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 1734d8f24eeSWei Wang * packet, enter pingpong mode. 1744d8f24eeSWei Wang */ 1754d8f24eeSWei Wang if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato) 1764d8f24eeSWei Wang inet_csk_enter_pingpong_mode(sk); 1771da177e4SLinus Torvalds } 1781da177e4SLinus Torvalds 17967edfef7SAndi Kleen /* Account for an ACK we sent. */ 18027cde44aSYuchung Cheng static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts, 18127cde44aSYuchung Cheng u32 rcv_nxt) 1821da177e4SLinus Torvalds { 1835d9f4262SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 1845d9f4262SEric Dumazet 1852b195850SEric Dumazet if (unlikely(tp->compressed_ack)) { 186200d95f4SEric Dumazet NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED, 1872b195850SEric Dumazet tp->compressed_ack); 1882b195850SEric Dumazet tp->compressed_ack = 0; 1895d9f4262SEric Dumazet if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1) 1905d9f4262SEric Dumazet __sock_put(sk); 1915d9f4262SEric Dumazet } 19227cde44aSYuchung Cheng 19327cde44aSYuchung Cheng if (unlikely(rcv_nxt != tp->rcv_nxt)) 19427cde44aSYuchung Cheng return; /* Special ACK sent by DCTCP to reflect ECN */ 195463c84b9SArnaldo Carvalho de Melo tcp_dec_quickack_mode(sk, pkts); 196463c84b9SArnaldo Carvalho de Melo inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); 1971da177e4SLinus Torvalds } 1981da177e4SLinus Torvalds 1991da177e4SLinus Torvalds /* Determine a window scaling and initial window to offer. 2001da177e4SLinus Torvalds * Based on the assumption that the given amount of space 2011da177e4SLinus Torvalds * will be offered. Store the results in the tp structure. 2021da177e4SLinus Torvalds * NOTE: for smooth operation initial space offering should 2031da177e4SLinus Torvalds * be a multiple of mss if possible. We assume here that mss >= 1. 2041da177e4SLinus Torvalds * This MUST be enforced by all callers. 2051da177e4SLinus Torvalds */ 206ceef9ab6SEric Dumazet void tcp_select_initial_window(const struct sock *sk, int __space, __u32 mss, 2071da177e4SLinus Torvalds __u32 *rcv_wnd, __u32 *window_clamp, 20831d12926Slaurent chavey int wscale_ok, __u8 *rcv_wscale, 20931d12926Slaurent chavey __u32 init_rcv_wnd) 2101da177e4SLinus Torvalds { 2111da177e4SLinus Torvalds unsigned int space = (__space < 0 ? 0 : __space); 2121da177e4SLinus Torvalds 2131da177e4SLinus Torvalds /* If no clamp set the clamp to the max possible scaled window */ 2141da177e4SLinus Torvalds if (*window_clamp == 0) 215589c49cbSGao Feng (*window_clamp) = (U16_MAX << TCP_MAX_WSCALE); 2161da177e4SLinus Torvalds 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 */ 241*02739545SKuniyuki Iwashima space = max_t(u32, space, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2])); 242f626300aSSoheil Hassas Yeganeh space = max_t(u32, space, sysctl_rmem_max); 243316c1592SStephen Hemminger 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 */ 248589c49cbSGao Feng (*window_clamp) = min_t(__u32, U16_MAX << (*rcv_wscale), *window_clamp); 2491da177e4SLinus Torvalds } 2504bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_select_initial_window); 2511da177e4SLinus Torvalds 2521da177e4SLinus Torvalds /* Chose a new window to advertise, update state in tcp_sock for the 2531da177e4SLinus Torvalds * socket, and return result with RFC1323 scaling applied. The return 2541da177e4SLinus Torvalds * value can be stuffed directly into th->window for an outgoing 2551da177e4SLinus Torvalds * frame. 2561da177e4SLinus Torvalds */ 25740efc6faSStephen Hemminger static u16 tcp_select_window(struct sock *sk) 2581da177e4SLinus Torvalds { 2591da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2608e165e20SFlorian Westphal u32 old_win = tp->rcv_wnd; 2611da177e4SLinus Torvalds u32 cur_win = tcp_receive_window(tp); 2621da177e4SLinus Torvalds u32 new_win = __tcp_select_window(sk); 2631da177e4SLinus Torvalds 2641da177e4SLinus Torvalds /* Never shrink the offered window */ 2651da177e4SLinus Torvalds if (new_win < cur_win) { 2661da177e4SLinus Torvalds /* Danger Will Robinson! 2671da177e4SLinus Torvalds * Don't update rcv_wup/rcv_wnd here or else 2681da177e4SLinus Torvalds * we will not be able to advertise a zero 2691da177e4SLinus Torvalds * window in time. --DaveM 2701da177e4SLinus Torvalds * 2711da177e4SLinus Torvalds * Relax Will Robinson. 2721da177e4SLinus Torvalds */ 2738e165e20SFlorian Westphal if (new_win == 0) 2748e165e20SFlorian Westphal NET_INC_STATS(sock_net(sk), 2758e165e20SFlorian Westphal LINUX_MIB_TCPWANTZEROWINDOWADV); 276607bfbf2SPatrick McHardy new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale); 2771da177e4SLinus Torvalds } 2781da177e4SLinus Torvalds tp->rcv_wnd = new_win; 2791da177e4SLinus Torvalds tp->rcv_wup = tp->rcv_nxt; 2801da177e4SLinus Torvalds 2811da177e4SLinus Torvalds /* Make sure we do not exceed the maximum possible 2821da177e4SLinus Torvalds * scaled window. 2831da177e4SLinus Torvalds */ 284ceef9ab6SEric Dumazet if (!tp->rx_opt.rcv_wscale && 2850f1e4d06SKuniyuki Iwashima READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows)) 2861da177e4SLinus Torvalds new_win = min(new_win, MAX_TCP_WINDOW); 2871da177e4SLinus Torvalds else 2881da177e4SLinus Torvalds new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); 2891da177e4SLinus Torvalds 2901da177e4SLinus Torvalds /* RFC1323 scaling applied */ 2911da177e4SLinus Torvalds new_win >>= tp->rx_opt.rcv_wscale; 2921da177e4SLinus Torvalds 29331770e34SFlorian Westphal /* If we advertise zero window, disable fast path. */ 2948e165e20SFlorian Westphal if (new_win == 0) { 29531770e34SFlorian Westphal tp->pred_flags = 0; 2968e165e20SFlorian Westphal if (old_win) 2978e165e20SFlorian Westphal NET_INC_STATS(sock_net(sk), 2988e165e20SFlorian Westphal LINUX_MIB_TCPTOZEROWINDOWADV); 2998e165e20SFlorian Westphal } else if (old_win == 0) { 3008e165e20SFlorian Westphal NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV); 3018e165e20SFlorian Westphal } 3021da177e4SLinus Torvalds 3031da177e4SLinus Torvalds return new_win; 3041da177e4SLinus Torvalds } 3051da177e4SLinus Torvalds 30667edfef7SAndi Kleen /* Packet ECN state for a SYN-ACK */ 307735d3831SFlorian Westphal static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb) 308bdf1ee5dSIlpo Järvinen { 30930e502a3SDaniel Borkmann const struct tcp_sock *tp = tcp_sk(sk); 31030e502a3SDaniel Borkmann 3114de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR; 312bdf1ee5dSIlpo Järvinen if (!(tp->ecn_flags & TCP_ECN_OK)) 3134de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE; 31491b5b21cSLawrence Brakmo else if (tcp_ca_needs_ecn(sk) || 31591b5b21cSLawrence Brakmo tcp_bpf_ca_needs_ecn(sk)) 31630e502a3SDaniel Borkmann INET_ECN_xmit(sk); 317bdf1ee5dSIlpo Järvinen } 318bdf1ee5dSIlpo Järvinen 31967edfef7SAndi Kleen /* Packet ECN state for a SYN. */ 320735d3831SFlorian Westphal static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb) 321bdf1ee5dSIlpo Järvinen { 322bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 32391b5b21cSLawrence Brakmo bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk); 3244785a667SKuniyuki Iwashima bool use_ecn = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_ecn) == 1 || 32591b5b21cSLawrence Brakmo tcp_ca_needs_ecn(sk) || bpf_needs_ecn; 326f7b3bec6SFlorian Westphal 327f7b3bec6SFlorian Westphal if (!use_ecn) { 328f7b3bec6SFlorian Westphal const struct dst_entry *dst = __sk_dst_get(sk); 329f7b3bec6SFlorian Westphal 330f7b3bec6SFlorian Westphal if (dst && dst_feature(dst, RTAX_FEATURE_ECN)) 331f7b3bec6SFlorian Westphal use_ecn = true; 332f7b3bec6SFlorian Westphal } 333bdf1ee5dSIlpo Järvinen 334bdf1ee5dSIlpo Järvinen tp->ecn_flags = 0; 335f7b3bec6SFlorian Westphal 336f7b3bec6SFlorian Westphal if (use_ecn) { 3374de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR; 338bdf1ee5dSIlpo Järvinen tp->ecn_flags = TCP_ECN_OK; 33991b5b21cSLawrence Brakmo if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn) 34030e502a3SDaniel Borkmann INET_ECN_xmit(sk); 341bdf1ee5dSIlpo Järvinen } 342bdf1ee5dSIlpo Järvinen } 343bdf1ee5dSIlpo Järvinen 34449213555SDaniel Borkmann static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb) 34549213555SDaniel Borkmann { 34612b8d9caSKuniyuki Iwashima if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback)) 34749213555SDaniel Borkmann /* tp->ecn_flags are cleared at a later point in time when 34849213555SDaniel Borkmann * SYN ACK is ultimatively being received. 34949213555SDaniel Borkmann */ 35049213555SDaniel Borkmann TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR); 35149213555SDaniel Borkmann } 35249213555SDaniel Borkmann 353735d3831SFlorian Westphal static void 3546ac705b1SEric Dumazet tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th) 355bdf1ee5dSIlpo Järvinen { 3566ac705b1SEric Dumazet if (inet_rsk(req)->ecn_ok) 357bdf1ee5dSIlpo Järvinen th->ece = 1; 358bdf1ee5dSIlpo Järvinen } 359bdf1ee5dSIlpo Järvinen 36067edfef7SAndi Kleen /* Set up ECN state for a packet on a ESTABLISHED socket that is about to 36167edfef7SAndi Kleen * be sent. 36267edfef7SAndi Kleen */ 363735d3831SFlorian Westphal static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb, 364ea1627c2SEric Dumazet struct tcphdr *th, int tcp_header_len) 365bdf1ee5dSIlpo Järvinen { 366bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 367bdf1ee5dSIlpo Järvinen 368bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_OK) { 369bdf1ee5dSIlpo Järvinen /* Not-retransmitted data segment: set ECT and inject CWR. */ 370bdf1ee5dSIlpo Järvinen if (skb->len != tcp_header_len && 371bdf1ee5dSIlpo Järvinen !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { 372bdf1ee5dSIlpo Järvinen INET_ECN_xmit(sk); 373bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) { 374bdf1ee5dSIlpo Järvinen tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; 375ea1627c2SEric Dumazet th->cwr = 1; 376bdf1ee5dSIlpo Järvinen skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; 377bdf1ee5dSIlpo Järvinen } 37830e502a3SDaniel Borkmann } else if (!tcp_ca_needs_ecn(sk)) { 379bdf1ee5dSIlpo Järvinen /* ACK or retransmitted segment: clear ECT|CE */ 380bdf1ee5dSIlpo Järvinen INET_ECN_dontxmit(sk); 381bdf1ee5dSIlpo Järvinen } 382bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) 383ea1627c2SEric Dumazet th->ece = 1; 384bdf1ee5dSIlpo Järvinen } 385bdf1ee5dSIlpo Järvinen } 386bdf1ee5dSIlpo Järvinen 387e870a8efSIlpo Järvinen /* Constructs common control bits of non-data skb. If SYN/FIN is present, 388e870a8efSIlpo Järvinen * auto increment end seqno. 389e870a8efSIlpo Järvinen */ 390e870a8efSIlpo Järvinen static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags) 391e870a8efSIlpo Järvinen { 3922e8e18efSDavid S. Miller skb->ip_summed = CHECKSUM_PARTIAL; 393e870a8efSIlpo Järvinen 3944de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags; 395e870a8efSIlpo Järvinen 396cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, 1); 397e870a8efSIlpo Järvinen 398e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->seq = seq; 399a3433f35SChangli Gao if (flags & (TCPHDR_SYN | TCPHDR_FIN)) 400e870a8efSIlpo Järvinen seq++; 401e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->end_seq = seq; 402e870a8efSIlpo Järvinen } 403e870a8efSIlpo Järvinen 404a2a385d6SEric Dumazet static inline bool tcp_urg_mode(const struct tcp_sock *tp) 40533f5f57eSIlpo Järvinen { 40633f5f57eSIlpo Järvinen return tp->snd_una != tp->snd_up; 40733f5f57eSIlpo Järvinen } 40833f5f57eSIlpo Järvinen 4093b65abb8SLeonard Crestez #define OPTION_SACK_ADVERTISE BIT(0) 4103b65abb8SLeonard Crestez #define OPTION_TS BIT(1) 4113b65abb8SLeonard Crestez #define OPTION_MD5 BIT(2) 4123b65abb8SLeonard Crestez #define OPTION_WSCALE BIT(3) 4133b65abb8SLeonard Crestez #define OPTION_FAST_OPEN_COOKIE BIT(8) 4143b65abb8SLeonard Crestez #define OPTION_SMC BIT(9) 4153b65abb8SLeonard Crestez #define OPTION_MPTCP BIT(10) 41660e2a778SUrsula Braun 41760e2a778SUrsula Braun static void smc_options_write(__be32 *ptr, u16 *options) 41860e2a778SUrsula Braun { 41960e2a778SUrsula Braun #if IS_ENABLED(CONFIG_SMC) 42060e2a778SUrsula Braun if (static_branch_unlikely(&tcp_have_smc)) { 42160e2a778SUrsula Braun if (unlikely(OPTION_SMC & *options)) { 42260e2a778SUrsula Braun *ptr++ = htonl((TCPOPT_NOP << 24) | 42360e2a778SUrsula Braun (TCPOPT_NOP << 16) | 42460e2a778SUrsula Braun (TCPOPT_EXP << 8) | 42560e2a778SUrsula Braun (TCPOLEN_EXP_SMC_BASE)); 42660e2a778SUrsula Braun *ptr++ = htonl(TCPOPT_SMC_MAGIC); 42760e2a778SUrsula Braun } 42860e2a778SUrsula Braun } 42960e2a778SUrsula Braun #endif 43060e2a778SUrsula Braun } 43133ad798cSAdam Langley 43233ad798cSAdam Langley struct tcp_out_options { 4332100c8d2SYuchung Cheng u16 options; /* bit field of OPTION_* */ 4342100c8d2SYuchung Cheng u16 mss; /* 0 to disable */ 43533ad798cSAdam Langley u8 ws; /* window scale, 0 to disable */ 43633ad798cSAdam Langley u8 num_sack_blocks; /* number of SACK blocks to include */ 437bd0388aeSWilliam Allen Simpson u8 hash_size; /* bytes in hash_location */ 438331fca43SMartin KaFai Lau u8 bpf_opt_len; /* length of BPF hdr option */ 439bd0388aeSWilliam Allen Simpson __u8 *hash_location; /* temporary pointer, overloaded */ 4402100c8d2SYuchung Cheng __u32 tsval, tsecr; /* need to include OPTION_TS */ 4412100c8d2SYuchung Cheng struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */ 442eda7acddSPeter Krystad struct mptcp_out_options mptcp; 44333ad798cSAdam Langley }; 44433ad798cSAdam Langley 445ea66758cSPaolo Abeni static void mptcp_options_write(struct tcphdr *th, __be32 *ptr, 446ea66758cSPaolo Abeni struct tcp_sock *tp, 447fa3fe2b1SFlorian Westphal struct tcp_out_options *opts) 448eda7acddSPeter Krystad { 449eda7acddSPeter Krystad #if IS_ENABLED(CONFIG_MPTCP) 450eda7acddSPeter Krystad if (unlikely(OPTION_MPTCP & opts->options)) 451ea66758cSPaolo Abeni mptcp_write_options(th, ptr, tp, &opts->mptcp); 452eda7acddSPeter Krystad #endif 453eda7acddSPeter Krystad } 454eda7acddSPeter Krystad 455331fca43SMartin KaFai Lau #ifdef CONFIG_CGROUP_BPF 4560813a841SMartin KaFai Lau static int bpf_skops_write_hdr_opt_arg0(struct sk_buff *skb, 4570813a841SMartin KaFai Lau enum tcp_synack_type synack_type) 4580813a841SMartin KaFai Lau { 4590813a841SMartin KaFai Lau if (unlikely(!skb)) 4600813a841SMartin KaFai Lau return BPF_WRITE_HDR_TCP_CURRENT_MSS; 4610813a841SMartin KaFai Lau 4620813a841SMartin KaFai Lau if (unlikely(synack_type == TCP_SYNACK_COOKIE)) 4630813a841SMartin KaFai Lau return BPF_WRITE_HDR_TCP_SYNACK_COOKIE; 4640813a841SMartin KaFai Lau 4650813a841SMartin KaFai Lau return 0; 4660813a841SMartin KaFai Lau } 4670813a841SMartin KaFai Lau 468331fca43SMartin KaFai Lau /* req, syn_skb and synack_type are used when writing synack */ 469331fca43SMartin KaFai Lau static void bpf_skops_hdr_opt_len(struct sock *sk, struct sk_buff *skb, 470331fca43SMartin KaFai Lau struct request_sock *req, 471331fca43SMartin KaFai Lau struct sk_buff *syn_skb, 472331fca43SMartin KaFai Lau enum tcp_synack_type synack_type, 473331fca43SMartin KaFai Lau struct tcp_out_options *opts, 474331fca43SMartin KaFai Lau unsigned int *remaining) 475331fca43SMartin KaFai Lau { 4760813a841SMartin KaFai Lau struct bpf_sock_ops_kern sock_ops; 4770813a841SMartin KaFai Lau int err; 4780813a841SMartin KaFai Lau 479331fca43SMartin KaFai Lau if (likely(!BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), 480331fca43SMartin KaFai Lau BPF_SOCK_OPS_WRITE_HDR_OPT_CB_FLAG)) || 481331fca43SMartin KaFai Lau !*remaining) 482331fca43SMartin KaFai Lau return; 483331fca43SMartin KaFai Lau 4840813a841SMartin KaFai Lau /* *remaining has already been aligned to 4 bytes, so *remaining >= 4 */ 4850813a841SMartin KaFai Lau 4860813a841SMartin KaFai Lau /* init sock_ops */ 4870813a841SMartin KaFai Lau memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp)); 4880813a841SMartin KaFai Lau 4890813a841SMartin KaFai Lau sock_ops.op = BPF_SOCK_OPS_HDR_OPT_LEN_CB; 4900813a841SMartin KaFai Lau 4910813a841SMartin KaFai Lau if (req) { 4920813a841SMartin KaFai Lau /* The listen "sk" cannot be passed here because 4930813a841SMartin KaFai Lau * it is not locked. It would not make too much 4940813a841SMartin KaFai Lau * sense to do bpf_setsockopt(listen_sk) based 4950813a841SMartin KaFai Lau * on individual connection request also. 4960813a841SMartin KaFai Lau * 4970813a841SMartin KaFai Lau * Thus, "req" is passed here and the cgroup-bpf-progs 4980813a841SMartin KaFai Lau * of the listen "sk" will be run. 4990813a841SMartin KaFai Lau * 5000813a841SMartin KaFai Lau * "req" is also used here for fastopen even the "sk" here is 5010813a841SMartin KaFai Lau * a fullsock "child" sk. It is to keep the behavior 5020813a841SMartin KaFai Lau * consistent between fastopen and non-fastopen on 5030813a841SMartin KaFai Lau * the bpf programming side. 504331fca43SMartin KaFai Lau */ 5050813a841SMartin KaFai Lau sock_ops.sk = (struct sock *)req; 5060813a841SMartin KaFai Lau sock_ops.syn_skb = syn_skb; 5070813a841SMartin KaFai Lau } else { 5080813a841SMartin KaFai Lau sock_owned_by_me(sk); 5090813a841SMartin KaFai Lau 5100813a841SMartin KaFai Lau sock_ops.is_fullsock = 1; 5110813a841SMartin KaFai Lau sock_ops.sk = sk; 5120813a841SMartin KaFai Lau } 5130813a841SMartin KaFai Lau 5140813a841SMartin KaFai Lau sock_ops.args[0] = bpf_skops_write_hdr_opt_arg0(skb, synack_type); 5150813a841SMartin KaFai Lau sock_ops.remaining_opt_len = *remaining; 5160813a841SMartin KaFai Lau /* tcp_current_mss() does not pass a skb */ 5170813a841SMartin KaFai Lau if (skb) 5180813a841SMartin KaFai Lau bpf_skops_init_skb(&sock_ops, skb, 0); 5190813a841SMartin KaFai Lau 5200813a841SMartin KaFai Lau err = BPF_CGROUP_RUN_PROG_SOCK_OPS_SK(&sock_ops, sk); 5210813a841SMartin KaFai Lau 5220813a841SMartin KaFai Lau if (err || sock_ops.remaining_opt_len == *remaining) 5230813a841SMartin KaFai Lau return; 5240813a841SMartin KaFai Lau 5250813a841SMartin KaFai Lau opts->bpf_opt_len = *remaining - sock_ops.remaining_opt_len; 5260813a841SMartin KaFai Lau /* round up to 4 bytes */ 5270813a841SMartin KaFai Lau opts->bpf_opt_len = (opts->bpf_opt_len + 3) & ~3; 5280813a841SMartin KaFai Lau 5290813a841SMartin KaFai Lau *remaining -= opts->bpf_opt_len; 530331fca43SMartin KaFai Lau } 531331fca43SMartin KaFai Lau 532331fca43SMartin KaFai Lau static void bpf_skops_write_hdr_opt(struct sock *sk, struct sk_buff *skb, 533331fca43SMartin KaFai Lau struct request_sock *req, 534331fca43SMartin KaFai Lau struct sk_buff *syn_skb, 535331fca43SMartin KaFai Lau enum tcp_synack_type synack_type, 536331fca43SMartin KaFai Lau struct tcp_out_options *opts) 537331fca43SMartin KaFai Lau { 5380813a841SMartin KaFai Lau u8 first_opt_off, nr_written, max_opt_len = opts->bpf_opt_len; 5390813a841SMartin KaFai Lau struct bpf_sock_ops_kern sock_ops; 5400813a841SMartin KaFai Lau int err; 5410813a841SMartin KaFai Lau 5420813a841SMartin KaFai Lau if (likely(!max_opt_len)) 543331fca43SMartin KaFai Lau return; 544331fca43SMartin KaFai Lau 5450813a841SMartin KaFai Lau memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp)); 5460813a841SMartin KaFai Lau 5470813a841SMartin KaFai Lau sock_ops.op = BPF_SOCK_OPS_WRITE_HDR_OPT_CB; 5480813a841SMartin KaFai Lau 5490813a841SMartin KaFai Lau if (req) { 5500813a841SMartin KaFai Lau sock_ops.sk = (struct sock *)req; 5510813a841SMartin KaFai Lau sock_ops.syn_skb = syn_skb; 5520813a841SMartin KaFai Lau } else { 5530813a841SMartin KaFai Lau sock_owned_by_me(sk); 5540813a841SMartin KaFai Lau 5550813a841SMartin KaFai Lau sock_ops.is_fullsock = 1; 5560813a841SMartin KaFai Lau sock_ops.sk = sk; 5570813a841SMartin KaFai Lau } 5580813a841SMartin KaFai Lau 5590813a841SMartin KaFai Lau sock_ops.args[0] = bpf_skops_write_hdr_opt_arg0(skb, synack_type); 5600813a841SMartin KaFai Lau sock_ops.remaining_opt_len = max_opt_len; 5610813a841SMartin KaFai Lau first_opt_off = tcp_hdrlen(skb) - max_opt_len; 5620813a841SMartin KaFai Lau bpf_skops_init_skb(&sock_ops, skb, first_opt_off); 5630813a841SMartin KaFai Lau 5640813a841SMartin KaFai Lau err = BPF_CGROUP_RUN_PROG_SOCK_OPS_SK(&sock_ops, sk); 5650813a841SMartin KaFai Lau 5660813a841SMartin KaFai Lau if (err) 5670813a841SMartin KaFai Lau nr_written = 0; 5680813a841SMartin KaFai Lau else 5690813a841SMartin KaFai Lau nr_written = max_opt_len - sock_ops.remaining_opt_len; 5700813a841SMartin KaFai Lau 5710813a841SMartin KaFai Lau if (nr_written < max_opt_len) 5720813a841SMartin KaFai Lau memset(skb->data + first_opt_off + nr_written, TCPOPT_NOP, 5730813a841SMartin KaFai Lau max_opt_len - nr_written); 574331fca43SMartin KaFai Lau } 575331fca43SMartin KaFai Lau #else 576331fca43SMartin KaFai Lau static void bpf_skops_hdr_opt_len(struct sock *sk, struct sk_buff *skb, 577331fca43SMartin KaFai Lau struct request_sock *req, 578331fca43SMartin KaFai Lau struct sk_buff *syn_skb, 579331fca43SMartin KaFai Lau enum tcp_synack_type synack_type, 580331fca43SMartin KaFai Lau struct tcp_out_options *opts, 581331fca43SMartin KaFai Lau unsigned int *remaining) 582331fca43SMartin KaFai Lau { 583331fca43SMartin KaFai Lau } 584331fca43SMartin KaFai Lau 585331fca43SMartin KaFai Lau static void bpf_skops_write_hdr_opt(struct sock *sk, struct sk_buff *skb, 586331fca43SMartin KaFai Lau struct request_sock *req, 587331fca43SMartin KaFai Lau struct sk_buff *syn_skb, 588331fca43SMartin KaFai Lau enum tcp_synack_type synack_type, 589331fca43SMartin KaFai Lau struct tcp_out_options *opts) 590331fca43SMartin KaFai Lau { 591331fca43SMartin KaFai Lau } 592331fca43SMartin KaFai Lau #endif 593331fca43SMartin KaFai Lau 59467edfef7SAndi Kleen /* Write previously computed TCP options to the packet. 59567edfef7SAndi Kleen * 59667edfef7SAndi Kleen * Beware: Something in the Internet is very sensitive to the ordering of 597fd6149d3SIlpo Järvinen * TCP options, we learned this through the hard way, so be careful here. 598fd6149d3SIlpo Järvinen * Luckily we can at least blame others for their non-compliance but from 5998e3bff96Sstephen hemminger * inter-operability perspective it seems that we're somewhat stuck with 600fd6149d3SIlpo Järvinen * the ordering which we have been using if we want to keep working with 601fd6149d3SIlpo Järvinen * those broken things (not that it currently hurts anybody as there isn't 602fd6149d3SIlpo Järvinen * particular reason why the ordering would need to be changed). 603fd6149d3SIlpo Järvinen * 604fd6149d3SIlpo Järvinen * At least SACK_PERM as the first option is known to lead to a disaster 605fd6149d3SIlpo Järvinen * (but it may well be that other scenarios fail similarly). 606fd6149d3SIlpo Järvinen */ 607ea66758cSPaolo Abeni static void tcp_options_write(struct tcphdr *th, struct tcp_sock *tp, 608bd0388aeSWilliam Allen Simpson struct tcp_out_options *opts) 609bd0388aeSWilliam Allen Simpson { 610ea66758cSPaolo Abeni __be32 *ptr = (__be32 *)(th + 1); 6112100c8d2SYuchung Cheng u16 options = opts->options; /* mungable copy */ 612bd0388aeSWilliam Allen Simpson 613bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_MD5 & options)) { 6141a2c6181SChristoph Paasch *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | 6151a2c6181SChristoph Paasch (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG); 616bd0388aeSWilliam Allen Simpson /* overload cookie hash location */ 617bd0388aeSWilliam Allen Simpson opts->hash_location = (__u8 *)ptr; 61833ad798cSAdam Langley ptr += 4; 61933ad798cSAdam Langley } 62033ad798cSAdam Langley 621fd6149d3SIlpo Järvinen if (unlikely(opts->mss)) { 622fd6149d3SIlpo Järvinen *ptr++ = htonl((TCPOPT_MSS << 24) | 623fd6149d3SIlpo Järvinen (TCPOLEN_MSS << 16) | 624fd6149d3SIlpo Järvinen opts->mss); 625fd6149d3SIlpo Järvinen } 626fd6149d3SIlpo Järvinen 627bd0388aeSWilliam Allen Simpson if (likely(OPTION_TS & options)) { 628bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 62933ad798cSAdam Langley *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | 63033ad798cSAdam Langley (TCPOLEN_SACK_PERM << 16) | 63133ad798cSAdam Langley (TCPOPT_TIMESTAMP << 8) | 63233ad798cSAdam Langley TCPOLEN_TIMESTAMP); 633bd0388aeSWilliam Allen Simpson options &= ~OPTION_SACK_ADVERTISE; 63433ad798cSAdam Langley } else { 635496c98dfSYOSHIFUJI Hideaki *ptr++ = htonl((TCPOPT_NOP << 24) | 63640efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 63740efc6faSStephen Hemminger (TCPOPT_TIMESTAMP << 8) | 63840efc6faSStephen Hemminger TCPOLEN_TIMESTAMP); 63940efc6faSStephen Hemminger } 64033ad798cSAdam Langley *ptr++ = htonl(opts->tsval); 64133ad798cSAdam Langley *ptr++ = htonl(opts->tsecr); 64233ad798cSAdam Langley } 64333ad798cSAdam Langley 644bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 64533ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 64633ad798cSAdam Langley (TCPOPT_NOP << 16) | 64733ad798cSAdam Langley (TCPOPT_SACK_PERM << 8) | 64833ad798cSAdam Langley TCPOLEN_SACK_PERM); 64933ad798cSAdam Langley } 65033ad798cSAdam Langley 651bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_WSCALE & options)) { 65233ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 65333ad798cSAdam Langley (TCPOPT_WINDOW << 16) | 65433ad798cSAdam Langley (TCPOLEN_WINDOW << 8) | 65533ad798cSAdam Langley opts->ws); 65633ad798cSAdam Langley } 65733ad798cSAdam Langley 65833ad798cSAdam Langley if (unlikely(opts->num_sack_blocks)) { 65933ad798cSAdam Langley struct tcp_sack_block *sp = tp->rx_opt.dsack ? 