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); 263*b650d953Smfreemon@cloudflare.com struct net *net = sock_net(sk); 2641da177e4SLinus Torvalds 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 */ 273*b650d953Smfreemon@cloudflare.com if (!READ_ONCE(net->ipv4.sysctl_tcp_shrink_window) || !tp->rx_opt.rcv_wscale) { 274*b650d953Smfreemon@cloudflare.com /* Never shrink the offered window */ 2758e165e20SFlorian Westphal if (new_win == 0) 276*b650d953Smfreemon@cloudflare.com NET_INC_STATS(net, LINUX_MIB_TCPWANTZEROWINDOWADV); 277607bfbf2SPatrick McHardy new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale); 2781da177e4SLinus Torvalds } 279*b650d953Smfreemon@cloudflare.com } 280*b650d953Smfreemon@cloudflare.com 2811da177e4SLinus Torvalds tp->rcv_wnd = new_win; 2821da177e4SLinus Torvalds tp->rcv_wup = tp->rcv_nxt; 2831da177e4SLinus Torvalds 2841da177e4SLinus Torvalds /* Make sure we do not exceed the maximum possible 2851da177e4SLinus Torvalds * scaled window. 2861da177e4SLinus Torvalds */ 287ceef9ab6SEric Dumazet if (!tp->rx_opt.rcv_wscale && 288*b650d953Smfreemon@cloudflare.com READ_ONCE(net->ipv4.sysctl_tcp_workaround_signed_windows)) 2891da177e4SLinus Torvalds new_win = min(new_win, MAX_TCP_WINDOW); 2901da177e4SLinus Torvalds else 2911da177e4SLinus Torvalds new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); 2921da177e4SLinus Torvalds 2931da177e4SLinus Torvalds /* RFC1323 scaling applied */ 2941da177e4SLinus Torvalds new_win >>= tp->rx_opt.rcv_wscale; 2951da177e4SLinus Torvalds 29631770e34SFlorian Westphal /* If we advertise zero window, disable fast path. */ 2978e165e20SFlorian Westphal if (new_win == 0) { 29831770e34SFlorian Westphal tp->pred_flags = 0; 2998e165e20SFlorian Westphal if (old_win) 300*b650d953Smfreemon@cloudflare.com NET_INC_STATS(net, LINUX_MIB_TCPTOZEROWINDOWADV); 3018e165e20SFlorian Westphal } else if (old_win == 0) { 302*b650d953Smfreemon@cloudflare.com NET_INC_STATS(net, LINUX_MIB_TCPFROMZEROWINDOWADV); 3038e165e20SFlorian Westphal } 3041da177e4SLinus Torvalds 3051da177e4SLinus Torvalds return new_win; 3061da177e4SLinus Torvalds } 3071da177e4SLinus Torvalds 30867edfef7SAndi Kleen /* Packet ECN state for a SYN-ACK */ 309735d3831SFlorian Westphal static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb) 310bdf1ee5dSIlpo Järvinen { 31130e502a3SDaniel Borkmann const struct tcp_sock *tp = tcp_sk(sk); 31230e502a3SDaniel Borkmann 3134de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR; 314bdf1ee5dSIlpo Järvinen if (!(tp->ecn_flags & TCP_ECN_OK)) 3154de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE; 31691b5b21cSLawrence Brakmo else if (tcp_ca_needs_ecn(sk) || 31791b5b21cSLawrence Brakmo tcp_bpf_ca_needs_ecn(sk)) 31830e502a3SDaniel Borkmann INET_ECN_xmit(sk); 319bdf1ee5dSIlpo Järvinen } 320bdf1ee5dSIlpo Järvinen 32167edfef7SAndi Kleen /* Packet ECN state for a SYN. */ 322735d3831SFlorian Westphal static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb) 323bdf1ee5dSIlpo Järvinen { 324bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 32591b5b21cSLawrence Brakmo bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk); 3264785a667SKuniyuki Iwashima bool use_ecn = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_ecn) == 1 || 32791b5b21cSLawrence Brakmo tcp_ca_needs_ecn(sk) || bpf_needs_ecn; 328f7b3bec6SFlorian Westphal 329f7b3bec6SFlorian Westphal if (!use_ecn) { 330f7b3bec6SFlorian Westphal const struct dst_entry *dst = __sk_dst_get(sk); 331f7b3bec6SFlorian Westphal 332f7b3bec6SFlorian Westphal if (dst && dst_feature(dst, RTAX_FEATURE_ECN)) 333f7b3bec6SFlorian Westphal use_ecn = true; 334f7b3bec6SFlorian Westphal } 335bdf1ee5dSIlpo Järvinen 336bdf1ee5dSIlpo Järvinen tp->ecn_flags = 0; 337f7b3bec6SFlorian Westphal 338f7b3bec6SFlorian Westphal if (use_ecn) { 3394de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR; 340bdf1ee5dSIlpo Järvinen tp->ecn_flags = TCP_ECN_OK; 34191b5b21cSLawrence Brakmo if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn) 34230e502a3SDaniel Borkmann INET_ECN_xmit(sk); 343bdf1ee5dSIlpo Järvinen } 344bdf1ee5dSIlpo Järvinen } 345bdf1ee5dSIlpo Järvinen 34649213555SDaniel Borkmann static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb) 34749213555SDaniel Borkmann { 34812b8d9caSKuniyuki Iwashima if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback)) 34949213555SDaniel Borkmann /* tp->ecn_flags are cleared at a later point in time when 35049213555SDaniel Borkmann * SYN ACK is ultimatively being received. 35149213555SDaniel Borkmann */ 35249213555SDaniel Borkmann TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR); 35349213555SDaniel Borkmann } 35449213555SDaniel Borkmann 355735d3831SFlorian Westphal static void 3566ac705b1SEric Dumazet tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th) 357bdf1ee5dSIlpo Järvinen { 3586ac705b1SEric Dumazet if (inet_rsk(req)->ecn_ok) 359bdf1ee5dSIlpo Järvinen th->ece = 1; 360bdf1ee5dSIlpo Järvinen } 361bdf1ee5dSIlpo Järvinen 36267edfef7SAndi Kleen /* Set up ECN state for a packet on a ESTABLISHED socket that is about to 36367edfef7SAndi Kleen * be sent. 36467edfef7SAndi Kleen */ 365735d3831SFlorian Westphal static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb, 366ea1627c2SEric Dumazet struct tcphdr *th, int tcp_header_len) 367bdf1ee5dSIlpo Järvinen { 368bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 369bdf1ee5dSIlpo Järvinen 370bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_OK) { 371bdf1ee5dSIlpo Järvinen /* Not-retransmitted data segment: set ECT and inject CWR. */ 372bdf1ee5dSIlpo Järvinen if (skb->len != tcp_header_len && 373bdf1ee5dSIlpo Järvinen !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { 374bdf1ee5dSIlpo Järvinen INET_ECN_xmit(sk); 375bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) { 376bdf1ee5dSIlpo Järvinen tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; 377ea1627c2SEric Dumazet th->cwr = 1; 378bdf1ee5dSIlpo Järvinen skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; 379bdf1ee5dSIlpo Järvinen } 38030e502a3SDaniel Borkmann } else if (!tcp_ca_needs_ecn(sk)) { 381bdf1ee5dSIlpo Järvinen /* ACK or retransmitted segment: clear ECT|CE */ 382bdf1ee5dSIlpo Järvinen INET_ECN_dontxmit(sk); 383bdf1ee5dSIlpo Järvinen } 384bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) 385ea1627c2SEric Dumazet th->ece = 1; 386bdf1ee5dSIlpo Järvinen } 387bdf1ee5dSIlpo Järvinen } 388bdf1ee5dSIlpo Järvinen 389e870a8efSIlpo Järvinen /* Constructs common control bits of non-data skb. If SYN/FIN is present, 390e870a8efSIlpo Järvinen * auto increment end seqno. 391e870a8efSIlpo Järvinen */ 392e870a8efSIlpo Järvinen static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags) 393e870a8efSIlpo Järvinen { 3942e8e18efSDavid S. Miller skb->ip_summed = CHECKSUM_PARTIAL; 395e870a8efSIlpo Järvinen 3964de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags; 397e870a8efSIlpo Järvinen 398cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, 1); 399e870a8efSIlpo Järvinen 400e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->seq = seq; 401a3433f35SChangli Gao if (flags & (TCPHDR_SYN | TCPHDR_FIN)) 402e870a8efSIlpo Järvinen seq++; 403e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->end_seq = seq; 404e870a8efSIlpo Järvinen } 405e870a8efSIlpo Järvinen 406a2a385d6SEric Dumazet static inline bool tcp_urg_mode(const struct tcp_sock *tp) 40733f5f57eSIlpo Järvinen { 40833f5f57eSIlpo Järvinen return tp->snd_una != tp->snd_up; 40933f5f57eSIlpo Järvinen } 41033f5f57eSIlpo Järvinen 4113b65abb8SLeonard Crestez #define OPTION_SACK_ADVERTISE BIT(0) 4123b65abb8SLeonard Crestez #define OPTION_TS BIT(1) 4133b65abb8SLeonard Crestez #define OPTION_MD5 BIT(2) 4143b65abb8SLeonard Crestez #define OPTION_WSCALE BIT(3) 4153b65abb8SLeonard Crestez #define OPTION_FAST_OPEN_COOKIE BIT(8) 4163b65abb8SLeonard Crestez #define OPTION_SMC BIT(9) 4173b65abb8SLeonard Crestez #define OPTION_MPTCP BIT(10) 41860e2a778SUrsula Braun 41960e2a778SUrsula Braun static void smc_options_write(__be32 *ptr, u16 *options) 42060e2a778SUrsula Braun { 42160e2a778SUrsula Braun #if IS_ENABLED(CONFIG_SMC) 42260e2a778SUrsula Braun if (static_branch_unlikely(&tcp_have_smc)) { 42360e2a778SUrsula Braun if (unlikely(OPTION_SMC & *options)) { 42460e2a778SUrsula Braun *ptr++ = htonl((TCPOPT_NOP << 24) | 42560e2a778SUrsula Braun (TCPOPT_NOP << 16) | 42660e2a778SUrsula Braun (TCPOPT_EXP << 8) | 42760e2a778SUrsula Braun (TCPOLEN_EXP_SMC_BASE)); 42860e2a778SUrsula Braun *ptr++ = htonl(TCPOPT_SMC_MAGIC); 42960e2a778SUrsula Braun } 43060e2a778SUrsula Braun } 43160e2a778SUrsula Braun #endif 43260e2a778SUrsula Braun } 43333ad798cSAdam Langley 43433ad798cSAdam Langley struct tcp_out_options { 4352100c8d2SYuchung Cheng u16 options; /* bit field of OPTION_* */ 4362100c8d2SYuchung Cheng u16 mss; /* 0 to disable */ 43733ad798cSAdam Langley u8 ws; /* window scale, 0 to disable */ 43833ad798cSAdam Langley u8 num_sack_blocks; /* number of SACK blocks to include */ 439bd0388aeSWilliam Allen Simpson u8 hash_size; /* bytes in hash_location */ 440331fca43SMartin KaFai Lau u8 bpf_opt_len; /* length of BPF hdr option */ 441bd0388aeSWilliam Allen Simpson __u8 *hash_location; /* temporary pointer, overloaded */ 4422100c8d2SYuchung Cheng __u32 tsval, tsecr; /* need to include OPTION_TS */ 4432100c8d2SYuchung Cheng struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */ 444eda7acddSPeter Krystad struct mptcp_out_options mptcp; 44533ad798cSAdam Langley }; 44633ad798cSAdam Langley 447ea66758cSPaolo Abeni static void mptcp_options_write(struct tcphdr *th, __be32 *ptr, 448ea66758cSPaolo Abeni struct tcp_sock *tp, 449fa3fe2b1SFlorian Westphal struct tcp_out_options *opts) 450eda7acddSPeter Krystad { 451eda7acddSPeter Krystad #if IS_ENABLED(CONFIG_MPTCP) 452eda7acddSPeter Krystad if (unlikely(OPTION_MPTCP & opts->options)) 453ea66758cSPaolo Abeni mptcp_write_options(th, ptr, tp, &opts->mptcp); 454eda7acddSPeter Krystad #endif 455eda7acddSPeter Krystad } 456eda7acddSPeter Krystad 457331fca43SMartin KaFai Lau #ifdef CONFIG_CGROUP_BPF 4580813a841SMartin KaFai Lau static int bpf_skops_write_hdr_opt_arg0(struct sk_buff *skb, 4590813a841SMartin KaFai Lau enum tcp_synack_type synack_type) 4600813a841SMartin KaFai Lau { 4610813a841SMartin KaFai Lau if (unlikely(!skb)) 4620813a841SMartin KaFai Lau return BPF_WRITE_HDR_TCP_CURRENT_MSS; 4630813a841SMartin KaFai Lau 4640813a841SMartin KaFai Lau if (unlikely(synack_type == TCP_SYNACK_COOKIE)) 4650813a841SMartin KaFai Lau return BPF_WRITE_HDR_TCP_SYNACK_COOKIE; 4660813a841SMartin KaFai Lau 4670813a841SMartin KaFai Lau return 0; 4680813a841SMartin KaFai Lau } 4690813a841SMartin KaFai Lau 470331fca43SMartin KaFai Lau /* req, syn_skb and synack_type are used when writing synack */ 471331fca43SMartin KaFai Lau static void bpf_skops_hdr_opt_len(struct sock *sk, struct sk_buff *skb, 472331fca43SMartin KaFai Lau struct request_sock *req, 473331fca43SMartin KaFai Lau struct sk_buff *syn_skb, 474331fca43SMartin KaFai Lau enum tcp_synack_type synack_type, 475331fca43SMartin KaFai Lau struct tcp_out_options *opts, 476331fca43SMartin KaFai Lau unsigned int *remaining) 477331fca43SMartin KaFai Lau { 4780813a841SMartin KaFai Lau struct bpf_sock_ops_kern sock_ops; 4790813a841SMartin KaFai Lau int err; 4800813a841SMartin KaFai Lau 481331fca43SMartin KaFai Lau if (likely(!BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), 482331fca43SMartin KaFai Lau BPF_SOCK_OPS_WRITE_HDR_OPT_CB_FLAG)) || 483331fca43SMartin KaFai Lau !*remaining) 484331fca43SMartin KaFai Lau return; 485331fca43SMartin KaFai Lau 4860813a841SMartin KaFai Lau /* *remaining has already been aligned to 4 bytes, so *remaining >= 4 */ 4870813a841SMartin KaFai Lau 4880813a841SMartin KaFai Lau /* init sock_ops */ 4890813a841SMartin KaFai Lau memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp)); 4900813a841SMartin KaFai Lau 4910813a841SMartin KaFai Lau sock_ops.op = BPF_SOCK_OPS_HDR_OPT_LEN_CB; 4920813a841SMartin KaFai Lau 4930813a841SMartin KaFai Lau if (req) { 4940813a841SMartin KaFai Lau /* The listen "sk" cannot be passed here because 4950813a841SMartin KaFai Lau * it is not locked. It would not make too much 4960813a841SMartin KaFai Lau * sense to do bpf_setsockopt(listen_sk) based 4970813a841SMartin KaFai Lau * on individual connection request also. 4980813a841SMartin KaFai Lau * 4990813a841SMartin KaFai Lau * Thus, "req" is passed here and the cgroup-bpf-progs 5000813a841SMartin KaFai Lau * of the listen "sk" will be run. 5010813a841SMartin KaFai Lau * 5020813a841SMartin KaFai Lau * "req" is also used here for fastopen even the "sk" here is 5030813a841SMartin KaFai Lau * a fullsock "child" sk. It is to keep the behavior 5040813a841SMartin KaFai Lau * consistent between fastopen and non-fastopen on 5050813a841SMartin KaFai Lau * the bpf programming side. 506331fca43SMartin KaFai Lau */ 5070813a841SMartin KaFai Lau sock_ops.sk = (struct sock *)req; 5080813a841SMartin KaFai Lau sock_ops.syn_skb = syn_skb; 5090813a841SMartin KaFai Lau } else { 5100813a841SMartin KaFai Lau sock_owned_by_me(sk); 5110813a841SMartin KaFai Lau 5120813a841SMartin KaFai Lau sock_ops.is_fullsock = 1; 5130813a841SMartin KaFai Lau sock_ops.sk = sk; 5140813a841SMartin KaFai Lau } 5150813a841SMartin KaFai Lau 5160813a841SMartin KaFai Lau sock_ops.args[0] = bpf_skops_write_hdr_opt_arg0(skb, synack_type); 5170813a841SMartin KaFai Lau sock_ops.remaining_opt_len = *remaining; 5180813a841SMartin KaFai Lau /* tcp_current_mss() does not pass a skb */ 5190813a841SMartin KaFai Lau if (skb) 5200813a841SMartin KaFai Lau bpf_skops_init_skb(&sock_ops, skb, 0); 5210813a841SMartin KaFai Lau 5220813a841SMartin KaFai Lau err = BPF_CGROUP_RUN_PROG_SOCK_OPS_SK(&sock_ops, sk); 5230813a841SMartin KaFai Lau 5240813a841SMartin KaFai Lau if (err || sock_ops.remaining_opt_len == *remaining) 5250813a841SMartin KaFai Lau return; 5260813a841SMartin KaFai Lau 5270813a841SMartin KaFai Lau opts->bpf_opt_len = *remaining - sock_ops.remaining_opt_len; 5280813a841SMartin KaFai Lau /* round up to 4 bytes */ 5290813a841SMartin KaFai Lau opts->bpf_opt_len = (opts->bpf_opt_len + 3) & ~3; 5300813a841SMartin KaFai Lau 5310813a841SMartin KaFai Lau *remaining -= opts->bpf_opt_len; 532331fca43SMartin KaFai Lau } 533331fca43SMartin KaFai Lau 534331fca43SMartin KaFai Lau static void bpf_skops_write_hdr_opt(struct sock *sk, struct sk_buff *skb, 535331fca43SMartin KaFai Lau struct request_sock *req, 536331fca43SMartin KaFai Lau struct sk_buff *syn_skb, 537331fca43SMartin KaFai Lau enum tcp_synack_type synack_type, 538331fca43SMartin KaFai Lau struct tcp_out_options *opts) 539331fca43SMartin KaFai Lau { 5400813a841SMartin KaFai Lau u8 first_opt_off, nr_written, max_opt_len = opts->bpf_opt_len; 5410813a841SMartin KaFai Lau struct bpf_sock_ops_kern sock_ops; 5420813a841SMartin KaFai Lau int err; 5430813a841SMartin KaFai Lau 5440813a841SMartin KaFai Lau if (likely(!max_opt_len)) 545331fca43SMartin KaFai Lau return; 546331fca43SMartin KaFai Lau 5470813a841SMartin KaFai Lau memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp)); 5480813a841SMartin KaFai Lau 5490813a841SMartin KaFai Lau sock_ops.op = BPF_SOCK_OPS_WRITE_HDR_OPT_CB; 5500813a841SMartin KaFai Lau 5510813a841SMartin KaFai Lau if (req) { 5520813a841SMartin KaFai Lau sock_ops.sk = (struct sock *)req; 5530813a841SMartin KaFai Lau sock_ops.syn_skb = syn_skb; 5540813a841SMartin KaFai Lau } else { 5550813a841SMartin KaFai Lau sock_owned_by_me(sk); 5560813a841SMartin KaFai Lau 5570813a841SMartin KaFai Lau sock_ops.is_fullsock = 1; 5580813a841SMartin KaFai Lau sock_ops.sk = sk; 5590813a841SMartin KaFai Lau } 5600813a841SMartin KaFai Lau 5610813a841SMartin KaFai Lau sock_ops.args[0] = bpf_skops_write_hdr_opt_arg0(skb, synack_type); 5620813a841SMartin KaFai Lau sock_ops.remaining_opt_len = max_opt_len; 5630813a841SMartin KaFai Lau first_opt_off = tcp_hdrlen(skb) - max_opt_len; 5640813a841SMartin KaFai Lau bpf_skops_init_skb(&sock_ops, skb, first_opt_off); 5650813a841SMartin KaFai Lau 5660813a841SMartin KaFai Lau err = BPF_CGROUP_RUN_PROG_SOCK_OPS_SK(&sock_ops, sk); 5670813a841SMartin KaFai Lau 5680813a841SMartin KaFai Lau if (err) 5690813a841SMartin KaFai Lau nr_written = 0; 5700813a841SMartin KaFai Lau else 5710813a841SMartin KaFai Lau nr_written = max_opt_len - sock_ops.remaining_opt_len; 5720813a841SMartin KaFai Lau 5730813a841SMartin KaFai Lau if (nr_written < max_opt_len) 5740813a841SMartin KaFai Lau memset(skb->data + first_opt_off + nr_written, TCPOPT_NOP, 5750813a841SMartin KaFai Lau max_opt_len - nr_written); 576331fca43SMartin KaFai Lau } 577331fca43SMartin KaFai Lau #else 578331fca43SMartin KaFai Lau static void bpf_skops_hdr_opt_len(struct sock *sk, struct sk_buff *skb, 579331fca43SMartin KaFai Lau struct request_sock *req, 580331fca43SMartin KaFai Lau struct sk_buff *syn_skb, 581331fca43SMartin KaFai Lau enum tcp_synack_type synack_type, 582331fca43SMartin KaFai Lau struct tcp_out_options *opts, 583331fca43SMartin KaFai Lau unsigned int *remaining) 584331fca43SMartin KaFai Lau { 585331fca43SMartin KaFai Lau } 586331fca43SMartin KaFai Lau 587331fca43SMartin KaFai Lau static void bpf_skops_write_hdr_opt(struct sock *sk, struct sk_buff *skb, 588331fca43SMartin KaFai Lau struct request_sock *req, 589331fca43SMartin KaFai Lau struct sk_buff *syn_skb, 590331fca43SMartin KaFai Lau enum tcp_synack_type synack_type, 591331fca43SMartin KaFai Lau struct tcp_out_options *opts) 592331fca43SMartin KaFai Lau { 593331fca43SMartin KaFai Lau } 594331fca43SMartin KaFai Lau #endif 595331fca43SMartin KaFai Lau 59667edfef7SAndi Kleen /* Write previously computed TCP options to the packet. 59767edfef7SAndi Kleen * 59867edfef7SAndi Kleen * Beware: Something in the Internet is very sensitive to the ordering of 599fd6149d3SIlpo Järvinen * TCP options, we learned this through the hard way, so be careful here. 600fd6149d3SIlpo Järvinen * Luckily we can at least blame others for their non-compliance but from 6018e3bff96Sstephen hemminger * inter-operability perspective it seems that we're somewhat stuck with 602fd6149d3SIlpo Järvinen * the ordering which we have been using if we want to keep working with 603fd6149d3SIlpo Järvinen * those broken things (not that it currently hurts anybody as there isn't 604fd6149d3SIlpo Järvinen * particular reason why the ordering would need to be changed). 605fd6149d3SIlpo Järvinen * 606fd6149d3SIlpo Järvinen * At least SACK_PERM as the first option is known to lead to a disaster 607fd6149d3SIlpo Järvinen * (but it may well be that other scenarios fail similarly). 608fd6149d3SIlpo Järvinen */ 609ea66758cSPaolo Abeni static void tcp_options_write(struct tcphdr *th, struct tcp_sock *tp, 610bd0388aeSWilliam Allen Simpson struct tcp_out_options *opts) 611bd0388aeSWilliam Allen Simpson { 612ea66758cSPaolo Abeni __be32 *ptr = (__be32 *)(th + 1); 6132100c8d2SYuchung Cheng u16 options = opts->options; /* mungable copy */ 614bd0388aeSWilliam Allen Simpson 615bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_MD5 & options)) { 6161a2c6181SChristoph Paasch *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | 6171a2c6181SChristoph Paasch (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG); 618bd0388aeSWilliam Allen Simpson /* overload cookie hash location */ 619bd0388aeSWilliam Allen Simpson opts->hash_location = (__u8 *)ptr; 62033ad798cSAdam Langley ptr += 4; 62133ad798cSAdam Langley } 62233ad798cSAdam Langley 623fd6149d3SIlpo Järvinen if (unlikely(opts->mss)) { 624fd6149d3SIlpo Järvinen *ptr++ = htonl((TCPOPT_MSS << 24) | 625fd6149d3SIlpo Järvinen (TCPOLEN_MSS << 16) | 626fd6149d3SIlpo Järvinen opts->mss); 627fd6149d3SIlpo Järvinen } 628fd6149d3SIlpo Järvinen 629bd0388aeSWilliam Allen Simpson if (likely(OPTION_TS & options)) { 630bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 63133ad798cSAdam Langley *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | 63233ad798cSAdam Langley (TCPOLEN_SACK_PERM << 16) | 63333ad798cSAdam Langley (TCPOPT_TIMESTAMP << 8) | 63433ad798cSAdam Langley TCPOLEN_TIMESTAMP); 635bd0388aeSWilliam Allen Simpson options &= ~OPTION_SACK_ADVERTISE; 63633ad798cSAdam Langley } else { 637496c98dfSYOSHIFUJI Hideaki *ptr++ = htonl((TCPOPT_NOP << 24) | 63840efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 63940efc6faSStephen Hemminger (TCPOPT_TIMESTAMP << 8) | 64040efc6faSStephen Hemminger TCPOLEN_TIMESTAMP); 64140efc6faSStephen Hemminger } 64233ad798cSAdam Langley *ptr++ = htonl(opts->tsval); 64333ad798cSAdam Langley *ptr++ = htonl(opts->tsecr); 64433ad798cSAdam Langley } 64533ad798cSAdam Langley 646bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 64733ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 64833ad798cSAdam Langley (TCPOPT_NOP << 16) | 64933ad798cSAdam Langley (TCPOPT_SACK_PERM << 8) | 65033ad798cSAdam Langley TCPOLEN_SACK_PERM); 65133ad798cSAdam Langley } 65233ad798cSAdam Langley 653bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_WSCALE & options)) { 65433ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 65533ad798cSAdam Langley (TCPOPT_WINDOW << 16) | 65633ad798cSAdam Langley (TCPOLEN_WINDOW << 8) | 65733ad798cSAdam Langley opts->ws); 65833ad798cSAdam Langley } 65933ad798cSAdam Langley 66033ad798cSAdam Langley if (unlikely(opts->num_sack_blocks)) { 66133ad798cSAdam Langley struct tcp_sack_block *sp = tp->rx_opt.dsack ? 66233ad798cSAdam Langley tp->duplicate_sack : tp->selective_acks; 66340efc6faSStephen Hemminger int this_sack; 66440efc6faSStephen Hemminger 66540efc6faSStephen Hemminger *ptr++ = htonl((TCPOPT_NOP << 24) | 66640efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 66740efc6faSStephen Hemminger (TCPOPT_SACK << 8) | 66833ad798cSAdam Langley (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * 66940efc6faSStephen Hemminger TCPOLEN_SACK_PERBLOCK))); 6702de979bdSStephen Hemminger 67133ad798cSAdam Langley for (this_sack = 0; this_sack < opts->num_sack_blocks; 67233ad798cSAdam Langley ++this_sack) { 67340efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].start_seq); 67440efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].end_seq); 67540efc6faSStephen Hemminger } 6762de979bdSStephen Hemminger 67740efc6faSStephen Hemminger tp->rx_opt.dsack = 0; 67840efc6faSStephen Hemminger } 6792100c8d2SYuchung Cheng 6802100c8d2SYuchung Cheng if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) { 6812100c8d2SYuchung Cheng struct tcp_fastopen_cookie *foc = opts->fastopen_cookie; 6827f9b838bSDaniel Lee u8 *p = (u8 *)ptr; 6837f9b838bSDaniel Lee u32 len; /* Fast Open option length */ 6842100c8d2SYuchung Cheng 6857f9b838bSDaniel Lee if (foc->exp) { 6867f9b838bSDaniel Lee len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len; 6877f9b838bSDaniel Lee *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) | 6882100c8d2SYuchung Cheng TCPOPT_FASTOPEN_MAGIC); 6897f9b838bSDaniel Lee p += TCPOLEN_EXP_FASTOPEN_BASE; 6907f9b838bSDaniel Lee } else { 6917f9b838bSDaniel Lee len = TCPOLEN_FASTOPEN_BASE + foc->len; 6927f9b838bSDaniel Lee *p++ = TCPOPT_FASTOPEN; 6937f9b838bSDaniel Lee *p++ = len; 6942100c8d2SYuchung Cheng } 6957f9b838bSDaniel Lee 6967f9b838bSDaniel Lee memcpy(p, foc->val, foc->len); 6977f9b838bSDaniel Lee if ((len & 3) == 2) { 6987f9b838bSDaniel Lee p[foc->len] = TCPOPT_NOP; 6997f9b838bSDaniel Lee p[foc->len + 1] = TCPOPT_NOP; 7007f9b838bSDaniel Lee } 7017f9b838bSDaniel Lee ptr += (len + 3) >> 2; 7022100c8d2SYuchung Cheng } 70360e2a778SUrsula Braun 70460e2a778SUrsula Braun smc_options_write(ptr, &options); 705eda7acddSPeter Krystad 706ea66758cSPaolo Abeni mptcp_options_write(th, ptr, tp, opts); 70760e2a778SUrsula Braun } 70860e2a778SUrsula Braun 70960e2a778SUrsula Braun static void smc_set_option(const struct tcp_sock *tp, 71060e2a778SUrsula Braun struct tcp_out_options *opts, 71160e2a778SUrsula Braun unsigned int *remaining) 71260e2a778SUrsula Braun { 71360e2a778SUrsula Braun #if IS_ENABLED(CONFIG_SMC) 71460e2a778SUrsula Braun if (static_branch_unlikely(&tcp_have_smc)) { 71560e2a778SUrsula Braun if (tp->syn_smc) { 71660e2a778SUrsula Braun if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) { 71760e2a778SUrsula Braun opts->options |= OPTION_SMC; 71860e2a778SUrsula Braun *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED; 71960e2a778SUrsula Braun } 72060e2a778SUrsula Braun } 72160e2a778SUrsula Braun } 72260e2a778SUrsula Braun #endif 72360e2a778SUrsula Braun } 72460e2a778SUrsula Braun 72560e2a778SUrsula Braun static void smc_set_option_cond(const struct tcp_sock *tp, 72660e2a778SUrsula Braun const struct inet_request_sock *ireq, 72760e2a778SUrsula Braun struct tcp_out_options *opts, 72860e2a778SUrsula Braun unsigned int *remaining) 72960e2a778SUrsula Braun { 73060e2a778SUrsula Braun #if IS_ENABLED(CONFIG_SMC) 73160e2a778SUrsula Braun if (static_branch_unlikely(&tcp_have_smc)) { 73260e2a778SUrsula Braun if (tp->syn_smc && ireq->smc_ok) { 73360e2a778SUrsula Braun if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) { 73460e2a778SUrsula Braun opts->options |= OPTION_SMC; 73560e2a778SUrsula Braun *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED; 73660e2a778SUrsula Braun } 73760e2a778SUrsula Braun } 73860e2a778SUrsula Braun } 73960e2a778SUrsula Braun #endif 74040efc6faSStephen Hemminger } 74140efc6faSStephen Hemminger 742cec37a6eSPeter Krystad static void mptcp_set_option_cond(const struct request_sock *req, 743cec37a6eSPeter Krystad struct tcp_out_options *opts, 744cec37a6eSPeter Krystad unsigned int *remaining) 745cec37a6eSPeter Krystad { 746cec37a6eSPeter Krystad if (rsk_is_mptcp(req)) { 747cec37a6eSPeter Krystad unsigned int size; 748cec37a6eSPeter Krystad 749cec37a6eSPeter Krystad if (mptcp_synack_options(req, &size, &opts->mptcp)) { 750cec37a6eSPeter Krystad if (*remaining >= size) { 751cec37a6eSPeter Krystad opts->options |= OPTION_MPTCP; 752cec37a6eSPeter Krystad *remaining -= size; 753cec37a6eSPeter Krystad } 754cec37a6eSPeter Krystad } 755cec37a6eSPeter Krystad } 756cec37a6eSPeter Krystad } 757cec37a6eSPeter Krystad 75867edfef7SAndi Kleen /* Compute TCP options for SYN packets. This is not the final 75967edfef7SAndi Kleen * network wire format yet. 76067edfef7SAndi Kleen */ 76195c96174SEric Dumazet static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb, 76233ad798cSAdam Langley struct tcp_out_options *opts, 763cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 764cf533ea5SEric Dumazet { 76533ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 76695c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 767783237e8SYuchung Cheng struct tcp_fastopen_request *fastopen = tp->fastopen_req; 76833ad798cSAdam Langley 7698c2320e8SEric Dumazet *md5 = NULL; 770cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 771459837b5SDmitry Safonov if (static_branch_unlikely(&tcp_md5_needed.key) && 7726015c71eSEric Dumazet rcu_access_pointer(tp->md5sig_info)) { 77333ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 77433ad798cSAdam Langley if (*md5) { 77533ad798cSAdam Langley opts->options |= OPTION_MD5; 776bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 777cfb6eeb4SYOSHIFUJI Hideaki } 7788c2320e8SEric Dumazet } 779cfb6eeb4SYOSHIFUJI Hideaki #endif 78033ad798cSAdam Langley 78133ad798cSAdam Langley /* We always get an MSS option. The option bytes which will be seen in 78233ad798cSAdam Langley * normal data packets should timestamps be used, must be in the MSS 78333ad798cSAdam Langley * advertised. But we subtract them from tp->mss_cache so that 78433ad798cSAdam Langley * calculations in tcp_sendmsg are simpler etc. So account for this 78533ad798cSAdam Langley * fact here if necessary. If we don't do this correctly, as a 78633ad798cSAdam Langley * receiver we won't recognize data packets as being full sized when we 78733ad798cSAdam Langley * should, and thus we won't abide by the delayed ACK rules correctly. 