required && needs_md5); 40810aadc739SDmitry Safonov } 40820aadc739SDmitry Safonov if (needs_md5 && needs_ao) 40830aadc739SDmitry Safonov return -EKEYREJECTED; 40840aadc739SDmitry Safonov 40850aadc739SDmitry Safonov /* If we have a matching md5 key and no matching tcp-ao key 40860aadc739SDmitry Safonov * then free up ao_info if allocated. 40870aadc739SDmitry Safonov */ 40880aadc739SDmitry Safonov if (needs_md5) { 4089decde258SDmitry Safonov tcp_ao_destroy_sock(sk, false); 40900aadc739SDmitry Safonov } else if (needs_ao) { 40910aadc739SDmitry Safonov tcp_clear_md5_list(sk); 40920aadc739SDmitry Safonov kfree(rcu_replace_pointer(tp->md5sig_info, NULL, 40930aadc739SDmitry Safonov lockdep_sock_is_held(sk))); 40940aadc739SDmitry Safonov } 40950aadc739SDmitry Safonov } 40960aadc739SDmitry Safonov #endif 40970aadc739SDmitry Safonov #ifdef CONFIG_TCP_AO 40980aadc739SDmitry Safonov if (unlikely(rcu_dereference_protected(tp->ao_info, 40990aadc739SDmitry Safonov lockdep_sock_is_held(sk)))) { 41000aadc739SDmitry Safonov /* Don't allow connecting if ao is configured but no 41010aadc739SDmitry Safonov * matching key is found. 41020aadc739SDmitry Safonov */ 41030aadc739SDmitry Safonov if (!tp->af_specific->ao_lookup(sk, sk, -1, -1)) 41040aadc739SDmitry Safonov return -EKEYREJECTED; 41050aadc739SDmitry Safonov } 41060aadc739SDmitry Safonov #endif 41070aadc739SDmitry Safonov 41088ba60924SEric Dumazet if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 41098ba60924SEric Dumazet return -EHOSTUNREACH; /* Routing failure or similar. */ 41108ba60924SEric Dumazet 41111da177e4SLinus Torvalds tcp_connect_init(sk); 41121da177e4SLinus Torvalds 41132b916477SAndrey Vagin if (unlikely(tp->repair)) { 41142b916477SAndrey Vagin tcp_finish_connect(sk, NULL); 41152b916477SAndrey Vagin return 0; 41162b916477SAndrey Vagin } 41172b916477SAndrey Vagin 41185882efffSEric Dumazet buff = tcp_stream_alloc_skb(sk, sk->sk_allocation, true); 4119355a901eSEric Dumazet if (unlikely(!buff)) 41201da177e4SLinus Torvalds return -ENOBUFS; 41211da177e4SLinus Torvalds 4122a3433f35SChangli Gao tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); 41239a568de4SEric Dumazet tcp_mstamp_refresh(tp); 41249d0c00f5SEric Dumazet tp->retrans_stamp = tcp_time_stamp_ts(tp); 4125783237e8SYuchung Cheng tcp_connect_queue_skb(sk, buff); 4126735d3831SFlorian Westphal tcp_ecn_send_syn(sk, buff); 412775c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, buff); 41281da177e4SLinus Torvalds 4129783237e8SYuchung Cheng /* Send off SYN; include data in Fast Open. */ 4130783237e8SYuchung Cheng err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : 4131783237e8SYuchung Cheng tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); 4132ee586811SEric Paris if (err == -ECONNREFUSED) 4133ee586811SEric Paris return err; 4134bd37a088SWei Yongjun 4135bd37a088SWei Yongjun /* We change tp->snd_nxt after the tcp_transmit_skb() call 4136bd37a088SWei Yongjun * in order to make this packet get counted in tcpOutSegs. 4137bd37a088SWei Yongjun */ 4138e0d694d6SEric Dumazet WRITE_ONCE(tp->snd_nxt, tp->write_seq); 4139bd37a088SWei Yongjun tp->pushed_seq = tp->write_seq; 4140b5b7db8dSEric Dumazet buff = tcp_send_head(sk); 4141b5b7db8dSEric Dumazet if (unlikely(buff)) { 4142e0d694d6SEric Dumazet WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(buff)->seq); 4143b5b7db8dSEric Dumazet tp->pushed_seq = TCP_SKB_CB(buff)->seq; 4144b5b7db8dSEric Dumazet } 414581cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); 41461da177e4SLinus Torvalds 41471da177e4SLinus Torvalds /* Timer for repeating the SYN until an answer. */ 41483f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 41493f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 41501da177e4SLinus Torvalds return 0; 