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> 47*f6d827b1SMina 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. */ 15055bbb432cSEric Dumazet static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now) 1506f6302d1dSDavid S. Miller { 15074a64fd6cSEric Dumazet if (skb->len <= mss_now) { 1508f6302d1dSDavid S. Miller /* Avoid the costly divide in the normal 1509f6302d1dSDavid S. Miller * non-TSO case. 1510f6302d1dSDavid S. Miller */ 1511cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, 1); 1512f69ad292SEric Dumazet TCP_SKB_CB(skb)->tcp_gso_size = 0; 1513f6302d1dSDavid S. Miller } else { 1514cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now)); 1515f69ad292SEric Dumazet TCP_SKB_CB(skb)->tcp_gso_size = mss_now; 15161da177e4SLinus Torvalds } 15171da177e4SLinus Torvalds } 15181da177e4SLinus Torvalds 1519797108d1SIlpo Järvinen /* Pcount in the middle of the write queue got changed, we need to do various 1520797108d1SIlpo Järvinen * tweaks to fix counters 1521797108d1SIlpo Järvinen */ 1522cf533ea5SEric Dumazet static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) 1523797108d1SIlpo Järvinen { 1524797108d1SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1525797108d1SIlpo Järvinen 1526797108d1SIlpo Järvinen tp->packets_out -= decr; 1527797108d1SIlpo Järvinen 1528797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 1529797108d1SIlpo Järvinen tp->sacked_out -= decr; 1530797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 1531797108d1SIlpo Järvinen tp->retrans_out -= decr; 1532797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 1533797108d1SIlpo Järvinen tp->lost_out -= decr; 1534797108d1SIlpo Järvinen 1535797108d1SIlpo Järvinen /* Reno case is special. Sigh... */ 1536797108d1SIlpo Järvinen if (tcp_is_reno(tp) && decr > 0) 1537797108d1SIlpo Järvinen tp->sacked_out -= min_t(u32, tp->sacked_out, decr); 1538797108d1SIlpo Järvinen 1539797108d1SIlpo Järvinen if (tp->lost_skb_hint && 1540797108d1SIlpo Järvinen before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && 1541713bafeaSYuchung Cheng (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) 1542797108d1SIlpo Järvinen tp->lost_cnt_hint -= decr; 1543797108d1SIlpo Järvinen 1544797108d1SIlpo Järvinen tcp_verify_left_out(tp); 1545797108d1SIlpo Järvinen } 1546797108d1SIlpo Järvinen 15470a2cf20cSSoheil Hassas Yeganeh static bool tcp_has_tx_tstamp(const struct sk_buff *skb) 15480a2cf20cSSoheil Hassas Yeganeh { 15490a2cf20cSSoheil Hassas Yeganeh return TCP_SKB_CB(skb)->txstamp_ack || 15500a2cf20cSSoheil Hassas Yeganeh (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP); 15510a2cf20cSSoheil Hassas Yeganeh } 15520a2cf20cSSoheil Hassas Yeganeh 1553490cc7d0SWillem de Bruijn static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2) 1554490cc7d0SWillem de Bruijn { 1555490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo = skb_shinfo(skb); 1556490cc7d0SWillem de Bruijn 15570a2cf20cSSoheil Hassas Yeganeh if (unlikely(tcp_has_tx_tstamp(skb)) && 1558490cc7d0SWillem de Bruijn !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) { 1559490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo2 = skb_shinfo(skb2); 1560490cc7d0SWillem de Bruijn u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP; 1561490cc7d0SWillem de Bruijn 1562490cc7d0SWillem de Bruijn shinfo->tx_flags &= ~tsflags; 1563490cc7d0SWillem de Bruijn shinfo2->tx_flags |= tsflags; 1564490cc7d0SWillem de Bruijn swap(shinfo->tskey, shinfo2->tskey); 1565b51e13faSMartin KaFai Lau TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack; 1566b51e13faSMartin KaFai Lau TCP_SKB_CB(skb)->txstamp_ack = 0; 1567490cc7d0SWillem de Bruijn } 1568490cc7d0SWillem de Bruijn } 1569490cc7d0SWillem de Bruijn 1570a166140eSMartin KaFai Lau static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2) 1571a166140eSMartin KaFai Lau { 1572a166140eSMartin KaFai Lau TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor; 1573a166140eSMartin KaFai Lau TCP_SKB_CB(skb)->eor = 0; 1574a166140eSMartin KaFai Lau } 1575a166140eSMartin KaFai Lau 157675c119afSEric Dumazet /* Insert buff after skb on the write or rtx queue of sk. */ 157775c119afSEric Dumazet static void tcp_insert_write_queue_after(struct sk_buff *skb, 157875c119afSEric Dumazet struct sk_buff *buff, 157975c119afSEric Dumazet struct sock *sk, 158075c119afSEric Dumazet enum tcp_queue tcp_queue) 158175c119afSEric Dumazet { 158275c119afSEric Dumazet if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE) 158375c119afSEric Dumazet __skb_queue_after(&sk->sk_write_queue, skb, buff); 158475c119afSEric Dumazet else 158575c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, buff); 158675c119afSEric Dumazet } 158775c119afSEric Dumazet 15881da177e4SLinus Torvalds /* Function to create two new TCP segments. Shrinks the given segment 15891da177e4SLinus Torvalds * to the specified size and appends a new segment with the rest of the 15901da177e4SLinus Torvalds * packet to the list. This won't be called frequently, I hope. 15911da177e4SLinus Torvalds * Remember, these are still headerless SKBs at this point. 15921da177e4SLinus Torvalds */ 159375c119afSEric Dumazet int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue, 159475c119afSEric Dumazet struct sk_buff *skb, u32 len, 15956cc55e09SOctavian Purdila unsigned int mss_now, gfp_t gfp) 15961da177e4SLinus Torvalds { 15971da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 15981da177e4SLinus Torvalds struct sk_buff *buff; 1599b4a24397SEric Dumazet int old_factor; 1600b617158dSEric Dumazet long limit; 1601b60b49eaSHerbert Xu int nlen; 16029ce01461SIlpo Järvinen u8 flags; 16031da177e4SLinus Torvalds 16042fceec13SIlpo Järvinen if (WARN_ON(len > skb->len)) 16052fceec13SIlpo Järvinen return -EINVAL; 16066a438bbeSStephen Hemminger 1607b4a24397SEric Dumazet DEBUG_NET_WARN_ON_ONCE(skb_headlen(skb)); 16081da177e4SLinus Torvalds 1609b617158dSEric Dumazet /* tcp_sendmsg() can overshoot sk_wmem_queued by one full size skb. 1610b617158dSEric Dumazet * We need some allowance to not penalize applications setting small 1611b617158dSEric Dumazet * SO_SNDBUF values. 1612b617158dSEric Dumazet * Also allow first and last skb in retransmit queue to be split. 1613b617158dSEric Dumazet */ 16147c4e983cSAlexander Duyck limit = sk->sk_sndbuf + 2 * SKB_TRUESIZE(GSO_LEGACY_MAX_SIZE); 1615b617158dSEric Dumazet if (unlikely((sk->sk_wmem_queued >> 1) > limit && 1616b617158dSEric Dumazet tcp_queue != TCP_FRAG_IN_WRITE_QUEUE && 1617b617158dSEric Dumazet skb != tcp_rtx_queue_head(sk) && 1618b617158dSEric Dumazet skb != tcp_rtx_queue_tail(sk))) { 1619f070ef2aSEric Dumazet NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPWQUEUETOOBIG); 1620f070ef2aSEric Dumazet return -ENOMEM; 1621f070ef2aSEric Dumazet } 1622f070ef2aSEric Dumazet 1623c4777efaSEric Dumazet if (skb_unclone_keeptruesize(skb, gfp)) 16241da177e4SLinus Torvalds return -ENOMEM; 16251da177e4SLinus Torvalds 16261da177e4SLinus Torvalds /* Get a new skb... force flag on. */ 16275882efffSEric Dumazet buff = tcp_stream_alloc_skb(sk, gfp, true); 162851456b29SIan Morris if (!buff) 16291da177e4SLinus Torvalds return -ENOMEM; /* We'll just try again later. */ 163041477662SJakub Kicinski skb_copy_decrypted(buff, skb); 16315a369ca6SPaolo Abeni mptcp_skb_ext_copy(buff, skb); 1632ef5cb973SHerbert Xu 1633ab4e846aSEric Dumazet sk_wmem_queued_add(sk, buff->truesize); 16343ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1635b4a24397SEric Dumazet nlen = skb->len - len; 1636b60b49eaSHerbert Xu buff->truesize += nlen; 1637b60b49eaSHerbert Xu skb->truesize -= nlen; 16381da177e4SLinus Torvalds 16391da177e4SLinus Torvalds /* Correct the sequence numbers. */ 16401da177e4SLinus Torvalds TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 16411da177e4SLinus Torvalds TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 16421da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 16431da177e4SLinus Torvalds 16441da177e4SLinus Torvalds /* PSH and FIN should only be set in the second packet. */ 16454de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 16464de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 16474de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1648e14c3cafSHerbert Xu TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; 1649a166140eSMartin KaFai Lau tcp_skb_fragment_eor(skb, buff); 16501da177e4SLinus Torvalds 16511da177e4SLinus Torvalds skb_split(skb, buff, len); 16521da177e4SLinus Torvalds 1653a1ac9c8aSMartin KaFai Lau skb_set_delivery_time(buff, skb->tstamp, true); 1654490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 16551da177e4SLinus Torvalds 16566475be16SDavid S. Miller old_factor = tcp_skb_pcount(skb); 16576475be16SDavid S. Miller 16581da177e4SLinus Torvalds /* Fix up tso_factor for both original and new SKB. */ 16595bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 16605bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 16611da177e4SLinus Torvalds 1662b9f64820SYuchung Cheng /* Update delivered info for the new segment */ 1663b9f64820SYuchung Cheng TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx; 1664b9f64820SYuchung Cheng 16656475be16SDavid S. Miller /* If this packet has been sent out already, we must 16666475be16SDavid S. Miller * adjust the various packet counters. 16676475be16SDavid S. Miller */ 1668cf0b450cSHerbert Xu if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 16696475be16SDavid S. Miller int diff = old_factor - tcp_skb_pcount(skb) - 16706475be16SDavid S. Miller tcp_skb_pcount(buff); 16711da177e4SLinus Torvalds 1672797108d1SIlpo Järvinen if (diff) 1673797108d1SIlpo Järvinen tcp_adjust_pcount(sk, skb, diff); 16741da177e4SLinus Torvalds } 16751da177e4SLinus Torvalds 16761da177e4SLinus Torvalds /* Link BUFF into the send queue. */ 1677f4a775d1SEric Dumazet __skb_header_release(buff); 167875c119afSEric Dumazet tcp_insert_write_queue_after(skb, buff, sk, tcp_queue); 1679f67971e6SEric Dumazet if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE) 1680e2080072SEric Dumazet list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor); 16811da177e4SLinus Torvalds 16821da177e4SLinus Torvalds return 0; 16831da177e4SLinus Torvalds } 16841da177e4SLinus Torvalds 1685f4d01666SEric Dumazet /* This is similar to __pskb_pull_tail(). The difference is that pulled 1686f4d01666SEric Dumazet * data is not copied, but immediately discarded. 16871da177e4SLinus Torvalds */ 16887162fb24SEric Dumazet static int __pskb_trim_head(struct sk_buff *skb, int len) 16891da177e4SLinus Torvalds { 16907b7fc97aSEric Dumazet struct skb_shared_info *shinfo; 16911da177e4SLinus Torvalds int i, k, eat; 16921da177e4SLinus Torvalds 1693b4a24397SEric Dumazet DEBUG_NET_WARN_ON_ONCE(skb_headlen(skb)); 16941da177e4SLinus Torvalds eat = len; 16951da177e4SLinus Torvalds k = 0; 16967b7fc97aSEric Dumazet shinfo = skb_shinfo(skb); 16977b7fc97aSEric Dumazet for (i = 0; i < shinfo->nr_frags; i++) { 16987b7fc97aSEric Dumazet int size = skb_frag_size(&shinfo->frags[i]); 16999e903e08SEric Dumazet 17009e903e08SEric Dumazet if (size <= eat) { 1701aff65da0SIan Campbell skb_frag_unref(skb, i); 17029e903e08SEric Dumazet eat -= size; 17031da177e4SLinus Torvalds } else { 17047b7fc97aSEric Dumazet shinfo->frags[k] = shinfo->frags[i]; 17051da177e4SLinus Torvalds if (eat) { 1706b54c9d5bSJonathan Lemon skb_frag_off_add(&shinfo->frags[k], eat); 17077b7fc97aSEric Dumazet skb_frag_size_sub(&shinfo->frags[k], eat); 17081da177e4SLinus Torvalds eat = 0; 17091da177e4SLinus Torvalds } 17101da177e4SLinus Torvalds k++; 17111da177e4SLinus Torvalds } 17121da177e4SLinus Torvalds } 17137b7fc97aSEric Dumazet shinfo->nr_frags = k; 17141da177e4SLinus Torvalds 17151da177e4SLinus Torvalds skb->data_len -= len; 17161da177e4SLinus Torvalds skb->len = skb->data_len; 17177162fb24SEric Dumazet return len; 17181da177e4SLinus Torvalds } 17191da177e4SLinus Torvalds 172067edfef7SAndi Kleen /* Remove acked data from a packet in the transmit queue. */ 17211da177e4SLinus Torvalds int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 17221da177e4SLinus Torvalds { 17237162fb24SEric Dumazet u32 delta_truesize; 17247162fb24SEric Dumazet 1725c4777efaSEric Dumazet if (skb_unclone_keeptruesize(skb, GFP_ATOMIC)) 17261da177e4SLinus Torvalds return -ENOMEM; 17271da177e4SLinus Torvalds 17287162fb24SEric Dumazet delta_truesize = __pskb_trim_head(skb, len); 17291da177e4SLinus Torvalds 17301da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq += len; 17311da177e4SLinus Torvalds 17327162fb24SEric Dumazet skb->truesize -= delta_truesize; 1733ab4e846aSEric Dumazet sk_wmem_queued_add(sk, -delta_truesize); 17349b65b17dSTalal Ahmad if (!skb_zcopy_pure(skb)) 17357162fb24SEric Dumazet sk_mem_uncharge(sk, delta_truesize); 17361da177e4SLinus Torvalds 17375b35e1e6SNeal Cardwell /* Any change of skb->len requires recalculation of tso factor. */ 17381da177e4SLinus Torvalds if (tcp_skb_pcount(skb) > 1) 17395bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb)); 17401da177e4SLinus Torvalds 17411da177e4SLinus Torvalds return 0; 17421da177e4SLinus Torvalds } 17431da177e4SLinus Torvalds 17441b63edd6SYuchung Cheng /* Calculate MSS not accounting any TCP options. */ 17451b63edd6SYuchung Cheng static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu) 17465d424d5aSJohn Heffner { 1747cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1748cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 17495d424d5aSJohn Heffner int mss_now; 17505d424d5aSJohn Heffner 17515d424d5aSJohn Heffner /* Calculate base mss without TCP options: 17525d424d5aSJohn Heffner It is MMS_S - sizeof(tcphdr) of rfc1122 17535d424d5aSJohn Heffner */ 17545d424d5aSJohn Heffner mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 17555d424d5aSJohn Heffner 17565d424d5aSJohn Heffner /* Clamp it (mss_clamp does not include tcp options) */ 17575d424d5aSJohn Heffner if (mss_now > tp->rx_opt.mss_clamp) 17585d424d5aSJohn Heffner mss_now = tp->rx_opt.mss_clamp; 17595d424d5aSJohn Heffner 17605d424d5aSJohn Heffner /* Now subtract optional transport overhead */ 17615d424d5aSJohn Heffner mss_now -= icsk->icsk_ext_hdr_len; 17625d424d5aSJohn Heffner 17635d424d5aSJohn Heffner /* Then reserve room for full set of TCP options and 8 bytes of data */ 176478eb166cSKuniyuki Iwashima mss_now = max(mss_now, 176578eb166cSKuniyuki Iwashima READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_snd_mss)); 17665d424d5aSJohn Heffner return mss_now; 17675d424d5aSJohn Heffner } 17685d424d5aSJohn Heffner 17691b63edd6SYuchung Cheng /* Calculate MSS. Not accounting for SACKs here. */ 17701b63edd6SYuchung Cheng int tcp_mtu_to_mss(struct sock *sk, int pmtu) 17711b63edd6SYuchung Cheng { 17721b63edd6SYuchung Cheng /* Subtract TCP options size, not including SACKs */ 17731b63edd6SYuchung Cheng return __tcp_mtu_to_mss(sk, pmtu) - 17741b63edd6SYuchung Cheng (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr)); 17751b63edd6SYuchung Cheng } 1776c7bb4b89SEric Dumazet EXPORT_SYMBOL(tcp_mtu_to_mss); 17771b63edd6SYuchung Cheng 17785d424d5aSJohn Heffner /* Inverse of above */ 177967469601SEric Dumazet int tcp_mss_to_mtu(struct sock *sk, int mss) 17805d424d5aSJohn Heffner { 1781cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1782cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 17835d424d5aSJohn Heffner 1784e57a3447SYan Zhai return mss + 17855d424d5aSJohn Heffner tp->tcp_header_len + 17865d424d5aSJohn Heffner icsk->icsk_ext_hdr_len + 17875d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 17885d424d5aSJohn Heffner } 1789556c6b46SNeal Cardwell EXPORT_SYMBOL(tcp_mss_to_mtu); 17905d424d5aSJohn Heffner 179167edfef7SAndi Kleen /* MTU probing init per socket */ 17925d424d5aSJohn Heffner void tcp_mtup_init(struct sock *sk) 17935d424d5aSJohn Heffner { 17945d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 17955d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 1796b0f9ca53SFan Du struct net *net = sock_net(sk); 17975d424d5aSJohn Heffner 1798f47d00e0SKuniyuki Iwashima icsk->icsk_mtup.enabled = READ_ONCE(net->ipv4.sysctl_tcp_mtu_probing) > 1; 17995d424d5aSJohn Heffner icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 18005d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 180188d78bc0SKuniyuki Iwashima icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, READ_ONCE(net->ipv4.sysctl_tcp_base_mss)); 18025d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 180305cbc0dbSFan Du if (icsk->icsk_mtup.enabled) 1804c74df29aSEric Dumazet icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; 18055d424d5aSJohn Heffner } 18064bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_mtup_init); 18075d424d5aSJohn Heffner 18081da177e4SLinus Torvalds /* This function synchronize snd mss to current pmtu/exthdr set. 18091da177e4SLinus Torvalds 18101da177e4SLinus Torvalds tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 18111da177e4SLinus Torvalds for TCP options, but includes only bare TCP header. 18121da177e4SLinus Torvalds 18131da177e4SLinus Torvalds tp->rx_opt.mss_clamp is mss negotiated at connection setup. 1814caa20d9aSStephen Hemminger It is minimum of user_mss and mss received with SYN. 18151da177e4SLinus Torvalds It also does not include TCP options. 18161da177e4SLinus Torvalds 1817d83d8461SArnaldo Carvalho de Melo inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. 18181da177e4SLinus Torvalds 18191da177e4SLinus Torvalds tp->mss_cache is current effective sending mss, including 18201da177e4SLinus Torvalds all tcp options except for SACKs. It is evaluated, 18211da177e4SLinus Torvalds taking into account current pmtu, but never exceeds 18221da177e4SLinus Torvalds tp->rx_opt.mss_clamp. 18231da177e4SLinus Torvalds 18241da177e4SLinus Torvalds NOTE1. rfc1122 clearly states that advertised MSS 18251da177e4SLinus Torvalds DOES NOT include either tcp or ip options. 18261da177e4SLinus Torvalds 1827d83d8461SArnaldo Carvalho de Melo NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache 1828d83d8461SArnaldo Carvalho de Melo are READ ONLY outside this function. --ANK (980731) 18291da177e4SLinus Torvalds */ 18301da177e4SLinus Torvalds unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 18311da177e4SLinus Torvalds { 18321da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1833d83d8461SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 18345d424d5aSJohn Heffner int mss_now; 18351da177e4SLinus Torvalds 18365d424d5aSJohn Heffner if (icsk->icsk_mtup.search_high > pmtu) 18375d424d5aSJohn Heffner icsk->icsk_mtup.search_high = pmtu; 18381da177e4SLinus Torvalds 18395d424d5aSJohn Heffner mss_now = tcp_mtu_to_mss(sk, pmtu); 1840409d22b4SIlpo Järvinen mss_now = tcp_bound_to_half_wnd(tp, mss_now); 18411da177e4SLinus Torvalds 18421da177e4SLinus Torvalds /* And store cached results */ 1843d83d8461SArnaldo Carvalho de Melo icsk->icsk_pmtu_cookie = pmtu; 18445d424d5aSJohn Heffner if (icsk->icsk_mtup.enabled) 18455d424d5aSJohn Heffner mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); 1846c1b4a7e6SDavid S. Miller tp->mss_cache = mss_now; 18471da177e4SLinus Torvalds 18481da177e4SLinus Torvalds return mss_now; 18491da177e4SLinus Torvalds } 18504bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_sync_mss); 18511da177e4SLinus Torvalds 18521da177e4SLinus Torvalds /* Compute the current effective MSS, taking SACKs and IP options, 18531da177e4SLinus Torvalds * and even PMTU discovery events into account. 