1457c8996SThomas Gleixner // SPDX-License-Identifier: GPL-2.0-only 21da177e4SLinus Torvalds /* 31da177e4SLinus Torvalds * INET An implementation of the TCP/IP protocol suite for the LINUX 41da177e4SLinus Torvalds * operating system. INET is implemented using the BSD Socket 51da177e4SLinus Torvalds * interface as the means of communication with the user level. 61da177e4SLinus Torvalds * 71da177e4SLinus Torvalds * Implementation of the Transmission Control Protocol(TCP). 81da177e4SLinus Torvalds * 902c30a84SJesper Juhl * Authors: Ross Biro 101da177e4SLinus Torvalds * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> 111da177e4SLinus Torvalds * Mark Evans, <evansmp@uhura.aston.ac.uk> 121da177e4SLinus Torvalds * Corey Minyard <wf-rch!minyard@relay.EU.net> 131da177e4SLinus Torvalds * Florian La Roche, <flla@stud.uni-sb.de> 141da177e4SLinus Torvalds * Charles Hedrick, <hedrick@klinzhai.rutgers.edu> 151da177e4SLinus Torvalds * Linus Torvalds, <torvalds@cs.helsinki.fi> 161da177e4SLinus Torvalds * Alan Cox, <gw4pts@gw4pts.ampr.org> 171da177e4SLinus Torvalds * Matthew Dillon, <dillon@apollo.west.oic.com> 181da177e4SLinus Torvalds * Arnt Gulbrandsen, <agulbra@nvg.unit.no> 191da177e4SLinus Torvalds * Jorge Cwik, <jorge@laser.satlink.net> 201da177e4SLinus Torvalds */ 211da177e4SLinus Torvalds 221da177e4SLinus Torvalds /* 231da177e4SLinus Torvalds * Changes: Pedro Roque : Retransmit queue handled by TCP. 241da177e4SLinus Torvalds * : Fragmentation on mtu decrease 251da177e4SLinus Torvalds * : Segment collapse on retransmit 261da177e4SLinus Torvalds * : AF independence 271da177e4SLinus Torvalds * 281da177e4SLinus Torvalds * Linus Torvalds : send_delayed_ack 291da177e4SLinus Torvalds * David S. Miller : Charge memory using the right skb 301da177e4SLinus Torvalds * during syn/ack processing. 311da177e4SLinus Torvalds * David S. Miller : Output engine completely rewritten. 321da177e4SLinus Torvalds * Andrea Arcangeli: SYNACK carry ts_recent in tsecr. 331da177e4SLinus Torvalds * Cacophonix Gaul : draft-minshall-nagle-01 341da177e4SLinus Torvalds * J Hadi Salim : ECN support 351da177e4SLinus Torvalds * 361da177e4SLinus Torvalds */ 371da177e4SLinus Torvalds 3891df42beSJoe Perches #define pr_fmt(fmt) "TCP: " fmt 3991df42beSJoe Perches 401da177e4SLinus Torvalds #include <net/tcp.h> 41eda7acddSPeter Krystad #include <net/mptcp.h> 421da177e4SLinus Torvalds 431da177e4SLinus Torvalds #include <linux/compiler.h> 445a0e3ad6STejun Heo #include <linux/gfp.h> 451da177e4SLinus Torvalds #include <linux/module.h> 4660e2a778SUrsula Braun #include <linux/static_key.h> 471da177e4SLinus Torvalds 48e086101bSCong Wang #include <trace/events/tcp.h> 4935089bb2SDavid S. Miller 509799ccb0SEric Dumazet /* Refresh clocks of a TCP socket, 519799ccb0SEric Dumazet * ensuring monotically increasing values. 529799ccb0SEric Dumazet */ 539799ccb0SEric Dumazet void tcp_mstamp_refresh(struct tcp_sock *tp) 549799ccb0SEric Dumazet { 559799ccb0SEric Dumazet u64 val = tcp_clock_ns(); 569799ccb0SEric Dumazet 575f6188a8SEric Dumazet tp->tcp_clock_cache = val; 58e6d14070SEric Dumazet tp->tcp_mstamp = div_u64(val, NSEC_PER_USEC); 599799ccb0SEric Dumazet } 609799ccb0SEric Dumazet 6146d3ceabSEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 6246d3ceabSEric Dumazet int push_one, gfp_t gfp); 63519855c5SWilliam Allen Simpson 6467edfef7SAndi Kleen /* Account for new data that has been sent to the network. */ 6575c119afSEric Dumazet static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb) 666ff03ac3SIlpo Järvinen { 676ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 686ff03ac3SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 6966f5fe62SIlpo Järvinen unsigned int prior_packets = tp->packets_out; 709e412ba7SIlpo Järvinen 71e0d694d6SEric Dumazet WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(skb)->end_seq); 728512430eSIlpo Järvinen 7375c119afSEric Dumazet __skb_unlink(skb, &sk->sk_write_queue); 7475c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, skb); 7575c119afSEric Dumazet 7685369750SCambda Zhu if (tp->highest_sack == NULL) 7785369750SCambda Zhu tp->highest_sack = skb; 7885369750SCambda Zhu 7966f5fe62SIlpo Järvinen tp->packets_out += tcp_skb_pcount(skb); 80bec41a11SYuchung Cheng if (!prior_packets || icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) 81750ea2baSYuchung Cheng tcp_rearm_rto(sk); 82f19c29e3SYuchung Cheng 83f7324acdSDavid S. Miller NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT, 84f19c29e3SYuchung Cheng tcp_skb_pcount(skb)); 854bfe744fSEric Dumazet tcp_check_space(sk); 866a5dc9e5SEric Dumazet } 871da177e4SLinus Torvalds 88a4ecb15aSCui, Cheng /* SND.NXT, if window was not shrunk or the amount of shrunk was less than one 89a4ecb15aSCui, Cheng * window scaling factor due to loss of precision. 901da177e4SLinus Torvalds * If window has been shrunk, what should we make? It is not clear at all. 911da177e4SLinus Torvalds * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-( 921da177e4SLinus Torvalds * Anything in between SND.UNA...SND.UNA+SND.WND also can be already 931da177e4SLinus Torvalds * invalid. OK, let's make this for now: 941da177e4SLinus Torvalds */ 95cf533ea5SEric Dumazet static inline __u32 tcp_acceptable_seq(const struct sock *sk) 961da177e4SLinus Torvalds { 97cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 989e412ba7SIlpo Järvinen 99a4ecb15aSCui, Cheng if (!before(tcp_wnd_end(tp), tp->snd_nxt) || 100a4ecb15aSCui, Cheng (tp->rx_opt.wscale_ok && 101a4ecb15aSCui, Cheng ((tp->snd_nxt - tcp_wnd_end(tp)) < (1 << tp->rx_opt.rcv_wscale)))) 1021da177e4SLinus Torvalds return tp->snd_nxt; 1031da177e4SLinus Torvalds else 10490840defSIlpo Järvinen return tcp_wnd_end(tp); 1051da177e4SLinus Torvalds } 1061da177e4SLinus Torvalds 1071da177e4SLinus Torvalds /* Calculate mss to advertise in SYN segment. 1081da177e4SLinus Torvalds * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that: 1091da177e4SLinus Torvalds * 1101da177e4SLinus Torvalds * 1. It is independent of path mtu. 1111da177e4SLinus Torvalds * 2. Ideally, it is maximal possible segment size i.e. 65535-40. 1121da177e4SLinus Torvalds * 3. For IPv4 it is reasonable to calculate it from maximal MTU of 1131da177e4SLinus Torvalds * attached devices, because some buggy hosts are confused by 1141da177e4SLinus Torvalds * large MSS. 1151da177e4SLinus Torvalds * 4. We do not make 3, we advertise MSS, calculated from first 1161da177e4SLinus Torvalds * hop device mtu, but allow to raise it to ip_rt_min_advmss. 1171da177e4SLinus Torvalds * This may be overridden via information stored in routing table. 1181da177e4SLinus Torvalds * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible, 1191da177e4SLinus Torvalds * probably even Jumbo". 1201da177e4SLinus Torvalds */ 1211da177e4SLinus Torvalds static __u16 tcp_advertise_mss(struct sock *sk) 1221da177e4SLinus Torvalds { 1231da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 124cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1251da177e4SLinus Torvalds int mss = tp->advmss; 1261da177e4SLinus Torvalds 1270dbaee3bSDavid S. Miller if (dst) { 1280dbaee3bSDavid S. Miller unsigned int metric = dst_metric_advmss(dst); 1290dbaee3bSDavid S. Miller 1300dbaee3bSDavid S. Miller if (metric < mss) { 1310dbaee3bSDavid S. Miller mss = metric; 1321da177e4SLinus Torvalds tp->advmss = mss; 1331da177e4SLinus Torvalds } 1340dbaee3bSDavid S. Miller } 1351da177e4SLinus Torvalds 1361da177e4SLinus Torvalds return (__u16)mss; 1371da177e4SLinus Torvalds } 1381da177e4SLinus Torvalds 1391da177e4SLinus Torvalds /* RFC2861. Reset CWND after idle period longer RTO to "restart window". 1406f021c62SEric Dumazet * This is the first part of cwnd validation mechanism. 1416f021c62SEric Dumazet */ 1426f021c62SEric Dumazet void tcp_cwnd_restart(struct sock *sk, s32 delta) 1431da177e4SLinus Torvalds { 144463c84b9SArnaldo Carvalho de Melo struct tcp_sock *tp = tcp_sk(sk); 1456f021c62SEric Dumazet u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk)); 14640570375SEric Dumazet u32 cwnd = tcp_snd_cwnd(tp); 1471da177e4SLinus Torvalds 1486687e988SArnaldo Carvalho de Melo tcp_ca_event(sk, CA_EVENT_CWND_RESTART); 1491da177e4SLinus Torvalds 1506687e988SArnaldo Carvalho de Melo tp->snd_ssthresh = tcp_current_ssthresh(sk); 1511da177e4SLinus Torvalds restart_cwnd = min(restart_cwnd, cwnd); 1521da177e4SLinus Torvalds 153463c84b9SArnaldo Carvalho de Melo while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd) 1541da177e4SLinus Torvalds cwnd >>= 1; 15540570375SEric Dumazet tcp_snd_cwnd_set(tp, max(cwnd, restart_cwnd)); 156c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 1571da177e4SLinus Torvalds tp->snd_cwnd_used = 0; 1581da177e4SLinus Torvalds } 1591da177e4SLinus Torvalds 16067edfef7SAndi Kleen /* Congestion state accounting after a packet has been sent. */ 16140efc6faSStephen Hemminger static void tcp_event_data_sent(struct tcp_sock *tp, 162cf533ea5SEric Dumazet struct sock *sk) 1631da177e4SLinus Torvalds { 164463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 165d635fbe2SEric Dumazet const u32 now = tcp_jiffies32; 1661da177e4SLinus Torvalds 16705c5a46dSNeal Cardwell if (tcp_packets_in_flight(tp) == 0) 16805c5a46dSNeal Cardwell tcp_ca_event(sk, CA_EVENT_TX_START); 16905c5a46dSNeal Cardwell 1704a41f453SWei Wang tp->lsndtime = now; 1714d8f24eeSWei Wang 1724d8f24eeSWei Wang /* If it is a reply for ato after last received 173562b1fdfSHaiyang Zhang * packet, increase pingpong count. 1744d8f24eeSWei Wang */ 1754d8f24eeSWei Wang if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato) 176562b1fdfSHaiyang Zhang inet_csk_inc_pingpong_cnt(sk); 1771da177e4SLinus Torvalds } 1781da177e4SLinus Torvalds 17967edfef7SAndi Kleen /* Account for an ACK we sent. */ 180059217c1SNeal Cardwell static inline void tcp_event_ack_sent(struct sock *sk, u32 rcv_nxt) 1811da177e4SLinus Torvalds { 1825d9f4262SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 1835d9f4262SEric Dumazet 1842b195850SEric Dumazet if (unlikely(tp->compressed_ack)) { 185200d95f4SEric Dumazet NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED, 1862b195850SEric Dumazet tp->compressed_ack); 1872b195850SEric Dumazet tp->compressed_ack = 0; 1885d9f4262SEric Dumazet if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1) 1895d9f4262SEric Dumazet __sock_put(sk); 1905d9f4262SEric Dumazet } 19127cde44aSYuchung Cheng 19227cde44aSYuchung Cheng if (unlikely(rcv_nxt != tp->rcv_nxt)) 19327cde44aSYuchung Cheng return; /* Special ACK sent by DCTCP to reflect ECN */ 194059217c1SNeal Cardwell tcp_dec_quickack_mode(sk); 195463c84b9SArnaldo Carvalho de Melo inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); 1961da177e4SLinus Torvalds } 1971da177e4SLinus Torvalds 1981da177e4SLinus Torvalds /* Determine a window scaling and initial window to offer. 1991da177e4SLinus Torvalds * Based on the assumption that the given amount of space 2001da177e4SLinus Torvalds * will be offered. Store the results in the tp structure. 2011da177e4SLinus Torvalds * NOTE: for smooth operation initial space offering should 2021da177e4SLinus Torvalds * be a multiple of mss if possible. We assume here that mss >= 1. 2031da177e4SLinus Torvalds * This MUST be enforced by all callers. 2041da177e4SLinus Torvalds */ 205ceef9ab6SEric Dumazet void tcp_select_initial_window(const struct sock *sk, int __space, __u32 mss, 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; 802d1a02ed6SEric Dumazet opts->tsval = tcp_skb_timestamp_ts(false, 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; 887d1a02ed6SEric Dumazet opts->tsval = tcp_skb_timestamp_ts(false, 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; 946d1a02ed6SEric Dumazet opts->tsval = skb ? tcp_skb_timestamp_ts(false, 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 | \ 1079133c4c0dSEric Dumazet TCPF_MTU_REDUCED_DEFERRED | \ 1080133c4c0dSEric Dumazet TCPF_ACK_DEFERRED) 108146d3ceabSEric Dumazet /** 108246d3ceabSEric Dumazet * tcp_release_cb - tcp release_sock() callback 108346d3ceabSEric Dumazet * @sk: socket 108446d3ceabSEric Dumazet * 108546d3ceabSEric Dumazet * called from release_sock() to perform protocol dependent 108646d3ceabSEric Dumazet * actions before socket release. 108746d3ceabSEric Dumazet */ 108846d3ceabSEric Dumazet void tcp_release_cb(struct sock *sk) 108946d3ceabSEric Dumazet { 1090fac30731SEric Dumazet unsigned long flags = smp_load_acquire(&sk->sk_tsq_flags); 1091fac30731SEric Dumazet unsigned long nflags; 109246d3ceabSEric Dumazet 10936f458dfbSEric Dumazet /* perform an atomic operation only if at least one flag is set */ 10946f458dfbSEric Dumazet do { 10956f458dfbSEric Dumazet if (!(flags & TCP_DEFERRED_ALL)) 10966f458dfbSEric Dumazet return; 10976f458dfbSEric Dumazet nflags = flags & ~TCP_DEFERRED_ALL; 1098fac30731SEric Dumazet } while (!try_cmpxchg(&sk->sk_tsq_flags, &flags, nflags)); 10996f458dfbSEric Dumazet 110073a6bab5SEric Dumazet if (flags & TCPF_TSQ_DEFERRED) { 110173a6bab5SEric Dumazet tcp_tsq_write(sk); 110273a6bab5SEric Dumazet __sock_put(sk); 110373a6bab5SEric Dumazet } 1104c3f9b018SEric Dumazet 110540fc3423SEric Dumazet if (flags & TCPF_WRITE_TIMER_DEFERRED) { 11066f458dfbSEric Dumazet tcp_write_timer_handler(sk); 1107144d56e9SEric Dumazet __sock_put(sk); 1108144d56e9SEric Dumazet } 110940fc3423SEric Dumazet if (flags & TCPF_DELACK_TIMER_DEFERRED) { 11106f458dfbSEric Dumazet tcp_delack_timer_handler(sk); 1111144d56e9SEric Dumazet __sock_put(sk); 1112144d56e9SEric Dumazet } 111340fc3423SEric Dumazet if (flags & TCPF_MTU_REDUCED_DEFERRED) { 11144fab9071SNeal Cardwell inet_csk(sk)->icsk_af_ops->mtu_reduced(sk); 1115144d56e9SEric Dumazet __sock_put(sk); 1116144d56e9SEric Dumazet } 1117133c4c0dSEric Dumazet if ((flags & TCPF_ACK_DEFERRED) && inet_csk_ack_scheduled(sk)) 1118133c4c0dSEric Dumazet tcp_send_ack(sk); 111946d3ceabSEric Dumazet } 112046d3ceabSEric Dumazet EXPORT_SYMBOL(tcp_release_cb); 112146d3ceabSEric Dumazet 112246d3ceabSEric Dumazet void __init tcp_tasklet_init(void) 112346d3ceabSEric Dumazet { 112446d3ceabSEric Dumazet int i; 112546d3ceabSEric Dumazet 112646d3ceabSEric Dumazet for_each_possible_cpu(i) { 112746d3ceabSEric Dumazet struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i); 112846d3ceabSEric Dumazet 112946d3ceabSEric Dumazet INIT_LIST_HEAD(&tsq->head); 1130c6533ca8SAllen Pais tasklet_setup(&tsq->tasklet, tcp_tasklet_func); 113146d3ceabSEric Dumazet } 113246d3ceabSEric Dumazet } 113346d3ceabSEric Dumazet 113446d3ceabSEric Dumazet /* 113546d3ceabSEric Dumazet * Write buffer destructor automatically called from kfree_skb. 11368e3bff96Sstephen hemminger * We can't xmit new skbs from this context, as we might already 113746d3ceabSEric Dumazet * hold qdisc lock. 113846d3ceabSEric Dumazet */ 1139d6a4a104SEric Dumazet void tcp_wfree(struct sk_buff *skb) 114046d3ceabSEric Dumazet { 114146d3ceabSEric Dumazet struct sock *sk = skb->sk; 114246d3ceabSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 1143408f0a6cSEric Dumazet unsigned long flags, nval, oval; 1144b548b17aSEric Dumazet struct tsq_tasklet *tsq; 1145b548b17aSEric Dumazet bool empty; 11469b462d02SEric Dumazet 11479b462d02SEric Dumazet /* Keep one reference on sk_wmem_alloc. 11489b462d02SEric Dumazet * Will be released by sk_free() from here or tcp_tasklet_func() 11499b462d02SEric Dumazet */ 115014afee4bSReshetova, Elena WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc)); 11519b462d02SEric Dumazet 11529b462d02SEric Dumazet /* If this softirq is serviced by ksoftirqd, we are likely under stress. 11539b462d02SEric Dumazet * Wait until our queues (qdisc + devices) are drained. 11549b462d02SEric Dumazet * This gives : 11559b462d02SEric Dumazet * - less callbacks to tcp_write_xmit(), reducing stress (batches) 11569b462d02SEric Dumazet * - chance for incoming ACK (processed by another cpu maybe) 11579b462d02SEric Dumazet * to migrate this flow (skb->ooo_okay will be eventually set) 11589b462d02SEric Dumazet */ 115914afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current) 11609b462d02SEric Dumazet goto out; 116146d3ceabSEric Dumazet 1162b548b17aSEric Dumazet oval = smp_load_acquire(&sk->sk_tsq_flags); 1163b548b17aSEric Dumazet do { 1164408f0a6cSEric Dumazet if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED)) 1165408f0a6cSEric Dumazet goto out; 1166408f0a6cSEric Dumazet 116773a6bab5SEric Dumazet nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED; 1168b548b17aSEric Dumazet } while (!try_cmpxchg(&sk->sk_tsq_flags, &oval, nval)); 1169408f0a6cSEric Dumazet 117046d3ceabSEric Dumazet /* queue this socket to tasklet queue */ 117146d3ceabSEric Dumazet local_irq_save(flags); 1172903ceff7SChristoph Lameter tsq = this_cpu_ptr(&tsq_tasklet); 1173a9b204d1SEric Dumazet empty = list_empty(&tsq->head); 117446d3ceabSEric Dumazet list_add(&tp->tsq_node, &tsq->head); 1175a9b204d1SEric Dumazet if (empty) 117646d3ceabSEric Dumazet tasklet_schedule(&tsq->tasklet); 117746d3ceabSEric Dumazet local_irq_restore(flags); 11789b462d02SEric Dumazet return; 11799b462d02SEric Dumazet out: 11809b462d02SEric Dumazet sk_free(sk); 118146d3ceabSEric Dumazet } 118246d3ceabSEric Dumazet 118373a6bab5SEric Dumazet /* Note: Called under soft irq. 118473a6bab5SEric Dumazet * We can call TCP stack right away, unless socket is owned by user. 1185218af599SEric Dumazet */ 1186218af599SEric Dumazet enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer) 1187218af599SEric Dumazet { 1188218af599SEric Dumazet struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer); 1189218af599SEric Dumazet struct sock *sk = (struct sock *)tp; 1190218af599SEric Dumazet 119173a6bab5SEric Dumazet tcp_tsq_handler(sk); 119273a6bab5SEric Dumazet sock_put(sk); 1193218af599SEric Dumazet 1194218af599SEric Dumazet return HRTIMER_NORESTART; 1195218af599SEric Dumazet } 1196218af599SEric Dumazet 1197a7a25630SEric Dumazet static void tcp_update_skb_after_send(struct sock *sk, struct sk_buff *skb, 1198a7a25630SEric Dumazet u64 prior_wstamp) 1199e2080072SEric Dumazet { 1200ab408b6dSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 1201ab408b6dSEric Dumazet 1202ab408b6dSEric Dumazet if (sk->sk_pacing_status != SK_PACING_NONE) { 120328b24f90SEric Dumazet unsigned long rate = READ_ONCE(sk->sk_pacing_rate); 1204ab408b6dSEric Dumazet 1205ab408b6dSEric Dumazet /* Original sch_fq does not pace first 10 MSS 1206ab408b6dSEric Dumazet * Note that tp->data_segs_out overflows after 2^32 packets, 1207ab408b6dSEric Dumazet * this is a minor annoyance. 1208ab408b6dSEric Dumazet */ 120976a9ebe8SEric Dumazet if (rate != ~0UL && rate && tp->data_segs_out >= 10) { 1210a7a25630SEric Dumazet u64 len_ns = div64_ul((u64)skb->len * NSEC_PER_SEC, rate); 1211a7a25630SEric Dumazet u64 credit = tp->tcp_wstamp_ns - prior_wstamp; 1212a7a25630SEric Dumazet 1213a7a25630SEric Dumazet /* take into account OS jitter */ 1214a7a25630SEric Dumazet len_ns -= min_t(u64, len_ns / 2, credit); 1215a7a25630SEric Dumazet tp->tcp_wstamp_ns += len_ns; 1216ab408b6dSEric Dumazet } 1217ab408b6dSEric Dumazet } 1218e2080072SEric Dumazet list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue); 1219e2080072SEric Dumazet } 1220e2080072SEric Dumazet 122105e22e83SEric Dumazet INDIRECT_CALLABLE_DECLARE(int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl)); 122205e22e83SEric Dumazet INDIRECT_CALLABLE_DECLARE(int inet6_csk_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl)); 1223dd2e0b86SEric Dumazet INDIRECT_CALLABLE_DECLARE(void tcp_v4_send_check(struct sock *sk, struct sk_buff *skb)); 122405e22e83SEric Dumazet 12251da177e4SLinus Torvalds /* This routine actually transmits TCP packets queued in by 12261da177e4SLinus Torvalds * tcp_do_sendmsg(). This is used by both the initial 12271da177e4SLinus Torvalds * transmission and possible later retransmissions. 12281da177e4SLinus Torvalds * All SKB's seen here are completely headerless. It is our 12291da177e4SLinus Torvalds * job to build the TCP header, and pass the packet down to 12301da177e4SLinus Torvalds * IP so it can do the same plus pass the packet off to the 12311da177e4SLinus Torvalds * device. 12321da177e4SLinus Torvalds * 12331da177e4SLinus Torvalds * We are working here with either a clone of the original 12341da177e4SLinus Torvalds * SKB, or a fresh unique copy made by the retransmit engine. 12351da177e4SLinus Torvalds */ 12362987babbSYuchung Cheng static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, 12372987babbSYuchung Cheng int clone_it, gfp_t gfp_mask, u32 rcv_nxt) 12381da177e4SLinus Torvalds { 12396687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 1240dfb4b9dcSDavid S. Miller struct inet_sock *inet; 1241dfb4b9dcSDavid S. Miller struct tcp_sock *tp; 1242dfb4b9dcSDavid S. Miller struct tcp_skb_cb *tcb; 124333ad798cSAdam Langley struct tcp_out_options opts; 124495c96174SEric Dumazet unsigned int tcp_options_size, tcp_header_size; 12458c72c65bSEric Dumazet struct sk_buff *oskb = NULL; 1246cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 12471da177e4SLinus Torvalds struct tcphdr *th; 1248a7a25630SEric Dumazet u64 prior_wstamp; 12491da177e4SLinus Torvalds int err; 12501da177e4SLinus Torvalds 1251dfb4b9dcSDavid S. Miller BUG_ON(!skb || !tcp_skb_pcount(skb)); 12526f094b9eSLawrence Brakmo tp = tcp_sk(sk); 12537f12422cSYuchung Cheng prior_wstamp = tp->tcp_wstamp_ns; 12547f12422cSYuchung Cheng tp->tcp_wstamp_ns = max(tp->tcp_wstamp_ns, tp->tcp_clock_cache); 1255a1ac9c8aSMartin KaFai Lau skb_set_delivery_time(skb, tp->tcp_wstamp_ns, true); 1256ccdbb6e9SEric Dumazet if (clone_it) { 12578c72c65bSEric Dumazet oskb = skb; 1258e2080072SEric Dumazet 1259e2080072SEric Dumazet tcp_skb_tsorted_save(oskb) { 1260e2080072SEric Dumazet if (unlikely(skb_cloned(oskb))) 1261e2080072SEric Dumazet skb = pskb_copy(oskb, gfp_mask); 1262dfb4b9dcSDavid S. Miller else 1263e2080072SEric Dumazet skb = skb_clone(oskb, gfp_mask); 1264e2080072SEric Dumazet } tcp_skb_tsorted_restore(oskb); 1265e2080072SEric Dumazet 1266dfb4b9dcSDavid S. Miller if (unlikely(!skb)) 1267dfb4b9dcSDavid S. Miller return -ENOBUFS; 1268b738a185SEric Dumazet /* retransmit skbs might have a non zero value in skb->dev 1269b738a185SEric Dumazet * because skb->dev is aliased with skb->rbnode.rb_left 1270b738a185SEric Dumazet */ 1271b738a185SEric Dumazet skb->dev = NULL; 1272dfb4b9dcSDavid S. Miller } 12735f6188a8SEric Dumazet 1274dfb4b9dcSDavid S. Miller inet = inet_sk(sk); 1275dfb4b9dcSDavid S. Miller tcb = TCP_SKB_CB(skb); 127633ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 12771da177e4SLinus Torvalds 1278051ba674SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) { 127933ad798cSAdam Langley tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); 1280051ba674SEric Dumazet } else { 128133ad798cSAdam Langley tcp_options_size = tcp_established_options(sk, skb, &opts, 128233ad798cSAdam Langley &md5); 1283051ba674SEric Dumazet /* Force a PSH flag on all (GSO) packets to expedite GRO flush 1284051ba674SEric Dumazet * at receiver : This slightly improve GRO performance. 1285051ba674SEric Dumazet * Note that we do not force the PSH flag for non GSO packets, 1286051ba674SEric Dumazet * because they might be sent under high congestion events, 1287051ba674SEric Dumazet * and in this case it is better to delay the delivery of 1-MSS 1288051ba674SEric Dumazet * packets and thus the corresponding ACK packet that would 1289051ba674SEric Dumazet * release the following packet. 