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. */ 180059217c1SNeal Cardwell static inline void tcp_event_ack_sent(struct sock *sk, u32 rcv_nxt) 1811da177e4SLinus Torvalds { 1825d9f4262SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 1835d9f4262SEric Dumazet 1842b195850SEric Dumazet if (unlikely(tp->compressed_ack)) { 185200d95f4SEric Dumazet NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED, 1862b195850SEric Dumazet tp->compressed_ack); 1872b195850SEric Dumazet tp->compressed_ack = 0; 1885d9f4262SEric Dumazet if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1) 1895d9f4262SEric Dumazet __sock_put(sk); 1905d9f4262SEric Dumazet } 19127cde44aSYuchung Cheng 19227cde44aSYuchung Cheng if (unlikely(rcv_nxt != tp->rcv_nxt)) 19327cde44aSYuchung Cheng return; /* Special ACK sent by DCTCP to reflect ECN */ 194059217c1SNeal Cardwell tcp_dec_quickack_mode(sk); 195463c84b9SArnaldo Carvalho de Melo inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); 1961da177e4SLinus Torvalds } 1971da177e4SLinus Torvalds 1981da177e4SLinus Torvalds /* Determine a window scaling and initial window to offer. 1991da177e4SLinus Torvalds * Based on the assumption that the given amount of space 2001da177e4SLinus Torvalds * will be offered. Store the results in the tp structure. 2011da177e4SLinus Torvalds * NOTE: for smooth operation initial space offering should 2021da177e4SLinus Torvalds * be a multiple of mss if possible. We assume here that mss >= 1. 2031da177e4SLinus Torvalds * This MUST be enforced by all callers. 2041da177e4SLinus Torvalds */ 205ceef9ab6SEric Dumazet void tcp_select_initial_window(const struct sock *sk, int __space, __u32 mss, 2061da177e4SLinus Torvalds __u32 *rcv_wnd, __u32 *window_clamp, 20731d12926Slaurent chavey int wscale_ok, __u8 *rcv_wscale, 20831d12926Slaurent chavey __u32 init_rcv_wnd) 2091da177e4SLinus Torvalds { 2101da177e4SLinus Torvalds unsigned int space = (__space < 0 ? 0 : __space); 2111da177e4SLinus Torvalds 2121da177e4SLinus Torvalds /* If no clamp set the clamp to the max possible scaled window */ 2131da177e4SLinus Torvalds if (*window_clamp == 0) 214589c49cbSGao Feng (*window_clamp) = (U16_MAX << TCP_MAX_WSCALE); 2151da177e4SLinus Torvalds space = min(*window_clamp, space); 2161da177e4SLinus Torvalds 2171da177e4SLinus Torvalds /* Quantize space offering to a multiple of mss if possible. */ 2181da177e4SLinus Torvalds if (space > mss) 219589c49cbSGao Feng space = rounddown(space, mss); 2201da177e4SLinus Torvalds 2211da177e4SLinus Torvalds /* NOTE: offering an initial window larger than 32767 22215d99e02SRick Jones * will break some buggy TCP stacks. If the admin tells us 22315d99e02SRick Jones * it is likely we could be speaking with such a buggy stack 22415d99e02SRick Jones * we will truncate our initial window offering to 32K-1 22515d99e02SRick Jones * unless the remote has sent us a window scaling option, 22615d99e02SRick Jones * which we interpret as a sign the remote TCP is not 22715d99e02SRick Jones * misinterpreting the window field as a signed quantity. 2281da177e4SLinus Torvalds */ 2290f1e4d06SKuniyuki Iwashima if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows)) 2301da177e4SLinus Torvalds (*rcv_wnd) = min(space, MAX_TCP_WINDOW); 23115d99e02SRick Jones else 232a337531bSYuchung Cheng (*rcv_wnd) = min_t(u32, space, U16_MAX); 233a337531bSYuchung Cheng 234a337531bSYuchung Cheng if (init_rcv_wnd) 235a337531bSYuchung Cheng *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss); 23615d99e02SRick Jones 23719bf6261SEric Dumazet *rcv_wscale = 0; 2381da177e4SLinus Torvalds if (wscale_ok) { 239589c49cbSGao Feng /* Set window scaling on max possible window */ 24002739545SKuniyuki Iwashima space = max_t(u32, space, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2])); 2411227c177SKuniyuki Iwashima space = max_t(u32, space, READ_ONCE(sysctl_rmem_max)); 242316c1592SStephen Hemminger space = min_t(u32, space, *window_clamp); 24319bf6261SEric Dumazet *rcv_wscale = clamp_t(int, ilog2(space) - 15, 24419bf6261SEric Dumazet 0, TCP_MAX_WSCALE); 2451da177e4SLinus Torvalds } 2461da177e4SLinus Torvalds /* Set the clamp no higher than max representable value */ 247589c49cbSGao Feng (*window_clamp) = min_t(__u32, U16_MAX << (*rcv_wscale), *window_clamp); 2481da177e4SLinus Torvalds } 2494bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_select_initial_window); 2501da177e4SLinus Torvalds 2511da177e4SLinus Torvalds /* Chose a new window to advertise, update state in tcp_sock for the 2521da177e4SLinus Torvalds * socket, and return result with RFC1323 scaling applied. The return 2531da177e4SLinus Torvalds * value can be stuffed directly into th->window for an outgoing 2541da177e4SLinus Torvalds * frame. 2551da177e4SLinus Torvalds */ 25640efc6faSStephen Hemminger static u16 tcp_select_window(struct sock *sk) 2571da177e4SLinus Torvalds { 2581da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 259b650d953Smfreemon@cloudflare.com struct net *net = sock_net(sk); 260e2142825SMenglong Dong u32 old_win = tp->rcv_wnd; 261e2142825SMenglong Dong u32 cur_win, new_win; 2621da177e4SLinus Torvalds 263e2142825SMenglong Dong /* Make the window 0 if we failed to queue the data because we 264e2142825SMenglong Dong * are out of memory. The window is temporary, so we don't store 265e2142825SMenglong Dong * it on the socket. 266e2142825SMenglong Dong */ 267e2142825SMenglong Dong if (unlikely(inet_csk(sk)->icsk_ack.pending & ICSK_ACK_NOMEM)) 268e2142825SMenglong Dong return 0; 269e2142825SMenglong Dong 270e2142825SMenglong Dong cur_win = tcp_receive_window(tp); 271e2142825SMenglong Dong new_win = __tcp_select_window(sk); 2721da177e4SLinus Torvalds if (new_win < cur_win) { 2731da177e4SLinus Torvalds /* Danger Will Robinson! 2741da177e4SLinus Torvalds * Don't update rcv_wup/rcv_wnd here or else 2751da177e4SLinus Torvalds * we will not be able to advertise a zero 2761da177e4SLinus Torvalds * window in time. --DaveM 2771da177e4SLinus Torvalds * 2781da177e4SLinus Torvalds * Relax Will Robinson. 2791da177e4SLinus Torvalds */ 280b650d953Smfreemon@cloudflare.com if (!READ_ONCE(net->ipv4.sysctl_tcp_shrink_window) || !tp->rx_opt.rcv_wscale) { 281b650d953Smfreemon@cloudflare.com /* Never shrink the offered window */ 2828e165e20SFlorian Westphal if (new_win == 0) 283b650d953Smfreemon@cloudflare.com NET_INC_STATS(net, LINUX_MIB_TCPWANTZEROWINDOWADV); 284607bfbf2SPatrick McHardy new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale); 2851da177e4SLinus Torvalds } 286b650d953Smfreemon@cloudflare.com } 287b650d953Smfreemon@cloudflare.com 2881da177e4SLinus Torvalds tp->rcv_wnd = new_win; 2891da177e4SLinus Torvalds tp->rcv_wup = tp->rcv_nxt; 2901da177e4SLinus Torvalds 2911da177e4SLinus Torvalds /* Make sure we do not exceed the maximum possible 2921da177e4SLinus Torvalds * scaled window. 2931da177e4SLinus Torvalds */ 294ceef9ab6SEric Dumazet if (!tp->rx_opt.rcv_wscale && 295b650d953Smfreemon@cloudflare.com READ_ONCE(net->ipv4.sysctl_tcp_workaround_signed_windows)) 2961da177e4SLinus Torvalds new_win = min(new_win, MAX_TCP_WINDOW); 2971da177e4SLinus Torvalds else 2981da177e4SLinus Torvalds new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); 2991da177e4SLinus Torvalds 3001da177e4SLinus Torvalds /* RFC1323 scaling applied */ 3011da177e4SLinus Torvalds new_win >>= tp->rx_opt.rcv_wscale; 3021da177e4SLinus Torvalds 30331770e34SFlorian Westphal /* If we advertise zero window, disable fast path. */ 3048e165e20SFlorian Westphal if (new_win == 0) { 30531770e34SFlorian Westphal tp->pred_flags = 0; 3068e165e20SFlorian Westphal if (old_win) 307b650d953Smfreemon@cloudflare.com NET_INC_STATS(net, LINUX_MIB_TCPTOZEROWINDOWADV); 3088e165e20SFlorian Westphal } else if (old_win == 0) { 309b650d953Smfreemon@cloudflare.com NET_INC_STATS(net, LINUX_MIB_TCPFROMZEROWINDOWADV); 3108e165e20SFlorian Westphal } 3111da177e4SLinus Torvalds 3121da177e4SLinus Torvalds return new_win; 3131da177e4SLinus Torvalds } 3141da177e4SLinus Torvalds 31567edfef7SAndi Kleen /* Packet ECN state for a SYN-ACK */ 316735d3831SFlorian Westphal static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb) 317bdf1ee5dSIlpo Järvinen { 31830e502a3SDaniel Borkmann const struct tcp_sock *tp = tcp_sk(sk); 31930e502a3SDaniel Borkmann 3204de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR; 321bdf1ee5dSIlpo Järvinen if (!(tp->ecn_flags & TCP_ECN_OK)) 3224de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE; 32391b5b21cSLawrence Brakmo else if (tcp_ca_needs_ecn(sk) || 32491b5b21cSLawrence Brakmo tcp_bpf_ca_needs_ecn(sk)) 32530e502a3SDaniel Borkmann INET_ECN_xmit(sk); 326bdf1ee5dSIlpo Järvinen } 327bdf1ee5dSIlpo Järvinen 32867edfef7SAndi Kleen /* Packet ECN state for a SYN. */ 329735d3831SFlorian Westphal static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb) 330bdf1ee5dSIlpo Järvinen { 331bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 33291b5b21cSLawrence Brakmo bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk); 3334785a667SKuniyuki Iwashima bool use_ecn = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_ecn) == 1 || 33491b5b21cSLawrence Brakmo tcp_ca_needs_ecn(sk) || bpf_needs_ecn; 335f7b3bec6SFlorian Westphal 336f7b3bec6SFlorian Westphal if (!use_ecn) { 337f7b3bec6SFlorian Westphal const struct dst_entry *dst = __sk_dst_get(sk); 338f7b3bec6SFlorian Westphal 339f7b3bec6SFlorian Westphal if (dst && dst_feature(dst, RTAX_FEATURE_ECN)) 340f7b3bec6SFlorian Westphal use_ecn = true; 341f7b3bec6SFlorian Westphal } 342bdf1ee5dSIlpo Järvinen 343bdf1ee5dSIlpo Järvinen tp->ecn_flags = 0; 344f7b3bec6SFlorian Westphal 345f7b3bec6SFlorian Westphal if (use_ecn) { 3464de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR; 347bdf1ee5dSIlpo Järvinen tp->ecn_flags = TCP_ECN_OK; 34891b5b21cSLawrence Brakmo if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn) 34930e502a3SDaniel Borkmann INET_ECN_xmit(sk); 350bdf1ee5dSIlpo Järvinen } 351bdf1ee5dSIlpo Järvinen } 352bdf1ee5dSIlpo Järvinen 35349213555SDaniel Borkmann static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb) 35449213555SDaniel Borkmann { 35512b8d9caSKuniyuki Iwashima if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback)) 35649213555SDaniel Borkmann /* tp->ecn_flags are cleared at a later point in time when 35749213555SDaniel Borkmann * SYN ACK is ultimatively being received. 35849213555SDaniel Borkmann */ 35949213555SDaniel Borkmann TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR); 36049213555SDaniel Borkmann } 36149213555SDaniel Borkmann 362735d3831SFlorian Westphal static void 3636ac705b1SEric Dumazet tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th) 364bdf1ee5dSIlpo Järvinen { 3656ac705b1SEric Dumazet if (inet_rsk(req)->ecn_ok) 366bdf1ee5dSIlpo Järvinen th->ece = 1; 367bdf1ee5dSIlpo Järvinen } 368bdf1ee5dSIlpo Järvinen 36967edfef7SAndi Kleen /* Set up ECN state for a packet on a ESTABLISHED socket that is about to 37067edfef7SAndi Kleen * be sent. 37167edfef7SAndi Kleen */ 372735d3831SFlorian Westphal static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb, 373ea1627c2SEric Dumazet struct tcphdr *th, int tcp_header_len) 374bdf1ee5dSIlpo Järvinen { 375bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 376bdf1ee5dSIlpo Järvinen 377bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_OK) { 378bdf1ee5dSIlpo Järvinen /* Not-retransmitted data segment: set ECT and inject CWR. */ 379bdf1ee5dSIlpo Järvinen if (skb->len != tcp_header_len && 380bdf1ee5dSIlpo Järvinen !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { 381bdf1ee5dSIlpo Järvinen INET_ECN_xmit(sk); 382bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) { 383bdf1ee5dSIlpo Järvinen tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; 384ea1627c2SEric Dumazet th->cwr = 1; 385bdf1ee5dSIlpo Järvinen skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; 386bdf1ee5dSIlpo Järvinen } 38730e502a3SDaniel Borkmann } else if (!tcp_ca_needs_ecn(sk)) { 388bdf1ee5dSIlpo Järvinen /* ACK or retransmitted segment: clear ECT|CE */ 389bdf1ee5dSIlpo Järvinen INET_ECN_dontxmit(sk); 390bdf1ee5dSIlpo Järvinen } 391bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) 392ea1627c2SEric Dumazet th->ece = 1; 393bdf1ee5dSIlpo Järvinen } 394bdf1ee5dSIlpo Järvinen } 395bdf1ee5dSIlpo Järvinen 396e870a8efSIlpo Järvinen /* Constructs common control bits of non-data skb. If SYN/FIN is present, 397e870a8efSIlpo Järvinen * auto increment end seqno. 398e870a8efSIlpo Järvinen */ 399e870a8efSIlpo Järvinen static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags) 400e870a8efSIlpo Järvinen { 4012e8e18efSDavid S. Miller skb->ip_summed = CHECKSUM_PARTIAL; 402e870a8efSIlpo Järvinen 4034de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags; 404e870a8efSIlpo Järvinen 405cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, 1); 406e870a8efSIlpo Järvinen 407e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->seq = seq; 408a3433f35SChangli Gao if (flags & (TCPHDR_SYN | TCPHDR_FIN)) 409e870a8efSIlpo Järvinen seq++; 410e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->end_seq = seq; 411e870a8efSIlpo Järvinen } 412e870a8efSIlpo Järvinen 413a2a385d6SEric Dumazet static inline bool tcp_urg_mode(const struct tcp_sock *tp) 41433f5f57eSIlpo Järvinen { 41533f5f57eSIlpo Järvinen return tp->snd_una != tp->snd_up; 41633f5f57eSIlpo Järvinen } 41733f5f57eSIlpo Järvinen 4183b65abb8SLeonard Crestez #define OPTION_SACK_ADVERTISE BIT(0) 4193b65abb8SLeonard Crestez #define OPTION_TS BIT(1) 4203b65abb8SLeonard Crestez #define OPTION_MD5 BIT(2) 4213b65abb8SLeonard Crestez #define OPTION_WSCALE BIT(3) 4223b65abb8SLeonard Crestez #define OPTION_FAST_OPEN_COOKIE BIT(8) 4233b65abb8SLeonard Crestez #define OPTION_SMC BIT(9) 4243b65abb8SLeonard Crestez #define OPTION_MPTCP BIT(10) 42560e2a778SUrsula Braun 42660e2a778SUrsula Braun static void smc_options_write(__be32 *ptr, u16 *options) 42760e2a778SUrsula Braun { 42860e2a778SUrsula Braun #if IS_ENABLED(CONFIG_SMC) 42960e2a778SUrsula Braun if (static_branch_unlikely(&tcp_have_smc)) { 43060e2a778SUrsula Braun if (unlikely(OPTION_SMC & *options)) { 43160e2a778SUrsula Braun *ptr++ = htonl((TCPOPT_NOP << 24) | 43260e2a778SUrsula Braun (TCPOPT_NOP << 16) | 43360e2a778SUrsula Braun (TCPOPT_EXP << 8) | 43460e2a778SUrsula Braun (TCPOLEN_EXP_SMC_BASE)); 43560e2a778SUrsula Braun *ptr++ = htonl(TCPOPT_SMC_MAGIC); 43660e2a778SUrsula Braun } 43760e2a778SUrsula Braun } 43860e2a778SUrsula Braun #endif 43960e2a778SUrsula Braun } 44033ad798cSAdam Langley 44133ad798cSAdam Langley struct tcp_out_options { 4422100c8d2SYuchung Cheng u16 options; /* bit field of OPTION_* */ 4432100c8d2SYuchung Cheng u16 mss; /* 0 to disable */ 44433ad798cSAdam Langley u8 ws; /* window scale, 0 to disable */ 44533ad798cSAdam Langley u8 num_sack_blocks; /* number of SACK blocks to include */ 446bd0388aeSWilliam Allen Simpson u8 hash_size; /* bytes in hash_location */ 447331fca43SMartin KaFai Lau u8 bpf_opt_len; /* length of BPF hdr option */ 448bd0388aeSWilliam Allen Simpson __u8 *hash_location; /* temporary pointer, overloaded */ 4492100c8d2SYuchung Cheng __u32 tsval, tsecr; /* need to include OPTION_TS */ 4502100c8d2SYuchung Cheng struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */ 451eda7acddSPeter Krystad struct mptcp_out_options mptcp; 45233ad798cSAdam Langley }; 45333ad798cSAdam Langley 454ea66758cSPaolo Abeni static void mptcp_options_write(struct tcphdr *th, __be32 *ptr, 455ea66758cSPaolo Abeni struct tcp_sock *tp, 456fa3fe2b1SFlorian Westphal struct tcp_out_options *opts) 457eda7acddSPeter Krystad { 458eda7acddSPeter Krystad #if IS_ENABLED(CONFIG_MPTCP) 459eda7acddSPeter Krystad if (unlikely(OPTION_MPTCP & opts->options)) 460ea66758cSPaolo Abeni mptcp_write_options(th, ptr, tp, &opts->mptcp); 461eda7acddSPeter Krystad #endif 462eda7acddSPeter Krystad } 463eda7acddSPeter Krystad 464331fca43SMartin KaFai Lau #ifdef CONFIG_CGROUP_BPF 4650813a841SMartin KaFai Lau static int bpf_skops_write_hdr_opt_arg0(struct sk_buff *skb, 4660813a841SMartin KaFai Lau enum tcp_synack_type synack_type) 4670813a841SMartin KaFai Lau { 4680813a841SMartin KaFai Lau if (unlikely(!skb)) 4690813a841SMartin KaFai Lau return BPF_WRITE_HDR_TCP_CURRENT_MSS; 4700813a841SMartin KaFai Lau 4710813a841SMartin KaFai Lau if (unlikely(synack_type == TCP_SYNACK_COOKIE)) 4720813a841SMartin KaFai Lau return BPF_WRITE_HDR_TCP_SYNACK_COOKIE; 4730813a841SMartin KaFai Lau 4740813a841SMartin KaFai Lau return 0; 4750813a841SMartin KaFai Lau } 4760813a841SMartin KaFai Lau 477331fca43SMartin KaFai Lau /* req, syn_skb and synack_type are used when writing synack */ 478331fca43SMartin KaFai Lau static void bpf_skops_hdr_opt_len(struct sock *sk, struct sk_buff *skb, 479331fca43SMartin KaFai Lau struct request_sock *req, 480331fca43SMartin KaFai Lau struct sk_buff *syn_skb, 481331fca43SMartin KaFai Lau enum tcp_synack_type synack_type, 482331fca43SMartin KaFai Lau struct tcp_out_options *opts, 483331fca43SMartin KaFai Lau unsigned int *remaining) 484331fca43SMartin KaFai Lau { 4850813a841SMartin KaFai Lau struct bpf_sock_ops_kern sock_ops; 4860813a841SMartin KaFai Lau int err; 4870813a841SMartin KaFai Lau 488331fca43SMartin KaFai Lau if (likely(!BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), 489331fca43SMartin KaFai Lau BPF_SOCK_OPS_WRITE_HDR_OPT_CB_FLAG)) || 490331fca43SMartin KaFai Lau !*remaining) 491331fca43SMartin KaFai Lau return; 492331fca43SMartin KaFai Lau 4930813a841SMartin KaFai Lau /* *remaining has already been aligned to 4 bytes, so *remaining >= 4 */ 4940813a841SMartin KaFai Lau 4950813a841SMartin KaFai Lau /* init sock_ops */ 4960813a841SMartin KaFai Lau memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp)); 4970813a841SMartin KaFai Lau 4980813a841SMartin KaFai Lau sock_ops.op = BPF_SOCK_OPS_HDR_OPT_LEN_CB; 4990813a841SMartin KaFai Lau 5000813a841SMartin KaFai Lau if (req) { 5010813a841SMartin KaFai Lau /* The listen "sk" cannot be passed here because 5020813a841SMartin KaFai Lau * it is not locked. It would not make too much 5030813a841SMartin KaFai Lau * sense to do bpf_setsockopt(listen_sk) based 5040813a841SMartin KaFai Lau * on individual connection request also. 5050813a841SMartin KaFai Lau * 5060813a841SMartin KaFai Lau * Thus, "req" is passed here and the cgroup-bpf-progs 5070813a841SMartin KaFai Lau * of the listen "sk" will be run. 5080813a841SMartin KaFai Lau * 5090813a841SMartin KaFai Lau * "req" is also used here for fastopen even the "sk" here is 5100813a841SMartin KaFai Lau * a fullsock "child" sk. It is to keep the behavior 5110813a841SMartin KaFai Lau * consistent between fastopen and non-fastopen on 5120813a841SMartin KaFai Lau * the bpf programming side. 513331fca43SMartin KaFai Lau */ 5140813a841SMartin KaFai Lau sock_ops.sk = (struct sock *)req; 5150813a841SMartin KaFai Lau sock_ops.syn_skb = syn_skb; 5160813a841SMartin KaFai Lau } else { 5170813a841SMartin KaFai Lau sock_owned_by_me(sk); 5180813a841SMartin KaFai Lau 5190813a841SMartin KaFai Lau sock_ops.is_fullsock = 1; 5200813a841SMartin KaFai Lau sock_ops.sk = sk; 5210813a841SMartin KaFai Lau } 5220813a841SMartin KaFai Lau 5230813a841SMartin KaFai Lau sock_ops.args[0] = bpf_skops_write_hdr_opt_arg0(skb, synack_type); 5240813a841SMartin KaFai Lau sock_ops.remaining_opt_len = *remaining; 5250813a841SMartin KaFai Lau /* tcp_current_mss() does not pass a skb */ 5260813a841SMartin KaFai Lau if (skb) 5270813a841SMartin KaFai Lau bpf_skops_init_skb(&sock_ops, skb, 0); 5280813a841SMartin KaFai Lau 5290813a841SMartin KaFai Lau err = BPF_CGROUP_RUN_PROG_SOCK_OPS_SK(&sock_ops, sk); 5300813a841SMartin KaFai Lau 5310813a841SMartin KaFai Lau if (err || sock_ops.remaining_opt_len == *remaining) 5320813a841SMartin KaFai Lau return; 5330813a841SMartin KaFai Lau 5340813a841SMartin KaFai Lau opts->bpf_opt_len = *remaining - sock_ops.remaining_opt_len; 5350813a841SMartin KaFai Lau /* round up to 4 bytes */ 5360813a841SMartin KaFai Lau opts->bpf_opt_len = (opts->bpf_opt_len + 3) & ~3; 5370813a841SMartin KaFai Lau 5380813a841SMartin KaFai Lau *remaining -= opts->bpf_opt_len; 539331fca43SMartin KaFai Lau } 540331fca43SMartin KaFai Lau 541331fca43SMartin KaFai Lau static void bpf_skops_write_hdr_opt(struct sock *sk, struct sk_buff *skb, 542331fca43SMartin KaFai Lau struct request_sock *req, 543331fca43SMartin KaFai Lau struct sk_buff *syn_skb, 544331fca43SMartin KaFai Lau enum tcp_synack_type synack_type, 545331fca43SMartin KaFai Lau struct tcp_out_options *opts) 546331fca43SMartin KaFai Lau { 5470813a841SMartin KaFai Lau u8 first_opt_off, nr_written, max_opt_len = opts->bpf_opt_len; 5480813a841SMartin KaFai Lau struct bpf_sock_ops_kern sock_ops; 5490813a841SMartin KaFai Lau int err; 5500813a841SMartin KaFai Lau 5510813a841SMartin KaFai Lau if (likely(!max_opt_len)) 552331fca43SMartin KaFai Lau return; 553331fca43SMartin KaFai Lau 5540813a841SMartin KaFai Lau memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp)); 5550813a841SMartin KaFai Lau 5560813a841SMartin KaFai Lau sock_ops.op = BPF_SOCK_OPS_WRITE_HDR_OPT_CB; 5570813a841SMartin KaFai Lau 5580813a841SMartin KaFai Lau if (req) { 5590813a841SMartin KaFai Lau sock_ops.sk = (struct sock *)req; 5600813a841SMartin KaFai Lau sock_ops.syn_skb = syn_skb; 5610813a841SMartin KaFai Lau } else { 5620813a841SMartin KaFai Lau sock_owned_by_me(sk); 5630813a841SMartin KaFai Lau 5640813a841SMartin KaFai Lau sock_ops.is_fullsock = 1; 5650813a841SMartin KaFai Lau sock_ops.sk = sk; 5660813a841SMartin KaFai Lau } 5670813a841SMartin KaFai Lau 5680813a841SMartin KaFai Lau sock_ops.args[0] = bpf_skops_write_hdr_opt_arg0(skb, synack_type); 5690813a841SMartin KaFai Lau sock_ops.remaining_opt_len = max_opt_len; 5700813a841SMartin KaFai Lau first_opt_off = tcp_hdrlen(skb) - max_opt_len; 5710813a841SMartin KaFai Lau bpf_skops_init_skb(&sock_ops, skb, first_opt_off); 5720813a841SMartin KaFai Lau 5730813a841SMartin KaFai Lau err = BPF_CGROUP_RUN_PROG_SOCK_OPS_SK(&sock_ops, sk); 5740813a841SMartin KaFai Lau 5750813a841SMartin KaFai Lau if (err) 5760813a841SMartin KaFai Lau nr_written = 0; 5770813a841SMartin KaFai Lau else 5780813a841SMartin KaFai Lau nr_written = max_opt_len - sock_ops.remaining_opt_len; 5790813a841SMartin KaFai Lau 5800813a841SMartin KaFai Lau if (nr_written < max_opt_len) 5810813a841SMartin KaFai Lau memset(skb->data + first_opt_off + nr_written, TCPOPT_NOP, 5820813a841SMartin KaFai Lau max_opt_len - nr_written); 583331fca43SMartin KaFai Lau } 584331fca43SMartin KaFai Lau #else 585331fca43SMartin KaFai Lau static void bpf_skops_hdr_opt_len(struct sock *sk, struct sk_buff *skb, 586331fca43SMartin KaFai Lau struct request_sock *req, 587331fca43SMartin KaFai Lau struct sk_buff *syn_skb, 588331fca43SMartin KaFai Lau enum tcp_synack_type synack_type, 589331fca43SMartin KaFai Lau struct tcp_out_options *opts, 590331fca43SMartin KaFai Lau unsigned int *remaining) 591331fca43SMartin KaFai Lau { 592331fca43SMartin KaFai Lau } 593331fca43SMartin KaFai Lau 594331fca43SMartin KaFai Lau static void bpf_skops_write_hdr_opt(struct sock *sk, struct sk_buff *skb, 595331fca43SMartin KaFai Lau struct request_sock *req, 596331fca43SMartin KaFai Lau struct sk_buff *syn_skb, 597331fca43SMartin KaFai Lau enum tcp_synack_type synack_type, 598331fca43SMartin KaFai Lau struct tcp_out_options *opts) 599331fca43SMartin KaFai Lau { 600331fca43SMartin KaFai Lau } 601331fca43SMartin KaFai Lau #endif 602331fca43SMartin KaFai Lau 60367edfef7SAndi Kleen /* Write previously computed TCP options to the packet. 60467edfef7SAndi Kleen * 60567edfef7SAndi Kleen * Beware: Something in the Internet is very sensitive to the ordering of 606fd6149d3SIlpo Järvinen * TCP options, we learned this through the hard way, so be careful here. 607fd6149d3SIlpo Järvinen * Luckily we can at least blame others for their non-compliance but from 6088e3bff96Sstephen hemminger * inter-operability perspective it seems that we're somewhat stuck with 609fd6149d3SIlpo Järvinen * the ordering which we have been using if we want to keep working with 610fd6149d3SIlpo Järvinen * those broken things (not that it currently hurts anybody as there isn't 611fd6149d3SIlpo Järvinen * particular reason why the ordering would need to be changed). 612fd6149d3SIlpo Järvinen * 613fd6149d3SIlpo Järvinen * At least SACK_PERM as the first option is known to lead to a disaster 614fd6149d3SIlpo Järvinen * (but it may well be that other scenarios fail similarly). 615fd6149d3SIlpo Järvinen */ 616ea66758cSPaolo Abeni static void tcp_options_write(struct tcphdr *th, struct tcp_sock *tp, 617bd0388aeSWilliam Allen Simpson struct tcp_out_options *opts) 618bd0388aeSWilliam Allen Simpson { 619ea66758cSPaolo Abeni __be32 *ptr = (__be32 *)(th + 1); 6202100c8d2SYuchung Cheng u16 options = opts->options; /* mungable copy */ 621bd0388aeSWilliam Allen Simpson 622bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_MD5 & options)) { 6231a2c6181SChristoph Paasch *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | 6241a2c6181SChristoph Paasch (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG); 625bd0388aeSWilliam Allen Simpson /* overload cookie hash location */ 626bd0388aeSWilliam Allen Simpson opts->hash_location = (__u8 *)ptr; 62733ad798cSAdam Langley ptr += 4; 62833ad798cSAdam Langley } 62933ad798cSAdam Langley 630fd6149d3SIlpo Järvinen if (unlikely(opts->mss)) { 631fd6149d3SIlpo Järvinen *ptr++ = htonl((TCPOPT_MSS << 24) | 632fd6149d3SIlpo Järvinen (TCPOLEN_MSS << 16) | 633fd6149d3SIlpo Järvinen opts->mss); 634fd6149d3SIlpo Järvinen } 635fd6149d3SIlpo Järvinen 636bd0388aeSWilliam Allen Simpson if (likely(OPTION_TS & options)) { 637bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 63833ad798cSAdam Langley *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | 63933ad798cSAdam Langley (TCPOLEN_SACK_PERM << 16) | 64033ad798cSAdam Langley (TCPOPT_TIMESTAMP << 8) | 64133ad798cSAdam Langley TCPOLEN_TIMESTAMP); 642bd0388aeSWilliam Allen Simpson options &= ~OPTION_SACK_ADVERTISE; 64333ad798cSAdam Langley } else { 644496c98dfSYOSHIFUJI Hideaki *ptr++ = htonl((TCPOPT_NOP << 24) | 64540efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 64640efc6faSStephen Hemminger (TCPOPT_TIMESTAMP << 8) | 64740efc6faSStephen Hemminger TCPOLEN_TIMESTAMP); 64840efc6faSStephen Hemminger } 64933ad798cSAdam Langley *ptr++ = htonl(opts->tsval); 65033ad798cSAdam Langley *ptr++ = htonl(opts->tsecr); 65133ad798cSAdam Langley } 65233ad798cSAdam Langley 653bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 65433ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 65533ad798cSAdam Langley (TCPOPT_NOP << 16) | 65633ad798cSAdam Langley (TCPOPT_SACK_PERM << 8) | 65733ad798cSAdam Langley TCPOLEN_SACK_PERM); 65833ad798cSAdam Langley } 65933ad798cSAdam Langley 660bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_WSCALE & options)) { 66133ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 66233ad798cSAdam Langley (TCPOPT_WINDOW << 16) | 66333ad798cSAdam Langley (TCPOLEN_WINDOW << 8) | 66433ad798cSAdam Langley opts->ws); 66533ad798cSAdam Langley } 66633ad798cSAdam Langley 66733ad798cSAdam Langley if (unlikely(opts->num_sack_blocks)) { 66833ad798cSAdam Langley struct tcp_sack_block *sp = tp->rx_opt.dsack ? 