18541da177e4SLinus Torvalds */ 18550c54b85fSIlpo Järvinen unsigned int tcp_current_mss(struct sock *sk) 18561da177e4SLinus Torvalds { 1857cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1858cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1859c1b4a7e6SDavid S. Miller u32 mss_now; 186095c96174SEric Dumazet unsigned int header_len; 186133ad798cSAdam Langley struct tcp_out_options opts; 18621e03d32bSDmitry Safonov struct tcp_key key; 18631da177e4SLinus Torvalds 1864c1b4a7e6SDavid S. Miller mss_now = tp->mss_cache; 1865c1b4a7e6SDavid S. Miller 18661da177e4SLinus Torvalds if (dst) { 18671da177e4SLinus Torvalds u32 mtu = dst_mtu(dst); 1868d83d8461SArnaldo Carvalho de Melo if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 18691da177e4SLinus Torvalds mss_now = tcp_sync_mss(sk, mtu); 18701da177e4SLinus Torvalds } 18711e03d32bSDmitry Safonov tcp_get_current_key(sk, &key); 18721e03d32bSDmitry Safonov header_len = tcp_established_options(sk, NULL, &opts, &key) + 187333ad798cSAdam Langley sizeof(struct tcphdr); 187433ad798cSAdam Langley /* The mss_cache is sized based on tp->tcp_header_len, which assumes 187533ad798cSAdam Langley * some common options. If this is an odd packet (because we have SACK 187633ad798cSAdam Langley * blocks etc) then our calculated header_len will be different, and 187733ad798cSAdam Langley * we have to adjust mss_now correspondingly */ 187833ad798cSAdam Langley if (header_len != tp->tcp_header_len) { 187933ad798cSAdam Langley int delta = (int) header_len - tp->tcp_header_len; 188033ad798cSAdam Langley mss_now -= delta; 188133ad798cSAdam Langley } 1882cfb6eeb4SYOSHIFUJI Hideaki 18831da177e4SLinus Torvalds return mss_now; 18841da177e4SLinus Torvalds } 18851da177e4SLinus Torvalds 188686fd14adSWeiping Pan /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. 188786fd14adSWeiping Pan * As additional protections, we do not touch cwnd in retransmission phases, 188886fd14adSWeiping Pan * and if application hit its sndbuf limit recently. 188986fd14adSWeiping Pan */ 189086fd14adSWeiping Pan static void tcp_cwnd_application_limited(struct sock *sk) 1891a762a980SDavid S. Miller { 18929e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1893a762a980SDavid S. Miller 189486fd14adSWeiping Pan if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && 189586fd14adSWeiping Pan sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 189686fd14adSWeiping Pan /* Limited by application or receiver window. */ 189786fd14adSWeiping Pan u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); 189886fd14adSWeiping Pan u32 win_used = max(tp->snd_cwnd_used, init_win); 189940570375SEric Dumazet if (win_used < tcp_snd_cwnd(tp)) { 190086fd14adSWeiping Pan tp->snd_ssthresh = tcp_current_ssthresh(sk); 190140570375SEric Dumazet tcp_snd_cwnd_set(tp, (tcp_snd_cwnd(tp) + win_used) >> 1); 190286fd14adSWeiping Pan } 190386fd14adSWeiping Pan tp->snd_cwnd_used = 0; 190486fd14adSWeiping Pan } 1905c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 190686fd14adSWeiping Pan } 190786fd14adSWeiping Pan 1908ca8a2263SNeal Cardwell static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited) 1909a762a980SDavid S. Miller { 19101b1fc3fdSWei Wang const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 1911a762a980SDavid S. Miller struct tcp_sock *tp = tcp_sk(sk); 1912a762a980SDavid S. Miller 1913f4ce91ceSNeal Cardwell /* Track the strongest available signal of the degree to which the cwnd 1914f4ce91ceSNeal Cardwell * is fully utilized. If cwnd-limited then remember that fact for the 1915f4ce91ceSNeal Cardwell * current window. If not cwnd-limited then track the maximum number of 1916f4ce91ceSNeal Cardwell * outstanding packets in the current window. (If cwnd-limited then we 1917f4ce91ceSNeal Cardwell * chose to not update tp->max_packets_out to avoid an extra else 1918f4ce91ceSNeal Cardwell * clause with no functional impact.) 1919ca8a2263SNeal Cardwell */ 1920f4ce91ceSNeal Cardwell if (!before(tp->snd_una, tp->cwnd_usage_seq) || 1921f4ce91ceSNeal Cardwell is_cwnd_limited || 1922f4ce91ceSNeal Cardwell (!tp->is_cwnd_limited && 1923f4ce91ceSNeal Cardwell tp->packets_out > tp->max_packets_out)) { 1924ca8a2263SNeal Cardwell tp->is_cwnd_limited = is_cwnd_limited; 1925f4ce91ceSNeal Cardwell tp->max_packets_out = tp->packets_out; 1926f4ce91ceSNeal Cardwell tp->cwnd_usage_seq = tp->snd_nxt; 1927ca8a2263SNeal Cardwell } 1928e114a710SEric Dumazet 192924901551SEric Dumazet if (tcp_is_cwnd_limited(sk)) { 1930a762a980SDavid S. Miller /* Network is feed fully. */ 1931a762a980SDavid S. Miller tp->snd_cwnd_used = 0; 1932c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 1933a762a980SDavid S. Miller } else { 1934a762a980SDavid S. Miller /* Network starves. */ 1935a762a980SDavid S. Miller if (tp->packets_out > tp->snd_cwnd_used) 1936a762a980SDavid S. Miller tp->snd_cwnd_used = tp->packets_out; 1937a762a980SDavid S. Miller 19384845b571SKuniyuki Iwashima if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle) && 1939c2203cf7SEric Dumazet (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto && 19401b1fc3fdSWei Wang !ca_ops->cong_control) 1941a762a980SDavid S. Miller tcp_cwnd_application_limited(sk); 1942b0f71bd3SFrancis Yan 1943b0f71bd3SFrancis Yan /* The following conditions together indicate the starvation 1944b0f71bd3SFrancis Yan * is caused by insufficient sender buffer: 1945b0f71bd3SFrancis Yan * 1) just sent some data (see tcp_write_xmit) 1946b0f71bd3SFrancis Yan * 2) not cwnd limited (this else condition) 194775c119afSEric Dumazet * 3) no more data to send (tcp_write_queue_empty()) 1948b0f71bd3SFrancis Yan * 4) application is hitting buffer limit (SOCK_NOSPACE) 1949b0f71bd3SFrancis Yan */ 195075c119afSEric Dumazet if (tcp_write_queue_empty(sk) && sk->sk_socket && 1951b0f71bd3SFrancis Yan test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) && 1952b0f71bd3SFrancis Yan (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) 1953b0f71bd3SFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED); 1954a762a980SDavid S. Miller } 1955a762a980SDavid S. Miller } 1956a762a980SDavid S. Miller 1957d4589926SEric Dumazet /* Minshall's variant of the Nagle send check. */ 1958d4589926SEric Dumazet static bool tcp_minshall_check(const struct tcp_sock *tp) 1959d4589926SEric Dumazet { 1960d4589926SEric Dumazet return after(tp->snd_sml, tp->snd_una) && 1961d4589926SEric Dumazet !after(tp->snd_sml, tp->snd_nxt); 1962d4589926SEric Dumazet } 1963d4589926SEric Dumazet 1964d4589926SEric Dumazet /* Update snd_sml if this skb is under mss 1965d4589926SEric Dumazet * Note that a TSO packet might end with a sub-mss segment 1966d4589926SEric Dumazet * The test is really : 1967d4589926SEric Dumazet * if ((skb->len % mss) != 0) 1968d4589926SEric Dumazet * tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1969d4589926SEric Dumazet * But we can avoid doing the divide again given we already have 1970d4589926SEric Dumazet * skb_pcount = skb->len / mss_now 19710e3a4803SIlpo Järvinen */ 1972d4589926SEric Dumazet static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now, 1973d4589926SEric Dumazet const struct sk_buff *skb) 1974d4589926SEric Dumazet { 1975d4589926SEric Dumazet if (skb->len < tcp_skb_pcount(skb) * mss_now) 1976d4589926SEric Dumazet tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1977d4589926SEric Dumazet } 1978d4589926SEric Dumazet 1979d4589926SEric Dumazet /* Return false, if packet can be sent now without violation Nagle's rules: 1980d4589926SEric Dumazet * 1. It is full sized. (provided by caller in %partial bool) 1981d4589926SEric Dumazet * 2. Or it contains FIN. (already checked by caller) 1982d4589926SEric Dumazet * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. 1983d4589926SEric Dumazet * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 1984d4589926SEric Dumazet * With Minshall's modification: all sent small packets are ACKed. 1985d4589926SEric Dumazet */ 1986d4589926SEric Dumazet static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp, 1987cc93fc51SPeter Pan(潘卫平) int nonagle) 1988d4589926SEric Dumazet { 1989d4589926SEric Dumazet return partial && 1990d4589926SEric Dumazet ((nonagle & TCP_NAGLE_CORK) || 1991d4589926SEric Dumazet (!nonagle && tp->packets_out && tcp_minshall_check(tp))); 1992d4589926SEric Dumazet } 1993605ad7f1SEric Dumazet 1994605ad7f1SEric Dumazet /* Return how many segs we'd like on a TSO packet, 199565466904SEric Dumazet * depending on current pacing rate, and how close the peer is. 199665466904SEric Dumazet * 199765466904SEric Dumazet * Rationale is: 199865466904SEric Dumazet * - For close peers, we rather send bigger packets to reduce 199965466904SEric Dumazet * cpu costs, because occasional losses will be repaired fast. 200065466904SEric Dumazet * - For long distance/rtt flows, we would like to get ACK clocking 200165466904SEric Dumazet * with 1 ACK per ms. 200265466904SEric Dumazet * 200365466904SEric Dumazet * Use min_rtt to help adapt TSO burst size, with smaller min_rtt resulting 200465466904SEric Dumazet * in bigger TSO bursts. We we cut the RTT-based allowance in half 200565466904SEric Dumazet * for every 2^9 usec (aka 512 us) of RTT, so that the RTT-based allowance 200665466904SEric Dumazet * is below 1500 bytes after 6 * ~500 usec = 3ms. 2007605ad7f1SEric Dumazet */ 2008dcb8c9b4SEric Dumazet static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now, 20091b3878caSNeal Cardwell int min_tso_segs) 2010605ad7f1SEric Dumazet { 201165466904SEric Dumazet unsigned long bytes; 201265466904SEric Dumazet u32 r; 2013605ad7f1SEric Dumazet 201428b24f90SEric Dumazet bytes = READ_ONCE(sk->sk_pacing_rate) >> READ_ONCE(sk->sk_pacing_shift); 2015605ad7f1SEric Dumazet 20162455e61bSKuniyuki Iwashima r = tcp_min_rtt(tcp_sk(sk)) >> READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_rtt_log); 201765466904SEric Dumazet if (r < BITS_PER_TYPE(sk->sk_gso_max_size)) 201865466904SEric Dumazet bytes += sk->sk_gso_max_size >> r; 2019605ad7f1SEric Dumazet 202065466904SEric Dumazet bytes = min_t(unsigned long, bytes, sk->sk_gso_max_size); 202165466904SEric Dumazet 202265466904SEric Dumazet return max_t(u32, bytes / mss_now, min_tso_segs); 2023605ad7f1SEric Dumazet } 2024605ad7f1SEric Dumazet 2025ed6e7268SNeal Cardwell /* Return the number of segments we want in the skb we are transmitting. 2026ed6e7268SNeal Cardwell * See if congestion control module wants to decide; otherwise, autosize. 2027ed6e7268SNeal Cardwell */ 2028ed6e7268SNeal Cardwell static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now) 2029ed6e7268SNeal Cardwell { 2030ed6e7268SNeal Cardwell const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 2031dcb8c9b4SEric Dumazet u32 min_tso, tso_segs; 2032ed6e7268SNeal Cardwell 2033dcb8c9b4SEric Dumazet min_tso = ca_ops->min_tso_segs ? 2034dcb8c9b4SEric Dumazet ca_ops->min_tso_segs(sk) : 2035e0bb4ab9SKuniyuki Iwashima READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs); 2036dcb8c9b4SEric Dumazet 2037dcb8c9b4SEric Dumazet tso_segs = tcp_tso_autosize(sk, mss_now, min_tso); 2038350c9f48SEric Dumazet return min_t(u32, tso_segs, sk->sk_gso_max_segs); 2039ed6e7268SNeal Cardwell } 2040ed6e7268SNeal Cardwell 2041d4589926SEric Dumazet /* Returns the portion of skb which can be sent right away */ 2042d4589926SEric Dumazet static unsigned int tcp_mss_split_point(const struct sock *sk, 2043d4589926SEric Dumazet const struct sk_buff *skb, 2044d4589926SEric Dumazet unsigned int mss_now, 2045d4589926SEric Dumazet unsigned int max_segs, 2046d4589926SEric Dumazet int nonagle) 2047c1b4a7e6SDavid S. Miller { 2048cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 2049d4589926SEric Dumazet u32 partial, needed, window, max_len; 2050c1b4a7e6SDavid S. Miller 205190840defSIlpo Järvinen window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 20521485348dSBen Hutchings max_len = mss_now * max_segs; 20530e3a4803SIlpo Järvinen 20541485348dSBen Hutchings if (likely(max_len <= window && skb != tcp_write_queue_tail(sk))) 20551485348dSBen Hutchings return max_len; 20560e3a4803SIlpo Järvinen 20575ea3a748SIlpo Järvinen needed = min(skb->len, window); 20585ea3a748SIlpo Järvinen 20591485348dSBen Hutchings if (max_len <= needed) 20601485348dSBen Hutchings return max_len; 20610e3a4803SIlpo Järvinen 2062d4589926SEric Dumazet partial = needed % mss_now; 2063d4589926SEric Dumazet /* If last segment is not a full MSS, check if Nagle rules allow us 2064d4589926SEric Dumazet * to include this last segment in this skb. 2065d4589926SEric Dumazet * Otherwise, we'll split the skb at last MSS boundary 2066d4589926SEric Dumazet */ 2067cc93fc51SPeter Pan(潘卫平) if (tcp_nagle_check(partial != 0, tp, nonagle)) 2068d4589926SEric Dumazet return needed - partial; 2069d4589926SEric Dumazet 2070d4589926SEric Dumazet return needed; 2071c1b4a7e6SDavid S. Miller } 2072c1b4a7e6SDavid S. Miller 2073c1b4a7e6SDavid S. Miller /* Can at least one segment of SKB be sent right now, according to the 2074c1b4a7e6SDavid S. Miller * congestion window rules? If so, return how many segments are allowed. 2075c1b4a7e6SDavid S. Miller */ 2076cf533ea5SEric Dumazet static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, 2077cf533ea5SEric Dumazet const struct sk_buff *skb) 2078c1b4a7e6SDavid S. Miller { 2079d649a7a8SEric Dumazet u32 in_flight, cwnd, halfcwnd; 2080c1b4a7e6SDavid S. Miller 2081c1b4a7e6SDavid S. Miller /* Don't be strict about the congestion window for the final FIN. */ 20824de075e0SEric Dumazet if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && 20834de075e0SEric Dumazet tcp_skb_pcount(skb) == 1) 2084c1b4a7e6SDavid S. Miller return 1; 2085c1b4a7e6SDavid S. Miller 2086c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 208740570375SEric Dumazet cwnd = tcp_snd_cwnd(tp); 2088d649a7a8SEric Dumazet if (in_flight >= cwnd) 2089c1b4a7e6SDavid S. Miller return 0; 2090d649a7a8SEric Dumazet 2091d649a7a8SEric Dumazet /* For better scheduling, ensure we have at least 2092d649a7a8SEric Dumazet * 2 GSO packets in flight. 2093d649a7a8SEric Dumazet */ 2094d649a7a8SEric Dumazet halfcwnd = max(cwnd >> 1, 1U); 2095d649a7a8SEric Dumazet return min(halfcwnd, cwnd - in_flight); 2096c1b4a7e6SDavid S. Miller } 2097c1b4a7e6SDavid S. Miller 2098b595076aSUwe Kleine-König /* Initialize TSO state of a skb. 209967edfef7SAndi Kleen * This must be invoked the first time we consider transmitting 2100c1b4a7e6SDavid S. Miller * SKB onto the wire. 2101c1b4a7e6SDavid S. Miller */ 21025bbb432cSEric Dumazet static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now) 2103c1b4a7e6SDavid S. Miller { 2104c1b4a7e6SDavid S. Miller int tso_segs = tcp_skb_pcount(skb); 2105c1b4a7e6SDavid S. Miller 2106f8269a49SIlpo Järvinen if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { 21075bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 2108c1b4a7e6SDavid S. Miller tso_segs = tcp_skb_pcount(skb); 2109c1b4a7e6SDavid S. Miller } 2110c1b4a7e6SDavid S. Miller return tso_segs; 2111c1b4a7e6SDavid S. Miller } 2112c1b4a7e6SDavid S. Miller 2113c1b4a7e6SDavid S. Miller 2114a2a385d6SEric Dumazet /* Return true if the Nagle test allows this packet to be 2115c1b4a7e6SDavid S. Miller * sent now. 2116c1b4a7e6SDavid S. Miller */ 2117a2a385d6SEric Dumazet static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, 2118c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 2119c1b4a7e6SDavid S. Miller { 2120c1b4a7e6SDavid S. Miller /* Nagle rule does not apply to frames, which sit in the middle of the 2121c1b4a7e6SDavid S. Miller * write_queue (they have no chances to get new data). 2122c1b4a7e6SDavid S. Miller * 2123c1b4a7e6SDavid S. Miller * This is implemented in the callers, where they modify the 'nonagle' 2124c1b4a7e6SDavid S. Miller * argument based upon the location of SKB in the send queue. 2125c1b4a7e6SDavid S. Miller */ 2126c1b4a7e6SDavid S. Miller if (nonagle & TCP_NAGLE_PUSH) 2127a2a385d6SEric Dumazet return true; 2128c1b4a7e6SDavid S. Miller 21299b44190dSYuchung Cheng /* Don't use the nagle rule for urgent data (or for the final FIN). */ 21309b44190dSYuchung Cheng if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) 2131a2a385d6SEric Dumazet return true; 2132c1b4a7e6SDavid S. Miller 2133cc93fc51SPeter Pan(潘卫平) if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle)) 2134a2a385d6SEric Dumazet return true; 2135c1b4a7e6SDavid S. Miller 2136a2a385d6SEric Dumazet return false; 2137c1b4a7e6SDavid S. Miller } 2138c1b4a7e6SDavid S. Miller 2139c1b4a7e6SDavid S. Miller /* Does at least the first segment of SKB fit into the send window? */ 2140a2a385d6SEric Dumazet static bool tcp_snd_wnd_test(const struct tcp_sock *tp, 2141a2a385d6SEric Dumazet const struct sk_buff *skb, 2142056834d9SIlpo Järvinen unsigned int cur_mss) 2143c1b4a7e6SDavid S. Miller { 2144c1b4a7e6SDavid S. Miller u32 end_seq = TCP_SKB_CB(skb)->end_seq; 2145c1b4a7e6SDavid S. Miller 2146c1b4a7e6SDavid S. Miller if (skb->len > cur_mss) 2147c1b4a7e6SDavid S. Miller end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 2148c1b4a7e6SDavid S. Miller 214990840defSIlpo Järvinen return !after(end_seq, tcp_wnd_end(tp)); 2150c1b4a7e6SDavid S. Miller } 2151c1b4a7e6SDavid S. Miller 2152c1b4a7e6SDavid S. Miller /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 2153c1b4a7e6SDavid S. Miller * which is put after SKB on the list. It is very much like 2154c1b4a7e6SDavid S. Miller * tcp_fragment() except that it may make several kinds of assumptions 2155c1b4a7e6SDavid S. Miller * in order to speed up the splitting operation. In particular, we 2156c1b4a7e6SDavid S. Miller * know that all the data is in scatter-gather pages, and that the 2157c1b4a7e6SDavid S. Miller * packet has never been sent out before (and thus is not cloned). 2158c1b4a7e6SDavid S. Miller */ 215956483341SEric Dumazet static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, 2160c4ead4c5SEric Dumazet unsigned int mss_now, gfp_t gfp) 2161c1b4a7e6SDavid S. Miller { 2162c1b4a7e6SDavid S. Miller int nlen = skb->len - len; 216356483341SEric Dumazet struct sk_buff *buff; 21649ce01461SIlpo Järvinen u8 flags; 2165c1b4a7e6SDavid S. Miller 2166c1b4a7e6SDavid S. Miller /* All of a TSO frame must be composed of paged data. */ 2167b4a24397SEric Dumazet DEBUG_NET_WARN_ON_ONCE(skb->len != skb->data_len); 2168c1b4a7e6SDavid S. Miller 21695882efffSEric Dumazet buff = tcp_stream_alloc_skb(sk, gfp, true); 217051456b29SIan Morris if (unlikely(!buff)) 2171c1b4a7e6SDavid S. Miller return -ENOMEM; 217241477662SJakub Kicinski skb_copy_decrypted(buff, skb); 21735a369ca6SPaolo Abeni mptcp_skb_ext_copy(buff, skb); 2174c1b4a7e6SDavid S. Miller 2175ab4e846aSEric Dumazet sk_wmem_queued_add(sk, buff->truesize); 21763ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 2177b60b49eaSHerbert Xu buff->truesize += nlen; 2178c1b4a7e6SDavid S. Miller skb->truesize -= nlen; 2179c1b4a7e6SDavid S. Miller 2180c1b4a7e6SDavid S. Miller /* Correct the sequence numbers. */ 2181c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 2182c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 2183c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 2184c1b4a7e6SDavid S. Miller 2185c1b4a7e6SDavid S. Miller /* PSH and FIN should only be set in the second packet. */ 21864de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 21874de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 21884de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 2189c1b4a7e6SDavid S. Miller 2190a166140eSMartin KaFai Lau tcp_skb_fragment_eor(skb, buff); 2191a166140eSMartin KaFai Lau 2192c1b4a7e6SDavid S. Miller skb_split(skb, buff, len); 2193490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 2194c1b4a7e6SDavid S. Miller 2195c1b4a7e6SDavid S. Miller /* Fix up tso_factor for both original and new SKB. */ 21965bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 21975bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 2198c1b4a7e6SDavid S. Miller 2199c1b4a7e6SDavid S. Miller /* Link BUFF into the send queue. */ 2200f4a775d1SEric Dumazet __skb_header_release(buff); 220156483341SEric Dumazet tcp_insert_write_queue_after(skb, buff, sk, TCP_FRAG_IN_WRITE_QUEUE); 2202c1b4a7e6SDavid S. Miller 2203c1b4a7e6SDavid S. Miller return 0; 2204c1b4a7e6SDavid S. Miller } 2205c1b4a7e6SDavid S. Miller 2206c1b4a7e6SDavid S. Miller /* Try to defer sending, if possible, in order to minimize the amount 2207c1b4a7e6SDavid S. Miller * of TSO splitting we do. View it as a kind of TSO Nagle test. 2208c1b4a7e6SDavid S. Miller * 2209c1b4a7e6SDavid S. Miller * This algorithm is from John Heffner. 