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 */ 24102739545SKuniyuki Iwashima space = max_t(u32, space, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2])); 2421227c177SKuniyuki Iwashima space = max_t(u32, space, READ_ONCE(sysctl_rmem_max)); 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 769459837b5SDmitry Safonov if (static_branch_unlikely(&tcp_md5_needed.key) && 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 925459837b5SDmitry Safonov if (static_branch_unlikely(&tcp_md5_needed.key) && 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 { 1080fac30731SEric Dumazet unsigned long flags = smp_load_acquire(&sk->sk_tsq_flags); 1081fac30731SEric Dumazet unsigned long nflags; 108246d3ceabSEric Dumazet 10836f458dfbSEric Dumazet /* perform an atomic operation only if at least one flag is set */ 10846f458dfbSEric Dumazet do { 10856f458dfbSEric Dumazet if (!(flags & TCP_DEFERRED_ALL)) 10866f458dfbSEric Dumazet return; 10876f458dfbSEric Dumazet nflags = flags & ~TCP_DEFERRED_ALL; 1088fac30731SEric Dumazet } while (!try_cmpxchg(&sk->sk_tsq_flags, &flags, nflags)); 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; 1142b548b17aSEric Dumazet struct tsq_tasklet *tsq; 1143b548b17aSEric Dumazet bool empty; 11449b462d02SEric Dumazet 11459b462d02SEric Dumazet /* Keep one reference on sk_wmem_alloc. 11469b462d02SEric Dumazet * Will be released by sk_free() from here or tcp_tasklet_func() 11479b462d02SEric Dumazet */ 114814afee4bSReshetova, Elena WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc)); 11499b462d02SEric Dumazet 11509b462d02SEric Dumazet /* If this softirq is serviced by ksoftirqd, we are likely under stress. 11519b462d02SEric Dumazet * Wait until our queues (qdisc + devices) are drained. 11529b462d02SEric Dumazet * This gives : 11539b462d02SEric Dumazet * - less callbacks to tcp_write_xmit(), reducing stress (batches) 11549b462d02SEric Dumazet * - chance for incoming ACK (processed by another cpu maybe) 11559b462d02SEric Dumazet * to migrate this flow (skb->ooo_okay will be eventually set) 11569b462d02SEric Dumazet */ 115714afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current) 11589b462d02SEric Dumazet goto out; 115946d3ceabSEric Dumazet 1160b548b17aSEric Dumazet oval = smp_load_acquire(&sk->sk_tsq_flags); 1161b548b17aSEric Dumazet do { 1162408f0a6cSEric Dumazet if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED)) 1163408f0a6cSEric Dumazet goto out; 1164408f0a6cSEric Dumazet 116573a6bab5SEric Dumazet nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED; 1166b548b17aSEric Dumazet } while (!try_cmpxchg(&sk->sk_tsq_flags, &oval, nval)); 1167408f0a6cSEric Dumazet 116846d3ceabSEric Dumazet /* queue this socket to tasklet queue */ 116946d3ceabSEric Dumazet local_irq_save(flags); 1170903ceff7SChristoph Lameter tsq = this_cpu_ptr(&tsq_tasklet); 1171a9b204d1SEric Dumazet empty = list_empty(&tsq->head); 117246d3ceabSEric Dumazet list_add(&tp->tsq_node, &tsq->head); 1173a9b204d1SEric Dumazet if (empty) 117446d3ceabSEric Dumazet tasklet_schedule(&tsq->tasklet); 117546d3ceabSEric Dumazet local_irq_restore(flags); 11769b462d02SEric Dumazet return; 11779b462d02SEric Dumazet out: 11789b462d02SEric Dumazet sk_free(sk); 117946d3ceabSEric Dumazet } 118046d3ceabSEric Dumazet 118173a6bab5SEric Dumazet /* Note: Called under soft irq. 118273a6bab5SEric Dumazet * We can call TCP stack right away, unless socket is owned by user. 1183218af599SEric Dumazet */ 1184218af599SEric Dumazet enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer) 1185218af599SEric Dumazet { 1186218af599SEric Dumazet struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer); 1187218af599SEric Dumazet struct sock *sk = (struct sock *)tp; 1188218af599SEric Dumazet 118973a6bab5SEric Dumazet tcp_tsq_handler(sk); 119073a6bab5SEric Dumazet sock_put(sk); 1191218af599SEric Dumazet 1192218af599SEric Dumazet return HRTIMER_NORESTART; 1193218af599SEric Dumazet } 1194218af599SEric Dumazet 1195a7a25630SEric Dumazet static void tcp_update_skb_after_send(struct sock *sk, struct sk_buff *skb, 1196a7a25630SEric Dumazet u64 prior_wstamp) 1197e2080072SEric Dumazet { 1198ab408b6dSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 1199ab408b6dSEric Dumazet 1200ab408b6dSEric Dumazet if (sk->sk_pacing_status != SK_PACING_NONE) { 120176a9ebe8SEric Dumazet unsigned long rate = sk->sk_pacing_rate; 1202ab408b6dSEric Dumazet 1203ab408b6dSEric Dumazet /* Original sch_fq does not pace first 10 MSS 1204ab408b6dSEric Dumazet * Note that tp->data_segs_out overflows after 2^32 packets, 1205ab408b6dSEric Dumazet * this is a minor annoyance. 1206ab408b6dSEric Dumazet */ 120776a9ebe8SEric Dumazet if (rate != ~0UL && rate && tp->data_segs_out >= 10) { 1208a7a25630SEric Dumazet u64 len_ns = div64_ul((u64)skb->len * NSEC_PER_SEC, rate); 1209a7a25630SEric Dumazet u64 credit = tp->tcp_wstamp_ns - prior_wstamp; 1210a7a25630SEric Dumazet 1211a7a25630SEric Dumazet /* take into account OS jitter */ 1212a7a25630SEric Dumazet len_ns -= min_t(u64, len_ns / 2, credit); 1213a7a25630SEric Dumazet tp->tcp_wstamp_ns += len_ns; 1214ab408b6dSEric Dumazet } 1215ab408b6dSEric Dumazet } 1216e2080072SEric Dumazet list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue); 1217e2080072SEric Dumazet } 1218e2080072SEric Dumazet 121905e22e83SEric Dumazet INDIRECT_CALLABLE_DECLARE(int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl)); 122005e22e83SEric Dumazet INDIRECT_CALLABLE_DECLARE(int inet6_csk_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl)); 1221dd2e0b86SEric Dumazet INDIRECT_CALLABLE_DECLARE(void tcp_v4_send_check(struct sock *sk, struct sk_buff *skb)); 122205e22e83SEric Dumazet 12231da177e4SLinus Torvalds /* This routine actually transmits TCP packets queued in by 12241da177e4SLinus Torvalds * tcp_do_sendmsg(). This is used by both the initial 12251da177e4SLinus Torvalds * transmission and possible later retransmissions. 12261da177e4SLinus Torvalds * All SKB's seen here are completely headerless. It is our 12271da177e4SLinus Torvalds * job to build the TCP header, and pass the packet down to 12281da177e4SLinus Torvalds * IP so it can do the same plus pass the packet off to the 12291da177e4SLinus Torvalds * device. 12301da177e4SLinus Torvalds * 12311da177e4SLinus Torvalds * We are working here with either a clone of the original 12321da177e4SLinus Torvalds * SKB, or a fresh unique copy made by the retransmit engine. 12331da177e4SLinus Torvalds */ 12342987babbSYuchung Cheng static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, 12352987babbSYuchung Cheng int clone_it, gfp_t gfp_mask, u32 rcv_nxt) 12361da177e4SLinus Torvalds { 12376687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 1238dfb4b9dcSDavid S. Miller struct inet_sock *inet; 1239dfb4b9dcSDavid S. Miller struct tcp_sock *tp; 1240dfb4b9dcSDavid S. Miller struct tcp_skb_cb *tcb; 124133ad798cSAdam Langley struct tcp_out_options opts; 124295c96174SEric Dumazet unsigned int tcp_options_size, tcp_header_size; 12438c72c65bSEric Dumazet struct sk_buff *oskb = NULL; 1244cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 12451da177e4SLinus Torvalds struct tcphdr *th; 1246a7a25630SEric Dumazet u64 prior_wstamp; 12471da177e4SLinus Torvalds int err; 12481da177e4SLinus Torvalds 1249dfb4b9dcSDavid S. Miller BUG_ON(!skb || !tcp_skb_pcount(skb)); 12506f094b9eSLawrence Brakmo tp = tcp_sk(sk); 12517f12422cSYuchung Cheng prior_wstamp = tp->tcp_wstamp_ns; 12527f12422cSYuchung Cheng tp->tcp_wstamp_ns = max(tp->tcp_wstamp_ns, tp->tcp_clock_cache); 1253a1ac9c8aSMartin KaFai Lau skb_set_delivery_time(skb, tp->tcp_wstamp_ns, true); 1254ccdbb6e9SEric Dumazet if (clone_it) { 12558c72c65bSEric Dumazet oskb = skb; 1256e2080072SEric Dumazet 1257e2080072SEric Dumazet tcp_skb_tsorted_save(oskb) { 1258e2080072SEric Dumazet if (unlikely(skb_cloned(oskb))) 1259e2080072SEric Dumazet skb = pskb_copy(oskb, gfp_mask); 1260dfb4b9dcSDavid S. Miller else 1261e2080072SEric Dumazet skb = skb_clone(oskb, gfp_mask); 1262e2080072SEric Dumazet } tcp_skb_tsorted_restore(oskb); 1263e2080072SEric Dumazet 1264dfb4b9dcSDavid S. Miller if (unlikely(!skb)) 1265dfb4b9dcSDavid S. Miller return -ENOBUFS; 1266b738a185SEric Dumazet /* retransmit skbs might have a non zero value in skb->dev 1267b738a185SEric Dumazet * because skb->dev is aliased with skb->rbnode.rb_left 1268b738a185SEric Dumazet */ 1269b738a185SEric Dumazet skb->dev = NULL; 1270dfb4b9dcSDavid S. Miller } 12715f6188a8SEric Dumazet 1272dfb4b9dcSDavid S. Miller inet = inet_sk(sk); 1273dfb4b9dcSDavid S. Miller tcb = TCP_SKB_CB(skb); 127433ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 12751da177e4SLinus Torvalds 1276051ba674SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) { 127733ad798cSAdam Langley tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); 1278051ba674SEric Dumazet } else { 127933ad798cSAdam Langley tcp_options_size = tcp_established_options(sk, skb, &opts, 128033ad798cSAdam Langley &md5); 1281051ba674SEric Dumazet /* Force a PSH flag on all (GSO) packets to expedite GRO flush 1282051ba674SEric Dumazet * at receiver : This slightly improve GRO performance. 1283051ba674SEric Dumazet * Note that we do not force the PSH flag for non GSO packets, 1284051ba674SEric Dumazet * because they might be sent under high congestion events, 1285051ba674SEric Dumazet * and in this case it is better to delay the delivery of 1-MSS 1286051ba674SEric Dumazet * packets and thus the corresponding ACK packet that would 1287051ba674SEric Dumazet * release the following packet. 1288051ba674SEric Dumazet */ 1289051ba674SEric Dumazet if (tcp_skb_pcount(skb) > 1) 1290051ba674SEric Dumazet tcb->tcp_flags |= TCPHDR_PSH; 1291051ba674SEric Dumazet } 129233ad798cSAdam Langley tcp_header_size = tcp_options_size + sizeof(struct tcphdr); 12931da177e4SLinus Torvalds 1294547669d4SEric Dumazet /* if no packet is in qdisc/device queue, then allow XPS to select 1295b2532eb9SEric Dumazet * another queue. We can be called from tcp_tsq_handler() 129673a6bab5SEric Dumazet * which holds one reference to sk. 1297b2532eb9SEric Dumazet * 1298b2532eb9SEric Dumazet * TODO: Ideally, in-flight pure ACK packets should not matter here. 1299b2532eb9SEric Dumazet * One way to get this would be to set skb->truesize = 2 on them. 1300547669d4SEric Dumazet */ 1301b2532eb9SEric Dumazet skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1); 13021da177e4SLinus Torvalds 130338ab52e8SEric Dumazet /* If we had to use memory reserve to allocate this skb, 130438ab52e8SEric Dumazet * this might cause drops if packet is looped back : 130538ab52e8SEric Dumazet * Other socket might not have SOCK_MEMALLOC. 130638ab52e8SEric Dumazet * Packets not looped back do not care about pfmemalloc. 130738ab52e8SEric Dumazet */ 130838ab52e8SEric Dumazet skb->pfmemalloc = 0; 130938ab52e8SEric Dumazet 1310aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 1311aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 131246d3ceabSEric Dumazet 131346d3ceabSEric Dumazet skb_orphan(skb); 131446d3ceabSEric Dumazet skb->sk = sk; 13151d2077acSEric Dumazet skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree; 131614afee4bSReshetova, Elena refcount_add(skb->truesize, &sk->sk_wmem_alloc); 13171da177e4SLinus Torvalds 1318c3a2e837SJulian Anastasov skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm); 1319c3a2e837SJulian Anastasov 13201da177e4SLinus Torvalds /* Build TCP header and checksum it. */ 1321ea1627c2SEric Dumazet th = (struct tcphdr *)skb->data; 1322c720c7e8SEric Dumazet th->source = inet->inet_sport; 1323c720c7e8SEric Dumazet th->dest = inet->inet_dport; 13241da177e4SLinus Torvalds th->seq = htonl(tcb->seq); 13252987babbSYuchung Cheng th->ack_seq = htonl(rcv_nxt); 1326df7a3b07SAl Viro *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | 13274de075e0SEric Dumazet tcb->tcp_flags); 1328dfb4b9dcSDavid S. Miller 13291da177e4SLinus Torvalds th->check = 0; 13301da177e4SLinus Torvalds th->urg_ptr = 0; 13311da177e4SLinus Torvalds 133233f5f57eSIlpo Järvinen /* The urg_mode check is necessary during a below snd_una win probe */ 13337691367dSHerbert Xu if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { 13347691367dSHerbert Xu if (before(tp->snd_up, tcb->seq + 0x10000)) { 13351da177e4SLinus Torvalds th->urg_ptr = htons(tp->snd_up - tcb->seq); 13361da177e4SLinus Torvalds th->urg = 1; 13377691367dSHerbert Xu } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { 13380eae88f3SEric Dumazet th->urg_ptr = htons(0xFFFF); 13397691367dSHerbert Xu th->urg = 1; 13407691367dSHerbert Xu } 13411da177e4SLinus Torvalds } 13421da177e4SLinus Torvalds 134351466a75SEric Dumazet skb_shinfo(skb)->gso_type = sk->sk_gso_type; 1344ea1627c2SEric Dumazet if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) { 1345ea1627c2SEric Dumazet th->window = htons(tcp_select_window(sk)); 1346ea1627c2SEric Dumazet tcp_ecn_send(sk, skb, th, tcp_header_size); 1347ea1627c2SEric Dumazet } else { 1348ea1627c2SEric Dumazet /* RFC1323: The window in SYN & SYN/ACK segments 1349ea1627c2SEric Dumazet * is never scaled. 1350ea1627c2SEric Dumazet */ 1351ea1627c2SEric Dumazet th->window = htons(min(tp->rcv_wnd, 65535U)); 1352ea1627c2SEric Dumazet } 1353fa3fe2b1SFlorian Westphal 1354ea66758cSPaolo Abeni tcp_options_write(th, tp, &opts); 1355fa3fe2b1SFlorian Westphal 1356cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 1357cfb6eeb4SYOSHIFUJI Hideaki /* Calculate the MD5 hash, as we have all we need now */ 1358cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 1359aba54656SEric Dumazet sk_gso_disable(sk); 1360bd0388aeSWilliam Allen Simpson tp->af_specific->calc_md5_hash(opts.hash_location, 136139f8e58eSEric Dumazet md5, sk, skb); 1362cfb6eeb4SYOSHIFUJI Hideaki } 1363cfb6eeb4SYOSHIFUJI Hideaki #endif 1364cfb6eeb4SYOSHIFUJI Hideaki 1365331fca43SMartin KaFai Lau /* BPF prog is the last one writing header option */ 1366331fca43SMartin KaFai Lau bpf_skops_write_hdr_opt(sk, skb, NULL, NULL, 0, &opts); 1367331fca43SMartin KaFai Lau 1368dd2e0b86SEric Dumazet INDIRECT_CALL_INET(icsk->icsk_af_ops->send_check, 1369dd2e0b86SEric Dumazet tcp_v6_send_check, tcp_v4_send_check, 1370dd2e0b86SEric Dumazet sk, skb); 13711da177e4SLinus Torvalds 13724de075e0SEric Dumazet if (likely(tcb->tcp_flags & TCPHDR_ACK)) 137327cde44aSYuchung Cheng tcp_event_ack_sent(sk, tcp_skb_pcount(skb), rcv_nxt); 13741da177e4SLinus Torvalds 1375a44d6eacSMartin KaFai Lau if (skb->len != tcp_header_size) { 1376cf533ea5SEric Dumazet tcp_event_data_sent(tp, sk); 1377a44d6eacSMartin KaFai Lau tp->data_segs_out += tcp_skb_pcount(skb); 1378ba113c3aSWei Wang tp->bytes_sent += skb->len - tcp_header_size; 1379a44d6eacSMartin KaFai Lau } 13801da177e4SLinus Torvalds 1381bd37a088SWei Yongjun if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) 1382aa2ea058STom Herbert TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, 1383aa2ea058STom Herbert tcp_skb_pcount(skb)); 13841da177e4SLinus Torvalds 13852efd055cSMarcelo Ricardo Leitner tp->segs_out += tcp_skb_pcount(skb); 13860ae5b43dSYuchung Cheng skb_set_hash_from_sk(skb, sk); 1387f69ad292SEric Dumazet /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */ 1388cd7d8498SEric Dumazet skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb); 1389f69ad292SEric Dumazet skb_shinfo(skb)->gso_size = tcp_skb_mss(skb); 1390cd7d8498SEric Dumazet 1391d3edd06eSEric Dumazet /* Leave earliest departure time in skb->tstamp (skb->skb_mstamp_ns) */ 1392971f10ecSEric Dumazet 1393971f10ecSEric Dumazet /* Cleanup our debris for IP stacks */ 1394971f10ecSEric Dumazet memset(skb->cb, 0, max(sizeof(struct inet_skb_parm), 1395971f10ecSEric Dumazet sizeof(struct inet6_skb_parm))); 1396971f10ecSEric Dumazet 1397a842fe14SEric Dumazet tcp_add_tx_delay(skb, tp); 1398a842fe14SEric Dumazet 139905e22e83SEric Dumazet err = INDIRECT_CALL_INET(icsk->icsk_af_ops->queue_xmit, 140005e22e83SEric Dumazet inet6_csk_xmit, ip_queue_xmit, 140105e22e83SEric Dumazet sk, skb, &inet->cork.fl); 14027faee5c0SEric Dumazet 14038c72c65bSEric Dumazet if (unlikely(err > 0)) { 14045ee2c941SChristoph Paasch tcp_enter_cwr(sk); 14058c72c65bSEric Dumazet err = net_xmit_eval(err); 14068c72c65bSEric Dumazet } 1407fc225799SEric Dumazet if (!err && oskb) { 1408a7a25630SEric Dumazet tcp_update_skb_after_send(sk, oskb, prior_wstamp); 1409fc225799SEric Dumazet tcp_rate_skb_sent(sk, oskb); 1410fc225799SEric Dumazet } 14118c72c65bSEric Dumazet return err; 14121da177e4SLinus Torvalds } 14131da177e4SLinus Torvalds 14142987babbSYuchung Cheng static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, 14152987babbSYuchung Cheng gfp_t gfp_mask) 14162987babbSYuchung Cheng { 14172987babbSYuchung Cheng return __tcp_transmit_skb(sk, skb, clone_it, gfp_mask, 14182987babbSYuchung Cheng tcp_sk(sk)->rcv_nxt); 14192987babbSYuchung Cheng } 14202987babbSYuchung Cheng 142167edfef7SAndi Kleen /* This routine just queues the buffer for sending. 14221da177e4SLinus Torvalds * 14231da177e4SLinus Torvalds * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, 14241da177e4SLinus Torvalds * otherwise socket can stall. 14251da177e4SLinus Torvalds */ 14261da177e4SLinus Torvalds static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) 14271da177e4SLinus Torvalds { 14281da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 14291da177e4SLinus Torvalds 14301da177e4SLinus Torvalds /* Advance write_seq and place onto the write_queue. */ 14310f317464SEric Dumazet WRITE_ONCE(tp->write_seq, TCP_SKB_CB(skb)->end_seq); 1432f4a775d1SEric Dumazet __skb_header_release(skb); 1433fe067e8aSDavid S. Miller tcp_add_write_queue_tail(sk, skb); 1434ab4e846aSEric Dumazet sk_wmem_queued_add(sk, skb->truesize); 14353ab224beSHideo Aoki sk_mem_charge(sk, skb->truesize); 14361da177e4SLinus Torvalds } 14371da177e4SLinus Torvalds 143867edfef7SAndi Kleen /* Initialize TSO segments for a packet. */ 14395bbb432cSEric Dumazet static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now) 1440f6302d1dSDavid S. Miller { 14414a64fd6cSEric Dumazet if (skb->len <= mss_now) { 1442f6302d1dSDavid S. Miller /* Avoid the costly divide in the normal 1443f6302d1dSDavid S. Miller * non-TSO case. 1444f6302d1dSDavid S. Miller */ 1445cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, 1); 1446f69ad292SEric Dumazet TCP_SKB_CB(skb)->tcp_gso_size = 0; 1447f6302d1dSDavid S. Miller } else { 1448cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now)); 1449f69ad292SEric Dumazet TCP_SKB_CB(skb)->tcp_gso_size = mss_now; 14501da177e4SLinus Torvalds } 14511da177e4SLinus Torvalds } 14521da177e4SLinus Torvalds 1453797108d1SIlpo Järvinen /* Pcount in the middle of the write queue got changed, we need to do various 1454797108d1SIlpo Järvinen * tweaks to fix counters 1455797108d1SIlpo Järvinen */ 1456cf533ea5SEric Dumazet static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) 1457797108d1SIlpo Järvinen { 1458797108d1SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1459797108d1SIlpo Järvinen 1460797108d1SIlpo Järvinen tp->packets_out -= decr; 1461797108d1SIlpo Järvinen 1462797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 1463797108d1SIlpo Järvinen tp->sacked_out -= decr; 1464797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 1465797108d1SIlpo Järvinen tp->retrans_out -= decr; 1466797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 1467797108d1SIlpo Järvinen tp->lost_out -= decr; 1468797108d1SIlpo Järvinen 1469797108d1SIlpo Järvinen /* Reno case is special. Sigh... */ 1470797108d1SIlpo Järvinen if (tcp_is_reno(tp) && decr > 0) 1471797108d1SIlpo Järvinen tp->sacked_out -= min_t(u32, tp->sacked_out, decr); 1472797108d1SIlpo Järvinen 1473797108d1SIlpo Järvinen if (tp->lost_skb_hint && 1474797108d1SIlpo Järvinen before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && 1475713bafeaSYuchung Cheng (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) 1476797108d1SIlpo Järvinen tp->lost_cnt_hint -= decr; 1477797108d1SIlpo Järvinen 1478797108d1SIlpo Järvinen tcp_verify_left_out(tp); 1479797108d1SIlpo Järvinen } 1480797108d1SIlpo Järvinen 14810a2cf20cSSoheil Hassas Yeganeh static bool tcp_has_tx_tstamp(const struct sk_buff *skb) 14820a2cf20cSSoheil Hassas Yeganeh { 14830a2cf20cSSoheil Hassas Yeganeh return TCP_SKB_CB(skb)->txstamp_ack || 14840a2cf20cSSoheil Hassas Yeganeh (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP); 14850a2cf20cSSoheil Hassas Yeganeh } 14860a2cf20cSSoheil Hassas Yeganeh 1487490cc7d0SWillem de Bruijn static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2) 1488490cc7d0SWillem de Bruijn { 1489490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo = skb_shinfo(skb); 1490490cc7d0SWillem de Bruijn 14910a2cf20cSSoheil Hassas Yeganeh if (unlikely(tcp_has_tx_tstamp(skb)) && 1492490cc7d0SWillem de Bruijn !