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> 471da177e4SLinus Torvalds 48e086101bSCong Wang #include <trace/events/tcp.h> 4935089bb2SDavid S. Miller 509799ccb0SEric Dumazet /* Refresh clocks of a TCP socket, 519799ccb0SEric Dumazet * ensuring monotically increasing values. 529799ccb0SEric Dumazet */ 539799ccb0SEric Dumazet void tcp_mstamp_refresh(struct tcp_sock *tp) 549799ccb0SEric Dumazet { 559799ccb0SEric Dumazet u64 val = tcp_clock_ns(); 569799ccb0SEric Dumazet 575f6188a8SEric Dumazet tp->tcp_clock_cache = val; 58e6d14070SEric Dumazet tp->tcp_mstamp = div_u64(val, NSEC_PER_USEC); 599799ccb0SEric Dumazet } 609799ccb0SEric Dumazet 6146d3ceabSEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 6246d3ceabSEric Dumazet int push_one, gfp_t gfp); 63519855c5SWilliam Allen Simpson 6467edfef7SAndi Kleen /* Account for new data that has been sent to the network. */ 6575c119afSEric Dumazet static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb) 666ff03ac3SIlpo Järvinen { 676ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 686ff03ac3SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 6966f5fe62SIlpo Järvinen unsigned int prior_packets = tp->packets_out; 709e412ba7SIlpo Järvinen 71e0d694d6SEric Dumazet WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(skb)->end_seq); 728512430eSIlpo Järvinen 7375c119afSEric Dumazet __skb_unlink(skb, &sk->sk_write_queue); 7475c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, skb); 7575c119afSEric Dumazet 7685369750SCambda Zhu if (tp->highest_sack == NULL) 7785369750SCambda Zhu tp->highest_sack = skb; 7885369750SCambda Zhu 7966f5fe62SIlpo Järvinen tp->packets_out += tcp_skb_pcount(skb); 80bec41a11SYuchung Cheng if (!prior_packets || icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) 81750ea2baSYuchung Cheng tcp_rearm_rto(sk); 82f19c29e3SYuchung Cheng 83f7324acdSDavid S. Miller NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT, 84f19c29e3SYuchung Cheng tcp_skb_pcount(skb)); 854bfe744fSEric Dumazet tcp_check_space(sk); 866a5dc9e5SEric Dumazet } 871da177e4SLinus Torvalds 88a4ecb15aSCui, Cheng /* SND.NXT, if window was not shrunk or the amount of shrunk was less than one 89a4ecb15aSCui, Cheng * window scaling factor due to loss of precision. 901da177e4SLinus Torvalds * If window has been shrunk, what should we make? It is not clear at all. 911da177e4SLinus Torvalds * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-( 921da177e4SLinus Torvalds * Anything in between SND.UNA...SND.UNA+SND.WND also can be already 931da177e4SLinus Torvalds * invalid. OK, let's make this for now: 941da177e4SLinus Torvalds */ 95cf533ea5SEric Dumazet static inline __u32 tcp_acceptable_seq(const struct sock *sk) 961da177e4SLinus Torvalds { 97cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 989e412ba7SIlpo Järvinen 99a4ecb15aSCui, Cheng if (!before(tcp_wnd_end(tp), tp->snd_nxt) || 100a4ecb15aSCui, Cheng (tp->rx_opt.wscale_ok && 101a4ecb15aSCui, Cheng ((tp->snd_nxt - tcp_wnd_end(tp)) < (1 << tp->rx_opt.rcv_wscale)))) 1021da177e4SLinus Torvalds return tp->snd_nxt; 1031da177e4SLinus Torvalds else 10490840defSIlpo Järvinen return tcp_wnd_end(tp); 1051da177e4SLinus Torvalds } 1061da177e4SLinus Torvalds 1071da177e4SLinus Torvalds /* Calculate mss to advertise in SYN segment. 1081da177e4SLinus Torvalds * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that: 1091da177e4SLinus Torvalds * 1101da177e4SLinus Torvalds * 1. It is independent of path mtu. 1111da177e4SLinus Torvalds * 2. Ideally, it is maximal possible segment size i.e. 65535-40. 1121da177e4SLinus Torvalds * 3. For IPv4 it is reasonable to calculate it from maximal MTU of 1131da177e4SLinus Torvalds * attached devices, because some buggy hosts are confused by 1141da177e4SLinus Torvalds * large MSS. 1151da177e4SLinus Torvalds * 4. We do not make 3, we advertise MSS, calculated from first 1161da177e4SLinus Torvalds * hop device mtu, but allow to raise it to ip_rt_min_advmss. 1171da177e4SLinus Torvalds * This may be overridden via information stored in routing table. 1181da177e4SLinus Torvalds * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible, 1191da177e4SLinus Torvalds * probably even Jumbo". 1201da177e4SLinus Torvalds */ 1211da177e4SLinus Torvalds static __u16 tcp_advertise_mss(struct sock *sk) 1221da177e4SLinus Torvalds { 1231da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 124cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1251da177e4SLinus Torvalds int mss = tp->advmss; 1261da177e4SLinus Torvalds 1270dbaee3bSDavid S. Miller if (dst) { 1280dbaee3bSDavid S. Miller unsigned int metric = dst_metric_advmss(dst); 1290dbaee3bSDavid S. Miller 1300dbaee3bSDavid S. Miller if (metric < mss) { 1310dbaee3bSDavid S. Miller mss = metric; 1321da177e4SLinus Torvalds tp->advmss = mss; 1331da177e4SLinus Torvalds } 1340dbaee3bSDavid S. Miller } 1351da177e4SLinus Torvalds 1361da177e4SLinus Torvalds return (__u16)mss; 1371da177e4SLinus Torvalds } 1381da177e4SLinus Torvalds 1391da177e4SLinus Torvalds /* RFC2861. Reset CWND after idle period longer RTO to "restart window". 1406f021c62SEric Dumazet * This is the first part of cwnd validation mechanism. 1416f021c62SEric Dumazet */ 1426f021c62SEric Dumazet void tcp_cwnd_restart(struct sock *sk, s32 delta) 1431da177e4SLinus Torvalds { 144463c84b9SArnaldo Carvalho de Melo struct tcp_sock *tp = tcp_sk(sk); 1456f021c62SEric Dumazet u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk)); 14640570375SEric Dumazet u32 cwnd = tcp_snd_cwnd(tp); 1471da177e4SLinus Torvalds 1486687e988SArnaldo Carvalho de Melo tcp_ca_event(sk, CA_EVENT_CWND_RESTART); 1491da177e4SLinus Torvalds 1506687e988SArnaldo Carvalho de Melo tp->snd_ssthresh = tcp_current_ssthresh(sk); 1511da177e4SLinus Torvalds restart_cwnd = min(restart_cwnd, cwnd); 1521da177e4SLinus Torvalds 153463c84b9SArnaldo Carvalho de Melo while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd) 1541da177e4SLinus Torvalds cwnd >>= 1; 15540570375SEric Dumazet tcp_snd_cwnd_set(tp, max(cwnd, restart_cwnd)); 156c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 1571da177e4SLinus Torvalds tp->snd_cwnd_used = 0; 1581da177e4SLinus Torvalds } 1591da177e4SLinus Torvalds 16067edfef7SAndi Kleen /* Congestion state accounting after a packet has been sent. */ 16140efc6faSStephen Hemminger static void tcp_event_data_sent(struct tcp_sock *tp, 162cf533ea5SEric Dumazet struct sock *sk) 1631da177e4SLinus Torvalds { 164463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 165d635fbe2SEric Dumazet const u32 now = tcp_jiffies32; 1661da177e4SLinus Torvalds 16705c5a46dSNeal Cardwell if (tcp_packets_in_flight(tp) == 0) 16805c5a46dSNeal Cardwell tcp_ca_event(sk, CA_EVENT_TX_START); 16905c5a46dSNeal Cardwell 1704a41f453SWei Wang tp->lsndtime = now; 1714d8f24eeSWei Wang 1724d8f24eeSWei Wang /* If it is a reply for ato after last received 173562b1fdfSHaiyang Zhang * packet, increase pingpong count. 1744d8f24eeSWei Wang */ 1754d8f24eeSWei Wang if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato) 176562b1fdfSHaiyang Zhang inet_csk_inc_pingpong_cnt(sk); 1771da177e4SLinus Torvalds } 1781da177e4SLinus Torvalds 17967edfef7SAndi Kleen /* Account for an ACK we sent. */ 180059217c1SNeal Cardwell static inline void tcp_event_ack_sent(struct sock *sk, u32 rcv_nxt) 1811da177e4SLinus Torvalds { 1825d9f4262SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 1835d9f4262SEric Dumazet 1842b195850SEric Dumazet if (unlikely(tp->compressed_ack)) { 185200d95f4SEric Dumazet NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED, 1862b195850SEric Dumazet tp->compressed_ack); 1872b195850SEric Dumazet tp->compressed_ack = 0; 1885d9f4262SEric Dumazet if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1) 1895d9f4262SEric Dumazet __sock_put(sk); 1905d9f4262SEric Dumazet } 19127cde44aSYuchung Cheng 19227cde44aSYuchung Cheng if (unlikely(rcv_nxt != tp->rcv_nxt)) 19327cde44aSYuchung Cheng return; /* Special ACK sent by DCTCP to reflect ECN */ 194059217c1SNeal Cardwell tcp_dec_quickack_mode(sk); 195463c84b9SArnaldo Carvalho de Melo inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); 1961da177e4SLinus Torvalds } 1971da177e4SLinus Torvalds 1981da177e4SLinus Torvalds /* Determine a window scaling and initial window to offer. 1991da177e4SLinus Torvalds * Based on the assumption that the given amount of space 2001da177e4SLinus Torvalds * will be offered. Store the results in the tp structure. 2011da177e4SLinus Torvalds * NOTE: for smooth operation initial space offering should 2021da177e4SLinus Torvalds * be a multiple of mss if possible. We assume here that mss >= 1. 2031da177e4SLinus Torvalds * This MUST be enforced by all callers. 2041da177e4SLinus Torvalds */ 205ceef9ab6SEric Dumazet void tcp_select_initial_window(const struct sock *sk, int __space, __u32 mss, 206*f410cbeaSEric Dumazet __u32 *rcv_wnd, __u32 *__window_clamp, 20731d12926Slaurent chavey int wscale_ok, __u8 *rcv_wscale, 20831d12926Slaurent chavey __u32 init_rcv_wnd) 2091da177e4SLinus Torvalds { 2101da177e4SLinus Torvalds unsigned int space = (__space < 0 ? 0 : __space); 211*f410cbeaSEric Dumazet u32 window_clamp = READ_ONCE(*__window_clamp); 2121da177e4SLinus Torvalds 2131da177e4SLinus Torvalds /* If no clamp set the clamp to the max possible scaled window */ 214*f410cbeaSEric Dumazet if (window_clamp == 0) 215*f410cbeaSEric Dumazet window_clamp = (U16_MAX << TCP_MAX_WSCALE); 216*f410cbeaSEric Dumazet space = min(window_clamp, space); 2171da177e4SLinus Torvalds 2181da177e4SLinus Torvalds /* Quantize space offering to a multiple of mss if possible. */ 2191da177e4SLinus Torvalds if (space > mss) 220589c49cbSGao Feng space = rounddown(space, mss); 2211da177e4SLinus Torvalds 2221da177e4SLinus Torvalds /* NOTE: offering an initial window larger than 32767 22315d99e02SRick Jones * will break some buggy TCP stacks. If the admin tells us 22415d99e02SRick Jones * it is likely we could be speaking with such a buggy stack 22515d99e02SRick Jones * we will truncate our initial window offering to 32K-1 22615d99e02SRick Jones * unless the remote has sent us a window scaling option, 22715d99e02SRick Jones * which we interpret as a sign the remote TCP is not 22815d99e02SRick Jones * misinterpreting the window field as a signed quantity. 2291da177e4SLinus Torvalds */ 2300f1e4d06SKuniyuki Iwashima if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows)) 2311da177e4SLinus Torvalds (*rcv_wnd) = min(space, MAX_TCP_WINDOW); 23215d99e02SRick Jones else 233a337531bSYuchung Cheng (*rcv_wnd) = min_t(u32, space, U16_MAX); 234a337531bSYuchung Cheng 235a337531bSYuchung Cheng if (init_rcv_wnd) 236a337531bSYuchung Cheng *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss); 23715d99e02SRick Jones 23819bf6261SEric Dumazet *rcv_wscale = 0; 2391da177e4SLinus Torvalds if (wscale_ok) { 240589c49cbSGao Feng /* Set window scaling on max possible window */ 24102739545SKuniyuki Iwashima space = max_t(u32, space, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2])); 2421227c177SKuniyuki Iwashima space = max_t(u32, space, READ_ONCE(sysctl_rmem_max)); 243*f410cbeaSEric Dumazet space = min_t(u32, space, window_clamp); 24419bf6261SEric Dumazet *rcv_wscale = clamp_t(int, ilog2(space) - 15, 24519bf6261SEric Dumazet 0, TCP_MAX_WSCALE); 2461da177e4SLinus Torvalds } 2471da177e4SLinus Torvalds /* Set the clamp no higher than max representable value */ 248*f410cbeaSEric Dumazet WRITE_ONCE(*__window_clamp, 249*f410cbeaSEric Dumazet min_t(__u32, U16_MAX << (*rcv_wscale), window_clamp)); 2501da177e4SLinus Torvalds } 2514bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_select_initial_window); 2521da177e4SLinus Torvalds 2531da177e4SLinus Torvalds /* Chose a new window to advertise, update state in tcp_sock for the 2541da177e4SLinus Torvalds * socket, and return result with RFC1323 scaling applied. The return 2551da177e4SLinus Torvalds * value can be stuffed directly into th->window for an outgoing 2561da177e4SLinus Torvalds * frame. 2571da177e4SLinus Torvalds */ 25840efc6faSStephen Hemminger static u16 tcp_select_window(struct sock *sk) 2591da177e4SLinus Torvalds { 2601da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 261b650d953Smfreemon@cloudflare.com struct net *net = sock_net(sk); 262e2142825SMenglong Dong u32 old_win = tp->rcv_wnd; 263e2142825SMenglong Dong u32 cur_win, new_win; 2641da177e4SLinus Torvalds 265e2142825SMenglong Dong /* Make the window 0 if we failed to queue the data because we 266e2142825SMenglong Dong * are out of memory. The window is temporary, so we don't store 267e2142825SMenglong Dong * it on the socket. 268e2142825SMenglong Dong */ 269e2142825SMenglong Dong if (unlikely(inet_csk(sk)->icsk_ack.pending & ICSK_ACK_NOMEM)) 270e2142825SMenglong Dong return 0; 271e2142825SMenglong Dong 272e2142825SMenglong Dong cur_win = tcp_receive_window(tp); 273e2142825SMenglong Dong new_win = __tcp_select_window(sk); 2741da177e4SLinus Torvalds if (new_win < cur_win) { 2751da177e4SLinus Torvalds /* Danger Will Robinson! 2761da177e4SLinus Torvalds * Don't update rcv_wup/rcv_wnd here or else 2771da177e4SLinus Torvalds * we will not be able to advertise a zero 2781da177e4SLinus Torvalds * window in time. --DaveM 2791da177e4SLinus Torvalds * 2801da177e4SLinus Torvalds * Relax Will Robinson. 2811da177e4SLinus Torvalds */ 282b650d953Smfreemon@cloudflare.com if (!READ_ONCE(net->ipv4.sysctl_tcp_shrink_window) || !tp->rx_opt.rcv_wscale) { 283b650d953Smfreemon@cloudflare.com /* Never shrink the offered window */ 2848e165e20SFlorian Westphal if (new_win == 0) 285b650d953Smfreemon@cloudflare.com NET_INC_STATS(net, LINUX_MIB_TCPWANTZEROWINDOWADV); 286607bfbf2SPatrick McHardy new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale); 2871da177e4SLinus Torvalds } 288b650d953Smfreemon@cloudflare.com } 289b650d953Smfreemon@cloudflare.com 2901da177e4SLinus Torvalds tp->rcv_wnd = new_win; 2911da177e4SLinus Torvalds tp->rcv_wup = tp->rcv_nxt; 2921da177e4SLinus Torvalds 2931da177e4SLinus Torvalds /* Make sure we do not exceed the maximum possible 2941da177e4SLinus Torvalds * scaled window. 2951da177e4SLinus Torvalds */ 296ceef9ab6SEric Dumazet if (!tp->rx_opt.rcv_wscale && 297b650d953Smfreemon@cloudflare.com READ_ONCE(net->ipv4.sysctl_tcp_workaround_signed_windows)) 2981da177e4SLinus Torvalds new_win = min(new_win, MAX_TCP_WINDOW); 2991da177e4SLinus Torvalds else 3001da177e4SLinus Torvalds new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); 3011da177e4SLinus Torvalds 3021da177e4SLinus Torvalds /* RFC1323 scaling applied */ 3031da177e4SLinus Torvalds new_win >>= tp->rx_opt.rcv_wscale; 3041da177e4SLinus Torvalds 30531770e34SFlorian Westphal /* If we advertise zero window, disable fast path. */ 3068e165e20SFlorian Westphal if (new_win == 0) { 30731770e34SFlorian Westphal tp->pred_flags = 0; 3088e165e20SFlorian Westphal if (old_win) 309b650d953Smfreemon@cloudflare.com NET_INC_STATS(net, LINUX_MIB_TCPTOZEROWINDOWADV); 3108e165e20SFlorian Westphal } else if (old_win == 0) { 311b650d953Smfreemon@cloudflare.com NET_INC_STATS(net, LINUX_MIB_TCPFROMZEROWINDOWADV); 3128e165e20SFlorian Westphal } 3131da177e4SLinus Torvalds 3141da177e4SLinus Torvalds return new_win; 3151da177e4SLinus Torvalds } 3161da177e4SLinus Torvalds 31767edfef7SAndi Kleen /* Packet ECN state for a SYN-ACK */ 318735d3831SFlorian Westphal static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb) 319bdf1ee5dSIlpo Järvinen { 32030e502a3SDaniel Borkmann const struct tcp_sock *tp = tcp_sk(sk); 32130e502a3SDaniel Borkmann 3224de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR; 323bdf1ee5dSIlpo Järvinen if (!(tp->ecn_flags & TCP_ECN_OK)) 3244de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE; 32591b5b21cSLawrence Brakmo else if (tcp_ca_needs_ecn(sk) || 32691b5b21cSLawrence Brakmo tcp_bpf_ca_needs_ecn(sk)) 32730e502a3SDaniel Borkmann INET_ECN_xmit(sk); 328bdf1ee5dSIlpo Järvinen } 329bdf1ee5dSIlpo Järvinen 33067edfef7SAndi Kleen /* Packet ECN state for a SYN. */ 331735d3831SFlorian Westphal static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb) 332bdf1ee5dSIlpo Järvinen { 333bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 33491b5b21cSLawrence Brakmo bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk); 3354785a667SKuniyuki Iwashima bool use_ecn = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_ecn) == 1 || 33691b5b21cSLawrence Brakmo tcp_ca_needs_ecn(sk) || bpf_needs_ecn; 337f7b3bec6SFlorian Westphal 338f7b3bec6SFlorian Westphal if (!use_ecn) { 339f7b3bec6SFlorian Westphal const struct dst_entry *dst = __sk_dst_get(sk); 340f7b3bec6SFlorian Westphal 341f7b3bec6SFlorian Westphal if (dst && dst_feature(dst, RTAX_FEATURE_ECN)) 342f7b3bec6SFlorian Westphal use_ecn = true; 343f7b3bec6SFlorian Westphal } 344bdf1ee5dSIlpo Järvinen 345bdf1ee5dSIlpo Järvinen tp->ecn_flags = 0; 346f7b3bec6SFlorian Westphal 347f7b3bec6SFlorian Westphal if (use_ecn) { 3484de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR; 349bdf1ee5dSIlpo Järvinen tp->ecn_flags = TCP_ECN_OK; 35091b5b21cSLawrence Brakmo if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn) 35130e502a3SDaniel Borkmann INET_ECN_xmit(sk); 352bdf1ee5dSIlpo Järvinen } 353bdf1ee5dSIlpo Järvinen } 354bdf1ee5dSIlpo Järvinen 35549213555SDaniel Borkmann static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb) 35649213555SDaniel Borkmann { 35712b8d9caSKuniyuki Iwashima if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback)) 35849213555SDaniel Borkmann /* tp->ecn_flags are cleared at a later point in time when 35949213555SDaniel Borkmann * SYN ACK is ultimatively being received. 36049213555SDaniel Borkmann */ 36149213555SDaniel Borkmann TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR); 36249213555SDaniel Borkmann } 36349213555SDaniel Borkmann 364735d3831SFlorian Westphal static void 3656ac705b1SEric Dumazet tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th) 366bdf1ee5dSIlpo Järvinen { 3676ac705b1SEric Dumazet if (inet_rsk(req)->ecn_ok) 368bdf1ee5dSIlpo Järvinen th->ece = 1; 369bdf1ee5dSIlpo Järvinen } 370bdf1ee5dSIlpo Järvinen 37167edfef7SAndi Kleen /* Set up ECN state for a packet on a ESTABLISHED socket that is about to 37267edfef7SAndi Kleen * be sent. 37367edfef7SAndi Kleen */ 374735d3831SFlorian Westphal static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb, 375ea1627c2SEric Dumazet struct tcphdr *th, int tcp_header_len) 376bdf1ee5dSIlpo Järvinen { 377bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 378bdf1ee5dSIlpo Järvinen 379bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_OK) { 380bdf1ee5dSIlpo Järvinen /* Not-retransmitted data segment: set ECT and inject CWR. */ 381bdf1ee5dSIlpo Järvinen if (skb->len != tcp_header_len && 382bdf1ee5dSIlpo Järvinen !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { 383bdf1ee5dSIlpo Järvinen INET_ECN_xmit(sk); 384bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) { 385bdf1ee5dSIlpo Järvinen tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; 386ea1627c2SEric Dumazet th->cwr = 1; 387bdf1ee5dSIlpo Järvinen skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; 388bdf1ee5dSIlpo Järvinen } 38930e502a3SDaniel Borkmann } else if (!tcp_ca_needs_ecn(sk)) { 390bdf1ee5dSIlpo Järvinen /* ACK or retransmitted segment: clear ECT|CE */ 391bdf1ee5dSIlpo Järvinen INET_ECN_dontxmit(sk); 392bdf1ee5dSIlpo Järvinen } 393bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) 394ea1627c2SEric Dumazet th->ece = 1; 395bdf1ee5dSIlpo Järvinen } 396bdf1ee5dSIlpo Järvinen } 397bdf1ee5dSIlpo Järvinen 398e870a8efSIlpo Järvinen /* Constructs common control bits of non-data skb. If SYN/FIN is present, 399e870a8efSIlpo Järvinen * auto increment end seqno. 400e870a8efSIlpo Järvinen */ 401e870a8efSIlpo Järvinen static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags) 402e870a8efSIlpo Järvinen { 4032e8e18efSDavid S. Miller skb->ip_summed = CHECKSUM_PARTIAL; 404e870a8efSIlpo Järvinen 4054de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags; 406e870a8efSIlpo Järvinen 407cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, 1); 408e870a8efSIlpo Järvinen 409e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->seq = seq; 410a3433f35SChangli Gao if (flags & (TCPHDR_SYN | TCPHDR_FIN)) 411e870a8efSIlpo Järvinen seq++; 412e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->end_seq = seq; 413e870a8efSIlpo Järvinen } 414e870a8efSIlpo Järvinen 415a2a385d6SEric Dumazet static inline bool tcp_urg_mode(const struct tcp_sock *tp) 41633f5f57eSIlpo Järvinen { 41733f5f57eSIlpo Järvinen return tp->snd_una != tp->snd_up; 41833f5f57eSIlpo Järvinen } 41933f5f57eSIlpo Järvinen 4203b65abb8SLeonard Crestez #define OPTION_SACK_ADVERTISE BIT(0) 4213b65abb8SLeonard Crestez #define OPTION_TS BIT(1) 4223b65abb8SLeonard Crestez #define OPTION_MD5 BIT(2) 4233b65abb8SLeonard Crestez #define OPTION_WSCALE BIT(3) 4243b65abb8SLeonard Crestez #define OPTION_FAST_OPEN_COOKIE BIT(8) 4253b65abb8SLeonard Crestez #define OPTION_SMC BIT(9) 4263b65abb8SLeonard Crestez #define OPTION_MPTCP BIT(10) 4271e03d32bSDmitry Safonov #define OPTION_AO BIT(11) 42860e2a778SUrsula Braun 42960e2a778SUrsula Braun static void smc_options_write(__be32 *ptr, u16 *options) 43060e2a778SUrsula Braun { 43160e2a778SUrsula Braun #if IS_ENABLED(CONFIG_SMC) 43260e2a778SUrsula Braun if (static_branch_unlikely(&tcp_have_smc)) { 43360e2a778SUrsula Braun if (unlikely(OPTION_SMC & *options)) { 43460e2a778SUrsula Braun *ptr++ = htonl((TCPOPT_NOP << 24) | 43560e2a778SUrsula Braun (TCPOPT_NOP << 16) | 43660e2a778SUrsula Braun (TCPOPT_EXP << 8) | 43760e2a778SUrsula Braun (TCPOLEN_EXP_SMC_BASE)); 43860e2a778SUrsula Braun *ptr++ = htonl(TCPOPT_SMC_MAGIC); 43960e2a778SUrsula Braun } 44060e2a778SUrsula Braun } 44160e2a778SUrsula Braun #endif 44260e2a778SUrsula Braun } 44333ad798cSAdam Langley 44433ad798cSAdam Langley struct tcp_out_options { 4452100c8d2SYuchung Cheng u16 options; /* bit field of OPTION_* */ 4462100c8d2SYuchung Cheng u16 mss; /* 0 to disable */ 44733ad798cSAdam Langley u8 ws; /* window scale, 0 to disable */ 44833ad798cSAdam Langley u8 num_sack_blocks; /* number of SACK blocks to include */ 449bd0388aeSWilliam Allen Simpson u8 hash_size; /* bytes in hash_location */ 450331fca43SMartin KaFai Lau u8 bpf_opt_len; /* length of BPF hdr option */ 451bd0388aeSWilliam Allen Simpson __u8 *hash_location; /* temporary pointer, overloaded */ 4522100c8d2SYuchung Cheng __u32 tsval, tsecr; /* need to include OPTION_TS */ 4532100c8d2SYuchung Cheng struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */ 454eda7acddSPeter Krystad struct mptcp_out_options mptcp; 45533ad798cSAdam Langley }; 45633ad798cSAdam Langley 457ea66758cSPaolo Abeni static void mptcp_options_write(struct tcphdr *th, __be32 *ptr, 458ea66758cSPaolo Abeni struct tcp_sock *tp, 459fa3fe2b1SFlorian Westphal struct tcp_out_options *opts) 460eda7acddSPeter Krystad { 461eda7acddSPeter Krystad #if IS_ENABLED(CONFIG_MPTCP) 462eda7acddSPeter Krystad if (unlikely(OPTION_MPTCP & opts->options)) 463ea66758cSPaolo Abeni mptcp_write_options(th, ptr, tp, &opts->mptcp); 464eda7acddSPeter Krystad #endif 465eda7acddSPeter Krystad } 466eda7acddSPeter Krystad 467331fca43SMartin KaFai Lau #ifdef CONFIG_CGROUP_BPF 4680813a841SMartin KaFai Lau static int bpf_skops_write_hdr_opt_arg0(struct sk_buff *skb, 4690813a841SMartin KaFai Lau enum tcp_synack_type synack_type) 4700813a841SMartin KaFai Lau { 4710813a841SMartin KaFai Lau if (unlikely(!skb)) 4720813a841SMartin KaFai Lau return BPF_WRITE_HDR_TCP_CURRENT_MSS; 4730813a841SMartin KaFai Lau 4740813a841SMartin KaFai Lau if (unlikely(synack_type == TCP_SYNACK_COOKIE)) 4750813a841SMartin KaFai Lau return BPF_WRITE_HDR_TCP_SYNACK_COOKIE; 4760813a841SMartin KaFai Lau 4770813a841SMartin KaFai Lau return 0; 4780813a841SMartin KaFai Lau } 4790813a841SMartin KaFai Lau 480331fca43SMartin KaFai Lau /* req, syn_skb and synack_type are used when writing synack */ 481331fca43SMartin KaFai Lau static void bpf_skops_hdr_opt_len(struct sock *sk, struct sk_buff *skb, 482331fca43SMartin KaFai Lau struct request_sock *req, 483331fca43SMartin KaFai Lau struct sk_buff *syn_skb, 484331fca43SMartin KaFai Lau enum tcp_synack_type synack_type, 485331fca43SMartin KaFai Lau struct tcp_out_options *opts, 486331fca43SMartin KaFai Lau unsigned int *remaining) 487331fca43SMartin KaFai Lau { 4880813a841SMartin KaFai Lau struct bpf_sock_ops_kern sock_ops; 4890813a841SMartin KaFai Lau int err; 4900813a841SMartin KaFai Lau 491331fca43SMartin KaFai Lau if (likely(!BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), 492331fca43SMartin KaFai Lau BPF_SOCK_OPS_WRITE_HDR_OPT_CB_FLAG)) || 493331fca43SMartin KaFai Lau !*remaining) 494331fca43SMartin KaFai Lau return; 495331fca43SMartin KaFai Lau 4960813a841SMartin KaFai Lau /* *remaining has already been aligned to 4 bytes, so *remaining >= 4 */ 4970813a841SMartin KaFai Lau 4980813a841SMartin KaFai Lau /* init sock_ops */ 4990813a841SMartin KaFai Lau memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp)); 5000813a841SMartin KaFai Lau 5010813a841SMartin KaFai Lau sock_ops.op = BPF_SOCK_OPS_HDR_OPT_LEN_CB; 5020813a841SMartin KaFai Lau 5030813a841SMartin KaFai Lau if (req) { 5040813a841SMartin KaFai Lau /* The listen "sk" cannot be passed here because 5050813a841SMartin KaFai Lau * it is not locked. It would not make too much 5060813a841SMartin KaFai Lau * sense to do bpf_setsockopt(listen_sk) based 5070813a841SMartin KaFai Lau * on individual connection request also. 5080813a841SMartin KaFai Lau * 5090813a841SMartin KaFai Lau * Thus, "req" is passed here and the cgroup-bpf-progs 5100813a841SMartin KaFai Lau * of the listen "sk" will be run. 5110813a841SMartin KaFai Lau * 5120813a841SMartin KaFai Lau * "req" is also used here for fastopen even the "sk" here is 5130813a841SMartin KaFai Lau * a fullsock "child" sk. It is to keep the behavior 5140813a841SMartin KaFai Lau * consistent between fastopen and non-fastopen on 5150813a841SMartin KaFai Lau * the bpf programming side. 516331fca43SMartin KaFai Lau */ 5170813a841SMartin KaFai Lau sock_ops.sk = (struct sock *)req; 5180813a841SMartin KaFai Lau sock_ops.syn_skb = syn_skb; 5190813a841SMartin KaFai Lau } else { 5200813a841SMartin KaFai Lau sock_owned_by_me(sk); 5210813a841SMartin KaFai Lau 5220813a841SMartin KaFai Lau sock_ops.is_fullsock = 1; 5230813a841SMartin KaFai Lau sock_ops.sk = sk; 5240813a841SMartin KaFai Lau } 5250813a841SMartin KaFai Lau 5260813a841SMartin KaFai Lau sock_ops.args[0] = bpf_skops_write_hdr_opt_arg0(skb, synack_type); 5270813a841SMartin KaFai Lau sock_ops.remaining_opt_len = *remaining; 5280813a841SMartin KaFai Lau /* tcp_current_mss() does not pass a skb */ 5290813a841SMartin KaFai Lau if (skb) 5300813a841SMartin KaFai Lau bpf_skops_init_skb(&sock_ops, skb, 0); 5310813a841SMartin KaFai Lau 5320813a841SMartin KaFai Lau err = BPF_CGROUP_RUN_PROG_SOCK_OPS_SK(&sock_ops, sk); 5330813a841SMartin KaFai Lau 5340813a841SMartin KaFai Lau if (err || sock_ops.remaining_opt_len == *remaining) 5350813a841SMartin KaFai Lau return; 5360813a841SMartin KaFai Lau 5370813a841SMartin KaFai Lau opts->bpf_opt_len = *remaining - sock_ops.remaining_opt_len; 5380813a841SMartin KaFai Lau /* round up to 4 bytes */ 5390813a841SMartin KaFai Lau opts->bpf_opt_len = (opts->bpf_opt_len + 3) & ~3; 5400813a841SMartin KaFai Lau 5410813a841SMartin KaFai Lau *remaining -= opts->bpf_opt_len; 542331fca43SMartin KaFai Lau } 543331fca43SMartin KaFai Lau 544331fca43SMartin KaFai Lau static void bpf_skops_write_hdr_opt(struct sock *sk, struct sk_buff *skb, 545331fca43SMartin KaFai Lau struct request_sock *req, 546331fca43SMartin KaFai Lau struct sk_buff *syn_skb, 547331fca43SMartin KaFai Lau enum tcp_synack_type synack_type, 548331fca43SMartin KaFai Lau struct tcp_out_options *opts) 549331fca43SMartin KaFai Lau { 5500813a841SMartin KaFai Lau u8 first_opt_off, nr_written, max_opt_len = opts->bpf_opt_len; 5510813a841SMartin KaFai Lau struct bpf_sock_ops_kern sock_ops; 5520813a841SMartin KaFai Lau int err; 5530813a841SMartin KaFai Lau 5540813a841SMartin KaFai Lau if (likely(!max_opt_len)) 555331fca43SMartin KaFai Lau return; 556331fca43SMartin KaFai Lau 5570813a841SMartin KaFai Lau memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp)); 5580813a841SMartin KaFai Lau 5590813a841SMartin KaFai Lau sock_ops.op = BPF_SOCK_OPS_WRITE_HDR_OPT_CB; 5600813a841SMartin KaFai Lau 5610813a841SMartin KaFai Lau if (req) { 5620813a841SMartin KaFai Lau sock_ops.sk = (struct sock *)req; 5630813a841SMartin KaFai Lau sock_ops.syn_skb = syn_skb; 5640813a841SMartin KaFai Lau } else { 5650813a841SMartin KaFai Lau sock_owned_by_me(sk); 5660813a841SMartin KaFai Lau 5670813a841SMartin KaFai Lau sock_ops.is_fullsock = 1; 5680813a841SMartin KaFai Lau sock_ops.sk = sk; 5690813a841SMartin KaFai Lau } 5700813a841SMartin KaFai Lau 5710813a841SMartin KaFai Lau sock_ops.args[0] = bpf_skops_write_hdr_opt_arg0(skb, synack_type); 5720813a841SMartin KaFai Lau sock_ops.remaining_opt_len = max_opt_len; 5730813a841SMartin KaFai Lau first_opt_off = tcp_hdrlen(skb) - max_opt_len; 5740813a841SMartin KaFai Lau bpf_skops_init_skb(&sock_ops, skb, first_opt_off); 5750813a841SMartin KaFai Lau 5760813a841SMartin KaFai Lau err = BPF_CGROUP_RUN_PROG_SOCK_OPS_SK(&sock_ops, sk); 5770813a841SMartin KaFai Lau 5780813a841SMartin KaFai Lau if (err) 5790813a841SMartin KaFai Lau nr_written = 0; 5800813a841SMartin KaFai Lau else 5810813a841SMartin KaFai Lau nr_written = max_opt_len - sock_ops.remaining_opt_len; 5820813a841SMartin KaFai Lau 5830813a841SMartin KaFai Lau if (nr_written < max_opt_len) 5840813a841SMartin KaFai Lau memset(skb->data + first_opt_off + nr_written, TCPOPT_NOP, 5850813a841SMartin KaFai Lau max_opt_len - nr_written); 586331fca43SMartin KaFai Lau } 587331fca43SMartin KaFai Lau #else 588331fca43SMartin KaFai Lau static void bpf_skops_hdr_opt_len(struct sock *sk, struct sk_buff *skb, 589331fca43SMartin KaFai Lau struct request_sock *req, 590331fca43SMartin KaFai Lau struct sk_buff *syn_skb, 591331fca43SMartin KaFai Lau enum tcp_synack_type synack_type, 592331fca43SMartin KaFai Lau struct tcp_out_options *opts, 593331fca43SMartin KaFai Lau unsigned int *remaining) 594331fca43SMartin KaFai Lau { 595331fca43SMartin KaFai Lau } 596331fca43SMartin KaFai Lau 597331fca43SMartin KaFai Lau static void bpf_skops_write_hdr_opt(struct sock *sk, struct sk_buff *skb, 598331fca43SMartin KaFai Lau struct request_sock *req, 599331fca43SMartin KaFai Lau struct sk_buff *syn_skb, 600331fca43SMartin KaFai Lau enum tcp_synack_type synack_type, 601331fca43SMartin KaFai Lau struct tcp_out_options *opts) 602331fca43SMartin KaFai Lau { 603331fca43SMartin KaFai Lau } 604331fca43SMartin KaFai Lau #endif 605331fca43SMartin KaFai Lau 6067425627bSNathan Chancellor static __be32 *process_tcp_ao_options(struct tcp_sock *tp, 6077425627bSNathan Chancellor const struct tcp_request_sock *tcprsk, 6087425627bSNathan Chancellor struct tcp_out_options *opts, 6097425627bSNathan Chancellor struct tcp_key *key, __be32 *ptr) 6107425627bSNathan Chancellor { 6117425627bSNathan Chancellor #ifdef CONFIG_TCP_AO 6127425627bSNathan Chancellor u8 maclen = tcp_ao_maclen(key->ao_key); 6137425627bSNathan Chancellor 6147425627bSNathan Chancellor if (tcprsk) { 6157425627bSNathan Chancellor u8 aolen = maclen + sizeof(struct tcp_ao_hdr); 6167425627bSNathan Chancellor 6177425627bSNathan Chancellor *ptr++ = htonl((TCPOPT_AO << 24) | (aolen << 16) | 6187425627bSNathan Chancellor (tcprsk->ao_keyid << 8) | 6197425627bSNathan Chancellor (tcprsk->ao_rcv_next)); 6207425627bSNathan Chancellor } else { 6217425627bSNathan Chancellor struct tcp_ao_key *rnext_key; 6227425627bSNathan Chancellor struct tcp_ao_info *ao_info; 6237425627bSNathan Chancellor 6247425627bSNathan Chancellor ao_info = rcu_dereference_check(tp->ao_info, 6257425627bSNathan Chancellor lockdep_sock_is_held(&tp->inet_conn.icsk_inet.sk)); 6267425627bSNathan Chancellor rnext_key = READ_ONCE(ao_info->rnext_key); 6277425627bSNathan Chancellor if (WARN_ON_ONCE(!rnext_key)) 6287425627bSNathan Chancellor return ptr; 6297425627bSNathan Chancellor *ptr++ = htonl((TCPOPT_AO << 24) | 6307425627bSNathan Chancellor (tcp_ao_len(key->ao_key) << 16) | 6317425627bSNathan Chancellor (key->ao_key->sndid << 8) | 6327425627bSNathan Chancellor (rnext_key->rcvid)); 6337425627bSNathan Chancellor } 6347425627bSNathan Chancellor opts->hash_location = (__u8 *)ptr; 6357425627bSNathan Chancellor ptr += maclen / sizeof(*ptr); 6367425627bSNathan Chancellor if (unlikely(maclen % sizeof(*ptr))) { 6377425627bSNathan Chancellor memset(ptr, TCPOPT_NOP, sizeof(*ptr)); 6387425627bSNathan Chancellor ptr++; 6397425627bSNathan Chancellor } 6407425627bSNathan Chancellor #endif 6417425627bSNathan Chancellor return ptr; 6427425627bSNathan Chancellor } 6437425627bSNathan Chancellor 64467edfef7SAndi Kleen /* Write previously computed TCP options to the packet. 