66933ad798cSAdam Langley tp->duplicate_sack : tp->selective_acks; 67040efc6faSStephen Hemminger int this_sack; 67140efc6faSStephen Hemminger 67240efc6faSStephen Hemminger *ptr++ = htonl((TCPOPT_NOP << 24) | 67340efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 67440efc6faSStephen Hemminger (TCPOPT_SACK << 8) | 67533ad798cSAdam Langley (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * 67640efc6faSStephen Hemminger TCPOLEN_SACK_PERBLOCK))); 6772de979bdSStephen Hemminger 67833ad798cSAdam Langley for (this_sack = 0; this_sack < opts->num_sack_blocks; 67933ad798cSAdam Langley ++this_sack) { 68040efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].start_seq); 68140efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].end_seq); 68240efc6faSStephen Hemminger } 6832de979bdSStephen Hemminger 68440efc6faSStephen Hemminger tp->rx_opt.dsack = 0; 68540efc6faSStephen Hemminger } 6862100c8d2SYuchung Cheng 6872100c8d2SYuchung Cheng if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) { 6882100c8d2SYuchung Cheng struct tcp_fastopen_cookie *foc = opts->fastopen_cookie; 6897f9b838bSDaniel Lee u8 *p = (u8 *)ptr; 6907f9b838bSDaniel Lee u32 len; /* Fast Open option length */ 6912100c8d2SYuchung Cheng 6927f9b838bSDaniel Lee if (foc->exp) { 6937f9b838bSDaniel Lee len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len; 6947f9b838bSDaniel Lee *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) | 6952100c8d2SYuchung Cheng TCPOPT_FASTOPEN_MAGIC); 6967f9b838bSDaniel Lee p += TCPOLEN_EXP_FASTOPEN_BASE; 6977f9b838bSDaniel Lee } else { 6987f9b838bSDaniel Lee len = TCPOLEN_FASTOPEN_BASE + foc->len; 6997f9b838bSDaniel Lee *p++ = TCPOPT_FASTOPEN; 7007f9b838bSDaniel Lee *p++ = len; 7012100c8d2SYuchung Cheng } 7027f9b838bSDaniel Lee 7037f9b838bSDaniel Lee memcpy(p, foc->val, foc->len); 7047f9b838bSDaniel Lee if ((len & 3) == 2) { 7057f9b838bSDaniel Lee p[foc->len] = TCPOPT_NOP; 7067f9b838bSDaniel Lee p[foc->len + 1] = TCPOPT_NOP; 7077f9b838bSDaniel Lee } 7087f9b838bSDaniel Lee ptr += (len + 3) >> 2; 7092100c8d2SYuchung Cheng } 71060e2a778SUrsula Braun 71160e2a778SUrsula Braun smc_options_write(ptr, &options); 712eda7acddSPeter Krystad 713ea66758cSPaolo Abeni mptcp_options_write(th, ptr, tp, opts); 71460e2a778SUrsula Braun } 71560e2a778SUrsula Braun 71660e2a778SUrsula Braun static void smc_set_option(const struct tcp_sock *tp, 71760e2a778SUrsula Braun struct tcp_out_options *opts, 71860e2a778SUrsula Braun unsigned int *remaining) 71960e2a778SUrsula Braun { 72060e2a778SUrsula Braun #if IS_ENABLED(CONFIG_SMC) 72160e2a778SUrsula Braun if (static_branch_unlikely(&tcp_have_smc)) { 72260e2a778SUrsula Braun if (tp->syn_smc) { 72360e2a778SUrsula Braun if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) { 72460e2a778SUrsula Braun opts->options |= OPTION_SMC; 72560e2a778SUrsula Braun *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED; 72660e2a778SUrsula Braun } 72760e2a778SUrsula Braun } 72860e2a778SUrsula Braun } 72960e2a778SUrsula Braun #endif 73060e2a778SUrsula Braun } 73160e2a778SUrsula Braun 73260e2a778SUrsula Braun static void smc_set_option_cond(const struct tcp_sock *tp, 73360e2a778SUrsula Braun const struct inet_request_sock *ireq, 73460e2a778SUrsula Braun struct tcp_out_options *opts, 73560e2a778SUrsula Braun unsigned int *remaining) 73660e2a778SUrsula Braun { 73760e2a778SUrsula Braun #if IS_ENABLED(CONFIG_SMC) 73860e2a778SUrsula Braun if (static_branch_unlikely(&tcp_have_smc)) { 73960e2a778SUrsula Braun if (tp->syn_smc && ireq->smc_ok) { 74060e2a778SUrsula Braun if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) { 74160e2a778SUrsula Braun opts->options |= OPTION_SMC; 74260e2a778SUrsula Braun *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED; 74360e2a778SUrsula Braun } 74460e2a778SUrsula Braun } 74560e2a778SUrsula Braun } 74660e2a778SUrsula Braun #endif 74740efc6faSStephen Hemminger } 74840efc6faSStephen Hemminger 749cec37a6eSPeter Krystad static void mptcp_set_option_cond(const struct request_sock *req, 750cec37a6eSPeter Krystad struct tcp_out_options *opts, 751cec37a6eSPeter Krystad unsigned int *remaining) 752cec37a6eSPeter Krystad { 753cec37a6eSPeter Krystad if (rsk_is_mptcp(req)) { 754cec37a6eSPeter Krystad unsigned int size; 755cec37a6eSPeter Krystad 756cec37a6eSPeter Krystad if (mptcp_synack_options(req, &size, &opts->mptcp)) { 757cec37a6eSPeter Krystad if (*remaining >= size) { 758cec37a6eSPeter Krystad opts->options |= OPTION_MPTCP; 759cec37a6eSPeter Krystad *remaining -= size; 760cec37a6eSPeter Krystad } 761cec37a6eSPeter Krystad } 762cec37a6eSPeter Krystad } 763cec37a6eSPeter Krystad } 764cec37a6eSPeter Krystad 76567edfef7SAndi Kleen /* Compute TCP options for SYN packets. This is not the final 76667edfef7SAndi Kleen * network wire format yet. 76767edfef7SAndi Kleen */ 76895c96174SEric Dumazet static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb, 76933ad798cSAdam Langley struct tcp_out_options *opts, 770cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 771cf533ea5SEric Dumazet { 77233ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 77395c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 774783237e8SYuchung Cheng struct tcp_fastopen_request *fastopen = tp->fastopen_req; 77533ad798cSAdam Langley 7768c2320e8SEric Dumazet *md5 = NULL; 777cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 778459837b5SDmitry Safonov if (static_branch_unlikely(&tcp_md5_needed.key) && 7796015c71eSEric Dumazet rcu_access_pointer(tp->md5sig_info)) { 78033ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 78133ad798cSAdam Langley if (*md5) { 78233ad798cSAdam Langley opts->options |= OPTION_MD5; 783bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 784cfb6eeb4SYOSHIFUJI Hideaki } 7858c2320e8SEric Dumazet } 786cfb6eeb4SYOSHIFUJI Hideaki #endif 78733ad798cSAdam Langley 78833ad798cSAdam Langley /* We always get an MSS option. The option bytes which will be seen in 78933ad798cSAdam Langley * normal data packets should timestamps be used, must be in the MSS 79033ad798cSAdam Langley * advertised. But we subtract them from tp->mss_cache so that 79133ad798cSAdam Langley * calculations in tcp_sendmsg are simpler etc. So account for this 79233ad798cSAdam Langley * fact here if necessary. If we don't do this correctly, as a 79333ad798cSAdam Langley * receiver we won't recognize data packets as being full sized when we 79433ad798cSAdam Langley * should, and thus we won't abide by the delayed ACK rules correctly. 79533ad798cSAdam Langley * SACKs don't matter, we never delay an ACK when we have any of those 79633ad798cSAdam Langley * going out. */ 79733ad798cSAdam Langley opts->mss = tcp_advertise_mss(sk); 798bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 79933ad798cSAdam Langley 8003666f666SKuniyuki Iwashima if (likely(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_timestamps) && !*md5)) { 80133ad798cSAdam Langley opts->options |= OPTION_TS; 8027faee5c0SEric Dumazet opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset; 80333ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 804bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 80533ad798cSAdam Langley } 8063666f666SKuniyuki Iwashima if (likely(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_window_scaling))) { 80733ad798cSAdam Langley opts->ws = tp->rx_opt.rcv_wscale; 80889e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 809bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 81033ad798cSAdam Langley } 8113666f666SKuniyuki Iwashima if (likely(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_sack))) { 81233ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 813b32d1310SDavid S. Miller if (unlikely(!(OPTION_TS & opts->options))) 814bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 81533ad798cSAdam Langley } 81633ad798cSAdam Langley 817783237e8SYuchung Cheng if (fastopen && fastopen->cookie.len >= 0) { 8182646c831SDaniel Lee u32 need = fastopen->cookie.len; 8192646c831SDaniel Lee 8202646c831SDaniel Lee need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE : 8212646c831SDaniel Lee TCPOLEN_FASTOPEN_BASE; 822783237e8SYuchung Cheng need = (need + 3) & ~3U; /* Align to 32 bits */ 823783237e8SYuchung Cheng if (remaining >= need) { 824783237e8SYuchung Cheng opts->options |= OPTION_FAST_OPEN_COOKIE; 825783237e8SYuchung Cheng opts->fastopen_cookie = &fastopen->cookie; 826783237e8SYuchung Cheng remaining -= need; 827783237e8SYuchung Cheng tp->syn_fastopen = 1; 8282646c831SDaniel Lee tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0; 829783237e8SYuchung Cheng } 830783237e8SYuchung Cheng } 831bd0388aeSWilliam Allen Simpson 83260e2a778SUrsula Braun smc_set_option(tp, opts, &remaining); 83360e2a778SUrsula Braun 834cec37a6eSPeter Krystad if (sk_is_mptcp(sk)) { 835cec37a6eSPeter Krystad unsigned int size; 836cec37a6eSPeter Krystad 837cc7972eaSChristoph Paasch if (mptcp_syn_options(sk, skb, &size, &opts->mptcp)) { 838cec37a6eSPeter Krystad opts->options |= OPTION_MPTCP; 839cec37a6eSPeter Krystad remaining -= size; 840cec37a6eSPeter Krystad } 841cec37a6eSPeter Krystad } 842cec37a6eSPeter Krystad 843331fca43SMartin KaFai Lau bpf_skops_hdr_opt_len(sk, skb, NULL, NULL, 0, opts, &remaining); 844331fca43SMartin KaFai Lau 845bd0388aeSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 84633ad798cSAdam Langley } 84733ad798cSAdam Langley 84867edfef7SAndi Kleen /* Set up TCP options for SYN-ACKs. */ 84960e2a778SUrsula Braun static unsigned int tcp_synack_options(const struct sock *sk, 85060e2a778SUrsula Braun struct request_sock *req, 85195c96174SEric Dumazet unsigned int mss, struct sk_buff *skb, 85233ad798cSAdam Langley struct tcp_out_options *opts, 85380f03e27SEric Dumazet const struct tcp_md5sig_key *md5, 854e114e1e8SEric Dumazet struct tcp_fastopen_cookie *foc, 855331fca43SMartin KaFai Lau enum tcp_synack_type synack_type, 856331fca43SMartin KaFai Lau struct sk_buff *syn_skb) 8574957faadSWilliam Allen Simpson { 85833ad798cSAdam Langley struct inet_request_sock *ireq = inet_rsk(req); 85995c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 86033ad798cSAdam Langley 86133ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 86280f03e27SEric Dumazet if (md5) { 86333ad798cSAdam Langley opts->options |= OPTION_MD5; 8644957faadSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 8654957faadSWilliam Allen Simpson 8664957faadSWilliam Allen Simpson /* We can't fit any SACK blocks in a packet with MD5 + TS 8674957faadSWilliam Allen Simpson * options. There was discussion about disabling SACK 8684957faadSWilliam Allen Simpson * rather than TS in order to fit in better with old, 8694957faadSWilliam Allen Simpson * buggy kernels, but that was deemed to be unnecessary. 8704957faadSWilliam Allen Simpson */ 871e114e1e8SEric Dumazet if (synack_type != TCP_SYNACK_COOKIE) 872de213e5eSEric Dumazet ireq->tstamp_ok &= !ireq->sack_ok; 87333ad798cSAdam Langley } 87433ad798cSAdam Langley #endif 87533ad798cSAdam Langley 8764957faadSWilliam Allen Simpson /* We always send an MSS option. */ 87733ad798cSAdam Langley opts->mss = mss; 8784957faadSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 87933ad798cSAdam Langley 88033ad798cSAdam Langley if (likely(ireq->wscale_ok)) { 88133ad798cSAdam Langley opts->ws = ireq->rcv_wscale; 88289e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 8834957faadSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 88433ad798cSAdam Langley } 885de213e5eSEric Dumazet if (likely(ireq->tstamp_ok)) { 88633ad798cSAdam Langley opts->options |= OPTION_TS; 88795a22caeSFlorian Westphal opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off; 888eba20811SEric Dumazet opts->tsecr = READ_ONCE(req->ts_recent); 8894957faadSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 89033ad798cSAdam Langley } 89133ad798cSAdam Langley if (likely(ireq->sack_ok)) { 89233ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 893de213e5eSEric Dumazet if (unlikely(!ireq->tstamp_ok)) 8944957faadSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 89533ad798cSAdam Langley } 8967f9b838bSDaniel Lee if (foc != NULL && foc->len >= 0) { 8977f9b838bSDaniel Lee u32 need = foc->len; 8987f9b838bSDaniel Lee 8997f9b838bSDaniel Lee need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE : 9007f9b838bSDaniel Lee TCPOLEN_FASTOPEN_BASE; 9018336886fSJerry Chu need = (need + 3) & ~3U; /* Align to 32 bits */ 9028336886fSJerry Chu if (remaining >= need) { 9038336886fSJerry Chu opts->options |= OPTION_FAST_OPEN_COOKIE; 9048336886fSJerry Chu opts->fastopen_cookie = foc; 9058336886fSJerry Chu remaining -= need; 9068336886fSJerry Chu } 9078336886fSJerry Chu } 9084957faadSWilliam Allen Simpson 909cec37a6eSPeter Krystad mptcp_set_option_cond(req, opts, &remaining); 910cec37a6eSPeter Krystad 91160e2a778SUrsula Braun smc_set_option_cond(tcp_sk(sk), ireq, opts, &remaining); 91260e2a778SUrsula Braun 913331fca43SMartin KaFai Lau bpf_skops_hdr_opt_len((struct sock *)sk, skb, req, syn_skb, 914331fca43SMartin KaFai Lau synack_type, opts, &remaining); 915331fca43SMartin KaFai Lau 9164957faadSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 91733ad798cSAdam Langley } 91833ad798cSAdam Langley 91967edfef7SAndi Kleen /* Compute TCP options for ESTABLISHED sockets. This is not the 92067edfef7SAndi Kleen * final wire format yet. 92167edfef7SAndi Kleen */ 92295c96174SEric Dumazet static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb, 92333ad798cSAdam Langley struct tcp_out_options *opts, 924cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 925cf533ea5SEric Dumazet { 92633ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 92795c96174SEric Dumazet unsigned int size = 0; 928cabeccbdSIlpo Järvinen unsigned int eff_sacks; 92933ad798cSAdam Langley 9305843ef42SAndi Kleen opts->options = 0; 9315843ef42SAndi Kleen 9328c2320e8SEric Dumazet *md5 = NULL; 93333ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 934459837b5SDmitry Safonov if (static_branch_unlikely(&tcp_md5_needed.key) && 9356015c71eSEric Dumazet rcu_access_pointer(tp->md5sig_info)) { 93633ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 9378c2320e8SEric Dumazet if (*md5) { 93833ad798cSAdam Langley opts->options |= OPTION_MD5; 93933ad798cSAdam Langley size += TCPOLEN_MD5SIG_ALIGNED; 94033ad798cSAdam Langley } 9418c2320e8SEric Dumazet } 94233ad798cSAdam Langley #endif 94333ad798cSAdam Langley 94433ad798cSAdam Langley if (likely(tp->rx_opt.tstamp_ok)) { 94533ad798cSAdam Langley opts->options |= OPTION_TS; 9467faee5c0SEric Dumazet opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0; 94733ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 94833ad798cSAdam Langley size += TCPOLEN_TSTAMP_ALIGNED; 94933ad798cSAdam Langley } 95033ad798cSAdam Langley 951cec37a6eSPeter Krystad /* MPTCP options have precedence over SACK for the limited TCP 952cec37a6eSPeter Krystad * option space because a MPTCP connection would be forced to 953cec37a6eSPeter Krystad * fall back to regular TCP if a required multipath option is 954cec37a6eSPeter Krystad * missing. SACK still gets a chance to use whatever space is 955cec37a6eSPeter Krystad * left. 956cec37a6eSPeter Krystad */ 957cec37a6eSPeter Krystad if (sk_is_mptcp(sk)) { 958cec37a6eSPeter Krystad unsigned int remaining = MAX_TCP_OPTION_SPACE - size; 959cec37a6eSPeter Krystad unsigned int opt_size = 0; 960cec37a6eSPeter Krystad 961cec37a6eSPeter Krystad if (mptcp_established_options(sk, skb, &opt_size, remaining, 962cec37a6eSPeter Krystad &opts->mptcp)) { 963cec37a6eSPeter Krystad opts->options |= OPTION_MPTCP; 964cec37a6eSPeter Krystad size += opt_size; 965cec37a6eSPeter Krystad } 966cec37a6eSPeter Krystad } 967cec37a6eSPeter Krystad 968cabeccbdSIlpo Järvinen eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; 969cabeccbdSIlpo Järvinen if (unlikely(eff_sacks)) { 97095c96174SEric Dumazet const unsigned int remaining = MAX_TCP_OPTION_SPACE - size; 9719cfcca23SMat Martineau if (unlikely(remaining < TCPOLEN_SACK_BASE_ALIGNED + 9729cfcca23SMat Martineau TCPOLEN_SACK_PERBLOCK)) 9739cfcca23SMat Martineau return size; 9749cfcca23SMat Martineau 97533ad798cSAdam Langley opts->num_sack_blocks = 97695c96174SEric Dumazet min_t(unsigned int, eff_sacks, 97733ad798cSAdam Langley (remaining - TCPOLEN_SACK_BASE_ALIGNED) / 97833ad798cSAdam Langley TCPOLEN_SACK_PERBLOCK); 9799cfcca23SMat Martineau 98033ad798cSAdam Langley size += TCPOLEN_SACK_BASE_ALIGNED + 98133ad798cSAdam Langley opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; 98233ad798cSAdam Langley } 98333ad798cSAdam Langley 984331fca43SMartin KaFai Lau if (unlikely(BPF_SOCK_OPS_TEST_FLAG(tp, 985331fca43SMartin KaFai Lau BPF_SOCK_OPS_WRITE_HDR_OPT_CB_FLAG))) { 986331fca43SMartin KaFai Lau unsigned int remaining = MAX_TCP_OPTION_SPACE - size; 987331fca43SMartin KaFai Lau 988331fca43SMartin KaFai Lau bpf_skops_hdr_opt_len(sk, skb, NULL, NULL, 0, opts, &remaining); 989331fca43SMartin KaFai Lau 990331fca43SMartin KaFai Lau size = MAX_TCP_OPTION_SPACE - remaining; 991331fca43SMartin KaFai Lau } 992331fca43SMartin KaFai Lau 99333ad798cSAdam Langley return size; 99440efc6faSStephen Hemminger } 9951da177e4SLinus Torvalds 99646d3ceabSEric Dumazet 99746d3ceabSEric Dumazet /* TCP SMALL QUEUES (TSQ) 99846d3ceabSEric Dumazet * 99946d3ceabSEric Dumazet * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev) 100046d3ceabSEric Dumazet * to reduce RTT and bufferbloat. 100146d3ceabSEric Dumazet * We do this using a special skb destructor (tcp_wfree). 100246d3ceabSEric Dumazet * 100346d3ceabSEric Dumazet * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb 100446d3ceabSEric Dumazet * needs to be reallocated in a driver. 10058e3bff96Sstephen hemminger * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc 100646d3ceabSEric Dumazet * 100746d3ceabSEric Dumazet * Since transmit from skb destructor is forbidden, we use a tasklet 100846d3ceabSEric Dumazet * to process all sockets that eventually need to send more skbs. 100946d3ceabSEric Dumazet * We use one tasklet per cpu, with its own queue of sockets. 101046d3ceabSEric Dumazet */ 101146d3ceabSEric Dumazet struct tsq_tasklet { 101246d3ceabSEric Dumazet struct tasklet_struct tasklet; 101346d3ceabSEric Dumazet struct list_head head; /* queue of tcp sockets */ 101446d3ceabSEric Dumazet }; 101546d3ceabSEric Dumazet static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet); 101646d3ceabSEric Dumazet 101773a6bab5SEric Dumazet static void tcp_tsq_write(struct sock *sk) 10186f458dfbSEric Dumazet { 10196f458dfbSEric Dumazet if ((1 << sk->sk_state) & 10206f458dfbSEric Dumazet (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING | 1021f9616c35SEric Dumazet TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) { 1022f9616c35SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 1023f9616c35SEric Dumazet 1024f9616c35SEric Dumazet if (tp->lost_out > tp->retrans_out && 102540570375SEric Dumazet tcp_snd_cwnd(tp) > tcp_packets_in_flight(tp)) { 10263a91d29fSKoichiro Den tcp_mstamp_refresh(tp); 1027f9616c35SEric Dumazet tcp_xmit_retransmit_queue(sk); 10283a91d29fSKoichiro Den } 1029f9616c35SEric Dumazet 1030f9616c35SEric Dumazet tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle, 1031bf06200eSJohn Ogness 0, GFP_ATOMIC); 10326f458dfbSEric Dumazet } 1033f9616c35SEric Dumazet } 103473a6bab5SEric Dumazet 103573a6bab5SEric Dumazet static void tcp_tsq_handler(struct sock *sk) 103673a6bab5SEric Dumazet { 103773a6bab5SEric Dumazet bh_lock_sock(sk); 103873a6bab5SEric Dumazet if (!sock_owned_by_user(sk)) 103973a6bab5SEric Dumazet tcp_tsq_write(sk); 104073a6bab5SEric Dumazet else if (!test_and_set_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags)) 104173a6bab5SEric Dumazet sock_hold(sk); 104273a6bab5SEric Dumazet bh_unlock_sock(sk); 104373a6bab5SEric Dumazet } 104446d3ceabSEric Dumazet /* 10458e3bff96Sstephen hemminger * One tasklet per cpu tries to send more skbs. 104646d3ceabSEric Dumazet * We run in tasklet context but need to disable irqs when 10478e3bff96Sstephen hemminger * transferring tsq->head because tcp_wfree() might 104846d3ceabSEric Dumazet * interrupt us (non NAPI drivers) 104946d3ceabSEric Dumazet */ 1050c6533ca8SAllen Pais static void tcp_tasklet_func(struct tasklet_struct *t) 105146d3ceabSEric Dumazet { 1052c6533ca8SAllen Pais struct tsq_tasklet *tsq = from_tasklet(tsq, t, tasklet); 105346d3ceabSEric Dumazet LIST_HEAD(list); 105446d3ceabSEric Dumazet unsigned long flags; 105546d3ceabSEric Dumazet struct list_head *q, *n; 105646d3ceabSEric Dumazet struct tcp_sock *tp; 105746d3ceabSEric Dumazet struct sock *sk; 105846d3ceabSEric Dumazet 105946d3ceabSEric Dumazet local_irq_save(flags); 106046d3ceabSEric Dumazet list_splice_init(&tsq->head, &list); 106146d3ceabSEric Dumazet local_irq_restore(flags); 106246d3ceabSEric Dumazet 106346d3ceabSEric Dumazet list_for_each_safe(q, n, &list) { 106446d3ceabSEric Dumazet tp = list_entry(q, struct tcp_sock, tsq_node); 106546d3ceabSEric Dumazet list_del(&tp->tsq_node); 106646d3ceabSEric Dumazet 106746d3ceabSEric Dumazet sk = (struct sock *)tp; 10680a9648f1SEric Dumazet smp_mb__before_atomic(); 10697aa5470cSEric Dumazet clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags); 10707aa5470cSEric Dumazet 10716f458dfbSEric Dumazet tcp_tsq_handler(sk); 107246d3ceabSEric Dumazet sk_free(sk); 107346d3ceabSEric Dumazet } 107446d3ceabSEric Dumazet } 107546d3ceabSEric Dumazet 107640fc3423SEric Dumazet #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \ 107740fc3423SEric Dumazet TCPF_WRITE_TIMER_DEFERRED | \ 107840fc3423SEric Dumazet TCPF_DELACK_TIMER_DEFERRED | \ 107940fc3423SEric Dumazet TCPF_MTU_REDUCED_DEFERRED) 108046d3ceabSEric Dumazet /** 108146d3ceabSEric Dumazet * tcp_release_cb - tcp release_sock() callback 108246d3ceabSEric Dumazet * @sk: socket 108346d3ceabSEric Dumazet * 108446d3ceabSEric Dumazet * called from release_sock() to perform protocol dependent 108546d3ceabSEric Dumazet * actions before socket release. 108646d3ceabSEric Dumazet */ 108746d3ceabSEric Dumazet void tcp_release_cb(struct sock *sk) 108846d3ceabSEric Dumazet { 1089fac30731SEric Dumazet unsigned long flags = smp_load_acquire(&sk->sk_tsq_flags); 1090fac30731SEric Dumazet unsigned long nflags; 109146d3ceabSEric Dumazet 10926f458dfbSEric Dumazet /* perform an atomic operation only if at least one flag is set */ 10936f458dfbSEric Dumazet do { 10946f458dfbSEric Dumazet if (!(flags & TCP_DEFERRED_ALL)) 10956f458dfbSEric Dumazet return; 10966f458dfbSEric Dumazet nflags = flags & ~TCP_DEFERRED_ALL; 1097fac30731SEric Dumazet } while (!try_cmpxchg(&sk->sk_tsq_flags, &flags, nflags)); 10986f458dfbSEric Dumazet 109973a6bab5SEric Dumazet if (flags & TCPF_TSQ_DEFERRED) { 110073a6bab5SEric Dumazet tcp_tsq_write(sk); 110173a6bab5SEric Dumazet __sock_put(sk); 110273a6bab5SEric Dumazet } 1103c3f9b018SEric Dumazet /* Here begins the tricky part : 1104c3f9b018SEric Dumazet * We are called from release_sock() with : 1105c3f9b018SEric Dumazet * 1) BH disabled 1106c3f9b018SEric Dumazet * 2) sk_lock.slock spinlock held 1107c3f9b018SEric Dumazet * 3) socket owned by us (sk->sk_lock.owned == 1) 1108c3f9b018SEric Dumazet * 1109c3f9b018SEric Dumazet * But following code is meant to be called from BH handlers, 1110c3f9b018SEric Dumazet * so we should keep BH disabled, but early release socket ownership 1111c3f9b018SEric Dumazet */ 1112c3f9b018SEric Dumazet sock_release_ownership(sk); 1113c3f9b018SEric Dumazet 111440fc3423SEric Dumazet if (flags & TCPF_WRITE_TIMER_DEFERRED) { 11156f458dfbSEric Dumazet tcp_write_timer_handler(sk); 1116144d56e9SEric Dumazet __sock_put(sk); 1117144d56e9SEric Dumazet } 111840fc3423SEric Dumazet if (flags & TCPF_DELACK_TIMER_DEFERRED) { 11196f458dfbSEric Dumazet tcp_delack_timer_handler(sk); 1120144d56e9SEric Dumazet __sock_put(sk); 1121144d56e9SEric Dumazet } 112240fc3423SEric Dumazet if (flags & TCPF_MTU_REDUCED_DEFERRED) { 11234fab9071SNeal Cardwell inet_csk(sk)->icsk_af_ops->mtu_reduced(sk); 1124144d56e9SEric Dumazet __sock_put(sk); 1125144d56e9SEric Dumazet } 112646d3ceabSEric Dumazet } 112746d3ceabSEric Dumazet EXPORT_SYMBOL(tcp_release_cb); 112846d3ceabSEric Dumazet 112946d3ceabSEric Dumazet void __init tcp_tasklet_init(void) 113046d3ceabSEric Dumazet { 113146d3ceabSEric Dumazet int i; 113246d3ceabSEric Dumazet 113346d3ceabSEric Dumazet for_each_possible_cpu(i) { 113446d3ceabSEric Dumazet struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i); 113546d3ceabSEric Dumazet 113646d3ceabSEric Dumazet INIT_LIST_HEAD(&tsq->head); 1137c6533ca8SAllen Pais tasklet_setup(&tsq->tasklet, tcp_tasklet_func); 113846d3ceabSEric Dumazet } 113946d3ceabSEric Dumazet } 114046d3ceabSEric Dumazet 114146d3ceabSEric Dumazet /* 114246d3ceabSEric Dumazet * Write buffer destructor automatically called from kfree_skb. 11438e3bff96Sstephen hemminger * We can't xmit new skbs from this context, as we might already 114446d3ceabSEric Dumazet * hold qdisc lock. 114546d3ceabSEric Dumazet */ 1146d6a4a104SEric Dumazet void tcp_wfree(struct sk_buff *skb) 114746d3ceabSEric Dumazet { 114846d3ceabSEric Dumazet struct sock *sk = skb->sk; 114946d3ceabSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 1150408f0a6cSEric Dumazet unsigned long flags, nval, oval; 1151b548b17aSEric Dumazet struct tsq_tasklet *tsq; 1152b548b17aSEric Dumazet bool empty; 11539b462d02SEric Dumazet 11549b462d02SEric Dumazet /* Keep one reference on sk_wmem_alloc. 11559b462d02SEric Dumazet * Will be released by sk_free() from here or tcp_tasklet_func() 11569b462d02SEric Dumazet */ 115714afee4bSReshetova, Elena WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc)); 11589b462d02SEric Dumazet 11599b462d02SEric Dumazet /* If this softirq is serviced by ksoftirqd, we are likely under stress. 11609b462d02SEric Dumazet * Wait until our queues (qdisc + devices) are drained. 11619b462d02SEric Dumazet * This gives : 11629b462d02SEric Dumazet * - less callbacks to tcp_write_xmit(), reducing stress (batches) 11639b462d02SEric Dumazet * - chance for incoming ACK (processed by another cpu maybe) 11649b462d02SEric Dumazet * to migrate this flow (skb->ooo_okay will be eventually set) 11659b462d02SEric Dumazet */ 116614afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current) 11679b462d02SEric Dumazet goto out; 116846d3ceabSEric Dumazet 1169b548b17aSEric Dumazet oval = smp_load_acquire(&sk->sk_tsq_flags); 1170b548b17aSEric Dumazet do { 1171408f0a6cSEric Dumazet if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED)) 1172408f0a6cSEric Dumazet goto out; 1173408f0a6cSEric Dumazet 117473a6bab5SEric Dumazet nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED; 1175b548b17aSEric Dumazet } while (!try_cmpxchg(&sk->sk_tsq_flags, &oval, nval)); 1176408f0a6cSEric Dumazet 117746d3ceabSEric Dumazet /* queue this socket to tasklet queue */ 117846d3ceabSEric Dumazet local_irq_save(flags); 1179903ceff7SChristoph Lameter tsq = this_cpu_ptr(&tsq_tasklet); 1180a9b204d1SEric Dumazet empty = list_empty(&tsq->head); 118146d3ceabSEric Dumazet list_add(&tp->tsq_node, &tsq->head); 1182a9b204d1SEric Dumazet if (empty) 118346d3ceabSEric Dumazet tasklet_schedule(&tsq->tasklet); 118446d3ceabSEric Dumazet local_irq_restore(flags); 11859b462d02SEric Dumazet return; 11869b462d02SEric Dumazet out: 11879b462d02SEric Dumazet sk_free(sk); 118846d3ceabSEric Dumazet } 118946d3ceabSEric Dumazet 119073a6bab5SEric Dumazet /* Note: Called under soft irq. 119173a6bab5SEric Dumazet * We can call TCP stack right away, unless socket is owned by user. 1192218af599SEric Dumazet */ 1193218af599SEric Dumazet enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer) 1194218af599SEric Dumazet { 1195218af599SEric Dumazet struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer); 1196218af599SEric Dumazet struct sock *sk = (struct sock *)tp; 1197218af599SEric Dumazet 119873a6bab5SEric Dumazet tcp_tsq_handler(sk); 119973a6bab5SEric Dumazet sock_put(sk); 1200218af599SEric Dumazet 1201218af599SEric Dumazet return HRTIMER_NORESTART; 1202218af599SEric Dumazet } 1203218af599SEric Dumazet 1204a7a25630SEric Dumazet static void tcp_update_skb_after_send(struct sock *sk, struct sk_buff *skb, 1205a7a25630SEric Dumazet u64 prior_wstamp) 1206e2080072SEric Dumazet { 1207ab408b6dSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 1208ab408b6dSEric Dumazet 1209ab408b6dSEric Dumazet if (sk->sk_pacing_status != SK_PACING_NONE) { 121076a9ebe8SEric Dumazet unsigned long rate = sk->sk_pacing_rate; 1211ab408b6dSEric Dumazet 1212ab408b6dSEric Dumazet /* Original sch_fq does not pace first 10 MSS 1213ab408b6dSEric Dumazet * Note that tp->data_segs_out overflows after 2^32 packets, 1214ab408b6dSEric Dumazet * this is a minor annoyance. 