1290051ba674SEric Dumazet */ 1291051ba674SEric Dumazet if (tcp_skb_pcount(skb) > 1) 1292051ba674SEric Dumazet tcb->tcp_flags |= TCPHDR_PSH; 1293051ba674SEric Dumazet } 129433ad798cSAdam Langley tcp_header_size = tcp_options_size + sizeof(struct tcphdr); 12951da177e4SLinus Torvalds 1296726e9e8bSEric Dumazet /* We set skb->ooo_okay to one if this packet can select 1297726e9e8bSEric Dumazet * a different TX queue than prior packets of this flow, 1298726e9e8bSEric Dumazet * to avoid self inflicted reorders. 1299726e9e8bSEric Dumazet * The 'other' queue decision is based on current cpu number 1300726e9e8bSEric Dumazet * if XPS is enabled, or sk->sk_txhash otherwise. 1301726e9e8bSEric Dumazet * We can switch to another (and better) queue if: 1302726e9e8bSEric Dumazet * 1) No packet with payload is in qdisc/device queues. 1303726e9e8bSEric Dumazet * Delays in TX completion can defeat the test 1304726e9e8bSEric Dumazet * even if packets were already sent. 1305726e9e8bSEric Dumazet * 2) Or rtx queue is empty. 1306726e9e8bSEric Dumazet * This mitigates above case if ACK packets for 1307726e9e8bSEric Dumazet * all prior packets were already processed. 1308547669d4SEric Dumazet */ 1309726e9e8bSEric Dumazet skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1) || 1310726e9e8bSEric Dumazet tcp_rtx_queue_empty(sk); 13111da177e4SLinus Torvalds 131238ab52e8SEric Dumazet /* If we had to use memory reserve to allocate this skb, 131338ab52e8SEric Dumazet * this might cause drops if packet is looped back : 131438ab52e8SEric Dumazet * Other socket might not have SOCK_MEMALLOC. 131538ab52e8SEric Dumazet * Packets not looped back do not care about pfmemalloc. 131638ab52e8SEric Dumazet */ 131738ab52e8SEric Dumazet skb->pfmemalloc = 0; 131838ab52e8SEric Dumazet 1319aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 1320aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 132146d3ceabSEric Dumazet 132246d3ceabSEric Dumazet skb_orphan(skb); 132346d3ceabSEric Dumazet skb->sk = sk; 13241d2077acSEric Dumazet skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree; 132514afee4bSReshetova, Elena refcount_add(skb->truesize, &sk->sk_wmem_alloc); 13261da177e4SLinus Torvalds 1327eb44ad4eSEric Dumazet skb_set_dst_pending_confirm(skb, READ_ONCE(sk->sk_dst_pending_confirm)); 1328c3a2e837SJulian Anastasov 13291da177e4SLinus Torvalds /* Build TCP header and checksum it. */ 1330ea1627c2SEric Dumazet th = (struct tcphdr *)skb->data; 1331c720c7e8SEric Dumazet th->source = inet->inet_sport; 1332c720c7e8SEric Dumazet th->dest = inet->inet_dport; 13331da177e4SLinus Torvalds th->seq = htonl(tcb->seq); 13342987babbSYuchung Cheng th->ack_seq = htonl(rcv_nxt); 1335df7a3b07SAl Viro *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | 13364de075e0SEric Dumazet tcb->tcp_flags); 1337dfb4b9dcSDavid S. Miller 13381da177e4SLinus Torvalds th->check = 0; 13391da177e4SLinus Torvalds th->urg_ptr = 0; 13401da177e4SLinus Torvalds 134133f5f57eSIlpo Järvinen /* The urg_mode check is necessary during a below snd_una win probe */ 13427691367dSHerbert Xu if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { 13437691367dSHerbert Xu if (before(tp->snd_up, tcb->seq + 0x10000)) { 13441da177e4SLinus Torvalds th->urg_ptr = htons(tp->snd_up - tcb->seq); 13451da177e4SLinus Torvalds th->urg = 1; 13467691367dSHerbert Xu } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { 13470eae88f3SEric Dumazet th->urg_ptr = htons(0xFFFF); 13487691367dSHerbert Xu th->urg = 1; 13497691367dSHerbert Xu } 13501da177e4SLinus Torvalds } 13511da177e4SLinus Torvalds 135251466a75SEric Dumazet skb_shinfo(skb)->gso_type = sk->sk_gso_type; 1353ea1627c2SEric Dumazet if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) { 1354ea1627c2SEric Dumazet th->window = htons(tcp_select_window(sk)); 1355ea1627c2SEric Dumazet tcp_ecn_send(sk, skb, th, tcp_header_size); 1356ea1627c2SEric Dumazet } else { 1357ea1627c2SEric Dumazet /* RFC1323: The window in SYN & SYN/ACK segments 1358ea1627c2SEric Dumazet * is never scaled. 1359ea1627c2SEric Dumazet */ 1360ea1627c2SEric Dumazet th->window = htons(min(tp->rcv_wnd, 65535U)); 1361ea1627c2SEric Dumazet } 1362fa3fe2b1SFlorian Westphal 1363ea66758cSPaolo Abeni tcp_options_write(th, tp, &opts); 1364fa3fe2b1SFlorian Westphal 1365cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 1366cfb6eeb4SYOSHIFUJI Hideaki /* Calculate the MD5 hash, as we have all we need now */ 1367cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 1368aba54656SEric Dumazet sk_gso_disable(sk); 1369bd0388aeSWilliam Allen Simpson tp->af_specific->calc_md5_hash(opts.hash_location, 137039f8e58eSEric Dumazet md5, sk, skb); 1371cfb6eeb4SYOSHIFUJI Hideaki } 1372cfb6eeb4SYOSHIFUJI Hideaki #endif 1373cfb6eeb4SYOSHIFUJI Hideaki 1374331fca43SMartin KaFai Lau /* BPF prog is the last one writing header option */ 1375331fca43SMartin KaFai Lau bpf_skops_write_hdr_opt(sk, skb, NULL, NULL, 0, &opts); 1376331fca43SMartin KaFai Lau 1377dd2e0b86SEric Dumazet INDIRECT_CALL_INET(icsk->icsk_af_ops->send_check, 1378dd2e0b86SEric Dumazet tcp_v6_send_check, tcp_v4_send_check, 1379dd2e0b86SEric Dumazet sk, skb); 13801da177e4SLinus Torvalds 13814de075e0SEric Dumazet if (likely(tcb->tcp_flags & TCPHDR_ACK)) 1382059217c1SNeal Cardwell tcp_event_ack_sent(sk, rcv_nxt); 13831da177e4SLinus Torvalds 1384a44d6eacSMartin KaFai Lau if (skb->len != tcp_header_size) { 1385cf533ea5SEric Dumazet tcp_event_data_sent(tp, sk); 1386a44d6eacSMartin KaFai Lau tp->data_segs_out += tcp_skb_pcount(skb); 1387ba113c3aSWei Wang tp->bytes_sent += skb->len - tcp_header_size; 1388a44d6eacSMartin KaFai Lau } 13891da177e4SLinus Torvalds 1390bd37a088SWei Yongjun if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) 1391aa2ea058STom Herbert TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, 1392aa2ea058STom Herbert tcp_skb_pcount(skb)); 13931da177e4SLinus Torvalds 13942efd055cSMarcelo Ricardo Leitner tp->segs_out += tcp_skb_pcount(skb); 13950ae5b43dSYuchung Cheng skb_set_hash_from_sk(skb, sk); 1396f69ad292SEric Dumazet /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */ 1397cd7d8498SEric Dumazet skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb); 1398f69ad292SEric Dumazet skb_shinfo(skb)->gso_size = tcp_skb_mss(skb); 1399cd7d8498SEric Dumazet 1400d3edd06eSEric Dumazet /* Leave earliest departure time in skb->tstamp (skb->skb_mstamp_ns) */ 1401971f10ecSEric Dumazet 1402971f10ecSEric Dumazet /* Cleanup our debris for IP stacks */ 1403971f10ecSEric Dumazet memset(skb->cb, 0, max(sizeof(struct inet_skb_parm), 1404971f10ecSEric Dumazet sizeof(struct inet6_skb_parm))); 1405971f10ecSEric Dumazet 1406a842fe14SEric Dumazet tcp_add_tx_delay(skb, tp); 1407a842fe14SEric Dumazet 140805e22e83SEric Dumazet err = INDIRECT_CALL_INET(icsk->icsk_af_ops->queue_xmit, 140905e22e83SEric Dumazet inet6_csk_xmit, ip_queue_xmit, 141005e22e83SEric Dumazet sk, skb, &inet->cork.fl); 14117faee5c0SEric Dumazet 14128c72c65bSEric Dumazet if (unlikely(err > 0)) { 14135ee2c941SChristoph Paasch tcp_enter_cwr(sk); 14148c72c65bSEric Dumazet err = net_xmit_eval(err); 14158c72c65bSEric Dumazet } 1416fc225799SEric Dumazet if (!err && oskb) { 1417a7a25630SEric Dumazet tcp_update_skb_after_send(sk, oskb, prior_wstamp); 1418fc225799SEric Dumazet tcp_rate_skb_sent(sk, oskb); 1419fc225799SEric Dumazet } 14208c72c65bSEric Dumazet return err; 14211da177e4SLinus Torvalds } 14221da177e4SLinus Torvalds 14232987babbSYuchung Cheng static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, 14242987babbSYuchung Cheng gfp_t gfp_mask) 14252987babbSYuchung Cheng { 14262987babbSYuchung Cheng return __tcp_transmit_skb(sk, skb, clone_it, gfp_mask, 14272987babbSYuchung Cheng tcp_sk(sk)->rcv_nxt); 14282987babbSYuchung Cheng } 14292987babbSYuchung Cheng 143067edfef7SAndi Kleen /* This routine just queues the buffer for sending. 14311da177e4SLinus Torvalds * 14321da177e4SLinus Torvalds * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, 14331da177e4SLinus Torvalds * otherwise socket can stall. 14341da177e4SLinus Torvalds */ 14351da177e4SLinus Torvalds static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) 14361da177e4SLinus Torvalds { 14371da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 14381da177e4SLinus Torvalds 14391da177e4SLinus Torvalds /* Advance write_seq and place onto the write_queue. */ 14400f317464SEric Dumazet WRITE_ONCE(tp->write_seq, TCP_SKB_CB(skb)->end_seq); 1441f4a775d1SEric Dumazet __skb_header_release(skb); 1442fe067e8aSDavid S. Miller tcp_add_write_queue_tail(sk, skb); 1443ab4e846aSEric Dumazet sk_wmem_queued_add(sk, skb->truesize); 14443ab224beSHideo Aoki sk_mem_charge(sk, skb->truesize); 14451da177e4SLinus Torvalds } 14461da177e4SLinus Torvalds 144767edfef7SAndi Kleen /* Initialize TSO segments for a packet. */ 14485bbb432cSEric Dumazet static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now) 1449f6302d1dSDavid S. Miller { 14504a64fd6cSEric Dumazet if (skb->len <= mss_now) { 1451f6302d1dSDavid S. Miller /* Avoid the costly divide in the normal 1452f6302d1dSDavid S. Miller * non-TSO case. 1453f6302d1dSDavid S. Miller */ 1454cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, 1); 1455f69ad292SEric Dumazet TCP_SKB_CB(skb)->tcp_gso_size = 0; 1456f6302d1dSDavid S. Miller } else { 1457cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now)); 1458f69ad292SEric Dumazet TCP_SKB_CB(skb)->tcp_gso_size = mss_now; 14591da177e4SLinus Torvalds } 14601da177e4SLinus Torvalds } 14611da177e4SLinus Torvalds 1462797108d1SIlpo Järvinen /* Pcount in the middle of the write queue got changed, we need to do various 1463797108d1SIlpo Järvinen * tweaks to fix counters 1464797108d1SIlpo Järvinen */ 1465cf533ea5SEric Dumazet static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) 1466797108d1SIlpo Järvinen { 1467797108d1SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1468797108d1SIlpo Järvinen 1469797108d1SIlpo Järvinen tp->packets_out -= decr; 1470797108d1SIlpo Järvinen 1471797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 1472797108d1SIlpo Järvinen tp->sacked_out -= decr; 1473797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 1474797108d1SIlpo Järvinen tp->retrans_out -= decr; 1475797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 1476797108d1SIlpo Järvinen tp->lost_out -= decr; 1477797108d1SIlpo Järvinen 1478797108d1SIlpo Järvinen /* Reno case is special. Sigh... */ 1479797108d1SIlpo Järvinen if (tcp_is_reno(tp) && decr > 0) 1480797108d1SIlpo Järvinen tp->sacked_out -= min_t(u32, tp->sacked_out, decr); 1481797108d1SIlpo Järvinen 1482797108d1SIlpo Järvinen if (tp->lost_skb_hint && 1483797108d1SIlpo Järvinen before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && 1484713bafeaSYuchung Cheng (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) 1485797108d1SIlpo Järvinen tp->lost_cnt_hint -= decr; 1486797108d1SIlpo Järvinen 1487797108d1SIlpo Järvinen tcp_verify_left_out(tp); 1488797108d1SIlpo Järvinen } 1489797108d1SIlpo Järvinen 14900a2cf20cSSoheil Hassas Yeganeh static bool tcp_has_tx_tstamp(const struct sk_buff *skb) 14910a2cf20cSSoheil Hassas Yeganeh { 14920a2cf20cSSoheil Hassas Yeganeh return TCP_SKB_CB(skb)->txstamp_ack || 14930a2cf20cSSoheil Hassas Yeganeh (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP); 14940a2cf20cSSoheil Hassas Yeganeh } 14950a2cf20cSSoheil Hassas Yeganeh 1496490cc7d0SWillem de Bruijn static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2) 1497490cc7d0SWillem de Bruijn { 1498490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo = skb_shinfo(skb); 1499490cc7d0SWillem de Bruijn 15000a2cf20cSSoheil Hassas Yeganeh if (unlikely(tcp_has_tx_tstamp(skb)) && 1501490cc7d0SWillem de Bruijn !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) { 1502490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo2 = skb_shinfo(skb2); 1503490cc7d0SWillem de Bruijn u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP; 1504490cc7d0SWillem de Bruijn 1505490cc7d0SWillem de Bruijn shinfo->tx_flags &= ~tsflags; 1506490cc7d0SWillem de Bruijn shinfo2->tx_flags |= tsflags; 1507490cc7d0SWillem de Bruijn swap(shinfo->tskey, shinfo2->tskey); 1508b51e13faSMartin KaFai Lau TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack; 1509b51e13faSMartin KaFai Lau TCP_SKB_CB(skb)->txstamp_ack = 0; 1510490cc7d0SWillem de Bruijn } 1511490cc7d0SWillem de Bruijn } 1512490cc7d0SWillem de Bruijn 1513a166140eSMartin KaFai Lau static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2) 1514a166140eSMartin KaFai Lau { 1515a166140eSMartin KaFai Lau TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor; 1516a166140eSMartin KaFai Lau TCP_SKB_CB(skb)->eor = 0; 1517a166140eSMartin KaFai Lau } 1518a166140eSMartin KaFai Lau 151975c119afSEric Dumazet /* Insert buff after skb on the write or rtx queue of sk. */ 152075c119afSEric Dumazet static void tcp_insert_write_queue_after(struct sk_buff *skb, 152175c119afSEric Dumazet struct sk_buff *buff, 152275c119afSEric Dumazet struct sock *sk, 152375c119afSEric Dumazet enum tcp_queue tcp_queue) 152475c119afSEric Dumazet { 152575c119afSEric Dumazet if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE) 152675c119afSEric Dumazet __skb_queue_after(&sk->sk_write_queue, skb, buff); 152775c119afSEric Dumazet else 152875c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, buff); 152975c119afSEric Dumazet } 153075c119afSEric Dumazet 15311da177e4SLinus Torvalds /* Function to create two new TCP segments. Shrinks the given segment 15321da177e4SLinus Torvalds * to the specified size and appends a new segment with the rest of the 15331da177e4SLinus Torvalds * packet to the list. This won't be called frequently, I hope. 15341da177e4SLinus Torvalds * Remember, these are still headerless SKBs at this point. 15351da177e4SLinus Torvalds */ 153675c119afSEric Dumazet int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue, 153775c119afSEric Dumazet struct sk_buff *skb, u32 len, 15386cc55e09SOctavian Purdila unsigned int mss_now, gfp_t gfp) 15391da177e4SLinus Torvalds { 15401da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 15411da177e4SLinus Torvalds struct sk_buff *buff; 1542b4a24397SEric Dumazet int old_factor; 1543b617158dSEric Dumazet long limit; 1544b60b49eaSHerbert Xu int nlen; 15459ce01461SIlpo Järvinen u8 flags; 15461da177e4SLinus Torvalds 15472fceec13SIlpo Järvinen if (WARN_ON(len > skb->len)) 15482fceec13SIlpo Järvinen return -EINVAL; 15496a438bbeSStephen Hemminger 1550b4a24397SEric Dumazet DEBUG_NET_WARN_ON_ONCE(skb_headlen(skb)); 15511da177e4SLinus Torvalds 1552b617158dSEric Dumazet /* tcp_sendmsg() can overshoot sk_wmem_queued by one full size skb. 1553b617158dSEric Dumazet * We need some allowance to not penalize applications setting small 1554b617158dSEric Dumazet * SO_SNDBUF values. 1555b617158dSEric Dumazet * Also allow first and last skb in retransmit queue to be split. 1556b617158dSEric Dumazet */ 15577c4e983cSAlexander Duyck limit = sk->sk_sndbuf + 2 * SKB_TRUESIZE(GSO_LEGACY_MAX_SIZE); 1558b617158dSEric Dumazet if (unlikely((sk->sk_wmem_queued >> 1) > limit && 1559b617158dSEric Dumazet tcp_queue != TCP_FRAG_IN_WRITE_QUEUE && 1560b617158dSEric Dumazet skb != tcp_rtx_queue_head(sk) && 1561b617158dSEric Dumazet skb != tcp_rtx_queue_tail(sk))) { 1562f070ef2aSEric Dumazet NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPWQUEUETOOBIG); 1563f070ef2aSEric Dumazet return -ENOMEM; 1564f070ef2aSEric Dumazet } 1565f070ef2aSEric Dumazet 1566c4777efaSEric Dumazet if (skb_unclone_keeptruesize(skb, gfp)) 15671da177e4SLinus Torvalds return -ENOMEM; 15681da177e4SLinus Torvalds 15691da177e4SLinus Torvalds /* Get a new skb... force flag on. */ 15705882efffSEric Dumazet buff = tcp_stream_alloc_skb(sk, gfp, true); 157151456b29SIan Morris if (!buff) 15721da177e4SLinus Torvalds return -ENOMEM; /* We'll just try again later. */ 157341477662SJakub Kicinski skb_copy_decrypted(buff, skb); 15745a369ca6SPaolo Abeni mptcp_skb_ext_copy(buff, skb); 1575ef5cb973SHerbert Xu 1576ab4e846aSEric Dumazet sk_wmem_queued_add(sk, buff->truesize); 15773ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1578b4a24397SEric Dumazet nlen = skb->len - len; 1579b60b49eaSHerbert Xu buff->truesize += nlen; 1580b60b49eaSHerbert Xu skb->truesize -= nlen; 15811da177e4SLinus Torvalds 15821da177e4SLinus Torvalds /* Correct the sequence numbers. */ 15831da177e4SLinus Torvalds TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 15841da177e4SLinus Torvalds TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 15851da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 15861da177e4SLinus Torvalds 15871da177e4SLinus Torvalds /* PSH and FIN should only be set in the second packet. */ 15884de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 15894de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 15904de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1591e14c3cafSHerbert Xu TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; 1592a166140eSMartin KaFai Lau tcp_skb_fragment_eor(skb, buff); 15931da177e4SLinus Torvalds 15941da177e4SLinus Torvalds skb_split(skb, buff, len); 15951da177e4SLinus Torvalds 1596a1ac9c8aSMartin KaFai Lau skb_set_delivery_time(buff, skb->tstamp, true); 1597490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 15981da177e4SLinus Torvalds 15996475be16SDavid S. Miller old_factor = tcp_skb_pcount(skb); 16006475be16SDavid S. Miller 16011da177e4SLinus Torvalds /* Fix up tso_factor for both original and new SKB. */ 16025bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 16035bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 16041da177e4SLinus Torvalds 1605b9f64820SYuchung Cheng /* Update delivered info for the new segment */ 1606b9f64820SYuchung Cheng TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx; 1607b9f64820SYuchung Cheng 16086475be16SDavid S. Miller /* If this packet has been sent out already, we must 16096475be16SDavid S. Miller * adjust the various packet counters. 16106475be16SDavid S. Miller */ 1611cf0b450cSHerbert Xu if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 16126475be16SDavid S. Miller int diff = old_factor - tcp_skb_pcount(skb) - 16136475be16SDavid S. Miller tcp_skb_pcount(buff); 16141da177e4SLinus Torvalds 1615797108d1SIlpo Järvinen if (diff) 1616797108d1SIlpo Järvinen tcp_adjust_pcount(sk, skb, diff); 16171da177e4SLinus Torvalds } 16181da177e4SLinus Torvalds 16191da177e4SLinus Torvalds /* Link BUFF into the send queue. */ 1620f4a775d1SEric Dumazet __skb_header_release(buff); 162175c119afSEric Dumazet tcp_insert_write_queue_after(skb, buff, sk, tcp_queue); 1622f67971e6SEric Dumazet if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE) 1623e2080072SEric Dumazet list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor); 16241da177e4SLinus Torvalds 16251da177e4SLinus Torvalds return 0; 16261da177e4SLinus Torvalds } 16271da177e4SLinus Torvalds 1628f4d01666SEric Dumazet /* This is similar to __pskb_pull_tail(). The difference is that pulled 1629f4d01666SEric Dumazet * data is not copied, but immediately discarded. 16301da177e4SLinus Torvalds */ 16317162fb24SEric Dumazet static int __pskb_trim_head(struct sk_buff *skb, int len) 16321da177e4SLinus Torvalds { 16337b7fc97aSEric Dumazet struct skb_shared_info *shinfo; 16341da177e4SLinus Torvalds int i, k, eat; 16351da177e4SLinus Torvalds 1636b4a24397SEric Dumazet DEBUG_NET_WARN_ON_ONCE(skb_headlen(skb)); 16371da177e4SLinus Torvalds eat = len; 16381da177e4SLinus Torvalds k = 0; 16397b7fc97aSEric Dumazet shinfo = skb_shinfo(skb); 16407b7fc97aSEric Dumazet for (i = 0; i < shinfo->nr_frags; i++) { 16417b7fc97aSEric Dumazet int size = skb_frag_size(&shinfo->frags[i]); 16429e903e08SEric Dumazet 16439e903e08SEric Dumazet if (size <= eat) { 1644aff65da0SIan Campbell skb_frag_unref(skb, i); 16459e903e08SEric Dumazet eat -= size; 16461da177e4SLinus Torvalds } else { 16477b7fc97aSEric Dumazet shinfo->frags[k] = shinfo->frags[i]; 16481da177e4SLinus Torvalds if (eat) { 1649b54c9d5bSJonathan Lemon skb_frag_off_add(&shinfo->frags[k], eat); 16507b7fc97aSEric Dumazet skb_frag_size_sub(&shinfo->frags[k], eat); 16511da177e4SLinus Torvalds eat = 0; 16521da177e4SLinus Torvalds } 16531da177e4SLinus Torvalds k++; 16541da177e4SLinus Torvalds } 16551da177e4SLinus Torvalds } 16567b7fc97aSEric Dumazet shinfo->nr_frags = k; 16571da177e4SLinus Torvalds 16581da177e4SLinus Torvalds skb->data_len -= len; 16591da177e4SLinus Torvalds skb->len = skb->data_len; 16607162fb24SEric Dumazet return len; 16611da177e4SLinus Torvalds } 16621da177e4SLinus Torvalds 166367edfef7SAndi Kleen /* Remove acked data from a packet in the transmit queue. */ 16641da177e4SLinus Torvalds int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 16651da177e4SLinus Torvalds { 16667162fb24SEric Dumazet u32 delta_truesize; 16677162fb24SEric Dumazet 1668c4777efaSEric Dumazet if (skb_unclone_keeptruesize(skb, GFP_ATOMIC)) 16691da177e4SLinus Torvalds return -ENOMEM; 16701da177e4SLinus Torvalds 16717162fb24SEric Dumazet delta_truesize = __pskb_trim_head(skb, len); 16721da177e4SLinus Torvalds 16731da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq += len; 16741da177e4SLinus Torvalds 16757162fb24SEric Dumazet skb->truesize -= delta_truesize; 1676ab4e846aSEric Dumazet sk_wmem_queued_add(sk, -delta_truesize); 16779b65b17dSTalal Ahmad if (!skb_zcopy_pure(skb)) 16787162fb24SEric Dumazet sk_mem_uncharge(sk, delta_truesize); 16791da177e4SLinus Torvalds 16805b35e1e6SNeal Cardwell /* Any change of skb->len requires recalculation of tso factor. */ 16811da177e4SLinus Torvalds if (tcp_skb_pcount(skb) > 1) 16825bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb)); 16831da177e4SLinus Torvalds 16841da177e4SLinus Torvalds return 0; 16851da177e4SLinus Torvalds } 16861da177e4SLinus Torvalds 16871b63edd6SYuchung Cheng /* Calculate MSS not accounting any TCP options. */ 16881b63edd6SYuchung Cheng static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu) 16895d424d5aSJohn Heffner { 1690cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1691cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 16925d424d5aSJohn Heffner int mss_now; 16935d424d5aSJohn Heffner 16945d424d5aSJohn Heffner /* Calculate base mss without TCP options: 16955d424d5aSJohn Heffner It is MMS_S - sizeof(tcphdr) of rfc1122 16965d424d5aSJohn Heffner */ 16975d424d5aSJohn Heffner mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 16985d424d5aSJohn Heffner 169967469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 170067469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 170167469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 170267469601SEric Dumazet 170367469601SEric Dumazet if (dst && dst_allfrag(dst)) 170467469601SEric Dumazet mss_now -= icsk->icsk_af_ops->net_frag_header_len; 170567469601SEric Dumazet } 170667469601SEric Dumazet 17075d424d5aSJohn Heffner /* Clamp it (mss_clamp does not include tcp options) */ 17085d424d5aSJohn Heffner if (mss_now > tp->rx_opt.mss_clamp) 17095d424d5aSJohn Heffner mss_now = tp->rx_opt.mss_clamp; 17105d424d5aSJohn Heffner 17115d424d5aSJohn Heffner /* Now subtract optional transport overhead */ 17125d424d5aSJohn Heffner mss_now -= icsk->icsk_ext_hdr_len; 17135d424d5aSJohn Heffner 17145d424d5aSJohn Heffner /* Then reserve room for full set of TCP options and 8 bytes of data */ 171578eb166cSKuniyuki Iwashima mss_now = max(mss_now, 171678eb166cSKuniyuki Iwashima READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_snd_mss)); 17175d424d5aSJohn Heffner return mss_now; 17185d424d5aSJohn Heffner } 17195d424d5aSJohn Heffner 17201b63edd6SYuchung Cheng /* Calculate MSS. Not accounting for SACKs here. */ 17211b63edd6SYuchung Cheng int tcp_mtu_to_mss(struct sock *sk, int pmtu) 17221b63edd6SYuchung Cheng { 17231b63edd6SYuchung Cheng /* Subtract TCP options size, not including SACKs */ 17241b63edd6SYuchung Cheng return __tcp_mtu_to_mss(sk, pmtu) - 17251b63edd6SYuchung Cheng (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr)); 17261b63edd6SYuchung Cheng } 1727c7bb4b89SEric Dumazet EXPORT_SYMBOL(tcp_mtu_to_mss); 17281b63edd6SYuchung Cheng 17295d424d5aSJohn Heffner /* Inverse of above */ 173067469601SEric Dumazet int tcp_mss_to_mtu(struct sock *sk, int mss) 17315d424d5aSJohn Heffner { 1732cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1733cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 17345d424d5aSJohn Heffner int mtu; 17355d424d5aSJohn Heffner 17365d424d5aSJohn Heffner mtu = mss + 17375d424d5aSJohn Heffner tp->tcp_header_len + 17385d424d5aSJohn Heffner icsk->icsk_ext_hdr_len + 17395d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 17405d424d5aSJohn Heffner 174167469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 174267469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 174367469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 174467469601SEric Dumazet 174567469601SEric Dumazet if (dst && dst_allfrag(dst)) 174667469601SEric Dumazet mtu += icsk->icsk_af_ops->net_frag_header_len; 174767469601SEric Dumazet } 17485d424d5aSJohn Heffner return mtu; 17495d424d5aSJohn Heffner } 1750556c6b46SNeal Cardwell EXPORT_SYMBOL(tcp_mss_to_mtu); 17515d424d5aSJohn Heffner 175267edfef7SAndi Kleen /* MTU probing init per socket */ 17535d424d5aSJohn Heffner void tcp_mtup_init(struct sock *sk) 17545d424d5aSJohn Heffner { 17555d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 17565d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 1757b0f9ca53SFan Du struct net *net = sock_net(sk); 17585d424d5aSJohn Heffner 1759f47d00e0SKuniyuki Iwashima icsk->icsk_mtup.enabled = READ_ONCE(net->ipv4.sysctl_tcp_mtu_probing) > 1; 17605d424d5aSJohn Heffner icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 17615d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 176288d78bc0SKuniyuki Iwashima icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, READ_ONCE(net->ipv4.sysctl_tcp_base_mss)); 17635d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 176405cbc0dbSFan Du if (icsk->icsk_mtup.enabled) 1765c74df29aSEric Dumazet icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; 17665d424d5aSJohn Heffner } 17674bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_mtup_init); 17685d424d5aSJohn Heffner 17691da177e4SLinus Torvalds /* This function synchronize snd mss to current pmtu/exthdr set. 17701da177e4SLinus Torvalds 17711da177e4SLinus Torvalds tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 17721da177e4SLinus Torvalds for TCP options, but includes only bare TCP header. 