64567edfef7SAndi Kleen * 64667edfef7SAndi Kleen * Beware: Something in the Internet is very sensitive to the ordering of 647fd6149d3SIlpo Järvinen * TCP options, we learned this through the hard way, so be careful here. 648fd6149d3SIlpo Järvinen * Luckily we can at least blame others for their non-compliance but from 6498e3bff96Sstephen hemminger * inter-operability perspective it seems that we're somewhat stuck with 650fd6149d3SIlpo Järvinen * the ordering which we have been using if we want to keep working with 651fd6149d3SIlpo Järvinen * those broken things (not that it currently hurts anybody as there isn't 652fd6149d3SIlpo Järvinen * particular reason why the ordering would need to be changed). 653fd6149d3SIlpo Järvinen * 654fd6149d3SIlpo Järvinen * At least SACK_PERM as the first option is known to lead to a disaster 655fd6149d3SIlpo Järvinen * (but it may well be that other scenarios fail similarly). 656fd6149d3SIlpo Järvinen */ 657ea66758cSPaolo Abeni static void tcp_options_write(struct tcphdr *th, struct tcp_sock *tp, 65806b22ef2SDmitry Safonov const struct tcp_request_sock *tcprsk, 6591e03d32bSDmitry Safonov struct tcp_out_options *opts, 6601e03d32bSDmitry Safonov struct tcp_key *key) 661bd0388aeSWilliam Allen Simpson { 662ea66758cSPaolo Abeni __be32 *ptr = (__be32 *)(th + 1); 6632100c8d2SYuchung Cheng u16 options = opts->options; /* mungable copy */ 664bd0388aeSWilliam Allen Simpson 6651e03d32bSDmitry Safonov if (tcp_key_is_md5(key)) { 6661a2c6181SChristoph Paasch *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | 6671a2c6181SChristoph Paasch (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG); 668bd0388aeSWilliam Allen Simpson /* overload cookie hash location */ 669bd0388aeSWilliam Allen Simpson opts->hash_location = (__u8 *)ptr; 67033ad798cSAdam Langley ptr += 4; 6711e03d32bSDmitry Safonov } else if (tcp_key_is_ao(key)) { 6727425627bSNathan Chancellor ptr = process_tcp_ao_options(tp, tcprsk, opts, key, ptr); 6731e03d32bSDmitry Safonov } 674fd6149d3SIlpo Järvinen if (unlikely(opts->mss)) { 675fd6149d3SIlpo Järvinen *ptr++ = htonl((TCPOPT_MSS << 24) | 676fd6149d3SIlpo Järvinen (TCPOLEN_MSS << 16) | 677fd6149d3SIlpo Järvinen opts->mss); 678fd6149d3SIlpo Järvinen } 679fd6149d3SIlpo Järvinen 680bd0388aeSWilliam Allen Simpson if (likely(OPTION_TS & options)) { 681bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 68233ad798cSAdam Langley *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | 68333ad798cSAdam Langley (TCPOLEN_SACK_PERM << 16) | 68433ad798cSAdam Langley (TCPOPT_TIMESTAMP << 8) | 68533ad798cSAdam Langley TCPOLEN_TIMESTAMP); 686bd0388aeSWilliam Allen Simpson options &= ~OPTION_SACK_ADVERTISE; 68733ad798cSAdam Langley } else { 688496c98dfSYOSHIFUJI Hideaki *ptr++ = htonl((TCPOPT_NOP << 24) | 68940efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 69040efc6faSStephen Hemminger (TCPOPT_TIMESTAMP << 8) | 69140efc6faSStephen Hemminger TCPOLEN_TIMESTAMP); 69240efc6faSStephen Hemminger } 69333ad798cSAdam Langley *ptr++ = htonl(opts->tsval); 69433ad798cSAdam Langley *ptr++ = htonl(opts->tsecr); 69533ad798cSAdam Langley } 69633ad798cSAdam Langley 697bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 69833ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 69933ad798cSAdam Langley (TCPOPT_NOP << 16) | 70033ad798cSAdam Langley (TCPOPT_SACK_PERM << 8) | 70133ad798cSAdam Langley TCPOLEN_SACK_PERM); 70233ad798cSAdam Langley } 70333ad798cSAdam Langley 704bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_WSCALE & options)) { 70533ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 70633ad798cSAdam Langley (TCPOPT_WINDOW << 16) | 70733ad798cSAdam Langley (TCPOLEN_WINDOW << 8) | 70833ad798cSAdam Langley opts->ws); 70933ad798cSAdam Langley } 71033ad798cSAdam Langley 71133ad798cSAdam Langley if (unlikely(opts->num_sack_blocks)) { 71233ad798cSAdam Langley struct tcp_sack_block *sp = tp->rx_opt.dsack ? 71333ad798cSAdam Langley tp->duplicate_sack : tp->selective_acks; 71440efc6faSStephen Hemminger int this_sack; 71540efc6faSStephen Hemminger 71640efc6faSStephen Hemminger *ptr++ = htonl((TCPOPT_NOP << 24) | 71740efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 71840efc6faSStephen Hemminger (TCPOPT_SACK << 8) | 71933ad798cSAdam Langley (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * 72040efc6faSStephen Hemminger TCPOLEN_SACK_PERBLOCK))); 7212de979bdSStephen Hemminger 72233ad798cSAdam Langley for (this_sack = 0; this_sack < opts->num_sack_blocks; 72333ad798cSAdam Langley ++this_sack) { 72440efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].start_seq); 72540efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].end_seq); 72640efc6faSStephen Hemminger } 7272de979bdSStephen Hemminger 72840efc6faSStephen Hemminger tp->rx_opt.dsack = 0; 72940efc6faSStephen Hemminger } 7302100c8d2SYuchung Cheng 7312100c8d2SYuchung Cheng if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) { 7322100c8d2SYuchung Cheng struct tcp_fastopen_cookie *foc = opts->fastopen_cookie; 7337f9b838bSDaniel Lee u8 *p = (u8 *)ptr; 7347f9b838bSDaniel Lee u32 len; /* Fast Open option length */ 7352100c8d2SYuchung Cheng 7367f9b838bSDaniel Lee if (foc->exp) { 7377f9b838bSDaniel Lee len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len; 7387f9b838bSDaniel Lee *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) | 7392100c8d2SYuchung Cheng TCPOPT_FASTOPEN_MAGIC); 7407f9b838bSDaniel Lee p += TCPOLEN_EXP_FASTOPEN_BASE; 7417f9b838bSDaniel Lee } else { 7427f9b838bSDaniel Lee len = TCPOLEN_FASTOPEN_BASE + foc->len; 7437f9b838bSDaniel Lee *p++ = TCPOPT_FASTOPEN; 7447f9b838bSDaniel Lee *p++ = len; 7452100c8d2SYuchung Cheng } 7467f9b838bSDaniel Lee 7477f9b838bSDaniel Lee memcpy(p, foc->val, foc->len); 7487f9b838bSDaniel Lee if ((len & 3) == 2) { 7497f9b838bSDaniel Lee p[foc->len] = TCPOPT_NOP; 7507f9b838bSDaniel Lee p[foc->len + 1] = TCPOPT_NOP; 7517f9b838bSDaniel Lee } 7527f9b838bSDaniel Lee ptr += (len + 3) >> 2; 7532100c8d2SYuchung Cheng } 75460e2a778SUrsula Braun 75560e2a778SUrsula Braun smc_options_write(ptr, &options); 756eda7acddSPeter Krystad 757ea66758cSPaolo Abeni mptcp_options_write(th, ptr, tp, opts); 75860e2a778SUrsula Braun } 75960e2a778SUrsula Braun 76060e2a778SUrsula Braun static void smc_set_option(const struct tcp_sock *tp, 76160e2a778SUrsula Braun struct tcp_out_options *opts, 76260e2a778SUrsula Braun unsigned int *remaining) 76360e2a778SUrsula Braun { 76460e2a778SUrsula Braun #if IS_ENABLED(CONFIG_SMC) 76560e2a778SUrsula Braun if (static_branch_unlikely(&tcp_have_smc)) { 76660e2a778SUrsula Braun if (tp->syn_smc) { 76760e2a778SUrsula Braun if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) { 76860e2a778SUrsula Braun opts->options |= OPTION_SMC; 76960e2a778SUrsula Braun *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED; 77060e2a778SUrsula Braun } 77160e2a778SUrsula Braun } 77260e2a778SUrsula Braun } 77360e2a778SUrsula Braun #endif 77460e2a778SUrsula Braun } 77560e2a778SUrsula Braun 77660e2a778SUrsula Braun static void smc_set_option_cond(const struct tcp_sock *tp, 77760e2a778SUrsula Braun const struct inet_request_sock *ireq, 77860e2a778SUrsula Braun struct tcp_out_options *opts, 77960e2a778SUrsula Braun unsigned int *remaining) 78060e2a778SUrsula Braun { 78160e2a778SUrsula Braun #if IS_ENABLED(CONFIG_SMC) 78260e2a778SUrsula Braun if (static_branch_unlikely(&tcp_have_smc)) { 78360e2a778SUrsula Braun if (tp->syn_smc && ireq->smc_ok) { 78460e2a778SUrsula Braun if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) { 78560e2a778SUrsula Braun opts->options |= OPTION_SMC; 78660e2a778SUrsula Braun *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED; 78760e2a778SUrsula Braun } 78860e2a778SUrsula Braun } 78960e2a778SUrsula Braun } 79060e2a778SUrsula Braun #endif 79140efc6faSStephen Hemminger } 79240efc6faSStephen Hemminger 793cec37a6eSPeter Krystad static void mptcp_set_option_cond(const struct request_sock *req, 794cec37a6eSPeter Krystad struct tcp_out_options *opts, 795cec37a6eSPeter Krystad unsigned int *remaining) 796cec37a6eSPeter Krystad { 797cec37a6eSPeter Krystad if (rsk_is_mptcp(req)) { 798cec37a6eSPeter Krystad unsigned int size; 799cec37a6eSPeter Krystad 800cec37a6eSPeter Krystad if (mptcp_synack_options(req, &size, &opts->mptcp)) { 801cec37a6eSPeter Krystad if (*remaining >= size) { 802cec37a6eSPeter Krystad opts->options |= OPTION_MPTCP; 803cec37a6eSPeter Krystad *remaining -= size; 804cec37a6eSPeter Krystad } 805cec37a6eSPeter Krystad } 806cec37a6eSPeter Krystad } 807cec37a6eSPeter Krystad } 808cec37a6eSPeter Krystad 80967edfef7SAndi Kleen /* Compute TCP options for SYN packets. This is not the final 81067edfef7SAndi Kleen * network wire format yet. 81167edfef7SAndi Kleen */ 81295c96174SEric Dumazet static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb, 81333ad798cSAdam Langley struct tcp_out_options *opts, 8141e03d32bSDmitry Safonov struct tcp_key *key) 815cf533ea5SEric Dumazet { 81633ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 81795c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 818783237e8SYuchung Cheng struct tcp_fastopen_request *fastopen = tp->fastopen_req; 8191e03d32bSDmitry Safonov bool timestamps; 82033ad798cSAdam Langley 8211e03d32bSDmitry Safonov /* Better than switch (key.type) as it has static branches */ 8221e03d32bSDmitry Safonov if (tcp_key_is_md5(key)) { 8231e03d32bSDmitry Safonov timestamps = false; 82433ad798cSAdam Langley opts->options |= OPTION_MD5; 825bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 8261e03d32bSDmitry Safonov } else { 8271e03d32bSDmitry Safonov timestamps = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_timestamps); 8281e03d32bSDmitry Safonov if (tcp_key_is_ao(key)) { 8291e03d32bSDmitry Safonov opts->options |= OPTION_AO; 830da7dfaa6SDmitry Safonov remaining -= tcp_ao_len_aligned(key->ao_key); 831cfb6eeb4SYOSHIFUJI Hideaki } 8328c2320e8SEric Dumazet } 83333ad798cSAdam Langley 83433ad798cSAdam Langley /* We always get an MSS option. The option bytes which will be seen in 83533ad798cSAdam Langley * normal data packets should timestamps be used, must be in the MSS 83633ad798cSAdam Langley * advertised. But we subtract them from tp->mss_cache so that 83733ad798cSAdam Langley * calculations in tcp_sendmsg are simpler etc. So account for this 83833ad798cSAdam Langley * fact here if necessary. If we don't do this correctly, as a 83933ad798cSAdam Langley * receiver we won't recognize data packets as being full sized when we 84033ad798cSAdam Langley * should, and thus we won't abide by the delayed ACK rules correctly. 84133ad798cSAdam Langley * SACKs don't matter, we never delay an ACK when we have any of those 84233ad798cSAdam Langley * going out. */ 84333ad798cSAdam Langley opts->mss = tcp_advertise_mss(sk); 844bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 84533ad798cSAdam Langley 8461e03d32bSDmitry Safonov if (likely(timestamps)) { 84733ad798cSAdam Langley opts->options |= OPTION_TS; 848614e8316SEric Dumazet opts->tsval = tcp_skb_timestamp_ts(tp->tcp_usec_ts, skb) + tp->tsoffset; 84933ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 850bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 85133ad798cSAdam Langley } 8523666f666SKuniyuki Iwashima if (likely(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_window_scaling))) { 85333ad798cSAdam Langley opts->ws = tp->rx_opt.rcv_wscale; 85489e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 855bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 85633ad798cSAdam Langley } 8573666f666SKuniyuki Iwashima if (likely(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_sack))) { 85833ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 859b32d1310SDavid S. Miller if (unlikely(!(OPTION_TS & opts->options))) 860bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 86133ad798cSAdam Langley } 86233ad798cSAdam Langley 863783237e8SYuchung Cheng if (fastopen && fastopen->cookie.len >= 0) { 8642646c831SDaniel Lee u32 need = fastopen->cookie.len; 8652646c831SDaniel Lee 8662646c831SDaniel Lee need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE : 8672646c831SDaniel Lee TCPOLEN_FASTOPEN_BASE; 868783237e8SYuchung Cheng need = (need + 3) & ~3U; /* Align to 32 bits */ 869783237e8SYuchung Cheng if (remaining >= need) { 870783237e8SYuchung Cheng opts->options |= OPTION_FAST_OPEN_COOKIE; 871783237e8SYuchung Cheng opts->fastopen_cookie = &fastopen->cookie; 872783237e8SYuchung Cheng remaining -= need; 873783237e8SYuchung Cheng tp->syn_fastopen = 1; 8742646c831SDaniel Lee tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0; 875783237e8SYuchung Cheng } 876783237e8SYuchung Cheng } 877bd0388aeSWilliam Allen Simpson 87860e2a778SUrsula Braun smc_set_option(tp, opts, &remaining); 87960e2a778SUrsula Braun 880cec37a6eSPeter Krystad if (sk_is_mptcp(sk)) { 881cec37a6eSPeter Krystad unsigned int size; 882cec37a6eSPeter Krystad 883cc7972eaSChristoph Paasch if (mptcp_syn_options(sk, skb, &size, &opts->mptcp)) { 884cec37a6eSPeter Krystad opts->options |= OPTION_MPTCP; 885cec37a6eSPeter Krystad remaining -= size; 886cec37a6eSPeter Krystad } 887cec37a6eSPeter Krystad } 888cec37a6eSPeter Krystad 889331fca43SMartin KaFai Lau bpf_skops_hdr_opt_len(sk, skb, NULL, NULL, 0, opts, &remaining); 890331fca43SMartin KaFai Lau 891bd0388aeSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 89233ad798cSAdam Langley } 89333ad798cSAdam Langley 89467edfef7SAndi Kleen /* Set up TCP options for SYN-ACKs. */ 89560e2a778SUrsula Braun static unsigned int tcp_synack_options(const struct sock *sk, 89660e2a778SUrsula Braun struct request_sock *req, 89795c96174SEric Dumazet unsigned int mss, struct sk_buff *skb, 89833ad798cSAdam Langley struct tcp_out_options *opts, 8999427c6aaSDmitry Safonov const struct tcp_key *key, 900e114e1e8SEric Dumazet struct tcp_fastopen_cookie *foc, 901331fca43SMartin KaFai Lau enum tcp_synack_type synack_type, 902331fca43SMartin KaFai Lau struct sk_buff *syn_skb) 9034957faadSWilliam Allen Simpson { 90433ad798cSAdam Langley struct inet_request_sock *ireq = inet_rsk(req); 90595c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 90633ad798cSAdam Langley 9079427c6aaSDmitry Safonov if (tcp_key_is_md5(key)) { 90833ad798cSAdam Langley opts->options |= OPTION_MD5; 9094957faadSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 9104957faadSWilliam Allen Simpson 9114957faadSWilliam Allen Simpson /* We can't fit any SACK blocks in a packet with MD5 + TS 9124957faadSWilliam Allen Simpson * options. There was discussion about disabling SACK 9134957faadSWilliam Allen Simpson * rather than TS in order to fit in better with old, 9144957faadSWilliam Allen Simpson * buggy kernels, but that was deemed to be unnecessary. 9154957faadSWilliam Allen Simpson */ 916e114e1e8SEric Dumazet if (synack_type != TCP_SYNACK_COOKIE) 917de213e5eSEric Dumazet ireq->tstamp_ok &= !ireq->sack_ok; 9189427c6aaSDmitry Safonov } else if (tcp_key_is_ao(key)) { 9199427c6aaSDmitry Safonov opts->options |= OPTION_AO; 920da7dfaa6SDmitry Safonov remaining -= tcp_ao_len_aligned(key->ao_key); 9219427c6aaSDmitry Safonov ireq->tstamp_ok &= !ireq->sack_ok; 92233ad798cSAdam Langley } 92333ad798cSAdam Langley 9244957faadSWilliam Allen Simpson /* We always send an MSS option. */ 92533ad798cSAdam Langley opts->mss = mss; 9264957faadSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 92733ad798cSAdam Langley 92833ad798cSAdam Langley if (likely(ireq->wscale_ok)) { 92933ad798cSAdam Langley opts->ws = ireq->rcv_wscale; 93089e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 9314957faadSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 93233ad798cSAdam Langley } 933de213e5eSEric Dumazet if (likely(ireq->tstamp_ok)) { 93433ad798cSAdam Langley opts->options |= OPTION_TS; 935614e8316SEric Dumazet opts->tsval = tcp_skb_timestamp_ts(tcp_rsk(req)->req_usec_ts, skb) + 936614e8316SEric Dumazet tcp_rsk(req)->ts_off; 937eba20811SEric Dumazet opts->tsecr = READ_ONCE(req->ts_recent); 9384957faadSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 93933ad798cSAdam Langley } 94033ad798cSAdam Langley if (likely(ireq->sack_ok)) { 94133ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 942de213e5eSEric Dumazet if (unlikely(!ireq->tstamp_ok)) 9434957faadSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 94433ad798cSAdam Langley } 9457f9b838bSDaniel Lee if (foc != NULL && foc->len >= 0) { 9467f9b838bSDaniel Lee u32 need = foc->len; 9477f9b838bSDaniel Lee 9487f9b838bSDaniel Lee need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE : 9497f9b838bSDaniel Lee TCPOLEN_FASTOPEN_BASE; 9508336886fSJerry Chu need = (need + 3) & ~3U; /* Align to 32 bits */ 9518336886fSJerry Chu if (remaining >= need) { 9528336886fSJerry Chu opts->options |= OPTION_FAST_OPEN_COOKIE; 9538336886fSJerry Chu opts->fastopen_cookie = foc; 9548336886fSJerry Chu remaining -= need; 9558336886fSJerry Chu } 9568336886fSJerry Chu } 9574957faadSWilliam Allen Simpson 958cec37a6eSPeter Krystad mptcp_set_option_cond(req, opts, &remaining); 959cec37a6eSPeter Krystad 96060e2a778SUrsula Braun smc_set_option_cond(tcp_sk(sk), ireq, opts, &remaining); 96160e2a778SUrsula Braun 962331fca43SMartin KaFai Lau bpf_skops_hdr_opt_len((struct sock *)sk, skb, req, syn_skb, 963331fca43SMartin KaFai Lau synack_type, opts, &remaining); 964331fca43SMartin KaFai Lau 9654957faadSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 96633ad798cSAdam Langley } 96733ad798cSAdam Langley 96867edfef7SAndi Kleen /* Compute TCP options for ESTABLISHED sockets. This is not the 96967edfef7SAndi Kleen * final wire format yet. 97067edfef7SAndi Kleen */ 97195c96174SEric Dumazet static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb, 97233ad798cSAdam Langley struct tcp_out_options *opts, 9731e03d32bSDmitry Safonov struct tcp_key *key) 974cf533ea5SEric Dumazet { 97533ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 97695c96174SEric Dumazet unsigned int size = 0; 977cabeccbdSIlpo Järvinen unsigned int eff_sacks; 97833ad798cSAdam Langley 9795843ef42SAndi Kleen opts->options = 0; 9805843ef42SAndi Kleen 9811e03d32bSDmitry Safonov /* Better than switch (key.type) as it has static branches */ 9821e03d32bSDmitry Safonov if (tcp_key_is_md5(key)) { 98333ad798cSAdam Langley opts->options |= OPTION_MD5; 98433ad798cSAdam Langley size += TCPOLEN_MD5SIG_ALIGNED; 9851e03d32bSDmitry Safonov } else if (tcp_key_is_ao(key)) { 9861e03d32bSDmitry Safonov opts->options |= OPTION_AO; 987da7dfaa6SDmitry Safonov size += tcp_ao_len_aligned(key->ao_key); 98833ad798cSAdam Langley } 98933ad798cSAdam Langley 99033ad798cSAdam Langley if (likely(tp->rx_opt.tstamp_ok)) { 99133ad798cSAdam Langley opts->options |= OPTION_TS; 992614e8316SEric Dumazet opts->tsval = skb ? tcp_skb_timestamp_ts(tp->tcp_usec_ts, skb) + 993614e8316SEric Dumazet tp->tsoffset : 0; 99433ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 99533ad798cSAdam Langley size += TCPOLEN_TSTAMP_ALIGNED; 99633ad798cSAdam Langley } 99733ad798cSAdam Langley 998cec37a6eSPeter Krystad /* MPTCP options have precedence over SACK for the limited TCP 999cec37a6eSPeter Krystad * option space because a MPTCP connection would be forced to 1000cec37a6eSPeter Krystad * fall back to regular TCP if a required multipath option is 1001cec37a6eSPeter Krystad * missing. SACK still gets a chance to use whatever space is 1002cec37a6eSPeter Krystad * left. 1003cec37a6eSPeter Krystad */ 1004cec37a6eSPeter Krystad if (sk_is_mptcp(sk)) { 1005cec37a6eSPeter Krystad unsigned int remaining = MAX_TCP_OPTION_SPACE - size; 1006cec37a6eSPeter Krystad unsigned int opt_size = 0; 1007cec37a6eSPeter Krystad 1008cec37a6eSPeter Krystad if (mptcp_established_options(sk, skb, &opt_size, remaining, 1009cec37a6eSPeter Krystad &opts->mptcp)) { 1010cec37a6eSPeter Krystad opts->options |= OPTION_MPTCP; 1011cec37a6eSPeter Krystad size += opt_size; 1012cec37a6eSPeter Krystad } 1013cec37a6eSPeter Krystad } 1014cec37a6eSPeter Krystad 1015cabeccbdSIlpo Järvinen eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; 1016cabeccbdSIlpo Järvinen if (unlikely(eff_sacks)) { 101795c96174SEric Dumazet const unsigned int remaining = MAX_TCP_OPTION_SPACE - size; 10189cfcca23SMat Martineau if (unlikely(remaining < TCPOLEN_SACK_BASE_ALIGNED + 10199cfcca23SMat Martineau TCPOLEN_SACK_PERBLOCK)) 10209cfcca23SMat Martineau return size; 10219cfcca23SMat Martineau 102233ad798cSAdam Langley opts->num_sack_blocks = 102395c96174SEric Dumazet min_t(unsigned int, eff_sacks, 102433ad798cSAdam Langley (remaining - TCPOLEN_SACK_BASE_ALIGNED) / 102533ad798cSAdam Langley TCPOLEN_SACK_PERBLOCK); 10269cfcca23SMat Martineau 102733ad798cSAdam Langley size += TCPOLEN_SACK_BASE_ALIGNED + 102833ad798cSAdam Langley opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; 102933ad798cSAdam Langley } 103033ad798cSAdam Langley 1031331fca43SMartin KaFai Lau if (unlikely(BPF_SOCK_OPS_TEST_FLAG(tp, 1032331fca43SMartin KaFai Lau BPF_SOCK_OPS_WRITE_HDR_OPT_CB_FLAG))) { 1033331fca43SMartin KaFai Lau unsigned int remaining = MAX_TCP_OPTION_SPACE - size; 1034331fca43SMartin KaFai Lau 1035331fca43SMartin KaFai Lau bpf_skops_hdr_opt_len(sk, skb, NULL, NULL, 0, opts, &remaining); 1036331fca43SMartin KaFai Lau 1037331fca43SMartin KaFai Lau size = MAX_TCP_OPTION_SPACE - remaining; 1038331fca43SMartin KaFai Lau } 1039331fca43SMartin KaFai Lau 104033ad798cSAdam Langley return size; 104140efc6faSStephen Hemminger } 10421da177e4SLinus Torvalds 104346d3ceabSEric Dumazet 104446d3ceabSEric Dumazet /* TCP SMALL QUEUES (TSQ) 104546d3ceabSEric Dumazet * 104646d3ceabSEric Dumazet * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev) 104746d3ceabSEric Dumazet * to reduce RTT and bufferbloat. 104846d3ceabSEric Dumazet * We do this using a special skb destructor (tcp_wfree). 104946d3ceabSEric Dumazet * 105046d3ceabSEric Dumazet * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb 105146d3ceabSEric Dumazet * needs to be reallocated in a driver. 10528e3bff96Sstephen hemminger * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc 105346d3ceabSEric Dumazet * 105446d3ceabSEric Dumazet * Since transmit from skb destructor is forbidden, we use a tasklet 105546d3ceabSEric Dumazet * to process all sockets that eventually need to send more skbs. 105646d3ceabSEric Dumazet * We use one tasklet per cpu, with its own queue of sockets. 105746d3ceabSEric Dumazet */ 105846d3ceabSEric Dumazet struct tsq_tasklet { 105946d3ceabSEric Dumazet struct tasklet_struct tasklet; 106046d3ceabSEric Dumazet struct list_head head; /* queue of tcp sockets */ 106146d3ceabSEric Dumazet }; 106246d3ceabSEric Dumazet static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet); 106346d3ceabSEric Dumazet 106473a6bab5SEric Dumazet static void tcp_tsq_write(struct sock *sk) 10656f458dfbSEric Dumazet { 10666f458dfbSEric Dumazet if ((1 << sk->sk_state) & 10676f458dfbSEric Dumazet (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING | 1068f9616c35SEric Dumazet TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) { 1069f9616c35SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 1070f9616c35SEric Dumazet 1071f9616c35SEric Dumazet if (tp->lost_out > tp->retrans_out && 107240570375SEric Dumazet tcp_snd_cwnd(tp) > tcp_packets_in_flight(tp)) { 10733a91d29fSKoichiro Den tcp_mstamp_refresh(tp); 1074f9616c35SEric Dumazet tcp_xmit_retransmit_queue(sk); 10753a91d29fSKoichiro Den } 1076f9616c35SEric Dumazet 1077f9616c35SEric Dumazet tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle, 1078bf06200eSJohn Ogness 0, GFP_ATOMIC); 10796f458dfbSEric Dumazet } 1080f9616c35SEric Dumazet } 108173a6bab5SEric Dumazet 108273a6bab5SEric Dumazet static void tcp_tsq_handler(struct sock *sk) 108373a6bab5SEric Dumazet { 108473a6bab5SEric Dumazet bh_lock_sock(sk); 108573a6bab5SEric Dumazet if (!sock_owned_by_user(sk)) 108673a6bab5SEric Dumazet tcp_tsq_write(sk); 108773a6bab5SEric Dumazet else if (!test_and_set_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags)) 108873a6bab5SEric Dumazet sock_hold(sk); 108973a6bab5SEric Dumazet bh_unlock_sock(sk); 109073a6bab5SEric Dumazet } 109146d3ceabSEric Dumazet /* 10928e3bff96Sstephen hemminger * One tasklet per cpu tries to send more skbs. 109346d3ceabSEric Dumazet * We run in tasklet context but need to disable irqs when 10948e3bff96Sstephen hemminger * transferring tsq->head because tcp_wfree() might 109546d3ceabSEric Dumazet * interrupt us (non NAPI drivers) 109646d3ceabSEric Dumazet */ 1097c6533ca8SAllen Pais static void tcp_tasklet_func(struct tasklet_struct *t) 109846d3ceabSEric Dumazet { 1099c6533ca8SAllen Pais struct tsq_tasklet *tsq = from_tasklet(tsq, t, tasklet); 110046d3ceabSEric Dumazet LIST_HEAD(list); 110146d3ceabSEric Dumazet unsigned long flags; 110246d3ceabSEric Dumazet struct list_head *q, *n; 110346d3ceabSEric Dumazet struct tcp_sock *tp; 110446d3ceabSEric Dumazet struct sock *sk; 110546d3ceabSEric Dumazet 110646d3ceabSEric Dumazet local_irq_save(flags); 110746d3ceabSEric Dumazet list_splice_init(&tsq->head, &list); 110846d3ceabSEric Dumazet local_irq_restore(flags); 110946d3ceabSEric Dumazet 111046d3ceabSEric Dumazet list_for_each_safe(q, n, &list) { 111146d3ceabSEric Dumazet tp = list_entry(q, struct tcp_sock, tsq_node); 111246d3ceabSEric Dumazet list_del(&tp->tsq_node); 111346d3ceabSEric Dumazet 111446d3ceabSEric Dumazet sk = (struct sock *)tp; 11150a9648f1SEric Dumazet smp_mb__before_atomic(); 11167aa5470cSEric Dumazet clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags); 11177aa5470cSEric Dumazet 11186f458dfbSEric Dumazet tcp_tsq_handler(sk); 111946d3ceabSEric Dumazet sk_free(sk); 112046d3ceabSEric Dumazet } 112146d3ceabSEric Dumazet } 112246d3ceabSEric Dumazet 112340fc3423SEric Dumazet #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \ 112440fc3423SEric Dumazet TCPF_WRITE_TIMER_DEFERRED | \ 112540fc3423SEric Dumazet TCPF_DELACK_TIMER_DEFERRED | \ 1126133c4c0dSEric Dumazet TCPF_MTU_REDUCED_DEFERRED | \ 1127133c4c0dSEric Dumazet TCPF_ACK_DEFERRED) 112846d3ceabSEric Dumazet /** 112946d3ceabSEric Dumazet * tcp_release_cb - tcp release_sock() callback 113046d3ceabSEric Dumazet * @sk: socket 113146d3ceabSEric Dumazet * 113246d3ceabSEric Dumazet * called from release_sock() to perform protocol dependent 113346d3ceabSEric Dumazet * actions before socket release. 113446d3ceabSEric Dumazet */ 113546d3ceabSEric Dumazet void tcp_release_cb(struct sock *sk) 113646d3ceabSEric Dumazet { 1137fac30731SEric Dumazet unsigned long flags = smp_load_acquire(&sk->sk_tsq_flags); 1138fac30731SEric Dumazet unsigned long nflags; 113946d3ceabSEric Dumazet 11406f458dfbSEric Dumazet /* perform an atomic operation only if at least one flag is set */ 11416f458dfbSEric Dumazet do { 11426f458dfbSEric Dumazet if (!