1215ab408b6dSEric Dumazet */ 121676a9ebe8SEric Dumazet if (rate != ~0UL && rate && tp->data_segs_out >= 10) { 1217a7a25630SEric Dumazet u64 len_ns = div64_ul((u64)skb->len * NSEC_PER_SEC, rate); 1218a7a25630SEric Dumazet u64 credit = tp->tcp_wstamp_ns - prior_wstamp; 1219a7a25630SEric Dumazet 1220a7a25630SEric Dumazet /* take into account OS jitter */ 1221a7a25630SEric Dumazet len_ns -= min_t(u64, len_ns / 2, credit); 1222a7a25630SEric Dumazet tp->tcp_wstamp_ns += len_ns; 1223ab408b6dSEric Dumazet } 1224ab408b6dSEric Dumazet } 1225e2080072SEric Dumazet list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue); 1226e2080072SEric Dumazet } 1227e2080072SEric Dumazet 122805e22e83SEric Dumazet INDIRECT_CALLABLE_DECLARE(int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl)); 122905e22e83SEric Dumazet INDIRECT_CALLABLE_DECLARE(int inet6_csk_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl)); 1230dd2e0b86SEric Dumazet INDIRECT_CALLABLE_DECLARE(void tcp_v4_send_check(struct sock *sk, struct sk_buff *skb)); 123105e22e83SEric Dumazet 12321da177e4SLinus Torvalds /* This routine actually transmits TCP packets queued in by 12331da177e4SLinus Torvalds * tcp_do_sendmsg(). This is used by both the initial 12341da177e4SLinus Torvalds * transmission and possible later retransmissions. 12351da177e4SLinus Torvalds * All SKB's seen here are completely headerless. It is our 12361da177e4SLinus Torvalds * job to build the TCP header, and pass the packet down to 12371da177e4SLinus Torvalds * IP so it can do the same plus pass the packet off to the 12381da177e4SLinus Torvalds * device. 12391da177e4SLinus Torvalds * 12401da177e4SLinus Torvalds * We are working here with either a clone of the original 12411da177e4SLinus Torvalds * SKB, or a fresh unique copy made by the retransmit engine. 12421da177e4SLinus Torvalds */ 12432987babbSYuchung Cheng static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, 12442987babbSYuchung Cheng int clone_it, gfp_t gfp_mask, u32 rcv_nxt) 12451da177e4SLinus Torvalds { 12466687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 1247dfb4b9dcSDavid S. Miller struct inet_sock *inet; 1248dfb4b9dcSDavid S. Miller struct tcp_sock *tp; 1249dfb4b9dcSDavid S. Miller struct tcp_skb_cb *tcb; 125033ad798cSAdam Langley struct tcp_out_options opts; 125195c96174SEric Dumazet unsigned int tcp_options_size, tcp_header_size; 12528c72c65bSEric Dumazet struct sk_buff *oskb = NULL; 1253cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 12541da177e4SLinus Torvalds struct tcphdr *th; 1255a7a25630SEric Dumazet u64 prior_wstamp; 12561da177e4SLinus Torvalds int err; 12571da177e4SLinus Torvalds 1258dfb4b9dcSDavid S. Miller BUG_ON(!skb || !tcp_skb_pcount(skb)); 12596f094b9eSLawrence Brakmo tp = tcp_sk(sk); 12607f12422cSYuchung Cheng prior_wstamp = tp->tcp_wstamp_ns; 12617f12422cSYuchung Cheng tp->tcp_wstamp_ns = max(tp->tcp_wstamp_ns, tp->tcp_clock_cache); 1262a1ac9c8aSMartin KaFai Lau skb_set_delivery_time(skb, tp->tcp_wstamp_ns, true); 1263ccdbb6e9SEric Dumazet if (clone_it) { 12648c72c65bSEric Dumazet oskb = skb; 1265e2080072SEric Dumazet 1266e2080072SEric Dumazet tcp_skb_tsorted_save(oskb) { 1267e2080072SEric Dumazet if (unlikely(skb_cloned(oskb))) 1268e2080072SEric Dumazet skb = pskb_copy(oskb, gfp_mask); 1269dfb4b9dcSDavid S. Miller else 1270e2080072SEric Dumazet skb = skb_clone(oskb, gfp_mask); 1271e2080072SEric Dumazet } tcp_skb_tsorted_restore(oskb); 1272e2080072SEric Dumazet 1273dfb4b9dcSDavid S. Miller if (unlikely(!skb)) 1274dfb4b9dcSDavid S. Miller return -ENOBUFS; 1275b738a185SEric Dumazet /* retransmit skbs might have a non zero value in skb->dev 1276b738a185SEric Dumazet * because skb->dev is aliased with skb->rbnode.rb_left 1277b738a185SEric Dumazet */ 1278b738a185SEric Dumazet skb->dev = NULL; 1279dfb4b9dcSDavid S. Miller } 12805f6188a8SEric Dumazet 1281dfb4b9dcSDavid S. Miller inet = inet_sk(sk); 1282dfb4b9dcSDavid S. Miller tcb = TCP_SKB_CB(skb); 128333ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 12841da177e4SLinus Torvalds 1285051ba674SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) { 128633ad798cSAdam Langley tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); 1287051ba674SEric Dumazet } else { 128833ad798cSAdam Langley tcp_options_size = tcp_established_options(sk, skb, &opts, 128933ad798cSAdam Langley &md5); 1290051ba674SEric Dumazet /* Force a PSH flag on all (GSO) packets to expedite GRO flush 1291051ba674SEric Dumazet * at receiver : This slightly improve GRO performance. 1292051ba674SEric Dumazet * Note that we do not force the PSH flag for non GSO packets, 1293051ba674SEric Dumazet * because they might be sent under high congestion events, 1294051ba674SEric Dumazet * and in this case it is better to delay the delivery of 1-MSS 1295051ba674SEric Dumazet * packets and thus the corresponding ACK packet that would 1296051ba674SEric Dumazet * release the following packet. 1297051ba674SEric Dumazet */ 1298051ba674SEric Dumazet if (tcp_skb_pcount(skb) > 1) 1299051ba674SEric Dumazet tcb->tcp_flags |= TCPHDR_PSH; 1300051ba674SEric Dumazet } 130133ad798cSAdam Langley tcp_header_size = tcp_options_size + sizeof(struct tcphdr); 13021da177e4SLinus Torvalds 1303726e9e8bSEric Dumazet /* We set skb->ooo_okay to one if this packet can select 1304726e9e8bSEric Dumazet * a different TX queue than prior packets of this flow, 1305726e9e8bSEric Dumazet * to avoid self inflicted reorders. 1306726e9e8bSEric Dumazet * The 'other' queue decision is based on current cpu number 1307726e9e8bSEric Dumazet * if XPS is enabled, or sk->sk_txhash otherwise. 1308726e9e8bSEric Dumazet * We can switch to another (and better) queue if: 1309726e9e8bSEric Dumazet * 1) No packet with payload is in qdisc/device queues. 1310726e9e8bSEric Dumazet * Delays in TX completion can defeat the test 1311726e9e8bSEric Dumazet * even if packets were already sent. 1312726e9e8bSEric Dumazet * 2) Or rtx queue is empty. 1313726e9e8bSEric Dumazet * This mitigates above case if ACK packets for 1314726e9e8bSEric Dumazet * all prior packets were already processed. 1315547669d4SEric Dumazet */ 1316726e9e8bSEric Dumazet skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1) || 1317726e9e8bSEric Dumazet tcp_rtx_queue_empty(sk); 13181da177e4SLinus Torvalds 131938ab52e8SEric Dumazet /* If we had to use memory reserve to allocate this skb, 132038ab52e8SEric Dumazet * this might cause drops if packet is looped back : 132138ab52e8SEric Dumazet * Other socket might not have SOCK_MEMALLOC. 132238ab52e8SEric Dumazet * Packets not looped back do not care about pfmemalloc. 132338ab52e8SEric Dumazet */ 132438ab52e8SEric Dumazet skb->pfmemalloc = 0; 132538ab52e8SEric Dumazet 1326aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 1327aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 132846d3ceabSEric Dumazet 132946d3ceabSEric Dumazet skb_orphan(skb); 133046d3ceabSEric Dumazet skb->sk = sk; 13311d2077acSEric Dumazet skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree; 133214afee4bSReshetova, Elena refcount_add(skb->truesize, &sk->sk_wmem_alloc); 13331da177e4SLinus Torvalds 1334c3a2e837SJulian Anastasov skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm); 1335c3a2e837SJulian Anastasov 13361da177e4SLinus Torvalds /* Build TCP header and checksum it. */ 1337ea1627c2SEric Dumazet th = (struct tcphdr *)skb->data; 1338c720c7e8SEric Dumazet th->source = inet->inet_sport; 1339c720c7e8SEric Dumazet th->dest = inet->inet_dport; 13401da177e4SLinus Torvalds th->seq = htonl(tcb->seq); 13412987babbSYuchung Cheng th->ack_seq = htonl(rcv_nxt); 1342df7a3b07SAl Viro *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | 13434de075e0SEric Dumazet tcb->tcp_flags); 1344dfb4b9dcSDavid S. Miller 13451da177e4SLinus Torvalds th->check = 0; 13461da177e4SLinus Torvalds th->urg_ptr = 0; 13471da177e4SLinus Torvalds 134833f5f57eSIlpo Järvinen /* The urg_mode check is necessary during a below snd_una win probe */ 13497691367dSHerbert Xu if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { 13507691367dSHerbert Xu if (before(tp->snd_up, tcb->seq + 0x10000)) { 13511da177e4SLinus Torvalds th->urg_ptr = htons(tp->snd_up - tcb->seq); 13521da177e4SLinus Torvalds th->urg = 1; 13537691367dSHerbert Xu } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { 13540eae88f3SEric Dumazet th->urg_ptr = htons(0xFFFF); 13557691367dSHerbert Xu th->urg = 1; 13567691367dSHerbert Xu } 13571da177e4SLinus Torvalds } 13581da177e4SLinus Torvalds 135951466a75SEric Dumazet skb_shinfo(skb)->gso_type = sk->sk_gso_type; 1360ea1627c2SEric Dumazet if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) { 1361ea1627c2SEric Dumazet th->window = htons(tcp_select_window(sk)); 1362ea1627c2SEric Dumazet tcp_ecn_send(sk, skb, th, tcp_header_size); 1363ea1627c2SEric Dumazet } else { 1364ea1627c2SEric Dumazet /* RFC1323: The window in SYN & SYN/ACK segments 1365ea1627c2SEric Dumazet * is never scaled. 1366ea1627c2SEric Dumazet */ 1367ea1627c2SEric Dumazet th->window = htons(min(tp->rcv_wnd, 65535U)); 1368ea1627c2SEric Dumazet } 1369fa3fe2b1SFlorian Westphal 1370ea66758cSPaolo Abeni tcp_options_write(th, tp, &opts); 1371fa3fe2b1SFlorian Westphal 1372cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 1373cfb6eeb4SYOSHIFUJI Hideaki /* Calculate the MD5 hash, as we have all we need now */ 1374cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 1375aba54656SEric Dumazet sk_gso_disable(sk); 1376bd0388aeSWilliam Allen Simpson tp->af_specific->calc_md5_hash(opts.hash_location, 137739f8e58eSEric Dumazet md5, sk, skb); 1378cfb6eeb4SYOSHIFUJI Hideaki } 1379cfb6eeb4SYOSHIFUJI Hideaki #endif 1380cfb6eeb4SYOSHIFUJI Hideaki 1381331fca43SMartin KaFai Lau /* BPF prog is the last one writing header option */ 1382331fca43SMartin KaFai Lau bpf_skops_write_hdr_opt(sk, skb, NULL, NULL, 0, &opts); 1383331fca43SMartin KaFai Lau 1384dd2e0b86SEric Dumazet INDIRECT_CALL_INET(icsk->icsk_af_ops->send_check, 1385dd2e0b86SEric Dumazet tcp_v6_send_check, tcp_v4_send_check, 1386dd2e0b86SEric Dumazet sk, skb); 13871da177e4SLinus Torvalds 13884de075e0SEric Dumazet if (likely(tcb->tcp_flags & TCPHDR_ACK)) 1389059217c1SNeal Cardwell tcp_event_ack_sent(sk, rcv_nxt); 13901da177e4SLinus Torvalds 1391a44d6eacSMartin KaFai Lau if (skb->len != tcp_header_size) { 1392cf533ea5SEric Dumazet tcp_event_data_sent(tp, sk); 1393a44d6eacSMartin KaFai Lau tp->data_segs_out += tcp_skb_pcount(skb); 1394ba113c3aSWei Wang tp->bytes_sent += skb->len - tcp_header_size; 1395a44d6eacSMartin KaFai Lau } 13961da177e4SLinus Torvalds 1397bd37a088SWei Yongjun if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) 1398aa2ea058STom Herbert TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, 1399aa2ea058STom Herbert tcp_skb_pcount(skb)); 14001da177e4SLinus Torvalds 14012efd055cSMarcelo Ricardo Leitner tp->segs_out += tcp_skb_pcount(skb); 14020ae5b43dSYuchung Cheng skb_set_hash_from_sk(skb, sk); 1403f69ad292SEric Dumazet /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */ 1404cd7d8498SEric Dumazet skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb); 1405f69ad292SEric Dumazet skb_shinfo(skb)->gso_size = tcp_skb_mss(skb); 1406cd7d8498SEric Dumazet 1407d3edd06eSEric Dumazet /* Leave earliest departure time in skb->tstamp (skb->skb_mstamp_ns) */ 1408971f10ecSEric Dumazet 1409971f10ecSEric Dumazet /* Cleanup our debris for IP stacks */ 1410971f10ecSEric Dumazet memset(skb->cb, 0, max(sizeof(struct inet_skb_parm), 1411971f10ecSEric Dumazet sizeof(struct inet6_skb_parm))); 1412971f10ecSEric Dumazet 1413a842fe14SEric Dumazet tcp_add_tx_delay(skb, tp); 1414a842fe14SEric Dumazet 141505e22e83SEric Dumazet err = INDIRECT_CALL_INET(icsk->icsk_af_ops->queue_xmit, 141605e22e83SEric Dumazet inet6_csk_xmit, ip_queue_xmit, 141705e22e83SEric Dumazet sk, skb, &inet->cork.fl); 14187faee5c0SEric Dumazet 14198c72c65bSEric Dumazet if (unlikely(err > 0)) { 14205ee2c941SChristoph Paasch tcp_enter_cwr(sk); 14218c72c65bSEric Dumazet err = net_xmit_eval(err); 14228c72c65bSEric Dumazet } 1423fc225799SEric Dumazet if (!err && oskb) { 1424a7a25630SEric Dumazet tcp_update_skb_after_send(sk, oskb, prior_wstamp); 1425fc225799SEric Dumazet tcp_rate_skb_sent(sk, oskb); 1426fc225799SEric Dumazet } 14278c72c65bSEric Dumazet return err; 14281da177e4SLinus Torvalds } 14291da177e4SLinus Torvalds 14302987babbSYuchung Cheng static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, 14312987babbSYuchung Cheng gfp_t gfp_mask) 14322987babbSYuchung Cheng { 14332987babbSYuchung Cheng return __tcp_transmit_skb(sk, skb, clone_it, gfp_mask, 14342987babbSYuchung Cheng tcp_sk(sk)->rcv_nxt); 14352987babbSYuchung Cheng } 14362987babbSYuchung Cheng 143767edfef7SAndi Kleen /* This routine just queues the buffer for sending. 14381da177e4SLinus Torvalds * 14391da177e4SLinus Torvalds * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, 14401da177e4SLinus Torvalds * otherwise socket can stall. 14411da177e4SLinus Torvalds */ 14421da177e4SLinus Torvalds static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) 14431da177e4SLinus Torvalds { 14441da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 14451da177e4SLinus Torvalds 14461da177e4SLinus Torvalds /* Advance write_seq and place onto the write_queue. */ 14470f317464SEric Dumazet WRITE_ONCE(tp->write_seq, TCP_SKB_CB(skb)->end_seq); 1448f4a775d1SEric Dumazet __skb_header_release(skb); 1449fe067e8aSDavid S. Miller tcp_add_write_queue_tail(sk, skb); 1450ab4e846aSEric Dumazet sk_wmem_queued_add(sk, skb->truesize); 14513ab224beSHideo Aoki sk_mem_charge(sk, skb->truesize); 14521da177e4SLinus Torvalds } 14531da177e4SLinus Torvalds 145467edfef7SAndi Kleen /* Initialize TSO segments for a packet. */ 14555bbb432cSEric Dumazet static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now) 1456f6302d1dSDavid S. Miller { 14574a64fd6cSEric Dumazet if (skb->len <= mss_now) { 1458f6302d1dSDavid S. Miller /* Avoid the costly divide in the normal 1459f6302d1dSDavid S. Miller * non-TSO case. 1460f6302d1dSDavid S. Miller */ 1461cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, 1); 1462f69ad292SEric Dumazet TCP_SKB_CB(skb)->tcp_gso_size = 0; 1463f6302d1dSDavid S. Miller } else { 1464cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now)); 1465f69ad292SEric Dumazet TCP_SKB_CB(skb)->tcp_gso_size = mss_now; 14661da177e4SLinus Torvalds } 14671da177e4SLinus Torvalds } 14681da177e4SLinus Torvalds 1469797108d1SIlpo Järvinen /* Pcount in the middle of the write queue got changed, we need to do various 1470797108d1SIlpo Järvinen * tweaks to fix counters 1471797108d1SIlpo Järvinen */ 1472cf533ea5SEric Dumazet static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) 1473797108d1SIlpo Järvinen { 1474797108d1SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1475797108d1SIlpo Järvinen 1476797108d1SIlpo Järvinen tp->packets_out -= decr; 1477797108d1SIlpo Järvinen 1478797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 1479797108d1SIlpo Järvinen tp->sacked_out -= decr; 1480797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 1481797108d1SIlpo Järvinen tp->retrans_out -= decr; 1482797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 1483797108d1SIlpo Järvinen tp->lost_out -= decr; 1484797108d1SIlpo Järvinen 1485797108d1SIlpo Järvinen /* Reno case is special. Sigh... */ 1486797108d1SIlpo Järvinen if (tcp_is_reno(tp) && decr > 0) 1487797108d1SIlpo Järvinen tp->sacked_out -= min_t(u32, tp->sacked_out, decr); 1488797108d1SIlpo Järvinen 1489797108d1SIlpo Järvinen if (tp->lost_skb_hint && 1490797108d1SIlpo Järvinen before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && 1491713bafeaSYuchung Cheng (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) 1492797108d1SIlpo Järvinen tp->lost_cnt_hint -= decr; 1493797108d1SIlpo Järvinen 1494797108d1SIlpo Järvinen tcp_verify_left_out(tp); 1495797108d1SIlpo Järvinen } 1496797108d1SIlpo Järvinen 14970a2cf20cSSoheil Hassas Yeganeh static bool tcp_has_tx_tstamp(const struct sk_buff *skb) 14980a2cf20cSSoheil Hassas Yeganeh { 14990a2cf20cSSoheil Hassas Yeganeh return TCP_SKB_CB(skb)->txstamp_ack || 15000a2cf20cSSoheil Hassas Yeganeh (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP); 15010a2cf20cSSoheil Hassas Yeganeh } 15020a2cf20cSSoheil Hassas Yeganeh 1503490cc7d0SWillem de Bruijn static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2) 1504490cc7d0SWillem de Bruijn { 1505490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo = skb_shinfo(skb); 1506490cc7d0SWillem de Bruijn 15070a2cf20cSSoheil Hassas Yeganeh if (unlikely(tcp_has_tx_tstamp(skb)) && 1508490cc7d0SWillem de Bruijn !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) { 1509490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo2 = skb_shinfo(skb2); 1510490cc7d0SWillem de Bruijn u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP; 1511490cc7d0SWillem de Bruijn 1512490cc7d0SWillem de Bruijn shinfo->tx_flags &= ~tsflags; 1513490cc7d0SWillem de Bruijn shinfo2->tx_flags |= tsflags; 1514490cc7d0SWillem de Bruijn swap(shinfo->tskey, shinfo2->tskey); 1515b51e13faSMartin KaFai Lau TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack; 1516b51e13faSMartin KaFai Lau TCP_SKB_CB(skb)->txstamp_ack = 0; 1517490cc7d0SWillem de Bruijn } 1518490cc7d0SWillem de Bruijn } 1519490cc7d0SWillem de Bruijn 1520a166140eSMartin KaFai Lau static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2) 1521a166140eSMartin KaFai Lau { 1522a166140eSMartin KaFai Lau TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor; 1523a166140eSMartin KaFai Lau TCP_SKB_CB(skb)->eor = 0; 1524a166140eSMartin KaFai Lau } 1525a166140eSMartin KaFai Lau 152675c119afSEric Dumazet /* Insert buff after skb on the write or rtx queue of sk. */ 152775c119afSEric Dumazet static void tcp_insert_write_queue_after(struct sk_buff *skb, 152875c119afSEric Dumazet struct sk_buff *buff, 152975c119afSEric Dumazet struct sock *sk, 153075c119afSEric Dumazet enum tcp_queue tcp_queue) 153175c119afSEric Dumazet { 153275c119afSEric Dumazet if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE) 153375c119afSEric Dumazet __skb_queue_after(&sk->sk_write_queue, skb, buff); 153475c119afSEric Dumazet else 153575c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, buff); 153675c119afSEric Dumazet } 153775c119afSEric Dumazet 15381da177e4SLinus Torvalds /* Function to create two new TCP segments. Shrinks the given segment 15391da177e4SLinus Torvalds * to the specified size and appends a new segment with the rest of the 15401da177e4SLinus Torvalds * packet to the list. This won't be called frequently, I hope. 15411da177e4SLinus Torvalds * Remember, these are still headerless SKBs at this point. 15421da177e4SLinus Torvalds */ 154375c119afSEric Dumazet int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue, 154475c119afSEric Dumazet struct sk_buff *skb, u32 len, 15456cc55e09SOctavian Purdila unsigned int mss_now, gfp_t gfp) 15461da177e4SLinus Torvalds { 15471da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 15481da177e4SLinus Torvalds struct sk_buff *buff; 1549b4a24397SEric Dumazet int old_factor; 1550b617158dSEric Dumazet long limit; 1551b60b49eaSHerbert Xu int nlen; 15529ce01461SIlpo Järvinen u8 flags; 15531da177e4SLinus Torvalds 15542fceec13SIlpo Järvinen if (WARN_ON(len > skb->len)) 15552fceec13SIlpo Järvinen return -EINVAL; 15566a438bbeSStephen Hemminger 1557b4a24397SEric Dumazet DEBUG_NET_WARN_ON_ONCE(skb_headlen(skb)); 15581da177e4SLinus Torvalds 1559b617158dSEric Dumazet /* tcp_sendmsg() can overshoot sk_wmem_queued by one full size skb. 1560b617158dSEric Dumazet * We need some allowance to not penalize applications setting small 1561b617158dSEric Dumazet * SO_SNDBUF values. 1562b617158dSEric Dumazet * Also allow first and last skb in retransmit queue to be split. 1563b617158dSEric Dumazet */ 15647c4e983cSAlexander Duyck limit = sk->sk_sndbuf + 2 * SKB_TRUESIZE(GSO_LEGACY_MAX_SIZE); 1565b617158dSEric Dumazet if (unlikely((sk->sk_wmem_queued >> 1) > limit && 1566b617158dSEric Dumazet tcp_queue != TCP_FRAG_IN_WRITE_QUEUE && 1567b617158dSEric Dumazet skb != tcp_rtx_queue_head(sk) && 1568b617158dSEric Dumazet skb != tcp_rtx_queue_tail(sk))) { 1569f070ef2aSEric Dumazet NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPWQUEUETOOBIG); 1570f070ef2aSEric Dumazet return -ENOMEM; 1571f070ef2aSEric Dumazet } 1572f070ef2aSEric Dumazet 1573c4777efaSEric Dumazet if (skb_unclone_keeptruesize(skb, gfp)) 15741da177e4SLinus Torvalds return -ENOMEM; 15751da177e4SLinus Torvalds 15761da177e4SLinus Torvalds /* Get a new skb... force flag on. */ 15775882efffSEric Dumazet buff = tcp_stream_alloc_skb(sk, gfp, true); 157851456b29SIan Morris if (!buff) 15791da177e4SLinus Torvalds return -ENOMEM; /* We'll just try again later. */ 158041477662SJakub Kicinski skb_copy_decrypted(buff, skb); 15815a369ca6SPaolo Abeni mptcp_skb_ext_copy(buff, skb); 1582ef5cb973SHerbert Xu 1583ab4e846aSEric Dumazet sk_wmem_queued_add(sk, buff->truesize); 15843ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1585b4a24397SEric Dumazet nlen = skb->len - len; 1586b60b49eaSHerbert Xu buff->truesize += nlen; 1587b60b49eaSHerbert Xu skb->truesize -= nlen; 15881da177e4SLinus Torvalds 15891da177e4SLinus Torvalds /* Correct the sequence numbers. */ 15901da177e4SLinus Torvalds TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 15911da177e4SLinus Torvalds TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 15921da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 15931da177e4SLinus Torvalds 15941da177e4SLinus Torvalds /* PSH and FIN should only be set in the second packet. */ 15954de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 15964de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 15974de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1598e14c3cafSHerbert Xu TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; 1599a166140eSMartin KaFai Lau tcp_skb_fragment_eor(skb, buff); 16001da177e4SLinus Torvalds 16011da177e4SLinus Torvalds skb_split(skb, buff, len); 16021da177e4SLinus Torvalds 1603a1ac9c8aSMartin KaFai Lau skb_set_delivery_time(buff, skb->tstamp, true); 1604490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 16051da177e4SLinus Torvalds 16066475be16SDavid S. Miller old_factor = tcp_skb_pcount(skb); 16076475be16SDavid S. Miller 16081da177e4SLinus Torvalds /* Fix up tso_factor for both original and new SKB. */ 16095bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 16105bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 16111da177e4SLinus Torvalds 1612b9f64820SYuchung Cheng /* Update delivered info for the new segment */ 1613b9f64820SYuchung Cheng TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx; 1614b9f64820SYuchung Cheng 16156475be16SDavid S. Miller /* If this packet has been sent out already, we must 16166475be16SDavid S. Miller * adjust the various packet counters. 16176475be16SDavid S. Miller */ 1618cf0b450cSHerbert Xu if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 16196475be16SDavid S. Miller int diff = old_factor - tcp_skb_pcount(skb) - 16206475be16SDavid S. Miller tcp_skb_pcount(buff); 16211da177e4SLinus Torvalds 1622797108d1SIlpo Järvinen if (diff) 1623797108d1SIlpo Järvinen tcp_adjust_pcount(sk, skb, diff); 16241da177e4SLinus Torvalds } 16251da177e4SLinus Torvalds 16261da177e4SLinus Torvalds /* Link BUFF into the send queue. */ 1627f4a775d1SEric Dumazet __skb_header_release(buff); 162875c119afSEric Dumazet tcp_insert_write_queue_after(skb, buff, sk, tcp_queue); 1629f67971e6SEric Dumazet if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE) 1630e2080072SEric Dumazet list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor); 16311da177e4SLinus Torvalds 16321da177e4SLinus Torvalds return 0; 16331da177e4SLinus Torvalds } 16341da177e4SLinus Torvalds 1635f4d01666SEric Dumazet /* This is similar to __pskb_pull_tail(). The difference is that pulled 1636f4d01666SEric Dumazet * data is not copied, but immediately discarded. 16371da177e4SLinus Torvalds */ 16387162fb24SEric Dumazet static int __pskb_trim_head(struct sk_buff *skb, int len) 16391da177e4SLinus Torvalds { 16407b7fc97aSEric Dumazet struct skb_shared_info *shinfo; 16411da177e4SLinus Torvalds int i, k, eat; 16421da177e4SLinus Torvalds 1643b4a24397SEric Dumazet DEBUG_NET_WARN_ON_ONCE(skb_headlen(skb)); 16441da177e4SLinus Torvalds eat = len; 16451da177e4SLinus Torvalds k = 0; 16467b7fc97aSEric Dumazet shinfo = skb_shinfo(skb); 16477b7fc97aSEric Dumazet for (i = 0; i < shinfo->nr_frags; i++) { 16487b7fc97aSEric Dumazet int size = skb_frag_size(&shinfo->frags[i]); 16499e903e08SEric Dumazet 16509e903e08SEric Dumazet if (size <= eat) { 1651aff65da0SIan Campbell skb_frag_unref(skb, i); 16529e903e08SEric Dumazet eat -= size; 16531da177e4SLinus Torvalds } else { 16547b7fc97aSEric Dumazet shinfo->frags[k] = shinfo->frags[i]; 16551da177e4SLinus Torvalds if (eat) { 1656b54c9d5bSJonathan Lemon skb_frag_off_add(&shinfo->frags[k], eat); 16577b7fc97aSEric Dumazet skb_frag_size_sub(&shinfo->frags[k], eat); 16581da177e4SLinus Torvalds eat = 0; 16591da177e4SLinus Torvalds } 16601da177e4SLinus Torvalds k++; 16611da177e4SLinus Torvalds } 16621da177e4SLinus Torvalds } 16637b7fc97aSEric Dumazet shinfo->nr_frags = k; 16641da177e4SLinus Torvalds 16651da177e4SLinus Torvalds skb->data_len -= len; 16661da177e4SLinus Torvalds skb->len = skb->data_len; 16677162fb24SEric Dumazet return len; 16681da177e4SLinus Torvalds } 16691da177e4SLinus Torvalds 167067edfef7SAndi Kleen /* Remove acked data from a packet in the transmit queue. */ 16711da177e4SLinus Torvalds int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 16721da177e4SLinus Torvalds { 16737162fb24SEric Dumazet u32 delta_truesize; 16747162fb24SEric Dumazet 1675c4777efaSEric Dumazet if (skb_unclone_keeptruesize(skb, GFP_ATOMIC)) 16761da177e4SLinus Torvalds return -ENOMEM; 16771da177e4SLinus Torvalds 16787162fb24SEric Dumazet delta_truesize = __pskb_trim_head(skb, len); 16791da177e4SLinus Torvalds 16801da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq += len; 16811da177e4SLinus Torvalds 16827162fb24SEric Dumazet skb->truesize -= delta_truesize; 1683ab4e846aSEric Dumazet sk_wmem_queued_add(sk, -delta_truesize); 16849b65b17dSTalal Ahmad if (!skb_zcopy_pure(skb)) 16857162fb24SEric Dumazet sk_mem_uncharge(sk, delta_truesize); 16861da177e4SLinus Torvalds 16875b35e1e6SNeal Cardwell /* Any change of skb->len requires recalculation of tso factor. */ 16881da177e4SLinus Torvalds if (tcp_skb_pcount(skb) > 1) 16895bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb)); 16901da177e4SLinus Torvalds 16911da177e4SLinus Torvalds return 0; 16921da177e4SLinus Torvalds } 16931da177e4SLinus Torvalds 16941b63edd6SYuchung Cheng /* Calculate MSS not accounting any TCP options. */ 16951b63edd6SYuchung Cheng static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu) 16965d424d5aSJohn Heffner { 1697cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1698cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 16995d424d5aSJohn Heffner int mss_now; 17005d424d5aSJohn Heffner 17015d424d5aSJohn Heffner /* Calculate base mss without TCP options: 17025d424d5aSJohn Heffner It is MMS_S - sizeof(tcphdr) of rfc1122 17035d424d5aSJohn Heffner */ 17045d424d5aSJohn Heffner mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 17055d424d5aSJohn Heffner 170667469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 170767469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 170867469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 170967469601SEric Dumazet 171067469601SEric Dumazet if (dst && dst_allfrag(dst)) 171167469601SEric Dumazet mss_now -= icsk->icsk_af_ops->net_frag_header_len; 171267469601SEric Dumazet } 171367469601SEric Dumazet 17145d424d5aSJohn Heffner /* Clamp it (mss_clamp does not include tcp options) */ 17155d424d5aSJohn Heffner if (mss_now > tp->rx_opt.mss_clamp) 17165d424d5aSJohn Heffner mss_now = tp->rx_opt.mss_clamp; 17175d424d5aSJohn Heffner 17185d424d5aSJohn Heffner /* Now subtract optional transport overhead */ 17195d424d5aSJohn Heffner mss_now -= icsk->icsk_ext_hdr_len; 17205d424d5aSJohn Heffner 17215d424d5aSJohn Heffner /* Then reserve room for full set of TCP options and 8 bytes of data */ 172278eb166cSKuniyuki Iwashima mss_now = max(mss_now, 172378eb166cSKuniyuki Iwashima READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_snd_mss)); 17245d424d5aSJohn Heffner return mss_now; 17255d424d5aSJohn Heffner } 17265d424d5aSJohn Heffner 17271b63edd6SYuchung Cheng /* Calculate MSS. Not accounting for SACKs here. */ 17281b63edd6SYuchung Cheng int tcp_mtu_to_mss(struct sock *sk, int pmtu) 17291b63edd6SYuchung Cheng { 17301b63edd6SYuchung Cheng /* Subtract TCP options size, not including SACKs */ 17311b63edd6SYuchung Cheng return __tcp_mtu_to_mss(sk, pmtu) - 17321b63edd6SYuchung Cheng (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr)); 17331b63edd6SYuchung Cheng } 1734c7bb4b89SEric Dumazet EXPORT_SYMBOL(tcp_mtu_to_mss); 17351b63edd6SYuchung Cheng 17365d424d5aSJohn Heffner /* Inverse of above */ 173767469601SEric Dumazet int tcp_mss_to_mtu(struct sock *sk, int mss) 17385d424d5aSJohn Heffner { 1739cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1740cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 17415d424d5aSJohn Heffner int mtu; 17425d424d5aSJohn Heffner 17435d424d5aSJohn Heffner mtu = mss + 17445d424d5aSJohn Heffner tp->tcp_header_len + 17455d424d5aSJohn Heffner icsk->icsk_ext_hdr_len + 17465d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 17475d424d5aSJohn Heffner 174867469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 174967469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 175067469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 175167469601SEric Dumazet 175267469601SEric Dumazet if (dst && dst_allfrag(dst)) 175367469601SEric Dumazet mtu += icsk->icsk_af_ops->net_frag_header_len; 175467469601SEric Dumazet } 17555d424d5aSJohn Heffner return mtu; 17565d424d5aSJohn Heffner } 1757556c6b46SNeal Cardwell EXPORT_SYMBOL(tcp_mss_to_mtu); 17585d424d5aSJohn Heffner 175967edfef7SAndi Kleen /* MTU probing init per socket */ 17605d424d5aSJohn Heffner void tcp_mtup_init(struct sock *sk) 17615d424d5aSJohn Heffner { 17625d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 17635d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 1764b0f9ca53SFan Du struct net *net = sock_net(sk); 17655d424d5aSJohn Heffner 1766f47d00e0SKuniyuki Iwashima icsk->icsk_mtup.enabled = READ_ONCE(net->ipv4.sysctl_tcp_mtu_probing) > 1; 17675d424d5aSJohn Heffner icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 17685d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 176988d78bc0SKuniyuki Iwashima icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, READ_ONCE(net->ipv4.sysctl_tcp_base_mss)); 17705d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 177105cbc0dbSFan Du if (icsk->icsk_mtup.enabled) 1772c74df29aSEric Dumazet icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; 17735d424d5aSJohn Heffner } 17744bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_mtup_init); 17755d424d5aSJohn Heffner 17761da177e4SLinus Torvalds /* This function synchronize snd mss to current pmtu/exthdr set. 17771da177e4SLinus Torvalds 17781da177e4SLinus Torvalds tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 17791da177e4SLinus Torvalds for TCP options, but includes only bare TCP header. 