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 /* If this is the first data packet sent in response to the 1714a41f453SWei Wang * previous received data, 1724a41f453SWei Wang * and it is a reply for ato after last received packet, 1734a41f453SWei Wang * increase pingpong count. 1741da177e4SLinus Torvalds */ 1754a41f453SWei Wang if (before(tp->lsndtime, icsk->icsk_ack.lrcvtime) && 1764a41f453SWei Wang (u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato) 1774a41f453SWei Wang inet_csk_inc_pingpong_cnt(sk); 1784a41f453SWei Wang 1794a41f453SWei Wang tp->lsndtime = now; 1801da177e4SLinus Torvalds } 1811da177e4SLinus Torvalds 18267edfef7SAndi Kleen /* Account for an ACK we sent. */ 18327cde44aSYuchung Cheng static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts, 18427cde44aSYuchung Cheng u32 rcv_nxt) 1851da177e4SLinus Torvalds { 1865d9f4262SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 1875d9f4262SEric Dumazet 1882b195850SEric Dumazet if (unlikely(tp->compressed_ack)) { 189200d95f4SEric Dumazet NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED, 1902b195850SEric Dumazet tp->compressed_ack); 1912b195850SEric Dumazet tp->compressed_ack = 0; 1925d9f4262SEric Dumazet if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1) 1935d9f4262SEric Dumazet __sock_put(sk); 1945d9f4262SEric Dumazet } 19527cde44aSYuchung Cheng 19627cde44aSYuchung Cheng if (unlikely(rcv_nxt != tp->rcv_nxt)) 19727cde44aSYuchung Cheng return; /* Special ACK sent by DCTCP to reflect ECN */ 198463c84b9SArnaldo Carvalho de Melo tcp_dec_quickack_mode(sk, pkts); 199463c84b9SArnaldo Carvalho de Melo inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); 2001da177e4SLinus Torvalds } 2011da177e4SLinus Torvalds 2021da177e4SLinus Torvalds /* Determine a window scaling and initial window to offer. 2031da177e4SLinus Torvalds * Based on the assumption that the given amount of space 2041da177e4SLinus Torvalds * will be offered. Store the results in the tp structure. 2051da177e4SLinus Torvalds * NOTE: for smooth operation initial space offering should 2061da177e4SLinus Torvalds * be a multiple of mss if possible. We assume here that mss >= 1. 2071da177e4SLinus Torvalds * This MUST be enforced by all callers. 2081da177e4SLinus Torvalds */ 209ceef9ab6SEric Dumazet void tcp_select_initial_window(const struct sock *sk, int __space, __u32 mss, 2101da177e4SLinus Torvalds __u32 *rcv_wnd, __u32 *window_clamp, 21131d12926Slaurent chavey int wscale_ok, __u8 *rcv_wscale, 21231d12926Slaurent chavey __u32 init_rcv_wnd) 2131da177e4SLinus Torvalds { 2141da177e4SLinus Torvalds unsigned int space = (__space < 0 ? 0 : __space); 2151da177e4SLinus Torvalds 2161da177e4SLinus Torvalds /* If no clamp set the clamp to the max possible scaled window */ 2171da177e4SLinus Torvalds if (*window_clamp == 0) 218589c49cbSGao Feng (*window_clamp) = (U16_MAX << TCP_MAX_WSCALE); 2191da177e4SLinus Torvalds space = min(*window_clamp, space); 2201da177e4SLinus Torvalds 2211da177e4SLinus Torvalds /* Quantize space offering to a multiple of mss if possible. */ 2221da177e4SLinus Torvalds if (space > mss) 223589c49cbSGao Feng space = rounddown(space, mss); 2241da177e4SLinus Torvalds 2251da177e4SLinus Torvalds /* NOTE: offering an initial window larger than 32767 22615d99e02SRick Jones * will break some buggy TCP stacks. If the admin tells us 22715d99e02SRick Jones * it is likely we could be speaking with such a buggy stack 22815d99e02SRick Jones * we will truncate our initial window offering to 32K-1 22915d99e02SRick Jones * unless the remote has sent us a window scaling option, 23015d99e02SRick Jones * which we interpret as a sign the remote TCP is not 23115d99e02SRick Jones * misinterpreting the window field as a signed quantity. 2321da177e4SLinus Torvalds */ 233ceef9ab6SEric Dumazet if (sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows) 2341da177e4SLinus Torvalds (*rcv_wnd) = min(space, MAX_TCP_WINDOW); 23515d99e02SRick Jones else 236a337531bSYuchung Cheng (*rcv_wnd) = min_t(u32, space, U16_MAX); 237a337531bSYuchung Cheng 238a337531bSYuchung Cheng if (init_rcv_wnd) 239a337531bSYuchung Cheng *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss); 24015d99e02SRick Jones 24119bf6261SEric Dumazet *rcv_wscale = 0; 2421da177e4SLinus Torvalds if (wscale_ok) { 243589c49cbSGao Feng /* Set window scaling on max possible window */ 244356d1833SEric Dumazet space = max_t(u32, space, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]); 245f626300aSSoheil Hassas Yeganeh space = max_t(u32, space, sysctl_rmem_max); 246316c1592SStephen Hemminger space = min_t(u32, space, *window_clamp); 24719bf6261SEric Dumazet *rcv_wscale = clamp_t(int, ilog2(space) - 15, 24819bf6261SEric Dumazet 0, TCP_MAX_WSCALE); 2491da177e4SLinus Torvalds } 2501da177e4SLinus Torvalds /* Set the clamp no higher than max representable value */ 251589c49cbSGao Feng (*window_clamp) = min_t(__u32, U16_MAX << (*rcv_wscale), *window_clamp); 2521da177e4SLinus Torvalds } 2534bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_select_initial_window); 2541da177e4SLinus Torvalds 2551da177e4SLinus Torvalds /* Chose a new window to advertise, update state in tcp_sock for the 2561da177e4SLinus Torvalds * socket, and return result with RFC1323 scaling applied. The return 2571da177e4SLinus Torvalds * value can be stuffed directly into th->window for an outgoing 2581da177e4SLinus Torvalds * frame. 2591da177e4SLinus Torvalds */ 26040efc6faSStephen Hemminger static u16 tcp_select_window(struct sock *sk) 2611da177e4SLinus Torvalds { 2621da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2638e165e20SFlorian Westphal u32 old_win = tp->rcv_wnd; 2641da177e4SLinus Torvalds u32 cur_win = tcp_receive_window(tp); 2651da177e4SLinus Torvalds u32 new_win = __tcp_select_window(sk); 2661da177e4SLinus Torvalds 2671da177e4SLinus Torvalds /* Never shrink the offered window */ 2681da177e4SLinus Torvalds if (new_win < cur_win) { 2691da177e4SLinus Torvalds /* Danger Will Robinson! 2701da177e4SLinus Torvalds * Don't update rcv_wup/rcv_wnd here or else 2711da177e4SLinus Torvalds * we will not be able to advertise a zero 2721da177e4SLinus Torvalds * window in time. --DaveM 2731da177e4SLinus Torvalds * 2741da177e4SLinus Torvalds * Relax Will Robinson. 2751da177e4SLinus Torvalds */ 2768e165e20SFlorian Westphal if (new_win == 0) 2778e165e20SFlorian Westphal NET_INC_STATS(sock_net(sk), 2788e165e20SFlorian Westphal LINUX_MIB_TCPWANTZEROWINDOWADV); 279607bfbf2SPatrick McHardy new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale); 2801da177e4SLinus Torvalds } 2811da177e4SLinus Torvalds tp->rcv_wnd = new_win; 2821da177e4SLinus Torvalds tp->rcv_wup = tp->rcv_nxt; 2831da177e4SLinus Torvalds 2841da177e4SLinus Torvalds /* Make sure we do not exceed the maximum possible 2851da177e4SLinus Torvalds * scaled window. 2861da177e4SLinus Torvalds */ 287ceef9ab6SEric Dumazet if (!tp->rx_opt.rcv_wscale && 288ceef9ab6SEric Dumazet sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows) 2891da177e4SLinus Torvalds new_win = min(new_win, MAX_TCP_WINDOW); 2901da177e4SLinus Torvalds else 2911da177e4SLinus Torvalds new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); 2921da177e4SLinus Torvalds 2931da177e4SLinus Torvalds /* RFC1323 scaling applied */ 2941da177e4SLinus Torvalds new_win >>= tp->rx_opt.rcv_wscale; 2951da177e4SLinus Torvalds 29631770e34SFlorian Westphal /* If we advertise zero window, disable fast path. */ 2978e165e20SFlorian Westphal if (new_win == 0) { 29831770e34SFlorian Westphal tp->pred_flags = 0; 2998e165e20SFlorian Westphal if (old_win) 3008e165e20SFlorian Westphal NET_INC_STATS(sock_net(sk), 3018e165e20SFlorian Westphal LINUX_MIB_TCPTOZEROWINDOWADV); 3028e165e20SFlorian Westphal } else if (old_win == 0) { 3038e165e20SFlorian Westphal NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV); 3048e165e20SFlorian Westphal } 3051da177e4SLinus Torvalds 3061da177e4SLinus Torvalds return new_win; 3071da177e4SLinus Torvalds } 3081da177e4SLinus Torvalds 30967edfef7SAndi Kleen /* Packet ECN state for a SYN-ACK */ 310735d3831SFlorian Westphal static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb) 311bdf1ee5dSIlpo Järvinen { 31230e502a3SDaniel Borkmann const struct tcp_sock *tp = tcp_sk(sk); 31330e502a3SDaniel Borkmann 3144de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR; 315bdf1ee5dSIlpo Järvinen if (!(tp->ecn_flags & TCP_ECN_OK)) 3164de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE; 31791b5b21cSLawrence Brakmo else if (tcp_ca_needs_ecn(sk) || 31891b5b21cSLawrence Brakmo tcp_bpf_ca_needs_ecn(sk)) 31930e502a3SDaniel Borkmann INET_ECN_xmit(sk); 320bdf1ee5dSIlpo Järvinen } 321bdf1ee5dSIlpo Järvinen 32267edfef7SAndi Kleen /* Packet ECN state for a SYN. */ 323735d3831SFlorian Westphal static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb) 324bdf1ee5dSIlpo Järvinen { 325bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 32691b5b21cSLawrence Brakmo bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk); 3274785a667SKuniyuki Iwashima bool use_ecn = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_ecn) == 1 || 32891b5b21cSLawrence Brakmo tcp_ca_needs_ecn(sk) || bpf_needs_ecn; 329f7b3bec6SFlorian Westphal 330f7b3bec6SFlorian Westphal if (!use_ecn) { 331f7b3bec6SFlorian Westphal const struct dst_entry *dst = __sk_dst_get(sk); 332f7b3bec6SFlorian Westphal 333f7b3bec6SFlorian Westphal if (dst && dst_feature(dst, RTAX_FEATURE_ECN)) 334f7b3bec6SFlorian Westphal use_ecn = true; 335f7b3bec6SFlorian Westphal } 336bdf1ee5dSIlpo Järvinen 337bdf1ee5dSIlpo Järvinen tp->ecn_flags = 0; 338f7b3bec6SFlorian Westphal 339f7b3bec6SFlorian Westphal if (use_ecn) { 3404de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR; 341bdf1ee5dSIlpo Järvinen tp->ecn_flags = TCP_ECN_OK; 34291b5b21cSLawrence Brakmo if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn) 34330e502a3SDaniel Borkmann INET_ECN_xmit(sk); 344bdf1ee5dSIlpo Järvinen } 345bdf1ee5dSIlpo Järvinen } 346bdf1ee5dSIlpo Järvinen 34749213555SDaniel Borkmann static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb) 34849213555SDaniel Borkmann { 34912b8d9caSKuniyuki Iwashima if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback)) 35049213555SDaniel Borkmann /* tp->ecn_flags are cleared at a later point in time when 35149213555SDaniel Borkmann * SYN ACK is ultimatively being received. 35249213555SDaniel Borkmann */ 35349213555SDaniel Borkmann TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR); 35449213555SDaniel Borkmann } 35549213555SDaniel Borkmann 356735d3831SFlorian Westphal static void 3576ac705b1SEric Dumazet tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th) 358bdf1ee5dSIlpo Järvinen { 3596ac705b1SEric Dumazet if (inet_rsk(req)->ecn_ok) 360bdf1ee5dSIlpo Järvinen th->ece = 1; 361bdf1ee5dSIlpo Järvinen } 362bdf1ee5dSIlpo Järvinen 36367edfef7SAndi Kleen /* Set up ECN state for a packet on a ESTABLISHED socket that is about to 36467edfef7SAndi Kleen * be sent. 36567edfef7SAndi Kleen */ 366735d3831SFlorian Westphal static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb, 367ea1627c2SEric Dumazet struct tcphdr *th, int tcp_header_len) 368bdf1ee5dSIlpo Järvinen { 369bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 370bdf1ee5dSIlpo Järvinen 371bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_OK) { 372bdf1ee5dSIlpo Järvinen /* Not-retransmitted data segment: set ECT and inject CWR. */ 373bdf1ee5dSIlpo Järvinen if (skb->len != tcp_header_len && 374bdf1ee5dSIlpo Järvinen !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { 375bdf1ee5dSIlpo Järvinen INET_ECN_xmit(sk); 376bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) { 377bdf1ee5dSIlpo Järvinen tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; 378ea1627c2SEric Dumazet th->cwr = 1; 379bdf1ee5dSIlpo Järvinen skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; 380bdf1ee5dSIlpo Järvinen } 38130e502a3SDaniel Borkmann } else if (!tcp_ca_needs_ecn(sk)) { 382bdf1ee5dSIlpo Järvinen /* ACK or retransmitted segment: clear ECT|CE */ 383bdf1ee5dSIlpo Järvinen INET_ECN_dontxmit(sk); 384bdf1ee5dSIlpo Järvinen } 385bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) 386ea1627c2SEric Dumazet th->ece = 1; 387bdf1ee5dSIlpo Järvinen } 388bdf1ee5dSIlpo Järvinen } 389bdf1ee5dSIlpo Järvinen 390e870a8efSIlpo Järvinen /* Constructs common control bits of non-data skb. If SYN/FIN is present, 391e870a8efSIlpo Järvinen * auto increment end seqno. 392e870a8efSIlpo Järvinen */ 393e870a8efSIlpo Järvinen static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags) 394e870a8efSIlpo Järvinen { 3952e8e18efSDavid S. Miller skb->ip_summed = CHECKSUM_PARTIAL; 396e870a8efSIlpo Järvinen 3974de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags; 398e870a8efSIlpo Järvinen 399cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, 1); 400e870a8efSIlpo Järvinen 401e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->seq = seq; 402a3433f35SChangli Gao if (flags & (TCPHDR_SYN | TCPHDR_FIN)) 403e870a8efSIlpo Järvinen seq++; 404e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->end_seq = seq; 405e870a8efSIlpo Järvinen } 406e870a8efSIlpo Järvinen 407a2a385d6SEric Dumazet static inline bool tcp_urg_mode(const struct tcp_sock *tp) 40833f5f57eSIlpo Järvinen { 40933f5f57eSIlpo Järvinen return tp->snd_una != tp->snd_up; 41033f5f57eSIlpo Järvinen } 41133f5f57eSIlpo Järvinen 4123b65abb8SLeonard Crestez #define OPTION_SACK_ADVERTISE BIT(0) 4133b65abb8SLeonard Crestez #define OPTION_TS BIT(1) 4143b65abb8SLeonard Crestez #define OPTION_MD5 BIT(2) 4153b65abb8SLeonard Crestez #define OPTION_WSCALE BIT(3) 4163b65abb8SLeonard Crestez #define OPTION_FAST_OPEN_COOKIE BIT(8) 4173b65abb8SLeonard Crestez #define OPTION_SMC BIT(9) 4183b65abb8SLeonard Crestez #define OPTION_MPTCP BIT(10) 41960e2a778SUrsula Braun 42060e2a778SUrsula Braun static void smc_options_write(__be32 *ptr, u16 *options) 42160e2a778SUrsula Braun { 42260e2a778SUrsula Braun #if IS_ENABLED(CONFIG_SMC) 42360e2a778SUrsula Braun if (static_branch_unlikely(&tcp_have_smc)) { 42460e2a778SUrsula Braun if (unlikely(OPTION_SMC & *options)) { 42560e2a778SUrsula Braun *ptr++ = htonl((TCPOPT_NOP << 24) | 42660e2a778SUrsula Braun (TCPOPT_NOP << 16) | 42760e2a778SUrsula Braun (TCPOPT_EXP << 8) | 42860e2a778SUrsula Braun (TCPOLEN_EXP_SMC_BASE)); 42960e2a778SUrsula Braun *ptr++ = htonl(TCPOPT_SMC_MAGIC); 43060e2a778SUrsula Braun } 43160e2a778SUrsula Braun } 43260e2a778SUrsula Braun #endif 43360e2a778SUrsula Braun } 43433ad798cSAdam Langley 43533ad798cSAdam Langley struct tcp_out_options { 4362100c8d2SYuchung Cheng u16 options; /* bit field of OPTION_* */ 4372100c8d2SYuchung Cheng u16 mss; /* 0 to disable */ 43833ad798cSAdam Langley u8 ws; /* window scale, 0 to disable */ 43933ad798cSAdam Langley u8 num_sack_blocks; /* number of SACK blocks to include */ 440bd0388aeSWilliam Allen Simpson u8 hash_size; /* bytes in hash_location */ 441331fca43SMartin KaFai Lau u8 bpf_opt_len; /* length of BPF hdr option */ 442bd0388aeSWilliam Allen Simpson __u8 *hash_location; /* temporary pointer, overloaded */ 4432100c8d2SYuchung Cheng __u32 tsval, tsecr; /* need to include OPTION_TS */ 4442100c8d2SYuchung Cheng struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */ 445eda7acddSPeter Krystad struct mptcp_out_options mptcp; 44633ad798cSAdam Langley }; 44733ad798cSAdam Langley 448ea66758cSPaolo Abeni static void mptcp_options_write(struct tcphdr *th, __be32 *ptr, 449ea66758cSPaolo Abeni struct tcp_sock *tp, 450fa3fe2b1SFlorian Westphal struct tcp_out_options *opts) 451eda7acddSPeter Krystad { 452eda7acddSPeter Krystad #if IS_ENABLED(CONFIG_MPTCP) 453eda7acddSPeter Krystad if (unlikely(OPTION_MPTCP & opts->options)) 454ea66758cSPaolo Abeni mptcp_write_options(th, ptr, tp, &opts->mptcp); 455eda7acddSPeter Krystad #endif 456eda7acddSPeter Krystad } 457eda7acddSPeter Krystad 458331fca43SMartin KaFai Lau #ifdef CONFIG_CGROUP_BPF 4590813a841SMartin KaFai Lau static int bpf_skops_write_hdr_opt_arg0(struct sk_buff *skb, 4600813a841SMartin KaFai Lau enum tcp_synack_type synack_type) 4610813a841SMartin KaFai Lau { 4620813a841SMartin KaFai Lau if (unlikely(!skb)) 4630813a841SMartin KaFai Lau return BPF_WRITE_HDR_TCP_CURRENT_MSS; 4640813a841SMartin KaFai Lau 4650813a841SMartin KaFai Lau if (unlikely(synack_type == TCP_SYNACK_COOKIE)) 4660813a841SMartin KaFai Lau return BPF_WRITE_HDR_TCP_SYNACK_COOKIE; 4670813a841SMartin KaFai Lau 4680813a841SMartin KaFai Lau return 0; 4690813a841SMartin KaFai Lau } 4700813a841SMartin KaFai Lau 471331fca43SMartin KaFai Lau /* req, syn_skb and synack_type are used when writing synack */ 472331fca43SMartin KaFai Lau static void bpf_skops_hdr_opt_len(struct sock *sk, struct sk_buff *skb, 473331fca43SMartin KaFai Lau struct request_sock *req, 474331fca43SMartin KaFai Lau struct sk_buff *syn_skb, 475331fca43SMartin KaFai Lau enum tcp_synack_type synack_type, 476331fca43SMartin KaFai Lau struct tcp_out_options *opts, 477331fca43SMartin KaFai Lau unsigned int *remaining) 478331fca43SMartin KaFai Lau { 4790813a841SMartin KaFai Lau struct bpf_sock_ops_kern sock_ops; 4800813a841SMartin KaFai Lau int err; 4810813a841SMartin KaFai Lau 482331fca43SMartin KaFai Lau if (likely(!BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), 483331fca43SMartin KaFai Lau BPF_SOCK_OPS_WRITE_HDR_OPT_CB_FLAG)) || 484331fca43SMartin KaFai Lau !*remaining) 485331fca43SMartin KaFai Lau return; 486331fca43SMartin KaFai Lau 4870813a841SMartin KaFai Lau /* *remaining has already been aligned to 4 bytes, so *remaining >= 4 */ 4880813a841SMartin KaFai Lau 4890813a841SMartin KaFai Lau /* init sock_ops */ 4900813a841SMartin KaFai Lau memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp)); 4910813a841SMartin KaFai Lau 4920813a841SMartin KaFai Lau sock_ops.op = BPF_SOCK_OPS_HDR_OPT_LEN_CB; 4930813a841SMartin KaFai Lau 4940813a841SMartin KaFai Lau if (req) { 4950813a841SMartin KaFai Lau /* The listen "sk" cannot be passed here because 4960813a841SMartin KaFai Lau * it is not locked. It would not make too much 4970813a841SMartin KaFai Lau * sense to do bpf_setsockopt(listen_sk) based 4980813a841SMartin KaFai Lau * on individual connection request also. 4990813a841SMartin KaFai Lau * 5000813a841SMartin KaFai Lau * Thus, "req" is passed here and the cgroup-bpf-progs 5010813a841SMartin KaFai Lau * of the listen "sk" will be run. 5020813a841SMartin KaFai Lau * 5030813a841SMartin KaFai Lau * "req" is also used here for fastopen even the "sk" here is 5040813a841SMartin KaFai Lau * a fullsock "child" sk. It is to keep the behavior 5050813a841SMartin KaFai Lau * consistent between fastopen and non-fastopen on 5060813a841SMartin KaFai Lau * the bpf programming side. 507331fca43SMartin KaFai Lau */ 5080813a841SMartin KaFai Lau sock_ops.sk = (struct sock *)req; 5090813a841SMartin KaFai Lau sock_ops.syn_skb = syn_skb; 5100813a841SMartin KaFai Lau } else { 5110813a841SMartin KaFai Lau sock_owned_by_me(sk); 5120813a841SMartin KaFai Lau 5130813a841SMartin KaFai Lau sock_ops.is_fullsock = 1; 5140813a841SMartin KaFai Lau sock_ops.sk = sk; 5150813a841SMartin KaFai Lau } 5160813a841SMartin KaFai Lau 5170813a841SMartin KaFai Lau sock_ops.args[0] = bpf_skops_write_hdr_opt_arg0(skb, synack_type); 5180813a841SMartin KaFai Lau sock_ops.remaining_opt_len = *remaining; 5190813a841SMartin KaFai Lau /* tcp_current_mss() does not pass a skb */ 5200813a841SMartin KaFai Lau if (skb) 5210813a841SMartin KaFai Lau bpf_skops_init_skb(&sock_ops, skb, 0); 5220813a841SMartin KaFai Lau 5230813a841SMartin KaFai Lau err = BPF_CGROUP_RUN_PROG_SOCK_OPS_SK(&sock_ops, sk); 5240813a841SMartin KaFai Lau 5250813a841SMartin KaFai Lau if (err || sock_ops.remaining_opt_len == *remaining) 5260813a841SMartin KaFai Lau return; 5270813a841SMartin KaFai Lau 5280813a841SMartin KaFai Lau opts->bpf_opt_len = *remaining - sock_ops.remaining_opt_len; 5290813a841SMartin KaFai Lau /* round up to 4 bytes */ 5300813a841SMartin KaFai Lau opts->bpf_opt_len = (opts->bpf_opt_len + 3) & ~3; 5310813a841SMartin KaFai Lau 5320813a841SMartin KaFai Lau *remaining -= opts->bpf_opt_len; 533331fca43SMartin KaFai Lau } 534331fca43SMartin KaFai Lau 535331fca43SMartin KaFai Lau static void bpf_skops_write_hdr_opt(struct sock *sk, struct sk_buff *skb, 536331fca43SMartin KaFai Lau struct request_sock *req, 537331fca43SMartin KaFai Lau struct sk_buff *syn_skb, 538331fca43SMartin KaFai Lau enum tcp_synack_type synack_type, 539331fca43SMartin KaFai Lau struct tcp_out_options *opts) 540331fca43SMartin KaFai Lau { 5410813a841SMartin KaFai Lau u8 first_opt_off, nr_written, max_opt_len = opts->bpf_opt_len; 5420813a841SMartin KaFai Lau struct bpf_sock_ops_kern sock_ops; 5430813a841SMartin KaFai Lau int err; 5440813a841SMartin KaFai Lau 5450813a841SMartin KaFai Lau if (likely(!max_opt_len)) 546331fca43SMartin KaFai Lau return; 547331fca43SMartin KaFai Lau 5480813a841SMartin KaFai Lau memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp)); 5490813a841SMartin KaFai Lau 5500813a841SMartin KaFai Lau sock_ops.op = BPF_SOCK_OPS_WRITE_HDR_OPT_CB; 5510813a841SMartin KaFai Lau 5520813a841SMartin KaFai Lau if (req) { 5530813a841SMartin KaFai Lau sock_ops.sk = (struct sock *)req; 5540813a841SMartin KaFai Lau sock_ops.syn_skb = syn_skb; 5550813a841SMartin KaFai Lau } else { 5560813a841SMartin KaFai Lau sock_owned_by_me(sk); 5570813a841SMartin KaFai Lau 5580813a841SMartin KaFai Lau sock_ops.is_fullsock = 1; 5590813a841SMartin KaFai Lau sock_ops.sk = sk; 5600813a841SMartin KaFai Lau } 5610813a841SMartin KaFai Lau 5620813a841SMartin KaFai Lau sock_ops.args[0] = bpf_skops_write_hdr_opt_arg0(skb, synack_type); 5630813a841SMartin KaFai Lau sock_ops.remaining_opt_len = max_opt_len; 5640813a841SMartin KaFai Lau first_opt_off = tcp_hdrlen(skb) - max_opt_len; 5650813a841SMartin KaFai Lau bpf_skops_init_skb(&sock_ops, skb, first_opt_off); 5660813a841SMartin KaFai Lau 5670813a841SMartin KaFai Lau err = BPF_CGROUP_RUN_PROG_SOCK_OPS_SK(&sock_ops, sk); 5680813a841SMartin KaFai Lau 5690813a841SMartin KaFai Lau if (err) 5700813a841SMartin KaFai Lau nr_written = 0; 5710813a841SMartin KaFai Lau else 5720813a841SMartin KaFai Lau nr_written = max_opt_len - sock_ops.remaining_opt_len; 5730813a841SMartin KaFai Lau 5740813a841SMartin KaFai Lau if (nr_written < max_opt_len) 5750813a841SMartin KaFai Lau memset(skb->data + first_opt_off + nr_written, TCPOPT_NOP, 5760813a841SMartin KaFai Lau max_opt_len - nr_written); 577331fca43SMartin KaFai Lau } 578331fca43SMartin KaFai Lau #else 579331fca43SMartin KaFai Lau static void bpf_skops_hdr_opt_len(struct sock *sk, struct sk_buff *skb, 580331fca43SMartin KaFai Lau struct request_sock *req, 581331fca43SMartin KaFai Lau struct sk_buff *syn_skb, 582331fca43SMartin KaFai Lau enum tcp_synack_type synack_type, 583331fca43SMartin KaFai Lau struct tcp_out_options *opts, 584331fca43SMartin KaFai Lau unsigned int *remaining) 585331fca43SMartin KaFai Lau { 586331fca43SMartin KaFai Lau } 587331fca43SMartin KaFai Lau 588331fca43SMartin KaFai Lau static void bpf_skops_write_hdr_opt(struct sock *sk, struct sk_buff *skb, 589331fca43SMartin KaFai Lau struct request_sock *req, 590331fca43SMartin KaFai Lau struct sk_buff *syn_skb, 591331fca43SMartin KaFai Lau enum tcp_synack_type synack_type, 592331fca43SMartin KaFai Lau struct tcp_out_options *opts) 593331fca43SMartin KaFai Lau { 594331fca43SMartin KaFai Lau } 595331fca43SMartin KaFai Lau #endif 596331fca43SMartin KaFai Lau 59767edfef7SAndi Kleen /* Write previously computed TCP options to the packet. 59867edfef7SAndi Kleen * 59967edfef7SAndi Kleen * Beware: Something in the Internet is very sensitive to the ordering of 600fd6149d3SIlpo Järvinen * TCP options, we learned this through the hard way, so be careful here. 601fd6149d3SIlpo Järvinen * Luckily we can at least blame others for their non-compliance but from 6028e3bff96Sstephen hemminger * inter-operability perspective it seems that we're somewhat stuck with 603fd6149d3SIlpo Järvinen * the ordering which we have been using if we want to keep working with 604fd6149d3SIlpo Järvinen * those broken things (not that it currently hurts anybody as there isn't 605fd6149d3SIlpo Järvinen * particular reason why the ordering would need to be changed). 606fd6149d3SIlpo Järvinen * 607fd6149d3SIlpo Järvinen * At least SACK_PERM as the first option is known to lead to a disaster 608fd6149d3SIlpo Järvinen * (but it may well be that other scenarios fail similarly). 609fd6149d3SIlpo Järvinen */ 610ea66758cSPaolo Abeni static void tcp_options_write(struct tcphdr *th, struct tcp_sock *tp, 611bd0388aeSWilliam Allen Simpson struct tcp_out_options *opts) 612bd0388aeSWilliam Allen Simpson { 613ea66758cSPaolo Abeni __be32 *ptr = (__be32 *)(th + 1); 6142100c8d2SYuchung Cheng u16 options = opts->options; /* mungable copy */ 615bd0388aeSWilliam Allen Simpson 616bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_MD5 & options)) { 6171a2c6181SChristoph Paasch *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | 6181a2c6181SChristoph Paasch (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG); 619bd0388aeSWilliam Allen Simpson /* overload cookie hash location */ 620bd0388aeSWilliam Allen Simpson opts->hash_location = (__u8 *)ptr; 62133ad798cSAdam Langley ptr += 4; 62233ad798cSAdam Langley } 62333ad798cSAdam Langley 624fd6149d3SIlpo Järvinen if (unlikely(opts->mss)) { 625fd6149d3SIlpo Järvinen *ptr++ = htonl((TCPOPT_MSS << 24) | 626fd6149d3SIlpo Järvinen (TCPOLEN_MSS << 16) | 627fd6149d3SIlpo Järvinen opts->mss); 628fd6149d3SIlpo Järvinen } 629fd6149d3SIlpo Järvinen 630bd0388aeSWilliam Allen Simpson if (likely(OPTION_TS & options)) { 631bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 63233ad798cSAdam Langley *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | 63333ad798cSAdam Langley (TCPOLEN_SACK_PERM << 16) | 63433ad798cSAdam Langley (TCPOPT_TIMESTAMP << 8) | 63533ad798cSAdam Langley TCPOLEN_TIMESTAMP); 636bd0388aeSWilliam Allen Simpson options &= ~OPTION_SACK_ADVERTISE; 63733ad798cSAdam Langley } else { 638496c98dfSYOSHIFUJI Hideaki *ptr++ = htonl((TCPOPT_NOP << 24) | 63940efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 64040efc6faSStephen Hemminger (TCPOPT_TIMESTAMP << 8) | 64140efc6faSStephen Hemminger TCPOLEN_TIMESTAMP); 64240efc6faSStephen Hemminger } 64333ad798cSAdam Langley *ptr++ = htonl(opts->tsval); 64433ad798cSAdam Langley *ptr++ = htonl(opts->tsecr); 64533ad798cSAdam Langley } 64633ad798cSAdam Langley 647bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 64833ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 64933ad798cSAdam Langley (TCPOPT_NOP << 16) | 65033ad798cSAdam Langley (TCPOPT_SACK_PERM << 8) | 65133ad798cSAdam Langley TCPOLEN_SACK_PERM); 65233ad798cSAdam Langley } 65333ad798cSAdam Langley 654bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_WSCALE & options)) { 65533ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 65633ad798cSAdam Langley (TCPOPT_WINDOW << 16) | 65733ad798cSAdam Langley (TCPOLEN_WINDOW << 8) | 65833ad798cSAdam Langley opts->ws); 65933ad798cSAdam Langley } 66033ad798cSAdam Langley 66133ad798cSAdam Langley if (unlikely(opts->num_sack_blocks)) { 66233ad798cSAdam Langley struct tcp_sack_block *sp = tp->rx_opt.dsack ? 