(flags & TCP_DEFERRED_ALL)) 11436f458dfbSEric Dumazet return; 11446f458dfbSEric Dumazet nflags = flags & ~TCP_DEFERRED_ALL; 1145fac30731SEric Dumazet } while (!try_cmpxchg(&sk->sk_tsq_flags, &flags, nflags)); 11466f458dfbSEric Dumazet 114773a6bab5SEric Dumazet if (flags & TCPF_TSQ_DEFERRED) { 114873a6bab5SEric Dumazet tcp_tsq_write(sk); 114973a6bab5SEric Dumazet __sock_put(sk); 115073a6bab5SEric Dumazet } 1151c3f9b018SEric Dumazet 115240fc3423SEric Dumazet if (flags & TCPF_WRITE_TIMER_DEFERRED) { 11536f458dfbSEric Dumazet tcp_write_timer_handler(sk); 1154144d56e9SEric Dumazet __sock_put(sk); 1155144d56e9SEric Dumazet } 115640fc3423SEric Dumazet if (flags & TCPF_DELACK_TIMER_DEFERRED) { 11576f458dfbSEric Dumazet tcp_delack_timer_handler(sk); 1158144d56e9SEric Dumazet __sock_put(sk); 1159144d56e9SEric Dumazet } 116040fc3423SEric Dumazet if (flags & TCPF_MTU_REDUCED_DEFERRED) { 11614fab9071SNeal Cardwell inet_csk(sk)->icsk_af_ops->mtu_reduced(sk); 1162144d56e9SEric Dumazet __sock_put(sk); 1163144d56e9SEric Dumazet } 1164133c4c0dSEric Dumazet if ((flags & TCPF_ACK_DEFERRED) && inet_csk_ack_scheduled(sk)) 1165133c4c0dSEric Dumazet tcp_send_ack(sk); 116646d3ceabSEric Dumazet } 116746d3ceabSEric Dumazet EXPORT_SYMBOL(tcp_release_cb); 116846d3ceabSEric Dumazet 116946d3ceabSEric Dumazet void __init tcp_tasklet_init(void) 117046d3ceabSEric Dumazet { 117146d3ceabSEric Dumazet int i; 117246d3ceabSEric Dumazet 117346d3ceabSEric Dumazet for_each_possible_cpu(i) { 117446d3ceabSEric Dumazet struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i); 117546d3ceabSEric Dumazet 117646d3ceabSEric Dumazet INIT_LIST_HEAD(&tsq->head); 1177c6533ca8SAllen Pais tasklet_setup(&tsq->tasklet, tcp_tasklet_func); 117846d3ceabSEric Dumazet } 117946d3ceabSEric Dumazet } 118046d3ceabSEric Dumazet 118146d3ceabSEric Dumazet /* 118246d3ceabSEric Dumazet * Write buffer destructor automatically called from kfree_skb. 11838e3bff96Sstephen hemminger * We can't xmit new skbs from this context, as we might already 118446d3ceabSEric Dumazet * hold qdisc lock. 118546d3ceabSEric Dumazet */ 1186d6a4a104SEric Dumazet void tcp_wfree(struct sk_buff *skb) 118746d3ceabSEric Dumazet { 118846d3ceabSEric Dumazet struct sock *sk = skb->sk; 118946d3ceabSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 1190408f0a6cSEric Dumazet unsigned long flags, nval, oval; 1191b548b17aSEric Dumazet struct tsq_tasklet *tsq; 1192b548b17aSEric Dumazet bool empty; 11939b462d02SEric Dumazet 11949b462d02SEric Dumazet /* Keep one reference on sk_wmem_alloc. 11959b462d02SEric Dumazet * Will be released by sk_free() from here or tcp_tasklet_func() 11969b462d02SEric Dumazet */ 119714afee4bSReshetova, Elena WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc)); 11989b462d02SEric Dumazet 11999b462d02SEric Dumazet /* If this softirq is serviced by ksoftirqd, we are likely under stress. 12009b462d02SEric Dumazet * Wait until our queues (qdisc + devices) are drained. 12019b462d02SEric Dumazet * This gives : 12029b462d02SEric Dumazet * - less callbacks to tcp_write_xmit(), reducing stress (batches) 12039b462d02SEric Dumazet * - chance for incoming ACK (processed by another cpu maybe) 12049b462d02SEric Dumazet * to migrate this flow (skb->ooo_okay will be eventually set) 12059b462d02SEric Dumazet */ 120614afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current) 12079b462d02SEric Dumazet goto out; 120846d3ceabSEric Dumazet 1209b548b17aSEric Dumazet oval = smp_load_acquire(&sk->sk_tsq_flags); 1210b548b17aSEric Dumazet do { 1211408f0a6cSEric Dumazet if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED)) 1212408f0a6cSEric Dumazet goto out; 1213408f0a6cSEric Dumazet 121473a6bab5SEric Dumazet nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED; 1215b548b17aSEric Dumazet } while (!try_cmpxchg(&sk->sk_tsq_flags, &oval, nval)); 1216408f0a6cSEric Dumazet 121746d3ceabSEric Dumazet /* queue this socket to tasklet queue */ 121846d3ceabSEric Dumazet local_irq_save(flags); 1219903ceff7SChristoph Lameter tsq = this_cpu_ptr(&tsq_tasklet); 1220a9b204d1SEric Dumazet empty = list_empty(&tsq->head); 122146d3ceabSEric Dumazet list_add(&tp->tsq_node, &tsq->head); 1222a9b204d1SEric Dumazet if (empty) 122346d3ceabSEric Dumazet tasklet_schedule(&tsq->tasklet); 122446d3ceabSEric Dumazet local_irq_restore(flags); 12259b462d02SEric Dumazet return; 12269b462d02SEric Dumazet out: 12279b462d02SEric Dumazet sk_free(sk); 122846d3ceabSEric Dumazet } 122946d3ceabSEric Dumazet 123073a6bab5SEric Dumazet /* Note: Called under soft irq. 123173a6bab5SEric Dumazet * We can call TCP stack right away, unless socket is owned by user. 1232218af599SEric Dumazet */ 1233218af599SEric Dumazet enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer) 1234218af599SEric Dumazet { 1235218af599SEric Dumazet struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer); 1236218af599SEric Dumazet struct sock *sk = (struct sock *)tp; 1237218af599SEric Dumazet 123873a6bab5SEric Dumazet tcp_tsq_handler(sk); 123973a6bab5SEric Dumazet sock_put(sk); 1240218af599SEric Dumazet 1241218af599SEric Dumazet return HRTIMER_NORESTART; 1242218af599SEric Dumazet } 1243218af599SEric Dumazet 1244a7a25630SEric Dumazet static void tcp_update_skb_after_send(struct sock *sk, struct sk_buff *skb, 1245a7a25630SEric Dumazet u64 prior_wstamp) 1246e2080072SEric Dumazet { 1247ab408b6dSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 1248ab408b6dSEric Dumazet 1249ab408b6dSEric Dumazet if (sk->sk_pacing_status != SK_PACING_NONE) { 125028b24f90SEric Dumazet unsigned long rate = READ_ONCE(sk->sk_pacing_rate); 1251ab408b6dSEric Dumazet 1252ab408b6dSEric Dumazet /* Original sch_fq does not pace first 10 MSS 1253ab408b6dSEric Dumazet * Note that tp->data_segs_out overflows after 2^32 packets, 1254ab408b6dSEric Dumazet * this is a minor annoyance. 1255ab408b6dSEric Dumazet */ 125676a9ebe8SEric Dumazet if (rate != ~0UL && rate && tp->data_segs_out >= 10) { 1257a7a25630SEric Dumazet u64 len_ns = div64_ul((u64)skb->len * NSEC_PER_SEC, rate); 1258a7a25630SEric Dumazet u64 credit = tp->tcp_wstamp_ns - prior_wstamp; 1259a7a25630SEric Dumazet 1260a7a25630SEric Dumazet /* take into account OS jitter */ 1261a7a25630SEric Dumazet len_ns -= min_t(u64, len_ns / 2, credit); 1262a7a25630SEric Dumazet tp->tcp_wstamp_ns += len_ns; 1263ab408b6dSEric Dumazet } 1264ab408b6dSEric Dumazet } 1265e2080072SEric Dumazet list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue); 1266e2080072SEric Dumazet } 1267e2080072SEric Dumazet 126805e22e83SEric Dumazet INDIRECT_CALLABLE_DECLARE(int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl)); 126905e22e83SEric Dumazet INDIRECT_CALLABLE_DECLARE(int inet6_csk_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl)); 1270dd2e0b86SEric Dumazet INDIRECT_CALLABLE_DECLARE(void tcp_v4_send_check(struct sock *sk, struct sk_buff *skb)); 127105e22e83SEric Dumazet 12721da177e4SLinus Torvalds /* This routine actually transmits TCP packets queued in by 12731da177e4SLinus Torvalds * tcp_do_sendmsg(). This is used by both the initial 12741da177e4SLinus Torvalds * transmission and possible later retransmissions. 12751da177e4SLinus Torvalds * All SKB's seen here are completely headerless. It is our 12761da177e4SLinus Torvalds * job to build the TCP header, and pass the packet down to 12771da177e4SLinus Torvalds * IP so it can do the same plus pass the packet off to the 12781da177e4SLinus Torvalds * device. 12791da177e4SLinus Torvalds * 12801da177e4SLinus Torvalds * We are working here with either a clone of the original 12811da177e4SLinus Torvalds * SKB, or a fresh unique copy made by the retransmit engine. 12821da177e4SLinus Torvalds */ 12832987babbSYuchung Cheng static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, 12842987babbSYuchung Cheng int clone_it, gfp_t gfp_mask, u32 rcv_nxt) 12851da177e4SLinus Torvalds { 12866687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 1287dfb4b9dcSDavid S. Miller struct inet_sock *inet; 1288dfb4b9dcSDavid S. Miller struct tcp_sock *tp; 1289dfb4b9dcSDavid S. Miller struct tcp_skb_cb *tcb; 129033ad798cSAdam Langley struct tcp_out_options opts; 129195c96174SEric Dumazet unsigned int tcp_options_size, tcp_header_size; 12928c72c65bSEric Dumazet struct sk_buff *oskb = NULL; 12931e03d32bSDmitry Safonov struct tcp_key key; 12941da177e4SLinus Torvalds struct tcphdr *th; 1295a7a25630SEric Dumazet u64 prior_wstamp; 12961da177e4SLinus Torvalds int err; 12971da177e4SLinus Torvalds 1298dfb4b9dcSDavid S. Miller BUG_ON(!skb || !tcp_skb_pcount(skb)); 12996f094b9eSLawrence Brakmo tp = tcp_sk(sk); 13007f12422cSYuchung Cheng prior_wstamp = tp->tcp_wstamp_ns; 13017f12422cSYuchung Cheng tp->tcp_wstamp_ns = max(tp->tcp_wstamp_ns, tp->tcp_clock_cache); 1302a1ac9c8aSMartin KaFai Lau skb_set_delivery_time(skb, tp->tcp_wstamp_ns, true); 1303ccdbb6e9SEric Dumazet if (clone_it) { 13048c72c65bSEric Dumazet oskb = skb; 1305e2080072SEric Dumazet 1306e2080072SEric Dumazet tcp_skb_tsorted_save(oskb) { 1307e2080072SEric Dumazet if (unlikely(skb_cloned(oskb))) 1308e2080072SEric Dumazet skb = pskb_copy(oskb, gfp_mask); 1309dfb4b9dcSDavid S. Miller else 1310e2080072SEric Dumazet skb = skb_clone(oskb, gfp_mask); 1311e2080072SEric Dumazet } tcp_skb_tsorted_restore(oskb); 1312e2080072SEric Dumazet 1313dfb4b9dcSDavid S. Miller if (unlikely(!skb)) 1314dfb4b9dcSDavid S. Miller return -ENOBUFS; 1315b738a185SEric Dumazet /* retransmit skbs might have a non zero value in skb->dev 1316b738a185SEric Dumazet * because skb->dev is aliased with skb->rbnode.rb_left 1317b738a185SEric Dumazet */ 1318b738a185SEric Dumazet skb->dev = NULL; 1319dfb4b9dcSDavid S. Miller } 13205f6188a8SEric Dumazet 1321dfb4b9dcSDavid S. Miller inet = inet_sk(sk); 1322dfb4b9dcSDavid S. Miller tcb = TCP_SKB_CB(skb); 132333ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 13241da177e4SLinus Torvalds 13251e03d32bSDmitry Safonov tcp_get_current_key(sk, &key); 1326051ba674SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) { 13271e03d32bSDmitry Safonov tcp_options_size = tcp_syn_options(sk, skb, &opts, &key); 1328051ba674SEric Dumazet } else { 13291e03d32bSDmitry Safonov tcp_options_size = tcp_established_options(sk, skb, &opts, &key); 1330051ba674SEric Dumazet /* Force a PSH flag on all (GSO) packets to expedite GRO flush 1331051ba674SEric Dumazet * at receiver : This slightly improve GRO performance. 1332051ba674SEric Dumazet * Note that we do not force the PSH flag for non GSO packets, 1333051ba674SEric Dumazet * because they might be sent under high congestion events, 1334051ba674SEric Dumazet * and in this case it is better to delay the delivery of 1-MSS 1335051ba674SEric Dumazet * packets and thus the corresponding ACK packet that would 1336051ba674SEric Dumazet * release the following packet. 1337051ba674SEric Dumazet */ 1338051ba674SEric Dumazet if (tcp_skb_pcount(skb) > 1) 1339051ba674SEric Dumazet tcb->tcp_flags |= TCPHDR_PSH; 1340051ba674SEric Dumazet } 134133ad798cSAdam Langley tcp_header_size = tcp_options_size + sizeof(struct tcphdr); 13421da177e4SLinus Torvalds 1343726e9e8bSEric Dumazet /* We set skb->ooo_okay to one if this packet can select 1344726e9e8bSEric Dumazet * a different TX queue than prior packets of this flow, 1345726e9e8bSEric Dumazet * to avoid self inflicted reorders. 1346726e9e8bSEric Dumazet * The 'other' queue decision is based on current cpu number 1347726e9e8bSEric Dumazet * if XPS is enabled, or sk->sk_txhash otherwise. 1348726e9e8bSEric Dumazet * We can switch to another (and better) queue if: 1349726e9e8bSEric Dumazet * 1) No packet with payload is in qdisc/device queues. 1350726e9e8bSEric Dumazet * Delays in TX completion can defeat the test 1351726e9e8bSEric Dumazet * even if packets were already sent. 1352726e9e8bSEric Dumazet * 2) Or rtx queue is empty. 1353726e9e8bSEric Dumazet * This mitigates above case if ACK packets for 1354726e9e8bSEric Dumazet * all prior packets were already processed. 1355547669d4SEric Dumazet */ 1356726e9e8bSEric Dumazet skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1) || 1357726e9e8bSEric Dumazet tcp_rtx_queue_empty(sk); 13581da177e4SLinus Torvalds 135938ab52e8SEric Dumazet /* If we had to use memory reserve to allocate this skb, 136038ab52e8SEric Dumazet * this might cause drops if packet is looped back : 136138ab52e8SEric Dumazet * Other socket might not have SOCK_MEMALLOC. 136238ab52e8SEric Dumazet * Packets not looped back do not care about pfmemalloc. 136338ab52e8SEric Dumazet */ 136438ab52e8SEric Dumazet skb->pfmemalloc = 0; 136538ab52e8SEric Dumazet 1366aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 1367aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 136846d3ceabSEric Dumazet 136946d3ceabSEric Dumazet skb_orphan(skb); 137046d3ceabSEric Dumazet skb->sk = sk; 13711d2077acSEric Dumazet skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree; 137214afee4bSReshetova, Elena refcount_add(skb->truesize, &sk->sk_wmem_alloc); 13731da177e4SLinus Torvalds 1374eb44ad4eSEric Dumazet skb_set_dst_pending_confirm(skb, READ_ONCE(sk->sk_dst_pending_confirm)); 1375c3a2e837SJulian Anastasov 13761da177e4SLinus Torvalds /* Build TCP header and checksum it. */ 1377ea1627c2SEric Dumazet th = (struct tcphdr *)skb->data; 1378c720c7e8SEric Dumazet th->source = inet->inet_sport; 1379c720c7e8SEric Dumazet th->dest = inet->inet_dport; 13801da177e4SLinus Torvalds th->seq = htonl(tcb->seq); 13812987babbSYuchung Cheng th->ack_seq = htonl(rcv_nxt); 1382df7a3b07SAl Viro *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | 13834de075e0SEric Dumazet tcb->tcp_flags); 1384dfb4b9dcSDavid S. Miller 13851da177e4SLinus Torvalds th->check = 0; 13861da177e4SLinus Torvalds th->urg_ptr = 0; 13871da177e4SLinus Torvalds 138833f5f57eSIlpo Järvinen /* The urg_mode check is necessary during a below snd_una win probe */ 13897691367dSHerbert Xu if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { 13907691367dSHerbert Xu if (before(tp->snd_up, tcb->seq + 0x10000)) { 13911da177e4SLinus Torvalds th->urg_ptr = htons(tp->snd_up - tcb->seq); 13921da177e4SLinus Torvalds th->urg = 1; 13937691367dSHerbert Xu } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { 13940eae88f3SEric Dumazet th->urg_ptr = htons(0xFFFF); 13957691367dSHerbert Xu th->urg = 1; 13967691367dSHerbert Xu } 13971da177e4SLinus Torvalds } 13981da177e4SLinus Torvalds 139951466a75SEric Dumazet skb_shinfo(skb)->gso_type = sk->sk_gso_type; 1400ea1627c2SEric Dumazet if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) { 1401ea1627c2SEric Dumazet th->window = htons(tcp_select_window(sk)); 1402ea1627c2SEric Dumazet tcp_ecn_send(sk, skb, th, tcp_header_size); 1403ea1627c2SEric Dumazet } else { 1404ea1627c2SEric Dumazet /* RFC1323: The window in SYN & SYN/ACK segments 1405ea1627c2SEric Dumazet * is never scaled. 1406ea1627c2SEric Dumazet */ 1407ea1627c2SEric Dumazet th->window = htons(min(tp->rcv_wnd, 65535U)); 1408ea1627c2SEric Dumazet } 1409fa3fe2b1SFlorian Westphal 141006b22ef2SDmitry Safonov tcp_options_write(th, tp, NULL, &opts, &key); 1411fa3fe2b1SFlorian Westphal 14121e03d32bSDmitry Safonov if (tcp_key_is_md5(&key)) { 1413cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 1414cfb6eeb4SYOSHIFUJI Hideaki /* Calculate the MD5 hash, as we have all we need now */ 1415aba54656SEric Dumazet sk_gso_disable(sk); 1416bd0388aeSWilliam Allen Simpson tp->af_specific->calc_md5_hash(opts.hash_location, 14171e03d32bSDmitry Safonov key.md5_key, sk, skb); 1418cfb6eeb4SYOSHIFUJI Hideaki #endif 14191e03d32bSDmitry Safonov } else if (tcp_key_is_ao(&key)) { 14201e03d32bSDmitry Safonov int err; 14211e03d32bSDmitry Safonov 14221e03d32bSDmitry Safonov err = tcp_ao_transmit_skb(sk, skb, key.ao_key, th, 14231e03d32bSDmitry Safonov opts.hash_location); 14241e03d32bSDmitry Safonov if (err) { 14251e03d32bSDmitry Safonov kfree_skb_reason(skb, SKB_DROP_REASON_NOT_SPECIFIED); 14261e03d32bSDmitry Safonov return -ENOMEM; 14271e03d32bSDmitry Safonov } 14281e03d32bSDmitry Safonov } 1429cfb6eeb4SYOSHIFUJI Hideaki 1430331fca43SMartin KaFai Lau /* BPF prog is the last one writing header option */ 1431331fca43SMartin KaFai Lau bpf_skops_write_hdr_opt(sk, skb, NULL, NULL, 0, &opts); 1432331fca43SMartin KaFai Lau 1433dd2e0b86SEric Dumazet INDIRECT_CALL_INET(icsk->icsk_af_ops->send_check, 1434dd2e0b86SEric Dumazet tcp_v6_send_check, tcp_v4_send_check, 1435dd2e0b86SEric Dumazet sk, skb); 14361da177e4SLinus Torvalds 14374de075e0SEric Dumazet if (likely(tcb->tcp_flags & TCPHDR_ACK)) 1438059217c1SNeal Cardwell tcp_event_ack_sent(sk, rcv_nxt); 14391da177e4SLinus Torvalds 1440a44d6eacSMartin KaFai Lau if (skb->len != tcp_header_size) { 1441cf533ea5SEric Dumazet tcp_event_data_sent(tp, sk); 1442a44d6eacSMartin KaFai Lau tp->data_segs_out += tcp_skb_pcount(skb); 1443ba113c3aSWei Wang tp->bytes_sent += skb->len - tcp_header_size; 1444a44d6eacSMartin KaFai Lau } 14451da177e4SLinus Torvalds 1446bd37a088SWei Yongjun if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) 1447aa2ea058STom Herbert TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, 1448aa2ea058STom Herbert tcp_skb_pcount(skb)); 14491da177e4SLinus Torvalds 14502efd055cSMarcelo Ricardo Leitner tp->segs_out += tcp_skb_pcount(skb); 14510ae5b43dSYuchung Cheng skb_set_hash_from_sk(skb, sk); 1452f69ad292SEric Dumazet /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */ 1453cd7d8498SEric Dumazet skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb); 1454f69ad292SEric Dumazet skb_shinfo(skb)->gso_size = tcp_skb_mss(skb); 1455cd7d8498SEric Dumazet 1456d3edd06eSEric Dumazet /* Leave earliest departure time in skb->tstamp (skb->skb_mstamp_ns) */ 1457971f10ecSEric Dumazet 1458971f10ecSEric Dumazet /* Cleanup our debris for IP stacks */ 1459971f10ecSEric Dumazet memset(skb->cb, 0, max(sizeof(struct inet_skb_parm), 1460971f10ecSEric Dumazet sizeof(struct inet6_skb_parm))); 1461971f10ecSEric Dumazet 1462a842fe14SEric Dumazet tcp_add_tx_delay(skb, tp); 1463a842fe14SEric Dumazet 146405e22e83SEric Dumazet err = INDIRECT_CALL_INET(icsk->icsk_af_ops->queue_xmit, 146505e22e83SEric Dumazet inet6_csk_xmit, ip_queue_xmit, 146605e22e83SEric Dumazet sk, skb, &inet->cork.fl); 14677faee5c0SEric Dumazet 14688c72c65bSEric Dumazet if (unlikely(err > 0)) { 14695ee2c941SChristoph Paasch tcp_enter_cwr(sk); 14708c72c65bSEric Dumazet err = net_xmit_eval(err); 14718c72c65bSEric Dumazet } 1472fc225799SEric Dumazet if (!err && oskb) { 1473a7a25630SEric Dumazet tcp_update_skb_after_send(sk, oskb, prior_wstamp); 1474fc225799SEric Dumazet tcp_rate_skb_sent(sk, oskb); 1475fc225799SEric Dumazet } 14768c72c65bSEric Dumazet return err; 14771da177e4SLinus Torvalds } 14781da177e4SLinus Torvalds 14792987babbSYuchung Cheng static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, 14802987babbSYuchung Cheng gfp_t gfp_mask) 14812987babbSYuchung Cheng { 14822987babbSYuchung Cheng return __tcp_transmit_skb(sk, skb, clone_it, gfp_mask, 14832987babbSYuchung Cheng tcp_sk(sk)->rcv_nxt); 14842987babbSYuchung Cheng } 14852987babbSYuchung Cheng 148667edfef7SAndi Kleen /* This routine just queues the buffer for sending. 14871da177e4SLinus Torvalds * 14881da177e4SLinus Torvalds * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, 14891da177e4SLinus Torvalds * otherwise socket can stall. 14901da177e4SLinus Torvalds */ 14911da177e4SLinus Torvalds static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) 14921da177e4SLinus Torvalds { 14931da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 14941da177e4SLinus Torvalds 14951da177e4SLinus Torvalds /* Advance write_seq and place onto the write_queue. */ 14960f317464SEric Dumazet WRITE_ONCE(tp->write_seq, TCP_SKB_CB(skb)->end_seq); 1497f4a775d1SEric Dumazet __skb_header_release(skb); 1498fe067e8aSDavid S. Miller tcp_add_write_queue_tail(sk, skb); 1499ab4e846aSEric Dumazet sk_wmem_queued_add(sk, skb->truesize); 15003ab224beSHideo Aoki sk_mem_charge(sk, skb->truesize); 15011da177e4SLinus Torvalds } 15021da177e4SLinus Torvalds 150367edfef7SAndi Kleen /* Initialize TSO segments for a packet. */ 15045bbb432cSEric Dumazet static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now) 1505f6302d1dSDavid S. Miller { 15064a64fd6cSEric Dumazet if (skb->len <= mss_now) { 1507f6302d1dSDavid S. Miller /* Avoid the costly divide in the normal 1508f6302d1dSDavid S. Miller * non-TSO case. 1509f6302d1dSDavid S. Miller */ 1510cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, 1); 1511f69ad292SEric Dumazet TCP_SKB_CB(skb)->tcp_gso_size = 0; 1512f6302d1dSDavid S. Miller } else { 1513cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now)); 1514f69ad292SEric Dumazet TCP_SKB_CB(skb)->tcp_gso_size = mss_now; 15151da177e4SLinus Torvalds } 15161da177e4SLinus Torvalds } 15171da177e4SLinus Torvalds 1518797108d1SIlpo Järvinen /* Pcount in the middle of the write queue got changed, we need to do various 1519797108d1SIlpo Järvinen * tweaks to fix counters 1520797108d1SIlpo Järvinen */ 1521cf533ea5SEric Dumazet static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) 1522797108d1SIlpo Järvinen { 1523797108d1SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1524797108d1SIlpo Järvinen 1525797108d1SIlpo Järvinen tp->packets_out -= decr; 1526797108d1SIlpo Järvinen 1527797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 1528797108d1SIlpo Järvinen tp->sacked_out -= decr; 1529797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 1530797108d1SIlpo Järvinen tp->retrans_out -= decr; 1531797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 1532797108d1SIlpo Järvinen tp->lost_out -= decr; 1533797108d1SIlpo Järvinen 1534797108d1SIlpo Järvinen /* Reno case is special. Sigh... */ 1535797108d1SIlpo Järvinen if (tcp_is_reno(tp) && decr > 0) 1536797108d1SIlpo Järvinen tp->sacked_out -= min_t(u32, tp->sacked_out, decr); 1537797108d1SIlpo Järvinen 1538797108d1SIlpo Järvinen if (tp->lost_skb_hint && 1539797108d1SIlpo Järvinen before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && 1540713bafeaSYuchung Cheng (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) 1541797108d1SIlpo Järvinen tp->lost_cnt_hint -= decr; 1542797108d1SIlpo Järvinen 1543797108d1SIlpo Järvinen tcp_verify_left_out(tp); 1544797108d1SIlpo Järvinen } 1545797108d1SIlpo Järvinen 15460a2cf20cSSoheil Hassas Yeganeh static bool tcp_has_tx_tstamp(const struct sk_buff *skb) 15470a2cf20cSSoheil Hassas Yeganeh { 15480a2cf20cSSoheil Hassas Yeganeh return TCP_SKB_CB(skb)->txstamp_ack || 15490a2cf20cSSoheil Hassas Yeganeh (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP); 15500a2cf20cSSoheil Hassas Yeganeh } 15510a2cf20cSSoheil Hassas Yeganeh 1552490cc7d0SWillem de Bruijn static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2) 1553490cc7d0SWillem de Bruijn { 1554490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo = skb_shinfo(skb); 1555490cc7d0SWillem de Bruijn 15560a2cf20cSSoheil Hassas Yeganeh if (unlikely(tcp_has_tx_tstamp(skb)) && 1557490cc7d0SWillem de Bruijn !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) { 1558490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo2 = skb_shinfo(skb2); 1559490cc7d0SWillem de Bruijn u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP; 1560490cc7d0SWillem de Bruijn 1561490cc7d0SWillem de Bruijn shinfo->tx_flags &= ~tsflags; 1562490cc7d0SWillem de Bruijn shinfo2->tx_flags |= tsflags; 1563490cc7d0SWillem de Bruijn swap(shinfo->tskey, shinfo2->tskey); 1564b51e13faSMartin KaFai Lau TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack; 1565b51e13faSMartin KaFai Lau TCP_SKB_CB(skb)->txstamp_ack = 0; 1566490cc7d0SWillem de Bruijn } 1567490cc7d0SWillem de Bruijn } 1568490cc7d0SWillem de Bruijn 1569a166140eSMartin KaFai Lau static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2) 1570a166140eSMartin KaFai Lau { 1571a166140eSMartin KaFai Lau TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor; 1572a166140eSMartin KaFai Lau TCP_SKB_CB(skb)->eor = 0; 1573a166140eSMartin KaFai Lau } 1574a166140eSMartin KaFai Lau 157575c119afSEric Dumazet /* Insert buff after skb on the write or rtx queue of sk. */ 157675c119afSEric Dumazet static void tcp_insert_write_queue_after(struct sk_buff *skb, 157775c119afSEric Dumazet struct sk_buff *buff, 157875c119afSEric Dumazet struct sock *sk, 157975c119afSEric Dumazet enum tcp_queue tcp_queue) 158075c119afSEric Dumazet { 158175c119afSEric Dumazet if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE) 158275c119afSEric Dumazet __skb_queue_after(&sk->sk_write_queue, skb, buff); 158375c119afSEric Dumazet else 158475c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, buff); 158575c119afSEric Dumazet } 158675c119afSEric Dumazet 15871da177e4SLinus Torvalds /* Function to create two new TCP segments. Shrinks the given segment 15881da177e4SLinus Torvalds * to the specified size and appends a new segment with the rest of the 15891da177e4SLinus Torvalds * packet to the list. This won't be called frequently, I hope. 15901da177e4SLinus Torvalds * Remember, these are still headerless SKBs at this point. 15911da177e4SLinus Torvalds */ 159275c119afSEric Dumazet int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue, 159375c119afSEric Dumazet struct sk_buff *skb, u32 len, 15946cc55e09SOctavian Purdila unsigned int mss_now, gfp_t gfp) 15951da177e4SLinus Torvalds { 15961da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 15971da177e4SLinus Torvalds struct sk_buff *buff; 1598b4a24397SEric Dumazet int old_factor; 1599b617158dSEric Dumazet long limit; 1600b60b49eaSHerbert Xu int nlen; 16019ce01461SIlpo Järvinen u8 flags; 16021da177e4SLinus Torvalds 16032fceec13SIlpo Järvinen if (WARN_ON(len > skb->len)) 16042fceec13SIlpo Järvinen return -EINVAL; 16056a438bbeSStephen Hemminger 1606b4a24397SEric Dumazet DEBUG_NET_WARN_ON_ONCE(skb_headlen(skb)); 16071da177e4SLinus Torvalds 1608b617158dSEric Dumazet /* tcp_sendmsg() can overshoot sk_wmem_queued by one full size skb. 1609b617158dSEric Dumazet * We need some allowance to not penalize applications setting small 1610b617158dSEric Dumazet * SO_SNDBUF values. 1611b617158dSEric Dumazet * Also allow first and last skb in retransmit queue to be split. 1612b617158dSEric Dumazet */ 16137c4e983cSAlexander Duyck limit = sk->sk_sndbuf + 2 * SKB_TRUESIZE(GSO_LEGACY_MAX_SIZE); 1614b617158dSEric Dumazet if (unlikely((sk->sk_wmem_queued >> 1) > limit && 1615b617158dSEric Dumazet tcp_queue != TCP_FRAG_IN_WRITE_QUEUE && 1616b617158dSEric Dumazet skb != tcp_rtx_queue_head(sk) && 1617b617158dSEric Dumazet skb != tcp_rtx_queue_tail(sk))) { 1618f070ef2aSEric Dumazet NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPWQUEUETOOBIG); 1619f070ef2aSEric Dumazet return -ENOMEM; 1620f070ef2aSEric Dumazet } 1621f070ef2aSEric Dumazet 1622c4777efaSEric Dumazet if (skb_unclone_keeptruesize(skb, gfp)) 16231da177e4SLinus Torvalds return -ENOMEM; 16241da177e4SLinus Torvalds 16251da177e4SLinus Torvalds /* Get a new skb... force flag on. */ 16265882efffSEric Dumazet buff = tcp_stream_alloc_skb(sk, gfp, true); 162751456b29SIan Morris if (!buff) 16281da177e4SLinus Torvalds return -ENOMEM; /* We'll just try again later. */ 162941477662SJakub Kicinski skb_copy_decrypted(buff, skb); 16305a369ca6SPaolo Abeni mptcp_skb_ext_copy(buff, skb); 1631ef5cb973SHerbert Xu 1632ab4e846aSEric Dumazet sk_wmem_queued_add(sk, buff->truesize); 16333ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1634b4a24397SEric Dumazet nlen = skb->len - len; 1635b60b49eaSHerbert Xu buff->truesize += nlen; 1636b60b49eaSHerbert Xu skb->truesize -= nlen; 16371da177e4SLinus Torvalds 16381da177e4SLinus Torvalds /* Correct the sequence numbers. */ 16391da177e4SLinus Torvalds TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 16401da177e4SLinus Torvalds TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 16411da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 16421da177e4SLinus Torvalds 16431da177e4SLinus Torvalds /* PSH and FIN should only be set in the second packet. */ 16444de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 16454de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 16464de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1647e14c3cafSHerbert Xu TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; 1648a166140eSMartin KaFai Lau tcp_skb_fragment_eor(skb, buff); 16491da177e4SLinus Torvalds 16501da177e4SLinus Torvalds skb_split(skb, buff, len); 16511da177e4SLinus Torvalds 1652a1ac9c8aSMartin KaFai Lau skb_set_delivery_time(buff, skb->tstamp, true); 1653490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 16541da177e4SLinus Torvalds 16556475be16SDavid S. Miller old_factor = tcp_skb_pcount(skb); 16566475be16SDavid S. Miller 16571da177e4SLinus Torvalds /* Fix up tso_factor for both original and new SKB. */ 16585bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 16595bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 16601da177e4SLinus Torvalds 1661b9f64820SYuchung Cheng /* Update delivered info for the new segment */ 1662b9f64820SYuchung Cheng TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx; 1663b9f64820SYuchung Cheng 16646475be16SDavid S. Miller /* If this packet has been sent out already, we must 16656475be16SDavid S. Miller * adjust the various packet counters. 16666475be16SDavid S. Miller */ 1667cf0b450cSHerbert Xu if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 16686475be16SDavid S. Miller int diff = old_factor - tcp_skb_pcount(skb) - 16696475be16SDavid S. Miller tcp_skb_pcount(buff); 16701da177e4SLinus Torvalds 1671797108d1SIlpo Järvinen if (diff) 1672797108d1SIlpo Järvinen tcp_adjust_pcount(sk, skb, diff); 16731da177e4SLinus Torvalds } 16741da177e4SLinus Torvalds 16751da177e4SLinus Torvalds /* Link BUFF into the send queue. */ 1676f4a775d1SEric Dumazet __skb_header_release(buff); 167775c119afSEric Dumazet tcp_insert_write_queue_after(skb, buff, sk, tcp_queue); 1678f67971e6SEric Dumazet if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE) 1679e2080072SEric Dumazet list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor); 16801da177e4SLinus Torvalds 16811da177e4SLinus Torvalds return 0; 16821da177e4SLinus Torvalds } 16831da177e4SLinus Torvalds 1684f4d01666SEric Dumazet /* This is similar to __pskb_pull_tail(). The difference is that pulled 1685f4d01666SEric Dumazet * data is not copied, but immediately discarded. 