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) { 1493490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo2 = skb_shinfo(skb2); 1494490cc7d0SWillem de Bruijn u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP; 1495490cc7d0SWillem de Bruijn 1496490cc7d0SWillem de Bruijn shinfo->tx_flags &= ~tsflags; 1497490cc7d0SWillem de Bruijn shinfo2->tx_flags |= tsflags; 1498490cc7d0SWillem de Bruijn swap(shinfo->tskey, shinfo2->tskey); 1499b51e13faSMartin KaFai Lau TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack; 1500b51e13faSMartin KaFai Lau TCP_SKB_CB(skb)->txstamp_ack = 0; 1501490cc7d0SWillem de Bruijn } 1502490cc7d0SWillem de Bruijn } 1503490cc7d0SWillem de Bruijn 1504a166140eSMartin KaFai Lau static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2) 1505a166140eSMartin KaFai Lau { 1506a166140eSMartin KaFai Lau TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor; 1507a166140eSMartin KaFai Lau TCP_SKB_CB(skb)->eor = 0; 1508a166140eSMartin KaFai Lau } 1509a166140eSMartin KaFai Lau 151075c119afSEric Dumazet /* Insert buff after skb on the write or rtx queue of sk. */ 151175c119afSEric Dumazet static void tcp_insert_write_queue_after(struct sk_buff *skb, 151275c119afSEric Dumazet struct sk_buff *buff, 151375c119afSEric Dumazet struct sock *sk, 151475c119afSEric Dumazet enum tcp_queue tcp_queue) 151575c119afSEric Dumazet { 151675c119afSEric Dumazet if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE) 151775c119afSEric Dumazet __skb_queue_after(&sk->sk_write_queue, skb, buff); 151875c119afSEric Dumazet else 151975c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, buff); 152075c119afSEric Dumazet } 152175c119afSEric Dumazet 15221da177e4SLinus Torvalds /* Function to create two new TCP segments. Shrinks the given segment 15231da177e4SLinus Torvalds * to the specified size and appends a new segment with the rest of the 15241da177e4SLinus Torvalds * packet to the list. This won't be called frequently, I hope. 15251da177e4SLinus Torvalds * Remember, these are still headerless SKBs at this point. 15261da177e4SLinus Torvalds */ 152775c119afSEric Dumazet int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue, 152875c119afSEric Dumazet struct sk_buff *skb, u32 len, 15296cc55e09SOctavian Purdila unsigned int mss_now, gfp_t gfp) 15301da177e4SLinus Torvalds { 15311da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 15321da177e4SLinus Torvalds struct sk_buff *buff; 15336475be16SDavid S. Miller int nsize, old_factor; 1534b617158dSEric Dumazet long limit; 1535b60b49eaSHerbert Xu int nlen; 15369ce01461SIlpo Järvinen u8 flags; 15371da177e4SLinus Torvalds 15382fceec13SIlpo Järvinen if (WARN_ON(len > skb->len)) 15392fceec13SIlpo Järvinen return -EINVAL; 15406a438bbeSStephen Hemminger 15411da177e4SLinus Torvalds nsize = skb_headlen(skb) - len; 15421da177e4SLinus Torvalds if (nsize < 0) 15431da177e4SLinus Torvalds nsize = 0; 15441da177e4SLinus Torvalds 1545b617158dSEric Dumazet /* tcp_sendmsg() can overshoot sk_wmem_queued by one full size skb. 1546b617158dSEric Dumazet * We need some allowance to not penalize applications setting small 1547b617158dSEric Dumazet * SO_SNDBUF values. 1548b617158dSEric Dumazet * Also allow first and last skb in retransmit queue to be split. 1549b617158dSEric Dumazet */ 15507c4e983cSAlexander Duyck limit = sk->sk_sndbuf + 2 * SKB_TRUESIZE(GSO_LEGACY_MAX_SIZE); 1551b617158dSEric Dumazet if (unlikely((sk->sk_wmem_queued >> 1) > limit && 1552b617158dSEric Dumazet tcp_queue != TCP_FRAG_IN_WRITE_QUEUE && 1553b617158dSEric Dumazet skb != tcp_rtx_queue_head(sk) && 1554b617158dSEric Dumazet skb != tcp_rtx_queue_tail(sk))) { 1555f070ef2aSEric Dumazet NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPWQUEUETOOBIG); 1556f070ef2aSEric Dumazet return -ENOMEM; 1557f070ef2aSEric Dumazet } 1558f070ef2aSEric Dumazet 1559c4777efaSEric Dumazet if (skb_unclone_keeptruesize(skb, gfp)) 15601da177e4SLinus Torvalds return -ENOMEM; 15611da177e4SLinus Torvalds 15621da177e4SLinus Torvalds /* Get a new skb... force flag on. */ 1563f8dd3b8dSEric Dumazet buff = tcp_stream_alloc_skb(sk, nsize, gfp, true); 156451456b29SIan Morris if (!buff) 15651da177e4SLinus Torvalds return -ENOMEM; /* We'll just try again later. */ 156641477662SJakub Kicinski skb_copy_decrypted(buff, skb); 15675a369ca6SPaolo Abeni mptcp_skb_ext_copy(buff, skb); 1568ef5cb973SHerbert Xu 1569ab4e846aSEric Dumazet sk_wmem_queued_add(sk, buff->truesize); 15703ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1571b60b49eaSHerbert Xu nlen = skb->len - len - nsize; 1572b60b49eaSHerbert Xu buff->truesize += nlen; 1573b60b49eaSHerbert Xu skb->truesize -= nlen; 15741da177e4SLinus Torvalds 15751da177e4SLinus Torvalds /* Correct the sequence numbers. */ 15761da177e4SLinus Torvalds TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 15771da177e4SLinus Torvalds TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 15781da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 15791da177e4SLinus Torvalds 15801da177e4SLinus Torvalds /* PSH and FIN should only be set in the second packet. */ 15814de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 15824de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 15834de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1584e14c3cafSHerbert Xu TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; 1585a166140eSMartin KaFai Lau tcp_skb_fragment_eor(skb, buff); 15861da177e4SLinus Torvalds 15871da177e4SLinus Torvalds skb_split(skb, buff, len); 15881da177e4SLinus Torvalds 1589a1ac9c8aSMartin KaFai Lau skb_set_delivery_time(buff, skb->tstamp, true); 1590490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 15911da177e4SLinus Torvalds 15926475be16SDavid S. Miller old_factor = tcp_skb_pcount(skb); 15936475be16SDavid S. Miller 15941da177e4SLinus Torvalds /* Fix up tso_factor for both original and new SKB. */ 15955bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 15965bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 15971da177e4SLinus Torvalds 1598b9f64820SYuchung Cheng /* Update delivered info for the new segment */ 1599b9f64820SYuchung Cheng TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx; 1600b9f64820SYuchung Cheng 16016475be16SDavid S. Miller /* If this packet has been sent out already, we must 16026475be16SDavid S. Miller * adjust the various packet counters. 16036475be16SDavid S. Miller */ 1604cf0b450cSHerbert Xu if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 16056475be16SDavid S. Miller int diff = old_factor - tcp_skb_pcount(skb) - 16066475be16SDavid S. Miller tcp_skb_pcount(buff); 16071da177e4SLinus Torvalds 1608797108d1SIlpo Järvinen if (diff) 1609797108d1SIlpo Järvinen tcp_adjust_pcount(sk, skb, diff); 16101da177e4SLinus Torvalds } 16111da177e4SLinus Torvalds 16121da177e4SLinus Torvalds /* Link BUFF into the send queue. */ 1613f4a775d1SEric Dumazet __skb_header_release(buff); 161475c119afSEric Dumazet tcp_insert_write_queue_after(skb, buff, sk, tcp_queue); 1615f67971e6SEric Dumazet if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE) 1616e2080072SEric Dumazet list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor); 16171da177e4SLinus Torvalds 16181da177e4SLinus Torvalds return 0; 16191da177e4SLinus Torvalds } 16201da177e4SLinus Torvalds 1621f4d01666SEric Dumazet /* This is similar to __pskb_pull_tail(). The difference is that pulled 1622f4d01666SEric Dumazet * data is not copied, but immediately discarded. 16231da177e4SLinus Torvalds */ 16247162fb24SEric Dumazet static int __pskb_trim_head(struct sk_buff *skb, int len) 16251da177e4SLinus Torvalds { 16267b7fc97aSEric Dumazet struct skb_shared_info *shinfo; 16271da177e4SLinus Torvalds int i, k, eat; 16281da177e4SLinus Torvalds 16294fa48bf3SEric Dumazet eat = min_t(int, len, skb_headlen(skb)); 16304fa48bf3SEric Dumazet if (eat) { 16314fa48bf3SEric Dumazet __skb_pull(skb, eat); 16324fa48bf3SEric Dumazet len -= eat; 16334fa48bf3SEric Dumazet if (!len) 16347162fb24SEric Dumazet return 0; 16354fa48bf3SEric Dumazet } 16361da177e4SLinus Torvalds eat = len; 16371da177e4SLinus Torvalds k = 0; 16387b7fc97aSEric Dumazet shinfo = skb_shinfo(skb); 16397b7fc97aSEric Dumazet for (i = 0; i < shinfo->nr_frags; i++) { 16407b7fc97aSEric Dumazet int size = skb_frag_size(&shinfo->frags[i]); 16419e903e08SEric Dumazet 16429e903e08SEric Dumazet if (size <= eat) { 1643aff65da0SIan Campbell skb_frag_unref(skb, i); 16449e903e08SEric Dumazet eat -= size; 16451da177e4SLinus Torvalds } else { 16467b7fc97aSEric Dumazet shinfo->frags[k] = shinfo->frags[i]; 16471da177e4SLinus Torvalds if (eat) { 1648b54c9d5bSJonathan Lemon skb_frag_off_add(&shinfo->frags[k], eat); 16497b7fc97aSEric Dumazet skb_frag_size_sub(&shinfo->frags[k], eat); 16501da177e4SLinus Torvalds eat = 0; 16511da177e4SLinus Torvalds } 16521da177e4SLinus Torvalds k++; 16531da177e4SLinus Torvalds } 16541da177e4SLinus Torvalds } 16557b7fc97aSEric Dumazet shinfo->nr_frags = k; 16561da177e4SLinus Torvalds 16571da177e4SLinus Torvalds skb->data_len -= len; 16581da177e4SLinus Torvalds skb->len = skb->data_len; 16597162fb24SEric Dumazet return len; 16601da177e4SLinus Torvalds } 16611da177e4SLinus Torvalds 166267edfef7SAndi Kleen /* Remove acked data from a packet in the transmit queue. */ 16631da177e4SLinus Torvalds int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 16641da177e4SLinus Torvalds { 16657162fb24SEric Dumazet u32 delta_truesize; 16667162fb24SEric Dumazet 1667c4777efaSEric Dumazet if (skb_unclone_keeptruesize(skb, GFP_ATOMIC)) 16681da177e4SLinus Torvalds return -ENOMEM; 16691da177e4SLinus Torvalds 16707162fb24SEric Dumazet delta_truesize = __pskb_trim_head(skb, len); 16711da177e4SLinus Torvalds 16721da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq += len; 16731da177e4SLinus Torvalds 16747162fb24SEric Dumazet if (delta_truesize) { 16757162fb24SEric Dumazet skb->truesize -= delta_truesize; 1676ab4e846aSEric Dumazet sk_wmem_queued_add(sk, -delta_truesize); 16779b65b17dSTalal Ahmad if (!skb_zcopy_pure(skb)) 16787162fb24SEric Dumazet sk_mem_uncharge(sk, delta_truesize); 16797162fb24SEric Dumazet } 16801da177e4SLinus Torvalds 16815b35e1e6SNeal Cardwell /* Any change of skb->len requires recalculation of tso factor. */ 16821da177e4SLinus Torvalds if (tcp_skb_pcount(skb) > 1) 16835bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb)); 16841da177e4SLinus Torvalds 16851da177e4SLinus Torvalds return 0; 16861da177e4SLinus Torvalds } 16871da177e4SLinus Torvalds 16881b63edd6SYuchung Cheng /* Calculate MSS not accounting any TCP options. */ 16891b63edd6SYuchung Cheng static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu) 16905d424d5aSJohn Heffner { 1691cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1692cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 16935d424d5aSJohn Heffner int mss_now; 16945d424d5aSJohn Heffner 16955d424d5aSJohn Heffner /* Calculate base mss without TCP options: 16965d424d5aSJohn Heffner It is MMS_S - sizeof(tcphdr) of rfc1122 16975d424d5aSJohn Heffner */ 16985d424d5aSJohn Heffner mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 16995d424d5aSJohn Heffner 170067469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 170167469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 170267469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 170367469601SEric Dumazet 170467469601SEric Dumazet if (dst && dst_allfrag(dst)) 170567469601SEric Dumazet mss_now -= icsk->icsk_af_ops->net_frag_header_len; 170667469601SEric Dumazet } 170767469601SEric Dumazet 17085d424d5aSJohn Heffner /* Clamp it (mss_clamp does not include tcp options) */ 17095d424d5aSJohn Heffner if (mss_now > tp->rx_opt.mss_clamp) 17105d424d5aSJohn Heffner mss_now = tp->rx_opt.mss_clamp; 17115d424d5aSJohn Heffner 17125d424d5aSJohn Heffner /* Now subtract optional transport overhead */ 17135d424d5aSJohn Heffner mss_now -= icsk->icsk_ext_hdr_len; 17145d424d5aSJohn Heffner 17155d424d5aSJohn Heffner /* Then reserve room for full set of TCP options and 8 bytes of data */ 171678eb166cSKuniyuki Iwashima mss_now = max(mss_now, 171778eb166cSKuniyuki Iwashima READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_snd_mss)); 17185d424d5aSJohn Heffner return mss_now; 17195d424d5aSJohn Heffner } 17205d424d5aSJohn Heffner 17211b63edd6SYuchung Cheng /* Calculate MSS. Not accounting for SACKs here. */ 17221b63edd6SYuchung Cheng int tcp_mtu_to_mss(struct sock *sk, int pmtu) 17231b63edd6SYuchung Cheng { 17241b63edd6SYuchung Cheng /* Subtract TCP options size, not including SACKs */ 17251b63edd6SYuchung Cheng return __tcp_mtu_to_mss(sk, pmtu) - 17261b63edd6SYuchung Cheng (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr)); 17271b63edd6SYuchung Cheng } 1728c7bb4b89SEric Dumazet EXPORT_SYMBOL(tcp_mtu_to_mss); 17291b63edd6SYuchung Cheng 17305d424d5aSJohn Heffner /* Inverse of above */ 173167469601SEric Dumazet int tcp_mss_to_mtu(struct sock *sk, int mss) 17325d424d5aSJohn Heffner { 1733cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1734cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 17355d424d5aSJohn Heffner int mtu; 17365d424d5aSJohn Heffner 17375d424d5aSJohn Heffner mtu = mss + 17385d424d5aSJohn Heffner tp->tcp_header_len + 17395d424d5aSJohn Heffner icsk->icsk_ext_hdr_len + 17405d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 17415d424d5aSJohn Heffner 174267469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 174367469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 174467469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 174567469601SEric Dumazet 174667469601SEric Dumazet if (dst && dst_allfrag(dst)) 174767469601SEric Dumazet mtu += icsk->icsk_af_ops->net_frag_header_len; 174867469601SEric Dumazet } 17495d424d5aSJohn Heffner return mtu; 17505d424d5aSJohn Heffner } 1751556c6b46SNeal Cardwell EXPORT_SYMBOL(tcp_mss_to_mtu); 17525d424d5aSJohn Heffner 175367edfef7SAndi Kleen /* MTU probing init per socket */ 17545d424d5aSJohn Heffner void tcp_mtup_init(struct sock *sk) 17555d424d5aSJohn Heffner { 17565d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 17575d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 1758b0f9ca53SFan Du struct net *net = sock_net(sk); 17595d424d5aSJohn Heffner 1760f47d00e0SKuniyuki Iwashima icsk->icsk_mtup.enabled = READ_ONCE(net->ipv4.sysctl_tcp_mtu_probing) > 1; 17615d424d5aSJohn Heffner icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 17625d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 176388d78bc0SKuniyuki Iwashima icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, READ_ONCE(net->ipv4.sysctl_tcp_base_mss)); 17645d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 176505cbc0dbSFan Du if (icsk->icsk_mtup.enabled) 1766c74df29aSEric Dumazet icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; 17675d424d5aSJohn Heffner } 17684bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_mtup_init); 17695d424d5aSJohn Heffner 17701da177e4SLinus Torvalds /* This function synchronize snd mss to current pmtu/exthdr set. 17711da177e4SLinus Torvalds 17721da177e4SLinus Torvalds tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 17731da177e4SLinus Torvalds for TCP options, but includes only bare TCP header. 17741da177e4SLinus Torvalds 17751da177e4SLinus Torvalds tp->rx_opt.mss_clamp is mss negotiated at connection setup. 1776caa20d9aSStephen Hemminger It is minimum of user_mss and mss received with SYN. 17771da177e4SLinus Torvalds It also does not include TCP options. 17781da177e4SLinus Torvalds 1779d83d8461SArnaldo Carvalho de Melo inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. 17801da177e4SLinus Torvalds 17811da177e4SLinus Torvalds tp->mss_cache is current effective sending mss, including 17821da177e4SLinus Torvalds all tcp options except for SACKs. It is evaluated, 17831da177e4SLinus Torvalds taking into account current pmtu, but never exceeds 17841da177e4SLinus Torvalds tp->rx_opt.mss_clamp. 17851da177e4SLinus Torvalds 17861da177e4SLinus Torvalds NOTE1. rfc1122 clearly states that advertised MSS 17871da177e4SLinus Torvalds DOES NOT include either tcp or ip options. 17881da177e4SLinus Torvalds 1789d83d8461SArnaldo Carvalho de Melo NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache 1790d83d8461SArnaldo Carvalho de Melo are READ ONLY outside this function. --ANK (980731) 17911da177e4SLinus Torvalds */ 17921da177e4SLinus Torvalds unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 17931da177e4SLinus Torvalds { 17941da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1795d83d8461SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 17965d424d5aSJohn Heffner int mss_now; 17971da177e4SLinus Torvalds 17985d424d5aSJohn Heffner if (icsk->icsk_mtup.search_high > pmtu) 17995d424d5aSJohn Heffner icsk->icsk_mtup.search_high = pmtu; 18001da177e4SLinus Torvalds 18015d424d5aSJohn Heffner mss_now = tcp_mtu_to_mss(sk, pmtu); 1802409d22b4SIlpo Järvinen mss_now = tcp_bound_to_half_wnd(tp, mss_now); 18031da177e4SLinus Torvalds 18041da177e4SLinus Torvalds /* And store cached results */ 1805d83d8461SArnaldo Carvalho de Melo icsk->icsk_pmtu_cookie = pmtu; 18065d424d5aSJohn Heffner if (icsk->icsk_mtup.enabled) 18075d424d5aSJohn Heffner mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); 1808c1b4a7e6SDavid S. Miller tp->mss_cache = mss_now; 18091da177e4SLinus Torvalds 18101da177e4SLinus Torvalds return mss_now; 18111da177e4SLinus Torvalds } 18124bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_sync_mss); 18131da177e4SLinus Torvalds 18141da177e4SLinus Torvalds /* Compute the current effective MSS, taking SACKs and IP options, 18151da177e4SLinus Torvalds * and even PMTU discovery events into account. 18161da177e4SLinus Torvalds */ 18170c54b85fSIlpo Järvinen unsigned int tcp_current_mss(struct sock *sk) 18181da177e4SLinus Torvalds { 1819cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1820cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1821c1b4a7e6SDavid S. Miller u32 mss_now; 182295c96174SEric Dumazet unsigned int header_len; 182333ad798cSAdam Langley struct tcp_out_options opts; 182433ad798cSAdam Langley struct tcp_md5sig_key *md5; 18251da177e4SLinus Torvalds 1826c1b4a7e6SDavid S. Miller mss_now = tp->mss_cache; 1827c1b4a7e6SDavid S. Miller 18281da177e4SLinus Torvalds if (dst) { 18291da177e4SLinus Torvalds u32 mtu = dst_mtu(dst); 1830d83d8461SArnaldo Carvalho de Melo if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 18311da177e4SLinus Torvalds mss_now = tcp_sync_mss(sk, mtu); 18321da177e4SLinus Torvalds } 18331da177e4SLinus Torvalds 183433ad798cSAdam Langley header_len = tcp_established_options(sk, NULL, &opts, &md5) + 183533ad798cSAdam Langley sizeof(struct tcphdr); 183633ad798cSAdam Langley /* The mss_cache is sized based on tp->tcp_header_len, which assumes 183733ad798cSAdam Langley * some common options. If this is an odd packet (because we have SACK 183833ad798cSAdam Langley * blocks etc) then our calculated header_len will be different, and 183933ad798cSAdam Langley * we have to adjust mss_now correspondingly */ 184033ad798cSAdam Langley if (header_len != tp->tcp_header_len) { 184133ad798cSAdam Langley int delta = (int) header_len - tp->tcp_header_len; 184233ad798cSAdam Langley mss_now -= delta; 184333ad798cSAdam Langley } 1844cfb6eeb4SYOSHIFUJI Hideaki 18451da177e4SLinus Torvalds return mss_now; 18461da177e4SLinus Torvalds } 18471da177e4SLinus Torvalds 184886fd14adSWeiping Pan /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. 184986fd14adSWeiping Pan * As additional protections, we do not touch cwnd in retransmission phases, 185086fd14adSWeiping Pan * and if application hit its sndbuf limit recently. 185186fd14adSWeiping Pan */ 185286fd14adSWeiping Pan static void tcp_cwnd_application_limited(struct sock *sk) 1853a762a980SDavid S. Miller { 18549e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1855a762a980SDavid S. Miller 185686fd14adSWeiping Pan if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && 185786fd14adSWeiping Pan sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 185886fd14adSWeiping Pan /* Limited by application or receiver window. */ 185986fd14adSWeiping Pan u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); 186086fd14adSWeiping Pan u32 win_used = max(tp->snd_cwnd_used, init_win); 186140570375SEric Dumazet if (win_used < tcp_snd_cwnd(tp)) { 186286fd14adSWeiping Pan tp->snd_ssthresh = tcp_current_ssthresh(sk); 186340570375SEric Dumazet tcp_snd_cwnd_set(tp, (tcp_snd_cwnd(tp) + win_used) >> 1); 186486fd14adSWeiping Pan } 186586fd14adSWeiping Pan tp->snd_cwnd_used = 0; 186686fd14adSWeiping Pan } 1867c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 186886fd14adSWeiping Pan } 186986fd14adSWeiping Pan 1870ca8a2263SNeal Cardwell static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited) 1871a762a980SDavid S. Miller { 18721b1fc3fdSWei Wang const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 1873a762a980SDavid S. Miller struct tcp_sock *tp = tcp_sk(sk); 1874a762a980SDavid S. Miller 1875f4ce91ceSNeal Cardwell /* Track the strongest available signal of the degree to which the cwnd 1876f4ce91ceSNeal Cardwell * is fully utilized. If cwnd-limited then remember that fact for the 1877f4ce91ceSNeal Cardwell * current window. If not cwnd-limited then track the maximum number of 1878f4ce91ceSNeal Cardwell * outstanding packets in the current window. (If cwnd-limited then we 1879f4ce91ceSNeal Cardwell * chose to not update tp->max_packets_out to avoid an extra else 1880f4ce91ceSNeal Cardwell * clause with no functional impact.) 