41511da177e4SLinus Torvalds } 41524bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_connect); 41531da177e4SLinus Torvalds 4154bbf80d71SEric Dumazet u32 tcp_delack_max(const struct sock *sk) 4155bbf80d71SEric Dumazet { 4156bbf80d71SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 4157bbf80d71SEric Dumazet u32 delack_max = inet_csk(sk)->icsk_delack_max; 4158bbf80d71SEric Dumazet 4159bbf80d71SEric Dumazet if (dst && dst_metric_locked(dst, RTAX_RTO_MIN)) { 4160bbf80d71SEric Dumazet u32 rto_min = dst_metric_rtt(dst, RTAX_RTO_MIN); 4161bbf80d71SEric Dumazet u32 delack_from_rto_min = max_t(int, 1, rto_min - 1); 4162bbf80d71SEric Dumazet 4163bbf80d71SEric Dumazet delack_max = min_t(u32, delack_max, delack_from_rto_min); 4164bbf80d71SEric Dumazet } 4165bbf80d71SEric Dumazet return delack_max; 4166bbf80d71SEric Dumazet } 4167bbf80d71SEric Dumazet 41681da177e4SLinus Torvalds /* Send out a delayed ack, the caller does the policy checking 41691da177e4SLinus Torvalds * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 41701da177e4SLinus Torvalds * for details. 41711da177e4SLinus Torvalds */ 41721da177e4SLinus Torvalds void tcp_send_delayed_ack(struct sock *sk) 41731da177e4SLinus Torvalds { 4174463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 4175463c84b9SArnaldo Carvalho de Melo int ato = icsk->icsk_ack.ato; 41761da177e4SLinus Torvalds unsigned long timeout; 41771da177e4SLinus Torvalds 41781da177e4SLinus Torvalds if (ato > TCP_DELACK_MIN) { 4179463c84b9SArnaldo Carvalho de Melo const struct tcp_sock *tp = tcp_sk(sk); 41801da177e4SLinus Torvalds int max_ato = HZ / 2; 41811da177e4SLinus Torvalds 418231954cd8SWei Wang if (inet_csk_in_pingpong_mode(sk) || 4183056834d9SIlpo Järvinen (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 41841da177e4SLinus Torvalds max_ato = TCP_DELACK_MAX; 41851da177e4SLinus Torvalds 41861da177e4SLinus Torvalds /* Slow path, intersegment interval is "high". */ 41871da177e4SLinus Torvalds 41881da177e4SLinus Torvalds /* If some rtt estimate is known, use it to bound delayed ack. 4189463c84b9SArnaldo Carvalho de Melo * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 41901da177e4SLinus Torvalds * directly. 41911da177e4SLinus Torvalds */ 4192740b0f18SEric Dumazet if (tp->srtt_us) { 4193740b0f18SEric Dumazet int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3), 4194740b0f18SEric Dumazet TCP_DELACK_MIN); 41951da177e4SLinus Torvalds 41961da177e4SLinus Torvalds if (rtt < max_ato) 41971da177e4SLinus Torvalds max_ato = rtt; 41981da177e4SLinus Torvalds } 41991da177e4SLinus Torvalds 42001da177e4SLinus Torvalds ato = min(ato, max_ato); 42011da177e4SLinus Torvalds } 42021da177e4SLinus Torvalds 4203bbf80d71SEric Dumazet ato = min_t(u32, ato, tcp_delack_max(sk)); 42042b8ee4f0SMartin KaFai Lau 42051da177e4SLinus Torvalds /* Stay within the limit we were given */ 42061da177e4SLinus Torvalds timeout = jiffies + ato; 42071da177e4SLinus Torvalds 42081da177e4SLinus Torvalds /* Use new timeout only if there wasn't a older one earlier. */ 4209463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 4210b6b6d653SEric Dumazet /* If delack timer is about to expire, send ACK now. */ 4211b6b6d653SEric Dumazet if (time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 42121da177e4SLinus Torvalds tcp_send_ack(sk); 42131da177e4SLinus Torvalds return; 42141da177e4SLinus Torvalds } 42151da177e4SLinus Torvalds 4216463c84b9SArnaldo Carvalho de Melo if (!time_before(timeout, icsk->icsk_ack.timeout)) 4217463c84b9SArnaldo Carvalho de Melo timeout = icsk->icsk_ack.timeout; 42181da177e4SLinus Torvalds } 4219463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 4220463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.timeout = timeout; 4221463c84b9SArnaldo Carvalho de Melo sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 42221da177e4SLinus Torvalds } 42231da177e4SLinus Torvalds 42241da177e4SLinus Torvalds /* This routine sends an ack and also updates the window. */ 42252987babbSYuchung Cheng void __tcp_send_ack(struct sock *sk, u32 rcv_nxt) 42261da177e4SLinus Torvalds { 42271da177e4SLinus Torvalds struct sk_buff *buff; 42281da177e4SLinus Torvalds 4229058dc334SIlpo Järvinen /* If we have been reset, we may not send again. */ 4230058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 4231058dc334SIlpo Järvinen return; 4232058dc334SIlpo Järvinen 42331da177e4SLinus Torvalds /* We are not putting this on the write queue, so 42341da177e4SLinus Torvalds * tcp_transmit_skb() will set the ownership to this 42351da177e4SLinus Torvalds * sock. 42361da177e4SLinus Torvalds */ 42377450aaf6SEric Dumazet buff = alloc_skb(MAX_TCP_HEADER, 42387450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 42397450aaf6SEric Dumazet if (unlikely(!buff)) { 4240a37c2134SEric Dumazet struct inet_connection_sock *icsk = inet_csk(sk); 4241a37c2134SEric Dumazet unsigned long delay; 4242a37c2134SEric Dumazet 4243a37c2134SEric Dumazet delay = TCP_DELACK_MAX << icsk->icsk_ack.retry; 4244a37c2134SEric Dumazet if (delay < TCP_RTO_MAX) 4245a37c2134SEric Dumazet icsk->icsk_ack.retry++; 4246463c84b9SArnaldo Carvalho de Melo inet_csk_schedule_ack(sk); 4247a37c2134SEric Dumazet icsk->icsk_ack.ato = TCP_ATO_MIN; 4248a37c2134SEric Dumazet inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, delay, TCP_RTO_MAX); 42491da177e4SLinus Torvalds return; 42501da177e4SLinus Torvalds } 42511da177e4SLinus Torvalds 42521da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 42531da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 4254a3433f35SChangli Gao tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); 42551da177e4SLinus Torvalds 425698781965SEric Dumazet /* We do not want pure acks influencing TCP Small Queues or fq/pacing 425798781965SEric Dumazet * too much. 425898781965SEric Dumazet * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784 425998781965SEric Dumazet */ 426098781965SEric Dumazet skb_set_tcp_pure_ack(buff); 426198781965SEric Dumazet 42621da177e4SLinus Torvalds /* Send it off, this clears delayed acks for us. */ 42632987babbSYuchung Cheng __tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0, rcv_nxt); 42641da177e4SLinus Torvalds } 426527cde44aSYuchung Cheng EXPORT_SYMBOL_GPL(__tcp_send_ack); 42662987babbSYuchung Cheng 42672987babbSYuchung Cheng void tcp_send_ack(struct sock *sk) 42682987babbSYuchung Cheng { 42692987babbSYuchung Cheng __tcp_send_ack(sk, tcp_sk(sk)->rcv_nxt); 42701da177e4SLinus Torvalds } 42711da177e4SLinus Torvalds 42721da177e4SLinus Torvalds /* This routine sends a packet with an out of date sequence 42731da177e4SLinus Torvalds * number. It assumes the other end will try to ack it. 42741da177e4SLinus Torvalds * 42751da177e4SLinus Torvalds * Question: what should we make while urgent mode? 42761da177e4SLinus Torvalds * 4.4BSD forces sending single byte of data. We cannot send 42771da177e4SLinus Torvalds * out of window data, because we have SND.NXT==SND.MAX... 42781da177e4SLinus Torvalds * 42791da177e4SLinus Torvalds * Current solution: to send TWO zero-length segments in urgent mode: 42801da177e4SLinus Torvalds * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 42811da177e4SLinus Torvalds * out-of-date with SND.UNA-1 to probe window. 