16861da177e4SLinus Torvalds */ 16877162fb24SEric Dumazet static int __pskb_trim_head(struct sk_buff *skb, int len) 16881da177e4SLinus Torvalds { 16897b7fc97aSEric Dumazet struct skb_shared_info *shinfo; 16901da177e4SLinus Torvalds int i, k, eat; 16911da177e4SLinus Torvalds 1692b4a24397SEric Dumazet DEBUG_NET_WARN_ON_ONCE(skb_headlen(skb)); 16931da177e4SLinus Torvalds eat = len; 16941da177e4SLinus Torvalds k = 0; 16957b7fc97aSEric Dumazet shinfo = skb_shinfo(skb); 16967b7fc97aSEric Dumazet for (i = 0; i < shinfo->nr_frags; i++) { 16977b7fc97aSEric Dumazet int size = skb_frag_size(&shinfo->frags[i]); 16989e903e08SEric Dumazet 16999e903e08SEric Dumazet if (size <= eat) { 1700aff65da0SIan Campbell skb_frag_unref(skb, i); 17019e903e08SEric Dumazet eat -= size; 17021da177e4SLinus Torvalds } else { 17037b7fc97aSEric Dumazet shinfo->frags[k] = shinfo->frags[i]; 17041da177e4SLinus Torvalds if (eat) { 1705b54c9d5bSJonathan Lemon skb_frag_off_add(&shinfo->frags[k], eat); 17067b7fc97aSEric Dumazet skb_frag_size_sub(&shinfo->frags[k], eat); 17071da177e4SLinus Torvalds eat = 0; 17081da177e4SLinus Torvalds } 17091da177e4SLinus Torvalds k++; 17101da177e4SLinus Torvalds } 17111da177e4SLinus Torvalds } 17127b7fc97aSEric Dumazet shinfo->nr_frags = k; 17131da177e4SLinus Torvalds 17141da177e4SLinus Torvalds skb->data_len -= len; 17151da177e4SLinus Torvalds skb->len = skb->data_len; 17167162fb24SEric Dumazet return len; 17171da177e4SLinus Torvalds } 17181da177e4SLinus Torvalds 171967edfef7SAndi Kleen /* Remove acked data from a packet in the transmit queue. */ 17201da177e4SLinus Torvalds int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 17211da177e4SLinus Torvalds { 17227162fb24SEric Dumazet u32 delta_truesize; 17237162fb24SEric Dumazet 1724c4777efaSEric Dumazet if (skb_unclone_keeptruesize(skb, GFP_ATOMIC)) 17251da177e4SLinus Torvalds return -ENOMEM; 17261da177e4SLinus Torvalds 17277162fb24SEric Dumazet delta_truesize = __pskb_trim_head(skb, len); 17281da177e4SLinus Torvalds 17291da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq += len; 17301da177e4SLinus Torvalds 17317162fb24SEric Dumazet skb->truesize -= delta_truesize; 1732ab4e846aSEric Dumazet sk_wmem_queued_add(sk, -delta_truesize); 17339b65b17dSTalal Ahmad if (!skb_zcopy_pure(skb)) 17347162fb24SEric Dumazet sk_mem_uncharge(sk, delta_truesize); 17351da177e4SLinus Torvalds 17365b35e1e6SNeal Cardwell /* Any change of skb->len requires recalculation of tso factor. */ 17371da177e4SLinus Torvalds if (tcp_skb_pcount(skb) > 1) 17385bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb)); 17391da177e4SLinus Torvalds 17401da177e4SLinus Torvalds return 0; 17411da177e4SLinus Torvalds } 17421da177e4SLinus Torvalds 17431b63edd6SYuchung Cheng /* Calculate MSS not accounting any TCP options. */ 17441b63edd6SYuchung Cheng static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu) 17455d424d5aSJohn Heffner { 1746cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1747cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 17485d424d5aSJohn Heffner int mss_now; 17495d424d5aSJohn Heffner 17505d424d5aSJohn Heffner /* Calculate base mss without TCP options: 17515d424d5aSJohn Heffner It is MMS_S - sizeof(tcphdr) of rfc1122 17525d424d5aSJohn Heffner */ 17535d424d5aSJohn Heffner mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 17545d424d5aSJohn Heffner 17555d424d5aSJohn Heffner /* Clamp it (mss_clamp does not include tcp options) */ 17565d424d5aSJohn Heffner if (mss_now > tp->rx_opt.mss_clamp) 17575d424d5aSJohn Heffner mss_now = tp->rx_opt.mss_clamp; 17585d424d5aSJohn Heffner 17595d424d5aSJohn Heffner /* Now subtract optional transport overhead */ 17605d424d5aSJohn Heffner mss_now -= icsk->icsk_ext_hdr_len; 17615d424d5aSJohn Heffner 17625d424d5aSJohn Heffner /* Then reserve room for full set of TCP options and 8 bytes of data */ 176378eb166cSKuniyuki Iwashima mss_now = max(mss_now, 176478eb166cSKuniyuki Iwashima READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_snd_mss)); 17655d424d5aSJohn Heffner return mss_now; 17665d424d5aSJohn Heffner } 17675d424d5aSJohn Heffner 17681b63edd6SYuchung Cheng /* Calculate MSS. Not accounting for SACKs here. */ 17691b63edd6SYuchung Cheng int tcp_mtu_to_mss(struct sock *sk, int pmtu) 17701b63edd6SYuchung Cheng { 17711b63edd6SYuchung Cheng /* Subtract TCP options size, not including SACKs */ 17721b63edd6SYuchung Cheng return __tcp_mtu_to_mss(sk, pmtu) - 17731b63edd6SYuchung Cheng (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr)); 17741b63edd6SYuchung Cheng } 1775c7bb4b89SEric Dumazet EXPORT_SYMBOL(tcp_mtu_to_mss); 17761b63edd6SYuchung Cheng 17775d424d5aSJohn Heffner /* Inverse of above */ 177867469601SEric Dumazet int tcp_mss_to_mtu(struct sock *sk, int mss) 17795d424d5aSJohn Heffner { 1780cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1781cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 17825d424d5aSJohn Heffner 1783e57a3447SYan Zhai return mss + 17845d424d5aSJohn Heffner tp->tcp_header_len + 17855d424d5aSJohn Heffner icsk->icsk_ext_hdr_len + 17865d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 17875d424d5aSJohn Heffner } 1788556c6b46SNeal Cardwell EXPORT_SYMBOL(tcp_mss_to_mtu); 17895d424d5aSJohn Heffner 179067edfef7SAndi Kleen /* MTU probing init per socket */ 17915d424d5aSJohn Heffner void tcp_mtup_init(struct sock *sk) 17925d424d5aSJohn Heffner { 17935d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 17945d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 1795b0f9ca53SFan Du struct net *net = sock_net(sk); 17965d424d5aSJohn Heffner 1797f47d00e0SKuniyuki Iwashima icsk->icsk_mtup.enabled = READ_ONCE(net->ipv4.sysctl_tcp_mtu_probing) > 1; 17985d424d5aSJohn Heffner icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 17995d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 180088d78bc0SKuniyuki Iwashima icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, READ_ONCE(net->ipv4.sysctl_tcp_base_mss)); 18015d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 180205cbc0dbSFan Du if (icsk->icsk_mtup.enabled) 1803c74df29aSEric Dumazet icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; 18045d424d5aSJohn Heffner } 18054bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_mtup_init); 18065d424d5aSJohn Heffner 18071da177e4SLinus Torvalds /* This function synchronize snd mss to current pmtu/exthdr set. 18081da177e4SLinus Torvalds 18091da177e4SLinus Torvalds tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 18101da177e4SLinus Torvalds for TCP options, but includes only bare TCP header. 18111da177e4SLinus Torvalds 18121da177e4SLinus Torvalds tp->rx_opt.mss_clamp is mss negotiated at connection setup. 1813caa20d9aSStephen Hemminger It is minimum of user_mss and mss received with SYN. 18141da177e4SLinus Torvalds It also does not include TCP options. 18151da177e4SLinus Torvalds 1816d83d8461SArnaldo Carvalho de Melo inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. 18171da177e4SLinus Torvalds 18181da177e4SLinus Torvalds tp->mss_cache is current effective sending mss, including 18191da177e4SLinus Torvalds all tcp options except for SACKs. It is evaluated, 18201da177e4SLinus Torvalds taking into account current pmtu, but never exceeds 18211da177e4SLinus Torvalds tp->rx_opt.mss_clamp. 18221da177e4SLinus Torvalds 18231da177e4SLinus Torvalds NOTE1. rfc1122 clearly states that advertised MSS 18241da177e4SLinus Torvalds DOES NOT include either tcp or ip options. 18251da177e4SLinus Torvalds 1826d83d8461SArnaldo Carvalho de Melo NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache 1827d83d8461SArnaldo Carvalho de Melo are READ ONLY outside this function. --ANK (980731) 18281da177e4SLinus Torvalds */ 18291da177e4SLinus Torvalds unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 18301da177e4SLinus Torvalds { 18311da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1832d83d8461SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 18335d424d5aSJohn Heffner int mss_now; 18341da177e4SLinus Torvalds 18355d424d5aSJohn Heffner if (icsk->icsk_mtup.search_high > pmtu) 18365d424d5aSJohn Heffner icsk->icsk_mtup.search_high = pmtu; 18371da177e4SLinus Torvalds 18385d424d5aSJohn Heffner mss_now = tcp_mtu_to_mss(sk, pmtu); 1839409d22b4SIlpo Järvinen mss_now = tcp_bound_to_half_wnd(tp, mss_now); 18401da177e4SLinus Torvalds 18411da177e4SLinus Torvalds /* And store cached results */ 1842d83d8461SArnaldo Carvalho de Melo icsk->icsk_pmtu_cookie = pmtu; 18435d424d5aSJohn Heffner if (icsk->icsk_mtup.enabled) 18445d424d5aSJohn Heffner mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); 1845c1b4a7e6SDavid S. Miller tp->mss_cache = mss_now; 18461da177e4SLinus Torvalds 18471da177e4SLinus Torvalds return mss_now; 18481da177e4SLinus Torvalds } 18494bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_sync_mss); 18501da177e4SLinus Torvalds 18511da177e4SLinus Torvalds /* Compute the current effective MSS, taking SACKs and IP options, 18521da177e4SLinus Torvalds * and even PMTU discovery events into account. 18531da177e4SLinus Torvalds */ 18540c54b85fSIlpo Järvinen unsigned int tcp_current_mss(struct sock *sk) 18551da177e4SLinus Torvalds { 1856cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1857cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1858c1b4a7e6SDavid S. Miller u32 mss_now; 185995c96174SEric Dumazet unsigned int header_len; 186033ad798cSAdam Langley struct tcp_out_options opts; 18611e03d32bSDmitry Safonov struct tcp_key key; 18621da177e4SLinus Torvalds 1863c1b4a7e6SDavid S. Miller mss_now = tp->mss_cache; 1864c1b4a7e6SDavid S. Miller 18651da177e4SLinus Torvalds if (dst) { 18661da177e4SLinus Torvalds u32 mtu = dst_mtu(dst); 1867d83d8461SArnaldo Carvalho de Melo if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 18681da177e4SLinus Torvalds mss_now = tcp_sync_mss(sk, mtu); 18691da177e4SLinus Torvalds } 18701e03d32bSDmitry Safonov tcp_get_current_key(sk, &key); 18711e03d32bSDmitry Safonov header_len = tcp_established_options(sk, NULL, &opts, &key) + 187233ad798cSAdam Langley sizeof(struct tcphdr); 187333ad798cSAdam Langley /* The mss_cache is sized based on tp->tcp_header_len, which assumes 187433ad798cSAdam Langley * some common options. If this is an odd packet (because we have SACK 187533ad798cSAdam Langley * blocks etc) then our calculated header_len will be different, and 187633ad798cSAdam Langley * we have to adjust mss_now correspondingly */ 187733ad798cSAdam Langley if (header_len != tp->tcp_header_len) { 187833ad798cSAdam Langley int delta = (int) header_len - tp->tcp_header_len; 187933ad798cSAdam Langley mss_now -= delta; 188033ad798cSAdam Langley } 1881cfb6eeb4SYOSHIFUJI Hideaki 18821da177e4SLinus Torvalds return mss_now; 18831da177e4SLinus Torvalds } 18841da177e4SLinus Torvalds 188586fd14adSWeiping Pan /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. 188686fd14adSWeiping Pan * As additional protections, we do not touch cwnd in retransmission phases, 188786fd14adSWeiping Pan * and if application hit its sndbuf limit recently. 188886fd14adSWeiping Pan */ 188986fd14adSWeiping Pan static void tcp_cwnd_application_limited(struct sock *sk) 1890a762a980SDavid S. Miller { 18919e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1892a762a980SDavid S. Miller 189386fd14adSWeiping Pan if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && 189486fd14adSWeiping Pan sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 189586fd14adSWeiping Pan /* Limited by application or receiver window. */ 189686fd14adSWeiping Pan u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); 189786fd14adSWeiping Pan u32 win_used = max(tp->snd_cwnd_used, init_win); 189840570375SEric Dumazet if (win_used < tcp_snd_cwnd(tp)) { 189986fd14adSWeiping Pan tp->snd_ssthresh = tcp_current_ssthresh(sk); 190040570375SEric Dumazet tcp_snd_cwnd_set(tp, (tcp_snd_cwnd(tp) + win_used) >> 1); 190186fd14adSWeiping Pan } 190286fd14adSWeiping Pan tp->snd_cwnd_used = 0; 190386fd14adSWeiping Pan } 1904c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 190586fd14adSWeiping Pan } 190686fd14adSWeiping Pan 1907ca8a2263SNeal Cardwell static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited) 1908a762a980SDavid S. Miller { 19091b1fc3fdSWei Wang const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 1910a762a980SDavid S. Miller struct tcp_sock *tp = tcp_sk(sk); 1911a762a980SDavid S. Miller 1912f4ce91ceSNeal Cardwell /* Track the strongest available signal of the degree to which the cwnd 1913f4ce91ceSNeal Cardwell * is fully utilized. If cwnd-limited then remember that fact for the 1914f4ce91ceSNeal Cardwell * current window. If not cwnd-limited then track the maximum number of 1915f4ce91ceSNeal Cardwell * outstanding packets in the current window. (If cwnd-limited then we 1916f4ce91ceSNeal Cardwell * chose to not update tp->max_packets_out to avoid an extra else 1917f4ce91ceSNeal Cardwell * clause with no functional impact.) 1918ca8a2263SNeal Cardwell */ 1919f4ce91ceSNeal Cardwell if (!before(tp->snd_una, tp->cwnd_usage_seq) || 1920f4ce91ceSNeal Cardwell is_cwnd_limited || 1921f4ce91ceSNeal Cardwell (!tp->is_cwnd_limited && 1922f4ce91ceSNeal Cardwell tp->packets_out > tp->max_packets_out)) { 1923ca8a2263SNeal Cardwell tp->is_cwnd_limited = is_cwnd_limited; 1924f4ce91ceSNeal Cardwell tp->max_packets_out = tp->packets_out; 1925f4ce91ceSNeal Cardwell tp->cwnd_usage_seq = tp->snd_nxt; 1926ca8a2263SNeal Cardwell } 1927e114a710SEric Dumazet 192824901551SEric Dumazet if (tcp_is_cwnd_limited(sk)) { 1929a762a980SDavid S. Miller /* Network is feed fully. */ 1930a762a980SDavid S. Miller tp->snd_cwnd_used = 0; 1931c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 1932a762a980SDavid S. Miller } else { 1933a762a980SDavid S. Miller /* Network starves. */ 1934a762a980SDavid S. Miller if (tp->packets_out > tp->snd_cwnd_used) 1935a762a980SDavid S. Miller tp->snd_cwnd_used = tp->packets_out; 1936a762a980SDavid S. Miller 19374845b571SKuniyuki Iwashima if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle) && 1938c2203cf7SEric Dumazet (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto && 19391b1fc3fdSWei Wang !ca_ops->cong_control) 1940a762a980SDavid S. Miller tcp_cwnd_application_limited(sk); 1941b0f71bd3SFrancis Yan 1942b0f71bd3SFrancis Yan /* The following conditions together indicate the starvation 1943b0f71bd3SFrancis Yan * is caused by insufficient sender buffer: 1944b0f71bd3SFrancis Yan * 1) just sent some data (see tcp_write_xmit) 1945b0f71bd3SFrancis Yan * 2) not cwnd limited (this else condition) 194675c119afSEric Dumazet * 3) no more data to send (tcp_write_queue_empty()) 1947b0f71bd3SFrancis Yan * 4) application is hitting buffer limit (SOCK_NOSPACE) 1948b0f71bd3SFrancis Yan */ 194975c119afSEric Dumazet if (tcp_write_queue_empty(sk) && sk->sk_socket && 1950b0f71bd3SFrancis Yan test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) && 1951b0f71bd3SFrancis Yan (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) 1952b0f71bd3SFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED); 1953a762a980SDavid S. Miller } 1954a762a980SDavid S. Miller } 1955a762a980SDavid S. Miller 1956d4589926SEric Dumazet /* Minshall's variant of the Nagle send check. */ 1957d4589926SEric Dumazet static bool tcp_minshall_check(const struct tcp_sock *tp) 1958d4589926SEric Dumazet { 1959d4589926SEric Dumazet return after(tp->snd_sml, tp->snd_una) && 1960d4589926SEric Dumazet !after(tp->snd_sml, tp->snd_nxt); 1961d4589926SEric Dumazet } 1962d4589926SEric Dumazet 1963d4589926SEric Dumazet /* Update snd_sml if this skb is under mss 1964d4589926SEric Dumazet * Note that a TSO packet might end with a sub-mss segment 1965d4589926SEric Dumazet * The test is really : 1966d4589926SEric Dumazet * if ((skb->len % mss) != 0) 1967d4589926SEric Dumazet * tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1968d4589926SEric Dumazet * But we can avoid doing the divide again given we already have 1969d4589926SEric Dumazet * skb_pcount = skb->len / mss_now 19700e3a4803SIlpo Järvinen */ 1971d4589926SEric Dumazet static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now, 1972d4589926SEric Dumazet const struct sk_buff *skb) 1973d4589926SEric Dumazet { 1974d4589926SEric Dumazet if (skb->len < tcp_skb_pcount(skb) * mss_now) 1975d4589926SEric Dumazet tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1976d4589926SEric Dumazet } 1977d4589926SEric Dumazet 1978d4589926SEric Dumazet /* Return false, if packet can be sent now without violation Nagle's rules: 1979d4589926SEric Dumazet * 1. It is full sized. (provided by caller in %partial bool) 1980d4589926SEric Dumazet * 2. Or it contains FIN. (already checked by caller) 1981d4589926SEric Dumazet * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. 1982d4589926SEric Dumazet * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 1983d4589926SEric Dumazet * With Minshall's modification: all sent small packets are ACKed. 1984d4589926SEric Dumazet */ 1985d4589926SEric Dumazet static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp, 1986cc93fc51SPeter Pan(潘卫平) int nonagle) 1987d4589926SEric Dumazet { 1988d4589926SEric Dumazet return partial && 1989d4589926SEric Dumazet ((nonagle & TCP_NAGLE_CORK) || 1990d4589926SEric Dumazet (!nonagle && tp->packets_out && tcp_minshall_check(tp))); 1991d4589926SEric Dumazet } 1992605ad7f1SEric Dumazet 1993605ad7f1SEric Dumazet /* Return how many segs we'd like on a TSO packet, 199465466904SEric Dumazet * depending on current pacing rate, and how close the peer is. 199565466904SEric Dumazet * 199665466904SEric Dumazet * Rationale is: 199765466904SEric Dumazet * - For close peers, we rather send bigger packets to reduce 199865466904SEric Dumazet * cpu costs, because occasional losses will be repaired fast. 199965466904SEric Dumazet * - For long distance/rtt flows, we would like to get ACK clocking 200065466904SEric Dumazet * with 1 ACK per ms. 200165466904SEric Dumazet * 200265466904SEric Dumazet * Use min_rtt to help adapt TSO burst size, with smaller min_rtt resulting 200365466904SEric Dumazet * in bigger TSO bursts. We we cut the RTT-based allowance in half 200465466904SEric Dumazet * for every 2^9 usec (aka 512 us) of RTT, so that the RTT-based allowance 200565466904SEric Dumazet * is below 1500 bytes after 6 * ~500 usec = 3ms. 2006605ad7f1SEric Dumazet */ 2007dcb8c9b4SEric Dumazet static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now, 20081b3878caSNeal Cardwell int min_tso_segs) 2009605ad7f1SEric Dumazet { 201065466904SEric Dumazet unsigned long bytes; 201165466904SEric Dumazet u32 r; 2012605ad7f1SEric Dumazet 201328b24f90SEric Dumazet bytes = READ_ONCE(sk->sk_pacing_rate) >> READ_ONCE(sk->sk_pacing_shift); 2014605ad7f1SEric Dumazet 20152455e61bSKuniyuki Iwashima r = tcp_min_rtt(tcp_sk(sk)) >> READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_rtt_log); 201665466904SEric Dumazet if (r < BITS_PER_TYPE(sk->sk_gso_max_size)) 201765466904SEric Dumazet bytes += sk->sk_gso_max_size >> r; 2018605ad7f1SEric Dumazet 201965466904SEric Dumazet bytes = min_t(unsigned long, bytes, sk->sk_gso_max_size); 202065466904SEric Dumazet 202165466904SEric Dumazet return max_t(u32, bytes / mss_now, min_tso_segs); 2022605ad7f1SEric Dumazet } 2023605ad7f1SEric Dumazet 2024ed6e7268SNeal Cardwell /* Return the number of segments we want in the skb we are transmitting. 2025ed6e7268SNeal Cardwell * See if congestion control module wants to decide; otherwise, autosize. 2026ed6e7268SNeal Cardwell */ 2027ed6e7268SNeal Cardwell static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now) 2028ed6e7268SNeal Cardwell { 2029ed6e7268SNeal Cardwell const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 2030dcb8c9b4SEric Dumazet u32 min_tso, tso_segs; 2031ed6e7268SNeal Cardwell 2032dcb8c9b4SEric Dumazet min_tso = ca_ops->min_tso_segs ? 2033dcb8c9b4SEric Dumazet ca_ops->min_tso_segs(sk) : 2034e0bb4ab9SKuniyuki Iwashima READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs); 2035dcb8c9b4SEric Dumazet 2036dcb8c9b4SEric Dumazet tso_segs = tcp_tso_autosize(sk, mss_now, min_tso); 2037350c9f48SEric Dumazet return min_t(u32, tso_segs, sk->sk_gso_max_segs); 2038ed6e7268SNeal Cardwell } 2039ed6e7268SNeal Cardwell 2040d4589926SEric Dumazet /* Returns the portion of skb which can be sent right away */ 2041d4589926SEric Dumazet static unsigned int tcp_mss_split_point(const struct sock *sk, 2042d4589926SEric Dumazet const struct sk_buff *skb, 2043d4589926SEric Dumazet unsigned int mss_now, 2044d4589926SEric Dumazet unsigned int max_segs, 2045d4589926SEric Dumazet int nonagle) 2046c1b4a7e6SDavid S. Miller { 2047cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 2048d4589926SEric Dumazet u32 partial, needed, window, max_len; 2049c1b4a7e6SDavid S. Miller 205090840defSIlpo Järvinen window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 20511485348dSBen Hutchings max_len = mss_now * max_segs; 20520e3a4803SIlpo Järvinen 20531485348dSBen Hutchings if (likely(max_len <= window && skb != tcp_write_queue_tail(sk))) 20541485348dSBen Hutchings return max_len; 20550e3a4803SIlpo Järvinen 20565ea3a748SIlpo Järvinen needed = min(skb->len, window); 20575ea3a748SIlpo Järvinen 20581485348dSBen Hutchings if (max_len <= needed) 20591485348dSBen Hutchings return max_len; 20600e3a4803SIlpo Järvinen 2061d4589926SEric Dumazet partial = needed % mss_now; 2062d4589926SEric Dumazet /* If last segment is not a full MSS, check if Nagle rules allow us 2063d4589926SEric Dumazet * to include this last segment in this skb. 2064d4589926SEric Dumazet * Otherwise, we'll split the skb at last MSS boundary 2065d4589926SEric Dumazet */ 2066cc93fc51SPeter Pan(潘卫平) if (tcp_nagle_check(partial != 0, tp, nonagle)) 2067d4589926SEric Dumazet return needed - partial; 2068d4589926SEric Dumazet 2069d4589926SEric Dumazet return needed; 2070c1b4a7e6SDavid S. Miller } 2071c1b4a7e6SDavid S. Miller 2072c1b4a7e6SDavid S. Miller /* Can at least one segment of SKB be sent right now, according to the 2073c1b4a7e6SDavid S. Miller * congestion window rules? If so, return how many segments are allowed. 2074c1b4a7e6SDavid S. Miller */ 2075cf533ea5SEric Dumazet static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, 2076cf533ea5SEric Dumazet const struct sk_buff *skb) 2077c1b4a7e6SDavid S. Miller { 2078d649a7a8SEric Dumazet u32 in_flight, cwnd, halfcwnd; 2079c1b4a7e6SDavid S. Miller 2080c1b4a7e6SDavid S. Miller /* Don't be strict about the congestion window for the final FIN. */ 20814de075e0SEric Dumazet if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && 20824de075e0SEric Dumazet tcp_skb_pcount(skb) == 1) 2083c1b4a7e6SDavid S. Miller return 1; 2084c1b4a7e6SDavid S. Miller 2085c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 208640570375SEric Dumazet cwnd = tcp_snd_cwnd(tp); 2087d649a7a8SEric Dumazet if (in_flight >= cwnd) 2088c1b4a7e6SDavid S. Miller return 0; 2089d649a7a8SEric Dumazet 2090d649a7a8SEric Dumazet /* For better scheduling, ensure we have at least 2091d649a7a8SEric Dumazet * 2 GSO packets in flight. 2092d649a7a8SEric Dumazet */ 2093d649a7a8SEric Dumazet halfcwnd = max(cwnd >> 1, 1U); 2094d649a7a8SEric Dumazet return min(halfcwnd, cwnd - in_flight); 2095c1b4a7e6SDavid S. Miller } 2096c1b4a7e6SDavid S. Miller 2097b595076aSUwe Kleine-König /* Initialize TSO state of a skb. 209867edfef7SAndi Kleen * This must be invoked the first time we consider transmitting 2099c1b4a7e6SDavid S. Miller * SKB onto the wire. 2100c1b4a7e6SDavid S. Miller */ 21015bbb432cSEric Dumazet static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now) 2102c1b4a7e6SDavid S. Miller { 2103c1b4a7e6SDavid S. Miller int tso_segs = tcp_skb_pcount(skb); 2104c1b4a7e6SDavid S. Miller 2105f8269a49SIlpo Järvinen if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { 21065bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 2107c1b4a7e6SDavid S. Miller tso_segs = tcp_skb_pcount(skb); 2108c1b4a7e6SDavid S. Miller } 2109c1b4a7e6SDavid S. Miller return tso_segs; 2110c1b4a7e6SDavid S. Miller } 2111c1b4a7e6SDavid S. Miller 2112c1b4a7e6SDavid S. Miller 2113a2a385d6SEric Dumazet /* Return true if the Nagle test allows this packet to be 2114c1b4a7e6SDavid S. Miller * sent now. 2115c1b4a7e6SDavid S. Miller */ 2116a2a385d6SEric Dumazet static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, 2117c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 2118c1b4a7e6SDavid S. Miller { 2119c1b4a7e6SDavid S. Miller /* Nagle rule does not apply to frames, which sit in the middle of the 2120c1b4a7e6SDavid S. Miller * write_queue (they have no chances to get new data). 2121c1b4a7e6SDavid S. Miller * 2122c1b4a7e6SDavid S. Miller * This is implemented in the callers, where they modify the 'nonagle' 2123c1b4a7e6SDavid S. Miller * argument based upon the location of SKB in the send queue. 2124c1b4a7e6SDavid S. Miller */ 2125c1b4a7e6SDavid S. Miller if (nonagle & TCP_NAGLE_PUSH) 2126a2a385d6SEric Dumazet return true; 2127c1b4a7e6SDavid S. Miller 21289b44190dSYuchung Cheng /* Don't use the nagle rule for urgent data (or for the final FIN). */ 21299b44190dSYuchung Cheng if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) 2130a2a385d6SEric Dumazet return true; 2131c1b4a7e6SDavid S. Miller 2132cc93fc51SPeter Pan(潘卫平) if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle)) 2133a2a385d6SEric Dumazet return true; 2134c1b4a7e6SDavid S. Miller 2135a2a385d6SEric Dumazet return false; 2136c1b4a7e6SDavid S. Miller } 2137c1b4a7e6SDavid S. Miller 2138c1b4a7e6SDavid S. Miller /* Does at least the first segment of SKB fit into the send window? */ 2139a2a385d6SEric Dumazet static bool tcp_snd_wnd_test(const struct tcp_sock *tp, 2140a2a385d6SEric Dumazet const struct sk_buff *skb, 2141056834d9SIlpo Järvinen unsigned int cur_mss) 2142c1b4a7e6SDavid S. Miller { 2143c1b4a7e6SDavid S. Miller u32 end_seq = TCP_SKB_CB(skb)->end_seq; 2144c1b4a7e6SDavid S. Miller 2145c1b4a7e6SDavid S. Miller if (skb->len > cur_mss) 2146c1b4a7e6SDavid S. Miller end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 2147c1b4a7e6SDavid S. Miller 214890840defSIlpo Järvinen return !after(end_seq, tcp_wnd_end(tp)); 2149c1b4a7e6SDavid S. Miller } 2150c1b4a7e6SDavid S. Miller 2151c1b4a7e6SDavid S. Miller /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 2152c1b4a7e6SDavid S. Miller * which is put after SKB on the list. It is very much like 2153c1b4a7e6SDavid S. Miller * tcp_fragment() except that it may make several kinds of assumptions 2154c1b4a7e6SDavid S. Miller * in order to speed up the splitting operation. In particular, we 2155c1b4a7e6SDavid S. Miller * know that all the data is in scatter-gather pages, and that the 2156c1b4a7e6SDavid S. Miller * packet has never been sent out before (and thus is not cloned). 2157c1b4a7e6SDavid S. Miller */ 215856483341SEric Dumazet static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, 2159c4ead4c5SEric Dumazet unsigned int mss_now, gfp_t gfp) 2160c1b4a7e6SDavid S. Miller { 2161c1b4a7e6SDavid S. Miller int nlen = skb->len - len; 216256483341SEric Dumazet struct sk_buff *buff; 21639ce01461SIlpo Järvinen u8 flags; 2164c1b4a7e6SDavid S. Miller 2165c1b4a7e6SDavid S. Miller /* All of a TSO frame must be composed of paged data. */ 2166b4a24397SEric Dumazet DEBUG_NET_WARN_ON_ONCE(skb->len != skb->data_len); 2167c1b4a7e6SDavid S. Miller 21685882efffSEric Dumazet buff = tcp_stream_alloc_skb(sk, gfp, true); 216951456b29SIan Morris if (unlikely(!buff)) 2170c1b4a7e6SDavid S. Miller return -ENOMEM; 217141477662SJakub Kicinski skb_copy_decrypted(buff, skb); 21725a369ca6SPaolo Abeni mptcp_skb_ext_copy(buff, skb); 2173c1b4a7e6SDavid S. Miller 2174ab4e846aSEric Dumazet sk_wmem_queued_add(sk, buff->truesize); 21753ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 2176b60b49eaSHerbert Xu buff->truesize += nlen; 2177c1b4a7e6SDavid S. Miller skb->truesize -= nlen; 2178c1b4a7e6SDavid S. Miller 2179c1b4a7e6SDavid S. Miller /* Correct the sequence numbers. */ 2180c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 2181c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 2182c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 2183c1b4a7e6SDavid S. Miller 2184c1b4a7e6SDavid S. Miller /* PSH and FIN should only be set in the second packet. */ 21854de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 21864de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 21874de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 2188c1b4a7e6SDavid S. Miller 2189a166140eSMartin KaFai Lau tcp_skb_fragment_eor(skb, buff); 2190a166140eSMartin KaFai Lau 2191c1b4a7e6SDavid S. Miller skb_split(skb, buff, len); 2192490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 2193c1b4a7e6SDavid S. Miller 2194c1b4a7e6SDavid S. Miller /* Fix up tso_factor for both original and new SKB. */ 21955bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 21965bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 2197c1b4a7e6SDavid S. Miller 2198c1b4a7e6SDavid S. Miller /* Link BUFF into the send queue. */ 2199f4a775d1SEric Dumazet __skb_header_release(buff); 220056483341SEric Dumazet tcp_insert_write_queue_after(skb, buff, sk, TCP_FRAG_IN_WRITE_QUEUE); 2201c1b4a7e6SDavid S. Miller 2202c1b4a7e6SDavid S. Miller return 0; 2203c1b4a7e6SDavid S. Miller } 2204c1b4a7e6SDavid S. Miller 2205c1b4a7e6SDavid S. Miller /* Try to defer sending, if possible, in order to minimize the amount 2206c1b4a7e6SDavid S. Miller * of TSO splitting we do. View it as a kind of TSO Nagle test. 2207c1b4a7e6SDavid S. Miller * 2208c1b4a7e6SDavid S. Miller * This algorithm is from John Heffner. 