66033ad798cSAdam Langley tp->duplicate_sack : tp->selective_acks; 66140efc6faSStephen Hemminger int this_sack; 66240efc6faSStephen Hemminger 66340efc6faSStephen Hemminger *ptr++ = htonl((TCPOPT_NOP << 24) | 66440efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 66540efc6faSStephen Hemminger (TCPOPT_SACK << 8) | 66633ad798cSAdam Langley (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * 66740efc6faSStephen Hemminger TCPOLEN_SACK_PERBLOCK))); 6682de979bdSStephen Hemminger 66933ad798cSAdam Langley for (this_sack = 0; this_sack < opts->num_sack_blocks; 67033ad798cSAdam Langley ++this_sack) { 67140efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].start_seq); 67240efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].end_seq); 67340efc6faSStephen Hemminger } 6742de979bdSStephen Hemminger 67540efc6faSStephen Hemminger tp->rx_opt.dsack = 0; 67640efc6faSStephen Hemminger } 6772100c8d2SYuchung Cheng 6782100c8d2SYuchung Cheng if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) { 6792100c8d2SYuchung Cheng struct tcp_fastopen_cookie *foc = opts->fastopen_cookie; 6807f9b838bSDaniel Lee u8 *p = (u8 *)ptr; 6817f9b838bSDaniel Lee u32 len; /* Fast Open option length */ 6822100c8d2SYuchung Cheng 6837f9b838bSDaniel Lee if (foc->exp) { 6847f9b838bSDaniel Lee len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len; 6857f9b838bSDaniel Lee *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) | 6862100c8d2SYuchung Cheng TCPOPT_FASTOPEN_MAGIC); 6877f9b838bSDaniel Lee p += TCPOLEN_EXP_FASTOPEN_BASE; 6887f9b838bSDaniel Lee } else { 6897f9b838bSDaniel Lee len = TCPOLEN_FASTOPEN_BASE + foc->len; 6907f9b838bSDaniel Lee *p++ = TCPOPT_FASTOPEN; 6917f9b838bSDaniel Lee *p++ = len; 6922100c8d2SYuchung Cheng } 6937f9b838bSDaniel Lee 6947f9b838bSDaniel Lee memcpy(p, foc->val, foc->len); 6957f9b838bSDaniel Lee if ((len & 3) == 2) { 6967f9b838bSDaniel Lee p[foc->len] = TCPOPT_NOP; 6977f9b838bSDaniel Lee p[foc->len + 1] = TCPOPT_NOP; 6987f9b838bSDaniel Lee } 6997f9b838bSDaniel Lee ptr += (len + 3) >> 2; 7002100c8d2SYuchung Cheng } 70160e2a778SUrsula Braun 70260e2a778SUrsula Braun smc_options_write(ptr, &options); 703eda7acddSPeter Krystad 704ea66758cSPaolo Abeni mptcp_options_write(th, ptr, tp, opts); 70560e2a778SUrsula Braun } 70660e2a778SUrsula Braun 70760e2a778SUrsula Braun static void smc_set_option(const struct tcp_sock *tp, 70860e2a778SUrsula Braun struct tcp_out_options *opts, 70960e2a778SUrsula Braun unsigned int *remaining) 71060e2a778SUrsula Braun { 71160e2a778SUrsula Braun #if IS_ENABLED(CONFIG_SMC) 71260e2a778SUrsula Braun if (static_branch_unlikely(&tcp_have_smc)) { 71360e2a778SUrsula Braun if (tp->syn_smc) { 71460e2a778SUrsula Braun if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) { 71560e2a778SUrsula Braun opts->options |= OPTION_SMC; 71660e2a778SUrsula Braun *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED; 71760e2a778SUrsula Braun } 71860e2a778SUrsula Braun } 71960e2a778SUrsula Braun } 72060e2a778SUrsula Braun #endif 72160e2a778SUrsula Braun } 72260e2a778SUrsula Braun 72360e2a778SUrsula Braun static void smc_set_option_cond(const struct tcp_sock *tp, 72460e2a778SUrsula Braun const struct inet_request_sock *ireq, 72560e2a778SUrsula Braun struct tcp_out_options *opts, 72660e2a778SUrsula Braun unsigned int *remaining) 72760e2a778SUrsula Braun { 72860e2a778SUrsula Braun #if IS_ENABLED(CONFIG_SMC) 72960e2a778SUrsula Braun if (static_branch_unlikely(&tcp_have_smc)) { 73060e2a778SUrsula Braun if (tp->syn_smc && ireq->smc_ok) { 73160e2a778SUrsula Braun if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) { 73260e2a778SUrsula Braun opts->options |= OPTION_SMC; 73360e2a778SUrsula Braun *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED; 73460e2a778SUrsula Braun } 73560e2a778SUrsula Braun } 73660e2a778SUrsula Braun } 73760e2a778SUrsula Braun #endif 73840efc6faSStephen Hemminger } 73940efc6faSStephen Hemminger 740cec37a6eSPeter Krystad static void mptcp_set_option_cond(const struct request_sock *req, 741cec37a6eSPeter Krystad struct tcp_out_options *opts, 742cec37a6eSPeter Krystad unsigned int *remaining) 743cec37a6eSPeter Krystad { 744cec37a6eSPeter Krystad if (rsk_is_mptcp(req)) { 745cec37a6eSPeter Krystad unsigned int size; 746cec37a6eSPeter Krystad 747cec37a6eSPeter Krystad if (mptcp_synack_options(req, &size, &opts->mptcp)) { 748cec37a6eSPeter Krystad if (*remaining >= size) { 749cec37a6eSPeter Krystad opts->options |= OPTION_MPTCP; 750cec37a6eSPeter Krystad *remaining -= size; 751cec37a6eSPeter Krystad } 752cec37a6eSPeter Krystad } 753cec37a6eSPeter Krystad } 754cec37a6eSPeter Krystad } 755cec37a6eSPeter Krystad 75667edfef7SAndi Kleen /* Compute TCP options for SYN packets. This is not the final 75767edfef7SAndi Kleen * network wire format yet. 75867edfef7SAndi Kleen */ 75995c96174SEric Dumazet static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb, 76033ad798cSAdam Langley struct tcp_out_options *opts, 761cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 762cf533ea5SEric Dumazet { 76333ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 76495c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 765783237e8SYuchung Cheng struct tcp_fastopen_request *fastopen = tp->fastopen_req; 76633ad798cSAdam Langley 7678c2320e8SEric Dumazet *md5 = NULL; 768cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 769921f9a0fSEric Dumazet if (static_branch_unlikely(&tcp_md5_needed) && 7706015c71eSEric Dumazet rcu_access_pointer(tp->md5sig_info)) { 77133ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 77233ad798cSAdam Langley if (*md5) { 77333ad798cSAdam Langley opts->options |= OPTION_MD5; 774bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 775cfb6eeb4SYOSHIFUJI Hideaki } 7768c2320e8SEric Dumazet } 777cfb6eeb4SYOSHIFUJI Hideaki #endif 77833ad798cSAdam Langley 77933ad798cSAdam Langley /* We always get an MSS option. The option bytes which will be seen in 78033ad798cSAdam Langley * normal data packets should timestamps be used, must be in the MSS 78133ad798cSAdam Langley * advertised. But we subtract them from tp->mss_cache so that 78233ad798cSAdam Langley * calculations in tcp_sendmsg are simpler etc. So account for this 78333ad798cSAdam Langley * fact here if necessary. If we don't do this correctly, as a 78433ad798cSAdam Langley * receiver we won't recognize data packets as being full sized when we 78533ad798cSAdam Langley * should, and thus we won't abide by the delayed ACK rules correctly. 78633ad798cSAdam Langley * SACKs don't matter, we never delay an ACK when we have any of those 78733ad798cSAdam Langley * going out. */ 78833ad798cSAdam Langley opts->mss = tcp_advertise_mss(sk); 789bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 79033ad798cSAdam Langley 7913666f666SKuniyuki Iwashima if (likely(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_timestamps) && !*md5)) { 79233ad798cSAdam Langley opts->options |= OPTION_TS; 7937faee5c0SEric Dumazet opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset; 79433ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 795bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 79633ad798cSAdam Langley } 7973666f666SKuniyuki Iwashima if (likely(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_window_scaling))) { 79833ad798cSAdam Langley opts->ws = tp->rx_opt.rcv_wscale; 79989e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 800bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 80133ad798cSAdam Langley } 8023666f666SKuniyuki Iwashima if (likely(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_sack))) { 80333ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 804b32d1310SDavid S. Miller if (unlikely(!(OPTION_TS & opts->options))) 805bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 80633ad798cSAdam Langley } 80733ad798cSAdam Langley 808783237e8SYuchung Cheng if (fastopen && fastopen->cookie.len >= 0) { 8092646c831SDaniel Lee u32 need = fastopen->cookie.len; 8102646c831SDaniel Lee 8112646c831SDaniel Lee need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE : 8122646c831SDaniel Lee TCPOLEN_FASTOPEN_BASE; 813783237e8SYuchung Cheng need = (need + 3) & ~3U; /* Align to 32 bits */ 814783237e8SYuchung Cheng if (remaining >= need) { 815783237e8SYuchung Cheng opts->options |= OPTION_FAST_OPEN_COOKIE; 816783237e8SYuchung Cheng opts->fastopen_cookie = &fastopen->cookie; 817783237e8SYuchung Cheng remaining -= need; 818783237e8SYuchung Cheng tp->syn_fastopen = 1; 8192646c831SDaniel Lee tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0; 820783237e8SYuchung Cheng } 821783237e8SYuchung Cheng } 822bd0388aeSWilliam Allen Simpson 82360e2a778SUrsula Braun smc_set_option(tp, opts, &remaining); 82460e2a778SUrsula Braun 825cec37a6eSPeter Krystad if (sk_is_mptcp(sk)) { 826cec37a6eSPeter Krystad unsigned int size; 827cec37a6eSPeter Krystad 828cc7972eaSChristoph Paasch if (mptcp_syn_options(sk, skb, &size, &opts->mptcp)) { 829cec37a6eSPeter Krystad opts->options |= OPTION_MPTCP; 830cec37a6eSPeter Krystad remaining -= size; 831cec37a6eSPeter Krystad } 832cec37a6eSPeter Krystad } 833cec37a6eSPeter Krystad 834331fca43SMartin KaFai Lau bpf_skops_hdr_opt_len(sk, skb, NULL, NULL, 0, opts, &remaining); 835331fca43SMartin KaFai Lau 836bd0388aeSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 83733ad798cSAdam Langley } 83833ad798cSAdam Langley 83967edfef7SAndi Kleen /* Set up TCP options for SYN-ACKs. */ 84060e2a778SUrsula Braun static unsigned int tcp_synack_options(const struct sock *sk, 84160e2a778SUrsula Braun struct request_sock *req, 84295c96174SEric Dumazet unsigned int mss, struct sk_buff *skb, 84333ad798cSAdam Langley struct tcp_out_options *opts, 84480f03e27SEric Dumazet const struct tcp_md5sig_key *md5, 845e114e1e8SEric Dumazet struct tcp_fastopen_cookie *foc, 846331fca43SMartin KaFai Lau enum tcp_synack_type synack_type, 847331fca43SMartin KaFai Lau struct sk_buff *syn_skb) 8484957faadSWilliam Allen Simpson { 84933ad798cSAdam Langley struct inet_request_sock *ireq = inet_rsk(req); 85095c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 85133ad798cSAdam Langley 85233ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 85380f03e27SEric Dumazet if (md5) { 85433ad798cSAdam Langley opts->options |= OPTION_MD5; 8554957faadSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 8564957faadSWilliam Allen Simpson 8574957faadSWilliam Allen Simpson /* We can't fit any SACK blocks in a packet with MD5 + TS 8584957faadSWilliam Allen Simpson * options. There was discussion about disabling SACK 8594957faadSWilliam Allen Simpson * rather than TS in order to fit in better with old, 8604957faadSWilliam Allen Simpson * buggy kernels, but that was deemed to be unnecessary. 8614957faadSWilliam Allen Simpson */ 862e114e1e8SEric Dumazet if (synack_type != TCP_SYNACK_COOKIE) 863de213e5eSEric Dumazet ireq->tstamp_ok &= !ireq->sack_ok; 86433ad798cSAdam Langley } 86533ad798cSAdam Langley #endif 86633ad798cSAdam Langley 8674957faadSWilliam Allen Simpson /* We always send an MSS option. */ 86833ad798cSAdam Langley opts->mss = mss; 8694957faadSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 87033ad798cSAdam Langley 87133ad798cSAdam Langley if (likely(ireq->wscale_ok)) { 87233ad798cSAdam Langley opts->ws = ireq->rcv_wscale; 87389e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 8744957faadSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 87533ad798cSAdam Langley } 876de213e5eSEric Dumazet if (likely(ireq->tstamp_ok)) { 87733ad798cSAdam Langley opts->options |= OPTION_TS; 87895a22caeSFlorian Westphal opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off; 87933ad798cSAdam Langley opts->tsecr = req->ts_recent; 8804957faadSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 88133ad798cSAdam Langley } 88233ad798cSAdam Langley if (likely(ireq->sack_ok)) { 88333ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 884de213e5eSEric Dumazet if (unlikely(!ireq->tstamp_ok)) 8854957faadSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 88633ad798cSAdam Langley } 8877f9b838bSDaniel Lee if (foc != NULL && foc->len >= 0) { 8887f9b838bSDaniel Lee u32 need = foc->len; 8897f9b838bSDaniel Lee 8907f9b838bSDaniel Lee need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE : 8917f9b838bSDaniel Lee TCPOLEN_FASTOPEN_BASE; 8928336886fSJerry Chu need = (need + 3) & ~3U; /* Align to 32 bits */ 8938336886fSJerry Chu if (remaining >= need) { 8948336886fSJerry Chu opts->options |= OPTION_FAST_OPEN_COOKIE; 8958336886fSJerry Chu opts->fastopen_cookie = foc; 8968336886fSJerry Chu remaining -= need; 8978336886fSJerry Chu } 8988336886fSJerry Chu } 8994957faadSWilliam Allen Simpson 900cec37a6eSPeter Krystad mptcp_set_option_cond(req, opts, &remaining); 901cec37a6eSPeter Krystad 90260e2a778SUrsula Braun smc_set_option_cond(tcp_sk(sk), ireq, opts, &remaining); 90360e2a778SUrsula Braun 904331fca43SMartin KaFai Lau bpf_skops_hdr_opt_len((struct sock *)sk, skb, req, syn_skb, 905331fca43SMartin KaFai Lau synack_type, opts, &remaining); 906331fca43SMartin KaFai Lau 9074957faadSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 90833ad798cSAdam Langley } 90933ad798cSAdam Langley 91067edfef7SAndi Kleen /* Compute TCP options for ESTABLISHED sockets. This is not the 91167edfef7SAndi Kleen * final wire format yet. 91267edfef7SAndi Kleen */ 91395c96174SEric Dumazet static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb, 91433ad798cSAdam Langley struct tcp_out_options *opts, 915cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 916cf533ea5SEric Dumazet { 91733ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 91895c96174SEric Dumazet unsigned int size = 0; 919cabeccbdSIlpo Järvinen unsigned int eff_sacks; 92033ad798cSAdam Langley 9215843ef42SAndi Kleen opts->options = 0; 9225843ef42SAndi Kleen 9238c2320e8SEric Dumazet *md5 = NULL; 92433ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 925921f9a0fSEric Dumazet if (static_branch_unlikely(&tcp_md5_needed) && 9266015c71eSEric Dumazet rcu_access_pointer(tp->md5sig_info)) { 92733ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 9288c2320e8SEric Dumazet if (*md5) { 92933ad798cSAdam Langley opts->options |= OPTION_MD5; 93033ad798cSAdam Langley size += TCPOLEN_MD5SIG_ALIGNED; 93133ad798cSAdam Langley } 9328c2320e8SEric Dumazet } 93333ad798cSAdam Langley #endif 93433ad798cSAdam Langley 93533ad798cSAdam Langley if (likely(tp->rx_opt.tstamp_ok)) { 93633ad798cSAdam Langley opts->options |= OPTION_TS; 9377faee5c0SEric Dumazet opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0; 93833ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 93933ad798cSAdam Langley size += TCPOLEN_TSTAMP_ALIGNED; 94033ad798cSAdam Langley } 94133ad798cSAdam Langley 942cec37a6eSPeter Krystad /* MPTCP options have precedence over SACK for the limited TCP 943cec37a6eSPeter Krystad * option space because a MPTCP connection would be forced to 944cec37a6eSPeter Krystad * fall back to regular TCP if a required multipath option is 945cec37a6eSPeter Krystad * missing. SACK still gets a chance to use whatever space is 946cec37a6eSPeter Krystad * left. 947cec37a6eSPeter Krystad */ 948cec37a6eSPeter Krystad if (sk_is_mptcp(sk)) { 949cec37a6eSPeter Krystad unsigned int remaining = MAX_TCP_OPTION_SPACE - size; 950cec37a6eSPeter Krystad unsigned int opt_size = 0; 951cec37a6eSPeter Krystad 952cec37a6eSPeter Krystad if (mptcp_established_options(sk, skb, &opt_size, remaining, 953cec37a6eSPeter Krystad &opts->mptcp)) { 954cec37a6eSPeter Krystad opts->options |= OPTION_MPTCP; 955cec37a6eSPeter Krystad size += opt_size; 956cec37a6eSPeter Krystad } 957cec37a6eSPeter Krystad } 958cec37a6eSPeter Krystad 959cabeccbdSIlpo Järvinen eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; 960cabeccbdSIlpo Järvinen if (unlikely(eff_sacks)) { 96195c96174SEric Dumazet const unsigned int remaining = MAX_TCP_OPTION_SPACE - size; 9629cfcca23SMat Martineau if (unlikely(remaining < TCPOLEN_SACK_BASE_ALIGNED + 9639cfcca23SMat Martineau TCPOLEN_SACK_PERBLOCK)) 9649cfcca23SMat Martineau return size; 9659cfcca23SMat Martineau 96633ad798cSAdam Langley opts->num_sack_blocks = 96795c96174SEric Dumazet min_t(unsigned int, eff_sacks, 96833ad798cSAdam Langley (remaining - TCPOLEN_SACK_BASE_ALIGNED) / 96933ad798cSAdam Langley TCPOLEN_SACK_PERBLOCK); 9709cfcca23SMat Martineau 97133ad798cSAdam Langley size += TCPOLEN_SACK_BASE_ALIGNED + 97233ad798cSAdam Langley opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; 97333ad798cSAdam Langley } 97433ad798cSAdam Langley 975331fca43SMartin KaFai Lau if (unlikely(BPF_SOCK_OPS_TEST_FLAG(tp, 976331fca43SMartin KaFai Lau BPF_SOCK_OPS_WRITE_HDR_OPT_CB_FLAG))) { 977331fca43SMartin KaFai Lau unsigned int remaining = MAX_TCP_OPTION_SPACE - size; 978331fca43SMartin KaFai Lau 979331fca43SMartin KaFai Lau bpf_skops_hdr_opt_len(sk, skb, NULL, NULL, 0, opts, &remaining); 980331fca43SMartin KaFai Lau 981331fca43SMartin KaFai Lau size = MAX_TCP_OPTION_SPACE - remaining; 982331fca43SMartin KaFai Lau } 983331fca43SMartin KaFai Lau 98433ad798cSAdam Langley return size; 98540efc6faSStephen Hemminger } 9861da177e4SLinus Torvalds 98746d3ceabSEric Dumazet 98846d3ceabSEric Dumazet /* TCP SMALL QUEUES (TSQ) 98946d3ceabSEric Dumazet * 99046d3ceabSEric Dumazet * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev) 99146d3ceabSEric Dumazet * to reduce RTT and bufferbloat. 99246d3ceabSEric Dumazet * We do this using a special skb destructor (tcp_wfree). 99346d3ceabSEric Dumazet * 99446d3ceabSEric Dumazet * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb 99546d3ceabSEric Dumazet * needs to be reallocated in a driver. 9968e3bff96Sstephen hemminger * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc 99746d3ceabSEric Dumazet * 99846d3ceabSEric Dumazet * Since transmit from skb destructor is forbidden, we use a tasklet 99946d3ceabSEric Dumazet * to process all sockets that eventually need to send more skbs. 100046d3ceabSEric Dumazet * We use one tasklet per cpu, with its own queue of sockets. 100146d3ceabSEric Dumazet */ 100246d3ceabSEric Dumazet struct tsq_tasklet { 100346d3ceabSEric Dumazet struct tasklet_struct tasklet; 100446d3ceabSEric Dumazet struct list_head head; /* queue of tcp sockets */ 100546d3ceabSEric Dumazet }; 100646d3ceabSEric Dumazet static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet); 100746d3ceabSEric Dumazet 100873a6bab5SEric Dumazet static void tcp_tsq_write(struct sock *sk) 10096f458dfbSEric Dumazet { 10106f458dfbSEric Dumazet if ((1 << sk->sk_state) & 10116f458dfbSEric Dumazet (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING | 1012f9616c35SEric Dumazet TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) { 1013f9616c35SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 1014f9616c35SEric Dumazet 1015f9616c35SEric Dumazet if (tp->lost_out > tp->retrans_out && 101640570375SEric Dumazet tcp_snd_cwnd(tp) > tcp_packets_in_flight(tp)) { 10173a91d29fSKoichiro Den tcp_mstamp_refresh(tp); 1018f9616c35SEric Dumazet tcp_xmit_retransmit_queue(sk); 10193a91d29fSKoichiro Den } 1020f9616c35SEric Dumazet 1021f9616c35SEric Dumazet tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle, 1022bf06200eSJohn Ogness 0, GFP_ATOMIC); 10236f458dfbSEric Dumazet } 1024f9616c35SEric Dumazet } 102573a6bab5SEric Dumazet 102673a6bab5SEric Dumazet static void tcp_tsq_handler(struct sock *sk) 102773a6bab5SEric Dumazet { 102873a6bab5SEric Dumazet bh_lock_sock(sk); 102973a6bab5SEric Dumazet if (!sock_owned_by_user(sk)) 103073a6bab5SEric Dumazet tcp_tsq_write(sk); 103173a6bab5SEric Dumazet else if (!test_and_set_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags)) 103273a6bab5SEric Dumazet sock_hold(sk); 103373a6bab5SEric Dumazet bh_unlock_sock(sk); 103473a6bab5SEric Dumazet } 103546d3ceabSEric Dumazet /* 10368e3bff96Sstephen hemminger * One tasklet per cpu tries to send more skbs. 103746d3ceabSEric Dumazet * We run in tasklet context but need to disable irqs when 10388e3bff96Sstephen hemminger * transferring tsq->head because tcp_wfree() might 103946d3ceabSEric Dumazet * interrupt us (non NAPI drivers) 104046d3ceabSEric Dumazet */ 1041c6533ca8SAllen Pais static void tcp_tasklet_func(struct tasklet_struct *t) 104246d3ceabSEric Dumazet { 1043c6533ca8SAllen Pais struct tsq_tasklet *tsq = from_tasklet(tsq, t, tasklet); 104446d3ceabSEric Dumazet LIST_HEAD(list); 104546d3ceabSEric Dumazet unsigned long flags; 104646d3ceabSEric Dumazet struct list_head *q, *n; 104746d3ceabSEric Dumazet struct tcp_sock *tp; 104846d3ceabSEric Dumazet struct sock *sk; 104946d3ceabSEric Dumazet 105046d3ceabSEric Dumazet local_irq_save(flags); 105146d3ceabSEric Dumazet list_splice_init(&tsq->head, &list); 105246d3ceabSEric Dumazet local_irq_restore(flags); 105346d3ceabSEric Dumazet 105446d3ceabSEric Dumazet list_for_each_safe(q, n, &list) { 105546d3ceabSEric Dumazet tp = list_entry(q, struct tcp_sock, tsq_node); 105646d3ceabSEric Dumazet list_del(&tp->tsq_node); 105746d3ceabSEric Dumazet 105846d3ceabSEric Dumazet sk = (struct sock *)tp; 10590a9648f1SEric Dumazet smp_mb__before_atomic(); 10607aa5470cSEric Dumazet clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags); 10617aa5470cSEric Dumazet 10626f458dfbSEric Dumazet tcp_tsq_handler(sk); 106346d3ceabSEric Dumazet sk_free(sk); 106446d3ceabSEric Dumazet } 106546d3ceabSEric Dumazet } 106646d3ceabSEric Dumazet 106740fc3423SEric Dumazet #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \ 106840fc3423SEric Dumazet TCPF_WRITE_TIMER_DEFERRED | \ 106940fc3423SEric Dumazet TCPF_DELACK_TIMER_DEFERRED | \ 107040fc3423SEric Dumazet TCPF_MTU_REDUCED_DEFERRED) 107146d3ceabSEric Dumazet /** 107246d3ceabSEric Dumazet * tcp_release_cb - tcp release_sock() callback 107346d3ceabSEric Dumazet * @sk: socket 107446d3ceabSEric Dumazet * 107546d3ceabSEric Dumazet * called from release_sock() to perform protocol dependent 107646d3ceabSEric Dumazet * actions before socket release. 107746d3ceabSEric Dumazet */ 107846d3ceabSEric Dumazet void tcp_release_cb(struct sock *sk) 107946d3ceabSEric Dumazet { 10806f458dfbSEric Dumazet unsigned long flags, nflags; 108146d3ceabSEric Dumazet 10826f458dfbSEric Dumazet /* perform an atomic operation only if at least one flag is set */ 10836f458dfbSEric Dumazet do { 10847aa5470cSEric Dumazet flags = sk->sk_tsq_flags; 10856f458dfbSEric Dumazet if (!(flags & TCP_DEFERRED_ALL)) 10866f458dfbSEric Dumazet return; 10876f458dfbSEric Dumazet nflags = flags & ~TCP_DEFERRED_ALL; 10887aa5470cSEric Dumazet } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags); 10896f458dfbSEric Dumazet 109073a6bab5SEric Dumazet if (flags & TCPF_TSQ_DEFERRED) { 109173a6bab5SEric Dumazet tcp_tsq_write(sk); 109273a6bab5SEric Dumazet __sock_put(sk); 109373a6bab5SEric Dumazet } 1094c3f9b018SEric Dumazet /* Here begins the tricky part : 1095c3f9b018SEric Dumazet * We are called from release_sock() with : 1096c3f9b018SEric Dumazet * 1) BH disabled 1097c3f9b018SEric Dumazet * 2) sk_lock.slock spinlock held 1098c3f9b018SEric Dumazet * 3) socket owned by us (sk->sk_lock.owned == 1) 1099c3f9b018SEric Dumazet * 1100c3f9b018SEric Dumazet * But following code is meant to be called from BH handlers, 1101c3f9b018SEric Dumazet * so we should keep BH disabled, but early release socket ownership 1102c3f9b018SEric Dumazet */ 1103c3f9b018SEric Dumazet sock_release_ownership(sk); 1104c3f9b018SEric Dumazet 110540fc3423SEric Dumazet if (flags & TCPF_WRITE_TIMER_DEFERRED) { 11066f458dfbSEric Dumazet tcp_write_timer_handler(sk); 1107144d56e9SEric Dumazet __sock_put(sk); 1108144d56e9SEric Dumazet } 110940fc3423SEric Dumazet if (flags & TCPF_DELACK_TIMER_DEFERRED) { 11106f458dfbSEric Dumazet tcp_delack_timer_handler(sk); 1111144d56e9SEric Dumazet __sock_put(sk); 1112144d56e9SEric Dumazet } 111340fc3423SEric Dumazet if (flags & TCPF_MTU_REDUCED_DEFERRED) { 11144fab9071SNeal Cardwell inet_csk(sk)->icsk_af_ops->mtu_reduced(sk); 1115144d56e9SEric Dumazet __sock_put(sk); 1116144d56e9SEric Dumazet } 111746d3ceabSEric Dumazet } 111846d3ceabSEric Dumazet EXPORT_SYMBOL(tcp_release_cb); 111946d3ceabSEric Dumazet 112046d3ceabSEric Dumazet void __init tcp_tasklet_init(void) 112146d3ceabSEric Dumazet { 112246d3ceabSEric Dumazet int i; 112346d3ceabSEric Dumazet 112446d3ceabSEric Dumazet for_each_possible_cpu(i) { 112546d3ceabSEric Dumazet struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i); 112646d3ceabSEric Dumazet 112746d3ceabSEric Dumazet INIT_LIST_HEAD(&tsq->head); 1128c6533ca8SAllen Pais tasklet_setup(&tsq->tasklet, tcp_tasklet_func); 112946d3ceabSEric Dumazet } 113046d3ceabSEric Dumazet } 113146d3ceabSEric Dumazet 113246d3ceabSEric Dumazet /* 113346d3ceabSEric Dumazet * Write buffer destructor automatically called from kfree_skb. 11348e3bff96Sstephen hemminger * We can't xmit new skbs from this context, as we might already 113546d3ceabSEric Dumazet * hold qdisc lock. 113646d3ceabSEric Dumazet */ 1137d6a4a104SEric Dumazet void tcp_wfree(struct sk_buff *skb) 113846d3ceabSEric Dumazet { 113946d3ceabSEric Dumazet struct sock *sk = skb->sk; 114046d3ceabSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 1141408f0a6cSEric Dumazet unsigned long flags, nval, oval; 11429b462d02SEric Dumazet 11439b462d02SEric Dumazet /* Keep one reference on sk_wmem_alloc. 11449b462d02SEric Dumazet * Will be released by sk_free() from here or tcp_tasklet_func() 11459b462d02SEric Dumazet */ 114614afee4bSReshetova, Elena WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc)); 11479b462d02SEric Dumazet 11489b462d02SEric Dumazet /* If this softirq is serviced by ksoftirqd, we are likely under stress. 11499b462d02SEric Dumazet * Wait until our queues (qdisc + devices) are drained. 11509b462d02SEric Dumazet * This gives : 11519b462d02SEric Dumazet * - less callbacks to tcp_write_xmit(), reducing stress (batches) 11529b462d02SEric Dumazet * - chance for incoming ACK (processed by another cpu maybe) 11539b462d02SEric Dumazet * to migrate this flow (skb->ooo_okay will be eventually set) 11549b462d02SEric Dumazet */ 115514afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current) 11569b462d02SEric Dumazet goto out; 115746d3ceabSEric Dumazet 11587aa5470cSEric Dumazet for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) { 115946d3ceabSEric Dumazet struct tsq_tasklet *tsq; 1160a9b204d1SEric Dumazet bool empty; 116146d3ceabSEric Dumazet 1162408f0a6cSEric Dumazet if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED)) 1163408f0a6cSEric Dumazet goto out; 1164408f0a6cSEric Dumazet 116573a6bab5SEric Dumazet nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED; 11667aa5470cSEric Dumazet nval = cmpxchg(&sk->sk_tsq_flags, oval, nval); 1167408f0a6cSEric Dumazet if (nval != oval) 1168408f0a6cSEric Dumazet continue; 1169408f0a6cSEric Dumazet 117046d3ceabSEric Dumazet /* queue this socket to tasklet queue */ 117146d3ceabSEric Dumazet local_irq_save(flags); 1172903ceff7SChristoph Lameter tsq = this_cpu_ptr(&tsq_tasklet); 1173a9b204d1SEric Dumazet empty = list_empty(&tsq->head); 117446d3ceabSEric Dumazet list_add(&tp->tsq_node, &tsq->head); 1175a9b204d1SEric Dumazet if (empty) 117646d3ceabSEric Dumazet tasklet_schedule(&tsq->tasklet); 117746d3ceabSEric Dumazet local_irq_restore(flags); 11789b462d02SEric Dumazet return; 117946d3ceabSEric Dumazet } 11809b462d02SEric Dumazet out: 11819b462d02SEric Dumazet sk_free(sk); 118246d3ceabSEric Dumazet } 118346d3ceabSEric Dumazet 118473a6bab5SEric Dumazet /* Note: Called under soft irq. 118573a6bab5SEric Dumazet * We can call TCP stack right away, unless socket is owned by user. 1186218af599SEric Dumazet */ 1187218af599SEric Dumazet enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer) 1188218af599SEric Dumazet { 1189218af599SEric Dumazet struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer); 1190218af599SEric Dumazet struct sock *sk = (struct sock *)tp; 1191218af599SEric Dumazet 119273a6bab5SEric Dumazet tcp_tsq_handler(sk); 119373a6bab5SEric Dumazet sock_put(sk); 1194218af599SEric Dumazet 1195218af599SEric Dumazet return HRTIMER_NORESTART; 1196218af599SEric Dumazet } 1197218af599SEric Dumazet 1198a7a25630SEric Dumazet static void tcp_update_skb_after_send(struct sock *sk, struct sk_buff *skb, 1199a7a25630SEric Dumazet u64 prior_wstamp) 1200e2080072SEric Dumazet { 1201ab408b6dSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 1202ab408b6dSEric Dumazet 1203ab408b6dSEric Dumazet if (sk->sk_pacing_status != SK_PACING_NONE) { 120476a9ebe8SEric Dumazet unsigned long rate = sk->sk_pacing_rate; 1205ab408b6dSEric Dumazet 1206ab408b6dSEric Dumazet /* Original sch_fq does not pace first 10 MSS 1207ab408b6dSEric Dumazet * Note that tp->data_segs_out overflows after 2^32 packets, 1208ab408b6dSEric Dumazet * this is a minor annoyance. 