66333ad798cSAdam Langley tp->duplicate_sack : tp->selective_acks; 66440efc6faSStephen Hemminger int this_sack; 66540efc6faSStephen Hemminger 66640efc6faSStephen Hemminger *ptr++ = htonl((TCPOPT_NOP << 24) | 66740efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 66840efc6faSStephen Hemminger (TCPOPT_SACK << 8) | 66933ad798cSAdam Langley (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * 67040efc6faSStephen Hemminger TCPOLEN_SACK_PERBLOCK))); 6712de979bdSStephen Hemminger 67233ad798cSAdam Langley for (this_sack = 0; this_sack < opts->num_sack_blocks; 67333ad798cSAdam Langley ++this_sack) { 67440efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].start_seq); 67540efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].end_seq); 67640efc6faSStephen Hemminger } 6772de979bdSStephen Hemminger 67840efc6faSStephen Hemminger tp->rx_opt.dsack = 0; 67940efc6faSStephen Hemminger } 6802100c8d2SYuchung Cheng 6812100c8d2SYuchung Cheng if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) { 6822100c8d2SYuchung Cheng struct tcp_fastopen_cookie *foc = opts->fastopen_cookie; 6837f9b838bSDaniel Lee u8 *p = (u8 *)ptr; 6847f9b838bSDaniel Lee u32 len; /* Fast Open option length */ 6852100c8d2SYuchung Cheng 6867f9b838bSDaniel Lee if (foc->exp) { 6877f9b838bSDaniel Lee len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len; 6887f9b838bSDaniel Lee *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) | 6892100c8d2SYuchung Cheng TCPOPT_FASTOPEN_MAGIC); 6907f9b838bSDaniel Lee p += TCPOLEN_EXP_FASTOPEN_BASE; 6917f9b838bSDaniel Lee } else { 6927f9b838bSDaniel Lee len = TCPOLEN_FASTOPEN_BASE + foc->len; 6937f9b838bSDaniel Lee *p++ = TCPOPT_FASTOPEN; 6947f9b838bSDaniel Lee *p++ = len; 6952100c8d2SYuchung Cheng } 6967f9b838bSDaniel Lee 6977f9b838bSDaniel Lee memcpy(p, foc->val, foc->len); 6987f9b838bSDaniel Lee if ((len & 3) == 2) { 6997f9b838bSDaniel Lee p[foc->len] = TCPOPT_NOP; 7007f9b838bSDaniel Lee p[foc->len + 1] = TCPOPT_NOP; 7017f9b838bSDaniel Lee } 7027f9b838bSDaniel Lee ptr += (len + 3) >> 2; 7032100c8d2SYuchung Cheng } 70460e2a778SUrsula Braun 70560e2a778SUrsula Braun smc_options_write(ptr, &options); 706eda7acddSPeter Krystad 707ea66758cSPaolo Abeni mptcp_options_write(th, ptr, tp, opts); 70860e2a778SUrsula Braun } 70960e2a778SUrsula Braun 71060e2a778SUrsula Braun static void smc_set_option(const struct tcp_sock *tp, 71160e2a778SUrsula Braun struct tcp_out_options *opts, 71260e2a778SUrsula Braun unsigned int *remaining) 71360e2a778SUrsula Braun { 71460e2a778SUrsula Braun #if IS_ENABLED(CONFIG_SMC) 71560e2a778SUrsula Braun if (static_branch_unlikely(&tcp_have_smc)) { 71660e2a778SUrsula Braun if (tp->syn_smc) { 71760e2a778SUrsula Braun if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) { 71860e2a778SUrsula Braun opts->options |= OPTION_SMC; 71960e2a778SUrsula Braun *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED; 72060e2a778SUrsula Braun } 72160e2a778SUrsula Braun } 72260e2a778SUrsula Braun } 72360e2a778SUrsula Braun #endif 72460e2a778SUrsula Braun } 72560e2a778SUrsula Braun 72660e2a778SUrsula Braun static void smc_set_option_cond(const struct tcp_sock *tp, 72760e2a778SUrsula Braun const struct inet_request_sock *ireq, 72860e2a778SUrsula Braun struct tcp_out_options *opts, 72960e2a778SUrsula Braun unsigned int *remaining) 73060e2a778SUrsula Braun { 73160e2a778SUrsula Braun #if IS_ENABLED(CONFIG_SMC) 73260e2a778SUrsula Braun if (static_branch_unlikely(&tcp_have_smc)) { 73360e2a778SUrsula Braun if (tp->syn_smc && ireq->smc_ok) { 73460e2a778SUrsula Braun if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) { 73560e2a778SUrsula Braun opts->options |= OPTION_SMC; 73660e2a778SUrsula Braun *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED; 73760e2a778SUrsula Braun } 73860e2a778SUrsula Braun } 73960e2a778SUrsula Braun } 74060e2a778SUrsula Braun #endif 74140efc6faSStephen Hemminger } 74240efc6faSStephen Hemminger 743cec37a6eSPeter Krystad static void mptcp_set_option_cond(const struct request_sock *req, 744cec37a6eSPeter Krystad struct tcp_out_options *opts, 745cec37a6eSPeter Krystad unsigned int *remaining) 746cec37a6eSPeter Krystad { 747cec37a6eSPeter Krystad if (rsk_is_mptcp(req)) { 748cec37a6eSPeter Krystad unsigned int size; 749cec37a6eSPeter Krystad 750cec37a6eSPeter Krystad if (mptcp_synack_options(req, &size, &opts->mptcp)) { 751cec37a6eSPeter Krystad if (*remaining >= size) { 752cec37a6eSPeter Krystad opts->options |= OPTION_MPTCP; 753cec37a6eSPeter Krystad *remaining -= size; 754cec37a6eSPeter Krystad } 755cec37a6eSPeter Krystad } 756cec37a6eSPeter Krystad } 757cec37a6eSPeter Krystad } 758cec37a6eSPeter Krystad 75967edfef7SAndi Kleen /* Compute TCP options for SYN packets. This is not the final 76067edfef7SAndi Kleen * network wire format yet. 76167edfef7SAndi Kleen */ 76295c96174SEric Dumazet static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb, 76333ad798cSAdam Langley struct tcp_out_options *opts, 764cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 765cf533ea5SEric Dumazet { 76633ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 76795c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 768783237e8SYuchung Cheng struct tcp_fastopen_request *fastopen = tp->fastopen_req; 76933ad798cSAdam Langley 7708c2320e8SEric Dumazet *md5 = NULL; 771cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 772921f9a0fSEric Dumazet if (static_branch_unlikely(&tcp_md5_needed) && 7736015c71eSEric Dumazet rcu_access_pointer(tp->md5sig_info)) { 77433ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 77533ad798cSAdam Langley if (*md5) { 77633ad798cSAdam Langley opts->options |= OPTION_MD5; 777bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 778cfb6eeb4SYOSHIFUJI Hideaki } 7798c2320e8SEric Dumazet } 780cfb6eeb4SYOSHIFUJI Hideaki #endif 78133ad798cSAdam Langley 78233ad798cSAdam Langley /* We always get an MSS option. The option bytes which will be seen in 78333ad798cSAdam Langley * normal data packets should timestamps be used, must be in the MSS 78433ad798cSAdam Langley * advertised. But we subtract them from tp->mss_cache so that 78533ad798cSAdam Langley * calculations in tcp_sendmsg are simpler etc. So account for this 78633ad798cSAdam Langley * fact here if necessary. If we don't do this correctly, as a 78733ad798cSAdam Langley * receiver we won't recognize data packets as being full sized when we 78833ad798cSAdam Langley * should, and thus we won't abide by the delayed ACK rules correctly. 78933ad798cSAdam Langley * SACKs don't matter, we never delay an ACK when we have any of those 79033ad798cSAdam Langley * going out. */ 79133ad798cSAdam Langley opts->mss = tcp_advertise_mss(sk); 792bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 79333ad798cSAdam Langley 7945d2ed052SEric Dumazet if (likely(sock_net(sk)->ipv4.sysctl_tcp_timestamps && !*md5)) { 79533ad798cSAdam Langley opts->options |= OPTION_TS; 7967faee5c0SEric Dumazet opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset; 79733ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 798bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 79933ad798cSAdam Langley } 8009bb37ef0SEric Dumazet if (likely(sock_net(sk)->ipv4.sysctl_tcp_window_scaling)) { 80133ad798cSAdam Langley opts->ws = tp->rx_opt.rcv_wscale; 80289e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 803bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 80433ad798cSAdam Langley } 805f9301034SEric Dumazet if (likely(sock_net(sk)->ipv4.sysctl_tcp_sack)) { 80633ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 807b32d1310SDavid S. Miller if (unlikely(!(OPTION_TS & opts->options))) 808bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 80933ad798cSAdam Langley } 81033ad798cSAdam Langley 811783237e8SYuchung Cheng if (fastopen && fastopen->cookie.len >= 0) { 8122646c831SDaniel Lee u32 need = fastopen->cookie.len; 8132646c831SDaniel Lee 8142646c831SDaniel Lee need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE : 8152646c831SDaniel Lee TCPOLEN_FASTOPEN_BASE; 816783237e8SYuchung Cheng need = (need + 3) & ~3U; /* Align to 32 bits */ 817783237e8SYuchung Cheng if (remaining >= need) { 818783237e8SYuchung Cheng opts->options |= OPTION_FAST_OPEN_COOKIE; 819783237e8SYuchung Cheng opts->fastopen_cookie = &fastopen->cookie; 820783237e8SYuchung Cheng remaining -= need; 821783237e8SYuchung Cheng tp->syn_fastopen = 1; 8222646c831SDaniel Lee tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0; 823783237e8SYuchung Cheng } 824783237e8SYuchung Cheng } 825bd0388aeSWilliam Allen Simpson 82660e2a778SUrsula Braun smc_set_option(tp, opts, &remaining); 82760e2a778SUrsula Braun 828cec37a6eSPeter Krystad if (sk_is_mptcp(sk)) { 829cec37a6eSPeter Krystad unsigned int size; 830cec37a6eSPeter Krystad 831cc7972eaSChristoph Paasch if (mptcp_syn_options(sk, skb, &size, &opts->mptcp)) { 832cec37a6eSPeter Krystad opts->options |= OPTION_MPTCP; 833cec37a6eSPeter Krystad remaining -= size; 834cec37a6eSPeter Krystad } 835cec37a6eSPeter Krystad } 836cec37a6eSPeter Krystad 837331fca43SMartin KaFai Lau bpf_skops_hdr_opt_len(sk, skb, NULL, NULL, 0, opts, &remaining); 838331fca43SMartin KaFai Lau 839bd0388aeSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 84033ad798cSAdam Langley } 84133ad798cSAdam Langley 84267edfef7SAndi Kleen /* Set up TCP options for SYN-ACKs. */ 84360e2a778SUrsula Braun static unsigned int tcp_synack_options(const struct sock *sk, 84460e2a778SUrsula Braun struct request_sock *req, 84595c96174SEric Dumazet unsigned int mss, struct sk_buff *skb, 84633ad798cSAdam Langley struct tcp_out_options *opts, 84780f03e27SEric Dumazet const struct tcp_md5sig_key *md5, 848e114e1e8SEric Dumazet struct tcp_fastopen_cookie *foc, 849331fca43SMartin KaFai Lau enum tcp_synack_type synack_type, 850331fca43SMartin KaFai Lau struct sk_buff *syn_skb) 8514957faadSWilliam Allen Simpson { 85233ad798cSAdam Langley struct inet_request_sock *ireq = inet_rsk(req); 85395c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 85433ad798cSAdam Langley 85533ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 85680f03e27SEric Dumazet if (md5) { 85733ad798cSAdam Langley opts->options |= OPTION_MD5; 8584957faadSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 8594957faadSWilliam Allen Simpson 8604957faadSWilliam Allen Simpson /* We can't fit any SACK blocks in a packet with MD5 + TS 8614957faadSWilliam Allen Simpson * options. There was discussion about disabling SACK 8624957faadSWilliam Allen Simpson * rather than TS in order to fit in better with old, 8634957faadSWilliam Allen Simpson * buggy kernels, but that was deemed to be unnecessary. 8644957faadSWilliam Allen Simpson */ 865e114e1e8SEric Dumazet if (synack_type != TCP_SYNACK_COOKIE) 866de213e5eSEric Dumazet ireq->tstamp_ok &= !ireq->sack_ok; 86733ad798cSAdam Langley } 86833ad798cSAdam Langley #endif 86933ad798cSAdam Langley 8704957faadSWilliam Allen Simpson /* We always send an MSS option. */ 87133ad798cSAdam Langley opts->mss = mss; 8724957faadSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 87333ad798cSAdam Langley 87433ad798cSAdam Langley if (likely(ireq->wscale_ok)) { 87533ad798cSAdam Langley opts->ws = ireq->rcv_wscale; 87689e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 8774957faadSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 87833ad798cSAdam Langley } 879de213e5eSEric Dumazet if (likely(ireq->tstamp_ok)) { 88033ad798cSAdam Langley opts->options |= OPTION_TS; 88195a22caeSFlorian Westphal opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off; 88233ad798cSAdam Langley opts->tsecr = req->ts_recent; 8834957faadSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 88433ad798cSAdam Langley } 88533ad798cSAdam Langley if (likely(ireq->sack_ok)) { 88633ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 887de213e5eSEric Dumazet if (unlikely(!ireq->tstamp_ok)) 8884957faadSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 88933ad798cSAdam Langley } 8907f9b838bSDaniel Lee if (foc != NULL && foc->len >= 0) { 8917f9b838bSDaniel Lee u32 need = foc->len; 8927f9b838bSDaniel Lee 8937f9b838bSDaniel Lee need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE : 8947f9b838bSDaniel Lee TCPOLEN_FASTOPEN_BASE; 8958336886fSJerry Chu need = (need + 3) & ~3U; /* Align to 32 bits */ 8968336886fSJerry Chu if (remaining >= need) { 8978336886fSJerry Chu opts->options |= OPTION_FAST_OPEN_COOKIE; 8988336886fSJerry Chu opts->fastopen_cookie = foc; 8998336886fSJerry Chu remaining -= need; 9008336886fSJerry Chu } 9018336886fSJerry Chu } 9024957faadSWilliam Allen Simpson 903cec37a6eSPeter Krystad mptcp_set_option_cond(req, opts, &remaining); 904cec37a6eSPeter Krystad 90560e2a778SUrsula Braun smc_set_option_cond(tcp_sk(sk), ireq, opts, &remaining); 90660e2a778SUrsula Braun 907331fca43SMartin KaFai Lau bpf_skops_hdr_opt_len((struct sock *)sk, skb, req, syn_skb, 908331fca43SMartin KaFai Lau synack_type, opts, &remaining); 909331fca43SMartin KaFai Lau 9104957faadSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 91133ad798cSAdam Langley } 91233ad798cSAdam Langley 91367edfef7SAndi Kleen /* Compute TCP options for ESTABLISHED sockets. This is not the 91467edfef7SAndi Kleen * final wire format yet. 91567edfef7SAndi Kleen */ 91695c96174SEric Dumazet static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb, 91733ad798cSAdam Langley struct tcp_out_options *opts, 918cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 919cf533ea5SEric Dumazet { 92033ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 92195c96174SEric Dumazet unsigned int size = 0; 922cabeccbdSIlpo Järvinen unsigned int eff_sacks; 92333ad798cSAdam Langley 9245843ef42SAndi Kleen opts->options = 0; 9255843ef42SAndi Kleen 9268c2320e8SEric Dumazet *md5 = NULL; 92733ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 928921f9a0fSEric Dumazet if (static_branch_unlikely(&tcp_md5_needed) && 9296015c71eSEric Dumazet rcu_access_pointer(tp->md5sig_info)) { 93033ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 9318c2320e8SEric Dumazet if (*md5) { 93233ad798cSAdam Langley opts->options |= OPTION_MD5; 93333ad798cSAdam Langley size += TCPOLEN_MD5SIG_ALIGNED; 93433ad798cSAdam Langley } 9358c2320e8SEric Dumazet } 93633ad798cSAdam Langley #endif 93733ad798cSAdam Langley 93833ad798cSAdam Langley if (likely(tp->rx_opt.tstamp_ok)) { 93933ad798cSAdam Langley opts->options |= OPTION_TS; 9407faee5c0SEric Dumazet opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0; 94133ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 94233ad798cSAdam Langley size += TCPOLEN_TSTAMP_ALIGNED; 94333ad798cSAdam Langley } 94433ad798cSAdam Langley 945cec37a6eSPeter Krystad /* MPTCP options have precedence over SACK for the limited TCP 946cec37a6eSPeter Krystad * option space because a MPTCP connection would be forced to 947cec37a6eSPeter Krystad * fall back to regular TCP if a required multipath option is 948cec37a6eSPeter Krystad * missing. SACK still gets a chance to use whatever space is 949cec37a6eSPeter Krystad * left. 950cec37a6eSPeter Krystad */ 951cec37a6eSPeter Krystad if (sk_is_mptcp(sk)) { 952cec37a6eSPeter Krystad unsigned int remaining = MAX_TCP_OPTION_SPACE - size; 953cec37a6eSPeter Krystad unsigned int opt_size = 0; 954cec37a6eSPeter Krystad 955cec37a6eSPeter Krystad if (mptcp_established_options(sk, skb, &opt_size, remaining, 956cec37a6eSPeter Krystad &opts->mptcp)) { 957cec37a6eSPeter Krystad opts->options |= OPTION_MPTCP; 958cec37a6eSPeter Krystad size += opt_size; 959cec37a6eSPeter Krystad } 960cec37a6eSPeter Krystad } 961cec37a6eSPeter Krystad 962cabeccbdSIlpo Järvinen eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; 963cabeccbdSIlpo Järvinen if (unlikely(eff_sacks)) { 96495c96174SEric Dumazet const unsigned int remaining = MAX_TCP_OPTION_SPACE - size; 9659cfcca23SMat Martineau if (unlikely(remaining < TCPOLEN_SACK_BASE_ALIGNED + 9669cfcca23SMat Martineau TCPOLEN_SACK_PERBLOCK)) 9679cfcca23SMat Martineau return size; 9689cfcca23SMat Martineau 96933ad798cSAdam Langley opts->num_sack_blocks = 97095c96174SEric Dumazet min_t(unsigned int, eff_sacks, 97133ad798cSAdam Langley (remaining - TCPOLEN_SACK_BASE_ALIGNED) / 97233ad798cSAdam Langley TCPOLEN_SACK_PERBLOCK); 9739cfcca23SMat Martineau 97433ad798cSAdam Langley size += TCPOLEN_SACK_BASE_ALIGNED + 97533ad798cSAdam Langley opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; 97633ad798cSAdam Langley } 97733ad798cSAdam Langley 978331fca43SMartin KaFai Lau if (unlikely(BPF_SOCK_OPS_TEST_FLAG(tp, 979331fca43SMartin KaFai Lau BPF_SOCK_OPS_WRITE_HDR_OPT_CB_FLAG))) { 980331fca43SMartin KaFai Lau unsigned int remaining = MAX_TCP_OPTION_SPACE - size; 981331fca43SMartin KaFai Lau 982331fca43SMartin KaFai Lau bpf_skops_hdr_opt_len(sk, skb, NULL, NULL, 0, opts, &remaining); 983331fca43SMartin KaFai Lau 984331fca43SMartin KaFai Lau size = MAX_TCP_OPTION_SPACE - remaining; 985331fca43SMartin KaFai Lau } 986331fca43SMartin KaFai Lau 98733ad798cSAdam Langley return size; 98840efc6faSStephen Hemminger } 9891da177e4SLinus Torvalds 99046d3ceabSEric Dumazet 99146d3ceabSEric Dumazet /* TCP SMALL QUEUES (TSQ) 99246d3ceabSEric Dumazet * 99346d3ceabSEric Dumazet * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev) 99446d3ceabSEric Dumazet * to reduce RTT and bufferbloat. 99546d3ceabSEric Dumazet * We do this using a special skb destructor (tcp_wfree). 99646d3ceabSEric Dumazet * 99746d3ceabSEric Dumazet * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb 99846d3ceabSEric Dumazet * needs to be reallocated in a driver. 9998e3bff96Sstephen hemminger * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc 100046d3ceabSEric Dumazet * 100146d3ceabSEric Dumazet * Since transmit from skb destructor is forbidden, we use a tasklet 100246d3ceabSEric Dumazet * to process all sockets that eventually need to send more skbs. 100346d3ceabSEric Dumazet * We use one tasklet per cpu, with its own queue of sockets. 100446d3ceabSEric Dumazet */ 100546d3ceabSEric Dumazet struct tsq_tasklet { 100646d3ceabSEric Dumazet struct tasklet_struct tasklet; 100746d3ceabSEric Dumazet struct list_head head; /* queue of tcp sockets */ 100846d3ceabSEric Dumazet }; 100946d3ceabSEric Dumazet static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet); 101046d3ceabSEric Dumazet 101173a6bab5SEric Dumazet static void tcp_tsq_write(struct sock *sk) 10126f458dfbSEric Dumazet { 10136f458dfbSEric Dumazet if ((1 << sk->sk_state) & 10146f458dfbSEric Dumazet (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING | 1015f9616c35SEric Dumazet TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) { 1016f9616c35SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 1017f9616c35SEric Dumazet 1018f9616c35SEric Dumazet if (tp->lost_out > tp->retrans_out && 101940570375SEric Dumazet tcp_snd_cwnd(tp) > tcp_packets_in_flight(tp)) { 10203a91d29fSKoichiro Den tcp_mstamp_refresh(tp); 1021f9616c35SEric Dumazet tcp_xmit_retransmit_queue(sk); 10223a91d29fSKoichiro Den } 1023f9616c35SEric Dumazet 1024f9616c35SEric Dumazet tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle, 1025bf06200eSJohn Ogness 0, GFP_ATOMIC); 10266f458dfbSEric Dumazet } 1027f9616c35SEric Dumazet } 102873a6bab5SEric Dumazet 102973a6bab5SEric Dumazet static void tcp_tsq_handler(struct sock *sk) 103073a6bab5SEric Dumazet { 103173a6bab5SEric Dumazet bh_lock_sock(sk); 103273a6bab5SEric Dumazet if (!sock_owned_by_user(sk)) 103373a6bab5SEric Dumazet tcp_tsq_write(sk); 103473a6bab5SEric Dumazet else if (!test_and_set_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags)) 103573a6bab5SEric Dumazet sock_hold(sk); 103673a6bab5SEric Dumazet bh_unlock_sock(sk); 103773a6bab5SEric Dumazet } 103846d3ceabSEric Dumazet /* 10398e3bff96Sstephen hemminger * One tasklet per cpu tries to send more skbs. 104046d3ceabSEric Dumazet * We run in tasklet context but need to disable irqs when 10418e3bff96Sstephen hemminger * transferring tsq->head because tcp_wfree() might 104246d3ceabSEric Dumazet * interrupt us (non NAPI drivers) 104346d3ceabSEric Dumazet */ 1044c6533ca8SAllen Pais static void tcp_tasklet_func(struct tasklet_struct *t) 104546d3ceabSEric Dumazet { 1046c6533ca8SAllen Pais struct tsq_tasklet *tsq = from_tasklet(tsq, t, tasklet); 104746d3ceabSEric Dumazet LIST_HEAD(list); 104846d3ceabSEric Dumazet unsigned long flags; 104946d3ceabSEric Dumazet struct list_head *q, *n; 105046d3ceabSEric Dumazet struct tcp_sock *tp; 105146d3ceabSEric Dumazet struct sock *sk; 105246d3ceabSEric Dumazet 105346d3ceabSEric Dumazet local_irq_save(flags); 105446d3ceabSEric Dumazet list_splice_init(&tsq->head, &list); 105546d3ceabSEric Dumazet local_irq_restore(flags); 105646d3ceabSEric Dumazet 105746d3ceabSEric Dumazet list_for_each_safe(q, n, &list) { 105846d3ceabSEric Dumazet tp = list_entry(q, struct tcp_sock, tsq_node); 105946d3ceabSEric Dumazet list_del(&tp->tsq_node); 106046d3ceabSEric Dumazet 106146d3ceabSEric Dumazet sk = (struct sock *)tp; 10620a9648f1SEric Dumazet smp_mb__before_atomic(); 10637aa5470cSEric Dumazet clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags); 10647aa5470cSEric Dumazet 10656f458dfbSEric Dumazet tcp_tsq_handler(sk); 106646d3ceabSEric Dumazet sk_free(sk); 106746d3ceabSEric Dumazet } 106846d3ceabSEric Dumazet } 106946d3ceabSEric Dumazet 107040fc3423SEric Dumazet #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \ 107140fc3423SEric Dumazet TCPF_WRITE_TIMER_DEFERRED | \ 107240fc3423SEric Dumazet TCPF_DELACK_TIMER_DEFERRED | \ 107340fc3423SEric Dumazet TCPF_MTU_REDUCED_DEFERRED) 107446d3ceabSEric Dumazet /** 107546d3ceabSEric Dumazet * tcp_release_cb - tcp release_sock() callback 107646d3ceabSEric Dumazet * @sk: socket 107746d3ceabSEric Dumazet * 107846d3ceabSEric Dumazet * called from release_sock() to perform protocol dependent 107946d3ceabSEric Dumazet * actions before socket release. 108046d3ceabSEric Dumazet */ 108146d3ceabSEric Dumazet void tcp_release_cb(struct sock *sk) 108246d3ceabSEric Dumazet { 10836f458dfbSEric Dumazet unsigned long flags, nflags; 108446d3ceabSEric Dumazet 10856f458dfbSEric Dumazet /* perform an atomic operation only if at least one flag is set */ 10866f458dfbSEric Dumazet do { 10877aa5470cSEric Dumazet flags = sk->sk_tsq_flags; 10886f458dfbSEric Dumazet if (!(flags & TCP_DEFERRED_ALL)) 10896f458dfbSEric Dumazet return; 10906f458dfbSEric Dumazet nflags = flags & ~TCP_DEFERRED_ALL; 10917aa5470cSEric Dumazet } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags); 10926f458dfbSEric Dumazet 109373a6bab5SEric Dumazet if (flags & TCPF_TSQ_DEFERRED) { 109473a6bab5SEric Dumazet tcp_tsq_write(sk); 109573a6bab5SEric Dumazet __sock_put(sk); 109673a6bab5SEric Dumazet } 1097c3f9b018SEric Dumazet /* Here begins the tricky part : 1098c3f9b018SEric Dumazet * We are called from release_sock() with : 1099c3f9b018SEric Dumazet * 1) BH disabled 1100c3f9b018SEric Dumazet * 2) sk_lock.slock spinlock held 1101c3f9b018SEric Dumazet * 3) socket owned by us (sk->sk_lock.owned == 1) 1102c3f9b018SEric Dumazet * 1103c3f9b018SEric Dumazet * But following code is meant to be called from BH handlers, 1104c3f9b018SEric Dumazet * so we should keep BH disabled, but early release socket ownership 1105c3f9b018SEric Dumazet */ 1106c3f9b018SEric Dumazet sock_release_ownership(sk); 1107c3f9b018SEric Dumazet 110840fc3423SEric Dumazet if (flags & TCPF_WRITE_TIMER_DEFERRED) { 11096f458dfbSEric Dumazet tcp_write_timer_handler(sk); 1110144d56e9SEric Dumazet __sock_put(sk); 1111144d56e9SEric Dumazet } 111240fc3423SEric Dumazet if (flags & TCPF_DELACK_TIMER_DEFERRED) { 11136f458dfbSEric Dumazet tcp_delack_timer_handler(sk); 1114144d56e9SEric Dumazet __sock_put(sk); 1115144d56e9SEric Dumazet } 111640fc3423SEric Dumazet if (flags & TCPF_MTU_REDUCED_DEFERRED) { 11174fab9071SNeal Cardwell inet_csk(sk)->icsk_af_ops->mtu_reduced(sk); 1118144d56e9SEric Dumazet __sock_put(sk); 1119144d56e9SEric Dumazet } 112046d3ceabSEric Dumazet } 112146d3ceabSEric Dumazet EXPORT_SYMBOL(tcp_release_cb); 112246d3ceabSEric Dumazet 112346d3ceabSEric Dumazet void __init tcp_tasklet_init(void) 112446d3ceabSEric Dumazet { 112546d3ceabSEric Dumazet int i; 112646d3ceabSEric Dumazet 112746d3ceabSEric Dumazet for_each_possible_cpu(i) { 112846d3ceabSEric Dumazet struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i); 112946d3ceabSEric Dumazet 113046d3ceabSEric Dumazet INIT_LIST_HEAD(&tsq->head); 1131c6533ca8SAllen Pais tasklet_setup(&tsq->tasklet, tcp_tasklet_func); 113246d3ceabSEric Dumazet } 113346d3ceabSEric Dumazet } 113446d3ceabSEric Dumazet 113546d3ceabSEric Dumazet /* 113646d3ceabSEric Dumazet * Write buffer destructor automatically called from kfree_skb. 11378e3bff96Sstephen hemminger * We can't xmit new skbs from this context, as we might already 113846d3ceabSEric Dumazet * hold qdisc lock. 113946d3ceabSEric Dumazet */ 1140d6a4a104SEric Dumazet void tcp_wfree(struct sk_buff *skb) 114146d3ceabSEric Dumazet { 114246d3ceabSEric Dumazet struct sock *sk = skb->sk; 114346d3ceabSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 1144408f0a6cSEric Dumazet unsigned long flags, nval, oval; 11459b462d02SEric Dumazet 11469b462d02SEric Dumazet /* Keep one reference on sk_wmem_alloc. 11479b462d02SEric Dumazet * Will be released by sk_free() from here or tcp_tasklet_func() 11489b462d02SEric Dumazet */ 114914afee4bSReshetova, Elena WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc)); 11509b462d02SEric Dumazet 11519b462d02SEric Dumazet /* If this softirq is serviced by ksoftirqd, we are likely under stress. 11529b462d02SEric Dumazet * Wait until our queues (qdisc + devices) are drained. 