2209c1b4a7e6SDavid S. Miller */ 2210ca8a2263SNeal Cardwell static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb, 2211f9bfe4e6SEric Dumazet bool *is_cwnd_limited, 2212f9bfe4e6SEric Dumazet bool *is_rwnd_limited, 2213f9bfe4e6SEric Dumazet u32 max_segs) 2214c1b4a7e6SDavid S. Miller { 22156687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 2216f1c6ea38SEric Dumazet u32 send_win, cong_win, limit, in_flight; 221750c8339eSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 221850c8339eSEric Dumazet struct sk_buff *head; 2219ad9f4f50SEric Dumazet int win_divisor; 2220f1c6ea38SEric Dumazet s64 delta; 2221c1b4a7e6SDavid S. Miller 222299d7662aSEric Dumazet if (icsk->icsk_ca_state >= TCP_CA_Recovery) 2223ae8064acSJohn Heffner goto send_now; 2224ae8064acSJohn Heffner 22255f852eb5SEric Dumazet /* Avoid bursty behavior by allowing defer 2226a682850aSEric Dumazet * only if the last write was recent (1 ms). 2227a682850aSEric Dumazet * Note that tp->tcp_wstamp_ns can be in the future if we have 2228a682850aSEric Dumazet * packets waiting in a qdisc or device for EDT delivery. 22295f852eb5SEric Dumazet */ 2230a682850aSEric Dumazet delta = tp->tcp_clock_cache - tp->tcp_wstamp_ns - NSEC_PER_MSEC; 2231a682850aSEric Dumazet if (delta > 0) 2232ae8064acSJohn Heffner goto send_now; 2233908a75c1SDavid S. Miller 2234c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 2235c1b4a7e6SDavid S. Miller 2236c8c9aeb5SStefano Brivio BUG_ON(tcp_skb_pcount(skb) <= 1); 223740570375SEric Dumazet BUG_ON(tcp_snd_cwnd(tp) <= in_flight); 2238c1b4a7e6SDavid S. Miller 223990840defSIlpo Järvinen send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 2240c1b4a7e6SDavid S. Miller 2241c1b4a7e6SDavid S. Miller /* From in_flight test above, we know that cwnd > in_flight. */ 224240570375SEric Dumazet cong_win = (tcp_snd_cwnd(tp) - in_flight) * tp->mss_cache; 2243c1b4a7e6SDavid S. Miller 2244c1b4a7e6SDavid S. Miller limit = min(send_win, cong_win); 2245c1b4a7e6SDavid S. Miller 2246ba244fe9SDavid S. Miller /* If a full-sized TSO skb can be sent, do it. */ 2247605ad7f1SEric Dumazet if (limit >= max_segs * tp->mss_cache) 2248ae8064acSJohn Heffner goto send_now; 2249ba244fe9SDavid S. Miller 225062ad2761SIlpo Järvinen /* Middle in queue won't get any more data, full sendable already? */ 225162ad2761SIlpo Järvinen if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) 225262ad2761SIlpo Järvinen goto send_now; 225362ad2761SIlpo Järvinen 22545bbcc0f5SLinus Torvalds win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor); 2255ad9f4f50SEric Dumazet if (win_divisor) { 225640570375SEric Dumazet u32 chunk = min(tp->snd_wnd, tcp_snd_cwnd(tp) * tp->mss_cache); 2257c1b4a7e6SDavid S. Miller 2258c1b4a7e6SDavid S. Miller /* If at least some fraction of a window is available, 2259c1b4a7e6SDavid S. Miller * just use it. 2260c1b4a7e6SDavid S. Miller */ 2261ad9f4f50SEric Dumazet chunk /= win_divisor; 2262c1b4a7e6SDavid S. Miller if (limit >= chunk) 2263ae8064acSJohn Heffner goto send_now; 2264c1b4a7e6SDavid S. Miller } else { 2265c1b4a7e6SDavid S. Miller /* Different approach, try not to defer past a single 2266c1b4a7e6SDavid S. Miller * ACK. Receiver should ACK every other full sized 2267c1b4a7e6SDavid S. Miller * frame, so if we have space for more than 3 frames 2268c1b4a7e6SDavid S. Miller * then send now. 2269c1b4a7e6SDavid S. Miller */ 22706b5a5c0dSNeal Cardwell if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) 2271ae8064acSJohn Heffner goto send_now; 2272c1b4a7e6SDavid S. Miller } 2273c1b4a7e6SDavid S. Miller 227475c119afSEric Dumazet /* TODO : use tsorted_sent_queue ? */ 227575c119afSEric Dumazet head = tcp_rtx_queue_head(sk); 227675c119afSEric Dumazet if (!head) 227775c119afSEric Dumazet goto send_now; 2278f1c6ea38SEric Dumazet delta = tp->tcp_clock_cache - head->tstamp; 227950c8339eSEric Dumazet /* If next ACK is likely to come too late (half srtt), do not defer */ 2280f1c6ea38SEric Dumazet if ((s64)(delta - (u64)NSEC_PER_USEC * (tp->srtt_us >> 4)) < 0) 228150c8339eSEric Dumazet goto send_now; 228250c8339eSEric Dumazet 2283f9bfe4e6SEric Dumazet /* Ok, it looks like it is advisable to defer. 2284f9bfe4e6SEric Dumazet * Three cases are tracked : 2285f9bfe4e6SEric Dumazet * 1) We are cwnd-limited 2286f9bfe4e6SEric Dumazet * 2) We are rwnd-limited 2287f9bfe4e6SEric Dumazet * 3) We are application limited. 2288f9bfe4e6SEric Dumazet */ 2289f9bfe4e6SEric Dumazet if (cong_win < send_win) { 2290f9bfe4e6SEric Dumazet if (cong_win <= skb->len) { 2291ca8a2263SNeal Cardwell *is_cwnd_limited = true; 2292f9bfe4e6SEric Dumazet return true; 2293f9bfe4e6SEric Dumazet } 2294f9bfe4e6SEric Dumazet } else { 2295f9bfe4e6SEric Dumazet if (send_win <= skb->len) { 2296f9bfe4e6SEric Dumazet *is_rwnd_limited = true; 2297f9bfe4e6SEric Dumazet return true; 2298f9bfe4e6SEric Dumazet } 2299f9bfe4e6SEric Dumazet } 2300f9bfe4e6SEric Dumazet 2301f9bfe4e6SEric Dumazet /* If this packet won't get more data, do not wait. */ 2302d8ed257fSEric Dumazet if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) || 2303d8ed257fSEric Dumazet TCP_SKB_CB(skb)->eor) 2304f9bfe4e6SEric Dumazet goto send_now; 2305ca8a2263SNeal Cardwell 2306a2a385d6SEric Dumazet return true; 2307ae8064acSJohn Heffner 2308ae8064acSJohn Heffner send_now: 2309a2a385d6SEric Dumazet return false; 2310c1b4a7e6SDavid S. Miller } 2311c1b4a7e6SDavid S. Miller 231205cbc0dbSFan Du static inline void tcp_mtu_check_reprobe(struct sock *sk) 231305cbc0dbSFan Du { 231405cbc0dbSFan Du struct inet_connection_sock *icsk = inet_csk(sk); 231505cbc0dbSFan Du struct tcp_sock *tp = tcp_sk(sk); 231605cbc0dbSFan Du struct net *net = sock_net(sk); 231705cbc0dbSFan Du u32 interval; 231805cbc0dbSFan Du s32 delta; 231905cbc0dbSFan Du 23202a85388fSKuniyuki Iwashima interval = READ_ONCE(net->ipv4.sysctl_tcp_probe_interval); 2321c74df29aSEric Dumazet delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp; 232205cbc0dbSFan Du if (unlikely(delta >= interval * HZ)) { 232305cbc0dbSFan Du int mss = tcp_current_mss(sk); 232405cbc0dbSFan Du 232505cbc0dbSFan Du /* Update current search range */ 232605cbc0dbSFan Du icsk->icsk_mtup.probe_size = 0; 232705cbc0dbSFan Du icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + 232805cbc0dbSFan Du sizeof(struct tcphdr) + 232905cbc0dbSFan Du icsk->icsk_af_ops->net_header_len; 233005cbc0dbSFan Du icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss); 233105cbc0dbSFan Du 233205cbc0dbSFan Du /* Update probe time stamp */ 2333c74df29aSEric Dumazet icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; 233405cbc0dbSFan Du } 233505cbc0dbSFan Du } 233605cbc0dbSFan Du 2337808cf9e3SIlya Lesokhin static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len) 2338808cf9e3SIlya Lesokhin { 2339808cf9e3SIlya Lesokhin struct sk_buff *skb, *next; 2340808cf9e3SIlya Lesokhin 2341808cf9e3SIlya Lesokhin skb = tcp_send_head(sk); 2342808cf9e3SIlya Lesokhin tcp_for_write_queue_from_safe(skb, next, sk) { 2343808cf9e3SIlya Lesokhin if (len <= skb->len) 2344808cf9e3SIlya Lesokhin break; 2345808cf9e3SIlya Lesokhin 23469b65b17dSTalal Ahmad if (unlikely(TCP_SKB_CB(skb)->eor) || 23479b65b17dSTalal Ahmad tcp_has_tx_tstamp(skb) || 23489b65b17dSTalal Ahmad !skb_pure_zcopy_same(skb, next)) 2349808cf9e3SIlya Lesokhin return false; 2350808cf9e3SIlya Lesokhin 2351808cf9e3SIlya Lesokhin len -= skb->len; 2352808cf9e3SIlya Lesokhin } 2353808cf9e3SIlya Lesokhin 2354808cf9e3SIlya Lesokhin return true; 2355808cf9e3SIlya Lesokhin } 2356808cf9e3SIlya Lesokhin 235773601329SEric Dumazet static int tcp_clone_payload(struct sock *sk, struct sk_buff *to, 235873601329SEric Dumazet int probe_size) 235973601329SEric Dumazet { 236073601329SEric Dumazet skb_frag_t *lastfrag = NULL, *fragto = skb_shinfo(to)->frags; 236173601329SEric Dumazet int i, todo, len = 0, nr_frags = 0; 236273601329SEric Dumazet const struct sk_buff *skb; 236373601329SEric Dumazet 236473601329SEric Dumazet if (!sk_wmem_schedule(sk, to->truesize + probe_size)) 236573601329SEric Dumazet return -ENOMEM; 236673601329SEric Dumazet 236773601329SEric Dumazet skb_queue_walk(&sk->sk_write_queue, skb) { 236873601329SEric Dumazet const skb_frag_t *fragfrom = skb_shinfo(skb)->frags; 236973601329SEric Dumazet 237073601329SEric Dumazet if (skb_headlen(skb)) 237173601329SEric Dumazet return -EINVAL; 237273601329SEric Dumazet 237373601329SEric Dumazet for (i = 0; i < skb_shinfo(skb)->nr_frags; i++, fragfrom++) { 237473601329SEric Dumazet if (len >= probe_size) 237573601329SEric Dumazet goto commit; 237673601329SEric Dumazet todo = min_t(int, skb_frag_size(fragfrom), 237773601329SEric Dumazet probe_size - len); 237873601329SEric Dumazet len += todo; 237973601329SEric Dumazet if (lastfrag && 238073601329SEric Dumazet skb_frag_page(fragfrom) == skb_frag_page(lastfrag) && 238173601329SEric Dumazet skb_frag_off(fragfrom) == skb_frag_off(lastfrag) + 238273601329SEric Dumazet skb_frag_size(lastfrag)) { 238373601329SEric Dumazet skb_frag_size_add(lastfrag, todo); 238473601329SEric Dumazet continue; 238573601329SEric Dumazet } 238673601329SEric Dumazet if (unlikely(nr_frags == MAX_SKB_FRAGS)) 238773601329SEric Dumazet return -E2BIG; 238873601329SEric Dumazet skb_frag_page_copy(fragto, fragfrom); 238973601329SEric Dumazet skb_frag_off_copy(fragto, fragfrom); 239073601329SEric Dumazet skb_frag_size_set(fragto, todo); 239173601329SEric Dumazet nr_frags++; 239273601329SEric Dumazet lastfrag = fragto++; 239373601329SEric Dumazet } 239473601329SEric Dumazet } 239573601329SEric Dumazet commit: 239673601329SEric Dumazet WARN_ON_ONCE(len != probe_size); 239773601329SEric Dumazet for (i = 0; i < nr_frags; i++) 239873601329SEric Dumazet skb_frag_ref(to, i); 239973601329SEric Dumazet 240073601329SEric Dumazet skb_shinfo(to)->nr_frags = nr_frags; 240173601329SEric Dumazet to->truesize += probe_size; 240273601329SEric Dumazet to->len += probe_size; 240373601329SEric Dumazet to->data_len += probe_size; 240473601329SEric Dumazet __skb_header_release(to); 240573601329SEric Dumazet return 0; 240673601329SEric Dumazet } 240773601329SEric Dumazet 24085d424d5aSJohn Heffner /* Create a new MTU probe if we are ready. 240967edfef7SAndi Kleen * MTU probe is regularly attempting to increase the path MTU by 241067edfef7SAndi Kleen * deliberately sending larger packets. This discovers routing 241167edfef7SAndi Kleen * changes resulting in larger path MTUs. 241267edfef7SAndi Kleen * 24135d424d5aSJohn Heffner * Returns 0 if we should wait to probe (no cwnd available), 24145d424d5aSJohn Heffner * 1 if a probe was sent, 2415056834d9SIlpo Järvinen * -1 otherwise 2416056834d9SIlpo Järvinen */ 24175d424d5aSJohn Heffner static int tcp_mtu_probe(struct sock *sk) 24185d424d5aSJohn Heffner { 24195d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 242012a59abcSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 24215d424d5aSJohn Heffner struct sk_buff *skb, *nskb, *next; 24226b58e0a5SFan Du struct net *net = sock_net(sk); 24235d424d5aSJohn Heffner int probe_size; 242491cc17c0SIlpo Järvinen int size_needed; 242512a59abcSEric Dumazet int copy, len; 24265d424d5aSJohn Heffner int mss_now; 24276b58e0a5SFan Du int interval; 24285d424d5aSJohn Heffner 24295d424d5aSJohn Heffner /* Not currently probing/verifying, 24305d424d5aSJohn Heffner * not in recovery, 24315d424d5aSJohn Heffner * have enough cwnd, and 243212a59abcSEric Dumazet * not SACKing (the variable headers throw things off) 243312a59abcSEric Dumazet */ 243412a59abcSEric Dumazet if (likely(!icsk->icsk_mtup.enabled || 24355d424d5aSJohn Heffner icsk->icsk_mtup.probe_size || 24365d424d5aSJohn Heffner inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 243740570375SEric Dumazet tcp_snd_cwnd(tp) < 11 || 243812a59abcSEric Dumazet tp->rx_opt.num_sacks || tp->rx_opt.dsack)) 24395d424d5aSJohn Heffner return -1; 24405d424d5aSJohn Heffner 24416b58e0a5SFan Du /* Use binary search for probe_size between tcp_mss_base, 24426b58e0a5SFan Du * and current mss_clamp. if (search_high - search_low) 24436b58e0a5SFan Du * smaller than a threshold, backoff from probing. 24446b58e0a5SFan Du */ 24450c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 24466b58e0a5SFan Du probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high + 24476b58e0a5SFan Du icsk->icsk_mtup.search_low) >> 1); 244891cc17c0SIlpo Järvinen size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; 24496b58e0a5SFan Du interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low; 245005cbc0dbSFan Du /* When misfortune happens, we are reprobing actively, 245105cbc0dbSFan Du * and then reprobe timer has expired. We stick with current 245205cbc0dbSFan Du * probing process by not resetting search range to its orignal. 245305cbc0dbSFan Du */ 24546b58e0a5SFan Du if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) || 245592c0aa41SKuniyuki Iwashima interval < READ_ONCE(net->ipv4.sysctl_tcp_probe_threshold)) { 245605cbc0dbSFan Du /* Check whether enough time has elaplased for 245705cbc0dbSFan Du * another round of probing. 245805cbc0dbSFan Du */ 245905cbc0dbSFan Du tcp_mtu_check_reprobe(sk); 24605d424d5aSJohn Heffner return -1; 24615d424d5aSJohn Heffner } 24625d424d5aSJohn Heffner 24635d424d5aSJohn Heffner /* Have enough data in the send queue to probe? */ 24647f9c33e5SIlpo Järvinen if (tp->write_seq - tp->snd_nxt < size_needed) 24655d424d5aSJohn Heffner return -1; 24665d424d5aSJohn Heffner 246791cc17c0SIlpo Järvinen if (tp->snd_wnd < size_needed) 24685d424d5aSJohn Heffner return -1; 246990840defSIlpo Järvinen if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) 24705d424d5aSJohn Heffner return 0; 24715d424d5aSJohn Heffner 2472d67c58e9SIlpo Järvinen /* Do we need to wait to drain cwnd? With none in flight, don't stall */ 247340570375SEric Dumazet if (tcp_packets_in_flight(tp) + 2 > tcp_snd_cwnd(tp)) { 2474d67c58e9SIlpo Järvinen if (!tcp_packets_in_flight(tp)) 24755d424d5aSJohn Heffner return -1; 24765d424d5aSJohn Heffner else 24775d424d5aSJohn Heffner return 0; 24785d424d5aSJohn Heffner } 24795d424d5aSJohn Heffner 2480808cf9e3SIlya Lesokhin if (!tcp_can_coalesce_send_queue_head(sk, probe_size)) 2481808cf9e3SIlya Lesokhin return -1; 2482808cf9e3SIlya Lesokhin 24835d424d5aSJohn Heffner /* We're allowed to probe. Build it now. */ 24845882efffSEric Dumazet nskb = tcp_stream_alloc_skb(sk, GFP_ATOMIC, false); 248551456b29SIan Morris if (!nskb) 24865d424d5aSJohn Heffner return -1; 248773601329SEric Dumazet 248873601329SEric Dumazet /* build the payload, and be prepared to abort if this fails. */ 248973601329SEric Dumazet if (tcp_clone_payload(sk, nskb, probe_size)) { 249071c299c7SJakub Kicinski tcp_skb_tsorted_anchor_cleanup(nskb); 249173601329SEric Dumazet consume_skb(nskb); 249273601329SEric Dumazet return -1; 249373601329SEric Dumazet } 2494ab4e846aSEric Dumazet sk_wmem_queued_add(sk, nskb->truesize); 24953ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 24965d424d5aSJohn Heffner 2497fe067e8aSDavid S. Miller skb = tcp_send_head(sk); 249841477662SJakub Kicinski skb_copy_decrypted(nskb, skb); 24995a369ca6SPaolo Abeni mptcp_skb_ext_copy(nskb, skb); 25005d424d5aSJohn Heffner 25015d424d5aSJohn Heffner TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 25025d424d5aSJohn Heffner TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 25034de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; 25045d424d5aSJohn Heffner 250550c4817eSIlpo Järvinen tcp_insert_write_queue_before(nskb, skb, sk); 25062b7cda9cSEric Dumazet tcp_highest_sack_replace(sk, skb, nskb); 250750c4817eSIlpo Järvinen 25085d424d5aSJohn Heffner len = 0; 2509234b6860SIlpo Järvinen tcp_for_write_queue_from_safe(skb, next, sk) { 25105d424d5aSJohn Heffner copy = min_t(int, skb->len, probe_size - len); 25115d424d5aSJohn Heffner 25125d424d5aSJohn Heffner if (skb->len <= copy) { 25135d424d5aSJohn Heffner /* We've eaten all the data from this skb. 25145d424d5aSJohn Heffner * Throw it away. */ 25154de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 2516808cf9e3SIlya Lesokhin /* If this is the last SKB we copy and eor is set 2517808cf9e3SIlya Lesokhin * we need to propagate it to the new skb. 2518808cf9e3SIlya Lesokhin */ 2519808cf9e3SIlya Lesokhin TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor; 2520888a5c53SWillem de Bruijn tcp_skb_collapse_tstamp(nskb, skb); 2521fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 252203271f3aSTalal Ahmad tcp_wmem_free_skb(sk, skb); 25235d424d5aSJohn Heffner } else { 25244de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & 2525a3433f35SChangli Gao ~(TCPHDR_FIN|TCPHDR_PSH); 25265d424d5aSJohn Heffner __pskb_trim_head(skb, copy); 25275bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 25285d424d5aSJohn Heffner TCP_SKB_CB(skb)->seq += copy; 25295d424d5aSJohn Heffner } 25305d424d5aSJohn Heffner 25315d424d5aSJohn Heffner len += copy; 2532234b6860SIlpo Järvinen 2533234b6860SIlpo Järvinen if (len >= probe_size) 2534234b6860SIlpo Järvinen break; 25355d424d5aSJohn Heffner } 25365bbb432cSEric Dumazet tcp_init_tso_segs(nskb, nskb->len); 25375d424d5aSJohn Heffner 25385d424d5aSJohn Heffner /* We're ready to send. If this fails, the probe will 25397faee5c0SEric Dumazet * be resegmented into mss-sized pieces by tcp_write_xmit(). 25407faee5c0SEric Dumazet */ 25415d424d5aSJohn Heffner if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 25425d424d5aSJohn Heffner /* Decrement cwnd here because we are sending 25435d424d5aSJohn Heffner * effectively two packets. */ 254440570375SEric Dumazet tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) - 1); 254566f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, nskb); 25465d424d5aSJohn Heffner 25475d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 25480e7b1368SJohn Heffner tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 25490e7b1368SJohn Heffner tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 25505d424d5aSJohn Heffner 25515d424d5aSJohn Heffner return 1; 25525d424d5aSJohn Heffner } 25535d424d5aSJohn Heffner 25545d424d5aSJohn Heffner return -1; 25555d424d5aSJohn Heffner } 25565d424d5aSJohn Heffner 2557864e5c09SEric Dumazet static bool tcp_pacing_check(struct sock *sk) 2558218af599SEric Dumazet { 2559864e5c09SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 2560864e5c09SEric Dumazet 2561864e5c09SEric Dumazet if (!tcp_needs_internal_pacing(sk)) 2562864e5c09SEric Dumazet return false; 2563864e5c09SEric Dumazet 2564864e5c09SEric Dumazet if (tp->tcp_wstamp_ns <= tp->tcp_clock_cache) 2565864e5c09SEric Dumazet return false; 2566864e5c09SEric Dumazet 2567864e5c09SEric Dumazet if (!hrtimer_is_queued(&tp->pacing_timer)) { 2568864e5c09SEric Dumazet hrtimer_start(&tp->pacing_timer, 2569864e5c09SEric Dumazet ns_to_ktime(tp->tcp_wstamp_ns), 2570864e5c09SEric Dumazet HRTIMER_MODE_ABS_PINNED_SOFT); 2571864e5c09SEric Dumazet sock_hold(sk); 2572864e5c09SEric Dumazet } 2573864e5c09SEric Dumazet return true; 2574218af599SEric Dumazet } 2575218af599SEric Dumazet 2576f921a4a5SEric Dumazet static bool tcp_rtx_queue_empty_or_single_skb(const struct sock *sk) 2577f921a4a5SEric Dumazet { 2578f921a4a5SEric Dumazet const struct rb_node *node = sk->tcp_rtx_queue.rb_node; 2579f921a4a5SEric Dumazet 2580f921a4a5SEric Dumazet /* No skb in the rtx queue. */ 2581f921a4a5SEric Dumazet if (!node) 2582f921a4a5SEric Dumazet return true; 2583f921a4a5SEric Dumazet 2584f921a4a5SEric Dumazet /* Only one skb in rtx queue. */ 2585f921a4a5SEric Dumazet return !node->rb_left && !node->rb_right; 2586f921a4a5SEric Dumazet } 2587f921a4a5SEric Dumazet 2588f9616c35SEric Dumazet /* TCP Small Queues : 2589f9616c35SEric Dumazet * Control number of packets in qdisc/devices to two packets / or ~1 ms. 2590f9616c35SEric Dumazet * (These limits are doubled for retransmits) 2591f9616c35SEric Dumazet * This allows for : 2592f9616c35SEric Dumazet * - better RTT estimation and ACK scheduling 2593f9616c35SEric Dumazet * - faster recovery 2594f9616c35SEric Dumazet * - high rates 2595f9616c35SEric Dumazet * Alas, some drivers / subsystems require a fair amount 2596f9616c35SEric Dumazet * of queued bytes to ensure line rate. 2597f9616c35SEric Dumazet * One example is wifi aggregation (802.11 AMPDU) 2598f9616c35SEric Dumazet */ 2599f9616c35SEric Dumazet static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb, 2600f9616c35SEric Dumazet unsigned int factor) 2601f9616c35SEric Dumazet { 260276a9ebe8SEric Dumazet unsigned long limit; 2603f9616c35SEric Dumazet 260476a9ebe8SEric Dumazet limit = max_t(unsigned long, 260576a9ebe8SEric Dumazet 2 * skb->truesize, 260628b24f90SEric Dumazet READ_ONCE(sk->sk_pacing_rate) >> READ_ONCE(sk->sk_pacing_shift)); 2607c73e5807SEric Dumazet if (sk->sk_pacing_status == SK_PACING_NONE) 260876a9ebe8SEric Dumazet limit = min_t(unsigned long, limit, 26099fb90193SKuniyuki Iwashima READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes)); 2610f9616c35SEric Dumazet limit <<= factor; 2611f9616c35SEric Dumazet 2612a842fe14SEric Dumazet if (static_branch_unlikely(&tcp_tx_delay_enabled) && 2613a842fe14SEric Dumazet tcp_sk(sk)->tcp_tx_delay) { 261428b24f90SEric Dumazet u64 extra_bytes = (u64)READ_ONCE(sk->sk_pacing_rate) * 261528b24f90SEric Dumazet tcp_sk(sk)->tcp_tx_delay; 2616a842fe14SEric Dumazet 2617a842fe14SEric Dumazet /* TSQ is based on skb truesize sum (sk_wmem_alloc), so we 2618a842fe14SEric Dumazet * approximate our needs assuming an ~100% skb->truesize overhead. 2619a842fe14SEric Dumazet * USEC_PER_SEC is approximated by 2^20. 2620a842fe14SEric Dumazet * do_div(extra_bytes, USEC_PER_SEC/2) is replaced by a right shift. 2621a842fe14SEric Dumazet */ 2622a842fe14SEric Dumazet extra_bytes >>= (20 - 1); 2623a842fe14SEric Dumazet limit += extra_bytes; 2624a842fe14SEric Dumazet } 262514afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) > limit) { 2626f921a4a5SEric Dumazet /* Always send skb if rtx queue is empty or has one skb. 262775eefc6cSEric Dumazet * No need to wait for TX completion to call us back, 262875eefc6cSEric Dumazet * after softirq/tasklet schedule. 262975eefc6cSEric Dumazet * This helps when TX completions are delayed too much. 263075eefc6cSEric Dumazet */ 2631f921a4a5SEric Dumazet if (tcp_rtx_queue_empty_or_single_skb(sk)) 263275eefc6cSEric Dumazet return false; 263375eefc6cSEric Dumazet 26347aa5470cSEric Dumazet set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags); 2635f9616c35SEric Dumazet /* It is possible TX completion already happened 2636f9616c35SEric Dumazet * before we set TSQ_THROTTLED, so we must 2637f9616c35SEric Dumazet * test again the condition. 2638f9616c35SEric Dumazet */ 2639f9616c35SEric Dumazet smp_mb__after_atomic(); 2640ce8299b6SEric Dumazet if (refcount_read(&sk->sk_wmem_alloc) > limit) 2641f9616c35SEric Dumazet return true; 2642f9616c35SEric Dumazet } 2643f9616c35SEric Dumazet return false; 2644f9616c35SEric Dumazet } 2645f9616c35SEric Dumazet 264605b055e8SFrancis Yan static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new) 264705b055e8SFrancis Yan { 2648628174ccSEric Dumazet const u32 now = tcp_jiffies32; 2649efe967cdSArnd Bergmann enum tcp_chrono old = tp->chrono_type; 265005b055e8SFrancis Yan 2651efe967cdSArnd Bergmann if (old > TCP_CHRONO_UNSPEC) 2652efe967cdSArnd Bergmann tp->chrono_stat[old - 1] += now - tp->chrono_start; 265305b055e8SFrancis Yan tp->chrono_start = now; 265405b055e8SFrancis Yan tp->chrono_type = new; 265505b055e8SFrancis Yan } 265605b055e8SFrancis Yan 265705b055e8SFrancis Yan void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type) 265805b055e8SFrancis Yan { 265905b055e8SFrancis Yan struct tcp_sock *tp = tcp_sk(sk); 266005b055e8SFrancis Yan 266105b055e8SFrancis Yan /* If there are multiple conditions worthy of tracking in a 26620f87230dSFrancis Yan * chronograph then the highest priority enum takes precedence 26630f87230dSFrancis Yan * over the other conditions. So that if something "more interesting" 266405b055e8SFrancis Yan * starts happening, stop the previous chrono and start a new one. 266505b055e8SFrancis Yan */ 266605b055e8SFrancis Yan if (type > tp->chrono_type) 266705b055e8SFrancis Yan tcp_chrono_set(tp, type); 266805b055e8SFrancis Yan } 266905b055e8SFrancis Yan 267005b055e8SFrancis Yan void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type) 267105b055e8SFrancis Yan { 267205b055e8SFrancis Yan struct tcp_sock *tp = tcp_sk(sk); 267305b055e8SFrancis Yan 26740f87230dSFrancis Yan 26750f87230dSFrancis Yan /* There are multiple conditions worthy of tracking in a 26760f87230dSFrancis Yan * chronograph, so that the highest priority enum takes 26770f87230dSFrancis Yan * precedence over the other conditions (see tcp_chrono_start). 26780f87230dSFrancis Yan * If a condition stops, we only stop chrono tracking if 26790f87230dSFrancis Yan * it's the "most interesting" or current chrono we are 26800f87230dSFrancis Yan * tracking and starts busy chrono if we have pending data. 26810f87230dSFrancis Yan */ 268275c119afSEric Dumazet if (tcp_rtx_and_write_queues_empty(sk)) 268305b055e8SFrancis Yan tcp_chrono_set(tp, TCP_CHRONO_UNSPEC); 26840f87230dSFrancis Yan else if (type == tp->chrono_type) 26850f87230dSFrancis Yan tcp_chrono_set(tp, TCP_CHRONO_BUSY); 268605b055e8SFrancis Yan } 268705b055e8SFrancis Yan 26881da177e4SLinus Torvalds /* This routine writes packets to the network. It advances the 26891da177e4SLinus Torvalds * send_head. This happens as incoming acks open up the remote 26901da177e4SLinus Torvalds * window for us. 26911da177e4SLinus Torvalds * 2692f8269a49SIlpo Järvinen * LARGESEND note: !tcp_urg_mode is overkill, only frames between 2693f8269a49SIlpo Järvinen * snd_up-64k-mss .. snd_up cannot be large. However, taking into 2694f8269a49SIlpo Järvinen * account rare use of URG, this is not a big flaw. 2695f8269a49SIlpo Järvinen * 26966ba8a3b1SNandita Dukkipati * Send at most one packet when push_one > 0. Temporarily ignore 26976ba8a3b1SNandita Dukkipati * cwnd limit to force at most one packet out when push_one == 2. 26986ba8a3b1SNandita Dukkipati 2699a2a385d6SEric Dumazet * Returns true, if no segments are in flight and we have queued segments, 2700a2a385d6SEric Dumazet * but cannot send anything now because of SWS or another problem. 27011da177e4SLinus Torvalds */ 2702a2a385d6SEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 2703d5dd9175SIlpo Järvinen int push_one, gfp_t gfp) 27041da177e4SLinus Torvalds { 27051da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 270692df7b51SDavid S. Miller struct sk_buff *skb; 2707c1b4a7e6SDavid S. Miller unsigned int tso_segs, sent_pkts; 2708c1b4a7e6SDavid S. Miller int cwnd_quota; 27095d424d5aSJohn Heffner int result; 27105615f886SFrancis Yan bool is_cwnd_limited = false, is_rwnd_limited = false; 2711605ad7f1SEric Dumazet u32 max_segs; 27121da177e4SLinus Torvalds 2713c1b4a7e6SDavid S. Miller sent_pkts = 0; 27145d424d5aSJohn Heffner 2715ee1836aeSEric Dumazet tcp_mstamp_refresh(tp); 2716d5dd9175SIlpo Järvinen if (!push_one) { 27175d424d5aSJohn Heffner /* Do MTU probing. */ 2718d5dd9175SIlpo Järvinen result = tcp_mtu_probe(sk); 2719d5dd9175SIlpo Järvinen if (!result) { 2720a2a385d6SEric Dumazet return false; 27215d424d5aSJohn Heffner } else if (result > 0) { 27225d424d5aSJohn Heffner sent_pkts = 1; 27235d424d5aSJohn Heffner } 2724d5dd9175SIlpo Järvinen } 27255d424d5aSJohn Heffner 2726ed6e7268SNeal Cardwell max_segs = tcp_tso_segs(sk, mss_now); 2727fe067e8aSDavid S. Miller while ((skb = tcp_send_head(sk))) { 2728c8ac3774SHerbert Xu unsigned int limit; 2729c8ac3774SHerbert Xu 273079861919SEric Dumazet if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) { 273179861919SEric Dumazet /* "skb_mstamp_ns" is used as a start point for the retransmit timer */ 2732a1ac9c8aSMartin KaFai Lau tp->tcp_wstamp_ns = tp->tcp_clock_cache; 2733a1ac9c8aSMartin KaFai Lau skb_set_delivery_time(skb, tp->tcp_wstamp_ns, true); 273479861919SEric Dumazet list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue); 2735bf50b606SEric Dumazet tcp_init_tso_segs(skb, mss_now); 273679861919SEric Dumazet goto repair; /* Skip network transmission */ 273779861919SEric Dumazet } 273879861919SEric Dumazet 2739218af599SEric Dumazet if (tcp_pacing_check(sk)) 2740218af599SEric Dumazet break; 2741218af599SEric Dumazet 27425bbb432cSEric Dumazet tso_segs = tcp_init_tso_segs(skb, mss_now); 2743c1b4a7e6SDavid S. Miller BUG_ON(!tso_segs); 2744c1b4a7e6SDavid S. Miller 2745b68e9f85SHerbert Xu cwnd_quota = tcp_cwnd_test(tp, skb); 27466ba8a3b1SNandita Dukkipati if (!cwnd_quota) { 27476ba8a3b1SNandita Dukkipati if (push_one == 2) 27486ba8a3b1SNandita Dukkipati /* Force out a loss probe pkt. */ 27496ba8a3b1SNandita Dukkipati cwnd_quota = 1; 27506ba8a3b1SNandita Dukkipati else 2751b68e9f85SHerbert Xu break; 27526ba8a3b1SNandita Dukkipati } 2753b68e9f85SHerbert Xu 27545615f886SFrancis Yan if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) { 27555615f886SFrancis Yan is_rwnd_limited = true; 2756b68e9f85SHerbert Xu break; 27575615f886SFrancis Yan } 2758b68e9f85SHerbert Xu 2759d6a4e26aSEric Dumazet if (tso_segs == 1) { 2760aa93466bSDavid S. Miller if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 2761aa93466bSDavid S. Miller (tcp_skb_is_last(sk, skb) ? 2762aa93466bSDavid S. Miller nonagle : TCP_NAGLE_PUSH)))) 2763aa93466bSDavid S. Miller break; 2764c1b4a7e6SDavid S. Miller } else { 2765ca8a2263SNeal Cardwell if (!push_one && 2766605ad7f1SEric Dumazet tcp_tso_should_defer(sk, skb, &is_cwnd_limited, 2767f9bfe4e6SEric Dumazet &is_rwnd_limited, max_segs)) 2768aa93466bSDavid S. Miller break; 2769c1b4a7e6SDavid S. Miller } 2770aa93466bSDavid S. Miller 2771605ad7f1SEric Dumazet limit = mss_now; 2772d6a4e26aSEric Dumazet if (tso_segs > 1 && !tcp_urg_mode(tp)) 2773605ad7f1SEric Dumazet limit = tcp_mss_split_point(sk, skb, mss_now, 2774605ad7f1SEric Dumazet min_t(unsigned int, 2775605ad7f1SEric Dumazet cwnd_quota, 2776605ad7f1SEric Dumazet max_segs), 2777605ad7f1SEric Dumazet nonagle); 2778605ad7f1SEric Dumazet 2779605ad7f1SEric Dumazet if (skb->len > limit && 278056483341SEric Dumazet unlikely(tso_fragment(sk, skb, limit, mss_now, gfp))) 2781605ad7f1SEric Dumazet break; 2782605ad7f1SEric Dumazet 2783f9616c35SEric Dumazet if (tcp_small_queue_check(sk, skb, 0)) 278446d3ceabSEric Dumazet break; 2785c9eeec26SEric Dumazet 27861f85e626SEric Dumazet /* Argh, we hit an empty skb(), presumably a thread 27871f85e626SEric Dumazet * is sleeping in sendmsg()/sk_stream_wait_memory(). 