2210c1b4a7e6SDavid S. Miller */ 2211ca8a2263SNeal Cardwell static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb, 2212f9bfe4e6SEric Dumazet bool *is_cwnd_limited, 2213f9bfe4e6SEric Dumazet bool *is_rwnd_limited, 2214f9bfe4e6SEric Dumazet u32 max_segs) 2215c1b4a7e6SDavid S. Miller { 22166687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 2217f1c6ea38SEric Dumazet u32 send_win, cong_win, limit, in_flight; 221850c8339eSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 221950c8339eSEric Dumazet struct sk_buff *head; 2220ad9f4f50SEric Dumazet int win_divisor; 2221f1c6ea38SEric Dumazet s64 delta; 2222c1b4a7e6SDavid S. Miller 222399d7662aSEric Dumazet if (icsk->icsk_ca_state >= TCP_CA_Recovery) 2224ae8064acSJohn Heffner goto send_now; 2225ae8064acSJohn Heffner 22265f852eb5SEric Dumazet /* Avoid bursty behavior by allowing defer 2227a682850aSEric Dumazet * only if the last write was recent (1 ms). 2228a682850aSEric Dumazet * Note that tp->tcp_wstamp_ns can be in the future if we have 2229a682850aSEric Dumazet * packets waiting in a qdisc or device for EDT delivery. 22305f852eb5SEric Dumazet */ 2231a682850aSEric Dumazet delta = tp->tcp_clock_cache - tp->tcp_wstamp_ns - NSEC_PER_MSEC; 2232a682850aSEric Dumazet if (delta > 0) 2233ae8064acSJohn Heffner goto send_now; 2234908a75c1SDavid S. Miller 2235c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 2236c1b4a7e6SDavid S. Miller 2237c8c9aeb5SStefano Brivio BUG_ON(tcp_skb_pcount(skb) <= 1); 223840570375SEric Dumazet BUG_ON(tcp_snd_cwnd(tp) <= in_flight); 2239c1b4a7e6SDavid S. Miller 224090840defSIlpo Järvinen send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 2241c1b4a7e6SDavid S. Miller 2242c1b4a7e6SDavid S. Miller /* From in_flight test above, we know that cwnd > in_flight. */ 224340570375SEric Dumazet cong_win = (tcp_snd_cwnd(tp) - in_flight) * tp->mss_cache; 2244c1b4a7e6SDavid S. Miller 2245c1b4a7e6SDavid S. Miller limit = min(send_win, cong_win); 2246c1b4a7e6SDavid S. Miller 2247ba244fe9SDavid S. Miller /* If a full-sized TSO skb can be sent, do it. */ 2248605ad7f1SEric Dumazet if (limit >= max_segs * tp->mss_cache) 2249ae8064acSJohn Heffner goto send_now; 2250ba244fe9SDavid S. Miller 225162ad2761SIlpo Järvinen /* Middle in queue won't get any more data, full sendable already? */ 225262ad2761SIlpo Järvinen if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) 225362ad2761SIlpo Järvinen goto send_now; 225462ad2761SIlpo Järvinen 22555bbcc0f5SLinus Torvalds win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor); 2256ad9f4f50SEric Dumazet if (win_divisor) { 225740570375SEric Dumazet u32 chunk = min(tp->snd_wnd, tcp_snd_cwnd(tp) * tp->mss_cache); 2258c1b4a7e6SDavid S. Miller 2259c1b4a7e6SDavid S. Miller /* If at least some fraction of a window is available, 2260c1b4a7e6SDavid S. Miller * just use it. 2261c1b4a7e6SDavid S. Miller */ 2262ad9f4f50SEric Dumazet chunk /= win_divisor; 2263c1b4a7e6SDavid S. Miller if (limit >= chunk) 2264ae8064acSJohn Heffner goto send_now; 2265c1b4a7e6SDavid S. Miller } else { 2266c1b4a7e6SDavid S. Miller /* Different approach, try not to defer past a single 2267c1b4a7e6SDavid S. Miller * ACK. Receiver should ACK every other full sized 2268c1b4a7e6SDavid S. Miller * frame, so if we have space for more than 3 frames 2269c1b4a7e6SDavid S. Miller * then send now. 2270c1b4a7e6SDavid S. Miller */ 22716b5a5c0dSNeal Cardwell if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) 2272ae8064acSJohn Heffner goto send_now; 2273c1b4a7e6SDavid S. Miller } 2274c1b4a7e6SDavid S. Miller 227575c119afSEric Dumazet /* TODO : use tsorted_sent_queue ? */ 227675c119afSEric Dumazet head = tcp_rtx_queue_head(sk); 227775c119afSEric Dumazet if (!head) 227875c119afSEric Dumazet goto send_now; 2279f1c6ea38SEric Dumazet delta = tp->tcp_clock_cache - head->tstamp; 228050c8339eSEric Dumazet /* If next ACK is likely to come too late (half srtt), do not defer */ 2281f1c6ea38SEric Dumazet if ((s64)(delta - (u64)NSEC_PER_USEC * (tp->srtt_us >> 4)) < 0) 228250c8339eSEric Dumazet goto send_now; 228350c8339eSEric Dumazet 2284f9bfe4e6SEric Dumazet /* Ok, it looks like it is advisable to defer. 2285f9bfe4e6SEric Dumazet * Three cases are tracked : 2286f9bfe4e6SEric Dumazet * 1) We are cwnd-limited 2287f9bfe4e6SEric Dumazet * 2) We are rwnd-limited 2288f9bfe4e6SEric Dumazet * 3) We are application limited. 2289f9bfe4e6SEric Dumazet */ 2290f9bfe4e6SEric Dumazet if (cong_win < send_win) { 2291f9bfe4e6SEric Dumazet if (cong_win <= skb->len) { 2292ca8a2263SNeal Cardwell *is_cwnd_limited = true; 2293f9bfe4e6SEric Dumazet return true; 2294f9bfe4e6SEric Dumazet } 2295f9bfe4e6SEric Dumazet } else { 2296f9bfe4e6SEric Dumazet if (send_win <= skb->len) { 2297f9bfe4e6SEric Dumazet *is_rwnd_limited = true; 2298f9bfe4e6SEric Dumazet return true; 2299f9bfe4e6SEric Dumazet } 2300f9bfe4e6SEric Dumazet } 2301f9bfe4e6SEric Dumazet 2302f9bfe4e6SEric Dumazet /* If this packet won't get more data, do not wait. */ 2303d8ed257fSEric Dumazet if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) || 2304d8ed257fSEric Dumazet TCP_SKB_CB(skb)->eor) 2305f9bfe4e6SEric Dumazet goto send_now; 2306ca8a2263SNeal Cardwell 2307a2a385d6SEric Dumazet return true; 2308ae8064acSJohn Heffner 2309ae8064acSJohn Heffner send_now: 2310a2a385d6SEric Dumazet return false; 2311c1b4a7e6SDavid S. Miller } 2312c1b4a7e6SDavid S. Miller 231305cbc0dbSFan Du static inline void tcp_mtu_check_reprobe(struct sock *sk) 231405cbc0dbSFan Du { 231505cbc0dbSFan Du struct inet_connection_sock *icsk = inet_csk(sk); 231605cbc0dbSFan Du struct tcp_sock *tp = tcp_sk(sk); 231705cbc0dbSFan Du struct net *net = sock_net(sk); 231805cbc0dbSFan Du u32 interval; 231905cbc0dbSFan Du s32 delta; 232005cbc0dbSFan Du 23212a85388fSKuniyuki Iwashima interval = READ_ONCE(net->ipv4.sysctl_tcp_probe_interval); 2322c74df29aSEric Dumazet delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp; 232305cbc0dbSFan Du if (unlikely(delta >= interval * HZ)) { 232405cbc0dbSFan Du int mss = tcp_current_mss(sk); 232505cbc0dbSFan Du 232605cbc0dbSFan Du /* Update current search range */ 232705cbc0dbSFan Du icsk->icsk_mtup.probe_size = 0; 232805cbc0dbSFan Du icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + 232905cbc0dbSFan Du sizeof(struct tcphdr) + 233005cbc0dbSFan Du icsk->icsk_af_ops->net_header_len; 233105cbc0dbSFan Du icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss); 233205cbc0dbSFan Du 233305cbc0dbSFan Du /* Update probe time stamp */ 2334c74df29aSEric Dumazet icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; 233505cbc0dbSFan Du } 233605cbc0dbSFan Du } 233705cbc0dbSFan Du 2338808cf9e3SIlya Lesokhin static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len) 2339808cf9e3SIlya Lesokhin { 2340808cf9e3SIlya Lesokhin struct sk_buff *skb, *next; 2341808cf9e3SIlya Lesokhin 2342808cf9e3SIlya Lesokhin skb = tcp_send_head(sk); 2343808cf9e3SIlya Lesokhin tcp_for_write_queue_from_safe(skb, next, sk) { 2344808cf9e3SIlya Lesokhin if (len <= skb->len) 2345808cf9e3SIlya Lesokhin break; 2346808cf9e3SIlya Lesokhin 23479b65b17dSTalal Ahmad if (unlikely(TCP_SKB_CB(skb)->eor) || 23489b65b17dSTalal Ahmad tcp_has_tx_tstamp(skb) || 23499b65b17dSTalal Ahmad !skb_pure_zcopy_same(skb, next)) 2350808cf9e3SIlya Lesokhin return false; 2351808cf9e3SIlya Lesokhin 2352808cf9e3SIlya Lesokhin len -= skb->len; 2353808cf9e3SIlya Lesokhin } 2354808cf9e3SIlya Lesokhin 2355808cf9e3SIlya Lesokhin return true; 2356808cf9e3SIlya Lesokhin } 2357808cf9e3SIlya Lesokhin 235873601329SEric Dumazet static int tcp_clone_payload(struct sock *sk, struct sk_buff *to, 235973601329SEric Dumazet int probe_size) 236073601329SEric Dumazet { 236173601329SEric Dumazet skb_frag_t *lastfrag = NULL, *fragto = skb_shinfo(to)->frags; 236273601329SEric Dumazet int i, todo, len = 0, nr_frags = 0; 236373601329SEric Dumazet const struct sk_buff *skb; 236473601329SEric Dumazet 236573601329SEric Dumazet if (!sk_wmem_schedule(sk, to->truesize + probe_size)) 236673601329SEric Dumazet return -ENOMEM; 236773601329SEric Dumazet 236873601329SEric Dumazet skb_queue_walk(&sk->sk_write_queue, skb) { 236973601329SEric Dumazet const skb_frag_t *fragfrom = skb_shinfo(skb)->frags; 237073601329SEric Dumazet 237173601329SEric Dumazet if (skb_headlen(skb)) 237273601329SEric Dumazet return -EINVAL; 237373601329SEric Dumazet 237473601329SEric Dumazet for (i = 0; i < skb_shinfo(skb)->nr_frags; i++, fragfrom++) { 237573601329SEric Dumazet if (len >= probe_size) 237673601329SEric Dumazet goto commit; 237773601329SEric Dumazet todo = min_t(int, skb_frag_size(fragfrom), 237873601329SEric Dumazet probe_size - len); 237973601329SEric Dumazet len += todo; 238073601329SEric Dumazet if (lastfrag && 238173601329SEric Dumazet skb_frag_page(fragfrom) == skb_frag_page(lastfrag) && 238273601329SEric Dumazet skb_frag_off(fragfrom) == skb_frag_off(lastfrag) + 238373601329SEric Dumazet skb_frag_size(lastfrag)) { 238473601329SEric Dumazet skb_frag_size_add(lastfrag, todo); 238573601329SEric Dumazet continue; 238673601329SEric Dumazet } 238773601329SEric Dumazet if (unlikely(nr_frags == MAX_SKB_FRAGS)) 238873601329SEric Dumazet return -E2BIG; 238973601329SEric Dumazet skb_frag_page_copy(fragto, fragfrom); 239073601329SEric Dumazet skb_frag_off_copy(fragto, fragfrom); 239173601329SEric Dumazet skb_frag_size_set(fragto, todo); 239273601329SEric Dumazet nr_frags++; 239373601329SEric Dumazet lastfrag = fragto++; 239473601329SEric Dumazet } 239573601329SEric Dumazet } 239673601329SEric Dumazet commit: 239773601329SEric Dumazet WARN_ON_ONCE(len != probe_size); 239873601329SEric Dumazet for (i = 0; i < nr_frags; i++) 239973601329SEric Dumazet skb_frag_ref(to, i); 240073601329SEric Dumazet 240173601329SEric Dumazet skb_shinfo(to)->nr_frags = nr_frags; 240273601329SEric Dumazet to->truesize += probe_size; 240373601329SEric Dumazet to->len += probe_size; 240473601329SEric Dumazet to->data_len += probe_size; 240573601329SEric Dumazet __skb_header_release(to); 240673601329SEric Dumazet return 0; 240773601329SEric Dumazet } 240873601329SEric Dumazet 24095d424d5aSJohn Heffner /* Create a new MTU probe if we are ready. 241067edfef7SAndi Kleen * MTU probe is regularly attempting to increase the path MTU by 241167edfef7SAndi Kleen * deliberately sending larger packets. This discovers routing 241267edfef7SAndi Kleen * changes resulting in larger path MTUs. 241367edfef7SAndi Kleen * 24145d424d5aSJohn Heffner * Returns 0 if we should wait to probe (no cwnd available), 24155d424d5aSJohn Heffner * 1 if a probe was sent, 2416056834d9SIlpo Järvinen * -1 otherwise 2417056834d9SIlpo Järvinen */ 24185d424d5aSJohn Heffner static int tcp_mtu_probe(struct sock *sk) 24195d424d5aSJohn Heffner { 24205d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 242112a59abcSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 24225d424d5aSJohn Heffner struct sk_buff *skb, *nskb, *next; 24236b58e0a5SFan Du struct net *net = sock_net(sk); 24245d424d5aSJohn Heffner int probe_size; 242591cc17c0SIlpo Järvinen int size_needed; 242612a59abcSEric Dumazet int copy, len; 24275d424d5aSJohn Heffner int mss_now; 24286b58e0a5SFan Du int interval; 24295d424d5aSJohn Heffner 24305d424d5aSJohn Heffner /* Not currently probing/verifying, 24315d424d5aSJohn Heffner * not in recovery, 24325d424d5aSJohn Heffner * have enough cwnd, and 243312a59abcSEric Dumazet * not SACKing (the variable headers throw things off) 243412a59abcSEric Dumazet */ 243512a59abcSEric Dumazet if (likely(!icsk->icsk_mtup.enabled || 24365d424d5aSJohn Heffner icsk->icsk_mtup.probe_size || 24375d424d5aSJohn Heffner inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 243840570375SEric Dumazet tcp_snd_cwnd(tp) < 11 || 243912a59abcSEric Dumazet tp->rx_opt.num_sacks || tp->rx_opt.dsack)) 24405d424d5aSJohn Heffner return -1; 24415d424d5aSJohn Heffner 24426b58e0a5SFan Du /* Use binary search for probe_size between tcp_mss_base, 24436b58e0a5SFan Du * and current mss_clamp. if (search_high - search_low) 24446b58e0a5SFan Du * smaller than a threshold, backoff from probing. 24456b58e0a5SFan Du */ 24460c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 24476b58e0a5SFan Du probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high + 24486b58e0a5SFan Du icsk->icsk_mtup.search_low) >> 1); 244991cc17c0SIlpo Järvinen size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; 24506b58e0a5SFan Du interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low; 245105cbc0dbSFan Du /* When misfortune happens, we are reprobing actively, 245205cbc0dbSFan Du * and then reprobe timer has expired. We stick with current 245305cbc0dbSFan Du * probing process by not resetting search range to its orignal. 245405cbc0dbSFan Du */ 24556b58e0a5SFan Du if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) || 245692c0aa41SKuniyuki Iwashima interval < READ_ONCE(net->ipv4.sysctl_tcp_probe_threshold)) { 245705cbc0dbSFan Du /* Check whether enough time has elaplased for 245805cbc0dbSFan Du * another round of probing. 245905cbc0dbSFan Du */ 246005cbc0dbSFan Du tcp_mtu_check_reprobe(sk); 24615d424d5aSJohn Heffner return -1; 24625d424d5aSJohn Heffner } 24635d424d5aSJohn Heffner 24645d424d5aSJohn Heffner /* Have enough data in the send queue to probe? */ 24657f9c33e5SIlpo Järvinen if (tp->write_seq - tp->snd_nxt < size_needed) 24665d424d5aSJohn Heffner return -1; 24675d424d5aSJohn Heffner 246891cc17c0SIlpo Järvinen if (tp->snd_wnd < size_needed) 24695d424d5aSJohn Heffner return -1; 247090840defSIlpo Järvinen if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) 24715d424d5aSJohn Heffner return 0; 24725d424d5aSJohn Heffner 2473d67c58e9SIlpo Järvinen /* Do we need to wait to drain cwnd? With none in flight, don't stall */ 247440570375SEric Dumazet if (tcp_packets_in_flight(tp) + 2 > tcp_snd_cwnd(tp)) { 2475d67c58e9SIlpo Järvinen if (!tcp_packets_in_flight(tp)) 24765d424d5aSJohn Heffner return -1; 24775d424d5aSJohn Heffner else 24785d424d5aSJohn Heffner return 0; 24795d424d5aSJohn Heffner } 24805d424d5aSJohn Heffner 2481808cf9e3SIlya Lesokhin if (!tcp_can_coalesce_send_queue_head(sk, probe_size)) 2482808cf9e3SIlya Lesokhin return -1; 2483808cf9e3SIlya Lesokhin 24845d424d5aSJohn Heffner /* We're allowed to probe. Build it now. */ 24855882efffSEric Dumazet nskb = tcp_stream_alloc_skb(sk, GFP_ATOMIC, false); 248651456b29SIan Morris if (!nskb) 24875d424d5aSJohn Heffner return -1; 248873601329SEric Dumazet 248973601329SEric Dumazet /* build the payload, and be prepared to abort if this fails. */ 249073601329SEric Dumazet if (tcp_clone_payload(sk, nskb, probe_size)) { 249171c299c7SJakub Kicinski tcp_skb_tsorted_anchor_cleanup(nskb); 249273601329SEric Dumazet consume_skb(nskb); 249373601329SEric Dumazet return -1; 249473601329SEric Dumazet } 2495ab4e846aSEric Dumazet sk_wmem_queued_add(sk, nskb->truesize); 24963ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 24975d424d5aSJohn Heffner 2498fe067e8aSDavid S. Miller skb = tcp_send_head(sk); 249941477662SJakub Kicinski skb_copy_decrypted(nskb, skb); 25005a369ca6SPaolo Abeni mptcp_skb_ext_copy(nskb, skb); 25015d424d5aSJohn Heffner 25025d424d5aSJohn Heffner TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 25035d424d5aSJohn Heffner TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 25044de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; 25055d424d5aSJohn Heffner 250650c4817eSIlpo Järvinen tcp_insert_write_queue_before(nskb, skb, sk); 25072b7cda9cSEric Dumazet tcp_highest_sack_replace(sk, skb, nskb); 250850c4817eSIlpo Järvinen 25095d424d5aSJohn Heffner len = 0; 2510234b6860SIlpo Järvinen tcp_for_write_queue_from_safe(skb, next, sk) { 25115d424d5aSJohn Heffner copy = min_t(int, skb->len, probe_size - len); 25125d424d5aSJohn Heffner 25135d424d5aSJohn Heffner if (skb->len <= copy) { 25145d424d5aSJohn Heffner /* We've eaten all the data from this skb. 25155d424d5aSJohn Heffner * Throw it away. */ 25164de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 2517808cf9e3SIlya Lesokhin /* If this is the last SKB we copy and eor is set 2518808cf9e3SIlya Lesokhin * we need to propagate it to the new skb. 2519808cf9e3SIlya Lesokhin */ 2520808cf9e3SIlya Lesokhin TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor; 2521888a5c53SWillem de Bruijn tcp_skb_collapse_tstamp(nskb, skb); 2522fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 252303271f3aSTalal Ahmad tcp_wmem_free_skb(sk, skb); 25245d424d5aSJohn Heffner } else { 25254de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & 2526a3433f35SChangli Gao ~(TCPHDR_FIN|TCPHDR_PSH); 25275d424d5aSJohn Heffner __pskb_trim_head(skb, copy); 25285bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 25295d424d5aSJohn Heffner TCP_SKB_CB(skb)->seq += copy; 25305d424d5aSJohn Heffner } 25315d424d5aSJohn Heffner 25325d424d5aSJohn Heffner len += copy; 2533234b6860SIlpo Järvinen 2534234b6860SIlpo Järvinen if (len >= probe_size) 2535234b6860SIlpo Järvinen break; 25365d424d5aSJohn Heffner } 25375bbb432cSEric Dumazet tcp_init_tso_segs(nskb, nskb->len); 25385d424d5aSJohn Heffner 25395d424d5aSJohn Heffner /* We're ready to send. If this fails, the probe will 25407faee5c0SEric Dumazet * be resegmented into mss-sized pieces by tcp_write_xmit(). 25417faee5c0SEric Dumazet */ 25425d424d5aSJohn Heffner if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 25435d424d5aSJohn Heffner /* Decrement cwnd here because we are sending 25445d424d5aSJohn Heffner * effectively two packets. */ 254540570375SEric Dumazet tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) - 1); 254666f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, nskb); 25475d424d5aSJohn Heffner 25485d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 25490e7b1368SJohn Heffner tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 25500e7b1368SJohn Heffner tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 25515d424d5aSJohn Heffner 25525d424d5aSJohn Heffner return 1; 25535d424d5aSJohn Heffner } 25545d424d5aSJohn Heffner 25555d424d5aSJohn Heffner return -1; 25565d424d5aSJohn Heffner } 25575d424d5aSJohn Heffner 2558864e5c09SEric Dumazet static bool tcp_pacing_check(struct sock *sk) 2559218af599SEric Dumazet { 2560864e5c09SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 2561864e5c09SEric Dumazet 2562864e5c09SEric Dumazet if (!tcp_needs_internal_pacing(sk)) 2563864e5c09SEric Dumazet return false; 2564864e5c09SEric Dumazet 2565864e5c09SEric Dumazet if (tp->tcp_wstamp_ns <= tp->tcp_clock_cache) 2566864e5c09SEric Dumazet return false; 2567864e5c09SEric Dumazet 2568864e5c09SEric Dumazet if (!hrtimer_is_queued(&tp->pacing_timer)) { 2569864e5c09SEric Dumazet hrtimer_start(&tp->pacing_timer, 2570864e5c09SEric Dumazet ns_to_ktime(tp->tcp_wstamp_ns), 2571864e5c09SEric Dumazet HRTIMER_MODE_ABS_PINNED_SOFT); 2572864e5c09SEric Dumazet sock_hold(sk); 2573864e5c09SEric Dumazet } 2574864e5c09SEric Dumazet return true; 2575218af599SEric Dumazet } 2576218af599SEric Dumazet 2577f921a4a5SEric Dumazet static bool tcp_rtx_queue_empty_or_single_skb(const struct sock *sk) 2578f921a4a5SEric Dumazet { 2579f921a4a5SEric Dumazet const struct rb_node *node = sk->tcp_rtx_queue.rb_node; 2580f921a4a5SEric Dumazet 2581f921a4a5SEric Dumazet /* No skb in the rtx queue. */ 2582f921a4a5SEric Dumazet if (!node) 2583f921a4a5SEric Dumazet return true; 2584f921a4a5SEric Dumazet 2585f921a4a5SEric Dumazet /* Only one skb in rtx queue. */ 2586f921a4a5SEric Dumazet return !node->rb_left && !node->rb_right; 2587f921a4a5SEric Dumazet } 2588f921a4a5SEric Dumazet 2589f9616c35SEric Dumazet /* TCP Small Queues : 2590f9616c35SEric Dumazet * Control number of packets in qdisc/devices to two packets / or ~1 ms. 2591f9616c35SEric Dumazet * (These limits are doubled for retransmits) 2592f9616c35SEric Dumazet * This allows for : 2593f9616c35SEric Dumazet * - better RTT estimation and ACK scheduling 2594f9616c35SEric Dumazet * - faster recovery 2595f9616c35SEric Dumazet * - high rates 2596f9616c35SEric Dumazet * Alas, some drivers / subsystems require a fair amount 2597f9616c35SEric Dumazet * of queued bytes to ensure line rate. 2598f9616c35SEric Dumazet * One example is wifi aggregation (802.11 AMPDU) 2599f9616c35SEric Dumazet */ 2600f9616c35SEric Dumazet static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb, 2601f9616c35SEric Dumazet unsigned int factor) 2602f9616c35SEric Dumazet { 260376a9ebe8SEric Dumazet unsigned long limit; 2604f9616c35SEric Dumazet 260576a9ebe8SEric Dumazet limit = max_t(unsigned long, 260676a9ebe8SEric Dumazet 2 * skb->truesize, 260728b24f90SEric Dumazet READ_ONCE(sk->sk_pacing_rate) >> READ_ONCE(sk->sk_pacing_shift)); 2608c73e5807SEric Dumazet if (sk->sk_pacing_status == SK_PACING_NONE) 260976a9ebe8SEric Dumazet limit = min_t(unsigned long, limit, 26109fb90193SKuniyuki Iwashima READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes)); 2611f9616c35SEric Dumazet limit <<= factor; 2612f9616c35SEric Dumazet 2613a842fe14SEric Dumazet if (static_branch_unlikely(&tcp_tx_delay_enabled) && 2614a842fe14SEric Dumazet tcp_sk(sk)->tcp_tx_delay) { 261528b24f90SEric Dumazet u64 extra_bytes = (u64)READ_ONCE(sk->sk_pacing_rate) * 261628b24f90SEric Dumazet tcp_sk(sk)->tcp_tx_delay; 2617a842fe14SEric Dumazet 2618a842fe14SEric Dumazet /* TSQ is based on skb truesize sum (sk_wmem_alloc), so we 2619a842fe14SEric Dumazet * approximate our needs assuming an ~100% skb->truesize overhead. 