17731da177e4SLinus Torvalds 17741da177e4SLinus Torvalds tp->rx_opt.mss_clamp is mss negotiated at connection setup. 1775caa20d9aSStephen Hemminger It is minimum of user_mss and mss received with SYN. 17761da177e4SLinus Torvalds It also does not include TCP options. 17771da177e4SLinus Torvalds 1778d83d8461SArnaldo Carvalho de Melo inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. 17791da177e4SLinus Torvalds 17801da177e4SLinus Torvalds tp->mss_cache is current effective sending mss, including 17811da177e4SLinus Torvalds all tcp options except for SACKs. It is evaluated, 17821da177e4SLinus Torvalds taking into account current pmtu, but never exceeds 17831da177e4SLinus Torvalds tp->rx_opt.mss_clamp. 17841da177e4SLinus Torvalds 17851da177e4SLinus Torvalds NOTE1. rfc1122 clearly states that advertised MSS 17861da177e4SLinus Torvalds DOES NOT include either tcp or ip options. 17871da177e4SLinus Torvalds 1788d83d8461SArnaldo Carvalho de Melo NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache 1789d83d8461SArnaldo Carvalho de Melo are READ ONLY outside this function. --ANK (980731) 17901da177e4SLinus Torvalds */ 17911da177e4SLinus Torvalds unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 17921da177e4SLinus Torvalds { 17931da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1794d83d8461SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 17955d424d5aSJohn Heffner int mss_now; 17961da177e4SLinus Torvalds 17975d424d5aSJohn Heffner if (icsk->icsk_mtup.search_high > pmtu) 17985d424d5aSJohn Heffner icsk->icsk_mtup.search_high = pmtu; 17991da177e4SLinus Torvalds 18005d424d5aSJohn Heffner mss_now = tcp_mtu_to_mss(sk, pmtu); 1801409d22b4SIlpo Järvinen mss_now = tcp_bound_to_half_wnd(tp, mss_now); 18021da177e4SLinus Torvalds 18031da177e4SLinus Torvalds /* And store cached results */ 1804d83d8461SArnaldo Carvalho de Melo icsk->icsk_pmtu_cookie = pmtu; 18055d424d5aSJohn Heffner if (icsk->icsk_mtup.enabled) 18065d424d5aSJohn Heffner mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); 1807c1b4a7e6SDavid S. Miller tp->mss_cache = mss_now; 18081da177e4SLinus Torvalds 18091da177e4SLinus Torvalds return mss_now; 18101da177e4SLinus Torvalds } 18114bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_sync_mss); 18121da177e4SLinus Torvalds 18131da177e4SLinus Torvalds /* Compute the current effective MSS, taking SACKs and IP options, 18141da177e4SLinus Torvalds * and even PMTU discovery events into account. 18151da177e4SLinus Torvalds */ 18160c54b85fSIlpo Järvinen unsigned int tcp_current_mss(struct sock *sk) 18171da177e4SLinus Torvalds { 1818cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1819cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1820c1b4a7e6SDavid S. Miller u32 mss_now; 182195c96174SEric Dumazet unsigned int header_len; 182233ad798cSAdam Langley struct tcp_out_options opts; 182333ad798cSAdam Langley struct tcp_md5sig_key *md5; 18241da177e4SLinus Torvalds 1825c1b4a7e6SDavid S. Miller mss_now = tp->mss_cache; 1826c1b4a7e6SDavid S. Miller 18271da177e4SLinus Torvalds if (dst) { 18281da177e4SLinus Torvalds u32 mtu = dst_mtu(dst); 1829d83d8461SArnaldo Carvalho de Melo if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 18301da177e4SLinus Torvalds mss_now = tcp_sync_mss(sk, mtu); 18311da177e4SLinus Torvalds } 18321da177e4SLinus Torvalds 183333ad798cSAdam Langley header_len = tcp_established_options(sk, NULL, &opts, &md5) + 183433ad798cSAdam Langley sizeof(struct tcphdr); 183533ad798cSAdam Langley /* The mss_cache is sized based on tp->tcp_header_len, which assumes 183633ad798cSAdam Langley * some common options. If this is an odd packet (because we have SACK 183733ad798cSAdam Langley * blocks etc) then our calculated header_len will be different, and 183833ad798cSAdam Langley * we have to adjust mss_now correspondingly */ 183933ad798cSAdam Langley if (header_len != tp->tcp_header_len) { 184033ad798cSAdam Langley int delta = (int) header_len - tp->tcp_header_len; 184133ad798cSAdam Langley mss_now -= delta; 184233ad798cSAdam Langley } 1843cfb6eeb4SYOSHIFUJI Hideaki 18441da177e4SLinus Torvalds return mss_now; 18451da177e4SLinus Torvalds } 18461da177e4SLinus Torvalds 184786fd14adSWeiping Pan /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. 184886fd14adSWeiping Pan * As additional protections, we do not touch cwnd in retransmission phases, 184986fd14adSWeiping Pan * and if application hit its sndbuf limit recently. 185086fd14adSWeiping Pan */ 185186fd14adSWeiping Pan static void tcp_cwnd_application_limited(struct sock *sk) 1852a762a980SDavid S. Miller { 18539e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1854a762a980SDavid S. Miller 185586fd14adSWeiping Pan if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && 185686fd14adSWeiping Pan sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 185786fd14adSWeiping Pan /* Limited by application or receiver window. */ 185886fd14adSWeiping Pan u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); 185986fd14adSWeiping Pan u32 win_used = max(tp->snd_cwnd_used, init_win); 186040570375SEric Dumazet if (win_used < tcp_snd_cwnd(tp)) { 186186fd14adSWeiping Pan tp->snd_ssthresh = tcp_current_ssthresh(sk); 186240570375SEric Dumazet tcp_snd_cwnd_set(tp, (tcp_snd_cwnd(tp) + win_used) >> 1); 186386fd14adSWeiping Pan } 186486fd14adSWeiping Pan tp->snd_cwnd_used = 0; 186586fd14adSWeiping Pan } 1866c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 186786fd14adSWeiping Pan } 186886fd14adSWeiping Pan 1869ca8a2263SNeal Cardwell static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited) 1870a762a980SDavid S. Miller { 18711b1fc3fdSWei Wang const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 1872a762a980SDavid S. Miller struct tcp_sock *tp = tcp_sk(sk); 1873a762a980SDavid S. Miller 1874f4ce91ceSNeal Cardwell /* Track the strongest available signal of the degree to which the cwnd 1875f4ce91ceSNeal Cardwell * is fully utilized. If cwnd-limited then remember that fact for the 1876f4ce91ceSNeal Cardwell * current window. If not cwnd-limited then track the maximum number of 1877f4ce91ceSNeal Cardwell * outstanding packets in the current window. (If cwnd-limited then we 1878f4ce91ceSNeal Cardwell * chose to not update tp->max_packets_out to avoid an extra else 1879f4ce91ceSNeal Cardwell * clause with no functional impact.) 1880ca8a2263SNeal Cardwell */ 1881f4ce91ceSNeal Cardwell if (!before(tp->snd_una, tp->cwnd_usage_seq) || 1882f4ce91ceSNeal Cardwell is_cwnd_limited || 1883f4ce91ceSNeal Cardwell (!tp->is_cwnd_limited && 1884f4ce91ceSNeal Cardwell tp->packets_out > tp->max_packets_out)) { 1885ca8a2263SNeal Cardwell tp->is_cwnd_limited = is_cwnd_limited; 1886f4ce91ceSNeal Cardwell tp->max_packets_out = tp->packets_out; 1887f4ce91ceSNeal Cardwell tp->cwnd_usage_seq = tp->snd_nxt; 1888ca8a2263SNeal Cardwell } 1889e114a710SEric Dumazet 189024901551SEric Dumazet if (tcp_is_cwnd_limited(sk)) { 1891a762a980SDavid S. Miller /* Network is feed fully. */ 1892a762a980SDavid S. Miller tp->snd_cwnd_used = 0; 1893c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 1894a762a980SDavid S. Miller } else { 1895a762a980SDavid S. Miller /* Network starves. */ 1896a762a980SDavid S. Miller if (tp->packets_out > tp->snd_cwnd_used) 1897a762a980SDavid S. Miller tp->snd_cwnd_used = tp->packets_out; 1898a762a980SDavid S. Miller 18994845b571SKuniyuki Iwashima if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle) && 1900c2203cf7SEric Dumazet (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto && 19011b1fc3fdSWei Wang !ca_ops->cong_control) 1902a762a980SDavid S. Miller tcp_cwnd_application_limited(sk); 1903b0f71bd3SFrancis Yan 1904b0f71bd3SFrancis Yan /* The following conditions together indicate the starvation 1905b0f71bd3SFrancis Yan * is caused by insufficient sender buffer: 1906b0f71bd3SFrancis Yan * 1) just sent some data (see tcp_write_xmit) 1907b0f71bd3SFrancis Yan * 2) not cwnd limited (this else condition) 190875c119afSEric Dumazet * 3) no more data to send (tcp_write_queue_empty()) 1909b0f71bd3SFrancis Yan * 4) application is hitting buffer limit (SOCK_NOSPACE) 1910b0f71bd3SFrancis Yan */ 191175c119afSEric Dumazet if (tcp_write_queue_empty(sk) && sk->sk_socket && 1912b0f71bd3SFrancis Yan test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) && 1913b0f71bd3SFrancis Yan (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) 1914b0f71bd3SFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED); 1915a762a980SDavid S. Miller } 1916a762a980SDavid S. Miller } 1917a762a980SDavid S. Miller 1918d4589926SEric Dumazet /* Minshall's variant of the Nagle send check. */ 1919d4589926SEric Dumazet static bool tcp_minshall_check(const struct tcp_sock *tp) 1920d4589926SEric Dumazet { 1921d4589926SEric Dumazet return after(tp->snd_sml, tp->snd_una) && 1922d4589926SEric Dumazet !after(tp->snd_sml, tp->snd_nxt); 1923d4589926SEric Dumazet } 1924d4589926SEric Dumazet 1925d4589926SEric Dumazet /* Update snd_sml if this skb is under mss 1926d4589926SEric Dumazet * Note that a TSO packet might end with a sub-mss segment 1927d4589926SEric Dumazet * The test is really : 1928d4589926SEric Dumazet * if ((skb->len % mss) != 0) 1929d4589926SEric Dumazet * tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1930d4589926SEric Dumazet * But we can avoid doing the divide again given we already have 1931d4589926SEric Dumazet * skb_pcount = skb->len / mss_now 19320e3a4803SIlpo Järvinen */ 1933d4589926SEric Dumazet static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now, 1934d4589926SEric Dumazet const struct sk_buff *skb) 1935d4589926SEric Dumazet { 1936d4589926SEric Dumazet if (skb->len < tcp_skb_pcount(skb) * mss_now) 1937d4589926SEric Dumazet tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1938d4589926SEric Dumazet } 1939d4589926SEric Dumazet 1940d4589926SEric Dumazet /* Return false, if packet can be sent now without violation Nagle's rules: 1941d4589926SEric Dumazet * 1. It is full sized. (provided by caller in %partial bool) 1942d4589926SEric Dumazet * 2. Or it contains FIN. (already checked by caller) 1943d4589926SEric Dumazet * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. 1944d4589926SEric Dumazet * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 1945d4589926SEric Dumazet * With Minshall's modification: all sent small packets are ACKed. 1946d4589926SEric Dumazet */ 1947d4589926SEric Dumazet static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp, 1948cc93fc51SPeter Pan(潘卫平) int nonagle) 1949d4589926SEric Dumazet { 1950d4589926SEric Dumazet return partial && 1951d4589926SEric Dumazet ((nonagle & TCP_NAGLE_CORK) || 1952d4589926SEric Dumazet (!nonagle && tp->packets_out && tcp_minshall_check(tp))); 1953d4589926SEric Dumazet } 1954605ad7f1SEric Dumazet 1955605ad7f1SEric Dumazet /* Return how many segs we'd like on a TSO packet, 195665466904SEric Dumazet * depending on current pacing rate, and how close the peer is. 195765466904SEric Dumazet * 195865466904SEric Dumazet * Rationale is: 195965466904SEric Dumazet * - For close peers, we rather send bigger packets to reduce 196065466904SEric Dumazet * cpu costs, because occasional losses will be repaired fast. 196165466904SEric Dumazet * - For long distance/rtt flows, we would like to get ACK clocking 196265466904SEric Dumazet * with 1 ACK per ms. 196365466904SEric Dumazet * 196465466904SEric Dumazet * Use min_rtt to help adapt TSO burst size, with smaller min_rtt resulting 196565466904SEric Dumazet * in bigger TSO bursts. We we cut the RTT-based allowance in half 196665466904SEric Dumazet * for every 2^9 usec (aka 512 us) of RTT, so that the RTT-based allowance 196765466904SEric Dumazet * is below 1500 bytes after 6 * ~500 usec = 3ms. 1968605ad7f1SEric Dumazet */ 1969dcb8c9b4SEric Dumazet static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now, 19701b3878caSNeal Cardwell int min_tso_segs) 1971605ad7f1SEric Dumazet { 197265466904SEric Dumazet unsigned long bytes; 197365466904SEric Dumazet u32 r; 1974605ad7f1SEric Dumazet 197528b24f90SEric Dumazet bytes = READ_ONCE(sk->sk_pacing_rate) >> READ_ONCE(sk->sk_pacing_shift); 1976605ad7f1SEric Dumazet 19772455e61bSKuniyuki Iwashima r = tcp_min_rtt(tcp_sk(sk)) >> READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_rtt_log); 197865466904SEric Dumazet if (r < BITS_PER_TYPE(sk->sk_gso_max_size)) 197965466904SEric Dumazet bytes += sk->sk_gso_max_size >> r; 1980605ad7f1SEric Dumazet 198165466904SEric Dumazet bytes = min_t(unsigned long, bytes, sk->sk_gso_max_size); 198265466904SEric Dumazet 198365466904SEric Dumazet return max_t(u32, bytes / mss_now, min_tso_segs); 1984605ad7f1SEric Dumazet } 1985605ad7f1SEric Dumazet 1986ed6e7268SNeal Cardwell /* Return the number of segments we want in the skb we are transmitting. 1987ed6e7268SNeal Cardwell * See if congestion control module wants to decide; otherwise, autosize. 1988ed6e7268SNeal Cardwell */ 1989ed6e7268SNeal Cardwell static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now) 1990ed6e7268SNeal Cardwell { 1991ed6e7268SNeal Cardwell const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 1992dcb8c9b4SEric Dumazet u32 min_tso, tso_segs; 1993ed6e7268SNeal Cardwell 1994dcb8c9b4SEric Dumazet min_tso = ca_ops->min_tso_segs ? 1995dcb8c9b4SEric Dumazet ca_ops->min_tso_segs(sk) : 1996e0bb4ab9SKuniyuki Iwashima READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs); 1997dcb8c9b4SEric Dumazet 1998dcb8c9b4SEric Dumazet tso_segs = tcp_tso_autosize(sk, mss_now, min_tso); 1999350c9f48SEric Dumazet return min_t(u32, tso_segs, sk->sk_gso_max_segs); 2000ed6e7268SNeal Cardwell } 2001ed6e7268SNeal Cardwell 2002d4589926SEric Dumazet /* Returns the portion of skb which can be sent right away */ 2003d4589926SEric Dumazet static unsigned int tcp_mss_split_point(const struct sock *sk, 2004d4589926SEric Dumazet const struct sk_buff *skb, 2005d4589926SEric Dumazet unsigned int mss_now, 2006d4589926SEric Dumazet unsigned int max_segs, 2007d4589926SEric Dumazet int nonagle) 2008c1b4a7e6SDavid S. Miller { 2009cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 2010d4589926SEric Dumazet u32 partial, needed, window, max_len; 2011c1b4a7e6SDavid S. Miller 201290840defSIlpo Järvinen window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 20131485348dSBen Hutchings max_len = mss_now * max_segs; 20140e3a4803SIlpo Järvinen 20151485348dSBen Hutchings if (likely(max_len <= window && skb != tcp_write_queue_tail(sk))) 20161485348dSBen Hutchings return max_len; 20170e3a4803SIlpo Järvinen 20185ea3a748SIlpo Järvinen needed = min(skb->len, window); 20195ea3a748SIlpo Järvinen 20201485348dSBen Hutchings if (max_len <= needed) 20211485348dSBen Hutchings return max_len; 20220e3a4803SIlpo Järvinen 2023d4589926SEric Dumazet partial = needed % mss_now; 2024d4589926SEric Dumazet /* If last segment is not a full MSS, check if Nagle rules allow us 2025d4589926SEric Dumazet * to include this last segment in this skb. 2026d4589926SEric Dumazet * Otherwise, we'll split the skb at last MSS boundary 2027d4589926SEric Dumazet */ 2028cc93fc51SPeter Pan(潘卫平) if (tcp_nagle_check(partial != 0, tp, nonagle)) 2029d4589926SEric Dumazet return needed - partial; 2030d4589926SEric Dumazet 2031d4589926SEric Dumazet return needed; 2032c1b4a7e6SDavid S. Miller } 2033c1b4a7e6SDavid S. Miller 2034c1b4a7e6SDavid S. Miller /* Can at least one segment of SKB be sent right now, according to the 2035c1b4a7e6SDavid S. Miller * congestion window rules? If so, return how many segments are allowed. 2036c1b4a7e6SDavid S. Miller */ 2037cf533ea5SEric Dumazet static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, 2038cf533ea5SEric Dumazet const struct sk_buff *skb) 2039c1b4a7e6SDavid S. Miller { 2040d649a7a8SEric Dumazet u32 in_flight, cwnd, halfcwnd; 2041c1b4a7e6SDavid S. Miller 2042c1b4a7e6SDavid S. Miller /* Don't be strict about the congestion window for the final FIN. */ 20434de075e0SEric Dumazet if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && 20444de075e0SEric Dumazet tcp_skb_pcount(skb) == 1) 2045c1b4a7e6SDavid S. Miller return 1; 2046c1b4a7e6SDavid S. Miller 2047c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 204840570375SEric Dumazet cwnd = tcp_snd_cwnd(tp); 2049d649a7a8SEric Dumazet if (in_flight >= cwnd) 2050c1b4a7e6SDavid S. Miller return 0; 2051d649a7a8SEric Dumazet 2052d649a7a8SEric Dumazet /* For better scheduling, ensure we have at least 2053d649a7a8SEric Dumazet * 2 GSO packets in flight. 2054d649a7a8SEric Dumazet */ 2055d649a7a8SEric Dumazet halfcwnd = max(cwnd >> 1, 1U); 2056d649a7a8SEric Dumazet return min(halfcwnd, cwnd - in_flight); 2057c1b4a7e6SDavid S. Miller } 2058c1b4a7e6SDavid S. Miller 2059b595076aSUwe Kleine-König /* Initialize TSO state of a skb. 206067edfef7SAndi Kleen * This must be invoked the first time we consider transmitting 2061c1b4a7e6SDavid S. Miller * SKB onto the wire. 2062c1b4a7e6SDavid S. Miller */ 20635bbb432cSEric Dumazet static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now) 2064c1b4a7e6SDavid S. Miller { 2065c1b4a7e6SDavid S. Miller int tso_segs = tcp_skb_pcount(skb); 2066c1b4a7e6SDavid S. Miller 2067f8269a49SIlpo Järvinen if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { 20685bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 2069c1b4a7e6SDavid S. Miller tso_segs = tcp_skb_pcount(skb); 2070c1b4a7e6SDavid S. Miller } 2071c1b4a7e6SDavid S. Miller return tso_segs; 2072c1b4a7e6SDavid S. Miller } 2073c1b4a7e6SDavid S. Miller 2074c1b4a7e6SDavid S. Miller 2075a2a385d6SEric Dumazet /* Return true if the Nagle test allows this packet to be 2076c1b4a7e6SDavid S. Miller * sent now. 2077c1b4a7e6SDavid S. Miller */ 2078a2a385d6SEric Dumazet static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, 2079c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 2080c1b4a7e6SDavid S. Miller { 2081c1b4a7e6SDavid S. Miller /* Nagle rule does not apply to frames, which sit in the middle of the 2082c1b4a7e6SDavid S. Miller * write_queue (they have no chances to get new data). 2083c1b4a7e6SDavid S. Miller * 2084c1b4a7e6SDavid S. Miller * This is implemented in the callers, where they modify the 'nonagle' 2085c1b4a7e6SDavid S. Miller * argument based upon the location of SKB in the send queue. 2086c1b4a7e6SDavid S. Miller */ 2087c1b4a7e6SDavid S. Miller if (nonagle & TCP_NAGLE_PUSH) 2088a2a385d6SEric Dumazet return true; 2089c1b4a7e6SDavid S. Miller 20909b44190dSYuchung Cheng /* Don't use the nagle rule for urgent data (or for the final FIN). */ 20919b44190dSYuchung Cheng if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) 2092a2a385d6SEric Dumazet return true; 2093c1b4a7e6SDavid S. Miller 2094cc93fc51SPeter Pan(潘卫平) if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle)) 2095a2a385d6SEric Dumazet return true; 2096c1b4a7e6SDavid S. Miller 2097a2a385d6SEric Dumazet return false; 2098c1b4a7e6SDavid S. Miller } 2099c1b4a7e6SDavid S. Miller 2100c1b4a7e6SDavid S. Miller /* Does at least the first segment of SKB fit into the send window? */ 2101a2a385d6SEric Dumazet static bool tcp_snd_wnd_test(const struct tcp_sock *tp, 2102a2a385d6SEric Dumazet const struct sk_buff *skb, 2103056834d9SIlpo Järvinen unsigned int cur_mss) 2104c1b4a7e6SDavid S. Miller { 2105c1b4a7e6SDavid S. Miller u32 end_seq = TCP_SKB_CB(skb)->end_seq; 2106c1b4a7e6SDavid S. Miller 2107c1b4a7e6SDavid S. Miller if (skb->len > cur_mss) 2108c1b4a7e6SDavid S. Miller end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 2109c1b4a7e6SDavid S. Miller 211090840defSIlpo Järvinen return !after(end_seq, tcp_wnd_end(tp)); 2111c1b4a7e6SDavid S. Miller } 2112c1b4a7e6SDavid S. Miller 2113c1b4a7e6SDavid S. Miller /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 2114c1b4a7e6SDavid S. Miller * which is put after SKB on the list. It is very much like 2115c1b4a7e6SDavid S. Miller * tcp_fragment() except that it may make several kinds of assumptions 2116c1b4a7e6SDavid S. Miller * in order to speed up the splitting operation. In particular, we 2117c1b4a7e6SDavid S. Miller * know that all the data is in scatter-gather pages, and that the 2118c1b4a7e6SDavid S. Miller * packet has never been sent out before (and thus is not cloned). 