1209ab408b6dSEric Dumazet */ 121076a9ebe8SEric Dumazet if (rate != ~0UL && rate && tp->data_segs_out >= 10) { 1211a7a25630SEric Dumazet u64 len_ns = div64_ul((u64)skb->len * NSEC_PER_SEC, rate); 1212a7a25630SEric Dumazet u64 credit = tp->tcp_wstamp_ns - prior_wstamp; 1213a7a25630SEric Dumazet 1214a7a25630SEric Dumazet /* take into account OS jitter */ 1215a7a25630SEric Dumazet len_ns -= min_t(u64, len_ns / 2, credit); 1216a7a25630SEric Dumazet tp->tcp_wstamp_ns += len_ns; 1217ab408b6dSEric Dumazet } 1218ab408b6dSEric Dumazet } 1219e2080072SEric Dumazet list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue); 1220e2080072SEric Dumazet } 1221e2080072SEric Dumazet 122205e22e83SEric Dumazet INDIRECT_CALLABLE_DECLARE(int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl)); 122305e22e83SEric Dumazet INDIRECT_CALLABLE_DECLARE(int inet6_csk_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl)); 1224dd2e0b86SEric Dumazet INDIRECT_CALLABLE_DECLARE(void tcp_v4_send_check(struct sock *sk, struct sk_buff *skb)); 122505e22e83SEric Dumazet 12261da177e4SLinus Torvalds /* This routine actually transmits TCP packets queued in by 12271da177e4SLinus Torvalds * tcp_do_sendmsg(). This is used by both the initial 12281da177e4SLinus Torvalds * transmission and possible later retransmissions. 12291da177e4SLinus Torvalds * All SKB's seen here are completely headerless. It is our 12301da177e4SLinus Torvalds * job to build the TCP header, and pass the packet down to 12311da177e4SLinus Torvalds * IP so it can do the same plus pass the packet off to the 12321da177e4SLinus Torvalds * device. 12331da177e4SLinus Torvalds * 12341da177e4SLinus Torvalds * We are working here with either a clone of the original 12351da177e4SLinus Torvalds * SKB, or a fresh unique copy made by the retransmit engine. 12361da177e4SLinus Torvalds */ 12372987babbSYuchung Cheng static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, 12382987babbSYuchung Cheng int clone_it, gfp_t gfp_mask, u32 rcv_nxt) 12391da177e4SLinus Torvalds { 12406687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 1241dfb4b9dcSDavid S. Miller struct inet_sock *inet; 1242dfb4b9dcSDavid S. Miller struct tcp_sock *tp; 1243dfb4b9dcSDavid S. Miller struct tcp_skb_cb *tcb; 124433ad798cSAdam Langley struct tcp_out_options opts; 124595c96174SEric Dumazet unsigned int tcp_options_size, tcp_header_size; 12468c72c65bSEric Dumazet struct sk_buff *oskb = NULL; 1247cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 12481da177e4SLinus Torvalds struct tcphdr *th; 1249a7a25630SEric Dumazet u64 prior_wstamp; 12501da177e4SLinus Torvalds int err; 12511da177e4SLinus Torvalds 1252dfb4b9dcSDavid S. Miller BUG_ON(!skb || !tcp_skb_pcount(skb)); 12536f094b9eSLawrence Brakmo tp = tcp_sk(sk); 12547f12422cSYuchung Cheng prior_wstamp = tp->tcp_wstamp_ns; 12557f12422cSYuchung Cheng tp->tcp_wstamp_ns = max(tp->tcp_wstamp_ns, tp->tcp_clock_cache); 1256a1ac9c8aSMartin KaFai Lau skb_set_delivery_time(skb, tp->tcp_wstamp_ns, true); 1257ccdbb6e9SEric Dumazet if (clone_it) { 12588c72c65bSEric Dumazet oskb = skb; 1259e2080072SEric Dumazet 1260e2080072SEric Dumazet tcp_skb_tsorted_save(oskb) { 1261e2080072SEric Dumazet if (unlikely(skb_cloned(oskb))) 1262e2080072SEric Dumazet skb = pskb_copy(oskb, gfp_mask); 1263dfb4b9dcSDavid S. Miller else 1264e2080072SEric Dumazet skb = skb_clone(oskb, gfp_mask); 1265e2080072SEric Dumazet } tcp_skb_tsorted_restore(oskb); 1266e2080072SEric Dumazet 1267dfb4b9dcSDavid S. Miller if (unlikely(!skb)) 1268dfb4b9dcSDavid S. Miller return -ENOBUFS; 1269b738a185SEric Dumazet /* retransmit skbs might have a non zero value in skb->dev 1270b738a185SEric Dumazet * because skb->dev is aliased with skb->rbnode.rb_left 1271b738a185SEric Dumazet */ 1272b738a185SEric Dumazet skb->dev = NULL; 1273dfb4b9dcSDavid S. Miller } 12745f6188a8SEric Dumazet 1275dfb4b9dcSDavid S. Miller inet = inet_sk(sk); 1276dfb4b9dcSDavid S. Miller tcb = TCP_SKB_CB(skb); 127733ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 12781da177e4SLinus Torvalds 1279051ba674SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) { 128033ad798cSAdam Langley tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); 1281051ba674SEric Dumazet } else { 128233ad798cSAdam Langley tcp_options_size = tcp_established_options(sk, skb, &opts, 128333ad798cSAdam Langley &md5); 1284051ba674SEric Dumazet /* Force a PSH flag on all (GSO) packets to expedite GRO flush 1285051ba674SEric Dumazet * at receiver : This slightly improve GRO performance. 1286051ba674SEric Dumazet * Note that we do not force the PSH flag for non GSO packets, 1287051ba674SEric Dumazet * because they might be sent under high congestion events, 1288051ba674SEric Dumazet * and in this case it is better to delay the delivery of 1-MSS 1289051ba674SEric Dumazet * packets and thus the corresponding ACK packet that would 1290051ba674SEric Dumazet * release the following packet. 1291051ba674SEric Dumazet */ 1292051ba674SEric Dumazet if (tcp_skb_pcount(skb) > 1) 1293051ba674SEric Dumazet tcb->tcp_flags |= TCPHDR_PSH; 1294051ba674SEric Dumazet } 129533ad798cSAdam Langley tcp_header_size = tcp_options_size + sizeof(struct tcphdr); 12961da177e4SLinus Torvalds 1297547669d4SEric Dumazet /* if no packet is in qdisc/device queue, then allow XPS to select 1298b2532eb9SEric Dumazet * another queue. We can be called from tcp_tsq_handler() 129973a6bab5SEric Dumazet * which holds one reference to sk. 1300b2532eb9SEric Dumazet * 1301b2532eb9SEric Dumazet * TODO: Ideally, in-flight pure ACK packets should not matter here. 1302b2532eb9SEric Dumazet * One way to get this would be to set skb->truesize = 2 on them. 1303547669d4SEric Dumazet */ 1304b2532eb9SEric Dumazet skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1); 13051da177e4SLinus Torvalds 130638ab52e8SEric Dumazet /* If we had to use memory reserve to allocate this skb, 130738ab52e8SEric Dumazet * this might cause drops if packet is looped back : 130838ab52e8SEric Dumazet * Other socket might not have SOCK_MEMALLOC. 130938ab52e8SEric Dumazet * Packets not looped back do not care about pfmemalloc. 131038ab52e8SEric Dumazet */ 131138ab52e8SEric Dumazet skb->pfmemalloc = 0; 131238ab52e8SEric Dumazet 1313aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 1314aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 131546d3ceabSEric Dumazet 131646d3ceabSEric Dumazet skb_orphan(skb); 131746d3ceabSEric Dumazet skb->sk = sk; 13181d2077acSEric Dumazet skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree; 131914afee4bSReshetova, Elena refcount_add(skb->truesize, &sk->sk_wmem_alloc); 13201da177e4SLinus Torvalds 1321c3a2e837SJulian Anastasov skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm); 1322c3a2e837SJulian Anastasov 13231da177e4SLinus Torvalds /* Build TCP header and checksum it. */ 1324ea1627c2SEric Dumazet th = (struct tcphdr *)skb->data; 1325c720c7e8SEric Dumazet th->source = inet->inet_sport; 1326c720c7e8SEric Dumazet th->dest = inet->inet_dport; 13271da177e4SLinus Torvalds th->seq = htonl(tcb->seq); 13282987babbSYuchung Cheng th->ack_seq = htonl(rcv_nxt); 1329df7a3b07SAl Viro *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | 13304de075e0SEric Dumazet tcb->tcp_flags); 1331dfb4b9dcSDavid S. Miller 13321da177e4SLinus Torvalds th->check = 0; 13331da177e4SLinus Torvalds th->urg_ptr = 0; 13341da177e4SLinus Torvalds 133533f5f57eSIlpo Järvinen /* The urg_mode check is necessary during a below snd_una win probe */ 13367691367dSHerbert Xu if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { 13377691367dSHerbert Xu if (before(tp->snd_up, tcb->seq + 0x10000)) { 13381da177e4SLinus Torvalds th->urg_ptr = htons(tp->snd_up - tcb->seq); 13391da177e4SLinus Torvalds th->urg = 1; 13407691367dSHerbert Xu } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { 13410eae88f3SEric Dumazet th->urg_ptr = htons(0xFFFF); 13427691367dSHerbert Xu th->urg = 1; 13437691367dSHerbert Xu } 13441da177e4SLinus Torvalds } 13451da177e4SLinus Torvalds 134651466a75SEric Dumazet skb_shinfo(skb)->gso_type = sk->sk_gso_type; 1347ea1627c2SEric Dumazet if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) { 1348ea1627c2SEric Dumazet th->window = htons(tcp_select_window(sk)); 1349ea1627c2SEric Dumazet tcp_ecn_send(sk, skb, th, tcp_header_size); 1350ea1627c2SEric Dumazet } else { 1351ea1627c2SEric Dumazet /* RFC1323: The window in SYN & SYN/ACK segments 1352ea1627c2SEric Dumazet * is never scaled. 1353ea1627c2SEric Dumazet */ 1354ea1627c2SEric Dumazet th->window = htons(min(tp->rcv_wnd, 65535U)); 1355ea1627c2SEric Dumazet } 1356fa3fe2b1SFlorian Westphal 1357ea66758cSPaolo Abeni tcp_options_write(th, tp, &opts); 1358fa3fe2b1SFlorian Westphal 1359cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 1360cfb6eeb4SYOSHIFUJI Hideaki /* Calculate the MD5 hash, as we have all we need now */ 1361cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 1362aba54656SEric Dumazet sk_gso_disable(sk); 1363bd0388aeSWilliam Allen Simpson tp->af_specific->calc_md5_hash(opts.hash_location, 136439f8e58eSEric Dumazet md5, sk, skb); 1365cfb6eeb4SYOSHIFUJI Hideaki } 1366cfb6eeb4SYOSHIFUJI Hideaki #endif 1367cfb6eeb4SYOSHIFUJI Hideaki 1368331fca43SMartin KaFai Lau /* BPF prog is the last one writing header option */ 1369331fca43SMartin KaFai Lau bpf_skops_write_hdr_opt(sk, skb, NULL, NULL, 0, &opts); 1370331fca43SMartin KaFai Lau 1371dd2e0b86SEric Dumazet INDIRECT_CALL_INET(icsk->icsk_af_ops->send_check, 1372dd2e0b86SEric Dumazet tcp_v6_send_check, tcp_v4_send_check, 1373dd2e0b86SEric Dumazet sk, skb); 13741da177e4SLinus Torvalds 13754de075e0SEric Dumazet if (likely(tcb->tcp_flags & TCPHDR_ACK)) 137627cde44aSYuchung Cheng tcp_event_ack_sent(sk, tcp_skb_pcount(skb), rcv_nxt); 13771da177e4SLinus Torvalds 1378a44d6eacSMartin KaFai Lau if (skb->len != tcp_header_size) { 1379cf533ea5SEric Dumazet tcp_event_data_sent(tp, sk); 1380a44d6eacSMartin KaFai Lau tp->data_segs_out += tcp_skb_pcount(skb); 1381ba113c3aSWei Wang tp->bytes_sent += skb->len - tcp_header_size; 1382a44d6eacSMartin KaFai Lau } 13831da177e4SLinus Torvalds 1384bd37a088SWei Yongjun if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) 1385aa2ea058STom Herbert TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, 1386aa2ea058STom Herbert tcp_skb_pcount(skb)); 13871da177e4SLinus Torvalds 13882efd055cSMarcelo Ricardo Leitner tp->segs_out += tcp_skb_pcount(skb); 13890ae5b43dSYuchung Cheng skb_set_hash_from_sk(skb, sk); 1390f69ad292SEric Dumazet /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */ 1391cd7d8498SEric Dumazet skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb); 1392f69ad292SEric Dumazet skb_shinfo(skb)->gso_size = tcp_skb_mss(skb); 1393cd7d8498SEric Dumazet 1394d3edd06eSEric Dumazet /* Leave earliest departure time in skb->tstamp (skb->skb_mstamp_ns) */ 1395971f10ecSEric Dumazet 1396971f10ecSEric Dumazet /* Cleanup our debris for IP stacks */ 1397971f10ecSEric Dumazet memset(skb->cb, 0, max(sizeof(struct inet_skb_parm), 1398971f10ecSEric Dumazet sizeof(struct inet6_skb_parm))); 1399971f10ecSEric Dumazet 1400a842fe14SEric Dumazet tcp_add_tx_delay(skb, tp); 1401a842fe14SEric Dumazet 140205e22e83SEric Dumazet err = INDIRECT_CALL_INET(icsk->icsk_af_ops->queue_xmit, 140305e22e83SEric Dumazet inet6_csk_xmit, ip_queue_xmit, 140405e22e83SEric Dumazet sk, skb, &inet->cork.fl); 14057faee5c0SEric Dumazet 14068c72c65bSEric Dumazet if (unlikely(err > 0)) { 14075ee2c941SChristoph Paasch tcp_enter_cwr(sk); 14088c72c65bSEric Dumazet err = net_xmit_eval(err); 14098c72c65bSEric Dumazet } 1410fc225799SEric Dumazet if (!err && oskb) { 1411a7a25630SEric Dumazet tcp_update_skb_after_send(sk, oskb, prior_wstamp); 1412fc225799SEric Dumazet tcp_rate_skb_sent(sk, oskb); 1413fc225799SEric Dumazet } 14148c72c65bSEric Dumazet return err; 14151da177e4SLinus Torvalds } 14161da177e4SLinus Torvalds 14172987babbSYuchung Cheng static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, 14182987babbSYuchung Cheng gfp_t gfp_mask) 14192987babbSYuchung Cheng { 14202987babbSYuchung Cheng return __tcp_transmit_skb(sk, skb, clone_it, gfp_mask, 14212987babbSYuchung Cheng tcp_sk(sk)->rcv_nxt); 14222987babbSYuchung Cheng } 14232987babbSYuchung Cheng 142467edfef7SAndi Kleen /* This routine just queues the buffer for sending. 14251da177e4SLinus Torvalds * 14261da177e4SLinus Torvalds * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, 14271da177e4SLinus Torvalds * otherwise socket can stall. 14281da177e4SLinus Torvalds */ 14291da177e4SLinus Torvalds static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) 14301da177e4SLinus Torvalds { 14311da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 14321da177e4SLinus Torvalds 14331da177e4SLinus Torvalds /* Advance write_seq and place onto the write_queue. */ 14340f317464SEric Dumazet WRITE_ONCE(tp->write_seq, TCP_SKB_CB(skb)->end_seq); 1435f4a775d1SEric Dumazet __skb_header_release(skb); 1436fe067e8aSDavid S. Miller tcp_add_write_queue_tail(sk, skb); 1437ab4e846aSEric Dumazet sk_wmem_queued_add(sk, skb->truesize); 14383ab224beSHideo Aoki sk_mem_charge(sk, skb->truesize); 14391da177e4SLinus Torvalds } 14401da177e4SLinus Torvalds 144167edfef7SAndi Kleen /* Initialize TSO segments for a packet. */ 14425bbb432cSEric Dumazet static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now) 1443f6302d1dSDavid S. Miller { 14444a64fd6cSEric Dumazet if (skb->len <= mss_now) { 1445f6302d1dSDavid S. Miller /* Avoid the costly divide in the normal 1446f6302d1dSDavid S. Miller * non-TSO case. 1447f6302d1dSDavid S. Miller */ 1448cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, 1); 1449f69ad292SEric Dumazet TCP_SKB_CB(skb)->tcp_gso_size = 0; 1450f6302d1dSDavid S. Miller } else { 1451cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now)); 1452f69ad292SEric Dumazet TCP_SKB_CB(skb)->tcp_gso_size = mss_now; 14531da177e4SLinus Torvalds } 14541da177e4SLinus Torvalds } 14551da177e4SLinus Torvalds 1456797108d1SIlpo Järvinen /* Pcount in the middle of the write queue got changed, we need to do various 1457797108d1SIlpo Järvinen * tweaks to fix counters 1458797108d1SIlpo Järvinen */ 1459cf533ea5SEric Dumazet static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) 1460797108d1SIlpo Järvinen { 1461797108d1SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1462797108d1SIlpo Järvinen 1463797108d1SIlpo Järvinen tp->packets_out -= decr; 1464797108d1SIlpo Järvinen 1465797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 1466797108d1SIlpo Järvinen tp->sacked_out -= decr; 1467797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 1468797108d1SIlpo Järvinen tp->retrans_out -= decr; 1469797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 1470797108d1SIlpo Järvinen tp->lost_out -= decr; 1471797108d1SIlpo Järvinen 1472797108d1SIlpo Järvinen /* Reno case is special. Sigh... */ 1473797108d1SIlpo Järvinen if (tcp_is_reno(tp) && decr > 0) 1474797108d1SIlpo Järvinen tp->sacked_out -= min_t(u32, tp->sacked_out, decr); 1475797108d1SIlpo Järvinen 1476797108d1SIlpo Järvinen if (tp->lost_skb_hint && 1477797108d1SIlpo Järvinen before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && 1478713bafeaSYuchung Cheng (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) 1479797108d1SIlpo Järvinen tp->lost_cnt_hint -= decr; 1480797108d1SIlpo Järvinen 1481797108d1SIlpo Järvinen tcp_verify_left_out(tp); 1482797108d1SIlpo Järvinen } 1483797108d1SIlpo Järvinen 14840a2cf20cSSoheil Hassas Yeganeh static bool tcp_has_tx_tstamp(const struct sk_buff *skb) 14850a2cf20cSSoheil Hassas Yeganeh { 14860a2cf20cSSoheil Hassas Yeganeh return TCP_SKB_CB(skb)->txstamp_ack || 14870a2cf20cSSoheil Hassas Yeganeh (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP); 14880a2cf20cSSoheil Hassas Yeganeh } 14890a2cf20cSSoheil Hassas Yeganeh 1490490cc7d0SWillem de Bruijn static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2) 1491490cc7d0SWillem de Bruijn { 1492490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo = skb_shinfo(skb); 1493490cc7d0SWillem de Bruijn 14940a2cf20cSSoheil Hassas Yeganeh if (unlikely(tcp_has_tx_tstamp(skb)) && 1495490cc7d0SWillem de Bruijn !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) { 1496490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo2 = skb_shinfo(skb2); 1497490cc7d0SWillem de Bruijn u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP; 1498490cc7d0SWillem de Bruijn 1499490cc7d0SWillem de Bruijn shinfo->tx_flags &= ~tsflags; 1500490cc7d0SWillem de Bruijn shinfo2->tx_flags |= tsflags; 1501490cc7d0SWillem de Bruijn swap(shinfo->tskey, shinfo2->tskey); 1502b51e13faSMartin KaFai Lau TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack; 1503b51e13faSMartin KaFai Lau TCP_SKB_CB(skb)->txstamp_ack = 0; 1504490cc7d0SWillem de Bruijn } 1505490cc7d0SWillem de Bruijn } 1506490cc7d0SWillem de Bruijn 1507a166140eSMartin KaFai Lau static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2) 1508a166140eSMartin KaFai Lau { 1509a166140eSMartin KaFai Lau TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor; 1510a166140eSMartin KaFai Lau TCP_SKB_CB(skb)->eor = 0; 1511a166140eSMartin KaFai Lau } 1512a166140eSMartin KaFai Lau 151375c119afSEric Dumazet /* Insert buff after skb on the write or rtx queue of sk. */ 151475c119afSEric Dumazet static void tcp_insert_write_queue_after(struct sk_buff *skb, 151575c119afSEric Dumazet struct sk_buff *buff, 151675c119afSEric Dumazet struct sock *sk, 151775c119afSEric Dumazet enum tcp_queue tcp_queue) 151875c119afSEric Dumazet { 151975c119afSEric Dumazet if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE) 152075c119afSEric Dumazet __skb_queue_after(&sk->sk_write_queue, skb, buff); 152175c119afSEric Dumazet else 152275c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, buff); 152375c119afSEric Dumazet } 152475c119afSEric Dumazet 15251da177e4SLinus Torvalds /* Function to create two new TCP segments. Shrinks the given segment 15261da177e4SLinus Torvalds * to the specified size and appends a new segment with the rest of the 15271da177e4SLinus Torvalds * packet to the list. This won't be called frequently, I hope. 15281da177e4SLinus Torvalds * Remember, these are still headerless SKBs at this point. 15291da177e4SLinus Torvalds */ 153075c119afSEric Dumazet int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue, 153175c119afSEric Dumazet struct sk_buff *skb, u32 len, 15326cc55e09SOctavian Purdila unsigned int mss_now, gfp_t gfp) 15331da177e4SLinus Torvalds { 15341da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 15351da177e4SLinus Torvalds struct sk_buff *buff; 15366475be16SDavid S. Miller int nsize, old_factor; 1537b617158dSEric Dumazet long limit; 1538b60b49eaSHerbert Xu int nlen; 15399ce01461SIlpo Järvinen u8 flags; 15401da177e4SLinus Torvalds 15412fceec13SIlpo Järvinen if (WARN_ON(len > skb->len)) 15422fceec13SIlpo Järvinen return -EINVAL; 15436a438bbeSStephen Hemminger 15441da177e4SLinus Torvalds nsize = skb_headlen(skb) - len; 15451da177e4SLinus Torvalds if (nsize < 0) 15461da177e4SLinus Torvalds nsize = 0; 15471da177e4SLinus Torvalds 1548b617158dSEric Dumazet /* tcp_sendmsg() can overshoot sk_wmem_queued by one full size skb. 1549b617158dSEric Dumazet * We need some allowance to not penalize applications setting small 1550b617158dSEric Dumazet * SO_SNDBUF values. 1551b617158dSEric Dumazet * Also allow first and last skb in retransmit queue to be split. 1552b617158dSEric Dumazet */ 15537c4e983cSAlexander Duyck limit = sk->sk_sndbuf + 2 * SKB_TRUESIZE(GSO_LEGACY_MAX_SIZE); 1554b617158dSEric Dumazet if (unlikely((sk->sk_wmem_queued >> 1) > limit && 1555b617158dSEric Dumazet tcp_queue != TCP_FRAG_IN_WRITE_QUEUE && 1556b617158dSEric Dumazet skb != tcp_rtx_queue_head(sk) && 1557b617158dSEric Dumazet skb != tcp_rtx_queue_tail(sk))) { 1558f070ef2aSEric Dumazet NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPWQUEUETOOBIG); 1559f070ef2aSEric Dumazet return -ENOMEM; 1560f070ef2aSEric Dumazet } 1561f070ef2aSEric Dumazet 1562c4777efaSEric Dumazet if (skb_unclone_keeptruesize(skb, gfp)) 15631da177e4SLinus Torvalds return -ENOMEM; 15641da177e4SLinus Torvalds 15651da177e4SLinus Torvalds /* Get a new skb... force flag on. */ 1566f8dd3b8dSEric Dumazet buff = tcp_stream_alloc_skb(sk, nsize, gfp, true); 156751456b29SIan Morris if (!buff) 15681da177e4SLinus Torvalds return -ENOMEM; /* We'll just try again later. */ 156941477662SJakub Kicinski skb_copy_decrypted(buff, skb); 15705a369ca6SPaolo Abeni mptcp_skb_ext_copy(buff, skb); 1571ef5cb973SHerbert Xu 1572ab4e846aSEric Dumazet sk_wmem_queued_add(sk, buff->truesize); 15733ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1574b60b49eaSHerbert Xu nlen = skb->len - len - nsize; 1575b60b49eaSHerbert Xu buff->truesize += nlen; 1576b60b49eaSHerbert Xu skb->truesize -= nlen; 15771da177e4SLinus Torvalds 15781da177e4SLinus Torvalds /* Correct the sequence numbers. */ 15791da177e4SLinus Torvalds TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 15801da177e4SLinus Torvalds TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 15811da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 15821da177e4SLinus Torvalds 15831da177e4SLinus Torvalds /* PSH and FIN should only be set in the second packet. */ 15844de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 15854de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 15864de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1587e14c3cafSHerbert Xu TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; 1588a166140eSMartin KaFai Lau tcp_skb_fragment_eor(skb, buff); 15891da177e4SLinus Torvalds 15901da177e4SLinus Torvalds skb_split(skb, buff, len); 15911da177e4SLinus Torvalds 1592a1ac9c8aSMartin KaFai Lau skb_set_delivery_time(buff, skb->tstamp, true); 1593490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 15941da177e4SLinus Torvalds 15956475be16SDavid S. Miller old_factor = tcp_skb_pcount(skb); 15966475be16SDavid S. Miller 15971da177e4SLinus Torvalds /* Fix up tso_factor for both original and new SKB. */ 15985bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 15995bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 16001da177e4SLinus Torvalds 1601b9f64820SYuchung Cheng /* Update delivered info for the new segment */ 1602b9f64820SYuchung Cheng TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx; 1603b9f64820SYuchung Cheng 16046475be16SDavid S. Miller /* If this packet has been sent out already, we must 16056475be16SDavid S. Miller * adjust the various packet counters. 16066475be16SDavid S. Miller */ 1607cf0b450cSHerbert Xu if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 16086475be16SDavid S. Miller int diff = old_factor - tcp_skb_pcount(skb) - 16096475be16SDavid S. Miller tcp_skb_pcount(buff); 16101da177e4SLinus Torvalds 1611797108d1SIlpo Järvinen if (diff) 1612797108d1SIlpo Järvinen tcp_adjust_pcount(sk, skb, diff); 16131da177e4SLinus Torvalds } 16141da177e4SLinus Torvalds 16151da177e4SLinus Torvalds /* Link BUFF into the send queue. */ 1616f4a775d1SEric Dumazet __skb_header_release(buff); 161775c119afSEric Dumazet tcp_insert_write_queue_after(skb, buff, sk, tcp_queue); 1618f67971e6SEric Dumazet if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE) 1619e2080072SEric Dumazet list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor); 16201da177e4SLinus Torvalds 16211da177e4SLinus Torvalds return 0; 16221da177e4SLinus Torvalds } 16231da177e4SLinus Torvalds 1624f4d01666SEric Dumazet /* This is similar to __pskb_pull_tail(). The difference is that pulled 1625f4d01666SEric Dumazet * data is not copied, but immediately discarded. 