1881ca8a2263SNeal Cardwell */ 1882f4ce91ceSNeal Cardwell if (!before(tp->snd_una, tp->cwnd_usage_seq) || 1883f4ce91ceSNeal Cardwell is_cwnd_limited || 1884f4ce91ceSNeal Cardwell (!tp->is_cwnd_limited && 1885f4ce91ceSNeal Cardwell tp->packets_out > tp->max_packets_out)) { 1886ca8a2263SNeal Cardwell tp->is_cwnd_limited = is_cwnd_limited; 1887f4ce91ceSNeal Cardwell tp->max_packets_out = tp->packets_out; 1888f4ce91ceSNeal Cardwell tp->cwnd_usage_seq = tp->snd_nxt; 1889ca8a2263SNeal Cardwell } 1890e114a710SEric Dumazet 189124901551SEric Dumazet if (tcp_is_cwnd_limited(sk)) { 1892a762a980SDavid S. Miller /* Network is feed fully. */ 1893a762a980SDavid S. Miller tp->snd_cwnd_used = 0; 1894c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 1895a762a980SDavid S. Miller } else { 1896a762a980SDavid S. Miller /* Network starves. */ 1897a762a980SDavid S. Miller if (tp->packets_out > tp->snd_cwnd_used) 1898a762a980SDavid S. Miller tp->snd_cwnd_used = tp->packets_out; 1899a762a980SDavid S. Miller 19004845b571SKuniyuki Iwashima if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle) && 1901c2203cf7SEric Dumazet (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto && 19021b1fc3fdSWei Wang !ca_ops->cong_control) 1903a762a980SDavid S. Miller tcp_cwnd_application_limited(sk); 1904b0f71bd3SFrancis Yan 1905b0f71bd3SFrancis Yan /* The following conditions together indicate the starvation 1906b0f71bd3SFrancis Yan * is caused by insufficient sender buffer: 1907b0f71bd3SFrancis Yan * 1) just sent some data (see tcp_write_xmit) 1908b0f71bd3SFrancis Yan * 2) not cwnd limited (this else condition) 190975c119afSEric Dumazet * 3) no more data to send (tcp_write_queue_empty()) 1910b0f71bd3SFrancis Yan * 4) application is hitting buffer limit (SOCK_NOSPACE) 1911b0f71bd3SFrancis Yan */ 191275c119afSEric Dumazet if (tcp_write_queue_empty(sk) && sk->sk_socket && 1913b0f71bd3SFrancis Yan test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) && 1914b0f71bd3SFrancis Yan (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) 1915b0f71bd3SFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED); 1916a762a980SDavid S. Miller } 1917a762a980SDavid S. Miller } 1918a762a980SDavid S. Miller 1919d4589926SEric Dumazet /* Minshall's variant of the Nagle send check. */ 1920d4589926SEric Dumazet static bool tcp_minshall_check(const struct tcp_sock *tp) 1921d4589926SEric Dumazet { 1922d4589926SEric Dumazet return after(tp->snd_sml, tp->snd_una) && 1923d4589926SEric Dumazet !after(tp->snd_sml, tp->snd_nxt); 1924d4589926SEric Dumazet } 1925d4589926SEric Dumazet 1926d4589926SEric Dumazet /* Update snd_sml if this skb is under mss 1927d4589926SEric Dumazet * Note that a TSO packet might end with a sub-mss segment 1928d4589926SEric Dumazet * The test is really : 1929d4589926SEric Dumazet * if ((skb->len % mss) != 0) 1930d4589926SEric Dumazet * tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1931d4589926SEric Dumazet * But we can avoid doing the divide again given we already have 1932d4589926SEric Dumazet * skb_pcount = skb->len / mss_now 19330e3a4803SIlpo Järvinen */ 1934d4589926SEric Dumazet static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now, 1935d4589926SEric Dumazet const struct sk_buff *skb) 1936d4589926SEric Dumazet { 1937d4589926SEric Dumazet if (skb->len < tcp_skb_pcount(skb) * mss_now) 1938d4589926SEric Dumazet tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1939d4589926SEric Dumazet } 1940d4589926SEric Dumazet 1941d4589926SEric Dumazet /* Return false, if packet can be sent now without violation Nagle's rules: 1942d4589926SEric Dumazet * 1. It is full sized. (provided by caller in %partial bool) 1943d4589926SEric Dumazet * 2. Or it contains FIN. (already checked by caller) 1944d4589926SEric Dumazet * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. 1945d4589926SEric Dumazet * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 1946d4589926SEric Dumazet * With Minshall's modification: all sent small packets are ACKed. 1947d4589926SEric Dumazet */ 1948d4589926SEric Dumazet static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp, 1949cc93fc51SPeter Pan(潘卫平) int nonagle) 1950d4589926SEric Dumazet { 1951d4589926SEric Dumazet return partial && 1952d4589926SEric Dumazet ((nonagle & TCP_NAGLE_CORK) || 1953d4589926SEric Dumazet (!nonagle && tp->packets_out && tcp_minshall_check(tp))); 1954d4589926SEric Dumazet } 1955605ad7f1SEric Dumazet 1956605ad7f1SEric Dumazet /* Return how many segs we'd like on a TSO packet, 195765466904SEric Dumazet * depending on current pacing rate, and how close the peer is. 195865466904SEric Dumazet * 195965466904SEric Dumazet * Rationale is: 196065466904SEric Dumazet * - For close peers, we rather send bigger packets to reduce 196165466904SEric Dumazet * cpu costs, because occasional losses will be repaired fast. 196265466904SEric Dumazet * - For long distance/rtt flows, we would like to get ACK clocking 196365466904SEric Dumazet * with 1 ACK per ms. 196465466904SEric Dumazet * 196565466904SEric Dumazet * Use min_rtt to help adapt TSO burst size, with smaller min_rtt resulting 196665466904SEric Dumazet * in bigger TSO bursts. We we cut the RTT-based allowance in half 196765466904SEric Dumazet * for every 2^9 usec (aka 512 us) of RTT, so that the RTT-based allowance 196865466904SEric Dumazet * is below 1500 bytes after 6 * ~500 usec = 3ms. 1969605ad7f1SEric Dumazet */ 1970dcb8c9b4SEric Dumazet static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now, 19711b3878caSNeal Cardwell int min_tso_segs) 1972605ad7f1SEric Dumazet { 197365466904SEric Dumazet unsigned long bytes; 197465466904SEric Dumazet u32 r; 1975605ad7f1SEric Dumazet 197665466904SEric Dumazet bytes = sk->sk_pacing_rate >> READ_ONCE(sk->sk_pacing_shift); 1977605ad7f1SEric Dumazet 19782455e61bSKuniyuki Iwashima r = tcp_min_rtt(tcp_sk(sk)) >> READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_rtt_log); 197965466904SEric Dumazet if (r < BITS_PER_TYPE(sk->sk_gso_max_size)) 198065466904SEric Dumazet bytes += sk->sk_gso_max_size >> r; 1981605ad7f1SEric Dumazet 198265466904SEric Dumazet bytes = min_t(unsigned long, bytes, sk->sk_gso_max_size); 198365466904SEric Dumazet 198465466904SEric Dumazet return max_t(u32, bytes / mss_now, min_tso_segs); 1985605ad7f1SEric Dumazet } 1986605ad7f1SEric Dumazet 1987ed6e7268SNeal Cardwell /* Return the number of segments we want in the skb we are transmitting. 1988ed6e7268SNeal Cardwell * See if congestion control module wants to decide; otherwise, autosize. 1989ed6e7268SNeal Cardwell */ 1990ed6e7268SNeal Cardwell static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now) 1991ed6e7268SNeal Cardwell { 1992ed6e7268SNeal Cardwell const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 1993dcb8c9b4SEric Dumazet u32 min_tso, tso_segs; 1994ed6e7268SNeal Cardwell 1995dcb8c9b4SEric Dumazet min_tso = ca_ops->min_tso_segs ? 1996dcb8c9b4SEric Dumazet ca_ops->min_tso_segs(sk) : 1997e0bb4ab9SKuniyuki Iwashima READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs); 1998dcb8c9b4SEric Dumazet 1999dcb8c9b4SEric Dumazet tso_segs = tcp_tso_autosize(sk, mss_now, min_tso); 2000350c9f48SEric Dumazet return min_t(u32, tso_segs, sk->sk_gso_max_segs); 2001ed6e7268SNeal Cardwell } 2002ed6e7268SNeal Cardwell 2003d4589926SEric Dumazet /* Returns the portion of skb which can be sent right away */ 2004d4589926SEric Dumazet static unsigned int tcp_mss_split_point(const struct sock *sk, 2005d4589926SEric Dumazet const struct sk_buff *skb, 2006d4589926SEric Dumazet unsigned int mss_now, 2007d4589926SEric Dumazet unsigned int max_segs, 2008d4589926SEric Dumazet int nonagle) 2009c1b4a7e6SDavid S. Miller { 2010cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 2011d4589926SEric Dumazet u32 partial, needed, window, max_len; 2012c1b4a7e6SDavid S. Miller 201390840defSIlpo Järvinen window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 20141485348dSBen Hutchings max_len = mss_now * max_segs; 20150e3a4803SIlpo Järvinen 20161485348dSBen Hutchings if (likely(max_len <= window && skb != tcp_write_queue_tail(sk))) 20171485348dSBen Hutchings return max_len; 20180e3a4803SIlpo Järvinen 20195ea3a748SIlpo Järvinen needed = min(skb->len, window); 20205ea3a748SIlpo Järvinen 20211485348dSBen Hutchings if (max_len <= needed) 20221485348dSBen Hutchings return max_len; 20230e3a4803SIlpo Järvinen 2024d4589926SEric Dumazet partial = needed % mss_now; 2025d4589926SEric Dumazet /* If last segment is not a full MSS, check if Nagle rules allow us 2026d4589926SEric Dumazet * to include this last segment in this skb. 2027d4589926SEric Dumazet * Otherwise, we'll split the skb at last MSS boundary 2028d4589926SEric Dumazet */ 2029cc93fc51SPeter Pan(潘卫平) if (tcp_nagle_check(partial != 0, tp, nonagle)) 2030d4589926SEric Dumazet return needed - partial; 2031d4589926SEric Dumazet 2032d4589926SEric Dumazet return needed; 2033c1b4a7e6SDavid S. Miller } 2034c1b4a7e6SDavid S. Miller 2035c1b4a7e6SDavid S. Miller /* Can at least one segment of SKB be sent right now, according to the 2036c1b4a7e6SDavid S. Miller * congestion window rules? If so, return how many segments are allowed. 2037c1b4a7e6SDavid S. Miller */ 2038cf533ea5SEric Dumazet static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, 2039cf533ea5SEric Dumazet const struct sk_buff *skb) 2040c1b4a7e6SDavid S. Miller { 2041d649a7a8SEric Dumazet u32 in_flight, cwnd, halfcwnd; 2042c1b4a7e6SDavid S. Miller 2043c1b4a7e6SDavid S. Miller /* Don't be strict about the congestion window for the final FIN. */ 20444de075e0SEric Dumazet if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && 20454de075e0SEric Dumazet tcp_skb_pcount(skb) == 1) 2046c1b4a7e6SDavid S. Miller return 1; 2047c1b4a7e6SDavid S. Miller 2048c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 204940570375SEric Dumazet cwnd = tcp_snd_cwnd(tp); 2050d649a7a8SEric Dumazet if (in_flight >= cwnd) 2051c1b4a7e6SDavid S. Miller return 0; 2052d649a7a8SEric Dumazet 2053d649a7a8SEric Dumazet /* For better scheduling, ensure we have at least 2054d649a7a8SEric Dumazet * 2 GSO packets in flight. 2055d649a7a8SEric Dumazet */ 2056d649a7a8SEric Dumazet halfcwnd = max(cwnd >> 1, 1U); 2057d649a7a8SEric Dumazet return min(halfcwnd, cwnd - in_flight); 2058c1b4a7e6SDavid S. Miller } 2059c1b4a7e6SDavid S. Miller 2060b595076aSUwe Kleine-König /* Initialize TSO state of a skb. 206167edfef7SAndi Kleen * This must be invoked the first time we consider transmitting 2062c1b4a7e6SDavid S. Miller * SKB onto the wire. 2063c1b4a7e6SDavid S. Miller */ 20645bbb432cSEric Dumazet static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now) 2065c1b4a7e6SDavid S. Miller { 2066c1b4a7e6SDavid S. Miller int tso_segs = tcp_skb_pcount(skb); 2067c1b4a7e6SDavid S. Miller 2068f8269a49SIlpo Järvinen if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { 20695bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 2070c1b4a7e6SDavid S. Miller tso_segs = tcp_skb_pcount(skb); 2071c1b4a7e6SDavid S. Miller } 2072c1b4a7e6SDavid S. Miller return tso_segs; 2073c1b4a7e6SDavid S. Miller } 2074c1b4a7e6SDavid S. Miller 2075c1b4a7e6SDavid S. Miller 2076a2a385d6SEric Dumazet /* Return true if the Nagle test allows this packet to be 2077c1b4a7e6SDavid S. Miller * sent now. 2078c1b4a7e6SDavid S. Miller */ 2079a2a385d6SEric Dumazet static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, 2080c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 2081c1b4a7e6SDavid S. Miller { 2082c1b4a7e6SDavid S. Miller /* Nagle rule does not apply to frames, which sit in the middle of the 2083c1b4a7e6SDavid S. Miller * write_queue (they have no chances to get new data). 2084c1b4a7e6SDavid S. Miller * 2085c1b4a7e6SDavid S. Miller * This is implemented in the callers, where they modify the 'nonagle' 2086c1b4a7e6SDavid S. Miller * argument based upon the location of SKB in the send queue. 2087c1b4a7e6SDavid S. Miller */ 2088c1b4a7e6SDavid S. Miller if (nonagle & TCP_NAGLE_PUSH) 2089a2a385d6SEric Dumazet return true; 2090c1b4a7e6SDavid S. Miller 20919b44190dSYuchung Cheng /* Don't use the nagle rule for urgent data (or for the final FIN). */ 20929b44190dSYuchung Cheng if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) 2093a2a385d6SEric Dumazet return true; 2094c1b4a7e6SDavid S. Miller 2095cc93fc51SPeter Pan(潘卫平) if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle)) 2096a2a385d6SEric Dumazet return true; 2097c1b4a7e6SDavid S. Miller 2098a2a385d6SEric Dumazet return false; 2099c1b4a7e6SDavid S. Miller } 2100c1b4a7e6SDavid S. Miller 2101c1b4a7e6SDavid S. Miller /* Does at least the first segment of SKB fit into the send window? */ 2102a2a385d6SEric Dumazet static bool tcp_snd_wnd_test(const struct tcp_sock *tp, 2103a2a385d6SEric Dumazet const struct sk_buff *skb, 2104056834d9SIlpo Järvinen unsigned int cur_mss) 2105c1b4a7e6SDavid S. Miller { 2106c1b4a7e6SDavid S. Miller u32 end_seq = TCP_SKB_CB(skb)->end_seq; 2107c1b4a7e6SDavid S. Miller 2108c1b4a7e6SDavid S. Miller if (skb->len > cur_mss) 2109c1b4a7e6SDavid S. Miller end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 2110c1b4a7e6SDavid S. Miller 211190840defSIlpo Järvinen return !after(end_seq, tcp_wnd_end(tp)); 2112c1b4a7e6SDavid S. Miller } 2113c1b4a7e6SDavid S. Miller 2114c1b4a7e6SDavid S. Miller /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 2115c1b4a7e6SDavid S. Miller * which is put after SKB on the list. It is very much like 2116c1b4a7e6SDavid S. Miller * tcp_fragment() except that it may make several kinds of assumptions 2117c1b4a7e6SDavid S. Miller * in order to speed up the splitting operation. In particular, we 2118c1b4a7e6SDavid S. Miller * know that all the data is in scatter-gather pages, and that the 2119c1b4a7e6SDavid S. Miller * packet has never been sent out before (and thus is not cloned). 2120c1b4a7e6SDavid S. Miller */ 212156483341SEric Dumazet static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, 2122c4ead4c5SEric Dumazet unsigned int mss_now, gfp_t gfp) 2123c1b4a7e6SDavid S. Miller { 2124c1b4a7e6SDavid S. Miller int nlen = skb->len - len; 212556483341SEric Dumazet struct sk_buff *buff; 21269ce01461SIlpo Järvinen u8 flags; 2127c1b4a7e6SDavid S. Miller 2128c1b4a7e6SDavid S. Miller /* All of a TSO frame must be composed of paged data. */ 2129c8ac3774SHerbert Xu if (skb->len != skb->data_len) 213056483341SEric Dumazet return tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE, 213156483341SEric Dumazet skb, len, mss_now, gfp); 2132c1b4a7e6SDavid S. Miller 2133f8dd3b8dSEric Dumazet buff = tcp_stream_alloc_skb(sk, 0, gfp, true); 213451456b29SIan Morris if (unlikely(!buff)) 2135c1b4a7e6SDavid S. Miller return -ENOMEM; 213641477662SJakub Kicinski skb_copy_decrypted(buff, skb); 21375a369ca6SPaolo Abeni mptcp_skb_ext_copy(buff, skb); 2138c1b4a7e6SDavid S. Miller 2139ab4e846aSEric Dumazet sk_wmem_queued_add(sk, buff->truesize); 21403ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 2141b60b49eaSHerbert Xu buff->truesize += nlen; 2142c1b4a7e6SDavid S. Miller skb->truesize -= nlen; 2143c1b4a7e6SDavid S. Miller 2144c1b4a7e6SDavid S. Miller /* Correct the sequence numbers. */ 2145c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 2146c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 2147c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 2148c1b4a7e6SDavid S. Miller 2149c1b4a7e6SDavid S. Miller /* PSH and FIN should only be set in the second packet. */ 21504de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 21514de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 21524de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 2153c1b4a7e6SDavid S. Miller 2154a166140eSMartin KaFai Lau tcp_skb_fragment_eor(skb, buff); 2155a166140eSMartin KaFai Lau 2156c1b4a7e6SDavid S. Miller skb_split(skb, buff, len); 2157490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 2158c1b4a7e6SDavid S. Miller 2159c1b4a7e6SDavid S. Miller /* Fix up tso_factor for both original and new SKB. */ 21605bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 21615bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 2162c1b4a7e6SDavid S. Miller 2163c1b4a7e6SDavid S. Miller /* Link BUFF into the send queue. */ 2164f4a775d1SEric Dumazet __skb_header_release(buff); 216556483341SEric Dumazet tcp_insert_write_queue_after(skb, buff, sk, TCP_FRAG_IN_WRITE_QUEUE); 2166c1b4a7e6SDavid S. Miller 2167c1b4a7e6SDavid S. Miller return 0; 2168c1b4a7e6SDavid S. Miller } 2169c1b4a7e6SDavid S. Miller 2170c1b4a7e6SDavid S. Miller /* Try to defer sending, if possible, in order to minimize the amount 2171c1b4a7e6SDavid S. Miller * of TSO splitting we do. View it as a kind of TSO Nagle test. 2172c1b4a7e6SDavid S. Miller * 2173c1b4a7e6SDavid S. Miller * This algorithm is from John Heffner. 2174c1b4a7e6SDavid S. Miller */ 2175ca8a2263SNeal Cardwell static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb, 2176f9bfe4e6SEric Dumazet bool *is_cwnd_limited, 2177f9bfe4e6SEric Dumazet bool *is_rwnd_limited, 2178f9bfe4e6SEric Dumazet u32 max_segs) 2179c1b4a7e6SDavid S. Miller { 21806687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 2181f1c6ea38SEric Dumazet u32 send_win, cong_win, limit, in_flight; 218250c8339eSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 218350c8339eSEric Dumazet struct sk_buff *head; 2184ad9f4f50SEric Dumazet int win_divisor; 2185f1c6ea38SEric Dumazet s64 delta; 2186c1b4a7e6SDavid S. Miller 218799d7662aSEric Dumazet if (icsk->icsk_ca_state >= TCP_CA_Recovery) 2188ae8064acSJohn Heffner goto send_now; 2189ae8064acSJohn Heffner 21905f852eb5SEric Dumazet /* Avoid bursty behavior by allowing defer 2191a682850aSEric Dumazet * only if the last write was recent (1 ms). 2192a682850aSEric Dumazet * Note that tp->tcp_wstamp_ns can be in the future if we have 2193a682850aSEric Dumazet * packets waiting in a qdisc or device for EDT delivery. 21945f852eb5SEric Dumazet */ 2195a682850aSEric Dumazet delta = tp->tcp_clock_cache - tp->tcp_wstamp_ns - NSEC_PER_MSEC; 2196a682850aSEric Dumazet if (delta > 0) 2197ae8064acSJohn Heffner goto send_now; 2198908a75c1SDavid S. Miller 2199c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 2200c1b4a7e6SDavid S. Miller 2201c8c9aeb5SStefano Brivio BUG_ON(tcp_skb_pcount(skb) <= 1); 220240570375SEric Dumazet BUG_ON(tcp_snd_cwnd(tp) <= in_flight); 2203c1b4a7e6SDavid S. Miller 220490840defSIlpo Järvinen send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 2205c1b4a7e6SDavid S. Miller 2206c1b4a7e6SDavid S. Miller /* From in_flight test above, we know that cwnd > in_flight. */ 220740570375SEric Dumazet cong_win = (tcp_snd_cwnd(tp) - in_flight) * tp->mss_cache; 2208c1b4a7e6SDavid S. Miller 2209c1b4a7e6SDavid S. Miller limit = min(send_win, cong_win); 2210c1b4a7e6SDavid S. Miller 2211ba244fe9SDavid S. Miller /* If a full-sized TSO skb can be sent, do it. */ 2212605ad7f1SEric Dumazet if (limit >= max_segs * tp->mss_cache) 2213ae8064acSJohn Heffner goto send_now; 2214ba244fe9SDavid S. Miller 221562ad2761SIlpo Järvinen /* Middle in queue won't get any more data, full sendable already? */ 221662ad2761SIlpo Järvinen if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) 221762ad2761SIlpo Järvinen goto send_now; 221862ad2761SIlpo Järvinen 22195bbcc0f5SLinus Torvalds win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor); 2220ad9f4f50SEric Dumazet if (win_divisor) { 222140570375SEric Dumazet u32 chunk = min(tp->snd_wnd, tcp_snd_cwnd(tp) * tp->mss_cache); 2222c1b4a7e6SDavid S. Miller 2223c1b4a7e6SDavid S. Miller /* If at least some fraction of a window is available, 2224c1b4a7e6SDavid S. Miller * just use it. 2225c1b4a7e6SDavid S. Miller */ 2226ad9f4f50SEric Dumazet chunk /= win_divisor; 2227c1b4a7e6SDavid S. Miller if (limit >= chunk) 2228ae8064acSJohn Heffner goto send_now; 2229c1b4a7e6SDavid S. Miller } else { 2230c1b4a7e6SDavid S. Miller /* Different approach, try not to defer past a single 2231c1b4a7e6SDavid S. Miller * ACK. Receiver should ACK every other full sized 2232c1b4a7e6SDavid S. Miller * frame, so if we have space for more than 3 frames 2233c1b4a7e6SDavid S. Miller * then send now. 2234c1b4a7e6SDavid S. Miller */ 22356b5a5c0dSNeal Cardwell if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) 2236ae8064acSJohn Heffner goto send_now; 2237c1b4a7e6SDavid S. Miller } 2238c1b4a7e6SDavid S. Miller 223975c119afSEric Dumazet /* TODO : use tsorted_sent_queue ? */ 224075c119afSEric Dumazet head = tcp_rtx_queue_head(sk); 224175c119afSEric Dumazet if (!head) 224275c119afSEric Dumazet goto send_now; 2243f1c6ea38SEric Dumazet delta = tp->tcp_clock_cache - head->tstamp; 224450c8339eSEric Dumazet /* If next ACK is likely to come too late (half srtt), do not defer */ 2245f1c6ea38SEric Dumazet if ((s64)(delta - (u64)NSEC_PER_USEC * (tp->srtt_us >> 4)) < 0) 224650c8339eSEric Dumazet goto send_now; 224750c8339eSEric Dumazet 2248f9bfe4e6SEric Dumazet /* Ok, it looks like it is advisable to defer. 2249f9bfe4e6SEric Dumazet * Three cases are tracked : 2250f9bfe4e6SEric Dumazet * 1) We are cwnd-limited 2251f9bfe4e6SEric Dumazet * 2) We are rwnd-limited 2252f9bfe4e6SEric Dumazet * 3) We are application limited. 