2119c1b4a7e6SDavid S. Miller */ 212056483341SEric Dumazet static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, 2121c4ead4c5SEric Dumazet unsigned int mss_now, gfp_t gfp) 2122c1b4a7e6SDavid S. Miller { 2123c1b4a7e6SDavid S. Miller int nlen = skb->len - len; 212456483341SEric Dumazet struct sk_buff *buff; 21259ce01461SIlpo Järvinen u8 flags; 2126c1b4a7e6SDavid S. Miller 2127c1b4a7e6SDavid S. Miller /* All of a TSO frame must be composed of paged data. */ 2128b4a24397SEric Dumazet DEBUG_NET_WARN_ON_ONCE(skb->len != skb->data_len); 2129c1b4a7e6SDavid S. Miller 21305882efffSEric Dumazet buff = tcp_stream_alloc_skb(sk, gfp, true); 213151456b29SIan Morris if (unlikely(!buff)) 2132c1b4a7e6SDavid S. Miller return -ENOMEM; 213341477662SJakub Kicinski skb_copy_decrypted(buff, skb); 21345a369ca6SPaolo Abeni mptcp_skb_ext_copy(buff, skb); 2135c1b4a7e6SDavid S. Miller 2136ab4e846aSEric Dumazet sk_wmem_queued_add(sk, buff->truesize); 21373ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 2138b60b49eaSHerbert Xu buff->truesize += nlen; 2139c1b4a7e6SDavid S. Miller skb->truesize -= nlen; 2140c1b4a7e6SDavid S. Miller 2141c1b4a7e6SDavid S. Miller /* Correct the sequence numbers. */ 2142c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 2143c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 2144c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 2145c1b4a7e6SDavid S. Miller 2146c1b4a7e6SDavid S. Miller /* PSH and FIN should only be set in the second packet. */ 21474de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 21484de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 21494de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 2150c1b4a7e6SDavid S. Miller 2151a166140eSMartin KaFai Lau tcp_skb_fragment_eor(skb, buff); 2152a166140eSMartin KaFai Lau 2153c1b4a7e6SDavid S. Miller skb_split(skb, buff, len); 2154490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 2155c1b4a7e6SDavid S. Miller 2156c1b4a7e6SDavid S. Miller /* Fix up tso_factor for both original and new SKB. */ 21575bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 21585bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 2159c1b4a7e6SDavid S. Miller 2160c1b4a7e6SDavid S. Miller /* Link BUFF into the send queue. */ 2161f4a775d1SEric Dumazet __skb_header_release(buff); 216256483341SEric Dumazet tcp_insert_write_queue_after(skb, buff, sk, TCP_FRAG_IN_WRITE_QUEUE); 2163c1b4a7e6SDavid S. Miller 2164c1b4a7e6SDavid S. Miller return 0; 2165c1b4a7e6SDavid S. Miller } 2166c1b4a7e6SDavid S. Miller 2167c1b4a7e6SDavid S. Miller /* Try to defer sending, if possible, in order to minimize the amount 2168c1b4a7e6SDavid S. Miller * of TSO splitting we do. View it as a kind of TSO Nagle test. 2169c1b4a7e6SDavid S. Miller * 2170c1b4a7e6SDavid S. Miller * This algorithm is from John Heffner. 2171c1b4a7e6SDavid S. Miller */ 2172ca8a2263SNeal Cardwell static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb, 2173f9bfe4e6SEric Dumazet bool *is_cwnd_limited, 2174f9bfe4e6SEric Dumazet bool *is_rwnd_limited, 2175f9bfe4e6SEric Dumazet u32 max_segs) 2176c1b4a7e6SDavid S. Miller { 21776687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 2178f1c6ea38SEric Dumazet u32 send_win, cong_win, limit, in_flight; 217950c8339eSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 218050c8339eSEric Dumazet struct sk_buff *head; 2181ad9f4f50SEric Dumazet int win_divisor; 2182f1c6ea38SEric Dumazet s64 delta; 2183c1b4a7e6SDavid S. Miller 218499d7662aSEric Dumazet if (icsk->icsk_ca_state >= TCP_CA_Recovery) 2185ae8064acSJohn Heffner goto send_now; 2186ae8064acSJohn Heffner 21875f852eb5SEric Dumazet /* Avoid bursty behavior by allowing defer 2188a682850aSEric Dumazet * only if the last write was recent (1 ms). 2189a682850aSEric Dumazet * Note that tp->tcp_wstamp_ns can be in the future if we have 2190a682850aSEric Dumazet * packets waiting in a qdisc or device for EDT delivery. 21915f852eb5SEric Dumazet */ 2192a682850aSEric Dumazet delta = tp->tcp_clock_cache - tp->tcp_wstamp_ns - NSEC_PER_MSEC; 2193a682850aSEric Dumazet if (delta > 0) 2194ae8064acSJohn Heffner goto send_now; 2195908a75c1SDavid S. Miller 2196c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 2197c1b4a7e6SDavid S. Miller 2198c8c9aeb5SStefano Brivio BUG_ON(tcp_skb_pcount(skb) <= 1); 219940570375SEric Dumazet BUG_ON(tcp_snd_cwnd(tp) <= in_flight); 2200c1b4a7e6SDavid S. Miller 220190840defSIlpo Järvinen send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 2202c1b4a7e6SDavid S. Miller 2203c1b4a7e6SDavid S. Miller /* From in_flight test above, we know that cwnd > in_flight. */ 220440570375SEric Dumazet cong_win = (tcp_snd_cwnd(tp) - in_flight) * tp->mss_cache; 2205c1b4a7e6SDavid S. Miller 2206c1b4a7e6SDavid S. Miller limit = min(send_win, cong_win); 2207c1b4a7e6SDavid S. Miller 2208ba244fe9SDavid S. Miller /* If a full-sized TSO skb can be sent, do it. */ 2209605ad7f1SEric Dumazet if (limit >= max_segs * tp->mss_cache) 2210ae8064acSJohn Heffner goto send_now; 2211ba244fe9SDavid S. Miller 221262ad2761SIlpo Järvinen /* Middle in queue won't get any more data, full sendable already? */ 221362ad2761SIlpo Järvinen if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) 221462ad2761SIlpo Järvinen goto send_now; 221562ad2761SIlpo Järvinen 22165bbcc0f5SLinus Torvalds win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor); 2217ad9f4f50SEric Dumazet if (win_divisor) { 221840570375SEric Dumazet u32 chunk = min(tp->snd_wnd, tcp_snd_cwnd(tp) * tp->mss_cache); 2219c1b4a7e6SDavid S. Miller 2220c1b4a7e6SDavid S. Miller /* If at least some fraction of a window is available, 2221c1b4a7e6SDavid S. Miller * just use it. 2222c1b4a7e6SDavid S. Miller */ 2223ad9f4f50SEric Dumazet chunk /= win_divisor; 2224c1b4a7e6SDavid S. Miller if (limit >= chunk) 2225ae8064acSJohn Heffner goto send_now; 2226c1b4a7e6SDavid S. Miller } else { 2227c1b4a7e6SDavid S. Miller /* Different approach, try not to defer past a single 2228c1b4a7e6SDavid S. Miller * ACK. Receiver should ACK every other full sized 2229c1b4a7e6SDavid S. Miller * frame, so if we have space for more than 3 frames 2230c1b4a7e6SDavid S. Miller * then send now. 2231c1b4a7e6SDavid S. Miller */ 22326b5a5c0dSNeal Cardwell if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) 2233ae8064acSJohn Heffner goto send_now; 2234c1b4a7e6SDavid S. Miller } 2235c1b4a7e6SDavid S. Miller 223675c119afSEric Dumazet /* TODO : use tsorted_sent_queue ? */ 223775c119afSEric Dumazet head = tcp_rtx_queue_head(sk); 223875c119afSEric Dumazet if (!head) 223975c119afSEric Dumazet goto send_now; 2240f1c6ea38SEric Dumazet delta = tp->tcp_clock_cache - head->tstamp; 224150c8339eSEric Dumazet /* If next ACK is likely to come too late (half srtt), do not defer */ 2242f1c6ea38SEric Dumazet if ((s64)(delta - (u64)NSEC_PER_USEC * (tp->srtt_us >> 4)) < 0) 224350c8339eSEric Dumazet goto send_now; 224450c8339eSEric Dumazet 2245f9bfe4e6SEric Dumazet /* Ok, it looks like it is advisable to defer. 2246f9bfe4e6SEric Dumazet * Three cases are tracked : 2247f9bfe4e6SEric Dumazet * 1) We are cwnd-limited 2248f9bfe4e6SEric Dumazet * 2) We are rwnd-limited 2249f9bfe4e6SEric Dumazet * 3) We are application limited. 2250f9bfe4e6SEric Dumazet */ 2251f9bfe4e6SEric Dumazet if (cong_win < send_win) { 2252f9bfe4e6SEric Dumazet if (cong_win <= skb->len) { 2253ca8a2263SNeal Cardwell *is_cwnd_limited = true; 2254f9bfe4e6SEric Dumazet return true; 2255f9bfe4e6SEric Dumazet } 2256f9bfe4e6SEric Dumazet } else { 2257f9bfe4e6SEric Dumazet if (send_win <= skb->len) { 2258f9bfe4e6SEric Dumazet *is_rwnd_limited = true; 2259f9bfe4e6SEric Dumazet return true; 2260f9bfe4e6SEric Dumazet } 2261f9bfe4e6SEric Dumazet } 2262f9bfe4e6SEric Dumazet 2263f9bfe4e6SEric Dumazet /* If this packet won't get more data, do not wait. */ 2264d8ed257fSEric Dumazet if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) || 2265d8ed257fSEric Dumazet TCP_SKB_CB(skb)->eor) 2266f9bfe4e6SEric Dumazet goto send_now; 2267ca8a2263SNeal Cardwell 2268a2a385d6SEric Dumazet return true; 2269ae8064acSJohn Heffner 2270ae8064acSJohn Heffner send_now: 2271a2a385d6SEric Dumazet return false; 2272c1b4a7e6SDavid S. Miller } 2273c1b4a7e6SDavid S. Miller 227405cbc0dbSFan Du static inline void tcp_mtu_check_reprobe(struct sock *sk) 227505cbc0dbSFan Du { 227605cbc0dbSFan Du struct inet_connection_sock *icsk = inet_csk(sk); 227705cbc0dbSFan Du struct tcp_sock *tp = tcp_sk(sk); 227805cbc0dbSFan Du struct net *net = sock_net(sk); 227905cbc0dbSFan Du u32 interval; 228005cbc0dbSFan Du s32 delta; 228105cbc0dbSFan Du 22822a85388fSKuniyuki Iwashima interval = READ_ONCE(net->ipv4.sysctl_tcp_probe_interval); 2283c74df29aSEric Dumazet delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp; 228405cbc0dbSFan Du if (unlikely(delta >= interval * HZ)) { 228505cbc0dbSFan Du int mss = tcp_current_mss(sk); 228605cbc0dbSFan Du 228705cbc0dbSFan Du /* Update current search range */ 228805cbc0dbSFan Du icsk->icsk_mtup.probe_size = 0; 228905cbc0dbSFan Du icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + 229005cbc0dbSFan Du sizeof(struct tcphdr) + 229105cbc0dbSFan Du icsk->icsk_af_ops->net_header_len; 229205cbc0dbSFan Du icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss); 229305cbc0dbSFan Du 229405cbc0dbSFan Du /* Update probe time stamp */ 2295c74df29aSEric Dumazet icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; 229605cbc0dbSFan Du } 229705cbc0dbSFan Du } 229805cbc0dbSFan Du 2299808cf9e3SIlya Lesokhin static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len) 2300808cf9e3SIlya Lesokhin { 2301808cf9e3SIlya Lesokhin struct sk_buff *skb, *next; 2302808cf9e3SIlya Lesokhin 2303808cf9e3SIlya Lesokhin skb = tcp_send_head(sk); 2304808cf9e3SIlya Lesokhin tcp_for_write_queue_from_safe(skb, next, sk) { 2305808cf9e3SIlya Lesokhin if (len <= skb->len) 2306808cf9e3SIlya Lesokhin break; 2307808cf9e3SIlya Lesokhin 23089b65b17dSTalal Ahmad if (unlikely(TCP_SKB_CB(skb)->eor) || 23099b65b17dSTalal Ahmad tcp_has_tx_tstamp(skb) || 23109b65b17dSTalal Ahmad !skb_pure_zcopy_same(skb, next)) 2311808cf9e3SIlya Lesokhin return false; 2312808cf9e3SIlya Lesokhin 2313808cf9e3SIlya Lesokhin len -= skb->len; 2314808cf9e3SIlya Lesokhin } 2315808cf9e3SIlya Lesokhin 2316808cf9e3SIlya Lesokhin return true; 2317808cf9e3SIlya Lesokhin } 2318808cf9e3SIlya Lesokhin 231973601329SEric Dumazet static int tcp_clone_payload(struct sock *sk, struct sk_buff *to, 232073601329SEric Dumazet int probe_size) 232173601329SEric Dumazet { 232273601329SEric Dumazet skb_frag_t *lastfrag = NULL, *fragto = skb_shinfo(to)->frags; 232373601329SEric Dumazet int i, todo, len = 0, nr_frags = 0; 232473601329SEric Dumazet const struct sk_buff *skb; 232573601329SEric Dumazet 232673601329SEric Dumazet if (!sk_wmem_schedule(sk, to->truesize + probe_size)) 232773601329SEric Dumazet return -ENOMEM; 232873601329SEric Dumazet 232973601329SEric Dumazet skb_queue_walk(&sk->sk_write_queue, skb) { 233073601329SEric Dumazet const skb_frag_t *fragfrom = skb_shinfo(skb)->frags; 233173601329SEric Dumazet 233273601329SEric Dumazet if (skb_headlen(skb)) 233373601329SEric Dumazet return -EINVAL; 233473601329SEric Dumazet 233573601329SEric Dumazet for (i = 0; i < skb_shinfo(skb)->nr_frags; i++, fragfrom++) { 233673601329SEric Dumazet if (len >= probe_size) 233773601329SEric Dumazet goto commit; 233873601329SEric Dumazet todo = min_t(int, skb_frag_size(fragfrom), 233973601329SEric Dumazet probe_size - len); 234073601329SEric Dumazet len += todo; 234173601329SEric Dumazet if (lastfrag && 234273601329SEric Dumazet skb_frag_page(fragfrom) == skb_frag_page(lastfrag) && 234373601329SEric Dumazet skb_frag_off(fragfrom) == skb_frag_off(lastfrag) + 234473601329SEric Dumazet skb_frag_size(lastfrag)) { 234573601329SEric Dumazet skb_frag_size_add(lastfrag, todo); 234673601329SEric Dumazet continue; 234773601329SEric Dumazet } 234873601329SEric Dumazet if (unlikely(nr_frags == MAX_SKB_FRAGS)) 234973601329SEric Dumazet return -E2BIG; 235073601329SEric Dumazet skb_frag_page_copy(fragto, fragfrom); 235173601329SEric Dumazet skb_frag_off_copy(fragto, fragfrom); 235273601329SEric Dumazet skb_frag_size_set(fragto, todo); 235373601329SEric Dumazet nr_frags++; 235473601329SEric Dumazet lastfrag = fragto++; 235573601329SEric Dumazet } 235673601329SEric Dumazet } 235773601329SEric Dumazet commit: 235873601329SEric Dumazet WARN_ON_ONCE(len != probe_size); 235973601329SEric Dumazet for (i = 0; i < nr_frags; i++) 236073601329SEric Dumazet skb_frag_ref(to, i); 236173601329SEric Dumazet 236273601329SEric Dumazet skb_shinfo(to)->nr_frags = nr_frags; 236373601329SEric Dumazet to->truesize += probe_size; 236473601329SEric Dumazet to->len += probe_size; 236573601329SEric Dumazet to->data_len += probe_size; 236673601329SEric Dumazet __skb_header_release(to); 236773601329SEric Dumazet return 0; 236873601329SEric Dumazet } 236973601329SEric Dumazet 23705d424d5aSJohn Heffner /* Create a new MTU probe if we are ready. 237167edfef7SAndi Kleen * MTU probe is regularly attempting to increase the path MTU by 237267edfef7SAndi Kleen * deliberately sending larger packets. This discovers routing 237367edfef7SAndi Kleen * changes resulting in larger path MTUs. 237467edfef7SAndi Kleen * 23755d424d5aSJohn Heffner * Returns 0 if we should wait to probe (no cwnd available), 23765d424d5aSJohn Heffner * 1 if a probe was sent, 2377056834d9SIlpo Järvinen * -1 otherwise 2378056834d9SIlpo Järvinen */ 23795d424d5aSJohn Heffner static int tcp_mtu_probe(struct sock *sk) 23805d424d5aSJohn Heffner { 23815d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 238212a59abcSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 23835d424d5aSJohn Heffner struct sk_buff *skb, *nskb, *next; 23846b58e0a5SFan Du struct net *net = sock_net(sk); 23855d424d5aSJohn Heffner int probe_size; 238691cc17c0SIlpo Järvinen int size_needed; 238712a59abcSEric Dumazet int copy, len; 23885d424d5aSJohn Heffner int mss_now; 23896b58e0a5SFan Du int interval; 23905d424d5aSJohn Heffner 23915d424d5aSJohn Heffner /* Not currently probing/verifying, 23925d424d5aSJohn Heffner * not in recovery, 23935d424d5aSJohn Heffner * have enough cwnd, and 239412a59abcSEric Dumazet * not SACKing (the variable headers throw things off) 239512a59abcSEric Dumazet */ 239612a59abcSEric Dumazet if (likely(!icsk->icsk_mtup.enabled || 23975d424d5aSJohn Heffner icsk->icsk_mtup.probe_size || 23985d424d5aSJohn Heffner inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 239940570375SEric Dumazet tcp_snd_cwnd(tp) < 11 || 240012a59abcSEric Dumazet tp->rx_opt.num_sacks || tp->rx_opt.dsack)) 24015d424d5aSJohn Heffner return -1; 24025d424d5aSJohn Heffner 24036b58e0a5SFan Du /* Use binary search for probe_size between tcp_mss_base, 24046b58e0a5SFan Du * and current mss_clamp. if (search_high - search_low) 24056b58e0a5SFan Du * smaller than a threshold, backoff from probing. 24066b58e0a5SFan Du */ 24070c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 24086b58e0a5SFan Du probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high + 24096b58e0a5SFan Du icsk->icsk_mtup.search_low) >> 1); 241091cc17c0SIlpo Järvinen size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; 24116b58e0a5SFan Du interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low; 241205cbc0dbSFan Du /* When misfortune happens, we are reprobing actively, 241305cbc0dbSFan Du * and then reprobe timer has expired. We stick with current 241405cbc0dbSFan Du * probing process by not resetting search range to its orignal. 241505cbc0dbSFan Du */ 24166b58e0a5SFan Du if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) || 241792c0aa41SKuniyuki Iwashima interval < READ_ONCE(net->ipv4.sysctl_tcp_probe_threshold)) { 241805cbc0dbSFan Du /* Check whether enough time has elaplased for 241905cbc0dbSFan Du * another round of probing. 242005cbc0dbSFan Du */ 242105cbc0dbSFan Du tcp_mtu_check_reprobe(sk); 24225d424d5aSJohn Heffner return -1; 24235d424d5aSJohn Heffner } 24245d424d5aSJohn Heffner 24255d424d5aSJohn Heffner /* Have enough data in the send queue to probe? */ 24267f9c33e5SIlpo Järvinen if (tp->write_seq - tp->snd_nxt < size_needed) 24275d424d5aSJohn Heffner return -1; 24285d424d5aSJohn Heffner 242991cc17c0SIlpo Järvinen if (tp->snd_wnd < size_needed) 24305d424d5aSJohn Heffner return -1; 243190840defSIlpo Järvinen if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) 24325d424d5aSJohn Heffner return 0; 24335d424d5aSJohn Heffner 2434d67c58e9SIlpo Järvinen /* Do we need to wait to drain cwnd? With none in flight, don't stall */ 243540570375SEric Dumazet if (tcp_packets_in_flight(tp) + 2 > tcp_snd_cwnd(tp)) { 2436d67c58e9SIlpo Järvinen if (!tcp_packets_in_flight(tp)) 24375d424d5aSJohn Heffner return -1; 24385d424d5aSJohn Heffner else 24395d424d5aSJohn Heffner return 0; 24405d424d5aSJohn Heffner } 24415d424d5aSJohn Heffner 2442808cf9e3SIlya Lesokhin if (!tcp_can_coalesce_send_queue_head(sk, probe_size)) 2443808cf9e3SIlya Lesokhin return -1; 2444808cf9e3SIlya Lesokhin 24455d424d5aSJohn Heffner /* We're allowed to probe. Build it now. */ 24465882efffSEric Dumazet nskb = tcp_stream_alloc_skb(sk, GFP_ATOMIC, false); 244751456b29SIan Morris if (!nskb) 24485d424d5aSJohn Heffner return -1; 244973601329SEric Dumazet 245073601329SEric Dumazet /* build the payload, and be prepared to abort if this fails. */ 245173601329SEric Dumazet if (tcp_clone_payload(sk, nskb, probe_size)) { 245271c299c7SJakub Kicinski tcp_skb_tsorted_anchor_cleanup(nskb); 245373601329SEric Dumazet consume_skb(nskb); 245473601329SEric Dumazet return -1; 245573601329SEric Dumazet } 2456ab4e846aSEric Dumazet sk_wmem_queued_add(sk, nskb->truesize); 24573ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 24585d424d5aSJohn Heffner 2459fe067e8aSDavid S. Miller skb = tcp_send_head(sk); 246041477662SJakub Kicinski skb_copy_decrypted(nskb, skb); 24615a369ca6SPaolo Abeni mptcp_skb_ext_copy(nskb, skb); 24625d424d5aSJohn Heffner 24635d424d5aSJohn Heffner TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 24645d424d5aSJohn Heffner TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 24654de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; 24665d424d5aSJohn Heffner 246750c4817eSIlpo Järvinen tcp_insert_write_queue_before(nskb, skb, sk); 24682b7cda9cSEric Dumazet tcp_highest_sack_replace(sk, skb, nskb); 246950c4817eSIlpo Järvinen 24705d424d5aSJohn Heffner len = 0; 2471234b6860SIlpo Järvinen tcp_for_write_queue_from_safe(skb, next, sk) { 24725d424d5aSJohn Heffner copy = min_t(int, skb->len, probe_size - len); 24735d424d5aSJohn Heffner 24745d424d5aSJohn Heffner if (skb->len <= copy) { 24755d424d5aSJohn Heffner /* We've eaten all the data from this skb. 24765d424d5aSJohn Heffner * Throw it away. */ 24774de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 2478808cf9e3SIlya Lesokhin /* If this is the last SKB we copy and eor is set 2479808cf9e3SIlya Lesokhin * we need to propagate it to the new skb. 2480808cf9e3SIlya Lesokhin */ 2481808cf9e3SIlya Lesokhin TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor; 2482888a5c53SWillem de Bruijn tcp_skb_collapse_tstamp(nskb, skb); 2483fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 248403271f3aSTalal Ahmad tcp_wmem_free_skb(sk, skb); 24855d424d5aSJohn Heffner } else { 24864de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & 2487a3433f35SChangli Gao ~(TCPHDR_FIN|TCPHDR_PSH); 24885d424d5aSJohn Heffner __pskb_trim_head(skb, copy); 24895bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 24905d424d5aSJohn Heffner TCP_SKB_CB(skb)->seq += copy; 24915d424d5aSJohn Heffner } 24925d424d5aSJohn Heffner 24935d424d5aSJohn Heffner len += copy; 2494234b6860SIlpo Järvinen 2495234b6860SIlpo Järvinen if (len >= probe_size) 2496234b6860SIlpo Järvinen break; 24975d424d5aSJohn Heffner } 24985bbb432cSEric Dumazet tcp_init_tso_segs(nskb, nskb->len); 24995d424d5aSJohn Heffner 25005d424d5aSJohn Heffner /* We're ready to send. If this fails, the probe will 25017faee5c0SEric Dumazet * be resegmented into mss-sized pieces by tcp_write_xmit(). 25027faee5c0SEric Dumazet */ 25035d424d5aSJohn Heffner if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 25045d424d5aSJohn Heffner /* Decrement cwnd here because we are sending 25055d424d5aSJohn Heffner * effectively two packets. */ 250640570375SEric Dumazet tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) - 1); 250766f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, nskb); 25085d424d5aSJohn Heffner 25095d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 25100e7b1368SJohn Heffner tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 25110e7b1368SJohn Heffner tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 25125d424d5aSJohn Heffner 25135d424d5aSJohn Heffner return 1; 25145d424d5aSJohn Heffner } 25155d424d5aSJohn Heffner 25165d424d5aSJohn Heffner return -1; 25175d424d5aSJohn Heffner } 25185d424d5aSJohn Heffner 2519864e5c09SEric Dumazet static bool tcp_pacing_check(struct sock *sk) 2520218af599SEric Dumazet { 2521864e5c09SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 2522864e5c09SEric Dumazet 2523864e5c09SEric Dumazet if (!tcp_needs_internal_pacing(sk)) 2524864e5c09SEric Dumazet return false; 2525864e5c09SEric Dumazet 2526864e5c09SEric Dumazet if (tp->tcp_wstamp_ns <= tp->tcp_clock_cache) 2527864e5c09SEric Dumazet return false; 2528864e5c09SEric Dumazet 2529864e5c09SEric Dumazet if (!hrtimer_is_queued(&tp->pacing_timer)) { 2530864e5c09SEric Dumazet hrtimer_start(&tp->pacing_timer, 2531864e5c09SEric Dumazet ns_to_ktime(tp->tcp_wstamp_ns), 2532864e5c09SEric Dumazet HRTIMER_MODE_ABS_PINNED_SOFT); 2533864e5c09SEric Dumazet sock_hold(sk); 2534864e5c09SEric Dumazet } 2535864e5c09SEric Dumazet return true; 2536218af599SEric Dumazet } 2537218af599SEric Dumazet 2538f921a4a5SEric Dumazet static bool tcp_rtx_queue_empty_or_single_skb(const struct sock *sk) 2539f921a4a5SEric Dumazet { 2540f921a4a5SEric Dumazet const struct rb_node *node = sk->tcp_rtx_queue.rb_node; 2541f921a4a5SEric Dumazet 2542f921a4a5SEric Dumazet /* No skb in the rtx queue. */ 2543f921a4a5SEric Dumazet if (!node) 2544f921a4a5SEric Dumazet return true; 2545f921a4a5SEric Dumazet 2546f921a4a5SEric Dumazet /* Only one skb in rtx queue. */ 2547f921a4a5SEric Dumazet return !node->rb_left && !node->rb_right; 2548f921a4a5SEric Dumazet } 2549f921a4a5SEric Dumazet 2550f9616c35SEric Dumazet /* TCP Small Queues : 2551f9616c35SEric Dumazet * Control number of packets in qdisc/devices to two packets / or ~1 ms. 2552f9616c35SEric Dumazet * (These limits are doubled for retransmits) 2553f9616c35SEric Dumazet * This allows for : 2554f9616c35SEric Dumazet * - better RTT estimation and ACK scheduling 2555f9616c35SEric Dumazet * - faster recovery 2556f9616c35SEric Dumazet * - high rates 2557f9616c35SEric Dumazet * Alas, some drivers / subsystems require a fair amount 2558f9616c35SEric Dumazet * of queued bytes to ensure line rate. 2559f9616c35SEric Dumazet * One example is wifi aggregation (802.11 AMPDU) 2560f9616c35SEric Dumazet */ 2561f9616c35SEric Dumazet static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb, 2562f9616c35SEric Dumazet unsigned int factor) 2563f9616c35SEric Dumazet { 256476a9ebe8SEric Dumazet unsigned long limit; 2565f9616c35SEric Dumazet 256676a9ebe8SEric Dumazet limit = max_t(unsigned long, 256776a9ebe8SEric Dumazet 2 * skb->truesize, 256828b24f90SEric Dumazet READ_ONCE(sk->sk_pacing_rate) >> READ_ONCE(sk->sk_pacing_shift)); 2569c73e5807SEric Dumazet if (sk->sk_pacing_status == SK_PACING_NONE) 257076a9ebe8SEric Dumazet limit = min_t(unsigned long, limit, 25719fb90193SKuniyuki Iwashima READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes)); 2572f9616c35SEric Dumazet limit <<= factor; 2573f9616c35SEric Dumazet 2574a842fe14SEric Dumazet if (static_branch_unlikely(&tcp_tx_delay_enabled) && 2575a842fe14SEric Dumazet tcp_sk(sk)->tcp_tx_delay) { 257628b24f90SEric Dumazet u64 extra_bytes = (u64)READ_ONCE(sk->sk_pacing_rate) * 257728b24f90SEric Dumazet tcp_sk(sk)->tcp_tx_delay; 2578a842fe14SEric Dumazet 2579a842fe14SEric Dumazet /* TSQ is based on skb truesize sum (sk_wmem_alloc), so we 2580a842fe14SEric Dumazet * approximate our needs assuming an ~100% skb->truesize overhead. 