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)); 856a5dc9e5SEric Dumazet } 861da177e4SLinus Torvalds 87a4ecb15aSCui, Cheng /* SND.NXT, if window was not shrunk or the amount of shrunk was less than one 88a4ecb15aSCui, Cheng * window scaling factor due to loss of precision. 891da177e4SLinus Torvalds * If window has been shrunk, what should we make? It is not clear at all. 901da177e4SLinus Torvalds * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-( 911da177e4SLinus Torvalds * Anything in between SND.UNA...SND.UNA+SND.WND also can be already 921da177e4SLinus Torvalds * invalid. OK, let's make this for now: 931da177e4SLinus Torvalds */ 94cf533ea5SEric Dumazet static inline __u32 tcp_acceptable_seq(const struct sock *sk) 951da177e4SLinus Torvalds { 96cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 979e412ba7SIlpo Järvinen 98a4ecb15aSCui, Cheng if (!before(tcp_wnd_end(tp), tp->snd_nxt) || 99a4ecb15aSCui, Cheng (tp->rx_opt.wscale_ok && 100a4ecb15aSCui, Cheng ((tp->snd_nxt - tcp_wnd_end(tp)) < (1 << tp->rx_opt.rcv_wscale)))) 1011da177e4SLinus Torvalds return tp->snd_nxt; 1021da177e4SLinus Torvalds else 10390840defSIlpo Järvinen return tcp_wnd_end(tp); 1041da177e4SLinus Torvalds } 1051da177e4SLinus Torvalds 1061da177e4SLinus Torvalds /* Calculate mss to advertise in SYN segment. 1071da177e4SLinus Torvalds * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that: 1081da177e4SLinus Torvalds * 1091da177e4SLinus Torvalds * 1. It is independent of path mtu. 1101da177e4SLinus Torvalds * 2. Ideally, it is maximal possible segment size i.e. 65535-40. 1111da177e4SLinus Torvalds * 3. For IPv4 it is reasonable to calculate it from maximal MTU of 1121da177e4SLinus Torvalds * attached devices, because some buggy hosts are confused by 1131da177e4SLinus Torvalds * large MSS. 1141da177e4SLinus Torvalds * 4. We do not make 3, we advertise MSS, calculated from first 1151da177e4SLinus Torvalds * hop device mtu, but allow to raise it to ip_rt_min_advmss. 1161da177e4SLinus Torvalds * This may be overridden via information stored in routing table. 1171da177e4SLinus Torvalds * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible, 1181da177e4SLinus Torvalds * probably even Jumbo". 1191da177e4SLinus Torvalds */ 1201da177e4SLinus Torvalds static __u16 tcp_advertise_mss(struct sock *sk) 1211da177e4SLinus Torvalds { 1221da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 123cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1241da177e4SLinus Torvalds int mss = tp->advmss; 1251da177e4SLinus Torvalds 1260dbaee3bSDavid S. Miller if (dst) { 1270dbaee3bSDavid S. Miller unsigned int metric = dst_metric_advmss(dst); 1280dbaee3bSDavid S. Miller 1290dbaee3bSDavid S. Miller if (metric < mss) { 1300dbaee3bSDavid S. Miller mss = metric; 1311da177e4SLinus Torvalds tp->advmss = mss; 1321da177e4SLinus Torvalds } 1330dbaee3bSDavid S. Miller } 1341da177e4SLinus Torvalds 1351da177e4SLinus Torvalds return (__u16)mss; 1361da177e4SLinus Torvalds } 1371da177e4SLinus Torvalds 1381da177e4SLinus Torvalds /* RFC2861. Reset CWND after idle period longer RTO to "restart window". 1396f021c62SEric Dumazet * This is the first part of cwnd validation mechanism. 1406f021c62SEric Dumazet */ 1416f021c62SEric Dumazet void tcp_cwnd_restart(struct sock *sk, s32 delta) 1421da177e4SLinus Torvalds { 143463c84b9SArnaldo Carvalho de Melo struct tcp_sock *tp = tcp_sk(sk); 1446f021c62SEric Dumazet u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk)); 1451da177e4SLinus Torvalds u32 cwnd = tp->snd_cwnd; 1461da177e4SLinus Torvalds 1476687e988SArnaldo Carvalho de Melo tcp_ca_event(sk, CA_EVENT_CWND_RESTART); 1481da177e4SLinus Torvalds 1496687e988SArnaldo Carvalho de Melo tp->snd_ssthresh = tcp_current_ssthresh(sk); 1501da177e4SLinus Torvalds restart_cwnd = min(restart_cwnd, cwnd); 1511da177e4SLinus Torvalds 152463c84b9SArnaldo Carvalho de Melo while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd) 1531da177e4SLinus Torvalds cwnd >>= 1; 1541da177e4SLinus Torvalds tp->snd_cwnd = max(cwnd, restart_cwnd); 155c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 1561da177e4SLinus Torvalds tp->snd_cwnd_used = 0; 1571da177e4SLinus Torvalds } 1581da177e4SLinus Torvalds 15967edfef7SAndi Kleen /* Congestion state accounting after a packet has been sent. */ 16040efc6faSStephen Hemminger static void tcp_event_data_sent(struct tcp_sock *tp, 161cf533ea5SEric Dumazet struct sock *sk) 1621da177e4SLinus Torvalds { 163463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 164d635fbe2SEric Dumazet const u32 now = tcp_jiffies32; 1651da177e4SLinus Torvalds 16605c5a46dSNeal Cardwell if (tcp_packets_in_flight(tp) == 0) 16705c5a46dSNeal Cardwell tcp_ca_event(sk, CA_EVENT_TX_START); 16805c5a46dSNeal Cardwell 1694a41f453SWei Wang /* If this is the first data packet sent in response to the 1704a41f453SWei Wang * previous received data, 1714a41f453SWei Wang * and it is a reply for ato after last received packet, 1724a41f453SWei Wang * increase pingpong count. 1731da177e4SLinus Torvalds */ 1744a41f453SWei Wang if (before(tp->lsndtime, icsk->icsk_ack.lrcvtime) && 1754a41f453SWei Wang (u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato) 1764a41f453SWei Wang inet_csk_inc_pingpong_cnt(sk); 1774a41f453SWei Wang 1784a41f453SWei Wang tp->lsndtime = now; 1791da177e4SLinus Torvalds } 1801da177e4SLinus Torvalds 18167edfef7SAndi Kleen /* Account for an ACK we sent. */ 18227cde44aSYuchung Cheng static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts, 18327cde44aSYuchung Cheng u32 rcv_nxt) 1841da177e4SLinus Torvalds { 1855d9f4262SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 1865d9f4262SEric Dumazet 1872b195850SEric Dumazet if (unlikely(tp->compressed_ack)) { 188200d95f4SEric Dumazet NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED, 1892b195850SEric Dumazet tp->compressed_ack); 1902b195850SEric Dumazet tp->compressed_ack = 0; 1915d9f4262SEric Dumazet if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1) 1925d9f4262SEric Dumazet __sock_put(sk); 1935d9f4262SEric Dumazet } 19427cde44aSYuchung Cheng 19527cde44aSYuchung Cheng if (unlikely(rcv_nxt != tp->rcv_nxt)) 19627cde44aSYuchung Cheng return; /* Special ACK sent by DCTCP to reflect ECN */ 197463c84b9SArnaldo Carvalho de Melo tcp_dec_quickack_mode(sk, pkts); 198463c84b9SArnaldo Carvalho de Melo inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); 1991da177e4SLinus Torvalds } 2001da177e4SLinus Torvalds 2011da177e4SLinus Torvalds /* Determine a window scaling and initial window to offer. 2021da177e4SLinus Torvalds * Based on the assumption that the given amount of space 2031da177e4SLinus Torvalds * will be offered. Store the results in the tp structure. 2041da177e4SLinus Torvalds * NOTE: for smooth operation initial space offering should 2051da177e4SLinus Torvalds * be a multiple of mss if possible. We assume here that mss >= 1. 2061da177e4SLinus Torvalds * This MUST be enforced by all callers. 2071da177e4SLinus Torvalds */ 208ceef9ab6SEric Dumazet void tcp_select_initial_window(const struct sock *sk, int __space, __u32 mss, 2091da177e4SLinus Torvalds __u32 *rcv_wnd, __u32 *window_clamp, 21031d12926Slaurent chavey int wscale_ok, __u8 *rcv_wscale, 21131d12926Slaurent chavey __u32 init_rcv_wnd) 2121da177e4SLinus Torvalds { 2131da177e4SLinus Torvalds unsigned int space = (__space < 0 ? 0 : __space); 2141da177e4SLinus Torvalds 2151da177e4SLinus Torvalds /* If no clamp set the clamp to the max possible scaled window */ 2161da177e4SLinus Torvalds if (*window_clamp == 0) 217589c49cbSGao Feng (*window_clamp) = (U16_MAX << TCP_MAX_WSCALE); 2181da177e4SLinus Torvalds space = min(*window_clamp, space); 2191da177e4SLinus Torvalds 2201da177e4SLinus Torvalds /* Quantize space offering to a multiple of mss if possible. */ 2211da177e4SLinus Torvalds if (space > mss) 222589c49cbSGao Feng space = rounddown(space, mss); 2231da177e4SLinus Torvalds 2241da177e4SLinus Torvalds /* NOTE: offering an initial window larger than 32767 22515d99e02SRick Jones * will break some buggy TCP stacks. If the admin tells us 22615d99e02SRick Jones * it is likely we could be speaking with such a buggy stack 22715d99e02SRick Jones * we will truncate our initial window offering to 32K-1 22815d99e02SRick Jones * unless the remote has sent us a window scaling option, 22915d99e02SRick Jones * which we interpret as a sign the remote TCP is not 23015d99e02SRick Jones * misinterpreting the window field as a signed quantity. 2311da177e4SLinus Torvalds */ 232ceef9ab6SEric Dumazet if (sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows) 2331da177e4SLinus Torvalds (*rcv_wnd) = min(space, MAX_TCP_WINDOW); 23415d99e02SRick Jones else 235a337531bSYuchung Cheng (*rcv_wnd) = min_t(u32, space, U16_MAX); 236a337531bSYuchung Cheng 237a337531bSYuchung Cheng if (init_rcv_wnd) 238a337531bSYuchung Cheng *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss); 23915d99e02SRick Jones 24019bf6261SEric Dumazet *rcv_wscale = 0; 2411da177e4SLinus Torvalds if (wscale_ok) { 242589c49cbSGao Feng /* Set window scaling on max possible window */ 243356d1833SEric Dumazet space = max_t(u32, space, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]); 244f626300aSSoheil Hassas Yeganeh space = max_t(u32, space, sysctl_rmem_max); 245316c1592SStephen Hemminger space = min_t(u32, space, *window_clamp); 24619bf6261SEric Dumazet *rcv_wscale = clamp_t(int, ilog2(space) - 15, 24719bf6261SEric Dumazet 0, TCP_MAX_WSCALE); 2481da177e4SLinus Torvalds } 2491da177e4SLinus Torvalds /* Set the clamp no higher than max representable value */ 250589c49cbSGao Feng (*window_clamp) = min_t(__u32, U16_MAX << (*rcv_wscale), *window_clamp); 2511da177e4SLinus Torvalds } 2524bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_select_initial_window); 2531da177e4SLinus Torvalds 2541da177e4SLinus Torvalds /* Chose a new window to advertise, update state in tcp_sock for the 2551da177e4SLinus Torvalds * socket, and return result with RFC1323 scaling applied. The return 2561da177e4SLinus Torvalds * value can be stuffed directly into th->window for an outgoing 2571da177e4SLinus Torvalds * frame. 2581da177e4SLinus Torvalds */ 25940efc6faSStephen Hemminger static u16 tcp_select_window(struct sock *sk) 2601da177e4SLinus Torvalds { 2611da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2628e165e20SFlorian Westphal u32 old_win = tp->rcv_wnd; 2631da177e4SLinus Torvalds u32 cur_win = tcp_receive_window(tp); 2641da177e4SLinus Torvalds u32 new_win = __tcp_select_window(sk); 2651da177e4SLinus Torvalds 2661da177e4SLinus Torvalds /* Never shrink the offered window */ 2671da177e4SLinus Torvalds if (new_win < cur_win) { 2681da177e4SLinus Torvalds /* Danger Will Robinson! 2691da177e4SLinus Torvalds * Don't update rcv_wup/rcv_wnd here or else 2701da177e4SLinus Torvalds * we will not be able to advertise a zero 2711da177e4SLinus Torvalds * window in time. --DaveM 2721da177e4SLinus Torvalds * 2731da177e4SLinus Torvalds * Relax Will Robinson. 