2253f9bfe4e6SEric Dumazet */ 2254f9bfe4e6SEric Dumazet if (cong_win < send_win) { 2255f9bfe4e6SEric Dumazet if (cong_win <= skb->len) { 2256ca8a2263SNeal Cardwell *is_cwnd_limited = true; 2257f9bfe4e6SEric Dumazet return true; 2258f9bfe4e6SEric Dumazet } 2259f9bfe4e6SEric Dumazet } else { 2260f9bfe4e6SEric Dumazet if (send_win <= skb->len) { 2261f9bfe4e6SEric Dumazet *is_rwnd_limited = true; 2262f9bfe4e6SEric Dumazet return true; 2263f9bfe4e6SEric Dumazet } 2264f9bfe4e6SEric Dumazet } 2265f9bfe4e6SEric Dumazet 2266f9bfe4e6SEric Dumazet /* If this packet won't get more data, do not wait. */ 2267d8ed257fSEric Dumazet if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) || 2268d8ed257fSEric Dumazet TCP_SKB_CB(skb)->eor) 2269f9bfe4e6SEric Dumazet goto send_now; 2270ca8a2263SNeal Cardwell 2271a2a385d6SEric Dumazet return true; 2272ae8064acSJohn Heffner 2273ae8064acSJohn Heffner send_now: 2274a2a385d6SEric Dumazet return false; 2275c1b4a7e6SDavid S. Miller } 2276c1b4a7e6SDavid S. Miller 227705cbc0dbSFan Du static inline void tcp_mtu_check_reprobe(struct sock *sk) 227805cbc0dbSFan Du { 227905cbc0dbSFan Du struct inet_connection_sock *icsk = inet_csk(sk); 228005cbc0dbSFan Du struct tcp_sock *tp = tcp_sk(sk); 228105cbc0dbSFan Du struct net *net = sock_net(sk); 228205cbc0dbSFan Du u32 interval; 228305cbc0dbSFan Du s32 delta; 228405cbc0dbSFan Du 22852a85388fSKuniyuki Iwashima interval = READ_ONCE(net->ipv4.sysctl_tcp_probe_interval); 2286c74df29aSEric Dumazet delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp; 228705cbc0dbSFan Du if (unlikely(delta >= interval * HZ)) { 228805cbc0dbSFan Du int mss = tcp_current_mss(sk); 228905cbc0dbSFan Du 229005cbc0dbSFan Du /* Update current search range */ 229105cbc0dbSFan Du icsk->icsk_mtup.probe_size = 0; 229205cbc0dbSFan Du icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + 229305cbc0dbSFan Du sizeof(struct tcphdr) + 229405cbc0dbSFan Du icsk->icsk_af_ops->net_header_len; 229505cbc0dbSFan Du icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss); 229605cbc0dbSFan Du 229705cbc0dbSFan Du /* Update probe time stamp */ 2298c74df29aSEric Dumazet icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; 229905cbc0dbSFan Du } 230005cbc0dbSFan Du } 230105cbc0dbSFan Du 2302808cf9e3SIlya Lesokhin static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len) 2303808cf9e3SIlya Lesokhin { 2304808cf9e3SIlya Lesokhin struct sk_buff *skb, *next; 2305808cf9e3SIlya Lesokhin 2306808cf9e3SIlya Lesokhin skb = tcp_send_head(sk); 2307808cf9e3SIlya Lesokhin tcp_for_write_queue_from_safe(skb, next, sk) { 2308808cf9e3SIlya Lesokhin if (len <= skb->len) 2309808cf9e3SIlya Lesokhin break; 2310808cf9e3SIlya Lesokhin 23119b65b17dSTalal Ahmad if (unlikely(TCP_SKB_CB(skb)->eor) || 23129b65b17dSTalal Ahmad tcp_has_tx_tstamp(skb) || 23139b65b17dSTalal Ahmad !skb_pure_zcopy_same(skb, next)) 2314808cf9e3SIlya Lesokhin return false; 2315808cf9e3SIlya Lesokhin 2316808cf9e3SIlya Lesokhin len -= skb->len; 2317808cf9e3SIlya Lesokhin } 2318808cf9e3SIlya Lesokhin 2319808cf9e3SIlya Lesokhin return true; 2320808cf9e3SIlya Lesokhin } 2321808cf9e3SIlya Lesokhin 232273601329SEric Dumazet static int tcp_clone_payload(struct sock *sk, struct sk_buff *to, 232373601329SEric Dumazet int probe_size) 232473601329SEric Dumazet { 232573601329SEric Dumazet skb_frag_t *lastfrag = NULL, *fragto = skb_shinfo(to)->frags; 232673601329SEric Dumazet int i, todo, len = 0, nr_frags = 0; 232773601329SEric Dumazet const struct sk_buff *skb; 232873601329SEric Dumazet 232973601329SEric Dumazet if (!sk_wmem_schedule(sk, to->truesize + probe_size)) 233073601329SEric Dumazet return -ENOMEM; 233173601329SEric Dumazet 233273601329SEric Dumazet skb_queue_walk(&sk->sk_write_queue, skb) { 233373601329SEric Dumazet const skb_frag_t *fragfrom = skb_shinfo(skb)->frags; 233473601329SEric Dumazet 233573601329SEric Dumazet if (skb_headlen(skb)) 233673601329SEric Dumazet return -EINVAL; 233773601329SEric Dumazet 233873601329SEric Dumazet for (i = 0; i < skb_shinfo(skb)->nr_frags; i++, fragfrom++) { 233973601329SEric Dumazet if (len >= probe_size) 234073601329SEric Dumazet goto commit; 234173601329SEric Dumazet todo = min_t(int, skb_frag_size(fragfrom), 234273601329SEric Dumazet probe_size - len); 234373601329SEric Dumazet len += todo; 234473601329SEric Dumazet if (lastfrag && 234573601329SEric Dumazet skb_frag_page(fragfrom) == skb_frag_page(lastfrag) && 234673601329SEric Dumazet skb_frag_off(fragfrom) == skb_frag_off(lastfrag) + 234773601329SEric Dumazet skb_frag_size(lastfrag)) { 234873601329SEric Dumazet skb_frag_size_add(lastfrag, todo); 234973601329SEric Dumazet continue; 235073601329SEric Dumazet } 235173601329SEric Dumazet if (unlikely(nr_frags == MAX_SKB_FRAGS)) 235273601329SEric Dumazet return -E2BIG; 235373601329SEric Dumazet skb_frag_page_copy(fragto, fragfrom); 235473601329SEric Dumazet skb_frag_off_copy(fragto, fragfrom); 235573601329SEric Dumazet skb_frag_size_set(fragto, todo); 235673601329SEric Dumazet nr_frags++; 235773601329SEric Dumazet lastfrag = fragto++; 235873601329SEric Dumazet } 235973601329SEric Dumazet } 236073601329SEric Dumazet commit: 236173601329SEric Dumazet WARN_ON_ONCE(len != probe_size); 236273601329SEric Dumazet for (i = 0; i < nr_frags; i++) 236373601329SEric Dumazet skb_frag_ref(to, i); 236473601329SEric Dumazet 236573601329SEric Dumazet skb_shinfo(to)->nr_frags = nr_frags; 236673601329SEric Dumazet to->truesize += probe_size; 236773601329SEric Dumazet to->len += probe_size; 236873601329SEric Dumazet to->data_len += probe_size; 236973601329SEric Dumazet __skb_header_release(to); 237073601329SEric Dumazet return 0; 237173601329SEric Dumazet } 237273601329SEric Dumazet 23735d424d5aSJohn Heffner /* Create a new MTU probe if we are ready. 237467edfef7SAndi Kleen * MTU probe is regularly attempting to increase the path MTU by 237567edfef7SAndi Kleen * deliberately sending larger packets. This discovers routing 237667edfef7SAndi Kleen * changes resulting in larger path MTUs. 237767edfef7SAndi Kleen * 23785d424d5aSJohn Heffner * Returns 0 if we should wait to probe (no cwnd available), 23795d424d5aSJohn Heffner * 1 if a probe was sent, 2380056834d9SIlpo Järvinen * -1 otherwise 2381056834d9SIlpo Järvinen */ 23825d424d5aSJohn Heffner static int tcp_mtu_probe(struct sock *sk) 23835d424d5aSJohn Heffner { 23845d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 238512a59abcSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 23865d424d5aSJohn Heffner struct sk_buff *skb, *nskb, *next; 23876b58e0a5SFan Du struct net *net = sock_net(sk); 23885d424d5aSJohn Heffner int probe_size; 238991cc17c0SIlpo Järvinen int size_needed; 239012a59abcSEric Dumazet int copy, len; 23915d424d5aSJohn Heffner int mss_now; 23926b58e0a5SFan Du int interval; 23935d424d5aSJohn Heffner 23945d424d5aSJohn Heffner /* Not currently probing/verifying, 23955d424d5aSJohn Heffner * not in recovery, 23965d424d5aSJohn Heffner * have enough cwnd, and 239712a59abcSEric Dumazet * not SACKing (the variable headers throw things off) 239812a59abcSEric Dumazet */ 239912a59abcSEric Dumazet if (likely(!icsk->icsk_mtup.enabled || 24005d424d5aSJohn Heffner icsk->icsk_mtup.probe_size || 24015d424d5aSJohn Heffner inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 240240570375SEric Dumazet tcp_snd_cwnd(tp) < 11 || 240312a59abcSEric Dumazet tp->rx_opt.num_sacks || tp->rx_opt.dsack)) 24045d424d5aSJohn Heffner return -1; 24055d424d5aSJohn Heffner 24066b58e0a5SFan Du /* Use binary search for probe_size between tcp_mss_base, 24076b58e0a5SFan Du * and current mss_clamp. if (search_high - search_low) 24086b58e0a5SFan Du * smaller than a threshold, backoff from probing. 24096b58e0a5SFan Du */ 24100c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 24116b58e0a5SFan Du probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high + 24126b58e0a5SFan Du icsk->icsk_mtup.search_low) >> 1); 241391cc17c0SIlpo Järvinen size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; 24146b58e0a5SFan Du interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low; 241505cbc0dbSFan Du /* When misfortune happens, we are reprobing actively, 241605cbc0dbSFan Du * and then reprobe timer has expired. We stick with current 241705cbc0dbSFan Du * probing process by not resetting search range to its orignal. 241805cbc0dbSFan Du */ 24196b58e0a5SFan Du if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) || 242092c0aa41SKuniyuki Iwashima interval < READ_ONCE(net->ipv4.sysctl_tcp_probe_threshold)) { 242105cbc0dbSFan Du /* Check whether enough time has elaplased for 242205cbc0dbSFan Du * another round of probing. 242305cbc0dbSFan Du */ 242405cbc0dbSFan Du tcp_mtu_check_reprobe(sk); 24255d424d5aSJohn Heffner return -1; 24265d424d5aSJohn Heffner } 24275d424d5aSJohn Heffner 24285d424d5aSJohn Heffner /* Have enough data in the send queue to probe? */ 24297f9c33e5SIlpo Järvinen if (tp->write_seq - tp->snd_nxt < size_needed) 24305d424d5aSJohn Heffner return -1; 24315d424d5aSJohn Heffner 243291cc17c0SIlpo Järvinen if (tp->snd_wnd < size_needed) 24335d424d5aSJohn Heffner return -1; 243490840defSIlpo Järvinen if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) 24355d424d5aSJohn Heffner return 0; 24365d424d5aSJohn Heffner 2437d67c58e9SIlpo Järvinen /* Do we need to wait to drain cwnd? With none in flight, don't stall */ 243840570375SEric Dumazet if (tcp_packets_in_flight(tp) + 2 > tcp_snd_cwnd(tp)) { 2439d67c58e9SIlpo Järvinen if (!tcp_packets_in_flight(tp)) 24405d424d5aSJohn Heffner return -1; 24415d424d5aSJohn Heffner else 24425d424d5aSJohn Heffner return 0; 24435d424d5aSJohn Heffner } 24445d424d5aSJohn Heffner 2445808cf9e3SIlya Lesokhin if (!tcp_can_coalesce_send_queue_head(sk, probe_size)) 2446808cf9e3SIlya Lesokhin return -1; 2447808cf9e3SIlya Lesokhin 24485d424d5aSJohn Heffner /* We're allowed to probe. Build it now. */ 244973601329SEric Dumazet nskb = tcp_stream_alloc_skb(sk, 0, GFP_ATOMIC, false); 245051456b29SIan Morris if (!nskb) 24515d424d5aSJohn Heffner return -1; 245273601329SEric Dumazet 245373601329SEric Dumazet /* build the payload, and be prepared to abort if this fails. */ 245473601329SEric Dumazet if (tcp_clone_payload(sk, nskb, probe_size)) { 245573601329SEric Dumazet consume_skb(nskb); 245673601329SEric Dumazet return -1; 245773601329SEric Dumazet } 2458ab4e846aSEric Dumazet sk_wmem_queued_add(sk, nskb->truesize); 24593ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 24605d424d5aSJohn Heffner 2461fe067e8aSDavid S. Miller skb = tcp_send_head(sk); 246241477662SJakub Kicinski skb_copy_decrypted(nskb, skb); 24635a369ca6SPaolo Abeni mptcp_skb_ext_copy(nskb, skb); 24645d424d5aSJohn Heffner 24655d424d5aSJohn Heffner TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 24665d424d5aSJohn Heffner TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 24674de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; 24685d424d5aSJohn Heffner 246950c4817eSIlpo Järvinen tcp_insert_write_queue_before(nskb, skb, sk); 24702b7cda9cSEric Dumazet tcp_highest_sack_replace(sk, skb, nskb); 247150c4817eSIlpo Järvinen 24725d424d5aSJohn Heffner len = 0; 2473234b6860SIlpo Järvinen tcp_for_write_queue_from_safe(skb, next, sk) { 24745d424d5aSJohn Heffner copy = min_t(int, skb->len, probe_size - len); 24755d424d5aSJohn Heffner 24765d424d5aSJohn Heffner if (skb->len <= copy) { 24775d424d5aSJohn Heffner /* We've eaten all the data from this skb. 24785d424d5aSJohn Heffner * Throw it away. */ 24794de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 2480808cf9e3SIlya Lesokhin /* If this is the last SKB we copy and eor is set 2481808cf9e3SIlya Lesokhin * we need to propagate it to the new skb. 2482808cf9e3SIlya Lesokhin */ 2483808cf9e3SIlya Lesokhin TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor; 2484888a5c53SWillem de Bruijn tcp_skb_collapse_tstamp(nskb, skb); 2485fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 248603271f3aSTalal Ahmad tcp_wmem_free_skb(sk, skb); 24875d424d5aSJohn Heffner } else { 24884de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & 2489a3433f35SChangli Gao ~(TCPHDR_FIN|TCPHDR_PSH); 24905d424d5aSJohn Heffner if (!skb_shinfo(skb)->nr_frags) { 24915d424d5aSJohn Heffner skb_pull(skb, copy); 24925d424d5aSJohn Heffner } else { 24935d424d5aSJohn Heffner __pskb_trim_head(skb, copy); 24945bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 24955d424d5aSJohn Heffner } 24965d424d5aSJohn Heffner TCP_SKB_CB(skb)->seq += copy; 24975d424d5aSJohn Heffner } 24985d424d5aSJohn Heffner 24995d424d5aSJohn Heffner len += copy; 2500234b6860SIlpo Järvinen 2501234b6860SIlpo Järvinen if (len >= probe_size) 2502234b6860SIlpo Järvinen break; 25035d424d5aSJohn Heffner } 25045bbb432cSEric Dumazet tcp_init_tso_segs(nskb, nskb->len); 25055d424d5aSJohn Heffner 25065d424d5aSJohn Heffner /* We're ready to send. If this fails, the probe will 25077faee5c0SEric Dumazet * be resegmented into mss-sized pieces by tcp_write_xmit(). 25087faee5c0SEric Dumazet */ 25095d424d5aSJohn Heffner if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 25105d424d5aSJohn Heffner /* Decrement cwnd here because we are sending 25115d424d5aSJohn Heffner * effectively two packets. */ 251240570375SEric Dumazet tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) - 1); 251366f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, nskb); 25145d424d5aSJohn Heffner 25155d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 25160e7b1368SJohn Heffner tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 25170e7b1368SJohn Heffner tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 25185d424d5aSJohn Heffner 25195d424d5aSJohn Heffner return 1; 25205d424d5aSJohn Heffner } 25215d424d5aSJohn Heffner 25225d424d5aSJohn Heffner return -1; 25235d424d5aSJohn Heffner } 25245d424d5aSJohn Heffner 2525864e5c09SEric Dumazet static bool tcp_pacing_check(struct sock *sk) 2526218af599SEric Dumazet { 2527864e5c09SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 2528864e5c09SEric Dumazet 2529864e5c09SEric Dumazet if (!tcp_needs_internal_pacing(sk)) 2530864e5c09SEric Dumazet return false; 2531864e5c09SEric Dumazet 2532864e5c09SEric Dumazet if (tp->tcp_wstamp_ns <= tp->tcp_clock_cache) 2533864e5c09SEric Dumazet return false; 2534864e5c09SEric Dumazet 2535864e5c09SEric Dumazet if (!hrtimer_is_queued(&tp->pacing_timer)) { 2536864e5c09SEric Dumazet hrtimer_start(&tp->pacing_timer, 2537864e5c09SEric Dumazet ns_to_ktime(tp->tcp_wstamp_ns), 2538864e5c09SEric Dumazet HRTIMER_MODE_ABS_PINNED_SOFT); 2539864e5c09SEric Dumazet sock_hold(sk); 2540864e5c09SEric Dumazet } 2541864e5c09SEric Dumazet return true; 2542218af599SEric Dumazet } 2543218af599SEric Dumazet 2544f9616c35SEric Dumazet /* TCP Small Queues : 2545f9616c35SEric Dumazet * Control number of packets in qdisc/devices to two packets / or ~1 ms. 2546f9616c35SEric Dumazet * (These limits are doubled for retransmits) 2547f9616c35SEric Dumazet * This allows for : 2548f9616c35SEric Dumazet * - better RTT estimation and ACK scheduling 2549f9616c35SEric Dumazet * - faster recovery 2550f9616c35SEric Dumazet * - high rates 2551f9616c35SEric Dumazet * Alas, some drivers / subsystems require a fair amount 2552f9616c35SEric Dumazet * of queued bytes to ensure line rate. 2553f9616c35SEric Dumazet * One example is wifi aggregation (802.11 AMPDU) 2554f9616c35SEric Dumazet */ 2555f9616c35SEric Dumazet static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb, 2556f9616c35SEric Dumazet unsigned int factor) 2557f9616c35SEric Dumazet { 255876a9ebe8SEric Dumazet unsigned long limit; 2559f9616c35SEric Dumazet 256076a9ebe8SEric Dumazet limit = max_t(unsigned long, 256176a9ebe8SEric Dumazet 2 * skb->truesize, 25627c68fa2bSEric Dumazet sk->sk_pacing_rate >> READ_ONCE(sk->sk_pacing_shift)); 2563c73e5807SEric Dumazet if (sk->sk_pacing_status == SK_PACING_NONE) 256476a9ebe8SEric Dumazet limit = min_t(unsigned long, limit, 25659fb90193SKuniyuki Iwashima READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes)); 2566f9616c35SEric Dumazet limit <<= factor; 2567f9616c35SEric Dumazet 2568a842fe14SEric Dumazet if (static_branch_unlikely(&tcp_tx_delay_enabled) && 2569a842fe14SEric Dumazet tcp_sk(sk)->tcp_tx_delay) { 2570a842fe14SEric Dumazet u64 extra_bytes = (u64)sk->sk_pacing_rate * tcp_sk(sk)->tcp_tx_delay; 2571a842fe14SEric Dumazet 2572a842fe14SEric Dumazet /* TSQ is based on skb truesize sum (sk_wmem_alloc), so we 2573a842fe14SEric Dumazet * approximate our needs assuming an ~100% skb->truesize overhead. 2574a842fe14SEric Dumazet * USEC_PER_SEC is approximated by 2^20. 2575a842fe14SEric Dumazet * do_div(extra_bytes, USEC_PER_SEC/2) is replaced by a right shift. 2576a842fe14SEric Dumazet */ 2577a842fe14SEric Dumazet extra_bytes >>= (20 - 1); 2578a842fe14SEric Dumazet limit += extra_bytes; 2579a842fe14SEric Dumazet } 258014afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) > limit) { 258175c119afSEric Dumazet /* Always send skb if rtx queue is empty. 258275eefc6cSEric Dumazet * No need to wait for TX completion to call us back, 258375eefc6cSEric Dumazet * after softirq/tasklet schedule. 258475eefc6cSEric Dumazet * This helps when TX completions are delayed too much. 258575eefc6cSEric Dumazet */ 258675c119afSEric Dumazet if (tcp_rtx_queue_empty(sk)) 258775eefc6cSEric Dumazet return false; 258875eefc6cSEric Dumazet 25897aa5470cSEric Dumazet set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags); 2590f9616c35SEric Dumazet /* It is possible TX completion already happened 2591f9616c35SEric Dumazet * before we set TSQ_THROTTLED, so we must 2592f9616c35SEric Dumazet * test again the condition. 2593f9616c35SEric Dumazet */ 2594f9616c35SEric Dumazet smp_mb__after_atomic(); 2595ce8299b6SEric Dumazet if (refcount_read(&sk->sk_wmem_alloc) > limit) 2596f9616c35SEric Dumazet return true; 2597f9616c35SEric Dumazet } 2598f9616c35SEric Dumazet return false; 2599f9616c35SEric Dumazet } 2600f9616c35SEric Dumazet 260105b055e8SFrancis Yan static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new) 260205b055e8SFrancis Yan { 2603628174ccSEric Dumazet const u32 now = tcp_jiffies32; 2604efe967cdSArnd Bergmann enum tcp_chrono old = tp->chrono_type; 260505b055e8SFrancis Yan 2606efe967cdSArnd Bergmann if (old > TCP_CHRONO_UNSPEC) 2607efe967cdSArnd Bergmann tp->chrono_stat[old - 1] += now - tp->chrono_start; 260805b055e8SFrancis Yan tp->chrono_start = now; 260905b055e8SFrancis Yan tp->chrono_type = new; 261005b055e8SFrancis Yan } 261105b055e8SFrancis Yan 261205b055e8SFrancis Yan void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type) 261305b055e8SFrancis Yan { 261405b055e8SFrancis Yan struct tcp_sock *tp = tcp_sk(sk); 261505b055e8SFrancis Yan 261605b055e8SFrancis Yan /* If there are multiple conditions worthy of tracking in a 26170f87230dSFrancis Yan * chronograph then the highest priority enum takes precedence 26180f87230dSFrancis Yan * over the other conditions. So that if something "more interesting" 261905b055e8SFrancis Yan * starts happening, stop the previous chrono and start a new one. 262005b055e8SFrancis Yan */ 262105b055e8SFrancis Yan if (type > tp->chrono_type) 262205b055e8SFrancis Yan tcp_chrono_set(tp, type); 262305b055e8SFrancis Yan } 262405b055e8SFrancis Yan 262505b055e8SFrancis Yan void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type) 262605b055e8SFrancis Yan { 262705b055e8SFrancis Yan struct tcp_sock *tp = tcp_sk(sk); 262805b055e8SFrancis Yan 26290f87230dSFrancis Yan 26300f87230dSFrancis Yan /* There are multiple conditions worthy of tracking in a 26310f87230dSFrancis Yan * chronograph, so that the highest priority enum takes 26320f87230dSFrancis Yan * precedence over the other conditions (see tcp_chrono_start). 26330f87230dSFrancis Yan * If a condition stops, we only stop chrono tracking if 26340f87230dSFrancis Yan * it's the "most interesting" or current chrono we are 26350f87230dSFrancis Yan * tracking and starts busy chrono if we have pending data. 26360f87230dSFrancis Yan */ 263775c119afSEric Dumazet if (tcp_rtx_and_write_queues_empty(sk)) 263805b055e8SFrancis Yan tcp_chrono_set(tp, TCP_CHRONO_UNSPEC); 26390f87230dSFrancis Yan else if (type == tp->chrono_type) 26400f87230dSFrancis Yan tcp_chrono_set(tp, TCP_CHRONO_BUSY); 264105b055e8SFrancis Yan } 264205b055e8SFrancis Yan 26431da177e4SLinus Torvalds /* This routine writes packets to the network. It advances the 26441da177e4SLinus Torvalds * send_head. This happens as incoming acks open up the remote 26451da177e4SLinus Torvalds * window for us. 26461da177e4SLinus Torvalds * 2647f8269a49SIlpo Järvinen * LARGESEND note: !tcp_urg_mode is overkill, only frames between 2648f8269a49SIlpo Järvinen * snd_up-64k-mss .. snd_up cannot be large. However, taking into 2649f8269a49SIlpo Järvinen * account rare use of URG, this is not a big flaw. 2650f8269a49SIlpo Järvinen * 26516ba8a3b1SNandita Dukkipati * Send at most one packet when push_one > 0. Temporarily ignore 26526ba8a3b1SNandita Dukkipati * cwnd limit to force at most one packet out when push_one == 2. 26536ba8a3b1SNandita Dukkipati 2654a2a385d6SEric Dumazet * Returns true, if no segments are in flight and we have queued segments, 2655a2a385d6SEric Dumazet * but cannot send anything now because of SWS or another problem. 