16261da177e4SLinus Torvalds */ 16277162fb24SEric Dumazet static int __pskb_trim_head(struct sk_buff *skb, int len) 16281da177e4SLinus Torvalds { 16297b7fc97aSEric Dumazet struct skb_shared_info *shinfo; 16301da177e4SLinus Torvalds int i, k, eat; 16311da177e4SLinus Torvalds 16324fa48bf3SEric Dumazet eat = min_t(int, len, skb_headlen(skb)); 16334fa48bf3SEric Dumazet if (eat) { 16344fa48bf3SEric Dumazet __skb_pull(skb, eat); 16354fa48bf3SEric Dumazet len -= eat; 16364fa48bf3SEric Dumazet if (!len) 16377162fb24SEric Dumazet return 0; 16384fa48bf3SEric Dumazet } 16391da177e4SLinus Torvalds eat = len; 16401da177e4SLinus Torvalds k = 0; 16417b7fc97aSEric Dumazet shinfo = skb_shinfo(skb); 16427b7fc97aSEric Dumazet for (i = 0; i < shinfo->nr_frags; i++) { 16437b7fc97aSEric Dumazet int size = skb_frag_size(&shinfo->frags[i]); 16449e903e08SEric Dumazet 16459e903e08SEric Dumazet if (size <= eat) { 1646aff65da0SIan Campbell skb_frag_unref(skb, i); 16479e903e08SEric Dumazet eat -= size; 16481da177e4SLinus Torvalds } else { 16497b7fc97aSEric Dumazet shinfo->frags[k] = shinfo->frags[i]; 16501da177e4SLinus Torvalds if (eat) { 1651b54c9d5bSJonathan Lemon skb_frag_off_add(&shinfo->frags[k], eat); 16527b7fc97aSEric Dumazet skb_frag_size_sub(&shinfo->frags[k], eat); 16531da177e4SLinus Torvalds eat = 0; 16541da177e4SLinus Torvalds } 16551da177e4SLinus Torvalds k++; 16561da177e4SLinus Torvalds } 16571da177e4SLinus Torvalds } 16587b7fc97aSEric Dumazet shinfo->nr_frags = k; 16591da177e4SLinus Torvalds 16601da177e4SLinus Torvalds skb->data_len -= len; 16611da177e4SLinus Torvalds skb->len = skb->data_len; 16627162fb24SEric Dumazet return len; 16631da177e4SLinus Torvalds } 16641da177e4SLinus Torvalds 166567edfef7SAndi Kleen /* Remove acked data from a packet in the transmit queue. */ 16661da177e4SLinus Torvalds int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 16671da177e4SLinus Torvalds { 16687162fb24SEric Dumazet u32 delta_truesize; 16697162fb24SEric Dumazet 1670c4777efaSEric Dumazet if (skb_unclone_keeptruesize(skb, GFP_ATOMIC)) 16711da177e4SLinus Torvalds return -ENOMEM; 16721da177e4SLinus Torvalds 16737162fb24SEric Dumazet delta_truesize = __pskb_trim_head(skb, len); 16741da177e4SLinus Torvalds 16751da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq += len; 16761da177e4SLinus Torvalds 16777162fb24SEric Dumazet if (delta_truesize) { 16787162fb24SEric Dumazet skb->truesize -= delta_truesize; 1679ab4e846aSEric Dumazet sk_wmem_queued_add(sk, -delta_truesize); 16809b65b17dSTalal Ahmad if (!skb_zcopy_pure(skb)) 16817162fb24SEric Dumazet sk_mem_uncharge(sk, delta_truesize); 16827162fb24SEric Dumazet } 16831da177e4SLinus Torvalds 16845b35e1e6SNeal Cardwell /* Any change of skb->len requires recalculation of tso factor. */ 16851da177e4SLinus Torvalds if (tcp_skb_pcount(skb) > 1) 16865bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb)); 16871da177e4SLinus Torvalds 16881da177e4SLinus Torvalds return 0; 16891da177e4SLinus Torvalds } 16901da177e4SLinus Torvalds 16911b63edd6SYuchung Cheng /* Calculate MSS not accounting any TCP options. */ 16921b63edd6SYuchung Cheng static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu) 16935d424d5aSJohn Heffner { 1694cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1695cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 16965d424d5aSJohn Heffner int mss_now; 16975d424d5aSJohn Heffner 16985d424d5aSJohn Heffner /* Calculate base mss without TCP options: 16995d424d5aSJohn Heffner It is MMS_S - sizeof(tcphdr) of rfc1122 17005d424d5aSJohn Heffner */ 17015d424d5aSJohn Heffner mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 17025d424d5aSJohn Heffner 170367469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 170467469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 170567469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 170667469601SEric Dumazet 170767469601SEric Dumazet if (dst && dst_allfrag(dst)) 170867469601SEric Dumazet mss_now -= icsk->icsk_af_ops->net_frag_header_len; 170967469601SEric Dumazet } 171067469601SEric Dumazet 17115d424d5aSJohn Heffner /* Clamp it (mss_clamp does not include tcp options) */ 17125d424d5aSJohn Heffner if (mss_now > tp->rx_opt.mss_clamp) 17135d424d5aSJohn Heffner mss_now = tp->rx_opt.mss_clamp; 17145d424d5aSJohn Heffner 17155d424d5aSJohn Heffner /* Now subtract optional transport overhead */ 17165d424d5aSJohn Heffner mss_now -= icsk->icsk_ext_hdr_len; 17175d424d5aSJohn Heffner 17185d424d5aSJohn Heffner /* Then reserve room for full set of TCP options and 8 bytes of data */ 171978eb166cSKuniyuki Iwashima mss_now = max(mss_now, 172078eb166cSKuniyuki Iwashima READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_snd_mss)); 17215d424d5aSJohn Heffner return mss_now; 17225d424d5aSJohn Heffner } 17235d424d5aSJohn Heffner 17241b63edd6SYuchung Cheng /* Calculate MSS. Not accounting for SACKs here. */ 17251b63edd6SYuchung Cheng int tcp_mtu_to_mss(struct sock *sk, int pmtu) 17261b63edd6SYuchung Cheng { 17271b63edd6SYuchung Cheng /* Subtract TCP options size, not including SACKs */ 17281b63edd6SYuchung Cheng return __tcp_mtu_to_mss(sk, pmtu) - 17291b63edd6SYuchung Cheng (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr)); 17301b63edd6SYuchung Cheng } 1731c7bb4b89SEric Dumazet EXPORT_SYMBOL(tcp_mtu_to_mss); 17321b63edd6SYuchung Cheng 17335d424d5aSJohn Heffner /* Inverse of above */ 173467469601SEric Dumazet int tcp_mss_to_mtu(struct sock *sk, int mss) 17355d424d5aSJohn Heffner { 1736cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1737cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 17385d424d5aSJohn Heffner int mtu; 17395d424d5aSJohn Heffner 17405d424d5aSJohn Heffner mtu = mss + 17415d424d5aSJohn Heffner tp->tcp_header_len + 17425d424d5aSJohn Heffner icsk->icsk_ext_hdr_len + 17435d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 17445d424d5aSJohn Heffner 174567469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 174667469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 174767469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 174867469601SEric Dumazet 174967469601SEric Dumazet if (dst && dst_allfrag(dst)) 175067469601SEric Dumazet mtu += icsk->icsk_af_ops->net_frag_header_len; 175167469601SEric Dumazet } 17525d424d5aSJohn Heffner return mtu; 17535d424d5aSJohn Heffner } 1754556c6b46SNeal Cardwell EXPORT_SYMBOL(tcp_mss_to_mtu); 17555d424d5aSJohn Heffner 175667edfef7SAndi Kleen /* MTU probing init per socket */ 17575d424d5aSJohn Heffner void tcp_mtup_init(struct sock *sk) 17585d424d5aSJohn Heffner { 17595d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 17605d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 1761b0f9ca53SFan Du struct net *net = sock_net(sk); 17625d424d5aSJohn Heffner 1763f47d00e0SKuniyuki Iwashima icsk->icsk_mtup.enabled = READ_ONCE(net->ipv4.sysctl_tcp_mtu_probing) > 1; 17645d424d5aSJohn Heffner icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 17655d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 176688d78bc0SKuniyuki Iwashima icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, READ_ONCE(net->ipv4.sysctl_tcp_base_mss)); 17675d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 176805cbc0dbSFan Du if (icsk->icsk_mtup.enabled) 1769c74df29aSEric Dumazet icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; 17705d424d5aSJohn Heffner } 17714bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_mtup_init); 17725d424d5aSJohn Heffner 17731da177e4SLinus Torvalds /* This function synchronize snd mss to current pmtu/exthdr set. 17741da177e4SLinus Torvalds 17751da177e4SLinus Torvalds tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 17761da177e4SLinus Torvalds for TCP options, but includes only bare TCP header. 17771da177e4SLinus Torvalds 17781da177e4SLinus Torvalds tp->rx_opt.mss_clamp is mss negotiated at connection setup. 1779caa20d9aSStephen Hemminger It is minimum of user_mss and mss received with SYN. 17801da177e4SLinus Torvalds It also does not include TCP options. 17811da177e4SLinus Torvalds 1782d83d8461SArnaldo Carvalho de Melo inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. 17831da177e4SLinus Torvalds 17841da177e4SLinus Torvalds tp->mss_cache is current effective sending mss, including 17851da177e4SLinus Torvalds all tcp options except for SACKs. It is evaluated, 17861da177e4SLinus Torvalds taking into account current pmtu, but never exceeds 17871da177e4SLinus Torvalds tp->rx_opt.mss_clamp. 17881da177e4SLinus Torvalds 17891da177e4SLinus Torvalds NOTE1. rfc1122 clearly states that advertised MSS 17901da177e4SLinus Torvalds DOES NOT include either tcp or ip options. 17911da177e4SLinus Torvalds 1792d83d8461SArnaldo Carvalho de Melo NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache 1793d83d8461SArnaldo Carvalho de Melo are READ ONLY outside this function. --ANK (980731) 17941da177e4SLinus Torvalds */ 17951da177e4SLinus Torvalds unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 17961da177e4SLinus Torvalds { 17971da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1798d83d8461SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 17995d424d5aSJohn Heffner int mss_now; 18001da177e4SLinus Torvalds 18015d424d5aSJohn Heffner if (icsk->icsk_mtup.search_high > pmtu) 18025d424d5aSJohn Heffner icsk->icsk_mtup.search_high = pmtu; 18031da177e4SLinus Torvalds 18045d424d5aSJohn Heffner mss_now = tcp_mtu_to_mss(sk, pmtu); 1805409d22b4SIlpo Järvinen mss_now = tcp_bound_to_half_wnd(tp, mss_now); 18061da177e4SLinus Torvalds 18071da177e4SLinus Torvalds /* And store cached results */ 1808d83d8461SArnaldo Carvalho de Melo icsk->icsk_pmtu_cookie = pmtu; 18095d424d5aSJohn Heffner if (icsk->icsk_mtup.enabled) 18105d424d5aSJohn Heffner mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); 1811c1b4a7e6SDavid S. Miller tp->mss_cache = mss_now; 18121da177e4SLinus Torvalds 18131da177e4SLinus Torvalds return mss_now; 18141da177e4SLinus Torvalds } 18154bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_sync_mss); 18161da177e4SLinus Torvalds 18171da177e4SLinus Torvalds /* Compute the current effective MSS, taking SACKs and IP options, 18181da177e4SLinus Torvalds * and even PMTU discovery events into account. 18191da177e4SLinus Torvalds */ 18200c54b85fSIlpo Järvinen unsigned int tcp_current_mss(struct sock *sk) 18211da177e4SLinus Torvalds { 1822cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1823cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1824c1b4a7e6SDavid S. Miller u32 mss_now; 182595c96174SEric Dumazet unsigned int header_len; 182633ad798cSAdam Langley struct tcp_out_options opts; 182733ad798cSAdam Langley struct tcp_md5sig_key *md5; 18281da177e4SLinus Torvalds 1829c1b4a7e6SDavid S. Miller mss_now = tp->mss_cache; 1830c1b4a7e6SDavid S. Miller 18311da177e4SLinus Torvalds if (dst) { 18321da177e4SLinus Torvalds u32 mtu = dst_mtu(dst); 1833d83d8461SArnaldo Carvalho de Melo if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 18341da177e4SLinus Torvalds mss_now = tcp_sync_mss(sk, mtu); 18351da177e4SLinus Torvalds } 18361da177e4SLinus Torvalds 183733ad798cSAdam Langley header_len = tcp_established_options(sk, NULL, &opts, &md5) + 183833ad798cSAdam Langley sizeof(struct tcphdr); 183933ad798cSAdam Langley /* The mss_cache is sized based on tp->tcp_header_len, which assumes 184033ad798cSAdam Langley * some common options. If this is an odd packet (because we have SACK 184133ad798cSAdam Langley * blocks etc) then our calculated header_len will be different, and 184233ad798cSAdam Langley * we have to adjust mss_now correspondingly */ 184333ad798cSAdam Langley if (header_len != tp->tcp_header_len) { 184433ad798cSAdam Langley int delta = (int) header_len - tp->tcp_header_len; 184533ad798cSAdam Langley mss_now -= delta; 184633ad798cSAdam Langley } 1847cfb6eeb4SYOSHIFUJI Hideaki 18481da177e4SLinus Torvalds return mss_now; 18491da177e4SLinus Torvalds } 18501da177e4SLinus Torvalds 185186fd14adSWeiping Pan /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. 185286fd14adSWeiping Pan * As additional protections, we do not touch cwnd in retransmission phases, 185386fd14adSWeiping Pan * and if application hit its sndbuf limit recently. 185486fd14adSWeiping Pan */ 185586fd14adSWeiping Pan static void tcp_cwnd_application_limited(struct sock *sk) 1856a762a980SDavid S. Miller { 18579e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1858a762a980SDavid S. Miller 185986fd14adSWeiping Pan if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && 186086fd14adSWeiping Pan sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 186186fd14adSWeiping Pan /* Limited by application or receiver window. */ 186286fd14adSWeiping Pan u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); 186386fd14adSWeiping Pan u32 win_used = max(tp->snd_cwnd_used, init_win); 186440570375SEric Dumazet if (win_used < tcp_snd_cwnd(tp)) { 186586fd14adSWeiping Pan tp->snd_ssthresh = tcp_current_ssthresh(sk); 186640570375SEric Dumazet tcp_snd_cwnd_set(tp, (tcp_snd_cwnd(tp) + win_used) >> 1); 186786fd14adSWeiping Pan } 186886fd14adSWeiping Pan tp->snd_cwnd_used = 0; 186986fd14adSWeiping Pan } 1870c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 187186fd14adSWeiping Pan } 187286fd14adSWeiping Pan 1873ca8a2263SNeal Cardwell static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited) 1874a762a980SDavid S. Miller { 18751b1fc3fdSWei Wang const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 1876a762a980SDavid S. Miller struct tcp_sock *tp = tcp_sk(sk); 1877a762a980SDavid S. Miller 1878ca8a2263SNeal Cardwell /* Track the maximum number of outstanding packets in each 1879ca8a2263SNeal Cardwell * window, and remember whether we were cwnd-limited then. 1880ca8a2263SNeal Cardwell */ 1881ca8a2263SNeal Cardwell if (!before(tp->snd_una, tp->max_packets_seq) || 1882299bcb55SNeal Cardwell tp->packets_out > tp->max_packets_out || 1883299bcb55SNeal Cardwell is_cwnd_limited) { 1884ca8a2263SNeal Cardwell tp->max_packets_out = tp->packets_out; 1885ca8a2263SNeal Cardwell tp->max_packets_seq = tp->snd_nxt; 1886ca8a2263SNeal Cardwell tp->is_cwnd_limited = is_cwnd_limited; 1887ca8a2263SNeal Cardwell } 1888e114a710SEric Dumazet 188924901551SEric Dumazet if (tcp_is_cwnd_limited(sk)) { 1890a762a980SDavid S. Miller /* Network is feed fully. */ 1891a762a980SDavid S. Miller tp->snd_cwnd_used = 0; 1892c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 1893a762a980SDavid S. Miller } else { 1894a762a980SDavid S. Miller /* Network starves. */ 1895a762a980SDavid S. Miller if (tp->packets_out > tp->snd_cwnd_used) 1896a762a980SDavid S. Miller tp->snd_cwnd_used = tp->packets_out; 1897a762a980SDavid S. Miller 18984845b571SKuniyuki Iwashima if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle) && 1899c2203cf7SEric Dumazet (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto && 19001b1fc3fdSWei Wang !ca_ops->cong_control) 1901a762a980SDavid S. Miller tcp_cwnd_application_limited(sk); 1902b0f71bd3SFrancis Yan 1903b0f71bd3SFrancis Yan /* The following conditions together indicate the starvation 1904b0f71bd3SFrancis Yan * is caused by insufficient sender buffer: 1905b0f71bd3SFrancis Yan * 1) just sent some data (see tcp_write_xmit) 1906b0f71bd3SFrancis Yan * 2) not cwnd limited (this else condition) 190775c119afSEric Dumazet * 3) no more data to send (tcp_write_queue_empty()) 1908b0f71bd3SFrancis Yan * 4) application is hitting buffer limit (SOCK_NOSPACE) 1909b0f71bd3SFrancis Yan */ 191075c119afSEric Dumazet if (tcp_write_queue_empty(sk) && sk->sk_socket && 1911b0f71bd3SFrancis Yan test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) && 1912b0f71bd3SFrancis Yan (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) 1913b0f71bd3SFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED); 1914a762a980SDavid S. Miller } 1915a762a980SDavid S. Miller } 1916a762a980SDavid S. Miller 1917d4589926SEric Dumazet /* Minshall's variant of the Nagle send check. */ 1918d4589926SEric Dumazet static bool tcp_minshall_check(const struct tcp_sock *tp) 1919d4589926SEric Dumazet { 1920d4589926SEric Dumazet return after(tp->snd_sml, tp->snd_una) && 1921d4589926SEric Dumazet !after(tp->snd_sml, tp->snd_nxt); 1922d4589926SEric Dumazet } 1923d4589926SEric Dumazet 1924d4589926SEric Dumazet /* Update snd_sml if this skb is under mss 1925d4589926SEric Dumazet * Note that a TSO packet might end with a sub-mss segment 1926d4589926SEric Dumazet * The test is really : 1927d4589926SEric Dumazet * if ((skb->len % mss) != 0) 1928d4589926SEric Dumazet * tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1929d4589926SEric Dumazet * But we can avoid doing the divide again given we already have 1930d4589926SEric Dumazet * skb_pcount = skb->len / mss_now 19310e3a4803SIlpo Järvinen */ 1932d4589926SEric Dumazet static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now, 1933d4589926SEric Dumazet const struct sk_buff *skb) 1934d4589926SEric Dumazet { 1935d4589926SEric Dumazet if (skb->len < tcp_skb_pcount(skb) * mss_now) 1936d4589926SEric Dumazet tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1937d4589926SEric Dumazet } 1938d4589926SEric Dumazet 1939d4589926SEric Dumazet /* Return false, if packet can be sent now without violation Nagle's rules: 1940d4589926SEric Dumazet * 1. It is full sized. (provided by caller in %partial bool) 1941d4589926SEric Dumazet * 2. Or it contains FIN. (already checked by caller) 1942d4589926SEric Dumazet * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. 1943d4589926SEric Dumazet * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 1944d4589926SEric Dumazet * With Minshall's modification: all sent small packets are ACKed. 1945d4589926SEric Dumazet */ 1946d4589926SEric Dumazet static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp, 1947cc93fc51SPeter Pan(潘卫平) int nonagle) 1948d4589926SEric Dumazet { 1949d4589926SEric Dumazet return partial && 1950d4589926SEric Dumazet ((nonagle & TCP_NAGLE_CORK) || 1951d4589926SEric Dumazet (!nonagle && tp->packets_out && tcp_minshall_check(tp))); 1952d4589926SEric Dumazet } 1953605ad7f1SEric Dumazet 1954605ad7f1SEric Dumazet /* Return how many segs we'd like on a TSO packet, 195565466904SEric Dumazet * depending on current pacing rate, and how close the peer is. 195665466904SEric Dumazet * 195765466904SEric Dumazet * Rationale is: 195865466904SEric Dumazet * - For close peers, we rather send bigger packets to reduce 195965466904SEric Dumazet * cpu costs, because occasional losses will be repaired fast. 196065466904SEric Dumazet * - For long distance/rtt flows, we would like to get ACK clocking 196165466904SEric Dumazet * with 1 ACK per ms. 196265466904SEric Dumazet * 196365466904SEric Dumazet * Use min_rtt to help adapt TSO burst size, with smaller min_rtt resulting 196465466904SEric Dumazet * in bigger TSO bursts. We we cut the RTT-based allowance in half 196565466904SEric Dumazet * for every 2^9 usec (aka 512 us) of RTT, so that the RTT-based allowance 196665466904SEric Dumazet * is below 1500 bytes after 6 * ~500 usec = 3ms. 1967605ad7f1SEric Dumazet */ 1968dcb8c9b4SEric Dumazet static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now, 19691b3878caSNeal Cardwell int min_tso_segs) 1970605ad7f1SEric Dumazet { 197165466904SEric Dumazet unsigned long bytes; 197265466904SEric Dumazet u32 r; 1973605ad7f1SEric Dumazet 197465466904SEric Dumazet bytes = sk->sk_pacing_rate >> READ_ONCE(sk->sk_pacing_shift); 1975605ad7f1SEric Dumazet 19762455e61bSKuniyuki Iwashima r = tcp_min_rtt(tcp_sk(sk)) >> READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_rtt_log); 197765466904SEric Dumazet if (r < BITS_PER_TYPE(sk->sk_gso_max_size)) 197865466904SEric Dumazet bytes += sk->sk_gso_max_size >> r; 1979605ad7f1SEric Dumazet 198065466904SEric Dumazet bytes = min_t(unsigned long, bytes, sk->sk_gso_max_size); 198165466904SEric Dumazet 198265466904SEric Dumazet return max_t(u32, bytes / mss_now, min_tso_segs); 1983605ad7f1SEric Dumazet } 1984605ad7f1SEric Dumazet 1985ed6e7268SNeal Cardwell /* Return the number of segments we want in the skb we are transmitting. 1986ed6e7268SNeal Cardwell * See if congestion control module wants to decide; otherwise, autosize. 1987ed6e7268SNeal Cardwell */ 1988ed6e7268SNeal Cardwell static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now) 1989ed6e7268SNeal Cardwell { 1990ed6e7268SNeal Cardwell const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 1991dcb8c9b4SEric Dumazet u32 min_tso, tso_segs; 1992ed6e7268SNeal Cardwell 1993dcb8c9b4SEric Dumazet min_tso = ca_ops->min_tso_segs ? 1994dcb8c9b4SEric Dumazet ca_ops->min_tso_segs(sk) : 1995e0bb4ab9SKuniyuki Iwashima READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs); 1996dcb8c9b4SEric Dumazet 1997dcb8c9b4SEric Dumazet tso_segs = tcp_tso_autosize(sk, mss_now, min_tso); 1998350c9f48SEric Dumazet return min_t(u32, tso_segs, sk->sk_gso_max_segs); 1999ed6e7268SNeal Cardwell } 2000ed6e7268SNeal Cardwell 2001d4589926SEric Dumazet /* Returns the portion of skb which can be sent right away */ 2002d4589926SEric Dumazet static unsigned int tcp_mss_split_point(const struct sock *sk, 2003d4589926SEric Dumazet const struct sk_buff *skb, 2004d4589926SEric Dumazet unsigned int mss_now, 2005d4589926SEric Dumazet unsigned int max_segs, 2006d4589926SEric Dumazet int nonagle) 2007c1b4a7e6SDavid S. Miller { 2008cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 2009d4589926SEric Dumazet u32 partial, needed, window, max_len; 2010c1b4a7e6SDavid S. Miller 201190840defSIlpo Järvinen window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 20121485348dSBen Hutchings max_len = mss_now * max_segs; 20130e3a4803SIlpo Järvinen 20141485348dSBen Hutchings if (likely(max_len <= window && skb != tcp_write_queue_tail(sk))) 20151485348dSBen Hutchings return max_len; 20160e3a4803SIlpo Järvinen 20175ea3a748SIlpo Järvinen needed = min(skb->len, window); 20185ea3a748SIlpo Järvinen 20191485348dSBen Hutchings if (max_len <= needed) 20201485348dSBen Hutchings return max_len; 20210e3a4803SIlpo Järvinen 2022d4589926SEric Dumazet partial = needed % mss_now; 2023d4589926SEric Dumazet /* If last segment is not a full MSS, check if Nagle rules allow us 2024d4589926SEric Dumazet * to include this last segment in this skb. 2025d4589926SEric Dumazet * Otherwise, we'll split the skb at last MSS boundary 2026d4589926SEric Dumazet */ 2027cc93fc51SPeter Pan(潘卫平) if (tcp_nagle_check(partial != 0, tp, nonagle)) 2028d4589926SEric Dumazet return needed - partial; 2029d4589926SEric Dumazet 2030d4589926SEric Dumazet return needed; 2031c1b4a7e6SDavid S. Miller } 2032c1b4a7e6SDavid S. Miller 2033c1b4a7e6SDavid S. Miller /* Can at least one segment of SKB be sent right now, according to the 2034c1b4a7e6SDavid S. Miller * congestion window rules? If so, return how many segments are allowed. 2035c1b4a7e6SDavid S. Miller */ 2036cf533ea5SEric Dumazet static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, 2037cf533ea5SEric Dumazet const struct sk_buff *skb) 2038c1b4a7e6SDavid S. Miller { 2039d649a7a8SEric Dumazet u32 in_flight, cwnd, halfcwnd; 2040c1b4a7e6SDavid S. Miller 2041c1b4a7e6SDavid S. Miller /* Don't be strict about the congestion window for the final FIN. */ 20424de075e0SEric Dumazet if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && 20434de075e0SEric Dumazet tcp_skb_pcount(skb) == 1) 2044c1b4a7e6SDavid S. Miller return 1; 2045c1b4a7e6SDavid S. Miller 2046c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 204740570375SEric Dumazet cwnd = tcp_snd_cwnd(tp); 2048d649a7a8SEric Dumazet if (in_flight >= cwnd) 2049c1b4a7e6SDavid S. Miller return 0; 2050d649a7a8SEric Dumazet 2051d649a7a8SEric Dumazet /* For better scheduling, ensure we have at least 2052d649a7a8SEric Dumazet * 2 GSO packets in flight. 2053d649a7a8SEric Dumazet */ 2054d649a7a8SEric Dumazet halfcwnd = max(cwnd >> 1, 1U); 2055d649a7a8SEric Dumazet return min(halfcwnd, cwnd - in_flight); 2056c1b4a7e6SDavid S. Miller } 2057c1b4a7e6SDavid S. Miller 2058b595076aSUwe Kleine-König /* Initialize TSO state of a skb. 205967edfef7SAndi Kleen * This must be invoked the first time we consider transmitting 2060c1b4a7e6SDavid S. Miller * SKB onto the wire. 2061c1b4a7e6SDavid S. Miller */ 20625bbb432cSEric Dumazet static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now) 2063c1b4a7e6SDavid S. Miller { 2064c1b4a7e6SDavid S. Miller int tso_segs = tcp_skb_pcount(skb); 2065c1b4a7e6SDavid S. Miller 2066f8269a49SIlpo Järvinen if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { 20675bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 2068c1b4a7e6SDavid S. Miller tso_segs = tcp_skb_pcount(skb); 2069c1b4a7e6SDavid S. Miller } 2070c1b4a7e6SDavid S. Miller return tso_segs; 2071c1b4a7e6SDavid S. Miller } 2072c1b4a7e6SDavid S. Miller 2073c1b4a7e6SDavid S. Miller 2074a2a385d6SEric Dumazet /* Return true if the Nagle test allows this packet to be 2075c1b4a7e6SDavid S. Miller * sent now. 2076c1b4a7e6SDavid S. Miller */ 2077a2a385d6SEric Dumazet static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, 2078c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 2079c1b4a7e6SDavid S. Miller { 2080c1b4a7e6SDavid S. Miller /* Nagle rule does not apply to frames, which sit in the middle of the 2081c1b4a7e6SDavid S. Miller * write_queue (they have no chances to get new data). 2082c1b4a7e6SDavid S. Miller * 2083c1b4a7e6SDavid S. Miller * This is implemented in the callers, where they modify the 'nonagle' 2084c1b4a7e6SDavid S. Miller * argument based upon the location of SKB in the send queue. 2085c1b4a7e6SDavid S. Miller */ 2086c1b4a7e6SDavid S. Miller if (nonagle & TCP_NAGLE_PUSH) 2087a2a385d6SEric Dumazet return true; 2088c1b4a7e6SDavid S. Miller 20899b44190dSYuchung Cheng /* Don't use the nagle rule for urgent data (or for the final FIN). */ 20909b44190dSYuchung Cheng if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) 2091a2a385d6SEric Dumazet return true; 2092c1b4a7e6SDavid S. Miller 2093cc93fc51SPeter Pan(潘卫平) if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle)) 2094a2a385d6SEric Dumazet return true; 2095c1b4a7e6SDavid S. Miller 2096a2a385d6SEric Dumazet return false; 2097c1b4a7e6SDavid S. Miller } 2098c1b4a7e6SDavid S. Miller 2099c1b4a7e6SDavid S. Miller /* Does at least the first segment of SKB fit into the send window? */ 2100a2a385d6SEric Dumazet static bool tcp_snd_wnd_test(const struct tcp_sock *tp, 2101a2a385d6SEric Dumazet const struct sk_buff *skb, 2102056834d9SIlpo Järvinen unsigned int cur_mss) 2103c1b4a7e6SDavid S. Miller { 2104c1b4a7e6SDavid S. Miller u32 end_seq = TCP_SKB_CB(skb)->end_seq; 2105c1b4a7e6SDavid S. Miller 2106c1b4a7e6SDavid S. Miller if (skb->len > cur_mss) 2107c1b4a7e6SDavid S. Miller end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 2108c1b4a7e6SDavid S. Miller 210990840defSIlpo Järvinen return !after(end_seq, tcp_wnd_end(tp)); 2110c1b4a7e6SDavid S. Miller } 2111c1b4a7e6SDavid S. Miller 2112c1b4a7e6SDavid S. Miller /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 2113c1b4a7e6SDavid S. Miller * which is put after SKB on the list. It is very much like 2114c1b4a7e6SDavid S. Miller * tcp_fragment() except that it may make several kinds of assumptions 2115c1b4a7e6SDavid S. Miller * in order to speed up the splitting operation. In particular, we 2116c1b4a7e6SDavid S. Miller * know that all the data is in scatter-gather pages, and that the 2117c1b4a7e6SDavid S. Miller * packet has never been sent out before (and thus is not cloned). 2118c1b4a7e6SDavid S. Miller */ 211956483341SEric Dumazet static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, 2120c4ead4c5SEric Dumazet unsigned int mss_now, gfp_t gfp) 2121c1b4a7e6SDavid S. Miller { 2122c1b4a7e6SDavid S. Miller int nlen = skb->len - len; 212356483341SEric Dumazet struct sk_buff *buff; 21249ce01461SIlpo Järvinen u8 flags; 2125c1b4a7e6SDavid S. Miller 2126c1b4a7e6SDavid S. Miller /* All of a TSO frame must be composed of paged data. */ 2127c8ac3774SHerbert Xu if (skb->len != skb->data_len) 212856483341SEric Dumazet return tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE, 212956483341SEric Dumazet skb, len, mss_now, gfp); 2130c1b4a7e6SDavid S. Miller 2131f8dd3b8dSEric Dumazet buff = tcp_stream_alloc_skb(sk, 0, gfp, true); 213251456b29SIan Morris if (unlikely(!buff)) 2133c1b4a7e6SDavid S. Miller return -ENOMEM; 213441477662SJakub Kicinski skb_copy_decrypted(buff, skb); 21355a369ca6SPaolo Abeni mptcp_skb_ext_copy(buff, skb); 2136c1b4a7e6SDavid S. Miller 2137ab4e846aSEric Dumazet sk_wmem_queued_add(sk, buff->truesize); 21383ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 2139b60b49eaSHerbert Xu buff->truesize += nlen; 2140c1b4a7e6SDavid S. Miller skb->truesize -= nlen; 2141c1b4a7e6SDavid S. Miller 2142c1b4a7e6SDavid S. Miller /* Correct the sequence numbers. */ 2143c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 2144c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 2145c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 2146c1b4a7e6SDavid S. Miller 2147c1b4a7e6SDavid S. Miller /* PSH and FIN should only be set in the second packet. */ 21484de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 21494de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 21504de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 2151c1b4a7e6SDavid S. Miller 2152a166140eSMartin KaFai Lau tcp_skb_fragment_eor(skb, buff); 2153a166140eSMartin KaFai Lau 2154c1b4a7e6SDavid S. Miller skb_split(skb, buff, len); 2155490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 2156c1b4a7e6SDavid S. Miller 2157c1b4a7e6SDavid S. Miller /* Fix up tso_factor for both original and new SKB. */ 21585bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 21595bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 2160c1b4a7e6SDavid S. Miller 2161c1b4a7e6SDavid S. Miller /* Link BUFF into the send queue. */ 2162f4a775d1SEric Dumazet __skb_header_release(buff); 216356483341SEric Dumazet tcp_insert_write_queue_after(skb, buff, sk, TCP_FRAG_IN_WRITE_QUEUE); 2164c1b4a7e6SDavid S. Miller 2165c1b4a7e6SDavid S. Miller return 0; 2166c1b4a7e6SDavid S. Miller } 2167c1b4a7e6SDavid S. Miller 2168c1b4a7e6SDavid S. Miller /* Try to defer sending, if possible, in order to minimize the amount 2169c1b4a7e6SDavid S. Miller * of TSO splitting we do. View it as a kind of TSO Nagle test. 2170c1b4a7e6SDavid S. Miller * 2171c1b4a7e6SDavid S. Miller * This algorithm is from John Heffner. 