2581a842fe14SEric Dumazet * USEC_PER_SEC is approximated by 2^20. 2582a842fe14SEric Dumazet * do_div(extra_bytes, USEC_PER_SEC/2) is replaced by a right shift. 2583a842fe14SEric Dumazet */ 2584a842fe14SEric Dumazet extra_bytes >>= (20 - 1); 2585a842fe14SEric Dumazet limit += extra_bytes; 2586a842fe14SEric Dumazet } 258714afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) > limit) { 2588f921a4a5SEric Dumazet /* Always send skb if rtx queue is empty or has one skb. 258975eefc6cSEric Dumazet * No need to wait for TX completion to call us back, 259075eefc6cSEric Dumazet * after softirq/tasklet schedule. 259175eefc6cSEric Dumazet * This helps when TX completions are delayed too much. 259275eefc6cSEric Dumazet */ 2593f921a4a5SEric Dumazet if (tcp_rtx_queue_empty_or_single_skb(sk)) 259475eefc6cSEric Dumazet return false; 259575eefc6cSEric Dumazet 25967aa5470cSEric Dumazet set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags); 2597f9616c35SEric Dumazet /* It is possible TX completion already happened 2598f9616c35SEric Dumazet * before we set TSQ_THROTTLED, so we must 2599f9616c35SEric Dumazet * test again the condition. 2600f9616c35SEric Dumazet */ 2601f9616c35SEric Dumazet smp_mb__after_atomic(); 2602ce8299b6SEric Dumazet if (refcount_read(&sk->sk_wmem_alloc) > limit) 2603f9616c35SEric Dumazet return true; 2604f9616c35SEric Dumazet } 2605f9616c35SEric Dumazet return false; 2606f9616c35SEric Dumazet } 2607f9616c35SEric Dumazet 260805b055e8SFrancis Yan static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new) 260905b055e8SFrancis Yan { 2610628174ccSEric Dumazet const u32 now = tcp_jiffies32; 2611efe967cdSArnd Bergmann enum tcp_chrono old = tp->chrono_type; 261205b055e8SFrancis Yan 2613efe967cdSArnd Bergmann if (old > TCP_CHRONO_UNSPEC) 2614efe967cdSArnd Bergmann tp->chrono_stat[old - 1] += now - tp->chrono_start; 261505b055e8SFrancis Yan tp->chrono_start = now; 261605b055e8SFrancis Yan tp->chrono_type = new; 261705b055e8SFrancis Yan } 261805b055e8SFrancis Yan 261905b055e8SFrancis Yan void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type) 262005b055e8SFrancis Yan { 262105b055e8SFrancis Yan struct tcp_sock *tp = tcp_sk(sk); 262205b055e8SFrancis Yan 262305b055e8SFrancis Yan /* If there are multiple conditions worthy of tracking in a 26240f87230dSFrancis Yan * chronograph then the highest priority enum takes precedence 26250f87230dSFrancis Yan * over the other conditions. So that if something "more interesting" 262605b055e8SFrancis Yan * starts happening, stop the previous chrono and start a new one. 262705b055e8SFrancis Yan */ 262805b055e8SFrancis Yan if (type > tp->chrono_type) 262905b055e8SFrancis Yan tcp_chrono_set(tp, type); 263005b055e8SFrancis Yan } 263105b055e8SFrancis Yan 263205b055e8SFrancis Yan void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type) 263305b055e8SFrancis Yan { 263405b055e8SFrancis Yan struct tcp_sock *tp = tcp_sk(sk); 263505b055e8SFrancis Yan 26360f87230dSFrancis Yan 26370f87230dSFrancis Yan /* There are multiple conditions worthy of tracking in a 26380f87230dSFrancis Yan * chronograph, so that the highest priority enum takes 26390f87230dSFrancis Yan * precedence over the other conditions (see tcp_chrono_start). 26400f87230dSFrancis Yan * If a condition stops, we only stop chrono tracking if 26410f87230dSFrancis Yan * it's the "most interesting" or current chrono we are 26420f87230dSFrancis Yan * tracking and starts busy chrono if we have pending data. 26430f87230dSFrancis Yan */ 264475c119afSEric Dumazet if (tcp_rtx_and_write_queues_empty(sk)) 264505b055e8SFrancis Yan tcp_chrono_set(tp, TCP_CHRONO_UNSPEC); 26460f87230dSFrancis Yan else if (type == tp->chrono_type) 26470f87230dSFrancis Yan tcp_chrono_set(tp, TCP_CHRONO_BUSY); 264805b055e8SFrancis Yan } 264905b055e8SFrancis Yan 26501da177e4SLinus Torvalds /* This routine writes packets to the network. It advances the 26511da177e4SLinus Torvalds * send_head. This happens as incoming acks open up the remote 26521da177e4SLinus Torvalds * window for us. 26531da177e4SLinus Torvalds * 2654f8269a49SIlpo Järvinen * LARGESEND note: !tcp_urg_mode is overkill, only frames between 2655f8269a49SIlpo Järvinen * snd_up-64k-mss .. snd_up cannot be large. However, taking into 2656f8269a49SIlpo Järvinen * account rare use of URG, this is not a big flaw. 2657f8269a49SIlpo Järvinen * 26586ba8a3b1SNandita Dukkipati * Send at most one packet when push_one > 0. Temporarily ignore 26596ba8a3b1SNandita Dukkipati * cwnd limit to force at most one packet out when push_one == 2. 26606ba8a3b1SNandita Dukkipati 2661a2a385d6SEric Dumazet * Returns true, if no segments are in flight and we have queued segments, 2662a2a385d6SEric Dumazet * but cannot send anything now because of SWS or another problem. 26631da177e4SLinus Torvalds */ 2664a2a385d6SEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 2665d5dd9175SIlpo Järvinen int push_one, gfp_t gfp) 26661da177e4SLinus Torvalds { 26671da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 266892df7b51SDavid S. Miller struct sk_buff *skb; 2669c1b4a7e6SDavid S. Miller unsigned int tso_segs, sent_pkts; 2670c1b4a7e6SDavid S. Miller int cwnd_quota; 26715d424d5aSJohn Heffner int result; 26725615f886SFrancis Yan bool is_cwnd_limited = false, is_rwnd_limited = false; 2673605ad7f1SEric Dumazet u32 max_segs; 26741da177e4SLinus Torvalds 2675c1b4a7e6SDavid S. Miller sent_pkts = 0; 26765d424d5aSJohn Heffner 2677ee1836aeSEric Dumazet tcp_mstamp_refresh(tp); 2678d5dd9175SIlpo Järvinen if (!push_one) { 26795d424d5aSJohn Heffner /* Do MTU probing. */ 2680d5dd9175SIlpo Järvinen result = tcp_mtu_probe(sk); 2681d5dd9175SIlpo Järvinen if (!result) { 2682a2a385d6SEric Dumazet return false; 26835d424d5aSJohn Heffner } else if (result > 0) { 26845d424d5aSJohn Heffner sent_pkts = 1; 26855d424d5aSJohn Heffner } 2686d5dd9175SIlpo Järvinen } 26875d424d5aSJohn Heffner 2688ed6e7268SNeal Cardwell max_segs = tcp_tso_segs(sk, mss_now); 2689fe067e8aSDavid S. Miller while ((skb = tcp_send_head(sk))) { 2690c8ac3774SHerbert Xu unsigned int limit; 2691c8ac3774SHerbert Xu 269279861919SEric Dumazet if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) { 269379861919SEric Dumazet /* "skb_mstamp_ns" is used as a start point for the retransmit timer */ 2694a1ac9c8aSMartin KaFai Lau tp->tcp_wstamp_ns = tp->tcp_clock_cache; 2695a1ac9c8aSMartin KaFai Lau skb_set_delivery_time(skb, tp->tcp_wstamp_ns, true); 269679861919SEric Dumazet list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue); 2697bf50b606SEric Dumazet tcp_init_tso_segs(skb, mss_now); 269879861919SEric Dumazet goto repair; /* Skip network transmission */ 269979861919SEric Dumazet } 270079861919SEric Dumazet 2701218af599SEric Dumazet if (tcp_pacing_check(sk)) 2702218af599SEric Dumazet break; 2703218af599SEric Dumazet 27045bbb432cSEric Dumazet tso_segs = tcp_init_tso_segs(skb, mss_now); 2705c1b4a7e6SDavid S. Miller BUG_ON(!tso_segs); 2706c1b4a7e6SDavid S. Miller 2707b68e9f85SHerbert Xu cwnd_quota = tcp_cwnd_test(tp, skb); 27086ba8a3b1SNandita Dukkipati if (!cwnd_quota) { 27096ba8a3b1SNandita Dukkipati if (push_one == 2) 27106ba8a3b1SNandita Dukkipati /* Force out a loss probe pkt. */ 27116ba8a3b1SNandita Dukkipati cwnd_quota = 1; 27126ba8a3b1SNandita Dukkipati else 2713b68e9f85SHerbert Xu break; 27146ba8a3b1SNandita Dukkipati } 2715b68e9f85SHerbert Xu 27165615f886SFrancis Yan if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) { 27175615f886SFrancis Yan is_rwnd_limited = true; 2718b68e9f85SHerbert Xu break; 27195615f886SFrancis Yan } 2720b68e9f85SHerbert Xu 2721d6a4e26aSEric Dumazet if (tso_segs == 1) { 2722aa93466bSDavid S. Miller if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 2723aa93466bSDavid S. Miller (tcp_skb_is_last(sk, skb) ? 2724aa93466bSDavid S. Miller nonagle : TCP_NAGLE_PUSH)))) 2725aa93466bSDavid S. Miller break; 2726c1b4a7e6SDavid S. Miller } else { 2727ca8a2263SNeal Cardwell if (!push_one && 2728605ad7f1SEric Dumazet tcp_tso_should_defer(sk, skb, &is_cwnd_limited, 2729f9bfe4e6SEric Dumazet &is_rwnd_limited, max_segs)) 2730aa93466bSDavid S. Miller break; 2731c1b4a7e6SDavid S. Miller } 2732aa93466bSDavid S. Miller 2733605ad7f1SEric Dumazet limit = mss_now; 2734d6a4e26aSEric Dumazet if (tso_segs > 1 && !tcp_urg_mode(tp)) 2735605ad7f1SEric Dumazet limit = tcp_mss_split_point(sk, skb, mss_now, 2736605ad7f1SEric Dumazet min_t(unsigned int, 2737605ad7f1SEric Dumazet cwnd_quota, 2738605ad7f1SEric Dumazet max_segs), 2739605ad7f1SEric Dumazet nonagle); 2740605ad7f1SEric Dumazet 2741605ad7f1SEric Dumazet if (skb->len > limit && 274256483341SEric Dumazet unlikely(tso_fragment(sk, skb, limit, mss_now, gfp))) 2743605ad7f1SEric Dumazet break; 2744605ad7f1SEric Dumazet 2745f9616c35SEric Dumazet if (tcp_small_queue_check(sk, skb, 0)) 274646d3ceabSEric Dumazet break; 2747c9eeec26SEric Dumazet 27481f85e626SEric Dumazet /* Argh, we hit an empty skb(), presumably a thread 27491f85e626SEric Dumazet * is sleeping in sendmsg()/sk_stream_wait_memory(). 27501f85e626SEric Dumazet * We do not want to send a pure-ack packet and have 27511f85e626SEric Dumazet * a strange looking rtx queue with empty packet(s). 27521f85e626SEric Dumazet */ 27531f85e626SEric Dumazet if (TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq) 27541f85e626SEric Dumazet break; 27551f85e626SEric Dumazet 2756d5dd9175SIlpo Järvinen if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) 27571da177e4SLinus Torvalds break; 27581da177e4SLinus Torvalds 2759ec342325SAndrew Vagin repair: 27601da177e4SLinus Torvalds /* Advance the send_head. This one is sent out. 27611da177e4SLinus Torvalds * This call will increment packets_out. 27621da177e4SLinus Torvalds */ 276366f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 27641da177e4SLinus Torvalds 27651da177e4SLinus Torvalds tcp_minshall_update(tp, mss_now, skb); 2766a262f0cdSNandita Dukkipati sent_pkts += tcp_skb_pcount(skb); 2767d5dd9175SIlpo Järvinen 2768d5dd9175SIlpo Järvinen if (push_one) 2769d5dd9175SIlpo Järvinen break; 27701da177e4SLinus Torvalds } 27711da177e4SLinus Torvalds 27725615f886SFrancis Yan if (is_rwnd_limited) 27735615f886SFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED); 27745615f886SFrancis Yan else 27755615f886SFrancis Yan tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED); 27765615f886SFrancis Yan 277740570375SEric Dumazet is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tcp_snd_cwnd(tp)); 2778299bcb55SNeal Cardwell if (likely(sent_pkts || is_cwnd_limited)) 2779299bcb55SNeal Cardwell tcp_cwnd_validate(sk, is_cwnd_limited); 2780299bcb55SNeal Cardwell 2781aa93466bSDavid S. Miller if (likely(sent_pkts)) { 2782684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 2783684bad11SYuchung Cheng tp->prr_out += sent_pkts; 27846ba8a3b1SNandita Dukkipati 27856ba8a3b1SNandita Dukkipati /* Send one loss probe per tail loss episode. */ 27866ba8a3b1SNandita Dukkipati if (push_one != 2) 2787ed66dfafSNeal Cardwell tcp_schedule_loss_probe(sk, false); 2788a2a385d6SEric Dumazet return false; 27891da177e4SLinus Torvalds } 279075c119afSEric Dumazet return !tp->packets_out && !tcp_write_queue_empty(sk); 27916ba8a3b1SNandita Dukkipati } 27926ba8a3b1SNandita Dukkipati 2793ed66dfafSNeal Cardwell bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto) 27946ba8a3b1SNandita Dukkipati { 27956ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 27966ba8a3b1SNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 27971c2709cfSNeal Cardwell u32 timeout, timeout_us, rto_delta_us; 27982ae21cf5SEric Dumazet int early_retrans; 27996ba8a3b1SNandita Dukkipati 28006ba8a3b1SNandita Dukkipati /* Don't do any loss probe on a Fast Open connection before 3WHS 28016ba8a3b1SNandita Dukkipati * finishes. 28026ba8a3b1SNandita Dukkipati */ 2803d983ea6fSEric Dumazet if (rcu_access_pointer(tp->fastopen_rsk)) 28046ba8a3b1SNandita Dukkipati return false; 28056ba8a3b1SNandita Dukkipati 280652e65865SKuniyuki Iwashima early_retrans = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_early_retrans); 28076ba8a3b1SNandita Dukkipati /* Schedule a loss probe in 2*RTT for SACK capable connections 2808b4f70c3dSNeal Cardwell * not in loss recovery, that are either limited by cwnd or application. 28096ba8a3b1SNandita Dukkipati */ 28102ae21cf5SEric Dumazet if ((early_retrans != 3 && early_retrans != 4) || 2811bec41a11SYuchung Cheng !tp->packets_out || !tcp_is_sack(tp) || 2812b4f70c3dSNeal Cardwell (icsk->icsk_ca_state != TCP_CA_Open && 2813b4f70c3dSNeal Cardwell icsk->icsk_ca_state != TCP_CA_CWR)) 28146ba8a3b1SNandita Dukkipati return false; 28156ba8a3b1SNandita Dukkipati 2816bb4d991aSYuchung Cheng /* Probe timeout is 2*rtt. Add minimum RTO to account 2817f9b99582SYuchung Cheng * for delayed ack when there's one outstanding packet. If no RTT 2818f9b99582SYuchung Cheng * sample is available then probe after TCP_TIMEOUT_INIT. 28196ba8a3b1SNandita Dukkipati */ 2820bb4d991aSYuchung Cheng if (tp->srtt_us) { 28211c2709cfSNeal Cardwell timeout_us = tp->srtt_us >> 2; 28226ba8a3b1SNandita Dukkipati if (tp->packets_out == 1) 28231c2709cfSNeal Cardwell timeout_us += tcp_rto_min_us(sk); 2824bb4d991aSYuchung Cheng else 28251c2709cfSNeal Cardwell timeout_us += TCP_TIMEOUT_MIN_US; 28261c2709cfSNeal Cardwell timeout = usecs_to_jiffies(timeout_us); 2827bb4d991aSYuchung Cheng } else { 2828bb4d991aSYuchung Cheng timeout = TCP_TIMEOUT_INIT; 2829bb4d991aSYuchung Cheng } 28306ba8a3b1SNandita Dukkipati 2831a2815817SNeal Cardwell /* If the RTO formula yields an earlier time, then use that time. */ 2832ed66dfafSNeal Cardwell rto_delta_us = advancing_rto ? 2833ed66dfafSNeal Cardwell jiffies_to_usecs(inet_csk(sk)->icsk_rto) : 2834ed66dfafSNeal Cardwell tcp_rto_delta_us(sk); /* How far in future is RTO? */ 2835a2815817SNeal Cardwell if (rto_delta_us > 0) 2836a2815817SNeal Cardwell timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us)); 28376ba8a3b1SNandita Dukkipati 28388dc242adSEric Dumazet tcp_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, TCP_RTO_MAX); 28396ba8a3b1SNandita Dukkipati return true; 28406ba8a3b1SNandita Dukkipati } 28416ba8a3b1SNandita Dukkipati 28421f3279aeSEric Dumazet /* Thanks to skb fast clones, we can detect if a prior transmit of 28431f3279aeSEric Dumazet * a packet is still in a qdisc or driver queue. 28441f3279aeSEric Dumazet * In this case, there is very little point doing a retransmit ! 28451f3279aeSEric Dumazet */ 2846f4dae54eSEric Dumazet static bool skb_still_in_host_queue(struct sock *sk, 28471f3279aeSEric Dumazet const struct sk_buff *skb) 28481f3279aeSEric Dumazet { 284939bb5e62SEric Dumazet if (unlikely(skb_fclone_busy(sk, skb))) { 2850f4dae54eSEric Dumazet set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags); 2851f4dae54eSEric Dumazet smp_mb__after_atomic(); 2852f4dae54eSEric Dumazet if (skb_fclone_busy(sk, skb)) { 2853c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), 28541f3279aeSEric Dumazet LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES); 28551f3279aeSEric Dumazet return true; 28561f3279aeSEric Dumazet } 2857f4dae54eSEric Dumazet } 28581f3279aeSEric Dumazet return false; 28591f3279aeSEric Dumazet } 28601f3279aeSEric Dumazet 2861b340b264SYuchung Cheng /* When probe timeout (PTO) fires, try send a new segment if possible, else 28626ba8a3b1SNandita Dukkipati * retransmit the last segment. 28636ba8a3b1SNandita Dukkipati */ 28646ba8a3b1SNandita Dukkipati void tcp_send_loss_probe(struct sock *sk) 28656ba8a3b1SNandita Dukkipati { 28669b717a8dSNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 28676ba8a3b1SNandita Dukkipati struct sk_buff *skb; 28686ba8a3b1SNandita Dukkipati int pcount; 28696ba8a3b1SNandita Dukkipati int mss = tcp_current_mss(sk); 28706ba8a3b1SNandita Dukkipati 287176be93fcSYuchung Cheng /* At most one outstanding TLP */ 287276be93fcSYuchung Cheng if (tp->tlp_high_seq) 287376be93fcSYuchung Cheng goto rearm_timer; 287476be93fcSYuchung Cheng 287576be93fcSYuchung Cheng tp->tlp_retrans = 0; 2876b340b264SYuchung Cheng skb = tcp_send_head(sk); 287775c119afSEric Dumazet if (skb && tcp_snd_wnd_test(tp, skb, mss)) { 2878b340b264SYuchung Cheng pcount = tp->packets_out; 2879b340b264SYuchung Cheng tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC); 2880b340b264SYuchung Cheng if (tp->packets_out > pcount) 2881b340b264SYuchung Cheng goto probe_sent; 28826ba8a3b1SNandita Dukkipati goto rearm_timer; 28836ba8a3b1SNandita Dukkipati } 288475c119afSEric Dumazet skb = skb_rb_last(&sk->tcp_rtx_queue); 2885b2b7af86SYuchung Cheng if (unlikely(!skb)) { 2886b2b7af86SYuchung Cheng WARN_ONCE(tp->packets_out, 2887b2b7af86SYuchung Cheng "invalid inflight: %u state %u cwnd %u mss %d\n", 288840570375SEric Dumazet tp->packets_out, sk->sk_state, tcp_snd_cwnd(tp), mss); 2889b2b7af86SYuchung Cheng inet_csk(sk)->icsk_pending = 0; 2890b2b7af86SYuchung Cheng return; 2891b2b7af86SYuchung Cheng } 28926ba8a3b1SNandita Dukkipati 28931f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 28941f3279aeSEric Dumazet goto rearm_timer; 28951f3279aeSEric Dumazet 28966ba8a3b1SNandita Dukkipati pcount = tcp_skb_pcount(skb); 28976ba8a3b1SNandita Dukkipati if (WARN_ON(!pcount)) 28986ba8a3b1SNandita Dukkipati goto rearm_timer; 28996ba8a3b1SNandita Dukkipati 29006ba8a3b1SNandita Dukkipati if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) { 290175c119afSEric Dumazet if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, 290275c119afSEric Dumazet (pcount - 1) * mss, mss, 29036cc55e09SOctavian Purdila GFP_ATOMIC))) 29046ba8a3b1SNandita Dukkipati goto rearm_timer; 290575c119afSEric Dumazet skb = skb_rb_next(skb); 29066ba8a3b1SNandita Dukkipati } 29076ba8a3b1SNandita Dukkipati 29086ba8a3b1SNandita Dukkipati if (WARN_ON(!skb || !tcp_skb_pcount(skb))) 29096ba8a3b1SNandita Dukkipati goto rearm_timer; 29106ba8a3b1SNandita Dukkipati 291110d3be56SEric Dumazet if (__tcp_retransmit_skb(sk, skb, 1)) 2912b340b264SYuchung Cheng goto rearm_timer; 29136ba8a3b1SNandita Dukkipati 291476be93fcSYuchung Cheng tp->tlp_retrans = 1; 291576be93fcSYuchung Cheng 291676be93fcSYuchung Cheng probe_sent: 29179b717a8dSNandita Dukkipati /* Record snd_nxt for loss detection. */ 29189b717a8dSNandita Dukkipati tp->tlp_high_seq = tp->snd_nxt; 29199b717a8dSNandita Dukkipati 2920c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES); 2921fcd16c0aSYuchung Cheng /* Reset s.t. tcp_rearm_rto will restart timer from now */ 2922fcd16c0aSYuchung Cheng inet_csk(sk)->icsk_pending = 0; 2923b340b264SYuchung Cheng rearm_timer: 2924fcd16c0aSYuchung Cheng tcp_rearm_rto(sk); 29251da177e4SLinus Torvalds } 29261da177e4SLinus Torvalds 2927a762a980SDavid S. Miller /* Push out any pending frames which were held back due to 2928a762a980SDavid S. Miller * TCP_CORK or attempt at coalescing tiny packets. 2929a762a980SDavid S. Miller * The socket must be locked by the caller. 2930a762a980SDavid S. Miller */ 29319e412ba7SIlpo Järvinen void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 29329e412ba7SIlpo Järvinen int nonagle) 2933a762a980SDavid S. Miller { 2934726e07a8SIlpo Järvinen /* If we are closed, the bytes will have to remain here. 2935726e07a8SIlpo Järvinen * In time closedown will finish, we empty the write queue and 2936726e07a8SIlpo Järvinen * all will be happy. 2937726e07a8SIlpo Järvinen */ 2938726e07a8SIlpo Järvinen if (unlikely(sk->sk_state == TCP_CLOSE)) 2939726e07a8SIlpo Järvinen return; 2940726e07a8SIlpo Järvinen 294199a1dec7SMel Gorman if (tcp_write_xmit(sk, cur_mss, nonagle, 0, 29427450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC))) 29439e412ba7SIlpo Järvinen tcp_check_probe_timer(sk); 2944a762a980SDavid S. Miller } 2945a762a980SDavid S. Miller 2946c1b4a7e6SDavid S. Miller /* Send _single_ skb sitting at the send head. This function requires 2947c1b4a7e6SDavid S. Miller * true push pending frames to setup probe timer etc. 2948c1b4a7e6SDavid S. Miller */ 2949c1b4a7e6SDavid S. Miller void tcp_push_one(struct sock *sk, unsigned int mss_now) 2950c1b4a7e6SDavid S. Miller { 2951fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 2952c1b4a7e6SDavid S. Miller 2953c1b4a7e6SDavid S. Miller BUG_ON(!skb || skb->len < mss_now); 2954c1b4a7e6SDavid S. Miller 2955d5dd9175SIlpo Järvinen tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); 2956c1b4a7e6SDavid S. Miller } 2957c1b4a7e6SDavid S. Miller 29581da177e4SLinus Torvalds /* This function returns the amount that we can raise the 29591da177e4SLinus Torvalds * usable window based on the following constraints 29601da177e4SLinus Torvalds * 29611da177e4SLinus Torvalds * 1. The window can never be shrunk once it is offered (RFC 793) 29621da177e4SLinus Torvalds * 2. We limit memory per socket 29631da177e4SLinus Torvalds * 29641da177e4SLinus Torvalds * RFC 1122: 29651da177e4SLinus Torvalds * "the suggested [SWS] avoidance algorithm for the receiver is to keep 29661da177e4SLinus Torvalds * RECV.NEXT + RCV.WIN fixed until: 29671da177e4SLinus Torvalds * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 29681da177e4SLinus Torvalds * 29691da177e4SLinus Torvalds * i.e. don't raise the right edge of the window until you can raise 29701da177e4SLinus Torvalds * it at least MSS bytes. 29711da177e4SLinus Torvalds * 29721da177e4SLinus Torvalds * Unfortunately, the recommended algorithm breaks header prediction, 29731da177e4SLinus Torvalds * since header prediction assumes th->window stays fixed. 29741da177e4SLinus Torvalds * 29751da177e4SLinus Torvalds * Strictly speaking, keeping th->window fixed violates the receiver 29761da177e4SLinus Torvalds * side SWS prevention criteria. The problem is that under this rule 29771da177e4SLinus Torvalds * a stream of single byte packets will cause the right side of the 29781da177e4SLinus Torvalds * window to always advance by a single byte. 29791da177e4SLinus Torvalds * 29801da177e4SLinus Torvalds * Of course, if the sender implements sender side SWS prevention 29811da177e4SLinus Torvalds * then this will not be a problem. 29821da177e4SLinus Torvalds * 29831da177e4SLinus Torvalds * BSD seems to make the following compromise: 29841da177e4SLinus Torvalds * 29851da177e4SLinus Torvalds * If the free space is less than the 1/4 of the maximum 29861da177e4SLinus Torvalds * space available and the free space is less than 1/2 mss, 29871da177e4SLinus Torvalds * then set the window to 0. 29881da177e4SLinus Torvalds * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 29891da177e4SLinus Torvalds * Otherwise, just prevent the window from shrinking 29901da177e4SLinus Torvalds * and from being larger than the largest representable value. 29911da177e4SLinus Torvalds * 29921da177e4SLinus Torvalds * This prevents incremental opening of the window in the regime 29931da177e4SLinus Torvalds * where TCP is limited by the speed of the reader side taking 29941da177e4SLinus Torvalds * data out of the TCP receive queue. It does nothing about 29951da177e4SLinus Torvalds * those cases where the window is constrained on the sender side 29961da177e4SLinus Torvalds * because the pipeline is full. 29971da177e4SLinus Torvalds * 29981da177e4SLinus Torvalds * BSD also seems to "accidentally" limit itself to windows that are a 29991da177e4SLinus Torvalds * multiple of MSS, at least until the free space gets quite small. 