26561da177e4SLinus Torvalds */ 2657a2a385d6SEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 2658d5dd9175SIlpo Järvinen int push_one, gfp_t gfp) 26591da177e4SLinus Torvalds { 26601da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 266192df7b51SDavid S. Miller struct sk_buff *skb; 2662c1b4a7e6SDavid S. Miller unsigned int tso_segs, sent_pkts; 2663c1b4a7e6SDavid S. Miller int cwnd_quota; 26645d424d5aSJohn Heffner int result; 26655615f886SFrancis Yan bool is_cwnd_limited = false, is_rwnd_limited = false; 2666605ad7f1SEric Dumazet u32 max_segs; 26671da177e4SLinus Torvalds 2668c1b4a7e6SDavid S. Miller sent_pkts = 0; 26695d424d5aSJohn Heffner 2670ee1836aeSEric Dumazet tcp_mstamp_refresh(tp); 2671d5dd9175SIlpo Järvinen if (!push_one) { 26725d424d5aSJohn Heffner /* Do MTU probing. */ 2673d5dd9175SIlpo Järvinen result = tcp_mtu_probe(sk); 2674d5dd9175SIlpo Järvinen if (!result) { 2675a2a385d6SEric Dumazet return false; 26765d424d5aSJohn Heffner } else if (result > 0) { 26775d424d5aSJohn Heffner sent_pkts = 1; 26785d424d5aSJohn Heffner } 2679d5dd9175SIlpo Järvinen } 26805d424d5aSJohn Heffner 2681ed6e7268SNeal Cardwell max_segs = tcp_tso_segs(sk, mss_now); 2682fe067e8aSDavid S. Miller while ((skb = tcp_send_head(sk))) { 2683c8ac3774SHerbert Xu unsigned int limit; 2684c8ac3774SHerbert Xu 268579861919SEric Dumazet if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) { 268679861919SEric Dumazet /* "skb_mstamp_ns" is used as a start point for the retransmit timer */ 2687a1ac9c8aSMartin KaFai Lau tp->tcp_wstamp_ns = tp->tcp_clock_cache; 2688a1ac9c8aSMartin KaFai Lau skb_set_delivery_time(skb, tp->tcp_wstamp_ns, true); 268979861919SEric Dumazet list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue); 2690bf50b606SEric Dumazet tcp_init_tso_segs(skb, mss_now); 269179861919SEric Dumazet goto repair; /* Skip network transmission */ 269279861919SEric Dumazet } 269379861919SEric Dumazet 2694218af599SEric Dumazet if (tcp_pacing_check(sk)) 2695218af599SEric Dumazet break; 2696218af599SEric Dumazet 26975bbb432cSEric Dumazet tso_segs = tcp_init_tso_segs(skb, mss_now); 2698c1b4a7e6SDavid S. Miller BUG_ON(!tso_segs); 2699c1b4a7e6SDavid S. Miller 2700b68e9f85SHerbert Xu cwnd_quota = tcp_cwnd_test(tp, skb); 27016ba8a3b1SNandita Dukkipati if (!cwnd_quota) { 27026ba8a3b1SNandita Dukkipati if (push_one == 2) 27036ba8a3b1SNandita Dukkipati /* Force out a loss probe pkt. */ 27046ba8a3b1SNandita Dukkipati cwnd_quota = 1; 27056ba8a3b1SNandita Dukkipati else 2706b68e9f85SHerbert Xu break; 27076ba8a3b1SNandita Dukkipati } 2708b68e9f85SHerbert Xu 27095615f886SFrancis Yan if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) { 27105615f886SFrancis Yan is_rwnd_limited = true; 2711b68e9f85SHerbert Xu break; 27125615f886SFrancis Yan } 2713b68e9f85SHerbert Xu 2714d6a4e26aSEric Dumazet if (tso_segs == 1) { 2715aa93466bSDavid S. Miller if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 2716aa93466bSDavid S. Miller (tcp_skb_is_last(sk, skb) ? 2717aa93466bSDavid S. Miller nonagle : TCP_NAGLE_PUSH)))) 2718aa93466bSDavid S. Miller break; 2719c1b4a7e6SDavid S. Miller } else { 2720ca8a2263SNeal Cardwell if (!push_one && 2721605ad7f1SEric Dumazet tcp_tso_should_defer(sk, skb, &is_cwnd_limited, 2722f9bfe4e6SEric Dumazet &is_rwnd_limited, max_segs)) 2723aa93466bSDavid S. Miller break; 2724c1b4a7e6SDavid S. Miller } 2725aa93466bSDavid S. Miller 2726605ad7f1SEric Dumazet limit = mss_now; 2727d6a4e26aSEric Dumazet if (tso_segs > 1 && !tcp_urg_mode(tp)) 2728605ad7f1SEric Dumazet limit = tcp_mss_split_point(sk, skb, mss_now, 2729605ad7f1SEric Dumazet min_t(unsigned int, 2730605ad7f1SEric Dumazet cwnd_quota, 2731605ad7f1SEric Dumazet max_segs), 2732605ad7f1SEric Dumazet nonagle); 2733605ad7f1SEric Dumazet 2734605ad7f1SEric Dumazet if (skb->len > limit && 273556483341SEric Dumazet unlikely(tso_fragment(sk, skb, limit, mss_now, gfp))) 2736605ad7f1SEric Dumazet break; 2737605ad7f1SEric Dumazet 2738f9616c35SEric Dumazet if (tcp_small_queue_check(sk, skb, 0)) 273946d3ceabSEric Dumazet break; 2740c9eeec26SEric Dumazet 27411f85e626SEric Dumazet /* Argh, we hit an empty skb(), presumably a thread 27421f85e626SEric Dumazet * is sleeping in sendmsg()/sk_stream_wait_memory(). 27431f85e626SEric Dumazet * We do not want to send a pure-ack packet and have 27441f85e626SEric Dumazet * a strange looking rtx queue with empty packet(s). 27451f85e626SEric Dumazet */ 27461f85e626SEric Dumazet if (TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq) 27471f85e626SEric Dumazet break; 27481f85e626SEric Dumazet 2749d5dd9175SIlpo Järvinen if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) 27501da177e4SLinus Torvalds break; 27511da177e4SLinus Torvalds 2752ec342325SAndrew Vagin repair: 27531da177e4SLinus Torvalds /* Advance the send_head. This one is sent out. 27541da177e4SLinus Torvalds * This call will increment packets_out. 27551da177e4SLinus Torvalds */ 275666f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 27571da177e4SLinus Torvalds 27581da177e4SLinus Torvalds tcp_minshall_update(tp, mss_now, skb); 2759a262f0cdSNandita Dukkipati sent_pkts += tcp_skb_pcount(skb); 2760d5dd9175SIlpo Järvinen 2761d5dd9175SIlpo Järvinen if (push_one) 2762d5dd9175SIlpo Järvinen break; 27631da177e4SLinus Torvalds } 27641da177e4SLinus Torvalds 27655615f886SFrancis Yan if (is_rwnd_limited) 27665615f886SFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED); 27675615f886SFrancis Yan else 27685615f886SFrancis Yan tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED); 27695615f886SFrancis Yan 277040570375SEric Dumazet is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tcp_snd_cwnd(tp)); 2771299bcb55SNeal Cardwell if (likely(sent_pkts || is_cwnd_limited)) 2772299bcb55SNeal Cardwell tcp_cwnd_validate(sk, is_cwnd_limited); 2773299bcb55SNeal Cardwell 2774aa93466bSDavid S. Miller if (likely(sent_pkts)) { 2775684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 2776684bad11SYuchung Cheng tp->prr_out += sent_pkts; 27776ba8a3b1SNandita Dukkipati 27786ba8a3b1SNandita Dukkipati /* Send one loss probe per tail loss episode. */ 27796ba8a3b1SNandita Dukkipati if (push_one != 2) 2780ed66dfafSNeal Cardwell tcp_schedule_loss_probe(sk, false); 2781a2a385d6SEric Dumazet return false; 27821da177e4SLinus Torvalds } 278375c119afSEric Dumazet return !tp->packets_out && !tcp_write_queue_empty(sk); 27846ba8a3b1SNandita Dukkipati } 27856ba8a3b1SNandita Dukkipati 2786ed66dfafSNeal Cardwell bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto) 27876ba8a3b1SNandita Dukkipati { 27886ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 27896ba8a3b1SNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 2790a2815817SNeal Cardwell u32 timeout, rto_delta_us; 27912ae21cf5SEric Dumazet int early_retrans; 27926ba8a3b1SNandita Dukkipati 27936ba8a3b1SNandita Dukkipati /* Don't do any loss probe on a Fast Open connection before 3WHS 27946ba8a3b1SNandita Dukkipati * finishes. 27956ba8a3b1SNandita Dukkipati */ 2796d983ea6fSEric Dumazet if (rcu_access_pointer(tp->fastopen_rsk)) 27976ba8a3b1SNandita Dukkipati return false; 27986ba8a3b1SNandita Dukkipati 279952e65865SKuniyuki Iwashima early_retrans = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_early_retrans); 28006ba8a3b1SNandita Dukkipati /* Schedule a loss probe in 2*RTT for SACK capable connections 2801b4f70c3dSNeal Cardwell * not in loss recovery, that are either limited by cwnd or application. 28026ba8a3b1SNandita Dukkipati */ 28032ae21cf5SEric Dumazet if ((early_retrans != 3 && early_retrans != 4) || 2804bec41a11SYuchung Cheng !tp->packets_out || !tcp_is_sack(tp) || 2805b4f70c3dSNeal Cardwell (icsk->icsk_ca_state != TCP_CA_Open && 2806b4f70c3dSNeal Cardwell icsk->icsk_ca_state != TCP_CA_CWR)) 28076ba8a3b1SNandita Dukkipati return false; 28086ba8a3b1SNandita Dukkipati 2809bb4d991aSYuchung Cheng /* Probe timeout is 2*rtt. Add minimum RTO to account 2810f9b99582SYuchung Cheng * for delayed ack when there's one outstanding packet. If no RTT 2811f9b99582SYuchung Cheng * sample is available then probe after TCP_TIMEOUT_INIT. 28126ba8a3b1SNandita Dukkipati */ 2813bb4d991aSYuchung Cheng if (tp->srtt_us) { 2814bb4d991aSYuchung Cheng timeout = usecs_to_jiffies(tp->srtt_us >> 2); 28156ba8a3b1SNandita Dukkipati if (tp->packets_out == 1) 2816bb4d991aSYuchung Cheng timeout += TCP_RTO_MIN; 2817bb4d991aSYuchung Cheng else 2818bb4d991aSYuchung Cheng timeout += TCP_TIMEOUT_MIN; 2819bb4d991aSYuchung Cheng } else { 2820bb4d991aSYuchung Cheng timeout = TCP_TIMEOUT_INIT; 2821bb4d991aSYuchung Cheng } 28226ba8a3b1SNandita Dukkipati 2823a2815817SNeal Cardwell /* If the RTO formula yields an earlier time, then use that time. */ 2824ed66dfafSNeal Cardwell rto_delta_us = advancing_rto ? 2825ed66dfafSNeal Cardwell jiffies_to_usecs(inet_csk(sk)->icsk_rto) : 2826ed66dfafSNeal Cardwell tcp_rto_delta_us(sk); /* How far in future is RTO? */ 2827a2815817SNeal Cardwell if (rto_delta_us > 0) 2828a2815817SNeal Cardwell timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us)); 28296ba8a3b1SNandita Dukkipati 28308dc242adSEric Dumazet tcp_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, TCP_RTO_MAX); 28316ba8a3b1SNandita Dukkipati return true; 28326ba8a3b1SNandita Dukkipati } 28336ba8a3b1SNandita Dukkipati 28341f3279aeSEric Dumazet /* Thanks to skb fast clones, we can detect if a prior transmit of 28351f3279aeSEric Dumazet * a packet is still in a qdisc or driver queue. 28361f3279aeSEric Dumazet * In this case, there is very little point doing a retransmit ! 28371f3279aeSEric Dumazet */ 2838f4dae54eSEric Dumazet static bool skb_still_in_host_queue(struct sock *sk, 28391f3279aeSEric Dumazet const struct sk_buff *skb) 28401f3279aeSEric Dumazet { 284139bb5e62SEric Dumazet if (unlikely(skb_fclone_busy(sk, skb))) { 2842f4dae54eSEric Dumazet set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags); 2843f4dae54eSEric Dumazet smp_mb__after_atomic(); 2844f4dae54eSEric Dumazet if (skb_fclone_busy(sk, skb)) { 2845c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), 28461f3279aeSEric Dumazet LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES); 28471f3279aeSEric Dumazet return true; 28481f3279aeSEric Dumazet } 2849f4dae54eSEric Dumazet } 28501f3279aeSEric Dumazet return false; 28511f3279aeSEric Dumazet } 28521f3279aeSEric Dumazet 2853b340b264SYuchung Cheng /* When probe timeout (PTO) fires, try send a new segment if possible, else 28546ba8a3b1SNandita Dukkipati * retransmit the last segment. 28556ba8a3b1SNandita Dukkipati */ 28566ba8a3b1SNandita Dukkipati void tcp_send_loss_probe(struct sock *sk) 28576ba8a3b1SNandita Dukkipati { 28589b717a8dSNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 28596ba8a3b1SNandita Dukkipati struct sk_buff *skb; 28606ba8a3b1SNandita Dukkipati int pcount; 28616ba8a3b1SNandita Dukkipati int mss = tcp_current_mss(sk); 28626ba8a3b1SNandita Dukkipati 286376be93fcSYuchung Cheng /* At most one outstanding TLP */ 286476be93fcSYuchung Cheng if (tp->tlp_high_seq) 286576be93fcSYuchung Cheng goto rearm_timer; 286676be93fcSYuchung Cheng 286776be93fcSYuchung Cheng tp->tlp_retrans = 0; 2868b340b264SYuchung Cheng skb = tcp_send_head(sk); 286975c119afSEric Dumazet if (skb && tcp_snd_wnd_test(tp, skb, mss)) { 2870b340b264SYuchung Cheng pcount = tp->packets_out; 2871b340b264SYuchung Cheng tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC); 2872b340b264SYuchung Cheng if (tp->packets_out > pcount) 2873b340b264SYuchung Cheng goto probe_sent; 28746ba8a3b1SNandita Dukkipati goto rearm_timer; 28756ba8a3b1SNandita Dukkipati } 287675c119afSEric Dumazet skb = skb_rb_last(&sk->tcp_rtx_queue); 2877b2b7af86SYuchung Cheng if (unlikely(!skb)) { 2878b2b7af86SYuchung Cheng WARN_ONCE(tp->packets_out, 2879b2b7af86SYuchung Cheng "invalid inflight: %u state %u cwnd %u mss %d\n", 288040570375SEric Dumazet tp->packets_out, sk->sk_state, tcp_snd_cwnd(tp), mss); 2881b2b7af86SYuchung Cheng inet_csk(sk)->icsk_pending = 0; 2882b2b7af86SYuchung Cheng return; 2883b2b7af86SYuchung Cheng } 28846ba8a3b1SNandita Dukkipati 28851f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 28861f3279aeSEric Dumazet goto rearm_timer; 28871f3279aeSEric Dumazet 28886ba8a3b1SNandita Dukkipati pcount = tcp_skb_pcount(skb); 28896ba8a3b1SNandita Dukkipati if (WARN_ON(!pcount)) 28906ba8a3b1SNandita Dukkipati goto rearm_timer; 28916ba8a3b1SNandita Dukkipati 28926ba8a3b1SNandita Dukkipati if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) { 289375c119afSEric Dumazet if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, 289475c119afSEric Dumazet (pcount - 1) * mss, mss, 28956cc55e09SOctavian Purdila GFP_ATOMIC))) 28966ba8a3b1SNandita Dukkipati goto rearm_timer; 289775c119afSEric Dumazet skb = skb_rb_next(skb); 28986ba8a3b1SNandita Dukkipati } 28996ba8a3b1SNandita Dukkipati 29006ba8a3b1SNandita Dukkipati if (WARN_ON(!skb || !tcp_skb_pcount(skb))) 29016ba8a3b1SNandita Dukkipati goto rearm_timer; 29026ba8a3b1SNandita Dukkipati 290310d3be56SEric Dumazet if (__tcp_retransmit_skb(sk, skb, 1)) 2904b340b264SYuchung Cheng goto rearm_timer; 29056ba8a3b1SNandita Dukkipati 290676be93fcSYuchung Cheng tp->tlp_retrans = 1; 290776be93fcSYuchung Cheng 290876be93fcSYuchung Cheng probe_sent: 29099b717a8dSNandita Dukkipati /* Record snd_nxt for loss detection. */ 29109b717a8dSNandita Dukkipati tp->tlp_high_seq = tp->snd_nxt; 29119b717a8dSNandita Dukkipati 2912c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES); 2913fcd16c0aSYuchung Cheng /* Reset s.t. tcp_rearm_rto will restart timer from now */ 2914fcd16c0aSYuchung Cheng inet_csk(sk)->icsk_pending = 0; 2915b340b264SYuchung Cheng rearm_timer: 2916fcd16c0aSYuchung Cheng tcp_rearm_rto(sk); 29171da177e4SLinus Torvalds } 29181da177e4SLinus Torvalds 2919a762a980SDavid S. Miller /* Push out any pending frames which were held back due to 2920a762a980SDavid S. Miller * TCP_CORK or attempt at coalescing tiny packets. 2921a762a980SDavid S. Miller * The socket must be locked by the caller. 2922a762a980SDavid S. Miller */ 29239e412ba7SIlpo Järvinen void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 29249e412ba7SIlpo Järvinen int nonagle) 2925a762a980SDavid S. Miller { 2926726e07a8SIlpo Järvinen /* If we are closed, the bytes will have to remain here. 2927726e07a8SIlpo Järvinen * In time closedown will finish, we empty the write queue and 2928726e07a8SIlpo Järvinen * all will be happy. 2929726e07a8SIlpo Järvinen */ 2930726e07a8SIlpo Järvinen if (unlikely(sk->sk_state == TCP_CLOSE)) 2931726e07a8SIlpo Järvinen return; 2932726e07a8SIlpo Järvinen 293399a1dec7SMel Gorman if (tcp_write_xmit(sk, cur_mss, nonagle, 0, 29347450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC))) 29359e412ba7SIlpo Järvinen tcp_check_probe_timer(sk); 2936a762a980SDavid S. Miller } 2937a762a980SDavid S. Miller 2938c1b4a7e6SDavid S. Miller /* Send _single_ skb sitting at the send head. This function requires 2939c1b4a7e6SDavid S. Miller * true push pending frames to setup probe timer etc. 2940c1b4a7e6SDavid S. Miller */ 2941c1b4a7e6SDavid S. Miller void tcp_push_one(struct sock *sk, unsigned int mss_now) 2942c1b4a7e6SDavid S. Miller { 2943fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 2944c1b4a7e6SDavid S. Miller 2945c1b4a7e6SDavid S. Miller BUG_ON(!skb || skb->len < mss_now); 2946c1b4a7e6SDavid S. Miller 2947d5dd9175SIlpo Järvinen tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); 2948c1b4a7e6SDavid S. Miller } 2949c1b4a7e6SDavid S. Miller 29501da177e4SLinus Torvalds /* This function returns the amount that we can raise the 29511da177e4SLinus Torvalds * usable window based on the following constraints 29521da177e4SLinus Torvalds * 29531da177e4SLinus Torvalds * 1. The window can never be shrunk once it is offered (RFC 793) 29541da177e4SLinus Torvalds * 2. We limit memory per socket 29551da177e4SLinus Torvalds * 29561da177e4SLinus Torvalds * RFC 1122: 29571da177e4SLinus Torvalds * "the suggested [SWS] avoidance algorithm for the receiver is to keep 29581da177e4SLinus Torvalds * RECV.NEXT + RCV.WIN fixed until: 29591da177e4SLinus Torvalds * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 29601da177e4SLinus Torvalds * 29611da177e4SLinus Torvalds * i.e. don't raise the right edge of the window until you can raise 29621da177e4SLinus Torvalds * it at least MSS bytes. 29631da177e4SLinus Torvalds * 29641da177e4SLinus Torvalds * Unfortunately, the recommended algorithm breaks header prediction, 29651da177e4SLinus Torvalds * since header prediction assumes th->window stays fixed. 29661da177e4SLinus Torvalds * 29671da177e4SLinus Torvalds * Strictly speaking, keeping th->window fixed violates the receiver 29681da177e4SLinus Torvalds * side SWS prevention criteria. The problem is that under this rule 29691da177e4SLinus Torvalds * a stream of single byte packets will cause the right side of the 29701da177e4SLinus Torvalds * window to always advance by a single byte. 29711da177e4SLinus Torvalds * 29721da177e4SLinus Torvalds * Of course, if the sender implements sender side SWS prevention 29731da177e4SLinus Torvalds * then this will not be a problem. 29741da177e4SLinus Torvalds * 29751da177e4SLinus Torvalds * BSD seems to make the following compromise: 29761da177e4SLinus Torvalds * 29771da177e4SLinus Torvalds * If the free space is less than the 1/4 of the maximum 29781da177e4SLinus Torvalds * space available and the free space is less than 1/2 mss, 29791da177e4SLinus Torvalds * then set the window to 0. 29801da177e4SLinus Torvalds * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 29811da177e4SLinus Torvalds * Otherwise, just prevent the window from shrinking 29821da177e4SLinus Torvalds * and from being larger than the largest representable value. 29831da177e4SLinus Torvalds * 29841da177e4SLinus Torvalds * This prevents incremental opening of the window in the regime 29851da177e4SLinus Torvalds * where TCP is limited by the speed of the reader side taking 29861da177e4SLinus Torvalds * data out of the TCP receive queue. It does nothing about 29871da177e4SLinus Torvalds * those cases where the window is constrained on the sender side 29881da177e4SLinus Torvalds * because the pipeline is full. 29891da177e4SLinus Torvalds * 29901da177e4SLinus Torvalds * BSD also seems to "accidentally" limit itself to windows that are a 29911da177e4SLinus Torvalds * multiple of MSS, at least until the free space gets quite small. 29921da177e4SLinus Torvalds * This would appear to be a side effect of the mbuf implementation. 29931da177e4SLinus Torvalds * Combining these two algorithms results in the observed behavior 29941da177e4SLinus Torvalds * of having a fixed window size at almost all times. 29951da177e4SLinus Torvalds * 29961da177e4SLinus Torvalds * Below we obtain similar behavior by forcing the offered window to 29971da177e4SLinus Torvalds * a multiple of the mss when it is feasible to do so. 29981da177e4SLinus Torvalds * 29991da177e4SLinus Torvalds * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 30001da177e4SLinus Torvalds * Regular options like TIMESTAMP are taken into account. 30011da177e4SLinus Torvalds */ 30021da177e4SLinus Torvalds u32 __tcp_select_window(struct sock *sk) 30031da177e4SLinus Torvalds { 3004463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 30051da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 3006caa20d9aSStephen Hemminger /* MSS for the peer's data. Previous versions used mss_clamp 30071da177e4SLinus Torvalds * here. I don't know if the value based on our guesses 30081da177e4SLinus Torvalds * of peer's MSS is better for the performance. It's more correct 30091da177e4SLinus Torvalds * but may be worse for the performance because of rcv_mss 30101da177e4SLinus Torvalds * fluctuations. --SAW 1998/11/1 30111da177e4SLinus Torvalds */ 3012463c84b9SArnaldo Carvalho de Melo int mss = icsk->icsk_ack.rcv_mss; 30131da177e4SLinus Torvalds int free_space = tcp_space(sk); 301486c1a045SFlorian Westphal int allowed_space = tcp_full_space(sk); 3015071c8ed6SFlorian Westphal int full_space, window; 3016071c8ed6SFlorian Westphal 3017071c8ed6SFlorian Westphal if (sk_is_mptcp(sk)) 3018071c8ed6SFlorian Westphal mptcp_space(sk, &free_space, &allowed_space); 3019071c8ed6SFlorian Westphal 3020071c8ed6SFlorian Westphal full_space = min_t(int, tp->window_clamp, allowed_space); 30211da177e4SLinus Torvalds 302206425c30SEric Dumazet if (unlikely(mss > full_space)) { 30231da177e4SLinus Torvalds mss = full_space; 302406425c30SEric Dumazet if (mss <= 0) 302506425c30SEric Dumazet return 0; 302606425c30SEric Dumazet } 3027b92edbe0SEric Dumazet if (free_space < (full_space >> 1)) { 3028463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.quick = 0; 30291da177e4SLinus Torvalds 3030b8da51ebSEric Dumazet if (tcp_under_memory_pressure(sk)) 3031053f3684SWei Wang tcp_adjust_rcv_ssthresh(sk); 30321da177e4SLinus Torvalds 303386c1a045SFlorian Westphal /* free_space might become our new window, make sure we don't 303486c1a045SFlorian Westphal * increase it due to wscale. 303586c1a045SFlorian Westphal */ 303686c1a045SFlorian Westphal free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale); 303786c1a045SFlorian Westphal 303886c1a045SFlorian Westphal /* if free space is less than mss estimate, or is below 1/16th 303986c1a045SFlorian Westphal * of the maximum allowed, try to move to zero-window, else 304086c1a045SFlorian Westphal * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and 304186c1a045SFlorian Westphal * new incoming data is dropped due to memory limits. 