2620a842fe14SEric Dumazet * USEC_PER_SEC is approximated by 2^20. 2621a842fe14SEric Dumazet * do_div(extra_bytes, USEC_PER_SEC/2) is replaced by a right shift. 2622a842fe14SEric Dumazet */ 2623a842fe14SEric Dumazet extra_bytes >>= (20 - 1); 2624a842fe14SEric Dumazet limit += extra_bytes; 2625a842fe14SEric Dumazet } 262614afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) > limit) { 2627f921a4a5SEric Dumazet /* Always send skb if rtx queue is empty or has one skb. 262875eefc6cSEric Dumazet * No need to wait for TX completion to call us back, 262975eefc6cSEric Dumazet * after softirq/tasklet schedule. 263075eefc6cSEric Dumazet * This helps when TX completions are delayed too much. 263175eefc6cSEric Dumazet */ 2632f921a4a5SEric Dumazet if (tcp_rtx_queue_empty_or_single_skb(sk)) 263375eefc6cSEric Dumazet return false; 263475eefc6cSEric Dumazet 26357aa5470cSEric Dumazet set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags); 2636f9616c35SEric Dumazet /* It is possible TX completion already happened 2637f9616c35SEric Dumazet * before we set TSQ_THROTTLED, so we must 2638f9616c35SEric Dumazet * test again the condition. 2639f9616c35SEric Dumazet */ 2640f9616c35SEric Dumazet smp_mb__after_atomic(); 2641ce8299b6SEric Dumazet if (refcount_read(&sk->sk_wmem_alloc) > limit) 2642f9616c35SEric Dumazet return true; 2643f9616c35SEric Dumazet } 2644f9616c35SEric Dumazet return false; 2645f9616c35SEric Dumazet } 2646f9616c35SEric Dumazet 264705b055e8SFrancis Yan static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new) 264805b055e8SFrancis Yan { 2649628174ccSEric Dumazet const u32 now = tcp_jiffies32; 2650efe967cdSArnd Bergmann enum tcp_chrono old = tp->chrono_type; 265105b055e8SFrancis Yan 2652efe967cdSArnd Bergmann if (old > TCP_CHRONO_UNSPEC) 2653efe967cdSArnd Bergmann tp->chrono_stat[old - 1] += now - tp->chrono_start; 265405b055e8SFrancis Yan tp->chrono_start = now; 265505b055e8SFrancis Yan tp->chrono_type = new; 265605b055e8SFrancis Yan } 265705b055e8SFrancis Yan 265805b055e8SFrancis Yan void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type) 265905b055e8SFrancis Yan { 266005b055e8SFrancis Yan struct tcp_sock *tp = tcp_sk(sk); 266105b055e8SFrancis Yan 266205b055e8SFrancis Yan /* If there are multiple conditions worthy of tracking in a 26630f87230dSFrancis Yan * chronograph then the highest priority enum takes precedence 26640f87230dSFrancis Yan * over the other conditions. So that if something "more interesting" 266505b055e8SFrancis Yan * starts happening, stop the previous chrono and start a new one. 266605b055e8SFrancis Yan */ 266705b055e8SFrancis Yan if (type > tp->chrono_type) 266805b055e8SFrancis Yan tcp_chrono_set(tp, type); 266905b055e8SFrancis Yan } 267005b055e8SFrancis Yan 267105b055e8SFrancis Yan void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type) 267205b055e8SFrancis Yan { 267305b055e8SFrancis Yan struct tcp_sock *tp = tcp_sk(sk); 267405b055e8SFrancis Yan 26750f87230dSFrancis Yan 26760f87230dSFrancis Yan /* There are multiple conditions worthy of tracking in a 26770f87230dSFrancis Yan * chronograph, so that the highest priority enum takes 26780f87230dSFrancis Yan * precedence over the other conditions (see tcp_chrono_start). 26790f87230dSFrancis Yan * If a condition stops, we only stop chrono tracking if 26800f87230dSFrancis Yan * it's the "most interesting" or current chrono we are 26810f87230dSFrancis Yan * tracking and starts busy chrono if we have pending data. 26820f87230dSFrancis Yan */ 268375c119afSEric Dumazet if (tcp_rtx_and_write_queues_empty(sk)) 268405b055e8SFrancis Yan tcp_chrono_set(tp, TCP_CHRONO_UNSPEC); 26850f87230dSFrancis Yan else if (type == tp->chrono_type) 26860f87230dSFrancis Yan tcp_chrono_set(tp, TCP_CHRONO_BUSY); 268705b055e8SFrancis Yan } 268805b055e8SFrancis Yan 26891da177e4SLinus Torvalds /* This routine writes packets to the network. It advances the 26901da177e4SLinus Torvalds * send_head. This happens as incoming acks open up the remote 26911da177e4SLinus Torvalds * window for us. 26921da177e4SLinus Torvalds * 2693f8269a49SIlpo Järvinen * LARGESEND note: !tcp_urg_mode is overkill, only frames between 2694f8269a49SIlpo Järvinen * snd_up-64k-mss .. snd_up cannot be large. However, taking into 2695f8269a49SIlpo Järvinen * account rare use of URG, this is not a big flaw. 2696f8269a49SIlpo Järvinen * 26976ba8a3b1SNandita Dukkipati * Send at most one packet when push_one > 0. Temporarily ignore 26986ba8a3b1SNandita Dukkipati * cwnd limit to force at most one packet out when push_one == 2. 26996ba8a3b1SNandita Dukkipati 2700a2a385d6SEric Dumazet * Returns true, if no segments are in flight and we have queued segments, 2701a2a385d6SEric Dumazet * but cannot send anything now because of SWS or another problem. 27021da177e4SLinus Torvalds */ 2703a2a385d6SEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 2704d5dd9175SIlpo Järvinen int push_one, gfp_t gfp) 27051da177e4SLinus Torvalds { 27061da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 270792df7b51SDavid S. Miller struct sk_buff *skb; 2708c1b4a7e6SDavid S. Miller unsigned int tso_segs, sent_pkts; 2709c1b4a7e6SDavid S. Miller int cwnd_quota; 27105d424d5aSJohn Heffner int result; 27115615f886SFrancis Yan bool is_cwnd_limited = false, is_rwnd_limited = false; 2712605ad7f1SEric Dumazet u32 max_segs; 27131da177e4SLinus Torvalds 2714c1b4a7e6SDavid S. Miller sent_pkts = 0; 27155d424d5aSJohn Heffner 2716ee1836aeSEric Dumazet tcp_mstamp_refresh(tp); 2717d5dd9175SIlpo Järvinen if (!push_one) { 27185d424d5aSJohn Heffner /* Do MTU probing. */ 2719d5dd9175SIlpo Järvinen result = tcp_mtu_probe(sk); 2720d5dd9175SIlpo Järvinen if (!result) { 2721a2a385d6SEric Dumazet return false; 27225d424d5aSJohn Heffner } else if (result > 0) { 27235d424d5aSJohn Heffner sent_pkts = 1; 27245d424d5aSJohn Heffner } 2725d5dd9175SIlpo Järvinen } 27265d424d5aSJohn Heffner 2727ed6e7268SNeal Cardwell max_segs = tcp_tso_segs(sk, mss_now); 2728fe067e8aSDavid S. Miller while ((skb = tcp_send_head(sk))) { 2729c8ac3774SHerbert Xu unsigned int limit; 2730c8ac3774SHerbert Xu 273179861919SEric Dumazet if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) { 273279861919SEric Dumazet /* "skb_mstamp_ns" is used as a start point for the retransmit timer */ 2733a1ac9c8aSMartin KaFai Lau tp->tcp_wstamp_ns = tp->tcp_clock_cache; 2734a1ac9c8aSMartin KaFai Lau skb_set_delivery_time(skb, tp->tcp_wstamp_ns, true); 273579861919SEric Dumazet list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue); 2736bf50b606SEric Dumazet tcp_init_tso_segs(skb, mss_now); 273779861919SEric Dumazet goto repair; /* Skip network transmission */ 273879861919SEric Dumazet } 273979861919SEric Dumazet 2740218af599SEric Dumazet if (tcp_pacing_check(sk)) 2741218af599SEric Dumazet break; 2742218af599SEric Dumazet 27435bbb432cSEric Dumazet tso_segs = tcp_init_tso_segs(skb, mss_now); 2744c1b4a7e6SDavid S. Miller BUG_ON(!tso_segs); 2745c1b4a7e6SDavid S. Miller 2746b68e9f85SHerbert Xu cwnd_quota = tcp_cwnd_test(tp, skb); 27476ba8a3b1SNandita Dukkipati if (!cwnd_quota) { 27486ba8a3b1SNandita Dukkipati if (push_one == 2) 27496ba8a3b1SNandita Dukkipati /* Force out a loss probe pkt. */ 27506ba8a3b1SNandita Dukkipati cwnd_quota = 1; 27516ba8a3b1SNandita Dukkipati else 2752b68e9f85SHerbert Xu break; 27536ba8a3b1SNandita Dukkipati } 2754b68e9f85SHerbert Xu 27555615f886SFrancis Yan if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) { 27565615f886SFrancis Yan is_rwnd_limited = true; 2757b68e9f85SHerbert Xu break; 27585615f886SFrancis Yan } 2759b68e9f85SHerbert Xu 2760d6a4e26aSEric Dumazet if (tso_segs == 1) { 2761aa93466bSDavid S. Miller if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 2762aa93466bSDavid S. Miller (tcp_skb_is_last(sk, skb) ? 2763aa93466bSDavid S. Miller nonagle : TCP_NAGLE_PUSH)))) 2764aa93466bSDavid S. Miller break; 2765c1b4a7e6SDavid S. Miller } else { 2766ca8a2263SNeal Cardwell if (!push_one && 2767605ad7f1SEric Dumazet tcp_tso_should_defer(sk, skb, &is_cwnd_limited, 2768f9bfe4e6SEric Dumazet &is_rwnd_limited, max_segs)) 2769aa93466bSDavid S. Miller break; 2770c1b4a7e6SDavid S. Miller } 2771aa93466bSDavid S. Miller 2772605ad7f1SEric Dumazet limit = mss_now; 2773d6a4e26aSEric Dumazet if (tso_segs > 1 && !tcp_urg_mode(tp)) 2774605ad7f1SEric Dumazet limit = tcp_mss_split_point(sk, skb, mss_now, 2775605ad7f1SEric Dumazet min_t(unsigned int, 2776605ad7f1SEric Dumazet cwnd_quota, 2777605ad7f1SEric Dumazet max_segs), 2778605ad7f1SEric Dumazet nonagle); 2779605ad7f1SEric Dumazet 2780605ad7f1SEric Dumazet if (skb->len > limit && 278156483341SEric Dumazet unlikely(tso_fragment(sk, skb, limit, mss_now, gfp))) 2782605ad7f1SEric Dumazet break; 2783605ad7f1SEric Dumazet 2784f9616c35SEric Dumazet if (tcp_small_queue_check(sk, skb, 0)) 278546d3ceabSEric Dumazet break; 2786c9eeec26SEric Dumazet 27871f85e626SEric Dumazet /* Argh, we hit an empty skb(), presumably a thread 27881f85e626SEric Dumazet * is sleeping in sendmsg()/sk_stream_wait_memory(). 27891f85e626SEric Dumazet * We do not want to send a pure-ack packet and have 27901f85e626SEric Dumazet * a strange looking rtx queue with empty packet(s). 27911f85e626SEric Dumazet */ 27921f85e626SEric Dumazet if (TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq) 27931f85e626SEric Dumazet break; 27941f85e626SEric Dumazet 2795d5dd9175SIlpo Järvinen if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) 27961da177e4SLinus Torvalds break; 27971da177e4SLinus Torvalds 2798ec342325SAndrew Vagin repair: 27991da177e4SLinus Torvalds /* Advance the send_head. This one is sent out. 28001da177e4SLinus Torvalds * This call will increment packets_out. 28011da177e4SLinus Torvalds */ 280266f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 28031da177e4SLinus Torvalds 28041da177e4SLinus Torvalds tcp_minshall_update(tp, mss_now, skb); 2805a262f0cdSNandita Dukkipati sent_pkts += tcp_skb_pcount(skb); 2806d5dd9175SIlpo Järvinen 2807d5dd9175SIlpo Järvinen if (push_one) 2808d5dd9175SIlpo Järvinen break; 28091da177e4SLinus Torvalds } 28101da177e4SLinus Torvalds 28115615f886SFrancis Yan if (is_rwnd_limited) 28125615f886SFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED); 28135615f886SFrancis Yan else 28145615f886SFrancis Yan tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED); 28155615f886SFrancis Yan 281640570375SEric Dumazet is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tcp_snd_cwnd(tp)); 2817299bcb55SNeal Cardwell if (likely(sent_pkts || is_cwnd_limited)) 2818299bcb55SNeal Cardwell tcp_cwnd_validate(sk, is_cwnd_limited); 2819299bcb55SNeal Cardwell 2820aa93466bSDavid S. Miller if (likely(sent_pkts)) { 2821684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 2822684bad11SYuchung Cheng tp->prr_out += sent_pkts; 28236ba8a3b1SNandita Dukkipati 28246ba8a3b1SNandita Dukkipati /* Send one loss probe per tail loss episode. */ 28256ba8a3b1SNandita Dukkipati if (push_one != 2) 2826ed66dfafSNeal Cardwell tcp_schedule_loss_probe(sk, false); 2827a2a385d6SEric Dumazet return false; 28281da177e4SLinus Torvalds } 282975c119afSEric Dumazet return !tp->packets_out && !tcp_write_queue_empty(sk); 28306ba8a3b1SNandita Dukkipati } 28316ba8a3b1SNandita Dukkipati 2832ed66dfafSNeal Cardwell bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto) 28336ba8a3b1SNandita Dukkipati { 28346ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 28356ba8a3b1SNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 28361c2709cfSNeal Cardwell u32 timeout, timeout_us, rto_delta_us; 28372ae21cf5SEric Dumazet int early_retrans; 28386ba8a3b1SNandita Dukkipati 28396ba8a3b1SNandita Dukkipati /* Don't do any loss probe on a Fast Open connection before 3WHS 28406ba8a3b1SNandita Dukkipati * finishes. 28416ba8a3b1SNandita Dukkipati */ 2842d983ea6fSEric Dumazet if (rcu_access_pointer(tp->fastopen_rsk)) 28436ba8a3b1SNandita Dukkipati return false; 28446ba8a3b1SNandita Dukkipati 284552e65865SKuniyuki Iwashima early_retrans = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_early_retrans); 28466ba8a3b1SNandita Dukkipati /* Schedule a loss probe in 2*RTT for SACK capable connections 2847b4f70c3dSNeal Cardwell * not in loss recovery, that are either limited by cwnd or application. 28486ba8a3b1SNandita Dukkipati */ 28492ae21cf5SEric Dumazet if ((early_retrans != 3 && early_retrans != 4) || 2850bec41a11SYuchung Cheng !tp->packets_out || !tcp_is_sack(tp) || 2851b4f70c3dSNeal Cardwell (icsk->icsk_ca_state != TCP_CA_Open && 2852b4f70c3dSNeal Cardwell icsk->icsk_ca_state != TCP_CA_CWR)) 28536ba8a3b1SNandita Dukkipati return false; 28546ba8a3b1SNandita Dukkipati 2855bb4d991aSYuchung Cheng /* Probe timeout is 2*rtt. Add minimum RTO to account 2856f9b99582SYuchung Cheng * for delayed ack when there's one outstanding packet. If no RTT 2857f9b99582SYuchung Cheng * sample is available then probe after TCP_TIMEOUT_INIT. 28586ba8a3b1SNandita Dukkipati */ 2859bb4d991aSYuchung Cheng if (tp->srtt_us) { 28601c2709cfSNeal Cardwell timeout_us = tp->srtt_us >> 2; 28616ba8a3b1SNandita Dukkipati if (tp->packets_out == 1) 28621c2709cfSNeal Cardwell timeout_us += tcp_rto_min_us(sk); 2863bb4d991aSYuchung Cheng else 28641c2709cfSNeal Cardwell timeout_us += TCP_TIMEOUT_MIN_US; 28651c2709cfSNeal Cardwell timeout = usecs_to_jiffies(timeout_us); 2866bb4d991aSYuchung Cheng } else { 2867bb4d991aSYuchung Cheng timeout = TCP_TIMEOUT_INIT; 2868bb4d991aSYuchung Cheng } 28696ba8a3b1SNandita Dukkipati 2870a2815817SNeal Cardwell /* If the RTO formula yields an earlier time, then use that time. */ 2871ed66dfafSNeal Cardwell rto_delta_us = advancing_rto ? 2872ed66dfafSNeal Cardwell jiffies_to_usecs(inet_csk(sk)->icsk_rto) : 2873ed66dfafSNeal Cardwell tcp_rto_delta_us(sk); /* How far in future is RTO? */ 2874a2815817SNeal Cardwell if (rto_delta_us > 0) 2875a2815817SNeal Cardwell timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us)); 28766ba8a3b1SNandita Dukkipati 28778dc242adSEric Dumazet tcp_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, TCP_RTO_MAX); 28786ba8a3b1SNandita Dukkipati return true; 28796ba8a3b1SNandita Dukkipati } 28806ba8a3b1SNandita Dukkipati 28811f3279aeSEric Dumazet /* Thanks to skb fast clones, we can detect if a prior transmit of 28821f3279aeSEric Dumazet * a packet is still in a qdisc or driver queue. 28831f3279aeSEric Dumazet * In this case, there is very little point doing a retransmit ! 28841f3279aeSEric Dumazet */ 2885f4dae54eSEric Dumazet static bool skb_still_in_host_queue(struct sock *sk, 28861f3279aeSEric Dumazet const struct sk_buff *skb) 28871f3279aeSEric Dumazet { 288839bb5e62SEric Dumazet if (unlikely(skb_fclone_busy(sk, skb))) { 2889f4dae54eSEric Dumazet set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags); 2890f4dae54eSEric Dumazet smp_mb__after_atomic(); 2891f4dae54eSEric Dumazet if (skb_fclone_busy(sk, skb)) { 2892c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), 28931f3279aeSEric Dumazet LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES); 28941f3279aeSEric Dumazet return true; 28951f3279aeSEric Dumazet } 2896f4dae54eSEric Dumazet } 28971f3279aeSEric Dumazet return false; 28981f3279aeSEric Dumazet } 28991f3279aeSEric Dumazet 2900b340b264SYuchung Cheng /* When probe timeout (PTO) fires, try send a new segment if possible, else 29016ba8a3b1SNandita Dukkipati * retransmit the last segment. 29026ba8a3b1SNandita Dukkipati */ 29036ba8a3b1SNandita Dukkipati void tcp_send_loss_probe(struct sock *sk) 29046ba8a3b1SNandita Dukkipati { 29059b717a8dSNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 29066ba8a3b1SNandita Dukkipati struct sk_buff *skb; 29076ba8a3b1SNandita Dukkipati int pcount; 29086ba8a3b1SNandita Dukkipati int mss = tcp_current_mss(sk); 29096ba8a3b1SNandita Dukkipati 291076be93fcSYuchung Cheng /* At most one outstanding TLP */ 291176be93fcSYuchung Cheng if (tp->tlp_high_seq) 291276be93fcSYuchung Cheng goto rearm_timer; 291376be93fcSYuchung Cheng 291476be93fcSYuchung Cheng tp->tlp_retrans = 0; 2915b340b264SYuchung Cheng skb = tcp_send_head(sk); 291675c119afSEric Dumazet if (skb && tcp_snd_wnd_test(tp, skb, mss)) { 2917b340b264SYuchung Cheng pcount = tp->packets_out; 2918b340b264SYuchung Cheng tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC); 2919b340b264SYuchung Cheng if (tp->packets_out > pcount) 2920b340b264SYuchung Cheng goto probe_sent; 29216ba8a3b1SNandita Dukkipati goto rearm_timer; 29226ba8a3b1SNandita Dukkipati } 292375c119afSEric Dumazet skb = skb_rb_last(&sk->tcp_rtx_queue); 2924b2b7af86SYuchung Cheng if (unlikely(!skb)) { 2925b2b7af86SYuchung Cheng WARN_ONCE(tp->packets_out, 2926b2b7af86SYuchung Cheng "invalid inflight: %u state %u cwnd %u mss %d\n", 292740570375SEric Dumazet tp->packets_out, sk->sk_state, tcp_snd_cwnd(tp), mss); 2928b2b7af86SYuchung Cheng inet_csk(sk)->icsk_pending = 0; 2929b2b7af86SYuchung Cheng return; 2930b2b7af86SYuchung Cheng } 29316ba8a3b1SNandita Dukkipati 29321f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 29331f3279aeSEric Dumazet goto rearm_timer; 29341f3279aeSEric Dumazet 29356ba8a3b1SNandita Dukkipati pcount = tcp_skb_pcount(skb); 29366ba8a3b1SNandita Dukkipati if (WARN_ON(!pcount)) 29376ba8a3b1SNandita Dukkipati goto rearm_timer; 29386ba8a3b1SNandita Dukkipati 29396ba8a3b1SNandita Dukkipati if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) { 294075c119afSEric Dumazet if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, 294175c119afSEric Dumazet (pcount - 1) * mss, mss, 29426cc55e09SOctavian Purdila GFP_ATOMIC))) 29436ba8a3b1SNandita Dukkipati goto rearm_timer; 294475c119afSEric Dumazet skb = skb_rb_next(skb); 29456ba8a3b1SNandita Dukkipati } 29466ba8a3b1SNandita Dukkipati 29476ba8a3b1SNandita Dukkipati if (WARN_ON(!skb || !tcp_skb_pcount(skb))) 29486ba8a3b1SNandita Dukkipati goto rearm_timer; 29496ba8a3b1SNandita Dukkipati 295010d3be56SEric Dumazet if (__tcp_retransmit_skb(sk, skb, 1)) 2951b340b264SYuchung Cheng goto rearm_timer; 29526ba8a3b1SNandita Dukkipati 295376be93fcSYuchung Cheng tp->tlp_retrans = 1; 295476be93fcSYuchung Cheng 295576be93fcSYuchung Cheng probe_sent: 29569b717a8dSNandita Dukkipati /* Record snd_nxt for loss detection. */ 29579b717a8dSNandita Dukkipati tp->tlp_high_seq = tp->snd_nxt; 29589b717a8dSNandita Dukkipati 2959c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES); 2960fcd16c0aSYuchung Cheng /* Reset s.t. tcp_rearm_rto will restart timer from now */ 2961fcd16c0aSYuchung Cheng inet_csk(sk)->icsk_pending = 0; 2962b340b264SYuchung Cheng rearm_timer: 2963fcd16c0aSYuchung Cheng tcp_rearm_rto(sk); 29641da177e4SLinus Torvalds } 29651da177e4SLinus Torvalds 2966a762a980SDavid S. Miller /* Push out any pending frames which were held back due to 2967a762a980SDavid S. Miller * TCP_CORK or attempt at coalescing tiny packets. 2968a762a980SDavid S. Miller * The socket must be locked by the caller. 