17801da177e4SLinus Torvalds 17811da177e4SLinus Torvalds tp->rx_opt.mss_clamp is mss negotiated at connection setup. 1782caa20d9aSStephen Hemminger It is minimum of user_mss and mss received with SYN. 17831da177e4SLinus Torvalds It also does not include TCP options. 17841da177e4SLinus Torvalds 1785d83d8461SArnaldo Carvalho de Melo inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. 17861da177e4SLinus Torvalds 17871da177e4SLinus Torvalds tp->mss_cache is current effective sending mss, including 17881da177e4SLinus Torvalds all tcp options except for SACKs. It is evaluated, 17891da177e4SLinus Torvalds taking into account current pmtu, but never exceeds 17901da177e4SLinus Torvalds tp->rx_opt.mss_clamp. 17911da177e4SLinus Torvalds 17921da177e4SLinus Torvalds NOTE1. rfc1122 clearly states that advertised MSS 17931da177e4SLinus Torvalds DOES NOT include either tcp or ip options. 17941da177e4SLinus Torvalds 1795d83d8461SArnaldo Carvalho de Melo NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache 1796d83d8461SArnaldo Carvalho de Melo are READ ONLY outside this function. --ANK (980731) 17971da177e4SLinus Torvalds */ 17981da177e4SLinus Torvalds unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 17991da177e4SLinus Torvalds { 18001da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1801d83d8461SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 18025d424d5aSJohn Heffner int mss_now; 18031da177e4SLinus Torvalds 18045d424d5aSJohn Heffner if (icsk->icsk_mtup.search_high > pmtu) 18055d424d5aSJohn Heffner icsk->icsk_mtup.search_high = pmtu; 18061da177e4SLinus Torvalds 18075d424d5aSJohn Heffner mss_now = tcp_mtu_to_mss(sk, pmtu); 1808409d22b4SIlpo Järvinen mss_now = tcp_bound_to_half_wnd(tp, mss_now); 18091da177e4SLinus Torvalds 18101da177e4SLinus Torvalds /* And store cached results */ 1811d83d8461SArnaldo Carvalho de Melo icsk->icsk_pmtu_cookie = pmtu; 18125d424d5aSJohn Heffner if (icsk->icsk_mtup.enabled) 18135d424d5aSJohn Heffner mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); 1814c1b4a7e6SDavid S. Miller tp->mss_cache = mss_now; 18151da177e4SLinus Torvalds 18161da177e4SLinus Torvalds return mss_now; 18171da177e4SLinus Torvalds } 18184bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_sync_mss); 18191da177e4SLinus Torvalds 18201da177e4SLinus Torvalds /* Compute the current effective MSS, taking SACKs and IP options, 18211da177e4SLinus Torvalds * and even PMTU discovery events into account. 18221da177e4SLinus Torvalds */ 18230c54b85fSIlpo Järvinen unsigned int tcp_current_mss(struct sock *sk) 18241da177e4SLinus Torvalds { 1825cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1826cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1827c1b4a7e6SDavid S. Miller u32 mss_now; 182895c96174SEric Dumazet unsigned int header_len; 182933ad798cSAdam Langley struct tcp_out_options opts; 183033ad798cSAdam Langley struct tcp_md5sig_key *md5; 18311da177e4SLinus Torvalds 1832c1b4a7e6SDavid S. Miller mss_now = tp->mss_cache; 1833c1b4a7e6SDavid S. Miller 18341da177e4SLinus Torvalds if (dst) { 18351da177e4SLinus Torvalds u32 mtu = dst_mtu(dst); 1836d83d8461SArnaldo Carvalho de Melo if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 18371da177e4SLinus Torvalds mss_now = tcp_sync_mss(sk, mtu); 18381da177e4SLinus Torvalds } 18391da177e4SLinus Torvalds 184033ad798cSAdam Langley header_len = tcp_established_options(sk, NULL, &opts, &md5) + 184133ad798cSAdam Langley sizeof(struct tcphdr); 184233ad798cSAdam Langley /* The mss_cache is sized based on tp->tcp_header_len, which assumes 184333ad798cSAdam Langley * some common options. If this is an odd packet (because we have SACK 184433ad798cSAdam Langley * blocks etc) then our calculated header_len will be different, and 184533ad798cSAdam Langley * we have to adjust mss_now correspondingly */ 184633ad798cSAdam Langley if (header_len != tp->tcp_header_len) { 184733ad798cSAdam Langley int delta = (int) header_len - tp->tcp_header_len; 184833ad798cSAdam Langley mss_now -= delta; 184933ad798cSAdam Langley } 1850cfb6eeb4SYOSHIFUJI Hideaki 18511da177e4SLinus Torvalds return mss_now; 18521da177e4SLinus Torvalds } 18531da177e4SLinus Torvalds 185486fd14adSWeiping Pan /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. 185586fd14adSWeiping Pan * As additional protections, we do not touch cwnd in retransmission phases, 185686fd14adSWeiping Pan * and if application hit its sndbuf limit recently. 185786fd14adSWeiping Pan */ 185886fd14adSWeiping Pan static void tcp_cwnd_application_limited(struct sock *sk) 1859a762a980SDavid S. Miller { 18609e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1861a762a980SDavid S. Miller 186286fd14adSWeiping Pan if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && 186386fd14adSWeiping Pan sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 186486fd14adSWeiping Pan /* Limited by application or receiver window. */ 186586fd14adSWeiping Pan u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); 186686fd14adSWeiping Pan u32 win_used = max(tp->snd_cwnd_used, init_win); 186740570375SEric Dumazet if (win_used < tcp_snd_cwnd(tp)) { 186886fd14adSWeiping Pan tp->snd_ssthresh = tcp_current_ssthresh(sk); 186940570375SEric Dumazet tcp_snd_cwnd_set(tp, (tcp_snd_cwnd(tp) + win_used) >> 1); 187086fd14adSWeiping Pan } 187186fd14adSWeiping Pan tp->snd_cwnd_used = 0; 187286fd14adSWeiping Pan } 1873c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 187486fd14adSWeiping Pan } 187586fd14adSWeiping Pan 1876ca8a2263SNeal Cardwell static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited) 1877a762a980SDavid S. Miller { 18781b1fc3fdSWei Wang const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 1879a762a980SDavid S. Miller struct tcp_sock *tp = tcp_sk(sk); 1880a762a980SDavid S. Miller 1881f4ce91ceSNeal Cardwell /* Track the strongest available signal of the degree to which the cwnd 1882f4ce91ceSNeal Cardwell * is fully utilized. If cwnd-limited then remember that fact for the 1883f4ce91ceSNeal Cardwell * current window. If not cwnd-limited then track the maximum number of 1884f4ce91ceSNeal Cardwell * outstanding packets in the current window. (If cwnd-limited then we 1885f4ce91ceSNeal Cardwell * chose to not update tp->max_packets_out to avoid an extra else 1886f4ce91ceSNeal Cardwell * clause with no functional impact.) 1887ca8a2263SNeal Cardwell */ 1888f4ce91ceSNeal Cardwell if (!before(tp->snd_una, tp->cwnd_usage_seq) || 1889f4ce91ceSNeal Cardwell is_cwnd_limited || 1890f4ce91ceSNeal Cardwell (!tp->is_cwnd_limited && 1891f4ce91ceSNeal Cardwell tp->packets_out > tp->max_packets_out)) { 1892ca8a2263SNeal Cardwell tp->is_cwnd_limited = is_cwnd_limited; 1893f4ce91ceSNeal Cardwell tp->max_packets_out = tp->packets_out; 1894f4ce91ceSNeal Cardwell tp->cwnd_usage_seq = tp->snd_nxt; 1895ca8a2263SNeal Cardwell } 1896e114a710SEric Dumazet 189724901551SEric Dumazet if (tcp_is_cwnd_limited(sk)) { 1898a762a980SDavid S. Miller /* Network is feed fully. */ 1899a762a980SDavid S. Miller tp->snd_cwnd_used = 0; 1900c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 1901a762a980SDavid S. Miller } else { 1902a762a980SDavid S. Miller /* Network starves. */ 1903a762a980SDavid S. Miller if (tp->packets_out > tp->snd_cwnd_used) 1904a762a980SDavid S. Miller tp->snd_cwnd_used = tp->packets_out; 1905a762a980SDavid S. Miller 19064845b571SKuniyuki Iwashima if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle) && 1907c2203cf7SEric Dumazet (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto && 19081b1fc3fdSWei Wang !ca_ops->cong_control) 1909a762a980SDavid S. Miller tcp_cwnd_application_limited(sk); 1910b0f71bd3SFrancis Yan 1911b0f71bd3SFrancis Yan /* The following conditions together indicate the starvation 1912b0f71bd3SFrancis Yan * is caused by insufficient sender buffer: 1913b0f71bd3SFrancis Yan * 1) just sent some data (see tcp_write_xmit) 1914b0f71bd3SFrancis Yan * 2) not cwnd limited (this else condition) 191575c119afSEric Dumazet * 3) no more data to send (tcp_write_queue_empty()) 1916b0f71bd3SFrancis Yan * 4) application is hitting buffer limit (SOCK_NOSPACE) 1917b0f71bd3SFrancis Yan */ 191875c119afSEric Dumazet if (tcp_write_queue_empty(sk) && sk->sk_socket && 1919b0f71bd3SFrancis Yan test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) && 1920b0f71bd3SFrancis Yan (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) 1921b0f71bd3SFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED); 1922a762a980SDavid S. Miller } 1923a762a980SDavid S. Miller } 1924a762a980SDavid S. Miller 1925d4589926SEric Dumazet /* Minshall's variant of the Nagle send check. */ 1926d4589926SEric Dumazet static bool tcp_minshall_check(const struct tcp_sock *tp) 1927d4589926SEric Dumazet { 1928d4589926SEric Dumazet return after(tp->snd_sml, tp->snd_una) && 1929d4589926SEric Dumazet !after(tp->snd_sml, tp->snd_nxt); 1930d4589926SEric Dumazet } 1931d4589926SEric Dumazet 1932d4589926SEric Dumazet /* Update snd_sml if this skb is under mss 1933d4589926SEric Dumazet * Note that a TSO packet might end with a sub-mss segment 1934d4589926SEric Dumazet * The test is really : 1935d4589926SEric Dumazet * if ((skb->len % mss) != 0) 1936d4589926SEric Dumazet * tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1937d4589926SEric Dumazet * But we can avoid doing the divide again given we already have 1938d4589926SEric Dumazet * skb_pcount = skb->len / mss_now 19390e3a4803SIlpo Järvinen */ 1940d4589926SEric Dumazet static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now, 1941d4589926SEric Dumazet const struct sk_buff *skb) 1942d4589926SEric Dumazet { 1943d4589926SEric Dumazet if (skb->len < tcp_skb_pcount(skb) * mss_now) 1944d4589926SEric Dumazet tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1945d4589926SEric Dumazet } 1946d4589926SEric Dumazet 1947d4589926SEric Dumazet /* Return false, if packet can be sent now without violation Nagle's rules: 1948d4589926SEric Dumazet * 1. It is full sized. (provided by caller in %partial bool) 1949d4589926SEric Dumazet * 2. Or it contains FIN. (already checked by caller) 1950d4589926SEric Dumazet * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. 1951d4589926SEric Dumazet * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 1952d4589926SEric Dumazet * With Minshall's modification: all sent small packets are ACKed. 1953d4589926SEric Dumazet */ 1954d4589926SEric Dumazet static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp, 1955cc93fc51SPeter Pan(潘卫平) int nonagle) 1956d4589926SEric Dumazet { 1957d4589926SEric Dumazet return partial && 1958d4589926SEric Dumazet ((nonagle & TCP_NAGLE_CORK) || 1959d4589926SEric Dumazet (!nonagle && tp->packets_out && tcp_minshall_check(tp))); 1960d4589926SEric Dumazet } 1961605ad7f1SEric Dumazet 1962605ad7f1SEric Dumazet /* Return how many segs we'd like on a TSO packet, 196365466904SEric Dumazet * depending on current pacing rate, and how close the peer is. 196465466904SEric Dumazet * 196565466904SEric Dumazet * Rationale is: 196665466904SEric Dumazet * - For close peers, we rather send bigger packets to reduce 196765466904SEric Dumazet * cpu costs, because occasional losses will be repaired fast. 196865466904SEric Dumazet * - For long distance/rtt flows, we would like to get ACK clocking 196965466904SEric Dumazet * with 1 ACK per ms. 197065466904SEric Dumazet * 197165466904SEric Dumazet * Use min_rtt to help adapt TSO burst size, with smaller min_rtt resulting 197265466904SEric Dumazet * in bigger TSO bursts. We we cut the RTT-based allowance in half 197365466904SEric Dumazet * for every 2^9 usec (aka 512 us) of RTT, so that the RTT-based allowance 197465466904SEric Dumazet * is below 1500 bytes after 6 * ~500 usec = 3ms. 1975605ad7f1SEric Dumazet */ 1976dcb8c9b4SEric Dumazet static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now, 19771b3878caSNeal Cardwell int min_tso_segs) 1978605ad7f1SEric Dumazet { 197965466904SEric Dumazet unsigned long bytes; 198065466904SEric Dumazet u32 r; 1981605ad7f1SEric Dumazet 198265466904SEric Dumazet bytes = sk->sk_pacing_rate >> READ_ONCE(sk->sk_pacing_shift); 1983605ad7f1SEric Dumazet 19842455e61bSKuniyuki Iwashima r = tcp_min_rtt(tcp_sk(sk)) >> READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_rtt_log); 198565466904SEric Dumazet if (r < BITS_PER_TYPE(sk->sk_gso_max_size)) 198665466904SEric Dumazet bytes += sk->sk_gso_max_size >> r; 1987605ad7f1SEric Dumazet 198865466904SEric Dumazet bytes = min_t(unsigned long, bytes, sk->sk_gso_max_size); 198965466904SEric Dumazet 199065466904SEric Dumazet return max_t(u32, bytes / mss_now, min_tso_segs); 1991605ad7f1SEric Dumazet } 1992605ad7f1SEric Dumazet 1993ed6e7268SNeal Cardwell /* Return the number of segments we want in the skb we are transmitting. 1994ed6e7268SNeal Cardwell * See if congestion control module wants to decide; otherwise, autosize. 1995ed6e7268SNeal Cardwell */ 1996ed6e7268SNeal Cardwell static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now) 1997ed6e7268SNeal Cardwell { 1998ed6e7268SNeal Cardwell const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 1999dcb8c9b4SEric Dumazet u32 min_tso, tso_segs; 2000ed6e7268SNeal Cardwell 2001dcb8c9b4SEric Dumazet min_tso = ca_ops->min_tso_segs ? 2002dcb8c9b4SEric Dumazet ca_ops->min_tso_segs(sk) : 2003e0bb4ab9SKuniyuki Iwashima READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs); 2004dcb8c9b4SEric Dumazet 2005dcb8c9b4SEric Dumazet tso_segs = tcp_tso_autosize(sk, mss_now, min_tso); 2006350c9f48SEric Dumazet return min_t(u32, tso_segs, sk->sk_gso_max_segs); 2007ed6e7268SNeal Cardwell } 2008ed6e7268SNeal Cardwell 2009d4589926SEric Dumazet /* Returns the portion of skb which can be sent right away */ 2010d4589926SEric Dumazet static unsigned int tcp_mss_split_point(const struct sock *sk, 2011d4589926SEric Dumazet const struct sk_buff *skb, 2012d4589926SEric Dumazet unsigned int mss_now, 2013d4589926SEric Dumazet unsigned int max_segs, 2014d4589926SEric Dumazet int nonagle) 2015c1b4a7e6SDavid S. Miller { 2016cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 2017d4589926SEric Dumazet u32 partial, needed, window, max_len; 2018c1b4a7e6SDavid S. Miller 201990840defSIlpo Järvinen window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 20201485348dSBen Hutchings max_len = mss_now * max_segs; 20210e3a4803SIlpo Järvinen 20221485348dSBen Hutchings if (likely(max_len <= window && skb != tcp_write_queue_tail(sk))) 20231485348dSBen Hutchings return max_len; 20240e3a4803SIlpo Järvinen 20255ea3a748SIlpo Järvinen needed = min(skb->len, window); 20265ea3a748SIlpo Järvinen 20271485348dSBen Hutchings if (max_len <= needed) 20281485348dSBen Hutchings return max_len; 20290e3a4803SIlpo Järvinen 2030d4589926SEric Dumazet partial = needed % mss_now; 2031d4589926SEric Dumazet /* If last segment is not a full MSS, check if Nagle rules allow us 2032d4589926SEric Dumazet * to include this last segment in this skb. 2033d4589926SEric Dumazet * Otherwise, we'll split the skb at last MSS boundary 2034d4589926SEric Dumazet */ 2035cc93fc51SPeter Pan(潘卫平) if (tcp_nagle_check(partial != 0, tp, nonagle)) 2036d4589926SEric Dumazet return needed - partial; 2037d4589926SEric Dumazet 2038d4589926SEric Dumazet return needed; 2039c1b4a7e6SDavid S. Miller } 2040c1b4a7e6SDavid S. Miller 2041c1b4a7e6SDavid S. Miller /* Can at least one segment of SKB be sent right now, according to the 2042c1b4a7e6SDavid S. Miller * congestion window rules? If so, return how many segments are allowed. 2043c1b4a7e6SDavid S. Miller */ 2044cf533ea5SEric Dumazet static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, 2045cf533ea5SEric Dumazet const struct sk_buff *skb) 2046c1b4a7e6SDavid S. Miller { 2047d649a7a8SEric Dumazet u32 in_flight, cwnd, halfcwnd; 2048c1b4a7e6SDavid S. Miller 2049c1b4a7e6SDavid S. Miller /* Don't be strict about the congestion window for the final FIN. */ 20504de075e0SEric Dumazet if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && 20514de075e0SEric Dumazet tcp_skb_pcount(skb) == 1) 2052c1b4a7e6SDavid S. Miller return 1; 2053c1b4a7e6SDavid S. Miller 2054c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 205540570375SEric Dumazet cwnd = tcp_snd_cwnd(tp); 2056d649a7a8SEric Dumazet if (in_flight >= cwnd) 2057c1b4a7e6SDavid S. Miller return 0; 2058d649a7a8SEric Dumazet 2059d649a7a8SEric Dumazet /* For better scheduling, ensure we have at least 2060d649a7a8SEric Dumazet * 2 GSO packets in flight. 2061d649a7a8SEric Dumazet */ 2062d649a7a8SEric Dumazet halfcwnd = max(cwnd >> 1, 1U); 2063d649a7a8SEric Dumazet return min(halfcwnd, cwnd - in_flight); 2064c1b4a7e6SDavid S. Miller } 2065c1b4a7e6SDavid S. Miller 2066b595076aSUwe Kleine-König /* Initialize TSO state of a skb. 206767edfef7SAndi Kleen * This must be invoked the first time we consider transmitting 2068c1b4a7e6SDavid S. Miller * SKB onto the wire. 2069c1b4a7e6SDavid S. Miller */ 20705bbb432cSEric Dumazet static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now) 2071c1b4a7e6SDavid S. Miller { 2072c1b4a7e6SDavid S. Miller int tso_segs = tcp_skb_pcount(skb); 2073c1b4a7e6SDavid S. Miller 2074f8269a49SIlpo Järvinen if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { 20755bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 2076c1b4a7e6SDavid S. Miller tso_segs = tcp_skb_pcount(skb); 2077c1b4a7e6SDavid S. Miller } 2078c1b4a7e6SDavid S. Miller return tso_segs; 2079c1b4a7e6SDavid S. Miller } 2080c1b4a7e6SDavid S. Miller 2081c1b4a7e6SDavid S. Miller 2082a2a385d6SEric Dumazet /* Return true if the Nagle test allows this packet to be 2083c1b4a7e6SDavid S. Miller * sent now. 2084c1b4a7e6SDavid S. Miller */ 2085a2a385d6SEric Dumazet static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, 2086c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 2087c1b4a7e6SDavid S. Miller { 2088c1b4a7e6SDavid S. Miller /* Nagle rule does not apply to frames, which sit in the middle of the 2089c1b4a7e6SDavid S. Miller * write_queue (they have no chances to get new data). 2090c1b4a7e6SDavid S. Miller * 2091c1b4a7e6SDavid S. Miller * This is implemented in the callers, where they modify the 'nonagle' 2092c1b4a7e6SDavid S. Miller * argument based upon the location of SKB in the send queue. 2093c1b4a7e6SDavid S. Miller */ 2094c1b4a7e6SDavid S. Miller if (nonagle & TCP_NAGLE_PUSH) 2095a2a385d6SEric Dumazet return true; 2096c1b4a7e6SDavid S. Miller 20979b44190dSYuchung Cheng /* Don't use the nagle rule for urgent data (or for the final FIN). */ 20989b44190dSYuchung Cheng if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) 2099a2a385d6SEric Dumazet return true; 2100c1b4a7e6SDavid S. Miller 2101cc93fc51SPeter Pan(潘卫平) if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle)) 2102a2a385d6SEric Dumazet return true; 2103c1b4a7e6SDavid S. Miller 2104a2a385d6SEric Dumazet return false; 2105c1b4a7e6SDavid S. Miller } 2106c1b4a7e6SDavid S. Miller 2107c1b4a7e6SDavid S. Miller /* Does at least the first segment of SKB fit into the send window? */ 2108a2a385d6SEric Dumazet static bool tcp_snd_wnd_test(const struct tcp_sock *tp, 2109a2a385d6SEric Dumazet const struct sk_buff *skb, 2110056834d9SIlpo Järvinen unsigned int cur_mss) 2111c1b4a7e6SDavid S. Miller { 2112c1b4a7e6SDavid S. Miller u32 end_seq = TCP_SKB_CB(skb)->end_seq; 2113c1b4a7e6SDavid S. Miller 2114c1b4a7e6SDavid S. Miller if (skb->len > cur_mss) 2115c1b4a7e6SDavid S. Miller end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 2116c1b4a7e6SDavid S. Miller 211790840defSIlpo Järvinen return !after(end_seq, tcp_wnd_end(tp)); 2118c1b4a7e6SDavid S. Miller } 2119c1b4a7e6SDavid S. Miller 2120c1b4a7e6SDavid S. Miller /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 2121c1b4a7e6SDavid S. Miller * which is put after SKB on the list. It is very much like 2122c1b4a7e6SDavid S. Miller * tcp_fragment() except that it may make several kinds of assumptions 2123c1b4a7e6SDavid S. Miller * in order to speed up the splitting operation. In particular, we 2124c1b4a7e6SDavid S. Miller * know that all the data is in scatter-gather pages, and that the 2125c1b4a7e6SDavid S. Miller * packet has never been sent out before (and thus is not cloned). 2126c1b4a7e6SDavid S. Miller */ 212756483341SEric Dumazet static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, 2128c4ead4c5SEric Dumazet unsigned int mss_now, gfp_t gfp) 2129c1b4a7e6SDavid S. Miller { 2130c1b4a7e6SDavid S. Miller int nlen = skb->len - len; 213156483341SEric Dumazet struct sk_buff *buff; 21329ce01461SIlpo Järvinen u8 flags; 2133c1b4a7e6SDavid S. Miller 2134c1b4a7e6SDavid S. Miller /* All of a TSO frame must be composed of paged data. */ 2135b4a24397SEric Dumazet DEBUG_NET_WARN_ON_ONCE(skb->len != skb->data_len); 2136c1b4a7e6SDavid S. Miller 21375882efffSEric Dumazet buff = tcp_stream_alloc_skb(sk, gfp, true); 213851456b29SIan Morris if (unlikely(!buff)) 2139c1b4a7e6SDavid S. Miller return -ENOMEM; 214041477662SJakub Kicinski skb_copy_decrypted(buff, skb); 21415a369ca6SPaolo Abeni mptcp_skb_ext_copy(buff, skb); 2142c1b4a7e6SDavid S. Miller 2143ab4e846aSEric Dumazet sk_wmem_queued_add(sk, buff->truesize); 21443ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 2145b60b49eaSHerbert Xu buff->truesize += nlen; 2146c1b4a7e6SDavid S. Miller skb->truesize -= nlen; 2147c1b4a7e6SDavid S. Miller 2148c1b4a7e6SDavid S. Miller /* Correct the sequence numbers. */ 2149c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 2150c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 2151c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 2152c1b4a7e6SDavid S. Miller 2153c1b4a7e6SDavid S. Miller /* PSH and FIN should only be set in the second packet. */ 21544de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 21554de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 21564de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 2157c1b4a7e6SDavid S. Miller 2158a166140eSMartin KaFai Lau tcp_skb_fragment_eor(skb, buff); 2159a166140eSMartin KaFai Lau 2160c1b4a7e6SDavid S. Miller skb_split(skb, buff, len); 2161490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 2162c1b4a7e6SDavid S. Miller 2163c1b4a7e6SDavid S. Miller /* Fix up tso_factor for both original and new SKB. */ 21645bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 21655bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 2166c1b4a7e6SDavid S. Miller 2167c1b4a7e6SDavid S. Miller /* Link BUFF into the send queue. */ 2168f4a775d1SEric Dumazet __skb_header_release(buff); 216956483341SEric Dumazet tcp_insert_write_queue_after(skb, buff, sk, TCP_FRAG_IN_WRITE_QUEUE); 2170c1b4a7e6SDavid S. Miller 2171c1b4a7e6SDavid S. Miller return 0; 2172c1b4a7e6SDavid S. Miller } 2173c1b4a7e6SDavid S. Miller 2174c1b4a7e6SDavid S. Miller /* Try to defer sending, if possible, in order to minimize the amount 2175c1b4a7e6SDavid S. Miller * of TSO splitting we do. View it as a kind of TSO Nagle test. 2176c1b4a7e6SDavid S. Miller * 2177c1b4a7e6SDavid S. Miller * This algorithm is from John Heffner. 2178c1b4a7e6SDavid S. Miller */ 2179ca8a2263SNeal Cardwell static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb, 2180f9bfe4e6SEric Dumazet bool *is_cwnd_limited, 2181f9bfe4e6SEric Dumazet bool *is_rwnd_limited, 2182f9bfe4e6SEric Dumazet u32 max_segs) 2183c1b4a7e6SDavid S. Miller { 21846687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 2185f1c6ea38SEric Dumazet u32 send_win, cong_win, limit, in_flight; 218650c8339eSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 218750c8339eSEric Dumazet struct sk_buff *head; 2188ad9f4f50SEric Dumazet int win_divisor; 2189f1c6ea38SEric Dumazet s64 delta; 2190c1b4a7e6SDavid S. Miller 219199d7662aSEric Dumazet if (icsk->icsk_ca_state >= TCP_CA_Recovery) 2192ae8064acSJohn Heffner goto send_now; 2193ae8064acSJohn Heffner 21945f852eb5SEric Dumazet /* Avoid bursty behavior by allowing defer 2195a682850aSEric Dumazet * only if the last write was recent (1 ms). 2196a682850aSEric Dumazet * Note that tp->tcp_wstamp_ns can be in the future if we have 2197a682850aSEric Dumazet * packets waiting in a qdisc or device for EDT delivery. 