11539b462d02SEric Dumazet * This gives : 11549b462d02SEric Dumazet * - less callbacks to tcp_write_xmit(), reducing stress (batches) 11559b462d02SEric Dumazet * - chance for incoming ACK (processed by another cpu maybe) 11569b462d02SEric Dumazet * to migrate this flow (skb->ooo_okay will be eventually set) 11579b462d02SEric Dumazet */ 115814afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current) 11599b462d02SEric Dumazet goto out; 116046d3ceabSEric Dumazet 11617aa5470cSEric Dumazet for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) { 116246d3ceabSEric Dumazet struct tsq_tasklet *tsq; 1163a9b204d1SEric Dumazet bool empty; 116446d3ceabSEric Dumazet 1165408f0a6cSEric Dumazet if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED)) 1166408f0a6cSEric Dumazet goto out; 1167408f0a6cSEric Dumazet 116873a6bab5SEric Dumazet nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED; 11697aa5470cSEric Dumazet nval = cmpxchg(&sk->sk_tsq_flags, oval, nval); 1170408f0a6cSEric Dumazet if (nval != oval) 1171408f0a6cSEric Dumazet continue; 1172408f0a6cSEric Dumazet 117346d3ceabSEric Dumazet /* queue this socket to tasklet queue */ 117446d3ceabSEric Dumazet local_irq_save(flags); 1175903ceff7SChristoph Lameter tsq = this_cpu_ptr(&tsq_tasklet); 1176a9b204d1SEric Dumazet empty = list_empty(&tsq->head); 117746d3ceabSEric Dumazet list_add(&tp->tsq_node, &tsq->head); 1178a9b204d1SEric Dumazet if (empty) 117946d3ceabSEric Dumazet tasklet_schedule(&tsq->tasklet); 118046d3ceabSEric Dumazet local_irq_restore(flags); 11819b462d02SEric Dumazet return; 118246d3ceabSEric Dumazet } 11839b462d02SEric Dumazet out: 11849b462d02SEric Dumazet sk_free(sk); 118546d3ceabSEric Dumazet } 118646d3ceabSEric Dumazet 118773a6bab5SEric Dumazet /* Note: Called under soft irq. 118873a6bab5SEric Dumazet * We can call TCP stack right away, unless socket is owned by user. 1189218af599SEric Dumazet */ 1190218af599SEric Dumazet enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer) 1191218af599SEric Dumazet { 1192218af599SEric Dumazet struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer); 1193218af599SEric Dumazet struct sock *sk = (struct sock *)tp; 1194218af599SEric Dumazet 119573a6bab5SEric Dumazet tcp_tsq_handler(sk); 119673a6bab5SEric Dumazet sock_put(sk); 1197218af599SEric Dumazet 1198218af599SEric Dumazet return HRTIMER_NORESTART; 1199218af599SEric Dumazet } 1200218af599SEric Dumazet 1201a7a25630SEric Dumazet static void tcp_update_skb_after_send(struct sock *sk, struct sk_buff *skb, 1202a7a25630SEric Dumazet u64 prior_wstamp) 1203e2080072SEric Dumazet { 1204ab408b6dSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 1205ab408b6dSEric Dumazet 1206ab408b6dSEric Dumazet if (sk->sk_pacing_status != SK_PACING_NONE) { 120776a9ebe8SEric Dumazet unsigned long rate = sk->sk_pacing_rate; 1208ab408b6dSEric Dumazet 1209ab408b6dSEric Dumazet /* Original sch_fq does not pace first 10 MSS 1210ab408b6dSEric Dumazet * Note that tp->data_segs_out overflows after 2^32 packets, 1211ab408b6dSEric Dumazet * this is a minor annoyance. 1212ab408b6dSEric Dumazet */ 121376a9ebe8SEric Dumazet if (rate != ~0UL && rate && tp->data_segs_out >= 10) { 1214a7a25630SEric Dumazet u64 len_ns = div64_ul((u64)skb->len * NSEC_PER_SEC, rate); 1215a7a25630SEric Dumazet u64 credit = tp->tcp_wstamp_ns - prior_wstamp; 1216a7a25630SEric Dumazet 1217a7a25630SEric Dumazet /* take into account OS jitter */ 1218a7a25630SEric Dumazet len_ns -= min_t(u64, len_ns / 2, credit); 1219a7a25630SEric Dumazet tp->tcp_wstamp_ns += len_ns; 1220ab408b6dSEric Dumazet } 1221ab408b6dSEric Dumazet } 1222e2080072SEric Dumazet list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue); 1223e2080072SEric Dumazet } 1224e2080072SEric Dumazet 122505e22e83SEric Dumazet INDIRECT_CALLABLE_DECLARE(int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl)); 122605e22e83SEric Dumazet INDIRECT_CALLABLE_DECLARE(int inet6_csk_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl)); 1227dd2e0b86SEric Dumazet INDIRECT_CALLABLE_DECLARE(void tcp_v4_send_check(struct sock *sk, struct sk_buff *skb)); 122805e22e83SEric Dumazet 12291da177e4SLinus Torvalds /* This routine actually transmits TCP packets queued in by 12301da177e4SLinus Torvalds * tcp_do_sendmsg(). This is used by both the initial 12311da177e4SLinus Torvalds * transmission and possible later retransmissions. 12321da177e4SLinus Torvalds * All SKB's seen here are completely headerless. It is our 12331da177e4SLinus Torvalds * job to build the TCP header, and pass the packet down to 12341da177e4SLinus Torvalds * IP so it can do the same plus pass the packet off to the 12351da177e4SLinus Torvalds * device. 12361da177e4SLinus Torvalds * 12371da177e4SLinus Torvalds * We are working here with either a clone of the original 12381da177e4SLinus Torvalds * SKB, or a fresh unique copy made by the retransmit engine. 12391da177e4SLinus Torvalds */ 12402987babbSYuchung Cheng static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, 12412987babbSYuchung Cheng int clone_it, gfp_t gfp_mask, u32 rcv_nxt) 12421da177e4SLinus Torvalds { 12436687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 1244dfb4b9dcSDavid S. Miller struct inet_sock *inet; 1245dfb4b9dcSDavid S. Miller struct tcp_sock *tp; 1246dfb4b9dcSDavid S. Miller struct tcp_skb_cb *tcb; 124733ad798cSAdam Langley struct tcp_out_options opts; 124895c96174SEric Dumazet unsigned int tcp_options_size, tcp_header_size; 12498c72c65bSEric Dumazet struct sk_buff *oskb = NULL; 1250cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 12511da177e4SLinus Torvalds struct tcphdr *th; 1252a7a25630SEric Dumazet u64 prior_wstamp; 12531da177e4SLinus Torvalds int err; 12541da177e4SLinus Torvalds 1255dfb4b9dcSDavid S. Miller BUG_ON(!skb || !tcp_skb_pcount(skb)); 12566f094b9eSLawrence Brakmo tp = tcp_sk(sk); 12577f12422cSYuchung Cheng prior_wstamp = tp->tcp_wstamp_ns; 12587f12422cSYuchung Cheng tp->tcp_wstamp_ns = max(tp->tcp_wstamp_ns, tp->tcp_clock_cache); 1259a1ac9c8aSMartin KaFai Lau skb_set_delivery_time(skb, tp->tcp_wstamp_ns, true); 1260ccdbb6e9SEric Dumazet if (clone_it) { 12618c72c65bSEric Dumazet oskb = skb; 1262e2080072SEric Dumazet 1263e2080072SEric Dumazet tcp_skb_tsorted_save(oskb) { 1264e2080072SEric Dumazet if (unlikely(skb_cloned(oskb))) 1265e2080072SEric Dumazet skb = pskb_copy(oskb, gfp_mask); 1266dfb4b9dcSDavid S. Miller else 1267e2080072SEric Dumazet skb = skb_clone(oskb, gfp_mask); 1268e2080072SEric Dumazet } tcp_skb_tsorted_restore(oskb); 1269e2080072SEric Dumazet 1270dfb4b9dcSDavid S. Miller if (unlikely(!skb)) 1271dfb4b9dcSDavid S. Miller return -ENOBUFS; 1272b738a185SEric Dumazet /* retransmit skbs might have a non zero value in skb->dev 1273b738a185SEric Dumazet * because skb->dev is aliased with skb->rbnode.rb_left 1274b738a185SEric Dumazet */ 1275b738a185SEric Dumazet skb->dev = NULL; 1276dfb4b9dcSDavid S. Miller } 12775f6188a8SEric Dumazet 1278dfb4b9dcSDavid S. Miller inet = inet_sk(sk); 1279dfb4b9dcSDavid S. Miller tcb = TCP_SKB_CB(skb); 128033ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 12811da177e4SLinus Torvalds 1282051ba674SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) { 128333ad798cSAdam Langley tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); 1284051ba674SEric Dumazet } else { 128533ad798cSAdam Langley tcp_options_size = tcp_established_options(sk, skb, &opts, 128633ad798cSAdam Langley &md5); 1287051ba674SEric Dumazet /* Force a PSH flag on all (GSO) packets to expedite GRO flush 1288051ba674SEric Dumazet * at receiver : This slightly improve GRO performance. 1289051ba674SEric Dumazet * Note that we do not force the PSH flag for non GSO packets, 1290051ba674SEric Dumazet * because they might be sent under high congestion events, 1291051ba674SEric Dumazet * and in this case it is better to delay the delivery of 1-MSS 1292051ba674SEric Dumazet * packets and thus the corresponding ACK packet that would 1293051ba674SEric Dumazet * release the following packet. 1294051ba674SEric Dumazet */ 1295051ba674SEric Dumazet if (tcp_skb_pcount(skb) > 1) 1296051ba674SEric Dumazet tcb->tcp_flags |= TCPHDR_PSH; 1297051ba674SEric Dumazet } 129833ad798cSAdam Langley tcp_header_size = tcp_options_size + sizeof(struct tcphdr); 12991da177e4SLinus Torvalds 1300547669d4SEric Dumazet /* if no packet is in qdisc/device queue, then allow XPS to select 1301b2532eb9SEric Dumazet * another queue. We can be called from tcp_tsq_handler() 130273a6bab5SEric Dumazet * which holds one reference to sk. 1303b2532eb9SEric Dumazet * 1304b2532eb9SEric Dumazet * TODO: Ideally, in-flight pure ACK packets should not matter here. 1305b2532eb9SEric Dumazet * One way to get this would be to set skb->truesize = 2 on them. 1306547669d4SEric Dumazet */ 1307b2532eb9SEric Dumazet skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1); 13081da177e4SLinus Torvalds 130938ab52e8SEric Dumazet /* If we had to use memory reserve to allocate this skb, 131038ab52e8SEric Dumazet * this might cause drops if packet is looped back : 131138ab52e8SEric Dumazet * Other socket might not have SOCK_MEMALLOC. 131238ab52e8SEric Dumazet * Packets not looped back do not care about pfmemalloc. 131338ab52e8SEric Dumazet */ 131438ab52e8SEric Dumazet skb->pfmemalloc = 0; 131538ab52e8SEric Dumazet 1316aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 1317aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 131846d3ceabSEric Dumazet 131946d3ceabSEric Dumazet skb_orphan(skb); 132046d3ceabSEric Dumazet skb->sk = sk; 13211d2077acSEric Dumazet skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree; 132214afee4bSReshetova, Elena refcount_add(skb->truesize, &sk->sk_wmem_alloc); 13231da177e4SLinus Torvalds 1324c3a2e837SJulian Anastasov skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm); 1325c3a2e837SJulian Anastasov 13261da177e4SLinus Torvalds /* Build TCP header and checksum it. */ 1327ea1627c2SEric Dumazet th = (struct tcphdr *)skb->data; 1328c720c7e8SEric Dumazet th->source = inet->inet_sport; 1329c720c7e8SEric Dumazet th->dest = inet->inet_dport; 13301da177e4SLinus Torvalds th->seq = htonl(tcb->seq); 13312987babbSYuchung Cheng th->ack_seq = htonl(rcv_nxt); 1332df7a3b07SAl Viro *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | 13334de075e0SEric Dumazet tcb->tcp_flags); 1334dfb4b9dcSDavid S. Miller 13351da177e4SLinus Torvalds th->check = 0; 13361da177e4SLinus Torvalds th->urg_ptr = 0; 13371da177e4SLinus Torvalds 133833f5f57eSIlpo Järvinen /* The urg_mode check is necessary during a below snd_una win probe */ 13397691367dSHerbert Xu if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { 13407691367dSHerbert Xu if (before(tp->snd_up, tcb->seq + 0x10000)) { 13411da177e4SLinus Torvalds th->urg_ptr = htons(tp->snd_up - tcb->seq); 13421da177e4SLinus Torvalds th->urg = 1; 13437691367dSHerbert Xu } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { 13440eae88f3SEric Dumazet th->urg_ptr = htons(0xFFFF); 13457691367dSHerbert Xu th->urg = 1; 13467691367dSHerbert Xu } 13471da177e4SLinus Torvalds } 13481da177e4SLinus Torvalds 134951466a75SEric Dumazet skb_shinfo(skb)->gso_type = sk->sk_gso_type; 1350ea1627c2SEric Dumazet if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) { 1351ea1627c2SEric Dumazet th->window = htons(tcp_select_window(sk)); 1352ea1627c2SEric Dumazet tcp_ecn_send(sk, skb, th, tcp_header_size); 1353ea1627c2SEric Dumazet } else { 1354ea1627c2SEric Dumazet /* RFC1323: The window in SYN & SYN/ACK segments 1355ea1627c2SEric Dumazet * is never scaled. 1356ea1627c2SEric Dumazet */ 1357ea1627c2SEric Dumazet th->window = htons(min(tp->rcv_wnd, 65535U)); 1358ea1627c2SEric Dumazet } 1359fa3fe2b1SFlorian Westphal 1360ea66758cSPaolo Abeni tcp_options_write(th, tp, &opts); 1361fa3fe2b1SFlorian Westphal 1362cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 1363cfb6eeb4SYOSHIFUJI Hideaki /* Calculate the MD5 hash, as we have all we need now */ 1364cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 1365aba54656SEric Dumazet sk_gso_disable(sk); 1366bd0388aeSWilliam Allen Simpson tp->af_specific->calc_md5_hash(opts.hash_location, 136739f8e58eSEric Dumazet md5, sk, skb); 1368cfb6eeb4SYOSHIFUJI Hideaki } 1369cfb6eeb4SYOSHIFUJI Hideaki #endif 1370cfb6eeb4SYOSHIFUJI Hideaki 1371331fca43SMartin KaFai Lau /* BPF prog is the last one writing header option */ 1372331fca43SMartin KaFai Lau bpf_skops_write_hdr_opt(sk, skb, NULL, NULL, 0, &opts); 1373331fca43SMartin KaFai Lau 1374dd2e0b86SEric Dumazet INDIRECT_CALL_INET(icsk->icsk_af_ops->send_check, 1375dd2e0b86SEric Dumazet tcp_v6_send_check, tcp_v4_send_check, 1376dd2e0b86SEric Dumazet sk, skb); 13771da177e4SLinus Torvalds 13784de075e0SEric Dumazet if (likely(tcb->tcp_flags & TCPHDR_ACK)) 137927cde44aSYuchung Cheng tcp_event_ack_sent(sk, tcp_skb_pcount(skb), rcv_nxt); 13801da177e4SLinus Torvalds 1381a44d6eacSMartin KaFai Lau if (skb->len != tcp_header_size) { 1382cf533ea5SEric Dumazet tcp_event_data_sent(tp, sk); 1383a44d6eacSMartin KaFai Lau tp->data_segs_out += tcp_skb_pcount(skb); 1384ba113c3aSWei Wang tp->bytes_sent += skb->len - tcp_header_size; 1385a44d6eacSMartin KaFai Lau } 13861da177e4SLinus Torvalds 1387bd37a088SWei Yongjun if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) 1388aa2ea058STom Herbert TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, 1389aa2ea058STom Herbert tcp_skb_pcount(skb)); 13901da177e4SLinus Torvalds 13912efd055cSMarcelo Ricardo Leitner tp->segs_out += tcp_skb_pcount(skb); 13920ae5b43dSYuchung Cheng skb_set_hash_from_sk(skb, sk); 1393f69ad292SEric Dumazet /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */ 1394cd7d8498SEric Dumazet skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb); 1395f69ad292SEric Dumazet skb_shinfo(skb)->gso_size = tcp_skb_mss(skb); 1396cd7d8498SEric Dumazet 1397d3edd06eSEric Dumazet /* Leave earliest departure time in skb->tstamp (skb->skb_mstamp_ns) */ 1398971f10ecSEric Dumazet 1399971f10ecSEric Dumazet /* Cleanup our debris for IP stacks */ 1400971f10ecSEric Dumazet memset(skb->cb, 0, max(sizeof(struct inet_skb_parm), 1401971f10ecSEric Dumazet sizeof(struct inet6_skb_parm))); 1402971f10ecSEric Dumazet 1403a842fe14SEric Dumazet tcp_add_tx_delay(skb, tp); 1404a842fe14SEric Dumazet 140505e22e83SEric Dumazet err = INDIRECT_CALL_INET(icsk->icsk_af_ops->queue_xmit, 140605e22e83SEric Dumazet inet6_csk_xmit, ip_queue_xmit, 140705e22e83SEric Dumazet sk, skb, &inet->cork.fl); 14087faee5c0SEric Dumazet 14098c72c65bSEric Dumazet if (unlikely(err > 0)) { 14105ee2c941SChristoph Paasch tcp_enter_cwr(sk); 14118c72c65bSEric Dumazet err = net_xmit_eval(err); 14128c72c65bSEric Dumazet } 1413fc225799SEric Dumazet if (!err && oskb) { 1414a7a25630SEric Dumazet tcp_update_skb_after_send(sk, oskb, prior_wstamp); 1415fc225799SEric Dumazet tcp_rate_skb_sent(sk, oskb); 1416fc225799SEric Dumazet } 14178c72c65bSEric Dumazet return err; 14181da177e4SLinus Torvalds } 14191da177e4SLinus Torvalds 14202987babbSYuchung Cheng static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, 14212987babbSYuchung Cheng gfp_t gfp_mask) 14222987babbSYuchung Cheng { 14232987babbSYuchung Cheng return __tcp_transmit_skb(sk, skb, clone_it, gfp_mask, 14242987babbSYuchung Cheng tcp_sk(sk)->rcv_nxt); 14252987babbSYuchung Cheng } 14262987babbSYuchung Cheng 142767edfef7SAndi Kleen /* This routine just queues the buffer for sending. 14281da177e4SLinus Torvalds * 14291da177e4SLinus Torvalds * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, 14301da177e4SLinus Torvalds * otherwise socket can stall. 14311da177e4SLinus Torvalds */ 14321da177e4SLinus Torvalds static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) 14331da177e4SLinus Torvalds { 14341da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 14351da177e4SLinus Torvalds 14361da177e4SLinus Torvalds /* Advance write_seq and place onto the write_queue. */ 14370f317464SEric Dumazet WRITE_ONCE(tp->write_seq, TCP_SKB_CB(skb)->end_seq); 1438f4a775d1SEric Dumazet __skb_header_release(skb); 1439fe067e8aSDavid S. Miller tcp_add_write_queue_tail(sk, skb); 1440ab4e846aSEric Dumazet sk_wmem_queued_add(sk, skb->truesize); 14413ab224beSHideo Aoki sk_mem_charge(sk, skb->truesize); 14421da177e4SLinus Torvalds } 14431da177e4SLinus Torvalds 144467edfef7SAndi Kleen /* Initialize TSO segments for a packet. */ 14455bbb432cSEric Dumazet static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now) 1446f6302d1dSDavid S. Miller { 14474a64fd6cSEric Dumazet if (skb->len <= mss_now) { 1448f6302d1dSDavid S. Miller /* Avoid the costly divide in the normal 1449f6302d1dSDavid S. Miller * non-TSO case. 1450f6302d1dSDavid S. Miller */ 1451cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, 1); 1452f69ad292SEric Dumazet TCP_SKB_CB(skb)->tcp_gso_size = 0; 1453f6302d1dSDavid S. Miller } else { 1454cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now)); 1455f69ad292SEric Dumazet TCP_SKB_CB(skb)->tcp_gso_size = mss_now; 14561da177e4SLinus Torvalds } 14571da177e4SLinus Torvalds } 14581da177e4SLinus Torvalds 1459797108d1SIlpo Järvinen /* Pcount in the middle of the write queue got changed, we need to do various 1460797108d1SIlpo Järvinen * tweaks to fix counters 1461797108d1SIlpo Järvinen */ 1462cf533ea5SEric Dumazet static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) 1463797108d1SIlpo Järvinen { 1464797108d1SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1465797108d1SIlpo Järvinen 1466797108d1SIlpo Järvinen tp->packets_out -= decr; 1467797108d1SIlpo Järvinen 1468797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 1469797108d1SIlpo Järvinen tp->sacked_out -= decr; 1470797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 1471797108d1SIlpo Järvinen tp->retrans_out -= decr; 1472797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 1473797108d1SIlpo Järvinen tp->lost_out -= decr; 1474797108d1SIlpo Järvinen 1475797108d1SIlpo Järvinen /* Reno case is special. Sigh... */ 1476797108d1SIlpo Järvinen if (tcp_is_reno(tp) && decr > 0) 1477797108d1SIlpo Järvinen tp->sacked_out -= min_t(u32, tp->sacked_out, decr); 1478797108d1SIlpo Järvinen 1479797108d1SIlpo Järvinen if (tp->lost_skb_hint && 1480797108d1SIlpo Järvinen before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && 1481713bafeaSYuchung Cheng (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) 1482797108d1SIlpo Järvinen tp->lost_cnt_hint -= decr; 1483797108d1SIlpo Järvinen 1484797108d1SIlpo Järvinen tcp_verify_left_out(tp); 1485797108d1SIlpo Järvinen } 1486797108d1SIlpo Järvinen 14870a2cf20cSSoheil Hassas Yeganeh static bool tcp_has_tx_tstamp(const struct sk_buff *skb) 14880a2cf20cSSoheil Hassas Yeganeh { 14890a2cf20cSSoheil Hassas Yeganeh return TCP_SKB_CB(skb)->txstamp_ack || 14900a2cf20cSSoheil Hassas Yeganeh (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP); 14910a2cf20cSSoheil Hassas Yeganeh } 14920a2cf20cSSoheil Hassas Yeganeh 1493490cc7d0SWillem de Bruijn static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2) 1494490cc7d0SWillem de Bruijn { 1495490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo = skb_shinfo(skb); 1496490cc7d0SWillem de Bruijn 14970a2cf20cSSoheil Hassas Yeganeh if (unlikely(tcp_has_tx_tstamp(skb)) && 1498490cc7d0SWillem de Bruijn !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) { 1499490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo2 = skb_shinfo(skb2); 1500490cc7d0SWillem de Bruijn u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP; 1501490cc7d0SWillem de Bruijn 1502490cc7d0SWillem de Bruijn shinfo->tx_flags &= ~tsflags; 1503490cc7d0SWillem de Bruijn shinfo2->tx_flags |= tsflags; 1504490cc7d0SWillem de Bruijn swap(shinfo->tskey, shinfo2->tskey); 1505b51e13faSMartin KaFai Lau TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack; 1506b51e13faSMartin KaFai Lau TCP_SKB_CB(skb)->txstamp_ack = 0; 1507490cc7d0SWillem de Bruijn } 1508490cc7d0SWillem de Bruijn } 1509490cc7d0SWillem de Bruijn 1510a166140eSMartin KaFai Lau static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2) 1511a166140eSMartin KaFai Lau { 1512a166140eSMartin KaFai Lau TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor; 1513a166140eSMartin KaFai Lau TCP_SKB_CB(skb)->eor = 0; 1514a166140eSMartin KaFai Lau } 1515a166140eSMartin KaFai Lau 151675c119afSEric Dumazet /* Insert buff after skb on the write or rtx queue of sk. */ 151775c119afSEric Dumazet static void tcp_insert_write_queue_after(struct sk_buff *skb, 151875c119afSEric Dumazet struct sk_buff *buff, 151975c119afSEric Dumazet struct sock *sk, 152075c119afSEric Dumazet enum tcp_queue tcp_queue) 152175c119afSEric Dumazet { 152275c119afSEric Dumazet if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE) 152375c119afSEric Dumazet __skb_queue_after(&sk->sk_write_queue, skb, buff); 152475c119afSEric Dumazet else 152575c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, buff); 152675c119afSEric Dumazet } 152775c119afSEric Dumazet 15281da177e4SLinus Torvalds /* Function to create two new TCP segments. Shrinks the given segment 15291da177e4SLinus Torvalds * to the specified size and appends a new segment with the rest of the 15301da177e4SLinus Torvalds * packet to the list. This won't be called frequently, I hope. 15311da177e4SLinus Torvalds * Remember, these are still headerless SKBs at this point. 15321da177e4SLinus Torvalds */ 153375c119afSEric Dumazet int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue, 153475c119afSEric Dumazet struct sk_buff *skb, u32 len, 15356cc55e09SOctavian Purdila unsigned int mss_now, gfp_t gfp) 15361da177e4SLinus Torvalds { 15371da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 15381da177e4SLinus Torvalds struct sk_buff *buff; 15396475be16SDavid S. Miller int nsize, old_factor; 1540b617158dSEric Dumazet long limit; 1541b60b49eaSHerbert Xu int nlen; 15429ce01461SIlpo Järvinen u8 flags; 15431da177e4SLinus Torvalds 15442fceec13SIlpo Järvinen if (WARN_ON(len > skb->len)) 15452fceec13SIlpo Järvinen return -EINVAL; 15466a438bbeSStephen Hemminger 15471da177e4SLinus Torvalds nsize = skb_headlen(skb) - len; 15481da177e4SLinus Torvalds if (nsize < 0) 15491da177e4SLinus Torvalds nsize = 0; 15501da177e4SLinus Torvalds 1551b617158dSEric Dumazet /* tcp_sendmsg() can overshoot sk_wmem_queued by one full size skb. 1552b617158dSEric Dumazet * We need some allowance to not penalize applications setting small 1553b617158dSEric Dumazet * SO_SNDBUF values. 1554b617158dSEric Dumazet * Also allow first and last skb in retransmit queue to be split. 1555b617158dSEric Dumazet */ 15567c4e983cSAlexander Duyck limit = sk->sk_sndbuf + 2 * SKB_TRUESIZE(GSO_LEGACY_MAX_SIZE); 1557b617158dSEric Dumazet if (unlikely((sk->sk_wmem_queued >> 1) > limit && 1558b617158dSEric Dumazet tcp_queue != TCP_FRAG_IN_WRITE_QUEUE && 1559b617158dSEric Dumazet skb != tcp_rtx_queue_head(sk) && 1560b617158dSEric Dumazet skb != tcp_rtx_queue_tail(sk))) { 1561f070ef2aSEric Dumazet NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPWQUEUETOOBIG); 1562f070ef2aSEric Dumazet return -ENOMEM; 1563f070ef2aSEric Dumazet } 1564f070ef2aSEric Dumazet 1565c4777efaSEric Dumazet if (skb_unclone_keeptruesize(skb, gfp)) 15661da177e4SLinus Torvalds return -ENOMEM; 15671da177e4SLinus Torvalds 15681da177e4SLinus Torvalds /* Get a new skb... force flag on. */ 1569f8dd3b8dSEric Dumazet buff = tcp_stream_alloc_skb(sk, nsize, gfp, true); 157051456b29SIan Morris if (!buff) 15711da177e4SLinus Torvalds return -ENOMEM; /* We'll just try again later. */ 157241477662SJakub Kicinski skb_copy_decrypted(buff, skb); 15735a369ca6SPaolo Abeni mptcp_skb_ext_copy(buff, skb); 1574ef5cb973SHerbert Xu 1575ab4e846aSEric Dumazet sk_wmem_queued_add(sk, buff->truesize); 15763ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1577b60b49eaSHerbert Xu nlen = skb->len - len - nsize; 1578b60b49eaSHerbert Xu buff->truesize += nlen; 1579b60b49eaSHerbert Xu skb->truesize -= nlen; 15801da177e4SLinus Torvalds 15811da177e4SLinus Torvalds /* Correct the sequence numbers. */ 15821da177e4SLinus Torvalds TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 15831da177e4SLinus Torvalds TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 15841da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 15851da177e4SLinus Torvalds 15861da177e4SLinus Torvalds /* PSH and FIN should only be set in the second packet. */ 15874de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 15884de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 15894de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1590e14c3cafSHerbert Xu TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; 1591a166140eSMartin KaFai Lau tcp_skb_fragment_eor(skb, buff); 15921da177e4SLinus Torvalds 15931da177e4SLinus Torvalds skb_split(skb, buff, len); 15941da177e4SLinus Torvalds 1595a1ac9c8aSMartin KaFai Lau skb_set_delivery_time(buff, skb->tstamp, true); 1596490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 15971da177e4SLinus Torvalds 15986475be16SDavid S. Miller old_factor = tcp_skb_pcount(skb); 15996475be16SDavid S. Miller 16001da177e4SLinus Torvalds /* Fix up tso_factor for both original and new SKB. */ 16015bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 16025bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 16031da177e4SLinus Torvalds 1604b9f64820SYuchung Cheng /* Update delivered info for the new segment */ 1605b9f64820SYuchung Cheng TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx; 1606b9f64820SYuchung Cheng 16076475be16SDavid S. Miller /* If this packet has been sent out already, we must 16086475be16SDavid S. Miller * adjust the various packet counters. 16096475be16SDavid S. Miller */ 1610cf0b450cSHerbert Xu if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 16116475be16SDavid S. Miller int diff = old_factor - tcp_skb_pcount(skb) - 16126475be16SDavid S. Miller tcp_skb_pcount(buff); 16131da177e4SLinus Torvalds 1614797108d1SIlpo Järvinen if (diff) 1615797108d1SIlpo Järvinen tcp_adjust_pcount(sk, skb, diff); 16161da177e4SLinus Torvalds } 16171da177e4SLinus Torvalds 16181da177e4SLinus Torvalds /* Link BUFF into the send queue. */ 1619f4a775d1SEric Dumazet __skb_header_release(buff); 162075c119afSEric Dumazet tcp_insert_write_queue_after(skb, buff, sk, tcp_queue); 1621f67971e6SEric Dumazet if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE) 1622e2080072SEric Dumazet list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor); 16231da177e4SLinus Torvalds 16241da177e4SLinus Torvalds return 0; 16251da177e4SLinus Torvalds } 16261da177e4SLinus Torvalds 1627f4d01666SEric Dumazet /* This is similar to __pskb_pull_tail(). The difference is that pulled 1628f4d01666SEric Dumazet * data is not copied, but immediately discarded. 