2741da177e4SLinus Torvalds */ 2758e165e20SFlorian Westphal if (new_win == 0) 2768e165e20SFlorian Westphal NET_INC_STATS(sock_net(sk), 2778e165e20SFlorian Westphal LINUX_MIB_TCPWANTZEROWINDOWADV); 278607bfbf2SPatrick McHardy new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale); 2791da177e4SLinus Torvalds } 2801da177e4SLinus Torvalds tp->rcv_wnd = new_win; 2811da177e4SLinus Torvalds tp->rcv_wup = tp->rcv_nxt; 2821da177e4SLinus Torvalds 2831da177e4SLinus Torvalds /* Make sure we do not exceed the maximum possible 2841da177e4SLinus Torvalds * scaled window. 2851da177e4SLinus Torvalds */ 286ceef9ab6SEric Dumazet if (!tp->rx_opt.rcv_wscale && 287ceef9ab6SEric Dumazet sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows) 2881da177e4SLinus Torvalds new_win = min(new_win, MAX_TCP_WINDOW); 2891da177e4SLinus Torvalds else 2901da177e4SLinus Torvalds new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); 2911da177e4SLinus Torvalds 2921da177e4SLinus Torvalds /* RFC1323 scaling applied */ 2931da177e4SLinus Torvalds new_win >>= tp->rx_opt.rcv_wscale; 2941da177e4SLinus Torvalds 29531770e34SFlorian Westphal /* If we advertise zero window, disable fast path. */ 2968e165e20SFlorian Westphal if (new_win == 0) { 29731770e34SFlorian Westphal tp->pred_flags = 0; 2988e165e20SFlorian Westphal if (old_win) 2998e165e20SFlorian Westphal NET_INC_STATS(sock_net(sk), 3008e165e20SFlorian Westphal LINUX_MIB_TCPTOZEROWINDOWADV); 3018e165e20SFlorian Westphal } else if (old_win == 0) { 3028e165e20SFlorian Westphal NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV); 3038e165e20SFlorian Westphal } 3041da177e4SLinus Torvalds 3051da177e4SLinus Torvalds return new_win; 3061da177e4SLinus Torvalds } 3071da177e4SLinus Torvalds 30867edfef7SAndi Kleen /* Packet ECN state for a SYN-ACK */ 309735d3831SFlorian Westphal static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb) 310bdf1ee5dSIlpo Järvinen { 31130e502a3SDaniel Borkmann const struct tcp_sock *tp = tcp_sk(sk); 31230e502a3SDaniel Borkmann 3134de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR; 314bdf1ee5dSIlpo Järvinen if (!(tp->ecn_flags & TCP_ECN_OK)) 3154de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE; 31691b5b21cSLawrence Brakmo else if (tcp_ca_needs_ecn(sk) || 31791b5b21cSLawrence Brakmo tcp_bpf_ca_needs_ecn(sk)) 31830e502a3SDaniel Borkmann INET_ECN_xmit(sk); 319bdf1ee5dSIlpo Järvinen } 320bdf1ee5dSIlpo Järvinen 32167edfef7SAndi Kleen /* Packet ECN state for a SYN. */ 322735d3831SFlorian Westphal static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb) 323bdf1ee5dSIlpo Järvinen { 324bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 32591b5b21cSLawrence Brakmo bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk); 326f7b3bec6SFlorian Westphal bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 || 32791b5b21cSLawrence Brakmo tcp_ca_needs_ecn(sk) || bpf_needs_ecn; 328f7b3bec6SFlorian Westphal 329f7b3bec6SFlorian Westphal if (!use_ecn) { 330f7b3bec6SFlorian Westphal const struct dst_entry *dst = __sk_dst_get(sk); 331f7b3bec6SFlorian Westphal 332f7b3bec6SFlorian Westphal if (dst && dst_feature(dst, RTAX_FEATURE_ECN)) 333f7b3bec6SFlorian Westphal use_ecn = true; 334f7b3bec6SFlorian Westphal } 335bdf1ee5dSIlpo Järvinen 336bdf1ee5dSIlpo Järvinen tp->ecn_flags = 0; 337f7b3bec6SFlorian Westphal 338f7b3bec6SFlorian Westphal if (use_ecn) { 3394de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR; 340bdf1ee5dSIlpo Järvinen tp->ecn_flags = TCP_ECN_OK; 34191b5b21cSLawrence Brakmo if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn) 34230e502a3SDaniel Borkmann INET_ECN_xmit(sk); 343bdf1ee5dSIlpo Järvinen } 344bdf1ee5dSIlpo Järvinen } 345bdf1ee5dSIlpo Järvinen 34649213555SDaniel Borkmann static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb) 34749213555SDaniel Borkmann { 34849213555SDaniel Borkmann if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback) 34949213555SDaniel Borkmann /* tp->ecn_flags are cleared at a later point in time when 35049213555SDaniel Borkmann * SYN ACK is ultimatively being received. 35149213555SDaniel Borkmann */ 35249213555SDaniel Borkmann TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR); 35349213555SDaniel Borkmann } 35449213555SDaniel Borkmann 355735d3831SFlorian Westphal static void 3566ac705b1SEric Dumazet tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th) 357bdf1ee5dSIlpo Järvinen { 3586ac705b1SEric Dumazet if (inet_rsk(req)->ecn_ok) 359bdf1ee5dSIlpo Järvinen th->ece = 1; 360bdf1ee5dSIlpo Järvinen } 361bdf1ee5dSIlpo Järvinen 36267edfef7SAndi Kleen /* Set up ECN state for a packet on a ESTABLISHED socket that is about to 36367edfef7SAndi Kleen * be sent. 36467edfef7SAndi Kleen */ 365735d3831SFlorian Westphal static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb, 366ea1627c2SEric Dumazet struct tcphdr *th, int tcp_header_len) 367bdf1ee5dSIlpo Järvinen { 368bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 369bdf1ee5dSIlpo Järvinen 370bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_OK) { 371bdf1ee5dSIlpo Järvinen /* Not-retransmitted data segment: set ECT and inject CWR. */ 372bdf1ee5dSIlpo Järvinen if (skb->len != tcp_header_len && 373bdf1ee5dSIlpo Järvinen !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { 374bdf1ee5dSIlpo Järvinen INET_ECN_xmit(sk); 375bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) { 376bdf1ee5dSIlpo Järvinen tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; 377ea1627c2SEric Dumazet th->cwr = 1; 378bdf1ee5dSIlpo Järvinen skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; 379bdf1ee5dSIlpo Järvinen } 38030e502a3SDaniel Borkmann } else if (!tcp_ca_needs_ecn(sk)) { 381bdf1ee5dSIlpo Järvinen /* ACK or retransmitted segment: clear ECT|CE */ 382bdf1ee5dSIlpo Järvinen INET_ECN_dontxmit(sk); 383bdf1ee5dSIlpo Järvinen } 384bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) 385ea1627c2SEric Dumazet th->ece = 1; 386bdf1ee5dSIlpo Järvinen } 387bdf1ee5dSIlpo Järvinen } 388bdf1ee5dSIlpo Järvinen 389e870a8efSIlpo Järvinen /* Constructs common control bits of non-data skb. If SYN/FIN is present, 390e870a8efSIlpo Järvinen * auto increment end seqno. 391e870a8efSIlpo Järvinen */ 392e870a8efSIlpo Järvinen static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags) 393e870a8efSIlpo Järvinen { 3942e8e18efSDavid S. Miller skb->ip_summed = CHECKSUM_PARTIAL; 395e870a8efSIlpo Järvinen 3964de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags; 397e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->sacked = 0; 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 41233ad798cSAdam Langley #define OPTION_SACK_ADVERTISE (1 << 0) 41333ad798cSAdam Langley #define OPTION_TS (1 << 1) 41433ad798cSAdam Langley #define OPTION_MD5 (1 << 2) 41589e95a61SOri Finkelman #define OPTION_WSCALE (1 << 3) 4162100c8d2SYuchung Cheng #define OPTION_FAST_OPEN_COOKIE (1 << 8) 41760e2a778SUrsula Braun #define OPTION_SMC (1 << 9) 418eda7acddSPeter Krystad #define OPTION_MPTCP (1 << 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 448fa3fe2b1SFlorian Westphal static void mptcp_options_write(__be32 *ptr, const struct tcp_sock *tp, 449fa3fe2b1SFlorian Westphal struct tcp_out_options *opts) 450eda7acddSPeter Krystad { 451eda7acddSPeter Krystad #if IS_ENABLED(CONFIG_MPTCP) 452eda7acddSPeter Krystad if (unlikely(OPTION_MPTCP & opts->options)) 453fa3fe2b1SFlorian Westphal mptcp_write_options(ptr, tp, &opts->mptcp); 454eda7acddSPeter Krystad #endif 455eda7acddSPeter Krystad } 456eda7acddSPeter Krystad 457331fca43SMartin KaFai Lau #ifdef CONFIG_CGROUP_BPF 4580813a841SMartin KaFai Lau static int bpf_skops_write_hdr_opt_arg0(struct sk_buff *skb, 4590813a841SMartin KaFai Lau enum tcp_synack_type synack_type) 4600813a841SMartin KaFai Lau { 4610813a841SMartin KaFai Lau if (unlikely(!skb)) 4620813a841SMartin KaFai Lau return BPF_WRITE_HDR_TCP_CURRENT_MSS; 4630813a841SMartin KaFai Lau 4640813a841SMartin KaFai Lau if (unlikely(synack_type == TCP_SYNACK_COOKIE)) 4650813a841SMartin KaFai Lau return BPF_WRITE_HDR_TCP_SYNACK_COOKIE; 4660813a841SMartin KaFai Lau 4670813a841SMartin KaFai Lau return 0; 4680813a841SMartin KaFai Lau } 4690813a841SMartin KaFai Lau 470331fca43SMartin KaFai Lau /* req, syn_skb and synack_type are used when writing synack */ 471331fca43SMartin KaFai Lau static void bpf_skops_hdr_opt_len(struct sock *sk, struct sk_buff *skb, 472331fca43SMartin KaFai Lau struct request_sock *req, 473331fca43SMartin KaFai Lau struct sk_buff *syn_skb, 474331fca43SMartin KaFai Lau enum tcp_synack_type synack_type, 475331fca43SMartin KaFai Lau struct tcp_out_options *opts, 476331fca43SMartin KaFai Lau unsigned int *remaining) 477331fca43SMartin KaFai Lau { 4780813a841SMartin KaFai Lau struct bpf_sock_ops_kern sock_ops; 4790813a841SMartin KaFai Lau int err; 4800813a841SMartin KaFai Lau 481331fca43SMartin KaFai Lau if (likely(!BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), 482331fca43SMartin KaFai Lau BPF_SOCK_OPS_WRITE_HDR_OPT_CB_FLAG)) || 483331fca43SMartin KaFai Lau !*remaining) 484331fca43SMartin KaFai Lau return; 485331fca43SMartin KaFai Lau 4860813a841SMartin KaFai Lau /* *remaining has already been aligned to 4 bytes, so *remaining >= 4 */ 4870813a841SMartin KaFai Lau 4880813a841SMartin KaFai Lau /* init sock_ops */ 4890813a841SMartin KaFai Lau memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp)); 4900813a841SMartin KaFai Lau 4910813a841SMartin KaFai Lau sock_ops.op = BPF_SOCK_OPS_HDR_OPT_LEN_CB; 4920813a841SMartin KaFai Lau 4930813a841SMartin KaFai Lau if (req) { 4940813a841SMartin KaFai Lau /* The listen "sk" cannot be passed here because 4950813a841SMartin KaFai Lau * it is not locked. It would not make too much 4960813a841SMartin KaFai Lau * sense to do bpf_setsockopt(listen_sk) based 4970813a841SMartin KaFai Lau * on individual connection request also. 4980813a841SMartin KaFai Lau * 4990813a841SMartin KaFai Lau * Thus, "req" is passed here and the cgroup-bpf-progs 5000813a841SMartin KaFai Lau * of the listen "sk" will be run. 5010813a841SMartin KaFai Lau * 5020813a841SMartin KaFai Lau * "req" is also used here for fastopen even the "sk" here is 5030813a841SMartin KaFai Lau * a fullsock "child" sk. It is to keep the behavior 5040813a841SMartin KaFai Lau * consistent between fastopen and non-fastopen on 5050813a841SMartin KaFai Lau * the bpf programming side. 506331fca43SMartin KaFai Lau */ 5070813a841SMartin KaFai Lau sock_ops.sk = (struct sock *)req; 5080813a841SMartin KaFai Lau sock_ops.syn_skb = syn_skb; 5090813a841SMartin KaFai Lau } else { 5100813a841SMartin KaFai Lau sock_owned_by_me(sk); 5110813a841SMartin KaFai Lau 5120813a841SMartin KaFai Lau sock_ops.is_fullsock = 1; 5130813a841SMartin KaFai Lau sock_ops.sk = sk; 5140813a841SMartin KaFai Lau } 5150813a841SMartin KaFai Lau 5160813a841SMartin KaFai Lau sock_ops.args[0] = bpf_skops_write_hdr_opt_arg0(skb, synack_type); 