21985f852eb5SEric Dumazet */ 2199a682850aSEric Dumazet delta = tp->tcp_clock_cache - tp->tcp_wstamp_ns - NSEC_PER_MSEC; 2200a682850aSEric Dumazet if (delta > 0) 2201ae8064acSJohn Heffner goto send_now; 2202908a75c1SDavid S. Miller 2203c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 2204c1b4a7e6SDavid S. Miller 2205c8c9aeb5SStefano Brivio BUG_ON(tcp_skb_pcount(skb) <= 1); 220640570375SEric Dumazet BUG_ON(tcp_snd_cwnd(tp) <= in_flight); 2207c1b4a7e6SDavid S. Miller 220890840defSIlpo Järvinen send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 2209c1b4a7e6SDavid S. Miller 2210c1b4a7e6SDavid S. Miller /* From in_flight test above, we know that cwnd > in_flight. */ 221140570375SEric Dumazet cong_win = (tcp_snd_cwnd(tp) - in_flight) * tp->mss_cache; 2212c1b4a7e6SDavid S. Miller 2213c1b4a7e6SDavid S. Miller limit = min(send_win, cong_win); 2214c1b4a7e6SDavid S. Miller 2215ba244fe9SDavid S. Miller /* If a full-sized TSO skb can be sent, do it. */ 2216605ad7f1SEric Dumazet if (limit >= max_segs * tp->mss_cache) 2217ae8064acSJohn Heffner goto send_now; 2218ba244fe9SDavid S. Miller 221962ad2761SIlpo Järvinen /* Middle in queue won't get any more data, full sendable already? */ 222062ad2761SIlpo Järvinen if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) 222162ad2761SIlpo Järvinen goto send_now; 222262ad2761SIlpo Järvinen 22235bbcc0f5SLinus Torvalds win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor); 2224ad9f4f50SEric Dumazet if (win_divisor) { 222540570375SEric Dumazet u32 chunk = min(tp->snd_wnd, tcp_snd_cwnd(tp) * tp->mss_cache); 2226c1b4a7e6SDavid S. Miller 2227c1b4a7e6SDavid S. Miller /* If at least some fraction of a window is available, 2228c1b4a7e6SDavid S. Miller * just use it. 2229c1b4a7e6SDavid S. Miller */ 2230ad9f4f50SEric Dumazet chunk /= win_divisor; 2231c1b4a7e6SDavid S. Miller if (limit >= chunk) 2232ae8064acSJohn Heffner goto send_now; 2233c1b4a7e6SDavid S. Miller } else { 2234c1b4a7e6SDavid S. Miller /* Different approach, try not to defer past a single 2235c1b4a7e6SDavid S. Miller * ACK. Receiver should ACK every other full sized 2236c1b4a7e6SDavid S. Miller * frame, so if we have space for more than 3 frames 2237c1b4a7e6SDavid S. Miller * then send now. 2238c1b4a7e6SDavid S. Miller */ 22396b5a5c0dSNeal Cardwell if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) 2240ae8064acSJohn Heffner goto send_now; 2241c1b4a7e6SDavid S. Miller } 2242c1b4a7e6SDavid S. Miller 224375c119afSEric Dumazet /* TODO : use tsorted_sent_queue ? */ 224475c119afSEric Dumazet head = tcp_rtx_queue_head(sk); 224575c119afSEric Dumazet if (!head) 224675c119afSEric Dumazet goto send_now; 2247f1c6ea38SEric Dumazet delta = tp->tcp_clock_cache - head->tstamp; 224850c8339eSEric Dumazet /* If next ACK is likely to come too late (half srtt), do not defer */ 2249f1c6ea38SEric Dumazet if ((s64)(delta - (u64)NSEC_PER_USEC * (tp->srtt_us >> 4)) < 0) 225050c8339eSEric Dumazet goto send_now; 225150c8339eSEric Dumazet 2252f9bfe4e6SEric Dumazet /* Ok, it looks like it is advisable to defer. 2253f9bfe4e6SEric Dumazet * Three cases are tracked : 2254f9bfe4e6SEric Dumazet * 1) We are cwnd-limited 2255f9bfe4e6SEric Dumazet * 2) We are rwnd-limited 2256f9bfe4e6SEric Dumazet * 3) We are application limited. 2257f9bfe4e6SEric Dumazet */ 2258f9bfe4e6SEric Dumazet if (cong_win < send_win) { 2259f9bfe4e6SEric Dumazet if (cong_win <= skb->len) { 2260ca8a2263SNeal Cardwell *is_cwnd_limited = true; 2261f9bfe4e6SEric Dumazet return true; 2262f9bfe4e6SEric Dumazet } 2263f9bfe4e6SEric Dumazet } else { 2264f9bfe4e6SEric Dumazet if (send_win <= skb->len) { 2265f9bfe4e6SEric Dumazet *is_rwnd_limited = true; 2266f9bfe4e6SEric Dumazet return true; 2267f9bfe4e6SEric Dumazet } 2268f9bfe4e6SEric Dumazet } 2269f9bfe4e6SEric Dumazet 2270f9bfe4e6SEric Dumazet /* If this packet won't get more data, do not wait. */ 2271d8ed257fSEric Dumazet if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) || 2272d8ed257fSEric Dumazet TCP_SKB_CB(skb)->eor) 2273f9bfe4e6SEric Dumazet goto send_now; 2274ca8a2263SNeal Cardwell 2275a2a385d6SEric Dumazet return true; 2276ae8064acSJohn Heffner 2277ae8064acSJohn Heffner send_now: 2278a2a385d6SEric Dumazet return false; 2279c1b4a7e6SDavid S. Miller } 2280c1b4a7e6SDavid S. Miller 228105cbc0dbSFan Du static inline void tcp_mtu_check_reprobe(struct sock *sk) 228205cbc0dbSFan Du { 228305cbc0dbSFan Du struct inet_connection_sock *icsk = inet_csk(sk); 228405cbc0dbSFan Du struct tcp_sock *tp = tcp_sk(sk); 228505cbc0dbSFan Du struct net *net = sock_net(sk); 228605cbc0dbSFan Du u32 interval; 228705cbc0dbSFan Du s32 delta; 228805cbc0dbSFan Du 22892a85388fSKuniyuki Iwashima interval = READ_ONCE(net->ipv4.sysctl_tcp_probe_interval); 2290c74df29aSEric Dumazet delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp; 229105cbc0dbSFan Du if (unlikely(delta >= interval * HZ)) { 229205cbc0dbSFan Du int mss = tcp_current_mss(sk); 229305cbc0dbSFan Du 229405cbc0dbSFan Du /* Update current search range */ 229505cbc0dbSFan Du icsk->icsk_mtup.probe_size = 0; 229605cbc0dbSFan Du icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + 229705cbc0dbSFan Du sizeof(struct tcphdr) + 229805cbc0dbSFan Du icsk->icsk_af_ops->net_header_len; 229905cbc0dbSFan Du icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss); 230005cbc0dbSFan Du 230105cbc0dbSFan Du /* Update probe time stamp */ 2302c74df29aSEric Dumazet icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; 230305cbc0dbSFan Du } 230405cbc0dbSFan Du } 230505cbc0dbSFan Du 2306808cf9e3SIlya Lesokhin static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len) 2307808cf9e3SIlya Lesokhin { 2308808cf9e3SIlya Lesokhin struct sk_buff *skb, *next; 2309808cf9e3SIlya Lesokhin 2310808cf9e3SIlya Lesokhin skb = tcp_send_head(sk); 2311808cf9e3SIlya Lesokhin tcp_for_write_queue_from_safe(skb, next, sk) { 2312808cf9e3SIlya Lesokhin if (len <= skb->len) 2313808cf9e3SIlya Lesokhin break; 2314808cf9e3SIlya Lesokhin 23159b65b17dSTalal Ahmad if (unlikely(TCP_SKB_CB(skb)->eor) || 23169b65b17dSTalal Ahmad tcp_has_tx_tstamp(skb) || 23179b65b17dSTalal Ahmad !skb_pure_zcopy_same(skb, next)) 2318808cf9e3SIlya Lesokhin return false; 2319808cf9e3SIlya Lesokhin 2320808cf9e3SIlya Lesokhin len -= skb->len; 2321808cf9e3SIlya Lesokhin } 2322808cf9e3SIlya Lesokhin 2323808cf9e3SIlya Lesokhin return true; 2324808cf9e3SIlya Lesokhin } 2325808cf9e3SIlya Lesokhin 232673601329SEric Dumazet static int tcp_clone_payload(struct sock *sk, struct sk_buff *to, 232773601329SEric Dumazet int probe_size) 232873601329SEric Dumazet { 232973601329SEric Dumazet skb_frag_t *lastfrag = NULL, *fragto = skb_shinfo(to)->frags; 233073601329SEric Dumazet int i, todo, len = 0, nr_frags = 0; 233173601329SEric Dumazet const struct sk_buff *skb; 233273601329SEric Dumazet 233373601329SEric Dumazet if (!sk_wmem_schedule(sk, to->truesize + probe_size)) 233473601329SEric Dumazet return -ENOMEM; 233573601329SEric Dumazet 233673601329SEric Dumazet skb_queue_walk(&sk->sk_write_queue, skb) { 233773601329SEric Dumazet const skb_frag_t *fragfrom = skb_shinfo(skb)->frags; 233873601329SEric Dumazet 233973601329SEric Dumazet if (skb_headlen(skb)) 234073601329SEric Dumazet return -EINVAL; 234173601329SEric Dumazet 234273601329SEric Dumazet for (i = 0; i < skb_shinfo(skb)->nr_frags; i++, fragfrom++) { 234373601329SEric Dumazet if (len >= probe_size) 234473601329SEric Dumazet goto commit; 234573601329SEric Dumazet todo = min_t(int, skb_frag_size(fragfrom), 234673601329SEric Dumazet probe_size - len); 234773601329SEric Dumazet len += todo; 234873601329SEric Dumazet if (lastfrag && 234973601329SEric Dumazet skb_frag_page(fragfrom) == skb_frag_page(lastfrag) && 235073601329SEric Dumazet skb_frag_off(fragfrom) == skb_frag_off(lastfrag) + 235173601329SEric Dumazet skb_frag_size(lastfrag)) { 235273601329SEric Dumazet skb_frag_size_add(lastfrag, todo); 235373601329SEric Dumazet continue; 235473601329SEric Dumazet } 235573601329SEric Dumazet if (unlikely(nr_frags == MAX_SKB_FRAGS)) 235673601329SEric Dumazet return -E2BIG; 235773601329SEric Dumazet skb_frag_page_copy(fragto, fragfrom); 235873601329SEric Dumazet skb_frag_off_copy(fragto, fragfrom); 235973601329SEric Dumazet skb_frag_size_set(fragto, todo); 236073601329SEric Dumazet nr_frags++; 236173601329SEric Dumazet lastfrag = fragto++; 236273601329SEric Dumazet } 236373601329SEric Dumazet } 236473601329SEric Dumazet commit: 236573601329SEric Dumazet WARN_ON_ONCE(len != probe_size); 236673601329SEric Dumazet for (i = 0; i < nr_frags; i++) 236773601329SEric Dumazet skb_frag_ref(to, i); 236873601329SEric Dumazet 236973601329SEric Dumazet skb_shinfo(to)->nr_frags = nr_frags; 237073601329SEric Dumazet to->truesize += probe_size; 237173601329SEric Dumazet to->len += probe_size; 237273601329SEric Dumazet to->data_len += probe_size; 237373601329SEric Dumazet __skb_header_release(to); 237473601329SEric Dumazet return 0; 237573601329SEric Dumazet } 237673601329SEric Dumazet 23775d424d5aSJohn Heffner /* Create a new MTU probe if we are ready. 237867edfef7SAndi Kleen * MTU probe is regularly attempting to increase the path MTU by 237967edfef7SAndi Kleen * deliberately sending larger packets. This discovers routing 238067edfef7SAndi Kleen * changes resulting in larger path MTUs. 238167edfef7SAndi Kleen * 23825d424d5aSJohn Heffner * Returns 0 if we should wait to probe (no cwnd available), 23835d424d5aSJohn Heffner * 1 if a probe was sent, 2384056834d9SIlpo Järvinen * -1 otherwise 2385056834d9SIlpo Järvinen */ 23865d424d5aSJohn Heffner static int tcp_mtu_probe(struct sock *sk) 23875d424d5aSJohn Heffner { 23885d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 238912a59abcSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 23905d424d5aSJohn Heffner struct sk_buff *skb, *nskb, *next; 23916b58e0a5SFan Du struct net *net = sock_net(sk); 23925d424d5aSJohn Heffner int probe_size; 239391cc17c0SIlpo Järvinen int size_needed; 239412a59abcSEric Dumazet int copy, len; 23955d424d5aSJohn Heffner int mss_now; 23966b58e0a5SFan Du int interval; 23975d424d5aSJohn Heffner 23985d424d5aSJohn Heffner /* Not currently probing/verifying, 23995d424d5aSJohn Heffner * not in recovery, 24005d424d5aSJohn Heffner * have enough cwnd, and 240112a59abcSEric Dumazet * not SACKing (the variable headers throw things off) 240212a59abcSEric Dumazet */ 240312a59abcSEric Dumazet if (likely(!icsk->icsk_mtup.enabled || 24045d424d5aSJohn Heffner icsk->icsk_mtup.probe_size || 24055d424d5aSJohn Heffner inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 240640570375SEric Dumazet tcp_snd_cwnd(tp) < 11 || 240712a59abcSEric Dumazet tp->rx_opt.num_sacks || tp->rx_opt.dsack)) 24085d424d5aSJohn Heffner return -1; 24095d424d5aSJohn Heffner 24106b58e0a5SFan Du /* Use binary search for probe_size between tcp_mss_base, 24116b58e0a5SFan Du * and current mss_clamp. if (search_high - search_low) 24126b58e0a5SFan Du * smaller than a threshold, backoff from probing. 24136b58e0a5SFan Du */ 24140c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 24156b58e0a5SFan Du probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high + 24166b58e0a5SFan Du icsk->icsk_mtup.search_low) >> 1); 241791cc17c0SIlpo Järvinen size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; 24186b58e0a5SFan Du interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low; 241905cbc0dbSFan Du /* When misfortune happens, we are reprobing actively, 242005cbc0dbSFan Du * and then reprobe timer has expired. We stick with current 242105cbc0dbSFan Du * probing process by not resetting search range to its orignal. 242205cbc0dbSFan Du */ 24236b58e0a5SFan Du if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) || 242492c0aa41SKuniyuki Iwashima interval < READ_ONCE(net->ipv4.sysctl_tcp_probe_threshold)) { 242505cbc0dbSFan Du /* Check whether enough time has elaplased for 242605cbc0dbSFan Du * another round of probing. 242705cbc0dbSFan Du */ 242805cbc0dbSFan Du tcp_mtu_check_reprobe(sk); 24295d424d5aSJohn Heffner return -1; 24305d424d5aSJohn Heffner } 24315d424d5aSJohn Heffner 24325d424d5aSJohn Heffner /* Have enough data in the send queue to probe? */ 24337f9c33e5SIlpo Järvinen if (tp->write_seq - tp->snd_nxt < size_needed) 24345d424d5aSJohn Heffner return -1; 24355d424d5aSJohn Heffner 243691cc17c0SIlpo Järvinen if (tp->snd_wnd < size_needed) 24375d424d5aSJohn Heffner return -1; 243890840defSIlpo Järvinen if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) 24395d424d5aSJohn Heffner return 0; 24405d424d5aSJohn Heffner 2441d67c58e9SIlpo Järvinen /* Do we need to wait to drain cwnd? With none in flight, don't stall */ 244240570375SEric Dumazet if (tcp_packets_in_flight(tp) + 2 > tcp_snd_cwnd(tp)) { 2443d67c58e9SIlpo Järvinen if (!tcp_packets_in_flight(tp)) 24445d424d5aSJohn Heffner return -1; 24455d424d5aSJohn Heffner else 24465d424d5aSJohn Heffner return 0; 24475d424d5aSJohn Heffner } 24485d424d5aSJohn Heffner 2449808cf9e3SIlya Lesokhin if (!tcp_can_coalesce_send_queue_head(sk, probe_size)) 2450808cf9e3SIlya Lesokhin return -1; 2451808cf9e3SIlya Lesokhin 24525d424d5aSJohn Heffner /* We're allowed to probe. Build it now. */ 24535882efffSEric Dumazet nskb = tcp_stream_alloc_skb(sk, GFP_ATOMIC, false); 245451456b29SIan Morris if (!nskb) 24555d424d5aSJohn Heffner return -1; 245673601329SEric Dumazet 245773601329SEric Dumazet /* build the payload, and be prepared to abort if this fails. */ 245873601329SEric Dumazet if (tcp_clone_payload(sk, nskb, probe_size)) { 245971c299c7SJakub Kicinski tcp_skb_tsorted_anchor_cleanup(nskb); 246073601329SEric Dumazet consume_skb(nskb); 246173601329SEric Dumazet return -1; 246273601329SEric Dumazet } 2463ab4e846aSEric Dumazet sk_wmem_queued_add(sk, nskb->truesize); 24643ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 24655d424d5aSJohn Heffner 2466fe067e8aSDavid S. Miller skb = tcp_send_head(sk); 246741477662SJakub Kicinski skb_copy_decrypted(nskb, skb); 24685a369ca6SPaolo Abeni mptcp_skb_ext_copy(nskb, skb); 24695d424d5aSJohn Heffner 24705d424d5aSJohn Heffner TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 24715d424d5aSJohn Heffner TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 24724de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; 24735d424d5aSJohn Heffner 247450c4817eSIlpo Järvinen tcp_insert_write_queue_before(nskb, skb, sk); 24752b7cda9cSEric Dumazet tcp_highest_sack_replace(sk, skb, nskb); 247650c4817eSIlpo Järvinen 24775d424d5aSJohn Heffner len = 0; 2478234b6860SIlpo Järvinen tcp_for_write_queue_from_safe(skb, next, sk) { 24795d424d5aSJohn Heffner copy = min_t(int, skb->len, probe_size - len); 24805d424d5aSJohn Heffner 24815d424d5aSJohn Heffner if (skb->len <= copy) { 24825d424d5aSJohn Heffner /* We've eaten all the data from this skb. 24835d424d5aSJohn Heffner * Throw it away. */ 24844de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 2485808cf9e3SIlya Lesokhin /* If this is the last SKB we copy and eor is set 2486808cf9e3SIlya Lesokhin * we need to propagate it to the new skb. 2487808cf9e3SIlya Lesokhin */ 2488808cf9e3SIlya Lesokhin TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor; 2489888a5c53SWillem de Bruijn tcp_skb_collapse_tstamp(nskb, skb); 2490fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 249103271f3aSTalal Ahmad tcp_wmem_free_skb(sk, skb); 24925d424d5aSJohn Heffner } else { 24934de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & 2494a3433f35SChangli Gao ~(TCPHDR_FIN|TCPHDR_PSH); 24955d424d5aSJohn Heffner __pskb_trim_head(skb, copy); 24965bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 24975d424d5aSJohn Heffner TCP_SKB_CB(skb)->seq += copy; 24985d424d5aSJohn Heffner } 24995d424d5aSJohn Heffner 25005d424d5aSJohn Heffner len += copy; 2501234b6860SIlpo Järvinen 2502234b6860SIlpo Järvinen if (len >= probe_size) 2503234b6860SIlpo Järvinen break; 25045d424d5aSJohn Heffner } 25055bbb432cSEric Dumazet tcp_init_tso_segs(nskb, nskb->len); 25065d424d5aSJohn Heffner 25075d424d5aSJohn Heffner /* We're ready to send. If this fails, the probe will 25087faee5c0SEric Dumazet * be resegmented into mss-sized pieces by tcp_write_xmit(). 25097faee5c0SEric Dumazet */ 25105d424d5aSJohn Heffner if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 25115d424d5aSJohn Heffner /* Decrement cwnd here because we are sending 25125d424d5aSJohn Heffner * effectively two packets. */ 251340570375SEric Dumazet tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) - 1); 251466f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, nskb); 25155d424d5aSJohn Heffner 25165d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 25170e7b1368SJohn Heffner tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 25180e7b1368SJohn Heffner tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 25195d424d5aSJohn Heffner 25205d424d5aSJohn Heffner return 1; 25215d424d5aSJohn Heffner } 25225d424d5aSJohn Heffner 25235d424d5aSJohn Heffner return -1; 25245d424d5aSJohn Heffner } 25255d424d5aSJohn Heffner 2526864e5c09SEric Dumazet static bool tcp_pacing_check(struct sock *sk) 2527218af599SEric Dumazet { 2528864e5c09SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 2529864e5c09SEric Dumazet 2530864e5c09SEric Dumazet if (!tcp_needs_internal_pacing(sk)) 2531864e5c09SEric Dumazet return false; 2532864e5c09SEric Dumazet 2533864e5c09SEric Dumazet if (tp->tcp_wstamp_ns <= tp->tcp_clock_cache) 2534864e5c09SEric Dumazet return false; 2535864e5c09SEric Dumazet 2536864e5c09SEric Dumazet if (!hrtimer_is_queued(&tp->pacing_timer)) { 2537864e5c09SEric Dumazet hrtimer_start(&tp->pacing_timer, 2538864e5c09SEric Dumazet ns_to_ktime(tp->tcp_wstamp_ns), 2539864e5c09SEric Dumazet HRTIMER_MODE_ABS_PINNED_SOFT); 2540864e5c09SEric Dumazet sock_hold(sk); 2541864e5c09SEric Dumazet } 2542864e5c09SEric Dumazet return true; 2543218af599SEric Dumazet } 2544218af599SEric Dumazet 2545*f921a4a5SEric Dumazet static bool tcp_rtx_queue_empty_or_single_skb(const struct sock *sk) 2546*f921a4a5SEric Dumazet { 2547*f921a4a5SEric Dumazet const struct rb_node *node = sk->tcp_rtx_queue.rb_node; 2548*f921a4a5SEric Dumazet 2549*f921a4a5SEric Dumazet /* No skb in the rtx queue. */ 2550*f921a4a5SEric Dumazet if (!node) 2551*f921a4a5SEric Dumazet return true; 2552*f921a4a5SEric Dumazet 2553*f921a4a5SEric Dumazet /* Only one skb in rtx queue. */ 2554*f921a4a5SEric Dumazet return !node->rb_left && !node->rb_right; 2555*f921a4a5SEric Dumazet } 2556*f921a4a5SEric Dumazet 2557f9616c35SEric Dumazet /* TCP Small Queues : 2558f9616c35SEric Dumazet * Control number of packets in qdisc/devices to two packets / or ~1 ms. 2559f9616c35SEric Dumazet * (These limits are doubled for retransmits) 2560f9616c35SEric Dumazet * This allows for : 2561f9616c35SEric Dumazet * - better RTT estimation and ACK scheduling 2562f9616c35SEric Dumazet * - faster recovery 2563f9616c35SEric Dumazet * - high rates 2564f9616c35SEric Dumazet * Alas, some drivers / subsystems require a fair amount 2565f9616c35SEric Dumazet * of queued bytes to ensure line rate. 2566f9616c35SEric Dumazet * One example is wifi aggregation (802.11 AMPDU) 2567f9616c35SEric Dumazet */ 2568f9616c35SEric Dumazet static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb, 2569f9616c35SEric Dumazet unsigned int factor) 2570f9616c35SEric Dumazet { 257176a9ebe8SEric Dumazet unsigned long limit; 2572f9616c35SEric Dumazet 257376a9ebe8SEric Dumazet limit = max_t(unsigned long, 257476a9ebe8SEric Dumazet 2 * skb->truesize, 25757c68fa2bSEric Dumazet sk->sk_pacing_rate >> READ_ONCE(sk->sk_pacing_shift)); 2576c73e5807SEric Dumazet if (sk->sk_pacing_status == SK_PACING_NONE) 257776a9ebe8SEric Dumazet limit = min_t(unsigned long, limit, 25789fb90193SKuniyuki Iwashima READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes)); 2579f9616c35SEric Dumazet limit <<= factor; 2580f9616c35SEric Dumazet 2581a842fe14SEric Dumazet if (static_branch_unlikely(&tcp_tx_delay_enabled) && 2582a842fe14SEric Dumazet tcp_sk(sk)->tcp_tx_delay) { 2583a842fe14SEric Dumazet u64 extra_bytes = (u64)sk->sk_pacing_rate * tcp_sk(sk)->tcp_tx_delay; 2584a842fe14SEric Dumazet 2585a842fe14SEric Dumazet /* TSQ is based on skb truesize sum (sk_wmem_alloc), so we 2586a842fe14SEric Dumazet * approximate our needs assuming an ~100% skb->truesize overhead. 2587a842fe14SEric Dumazet * USEC_PER_SEC is approximated by 2^20. 2588a842fe14SEric Dumazet * do_div(extra_bytes, USEC_PER_SEC/2) is replaced by a right shift. 2589a842fe14SEric Dumazet */ 2590a842fe14SEric Dumazet extra_bytes >>= (20 - 1); 2591a842fe14SEric Dumazet limit += extra_bytes; 2592a842fe14SEric Dumazet } 259314afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) > limit) { 2594*f921a4a5SEric Dumazet /* Always send skb if rtx queue is empty or has one skb. 259575eefc6cSEric Dumazet * No need to wait for TX completion to call us back, 259675eefc6cSEric Dumazet * after softirq/tasklet schedule. 259775eefc6cSEric Dumazet * This helps when TX completions are delayed too much. 259875eefc6cSEric Dumazet */ 2599*f921a4a5SEric Dumazet if (tcp_rtx_queue_empty_or_single_skb(sk)) 260075eefc6cSEric Dumazet return false; 260175eefc6cSEric Dumazet 26027aa5470cSEric Dumazet set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags); 2603f9616c35SEric Dumazet /* It is possible TX completion already happened 2604f9616c35SEric Dumazet * before we set TSQ_THROTTLED, so we must 2605f9616c35SEric Dumazet * test again the condition. 2606f9616c35SEric Dumazet */ 2607f9616c35SEric Dumazet smp_mb__after_atomic(); 2608ce8299b6SEric Dumazet if (refcount_read(&sk->sk_wmem_alloc) > limit) 2609f9616c35SEric Dumazet return true; 2610f9616c35SEric Dumazet } 2611f9616c35SEric Dumazet return false; 2612f9616c35SEric Dumazet } 2613f9616c35SEric Dumazet 261405b055e8SFrancis Yan static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new) 261505b055e8SFrancis Yan { 2616628174ccSEric Dumazet const u32 now = tcp_jiffies32; 2617efe967cdSArnd Bergmann enum tcp_chrono old = tp->chrono_type; 261805b055e8SFrancis Yan 2619efe967cdSArnd Bergmann if (old > TCP_CHRONO_UNSPEC) 2620efe967cdSArnd Bergmann tp->chrono_stat[old - 1] += now - tp->chrono_start; 262105b055e8SFrancis Yan tp->chrono_start = now; 262205b055e8SFrancis Yan tp->chrono_type = new; 262305b055e8SFrancis Yan } 262405b055e8SFrancis Yan 262505b055e8SFrancis Yan void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type) 262605b055e8SFrancis Yan { 262705b055e8SFrancis Yan struct tcp_sock *tp = tcp_sk(sk); 262805b055e8SFrancis Yan 262905b055e8SFrancis Yan /* If there are multiple conditions worthy of tracking in a 26300f87230dSFrancis Yan * chronograph then the highest priority enum takes precedence 26310f87230dSFrancis Yan * over the other conditions. So that if something "more interesting" 263205b055e8SFrancis Yan * starts happening, stop the previous chrono and start a new one. 263305b055e8SFrancis Yan */ 263405b055e8SFrancis Yan if (type > tp->chrono_type) 263505b055e8SFrancis Yan tcp_chrono_set(tp, type); 263605b055e8SFrancis Yan } 263705b055e8SFrancis Yan 263805b055e8SFrancis Yan void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type) 263905b055e8SFrancis Yan { 264005b055e8SFrancis Yan struct tcp_sock *tp = tcp_sk(sk); 264105b055e8SFrancis Yan 26420f87230dSFrancis Yan 26430f87230dSFrancis Yan /* There are multiple conditions worthy of tracking in a 26440f87230dSFrancis Yan * chronograph, so that the highest priority enum takes 26450f87230dSFrancis Yan * precedence over the other conditions (see tcp_chrono_start). 26460f87230dSFrancis Yan * If a condition stops, we only stop chrono tracking if 26470f87230dSFrancis Yan * it's the "most interesting" or current chrono we are 26480f87230dSFrancis Yan * tracking and starts busy chrono if we have pending data. 26490f87230dSFrancis Yan */ 265075c119afSEric Dumazet if (tcp_rtx_and_write_queues_empty(sk)) 265105b055e8SFrancis Yan tcp_chrono_set(tp, TCP_CHRONO_UNSPEC); 26520f87230dSFrancis Yan else if (type == tp->chrono_type) 26530f87230dSFrancis Yan tcp_chrono_set(tp, TCP_CHRONO_BUSY); 265405b055e8SFrancis Yan } 265505b055e8SFrancis Yan 26561da177e4SLinus Torvalds /* This routine writes packets to the network. It advances the 26571da177e4SLinus Torvalds * send_head. This happens as incoming acks open up the remote 26581da177e4SLinus Torvalds * window for us. 26591da177e4SLinus Torvalds * 2660f8269a49SIlpo Järvinen * LARGESEND note: !tcp_urg_mode is overkill, only frames between 2661f8269a49SIlpo Järvinen * snd_up-64k-mss .. snd_up cannot be large. However, taking into 2662f8269a49SIlpo Järvinen * account rare use of URG, this is not a big flaw. 2663f8269a49SIlpo Järvinen * 26646ba8a3b1SNandita Dukkipati * Send at most one packet when push_one > 0. Temporarily ignore 26656ba8a3b1SNandita Dukkipati * cwnd limit to force at most one packet out when push_one == 2. 26666ba8a3b1SNandita Dukkipati 2667a2a385d6SEric Dumazet * Returns true, if no segments are in flight and we have queued segments, 2668a2a385d6SEric Dumazet * but cannot send anything now because of SWS or another problem. 26691da177e4SLinus Torvalds */ 2670a2a385d6SEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 2671d5dd9175SIlpo Järvinen int push_one, gfp_t gfp) 26721da177e4SLinus Torvalds { 26731da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 267492df7b51SDavid S. Miller struct sk_buff *skb; 2675c1b4a7e6SDavid S. Miller unsigned int tso_segs, sent_pkts; 2676c1b4a7e6SDavid S. Miller int cwnd_quota; 26775d424d5aSJohn Heffner int result; 26785615f886SFrancis Yan bool is_cwnd_limited = false, is_rwnd_limited = false; 2679605ad7f1SEric Dumazet u32 max_segs; 26801da177e4SLinus Torvalds 2681c1b4a7e6SDavid S. Miller sent_pkts = 0; 26825d424d5aSJohn Heffner 2683ee1836aeSEric Dumazet tcp_mstamp_refresh(tp); 2684d5dd9175SIlpo Järvinen if (!push_one) { 26855d424d5aSJohn Heffner /* Do MTU probing. */ 2686d5dd9175SIlpo Järvinen result = tcp_mtu_probe(sk); 2687d5dd9175SIlpo Järvinen if (!result) { 2688a2a385d6SEric Dumazet return false; 26895d424d5aSJohn Heffner } else if (result > 0) { 26905d424d5aSJohn Heffner sent_pkts = 1; 26915d424d5aSJohn Heffner } 2692d5dd9175SIlpo Järvinen } 26935d424d5aSJohn Heffner 2694ed6e7268SNeal Cardwell max_segs = tcp_tso_segs(sk, mss_now); 2695fe067e8aSDavid S. Miller while ((skb = tcp_send_head(sk))) { 2696c8ac3774SHerbert Xu unsigned int limit; 2697c8ac3774SHerbert Xu 269879861919SEric Dumazet if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) { 269979861919SEric Dumazet /* "skb_mstamp_ns" is used as a start point for the retransmit timer */ 2700a1ac9c8aSMartin KaFai Lau tp->tcp_wstamp_ns = tp->tcp_clock_cache; 2701a1ac9c8aSMartin KaFai Lau skb_set_delivery_time(skb, tp->tcp_wstamp_ns, true); 270279861919SEric Dumazet list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue); 2703bf50b606SEric Dumazet tcp_init_tso_segs(skb, mss_now); 270479861919SEric Dumazet goto repair; /* Skip network transmission */ 270579861919SEric Dumazet } 270679861919SEric Dumazet 2707218af599SEric Dumazet if (tcp_pacing_check(sk)) 2708218af599SEric Dumazet break; 2709218af599SEric Dumazet 27105bbb432cSEric Dumazet tso_segs = tcp_init_tso_segs(skb, mss_now); 2711c1b4a7e6SDavid S. Miller BUG_ON(!tso_segs); 2712c1b4a7e6SDavid S. Miller 2713b68e9f85SHerbert Xu cwnd_quota = tcp_cwnd_test(tp, skb); 27146ba8a3b1SNandita Dukkipati if (!cwnd_quota) { 27156ba8a3b1SNandita Dukkipati if (push_one == 2) 27166ba8a3b1SNandita Dukkipati /* Force out a loss probe pkt. */ 27176ba8a3b1SNandita Dukkipati cwnd_quota = 1; 27186ba8a3b1SNandita Dukkipati else 2719b68e9f85SHerbert Xu break; 27206ba8a3b1SNandita Dukkipati } 2721b68e9f85SHerbert Xu 27225615f886SFrancis Yan if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) { 27235615f886SFrancis Yan is_rwnd_limited = true; 2724b68e9f85SHerbert Xu break; 27255615f886SFrancis Yan } 2726b68e9f85SHerbert Xu 2727d6a4e26aSEric Dumazet if (tso_segs == 1) { 2728aa93466bSDavid S. Miller if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 2729aa93466bSDavid S. Miller (tcp_skb_is_last(sk, skb) ? 2730aa93466bSDavid S. Miller nonagle : TCP_NAGLE_PUSH)))) 2731aa93466bSDavid S. Miller break; 2732c1b4a7e6SDavid S. Miller } else { 2733ca8a2263SNeal Cardwell if (!push_one && 2734605ad7f1SEric Dumazet tcp_tso_should_defer(sk, skb, &is_cwnd_limited, 2735f9bfe4e6SEric Dumazet &is_rwnd_limited, max_segs)) 2736aa93466bSDavid S. Miller break; 2737c1b4a7e6SDavid S. Miller } 2738aa93466bSDavid S. Miller 2739605ad7f1SEric Dumazet limit = mss_now; 2740d6a4e26aSEric Dumazet if (tso_segs > 1 && !tcp_urg_mode(tp)) 2741605ad7f1SEric Dumazet limit = tcp_mss_split_point(sk, skb, mss_now, 2742605ad7f1SEric Dumazet min_t(unsigned int, 2743605ad7f1SEric Dumazet cwnd_quota, 2744605ad7f1SEric Dumazet max_segs), 2745605ad7f1SEric Dumazet nonagle); 2746605ad7f1SEric Dumazet 2747605ad7f1SEric Dumazet if (skb->len > limit && 274856483341SEric Dumazet unlikely(tso_fragment(sk, skb, limit, mss_now, gfp))) 2749605ad7f1SEric Dumazet break; 2750605ad7f1SEric Dumazet 2751f9616c35SEric Dumazet if (tcp_small_queue_check(sk, skb, 0)) 275246d3ceabSEric Dumazet break; 2753c9eeec26SEric Dumazet 27541f85e626SEric Dumazet /* Argh, we hit an empty skb(), presumably a thread 27551f85e626SEric Dumazet * is sleeping in sendmsg()/sk_stream_wait_memory(). 