16291da177e4SLinus Torvalds */ 16307162fb24SEric Dumazet static int __pskb_trim_head(struct sk_buff *skb, int len) 16311da177e4SLinus Torvalds { 16327b7fc97aSEric Dumazet struct skb_shared_info *shinfo; 16331da177e4SLinus Torvalds int i, k, eat; 16341da177e4SLinus Torvalds 16354fa48bf3SEric Dumazet eat = min_t(int, len, skb_headlen(skb)); 16364fa48bf3SEric Dumazet if (eat) { 16374fa48bf3SEric Dumazet __skb_pull(skb, eat); 16384fa48bf3SEric Dumazet len -= eat; 16394fa48bf3SEric Dumazet if (!len) 16407162fb24SEric Dumazet return 0; 16414fa48bf3SEric Dumazet } 16421da177e4SLinus Torvalds eat = len; 16431da177e4SLinus Torvalds k = 0; 16447b7fc97aSEric Dumazet shinfo = skb_shinfo(skb); 16457b7fc97aSEric Dumazet for (i = 0; i < shinfo->nr_frags; i++) { 16467b7fc97aSEric Dumazet int size = skb_frag_size(&shinfo->frags[i]); 16479e903e08SEric Dumazet 16489e903e08SEric Dumazet if (size <= eat) { 1649aff65da0SIan Campbell skb_frag_unref(skb, i); 16509e903e08SEric Dumazet eat -= size; 16511da177e4SLinus Torvalds } else { 16527b7fc97aSEric Dumazet shinfo->frags[k] = shinfo->frags[i]; 16531da177e4SLinus Torvalds if (eat) { 1654b54c9d5bSJonathan Lemon skb_frag_off_add(&shinfo->frags[k], eat); 16557b7fc97aSEric Dumazet skb_frag_size_sub(&shinfo->frags[k], eat); 16561da177e4SLinus Torvalds eat = 0; 16571da177e4SLinus Torvalds } 16581da177e4SLinus Torvalds k++; 16591da177e4SLinus Torvalds } 16601da177e4SLinus Torvalds } 16617b7fc97aSEric Dumazet shinfo->nr_frags = k; 16621da177e4SLinus Torvalds 16631da177e4SLinus Torvalds skb->data_len -= len; 16641da177e4SLinus Torvalds skb->len = skb->data_len; 16657162fb24SEric Dumazet return len; 16661da177e4SLinus Torvalds } 16671da177e4SLinus Torvalds 166867edfef7SAndi Kleen /* Remove acked data from a packet in the transmit queue. */ 16691da177e4SLinus Torvalds int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 16701da177e4SLinus Torvalds { 16717162fb24SEric Dumazet u32 delta_truesize; 16727162fb24SEric Dumazet 1673c4777efaSEric Dumazet if (skb_unclone_keeptruesize(skb, GFP_ATOMIC)) 16741da177e4SLinus Torvalds return -ENOMEM; 16751da177e4SLinus Torvalds 16767162fb24SEric Dumazet delta_truesize = __pskb_trim_head(skb, len); 16771da177e4SLinus Torvalds 16781da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq += len; 16791da177e4SLinus Torvalds 16807162fb24SEric Dumazet if (delta_truesize) { 16817162fb24SEric Dumazet skb->truesize -= delta_truesize; 1682ab4e846aSEric Dumazet sk_wmem_queued_add(sk, -delta_truesize); 16839b65b17dSTalal Ahmad if (!skb_zcopy_pure(skb)) 16847162fb24SEric Dumazet sk_mem_uncharge(sk, delta_truesize); 16857162fb24SEric Dumazet } 16861da177e4SLinus Torvalds 16875b35e1e6SNeal Cardwell /* Any change of skb->len requires recalculation of tso factor. */ 16881da177e4SLinus Torvalds if (tcp_skb_pcount(skb) > 1) 16895bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb)); 16901da177e4SLinus Torvalds 16911da177e4SLinus Torvalds return 0; 16921da177e4SLinus Torvalds } 16931da177e4SLinus Torvalds 16941b63edd6SYuchung Cheng /* Calculate MSS not accounting any TCP options. */ 16951b63edd6SYuchung Cheng static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu) 16965d424d5aSJohn Heffner { 1697cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1698cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 16995d424d5aSJohn Heffner int mss_now; 17005d424d5aSJohn Heffner 17015d424d5aSJohn Heffner /* Calculate base mss without TCP options: 17025d424d5aSJohn Heffner It is MMS_S - sizeof(tcphdr) of rfc1122 17035d424d5aSJohn Heffner */ 17045d424d5aSJohn Heffner mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 17055d424d5aSJohn Heffner 170667469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 170767469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 170867469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 170967469601SEric Dumazet 171067469601SEric Dumazet if (dst && dst_allfrag(dst)) 171167469601SEric Dumazet mss_now -= icsk->icsk_af_ops->net_frag_header_len; 171267469601SEric Dumazet } 171367469601SEric Dumazet 17145d424d5aSJohn Heffner /* Clamp it (mss_clamp does not include tcp options) */ 17155d424d5aSJohn Heffner if (mss_now > tp->rx_opt.mss_clamp) 17165d424d5aSJohn Heffner mss_now = tp->rx_opt.mss_clamp; 17175d424d5aSJohn Heffner 17185d424d5aSJohn Heffner /* Now subtract optional transport overhead */ 17195d424d5aSJohn Heffner mss_now -= icsk->icsk_ext_hdr_len; 17205d424d5aSJohn Heffner 17215d424d5aSJohn Heffner /* Then reserve room for full set of TCP options and 8 bytes of data */ 1722*78eb166cSKuniyuki Iwashima mss_now = max(mss_now, 1723*78eb166cSKuniyuki Iwashima READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_snd_mss)); 17245d424d5aSJohn Heffner return mss_now; 17255d424d5aSJohn Heffner } 17265d424d5aSJohn Heffner 17271b63edd6SYuchung Cheng /* Calculate MSS. Not accounting for SACKs here. */ 17281b63edd6SYuchung Cheng int tcp_mtu_to_mss(struct sock *sk, int pmtu) 17291b63edd6SYuchung Cheng { 17301b63edd6SYuchung Cheng /* Subtract TCP options size, not including SACKs */ 17311b63edd6SYuchung Cheng return __tcp_mtu_to_mss(sk, pmtu) - 17321b63edd6SYuchung Cheng (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr)); 17331b63edd6SYuchung Cheng } 1734c7bb4b89SEric Dumazet EXPORT_SYMBOL(tcp_mtu_to_mss); 17351b63edd6SYuchung Cheng 17365d424d5aSJohn Heffner /* Inverse of above */ 173767469601SEric Dumazet int tcp_mss_to_mtu(struct sock *sk, int mss) 17385d424d5aSJohn Heffner { 1739cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1740cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 17415d424d5aSJohn Heffner int mtu; 17425d424d5aSJohn Heffner 17435d424d5aSJohn Heffner mtu = mss + 17445d424d5aSJohn Heffner tp->tcp_header_len + 17455d424d5aSJohn Heffner icsk->icsk_ext_hdr_len + 17465d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 17475d424d5aSJohn Heffner 174867469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 174967469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 175067469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 175167469601SEric Dumazet 175267469601SEric Dumazet if (dst && dst_allfrag(dst)) 175367469601SEric Dumazet mtu += icsk->icsk_af_ops->net_frag_header_len; 175467469601SEric Dumazet } 17555d424d5aSJohn Heffner return mtu; 17565d424d5aSJohn Heffner } 1757556c6b46SNeal Cardwell EXPORT_SYMBOL(tcp_mss_to_mtu); 17585d424d5aSJohn Heffner 175967edfef7SAndi Kleen /* MTU probing init per socket */ 17605d424d5aSJohn Heffner void tcp_mtup_init(struct sock *sk) 17615d424d5aSJohn Heffner { 17625d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 17635d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 1764b0f9ca53SFan Du struct net *net = sock_net(sk); 17655d424d5aSJohn Heffner 1766f47d00e0SKuniyuki Iwashima icsk->icsk_mtup.enabled = READ_ONCE(net->ipv4.sysctl_tcp_mtu_probing) > 1; 17675d424d5aSJohn Heffner icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 17685d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 176988d78bc0SKuniyuki Iwashima icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, READ_ONCE(net->ipv4.sysctl_tcp_base_mss)); 17705d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 177105cbc0dbSFan Du if (icsk->icsk_mtup.enabled) 1772c74df29aSEric Dumazet icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; 17735d424d5aSJohn Heffner } 17744bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_mtup_init); 17755d424d5aSJohn Heffner 17761da177e4SLinus Torvalds /* This function synchronize snd mss to current pmtu/exthdr set. 17771da177e4SLinus Torvalds 17781da177e4SLinus Torvalds tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 17791da177e4SLinus Torvalds for TCP options, but includes only bare TCP header. 17801da177e4SLinus Torvalds 17811da177e4SLinus Torvalds tp->rx_opt.mss_clamp is mss negotiated at connection setup. 1782caa20d9aSStephen Hemminger It is minimum of user_mss and mss received with SYN. 17831da177e4SLinus Torvalds It also does not include TCP options. 17841da177e4SLinus Torvalds 1785d83d8461SArnaldo Carvalho de Melo inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. 17861da177e4SLinus Torvalds 17871da177e4SLinus Torvalds tp->mss_cache is current effective sending mss, including 17881da177e4SLinus Torvalds all tcp options except for SACKs. It is evaluated, 17891da177e4SLinus Torvalds taking into account current pmtu, but never exceeds 17901da177e4SLinus Torvalds tp->rx_opt.mss_clamp. 17911da177e4SLinus Torvalds 17921da177e4SLinus Torvalds NOTE1. rfc1122 clearly states that advertised MSS 17931da177e4SLinus Torvalds DOES NOT include either tcp or ip options. 17941da177e4SLinus Torvalds 1795d83d8461SArnaldo Carvalho de Melo NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache 1796d83d8461SArnaldo Carvalho de Melo are READ ONLY outside this function. --ANK (980731) 17971da177e4SLinus Torvalds */ 17981da177e4SLinus Torvalds unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 17991da177e4SLinus Torvalds { 18001da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1801d83d8461SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 18025d424d5aSJohn Heffner int mss_now; 18031da177e4SLinus Torvalds 18045d424d5aSJohn Heffner if (icsk->icsk_mtup.search_high > pmtu) 18055d424d5aSJohn Heffner icsk->icsk_mtup.search_high = pmtu; 18061da177e4SLinus Torvalds 18075d424d5aSJohn Heffner mss_now = tcp_mtu_to_mss(sk, pmtu); 1808409d22b4SIlpo Järvinen mss_now = tcp_bound_to_half_wnd(tp, mss_now); 18091da177e4SLinus Torvalds 18101da177e4SLinus Torvalds /* And store cached results */ 1811d83d8461SArnaldo Carvalho de Melo icsk->icsk_pmtu_cookie = pmtu; 18125d424d5aSJohn Heffner if (icsk->icsk_mtup.enabled) 18135d424d5aSJohn Heffner mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); 1814c1b4a7e6SDavid S. Miller tp->mss_cache = mss_now; 18151da177e4SLinus Torvalds 18161da177e4SLinus Torvalds return mss_now; 18171da177e4SLinus Torvalds } 18184bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_sync_mss); 18191da177e4SLinus Torvalds 18201da177e4SLinus Torvalds /* Compute the current effective MSS, taking SACKs and IP options, 18211da177e4SLinus Torvalds * and even PMTU discovery events into account. 18221da177e4SLinus Torvalds */ 18230c54b85fSIlpo Järvinen unsigned int tcp_current_mss(struct sock *sk) 18241da177e4SLinus Torvalds { 1825cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1826cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1827c1b4a7e6SDavid S. Miller u32 mss_now; 182895c96174SEric Dumazet unsigned int header_len; 182933ad798cSAdam Langley struct tcp_out_options opts; 183033ad798cSAdam Langley struct tcp_md5sig_key *md5; 18311da177e4SLinus Torvalds 1832c1b4a7e6SDavid S. Miller mss_now = tp->mss_cache; 1833c1b4a7e6SDavid S. Miller 18341da177e4SLinus Torvalds if (dst) { 18351da177e4SLinus Torvalds u32 mtu = dst_mtu(dst); 1836d83d8461SArnaldo Carvalho de Melo if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 18371da177e4SLinus Torvalds mss_now = tcp_sync_mss(sk, mtu); 18381da177e4SLinus Torvalds } 18391da177e4SLinus Torvalds 184033ad798cSAdam Langley header_len = tcp_established_options(sk, NULL, &opts, &md5) + 184133ad798cSAdam Langley sizeof(struct tcphdr); 184233ad798cSAdam Langley /* The mss_cache is sized based on tp->tcp_header_len, which assumes 184333ad798cSAdam Langley * some common options. If this is an odd packet (because we have SACK 184433ad798cSAdam Langley * blocks etc) then our calculated header_len will be different, and 184533ad798cSAdam Langley * we have to adjust mss_now correspondingly */ 184633ad798cSAdam Langley if (header_len != tp->tcp_header_len) { 184733ad798cSAdam Langley int delta = (int) header_len - tp->tcp_header_len; 184833ad798cSAdam Langley mss_now -= delta; 184933ad798cSAdam Langley } 1850cfb6eeb4SYOSHIFUJI Hideaki 18511da177e4SLinus Torvalds return mss_now; 18521da177e4SLinus Torvalds } 18531da177e4SLinus Torvalds 185486fd14adSWeiping Pan /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. 185586fd14adSWeiping Pan * As additional protections, we do not touch cwnd in retransmission phases, 185686fd14adSWeiping Pan * and if application hit its sndbuf limit recently. 185786fd14adSWeiping Pan */ 185886fd14adSWeiping Pan static void tcp_cwnd_application_limited(struct sock *sk) 1859a762a980SDavid S. Miller { 18609e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1861a762a980SDavid S. Miller 186286fd14adSWeiping Pan if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && 186386fd14adSWeiping Pan sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 186486fd14adSWeiping Pan /* Limited by application or receiver window. */ 186586fd14adSWeiping Pan u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); 186686fd14adSWeiping Pan u32 win_used = max(tp->snd_cwnd_used, init_win); 186740570375SEric Dumazet if (win_used < tcp_snd_cwnd(tp)) { 186886fd14adSWeiping Pan tp->snd_ssthresh = tcp_current_ssthresh(sk); 186940570375SEric Dumazet tcp_snd_cwnd_set(tp, (tcp_snd_cwnd(tp) + win_used) >> 1); 187086fd14adSWeiping Pan } 187186fd14adSWeiping Pan tp->snd_cwnd_used = 0; 187286fd14adSWeiping Pan } 1873c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 187486fd14adSWeiping Pan } 187586fd14adSWeiping Pan 1876ca8a2263SNeal Cardwell static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited) 1877a762a980SDavid S. Miller { 18781b1fc3fdSWei Wang const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 1879a762a980SDavid S. Miller struct tcp_sock *tp = tcp_sk(sk); 1880a762a980SDavid S. Miller 1881ca8a2263SNeal Cardwell /* Track the maximum number of outstanding packets in each 1882ca8a2263SNeal Cardwell * window, and remember whether we were cwnd-limited then. 1883ca8a2263SNeal Cardwell */ 1884ca8a2263SNeal Cardwell if (!before(tp->snd_una, tp->max_packets_seq) || 1885299bcb55SNeal Cardwell tp->packets_out > tp->max_packets_out || 1886299bcb55SNeal Cardwell is_cwnd_limited) { 1887ca8a2263SNeal Cardwell tp->max_packets_out = tp->packets_out; 1888ca8a2263SNeal Cardwell tp->max_packets_seq = tp->snd_nxt; 1889ca8a2263SNeal Cardwell tp->is_cwnd_limited = is_cwnd_limited; 1890ca8a2263SNeal Cardwell } 1891e114a710SEric Dumazet 189224901551SEric Dumazet if (tcp_is_cwnd_limited(sk)) { 1893a762a980SDavid S. Miller /* Network is feed fully. */ 1894a762a980SDavid S. Miller tp->snd_cwnd_used = 0; 1895c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 1896a762a980SDavid S. Miller } else { 1897a762a980SDavid S. Miller /* Network starves. */ 1898a762a980SDavid S. Miller if (tp->packets_out > tp->snd_cwnd_used) 1899a762a980SDavid S. Miller tp->snd_cwnd_used = tp->packets_out; 1900a762a980SDavid S. Miller 1901b510f0d2SEric Dumazet if (sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle && 1902c2203cf7SEric Dumazet (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto && 19031b1fc3fdSWei Wang !ca_ops->cong_control) 1904a762a980SDavid S. Miller tcp_cwnd_application_limited(sk); 1905b0f71bd3SFrancis Yan 1906b0f71bd3SFrancis Yan /* The following conditions together indicate the starvation 1907b0f71bd3SFrancis Yan * is caused by insufficient sender buffer: 1908b0f71bd3SFrancis Yan * 1) just sent some data (see tcp_write_xmit) 1909b0f71bd3SFrancis Yan * 2) not cwnd limited (this else condition) 191075c119afSEric Dumazet * 3) no more data to send (tcp_write_queue_empty()) 1911b0f71bd3SFrancis Yan * 4) application is hitting buffer limit (SOCK_NOSPACE) 1912b0f71bd3SFrancis Yan */ 191375c119afSEric Dumazet if (tcp_write_queue_empty(sk) && sk->sk_socket && 1914b0f71bd3SFrancis Yan test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) && 1915b0f71bd3SFrancis Yan (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) 1916b0f71bd3SFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED); 1917a762a980SDavid S. Miller } 1918a762a980SDavid S. Miller } 1919a762a980SDavid S. Miller 1920d4589926SEric Dumazet /* Minshall's variant of the Nagle send check. */ 1921d4589926SEric Dumazet static bool tcp_minshall_check(const struct tcp_sock *tp) 1922d4589926SEric Dumazet { 1923d4589926SEric Dumazet return after(tp->snd_sml, tp->snd_una) && 1924d4589926SEric Dumazet !after(tp->snd_sml, tp->snd_nxt); 1925d4589926SEric Dumazet } 1926d4589926SEric Dumazet 1927d4589926SEric Dumazet /* Update snd_sml if this skb is under mss 1928d4589926SEric Dumazet * Note that a TSO packet might end with a sub-mss segment 1929d4589926SEric Dumazet * The test is really : 1930d4589926SEric Dumazet * if ((skb->len % mss) != 0) 1931d4589926SEric Dumazet * tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1932d4589926SEric Dumazet * But we can avoid doing the divide again given we already have 1933d4589926SEric Dumazet * skb_pcount = skb->len / mss_now 19340e3a4803SIlpo Järvinen */ 1935d4589926SEric Dumazet static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now, 1936d4589926SEric Dumazet const struct sk_buff *skb) 1937d4589926SEric Dumazet { 1938d4589926SEric Dumazet if (skb->len < tcp_skb_pcount(skb) * mss_now) 1939d4589926SEric Dumazet tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1940d4589926SEric Dumazet } 1941d4589926SEric Dumazet 1942d4589926SEric Dumazet /* Return false, if packet can be sent now without violation Nagle's rules: 1943d4589926SEric Dumazet * 1. It is full sized. (provided by caller in %partial bool) 1944d4589926SEric Dumazet * 2. Or it contains FIN. (already checked by caller) 1945d4589926SEric Dumazet * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. 1946d4589926SEric Dumazet * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 1947d4589926SEric Dumazet * With Minshall's modification: all sent small packets are ACKed. 1948d4589926SEric Dumazet */ 1949d4589926SEric Dumazet static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp, 1950cc93fc51SPeter Pan(潘卫平) int nonagle) 1951d4589926SEric Dumazet { 1952d4589926SEric Dumazet return partial && 1953d4589926SEric Dumazet ((nonagle & TCP_NAGLE_CORK) || 1954d4589926SEric Dumazet (!nonagle && tp->packets_out && tcp_minshall_check(tp))); 1955d4589926SEric Dumazet } 1956605ad7f1SEric Dumazet 1957605ad7f1SEric Dumazet /* Return how many segs we'd like on a TSO packet, 195865466904SEric Dumazet * depending on current pacing rate, and how close the peer is. 195965466904SEric Dumazet * 196065466904SEric Dumazet * Rationale is: 196165466904SEric Dumazet * - For close peers, we rather send bigger packets to reduce 196265466904SEric Dumazet * cpu costs, because occasional losses will be repaired fast. 196365466904SEric Dumazet * - For long distance/rtt flows, we would like to get ACK clocking 196465466904SEric Dumazet * with 1 ACK per ms. 196565466904SEric Dumazet * 196665466904SEric Dumazet * Use min_rtt to help adapt TSO burst size, with smaller min_rtt resulting 196765466904SEric Dumazet * in bigger TSO bursts. We we cut the RTT-based allowance in half 196865466904SEric Dumazet * for every 2^9 usec (aka 512 us) of RTT, so that the RTT-based allowance 196965466904SEric Dumazet * is below 1500 bytes after 6 * ~500 usec = 3ms. 1970605ad7f1SEric Dumazet */ 1971dcb8c9b4SEric Dumazet static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now, 19721b3878caSNeal Cardwell int min_tso_segs) 1973605ad7f1SEric Dumazet { 197465466904SEric Dumazet unsigned long bytes; 197565466904SEric Dumazet u32 r; 1976605ad7f1SEric Dumazet 197765466904SEric Dumazet bytes = sk->sk_pacing_rate >> READ_ONCE(sk->sk_pacing_shift); 1978605ad7f1SEric Dumazet 197965466904SEric Dumazet r = tcp_min_rtt(tcp_sk(sk)) >> sock_net(sk)->ipv4.sysctl_tcp_tso_rtt_log; 198065466904SEric Dumazet if (r < BITS_PER_TYPE(sk->sk_gso_max_size)) 198165466904SEric Dumazet bytes += sk->sk_gso_max_size >> r; 1982605ad7f1SEric Dumazet 198365466904SEric Dumazet bytes = min_t(unsigned long, bytes, sk->sk_gso_max_size); 198465466904SEric Dumazet 198565466904SEric Dumazet return max_t(u32, bytes / mss_now, min_tso_segs); 1986605ad7f1SEric Dumazet } 1987605ad7f1SEric Dumazet 1988ed6e7268SNeal Cardwell /* Return the number of segments we want in the skb we are transmitting. 1989ed6e7268SNeal Cardwell * See if congestion control module wants to decide; otherwise, autosize. 1990ed6e7268SNeal Cardwell */ 1991ed6e7268SNeal Cardwell static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now) 1992ed6e7268SNeal Cardwell { 1993ed6e7268SNeal Cardwell const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 1994dcb8c9b4SEric Dumazet u32 min_tso, tso_segs; 1995ed6e7268SNeal Cardwell 1996dcb8c9b4SEric Dumazet min_tso = ca_ops->min_tso_segs ? 1997dcb8c9b4SEric Dumazet ca_ops->min_tso_segs(sk) : 1998dcb8c9b4SEric Dumazet sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs; 1999dcb8c9b4SEric Dumazet 2000dcb8c9b4SEric Dumazet tso_segs = tcp_tso_autosize(sk, mss_now, min_tso); 2001350c9f48SEric Dumazet return min_t(u32, tso_segs, sk->sk_gso_max_segs); 2002ed6e7268SNeal Cardwell } 2003ed6e7268SNeal Cardwell 2004d4589926SEric Dumazet /* Returns the portion of skb which can be sent right away */ 2005d4589926SEric Dumazet static unsigned int tcp_mss_split_point(const struct sock *sk, 2006d4589926SEric Dumazet const struct sk_buff *skb, 2007d4589926SEric Dumazet unsigned int mss_now, 2008d4589926SEric Dumazet unsigned int max_segs, 2009d4589926SEric Dumazet int nonagle) 2010c1b4a7e6SDavid S. Miller { 2011cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 2012d4589926SEric Dumazet u32 partial, needed, window, max_len; 2013c1b4a7e6SDavid S. Miller 201490840defSIlpo Järvinen window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 20151485348dSBen Hutchings max_len = mss_now * max_segs; 20160e3a4803SIlpo Järvinen 20171485348dSBen Hutchings if (likely(max_len <= window && skb != tcp_write_queue_tail(sk))) 20181485348dSBen Hutchings return max_len; 20190e3a4803SIlpo Järvinen 20205ea3a748SIlpo Järvinen needed = min(skb->len, window); 20215ea3a748SIlpo Järvinen 20221485348dSBen Hutchings if (max_len <= needed) 20231485348dSBen Hutchings return max_len; 20240e3a4803SIlpo Järvinen 2025d4589926SEric Dumazet partial = needed % mss_now; 2026d4589926SEric Dumazet /* If last segment is not a full MSS, check if Nagle rules allow us 2027d4589926SEric Dumazet * to include this last segment in this skb. 2028d4589926SEric Dumazet * Otherwise, we'll split the skb at last MSS boundary 2029d4589926SEric Dumazet */ 2030cc93fc51SPeter Pan(潘卫平) if (tcp_nagle_check(partial != 0, tp, nonagle)) 2031d4589926SEric Dumazet return needed - partial; 2032d4589926SEric Dumazet 2033d4589926SEric Dumazet return needed; 2034c1b4a7e6SDavid S. Miller } 2035c1b4a7e6SDavid S. Miller 2036c1b4a7e6SDavid S. Miller /* Can at least one segment of SKB be sent right now, according to the 2037c1b4a7e6SDavid S. Miller * congestion window rules? If so, return how many segments are allowed. 2038c1b4a7e6SDavid S. Miller */ 2039cf533ea5SEric Dumazet static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, 2040cf533ea5SEric Dumazet const struct sk_buff *skb) 2041c1b4a7e6SDavid S. Miller { 2042d649a7a8SEric Dumazet u32 in_flight, cwnd, halfcwnd; 2043c1b4a7e6SDavid S. Miller 2044c1b4a7e6SDavid S. Miller /* Don't be strict about the congestion window for the final FIN. */ 20454de075e0SEric Dumazet if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && 20464de075e0SEric Dumazet tcp_skb_pcount(skb) == 1) 2047c1b4a7e6SDavid S. Miller return 1; 2048c1b4a7e6SDavid S. Miller 2049c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 205040570375SEric Dumazet cwnd = tcp_snd_cwnd(tp); 2051d649a7a8SEric Dumazet if (in_flight >= cwnd) 2052c1b4a7e6SDavid S. Miller return 0; 2053d649a7a8SEric Dumazet 2054d649a7a8SEric Dumazet /* For better scheduling, ensure we have at least 2055d649a7a8SEric Dumazet * 2 GSO packets in flight. 2056d649a7a8SEric Dumazet */ 2057d649a7a8SEric Dumazet halfcwnd = max(cwnd >> 1, 1U); 2058d649a7a8SEric Dumazet return min(halfcwnd, cwnd - in_flight); 2059c1b4a7e6SDavid S. Miller } 2060c1b4a7e6SDavid S. Miller 2061b595076aSUwe Kleine-König /* Initialize TSO state of a skb. 206267edfef7SAndi Kleen * This must be invoked the first time we consider transmitting 2063c1b4a7e6SDavid S. Miller * SKB onto the wire. 2064c1b4a7e6SDavid S. Miller */ 20655bbb432cSEric Dumazet static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now) 2066c1b4a7e6SDavid S. Miller { 2067c1b4a7e6SDavid S. Miller int tso_segs = tcp_skb_pcount(skb); 2068c1b4a7e6SDavid S. Miller 2069f8269a49SIlpo Järvinen if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { 20705bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 2071c1b4a7e6SDavid S. Miller tso_segs = tcp_skb_pcount(skb); 2072c1b4a7e6SDavid S. Miller } 2073c1b4a7e6SDavid S. Miller return tso_segs; 2074c1b4a7e6SDavid S. Miller } 2075c1b4a7e6SDavid S. Miller 2076c1b4a7e6SDavid S. Miller 2077a2a385d6SEric Dumazet /* Return true if the Nagle test allows this packet to be 2078c1b4a7e6SDavid S. Miller * sent now. 2079c1b4a7e6SDavid S. Miller */ 2080a2a385d6SEric Dumazet static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, 2081c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 2082c1b4a7e6SDavid S. Miller { 2083c1b4a7e6SDavid S. Miller /* Nagle rule does not apply to frames, which sit in the middle of the 2084c1b4a7e6SDavid S. Miller * write_queue (they have no chances to get new data). 2085c1b4a7e6SDavid S. Miller * 2086c1b4a7e6SDavid S. Miller * This is implemented in the callers, where they modify the 'nonagle' 2087c1b4a7e6SDavid S. Miller * argument based upon the location of SKB in the send queue. 2088c1b4a7e6SDavid S. Miller */ 2089c1b4a7e6SDavid S. Miller if (nonagle & TCP_NAGLE_PUSH) 2090a2a385d6SEric Dumazet return true; 2091c1b4a7e6SDavid S. Miller 20929b44190dSYuchung Cheng /* Don't use the nagle rule for urgent data (or for the final FIN). */ 20939b44190dSYuchung Cheng if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) 2094a2a385d6SEric Dumazet return true; 2095c1b4a7e6SDavid S. Miller 2096cc93fc51SPeter Pan(潘卫平) if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle)) 2097a2a385d6SEric Dumazet return true; 2098c1b4a7e6SDavid S. Miller 2099a2a385d6SEric Dumazet return false; 2100c1b4a7e6SDavid S. Miller } 2101c1b4a7e6SDavid S. Miller 2102c1b4a7e6SDavid S. Miller /* Does at least the first segment of SKB fit into the send window? */ 2103a2a385d6SEric Dumazet static bool tcp_snd_wnd_test(const struct tcp_sock *tp, 2104a2a385d6SEric Dumazet const struct sk_buff *skb, 2105056834d9SIlpo Järvinen unsigned int cur_mss) 2106c1b4a7e6SDavid S. Miller { 2107c1b4a7e6SDavid S. Miller u32 end_seq = TCP_SKB_CB(skb)->end_seq; 2108c1b4a7e6SDavid S. Miller 2109c1b4a7e6SDavid S. Miller if (skb->len > cur_mss) 2110c1b4a7e6SDavid S. Miller end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 2111c1b4a7e6SDavid S. Miller 211290840defSIlpo Järvinen return !after(end_seq, tcp_wnd_end(tp)); 2113c1b4a7e6SDavid S. Miller } 2114c1b4a7e6SDavid S. Miller 2115c1b4a7e6SDavid S. Miller /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 2116c1b4a7e6SDavid S. Miller * which is put after SKB on the list. It is very much like 2117c1b4a7e6SDavid S. Miller * tcp_fragment() except that it may make several kinds of assumptions 2118c1b4a7e6SDavid S. Miller * in order to speed up the splitting operation. In particular, we 2119c1b4a7e6SDavid S. Miller * know that all the data is in scatter-gather pages, and that the 2120c1b4a7e6SDavid S. Miller * packet has never been sent out before (and thus is not cloned). 2121c1b4a7e6SDavid S. Miller */ 212256483341SEric Dumazet static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, 2123c4ead4c5SEric Dumazet unsigned int mss_now, gfp_t gfp) 2124c1b4a7e6SDavid S. Miller { 2125c1b4a7e6SDavid S. Miller int nlen = skb->len - len; 212656483341SEric Dumazet struct sk_buff *buff; 21279ce01461SIlpo Järvinen u8 flags; 2128c1b4a7e6SDavid S. Miller 2129c1b4a7e6SDavid S. Miller /* All of a TSO frame must be composed of paged data. */ 2130c8ac3774SHerbert Xu if (skb->len != skb->data_len) 213156483341SEric Dumazet return tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE, 213256483341SEric Dumazet skb, len, mss_now, gfp); 2133c1b4a7e6SDavid S. Miller 2134f8dd3b8dSEric Dumazet buff = tcp_stream_alloc_skb(sk, 0, gfp, true); 213551456b29SIan Morris if (unlikely(!buff)) 2136c1b4a7e6SDavid S. Miller return -ENOMEM; 213741477662SJakub Kicinski skb_copy_decrypted(buff, skb); 21385a369ca6SPaolo Abeni mptcp_skb_ext_copy(buff, skb); 2139c1b4a7e6SDavid S. Miller 2140ab4e846aSEric Dumazet sk_wmem_queued_add(sk, buff->truesize); 21413ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 2142b60b49eaSHerbert Xu buff->truesize += nlen; 2143c1b4a7e6SDavid S. Miller skb->truesize -= nlen; 2144c1b4a7e6SDavid S. Miller 2145c1b4a7e6SDavid S. Miller /* Correct the sequence numbers. */ 2146c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 2147c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 2148c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 2149c1b4a7e6SDavid S. Miller 2150c1b4a7e6SDavid S. Miller /* PSH and FIN should only be set in the second packet. */ 21514de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 21524de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 21534de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 2154c1b4a7e6SDavid S. Miller 2155a166140eSMartin KaFai Lau tcp_skb_fragment_eor(skb, buff); 2156a166140eSMartin KaFai Lau 2157c1b4a7e6SDavid S. Miller skb_split(skb, buff, len); 2158490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 2159c1b4a7e6SDavid S. Miller 2160c1b4a7e6SDavid S. Miller /* Fix up tso_factor for both original and new SKB. */ 21615bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 21625bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 2163c1b4a7e6SDavid S. Miller 2164c1b4a7e6SDavid S. Miller /* Link BUFF into the send queue. */ 2165f4a775d1SEric Dumazet __skb_header_release(buff); 216656483341SEric Dumazet tcp_insert_write_queue_after(skb, buff, sk, TCP_FRAG_IN_WRITE_QUEUE); 2167c1b4a7e6SDavid S. Miller 2168c1b4a7e6SDavid S. Miller return 0; 2169c1b4a7e6SDavid S. Miller } 2170c1b4a7e6SDavid S. Miller 2171c1b4a7e6SDavid S. Miller /* Try to defer sending, if possible, in order to minimize the amount 2172c1b4a7e6SDavid S. Miller * of TSO splitting we do. View it as a kind of TSO Nagle test. 2173c1b4a7e6SDavid S. Miller * 2174c1b4a7e6SDavid S. Miller * This algorithm is from John Heffner. 