5170813a841SMartin KaFai Lau sock_ops.remaining_opt_len = *remaining; 5180813a841SMartin KaFai Lau /* tcp_current_mss() does not pass a skb */ 5190813a841SMartin KaFai Lau if (skb) 5200813a841SMartin KaFai Lau bpf_skops_init_skb(&sock_ops, skb, 0); 5210813a841SMartin KaFai Lau 5220813a841SMartin KaFai Lau err = BPF_CGROUP_RUN_PROG_SOCK_OPS_SK(&sock_ops, sk); 5230813a841SMartin KaFai Lau 5240813a841SMartin KaFai Lau if (err || sock_ops.remaining_opt_len == *remaining) 5250813a841SMartin KaFai Lau return; 5260813a841SMartin KaFai Lau 5270813a841SMartin KaFai Lau opts->bpf_opt_len = *remaining - sock_ops.remaining_opt_len; 5280813a841SMartin KaFai Lau /* round up to 4 bytes */ 5290813a841SMartin KaFai Lau opts->bpf_opt_len = (opts->bpf_opt_len + 3) & ~3; 5300813a841SMartin KaFai Lau 5310813a841SMartin KaFai Lau *remaining -= opts->bpf_opt_len; 532331fca43SMartin KaFai Lau } 533331fca43SMartin KaFai Lau 534331fca43SMartin KaFai Lau static void bpf_skops_write_hdr_opt(struct sock *sk, struct sk_buff *skb, 535331fca43SMartin KaFai Lau struct request_sock *req, 536331fca43SMartin KaFai Lau struct sk_buff *syn_skb, 537331fca43SMartin KaFai Lau enum tcp_synack_type synack_type, 538331fca43SMartin KaFai Lau struct tcp_out_options *opts) 539331fca43SMartin KaFai Lau { 5400813a841SMartin KaFai Lau u8 first_opt_off, nr_written, max_opt_len = opts->bpf_opt_len; 5410813a841SMartin KaFai Lau struct bpf_sock_ops_kern sock_ops; 5420813a841SMartin KaFai Lau int err; 5430813a841SMartin KaFai Lau 5440813a841SMartin KaFai Lau if (likely(!max_opt_len)) 545331fca43SMartin KaFai Lau return; 546331fca43SMartin KaFai Lau 5470813a841SMartin KaFai Lau memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp)); 5480813a841SMartin KaFai Lau 5490813a841SMartin KaFai Lau sock_ops.op = BPF_SOCK_OPS_WRITE_HDR_OPT_CB; 5500813a841SMartin KaFai Lau 5510813a841SMartin KaFai Lau if (req) { 5520813a841SMartin KaFai Lau sock_ops.sk = (struct sock *)req; 5530813a841SMartin KaFai Lau sock_ops.syn_skb = syn_skb; 5540813a841SMartin KaFai Lau } else { 5550813a841SMartin KaFai Lau sock_owned_by_me(sk); 5560813a841SMartin KaFai Lau 5570813a841SMartin KaFai Lau sock_ops.is_fullsock = 1; 5580813a841SMartin KaFai Lau sock_ops.sk = sk; 5590813a841SMartin KaFai Lau } 5600813a841SMartin KaFai Lau 5610813a841SMartin KaFai Lau sock_ops.args[0] = bpf_skops_write_hdr_opt_arg0(skb, synack_type); 5620813a841SMartin KaFai Lau sock_ops.remaining_opt_len = max_opt_len; 5630813a841SMartin KaFai Lau first_opt_off = tcp_hdrlen(skb) - max_opt_len; 5640813a841SMartin KaFai Lau bpf_skops_init_skb(&sock_ops, skb, first_opt_off); 5650813a841SMartin KaFai Lau 5660813a841SMartin KaFai Lau err = BPF_CGROUP_RUN_PROG_SOCK_OPS_SK(&sock_ops, sk); 5670813a841SMartin KaFai Lau 5680813a841SMartin KaFai Lau if (err) 5690813a841SMartin KaFai Lau nr_written = 0; 5700813a841SMartin KaFai Lau else 5710813a841SMartin KaFai Lau nr_written = max_opt_len - sock_ops.remaining_opt_len; 5720813a841SMartin KaFai Lau 5730813a841SMartin KaFai Lau if (nr_written < max_opt_len) 5740813a841SMartin KaFai Lau memset(skb->data + first_opt_off + nr_written, TCPOPT_NOP, 5750813a841SMartin KaFai Lau max_opt_len - nr_written); 576331fca43SMartin KaFai Lau } 577331fca43SMartin KaFai Lau #else 578331fca43SMartin KaFai Lau static void bpf_skops_hdr_opt_len(struct sock *sk, struct sk_buff *skb, 579331fca43SMartin KaFai Lau struct request_sock *req, 580331fca43SMartin KaFai Lau struct sk_buff *syn_skb, 581331fca43SMartin KaFai Lau enum tcp_synack_type synack_type, 582331fca43SMartin KaFai Lau struct tcp_out_options *opts, 583331fca43SMartin KaFai Lau unsigned int *remaining) 584331fca43SMartin KaFai Lau { 585331fca43SMartin KaFai Lau } 586331fca43SMartin KaFai Lau 587331fca43SMartin KaFai Lau static void bpf_skops_write_hdr_opt(struct sock *sk, struct sk_buff *skb, 588331fca43SMartin KaFai Lau struct request_sock *req, 589331fca43SMartin KaFai Lau struct sk_buff *syn_skb, 590331fca43SMartin KaFai Lau enum tcp_synack_type synack_type, 591331fca43SMartin KaFai Lau struct tcp_out_options *opts) 592331fca43SMartin KaFai Lau { 593331fca43SMartin KaFai Lau } 594331fca43SMartin KaFai Lau #endif 595331fca43SMartin KaFai Lau 59667edfef7SAndi Kleen /* Write previously computed TCP options to the packet. 59767edfef7SAndi Kleen * 59867edfef7SAndi Kleen * Beware: Something in the Internet is very sensitive to the ordering of 599fd6149d3SIlpo Järvinen * TCP options, we learned this through the hard way, so be careful here. 600fd6149d3SIlpo Järvinen * Luckily we can at least blame others for their non-compliance but from 6018e3bff96Sstephen hemminger * inter-operability perspective it seems that we're somewhat stuck with 602fd6149d3SIlpo Järvinen * the ordering which we have been using if we want to keep working with 603fd6149d3SIlpo Järvinen * those broken things (not that it currently hurts anybody as there isn't 604fd6149d3SIlpo Järvinen * particular reason why the ordering would need to be changed). 605fd6149d3SIlpo Järvinen * 606fd6149d3SIlpo Järvinen * At least SACK_PERM as the first option is known to lead to a disaster 607fd6149d3SIlpo Järvinen * (but it may well be that other scenarios fail similarly). 608fd6149d3SIlpo Järvinen */ 60933ad798cSAdam Langley static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp, 610bd0388aeSWilliam Allen Simpson struct tcp_out_options *opts) 611bd0388aeSWilliam Allen Simpson { 6122100c8d2SYuchung Cheng u16 options = opts->options; /* mungable copy */ 613bd0388aeSWilliam Allen Simpson 614bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_MD5 & options)) { 6151a2c6181SChristoph Paasch *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | 6161a2c6181SChristoph Paasch (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG); 617bd0388aeSWilliam Allen Simpson /* overload cookie hash location */ 618bd0388aeSWilliam Allen Simpson opts->hash_location = (__u8 *)ptr; 61933ad798cSAdam Langley ptr += 4; 62033ad798cSAdam Langley } 62133ad798cSAdam Langley 622fd6149d3SIlpo Järvinen if (unlikely(opts->mss)) { 623fd6149d3SIlpo Järvinen *ptr++ = htonl((TCPOPT_MSS << 24) | 624fd6149d3SIlpo Järvinen (TCPOLEN_MSS << 16) | 625fd6149d3SIlpo Järvinen opts->mss); 626fd6149d3SIlpo Järvinen } 627fd6149d3SIlpo Järvinen 628bd0388aeSWilliam Allen Simpson if (likely(OPTION_TS & options)) { 629bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 63033ad798cSAdam Langley *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | 63133ad798cSAdam Langley (TCPOLEN_SACK_PERM << 16) | 63233ad798cSAdam Langley (TCPOPT_TIMESTAMP << 8) | 63333ad798cSAdam Langley TCPOLEN_TIMESTAMP); 634bd0388aeSWilliam Allen Simpson options &= ~OPTION_SACK_ADVERTISE; 63533ad798cSAdam Langley } else { 636496c98dfSYOSHIFUJI Hideaki *ptr++ = htonl((TCPOPT_NOP << 24) | 63740efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 63840efc6faSStephen Hemminger (TCPOPT_TIMESTAMP << 8) | 63940efc6faSStephen Hemminger TCPOLEN_TIMESTAMP); 64040efc6faSStephen Hemminger } 64133ad798cSAdam Langley *ptr++ = htonl(opts->tsval); 64233ad798cSAdam Langley *ptr++ = htonl(opts->tsecr); 64333ad798cSAdam Langley } 64433ad798cSAdam Langley 645bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 64633ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 64733ad798cSAdam Langley (TCPOPT_NOP << 16) | 64833ad798cSAdam Langley (TCPOPT_SACK_PERM << 8) | 64933ad798cSAdam Langley TCPOLEN_SACK_PERM); 65033ad798cSAdam Langley } 65133ad798cSAdam Langley 652bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_WSCALE & options)) { 65333ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 65433ad798cSAdam Langley (TCPOPT_WINDOW << 16) | 65533ad798cSAdam Langley (TCPOLEN_WINDOW << 8) | 65633ad798cSAdam Langley opts->ws); 65733ad798cSAdam Langley } 65833ad798cSAdam Langley 65933ad798cSAdam Langley if (unlikely(opts->num_sack_blocks)) { 66033ad798cSAdam Langley struct tcp_sack_block *sp = tp->rx_opt.dsack ? 66133ad798cSAdam Langley tp->duplicate_sack : tp->selective_acks; 66240efc6faSStephen Hemminger int this_sack; 66340efc6faSStephen Hemminger 66440efc6faSStephen Hemminger *ptr++ = htonl((TCPOPT_NOP << 24) | 66540efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 66640efc6faSStephen Hemminger (TCPOPT_SACK << 8) | 66733ad798cSAdam Langley (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * 66840efc6faSStephen Hemminger TCPOLEN_SACK_PERBLOCK))); 6692de979bdSStephen Hemminger 67033ad798cSAdam Langley for (this_sack = 0; this_sack < opts->num_sack_blocks; 67133ad798cSAdam Langley ++this_sack) { 67240efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].start_seq); 67340efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].end_seq); 67440efc6faSStephen Hemminger } 6752de979bdSStephen Hemminger 67640efc6faSStephen Hemminger tp->rx_opt.dsack = 0; 67740efc6faSStephen Hemminger } 6782100c8d2SYuchung Cheng 6792100c8d2SYuchung Cheng if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) { 6802100c8d2SYuchung Cheng struct tcp_fastopen_cookie *foc = opts->fastopen_cookie; 6817f9b838bSDaniel Lee u8 *p = (u8 *)ptr; 6827f9b838bSDaniel Lee u32 len; /* Fast Open option length */ 6832100c8d2SYuchung Cheng 6847f9b838bSDaniel Lee if (foc->exp) { 6857f9b838bSDaniel Lee len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len; 6867f9b838bSDaniel Lee *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) | 6872100c8d2SYuchung Cheng TCPOPT_FASTOPEN_MAGIC); 6887f9b838bSDaniel Lee p += TCPOLEN_EXP_FASTOPEN_BASE; 6897f9b838bSDaniel Lee } else { 6907f9b838bSDaniel Lee len = TCPOLEN_FASTOPEN_BASE + foc->len; 6917f9b838bSDaniel Lee *p++ = TCPOPT_FASTOPEN; 6927f9b838bSDaniel Lee *p++ = len; 6932100c8d2SYuchung Cheng } 6947f9b838bSDaniel Lee 6957f9b838bSDaniel Lee memcpy(p, foc->val, foc->len); 6967f9b838bSDaniel Lee if ((len & 3) == 2) { 6977f9b838bSDaniel Lee p[foc->len] = TCPOPT_NOP; 6987f9b838bSDaniel Lee p[foc->len + 1] = TCPOPT_NOP; 6997f9b838bSDaniel Lee } 7007f9b838bSDaniel Lee ptr += (len + 3) >> 2; 7012100c8d2SYuchung Cheng } 70260e2a778SUrsula Braun 70360e2a778SUrsula Braun smc_options_write(ptr, &options); 704eda7acddSPeter Krystad 705fa3fe2b1SFlorian Westphal mptcp_options_write(ptr, tp, opts); 70660e2a778SUrsula Braun } 70760e2a778SUrsula Braun 70860e2a778SUrsula Braun static void smc_set_option(const struct tcp_sock *tp, 70960e2a778SUrsula Braun struct tcp_out_options *opts, 71060e2a778SUrsula Braun unsigned int *remaining) 71160e2a778SUrsula Braun { 71260e2a778SUrsula Braun #if IS_ENABLED(CONFIG_SMC) 71360e2a778SUrsula Braun if (static_branch_unlikely(&tcp_have_smc)) { 71460e2a778SUrsula Braun if (tp->syn_smc) { 71560e2a778SUrsula Braun if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) { 71660e2a778SUrsula Braun opts->options |= OPTION_SMC; 71760e2a778SUrsula Braun *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED; 71860e2a778SUrsula Braun } 71960e2a778SUrsula Braun } 72060e2a778SUrsula Braun } 