30001da177e4SLinus Torvalds * This would appear to be a side effect of the mbuf implementation. 30011da177e4SLinus Torvalds * Combining these two algorithms results in the observed behavior 30021da177e4SLinus Torvalds * of having a fixed window size at almost all times. 30031da177e4SLinus Torvalds * 30041da177e4SLinus Torvalds * Below we obtain similar behavior by forcing the offered window to 30051da177e4SLinus Torvalds * a multiple of the mss when it is feasible to do so. 30061da177e4SLinus Torvalds * 30071da177e4SLinus Torvalds * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 30081da177e4SLinus Torvalds * Regular options like TIMESTAMP are taken into account. 30091da177e4SLinus Torvalds */ 30101da177e4SLinus Torvalds u32 __tcp_select_window(struct sock *sk) 30111da177e4SLinus Torvalds { 3012463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 30131da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 3014b650d953Smfreemon@cloudflare.com struct net *net = sock_net(sk); 3015caa20d9aSStephen Hemminger /* MSS for the peer's data. Previous versions used mss_clamp 30161da177e4SLinus Torvalds * here. I don't know if the value based on our guesses 30171da177e4SLinus Torvalds * of peer's MSS is better for the performance. It's more correct 30181da177e4SLinus Torvalds * but may be worse for the performance because of rcv_mss 30191da177e4SLinus Torvalds * fluctuations. --SAW 1998/11/1 30201da177e4SLinus Torvalds */ 3021463c84b9SArnaldo Carvalho de Melo int mss = icsk->icsk_ack.rcv_mss; 30221da177e4SLinus Torvalds int free_space = tcp_space(sk); 302386c1a045SFlorian Westphal int allowed_space = tcp_full_space(sk); 3024071c8ed6SFlorian Westphal int full_space, window; 3025071c8ed6SFlorian Westphal 3026071c8ed6SFlorian Westphal if (sk_is_mptcp(sk)) 3027071c8ed6SFlorian Westphal mptcp_space(sk, &free_space, &allowed_space); 3028071c8ed6SFlorian Westphal 3029071c8ed6SFlorian Westphal full_space = min_t(int, tp->window_clamp, allowed_space); 30301da177e4SLinus Torvalds 303106425c30SEric Dumazet if (unlikely(mss > full_space)) { 30321da177e4SLinus Torvalds mss = full_space; 303306425c30SEric Dumazet if (mss <= 0) 303406425c30SEric Dumazet return 0; 303506425c30SEric Dumazet } 3036b650d953Smfreemon@cloudflare.com 3037b650d953Smfreemon@cloudflare.com /* Only allow window shrink if the sysctl is enabled and we have 3038b650d953Smfreemon@cloudflare.com * a non-zero scaling factor in effect. 3039b650d953Smfreemon@cloudflare.com */ 3040b650d953Smfreemon@cloudflare.com if (READ_ONCE(net->ipv4.sysctl_tcp_shrink_window) && tp->rx_opt.rcv_wscale) 3041b650d953Smfreemon@cloudflare.com goto shrink_window_allowed; 3042b650d953Smfreemon@cloudflare.com 3043b650d953Smfreemon@cloudflare.com /* do not allow window to shrink */ 3044b650d953Smfreemon@cloudflare.com 3045b92edbe0SEric Dumazet if (free_space < (full_space >> 1)) { 3046463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.quick = 0; 30471da177e4SLinus Torvalds 3048b8da51ebSEric Dumazet if (tcp_under_memory_pressure(sk)) 3049053f3684SWei Wang tcp_adjust_rcv_ssthresh(sk); 30501da177e4SLinus Torvalds 305186c1a045SFlorian Westphal /* free_space might become our new window, make sure we don't 305286c1a045SFlorian Westphal * increase it due to wscale. 305386c1a045SFlorian Westphal */ 305486c1a045SFlorian Westphal free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale); 305586c1a045SFlorian Westphal 305686c1a045SFlorian Westphal /* if free space is less than mss estimate, or is below 1/16th 305786c1a045SFlorian Westphal * of the maximum allowed, try to move to zero-window, else 305886c1a045SFlorian Westphal * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and 305986c1a045SFlorian Westphal * new incoming data is dropped due to memory limits. 306086c1a045SFlorian Westphal * With large window, mss test triggers way too late in order 306186c1a045SFlorian Westphal * to announce zero window in time before rmem limit kicks in. 306286c1a045SFlorian Westphal */ 306386c1a045SFlorian Westphal if (free_space < (allowed_space >> 4) || free_space < mss) 30641da177e4SLinus Torvalds return 0; 30651da177e4SLinus Torvalds } 30661da177e4SLinus Torvalds 30671da177e4SLinus Torvalds if (free_space > tp->rcv_ssthresh) 30681da177e4SLinus Torvalds free_space = tp->rcv_ssthresh; 30691da177e4SLinus Torvalds 30701da177e4SLinus Torvalds /* Don't do rounding if we are using window scaling, since the 30711da177e4SLinus Torvalds * scaled window will not line up with the MSS boundary anyway. 30721da177e4SLinus Torvalds */ 30731da177e4SLinus Torvalds if (tp->rx_opt.rcv_wscale) { 30741da177e4SLinus Torvalds window = free_space; 30751da177e4SLinus Torvalds 30761da177e4SLinus Torvalds /* Advertise enough space so that it won't get scaled away. 30771da177e4SLinus Torvalds * Import case: prevent zero window announcement if 30781da177e4SLinus Torvalds * 1<<rcv_wscale > mss. 30791da177e4SLinus Torvalds */ 30801935299dSGao Feng window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale)); 30811da177e4SLinus Torvalds } else { 30821935299dSGao Feng window = tp->rcv_wnd; 30831da177e4SLinus Torvalds /* Get the largest window that is a nice multiple of mss. 30841da177e4SLinus Torvalds * Window clamp already applied above. 30851da177e4SLinus Torvalds * If our current window offering is within 1 mss of the 30861da177e4SLinus Torvalds * free space we just keep it. This prevents the divide 30871da177e4SLinus Torvalds * and multiply from happening most of the time. 30881da177e4SLinus Torvalds * We also don't do any window rounding when the free space 30891da177e4SLinus Torvalds * is too small. 30901da177e4SLinus Torvalds */ 30911da177e4SLinus Torvalds if (window <= free_space - mss || window > free_space) 30921935299dSGao Feng window = rounddown(free_space, mss); 309384565070SJohn Heffner else if (mss == full_space && 3094b92edbe0SEric Dumazet free_space > window + (full_space >> 1)) 309584565070SJohn Heffner window = free_space; 30961da177e4SLinus Torvalds } 30971da177e4SLinus Torvalds 30981da177e4SLinus Torvalds return window; 3099b650d953Smfreemon@cloudflare.com 3100b650d953Smfreemon@cloudflare.com shrink_window_allowed: 3101b650d953Smfreemon@cloudflare.com /* new window should always be an exact multiple of scaling factor */ 3102b650d953Smfreemon@cloudflare.com free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale); 3103b650d953Smfreemon@cloudflare.com 3104b650d953Smfreemon@cloudflare.com if (free_space < (full_space >> 1)) { 3105b650d953Smfreemon@cloudflare.com icsk->icsk_ack.quick = 0; 3106b650d953Smfreemon@cloudflare.com 3107b650d953Smfreemon@cloudflare.com if (tcp_under_memory_pressure(sk)) 3108b650d953Smfreemon@cloudflare.com tcp_adjust_rcv_ssthresh(sk); 3109b650d953Smfreemon@cloudflare.com 3110b650d953Smfreemon@cloudflare.com /* if free space is too low, return a zero window */ 3111b650d953Smfreemon@cloudflare.com if (free_space < (allowed_space >> 4) || free_space < mss || 3112b650d953Smfreemon@cloudflare.com free_space < (1 << tp->rx_opt.rcv_wscale)) 3113b650d953Smfreemon@cloudflare.com return 0; 3114b650d953Smfreemon@cloudflare.com } 3115b650d953Smfreemon@cloudflare.com 3116b650d953Smfreemon@cloudflare.com if (free_space > tp->rcv_ssthresh) { 3117b650d953Smfreemon@cloudflare.com free_space = tp->rcv_ssthresh; 3118b650d953Smfreemon@cloudflare.com /* new window should always be an exact multiple of scaling factor 3119b650d953Smfreemon@cloudflare.com * 3120b650d953Smfreemon@cloudflare.com * For this case, we ALIGN "up" (increase free_space) because 3121b650d953Smfreemon@cloudflare.com * we know free_space is not zero here, it has been reduced from 3122b650d953Smfreemon@cloudflare.com * the memory-based limit, and rcv_ssthresh is not a hard limit 3123b650d953Smfreemon@cloudflare.com * (unlike sk_rcvbuf). 3124b650d953Smfreemon@cloudflare.com */ 3125b650d953Smfreemon@cloudflare.com free_space = ALIGN(free_space, (1 << tp->rx_opt.rcv_wscale)); 3126b650d953Smfreemon@cloudflare.com } 3127b650d953Smfreemon@cloudflare.com 3128b650d953Smfreemon@cloudflare.com return free_space; 31291da177e4SLinus Torvalds } 31301da177e4SLinus Torvalds 3131cfea5a68SMartin KaFai Lau void tcp_skb_collapse_tstamp(struct sk_buff *skb, 3132082ac2d5SMartin KaFai Lau const struct sk_buff *next_skb) 3133082ac2d5SMartin KaFai Lau { 31340a2cf20cSSoheil Hassas Yeganeh if (unlikely(tcp_has_tx_tstamp(next_skb))) { 31350a2cf20cSSoheil Hassas Yeganeh const struct skb_shared_info *next_shinfo = 31360a2cf20cSSoheil Hassas Yeganeh skb_shinfo(next_skb); 3137082ac2d5SMartin KaFai Lau struct skb_shared_info *shinfo = skb_shinfo(skb); 3138082ac2d5SMartin KaFai Lau 31390a2cf20cSSoheil Hassas Yeganeh shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP; 3140082ac2d5SMartin KaFai Lau shinfo->tskey = next_shinfo->tskey; 31412de8023eSMartin KaFai Lau TCP_SKB_CB(skb)->txstamp_ack |= 31422de8023eSMartin KaFai Lau TCP_SKB_CB(next_skb)->txstamp_ack; 3143082ac2d5SMartin KaFai Lau } 3144082ac2d5SMartin KaFai Lau } 3145082ac2d5SMartin KaFai Lau 31464a17fc3aSIlpo Järvinen /* Collapses two adjacent SKB's during retransmission. */ 3147f8071cdeSEric Dumazet static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) 31481da177e4SLinus Torvalds { 31491da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 315075c119afSEric Dumazet struct sk_buff *next_skb = skb_rb_next(skb); 315113dde04fSWei Yongjun int next_skb_size; 31521da177e4SLinus Torvalds 3153058dc334SIlpo Järvinen next_skb_size = next_skb->len; 31541da177e4SLinus Torvalds 3155058dc334SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); 31561da177e4SLinus Torvalds 3157bd446314SEric Dumazet if (next_skb_size && !tcp_skb_shift(skb, next_skb, 1, next_skb_size)) 3158f8071cdeSEric Dumazet return false; 3159bd446314SEric Dumazet 31602b7cda9cSEric Dumazet tcp_highest_sack_replace(sk, next_skb, skb); 3161a6963a6bSIlpo Järvinen 31621da177e4SLinus Torvalds /* Update sequence range on original skb. */ 31631da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 31641da177e4SLinus Torvalds 3165e6c7d085SIlpo Järvinen /* Merge over control information. This moves PSH/FIN etc. over */ 31664de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; 31671da177e4SLinus Torvalds 31681da177e4SLinus Torvalds /* All done, get rid of second SKB and account for it so 31691da177e4SLinus Torvalds * packet counting does not break. 31701da177e4SLinus Torvalds */ 31714828e7f4SIlpo Järvinen TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; 3172a643b5d4SMartin KaFai Lau TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor; 3173b7689205SIlpo Järvinen 3174b7689205SIlpo Järvinen /* changed transmit queue under us so clear hints */ 3175ef9da47cSIlpo Järvinen tcp_clear_retrans_hints_partial(tp); 3176ef9da47cSIlpo Järvinen if (next_skb == tp->retransmit_skb_hint) 3177ef9da47cSIlpo Järvinen tp->retransmit_skb_hint = skb; 3178b7689205SIlpo Järvinen 3179797108d1SIlpo Järvinen tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); 3180797108d1SIlpo Järvinen 3181082ac2d5SMartin KaFai Lau tcp_skb_collapse_tstamp(skb, next_skb); 3182082ac2d5SMartin KaFai Lau 318375c119afSEric Dumazet tcp_rtx_queue_unlink_and_free(next_skb, sk); 3184f8071cdeSEric Dumazet return true; 31851da177e4SLinus Torvalds } 31861da177e4SLinus Torvalds 318767edfef7SAndi Kleen /* Check if coalescing SKBs is legal. */ 3188a2a385d6SEric Dumazet static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) 31894a17fc3aSIlpo Järvinen { 31904a17fc3aSIlpo Järvinen if (tcp_skb_pcount(skb) > 1) 3191a2a385d6SEric Dumazet return false; 31924a17fc3aSIlpo Järvinen if (skb_cloned(skb)) 3193a2a385d6SEric Dumazet return false; 31942331ccc5SEric Dumazet /* Some heuristics for collapsing over SACK'd could be invented */ 31954a17fc3aSIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 3196a2a385d6SEric Dumazet return false; 31974a17fc3aSIlpo Järvinen 3198a2a385d6SEric Dumazet return true; 31994a17fc3aSIlpo Järvinen } 32004a17fc3aSIlpo Järvinen 320167edfef7SAndi Kleen /* Collapse packets in the retransmit queue to make to create 320267edfef7SAndi Kleen * less packets on the wire. This is only done on retransmission. 320367edfef7SAndi Kleen */ 32044a17fc3aSIlpo Järvinen static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, 32054a17fc3aSIlpo Järvinen int space) 32064a17fc3aSIlpo Järvinen { 32074a17fc3aSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 32084a17fc3aSIlpo Järvinen struct sk_buff *skb = to, *tmp; 3209a2a385d6SEric Dumazet bool first = true; 32104a17fc3aSIlpo Järvinen 32111a63cb91SKuniyuki Iwashima if (!READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse)) 32124a17fc3aSIlpo Järvinen return; 32134de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 32144a17fc3aSIlpo Järvinen return; 32154a17fc3aSIlpo Järvinen 321675c119afSEric Dumazet skb_rbtree_walk_from_safe(skb, tmp) { 32174a17fc3aSIlpo Järvinen if (!tcp_can_collapse(sk, skb)) 32184a17fc3aSIlpo Järvinen break; 32194a17fc3aSIlpo Järvinen 322085712484SMat Martineau if (!tcp_skb_can_collapse(to, skb)) 3221a643b5d4SMartin KaFai Lau break; 3222a643b5d4SMartin KaFai Lau 32234a17fc3aSIlpo Järvinen space -= skb->len; 32244a17fc3aSIlpo Järvinen 32254a17fc3aSIlpo Järvinen if (first) { 3226a2a385d6SEric Dumazet first = false; 32274a17fc3aSIlpo Järvinen continue; 32284a17fc3aSIlpo Järvinen } 32294a17fc3aSIlpo Järvinen 32304a17fc3aSIlpo Järvinen if (space < 0) 32314a17fc3aSIlpo Järvinen break; 32324a17fc3aSIlpo Järvinen 32334a17fc3aSIlpo Järvinen if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) 32344a17fc3aSIlpo Järvinen break; 32354a17fc3aSIlpo Järvinen 3236f8071cdeSEric Dumazet if (!tcp_collapse_retrans(sk, to)) 3237f8071cdeSEric Dumazet break; 32384a17fc3aSIlpo Järvinen } 32394a17fc3aSIlpo Järvinen } 32404a17fc3aSIlpo Järvinen 32411da177e4SLinus Torvalds /* This retransmits one SKB. Policy decisions and retransmit queue 32421da177e4SLinus Torvalds * state updates are done by the caller. Returns non-zero if an 32431da177e4SLinus Torvalds * error occurred which prevented the send. 32441da177e4SLinus Torvalds */ 324510d3be56SEric Dumazet int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) 32461da177e4SLinus Torvalds { 32475d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 324810d3be56SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 32497d227cd2SSridhar Samudrala unsigned int cur_mss; 325010d3be56SEric Dumazet int diff, len, err; 3251536a6c8eSYonglong Li int avail_wnd; 325210d3be56SEric Dumazet 325310d3be56SEric Dumazet /* Inconclusive MTU probe */ 325410d3be56SEric Dumazet if (icsk->icsk_mtup.probe_size) 32555d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 32565d424d5aSJohn Heffner 32571f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 32581f3279aeSEric Dumazet return -EBUSY; 32591f3279aeSEric Dumazet 32601da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 32617f582b24SEric Dumazet if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) { 32627f582b24SEric Dumazet WARN_ON_ONCE(1); 32637f582b24SEric Dumazet return -EINVAL; 32647f582b24SEric Dumazet } 32651da177e4SLinus Torvalds if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 32661da177e4SLinus Torvalds return -ENOMEM; 32671da177e4SLinus Torvalds } 32681da177e4SLinus Torvalds 32697d227cd2SSridhar Samudrala if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 32707d227cd2SSridhar Samudrala return -EHOSTUNREACH; /* Routing failure or similar. */ 32717d227cd2SSridhar Samudrala 32720c54b85fSIlpo Järvinen cur_mss = tcp_current_mss(sk); 3273536a6c8eSYonglong Li avail_wnd = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 32747d227cd2SSridhar Samudrala 32751da177e4SLinus Torvalds /* If receiver has shrunk his window, and skb is out of 32761da177e4SLinus Torvalds * new window, do not retransmit it. The exception is the 32771da177e4SLinus Torvalds * case, when window is shrunk to zero. In this case 3278536a6c8eSYonglong Li * our retransmit of one segment serves as a zero window probe. 32791da177e4SLinus Torvalds */ 3280536a6c8eSYonglong Li if (avail_wnd <= 0) { 3281536a6c8eSYonglong Li if (TCP_SKB_CB(skb)->seq != tp->snd_una) 32821da177e4SLinus Torvalds return -EAGAIN; 3283536a6c8eSYonglong Li avail_wnd = cur_mss; 3284536a6c8eSYonglong Li } 32851da177e4SLinus Torvalds 328610d3be56SEric Dumazet len = cur_mss * segs; 3287536a6c8eSYonglong Li if (len > avail_wnd) { 3288536a6c8eSYonglong Li len = rounddown(avail_wnd, cur_mss); 3289536a6c8eSYonglong Li if (!len) 3290536a6c8eSYonglong Li len = avail_wnd; 3291536a6c8eSYonglong Li } 329210d3be56SEric Dumazet if (skb->len > len) { 329375c119afSEric Dumazet if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len, 329475c119afSEric Dumazet cur_mss, GFP_ATOMIC)) 32951da177e4SLinus Torvalds return -ENOMEM; /* We'll try again later. */ 329602276f3cSIlpo Järvinen } else { 3297c4777efaSEric Dumazet if (skb_unclone_keeptruesize(skb, GFP_ATOMIC)) 3298c52e2421SEric Dumazet return -ENOMEM; 329910d3be56SEric Dumazet 330010d3be56SEric Dumazet diff = tcp_skb_pcount(skb); 330110d3be56SEric Dumazet tcp_set_skb_tso_segs(skb, cur_mss); 330210d3be56SEric Dumazet diff -= tcp_skb_pcount(skb); 330310d3be56SEric Dumazet if (diff) 330410d3be56SEric Dumazet tcp_adjust_pcount(sk, skb, diff); 3305536a6c8eSYonglong Li avail_wnd = min_t(int, avail_wnd, cur_mss); 3306536a6c8eSYonglong Li if (skb->len < avail_wnd) 3307536a6c8eSYonglong Li tcp_retrans_try_collapse(sk, skb, avail_wnd); 33081da177e4SLinus Torvalds } 33091da177e4SLinus Torvalds 331049213555SDaniel Borkmann /* RFC3168, section 6.1.1.1. ECN fallback */ 331149213555SDaniel Borkmann if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN) 331249213555SDaniel Borkmann tcp_ecn_clear_syn(sk, skb); 331349213555SDaniel Borkmann 3314678550c6SYuchung Cheng /* Update global and local TCP statistics. */ 3315678550c6SYuchung Cheng segs = tcp_skb_pcount(skb); 3316678550c6SYuchung Cheng TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs); 3317678550c6SYuchung Cheng if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 3318678550c6SYuchung Cheng __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 3319678550c6SYuchung Cheng tp->total_retrans += segs; 3320fb31c9b9SWei Wang tp->bytes_retrans += skb->len; 3321678550c6SYuchung Cheng 332250bceae9SThomas Graf /* make sure skb->data is aligned on arches that require it 332350bceae9SThomas Graf * and check if ack-trimming & collapsing extended the headroom 332450bceae9SThomas Graf * beyond what csum_start can cover. 332550bceae9SThomas Graf */ 332650bceae9SThomas Graf if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) || 332750bceae9SThomas Graf skb_headroom(skb) >= 0xFFFF)) { 332810a81980SEric Dumazet struct sk_buff *nskb; 332910a81980SEric Dumazet 3330e2080072SEric Dumazet tcp_skb_tsorted_save(skb) { 333110a81980SEric Dumazet nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC); 333207f8e4d0SFlorian Westphal if (nskb) { 333307f8e4d0SFlorian Westphal nskb->dev = NULL; 333407f8e4d0SFlorian Westphal err = tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC); 333507f8e4d0SFlorian Westphal } else { 333607f8e4d0SFlorian Westphal err = -ENOBUFS; 333707f8e4d0SFlorian Westphal } 3338e2080072SEric Dumazet } tcp_skb_tsorted_restore(skb); 3339e2080072SEric Dumazet 33405889e2c0SYousuk Seung if (!err) { 3341a7a25630SEric Dumazet tcp_update_skb_after_send(sk, skb, tp->tcp_wstamp_ns); 33425889e2c0SYousuk Seung tcp_rate_skb_sent(sk, skb); 33435889e2c0SYousuk Seung } 3344117632e6SEric Dumazet } else { 3345c84a5711SYuchung Cheng err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 3346117632e6SEric Dumazet } 3347c84a5711SYuchung Cheng 33487f12422cSYuchung Cheng /* To avoid taking spuriously low RTT samples based on a timestamp 33497f12422cSYuchung Cheng * for a transmit that never happened, always mark EVER_RETRANS 33507f12422cSYuchung Cheng */ 33517f12422cSYuchung Cheng TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS; 33527f12422cSYuchung Cheng 3353a31ad29eSLawrence Brakmo if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RETRANS_CB_FLAG)) 3354a31ad29eSLawrence Brakmo tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RETRANS_CB, 3355a31ad29eSLawrence Brakmo TCP_SKB_CB(skb)->seq, segs, err); 3356a31ad29eSLawrence Brakmo 3357fc9f3501SEric Dumazet if (likely(!err)) { 3358e086101bSCong Wang trace_tcp_retransmit_skb(sk, skb); 3359678550c6SYuchung Cheng } else if (err != -EBUSY) { 3360ec641b39SYuchung Cheng NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL, segs); 3361fc9f3501SEric Dumazet } 3362c84a5711SYuchung Cheng return err; 336393b174adSYuchung Cheng } 336493b174adSYuchung Cheng 336510d3be56SEric Dumazet int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) 336693b174adSYuchung Cheng { 336793b174adSYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 336810d3be56SEric Dumazet int err = __tcp_retransmit_skb(sk, skb, segs); 33691da177e4SLinus Torvalds 33701da177e4SLinus Torvalds if (err == 0) { 33711da177e4SLinus Torvalds #if FASTRETRANS_DEBUG > 0 33721da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 3373e87cc472SJoe Perches net_dbg_ratelimited("retrans_out leaked\n"); 33741da177e4SLinus Torvalds } 33751da177e4SLinus Torvalds #endif 33761da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 33771da177e4SLinus Torvalds tp->retrans_out += tcp_skb_pcount(skb); 33787ae18975SYuchung Cheng } 33791da177e4SLinus Torvalds 33807ae18975SYuchung Cheng /* Save stamp of the first (attempted) retransmit. */ 33811da177e4SLinus Torvalds if (!tp->retrans_stamp) 3382d1a02ed6SEric Dumazet tp->retrans_stamp = tcp_skb_timestamp_ts(false, skb); 33831da177e4SLinus Torvalds 33846e08d5e3SYuchung Cheng if (tp->undo_retrans < 0) 33856e08d5e3SYuchung Cheng tp->undo_retrans = 0; 33866e08d5e3SYuchung Cheng tp->undo_retrans += tcp_skb_pcount(skb); 33871da177e4SLinus Torvalds return err; 33881da177e4SLinus Torvalds } 33891da177e4SLinus Torvalds 33901da177e4SLinus Torvalds /* This gets called after a retransmit timeout, and the initially 33911da177e4SLinus Torvalds * retransmitted data is acknowledged. It tries to continue 33921da177e4SLinus Torvalds * resending the rest of the retransmit queue, until either 33931da177e4SLinus Torvalds * we've sent it all or the congestion window limit is reached. 33941da177e4SLinus Torvalds */ 33951da177e4SLinus Torvalds void tcp_xmit_retransmit_queue(struct sock *sk) 33961da177e4SLinus Torvalds { 33976687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 3398b9f1f1ceSEric Dumazet struct sk_buff *skb, *rtx_head, *hole = NULL; 33991da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 3400916e6d1aSEric Dumazet bool rearm_timer = false; 3401840a3cbeSYuchung Cheng u32 max_segs; 340261eb55f4SIlpo Järvinen int mib_idx; 34036a438bbeSStephen Hemminger 340445e77d31SIlpo Järvinen if (!tp->packets_out) 340545e77d31SIlpo Järvinen return; 340645e77d31SIlpo Järvinen 340775c119afSEric Dumazet rtx_head = tcp_rtx_queue_head(sk); 3408b9f1f1ceSEric Dumazet skb = tp->retransmit_skb_hint ?: rtx_head; 3409ed6e7268SNeal Cardwell max_segs = tcp_tso_segs(sk, tcp_current_mss(sk)); 341075c119afSEric Dumazet skb_rbtree_walk_from(skb) { 3411dca0aaf8SEric Dumazet __u8 sacked; 341210d3be56SEric Dumazet int segs; 34131da177e4SLinus Torvalds 3414218af599SEric Dumazet if (tcp_pacing_check(sk)) 3415218af599SEric Dumazet break; 3416218af599SEric Dumazet 34176a438bbeSStephen Hemminger /* we could do better than to assign each time */ 341851456b29SIan Morris if (!hole) 34196a438bbeSStephen Hemminger tp->retransmit_skb_hint = skb; 34206a438bbeSStephen Hemminger 342140570375SEric Dumazet segs = tcp_snd_cwnd(tp) - tcp_packets_in_flight(tp); 342210d3be56SEric Dumazet if (segs <= 0) 3423916e6d1aSEric Dumazet break; 3424dca0aaf8SEric Dumazet sacked = TCP_SKB_CB(skb)->sacked; 3425a3d2e9f8SEric Dumazet /* In case tcp_shift_skb_data() have aggregated large skbs, 3426a3d2e9f8SEric Dumazet * we need to make sure not sending too bigs TSO packets 3427a3d2e9f8SEric Dumazet */ 3428a3d2e9f8SEric Dumazet segs = min_t(int, segs, max_segs); 34290e1c54c2SIlpo Järvinen 3430840a3cbeSYuchung Cheng if (tp->retrans_out >= tp->lost_out) { 3431006f582cSIlpo Järvinen break; 34320e1c54c2SIlpo Järvinen } else if (!