304286c1a045SFlorian Westphal * With large window, mss test triggers way too late in order 304386c1a045SFlorian Westphal * to announce zero window in time before rmem limit kicks in. 304486c1a045SFlorian Westphal */ 304586c1a045SFlorian Westphal if (free_space < (allowed_space >> 4) || free_space < mss) 30461da177e4SLinus Torvalds return 0; 30471da177e4SLinus Torvalds } 30481da177e4SLinus Torvalds 30491da177e4SLinus Torvalds if (free_space > tp->rcv_ssthresh) 30501da177e4SLinus Torvalds free_space = tp->rcv_ssthresh; 30511da177e4SLinus Torvalds 30521da177e4SLinus Torvalds /* Don't do rounding if we are using window scaling, since the 30531da177e4SLinus Torvalds * scaled window will not line up with the MSS boundary anyway. 30541da177e4SLinus Torvalds */ 30551da177e4SLinus Torvalds if (tp->rx_opt.rcv_wscale) { 30561da177e4SLinus Torvalds window = free_space; 30571da177e4SLinus Torvalds 30581da177e4SLinus Torvalds /* Advertise enough space so that it won't get scaled away. 30591da177e4SLinus Torvalds * Import case: prevent zero window announcement if 30601da177e4SLinus Torvalds * 1<<rcv_wscale > mss. 30611da177e4SLinus Torvalds */ 30621935299dSGao Feng window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale)); 30631da177e4SLinus Torvalds } else { 30641935299dSGao Feng window = tp->rcv_wnd; 30651da177e4SLinus Torvalds /* Get the largest window that is a nice multiple of mss. 30661da177e4SLinus Torvalds * Window clamp already applied above. 30671da177e4SLinus Torvalds * If our current window offering is within 1 mss of the 30681da177e4SLinus Torvalds * free space we just keep it. This prevents the divide 30691da177e4SLinus Torvalds * and multiply from happening most of the time. 30701da177e4SLinus Torvalds * We also don't do any window rounding when the free space 30711da177e4SLinus Torvalds * is too small. 30721da177e4SLinus Torvalds */ 30731da177e4SLinus Torvalds if (window <= free_space - mss || window > free_space) 30741935299dSGao Feng window = rounddown(free_space, mss); 307584565070SJohn Heffner else if (mss == full_space && 3076b92edbe0SEric Dumazet free_space > window + (full_space >> 1)) 307784565070SJohn Heffner window = free_space; 30781da177e4SLinus Torvalds } 30791da177e4SLinus Torvalds 30801da177e4SLinus Torvalds return window; 30811da177e4SLinus Torvalds } 30821da177e4SLinus Torvalds 3083cfea5a68SMartin KaFai Lau void tcp_skb_collapse_tstamp(struct sk_buff *skb, 3084082ac2d5SMartin KaFai Lau const struct sk_buff *next_skb) 3085082ac2d5SMartin KaFai Lau { 30860a2cf20cSSoheil Hassas Yeganeh if (unlikely(tcp_has_tx_tstamp(next_skb))) { 30870a2cf20cSSoheil Hassas Yeganeh const struct skb_shared_info *next_shinfo = 30880a2cf20cSSoheil Hassas Yeganeh skb_shinfo(next_skb); 3089082ac2d5SMartin KaFai Lau struct skb_shared_info *shinfo = skb_shinfo(skb); 3090082ac2d5SMartin KaFai Lau 30910a2cf20cSSoheil Hassas Yeganeh shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP; 3092082ac2d5SMartin KaFai Lau shinfo->tskey = next_shinfo->tskey; 30932de8023eSMartin KaFai Lau TCP_SKB_CB(skb)->txstamp_ack |= 30942de8023eSMartin KaFai Lau TCP_SKB_CB(next_skb)->txstamp_ack; 3095082ac2d5SMartin KaFai Lau } 3096082ac2d5SMartin KaFai Lau } 3097082ac2d5SMartin KaFai Lau 30984a17fc3aSIlpo Järvinen /* Collapses two adjacent SKB's during retransmission. */ 3099f8071cdeSEric Dumazet static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) 31001da177e4SLinus Torvalds { 31011da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 310275c119afSEric Dumazet struct sk_buff *next_skb = skb_rb_next(skb); 310313dde04fSWei Yongjun int next_skb_size; 31041da177e4SLinus Torvalds 3105058dc334SIlpo Järvinen next_skb_size = next_skb->len; 31061da177e4SLinus Torvalds 3107058dc334SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); 31081da177e4SLinus Torvalds 3109bd446314SEric Dumazet if (next_skb_size && !tcp_skb_shift(skb, next_skb, 1, next_skb_size)) 3110f8071cdeSEric Dumazet return false; 3111bd446314SEric Dumazet 31122b7cda9cSEric Dumazet tcp_highest_sack_replace(sk, next_skb, skb); 3113a6963a6bSIlpo Järvinen 31141da177e4SLinus Torvalds /* Update sequence range on original skb. */ 31151da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 31161da177e4SLinus Torvalds 3117e6c7d085SIlpo Järvinen /* Merge over control information. This moves PSH/FIN etc. over */ 31184de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; 31191da177e4SLinus Torvalds 31201da177e4SLinus Torvalds /* All done, get rid of second SKB and account for it so 31211da177e4SLinus Torvalds * packet counting does not break. 31221da177e4SLinus Torvalds */ 31234828e7f4SIlpo Järvinen TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; 3124a643b5d4SMartin KaFai Lau TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor; 3125b7689205SIlpo Järvinen 3126b7689205SIlpo Järvinen /* changed transmit queue under us so clear hints */ 3127ef9da47cSIlpo Järvinen tcp_clear_retrans_hints_partial(tp); 3128ef9da47cSIlpo Järvinen if (next_skb == tp->retransmit_skb_hint) 3129ef9da47cSIlpo Järvinen tp->retransmit_skb_hint = skb; 3130b7689205SIlpo Järvinen 3131797108d1SIlpo Järvinen tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); 3132797108d1SIlpo Järvinen 3133082ac2d5SMartin KaFai Lau tcp_skb_collapse_tstamp(skb, next_skb); 3134082ac2d5SMartin KaFai Lau 313575c119afSEric Dumazet tcp_rtx_queue_unlink_and_free(next_skb, sk); 3136f8071cdeSEric Dumazet return true; 31371da177e4SLinus Torvalds } 31381da177e4SLinus Torvalds 313967edfef7SAndi Kleen /* Check if coalescing SKBs is legal. */ 3140a2a385d6SEric Dumazet static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) 31414a17fc3aSIlpo Järvinen { 31424a17fc3aSIlpo Järvinen if (tcp_skb_pcount(skb) > 1) 3143a2a385d6SEric Dumazet return false; 31444a17fc3aSIlpo Järvinen if (skb_cloned(skb)) 3145a2a385d6SEric Dumazet return false; 31462331ccc5SEric Dumazet /* Some heuristics for collapsing over SACK'd could be invented */ 31474a17fc3aSIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 3148a2a385d6SEric Dumazet return false; 31494a17fc3aSIlpo Järvinen 3150a2a385d6SEric Dumazet return true; 31514a17fc3aSIlpo Järvinen } 31524a17fc3aSIlpo Järvinen 315367edfef7SAndi Kleen /* Collapse packets in the retransmit queue to make to create 315467edfef7SAndi Kleen * less packets on the wire. This is only done on retransmission. 315567edfef7SAndi Kleen */ 31564a17fc3aSIlpo Järvinen static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, 31574a17fc3aSIlpo Järvinen int space) 31584a17fc3aSIlpo Järvinen { 31594a17fc3aSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 31604a17fc3aSIlpo Järvinen struct sk_buff *skb = to, *tmp; 3161a2a385d6SEric Dumazet bool first = true; 31624a17fc3aSIlpo Järvinen 31631a63cb91SKuniyuki Iwashima if (!READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse)) 31644a17fc3aSIlpo Järvinen return; 31654de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 31664a17fc3aSIlpo Järvinen return; 31674a17fc3aSIlpo Järvinen 316875c119afSEric Dumazet skb_rbtree_walk_from_safe(skb, tmp) { 31694a17fc3aSIlpo Järvinen if (!tcp_can_collapse(sk, skb)) 31704a17fc3aSIlpo Järvinen break; 31714a17fc3aSIlpo Järvinen 317285712484SMat Martineau if (!tcp_skb_can_collapse(to, skb)) 3173a643b5d4SMartin KaFai Lau break; 3174a643b5d4SMartin KaFai Lau 31754a17fc3aSIlpo Järvinen space -= skb->len; 31764a17fc3aSIlpo Järvinen 31774a17fc3aSIlpo Järvinen if (first) { 3178a2a385d6SEric Dumazet first = false; 31794a17fc3aSIlpo Järvinen continue; 31804a17fc3aSIlpo Järvinen } 31814a17fc3aSIlpo Järvinen 31824a17fc3aSIlpo Järvinen if (space < 0) 31834a17fc3aSIlpo Järvinen break; 31844a17fc3aSIlpo Järvinen 31854a17fc3aSIlpo Järvinen if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) 31864a17fc3aSIlpo Järvinen break; 31874a17fc3aSIlpo Järvinen 3188f8071cdeSEric Dumazet if (!tcp_collapse_retrans(sk, to)) 3189f8071cdeSEric Dumazet break; 31904a17fc3aSIlpo Järvinen } 31914a17fc3aSIlpo Järvinen } 31924a17fc3aSIlpo Järvinen 31931da177e4SLinus Torvalds /* This retransmits one SKB. Policy decisions and retransmit queue 31941da177e4SLinus Torvalds * state updates are done by the caller. Returns non-zero if an 31951da177e4SLinus Torvalds * error occurred which prevented the send. 31961da177e4SLinus Torvalds */ 319710d3be56SEric Dumazet int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) 31981da177e4SLinus Torvalds { 31995d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 320010d3be56SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 32017d227cd2SSridhar Samudrala unsigned int cur_mss; 320210d3be56SEric Dumazet int diff, len, err; 3203536a6c8eSYonglong Li int avail_wnd; 320410d3be56SEric Dumazet 320510d3be56SEric Dumazet /* Inconclusive MTU probe */ 320610d3be56SEric Dumazet if (icsk->icsk_mtup.probe_size) 32075d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 32085d424d5aSJohn Heffner 32091f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 32101f3279aeSEric Dumazet return -EBUSY; 32111f3279aeSEric Dumazet 32121da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 32137f582b24SEric Dumazet if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) { 32147f582b24SEric Dumazet WARN_ON_ONCE(1); 32157f582b24SEric Dumazet return -EINVAL; 32167f582b24SEric Dumazet } 32171da177e4SLinus Torvalds if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 32181da177e4SLinus Torvalds return -ENOMEM; 32191da177e4SLinus Torvalds } 32201da177e4SLinus Torvalds 32217d227cd2SSridhar Samudrala if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 32227d227cd2SSridhar Samudrala return -EHOSTUNREACH; /* Routing failure or similar. */ 32237d227cd2SSridhar Samudrala 32240c54b85fSIlpo Järvinen cur_mss = tcp_current_mss(sk); 3225536a6c8eSYonglong Li avail_wnd = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 32267d227cd2SSridhar Samudrala 32271da177e4SLinus Torvalds /* If receiver has shrunk his window, and skb is out of 32281da177e4SLinus Torvalds * new window, do not retransmit it. The exception is the 32291da177e4SLinus Torvalds * case, when window is shrunk to zero. In this case 3230536a6c8eSYonglong Li * our retransmit of one segment serves as a zero window probe. 32311da177e4SLinus Torvalds */ 3232536a6c8eSYonglong Li if (avail_wnd <= 0) { 3233536a6c8eSYonglong Li if (TCP_SKB_CB(skb)->seq != tp->snd_una) 32341da177e4SLinus Torvalds return -EAGAIN; 3235536a6c8eSYonglong Li avail_wnd = cur_mss; 3236536a6c8eSYonglong Li } 32371da177e4SLinus Torvalds 323810d3be56SEric Dumazet len = cur_mss * segs; 3239536a6c8eSYonglong Li if (len > avail_wnd) { 3240536a6c8eSYonglong Li len = rounddown(avail_wnd, cur_mss); 3241536a6c8eSYonglong Li if (!len) 3242536a6c8eSYonglong Li len = avail_wnd; 3243536a6c8eSYonglong Li } 324410d3be56SEric Dumazet if (skb->len > len) { 324575c119afSEric Dumazet if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len, 324675c119afSEric Dumazet cur_mss, GFP_ATOMIC)) 32471da177e4SLinus Torvalds return -ENOMEM; /* We'll try again later. */ 324802276f3cSIlpo Järvinen } else { 3249c4777efaSEric Dumazet if (skb_unclone_keeptruesize(skb, GFP_ATOMIC)) 3250c52e2421SEric Dumazet return -ENOMEM; 325110d3be56SEric Dumazet 325210d3be56SEric Dumazet diff = tcp_skb_pcount(skb); 325310d3be56SEric Dumazet tcp_set_skb_tso_segs(skb, cur_mss); 325410d3be56SEric Dumazet diff -= tcp_skb_pcount(skb); 325510d3be56SEric Dumazet if (diff) 325610d3be56SEric Dumazet tcp_adjust_pcount(sk, skb, diff); 3257536a6c8eSYonglong Li avail_wnd = min_t(int, avail_wnd, cur_mss); 3258536a6c8eSYonglong Li if (skb->len < avail_wnd) 3259536a6c8eSYonglong Li tcp_retrans_try_collapse(sk, skb, avail_wnd); 32601da177e4SLinus Torvalds } 32611da177e4SLinus Torvalds 326249213555SDaniel Borkmann /* RFC3168, section 6.1.1.1. ECN fallback */ 326349213555SDaniel Borkmann if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN) 326449213555SDaniel Borkmann tcp_ecn_clear_syn(sk, skb); 326549213555SDaniel Borkmann 3266678550c6SYuchung Cheng /* Update global and local TCP statistics. */ 3267678550c6SYuchung Cheng segs = tcp_skb_pcount(skb); 3268678550c6SYuchung Cheng TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs); 3269678550c6SYuchung Cheng if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 3270678550c6SYuchung Cheng __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 3271678550c6SYuchung Cheng tp->total_retrans += segs; 3272fb31c9b9SWei Wang tp->bytes_retrans += skb->len; 3273678550c6SYuchung Cheng 327450bceae9SThomas Graf /* make sure skb->data is aligned on arches that require it 327550bceae9SThomas Graf * and check if ack-trimming & collapsing extended the headroom 327650bceae9SThomas Graf * beyond what csum_start can cover. 327750bceae9SThomas Graf */ 327850bceae9SThomas Graf if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) || 327950bceae9SThomas Graf skb_headroom(skb) >= 0xFFFF)) { 328010a81980SEric Dumazet struct sk_buff *nskb; 328110a81980SEric Dumazet 3282e2080072SEric Dumazet tcp_skb_tsorted_save(skb) { 328310a81980SEric Dumazet nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC); 328407f8e4d0SFlorian Westphal if (nskb) { 328507f8e4d0SFlorian Westphal nskb->dev = NULL; 328607f8e4d0SFlorian Westphal err = tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC); 328707f8e4d0SFlorian Westphal } else { 328807f8e4d0SFlorian Westphal err = -ENOBUFS; 328907f8e4d0SFlorian Westphal } 3290e2080072SEric Dumazet } tcp_skb_tsorted_restore(skb); 3291e2080072SEric Dumazet 32925889e2c0SYousuk Seung if (!err) { 3293a7a25630SEric Dumazet tcp_update_skb_after_send(sk, skb, tp->tcp_wstamp_ns); 32945889e2c0SYousuk Seung tcp_rate_skb_sent(sk, skb); 32955889e2c0SYousuk Seung } 3296117632e6SEric Dumazet } else { 3297c84a5711SYuchung Cheng err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 3298117632e6SEric Dumazet } 3299c84a5711SYuchung Cheng 33007f12422cSYuchung Cheng /* To avoid taking spuriously low RTT samples based on a timestamp 33017f12422cSYuchung Cheng * for a transmit that never happened, always mark EVER_RETRANS 33027f12422cSYuchung Cheng */ 33037f12422cSYuchung Cheng TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS; 33047f12422cSYuchung Cheng 3305a31ad29eSLawrence Brakmo if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RETRANS_CB_FLAG)) 3306a31ad29eSLawrence Brakmo tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RETRANS_CB, 3307a31ad29eSLawrence Brakmo TCP_SKB_CB(skb)->seq, segs, err); 3308a31ad29eSLawrence Brakmo 3309fc9f3501SEric Dumazet if (likely(!err)) { 3310e086101bSCong Wang trace_tcp_retransmit_skb(sk, skb); 3311678550c6SYuchung Cheng } else if (err != -EBUSY) { 3312ec641b39SYuchung Cheng NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL, segs); 3313fc9f3501SEric Dumazet } 3314c84a5711SYuchung Cheng return err; 331593b174adSYuchung Cheng } 331693b174adSYuchung Cheng 331710d3be56SEric Dumazet int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) 331893b174adSYuchung Cheng { 331993b174adSYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 332010d3be56SEric Dumazet int err = __tcp_retransmit_skb(sk, skb, segs); 33211da177e4SLinus Torvalds 33221da177e4SLinus Torvalds if (err == 0) { 33231da177e4SLinus Torvalds #if FASTRETRANS_DEBUG > 0 33241da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 3325e87cc472SJoe Perches net_dbg_ratelimited("retrans_out leaked\n"); 33261da177e4SLinus Torvalds } 33271da177e4SLinus Torvalds #endif 33281da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 33291da177e4SLinus Torvalds tp->retrans_out += tcp_skb_pcount(skb); 33307ae18975SYuchung Cheng } 33311da177e4SLinus Torvalds 33327ae18975SYuchung Cheng /* Save stamp of the first (attempted) retransmit. */ 33331da177e4SLinus Torvalds if (!tp->retrans_stamp) 33347faee5c0SEric Dumazet tp->retrans_stamp = tcp_skb_timestamp(skb); 33351da177e4SLinus Torvalds 33366e08d5e3SYuchung Cheng if (tp->undo_retrans < 0) 33376e08d5e3SYuchung Cheng tp->undo_retrans = 0; 33386e08d5e3SYuchung Cheng tp->undo_retrans += tcp_skb_pcount(skb); 33391da177e4SLinus Torvalds return err; 33401da177e4SLinus Torvalds } 33411da177e4SLinus Torvalds 33421da177e4SLinus Torvalds /* This gets called after a retransmit timeout, and the initially 33431da177e4SLinus Torvalds * retransmitted data is acknowledged. It tries to continue 33441da177e4SLinus Torvalds * resending the rest of the retransmit queue, until either 33451da177e4SLinus Torvalds * we've sent it all or the congestion window limit is reached. 33461da177e4SLinus Torvalds */ 33471da177e4SLinus Torvalds void tcp_xmit_retransmit_queue(struct sock *sk) 33481da177e4SLinus Torvalds { 33496687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 3350b9f1f1ceSEric Dumazet struct sk_buff *skb, *rtx_head, *hole = NULL; 33511da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 3352916e6d1aSEric Dumazet bool rearm_timer = false; 3353840a3cbeSYuchung Cheng u32 max_segs; 335461eb55f4SIlpo Järvinen int mib_idx; 33556a438bbeSStephen Hemminger 335645e77d31SIlpo Järvinen if (!tp->packets_out) 335745e77d31SIlpo Järvinen return; 335845e77d31SIlpo Järvinen 335975c119afSEric Dumazet rtx_head = tcp_rtx_queue_head(sk); 3360b9f1f1ceSEric Dumazet skb = tp->retransmit_skb_hint ?: rtx_head; 3361ed6e7268SNeal Cardwell max_segs = tcp_tso_segs(sk, tcp_current_mss(sk)); 336275c119afSEric Dumazet skb_rbtree_walk_from(skb) { 3363dca0aaf8SEric Dumazet __u8 sacked; 336410d3be56SEric Dumazet int segs; 33651da177e4SLinus Torvalds 3366218af599SEric Dumazet if (tcp_pacing_check(sk)) 3367218af599SEric Dumazet break; 3368218af599SEric Dumazet 33696a438bbeSStephen Hemminger /* we could do better than to assign each time */ 337051456b29SIan Morris if (!hole) 33716a438bbeSStephen Hemminger tp->retransmit_skb_hint = skb; 33726a438bbeSStephen Hemminger 337340570375SEric Dumazet segs = tcp_snd_cwnd(tp) - tcp_packets_in_flight(tp); 337410d3be56SEric Dumazet if (segs <= 0) 3375916e6d1aSEric Dumazet break; 3376dca0aaf8SEric Dumazet sacked = TCP_SKB_CB(skb)->sacked; 3377a3d2e9f8SEric Dumazet /* In case tcp_shift_skb_data() have aggregated large skbs, 3378a3d2e9f8SEric Dumazet * we need to make sure not sending too bigs TSO packets 3379a3d2e9f8SEric Dumazet */ 3380a3d2e9f8SEric Dumazet segs = min_t(int, segs, max_segs); 33810e1c54c2SIlpo Järvinen 3382840a3cbeSYuchung Cheng if (tp->retrans_out >= tp->lost_out) { 3383006f582cSIlpo Järvinen break; 33840e1c54c2SIlpo Järvinen } else if (!(sacked & TCPCB_LOST)) { 338551456b29SIan Morris if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) 33860e1c54c2SIlpo Järvinen hole = skb; 338761eb55f4SIlpo Järvinen continue; 33881da177e4SLinus Torvalds 33890e1c54c2SIlpo Järvinen } else { 33900e1c54c2SIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Loss) 33910e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFASTRETRANS; 33920e1c54c2SIlpo Järvinen else 33930e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; 33940e1c54c2SIlpo Järvinen } 33950e1c54c2SIlpo Järvinen 33960e1c54c2SIlpo Järvinen if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) 339761eb55f4SIlpo Järvinen continue; 339840b215e5SPavel Emelyanov 3399f9616c35SEric Dumazet if (tcp_small_queue_check(sk, skb, 1)) 3400916e6d1aSEric Dumazet break; 3401f9616c35SEric Dumazet 340210d3be56SEric Dumazet if (tcp_retransmit_skb(sk, skb, segs)) 3403916e6d1aSEric Dumazet break; 340424ab6becSYuchung Cheng 3405de1d6578SYuchung Cheng NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb)); 34061da177e4SLinus Torvalds 3407684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 3408a262f0cdSNandita Dukkipati tp->prr_out += tcp_skb_pcount(skb); 3409a262f0cdSNandita Dukkipati 341075c119afSEric Dumazet if (skb == rtx_head && 341157dde7f7SYuchung Cheng icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT) 3412916e6d1aSEric Dumazet rearm_timer = true; 3413916e6d1aSEric Dumazet 3414916e6d1aSEric Dumazet } 3415916e6d1aSEric Dumazet if (rearm_timer) 34163f80e08fSEric Dumazet tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 34173f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, 34188dc242adSEric Dumazet TCP_RTO_MAX); 34191da177e4SLinus Torvalds } 34201da177e4SLinus Torvalds 3421d83769a5SEric Dumazet /* We allow to exceed memory limits for FIN packets to expedite 3422d83769a5SEric Dumazet * connection tear down and (memory) recovery. 3423845704a5SEric Dumazet * Otherwise tcp_send_fin() could be tempted to either delay FIN 3424845704a5SEric Dumazet * or even be forced to close flow without any FIN. 3425a6c5ea4cSEric Dumazet * In general, we want to allow one skb per socket to avoid hangs 3426a6c5ea4cSEric Dumazet * with edge trigger epoll() 3427d83769a5SEric Dumazet */ 3428a6c5ea4cSEric Dumazet void sk_forced_mem_schedule(struct sock *sk, int size) 3429d83769a5SEric Dumazet { 3430c4ee1185SEric Dumazet int delta, amt; 3431d83769a5SEric Dumazet 3432c4ee1185SEric Dumazet delta = size - sk->sk_forward_alloc; 3433c4ee1185SEric Dumazet if (delta <= 0) 3434d83769a5SEric Dumazet return; 3435c4ee1185SEric Dumazet amt = sk_mem_pages(delta); 3436100fdd1fSEric Dumazet sk->sk_forward_alloc += amt << PAGE_SHIFT; 3437e805605cSJohannes Weiner sk_memory_allocated_add(sk, amt); 3438e805605cSJohannes Weiner 3439baac50bbSJohannes Weiner if (mem_cgroup_sockets_enabled && sk->sk_memcg) 34404b1327beSWei Wang mem_cgroup_charge_skmem(sk->sk_memcg, amt, 34414b1327beSWei Wang gfp_memcg_charge() | __GFP_NOFAIL); 3442d83769a5SEric Dumazet } 3443d83769a5SEric Dumazet 3444845704a5SEric Dumazet /* Send a FIN. The caller locks the socket for us. 3445845704a5SEric Dumazet * We should try to send a FIN packet really hard, but eventually give up. 34461da177e4SLinus Torvalds */ 34471da177e4SLinus Torvalds void tcp_send_fin(struct sock *sk) 34481da177e4SLinus Torvalds { 3449ee2aabd3SEric Dumazet struct sk_buff *skb, *tskb, *tail = tcp_write_queue_tail(sk); 34501da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 34511da177e4SLinus Torvalds 3452845704a5SEric Dumazet /* Optimization, tack on the FIN if we have one skb in write queue and 3453845704a5SEric Dumazet * this skb was not yet sent, or we are under memory pressure. 