2172c1b4a7e6SDavid S. Miller */ 2173ca8a2263SNeal Cardwell static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb, 2174f9bfe4e6SEric Dumazet bool *is_cwnd_limited, 2175f9bfe4e6SEric Dumazet bool *is_rwnd_limited, 2176f9bfe4e6SEric Dumazet u32 max_segs) 2177c1b4a7e6SDavid S. Miller { 21786687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 2179f1c6ea38SEric Dumazet u32 send_win, cong_win, limit, in_flight; 218050c8339eSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 218150c8339eSEric Dumazet struct sk_buff *head; 2182ad9f4f50SEric Dumazet int win_divisor; 2183f1c6ea38SEric Dumazet s64 delta; 2184c1b4a7e6SDavid S. Miller 218599d7662aSEric Dumazet if (icsk->icsk_ca_state >= TCP_CA_Recovery) 2186ae8064acSJohn Heffner goto send_now; 2187ae8064acSJohn Heffner 21885f852eb5SEric Dumazet /* Avoid bursty behavior by allowing defer 2189a682850aSEric Dumazet * only if the last write was recent (1 ms). 2190a682850aSEric Dumazet * Note that tp->tcp_wstamp_ns can be in the future if we have 2191a682850aSEric Dumazet * packets waiting in a qdisc or device for EDT delivery. 21925f852eb5SEric Dumazet */ 2193a682850aSEric Dumazet delta = tp->tcp_clock_cache - tp->tcp_wstamp_ns - NSEC_PER_MSEC; 2194a682850aSEric Dumazet if (delta > 0) 2195ae8064acSJohn Heffner goto send_now; 2196908a75c1SDavid S. Miller 2197c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 2198c1b4a7e6SDavid S. Miller 2199c8c9aeb5SStefano Brivio BUG_ON(tcp_skb_pcount(skb) <= 1); 220040570375SEric Dumazet BUG_ON(tcp_snd_cwnd(tp) <= in_flight); 2201c1b4a7e6SDavid S. Miller 220290840defSIlpo Järvinen send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 2203c1b4a7e6SDavid S. Miller 2204c1b4a7e6SDavid S. Miller /* From in_flight test above, we know that cwnd > in_flight. */ 220540570375SEric Dumazet cong_win = (tcp_snd_cwnd(tp) - in_flight) * tp->mss_cache; 2206c1b4a7e6SDavid S. Miller 2207c1b4a7e6SDavid S. Miller limit = min(send_win, cong_win); 2208c1b4a7e6SDavid S. Miller 2209ba244fe9SDavid S. Miller /* If a full-sized TSO skb can be sent, do it. */ 2210605ad7f1SEric Dumazet if (limit >= max_segs * tp->mss_cache) 2211ae8064acSJohn Heffner goto send_now; 2212ba244fe9SDavid S. Miller 221362ad2761SIlpo Järvinen /* Middle in queue won't get any more data, full sendable already? */ 221462ad2761SIlpo Järvinen if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) 221562ad2761SIlpo Järvinen goto send_now; 221662ad2761SIlpo Järvinen 22175bbcc0f5SLinus Torvalds win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor); 2218ad9f4f50SEric Dumazet if (win_divisor) { 221940570375SEric Dumazet u32 chunk = min(tp->snd_wnd, tcp_snd_cwnd(tp) * tp->mss_cache); 2220c1b4a7e6SDavid S. Miller 2221c1b4a7e6SDavid S. Miller /* If at least some fraction of a window is available, 2222c1b4a7e6SDavid S. Miller * just use it. 2223c1b4a7e6SDavid S. Miller */ 2224ad9f4f50SEric Dumazet chunk /= win_divisor; 2225c1b4a7e6SDavid S. Miller if (limit >= chunk) 2226ae8064acSJohn Heffner goto send_now; 2227c1b4a7e6SDavid S. Miller } else { 2228c1b4a7e6SDavid S. Miller /* Different approach, try not to defer past a single 2229c1b4a7e6SDavid S. Miller * ACK. Receiver should ACK every other full sized 2230c1b4a7e6SDavid S. Miller * frame, so if we have space for more than 3 frames 2231c1b4a7e6SDavid S. Miller * then send now. 2232c1b4a7e6SDavid S. Miller */ 22336b5a5c0dSNeal Cardwell if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) 2234ae8064acSJohn Heffner goto send_now; 2235c1b4a7e6SDavid S. Miller } 2236c1b4a7e6SDavid S. Miller 223775c119afSEric Dumazet /* TODO : use tsorted_sent_queue ? */ 223875c119afSEric Dumazet head = tcp_rtx_queue_head(sk); 223975c119afSEric Dumazet if (!head) 224075c119afSEric Dumazet goto send_now; 2241f1c6ea38SEric Dumazet delta = tp->tcp_clock_cache - head->tstamp; 224250c8339eSEric Dumazet /* If next ACK is likely to come too late (half srtt), do not defer */ 2243f1c6ea38SEric Dumazet if ((s64)(delta - (u64)NSEC_PER_USEC * (tp->srtt_us >> 4)) < 0) 224450c8339eSEric Dumazet goto send_now; 224550c8339eSEric Dumazet 2246f9bfe4e6SEric Dumazet /* Ok, it looks like it is advisable to defer. 2247f9bfe4e6SEric Dumazet * Three cases are tracked : 2248f9bfe4e6SEric Dumazet * 1) We are cwnd-limited 2249f9bfe4e6SEric Dumazet * 2) We are rwnd-limited 2250f9bfe4e6SEric Dumazet * 3) We are application limited. 2251f9bfe4e6SEric Dumazet */ 2252f9bfe4e6SEric Dumazet if (cong_win < send_win) { 2253f9bfe4e6SEric Dumazet if (cong_win <= skb->len) { 2254ca8a2263SNeal Cardwell *is_cwnd_limited = true; 2255f9bfe4e6SEric Dumazet return true; 2256f9bfe4e6SEric Dumazet } 2257f9bfe4e6SEric Dumazet } else { 2258f9bfe4e6SEric Dumazet if (send_win <= skb->len) { 2259f9bfe4e6SEric Dumazet *is_rwnd_limited = true; 2260f9bfe4e6SEric Dumazet return true; 2261f9bfe4e6SEric Dumazet } 2262f9bfe4e6SEric Dumazet } 2263f9bfe4e6SEric Dumazet 2264f9bfe4e6SEric Dumazet /* If this packet won't get more data, do not wait. */ 2265d8ed257fSEric Dumazet if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) || 2266d8ed257fSEric Dumazet TCP_SKB_CB(skb)->eor) 2267f9bfe4e6SEric Dumazet goto send_now; 2268ca8a2263SNeal Cardwell 2269a2a385d6SEric Dumazet return true; 2270ae8064acSJohn Heffner 2271ae8064acSJohn Heffner send_now: 2272a2a385d6SEric Dumazet return false; 2273c1b4a7e6SDavid S. Miller } 2274c1b4a7e6SDavid S. Miller 227505cbc0dbSFan Du static inline void tcp_mtu_check_reprobe(struct sock *sk) 227605cbc0dbSFan Du { 227705cbc0dbSFan Du struct inet_connection_sock *icsk = inet_csk(sk); 227805cbc0dbSFan Du struct tcp_sock *tp = tcp_sk(sk); 227905cbc0dbSFan Du struct net *net = sock_net(sk); 228005cbc0dbSFan Du u32 interval; 228105cbc0dbSFan Du s32 delta; 228205cbc0dbSFan Du 22832a85388fSKuniyuki Iwashima interval = READ_ONCE(net->ipv4.sysctl_tcp_probe_interval); 2284c74df29aSEric Dumazet delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp; 228505cbc0dbSFan Du if (unlikely(delta >= interval * HZ)) { 228605cbc0dbSFan Du int mss = tcp_current_mss(sk); 228705cbc0dbSFan Du 228805cbc0dbSFan Du /* Update current search range */ 228905cbc0dbSFan Du icsk->icsk_mtup.probe_size = 0; 229005cbc0dbSFan Du icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + 229105cbc0dbSFan Du sizeof(struct tcphdr) + 229205cbc0dbSFan Du icsk->icsk_af_ops->net_header_len; 229305cbc0dbSFan Du icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss); 229405cbc0dbSFan Du 229505cbc0dbSFan Du /* Update probe time stamp */ 2296c74df29aSEric Dumazet icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; 229705cbc0dbSFan Du } 229805cbc0dbSFan Du } 229905cbc0dbSFan Du 2300808cf9e3SIlya Lesokhin static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len) 2301808cf9e3SIlya Lesokhin { 2302808cf9e3SIlya Lesokhin struct sk_buff *skb, *next; 2303808cf9e3SIlya Lesokhin 2304808cf9e3SIlya Lesokhin skb = tcp_send_head(sk); 2305808cf9e3SIlya Lesokhin tcp_for_write_queue_from_safe(skb, next, sk) { 2306808cf9e3SIlya Lesokhin if (len <= skb->len) 2307808cf9e3SIlya Lesokhin break; 2308808cf9e3SIlya Lesokhin 23099b65b17dSTalal Ahmad if (unlikely(TCP_SKB_CB(skb)->eor) || 23109b65b17dSTalal Ahmad tcp_has_tx_tstamp(skb) || 23119b65b17dSTalal Ahmad !skb_pure_zcopy_same(skb, next)) 2312808cf9e3SIlya Lesokhin return false; 2313808cf9e3SIlya Lesokhin 2314808cf9e3SIlya Lesokhin len -= skb->len; 2315808cf9e3SIlya Lesokhin } 2316808cf9e3SIlya Lesokhin 2317808cf9e3SIlya Lesokhin return true; 2318808cf9e3SIlya Lesokhin } 2319808cf9e3SIlya Lesokhin 23205d424d5aSJohn Heffner /* Create a new MTU probe if we are ready. 232167edfef7SAndi Kleen * MTU probe is regularly attempting to increase the path MTU by 232267edfef7SAndi Kleen * deliberately sending larger packets. This discovers routing 232367edfef7SAndi Kleen * changes resulting in larger path MTUs. 232467edfef7SAndi Kleen * 23255d424d5aSJohn Heffner * Returns 0 if we should wait to probe (no cwnd available), 23265d424d5aSJohn Heffner * 1 if a probe was sent, 2327056834d9SIlpo Järvinen * -1 otherwise 2328056834d9SIlpo Järvinen */ 23295d424d5aSJohn Heffner static int tcp_mtu_probe(struct sock *sk) 23305d424d5aSJohn Heffner { 23315d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 233212a59abcSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 23335d424d5aSJohn Heffner struct sk_buff *skb, *nskb, *next; 23346b58e0a5SFan Du struct net *net = sock_net(sk); 23355d424d5aSJohn Heffner int probe_size; 233691cc17c0SIlpo Järvinen int size_needed; 233712a59abcSEric Dumazet int copy, len; 23385d424d5aSJohn Heffner int mss_now; 23396b58e0a5SFan Du int interval; 23405d424d5aSJohn Heffner 23415d424d5aSJohn Heffner /* Not currently probing/verifying, 23425d424d5aSJohn Heffner * not in recovery, 23435d424d5aSJohn Heffner * have enough cwnd, and 234412a59abcSEric Dumazet * not SACKing (the variable headers throw things off) 234512a59abcSEric Dumazet */ 234612a59abcSEric Dumazet if (likely(!icsk->icsk_mtup.enabled || 23475d424d5aSJohn Heffner icsk->icsk_mtup.probe_size || 23485d424d5aSJohn Heffner inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 234940570375SEric Dumazet tcp_snd_cwnd(tp) < 11 || 235012a59abcSEric Dumazet tp->rx_opt.num_sacks || tp->rx_opt.dsack)) 23515d424d5aSJohn Heffner return -1; 23525d424d5aSJohn Heffner 23536b58e0a5SFan Du /* Use binary search for probe_size between tcp_mss_base, 23546b58e0a5SFan Du * and current mss_clamp. if (search_high - search_low) 23556b58e0a5SFan Du * smaller than a threshold, backoff from probing. 23566b58e0a5SFan Du */ 23570c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 23586b58e0a5SFan Du probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high + 23596b58e0a5SFan Du icsk->icsk_mtup.search_low) >> 1); 236091cc17c0SIlpo Järvinen size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; 23616b58e0a5SFan Du interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low; 236205cbc0dbSFan Du /* When misfortune happens, we are reprobing actively, 236305cbc0dbSFan Du * and then reprobe timer has expired. We stick with current 236405cbc0dbSFan Du * probing process by not resetting search range to its orignal. 236505cbc0dbSFan Du */ 23666b58e0a5SFan Du if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) || 236792c0aa41SKuniyuki Iwashima interval < READ_ONCE(net->ipv4.sysctl_tcp_probe_threshold)) { 236805cbc0dbSFan Du /* Check whether enough time has elaplased for 236905cbc0dbSFan Du * another round of probing. 237005cbc0dbSFan Du */ 237105cbc0dbSFan Du tcp_mtu_check_reprobe(sk); 23725d424d5aSJohn Heffner return -1; 23735d424d5aSJohn Heffner } 23745d424d5aSJohn Heffner 23755d424d5aSJohn Heffner /* Have enough data in the send queue to probe? */ 23767f9c33e5SIlpo Järvinen if (tp->write_seq - tp->snd_nxt < size_needed) 23775d424d5aSJohn Heffner return -1; 23785d424d5aSJohn Heffner 237991cc17c0SIlpo Järvinen if (tp->snd_wnd < size_needed) 23805d424d5aSJohn Heffner return -1; 238190840defSIlpo Järvinen if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) 23825d424d5aSJohn Heffner return 0; 23835d424d5aSJohn Heffner 2384d67c58e9SIlpo Järvinen /* Do we need to wait to drain cwnd? With none in flight, don't stall */ 238540570375SEric Dumazet if (tcp_packets_in_flight(tp) + 2 > tcp_snd_cwnd(tp)) { 2386d67c58e9SIlpo Järvinen if (!tcp_packets_in_flight(tp)) 23875d424d5aSJohn Heffner return -1; 23885d424d5aSJohn Heffner else 23895d424d5aSJohn Heffner return 0; 23905d424d5aSJohn Heffner } 23915d424d5aSJohn Heffner 2392808cf9e3SIlya Lesokhin if (!tcp_can_coalesce_send_queue_head(sk, probe_size)) 2393808cf9e3SIlya Lesokhin return -1; 2394808cf9e3SIlya Lesokhin 23955d424d5aSJohn Heffner /* We're allowed to probe. Build it now. */ 2396f8dd3b8dSEric Dumazet nskb = tcp_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false); 239751456b29SIan Morris if (!nskb) 23985d424d5aSJohn Heffner return -1; 2399ab4e846aSEric Dumazet sk_wmem_queued_add(sk, nskb->truesize); 24003ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 24015d424d5aSJohn Heffner 2402fe067e8aSDavid S. Miller skb = tcp_send_head(sk); 240341477662SJakub Kicinski skb_copy_decrypted(nskb, skb); 24045a369ca6SPaolo Abeni mptcp_skb_ext_copy(nskb, skb); 24055d424d5aSJohn Heffner 24065d424d5aSJohn Heffner TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 24075d424d5aSJohn Heffner TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 24084de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; 24095d424d5aSJohn Heffner 241050c4817eSIlpo Järvinen tcp_insert_write_queue_before(nskb, skb, sk); 24112b7cda9cSEric Dumazet tcp_highest_sack_replace(sk, skb, nskb); 241250c4817eSIlpo Järvinen 24135d424d5aSJohn Heffner len = 0; 2414234b6860SIlpo Järvinen tcp_for_write_queue_from_safe(skb, next, sk) { 24155d424d5aSJohn Heffner copy = min_t(int, skb->len, probe_size - len); 24165d424d5aSJohn Heffner skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); 24175d424d5aSJohn Heffner 24185d424d5aSJohn Heffner if (skb->len <= copy) { 24195d424d5aSJohn Heffner /* We've eaten all the data from this skb. 24205d424d5aSJohn Heffner * Throw it away. */ 24214de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 2422808cf9e3SIlya Lesokhin /* If this is the last SKB we copy and eor is set 2423808cf9e3SIlya Lesokhin * we need to propagate it to the new skb. 2424808cf9e3SIlya Lesokhin */ 2425808cf9e3SIlya Lesokhin TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor; 2426888a5c53SWillem de Bruijn tcp_skb_collapse_tstamp(nskb, skb); 2427fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 242803271f3aSTalal Ahmad tcp_wmem_free_skb(sk, skb); 24295d424d5aSJohn Heffner } else { 24304de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & 2431a3433f35SChangli Gao ~(TCPHDR_FIN|TCPHDR_PSH); 24325d424d5aSJohn Heffner if (!skb_shinfo(skb)->nr_frags) { 24335d424d5aSJohn Heffner skb_pull(skb, copy); 24345d424d5aSJohn Heffner } else { 24355d424d5aSJohn Heffner __pskb_trim_head(skb, copy); 24365bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 24375d424d5aSJohn Heffner } 24385d424d5aSJohn Heffner TCP_SKB_CB(skb)->seq += copy; 24395d424d5aSJohn Heffner } 24405d424d5aSJohn Heffner 24415d424d5aSJohn Heffner len += copy; 2442234b6860SIlpo Järvinen 2443234b6860SIlpo Järvinen if (len >= probe_size) 2444234b6860SIlpo Järvinen break; 24455d424d5aSJohn Heffner } 24465bbb432cSEric Dumazet tcp_init_tso_segs(nskb, nskb->len); 24475d424d5aSJohn Heffner 24485d424d5aSJohn Heffner /* We're ready to send. If this fails, the probe will 24497faee5c0SEric Dumazet * be resegmented into mss-sized pieces by tcp_write_xmit(). 24507faee5c0SEric Dumazet */ 24515d424d5aSJohn Heffner if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 24525d424d5aSJohn Heffner /* Decrement cwnd here because we are sending 24535d424d5aSJohn Heffner * effectively two packets. */ 245440570375SEric Dumazet tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) - 1); 245566f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, nskb); 24565d424d5aSJohn Heffner 24575d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 24580e7b1368SJohn Heffner tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 24590e7b1368SJohn Heffner tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 24605d424d5aSJohn Heffner 24615d424d5aSJohn Heffner return 1; 24625d424d5aSJohn Heffner } 24635d424d5aSJohn Heffner 24645d424d5aSJohn Heffner return -1; 24655d424d5aSJohn Heffner } 24665d424d5aSJohn Heffner 2467864e5c09SEric Dumazet static bool tcp_pacing_check(struct sock *sk) 2468218af599SEric Dumazet { 2469864e5c09SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 2470864e5c09SEric Dumazet 2471864e5c09SEric Dumazet if (!tcp_needs_internal_pacing(sk)) 2472864e5c09SEric Dumazet return false; 2473864e5c09SEric Dumazet 2474864e5c09SEric Dumazet if (tp->tcp_wstamp_ns <= tp->tcp_clock_cache) 2475864e5c09SEric Dumazet return false; 2476864e5c09SEric Dumazet 2477864e5c09SEric Dumazet if (!hrtimer_is_queued(&tp->pacing_timer)) { 2478864e5c09SEric Dumazet hrtimer_start(&tp->pacing_timer, 2479864e5c09SEric Dumazet ns_to_ktime(tp->tcp_wstamp_ns), 2480864e5c09SEric Dumazet HRTIMER_MODE_ABS_PINNED_SOFT); 2481864e5c09SEric Dumazet sock_hold(sk); 2482864e5c09SEric Dumazet } 2483864e5c09SEric Dumazet return true; 2484218af599SEric Dumazet } 2485218af599SEric Dumazet 2486f9616c35SEric Dumazet /* TCP Small Queues : 2487f9616c35SEric Dumazet * Control number of packets in qdisc/devices to two packets / or ~1 ms. 2488f9616c35SEric Dumazet * (These limits are doubled for retransmits) 2489f9616c35SEric Dumazet * This allows for : 2490f9616c35SEric Dumazet * - better RTT estimation and ACK scheduling 2491f9616c35SEric Dumazet * - faster recovery 2492f9616c35SEric Dumazet * - high rates 2493f9616c35SEric Dumazet * Alas, some drivers / subsystems require a fair amount 2494f9616c35SEric Dumazet * of queued bytes to ensure line rate. 2495f9616c35SEric Dumazet * One example is wifi aggregation (802.11 AMPDU) 2496f9616c35SEric Dumazet */ 2497f9616c35SEric Dumazet static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb, 2498f9616c35SEric Dumazet unsigned int factor) 2499f9616c35SEric Dumazet { 250076a9ebe8SEric Dumazet unsigned long limit; 2501f9616c35SEric Dumazet 250276a9ebe8SEric Dumazet limit = max_t(unsigned long, 250376a9ebe8SEric Dumazet 2 * skb->truesize, 25047c68fa2bSEric Dumazet sk->sk_pacing_rate >> READ_ONCE(sk->sk_pacing_shift)); 2505c73e5807SEric Dumazet if (sk->sk_pacing_status == SK_PACING_NONE) 250676a9ebe8SEric Dumazet limit = min_t(unsigned long, limit, 25079fb90193SKuniyuki Iwashima READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes)); 2508f9616c35SEric Dumazet limit <<= factor; 2509f9616c35SEric Dumazet 2510a842fe14SEric Dumazet if (static_branch_unlikely(&tcp_tx_delay_enabled) && 2511a842fe14SEric Dumazet tcp_sk(sk)->tcp_tx_delay) { 2512a842fe14SEric Dumazet u64 extra_bytes = (u64)sk->sk_pacing_rate * tcp_sk(sk)->tcp_tx_delay; 2513a842fe14SEric Dumazet 2514a842fe14SEric Dumazet /* TSQ is based on skb truesize sum (sk_wmem_alloc), so we 2515a842fe14SEric Dumazet * approximate our needs assuming an ~100% skb->truesize overhead. 2516a842fe14SEric Dumazet * USEC_PER_SEC is approximated by 2^20. 2517a842fe14SEric Dumazet * do_div(extra_bytes, USEC_PER_SEC/2) is replaced by a right shift. 2518a842fe14SEric Dumazet */ 2519a842fe14SEric Dumazet extra_bytes >>= (20 - 1); 2520a842fe14SEric Dumazet limit += extra_bytes; 2521a842fe14SEric Dumazet } 252214afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) > limit) { 252375c119afSEric Dumazet /* Always send skb if rtx queue is empty. 252475eefc6cSEric Dumazet * No need to wait for TX completion to call us back, 252575eefc6cSEric Dumazet * after softirq/tasklet schedule. 252675eefc6cSEric Dumazet * This helps when TX completions are delayed too much. 252775eefc6cSEric Dumazet */ 252875c119afSEric Dumazet if (tcp_rtx_queue_empty(sk)) 252975eefc6cSEric Dumazet return false; 253075eefc6cSEric Dumazet 25317aa5470cSEric Dumazet set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags); 2532f9616c35SEric Dumazet /* It is possible TX completion already happened 2533f9616c35SEric Dumazet * before we set TSQ_THROTTLED, so we must 2534f9616c35SEric Dumazet * test again the condition. 2535f9616c35SEric Dumazet */ 2536f9616c35SEric Dumazet smp_mb__after_atomic(); 2537ce8299b6SEric Dumazet if (refcount_read(&sk->sk_wmem_alloc) > limit) 2538f9616c35SEric Dumazet return true; 2539f9616c35SEric Dumazet } 2540f9616c35SEric Dumazet return false; 2541f9616c35SEric Dumazet } 2542f9616c35SEric Dumazet 254305b055e8SFrancis Yan static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new) 254405b055e8SFrancis Yan { 2545628174ccSEric Dumazet const u32 now = tcp_jiffies32; 2546efe967cdSArnd Bergmann enum tcp_chrono old = tp->chrono_type; 254705b055e8SFrancis Yan 2548efe967cdSArnd Bergmann if (old > TCP_CHRONO_UNSPEC) 2549efe967cdSArnd Bergmann tp->chrono_stat[old - 1] += now - tp->chrono_start; 255005b055e8SFrancis Yan tp->chrono_start = now; 255105b055e8SFrancis Yan tp->chrono_type = new; 255205b055e8SFrancis Yan } 255305b055e8SFrancis Yan 255405b055e8SFrancis Yan void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type) 255505b055e8SFrancis Yan { 255605b055e8SFrancis Yan struct tcp_sock *tp = tcp_sk(sk); 255705b055e8SFrancis Yan 255805b055e8SFrancis Yan /* If there are multiple conditions worthy of tracking in a 25590f87230dSFrancis Yan * chronograph then the highest priority enum takes precedence 25600f87230dSFrancis Yan * over the other conditions. So that if something "more interesting" 256105b055e8SFrancis Yan * starts happening, stop the previous chrono and start a new one. 256205b055e8SFrancis Yan */ 256305b055e8SFrancis Yan if (type > tp->chrono_type) 256405b055e8SFrancis Yan tcp_chrono_set(tp, type); 256505b055e8SFrancis Yan } 256605b055e8SFrancis Yan 256705b055e8SFrancis Yan void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type) 256805b055e8SFrancis Yan { 256905b055e8SFrancis Yan struct tcp_sock *tp = tcp_sk(sk); 257005b055e8SFrancis Yan 25710f87230dSFrancis Yan 25720f87230dSFrancis Yan /* There are multiple conditions worthy of tracking in a 25730f87230dSFrancis Yan * chronograph, so that the highest priority enum takes 25740f87230dSFrancis Yan * precedence over the other conditions (see tcp_chrono_start). 25750f87230dSFrancis Yan * If a condition stops, we only stop chrono tracking if 25760f87230dSFrancis Yan * it's the "most interesting" or current chrono we are 25770f87230dSFrancis Yan * tracking and starts busy chrono if we have pending data. 25780f87230dSFrancis Yan */ 257975c119afSEric Dumazet if (tcp_rtx_and_write_queues_empty(sk)) 258005b055e8SFrancis Yan tcp_chrono_set(tp, TCP_CHRONO_UNSPEC); 25810f87230dSFrancis Yan else if (type == tp->chrono_type) 25820f87230dSFrancis Yan tcp_chrono_set(tp, TCP_CHRONO_BUSY); 258305b055e8SFrancis Yan } 258405b055e8SFrancis Yan 25851da177e4SLinus Torvalds /* This routine writes packets to the network. It advances the 25861da177e4SLinus Torvalds * send_head. This happens as incoming acks open up the remote 25871da177e4SLinus Torvalds * window for us. 25881da177e4SLinus Torvalds * 2589f8269a49SIlpo Järvinen * LARGESEND note: !tcp_urg_mode is overkill, only frames between 2590f8269a49SIlpo Järvinen * snd_up-64k-mss .. snd_up cannot be large. However, taking into 2591f8269a49SIlpo Järvinen * account rare use of URG, this is not a big flaw. 2592f8269a49SIlpo Järvinen * 25936ba8a3b1SNandita Dukkipati * Send at most one packet when push_one > 0. Temporarily ignore 25946ba8a3b1SNandita Dukkipati * cwnd limit to force at most one packet out when push_one == 2. 25956ba8a3b1SNandita Dukkipati 2596a2a385d6SEric Dumazet * Returns true, if no segments are in flight and we have queued segments, 2597a2a385d6SEric Dumazet * but cannot send anything now because of SWS or another problem. 25981da177e4SLinus Torvalds */ 2599a2a385d6SEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 2600d5dd9175SIlpo Järvinen int push_one, gfp_t gfp) 26011da177e4SLinus Torvalds { 26021da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 260392df7b51SDavid S. Miller struct sk_buff *skb; 2604c1b4a7e6SDavid S. Miller unsigned int tso_segs, sent_pkts; 2605c1b4a7e6SDavid S. Miller int cwnd_quota; 26065d424d5aSJohn Heffner int result; 26075615f886SFrancis Yan bool is_cwnd_limited = false, is_rwnd_limited = false; 2608605ad7f1SEric Dumazet u32 max_segs; 26091da177e4SLinus Torvalds 2610c1b4a7e6SDavid S. Miller sent_pkts = 0; 26115d424d5aSJohn Heffner 2612ee1836aeSEric Dumazet tcp_mstamp_refresh(tp); 2613d5dd9175SIlpo Järvinen if (!push_one) { 26145d424d5aSJohn Heffner /* Do MTU probing. */ 2615d5dd9175SIlpo Järvinen result = tcp_mtu_probe(sk); 2616d5dd9175SIlpo Järvinen if (!result) { 2617a2a385d6SEric Dumazet return false; 26185d424d5aSJohn Heffner } else if (result > 0) { 26195d424d5aSJohn Heffner sent_pkts = 1; 26205d424d5aSJohn Heffner } 2621d5dd9175SIlpo Järvinen } 26225d424d5aSJohn Heffner 2623ed6e7268SNeal Cardwell max_segs = tcp_tso_segs(sk, mss_now); 2624fe067e8aSDavid S. Miller while ((skb = tcp_send_head(sk))) { 2625c8ac3774SHerbert Xu unsigned int limit; 2626c8ac3774SHerbert Xu 262779861919SEric Dumazet if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) { 262879861919SEric Dumazet /* "skb_mstamp_ns" is used as a start point for the retransmit timer */ 2629a1ac9c8aSMartin KaFai Lau tp->tcp_wstamp_ns = tp->tcp_clock_cache; 2630a1ac9c8aSMartin KaFai Lau skb_set_delivery_time(skb, tp->tcp_wstamp_ns, true); 263179861919SEric Dumazet list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue); 2632bf50b606SEric Dumazet tcp_init_tso_segs(skb, mss_now); 263379861919SEric Dumazet goto repair; /* Skip network transmission */ 263479861919SEric Dumazet } 263579861919SEric Dumazet 2636218af599SEric Dumazet if (tcp_pacing_check(sk)) 2637218af599SEric Dumazet break; 2638218af599SEric Dumazet 26395bbb432cSEric Dumazet tso_segs = tcp_init_tso_segs(skb, mss_now); 2640c1b4a7e6SDavid S. Miller BUG_ON(!tso_segs); 2641c1b4a7e6SDavid S. Miller 2642b68e9f85SHerbert Xu cwnd_quota = tcp_cwnd_test(tp, skb); 26436ba8a3b1SNandita Dukkipati if (!cwnd_quota) { 26446ba8a3b1SNandita Dukkipati if (push_one == 2) 26456ba8a3b1SNandita Dukkipati /* Force out a loss probe pkt. */ 26466ba8a3b1SNandita Dukkipati cwnd_quota = 1; 26476ba8a3b1SNandita Dukkipati else 2648b68e9f85SHerbert Xu break; 26496ba8a3b1SNandita Dukkipati } 2650b68e9f85SHerbert Xu 26515615f886SFrancis Yan if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) { 26525615f886SFrancis Yan is_rwnd_limited = true; 2653b68e9f85SHerbert Xu break; 26545615f886SFrancis Yan } 2655b68e9f85SHerbert Xu 2656d6a4e26aSEric Dumazet if (tso_segs == 1) { 2657aa93466bSDavid S. Miller if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 2658aa93466bSDavid S. Miller (tcp_skb_is_last(sk, skb) ? 2659aa93466bSDavid S. Miller nonagle : TCP_NAGLE_PUSH)))) 2660aa93466bSDavid S. Miller break; 2661c1b4a7e6SDavid S. Miller } else { 2662ca8a2263SNeal Cardwell if (!push_one && 2663605ad7f1SEric Dumazet tcp_tso_should_defer(sk, skb, &is_cwnd_limited, 2664f9bfe4e6SEric Dumazet &is_rwnd_limited, max_segs)) 2665aa93466bSDavid S. Miller break; 2666c1b4a7e6SDavid S. Miller } 2667aa93466bSDavid S. Miller 2668605ad7f1SEric Dumazet limit = mss_now; 2669d6a4e26aSEric Dumazet if (tso_segs > 1 && !tcp_urg_mode(tp)) 2670605ad7f1SEric Dumazet limit = tcp_mss_split_point(sk, skb, mss_now, 2671605ad7f1SEric Dumazet min_t(unsigned int, 2672605ad7f1SEric Dumazet cwnd_quota, 2673605ad7f1SEric Dumazet max_segs), 2674605ad7f1SEric Dumazet nonagle); 2675605ad7f1SEric Dumazet 2676605ad7f1SEric Dumazet if (skb->len > limit && 267756483341SEric Dumazet unlikely(tso_fragment(sk, skb, limit, mss_now, gfp))) 2678605ad7f1SEric Dumazet break; 2679605ad7f1SEric Dumazet 2680f9616c35SEric Dumazet if (tcp_small_queue_check(sk, skb, 0)) 268146d3ceabSEric Dumazet break; 2682c9eeec26SEric Dumazet 26831f85e626SEric Dumazet /* Argh, we hit an empty skb(), presumably a thread 26841f85e626SEric Dumazet * is sleeping in sendmsg()/sk_stream_wait_memory(). 