78833ad798cSAdam Langley * SACKs don't matter, we never delay an ACK when we have any of those 78933ad798cSAdam Langley * going out. */ 79033ad798cSAdam Langley opts->mss = tcp_advertise_mss(sk); 791bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 79233ad798cSAdam Langley 7933666f666SKuniyuki Iwashima if (likely(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_timestamps) && !*md5)) { 79433ad798cSAdam Langley opts->options |= OPTION_TS; 7957faee5c0SEric Dumazet opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset; 79633ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 797bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 79833ad798cSAdam Langley } 7993666f666SKuniyuki Iwashima if (likely(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_window_scaling))) { 80033ad798cSAdam Langley opts->ws = tp->rx_opt.rcv_wscale; 80189e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 802bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 80333ad798cSAdam Langley } 8043666f666SKuniyuki Iwashima if (likely(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_sack))) { 80533ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 806b32d1310SDavid S. Miller if (unlikely(!(OPTION_TS & opts->options))) 807bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 80833ad798cSAdam Langley } 80933ad798cSAdam Langley 810783237e8SYuchung Cheng if (fastopen && fastopen->cookie.len >= 0) { 8112646c831SDaniel Lee u32 need = fastopen->cookie.len; 8122646c831SDaniel Lee 8132646c831SDaniel Lee need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE : 8142646c831SDaniel Lee TCPOLEN_FASTOPEN_BASE; 815783237e8SYuchung Cheng need = (need + 3) & ~3U; /* Align to 32 bits */ 816783237e8SYuchung Cheng if (remaining >= need) { 817783237e8SYuchung Cheng opts->options |= OPTION_FAST_OPEN_COOKIE; 818783237e8SYuchung Cheng opts->fastopen_cookie = &fastopen->cookie; 819783237e8SYuchung Cheng remaining -= need; 820783237e8SYuchung Cheng tp->syn_fastopen = 1; 8212646c831SDaniel Lee tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0; 822783237e8SYuchung Cheng } 823783237e8SYuchung Cheng } 824bd0388aeSWilliam Allen Simpson 82560e2a778SUrsula Braun smc_set_option(tp, opts, &remaining); 82660e2a778SUrsula Braun 827cec37a6eSPeter Krystad if (sk_is_mptcp(sk)) { 828cec37a6eSPeter Krystad unsigned int size; 829cec37a6eSPeter Krystad 830cc7972eaSChristoph Paasch if (mptcp_syn_options(sk, skb, &size, &opts->mptcp)) { 831cec37a6eSPeter Krystad opts->options |= OPTION_MPTCP; 832cec37a6eSPeter Krystad remaining -= size; 833cec37a6eSPeter Krystad } 834cec37a6eSPeter Krystad } 835cec37a6eSPeter Krystad 836331fca43SMartin KaFai Lau bpf_skops_hdr_opt_len(sk, skb, NULL, NULL, 0, opts, &remaining); 837331fca43SMartin KaFai Lau 838bd0388aeSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 83933ad798cSAdam Langley } 84033ad798cSAdam Langley 84167edfef7SAndi Kleen /* Set up TCP options for SYN-ACKs. */ 84260e2a778SUrsula Braun static unsigned int tcp_synack_options(const struct sock *sk, 84360e2a778SUrsula Braun struct request_sock *req, 84495c96174SEric Dumazet unsigned int mss, struct sk_buff *skb, 84533ad798cSAdam Langley struct tcp_out_options *opts, 84680f03e27SEric Dumazet const struct tcp_md5sig_key *md5, 847e114e1e8SEric Dumazet struct tcp_fastopen_cookie *foc, 848331fca43SMartin KaFai Lau enum tcp_synack_type synack_type, 849331fca43SMartin KaFai Lau struct sk_buff *syn_skb) 8504957faadSWilliam Allen Simpson { 85133ad798cSAdam Langley struct inet_request_sock *ireq = inet_rsk(req); 85295c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 85333ad798cSAdam Langley 85433ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 85580f03e27SEric Dumazet if (md5) { 85633ad798cSAdam Langley opts->options |= OPTION_MD5; 8574957faadSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 8584957faadSWilliam Allen Simpson 8594957faadSWilliam Allen Simpson /* We can't fit any SACK blocks in a packet with MD5 + TS 8604957faadSWilliam Allen Simpson * options. There was discussion about disabling SACK 8614957faadSWilliam Allen Simpson * rather than TS in order to fit in better with old, 8624957faadSWilliam Allen Simpson * buggy kernels, but that was deemed to be unnecessary. 8634957faadSWilliam Allen Simpson */ 864e114e1e8SEric Dumazet if (synack_type != TCP_SYNACK_COOKIE) 865de213e5eSEric Dumazet ireq->tstamp_ok &= !ireq->sack_ok; 86633ad798cSAdam Langley } 86733ad798cSAdam Langley #endif 86833ad798cSAdam Langley 8694957faadSWilliam Allen Simpson /* We always send an MSS option. */ 87033ad798cSAdam Langley opts->mss = mss; 8714957faadSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 87233ad798cSAdam Langley 87333ad798cSAdam Langley if (likely(ireq->wscale_ok)) { 87433ad798cSAdam Langley opts->ws = ireq->rcv_wscale; 87589e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 8764957faadSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 87733ad798cSAdam Langley } 878de213e5eSEric Dumazet if (likely(ireq->tstamp_ok)) { 87933ad798cSAdam Langley opts->options |= OPTION_TS; 88095a22caeSFlorian Westphal opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off; 88133ad798cSAdam Langley opts->tsecr = req->ts_recent; 8824957faadSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 88333ad798cSAdam Langley } 88433ad798cSAdam Langley if (likely(ireq->sack_ok)) { 88533ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 886de213e5eSEric Dumazet if (unlikely(!ireq->tstamp_ok)) 8874957faadSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 88833ad798cSAdam Langley } 8897f9b838bSDaniel Lee if (foc != NULL && foc->len >= 0) { 8907f9b838bSDaniel Lee u32 need = foc->len; 8917f9b838bSDaniel Lee 8927f9b838bSDaniel Lee need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE : 8937f9b838bSDaniel Lee TCPOLEN_FASTOPEN_BASE; 8948336886fSJerry Chu need = (need + 3) & ~3U; /* Align to 32 bits */ 8958336886fSJerry Chu if (remaining >= need) { 8968336886fSJerry Chu opts->options |= OPTION_FAST_OPEN_COOKIE; 8978336886fSJerry Chu opts->fastopen_cookie = foc; 8988336886fSJerry Chu remaining -= need; 8998336886fSJerry Chu } 9008336886fSJerry Chu } 9014957faadSWilliam Allen Simpson 902cec37a6eSPeter Krystad mptcp_set_option_cond(req, opts, &remaining); 903cec37a6eSPeter Krystad 90460e2a778SUrsula Braun smc_set_option_cond(tcp_sk(sk), ireq, opts, &remaining); 90560e2a778SUrsula Braun 906331fca43SMartin KaFai Lau bpf_skops_hdr_opt_len((struct sock *)sk, skb, req, syn_skb, 907331fca43SMartin KaFai Lau synack_type, opts, &remaining); 908331fca43SMartin KaFai Lau 9094957faadSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 91033ad798cSAdam Langley } 91133ad798cSAdam Langley 91267edfef7SAndi Kleen /* Compute TCP options for ESTABLISHED sockets. This is not the 91367edfef7SAndi Kleen * final wire format yet. 91467edfef7SAndi Kleen */ 91595c96174SEric Dumazet static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb, 91633ad798cSAdam Langley struct tcp_out_options *opts, 917cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 918cf533ea5SEric Dumazet { 91933ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 92095c96174SEric Dumazet unsigned int size = 0; 921cabeccbdSIlpo Järvinen unsigned int eff_sacks; 92233ad798cSAdam Langley 9235843ef42SAndi Kleen opts->options = 0; 9245843ef42SAndi Kleen 9258c2320e8SEric Dumazet *md5 = NULL; 92633ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 927459837b5SDmitry Safonov if (static_branch_unlikely(&tcp_md5_needed.key) && 9286015c71eSEric Dumazet rcu_access_pointer(tp->md5sig_info)) { 92933ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 9308c2320e8SEric Dumazet if (*md5) { 93133ad798cSAdam Langley opts->options |= OPTION_MD5; 93233ad798cSAdam Langley size += TCPOLEN_MD5SIG_ALIGNED; 93333ad798cSAdam Langley } 9348c2320e8SEric Dumazet } 93533ad798cSAdam Langley #endif 93633ad798cSAdam Langley 93733ad798cSAdam Langley if (likely(tp->rx_opt.tstamp_ok)) { 93833ad798cSAdam Langley opts->options |= OPTION_TS; 9397faee5c0SEric Dumazet opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0; 94033ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 94133ad798cSAdam Langley size += TCPOLEN_TSTAMP_ALIGNED; 94233ad798cSAdam Langley } 94333ad798cSAdam Langley 944cec37a6eSPeter Krystad /* MPTCP options have precedence over SACK for the limited TCP 945cec37a6eSPeter Krystad * option space because a MPTCP connection would be forced to 946cec37a6eSPeter Krystad * fall back to regular TCP if a required multipath option is 947cec37a6eSPeter Krystad * missing. SACK still gets a chance to use whatever space is 948cec37a6eSPeter Krystad * left. 949cec37a6eSPeter Krystad */ 950cec37a6eSPeter Krystad if (sk_is_mptcp(sk)) { 951cec37a6eSPeter Krystad unsigned int remaining = MAX_TCP_OPTION_SPACE - size; 952cec37a6eSPeter Krystad unsigned int opt_size = 0; 953cec37a6eSPeter Krystad 954cec37a6eSPeter Krystad if (mptcp_established_options(sk, skb, &opt_size, remaining, 955cec37a6eSPeter Krystad &opts->mptcp)) { 956cec37a6eSPeter Krystad opts->options |= OPTION_MPTCP; 957cec37a6eSPeter Krystad size += opt_size; 958cec37a6eSPeter Krystad } 959cec37a6eSPeter Krystad } 960cec37a6eSPeter Krystad 961cabeccbdSIlpo Järvinen eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; 962cabeccbdSIlpo Järvinen if (unlikely(eff_sacks)) { 96395c96174SEric Dumazet const unsigned int remaining = MAX_TCP_OPTION_SPACE - size; 9649cfcca23SMat Martineau if (unlikely(remaining < TCPOLEN_SACK_BASE_ALIGNED + 9659cfcca23SMat Martineau TCPOLEN_SACK_PERBLOCK)) 9669cfcca23SMat Martineau return size; 9679cfcca23SMat Martineau 96833ad798cSAdam Langley opts->num_sack_blocks = 96995c96174SEric Dumazet min_t(unsigned int, eff_sacks, 97033ad798cSAdam Langley (remaining - TCPOLEN_SACK_BASE_ALIGNED) / 97133ad798cSAdam Langley TCPOLEN_SACK_PERBLOCK); 9729cfcca23SMat Martineau 97333ad798cSAdam Langley size += TCPOLEN_SACK_BASE_ALIGNED + 97433ad798cSAdam Langley opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; 97533ad798cSAdam Langley } 97633ad798cSAdam Langley 977331fca43SMartin KaFai Lau if (unlikely(BPF_SOCK_OPS_TEST_FLAG(tp, 978331fca43SMartin KaFai Lau BPF_SOCK_OPS_WRITE_HDR_OPT_CB_FLAG))) { 979331fca43SMartin KaFai Lau unsigned int remaining = MAX_TCP_OPTION_SPACE - size; 980331fca43SMartin KaFai Lau 981331fca43SMartin KaFai Lau bpf_skops_hdr_opt_len(sk, skb, NULL, NULL, 0, opts, &remaining); 982331fca43SMartin KaFai Lau 983331fca43SMartin KaFai Lau size = MAX_TCP_OPTION_SPACE - remaining; 984331fca43SMartin KaFai Lau } 985331fca43SMartin KaFai Lau 98633ad798cSAdam Langley return size; 98740efc6faSStephen Hemminger } 9881da177e4SLinus Torvalds 98946d3ceabSEric Dumazet 99046d3ceabSEric Dumazet /* TCP SMALL QUEUES (TSQ) 99146d3ceabSEric Dumazet * 99246d3ceabSEric Dumazet * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev) 99346d3ceabSEric Dumazet * to reduce RTT and bufferbloat. 99446d3ceabSEric Dumazet * We do this using a special skb destructor (tcp_wfree). 99546d3ceabSEric Dumazet * 99646d3ceabSEric Dumazet * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb 99746d3ceabSEric Dumazet * needs to be reallocated in a driver. 9988e3bff96Sstephen hemminger * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc 99946d3ceabSEric Dumazet * 100046d3ceabSEric Dumazet * Since transmit from skb destructor is forbidden, we use a tasklet 100146d3ceabSEric Dumazet * to process all sockets that eventually need to send more skbs. 100246d3ceabSEric Dumazet * We use one tasklet per cpu, with its own queue of sockets. 100346d3ceabSEric Dumazet */ 100446d3ceabSEric Dumazet struct tsq_tasklet { 100546d3ceabSEric Dumazet struct tasklet_struct tasklet; 100646d3ceabSEric Dumazet struct list_head head; /* queue of tcp sockets */ 100746d3ceabSEric Dumazet }; 100846d3ceabSEric Dumazet static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet); 100946d3ceabSEric Dumazet 101073a6bab5SEric Dumazet static void tcp_tsq_write(struct sock *sk) 10116f458dfbSEric Dumazet { 10126f458dfbSEric Dumazet if ((1 << sk->sk_state) & 10136f458dfbSEric Dumazet (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING | 1014f9616c35SEric Dumazet TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) { 1015f9616c35SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 1016f9616c35SEric Dumazet 1017f9616c35SEric Dumazet if (tp->lost_out > tp->retrans_out && 101840570375SEric Dumazet tcp_snd_cwnd(tp) > tcp_packets_in_flight(tp)) { 10193a91d29fSKoichiro Den tcp_mstamp_refresh(tp); 1020f9616c35SEric Dumazet tcp_xmit_retransmit_queue(sk); 10213a91d29fSKoichiro Den } 1022f9616c35SEric Dumazet 1023f9616c35SEric Dumazet tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle, 1024bf06200eSJohn Ogness 0, GFP_ATOMIC); 10256f458dfbSEric Dumazet } 1026f9616c35SEric Dumazet } 102773a6bab5SEric Dumazet 102873a6bab5SEric Dumazet static void tcp_tsq_handler(struct sock *sk) 102973a6bab5SEric Dumazet { 103073a6bab5SEric Dumazet bh_lock_sock(sk); 103173a6bab5SEric Dumazet if (!sock_owned_by_user(sk)) 103273a6bab5SEric Dumazet tcp_tsq_write(sk); 103373a6bab5SEric Dumazet else if (!test_and_set_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags)) 103473a6bab5SEric Dumazet sock_hold(sk); 103573a6bab5SEric Dumazet bh_unlock_sock(sk); 103673a6bab5SEric Dumazet } 103746d3ceabSEric Dumazet /* 10388e3bff96Sstephen hemminger * One tasklet per cpu tries to send more skbs. 103946d3ceabSEric Dumazet * We run in tasklet context but need to disable irqs when 10408e3bff96Sstephen hemminger * transferring tsq->head because tcp_wfree() might 104146d3ceabSEric Dumazet * interrupt us (non NAPI drivers) 104246d3ceabSEric Dumazet */ 1043c6533ca8SAllen Pais static void tcp_tasklet_func(struct tasklet_struct *t) 104446d3ceabSEric Dumazet { 1045c6533ca8SAllen Pais struct tsq_tasklet *tsq = from_tasklet(tsq, t, tasklet); 104646d3ceabSEric Dumazet LIST_HEAD(list); 104746d3ceabSEric Dumazet unsigned long flags; 104846d3ceabSEric Dumazet struct list_head *q, *n; 104946d3ceabSEric Dumazet struct tcp_sock *tp; 105046d3ceabSEric Dumazet struct sock *sk; 105146d3ceabSEric Dumazet 105246d3ceabSEric Dumazet local_irq_save(flags); 105346d3ceabSEric Dumazet list_splice_init(&tsq->head, &list); 105446d3ceabSEric Dumazet local_irq_restore(flags); 105546d3ceabSEric Dumazet 105646d3ceabSEric Dumazet list_for_each_safe(q, n, &list) { 105746d3ceabSEric Dumazet tp = list_entry(q, struct tcp_sock, tsq_node); 105846d3ceabSEric Dumazet list_del(&tp->tsq_node); 105946d3ceabSEric Dumazet 106046d3ceabSEric Dumazet sk = (struct sock *)tp; 10610a9648f1SEric Dumazet smp_mb__before_atomic(); 10627aa5470cSEric Dumazet clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags); 10637aa5470cSEric Dumazet 10646f458dfbSEric Dumazet tcp_tsq_handler(sk); 106546d3ceabSEric Dumazet sk_free(sk); 106646d3ceabSEric Dumazet } 106746d3ceabSEric Dumazet } 106846d3ceabSEric Dumazet 106940fc3423SEric Dumazet #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \ 107040fc3423SEric Dumazet TCPF_WRITE_TIMER_DEFERRED | \ 107140fc3423SEric Dumazet TCPF_DELACK_TIMER_DEFERRED | \ 107240fc3423SEric Dumazet TCPF_MTU_REDUCED_DEFERRED) 107346d3ceabSEric Dumazet /** 107446d3ceabSEric Dumazet * tcp_release_cb - tcp release_sock() callback 107546d3ceabSEric Dumazet * @sk: socket 107646d3ceabSEric Dumazet * 107746d3ceabSEric Dumazet * called from release_sock() to perform protocol dependent 107846d3ceabSEric Dumazet * actions before socket release. 107946d3ceabSEric Dumazet */ 108046d3ceabSEric Dumazet void tcp_release_cb(struct sock *sk) 108146d3ceabSEric Dumazet { 1082fac30731SEric Dumazet unsigned long flags = smp_load_acquire(&sk->sk_tsq_flags); 1083fac30731SEric Dumazet unsigned long nflags; 108446d3ceabSEric Dumazet 10856f458dfbSEric Dumazet /* perform an atomic operation only if at least one flag is set */ 10866f458dfbSEric Dumazet do { 10876f458dfbSEric Dumazet if (!(flags & TCP_DEFERRED_ALL)) 10886f458dfbSEric Dumazet return; 10896f458dfbSEric Dumazet nflags = flags & ~TCP_DEFERRED_ALL; 1090fac30731SEric Dumazet } while (!try_cmpxchg(&sk->sk_tsq_flags, &flags, nflags)); 10916f458dfbSEric Dumazet 109273a6bab5SEric Dumazet if (flags & TCPF_TSQ_DEFERRED) { 109373a6bab5SEric Dumazet tcp_tsq_write(sk); 109473a6bab5SEric Dumazet __sock_put(sk); 109573a6bab5SEric Dumazet } 1096c3f9b018SEric Dumazet /* Here begins the tricky part : 1097c3f9b018SEric Dumazet * We are called from release_sock() with : 1098c3f9b018SEric Dumazet * 1) BH disabled 1099c3f9b018SEric Dumazet * 2) sk_lock.slock spinlock held 1100c3f9b018SEric Dumazet * 3) socket owned by us (sk->sk_lock.owned == 1) 1101c3f9b018SEric Dumazet * 1102c3f9b018SEric Dumazet * But following code is meant to be called from BH handlers, 1103c3f9b018SEric Dumazet * so we should keep BH disabled, but early release socket ownership 1104c3f9b018SEric Dumazet */ 1105c3f9b018SEric Dumazet sock_release_ownership(sk); 1106c3f9b018SEric Dumazet 110740fc3423SEric Dumazet if (flags & TCPF_WRITE_TIMER_DEFERRED) { 11086f458dfbSEric Dumazet tcp_write_timer_handler(sk); 1109144d56e9SEric Dumazet __sock_put(sk); 1110144d56e9SEric Dumazet } 111140fc3423SEric Dumazet if (flags & TCPF_DELACK_TIMER_DEFERRED) { 11126f458dfbSEric Dumazet tcp_delack_timer_handler(sk); 1113144d56e9SEric Dumazet __sock_put(sk); 1114144d56e9SEric Dumazet } 111540fc3423SEric Dumazet if (flags & TCPF_MTU_REDUCED_DEFERRED) { 11164fab9071SNeal Cardwell inet_csk(sk)->icsk_af_ops->mtu_reduced(sk); 1117144d56e9SEric Dumazet __sock_put(sk); 1118144d56e9SEric Dumazet } 111946d3ceabSEric Dumazet } 112046d3ceabSEric Dumazet EXPORT_SYMBOL(tcp_release_cb); 112146d3ceabSEric Dumazet 112246d3ceabSEric Dumazet void __init tcp_tasklet_init(void) 112346d3ceabSEric Dumazet { 112446d3ceabSEric Dumazet int i; 112546d3ceabSEric Dumazet 112646d3ceabSEric Dumazet for_each_possible_cpu(i) { 112746d3ceabSEric Dumazet struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i); 112846d3ceabSEric Dumazet 112946d3ceabSEric Dumazet INIT_LIST_HEAD(&tsq->head); 1130c6533ca8SAllen Pais tasklet_setup(&tsq->tasklet, tcp_tasklet_func); 113146d3ceabSEric Dumazet } 113246d3ceabSEric Dumazet } 113346d3ceabSEric Dumazet 113446d3ceabSEric Dumazet /* 113546d3ceabSEric Dumazet * Write buffer destructor automatically called from kfree_skb. 11368e3bff96Sstephen hemminger * We can't xmit new skbs from this context, as we might already 113746d3ceabSEric Dumazet * hold qdisc lock. 113846d3ceabSEric Dumazet */ 1139d6a4a104SEric Dumazet void tcp_wfree(struct sk_buff *skb) 114046d3ceabSEric Dumazet { 114146d3ceabSEric Dumazet struct sock *sk = skb->sk; 114246d3ceabSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 1143408f0a6cSEric Dumazet unsigned long flags, nval, oval; 1144b548b17aSEric Dumazet struct tsq_tasklet *tsq; 1145b548b17aSEric Dumazet bool empty; 11469b462d02SEric Dumazet 11479b462d02SEric Dumazet /* Keep one reference on sk_wmem_alloc. 11489b462d02SEric Dumazet * Will be released by sk_free() from here or tcp_tasklet_func() 11499b462d02SEric Dumazet */ 115014afee4bSReshetova, Elena WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc)); 11519b462d02SEric Dumazet 11529b462d02SEric Dumazet /* If this softirq is serviced by ksoftirqd, we are likely under stress. 11539b462d02SEric Dumazet * Wait until our queues (qdisc + devices) are drained. 11549b462d02SEric Dumazet * This gives : 11559b462d02SEric Dumazet * - less callbacks to tcp_write_xmit(), reducing stress (batches) 11569b462d02SEric Dumazet * - chance for incoming ACK (processed by another cpu maybe) 11579b462d02SEric Dumazet * to migrate this flow (skb->ooo_okay will be eventually set) 11589b462d02SEric Dumazet */ 115914afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current) 11609b462d02SEric Dumazet goto out; 116146d3ceabSEric Dumazet 1162b548b17aSEric Dumazet oval = smp_load_acquire(&sk->sk_tsq_flags); 1163b548b17aSEric Dumazet do { 1164408f0a6cSEric Dumazet if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED)) 1165408f0a6cSEric Dumazet goto out; 1166408f0a6cSEric Dumazet 116773a6bab5SEric Dumazet nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED; 1168b548b17aSEric Dumazet } while (!try_cmpxchg(&sk->sk_tsq_flags, &oval, nval)); 1169408f0a6cSEric Dumazet 117046d3ceabSEric Dumazet /* queue this socket to tasklet queue */ 117146d3ceabSEric Dumazet local_irq_save(flags); 1172903ceff7SChristoph Lameter tsq = this_cpu_ptr(&tsq_tasklet); 1173a9b204d1SEric Dumazet empty = list_empty(&tsq->head); 117446d3ceabSEric Dumazet list_add(&tp->tsq_node, &tsq->head); 1175a9b204d1SEric Dumazet if (empty) 117646d3ceabSEric Dumazet tasklet_schedule(&tsq->tasklet); 117746d3ceabSEric Dumazet local_irq_restore(flags); 11789b462d02SEric Dumazet return; 11799b462d02SEric Dumazet out: 11809b462d02SEric Dumazet sk_free(sk); 118146d3ceabSEric Dumazet } 118246d3ceabSEric Dumazet 118373a6bab5SEric Dumazet /* Note: Called under soft irq. 118473a6bab5SEric Dumazet * We can call TCP stack right away, unless socket is owned by user. 1185218af599SEric Dumazet */ 1186218af599SEric Dumazet enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer) 1187218af599SEric Dumazet { 1188218af599SEric Dumazet struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer); 1189218af599SEric Dumazet struct sock *sk = (struct sock *)tp; 1190218af599SEric Dumazet 119173a6bab5SEric Dumazet tcp_tsq_handler(sk); 119273a6bab5SEric Dumazet sock_put(sk); 1193218af599SEric Dumazet 1194218af599SEric Dumazet return HRTIMER_NORESTART; 1195218af599SEric Dumazet } 1196218af599SEric Dumazet 1197a7a25630SEric Dumazet static void tcp_update_skb_after_send(struct sock *sk, struct sk_buff *skb, 1198a7a25630SEric Dumazet u64 prior_wstamp) 1199e2080072SEric Dumazet { 1200ab408b6dSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 1201ab408b6dSEric Dumazet 1202ab408b6dSEric Dumazet if (sk->sk_pacing_status != SK_PACING_NONE) { 120376a9ebe8SEric Dumazet unsigned long rate = sk->sk_pacing_rate; 1204ab408b6dSEric Dumazet 1205ab408b6dSEric Dumazet /* Original sch_fq does not pace first 10 MSS 1206ab408b6dSEric Dumazet * Note that tp->data_segs_out overflows after 2^32 packets, 1207ab408b6dSEric Dumazet * this is a minor annoyance. 1208ab408b6dSEric Dumazet */ 120976a9ebe8SEric Dumazet if (rate != ~0UL && rate && tp->data_segs_out >= 10) { 1210a7a25630SEric Dumazet u64 len_ns = div64_ul((u64)skb->len * NSEC_PER_SEC, rate); 1211a7a25630SEric Dumazet u64 credit = tp->tcp_wstamp_ns - prior_wstamp; 1212a7a25630SEric Dumazet 1213a7a25630SEric Dumazet /* take into account OS jitter */ 1214a7a25630SEric Dumazet len_ns -= min_t(u64, len_ns / 2, credit); 1215a7a25630SEric Dumazet tp->tcp_wstamp_ns += len_ns; 1216ab408b6dSEric Dumazet } 1217ab408b6dSEric Dumazet } 1218e2080072SEric Dumazet list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue); 1219e2080072SEric Dumazet } 1220e2080072SEric Dumazet 122105e22e83SEric Dumazet INDIRECT_CALLABLE_DECLARE(int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl)); 122205e22e83SEric Dumazet INDIRECT_CALLABLE_DECLARE(int inet6_csk_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl)); 1223dd2e0b86SEric Dumazet INDIRECT_CALLABLE_DECLARE(void tcp_v4_send_check(struct sock *sk, struct sk_buff *skb)); 122405e22e83SEric Dumazet 12251da177e4SLinus Torvalds /* This routine actually transmits TCP packets queued in by 12261da177e4SLinus Torvalds * tcp_do_sendmsg(). This is used by both the initial 12271da177e4SLinus Torvalds * transmission and possible later retransmissions. 12281da177e4SLinus Torvalds * All SKB's seen here are completely headerless. It is our 12291da177e4SLinus Torvalds * job to build the TCP header, and pass the packet down to 12301da177e4SLinus Torvalds * IP so it can do the same plus pass the packet off to the 12311da177e4SLinus Torvalds * device. 12321da177e4SLinus Torvalds * 12331da177e4SLinus Torvalds * We are working here with either a clone of the original 12341da177e4SLinus Torvalds * SKB, or a fresh unique copy made by the retransmit engine. 12351da177e4SLinus Torvalds */ 12362987babbSYuchung Cheng static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, 12372987babbSYuchung Cheng int clone_it, gfp_t gfp_mask, u32 rcv_nxt) 12381da177e4SLinus Torvalds { 12396687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 1240dfb4b9dcSDavid S. Miller struct inet_sock *inet; 1241dfb4b9dcSDavid S. Miller struct tcp_sock *tp; 1242dfb4b9dcSDavid S. Miller struct tcp_skb_cb *tcb; 124333ad798cSAdam Langley struct tcp_out_options opts; 124495c96174SEric Dumazet unsigned int tcp_options_size, tcp_header_size; 12458c72c65bSEric Dumazet struct sk_buff *oskb = NULL; 1246cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 12471da177e4SLinus Torvalds struct tcphdr *th; 1248a7a25630SEric Dumazet u64 prior_wstamp; 12491da177e4SLinus Torvalds int err; 12501da177e4SLinus Torvalds 1251dfb4b9dcSDavid S. Miller BUG_ON(!skb || !tcp_skb_pcount(skb)); 12526f094b9eSLawrence Brakmo tp = tcp_sk(sk); 12537f12422cSYuchung Cheng prior_wstamp = tp->tcp_wstamp_ns; 12547f12422cSYuchung Cheng tp->tcp_wstamp_ns = max(tp->tcp_wstamp_ns, tp->tcp_clock_cache); 1255a1ac9c8aSMartin KaFai Lau skb_set_delivery_time(skb, tp->tcp_wstamp_ns, true); 1256ccdbb6e9SEric Dumazet if (clone_it) { 12578c72c65bSEric Dumazet oskb = skb; 1258e2080072SEric Dumazet 1259e2080072SEric Dumazet tcp_skb_tsorted_save(oskb) { 1260e2080072SEric Dumazet if (unlikely(skb_cloned(oskb))) 1261e2080072SEric Dumazet skb = pskb_copy(oskb, gfp_mask); 1262dfb4b9dcSDavid S. Miller else 1263e2080072SEric Dumazet skb = skb_clone(oskb, gfp_mask); 1264e2080072SEric Dumazet } tcp_skb_tsorted_restore(oskb); 1265e2080072SEric Dumazet 1266dfb4b9dcSDavid S. Miller if (unlikely(!skb)) 1267dfb4b9dcSDavid S. Miller return -ENOBUFS; 1268b738a185SEric Dumazet /* retransmit skbs might have a non zero value in skb->dev 1269b738a185SEric Dumazet * because skb->dev is aliased with skb->rbnode.rb_left 1270b738a185SEric Dumazet */ 1271b738a185SEric Dumazet skb->dev = NULL; 1272dfb4b9dcSDavid S. Miller } 12735f6188a8SEric Dumazet 1274dfb4b9dcSDavid S. Miller inet = inet_sk(sk); 1275dfb4b9dcSDavid S. Miller tcb = TCP_SKB_CB(skb); 127633ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 12771da177e4SLinus Torvalds 1278051ba674SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) { 127933ad798cSAdam Langley tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); 1280051ba674SEric Dumazet } else { 128133ad798cSAdam Langley tcp_options_size = tcp_established_options(sk, skb, &opts, 128233ad798cSAdam Langley &md5); 1283051ba674SEric Dumazet /* Force a PSH flag on all (GSO) packets to expedite GRO flush 1284051ba674SEric Dumazet * at receiver : This slightly improve GRO performance. 