2175c1b4a7e6SDavid S. Miller */ 2176ca8a2263SNeal Cardwell static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb, 2177f9bfe4e6SEric Dumazet bool *is_cwnd_limited, 2178f9bfe4e6SEric Dumazet bool *is_rwnd_limited, 2179f9bfe4e6SEric Dumazet u32 max_segs) 2180c1b4a7e6SDavid S. Miller { 21816687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 2182f1c6ea38SEric Dumazet u32 send_win, cong_win, limit, in_flight; 218350c8339eSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 218450c8339eSEric Dumazet struct sk_buff *head; 2185ad9f4f50SEric Dumazet int win_divisor; 2186f1c6ea38SEric Dumazet s64 delta; 2187c1b4a7e6SDavid S. Miller 218899d7662aSEric Dumazet if (icsk->icsk_ca_state >= TCP_CA_Recovery) 2189ae8064acSJohn Heffner goto send_now; 2190ae8064acSJohn Heffner 21915f852eb5SEric Dumazet /* Avoid bursty behavior by allowing defer 2192a682850aSEric Dumazet * only if the last write was recent (1 ms). 2193a682850aSEric Dumazet * Note that tp->tcp_wstamp_ns can be in the future if we have 2194a682850aSEric Dumazet * packets waiting in a qdisc or device for EDT delivery. 21955f852eb5SEric Dumazet */ 2196a682850aSEric Dumazet delta = tp->tcp_clock_cache - tp->tcp_wstamp_ns - NSEC_PER_MSEC; 2197a682850aSEric Dumazet if (delta > 0) 2198ae8064acSJohn Heffner goto send_now; 2199908a75c1SDavid S. Miller 2200c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 2201c1b4a7e6SDavid S. Miller 2202c8c9aeb5SStefano Brivio BUG_ON(tcp_skb_pcount(skb) <= 1); 220340570375SEric Dumazet BUG_ON(tcp_snd_cwnd(tp) <= in_flight); 2204c1b4a7e6SDavid S. Miller 220590840defSIlpo Järvinen send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 2206c1b4a7e6SDavid S. Miller 2207c1b4a7e6SDavid S. Miller /* From in_flight test above, we know that cwnd > in_flight. */ 220840570375SEric Dumazet cong_win = (tcp_snd_cwnd(tp) - in_flight) * tp->mss_cache; 2209c1b4a7e6SDavid S. Miller 2210c1b4a7e6SDavid S. Miller limit = min(send_win, cong_win); 2211c1b4a7e6SDavid S. Miller 2212ba244fe9SDavid S. Miller /* If a full-sized TSO skb can be sent, do it. */ 2213605ad7f1SEric Dumazet if (limit >= max_segs * tp->mss_cache) 2214ae8064acSJohn Heffner goto send_now; 2215ba244fe9SDavid S. Miller 221662ad2761SIlpo Järvinen /* Middle in queue won't get any more data, full sendable already? */ 221762ad2761SIlpo Järvinen if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) 221862ad2761SIlpo Järvinen goto send_now; 221962ad2761SIlpo Järvinen 22205bbcc0f5SLinus Torvalds win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor); 2221ad9f4f50SEric Dumazet if (win_divisor) { 222240570375SEric Dumazet u32 chunk = min(tp->snd_wnd, tcp_snd_cwnd(tp) * tp->mss_cache); 2223c1b4a7e6SDavid S. Miller 2224c1b4a7e6SDavid S. Miller /* If at least some fraction of a window is available, 2225c1b4a7e6SDavid S. Miller * just use it. 2226c1b4a7e6SDavid S. Miller */ 2227ad9f4f50SEric Dumazet chunk /= win_divisor; 2228c1b4a7e6SDavid S. Miller if (limit >= chunk) 2229ae8064acSJohn Heffner goto send_now; 2230c1b4a7e6SDavid S. Miller } else { 2231c1b4a7e6SDavid S. Miller /* Different approach, try not to defer past a single 2232c1b4a7e6SDavid S. Miller * ACK. Receiver should ACK every other full sized 2233c1b4a7e6SDavid S. Miller * frame, so if we have space for more than 3 frames 2234c1b4a7e6SDavid S. Miller * then send now. 2235c1b4a7e6SDavid S. Miller */ 22366b5a5c0dSNeal Cardwell if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) 2237ae8064acSJohn Heffner goto send_now; 2238c1b4a7e6SDavid S. Miller } 2239c1b4a7e6SDavid S. Miller 224075c119afSEric Dumazet /* TODO : use tsorted_sent_queue ? */ 224175c119afSEric Dumazet head = tcp_rtx_queue_head(sk); 224275c119afSEric Dumazet if (!head) 224375c119afSEric Dumazet goto send_now; 2244f1c6ea38SEric Dumazet delta = tp->tcp_clock_cache - head->tstamp; 224550c8339eSEric Dumazet /* If next ACK is likely to come too late (half srtt), do not defer */ 2246f1c6ea38SEric Dumazet if ((s64)(delta - (u64)NSEC_PER_USEC * (tp->srtt_us >> 4)) < 0) 224750c8339eSEric Dumazet goto send_now; 224850c8339eSEric Dumazet 2249f9bfe4e6SEric Dumazet /* Ok, it looks like it is advisable to defer. 2250f9bfe4e6SEric Dumazet * Three cases are tracked : 2251f9bfe4e6SEric Dumazet * 1) We are cwnd-limited 2252f9bfe4e6SEric Dumazet * 2) We are rwnd-limited 2253f9bfe4e6SEric Dumazet * 3) We are application limited. 2254f9bfe4e6SEric Dumazet */ 2255f9bfe4e6SEric Dumazet if (cong_win < send_win) { 2256f9bfe4e6SEric Dumazet if (cong_win <= skb->len) { 2257ca8a2263SNeal Cardwell *is_cwnd_limited = true; 2258f9bfe4e6SEric Dumazet return true; 2259f9bfe4e6SEric Dumazet } 2260f9bfe4e6SEric Dumazet } else { 2261f9bfe4e6SEric Dumazet if (send_win <= skb->len) { 2262f9bfe4e6SEric Dumazet *is_rwnd_limited = true; 2263f9bfe4e6SEric Dumazet return true; 2264f9bfe4e6SEric Dumazet } 2265f9bfe4e6SEric Dumazet } 2266f9bfe4e6SEric Dumazet 2267f9bfe4e6SEric Dumazet /* If this packet won't get more data, do not wait. */ 2268d8ed257fSEric Dumazet if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) || 2269d8ed257fSEric Dumazet TCP_SKB_CB(skb)->eor) 2270f9bfe4e6SEric Dumazet goto send_now; 2271ca8a2263SNeal Cardwell 2272a2a385d6SEric Dumazet return true; 2273ae8064acSJohn Heffner 2274ae8064acSJohn Heffner send_now: 2275a2a385d6SEric Dumazet return false; 2276c1b4a7e6SDavid S. Miller } 2277c1b4a7e6SDavid S. Miller 227805cbc0dbSFan Du static inline void tcp_mtu_check_reprobe(struct sock *sk) 227905cbc0dbSFan Du { 228005cbc0dbSFan Du struct inet_connection_sock *icsk = inet_csk(sk); 228105cbc0dbSFan Du struct tcp_sock *tp = tcp_sk(sk); 228205cbc0dbSFan Du struct net *net = sock_net(sk); 228305cbc0dbSFan Du u32 interval; 228405cbc0dbSFan Du s32 delta; 228505cbc0dbSFan Du 228605cbc0dbSFan Du interval = net->ipv4.sysctl_tcp_probe_interval; 2287c74df29aSEric Dumazet delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp; 228805cbc0dbSFan Du if (unlikely(delta >= interval * HZ)) { 228905cbc0dbSFan Du int mss = tcp_current_mss(sk); 229005cbc0dbSFan Du 229105cbc0dbSFan Du /* Update current search range */ 229205cbc0dbSFan Du icsk->icsk_mtup.probe_size = 0; 229305cbc0dbSFan Du icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + 229405cbc0dbSFan Du sizeof(struct tcphdr) + 229505cbc0dbSFan Du icsk->icsk_af_ops->net_header_len; 229605cbc0dbSFan Du icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss); 229705cbc0dbSFan Du 229805cbc0dbSFan Du /* Update probe time stamp */ 2299c74df29aSEric Dumazet icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; 230005cbc0dbSFan Du } 230105cbc0dbSFan Du } 230205cbc0dbSFan Du 2303808cf9e3SIlya Lesokhin static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len) 2304808cf9e3SIlya Lesokhin { 2305808cf9e3SIlya Lesokhin struct sk_buff *skb, *next; 2306808cf9e3SIlya Lesokhin 2307808cf9e3SIlya Lesokhin skb = tcp_send_head(sk); 2308808cf9e3SIlya Lesokhin tcp_for_write_queue_from_safe(skb, next, sk) { 2309808cf9e3SIlya Lesokhin if (len <= skb->len) 2310808cf9e3SIlya Lesokhin break; 2311808cf9e3SIlya Lesokhin 23129b65b17dSTalal Ahmad if (unlikely(TCP_SKB_CB(skb)->eor) || 23139b65b17dSTalal Ahmad tcp_has_tx_tstamp(skb) || 23149b65b17dSTalal Ahmad !skb_pure_zcopy_same(skb, next)) 2315808cf9e3SIlya Lesokhin return false; 2316808cf9e3SIlya Lesokhin 2317808cf9e3SIlya Lesokhin len -= skb->len; 2318808cf9e3SIlya Lesokhin } 2319808cf9e3SIlya Lesokhin 2320808cf9e3SIlya Lesokhin return true; 2321808cf9e3SIlya Lesokhin } 2322808cf9e3SIlya Lesokhin 23235d424d5aSJohn Heffner /* Create a new MTU probe if we are ready. 232467edfef7SAndi Kleen * MTU probe is regularly attempting to increase the path MTU by 232567edfef7SAndi Kleen * deliberately sending larger packets. This discovers routing 232667edfef7SAndi Kleen * changes resulting in larger path MTUs. 232767edfef7SAndi Kleen * 23285d424d5aSJohn Heffner * Returns 0 if we should wait to probe (no cwnd available), 23295d424d5aSJohn Heffner * 1 if a probe was sent, 2330056834d9SIlpo Järvinen * -1 otherwise 2331056834d9SIlpo Järvinen */ 23325d424d5aSJohn Heffner static int tcp_mtu_probe(struct sock *sk) 23335d424d5aSJohn Heffner { 23345d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 233512a59abcSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 23365d424d5aSJohn Heffner struct sk_buff *skb, *nskb, *next; 23376b58e0a5SFan Du struct net *net = sock_net(sk); 23385d424d5aSJohn Heffner int probe_size; 233991cc17c0SIlpo Järvinen int size_needed; 234012a59abcSEric Dumazet int copy, len; 23415d424d5aSJohn Heffner int mss_now; 23426b58e0a5SFan Du int interval; 23435d424d5aSJohn Heffner 23445d424d5aSJohn Heffner /* Not currently probing/verifying, 23455d424d5aSJohn Heffner * not in recovery, 23465d424d5aSJohn Heffner * have enough cwnd, and 234712a59abcSEric Dumazet * not SACKing (the variable headers throw things off) 234812a59abcSEric Dumazet */ 234912a59abcSEric Dumazet if (likely(!icsk->icsk_mtup.enabled || 23505d424d5aSJohn Heffner icsk->icsk_mtup.probe_size || 23515d424d5aSJohn Heffner inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 235240570375SEric Dumazet tcp_snd_cwnd(tp) < 11 || 235312a59abcSEric Dumazet tp->rx_opt.num_sacks || tp->rx_opt.dsack)) 23545d424d5aSJohn Heffner return -1; 23555d424d5aSJohn Heffner 23566b58e0a5SFan Du /* Use binary search for probe_size between tcp_mss_base, 23576b58e0a5SFan Du * and current mss_clamp. if (search_high - search_low) 23586b58e0a5SFan Du * smaller than a threshold, backoff from probing. 23596b58e0a5SFan Du */ 23600c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 23616b58e0a5SFan Du probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high + 23626b58e0a5SFan Du icsk->icsk_mtup.search_low) >> 1); 236391cc17c0SIlpo Järvinen size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; 23646b58e0a5SFan Du interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low; 236505cbc0dbSFan Du /* When misfortune happens, we are reprobing actively, 236605cbc0dbSFan Du * and then reprobe timer has expired. We stick with current 236705cbc0dbSFan Du * probing process by not resetting search range to its orignal. 236805cbc0dbSFan Du */ 23696b58e0a5SFan Du if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) || 237005cbc0dbSFan Du interval < net->ipv4.sysctl_tcp_probe_threshold) { 237105cbc0dbSFan Du /* Check whether enough time has elaplased for 237205cbc0dbSFan Du * another round of probing. 237305cbc0dbSFan Du */ 237405cbc0dbSFan Du tcp_mtu_check_reprobe(sk); 23755d424d5aSJohn Heffner return -1; 23765d424d5aSJohn Heffner } 23775d424d5aSJohn Heffner 23785d424d5aSJohn Heffner /* Have enough data in the send queue to probe? */ 23797f9c33e5SIlpo Järvinen if (tp->write_seq - tp->snd_nxt < size_needed) 23805d424d5aSJohn Heffner return -1; 23815d424d5aSJohn Heffner 238291cc17c0SIlpo Järvinen if (tp->snd_wnd < size_needed) 23835d424d5aSJohn Heffner return -1; 238490840defSIlpo Järvinen if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) 23855d424d5aSJohn Heffner return 0; 23865d424d5aSJohn Heffner 2387d67c58e9SIlpo Järvinen /* Do we need to wait to drain cwnd? With none in flight, don't stall */ 238840570375SEric Dumazet if (tcp_packets_in_flight(tp) + 2 > tcp_snd_cwnd(tp)) { 2389d67c58e9SIlpo Järvinen if (!tcp_packets_in_flight(tp)) 23905d424d5aSJohn Heffner return -1; 23915d424d5aSJohn Heffner else 23925d424d5aSJohn Heffner return 0; 23935d424d5aSJohn Heffner } 23945d424d5aSJohn Heffner 2395808cf9e3SIlya Lesokhin if (!tcp_can_coalesce_send_queue_head(sk, probe_size)) 2396808cf9e3SIlya Lesokhin return -1; 2397808cf9e3SIlya Lesokhin 23985d424d5aSJohn Heffner /* We're allowed to probe. Build it now. */ 2399f8dd3b8dSEric Dumazet nskb = tcp_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false); 240051456b29SIan Morris if (!nskb) 24015d424d5aSJohn Heffner return -1; 2402ab4e846aSEric Dumazet sk_wmem_queued_add(sk, nskb->truesize); 24033ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 24045d424d5aSJohn Heffner 2405fe067e8aSDavid S. Miller skb = tcp_send_head(sk); 240641477662SJakub Kicinski skb_copy_decrypted(nskb, skb); 24075a369ca6SPaolo Abeni mptcp_skb_ext_copy(nskb, skb); 24085d424d5aSJohn Heffner 24095d424d5aSJohn Heffner TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 24105d424d5aSJohn Heffner TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 24114de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; 24125d424d5aSJohn Heffner 241350c4817eSIlpo Järvinen tcp_insert_write_queue_before(nskb, skb, sk); 24142b7cda9cSEric Dumazet tcp_highest_sack_replace(sk, skb, nskb); 241550c4817eSIlpo Järvinen 24165d424d5aSJohn Heffner len = 0; 2417234b6860SIlpo Järvinen tcp_for_write_queue_from_safe(skb, next, sk) { 24185d424d5aSJohn Heffner copy = min_t(int, skb->len, probe_size - len); 24195d424d5aSJohn Heffner skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); 24205d424d5aSJohn Heffner 24215d424d5aSJohn Heffner if (skb->len <= copy) { 24225d424d5aSJohn Heffner /* We've eaten all the data from this skb. 24235d424d5aSJohn Heffner * Throw it away. */ 24244de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 2425808cf9e3SIlya Lesokhin /* If this is the last SKB we copy and eor is set 2426808cf9e3SIlya Lesokhin * we need to propagate it to the new skb. 2427808cf9e3SIlya Lesokhin */ 2428808cf9e3SIlya Lesokhin TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor; 2429888a5c53SWillem de Bruijn tcp_skb_collapse_tstamp(nskb, skb); 2430fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 243103271f3aSTalal Ahmad tcp_wmem_free_skb(sk, skb); 24325d424d5aSJohn Heffner } else { 24334de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & 2434a3433f35SChangli Gao ~(TCPHDR_FIN|TCPHDR_PSH); 24355d424d5aSJohn Heffner if (!skb_shinfo(skb)->nr_frags) { 24365d424d5aSJohn Heffner skb_pull(skb, copy); 24375d424d5aSJohn Heffner } else { 24385d424d5aSJohn Heffner __pskb_trim_head(skb, copy); 24395bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 24405d424d5aSJohn Heffner } 24415d424d5aSJohn Heffner TCP_SKB_CB(skb)->seq += copy; 24425d424d5aSJohn Heffner } 24435d424d5aSJohn Heffner 24445d424d5aSJohn Heffner len += copy; 2445234b6860SIlpo Järvinen 2446234b6860SIlpo Järvinen if (len >= probe_size) 2447234b6860SIlpo Järvinen break; 24485d424d5aSJohn Heffner } 24495bbb432cSEric Dumazet tcp_init_tso_segs(nskb, nskb->len); 24505d424d5aSJohn Heffner 24515d424d5aSJohn Heffner /* We're ready to send. If this fails, the probe will 24527faee5c0SEric Dumazet * be resegmented into mss-sized pieces by tcp_write_xmit(). 24537faee5c0SEric Dumazet */ 24545d424d5aSJohn Heffner if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 24555d424d5aSJohn Heffner /* Decrement cwnd here because we are sending 24565d424d5aSJohn Heffner * effectively two packets. */ 245740570375SEric Dumazet tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) - 1); 245866f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, nskb); 24595d424d5aSJohn Heffner 24605d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 24610e7b1368SJohn Heffner tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 24620e7b1368SJohn Heffner tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 24635d424d5aSJohn Heffner 24645d424d5aSJohn Heffner return 1; 24655d424d5aSJohn Heffner } 24665d424d5aSJohn Heffner 24675d424d5aSJohn Heffner return -1; 24685d424d5aSJohn Heffner } 24695d424d5aSJohn Heffner 2470864e5c09SEric Dumazet static bool tcp_pacing_check(struct sock *sk) 2471218af599SEric Dumazet { 2472864e5c09SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 2473864e5c09SEric Dumazet 2474864e5c09SEric Dumazet if (!tcp_needs_internal_pacing(sk)) 2475864e5c09SEric Dumazet return false; 2476864e5c09SEric Dumazet 2477864e5c09SEric Dumazet if (tp->tcp_wstamp_ns <= tp->tcp_clock_cache) 2478864e5c09SEric Dumazet return false; 2479864e5c09SEric Dumazet 2480864e5c09SEric Dumazet if (!hrtimer_is_queued(&tp->pacing_timer)) { 2481864e5c09SEric Dumazet hrtimer_start(&tp->pacing_timer, 2482864e5c09SEric Dumazet ns_to_ktime(tp->tcp_wstamp_ns), 2483864e5c09SEric Dumazet HRTIMER_MODE_ABS_PINNED_SOFT); 2484864e5c09SEric Dumazet sock_hold(sk); 2485864e5c09SEric Dumazet } 2486864e5c09SEric Dumazet return true; 2487218af599SEric Dumazet } 2488218af599SEric Dumazet 2489f9616c35SEric Dumazet /* TCP Small Queues : 2490f9616c35SEric Dumazet * Control number of packets in qdisc/devices to two packets / or ~1 ms. 2491f9616c35SEric Dumazet * (These limits are doubled for retransmits) 2492f9616c35SEric Dumazet * This allows for : 2493f9616c35SEric Dumazet * - better RTT estimation and ACK scheduling 2494f9616c35SEric Dumazet * - faster recovery 2495f9616c35SEric Dumazet * - high rates 2496f9616c35SEric Dumazet * Alas, some drivers / subsystems require a fair amount 2497f9616c35SEric Dumazet * of queued bytes to ensure line rate. 2498f9616c35SEric Dumazet * One example is wifi aggregation (802.11 AMPDU) 2499f9616c35SEric Dumazet */ 2500f9616c35SEric Dumazet static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb, 2501f9616c35SEric Dumazet unsigned int factor) 2502f9616c35SEric Dumazet { 250376a9ebe8SEric Dumazet unsigned long limit; 2504f9616c35SEric Dumazet 250576a9ebe8SEric Dumazet limit = max_t(unsigned long, 250676a9ebe8SEric Dumazet 2 * skb->truesize, 25077c68fa2bSEric Dumazet sk->sk_pacing_rate >> READ_ONCE(sk->sk_pacing_shift)); 2508c73e5807SEric Dumazet if (sk->sk_pacing_status == SK_PACING_NONE) 250976a9ebe8SEric Dumazet limit = min_t(unsigned long, limit, 25109184d8bbSEric Dumazet sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes); 2511f9616c35SEric Dumazet limit <<= factor; 2512f9616c35SEric Dumazet 2513a842fe14SEric Dumazet if (static_branch_unlikely(&tcp_tx_delay_enabled) && 2514a842fe14SEric Dumazet tcp_sk(sk)->tcp_tx_delay) { 2515a842fe14SEric Dumazet u64 extra_bytes = (u64)sk->sk_pacing_rate * tcp_sk(sk)->tcp_tx_delay; 2516a842fe14SEric Dumazet 2517a842fe14SEric Dumazet /* TSQ is based on skb truesize sum (sk_wmem_alloc), so we 2518a842fe14SEric Dumazet * approximate our needs assuming an ~100% skb->truesize overhead. 2519a842fe14SEric Dumazet * USEC_PER_SEC is approximated by 2^20. 2520a842fe14SEric Dumazet * do_div(extra_bytes, USEC_PER_SEC/2) is replaced by a right shift. 2521a842fe14SEric Dumazet */ 2522a842fe14SEric Dumazet extra_bytes >>= (20 - 1); 2523a842fe14SEric Dumazet limit += extra_bytes; 2524a842fe14SEric Dumazet } 252514afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) > limit) { 252675c119afSEric Dumazet /* Always send skb if rtx queue is empty. 252775eefc6cSEric Dumazet * No need to wait for TX completion to call us back, 252875eefc6cSEric Dumazet * after softirq/tasklet schedule. 252975eefc6cSEric Dumazet * This helps when TX completions are delayed too much. 253075eefc6cSEric Dumazet */ 253175c119afSEric Dumazet if (tcp_rtx_queue_empty(sk)) 253275eefc6cSEric Dumazet return false; 253375eefc6cSEric Dumazet 25347aa5470cSEric Dumazet set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags); 2535f9616c35SEric Dumazet /* It is possible TX completion already happened 2536f9616c35SEric Dumazet * before we set TSQ_THROTTLED, so we must 2537f9616c35SEric Dumazet * test again the condition. 2538f9616c35SEric Dumazet */ 2539f9616c35SEric Dumazet smp_mb__after_atomic(); 2540ce8299b6SEric Dumazet if (refcount_read(&sk->sk_wmem_alloc) > limit) 2541f9616c35SEric Dumazet return true; 2542f9616c35SEric Dumazet } 2543f9616c35SEric Dumazet return false; 2544f9616c35SEric Dumazet } 2545f9616c35SEric Dumazet 254605b055e8SFrancis Yan static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new) 254705b055e8SFrancis Yan { 2548628174ccSEric Dumazet const u32 now = tcp_jiffies32; 2549efe967cdSArnd Bergmann enum tcp_chrono old = tp->chrono_type; 255005b055e8SFrancis Yan 2551efe967cdSArnd Bergmann if (old > TCP_CHRONO_UNSPEC) 2552efe967cdSArnd Bergmann tp->chrono_stat[old - 1] += now - tp->chrono_start; 255305b055e8SFrancis Yan tp->chrono_start = now; 255405b055e8SFrancis Yan tp->chrono_type = new; 255505b055e8SFrancis Yan } 255605b055e8SFrancis Yan 255705b055e8SFrancis Yan void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type) 255805b055e8SFrancis Yan { 255905b055e8SFrancis Yan struct tcp_sock *tp = tcp_sk(sk); 256005b055e8SFrancis Yan 256105b055e8SFrancis Yan /* If there are multiple conditions worthy of tracking in a 25620f87230dSFrancis Yan * chronograph then the highest priority enum takes precedence 25630f87230dSFrancis Yan * over the other conditions. So that if something "more interesting" 256405b055e8SFrancis Yan * starts happening, stop the previous chrono and start a new one. 256505b055e8SFrancis Yan */ 256605b055e8SFrancis Yan if (type > tp->chrono_type) 256705b055e8SFrancis Yan tcp_chrono_set(tp, type); 256805b055e8SFrancis Yan } 256905b055e8SFrancis Yan 257005b055e8SFrancis Yan void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type) 257105b055e8SFrancis Yan { 257205b055e8SFrancis Yan struct tcp_sock *tp = tcp_sk(sk); 257305b055e8SFrancis Yan 25740f87230dSFrancis Yan 25750f87230dSFrancis Yan /* There are multiple conditions worthy of tracking in a 25760f87230dSFrancis Yan * chronograph, so that the highest priority enum takes 25770f87230dSFrancis Yan * precedence over the other conditions (see tcp_chrono_start). 25780f87230dSFrancis Yan * If a condition stops, we only stop chrono tracking if 25790f87230dSFrancis Yan * it's the "most interesting" or current chrono we are 25800f87230dSFrancis Yan * tracking and starts busy chrono if we have pending data. 25810f87230dSFrancis Yan */ 258275c119afSEric Dumazet if (tcp_rtx_and_write_queues_empty(sk)) 258305b055e8SFrancis Yan tcp_chrono_set(tp, TCP_CHRONO_UNSPEC); 25840f87230dSFrancis Yan else if (type == tp->chrono_type) 25850f87230dSFrancis Yan tcp_chrono_set(tp, TCP_CHRONO_BUSY); 258605b055e8SFrancis Yan } 258705b055e8SFrancis Yan 25881da177e4SLinus Torvalds /* This routine writes packets to the network. It advances the 25891da177e4SLinus Torvalds * send_head. This happens as incoming acks open up the remote 25901da177e4SLinus Torvalds * window for us. 25911da177e4SLinus Torvalds * 2592f8269a49SIlpo Järvinen * LARGESEND note: !tcp_urg_mode is overkill, only frames between 2593f8269a49SIlpo Järvinen * snd_up-64k-mss .. snd_up cannot be large. However, taking into 2594f8269a49SIlpo Järvinen * account rare use of URG, this is not a big flaw. 2595f8269a49SIlpo Järvinen * 25966ba8a3b1SNandita Dukkipati * Send at most one packet when push_one > 0. Temporarily ignore 25976ba8a3b1SNandita Dukkipati * cwnd limit to force at most one packet out when push_one == 2. 25986ba8a3b1SNandita Dukkipati 2599a2a385d6SEric Dumazet * Returns true, if no segments are in flight and we have queued segments, 2600a2a385d6SEric Dumazet * but cannot send anything now because of SWS or another problem. 26011da177e4SLinus Torvalds */ 2602a2a385d6SEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 2603d5dd9175SIlpo Järvinen int push_one, gfp_t gfp) 26041da177e4SLinus Torvalds { 26051da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 260692df7b51SDavid S. Miller struct sk_buff *skb; 2607c1b4a7e6SDavid S. Miller unsigned int tso_segs, sent_pkts; 2608c1b4a7e6SDavid S. Miller int cwnd_quota; 26095d424d5aSJohn Heffner int result; 26105615f886SFrancis Yan bool is_cwnd_limited = false, is_rwnd_limited = false; 2611605ad7f1SEric Dumazet u32 max_segs; 26121da177e4SLinus Torvalds 2613c1b4a7e6SDavid S. Miller sent_pkts = 0; 26145d424d5aSJohn Heffner 2615ee1836aeSEric Dumazet tcp_mstamp_refresh(tp); 2616d5dd9175SIlpo Järvinen if (!push_one) { 26175d424d5aSJohn Heffner /* Do MTU probing. */ 2618d5dd9175SIlpo Järvinen result = tcp_mtu_probe(sk); 2619d5dd9175SIlpo Järvinen if (!result) { 2620a2a385d6SEric Dumazet return false; 26215d424d5aSJohn Heffner } else if (result > 0) { 26225d424d5aSJohn Heffner sent_pkts = 1; 26235d424d5aSJohn Heffner } 2624d5dd9175SIlpo Järvinen } 26255d424d5aSJohn Heffner 2626ed6e7268SNeal Cardwell max_segs = tcp_tso_segs(sk, mss_now); 2627fe067e8aSDavid S. Miller while ((skb = tcp_send_head(sk))) { 2628c8ac3774SHerbert Xu unsigned int limit; 2629c8ac3774SHerbert Xu 263079861919SEric Dumazet if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) { 263179861919SEric Dumazet /* "skb_mstamp_ns" is used as a start point for the retransmit timer */ 2632a1ac9c8aSMartin KaFai Lau tp->tcp_wstamp_ns = tp->tcp_clock_cache; 2633a1ac9c8aSMartin KaFai Lau skb_set_delivery_time(skb, tp->tcp_wstamp_ns, true); 263479861919SEric Dumazet list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue); 2635bf50b606SEric Dumazet tcp_init_tso_segs(skb, mss_now); 263679861919SEric Dumazet goto repair; /* Skip network transmission */ 263779861919SEric Dumazet } 263879861919SEric Dumazet 2639218af599SEric Dumazet if (tcp_pacing_check(sk)) 2640218af599SEric Dumazet break; 2641218af599SEric Dumazet 26425bbb432cSEric Dumazet tso_segs = tcp_init_tso_segs(skb, mss_now); 2643c1b4a7e6SDavid S. Miller BUG_ON(!tso_segs); 2644c1b4a7e6SDavid S. Miller 2645b68e9f85SHerbert Xu cwnd_quota = tcp_cwnd_test(tp, skb); 26466ba8a3b1SNandita Dukkipati if (!cwnd_quota) { 26476ba8a3b1SNandita Dukkipati if (push_one == 2) 26486ba8a3b1SNandita Dukkipati /* Force out a loss probe pkt. */ 26496ba8a3b1SNandita Dukkipati cwnd_quota = 1; 26506ba8a3b1SNandita Dukkipati else 2651b68e9f85SHerbert Xu break; 26526ba8a3b1SNandita Dukkipati } 2653b68e9f85SHerbert Xu 26545615f886SFrancis Yan if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) { 26555615f886SFrancis Yan is_rwnd_limited = true; 2656b68e9f85SHerbert Xu break; 26575615f886SFrancis Yan } 2658b68e9f85SHerbert Xu 2659d6a4e26aSEric Dumazet if (tso_segs == 1) { 2660aa93466bSDavid S. Miller if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 2661aa93466bSDavid S. Miller (tcp_skb_is_last(sk, skb) ? 