(sacked & TCPCB_LOST)) { 343351456b29SIan Morris if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) 34340e1c54c2SIlpo Järvinen hole = skb; 343561eb55f4SIlpo Järvinen continue; 34361da177e4SLinus Torvalds 34370e1c54c2SIlpo Järvinen } else { 34380e1c54c2SIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Loss) 34390e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFASTRETRANS; 34400e1c54c2SIlpo Järvinen else 34410e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; 34420e1c54c2SIlpo Järvinen } 34430e1c54c2SIlpo Järvinen 34440e1c54c2SIlpo Järvinen if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) 344561eb55f4SIlpo Järvinen continue; 344640b215e5SPavel Emelyanov 3447f9616c35SEric Dumazet if (tcp_small_queue_check(sk, skb, 1)) 3448916e6d1aSEric Dumazet break; 3449f9616c35SEric Dumazet 345010d3be56SEric Dumazet if (tcp_retransmit_skb(sk, skb, segs)) 3451916e6d1aSEric Dumazet break; 345224ab6becSYuchung Cheng 3453de1d6578SYuchung Cheng NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb)); 34541da177e4SLinus Torvalds 3455684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 3456a262f0cdSNandita Dukkipati tp->prr_out += tcp_skb_pcount(skb); 3457a262f0cdSNandita Dukkipati 345875c119afSEric Dumazet if (skb == rtx_head && 345957dde7f7SYuchung Cheng icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT) 3460916e6d1aSEric Dumazet rearm_timer = true; 3461916e6d1aSEric Dumazet 3462916e6d1aSEric Dumazet } 3463916e6d1aSEric Dumazet if (rearm_timer) 34643f80e08fSEric Dumazet tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 34653f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, 34668dc242adSEric Dumazet TCP_RTO_MAX); 34671da177e4SLinus Torvalds } 34681da177e4SLinus Torvalds 3469d83769a5SEric Dumazet /* We allow to exceed memory limits for FIN packets to expedite 3470d83769a5SEric Dumazet * connection tear down and (memory) recovery. 3471845704a5SEric Dumazet * Otherwise tcp_send_fin() could be tempted to either delay FIN 3472845704a5SEric Dumazet * or even be forced to close flow without any FIN. 3473a6c5ea4cSEric Dumazet * In general, we want to allow one skb per socket to avoid hangs 3474a6c5ea4cSEric Dumazet * with edge trigger epoll() 3475d83769a5SEric Dumazet */ 3476a6c5ea4cSEric Dumazet void sk_forced_mem_schedule(struct sock *sk, int size) 3477d83769a5SEric Dumazet { 3478c4ee1185SEric Dumazet int delta, amt; 3479d83769a5SEric Dumazet 3480c4ee1185SEric Dumazet delta = size - sk->sk_forward_alloc; 3481c4ee1185SEric Dumazet if (delta <= 0) 3482d83769a5SEric Dumazet return; 3483c4ee1185SEric Dumazet amt = sk_mem_pages(delta); 34845e6300e7SEric Dumazet sk_forward_alloc_add(sk, amt << PAGE_SHIFT); 3485e805605cSJohannes Weiner sk_memory_allocated_add(sk, amt); 3486e805605cSJohannes Weiner 3487baac50bbSJohannes Weiner if (mem_cgroup_sockets_enabled && sk->sk_memcg) 34884b1327beSWei Wang mem_cgroup_charge_skmem(sk->sk_memcg, amt, 34894b1327beSWei Wang gfp_memcg_charge() | __GFP_NOFAIL); 3490d83769a5SEric Dumazet } 3491d83769a5SEric Dumazet 3492845704a5SEric Dumazet /* Send a FIN. The caller locks the socket for us. 3493845704a5SEric Dumazet * We should try to send a FIN packet really hard, but eventually give up. 34941da177e4SLinus Torvalds */ 34951da177e4SLinus Torvalds void tcp_send_fin(struct sock *sk) 34961da177e4SLinus Torvalds { 3497ee2aabd3SEric Dumazet struct sk_buff *skb, *tskb, *tail = tcp_write_queue_tail(sk); 34981da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 34991da177e4SLinus Torvalds 3500845704a5SEric Dumazet /* Optimization, tack on the FIN if we have one skb in write queue and 3501845704a5SEric Dumazet * this skb was not yet sent, or we are under memory pressure. 3502845704a5SEric Dumazet * Note: in the latter case, FIN packet will be sent after a timeout, 3503845704a5SEric Dumazet * as TCP stack thinks it has already been transmitted. 35041da177e4SLinus Torvalds */ 3505ee2aabd3SEric Dumazet tskb = tail; 350675c119afSEric Dumazet if (!tskb && tcp_under_memory_pressure(sk)) 350775c119afSEric Dumazet tskb = skb_rb_last(&sk->tcp_rtx_queue); 350875c119afSEric Dumazet 350975c119afSEric Dumazet if (tskb) { 3510845704a5SEric Dumazet TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN; 3511845704a5SEric Dumazet TCP_SKB_CB(tskb)->end_seq++; 35121da177e4SLinus Torvalds tp->write_seq++; 3513ee2aabd3SEric Dumazet if (!tail) { 3514845704a5SEric Dumazet /* This means tskb was already sent. 3515845704a5SEric Dumazet * Pretend we included the FIN on previous transmit. 3516845704a5SEric Dumazet * We need to set tp->snd_nxt to the value it would have 3517845704a5SEric Dumazet * if FIN had been sent. This is because retransmit path 3518845704a5SEric Dumazet * does not change tp->snd_nxt. 3519845704a5SEric Dumazet */ 3520e0d694d6SEric Dumazet WRITE_ONCE(tp->snd_nxt, tp->snd_nxt + 1); 3521845704a5SEric Dumazet return; 3522845704a5SEric Dumazet } 35231da177e4SLinus Torvalds } else { 3524845704a5SEric Dumazet skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation); 3525d1edc085SColin Ian King if (unlikely(!skb)) 3526845704a5SEric Dumazet return; 3527d1edc085SColin Ian King 3528e2080072SEric Dumazet INIT_LIST_HEAD(&skb->tcp_tsorted_anchor); 3529d83769a5SEric Dumazet skb_reserve(skb, MAX_TCP_HEADER); 3530a6c5ea4cSEric Dumazet sk_forced_mem_schedule(sk, skb->truesize); 35311da177e4SLinus Torvalds /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 3532e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tp->write_seq, 3533a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_FIN); 35341da177e4SLinus Torvalds tcp_queue_skb(sk, skb); 35351da177e4SLinus Torvalds } 3536845704a5SEric Dumazet __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF); 35371da177e4SLinus Torvalds } 35381da177e4SLinus Torvalds 35391da177e4SLinus Torvalds /* We get here when a process closes a file descriptor (either due to 35401da177e4SLinus Torvalds * an explicit close() or as a byproduct of exit()'ing) and there 35411da177e4SLinus Torvalds * was unread data in the receive queue. This behavior is recommended 354265bb723cSGerrit Renker * by RFC 2525, section 2.17. -DaveM 35431da177e4SLinus Torvalds */ 3544dd0fc66fSAl Viro void tcp_send_active_reset(struct sock *sk, gfp_t priority) 35451da177e4SLinus Torvalds { 35461da177e4SLinus Torvalds struct sk_buff *skb; 35471da177e4SLinus Torvalds 35487cc2b043SGao Feng TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); 35497cc2b043SGao Feng 35501da177e4SLinus Torvalds /* NOTE: No TCP options attached and we never retransmit this. */ 35511da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, priority); 35521da177e4SLinus Torvalds if (!skb) { 35534e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 35541da177e4SLinus Torvalds return; 35551da177e4SLinus Torvalds } 35561da177e4SLinus Torvalds 35571da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 35581da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 3559e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), 3560a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_RST); 35619a568de4SEric Dumazet tcp_mstamp_refresh(tcp_sk(sk)); 35621da177e4SLinus Torvalds /* Send it off. */ 3563dfb4b9dcSDavid S. Miller if (tcp_transmit_skb(sk, skb, 0, priority)) 35644e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 3565c24b14c4SSong Liu 3566c24b14c4SSong Liu /* skb of trace_tcp_send_reset() keeps the skb that caused RST, 3567c24b14c4SSong Liu * skb here is different to the troublesome skb, so use NULL 3568c24b14c4SSong Liu */ 3569c24b14c4SSong Liu trace_tcp_send_reset(sk, NULL); 35701da177e4SLinus Torvalds } 35711da177e4SLinus Torvalds 357267edfef7SAndi Kleen /* Send a crossed SYN-ACK during socket establishment. 357367edfef7SAndi Kleen * WARNING: This routine must only be called when we have already sent 35741da177e4SLinus Torvalds * a SYN packet that crossed the incoming SYN that caused this routine 35751da177e4SLinus Torvalds * to get called. If this assumption fails then the initial rcv_wnd 35761da177e4SLinus Torvalds * and rcv_wscale values will not be correct. 35771da177e4SLinus Torvalds */ 35781da177e4SLinus Torvalds int tcp_send_synack(struct sock *sk) 35791da177e4SLinus Torvalds { 35801da177e4SLinus Torvalds struct sk_buff *skb; 35811da177e4SLinus Torvalds 358275c119afSEric Dumazet skb = tcp_rtx_queue_head(sk); 358351456b29SIan Morris if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 358475c119afSEric Dumazet pr_err("%s: wrong queue state\n", __func__); 35851da177e4SLinus Torvalds return -EFAULT; 35861da177e4SLinus Torvalds } 35874de075e0SEric Dumazet if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { 35881da177e4SLinus Torvalds if (skb_cloned(skb)) { 3589e2080072SEric Dumazet struct sk_buff *nskb; 3590e2080072SEric Dumazet 3591e2080072SEric Dumazet tcp_skb_tsorted_save(skb) { 3592e2080072SEric Dumazet nskb = skb_copy(skb, GFP_ATOMIC); 3593e2080072SEric Dumazet } tcp_skb_tsorted_restore(skb); 359451456b29SIan Morris if (!nskb) 35951da177e4SLinus Torvalds return -ENOMEM; 3596e2080072SEric Dumazet INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor); 35972bec445fSEric Dumazet tcp_highest_sack_replace(sk, skb, nskb); 359875c119afSEric Dumazet tcp_rtx_queue_unlink_and_free(skb, sk); 3599f4a775d1SEric Dumazet __skb_header_release(nskb); 360075c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb); 3601ab4e846aSEric Dumazet sk_wmem_queued_add(sk, nskb->truesize); 36023ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 36031da177e4SLinus Torvalds skb = nskb; 36041da177e4SLinus Torvalds } 36051da177e4SLinus Torvalds 36064de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; 3607735d3831SFlorian Westphal tcp_ecn_send_synack(sk, skb); 36081da177e4SLinus Torvalds } 3609dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 36101da177e4SLinus Torvalds } 36111da177e4SLinus Torvalds 36124aea39c1SEric Dumazet /** 3613331fca43SMartin KaFai Lau * tcp_make_synack - Allocate one skb and build a SYNACK packet. 3614331fca43SMartin KaFai Lau * @sk: listener socket 3615331fca43SMartin KaFai Lau * @dst: dst entry attached to the SYNACK. It is consumed and caller 3616331fca43SMartin KaFai Lau * should not use it again. 3617331fca43SMartin KaFai Lau * @req: request_sock pointer 3618331fca43SMartin KaFai Lau * @foc: cookie for tcp fast open 3619331fca43SMartin KaFai Lau * @synack_type: Type of synack to prepare 3620331fca43SMartin KaFai Lau * @syn_skb: SYN packet just received. It could be NULL for rtx case. 36214aea39c1SEric Dumazet */ 36225d062de7SEric Dumazet struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst, 3623e6b4d113SWilliam Allen Simpson struct request_sock *req, 3624ca6fb065SEric Dumazet struct tcp_fastopen_cookie *foc, 3625331fca43SMartin KaFai Lau enum tcp_synack_type synack_type, 3626331fca43SMartin KaFai Lau struct sk_buff *syn_skb) 36271da177e4SLinus Torvalds { 36282e6599cbSArnaldo Carvalho de Melo struct inet_request_sock *ireq = inet_rsk(req); 36295d062de7SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 363080f03e27SEric Dumazet struct tcp_md5sig_key *md5 = NULL; 36315d062de7SEric Dumazet struct tcp_out_options opts; 36325d062de7SEric Dumazet struct sk_buff *skb; 3633bd0388aeSWilliam Allen Simpson int tcp_header_size; 36345d062de7SEric Dumazet struct tcphdr *th; 3635f5fff5dcSTom Quetchenbach int mss; 3636a842fe14SEric Dumazet u64 now; 36371da177e4SLinus Torvalds 3638ca6fb065SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 36394aea39c1SEric Dumazet if (unlikely(!skb)) { 36404aea39c1SEric Dumazet dst_release(dst); 36411da177e4SLinus Torvalds return NULL; 36424aea39c1SEric Dumazet } 36431da177e4SLinus Torvalds /* Reserve space for headers. */ 36441da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 36451da177e4SLinus Torvalds 3646b3d05147SEric Dumazet switch (synack_type) { 3647b3d05147SEric Dumazet case TCP_SYNACK_NORMAL: 36489e17f8a4SEric Dumazet skb_set_owner_w(skb, req_to_sk(req)); 3649b3d05147SEric Dumazet break; 3650b3d05147SEric Dumazet case TCP_SYNACK_COOKIE: 3651b3d05147SEric Dumazet /* Under synflood, we do not attach skb to a socket, 3652b3d05147SEric Dumazet * to avoid false sharing. 3653b3d05147SEric Dumazet */ 3654b3d05147SEric Dumazet break; 3655b3d05147SEric Dumazet case TCP_SYNACK_FASTOPEN: 3656ca6fb065SEric Dumazet /* sk is a const pointer, because we want to express multiple 3657ca6fb065SEric Dumazet * cpu might call us concurrently. 3658ca6fb065SEric Dumazet * sk->sk_wmem_alloc in an atomic, we can promote to rw. 3659ca6fb065SEric Dumazet */ 3660ca6fb065SEric Dumazet skb_set_owner_w(skb, (struct sock *)sk); 3661b3d05147SEric Dumazet break; 3662ca6fb065SEric Dumazet } 36634aea39c1SEric Dumazet skb_dst_set(skb, dst); 36641da177e4SLinus Torvalds 36653541f9e8SEric Dumazet mss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 3666f5fff5dcSTom Quetchenbach 366733ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 3668a842fe14SEric Dumazet now = tcp_clock_ns(); 36698b5f12d0SFlorian Westphal #ifdef CONFIG_SYN_COOKIES 3670f8ace8d9SFlorian Westphal if (unlikely(synack_type == TCP_SYNACK_COOKIE && ireq->tstamp_ok)) 3671a1ac9c8aSMartin KaFai Lau skb_set_delivery_time(skb, cookie_init_timestamp(req, now), 3672a1ac9c8aSMartin KaFai Lau true); 36738b5f12d0SFlorian Westphal else 36748b5f12d0SFlorian Westphal #endif 36759e450c1eSYuchung Cheng { 3676a1ac9c8aSMartin KaFai Lau skb_set_delivery_time(skb, now, true); 36779e450c1eSYuchung Cheng if (!tcp_rsk(req)->snt_synack) /* Timestamp first SYNACK */ 36789e450c1eSYuchung Cheng tcp_rsk(req)->snt_synack = tcp_skb_timestamp_us(skb); 36799e450c1eSYuchung Cheng } 368080f03e27SEric Dumazet 368180f03e27SEric Dumazet #ifdef CONFIG_TCP_MD5SIG 368280f03e27SEric Dumazet rcu_read_lock(); 3683fd3a154aSEric Dumazet md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req)); 368480f03e27SEric Dumazet #endif 36855e526552SEric Dumazet skb_set_hash(skb, READ_ONCE(tcp_rsk(req)->txhash), PKT_HASH_TYPE_L4); 3686331fca43SMartin KaFai Lau /* bpf program will be interested in the tcp_flags */ 3687331fca43SMartin KaFai Lau TCP_SKB_CB(skb)->tcp_flags = TCPHDR_SYN | TCPHDR_ACK; 368860e2a778SUrsula Braun tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5, 3689331fca43SMartin KaFai Lau foc, synack_type, 3690331fca43SMartin KaFai Lau syn_skb) + sizeof(*th); 369133ad798cSAdam Langley 3692aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 3693aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 36941da177e4SLinus Torvalds 3695ea1627c2SEric Dumazet th = (struct tcphdr *)skb->data; 36961da177e4SLinus Torvalds memset(th, 0, sizeof(struct tcphdr)); 36971da177e4SLinus Torvalds th->syn = 1; 36981da177e4SLinus Torvalds th->ack = 1; 36996ac705b1SEric Dumazet tcp_ecn_make_synack(req, th); 3700b44084c2SEric Dumazet th->source = htons(ireq->ir_num); 3701634fb979SEric Dumazet th->dest = ireq->ir_rmt_port; 3702e05a90ecSJamal Hadi Salim skb->mark = ireq->ir_mark; 37033b117750SEric Dumazet skb->ip_summed = CHECKSUM_PARTIAL; 37043b117750SEric Dumazet th->seq = htonl(tcp_rsk(req)->snt_isn); 37058336886fSJerry Chu /* XXX data is queued and acked as is. No buffer/window check */ 37068336886fSJerry Chu th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt); 37071da177e4SLinus Torvalds 37081da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 3709ed53d0abSEric Dumazet th->window = htons(min(req->rsk_rcv_wnd, 65535U)); 3710ea66758cSPaolo Abeni tcp_options_write(th, NULL, &opts); 37111da177e4SLinus Torvalds th->doff = (tcp_header_size >> 2); 3712bced3f7dSBreno Leitao TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS); 3713cfb6eeb4SYOSHIFUJI Hideaki 3714cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 3715cfb6eeb4SYOSHIFUJI Hideaki /* Okay, we have all we need - do the md5 hash if needed */ 371680f03e27SEric Dumazet if (md5) 3717bd0388aeSWilliam Allen Simpson tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, 371839f8e58eSEric Dumazet md5, req_to_sk(req), skb); 371980f03e27SEric Dumazet rcu_read_unlock(); 3720cfb6eeb4SYOSHIFUJI Hideaki #endif 3721cfb6eeb4SYOSHIFUJI Hideaki 3722331fca43SMartin KaFai Lau bpf_skops_write_hdr_opt((struct sock *)sk, skb, req, syn_skb, 3723331fca43SMartin KaFai Lau synack_type, &opts); 3724331fca43SMartin KaFai Lau 3725a1ac9c8aSMartin KaFai Lau skb_set_delivery_time(skb, now, true); 3726a842fe14SEric Dumazet tcp_add_tx_delay(skb, tp); 3727a842fe14SEric Dumazet 37281da177e4SLinus Torvalds return skb; 37291da177e4SLinus Torvalds } 37304bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_make_synack); 37311da177e4SLinus Torvalds 373281164413SDaniel Borkmann static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst) 373381164413SDaniel Borkmann { 373481164413SDaniel Borkmann struct inet_connection_sock *icsk = inet_csk(sk); 373581164413SDaniel Borkmann const struct tcp_congestion_ops *ca; 373681164413SDaniel Borkmann u32 ca_key = dst_metric(dst, RTAX_CC_ALGO); 373781164413SDaniel Borkmann 373881164413SDaniel Borkmann if (ca_key == TCP_CA_UNSPEC) 373981164413SDaniel Borkmann return; 374081164413SDaniel Borkmann 374181164413SDaniel Borkmann rcu_read_lock(); 374281164413SDaniel Borkmann ca = tcp_ca_find_key(ca_key); 37430baf26b0SMartin KaFai Lau if (likely(ca && bpf_try_module_get(ca, ca->owner))) { 37440baf26b0SMartin KaFai Lau bpf_module_put(icsk->icsk_ca_ops, icsk->icsk_ca_ops->owner); 374581164413SDaniel Borkmann icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst); 374681164413SDaniel Borkmann icsk->icsk_ca_ops = ca; 374781164413SDaniel Borkmann } 374881164413SDaniel Borkmann rcu_read_unlock(); 374981164413SDaniel Borkmann } 375081164413SDaniel Borkmann 375167edfef7SAndi Kleen /* Do all connect socket setups that can be done AF independent. */ 3752f7e56a76Sstephen hemminger static void tcp_connect_init(struct sock *sk) 37531da177e4SLinus Torvalds { 3754cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 37551da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 37561da177e4SLinus Torvalds __u8 rcv_wscale; 375713d3b1ebSLawrence Brakmo u32 rcv_wnd; 37581da177e4SLinus Torvalds 37591da177e4SLinus Torvalds /* We'll fix this up when we get a response from the other end. 37601da177e4SLinus Torvalds * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 37611da177e4SLinus Torvalds */ 37625d2ed052SEric Dumazet tp->tcp_header_len = sizeof(struct tcphdr); 37633666f666SKuniyuki Iwashima if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_timestamps)) 37645d2ed052SEric Dumazet tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED; 37651da177e4SLinus Torvalds 37661da177e4SLinus Torvalds /* If user gave his TCP_MAXSEG, record it to clamp */ 37671da177e4SLinus Torvalds if (tp->rx_opt.user_mss) 37681da177e4SLinus Torvalds tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 37691da177e4SLinus Torvalds tp->max_window = 0; 37705d424d5aSJohn Heffner tcp_mtup_init(sk); 37711da177e4SLinus Torvalds tcp_sync_mss(sk, dst_mtu(dst)); 37721da177e4SLinus Torvalds 377381164413SDaniel Borkmann tcp_ca_dst_init(sk, dst); 377481164413SDaniel Borkmann 37751da177e4SLinus Torvalds if (!tp->window_clamp) 37761da177e4SLinus Torvalds tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 37773541f9e8SEric Dumazet tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 3778f5fff5dcSTom Quetchenbach 37791da177e4SLinus Torvalds tcp_initialize_rcv_mss(sk); 37801da177e4SLinus Torvalds 3781e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 3782e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 3783e88c64f0SHagen Paul Pfeifer (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) 3784e88c64f0SHagen Paul Pfeifer tp->window_clamp = tcp_full_space(sk); 3785e88c64f0SHagen Paul Pfeifer 378613d3b1ebSLawrence Brakmo rcv_wnd = tcp_rwnd_init_bpf(sk); 378713d3b1ebSLawrence Brakmo if (rcv_wnd == 0) 378813d3b1ebSLawrence Brakmo rcv_wnd = dst_metric(dst, RTAX_INITRWND); 378913d3b1ebSLawrence Brakmo 3790ceef9ab6SEric Dumazet tcp_select_initial_window(sk, tcp_full_space(sk), 37911da177e4SLinus Torvalds tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 37921da177e4SLinus Torvalds &tp->rcv_wnd, 37931da177e4SLinus Torvalds &tp->window_clamp, 37943666f666SKuniyuki Iwashima READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_window_scaling), 379531d12926Slaurent chavey &rcv_wscale, 379613d3b1ebSLawrence Brakmo rcv_wnd); 37971da177e4SLinus Torvalds 37981da177e4SLinus Torvalds tp->rx_opt.rcv_wscale = rcv_wscale; 37991da177e4SLinus Torvalds tp->rcv_ssthresh = tp->rcv_wnd; 38001da177e4SLinus Torvalds 3801e13ec3daSEric Dumazet WRITE_ONCE(sk->sk_err, 0); 38021da177e4SLinus Torvalds sock_reset_flag(sk, SOCK_DONE); 38031da177e4SLinus Torvalds tp->snd_wnd = 0; 3804ee7537b6SHantzis Fotis tcp_init_wl(tp, 0); 38057f582b24SEric Dumazet tcp_write_queue_purge(sk); 38061da177e4SLinus Torvalds tp->snd_una = tp->write_seq; 38071da177e4SLinus Torvalds tp->snd_sml = tp->write_seq; 380833f5f57eSIlpo Järvinen tp->snd_up = tp->write_seq; 3809e0d694d6SEric Dumazet WRITE_ONCE(tp->snd_nxt, tp->write_seq); 3810ee995283SPavel Emelyanov 3811ee995283SPavel Emelyanov if (likely(!tp->repair)) 38121da177e4SLinus Torvalds tp->rcv_nxt = 0; 3813c7781a6eSAndrew Vagin else 381470eabf0eSEric Dumazet tp->rcv_tstamp = tcp_jiffies32; 3815ee995283SPavel Emelyanov tp->rcv_wup = tp->rcv_nxt; 38167db48e98SEric Dumazet WRITE_ONCE(tp->copied_seq, tp->rcv_nxt); 38171da177e4SLinus Torvalds 38188550f328SLawrence Brakmo inet_csk(sk)->icsk_rto = tcp_timeout_init(sk); 3819463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_retransmits = 0; 38201da177e4SLinus Torvalds tcp_clear_retrans(tp); 38211da177e4SLinus Torvalds } 38221da177e4SLinus Torvalds 3823783237e8SYuchung Cheng static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) 3824783237e8SYuchung Cheng { 3825783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3826783237e8SYuchung Cheng struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 3827783237e8SYuchung Cheng 3828783237e8SYuchung Cheng tcb->end_seq += skb->len; 3829f4a775d1SEric Dumazet __skb_header_release(skb); 3830ab4e846aSEric Dumazet sk_wmem_queued_add(sk, skb->truesize); 3831783237e8SYuchung Cheng sk_mem_charge(sk, skb->truesize); 38320f317464SEric Dumazet WRITE_ONCE(tp->write_seq, tcb->end_seq); 3833783237e8SYuchung Cheng tp->packets_out += tcp_skb_pcount(skb); 3834783237e8SYuchung Cheng } 3835783237e8SYuchung Cheng 3836783237e8SYuchung Cheng /* Build and send a SYN with data and (cached) Fast Open cookie. However, 3837783237e8SYuchung Cheng * queue a data-only packet after the regular SYN, such that regular SYNs 3838783237e8SYuchung Cheng * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges 3839783237e8SYuchung Cheng * only the SYN sequence, the data are retransmitted in the first ACK. 3840783237e8SYuchung Cheng * If cookie is not cached or other error occurs, falls back to send a 3841783237e8SYuchung Cheng * regular SYN with Fast Open cookie request option. 