27881f85e626SEric Dumazet * We do not want to send a pure-ack packet and have 27891f85e626SEric Dumazet * a strange looking rtx queue with empty packet(s). 27901f85e626SEric Dumazet */ 27911f85e626SEric Dumazet if (TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq) 27921f85e626SEric Dumazet break; 27931f85e626SEric Dumazet 2794d5dd9175SIlpo Järvinen if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) 27951da177e4SLinus Torvalds break; 27961da177e4SLinus Torvalds 2797ec342325SAndrew Vagin repair: 27981da177e4SLinus Torvalds /* Advance the send_head. This one is sent out. 27991da177e4SLinus Torvalds * This call will increment packets_out. 28001da177e4SLinus Torvalds */ 280166f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 28021da177e4SLinus Torvalds 28031da177e4SLinus Torvalds tcp_minshall_update(tp, mss_now, skb); 2804a262f0cdSNandita Dukkipati sent_pkts += tcp_skb_pcount(skb); 2805d5dd9175SIlpo Järvinen 2806d5dd9175SIlpo Järvinen if (push_one) 2807d5dd9175SIlpo Järvinen break; 28081da177e4SLinus Torvalds } 28091da177e4SLinus Torvalds 28105615f886SFrancis Yan if (is_rwnd_limited) 28115615f886SFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED); 28125615f886SFrancis Yan else 28135615f886SFrancis Yan tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED); 28145615f886SFrancis Yan 281540570375SEric Dumazet is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tcp_snd_cwnd(tp)); 2816299bcb55SNeal Cardwell if (likely(sent_pkts || is_cwnd_limited)) 2817299bcb55SNeal Cardwell tcp_cwnd_validate(sk, is_cwnd_limited); 2818299bcb55SNeal Cardwell 2819aa93466bSDavid S. Miller if (likely(sent_pkts)) { 2820684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 2821684bad11SYuchung Cheng tp->prr_out += sent_pkts; 28226ba8a3b1SNandita Dukkipati 28236ba8a3b1SNandita Dukkipati /* Send one loss probe per tail loss episode. */ 28246ba8a3b1SNandita Dukkipati if (push_one != 2) 2825ed66dfafSNeal Cardwell tcp_schedule_loss_probe(sk, false); 2826a2a385d6SEric Dumazet return false; 28271da177e4SLinus Torvalds } 282875c119afSEric Dumazet return !tp->packets_out && !tcp_write_queue_empty(sk); 28296ba8a3b1SNandita Dukkipati } 28306ba8a3b1SNandita Dukkipati 2831ed66dfafSNeal Cardwell bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto) 28326ba8a3b1SNandita Dukkipati { 28336ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 28346ba8a3b1SNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 28351c2709cfSNeal Cardwell u32 timeout, timeout_us, rto_delta_us; 28362ae21cf5SEric Dumazet int early_retrans; 28376ba8a3b1SNandita Dukkipati 28386ba8a3b1SNandita Dukkipati /* Don't do any loss probe on a Fast Open connection before 3WHS 28396ba8a3b1SNandita Dukkipati * finishes. 28406ba8a3b1SNandita Dukkipati */ 2841d983ea6fSEric Dumazet if (rcu_access_pointer(tp->fastopen_rsk)) 28426ba8a3b1SNandita Dukkipati return false; 28436ba8a3b1SNandita Dukkipati 284452e65865SKuniyuki Iwashima early_retrans = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_early_retrans); 28456ba8a3b1SNandita Dukkipati /* Schedule a loss probe in 2*RTT for SACK capable connections 2846b4f70c3dSNeal Cardwell * not in loss recovery, that are either limited by cwnd or application. 28476ba8a3b1SNandita Dukkipati */ 28482ae21cf5SEric Dumazet if ((early_retrans != 3 && early_retrans != 4) || 2849bec41a11SYuchung Cheng !tp->packets_out || !tcp_is_sack(tp) || 2850b4f70c3dSNeal Cardwell (icsk->icsk_ca_state != TCP_CA_Open && 2851b4f70c3dSNeal Cardwell icsk->icsk_ca_state != TCP_CA_CWR)) 28526ba8a3b1SNandita Dukkipati return false; 28536ba8a3b1SNandita Dukkipati 2854bb4d991aSYuchung Cheng /* Probe timeout is 2*rtt. Add minimum RTO to account 2855f9b99582SYuchung Cheng * for delayed ack when there's one outstanding packet. If no RTT 2856f9b99582SYuchung Cheng * sample is available then probe after TCP_TIMEOUT_INIT. 28576ba8a3b1SNandita Dukkipati */ 2858bb4d991aSYuchung Cheng if (tp->srtt_us) { 28591c2709cfSNeal Cardwell timeout_us = tp->srtt_us >> 2; 28606ba8a3b1SNandita Dukkipati if (tp->packets_out == 1) 28611c2709cfSNeal Cardwell timeout_us += tcp_rto_min_us(sk); 2862bb4d991aSYuchung Cheng else 28631c2709cfSNeal Cardwell timeout_us += TCP_TIMEOUT_MIN_US; 28641c2709cfSNeal Cardwell timeout = usecs_to_jiffies(timeout_us); 2865bb4d991aSYuchung Cheng } else { 2866bb4d991aSYuchung Cheng timeout = TCP_TIMEOUT_INIT; 2867bb4d991aSYuchung Cheng } 28686ba8a3b1SNandita Dukkipati 2869a2815817SNeal Cardwell /* If the RTO formula yields an earlier time, then use that time. */ 2870ed66dfafSNeal Cardwell rto_delta_us = advancing_rto ? 2871ed66dfafSNeal Cardwell jiffies_to_usecs(inet_csk(sk)->icsk_rto) : 2872ed66dfafSNeal Cardwell tcp_rto_delta_us(sk); /* How far in future is RTO? */ 2873a2815817SNeal Cardwell if (rto_delta_us > 0) 2874a2815817SNeal Cardwell timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us)); 28756ba8a3b1SNandita Dukkipati 28768dc242adSEric Dumazet tcp_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, TCP_RTO_MAX); 28776ba8a3b1SNandita Dukkipati return true; 28786ba8a3b1SNandita Dukkipati } 28796ba8a3b1SNandita Dukkipati 28801f3279aeSEric Dumazet /* Thanks to skb fast clones, we can detect if a prior transmit of 28811f3279aeSEric Dumazet * a packet is still in a qdisc or driver queue. 28821f3279aeSEric Dumazet * In this case, there is very little point doing a retransmit ! 28831f3279aeSEric Dumazet */ 2884f4dae54eSEric Dumazet static bool skb_still_in_host_queue(struct sock *sk, 28851f3279aeSEric Dumazet const struct sk_buff *skb) 28861f3279aeSEric Dumazet { 288739bb5e62SEric Dumazet if (unlikely(skb_fclone_busy(sk, skb))) { 2888f4dae54eSEric Dumazet set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags); 2889f4dae54eSEric Dumazet smp_mb__after_atomic(); 2890f4dae54eSEric Dumazet if (skb_fclone_busy(sk, skb)) { 2891c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), 28921f3279aeSEric Dumazet LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES); 28931f3279aeSEric Dumazet return true; 28941f3279aeSEric Dumazet } 2895f4dae54eSEric Dumazet } 28961f3279aeSEric Dumazet return false; 28971f3279aeSEric Dumazet } 28981f3279aeSEric Dumazet 2899b340b264SYuchung Cheng /* When probe timeout (PTO) fires, try send a new segment if possible, else 29006ba8a3b1SNandita Dukkipati * retransmit the last segment. 29016ba8a3b1SNandita Dukkipati */ 29026ba8a3b1SNandita Dukkipati void tcp_send_loss_probe(struct sock *sk) 29036ba8a3b1SNandita Dukkipati { 29049b717a8dSNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 29056ba8a3b1SNandita Dukkipati struct sk_buff *skb; 29066ba8a3b1SNandita Dukkipati int pcount; 29076ba8a3b1SNandita Dukkipati int mss = tcp_current_mss(sk); 29086ba8a3b1SNandita Dukkipati 290976be93fcSYuchung Cheng /* At most one outstanding TLP */ 291076be93fcSYuchung Cheng if (tp->tlp_high_seq) 291176be93fcSYuchung Cheng goto rearm_timer; 291276be93fcSYuchung Cheng 291376be93fcSYuchung Cheng tp->tlp_retrans = 0; 2914b340b264SYuchung Cheng skb = tcp_send_head(sk); 291575c119afSEric Dumazet if (skb && tcp_snd_wnd_test(tp, skb, mss)) { 2916b340b264SYuchung Cheng pcount = tp->packets_out; 2917b340b264SYuchung Cheng tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC); 2918b340b264SYuchung Cheng if (tp->packets_out > pcount) 2919b340b264SYuchung Cheng goto probe_sent; 29206ba8a3b1SNandita Dukkipati goto rearm_timer; 29216ba8a3b1SNandita Dukkipati } 292275c119afSEric Dumazet skb = skb_rb_last(&sk->tcp_rtx_queue); 2923b2b7af86SYuchung Cheng if (unlikely(!skb)) { 2924b2b7af86SYuchung Cheng WARN_ONCE(tp->packets_out, 2925b2b7af86SYuchung Cheng "invalid inflight: %u state %u cwnd %u mss %d\n", 292640570375SEric Dumazet tp->packets_out, sk->sk_state, tcp_snd_cwnd(tp), mss); 2927b2b7af86SYuchung Cheng inet_csk(sk)->icsk_pending = 0; 2928b2b7af86SYuchung Cheng return; 2929b2b7af86SYuchung Cheng } 29306ba8a3b1SNandita Dukkipati 29311f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 29321f3279aeSEric Dumazet goto rearm_timer; 29331f3279aeSEric Dumazet 29346ba8a3b1SNandita Dukkipati pcount = tcp_skb_pcount(skb); 29356ba8a3b1SNandita Dukkipati if (WARN_ON(!pcount)) 29366ba8a3b1SNandita Dukkipati goto rearm_timer; 29376ba8a3b1SNandita Dukkipati 29386ba8a3b1SNandita Dukkipati if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) { 293975c119afSEric Dumazet if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, 294075c119afSEric Dumazet (pcount - 1) * mss, mss, 29416cc55e09SOctavian Purdila GFP_ATOMIC))) 29426ba8a3b1SNandita Dukkipati goto rearm_timer; 294375c119afSEric Dumazet skb = skb_rb_next(skb); 29446ba8a3b1SNandita Dukkipati } 29456ba8a3b1SNandita Dukkipati 29466ba8a3b1SNandita Dukkipati if (WARN_ON(!skb || !tcp_skb_pcount(skb))) 29476ba8a3b1SNandita Dukkipati goto rearm_timer; 29486ba8a3b1SNandita Dukkipati 294910d3be56SEric Dumazet if (__tcp_retransmit_skb(sk, skb, 1)) 2950b340b264SYuchung Cheng goto rearm_timer; 29516ba8a3b1SNandita Dukkipati 295276be93fcSYuchung Cheng tp->tlp_retrans = 1; 295376be93fcSYuchung Cheng 295476be93fcSYuchung Cheng probe_sent: 29559b717a8dSNandita Dukkipati /* Record snd_nxt for loss detection. */ 29569b717a8dSNandita Dukkipati tp->tlp_high_seq = tp->snd_nxt; 29579b717a8dSNandita Dukkipati 2958c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES); 2959fcd16c0aSYuchung Cheng /* Reset s.t. tcp_rearm_rto will restart timer from now */ 2960fcd16c0aSYuchung Cheng inet_csk(sk)->icsk_pending = 0; 2961b340b264SYuchung Cheng rearm_timer: 2962fcd16c0aSYuchung Cheng tcp_rearm_rto(sk); 29631da177e4SLinus Torvalds } 29641da177e4SLinus Torvalds 2965a762a980SDavid S. Miller /* Push out any pending frames which were held back due to 2966a762a980SDavid S. Miller * TCP_CORK or attempt at coalescing tiny packets. 2967a762a980SDavid S. Miller * The socket must be locked by the caller. 2968a762a980SDavid S. Miller */ 29699e412ba7SIlpo Järvinen void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 29709e412ba7SIlpo Järvinen int nonagle) 2971a762a980SDavid S. Miller { 2972726e07a8SIlpo Järvinen /* If we are closed, the bytes will have to remain here. 2973726e07a8SIlpo Järvinen * In time closedown will finish, we empty the write queue and 2974726e07a8SIlpo Järvinen * all will be happy. 2975726e07a8SIlpo Järvinen */ 2976726e07a8SIlpo Järvinen if (unlikely(sk->sk_state == TCP_CLOSE)) 2977726e07a8SIlpo Järvinen return; 2978726e07a8SIlpo Järvinen 297999a1dec7SMel Gorman if (tcp_write_xmit(sk, cur_mss, nonagle, 0, 29807450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC))) 29819e412ba7SIlpo Järvinen tcp_check_probe_timer(sk); 2982a762a980SDavid S. Miller } 2983a762a980SDavid S. Miller 2984c1b4a7e6SDavid S. Miller /* Send _single_ skb sitting at the send head. This function requires 2985c1b4a7e6SDavid S. Miller * true push pending frames to setup probe timer etc. 2986c1b4a7e6SDavid S. Miller */ 2987c1b4a7e6SDavid S. Miller void tcp_push_one(struct sock *sk, unsigned int mss_now) 2988c1b4a7e6SDavid S. Miller { 2989fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 2990c1b4a7e6SDavid S. Miller 2991c1b4a7e6SDavid S. Miller BUG_ON(!skb || skb->len < mss_now); 2992c1b4a7e6SDavid S. Miller 2993d5dd9175SIlpo Järvinen tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); 2994c1b4a7e6SDavid S. Miller } 2995c1b4a7e6SDavid S. Miller 29961da177e4SLinus Torvalds /* This function returns the amount that we can raise the 29971da177e4SLinus Torvalds * usable window based on the following constraints 29981da177e4SLinus Torvalds * 29991da177e4SLinus Torvalds * 1. The window can never be shrunk once it is offered (RFC 793) 30001da177e4SLinus Torvalds * 2. We limit memory per socket 30011da177e4SLinus Torvalds * 30021da177e4SLinus Torvalds * RFC 1122: 30031da177e4SLinus Torvalds * "the suggested [SWS] avoidance algorithm for the receiver is to keep 30041da177e4SLinus Torvalds * RECV.NEXT + RCV.WIN fixed until: 30051da177e4SLinus Torvalds * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 30061da177e4SLinus Torvalds * 30071da177e4SLinus Torvalds * i.e. don't raise the right edge of the window until you can raise 30081da177e4SLinus Torvalds * it at least MSS bytes. 30091da177e4SLinus Torvalds * 30101da177e4SLinus Torvalds * Unfortunately, the recommended algorithm breaks header prediction, 30111da177e4SLinus Torvalds * since header prediction assumes th->window stays fixed. 30121da177e4SLinus Torvalds * 30131da177e4SLinus Torvalds * Strictly speaking, keeping th->window fixed violates the receiver 30141da177e4SLinus Torvalds * side SWS prevention criteria. The problem is that under this rule 30151da177e4SLinus Torvalds * a stream of single byte packets will cause the right side of the 30161da177e4SLinus Torvalds * window to always advance by a single byte. 30171da177e4SLinus Torvalds * 30181da177e4SLinus Torvalds * Of course, if the sender implements sender side SWS prevention 30191da177e4SLinus Torvalds * then this will not be a problem. 30201da177e4SLinus Torvalds * 30211da177e4SLinus Torvalds * BSD seems to make the following compromise: 30221da177e4SLinus Torvalds * 30231da177e4SLinus Torvalds * If the free space is less than the 1/4 of the maximum 30241da177e4SLinus Torvalds * space available and the free space is less than 1/2 mss, 30251da177e4SLinus Torvalds * then set the window to 0. 30261da177e4SLinus Torvalds * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 30271da177e4SLinus Torvalds * Otherwise, just prevent the window from shrinking 30281da177e4SLinus Torvalds * and from being larger than the largest representable value. 30291da177e4SLinus Torvalds * 30301da177e4SLinus Torvalds * This prevents incremental opening of the window in the regime 30311da177e4SLinus Torvalds * where TCP is limited by the speed of the reader side taking 30321da177e4SLinus Torvalds * data out of the TCP receive queue. It does nothing about 30331da177e4SLinus Torvalds * those cases where the window is constrained on the sender side 30341da177e4SLinus Torvalds * because the pipeline is full. 30351da177e4SLinus Torvalds * 30361da177e4SLinus Torvalds * BSD also seems to "accidentally" limit itself to windows that are a 30371da177e4SLinus Torvalds * multiple of MSS, at least until the free space gets quite small. 30381da177e4SLinus Torvalds * This would appear to be a side effect of the mbuf implementation. 30391da177e4SLinus Torvalds * Combining these two algorithms results in the observed behavior 30401da177e4SLinus Torvalds * of having a fixed window size at almost all times. 30411da177e4SLinus Torvalds * 30421da177e4SLinus Torvalds * Below we obtain similar behavior by forcing the offered window to 30431da177e4SLinus Torvalds * a multiple of the mss when it is feasible to do so. 30441da177e4SLinus Torvalds * 30451da177e4SLinus Torvalds * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 30461da177e4SLinus Torvalds * Regular options like TIMESTAMP are taken into account. 30471da177e4SLinus Torvalds */ 30481da177e4SLinus Torvalds u32 __tcp_select_window(struct sock *sk) 30491da177e4SLinus Torvalds { 3050463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 30511da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 3052b650d953Smfreemon@cloudflare.com struct net *net = sock_net(sk); 3053caa20d9aSStephen Hemminger /* MSS for the peer's data. Previous versions used mss_clamp 30541da177e4SLinus Torvalds * here. I don't know if the value based on our guesses 30551da177e4SLinus Torvalds * of peer's MSS is better for the performance. It's more correct 30561da177e4SLinus Torvalds * but may be worse for the performance because of rcv_mss 30571da177e4SLinus Torvalds * fluctuations. --SAW 1998/11/1 30581da177e4SLinus Torvalds */ 3059463c84b9SArnaldo Carvalho de Melo int mss = icsk->icsk_ack.rcv_mss; 30601da177e4SLinus Torvalds int free_space = tcp_space(sk); 306186c1a045SFlorian Westphal int allowed_space = tcp_full_space(sk); 3062071c8ed6SFlorian Westphal int full_space, window; 3063071c8ed6SFlorian Westphal 3064071c8ed6SFlorian Westphal if (sk_is_mptcp(sk)) 3065071c8ed6SFlorian Westphal mptcp_space(sk, &free_space, &allowed_space); 3066071c8ed6SFlorian Westphal 3067071c8ed6SFlorian Westphal full_space = min_t(int, tp->window_clamp, allowed_space); 30681da177e4SLinus Torvalds 306906425c30SEric Dumazet if (unlikely(mss > full_space)) { 30701da177e4SLinus Torvalds mss = full_space; 307106425c30SEric Dumazet if (mss <= 0) 307206425c30SEric Dumazet return 0; 307306425c30SEric Dumazet } 3074b650d953Smfreemon@cloudflare.com 3075b650d953Smfreemon@cloudflare.com /* Only allow window shrink if the sysctl is enabled and we have 3076b650d953Smfreemon@cloudflare.com * a non-zero scaling factor in effect. 3077b650d953Smfreemon@cloudflare.com */ 3078b650d953Smfreemon@cloudflare.com if (READ_ONCE(net->ipv4.sysctl_tcp_shrink_window) && tp->rx_opt.rcv_wscale) 3079b650d953Smfreemon@cloudflare.com goto shrink_window_allowed; 3080b650d953Smfreemon@cloudflare.com 3081b650d953Smfreemon@cloudflare.com /* do not allow window to shrink */ 3082b650d953Smfreemon@cloudflare.com 3083b92edbe0SEric Dumazet if (free_space < (full_space >> 1)) { 3084463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.quick = 0; 30851da177e4SLinus Torvalds 3086b8da51ebSEric Dumazet if (tcp_under_memory_pressure(sk)) 3087053f3684SWei Wang tcp_adjust_rcv_ssthresh(sk); 30881da177e4SLinus Torvalds 308986c1a045SFlorian Westphal /* free_space might become our new window, make sure we don't 309086c1a045SFlorian Westphal * increase it due to wscale. 309186c1a045SFlorian Westphal */ 309286c1a045SFlorian Westphal free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale); 309386c1a045SFlorian Westphal 309486c1a045SFlorian Westphal /* if free space is less than mss estimate, or is below 1/16th 309586c1a045SFlorian Westphal * of the maximum allowed, try to move to zero-window, else 309686c1a045SFlorian Westphal * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and 309786c1a045SFlorian Westphal * new incoming data is dropped due to memory limits. 309886c1a045SFlorian Westphal * With large window, mss test triggers way too late in order 309986c1a045SFlorian Westphal * to announce zero window in time before rmem limit kicks in. 310086c1a045SFlorian Westphal */ 310186c1a045SFlorian Westphal if (free_space < (allowed_space >> 4) || free_space < mss) 31021da177e4SLinus Torvalds return 0; 31031da177e4SLinus Torvalds } 31041da177e4SLinus Torvalds 31051da177e4SLinus Torvalds if (free_space > tp->rcv_ssthresh) 31061da177e4SLinus Torvalds free_space = tp->rcv_ssthresh; 31071da177e4SLinus Torvalds 31081da177e4SLinus Torvalds /* Don't do rounding if we are using window scaling, since the 31091da177e4SLinus Torvalds * scaled window will not line up with the MSS boundary anyway. 31101da177e4SLinus Torvalds */ 31111da177e4SLinus Torvalds if (tp->rx_opt.rcv_wscale) { 31121da177e4SLinus Torvalds window = free_space; 31131da177e4SLinus Torvalds 31141da177e4SLinus Torvalds /* Advertise enough space so that it won't get scaled away. 31151da177e4SLinus Torvalds * Import case: prevent zero window announcement if 31161da177e4SLinus Torvalds * 1<<rcv_wscale > mss. 31171da177e4SLinus Torvalds */ 31181935299dSGao Feng window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale)); 31191da177e4SLinus Torvalds } else { 31201935299dSGao Feng window = tp->rcv_wnd; 31211da177e4SLinus Torvalds /* Get the largest window that is a nice multiple of mss. 31221da177e4SLinus Torvalds * Window clamp already applied above. 31231da177e4SLinus Torvalds * If our current window offering is within 1 mss of the 31241da177e4SLinus Torvalds * free space we just keep it. This prevents the divide 31251da177e4SLinus Torvalds * and multiply from happening most of the time. 31261da177e4SLinus Torvalds * We also don't do any window rounding when the free space 31271da177e4SLinus Torvalds * is too small. 31281da177e4SLinus Torvalds */ 31291da177e4SLinus Torvalds if (window <= free_space - mss || window > free_space) 31301935299dSGao Feng window = rounddown(free_space, mss); 313184565070SJohn Heffner else if (mss == full_space && 3132b92edbe0SEric Dumazet free_space > window + (full_space >> 1)) 313384565070SJohn Heffner window = free_space; 31341da177e4SLinus Torvalds } 31351da177e4SLinus Torvalds 31361da177e4SLinus Torvalds return window; 3137b650d953Smfreemon@cloudflare.com 3138b650d953Smfreemon@cloudflare.com shrink_window_allowed: 3139b650d953Smfreemon@cloudflare.com /* new window should always be an exact multiple of scaling factor */ 3140b650d953Smfreemon@cloudflare.com free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale); 3141b650d953Smfreemon@cloudflare.com 3142b650d953Smfreemon@cloudflare.com if (free_space < (full_space >> 1)) { 3143b650d953Smfreemon@cloudflare.com icsk->icsk_ack.quick = 0; 3144b650d953Smfreemon@cloudflare.com 3145b650d953Smfreemon@cloudflare.com if (tcp_under_memory_pressure(sk)) 3146b650d953Smfreemon@cloudflare.com tcp_adjust_rcv_ssthresh(sk); 3147b650d953Smfreemon@cloudflare.com 3148b650d953Smfreemon@cloudflare.com /* if free space is too low, return a zero window */ 3149b650d953Smfreemon@cloudflare.com if (free_space < (allowed_space >> 4) || free_space < mss || 3150b650d953Smfreemon@cloudflare.com free_space < (1 << tp->rx_opt.rcv_wscale)) 3151b650d953Smfreemon@cloudflare.com return 0; 3152b650d953Smfreemon@cloudflare.com } 3153b650d953Smfreemon@cloudflare.com 3154b650d953Smfreemon@cloudflare.com if (free_space > tp->rcv_ssthresh) { 3155b650d953Smfreemon@cloudflare.com free_space = tp->rcv_ssthresh; 3156b650d953Smfreemon@cloudflare.com /* new window should always be an exact multiple of scaling factor 3157b650d953Smfreemon@cloudflare.com * 3158b650d953Smfreemon@cloudflare.com * For this case, we ALIGN "up" (increase free_space) because 3159b650d953Smfreemon@cloudflare.com * we know free_space is not zero here, it has been reduced from 3160b650d953Smfreemon@cloudflare.com * the memory-based limit, and rcv_ssthresh is not a hard limit 3161b650d953Smfreemon@cloudflare.com * (unlike sk_rcvbuf). 3162b650d953Smfreemon@cloudflare.com */ 3163b650d953Smfreemon@cloudflare.com free_space = ALIGN(free_space, (1 << tp->rx_opt.rcv_wscale)); 3164b650d953Smfreemon@cloudflare.com } 3165b650d953Smfreemon@cloudflare.com 3166b650d953Smfreemon@cloudflare.com return free_space; 31671da177e4SLinus Torvalds } 31681da177e4SLinus Torvalds 3169cfea5a68SMartin KaFai Lau void tcp_skb_collapse_tstamp(struct sk_buff *skb, 3170082ac2d5SMartin KaFai Lau const struct sk_buff *next_skb) 3171082ac2d5SMartin KaFai Lau { 31720a2cf20cSSoheil Hassas Yeganeh if (unlikely(tcp_has_tx_tstamp(next_skb))) { 31730a2cf20cSSoheil Hassas Yeganeh const struct skb_shared_info *next_shinfo = 31740a2cf20cSSoheil Hassas Yeganeh skb_shinfo(next_skb); 3175082ac2d5SMartin KaFai Lau struct skb_shared_info *shinfo = skb_shinfo(skb); 3176082ac2d5SMartin KaFai Lau 31770a2cf20cSSoheil Hassas Yeganeh shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP; 3178082ac2d5SMartin KaFai Lau shinfo->tskey = next_shinfo->tskey; 31792de8023eSMartin KaFai Lau TCP_SKB_CB(skb)->txstamp_ack |= 31802de8023eSMartin KaFai Lau TCP_SKB_CB(next_skb)->txstamp_ack; 3181082ac2d5SMartin KaFai Lau } 3182082ac2d5SMartin KaFai Lau } 3183082ac2d5SMartin KaFai Lau 31844a17fc3aSIlpo Järvinen /* Collapses two adjacent SKB's during retransmission. */ 3185f8071cdeSEric Dumazet static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) 31861da177e4SLinus Torvalds { 31871da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 318875c119afSEric Dumazet struct sk_buff *next_skb = skb_rb_next(skb); 318913dde04fSWei Yongjun int next_skb_size; 31901da177e4SLinus Torvalds 3191058dc334SIlpo Järvinen next_skb_size = next_skb->len; 31921da177e4SLinus Torvalds 3193058dc334SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); 31941da177e4SLinus Torvalds 3195bd446314SEric Dumazet if (next_skb_size && !tcp_skb_shift(skb, next_skb, 1, next_skb_size)) 3196f8071cdeSEric Dumazet return false; 3197bd446314SEric Dumazet 31982b7cda9cSEric Dumazet tcp_highest_sack_replace(sk, next_skb, skb); 3199a6963a6bSIlpo Järvinen 32001da177e4SLinus Torvalds /* Update sequence range on original skb. */ 32011da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 32021da177e4SLinus Torvalds 3203e6c7d085SIlpo Järvinen /* Merge over control information. This moves PSH/FIN etc. over */ 32044de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; 32051da177e4SLinus Torvalds 32061da177e4SLinus Torvalds /* All done, get rid of second SKB and account for it so 32071da177e4SLinus Torvalds * packet counting does not break. 32081da177e4SLinus Torvalds */ 32094828e7f4SIlpo Järvinen TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; 3210a643b5d4SMartin KaFai Lau TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor; 3211b7689205SIlpo Järvinen 3212b7689205SIlpo Järvinen /* changed transmit queue under us so clear hints */ 3213ef9da47cSIlpo Järvinen tcp_clear_retrans_hints_partial(tp); 3214ef9da47cSIlpo Järvinen if (next_skb == tp->retransmit_skb_hint) 3215ef9da47cSIlpo Järvinen tp->retransmit_skb_hint = skb; 3216b7689205SIlpo Järvinen 3217797108d1SIlpo Järvinen tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); 3218797108d1SIlpo Järvinen 3219082ac2d5SMartin KaFai Lau tcp_skb_collapse_tstamp(skb, next_skb); 3220082ac2d5SMartin KaFai Lau 322175c119afSEric Dumazet tcp_rtx_queue_unlink_and_free(next_skb, sk); 3222f8071cdeSEric Dumazet return true; 32231da177e4SLinus Torvalds } 32241da177e4SLinus Torvalds 322567edfef7SAndi Kleen /* Check if coalescing SKBs is legal. */ 3226a2a385d6SEric Dumazet static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) 32274a17fc3aSIlpo Järvinen { 32284a17fc3aSIlpo Järvinen if (tcp_skb_pcount(skb) > 1) 3229a2a385d6SEric Dumazet return false; 32304a17fc3aSIlpo Järvinen if (skb_cloned(skb)) 3231a2a385d6SEric Dumazet return false; 32322331ccc5SEric Dumazet /* Some heuristics for collapsing over SACK'd could be invented */ 32334a17fc3aSIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 3234a2a385d6SEric Dumazet return false; 32354a17fc3aSIlpo Järvinen 3236a2a385d6SEric Dumazet return true; 32374a17fc3aSIlpo Järvinen } 32384a17fc3aSIlpo Järvinen 323967edfef7SAndi Kleen /* Collapse packets in the retransmit queue to make to create 324067edfef7SAndi Kleen * less packets on the wire. This is only done on retransmission. 324167edfef7SAndi Kleen */ 32424a17fc3aSIlpo Järvinen static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, 32434a17fc3aSIlpo Järvinen int space) 32444a17fc3aSIlpo Järvinen { 32454a17fc3aSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 32464a17fc3aSIlpo Järvinen struct sk_buff *skb = to, *tmp; 3247a2a385d6SEric Dumazet bool first = true; 32484a17fc3aSIlpo Järvinen 32491a63cb91SKuniyuki Iwashima if (!READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse)) 32504a17fc3aSIlpo Järvinen return; 32514de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 32524a17fc3aSIlpo Järvinen return; 32534a17fc3aSIlpo Järvinen 325475c119afSEric Dumazet skb_rbtree_walk_from_safe(skb, tmp) { 32554a17fc3aSIlpo Järvinen if (!tcp_can_collapse(sk, skb)) 32564a17fc3aSIlpo Järvinen break; 32574a17fc3aSIlpo Järvinen 325885712484SMat Martineau if (!tcp_skb_can_collapse(to, skb)) 3259a643b5d4SMartin KaFai Lau break; 3260a643b5d4SMartin KaFai Lau 32614a17fc3aSIlpo Järvinen space -= skb->len; 32624a17fc3aSIlpo Järvinen 32634a17fc3aSIlpo Järvinen if (first) { 3264a2a385d6SEric Dumazet first = false; 32654a17fc3aSIlpo Järvinen continue; 32664a17fc3aSIlpo Järvinen } 32674a17fc3aSIlpo Järvinen 32684a17fc3aSIlpo Järvinen if (space < 0) 32694a17fc3aSIlpo Järvinen break; 32704a17fc3aSIlpo Järvinen 32714a17fc3aSIlpo Järvinen if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) 32724a17fc3aSIlpo Järvinen break; 32734a17fc3aSIlpo Järvinen 3274f8071cdeSEric Dumazet if (!tcp_collapse_retrans(sk, to)) 3275f8071cdeSEric Dumazet break; 32764a17fc3aSIlpo Järvinen } 32774a17fc3aSIlpo Järvinen } 32784a17fc3aSIlpo Järvinen 32791da177e4SLinus Torvalds /* This retransmits one SKB. Policy decisions and retransmit queue 32801da177e4SLinus Torvalds * state updates are done by the caller. Returns non-zero if an 32811da177e4SLinus Torvalds * error occurred which prevented the send. 32821da177e4SLinus Torvalds */ 328310d3be56SEric Dumazet int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) 32841da177e4SLinus Torvalds { 32855d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 328610d3be56SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 32877d227cd2SSridhar Samudrala unsigned int cur_mss; 328810d3be56SEric Dumazet int diff, len, err; 3289536a6c8eSYonglong Li int avail_wnd; 329010d3be56SEric Dumazet 329110d3be56SEric Dumazet /* Inconclusive MTU probe */ 329210d3be56SEric Dumazet if (icsk->icsk_mtup.probe_size) 32935d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 32945d424d5aSJohn Heffner 32951f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 32961f3279aeSEric Dumazet return -EBUSY; 32971f3279aeSEric Dumazet 3298f99cd562SDong Chenchen start: 32991da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 3300f99cd562SDong Chenchen if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 3301f99cd562SDong Chenchen TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_SYN; 3302f99cd562SDong Chenchen TCP_SKB_CB(skb)->seq++; 3303f99cd562SDong Chenchen goto start; 3304f99cd562SDong Chenchen } 33057f582b24SEric Dumazet if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) { 33067f582b24SEric Dumazet WARN_ON_ONCE(1); 33077f582b24SEric Dumazet return -EINVAL; 33087f582b24SEric Dumazet } 33091da177e4SLinus Torvalds if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 33101da177e4SLinus Torvalds return -ENOMEM; 33111da177e4SLinus Torvalds } 33121da177e4SLinus Torvalds 33137d227cd2SSridhar Samudrala if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 33147d227cd2SSridhar Samudrala return -EHOSTUNREACH; /* Routing failure or similar. */ 33157d227cd2SSridhar Samudrala 33160c54b85fSIlpo Järvinen cur_mss = tcp_current_mss(sk); 3317536a6c8eSYonglong Li avail_wnd = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 33187d227cd2SSridhar Samudrala 33191da177e4SLinus Torvalds /* If receiver has shrunk his window, and skb is out of 33201da177e4SLinus Torvalds * new window, do not retransmit it. The exception is the 33211da177e4SLinus Torvalds * case, when window is shrunk to zero. In this case 3322536a6c8eSYonglong Li * our retransmit of one segment serves as a zero window probe. 