1285051ba674SEric Dumazet * Note that we do not force the PSH flag for non GSO packets, 1286051ba674SEric Dumazet * because they might be sent under high congestion events, 1287051ba674SEric Dumazet * and in this case it is better to delay the delivery of 1-MSS 1288051ba674SEric Dumazet * packets and thus the corresponding ACK packet that would 1289051ba674SEric Dumazet * release the following packet. 1290051ba674SEric Dumazet */ 1291051ba674SEric Dumazet if (tcp_skb_pcount(skb) > 1) 1292051ba674SEric Dumazet tcb->tcp_flags |= TCPHDR_PSH; 1293051ba674SEric Dumazet } 129433ad798cSAdam Langley tcp_header_size = tcp_options_size + sizeof(struct tcphdr); 12951da177e4SLinus Torvalds 1296547669d4SEric Dumazet /* if no packet is in qdisc/device queue, then allow XPS to select 1297b2532eb9SEric Dumazet * another queue. We can be called from tcp_tsq_handler() 129873a6bab5SEric Dumazet * which holds one reference to sk. 1299b2532eb9SEric Dumazet * 1300b2532eb9SEric Dumazet * TODO: Ideally, in-flight pure ACK packets should not matter here. 1301b2532eb9SEric Dumazet * One way to get this would be to set skb->truesize = 2 on them. 1302547669d4SEric Dumazet */ 1303b2532eb9SEric Dumazet skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1); 13041da177e4SLinus Torvalds 130538ab52e8SEric Dumazet /* If we had to use memory reserve to allocate this skb, 130638ab52e8SEric Dumazet * this might cause drops if packet is looped back : 130738ab52e8SEric Dumazet * Other socket might not have SOCK_MEMALLOC. 130838ab52e8SEric Dumazet * Packets not looped back do not care about pfmemalloc. 130938ab52e8SEric Dumazet */ 131038ab52e8SEric Dumazet skb->pfmemalloc = 0; 131138ab52e8SEric Dumazet 1312aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 1313aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 131446d3ceabSEric Dumazet 131546d3ceabSEric Dumazet skb_orphan(skb); 131646d3ceabSEric Dumazet skb->sk = sk; 13171d2077acSEric Dumazet skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree; 131814afee4bSReshetova, Elena refcount_add(skb->truesize, &sk->sk_wmem_alloc); 13191da177e4SLinus Torvalds 1320c3a2e837SJulian Anastasov skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm); 1321c3a2e837SJulian Anastasov 13221da177e4SLinus Torvalds /* Build TCP header and checksum it. */ 1323ea1627c2SEric Dumazet th = (struct tcphdr *)skb->data; 1324c720c7e8SEric Dumazet th->source = inet->inet_sport; 1325c720c7e8SEric Dumazet th->dest = inet->inet_dport; 13261da177e4SLinus Torvalds th->seq = htonl(tcb->seq); 13272987babbSYuchung Cheng th->ack_seq = htonl(rcv_nxt); 1328df7a3b07SAl Viro *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | 13294de075e0SEric Dumazet tcb->tcp_flags); 1330dfb4b9dcSDavid S. Miller 13311da177e4SLinus Torvalds th->check = 0; 13321da177e4SLinus Torvalds th->urg_ptr = 0; 13331da177e4SLinus Torvalds 133433f5f57eSIlpo Järvinen /* The urg_mode check is necessary during a below snd_una win probe */ 13357691367dSHerbert Xu if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { 13367691367dSHerbert Xu if (before(tp->snd_up, tcb->seq + 0x10000)) { 13371da177e4SLinus Torvalds th->urg_ptr = htons(tp->snd_up - tcb->seq); 13381da177e4SLinus Torvalds th->urg = 1; 13397691367dSHerbert Xu } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { 13400eae88f3SEric Dumazet th->urg_ptr = htons(0xFFFF); 13417691367dSHerbert Xu th->urg = 1; 13427691367dSHerbert Xu } 13431da177e4SLinus Torvalds } 13441da177e4SLinus Torvalds 134551466a75SEric Dumazet skb_shinfo(skb)->gso_type = sk->sk_gso_type; 1346ea1627c2SEric Dumazet if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) { 1347ea1627c2SEric Dumazet th->window = htons(tcp_select_window(sk)); 1348ea1627c2SEric Dumazet tcp_ecn_send(sk, skb, th, tcp_header_size); 1349ea1627c2SEric Dumazet } else { 1350ea1627c2SEric Dumazet /* RFC1323: The window in SYN & SYN/ACK segments 1351ea1627c2SEric Dumazet * is never scaled. 1352ea1627c2SEric Dumazet */ 1353ea1627c2SEric Dumazet th->window = htons(min(tp->rcv_wnd, 65535U)); 1354ea1627c2SEric Dumazet } 1355fa3fe2b1SFlorian Westphal 1356ea66758cSPaolo Abeni tcp_options_write(th, tp, &opts); 1357fa3fe2b1SFlorian Westphal 1358cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 1359cfb6eeb4SYOSHIFUJI Hideaki /* Calculate the MD5 hash, as we have all we need now */ 1360cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 1361aba54656SEric Dumazet sk_gso_disable(sk); 1362bd0388aeSWilliam Allen Simpson tp->af_specific->calc_md5_hash(opts.hash_location, 136339f8e58eSEric Dumazet md5, sk, skb); 1364cfb6eeb4SYOSHIFUJI Hideaki } 1365cfb6eeb4SYOSHIFUJI Hideaki #endif 1366cfb6eeb4SYOSHIFUJI Hideaki 1367331fca43SMartin KaFai Lau /* BPF prog is the last one writing header option */ 1368331fca43SMartin KaFai Lau bpf_skops_write_hdr_opt(sk, skb, NULL, NULL, 0, &opts); 1369331fca43SMartin KaFai Lau 1370dd2e0b86SEric Dumazet INDIRECT_CALL_INET(icsk->icsk_af_ops->send_check, 1371dd2e0b86SEric Dumazet tcp_v6_send_check, tcp_v4_send_check, 1372dd2e0b86SEric Dumazet sk, skb); 13731da177e4SLinus Torvalds 13744de075e0SEric Dumazet if (likely(tcb->tcp_flags & TCPHDR_ACK)) 137527cde44aSYuchung Cheng tcp_event_ack_sent(sk, tcp_skb_pcount(skb), rcv_nxt); 13761da177e4SLinus Torvalds 1377a44d6eacSMartin KaFai Lau if (skb->len != tcp_header_size) { 1378cf533ea5SEric Dumazet tcp_event_data_sent(tp, sk); 1379a44d6eacSMartin KaFai Lau tp->data_segs_out += tcp_skb_pcount(skb); 1380ba113c3aSWei Wang tp->bytes_sent += skb->len - tcp_header_size; 1381a44d6eacSMartin KaFai Lau } 13821da177e4SLinus Torvalds 1383bd37a088SWei Yongjun if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) 1384aa2ea058STom Herbert TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, 1385aa2ea058STom Herbert tcp_skb_pcount(skb)); 13861da177e4SLinus Torvalds 13872efd055cSMarcelo Ricardo Leitner tp->segs_out += tcp_skb_pcount(skb); 13880ae5b43dSYuchung Cheng skb_set_hash_from_sk(skb, sk); 1389f69ad292SEric Dumazet /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */ 1390cd7d8498SEric Dumazet skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb); 1391f69ad292SEric Dumazet skb_shinfo(skb)->gso_size = tcp_skb_mss(skb); 1392cd7d8498SEric Dumazet 1393d3edd06eSEric Dumazet /* Leave earliest departure time in skb->tstamp (skb->skb_mstamp_ns) */ 1394971f10ecSEric Dumazet 1395971f10ecSEric Dumazet /* Cleanup our debris for IP stacks */ 1396971f10ecSEric Dumazet memset(skb->cb, 0, max(sizeof(struct inet_skb_parm), 1397971f10ecSEric Dumazet sizeof(struct inet6_skb_parm))); 1398971f10ecSEric Dumazet 1399a842fe14SEric Dumazet tcp_add_tx_delay(skb, tp); 1400a842fe14SEric Dumazet 140105e22e83SEric Dumazet err = INDIRECT_CALL_INET(icsk->icsk_af_ops->queue_xmit, 140205e22e83SEric Dumazet inet6_csk_xmit, ip_queue_xmit, 140305e22e83SEric Dumazet sk, skb, &inet->cork.fl); 14047faee5c0SEric Dumazet 14058c72c65bSEric Dumazet if (unlikely(err > 0)) { 14065ee2c941SChristoph Paasch tcp_enter_cwr(sk); 14078c72c65bSEric Dumazet err = net_xmit_eval(err); 14088c72c65bSEric Dumazet } 1409fc225799SEric Dumazet if (!err && oskb) { 1410a7a25630SEric Dumazet tcp_update_skb_after_send(sk, oskb, prior_wstamp); 1411fc225799SEric Dumazet tcp_rate_skb_sent(sk, oskb); 1412fc225799SEric Dumazet } 14138c72c65bSEric Dumazet return err; 14141da177e4SLinus Torvalds } 14151da177e4SLinus Torvalds 14162987babbSYuchung Cheng static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, 14172987babbSYuchung Cheng gfp_t gfp_mask) 14182987babbSYuchung Cheng { 14192987babbSYuchung Cheng return __tcp_transmit_skb(sk, skb, clone_it, gfp_mask, 14202987babbSYuchung Cheng tcp_sk(sk)->rcv_nxt); 14212987babbSYuchung Cheng } 14222987babbSYuchung Cheng 142367edfef7SAndi Kleen /* This routine just queues the buffer for sending. 14241da177e4SLinus Torvalds * 14251da177e4SLinus Torvalds * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, 14261da177e4SLinus Torvalds * otherwise socket can stall. 14271da177e4SLinus Torvalds */ 14281da177e4SLinus Torvalds static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) 14291da177e4SLinus Torvalds { 14301da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 14311da177e4SLinus Torvalds 14321da177e4SLinus Torvalds /* Advance write_seq and place onto the write_queue. */ 14330f317464SEric Dumazet WRITE_ONCE(tp->write_seq, TCP_SKB_CB(skb)->end_seq); 1434f4a775d1SEric Dumazet __skb_header_release(skb); 1435fe067e8aSDavid S. Miller tcp_add_write_queue_tail(sk, skb); 1436ab4e846aSEric Dumazet sk_wmem_queued_add(sk, skb->truesize); 14373ab224beSHideo Aoki sk_mem_charge(sk, skb->truesize); 14381da177e4SLinus Torvalds } 14391da177e4SLinus Torvalds 144067edfef7SAndi Kleen /* Initialize TSO segments for a packet. */ 14415bbb432cSEric Dumazet static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now) 1442f6302d1dSDavid S. Miller { 14434a64fd6cSEric Dumazet if (skb->len <= mss_now) { 1444f6302d1dSDavid S. Miller /* Avoid the costly divide in the normal 1445f6302d1dSDavid S. Miller * non-TSO case. 1446f6302d1dSDavid S. Miller */ 1447cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, 1); 1448f69ad292SEric Dumazet TCP_SKB_CB(skb)->tcp_gso_size = 0; 1449f6302d1dSDavid S. Miller } else { 1450cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now)); 1451f69ad292SEric Dumazet TCP_SKB_CB(skb)->tcp_gso_size = mss_now; 14521da177e4SLinus Torvalds } 14531da177e4SLinus Torvalds } 14541da177e4SLinus Torvalds 1455797108d1SIlpo Järvinen /* Pcount in the middle of the write queue got changed, we need to do various 1456797108d1SIlpo Järvinen * tweaks to fix counters 1457797108d1SIlpo Järvinen */ 1458cf533ea5SEric Dumazet static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) 1459797108d1SIlpo Järvinen { 1460797108d1SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1461797108d1SIlpo Järvinen 1462797108d1SIlpo Järvinen tp->packets_out -= decr; 1463797108d1SIlpo Järvinen 1464797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 1465797108d1SIlpo Järvinen tp->sacked_out -= decr; 1466797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 1467797108d1SIlpo Järvinen tp->retrans_out -= decr; 1468797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 1469797108d1SIlpo Järvinen tp->lost_out -= decr; 1470797108d1SIlpo Järvinen 1471797108d1SIlpo Järvinen /* Reno case is special. Sigh... */ 1472797108d1SIlpo Järvinen if (tcp_is_reno(tp) && decr > 0) 1473797108d1SIlpo Järvinen tp->sacked_out -= min_t(u32, tp->sacked_out, decr); 1474797108d1SIlpo Järvinen 1475797108d1SIlpo Järvinen if (tp->lost_skb_hint && 1476797108d1SIlpo Järvinen before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && 1477713bafeaSYuchung Cheng (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) 1478797108d1SIlpo Järvinen tp->lost_cnt_hint -= decr; 1479797108d1SIlpo Järvinen 1480797108d1SIlpo Järvinen tcp_verify_left_out(tp); 1481797108d1SIlpo Järvinen } 1482797108d1SIlpo Järvinen 14830a2cf20cSSoheil Hassas Yeganeh static bool tcp_has_tx_tstamp(const struct sk_buff *skb) 14840a2cf20cSSoheil Hassas Yeganeh { 14850a2cf20cSSoheil Hassas Yeganeh return TCP_SKB_CB(skb)->txstamp_ack || 14860a2cf20cSSoheil Hassas Yeganeh (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP); 14870a2cf20cSSoheil Hassas Yeganeh } 14880a2cf20cSSoheil Hassas Yeganeh 1489490cc7d0SWillem de Bruijn static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2) 1490490cc7d0SWillem de Bruijn { 1491490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo = skb_shinfo(skb); 1492490cc7d0SWillem de Bruijn 14930a2cf20cSSoheil Hassas Yeganeh if (unlikely(tcp_has_tx_tstamp(skb)) && 1494490cc7d0SWillem de Bruijn !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) { 1495490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo2 = skb_shinfo(skb2); 1496490cc7d0SWillem de Bruijn u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP; 1497490cc7d0SWillem de Bruijn 1498490cc7d0SWillem de Bruijn shinfo->tx_flags &= ~tsflags; 1499490cc7d0SWillem de Bruijn shinfo2->tx_flags |= tsflags; 1500490cc7d0SWillem de Bruijn swap(shinfo->tskey, shinfo2->tskey); 1501b51e13faSMartin KaFai Lau TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack; 1502b51e13faSMartin KaFai Lau TCP_SKB_CB(skb)->txstamp_ack = 0; 1503490cc7d0SWillem de Bruijn } 1504490cc7d0SWillem de Bruijn } 1505490cc7d0SWillem de Bruijn 1506a166140eSMartin KaFai Lau static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2) 1507a166140eSMartin KaFai Lau { 1508a166140eSMartin KaFai Lau TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor; 1509a166140eSMartin KaFai Lau TCP_SKB_CB(skb)->eor = 0; 1510a166140eSMartin KaFai Lau } 1511a166140eSMartin KaFai Lau 151275c119afSEric Dumazet /* Insert buff after skb on the write or rtx queue of sk. */ 151375c119afSEric Dumazet static void tcp_insert_write_queue_after(struct sk_buff *skb, 151475c119afSEric Dumazet struct sk_buff *buff, 151575c119afSEric Dumazet struct sock *sk, 151675c119afSEric Dumazet enum tcp_queue tcp_queue) 151775c119afSEric Dumazet { 151875c119afSEric Dumazet if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE) 151975c119afSEric Dumazet __skb_queue_after(&sk->sk_write_queue, skb, buff); 152075c119afSEric Dumazet else 152175c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, buff); 152275c119afSEric Dumazet } 152375c119afSEric Dumazet 15241da177e4SLinus Torvalds /* Function to create two new TCP segments. Shrinks the given segment 15251da177e4SLinus Torvalds * to the specified size and appends a new segment with the rest of the 15261da177e4SLinus Torvalds * packet to the list. This won't be called frequently, I hope. 15271da177e4SLinus Torvalds * Remember, these are still headerless SKBs at this point. 15281da177e4SLinus Torvalds */ 152975c119afSEric Dumazet int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue, 153075c119afSEric Dumazet struct sk_buff *skb, u32 len, 15316cc55e09SOctavian Purdila unsigned int mss_now, gfp_t gfp) 15321da177e4SLinus Torvalds { 15331da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 15341da177e4SLinus Torvalds struct sk_buff *buff; 1535b4a24397SEric Dumazet int old_factor; 1536b617158dSEric Dumazet long limit; 1537b60b49eaSHerbert Xu int nlen; 15389ce01461SIlpo Järvinen u8 flags; 15391da177e4SLinus Torvalds 15402fceec13SIlpo Järvinen if (WARN_ON(len > skb->len)) 15412fceec13SIlpo Järvinen return -EINVAL; 15426a438bbeSStephen Hemminger 1543b4a24397SEric Dumazet DEBUG_NET_WARN_ON_ONCE(skb_headlen(skb)); 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. */ 15635882efffSEric Dumazet buff = tcp_stream_alloc_skb(sk, 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); 1571b4a24397SEric Dumazet nlen = skb->len - len; 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 1629b4a24397SEric Dumazet DEBUG_NET_WARN_ON_ONCE(skb_headlen(skb)); 16301da177e4SLinus Torvalds eat = len; 16311da177e4SLinus Torvalds k = 0; 16327b7fc97aSEric Dumazet shinfo = skb_shinfo(skb); 16337b7fc97aSEric Dumazet for (i = 0; i < shinfo->nr_frags; i++) { 16347b7fc97aSEric Dumazet int size = skb_frag_size(&shinfo->frags[i]); 16359e903e08SEric Dumazet 16369e903e08SEric Dumazet if (size <= eat) { 1637aff65da0SIan Campbell skb_frag_unref(skb, i); 16389e903e08SEric Dumazet eat -= size; 16391da177e4SLinus Torvalds } else { 16407b7fc97aSEric Dumazet shinfo->frags[k] = shinfo->frags[i]; 16411da177e4SLinus Torvalds if (eat) { 1642b54c9d5bSJonathan Lemon skb_frag_off_add(&shinfo->frags[k], eat); 16437b7fc97aSEric Dumazet skb_frag_size_sub(&shinfo->frags[k], eat); 16441da177e4SLinus Torvalds eat = 0; 16451da177e4SLinus Torvalds } 16461da177e4SLinus Torvalds k++; 16471da177e4SLinus Torvalds } 16481da177e4SLinus Torvalds } 16497b7fc97aSEric Dumazet shinfo->nr_frags = k; 16501da177e4SLinus Torvalds 16511da177e4SLinus Torvalds skb->data_len -= len; 16521da177e4SLinus Torvalds skb->len = skb->data_len; 16537162fb24SEric Dumazet return len; 16541da177e4SLinus Torvalds } 16551da177e4SLinus Torvalds 165667edfef7SAndi Kleen /* Remove acked data from a packet in the transmit queue. */ 16571da177e4SLinus Torvalds int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 16581da177e4SLinus Torvalds { 16597162fb24SEric Dumazet u32 delta_truesize; 16607162fb24SEric Dumazet 1661c4777efaSEric Dumazet if (skb_unclone_keeptruesize(skb, GFP_ATOMIC)) 16621da177e4SLinus Torvalds return -ENOMEM; 16631da177e4SLinus Torvalds 16647162fb24SEric Dumazet delta_truesize = __pskb_trim_head(skb, len); 16651da177e4SLinus Torvalds 16661da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq += len; 16671da177e4SLinus Torvalds 16687162fb24SEric Dumazet skb->truesize -= delta_truesize; 1669ab4e846aSEric Dumazet sk_wmem_queued_add(sk, -delta_truesize); 16709b65b17dSTalal Ahmad if (!skb_zcopy_pure(skb)) 16717162fb24SEric Dumazet sk_mem_uncharge(sk, delta_truesize); 16721da177e4SLinus Torvalds 16735b35e1e6SNeal Cardwell /* Any change of skb->len requires recalculation of tso factor. */ 16741da177e4SLinus Torvalds if (tcp_skb_pcount(skb) > 1) 16755bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb)); 16761da177e4SLinus Torvalds 16771da177e4SLinus Torvalds return 0; 16781da177e4SLinus Torvalds } 16791da177e4SLinus Torvalds 16801b63edd6SYuchung Cheng /* Calculate MSS not accounting any TCP options. */ 16811b63edd6SYuchung Cheng static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu) 16825d424d5aSJohn Heffner { 1683cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1684cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 16855d424d5aSJohn Heffner int mss_now; 16865d424d5aSJohn Heffner 16875d424d5aSJohn Heffner /* Calculate base mss without TCP options: 16885d424d5aSJohn Heffner It is MMS_S - sizeof(tcphdr) of rfc1122 16895d424d5aSJohn Heffner */ 16905d424d5aSJohn Heffner mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 16915d424d5aSJohn Heffner 169267469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 169367469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 169467469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 169567469601SEric Dumazet 169667469601SEric Dumazet if (dst && dst_allfrag(dst)) 169767469601SEric Dumazet mss_now -= icsk->icsk_af_ops->net_frag_header_len; 169867469601SEric Dumazet } 169967469601SEric Dumazet 17005d424d5aSJohn Heffner /* Clamp it (mss_clamp does not include tcp options) */ 17015d424d5aSJohn Heffner if (mss_now > tp->rx_opt.mss_clamp) 17025d424d5aSJohn Heffner mss_now = tp->rx_opt.mss_clamp; 17035d424d5aSJohn Heffner 17045d424d5aSJohn Heffner /* Now subtract optional transport overhead */ 17055d424d5aSJohn Heffner mss_now -= icsk->icsk_ext_hdr_len; 17065d424d5aSJohn Heffner 17075d424d5aSJohn Heffner /* Then reserve room for full set of TCP options and 8 bytes of data */ 170878eb166cSKuniyuki Iwashima mss_now = max(mss_now, 170978eb166cSKuniyuki Iwashima READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_snd_mss)); 17105d424d5aSJohn Heffner return mss_now; 17115d424d5aSJohn Heffner } 17125d424d5aSJohn Heffner 17131b63edd6SYuchung Cheng /* Calculate MSS. Not accounting for SACKs here. */ 17141b63edd6SYuchung Cheng int tcp_mtu_to_mss(struct sock *sk, int pmtu) 17151b63edd6SYuchung Cheng { 17161b63edd6SYuchung Cheng /* Subtract TCP options size, not including SACKs */ 17171b63edd6SYuchung Cheng return __tcp_mtu_to_mss(sk, pmtu) - 17181b63edd6SYuchung Cheng (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr)); 17191b63edd6SYuchung Cheng } 1720c7bb4b89SEric Dumazet EXPORT_SYMBOL(tcp_mtu_to_mss); 17211b63edd6SYuchung Cheng 17225d424d5aSJohn Heffner /* Inverse of above */ 172367469601SEric Dumazet int tcp_mss_to_mtu(struct sock *sk, int mss) 17245d424d5aSJohn Heffner { 1725cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1726cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 17275d424d5aSJohn Heffner int mtu; 17285d424d5aSJohn Heffner 17295d424d5aSJohn Heffner mtu = mss + 17305d424d5aSJohn Heffner tp->tcp_header_len + 17315d424d5aSJohn Heffner icsk->icsk_ext_hdr_len + 17325d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 17335d424d5aSJohn Heffner 173467469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 173567469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 173667469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 173767469601SEric Dumazet 173867469601SEric Dumazet if (dst && dst_allfrag(dst)) 173967469601SEric Dumazet mtu += icsk->icsk_af_ops->net_frag_header_len; 174067469601SEric Dumazet } 17415d424d5aSJohn Heffner return mtu; 17425d424d5aSJohn Heffner } 1743556c6b46SNeal Cardwell EXPORT_SYMBOL(tcp_mss_to_mtu); 17445d424d5aSJohn Heffner 174567edfef7SAndi Kleen /* MTU probing init per socket */ 17465d424d5aSJohn Heffner void tcp_mtup_init(struct sock *sk) 17475d424d5aSJohn Heffner { 17485d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 17495d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 1750b0f9ca53SFan Du struct net *net = sock_net(sk); 17515d424d5aSJohn Heffner 1752f47d00e0SKuniyuki Iwashima icsk->icsk_mtup.enabled = READ_ONCE(net->ipv4.sysctl_tcp_mtu_probing) > 1; 17535d424d5aSJohn Heffner icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 17545d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 175588d78bc0SKuniyuki Iwashima icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, READ_ONCE(net->ipv4.sysctl_tcp_base_mss)); 17565d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 175705cbc0dbSFan Du if (icsk->icsk_mtup.enabled) 1758c74df29aSEric Dumazet icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; 17595d424d5aSJohn Heffner } 17604bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_mtup_init); 17615d424d5aSJohn Heffner 17621da177e4SLinus Torvalds /* This function synchronize snd mss to current pmtu/exthdr set. 17631da177e4SLinus Torvalds 17641da177e4SLinus Torvalds tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 17651da177e4SLinus Torvalds for TCP options, but includes only bare TCP header. 17661da177e4SLinus Torvalds 17671da177e4SLinus Torvalds tp->rx_opt.mss_clamp is mss negotiated at connection setup. 1768caa20d9aSStephen Hemminger It is minimum of user_mss and mss received with SYN. 17691da177e4SLinus Torvalds It also does not include TCP options. 17701da177e4SLinus Torvalds 1771d83d8461SArnaldo Carvalho de Melo inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. 17721da177e4SLinus Torvalds 17731da177e4SLinus Torvalds tp->mss_cache is current effective sending mss, including 17741da177e4SLinus Torvalds all tcp options except for SACKs. It is evaluated, 17751da177e4SLinus Torvalds taking into account current pmtu, but never exceeds 17761da177e4SLinus Torvalds tp->rx_opt.mss_clamp. 17771da177e4SLinus Torvalds 17781da177e4SLinus Torvalds NOTE1. rfc1122 clearly states that advertised MSS 17791da177e4SLinus Torvalds DOES NOT include either tcp or ip options. 17801da177e4SLinus Torvalds 1781d83d8461SArnaldo Carvalho de Melo NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache 1782d83d8461SArnaldo Carvalho de Melo are READ ONLY outside this function. --ANK (980731) 17831da177e4SLinus Torvalds */ 17841da177e4SLinus Torvalds unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 17851da177e4SLinus Torvalds { 17861da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1787d83d8461SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 17885d424d5aSJohn Heffner int mss_now; 17891da177e4SLinus Torvalds 17905d424d5aSJohn Heffner if (icsk->icsk_mtup.search_high > pmtu) 17915d424d5aSJohn Heffner icsk->icsk_mtup.search_high = pmtu; 17921da177e4SLinus Torvalds 17935d424d5aSJohn Heffner mss_now = tcp_mtu_to_mss(sk, pmtu); 1794409d22b4SIlpo Järvinen mss_now = tcp_bound_to_half_wnd(tp, mss_now); 17951da177e4SLinus Torvalds 17961da177e4SLinus Torvalds /* And store cached results */ 1797d83d8461SArnaldo Carvalho de Melo icsk->icsk_pmtu_cookie = pmtu; 17985d424d5aSJohn Heffner if (icsk->icsk_mtup.enabled) 17995d424d5aSJohn Heffner mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); 1800c1b4a7e6SDavid S. Miller tp->mss_cache = mss_now; 18011da177e4SLinus Torvalds 18021da177e4SLinus Torvalds return mss_now; 18031da177e4SLinus Torvalds } 18044bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_sync_mss); 18051da177e4SLinus Torvalds 18061da177e4SLinus Torvalds /* Compute the current effective MSS, taking SACKs and IP options, 18071da177e4SLinus Torvalds * and even PMTU discovery events into account. 18081da177e4SLinus Torvalds */ 18090c54b85fSIlpo Järvinen unsigned int tcp_current_mss(struct sock *sk) 18101da177e4SLinus Torvalds { 1811cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1812cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1813c1b4a7e6SDavid S. Miller u32 mss_now; 181495c96174SEric Dumazet unsigned int header_len; 181533ad798cSAdam Langley struct tcp_out_options opts; 181633ad798cSAdam Langley struct tcp_md5sig_key *md5; 18171da177e4SLinus Torvalds 1818c1b4a7e6SDavid S. Miller mss_now = tp->mss_cache; 1819c1b4a7e6SDavid S. Miller 18201da177e4SLinus Torvalds if (dst) { 18211da177e4SLinus Torvalds u32 mtu = dst_mtu(dst); 1822d83d8461SArnaldo Carvalho de Melo if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 18231da177e4SLinus Torvalds mss_now = tcp_sync_mss(sk, mtu); 18241da177e4SLinus Torvalds } 18251da177e4SLinus Torvalds 182633ad798cSAdam Langley header_len = tcp_established_options(sk, NULL, &opts, &md5) + 182733ad798cSAdam Langley sizeof(struct tcphdr); 182833ad798cSAdam Langley /* The mss_cache is sized based on tp->tcp_header_len, which assumes 182933ad798cSAdam Langley * some common options. If this is an odd packet (because we have SACK 183033ad798cSAdam Langley * blocks etc) then our calculated header_len will be different, and 183133ad798cSAdam Langley * we have to adjust mss_now correspondingly */ 183233ad798cSAdam Langley if (header_len != tp->tcp_header_len) { 183333ad798cSAdam Langley int delta = (int) header_len - tp->tcp_header_len; 183433ad798cSAdam Langley mss_now -= delta; 183533ad798cSAdam Langley } 1836cfb6eeb4SYOSHIFUJI Hideaki 18371da177e4SLinus Torvalds return mss_now; 18381da177e4SLinus Torvalds } 18391da177e4SLinus Torvalds 184086fd14adSWeiping Pan /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. 