2662aa93466bSDavid S. Miller nonagle : TCP_NAGLE_PUSH)))) 2663aa93466bSDavid S. Miller break; 2664c1b4a7e6SDavid S. Miller } else { 2665ca8a2263SNeal Cardwell if (!push_one && 2666605ad7f1SEric Dumazet tcp_tso_should_defer(sk, skb, &is_cwnd_limited, 2667f9bfe4e6SEric Dumazet &is_rwnd_limited, max_segs)) 2668aa93466bSDavid S. Miller break; 2669c1b4a7e6SDavid S. Miller } 2670aa93466bSDavid S. Miller 2671605ad7f1SEric Dumazet limit = mss_now; 2672d6a4e26aSEric Dumazet if (tso_segs > 1 && !tcp_urg_mode(tp)) 2673605ad7f1SEric Dumazet limit = tcp_mss_split_point(sk, skb, mss_now, 2674605ad7f1SEric Dumazet min_t(unsigned int, 2675605ad7f1SEric Dumazet cwnd_quota, 2676605ad7f1SEric Dumazet max_segs), 2677605ad7f1SEric Dumazet nonagle); 2678605ad7f1SEric Dumazet 2679605ad7f1SEric Dumazet if (skb->len > limit && 268056483341SEric Dumazet unlikely(tso_fragment(sk, skb, limit, mss_now, gfp))) 2681605ad7f1SEric Dumazet break; 2682605ad7f1SEric Dumazet 2683f9616c35SEric Dumazet if (tcp_small_queue_check(sk, skb, 0)) 268446d3ceabSEric Dumazet break; 2685c9eeec26SEric Dumazet 26861f85e626SEric Dumazet /* Argh, we hit an empty skb(), presumably a thread 26871f85e626SEric Dumazet * is sleeping in sendmsg()/sk_stream_wait_memory(). 26881f85e626SEric Dumazet * We do not want to send a pure-ack packet and have 26891f85e626SEric Dumazet * a strange looking rtx queue with empty packet(s). 26901f85e626SEric Dumazet */ 26911f85e626SEric Dumazet if (TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq) 26921f85e626SEric Dumazet break; 26931f85e626SEric Dumazet 2694d5dd9175SIlpo Järvinen if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) 26951da177e4SLinus Torvalds break; 26961da177e4SLinus Torvalds 2697ec342325SAndrew Vagin repair: 26981da177e4SLinus Torvalds /* Advance the send_head. This one is sent out. 26991da177e4SLinus Torvalds * This call will increment packets_out. 27001da177e4SLinus Torvalds */ 270166f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 27021da177e4SLinus Torvalds 27031da177e4SLinus Torvalds tcp_minshall_update(tp, mss_now, skb); 2704a262f0cdSNandita Dukkipati sent_pkts += tcp_skb_pcount(skb); 2705d5dd9175SIlpo Järvinen 2706d5dd9175SIlpo Järvinen if (push_one) 2707d5dd9175SIlpo Järvinen break; 27081da177e4SLinus Torvalds } 27091da177e4SLinus Torvalds 27105615f886SFrancis Yan if (is_rwnd_limited) 27115615f886SFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED); 27125615f886SFrancis Yan else 27135615f886SFrancis Yan tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED); 27145615f886SFrancis Yan 271540570375SEric Dumazet is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tcp_snd_cwnd(tp)); 2716299bcb55SNeal Cardwell if (likely(sent_pkts || is_cwnd_limited)) 2717299bcb55SNeal Cardwell tcp_cwnd_validate(sk, is_cwnd_limited); 2718299bcb55SNeal Cardwell 2719aa93466bSDavid S. Miller if (likely(sent_pkts)) { 2720684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 2721684bad11SYuchung Cheng tp->prr_out += sent_pkts; 27226ba8a3b1SNandita Dukkipati 27236ba8a3b1SNandita Dukkipati /* Send one loss probe per tail loss episode. */ 27246ba8a3b1SNandita Dukkipati if (push_one != 2) 2725ed66dfafSNeal Cardwell tcp_schedule_loss_probe(sk, false); 2726a2a385d6SEric Dumazet return false; 27271da177e4SLinus Torvalds } 272875c119afSEric Dumazet return !tp->packets_out && !tcp_write_queue_empty(sk); 27296ba8a3b1SNandita Dukkipati } 27306ba8a3b1SNandita Dukkipati 2731ed66dfafSNeal Cardwell bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto) 27326ba8a3b1SNandita Dukkipati { 27336ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 27346ba8a3b1SNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 2735a2815817SNeal Cardwell u32 timeout, rto_delta_us; 27362ae21cf5SEric Dumazet int early_retrans; 27376ba8a3b1SNandita Dukkipati 27386ba8a3b1SNandita Dukkipati /* Don't do any loss probe on a Fast Open connection before 3WHS 27396ba8a3b1SNandita Dukkipati * finishes. 27406ba8a3b1SNandita Dukkipati */ 2741d983ea6fSEric Dumazet if (rcu_access_pointer(tp->fastopen_rsk)) 27426ba8a3b1SNandita Dukkipati return false; 27436ba8a3b1SNandita Dukkipati 27442ae21cf5SEric Dumazet early_retrans = sock_net(sk)->ipv4.sysctl_tcp_early_retrans; 27456ba8a3b1SNandita Dukkipati /* Schedule a loss probe in 2*RTT for SACK capable connections 2746b4f70c3dSNeal Cardwell * not in loss recovery, that are either limited by cwnd or application. 27476ba8a3b1SNandita Dukkipati */ 27482ae21cf5SEric Dumazet if ((early_retrans != 3 && early_retrans != 4) || 2749bec41a11SYuchung Cheng !tp->packets_out || !tcp_is_sack(tp) || 2750b4f70c3dSNeal Cardwell (icsk->icsk_ca_state != TCP_CA_Open && 2751b4f70c3dSNeal Cardwell icsk->icsk_ca_state != TCP_CA_CWR)) 27526ba8a3b1SNandita Dukkipati return false; 27536ba8a3b1SNandita Dukkipati 2754bb4d991aSYuchung Cheng /* Probe timeout is 2*rtt. Add minimum RTO to account 2755f9b99582SYuchung Cheng * for delayed ack when there's one outstanding packet. If no RTT 2756f9b99582SYuchung Cheng * sample is available then probe after TCP_TIMEOUT_INIT. 27576ba8a3b1SNandita Dukkipati */ 2758bb4d991aSYuchung Cheng if (tp->srtt_us) { 2759bb4d991aSYuchung Cheng timeout = usecs_to_jiffies(tp->srtt_us >> 2); 27606ba8a3b1SNandita Dukkipati if (tp->packets_out == 1) 2761bb4d991aSYuchung Cheng timeout += TCP_RTO_MIN; 2762bb4d991aSYuchung Cheng else 2763bb4d991aSYuchung Cheng timeout += TCP_TIMEOUT_MIN; 2764bb4d991aSYuchung Cheng } else { 2765bb4d991aSYuchung Cheng timeout = TCP_TIMEOUT_INIT; 2766bb4d991aSYuchung Cheng } 27676ba8a3b1SNandita Dukkipati 2768a2815817SNeal Cardwell /* If the RTO formula yields an earlier time, then use that time. */ 2769ed66dfafSNeal Cardwell rto_delta_us = advancing_rto ? 2770ed66dfafSNeal Cardwell jiffies_to_usecs(inet_csk(sk)->icsk_rto) : 2771ed66dfafSNeal Cardwell tcp_rto_delta_us(sk); /* How far in future is RTO? */ 2772a2815817SNeal Cardwell if (rto_delta_us > 0) 2773a2815817SNeal Cardwell timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us)); 27746ba8a3b1SNandita Dukkipati 27758dc242adSEric Dumazet tcp_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, TCP_RTO_MAX); 27766ba8a3b1SNandita Dukkipati return true; 27776ba8a3b1SNandita Dukkipati } 27786ba8a3b1SNandita Dukkipati 27791f3279aeSEric Dumazet /* Thanks to skb fast clones, we can detect if a prior transmit of 27801f3279aeSEric Dumazet * a packet is still in a qdisc or driver queue. 27811f3279aeSEric Dumazet * In this case, there is very little point doing a retransmit ! 27821f3279aeSEric Dumazet */ 2783f4dae54eSEric Dumazet static bool skb_still_in_host_queue(struct sock *sk, 27841f3279aeSEric Dumazet const struct sk_buff *skb) 27851f3279aeSEric Dumazet { 278639bb5e62SEric Dumazet if (unlikely(skb_fclone_busy(sk, skb))) { 2787f4dae54eSEric Dumazet set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags); 2788f4dae54eSEric Dumazet smp_mb__after_atomic(); 2789f4dae54eSEric Dumazet if (skb_fclone_busy(sk, skb)) { 2790c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), 27911f3279aeSEric Dumazet LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES); 27921f3279aeSEric Dumazet return true; 27931f3279aeSEric Dumazet } 2794f4dae54eSEric Dumazet } 27951f3279aeSEric Dumazet return false; 27961f3279aeSEric Dumazet } 27971f3279aeSEric Dumazet 2798b340b264SYuchung Cheng /* When probe timeout (PTO) fires, try send a new segment if possible, else 27996ba8a3b1SNandita Dukkipati * retransmit the last segment. 28006ba8a3b1SNandita Dukkipati */ 28016ba8a3b1SNandita Dukkipati void tcp_send_loss_probe(struct sock *sk) 28026ba8a3b1SNandita Dukkipati { 28039b717a8dSNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 28046ba8a3b1SNandita Dukkipati struct sk_buff *skb; 28056ba8a3b1SNandita Dukkipati int pcount; 28066ba8a3b1SNandita Dukkipati int mss = tcp_current_mss(sk); 28076ba8a3b1SNandita Dukkipati 280876be93fcSYuchung Cheng /* At most one outstanding TLP */ 280976be93fcSYuchung Cheng if (tp->tlp_high_seq) 281076be93fcSYuchung Cheng goto rearm_timer; 281176be93fcSYuchung Cheng 281276be93fcSYuchung Cheng tp->tlp_retrans = 0; 2813b340b264SYuchung Cheng skb = tcp_send_head(sk); 281475c119afSEric Dumazet if (skb && tcp_snd_wnd_test(tp, skb, mss)) { 2815b340b264SYuchung Cheng pcount = tp->packets_out; 2816b340b264SYuchung Cheng tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC); 2817b340b264SYuchung Cheng if (tp->packets_out > pcount) 2818b340b264SYuchung Cheng goto probe_sent; 28196ba8a3b1SNandita Dukkipati goto rearm_timer; 28206ba8a3b1SNandita Dukkipati } 282175c119afSEric Dumazet skb = skb_rb_last(&sk->tcp_rtx_queue); 2822b2b7af86SYuchung Cheng if (unlikely(!skb)) { 2823b2b7af86SYuchung Cheng WARN_ONCE(tp->packets_out, 2824b2b7af86SYuchung Cheng "invalid inflight: %u state %u cwnd %u mss %d\n", 282540570375SEric Dumazet tp->packets_out, sk->sk_state, tcp_snd_cwnd(tp), mss); 2826b2b7af86SYuchung Cheng inet_csk(sk)->icsk_pending = 0; 2827b2b7af86SYuchung Cheng return; 2828b2b7af86SYuchung Cheng } 28296ba8a3b1SNandita Dukkipati 28301f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 28311f3279aeSEric Dumazet goto rearm_timer; 28321f3279aeSEric Dumazet 28336ba8a3b1SNandita Dukkipati pcount = tcp_skb_pcount(skb); 28346ba8a3b1SNandita Dukkipati if (WARN_ON(!pcount)) 28356ba8a3b1SNandita Dukkipati goto rearm_timer; 28366ba8a3b1SNandita Dukkipati 28376ba8a3b1SNandita Dukkipati if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) { 283875c119afSEric Dumazet if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, 283975c119afSEric Dumazet (pcount - 1) * mss, mss, 28406cc55e09SOctavian Purdila GFP_ATOMIC))) 28416ba8a3b1SNandita Dukkipati goto rearm_timer; 284275c119afSEric Dumazet skb = skb_rb_next(skb); 28436ba8a3b1SNandita Dukkipati } 28446ba8a3b1SNandita Dukkipati 28456ba8a3b1SNandita Dukkipati if (WARN_ON(!skb || !tcp_skb_pcount(skb))) 28466ba8a3b1SNandita Dukkipati goto rearm_timer; 28476ba8a3b1SNandita Dukkipati 284810d3be56SEric Dumazet if (__tcp_retransmit_skb(sk, skb, 1)) 2849b340b264SYuchung Cheng goto rearm_timer; 28506ba8a3b1SNandita Dukkipati 285176be93fcSYuchung Cheng tp->tlp_retrans = 1; 285276be93fcSYuchung Cheng 285376be93fcSYuchung Cheng probe_sent: 28549b717a8dSNandita Dukkipati /* Record snd_nxt for loss detection. */ 28559b717a8dSNandita Dukkipati tp->tlp_high_seq = tp->snd_nxt; 28569b717a8dSNandita Dukkipati 2857c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES); 2858fcd16c0aSYuchung Cheng /* Reset s.t. tcp_rearm_rto will restart timer from now */ 2859fcd16c0aSYuchung Cheng inet_csk(sk)->icsk_pending = 0; 2860b340b264SYuchung Cheng rearm_timer: 2861fcd16c0aSYuchung Cheng tcp_rearm_rto(sk); 28621da177e4SLinus Torvalds } 28631da177e4SLinus Torvalds 2864a762a980SDavid S. Miller /* Push out any pending frames which were held back due to 2865a762a980SDavid S. Miller * TCP_CORK or attempt at coalescing tiny packets. 2866a762a980SDavid S. Miller * The socket must be locked by the caller. 2867a762a980SDavid S. Miller */ 28689e412ba7SIlpo Järvinen void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 28699e412ba7SIlpo Järvinen int nonagle) 2870a762a980SDavid S. Miller { 2871726e07a8SIlpo Järvinen /* If we are closed, the bytes will have to remain here. 2872726e07a8SIlpo Järvinen * In time closedown will finish, we empty the write queue and 2873726e07a8SIlpo Järvinen * all will be happy. 2874726e07a8SIlpo Järvinen */ 2875726e07a8SIlpo Järvinen if (unlikely(sk->sk_state == TCP_CLOSE)) 2876726e07a8SIlpo Järvinen return; 2877726e07a8SIlpo Järvinen 287899a1dec7SMel Gorman if (tcp_write_xmit(sk, cur_mss, nonagle, 0, 28797450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC))) 28809e412ba7SIlpo Järvinen tcp_check_probe_timer(sk); 2881a762a980SDavid S. Miller } 2882a762a980SDavid S. Miller 2883c1b4a7e6SDavid S. Miller /* Send _single_ skb sitting at the send head. This function requires 2884c1b4a7e6SDavid S. Miller * true push pending frames to setup probe timer etc. 2885c1b4a7e6SDavid S. Miller */ 2886c1b4a7e6SDavid S. Miller void tcp_push_one(struct sock *sk, unsigned int mss_now) 2887c1b4a7e6SDavid S. Miller { 2888fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 2889c1b4a7e6SDavid S. Miller 2890c1b4a7e6SDavid S. Miller BUG_ON(!skb || skb->len < mss_now); 2891c1b4a7e6SDavid S. Miller 2892d5dd9175SIlpo Järvinen tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); 2893c1b4a7e6SDavid S. Miller } 2894c1b4a7e6SDavid S. Miller 28951da177e4SLinus Torvalds /* This function returns the amount that we can raise the 28961da177e4SLinus Torvalds * usable window based on the following constraints 28971da177e4SLinus Torvalds * 28981da177e4SLinus Torvalds * 1. The window can never be shrunk once it is offered (RFC 793) 28991da177e4SLinus Torvalds * 2. We limit memory per socket 29001da177e4SLinus Torvalds * 29011da177e4SLinus Torvalds * RFC 1122: 29021da177e4SLinus Torvalds * "the suggested [SWS] avoidance algorithm for the receiver is to keep 29031da177e4SLinus Torvalds * RECV.NEXT + RCV.WIN fixed until: 29041da177e4SLinus Torvalds * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 29051da177e4SLinus Torvalds * 29061da177e4SLinus Torvalds * i.e. don't raise the right edge of the window until you can raise 29071da177e4SLinus Torvalds * it at least MSS bytes. 29081da177e4SLinus Torvalds * 29091da177e4SLinus Torvalds * Unfortunately, the recommended algorithm breaks header prediction, 29101da177e4SLinus Torvalds * since header prediction assumes th->window stays fixed. 29111da177e4SLinus Torvalds * 29121da177e4SLinus Torvalds * Strictly speaking, keeping th->window fixed violates the receiver 29131da177e4SLinus Torvalds * side SWS prevention criteria. The problem is that under this rule 29141da177e4SLinus Torvalds * a stream of single byte packets will cause the right side of the 29151da177e4SLinus Torvalds * window to always advance by a single byte. 29161da177e4SLinus Torvalds * 29171da177e4SLinus Torvalds * Of course, if the sender implements sender side SWS prevention 29181da177e4SLinus Torvalds * then this will not be a problem. 29191da177e4SLinus Torvalds * 29201da177e4SLinus Torvalds * BSD seems to make the following compromise: 29211da177e4SLinus Torvalds * 29221da177e4SLinus Torvalds * If the free space is less than the 1/4 of the maximum 29231da177e4SLinus Torvalds * space available and the free space is less than 1/2 mss, 29241da177e4SLinus Torvalds * then set the window to 0. 29251da177e4SLinus Torvalds * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 29261da177e4SLinus Torvalds * Otherwise, just prevent the window from shrinking 29271da177e4SLinus Torvalds * and from being larger than the largest representable value. 29281da177e4SLinus Torvalds * 29291da177e4SLinus Torvalds * This prevents incremental opening of the window in the regime 29301da177e4SLinus Torvalds * where TCP is limited by the speed of the reader side taking 29311da177e4SLinus Torvalds * data out of the TCP receive queue. It does nothing about 29321da177e4SLinus Torvalds * those cases where the window is constrained on the sender side 29331da177e4SLinus Torvalds * because the pipeline is full. 29341da177e4SLinus Torvalds * 29351da177e4SLinus Torvalds * BSD also seems to "accidentally" limit itself to windows that are a 29361da177e4SLinus Torvalds * multiple of MSS, at least until the free space gets quite small. 29371da177e4SLinus Torvalds * This would appear to be a side effect of the mbuf implementation. 29381da177e4SLinus Torvalds * Combining these two algorithms results in the observed behavior 29391da177e4SLinus Torvalds * of having a fixed window size at almost all times. 29401da177e4SLinus Torvalds * 29411da177e4SLinus Torvalds * Below we obtain similar behavior by forcing the offered window to 29421da177e4SLinus Torvalds * a multiple of the mss when it is feasible to do so. 29431da177e4SLinus Torvalds * 29441da177e4SLinus Torvalds * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 29451da177e4SLinus Torvalds * Regular options like TIMESTAMP are taken into account. 29461da177e4SLinus Torvalds */ 29471da177e4SLinus Torvalds u32 __tcp_select_window(struct sock *sk) 29481da177e4SLinus Torvalds { 2949463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 29501da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2951caa20d9aSStephen Hemminger /* MSS for the peer's data. Previous versions used mss_clamp 29521da177e4SLinus Torvalds * here. I don't know if the value based on our guesses 29531da177e4SLinus Torvalds * of peer's MSS is better for the performance. It's more correct 29541da177e4SLinus Torvalds * but may be worse for the performance because of rcv_mss 29551da177e4SLinus Torvalds * fluctuations. --SAW 1998/11/1 29561da177e4SLinus Torvalds */ 2957463c84b9SArnaldo Carvalho de Melo int mss = icsk->icsk_ack.rcv_mss; 29581da177e4SLinus Torvalds int free_space = tcp_space(sk); 295986c1a045SFlorian Westphal int allowed_space = tcp_full_space(sk); 2960071c8ed6SFlorian Westphal int full_space, window; 2961071c8ed6SFlorian Westphal 2962071c8ed6SFlorian Westphal if (sk_is_mptcp(sk)) 2963071c8ed6SFlorian Westphal mptcp_space(sk, &free_space, &allowed_space); 2964071c8ed6SFlorian Westphal 2965071c8ed6SFlorian Westphal full_space = min_t(int, tp->window_clamp, allowed_space); 29661da177e4SLinus Torvalds 296706425c30SEric Dumazet if (unlikely(mss > full_space)) { 29681da177e4SLinus Torvalds mss = full_space; 296906425c30SEric Dumazet if (mss <= 0) 297006425c30SEric Dumazet return 0; 297106425c30SEric Dumazet } 2972b92edbe0SEric Dumazet if (free_space < (full_space >> 1)) { 2973463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.quick = 0; 29741da177e4SLinus Torvalds 2975b8da51ebSEric Dumazet if (tcp_under_memory_pressure(sk)) 2976053f3684SWei Wang tcp_adjust_rcv_ssthresh(sk); 29771da177e4SLinus Torvalds 297886c1a045SFlorian Westphal /* free_space might become our new window, make sure we don't 297986c1a045SFlorian Westphal * increase it due to wscale. 298086c1a045SFlorian Westphal */ 298186c1a045SFlorian Westphal free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale); 298286c1a045SFlorian Westphal 298386c1a045SFlorian Westphal /* if free space is less than mss estimate, or is below 1/16th 298486c1a045SFlorian Westphal * of the maximum allowed, try to move to zero-window, else 298586c1a045SFlorian Westphal * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and 298686c1a045SFlorian Westphal * new incoming data is dropped due to memory limits. 298786c1a045SFlorian Westphal * With large window, mss test triggers way too late in order 298886c1a045SFlorian Westphal * to announce zero window in time before rmem limit kicks in. 298986c1a045SFlorian Westphal */ 299086c1a045SFlorian Westphal if (free_space < (allowed_space >> 4) || free_space < mss) 29911da177e4SLinus Torvalds return 0; 29921da177e4SLinus Torvalds } 29931da177e4SLinus Torvalds 29941da177e4SLinus Torvalds if (free_space > tp->rcv_ssthresh) 29951da177e4SLinus Torvalds free_space = tp->rcv_ssthresh; 29961da177e4SLinus Torvalds 29971da177e4SLinus Torvalds /* Don't do rounding if we are using window scaling, since the 29981da177e4SLinus Torvalds * scaled window will not line up with the MSS boundary anyway. 29991da177e4SLinus Torvalds */ 30001da177e4SLinus Torvalds if (tp->rx_opt.rcv_wscale) { 30011da177e4SLinus Torvalds window = free_space; 30021da177e4SLinus Torvalds 30031da177e4SLinus Torvalds /* Advertise enough space so that it won't get scaled away. 30041da177e4SLinus Torvalds * Import case: prevent zero window announcement if 30051da177e4SLinus Torvalds * 1<<rcv_wscale > mss. 30061da177e4SLinus Torvalds */ 30071935299dSGao Feng window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale)); 30081da177e4SLinus Torvalds } else { 30091935299dSGao Feng window = tp->rcv_wnd; 30101da177e4SLinus Torvalds /* Get the largest window that is a nice multiple of mss. 30111da177e4SLinus Torvalds * Window clamp already applied above. 30121da177e4SLinus Torvalds * If our current window offering is within 1 mss of the 30131da177e4SLinus Torvalds * free space we just keep it. This prevents the divide 30141da177e4SLinus Torvalds * and multiply from happening most of the time. 30151da177e4SLinus Torvalds * We also don't do any window rounding when the free space 30161da177e4SLinus Torvalds * is too small. 30171da177e4SLinus Torvalds */ 30181da177e4SLinus Torvalds if (window <= free_space - mss || window > free_space) 30191935299dSGao Feng window = rounddown(free_space, mss); 302084565070SJohn Heffner else if (mss == full_space && 3021b92edbe0SEric Dumazet free_space > window + (full_space >> 1)) 302284565070SJohn Heffner window = free_space; 30231da177e4SLinus Torvalds } 30241da177e4SLinus Torvalds 30251da177e4SLinus Torvalds return window; 30261da177e4SLinus Torvalds } 30271da177e4SLinus Torvalds 3028cfea5a68SMartin KaFai Lau void tcp_skb_collapse_tstamp(struct sk_buff *skb, 3029082ac2d5SMartin KaFai Lau const struct sk_buff *next_skb) 3030082ac2d5SMartin KaFai Lau { 30310a2cf20cSSoheil Hassas Yeganeh if (unlikely(tcp_has_tx_tstamp(next_skb))) { 30320a2cf20cSSoheil Hassas Yeganeh const struct skb_shared_info *next_shinfo = 30330a2cf20cSSoheil Hassas Yeganeh skb_shinfo(next_skb); 3034082ac2d5SMartin KaFai Lau struct skb_shared_info *shinfo = skb_shinfo(skb); 3035082ac2d5SMartin KaFai Lau 30360a2cf20cSSoheil Hassas Yeganeh shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP; 3037082ac2d5SMartin KaFai Lau shinfo->tskey = next_shinfo->tskey; 30382de8023eSMartin KaFai Lau TCP_SKB_CB(skb)->txstamp_ack |= 30392de8023eSMartin KaFai Lau TCP_SKB_CB(next_skb)->txstamp_ack; 3040082ac2d5SMartin KaFai Lau } 3041082ac2d5SMartin KaFai Lau } 3042082ac2d5SMartin KaFai Lau 30434a17fc3aSIlpo Järvinen /* Collapses two adjacent SKB's during retransmission. */ 3044f8071cdeSEric Dumazet static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) 30451da177e4SLinus Torvalds { 30461da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 304775c119afSEric Dumazet struct sk_buff *next_skb = skb_rb_next(skb); 304813dde04fSWei Yongjun int next_skb_size; 30491da177e4SLinus Torvalds 3050058dc334SIlpo Järvinen next_skb_size = next_skb->len; 30511da177e4SLinus Torvalds 3052058dc334SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); 30531da177e4SLinus Torvalds 3054bd446314SEric Dumazet if (next_skb_size && !tcp_skb_shift(skb, next_skb, 1, next_skb_size)) 3055f8071cdeSEric Dumazet return false; 3056bd446314SEric Dumazet 30572b7cda9cSEric Dumazet tcp_highest_sack_replace(sk, next_skb, skb); 3058a6963a6bSIlpo Järvinen 30591da177e4SLinus Torvalds /* Update sequence range on original skb. */ 30601da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 30611da177e4SLinus Torvalds 3062e6c7d085SIlpo Järvinen /* Merge over control information. This moves PSH/FIN etc. over */ 30634de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; 30641da177e4SLinus Torvalds 30651da177e4SLinus Torvalds /* All done, get rid of second SKB and account for it so 30661da177e4SLinus Torvalds * packet counting does not break. 30671da177e4SLinus Torvalds */ 30684828e7f4SIlpo Järvinen TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; 3069a643b5d4SMartin KaFai Lau TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor; 3070b7689205SIlpo Järvinen 3071b7689205SIlpo Järvinen /* changed transmit queue under us so clear hints */ 3072ef9da47cSIlpo Järvinen tcp_clear_retrans_hints_partial(tp); 3073ef9da47cSIlpo Järvinen if (next_skb == tp->retransmit_skb_hint) 3074ef9da47cSIlpo Järvinen tp->retransmit_skb_hint = skb; 3075b7689205SIlpo Järvinen 3076797108d1SIlpo Järvinen tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); 3077797108d1SIlpo Järvinen 3078082ac2d5SMartin KaFai Lau tcp_skb_collapse_tstamp(skb, next_skb); 3079082ac2d5SMartin KaFai Lau 308075c119afSEric Dumazet tcp_rtx_queue_unlink_and_free(next_skb, sk); 3081f8071cdeSEric Dumazet return true; 30821da177e4SLinus Torvalds } 30831da177e4SLinus Torvalds 308467edfef7SAndi Kleen /* Check if coalescing SKBs is legal. */ 3085a2a385d6SEric Dumazet static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) 30864a17fc3aSIlpo Järvinen { 30874a17fc3aSIlpo Järvinen if (tcp_skb_pcount(skb) > 1) 3088a2a385d6SEric Dumazet return false; 30894a17fc3aSIlpo Järvinen if (skb_cloned(skb)) 3090a2a385d6SEric Dumazet return false; 30912331ccc5SEric Dumazet /* Some heuristics for collapsing over SACK'd could be invented */ 30924a17fc3aSIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 3093a2a385d6SEric Dumazet return false; 30944a17fc3aSIlpo Järvinen 3095a2a385d6SEric Dumazet return true; 30964a17fc3aSIlpo Järvinen } 30974a17fc3aSIlpo Järvinen 309867edfef7SAndi Kleen /* Collapse packets in the retransmit queue to make to create 309967edfef7SAndi Kleen * less packets on the wire. This is only done on retransmission. 310067edfef7SAndi Kleen */ 31014a17fc3aSIlpo Järvinen static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, 31024a17fc3aSIlpo Järvinen int space) 31034a17fc3aSIlpo Järvinen { 31044a17fc3aSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 31054a17fc3aSIlpo Järvinen struct sk_buff *skb = to, *tmp; 3106a2a385d6SEric Dumazet bool first = true; 31074a17fc3aSIlpo Järvinen 3108e0a1e5b5SEric Dumazet if (!sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse) 31094a17fc3aSIlpo Järvinen return; 31104de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 31114a17fc3aSIlpo Järvinen return; 31124a17fc3aSIlpo Järvinen 311375c119afSEric Dumazet skb_rbtree_walk_from_safe(skb, tmp) { 31144a17fc3aSIlpo Järvinen if (!tcp_can_collapse(sk, skb)) 31154a17fc3aSIlpo Järvinen break; 31164a17fc3aSIlpo Järvinen 311785712484SMat Martineau if (!tcp_skb_can_collapse(to, skb)) 3118a643b5d4SMartin KaFai Lau break; 3119a643b5d4SMartin KaFai Lau 31204a17fc3aSIlpo Järvinen space -= skb->len; 31214a17fc3aSIlpo Järvinen 31224a17fc3aSIlpo Järvinen if (first) { 3123a2a385d6SEric Dumazet first = false; 31244a17fc3aSIlpo Järvinen continue; 31254a17fc3aSIlpo Järvinen } 31264a17fc3aSIlpo Järvinen 31274a17fc3aSIlpo Järvinen if (space < 0) 31284a17fc3aSIlpo Järvinen break; 31294a17fc3aSIlpo Järvinen 31304a17fc3aSIlpo Järvinen if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) 31314a17fc3aSIlpo Järvinen break; 31324a17fc3aSIlpo Järvinen 3133f8071cdeSEric Dumazet if (!tcp_collapse_retrans(sk, to)) 3134f8071cdeSEric Dumazet break; 31354a17fc3aSIlpo Järvinen } 31364a17fc3aSIlpo Järvinen } 31374a17fc3aSIlpo Järvinen 31381da177e4SLinus Torvalds /* This retransmits one SKB. Policy decisions and retransmit queue 31391da177e4SLinus Torvalds * state updates are done by the caller. Returns non-zero if an 31401da177e4SLinus Torvalds * error occurred which prevented the send. 31411da177e4SLinus Torvalds */ 314210d3be56SEric Dumazet int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) 31431da177e4SLinus Torvalds { 31445d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 314510d3be56SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 31467d227cd2SSridhar Samudrala unsigned int cur_mss; 314710d3be56SEric Dumazet int diff, len, err; 31481da177e4SLinus Torvalds 314910d3be56SEric Dumazet 315010d3be56SEric Dumazet /* Inconclusive MTU probe */ 315110d3be56SEric Dumazet if (icsk->icsk_mtup.probe_size) 31525d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 31535d424d5aSJohn Heffner 31541f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 31551f3279aeSEric Dumazet return -EBUSY; 31561f3279aeSEric Dumazet 31571da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 31587f582b24SEric Dumazet if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) { 31597f582b24SEric Dumazet WARN_ON_ONCE(1); 31607f582b24SEric Dumazet return -EINVAL; 31617f582b24SEric Dumazet } 31621da177e4SLinus Torvalds if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 31631da177e4SLinus Torvalds return -ENOMEM; 31641da177e4SLinus Torvalds } 31651da177e4SLinus Torvalds 31667d227cd2SSridhar Samudrala if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 31677d227cd2SSridhar Samudrala return -EHOSTUNREACH; /* Routing failure or similar. */ 31687d227cd2SSridhar Samudrala 31690c54b85fSIlpo Järvinen cur_mss = tcp_current_mss(sk); 31707d227cd2SSridhar Samudrala 31711da177e4SLinus Torvalds /* If receiver has shrunk his window, and skb is out of 31721da177e4SLinus Torvalds * new window, do not retransmit it. The exception is the 31731da177e4SLinus Torvalds * case, when window is shrunk to zero. In this case 31741da177e4SLinus Torvalds * our retransmit serves as a zero window probe. 