42821da177e4SLinus Torvalds */ 4283e520af48SEric Dumazet static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib) 42841da177e4SLinus Torvalds { 42851da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 42861da177e4SLinus Torvalds struct sk_buff *skb; 42871da177e4SLinus Torvalds 42881da177e4SLinus Torvalds /* We don't queue it, tcp_transmit_skb() sets ownership. */ 42897450aaf6SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, 42907450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 429151456b29SIan Morris if (!skb) 42921da177e4SLinus Torvalds return -1; 42931da177e4SLinus Torvalds 42941da177e4SLinus Torvalds /* Reserve space for headers and set control bits. */ 42951da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 42961da177e4SLinus Torvalds /* Use a previous sequence. This should cause the other 42971da177e4SLinus Torvalds * end to send an ack. Don't queue or clone SKB, just 42981da177e4SLinus Torvalds * send it. 42991da177e4SLinus Torvalds */ 4300a3433f35SChangli Gao tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); 4301e2e8009fSRenato Westphal NET_INC_STATS(sock_net(sk), mib); 43027450aaf6SEric Dumazet return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0); 43031da177e4SLinus Torvalds } 43041da177e4SLinus Torvalds 4305385e2070SEric Dumazet /* Called from setsockopt( ... TCP_REPAIR ) */ 4306ee995283SPavel Emelyanov void tcp_send_window_probe(struct sock *sk) 4307ee995283SPavel Emelyanov { 4308ee995283SPavel Emelyanov if (sk->sk_state == TCP_ESTABLISHED) { 4309ee995283SPavel Emelyanov tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; 43109a568de4SEric Dumazet tcp_mstamp_refresh(tcp_sk(sk)); 4311e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE); 4312ee995283SPavel Emelyanov } 4313ee995283SPavel Emelyanov } 4314ee995283SPavel Emelyanov 431567edfef7SAndi Kleen /* Initiate keepalive or window probe from timer. */ 4316e520af48SEric Dumazet int tcp_write_wakeup(struct sock *sk, int mib) 43171da177e4SLinus Torvalds { 43181da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 43191da177e4SLinus Torvalds struct sk_buff *skb; 43201da177e4SLinus Torvalds 4321058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 4322058dc334SIlpo Järvinen return -1; 4323058dc334SIlpo Järvinen 432400db4124SIan Morris skb = tcp_send_head(sk); 432500db4124SIan Morris if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { 43261da177e4SLinus Torvalds int err; 43270c54b85fSIlpo Järvinen unsigned int mss = tcp_current_mss(sk); 432890840defSIlpo Järvinen unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 43291da177e4SLinus Torvalds 43301da177e4SLinus Torvalds if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 43311da177e4SLinus Torvalds tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 43321da177e4SLinus Torvalds 43331da177e4SLinus Torvalds /* We are probing the opening of a window 43341da177e4SLinus Torvalds * but the window size is != 0 43351da177e4SLinus Torvalds * must have been a result SWS avoidance ( sender ) 43361da177e4SLinus Torvalds */ 43371da177e4SLinus Torvalds if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 43381da177e4SLinus Torvalds skb->len > mss) { 43391da177e4SLinus Torvalds seg_size = min(seg_size, mss); 43404de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 434175c119afSEric Dumazet if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE, 434275c119afSEric Dumazet skb, seg_size, mss, GFP_ATOMIC)) 43431da177e4SLinus Torvalds return -1; 43441da177e4SLinus Torvalds } else if (!tcp_skb_pcount(skb)) 43455bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss); 43461da177e4SLinus Torvalds 43474de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 4348dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 434966f5fe62SIlpo Järvinen if (!err) 435066f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 43511da177e4SLinus Torvalds return err; 43521da177e4SLinus Torvalds } else { 435333f5f57eSIlpo Järvinen if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) 4354e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 1, mib); 4355e520af48SEric Dumazet return tcp_xmit_probe_skb(sk, 0, mib); 43561da177e4SLinus Torvalds } 43571da177e4SLinus Torvalds } 43581da177e4SLinus Torvalds 43591da177e4SLinus Torvalds /* A window probe timeout has occurred. If window is not closed send 43601da177e4SLinus Torvalds * a partial packet else a zero probe. 