72160e2a778SUrsula Braun #endif 72260e2a778SUrsula Braun } 72360e2a778SUrsula Braun 72460e2a778SUrsula Braun static void smc_set_option_cond(const struct tcp_sock *tp, 72560e2a778SUrsula Braun const struct inet_request_sock *ireq, 72660e2a778SUrsula Braun struct tcp_out_options *opts, 72760e2a778SUrsula Braun unsigned int *remaining) 72860e2a778SUrsula Braun { 72960e2a778SUrsula Braun #if IS_ENABLED(CONFIG_SMC) 73060e2a778SUrsula Braun if (static_branch_unlikely(&tcp_have_smc)) { 73160e2a778SUrsula Braun if (tp->syn_smc && ireq->smc_ok) { 73260e2a778SUrsula Braun if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) { 73360e2a778SUrsula Braun opts->options |= OPTION_SMC; 73460e2a778SUrsula Braun *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED; 73560e2a778SUrsula Braun } 73660e2a778SUrsula Braun } 73760e2a778SUrsula Braun } 73860e2a778SUrsula Braun #endif 73940efc6faSStephen Hemminger } 74040efc6faSStephen Hemminger 741cec37a6eSPeter Krystad static void mptcp_set_option_cond(const struct request_sock *req, 742cec37a6eSPeter Krystad struct tcp_out_options *opts, 743cec37a6eSPeter Krystad unsigned int *remaining) 744cec37a6eSPeter Krystad { 745cec37a6eSPeter Krystad if (rsk_is_mptcp(req)) { 746cec37a6eSPeter Krystad unsigned int size; 747cec37a6eSPeter Krystad 748cec37a6eSPeter Krystad if (mptcp_synack_options(req, &size, &opts->mptcp)) { 749cec37a6eSPeter Krystad if (*remaining >= size) { 750cec37a6eSPeter Krystad opts->options |= OPTION_MPTCP; 751cec37a6eSPeter Krystad *remaining -= size; 752cec37a6eSPeter Krystad } 753cec37a6eSPeter Krystad } 754cec37a6eSPeter Krystad } 755cec37a6eSPeter Krystad } 756cec37a6eSPeter Krystad 75767edfef7SAndi Kleen /* Compute TCP options for SYN packets. This is not the final 75867edfef7SAndi Kleen * network wire format yet. 75967edfef7SAndi Kleen */ 76095c96174SEric Dumazet static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb, 76133ad798cSAdam Langley struct tcp_out_options *opts, 762cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 763cf533ea5SEric Dumazet { 76433ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 76595c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 766783237e8SYuchung Cheng struct tcp_fastopen_request *fastopen = tp->fastopen_req; 76733ad798cSAdam Langley 7688c2320e8SEric Dumazet *md5 = NULL; 769cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 770921f9a0fSEric Dumazet if (static_branch_unlikely(&tcp_md5_needed) && 7716015c71eSEric Dumazet rcu_access_pointer(tp->md5sig_info)) { 77233ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 77333ad798cSAdam Langley if (*md5) { 77433ad798cSAdam Langley opts->options |= OPTION_MD5; 775bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 776cfb6eeb4SYOSHIFUJI Hideaki } 7778c2320e8SEric Dumazet } 778cfb6eeb4SYOSHIFUJI Hideaki #endif 77933ad798cSAdam Langley 78033ad798cSAdam Langley /* We always get an MSS option. The option bytes which will be seen in 78133ad798cSAdam Langley * normal data packets should timestamps be used, must be in the MSS 78233ad798cSAdam Langley * advertised. But we subtract them from tp->mss_cache so that 78333ad798cSAdam Langley * calculations in tcp_sendmsg are simpler etc. So account for this 78433ad798cSAdam Langley * fact here if necessary. If we don't do this correctly, as a 78533ad798cSAdam Langley * receiver we won't recognize data packets as being full sized when we 78633ad798cSAdam Langley * should, and thus we won't abide by the delayed ACK rules correctly. 78733ad798cSAdam Langley * SACKs don't matter, we never delay an ACK when we have any of those 78833ad798cSAdam Langley * going out. */ 78933ad798cSAdam Langley opts->mss = tcp_advertise_mss(sk); 790bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 79133ad798cSAdam Langley 7925d2ed052SEric Dumazet if (likely(sock_net(sk)->ipv4.sysctl_tcp_timestamps && !*md5)) { 79333ad798cSAdam Langley opts->options |= OPTION_TS; 7947faee5c0SEric Dumazet opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset; 79533ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 796bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 79733ad798cSAdam Langley } 7989bb37ef0SEric Dumazet if (likely(sock_net(sk)->ipv4.sysctl_tcp_window_scaling)) { 79933ad798cSAdam Langley opts->ws = tp->rx_opt.rcv_wscale; 80089e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 801bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 80233ad798cSAdam Langley } 803f9301034SEric Dumazet if (likely(sock_net(sk)->ipv4.sysctl_tcp_sack)) { 80433ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 805b32d1310SDavid S. Miller if (unlikely(!(OPTION_TS & opts->options))) 806bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 80733ad798cSAdam Langley } 80833ad798cSAdam Langley 809783237e8SYuchung Cheng if (fastopen && fastopen->cookie.len >= 0) { 8102646c831SDaniel Lee u32 need = fastopen->cookie.len; 8112646c831SDaniel Lee 8122646c831SDaniel Lee need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE : 8132646c831SDaniel Lee TCPOLEN_FASTOPEN_BASE; 814783237e8SYuchung Cheng need = (need + 3) & ~3U; /* Align to 32 bits */ 815783237e8SYuchung Cheng if (remaining >= need) { 816783237e8SYuchung Cheng opts->options |= OPTION_FAST_OPEN_COOKIE; 817783237e8SYuchung Cheng opts->fastopen_cookie = &fastopen->cookie; 818783237e8SYuchung Cheng remaining -= need; 819783237e8SYuchung Cheng tp->syn_fastopen = 1; 8202646c831SDaniel Lee tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0; 821783237e8SYuchung Cheng } 822783237e8SYuchung Cheng } 823bd0388aeSWilliam Allen Simpson 82460e2a778SUrsula Braun smc_set_option(tp, opts, &remaining); 82560e2a778SUrsula Braun 826cec37a6eSPeter Krystad if (sk_is_mptcp(sk)) { 827cec37a6eSPeter Krystad unsigned int size; 828cec37a6eSPeter Krystad 829cc7972eaSChristoph Paasch if (mptcp_syn_options(sk, skb, &size, &opts->mptcp)) { 830cec37a6eSPeter Krystad opts->options |= OPTION_MPTCP; 831cec37a6eSPeter Krystad remaining -= size; 832cec37a6eSPeter Krystad } 833cec37a6eSPeter Krystad } 834cec37a6eSPeter Krystad 835331fca43SMartin KaFai Lau bpf_skops_hdr_opt_len(sk, skb, NULL, NULL, 0, opts, &remaining); 836331fca43SMartin KaFai Lau 837bd0388aeSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 83833ad798cSAdam Langley } 83933ad798cSAdam Langley 84067edfef7SAndi Kleen /* Set up TCP options for SYN-ACKs. */ 84160e2a778SUrsula Braun static unsigned int tcp_synack_options(const struct sock *sk, 84260e2a778SUrsula Braun struct request_sock *req, 84395c96174SEric Dumazet unsigned int mss, struct sk_buff *skb, 84433ad798cSAdam Langley struct tcp_out_options *opts, 84580f03e27SEric Dumazet const struct tcp_md5sig_key *md5, 846e114e1e8SEric Dumazet struct tcp_fastopen_cookie *foc, 847331fca43SMartin KaFai Lau enum tcp_synack_type synack_type, 848331fca43SMartin KaFai Lau struct sk_buff *syn_skb) 8494957faadSWilliam Allen Simpson { 85033ad798cSAdam Langley struct inet_request_sock *ireq = inet_rsk(req); 85195c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 85233ad798cSAdam Langley 85333ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 85480f03e27SEric Dumazet if (md5) { 85533ad798cSAdam Langley opts->options |= OPTION_MD5; 8564957faadSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 8574957faadSWilliam Allen Simpson 8584957faadSWilliam Allen Simpson /* We can't fit any SACK blocks in a packet with MD5 + TS 8594957faadSWilliam Allen Simpson * options. There was discussion about disabling SACK 8604957faadSWilliam Allen Simpson * rather than TS in order to fit in better with old, 8614957faadSWilliam Allen Simpson * buggy kernels, but that was deemed to be unnecessary. 8624957faadSWilliam Allen Simpson */ 863e114e1e8SEric Dumazet if (synack_type != TCP_SYNACK_COOKIE) 864de213e5eSEric Dumazet ireq->tstamp_ok &= !ireq->sack_ok; 86533ad798cSAdam Langley } 86633ad798cSAdam Langley #endif 86733ad798cSAdam Langley 8684957faadSWilliam Allen Simpson /* We always send an MSS option. */ 86933ad798cSAdam Langley opts->mss = mss; 8704957faadSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 87133ad798cSAdam Langley 87233ad798cSAdam Langley if (likely(ireq->wscale_ok)) { 87333ad798cSAdam Langley opts->ws = ireq->rcv_wscale; 87489e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 8754957faadSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 87633ad798cSAdam Langley } 877de213e5eSEric Dumazet if (likely(ireq->tstamp_ok)) { 87833ad798cSAdam Langley opts->options |= OPTION_TS; 87995a22caeSFlorian Westphal opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off; 88033ad798cSAdam Langley opts->tsecr = req->ts_recent; 8814957faadSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 88233ad798cSAdam Langley } 88333ad798cSAdam Langley if (likely(ireq->sack_ok)) { 88433ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 885de213e5eSEric Dumazet if (unlikely(!ireq->tstamp_ok)) 8864957faadSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 88733ad798cSAdam Langley } 8887f9b838bSDaniel Lee if (foc != NULL && foc->len >= 0) { 8897f9b838bSDaniel Lee u32 need = foc->len; 8907f9b838bSDaniel Lee 8917f9b838bSDaniel Lee need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE : 8927f9b838bSDaniel Lee TCPOLEN_FASTOPEN_BASE; 8938336886fSJerry Chu need = (need + 3) & ~3U; /* Align to 32 bits */ 8948336886fSJerry Chu if (remaining >= need) { 8958336886fSJerry Chu opts->options |= OPTION_FAST_OPEN_COOKIE; 8968336886fSJerry Chu opts->fastopen_cookie = foc; 8978336886fSJerry Chu remaining -= need; 8988336886fSJerry Chu } 8998336886fSJerry Chu } 9004957faadSWilliam Allen Simpson 901cec37a6eSPeter Krystad mptcp_set_option_cond(req, opts, &remaining); 902cec37a6eSPeter Krystad 90360e2a778SUrsula Braun smc_set_option_cond(tcp_sk(sk), ireq, opts, &remaining); 90460e2a778SUrsula Braun 905331fca43SMartin KaFai Lau bpf_skops_hdr_opt_len((struct sock *)sk, skb, req, syn_skb, 906331fca43SMartin KaFai Lau synack_type, opts, &remaining); 907331fca43SMartin KaFai Lau 9084957faadSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 90933ad798cSAdam Langley } 91033ad798cSAdam Langley 91167edfef7SAndi Kleen /* Compute TCP options for ESTABLISHED sockets. This is not the 91267edfef7SAndi Kleen * final wire format yet. 91367edfef7SAndi Kleen */ 91495c96174SEric Dumazet static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb, 91533ad798cSAdam Langley struct tcp_out_options *opts, 916cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 917cf533ea5SEric Dumazet { 91833ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 91995c96174SEric Dumazet unsigned int size = 0; 920cabeccbdSIlpo Järvinen unsigned int eff_sacks; 92133ad798cSAdam Langley 9225843ef42SAndi Kleen opts->options = 0; 9235843ef42SAndi Kleen 9248c2320e8SEric Dumazet *md5 = NULL; 92533ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 926921f9a0fSEric Dumazet if (static_branch_unlikely(&tcp_md5_needed) && 9276015c71eSEric Dumazet