26851f85e626SEric Dumazet * We do not want to send a pure-ack packet and have 26861f85e626SEric Dumazet * a strange looking rtx queue with empty packet(s). 26871f85e626SEric Dumazet */ 26881f85e626SEric Dumazet if (TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq) 26891f85e626SEric Dumazet break; 26901f85e626SEric Dumazet 2691d5dd9175SIlpo Järvinen if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) 26921da177e4SLinus Torvalds break; 26931da177e4SLinus Torvalds 2694ec342325SAndrew Vagin repair: 26951da177e4SLinus Torvalds /* Advance the send_head. This one is sent out. 26961da177e4SLinus Torvalds * This call will increment packets_out. 26971da177e4SLinus Torvalds */ 269866f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 26991da177e4SLinus Torvalds 27001da177e4SLinus Torvalds tcp_minshall_update(tp, mss_now, skb); 2701a262f0cdSNandita Dukkipati sent_pkts += tcp_skb_pcount(skb); 2702d5dd9175SIlpo Järvinen 2703d5dd9175SIlpo Järvinen if (push_one) 2704d5dd9175SIlpo Järvinen break; 27051da177e4SLinus Torvalds } 27061da177e4SLinus Torvalds 27075615f886SFrancis Yan if (is_rwnd_limited) 27085615f886SFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED); 27095615f886SFrancis Yan else 27105615f886SFrancis Yan tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED); 27115615f886SFrancis Yan 271240570375SEric Dumazet is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tcp_snd_cwnd(tp)); 2713299bcb55SNeal Cardwell if (likely(sent_pkts || is_cwnd_limited)) 2714299bcb55SNeal Cardwell tcp_cwnd_validate(sk, is_cwnd_limited); 2715299bcb55SNeal Cardwell 2716aa93466bSDavid S. Miller if (likely(sent_pkts)) { 2717684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 2718684bad11SYuchung Cheng tp->prr_out += sent_pkts; 27196ba8a3b1SNandita Dukkipati 27206ba8a3b1SNandita Dukkipati /* Send one loss probe per tail loss episode. */ 27216ba8a3b1SNandita Dukkipati if (push_one != 2) 2722ed66dfafSNeal Cardwell tcp_schedule_loss_probe(sk, false); 2723a2a385d6SEric Dumazet return false; 27241da177e4SLinus Torvalds } 272575c119afSEric Dumazet return !tp->packets_out && !tcp_write_queue_empty(sk); 27266ba8a3b1SNandita Dukkipati } 27276ba8a3b1SNandita Dukkipati 2728ed66dfafSNeal Cardwell bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto) 27296ba8a3b1SNandita Dukkipati { 27306ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 27316ba8a3b1SNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 2732a2815817SNeal Cardwell u32 timeout, rto_delta_us; 27332ae21cf5SEric Dumazet int early_retrans; 27346ba8a3b1SNandita Dukkipati 27356ba8a3b1SNandita Dukkipati /* Don't do any loss probe on a Fast Open connection before 3WHS 27366ba8a3b1SNandita Dukkipati * finishes. 27376ba8a3b1SNandita Dukkipati */ 2738d983ea6fSEric Dumazet if (rcu_access_pointer(tp->fastopen_rsk)) 27396ba8a3b1SNandita Dukkipati return false; 27406ba8a3b1SNandita Dukkipati 274152e65865SKuniyuki Iwashima early_retrans = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_early_retrans); 27426ba8a3b1SNandita Dukkipati /* Schedule a loss probe in 2*RTT for SACK capable connections 2743b4f70c3dSNeal Cardwell * not in loss recovery, that are either limited by cwnd or application. 27446ba8a3b1SNandita Dukkipati */ 27452ae21cf5SEric Dumazet if ((early_retrans != 3 && early_retrans != 4) || 2746bec41a11SYuchung Cheng !tp->packets_out || !tcp_is_sack(tp) || 2747b4f70c3dSNeal Cardwell (icsk->icsk_ca_state != TCP_CA_Open && 2748b4f70c3dSNeal Cardwell icsk->icsk_ca_state != TCP_CA_CWR)) 27496ba8a3b1SNandita Dukkipati return false; 27506ba8a3b1SNandita Dukkipati 2751bb4d991aSYuchung Cheng /* Probe timeout is 2*rtt. Add minimum RTO to account 2752f9b99582SYuchung Cheng * for delayed ack when there's one outstanding packet. If no RTT 2753f9b99582SYuchung Cheng * sample is available then probe after TCP_TIMEOUT_INIT. 27546ba8a3b1SNandita Dukkipati */ 2755bb4d991aSYuchung Cheng if (tp->srtt_us) { 2756bb4d991aSYuchung Cheng timeout = usecs_to_jiffies(tp->srtt_us >> 2); 27576ba8a3b1SNandita Dukkipati if (tp->packets_out == 1) 2758bb4d991aSYuchung Cheng timeout += TCP_RTO_MIN; 2759bb4d991aSYuchung Cheng else 2760bb4d991aSYuchung Cheng timeout += TCP_TIMEOUT_MIN; 2761bb4d991aSYuchung Cheng } else { 2762bb4d991aSYuchung Cheng timeout = TCP_TIMEOUT_INIT; 2763bb4d991aSYuchung Cheng } 27646ba8a3b1SNandita Dukkipati 2765a2815817SNeal Cardwell /* If the RTO formula yields an earlier time, then use that time. */ 2766ed66dfafSNeal Cardwell rto_delta_us = advancing_rto ? 2767ed66dfafSNeal Cardwell jiffies_to_usecs(inet_csk(sk)->icsk_rto) : 2768ed66dfafSNeal Cardwell tcp_rto_delta_us(sk); /* How far in future is RTO? */ 2769a2815817SNeal Cardwell if (rto_delta_us > 0) 2770a2815817SNeal Cardwell timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us)); 27716ba8a3b1SNandita Dukkipati 27728dc242adSEric Dumazet tcp_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, TCP_RTO_MAX); 27736ba8a3b1SNandita Dukkipati return true; 27746ba8a3b1SNandita Dukkipati } 27756ba8a3b1SNandita Dukkipati 27761f3279aeSEric Dumazet /* Thanks to skb fast clones, we can detect if a prior transmit of 27771f3279aeSEric Dumazet * a packet is still in a qdisc or driver queue. 27781f3279aeSEric Dumazet * In this case, there is very little point doing a retransmit ! 27791f3279aeSEric Dumazet */ 2780f4dae54eSEric Dumazet static bool skb_still_in_host_queue(struct sock *sk, 27811f3279aeSEric Dumazet const struct sk_buff *skb) 27821f3279aeSEric Dumazet { 278339bb5e62SEric Dumazet if (unlikely(skb_fclone_busy(sk, skb))) { 2784f4dae54eSEric Dumazet set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags); 2785f4dae54eSEric Dumazet smp_mb__after_atomic(); 2786f4dae54eSEric Dumazet if (skb_fclone_busy(sk, skb)) { 2787c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), 27881f3279aeSEric Dumazet LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES); 27891f3279aeSEric Dumazet return true; 27901f3279aeSEric Dumazet } 2791f4dae54eSEric Dumazet } 27921f3279aeSEric Dumazet return false; 27931f3279aeSEric Dumazet } 27941f3279aeSEric Dumazet 2795b340b264SYuchung Cheng /* When probe timeout (PTO) fires, try send a new segment if possible, else 27966ba8a3b1SNandita Dukkipati * retransmit the last segment. 27976ba8a3b1SNandita Dukkipati */ 27986ba8a3b1SNandita Dukkipati void tcp_send_loss_probe(struct sock *sk) 27996ba8a3b1SNandita Dukkipati { 28009b717a8dSNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 28016ba8a3b1SNandita Dukkipati struct sk_buff *skb; 28026ba8a3b1SNandita Dukkipati int pcount; 28036ba8a3b1SNandita Dukkipati int mss = tcp_current_mss(sk); 28046ba8a3b1SNandita Dukkipati 280576be93fcSYuchung Cheng /* At most one outstanding TLP */ 280676be93fcSYuchung Cheng if (tp->tlp_high_seq) 280776be93fcSYuchung Cheng goto rearm_timer; 280876be93fcSYuchung Cheng 280976be93fcSYuchung Cheng tp->tlp_retrans = 0; 2810b340b264SYuchung Cheng skb = tcp_send_head(sk); 281175c119afSEric Dumazet if (skb && tcp_snd_wnd_test(tp, skb, mss)) { 2812b340b264SYuchung Cheng pcount = tp->packets_out; 2813b340b264SYuchung Cheng tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC); 2814b340b264SYuchung Cheng if (tp->packets_out > pcount) 2815b340b264SYuchung Cheng goto probe_sent; 28166ba8a3b1SNandita Dukkipati goto rearm_timer; 28176ba8a3b1SNandita Dukkipati } 281875c119afSEric Dumazet skb = skb_rb_last(&sk->tcp_rtx_queue); 2819b2b7af86SYuchung Cheng if (unlikely(!skb)) { 2820b2b7af86SYuchung Cheng WARN_ONCE(tp->packets_out, 2821b2b7af86SYuchung Cheng "invalid inflight: %u state %u cwnd %u mss %d\n", 282240570375SEric Dumazet tp->packets_out, sk->sk_state, tcp_snd_cwnd(tp), mss); 2823b2b7af86SYuchung Cheng inet_csk(sk)->icsk_pending = 0; 2824b2b7af86SYuchung Cheng return; 2825b2b7af86SYuchung Cheng } 28266ba8a3b1SNandita Dukkipati 28271f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 28281f3279aeSEric Dumazet goto rearm_timer; 28291f3279aeSEric Dumazet 28306ba8a3b1SNandita Dukkipati pcount = tcp_skb_pcount(skb); 28316ba8a3b1SNandita Dukkipati if (WARN_ON(!pcount)) 28326ba8a3b1SNandita Dukkipati goto rearm_timer; 28336ba8a3b1SNandita Dukkipati 28346ba8a3b1SNandita Dukkipati if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) { 283575c119afSEric Dumazet if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, 283675c119afSEric Dumazet (pcount - 1) * mss, mss, 28376cc55e09SOctavian Purdila GFP_ATOMIC))) 28386ba8a3b1SNandita Dukkipati goto rearm_timer; 283975c119afSEric Dumazet skb = skb_rb_next(skb); 28406ba8a3b1SNandita Dukkipati } 28416ba8a3b1SNandita Dukkipati 28426ba8a3b1SNandita Dukkipati if (WARN_ON(!skb || !tcp_skb_pcount(skb))) 28436ba8a3b1SNandita Dukkipati goto rearm_timer; 28446ba8a3b1SNandita Dukkipati 284510d3be56SEric Dumazet if (__tcp_retransmit_skb(sk, skb, 1)) 2846b340b264SYuchung Cheng goto rearm_timer; 28476ba8a3b1SNandita Dukkipati 284876be93fcSYuchung Cheng tp->tlp_retrans = 1; 284976be93fcSYuchung Cheng 285076be93fcSYuchung Cheng probe_sent: 28519b717a8dSNandita Dukkipati /* Record snd_nxt for loss detection. */ 28529b717a8dSNandita Dukkipati tp->tlp_high_seq = tp->snd_nxt; 28539b717a8dSNandita Dukkipati 2854c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES); 2855fcd16c0aSYuchung Cheng /* Reset s.t. tcp_rearm_rto will restart timer from now */ 2856fcd16c0aSYuchung Cheng inet_csk(sk)->icsk_pending = 0; 2857b340b264SYuchung Cheng rearm_timer: 2858fcd16c0aSYuchung Cheng tcp_rearm_rto(sk); 28591da177e4SLinus Torvalds } 28601da177e4SLinus Torvalds 2861a762a980SDavid S. Miller /* Push out any pending frames which were held back due to 2862a762a980SDavid S. Miller * TCP_CORK or attempt at coalescing tiny packets. 2863a762a980SDavid S. Miller * The socket must be locked by the caller. 2864a762a980SDavid S. Miller */ 28659e412ba7SIlpo Järvinen void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 28669e412ba7SIlpo Järvinen int nonagle) 2867a762a980SDavid S. Miller { 2868726e07a8SIlpo Järvinen /* If we are closed, the bytes will have to remain here. 2869726e07a8SIlpo Järvinen * In time closedown will finish, we empty the write queue and 2870726e07a8SIlpo Järvinen * all will be happy. 2871726e07a8SIlpo Järvinen */ 2872726e07a8SIlpo Järvinen if (unlikely(sk->sk_state == TCP_CLOSE)) 2873726e07a8SIlpo Järvinen return; 2874726e07a8SIlpo Järvinen 287599a1dec7SMel Gorman if (tcp_write_xmit(sk, cur_mss, nonagle, 0, 28767450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC))) 28779e412ba7SIlpo Järvinen tcp_check_probe_timer(sk); 2878a762a980SDavid S. Miller } 2879a762a980SDavid S. Miller 2880c1b4a7e6SDavid S. Miller /* Send _single_ skb sitting at the send head. This function requires 2881c1b4a7e6SDavid S. Miller * true push pending frames to setup probe timer etc. 2882c1b4a7e6SDavid S. Miller */ 2883c1b4a7e6SDavid S. Miller void tcp_push_one(struct sock *sk, unsigned int mss_now) 2884c1b4a7e6SDavid S. Miller { 2885fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 2886c1b4a7e6SDavid S. Miller 2887c1b4a7e6SDavid S. Miller BUG_ON(!skb || skb->len < mss_now); 2888c1b4a7e6SDavid S. Miller 2889d5dd9175SIlpo Järvinen tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); 2890c1b4a7e6SDavid S. Miller } 2891c1b4a7e6SDavid S. Miller 28921da177e4SLinus Torvalds /* This function returns the amount that we can raise the 28931da177e4SLinus Torvalds * usable window based on the following constraints 28941da177e4SLinus Torvalds * 28951da177e4SLinus Torvalds * 1. The window can never be shrunk once it is offered (RFC 793) 28961da177e4SLinus Torvalds * 2. We limit memory per socket 28971da177e4SLinus Torvalds * 28981da177e4SLinus Torvalds * RFC 1122: 28991da177e4SLinus Torvalds * "the suggested [SWS] avoidance algorithm for the receiver is to keep 29001da177e4SLinus Torvalds * RECV.NEXT + RCV.WIN fixed until: 29011da177e4SLinus Torvalds * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 29021da177e4SLinus Torvalds * 29031da177e4SLinus Torvalds * i.e. don't raise the right edge of the window until you can raise 29041da177e4SLinus Torvalds * it at least MSS bytes. 29051da177e4SLinus Torvalds * 29061da177e4SLinus Torvalds * Unfortunately, the recommended algorithm breaks header prediction, 29071da177e4SLinus Torvalds * since header prediction assumes th->window stays fixed. 29081da177e4SLinus Torvalds * 29091da177e4SLinus Torvalds * Strictly speaking, keeping th->window fixed violates the receiver 29101da177e4SLinus Torvalds * side SWS prevention criteria. The problem is that under this rule 29111da177e4SLinus Torvalds * a stream of single byte packets will cause the right side of the 29121da177e4SLinus Torvalds * window to always advance by a single byte. 29131da177e4SLinus Torvalds * 29141da177e4SLinus Torvalds * Of course, if the sender implements sender side SWS prevention 29151da177e4SLinus Torvalds * then this will not be a problem. 29161da177e4SLinus Torvalds * 29171da177e4SLinus Torvalds * BSD seems to make the following compromise: 29181da177e4SLinus Torvalds * 29191da177e4SLinus Torvalds * If the free space is less than the 1/4 of the maximum 29201da177e4SLinus Torvalds * space available and the free space is less than 1/2 mss, 29211da177e4SLinus Torvalds * then set the window to 0. 29221da177e4SLinus Torvalds * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 29231da177e4SLinus Torvalds * Otherwise, just prevent the window from shrinking 29241da177e4SLinus Torvalds * and from being larger than the largest representable value. 29251da177e4SLinus Torvalds * 29261da177e4SLinus Torvalds * This prevents incremental opening of the window in the regime 29271da177e4SLinus Torvalds * where TCP is limited by the speed of the reader side taking 29281da177e4SLinus Torvalds * data out of the TCP receive queue. It does nothing about 29291da177e4SLinus Torvalds * those cases where the window is constrained on the sender side 29301da177e4SLinus Torvalds * because the pipeline is full. 29311da177e4SLinus Torvalds * 29321da177e4SLinus Torvalds * BSD also seems to "accidentally" limit itself to windows that are a 29331da177e4SLinus Torvalds * multiple of MSS, at least until the free space gets quite small. 29341da177e4SLinus Torvalds * This would appear to be a side effect of the mbuf implementation. 29351da177e4SLinus Torvalds * Combining these two algorithms results in the observed behavior 29361da177e4SLinus Torvalds * of having a fixed window size at almost all times. 29371da177e4SLinus Torvalds * 29381da177e4SLinus Torvalds * Below we obtain similar behavior by forcing the offered window to 29391da177e4SLinus Torvalds * a multiple of the mss when it is feasible to do so. 29401da177e4SLinus Torvalds * 29411da177e4SLinus Torvalds * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 29421da177e4SLinus Torvalds * Regular options like TIMESTAMP are taken into account. 29431da177e4SLinus Torvalds */ 29441da177e4SLinus Torvalds u32 __tcp_select_window(struct sock *sk) 29451da177e4SLinus Torvalds { 2946463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 29471da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2948caa20d9aSStephen Hemminger /* MSS for the peer's data. Previous versions used mss_clamp 29491da177e4SLinus Torvalds * here. I don't know if the value based on our guesses 29501da177e4SLinus Torvalds * of peer's MSS is better for the performance. It's more correct 29511da177e4SLinus Torvalds * but may be worse for the performance because of rcv_mss 29521da177e4SLinus Torvalds * fluctuations. --SAW 1998/11/1 29531da177e4SLinus Torvalds */ 2954463c84b9SArnaldo Carvalho de Melo int mss = icsk->icsk_ack.rcv_mss; 29551da177e4SLinus Torvalds int free_space = tcp_space(sk); 295686c1a045SFlorian Westphal int allowed_space = tcp_full_space(sk); 2957071c8ed6SFlorian Westphal int full_space, window; 2958071c8ed6SFlorian Westphal 2959071c8ed6SFlorian Westphal if (sk_is_mptcp(sk)) 2960071c8ed6SFlorian Westphal mptcp_space(sk, &free_space, &allowed_space); 2961071c8ed6SFlorian Westphal 2962071c8ed6SFlorian Westphal full_space = min_t(int, tp->window_clamp, allowed_space); 29631da177e4SLinus Torvalds 296406425c30SEric Dumazet if (unlikely(mss > full_space)) { 29651da177e4SLinus Torvalds mss = full_space; 296606425c30SEric Dumazet if (mss <= 0) 296706425c30SEric Dumazet return 0; 296806425c30SEric Dumazet } 2969b92edbe0SEric Dumazet if (free_space < (full_space >> 1)) { 2970463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.quick = 0; 29711da177e4SLinus Torvalds 2972b8da51ebSEric Dumazet if (tcp_under_memory_pressure(sk)) 2973053f3684SWei Wang tcp_adjust_rcv_ssthresh(sk); 29741da177e4SLinus Torvalds 297586c1a045SFlorian Westphal /* free_space might become our new window, make sure we don't 297686c1a045SFlorian Westphal * increase it due to wscale. 297786c1a045SFlorian Westphal */ 297886c1a045SFlorian Westphal free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale); 297986c1a045SFlorian Westphal 298086c1a045SFlorian Westphal /* if free space is less than mss estimate, or is below 1/16th 298186c1a045SFlorian Westphal * of the maximum allowed, try to move to zero-window, else 298286c1a045SFlorian Westphal * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and 298386c1a045SFlorian Westphal * new incoming data is dropped due to memory limits. 298486c1a045SFlorian Westphal * With large window, mss test triggers way too late in order 298586c1a045SFlorian Westphal * to announce zero window in time before rmem limit kicks in. 298686c1a045SFlorian Westphal */ 298786c1a045SFlorian Westphal if (free_space < (allowed_space >> 4) || free_space < mss) 29881da177e4SLinus Torvalds return 0; 29891da177e4SLinus Torvalds } 29901da177e4SLinus Torvalds 29911da177e4SLinus Torvalds if (free_space > tp->rcv_ssthresh) 29921da177e4SLinus Torvalds free_space = tp->rcv_ssthresh; 29931da177e4SLinus Torvalds 29941da177e4SLinus Torvalds /* Don't do rounding if we are using window scaling, since the 29951da177e4SLinus Torvalds * scaled window will not line up with the MSS boundary anyway. 29961da177e4SLinus Torvalds */ 29971da177e4SLinus Torvalds if (tp->rx_opt.rcv_wscale) { 29981da177e4SLinus Torvalds window = free_space; 29991da177e4SLinus Torvalds 30001da177e4SLinus Torvalds /* Advertise enough space so that it won't get scaled away. 30011da177e4SLinus Torvalds * Import case: prevent zero window announcement if 30021da177e4SLinus Torvalds * 1<<rcv_wscale > mss. 30031da177e4SLinus Torvalds */ 30041935299dSGao Feng window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale)); 30051da177e4SLinus Torvalds } else { 30061935299dSGao Feng window = tp->rcv_wnd; 30071da177e4SLinus Torvalds /* Get the largest window that is a nice multiple of mss. 30081da177e4SLinus Torvalds * Window clamp already applied above. 30091da177e4SLinus Torvalds * If our current window offering is within 1 mss of the 30101da177e4SLinus Torvalds * free space we just keep it. This prevents the divide 30111da177e4SLinus Torvalds * and multiply from happening most of the time. 30121da177e4SLinus Torvalds * We also don't do any window rounding when the free space 30131da177e4SLinus Torvalds * is too small. 30141da177e4SLinus Torvalds */ 30151da177e4SLinus Torvalds if (window <= free_space - mss || window > free_space) 30161935299dSGao Feng window = rounddown(free_space, mss); 301784565070SJohn Heffner else if (mss == full_space && 3018b92edbe0SEric Dumazet free_space > window + (full_space >> 1)) 301984565070SJohn Heffner window = free_space; 30201da177e4SLinus Torvalds } 30211da177e4SLinus Torvalds 30221da177e4SLinus Torvalds return window; 30231da177e4SLinus Torvalds } 30241da177e4SLinus Torvalds 3025cfea5a68SMartin KaFai Lau void tcp_skb_collapse_tstamp(struct sk_buff *skb, 3026082ac2d5SMartin KaFai Lau const struct sk_buff *next_skb) 3027082ac2d5SMartin KaFai Lau { 30280a2cf20cSSoheil Hassas Yeganeh if (unlikely(tcp_has_tx_tstamp(next_skb))) { 30290a2cf20cSSoheil Hassas Yeganeh const struct skb_shared_info *next_shinfo = 30300a2cf20cSSoheil Hassas Yeganeh skb_shinfo(next_skb); 3031082ac2d5SMartin KaFai Lau struct skb_shared_info *shinfo = skb_shinfo(skb); 3032082ac2d5SMartin KaFai Lau 30330a2cf20cSSoheil Hassas Yeganeh shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP; 3034082ac2d5SMartin KaFai Lau shinfo->tskey = next_shinfo->tskey; 30352de8023eSMartin KaFai Lau TCP_SKB_CB(skb)->txstamp_ack |= 30362de8023eSMartin KaFai Lau TCP_SKB_CB(next_skb)->txstamp_ack; 3037082ac2d5SMartin KaFai Lau } 3038082ac2d5SMartin KaFai Lau } 3039082ac2d5SMartin KaFai Lau 30404a17fc3aSIlpo Järvinen /* Collapses two adjacent SKB's during retransmission. */ 3041f8071cdeSEric Dumazet static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) 30421da177e4SLinus Torvalds { 30431da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 304475c119afSEric Dumazet struct sk_buff *next_skb = skb_rb_next(skb); 304513dde04fSWei Yongjun int next_skb_size; 30461da177e4SLinus Torvalds 3047058dc334SIlpo Järvinen next_skb_size = next_skb->len; 30481da177e4SLinus Torvalds 3049058dc334SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); 30501da177e4SLinus Torvalds 3051bd446314SEric Dumazet if (next_skb_size && !tcp_skb_shift(skb, next_skb, 1, next_skb_size)) 3052f8071cdeSEric Dumazet return false; 3053bd446314SEric Dumazet 30542b7cda9cSEric Dumazet tcp_highest_sack_replace(sk, next_skb, skb); 3055a6963a6bSIlpo Järvinen 30561da177e4SLinus Torvalds /* Update sequence range on original skb. */ 30571da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 30581da177e4SLinus Torvalds 3059e6c7d085SIlpo Järvinen /* Merge over control information. This moves PSH/FIN etc. over */ 30604de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; 30611da177e4SLinus Torvalds 30621da177e4SLinus Torvalds /* All done, get rid of second SKB and account for it so 30631da177e4SLinus Torvalds * packet counting does not break. 30641da177e4SLinus Torvalds */ 30654828e7f4SIlpo Järvinen TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; 3066a643b5d4SMartin KaFai Lau TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor; 3067b7689205SIlpo Järvinen 3068b7689205SIlpo Järvinen /* changed transmit queue under us so clear hints */ 3069ef9da47cSIlpo Järvinen tcp_clear_retrans_hints_partial(tp); 3070ef9da47cSIlpo Järvinen if (next_skb == tp->retransmit_skb_hint) 3071ef9da47cSIlpo Järvinen tp->retransmit_skb_hint = skb; 3072b7689205SIlpo Järvinen 3073797108d1SIlpo Järvinen tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); 3074797108d1SIlpo Järvinen 3075082ac2d5SMartin KaFai Lau tcp_skb_collapse_tstamp(skb, next_skb); 3076082ac2d5SMartin KaFai Lau 307775c119afSEric Dumazet tcp_rtx_queue_unlink_and_free(next_skb, sk); 3078f8071cdeSEric Dumazet return true; 30791da177e4SLinus Torvalds } 30801da177e4SLinus Torvalds 308167edfef7SAndi Kleen /* Check if coalescing SKBs is legal. */ 3082a2a385d6SEric Dumazet static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) 30834a17fc3aSIlpo Järvinen { 30844a17fc3aSIlpo Järvinen if (tcp_skb_pcount(skb) > 1) 3085a2a385d6SEric Dumazet return false; 30864a17fc3aSIlpo Järvinen if (skb_cloned(skb)) 3087a2a385d6SEric Dumazet return false; 30882331ccc5SEric Dumazet /* Some heuristics for collapsing over SACK'd could be invented */ 30894a17fc3aSIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 3090a2a385d6SEric Dumazet return false; 30914a17fc3aSIlpo Järvinen 3092a2a385d6SEric Dumazet return true; 30934a17fc3aSIlpo Järvinen } 30944a17fc3aSIlpo Järvinen 309567edfef7SAndi Kleen /* Collapse packets in the retransmit queue to make to create 309667edfef7SAndi Kleen * less packets on the wire. This is only done on retransmission. 309767edfef7SAndi Kleen */ 30984a17fc3aSIlpo Järvinen static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, 30994a17fc3aSIlpo Järvinen int space) 31004a17fc3aSIlpo Järvinen { 31014a17fc3aSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 31024a17fc3aSIlpo Järvinen struct sk_buff *skb = to, *tmp; 3103a2a385d6SEric Dumazet bool first = true; 31044a17fc3aSIlpo Järvinen 31051a63cb91SKuniyuki Iwashima if (!READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse)) 31064a17fc3aSIlpo Järvinen return; 31074de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 31084a17fc3aSIlpo Järvinen return; 31094a17fc3aSIlpo Järvinen 311075c119afSEric Dumazet skb_rbtree_walk_from_safe(skb, tmp) { 31114a17fc3aSIlpo Järvinen if (!tcp_can_collapse(sk, skb)) 31124a17fc3aSIlpo Järvinen break; 31134a17fc3aSIlpo Järvinen 311485712484SMat Martineau if (!tcp_skb_can_collapse(to, skb)) 3115a643b5d4SMartin KaFai Lau break; 3116a643b5d4SMartin KaFai Lau 31174a17fc3aSIlpo Järvinen space -= skb->len; 31184a17fc3aSIlpo Järvinen 31194a17fc3aSIlpo Järvinen if (first) { 3120a2a385d6SEric Dumazet first = false; 31214a17fc3aSIlpo Järvinen continue; 31224a17fc3aSIlpo Järvinen } 31234a17fc3aSIlpo Järvinen 31244a17fc3aSIlpo Järvinen if (space < 0) 31254a17fc3aSIlpo Järvinen break; 31264a17fc3aSIlpo Järvinen 31274a17fc3aSIlpo Järvinen if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) 31284a17fc3aSIlpo Järvinen break; 31294a17fc3aSIlpo Järvinen 3130f8071cdeSEric Dumazet if (!tcp_collapse_retrans(sk, to)) 3131f8071cdeSEric Dumazet break; 31324a17fc3aSIlpo Järvinen } 31334a17fc3aSIlpo Järvinen } 31344a17fc3aSIlpo Järvinen 31351da177e4SLinus Torvalds /* This retransmits one SKB. Policy decisions and retransmit queue 31361da177e4SLinus Torvalds * state updates are done by the caller. Returns non-zero if an 31371da177e4SLinus Torvalds * error occurred which prevented the send. 31381da177e4SLinus Torvalds */ 313910d3be56SEric Dumazet int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) 31401da177e4SLinus Torvalds { 31415d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 314210d3be56SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 31437d227cd2SSridhar Samudrala unsigned int cur_mss; 314410d3be56SEric Dumazet int diff, len, err; 31451da177e4SLinus Torvalds 314610d3be56SEric Dumazet 314710d3be56SEric Dumazet /* Inconclusive MTU probe */ 314810d3be56SEric Dumazet if (icsk->icsk_mtup.probe_size) 31495d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 31505d424d5aSJohn Heffner 31511f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 31521f3279aeSEric Dumazet return -EBUSY; 31531f3279aeSEric Dumazet 31541da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 31557f582b24SEric Dumazet if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) { 31567f582b24SEric Dumazet WARN_ON_ONCE(1); 31577f582b24SEric Dumazet return -EINVAL; 31587f582b24SEric Dumazet } 31591da177e4SLinus Torvalds if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 31601da177e4SLinus Torvalds return -ENOMEM; 31611da177e4SLinus Torvalds } 31621da177e4SLinus Torvalds 31637d227cd2SSridhar Samudrala if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 31647d227cd2SSridhar Samudrala return -EHOSTUNREACH; /* Routing failure or similar. */ 31657d227cd2SSridhar Samudrala 31660c54b85fSIlpo Järvinen cur_mss = tcp_current_mss(sk); 31677d227cd2SSridhar Samudrala 31681da177e4SLinus Torvalds /* If receiver has shrunk his window, and skb is out of 31691da177e4SLinus Torvalds * new window, do not retransmit it. The exception is the 31701da177e4SLinus Torvalds * case, when window is shrunk to zero. In this case 31711da177e4SLinus Torvalds * our retransmit serves as a zero window probe. 