33231da177e4SLinus Torvalds */ 3324536a6c8eSYonglong Li if (avail_wnd <= 0) { 3325536a6c8eSYonglong Li if (TCP_SKB_CB(skb)->seq != tp->snd_una) 33261da177e4SLinus Torvalds return -EAGAIN; 3327536a6c8eSYonglong Li avail_wnd = cur_mss; 3328536a6c8eSYonglong Li } 33291da177e4SLinus Torvalds 333010d3be56SEric Dumazet len = cur_mss * segs; 3331536a6c8eSYonglong Li if (len > avail_wnd) { 3332536a6c8eSYonglong Li len = rounddown(avail_wnd, cur_mss); 3333536a6c8eSYonglong Li if (!len) 3334536a6c8eSYonglong Li len = avail_wnd; 3335536a6c8eSYonglong Li } 333610d3be56SEric Dumazet if (skb->len > len) { 333775c119afSEric Dumazet if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len, 333875c119afSEric Dumazet cur_mss, GFP_ATOMIC)) 33391da177e4SLinus Torvalds return -ENOMEM; /* We'll try again later. */ 334002276f3cSIlpo Järvinen } else { 3341c4777efaSEric Dumazet if (skb_unclone_keeptruesize(skb, GFP_ATOMIC)) 3342c52e2421SEric Dumazet return -ENOMEM; 334310d3be56SEric Dumazet 334410d3be56SEric Dumazet diff = tcp_skb_pcount(skb); 334510d3be56SEric Dumazet tcp_set_skb_tso_segs(skb, cur_mss); 334610d3be56SEric Dumazet diff -= tcp_skb_pcount(skb); 334710d3be56SEric Dumazet if (diff) 334810d3be56SEric Dumazet tcp_adjust_pcount(sk, skb, diff); 3349536a6c8eSYonglong Li avail_wnd = min_t(int, avail_wnd, cur_mss); 3350536a6c8eSYonglong Li if (skb->len < avail_wnd) 3351536a6c8eSYonglong Li tcp_retrans_try_collapse(sk, skb, avail_wnd); 33521da177e4SLinus Torvalds } 33531da177e4SLinus Torvalds 335449213555SDaniel Borkmann /* RFC3168, section 6.1.1.1. ECN fallback */ 335549213555SDaniel Borkmann if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN) 335649213555SDaniel Borkmann tcp_ecn_clear_syn(sk, skb); 335749213555SDaniel Borkmann 3358678550c6SYuchung Cheng /* Update global and local TCP statistics. */ 3359678550c6SYuchung Cheng segs = tcp_skb_pcount(skb); 3360678550c6SYuchung Cheng TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs); 3361678550c6SYuchung Cheng if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 3362678550c6SYuchung Cheng __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 3363678550c6SYuchung Cheng tp->total_retrans += segs; 3364fb31c9b9SWei Wang tp->bytes_retrans += skb->len; 3365678550c6SYuchung Cheng 336650bceae9SThomas Graf /* make sure skb->data is aligned on arches that require it 336750bceae9SThomas Graf * and check if ack-trimming & collapsing extended the headroom 336850bceae9SThomas Graf * beyond what csum_start can cover. 336950bceae9SThomas Graf */ 337050bceae9SThomas Graf if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) || 337150bceae9SThomas Graf skb_headroom(skb) >= 0xFFFF)) { 337210a81980SEric Dumazet struct sk_buff *nskb; 337310a81980SEric Dumazet 3374e2080072SEric Dumazet tcp_skb_tsorted_save(skb) { 337510a81980SEric Dumazet nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC); 337607f8e4d0SFlorian Westphal if (nskb) { 337707f8e4d0SFlorian Westphal nskb->dev = NULL; 337807f8e4d0SFlorian Westphal err = tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC); 337907f8e4d0SFlorian Westphal } else { 338007f8e4d0SFlorian Westphal err = -ENOBUFS; 338107f8e4d0SFlorian Westphal } 3382e2080072SEric Dumazet } tcp_skb_tsorted_restore(skb); 3383e2080072SEric Dumazet 33845889e2c0SYousuk Seung if (!err) { 3385a7a25630SEric Dumazet tcp_update_skb_after_send(sk, skb, tp->tcp_wstamp_ns); 33865889e2c0SYousuk Seung tcp_rate_skb_sent(sk, skb); 33875889e2c0SYousuk Seung } 3388117632e6SEric Dumazet } else { 3389c84a5711SYuchung Cheng err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 3390117632e6SEric Dumazet } 3391c84a5711SYuchung Cheng 33927f12422cSYuchung Cheng /* To avoid taking spuriously low RTT samples based on a timestamp 33937f12422cSYuchung Cheng * for a transmit that never happened, always mark EVER_RETRANS 33947f12422cSYuchung Cheng */ 33957f12422cSYuchung Cheng TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS; 33967f12422cSYuchung Cheng 3397a31ad29eSLawrence Brakmo if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RETRANS_CB_FLAG)) 3398a31ad29eSLawrence Brakmo tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RETRANS_CB, 3399a31ad29eSLawrence Brakmo TCP_SKB_CB(skb)->seq, segs, err); 3400a31ad29eSLawrence Brakmo 3401fc9f3501SEric Dumazet if (likely(!err)) { 3402e086101bSCong Wang trace_tcp_retransmit_skb(sk, skb); 3403678550c6SYuchung Cheng } else if (err != -EBUSY) { 3404ec641b39SYuchung Cheng NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL, segs); 3405fc9f3501SEric Dumazet } 3406c84a5711SYuchung Cheng return err; 340793b174adSYuchung Cheng } 340893b174adSYuchung Cheng 340910d3be56SEric Dumazet int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) 341093b174adSYuchung Cheng { 341193b174adSYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 341210d3be56SEric Dumazet int err = __tcp_retransmit_skb(sk, skb, segs); 34131da177e4SLinus Torvalds 34141da177e4SLinus Torvalds if (err == 0) { 34151da177e4SLinus Torvalds #if FASTRETRANS_DEBUG > 0 34161da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 3417e87cc472SJoe Perches net_dbg_ratelimited("retrans_out leaked\n"); 34181da177e4SLinus Torvalds } 34191da177e4SLinus Torvalds #endif 34201da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 34211da177e4SLinus Torvalds tp->retrans_out += tcp_skb_pcount(skb); 34227ae18975SYuchung Cheng } 34231da177e4SLinus Torvalds 34247ae18975SYuchung Cheng /* Save stamp of the first (attempted) retransmit. */ 34251da177e4SLinus Torvalds if (!tp->retrans_stamp) 3426614e8316SEric Dumazet tp->retrans_stamp = tcp_skb_timestamp_ts(tp->tcp_usec_ts, skb); 34271da177e4SLinus Torvalds 34286e08d5e3SYuchung Cheng if (tp->undo_retrans < 0) 34296e08d5e3SYuchung Cheng tp->undo_retrans = 0; 34306e08d5e3SYuchung Cheng tp->undo_retrans += tcp_skb_pcount(skb); 34311da177e4SLinus Torvalds return err; 34321da177e4SLinus Torvalds } 34331da177e4SLinus Torvalds 34341da177e4SLinus Torvalds /* This gets called after a retransmit timeout, and the initially 34351da177e4SLinus Torvalds * retransmitted data is acknowledged. It tries to continue 34361da177e4SLinus Torvalds * resending the rest of the retransmit queue, until either 34371da177e4SLinus Torvalds * we've sent it all or the congestion window limit is reached. 34381da177e4SLinus Torvalds */ 34391da177e4SLinus Torvalds void tcp_xmit_retransmit_queue(struct sock *sk) 34401da177e4SLinus Torvalds { 34416687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 3442b9f1f1ceSEric Dumazet struct sk_buff *skb, *rtx_head, *hole = NULL; 34431da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 3444916e6d1aSEric Dumazet bool rearm_timer = false; 3445840a3cbeSYuchung Cheng u32 max_segs; 344661eb55f4SIlpo Järvinen int mib_idx; 34476a438bbeSStephen Hemminger 344845e77d31SIlpo Järvinen if (!tp->packets_out) 344945e77d31SIlpo Järvinen return; 345045e77d31SIlpo Järvinen 345175c119afSEric Dumazet rtx_head = tcp_rtx_queue_head(sk); 3452b9f1f1ceSEric Dumazet skb = tp->retransmit_skb_hint ?: rtx_head; 3453ed6e7268SNeal Cardwell max_segs = tcp_tso_segs(sk, tcp_current_mss(sk)); 345475c119afSEric Dumazet skb_rbtree_walk_from(skb) { 3455dca0aaf8SEric Dumazet __u8 sacked; 345610d3be56SEric Dumazet int segs; 34571da177e4SLinus Torvalds 3458218af599SEric Dumazet if (tcp_pacing_check(sk)) 3459218af599SEric Dumazet break; 3460218af599SEric Dumazet 34616a438bbeSStephen Hemminger /* we could do better than to assign each time */ 346251456b29SIan Morris if (!hole) 34636a438bbeSStephen Hemminger tp->retransmit_skb_hint = skb; 34646a438bbeSStephen Hemminger 346540570375SEric Dumazet segs = tcp_snd_cwnd(tp) - tcp_packets_in_flight(tp); 346610d3be56SEric Dumazet if (segs <= 0) 3467916e6d1aSEric Dumazet break; 3468dca0aaf8SEric Dumazet sacked = TCP_SKB_CB(skb)->sacked; 3469a3d2e9f8SEric Dumazet /* In case tcp_shift_skb_data() have aggregated large skbs, 3470a3d2e9f8SEric Dumazet * we need to make sure not sending too bigs TSO packets 3471a3d2e9f8SEric Dumazet */ 3472a3d2e9f8SEric Dumazet segs = min_t(int, segs, max_segs); 34730e1c54c2SIlpo Järvinen 3474840a3cbeSYuchung Cheng if (tp->retrans_out >= tp->lost_out) { 3475006f582cSIlpo Järvinen break; 34760e1c54c2SIlpo Järvinen } else if (!(sacked & TCPCB_LOST)) { 347751456b29SIan Morris if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) 34780e1c54c2SIlpo Järvinen hole = skb; 347961eb55f4SIlpo Järvinen continue; 34801da177e4SLinus Torvalds 34810e1c54c2SIlpo Järvinen } else { 34820e1c54c2SIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Loss) 34830e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFASTRETRANS; 34840e1c54c2SIlpo Järvinen else 34850e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; 34860e1c54c2SIlpo Järvinen } 34870e1c54c2SIlpo Järvinen 34880e1c54c2SIlpo Järvinen if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) 348961eb55f4SIlpo Järvinen continue; 349040b215e5SPavel Emelyanov 3491f9616c35SEric Dumazet if (tcp_small_queue_check(sk, skb, 1)) 3492916e6d1aSEric Dumazet break; 3493f9616c35SEric Dumazet 349410d3be56SEric Dumazet if (tcp_retransmit_skb(sk, skb, segs)) 3495916e6d1aSEric Dumazet break; 349624ab6becSYuchung Cheng 3497de1d6578SYuchung Cheng NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb)); 34981da177e4SLinus Torvalds 3499684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 3500a262f0cdSNandita Dukkipati tp->prr_out += tcp_skb_pcount(skb); 3501a262f0cdSNandita Dukkipati 350275c119afSEric Dumazet if (skb == rtx_head && 350357dde7f7SYuchung Cheng icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT) 3504916e6d1aSEric Dumazet rearm_timer = true; 3505916e6d1aSEric Dumazet 3506916e6d1aSEric Dumazet } 3507916e6d1aSEric Dumazet if (rearm_timer) 35083f80e08fSEric Dumazet tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 35093f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, 35108dc242adSEric Dumazet TCP_RTO_MAX); 35111da177e4SLinus Torvalds } 35121da177e4SLinus Torvalds 3513d83769a5SEric Dumazet /* We allow to exceed memory limits for FIN packets to expedite 3514d83769a5SEric Dumazet * connection tear down and (memory) recovery. 3515845704a5SEric Dumazet * Otherwise tcp_send_fin() could be tempted to either delay FIN 3516845704a5SEric Dumazet * or even be forced to close flow without any FIN. 3517a6c5ea4cSEric Dumazet * In general, we want to allow one skb per socket to avoid hangs 3518a6c5ea4cSEric Dumazet * with edge trigger epoll() 3519d83769a5SEric Dumazet */ 3520a6c5ea4cSEric Dumazet void sk_forced_mem_schedule(struct sock *sk, int size) 3521d83769a5SEric Dumazet { 3522c4ee1185SEric Dumazet int delta, amt; 3523d83769a5SEric Dumazet 3524c4ee1185SEric Dumazet delta = size - sk->sk_forward_alloc; 3525c4ee1185SEric Dumazet if (delta <= 0) 3526d83769a5SEric Dumazet return; 3527c4ee1185SEric Dumazet amt = sk_mem_pages(delta); 35285e6300e7SEric Dumazet sk_forward_alloc_add(sk, amt << PAGE_SHIFT); 3529e805605cSJohannes Weiner sk_memory_allocated_add(sk, amt); 3530e805605cSJohannes Weiner 3531baac50bbSJohannes Weiner if (mem_cgroup_sockets_enabled && sk->sk_memcg) 35324b1327beSWei Wang mem_cgroup_charge_skmem(sk->sk_memcg, amt, 35334b1327beSWei Wang gfp_memcg_charge() | __GFP_NOFAIL); 3534d83769a5SEric Dumazet } 3535d83769a5SEric Dumazet 3536845704a5SEric Dumazet /* Send a FIN. The caller locks the socket for us. 3537845704a5SEric Dumazet * We should try to send a FIN packet really hard, but eventually give up. 35381da177e4SLinus Torvalds */ 35391da177e4SLinus Torvalds void tcp_send_fin(struct sock *sk) 35401da177e4SLinus Torvalds { 3541ee2aabd3SEric Dumazet struct sk_buff *skb, *tskb, *tail = tcp_write_queue_tail(sk); 35421da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 35431da177e4SLinus Torvalds 3544845704a5SEric Dumazet /* Optimization, tack on the FIN if we have one skb in write queue and 3545845704a5SEric Dumazet * this skb was not yet sent, or we are under memory pressure. 3546845704a5SEric Dumazet * Note: in the latter case, FIN packet will be sent after a timeout, 3547845704a5SEric Dumazet * as TCP stack thinks it has already been transmitted. 35481da177e4SLinus Torvalds */ 3549ee2aabd3SEric Dumazet tskb = tail; 355075c119afSEric Dumazet if (!tskb && tcp_under_memory_pressure(sk)) 355175c119afSEric Dumazet tskb = skb_rb_last(&sk->tcp_rtx_queue); 355275c119afSEric Dumazet 355375c119afSEric Dumazet if (tskb) { 3554845704a5SEric Dumazet TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN; 3555845704a5SEric Dumazet TCP_SKB_CB(tskb)->end_seq++; 35561da177e4SLinus Torvalds tp->write_seq++; 3557ee2aabd3SEric Dumazet if (!tail) { 3558845704a5SEric Dumazet /* This means tskb was already sent. 3559845704a5SEric Dumazet * Pretend we included the FIN on previous transmit. 3560845704a5SEric Dumazet * We need to set tp->snd_nxt to the value it would have 3561845704a5SEric Dumazet * if FIN had been sent. This is because retransmit path 3562845704a5SEric Dumazet * does not change tp->snd_nxt. 3563845704a5SEric Dumazet */ 3564e0d694d6SEric Dumazet WRITE_ONCE(tp->snd_nxt, tp->snd_nxt + 1); 3565845704a5SEric Dumazet return; 3566845704a5SEric Dumazet } 35671da177e4SLinus Torvalds } else { 3568845704a5SEric Dumazet skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation); 3569d1edc085SColin Ian King if (unlikely(!skb)) 3570845704a5SEric Dumazet return; 3571d1edc085SColin Ian King 3572e2080072SEric Dumazet INIT_LIST_HEAD(&skb->tcp_tsorted_anchor); 3573d83769a5SEric Dumazet skb_reserve(skb, MAX_TCP_HEADER); 3574a6c5ea4cSEric Dumazet sk_forced_mem_schedule(sk, skb->truesize); 35751da177e4SLinus Torvalds /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 3576e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tp->write_seq, 3577a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_FIN); 35781da177e4SLinus Torvalds tcp_queue_skb(sk, skb); 35791da177e4SLinus Torvalds } 3580845704a5SEric Dumazet __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF); 35811da177e4SLinus Torvalds } 35821da177e4SLinus Torvalds 35831da177e4SLinus Torvalds /* We get here when a process closes a file descriptor (either due to 35841da177e4SLinus Torvalds * an explicit close() or as a byproduct of exit()'ing) and there 35851da177e4SLinus Torvalds * was unread data in the receive queue. This behavior is recommended 358665bb723cSGerrit Renker * by RFC 2525, section 2.17. -DaveM 35871da177e4SLinus Torvalds */ 3588dd0fc66fSAl Viro void tcp_send_active_reset(struct sock *sk, gfp_t priority) 35891da177e4SLinus Torvalds { 35901da177e4SLinus Torvalds struct sk_buff *skb; 35911da177e4SLinus Torvalds 35927cc2b043SGao Feng TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); 35937cc2b043SGao Feng 35941da177e4SLinus Torvalds /* NOTE: No TCP options attached and we never retransmit this. */ 35951da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, priority); 35961da177e4SLinus Torvalds if (!skb) { 35974e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 35981da177e4SLinus Torvalds return; 35991da177e4SLinus Torvalds } 36001da177e4SLinus Torvalds 36011da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 36021da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 3603e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), 3604a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_RST); 36059a568de4SEric Dumazet tcp_mstamp_refresh(tcp_sk(sk)); 36061da177e4SLinus Torvalds /* Send it off. */ 3607dfb4b9dcSDavid S. Miller if (tcp_transmit_skb(sk, skb, 0, priority)) 36084e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 3609c24b14c4SSong Liu 3610c24b14c4SSong Liu /* skb of trace_tcp_send_reset() keeps the skb that caused RST, 3611c24b14c4SSong Liu * skb here is different to the troublesome skb, so use NULL 3612c24b14c4SSong Liu */ 3613c24b14c4SSong Liu trace_tcp_send_reset(sk, NULL); 36141da177e4SLinus Torvalds } 36151da177e4SLinus Torvalds 361667edfef7SAndi Kleen /* Send a crossed SYN-ACK during socket establishment. 361767edfef7SAndi Kleen * WARNING: This routine must only be called when we have already sent 36181da177e4SLinus Torvalds * a SYN packet that crossed the incoming SYN that caused this routine 36191da177e4SLinus Torvalds * to get called. If this assumption fails then the initial rcv_wnd 36201da177e4SLinus Torvalds * and rcv_wscale values will not be correct. 36211da177e4SLinus Torvalds */ 36221da177e4SLinus Torvalds int tcp_send_synack(struct sock *sk) 36231da177e4SLinus Torvalds { 36241da177e4SLinus Torvalds struct sk_buff *skb; 36251da177e4SLinus Torvalds 362675c119afSEric Dumazet skb = tcp_rtx_queue_head(sk); 362751456b29SIan Morris if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 362875c119afSEric Dumazet pr_err("%s: wrong queue state\n", __func__); 36291da177e4SLinus Torvalds return -EFAULT; 36301da177e4SLinus Torvalds } 36314de075e0SEric Dumazet if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { 36321da177e4SLinus Torvalds if (skb_cloned(skb)) { 3633e2080072SEric Dumazet struct sk_buff *nskb; 3634e2080072SEric Dumazet 3635e2080072SEric Dumazet tcp_skb_tsorted_save(skb) { 3636e2080072SEric Dumazet nskb = skb_copy(skb, GFP_ATOMIC); 3637e2080072SEric Dumazet } tcp_skb_tsorted_restore(skb); 363851456b29SIan Morris if (!nskb) 36391da177e4SLinus Torvalds return -ENOMEM; 3640e2080072SEric Dumazet INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor); 36412bec445fSEric Dumazet tcp_highest_sack_replace(sk, skb, nskb); 364275c119afSEric Dumazet tcp_rtx_queue_unlink_and_free(skb, sk); 3643f4a775d1SEric Dumazet __skb_header_release(nskb); 364475c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb); 3645ab4e846aSEric Dumazet sk_wmem_queued_add(sk, nskb->truesize); 36463ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 36471da177e4SLinus Torvalds skb = nskb; 36481da177e4SLinus Torvalds } 36491da177e4SLinus Torvalds 36504de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; 3651735d3831SFlorian Westphal tcp_ecn_send_synack(sk, skb); 36521da177e4SLinus Torvalds } 3653dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 36541da177e4SLinus Torvalds } 36551da177e4SLinus Torvalds 36564aea39c1SEric Dumazet /** 3657331fca43SMartin KaFai Lau * tcp_make_synack - Allocate one skb and build a SYNACK packet. 3658331fca43SMartin KaFai Lau * @sk: listener socket 3659331fca43SMartin KaFai Lau * @dst: dst entry attached to the SYNACK. It is consumed and caller 3660331fca43SMartin KaFai Lau * should not use it again. 3661331fca43SMartin KaFai Lau * @req: request_sock pointer 3662331fca43SMartin KaFai Lau * @foc: cookie for tcp fast open 3663331fca43SMartin KaFai Lau * @synack_type: Type of synack to prepare 3664331fca43SMartin KaFai Lau * @syn_skb: SYN packet just received. It could be NULL for rtx case. 36654aea39c1SEric Dumazet */ 36665d062de7SEric Dumazet struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst, 3667e6b4d113SWilliam Allen Simpson struct request_sock *req, 3668ca6fb065SEric Dumazet struct tcp_fastopen_cookie *foc, 3669331fca43SMartin KaFai Lau enum tcp_synack_type synack_type, 3670331fca43SMartin KaFai Lau struct sk_buff *syn_skb) 36711da177e4SLinus Torvalds { 36722e6599cbSArnaldo Carvalho de Melo struct inet_request_sock *ireq = inet_rsk(req); 36735d062de7SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 36745d062de7SEric Dumazet struct tcp_out_options opts; 36751e03d32bSDmitry Safonov struct tcp_key key = {}; 36765d062de7SEric Dumazet struct sk_buff *skb; 3677bd0388aeSWilliam Allen Simpson int tcp_header_size; 36785d062de7SEric Dumazet struct tcphdr *th; 3679f5fff5dcSTom Quetchenbach int mss; 3680a842fe14SEric Dumazet u64 now; 36811da177e4SLinus Torvalds 3682ca6fb065SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 36834aea39c1SEric Dumazet if (unlikely(!skb)) { 36844aea39c1SEric Dumazet dst_release(dst); 36851da177e4SLinus Torvalds return NULL; 36864aea39c1SEric Dumazet } 36871da177e4SLinus Torvalds /* Reserve space for headers. */ 36881da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 36891da177e4SLinus Torvalds 3690b3d05147SEric Dumazet switch (synack_type) { 3691b3d05147SEric Dumazet case TCP_SYNACK_NORMAL: 36929e17f8a4SEric Dumazet skb_set_owner_w(skb, req_to_sk(req)); 3693b3d05147SEric Dumazet break; 3694b3d05147SEric Dumazet case TCP_SYNACK_COOKIE: 3695b3d05147SEric Dumazet /* Under synflood, we do not attach skb to a socket, 3696b3d05147SEric Dumazet * to avoid false sharing. 3697b3d05147SEric Dumazet */ 3698b3d05147SEric Dumazet break; 3699b3d05147SEric Dumazet case TCP_SYNACK_FASTOPEN: 3700ca6fb065SEric Dumazet /* sk is a const pointer, because we want to express multiple 3701ca6fb065SEric Dumazet * cpu might call us concurrently. 3702ca6fb065SEric Dumazet * sk->sk_wmem_alloc in an atomic, we can promote to rw. 3703ca6fb065SEric Dumazet */ 3704ca6fb065SEric Dumazet skb_set_owner_w(skb, (struct sock *)sk); 3705b3d05147SEric Dumazet break; 3706ca6fb065SEric Dumazet } 37074aea39c1SEric Dumazet skb_dst_set(skb, dst); 37081da177e4SLinus Torvalds 37093541f9e8SEric Dumazet mss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 3710f5fff5dcSTom Quetchenbach 371133ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 3712a842fe14SEric Dumazet now = tcp_clock_ns(); 37138b5f12d0SFlorian Westphal #ifdef CONFIG_SYN_COOKIES 3714f8ace8d9SFlorian Westphal if (unlikely(synack_type == TCP_SYNACK_COOKIE && ireq->tstamp_ok)) 3715a1ac9c8aSMartin KaFai Lau skb_set_delivery_time(skb, cookie_init_timestamp(req, now), 3716a1ac9c8aSMartin KaFai Lau true); 37178b5f12d0SFlorian Westphal else 37188b5f12d0SFlorian Westphal #endif 37199e450c1eSYuchung Cheng { 3720a1ac9c8aSMartin KaFai Lau skb_set_delivery_time(skb, now, true); 37219e450c1eSYuchung Cheng if (!tcp_rsk(req)->snt_synack) /* Timestamp first SYNACK */ 37229e450c1eSYuchung Cheng tcp_rsk(req)->snt_synack = tcp_skb_timestamp_us(skb); 37239e450c1eSYuchung Cheng } 372480f03e27SEric Dumazet 37259427c6aaSDmitry Safonov #if defined(CONFIG_TCP_MD5SIG) || defined(CONFIG_TCP_AO) 372680f03e27SEric Dumazet rcu_read_lock(); 37279427c6aaSDmitry Safonov #endif 37289427c6aaSDmitry Safonov if (tcp_rsk_used_ao(req)) { 37299427c6aaSDmitry Safonov #ifdef CONFIG_TCP_AO 37309427c6aaSDmitry Safonov struct tcp_ao_key *ao_key = NULL; 37319427c6aaSDmitry Safonov u8 keyid = tcp_rsk(req)->ao_keyid; 37329427c6aaSDmitry Safonov 37339427c6aaSDmitry Safonov ao_key = tcp_sk(sk)->af_specific->ao_lookup(sk, req_to_sk(req), 37349427c6aaSDmitry Safonov keyid, -1); 37359427c6aaSDmitry Safonov /* If there is no matching key - avoid sending anything, 37369427c6aaSDmitry Safonov * especially usigned segments. It could try harder and lookup 37379427c6aaSDmitry Safonov * for another peer-matching key, but the peer has requested 37389427c6aaSDmitry Safonov * ao_keyid (RFC5925 RNextKeyID), so let's keep it simple here. 37399427c6aaSDmitry Safonov */ 37409396c4eeSDmitry Safonov if (unlikely(!ao_key)) { 37419427c6aaSDmitry Safonov rcu_read_unlock(); 37429427c6aaSDmitry Safonov kfree_skb(skb); 37439396c4eeSDmitry Safonov net_warn_ratelimited("TCP-AO: the keyid %u from SYN packet is not present - not sending SYNACK\n", 37449396c4eeSDmitry Safonov keyid); 37459427c6aaSDmitry Safonov return NULL; 37469427c6aaSDmitry Safonov } 37479427c6aaSDmitry Safonov key.ao_key = ao_key; 37489427c6aaSDmitry Safonov key.type = TCP_KEY_AO; 37499427c6aaSDmitry Safonov #endif 37509427c6aaSDmitry Safonov } else { 37519427c6aaSDmitry Safonov #ifdef CONFIG_TCP_MD5SIG 37529427c6aaSDmitry Safonov key.md5_key = tcp_rsk(req)->af_specific->req_md5_lookup(sk, 37539427c6aaSDmitry Safonov req_to_sk(req)); 37549427c6aaSDmitry Safonov if (key.md5_key) 37551e03d32bSDmitry Safonov key.type = TCP_KEY_MD5; 375680f03e27SEric Dumazet #endif 37579427c6aaSDmitry Safonov } 37585e526552SEric Dumazet skb_set_hash(skb, READ_ONCE(tcp_rsk(req)->txhash), PKT_HASH_TYPE_L4); 3759331fca43SMartin KaFai Lau /* bpf program will be interested in the tcp_flags */ 3760331fca43SMartin KaFai Lau TCP_SKB_CB(skb)->tcp_flags = TCPHDR_SYN | TCPHDR_ACK; 37619427c6aaSDmitry Safonov tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, 37629427c6aaSDmitry Safonov &key, foc, synack_type, syn_skb) 37639427c6aaSDmitry Safonov + sizeof(*th); 376433ad798cSAdam Langley 3765aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 3766aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 37671da177e4SLinus Torvalds 3768ea1627c2SEric Dumazet th = (struct tcphdr *)skb->data; 37691da177e4SLinus Torvalds memset(th, 0, sizeof(struct tcphdr)); 37701da177e4SLinus Torvalds th->syn = 1; 37711da177e4SLinus Torvalds th->ack = 1; 37726ac705b1SEric Dumazet tcp_ecn_make_synack(req, th); 3773b44084c2SEric Dumazet th->source = htons(ireq->ir_num); 3774634fb979SEric Dumazet th->dest = ireq->ir_rmt_port; 3775e05a90ecSJamal Hadi Salim skb->mark = ireq->ir_mark; 37763b117750SEric Dumazet skb->ip_summed = CHECKSUM_PARTIAL; 37773b117750SEric Dumazet th->seq = htonl(tcp_rsk(req)->snt_isn); 37788336886fSJerry Chu /* XXX data is queued and acked as is. No buffer/window check */ 37798336886fSJerry Chu th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt); 37801da177e4SLinus Torvalds 37811da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 3782ed53d0abSEric Dumazet th->window = htons(min(req->rsk_rcv_wnd, 65535U)); 37839427c6aaSDmitry Safonov tcp_options_write(th, NULL, tcp_rsk(req), &opts, &key); 37841da177e4SLinus Torvalds th->doff = (tcp_header_size >> 2); 3785bced3f7dSBreno Leitao TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS); 3786cfb6eeb4SYOSHIFUJI Hideaki 3787cfb6eeb4SYOSHIFUJI Hideaki /* Okay, we have all we need - do the md5 hash if needed */ 37889427c6aaSDmitry Safonov if (tcp_key_is_md5(&key)) { 37899427c6aaSDmitry Safonov #ifdef CONFIG_TCP_MD5SIG 3790bd0388aeSWilliam Allen Simpson tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, 37919427c6aaSDmitry Safonov key.md5_key, req_to_sk(req), skb); 37929427c6aaSDmitry Safonov #endif 37939427c6aaSDmitry Safonov } else if (tcp_key_is_ao(&key)) { 37949427c6aaSDmitry Safonov #ifdef CONFIG_TCP_AO 37959427c6aaSDmitry Safonov tcp_rsk(req)->af_specific->ao_synack_hash(opts.hash_location, 37969427c6aaSDmitry Safonov key.ao_key, req, skb, 37979427c6aaSDmitry Safonov opts.hash_location - (u8 *)th, 0); 37989427c6aaSDmitry Safonov #endif 37999427c6aaSDmitry Safonov } 38009427c6aaSDmitry Safonov #if defined(CONFIG_TCP_MD5SIG) || defined(CONFIG_TCP_AO) 380180f03e27SEric Dumazet rcu_read_unlock(); 3802cfb6eeb4SYOSHIFUJI Hideaki #endif 3803cfb6eeb4SYOSHIFUJI Hideaki 3804331fca43SMartin KaFai Lau bpf_skops_write_hdr_opt((struct sock *)sk, skb, req, syn_skb, 3805331fca43SMartin KaFai Lau synack_type, &opts); 3806331fca43SMartin KaFai Lau 3807a1ac9c8aSMartin KaFai Lau skb_set_delivery_time(skb, now, true); 3808a842fe14SEric Dumazet tcp_add_tx_delay(skb, tp); 3809a842fe14SEric Dumazet 38101da177e4SLinus Torvalds return skb; 38111da177e4SLinus Torvalds } 38124bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_make_synack); 38131da177e4SLinus Torvalds 381481164413SDaniel Borkmann static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst) 381581164413SDaniel Borkmann { 381681164413SDaniel Borkmann struct inet_connection_sock *icsk = inet_csk(sk); 381781164413SDaniel Borkmann const struct tcp_congestion_ops *ca; 381881164413SDaniel Borkmann u32 ca_key = dst_metric(dst, RTAX_CC_ALGO); 381981164413SDaniel Borkmann 382081164413SDaniel Borkmann if (ca_key == TCP_CA_UNSPEC) 382181164413SDaniel Borkmann return; 382281164413SDaniel Borkmann 382381164413SDaniel Borkmann rcu_read_lock(); 382481164413SDaniel Borkmann ca = tcp_ca_find_key(ca_key); 38250baf26b0SMartin KaFai Lau if (likely(ca && bpf_try_module_get(ca, ca->owner))) { 38260baf26b0SMartin KaFai Lau bpf_module_put(icsk->icsk_ca_ops, icsk->icsk_ca_ops->owner); 382781164413SDaniel Borkmann icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst); 382881164413SDaniel Borkmann icsk->icsk_ca_ops = ca; 382981164413SDaniel Borkmann } 383081164413SDaniel Borkmann rcu_read_unlock(); 383181164413SDaniel Borkmann } 383281164413SDaniel Borkmann 383367edfef7SAndi Kleen /* Do all connect socket setups that can be done AF independent. */ 3834f7e56a76Sstephen hemminger static void tcp_connect_init(struct sock *sk) 38351da177e4SLinus Torvalds { 3836cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 38371da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 38381da177e4SLinus Torvalds __u8 rcv_wscale; 383913d3b1ebSLawrence Brakmo u32 rcv_wnd; 38401da177e4SLinus Torvalds 38411da177e4SLinus Torvalds /* We'll fix this up when we get a response from the other end. 38421da177e4SLinus Torvalds * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 38431da177e4SLinus Torvalds */ 38445d2ed052SEric Dumazet tp->tcp_header_len = sizeof(struct tcphdr); 38453666f666SKuniyuki Iwashima if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_timestamps)) 38465d2ed052SEric Dumazet tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED; 38471da177e4SLinus Torvalds 38487c2ffaf2SDmitry Safonov tcp_ao_connect_init(sk); 38497c2ffaf2SDmitry Safonov 38501da177e4SLinus Torvalds /* If user gave his TCP_MAXSEG, record it to clamp */ 38511da177e4SLinus Torvalds if (tp->rx_opt.user_mss) 38521da177e4SLinus Torvalds tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 38531da177e4SLinus Torvalds tp->max_window = 0; 38545d424d5aSJohn Heffner tcp_mtup_init(sk); 38551da177e4SLinus Torvalds tcp_sync_mss(sk, dst_mtu(dst)); 38561da177e4SLinus Torvalds 385781164413SDaniel Borkmann tcp_ca_dst_init(sk, dst); 385881164413SDaniel Borkmann 38591da177e4SLinus Torvalds if (!tp->window_clamp) 3860*f410cbeaSEric Dumazet WRITE_ONCE(tp->window_clamp, dst_metric(dst, RTAX_WINDOW)); 38613541f9e8SEric Dumazet tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 3862f5fff5dcSTom Quetchenbach 38631da177e4SLinus Torvalds tcp_initialize_rcv_mss(sk); 38641da177e4SLinus Torvalds 3865e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 3866e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 3867e88c64f0SHagen Paul Pfeifer (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) 3868*f410cbeaSEric Dumazet WRITE_ONCE(tp->window_clamp, tcp_full_space(sk)); 3869e88c64f0SHagen Paul Pfeifer 387013d3b1ebSLawrence Brakmo rcv_wnd = tcp_rwnd_init_bpf(sk); 387113d3b1ebSLawrence Brakmo if (rcv_wnd == 0) 387213d3b1ebSLawrence Brakmo rcv_wnd = dst_metric(dst, RTAX_INITRWND); 387313d3b1ebSLawrence Brakmo 3874ceef9ab6SEric Dumazet tcp_select_initial_window(sk, tcp_full_space(sk), 38751da177e4SLinus Torvalds tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 38761da177e4SLinus Torvalds &tp->rcv_wnd, 38771da177e4SLinus Torvalds &tp->window_clamp, 38783666f666SKuniyuki Iwashima READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_window_scaling), 387931d12926Slaurent chavey &rcv_wscale, 388013d3b1ebSLawrence Brakmo rcv_wnd); 38811da177e4SLinus Torvalds 38821da177e4SLinus Torvalds tp->rx_opt.rcv_wscale = rcv_wscale; 38831da177e4SLinus Torvalds tp->rcv_ssthresh = tp->rcv_wnd; 38841da177e4SLinus Torvalds 3885e13ec3daSEric Dumazet WRITE_ONCE(sk->sk_err, 0); 38861da177e4SLinus Torvalds sock_reset_flag(sk, SOCK_DONE); 38871da177e4SLinus Torvalds tp->snd_wnd = 0; 3888ee7537b6SHantzis Fotis tcp_init_wl(tp, 0); 38897f582b24SEric Dumazet tcp_write_queue_purge(sk); 38901da177e4SLinus Torvalds tp->snd_una = tp->write_seq; 38911da177e4SLinus Torvalds tp->snd_sml = tp->write_seq; 389233f5f57eSIlpo Järvinen tp->snd_up = tp->write_seq; 3893e0d694d6SEric Dumazet WRITE_ONCE(tp->snd_nxt, tp->write_seq); 3894ee995283SPavel Emelyanov 3895ee995283SPavel Emelyanov if (likely(!tp->repair)) 38961da177e4SLinus Torvalds tp->rcv_nxt = 0; 3897c7781a6eSAndrew Vagin else 389870eabf0eSEric Dumazet tp->rcv_tstamp = tcp_jiffies32; 3899ee995283SPavel Emelyanov tp->rcv_wup = tp->rcv_nxt; 39007db48e98SEric Dumazet WRITE_ONCE(tp->copied_seq, tp->rcv_nxt); 39011da177e4SLinus Torvalds 39028550f328SLawrence Brakmo inet_csk(sk)->icsk_rto = tcp_timeout_init(sk); 3903463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_retransmits = 0; 39041da177e4SLinus Torvalds tcp_clear_retrans(tp); 39051da177e4SLinus Torvalds } 39061da177e4SLinus Torvalds 3907783237e8SYuchung Cheng static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) 3908783237e8SYuchung Cheng { 3909783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3910783237e8SYuchung Cheng struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 3911783237e8SYuchung Cheng 3912783237e8SYuchung Cheng tcb->end_seq += skb->len; 3913f4a775d1SEric Dumazet __skb_header_release(skb); 3914ab4e846aSEric Dumazet sk_wmem_queued_add(sk, skb->truesize); 3915783237e8SYuchung Cheng sk_mem_charge(sk, skb->truesize); 39160f317464SEric Dumazet WRITE_ONCE(tp->write_seq, tcb->end_seq); 3917783237e8SYuchung Cheng tp->packets_out += tcp_skb_pcount(skb); 3918783237e8SYuchung Cheng } 3919783237e8SYuchung Cheng 3920783237e8SYuchung Cheng /* Build and send a SYN with data and (cached) Fast Open cookie. However, 3921783237e8SYuchung Cheng * queue a data-only packet after the regular SYN, such that regular SYNs 3922783237e8SYuchung Cheng * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges 3923783237e8SYuchung Cheng * only the SYN sequence, the data are retransmitted in the first ACK. 3924783237e8SYuchung Cheng * If cookie is not cached or other error occurs, falls back to send a 3925783237e8SYuchung Cheng * regular SYN with Fast Open cookie request option. 