rcu_access_pointer(tp->md5sig_info)) { 92833ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 9298c2320e8SEric Dumazet if (*md5) { 93033ad798cSAdam Langley opts->options |= OPTION_MD5; 93133ad798cSAdam Langley size += TCPOLEN_MD5SIG_ALIGNED; 93233ad798cSAdam Langley } 9338c2320e8SEric Dumazet } 93433ad798cSAdam Langley #endif 93533ad798cSAdam Langley 93633ad798cSAdam Langley if (likely(tp->rx_opt.tstamp_ok)) { 93733ad798cSAdam Langley opts->options |= OPTION_TS; 9387faee5c0SEric Dumazet opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0; 93933ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 94033ad798cSAdam Langley size += TCPOLEN_TSTAMP_ALIGNED; 94133ad798cSAdam Langley } 94233ad798cSAdam Langley 943cec37a6eSPeter Krystad /* MPTCP options have precedence over SACK for the limited TCP 944cec37a6eSPeter Krystad * option space because a MPTCP connection would be forced to 945cec37a6eSPeter Krystad * fall back to regular TCP if a required multipath option is 946cec37a6eSPeter Krystad * missing. SACK still gets a chance to use whatever space is 947cec37a6eSPeter Krystad * left. 948cec37a6eSPeter Krystad */ 949cec37a6eSPeter Krystad if (sk_is_mptcp(sk)) { 950cec37a6eSPeter Krystad unsigned int remaining = MAX_TCP_OPTION_SPACE - size; 951cec37a6eSPeter Krystad unsigned int opt_size = 0; 952cec37a6eSPeter Krystad 953cec37a6eSPeter Krystad if (mptcp_established_options(sk, skb, &opt_size, remaining, 954cec37a6eSPeter Krystad &opts->mptcp)) { 955cec37a6eSPeter Krystad opts->options |= OPTION_MPTCP; 956cec37a6eSPeter Krystad size += opt_size; 957cec37a6eSPeter Krystad } 958cec37a6eSPeter Krystad } 959cec37a6eSPeter Krystad 960cabeccbdSIlpo Järvinen eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; 961cabeccbdSIlpo Järvinen if (unlikely(eff_sacks)) { 96295c96174SEric Dumazet const unsigned int remaining = MAX_TCP_OPTION_SPACE - size; 9639cfcca23SMat Martineau if (unlikely(remaining < TCPOLEN_SACK_BASE_ALIGNED + 9649cfcca23SMat Martineau TCPOLEN_SACK_PERBLOCK)) 9659cfcca23SMat Martineau return size; 9669cfcca23SMat Martineau 96733ad798cSAdam Langley opts->num_sack_blocks = 96895c96174SEric Dumazet min_t(unsigned int, eff_sacks, 96933ad798cSAdam Langley (remaining - TCPOLEN_SACK_BASE_ALIGNED) / 97033ad798cSAdam Langley TCPOLEN_SACK_PERBLOCK); 9719cfcca23SMat Martineau 97233ad798cSAdam Langley size += TCPOLEN_SACK_BASE_ALIGNED + 97333ad798cSAdam Langley opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; 97433ad798cSAdam Langley } 97533ad798cSAdam Langley 976331fca43SMartin KaFai Lau if (unlikely(BPF_SOCK_OPS_TEST_FLAG(tp, 977331fca43SMartin KaFai Lau BPF_SOCK_OPS_WRITE_HDR_OPT_CB_FLAG))) { 978331fca43SMartin KaFai Lau unsigned int remaining = MAX_TCP_OPTION_SPACE - size; 979331fca43SMartin KaFai Lau 980331fca43SMartin KaFai Lau bpf_skops_hdr_opt_len(sk, skb, NULL, NULL, 0, opts, &remaining); 981331fca43SMartin KaFai Lau 982331fca43SMartin KaFai Lau size = MAX_TCP_OPTION_SPACE - remaining; 983331fca43SMartin KaFai Lau } 984331fca43SMartin KaFai Lau 98533ad798cSAdam Langley return size; 98640efc6faSStephen Hemminger } 9871da177e4SLinus Torvalds 98846d3ceabSEric Dumazet 98946d3ceabSEric Dumazet /* TCP SMALL QUEUES (TSQ) 99046d3ceabSEric Dumazet * 99146d3ceabSEric Dumazet * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev) 99246d3ceabSEric Dumazet * to reduce RTT and bufferbloat. 99346d3ceabSEric Dumazet * We do this using a special skb destructor (tcp_wfree). 99446d3ceabSEric Dumazet * 99546d3ceabSEric Dumazet * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb 99646d3ceabSEric Dumazet * needs to be reallocated in a driver. 9978e3bff96Sstephen hemminger * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc 99846d3ceabSEric Dumazet * 99946d3ceabSEric Dumazet * Since transmit from skb destructor is forbidden, we use a tasklet 100046d3ceabSEric Dumazet * to process all sockets that eventually need to send more skbs. 100146d3ceabSEric Dumazet * We use one tasklet per cpu, with its own queue of sockets. 100246d3ceabSEric Dumazet */ 100346d3ceabSEric Dumazet struct tsq_tasklet { 100446d3ceabSEric Dumazet struct tasklet_struct tasklet; 100546d3ceabSEric Dumazet struct list_head head; /* queue of tcp sockets */ 100646d3ceabSEric Dumazet }; 100746d3ceabSEric Dumazet static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet); 100846d3ceabSEric Dumazet 100973a6bab5SEric Dumazet static void tcp_tsq_write(struct sock *sk) 10106f458dfbSEric Dumazet { 10116f458dfbSEric Dumazet if ((1 << sk->sk_state) & 10126f458dfbSEric Dumazet (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING | 1013f9616c35SEric Dumazet TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) { 1014f9616c35SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 1015f9616c35SEric Dumazet 1016f9616c35SEric Dumazet if (tp->lost_out > tp->retrans_out && 10173a91d29fSKoichiro Den tp->snd_cwnd > tcp_packets_in_flight(tp)) { 10183a91d29fSKoichiro Den tcp_mstamp_refresh(tp); 1019f9616c35SEric Dumazet tcp_xmit_retransmit_queue(sk); 10203a91d29fSKoichiro Den } 1021f9616c35SEric Dumazet 1022f9616c35SEric Dumazet tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle, 1023bf06200eSJohn Ogness 0, GFP_ATOMIC); 10246f458dfbSEric Dumazet } 1025f9616c35SEric Dumazet } 102673a6bab5SEric Dumazet 102773a6bab5SEric Dumazet static void tcp_tsq_handler(struct sock *sk) 102873a6bab5SEric Dumazet { 102973a6bab5SEric Dumazet bh_lock_sock(sk); 103073a6bab5SEric Dumazet if (!sock_owned_by_user(sk)) 103173a6bab5SEric Dumazet tcp_tsq_write(sk); 103273a6bab5SEric Dumazet else if (!test_and_set_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags)) 103373a6bab5SEric Dumazet sock_hold(sk); 103473a6bab5SEric Dumazet bh_unlock_sock(sk); 103573a6bab5SEric Dumazet } 103646d3ceabSEric Dumazet /* 10378e3bff96Sstephen hemminger * One tasklet per cpu tries to send more skbs. 103846d3ceabSEric Dumazet * We run in tasklet context but need to disable irqs when 10398e3bff96Sstephen hemminger * transferring tsq->head because tcp_wfree() might 104046d3ceabSEric Dumazet * interrupt us (non NAPI drivers) 104146d3ceabSEric Dumazet */ 1042c6533ca8SAllen Pais static void tcp_tasklet_func(struct tasklet_struct *t) 104346d3ceabSEric Dumazet { 1044c6533ca8SAllen Pais struct tsq_tasklet *tsq = from_tasklet(tsq, t, tasklet); 104546d3ceabSEric Dumazet LIST_HEAD(list); 104646d3ceabSEric Dumazet unsigned long flags; 104746d3ceabSEric Dumazet struct list_head *q, *n; 104846d3ceabSEric Dumazet struct tcp_sock *tp; 104946d3ceabSEric Dumazet struct sock *sk; 105046d3ceabSEric Dumazet 105146d3ceabSEric Dumazet local_irq_save(flags); 105246d3ceabSEric Dumazet list_splice_init(&tsq->head, &list); 105346d3ceabSEric Dumazet local_irq_restore(flags); 105446d3ceabSEric Dumazet 105546d3ceabSEric Dumazet list_for_each_safe(q, n, &list) { 105646d3ceabSEric Dumazet tp = list_entry(q, struct tcp_sock, tsq_node); 105746d3ceabSEric Dumazet list_del(&tp->tsq_node); 105846d3ceabSEric Dumazet 105946d3ceabSEric Dumazet sk = (struct sock *)tp; 10600a9648f1SEric Dumazet smp_mb__before_atomic(); 10617aa5470cSEric Dumazet clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags); 10627aa5470cSEric Dumazet 10636f458dfbSEric Dumazet tcp_tsq_handler(sk); 106446d3ceabSEric Dumazet sk_free(sk); 106546d3ceabSEric Dumazet } 106646d3ceabSEric Dumazet } 106746d3ceabSEric Dumazet 106840fc3423SEric Dumazet #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \ 106940fc3423SEric Dumazet TCPF_WRITE_TIMER_DEFERRED | \ 107040fc3423SEric Dumazet TCPF_DELACK_TIMER_DEFERRED | \ 107140fc3423SEric Dumazet TCPF_MTU_REDUCED_DEFERRED) 107246d3ceabSEric Dumazet /** 107346d3ceabSEric Dumazet * tcp_release_cb - tcp release_sock() callback 107446d3ceabSEric Dumazet * @sk: socket 107546d3ceabSEric Dumazet * 107646d3ceabSEric Dumazet * called from release_sock() to perform protocol dependent 107746d3ceabSEric Dumazet * actions before socket release. 107846d3ceabSEric Dumazet */ 107946d3ceabSEric Dumazet void tcp_release_cb(struct sock *sk) 108046d3ceabSEric Dumazet { 10816f458dfbSEric Dumazet unsigned long flags, nflags; 108246d3ceabSEric Dumazet 10836f458dfbSEric Dumazet /* perform an atomic operation only if at least one flag is set */ 10846f458dfbSEric Dumazet do { 10857aa5470cSEric Dumazet flags = sk->sk_tsq_flags; 10866f458dfbSEric Dumazet if (!(flags & TCP_DEFERRED_ALL)) 10876f458dfbSEric Dumazet return; 10886f458dfbSEric Dumazet nflags = flags & ~TCP_DEFERRED_ALL; 10897aa5470cSEric Dumazet } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags); 10906f458dfbSEric Dumazet 109173a6bab5SEric Dumazet if (flags & TCPF_TSQ_DEFERRED) { 109273a6bab5SEric Dumazet tcp_tsq_write(sk); 109373a6bab5SEric Dumazet __sock_put(sk); 109473a6bab5SEric Dumazet } 1095c3f9b018SEric Dumazet /* Here begins the tricky part : 1096c3f9b018SEric Dumazet * We are called from release_sock() with : 1097c3f9b018SEric Dumazet * 1) BH disabled 1098c3f9b018SEric Dumazet * 2) sk_lock.slock spinlock held 1099c3f9b018SEric Dumazet * 3) socket owned by us (sk->sk_lock.owned == 1) 1100c3f9b018SEric Dumazet * 1101c3f9b018SEric Dumazet * But following code is meant to be called from BH handlers, 1102c3f9b018SEric Dumazet * so we should keep BH disabled, but early release socket ownership 1103c3f9b018SEric Dumazet */ 1104c3f9b018SEric Dumazet sock_release_ownership(sk); 1105c3f9b018SEric Dumazet 110640fc3423SEric Dumazet if (flags & TCPF_WRITE_TIMER_DEFERRED) { 11076f458dfbSEric Dumazet tcp_write_timer_handler(sk); 1108144d56e9SEric Dumazet __sock_put(sk); 1109144d56e9SEric Dumazet } 111040fc3423SEric Dumazet if (flags & TCPF_DELACK_TIMER_DEFERRED) { 11116f458dfbSEric Dumazet tcp_delack_timer_handler(sk); 1112144d56e9SEric Dumazet __sock_put(sk); 1113144d56e9SEric Dumazet } 111440fc3423SEric Dumazet if (flags & TCPF_MTU_REDUCED_DEFERRED) { 11154fab9071SNeal Cardwell inet_csk(sk)->icsk_af_ops->mtu_reduced(sk); 1116144d56e9SEric Dumazet __sock_put(sk); 1117144d56e9SEric Dumazet } 111846d3ceabSEric Dumazet } 111946d3ceabSEric Dumazet EXPORT_SYMBOL(tcp_release_cb); 112046d3ceabSEric Dumazet 112146d3ceabSEric Dumazet void __init tcp_tasklet_init(void) 112246d3ceabSEric Dumazet { 112346d3ceabSEric Dumazet int i; 112446d3ceabSEric Dumazet 112546d3ceabSEric Dumazet for_each_possible_cpu(i) { 112646d3ceabSEric Dumazet struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i); 112746d3ceabSEric Dumazet 112846d3ceabSEric Dumazet INIT_LIST_HEAD(&tsq->head); 1129c6533ca8SAllen Pais tasklet_setup(&tsq->tasklet, tcp_tasklet_func); 113046d3ceabSEric Dumazet } 113146d3ceabSEric Dumazet } 113246d3ceabSEric Dumazet 113346d3ceabSEric Dumazet /* 113446d3ceabSEric Dumazet * Write buffer destructor automatically called from kfree_skb. 11358e3bff96Sstephen hemminger * We can't xmit new skbs from this context, as we might already 113646d3ceabSEric Dumazet * hold qdisc lock. 113746d3ceabSEric Dumazet */ 1138d6a4a104SEric Dumazet void tcp_wfree(struct sk_buff *skb) 113946d3ceabSEric Dumazet { 114046d3ceabSEric Dumazet struct sock *sk = skb->sk; 114146d3ceabSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 1142408f0a6cSEric Dumazet unsigned long flags, nval, oval; 11439b462d02SEric Dumazet 11449b462d02SEric Dumazet /* Keep one reference on sk_wmem_alloc. 