2969a762a980SDavid S. Miller */ 29709e412ba7SIlpo Järvinen void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 29719e412ba7SIlpo Järvinen int nonagle) 2972a762a980SDavid S. Miller { 2973726e07a8SIlpo Järvinen /* If we are closed, the bytes will have to remain here. 2974726e07a8SIlpo Järvinen * In time closedown will finish, we empty the write queue and 2975726e07a8SIlpo Järvinen * all will be happy. 2976726e07a8SIlpo Järvinen */ 2977726e07a8SIlpo Järvinen if (unlikely(sk->sk_state == TCP_CLOSE)) 2978726e07a8SIlpo Järvinen return; 2979726e07a8SIlpo Järvinen 298099a1dec7SMel Gorman if (tcp_write_xmit(sk, cur_mss, nonagle, 0, 29817450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC))) 29829e412ba7SIlpo Järvinen tcp_check_probe_timer(sk); 2983a762a980SDavid S. Miller } 2984a762a980SDavid S. Miller 2985c1b4a7e6SDavid S. Miller /* Send _single_ skb sitting at the send head. This function requires 2986c1b4a7e6SDavid S. Miller * true push pending frames to setup probe timer etc. 2987c1b4a7e6SDavid S. Miller */ 2988c1b4a7e6SDavid S. Miller void tcp_push_one(struct sock *sk, unsigned int mss_now) 2989c1b4a7e6SDavid S. Miller { 2990fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 2991c1b4a7e6SDavid S. Miller 2992c1b4a7e6SDavid S. Miller BUG_ON(!skb || skb->len < mss_now); 2993c1b4a7e6SDavid S. Miller 2994d5dd9175SIlpo Järvinen tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); 2995c1b4a7e6SDavid S. Miller } 2996c1b4a7e6SDavid S. Miller 29971da177e4SLinus Torvalds /* This function returns the amount that we can raise the 29981da177e4SLinus Torvalds * usable window based on the following constraints 29991da177e4SLinus Torvalds * 30001da177e4SLinus Torvalds * 1. The window can never be shrunk once it is offered (RFC 793) 30011da177e4SLinus Torvalds * 2. We limit memory per socket 30021da177e4SLinus Torvalds * 30031da177e4SLinus Torvalds * RFC 1122: 30041da177e4SLinus Torvalds * "the suggested [SWS] avoidance algorithm for the receiver is to keep 30051da177e4SLinus Torvalds * RECV.NEXT + RCV.WIN fixed until: 30061da177e4SLinus Torvalds * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 30071da177e4SLinus Torvalds * 30081da177e4SLinus Torvalds * i.e. don't raise the right edge of the window until you can raise 30091da177e4SLinus Torvalds * it at least MSS bytes. 30101da177e4SLinus Torvalds * 30111da177e4SLinus Torvalds * Unfortunately, the recommended algorithm breaks header prediction, 30121da177e4SLinus Torvalds * since header prediction assumes th->window stays fixed. 30131da177e4SLinus Torvalds * 30141da177e4SLinus Torvalds * Strictly speaking, keeping th->window fixed violates the receiver 30151da177e4SLinus Torvalds * side SWS prevention criteria. The problem is that under this rule 30161da177e4SLinus Torvalds * a stream of single byte packets will cause the right side of the 30171da177e4SLinus Torvalds * window to always advance by a single byte. 30181da177e4SLinus Torvalds * 30191da177e4SLinus Torvalds * Of course, if the sender implements sender side SWS prevention 30201da177e4SLinus Torvalds * then this will not be a problem. 30211da177e4SLinus Torvalds * 30221da177e4SLinus Torvalds * BSD seems to make the following compromise: 30231da177e4SLinus Torvalds * 30241da177e4SLinus Torvalds * If the free space is less than the 1/4 of the maximum 30251da177e4SLinus Torvalds * space available and the free space is less than 1/2 mss, 30261da177e4SLinus Torvalds * then set the window to 0. 30271da177e4SLinus Torvalds * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 30281da177e4SLinus Torvalds * Otherwise, just prevent the window from shrinking 30291da177e4SLinus Torvalds * and from being larger than the largest representable value. 30301da177e4SLinus Torvalds * 30311da177e4SLinus Torvalds * This prevents incremental opening of the window in the regime 30321da177e4SLinus Torvalds * where TCP is limited by the speed of the reader side taking 30331da177e4SLinus Torvalds * data out of the TCP receive queue. It does nothing about 30341da177e4SLinus Torvalds * those cases where the window is constrained on the sender side 30351da177e4SLinus Torvalds * because the pipeline is full. 30361da177e4SLinus Torvalds * 30371da177e4SLinus Torvalds * BSD also seems to "accidentally" limit itself to windows that are a 30381da177e4SLinus Torvalds * multiple of MSS, at least until the free space gets quite small. 30391da177e4SLinus Torvalds * This would appear to be a side effect of the mbuf implementation. 30401da177e4SLinus Torvalds * Combining these two algorithms results in the observed behavior 30411da177e4SLinus Torvalds * of having a fixed window size at almost all times. 30421da177e4SLinus Torvalds * 30431da177e4SLinus Torvalds * Below we obtain similar behavior by forcing the offered window to 30441da177e4SLinus Torvalds * a multiple of the mss when it is feasible to do so. 30451da177e4SLinus Torvalds * 30461da177e4SLinus Torvalds * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 30471da177e4SLinus Torvalds * Regular options like TIMESTAMP are taken into account. 30481da177e4SLinus Torvalds */ 30491da177e4SLinus Torvalds u32 __tcp_select_window(struct sock *sk) 30501da177e4SLinus Torvalds { 3051463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 30521da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 3053b650d953Smfreemon@cloudflare.com struct net *net = sock_net(sk); 3054caa20d9aSStephen Hemminger /* MSS for the peer's data. Previous versions used mss_clamp 30551da177e4SLinus Torvalds * here. I don't know if the value based on our guesses 30561da177e4SLinus Torvalds * of peer's MSS is better for the performance. It's more correct 30571da177e4SLinus Torvalds * but may be worse for the performance because of rcv_mss 30581da177e4SLinus Torvalds * fluctuations. --SAW 1998/11/1 30591da177e4SLinus Torvalds */ 3060463c84b9SArnaldo Carvalho de Melo int mss = icsk->icsk_ack.rcv_mss; 30611da177e4SLinus Torvalds int free_space = tcp_space(sk); 306286c1a045SFlorian Westphal int allowed_space = tcp_full_space(sk); 3063071c8ed6SFlorian Westphal int full_space, window; 3064071c8ed6SFlorian Westphal 3065071c8ed6SFlorian Westphal if (sk_is_mptcp(sk)) 3066071c8ed6SFlorian Westphal mptcp_space(sk, &free_space, &allowed_space); 3067071c8ed6SFlorian Westphal 3068071c8ed6SFlorian Westphal full_space = min_t(int, tp->window_clamp, allowed_space); 30691da177e4SLinus Torvalds 307006425c30SEric Dumazet if (unlikely(mss > full_space)) { 30711da177e4SLinus Torvalds mss = full_space; 307206425c30SEric Dumazet if (mss <= 0) 307306425c30SEric Dumazet return 0; 307406425c30SEric Dumazet } 3075b650d953Smfreemon@cloudflare.com 3076b650d953Smfreemon@cloudflare.com /* Only allow window shrink if the sysctl is enabled and we have 3077b650d953Smfreemon@cloudflare.com * a non-zero scaling factor in effect. 3078b650d953Smfreemon@cloudflare.com */ 3079b650d953Smfreemon@cloudflare.com if (READ_ONCE(net->ipv4.sysctl_tcp_shrink_window) && tp->rx_opt.rcv_wscale) 3080b650d953Smfreemon@cloudflare.com goto shrink_window_allowed; 3081b650d953Smfreemon@cloudflare.com 3082b650d953Smfreemon@cloudflare.com /* do not allow window to shrink */ 3083b650d953Smfreemon@cloudflare.com 3084b92edbe0SEric Dumazet if (free_space < (full_space >> 1)) { 3085463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.quick = 0; 30861da177e4SLinus Torvalds 3087b8da51ebSEric Dumazet if (tcp_under_memory_pressure(sk)) 3088053f3684SWei Wang tcp_adjust_rcv_ssthresh(sk); 30891da177e4SLinus Torvalds 309086c1a045SFlorian Westphal /* free_space might become our new window, make sure we don't 309186c1a045SFlorian Westphal * increase it due to wscale. 309286c1a045SFlorian Westphal */ 309386c1a045SFlorian Westphal free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale); 309486c1a045SFlorian Westphal 309586c1a045SFlorian Westphal /* if free space is less than mss estimate, or is below 1/16th 309686c1a045SFlorian Westphal * of the maximum allowed, try to move to zero-window, else 309786c1a045SFlorian Westphal * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and 309886c1a045SFlorian Westphal * new incoming data is dropped due to memory limits. 309986c1a045SFlorian Westphal * With large window, mss test triggers way too late in order 310086c1a045SFlorian Westphal * to announce zero window in time before rmem limit kicks in. 310186c1a045SFlorian Westphal */ 310286c1a045SFlorian Westphal if (free_space < (allowed_space >> 4) || free_space < mss) 31031da177e4SLinus Torvalds return 0; 31041da177e4SLinus Torvalds } 31051da177e4SLinus Torvalds 31061da177e4SLinus Torvalds if (free_space > tp->rcv_ssthresh) 31071da177e4SLinus Torvalds free_space = tp->rcv_ssthresh; 31081da177e4SLinus Torvalds 31091da177e4SLinus Torvalds /* Don't do rounding if we are using window scaling, since the 31101da177e4SLinus Torvalds * scaled window will not line up with the MSS boundary anyway. 31111da177e4SLinus Torvalds */ 31121da177e4SLinus Torvalds if (tp->rx_opt.rcv_wscale) { 31131da177e4SLinus Torvalds window = free_space; 31141da177e4SLinus Torvalds 31151da177e4SLinus Torvalds /* Advertise enough space so that it won't get scaled away. 31161da177e4SLinus Torvalds * Import case: prevent zero window announcement if 31171da177e4SLinus Torvalds * 1<<rcv_wscale > mss. 31181da177e4SLinus Torvalds */ 31191935299dSGao Feng window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale)); 31201da177e4SLinus Torvalds } else { 31211935299dSGao Feng window = tp->rcv_wnd; 31221da177e4SLinus Torvalds /* Get the largest window that is a nice multiple of mss. 31231da177e4SLinus Torvalds * Window clamp already applied above. 31241da177e4SLinus Torvalds * If our current window offering is within 1 mss of the 31251da177e4SLinus Torvalds * free space we just keep it. This prevents the divide 31261da177e4SLinus Torvalds * and multiply from happening most of the time. 31271da177e4SLinus Torvalds * We also don't do any window rounding when the free space 31281da177e4SLinus Torvalds * is too small. 31291da177e4SLinus Torvalds */ 31301da177e4SLinus Torvalds if (window <= free_space - mss || window > free_space) 31311935299dSGao Feng window = rounddown(free_space, mss); 313284565070SJohn Heffner else if (mss == full_space && 3133b92edbe0SEric Dumazet free_space > window + (full_space >> 1)) 313484565070SJohn Heffner window = free_space; 31351da177e4SLinus Torvalds } 31361da177e4SLinus Torvalds 31371da177e4SLinus Torvalds return window; 3138b650d953Smfreemon@cloudflare.com 3139b650d953Smfreemon@cloudflare.com shrink_window_allowed: 3140b650d953Smfreemon@cloudflare.com /* new window should always be an exact multiple of scaling factor */ 3141b650d953Smfreemon@cloudflare.com free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale); 3142b650d953Smfreemon@cloudflare.com 3143b650d953Smfreemon@cloudflare.com if (free_space < (full_space >> 1)) { 3144b650d953Smfreemon@cloudflare.com icsk->icsk_ack.quick = 0; 3145b650d953Smfreemon@cloudflare.com 3146b650d953Smfreemon@cloudflare.com if (tcp_under_memory_pressure(sk)) 3147b650d953Smfreemon@cloudflare.com tcp_adjust_rcv_ssthresh(sk); 3148b650d953Smfreemon@cloudflare.com 3149b650d953Smfreemon@cloudflare.com /* if free space is too low, return a zero window */ 3150b650d953Smfreemon@cloudflare.com if (free_space < (allowed_space >> 4) || free_space < mss || 3151b650d953Smfreemon@cloudflare.com free_space < (1 << tp->rx_opt.rcv_wscale)) 3152b650d953Smfreemon@cloudflare.com return 0; 3153b650d953Smfreemon@cloudflare.com } 3154b650d953Smfreemon@cloudflare.com 3155b650d953Smfreemon@cloudflare.com if (free_space > tp->rcv_ssthresh) { 3156b650d953Smfreemon@cloudflare.com free_space = tp->rcv_ssthresh; 3157b650d953Smfreemon@cloudflare.com /* new window should always be an exact multiple of scaling factor 3158b650d953Smfreemon@cloudflare.com * 3159b650d953Smfreemon@cloudflare.com * For this case, we ALIGN "up" (increase free_space) because 3160b650d953Smfreemon@cloudflare.com * we know free_space is not zero here, it has been reduced from 3161b650d953Smfreemon@cloudflare.com * the memory-based limit, and rcv_ssthresh is not a hard limit 3162b650d953Smfreemon@cloudflare.com * (unlike sk_rcvbuf). 3163b650d953Smfreemon@cloudflare.com */ 3164b650d953Smfreemon@cloudflare.com free_space = ALIGN(free_space, (1 << tp->rx_opt.rcv_wscale)); 3165b650d953Smfreemon@cloudflare.com } 3166b650d953Smfreemon@cloudflare.com 3167b650d953Smfreemon@cloudflare.com return free_space; 31681da177e4SLinus Torvalds } 31691da177e4SLinus Torvalds 3170cfea5a68SMartin KaFai Lau void tcp_skb_collapse_tstamp(struct sk_buff *skb, 3171082ac2d5SMartin KaFai Lau const struct sk_buff *next_skb) 3172082ac2d5SMartin KaFai Lau { 31730a2cf20cSSoheil Hassas Yeganeh if (unlikely(tcp_has_tx_tstamp(next_skb))) { 31740a2cf20cSSoheil Hassas Yeganeh const struct skb_shared_info *next_shinfo = 31750a2cf20cSSoheil Hassas Yeganeh skb_shinfo(next_skb); 3176082ac2d5SMartin KaFai Lau struct skb_shared_info *shinfo = skb_shinfo(skb); 3177082ac2d5SMartin KaFai Lau 31780a2cf20cSSoheil Hassas Yeganeh shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP; 3179082ac2d5SMartin KaFai Lau shinfo->tskey = next_shinfo->tskey; 31802de8023eSMartin KaFai Lau TCP_SKB_CB(skb)->txstamp_ack |= 31812de8023eSMartin KaFai Lau TCP_SKB_CB(next_skb)->txstamp_ack; 3182082ac2d5SMartin KaFai Lau } 3183082ac2d5SMartin KaFai Lau } 3184082ac2d5SMartin KaFai Lau 31854a17fc3aSIlpo Järvinen /* Collapses two adjacent SKB's during retransmission. */ 3186f8071cdeSEric Dumazet static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) 31871da177e4SLinus Torvalds { 31881da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 318975c119afSEric Dumazet struct sk_buff *next_skb = skb_rb_next(skb); 319013dde04fSWei Yongjun int next_skb_size; 31911da177e4SLinus Torvalds 3192058dc334SIlpo Järvinen next_skb_size = next_skb->len; 31931da177e4SLinus Torvalds 3194058dc334SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); 31951da177e4SLinus Torvalds 3196bd446314SEric Dumazet if (next_skb_size && !tcp_skb_shift(skb, next_skb, 1, next_skb_size)) 3197f8071cdeSEric Dumazet return false; 3198bd446314SEric Dumazet 31992b7cda9cSEric Dumazet tcp_highest_sack_replace(sk, next_skb, skb); 3200a6963a6bSIlpo Järvinen 32011da177e4SLinus Torvalds /* Update sequence range on original skb. */ 32021da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 32031da177e4SLinus Torvalds 3204e6c7d085SIlpo Järvinen /* Merge over control information. This moves PSH/FIN etc. over */ 32054de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; 32061da177e4SLinus Torvalds 32071da177e4SLinus Torvalds /* All done, get rid of second SKB and account for it so 32081da177e4SLinus Torvalds * packet counting does not break. 32091da177e4SLinus Torvalds */ 32104828e7f4SIlpo Järvinen TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; 3211a643b5d4SMartin KaFai Lau TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor; 3212b7689205SIlpo Järvinen 3213b7689205SIlpo Järvinen /* changed transmit queue under us so clear hints */ 3214ef9da47cSIlpo Järvinen tcp_clear_retrans_hints_partial(tp); 3215ef9da47cSIlpo Järvinen if (next_skb == tp->retransmit_skb_hint) 3216ef9da47cSIlpo Järvinen tp->retransmit_skb_hint = skb; 3217b7689205SIlpo Järvinen 3218797108d1SIlpo Järvinen tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); 3219797108d1SIlpo Järvinen 3220082ac2d5SMartin KaFai Lau tcp_skb_collapse_tstamp(skb, next_skb); 3221082ac2d5SMartin KaFai Lau 322275c119afSEric Dumazet tcp_rtx_queue_unlink_and_free(next_skb, sk); 3223f8071cdeSEric Dumazet return true; 32241da177e4SLinus Torvalds } 32251da177e4SLinus Torvalds 322667edfef7SAndi Kleen /* Check if coalescing SKBs is legal. */ 3227a2a385d6SEric Dumazet static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) 32284a17fc3aSIlpo Järvinen { 32294a17fc3aSIlpo Järvinen if (tcp_skb_pcount(skb) > 1) 3230a2a385d6SEric Dumazet return false; 32314a17fc3aSIlpo Järvinen if (skb_cloned(skb)) 3232a2a385d6SEric Dumazet return false; 32332331ccc5SEric Dumazet /* Some heuristics for collapsing over SACK'd could be invented */ 32344a17fc3aSIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 3235a2a385d6SEric Dumazet return false; 32364a17fc3aSIlpo Järvinen 3237a2a385d6SEric Dumazet return true; 32384a17fc3aSIlpo Järvinen } 32394a17fc3aSIlpo Järvinen 324067edfef7SAndi Kleen /* Collapse packets in the retransmit queue to make to create 324167edfef7SAndi Kleen * less packets on the wire. This is only done on retransmission. 324267edfef7SAndi Kleen */ 32434a17fc3aSIlpo Järvinen static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, 32444a17fc3aSIlpo Järvinen int space) 32454a17fc3aSIlpo Järvinen { 32464a17fc3aSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 32474a17fc3aSIlpo Järvinen struct sk_buff *skb = to, *tmp; 3248a2a385d6SEric Dumazet bool first = true; 32494a17fc3aSIlpo Järvinen 32501a63cb91SKuniyuki Iwashima if (!READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse)) 32514a17fc3aSIlpo Järvinen return; 32524de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 32534a17fc3aSIlpo Järvinen return; 32544a17fc3aSIlpo Järvinen 325575c119afSEric Dumazet skb_rbtree_walk_from_safe(skb, tmp) { 32564a17fc3aSIlpo Järvinen if (!tcp_can_collapse(sk, skb)) 32574a17fc3aSIlpo Järvinen break; 32584a17fc3aSIlpo Järvinen 325985712484SMat Martineau if (!tcp_skb_can_collapse(to, skb)) 3260a643b5d4SMartin KaFai Lau break; 3261a643b5d4SMartin KaFai Lau 32624a17fc3aSIlpo Järvinen space -= skb->len; 32634a17fc3aSIlpo Järvinen 32644a17fc3aSIlpo Järvinen if (first) { 3265a2a385d6SEric Dumazet first = false; 32664a17fc3aSIlpo Järvinen continue; 32674a17fc3aSIlpo Järvinen } 32684a17fc3aSIlpo Järvinen 32694a17fc3aSIlpo Järvinen if (space < 0) 32704a17fc3aSIlpo Järvinen break; 32714a17fc3aSIlpo Järvinen 32724a17fc3aSIlpo Järvinen if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) 32734a17fc3aSIlpo Järvinen break; 32744a17fc3aSIlpo Järvinen 3275f8071cdeSEric Dumazet if (!tcp_collapse_retrans(sk, to)) 3276f8071cdeSEric Dumazet break; 32774a17fc3aSIlpo Järvinen } 32784a17fc3aSIlpo Järvinen } 32794a17fc3aSIlpo Järvinen 32801da177e4SLinus Torvalds /* This retransmits one SKB. Policy decisions and retransmit queue 32811da177e4SLinus Torvalds * state updates are done by the caller. Returns non-zero if an 32821da177e4SLinus Torvalds * error occurred which prevented the send. 32831da177e4SLinus Torvalds */ 328410d3be56SEric Dumazet int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) 32851da177e4SLinus Torvalds { 32865d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 328710d3be56SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 32887d227cd2SSridhar Samudrala unsigned int cur_mss; 328910d3be56SEric Dumazet int diff, len, err; 3290536a6c8eSYonglong Li int avail_wnd; 329110d3be56SEric Dumazet 329210d3be56SEric Dumazet /* Inconclusive MTU probe */ 329310d3be56SEric Dumazet if (icsk->icsk_mtup.probe_size) 32945d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 32955d424d5aSJohn Heffner 32961f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 32971f3279aeSEric Dumazet return -EBUSY; 32981f3279aeSEric Dumazet 3299f99cd562SDong Chenchen start: 33001da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 3301f99cd562SDong Chenchen if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 3302f99cd562SDong Chenchen TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_SYN; 3303f99cd562SDong Chenchen TCP_SKB_CB(skb)->seq++; 3304f99cd562SDong Chenchen goto start; 3305f99cd562SDong Chenchen } 33067f582b24SEric Dumazet if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) { 33077f582b24SEric Dumazet WARN_ON_ONCE(1); 33087f582b24SEric Dumazet return -EINVAL; 33097f582b24SEric Dumazet } 33101da177e4SLinus Torvalds if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 33111da177e4SLinus Torvalds return -ENOMEM; 33121da177e4SLinus Torvalds } 33131da177e4SLinus Torvalds 33147d227cd2SSridhar Samudrala if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 33157d227cd2SSridhar Samudrala return -EHOSTUNREACH; /* Routing failure or similar. */ 33167d227cd2SSridhar Samudrala 33170c54b85fSIlpo Järvinen cur_mss = tcp_current_mss(sk); 3318536a6c8eSYonglong Li avail_wnd = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 33197d227cd2SSridhar Samudrala 33201da177e4SLinus Torvalds /* If receiver has shrunk his window, and skb is out of 33211da177e4SLinus Torvalds * new window, do not retransmit it. The exception is the 33221da177e4SLinus Torvalds * case, when window is shrunk to zero. In this case 3323536a6c8eSYonglong Li * our retransmit of one segment serves as a zero window probe. 