3926783237e8SYuchung Cheng */ 3927783237e8SYuchung Cheng static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) 3928783237e8SYuchung Cheng { 3929ed0c99dcSJakub Kicinski struct inet_connection_sock *icsk = inet_csk(sk); 3930783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3931783237e8SYuchung Cheng struct tcp_fastopen_request *fo = tp->fastopen_req; 3932fbf93406SEric Dumazet struct page_frag *pfrag = sk_page_frag(sk); 3933355a901eSEric Dumazet struct sk_buff *syn_data; 3934fbf93406SEric Dumazet int space, err = 0; 3935783237e8SYuchung Cheng 393667da22d2SYuchung Cheng tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ 3937065263f4SWei Wang if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie)) 3938783237e8SYuchung Cheng goto fallback; 3939783237e8SYuchung Cheng 3940783237e8SYuchung Cheng /* MSS for SYN-data is based on cached MSS and bounded by PMTU and 3941783237e8SYuchung Cheng * user-MSS. Reserve maximum option space for middleboxes that add 3942783237e8SYuchung Cheng * private TCP options. The cost is reduced data space in SYN :( 3943783237e8SYuchung Cheng */ 39443541f9e8SEric Dumazet tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp); 3945ed0c99dcSJakub Kicinski /* Sync mss_cache after updating the mss_clamp */ 3946ed0c99dcSJakub Kicinski tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 39473541f9e8SEric Dumazet 3948ed0c99dcSJakub Kicinski space = __tcp_mtu_to_mss(sk, icsk->icsk_pmtu_cookie) - 3949783237e8SYuchung Cheng MAX_TCP_OPTION_SPACE; 3950783237e8SYuchung Cheng 3951f5ddcbbbSEric Dumazet space = min_t(size_t, space, fo->size); 3952f5ddcbbbSEric Dumazet 3953fbf93406SEric Dumazet if (space && 3954fbf93406SEric Dumazet !skb_page_frag_refill(min_t(size_t, space, PAGE_SIZE), 3955fbf93406SEric Dumazet pfrag, sk->sk_allocation)) 3956fbf93406SEric Dumazet goto fallback; 39575882efffSEric Dumazet syn_data = tcp_stream_alloc_skb(sk, sk->sk_allocation, false); 3958355a901eSEric Dumazet if (!syn_data) 3959783237e8SYuchung Cheng goto fallback; 3960355a901eSEric Dumazet memcpy(syn_data->cb, syn->cb, sizeof(syn->cb)); 396107e100f9SEric Dumazet if (space) { 3962fbf93406SEric Dumazet space = min_t(size_t, space, pfrag->size - pfrag->offset); 3963fbf93406SEric Dumazet space = tcp_wmem_schedule(sk, space); 3964fbf93406SEric Dumazet } 3965fbf93406SEric Dumazet if (space) { 3966fbf93406SEric Dumazet space = copy_page_from_iter(pfrag->page, pfrag->offset, 3967fbf93406SEric Dumazet space, &fo->data->msg_iter); 3968fbf93406SEric Dumazet if (unlikely(!space)) { 3969ba233b34SEric Dumazet tcp_skb_tsorted_anchor_cleanup(syn_data); 3970355a901eSEric Dumazet kfree_skb(syn_data); 3971783237e8SYuchung Cheng goto fallback; 3972783237e8SYuchung Cheng } 3973fbf93406SEric Dumazet skb_fill_page_desc(syn_data, 0, pfrag->page, 3974fbf93406SEric Dumazet pfrag->offset, space); 3975fbf93406SEric Dumazet page_ref_inc(pfrag->page); 3976fbf93406SEric Dumazet pfrag->offset += space; 3977fbf93406SEric Dumazet skb_len_add(syn_data, space); 3978f859a448SWillem de Bruijn skb_zcopy_set(syn_data, fo->uarg, NULL); 397907e100f9SEric Dumazet } 3980355a901eSEric Dumazet /* No more data pending in inet_wait_for_connect() */ 3981355a901eSEric Dumazet if (space == fo->size) 3982355a901eSEric Dumazet fo->data = NULL; 3983355a901eSEric Dumazet fo->copied = space; 3984783237e8SYuchung Cheng 3985355a901eSEric Dumazet tcp_connect_queue_skb(sk, syn_data); 39860f87230dSFrancis Yan if (syn_data->len) 39870f87230dSFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_BUSY); 3988355a901eSEric Dumazet 3989355a901eSEric Dumazet err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation); 3990355a901eSEric Dumazet 3991a1ac9c8aSMartin KaFai Lau skb_set_delivery_time(syn, syn_data->skb_mstamp_ns, true); 3992355a901eSEric Dumazet 3993355a901eSEric Dumazet /* Now full SYN+DATA was cloned and sent (or not), 3994355a901eSEric Dumazet * remove the SYN from the original skb (syn_data) 3995355a901eSEric Dumazet * we keep in write queue in case of a retransmit, as we 3996355a901eSEric Dumazet * also have the SYN packet (with no data) in the same queue. 3997431a9124SEric Dumazet */ 3998355a901eSEric Dumazet TCP_SKB_CB(syn_data)->seq++; 3999355a901eSEric Dumazet TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH; 4000355a901eSEric Dumazet if (!err) { 400167da22d2SYuchung Cheng tp->syn_data = (fo->copied > 0); 400275c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data); 4003f19c29e3SYuchung Cheng NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT); 4004783237e8SYuchung Cheng goto done; 4005783237e8SYuchung Cheng } 4006783237e8SYuchung Cheng 400775c119afSEric Dumazet /* data was not sent, put it in write_queue */ 400875c119afSEric Dumazet __skb_queue_tail(&sk->sk_write_queue, syn_data); 4009b5b7db8dSEric Dumazet tp->packets_out -= tcp_skb_pcount(syn_data); 4010b5b7db8dSEric Dumazet 4011783237e8SYuchung Cheng fallback: 4012783237e8SYuchung Cheng /* Send a regular SYN with Fast Open cookie request option */ 4013783237e8SYuchung Cheng if (fo->cookie.len > 0) 4014783237e8SYuchung Cheng fo->cookie.len = 0; 4015783237e8SYuchung Cheng err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); 4016783237e8SYuchung Cheng if (err) 4017783237e8SYuchung Cheng tp->syn_fastopen = 0; 4018783237e8SYuchung Cheng done: 4019783237e8SYuchung Cheng fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ 4020783237e8SYuchung Cheng return err; 4021783237e8SYuchung Cheng } 4022783237e8SYuchung Cheng 402367edfef7SAndi Kleen /* Build a SYN and send it off. */ 40241da177e4SLinus Torvalds int tcp_connect(struct sock *sk) 40251da177e4SLinus Torvalds { 40261da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 40271da177e4SLinus Torvalds struct sk_buff *buff; 4028ee586811SEric Paris int err; 40291da177e4SLinus Torvalds 4030de525be2SLawrence Brakmo tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB, 0, NULL); 40318ba60924SEric Dumazet 40320aadc739SDmitry Safonov #if defined(CONFIG_TCP_MD5SIG) && defined(CONFIG_TCP_AO) 40330aadc739SDmitry Safonov /* Has to be checked late, after setting daddr/saddr/ops. 40340aadc739SDmitry Safonov * Return error if the peer has both a md5 and a tcp-ao key 40350aadc739SDmitry Safonov * configured as this is ambiguous. 40360aadc739SDmitry Safonov */ 40370aadc739SDmitry Safonov if (unlikely(rcu_dereference_protected(tp->md5sig_info, 40380aadc739SDmitry Safonov lockdep_sock_is_held(sk)))) { 40390aadc739SDmitry Safonov bool needs_ao = !!tp->af_specific->ao_lookup(sk, sk, -1, -1); 40400aadc739SDmitry Safonov bool needs_md5 = !!tp->af_specific->md5_lookup(sk, sk); 40410aadc739SDmitry Safonov struct tcp_ao_info *ao_info; 40420aadc739SDmitry Safonov 40430aadc739SDmitry Safonov ao_info = rcu_dereference_check(tp->ao_info, 40440aadc739SDmitry Safonov lockdep_sock_is_held(sk)); 40450aadc739SDmitry Safonov if (ao_info) { 40460aadc739SDmitry Safonov /* This is an extra check: tcp_ao_required() in 40470aadc739SDmitry Safonov * tcp_v{4,6}_parse_md5_keys() should prevent adding 40480aadc739SDmitry Safonov * md5 keys on ao_required socket. 40490aadc739SDmitry Safonov */ 40500aadc739SDmitry Safonov needs_ao |= ao_info->ao_required; 40510aadc739SDmitry Safonov WARN_ON_ONCE(ao_info->ao_required && needs_md5); 40520aadc739SDmitry Safonov } 40530aadc739SDmitry Safonov if (needs_md5 && needs_ao) 40540aadc739SDmitry Safonov return -EKEYREJECTED; 40550aadc739SDmitry Safonov 40560aadc739SDmitry Safonov /* If we have a matching md5 key and no matching tcp-ao key 40570aadc739SDmitry Safonov * then free up ao_info if allocated. 40580aadc739SDmitry Safonov */ 40590aadc739SDmitry Safonov if (needs_md5) { 4060decde258SDmitry Safonov tcp_ao_destroy_sock(sk, false); 40610aadc739SDmitry Safonov } else if (needs_ao) { 40620aadc739SDmitry Safonov tcp_clear_md5_list(sk); 40630aadc739SDmitry Safonov kfree(rcu_replace_pointer(tp->md5sig_info, NULL, 40640aadc739SDmitry Safonov lockdep_sock_is_held(sk))); 40650aadc739SDmitry Safonov } 40660aadc739SDmitry Safonov } 40670aadc739SDmitry Safonov #endif 40680aadc739SDmitry Safonov #ifdef CONFIG_TCP_AO 40690aadc739SDmitry Safonov if (unlikely(rcu_dereference_protected(tp->ao_info, 40700aadc739SDmitry Safonov lockdep_sock_is_held(sk)))) { 40710aadc739SDmitry Safonov /* Don't allow connecting if ao is configured but no 40720aadc739SDmitry Safonov * matching key is found. 40730aadc739SDmitry Safonov */ 40740aadc739SDmitry Safonov if (!tp->af_specific->ao_lookup(sk, sk, -1, -1)) 40750aadc739SDmitry Safonov return -EKEYREJECTED; 40760aadc739SDmitry Safonov } 40770aadc739SDmitry Safonov #endif 40780aadc739SDmitry Safonov 40798ba60924SEric Dumazet if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 40808ba60924SEric Dumazet return -EHOSTUNREACH; /* Routing failure or similar. */ 40818ba60924SEric Dumazet 40821da177e4SLinus Torvalds tcp_connect_init(sk); 40831da177e4SLinus Torvalds 40842b916477SAndrey Vagin if (unlikely(tp->repair)) { 40852b916477SAndrey Vagin tcp_finish_connect(sk, NULL); 40862b916477SAndrey Vagin return 0; 40872b916477SAndrey Vagin } 40882b916477SAndrey Vagin 40895882efffSEric Dumazet buff = tcp_stream_alloc_skb(sk, sk->sk_allocation, true); 4090355a901eSEric Dumazet if (unlikely(!buff)) 40911da177e4SLinus Torvalds return -ENOBUFS; 40921da177e4SLinus Torvalds 4093a3433f35SChangli Gao tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); 40949a568de4SEric Dumazet tcp_mstamp_refresh(tp); 40959d0c00f5SEric Dumazet tp->retrans_stamp = tcp_time_stamp_ts(tp); 4096783237e8SYuchung Cheng tcp_connect_queue_skb(sk, buff); 4097735d3831SFlorian Westphal tcp_ecn_send_syn(sk, buff); 409875c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, buff); 40991da177e4SLinus Torvalds 4100783237e8SYuchung Cheng /* Send off SYN; include data in Fast Open. */ 4101783237e8SYuchung Cheng err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : 4102783237e8SYuchung Cheng tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); 4103ee586811SEric Paris if (err == -ECONNREFUSED) 4104ee586811SEric Paris return err; 4105bd37a088SWei Yongjun 4106bd37a088SWei Yongjun /* We change tp->snd_nxt after the tcp_transmit_skb() call 4107bd37a088SWei Yongjun * in order to make this packet get counted in tcpOutSegs. 4108bd37a088SWei Yongjun */ 4109e0d694d6SEric Dumazet WRITE_ONCE(tp->snd_nxt, tp->write_seq); 4110bd37a088SWei Yongjun tp->pushed_seq = tp->write_seq; 4111b5b7db8dSEric Dumazet buff = tcp_send_head(sk); 4112b5b7db8dSEric Dumazet if (unlikely(buff)) { 4113e0d694d6SEric Dumazet WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(buff)->seq); 4114b5b7db8dSEric Dumazet tp->pushed_seq = TCP_SKB_CB(buff)->seq; 4115b5b7db8dSEric Dumazet } 411681cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); 41171da177e4SLinus Torvalds 41181da177e4SLinus Torvalds /* Timer for repeating the SYN until an answer. */ 41193f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 41203f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 41211da177e4SLinus Torvalds return 0; 41221da177e4SLinus Torvalds } 41234bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_connect); 41241da177e4SLinus Torvalds 4125bbf80d71SEric Dumazet u32 tcp_delack_max(const struct sock *sk) 4126bbf80d71SEric Dumazet { 4127bbf80d71SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 4128bbf80d71SEric Dumazet u32 delack_max = inet_csk(sk)->icsk_delack_max; 4129bbf80d71SEric Dumazet 4130bbf80d71SEric Dumazet if (dst && dst_metric_locked(dst, RTAX_RTO_MIN)) { 4131bbf80d71SEric Dumazet u32 rto_min = dst_metric_rtt(dst, RTAX_RTO_MIN); 4132bbf80d71SEric Dumazet u32 delack_from_rto_min = max_t(int, 1, rto_min - 1); 4133bbf80d71SEric Dumazet 4134bbf80d71SEric Dumazet delack_max = min_t(u32, delack_max, delack_from_rto_min); 4135bbf80d71SEric Dumazet } 4136bbf80d71SEric Dumazet return delack_max; 4137bbf80d71SEric Dumazet } 4138bbf80d71SEric Dumazet 41391da177e4SLinus Torvalds /* Send out a delayed ack, the caller does the policy checking 41401da177e4SLinus Torvalds * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 41411da177e4SLinus Torvalds * for details. 41421da177e4SLinus Torvalds */ 41431da177e4SLinus Torvalds void tcp_send_delayed_ack(struct sock *sk) 41441da177e4SLinus Torvalds { 4145463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 4146463c84b9SArnaldo Carvalho de Melo int ato = icsk->icsk_ack.ato; 41471da177e4SLinus Torvalds unsigned long timeout; 41481da177e4SLinus Torvalds 41491da177e4SLinus Torvalds if (ato > TCP_DELACK_MIN) { 4150463c84b9SArnaldo Carvalho de Melo const struct tcp_sock *tp = tcp_sk(sk); 41511da177e4SLinus Torvalds int max_ato = HZ / 2; 41521da177e4SLinus Torvalds 415331954cd8SWei Wang if (inet_csk_in_pingpong_mode(sk) || 4154056834d9SIlpo Järvinen (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 41551da177e4SLinus Torvalds max_ato = TCP_DELACK_MAX; 41561da177e4SLinus Torvalds 41571da177e4SLinus Torvalds /* Slow path, intersegment interval is "high". */ 41581da177e4SLinus Torvalds 41591da177e4SLinus Torvalds /* If some rtt estimate is known, use it to bound delayed ack. 4160463c84b9SArnaldo Carvalho de Melo * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 41611da177e4SLinus Torvalds * directly. 41621da177e4SLinus Torvalds */ 4163740b0f18SEric Dumazet if (tp->srtt_us) { 4164740b0f18SEric Dumazet int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3), 4165740b0f18SEric Dumazet TCP_DELACK_MIN); 41661da177e4SLinus Torvalds 41671da177e4SLinus Torvalds if (rtt < max_ato) 41681da177e4SLinus Torvalds max_ato = rtt; 41691da177e4SLinus Torvalds } 41701da177e4SLinus Torvalds 41711da177e4SLinus Torvalds ato = min(ato, max_ato); 41721da177e4SLinus Torvalds } 41731da177e4SLinus Torvalds 4174bbf80d71SEric Dumazet ato = min_t(u32, ato, tcp_delack_max(sk)); 41752b8ee4f0SMartin KaFai Lau 41761da177e4SLinus Torvalds /* Stay within the limit we were given */ 41771da177e4SLinus Torvalds timeout = jiffies + ato; 41781da177e4SLinus Torvalds 41791da177e4SLinus Torvalds /* Use new timeout only if there wasn't a older one earlier. */ 4180463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 4181b6b6d653SEric Dumazet /* If delack timer is about to expire, send ACK now. */ 4182b6b6d653SEric Dumazet if (time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 41831da177e4SLinus Torvalds tcp_send_ack(sk); 41841da177e4SLinus Torvalds return; 41851da177e4SLinus Torvalds } 41861da177e4SLinus Torvalds 4187463c84b9SArnaldo Carvalho de Melo if (!time_before(timeout, icsk->icsk_ack.timeout)) 4188463c84b9SArnaldo Carvalho de Melo timeout = icsk->icsk_ack.timeout; 41891da177e4SLinus Torvalds } 4190463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 4191463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.timeout = timeout; 4192463c84b9SArnaldo Carvalho de Melo sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 41931da177e4SLinus Torvalds } 41941da177e4SLinus Torvalds 41951da177e4SLinus Torvalds /* This routine sends an ack and also updates the window. */ 41962987babbSYuchung Cheng void __tcp_send_ack(struct sock *sk, u32 rcv_nxt) 41971da177e4SLinus Torvalds { 41981da177e4SLinus Torvalds struct sk_buff *buff; 41991da177e4SLinus Torvalds 4200058dc334SIlpo Järvinen /* If we have been reset, we may not send again. */ 4201058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 4202058dc334SIlpo Järvinen return; 4203058dc334SIlpo Järvinen 42041da177e4SLinus Torvalds /* We are not putting this on the write queue, so 42051da177e4SLinus Torvalds * tcp_transmit_skb() will set the ownership to this 42061da177e4SLinus Torvalds * sock. 42071da177e4SLinus Torvalds */ 42087450aaf6SEric Dumazet buff = alloc_skb(MAX_TCP_HEADER, 42097450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 42107450aaf6SEric Dumazet if (unlikely(!buff)) { 4211a37c2134SEric Dumazet struct inet_connection_sock *icsk = inet_csk(sk); 4212a37c2134SEric Dumazet unsigned long delay; 4213a37c2134SEric Dumazet 4214a37c2134SEric Dumazet delay = TCP_DELACK_MAX << icsk->icsk_ack.retry; 4215a37c2134SEric Dumazet if (delay < TCP_RTO_MAX) 4216a37c2134SEric Dumazet icsk->icsk_ack.retry++; 4217463c84b9SArnaldo Carvalho de Melo inet_csk_schedule_ack(sk); 4218a37c2134SEric Dumazet icsk->icsk_ack.ato = TCP_ATO_MIN; 4219a37c2134SEric Dumazet inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, delay, TCP_RTO_MAX); 42201da177e4SLinus Torvalds return; 42211da177e4SLinus Torvalds } 42221da177e4SLinus Torvalds 42231da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 42241da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 4225a3433f35SChangli Gao tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); 42261da177e4SLinus Torvalds 422798781965SEric Dumazet /* We do not want pure acks influencing TCP Small Queues or fq/pacing 422898781965SEric Dumazet * too much. 422998781965SEric Dumazet * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784 423098781965SEric Dumazet */ 423198781965SEric Dumazet skb_set_tcp_pure_ack(buff); 423298781965SEric Dumazet 42331da177e4SLinus Torvalds /* Send it off, this clears delayed acks for us. */ 42342987babbSYuchung Cheng __tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0, rcv_nxt); 42351da177e4SLinus Torvalds } 423627cde44aSYuchung Cheng EXPORT_SYMBOL_GPL(__tcp_send_ack); 42372987babbSYuchung Cheng 42382987babbSYuchung Cheng void tcp_send_ack(struct sock *sk) 42392987babbSYuchung Cheng { 42402987babbSYuchung Cheng __tcp_send_ack(sk, tcp_sk(sk)->rcv_nxt); 42411da177e4SLinus Torvalds } 42421da177e4SLinus Torvalds 42431da177e4SLinus Torvalds /* This routine sends a packet with an out of date sequence 42441da177e4SLinus Torvalds * number. It assumes the other end will try to ack it. 42451da177e4SLinus Torvalds * 42461da177e4SLinus Torvalds * Question: what should we make while urgent mode? 42471da177e4SLinus Torvalds * 4.4BSD forces sending single byte of data. We cannot send 42481da177e4SLinus Torvalds * out of window data, because we have SND.NXT==SND.MAX... 42491da177e4SLinus Torvalds * 42501da177e4SLinus Torvalds * Current solution: to send TWO zero-length segments in urgent mode: 42511da177e4SLinus Torvalds * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 42521da177e4SLinus Torvalds * out-of-date with SND.UNA-1 to probe window. 42531da177e4SLinus Torvalds */ 4254e520af48SEric Dumazet static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib) 42551da177e4SLinus Torvalds { 42561da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 42571da177e4SLinus Torvalds struct sk_buff *skb; 42581da177e4SLinus Torvalds 42591da177e4SLinus Torvalds /* We don't queue it, tcp_transmit_skb() sets ownership. */ 42607450aaf6SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, 42617450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 426251456b29SIan Morris if (!skb) 42631da177e4SLinus Torvalds return -1; 42641da177e4SLinus Torvalds 42651da177e4SLinus Torvalds /* Reserve space for headers and set control bits. */ 42661da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 42671da177e4SLinus Torvalds /* Use a previous sequence. This should cause the other 42681da177e4SLinus Torvalds * end to send an ack. Don't queue or clone SKB, just 42691da177e4SLinus Torvalds * send it. 42701da177e4SLinus Torvalds */ 4271a3433f35SChangli Gao tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); 4272e2e8009fSRenato Westphal NET_INC_STATS(sock_net(sk), mib); 42737450aaf6SEric Dumazet return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0); 42741da177e4SLinus Torvalds } 42751da177e4SLinus Torvalds 4276385e2070SEric Dumazet /* Called from setsockopt( ... TCP_REPAIR ) */ 4277ee995283SPavel Emelyanov void tcp_send_window_probe(struct sock *sk) 4278ee995283SPavel Emelyanov { 4279ee995283SPavel Emelyanov if (sk->sk_state == TCP_ESTABLISHED) { 4280ee995283SPavel Emelyanov tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; 42819a568de4SEric Dumazet tcp_mstamp_refresh(tcp_sk(sk)); 4282e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE); 4283ee995283SPavel Emelyanov } 4284ee995283SPavel Emelyanov } 4285ee995283SPavel Emelyanov 428667edfef7SAndi Kleen /* Initiate keepalive or window probe from timer. */ 4287e520af48SEric Dumazet int tcp_write_wakeup(struct sock *sk, int mib) 42881da177e4SLinus Torvalds { 42891da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 42901da177e4SLinus Torvalds struct sk_buff *skb; 42911da177e4SLinus Torvalds 4292058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 4293058dc334SIlpo Järvinen return -1; 4294058dc334SIlpo Järvinen 429500db4124SIan Morris skb = tcp_send_head(sk); 429600db4124SIan Morris if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { 42971da177e4SLinus Torvalds int err; 42980c54b85fSIlpo Järvinen unsigned int mss = tcp_current_mss(sk); 429990840defSIlpo Järvinen unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 43001da177e4SLinus Torvalds 43011da177e4SLinus Torvalds if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 43021da177e4SLinus Torvalds tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 43031da177e4SLinus Torvalds 43041da177e4SLinus Torvalds /* We are probing the opening of a window 43051da177e4SLinus Torvalds * but the window size is != 0 43061da177e4SLinus Torvalds * must have been a result SWS avoidance ( sender ) 43071da177e4SLinus Torvalds */ 43081da177e4SLinus Torvalds if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 43091da177e4SLinus Torvalds skb->len > mss) { 43101da177e4SLinus Torvalds seg_size = min(seg_size, mss); 43114de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 431275c119afSEric Dumazet if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE, 431375c119afSEric Dumazet skb, seg_size, mss, GFP_ATOMIC)) 43141da177e4SLinus Torvalds return -1; 43151da177e4SLinus Torvalds } else if (!tcp_skb_pcount(skb)) 43165bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss); 43171da177e4SLinus Torvalds 43184de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 4319dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 432066f5fe62SIlpo Järvinen if (!err) 432166f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 43221da177e4SLinus Torvalds return err; 43231da177e4SLinus Torvalds } else { 432433f5f57eSIlpo Järvinen if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) 4325e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 1, mib); 4326e520af48SEric Dumazet return tcp_xmit_probe_skb(sk, 0, mib); 43271da177e4SLinus Torvalds } 43281da177e4SLinus Torvalds } 43291da177e4SLinus Torvalds 43301da177e4SLinus Torvalds /* A window probe timeout has occurred. If window is not closed send 43311da177e4SLinus Torvalds * a partial packet else a zero probe. 43321da177e4SLinus Torvalds */ 43331da177e4SLinus Torvalds void tcp_send_probe0(struct sock *sk) 43341da177e4SLinus Torvalds { 4335463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 43361da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 4337c6214a97SNikolay Borisov struct net *net = sock_net(sk); 4338c1d5674fSYuchung Cheng unsigned long timeout; 43391da177e4SLinus Torvalds int err; 43401da177e4SLinus Torvalds 4341e520af48SEric Dumazet err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE); 43421da177e4SLinus Torvalds 434375c119afSEric Dumazet if (tp->packets_out || tcp_write_queue_empty(sk)) { 43441da177e4SLinus Torvalds /* Cancel probe timer, if it is not required. */ 43456687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 0; 4346463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff = 0; 43479d9b1ee0SEnke Chen icsk->icsk_probes_tstamp = 0; 43481da177e4SLinus Torvalds return; 43491da177e4SLinus Torvalds } 43501da177e4SLinus Torvalds 4351c1d5674fSYuchung Cheng icsk->icsk_probes_out++; 43521da177e4SLinus Torvalds if (err <= 0) { 435339e24435SKuniyuki Iwashima if (icsk->icsk_backoff < READ_ONCE(net->ipv4.sysctl_tcp_retries2)) 4354463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff++; 4355c1d5674fSYuchung Cheng timeout = tcp_probe0_when(sk, TCP_RTO_MAX); 43561da177e4SLinus Torvalds } else { 43571da177e4SLinus Torvalds /* If packet was not sent due to local congestion, 4358c1d5674fSYuchung Cheng * Let senders fight for local resources conservatively. 43591da177e4SLinus Torvalds */ 4360c1d5674fSYuchung Cheng timeout = TCP_RESOURCE_PROBE_INTERVAL; 43611da177e4SLinus Torvalds } 4362344db93aSEnke Chen 4363344db93aSEnke Chen timeout = tcp_clamp_probe0_to_user_timeout(sk, timeout); 43648dc242adSEric Dumazet tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, timeout, TCP_RTO_MAX); 43651da177e4SLinus Torvalds } 43665db92c99SOctavian Purdila 4367ea3bea3aSEric Dumazet int tcp_rtx_synack(const struct sock *sk, struct request_sock *req) 43685db92c99SOctavian Purdila { 43695db92c99SOctavian Purdila const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific; 43705db92c99SOctavian Purdila struct flowi fl; 43715db92c99SOctavian Purdila int res; 43725db92c99SOctavian Purdila 4373cb6cd2ceSAkhmat Karakotov /* Paired with WRITE_ONCE() in sock_setsockopt() */ 4374cb6cd2ceSAkhmat Karakotov if (READ_ONCE(sk->sk_txrehash) == SOCK_TXREHASH_ENABLED) 43755e526552SEric Dumazet WRITE_ONCE(tcp_rsk(req)->txhash, net_tx_rndhash()); 4376331fca43SMartin KaFai Lau res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL, 4377331fca43SMartin KaFai Lau NULL); 43785db92c99SOctavian Purdila if (!res) { 43790a375c82SEric Dumazet TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); 43800a375c82SEric Dumazet NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 4381e9d9da91SEric Dumazet if (unlikely(tcp_passive_fastopen(sk))) { 4382e9d9da91SEric Dumazet /* sk has const attribute because listeners are lockless. 4383e9d9da91SEric Dumazet * However in this case, we are dealing with a passive fastopen 4384e9d9da91SEric Dumazet * socket thus we can change total_retrans value. 4385e9d9da91SEric Dumazet */ 4386e9d9da91SEric Dumazet tcp_sk_rw(sk)->total_retrans++; 4387e9d9da91SEric Dumazet } 4388cf34ce3dSSong Liu trace_tcp_retransmit_synack(sk, req); 43895db92c99SOctavian Purdila } 43905db92c99SOctavian Purdila return res; 43915db92c99SOctavian Purdila } 43925db92c99SOctavian Purdila EXPORT_SYMBOL(tcp_rtx_synack); 4393