184186fd14adSWeiping Pan * As additional protections, we do not touch cwnd in retransmission phases, 184286fd14adSWeiping Pan * and if application hit its sndbuf limit recently. 184386fd14adSWeiping Pan */ 184486fd14adSWeiping Pan static void tcp_cwnd_application_limited(struct sock *sk) 1845a762a980SDavid S. Miller { 18469e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1847a762a980SDavid S. Miller 184886fd14adSWeiping Pan if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && 184986fd14adSWeiping Pan sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 185086fd14adSWeiping Pan /* Limited by application or receiver window. */ 185186fd14adSWeiping Pan u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); 185286fd14adSWeiping Pan u32 win_used = max(tp->snd_cwnd_used, init_win); 185340570375SEric Dumazet if (win_used < tcp_snd_cwnd(tp)) { 185486fd14adSWeiping Pan tp->snd_ssthresh = tcp_current_ssthresh(sk); 185540570375SEric Dumazet tcp_snd_cwnd_set(tp, (tcp_snd_cwnd(tp) + win_used) >> 1); 185686fd14adSWeiping Pan } 185786fd14adSWeiping Pan tp->snd_cwnd_used = 0; 185886fd14adSWeiping Pan } 1859c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 186086fd14adSWeiping Pan } 186186fd14adSWeiping Pan 1862ca8a2263SNeal Cardwell static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited) 1863a762a980SDavid S. Miller { 18641b1fc3fdSWei Wang const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 1865a762a980SDavid S. Miller struct tcp_sock *tp = tcp_sk(sk); 1866a762a980SDavid S. Miller 1867f4ce91ceSNeal Cardwell /* Track the strongest available signal of the degree to which the cwnd 1868f4ce91ceSNeal Cardwell * is fully utilized. If cwnd-limited then remember that fact for the 1869f4ce91ceSNeal Cardwell * current window. If not cwnd-limited then track the maximum number of 1870f4ce91ceSNeal Cardwell * outstanding packets in the current window. (If cwnd-limited then we 1871f4ce91ceSNeal Cardwell * chose to not update tp->max_packets_out to avoid an extra else 1872f4ce91ceSNeal Cardwell * clause with no functional impact.) 1873ca8a2263SNeal Cardwell */ 1874f4ce91ceSNeal Cardwell if (!before(tp->snd_una, tp->cwnd_usage_seq) || 1875f4ce91ceSNeal Cardwell is_cwnd_limited || 1876f4ce91ceSNeal Cardwell (!tp->is_cwnd_limited && 1877f4ce91ceSNeal Cardwell tp->packets_out > tp->max_packets_out)) { 1878ca8a2263SNeal Cardwell tp->is_cwnd_limited = is_cwnd_limited; 1879f4ce91ceSNeal Cardwell tp->max_packets_out = tp->packets_out; 1880f4ce91ceSNeal Cardwell tp->cwnd_usage_seq = tp->snd_nxt; 1881ca8a2263SNeal Cardwell } 1882e114a710SEric Dumazet 188324901551SEric Dumazet if (tcp_is_cwnd_limited(sk)) { 1884a762a980SDavid S. Miller /* Network is feed fully. */ 1885a762a980SDavid S. Miller tp->snd_cwnd_used = 0; 1886c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 1887a762a980SDavid S. Miller } else { 1888a762a980SDavid S. Miller /* Network starves. */ 1889a762a980SDavid S. Miller if (tp->packets_out > tp->snd_cwnd_used) 1890a762a980SDavid S. Miller tp->snd_cwnd_used = tp->packets_out; 1891a762a980SDavid S. Miller 18924845b571SKuniyuki Iwashima if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle) && 1893c2203cf7SEric Dumazet (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto && 18941b1fc3fdSWei Wang !ca_ops->cong_control) 1895a762a980SDavid S. Miller tcp_cwnd_application_limited(sk); 1896b0f71bd3SFrancis Yan 1897b0f71bd3SFrancis Yan /* The following conditions together indicate the starvation 1898b0f71bd3SFrancis Yan * is caused by insufficient sender buffer: 1899b0f71bd3SFrancis Yan * 1) just sent some data (see tcp_write_xmit) 1900b0f71bd3SFrancis Yan * 2) not cwnd limited (this else condition) 190175c119afSEric Dumazet * 3) no more data to send (tcp_write_queue_empty()) 1902b0f71bd3SFrancis Yan * 4) application is hitting buffer limit (SOCK_NOSPACE) 1903b0f71bd3SFrancis Yan */ 190475c119afSEric Dumazet if (tcp_write_queue_empty(sk) && sk->sk_socket && 1905b0f71bd3SFrancis Yan test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) && 1906b0f71bd3SFrancis Yan (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) 1907b0f71bd3SFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED); 1908a762a980SDavid S. Miller } 1909a762a980SDavid S. Miller } 1910a762a980SDavid S. Miller 1911d4589926SEric Dumazet /* Minshall's variant of the Nagle send check. */ 1912d4589926SEric Dumazet static bool tcp_minshall_check(const struct tcp_sock *tp) 1913d4589926SEric Dumazet { 1914d4589926SEric Dumazet return after(tp->snd_sml, tp->snd_una) && 1915d4589926SEric Dumazet !after(tp->snd_sml, tp->snd_nxt); 1916d4589926SEric Dumazet } 1917d4589926SEric Dumazet 1918d4589926SEric Dumazet /* Update snd_sml if this skb is under mss 1919d4589926SEric Dumazet * Note that a TSO packet might end with a sub-mss segment 1920d4589926SEric Dumazet * The test is really : 1921d4589926SEric Dumazet * if ((skb->len % mss) != 0) 1922d4589926SEric Dumazet * tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1923d4589926SEric Dumazet * But we can avoid doing the divide again given we already have 1924d4589926SEric Dumazet * skb_pcount = skb->len / mss_now 19250e3a4803SIlpo Järvinen */ 1926d4589926SEric Dumazet static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now, 1927d4589926SEric Dumazet const struct sk_buff *skb) 1928d4589926SEric Dumazet { 1929d4589926SEric Dumazet if (skb->len < tcp_skb_pcount(skb) * mss_now) 1930d4589926SEric Dumazet tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1931d4589926SEric Dumazet } 1932d4589926SEric Dumazet 1933d4589926SEric Dumazet /* Return false, if packet can be sent now without violation Nagle's rules: 1934d4589926SEric Dumazet * 1. It is full sized. (provided by caller in %partial bool) 1935d4589926SEric Dumazet * 2. Or it contains FIN. (already checked by caller) 1936d4589926SEric Dumazet * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. 1937d4589926SEric Dumazet * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 1938d4589926SEric Dumazet * With Minshall's modification: all sent small packets are ACKed. 1939d4589926SEric Dumazet */ 1940d4589926SEric Dumazet static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp, 1941cc93fc51SPeter Pan(潘卫平) int nonagle) 1942d4589926SEric Dumazet { 1943d4589926SEric Dumazet return partial && 1944d4589926SEric Dumazet ((nonagle & TCP_NAGLE_CORK) || 1945d4589926SEric Dumazet (!nonagle && tp->packets_out && tcp_minshall_check(tp))); 1946d4589926SEric Dumazet } 1947605ad7f1SEric Dumazet 1948605ad7f1SEric Dumazet /* Return how many segs we'd like on a TSO packet, 194965466904SEric Dumazet * depending on current pacing rate, and how close the peer is. 195065466904SEric Dumazet * 195165466904SEric Dumazet * Rationale is: 195265466904SEric Dumazet * - For close peers, we rather send bigger packets to reduce 195365466904SEric Dumazet * cpu costs, because occasional losses will be repaired fast. 195465466904SEric Dumazet * - For long distance/rtt flows, we would like to get ACK clocking 195565466904SEric Dumazet * with 1 ACK per ms. 195665466904SEric Dumazet * 195765466904SEric Dumazet * Use min_rtt to help adapt TSO burst size, with smaller min_rtt resulting 195865466904SEric Dumazet * in bigger TSO bursts. We we cut the RTT-based allowance in half 195965466904SEric Dumazet * for every 2^9 usec (aka 512 us) of RTT, so that the RTT-based allowance 196065466904SEric Dumazet * is below 1500 bytes after 6 * ~500 usec = 3ms. 1961605ad7f1SEric Dumazet */ 1962dcb8c9b4SEric Dumazet static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now, 19631b3878caSNeal Cardwell int min_tso_segs) 1964605ad7f1SEric Dumazet { 196565466904SEric Dumazet unsigned long bytes; 196665466904SEric Dumazet u32 r; 1967605ad7f1SEric Dumazet 196865466904SEric Dumazet bytes = sk->sk_pacing_rate >> READ_ONCE(sk->sk_pacing_shift); 1969605ad7f1SEric Dumazet 19702455e61bSKuniyuki Iwashima r = tcp_min_rtt(tcp_sk(sk)) >> READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_rtt_log); 197165466904SEric Dumazet if (r < BITS_PER_TYPE(sk->sk_gso_max_size)) 197265466904SEric Dumazet bytes += sk->sk_gso_max_size >> r; 1973605ad7f1SEric Dumazet 197465466904SEric Dumazet bytes = min_t(unsigned long, bytes, sk->sk_gso_max_size); 197565466904SEric Dumazet 197665466904SEric Dumazet return max_t(u32, bytes / mss_now, min_tso_segs); 1977605ad7f1SEric Dumazet } 1978605ad7f1SEric Dumazet 1979ed6e7268SNeal Cardwell /* Return the number of segments we want in the skb we are transmitting. 1980ed6e7268SNeal Cardwell * See if congestion control module wants to decide; otherwise, autosize. 1981ed6e7268SNeal Cardwell */ 1982ed6e7268SNeal Cardwell static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now) 1983ed6e7268SNeal Cardwell { 1984ed6e7268SNeal Cardwell const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 1985dcb8c9b4SEric Dumazet u32 min_tso, tso_segs; 1986ed6e7268SNeal Cardwell 1987dcb8c9b4SEric Dumazet min_tso = ca_ops->min_tso_segs ? 1988dcb8c9b4SEric Dumazet ca_ops->min_tso_segs(sk) : 1989e0bb4ab9SKuniyuki Iwashima READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs); 1990dcb8c9b4SEric Dumazet 1991dcb8c9b4SEric Dumazet tso_segs = tcp_tso_autosize(sk, mss_now, min_tso); 1992350c9f48SEric Dumazet return min_t(u32, tso_segs, sk->sk_gso_max_segs); 1993ed6e7268SNeal Cardwell } 1994ed6e7268SNeal Cardwell 1995d4589926SEric Dumazet /* Returns the portion of skb which can be sent right away */ 1996d4589926SEric Dumazet static unsigned int tcp_mss_split_point(const struct sock *sk, 1997d4589926SEric Dumazet const struct sk_buff *skb, 1998d4589926SEric Dumazet unsigned int mss_now, 1999d4589926SEric Dumazet unsigned int max_segs, 2000d4589926SEric Dumazet int nonagle) 2001c1b4a7e6SDavid S. Miller { 2002cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 2003d4589926SEric Dumazet u32 partial, needed, window, max_len; 2004c1b4a7e6SDavid S. Miller 200590840defSIlpo Järvinen window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 20061485348dSBen Hutchings max_len = mss_now * max_segs; 20070e3a4803SIlpo Järvinen 20081485348dSBen Hutchings if (likely(max_len <= window && skb != tcp_write_queue_tail(sk))) 20091485348dSBen Hutchings return max_len; 20100e3a4803SIlpo Järvinen 20115ea3a748SIlpo Järvinen needed = min(skb->len, window); 20125ea3a748SIlpo Järvinen 20131485348dSBen Hutchings if (max_len <= needed) 20141485348dSBen Hutchings return max_len; 20150e3a4803SIlpo Järvinen 2016d4589926SEric Dumazet partial = needed % mss_now; 2017d4589926SEric Dumazet /* If last segment is not a full MSS, check if Nagle rules allow us 2018d4589926SEric Dumazet * to include this last segment in this skb. 2019d4589926SEric Dumazet * Otherwise, we'll split the skb at last MSS boundary 2020d4589926SEric Dumazet */ 2021cc93fc51SPeter Pan(潘卫平) if (tcp_nagle_check(partial != 0, tp, nonagle)) 2022d4589926SEric Dumazet return needed - partial; 2023d4589926SEric Dumazet 2024d4589926SEric Dumazet return needed; 2025c1b4a7e6SDavid S. Miller } 2026c1b4a7e6SDavid S. Miller 2027c1b4a7e6SDavid S. Miller /* Can at least one segment of SKB be sent right now, according to the 2028c1b4a7e6SDavid S. Miller * congestion window rules? If so, return how many segments are allowed. 2029c1b4a7e6SDavid S. Miller */ 2030cf533ea5SEric Dumazet static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, 2031cf533ea5SEric Dumazet const struct sk_buff *skb) 2032c1b4a7e6SDavid S. Miller { 2033d649a7a8SEric Dumazet u32 in_flight, cwnd, halfcwnd; 2034c1b4a7e6SDavid S. Miller 2035c1b4a7e6SDavid S. Miller /* Don't be strict about the congestion window for the final FIN. */ 20364de075e0SEric Dumazet if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && 20374de075e0SEric Dumazet tcp_skb_pcount(skb) == 1) 2038c1b4a7e6SDavid S. Miller return 1; 2039c1b4a7e6SDavid S. Miller 2040c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 204140570375SEric Dumazet cwnd = tcp_snd_cwnd(tp); 2042d649a7a8SEric Dumazet if (in_flight >= cwnd) 2043c1b4a7e6SDavid S. Miller return 0; 2044d649a7a8SEric Dumazet 2045d649a7a8SEric Dumazet /* For better scheduling, ensure we have at least 2046d649a7a8SEric Dumazet * 2 GSO packets in flight. 2047d649a7a8SEric Dumazet */ 2048d649a7a8SEric Dumazet halfcwnd = max(cwnd >> 1, 1U); 2049d649a7a8SEric Dumazet return min(halfcwnd, cwnd - in_flight); 2050c1b4a7e6SDavid S. Miller } 2051c1b4a7e6SDavid S. Miller 2052b595076aSUwe Kleine-König /* Initialize TSO state of a skb. 205367edfef7SAndi Kleen * This must be invoked the first time we consider transmitting 2054c1b4a7e6SDavid S. Miller * SKB onto the wire. 2055c1b4a7e6SDavid S. Miller */ 20565bbb432cSEric Dumazet static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now) 2057c1b4a7e6SDavid S. Miller { 2058c1b4a7e6SDavid S. Miller int tso_segs = tcp_skb_pcount(skb); 2059c1b4a7e6SDavid S. Miller 2060f8269a49SIlpo Järvinen if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { 20615bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 2062c1b4a7e6SDavid S. Miller tso_segs = tcp_skb_pcount(skb); 2063c1b4a7e6SDavid S. Miller } 2064c1b4a7e6SDavid S. Miller return tso_segs; 2065c1b4a7e6SDavid S. Miller } 2066c1b4a7e6SDavid S. Miller 2067c1b4a7e6SDavid S. Miller 2068a2a385d6SEric Dumazet /* Return true if the Nagle test allows this packet to be 2069c1b4a7e6SDavid S. Miller * sent now. 2070c1b4a7e6SDavid S. Miller */ 2071a2a385d6SEric Dumazet static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, 2072c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 2073c1b4a7e6SDavid S. Miller { 2074c1b4a7e6SDavid S. Miller /* Nagle rule does not apply to frames, which sit in the middle of the 2075c1b4a7e6SDavid S. Miller * write_queue (they have no chances to get new data). 2076c1b4a7e6SDavid S. Miller * 2077c1b4a7e6SDavid S. Miller * This is implemented in the callers, where they modify the 'nonagle' 2078c1b4a7e6SDavid S. Miller * argument based upon the location of SKB in the send queue. 2079c1b4a7e6SDavid S. Miller */ 2080c1b4a7e6SDavid S. Miller if (nonagle & TCP_NAGLE_PUSH) 2081a2a385d6SEric Dumazet return true; 2082c1b4a7e6SDavid S. Miller 20839b44190dSYuchung Cheng /* Don't use the nagle rule for urgent data (or for the final FIN). */ 20849b44190dSYuchung Cheng if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) 2085a2a385d6SEric Dumazet return true; 2086c1b4a7e6SDavid S. Miller 2087cc93fc51SPeter Pan(潘卫平) if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle)) 2088a2a385d6SEric Dumazet return true; 2089c1b4a7e6SDavid S. Miller 2090a2a385d6SEric Dumazet return false; 2091c1b4a7e6SDavid S. Miller } 2092c1b4a7e6SDavid S. Miller 2093c1b4a7e6SDavid S. Miller /* Does at least the first segment of SKB fit into the send window? */ 2094a2a385d6SEric Dumazet static bool tcp_snd_wnd_test(const struct tcp_sock *tp, 2095a2a385d6SEric Dumazet const struct sk_buff *skb, 2096056834d9SIlpo Järvinen unsigned int cur_mss) 2097c1b4a7e6SDavid S. Miller { 2098c1b4a7e6SDavid S. Miller u32 end_seq = TCP_SKB_CB(skb)->end_seq; 2099c1b4a7e6SDavid S. Miller 2100c1b4a7e6SDavid S. Miller if (skb->len > cur_mss) 2101c1b4a7e6SDavid S. Miller end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 2102c1b4a7e6SDavid S. Miller 210390840defSIlpo Järvinen return !after(end_seq, tcp_wnd_end(tp)); 2104c1b4a7e6SDavid S. Miller } 2105c1b4a7e6SDavid S. Miller 2106c1b4a7e6SDavid S. Miller /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 2107c1b4a7e6SDavid S. Miller * which is put after SKB on the list. It is very much like 2108c1b4a7e6SDavid S. Miller * tcp_fragment() except that it may make several kinds of assumptions 2109c1b4a7e6SDavid S. Miller * in order to speed up the splitting operation. In particular, we 2110c1b4a7e6SDavid S. Miller * know that all the data is in scatter-gather pages, and that the 2111c1b4a7e6SDavid S. Miller * packet has never been sent out before (and thus is not cloned). 2112c1b4a7e6SDavid S. Miller */ 211356483341SEric Dumazet static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, 2114c4ead4c5SEric Dumazet unsigned int mss_now, gfp_t gfp) 2115c1b4a7e6SDavid S. Miller { 2116c1b4a7e6SDavid S. Miller int nlen = skb->len - len; 211756483341SEric Dumazet struct sk_buff *buff; 21189ce01461SIlpo Järvinen u8 flags; 2119c1b4a7e6SDavid S. Miller 2120c1b4a7e6SDavid S. Miller /* All of a TSO frame must be composed of paged data. */ 2121b4a24397SEric Dumazet DEBUG_NET_WARN_ON_ONCE(skb->len != skb->data_len); 2122c1b4a7e6SDavid S. Miller 21235882efffSEric Dumazet buff = tcp_stream_alloc_skb(sk, gfp, true); 212451456b29SIan Morris if (unlikely(!buff)) 2125c1b4a7e6SDavid S. Miller return -ENOMEM; 212641477662SJakub Kicinski skb_copy_decrypted(buff, skb); 21275a369ca6SPaolo Abeni mptcp_skb_ext_copy(buff, skb); 2128c1b4a7e6SDavid S. Miller 2129ab4e846aSEric Dumazet sk_wmem_queued_add(sk, buff->truesize); 21303ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 2131b60b49eaSHerbert Xu buff->truesize += nlen; 2132c1b4a7e6SDavid S. Miller skb->truesize -= nlen; 2133c1b4a7e6SDavid S. Miller 2134c1b4a7e6SDavid S. Miller /* Correct the sequence numbers. */ 2135c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 2136c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 2137c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 2138c1b4a7e6SDavid S. Miller 2139c1b4a7e6SDavid S. Miller /* PSH and FIN should only be set in the second packet. */ 21404de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 21414de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 21424de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 2143c1b4a7e6SDavid S. Miller 2144a166140eSMartin KaFai Lau tcp_skb_fragment_eor(skb, buff); 2145a166140eSMartin KaFai Lau 2146c1b4a7e6SDavid S. Miller skb_split(skb, buff, len); 2147490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 2148c1b4a7e6SDavid S. Miller 2149c1b4a7e6SDavid S. Miller /* Fix up tso_factor for both original and new SKB. */ 21505bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 21515bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 2152c1b4a7e6SDavid S. Miller 2153c1b4a7e6SDavid S. Miller /* Link BUFF into the send queue. */ 2154f4a775d1SEric Dumazet __skb_header_release(buff); 215556483341SEric Dumazet tcp_insert_write_queue_after(skb, buff, sk, TCP_FRAG_IN_WRITE_QUEUE); 2156c1b4a7e6SDavid S. Miller 2157c1b4a7e6SDavid S. Miller return 0; 2158c1b4a7e6SDavid S. Miller } 2159c1b4a7e6SDavid S. Miller 2160c1b4a7e6SDavid S. Miller /* Try to defer sending, if possible, in order to minimize the amount 2161c1b4a7e6SDavid S. Miller * of TSO splitting we do. View it as a kind of TSO Nagle test. 2162c1b4a7e6SDavid S. Miller * 2163c1b4a7e6SDavid S. Miller * This algorithm is from John Heffner. 2164c1b4a7e6SDavid S. Miller */ 2165ca8a2263SNeal Cardwell static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb, 2166f9bfe4e6SEric Dumazet bool *is_cwnd_limited, 2167f9bfe4e6SEric Dumazet bool *is_rwnd_limited, 2168f9bfe4e6SEric Dumazet u32 max_segs) 2169c1b4a7e6SDavid S. Miller { 21706687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 2171f1c6ea38SEric Dumazet u32 send_win, cong_win, limit, in_flight; 217250c8339eSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 217350c8339eSEric Dumazet struct sk_buff *head; 2174ad9f4f50SEric Dumazet int win_divisor; 2175f1c6ea38SEric Dumazet s64 delta; 2176c1b4a7e6SDavid S. Miller 217799d7662aSEric Dumazet if (icsk->icsk_ca_state >= TCP_CA_Recovery) 2178ae8064acSJohn Heffner goto send_now; 2179ae8064acSJohn Heffner 21805f852eb5SEric Dumazet /* Avoid bursty behavior by allowing defer 2181a682850aSEric Dumazet * only if the last write was recent (1 ms). 2182a682850aSEric Dumazet * Note that tp->tcp_wstamp_ns can be in the future if we have 2183a682850aSEric Dumazet * packets waiting in a qdisc or device for EDT delivery. 21845f852eb5SEric Dumazet */ 2185a682850aSEric Dumazet delta = tp->tcp_clock_cache - tp->tcp_wstamp_ns - NSEC_PER_MSEC; 2186a682850aSEric Dumazet if (delta > 0) 2187ae8064acSJohn Heffner goto send_now; 2188908a75c1SDavid S. Miller 2189c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 2190c1b4a7e6SDavid S. Miller 2191c8c9aeb5SStefano Brivio BUG_ON(tcp_skb_pcount(skb) <= 1); 219240570375SEric Dumazet BUG_ON(tcp_snd_cwnd(tp) <= in_flight); 2193c1b4a7e6SDavid S. Miller 219490840defSIlpo Järvinen send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 2195c1b4a7e6SDavid S. Miller 2196c1b4a7e6SDavid S. Miller /* From in_flight test above, we know that cwnd > in_flight. */ 219740570375SEric Dumazet cong_win = (tcp_snd_cwnd(tp) - in_flight) * tp->mss_cache; 2198c1b4a7e6SDavid S. Miller 2199c1b4a7e6SDavid S. Miller limit = min(send_win, cong_win); 2200c1b4a7e6SDavid S. Miller 2201ba244fe9SDavid S. Miller /* If a full-sized TSO skb can be sent, do it. */ 2202605ad7f1SEric Dumazet if (limit >= max_segs * tp->mss_cache) 2203ae8064acSJohn Heffner goto send_now; 2204ba244fe9SDavid S. Miller 220562ad2761SIlpo Järvinen /* Middle in queue won't get any more data, full sendable already? */ 220662ad2761SIlpo Järvinen if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) 220762ad2761SIlpo Järvinen goto send_now; 220862ad2761SIlpo Järvinen 22095bbcc0f5SLinus Torvalds win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor); 2210ad9f4f50SEric Dumazet if (win_divisor) { 221140570375SEric Dumazet u32 chunk = min(tp->snd_wnd, tcp_snd_cwnd(tp) * tp->mss_cache); 2212c1b4a7e6SDavid S. Miller 2213c1b4a7e6SDavid S. Miller /* If at least some fraction of a window is available, 2214c1b4a7e6SDavid S. Miller * just use it. 2215c1b4a7e6SDavid S. Miller */ 2216ad9f4f50SEric Dumazet chunk /= win_divisor; 2217c1b4a7e6SDavid S. Miller if (limit >= chunk) 2218ae8064acSJohn Heffner goto send_now; 2219c1b4a7e6SDavid S. Miller } else { 2220c1b4a7e6SDavid S. Miller /* Different approach, try not to defer past a single 2221c1b4a7e6SDavid S. Miller * ACK. Receiver should ACK every other full sized 2222c1b4a7e6SDavid S. Miller * frame, so if we have space for more than 3 frames 2223c1b4a7e6SDavid S. Miller * then send now. 2224c1b4a7e6SDavid S. Miller */ 22256b5a5c0dSNeal Cardwell if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) 2226ae8064acSJohn Heffner goto send_now; 2227c1b4a7e6SDavid S. Miller } 2228c1b4a7e6SDavid S. Miller 222975c119afSEric Dumazet /* TODO : use tsorted_sent_queue ? */ 223075c119afSEric Dumazet head = tcp_rtx_queue_head(sk); 223175c119afSEric Dumazet if (!head) 223275c119afSEric Dumazet goto send_now; 2233f1c6ea38SEric Dumazet delta = tp->tcp_clock_cache - head->tstamp; 223450c8339eSEric Dumazet /* If next ACK is likely to come too late (half srtt), do not defer */ 2235f1c6ea38SEric Dumazet if ((s64)(delta - (u64)NSEC_PER_USEC * (tp->srtt_us >> 4)) < 0) 223650c8339eSEric Dumazet goto send_now; 223750c8339eSEric Dumazet 2238f9bfe4e6SEric Dumazet /* Ok, it looks like it is advisable to defer. 2239f9bfe4e6SEric Dumazet * Three cases are tracked : 2240f9bfe4e6SEric Dumazet * 1) We are cwnd-limited 2241f9bfe4e6SEric Dumazet * 2) We are rwnd-limited 2242f9bfe4e6SEric Dumazet * 3) We are application limited. 