27561f85e626SEric Dumazet * We do not want to send a pure-ack packet and have 27571f85e626SEric Dumazet * a strange looking rtx queue with empty packet(s). 27581f85e626SEric Dumazet */ 27591f85e626SEric Dumazet if (TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq) 27601f85e626SEric Dumazet break; 27611f85e626SEric Dumazet 2762d5dd9175SIlpo Järvinen if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) 27631da177e4SLinus Torvalds break; 27641da177e4SLinus Torvalds 2765ec342325SAndrew Vagin repair: 27661da177e4SLinus Torvalds /* Advance the send_head. This one is sent out. 27671da177e4SLinus Torvalds * This call will increment packets_out. 27681da177e4SLinus Torvalds */ 276966f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 27701da177e4SLinus Torvalds 27711da177e4SLinus Torvalds tcp_minshall_update(tp, mss_now, skb); 2772a262f0cdSNandita Dukkipati sent_pkts += tcp_skb_pcount(skb); 2773d5dd9175SIlpo Järvinen 2774d5dd9175SIlpo Järvinen if (push_one) 2775d5dd9175SIlpo Järvinen break; 27761da177e4SLinus Torvalds } 27771da177e4SLinus Torvalds 27785615f886SFrancis Yan if (is_rwnd_limited) 27795615f886SFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED); 27805615f886SFrancis Yan else 27815615f886SFrancis Yan tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED); 27825615f886SFrancis Yan 278340570375SEric Dumazet is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tcp_snd_cwnd(tp)); 2784299bcb55SNeal Cardwell if (likely(sent_pkts || is_cwnd_limited)) 2785299bcb55SNeal Cardwell tcp_cwnd_validate(sk, is_cwnd_limited); 2786299bcb55SNeal Cardwell 2787aa93466bSDavid S. Miller if (likely(sent_pkts)) { 2788684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 2789684bad11SYuchung Cheng tp->prr_out += sent_pkts; 27906ba8a3b1SNandita Dukkipati 27916ba8a3b1SNandita Dukkipati /* Send one loss probe per tail loss episode. */ 27926ba8a3b1SNandita Dukkipati if (push_one != 2) 2793ed66dfafSNeal Cardwell tcp_schedule_loss_probe(sk, false); 2794a2a385d6SEric Dumazet return false; 27951da177e4SLinus Torvalds } 279675c119afSEric Dumazet return !tp->packets_out && !tcp_write_queue_empty(sk); 27976ba8a3b1SNandita Dukkipati } 27986ba8a3b1SNandita Dukkipati 2799ed66dfafSNeal Cardwell bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto) 28006ba8a3b1SNandita Dukkipati { 28016ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 28026ba8a3b1SNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 28031c2709cfSNeal Cardwell u32 timeout, timeout_us, rto_delta_us; 28042ae21cf5SEric Dumazet int early_retrans; 28056ba8a3b1SNandita Dukkipati 28066ba8a3b1SNandita Dukkipati /* Don't do any loss probe on a Fast Open connection before 3WHS 28076ba8a3b1SNandita Dukkipati * finishes. 28086ba8a3b1SNandita Dukkipati */ 2809d983ea6fSEric Dumazet if (rcu_access_pointer(tp->fastopen_rsk)) 28106ba8a3b1SNandita Dukkipati return false; 28116ba8a3b1SNandita Dukkipati 281252e65865SKuniyuki Iwashima early_retrans = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_early_retrans); 28136ba8a3b1SNandita Dukkipati /* Schedule a loss probe in 2*RTT for SACK capable connections 2814b4f70c3dSNeal Cardwell * not in loss recovery, that are either limited by cwnd or application. 28156ba8a3b1SNandita Dukkipati */ 28162ae21cf5SEric Dumazet if ((early_retrans != 3 && early_retrans != 4) || 2817bec41a11SYuchung Cheng !tp->packets_out || !tcp_is_sack(tp) || 2818b4f70c3dSNeal Cardwell (icsk->icsk_ca_state != TCP_CA_Open && 2819b4f70c3dSNeal Cardwell icsk->icsk_ca_state != TCP_CA_CWR)) 28206ba8a3b1SNandita Dukkipati return false; 28216ba8a3b1SNandita Dukkipati 2822bb4d991aSYuchung Cheng /* Probe timeout is 2*rtt. Add minimum RTO to account 2823f9b99582SYuchung Cheng * for delayed ack when there's one outstanding packet. If no RTT 2824f9b99582SYuchung Cheng * sample is available then probe after TCP_TIMEOUT_INIT. 28256ba8a3b1SNandita Dukkipati */ 2826bb4d991aSYuchung Cheng if (tp->srtt_us) { 28271c2709cfSNeal Cardwell timeout_us = tp->srtt_us >> 2; 28286ba8a3b1SNandita Dukkipati if (tp->packets_out == 1) 28291c2709cfSNeal Cardwell timeout_us += tcp_rto_min_us(sk); 2830bb4d991aSYuchung Cheng else 28311c2709cfSNeal Cardwell timeout_us += TCP_TIMEOUT_MIN_US; 28321c2709cfSNeal Cardwell timeout = usecs_to_jiffies(timeout_us); 2833bb4d991aSYuchung Cheng } else { 2834bb4d991aSYuchung Cheng timeout = TCP_TIMEOUT_INIT; 2835bb4d991aSYuchung Cheng } 28366ba8a3b1SNandita Dukkipati 2837a2815817SNeal Cardwell /* If the RTO formula yields an earlier time, then use that time. */ 2838ed66dfafSNeal Cardwell rto_delta_us = advancing_rto ? 2839ed66dfafSNeal Cardwell jiffies_to_usecs(inet_csk(sk)->icsk_rto) : 2840ed66dfafSNeal Cardwell tcp_rto_delta_us(sk); /* How far in future is RTO? */ 2841a2815817SNeal Cardwell if (rto_delta_us > 0) 2842a2815817SNeal Cardwell timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us)); 28436ba8a3b1SNandita Dukkipati 28448dc242adSEric Dumazet tcp_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, TCP_RTO_MAX); 28456ba8a3b1SNandita Dukkipati return true; 28466ba8a3b1SNandita Dukkipati } 28476ba8a3b1SNandita Dukkipati 28481f3279aeSEric Dumazet /* Thanks to skb fast clones, we can detect if a prior transmit of 28491f3279aeSEric Dumazet * a packet is still in a qdisc or driver queue. 28501f3279aeSEric Dumazet * In this case, there is very little point doing a retransmit ! 28511f3279aeSEric Dumazet */ 2852f4dae54eSEric Dumazet static bool skb_still_in_host_queue(struct sock *sk, 28531f3279aeSEric Dumazet const struct sk_buff *skb) 28541f3279aeSEric Dumazet { 285539bb5e62SEric Dumazet if (unlikely(skb_fclone_busy(sk, skb))) { 2856f4dae54eSEric Dumazet set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags); 2857f4dae54eSEric Dumazet smp_mb__after_atomic(); 2858f4dae54eSEric Dumazet if (skb_fclone_busy(sk, skb)) { 2859c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), 28601f3279aeSEric Dumazet LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES); 28611f3279aeSEric Dumazet return true; 28621f3279aeSEric Dumazet } 2863f4dae54eSEric Dumazet } 28641f3279aeSEric Dumazet return false; 28651f3279aeSEric Dumazet } 28661f3279aeSEric Dumazet 2867b340b264SYuchung Cheng /* When probe timeout (PTO) fires, try send a new segment if possible, else 28686ba8a3b1SNandita Dukkipati * retransmit the last segment. 28696ba8a3b1SNandita Dukkipati */ 28706ba8a3b1SNandita Dukkipati void tcp_send_loss_probe(struct sock *sk) 28716ba8a3b1SNandita Dukkipati { 28729b717a8dSNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 28736ba8a3b1SNandita Dukkipati struct sk_buff *skb; 28746ba8a3b1SNandita Dukkipati int pcount; 28756ba8a3b1SNandita Dukkipati int mss = tcp_current_mss(sk); 28766ba8a3b1SNandita Dukkipati 287776be93fcSYuchung Cheng /* At most one outstanding TLP */ 287876be93fcSYuchung Cheng if (tp->tlp_high_seq) 287976be93fcSYuchung Cheng goto rearm_timer; 288076be93fcSYuchung Cheng 288176be93fcSYuchung Cheng tp->tlp_retrans = 0; 2882b340b264SYuchung Cheng skb = tcp_send_head(sk); 288375c119afSEric Dumazet if (skb && tcp_snd_wnd_test(tp, skb, mss)) { 2884b340b264SYuchung Cheng pcount = tp->packets_out; 2885b340b264SYuchung Cheng tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC); 2886b340b264SYuchung Cheng if (tp->packets_out > pcount) 2887b340b264SYuchung Cheng goto probe_sent; 28886ba8a3b1SNandita Dukkipati goto rearm_timer; 28896ba8a3b1SNandita Dukkipati } 289075c119afSEric Dumazet skb = skb_rb_last(&sk->tcp_rtx_queue); 2891b2b7af86SYuchung Cheng if (unlikely(!skb)) { 2892b2b7af86SYuchung Cheng WARN_ONCE(tp->packets_out, 2893b2b7af86SYuchung Cheng "invalid inflight: %u state %u cwnd %u mss %d\n", 289440570375SEric Dumazet tp->packets_out, sk->sk_state, tcp_snd_cwnd(tp), mss); 2895b2b7af86SYuchung Cheng inet_csk(sk)->icsk_pending = 0; 2896b2b7af86SYuchung Cheng return; 2897b2b7af86SYuchung Cheng } 28986ba8a3b1SNandita Dukkipati 28991f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 29001f3279aeSEric Dumazet goto rearm_timer; 29011f3279aeSEric Dumazet 29026ba8a3b1SNandita Dukkipati pcount = tcp_skb_pcount(skb); 29036ba8a3b1SNandita Dukkipati if (WARN_ON(!pcount)) 29046ba8a3b1SNandita Dukkipati goto rearm_timer; 29056ba8a3b1SNandita Dukkipati 29066ba8a3b1SNandita Dukkipati if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) { 290775c119afSEric Dumazet if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, 290875c119afSEric Dumazet (pcount - 1) * mss, mss, 29096cc55e09SOctavian Purdila GFP_ATOMIC))) 29106ba8a3b1SNandita Dukkipati goto rearm_timer; 291175c119afSEric Dumazet skb = skb_rb_next(skb); 29126ba8a3b1SNandita Dukkipati } 29136ba8a3b1SNandita Dukkipati 29146ba8a3b1SNandita Dukkipati if (WARN_ON(!skb || !tcp_skb_pcount(skb))) 29156ba8a3b1SNandita Dukkipati goto rearm_timer; 29166ba8a3b1SNandita Dukkipati 291710d3be56SEric Dumazet if (__tcp_retransmit_skb(sk, skb, 1)) 2918b340b264SYuchung Cheng goto rearm_timer; 29196ba8a3b1SNandita Dukkipati 292076be93fcSYuchung Cheng tp->tlp_retrans = 1; 292176be93fcSYuchung Cheng 292276be93fcSYuchung Cheng probe_sent: 29239b717a8dSNandita Dukkipati /* Record snd_nxt for loss detection. */ 29249b717a8dSNandita Dukkipati tp->tlp_high_seq = tp->snd_nxt; 29259b717a8dSNandita Dukkipati 2926c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES); 2927fcd16c0aSYuchung Cheng /* Reset s.t. tcp_rearm_rto will restart timer from now */ 2928fcd16c0aSYuchung Cheng inet_csk(sk)->icsk_pending = 0; 2929b340b264SYuchung Cheng rearm_timer: 2930fcd16c0aSYuchung Cheng tcp_rearm_rto(sk); 29311da177e4SLinus Torvalds } 29321da177e4SLinus Torvalds 2933a762a980SDavid S. Miller /* Push out any pending frames which were held back due to 2934a762a980SDavid S. Miller * TCP_CORK or attempt at coalescing tiny packets. 2935a762a980SDavid S. Miller * The socket must be locked by the caller. 2936a762a980SDavid S. Miller */ 29379e412ba7SIlpo Järvinen void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 29389e412ba7SIlpo Järvinen int nonagle) 2939a762a980SDavid S. Miller { 2940726e07a8SIlpo Järvinen /* If we are closed, the bytes will have to remain here. 2941726e07a8SIlpo Järvinen * In time closedown will finish, we empty the write queue and 2942726e07a8SIlpo Järvinen * all will be happy. 2943726e07a8SIlpo Järvinen */ 2944726e07a8SIlpo Järvinen if (unlikely(sk->sk_state == TCP_CLOSE)) 2945726e07a8SIlpo Järvinen return; 2946726e07a8SIlpo Järvinen 294799a1dec7SMel Gorman if (tcp_write_xmit(sk, cur_mss, nonagle, 0, 29487450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC))) 29499e412ba7SIlpo Järvinen tcp_check_probe_timer(sk); 2950a762a980SDavid S. Miller } 2951a762a980SDavid S. Miller 2952c1b4a7e6SDavid S. Miller /* Send _single_ skb sitting at the send head. This function requires 2953c1b4a7e6SDavid S. Miller * true push pending frames to setup probe timer etc. 2954c1b4a7e6SDavid S. Miller */ 2955c1b4a7e6SDavid S. Miller void tcp_push_one(struct sock *sk, unsigned int mss_now) 2956c1b4a7e6SDavid S. Miller { 2957fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 2958c1b4a7e6SDavid S. Miller 2959c1b4a7e6SDavid S. Miller BUG_ON(!skb || skb->len < mss_now); 2960c1b4a7e6SDavid S. Miller 2961d5dd9175SIlpo Järvinen tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); 2962c1b4a7e6SDavid S. Miller } 2963c1b4a7e6SDavid S. Miller 29641da177e4SLinus Torvalds /* This function returns the amount that we can raise the 29651da177e4SLinus Torvalds * usable window based on the following constraints 29661da177e4SLinus Torvalds * 29671da177e4SLinus Torvalds * 1. The window can never be shrunk once it is offered (RFC 793) 29681da177e4SLinus Torvalds * 2. We limit memory per socket 29691da177e4SLinus Torvalds * 29701da177e4SLinus Torvalds * RFC 1122: 29711da177e4SLinus Torvalds * "the suggested [SWS] avoidance algorithm for the receiver is to keep 29721da177e4SLinus Torvalds * RECV.NEXT + RCV.WIN fixed until: 29731da177e4SLinus Torvalds * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 29741da177e4SLinus Torvalds * 29751da177e4SLinus Torvalds * i.e. don't raise the right edge of the window until you can raise 29761da177e4SLinus Torvalds * it at least MSS bytes. 29771da177e4SLinus Torvalds * 29781da177e4SLinus Torvalds * Unfortunately, the recommended algorithm breaks header prediction, 29791da177e4SLinus Torvalds * since header prediction assumes th->window stays fixed. 29801da177e4SLinus Torvalds * 29811da177e4SLinus Torvalds * Strictly speaking, keeping th->window fixed violates the receiver 29821da177e4SLinus Torvalds * side SWS prevention criteria. The problem is that under this rule 29831da177e4SLinus Torvalds * a stream of single byte packets will cause the right side of the 29841da177e4SLinus Torvalds * window to always advance by a single byte. 29851da177e4SLinus Torvalds * 29861da177e4SLinus Torvalds * Of course, if the sender implements sender side SWS prevention 29871da177e4SLinus Torvalds * then this will not be a problem. 29881da177e4SLinus Torvalds * 29891da177e4SLinus Torvalds * BSD seems to make the following compromise: 29901da177e4SLinus Torvalds * 29911da177e4SLinus Torvalds * If the free space is less than the 1/4 of the maximum 29921da177e4SLinus Torvalds * space available and the free space is less than 1/2 mss, 29931da177e4SLinus Torvalds * then set the window to 0. 29941da177e4SLinus Torvalds * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 29951da177e4SLinus Torvalds * Otherwise, just prevent the window from shrinking 29961da177e4SLinus Torvalds * and from being larger than the largest representable value. 29971da177e4SLinus Torvalds * 29981da177e4SLinus Torvalds * This prevents incremental opening of the window in the regime 29991da177e4SLinus Torvalds * where TCP is limited by the speed of the reader side taking 30001da177e4SLinus Torvalds * data out of the TCP receive queue. It does nothing about 30011da177e4SLinus Torvalds * those cases where the window is constrained on the sender side 30021da177e4SLinus Torvalds * because the pipeline is full. 30031da177e4SLinus Torvalds * 30041da177e4SLinus Torvalds * BSD also seems to "accidentally" limit itself to windows that are a 30051da177e4SLinus Torvalds * multiple of MSS, at least until the free space gets quite small. 30061da177e4SLinus Torvalds * This would appear to be a side effect of the mbuf implementation. 30071da177e4SLinus Torvalds * Combining these two algorithms results in the observed behavior 30081da177e4SLinus Torvalds * of having a fixed window size at almost all times. 30091da177e4SLinus Torvalds * 30101da177e4SLinus Torvalds * Below we obtain similar behavior by forcing the offered window to 30111da177e4SLinus Torvalds * a multiple of the mss when it is feasible to do so. 30121da177e4SLinus Torvalds * 30131da177e4SLinus Torvalds * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 30141da177e4SLinus Torvalds * Regular options like TIMESTAMP are taken into account. 30151da177e4SLinus Torvalds */ 30161da177e4SLinus Torvalds u32 __tcp_select_window(struct sock *sk) 30171da177e4SLinus Torvalds { 3018463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 30191da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 3020b650d953Smfreemon@cloudflare.com struct net *net = sock_net(sk); 3021caa20d9aSStephen Hemminger /* MSS for the peer's data. Previous versions used mss_clamp 30221da177e4SLinus Torvalds * here. I don't know if the value based on our guesses 30231da177e4SLinus Torvalds * of peer's MSS is better for the performance. It's more correct 30241da177e4SLinus Torvalds * but may be worse for the performance because of rcv_mss 30251da177e4SLinus Torvalds * fluctuations. --SAW 1998/11/1 30261da177e4SLinus Torvalds */ 3027463c84b9SArnaldo Carvalho de Melo int mss = icsk->icsk_ack.rcv_mss; 30281da177e4SLinus Torvalds int free_space = tcp_space(sk); 302986c1a045SFlorian Westphal int allowed_space = tcp_full_space(sk); 3030071c8ed6SFlorian Westphal int full_space, window; 3031071c8ed6SFlorian Westphal 3032071c8ed6SFlorian Westphal if (sk_is_mptcp(sk)) 3033071c8ed6SFlorian Westphal mptcp_space(sk, &free_space, &allowed_space); 3034071c8ed6SFlorian Westphal 3035071c8ed6SFlorian Westphal full_space = min_t(int, tp->window_clamp, allowed_space); 30361da177e4SLinus Torvalds 303706425c30SEric Dumazet if (unlikely(mss > full_space)) { 30381da177e4SLinus Torvalds mss = full_space; 303906425c30SEric Dumazet if (mss <= 0) 304006425c30SEric Dumazet return 0; 304106425c30SEric Dumazet } 3042b650d953Smfreemon@cloudflare.com 3043b650d953Smfreemon@cloudflare.com /* Only allow window shrink if the sysctl is enabled and we have 3044b650d953Smfreemon@cloudflare.com * a non-zero scaling factor in effect. 3045b650d953Smfreemon@cloudflare.com */ 3046b650d953Smfreemon@cloudflare.com if (READ_ONCE(net->ipv4.sysctl_tcp_shrink_window) && tp->rx_opt.rcv_wscale) 3047b650d953Smfreemon@cloudflare.com goto shrink_window_allowed; 3048b650d953Smfreemon@cloudflare.com 3049b650d953Smfreemon@cloudflare.com /* do not allow window to shrink */ 3050b650d953Smfreemon@cloudflare.com 3051b92edbe0SEric Dumazet if (free_space < (full_space >> 1)) { 3052463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.quick = 0; 30531da177e4SLinus Torvalds 3054b8da51ebSEric Dumazet if (tcp_under_memory_pressure(sk)) 3055053f3684SWei Wang tcp_adjust_rcv_ssthresh(sk); 30561da177e4SLinus Torvalds 305786c1a045SFlorian Westphal /* free_space might become our new window, make sure we don't 305886c1a045SFlorian Westphal * increase it due to wscale. 305986c1a045SFlorian Westphal */ 306086c1a045SFlorian Westphal free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale); 306186c1a045SFlorian Westphal 306286c1a045SFlorian Westphal /* if free space is less than mss estimate, or is below 1/16th 306386c1a045SFlorian Westphal * of the maximum allowed, try to move to zero-window, else 306486c1a045SFlorian Westphal * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and 306586c1a045SFlorian Westphal * new incoming data is dropped due to memory limits. 306686c1a045SFlorian Westphal * With large window, mss test triggers way too late in order 306786c1a045SFlorian Westphal * to announce zero window in time before rmem limit kicks in. 306886c1a045SFlorian Westphal */ 306986c1a045SFlorian Westphal if (free_space < (allowed_space >> 4) || free_space < mss) 30701da177e4SLinus Torvalds return 0; 30711da177e4SLinus Torvalds } 30721da177e4SLinus Torvalds 30731da177e4SLinus Torvalds if (free_space > tp->rcv_ssthresh) 30741da177e4SLinus Torvalds free_space = tp->rcv_ssthresh; 30751da177e4SLinus Torvalds 30761da177e4SLinus Torvalds /* Don't do rounding if we are using window scaling, since the 30771da177e4SLinus Torvalds * scaled window will not line up with the MSS boundary anyway. 30781da177e4SLinus Torvalds */ 30791da177e4SLinus Torvalds if (tp->rx_opt.rcv_wscale) { 30801da177e4SLinus Torvalds window = free_space; 30811da177e4SLinus Torvalds 30821da177e4SLinus Torvalds /* Advertise enough space so that it won't get scaled away. 30831da177e4SLinus Torvalds * Import case: prevent zero window announcement if 30841da177e4SLinus Torvalds * 1<<rcv_wscale > mss. 30851da177e4SLinus Torvalds */ 30861935299dSGao Feng window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale)); 30871da177e4SLinus Torvalds } else { 30881935299dSGao Feng window = tp->rcv_wnd; 30891da177e4SLinus Torvalds /* Get the largest window that is a nice multiple of mss. 30901da177e4SLinus Torvalds * Window clamp already applied above. 30911da177e4SLinus Torvalds * If our current window offering is within 1 mss of the 30921da177e4SLinus Torvalds * free space we just keep it. This prevents the divide 30931da177e4SLinus Torvalds * and multiply from happening most of the time. 30941da177e4SLinus Torvalds * We also don't do any window rounding when the free space 30951da177e4SLinus Torvalds * is too small. 30961da177e4SLinus Torvalds */ 30971da177e4SLinus Torvalds if (window <= free_space - mss || window > free_space) 30981935299dSGao Feng window = rounddown(free_space, mss); 309984565070SJohn Heffner else if (mss == full_space && 3100b92edbe0SEric Dumazet free_space > window + (full_space >> 1)) 310184565070SJohn Heffner window = free_space; 31021da177e4SLinus Torvalds } 31031da177e4SLinus Torvalds 31041da177e4SLinus Torvalds return window; 3105b650d953Smfreemon@cloudflare.com 3106b650d953Smfreemon@cloudflare.com shrink_window_allowed: 3107b650d953Smfreemon@cloudflare.com /* new window should always be an exact multiple of scaling factor */ 3108b650d953Smfreemon@cloudflare.com free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale); 3109b650d953Smfreemon@cloudflare.com 3110b650d953Smfreemon@cloudflare.com if (free_space < (full_space >> 1)) { 3111b650d953Smfreemon@cloudflare.com icsk->icsk_ack.quick = 0; 3112b650d953Smfreemon@cloudflare.com 3113b650d953Smfreemon@cloudflare.com if (tcp_under_memory_pressure(sk)) 3114b650d953Smfreemon@cloudflare.com tcp_adjust_rcv_ssthresh(sk); 3115b650d953Smfreemon@cloudflare.com 3116b650d953Smfreemon@cloudflare.com /* if free space is too low, return a zero window */ 3117b650d953Smfreemon@cloudflare.com if (free_space < (allowed_space >> 4) || free_space < mss || 3118b650d953Smfreemon@cloudflare.com free_space < (1 << tp->rx_opt.rcv_wscale)) 3119b650d953Smfreemon@cloudflare.com return 0; 3120b650d953Smfreemon@cloudflare.com } 3121b650d953Smfreemon@cloudflare.com 3122b650d953Smfreemon@cloudflare.com if (free_space > tp->rcv_ssthresh) { 3123b650d953Smfreemon@cloudflare.com free_space = tp->rcv_ssthresh; 3124b650d953Smfreemon@cloudflare.com /* new window should always be an exact multiple of scaling factor 3125b650d953Smfreemon@cloudflare.com * 3126b650d953Smfreemon@cloudflare.com * For this case, we ALIGN "up" (increase free_space) because 3127b650d953Smfreemon@cloudflare.com * we know free_space is not zero here, it has been reduced from 3128b650d953Smfreemon@cloudflare.com * the memory-based limit, and rcv_ssthresh is not a hard limit 3129b650d953Smfreemon@cloudflare.com * (unlike sk_rcvbuf). 3130b650d953Smfreemon@cloudflare.com */ 3131b650d953Smfreemon@cloudflare.com free_space = ALIGN(free_space, (1 << tp->rx_opt.rcv_wscale)); 3132b650d953Smfreemon@cloudflare.com } 3133b650d953Smfreemon@cloudflare.com 3134b650d953Smfreemon@cloudflare.com return free_space; 31351da177e4SLinus Torvalds } 31361da177e4SLinus Torvalds 3137cfea5a68SMartin KaFai Lau void tcp_skb_collapse_tstamp(struct sk_buff *skb, 3138082ac2d5SMartin KaFai Lau const struct sk_buff *next_skb) 3139082ac2d5SMartin KaFai Lau { 31400a2cf20cSSoheil Hassas Yeganeh if (unlikely(tcp_has_tx_tstamp(next_skb))) { 31410a2cf20cSSoheil Hassas Yeganeh const struct skb_shared_info *next_shinfo = 31420a2cf20cSSoheil Hassas Yeganeh skb_shinfo(next_skb); 3143082ac2d5SMartin KaFai Lau struct skb_shared_info *shinfo = skb_shinfo(skb); 3144082ac2d5SMartin KaFai Lau 31450a2cf20cSSoheil Hassas Yeganeh shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP; 3146082ac2d5SMartin KaFai Lau shinfo->tskey = next_shinfo->tskey; 31472de8023eSMartin KaFai Lau TCP_SKB_CB(skb)->txstamp_ack |= 31482de8023eSMartin KaFai Lau TCP_SKB_CB(next_skb)->txstamp_ack; 3149082ac2d5SMartin KaFai Lau } 3150082ac2d5SMartin KaFai Lau } 3151082ac2d5SMartin KaFai Lau 31524a17fc3aSIlpo Järvinen /* Collapses two adjacent SKB's during retransmission. */ 3153f8071cdeSEric Dumazet static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) 31541da177e4SLinus Torvalds { 31551da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 315675c119afSEric Dumazet struct sk_buff *next_skb = skb_rb_next(skb); 315713dde04fSWei Yongjun int next_skb_size; 31581da177e4SLinus Torvalds 3159058dc334SIlpo Järvinen next_skb_size = next_skb->len; 31601da177e4SLinus Torvalds 3161058dc334SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); 31621da177e4SLinus Torvalds 3163bd446314SEric Dumazet if (next_skb_size && !tcp_skb_shift(skb, next_skb, 1, next_skb_size)) 3164f8071cdeSEric Dumazet return false; 3165bd446314SEric Dumazet 31662b7cda9cSEric Dumazet tcp_highest_sack_replace(sk, next_skb, skb); 3167a6963a6bSIlpo Järvinen 31681da177e4SLinus Torvalds /* Update sequence range on original skb. */ 31691da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 31701da177e4SLinus Torvalds 3171e6c7d085SIlpo Järvinen /* Merge over control information. This moves PSH/FIN etc. over */ 31724de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; 31731da177e4SLinus Torvalds 31741da177e4SLinus Torvalds /* All done, get rid of second SKB and account for it so 31751da177e4SLinus Torvalds * packet counting does not break. 31761da177e4SLinus Torvalds */ 31774828e7f4SIlpo Järvinen TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; 3178a643b5d4SMartin KaFai Lau TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor; 3179b7689205SIlpo Järvinen 3180b7689205SIlpo Järvinen /* changed transmit queue under us so clear hints */ 3181ef9da47cSIlpo Järvinen tcp_clear_retrans_hints_partial(tp); 3182ef9da47cSIlpo Järvinen if (next_skb == tp->retransmit_skb_hint) 3183ef9da47cSIlpo Järvinen tp->retransmit_skb_hint = skb; 3184b7689205SIlpo Järvinen 3185797108d1SIlpo Järvinen tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); 3186797108d1SIlpo Järvinen 3187082ac2d5SMartin KaFai Lau tcp_skb_collapse_tstamp(skb, next_skb); 3188082ac2d5SMartin KaFai Lau 318975c119afSEric Dumazet tcp_rtx_queue_unlink_and_free(next_skb, sk); 3190f8071cdeSEric Dumazet return true; 31911da177e4SLinus Torvalds } 31921da177e4SLinus Torvalds 319367edfef7SAndi Kleen /* Check if coalescing SKBs is legal. */ 3194a2a385d6SEric Dumazet static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) 31954a17fc3aSIlpo Järvinen { 31964a17fc3aSIlpo Järvinen if (tcp_skb_pcount(skb) > 1) 3197a2a385d6SEric Dumazet return false; 31984a17fc3aSIlpo Järvinen if (skb_cloned(skb)) 3199a2a385d6SEric Dumazet return false; 32002331ccc5SEric Dumazet /* Some heuristics for collapsing over SACK'd could be invented */ 32014a17fc3aSIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 3202a2a385d6SEric Dumazet return false; 32034a17fc3aSIlpo Järvinen 3204a2a385d6SEric Dumazet return true; 32054a17fc3aSIlpo Järvinen } 32064a17fc3aSIlpo Järvinen 320767edfef7SAndi Kleen /* Collapse packets in the retransmit queue to make to create 320867edfef7SAndi Kleen * less packets on the wire. This is only done on retransmission. 