33241da177e4SLinus Torvalds */ 3325536a6c8eSYonglong Li if (avail_wnd <= 0) { 3326536a6c8eSYonglong Li if (TCP_SKB_CB(skb)->seq != tp->snd_una) 33271da177e4SLinus Torvalds return -EAGAIN; 3328536a6c8eSYonglong Li avail_wnd = cur_mss; 3329536a6c8eSYonglong Li } 33301da177e4SLinus Torvalds 333110d3be56SEric Dumazet len = cur_mss * segs; 3332536a6c8eSYonglong Li if (len > avail_wnd) { 3333536a6c8eSYonglong Li len = rounddown(avail_wnd, cur_mss); 3334536a6c8eSYonglong Li if (!len) 3335536a6c8eSYonglong Li len = avail_wnd; 3336536a6c8eSYonglong Li } 333710d3be56SEric Dumazet if (skb->len > len) { 333875c119afSEric Dumazet if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len, 333975c119afSEric Dumazet cur_mss, GFP_ATOMIC)) 33401da177e4SLinus Torvalds return -ENOMEM; /* We'll try again later. */ 334102276f3cSIlpo Järvinen } else { 3342c4777efaSEric Dumazet if (skb_unclone_keeptruesize(skb, GFP_ATOMIC)) 3343c52e2421SEric Dumazet return -ENOMEM; 334410d3be56SEric Dumazet 334510d3be56SEric Dumazet diff = tcp_skb_pcount(skb); 334610d3be56SEric Dumazet tcp_set_skb_tso_segs(skb, cur_mss); 334710d3be56SEric Dumazet diff -= tcp_skb_pcount(skb); 334810d3be56SEric Dumazet if (diff) 334910d3be56SEric Dumazet tcp_adjust_pcount(sk, skb, diff); 3350536a6c8eSYonglong Li avail_wnd = min_t(int, avail_wnd, cur_mss); 3351536a6c8eSYonglong Li if (skb->len < avail_wnd) 3352536a6c8eSYonglong Li tcp_retrans_try_collapse(sk, skb, avail_wnd); 33531da177e4SLinus Torvalds } 33541da177e4SLinus Torvalds 335549213555SDaniel Borkmann /* RFC3168, section 6.1.1.1. ECN fallback */ 335649213555SDaniel Borkmann if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN) 335749213555SDaniel Borkmann tcp_ecn_clear_syn(sk, skb); 335849213555SDaniel Borkmann 3359678550c6SYuchung Cheng /* Update global and local TCP statistics. */ 3360678550c6SYuchung Cheng segs = tcp_skb_pcount(skb); 3361678550c6SYuchung Cheng TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs); 3362678550c6SYuchung Cheng if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 3363678550c6SYuchung Cheng __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 3364678550c6SYuchung Cheng tp->total_retrans += segs; 3365fb31c9b9SWei Wang tp->bytes_retrans += skb->len; 3366678550c6SYuchung Cheng 336750bceae9SThomas Graf /* make sure skb->data is aligned on arches that require it 336850bceae9SThomas Graf * and check if ack-trimming & collapsing extended the headroom 336950bceae9SThomas Graf * beyond what csum_start can cover. 337050bceae9SThomas Graf */ 337150bceae9SThomas Graf if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) || 337250bceae9SThomas Graf skb_headroom(skb) >= 0xFFFF)) { 337310a81980SEric Dumazet struct sk_buff *nskb; 337410a81980SEric Dumazet 3375e2080072SEric Dumazet tcp_skb_tsorted_save(skb) { 337610a81980SEric Dumazet nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC); 337707f8e4d0SFlorian Westphal if (nskb) { 337807f8e4d0SFlorian Westphal nskb->dev = NULL; 337907f8e4d0SFlorian Westphal err = tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC); 338007f8e4d0SFlorian Westphal } else { 338107f8e4d0SFlorian Westphal err = -ENOBUFS; 338207f8e4d0SFlorian Westphal } 3383e2080072SEric Dumazet } tcp_skb_tsorted_restore(skb); 3384e2080072SEric Dumazet 33855889e2c0SYousuk Seung if (!err) { 3386a7a25630SEric Dumazet tcp_update_skb_after_send(sk, skb, tp->tcp_wstamp_ns); 33875889e2c0SYousuk Seung tcp_rate_skb_sent(sk, skb); 33885889e2c0SYousuk Seung } 3389117632e6SEric Dumazet } else { 3390c84a5711SYuchung Cheng err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 3391117632e6SEric Dumazet } 3392c84a5711SYuchung Cheng 33937f12422cSYuchung Cheng /* To avoid taking spuriously low RTT samples based on a timestamp 33947f12422cSYuchung Cheng * for a transmit that never happened, always mark EVER_RETRANS 33957f12422cSYuchung Cheng */ 33967f12422cSYuchung Cheng TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS; 33977f12422cSYuchung Cheng 3398a31ad29eSLawrence Brakmo if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RETRANS_CB_FLAG)) 3399a31ad29eSLawrence Brakmo tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RETRANS_CB, 3400a31ad29eSLawrence Brakmo TCP_SKB_CB(skb)->seq, segs, err); 3401a31ad29eSLawrence Brakmo 3402fc9f3501SEric Dumazet if (likely(!err)) { 3403e086101bSCong Wang trace_tcp_retransmit_skb(sk, skb); 3404678550c6SYuchung Cheng } else if (err != -EBUSY) { 3405ec641b39SYuchung Cheng NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL, segs); 3406fc9f3501SEric Dumazet } 3407c84a5711SYuchung Cheng return err; 340893b174adSYuchung Cheng } 340993b174adSYuchung Cheng 341010d3be56SEric Dumazet int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) 341193b174adSYuchung Cheng { 341293b174adSYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 341310d3be56SEric Dumazet int err = __tcp_retransmit_skb(sk, skb, segs); 34141da177e4SLinus Torvalds 34151da177e4SLinus Torvalds if (err == 0) { 34161da177e4SLinus Torvalds #if FASTRETRANS_DEBUG > 0 34171da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 3418e87cc472SJoe Perches net_dbg_ratelimited("retrans_out leaked\n"); 34191da177e4SLinus Torvalds } 34201da177e4SLinus Torvalds #endif 34211da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 34221da177e4SLinus Torvalds tp->retrans_out += tcp_skb_pcount(skb); 34237ae18975SYuchung Cheng } 34241da177e4SLinus Torvalds 34257ae18975SYuchung Cheng /* Save stamp of the first (attempted) retransmit. */ 34261da177e4SLinus Torvalds if (!tp->retrans_stamp) 3427614e8316SEric Dumazet tp->retrans_stamp = tcp_skb_timestamp_ts(tp->tcp_usec_ts, skb); 34281da177e4SLinus Torvalds 34296e08d5e3SYuchung Cheng if (tp->undo_retrans < 0) 34306e08d5e3SYuchung Cheng tp->undo_retrans = 0; 34316e08d5e3SYuchung Cheng tp->undo_retrans += tcp_skb_pcount(skb); 34321da177e4SLinus Torvalds return err; 34331da177e4SLinus Torvalds } 34341da177e4SLinus Torvalds 34351da177e4SLinus Torvalds /* This gets called after a retransmit timeout, and the initially 34361da177e4SLinus Torvalds * retransmitted data is acknowledged. It tries to continue 34371da177e4SLinus Torvalds * resending the rest of the retransmit queue, until either 34381da177e4SLinus Torvalds * we've sent it all or the congestion window limit is reached. 34391da177e4SLinus Torvalds */ 34401da177e4SLinus Torvalds void tcp_xmit_retransmit_queue(struct sock *sk) 34411da177e4SLinus Torvalds { 34426687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 3443b9f1f1ceSEric Dumazet struct sk_buff *skb, *rtx_head, *hole = NULL; 34441da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 3445916e6d1aSEric Dumazet bool rearm_timer = false; 3446840a3cbeSYuchung Cheng u32 max_segs; 344761eb55f4SIlpo Järvinen int mib_idx; 34486a438bbeSStephen Hemminger 344945e77d31SIlpo Järvinen if (!tp->packets_out) 345045e77d31SIlpo Järvinen return; 345145e77d31SIlpo Järvinen 345275c119afSEric Dumazet rtx_head = tcp_rtx_queue_head(sk); 3453b9f1f1ceSEric Dumazet skb = tp->retransmit_skb_hint ?: rtx_head; 3454ed6e7268SNeal Cardwell max_segs = tcp_tso_segs(sk, tcp_current_mss(sk)); 345575c119afSEric Dumazet skb_rbtree_walk_from(skb) { 3456dca0aaf8SEric Dumazet __u8 sacked; 345710d3be56SEric Dumazet int segs; 34581da177e4SLinus Torvalds 3459218af599SEric Dumazet if (tcp_pacing_check(sk)) 3460218af599SEric Dumazet break; 3461218af599SEric Dumazet 34626a438bbeSStephen Hemminger /* we could do better than to assign each time */ 346351456b29SIan Morris if (!hole) 34646a438bbeSStephen Hemminger tp->retransmit_skb_hint = skb; 34656a438bbeSStephen Hemminger 346640570375SEric Dumazet segs = tcp_snd_cwnd(tp) - tcp_packets_in_flight(tp); 346710d3be56SEric Dumazet if (segs <= 0) 3468916e6d1aSEric Dumazet break; 3469dca0aaf8SEric Dumazet sacked = TCP_SKB_CB(skb)->sacked; 3470a3d2e9f8SEric Dumazet /* In case tcp_shift_skb_data() have aggregated large skbs, 3471a3d2e9f8SEric Dumazet * we need to make sure not sending too bigs TSO packets 3472a3d2e9f8SEric Dumazet */ 3473a3d2e9f8SEric Dumazet segs = min_t(int, segs, max_segs); 34740e1c54c2SIlpo Järvinen 3475840a3cbeSYuchung Cheng if (tp->retrans_out >= tp->lost_out) { 3476006f582cSIlpo Järvinen break; 34770e1c54c2SIlpo Järvinen } else if (!(sacked & TCPCB_LOST)) { 347851456b29SIan Morris if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) 34790e1c54c2SIlpo Järvinen hole = skb; 348061eb55f4SIlpo Järvinen continue; 34811da177e4SLinus Torvalds 34820e1c54c2SIlpo Järvinen } else { 34830e1c54c2SIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Loss) 34840e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFASTRETRANS; 34850e1c54c2SIlpo Järvinen else 34860e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; 34870e1c54c2SIlpo Järvinen } 34880e1c54c2SIlpo Järvinen 34890e1c54c2SIlpo Järvinen if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) 349061eb55f4SIlpo Järvinen continue; 349140b215e5SPavel Emelyanov 3492f9616c35SEric Dumazet if (tcp_small_queue_check(sk, skb, 1)) 3493916e6d1aSEric Dumazet break; 3494f9616c35SEric Dumazet 349510d3be56SEric Dumazet if (tcp_retransmit_skb(sk, skb, segs)) 3496916e6d1aSEric Dumazet break; 349724ab6becSYuchung Cheng 3498de1d6578SYuchung Cheng NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb)); 34991da177e4SLinus Torvalds 3500684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 3501a262f0cdSNandita Dukkipati tp->prr_out += tcp_skb_pcount(skb); 3502a262f0cdSNandita Dukkipati 350375c119afSEric Dumazet if (skb == rtx_head && 350457dde7f7SYuchung Cheng icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT) 3505916e6d1aSEric Dumazet rearm_timer = true; 3506916e6d1aSEric Dumazet 3507916e6d1aSEric Dumazet } 3508916e6d1aSEric Dumazet if (rearm_timer) 35093f80e08fSEric Dumazet tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 35103f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, 35118dc242adSEric Dumazet TCP_RTO_MAX); 35121da177e4SLinus Torvalds } 35131da177e4SLinus Torvalds 3514d83769a5SEric Dumazet /* We allow to exceed memory limits for FIN packets to expedite 3515d83769a5SEric Dumazet * connection tear down and (memory) recovery. 3516845704a5SEric Dumazet * Otherwise tcp_send_fin() could be tempted to either delay FIN 3517845704a5SEric Dumazet * or even be forced to close flow without any FIN. 3518a6c5ea4cSEric Dumazet * In general, we want to allow one skb per socket to avoid hangs 3519a6c5ea4cSEric Dumazet * with edge trigger epoll() 3520d83769a5SEric Dumazet */ 3521a6c5ea4cSEric Dumazet void sk_forced_mem_schedule(struct sock *sk, int size) 3522d83769a5SEric Dumazet { 3523c4ee1185SEric Dumazet int delta, amt; 3524d83769a5SEric Dumazet 3525c4ee1185SEric Dumazet delta = size - sk->sk_forward_alloc; 3526c4ee1185SEric Dumazet if (delta <= 0) 3527d83769a5SEric Dumazet return; 3528c4ee1185SEric Dumazet amt = sk_mem_pages(delta); 35295e6300e7SEric Dumazet sk_forward_alloc_add(sk, amt << PAGE_SHIFT); 3530e805605cSJohannes Weiner sk_memory_allocated_add(sk, amt); 3531e805605cSJohannes Weiner 3532baac50bbSJohannes Weiner if (mem_cgroup_sockets_enabled && sk->sk_memcg) 35334b1327beSWei Wang mem_cgroup_charge_skmem(sk->sk_memcg, amt, 35344b1327beSWei Wang gfp_memcg_charge() | __GFP_NOFAIL); 3535d83769a5SEric Dumazet } 3536d83769a5SEric Dumazet 3537845704a5SEric Dumazet /* Send a FIN. The caller locks the socket for us. 3538845704a5SEric Dumazet * We should try to send a FIN packet really hard, but eventually give up. 35391da177e4SLinus Torvalds */ 35401da177e4SLinus Torvalds void tcp_send_fin(struct sock *sk) 35411da177e4SLinus Torvalds { 3542ee2aabd3SEric Dumazet struct sk_buff *skb, *tskb, *tail = tcp_write_queue_tail(sk); 35431da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 35441da177e4SLinus Torvalds 3545845704a5SEric Dumazet /* Optimization, tack on the FIN if we have one skb in write queue and 3546845704a5SEric Dumazet * this skb was not yet sent, or we are under memory pressure. 3547845704a5SEric Dumazet * Note: in the latter case, FIN packet will be sent after a timeout, 3548845704a5SEric Dumazet * as TCP stack thinks it has already been transmitted. 35491da177e4SLinus Torvalds */ 3550ee2aabd3SEric Dumazet tskb = tail; 355175c119afSEric Dumazet if (!tskb && tcp_under_memory_pressure(sk)) 355275c119afSEric Dumazet tskb = skb_rb_last(&sk->tcp_rtx_queue); 355375c119afSEric Dumazet 355475c119afSEric Dumazet if (tskb) { 3555845704a5SEric Dumazet TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN; 3556845704a5SEric Dumazet TCP_SKB_CB(tskb)->end_seq++; 35571da177e4SLinus Torvalds tp->write_seq++; 3558ee2aabd3SEric Dumazet if (!tail) { 3559845704a5SEric Dumazet /* This means tskb was already sent. 3560845704a5SEric Dumazet * Pretend we included the FIN on previous transmit. 3561845704a5SEric Dumazet * We need to set tp->snd_nxt to the value it would have 3562845704a5SEric Dumazet * if FIN had been sent. This is because retransmit path 3563845704a5SEric Dumazet * does not change tp->snd_nxt. 3564845704a5SEric Dumazet */ 3565e0d694d6SEric Dumazet WRITE_ONCE(tp->snd_nxt, tp->snd_nxt + 1); 3566845704a5SEric Dumazet return; 3567845704a5SEric Dumazet } 35681da177e4SLinus Torvalds } else { 3569845704a5SEric Dumazet skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation); 3570d1edc085SColin Ian King if (unlikely(!skb)) 3571845704a5SEric Dumazet return; 3572d1edc085SColin Ian King 3573e2080072SEric Dumazet INIT_LIST_HEAD(&skb->tcp_tsorted_anchor); 3574d83769a5SEric Dumazet skb_reserve(skb, MAX_TCP_HEADER); 3575a6c5ea4cSEric Dumazet sk_forced_mem_schedule(sk, skb->truesize); 35761da177e4SLinus Torvalds /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 3577e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tp->write_seq, 3578a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_FIN); 35791da177e4SLinus Torvalds tcp_queue_skb(sk, skb); 35801da177e4SLinus Torvalds } 3581845704a5SEric Dumazet __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF); 35821da177e4SLinus Torvalds } 35831da177e4SLinus Torvalds 35841da177e4SLinus Torvalds /* We get here when a process closes a file descriptor (either due to 35851da177e4SLinus Torvalds * an explicit close() or as a byproduct of exit()'ing) and there 35861da177e4SLinus Torvalds * was unread data in the receive queue. This behavior is recommended 358765bb723cSGerrit Renker * by RFC 2525, section 2.17. -DaveM 35881da177e4SLinus Torvalds */ 3589dd0fc66fSAl Viro void tcp_send_active_reset(struct sock *sk, gfp_t priority) 35901da177e4SLinus Torvalds { 35911da177e4SLinus Torvalds struct sk_buff *skb; 35921da177e4SLinus Torvalds 35937cc2b043SGao Feng TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); 35947cc2b043SGao Feng 35951da177e4SLinus Torvalds /* NOTE: No TCP options attached and we never retransmit this. */ 35961da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, priority); 35971da177e4SLinus Torvalds if (!skb) { 35984e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 35991da177e4SLinus Torvalds return; 36001da177e4SLinus Torvalds } 36011da177e4SLinus Torvalds 36021da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 36031da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 3604e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), 3605a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_RST); 36069a568de4SEric Dumazet tcp_mstamp_refresh(tcp_sk(sk)); 36071da177e4SLinus Torvalds /* Send it off. */ 3608dfb4b9dcSDavid S. Miller if (tcp_transmit_skb(sk, skb, 0, priority)) 36094e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 3610c24b14c4SSong Liu 3611c24b14c4SSong Liu /* skb of trace_tcp_send_reset() keeps the skb that caused RST, 3612c24b14c4SSong Liu * skb here is different to the troublesome skb, so use NULL 3613c24b14c4SSong Liu */ 3614c24b14c4SSong Liu trace_tcp_send_reset(sk, NULL); 36151da177e4SLinus Torvalds } 36161da177e4SLinus Torvalds 361767edfef7SAndi Kleen /* Send a crossed SYN-ACK during socket establishment. 361867edfef7SAndi Kleen * WARNING: This routine must only be called when we have already sent 36191da177e4SLinus Torvalds * a SYN packet that crossed the incoming SYN that caused this routine 36201da177e4SLinus Torvalds * to get called. If this assumption fails then the initial rcv_wnd 36211da177e4SLinus Torvalds * and rcv_wscale values will not be correct. 36221da177e4SLinus Torvalds */ 36231da177e4SLinus Torvalds int tcp_send_synack(struct sock *sk) 36241da177e4SLinus Torvalds { 36251da177e4SLinus Torvalds struct sk_buff *skb; 36261da177e4SLinus Torvalds 362775c119afSEric Dumazet skb = tcp_rtx_queue_head(sk); 362851456b29SIan Morris if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 362975c119afSEric Dumazet pr_err("%s: wrong queue state\n", __func__); 36301da177e4SLinus Torvalds return -EFAULT; 36311da177e4SLinus Torvalds } 36324de075e0SEric Dumazet if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { 36331da177e4SLinus Torvalds if (skb_cloned(skb)) { 3634e2080072SEric Dumazet struct sk_buff *nskb; 3635e2080072SEric Dumazet 3636e2080072SEric Dumazet tcp_skb_tsorted_save(skb) { 3637e2080072SEric Dumazet nskb = skb_copy(skb, GFP_ATOMIC); 3638e2080072SEric Dumazet } tcp_skb_tsorted_restore(skb); 363951456b29SIan Morris if (!nskb) 36401da177e4SLinus Torvalds return -ENOMEM; 3641e2080072SEric Dumazet INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor); 36422bec445fSEric Dumazet tcp_highest_sack_replace(sk, skb, nskb); 364375c119afSEric Dumazet tcp_rtx_queue_unlink_and_free(skb, sk); 3644f4a775d1SEric Dumazet __skb_header_release(nskb); 364575c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb); 3646ab4e846aSEric Dumazet sk_wmem_queued_add(sk, nskb->truesize); 36473ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 36481da177e4SLinus Torvalds skb = nskb; 36491da177e4SLinus Torvalds } 36501da177e4SLinus Torvalds 36514de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; 3652735d3831SFlorian Westphal tcp_ecn_send_synack(sk, skb); 36531da177e4SLinus Torvalds } 3654dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 36551da177e4SLinus Torvalds } 36561da177e4SLinus Torvalds 36574aea39c1SEric Dumazet /** 3658331fca43SMartin KaFai Lau * tcp_make_synack - Allocate one skb and build a SYNACK packet. 3659331fca43SMartin KaFai Lau * @sk: listener socket 3660331fca43SMartin KaFai Lau * @dst: dst entry attached to the SYNACK. It is consumed and caller 3661331fca43SMartin KaFai Lau * should not use it again. 3662331fca43SMartin KaFai Lau * @req: request_sock pointer 3663331fca43SMartin KaFai Lau * @foc: cookie for tcp fast open 3664331fca43SMartin KaFai Lau * @synack_type: Type of synack to prepare 3665331fca43SMartin KaFai Lau * @syn_skb: SYN packet just received. It could be NULL for rtx case. 36664aea39c1SEric Dumazet */ 36675d062de7SEric Dumazet struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst, 3668e6b4d113SWilliam Allen Simpson struct request_sock *req, 3669ca6fb065SEric Dumazet struct tcp_fastopen_cookie *foc, 3670331fca43SMartin KaFai Lau enum tcp_synack_type synack_type, 3671331fca43SMartin KaFai Lau struct sk_buff *syn_skb) 36721da177e4SLinus Torvalds { 36732e6599cbSArnaldo Carvalho de Melo struct inet_request_sock *ireq = inet_rsk(req); 36745d062de7SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 36755d062de7SEric Dumazet struct tcp_out_options opts; 36761e03d32bSDmitry Safonov struct tcp_key key = {}; 36775d062de7SEric Dumazet struct sk_buff *skb; 3678bd0388aeSWilliam Allen Simpson int tcp_header_size; 36795d062de7SEric Dumazet struct tcphdr *th; 3680f5fff5dcSTom Quetchenbach int mss; 3681a842fe14SEric Dumazet u64 now; 36821da177e4SLinus Torvalds 3683ca6fb065SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 36844aea39c1SEric Dumazet if (unlikely(!skb)) { 36854aea39c1SEric Dumazet dst_release(dst); 36861da177e4SLinus Torvalds return NULL; 36874aea39c1SEric Dumazet } 36881da177e4SLinus Torvalds /* Reserve space for headers. */ 36891da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 36901da177e4SLinus Torvalds 3691b3d05147SEric Dumazet switch (synack_type) { 3692b3d05147SEric Dumazet case TCP_SYNACK_NORMAL: 36939e17f8a4SEric Dumazet skb_set_owner_w(skb, req_to_sk(req)); 3694b3d05147SEric Dumazet break; 3695b3d05147SEric Dumazet case TCP_SYNACK_COOKIE: 3696b3d05147SEric Dumazet /* Under synflood, we do not attach skb to a socket, 3697b3d05147SEric Dumazet * to avoid false sharing. 