2243f9bfe4e6SEric Dumazet */ 2244f9bfe4e6SEric Dumazet if (cong_win < send_win) { 2245f9bfe4e6SEric Dumazet if (cong_win <= skb->len) { 2246ca8a2263SNeal Cardwell *is_cwnd_limited = true; 2247f9bfe4e6SEric Dumazet return true; 2248f9bfe4e6SEric Dumazet } 2249f9bfe4e6SEric Dumazet } else { 2250f9bfe4e6SEric Dumazet if (send_win <= skb->len) { 2251f9bfe4e6SEric Dumazet *is_rwnd_limited = true; 2252f9bfe4e6SEric Dumazet return true; 2253f9bfe4e6SEric Dumazet } 2254f9bfe4e6SEric Dumazet } 2255f9bfe4e6SEric Dumazet 2256f9bfe4e6SEric Dumazet /* If this packet won't get more data, do not wait. */ 2257d8ed257fSEric Dumazet if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) || 2258d8ed257fSEric Dumazet TCP_SKB_CB(skb)->eor) 2259f9bfe4e6SEric Dumazet goto send_now; 2260ca8a2263SNeal Cardwell 2261a2a385d6SEric Dumazet return true; 2262ae8064acSJohn Heffner 2263ae8064acSJohn Heffner send_now: 2264a2a385d6SEric Dumazet return false; 2265c1b4a7e6SDavid S. Miller } 2266c1b4a7e6SDavid S. Miller 226705cbc0dbSFan Du static inline void tcp_mtu_check_reprobe(struct sock *sk) 226805cbc0dbSFan Du { 226905cbc0dbSFan Du struct inet_connection_sock *icsk = inet_csk(sk); 227005cbc0dbSFan Du struct tcp_sock *tp = tcp_sk(sk); 227105cbc0dbSFan Du struct net *net = sock_net(sk); 227205cbc0dbSFan Du u32 interval; 227305cbc0dbSFan Du s32 delta; 227405cbc0dbSFan Du 22752a85388fSKuniyuki Iwashima interval = READ_ONCE(net->ipv4.sysctl_tcp_probe_interval); 2276c74df29aSEric Dumazet delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp; 227705cbc0dbSFan Du if (unlikely(delta >= interval * HZ)) { 227805cbc0dbSFan Du int mss = tcp_current_mss(sk); 227905cbc0dbSFan Du 228005cbc0dbSFan Du /* Update current search range */ 228105cbc0dbSFan Du icsk->icsk_mtup.probe_size = 0; 228205cbc0dbSFan Du icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + 228305cbc0dbSFan Du sizeof(struct tcphdr) + 228405cbc0dbSFan Du icsk->icsk_af_ops->net_header_len; 228505cbc0dbSFan Du icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss); 228605cbc0dbSFan Du 228705cbc0dbSFan Du /* Update probe time stamp */ 2288c74df29aSEric Dumazet icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; 228905cbc0dbSFan Du } 229005cbc0dbSFan Du } 229105cbc0dbSFan Du 2292808cf9e3SIlya Lesokhin static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len) 2293808cf9e3SIlya Lesokhin { 2294808cf9e3SIlya Lesokhin struct sk_buff *skb, *next; 2295808cf9e3SIlya Lesokhin 2296808cf9e3SIlya Lesokhin skb = tcp_send_head(sk); 2297808cf9e3SIlya Lesokhin tcp_for_write_queue_from_safe(skb, next, sk) { 2298808cf9e3SIlya Lesokhin if (len <= skb->len) 2299808cf9e3SIlya Lesokhin break; 2300808cf9e3SIlya Lesokhin 23019b65b17dSTalal Ahmad if (unlikely(TCP_SKB_CB(skb)->eor) || 23029b65b17dSTalal Ahmad tcp_has_tx_tstamp(skb) || 23039b65b17dSTalal Ahmad !skb_pure_zcopy_same(skb, next)) 2304808cf9e3SIlya Lesokhin return false; 2305808cf9e3SIlya Lesokhin 2306808cf9e3SIlya Lesokhin len -= skb->len; 2307808cf9e3SIlya Lesokhin } 2308808cf9e3SIlya Lesokhin 2309808cf9e3SIlya Lesokhin return true; 2310808cf9e3SIlya Lesokhin } 2311808cf9e3SIlya Lesokhin 231273601329SEric Dumazet static int tcp_clone_payload(struct sock *sk, struct sk_buff *to, 231373601329SEric Dumazet int probe_size) 231473601329SEric Dumazet { 231573601329SEric Dumazet skb_frag_t *lastfrag = NULL, *fragto = skb_shinfo(to)->frags; 231673601329SEric Dumazet int i, todo, len = 0, nr_frags = 0; 231773601329SEric Dumazet const struct sk_buff *skb; 231873601329SEric Dumazet 231973601329SEric Dumazet if (!sk_wmem_schedule(sk, to->truesize + probe_size)) 232073601329SEric Dumazet return -ENOMEM; 232173601329SEric Dumazet 232273601329SEric Dumazet skb_queue_walk(&sk->sk_write_queue, skb) { 232373601329SEric Dumazet const skb_frag_t *fragfrom = skb_shinfo(skb)->frags; 232473601329SEric Dumazet 232573601329SEric Dumazet if (skb_headlen(skb)) 232673601329SEric Dumazet return -EINVAL; 232773601329SEric Dumazet 232873601329SEric Dumazet for (i = 0; i < skb_shinfo(skb)->nr_frags; i++, fragfrom++) { 232973601329SEric Dumazet if (len >= probe_size) 233073601329SEric Dumazet goto commit; 233173601329SEric Dumazet todo = min_t(int, skb_frag_size(fragfrom), 233273601329SEric Dumazet probe_size - len); 233373601329SEric Dumazet len += todo; 233473601329SEric Dumazet if (lastfrag && 233573601329SEric Dumazet skb_frag_page(fragfrom) == skb_frag_page(lastfrag) && 233673601329SEric Dumazet skb_frag_off(fragfrom) == skb_frag_off(lastfrag) + 233773601329SEric Dumazet skb_frag_size(lastfrag)) { 233873601329SEric Dumazet skb_frag_size_add(lastfrag, todo); 233973601329SEric Dumazet continue; 234073601329SEric Dumazet } 234173601329SEric Dumazet if (unlikely(nr_frags == MAX_SKB_FRAGS)) 234273601329SEric Dumazet return -E2BIG; 234373601329SEric Dumazet skb_frag_page_copy(fragto, fragfrom); 234473601329SEric Dumazet skb_frag_off_copy(fragto, fragfrom); 234573601329SEric Dumazet skb_frag_size_set(fragto, todo); 234673601329SEric Dumazet nr_frags++; 234773601329SEric Dumazet lastfrag = fragto++; 234873601329SEric Dumazet } 234973601329SEric Dumazet } 235073601329SEric Dumazet commit: 235173601329SEric Dumazet WARN_ON_ONCE(len != probe_size); 235273601329SEric Dumazet for (i = 0; i < nr_frags; i++) 235373601329SEric Dumazet skb_frag_ref(to, i); 235473601329SEric Dumazet 235573601329SEric Dumazet skb_shinfo(to)->nr_frags = nr_frags; 235673601329SEric Dumazet to->truesize += probe_size; 235773601329SEric Dumazet to->len += probe_size; 235873601329SEric Dumazet to->data_len += probe_size; 235973601329SEric Dumazet __skb_header_release(to); 236073601329SEric Dumazet return 0; 236173601329SEric Dumazet } 236273601329SEric Dumazet 23635d424d5aSJohn Heffner /* Create a new MTU probe if we are ready. 236467edfef7SAndi Kleen * MTU probe is regularly attempting to increase the path MTU by 236567edfef7SAndi Kleen * deliberately sending larger packets. This discovers routing 236667edfef7SAndi Kleen * changes resulting in larger path MTUs. 236767edfef7SAndi Kleen * 23685d424d5aSJohn Heffner * Returns 0 if we should wait to probe (no cwnd available), 23695d424d5aSJohn Heffner * 1 if a probe was sent, 2370056834d9SIlpo Järvinen * -1 otherwise 2371056834d9SIlpo Järvinen */ 23725d424d5aSJohn Heffner static int tcp_mtu_probe(struct sock *sk) 23735d424d5aSJohn Heffner { 23745d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 237512a59abcSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 23765d424d5aSJohn Heffner struct sk_buff *skb, *nskb, *next; 23776b58e0a5SFan Du struct net *net = sock_net(sk); 23785d424d5aSJohn Heffner int probe_size; 237991cc17c0SIlpo Järvinen int size_needed; 238012a59abcSEric Dumazet int copy, len; 23815d424d5aSJohn Heffner int mss_now; 23826b58e0a5SFan Du int interval; 23835d424d5aSJohn Heffner 23845d424d5aSJohn Heffner /* Not currently probing/verifying, 23855d424d5aSJohn Heffner * not in recovery, 23865d424d5aSJohn Heffner * have enough cwnd, and 238712a59abcSEric Dumazet * not SACKing (the variable headers throw things off) 238812a59abcSEric Dumazet */ 238912a59abcSEric Dumazet if (likely(!icsk->icsk_mtup.enabled || 23905d424d5aSJohn Heffner icsk->icsk_mtup.probe_size || 23915d424d5aSJohn Heffner inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 239240570375SEric Dumazet tcp_snd_cwnd(tp) < 11 || 239312a59abcSEric Dumazet tp->rx_opt.num_sacks || tp->rx_opt.dsack)) 23945d424d5aSJohn Heffner return -1; 23955d424d5aSJohn Heffner 23966b58e0a5SFan Du /* Use binary search for probe_size between tcp_mss_base, 23976b58e0a5SFan Du * and current mss_clamp. if (search_high - search_low) 23986b58e0a5SFan Du * smaller than a threshold, backoff from probing. 23996b58e0a5SFan Du */ 24000c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 24016b58e0a5SFan Du probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high + 24026b58e0a5SFan Du icsk->icsk_mtup.search_low) >> 1); 240391cc17c0SIlpo Järvinen size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; 24046b58e0a5SFan Du interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low; 240505cbc0dbSFan Du /* When misfortune happens, we are reprobing actively, 240605cbc0dbSFan Du * and then reprobe timer has expired. We stick with current 240705cbc0dbSFan Du * probing process by not resetting search range to its orignal. 240805cbc0dbSFan Du */ 24096b58e0a5SFan Du if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) || 241092c0aa41SKuniyuki Iwashima interval < READ_ONCE(net->ipv4.sysctl_tcp_probe_threshold)) { 241105cbc0dbSFan Du /* Check whether enough time has elaplased for 241205cbc0dbSFan Du * another round of probing. 241305cbc0dbSFan Du */ 241405cbc0dbSFan Du tcp_mtu_check_reprobe(sk); 24155d424d5aSJohn Heffner return -1; 24165d424d5aSJohn Heffner } 24175d424d5aSJohn Heffner 24185d424d5aSJohn Heffner /* Have enough data in the send queue to probe? */ 24197f9c33e5SIlpo Järvinen if (tp->write_seq - tp->snd_nxt < size_needed) 24205d424d5aSJohn Heffner return -1; 24215d424d5aSJohn Heffner 242291cc17c0SIlpo Järvinen if (tp->snd_wnd < size_needed) 24235d424d5aSJohn Heffner return -1; 242490840defSIlpo Järvinen if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) 24255d424d5aSJohn Heffner return 0; 24265d424d5aSJohn Heffner 2427d67c58e9SIlpo Järvinen /* Do we need to wait to drain cwnd? With none in flight, don't stall */ 242840570375SEric Dumazet if (tcp_packets_in_flight(tp) + 2 > tcp_snd_cwnd(tp)) { 2429d67c58e9SIlpo Järvinen if (!tcp_packets_in_flight(tp)) 24305d424d5aSJohn Heffner return -1; 24315d424d5aSJohn Heffner else 24325d424d5aSJohn Heffner return 0; 24335d424d5aSJohn Heffner } 24345d424d5aSJohn Heffner 2435808cf9e3SIlya Lesokhin if (!tcp_can_coalesce_send_queue_head(sk, probe_size)) 2436808cf9e3SIlya Lesokhin return -1; 2437808cf9e3SIlya Lesokhin 24385d424d5aSJohn Heffner /* We're allowed to probe. Build it now. */ 24395882efffSEric Dumazet nskb = tcp_stream_alloc_skb(sk, GFP_ATOMIC, false); 244051456b29SIan Morris if (!nskb) 24415d424d5aSJohn Heffner return -1; 244273601329SEric Dumazet 244373601329SEric Dumazet /* build the payload, and be prepared to abort if this fails. */ 244473601329SEric Dumazet if (tcp_clone_payload(sk, nskb, probe_size)) { 244573601329SEric Dumazet consume_skb(nskb); 244673601329SEric Dumazet return -1; 244773601329SEric Dumazet } 2448ab4e846aSEric Dumazet sk_wmem_queued_add(sk, nskb->truesize); 24493ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 24505d424d5aSJohn Heffner 2451fe067e8aSDavid S. Miller skb = tcp_send_head(sk); 245241477662SJakub Kicinski skb_copy_decrypted(nskb, skb); 24535a369ca6SPaolo Abeni mptcp_skb_ext_copy(nskb, skb); 24545d424d5aSJohn Heffner 24555d424d5aSJohn Heffner TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 24565d424d5aSJohn Heffner TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 24574de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; 24585d424d5aSJohn Heffner 245950c4817eSIlpo Järvinen tcp_insert_write_queue_before(nskb, skb, sk); 24602b7cda9cSEric Dumazet tcp_highest_sack_replace(sk, skb, nskb); 246150c4817eSIlpo Järvinen 24625d424d5aSJohn Heffner len = 0; 2463234b6860SIlpo Järvinen tcp_for_write_queue_from_safe(skb, next, sk) { 24645d424d5aSJohn Heffner copy = min_t(int, skb->len, probe_size - len); 24655d424d5aSJohn Heffner 24665d424d5aSJohn Heffner if (skb->len <= copy) { 24675d424d5aSJohn Heffner /* We've eaten all the data from this skb. 24685d424d5aSJohn Heffner * Throw it away. */ 24694de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 2470808cf9e3SIlya Lesokhin /* If this is the last SKB we copy and eor is set 2471808cf9e3SIlya Lesokhin * we need to propagate it to the new skb. 2472808cf9e3SIlya Lesokhin */ 2473808cf9e3SIlya Lesokhin TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor; 2474888a5c53SWillem de Bruijn tcp_skb_collapse_tstamp(nskb, skb); 2475fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 247603271f3aSTalal Ahmad tcp_wmem_free_skb(sk, skb); 24775d424d5aSJohn Heffner } else { 24784de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & 2479a3433f35SChangli Gao ~(TCPHDR_FIN|TCPHDR_PSH); 24805d424d5aSJohn Heffner __pskb_trim_head(skb, copy); 24815bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 24825d424d5aSJohn Heffner TCP_SKB_CB(skb)->seq += copy; 24835d424d5aSJohn Heffner } 24845d424d5aSJohn Heffner 24855d424d5aSJohn Heffner len += copy; 2486234b6860SIlpo Järvinen 2487234b6860SIlpo Järvinen if (len >= probe_size) 2488234b6860SIlpo Järvinen break; 24895d424d5aSJohn Heffner } 24905bbb432cSEric Dumazet tcp_init_tso_segs(nskb, nskb->len); 24915d424d5aSJohn Heffner 24925d424d5aSJohn Heffner /* We're ready to send. If this fails, the probe will 24937faee5c0SEric Dumazet * be resegmented into mss-sized pieces by tcp_write_xmit(). 24947faee5c0SEric Dumazet */ 24955d424d5aSJohn Heffner if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 24965d424d5aSJohn Heffner /* Decrement cwnd here because we are sending 24975d424d5aSJohn Heffner * effectively two packets. */ 249840570375SEric Dumazet tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) - 1); 249966f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, nskb); 25005d424d5aSJohn Heffner 25015d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 25020e7b1368SJohn Heffner tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 25030e7b1368SJohn Heffner tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 25045d424d5aSJohn Heffner 25055d424d5aSJohn Heffner return 1; 25065d424d5aSJohn Heffner } 25075d424d5aSJohn Heffner 25085d424d5aSJohn Heffner return -1; 25095d424d5aSJohn Heffner } 25105d424d5aSJohn Heffner 2511864e5c09SEric Dumazet static bool tcp_pacing_check(struct sock *sk) 2512218af599SEric Dumazet { 2513864e5c09SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 2514864e5c09SEric Dumazet 2515864e5c09SEric Dumazet if (!tcp_needs_internal_pacing(sk)) 2516864e5c09SEric Dumazet return false; 2517864e5c09SEric Dumazet 2518864e5c09SEric Dumazet if (tp->tcp_wstamp_ns <= tp->tcp_clock_cache) 2519864e5c09SEric Dumazet return false; 2520864e5c09SEric Dumazet 2521864e5c09SEric Dumazet if (!hrtimer_is_queued(&tp->pacing_timer)) { 2522864e5c09SEric Dumazet hrtimer_start(&tp->pacing_timer, 2523864e5c09SEric Dumazet ns_to_ktime(tp->tcp_wstamp_ns), 2524864e5c09SEric Dumazet HRTIMER_MODE_ABS_PINNED_SOFT); 2525864e5c09SEric Dumazet sock_hold(sk); 2526864e5c09SEric Dumazet } 2527864e5c09SEric Dumazet return true; 2528218af599SEric Dumazet } 2529218af599SEric Dumazet 2530f9616c35SEric Dumazet /* TCP Small Queues : 2531f9616c35SEric Dumazet * Control number of packets in qdisc/devices to two packets / or ~1 ms. 2532f9616c35SEric Dumazet * (These limits are doubled for retransmits) 2533f9616c35SEric Dumazet * This allows for : 2534f9616c35SEric Dumazet * - better RTT estimation and ACK scheduling 2535f9616c35SEric Dumazet * - faster recovery 2536f9616c35SEric Dumazet * - high rates 2537f9616c35SEric Dumazet * Alas, some drivers / subsystems require a fair amount 2538f9616c35SEric Dumazet * of queued bytes to ensure line rate. 2539f9616c35SEric Dumazet * One example is wifi aggregation (802.11 AMPDU) 2540f9616c35SEric Dumazet */ 2541f9616c35SEric Dumazet static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb, 2542f9616c35SEric Dumazet unsigned int factor) 2543f9616c35SEric Dumazet { 254476a9ebe8SEric Dumazet unsigned long limit; 2545f9616c35SEric Dumazet 254676a9ebe8SEric Dumazet limit = max_t(unsigned long, 254776a9ebe8SEric Dumazet 2 * skb->truesize, 25487c68fa2bSEric Dumazet sk->sk_pacing_rate >> READ_ONCE(sk->sk_pacing_shift)); 2549c73e5807SEric Dumazet if (sk->sk_pacing_status == SK_PACING_NONE) 255076a9ebe8SEric Dumazet limit = min_t(unsigned long, limit, 25519fb90193SKuniyuki Iwashima READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes)); 2552f9616c35SEric Dumazet limit <<= factor; 2553f9616c35SEric Dumazet 2554a842fe14SEric Dumazet if (static_branch_unlikely(&tcp_tx_delay_enabled) && 2555a842fe14SEric Dumazet tcp_sk(sk)->tcp_tx_delay) { 2556a842fe14SEric Dumazet u64 extra_bytes = (u64)sk->sk_pacing_rate * tcp_sk(sk)->tcp_tx_delay; 2557a842fe14SEric Dumazet 2558a842fe14SEric Dumazet /* TSQ is based on skb truesize sum (sk_wmem_alloc), so we 2559a842fe14SEric Dumazet * approximate our needs assuming an ~100% skb->truesize overhead. 2560a842fe14SEric Dumazet * USEC_PER_SEC is approximated by 2^20. 2561a842fe14SEric Dumazet * do_div(extra_bytes, USEC_PER_SEC/2) is replaced by a right shift. 2562a842fe14SEric Dumazet */ 2563a842fe14SEric Dumazet extra_bytes >>= (20 - 1); 2564a842fe14SEric Dumazet limit += extra_bytes; 2565a842fe14SEric Dumazet } 256614afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) > limit) { 256775c119afSEric Dumazet /* Always send skb if rtx queue is empty. 256875eefc6cSEric Dumazet * No need to wait for TX completion to call us back, 256975eefc6cSEric Dumazet * after softirq/tasklet schedule. 257075eefc6cSEric Dumazet * This helps when TX completions are delayed too much. 257175eefc6cSEric Dumazet */ 257275c119afSEric Dumazet if (tcp_rtx_queue_empty(sk)) 257375eefc6cSEric Dumazet return false; 257475eefc6cSEric Dumazet 25757aa5470cSEric Dumazet set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags); 2576f9616c35SEric Dumazet /* It is possible TX completion already happened 2577f9616c35SEric Dumazet * before we set TSQ_THROTTLED, so we must 2578f9616c35SEric Dumazet * test again the condition. 2579f9616c35SEric Dumazet */ 2580f9616c35SEric Dumazet smp_mb__after_atomic(); 2581ce8299b6SEric Dumazet if (refcount_read(&sk->sk_wmem_alloc) > limit) 2582f9616c35SEric Dumazet return true; 2583f9616c35SEric Dumazet } 2584f9616c35SEric Dumazet return false; 2585f9616c35SEric Dumazet } 2586f9616c35SEric Dumazet 258705b055e8SFrancis Yan static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new) 258805b055e8SFrancis Yan { 2589628174ccSEric Dumazet const u32 now = tcp_jiffies32; 2590efe967cdSArnd Bergmann enum tcp_chrono old = tp->chrono_type; 259105b055e8SFrancis Yan 2592efe967cdSArnd Bergmann if (old > TCP_CHRONO_UNSPEC) 2593efe967cdSArnd Bergmann tp->chrono_stat[old - 1] += now - tp->chrono_start; 259405b055e8SFrancis Yan tp->chrono_start = now; 259505b055e8SFrancis Yan tp->chrono_type = new; 259605b055e8SFrancis Yan } 259705b055e8SFrancis Yan 259805b055e8SFrancis Yan void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type) 259905b055e8SFrancis Yan { 260005b055e8SFrancis Yan struct tcp_sock *tp = tcp_sk(sk); 260105b055e8SFrancis Yan 260205b055e8SFrancis Yan /* If there are multiple conditions worthy of tracking in a 26030f87230dSFrancis Yan * chronograph then the highest priority enum takes precedence 26040f87230dSFrancis Yan * over the other conditions. So that if something "more interesting" 260505b055e8SFrancis Yan * starts happening, stop the previous chrono and start a new one. 260605b055e8SFrancis Yan */ 260705b055e8SFrancis Yan if (type > tp->chrono_type) 260805b055e8SFrancis Yan tcp_chrono_set(tp, type); 260905b055e8SFrancis Yan } 261005b055e8SFrancis Yan 261105b055e8SFrancis Yan void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type) 261205b055e8SFrancis Yan { 261305b055e8SFrancis Yan struct tcp_sock *tp = tcp_sk(sk); 261405b055e8SFrancis Yan 26150f87230dSFrancis Yan 26160f87230dSFrancis Yan /* There are multiple conditions worthy of tracking in a 26170f87230dSFrancis Yan * chronograph, so that the highest priority enum takes 26180f87230dSFrancis Yan * precedence over the other conditions (see tcp_chrono_start). 26190f87230dSFrancis Yan * If a condition stops, we only stop chrono tracking if 26200f87230dSFrancis Yan * it's the "most interesting" or current chrono we are 26210f87230dSFrancis Yan * tracking and starts busy chrono if we have pending data. 26220f87230dSFrancis Yan */ 262375c119afSEric Dumazet if (tcp_rtx_and_write_queues_empty(sk)) 262405b055e8SFrancis Yan tcp_chrono_set(tp, TCP_CHRONO_UNSPEC); 26250f87230dSFrancis Yan else if (type == tp->chrono_type) 26260f87230dSFrancis Yan tcp_chrono_set(tp, TCP_CHRONO_BUSY); 262705b055e8SFrancis Yan } 262805b055e8SFrancis Yan 26291da177e4SLinus Torvalds /* This routine writes packets to the network. It advances the 26301da177e4SLinus Torvalds * send_head. This happens as incoming acks open up the remote 26311da177e4SLinus Torvalds * window for us. 26321da177e4SLinus Torvalds * 2633f8269a49SIlpo Järvinen * LARGESEND note: !tcp_urg_mode is overkill, only frames between 2634f8269a49SIlpo Järvinen * snd_up-64k-mss .. snd_up cannot be large. However, taking into 2635f8269a49SIlpo Järvinen * account rare use of URG, this is not a big flaw. 2636f8269a49SIlpo Järvinen * 26376ba8a3b1SNandita Dukkipati * Send at most one packet when push_one > 0. Temporarily ignore 26386ba8a3b1SNandita Dukkipati * cwnd limit to force at most one packet out when push_one == 2. 26396ba8a3b1SNandita Dukkipati 2640a2a385d6SEric Dumazet * Returns true, if no segments are in flight and we have queued segments, 2641a2a385d6SEric Dumazet * but cannot send anything now because of SWS or another problem. 26421da177e4SLinus Torvalds */ 2643a2a385d6SEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 2644d5dd9175SIlpo Järvinen int push_one, gfp_t gfp) 26451da177e4SLinus Torvalds { 26461da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 264792df7b51SDavid S. Miller struct sk_buff *skb; 2648c1b4a7e6SDavid S. Miller unsigned int tso_segs, sent_pkts; 2649c1b4a7e6SDavid S. Miller int cwnd_quota; 26505d424d5aSJohn Heffner int result; 26515615f886SFrancis Yan bool is_cwnd_limited = false, is_rwnd_limited = false; 2652605ad7f1SEric Dumazet u32 max_segs; 26531da177e4SLinus Torvalds 2654c1b4a7e6SDavid S. Miller sent_pkts = 0; 26555d424d5aSJohn Heffner 2656ee1836aeSEric Dumazet tcp_mstamp_refresh(tp); 2657d5dd9175SIlpo Järvinen if (!push_one) { 26585d424d5aSJohn Heffner /* Do MTU probing. */ 2659d5dd9175SIlpo Järvinen result = tcp_mtu_probe(sk); 2660d5dd9175SIlpo Järvinen if (!result) { 2661a2a385d6SEric Dumazet return false; 26625d424d5aSJohn Heffner } else if (result > 0) { 26635d424d5aSJohn Heffner sent_pkts = 1; 26645d424d5aSJohn Heffner } 2665d5dd9175SIlpo Järvinen } 26665d424d5aSJohn Heffner 2667ed6e7268SNeal Cardwell max_segs = tcp_tso_segs(sk, mss_now); 2668fe067e8aSDavid S. Miller while ((skb = tcp_send_head(sk))) { 2669c8ac3774SHerbert Xu unsigned int limit; 2670c8ac3774SHerbert Xu 267179861919SEric Dumazet if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) { 267279861919SEric Dumazet /* "skb_mstamp_ns" is used as a start point for the retransmit timer */ 2673a1ac9c8aSMartin KaFai Lau tp->tcp_wstamp_ns = tp->tcp_clock_cache; 2674a1ac9c8aSMartin KaFai Lau skb_set_delivery_time(skb, tp->tcp_wstamp_ns, true); 267579861919SEric Dumazet list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue); 2676bf50b606SEric Dumazet tcp_init_tso_segs(skb, mss_now); 267779861919SEric Dumazet goto repair; /* Skip network transmission */ 267879861919SEric Dumazet } 267979861919SEric Dumazet 2680218af599SEric Dumazet if (tcp_pacing_check(sk)) 2681218af599SEric Dumazet break; 2682218af599SEric Dumazet 26835bbb432cSEric Dumazet tso_segs = tcp_init_tso_segs(skb, mss_now); 2684c1b4a7e6SDavid S. Miller BUG_ON(!tso_segs); 2685c1b4a7e6SDavid S. Miller 2686b68e9f85SHerbert Xu cwnd_quota = tcp_cwnd_test(tp, skb); 26876ba8a3b1SNandita Dukkipati if (!cwnd_quota) { 26886ba8a3b1SNandita Dukkipati if (push_one == 2) 26896ba8a3b1SNandita Dukkipati /* Force out a loss probe pkt. */ 26906ba8a3b1SNandita Dukkipati cwnd_quota = 1; 26916ba8a3b1SNandita Dukkipati else 2692b68e9f85SHerbert Xu break; 26936ba8a3b1SNandita Dukkipati } 2694b68e9f85SHerbert Xu 26955615f886SFrancis Yan if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) { 26965615f886SFrancis Yan is_rwnd_limited = true; 2697b68e9f85SHerbert Xu break; 26985615f886SFrancis Yan } 2699b68e9f85SHerbert Xu 2700d6a4e26aSEric Dumazet if (tso_segs == 1) { 2701aa93466bSDavid S. Miller if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 2702aa93466bSDavid S. Miller (tcp_skb_is_last(sk, skb) ? 2703aa93466bSDavid S. Miller nonagle : TCP_NAGLE_PUSH)))) 2704aa93466bSDavid S. Miller break; 2705c1b4a7e6SDavid S. Miller } else { 2706ca8a2263SNeal Cardwell if (!push_one && 2707605ad7f1SEric Dumazet tcp_tso_should_defer(sk, skb, &is_cwnd_limited, 2708f9bfe4e6SEric Dumazet &is_rwnd_limited, max_segs)) 2709aa93466bSDavid S. Miller break; 2710c1b4a7e6SDavid S. Miller } 2711aa93466bSDavid S. Miller 2712605ad7f1SEric Dumazet limit = mss_now; 2713d6a4e26aSEric Dumazet if (tso_segs > 1 && !tcp_urg_mode(tp)) 2714605ad7f1SEric Dumazet limit = tcp_mss_split_point(sk, skb, mss_now, 2715605ad7f1SEric Dumazet min_t(unsigned int, 2716605ad7f1SEric Dumazet cwnd_quota, 2717605ad7f1SEric Dumazet max_segs), 2718605ad7f1SEric Dumazet nonagle); 2719605ad7f1SEric Dumazet 2720605ad7f1SEric Dumazet if (skb->len > limit && 272156483341SEric Dumazet unlikely(tso_fragment(sk, skb, limit, mss_now, gfp))) 2722605ad7f1SEric Dumazet break; 2723605ad7f1SEric Dumazet 2724f9616c35SEric Dumazet if (tcp_small_queue_check(sk, skb, 0)) 272546d3ceabSEric Dumazet break; 2726c9eeec26SEric Dumazet 27271f85e626SEric Dumazet /* Argh, we hit an empty skb(), presumably a thread 27281f85e626SEric Dumazet * is sleeping in sendmsg()/sk_stream_wait_memory(). 27291f85e626SEric Dumazet * We do not want to send a pure-ack packet and have 27301f85e626SEric Dumazet * a strange looking rtx queue with empty packet(s). 27311f85e626SEric Dumazet */ 27321f85e626SEric Dumazet if (TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq) 27331f85e626SEric Dumazet break; 27341f85e626SEric Dumazet 2735d5dd9175SIlpo Järvinen if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) 27361da177e4SLinus Torvalds break; 27371da177e4SLinus Torvalds 2738ec342325SAndrew Vagin repair: 27391da177e4SLinus Torvalds /* Advance the send_head. This one is sent out. 27401da177e4SLinus Torvalds * This call will increment packets_out. 27411da177e4SLinus Torvalds */ 274266f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 27431da177e4SLinus Torvalds 27441da177e4SLinus Torvalds tcp_minshall_update(tp, mss_now, skb); 2745a262f0cdSNandita Dukkipati sent_pkts += tcp_skb_pcount(skb); 2746d5dd9175SIlpo Järvinen 2747d5dd9175SIlpo Järvinen if (push_one) 2748d5dd9175SIlpo Järvinen break; 27491da177e4SLinus Torvalds } 27501da177e4SLinus Torvalds 27515615f886SFrancis Yan if (is_rwnd_limited) 27525615f886SFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED); 27535615f886SFrancis Yan else 27545615f886SFrancis Yan tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED); 27555615f886SFrancis Yan 275640570375SEric Dumazet is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tcp_snd_cwnd(tp)); 2757299bcb55SNeal Cardwell if (likely(sent_pkts || is_cwnd_limited)) 2758299bcb55SNeal Cardwell tcp_cwnd_validate(sk, is_cwnd_limited); 2759299bcb55SNeal Cardwell 2760aa93466bSDavid S. Miller if (likely(sent_pkts)) { 2761684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 2762684bad11SYuchung Cheng tp->prr_out += sent_pkts; 27636ba8a3b1SNandita Dukkipati 27646ba8a3b1SNandita Dukkipati /* Send one loss probe per tail loss episode. */ 27656ba8a3b1SNandita Dukkipati if (push_one != 2) 2766ed66dfafSNeal Cardwell tcp_schedule_loss_probe(sk, false); 2767a2a385d6SEric Dumazet return false; 27681da177e4SLinus Torvalds } 276975c119afSEric Dumazet return !tp->packets_out && !tcp_write_queue_empty(sk); 27706ba8a3b1SNandita Dukkipati } 27716ba8a3b1SNandita Dukkipati 2772ed66dfafSNeal Cardwell bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto) 27736ba8a3b1SNandita Dukkipati { 27746ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 27756ba8a3b1SNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 2776a2815817SNeal Cardwell u32 timeout, rto_delta_us; 27772ae21cf5SEric Dumazet int early_retrans; 27786ba8a3b1SNandita Dukkipati 27796ba8a3b1SNandita Dukkipati /* Don't do any loss probe on a Fast Open connection before 3WHS 27806ba8a3b1SNandita Dukkipati * finishes. 27816ba8a3b1SNandita Dukkipati */ 2782d983ea6fSEric Dumazet if (rcu_access_pointer(tp->fastopen_rsk)) 27836ba8a3b1SNandita Dukkipati return false; 27846ba8a3b1SNandita Dukkipati 278552e65865SKuniyuki Iwashima early_retrans = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_early_retrans); 27866ba8a3b1SNandita Dukkipati /* Schedule a loss probe in 2*RTT for SACK capable connections 2787b4f70c3dSNeal Cardwell * not in loss recovery, that are either limited by cwnd or application. 27886ba8a3b1SNandita Dukkipati */ 27892ae21cf5SEric Dumazet if ((early_retrans != 3 && early_retrans != 4) || 2790bec41a11SYuchung Cheng !tp->packets_out || !tcp_is_sack(tp) || 2791b4f70c3dSNeal Cardwell (icsk->icsk_ca_state != TCP_CA_Open && 2792b4f70c3dSNeal Cardwell icsk->icsk_ca_state != TCP_CA_CWR)) 27936ba8a3b1SNandita Dukkipati return false; 27946ba8a3b1SNandita Dukkipati 2795bb4d991aSYuchung Cheng /* Probe timeout is 2*rtt. Add minimum RTO to account 2796f9b99582SYuchung Cheng * for delayed ack when there's one outstanding packet. If no RTT 2797f9b99582SYuchung Cheng * sample is available then probe after TCP_TIMEOUT_INIT. 27986ba8a3b1SNandita Dukkipati */ 2799bb4d991aSYuchung Cheng if (tp->srtt_us) { 2800bb4d991aSYuchung Cheng timeout = usecs_to_jiffies(tp->srtt_us >> 2); 28016ba8a3b1SNandita Dukkipati if (tp->packets_out == 1) 2802bb4d991aSYuchung Cheng timeout += TCP_RTO_MIN; 2803bb4d991aSYuchung Cheng else 2804bb4d991aSYuchung Cheng timeout += TCP_TIMEOUT_MIN; 2805bb4d991aSYuchung Cheng } else { 2806bb4d991aSYuchung Cheng timeout = TCP_TIMEOUT_INIT; 2807bb4d991aSYuchung Cheng } 28086ba8a3b1SNandita Dukkipati 2809a2815817SNeal Cardwell /* If the RTO formula yields an earlier time, then use that time. */ 2810ed66dfafSNeal Cardwell rto_delta_us = advancing_rto ? 