320967edfef7SAndi Kleen */ 32104a17fc3aSIlpo Järvinen static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, 32114a17fc3aSIlpo Järvinen int space) 32124a17fc3aSIlpo Järvinen { 32134a17fc3aSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 32144a17fc3aSIlpo Järvinen struct sk_buff *skb = to, *tmp; 3215a2a385d6SEric Dumazet bool first = true; 32164a17fc3aSIlpo Järvinen 32171a63cb91SKuniyuki Iwashima if (!READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse)) 32184a17fc3aSIlpo Järvinen return; 32194de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 32204a17fc3aSIlpo Järvinen return; 32214a17fc3aSIlpo Järvinen 322275c119afSEric Dumazet skb_rbtree_walk_from_safe(skb, tmp) { 32234a17fc3aSIlpo Järvinen if (!tcp_can_collapse(sk, skb)) 32244a17fc3aSIlpo Järvinen break; 32254a17fc3aSIlpo Järvinen 322685712484SMat Martineau if (!tcp_skb_can_collapse(to, skb)) 3227a643b5d4SMartin KaFai Lau break; 3228a643b5d4SMartin KaFai Lau 32294a17fc3aSIlpo Järvinen space -= skb->len; 32304a17fc3aSIlpo Järvinen 32314a17fc3aSIlpo Järvinen if (first) { 3232a2a385d6SEric Dumazet first = false; 32334a17fc3aSIlpo Järvinen continue; 32344a17fc3aSIlpo Järvinen } 32354a17fc3aSIlpo Järvinen 32364a17fc3aSIlpo Järvinen if (space < 0) 32374a17fc3aSIlpo Järvinen break; 32384a17fc3aSIlpo Järvinen 32394a17fc3aSIlpo Järvinen if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) 32404a17fc3aSIlpo Järvinen break; 32414a17fc3aSIlpo Järvinen 3242f8071cdeSEric Dumazet if (!tcp_collapse_retrans(sk, to)) 3243f8071cdeSEric Dumazet break; 32444a17fc3aSIlpo Järvinen } 32454a17fc3aSIlpo Järvinen } 32464a17fc3aSIlpo Järvinen 32471da177e4SLinus Torvalds /* This retransmits one SKB. Policy decisions and retransmit queue 32481da177e4SLinus Torvalds * state updates are done by the caller. Returns non-zero if an 32491da177e4SLinus Torvalds * error occurred which prevented the send. 32501da177e4SLinus Torvalds */ 325110d3be56SEric Dumazet int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) 32521da177e4SLinus Torvalds { 32535d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 325410d3be56SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 32557d227cd2SSridhar Samudrala unsigned int cur_mss; 325610d3be56SEric Dumazet int diff, len, err; 3257536a6c8eSYonglong Li int avail_wnd; 325810d3be56SEric Dumazet 325910d3be56SEric Dumazet /* Inconclusive MTU probe */ 326010d3be56SEric Dumazet if (icsk->icsk_mtup.probe_size) 32615d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 32625d424d5aSJohn Heffner 32631f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 32641f3279aeSEric Dumazet return -EBUSY; 32651f3279aeSEric Dumazet 32661da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 32677f582b24SEric Dumazet if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) { 32687f582b24SEric Dumazet WARN_ON_ONCE(1); 32697f582b24SEric Dumazet return -EINVAL; 32707f582b24SEric Dumazet } 32711da177e4SLinus Torvalds if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 32721da177e4SLinus Torvalds return -ENOMEM; 32731da177e4SLinus Torvalds } 32741da177e4SLinus Torvalds 32757d227cd2SSridhar Samudrala if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 32767d227cd2SSridhar Samudrala return -EHOSTUNREACH; /* Routing failure or similar. */ 32777d227cd2SSridhar Samudrala 32780c54b85fSIlpo Järvinen cur_mss = tcp_current_mss(sk); 3279536a6c8eSYonglong Li avail_wnd = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 32807d227cd2SSridhar Samudrala 32811da177e4SLinus Torvalds /* If receiver has shrunk his window, and skb is out of 32821da177e4SLinus Torvalds * new window, do not retransmit it. The exception is the 32831da177e4SLinus Torvalds * case, when window is shrunk to zero. In this case 3284536a6c8eSYonglong Li * our retransmit of one segment serves as a zero window probe. 32851da177e4SLinus Torvalds */ 3286536a6c8eSYonglong Li if (avail_wnd <= 0) { 3287536a6c8eSYonglong Li if (TCP_SKB_CB(skb)->seq != tp->snd_una) 32881da177e4SLinus Torvalds return -EAGAIN; 3289536a6c8eSYonglong Li avail_wnd = cur_mss; 3290536a6c8eSYonglong Li } 32911da177e4SLinus Torvalds 329210d3be56SEric Dumazet len = cur_mss * segs; 3293536a6c8eSYonglong Li if (len > avail_wnd) { 3294536a6c8eSYonglong Li len = rounddown(avail_wnd, cur_mss); 3295536a6c8eSYonglong Li if (!len) 3296536a6c8eSYonglong Li len = avail_wnd; 3297536a6c8eSYonglong Li } 329810d3be56SEric Dumazet if (skb->len > len) { 329975c119afSEric Dumazet if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len, 330075c119afSEric Dumazet cur_mss, GFP_ATOMIC)) 33011da177e4SLinus Torvalds return -ENOMEM; /* We'll try again later. */ 330202276f3cSIlpo Järvinen } else { 3303c4777efaSEric Dumazet if (skb_unclone_keeptruesize(skb, GFP_ATOMIC)) 3304c52e2421SEric Dumazet return -ENOMEM; 330510d3be56SEric Dumazet 330610d3be56SEric Dumazet diff = tcp_skb_pcount(skb); 330710d3be56SEric Dumazet tcp_set_skb_tso_segs(skb, cur_mss); 330810d3be56SEric Dumazet diff -= tcp_skb_pcount(skb); 330910d3be56SEric Dumazet if (diff) 331010d3be56SEric Dumazet tcp_adjust_pcount(sk, skb, diff); 3311536a6c8eSYonglong Li avail_wnd = min_t(int, avail_wnd, cur_mss); 3312536a6c8eSYonglong Li if (skb->len < avail_wnd) 3313536a6c8eSYonglong Li tcp_retrans_try_collapse(sk, skb, avail_wnd); 33141da177e4SLinus Torvalds } 33151da177e4SLinus Torvalds 331649213555SDaniel Borkmann /* RFC3168, section 6.1.1.1. ECN fallback */ 331749213555SDaniel Borkmann if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN) 331849213555SDaniel Borkmann tcp_ecn_clear_syn(sk, skb); 331949213555SDaniel Borkmann 3320678550c6SYuchung Cheng /* Update global and local TCP statistics. */ 3321678550c6SYuchung Cheng segs = tcp_skb_pcount(skb); 3322678550c6SYuchung Cheng TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs); 3323678550c6SYuchung Cheng if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 3324678550c6SYuchung Cheng __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 3325678550c6SYuchung Cheng tp->total_retrans += segs; 3326fb31c9b9SWei Wang tp->bytes_retrans += skb->len; 3327678550c6SYuchung Cheng 332850bceae9SThomas Graf /* make sure skb->data is aligned on arches that require it 332950bceae9SThomas Graf * and check if ack-trimming & collapsing extended the headroom 333050bceae9SThomas Graf * beyond what csum_start can cover. 333150bceae9SThomas Graf */ 333250bceae9SThomas Graf if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) || 333350bceae9SThomas Graf skb_headroom(skb) >= 0xFFFF)) { 333410a81980SEric Dumazet struct sk_buff *nskb; 333510a81980SEric Dumazet 3336e2080072SEric Dumazet tcp_skb_tsorted_save(skb) { 333710a81980SEric Dumazet nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC); 333807f8e4d0SFlorian Westphal if (nskb) { 333907f8e4d0SFlorian Westphal nskb->dev = NULL; 334007f8e4d0SFlorian Westphal err = tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC); 334107f8e4d0SFlorian Westphal } else { 334207f8e4d0SFlorian Westphal err = -ENOBUFS; 334307f8e4d0SFlorian Westphal } 3344e2080072SEric Dumazet } tcp_skb_tsorted_restore(skb); 3345e2080072SEric Dumazet 33465889e2c0SYousuk Seung if (!err) { 3347a7a25630SEric Dumazet tcp_update_skb_after_send(sk, skb, tp->tcp_wstamp_ns); 33485889e2c0SYousuk Seung tcp_rate_skb_sent(sk, skb); 33495889e2c0SYousuk Seung } 3350117632e6SEric Dumazet } else { 3351c84a5711SYuchung Cheng err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 3352117632e6SEric Dumazet } 3353c84a5711SYuchung Cheng 33547f12422cSYuchung Cheng /* To avoid taking spuriously low RTT samples based on a timestamp 33557f12422cSYuchung Cheng * for a transmit that never happened, always mark EVER_RETRANS 33567f12422cSYuchung Cheng */ 33577f12422cSYuchung Cheng TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS; 33587f12422cSYuchung Cheng 3359a31ad29eSLawrence Brakmo if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RETRANS_CB_FLAG)) 3360a31ad29eSLawrence Brakmo tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RETRANS_CB, 3361a31ad29eSLawrence Brakmo TCP_SKB_CB(skb)->seq, segs, err); 3362a31ad29eSLawrence Brakmo 3363fc9f3501SEric Dumazet if (likely(!err)) { 3364e086101bSCong Wang trace_tcp_retransmit_skb(sk, skb); 3365678550c6SYuchung Cheng } else if (err != -EBUSY) { 3366ec641b39SYuchung Cheng NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL, segs); 3367fc9f3501SEric Dumazet } 3368c84a5711SYuchung Cheng return err; 336993b174adSYuchung Cheng } 337093b174adSYuchung Cheng 337110d3be56SEric Dumazet int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) 337293b174adSYuchung Cheng { 337393b174adSYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 337410d3be56SEric Dumazet int err = __tcp_retransmit_skb(sk, skb, segs); 33751da177e4SLinus Torvalds 33761da177e4SLinus Torvalds if (err == 0) { 33771da177e4SLinus Torvalds #if FASTRETRANS_DEBUG > 0 33781da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 3379e87cc472SJoe Perches net_dbg_ratelimited("retrans_out leaked\n"); 33801da177e4SLinus Torvalds } 33811da177e4SLinus Torvalds #endif 33821da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 33831da177e4SLinus Torvalds tp->retrans_out += tcp_skb_pcount(skb); 33847ae18975SYuchung Cheng } 33851da177e4SLinus Torvalds 33867ae18975SYuchung Cheng /* Save stamp of the first (attempted) retransmit. */ 33871da177e4SLinus Torvalds if (!tp->retrans_stamp) 33887faee5c0SEric Dumazet tp->retrans_stamp = tcp_skb_timestamp(skb); 33891da177e4SLinus Torvalds 33906e08d5e3SYuchung Cheng if (tp->undo_retrans < 0) 33916e08d5e3SYuchung Cheng tp->undo_retrans = 0; 33926e08d5e3SYuchung Cheng tp->undo_retrans += tcp_skb_pcount(skb); 33931da177e4SLinus Torvalds return err; 33941da177e4SLinus Torvalds } 33951da177e4SLinus Torvalds 33961da177e4SLinus Torvalds /* This gets called after a retransmit timeout, and the initially 33971da177e4SLinus Torvalds * retransmitted data is acknowledged. It tries to continue 33981da177e4SLinus Torvalds * resending the rest of the retransmit queue, until either 33991da177e4SLinus Torvalds * we've sent it all or the congestion window limit is reached. 34001da177e4SLinus Torvalds */ 34011da177e4SLinus Torvalds void tcp_xmit_retransmit_queue(struct sock *sk) 34021da177e4SLinus Torvalds { 34036687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 3404b9f1f1ceSEric Dumazet struct sk_buff *skb, *rtx_head, *hole = NULL; 34051da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 3406916e6d1aSEric Dumazet bool rearm_timer = false; 3407840a3cbeSYuchung Cheng u32 max_segs; 340861eb55f4SIlpo Järvinen int mib_idx; 34096a438bbeSStephen Hemminger 341045e77d31SIlpo Järvinen if (!tp->packets_out) 341145e77d31SIlpo Järvinen return; 341245e77d31SIlpo Järvinen 341375c119afSEric Dumazet rtx_head = tcp_rtx_queue_head(sk); 3414b9f1f1ceSEric Dumazet skb = tp->retransmit_skb_hint ?: rtx_head; 3415ed6e7268SNeal Cardwell max_segs = tcp_tso_segs(sk, tcp_current_mss(sk)); 341675c119afSEric Dumazet skb_rbtree_walk_from(skb) { 3417dca0aaf8SEric Dumazet __u8 sacked; 341810d3be56SEric Dumazet int segs; 34191da177e4SLinus Torvalds 3420218af599SEric Dumazet if (tcp_pacing_check(sk)) 3421218af599SEric Dumazet break; 3422218af599SEric Dumazet 34236a438bbeSStephen Hemminger /* we could do better than to assign each time */ 342451456b29SIan Morris if (!hole) 34256a438bbeSStephen Hemminger tp->retransmit_skb_hint = skb; 34266a438bbeSStephen Hemminger 342740570375SEric Dumazet segs = tcp_snd_cwnd(tp) - tcp_packets_in_flight(tp); 342810d3be56SEric Dumazet if (segs <= 0) 3429916e6d1aSEric Dumazet break; 3430dca0aaf8SEric Dumazet sacked = TCP_SKB_CB(skb)->sacked; 3431a3d2e9f8SEric Dumazet /* In case tcp_shift_skb_data() have aggregated large skbs, 3432a3d2e9f8SEric Dumazet * we need to make sure not sending too bigs TSO packets 3433a3d2e9f8SEric Dumazet */ 3434a3d2e9f8SEric Dumazet segs = min_t(int, segs, max_segs); 34350e1c54c2SIlpo Järvinen 3436840a3cbeSYuchung Cheng if (tp->retrans_out >= tp->lost_out) { 3437006f582cSIlpo Järvinen break; 34380e1c54c2SIlpo Järvinen } else if (!(sacked & TCPCB_LOST)) { 343951456b29SIan Morris if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) 34400e1c54c2SIlpo Järvinen hole = skb; 344161eb55f4SIlpo Järvinen continue; 34421da177e4SLinus Torvalds 34430e1c54c2SIlpo Järvinen } else { 34440e1c54c2SIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Loss) 34450e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFASTRETRANS; 34460e1c54c2SIlpo Järvinen else 34470e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; 34480e1c54c2SIlpo Järvinen } 34490e1c54c2SIlpo Järvinen 34500e1c54c2SIlpo Järvinen if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) 345161eb55f4SIlpo Järvinen continue; 345240b215e5SPavel Emelyanov 3453f9616c35SEric Dumazet if (tcp_small_queue_check(sk, skb, 1)) 3454916e6d1aSEric Dumazet break; 3455f9616c35SEric Dumazet 345610d3be56SEric Dumazet if (tcp_retransmit_skb(sk, skb, segs)) 3457916e6d1aSEric Dumazet break; 345824ab6becSYuchung Cheng 3459de1d6578SYuchung Cheng NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb)); 34601da177e4SLinus Torvalds 3461684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 3462a262f0cdSNandita Dukkipati tp->prr_out += tcp_skb_pcount(skb); 3463a262f0cdSNandita Dukkipati 346475c119afSEric Dumazet if (skb == rtx_head && 346557dde7f7SYuchung Cheng icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT) 3466916e6d1aSEric Dumazet rearm_timer = true; 3467916e6d1aSEric Dumazet 3468916e6d1aSEric Dumazet } 3469916e6d1aSEric Dumazet if (rearm_timer) 34703f80e08fSEric Dumazet tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 34713f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, 34728dc242adSEric Dumazet TCP_RTO_MAX); 34731da177e4SLinus Torvalds } 34741da177e4SLinus Torvalds 3475d83769a5SEric Dumazet /* We allow to exceed memory limits for FIN packets to expedite 3476d83769a5SEric Dumazet * connection tear down and (memory) recovery. 3477845704a5SEric Dumazet * Otherwise tcp_send_fin() could be tempted to either delay FIN 3478845704a5SEric Dumazet * or even be forced to close flow without any FIN. 3479a6c5ea4cSEric Dumazet * In general, we want to allow one skb per socket to avoid hangs 3480a6c5ea4cSEric Dumazet * with edge trigger epoll() 3481d83769a5SEric Dumazet */ 3482a6c5ea4cSEric Dumazet void sk_forced_mem_schedule(struct sock *sk, int size) 3483d83769a5SEric Dumazet { 3484c4ee1185SEric Dumazet int delta, amt; 3485d83769a5SEric Dumazet 3486c4ee1185SEric Dumazet delta = size - sk->sk_forward_alloc; 3487c4ee1185SEric Dumazet if (delta <= 0) 3488d83769a5SEric Dumazet return; 3489c4ee1185SEric Dumazet amt = sk_mem_pages(delta); 34905e6300e7SEric Dumazet sk_forward_alloc_add(sk, amt << PAGE_SHIFT); 3491e805605cSJohannes Weiner sk_memory_allocated_add(sk, amt); 3492e805605cSJohannes Weiner 3493baac50bbSJohannes Weiner if (mem_cgroup_sockets_enabled && sk->sk_memcg) 34944b1327beSWei Wang mem_cgroup_charge_skmem(sk->sk_memcg, amt, 34954b1327beSWei Wang gfp_memcg_charge() | __GFP_NOFAIL); 3496d83769a5SEric Dumazet } 3497d83769a5SEric Dumazet 3498845704a5SEric Dumazet /* Send a FIN. The caller locks the socket for us. 3499845704a5SEric Dumazet * We should try to send a FIN packet really hard, but eventually give up. 35001da177e4SLinus Torvalds */ 35011da177e4SLinus Torvalds void tcp_send_fin(struct sock *sk) 35021da177e4SLinus Torvalds { 3503ee2aabd3SEric Dumazet struct sk_buff *skb, *tskb, *tail = tcp_write_queue_tail(sk); 35041da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 35051da177e4SLinus Torvalds 3506845704a5SEric Dumazet /* Optimization, tack on the FIN if we have one skb in write queue and 3507845704a5SEric Dumazet * this skb was not yet sent, or we are under memory pressure. 3508845704a5SEric Dumazet * Note: in the latter case, FIN packet will be sent after a timeout, 3509845704a5SEric Dumazet * as TCP stack thinks it has already been transmitted. 35101da177e4SLinus Torvalds */ 3511ee2aabd3SEric Dumazet tskb = tail; 351275c119afSEric Dumazet if (!tskb && tcp_under_memory_pressure(sk)) 351375c119afSEric Dumazet tskb = skb_rb_last(&sk->tcp_rtx_queue); 351475c119afSEric Dumazet 351575c119afSEric Dumazet if (tskb) { 3516845704a5SEric Dumazet TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN; 3517845704a5SEric Dumazet TCP_SKB_CB(tskb)->end_seq++; 35181da177e4SLinus Torvalds tp->write_seq++; 3519ee2aabd3SEric Dumazet if (!tail) { 3520845704a5SEric Dumazet /* This means tskb was already sent. 3521845704a5SEric Dumazet * Pretend we included the FIN on previous transmit. 3522845704a5SEric Dumazet * We need to set tp->snd_nxt to the value it would have 3523845704a5SEric Dumazet * if FIN had been sent. This is because retransmit path 3524845704a5SEric Dumazet * does not change tp->snd_nxt. 3525845704a5SEric Dumazet */ 3526e0d694d6SEric Dumazet WRITE_ONCE(tp->snd_nxt, tp->snd_nxt + 1); 3527845704a5SEric Dumazet return; 3528845704a5SEric Dumazet } 35291da177e4SLinus Torvalds } else { 3530845704a5SEric Dumazet skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation); 3531d1edc085SColin Ian King if (unlikely(!skb)) 3532845704a5SEric Dumazet return; 3533d1edc085SColin Ian King 3534e2080072SEric Dumazet INIT_LIST_HEAD(&skb->tcp_tsorted_anchor); 3535d83769a5SEric Dumazet skb_reserve(skb, MAX_TCP_HEADER); 3536a6c5ea4cSEric Dumazet sk_forced_mem_schedule(sk, skb->truesize); 35371da177e4SLinus Torvalds /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 3538e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tp->write_seq, 3539a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_FIN); 35401da177e4SLinus Torvalds tcp_queue_skb(sk, skb); 35411da177e4SLinus Torvalds } 3542845704a5SEric Dumazet __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF); 35431da177e4SLinus Torvalds } 35441da177e4SLinus Torvalds 35451da177e4SLinus Torvalds /* We get here when a process closes a file descriptor (either due to 35461da177e4SLinus Torvalds * an explicit close() or as a byproduct of exit()'ing) and there 35471da177e4SLinus Torvalds * was unread data in the receive queue. This behavior is recommended 354865bb723cSGerrit Renker * by RFC 2525, section 2.17. -DaveM 35491da177e4SLinus Torvalds */ 3550dd0fc66fSAl Viro void tcp_send_active_reset(struct sock *sk, gfp_t priority) 35511da177e4SLinus Torvalds { 35521da177e4SLinus Torvalds struct sk_buff *skb; 35531da177e4SLinus Torvalds 35547cc2b043SGao Feng TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); 35557cc2b043SGao Feng 35561da177e4SLinus Torvalds /* NOTE: No TCP options attached and we never retransmit this. */ 35571da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, priority); 35581da177e4SLinus Torvalds if (!skb) { 35594e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 35601da177e4SLinus Torvalds return; 35611da177e4SLinus Torvalds } 35621da177e4SLinus Torvalds 35631da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 35641da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 3565e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), 3566a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_RST); 35679a568de4SEric Dumazet tcp_mstamp_refresh(tcp_sk(sk)); 35681da177e4SLinus Torvalds /* Send it off. */ 3569dfb4b9dcSDavid S. Miller if (tcp_transmit_skb(sk, skb, 0, priority)) 35704e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 3571c24b14c4SSong Liu 3572c24b14c4SSong Liu /* skb of trace_tcp_send_reset() keeps the skb that caused RST, 3573c24b14c4SSong Liu * skb here is different to the troublesome skb, so use NULL 3574c24b14c4SSong Liu */ 3575c24b14c4SSong Liu trace_tcp_send_reset(sk, NULL); 35761da177e4SLinus Torvalds } 35771da177e4SLinus Torvalds 357867edfef7SAndi Kleen /* Send a crossed SYN-ACK during socket establishment. 357967edfef7SAndi Kleen * WARNING: This routine must only be called when we have already sent 35801da177e4SLinus Torvalds * a SYN packet that crossed the incoming SYN that caused this routine 35811da177e4SLinus Torvalds * to get called. If this assumption fails then the initial rcv_wnd 35821da177e4SLinus Torvalds * and rcv_wscale values will not be correct. 35831da177e4SLinus Torvalds */ 35841da177e4SLinus Torvalds int tcp_send_synack(struct sock *sk) 35851da177e4SLinus Torvalds { 35861da177e4SLinus Torvalds struct sk_buff *skb; 35871da177e4SLinus Torvalds 358875c119afSEric Dumazet skb = tcp_rtx_queue_head(sk); 358951456b29SIan Morris if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 359075c119afSEric Dumazet pr_err("%s: wrong queue state\n", __func__); 35911da177e4SLinus Torvalds return -EFAULT; 35921da177e4SLinus Torvalds } 35934de075e0SEric Dumazet if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { 35941da177e4SLinus Torvalds if (skb_cloned(skb)) { 3595e2080072SEric Dumazet struct sk_buff *nskb; 3596e2080072SEric Dumazet 3597e2080072SEric Dumazet tcp_skb_tsorted_save(skb) { 3598e2080072SEric Dumazet nskb = skb_copy(skb, GFP_ATOMIC); 3599e2080072SEric Dumazet } tcp_skb_tsorted_restore(skb); 360051456b29SIan Morris if (!nskb) 36011da177e4SLinus Torvalds return -ENOMEM; 3602e2080072SEric Dumazet INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor); 36032bec445fSEric Dumazet tcp_highest_sack_replace(sk, skb, nskb); 360475c119afSEric Dumazet tcp_rtx_queue_unlink_and_free(skb, sk); 3605f4a775d1SEric Dumazet __skb_header_release(nskb); 360675c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb); 3607ab4e846aSEric Dumazet sk_wmem_queued_add(sk, nskb->truesize); 36083ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 36091da177e4SLinus Torvalds skb = nskb; 36101da177e4SLinus Torvalds } 36111da177e4SLinus Torvalds 36124de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; 3613735d3831SFlorian Westphal tcp_ecn_send_synack(sk, skb); 36141da177e4SLinus Torvalds } 3615dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 36161da177e4SLinus Torvalds } 36171da177e4SLinus Torvalds 36184aea39c1SEric Dumazet /** 3619331fca43SMartin KaFai Lau * tcp_make_synack - Allocate one skb and build a SYNACK packet. 3620331fca43SMartin KaFai Lau * @sk: listener socket 3621331fca43SMartin KaFai Lau * @dst: dst entry attached to the SYNACK. It is consumed and caller 3622331fca43SMartin KaFai Lau * should not use it again. 3623331fca43SMartin KaFai Lau * @req: request_sock pointer 3624331fca43SMartin KaFai Lau * @foc: cookie for tcp fast open 3625331fca43SMartin KaFai Lau * @synack_type: Type of synack to prepare 3626331fca43SMartin KaFai Lau * @syn_skb: SYN packet just received. It could be NULL for rtx case. 36274aea39c1SEric Dumazet */ 36285d062de7SEric Dumazet struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst, 3629e6b4d113SWilliam Allen Simpson struct request_sock *req, 3630ca6fb065SEric Dumazet struct tcp_fastopen_cookie *foc, 3631331fca43SMartin KaFai Lau enum tcp_synack_type synack_type, 3632331fca43SMartin KaFai Lau struct sk_buff *syn_skb) 36331da177e4SLinus Torvalds { 36342e6599cbSArnaldo Carvalho de Melo struct inet_request_sock *ireq = inet_rsk(req); 36355d062de7SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 363680f03e27SEric Dumazet struct tcp_md5sig_key *md5 = NULL; 36375d062de7SEric Dumazet struct tcp_out_options opts; 36385d062de7SEric Dumazet struct sk_buff *skb; 3639bd0388aeSWilliam Allen Simpson int tcp_header_size; 36405d062de7SEric Dumazet struct tcphdr *th; 3641f5fff5dcSTom Quetchenbach int mss; 3642a842fe14SEric Dumazet u64 now; 36431da177e4SLinus Torvalds 3644ca6fb065SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 36454aea39c1SEric Dumazet if (unlikely(!skb)) { 36464aea39c1SEric Dumazet dst_release(dst); 36471da177e4SLinus Torvalds return NULL; 36484aea39c1SEric Dumazet } 36491da177e4SLinus Torvalds /* Reserve space for headers. */ 36501da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 36511da177e4SLinus Torvalds 3652b3d05147SEric Dumazet switch (synack_type) { 3653b3d05147SEric Dumazet case TCP_SYNACK_NORMAL: 36549e17f8a4SEric Dumazet skb_set_owner_w(skb, req_to_sk(req)); 3655b3d05147SEric Dumazet break; 3656b3d05147SEric Dumazet case TCP_SYNACK_COOKIE: 3657b3d05147SEric Dumazet /* Under synflood, we do not attach skb to a socket, 3658b3d05147SEric Dumazet * to avoid false sharing. 3659b3d05147SEric Dumazet */ 3660b3d05147SEric Dumazet break; 3661b3d05147SEric Dumazet case TCP_SYNACK_FASTOPEN: 3662ca6fb065SEric Dumazet /* sk is a const pointer, because we want to express multiple 3663ca6fb065SEric Dumazet * cpu might call us concurrently. 3664ca6fb065SEric Dumazet * sk->sk_wmem_alloc in an atomic, we can promote to rw. 3665ca6fb065SEric Dumazet */ 3666ca6fb065SEric Dumazet skb_set_owner_w(skb, (struct sock *)sk); 3667b3d05147SEric Dumazet break; 3668ca6fb065SEric Dumazet } 36694aea39c1SEric Dumazet skb_dst_set(skb, dst); 36701da177e4SLinus Torvalds 36713541f9e8SEric Dumazet mss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 3672f5fff5dcSTom Quetchenbach 367333ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 3674a842fe14SEric Dumazet now = tcp_clock_ns(); 36758b5f12d0SFlorian Westphal #ifdef CONFIG_SYN_COOKIES 3676f8ace8d9SFlorian Westphal if (unlikely(synack_type == TCP_SYNACK_COOKIE && ireq->tstamp_ok)) 3677a1ac9c8aSMartin KaFai Lau skb_set_delivery_time(skb, cookie_init_timestamp(req, now), 3678a1ac9c8aSMartin KaFai Lau true); 36798b5f12d0SFlorian Westphal else 36808b5f12d0SFlorian Westphal #endif 36819e450c1eSYuchung Cheng { 3682a1ac9c8aSMartin KaFai Lau skb_set_delivery_time(skb, now, true); 36839e450c1eSYuchung Cheng if (!tcp_rsk(req)->snt_synack) /* Timestamp first SYNACK */ 36849e450c1eSYuchung Cheng tcp_rsk(req)->snt_synack = tcp_skb_timestamp_us(skb); 36859e450c1eSYuchung Cheng } 368680f03e27SEric Dumazet 368780f03e27SEric Dumazet #ifdef CONFIG_TCP_MD5SIG 368880f03e27SEric Dumazet rcu_read_lock(); 3689fd3a154aSEric Dumazet md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req)); 369080f03e27SEric Dumazet #endif 36915e526552SEric Dumazet skb_set_hash(skb, READ_ONCE(tcp_rsk(req)->txhash), PKT_HASH_TYPE_L4); 3692331fca43SMartin KaFai Lau /* bpf program will be interested in the tcp_flags */ 3693331fca43SMartin KaFai Lau TCP_SKB_CB(skb)->tcp_flags = TCPHDR_SYN | TCPHDR_ACK; 369460e2a778SUrsula Braun tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5, 3695331fca43SMartin KaFai Lau foc, synack_type, 3696331fca43SMartin KaFai Lau syn_skb) + sizeof(*th); 369733ad798cSAdam Langley 3698aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 3699aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 37001da177e4SLinus Torvalds 3701ea1627c2SEric Dumazet th = (struct tcphdr *)skb->data; 37021da177e4SLinus Torvalds memset(th, 0, sizeof(struct tcphdr)); 37031da177e4SLinus Torvalds th->syn = 1; 37041da177e4SLinus Torvalds th->ack = 1; 37056ac705b1SEric Dumazet tcp_ecn_make_synack(req, th); 3706b44084c2SEric Dumazet th->source = htons(ireq->ir_num); 3707634fb979SEric Dumazet th->dest = ireq->ir_rmt_port; 3708e05a90ecSJamal Hadi Salim skb->mark = ireq->ir_mark; 37093b117750SEric Dumazet skb->ip_summed = CHECKSUM_PARTIAL; 37103b117750SEric Dumazet th->seq = htonl(tcp_rsk(req)->snt_isn); 37118336886fSJerry Chu /* XXX data is queued and acked as is. No buffer/window check */ 37128336886fSJerry Chu th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt); 37131da177e4SLinus Torvalds 37141da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 3715ed53d0abSEric Dumazet th->window = htons(min(req->rsk_rcv_wnd, 65535U)); 3716ea66758cSPaolo Abeni tcp_options_write(th, NULL, &opts); 37171da177e4SLinus Torvalds th->doff = (tcp_header_size >> 2); 3718bced3f7dSBreno Leitao TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS); 3719cfb6eeb4SYOSHIFUJI Hideaki 3720cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 3721cfb6eeb4SYOSHIFUJI Hideaki /* Okay, we have all we need - do the md5 hash if needed */ 372280f03e27SEric Dumazet if (md5) 3723bd0388aeSWilliam Allen Simpson tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, 372439f8e58eSEric Dumazet md5, req_to_sk(req), skb); 372580f03e27SEric Dumazet rcu_read_unlock(); 3726cfb6eeb4SYOSHIFUJI Hideaki #endif 3727cfb6eeb4SYOSHIFUJI Hideaki 3728331fca43SMartin KaFai Lau bpf_skops_write_hdr_opt((struct sock *)sk, skb, req, syn_skb, 3729331fca43SMartin KaFai Lau synack_type, &opts); 3730331fca43SMartin KaFai Lau 3731a1ac9c8aSMartin KaFai Lau skb_set_delivery_time(skb, now, true); 3732a842fe14SEric Dumazet tcp_add_tx_delay(skb, tp); 3733a842fe14SEric Dumazet 37341da177e4SLinus Torvalds return skb; 37351da177e4SLinus Torvalds } 37364bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_make_synack); 37371da177e4SLinus Torvalds 373881164413SDaniel Borkmann static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst) 373981164413SDaniel Borkmann { 374081164413SDaniel Borkmann struct inet_connection_sock *icsk = inet_csk(sk); 374181164413SDaniel Borkmann const struct tcp_congestion_ops *ca; 374281164413SDaniel Borkmann u32 ca_key = dst_metric(dst, RTAX_CC_ALGO); 374381164413SDaniel Borkmann 374481164413SDaniel Borkmann if (ca_key == TCP_CA_UNSPEC) 374581164413SDaniel Borkmann return; 374681164413SDaniel Borkmann 374781164413SDaniel Borkmann rcu_read_lock(); 374881164413SDaniel Borkmann ca = tcp_ca_find_key(ca_key); 37490baf26b0SMartin KaFai Lau if (likely(ca && bpf_try_module_get(ca, ca->owner))) { 37500baf26b0SMartin KaFai Lau bpf_module_put(icsk->icsk_ca_ops, icsk->icsk_ca_ops->owner); 375181164413SDaniel Borkmann icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst); 375281164413SDaniel Borkmann icsk->icsk_ca_ops = ca; 375381164413SDaniel Borkmann } 375481164413SDaniel Borkmann