3698b3d05147SEric Dumazet */ 3699b3d05147SEric Dumazet break; 3700b3d05147SEric Dumazet case TCP_SYNACK_FASTOPEN: 3701ca6fb065SEric Dumazet /* sk is a const pointer, because we want to express multiple 3702ca6fb065SEric Dumazet * cpu might call us concurrently. 3703ca6fb065SEric Dumazet * sk->sk_wmem_alloc in an atomic, we can promote to rw. 3704ca6fb065SEric Dumazet */ 3705ca6fb065SEric Dumazet skb_set_owner_w(skb, (struct sock *)sk); 3706b3d05147SEric Dumazet break; 3707ca6fb065SEric Dumazet } 37084aea39c1SEric Dumazet skb_dst_set(skb, dst); 37091da177e4SLinus Torvalds 37103541f9e8SEric Dumazet mss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 3711f5fff5dcSTom Quetchenbach 371233ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 3713a842fe14SEric Dumazet now = tcp_clock_ns(); 37148b5f12d0SFlorian Westphal #ifdef CONFIG_SYN_COOKIES 3715f8ace8d9SFlorian Westphal if (unlikely(synack_type == TCP_SYNACK_COOKIE && ireq->tstamp_ok)) 3716a1ac9c8aSMartin KaFai Lau skb_set_delivery_time(skb, cookie_init_timestamp(req, now), 3717a1ac9c8aSMartin KaFai Lau true); 37188b5f12d0SFlorian Westphal else 37198b5f12d0SFlorian Westphal #endif 37209e450c1eSYuchung Cheng { 3721a1ac9c8aSMartin KaFai Lau skb_set_delivery_time(skb, now, true); 37229e450c1eSYuchung Cheng if (!tcp_rsk(req)->snt_synack) /* Timestamp first SYNACK */ 37239e450c1eSYuchung Cheng tcp_rsk(req)->snt_synack = tcp_skb_timestamp_us(skb); 37249e450c1eSYuchung Cheng } 372580f03e27SEric Dumazet 37269427c6aaSDmitry Safonov #if defined(CONFIG_TCP_MD5SIG) || defined(CONFIG_TCP_AO) 372780f03e27SEric Dumazet rcu_read_lock(); 37289427c6aaSDmitry Safonov #endif 37299427c6aaSDmitry Safonov if (tcp_rsk_used_ao(req)) { 37309427c6aaSDmitry Safonov #ifdef CONFIG_TCP_AO 37319427c6aaSDmitry Safonov struct tcp_ao_key *ao_key = NULL; 37329427c6aaSDmitry Safonov u8 keyid = tcp_rsk(req)->ao_keyid; 37339427c6aaSDmitry Safonov 37349427c6aaSDmitry Safonov ao_key = tcp_sk(sk)->af_specific->ao_lookup(sk, req_to_sk(req), 37359427c6aaSDmitry Safonov keyid, -1); 37369427c6aaSDmitry Safonov /* If there is no matching key - avoid sending anything, 37379427c6aaSDmitry Safonov * especially usigned segments. It could try harder and lookup 37389427c6aaSDmitry Safonov * for another peer-matching key, but the peer has requested 37399427c6aaSDmitry Safonov * ao_keyid (RFC5925 RNextKeyID), so let's keep it simple here. 37409427c6aaSDmitry Safonov */ 37419396c4eeSDmitry Safonov if (unlikely(!ao_key)) { 37429427c6aaSDmitry Safonov rcu_read_unlock(); 37439427c6aaSDmitry Safonov kfree_skb(skb); 37449396c4eeSDmitry Safonov net_warn_ratelimited("TCP-AO: the keyid %u from SYN packet is not present - not sending SYNACK\n", 37459396c4eeSDmitry Safonov keyid); 37469427c6aaSDmitry Safonov return NULL; 37479427c6aaSDmitry Safonov } 37489427c6aaSDmitry Safonov key.ao_key = ao_key; 37499427c6aaSDmitry Safonov key.type = TCP_KEY_AO; 37509427c6aaSDmitry Safonov #endif 37519427c6aaSDmitry Safonov } else { 37529427c6aaSDmitry Safonov #ifdef CONFIG_TCP_MD5SIG 37539427c6aaSDmitry Safonov key.md5_key = tcp_rsk(req)->af_specific->req_md5_lookup(sk, 37549427c6aaSDmitry Safonov req_to_sk(req)); 37559427c6aaSDmitry Safonov if (key.md5_key) 37561e03d32bSDmitry Safonov key.type = TCP_KEY_MD5; 375780f03e27SEric Dumazet #endif 37589427c6aaSDmitry Safonov } 37595e526552SEric Dumazet skb_set_hash(skb, READ_ONCE(tcp_rsk(req)->txhash), PKT_HASH_TYPE_L4); 3760331fca43SMartin KaFai Lau /* bpf program will be interested in the tcp_flags */ 3761331fca43SMartin KaFai Lau TCP_SKB_CB(skb)->tcp_flags = TCPHDR_SYN | TCPHDR_ACK; 37629427c6aaSDmitry Safonov tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, 37639427c6aaSDmitry Safonov &key, foc, synack_type, syn_skb) 37649427c6aaSDmitry Safonov + sizeof(*th); 376533ad798cSAdam Langley 3766aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 3767aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 37681da177e4SLinus Torvalds 3769ea1627c2SEric Dumazet th = (struct tcphdr *)skb->data; 37701da177e4SLinus Torvalds memset(th, 0, sizeof(struct tcphdr)); 37711da177e4SLinus Torvalds th->syn = 1; 37721da177e4SLinus Torvalds th->ack = 1; 37736ac705b1SEric Dumazet tcp_ecn_make_synack(req, th); 3774b44084c2SEric Dumazet th->source = htons(ireq->ir_num); 3775634fb979SEric Dumazet th->dest = ireq->ir_rmt_port; 3776e05a90ecSJamal Hadi Salim skb->mark = ireq->ir_mark; 37773b117750SEric Dumazet skb->ip_summed = CHECKSUM_PARTIAL; 37783b117750SEric Dumazet th->seq = htonl(tcp_rsk(req)->snt_isn); 37798336886fSJerry Chu /* XXX data is queued and acked as is. No buffer/window check */ 37808336886fSJerry Chu th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt); 37811da177e4SLinus Torvalds 37821da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 3783ed53d0abSEric Dumazet th->window = htons(min(req->rsk_rcv_wnd, 65535U)); 37849427c6aaSDmitry Safonov tcp_options_write(th, NULL, tcp_rsk(req), &opts, &key); 37851da177e4SLinus Torvalds th->doff = (tcp_header_size >> 2); 3786bced3f7dSBreno Leitao TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS); 3787cfb6eeb4SYOSHIFUJI Hideaki 3788cfb6eeb4SYOSHIFUJI Hideaki /* Okay, we have all we need - do the md5 hash if needed */ 37899427c6aaSDmitry Safonov if (tcp_key_is_md5(&key)) { 37909427c6aaSDmitry Safonov #ifdef CONFIG_TCP_MD5SIG 3791bd0388aeSWilliam Allen Simpson tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, 37929427c6aaSDmitry Safonov key.md5_key, req_to_sk(req), skb); 37939427c6aaSDmitry Safonov #endif 37949427c6aaSDmitry Safonov } else if (tcp_key_is_ao(&key)) { 37959427c6aaSDmitry Safonov #ifdef CONFIG_TCP_AO 37969427c6aaSDmitry Safonov tcp_rsk(req)->af_specific->ao_synack_hash(opts.hash_location, 37979427c6aaSDmitry Safonov key.ao_key, req, skb, 37989427c6aaSDmitry Safonov opts.hash_location - (u8 *)th, 0); 37999427c6aaSDmitry Safonov #endif 38009427c6aaSDmitry Safonov } 38019427c6aaSDmitry Safonov #if defined(CONFIG_TCP_MD5SIG) || defined(CONFIG_TCP_AO) 380280f03e27SEric Dumazet rcu_read_unlock(); 3803cfb6eeb4SYOSHIFUJI Hideaki #endif 3804cfb6eeb4SYOSHIFUJI Hideaki 3805331fca43SMartin KaFai Lau bpf_skops_write_hdr_opt((struct sock *)sk, skb, req, syn_skb, 3806331fca43SMartin KaFai Lau synack_type, &opts); 3807331fca43SMartin KaFai Lau 3808a1ac9c8aSMartin KaFai Lau skb_set_delivery_time(skb, now, true); 3809a842fe14SEric Dumazet tcp_add_tx_delay(skb, tp); 3810a842fe14SEric Dumazet 38111da177e4SLinus Torvalds return skb; 38121da177e4SLinus Torvalds } 38134bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_make_synack); 38141da177e4SLinus Torvalds 381581164413SDaniel Borkmann static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst) 381681164413SDaniel Borkmann { 381781164413SDaniel Borkmann struct inet_connection_sock *icsk = inet_csk(sk); 381881164413SDaniel Borkmann const struct tcp_congestion_ops *ca; 381981164413SDaniel Borkmann u32 ca_key = dst_metric(dst, RTAX_CC_ALGO); 382081164413SDaniel Borkmann 382181164413SDaniel Borkmann if (ca_key == TCP_CA_UNSPEC) 382281164413SDaniel Borkmann return; 382381164413SDaniel Borkmann 382481164413SDaniel Borkmann rcu_read_lock(); 382581164413SDaniel Borkmann ca = tcp_ca_find_key(ca_key); 38260baf26b0SMartin KaFai Lau if (likely(ca && bpf_try_module_get(ca, ca->owner))) { 38270baf26b0SMartin KaFai Lau bpf_module_put(icsk->icsk_ca_ops, icsk->icsk_ca_ops->owner); 382881164413SDaniel Borkmann icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst); 382981164413SDaniel Borkmann icsk->icsk_ca_ops = ca; 383081164413SDaniel Borkmann } 383181164413SDaniel Borkmann rcu_read_unlock(); 383281164413SDaniel Borkmann } 383381164413SDaniel Borkmann 383467edfef7SAndi Kleen /* Do all connect socket setups that can be done AF independent. */ 3835f7e56a76Sstephen hemminger static void tcp_connect_init(struct sock *sk) 38361da177e4SLinus Torvalds { 3837cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 38381da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 38391da177e4SLinus Torvalds __u8 rcv_wscale; 384013d3b1ebSLawrence Brakmo u32 rcv_wnd; 38411da177e4SLinus Torvalds 38421da177e4SLinus Torvalds /* We'll fix this up when we get a response from the other end. 38431da177e4SLinus Torvalds * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 38441da177e4SLinus Torvalds */ 38455d2ed052SEric Dumazet tp->tcp_header_len = sizeof(struct tcphdr); 38463666f666SKuniyuki Iwashima if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_timestamps)) 38475d2ed052SEric Dumazet tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED; 38481da177e4SLinus Torvalds 38497c2ffaf2SDmitry Safonov tcp_ao_connect_init(sk); 38507c2ffaf2SDmitry Safonov 38511da177e4SLinus Torvalds /* If user gave his TCP_MAXSEG, record it to clamp */ 38521da177e4SLinus Torvalds if (tp->rx_opt.user_mss) 38531da177e4SLinus Torvalds tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 38541da177e4SLinus Torvalds tp->max_window = 0; 38555d424d5aSJohn Heffner tcp_mtup_init(sk); 38561da177e4SLinus Torvalds tcp_sync_mss(sk, dst_mtu(dst)); 38571da177e4SLinus Torvalds 385881164413SDaniel Borkmann tcp_ca_dst_init(sk, dst); 385981164413SDaniel Borkmann 38601da177e4SLinus Torvalds if (!tp->window_clamp) 3861f410cbeaSEric Dumazet WRITE_ONCE(tp->window_clamp, dst_metric(dst, RTAX_WINDOW)); 38623541f9e8SEric Dumazet tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 3863f5fff5dcSTom Quetchenbach 38641da177e4SLinus Torvalds tcp_initialize_rcv_mss(sk); 38651da177e4SLinus Torvalds 3866e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 3867e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 3868e88c64f0SHagen Paul Pfeifer (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) 3869f410cbeaSEric Dumazet WRITE_ONCE(tp->window_clamp, tcp_full_space(sk)); 3870e88c64f0SHagen Paul Pfeifer 387113d3b1ebSLawrence Brakmo rcv_wnd = tcp_rwnd_init_bpf(sk); 387213d3b1ebSLawrence Brakmo if (rcv_wnd == 0) 387313d3b1ebSLawrence Brakmo rcv_wnd = dst_metric(dst, RTAX_INITRWND); 387413d3b1ebSLawrence Brakmo 3875ceef9ab6SEric Dumazet tcp_select_initial_window(sk, tcp_full_space(sk), 38761da177e4SLinus Torvalds tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 38771da177e4SLinus Torvalds &tp->rcv_wnd, 38781da177e4SLinus Torvalds &tp->window_clamp, 38793666f666SKuniyuki Iwashima READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_window_scaling), 388031d12926Slaurent chavey &rcv_wscale, 388113d3b1ebSLawrence Brakmo rcv_wnd); 38821da177e4SLinus Torvalds 38831da177e4SLinus Torvalds tp->rx_opt.rcv_wscale = rcv_wscale; 38841da177e4SLinus Torvalds tp->rcv_ssthresh = tp->rcv_wnd; 38851da177e4SLinus Torvalds 3886e13ec3daSEric Dumazet WRITE_ONCE(sk->sk_err, 0); 38871da177e4SLinus Torvalds sock_reset_flag(sk, SOCK_DONE); 38881da177e4SLinus Torvalds tp->snd_wnd = 0; 3889ee7537b6SHantzis Fotis tcp_init_wl(tp, 0); 38907f582b24SEric Dumazet tcp_write_queue_purge(sk); 38911da177e4SLinus Torvalds tp->snd_una = tp->write_seq; 38921da177e4SLinus Torvalds tp->snd_sml = tp->write_seq; 389333f5f57eSIlpo Järvinen tp->snd_up = tp->write_seq; 3894e0d694d6SEric Dumazet WRITE_ONCE(tp->snd_nxt, tp->write_seq); 3895ee995283SPavel Emelyanov 3896ee995283SPavel Emelyanov if (likely(!tp->repair)) 38971da177e4SLinus Torvalds tp->rcv_nxt = 0; 3898c7781a6eSAndrew Vagin else 389970eabf0eSEric Dumazet tp->rcv_tstamp = tcp_jiffies32; 3900ee995283SPavel Emelyanov tp->rcv_wup = tp->rcv_nxt; 39017db48e98SEric Dumazet WRITE_ONCE(tp->copied_seq, tp->rcv_nxt); 39021da177e4SLinus Torvalds 39038550f328SLawrence Brakmo inet_csk(sk)->icsk_rto = tcp_timeout_init(sk); 3904463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_retransmits = 0; 39051da177e4SLinus Torvalds tcp_clear_retrans(tp); 39061da177e4SLinus Torvalds } 39071da177e4SLinus Torvalds 3908783237e8SYuchung Cheng static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) 3909783237e8SYuchung Cheng { 3910783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3911783237e8SYuchung Cheng struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 3912783237e8SYuchung Cheng 3913783237e8SYuchung Cheng tcb->end_seq += skb->len; 3914f4a775d1SEric Dumazet __skb_header_release(skb); 3915ab4e846aSEric Dumazet sk_wmem_queued_add(sk, skb->truesize); 3916783237e8SYuchung Cheng sk_mem_charge(sk, skb->truesize); 39170f317464SEric Dumazet WRITE_ONCE(tp->write_seq, tcb->end_seq); 3918783237e8SYuchung Cheng tp->packets_out += tcp_skb_pcount(skb); 3919783237e8SYuchung Cheng } 3920783237e8SYuchung Cheng 3921783237e8SYuchung Cheng /* Build and send a SYN with data and (cached) Fast Open cookie. However, 3922783237e8SYuchung Cheng * queue a data-only packet after the regular SYN, such that regular SYNs 3923783237e8SYuchung Cheng * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges 3924783237e8SYuchung Cheng * only the SYN sequence, the data are retransmitted in the first ACK. 3925783237e8SYuchung Cheng * If cookie is not cached or other error occurs, falls back to send a 3926783237e8SYuchung Cheng * regular SYN with Fast Open cookie request option. 3927783237e8SYuchung Cheng */ 3928783237e8SYuchung Cheng static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) 3929783237e8SYuchung Cheng { 3930ed0c99dcSJakub Kicinski struct inet_connection_sock *icsk = inet_csk(sk); 3931783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3932783237e8SYuchung Cheng struct tcp_fastopen_request *fo = tp->fastopen_req; 3933fbf93406SEric Dumazet struct page_frag *pfrag = sk_page_frag(sk); 3934355a901eSEric Dumazet struct sk_buff *syn_data; 3935fbf93406SEric Dumazet int space, err = 0; 3936783237e8SYuchung Cheng 393767da22d2SYuchung Cheng tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ 3938065263f4SWei Wang if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie)) 3939783237e8SYuchung Cheng goto fallback; 3940783237e8SYuchung Cheng 3941783237e8SYuchung Cheng /* MSS for SYN-data is based on cached MSS and bounded by PMTU and 3942783237e8SYuchung Cheng * user-MSS. Reserve maximum option space for middleboxes that add 3943783237e8SYuchung Cheng * private TCP options. The cost is reduced data space in SYN :( 3944783237e8SYuchung Cheng */ 39453541f9e8SEric Dumazet tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp); 3946ed0c99dcSJakub Kicinski /* Sync mss_cache after updating the mss_clamp */ 3947ed0c99dcSJakub Kicinski tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 39483541f9e8SEric Dumazet 3949ed0c99dcSJakub Kicinski space = __tcp_mtu_to_mss(sk, icsk->icsk_pmtu_cookie) - 3950783237e8SYuchung Cheng MAX_TCP_OPTION_SPACE; 3951783237e8SYuchung Cheng 3952f5ddcbbbSEric Dumazet space = min_t(size_t, space, fo->size); 3953f5ddcbbbSEric Dumazet 3954fbf93406SEric Dumazet if (space && 3955fbf93406SEric Dumazet !skb_page_frag_refill(min_t(size_t, space, PAGE_SIZE), 3956fbf93406SEric Dumazet pfrag, sk->sk_allocation)) 3957fbf93406SEric Dumazet goto fallback; 39585882efffSEric Dumazet syn_data = tcp_stream_alloc_skb(sk, sk->sk_allocation, false); 3959355a901eSEric Dumazet if (!syn_data) 3960783237e8SYuchung Cheng goto fallback; 3961355a901eSEric Dumazet memcpy(syn_data->cb, syn->cb, sizeof(syn->cb)); 396207e100f9SEric Dumazet if (space) { 3963fbf93406SEric Dumazet space = min_t(size_t, space, pfrag->size - pfrag->offset); 3964fbf93406SEric Dumazet space = tcp_wmem_schedule(sk, space); 3965fbf93406SEric Dumazet } 3966fbf93406SEric Dumazet if (space) { 3967fbf93406SEric Dumazet space = copy_page_from_iter(pfrag->page, pfrag->offset, 3968fbf93406SEric Dumazet space, &fo->data->msg_iter); 3969fbf93406SEric Dumazet if (unlikely(!space)) { 3970ba233b34SEric Dumazet tcp_skb_tsorted_anchor_cleanup(syn_data); 3971355a901eSEric Dumazet kfree_skb(syn_data); 3972783237e8SYuchung Cheng goto fallback; 3973783237e8SYuchung Cheng } 3974fbf93406SEric Dumazet skb_fill_page_desc(syn_data, 0, pfrag->page, 3975fbf93406SEric Dumazet pfrag->offset, space); 3976fbf93406SEric Dumazet page_ref_inc(pfrag->page); 3977fbf93406SEric Dumazet pfrag->offset += space; 3978fbf93406SEric Dumazet skb_len_add(syn_data, space); 3979f859a448SWillem de Bruijn skb_zcopy_set(syn_data, fo->uarg, NULL); 398007e100f9SEric Dumazet } 3981355a901eSEric Dumazet /* No more data pending in inet_wait_for_connect() */ 3982355a901eSEric Dumazet if (space == fo->size) 3983355a901eSEric Dumazet fo->data = NULL; 3984355a901eSEric Dumazet fo->copied = space; 3985783237e8SYuchung Cheng 3986355a901eSEric Dumazet tcp_connect_queue_skb(sk, syn_data); 39870f87230dSFrancis Yan if (syn_data->len) 39880f87230dSFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_BUSY); 3989355a901eSEric Dumazet 3990355a901eSEric Dumazet err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation); 3991355a901eSEric Dumazet 3992a1ac9c8aSMartin KaFai Lau skb_set_delivery_time(syn, syn_data->skb_mstamp_ns, true); 3993355a901eSEric Dumazet 3994355a901eSEric Dumazet /* Now full SYN+DATA was cloned and sent (or not), 3995355a901eSEric Dumazet * remove the SYN from the original skb (syn_data) 3996355a901eSEric Dumazet * we keep in write queue in case of a retransmit, as we 3997355a901eSEric Dumazet * also have the SYN packet (with no data) in the same queue. 3998431a9124SEric Dumazet */ 3999355a901eSEric Dumazet TCP_SKB_CB(syn_data)->seq++; 4000355a901eSEric Dumazet TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH; 4001355a901eSEric Dumazet if (!err) { 400267da22d2SYuchung Cheng tp->syn_data = (fo->copied > 0); 400375c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data); 4004f19c29e3SYuchung Cheng NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT); 4005783237e8SYuchung Cheng goto done; 4006783237e8SYuchung Cheng } 4007783237e8SYuchung Cheng 400875c119afSEric Dumazet /* data was not sent, put it in write_queue */ 400975c119afSEric Dumazet __skb_queue_tail(&sk->sk_write_queue, syn_data); 4010b5b7db8dSEric Dumazet tp->packets_out -= tcp_skb_pcount(syn_data); 4011b5b7db8dSEric Dumazet 4012783237e8SYuchung Cheng fallback: 4013783237e8SYuchung Cheng /* Send a regular SYN with Fast Open cookie request option */ 4014783237e8SYuchung Cheng if (fo->cookie.len > 0) 4015783237e8SYuchung Cheng fo->cookie.len = 0; 4016783237e8SYuchung Cheng err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); 4017783237e8SYuchung Cheng if (err) 4018783237e8SYuchung Cheng tp->syn_fastopen = 0; 4019783237e8SYuchung Cheng done: 4020783237e8SYuchung Cheng fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ 4021783237e8SYuchung Cheng return err; 4022783237e8SYuchung Cheng } 4023783237e8SYuchung Cheng 402467edfef7SAndi Kleen /* Build a SYN and send it off. */ 40251da177e4SLinus Torvalds int tcp_connect(struct sock *sk) 40261da177e4SLinus Torvalds { 40271da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 40281da177e4SLinus Torvalds struct sk_buff *buff; 4029ee586811SEric Paris int err; 40301da177e4SLinus Torvalds 4031de525be2SLawrence Brakmo tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB, 0, NULL); 40328ba60924SEric Dumazet 40330aadc739SDmitry Safonov #if defined(CONFIG_TCP_MD5SIG) && defined(CONFIG_TCP_AO) 40340aadc739SDmitry Safonov /* Has to be checked late, after setting daddr/saddr/ops. 40350aadc739SDmitry Safonov * Return error if the peer has both a md5 and a tcp-ao key 40360aadc739SDmitry Safonov * configured as this is ambiguous. 