2811ed66dfafSNeal Cardwell jiffies_to_usecs(inet_csk(sk)->icsk_rto) : 2812ed66dfafSNeal Cardwell tcp_rto_delta_us(sk); /* How far in future is RTO? */ 2813a2815817SNeal Cardwell if (rto_delta_us > 0) 2814a2815817SNeal Cardwell timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us)); 28156ba8a3b1SNandita Dukkipati 28168dc242adSEric Dumazet tcp_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, TCP_RTO_MAX); 28176ba8a3b1SNandita Dukkipati return true; 28186ba8a3b1SNandita Dukkipati } 28196ba8a3b1SNandita Dukkipati 28201f3279aeSEric Dumazet /* Thanks to skb fast clones, we can detect if a prior transmit of 28211f3279aeSEric Dumazet * a packet is still in a qdisc or driver queue. 28221f3279aeSEric Dumazet * In this case, there is very little point doing a retransmit ! 28231f3279aeSEric Dumazet */ 2824f4dae54eSEric Dumazet static bool skb_still_in_host_queue(struct sock *sk, 28251f3279aeSEric Dumazet const struct sk_buff *skb) 28261f3279aeSEric Dumazet { 282739bb5e62SEric Dumazet if (unlikely(skb_fclone_busy(sk, skb))) { 2828f4dae54eSEric Dumazet set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags); 2829f4dae54eSEric Dumazet smp_mb__after_atomic(); 2830f4dae54eSEric Dumazet if (skb_fclone_busy(sk, skb)) { 2831c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), 28321f3279aeSEric Dumazet LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES); 28331f3279aeSEric Dumazet return true; 28341f3279aeSEric Dumazet } 2835f4dae54eSEric Dumazet } 28361f3279aeSEric Dumazet return false; 28371f3279aeSEric Dumazet } 28381f3279aeSEric Dumazet 2839b340b264SYuchung Cheng /* When probe timeout (PTO) fires, try send a new segment if possible, else 28406ba8a3b1SNandita Dukkipati * retransmit the last segment. 28416ba8a3b1SNandita Dukkipati */ 28426ba8a3b1SNandita Dukkipati void tcp_send_loss_probe(struct sock *sk) 28436ba8a3b1SNandita Dukkipati { 28449b717a8dSNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 28456ba8a3b1SNandita Dukkipati struct sk_buff *skb; 28466ba8a3b1SNandita Dukkipati int pcount; 28476ba8a3b1SNandita Dukkipati int mss = tcp_current_mss(sk); 28486ba8a3b1SNandita Dukkipati 284976be93fcSYuchung Cheng /* At most one outstanding TLP */ 285076be93fcSYuchung Cheng if (tp->tlp_high_seq) 285176be93fcSYuchung Cheng goto rearm_timer; 285276be93fcSYuchung Cheng 285376be93fcSYuchung Cheng tp->tlp_retrans = 0; 2854b340b264SYuchung Cheng skb = tcp_send_head(sk); 285575c119afSEric Dumazet if (skb && tcp_snd_wnd_test(tp, skb, mss)) { 2856b340b264SYuchung Cheng pcount = tp->packets_out; 2857b340b264SYuchung Cheng tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC); 2858b340b264SYuchung Cheng if (tp->packets_out > pcount) 2859b340b264SYuchung Cheng goto probe_sent; 28606ba8a3b1SNandita Dukkipati goto rearm_timer; 28616ba8a3b1SNandita Dukkipati } 286275c119afSEric Dumazet skb = skb_rb_last(&sk->tcp_rtx_queue); 2863b2b7af86SYuchung Cheng if (unlikely(!skb)) { 2864b2b7af86SYuchung Cheng WARN_ONCE(tp->packets_out, 2865b2b7af86SYuchung Cheng "invalid inflight: %u state %u cwnd %u mss %d\n", 286640570375SEric Dumazet tp->packets_out, sk->sk_state, tcp_snd_cwnd(tp), mss); 2867b2b7af86SYuchung Cheng inet_csk(sk)->icsk_pending = 0; 2868b2b7af86SYuchung Cheng return; 2869b2b7af86SYuchung Cheng } 28706ba8a3b1SNandita Dukkipati 28711f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 28721f3279aeSEric Dumazet goto rearm_timer; 28731f3279aeSEric Dumazet 28746ba8a3b1SNandita Dukkipati pcount = tcp_skb_pcount(skb); 28756ba8a3b1SNandita Dukkipati if (WARN_ON(!pcount)) 28766ba8a3b1SNandita Dukkipati goto rearm_timer; 28776ba8a3b1SNandita Dukkipati 28786ba8a3b1SNandita Dukkipati if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) { 287975c119afSEric Dumazet if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, 288075c119afSEric Dumazet (pcount - 1) * mss, mss, 28816cc55e09SOctavian Purdila GFP_ATOMIC))) 28826ba8a3b1SNandita Dukkipati goto rearm_timer; 288375c119afSEric Dumazet skb = skb_rb_next(skb); 28846ba8a3b1SNandita Dukkipati } 28856ba8a3b1SNandita Dukkipati 28866ba8a3b1SNandita Dukkipati if (WARN_ON(!skb || !tcp_skb_pcount(skb))) 28876ba8a3b1SNandita Dukkipati goto rearm_timer; 28886ba8a3b1SNandita Dukkipati 288910d3be56SEric Dumazet if (__tcp_retransmit_skb(sk, skb, 1)) 2890b340b264SYuchung Cheng goto rearm_timer; 28916ba8a3b1SNandita Dukkipati 289276be93fcSYuchung Cheng tp->tlp_retrans = 1; 289376be93fcSYuchung Cheng 289476be93fcSYuchung Cheng probe_sent: 28959b717a8dSNandita Dukkipati /* Record snd_nxt for loss detection. */ 28969b717a8dSNandita Dukkipati tp->tlp_high_seq = tp->snd_nxt; 28979b717a8dSNandita Dukkipati 2898c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES); 2899fcd16c0aSYuchung Cheng /* Reset s.t. tcp_rearm_rto will restart timer from now */ 2900fcd16c0aSYuchung Cheng inet_csk(sk)->icsk_pending = 0; 2901b340b264SYuchung Cheng rearm_timer: 2902fcd16c0aSYuchung Cheng tcp_rearm_rto(sk); 29031da177e4SLinus Torvalds } 29041da177e4SLinus Torvalds 2905a762a980SDavid S. Miller /* Push out any pending frames which were held back due to 2906a762a980SDavid S. Miller * TCP_CORK or attempt at coalescing tiny packets. 2907a762a980SDavid S. Miller * The socket must be locked by the caller. 2908a762a980SDavid S. Miller */ 29099e412ba7SIlpo Järvinen void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 29109e412ba7SIlpo Järvinen int nonagle) 2911a762a980SDavid S. Miller { 2912726e07a8SIlpo Järvinen /* If we are closed, the bytes will have to remain here. 2913726e07a8SIlpo Järvinen * In time closedown will finish, we empty the write queue and 2914726e07a8SIlpo Järvinen * all will be happy. 2915726e07a8SIlpo Järvinen */ 2916726e07a8SIlpo Järvinen if (unlikely(sk->sk_state == TCP_CLOSE)) 2917726e07a8SIlpo Järvinen return; 2918726e07a8SIlpo Järvinen 291999a1dec7SMel Gorman if (tcp_write_xmit(sk, cur_mss, nonagle, 0, 29207450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC))) 29219e412ba7SIlpo Järvinen tcp_check_probe_timer(sk); 2922a762a980SDavid S. Miller } 2923a762a980SDavid S. Miller 2924c1b4a7e6SDavid S. Miller /* Send _single_ skb sitting at the send head. This function requires 2925c1b4a7e6SDavid S. Miller * true push pending frames to setup probe timer etc. 2926c1b4a7e6SDavid S. Miller */ 2927c1b4a7e6SDavid S. Miller void tcp_push_one(struct sock *sk, unsigned int mss_now) 2928c1b4a7e6SDavid S. Miller { 2929fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 2930c1b4a7e6SDavid S. Miller 2931c1b4a7e6SDavid S. Miller BUG_ON(!skb || skb->len < mss_now); 2932c1b4a7e6SDavid S. Miller 2933d5dd9175SIlpo Järvinen tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); 2934c1b4a7e6SDavid S. Miller } 2935c1b4a7e6SDavid S. Miller 29361da177e4SLinus Torvalds /* This function returns the amount that we can raise the 29371da177e4SLinus Torvalds * usable window based on the following constraints 29381da177e4SLinus Torvalds * 29391da177e4SLinus Torvalds * 1. The window can never be shrunk once it is offered (RFC 793) 29401da177e4SLinus Torvalds * 2. We limit memory per socket 29411da177e4SLinus Torvalds * 29421da177e4SLinus Torvalds * RFC 1122: 29431da177e4SLinus Torvalds * "the suggested [SWS] avoidance algorithm for the receiver is to keep 29441da177e4SLinus Torvalds * RECV.NEXT + RCV.WIN fixed until: 29451da177e4SLinus Torvalds * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 29461da177e4SLinus Torvalds * 29471da177e4SLinus Torvalds * i.e. don't raise the right edge of the window until you can raise 29481da177e4SLinus Torvalds * it at least MSS bytes. 29491da177e4SLinus Torvalds * 29501da177e4SLinus Torvalds * Unfortunately, the recommended algorithm breaks header prediction, 29511da177e4SLinus Torvalds * since header prediction assumes th->window stays fixed. 29521da177e4SLinus Torvalds * 29531da177e4SLinus Torvalds * Strictly speaking, keeping th->window fixed violates the receiver 29541da177e4SLinus Torvalds * side SWS prevention criteria. The problem is that under this rule 29551da177e4SLinus Torvalds * a stream of single byte packets will cause the right side of the 29561da177e4SLinus Torvalds * window to always advance by a single byte. 29571da177e4SLinus Torvalds * 29581da177e4SLinus Torvalds * Of course, if the sender implements sender side SWS prevention 29591da177e4SLinus Torvalds * then this will not be a problem. 29601da177e4SLinus Torvalds * 29611da177e4SLinus Torvalds * BSD seems to make the following compromise: 29621da177e4SLinus Torvalds * 29631da177e4SLinus Torvalds * If the free space is less than the 1/4 of the maximum 29641da177e4SLinus Torvalds * space available and the free space is less than 1/2 mss, 29651da177e4SLinus Torvalds * then set the window to 0. 29661da177e4SLinus Torvalds * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 29671da177e4SLinus Torvalds * Otherwise, just prevent the window from shrinking 29681da177e4SLinus Torvalds * and from being larger than the largest representable value. 29691da177e4SLinus Torvalds * 29701da177e4SLinus Torvalds * This prevents incremental opening of the window in the regime 29711da177e4SLinus Torvalds * where TCP is limited by the speed of the reader side taking 29721da177e4SLinus Torvalds * data out of the TCP receive queue. It does nothing about 29731da177e4SLinus Torvalds * those cases where the window is constrained on the sender side 29741da177e4SLinus Torvalds * because the pipeline is full. 29751da177e4SLinus Torvalds * 29761da177e4SLinus Torvalds * BSD also seems to "accidentally" limit itself to windows that are a 29771da177e4SLinus Torvalds * multiple of MSS, at least until the free space gets quite small. 29781da177e4SLinus Torvalds * This would appear to be a side effect of the mbuf implementation. 29791da177e4SLinus Torvalds * Combining these two algorithms results in the observed behavior 29801da177e4SLinus Torvalds * of having a fixed window size at almost all times. 29811da177e4SLinus Torvalds * 29821da177e4SLinus Torvalds * Below we obtain similar behavior by forcing the offered window to 29831da177e4SLinus Torvalds * a multiple of the mss when it is feasible to do so. 29841da177e4SLinus Torvalds * 29851da177e4SLinus Torvalds * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 29861da177e4SLinus Torvalds * Regular options like TIMESTAMP are taken into account. 29871da177e4SLinus Torvalds */ 29881da177e4SLinus Torvalds u32 __tcp_select_window(struct sock *sk) 29891da177e4SLinus Torvalds { 2990463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 29911da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2992*b650d953Smfreemon@cloudflare.com struct net *net = sock_net(sk); 2993caa20d9aSStephen Hemminger /* MSS for the peer's data. Previous versions used mss_clamp 29941da177e4SLinus Torvalds * here. I don't know if the value based on our guesses 29951da177e4SLinus Torvalds * of peer's MSS is better for the performance. It's more correct 29961da177e4SLinus Torvalds * but may be worse for the performance because of rcv_mss 29971da177e4SLinus Torvalds * fluctuations. --SAW 1998/11/1 29981da177e4SLinus Torvalds */ 2999463c84b9SArnaldo Carvalho de Melo int mss = icsk->icsk_ack.rcv_mss; 30001da177e4SLinus Torvalds int free_space = tcp_space(sk); 300186c1a045SFlorian Westphal int allowed_space = tcp_full_space(sk); 3002071c8ed6SFlorian Westphal int full_space, window; 3003071c8ed6SFlorian Westphal 3004071c8ed6SFlorian Westphal if (sk_is_mptcp(sk)) 3005071c8ed6SFlorian Westphal mptcp_space(sk, &free_space, &allowed_space); 3006071c8ed6SFlorian Westphal 3007071c8ed6SFlorian Westphal full_space = min_t(int, tp->window_clamp, allowed_space); 30081da177e4SLinus Torvalds 300906425c30SEric Dumazet if (unlikely(mss > full_space)) { 30101da177e4SLinus Torvalds mss = full_space; 301106425c30SEric Dumazet if (mss <= 0) 301206425c30SEric Dumazet return 0; 301306425c30SEric Dumazet } 3014*b650d953Smfreemon@cloudflare.com 3015*b650d953Smfreemon@cloudflare.com /* Only allow window shrink if the sysctl is enabled and we have 3016*b650d953Smfreemon@cloudflare.com * a non-zero scaling factor in effect. 3017*b650d953Smfreemon@cloudflare.com */ 3018*b650d953Smfreemon@cloudflare.com if (READ_ONCE(net->ipv4.sysctl_tcp_shrink_window) && tp->rx_opt.rcv_wscale) 3019*b650d953Smfreemon@cloudflare.com goto shrink_window_allowed; 3020*b650d953Smfreemon@cloudflare.com 3021*b650d953Smfreemon@cloudflare.com /* do not allow window to shrink */ 3022*b650d953Smfreemon@cloudflare.com 3023b92edbe0SEric Dumazet if (free_space < (full_space >> 1)) { 3024463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.quick = 0; 30251da177e4SLinus Torvalds 3026b8da51ebSEric Dumazet if (tcp_under_memory_pressure(sk)) 3027053f3684SWei Wang tcp_adjust_rcv_ssthresh(sk); 30281da177e4SLinus Torvalds 302986c1a045SFlorian Westphal /* free_space might become our new window, make sure we don't 303086c1a045SFlorian Westphal * increase it due to wscale. 303186c1a045SFlorian Westphal */ 303286c1a045SFlorian Westphal free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale); 303386c1a045SFlorian Westphal 303486c1a045SFlorian Westphal /* if free space is less than mss estimate, or is below 1/16th 303586c1a045SFlorian Westphal * of the maximum allowed, try to move to zero-window, else 303686c1a045SFlorian Westphal * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and 303786c1a045SFlorian Westphal * new incoming data is dropped due to memory limits. 303886c1a045SFlorian Westphal * With large window, mss test triggers way too late in order 303986c1a045SFlorian Westphal * to announce zero window in time before rmem limit kicks in. 304086c1a045SFlorian Westphal */ 304186c1a045SFlorian Westphal if (free_space < (allowed_space >> 4) || free_space < mss) 30421da177e4SLinus Torvalds return 0; 30431da177e4SLinus Torvalds } 30441da177e4SLinus Torvalds 30451da177e4SLinus Torvalds if (free_space > tp->rcv_ssthresh) 30461da177e4SLinus Torvalds free_space = tp->rcv_ssthresh; 30471da177e4SLinus Torvalds 30481da177e4SLinus Torvalds /* Don't do rounding if we are using window scaling, since the 30491da177e4SLinus Torvalds * scaled window will not line up with the MSS boundary anyway. 30501da177e4SLinus Torvalds */ 30511da177e4SLinus Torvalds if (tp->rx_opt.rcv_wscale) { 30521da177e4SLinus Torvalds window = free_space; 30531da177e4SLinus Torvalds 30541da177e4SLinus Torvalds /* Advertise enough space so that it won't get scaled away. 30551da177e4SLinus Torvalds * Import case: prevent zero window announcement if 30561da177e4SLinus Torvalds * 1<<rcv_wscale > mss. 30571da177e4SLinus Torvalds */ 30581935299dSGao Feng window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale)); 30591da177e4SLinus Torvalds } else { 30601935299dSGao Feng window = tp->rcv_wnd; 30611da177e4SLinus Torvalds /* Get the largest window that is a nice multiple of mss. 30621da177e4SLinus Torvalds * Window clamp already applied above. 30631da177e4SLinus Torvalds * If our current window offering is within 1 mss of the 30641da177e4SLinus Torvalds * free space we just keep it. This prevents the divide 30651da177e4SLinus Torvalds * and multiply from happening most of the time. 30661da177e4SLinus Torvalds * We also don't do any window rounding when the free space 30671da177e4SLinus Torvalds * is too small. 30681da177e4SLinus Torvalds */ 30691da177e4SLinus Torvalds if (window <= free_space - mss || window > free_space) 30701935299dSGao Feng window = rounddown(free_space, mss); 307184565070SJohn Heffner else if (mss == full_space && 3072b92edbe0SEric Dumazet free_space > window + (full_space >> 1)) 307384565070SJohn Heffner window = free_space; 30741da177e4SLinus Torvalds } 30751da177e4SLinus Torvalds 30761da177e4SLinus Torvalds return window; 3077*b650d953Smfreemon@cloudflare.com 3078*b650d953Smfreemon@cloudflare.com shrink_window_allowed: 3079*b650d953Smfreemon@cloudflare.com /* new window should always be an exact multiple of scaling factor */ 3080*b650d953Smfreemon@cloudflare.com free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale); 3081*b650d953Smfreemon@cloudflare.com 3082*b650d953Smfreemon@cloudflare.com if (free_space < (full_space >> 1)) { 3083*b650d953Smfreemon@cloudflare.com icsk->icsk_ack.quick = 0; 3084*b650d953Smfreemon@cloudflare.com 3085*b650d953Smfreemon@cloudflare.com if (tcp_under_memory_pressure(sk)) 3086*b650d953Smfreemon@cloudflare.com tcp_adjust_rcv_ssthresh(sk); 3087*b650d953Smfreemon@cloudflare.com 3088*b650d953Smfreemon@cloudflare.com /* if free space is too low, return a zero window */ 3089*b650d953Smfreemon@cloudflare.com if (free_space < (allowed_space >> 4) || free_space < mss || 3090*b650d953Smfreemon@cloudflare.com free_space < (1 << tp->rx_opt.rcv_wscale)) 3091*b650d953Smfreemon@cloudflare.com return 0; 3092*b650d953Smfreemon@cloudflare.com } 3093*b650d953Smfreemon@cloudflare.com 3094*b650d953Smfreemon@cloudflare.com if (free_space > tp->rcv_ssthresh) { 3095*b650d953Smfreemon@cloudflare.com free_space = tp->rcv_ssthresh; 3096*b650d953Smfreemon@cloudflare.com /* new window should always be an exact multiple of scaling factor 3097*b650d953Smfreemon@cloudflare.com * 3098*b650d953Smfreemon@cloudflare.com * For this case, we ALIGN "up" (increase free_space) because 3099*b650d953Smfreemon@cloudflare.com * we know free_space is not zero here, it has been reduced from 3100*b650d953Smfreemon@cloudflare.com * the memory-based limit, and rcv_ssthresh is not a hard limit 3101*b650d953Smfreemon@cloudflare.com * (unlike sk_rcvbuf). 3102*b650d953Smfreemon@cloudflare.com */ 3103*b650d953Smfreemon@cloudflare.com free_space = ALIGN(free_space, (1 << tp->rx_opt.rcv_wscale)); 3104*b650d953Smfreemon@cloudflare.com } 3105*b650d953Smfreemon@cloudflare.com 3106*b650d953Smfreemon@cloudflare.com return free_space; 31071da177e4SLinus Torvalds } 31081da177e4SLinus Torvalds 3109cfea5a68SMartin KaFai Lau void tcp_skb_collapse_tstamp(struct sk_buff *skb, 3110082ac2d5SMartin KaFai Lau const struct sk_buff *next_skb) 3111082ac2d5SMartin KaFai Lau { 31120a2cf20cSSoheil Hassas Yeganeh if (unlikely(tcp_has_tx_tstamp(next_skb))) { 31130a2cf20cSSoheil Hassas Yeganeh const struct skb_shared_info *next_shinfo = 31140a2cf20cSSoheil Hassas Yeganeh skb_shinfo(next_skb); 3115082ac2d5SMartin KaFai Lau struct skb_shared_info *shinfo = skb_shinfo(skb); 3116082ac2d5SMartin KaFai Lau 31170a2cf20cSSoheil Hassas Yeganeh shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP; 3118082ac2d5SMartin KaFai Lau shinfo->tskey = next_shinfo->tskey; 31192de8023eSMartin KaFai Lau TCP_SKB_CB(skb)->txstamp_ack |= 31202de8023eSMartin KaFai Lau TCP_SKB_CB(next_skb)->txstamp_ack; 3121082ac2d5SMartin KaFai Lau } 3122082ac2d5SMartin KaFai Lau } 3123082ac2d5SMartin KaFai Lau 31244a17fc3aSIlpo Järvinen /* Collapses two adjacent SKB's during retransmission. */ 3125f8071cdeSEric Dumazet static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) 31261da177e4SLinus Torvalds { 31271da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 312875c119afSEric Dumazet struct sk_buff *next_skb = skb_rb_next(skb); 312913dde04fSWei Yongjun int next_skb_size; 31301da177e4SLinus Torvalds 3131058dc334SIlpo Järvinen next_skb_size = next_skb->len; 31321da177e4SLinus Torvalds 3133058dc334SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); 31341da177e4SLinus Torvalds 3135bd446314SEric Dumazet if (next_skb_size && !tcp_skb_shift(skb, next_skb, 1, next_skb_size)) 3136f8071cdeSEric Dumazet return false; 3137bd446314SEric Dumazet 31382b7cda9cSEric Dumazet tcp_highest_sack_replace(sk, next_skb, skb); 3139a6963a6bSIlpo Järvinen 31401da177e4SLinus Torvalds /* Update sequence range on original skb. */ 31411da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 31421da177e4SLinus Torvalds 3143e6c7d085SIlpo Järvinen /* Merge over control information. This moves PSH/FIN etc. over */ 31444de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; 31451da177e4SLinus Torvalds 31461da177e4SLinus Torvalds /* All done, get rid of second SKB and account for it so 31471da177e4SLinus Torvalds * packet counting does not break. 31481da177e4SLinus Torvalds */ 31494828e7f4SIlpo Järvinen TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; 3150a643b5d4SMartin KaFai Lau TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor; 3151b7689205SIlpo Järvinen 3152b7689205SIlpo Järvinen /* changed transmit queue under us so clear hints */ 3153ef9da47cSIlpo Järvinen tcp_clear_retrans_hints_partial(tp); 3154ef9da47cSIlpo Järvinen if (next_skb == tp->retransmit_skb_hint) 3155ef9da47cSIlpo Järvinen tp->retransmit_skb_hint = skb; 3156b7689205SIlpo Järvinen 3157797108d1SIlpo Järvinen tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); 3158797108d1SIlpo Järvinen 3159082ac2d5SMartin KaFai Lau tcp_skb_collapse_tstamp(skb, next_skb); 3160082ac2d5SMartin KaFai Lau 316175c119afSEric Dumazet tcp_rtx_queue_unlink_and_free(next_skb, sk); 3162f8071cdeSEric Dumazet return true; 31631da177e4SLinus Torvalds } 31641da177e4SLinus Torvalds 316567edfef7SAndi Kleen /* Check if coalescing SKBs is legal. */ 3166a2a385d6SEric Dumazet static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) 31674a17fc3aSIlpo Järvinen { 31684a17fc3aSIlpo Järvinen if (tcp_skb_pcount(skb) > 1) 3169a2a385d6SEric Dumazet return false; 31704a17fc3aSIlpo Järvinen if (skb_cloned(skb)) 3171a2a385d6SEric Dumazet return false; 31722331ccc5SEric Dumazet /* Some heuristics for collapsing over SACK'd could be invented */ 31734a17fc3aSIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 3174a2a385d6SEric Dumazet return false; 31754a17fc3aSIlpo Järvinen 3176a2a385d6SEric Dumazet return true; 31774a17fc3aSIlpo Järvinen } 31784a17fc3aSIlpo Järvinen 317967edfef7SAndi Kleen /* Collapse packets in the retransmit queue to make to create 318067edfef7SAndi Kleen * less packets on the wire. This is only done on retransmission. 318167edfef7SAndi Kleen */ 31824a17fc3aSIlpo Järvinen static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, 31834a17fc3aSIlpo Järvinen int space) 31844a17fc3aSIlpo Järvinen { 31854a17fc3aSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 31864a17fc3aSIlpo Järvinen struct sk_buff *skb = to, *tmp; 3187a2a385d6SEric Dumazet bool first = true; 31884a17fc3aSIlpo Järvinen 31891a63cb91SKuniyuki Iwashima if (!READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse)) 31904a17fc3aSIlpo Järvinen return; 31914de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 31924a17fc3aSIlpo Järvinen return; 31934a17fc3aSIlpo Järvinen 319475c119afSEric Dumazet skb_rbtree_walk_from_safe(skb, tmp) { 31954a17fc3aSIlpo Järvinen if (!tcp_can_collapse(sk, skb)) 31964a17fc3aSIlpo Järvinen break; 31974a17fc3aSIlpo Järvinen 319885712484SMat Martineau if (!tcp_skb_can_collapse(to, skb)) 3199a643b5d4SMartin KaFai Lau break; 3200a643b5d4SMartin KaFai Lau 32014a17fc3aSIlpo Järvinen space -= skb->len; 32024a17fc3aSIlpo Järvinen 32034a17fc3aSIlpo Järvinen if (first) { 3204a2a385d6SEric Dumazet first = false; 32054a17fc3aSIlpo Järvinen continue; 32064a17fc3aSIlpo Järvinen } 32074a17fc3aSIlpo Järvinen 32084a17fc3aSIlpo Järvinen if (space < 0) 32094a17fc3aSIlpo Järvinen break; 32104a17fc3aSIlpo Järvinen 32114a17fc3aSIlpo Järvinen if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) 32124a17fc3aSIlpo Järvinen break; 32134a17fc3aSIlpo Järvinen 3214f8071cdeSEric Dumazet if (!tcp_collapse_retrans(sk, to)) 3215f8071cdeSEric Dumazet break; 32164a17fc3aSIlpo Järvinen } 32174a17fc3aSIlpo Järvinen } 32184a17fc3aSIlpo Järvinen 32191da177e4SLinus Torvalds /* This retransmits one SKB. Policy decisions and retransmit queue 32201da177e4SLinus Torvalds * state updates are done by the caller. Returns non-zero if an 32211da177e4SLinus Torvalds * error occurred which prevented the send. 32221da177e4SLinus Torvalds */ 322310d3be56SEric Dumazet int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) 32241da177e4SLinus Torvalds { 32255d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 322610d3be56SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 32277d227cd2SSridhar Samudrala unsigned int cur_mss; 322810d3be56SEric Dumazet int diff, len, err; 3229536a6c8eSYonglong Li int avail_wnd; 323010d3be56SEric Dumazet 323110d3be56SEric Dumazet /* Inconclusive MTU probe */ 323210d3be56SEric Dumazet if (icsk->icsk_mtup.probe_size) 32335d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 32345d424d5aSJohn Heffner 32351f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 32361f3279aeSEric Dumazet return -EBUSY; 32371f3279aeSEric Dumazet 32381da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 32397f582b24SEric Dumazet if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) { 32407f582b24SEric Dumazet WARN_ON_ONCE(1); 32417f582b24SEric Dumazet return -EINVAL; 32427f582b24SEric Dumazet } 32431da177e4SLinus Torvalds if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 32441da177e4SLinus Torvalds return -ENOMEM; 32451da177e4SLinus Torvalds } 32461da177e4SLinus Torvalds 32477d227cd2SSridhar Samudrala if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 32487d227cd2SSridhar Samudrala return -EHOSTUNREACH; /* Routing failure or similar. */ 32497d227cd2SSridhar Samudrala 32500c54b85fSIlpo Järvinen cur_mss = tcp_current_mss(sk); 3251536a6c8eSYonglong Li avail_wnd = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 32527d227cd2SSridhar Samudrala 32531da177e4SLinus Torvalds /* If receiver has shrunk his window, and skb is out of 32541da177e4SLinus Torvalds * new window, do not retransmit it. The exception is the 32551da177e4SLinus Torvalds * case, when window is shrunk to zero. In this case 3256536a6c8eSYonglong Li * our retransmit of one segment serves as a zero window probe. 