11459b462d02SEric Dumazet * Will be released by sk_free() from here or tcp_tasklet_func() 11469b462d02SEric Dumazet */ 114714afee4bSReshetova, Elena WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc)); 11489b462d02SEric Dumazet 11499b462d02SEric Dumazet /* If this softirq is serviced by ksoftirqd, we are likely under stress. 11509b462d02SEric Dumazet * Wait until our queues (qdisc + devices) are drained. 11519b462d02SEric Dumazet * This gives : 11529b462d02SEric Dumazet * - less callbacks to tcp_write_xmit(), reducing stress (batches) 11539b462d02SEric Dumazet * - chance for incoming ACK (processed by another cpu maybe) 11549b462d02SEric Dumazet * to migrate this flow (skb->ooo_okay will be eventually set) 11559b462d02SEric Dumazet */ 115614afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current) 11579b462d02SEric Dumazet goto out; 115846d3ceabSEric Dumazet 11597aa5470cSEric Dumazet for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) { 116046d3ceabSEric Dumazet struct tsq_tasklet *tsq; 1161a9b204d1SEric Dumazet bool empty; 116246d3ceabSEric Dumazet 1163408f0a6cSEric Dumazet if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED)) 1164408f0a6cSEric Dumazet goto out; 1165408f0a6cSEric Dumazet 116673a6bab5SEric Dumazet nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED; 11677aa5470cSEric Dumazet nval = cmpxchg(&sk->sk_tsq_flags, oval, nval); 1168408f0a6cSEric Dumazet if (nval != oval) 1169408f0a6cSEric Dumazet continue; 1170408f0a6cSEric Dumazet 117146d3ceabSEric Dumazet /* queue this socket to tasklet queue */ 117246d3ceabSEric Dumazet local_irq_save(flags); 1173903ceff7SChristoph Lameter tsq = this_cpu_ptr(&tsq_tasklet); 1174a9b204d1SEric Dumazet empty = list_empty(&tsq->head); 117546d3ceabSEric Dumazet list_add(&tp->tsq_node, &tsq->head); 1176a9b204d1SEric Dumazet if (empty) 117746d3ceabSEric Dumazet tasklet_schedule(&tsq->tasklet); 117846d3ceabSEric Dumazet local_irq_restore(flags); 11799b462d02SEric Dumazet return; 118046d3ceabSEric Dumazet } 11819b462d02SEric Dumazet out: 11829b462d02SEric Dumazet sk_free(sk); 118346d3ceabSEric Dumazet } 118446d3ceabSEric Dumazet 118573a6bab5SEric Dumazet /* Note: Called under soft irq. 118673a6bab5SEric Dumazet * We can call TCP stack right away, unless socket is owned by user. 1187218af599SEric Dumazet */ 1188218af599SEric Dumazet enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer) 1189218af599SEric Dumazet { 1190218af599SEric Dumazet struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer); 1191218af599SEric Dumazet struct sock *sk = (struct sock *)tp; 1192218af599SEric Dumazet 119373a6bab5SEric Dumazet tcp_tsq_handler(sk); 119473a6bab5SEric Dumazet sock_put(sk); 1195218af599SEric Dumazet 1196218af599SEric Dumazet return HRTIMER_NORESTART; 1197218af599SEric Dumazet } 1198218af599SEric Dumazet 1199a7a25630SEric Dumazet static void tcp_update_skb_after_send(struct sock *sk, struct sk_buff *skb, 1200a7a25630SEric Dumazet u64 prior_wstamp) 1201e2080072SEric Dumazet { 1202ab408b6dSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 1203ab408b6dSEric Dumazet 1204ab408b6dSEric Dumazet if (sk->sk_pacing_status != SK_PACING_NONE) { 120576a9ebe8SEric Dumazet unsigned long rate = sk->sk_pacing_rate; 1206ab408b6dSEric Dumazet 1207ab408b6dSEric Dumazet /* Original sch_fq does not pace first 10 MSS 1208ab408b6dSEric Dumazet * Note that tp->data_segs_out overflows after 2^32 packets, 1209ab408b6dSEric Dumazet * this is a minor annoyance. 1210ab408b6dSEric Dumazet */ 121176a9ebe8SEric Dumazet if (rate != ~0UL && rate && tp->data_segs_out >= 10) { 1212a7a25630SEric Dumazet u64 len_ns = div64_ul((u64)skb->len * NSEC_PER_SEC, rate); 1213a7a25630SEric Dumazet u64 credit = tp->tcp_wstamp_ns - prior_wstamp; 1214a7a25630SEric Dumazet 1215a7a25630SEric Dumazet /* take into account OS jitter */ 1216a7a25630SEric Dumazet len_ns -= min_t(u64, len_ns / 2, credit); 1217a7a25630SEric Dumazet tp->tcp_wstamp_ns += len_ns; 1218ab408b6dSEric Dumazet } 1219ab408b6dSEric Dumazet } 1220e2080072SEric Dumazet list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue); 1221e2080072SEric Dumazet } 1222e2080072SEric Dumazet 122305e22e83SEric Dumazet INDIRECT_CALLABLE_DECLARE(int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl)); 122405e22e83SEric Dumazet INDIRECT_CALLABLE_DECLARE(int inet6_csk_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl)); 1225dd2e0b86SEric Dumazet INDIRECT_CALLABLE_DECLARE(void tcp_v4_send_check(struct sock *sk, struct sk_buff *skb)); 122605e22e83SEric Dumazet 12271da177e4SLinus Torvalds /* This routine actually transmits TCP packets queued in by 12281da177e4SLinus Torvalds * tcp_do_sendmsg(). This is used by both the initial 12291da177e4SLinus Torvalds * transmission and possible later retransmissions. 12301da177e4SLinus Torvalds * All SKB's seen here are completely headerless. It is our 12311da177e4SLinus Torvalds * job to build the TCP header, and pass the packet down to 12321da177e4SLinus Torvalds * IP so it can do the same plus pass the packet off to the 12331da177e4SLinus Torvalds * device. 12341da177e4SLinus Torvalds * 12351da177e4SLinus Torvalds * We are working here with either a clone of the original 12361da177e4SLinus Torvalds * SKB, or a fresh unique copy made by the retransmit engine. 12371da177e4SLinus Torvalds */ 12382987babbSYuchung Cheng static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, 12392987babbSYuchung Cheng int clone_it, gfp_t gfp_mask, u32 rcv_nxt) 12401da177e4SLinus Torvalds { 12416687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 1242dfb4b9dcSDavid S. Miller struct inet_sock *inet; 1243dfb4b9dcSDavid S. Miller struct tcp_sock *tp; 1244dfb4b9dcSDavid S. Miller struct tcp_skb_cb *tcb; 124533ad798cSAdam Langley struct tcp_out_options opts; 124695c96174SEric Dumazet unsigned int tcp_options_size, tcp_header_size; 12478c72c65bSEric Dumazet struct sk_buff *oskb = NULL; 1248cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 12491da177e4SLinus Torvalds struct tcphdr *th; 1250a7a25630SEric Dumazet u64 prior_wstamp; 12511da177e4SLinus Torvalds int err; 12521da177e4SLinus Torvalds 1253dfb4b9dcSDavid S. Miller BUG_ON(!skb || !tcp_skb_pcount(skb)); 12546f094b9eSLawrence Brakmo tp = tcp_sk(sk); 12557f12422cSYuchung Cheng prior_wstamp = tp->tcp_wstamp_ns; 12567f12422cSYuchung Cheng tp->tcp_wstamp_ns = max(tp->tcp_wstamp_ns, tp->tcp_clock_cache); 12577f12422cSYuchung Cheng skb->skb_mstamp_ns = tp->tcp_wstamp_ns; 1258ccdbb6e9SEric Dumazet if (clone_it) { 12596f094b9eSLawrence Brakmo TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq 12606f094b9eSLawrence Brakmo - tp->snd_una; 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 1360fa3fe2b1SFlorian Westphal tcp_options_write((__be32 *)(th + 1), 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) { 1365a465419bSEric Dumazet sk_nocaps_add(sk, NETIF_F_GSO_MASK); 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 */ 1556b617158dSEric Dumazet limit = sk->sk_sndbuf + 2 * SKB_TRUESIZE(GSO_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 15656cc55e09SOctavian Purdila if (skb_unclone(skb, gfp)) 15661da177e4SLinus Torvalds return -ENOMEM; 15671da177e4SLinus Torvalds 15681da177e4SLinus Torvalds /* Get a new skb... force flag on. */ 1569eb934478SEric Dumazet buff = sk_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 159598be9b12SEric Dumazet buff->ip_summed = CHECKSUM_PARTIAL; 15961da177e4SLinus Torvalds 1597a61bbcf2SPatrick McHardy buff->tstamp = skb->tstamp; 1598490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 15991da177e4SLinus Torvalds 16006475be16SDavid S. Miller old_factor = tcp_skb_pcount(skb); 16016475be16SDavid S. Miller 16021da177e4SLinus Torvalds /* Fix up tso_factor for both original and new SKB. */ 16035bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 16045bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 16051da177e4SLinus Torvalds 1606b9f64820SYuchung Cheng /* Update delivered info for the new segment */ 1607b9f64820SYuchung Cheng TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx; 1608b9f64820SYuchung Cheng 16096475be16SDavid S. Miller /* If this packet has been sent out already, we must 16106475be16SDavid S. Miller * adjust the various packet counters. 16116475be16SDavid S. Miller */ 1612cf0b450cSHerbert Xu if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 16136475be16SDavid S. Miller int diff = old_factor - tcp_skb_pcount(skb) - 16146475be16SDavid S. Miller tcp_skb_pcount(buff); 16151da177e4SLinus Torvalds 1616797108d1SIlpo Järvinen if (diff) 1617797108d1SIlpo Järvinen tcp_adjust_pcount(sk, skb, diff); 16181da177e4SLinus Torvalds } 16191da177e4SLinus Torvalds 16201da177e4SLinus Torvalds /* Link BUFF into the send queue. */ 1621f4a775d1SEric Dumazet __skb_header_release(buff); 162275c119afSEric Dumazet tcp_insert_write_queue_after(skb, buff, sk, tcp_queue); 1623f67971e6SEric Dumazet if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE) 1624e2080072SEric Dumazet list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor); 16251da177e4SLinus Torvalds 16261da177e4SLinus Torvalds return 0; 16271da177e4SLinus Torvalds } 16281da177e4SLinus Torvalds 1629f4d01666SEric Dumazet /* This is similar to __pskb_pull_tail(). The difference is that pulled 1630f4d01666SEric Dumazet * data is not copied, but immediately discarded. 