rcu_read_unlock(); 375581164413SDaniel Borkmann } 375681164413SDaniel Borkmann 375767edfef7SAndi Kleen /* Do all connect socket setups that can be done AF independent. */ 3758f7e56a76Sstephen hemminger static void tcp_connect_init(struct sock *sk) 37591da177e4SLinus Torvalds { 3760cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 37611da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 37621da177e4SLinus Torvalds __u8 rcv_wscale; 376313d3b1ebSLawrence Brakmo u32 rcv_wnd; 37641da177e4SLinus Torvalds 37651da177e4SLinus Torvalds /* We'll fix this up when we get a response from the other end. 37661da177e4SLinus Torvalds * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 37671da177e4SLinus Torvalds */ 37685d2ed052SEric Dumazet tp->tcp_header_len = sizeof(struct tcphdr); 37693666f666SKuniyuki Iwashima if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_timestamps)) 37705d2ed052SEric Dumazet tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED; 37711da177e4SLinus Torvalds 37721da177e4SLinus Torvalds /* If user gave his TCP_MAXSEG, record it to clamp */ 37731da177e4SLinus Torvalds if (tp->rx_opt.user_mss) 37741da177e4SLinus Torvalds tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 37751da177e4SLinus Torvalds tp->max_window = 0; 37765d424d5aSJohn Heffner tcp_mtup_init(sk); 37771da177e4SLinus Torvalds tcp_sync_mss(sk, dst_mtu(dst)); 37781da177e4SLinus Torvalds 377981164413SDaniel Borkmann tcp_ca_dst_init(sk, dst); 378081164413SDaniel Borkmann 37811da177e4SLinus Torvalds if (!tp->window_clamp) 37821da177e4SLinus Torvalds tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 37833541f9e8SEric Dumazet tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 3784f5fff5dcSTom Quetchenbach 37851da177e4SLinus Torvalds tcp_initialize_rcv_mss(sk); 37861da177e4SLinus Torvalds 3787e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 3788e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 3789e88c64f0SHagen Paul Pfeifer (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) 3790e88c64f0SHagen Paul Pfeifer tp->window_clamp = tcp_full_space(sk); 3791e88c64f0SHagen Paul Pfeifer 379213d3b1ebSLawrence Brakmo rcv_wnd = tcp_rwnd_init_bpf(sk); 379313d3b1ebSLawrence Brakmo if (rcv_wnd == 0) 379413d3b1ebSLawrence Brakmo rcv_wnd = dst_metric(dst, RTAX_INITRWND); 379513d3b1ebSLawrence Brakmo 3796ceef9ab6SEric Dumazet tcp_select_initial_window(sk, tcp_full_space(sk), 37971da177e4SLinus Torvalds tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 37981da177e4SLinus Torvalds &tp->rcv_wnd, 37991da177e4SLinus Torvalds &tp->window_clamp, 38003666f666SKuniyuki Iwashima READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_window_scaling), 380131d12926Slaurent chavey &rcv_wscale, 380213d3b1ebSLawrence Brakmo rcv_wnd); 38031da177e4SLinus Torvalds 38041da177e4SLinus Torvalds tp->rx_opt.rcv_wscale = rcv_wscale; 38051da177e4SLinus Torvalds tp->rcv_ssthresh = tp->rcv_wnd; 38061da177e4SLinus Torvalds 3807e13ec3daSEric Dumazet WRITE_ONCE(sk->sk_err, 0); 38081da177e4SLinus Torvalds sock_reset_flag(sk, SOCK_DONE); 38091da177e4SLinus Torvalds tp->snd_wnd = 0; 3810ee7537b6SHantzis Fotis tcp_init_wl(tp, 0); 38117f582b24SEric Dumazet tcp_write_queue_purge(sk); 38121da177e4SLinus Torvalds tp->snd_una = tp->write_seq; 38131da177e4SLinus Torvalds tp->snd_sml = tp->write_seq; 381433f5f57eSIlpo Järvinen tp->snd_up = tp->write_seq; 3815e0d694d6SEric Dumazet WRITE_ONCE(tp->snd_nxt, tp->write_seq); 3816ee995283SPavel Emelyanov 3817ee995283SPavel Emelyanov if (likely(!tp->repair)) 38181da177e4SLinus Torvalds tp->rcv_nxt = 0; 3819c7781a6eSAndrew Vagin else 382070eabf0eSEric Dumazet tp->rcv_tstamp = tcp_jiffies32; 3821ee995283SPavel Emelyanov tp->rcv_wup = tp->rcv_nxt; 38227db48e98SEric Dumazet WRITE_ONCE(tp->copied_seq, tp->rcv_nxt); 38231da177e4SLinus Torvalds 38248550f328SLawrence Brakmo inet_csk(sk)->icsk_rto = tcp_timeout_init(sk); 3825463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_retransmits = 0; 38261da177e4SLinus Torvalds tcp_clear_retrans(tp); 38271da177e4SLinus Torvalds } 38281da177e4SLinus Torvalds 3829783237e8SYuchung Cheng static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) 3830783237e8SYuchung Cheng { 3831783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3832783237e8SYuchung Cheng struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 3833783237e8SYuchung Cheng 3834783237e8SYuchung Cheng tcb->end_seq += skb->len; 3835f4a775d1SEric Dumazet __skb_header_release(skb); 3836ab4e846aSEric Dumazet sk_wmem_queued_add(sk, skb->truesize); 3837783237e8SYuchung Cheng sk_mem_charge(sk, skb->truesize); 38380f317464SEric Dumazet WRITE_ONCE(tp->write_seq, tcb->end_seq); 3839783237e8SYuchung Cheng tp->packets_out += tcp_skb_pcount(skb); 3840783237e8SYuchung Cheng } 3841783237e8SYuchung Cheng 3842783237e8SYuchung Cheng /* Build and send a SYN with data and (cached) Fast Open cookie. However, 3843783237e8SYuchung Cheng * queue a data-only packet after the regular SYN, such that regular SYNs 3844783237e8SYuchung Cheng * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges 3845783237e8SYuchung Cheng * only the SYN sequence, the data are retransmitted in the first ACK. 3846783237e8SYuchung Cheng * If cookie is not cached or other error occurs, falls back to send a 3847783237e8SYuchung Cheng * regular SYN with Fast Open cookie request option. 3848783237e8SYuchung Cheng */ 3849783237e8SYuchung Cheng static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) 3850783237e8SYuchung Cheng { 3851ed0c99dcSJakub Kicinski struct inet_connection_sock *icsk = inet_csk(sk); 3852783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3853783237e8SYuchung Cheng struct tcp_fastopen_request *fo = tp->fastopen_req; 3854fbf93406SEric Dumazet struct page_frag *pfrag = sk_page_frag(sk); 3855355a901eSEric Dumazet struct sk_buff *syn_data; 3856fbf93406SEric Dumazet int space, err = 0; 3857783237e8SYuchung Cheng 385867da22d2SYuchung Cheng tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ 3859065263f4SWei Wang if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie)) 3860783237e8SYuchung Cheng goto fallback; 3861783237e8SYuchung Cheng 3862783237e8SYuchung Cheng /* MSS for SYN-data is based on cached MSS and bounded by PMTU and 3863783237e8SYuchung Cheng * user-MSS. Reserve maximum option space for middleboxes that add 3864783237e8SYuchung Cheng * private TCP options. The cost is reduced data space in SYN :( 3865783237e8SYuchung Cheng */ 38663541f9e8SEric Dumazet tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp); 3867ed0c99dcSJakub Kicinski /* Sync mss_cache after updating the mss_clamp */ 3868ed0c99dcSJakub Kicinski tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 38693541f9e8SEric Dumazet 3870ed0c99dcSJakub Kicinski space = __tcp_mtu_to_mss(sk, icsk->icsk_pmtu_cookie) - 3871783237e8SYuchung Cheng MAX_TCP_OPTION_SPACE; 3872783237e8SYuchung Cheng 3873f5ddcbbbSEric Dumazet space = min_t(size_t, space, fo->size); 3874f5ddcbbbSEric Dumazet 3875fbf93406SEric Dumazet if (space && 3876fbf93406SEric Dumazet !skb_page_frag_refill(min_t(size_t, space, PAGE_SIZE), 3877fbf93406SEric Dumazet pfrag, sk->sk_allocation)) 3878fbf93406SEric Dumazet goto fallback; 38795882efffSEric Dumazet syn_data = tcp_stream_alloc_skb(sk, sk->sk_allocation, false); 3880355a901eSEric Dumazet if (!syn_data) 3881783237e8SYuchung Cheng goto fallback; 3882355a901eSEric Dumazet memcpy(syn_data->cb, syn->cb, sizeof(syn->cb)); 388307e100f9SEric Dumazet if (space) { 3884fbf93406SEric Dumazet space = min_t(size_t, space, pfrag->size - pfrag->offset); 3885fbf93406SEric Dumazet space = tcp_wmem_schedule(sk, space); 3886fbf93406SEric Dumazet } 3887fbf93406SEric Dumazet if (space) { 3888fbf93406SEric Dumazet space = copy_page_from_iter(pfrag->page, pfrag->offset, 3889fbf93406SEric Dumazet space, &fo->data->msg_iter); 3890fbf93406SEric Dumazet if (unlikely(!space)) { 3891ba233b34SEric Dumazet tcp_skb_tsorted_anchor_cleanup(syn_data); 3892355a901eSEric Dumazet kfree_skb(syn_data); 3893783237e8SYuchung Cheng goto fallback; 3894783237e8SYuchung Cheng } 3895fbf93406SEric Dumazet skb_fill_page_desc(syn_data, 0, pfrag->page, 3896fbf93406SEric Dumazet pfrag->offset, space); 3897fbf93406SEric Dumazet page_ref_inc(pfrag->page); 3898fbf93406SEric Dumazet pfrag->offset += space; 3899fbf93406SEric Dumazet skb_len_add(syn_data, space); 3900f859a448SWillem de Bruijn skb_zcopy_set(syn_data, fo->uarg, NULL); 390107e100f9SEric Dumazet } 3902355a901eSEric Dumazet /* No more data pending in inet_wait_for_connect() */ 3903355a901eSEric Dumazet if (space == fo->size) 3904355a901eSEric Dumazet fo->data = NULL; 3905355a901eSEric Dumazet fo->copied = space; 3906783237e8SYuchung Cheng 3907355a901eSEric Dumazet tcp_connect_queue_skb(sk, syn_data); 39080f87230dSFrancis Yan if (syn_data->len) 39090f87230dSFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_BUSY); 3910355a901eSEric Dumazet 3911355a901eSEric Dumazet err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation); 3912355a901eSEric Dumazet 3913a1ac9c8aSMartin KaFai Lau skb_set_delivery_time(syn, syn_data->skb_mstamp_ns, true); 3914355a901eSEric Dumazet 3915355a901eSEric Dumazet /* Now full SYN+DATA was cloned and sent (or not), 3916355a901eSEric Dumazet * remove the SYN from the original skb (syn_data) 3917355a901eSEric Dumazet * we keep in write queue in case of a retransmit, as we 3918355a901eSEric Dumazet * also have the SYN packet (with no data) in the same queue. 3919431a9124SEric Dumazet */ 3920355a901eSEric Dumazet TCP_SKB_CB(syn_data)->seq++; 3921355a901eSEric Dumazet TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH; 3922355a901eSEric Dumazet if (!err) { 392367da22d2SYuchung Cheng tp->syn_data = (fo->copied > 0); 392475c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data); 3925f19c29e3SYuchung Cheng NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT); 3926783237e8SYuchung Cheng goto done; 3927783237e8SYuchung Cheng } 3928783237e8SYuchung Cheng 392975c119afSEric Dumazet /* data was not sent, put it in write_queue */ 393075c119afSEric Dumazet __skb_queue_tail(&sk->sk_write_queue, syn_data); 3931b5b7db8dSEric Dumazet tp->packets_out -= tcp_skb_pcount(syn_data); 3932b5b7db8dSEric Dumazet 3933783237e8SYuchung Cheng fallback: 3934783237e8SYuchung Cheng /* Send a regular SYN with Fast Open cookie request option */ 3935783237e8SYuchung Cheng if (fo->cookie.len > 0) 3936783237e8SYuchung Cheng fo->cookie.len = 0; 3937783237e8SYuchung Cheng err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); 3938783237e8SYuchung Cheng if (err) 3939783237e8SYuchung Cheng tp->syn_fastopen = 0; 3940783237e8SYuchung Cheng done: 3941783237e8SYuchung Cheng fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ 3942783237e8SYuchung Cheng return err; 3943783237e8SYuchung Cheng } 3944783237e8SYuchung Cheng 394567edfef7SAndi Kleen /* Build a SYN and send it off. */ 39461da177e4SLinus Torvalds int tcp_connect(struct sock *sk) 39471da177e4SLinus Torvalds { 39481da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 39491da177e4SLinus Torvalds struct sk_buff *buff; 3950ee586811SEric Paris int err; 39511da177e4SLinus Torvalds 3952de525be2SLawrence Brakmo tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB, 0, NULL); 39538ba60924SEric Dumazet 39548ba60924SEric Dumazet if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 39558ba60924SEric Dumazet return -EHOSTUNREACH; /* Routing failure or similar. */ 39568ba60924SEric Dumazet 39571da177e4SLinus Torvalds tcp_connect_init(sk); 39581da177e4SLinus Torvalds 39592b916477SAndrey Vagin if (unlikely(tp->repair)) { 39602b916477SAndrey Vagin tcp_finish_connect(sk, NULL); 39612b916477SAndrey Vagin return 0; 39622b916477SAndrey Vagin } 39632b916477SAndrey Vagin 39645882efffSEric Dumazet buff = tcp_stream_alloc_skb(sk, sk->sk_allocation, true); 3965355a901eSEric Dumazet if (unlikely(!buff)) 39661da177e4SLinus Torvalds return -ENOBUFS; 39671da177e4SLinus Torvalds 3968a3433f35SChangli Gao tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); 39699a568de4SEric Dumazet tcp_mstamp_refresh(tp); 39709a568de4SEric Dumazet tp->retrans_stamp = tcp_time_stamp(tp); 3971783237e8SYuchung Cheng tcp_connect_queue_skb(sk, buff); 3972735d3831SFlorian Westphal tcp_ecn_send_syn(sk, buff); 397375c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, buff); 39741da177e4SLinus Torvalds 3975783237e8SYuchung Cheng /* Send off SYN; include data in Fast Open. */ 3976783237e8SYuchung Cheng err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : 3977783237e8SYuchung Cheng tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); 3978ee586811SEric Paris if (err == -ECONNREFUSED) 3979ee586811SEric Paris return err; 3980bd37a088SWei Yongjun 3981bd37a088SWei Yongjun /* We change tp->snd_nxt after the tcp_transmit_skb() call 3982bd37a088SWei Yongjun * in order to make this packet get counted in tcpOutSegs. 3983bd37a088SWei Yongjun */ 3984e0d694d6SEric Dumazet WRITE_ONCE(tp->snd_nxt, tp->write_seq); 3985bd37a088SWei Yongjun tp->pushed_seq = tp->write_seq; 3986b5b7db8dSEric Dumazet buff = tcp_send_head(sk); 3987b5b7db8dSEric Dumazet if (unlikely(buff)) { 3988e0d694d6SEric Dumazet WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(buff)->seq); 3989b5b7db8dSEric Dumazet tp->pushed_seq = TCP_SKB_CB(buff)->seq; 3990b5b7db8dSEric Dumazet } 399181cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); 39921da177e4SLinus Torvalds 39931da177e4SLinus Torvalds /* Timer for repeating the SYN until an answer. */ 39943f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 39953f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 39961da177e4SLinus Torvalds return 0; 39971da177e4SLinus Torvalds } 39984bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_connect); 39991da177e4SLinus Torvalds 40001da177e4SLinus Torvalds /* Send out a delayed ack, the caller does the policy checking 40011da177e4SLinus Torvalds * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 40021da177e4SLinus Torvalds * for details. 40031da177e4SLinus Torvalds */ 40041da177e4SLinus Torvalds void tcp_send_delayed_ack(struct sock *sk) 40051da177e4SLinus Torvalds { 4006463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 4007463c84b9SArnaldo Carvalho de Melo int ato = icsk->icsk_ack.ato; 40081da177e4SLinus Torvalds unsigned long timeout; 40091da177e4SLinus Torvalds 40101da177e4SLinus Torvalds if (ato > TCP_DELACK_MIN) { 4011463c84b9SArnaldo Carvalho de Melo const struct tcp_sock *tp = tcp_sk(sk); 40121da177e4SLinus Torvalds int max_ato = HZ / 2; 40131da177e4SLinus Torvalds 401431954cd8SWei Wang if (inet_csk_in_pingpong_mode(sk) || 4015056834d9SIlpo Järvinen (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 40161da177e4SLinus Torvalds max_ato = TCP_DELACK_MAX; 40171da177e4SLinus Torvalds 40181da177e4SLinus Torvalds /* Slow path, intersegment interval is "high". */ 40191da177e4SLinus Torvalds 40201da177e4SLinus Torvalds /* If some rtt estimate is known, use it to bound delayed ack. 4021463c84b9SArnaldo Carvalho de Melo * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 40221da177e4SLinus Torvalds * directly. 40231da177e4SLinus Torvalds */ 4024740b0f18SEric Dumazet if (tp->srtt_us) { 4025740b0f18SEric Dumazet int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3), 4026740b0f18SEric Dumazet TCP_DELACK_MIN); 40271da177e4SLinus Torvalds 40281da177e4SLinus Torvalds if (rtt < max_ato) 40291da177e4SLinus Torvalds max_ato = rtt; 40301da177e4SLinus Torvalds } 40311da177e4SLinus Torvalds 40321da177e4SLinus Torvalds ato = min(ato, max_ato); 40331da177e4SLinus Torvalds } 40341da177e4SLinus Torvalds 40352b8ee4f0SMartin KaFai Lau ato = min_t(u32, ato, inet_csk(sk)->icsk_delack_max); 40362b8ee4f0SMartin KaFai Lau 40371da177e4SLinus Torvalds /* Stay within the limit we were given */ 40381da177e4SLinus Torvalds timeout = jiffies + ato; 40391da177e4SLinus Torvalds 40401da177e4SLinus Torvalds /* Use new timeout only if there wasn't a older one earlier. */ 4041463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 4042b6b6d653SEric Dumazet /* If delack timer is about to expire, send ACK now. */ 4043b6b6d653SEric Dumazet if (time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 40441da177e4SLinus Torvalds tcp_send_ack(sk); 40451da177e4SLinus Torvalds return; 40461da177e4SLinus Torvalds } 40471da177e4SLinus Torvalds 4048463c84b9SArnaldo Carvalho de Melo if (!time_before(timeout, icsk->icsk_ack.timeout)) 4049463c84b9SArnaldo Carvalho de Melo timeout = icsk->icsk_ack.timeout; 40501da177e4SLinus Torvalds } 4051463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 4052463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.timeout = timeout; 4053463c84b9SArnaldo Carvalho de Melo sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 40541da177e4SLinus Torvalds } 40551da177e4SLinus Torvalds 40561da177e4SLinus Torvalds /* This routine sends an ack and also updates the window. */ 40572987babbSYuchung Cheng void __tcp_send_ack(struct sock *sk, u32 rcv_nxt) 40581da177e4SLinus Torvalds { 40591da177e4SLinus Torvalds struct sk_buff *buff; 40601da177e4SLinus Torvalds 4061058dc334SIlpo Järvinen /* If we have been reset, we may not send again. */ 4062058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 4063058dc334SIlpo Järvinen return; 4064058dc334SIlpo Järvinen 40651da177e4SLinus Torvalds /* We are not putting this on the write queue, so 40661da177e4SLinus Torvalds * tcp_transmit_skb() will set the ownership to this 40671da177e4SLinus Torvalds * sock. 40681da177e4SLinus Torvalds */ 40697450aaf6SEric Dumazet buff = alloc_skb(MAX_TCP_HEADER, 40707450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 40717450aaf6SEric Dumazet if (unlikely(!buff)) { 4072a37c2134SEric Dumazet struct inet_connection_sock *icsk = inet_csk(sk); 4073a37c2134SEric Dumazet unsigned long delay; 4074a37c2134SEric Dumazet 4075a37c2134SEric Dumazet delay = TCP_DELACK_MAX << icsk->icsk_ack.retry; 4076a37c2134SEric Dumazet if (delay < TCP_RTO_MAX) 4077a37c2134SEric Dumazet icsk->icsk_ack.retry++; 4078463c84b9SArnaldo Carvalho de Melo inet_csk_schedule_ack(sk); 4079a37c2134SEric Dumazet icsk->icsk_ack.ato = TCP_ATO_MIN; 4080a37c2134SEric Dumazet inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, delay, TCP_RTO_MAX); 40811da177e4SLinus Torvalds return; 40821da177e4SLinus Torvalds } 40831da177e4SLinus Torvalds 40841da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 40851da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 4086a3433f35SChangli Gao tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); 40871da177e4SLinus Torvalds 408898781965SEric Dumazet /* We do not want pure acks influencing TCP Small Queues or fq/pacing 408998781965SEric Dumazet * too much. 409098781965SEric Dumazet * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784 409198781965SEric Dumazet */ 409298781965SEric Dumazet skb_set_tcp_pure_ack(buff); 409398781965SEric Dumazet 40941da177e4SLinus Torvalds /* Send it off, this clears delayed acks for us. */ 40952987babbSYuchung Cheng __tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0, rcv_nxt); 40961da177e4SLinus Torvalds } 409727cde44aSYuchung Cheng EXPORT_SYMBOL_GPL(__tcp_send_ack); 40982987babbSYuchung Cheng 40992987babbSYuchung Cheng void tcp_send_ack(struct sock *sk) 41002987babbSYuchung Cheng { 41012987babbSYuchung Cheng __tcp_send_ack(sk, tcp_sk(sk)->rcv_nxt); 41021da177e4SLinus Torvalds } 41031da177e4SLinus Torvalds 41041da177e4SLinus Torvalds /* This routine sends a packet with an out of date sequence 41051da177e4SLinus Torvalds * number. It assumes the other end will try to ack it. 41061da177e4SLinus Torvalds * 41071da177e4SLinus Torvalds * Question: what should we make while urgent mode? 41081da177e4SLinus Torvalds * 4.4BSD forces sending single byte of data. We cannot send 41091da177e4SLinus Torvalds * out of window data, because we have SND.NXT==SND.MAX... 41101da177e4SLinus Torvalds * 41111da177e4SLinus Torvalds * Current solution: to send TWO zero-length segments in urgent mode: 41121da177e4SLinus Torvalds * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 41131da177e4SLinus Torvalds * out-of-date with SND.UNA-1 to probe window. 41141da177e4SLinus Torvalds */ 4115e520af48SEric Dumazet static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib) 41161da177e4SLinus Torvalds { 41171da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 41181da177e4SLinus Torvalds struct sk_buff *skb; 41191da177e4SLinus Torvalds 41201da177e4SLinus Torvalds /* We don't queue it, tcp_transmit_skb() sets ownership. */ 41217450aaf6SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, 41227450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 412351456b29SIan Morris if (!skb) 41241da177e4SLinus Torvalds return -1; 41251da177e4SLinus Torvalds 41261da177e4SLinus Torvalds /* Reserve space for headers and set control bits. */ 41271da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 41281da177e4SLinus Torvalds /* Use a previous sequence. This should cause the other 41291da177e4SLinus Torvalds * end to send an ack. Don't queue or clone SKB, just 41301da177e4SLinus Torvalds * send it. 41311da177e4SLinus Torvalds */ 4132a3433f35SChangli Gao tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); 4133e2e8009fSRenato Westphal NET_INC_STATS(sock_net(sk), mib); 41347450aaf6SEric Dumazet return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0); 41351da177e4SLinus Torvalds } 41361da177e4SLinus Torvalds 4137385e2070SEric Dumazet /* Called from setsockopt( ... TCP_REPAIR ) */ 4138ee995283SPavel Emelyanov void tcp_send_window_probe(struct sock *sk) 4139ee995283SPavel Emelyanov { 4140ee995283SPavel Emelyanov if (sk->sk_state == TCP_ESTABLISHED) { 4141ee995283SPavel Emelyanov tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; 41429a568de4SEric Dumazet tcp_mstamp_refresh(tcp_sk(sk)); 4143e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE); 4144ee995283SPavel Emelyanov } 4145ee995283SPavel Emelyanov } 4146ee995283SPavel Emelyanov 414767edfef7SAndi Kleen /* Initiate keepalive or window probe from timer. */ 4148e520af48SEric Dumazet int tcp_write_wakeup(struct sock *sk, int mib) 41491da177e4SLinus Torvalds { 41501da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 41511da177e4SLinus Torvalds struct sk_buff *skb; 41521da177e4SLinus Torvalds 4153058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 4154058dc334SIlpo Järvinen return -1; 4155058dc334SIlpo Järvinen 415600db4124SIan Morris skb = tcp_send_head(sk); 415700db4124SIan Morris if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { 41581da177e4SLinus Torvalds int err; 41590c54b85fSIlpo Järvinen unsigned int mss = tcp_current_mss(sk); 416090840defSIlpo Järvinen unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 41611da177e4SLinus Torvalds 41621da177e4SLinus Torvalds if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 41631da177e4SLinus Torvalds tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 41641da177e4SLinus Torvalds 41651da177e4SLinus Torvalds /* We are probing the opening of a window 41661da177e4SLinus Torvalds * but the window size is != 0 41671da177e4SLinus Torvalds * must have been a result SWS avoidance ( sender ) 41681da177e4SLinus Torvalds */ 41691da177e4SLinus Torvalds if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 41701da177e4SLinus Torvalds skb->len > mss) { 41711da177e4SLinus Torvalds seg_size = min(seg_size, mss); 41724de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 417375c119afSEric Dumazet if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE, 417475c119afSEric Dumazet skb, seg_size, mss, GFP_ATOMIC)) 41751da177e4SLinus Torvalds return -1; 41761da177e4SLinus Torvalds } else if (!tcp_skb_pcount(skb)) 41775bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss); 41781da177e4SLinus Torvalds 41794de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 4180dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 418166f5fe62SIlpo Järvinen if (!err) 418266f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 41831da177e4SLinus Torvalds return err; 41841da177e4SLinus Torvalds } else { 418533f5f57eSIlpo Järvinen if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) 4186e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 1, mib); 4187e520af48SEric Dumazet return tcp_xmit_probe_skb(sk, 0, mib); 41881da177e4SLinus Torvalds } 41891da177e4SLinus Torvalds } 41901da177e4SLinus Torvalds 41911da177e4SLinus Torvalds /* A window probe timeout has occurred. If window is not closed send 41921da177e4SLinus Torvalds * a partial packet else a zero probe. 41931da177e4SLinus Torvalds */ 41941da177e4SLinus Torvalds void tcp_send_probe0(struct sock *sk) 41951da177e4SLinus Torvalds { 4196463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 41971da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 4198c6214a97SNikolay Borisov struct net *net = sock_net(sk); 4199c1d5674fSYuchung Cheng unsigned long timeout; 42001da177e4SLinus Torvalds int err; 42011da177e4SLinus Torvalds 4202e520af48SEric Dumazet err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE); 42031da177e4SLinus Torvalds 420475c119afSEric Dumazet if (tp->packets_out || tcp_write_queue_empty(sk)) { 42051da177e4SLinus Torvalds /* Cancel probe timer, if it is not required. */ 42066687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 0; 4207463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff = 0; 42089d9b1ee0SEnke Chen icsk->icsk_probes_tstamp = 0; 42091da177e4SLinus Torvalds return; 42101da177e4SLinus Torvalds } 42111da177e4SLinus Torvalds 4212c1d5674fSYuchung Cheng icsk->icsk_probes_out++; 42131da177e4SLinus Torvalds if (err <= 0) { 421439e24435SKuniyuki Iwashima if (icsk->icsk_backoff < READ_ONCE(net->ipv4.sysctl_tcp_retries2)) 4215463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff++; 4216c1d5674fSYuchung Cheng timeout = tcp_probe0_when(sk, TCP_RTO_MAX); 42171da177e4SLinus Torvalds } else { 42181da177e4SLinus Torvalds /* If packet was not sent due to local congestion, 4219c1d5674fSYuchung Cheng * Let senders fight for local resources conservatively. 42201da177e4SLinus Torvalds */ 4221c1d5674fSYuchung Cheng timeout = TCP_RESOURCE_PROBE_INTERVAL; 42221da177e4SLinus Torvalds } 4223344db93aSEnke Chen 4224344db93aSEnke Chen timeout = tcp_clamp_probe0_to_user_timeout(sk, timeout); 42258dc242adSEric Dumazet tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, timeout, TCP_RTO_MAX); 42261da177e4SLinus Torvalds } 42275db92c99SOctavian Purdila 4228ea3bea3aSEric Dumazet int tcp_rtx_synack(const struct sock *sk, struct request_sock *req) 42295db92c99SOctavian Purdila { 42305db92c99SOctavian Purdila const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific; 42315db92c99SOctavian Purdila struct flowi fl; 42325db92c99SOctavian Purdila int res; 42335db92c99SOctavian Purdila 4234cb6cd2ceSAkhmat Karakotov /* Paired with WRITE_ONCE() in sock_setsockopt() */ 4235cb6cd2ceSAkhmat Karakotov if (READ_ONCE(sk->sk_txrehash) == SOCK_TXREHASH_ENABLED) 42365e526552SEric Dumazet WRITE_ONCE(tcp_rsk(req)->txhash, net_tx_rndhash()); 4237331fca43SMartin KaFai Lau res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL, 4238331fca43SMartin KaFai Lau NULL); 42395db92c99SOctavian Purdila if (!res) { 42400a375c82SEric Dumazet TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); 42410a375c82SEric Dumazet NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 4242e9d9da91SEric Dumazet if (unlikely(tcp_passive_fastopen(sk))) { 4243e9d9da91SEric Dumazet /* sk has const attribute because listeners are lockless. 4244e9d9da91SEric Dumazet * However in this case, we are dealing with a passive fastopen 4245e9d9da91SEric Dumazet * socket thus we can change total_retrans value. 4246e9d9da91SEric Dumazet */ 4247e9d9da91SEric Dumazet tcp_sk_rw(sk)->total_retrans++; 4248e9d9da91SEric Dumazet } 4249cf34ce3dSSong Liu trace_tcp_retransmit_synack(sk, req); 42505db92c99SOctavian Purdila } 42515db92c99SOctavian Purdila return res; 42525db92c99SOctavian Purdila } 42535db92c99SOctavian Purdila EXPORT_SYMBOL(tcp_rtx_synack); 4254