3454845704a5SEric Dumazet * Note: in the latter case, FIN packet will be sent after a timeout, 3455845704a5SEric Dumazet * as TCP stack thinks it has already been transmitted. 34561da177e4SLinus Torvalds */ 3457ee2aabd3SEric Dumazet tskb = tail; 345875c119afSEric Dumazet if (!tskb && tcp_under_memory_pressure(sk)) 345975c119afSEric Dumazet tskb = skb_rb_last(&sk->tcp_rtx_queue); 346075c119afSEric Dumazet 346175c119afSEric Dumazet if (tskb) { 3462845704a5SEric Dumazet TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN; 3463845704a5SEric Dumazet TCP_SKB_CB(tskb)->end_seq++; 34641da177e4SLinus Torvalds tp->write_seq++; 3465ee2aabd3SEric Dumazet if (!tail) { 3466845704a5SEric Dumazet /* This means tskb was already sent. 3467845704a5SEric Dumazet * Pretend we included the FIN on previous transmit. 3468845704a5SEric Dumazet * We need to set tp->snd_nxt to the value it would have 3469845704a5SEric Dumazet * if FIN had been sent. This is because retransmit path 3470845704a5SEric Dumazet * does not change tp->snd_nxt. 3471845704a5SEric Dumazet */ 3472e0d694d6SEric Dumazet WRITE_ONCE(tp->snd_nxt, tp->snd_nxt + 1); 3473845704a5SEric Dumazet return; 3474845704a5SEric Dumazet } 34751da177e4SLinus Torvalds } else { 3476845704a5SEric Dumazet skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation); 3477d1edc085SColin Ian King if (unlikely(!skb)) 3478845704a5SEric Dumazet return; 3479d1edc085SColin Ian King 3480e2080072SEric Dumazet INIT_LIST_HEAD(&skb->tcp_tsorted_anchor); 3481d83769a5SEric Dumazet skb_reserve(skb, MAX_TCP_HEADER); 3482a6c5ea4cSEric Dumazet sk_forced_mem_schedule(sk, skb->truesize); 34831da177e4SLinus Torvalds /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 3484e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tp->write_seq, 3485a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_FIN); 34861da177e4SLinus Torvalds tcp_queue_skb(sk, skb); 34871da177e4SLinus Torvalds } 3488845704a5SEric Dumazet __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF); 34891da177e4SLinus Torvalds } 34901da177e4SLinus Torvalds 34911da177e4SLinus Torvalds /* We get here when a process closes a file descriptor (either due to 34921da177e4SLinus Torvalds * an explicit close() or as a byproduct of exit()'ing) and there 34931da177e4SLinus Torvalds * was unread data in the receive queue. This behavior is recommended 349465bb723cSGerrit Renker * by RFC 2525, section 2.17. -DaveM 34951da177e4SLinus Torvalds */ 3496dd0fc66fSAl Viro void tcp_send_active_reset(struct sock *sk, gfp_t priority) 34971da177e4SLinus Torvalds { 34981da177e4SLinus Torvalds struct sk_buff *skb; 34991da177e4SLinus Torvalds 35007cc2b043SGao Feng TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); 35017cc2b043SGao Feng 35021da177e4SLinus Torvalds /* NOTE: No TCP options attached and we never retransmit this. */ 35031da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, priority); 35041da177e4SLinus Torvalds if (!skb) { 35054e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 35061da177e4SLinus Torvalds return; 35071da177e4SLinus Torvalds } 35081da177e4SLinus Torvalds 35091da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 35101da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 3511e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), 3512a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_RST); 35139a568de4SEric Dumazet tcp_mstamp_refresh(tcp_sk(sk)); 35141da177e4SLinus Torvalds /* Send it off. */ 3515dfb4b9dcSDavid S. Miller if (tcp_transmit_skb(sk, skb, 0, priority)) 35164e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 3517c24b14c4SSong Liu 3518c24b14c4SSong Liu /* skb of trace_tcp_send_reset() keeps the skb that caused RST, 3519c24b14c4SSong Liu * skb here is different to the troublesome skb, so use NULL 3520c24b14c4SSong Liu */ 3521c24b14c4SSong Liu trace_tcp_send_reset(sk, NULL); 35221da177e4SLinus Torvalds } 35231da177e4SLinus Torvalds 352467edfef7SAndi Kleen /* Send a crossed SYN-ACK during socket establishment. 352567edfef7SAndi Kleen * WARNING: This routine must only be called when we have already sent 35261da177e4SLinus Torvalds * a SYN packet that crossed the incoming SYN that caused this routine 35271da177e4SLinus Torvalds * to get called. If this assumption fails then the initial rcv_wnd 35281da177e4SLinus Torvalds * and rcv_wscale values will not be correct. 35291da177e4SLinus Torvalds */ 35301da177e4SLinus Torvalds int tcp_send_synack(struct sock *sk) 35311da177e4SLinus Torvalds { 35321da177e4SLinus Torvalds struct sk_buff *skb; 35331da177e4SLinus Torvalds 353475c119afSEric Dumazet skb = tcp_rtx_queue_head(sk); 353551456b29SIan Morris if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 353675c119afSEric Dumazet pr_err("%s: wrong queue state\n", __func__); 35371da177e4SLinus Torvalds return -EFAULT; 35381da177e4SLinus Torvalds } 35394de075e0SEric Dumazet if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { 35401da177e4SLinus Torvalds if (skb_cloned(skb)) { 3541e2080072SEric Dumazet struct sk_buff *nskb; 3542e2080072SEric Dumazet 3543e2080072SEric Dumazet tcp_skb_tsorted_save(skb) { 3544e2080072SEric Dumazet nskb = skb_copy(skb, GFP_ATOMIC); 3545e2080072SEric Dumazet } tcp_skb_tsorted_restore(skb); 354651456b29SIan Morris if (!nskb) 35471da177e4SLinus Torvalds return -ENOMEM; 3548e2080072SEric Dumazet INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor); 35492bec445fSEric Dumazet tcp_highest_sack_replace(sk, skb, nskb); 355075c119afSEric Dumazet tcp_rtx_queue_unlink_and_free(skb, sk); 3551f4a775d1SEric Dumazet __skb_header_release(nskb); 355275c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb); 3553ab4e846aSEric Dumazet sk_wmem_queued_add(sk, nskb->truesize); 35543ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 35551da177e4SLinus Torvalds skb = nskb; 35561da177e4SLinus Torvalds } 35571da177e4SLinus Torvalds 35584de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; 3559735d3831SFlorian Westphal tcp_ecn_send_synack(sk, skb); 35601da177e4SLinus Torvalds } 3561dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 35621da177e4SLinus Torvalds } 35631da177e4SLinus Torvalds 35644aea39c1SEric Dumazet /** 3565331fca43SMartin KaFai Lau * tcp_make_synack - Allocate one skb and build a SYNACK packet. 3566331fca43SMartin KaFai Lau * @sk: listener socket 3567331fca43SMartin KaFai Lau * @dst: dst entry attached to the SYNACK. It is consumed and caller 3568331fca43SMartin KaFai Lau * should not use it again. 3569331fca43SMartin KaFai Lau * @req: request_sock pointer 3570331fca43SMartin KaFai Lau * @foc: cookie for tcp fast open 3571331fca43SMartin KaFai Lau * @synack_type: Type of synack to prepare 3572331fca43SMartin KaFai Lau * @syn_skb: SYN packet just received. It could be NULL for rtx case. 35734aea39c1SEric Dumazet */ 35745d062de7SEric Dumazet struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst, 3575e6b4d113SWilliam Allen Simpson struct request_sock *req, 3576ca6fb065SEric Dumazet struct tcp_fastopen_cookie *foc, 3577331fca43SMartin KaFai Lau enum tcp_synack_type synack_type, 3578331fca43SMartin KaFai Lau struct sk_buff *syn_skb) 35791da177e4SLinus Torvalds { 35802e6599cbSArnaldo Carvalho de Melo struct inet_request_sock *ireq = inet_rsk(req); 35815d062de7SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 358280f03e27SEric Dumazet struct tcp_md5sig_key *md5 = NULL; 35835d062de7SEric Dumazet struct tcp_out_options opts; 35845d062de7SEric Dumazet struct sk_buff *skb; 3585bd0388aeSWilliam Allen Simpson int tcp_header_size; 35865d062de7SEric Dumazet struct tcphdr *th; 3587f5fff5dcSTom Quetchenbach int mss; 3588a842fe14SEric Dumazet u64 now; 35891da177e4SLinus Torvalds 3590ca6fb065SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 35914aea39c1SEric Dumazet if (unlikely(!skb)) { 35924aea39c1SEric Dumazet dst_release(dst); 35931da177e4SLinus Torvalds return NULL; 35944aea39c1SEric Dumazet } 35951da177e4SLinus Torvalds /* Reserve space for headers. */ 35961da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 35971da177e4SLinus Torvalds 3598b3d05147SEric Dumazet switch (synack_type) { 3599b3d05147SEric Dumazet case TCP_SYNACK_NORMAL: 36009e17f8a4SEric Dumazet skb_set_owner_w(skb, req_to_sk(req)); 3601b3d05147SEric Dumazet break; 3602b3d05147SEric Dumazet case TCP_SYNACK_COOKIE: 3603b3d05147SEric Dumazet /* Under synflood, we do not attach skb to a socket, 3604b3d05147SEric Dumazet * to avoid false sharing. 3605b3d05147SEric Dumazet */ 3606b3d05147SEric Dumazet break; 3607b3d05147SEric Dumazet case TCP_SYNACK_FASTOPEN: 3608ca6fb065SEric Dumazet /* sk is a const pointer, because we want to express multiple 3609ca6fb065SEric Dumazet * cpu might call us concurrently. 3610ca6fb065SEric Dumazet * sk->sk_wmem_alloc in an atomic, we can promote to rw. 3611ca6fb065SEric Dumazet */ 3612ca6fb065SEric Dumazet skb_set_owner_w(skb, (struct sock *)sk); 3613b3d05147SEric Dumazet break; 3614ca6fb065SEric Dumazet } 36154aea39c1SEric Dumazet skb_dst_set(skb, dst); 36161da177e4SLinus Torvalds 36173541f9e8SEric Dumazet mss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 3618f5fff5dcSTom Quetchenbach 361933ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 3620a842fe14SEric Dumazet now = tcp_clock_ns(); 36218b5f12d0SFlorian Westphal #ifdef CONFIG_SYN_COOKIES 3622f8ace8d9SFlorian Westphal if (unlikely(synack_type == TCP_SYNACK_COOKIE && ireq->tstamp_ok)) 3623a1ac9c8aSMartin KaFai Lau skb_set_delivery_time(skb, cookie_init_timestamp(req, now), 3624a1ac9c8aSMartin KaFai Lau true); 36258b5f12d0SFlorian Westphal else 36268b5f12d0SFlorian Westphal #endif 36279e450c1eSYuchung Cheng { 3628a1ac9c8aSMartin KaFai Lau skb_set_delivery_time(skb, now, true); 36299e450c1eSYuchung Cheng if (!tcp_rsk(req)->snt_synack) /* Timestamp first SYNACK */ 36309e450c1eSYuchung Cheng tcp_rsk(req)->snt_synack = tcp_skb_timestamp_us(skb); 36319e450c1eSYuchung Cheng } 363280f03e27SEric Dumazet 363380f03e27SEric Dumazet #ifdef CONFIG_TCP_MD5SIG 363480f03e27SEric Dumazet rcu_read_lock(); 3635fd3a154aSEric Dumazet md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req)); 363680f03e27SEric Dumazet #endif 363758d607d3SEric Dumazet skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4); 3638331fca43SMartin KaFai Lau /* bpf program will be interested in the tcp_flags */ 3639331fca43SMartin KaFai Lau TCP_SKB_CB(skb)->tcp_flags = TCPHDR_SYN | TCPHDR_ACK; 364060e2a778SUrsula Braun tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5, 3641331fca43SMartin KaFai Lau foc, synack_type, 3642331fca43SMartin KaFai Lau syn_skb) + sizeof(*th); 364333ad798cSAdam Langley 3644aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 3645aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 36461da177e4SLinus Torvalds 3647ea1627c2SEric Dumazet th = (struct tcphdr *)skb->data; 36481da177e4SLinus Torvalds memset(th, 0, sizeof(struct tcphdr)); 36491da177e4SLinus Torvalds th->syn = 1; 36501da177e4SLinus Torvalds th->ack = 1; 36516ac705b1SEric Dumazet tcp_ecn_make_synack(req, th); 3652b44084c2SEric Dumazet th->source = htons(ireq->ir_num); 3653634fb979SEric Dumazet th->dest = ireq->ir_rmt_port; 3654e05a90ecSJamal Hadi Salim skb->mark = ireq->ir_mark; 36553b117750SEric Dumazet skb->ip_summed = CHECKSUM_PARTIAL; 36563b117750SEric Dumazet th->seq = htonl(tcp_rsk(req)->snt_isn); 36578336886fSJerry Chu /* XXX data is queued and acked as is. No buffer/window check */ 36588336886fSJerry Chu th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt); 36591da177e4SLinus Torvalds 36601da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 3661ed53d0abSEric Dumazet th->window = htons(min(req->rsk_rcv_wnd, 65535U)); 3662ea66758cSPaolo Abeni tcp_options_write(th, NULL, &opts); 36631da177e4SLinus Torvalds th->doff = (tcp_header_size >> 2); 3664bced3f7dSBreno Leitao TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS); 3665cfb6eeb4SYOSHIFUJI Hideaki 3666cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 3667cfb6eeb4SYOSHIFUJI Hideaki /* Okay, we have all we need - do the md5 hash if needed */ 366880f03e27SEric Dumazet if (md5) 3669bd0388aeSWilliam Allen Simpson tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, 367039f8e58eSEric Dumazet md5, req_to_sk(req), skb); 367180f03e27SEric Dumazet rcu_read_unlock(); 3672cfb6eeb4SYOSHIFUJI Hideaki #endif 3673cfb6eeb4SYOSHIFUJI Hideaki 3674331fca43SMartin KaFai Lau bpf_skops_write_hdr_opt((struct sock *)sk, skb, req, syn_skb, 3675331fca43SMartin KaFai Lau synack_type, &opts); 3676331fca43SMartin KaFai Lau 3677a1ac9c8aSMartin KaFai Lau skb_set_delivery_time(skb, now, true); 3678a842fe14SEric Dumazet tcp_add_tx_delay(skb, tp); 3679a842fe14SEric Dumazet 36801da177e4SLinus Torvalds return skb; 36811da177e4SLinus Torvalds } 36824bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_make_synack); 36831da177e4SLinus Torvalds 368481164413SDaniel Borkmann static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst) 368581164413SDaniel Borkmann { 368681164413SDaniel Borkmann struct inet_connection_sock *icsk = inet_csk(sk); 368781164413SDaniel Borkmann const struct tcp_congestion_ops *ca; 368881164413SDaniel Borkmann u32 ca_key = dst_metric(dst, RTAX_CC_ALGO); 368981164413SDaniel Borkmann 369081164413SDaniel Borkmann if (ca_key == TCP_CA_UNSPEC) 369181164413SDaniel Borkmann return; 369281164413SDaniel Borkmann 369381164413SDaniel Borkmann rcu_read_lock(); 369481164413SDaniel Borkmann ca = tcp_ca_find_key(ca_key); 36950baf26b0SMartin KaFai Lau if (likely(ca && bpf_try_module_get(ca, ca->owner))) { 36960baf26b0SMartin KaFai Lau bpf_module_put(icsk->icsk_ca_ops, icsk->icsk_ca_ops->owner); 369781164413SDaniel Borkmann icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst); 369881164413SDaniel Borkmann icsk->icsk_ca_ops = ca; 369981164413SDaniel Borkmann } 370081164413SDaniel Borkmann rcu_read_unlock(); 370181164413SDaniel Borkmann } 370281164413SDaniel Borkmann 370367edfef7SAndi Kleen /* Do all connect socket setups that can be done AF independent. */ 3704f7e56a76Sstephen hemminger static void tcp_connect_init(struct sock *sk) 37051da177e4SLinus Torvalds { 3706cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 37071da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 37081da177e4SLinus Torvalds __u8 rcv_wscale; 370913d3b1ebSLawrence Brakmo u32 rcv_wnd; 37101da177e4SLinus Torvalds 37111da177e4SLinus Torvalds /* We'll fix this up when we get a response from the other end. 37121da177e4SLinus Torvalds * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 37131da177e4SLinus Torvalds */ 37145d2ed052SEric Dumazet tp->tcp_header_len = sizeof(struct tcphdr); 37153666f666SKuniyuki Iwashima if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_timestamps)) 37165d2ed052SEric Dumazet tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED; 37171da177e4SLinus Torvalds 3718cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 371900db4124SIan Morris if (tp->af_specific->md5_lookup(sk, sk)) 3720cfb6eeb4SYOSHIFUJI Hideaki tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 3721cfb6eeb4SYOSHIFUJI Hideaki #endif 3722cfb6eeb4SYOSHIFUJI Hideaki 37231da177e4SLinus Torvalds /* If user gave his TCP_MAXSEG, record it to clamp */ 37241da177e4SLinus Torvalds if (tp->rx_opt.user_mss) 37251da177e4SLinus Torvalds tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 37261da177e4SLinus Torvalds tp->max_window = 0; 37275d424d5aSJohn Heffner tcp_mtup_init(sk); 37281da177e4SLinus Torvalds tcp_sync_mss(sk, dst_mtu(dst)); 37291da177e4SLinus Torvalds 373081164413SDaniel Borkmann tcp_ca_dst_init(sk, dst); 373181164413SDaniel Borkmann 37321da177e4SLinus Torvalds if (!tp->window_clamp) 37331da177e4SLinus Torvalds tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 37343541f9e8SEric Dumazet tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 3735f5fff5dcSTom Quetchenbach 37361da177e4SLinus Torvalds tcp_initialize_rcv_mss(sk); 37371da177e4SLinus Torvalds 3738e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 3739e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 3740e88c64f0SHagen Paul Pfeifer (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) 3741e88c64f0SHagen Paul Pfeifer tp->window_clamp = tcp_full_space(sk); 3742e88c64f0SHagen Paul Pfeifer 374313d3b1ebSLawrence Brakmo rcv_wnd = tcp_rwnd_init_bpf(sk); 374413d3b1ebSLawrence Brakmo if (rcv_wnd == 0) 374513d3b1ebSLawrence Brakmo rcv_wnd = dst_metric(dst, RTAX_INITRWND); 374613d3b1ebSLawrence Brakmo 3747ceef9ab6SEric Dumazet tcp_select_initial_window(sk, tcp_full_space(sk), 37481da177e4SLinus Torvalds tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 37491da177e4SLinus Torvalds &tp->rcv_wnd, 37501da177e4SLinus Torvalds &tp->window_clamp, 37513666f666SKuniyuki Iwashima READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_window_scaling), 375231d12926Slaurent chavey &rcv_wscale, 375313d3b1ebSLawrence Brakmo rcv_wnd); 37541da177e4SLinus Torvalds 37551da177e4SLinus Torvalds tp->rx_opt.rcv_wscale = rcv_wscale; 37561da177e4SLinus Torvalds tp->rcv_ssthresh = tp->rcv_wnd; 37571da177e4SLinus Torvalds 3758e13ec3daSEric Dumazet WRITE_ONCE(sk->sk_err, 0); 37591da177e4SLinus Torvalds sock_reset_flag(sk, SOCK_DONE); 37601da177e4SLinus Torvalds tp->snd_wnd = 0; 3761ee7537b6SHantzis Fotis tcp_init_wl(tp, 0); 37627f582b24SEric Dumazet tcp_write_queue_purge(sk); 37631da177e4SLinus Torvalds tp->snd_una = tp->write_seq; 37641da177e4SLinus Torvalds tp->snd_sml = tp->write_seq; 376533f5f57eSIlpo Järvinen tp->snd_up = tp->write_seq; 3766e0d694d6SEric Dumazet WRITE_ONCE(tp->snd_nxt, tp->write_seq); 3767ee995283SPavel Emelyanov 3768ee995283SPavel Emelyanov if (likely(!tp->repair)) 37691da177e4SLinus Torvalds tp->rcv_nxt = 0; 3770c7781a6eSAndrew Vagin else 377170eabf0eSEric Dumazet tp->rcv_tstamp = tcp_jiffies32; 3772ee995283SPavel Emelyanov tp->rcv_wup = tp->rcv_nxt; 37737db48e98SEric Dumazet WRITE_ONCE(tp->copied_seq, tp->rcv_nxt); 37741da177e4SLinus Torvalds 37758550f328SLawrence Brakmo inet_csk(sk)->icsk_rto = tcp_timeout_init(sk); 3776463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_retransmits = 0; 37771da177e4SLinus Torvalds tcp_clear_retrans(tp); 37781da177e4SLinus Torvalds } 37791da177e4SLinus Torvalds 3780783237e8SYuchung Cheng static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) 3781783237e8SYuchung Cheng { 3782783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3783783237e8SYuchung Cheng struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 3784783237e8SYuchung Cheng 3785783237e8SYuchung Cheng tcb->end_seq += skb->len; 3786f4a775d1SEric Dumazet __skb_header_release(skb); 3787ab4e846aSEric Dumazet sk_wmem_queued_add(sk, skb->truesize); 3788783237e8SYuchung Cheng sk_mem_charge(sk, skb->truesize); 37890f317464SEric Dumazet WRITE_ONCE(tp->write_seq, tcb->end_seq); 3790783237e8SYuchung Cheng tp->packets_out += tcp_skb_pcount(skb); 3791783237e8SYuchung Cheng } 3792783237e8SYuchung Cheng 3793783237e8SYuchung Cheng /* Build and send a SYN with data and (cached) Fast Open cookie. However, 3794783237e8SYuchung Cheng * queue a data-only packet after the regular SYN, such that regular SYNs 3795783237e8SYuchung Cheng * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges 3796783237e8SYuchung Cheng * only the SYN sequence, the data are retransmitted in the first ACK. 3797783237e8SYuchung Cheng * If cookie is not cached or other error occurs, falls back to send a 3798783237e8SYuchung Cheng * regular SYN with Fast Open cookie request option. 