16311da177e4SLinus Torvalds */ 16327162fb24SEric Dumazet static int __pskb_trim_head(struct sk_buff *skb, int len) 16331da177e4SLinus Torvalds { 16347b7fc97aSEric Dumazet struct skb_shared_info *shinfo; 16351da177e4SLinus Torvalds int i, k, eat; 16361da177e4SLinus Torvalds 16374fa48bf3SEric Dumazet eat = min_t(int, len, skb_headlen(skb)); 16384fa48bf3SEric Dumazet if (eat) { 16394fa48bf3SEric Dumazet __skb_pull(skb, eat); 16404fa48bf3SEric Dumazet len -= eat; 16414fa48bf3SEric Dumazet if (!len) 16427162fb24SEric Dumazet return 0; 16434fa48bf3SEric Dumazet } 16441da177e4SLinus Torvalds eat = len; 16451da177e4SLinus Torvalds k = 0; 16467b7fc97aSEric Dumazet shinfo = skb_shinfo(skb); 16477b7fc97aSEric Dumazet for (i = 0; i < shinfo->nr_frags; i++) { 16487b7fc97aSEric Dumazet int size = skb_frag_size(&shinfo->frags[i]); 16499e903e08SEric Dumazet 16509e903e08SEric Dumazet if (size <= eat) { 1651aff65da0SIan Campbell skb_frag_unref(skb, i); 16529e903e08SEric Dumazet eat -= size; 16531da177e4SLinus Torvalds } else { 16547b7fc97aSEric Dumazet shinfo->frags[k] = shinfo->frags[i]; 16551da177e4SLinus Torvalds if (eat) { 1656b54c9d5bSJonathan Lemon skb_frag_off_add(&shinfo->frags[k], eat); 16577b7fc97aSEric Dumazet skb_frag_size_sub(&shinfo->frags[k], eat); 16581da177e4SLinus Torvalds eat = 0; 16591da177e4SLinus Torvalds } 16601da177e4SLinus Torvalds k++; 16611da177e4SLinus Torvalds } 16621da177e4SLinus Torvalds } 16637b7fc97aSEric Dumazet shinfo->nr_frags = k; 16641da177e4SLinus Torvalds 16651da177e4SLinus Torvalds skb->data_len -= len; 16661da177e4SLinus Torvalds skb->len = skb->data_len; 16677162fb24SEric Dumazet return len; 16681da177e4SLinus Torvalds } 16691da177e4SLinus Torvalds 167067edfef7SAndi Kleen /* Remove acked data from a packet in the transmit queue. */ 16711da177e4SLinus Torvalds int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 16721da177e4SLinus Torvalds { 16737162fb24SEric Dumazet u32 delta_truesize; 16747162fb24SEric Dumazet 167514bbd6a5SPravin B Shelar if (skb_unclone(skb, GFP_ATOMIC)) 16761da177e4SLinus Torvalds return -ENOMEM; 16771da177e4SLinus Torvalds 16787162fb24SEric Dumazet delta_truesize = __pskb_trim_head(skb, len); 16791da177e4SLinus Torvalds 16801da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq += len; 168184fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 16821da177e4SLinus Torvalds 16837162fb24SEric Dumazet if (delta_truesize) { 16847162fb24SEric Dumazet skb->truesize -= delta_truesize; 1685ab4e846aSEric Dumazet sk_wmem_queued_add(sk, -delta_truesize); 16867162fb24SEric Dumazet sk_mem_uncharge(sk, delta_truesize); 16877162fb24SEric Dumazet } 16881da177e4SLinus Torvalds 16895b35e1e6SNeal Cardwell /* Any change of skb->len requires recalculation of tso factor. */ 16901da177e4SLinus Torvalds if (tcp_skb_pcount(skb) > 1) 16915bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb)); 16921da177e4SLinus Torvalds 16931da177e4SLinus Torvalds return 0; 16941da177e4SLinus Torvalds } 16951da177e4SLinus Torvalds 16961b63edd6SYuchung Cheng /* Calculate MSS not accounting any TCP options. */ 16971b63edd6SYuchung Cheng static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu) 16985d424d5aSJohn Heffner { 1699cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1700cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 17015d424d5aSJohn Heffner int mss_now; 17025d424d5aSJohn Heffner 17035d424d5aSJohn Heffner /* Calculate base mss without TCP options: 17045d424d5aSJohn Heffner It is MMS_S - sizeof(tcphdr) of rfc1122 17055d424d5aSJohn Heffner */ 17065d424d5aSJohn Heffner mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 17075d424d5aSJohn Heffner 170867469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 170967469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 171067469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 171167469601SEric Dumazet 171267469601SEric Dumazet if (dst && dst_allfrag(dst)) 171367469601SEric Dumazet mss_now -= icsk->icsk_af_ops->net_frag_header_len; 171467469601SEric Dumazet } 171567469601SEric Dumazet 17165d424d5aSJohn Heffner /* Clamp it (mss_clamp does not include tcp options) */ 17175d424d5aSJohn Heffner if (mss_now > tp->rx_opt.mss_clamp) 17185d424d5aSJohn Heffner mss_now = tp->rx_opt.mss_clamp; 17195d424d5aSJohn Heffner 17205d424d5aSJohn Heffner /* Now subtract optional transport overhead */ 17215d424d5aSJohn Heffner mss_now -= icsk->icsk_ext_hdr_len; 17225d424d5aSJohn Heffner 17235d424d5aSJohn Heffner /* Then reserve room for full set of TCP options and 8 bytes of data */ 17245f3e2bf0SEric Dumazet mss_now = max(mss_now, sock_net(sk)->ipv4.sysctl_tcp_min_snd_mss); 17255d424d5aSJohn Heffner return mss_now; 17265d424d5aSJohn Heffner } 17275d424d5aSJohn Heffner 17281b63edd6SYuchung Cheng /* Calculate MSS. Not accounting for SACKs here. */ 17291b63edd6SYuchung Cheng int tcp_mtu_to_mss(struct sock *sk, int pmtu) 17301b63edd6SYuchung Cheng { 17311b63edd6SYuchung Cheng /* Subtract TCP options size, not including SACKs */ 17321b63edd6SYuchung Cheng return __tcp_mtu_to_mss(sk, pmtu) - 17331b63edd6SYuchung Cheng (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr)); 17341b63edd6SYuchung Cheng } 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 1766b0f9ca53SFan Du icsk->icsk_mtup.enabled = 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; 1769b0f9ca53SFan Du icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, 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); 186786fd14adSWeiping Pan if (win_used < tp->snd_cwnd) { 186886fd14adSWeiping Pan tp->snd_ssthresh = tcp_current_ssthresh(sk); 186986fd14adSWeiping Pan tp->snd_cwnd = (tp->snd_cwnd + 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, 1958605ad7f1SEric Dumazet * to send one TSO packet per ms 1959605ad7f1SEric Dumazet */ 1960dcb8c9b4SEric Dumazet static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now, 19611b3878caSNeal Cardwell int min_tso_segs) 1962605ad7f1SEric Dumazet { 1963605ad7f1SEric Dumazet u32 bytes, segs; 1964605ad7f1SEric Dumazet 196576a9ebe8SEric Dumazet bytes = min_t(unsigned long, 19667c68fa2bSEric Dumazet sk->sk_pacing_rate >> READ_ONCE(sk->sk_pacing_shift), 1967605ad7f1SEric Dumazet sk->sk_gso_max_size - 1 - MAX_TCP_HEADER); 1968605ad7f1SEric Dumazet 1969605ad7f1SEric Dumazet /* Goal is to send at least one packet per ms, 1970605ad7f1SEric Dumazet * not one big TSO packet every 100 ms. 1971605ad7f1SEric Dumazet * This preserves ACK clocking and is consistent 1972605ad7f1SEric Dumazet * with tcp_tso_should_defer() heuristic. 1973605ad7f1SEric Dumazet */ 19741b3878caSNeal Cardwell segs = max_t(u32, bytes / mss_now, min_tso_segs); 1975605ad7f1SEric Dumazet 1976350c9f48SEric Dumazet return segs; 1977605ad7f1SEric Dumazet } 1978605ad7f1SEric Dumazet 1979ed6e7268SNeal Cardwell /* Return the number of segments we want in the skb we are transmitting. 1980ed6e7268SNeal Cardwell * See if congestion control module wants to decide; otherwise, autosize. 1981ed6e7268SNeal Cardwell */ 1982ed6e7268SNeal Cardwell static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now) 1983ed6e7268SNeal Cardwell { 1984ed6e7268SNeal Cardwell const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 1985dcb8c9b4SEric Dumazet u32 min_tso, tso_segs; 1986ed6e7268SNeal Cardwell 1987dcb8c9b4SEric Dumazet min_tso = ca_ops->min_tso_segs ? 1988dcb8c9b4SEric Dumazet ca_ops->min_tso_segs(sk) : 1989dcb8c9b4SEric Dumazet sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs; 1990dcb8c9b4SEric Dumazet 1991dcb8c9b4SEric Dumazet tso_segs = tcp_tso_autosize(sk, mss_now, min_tso); 1992350c9f48SEric Dumazet return min_t(u32, tso_segs, sk->sk_gso_max_segs); 1993ed6e7268SNeal Cardwell } 1994ed6e7268SNeal Cardwell 1995d4589926SEric Dumazet /* Returns the portion of skb which can be sent right away */ 1996d4589926SEric Dumazet static unsigned int tcp_mss_split_point(const struct sock *sk, 1997d4589926SEric Dumazet const struct sk_buff *skb, 1998d4589926SEric Dumazet unsigned int mss_now, 1999d4589926SEric Dumazet unsigned int max_segs, 2000d4589926SEric Dumazet int nonagle) 2001c1b4a7e6SDavid S. Miller { 2002cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 2003d4589926SEric Dumazet u32 partial, needed, window, max_len; 2004c1b4a7e6SDavid S. Miller 200590840defSIlpo Järvinen window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 20061485348dSBen Hutchings max_len = mss_now * max_segs; 20070e3a4803SIlpo Järvinen 20081485348dSBen Hutchings if (likely(max_len <= window && skb != tcp_write_queue_tail(sk))) 20091485348dSBen Hutchings return max_len; 20100e3a4803SIlpo Järvinen 20115ea3a748SIlpo Järvinen needed = min(skb->len, window); 20125ea3a748SIlpo Järvinen 20131485348dSBen Hutchings if (max_len <= needed) 20141485348dSBen Hutchings return max_len; 20150e3a4803SIlpo Järvinen 2016d4589926SEric Dumazet partial = needed % mss_now; 2017d4589926SEric Dumazet /* If last segment is not a full MSS, check if Nagle rules allow us 2018d4589926SEric Dumazet * to include this last segment in this skb. 2019d4589926SEric Dumazet * Otherwise, we'll split the skb at last MSS boundary 2020d4589926SEric Dumazet */ 2021cc93fc51SPeter Pan(潘卫平) if (tcp_nagle_check(partial != 0, tp, nonagle)) 2022d4589926SEric Dumazet return needed - partial; 2023d4589926SEric Dumazet 2024d4589926SEric Dumazet return needed; 2025c1b4a7e6SDavid S. Miller } 2026c1b4a7e6SDavid S. Miller 2027c1b4a7e6SDavid S. Miller /* Can at least one segment of SKB be sent right now, according to the 2028c1b4a7e6SDavid S. Miller * congestion window rules? If so, return how many segments are allowed. 2029c1b4a7e6SDavid S. Miller */ 2030cf533ea5SEric Dumazet static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, 2031cf533ea5SEric Dumazet const struct sk_buff *skb) 2032c1b4a7e6SDavid S. Miller { 2033d649a7a8SEric Dumazet u32 in_flight, cwnd, halfcwnd; 2034c1b4a7e6SDavid S. Miller 2035c1b4a7e6SDavid S. Miller /* Don't be strict about the congestion window for the final FIN. */ 20364de075e0SEric Dumazet if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && 20374de075e0SEric Dumazet tcp_skb_pcount(skb) == 1) 2038c1b4a7e6SDavid S. Miller return 1; 2039c1b4a7e6SDavid S. Miller 2040c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 2041c1b4a7e6SDavid S. Miller cwnd = tp->snd_cwnd; 2042d649a7a8SEric Dumazet if (in_flight >= cwnd) 2043c1b4a7e6SDavid S. Miller return 0; 2044d649a7a8SEric Dumazet 2045d649a7a8SEric Dumazet /* For better scheduling, ensure we have at least 2046d649a7a8SEric Dumazet * 2 GSO packets in flight. 2047d649a7a8SEric Dumazet */ 2048d649a7a8SEric Dumazet halfcwnd = max(cwnd >> 1, 1U); 2049d649a7a8SEric Dumazet return min(halfcwnd, cwnd - in_flight); 2050c1b4a7e6SDavid S. Miller } 2051c1b4a7e6SDavid S. Miller 2052b595076aSUwe Kleine-König /* Initialize TSO state of a skb. 205367edfef7SAndi Kleen * This must be invoked the first time we consider transmitting 2054c1b4a7e6SDavid S. Miller * SKB onto the wire. 2055c1b4a7e6SDavid S. Miller */ 20565bbb432cSEric Dumazet static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now) 2057c1b4a7e6SDavid S. Miller { 2058c1b4a7e6SDavid S. Miller int tso_segs = tcp_skb_pcount(skb); 2059c1b4a7e6SDavid S. Miller 2060f8269a49SIlpo Järvinen if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { 20615bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 2062c1b4a7e6SDavid S. Miller tso_segs = tcp_skb_pcount(skb); 2063c1b4a7e6SDavid S. Miller } 2064c1b4a7e6SDavid S. Miller return tso_segs; 2065c1b4a7e6SDavid S. Miller } 2066c1b4a7e6SDavid S. Miller 2067c1b4a7e6SDavid S. Miller 2068a2a385d6SEric Dumazet /* Return true if the Nagle test allows this packet to be 2069c1b4a7e6SDavid S. Miller * sent now. 2070c1b4a7e6SDavid S. Miller */ 2071a2a385d6SEric Dumazet static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, 2072c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 2073c1b4a7e6SDavid S. Miller { 2074c1b4a7e6SDavid S. Miller /* Nagle rule does not apply to frames, which sit in the middle of the 2075c1b4a7e6SDavid S. Miller * write_queue (they have no chances to get new data). 2076c1b4a7e6SDavid S. Miller * 2077c1b4a7e6SDavid S. Miller * This is implemented in the callers, where they modify the 'nonagle' 2078c1b4a7e6SDavid S. Miller * argument based upon the location of SKB in the send queue. 