32571da177e4SLinus Torvalds */ 3258536a6c8eSYonglong Li if (avail_wnd <= 0) { 3259536a6c8eSYonglong Li if (TCP_SKB_CB(skb)->seq != tp->snd_una) 32601da177e4SLinus Torvalds return -EAGAIN; 3261536a6c8eSYonglong Li avail_wnd = cur_mss; 3262536a6c8eSYonglong Li } 32631da177e4SLinus Torvalds 326410d3be56SEric Dumazet len = cur_mss * segs; 3265536a6c8eSYonglong Li if (len > avail_wnd) { 3266536a6c8eSYonglong Li len = rounddown(avail_wnd, cur_mss); 3267536a6c8eSYonglong Li if (!len) 3268536a6c8eSYonglong Li len = avail_wnd; 3269536a6c8eSYonglong Li } 327010d3be56SEric Dumazet if (skb->len > len) { 327175c119afSEric Dumazet if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len, 327275c119afSEric Dumazet cur_mss, GFP_ATOMIC)) 32731da177e4SLinus Torvalds return -ENOMEM; /* We'll try again later. */ 327402276f3cSIlpo Järvinen } else { 3275c4777efaSEric Dumazet if (skb_unclone_keeptruesize(skb, GFP_ATOMIC)) 3276c52e2421SEric Dumazet return -ENOMEM; 327710d3be56SEric Dumazet 327810d3be56SEric Dumazet diff = tcp_skb_pcount(skb); 327910d3be56SEric Dumazet tcp_set_skb_tso_segs(skb, cur_mss); 328010d3be56SEric Dumazet diff -= tcp_skb_pcount(skb); 328110d3be56SEric Dumazet if (diff) 328210d3be56SEric Dumazet tcp_adjust_pcount(sk, skb, diff); 3283536a6c8eSYonglong Li avail_wnd = min_t(int, avail_wnd, cur_mss); 3284536a6c8eSYonglong Li if (skb->len < avail_wnd) 3285536a6c8eSYonglong Li tcp_retrans_try_collapse(sk, skb, avail_wnd); 32861da177e4SLinus Torvalds } 32871da177e4SLinus Torvalds 328849213555SDaniel Borkmann /* RFC3168, section 6.1.1.1. ECN fallback */ 328949213555SDaniel Borkmann if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN) 329049213555SDaniel Borkmann tcp_ecn_clear_syn(sk, skb); 329149213555SDaniel Borkmann 3292678550c6SYuchung Cheng /* Update global and local TCP statistics. */ 3293678550c6SYuchung Cheng segs = tcp_skb_pcount(skb); 3294678550c6SYuchung Cheng TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs); 3295678550c6SYuchung Cheng if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 3296678550c6SYuchung Cheng __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 3297678550c6SYuchung Cheng tp->total_retrans += segs; 3298fb31c9b9SWei Wang tp->bytes_retrans += skb->len; 3299678550c6SYuchung Cheng 330050bceae9SThomas Graf /* make sure skb->data is aligned on arches that require it 330150bceae9SThomas Graf * and check if ack-trimming & collapsing extended the headroom 330250bceae9SThomas Graf * beyond what csum_start can cover. 330350bceae9SThomas Graf */ 330450bceae9SThomas Graf if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) || 330550bceae9SThomas Graf skb_headroom(skb) >= 0xFFFF)) { 330610a81980SEric Dumazet struct sk_buff *nskb; 330710a81980SEric Dumazet 3308e2080072SEric Dumazet tcp_skb_tsorted_save(skb) { 330910a81980SEric Dumazet nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC); 331007f8e4d0SFlorian Westphal if (nskb) { 331107f8e4d0SFlorian Westphal nskb->dev = NULL; 331207f8e4d0SFlorian Westphal err = tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC); 331307f8e4d0SFlorian Westphal } else { 331407f8e4d0SFlorian Westphal err = -ENOBUFS; 331507f8e4d0SFlorian Westphal } 3316e2080072SEric Dumazet } tcp_skb_tsorted_restore(skb); 3317e2080072SEric Dumazet 33185889e2c0SYousuk Seung if (!err) { 3319a7a25630SEric Dumazet tcp_update_skb_after_send(sk, skb, tp->tcp_wstamp_ns); 33205889e2c0SYousuk Seung tcp_rate_skb_sent(sk, skb); 33215889e2c0SYousuk Seung } 3322117632e6SEric Dumazet } else { 3323c84a5711SYuchung Cheng err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 3324117632e6SEric Dumazet } 3325c84a5711SYuchung Cheng 33267f12422cSYuchung Cheng /* To avoid taking spuriously low RTT samples based on a timestamp 33277f12422cSYuchung Cheng * for a transmit that never happened, always mark EVER_RETRANS 33287f12422cSYuchung Cheng */ 33297f12422cSYuchung Cheng TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS; 33307f12422cSYuchung Cheng 3331a31ad29eSLawrence Brakmo if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RETRANS_CB_FLAG)) 3332a31ad29eSLawrence Brakmo tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RETRANS_CB, 3333a31ad29eSLawrence Brakmo TCP_SKB_CB(skb)->seq, segs, err); 3334a31ad29eSLawrence Brakmo 3335fc9f3501SEric Dumazet if (likely(!err)) { 3336e086101bSCong Wang trace_tcp_retransmit_skb(sk, skb); 3337678550c6SYuchung Cheng } else if (err != -EBUSY) { 3338ec641b39SYuchung Cheng NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL, segs); 3339fc9f3501SEric Dumazet } 3340c84a5711SYuchung Cheng return err; 334193b174adSYuchung Cheng } 334293b174adSYuchung Cheng 334310d3be56SEric Dumazet int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) 334493b174adSYuchung Cheng { 334593b174adSYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 334610d3be56SEric Dumazet int err = __tcp_retransmit_skb(sk, skb, segs); 33471da177e4SLinus Torvalds 33481da177e4SLinus Torvalds if (err == 0) { 33491da177e4SLinus Torvalds #if FASTRETRANS_DEBUG > 0 33501da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 3351e87cc472SJoe Perches net_dbg_ratelimited("retrans_out leaked\n"); 33521da177e4SLinus Torvalds } 33531da177e4SLinus Torvalds #endif 33541da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 33551da177e4SLinus Torvalds tp->retrans_out += tcp_skb_pcount(skb); 33567ae18975SYuchung Cheng } 33571da177e4SLinus Torvalds 33587ae18975SYuchung Cheng /* Save stamp of the first (attempted) retransmit. */ 33591da177e4SLinus Torvalds if (!tp->retrans_stamp) 33607faee5c0SEric Dumazet tp->retrans_stamp = tcp_skb_timestamp(skb); 33611da177e4SLinus Torvalds 33626e08d5e3SYuchung Cheng if (tp->undo_retrans < 0) 33636e08d5e3SYuchung Cheng tp->undo_retrans = 0; 33646e08d5e3SYuchung Cheng tp->undo_retrans += tcp_skb_pcount(skb); 33651da177e4SLinus Torvalds return err; 33661da177e4SLinus Torvalds } 33671da177e4SLinus Torvalds 33681da177e4SLinus Torvalds /* This gets called after a retransmit timeout, and the initially 33691da177e4SLinus Torvalds * retransmitted data is acknowledged. It tries to continue 33701da177e4SLinus Torvalds * resending the rest of the retransmit queue, until either 33711da177e4SLinus Torvalds * we've sent it all or the congestion window limit is reached. 33721da177e4SLinus Torvalds */ 33731da177e4SLinus Torvalds void tcp_xmit_retransmit_queue(struct sock *sk) 33741da177e4SLinus Torvalds { 33756687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 3376b9f1f1ceSEric Dumazet struct sk_buff *skb, *rtx_head, *hole = NULL; 33771da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 3378916e6d1aSEric Dumazet bool rearm_timer = false; 3379840a3cbeSYuchung Cheng u32 max_segs; 338061eb55f4SIlpo Järvinen int mib_idx; 33816a438bbeSStephen Hemminger 338245e77d31SIlpo Järvinen if (!tp->packets_out) 338345e77d31SIlpo Järvinen return; 338445e77d31SIlpo Järvinen 338575c119afSEric Dumazet rtx_head = tcp_rtx_queue_head(sk); 3386b9f1f1ceSEric Dumazet skb = tp->retransmit_skb_hint ?: rtx_head; 3387ed6e7268SNeal Cardwell max_segs = tcp_tso_segs(sk, tcp_current_mss(sk)); 338875c119afSEric Dumazet skb_rbtree_walk_from(skb) { 3389dca0aaf8SEric Dumazet __u8 sacked; 339010d3be56SEric Dumazet int segs; 33911da177e4SLinus Torvalds 3392218af599SEric Dumazet if (tcp_pacing_check(sk)) 3393218af599SEric Dumazet break; 3394218af599SEric Dumazet 33956a438bbeSStephen Hemminger /* we could do better than to assign each time */ 339651456b29SIan Morris if (!hole) 33976a438bbeSStephen Hemminger tp->retransmit_skb_hint = skb; 33986a438bbeSStephen Hemminger 339940570375SEric Dumazet segs = tcp_snd_cwnd(tp) - tcp_packets_in_flight(tp); 340010d3be56SEric Dumazet if (segs <= 0) 3401916e6d1aSEric Dumazet break; 3402dca0aaf8SEric Dumazet sacked = TCP_SKB_CB(skb)->sacked; 3403a3d2e9f8SEric Dumazet /* In case tcp_shift_skb_data() have aggregated large skbs, 3404a3d2e9f8SEric Dumazet * we need to make sure not sending too bigs TSO packets 3405a3d2e9f8SEric Dumazet */ 3406a3d2e9f8SEric Dumazet segs = min_t(int, segs, max_segs); 34070e1c54c2SIlpo Järvinen 3408840a3cbeSYuchung Cheng if (tp->retrans_out >= tp->lost_out) { 3409006f582cSIlpo Järvinen break; 34100e1c54c2SIlpo Järvinen } else if (!(sacked & TCPCB_LOST)) { 341151456b29SIan Morris if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) 34120e1c54c2SIlpo Järvinen hole = skb; 341361eb55f4SIlpo Järvinen continue; 34141da177e4SLinus Torvalds 34150e1c54c2SIlpo Järvinen } else { 34160e1c54c2SIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Loss) 34170e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFASTRETRANS; 34180e1c54c2SIlpo Järvinen else 34190e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; 34200e1c54c2SIlpo Järvinen } 34210e1c54c2SIlpo Järvinen 34220e1c54c2SIlpo Järvinen if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) 342361eb55f4SIlpo Järvinen continue; 342440b215e5SPavel Emelyanov 3425f9616c35SEric Dumazet if (tcp_small_queue_check(sk, skb, 1)) 3426916e6d1aSEric Dumazet break; 3427f9616c35SEric Dumazet 342810d3be56SEric Dumazet if (tcp_retransmit_skb(sk, skb, segs)) 3429916e6d1aSEric Dumazet break; 343024ab6becSYuchung Cheng 3431de1d6578SYuchung Cheng NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb)); 34321da177e4SLinus Torvalds 3433684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 3434a262f0cdSNandita Dukkipati tp->prr_out += tcp_skb_pcount(skb); 3435a262f0cdSNandita Dukkipati 343675c119afSEric Dumazet if (skb == rtx_head && 343757dde7f7SYuchung Cheng icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT) 3438916e6d1aSEric Dumazet rearm_timer = true; 3439916e6d1aSEric Dumazet 3440916e6d1aSEric Dumazet } 3441916e6d1aSEric Dumazet if (rearm_timer) 34423f80e08fSEric Dumazet tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 34433f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, 34448dc242adSEric Dumazet TCP_RTO_MAX); 34451da177e4SLinus Torvalds } 34461da177e4SLinus Torvalds 3447d83769a5SEric Dumazet /* We allow to exceed memory limits for FIN packets to expedite 3448d83769a5SEric Dumazet * connection tear down and (memory) recovery. 3449845704a5SEric Dumazet * Otherwise tcp_send_fin() could be tempted to either delay FIN 3450845704a5SEric Dumazet * or even be forced to close flow without any FIN. 3451a6c5ea4cSEric Dumazet * In general, we want to allow one skb per socket to avoid hangs 3452a6c5ea4cSEric Dumazet * with edge trigger epoll() 3453d83769a5SEric Dumazet */ 3454a6c5ea4cSEric Dumazet void sk_forced_mem_schedule(struct sock *sk, int size) 3455d83769a5SEric Dumazet { 3456c4ee1185SEric Dumazet int delta, amt; 3457d83769a5SEric Dumazet 3458c4ee1185SEric Dumazet delta = size - sk->sk_forward_alloc; 3459c4ee1185SEric Dumazet if (delta <= 0) 3460d83769a5SEric Dumazet return; 3461c4ee1185SEric Dumazet amt = sk_mem_pages(delta); 3462100fdd1fSEric Dumazet sk->sk_forward_alloc += amt << PAGE_SHIFT; 3463e805605cSJohannes Weiner sk_memory_allocated_add(sk, amt); 3464e805605cSJohannes Weiner 3465baac50bbSJohannes Weiner if (mem_cgroup_sockets_enabled && sk->sk_memcg) 34664b1327beSWei Wang mem_cgroup_charge_skmem(sk->sk_memcg, amt, 34674b1327beSWei Wang gfp_memcg_charge() | __GFP_NOFAIL); 3468d83769a5SEric Dumazet } 3469d83769a5SEric Dumazet 3470845704a5SEric Dumazet /* Send a FIN. The caller locks the socket for us. 3471845704a5SEric Dumazet * We should try to send a FIN packet really hard, but eventually give up. 34721da177e4SLinus Torvalds */ 34731da177e4SLinus Torvalds void tcp_send_fin(struct sock *sk) 34741da177e4SLinus Torvalds { 3475ee2aabd3SEric Dumazet struct sk_buff *skb, *tskb, *tail = tcp_write_queue_tail(sk); 34761da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 34771da177e4SLinus Torvalds 3478845704a5SEric Dumazet /* Optimization, tack on the FIN if we have one skb in write queue and 3479845704a5SEric Dumazet * this skb was not yet sent, or we are under memory pressure. 3480845704a5SEric Dumazet * Note: in the latter case, FIN packet will be sent after a timeout, 3481845704a5SEric Dumazet * as TCP stack thinks it has already been transmitted. 34821da177e4SLinus Torvalds */ 3483ee2aabd3SEric Dumazet tskb = tail; 348475c119afSEric Dumazet if (!tskb && tcp_under_memory_pressure(sk)) 348575c119afSEric Dumazet tskb = skb_rb_last(&sk->tcp_rtx_queue); 348675c119afSEric Dumazet 348775c119afSEric Dumazet if (tskb) { 3488845704a5SEric Dumazet TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN; 3489845704a5SEric Dumazet TCP_SKB_CB(tskb)->end_seq++; 34901da177e4SLinus Torvalds tp->write_seq++; 3491ee2aabd3SEric Dumazet if (!tail) { 3492845704a5SEric Dumazet /* This means tskb was already sent. 3493845704a5SEric Dumazet * Pretend we included the FIN on previous transmit. 3494845704a5SEric Dumazet * We need to set tp->snd_nxt to the value it would have 3495845704a5SEric Dumazet * if FIN had been sent. This is because retransmit path 3496845704a5SEric Dumazet * does not change tp->snd_nxt. 3497845704a5SEric Dumazet */ 3498e0d694d6SEric Dumazet WRITE_ONCE(tp->snd_nxt, tp->snd_nxt + 1); 3499845704a5SEric Dumazet return; 3500845704a5SEric Dumazet } 35011da177e4SLinus Torvalds } else { 3502845704a5SEric Dumazet skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation); 3503d1edc085SColin Ian King if (unlikely(!skb)) 3504845704a5SEric Dumazet return; 3505d1edc085SColin Ian King 3506e2080072SEric Dumazet INIT_LIST_HEAD(&skb->tcp_tsorted_anchor); 3507d83769a5SEric Dumazet skb_reserve(skb, MAX_TCP_HEADER); 3508a6c5ea4cSEric Dumazet sk_forced_mem_schedule(sk, skb->truesize); 35091da177e4SLinus Torvalds /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 3510e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tp->write_seq, 3511a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_FIN); 35121da177e4SLinus Torvalds tcp_queue_skb(sk, skb); 35131da177e4SLinus Torvalds } 3514845704a5SEric Dumazet __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF); 35151da177e4SLinus Torvalds } 35161da177e4SLinus Torvalds 35171da177e4SLinus Torvalds /* We get here when a process closes a file descriptor (either due to 35181da177e4SLinus Torvalds * an explicit close() or as a byproduct of exit()'ing) and there 35191da177e4SLinus Torvalds * was unread data in the receive queue. This behavior is recommended 352065bb723cSGerrit Renker * by RFC 2525, section 2.17. -DaveM 35211da177e4SLinus Torvalds */ 3522dd0fc66fSAl Viro void tcp_send_active_reset(struct sock *sk, gfp_t priority) 35231da177e4SLinus Torvalds { 35241da177e4SLinus Torvalds struct sk_buff *skb; 35251da177e4SLinus Torvalds 35267cc2b043SGao Feng TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); 35277cc2b043SGao Feng 35281da177e4SLinus Torvalds /* NOTE: No TCP options attached and we never retransmit this. */ 35291da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, priority); 35301da177e4SLinus Torvalds if (!skb) { 35314e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 35321da177e4SLinus Torvalds return; 35331da177e4SLinus Torvalds } 35341da177e4SLinus Torvalds 35351da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 35361da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 3537e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), 3538a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_RST); 35399a568de4SEric Dumazet tcp_mstamp_refresh(tcp_sk(sk)); 35401da177e4SLinus Torvalds /* Send it off. */ 3541dfb4b9dcSDavid S. Miller if (tcp_transmit_skb(sk, skb, 0, priority)) 35424e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 3543c24b14c4SSong Liu 3544c24b14c4SSong Liu /* skb of trace_tcp_send_reset() keeps the skb that caused RST, 3545c24b14c4SSong Liu * skb here is different to the troublesome skb, so use NULL 3546c24b14c4SSong Liu */ 3547c24b14c4SSong Liu trace_tcp_send_reset(sk, NULL); 35481da177e4SLinus Torvalds } 35491da177e4SLinus Torvalds 355067edfef7SAndi Kleen /* Send a crossed SYN-ACK during socket establishment. 355167edfef7SAndi Kleen * WARNING: This routine must only be called when we have already sent 35521da177e4SLinus Torvalds * a SYN packet that crossed the incoming SYN that caused this routine 35531da177e4SLinus Torvalds * to get called. If this assumption fails then the initial rcv_wnd 35541da177e4SLinus Torvalds * and rcv_wscale values will not be correct. 35551da177e4SLinus Torvalds */ 35561da177e4SLinus Torvalds int tcp_send_synack(struct sock *sk) 35571da177e4SLinus Torvalds { 35581da177e4SLinus Torvalds struct sk_buff *skb; 35591da177e4SLinus Torvalds 356075c119afSEric Dumazet skb = tcp_rtx_queue_head(sk); 356151456b29SIan Morris if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 356275c119afSEric Dumazet pr_err("%s: wrong queue state\n", __func__); 35631da177e4SLinus Torvalds return -EFAULT; 35641da177e4SLinus Torvalds } 35654de075e0SEric Dumazet if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { 35661da177e4SLinus Torvalds if (skb_cloned(skb)) { 3567e2080072SEric Dumazet struct sk_buff *nskb; 3568e2080072SEric Dumazet 3569e2080072SEric Dumazet tcp_skb_tsorted_save(skb) { 3570e2080072SEric Dumazet nskb = skb_copy(skb, GFP_ATOMIC); 3571e2080072SEric Dumazet } tcp_skb_tsorted_restore(skb); 357251456b29SIan Morris if (!nskb) 35731da177e4SLinus Torvalds return -ENOMEM; 3574e2080072SEric Dumazet INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor); 35752bec445fSEric Dumazet tcp_highest_sack_replace(sk, skb, nskb); 357675c119afSEric Dumazet tcp_rtx_queue_unlink_and_free(skb, sk); 3577f4a775d1SEric Dumazet __skb_header_release(nskb); 357875c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb); 3579ab4e846aSEric Dumazet sk_wmem_queued_add(sk, nskb->truesize); 35803ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 35811da177e4SLinus Torvalds skb = nskb; 35821da177e4SLinus Torvalds } 35831da177e4SLinus Torvalds 35844de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; 3585735d3831SFlorian Westphal tcp_ecn_send_synack(sk, skb); 35861da177e4SLinus Torvalds } 3587dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 35881da177e4SLinus Torvalds } 35891da177e4SLinus Torvalds 35904aea39c1SEric Dumazet /** 3591331fca43SMartin KaFai Lau * tcp_make_synack - Allocate one skb and build a SYNACK packet. 3592331fca43SMartin KaFai Lau * @sk: listener socket 3593331fca43SMartin KaFai Lau * @dst: dst entry attached to the SYNACK. It is consumed and caller 3594331fca43SMartin KaFai Lau * should not use it again. 3595331fca43SMartin KaFai Lau * @req: request_sock pointer 3596331fca43SMartin KaFai Lau * @foc: cookie for tcp fast open 3597331fca43SMartin KaFai Lau * @synack_type: Type of synack to prepare 3598331fca43SMartin KaFai Lau * @syn_skb: SYN packet just received. It could be NULL for rtx case. 35994aea39c1SEric Dumazet */ 36005d062de7SEric Dumazet struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst, 3601e6b4d113SWilliam Allen Simpson struct request_sock *req, 3602ca6fb065SEric Dumazet struct tcp_fastopen_cookie *foc, 3603331fca43SMartin KaFai Lau enum tcp_synack_type synack_type, 3604331fca43SMartin KaFai Lau struct sk_buff *syn_skb) 36051da177e4SLinus Torvalds { 36062e6599cbSArnaldo Carvalho de Melo struct inet_request_sock *ireq = inet_rsk(req); 36075d062de7SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 360880f03e27SEric Dumazet struct tcp_md5sig_key *md5 = NULL; 36095d062de7SEric Dumazet struct tcp_out_options opts; 36105d062de7SEric Dumazet struct sk_buff *skb; 3611bd0388aeSWilliam Allen Simpson int tcp_header_size; 36125d062de7SEric Dumazet struct tcphdr *th; 3613f5fff5dcSTom Quetchenbach int mss; 3614a842fe14SEric Dumazet u64 now; 36151da177e4SLinus Torvalds 3616ca6fb065SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 36174aea39c1SEric Dumazet if (unlikely(!skb)) { 36184aea39c1SEric Dumazet dst_release(dst); 36191da177e4SLinus Torvalds return NULL; 36204aea39c1SEric Dumazet } 36211da177e4SLinus Torvalds /* Reserve space for headers. */ 36221da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 36231da177e4SLinus Torvalds 3624b3d05147SEric Dumazet switch (synack_type) { 3625b3d05147SEric Dumazet case TCP_SYNACK_NORMAL: 36269e17f8a4SEric Dumazet skb_set_owner_w(skb, req_to_sk(req)); 3627b3d05147SEric Dumazet break; 3628b3d05147SEric Dumazet case TCP_SYNACK_COOKIE: 3629b3d05147SEric Dumazet /* Under synflood, we do not attach skb to a socket, 3630b3d05147SEric Dumazet * to avoid false sharing. 3631b3d05147SEric Dumazet */ 3632b3d05147SEric Dumazet break; 3633b3d05147SEric Dumazet case TCP_SYNACK_FASTOPEN: 3634ca6fb065SEric Dumazet /* sk is a const pointer, because we want to express multiple 3635ca6fb065SEric Dumazet * cpu might call us concurrently. 3636ca6fb065SEric Dumazet * sk->sk_wmem_alloc in an atomic, we can promote to rw. 3637ca6fb065SEric Dumazet */ 3638ca6fb065SEric Dumazet skb_set_owner_w(skb, (struct sock *)sk); 3639b3d05147SEric Dumazet break; 3640ca6fb065SEric Dumazet } 36414aea39c1SEric Dumazet skb_dst_set(skb, dst); 36421da177e4SLinus Torvalds 36433541f9e8SEric Dumazet mss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 3644f5fff5dcSTom Quetchenbach 364533ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 3646a842fe14SEric Dumazet now = tcp_clock_ns(); 36478b5f12d0SFlorian Westphal #ifdef CONFIG_SYN_COOKIES 3648f8ace8d9SFlorian Westphal if (unlikely(synack_type == TCP_SYNACK_COOKIE && ireq->tstamp_ok)) 3649a1ac9c8aSMartin KaFai Lau skb_set_delivery_time(skb, cookie_init_timestamp(req, now), 3650a1ac9c8aSMartin KaFai Lau true); 36518b5f12d0SFlorian Westphal else 36528b5f12d0SFlorian Westphal #endif 36539e450c1eSYuchung Cheng { 3654a1ac9c8aSMartin KaFai Lau skb_set_delivery_time(skb, now, true); 36559e450c1eSYuchung Cheng if (!tcp_rsk(req)->snt_synack) /* Timestamp first SYNACK */ 36569e450c1eSYuchung Cheng tcp_rsk(req)->snt_synack = tcp_skb_timestamp_us(skb); 36579e450c1eSYuchung Cheng } 365880f03e27SEric Dumazet 365980f03e27SEric Dumazet #ifdef CONFIG_TCP_MD5SIG 366080f03e27SEric Dumazet rcu_read_lock(); 3661fd3a154aSEric Dumazet md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req)); 366280f03e27SEric Dumazet #endif 366358d607d3SEric Dumazet skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4); 3664331fca43SMartin KaFai Lau /* bpf program will be interested in the tcp_flags */ 3665331fca43SMartin KaFai Lau TCP_SKB_CB(skb)->tcp_flags = TCPHDR_SYN | TCPHDR_ACK; 366660e2a778SUrsula Braun tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5, 3667331fca43SMartin KaFai Lau foc, synack_type, 3668331fca43SMartin KaFai Lau syn_skb) + sizeof(*th); 366933ad798cSAdam Langley 3670aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 3671aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 36721da177e4SLinus Torvalds 3673ea1627c2SEric Dumazet th = (struct tcphdr *)skb->data; 36741da177e4SLinus Torvalds memset(th, 0, sizeof(struct tcphdr)); 36751da177e4SLinus Torvalds th->syn = 1; 36761da177e4SLinus Torvalds th->ack = 1; 36776ac705b1SEric Dumazet tcp_ecn_make_synack(req, th); 3678b44084c2SEric Dumazet th->source = htons(ireq->ir_num); 3679634fb979SEric Dumazet th->dest = ireq->ir_rmt_port; 3680e05a90ecSJamal Hadi Salim skb->mark = ireq->ir_mark; 36813b117750SEric Dumazet skb->ip_summed = CHECKSUM_PARTIAL; 36823b117750SEric Dumazet th->seq = htonl(tcp_rsk(req)->snt_isn); 36838336886fSJerry Chu /* XXX data is queued and acked as is. No buffer/window check */ 36848336886fSJerry Chu th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt); 36851da177e4SLinus Torvalds 36861da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 3687ed53d0abSEric Dumazet th->window = htons(min(req->rsk_rcv_wnd, 65535U)); 3688ea66758cSPaolo Abeni tcp_options_write(th, NULL, &opts); 36891da177e4SLinus Torvalds th->doff = (tcp_header_size >> 2); 3690bced3f7dSBreno Leitao TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS); 3691cfb6eeb4SYOSHIFUJI Hideaki 3692cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 3693cfb6eeb4SYOSHIFUJI Hideaki /* Okay, we have all we need - do the md5 hash if needed */ 369480f03e27SEric Dumazet if (md5) 3695bd0388aeSWilliam Allen Simpson tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, 369639f8e58eSEric Dumazet md5, req_to_sk(req), skb); 369780f03e27SEric Dumazet rcu_read_unlock(); 3698cfb6eeb4SYOSHIFUJI Hideaki #endif 3699cfb6eeb4SYOSHIFUJI Hideaki 3700331fca43SMartin KaFai Lau bpf_skops_write_hdr_opt((struct sock *)sk, skb, req, syn_skb, 3701331fca43SMartin KaFai Lau synack_type, &opts); 3702331fca43SMartin KaFai Lau 3703a1ac9c8aSMartin KaFai Lau skb_set_delivery_time(skb, now, true); 3704a842fe14SEric Dumazet tcp_add_tx_delay(skb, tp); 3705a842fe14SEric Dumazet 37061da177e4SLinus Torvalds return skb; 37071da177e4SLinus Torvalds } 37084bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_make_synack); 37091da177e4SLinus Torvalds 371081164413SDaniel Borkmann static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst) 371181164413SDaniel Borkmann { 371281164413SDaniel Borkmann struct inet_connection_sock *icsk = inet_csk(sk); 371381164413SDaniel Borkmann const struct tcp_congestion_ops *ca; 371481164413SDaniel Borkmann u32 ca_key = dst_metric(dst, RTAX_CC_ALGO); 371581164413SDaniel Borkmann 371681164413SDaniel Borkmann if (ca_key == TCP_CA_UNSPEC) 371781164413SDaniel Borkmann return; 371881164413SDaniel Borkmann 371981164413SDaniel Borkmann rcu_read_lock(); 372081164413SDaniel Borkmann ca = tcp_ca_find_key(ca_key); 37210baf26b0SMartin KaFai Lau if (likely(ca && bpf_try_module_get(ca, ca->owner))) { 37220baf26b0SMartin KaFai Lau bpf_module_put(icsk->icsk_ca_ops, icsk->icsk_ca_ops->owner); 372381164413SDaniel Borkmann icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst); 372481164413SDaniel Borkmann icsk->icsk_ca_ops = ca; 372581164413SDaniel Borkmann } 372681164413SDaniel Borkmann rcu_read_unlock(); 372781164413SDaniel Borkmann } 372881164413SDaniel Borkmann 372967edfef7SAndi Kleen /* Do all connect socket setups that can be done AF independent. */ 3730f7e56a76Sstephen hemminger static void tcp_connect_init(struct sock *sk) 37311da177e4SLinus Torvalds { 3732cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 37331da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 37341da177e4SLinus Torvalds __u8 rcv_wscale; 373513d3b1ebSLawrence Brakmo u32 rcv_wnd; 37361da177e4SLinus Torvalds 37371da177e4SLinus Torvalds /* We'll fix this up when we get a response from the other end. 37381da177e4SLinus Torvalds * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 37391da177e4SLinus Torvalds */ 37405d2ed052SEric Dumazet tp->tcp_header_len = sizeof(struct tcphdr); 37413666f666SKuniyuki Iwashima if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_timestamps)) 37425d2ed052SEric Dumazet tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED; 37431da177e4SLinus Torvalds 3744cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 374500db4124SIan Morris if (tp->af_specific->md5_lookup(sk, sk)) 3746cfb6eeb4SYOSHIFUJI Hideaki tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 3747cfb6eeb4SYOSHIFUJI Hideaki #endif 3748cfb6eeb4SYOSHIFUJI Hideaki 37491da177e4SLinus Torvalds /* If user gave his TCP_MAXSEG, record it to clamp */ 37501da177e4SLinus Torvalds if (tp->rx_opt.user_mss) 37511da177e4SLinus Torvalds tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 37521da177e4SLinus Torvalds tp->max_window = 0; 37535d424d5aSJohn Heffner tcp_mtup_init(sk); 37541da177e4SLinus Torvalds tcp_sync_mss(sk, dst_mtu(dst)); 37551da177e4SLinus Torvalds 375681164413SDaniel Borkmann tcp_ca_dst_init(sk, dst); 375781164413SDaniel Borkmann 37581da177e4SLinus Torvalds if (!tp->window_clamp) 37591da177e4SLinus Torvalds tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 37603541f9e8SEric Dumazet tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 3761f5fff5dcSTom Quetchenbach 37621da177e4SLinus Torvalds tcp_initialize_rcv_mss(sk); 37631da177e4SLinus Torvalds 3764e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 3765e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 3766e88c64f0SHagen Paul Pfeifer (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) 3767e88c64f0SHagen Paul Pfeifer tp->window_clamp = tcp_full_space(sk); 3768e88c64f0SHagen Paul Pfeifer 376913d3b1ebSLawrence Brakmo rcv_wnd = tcp_rwnd_init_bpf(sk); 377013d3b1ebSLawrence Brakmo if (rcv_wnd == 0) 377113d3b1ebSLawrence Brakmo rcv_wnd = dst_metric(dst, RTAX_INITRWND); 377213d3b1ebSLawrence Brakmo 3773ceef9ab6SEric Dumazet tcp_select_initial_window(sk, tcp_full_space(sk), 37741da177e4SLinus Torvalds tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 37751da177e4SLinus Torvalds &tp->rcv_wnd, 37761da177e4SLinus Torvalds &tp->window_clamp, 37773666f666SKuniyuki Iwashima READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_window_scaling), 377831d12926Slaurent chavey &rcv_wscale, 377913d3b1ebSLawrence Brakmo rcv_wnd); 37801da177e4SLinus Torvalds 37811da177e4SLinus Torvalds tp->rx_opt.rcv_wscale = rcv_wscale; 37821da177e4SLinus Torvalds tp->rcv_ssthresh = tp->rcv_wnd; 37831da177e4SLinus Torvalds 3784e13ec3daSEric Dumazet WRITE_ONCE(sk->sk_err, 0); 37851da177e4SLinus Torvalds sock_reset_flag(sk, SOCK_DONE); 37861da177e4SLinus Torvalds tp->snd_wnd = 0; 3787ee7537b6SHantzis Fotis tcp_init_wl(tp, 0); 37887f582b24SEric Dumazet tcp_write_queue_purge(sk); 37891da177e4SLinus Torvalds tp->snd_una = tp->write_seq; 37901da177e4SLinus Torvalds tp->snd_sml = tp->write_seq; 379133f5f57eSIlpo Järvinen tp->snd_up = tp->write_seq; 3792e0d694d6SEric Dumazet WRITE_ONCE(tp->snd_nxt, tp->write_seq); 3793ee995283SPavel Emelyanov 3794ee995283SPavel Emelyanov if (likely(!tp->repair)) 37951da177e4SLinus Torvalds tp->rcv_nxt = 0; 3796c7781a6eSAndrew Vagin else 379770eabf0eSEric Dumazet tp->rcv_tstamp = tcp_jiffies32; 3798ee995283SPavel Emelyanov tp->rcv_wup = tp->rcv_nxt; 37997db48e98SEric Dumazet WRITE_ONCE(tp->copied_seq, tp->rcv_nxt); 38001da177e4SLinus Torvalds 38018550f328SLawrence Brakmo inet_csk(sk)->icsk_rto = tcp_timeout_init(sk); 3802463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_retransmits = 0; 38031da177e4SLinus Torvalds tcp_clear_retrans(tp); 38041da177e4SLinus Torvalds } 38051da177e4SLinus Torvalds 3806783237e8SYuchung Cheng static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) 3807783237e8SYuchung Cheng { 3808783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3809783237e8SYuchung Cheng struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 3810783237e8SYuchung Cheng 3811783237e8SYuchung Cheng tcb->end_seq += skb->len; 3812f4a775d1SEric Dumazet __skb_header_release(skb); 3813ab4e846aSEric Dumazet sk_wmem_queued_add(sk, skb->truesize); 3814783237e8SYuchung Cheng sk_mem_charge(sk, skb->truesize); 38150f317464SEric Dumazet WRITE_ONCE(tp->write_seq, tcb->end_seq); 3816783237e8SYuchung Cheng tp->packets_out += tcp_skb_pcount(skb); 3817783237e8SYuchung Cheng } 3818783237e8SYuchung Cheng 3819783237e8SYuchung Cheng /* Build and send a SYN with data and (cached) Fast Open cookie. However, 3820783237e8SYuchung Cheng * queue a data-only packet after the regular SYN, such that regular SYNs 3821783237e8SYuchung Cheng * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges 3822783237e8SYuchung Cheng * only the SYN sequence, the data are retransmitted in the first ACK. 3823783237e8SYuchung Cheng * If cookie is not cached or other error occurs, falls back to send a 3824783237e8SYuchung Cheng * regular SYN with Fast Open cookie request option. 