31751da177e4SLinus Torvalds */ 31769d4fb27dSJoe Perches if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && 31779d4fb27dSJoe Perches TCP_SKB_CB(skb)->seq != tp->snd_una) 31781da177e4SLinus Torvalds return -EAGAIN; 31791da177e4SLinus Torvalds 318010d3be56SEric Dumazet len = cur_mss * segs; 318110d3be56SEric Dumazet if (skb->len > len) { 318275c119afSEric Dumazet if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len, 318375c119afSEric Dumazet cur_mss, GFP_ATOMIC)) 31841da177e4SLinus Torvalds return -ENOMEM; /* We'll try again later. */ 318502276f3cSIlpo Järvinen } else { 3186c4777efaSEric Dumazet if (skb_unclone_keeptruesize(skb, GFP_ATOMIC)) 3187c52e2421SEric Dumazet return -ENOMEM; 318810d3be56SEric Dumazet 318910d3be56SEric Dumazet diff = tcp_skb_pcount(skb); 319010d3be56SEric Dumazet tcp_set_skb_tso_segs(skb, cur_mss); 319110d3be56SEric Dumazet diff -= tcp_skb_pcount(skb); 319210d3be56SEric Dumazet if (diff) 319310d3be56SEric Dumazet tcp_adjust_pcount(sk, skb, diff); 319410d3be56SEric Dumazet if (skb->len < cur_mss) 319510d3be56SEric Dumazet tcp_retrans_try_collapse(sk, skb, cur_mss); 31961da177e4SLinus Torvalds } 31971da177e4SLinus Torvalds 319849213555SDaniel Borkmann /* RFC3168, section 6.1.1.1. ECN fallback */ 319949213555SDaniel Borkmann if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN) 320049213555SDaniel Borkmann tcp_ecn_clear_syn(sk, skb); 320149213555SDaniel Borkmann 3202678550c6SYuchung Cheng /* Update global and local TCP statistics. */ 3203678550c6SYuchung Cheng segs = tcp_skb_pcount(skb); 3204678550c6SYuchung Cheng TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs); 3205678550c6SYuchung Cheng if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 3206678550c6SYuchung Cheng __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 3207678550c6SYuchung Cheng tp->total_retrans += segs; 3208fb31c9b9SWei Wang tp->bytes_retrans += skb->len; 3209678550c6SYuchung Cheng 321050bceae9SThomas Graf /* make sure skb->data is aligned on arches that require it 321150bceae9SThomas Graf * and check if ack-trimming & collapsing extended the headroom 321250bceae9SThomas Graf * beyond what csum_start can cover. 321350bceae9SThomas Graf */ 321450bceae9SThomas Graf if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) || 321550bceae9SThomas Graf skb_headroom(skb) >= 0xFFFF)) { 321610a81980SEric Dumazet struct sk_buff *nskb; 321710a81980SEric Dumazet 3218e2080072SEric Dumazet tcp_skb_tsorted_save(skb) { 321910a81980SEric Dumazet nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC); 322007f8e4d0SFlorian Westphal if (nskb) { 322107f8e4d0SFlorian Westphal nskb->dev = NULL; 322207f8e4d0SFlorian Westphal err = tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC); 322307f8e4d0SFlorian Westphal } else { 322407f8e4d0SFlorian Westphal err = -ENOBUFS; 322507f8e4d0SFlorian Westphal } 3226e2080072SEric Dumazet } tcp_skb_tsorted_restore(skb); 3227e2080072SEric Dumazet 32285889e2c0SYousuk Seung if (!err) { 3229a7a25630SEric Dumazet tcp_update_skb_after_send(sk, skb, tp->tcp_wstamp_ns); 32305889e2c0SYousuk Seung tcp_rate_skb_sent(sk, skb); 32315889e2c0SYousuk Seung } 3232117632e6SEric Dumazet } else { 3233c84a5711SYuchung Cheng err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 3234117632e6SEric Dumazet } 3235c84a5711SYuchung Cheng 32367f12422cSYuchung Cheng /* To avoid taking spuriously low RTT samples based on a timestamp 32377f12422cSYuchung Cheng * for a transmit that never happened, always mark EVER_RETRANS 32387f12422cSYuchung Cheng */ 32397f12422cSYuchung Cheng TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS; 32407f12422cSYuchung Cheng 3241a31ad29eSLawrence Brakmo if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RETRANS_CB_FLAG)) 3242a31ad29eSLawrence Brakmo tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RETRANS_CB, 3243a31ad29eSLawrence Brakmo TCP_SKB_CB(skb)->seq, segs, err); 3244a31ad29eSLawrence Brakmo 3245fc9f3501SEric Dumazet if (likely(!err)) { 3246e086101bSCong Wang trace_tcp_retransmit_skb(sk, skb); 3247678550c6SYuchung Cheng } else if (err != -EBUSY) { 3248ec641b39SYuchung Cheng NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL, segs); 3249fc9f3501SEric Dumazet } 3250c84a5711SYuchung Cheng return err; 325193b174adSYuchung Cheng } 325293b174adSYuchung Cheng 325310d3be56SEric Dumazet int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) 325493b174adSYuchung Cheng { 325593b174adSYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 325610d3be56SEric Dumazet int err = __tcp_retransmit_skb(sk, skb, segs); 32571da177e4SLinus Torvalds 32581da177e4SLinus Torvalds if (err == 0) { 32591da177e4SLinus Torvalds #if FASTRETRANS_DEBUG > 0 32601da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 3261e87cc472SJoe Perches net_dbg_ratelimited("retrans_out leaked\n"); 32621da177e4SLinus Torvalds } 32631da177e4SLinus Torvalds #endif 32641da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 32651da177e4SLinus Torvalds tp->retrans_out += tcp_skb_pcount(skb); 32667ae18975SYuchung Cheng } 32671da177e4SLinus Torvalds 32687ae18975SYuchung Cheng /* Save stamp of the first (attempted) retransmit. */ 32691da177e4SLinus Torvalds if (!tp->retrans_stamp) 32707faee5c0SEric Dumazet tp->retrans_stamp = tcp_skb_timestamp(skb); 32711da177e4SLinus Torvalds 32726e08d5e3SYuchung Cheng if (tp->undo_retrans < 0) 32736e08d5e3SYuchung Cheng tp->undo_retrans = 0; 32746e08d5e3SYuchung Cheng tp->undo_retrans += tcp_skb_pcount(skb); 32751da177e4SLinus Torvalds return err; 32761da177e4SLinus Torvalds } 32771da177e4SLinus Torvalds 32781da177e4SLinus Torvalds /* This gets called after a retransmit timeout, and the initially 32791da177e4SLinus Torvalds * retransmitted data is acknowledged. It tries to continue 32801da177e4SLinus Torvalds * resending the rest of the retransmit queue, until either 32811da177e4SLinus Torvalds * we've sent it all or the congestion window limit is reached. 32821da177e4SLinus Torvalds */ 32831da177e4SLinus Torvalds void tcp_xmit_retransmit_queue(struct sock *sk) 32841da177e4SLinus Torvalds { 32856687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 3286b9f1f1ceSEric Dumazet struct sk_buff *skb, *rtx_head, *hole = NULL; 32871da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 3288916e6d1aSEric Dumazet bool rearm_timer = false; 3289840a3cbeSYuchung Cheng u32 max_segs; 329061eb55f4SIlpo Järvinen int mib_idx; 32916a438bbeSStephen Hemminger 329245e77d31SIlpo Järvinen if (!tp->packets_out) 329345e77d31SIlpo Järvinen return; 329445e77d31SIlpo Järvinen 329575c119afSEric Dumazet rtx_head = tcp_rtx_queue_head(sk); 3296b9f1f1ceSEric Dumazet skb = tp->retransmit_skb_hint ?: rtx_head; 3297ed6e7268SNeal Cardwell max_segs = tcp_tso_segs(sk, tcp_current_mss(sk)); 329875c119afSEric Dumazet skb_rbtree_walk_from(skb) { 3299dca0aaf8SEric Dumazet __u8 sacked; 330010d3be56SEric Dumazet int segs; 33011da177e4SLinus Torvalds 3302218af599SEric Dumazet if (tcp_pacing_check(sk)) 3303218af599SEric Dumazet break; 3304218af599SEric Dumazet 33056a438bbeSStephen Hemminger /* we could do better than to assign each time */ 330651456b29SIan Morris if (!hole) 33076a438bbeSStephen Hemminger tp->retransmit_skb_hint = skb; 33086a438bbeSStephen Hemminger 330940570375SEric Dumazet segs = tcp_snd_cwnd(tp) - tcp_packets_in_flight(tp); 331010d3be56SEric Dumazet if (segs <= 0) 3311916e6d1aSEric Dumazet break; 3312dca0aaf8SEric Dumazet sacked = TCP_SKB_CB(skb)->sacked; 3313a3d2e9f8SEric Dumazet /* In case tcp_shift_skb_data() have aggregated large skbs, 3314a3d2e9f8SEric Dumazet * we need to make sure not sending too bigs TSO packets 3315a3d2e9f8SEric Dumazet */ 3316a3d2e9f8SEric Dumazet segs = min_t(int, segs, max_segs); 33170e1c54c2SIlpo Järvinen 3318840a3cbeSYuchung Cheng if (tp->retrans_out >= tp->lost_out) { 3319006f582cSIlpo Järvinen break; 33200e1c54c2SIlpo Järvinen } else if (!(sacked & TCPCB_LOST)) { 332151456b29SIan Morris if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) 33220e1c54c2SIlpo Järvinen hole = skb; 332361eb55f4SIlpo Järvinen continue; 33241da177e4SLinus Torvalds 33250e1c54c2SIlpo Järvinen } else { 33260e1c54c2SIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Loss) 33270e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFASTRETRANS; 33280e1c54c2SIlpo Järvinen else 33290e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; 33300e1c54c2SIlpo Järvinen } 33310e1c54c2SIlpo Järvinen 33320e1c54c2SIlpo Järvinen if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) 333361eb55f4SIlpo Järvinen continue; 333440b215e5SPavel Emelyanov 3335f9616c35SEric Dumazet if (tcp_small_queue_check(sk, skb, 1)) 3336916e6d1aSEric Dumazet break; 3337f9616c35SEric Dumazet 333810d3be56SEric Dumazet if (tcp_retransmit_skb(sk, skb, segs)) 3339916e6d1aSEric Dumazet break; 334024ab6becSYuchung Cheng 3341de1d6578SYuchung Cheng NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb)); 33421da177e4SLinus Torvalds 3343684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 3344a262f0cdSNandita Dukkipati tp->prr_out += tcp_skb_pcount(skb); 3345a262f0cdSNandita Dukkipati 334675c119afSEric Dumazet if (skb == rtx_head && 334757dde7f7SYuchung Cheng icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT) 3348916e6d1aSEric Dumazet rearm_timer = true; 3349916e6d1aSEric Dumazet 3350916e6d1aSEric Dumazet } 3351916e6d1aSEric Dumazet if (rearm_timer) 33523f80e08fSEric Dumazet tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 33533f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, 33548dc242adSEric Dumazet TCP_RTO_MAX); 33551da177e4SLinus Torvalds } 33561da177e4SLinus Torvalds 3357d83769a5SEric Dumazet /* We allow to exceed memory limits for FIN packets to expedite 3358d83769a5SEric Dumazet * connection tear down and (memory) recovery. 3359845704a5SEric Dumazet * Otherwise tcp_send_fin() could be tempted to either delay FIN 3360845704a5SEric Dumazet * or even be forced to close flow without any FIN. 3361a6c5ea4cSEric Dumazet * In general, we want to allow one skb per socket to avoid hangs 3362a6c5ea4cSEric Dumazet * with edge trigger epoll() 3363d83769a5SEric Dumazet */ 3364a6c5ea4cSEric Dumazet void sk_forced_mem_schedule(struct sock *sk, int size) 3365d83769a5SEric Dumazet { 3366e805605cSJohannes Weiner int amt; 3367d83769a5SEric Dumazet 3368d83769a5SEric Dumazet if (size <= sk->sk_forward_alloc) 3369d83769a5SEric Dumazet return; 3370d83769a5SEric Dumazet amt = sk_mem_pages(size); 3371d83769a5SEric Dumazet sk->sk_forward_alloc += amt * SK_MEM_QUANTUM; 3372e805605cSJohannes Weiner sk_memory_allocated_add(sk, amt); 3373e805605cSJohannes Weiner 3374baac50bbSJohannes Weiner if (mem_cgroup_sockets_enabled && sk->sk_memcg) 33754b1327beSWei Wang mem_cgroup_charge_skmem(sk->sk_memcg, amt, 33764b1327beSWei Wang gfp_memcg_charge() | __GFP_NOFAIL); 3377d83769a5SEric Dumazet } 3378d83769a5SEric Dumazet 3379845704a5SEric Dumazet /* Send a FIN. The caller locks the socket for us. 3380845704a5SEric Dumazet * We should try to send a FIN packet really hard, but eventually give up. 33811da177e4SLinus Torvalds */ 33821da177e4SLinus Torvalds void tcp_send_fin(struct sock *sk) 33831da177e4SLinus Torvalds { 3384ee2aabd3SEric Dumazet struct sk_buff *skb, *tskb, *tail = tcp_write_queue_tail(sk); 33851da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 33861da177e4SLinus Torvalds 3387845704a5SEric Dumazet /* Optimization, tack on the FIN if we have one skb in write queue and 3388845704a5SEric Dumazet * this skb was not yet sent, or we are under memory pressure. 3389845704a5SEric Dumazet * Note: in the latter case, FIN packet will be sent after a timeout, 3390845704a5SEric Dumazet * as TCP stack thinks it has already been transmitted. 33911da177e4SLinus Torvalds */ 3392ee2aabd3SEric Dumazet tskb = tail; 339375c119afSEric Dumazet if (!tskb && tcp_under_memory_pressure(sk)) 339475c119afSEric Dumazet tskb = skb_rb_last(&sk->tcp_rtx_queue); 339575c119afSEric Dumazet 339675c119afSEric Dumazet if (tskb) { 3397845704a5SEric Dumazet TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN; 3398845704a5SEric Dumazet TCP_SKB_CB(tskb)->end_seq++; 33991da177e4SLinus Torvalds tp->write_seq++; 3400ee2aabd3SEric Dumazet if (!tail) { 3401845704a5SEric Dumazet /* This means tskb was already sent. 3402845704a5SEric Dumazet * Pretend we included the FIN on previous transmit. 3403845704a5SEric Dumazet * We need to set tp->snd_nxt to the value it would have 3404845704a5SEric Dumazet * if FIN had been sent. This is because retransmit path 3405845704a5SEric Dumazet * does not change tp->snd_nxt. 3406845704a5SEric Dumazet */ 3407e0d694d6SEric Dumazet WRITE_ONCE(tp->snd_nxt, tp->snd_nxt + 1); 3408845704a5SEric Dumazet return; 3409845704a5SEric Dumazet } 34101da177e4SLinus Torvalds } else { 3411845704a5SEric Dumazet skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation); 3412d1edc085SColin Ian King if (unlikely(!skb)) 3413845704a5SEric Dumazet return; 3414d1edc085SColin Ian King 3415e2080072SEric Dumazet INIT_LIST_HEAD(&skb->tcp_tsorted_anchor); 3416d83769a5SEric Dumazet skb_reserve(skb, MAX_TCP_HEADER); 3417a6c5ea4cSEric Dumazet sk_forced_mem_schedule(sk, skb->truesize); 34181da177e4SLinus Torvalds /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 3419e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tp->write_seq, 3420a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_FIN); 34211da177e4SLinus Torvalds tcp_queue_skb(sk, skb); 34221da177e4SLinus Torvalds } 3423845704a5SEric Dumazet __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF); 34241da177e4SLinus Torvalds } 34251da177e4SLinus Torvalds 34261da177e4SLinus Torvalds /* We get here when a process closes a file descriptor (either due to 34271da177e4SLinus Torvalds * an explicit close() or as a byproduct of exit()'ing) and there 34281da177e4SLinus Torvalds * was unread data in the receive queue. This behavior is recommended 342965bb723cSGerrit Renker * by RFC 2525, section 2.17. -DaveM 34301da177e4SLinus Torvalds */ 3431dd0fc66fSAl Viro void tcp_send_active_reset(struct sock *sk, gfp_t priority) 34321da177e4SLinus Torvalds { 34331da177e4SLinus Torvalds struct sk_buff *skb; 34341da177e4SLinus Torvalds 34357cc2b043SGao Feng TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); 34367cc2b043SGao Feng 34371da177e4SLinus Torvalds /* NOTE: No TCP options attached and we never retransmit this. */ 34381da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, priority); 34391da177e4SLinus Torvalds if (!skb) { 34404e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 34411da177e4SLinus Torvalds return; 34421da177e4SLinus Torvalds } 34431da177e4SLinus Torvalds 34441da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 34451da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 3446e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), 3447a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_RST); 34489a568de4SEric Dumazet tcp_mstamp_refresh(tcp_sk(sk)); 34491da177e4SLinus Torvalds /* Send it off. */ 3450dfb4b9dcSDavid S. Miller if (tcp_transmit_skb(sk, skb, 0, priority)) 34514e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 3452c24b14c4SSong Liu 3453c24b14c4SSong Liu /* skb of trace_tcp_send_reset() keeps the skb that caused RST, 3454c24b14c4SSong Liu * skb here is different to the troublesome skb, so use NULL 3455c24b14c4SSong Liu */ 3456c24b14c4SSong Liu trace_tcp_send_reset(sk, NULL); 34571da177e4SLinus Torvalds } 34581da177e4SLinus Torvalds 345967edfef7SAndi Kleen /* Send a crossed SYN-ACK during socket establishment. 346067edfef7SAndi Kleen * WARNING: This routine must only be called when we have already sent 34611da177e4SLinus Torvalds * a SYN packet that crossed the incoming SYN that caused this routine 34621da177e4SLinus Torvalds * to get called. If this assumption fails then the initial rcv_wnd 34631da177e4SLinus Torvalds * and rcv_wscale values will not be correct. 34641da177e4SLinus Torvalds */ 34651da177e4SLinus Torvalds int tcp_send_synack(struct sock *sk) 34661da177e4SLinus Torvalds { 34671da177e4SLinus Torvalds struct sk_buff *skb; 34681da177e4SLinus Torvalds 346975c119afSEric Dumazet skb = tcp_rtx_queue_head(sk); 347051456b29SIan Morris if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 347175c119afSEric Dumazet pr_err("%s: wrong queue state\n", __func__); 34721da177e4SLinus Torvalds return -EFAULT; 34731da177e4SLinus Torvalds } 34744de075e0SEric Dumazet if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { 34751da177e4SLinus Torvalds if (skb_cloned(skb)) { 3476e2080072SEric Dumazet struct sk_buff *nskb; 3477e2080072SEric Dumazet 3478e2080072SEric Dumazet tcp_skb_tsorted_save(skb) { 3479e2080072SEric Dumazet nskb = skb_copy(skb, GFP_ATOMIC); 3480e2080072SEric Dumazet } tcp_skb_tsorted_restore(skb); 348151456b29SIan Morris if (!nskb) 34821da177e4SLinus Torvalds return -ENOMEM; 3483e2080072SEric Dumazet INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor); 34842bec445fSEric Dumazet tcp_highest_sack_replace(sk, skb, nskb); 348575c119afSEric Dumazet tcp_rtx_queue_unlink_and_free(skb, sk); 3486f4a775d1SEric Dumazet __skb_header_release(nskb); 348775c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb); 3488ab4e846aSEric Dumazet sk_wmem_queued_add(sk, nskb->truesize); 34893ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 34901da177e4SLinus Torvalds skb = nskb; 34911da177e4SLinus Torvalds } 34921da177e4SLinus Torvalds 34934de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; 3494735d3831SFlorian Westphal tcp_ecn_send_synack(sk, skb); 34951da177e4SLinus Torvalds } 3496dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 34971da177e4SLinus Torvalds } 34981da177e4SLinus Torvalds 34994aea39c1SEric Dumazet /** 3500331fca43SMartin KaFai Lau * tcp_make_synack - Allocate one skb and build a SYNACK packet. 3501331fca43SMartin KaFai Lau * @sk: listener socket 3502331fca43SMartin KaFai Lau * @dst: dst entry attached to the SYNACK. It is consumed and caller 3503331fca43SMartin KaFai Lau * should not use it again. 3504331fca43SMartin KaFai Lau * @req: request_sock pointer 3505331fca43SMartin KaFai Lau * @foc: cookie for tcp fast open 3506331fca43SMartin KaFai Lau * @synack_type: Type of synack to prepare 3507331fca43SMartin KaFai Lau * @syn_skb: SYN packet just received. It could be NULL for rtx case. 35084aea39c1SEric Dumazet */ 35095d062de7SEric Dumazet struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst, 3510e6b4d113SWilliam Allen Simpson struct request_sock *req, 3511ca6fb065SEric Dumazet struct tcp_fastopen_cookie *foc, 3512331fca43SMartin KaFai Lau enum tcp_synack_type synack_type, 3513331fca43SMartin KaFai Lau struct sk_buff *syn_skb) 35141da177e4SLinus Torvalds { 35152e6599cbSArnaldo Carvalho de Melo struct inet_request_sock *ireq = inet_rsk(req); 35165d062de7SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 351780f03e27SEric Dumazet struct tcp_md5sig_key *md5 = NULL; 35185d062de7SEric Dumazet struct tcp_out_options opts; 35195d062de7SEric Dumazet struct sk_buff *skb; 3520bd0388aeSWilliam Allen Simpson int tcp_header_size; 35215d062de7SEric Dumazet struct tcphdr *th; 3522f5fff5dcSTom Quetchenbach int mss; 3523a842fe14SEric Dumazet u64 now; 35241da177e4SLinus Torvalds 3525ca6fb065SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 35264aea39c1SEric Dumazet if (unlikely(!skb)) { 35274aea39c1SEric Dumazet dst_release(dst); 35281da177e4SLinus Torvalds return NULL; 35294aea39c1SEric Dumazet } 35301da177e4SLinus Torvalds /* Reserve space for headers. */ 35311da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 35321da177e4SLinus Torvalds 3533b3d05147SEric Dumazet switch (synack_type) { 3534b3d05147SEric Dumazet case TCP_SYNACK_NORMAL: 35359e17f8a4SEric Dumazet skb_set_owner_w(skb, req_to_sk(req)); 3536b3d05147SEric Dumazet break; 3537b3d05147SEric Dumazet case TCP_SYNACK_COOKIE: 3538b3d05147SEric Dumazet /* Under synflood, we do not attach skb to a socket, 3539b3d05147SEric Dumazet * to avoid false sharing. 3540b3d05147SEric Dumazet */ 3541b3d05147SEric Dumazet break; 3542b3d05147SEric Dumazet case TCP_SYNACK_FASTOPEN: 3543ca6fb065SEric Dumazet /* sk is a const pointer, because we want to express multiple 3544ca6fb065SEric Dumazet * cpu might call us concurrently. 3545ca6fb065SEric Dumazet * sk->sk_wmem_alloc in an atomic, we can promote to rw. 3546ca6fb065SEric Dumazet */ 3547ca6fb065SEric Dumazet skb_set_owner_w(skb, (struct sock *)sk); 3548b3d05147SEric Dumazet break; 3549ca6fb065SEric Dumazet } 35504aea39c1SEric Dumazet skb_dst_set(skb, dst); 35511da177e4SLinus Torvalds 35523541f9e8SEric Dumazet mss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 3553f5fff5dcSTom Quetchenbach 355433ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 3555a842fe14SEric Dumazet now = tcp_clock_ns(); 35568b5f12d0SFlorian Westphal #ifdef CONFIG_SYN_COOKIES 3557f8ace8d9SFlorian Westphal if (unlikely(synack_type == TCP_SYNACK_COOKIE && ireq->tstamp_ok)) 3558a1ac9c8aSMartin KaFai Lau skb_set_delivery_time(skb, cookie_init_timestamp(req, now), 3559a1ac9c8aSMartin KaFai Lau true); 35608b5f12d0SFlorian Westphal else 35618b5f12d0SFlorian Westphal #endif 35629e450c1eSYuchung Cheng { 3563a1ac9c8aSMartin KaFai Lau skb_set_delivery_time(skb, now, true); 35649e450c1eSYuchung Cheng if (!tcp_rsk(req)->snt_synack) /* Timestamp first SYNACK */ 35659e450c1eSYuchung Cheng tcp_rsk(req)->snt_synack = tcp_skb_timestamp_us(skb); 35669e450c1eSYuchung Cheng } 356780f03e27SEric Dumazet 356880f03e27SEric Dumazet #ifdef CONFIG_TCP_MD5SIG 356980f03e27SEric Dumazet rcu_read_lock(); 3570fd3a154aSEric Dumazet md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req)); 357180f03e27SEric Dumazet #endif 357258d607d3SEric Dumazet skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4); 3573331fca43SMartin KaFai Lau /* bpf program will be interested in the tcp_flags */ 3574331fca43SMartin KaFai Lau TCP_SKB_CB(skb)->tcp_flags = TCPHDR_SYN | TCPHDR_ACK; 357560e2a778SUrsula Braun tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5, 3576331fca43SMartin KaFai Lau foc, synack_type, 3577331fca43SMartin KaFai Lau syn_skb) + sizeof(*th); 357833ad798cSAdam Langley 3579aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 3580aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 35811da177e4SLinus Torvalds 3582ea1627c2SEric Dumazet th = (struct tcphdr *)skb->data; 35831da177e4SLinus Torvalds memset(th, 0, sizeof(struct tcphdr)); 35841da177e4SLinus Torvalds th->syn = 1; 35851da177e4SLinus Torvalds th->ack = 1; 35866ac705b1SEric Dumazet tcp_ecn_make_synack(req, th); 3587b44084c2SEric Dumazet th->source = htons(ireq->ir_num); 3588634fb979SEric Dumazet th->dest = ireq->ir_rmt_port; 3589e05a90ecSJamal Hadi Salim skb->mark = ireq->ir_mark; 35903b117750SEric Dumazet skb->ip_summed = CHECKSUM_PARTIAL; 35913b117750SEric Dumazet th->seq = htonl(tcp_rsk(req)->snt_isn); 35928336886fSJerry Chu /* XXX data is queued and acked as is. No buffer/window check */ 35938336886fSJerry Chu th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt); 35941da177e4SLinus Torvalds 35951da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 3596ed53d0abSEric Dumazet th->window = htons(min(req->rsk_rcv_wnd, 65535U)); 3597ea66758cSPaolo Abeni tcp_options_write(th, NULL, &opts); 35981da177e4SLinus Torvalds th->doff = (tcp_header_size >> 2); 359990bbcc60SEric Dumazet __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS); 3600cfb6eeb4SYOSHIFUJI Hideaki 3601cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 3602cfb6eeb4SYOSHIFUJI Hideaki /* Okay, we have all we need - do the md5 hash if needed */ 360380f03e27SEric Dumazet if (md5) 3604bd0388aeSWilliam Allen Simpson tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, 360539f8e58eSEric Dumazet md5, req_to_sk(req), skb); 360680f03e27SEric Dumazet rcu_read_unlock(); 3607cfb6eeb4SYOSHIFUJI Hideaki #endif 3608cfb6eeb4SYOSHIFUJI Hideaki 3609331fca43SMartin KaFai Lau bpf_skops_write_hdr_opt((struct sock *)sk, skb, req, syn_skb, 3610331fca43SMartin KaFai Lau synack_type, &opts); 3611331fca43SMartin KaFai Lau 3612a1ac9c8aSMartin KaFai Lau skb_set_delivery_time(skb, now, true); 3613a842fe14SEric Dumazet tcp_add_tx_delay(skb, tp); 3614a842fe14SEric Dumazet 36151da177e4SLinus Torvalds return skb; 36161da177e4SLinus Torvalds } 36174bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_make_synack); 36181da177e4SLinus Torvalds 361981164413SDaniel Borkmann static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst) 362081164413SDaniel Borkmann { 362181164413SDaniel Borkmann struct inet_connection_sock *icsk = inet_csk(sk); 362281164413SDaniel Borkmann const struct tcp_congestion_ops *ca; 362381164413SDaniel Borkmann u32 ca_key = dst_metric(dst, RTAX_CC_ALGO); 362481164413SDaniel Borkmann 362581164413SDaniel Borkmann if (ca_key == TCP_CA_UNSPEC) 362681164413SDaniel Borkmann return; 362781164413SDaniel Borkmann 362881164413SDaniel Borkmann rcu_read_lock(); 362981164413SDaniel Borkmann ca = tcp_ca_find_key(ca_key); 36300baf26b0SMartin KaFai Lau if (likely(ca && bpf_try_module_get(ca, ca->owner))) { 36310baf26b0SMartin KaFai Lau bpf_module_put(icsk->icsk_ca_ops, icsk->icsk_ca_ops->owner); 363281164413SDaniel Borkmann icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst); 363381164413SDaniel Borkmann icsk->icsk_ca_ops = ca; 363481164413SDaniel Borkmann } 363581164413SDaniel Borkmann rcu_read_unlock(); 363681164413SDaniel Borkmann } 363781164413SDaniel Borkmann 363867edfef7SAndi Kleen /* Do all connect socket setups that can be done AF independent. */ 3639f7e56a76Sstephen hemminger static void tcp_connect_init(struct sock *sk) 36401da177e4SLinus Torvalds { 3641cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 36421da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 