43611da177e4SLinus Torvalds */ 43621da177e4SLinus Torvalds void tcp_send_probe0(struct sock *sk) 43631da177e4SLinus Torvalds { 4364463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 43651da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 4366c6214a97SNikolay Borisov struct net *net = sock_net(sk); 4367c1d5674fSYuchung Cheng unsigned long timeout; 43681da177e4SLinus Torvalds int err; 43691da177e4SLinus Torvalds 4370e520af48SEric Dumazet err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE); 43711da177e4SLinus Torvalds 437275c119afSEric Dumazet if (tp->packets_out || tcp_write_queue_empty(sk)) { 43731da177e4SLinus Torvalds /* Cancel probe timer, if it is not required. */ 43746687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 0; 4375463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff = 0; 43769d9b1ee0SEnke Chen icsk->icsk_probes_tstamp = 0; 43771da177e4SLinus Torvalds return; 43781da177e4SLinus Torvalds } 43791da177e4SLinus Torvalds 4380c1d5674fSYuchung Cheng icsk->icsk_probes_out++; 43811da177e4SLinus Torvalds if (err <= 0) { 438239e24435SKuniyuki Iwashima if (icsk->icsk_backoff < READ_ONCE(net->ipv4.sysctl_tcp_retries2)) 4383463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff++; 4384c1d5674fSYuchung Cheng timeout = tcp_probe0_when(sk, TCP_RTO_MAX); 43851da177e4SLinus Torvalds } else { 43861da177e4SLinus Torvalds /* If packet was not sent due to local congestion, 4387c1d5674fSYuchung Cheng * Let senders fight for local resources conservatively. 43881da177e4SLinus Torvalds */ 4389c1d5674fSYuchung Cheng timeout = TCP_RESOURCE_PROBE_INTERVAL; 43901da177e4SLinus Torvalds } 4391344db93aSEnke Chen 4392344db93aSEnke Chen timeout = tcp_clamp_probe0_to_user_timeout(sk, timeout); 43938dc242adSEric Dumazet tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, timeout, TCP_RTO_MAX); 43941da177e4SLinus Torvalds } 43955db92c99SOctavian Purdila 4396ea3bea3aSEric Dumazet int tcp_rtx_synack(const struct sock *sk, struct request_sock *req) 43975db92c99SOctavian Purdila { 43985db92c99SOctavian Purdila const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific; 43995db92c99SOctavian Purdila struct flowi fl; 44005db92c99SOctavian Purdila int res; 44015db92c99SOctavian Purdila 4402cb6cd2ceSAkhmat Karakotov /* Paired with WRITE_ONCE() in sock_setsockopt() */ 4403cb6cd2ceSAkhmat Karakotov if (READ_ONCE(sk->sk_txrehash) == SOCK_TXREHASH_ENABLED) 44045e526552SEric Dumazet WRITE_ONCE(tcp_rsk(req)->txhash, net_tx_rndhash()); 4405331fca43SMartin KaFai Lau res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL, 4406331fca43SMartin KaFai Lau NULL); 44075db92c99SOctavian Purdila if (!res) { 44080a375c82SEric Dumazet TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); 44090a375c82SEric Dumazet NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 4410e9d9da91SEric Dumazet if (unlikely(tcp_passive_fastopen(sk))) { 4411e9d9da91SEric Dumazet /* sk has const attribute because listeners are lockless. 4412e9d9da91SEric Dumazet * However in this case, we are dealing with a passive fastopen 4413e9d9da91SEric Dumazet * socket thus we can change total_retrans value. 4414e9d9da91SEric Dumazet */ 4415e9d9da91SEric Dumazet tcp_sk_rw(sk)->total_retrans++; 4416e9d9da91SEric Dumazet } 4417cf34ce3dSSong Liu trace_tcp_retransmit_synack(sk, req); 44185db92c99SOctavian Purdila } 44195db92c99SOctavian Purdila return res; 44205db92c99SOctavian Purdila } 44215db92c99SOctavian Purdila EXPORT_SYMBOL(tcp_rtx_synack); 4422