3842783237e8SYuchung Cheng */ 3843783237e8SYuchung Cheng static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) 3844783237e8SYuchung Cheng { 3845ed0c99dcSJakub Kicinski struct inet_connection_sock *icsk = inet_csk(sk); 3846783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3847783237e8SYuchung Cheng struct tcp_fastopen_request *fo = tp->fastopen_req; 3848fbf93406SEric Dumazet struct page_frag *pfrag = sk_page_frag(sk); 3849355a901eSEric Dumazet struct sk_buff *syn_data; 3850fbf93406SEric Dumazet int space, err = 0; 3851783237e8SYuchung Cheng 385267da22d2SYuchung Cheng tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ 3853065263f4SWei Wang if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie)) 3854783237e8SYuchung Cheng goto fallback; 3855783237e8SYuchung Cheng 3856783237e8SYuchung Cheng /* MSS for SYN-data is based on cached MSS and bounded by PMTU and 3857783237e8SYuchung Cheng * user-MSS. Reserve maximum option space for middleboxes that add 3858783237e8SYuchung Cheng * private TCP options. The cost is reduced data space in SYN :( 3859783237e8SYuchung Cheng */ 38603541f9e8SEric Dumazet tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp); 3861ed0c99dcSJakub Kicinski /* Sync mss_cache after updating the mss_clamp */ 3862ed0c99dcSJakub Kicinski tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 38633541f9e8SEric Dumazet 3864ed0c99dcSJakub Kicinski space = __tcp_mtu_to_mss(sk, icsk->icsk_pmtu_cookie) - 3865783237e8SYuchung Cheng MAX_TCP_OPTION_SPACE; 3866783237e8SYuchung Cheng 3867f5ddcbbbSEric Dumazet space = min_t(size_t, space, fo->size); 3868f5ddcbbbSEric Dumazet 3869fbf93406SEric Dumazet if (space && 3870fbf93406SEric Dumazet !skb_page_frag_refill(min_t(size_t, space, PAGE_SIZE), 3871fbf93406SEric Dumazet pfrag, sk->sk_allocation)) 3872fbf93406SEric Dumazet goto fallback; 38735882efffSEric Dumazet syn_data = tcp_stream_alloc_skb(sk, sk->sk_allocation, false); 3874355a901eSEric Dumazet if (!syn_data) 3875783237e8SYuchung Cheng goto fallback; 3876355a901eSEric Dumazet memcpy(syn_data->cb, syn->cb, sizeof(syn->cb)); 387707e100f9SEric Dumazet if (space) { 3878fbf93406SEric Dumazet space = min_t(size_t, space, pfrag->size - pfrag->offset); 3879fbf93406SEric Dumazet space = tcp_wmem_schedule(sk, space); 3880fbf93406SEric Dumazet } 3881fbf93406SEric Dumazet if (space) { 3882fbf93406SEric Dumazet space = copy_page_from_iter(pfrag->page, pfrag->offset, 3883fbf93406SEric Dumazet space, &fo->data->msg_iter); 3884fbf93406SEric Dumazet if (unlikely(!space)) { 3885ba233b34SEric Dumazet tcp_skb_tsorted_anchor_cleanup(syn_data); 3886355a901eSEric Dumazet kfree_skb(syn_data); 3887783237e8SYuchung Cheng goto fallback; 3888783237e8SYuchung Cheng } 3889fbf93406SEric Dumazet skb_fill_page_desc(syn_data, 0, pfrag->page, 3890fbf93406SEric Dumazet pfrag->offset, space); 3891fbf93406SEric Dumazet page_ref_inc(pfrag->page); 3892fbf93406SEric Dumazet pfrag->offset += space; 3893fbf93406SEric Dumazet skb_len_add(syn_data, space); 3894f859a448SWillem de Bruijn skb_zcopy_set(syn_data, fo->uarg, NULL); 389507e100f9SEric Dumazet } 3896355a901eSEric Dumazet /* No more data pending in inet_wait_for_connect() */ 3897355a901eSEric Dumazet if (space == fo->size) 3898355a901eSEric Dumazet fo->data = NULL; 3899355a901eSEric Dumazet fo->copied = space; 3900783237e8SYuchung Cheng 3901355a901eSEric Dumazet tcp_connect_queue_skb(sk, syn_data); 39020f87230dSFrancis Yan if (syn_data->len) 39030f87230dSFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_BUSY); 3904355a901eSEric Dumazet 3905355a901eSEric Dumazet err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation); 3906355a901eSEric Dumazet 3907a1ac9c8aSMartin KaFai Lau skb_set_delivery_time(syn, syn_data->skb_mstamp_ns, true); 3908355a901eSEric Dumazet 3909355a901eSEric Dumazet /* Now full SYN+DATA was cloned and sent (or not), 3910355a901eSEric Dumazet * remove the SYN from the original skb (syn_data) 3911355a901eSEric Dumazet * we keep in write queue in case of a retransmit, as we 3912355a901eSEric Dumazet * also have the SYN packet (with no data) in the same queue. 3913431a9124SEric Dumazet */ 3914355a901eSEric Dumazet TCP_SKB_CB(syn_data)->seq++; 3915355a901eSEric Dumazet TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH; 3916355a901eSEric Dumazet if (!err) { 391767da22d2SYuchung Cheng tp->syn_data = (fo->copied > 0); 391875c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data); 3919f19c29e3SYuchung Cheng NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT); 3920783237e8SYuchung Cheng goto done; 3921783237e8SYuchung Cheng } 3922783237e8SYuchung Cheng 392375c119afSEric Dumazet /* data was not sent, put it in write_queue */ 392475c119afSEric Dumazet __skb_queue_tail(&sk->sk_write_queue, syn_data); 3925b5b7db8dSEric Dumazet tp->packets_out -= tcp_skb_pcount(syn_data); 3926b5b7db8dSEric Dumazet 3927783237e8SYuchung Cheng fallback: 3928783237e8SYuchung Cheng /* Send a regular SYN with Fast Open cookie request option */ 3929783237e8SYuchung Cheng if (fo->cookie.len > 0) 3930783237e8SYuchung Cheng fo->cookie.len = 0; 3931783237e8SYuchung Cheng err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); 3932783237e8SYuchung Cheng if (err) 3933783237e8SYuchung Cheng tp->syn_fastopen = 0; 3934783237e8SYuchung Cheng done: 3935783237e8SYuchung Cheng fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ 3936783237e8SYuchung Cheng return err; 3937783237e8SYuchung Cheng } 3938783237e8SYuchung Cheng 393967edfef7SAndi Kleen /* Build a SYN and send it off. */ 39401da177e4SLinus Torvalds int tcp_connect(struct sock *sk) 39411da177e4SLinus Torvalds { 39421da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 39431da177e4SLinus Torvalds struct sk_buff *buff; 3944ee586811SEric Paris int err; 39451da177e4SLinus Torvalds 3946de525be2SLawrence Brakmo tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB, 0, NULL); 39478ba60924SEric Dumazet 39488ba60924SEric Dumazet if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 39498ba60924SEric Dumazet return -EHOSTUNREACH; /* Routing failure or similar. */ 39508ba60924SEric Dumazet 39511da177e4SLinus Torvalds tcp_connect_init(sk); 39521da177e4SLinus Torvalds 39532b916477SAndrey Vagin if (unlikely(tp->repair)) { 39542b916477SAndrey Vagin tcp_finish_connect(sk, NULL); 39552b916477SAndrey Vagin return 0; 39562b916477SAndrey Vagin } 39572b916477SAndrey Vagin 39585882efffSEric Dumazet buff = tcp_stream_alloc_skb(sk, sk->sk_allocation, true); 3959355a901eSEric Dumazet if (unlikely(!buff)) 39601da177e4SLinus Torvalds return -ENOBUFS; 39611da177e4SLinus Torvalds 3962a3433f35SChangli Gao tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); 39639a568de4SEric Dumazet tcp_mstamp_refresh(tp); 3964*9d0c00f5SEric Dumazet tp->retrans_stamp = tcp_time_stamp_ts(tp); 3965783237e8SYuchung Cheng tcp_connect_queue_skb(sk, buff); 3966735d3831SFlorian Westphal tcp_ecn_send_syn(sk, buff); 396775c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, buff); 39681da177e4SLinus Torvalds 3969783237e8SYuchung Cheng /* Send off SYN; include data in Fast Open. */ 3970783237e8SYuchung Cheng err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : 3971783237e8SYuchung Cheng tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); 3972ee586811SEric Paris if (err == -ECONNREFUSED) 3973ee586811SEric Paris return err; 3974bd37a088SWei Yongjun 3975bd37a088SWei Yongjun /* We change tp->snd_nxt after the tcp_transmit_skb() call 3976bd37a088SWei Yongjun * in order to make this packet get counted in tcpOutSegs. 3977bd37a088SWei Yongjun */ 3978e0d694d6SEric Dumazet WRITE_ONCE(tp->snd_nxt, tp->write_seq); 3979bd37a088SWei Yongjun tp->pushed_seq = tp->write_seq; 3980b5b7db8dSEric Dumazet buff = tcp_send_head(sk); 3981b5b7db8dSEric Dumazet if (unlikely(buff)) { 3982e0d694d6SEric Dumazet WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(buff)->seq); 3983b5b7db8dSEric Dumazet tp->pushed_seq = TCP_SKB_CB(buff)->seq; 3984b5b7db8dSEric Dumazet } 398581cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); 39861da177e4SLinus Torvalds 39871da177e4SLinus Torvalds /* Timer for repeating the SYN until an answer. */ 39883f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 39893f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 39901da177e4SLinus Torvalds return 0; 39911da177e4SLinus Torvalds } 39924bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_connect); 39931da177e4SLinus Torvalds 3994bbf80d71SEric Dumazet u32 tcp_delack_max(const struct sock *sk) 3995bbf80d71SEric Dumazet { 3996bbf80d71SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 3997bbf80d71SEric Dumazet u32 delack_max = inet_csk(sk)->icsk_delack_max; 3998bbf80d71SEric Dumazet 3999bbf80d71SEric Dumazet if (dst && dst_metric_locked(dst, RTAX_RTO_MIN)) { 4000bbf80d71SEric Dumazet u32 rto_min = dst_metric_rtt(dst, RTAX_RTO_MIN); 4001bbf80d71SEric Dumazet u32 delack_from_rto_min = max_t(int, 1, rto_min - 1); 4002bbf80d71SEric Dumazet 4003bbf80d71SEric Dumazet delack_max = min_t(u32, delack_max, delack_from_rto_min); 4004bbf80d71SEric Dumazet } 4005bbf80d71SEric Dumazet return delack_max; 4006bbf80d71SEric Dumazet } 4007bbf80d71SEric Dumazet 40081da177e4SLinus Torvalds /* Send out a delayed ack, the caller does the policy checking 40091da177e4SLinus Torvalds * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 40101da177e4SLinus Torvalds * for details. 40111da177e4SLinus Torvalds */ 40121da177e4SLinus Torvalds void tcp_send_delayed_ack(struct sock *sk) 40131da177e4SLinus Torvalds { 4014463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 4015463c84b9SArnaldo Carvalho de Melo int ato = icsk->icsk_ack.ato; 40161da177e4SLinus Torvalds unsigned long timeout; 40171da177e4SLinus Torvalds 40181da177e4SLinus Torvalds if (ato > TCP_DELACK_MIN) { 4019463c84b9SArnaldo Carvalho de Melo const struct tcp_sock *tp = tcp_sk(sk); 40201da177e4SLinus Torvalds int max_ato = HZ / 2; 40211da177e4SLinus Torvalds 402231954cd8SWei Wang if (inet_csk_in_pingpong_mode(sk) || 4023056834d9SIlpo Järvinen (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 40241da177e4SLinus Torvalds max_ato = TCP_DELACK_MAX; 40251da177e4SLinus Torvalds 40261da177e4SLinus Torvalds /* Slow path, intersegment interval is "high". */ 40271da177e4SLinus Torvalds 40281da177e4SLinus Torvalds /* If some rtt estimate is known, use it to bound delayed ack. 4029463c84b9SArnaldo Carvalho de Melo * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 40301da177e4SLinus Torvalds * directly. 40311da177e4SLinus Torvalds */ 4032740b0f18SEric Dumazet if (tp->srtt_us) { 4033740b0f18SEric Dumazet int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3), 4034740b0f18SEric Dumazet TCP_DELACK_MIN); 40351da177e4SLinus Torvalds 40361da177e4SLinus Torvalds if (rtt < max_ato) 40371da177e4SLinus Torvalds max_ato = rtt; 40381da177e4SLinus Torvalds } 40391da177e4SLinus Torvalds 40401da177e4SLinus Torvalds ato = min(ato, max_ato); 40411da177e4SLinus Torvalds } 40421da177e4SLinus Torvalds 4043bbf80d71SEric Dumazet ato = min_t(u32, ato, tcp_delack_max(sk)); 40442b8ee4f0SMartin KaFai Lau 40451da177e4SLinus Torvalds /* Stay within the limit we were given */ 40461da177e4SLinus Torvalds timeout = jiffies + ato; 40471da177e4SLinus Torvalds 40481da177e4SLinus Torvalds /* Use new timeout only if there wasn't a older one earlier. */ 4049463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 4050b6b6d653SEric Dumazet /* If delack timer is about to expire, send ACK now. */ 4051b6b6d653SEric Dumazet if (time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 40521da177e4SLinus Torvalds tcp_send_ack(sk); 40531da177e4SLinus Torvalds return; 40541da177e4SLinus Torvalds } 40551da177e4SLinus Torvalds 4056463c84b9SArnaldo Carvalho de Melo if (!time_before(timeout, icsk->icsk_ack.timeout)) 4057463c84b9SArnaldo Carvalho de Melo timeout = icsk->icsk_ack.timeout; 40581da177e4SLinus Torvalds } 4059463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 4060463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.timeout = timeout; 4061463c84b9SArnaldo Carvalho de Melo sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 40621da177e4SLinus Torvalds } 40631da177e4SLinus Torvalds 40641da177e4SLinus Torvalds /* This routine sends an ack and also updates the window. */ 40652987babbSYuchung Cheng void __tcp_send_ack(struct sock *sk, u32 rcv_nxt) 40661da177e4SLinus Torvalds { 40671da177e4SLinus Torvalds struct sk_buff *buff; 40681da177e4SLinus Torvalds 4069058dc334SIlpo Järvinen /* If we have been reset, we may not send again. */ 4070058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 4071058dc334SIlpo Järvinen return; 4072058dc334SIlpo Järvinen 40731da177e4SLinus Torvalds /* We are not putting this on the write queue, so 40741da177e4SLinus Torvalds * tcp_transmit_skb() will set the ownership to this 40751da177e4SLinus Torvalds * sock. 40761da177e4SLinus Torvalds */ 40777450aaf6SEric Dumazet buff = alloc_skb(MAX_TCP_HEADER, 40787450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 40797450aaf6SEric Dumazet if (unlikely(!buff)) { 4080a37c2134SEric Dumazet struct inet_connection_sock *icsk = inet_csk(sk); 4081a37c2134SEric Dumazet unsigned long delay; 4082a37c2134SEric Dumazet 4083a37c2134SEric Dumazet delay = TCP_DELACK_MAX << icsk->icsk_ack.retry; 4084a37c2134SEric Dumazet if (delay < TCP_RTO_MAX) 4085a37c2134SEric Dumazet icsk->icsk_ack.retry++; 4086463c84b9SArnaldo Carvalho de Melo inet_csk_schedule_ack(sk); 4087a37c2134SEric Dumazet icsk->icsk_ack.ato = TCP_ATO_MIN; 4088a37c2134SEric Dumazet inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, delay, TCP_RTO_MAX); 40891da177e4SLinus Torvalds return; 40901da177e4SLinus Torvalds } 40911da177e4SLinus Torvalds 40921da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 40931da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 4094a3433f35SChangli Gao tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); 40951da177e4SLinus Torvalds 409698781965SEric Dumazet /* We do not want pure acks influencing TCP Small Queues or fq/pacing 409798781965SEric Dumazet * too much. 409898781965SEric Dumazet * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784 409998781965SEric Dumazet */ 410098781965SEric Dumazet skb_set_tcp_pure_ack(buff); 410198781965SEric Dumazet 41021da177e4SLinus Torvalds /* Send it off, this clears delayed acks for us. */ 41032987babbSYuchung Cheng __tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0, rcv_nxt); 41041da177e4SLinus Torvalds } 410527cde44aSYuchung Cheng EXPORT_SYMBOL_GPL(__tcp_send_ack); 41062987babbSYuchung Cheng 41072987babbSYuchung Cheng void tcp_send_ack(struct sock *sk) 41082987babbSYuchung Cheng { 41092987babbSYuchung Cheng __tcp_send_ack(sk, tcp_sk(sk)->rcv_nxt); 41101da177e4SLinus Torvalds } 41111da177e4SLinus Torvalds 41121da177e4SLinus Torvalds /* This routine sends a packet with an out of date sequence 41131da177e4SLinus Torvalds * number. It assumes the other end will try to ack it. 41141da177e4SLinus Torvalds * 41151da177e4SLinus Torvalds * Question: what should we make while urgent mode? 41161da177e4SLinus Torvalds * 4.4BSD forces sending single byte of data. We cannot send 41171da177e4SLinus Torvalds * out of window data, because we have SND.NXT==SND.MAX... 41181da177e4SLinus Torvalds * 41191da177e4SLinus Torvalds * Current solution: to send TWO zero-length segments in urgent mode: 41201da177e4SLinus Torvalds * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 41211da177e4SLinus Torvalds * out-of-date with SND.UNA-1 to probe window. 41221da177e4SLinus Torvalds */ 4123e520af48SEric Dumazet static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib) 41241da177e4SLinus Torvalds { 41251da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 41261da177e4SLinus Torvalds struct sk_buff *skb; 41271da177e4SLinus Torvalds 41281da177e4SLinus Torvalds /* We don't queue it, tcp_transmit_skb() sets ownership. */ 41297450aaf6SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, 41307450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 413151456b29SIan Morris if (!skb) 41321da177e4SLinus Torvalds return -1; 41331da177e4SLinus Torvalds 41341da177e4SLinus Torvalds /* Reserve space for headers and set control bits. */ 41351da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 41361da177e4SLinus Torvalds /* Use a previous sequence. This should cause the other 41371da177e4SLinus Torvalds * end to send an ack. Don't queue or clone SKB, just 41381da177e4SLinus Torvalds * send it. 41391da177e4SLinus Torvalds */ 4140a3433f35SChangli Gao tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); 4141e2e8009fSRenato Westphal NET_INC_STATS(sock_net(sk), mib); 41427450aaf6SEric Dumazet return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0); 41431da177e4SLinus Torvalds } 41441da177e4SLinus Torvalds 4145385e2070SEric Dumazet /* Called from setsockopt( ... TCP_REPAIR ) */ 4146ee995283SPavel Emelyanov void tcp_send_window_probe(struct sock *sk) 4147ee995283SPavel Emelyanov { 4148ee995283SPavel Emelyanov if (sk->sk_state == TCP_ESTABLISHED) { 4149ee995283SPavel Emelyanov tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; 41509a568de4SEric Dumazet tcp_mstamp_refresh(tcp_sk(sk)); 4151e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE); 4152ee995283SPavel Emelyanov } 4153ee995283SPavel Emelyanov } 4154ee995283SPavel Emelyanov 415567edfef7SAndi Kleen /* Initiate keepalive or window probe from timer. */ 4156e520af48SEric Dumazet int tcp_write_wakeup(struct sock *sk, int mib) 41571da177e4SLinus Torvalds { 41581da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 41591da177e4SLinus Torvalds struct sk_buff *skb; 41601da177e4SLinus Torvalds 4161058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 4162058dc334SIlpo Järvinen return -1; 4163058dc334SIlpo Järvinen 416400db4124SIan Morris skb = tcp_send_head(sk); 416500db4124SIan Morris if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { 41661da177e4SLinus Torvalds int err; 41670c54b85fSIlpo Järvinen unsigned int mss = tcp_current_mss(sk); 416890840defSIlpo Järvinen unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 41691da177e4SLinus Torvalds 41701da177e4SLinus Torvalds if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 41711da177e4SLinus Torvalds tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 41721da177e4SLinus Torvalds 41731da177e4SLinus Torvalds /* We are probing the opening of a window 41741da177e4SLinus Torvalds * but the window size is != 0 41751da177e4SLinus Torvalds * must have been a result SWS avoidance ( sender ) 41761da177e4SLinus Torvalds */ 41771da177e4SLinus Torvalds if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 41781da177e4SLinus Torvalds skb->len > mss) { 41791da177e4SLinus Torvalds seg_size = min(seg_size, mss); 41804de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 418175c119afSEric Dumazet if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE, 418275c119afSEric Dumazet skb, seg_size, mss, GFP_ATOMIC)) 41831da177e4SLinus Torvalds return -1; 41841da177e4SLinus Torvalds } else if (!tcp_skb_pcount(skb)) 41855bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss); 41861da177e4SLinus Torvalds 41874de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 4188dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 418966f5fe62SIlpo Järvinen if (!err) 419066f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 41911da177e4SLinus Torvalds return err; 41921da177e4SLinus Torvalds } else { 419333f5f57eSIlpo Järvinen if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) 4194e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 1, mib); 4195e520af48SEric Dumazet return tcp_xmit_probe_skb(sk, 0, mib); 41961da177e4SLinus Torvalds } 41971da177e4SLinus Torvalds } 41981da177e4SLinus Torvalds 41991da177e4SLinus Torvalds /* A window probe timeout has occurred. If window is not closed send 42001da177e4SLinus Torvalds * a partial packet else a zero probe. 42011da177e4SLinus Torvalds */ 42021da177e4SLinus Torvalds void tcp_send_probe0(struct sock *sk) 42031da177e4SLinus Torvalds { 4204463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 42051da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 4206c6214a97SNikolay Borisov struct net *net = sock_net(sk); 4207c1d5674fSYuchung Cheng unsigned long timeout; 42081da177e4SLinus Torvalds int err; 42091da177e4SLinus Torvalds 4210e520af48SEric Dumazet err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE); 42111da177e4SLinus Torvalds 421275c119afSEric Dumazet if (tp->packets_out || tcp_write_queue_empty(sk)) { 42131da177e4SLinus Torvalds /* Cancel probe timer, if it is not required. */ 42146687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 0; 4215463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff = 0; 42169d9b1ee0SEnke Chen icsk->icsk_probes_tstamp = 0; 42171da177e4SLinus Torvalds return; 42181da177e4SLinus Torvalds } 42191da177e4SLinus Torvalds 4220c1d5674fSYuchung Cheng icsk->icsk_probes_out++; 42211da177e4SLinus Torvalds if (err <= 0) { 422239e24435SKuniyuki Iwashima if (icsk->icsk_backoff < READ_ONCE(net->ipv4.sysctl_tcp_retries2)) 4223463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff++; 4224c1d5674fSYuchung Cheng timeout = tcp_probe0_when(sk, TCP_RTO_MAX); 42251da177e4SLinus Torvalds } else { 42261da177e4SLinus Torvalds /* If packet was not sent due to local congestion, 4227c1d5674fSYuchung Cheng * Let senders fight for local resources conservatively. 42281da177e4SLinus Torvalds */ 4229c1d5674fSYuchung Cheng timeout = TCP_RESOURCE_PROBE_INTERVAL; 42301da177e4SLinus Torvalds } 4231344db93aSEnke Chen 4232344db93aSEnke Chen timeout = tcp_clamp_probe0_to_user_timeout(sk, timeout); 42338dc242adSEric Dumazet tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, timeout, TCP_RTO_MAX); 42341da177e4SLinus Torvalds } 42355db92c99SOctavian Purdila 4236ea3bea3aSEric Dumazet int tcp_rtx_synack(const struct sock *sk, struct request_sock *req) 42375db92c99SOctavian Purdila { 42385db92c99SOctavian Purdila const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific; 42395db92c99SOctavian Purdila struct flowi fl; 42405db92c99SOctavian Purdila int res; 42415db92c99SOctavian Purdila 4242cb6cd2ceSAkhmat Karakotov /* Paired with WRITE_ONCE() in sock_setsockopt() */ 4243cb6cd2ceSAkhmat Karakotov if (READ_ONCE(sk->sk_txrehash) == SOCK_TXREHASH_ENABLED) 42445e526552SEric Dumazet WRITE_ONCE(tcp_rsk(req)->txhash, net_tx_rndhash()); 4245331fca43SMartin KaFai Lau res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL, 4246331fca43SMartin KaFai Lau NULL); 42475db92c99SOctavian Purdila if (!res) { 42480a375c82SEric Dumazet TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); 42490a375c82SEric Dumazet NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 4250e9d9da91SEric Dumazet if (unlikely(tcp_passive_fastopen(sk))) { 4251e9d9da91SEric Dumazet /* sk has const attribute because listeners are lockless. 4252e9d9da91SEric Dumazet * However in this case, we are dealing with a passive fastopen 4253e9d9da91SEric Dumazet * socket thus we can change total_retrans value. 4254e9d9da91SEric Dumazet */ 4255e9d9da91SEric Dumazet tcp_sk_rw(sk)->total_retrans++; 4256e9d9da91SEric Dumazet } 4257cf34ce3dSSong Liu trace_tcp_retransmit_synack(sk, req); 42585db92c99SOctavian Purdila } 42595db92c99SOctavian Purdila return res; 42605db92c99SOctavian Purdila } 42615db92c99SOctavian Purdila EXPORT_SYMBOL(tcp_rtx_synack); 4262