31721da177e4SLinus Torvalds */ 31739d4fb27dSJoe Perches if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && 31749d4fb27dSJoe Perches TCP_SKB_CB(skb)->seq != tp->snd_una) 31751da177e4SLinus Torvalds return -EAGAIN; 31761da177e4SLinus Torvalds 317710d3be56SEric Dumazet len = cur_mss * segs; 317810d3be56SEric Dumazet if (skb->len > len) { 317975c119afSEric Dumazet if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len, 318075c119afSEric Dumazet cur_mss, GFP_ATOMIC)) 31811da177e4SLinus Torvalds return -ENOMEM; /* We'll try again later. */ 318202276f3cSIlpo Järvinen } else { 3183c4777efaSEric Dumazet if (skb_unclone_keeptruesize(skb, GFP_ATOMIC)) 3184c52e2421SEric Dumazet return -ENOMEM; 318510d3be56SEric Dumazet 318610d3be56SEric Dumazet diff = tcp_skb_pcount(skb); 318710d3be56SEric Dumazet tcp_set_skb_tso_segs(skb, cur_mss); 318810d3be56SEric Dumazet diff -= tcp_skb_pcount(skb); 318910d3be56SEric Dumazet if (diff) 319010d3be56SEric Dumazet tcp_adjust_pcount(sk, skb, diff); 319110d3be56SEric Dumazet if (skb->len < cur_mss) 319210d3be56SEric Dumazet tcp_retrans_try_collapse(sk, skb, cur_mss); 31931da177e4SLinus Torvalds } 31941da177e4SLinus Torvalds 319549213555SDaniel Borkmann /* RFC3168, section 6.1.1.1. ECN fallback */ 319649213555SDaniel Borkmann if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN) 319749213555SDaniel Borkmann tcp_ecn_clear_syn(sk, skb); 319849213555SDaniel Borkmann 3199678550c6SYuchung Cheng /* Update global and local TCP statistics. */ 3200678550c6SYuchung Cheng segs = tcp_skb_pcount(skb); 3201678550c6SYuchung Cheng TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs); 3202678550c6SYuchung Cheng if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 3203678550c6SYuchung Cheng __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 3204678550c6SYuchung Cheng tp->total_retrans += segs; 3205fb31c9b9SWei Wang tp->bytes_retrans += skb->len; 3206678550c6SYuchung Cheng 320750bceae9SThomas Graf /* make sure skb->data is aligned on arches that require it 320850bceae9SThomas Graf * and check if ack-trimming & collapsing extended the headroom 320950bceae9SThomas Graf * beyond what csum_start can cover. 321050bceae9SThomas Graf */ 321150bceae9SThomas Graf if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) || 321250bceae9SThomas Graf skb_headroom(skb) >= 0xFFFF)) { 321310a81980SEric Dumazet struct sk_buff *nskb; 321410a81980SEric Dumazet 3215e2080072SEric Dumazet tcp_skb_tsorted_save(skb) { 321610a81980SEric Dumazet nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC); 321707f8e4d0SFlorian Westphal if (nskb) { 321807f8e4d0SFlorian Westphal nskb->dev = NULL; 321907f8e4d0SFlorian Westphal err = tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC); 322007f8e4d0SFlorian Westphal } else { 322107f8e4d0SFlorian Westphal err = -ENOBUFS; 322207f8e4d0SFlorian Westphal } 3223e2080072SEric Dumazet } tcp_skb_tsorted_restore(skb); 3224e2080072SEric Dumazet 32255889e2c0SYousuk Seung if (!err) { 3226a7a25630SEric Dumazet tcp_update_skb_after_send(sk, skb, tp->tcp_wstamp_ns); 32275889e2c0SYousuk Seung tcp_rate_skb_sent(sk, skb); 32285889e2c0SYousuk Seung } 3229117632e6SEric Dumazet } else { 3230c84a5711SYuchung Cheng err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 3231117632e6SEric Dumazet } 3232c84a5711SYuchung Cheng 32337f12422cSYuchung Cheng /* To avoid taking spuriously low RTT samples based on a timestamp 32347f12422cSYuchung Cheng * for a transmit that never happened, always mark EVER_RETRANS 32357f12422cSYuchung Cheng */ 32367f12422cSYuchung Cheng TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS; 32377f12422cSYuchung Cheng 3238a31ad29eSLawrence Brakmo if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RETRANS_CB_FLAG)) 3239a31ad29eSLawrence Brakmo tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RETRANS_CB, 3240a31ad29eSLawrence Brakmo TCP_SKB_CB(skb)->seq, segs, err); 3241a31ad29eSLawrence Brakmo 3242fc9f3501SEric Dumazet if (likely(!err)) { 3243e086101bSCong Wang trace_tcp_retransmit_skb(sk, skb); 3244678550c6SYuchung Cheng } else if (err != -EBUSY) { 3245ec641b39SYuchung Cheng NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL, segs); 3246fc9f3501SEric Dumazet } 3247c84a5711SYuchung Cheng return err; 324893b174adSYuchung Cheng } 324993b174adSYuchung Cheng 325010d3be56SEric Dumazet int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) 325193b174adSYuchung Cheng { 325293b174adSYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 325310d3be56SEric Dumazet int err = __tcp_retransmit_skb(sk, skb, segs); 32541da177e4SLinus Torvalds 32551da177e4SLinus Torvalds if (err == 0) { 32561da177e4SLinus Torvalds #if FASTRETRANS_DEBUG > 0 32571da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 3258e87cc472SJoe Perches net_dbg_ratelimited("retrans_out leaked\n"); 32591da177e4SLinus Torvalds } 32601da177e4SLinus Torvalds #endif 32611da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 32621da177e4SLinus Torvalds tp->retrans_out += tcp_skb_pcount(skb); 32637ae18975SYuchung Cheng } 32641da177e4SLinus Torvalds 32657ae18975SYuchung Cheng /* Save stamp of the first (attempted) retransmit. */ 32661da177e4SLinus Torvalds if (!tp->retrans_stamp) 32677faee5c0SEric Dumazet tp->retrans_stamp = tcp_skb_timestamp(skb); 32681da177e4SLinus Torvalds 32696e08d5e3SYuchung Cheng if (tp->undo_retrans < 0) 32706e08d5e3SYuchung Cheng tp->undo_retrans = 0; 32716e08d5e3SYuchung Cheng tp->undo_retrans += tcp_skb_pcount(skb); 32721da177e4SLinus Torvalds return err; 32731da177e4SLinus Torvalds } 32741da177e4SLinus Torvalds 32751da177e4SLinus Torvalds /* This gets called after a retransmit timeout, and the initially 32761da177e4SLinus Torvalds * retransmitted data is acknowledged. It tries to continue 32771da177e4SLinus Torvalds * resending the rest of the retransmit queue, until either 32781da177e4SLinus Torvalds * we've sent it all or the congestion window limit is reached. 32791da177e4SLinus Torvalds */ 32801da177e4SLinus Torvalds void tcp_xmit_retransmit_queue(struct sock *sk) 32811da177e4SLinus Torvalds { 32826687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 3283b9f1f1ceSEric Dumazet struct sk_buff *skb, *rtx_head, *hole = NULL; 32841da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 3285916e6d1aSEric Dumazet bool rearm_timer = false; 3286840a3cbeSYuchung Cheng u32 max_segs; 328761eb55f4SIlpo Järvinen int mib_idx; 32886a438bbeSStephen Hemminger 328945e77d31SIlpo Järvinen if (!tp->packets_out) 329045e77d31SIlpo Järvinen return; 329145e77d31SIlpo Järvinen 329275c119afSEric Dumazet rtx_head = tcp_rtx_queue_head(sk); 3293b9f1f1ceSEric Dumazet skb = tp->retransmit_skb_hint ?: rtx_head; 3294ed6e7268SNeal Cardwell max_segs = tcp_tso_segs(sk, tcp_current_mss(sk)); 329575c119afSEric Dumazet skb_rbtree_walk_from(skb) { 3296dca0aaf8SEric Dumazet __u8 sacked; 329710d3be56SEric Dumazet int segs; 32981da177e4SLinus Torvalds 3299218af599SEric Dumazet if (tcp_pacing_check(sk)) 3300218af599SEric Dumazet break; 3301218af599SEric Dumazet 33026a438bbeSStephen Hemminger /* we could do better than to assign each time */ 330351456b29SIan Morris if (!hole) 33046a438bbeSStephen Hemminger tp->retransmit_skb_hint = skb; 33056a438bbeSStephen Hemminger 330640570375SEric Dumazet segs = tcp_snd_cwnd(tp) - tcp_packets_in_flight(tp); 330710d3be56SEric Dumazet if (segs <= 0) 3308916e6d1aSEric Dumazet break; 3309dca0aaf8SEric Dumazet sacked = TCP_SKB_CB(skb)->sacked; 3310a3d2e9f8SEric Dumazet /* In case tcp_shift_skb_data() have aggregated large skbs, 3311a3d2e9f8SEric Dumazet * we need to make sure not sending too bigs TSO packets 3312a3d2e9f8SEric Dumazet */ 3313a3d2e9f8SEric Dumazet segs = min_t(int, segs, max_segs); 33140e1c54c2SIlpo Järvinen 3315840a3cbeSYuchung Cheng if (tp->retrans_out >= tp->lost_out) { 3316006f582cSIlpo Järvinen break; 33170e1c54c2SIlpo Järvinen } else if (!(sacked & TCPCB_LOST)) { 331851456b29SIan Morris if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) 33190e1c54c2SIlpo Järvinen hole = skb; 332061eb55f4SIlpo Järvinen continue; 33211da177e4SLinus Torvalds 33220e1c54c2SIlpo Järvinen } else { 33230e1c54c2SIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Loss) 33240e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFASTRETRANS; 33250e1c54c2SIlpo Järvinen else 33260e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; 33270e1c54c2SIlpo Järvinen } 33280e1c54c2SIlpo Järvinen 33290e1c54c2SIlpo Järvinen if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) 333061eb55f4SIlpo Järvinen continue; 333140b215e5SPavel Emelyanov 3332f9616c35SEric Dumazet if (tcp_small_queue_check(sk, skb, 1)) 3333916e6d1aSEric Dumazet break; 3334f9616c35SEric Dumazet 333510d3be56SEric Dumazet if (tcp_retransmit_skb(sk, skb, segs)) 3336916e6d1aSEric Dumazet break; 333724ab6becSYuchung Cheng 3338de1d6578SYuchung Cheng NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb)); 33391da177e4SLinus Torvalds 3340684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 3341a262f0cdSNandita Dukkipati tp->prr_out += tcp_skb_pcount(skb); 3342a262f0cdSNandita Dukkipati 334375c119afSEric Dumazet if (skb == rtx_head && 334457dde7f7SYuchung Cheng icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT) 3345916e6d1aSEric Dumazet rearm_timer = true; 3346916e6d1aSEric Dumazet 3347916e6d1aSEric Dumazet } 3348916e6d1aSEric Dumazet if (rearm_timer) 33493f80e08fSEric Dumazet tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 33503f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, 33518dc242adSEric Dumazet TCP_RTO_MAX); 33521da177e4SLinus Torvalds } 33531da177e4SLinus Torvalds 3354d83769a5SEric Dumazet /* We allow to exceed memory limits for FIN packets to expedite 3355d83769a5SEric Dumazet * connection tear down and (memory) recovery. 3356845704a5SEric Dumazet * Otherwise tcp_send_fin() could be tempted to either delay FIN 3357845704a5SEric Dumazet * or even be forced to close flow without any FIN. 3358a6c5ea4cSEric Dumazet * In general, we want to allow one skb per socket to avoid hangs 3359a6c5ea4cSEric Dumazet * with edge trigger epoll() 3360d83769a5SEric Dumazet */ 3361a6c5ea4cSEric Dumazet void sk_forced_mem_schedule(struct sock *sk, int size) 3362d83769a5SEric Dumazet { 3363e805605cSJohannes Weiner int amt; 3364d83769a5SEric Dumazet 3365d83769a5SEric Dumazet if (size <= sk->sk_forward_alloc) 3366d83769a5SEric Dumazet return; 3367d83769a5SEric Dumazet amt = sk_mem_pages(size); 3368d83769a5SEric Dumazet sk->sk_forward_alloc += amt * SK_MEM_QUANTUM; 3369e805605cSJohannes Weiner sk_memory_allocated_add(sk, amt); 3370e805605cSJohannes Weiner 3371baac50bbSJohannes Weiner if (mem_cgroup_sockets_enabled && sk->sk_memcg) 33724b1327beSWei Wang mem_cgroup_charge_skmem(sk->sk_memcg, amt, 33734b1327beSWei Wang gfp_memcg_charge() | __GFP_NOFAIL); 3374d83769a5SEric Dumazet } 3375d83769a5SEric Dumazet 3376845704a5SEric Dumazet /* Send a FIN. The caller locks the socket for us. 3377845704a5SEric Dumazet * We should try to send a FIN packet really hard, but eventually give up. 33781da177e4SLinus Torvalds */ 33791da177e4SLinus Torvalds void tcp_send_fin(struct sock *sk) 33801da177e4SLinus Torvalds { 3381ee2aabd3SEric Dumazet struct sk_buff *skb, *tskb, *tail = tcp_write_queue_tail(sk); 33821da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 33831da177e4SLinus Torvalds 3384845704a5SEric Dumazet /* Optimization, tack on the FIN if we have one skb in write queue and 3385845704a5SEric Dumazet * this skb was not yet sent, or we are under memory pressure. 3386845704a5SEric Dumazet * Note: in the latter case, FIN packet will be sent after a timeout, 3387845704a5SEric Dumazet * as TCP stack thinks it has already been transmitted. 33881da177e4SLinus Torvalds */ 3389ee2aabd3SEric Dumazet tskb = tail; 339075c119afSEric Dumazet if (!tskb && tcp_under_memory_pressure(sk)) 339175c119afSEric Dumazet tskb = skb_rb_last(&sk->tcp_rtx_queue); 339275c119afSEric Dumazet 339375c119afSEric Dumazet if (tskb) { 3394845704a5SEric Dumazet TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN; 3395845704a5SEric Dumazet TCP_SKB_CB(tskb)->end_seq++; 33961da177e4SLinus Torvalds tp->write_seq++; 3397ee2aabd3SEric Dumazet if (!tail) { 3398845704a5SEric Dumazet /* This means tskb was already sent. 3399845704a5SEric Dumazet * Pretend we included the FIN on previous transmit. 3400845704a5SEric Dumazet * We need to set tp->snd_nxt to the value it would have 3401845704a5SEric Dumazet * if FIN had been sent. This is because retransmit path 3402845704a5SEric Dumazet * does not change tp->snd_nxt. 3403845704a5SEric Dumazet */ 3404e0d694d6SEric Dumazet WRITE_ONCE(tp->snd_nxt, tp->snd_nxt + 1); 3405845704a5SEric Dumazet return; 3406845704a5SEric Dumazet } 34071da177e4SLinus Torvalds } else { 3408845704a5SEric Dumazet skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation); 3409d1edc085SColin Ian King if (unlikely(!skb)) 3410845704a5SEric Dumazet return; 3411d1edc085SColin Ian King 3412e2080072SEric Dumazet INIT_LIST_HEAD(&skb->tcp_tsorted_anchor); 3413d83769a5SEric Dumazet skb_reserve(skb, MAX_TCP_HEADER); 3414a6c5ea4cSEric Dumazet sk_forced_mem_schedule(sk, skb->truesize); 34151da177e4SLinus Torvalds /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 3416e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tp->write_seq, 3417a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_FIN); 34181da177e4SLinus Torvalds tcp_queue_skb(sk, skb); 34191da177e4SLinus Torvalds } 3420845704a5SEric Dumazet __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF); 34211da177e4SLinus Torvalds } 34221da177e4SLinus Torvalds 34231da177e4SLinus Torvalds /* We get here when a process closes a file descriptor (either due to 34241da177e4SLinus Torvalds * an explicit close() or as a byproduct of exit()'ing) and there 34251da177e4SLinus Torvalds * was unread data in the receive queue. This behavior is recommended 342665bb723cSGerrit Renker * by RFC 2525, section 2.17. -DaveM 34271da177e4SLinus Torvalds */ 3428dd0fc66fSAl Viro void tcp_send_active_reset(struct sock *sk, gfp_t priority) 34291da177e4SLinus Torvalds { 34301da177e4SLinus Torvalds struct sk_buff *skb; 34311da177e4SLinus Torvalds 34327cc2b043SGao Feng TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); 34337cc2b043SGao Feng 34341da177e4SLinus Torvalds /* NOTE: No TCP options attached and we never retransmit this. */ 34351da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, priority); 34361da177e4SLinus Torvalds if (!skb) { 34374e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 34381da177e4SLinus Torvalds return; 34391da177e4SLinus Torvalds } 34401da177e4SLinus Torvalds 34411da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 34421da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 3443e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), 3444a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_RST); 34459a568de4SEric Dumazet tcp_mstamp_refresh(tcp_sk(sk)); 34461da177e4SLinus Torvalds /* Send it off. */ 3447dfb4b9dcSDavid S. Miller if (tcp_transmit_skb(sk, skb, 0, priority)) 34484e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 3449c24b14c4SSong Liu 3450c24b14c4SSong Liu /* skb of trace_tcp_send_reset() keeps the skb that caused RST, 3451c24b14c4SSong Liu * skb here is different to the troublesome skb, so use NULL 3452c24b14c4SSong Liu */ 3453c24b14c4SSong Liu trace_tcp_send_reset(sk, NULL); 34541da177e4SLinus Torvalds } 34551da177e4SLinus Torvalds 345667edfef7SAndi Kleen /* Send a crossed SYN-ACK during socket establishment. 345767edfef7SAndi Kleen * WARNING: This routine must only be called when we have already sent 34581da177e4SLinus Torvalds * a SYN packet that crossed the incoming SYN that caused this routine 34591da177e4SLinus Torvalds * to get called. If this assumption fails then the initial rcv_wnd 34601da177e4SLinus Torvalds * and rcv_wscale values will not be correct. 34611da177e4SLinus Torvalds */ 34621da177e4SLinus Torvalds int tcp_send_synack(struct sock *sk) 34631da177e4SLinus Torvalds { 34641da177e4SLinus Torvalds struct sk_buff *skb; 34651da177e4SLinus Torvalds 346675c119afSEric Dumazet skb = tcp_rtx_queue_head(sk); 346751456b29SIan Morris if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 346875c119afSEric Dumazet pr_err("%s: wrong queue state\n", __func__); 34691da177e4SLinus Torvalds return -EFAULT; 34701da177e4SLinus Torvalds } 34714de075e0SEric Dumazet if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { 34721da177e4SLinus Torvalds if (skb_cloned(skb)) { 3473e2080072SEric Dumazet struct sk_buff *nskb; 3474e2080072SEric Dumazet 3475e2080072SEric Dumazet tcp_skb_tsorted_save(skb) { 3476e2080072SEric Dumazet nskb = skb_copy(skb, GFP_ATOMIC); 3477e2080072SEric Dumazet } tcp_skb_tsorted_restore(skb); 347851456b29SIan Morris if (!nskb) 34791da177e4SLinus Torvalds return -ENOMEM; 3480e2080072SEric Dumazet INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor); 34812bec445fSEric Dumazet tcp_highest_sack_replace(sk, skb, nskb); 348275c119afSEric Dumazet tcp_rtx_queue_unlink_and_free(skb, sk); 3483f4a775d1SEric Dumazet __skb_header_release(nskb); 348475c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb); 3485ab4e846aSEric Dumazet sk_wmem_queued_add(sk, nskb->truesize); 34863ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 34871da177e4SLinus Torvalds skb = nskb; 34881da177e4SLinus Torvalds } 34891da177e4SLinus Torvalds 34904de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; 3491735d3831SFlorian Westphal tcp_ecn_send_synack(sk, skb); 34921da177e4SLinus Torvalds } 3493dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 34941da177e4SLinus Torvalds } 34951da177e4SLinus Torvalds 34964aea39c1SEric Dumazet /** 3497331fca43SMartin KaFai Lau * tcp_make_synack - Allocate one skb and build a SYNACK packet. 3498331fca43SMartin KaFai Lau * @sk: listener socket 3499331fca43SMartin KaFai Lau * @dst: dst entry attached to the SYNACK. It is consumed and caller 3500331fca43SMartin KaFai Lau * should not use it again. 3501331fca43SMartin KaFai Lau * @req: request_sock pointer 3502331fca43SMartin KaFai Lau * @foc: cookie for tcp fast open 3503331fca43SMartin KaFai Lau * @synack_type: Type of synack to prepare 3504331fca43SMartin KaFai Lau * @syn_skb: SYN packet just received. It could be NULL for rtx case. 35054aea39c1SEric Dumazet */ 35065d062de7SEric Dumazet struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst, 3507e6b4d113SWilliam Allen Simpson struct request_sock *req, 3508ca6fb065SEric Dumazet struct tcp_fastopen_cookie *foc, 3509331fca43SMartin KaFai Lau enum tcp_synack_type synack_type, 3510331fca43SMartin KaFai Lau struct sk_buff *syn_skb) 35111da177e4SLinus Torvalds { 35122e6599cbSArnaldo Carvalho de Melo struct inet_request_sock *ireq = inet_rsk(req); 35135d062de7SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 351480f03e27SEric Dumazet struct tcp_md5sig_key *md5 = NULL; 35155d062de7SEric Dumazet struct tcp_out_options opts; 35165d062de7SEric Dumazet struct sk_buff *skb; 3517bd0388aeSWilliam Allen Simpson int tcp_header_size; 35185d062de7SEric Dumazet struct tcphdr *th; 3519f5fff5dcSTom Quetchenbach int mss; 3520a842fe14SEric Dumazet u64 now; 35211da177e4SLinus Torvalds 3522ca6fb065SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 35234aea39c1SEric Dumazet if (unlikely(!skb)) { 35244aea39c1SEric Dumazet dst_release(dst); 35251da177e4SLinus Torvalds return NULL; 35264aea39c1SEric Dumazet } 35271da177e4SLinus Torvalds /* Reserve space for headers. */ 35281da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 35291da177e4SLinus Torvalds 3530b3d05147SEric Dumazet switch (synack_type) { 3531b3d05147SEric Dumazet case TCP_SYNACK_NORMAL: 35329e17f8a4SEric Dumazet skb_set_owner_w(skb, req_to_sk(req)); 3533b3d05147SEric Dumazet break; 3534b3d05147SEric Dumazet case TCP_SYNACK_COOKIE: 3535b3d05147SEric Dumazet /* Under synflood, we do not attach skb to a socket, 3536b3d05147SEric Dumazet * to avoid false sharing. 3537b3d05147SEric Dumazet */ 3538b3d05147SEric Dumazet break; 3539b3d05147SEric Dumazet case TCP_SYNACK_FASTOPEN: 3540ca6fb065SEric Dumazet /* sk is a const pointer, because we want to express multiple 3541ca6fb065SEric Dumazet * cpu might call us concurrently. 3542ca6fb065SEric Dumazet * sk->sk_wmem_alloc in an atomic, we can promote to rw. 3543ca6fb065SEric Dumazet */ 3544ca6fb065SEric Dumazet skb_set_owner_w(skb, (struct sock *)sk); 3545b3d05147SEric Dumazet break; 3546ca6fb065SEric Dumazet } 35474aea39c1SEric Dumazet skb_dst_set(skb, dst); 35481da177e4SLinus Torvalds 35493541f9e8SEric Dumazet mss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 3550f5fff5dcSTom Quetchenbach 355133ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 3552a842fe14SEric Dumazet now = tcp_clock_ns(); 35538b5f12d0SFlorian Westphal #ifdef CONFIG_SYN_COOKIES 3554f8ace8d9SFlorian Westphal if (unlikely(synack_type == TCP_SYNACK_COOKIE && ireq->tstamp_ok)) 3555a1ac9c8aSMartin KaFai Lau skb_set_delivery_time(skb, cookie_init_timestamp(req, now), 3556a1ac9c8aSMartin KaFai Lau true); 35578b5f12d0SFlorian Westphal else 35588b5f12d0SFlorian Westphal #endif 35599e450c1eSYuchung Cheng { 3560a1ac9c8aSMartin KaFai Lau skb_set_delivery_time(skb, now, true); 35619e450c1eSYuchung Cheng if (!tcp_rsk(req)->snt_synack) /* Timestamp first SYNACK */ 35629e450c1eSYuchung Cheng tcp_rsk(req)->snt_synack = tcp_skb_timestamp_us(skb); 35639e450c1eSYuchung Cheng } 356480f03e27SEric Dumazet 356580f03e27SEric Dumazet #ifdef CONFIG_TCP_MD5SIG 356680f03e27SEric Dumazet rcu_read_lock(); 3567fd3a154aSEric Dumazet md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req)); 356880f03e27SEric Dumazet #endif 356958d607d3SEric Dumazet skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4); 3570331fca43SMartin KaFai Lau /* bpf program will be interested in the tcp_flags */ 3571331fca43SMartin KaFai Lau TCP_SKB_CB(skb)->tcp_flags = TCPHDR_SYN | TCPHDR_ACK; 357260e2a778SUrsula Braun tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5, 3573331fca43SMartin KaFai Lau foc, synack_type, 3574331fca43SMartin KaFai Lau syn_skb) + sizeof(*th); 357533ad798cSAdam Langley 3576aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 3577aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 35781da177e4SLinus Torvalds 3579ea1627c2SEric Dumazet th = (struct tcphdr *)skb->data; 35801da177e4SLinus Torvalds memset(th, 0, sizeof(struct tcphdr)); 35811da177e4SLinus Torvalds th->syn = 1; 35821da177e4SLinus Torvalds th->ack = 1; 35836ac705b1SEric Dumazet tcp_ecn_make_synack(req, th); 3584b44084c2SEric Dumazet th->source = htons(ireq->ir_num); 3585634fb979SEric Dumazet th->dest = ireq->ir_rmt_port; 3586e05a90ecSJamal Hadi Salim skb->mark = ireq->ir_mark; 35873b117750SEric Dumazet skb->ip_summed = CHECKSUM_PARTIAL; 35883b117750SEric Dumazet th->seq = htonl(tcp_rsk(req)->snt_isn); 35898336886fSJerry Chu /* XXX data is queued and acked as is. No buffer/window check */ 35908336886fSJerry Chu th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt); 35911da177e4SLinus Torvalds 35921da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 3593ed53d0abSEric Dumazet th->window = htons(min(req->rsk_rcv_wnd, 65535U)); 3594ea66758cSPaolo Abeni tcp_options_write(th, NULL, &opts); 35951da177e4SLinus Torvalds th->doff = (tcp_header_size >> 2); 359690bbcc60SEric Dumazet __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS); 3597cfb6eeb4SYOSHIFUJI Hideaki 3598cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 3599cfb6eeb4SYOSHIFUJI Hideaki /* Okay, we have all we need - do the md5 hash if needed */ 360080f03e27SEric Dumazet if (md5) 3601bd0388aeSWilliam Allen Simpson tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, 360239f8e58eSEric Dumazet md5, req_to_sk(req), skb); 360380f03e27SEric Dumazet rcu_read_unlock(); 3604cfb6eeb4SYOSHIFUJI Hideaki #endif 3605cfb6eeb4SYOSHIFUJI Hideaki 3606331fca43SMartin KaFai Lau bpf_skops_write_hdr_opt((struct sock *)sk, skb, req, syn_skb, 3607331fca43SMartin KaFai Lau synack_type, &opts); 3608331fca43SMartin KaFai Lau 3609a1ac9c8aSMartin KaFai Lau skb_set_delivery_time(skb, now, true); 3610a842fe14SEric Dumazet tcp_add_tx_delay(skb, tp); 3611a842fe14SEric Dumazet 36121da177e4SLinus Torvalds return skb; 36131da177e4SLinus Torvalds } 36144bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_make_synack); 36151da177e4SLinus Torvalds 361681164413SDaniel Borkmann static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst) 361781164413SDaniel Borkmann { 361881164413SDaniel Borkmann struct inet_connection_sock *icsk = inet_csk(sk); 361981164413SDaniel Borkmann const struct tcp_congestion_ops *ca; 362081164413SDaniel Borkmann u32 ca_key = dst_metric(dst, RTAX_CC_ALGO); 362181164413SDaniel Borkmann 362281164413SDaniel Borkmann if (ca_key == TCP_CA_UNSPEC) 362381164413SDaniel Borkmann return; 362481164413SDaniel Borkmann 362581164413SDaniel Borkmann rcu_read_lock(); 362681164413SDaniel Borkmann ca = tcp_ca_find_key(ca_key); 36270baf26b0SMartin KaFai Lau if (likely(ca && bpf_try_module_get(ca, ca->owner))) { 36280baf26b0SMartin KaFai Lau bpf_module_put(icsk->icsk_ca_ops, icsk->icsk_ca_ops->owner); 362981164413SDaniel Borkmann icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst); 363081164413SDaniel Borkmann icsk->icsk_ca_ops = ca; 363181164413SDaniel Borkmann } 363281164413SDaniel Borkmann rcu_read_unlock(); 363381164413SDaniel Borkmann } 363481164413SDaniel Borkmann 363567edfef7SAndi Kleen /* Do all connect socket setups that can be done AF independent. */ 3636f7e56a76Sstephen hemminger static void tcp_connect_init(struct sock *sk) 36371da177e4SLinus Torvalds { 3638cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 36391da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 36401da177e4SLinus Torvalds __u8 rcv_wscale; 364113d3b1ebSLawrence Brakmo u32 rcv_wnd; 36421da177e4SLinus Torvalds 36431da177e4SLinus Torvalds /* We'll fix this up when we get a response from the other end. 36441da177e4SLinus Torvalds * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 36451da177e4SLinus Torvalds */ 36465d2ed052SEric Dumazet tp->tcp_header_len = sizeof(struct tcphdr); 36473666f666SKuniyuki Iwashima if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_timestamps)) 36485d2ed052SEric Dumazet tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED; 36491da177e4SLinus Torvalds 3650cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 365100db4124SIan Morris if (tp->af_specific->md5_lookup(sk, sk)) 3652cfb6eeb4SYOSHIFUJI Hideaki tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 3653cfb6eeb4SYOSHIFUJI Hideaki #endif 3654cfb6eeb4SYOSHIFUJI Hideaki 36551da177e4SLinus Torvalds /* If user gave his TCP_MAXSEG, record it to clamp */ 36561da177e4SLinus Torvalds if (tp->rx_opt.user_mss) 36571da177e4SLinus Torvalds tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 36581da177e4SLinus Torvalds tp->max_window = 0; 36595d424d5aSJohn Heffner tcp_mtup_init(sk); 36601da177e4SLinus Torvalds tcp_sync_mss(sk, dst_mtu(dst)); 36611da177e4SLinus Torvalds 366281164413SDaniel Borkmann tcp_ca_dst_init(sk, dst); 366381164413SDaniel Borkmann 36641da177e4SLinus Torvalds if (!tp->window_clamp) 36651da177e4SLinus Torvalds tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 36663541f9e8SEric Dumazet tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 3667f5fff5dcSTom Quetchenbach 36681da177e4SLinus Torvalds tcp_initialize_rcv_mss(sk); 36691da177e4SLinus Torvalds 3670e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 3671e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 3672e88c64f0SHagen Paul Pfeifer (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) 3673e88c64f0SHagen Paul Pfeifer tp->window_clamp = tcp_full_space(sk); 3674e88c64f0SHagen Paul Pfeifer 367513d3b1ebSLawrence Brakmo rcv_wnd = tcp_rwnd_init_bpf(sk); 367613d3b1ebSLawrence Brakmo if (rcv_wnd == 0) 367713d3b1ebSLawrence Brakmo rcv_wnd = dst_metric(dst, RTAX_INITRWND); 367813d3b1ebSLawrence Brakmo 3679ceef9ab6SEric Dumazet tcp_select_initial_window(sk, tcp_full_space(sk), 36801da177e4SLinus Torvalds tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 36811da177e4SLinus Torvalds &tp->rcv_wnd, 36821da177e4SLinus Torvalds &tp->window_clamp, 36833666f666SKuniyuki Iwashima READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_window_scaling), 368431d12926Slaurent chavey &rcv_wscale, 368513d3b1ebSLawrence Brakmo rcv_wnd); 36861da177e4SLinus Torvalds 36871da177e4SLinus Torvalds tp->rx_opt.rcv_wscale = rcv_wscale; 36881da177e4SLinus Torvalds tp->rcv_ssthresh = tp->rcv_wnd; 36891da177e4SLinus Torvalds 36901da177e4SLinus Torvalds sk->sk_err = 0; 36911da177e4SLinus Torvalds sock_reset_flag(sk, SOCK_DONE); 36921da177e4SLinus Torvalds tp->snd_wnd = 0; 3693ee7537b6SHantzis Fotis tcp_init_wl(tp, 0); 36947f582b24SEric Dumazet tcp_write_queue_purge(sk); 36951da177e4SLinus Torvalds tp->snd_una = tp->write_seq; 36961da177e4SLinus Torvalds tp->snd_sml = tp->write_seq; 369733f5f57eSIlpo Järvinen tp->snd_up = tp->write_seq; 3698e0d694d6SEric Dumazet WRITE_ONCE(tp->snd_nxt, tp->write_seq); 3699ee995283SPavel Emelyanov 3700ee995283SPavel Emelyanov if (likely(!tp->repair)) 37011da177e4SLinus Torvalds tp->rcv_nxt = 0; 3702c7781a6eSAndrew Vagin else 370370eabf0eSEric Dumazet tp->rcv_tstamp = tcp_jiffies32; 3704ee995283SPavel Emelyanov tp->rcv_wup = tp->rcv_nxt; 37057db48e98SEric Dumazet WRITE_ONCE(tp->copied_seq, tp->rcv_nxt); 37061da177e4SLinus Torvalds 37078550f328SLawrence Brakmo inet_csk(sk)->icsk_rto = tcp_timeout_init(sk); 3708463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_retransmits = 0; 37091da177e4SLinus Torvalds tcp_clear_retrans(tp); 37101da177e4SLinus Torvalds } 37111da177e4SLinus Torvalds 3712783237e8SYuchung Cheng static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) 3713783237e8SYuchung Cheng { 3714783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3715783237e8SYuchung Cheng struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 3716783237e8SYuchung Cheng 3717783237e8SYuchung Cheng tcb->end_seq += skb->len; 3718f4a775d1SEric Dumazet __skb_header_release(skb); 3719ab4e846aSEric Dumazet sk_wmem_queued_add(sk, skb->truesize); 3720783237e8SYuchung Cheng sk_mem_charge(sk, skb->truesize); 37210f317464SEric Dumazet WRITE_ONCE(tp->write_seq, tcb->end_seq); 3722783237e8SYuchung Cheng tp->packets_out += tcp_skb_pcount(skb); 3723783237e8SYuchung Cheng } 3724783237e8SYuchung Cheng 3725783237e8SYuchung Cheng /* Build and send a SYN with data and (cached) Fast Open cookie. However, 3726783237e8SYuchung Cheng * queue a data-only packet after the regular SYN, such that regular SYNs 3727783237e8SYuchung Cheng * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges 3728783237e8SYuchung Cheng * only the SYN sequence, the data are retransmitted in the first ACK. 3729783237e8SYuchung Cheng * If cookie is not cached or other error occurs, falls back to send a 3730783237e8SYuchung Cheng * regular SYN with Fast Open cookie request option. 3731783237e8SYuchung Cheng */ 3732783237e8SYuchung Cheng static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) 3733783237e8SYuchung Cheng { 3734ed0c99dcSJakub Kicinski struct inet_connection_sock *icsk = inet_csk(sk); 3735783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3736783237e8SYuchung Cheng struct tcp_fastopen_request *fo = tp->fastopen_req; 3737065263f4SWei Wang int space, err = 0; 3738355a901eSEric Dumazet struct sk_buff *syn_data; 3739783237e8SYuchung Cheng 374067da22d2SYuchung Cheng tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ 3741065263f4SWei Wang if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie)) 3742783237e8SYuchung Cheng goto fallback; 3743783237e8SYuchung Cheng 3744783237e8SYuchung Cheng /* MSS for SYN-data is based on cached MSS and bounded by PMTU and 3745783237e8SYuchung Cheng * user-MSS. Reserve maximum option space for middleboxes that add 3746783237e8SYuchung Cheng * private TCP options. The cost is reduced data space in SYN :( 3747783237e8SYuchung Cheng */ 37483541f9e8SEric Dumazet tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp); 3749ed0c99dcSJakub Kicinski /* Sync mss_cache after updating the mss_clamp */ 3750ed0c99dcSJakub Kicinski tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 37513541f9e8SEric Dumazet 3752ed0c99dcSJakub Kicinski space = __tcp_mtu_to_mss(sk, icsk->icsk_pmtu_cookie) - 3753783237e8SYuchung Cheng MAX_TCP_OPTION_SPACE; 3754783237e8SYuchung Cheng 3755f5ddcbbbSEric Dumazet space = min_t(size_t, space, fo->size); 3756f5ddcbbbSEric Dumazet 3757f5ddcbbbSEric Dumazet /* limit to order-0 allocations */ 3758f5ddcbbbSEric Dumazet space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER)); 3759f5ddcbbbSEric Dumazet 3760f8dd3b8dSEric Dumazet syn_data = tcp_stream_alloc_skb(sk, space, sk->sk_allocation, false); 3761355a901eSEric Dumazet if (!syn_data) 3762783237e8SYuchung Cheng goto fallback; 3763355a901eSEric Dumazet memcpy(syn_data->cb, syn->cb, sizeof(syn->cb)); 376407e100f9SEric Dumazet if (space) { 376507e100f9SEric Dumazet int copied = copy_from_iter(skb_put(syn_data, space), space, 376657be5bdaSAl Viro &fo->data->msg_iter); 376757be5bdaSAl Viro if (unlikely(!copied)) { 3768ba233b34SEric Dumazet tcp_skb_tsorted_anchor_cleanup(syn_data); 3769355a901eSEric Dumazet kfree_skb(syn_data); 3770783237e8SYuchung Cheng goto fallback; 3771783237e8SYuchung Cheng } 377257be5bdaSAl Viro if (copied != space) { 377357be5bdaSAl Viro skb_trim(syn_data, copied); 377457be5bdaSAl Viro space = copied; 377557be5bdaSAl Viro } 3776f859a448SWillem de Bruijn skb_zcopy_set(syn_data, fo->uarg, NULL); 377707e100f9SEric Dumazet } 3778355a901eSEric Dumazet /* No more data pending in inet_wait_for_connect() */ 3779355a901eSEric Dumazet if (space == fo->size) 3780355a901eSEric Dumazet fo->data = NULL; 3781355a901eSEric Dumazet fo->copied = space; 3782783237e8SYuchung Cheng 3783355a901eSEric Dumazet tcp_connect_queue_skb(sk, syn_data); 37840f87230dSFrancis Yan if (syn_data->len) 37850f87230dSFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_BUSY); 3786355a901eSEric Dumazet 3787355a901eSEric Dumazet err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation); 3788355a901eSEric Dumazet 3789a1ac9c8aSMartin KaFai Lau skb_set_delivery_time(syn, syn_data->skb_mstamp_ns, true); 3790355a901eSEric Dumazet 3791355a901eSEric Dumazet /* Now full SYN+DATA was cloned and sent (or not), 3792355a901eSEric Dumazet * remove the SYN from the original skb (syn_data) 3793355a901eSEric Dumazet * we keep in write queue in case of a retransmit, as we 3794355a901eSEric Dumazet * also have the SYN packet (with no data) in the same queue. 3795431a9124SEric Dumazet */ 3796355a901eSEric Dumazet TCP_SKB_CB(syn_data)->seq++; 3797355a901eSEric Dumazet TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH; 3798355a901eSEric Dumazet if (!err) { 379967da22d2SYuchung Cheng tp->syn_data = (fo->copied > 0); 380075c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data); 3801f19c29e3SYuchung Cheng NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT); 3802783237e8SYuchung Cheng goto done; 3803783237e8SYuchung Cheng } 3804783237e8SYuchung Cheng 380575c119afSEric Dumazet /* data was not sent, put it in write_queue */ 380675c119afSEric Dumazet __skb_queue_tail(&sk->sk_write_queue, syn_data); 3807b5b7db8dSEric Dumazet tp->packets_out -= tcp_skb_pcount(syn_data); 3808b5b7db8dSEric Dumazet 3809783237e8SYuchung Cheng fallback: 3810783237e8SYuchung Cheng /* Send a regular SYN with Fast Open cookie request option */ 3811783237e8SYuchung Cheng if (fo->cookie.len > 0) 3812783237e8SYuchung Cheng fo->cookie.len = 0; 3813783237e8SYuchung Cheng err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); 3814783237e8SYuchung Cheng if (err) 3815783237e8SYuchung Cheng tp->syn_fastopen = 0; 3816783237e8SYuchung Cheng done: 3817783237e8SYuchung Cheng fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ 3818783237e8SYuchung Cheng return err; 3819783237e8SYuchung Cheng } 3820783237e8SYuchung Cheng 382167edfef7SAndi Kleen /* Build a SYN and send it off. */ 38221da177e4SLinus Torvalds int tcp_connect(struct sock *sk) 38231da177e4SLinus Torvalds { 38241da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 38251da177e4SLinus Torvalds struct sk_buff *buff; 3826ee586811SEric Paris int err; 38271da177e4SLinus Torvalds 3828de525be2SLawrence Brakmo tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB, 0, NULL); 38298ba60924SEric Dumazet 38308ba60924SEric Dumazet if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 38318ba60924SEric Dumazet return -EHOSTUNREACH; /* Routing failure or similar. */ 38328ba60924SEric Dumazet 38331da177e4SLinus Torvalds tcp_connect_init(sk); 38341da177e4SLinus Torvalds 38352b916477SAndrey Vagin if (unlikely(tp->repair)) { 38362b916477SAndrey Vagin tcp_finish_connect(sk, NULL); 38372b916477SAndrey Vagin return 0; 38382b916477SAndrey Vagin } 38392b916477SAndrey Vagin 3840f8dd3b8dSEric Dumazet buff = tcp_stream_alloc_skb(sk, 0, sk->sk_allocation, true); 3841355a901eSEric Dumazet if (unlikely(!buff)) 38421da177e4SLinus Torvalds return -ENOBUFS; 38431da177e4SLinus Torvalds 3844a3433f35SChangli Gao tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); 38459a568de4SEric Dumazet tcp_mstamp_refresh(tp); 38469a568de4SEric Dumazet tp->retrans_stamp = tcp_time_stamp(tp); 3847783237e8SYuchung Cheng tcp_connect_queue_skb(sk, buff); 3848735d3831SFlorian Westphal tcp_ecn_send_syn(sk, buff); 384975c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, buff); 38501da177e4SLinus Torvalds 3851783237e8SYuchung Cheng /* Send off SYN; include data in Fast Open. */ 3852783237e8SYuchung Cheng err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : 3853783237e8SYuchung Cheng tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); 3854ee586811SEric Paris if (err == -ECONNREFUSED) 3855ee586811SEric Paris return err; 3856bd37a088SWei Yongjun 3857bd37a088SWei Yongjun /* We change tp->snd_nxt after the tcp_transmit_skb() call 3858bd37a088SWei Yongjun * in order to make this packet get counted in tcpOutSegs. 3859bd37a088SWei Yongjun */ 3860e0d694d6SEric Dumazet WRITE_ONCE(tp->snd_nxt, tp->write_seq); 3861bd37a088SWei Yongjun tp->pushed_seq = tp->write_seq; 3862b5b7db8dSEric Dumazet buff = tcp_send_head(sk); 3863b5b7db8dSEric Dumazet if (unlikely(buff)) { 3864e0d694d6SEric Dumazet WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(buff)->seq); 3865b5b7db8dSEric Dumazet tp->pushed_seq = TCP_SKB_CB(buff)->seq; 3866b5b7db8dSEric Dumazet } 386781cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); 38681da177e4SLinus Torvalds 38691da177e4SLinus Torvalds /* Timer for repeating the SYN until an answer. */ 38703f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 38713f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 38721da177e4SLinus Torvalds return 0; 38731da177e4SLinus Torvalds } 38744bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_connect); 38751da177e4SLinus Torvalds 38761da177e4SLinus Torvalds /* Send out a delayed ack, the caller does the policy checking 38771da177e4SLinus Torvalds * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 38781da177e4SLinus Torvalds * for details. 38791da177e4SLinus Torvalds */ 38801da177e4SLinus Torvalds void tcp_send_delayed_ack(struct sock *sk) 38811da177e4SLinus Torvalds { 3882463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 3883463c84b9SArnaldo Carvalho de Melo int ato = icsk->icsk_ack.ato; 38841da177e4SLinus Torvalds unsigned long timeout; 38851da177e4SLinus Torvalds 38861da177e4SLinus Torvalds if (ato > TCP_DELACK_MIN) { 3887463c84b9SArnaldo Carvalho de Melo const struct tcp_sock *tp = tcp_sk(sk); 38881da177e4SLinus Torvalds int max_ato = HZ / 2; 38891da177e4SLinus Torvalds 389031954cd8SWei Wang if (inet_csk_in_pingpong_mode(sk) || 3891056834d9SIlpo Järvinen (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 38921da177e4SLinus Torvalds max_ato = TCP_DELACK_MAX; 38931da177e4SLinus Torvalds 38941da177e4SLinus Torvalds /* Slow path, intersegment interval is "high". */ 38951da177e4SLinus Torvalds 38961da177e4SLinus Torvalds /* If some rtt estimate is known, use it to bound delayed ack. 3897463c84b9SArnaldo Carvalho de Melo * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 38981da177e4SLinus Torvalds * directly. 38991da177e4SLinus Torvalds */ 3900740b0f18SEric Dumazet if (tp->srtt_us) { 3901740b0f18SEric Dumazet int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3), 3902740b0f18SEric Dumazet TCP_DELACK_MIN); 39031da177e4SLinus Torvalds 39041da177e4SLinus Torvalds if (rtt < max_ato) 39051da177e4SLinus Torvalds max_ato = rtt; 39061da177e4SLinus Torvalds } 39071da177e4SLinus Torvalds 39081da177e4SLinus Torvalds ato = min(ato, max_ato); 39091da177e4SLinus Torvalds } 39101da177e4SLinus Torvalds 39112b8ee4f0SMartin KaFai Lau ato = min_t(u32, ato, inet_csk(sk)->icsk_delack_max); 39122b8ee4f0SMartin KaFai Lau 39131da177e4SLinus Torvalds /* Stay within the limit we were given */ 39141da177e4SLinus Torvalds timeout = jiffies + ato; 39151da177e4SLinus Torvalds 39161da177e4SLinus Torvalds /* Use new timeout only if there wasn't a older one earlier. */ 3917463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 3918b6b6d653SEric Dumazet /* If delack timer is about to expire, send ACK now. */ 3919b6b6d653SEric Dumazet if (time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 39201da177e4SLinus Torvalds tcp_send_ack(sk); 39211da177e4SLinus Torvalds return; 39221da177e4SLinus Torvalds } 39231da177e4SLinus Torvalds 3924463c84b9SArnaldo Carvalho de Melo if (!time_before(timeout, icsk->icsk_ack.timeout)) 3925463c84b9SArnaldo Carvalho de Melo timeout = icsk->icsk_ack.timeout; 39261da177e4SLinus Torvalds } 3927463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 3928463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.timeout = timeout; 3929463c84b9SArnaldo Carvalho de Melo sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 39301da177e4SLinus Torvalds } 39311da177e4SLinus Torvalds 39321da177e4SLinus Torvalds /* This routine sends an ack and also updates the window. */ 39332987babbSYuchung Cheng void __tcp_send_ack(struct sock *sk, u32 rcv_nxt) 39341da177e4SLinus Torvalds { 39351da177e4SLinus Torvalds struct sk_buff *buff; 39361da177e4SLinus Torvalds 3937058dc334SIlpo Järvinen /* If we have been reset, we may not send again. */ 3938058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3939058dc334SIlpo Järvinen return; 3940058dc334SIlpo Järvinen 39411da177e4SLinus Torvalds /* We are not putting this on the write queue, so 39421da177e4SLinus Torvalds * tcp_transmit_skb() will set the ownership to this 39431da177e4SLinus Torvalds * sock. 39441da177e4SLinus Torvalds */ 39457450aaf6SEric Dumazet buff = alloc_skb(MAX_TCP_HEADER, 39467450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 39477450aaf6SEric Dumazet if (unlikely(!buff)) { 3948a37c2134SEric Dumazet struct inet_connection_sock *icsk = inet_csk(sk); 3949a37c2134SEric Dumazet unsigned long delay; 3950a37c2134SEric Dumazet 3951a37c2134SEric Dumazet delay = TCP_DELACK_MAX << icsk->icsk_ack.retry; 3952a37c2134SEric Dumazet if (delay < TCP_RTO_MAX) 3953a37c2134SEric Dumazet icsk->icsk_ack.retry++; 3954463c84b9SArnaldo Carvalho de Melo inet_csk_schedule_ack(sk); 3955a37c2134SEric Dumazet icsk->icsk_ack.ato = TCP_ATO_MIN; 3956a37c2134SEric Dumazet inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, delay, TCP_RTO_MAX); 39571da177e4SLinus Torvalds return; 39581da177e4SLinus Torvalds } 39591da177e4SLinus Torvalds 39601da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 39611da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 3962a3433f35SChangli Gao tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); 39631da177e4SLinus Torvalds 396498781965SEric Dumazet /* We do not want pure acks influencing TCP Small Queues or fq/pacing 396598781965SEric Dumazet * too much. 396698781965SEric Dumazet * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784 396798781965SEric Dumazet */ 396898781965SEric Dumazet skb_set_tcp_pure_ack(buff); 396998781965SEric Dumazet 39701da177e4SLinus Torvalds /* Send it off, this clears delayed acks for us. */ 39712987babbSYuchung Cheng __tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0, rcv_nxt); 39721da177e4SLinus Torvalds } 397327cde44aSYuchung Cheng EXPORT_SYMBOL_GPL(__tcp_send_ack); 39742987babbSYuchung Cheng 39752987babbSYuchung Cheng void tcp_send_ack(struct sock *sk) 39762987babbSYuchung Cheng { 39772987babbSYuchung Cheng __tcp_send_ack(sk, tcp_sk(sk)->rcv_nxt); 39781da177e4SLinus Torvalds } 39791da177e4SLinus Torvalds 39801da177e4SLinus Torvalds /* This routine sends a packet with an out of date sequence 39811da177e4SLinus Torvalds * number. It assumes the other end will try to ack it. 39821da177e4SLinus Torvalds * 39831da177e4SLinus Torvalds * Question: what should we make while urgent mode? 39841da177e4SLinus Torvalds * 4.4BSD forces sending single byte of data. We cannot send 39851da177e4SLinus Torvalds * out of window data, because we have SND.NXT==SND.MAX... 39861da177e4SLinus Torvalds * 39871da177e4SLinus Torvalds * Current solution: to send TWO zero-length segments in urgent mode: 39881da177e4SLinus Torvalds * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 39891da177e4SLinus Torvalds * out-of-date with SND.UNA-1 to probe window. 39901da177e4SLinus Torvalds */ 3991e520af48SEric Dumazet static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib) 39921da177e4SLinus Torvalds { 39931da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 39941da177e4SLinus Torvalds struct sk_buff *skb; 39951da177e4SLinus Torvalds 39961da177e4SLinus Torvalds /* We don't queue it, tcp_transmit_skb() sets ownership. */ 39977450aaf6SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, 39987450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 399951456b29SIan Morris if (!skb) 40001da177e4SLinus Torvalds return -1; 40011da177e4SLinus Torvalds 40021da177e4SLinus Torvalds /* Reserve space for headers and set control bits. */ 40031da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 40041da177e4SLinus Torvalds /* Use a previous sequence. This should cause the other 40051da177e4SLinus Torvalds * end to send an ack. Don't queue or clone SKB, just 40061da177e4SLinus Torvalds * send it. 40071da177e4SLinus Torvalds */ 4008a3433f35SChangli Gao tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); 4009e2e8009fSRenato Westphal NET_INC_STATS(sock_net(sk), mib); 40107450aaf6SEric Dumazet return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0); 40111da177e4SLinus Torvalds } 40121da177e4SLinus Torvalds 4013385e2070SEric Dumazet /* Called from setsockopt( ... TCP_REPAIR ) */ 4014ee995283SPavel Emelyanov void tcp_send_window_probe(struct sock *sk) 4015ee995283SPavel Emelyanov { 4016ee995283SPavel Emelyanov if (sk->sk_state == TCP_ESTABLISHED) { 4017ee995283SPavel Emelyanov tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; 40189a568de4SEric Dumazet tcp_mstamp_refresh(tcp_sk(sk)); 4019e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE); 4020ee995283SPavel Emelyanov } 4021ee995283SPavel Emelyanov } 4022ee995283SPavel Emelyanov 402367edfef7SAndi Kleen /* Initiate keepalive or window probe from timer. */ 4024e520af48SEric Dumazet int tcp_write_wakeup(struct sock *sk, int mib) 40251da177e4SLinus Torvalds { 40261da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 40271da177e4SLinus Torvalds struct sk_buff *skb; 40281da177e4SLinus Torvalds 4029058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 4030058dc334SIlpo Järvinen return -1; 4031058dc334SIlpo Järvinen 403200db4124SIan Morris skb = tcp_send_head(sk); 403300db4124SIan Morris if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { 40341da177e4SLinus Torvalds int err; 40350c54b85fSIlpo Järvinen unsigned int mss = tcp_current_mss(sk); 403690840defSIlpo Järvinen unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 40371da177e4SLinus Torvalds 40381da177e4SLinus Torvalds if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 40391da177e4SLinus Torvalds tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 40401da177e4SLinus Torvalds 40411da177e4SLinus Torvalds /* We are probing the opening of a window 40421da177e4SLinus Torvalds * but the window size is != 0 40431da177e4SLinus Torvalds * must have been a result SWS avoidance ( sender ) 40441da177e4SLinus Torvalds */ 40451da177e4SLinus Torvalds if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 40461da177e4SLinus Torvalds skb->len > mss) { 40471da177e4SLinus Torvalds seg_size = min(seg_size, mss); 40484de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 404975c119afSEric Dumazet if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE, 405075c119afSEric Dumazet skb, seg_size, mss, GFP_ATOMIC)) 40511da177e4SLinus Torvalds return -1; 40521da177e4SLinus Torvalds } else if (!tcp_skb_pcount(skb)) 40535bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss); 40541da177e4SLinus Torvalds 40554de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 4056dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 405766f5fe62SIlpo Järvinen if (!err) 405866f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 40591da177e4SLinus Torvalds return err; 40601da177e4SLinus Torvalds } else { 406133f5f57eSIlpo Järvinen if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) 4062e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 1, mib); 4063e520af48SEric Dumazet return tcp_xmit_probe_skb(sk, 0, mib); 40641da177e4SLinus Torvalds } 40651da177e4SLinus Torvalds } 40661da177e4SLinus Torvalds 40671da177e4SLinus Torvalds /* A window probe timeout has occurred. If window is not closed send 40681da177e4SLinus Torvalds * a partial packet else a zero probe. 40691da177e4SLinus Torvalds */ 40701da177e4SLinus Torvalds void tcp_send_probe0(struct sock *sk) 40711da177e4SLinus Torvalds { 4072463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 40731da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 4074c6214a97SNikolay Borisov struct net *net = sock_net(sk); 4075c1d5674fSYuchung Cheng unsigned long timeout; 40761da177e4SLinus Torvalds int err; 40771da177e4SLinus Torvalds 4078e520af48SEric Dumazet err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE); 40791da177e4SLinus Torvalds 408075c119afSEric Dumazet if (tp->packets_out || tcp_write_queue_empty(sk)) { 40811da177e4SLinus Torvalds /* Cancel probe timer, if it is not required. */ 40826687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 0; 4083463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff = 0; 40849d9b1ee0SEnke Chen icsk->icsk_probes_tstamp = 0; 40851da177e4SLinus Torvalds return; 40861da177e4SLinus Torvalds } 40871da177e4SLinus Torvalds 4088c1d5674fSYuchung Cheng icsk->icsk_probes_out++; 40891da177e4SLinus Torvalds if (err <= 0) { 409039e24435SKuniyuki Iwashima if (icsk->icsk_backoff < READ_ONCE(net->ipv4.sysctl_tcp_retries2)) 4091463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff++; 4092c1d5674fSYuchung Cheng timeout = tcp_probe0_when(sk, TCP_RTO_MAX); 40931da177e4SLinus Torvalds } else { 40941da177e4SLinus Torvalds /* If packet was not sent due to local congestion, 4095c1d5674fSYuchung Cheng * Let senders fight for local resources conservatively. 40961da177e4SLinus Torvalds */ 4097c1d5674fSYuchung Cheng timeout = TCP_RESOURCE_PROBE_INTERVAL; 40981da177e4SLinus Torvalds } 4099344db93aSEnke Chen 4100344db93aSEnke Chen timeout = tcp_clamp_probe0_to_user_timeout(sk, timeout); 41018dc242adSEric Dumazet tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, timeout, TCP_RTO_MAX); 41021da177e4SLinus Torvalds } 41035db92c99SOctavian Purdila 4104ea3bea3aSEric Dumazet int tcp_rtx_synack(const struct sock *sk, struct request_sock *req) 41055db92c99SOctavian Purdila { 41065db92c99SOctavian Purdila const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific; 41075db92c99SOctavian Purdila struct flowi fl; 41085db92c99SOctavian Purdila int res; 41095db92c99SOctavian Purdila 4110cb6cd2ceSAkhmat Karakotov /* Paired with WRITE_ONCE() in sock_setsockopt() */ 4111cb6cd2ceSAkhmat Karakotov if (READ_ONCE(sk->sk_txrehash) == SOCK_TXREHASH_ENABLED) 411258d607d3SEric Dumazet tcp_rsk(req)->txhash = net_tx_rndhash(); 4113331fca43SMartin KaFai Lau res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL, 4114331fca43SMartin KaFai Lau NULL); 41155db92c99SOctavian Purdila if (!res) { 41160a375c82SEric Dumazet TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); 41170a375c82SEric Dumazet NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 41187e32b443SYuchung Cheng if (unlikely(tcp_passive_fastopen(sk))) 41197e32b443SYuchung Cheng tcp_sk(sk)->total_retrans++; 4120cf34ce3dSSong Liu trace_tcp_retransmit_synack(sk, req); 41215db92c99SOctavian Purdila } 41225db92c99SOctavian Purdila return res; 41235db92c99SOctavian Purdila } 41245db92c99SOctavian Purdila EXPORT_SYMBOL(tcp_rtx_synack); 4125