3825783237e8SYuchung Cheng */ 3826783237e8SYuchung Cheng static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) 3827783237e8SYuchung Cheng { 3828ed0c99dcSJakub Kicinski struct inet_connection_sock *icsk = inet_csk(sk); 3829783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3830783237e8SYuchung Cheng struct tcp_fastopen_request *fo = tp->fastopen_req; 3831fbf93406SEric Dumazet struct page_frag *pfrag = sk_page_frag(sk); 3832355a901eSEric Dumazet struct sk_buff *syn_data; 3833fbf93406SEric Dumazet int space, err = 0; 3834783237e8SYuchung Cheng 383567da22d2SYuchung Cheng tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ 3836065263f4SWei Wang if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie)) 3837783237e8SYuchung Cheng goto fallback; 3838783237e8SYuchung Cheng 3839783237e8SYuchung Cheng /* MSS for SYN-data is based on cached MSS and bounded by PMTU and 3840783237e8SYuchung Cheng * user-MSS. Reserve maximum option space for middleboxes that add 3841783237e8SYuchung Cheng * private TCP options. The cost is reduced data space in SYN :( 3842783237e8SYuchung Cheng */ 38433541f9e8SEric Dumazet tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp); 3844ed0c99dcSJakub Kicinski /* Sync mss_cache after updating the mss_clamp */ 3845ed0c99dcSJakub Kicinski tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 38463541f9e8SEric Dumazet 3847ed0c99dcSJakub Kicinski space = __tcp_mtu_to_mss(sk, icsk->icsk_pmtu_cookie) - 3848783237e8SYuchung Cheng MAX_TCP_OPTION_SPACE; 3849783237e8SYuchung Cheng 3850f5ddcbbbSEric Dumazet space = min_t(size_t, space, fo->size); 3851f5ddcbbbSEric Dumazet 3852fbf93406SEric Dumazet if (space && 3853fbf93406SEric Dumazet !skb_page_frag_refill(min_t(size_t, space, PAGE_SIZE), 3854fbf93406SEric Dumazet pfrag, sk->sk_allocation)) 3855fbf93406SEric Dumazet goto fallback; 38565882efffSEric Dumazet syn_data = tcp_stream_alloc_skb(sk, sk->sk_allocation, false); 3857355a901eSEric Dumazet if (!syn_data) 3858783237e8SYuchung Cheng goto fallback; 3859355a901eSEric Dumazet memcpy(syn_data->cb, syn->cb, sizeof(syn->cb)); 386007e100f9SEric Dumazet if (space) { 3861fbf93406SEric Dumazet space = min_t(size_t, space, pfrag->size - pfrag->offset); 3862fbf93406SEric Dumazet space = tcp_wmem_schedule(sk, space); 3863fbf93406SEric Dumazet } 3864fbf93406SEric Dumazet if (space) { 3865fbf93406SEric Dumazet space = copy_page_from_iter(pfrag->page, pfrag->offset, 3866fbf93406SEric Dumazet space, &fo->data->msg_iter); 3867fbf93406SEric Dumazet if (unlikely(!space)) { 3868ba233b34SEric Dumazet tcp_skb_tsorted_anchor_cleanup(syn_data); 3869355a901eSEric Dumazet kfree_skb(syn_data); 3870783237e8SYuchung Cheng goto fallback; 3871783237e8SYuchung Cheng } 3872fbf93406SEric Dumazet skb_fill_page_desc(syn_data, 0, pfrag->page, 3873fbf93406SEric Dumazet pfrag->offset, space); 3874fbf93406SEric Dumazet page_ref_inc(pfrag->page); 3875fbf93406SEric Dumazet pfrag->offset += space; 3876fbf93406SEric Dumazet skb_len_add(syn_data, space); 3877f859a448SWillem de Bruijn skb_zcopy_set(syn_data, fo->uarg, NULL); 387807e100f9SEric Dumazet } 3879355a901eSEric Dumazet /* No more data pending in inet_wait_for_connect() */ 3880355a901eSEric Dumazet if (space == fo->size) 3881355a901eSEric Dumazet fo->data = NULL; 3882355a901eSEric Dumazet fo->copied = space; 3883783237e8SYuchung Cheng 3884355a901eSEric Dumazet tcp_connect_queue_skb(sk, syn_data); 38850f87230dSFrancis Yan if (syn_data->len) 38860f87230dSFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_BUSY); 3887355a901eSEric Dumazet 3888355a901eSEric Dumazet err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation); 3889355a901eSEric Dumazet 3890a1ac9c8aSMartin KaFai Lau skb_set_delivery_time(syn, syn_data->skb_mstamp_ns, true); 3891355a901eSEric Dumazet 3892355a901eSEric Dumazet /* Now full SYN+DATA was cloned and sent (or not), 3893355a901eSEric Dumazet * remove the SYN from the original skb (syn_data) 3894355a901eSEric Dumazet * we keep in write queue in case of a retransmit, as we 3895355a901eSEric Dumazet * also have the SYN packet (with no data) in the same queue. 3896431a9124SEric Dumazet */ 3897355a901eSEric Dumazet TCP_SKB_CB(syn_data)->seq++; 3898355a901eSEric Dumazet TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH; 3899355a901eSEric Dumazet if (!err) { 390067da22d2SYuchung Cheng tp->syn_data = (fo->copied > 0); 390175c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data); 3902f19c29e3SYuchung Cheng NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT); 3903783237e8SYuchung Cheng goto done; 3904783237e8SYuchung Cheng } 3905783237e8SYuchung Cheng 390675c119afSEric Dumazet /* data was not sent, put it in write_queue */ 390775c119afSEric Dumazet __skb_queue_tail(&sk->sk_write_queue, syn_data); 3908b5b7db8dSEric Dumazet tp->packets_out -= tcp_skb_pcount(syn_data); 3909b5b7db8dSEric Dumazet 3910783237e8SYuchung Cheng fallback: 3911783237e8SYuchung Cheng /* Send a regular SYN with Fast Open cookie request option */ 3912783237e8SYuchung Cheng if (fo->cookie.len > 0) 3913783237e8SYuchung Cheng fo->cookie.len = 0; 3914783237e8SYuchung Cheng err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); 3915783237e8SYuchung Cheng if (err) 3916783237e8SYuchung Cheng tp->syn_fastopen = 0; 3917783237e8SYuchung Cheng done: 3918783237e8SYuchung Cheng fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ 3919783237e8SYuchung Cheng return err; 3920783237e8SYuchung Cheng } 3921783237e8SYuchung Cheng 392267edfef7SAndi Kleen /* Build a SYN and send it off. */ 39231da177e4SLinus Torvalds int tcp_connect(struct sock *sk) 39241da177e4SLinus Torvalds { 39251da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 39261da177e4SLinus Torvalds struct sk_buff *buff; 3927ee586811SEric Paris int err; 39281da177e4SLinus Torvalds 3929de525be2SLawrence Brakmo tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB, 0, NULL); 39308ba60924SEric Dumazet 39318ba60924SEric Dumazet if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 39328ba60924SEric Dumazet return -EHOSTUNREACH; /* Routing failure or similar. */ 39338ba60924SEric Dumazet 39341da177e4SLinus Torvalds tcp_connect_init(sk); 39351da177e4SLinus Torvalds 39362b916477SAndrey Vagin if (unlikely(tp->repair)) { 39372b916477SAndrey Vagin tcp_finish_connect(sk, NULL); 39382b916477SAndrey Vagin return 0; 39392b916477SAndrey Vagin } 39402b916477SAndrey Vagin 39415882efffSEric Dumazet buff = tcp_stream_alloc_skb(sk, sk->sk_allocation, true); 3942355a901eSEric Dumazet if (unlikely(!buff)) 39431da177e4SLinus Torvalds return -ENOBUFS; 39441da177e4SLinus Torvalds 3945a3433f35SChangli Gao tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); 39469a568de4SEric Dumazet tcp_mstamp_refresh(tp); 39479a568de4SEric Dumazet tp->retrans_stamp = tcp_time_stamp(tp); 3948783237e8SYuchung Cheng tcp_connect_queue_skb(sk, buff); 3949735d3831SFlorian Westphal tcp_ecn_send_syn(sk, buff); 395075c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, buff); 39511da177e4SLinus Torvalds 3952783237e8SYuchung Cheng /* Send off SYN; include data in Fast Open. */ 3953783237e8SYuchung Cheng err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : 3954783237e8SYuchung Cheng tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); 3955ee586811SEric Paris if (err == -ECONNREFUSED) 3956ee586811SEric Paris return err; 3957bd37a088SWei Yongjun 3958bd37a088SWei Yongjun /* We change tp->snd_nxt after the tcp_transmit_skb() call 3959bd37a088SWei Yongjun * in order to make this packet get counted in tcpOutSegs. 3960bd37a088SWei Yongjun */ 3961e0d694d6SEric Dumazet WRITE_ONCE(tp->snd_nxt, tp->write_seq); 3962bd37a088SWei Yongjun tp->pushed_seq = tp->write_seq; 3963b5b7db8dSEric Dumazet buff = tcp_send_head(sk); 3964b5b7db8dSEric Dumazet if (unlikely(buff)) { 3965e0d694d6SEric Dumazet WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(buff)->seq); 3966b5b7db8dSEric Dumazet tp->pushed_seq = TCP_SKB_CB(buff)->seq; 3967b5b7db8dSEric Dumazet } 396881cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); 39691da177e4SLinus Torvalds 39701da177e4SLinus Torvalds /* Timer for repeating the SYN until an answer. */ 39713f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 39723f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 39731da177e4SLinus Torvalds return 0; 39741da177e4SLinus Torvalds } 39754bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_connect); 39761da177e4SLinus Torvalds 39771da177e4SLinus Torvalds /* Send out a delayed ack, the caller does the policy checking 39781da177e4SLinus Torvalds * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 39791da177e4SLinus Torvalds * for details. 39801da177e4SLinus Torvalds */ 39811da177e4SLinus Torvalds void tcp_send_delayed_ack(struct sock *sk) 39821da177e4SLinus Torvalds { 3983463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 3984463c84b9SArnaldo Carvalho de Melo int ato = icsk->icsk_ack.ato; 39851da177e4SLinus Torvalds unsigned long timeout; 39861da177e4SLinus Torvalds 39871da177e4SLinus Torvalds if (ato > TCP_DELACK_MIN) { 3988463c84b9SArnaldo Carvalho de Melo const struct tcp_sock *tp = tcp_sk(sk); 39891da177e4SLinus Torvalds int max_ato = HZ / 2; 39901da177e4SLinus Torvalds 399131954cd8SWei Wang if (inet_csk_in_pingpong_mode(sk) || 3992056834d9SIlpo Järvinen (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 39931da177e4SLinus Torvalds max_ato = TCP_DELACK_MAX; 39941da177e4SLinus Torvalds 39951da177e4SLinus Torvalds /* Slow path, intersegment interval is "high". */ 39961da177e4SLinus Torvalds 39971da177e4SLinus Torvalds /* If some rtt estimate is known, use it to bound delayed ack. 3998463c84b9SArnaldo Carvalho de Melo * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 39991da177e4SLinus Torvalds * directly. 40001da177e4SLinus Torvalds */ 4001740b0f18SEric Dumazet if (tp->srtt_us) { 4002740b0f18SEric Dumazet int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3), 4003740b0f18SEric Dumazet TCP_DELACK_MIN); 40041da177e4SLinus Torvalds 40051da177e4SLinus Torvalds if (rtt < max_ato) 40061da177e4SLinus Torvalds max_ato = rtt; 40071da177e4SLinus Torvalds } 40081da177e4SLinus Torvalds 40091da177e4SLinus Torvalds ato = min(ato, max_ato); 40101da177e4SLinus Torvalds } 40111da177e4SLinus Torvalds 40122b8ee4f0SMartin KaFai Lau ato = min_t(u32, ato, inet_csk(sk)->icsk_delack_max); 40132b8ee4f0SMartin KaFai Lau 40141da177e4SLinus Torvalds /* Stay within the limit we were given */ 40151da177e4SLinus Torvalds timeout = jiffies + ato; 40161da177e4SLinus Torvalds 40171da177e4SLinus Torvalds /* Use new timeout only if there wasn't a older one earlier. */ 4018463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 4019b6b6d653SEric Dumazet /* If delack timer is about to expire, send ACK now. */ 4020b6b6d653SEric Dumazet if (time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 40211da177e4SLinus Torvalds tcp_send_ack(sk); 40221da177e4SLinus Torvalds return; 40231da177e4SLinus Torvalds } 40241da177e4SLinus Torvalds 4025463c84b9SArnaldo Carvalho de Melo if (!time_before(timeout, icsk->icsk_ack.timeout)) 4026463c84b9SArnaldo Carvalho de Melo timeout = icsk->icsk_ack.timeout; 40271da177e4SLinus Torvalds } 4028463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 4029463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.timeout = timeout; 4030463c84b9SArnaldo Carvalho de Melo sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 40311da177e4SLinus Torvalds } 40321da177e4SLinus Torvalds 40331da177e4SLinus Torvalds /* This routine sends an ack and also updates the window. */ 40342987babbSYuchung Cheng void __tcp_send_ack(struct sock *sk, u32 rcv_nxt) 40351da177e4SLinus Torvalds { 40361da177e4SLinus Torvalds struct sk_buff *buff; 40371da177e4SLinus Torvalds 4038058dc334SIlpo Järvinen /* If we have been reset, we may not send again. */ 4039058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 4040058dc334SIlpo Järvinen return; 4041058dc334SIlpo Järvinen 40421da177e4SLinus Torvalds /* We are not putting this on the write queue, so 40431da177e4SLinus Torvalds * tcp_transmit_skb() will set the ownership to this 40441da177e4SLinus Torvalds * sock. 40451da177e4SLinus Torvalds */ 40467450aaf6SEric Dumazet buff = alloc_skb(MAX_TCP_HEADER, 40477450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 40487450aaf6SEric Dumazet if (unlikely(!buff)) { 4049a37c2134SEric Dumazet struct inet_connection_sock *icsk = inet_csk(sk); 4050a37c2134SEric Dumazet unsigned long delay; 4051a37c2134SEric Dumazet 4052a37c2134SEric Dumazet delay = TCP_DELACK_MAX << icsk->icsk_ack.retry; 4053a37c2134SEric Dumazet if (delay < TCP_RTO_MAX) 4054a37c2134SEric Dumazet icsk->icsk_ack.retry++; 4055463c84b9SArnaldo Carvalho de Melo inet_csk_schedule_ack(sk); 4056a37c2134SEric Dumazet icsk->icsk_ack.ato = TCP_ATO_MIN; 4057a37c2134SEric Dumazet inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, delay, TCP_RTO_MAX); 40581da177e4SLinus Torvalds return; 40591da177e4SLinus Torvalds } 40601da177e4SLinus Torvalds 40611da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 40621da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 4063a3433f35SChangli Gao tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); 40641da177e4SLinus Torvalds 406598781965SEric Dumazet /* We do not want pure acks influencing TCP Small Queues or fq/pacing 406698781965SEric Dumazet * too much. 406798781965SEric Dumazet * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784 406898781965SEric Dumazet */ 406998781965SEric Dumazet skb_set_tcp_pure_ack(buff); 407098781965SEric Dumazet 40711da177e4SLinus Torvalds /* Send it off, this clears delayed acks for us. */ 40722987babbSYuchung Cheng __tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0, rcv_nxt); 40731da177e4SLinus Torvalds } 407427cde44aSYuchung Cheng EXPORT_SYMBOL_GPL(__tcp_send_ack); 40752987babbSYuchung Cheng 40762987babbSYuchung Cheng void tcp_send_ack(struct sock *sk) 40772987babbSYuchung Cheng { 40782987babbSYuchung Cheng __tcp_send_ack(sk, tcp_sk(sk)->rcv_nxt); 40791da177e4SLinus Torvalds } 40801da177e4SLinus Torvalds 40811da177e4SLinus Torvalds /* This routine sends a packet with an out of date sequence 40821da177e4SLinus Torvalds * number. It assumes the other end will try to ack it. 40831da177e4SLinus Torvalds * 40841da177e4SLinus Torvalds * Question: what should we make while urgent mode? 40851da177e4SLinus Torvalds * 4.4BSD forces sending single byte of data. We cannot send 40861da177e4SLinus Torvalds * out of window data, because we have SND.NXT==SND.MAX... 40871da177e4SLinus Torvalds * 40881da177e4SLinus Torvalds * Current solution: to send TWO zero-length segments in urgent mode: 40891da177e4SLinus Torvalds * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 40901da177e4SLinus Torvalds * out-of-date with SND.UNA-1 to probe window. 40911da177e4SLinus Torvalds */ 4092e520af48SEric Dumazet static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib) 40931da177e4SLinus Torvalds { 40941da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 40951da177e4SLinus Torvalds struct sk_buff *skb; 40961da177e4SLinus Torvalds 40971da177e4SLinus Torvalds /* We don't queue it, tcp_transmit_skb() sets ownership. */ 40987450aaf6SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, 40997450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 410051456b29SIan Morris if (!skb) 41011da177e4SLinus Torvalds return -1; 41021da177e4SLinus Torvalds 41031da177e4SLinus Torvalds /* Reserve space for headers and set control bits. */ 41041da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 41051da177e4SLinus Torvalds /* Use a previous sequence. This should cause the other 41061da177e4SLinus Torvalds * end to send an ack. Don't queue or clone SKB, just 41071da177e4SLinus Torvalds * send it. 41081da177e4SLinus Torvalds */ 4109a3433f35SChangli Gao tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); 4110e2e8009fSRenato Westphal NET_INC_STATS(sock_net(sk), mib); 41117450aaf6SEric Dumazet return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0); 41121da177e4SLinus Torvalds } 41131da177e4SLinus Torvalds 4114385e2070SEric Dumazet /* Called from setsockopt( ... TCP_REPAIR ) */ 4115ee995283SPavel Emelyanov void tcp_send_window_probe(struct sock *sk) 4116ee995283SPavel Emelyanov { 4117ee995283SPavel Emelyanov if (sk->sk_state == TCP_ESTABLISHED) { 4118ee995283SPavel Emelyanov tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; 41199a568de4SEric Dumazet tcp_mstamp_refresh(tcp_sk(sk)); 4120e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE); 4121ee995283SPavel Emelyanov } 4122ee995283SPavel Emelyanov } 4123ee995283SPavel Emelyanov 412467edfef7SAndi Kleen /* Initiate keepalive or window probe from timer. */ 4125e520af48SEric Dumazet int tcp_write_wakeup(struct sock *sk, int mib) 41261da177e4SLinus Torvalds { 41271da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 41281da177e4SLinus Torvalds struct sk_buff *skb; 41291da177e4SLinus Torvalds 4130058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 4131058dc334SIlpo Järvinen return -1; 4132058dc334SIlpo Järvinen 413300db4124SIan Morris skb = tcp_send_head(sk); 413400db4124SIan Morris if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { 41351da177e4SLinus Torvalds int err; 41360c54b85fSIlpo Järvinen unsigned int mss = tcp_current_mss(sk); 413790840defSIlpo Järvinen unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 41381da177e4SLinus Torvalds 41391da177e4SLinus Torvalds if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 41401da177e4SLinus Torvalds tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 41411da177e4SLinus Torvalds 41421da177e4SLinus Torvalds /* We are probing the opening of a window 41431da177e4SLinus Torvalds * but the window size is != 0 41441da177e4SLinus Torvalds * must have been a result SWS avoidance ( sender ) 41451da177e4SLinus Torvalds */ 41461da177e4SLinus Torvalds if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 41471da177e4SLinus Torvalds skb->len > mss) { 41481da177e4SLinus Torvalds seg_size = min(seg_size, mss); 41494de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 415075c119afSEric Dumazet if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE, 415175c119afSEric Dumazet skb, seg_size, mss, GFP_ATOMIC)) 41521da177e4SLinus Torvalds return -1; 41531da177e4SLinus Torvalds } else if (!tcp_skb_pcount(skb)) 41545bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss); 41551da177e4SLinus Torvalds 41564de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 4157dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 415866f5fe62SIlpo Järvinen if (!err) 415966f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 41601da177e4SLinus Torvalds return err; 41611da177e4SLinus Torvalds } else { 416233f5f57eSIlpo Järvinen if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) 4163e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 1, mib); 4164e520af48SEric Dumazet return tcp_xmit_probe_skb(sk, 0, mib); 41651da177e4SLinus Torvalds } 41661da177e4SLinus Torvalds } 41671da177e4SLinus Torvalds 41681da177e4SLinus Torvalds /* A window probe timeout has occurred. If window is not closed send 41691da177e4SLinus Torvalds * a partial packet else a zero probe. 41701da177e4SLinus Torvalds */ 41711da177e4SLinus Torvalds void tcp_send_probe0(struct sock *sk) 41721da177e4SLinus Torvalds { 4173463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 41741da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 4175c6214a97SNikolay Borisov struct net *net = sock_net(sk); 4176c1d5674fSYuchung Cheng unsigned long timeout; 41771da177e4SLinus Torvalds int err; 41781da177e4SLinus Torvalds 4179e520af48SEric Dumazet err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE); 41801da177e4SLinus Torvalds 418175c119afSEric Dumazet if (tp->packets_out || tcp_write_queue_empty(sk)) { 41821da177e4SLinus Torvalds /* Cancel probe timer, if it is not required. */ 41836687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 0; 4184463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff = 0; 41859d9b1ee0SEnke Chen icsk->icsk_probes_tstamp = 0; 41861da177e4SLinus Torvalds return; 41871da177e4SLinus Torvalds } 41881da177e4SLinus Torvalds 4189c1d5674fSYuchung Cheng icsk->icsk_probes_out++; 41901da177e4SLinus Torvalds if (err <= 0) { 419139e24435SKuniyuki Iwashima if (icsk->icsk_backoff < READ_ONCE(net->ipv4.sysctl_tcp_retries2)) 4192463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff++; 4193c1d5674fSYuchung Cheng timeout = tcp_probe0_when(sk, TCP_RTO_MAX); 41941da177e4SLinus Torvalds } else { 41951da177e4SLinus Torvalds /* If packet was not sent due to local congestion, 4196c1d5674fSYuchung Cheng * Let senders fight for local resources conservatively. 41971da177e4SLinus Torvalds */ 4198c1d5674fSYuchung Cheng timeout = TCP_RESOURCE_PROBE_INTERVAL; 41991da177e4SLinus Torvalds } 4200344db93aSEnke Chen 4201344db93aSEnke Chen timeout = tcp_clamp_probe0_to_user_timeout(sk, timeout); 42028dc242adSEric Dumazet tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, timeout, TCP_RTO_MAX); 42031da177e4SLinus Torvalds } 42045db92c99SOctavian Purdila 4205ea3bea3aSEric Dumazet int tcp_rtx_synack(const struct sock *sk, struct request_sock *req) 42065db92c99SOctavian Purdila { 42075db92c99SOctavian Purdila const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific; 42085db92c99SOctavian Purdila struct flowi fl; 42095db92c99SOctavian Purdila int res; 42105db92c99SOctavian Purdila 4211cb6cd2ceSAkhmat Karakotov /* Paired with WRITE_ONCE() in sock_setsockopt() */ 4212cb6cd2ceSAkhmat Karakotov if (READ_ONCE(sk->sk_txrehash) == SOCK_TXREHASH_ENABLED) 421358d607d3SEric Dumazet tcp_rsk(req)->txhash = net_tx_rndhash(); 4214331fca43SMartin KaFai Lau res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL, 4215331fca43SMartin KaFai Lau NULL); 42165db92c99SOctavian Purdila if (!res) { 42170a375c82SEric Dumazet TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); 42180a375c82SEric Dumazet NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 4219e9d9da91SEric Dumazet if (unlikely(tcp_passive_fastopen(sk))) { 4220e9d9da91SEric Dumazet /* sk has const attribute because listeners are lockless. 4221e9d9da91SEric Dumazet * However in this case, we are dealing with a passive fastopen 4222e9d9da91SEric Dumazet * socket thus we can change total_retrans value. 4223e9d9da91SEric Dumazet */ 4224e9d9da91SEric Dumazet tcp_sk_rw(sk)->total_retrans++; 4225e9d9da91SEric Dumazet } 4226cf34ce3dSSong Liu trace_tcp_retransmit_synack(sk, req); 42275db92c99SOctavian Purdila } 42285db92c99SOctavian Purdila return res; 42295db92c99SOctavian Purdila } 42305db92c99SOctavian Purdila EXPORT_SYMBOL(tcp_rtx_synack); 4231