36431da177e4SLinus Torvalds __u8 rcv_wscale; 364413d3b1ebSLawrence Brakmo u32 rcv_wnd; 36451da177e4SLinus Torvalds 36461da177e4SLinus Torvalds /* We'll fix this up when we get a response from the other end. 36471da177e4SLinus Torvalds * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 36481da177e4SLinus Torvalds */ 36495d2ed052SEric Dumazet tp->tcp_header_len = sizeof(struct tcphdr); 36505d2ed052SEric Dumazet if (sock_net(sk)->ipv4.sysctl_tcp_timestamps) 36515d2ed052SEric Dumazet tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED; 36521da177e4SLinus Torvalds 3653cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 365400db4124SIan Morris if (tp->af_specific->md5_lookup(sk, sk)) 3655cfb6eeb4SYOSHIFUJI Hideaki tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 3656cfb6eeb4SYOSHIFUJI Hideaki #endif 3657cfb6eeb4SYOSHIFUJI Hideaki 36581da177e4SLinus Torvalds /* If user gave his TCP_MAXSEG, record it to clamp */ 36591da177e4SLinus Torvalds if (tp->rx_opt.user_mss) 36601da177e4SLinus Torvalds tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 36611da177e4SLinus Torvalds tp->max_window = 0; 36625d424d5aSJohn Heffner tcp_mtup_init(sk); 36631da177e4SLinus Torvalds tcp_sync_mss(sk, dst_mtu(dst)); 36641da177e4SLinus Torvalds 366581164413SDaniel Borkmann tcp_ca_dst_init(sk, dst); 366681164413SDaniel Borkmann 36671da177e4SLinus Torvalds if (!tp->window_clamp) 36681da177e4SLinus Torvalds tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 36693541f9e8SEric Dumazet tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 3670f5fff5dcSTom Quetchenbach 36711da177e4SLinus Torvalds tcp_initialize_rcv_mss(sk); 36721da177e4SLinus Torvalds 3673e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 3674e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 3675e88c64f0SHagen Paul Pfeifer (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) 3676e88c64f0SHagen Paul Pfeifer tp->window_clamp = tcp_full_space(sk); 3677e88c64f0SHagen Paul Pfeifer 367813d3b1ebSLawrence Brakmo rcv_wnd = tcp_rwnd_init_bpf(sk); 367913d3b1ebSLawrence Brakmo if (rcv_wnd == 0) 368013d3b1ebSLawrence Brakmo rcv_wnd = dst_metric(dst, RTAX_INITRWND); 368113d3b1ebSLawrence Brakmo 3682ceef9ab6SEric Dumazet tcp_select_initial_window(sk, tcp_full_space(sk), 36831da177e4SLinus Torvalds tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 36841da177e4SLinus Torvalds &tp->rcv_wnd, 36851da177e4SLinus Torvalds &tp->window_clamp, 36869bb37ef0SEric Dumazet sock_net(sk)->ipv4.sysctl_tcp_window_scaling, 368731d12926Slaurent chavey &rcv_wscale, 368813d3b1ebSLawrence Brakmo rcv_wnd); 36891da177e4SLinus Torvalds 36901da177e4SLinus Torvalds tp->rx_opt.rcv_wscale = rcv_wscale; 36911da177e4SLinus Torvalds tp->rcv_ssthresh = tp->rcv_wnd; 36921da177e4SLinus Torvalds 36931da177e4SLinus Torvalds sk->sk_err = 0; 36941da177e4SLinus Torvalds sock_reset_flag(sk, SOCK_DONE); 36951da177e4SLinus Torvalds tp->snd_wnd = 0; 3696ee7537b6SHantzis Fotis tcp_init_wl(tp, 0); 36977f582b24SEric Dumazet tcp_write_queue_purge(sk); 36981da177e4SLinus Torvalds tp->snd_una = tp->write_seq; 36991da177e4SLinus Torvalds tp->snd_sml = tp->write_seq; 370033f5f57eSIlpo Järvinen tp->snd_up = tp->write_seq; 3701e0d694d6SEric Dumazet WRITE_ONCE(tp->snd_nxt, tp->write_seq); 3702ee995283SPavel Emelyanov 3703ee995283SPavel Emelyanov if (likely(!tp->repair)) 37041da177e4SLinus Torvalds tp->rcv_nxt = 0; 3705c7781a6eSAndrew Vagin else 370670eabf0eSEric Dumazet tp->rcv_tstamp = tcp_jiffies32; 3707ee995283SPavel Emelyanov tp->rcv_wup = tp->rcv_nxt; 37087db48e98SEric Dumazet WRITE_ONCE(tp->copied_seq, tp->rcv_nxt); 37091da177e4SLinus Torvalds 37108550f328SLawrence Brakmo inet_csk(sk)->icsk_rto = tcp_timeout_init(sk); 3711463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_retransmits = 0; 37121da177e4SLinus Torvalds tcp_clear_retrans(tp); 37131da177e4SLinus Torvalds } 37141da177e4SLinus Torvalds 3715783237e8SYuchung Cheng static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) 3716783237e8SYuchung Cheng { 3717783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3718783237e8SYuchung Cheng struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 3719783237e8SYuchung Cheng 3720783237e8SYuchung Cheng tcb->end_seq += skb->len; 3721f4a775d1SEric Dumazet __skb_header_release(skb); 3722ab4e846aSEric Dumazet sk_wmem_queued_add(sk, skb->truesize); 3723783237e8SYuchung Cheng sk_mem_charge(sk, skb->truesize); 37240f317464SEric Dumazet WRITE_ONCE(tp->write_seq, tcb->end_seq); 3725783237e8SYuchung Cheng tp->packets_out += tcp_skb_pcount(skb); 3726783237e8SYuchung Cheng } 3727783237e8SYuchung Cheng 3728783237e8SYuchung Cheng /* Build and send a SYN with data and (cached) Fast Open cookie. However, 3729783237e8SYuchung Cheng * queue a data-only packet after the regular SYN, such that regular SYNs 3730783237e8SYuchung Cheng * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges 3731783237e8SYuchung Cheng * only the SYN sequence, the data are retransmitted in the first ACK. 3732783237e8SYuchung Cheng * If cookie is not cached or other error occurs, falls back to send a 3733783237e8SYuchung Cheng * regular SYN with Fast Open cookie request option. 3734783237e8SYuchung Cheng */ 3735783237e8SYuchung Cheng static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) 3736783237e8SYuchung Cheng { 3737ed0c99dcSJakub Kicinski struct inet_connection_sock *icsk = inet_csk(sk); 3738783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3739783237e8SYuchung Cheng struct tcp_fastopen_request *fo = tp->fastopen_req; 3740065263f4SWei Wang int space, err = 0; 3741355a901eSEric Dumazet struct sk_buff *syn_data; 3742783237e8SYuchung Cheng 374367da22d2SYuchung Cheng tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ 3744065263f4SWei Wang if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie)) 3745783237e8SYuchung Cheng goto fallback; 3746783237e8SYuchung Cheng 3747783237e8SYuchung Cheng /* MSS for SYN-data is based on cached MSS and bounded by PMTU and 3748783237e8SYuchung Cheng * user-MSS. Reserve maximum option space for middleboxes that add 3749783237e8SYuchung Cheng * private TCP options. The cost is reduced data space in SYN :( 3750783237e8SYuchung Cheng */ 37513541f9e8SEric Dumazet tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp); 3752ed0c99dcSJakub Kicinski /* Sync mss_cache after updating the mss_clamp */ 3753ed0c99dcSJakub Kicinski tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 37543541f9e8SEric Dumazet 3755ed0c99dcSJakub Kicinski space = __tcp_mtu_to_mss(sk, icsk->icsk_pmtu_cookie) - 3756783237e8SYuchung Cheng MAX_TCP_OPTION_SPACE; 3757783237e8SYuchung Cheng 3758f5ddcbbbSEric Dumazet space = min_t(size_t, space, fo->size); 3759f5ddcbbbSEric Dumazet 3760f5ddcbbbSEric Dumazet /* limit to order-0 allocations */ 3761f5ddcbbbSEric Dumazet space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER)); 3762f5ddcbbbSEric Dumazet 3763f8dd3b8dSEric Dumazet syn_data = tcp_stream_alloc_skb(sk, space, sk->sk_allocation, false); 3764355a901eSEric Dumazet if (!syn_data) 3765783237e8SYuchung Cheng goto fallback; 3766355a901eSEric Dumazet memcpy(syn_data->cb, syn->cb, sizeof(syn->cb)); 376707e100f9SEric Dumazet if (space) { 376807e100f9SEric Dumazet int copied = copy_from_iter(skb_put(syn_data, space), space, 376957be5bdaSAl Viro &fo->data->msg_iter); 377057be5bdaSAl Viro if (unlikely(!copied)) { 3771ba233b34SEric Dumazet tcp_skb_tsorted_anchor_cleanup(syn_data); 3772355a901eSEric Dumazet kfree_skb(syn_data); 3773783237e8SYuchung Cheng goto fallback; 3774783237e8SYuchung Cheng } 377557be5bdaSAl Viro if (copied != space) { 377657be5bdaSAl Viro skb_trim(syn_data, copied); 377757be5bdaSAl Viro space = copied; 377857be5bdaSAl Viro } 3779f859a448SWillem de Bruijn skb_zcopy_set(syn_data, fo->uarg, NULL); 378007e100f9SEric Dumazet } 3781355a901eSEric Dumazet /* No more data pending in inet_wait_for_connect() */ 3782355a901eSEric Dumazet if (space == fo->size) 3783355a901eSEric Dumazet fo->data = NULL; 3784355a901eSEric Dumazet fo->copied = space; 3785783237e8SYuchung Cheng 3786355a901eSEric Dumazet tcp_connect_queue_skb(sk, syn_data); 37870f87230dSFrancis Yan if (syn_data->len) 37880f87230dSFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_BUSY); 3789355a901eSEric Dumazet 3790355a901eSEric Dumazet err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation); 3791355a901eSEric Dumazet 3792a1ac9c8aSMartin KaFai Lau skb_set_delivery_time(syn, syn_data->skb_mstamp_ns, true); 3793355a901eSEric Dumazet 3794355a901eSEric Dumazet /* Now full SYN+DATA was cloned and sent (or not), 3795355a901eSEric Dumazet * remove the SYN from the original skb (syn_data) 3796355a901eSEric Dumazet * we keep in write queue in case of a retransmit, as we 3797355a901eSEric Dumazet * also have the SYN packet (with no data) in the same queue. 3798431a9124SEric Dumazet */ 3799355a901eSEric Dumazet TCP_SKB_CB(syn_data)->seq++; 3800355a901eSEric Dumazet TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH; 3801355a901eSEric Dumazet if (!err) { 380267da22d2SYuchung Cheng tp->syn_data = (fo->copied > 0); 380375c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data); 3804f19c29e3SYuchung Cheng NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT); 3805783237e8SYuchung Cheng goto done; 3806783237e8SYuchung Cheng } 3807783237e8SYuchung Cheng 380875c119afSEric Dumazet /* data was not sent, put it in write_queue */ 380975c119afSEric Dumazet __skb_queue_tail(&sk->sk_write_queue, syn_data); 3810b5b7db8dSEric Dumazet tp->packets_out -= tcp_skb_pcount(syn_data); 3811b5b7db8dSEric Dumazet 3812783237e8SYuchung Cheng fallback: 3813783237e8SYuchung Cheng /* Send a regular SYN with Fast Open cookie request option */ 3814783237e8SYuchung Cheng if (fo->cookie.len > 0) 3815783237e8SYuchung Cheng fo->cookie.len = 0; 3816783237e8SYuchung Cheng err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); 3817783237e8SYuchung Cheng if (err) 3818783237e8SYuchung Cheng tp->syn_fastopen = 0; 3819783237e8SYuchung Cheng done: 3820783237e8SYuchung Cheng fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ 3821783237e8SYuchung Cheng return err; 3822783237e8SYuchung Cheng } 3823783237e8SYuchung Cheng 382467edfef7SAndi Kleen /* Build a SYN and send it off. */ 38251da177e4SLinus Torvalds int tcp_connect(struct sock *sk) 38261da177e4SLinus Torvalds { 38271da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 38281da177e4SLinus Torvalds struct sk_buff *buff; 3829ee586811SEric Paris int err; 38301da177e4SLinus Torvalds 3831de525be2SLawrence Brakmo tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB, 0, NULL); 38328ba60924SEric Dumazet 38338ba60924SEric Dumazet if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 38348ba60924SEric Dumazet return -EHOSTUNREACH; /* Routing failure or similar. */ 38358ba60924SEric Dumazet 38361da177e4SLinus Torvalds tcp_connect_init(sk); 38371da177e4SLinus Torvalds 38382b916477SAndrey Vagin if (unlikely(tp->repair)) { 38392b916477SAndrey Vagin tcp_finish_connect(sk, NULL); 38402b916477SAndrey Vagin return 0; 38412b916477SAndrey Vagin } 38422b916477SAndrey Vagin 3843f8dd3b8dSEric Dumazet buff = tcp_stream_alloc_skb(sk, 0, sk->sk_allocation, true); 3844355a901eSEric Dumazet if (unlikely(!buff)) 38451da177e4SLinus Torvalds return -ENOBUFS; 38461da177e4SLinus Torvalds 3847a3433f35SChangli Gao tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); 38489a568de4SEric Dumazet tcp_mstamp_refresh(tp); 38499a568de4SEric Dumazet tp->retrans_stamp = tcp_time_stamp(tp); 3850783237e8SYuchung Cheng tcp_connect_queue_skb(sk, buff); 3851735d3831SFlorian Westphal tcp_ecn_send_syn(sk, buff); 385275c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, buff); 38531da177e4SLinus Torvalds 3854783237e8SYuchung Cheng /* Send off SYN; include data in Fast Open. */ 3855783237e8SYuchung Cheng err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : 3856783237e8SYuchung Cheng tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); 3857ee586811SEric Paris if (err == -ECONNREFUSED) 3858ee586811SEric Paris return err; 3859bd37a088SWei Yongjun 3860bd37a088SWei Yongjun /* We change tp->snd_nxt after the tcp_transmit_skb() call 3861bd37a088SWei Yongjun * in order to make this packet get counted in tcpOutSegs. 3862bd37a088SWei Yongjun */ 3863e0d694d6SEric Dumazet WRITE_ONCE(tp->snd_nxt, tp->write_seq); 3864bd37a088SWei Yongjun tp->pushed_seq = tp->write_seq; 3865b5b7db8dSEric Dumazet buff = tcp_send_head(sk); 3866b5b7db8dSEric Dumazet if (unlikely(buff)) { 3867e0d694d6SEric Dumazet WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(buff)->seq); 3868b5b7db8dSEric Dumazet tp->pushed_seq = TCP_SKB_CB(buff)->seq; 3869b5b7db8dSEric Dumazet } 387081cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); 38711da177e4SLinus Torvalds 38721da177e4SLinus Torvalds /* Timer for repeating the SYN until an answer. */ 38733f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 38743f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 38751da177e4SLinus Torvalds return 0; 38761da177e4SLinus Torvalds } 38774bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_connect); 38781da177e4SLinus Torvalds 38791da177e4SLinus Torvalds /* Send out a delayed ack, the caller does the policy checking 38801da177e4SLinus Torvalds * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 38811da177e4SLinus Torvalds * for details. 38821da177e4SLinus Torvalds */ 38831da177e4SLinus Torvalds void tcp_send_delayed_ack(struct sock *sk) 38841da177e4SLinus Torvalds { 3885463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 3886463c84b9SArnaldo Carvalho de Melo int ato = icsk->icsk_ack.ato; 38871da177e4SLinus Torvalds unsigned long timeout; 38881da177e4SLinus Torvalds 38891da177e4SLinus Torvalds if (ato > TCP_DELACK_MIN) { 3890463c84b9SArnaldo Carvalho de Melo const struct tcp_sock *tp = tcp_sk(sk); 38911da177e4SLinus Torvalds int max_ato = HZ / 2; 38921da177e4SLinus Torvalds 389331954cd8SWei Wang if (inet_csk_in_pingpong_mode(sk) || 3894056834d9SIlpo Järvinen (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 38951da177e4SLinus Torvalds max_ato = TCP_DELACK_MAX; 38961da177e4SLinus Torvalds 38971da177e4SLinus Torvalds /* Slow path, intersegment interval is "high". */ 38981da177e4SLinus Torvalds 38991da177e4SLinus Torvalds /* If some rtt estimate is known, use it to bound delayed ack. 3900463c84b9SArnaldo Carvalho de Melo * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 39011da177e4SLinus Torvalds * directly. 39021da177e4SLinus Torvalds */ 3903740b0f18SEric Dumazet if (tp->srtt_us) { 3904740b0f18SEric Dumazet int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3), 3905740b0f18SEric Dumazet TCP_DELACK_MIN); 39061da177e4SLinus Torvalds 39071da177e4SLinus Torvalds if (rtt < max_ato) 39081da177e4SLinus Torvalds max_ato = rtt; 39091da177e4SLinus Torvalds } 39101da177e4SLinus Torvalds 39111da177e4SLinus Torvalds ato = min(ato, max_ato); 39121da177e4SLinus Torvalds } 39131da177e4SLinus Torvalds 39142b8ee4f0SMartin KaFai Lau ato = min_t(u32, ato, inet_csk(sk)->icsk_delack_max); 39152b8ee4f0SMartin KaFai Lau 39161da177e4SLinus Torvalds /* Stay within the limit we were given */ 39171da177e4SLinus Torvalds timeout = jiffies + ato; 39181da177e4SLinus Torvalds 39191da177e4SLinus Torvalds /* Use new timeout only if there wasn't a older one earlier. */ 3920463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 3921b6b6d653SEric Dumazet /* If delack timer is about to expire, send ACK now. */ 3922b6b6d653SEric Dumazet if (time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 39231da177e4SLinus Torvalds tcp_send_ack(sk); 39241da177e4SLinus Torvalds return; 39251da177e4SLinus Torvalds } 39261da177e4SLinus Torvalds 3927463c84b9SArnaldo Carvalho de Melo if (!time_before(timeout, icsk->icsk_ack.timeout)) 3928463c84b9SArnaldo Carvalho de Melo timeout = icsk->icsk_ack.timeout; 39291da177e4SLinus Torvalds } 3930463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 3931463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.timeout = timeout; 3932463c84b9SArnaldo Carvalho de Melo sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 39331da177e4SLinus Torvalds } 39341da177e4SLinus Torvalds 39351da177e4SLinus Torvalds /* This routine sends an ack and also updates the window. */ 39362987babbSYuchung Cheng void __tcp_send_ack(struct sock *sk, u32 rcv_nxt) 39371da177e4SLinus Torvalds { 39381da177e4SLinus Torvalds struct sk_buff *buff; 39391da177e4SLinus Torvalds 3940058dc334SIlpo Järvinen /* If we have been reset, we may not send again. */ 3941058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3942058dc334SIlpo Järvinen return; 3943058dc334SIlpo Järvinen 39441da177e4SLinus Torvalds /* We are not putting this on the write queue, so 39451da177e4SLinus Torvalds * tcp_transmit_skb() will set the ownership to this 39461da177e4SLinus Torvalds * sock. 39471da177e4SLinus Torvalds */ 39487450aaf6SEric Dumazet buff = alloc_skb(MAX_TCP_HEADER, 39497450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 39507450aaf6SEric Dumazet if (unlikely(!buff)) { 3951a37c2134SEric Dumazet struct inet_connection_sock *icsk = inet_csk(sk); 3952a37c2134SEric Dumazet unsigned long delay; 3953a37c2134SEric Dumazet 3954a37c2134SEric Dumazet delay = TCP_DELACK_MAX << icsk->icsk_ack.retry; 3955a37c2134SEric Dumazet if (delay < TCP_RTO_MAX) 3956a37c2134SEric Dumazet icsk->icsk_ack.retry++; 3957463c84b9SArnaldo Carvalho de Melo inet_csk_schedule_ack(sk); 3958a37c2134SEric Dumazet icsk->icsk_ack.ato = TCP_ATO_MIN; 3959a37c2134SEric Dumazet inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, delay, TCP_RTO_MAX); 39601da177e4SLinus Torvalds return; 39611da177e4SLinus Torvalds } 39621da177e4SLinus Torvalds 39631da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 39641da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 3965a3433f35SChangli Gao tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); 39661da177e4SLinus Torvalds 396798781965SEric Dumazet /* We do not want pure acks influencing TCP Small Queues or fq/pacing 396898781965SEric Dumazet * too much. 396998781965SEric Dumazet * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784 397098781965SEric Dumazet */ 397198781965SEric Dumazet skb_set_tcp_pure_ack(buff); 397298781965SEric Dumazet 39731da177e4SLinus Torvalds /* Send it off, this clears delayed acks for us. */ 39742987babbSYuchung Cheng __tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0, rcv_nxt); 39751da177e4SLinus Torvalds } 397627cde44aSYuchung Cheng EXPORT_SYMBOL_GPL(__tcp_send_ack); 39772987babbSYuchung Cheng 39782987babbSYuchung Cheng void tcp_send_ack(struct sock *sk) 39792987babbSYuchung Cheng { 39802987babbSYuchung Cheng __tcp_send_ack(sk, tcp_sk(sk)->rcv_nxt); 39811da177e4SLinus Torvalds } 39821da177e4SLinus Torvalds 39831da177e4SLinus Torvalds /* This routine sends a packet with an out of date sequence 39841da177e4SLinus Torvalds * number. It assumes the other end will try to ack it. 39851da177e4SLinus Torvalds * 39861da177e4SLinus Torvalds * Question: what should we make while urgent mode? 39871da177e4SLinus Torvalds * 4.4BSD forces sending single byte of data. We cannot send 39881da177e4SLinus Torvalds * out of window data, because we have SND.NXT==SND.MAX... 39891da177e4SLinus Torvalds * 39901da177e4SLinus Torvalds * Current solution: to send TWO zero-length segments in urgent mode: 39911da177e4SLinus Torvalds * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 39921da177e4SLinus Torvalds * out-of-date with SND.UNA-1 to probe window. 39931da177e4SLinus Torvalds */ 3994e520af48SEric Dumazet static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib) 39951da177e4SLinus Torvalds { 39961da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 39971da177e4SLinus Torvalds struct sk_buff *skb; 39981da177e4SLinus Torvalds 39991da177e4SLinus Torvalds /* We don't queue it, tcp_transmit_skb() sets ownership. */ 40007450aaf6SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, 40017450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 400251456b29SIan Morris if (!skb) 40031da177e4SLinus Torvalds return -1; 40041da177e4SLinus Torvalds 40051da177e4SLinus Torvalds /* Reserve space for headers and set control bits. */ 40061da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 40071da177e4SLinus Torvalds /* Use a previous sequence. This should cause the other 40081da177e4SLinus Torvalds * end to send an ack. Don't queue or clone SKB, just 40091da177e4SLinus Torvalds * send it. 40101da177e4SLinus Torvalds */ 4011a3433f35SChangli Gao tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); 4012e2e8009fSRenato Westphal NET_INC_STATS(sock_net(sk), mib); 40137450aaf6SEric Dumazet return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0); 40141da177e4SLinus Torvalds } 40151da177e4SLinus Torvalds 4016385e2070SEric Dumazet /* Called from setsockopt( ... TCP_REPAIR ) */ 4017ee995283SPavel Emelyanov void tcp_send_window_probe(struct sock *sk) 4018ee995283SPavel Emelyanov { 4019ee995283SPavel Emelyanov if (sk->sk_state == TCP_ESTABLISHED) { 4020ee995283SPavel Emelyanov tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; 40219a568de4SEric Dumazet tcp_mstamp_refresh(tcp_sk(sk)); 4022e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE); 4023ee995283SPavel Emelyanov } 4024ee995283SPavel Emelyanov } 4025ee995283SPavel Emelyanov 402667edfef7SAndi Kleen /* Initiate keepalive or window probe from timer. */ 4027e520af48SEric Dumazet int tcp_write_wakeup(struct sock *sk, int mib) 40281da177e4SLinus Torvalds { 40291da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 40301da177e4SLinus Torvalds struct sk_buff *skb; 40311da177e4SLinus Torvalds 4032058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 4033058dc334SIlpo Järvinen return -1; 4034058dc334SIlpo Järvinen 403500db4124SIan Morris skb = tcp_send_head(sk); 403600db4124SIan Morris if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { 40371da177e4SLinus Torvalds int err; 40380c54b85fSIlpo Järvinen unsigned int mss = tcp_current_mss(sk); 403990840defSIlpo Järvinen unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 40401da177e4SLinus Torvalds 40411da177e4SLinus Torvalds if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 40421da177e4SLinus Torvalds tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 40431da177e4SLinus Torvalds 40441da177e4SLinus Torvalds /* We are probing the opening of a window 40451da177e4SLinus Torvalds * but the window size is != 0 40461da177e4SLinus Torvalds * must have been a result SWS avoidance ( sender ) 40471da177e4SLinus Torvalds */ 40481da177e4SLinus Torvalds if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 40491da177e4SLinus Torvalds skb->len > mss) { 40501da177e4SLinus Torvalds seg_size = min(seg_size, mss); 40514de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 405275c119afSEric Dumazet if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE, 405375c119afSEric Dumazet skb, seg_size, mss, GFP_ATOMIC)) 40541da177e4SLinus Torvalds return -1; 40551da177e4SLinus Torvalds } else if (!tcp_skb_pcount(skb)) 40565bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss); 40571da177e4SLinus Torvalds 40584de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 4059dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 406066f5fe62SIlpo Järvinen if (!err) 406166f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 40621da177e4SLinus Torvalds return err; 40631da177e4SLinus Torvalds } else { 406433f5f57eSIlpo Järvinen if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) 4065e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 1, mib); 4066e520af48SEric Dumazet return tcp_xmit_probe_skb(sk, 0, mib); 40671da177e4SLinus Torvalds } 40681da177e4SLinus Torvalds } 40691da177e4SLinus Torvalds 40701da177e4SLinus Torvalds /* A window probe timeout has occurred. If window is not closed send 40711da177e4SLinus Torvalds * a partial packet else a zero probe. 40721da177e4SLinus Torvalds */ 40731da177e4SLinus Torvalds void tcp_send_probe0(struct sock *sk) 40741da177e4SLinus Torvalds { 4075463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 40761da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 4077c6214a97SNikolay Borisov struct net *net = sock_net(sk); 4078c1d5674fSYuchung Cheng unsigned long timeout; 40791da177e4SLinus Torvalds int err; 40801da177e4SLinus Torvalds 4081e520af48SEric Dumazet err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE); 40821da177e4SLinus Torvalds 408375c119afSEric Dumazet if (tp->packets_out || tcp_write_queue_empty(sk)) { 40841da177e4SLinus Torvalds /* Cancel probe timer, if it is not required. */ 40856687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 0; 4086463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff = 0; 40879d9b1ee0SEnke Chen icsk->icsk_probes_tstamp = 0; 40881da177e4SLinus Torvalds return; 40891da177e4SLinus Torvalds } 40901da177e4SLinus Torvalds 4091c1d5674fSYuchung Cheng icsk->icsk_probes_out++; 40921da177e4SLinus Torvalds if (err <= 0) { 4093c6214a97SNikolay Borisov if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2) 4094463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff++; 4095c1d5674fSYuchung Cheng timeout = tcp_probe0_when(sk, TCP_RTO_MAX); 40961da177e4SLinus Torvalds } else { 40971da177e4SLinus Torvalds /* If packet was not sent due to local congestion, 4098c1d5674fSYuchung Cheng * Let senders fight for local resources conservatively. 40991da177e4SLinus Torvalds */ 4100c1d5674fSYuchung Cheng timeout = TCP_RESOURCE_PROBE_INTERVAL; 41011da177e4SLinus Torvalds } 4102344db93aSEnke Chen 4103344db93aSEnke Chen timeout = tcp_clamp_probe0_to_user_timeout(sk, timeout); 41048dc242adSEric Dumazet tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, timeout, TCP_RTO_MAX); 41051da177e4SLinus Torvalds } 41065db92c99SOctavian Purdila 4107ea3bea3aSEric Dumazet int tcp_rtx_synack(const struct sock *sk, struct request_sock *req) 41085db92c99SOctavian Purdila { 41095db92c99SOctavian Purdila const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific; 41105db92c99SOctavian Purdila struct flowi fl; 41115db92c99SOctavian Purdila int res; 41125db92c99SOctavian Purdila 4113cb6cd2ceSAkhmat Karakotov /* Paired with WRITE_ONCE() in sock_setsockopt() */ 4114cb6cd2ceSAkhmat Karakotov if (READ_ONCE(sk->sk_txrehash) == SOCK_TXREHASH_ENABLED) 411558d607d3SEric Dumazet tcp_rsk(req)->txhash = net_tx_rndhash(); 4116331fca43SMartin KaFai Lau res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL, 4117331fca43SMartin KaFai Lau NULL); 41185db92c99SOctavian Purdila if (!res) { 41190a375c82SEric Dumazet TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); 41200a375c82SEric Dumazet NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 41217e32b443SYuchung Cheng if (unlikely(tcp_passive_fastopen(sk))) 41227e32b443SYuchung Cheng tcp_sk(sk)->total_retrans++; 4123cf34ce3dSSong Liu trace_tcp_retransmit_synack(sk, req); 41245db92c99SOctavian Purdila } 41255db92c99SOctavian Purdila return res; 41265db92c99SOctavian Purdila } 41275db92c99SOctavian Purdila EXPORT_SYMBOL(tcp_rtx_synack); 4128