40370aadc739SDmitry Safonov */ 40380aadc739SDmitry Safonov if (unlikely(rcu_dereference_protected(tp->md5sig_info, 40390aadc739SDmitry Safonov lockdep_sock_is_held(sk)))) { 40400aadc739SDmitry Safonov bool needs_ao = !!tp->af_specific->ao_lookup(sk, sk, -1, -1); 40410aadc739SDmitry Safonov bool needs_md5 = !!tp->af_specific->md5_lookup(sk, sk); 40420aadc739SDmitry Safonov struct tcp_ao_info *ao_info; 40430aadc739SDmitry Safonov 40440aadc739SDmitry Safonov ao_info = rcu_dereference_check(tp->ao_info, 40450aadc739SDmitry Safonov lockdep_sock_is_held(sk)); 40460aadc739SDmitry Safonov if (ao_info) { 40470aadc739SDmitry Safonov /* This is an extra check: tcp_ao_required() in 40480aadc739SDmitry Safonov * tcp_v{4,6}_parse_md5_keys() should prevent adding 40490aadc739SDmitry Safonov * md5 keys on ao_required socket. 40500aadc739SDmitry Safonov */ 40510aadc739SDmitry Safonov needs_ao |= ao_info->ao_required; 40520aadc739SDmitry Safonov WARN_ON_ONCE(ao_info->ao_required && needs_md5); 40530aadc739SDmitry Safonov } 40540aadc739SDmitry Safonov if (needs_md5 && needs_ao) 40550aadc739SDmitry Safonov return -EKEYREJECTED; 40560aadc739SDmitry Safonov 40570aadc739SDmitry Safonov /* If we have a matching md5 key and no matching tcp-ao key 40580aadc739SDmitry Safonov * then free up ao_info if allocated. 40590aadc739SDmitry Safonov */ 40600aadc739SDmitry Safonov if (needs_md5) { 4061decde258SDmitry Safonov tcp_ao_destroy_sock(sk, false); 40620aadc739SDmitry Safonov } else if (needs_ao) { 40630aadc739SDmitry Safonov tcp_clear_md5_list(sk); 40640aadc739SDmitry Safonov kfree(rcu_replace_pointer(tp->md5sig_info, NULL, 40650aadc739SDmitry Safonov lockdep_sock_is_held(sk))); 40660aadc739SDmitry Safonov } 40670aadc739SDmitry Safonov } 40680aadc739SDmitry Safonov #endif 40690aadc739SDmitry Safonov #ifdef CONFIG_TCP_AO 40700aadc739SDmitry Safonov if (unlikely(rcu_dereference_protected(tp->ao_info, 40710aadc739SDmitry Safonov lockdep_sock_is_held(sk)))) { 40720aadc739SDmitry Safonov /* Don't allow connecting if ao is configured but no 40730aadc739SDmitry Safonov * matching key is found. 40740aadc739SDmitry Safonov */ 40750aadc739SDmitry Safonov if (!tp->af_specific->ao_lookup(sk, sk, -1, -1)) 40760aadc739SDmitry Safonov return -EKEYREJECTED; 40770aadc739SDmitry Safonov } 40780aadc739SDmitry Safonov #endif 40790aadc739SDmitry Safonov 40808ba60924SEric Dumazet if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 40818ba60924SEric Dumazet return -EHOSTUNREACH; /* Routing failure or similar. */ 40828ba60924SEric Dumazet 40831da177e4SLinus Torvalds tcp_connect_init(sk); 40841da177e4SLinus Torvalds 40852b916477SAndrey Vagin if (unlikely(tp->repair)) { 40862b916477SAndrey Vagin tcp_finish_connect(sk, NULL); 40872b916477SAndrey Vagin return 0; 40882b916477SAndrey Vagin } 40892b916477SAndrey Vagin 40905882efffSEric Dumazet buff = tcp_stream_alloc_skb(sk, sk->sk_allocation, true); 4091355a901eSEric Dumazet if (unlikely(!buff)) 40921da177e4SLinus Torvalds return -ENOBUFS; 40931da177e4SLinus Torvalds 4094a3433f35SChangli Gao tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); 40959a568de4SEric Dumazet tcp_mstamp_refresh(tp); 40969d0c00f5SEric Dumazet tp->retrans_stamp = tcp_time_stamp_ts(tp); 4097783237e8SYuchung Cheng tcp_connect_queue_skb(sk, buff); 4098735d3831SFlorian Westphal tcp_ecn_send_syn(sk, buff); 409975c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, buff); 41001da177e4SLinus Torvalds 4101783237e8SYuchung Cheng /* Send off SYN; include data in Fast Open. */ 4102783237e8SYuchung Cheng err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : 4103783237e8SYuchung Cheng tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); 4104ee586811SEric Paris if (err == -ECONNREFUSED) 4105ee586811SEric Paris return err; 4106bd37a088SWei Yongjun 4107bd37a088SWei Yongjun /* We change tp->snd_nxt after the tcp_transmit_skb() call 4108bd37a088SWei Yongjun * in order to make this packet get counted in tcpOutSegs. 4109bd37a088SWei Yongjun */ 4110e0d694d6SEric Dumazet WRITE_ONCE(tp->snd_nxt, tp->write_seq); 4111bd37a088SWei Yongjun tp->pushed_seq = tp->write_seq; 4112b5b7db8dSEric Dumazet buff = tcp_send_head(sk); 4113b5b7db8dSEric Dumazet if (unlikely(buff)) { 4114e0d694d6SEric Dumazet WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(buff)->seq); 4115b5b7db8dSEric Dumazet tp->pushed_seq = TCP_SKB_CB(buff)->seq; 4116b5b7db8dSEric Dumazet } 411781cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); 41181da177e4SLinus Torvalds 41191da177e4SLinus Torvalds /* Timer for repeating the SYN until an answer. */ 41203f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 41213f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 41221da177e4SLinus Torvalds return 0; 41231da177e4SLinus Torvalds } 41244bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_connect); 41251da177e4SLinus Torvalds 4126bbf80d71SEric Dumazet u32 tcp_delack_max(const struct sock *sk) 4127bbf80d71SEric Dumazet { 4128bbf80d71SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 4129bbf80d71SEric Dumazet u32 delack_max = inet_csk(sk)->icsk_delack_max; 4130bbf80d71SEric Dumazet 4131bbf80d71SEric Dumazet if (dst && dst_metric_locked(dst, RTAX_RTO_MIN)) { 4132bbf80d71SEric Dumazet u32 rto_min = dst_metric_rtt(dst, RTAX_RTO_MIN); 4133bbf80d71SEric Dumazet u32 delack_from_rto_min = max_t(int, 1, rto_min - 1); 4134bbf80d71SEric Dumazet 4135bbf80d71SEric Dumazet delack_max = min_t(u32, delack_max, delack_from_rto_min); 4136bbf80d71SEric Dumazet } 4137bbf80d71SEric Dumazet return delack_max; 4138bbf80d71SEric Dumazet } 4139bbf80d71SEric Dumazet 41401da177e4SLinus Torvalds /* Send out a delayed ack, the caller does the policy checking 41411da177e4SLinus Torvalds * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 41421da177e4SLinus Torvalds * for details. 41431da177e4SLinus Torvalds */ 41441da177e4SLinus Torvalds void tcp_send_delayed_ack(struct sock *sk) 41451da177e4SLinus Torvalds { 4146463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 4147463c84b9SArnaldo Carvalho de Melo int ato = icsk->icsk_ack.ato; 41481da177e4SLinus Torvalds unsigned long timeout; 41491da177e4SLinus Torvalds 41501da177e4SLinus Torvalds if (ato > TCP_DELACK_MIN) { 4151463c84b9SArnaldo Carvalho de Melo const struct tcp_sock *tp = tcp_sk(sk); 41521da177e4SLinus Torvalds int max_ato = HZ / 2; 41531da177e4SLinus Torvalds 415431954cd8SWei Wang if (inet_csk_in_pingpong_mode(sk) || 4155056834d9SIlpo Järvinen (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 41561da177e4SLinus Torvalds max_ato = TCP_DELACK_MAX; 41571da177e4SLinus Torvalds 41581da177e4SLinus Torvalds /* Slow path, intersegment interval is "high". */ 41591da177e4SLinus Torvalds 41601da177e4SLinus Torvalds /* If some rtt estimate is known, use it to bound delayed ack. 4161463c84b9SArnaldo Carvalho de Melo * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 41621da177e4SLinus Torvalds * directly. 41631da177e4SLinus Torvalds */ 4164740b0f18SEric Dumazet if (tp->srtt_us) { 4165740b0f18SEric Dumazet int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3), 4166740b0f18SEric Dumazet TCP_DELACK_MIN); 41671da177e4SLinus Torvalds 41681da177e4SLinus Torvalds if (rtt < max_ato) 41691da177e4SLinus Torvalds max_ato = rtt; 41701da177e4SLinus Torvalds } 41711da177e4SLinus Torvalds 41721da177e4SLinus Torvalds ato = min(ato, max_ato); 41731da177e4SLinus Torvalds } 41741da177e4SLinus Torvalds 4175bbf80d71SEric Dumazet ato = min_t(u32, ato, tcp_delack_max(sk)); 41762b8ee4f0SMartin KaFai Lau 41771da177e4SLinus Torvalds /* Stay within the limit we were given */ 41781da177e4SLinus Torvalds timeout = jiffies + ato; 41791da177e4SLinus Torvalds 41801da177e4SLinus Torvalds /* Use new timeout only if there wasn't a older one earlier. */ 4181463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 4182b6b6d653SEric Dumazet /* If delack timer is about to expire, send ACK now. */ 4183b6b6d653SEric Dumazet if (time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 41841da177e4SLinus Torvalds tcp_send_ack(sk); 41851da177e4SLinus Torvalds return; 41861da177e4SLinus Torvalds } 41871da177e4SLinus Torvalds 4188463c84b9SArnaldo Carvalho de Melo if (!time_before(timeout, icsk->icsk_ack.timeout)) 4189463c84b9SArnaldo Carvalho de Melo timeout = icsk->icsk_ack.timeout; 41901da177e4SLinus Torvalds } 4191463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 4192463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.timeout = timeout; 4193463c84b9SArnaldo Carvalho de Melo sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 41941da177e4SLinus Torvalds } 41951da177e4SLinus Torvalds 41961da177e4SLinus Torvalds /* This routine sends an ack and also updates the window. */ 41972987babbSYuchung Cheng void __tcp_send_ack(struct sock *sk, u32 rcv_nxt) 41981da177e4SLinus Torvalds { 41991da177e4SLinus Torvalds struct sk_buff *buff; 42001da177e4SLinus Torvalds 4201058dc334SIlpo Järvinen /* If we have been reset, we may not send again. */ 4202058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 4203058dc334SIlpo Järvinen return; 4204058dc334SIlpo Järvinen 42051da177e4SLinus Torvalds /* We are not putting this on the write queue, so 42061da177e4SLinus Torvalds * tcp_transmit_skb() will set the ownership to this 42071da177e4SLinus Torvalds * sock. 42081da177e4SLinus Torvalds */ 42097450aaf6SEric Dumazet buff = alloc_skb(MAX_TCP_HEADER, 42107450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 42117450aaf6SEric Dumazet if (unlikely(!buff)) { 4212a37c2134SEric Dumazet struct inet_connection_sock *icsk = inet_csk(sk); 4213a37c2134SEric Dumazet unsigned long delay; 4214a37c2134SEric Dumazet 4215a37c2134SEric Dumazet delay = TCP_DELACK_MAX << icsk->icsk_ack.retry; 4216a37c2134SEric Dumazet if (delay < TCP_RTO_MAX) 4217a37c2134SEric Dumazet icsk->icsk_ack.retry++; 4218463c84b9SArnaldo Carvalho de Melo inet_csk_schedule_ack(sk); 4219a37c2134SEric Dumazet icsk->icsk_ack.ato = TCP_ATO_MIN; 4220a37c2134SEric Dumazet inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, delay, TCP_RTO_MAX); 42211da177e4SLinus Torvalds return; 42221da177e4SLinus Torvalds } 42231da177e4SLinus Torvalds 42241da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 42251da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 4226a3433f35SChangli Gao tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); 42271da177e4SLinus Torvalds 422898781965SEric Dumazet /* We do not want pure acks influencing TCP Small Queues or fq/pacing 422998781965SEric Dumazet * too much. 423098781965SEric Dumazet * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784 423198781965SEric Dumazet */ 423298781965SEric Dumazet skb_set_tcp_pure_ack(buff); 423398781965SEric Dumazet 42341da177e4SLinus Torvalds /* Send it off, this clears delayed acks for us. */ 42352987babbSYuchung Cheng __tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0, rcv_nxt); 42361da177e4SLinus Torvalds } 423727cde44aSYuchung Cheng EXPORT_SYMBOL_GPL(__tcp_send_ack); 42382987babbSYuchung Cheng 42392987babbSYuchung Cheng void tcp_send_ack(struct sock *sk) 42402987babbSYuchung Cheng { 42412987babbSYuchung Cheng __tcp_send_ack(sk, tcp_sk(sk)->rcv_nxt); 42421da177e4SLinus Torvalds } 42431da177e4SLinus Torvalds 42441da177e4SLinus Torvalds /* This routine sends a packet with an out of date sequence 42451da177e4SLinus Torvalds * number. It assumes the other end will try to ack it. 42461da177e4SLinus Torvalds * 42471da177e4SLinus Torvalds * Question: what should we make while urgent mode? 42481da177e4SLinus Torvalds * 4.4BSD forces sending single byte of data. We cannot send 42491da177e4SLinus Torvalds * out of window data, because we have SND.NXT==SND.MAX... 42501da177e4SLinus Torvalds * 42511da177e4SLinus Torvalds * Current solution: to send TWO zero-length segments in urgent mode: 42521da177e4SLinus Torvalds * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 42531da177e4SLinus Torvalds * out-of-date with SND.UNA-1 to probe window. 42541da177e4SLinus Torvalds */ 4255e520af48SEric Dumazet static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib) 42561da177e4SLinus Torvalds { 42571da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 42581da177e4SLinus Torvalds struct sk_buff *skb; 42591da177e4SLinus Torvalds 42601da177e4SLinus Torvalds /* We don't queue it, tcp_transmit_skb() sets ownership. */ 42617450aaf6SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, 42627450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 426351456b29SIan Morris if (!skb) 42641da177e4SLinus Torvalds return -1; 42651da177e4SLinus Torvalds 42661da177e4SLinus Torvalds /* Reserve space for headers and set control bits. */ 42671da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 42681da177e4SLinus Torvalds /* Use a previous sequence. This should cause the other 42691da177e4SLinus Torvalds * end to send an ack. Don't queue or clone SKB, just 42701da177e4SLinus Torvalds * send it. 42711da177e4SLinus Torvalds */ 4272a3433f35SChangli Gao tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); 4273e2e8009fSRenato Westphal NET_INC_STATS(sock_net(sk), mib); 42747450aaf6SEric Dumazet return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0); 42751da177e4SLinus Torvalds } 42761da177e4SLinus Torvalds 4277385e2070SEric Dumazet /* Called from setsockopt( ... TCP_REPAIR ) */ 4278ee995283SPavel Emelyanov void tcp_send_window_probe(struct sock *sk) 4279ee995283SPavel Emelyanov { 4280ee995283SPavel Emelyanov if (sk->sk_state == TCP_ESTABLISHED) { 4281ee995283SPavel Emelyanov tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; 42829a568de4SEric Dumazet tcp_mstamp_refresh(tcp_sk(sk)); 4283e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE); 4284ee995283SPavel Emelyanov } 4285ee995283SPavel Emelyanov } 4286ee995283SPavel Emelyanov 428767edfef7SAndi Kleen /* Initiate keepalive or window probe from timer. */ 4288e520af48SEric Dumazet int tcp_write_wakeup(struct sock *sk, int mib) 42891da177e4SLinus Torvalds { 42901da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 42911da177e4SLinus Torvalds struct sk_buff *skb; 42921da177e4SLinus Torvalds 4293058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 4294058dc334SIlpo Järvinen return -1; 4295058dc334SIlpo Järvinen 429600db4124SIan Morris skb = tcp_send_head(sk); 429700db4124SIan Morris if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { 42981da177e4SLinus Torvalds int err; 42990c54b85fSIlpo Järvinen unsigned int mss = tcp_current_mss(sk); 430090840defSIlpo Järvinen unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 43011da177e4SLinus Torvalds 43021da177e4SLinus Torvalds if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 43031da177e4SLinus Torvalds tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 43041da177e4SLinus Torvalds 43051da177e4SLinus Torvalds /* We are probing the opening of a window 43061da177e4SLinus Torvalds * but the window size is != 0 43071da177e4SLinus Torvalds * must have been a result SWS avoidance ( sender ) 43081da177e4SLinus Torvalds */ 43091da177e4SLinus Torvalds if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 43101da177e4SLinus Torvalds skb->len > mss) { 43111da177e4SLinus Torvalds seg_size = min(seg_size, mss); 43124de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 431375c119afSEric Dumazet if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE, 431475c119afSEric Dumazet skb, seg_size, mss, GFP_ATOMIC)) 43151da177e4SLinus Torvalds return -1; 43161da177e4SLinus Torvalds } else if (!tcp_skb_pcount(skb)) 43175bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss); 43181da177e4SLinus Torvalds 43194de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 4320dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 432166f5fe62SIlpo Järvinen if (!err) 432266f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 43231da177e4SLinus Torvalds return err; 43241da177e4SLinus Torvalds } else { 432533f5f57eSIlpo Järvinen if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) 4326e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 1, mib); 4327e520af48SEric Dumazet return tcp_xmit_probe_skb(sk, 0, mib); 43281da177e4SLinus Torvalds } 43291da177e4SLinus Torvalds } 43301da177e4SLinus Torvalds 43311da177e4SLinus Torvalds /* A window probe timeout has occurred. If window is not closed send 43321da177e4SLinus Torvalds * a partial packet else a zero probe. 43331da177e4SLinus Torvalds */ 43341da177e4SLinus Torvalds void tcp_send_probe0(struct sock *sk) 43351da177e4SLinus Torvalds { 4336463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 43371da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 4338c6214a97SNikolay Borisov struct net *net = sock_net(sk); 4339c1d5674fSYuchung Cheng unsigned long timeout; 43401da177e4SLinus Torvalds int err; 43411da177e4SLinus Torvalds 4342e520af48SEric Dumazet err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE); 43431da177e4SLinus Torvalds 434475c119afSEric Dumazet if (tp->packets_out || tcp_write_queue_empty(sk)) { 43451da177e4SLinus Torvalds /* Cancel probe timer, if it is not required. */ 43466687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 0; 4347463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff = 0; 43489d9b1ee0SEnke Chen icsk->icsk_probes_tstamp = 0; 43491da177e4SLinus Torvalds return; 43501da177e4SLinus Torvalds } 43511da177e4SLinus Torvalds 4352c1d5674fSYuchung Cheng icsk->icsk_probes_out++; 43531da177e4SLinus Torvalds if (err <= 0) { 435439e24435SKuniyuki Iwashima if (icsk->icsk_backoff < READ_ONCE(net->ipv4.sysctl_tcp_retries2)) 4355463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff++; 4356c1d5674fSYuchung Cheng timeout = tcp_probe0_when(sk, TCP_RTO_MAX); 43571da177e4SLinus Torvalds } else { 43581da177e4SLinus Torvalds /* If packet was not sent due to local congestion, 4359c1d5674fSYuchung Cheng * Let senders fight for local resources conservatively. 43601da177e4SLinus Torvalds */ 4361c1d5674fSYuchung Cheng timeout = TCP_RESOURCE_PROBE_INTERVAL; 43621da177e4SLinus Torvalds } 4363344db93aSEnke Chen 4364344db93aSEnke Chen timeout = tcp_clamp_probe0_to_user_timeout(sk, timeout); 43658dc242adSEric Dumazet tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, timeout, TCP_RTO_MAX); 43661da177e4SLinus Torvalds } 43675db92c99SOctavian Purdila 4368ea3bea3aSEric Dumazet int tcp_rtx_synack(const struct sock *sk, struct request_sock *req) 43695db92c99SOctavian Purdila { 43705db92c99SOctavian Purdila const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific; 43715db92c99SOctavian Purdila struct flowi fl; 43725db92c99SOctavian Purdila int res; 43735db92c99SOctavian Purdila 4374cb6cd2ceSAkhmat Karakotov /* Paired with WRITE_ONCE() in sock_setsockopt() */ 4375cb6cd2ceSAkhmat Karakotov if (READ_ONCE(sk->sk_txrehash) == SOCK_TXREHASH_ENABLED) 43765e526552SEric Dumazet WRITE_ONCE(tcp_rsk(req)->txhash, net_tx_rndhash()); 4377331fca43SMartin KaFai Lau res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL, 4378331fca43SMartin KaFai Lau NULL); 43795db92c99SOctavian Purdila if (!res) { 43800a375c82SEric Dumazet TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); 43810a375c82SEric Dumazet NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 4382e9d9da91SEric Dumazet if (unlikely(tcp_passive_fastopen(sk))) { 4383e9d9da91SEric Dumazet /* sk has const attribute because listeners are lockless. 4384e9d9da91SEric Dumazet * However in this case, we are dealing with a passive fastopen 4385e9d9da91SEric Dumazet * socket thus we can change total_retrans value. 4386e9d9da91SEric Dumazet */ 4387e9d9da91SEric Dumazet tcp_sk_rw(sk)->total_retrans++; 4388e9d9da91SEric Dumazet } 4389cf34ce3dSSong Liu trace_tcp_retransmit_synack(sk, req); 43905db92c99SOctavian Purdila } 43915db92c99SOctavian Purdila return res; 43925db92c99SOctavian Purdila } 43935db92c99SOctavian Purdila EXPORT_SYMBOL(tcp_rtx_synack); 4394