3799783237e8SYuchung Cheng */ 3800783237e8SYuchung Cheng static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) 3801783237e8SYuchung Cheng { 3802ed0c99dcSJakub Kicinski struct inet_connection_sock *icsk = inet_csk(sk); 3803783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3804783237e8SYuchung Cheng struct tcp_fastopen_request *fo = tp->fastopen_req; 3805*fbf93406SEric Dumazet struct page_frag *pfrag = sk_page_frag(sk); 3806355a901eSEric Dumazet struct sk_buff *syn_data; 3807*fbf93406SEric Dumazet int space, err = 0; 3808783237e8SYuchung Cheng 380967da22d2SYuchung Cheng tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ 3810065263f4SWei Wang if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie)) 3811783237e8SYuchung Cheng goto fallback; 3812783237e8SYuchung Cheng 3813783237e8SYuchung Cheng /* MSS for SYN-data is based on cached MSS and bounded by PMTU and 3814783237e8SYuchung Cheng * user-MSS. Reserve maximum option space for middleboxes that add 3815783237e8SYuchung Cheng * private TCP options. The cost is reduced data space in SYN :( 3816783237e8SYuchung Cheng */ 38173541f9e8SEric Dumazet tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp); 3818ed0c99dcSJakub Kicinski /* Sync mss_cache after updating the mss_clamp */ 3819ed0c99dcSJakub Kicinski tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 38203541f9e8SEric Dumazet 3821ed0c99dcSJakub Kicinski space = __tcp_mtu_to_mss(sk, icsk->icsk_pmtu_cookie) - 3822783237e8SYuchung Cheng MAX_TCP_OPTION_SPACE; 3823783237e8SYuchung Cheng 3824f5ddcbbbSEric Dumazet space = min_t(size_t, space, fo->size); 3825f5ddcbbbSEric Dumazet 3826*fbf93406SEric Dumazet if (space && 3827*fbf93406SEric Dumazet !skb_page_frag_refill(min_t(size_t, space, PAGE_SIZE), 3828*fbf93406SEric Dumazet pfrag, sk->sk_allocation)) 3829*fbf93406SEric Dumazet goto fallback; 3830*fbf93406SEric Dumazet syn_data = tcp_stream_alloc_skb(sk, 0, sk->sk_allocation, false); 3831355a901eSEric Dumazet if (!syn_data) 3832783237e8SYuchung Cheng goto fallback; 3833355a901eSEric Dumazet memcpy(syn_data->cb, syn->cb, sizeof(syn->cb)); 383407e100f9SEric Dumazet if (space) { 3835*fbf93406SEric Dumazet space = min_t(size_t, space, pfrag->size - pfrag->offset); 3836*fbf93406SEric Dumazet space = tcp_wmem_schedule(sk, space); 3837*fbf93406SEric Dumazet } 3838*fbf93406SEric Dumazet if (space) { 3839*fbf93406SEric Dumazet space = copy_page_from_iter(pfrag->page, pfrag->offset, 3840*fbf93406SEric Dumazet space, &fo->data->msg_iter); 3841*fbf93406SEric Dumazet if (unlikely(!space)) { 3842ba233b34SEric Dumazet tcp_skb_tsorted_anchor_cleanup(syn_data); 3843355a901eSEric Dumazet kfree_skb(syn_data); 3844783237e8SYuchung Cheng goto fallback; 3845783237e8SYuchung Cheng } 3846*fbf93406SEric Dumazet skb_fill_page_desc(syn_data, 0, pfrag->page, 3847*fbf93406SEric Dumazet pfrag->offset, space); 3848*fbf93406SEric Dumazet page_ref_inc(pfrag->page); 3849*fbf93406SEric Dumazet pfrag->offset += space; 3850*fbf93406SEric Dumazet skb_len_add(syn_data, space); 3851f859a448SWillem de Bruijn skb_zcopy_set(syn_data, fo->uarg, NULL); 385207e100f9SEric Dumazet } 3853355a901eSEric Dumazet /* No more data pending in inet_wait_for_connect() */ 3854355a901eSEric Dumazet if (space == fo->size) 3855355a901eSEric Dumazet fo->data = NULL; 3856355a901eSEric Dumazet fo->copied = space; 3857783237e8SYuchung Cheng 3858355a901eSEric Dumazet tcp_connect_queue_skb(sk, syn_data); 38590f87230dSFrancis Yan if (syn_data->len) 38600f87230dSFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_BUSY); 3861355a901eSEric Dumazet 3862355a901eSEric Dumazet err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation); 3863355a901eSEric Dumazet 3864a1ac9c8aSMartin KaFai Lau skb_set_delivery_time(syn, syn_data->skb_mstamp_ns, true); 3865355a901eSEric Dumazet 3866355a901eSEric Dumazet /* Now full SYN+DATA was cloned and sent (or not), 3867355a901eSEric Dumazet * remove the SYN from the original skb (syn_data) 3868355a901eSEric Dumazet * we keep in write queue in case of a retransmit, as we 3869355a901eSEric Dumazet * also have the SYN packet (with no data) in the same queue. 3870431a9124SEric Dumazet */ 3871355a901eSEric Dumazet TCP_SKB_CB(syn_data)->seq++; 3872355a901eSEric Dumazet TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH; 3873355a901eSEric Dumazet if (!err) { 387467da22d2SYuchung Cheng tp->syn_data = (fo->copied > 0); 387575c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data); 3876f19c29e3SYuchung Cheng NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT); 3877783237e8SYuchung Cheng goto done; 3878783237e8SYuchung Cheng } 3879783237e8SYuchung Cheng 388075c119afSEric Dumazet /* data was not sent, put it in write_queue */ 388175c119afSEric Dumazet __skb_queue_tail(&sk->sk_write_queue, syn_data); 3882b5b7db8dSEric Dumazet tp->packets_out -= tcp_skb_pcount(syn_data); 3883b5b7db8dSEric Dumazet 3884783237e8SYuchung Cheng fallback: 3885783237e8SYuchung Cheng /* Send a regular SYN with Fast Open cookie request option */ 3886783237e8SYuchung Cheng if (fo->cookie.len > 0) 3887783237e8SYuchung Cheng fo->cookie.len = 0; 3888783237e8SYuchung Cheng err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); 3889783237e8SYuchung Cheng if (err) 3890783237e8SYuchung Cheng tp->syn_fastopen = 0; 3891783237e8SYuchung Cheng done: 3892783237e8SYuchung Cheng fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ 3893783237e8SYuchung Cheng return err; 3894783237e8SYuchung Cheng } 3895783237e8SYuchung Cheng 389667edfef7SAndi Kleen /* Build a SYN and send it off. */ 38971da177e4SLinus Torvalds int tcp_connect(struct sock *sk) 38981da177e4SLinus Torvalds { 38991da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 39001da177e4SLinus Torvalds struct sk_buff *buff; 3901ee586811SEric Paris int err; 39021da177e4SLinus Torvalds 3903de525be2SLawrence Brakmo tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB, 0, NULL); 39048ba60924SEric Dumazet 39058ba60924SEric Dumazet if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 39068ba60924SEric Dumazet return -EHOSTUNREACH; /* Routing failure or similar. */ 39078ba60924SEric Dumazet 39081da177e4SLinus Torvalds tcp_connect_init(sk); 39091da177e4SLinus Torvalds 39102b916477SAndrey Vagin if (unlikely(tp->repair)) { 39112b916477SAndrey Vagin tcp_finish_connect(sk, NULL); 39122b916477SAndrey Vagin return 0; 39132b916477SAndrey Vagin } 39142b916477SAndrey Vagin 3915f8dd3b8dSEric Dumazet buff = tcp_stream_alloc_skb(sk, 0, sk->sk_allocation, true); 3916355a901eSEric Dumazet if (unlikely(!buff)) 39171da177e4SLinus Torvalds return -ENOBUFS; 39181da177e4SLinus Torvalds 3919a3433f35SChangli Gao tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); 39209a568de4SEric Dumazet tcp_mstamp_refresh(tp); 39219a568de4SEric Dumazet tp->retrans_stamp = tcp_time_stamp(tp); 3922783237e8SYuchung Cheng tcp_connect_queue_skb(sk, buff); 3923735d3831SFlorian Westphal tcp_ecn_send_syn(sk, buff); 392475c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, buff); 39251da177e4SLinus Torvalds 3926783237e8SYuchung Cheng /* Send off SYN; include data in Fast Open. */ 3927783237e8SYuchung Cheng err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : 3928783237e8SYuchung Cheng tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); 3929ee586811SEric Paris if (err == -ECONNREFUSED) 3930ee586811SEric Paris return err; 3931bd37a088SWei Yongjun 3932bd37a088SWei Yongjun /* We change tp->snd_nxt after the tcp_transmit_skb() call 3933bd37a088SWei Yongjun * in order to make this packet get counted in tcpOutSegs. 3934bd37a088SWei Yongjun */ 3935e0d694d6SEric Dumazet WRITE_ONCE(tp->snd_nxt, tp->write_seq); 3936bd37a088SWei Yongjun tp->pushed_seq = tp->write_seq; 3937b5b7db8dSEric Dumazet buff = tcp_send_head(sk); 3938b5b7db8dSEric Dumazet if (unlikely(buff)) { 3939e0d694d6SEric Dumazet WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(buff)->seq); 3940b5b7db8dSEric Dumazet tp->pushed_seq = TCP_SKB_CB(buff)->seq; 3941b5b7db8dSEric Dumazet } 394281cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); 39431da177e4SLinus Torvalds 39441da177e4SLinus Torvalds /* Timer for repeating the SYN until an answer. */ 39453f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 39463f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 39471da177e4SLinus Torvalds return 0; 39481da177e4SLinus Torvalds } 39494bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_connect); 39501da177e4SLinus Torvalds 39511da177e4SLinus Torvalds /* Send out a delayed ack, the caller does the policy checking 39521da177e4SLinus Torvalds * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 39531da177e4SLinus Torvalds * for details. 39541da177e4SLinus Torvalds */ 39551da177e4SLinus Torvalds void tcp_send_delayed_ack(struct sock *sk) 39561da177e4SLinus Torvalds { 3957463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 3958463c84b9SArnaldo Carvalho de Melo int ato = icsk->icsk_ack.ato; 39591da177e4SLinus Torvalds unsigned long timeout; 39601da177e4SLinus Torvalds 39611da177e4SLinus Torvalds if (ato > TCP_DELACK_MIN) { 3962463c84b9SArnaldo Carvalho de Melo const struct tcp_sock *tp = tcp_sk(sk); 39631da177e4SLinus Torvalds int max_ato = HZ / 2; 39641da177e4SLinus Torvalds 396531954cd8SWei Wang if (inet_csk_in_pingpong_mode(sk) || 3966056834d9SIlpo Järvinen (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 39671da177e4SLinus Torvalds max_ato = TCP_DELACK_MAX; 39681da177e4SLinus Torvalds 39691da177e4SLinus Torvalds /* Slow path, intersegment interval is "high". */ 39701da177e4SLinus Torvalds 39711da177e4SLinus Torvalds /* If some rtt estimate is known, use it to bound delayed ack. 3972463c84b9SArnaldo Carvalho de Melo * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 39731da177e4SLinus Torvalds * directly. 39741da177e4SLinus Torvalds */ 3975740b0f18SEric Dumazet if (tp->srtt_us) { 3976740b0f18SEric Dumazet int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3), 3977740b0f18SEric Dumazet TCP_DELACK_MIN); 39781da177e4SLinus Torvalds 39791da177e4SLinus Torvalds if (rtt < max_ato) 39801da177e4SLinus Torvalds max_ato = rtt; 39811da177e4SLinus Torvalds } 39821da177e4SLinus Torvalds 39831da177e4SLinus Torvalds ato = min(ato, max_ato); 39841da177e4SLinus Torvalds } 39851da177e4SLinus Torvalds 39862b8ee4f0SMartin KaFai Lau ato = min_t(u32, ato, inet_csk(sk)->icsk_delack_max); 39872b8ee4f0SMartin KaFai Lau 39881da177e4SLinus Torvalds /* Stay within the limit we were given */ 39891da177e4SLinus Torvalds timeout = jiffies + ato; 39901da177e4SLinus Torvalds 39911da177e4SLinus Torvalds /* Use new timeout only if there wasn't a older one earlier. */ 3992463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 3993b6b6d653SEric Dumazet /* If delack timer is about to expire, send ACK now. */ 3994b6b6d653SEric Dumazet if (time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 39951da177e4SLinus Torvalds tcp_send_ack(sk); 39961da177e4SLinus Torvalds return; 39971da177e4SLinus Torvalds } 39981da177e4SLinus Torvalds 3999463c84b9SArnaldo Carvalho de Melo if (!time_before(timeout, icsk->icsk_ack.timeout)) 4000463c84b9SArnaldo Carvalho de Melo timeout = icsk->icsk_ack.timeout; 40011da177e4SLinus Torvalds } 4002463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 4003463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.timeout = timeout; 4004463c84b9SArnaldo Carvalho de Melo sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 40051da177e4SLinus Torvalds } 40061da177e4SLinus Torvalds 40071da177e4SLinus Torvalds /* This routine sends an ack and also updates the window. */ 40082987babbSYuchung Cheng void __tcp_send_ack(struct sock *sk, u32 rcv_nxt) 40091da177e4SLinus Torvalds { 40101da177e4SLinus Torvalds struct sk_buff *buff; 40111da177e4SLinus Torvalds 4012058dc334SIlpo Järvinen /* If we have been reset, we may not send again. */ 4013058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 4014058dc334SIlpo Järvinen return; 4015058dc334SIlpo Järvinen 40161da177e4SLinus Torvalds /* We are not putting this on the write queue, so 40171da177e4SLinus Torvalds * tcp_transmit_skb() will set the ownership to this 40181da177e4SLinus Torvalds * sock. 40191da177e4SLinus Torvalds */ 40207450aaf6SEric Dumazet buff = alloc_skb(MAX_TCP_HEADER, 40217450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 40227450aaf6SEric Dumazet if (unlikely(!buff)) { 4023a37c2134SEric Dumazet struct inet_connection_sock *icsk = inet_csk(sk); 4024a37c2134SEric Dumazet unsigned long delay; 4025a37c2134SEric Dumazet 4026a37c2134SEric Dumazet delay = TCP_DELACK_MAX << icsk->icsk_ack.retry; 4027a37c2134SEric Dumazet if (delay < TCP_RTO_MAX) 4028a37c2134SEric Dumazet icsk->icsk_ack.retry++; 4029463c84b9SArnaldo Carvalho de Melo inet_csk_schedule_ack(sk); 4030a37c2134SEric Dumazet icsk->icsk_ack.ato = TCP_ATO_MIN; 4031a37c2134SEric Dumazet inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, delay, TCP_RTO_MAX); 40321da177e4SLinus Torvalds return; 40331da177e4SLinus Torvalds } 40341da177e4SLinus Torvalds 40351da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 40361da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 4037a3433f35SChangli Gao tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); 40381da177e4SLinus Torvalds 403998781965SEric Dumazet /* We do not want pure acks influencing TCP Small Queues or fq/pacing 404098781965SEric Dumazet * too much. 404198781965SEric Dumazet * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784 404298781965SEric Dumazet */ 404398781965SEric Dumazet skb_set_tcp_pure_ack(buff); 404498781965SEric Dumazet 40451da177e4SLinus Torvalds /* Send it off, this clears delayed acks for us. */ 40462987babbSYuchung Cheng __tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0, rcv_nxt); 40471da177e4SLinus Torvalds } 404827cde44aSYuchung Cheng EXPORT_SYMBOL_GPL(__tcp_send_ack); 40492987babbSYuchung Cheng 40502987babbSYuchung Cheng void tcp_send_ack(struct sock *sk) 40512987babbSYuchung Cheng { 40522987babbSYuchung Cheng __tcp_send_ack(sk, tcp_sk(sk)->rcv_nxt); 40531da177e4SLinus Torvalds } 40541da177e4SLinus Torvalds 40551da177e4SLinus Torvalds /* This routine sends a packet with an out of date sequence 40561da177e4SLinus Torvalds * number. It assumes the other end will try to ack it. 40571da177e4SLinus Torvalds * 40581da177e4SLinus Torvalds * Question: what should we make while urgent mode? 40591da177e4SLinus Torvalds * 4.4BSD forces sending single byte of data. We cannot send 40601da177e4SLinus Torvalds * out of window data, because we have SND.NXT==SND.MAX... 40611da177e4SLinus Torvalds * 40621da177e4SLinus Torvalds * Current solution: to send TWO zero-length segments in urgent mode: 40631da177e4SLinus Torvalds * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 40641da177e4SLinus Torvalds * out-of-date with SND.UNA-1 to probe window. 40651da177e4SLinus Torvalds */ 4066e520af48SEric Dumazet static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib) 40671da177e4SLinus Torvalds { 40681da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 40691da177e4SLinus Torvalds struct sk_buff *skb; 40701da177e4SLinus Torvalds 40711da177e4SLinus Torvalds /* We don't queue it, tcp_transmit_skb() sets ownership. */ 40727450aaf6SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, 40737450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 407451456b29SIan Morris if (!skb) 40751da177e4SLinus Torvalds return -1; 40761da177e4SLinus Torvalds 40771da177e4SLinus Torvalds /* Reserve space for headers and set control bits. */ 40781da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 40791da177e4SLinus Torvalds /* Use a previous sequence. This should cause the other 40801da177e4SLinus Torvalds * end to send an ack. Don't queue or clone SKB, just 40811da177e4SLinus Torvalds * send it. 40821da177e4SLinus Torvalds */ 4083a3433f35SChangli Gao tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); 4084e2e8009fSRenato Westphal NET_INC_STATS(sock_net(sk), mib); 40857450aaf6SEric Dumazet return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0); 40861da177e4SLinus Torvalds } 40871da177e4SLinus Torvalds 4088385e2070SEric Dumazet /* Called from setsockopt( ... TCP_REPAIR ) */ 4089ee995283SPavel Emelyanov void tcp_send_window_probe(struct sock *sk) 4090ee995283SPavel Emelyanov { 4091ee995283SPavel Emelyanov if (sk->sk_state == TCP_ESTABLISHED) { 4092ee995283SPavel Emelyanov tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; 40939a568de4SEric Dumazet tcp_mstamp_refresh(tcp_sk(sk)); 4094e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE); 4095ee995283SPavel Emelyanov } 4096ee995283SPavel Emelyanov } 4097ee995283SPavel Emelyanov 409867edfef7SAndi Kleen /* Initiate keepalive or window probe from timer. */ 4099e520af48SEric Dumazet int tcp_write_wakeup(struct sock *sk, int mib) 41001da177e4SLinus Torvalds { 41011da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 41021da177e4SLinus Torvalds struct sk_buff *skb; 41031da177e4SLinus Torvalds 4104058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 4105058dc334SIlpo Järvinen return -1; 4106058dc334SIlpo Järvinen 410700db4124SIan Morris skb = tcp_send_head(sk); 410800db4124SIan Morris if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { 41091da177e4SLinus Torvalds int err; 41100c54b85fSIlpo Järvinen unsigned int mss = tcp_current_mss(sk); 411190840defSIlpo Järvinen unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 41121da177e4SLinus Torvalds 41131da177e4SLinus Torvalds if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 41141da177e4SLinus Torvalds tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 41151da177e4SLinus Torvalds 41161da177e4SLinus Torvalds /* We are probing the opening of a window 41171da177e4SLinus Torvalds * but the window size is != 0 41181da177e4SLinus Torvalds * must have been a result SWS avoidance ( sender ) 41191da177e4SLinus Torvalds */ 41201da177e4SLinus Torvalds if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 41211da177e4SLinus Torvalds skb->len > mss) { 41221da177e4SLinus Torvalds seg_size = min(seg_size, mss); 41234de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 412475c119afSEric Dumazet if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE, 412575c119afSEric Dumazet skb, seg_size, mss, GFP_ATOMIC)) 41261da177e4SLinus Torvalds return -1; 41271da177e4SLinus Torvalds } else if (!tcp_skb_pcount(skb)) 41285bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss); 41291da177e4SLinus Torvalds 41304de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 4131dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 413266f5fe62SIlpo Järvinen if (!err) 413366f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 41341da177e4SLinus Torvalds return err; 41351da177e4SLinus Torvalds } else { 413633f5f57eSIlpo Järvinen if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) 4137e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 1, mib); 4138e520af48SEric Dumazet return tcp_xmit_probe_skb(sk, 0, mib); 41391da177e4SLinus Torvalds } 41401da177e4SLinus Torvalds } 41411da177e4SLinus Torvalds 41421da177e4SLinus Torvalds /* A window probe timeout has occurred. If window is not closed send 41431da177e4SLinus Torvalds * a partial packet else a zero probe. 41441da177e4SLinus Torvalds */ 41451da177e4SLinus Torvalds void tcp_send_probe0(struct sock *sk) 41461da177e4SLinus Torvalds { 4147463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 41481da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 4149c6214a97SNikolay Borisov struct net *net = sock_net(sk); 4150c1d5674fSYuchung Cheng unsigned long timeout; 41511da177e4SLinus Torvalds int err; 41521da177e4SLinus Torvalds 4153e520af48SEric Dumazet err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE); 41541da177e4SLinus Torvalds 415575c119afSEric Dumazet if (tp->packets_out || tcp_write_queue_empty(sk)) { 41561da177e4SLinus Torvalds /* Cancel probe timer, if it is not required. */ 41576687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 0; 4158463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff = 0; 41599d9b1ee0SEnke Chen icsk->icsk_probes_tstamp = 0; 41601da177e4SLinus Torvalds return; 41611da177e4SLinus Torvalds } 41621da177e4SLinus Torvalds 4163c1d5674fSYuchung Cheng icsk->icsk_probes_out++; 41641da177e4SLinus Torvalds if (err <= 0) { 416539e24435SKuniyuki Iwashima if (icsk->icsk_backoff < READ_ONCE(net->ipv4.sysctl_tcp_retries2)) 4166463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff++; 4167c1d5674fSYuchung Cheng timeout = tcp_probe0_when(sk, TCP_RTO_MAX); 41681da177e4SLinus Torvalds } else { 41691da177e4SLinus Torvalds /* If packet was not sent due to local congestion, 4170c1d5674fSYuchung Cheng * Let senders fight for local resources conservatively. 41711da177e4SLinus Torvalds */ 4172c1d5674fSYuchung Cheng timeout = TCP_RESOURCE_PROBE_INTERVAL; 41731da177e4SLinus Torvalds } 4174344db93aSEnke Chen 4175344db93aSEnke Chen timeout = tcp_clamp_probe0_to_user_timeout(sk, timeout); 41768dc242adSEric Dumazet tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, timeout, TCP_RTO_MAX); 41771da177e4SLinus Torvalds } 41785db92c99SOctavian Purdila 4179ea3bea3aSEric Dumazet int tcp_rtx_synack(const struct sock *sk, struct request_sock *req) 41805db92c99SOctavian Purdila { 41815db92c99SOctavian Purdila const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific; 41825db92c99SOctavian Purdila struct flowi fl; 41835db92c99SOctavian Purdila int res; 41845db92c99SOctavian Purdila 4185cb6cd2ceSAkhmat Karakotov /* Paired with WRITE_ONCE() in sock_setsockopt() */ 4186cb6cd2ceSAkhmat Karakotov if (READ_ONCE(sk->sk_txrehash) == SOCK_TXREHASH_ENABLED) 418758d607d3SEric Dumazet tcp_rsk(req)->txhash = net_tx_rndhash(); 4188331fca43SMartin KaFai Lau res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL, 4189331fca43SMartin KaFai Lau NULL); 41905db92c99SOctavian Purdila if (!res) { 41910a375c82SEric Dumazet TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); 41920a375c82SEric Dumazet NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 4193e9d9da91SEric Dumazet if (unlikely(tcp_passive_fastopen(sk))) { 4194e9d9da91SEric Dumazet /* sk has const attribute because listeners are lockless. 4195e9d9da91SEric Dumazet * However in this case, we are dealing with a passive fastopen 4196e9d9da91SEric Dumazet * socket thus we can change total_retrans value. 4197e9d9da91SEric Dumazet */ 4198e9d9da91SEric Dumazet tcp_sk_rw(sk)->total_retrans++; 4199e9d9da91SEric Dumazet } 4200cf34ce3dSSong Liu trace_tcp_retransmit_synack(sk, req); 42015db92c99SOctavian Purdila } 42025db92c99SOctavian Purdila return res; 42035db92c99SOctavian Purdila } 42045db92c99SOctavian Purdila EXPORT_SYMBOL(tcp_rtx_synack); 4205