2079c1b4a7e6SDavid S. Miller */ 2080c1b4a7e6SDavid S. Miller if (nonagle & TCP_NAGLE_PUSH) 2081a2a385d6SEric Dumazet return true; 2082c1b4a7e6SDavid S. Miller 20839b44190dSYuchung Cheng /* Don't use the nagle rule for urgent data (or for the final FIN). */ 20849b44190dSYuchung Cheng if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) 2085a2a385d6SEric Dumazet return true; 2086c1b4a7e6SDavid S. Miller 2087cc93fc51SPeter Pan(潘卫平) if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle)) 2088a2a385d6SEric Dumazet return true; 2089c1b4a7e6SDavid S. Miller 2090a2a385d6SEric Dumazet return false; 2091c1b4a7e6SDavid S. Miller } 2092c1b4a7e6SDavid S. Miller 2093c1b4a7e6SDavid S. Miller /* Does at least the first segment of SKB fit into the send window? */ 2094a2a385d6SEric Dumazet static bool tcp_snd_wnd_test(const struct tcp_sock *tp, 2095a2a385d6SEric Dumazet const struct sk_buff *skb, 2096056834d9SIlpo Järvinen unsigned int cur_mss) 2097c1b4a7e6SDavid S. Miller { 2098c1b4a7e6SDavid S. Miller u32 end_seq = TCP_SKB_CB(skb)->end_seq; 2099c1b4a7e6SDavid S. Miller 2100c1b4a7e6SDavid S. Miller if (skb->len > cur_mss) 2101c1b4a7e6SDavid S. Miller end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 2102c1b4a7e6SDavid S. Miller 210390840defSIlpo Järvinen return !after(end_seq, tcp_wnd_end(tp)); 2104c1b4a7e6SDavid S. Miller } 2105c1b4a7e6SDavid S. Miller 2106c1b4a7e6SDavid S. Miller /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 2107c1b4a7e6SDavid S. Miller * which is put after SKB on the list. It is very much like 2108c1b4a7e6SDavid S. Miller * tcp_fragment() except that it may make several kinds of assumptions 2109c1b4a7e6SDavid S. Miller * in order to speed up the splitting operation. In particular, we 2110c1b4a7e6SDavid S. Miller * know that all the data is in scatter-gather pages, and that the 2111c1b4a7e6SDavid S. Miller * packet has never been sent out before (and thus is not cloned). 2112c1b4a7e6SDavid S. Miller */ 211356483341SEric Dumazet static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, 2114c4ead4c5SEric Dumazet unsigned int mss_now, gfp_t gfp) 2115c1b4a7e6SDavid S. Miller { 2116c1b4a7e6SDavid S. Miller int nlen = skb->len - len; 211756483341SEric Dumazet struct sk_buff *buff; 21189ce01461SIlpo Järvinen u8 flags; 2119c1b4a7e6SDavid S. Miller 2120c1b4a7e6SDavid S. Miller /* All of a TSO frame must be composed of paged data. */ 2121c8ac3774SHerbert Xu if (skb->len != skb->data_len) 212256483341SEric Dumazet return tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE, 212356483341SEric Dumazet skb, len, mss_now, gfp); 2124c1b4a7e6SDavid S. Miller 2125eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, gfp, true); 212651456b29SIan Morris if (unlikely(!buff)) 2127c1b4a7e6SDavid S. Miller return -ENOMEM; 212841477662SJakub Kicinski skb_copy_decrypted(buff, skb); 21295a369ca6SPaolo Abeni mptcp_skb_ext_copy(buff, skb); 2130c1b4a7e6SDavid S. Miller 2131ab4e846aSEric Dumazet sk_wmem_queued_add(sk, buff->truesize); 21323ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 2133b60b49eaSHerbert Xu buff->truesize += nlen; 2134c1b4a7e6SDavid S. Miller skb->truesize -= nlen; 2135c1b4a7e6SDavid S. Miller 2136c1b4a7e6SDavid S. Miller /* Correct the sequence numbers. */ 2137c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 2138c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 2139c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 2140c1b4a7e6SDavid S. Miller 2141c1b4a7e6SDavid S. Miller /* PSH and FIN should only be set in the second packet. */ 21424de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 21434de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 21444de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 2145c1b4a7e6SDavid S. Miller 2146c1b4a7e6SDavid S. Miller /* This packet was never sent out yet, so no SACK bits. */ 2147c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->sacked = 0; 2148c1b4a7e6SDavid S. Miller 2149a166140eSMartin KaFai Lau tcp_skb_fragment_eor(skb, buff); 2150a166140eSMartin KaFai Lau 215198be9b12SEric Dumazet buff->ip_summed = CHECKSUM_PARTIAL; 2152c1b4a7e6SDavid S. Miller skb_split(skb, buff, len); 2153490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 2154c1b4a7e6SDavid S. Miller 2155c1b4a7e6SDavid S. Miller /* Fix up tso_factor for both original and new SKB. */ 21565bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 21575bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 2158c1b4a7e6SDavid S. Miller 2159c1b4a7e6SDavid S. Miller /* Link BUFF into the send queue. */ 2160f4a775d1SEric Dumazet __skb_header_release(buff); 216156483341SEric Dumazet tcp_insert_write_queue_after(skb, buff, sk, TCP_FRAG_IN_WRITE_QUEUE); 2162c1b4a7e6SDavid S. Miller 2163c1b4a7e6SDavid S. Miller return 0; 2164c1b4a7e6SDavid S. Miller } 2165c1b4a7e6SDavid S. Miller 2166c1b4a7e6SDavid S. Miller /* Try to defer sending, if possible, in order to minimize the amount 2167c1b4a7e6SDavid S. Miller * of TSO splitting we do. View it as a kind of TSO Nagle test. 2168c1b4a7e6SDavid S. Miller * 2169c1b4a7e6SDavid S. Miller * This algorithm is from John Heffner. 2170c1b4a7e6SDavid S. Miller */ 2171ca8a2263SNeal Cardwell static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb, 2172f9bfe4e6SEric Dumazet bool *is_cwnd_limited, 2173f9bfe4e6SEric Dumazet bool *is_rwnd_limited, 2174f9bfe4e6SEric Dumazet u32 max_segs) 2175c1b4a7e6SDavid S. Miller { 21766687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 2177f1c6ea38SEric Dumazet u32 send_win, cong_win, limit, in_flight; 217850c8339eSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 217950c8339eSEric Dumazet struct sk_buff *head; 2180ad9f4f50SEric Dumazet int win_divisor; 2181f1c6ea38SEric Dumazet s64 delta; 2182c1b4a7e6SDavid S. Miller 218399d7662aSEric Dumazet if (icsk->icsk_ca_state >= TCP_CA_Recovery) 2184ae8064acSJohn Heffner goto send_now; 2185ae8064acSJohn Heffner 21865f852eb5SEric Dumazet /* Avoid bursty behavior by allowing defer 2187a682850aSEric Dumazet * only if the last write was recent (1 ms). 2188a682850aSEric Dumazet * Note that tp->tcp_wstamp_ns can be in the future if we have 2189a682850aSEric Dumazet * packets waiting in a qdisc or device for EDT delivery. 21905f852eb5SEric Dumazet */ 2191a682850aSEric Dumazet delta = tp->tcp_clock_cache - tp->tcp_wstamp_ns - NSEC_PER_MSEC; 2192a682850aSEric Dumazet if (delta > 0) 2193ae8064acSJohn Heffner goto send_now; 2194908a75c1SDavid S. Miller 2195c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 2196c1b4a7e6SDavid S. Miller 2197c8c9aeb5SStefano Brivio BUG_ON(tcp_skb_pcount(skb) <= 1); 2198c8c9aeb5SStefano Brivio BUG_ON(tp->snd_cwnd <= in_flight); 2199c1b4a7e6SDavid S. Miller 220090840defSIlpo Järvinen send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 2201c1b4a7e6SDavid S. Miller 2202c1b4a7e6SDavid S. Miller /* From in_flight test above, we know that cwnd > in_flight. */ 2203c1b4a7e6SDavid S. Miller cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; 2204c1b4a7e6SDavid S. Miller 2205c1b4a7e6SDavid S. Miller limit = min(send_win, cong_win); 2206c1b4a7e6SDavid S. Miller 2207ba244fe9SDavid S. Miller /* If a full-sized TSO skb can be sent, do it. */ 2208605ad7f1SEric Dumazet if (limit >= max_segs * tp->mss_cache) 2209ae8064acSJohn Heffner goto send_now; 2210ba244fe9SDavid S. Miller 221162ad2761SIlpo Järvinen /* Middle in queue won't get any more data, full sendable already? */ 221262ad2761SIlpo Järvinen if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) 221362ad2761SIlpo Järvinen goto send_now; 221462ad2761SIlpo Järvinen 22155bbcc0f5SLinus Torvalds win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor); 2216ad9f4f50SEric Dumazet if (win_divisor) { 2217c1b4a7e6SDavid S. Miller u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); 2218c1b4a7e6SDavid S. Miller 2219c1b4a7e6SDavid S. Miller /* If at least some fraction of a window is available, 2220c1b4a7e6SDavid S. Miller * just use it. 2221c1b4a7e6SDavid S. Miller */ 2222ad9f4f50SEric Dumazet chunk /= win_divisor; 2223c1b4a7e6SDavid S. Miller if (limit >= chunk) 2224ae8064acSJohn Heffner goto send_now; 2225c1b4a7e6SDavid S. Miller } else { 2226c1b4a7e6SDavid S. Miller /* Different approach, try not to defer past a single 2227c1b4a7e6SDavid S. Miller * ACK. Receiver should ACK every other full sized 2228c1b4a7e6SDavid S. Miller * frame, so if we have space for more than 3 frames 2229c1b4a7e6SDavid S. Miller * then send now. 2230c1b4a7e6SDavid S. Miller */ 22316b5a5c0dSNeal Cardwell if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) 2232ae8064acSJohn Heffner goto send_now; 2233c1b4a7e6SDavid S. Miller } 2234c1b4a7e6SDavid S. Miller 223575c119afSEric Dumazet /* TODO : use tsorted_sent_queue ? */ 223675c119afSEric Dumazet head = tcp_rtx_queue_head(sk); 223775c119afSEric Dumazet if (!head) 223875c119afSEric Dumazet goto send_now; 2239f1c6ea38SEric Dumazet delta = tp->tcp_clock_cache - head->tstamp; 224050c8339eSEric Dumazet /* If next ACK is likely to come too late (half srtt), do not defer */ 2241f1c6ea38SEric Dumazet if ((s64)(delta - (u64)NSEC_PER_USEC * (tp->srtt_us >> 4)) < 0) 224250c8339eSEric Dumazet goto send_now; 224350c8339eSEric Dumazet 2244f9bfe4e6SEric Dumazet /* Ok, it looks like it is advisable to defer. 2245f9bfe4e6SEric Dumazet * Three cases are tracked : 2246f9bfe4e6SEric Dumazet * 1) We are cwnd-limited 2247f9bfe4e6SEric Dumazet * 2) We are rwnd-limited 2248f9bfe4e6SEric Dumazet * 3) We are application limited. 2249f9bfe4e6SEric Dumazet */ 2250f9bfe4e6SEric Dumazet if (cong_win < send_win) { 2251f9bfe4e6SEric Dumazet if (cong_win <= skb->len) { 2252ca8a2263SNeal Cardwell *is_cwnd_limited = true; 2253f9bfe4e6SEric Dumazet return true; 2254f9bfe4e6SEric Dumazet } 2255f9bfe4e6SEric Dumazet } else { 2256f9bfe4e6SEric Dumazet if (send_win <= skb->len) { 2257f9bfe4e6SEric Dumazet *is_rwnd_limited = true; 2258f9bfe4e6SEric Dumazet return true; 2259f9bfe4e6SEric Dumazet } 2260f9bfe4e6SEric Dumazet } 2261f9bfe4e6SEric Dumazet 2262f9bfe4e6SEric Dumazet /* If this packet won't get more data, do not wait. */ 2263d8ed257fSEric Dumazet if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) || 2264d8ed257fSEric Dumazet TCP_SKB_CB(skb)->eor) 2265f9bfe4e6SEric Dumazet goto send_now; 2266ca8a2263SNeal Cardwell 2267a2a385d6SEric Dumazet return true; 2268ae8064acSJohn Heffner 2269ae8064acSJohn Heffner send_now: 2270a2a385d6SEric Dumazet return false; 2271c1b4a7e6SDavid S. Miller } 2272c1b4a7e6SDavid S. Miller 227305cbc0dbSFan Du static inline void tcp_mtu_check_reprobe(struct sock *sk) 227405cbc0dbSFan Du { 227505cbc0dbSFan Du struct inet_connection_sock *icsk = inet_csk(sk); 227605cbc0dbSFan Du struct tcp_sock *tp = tcp_sk(sk); 227705cbc0dbSFan Du struct net *net = sock_net(sk); 227805cbc0dbSFan Du u32 interval; 227905cbc0dbSFan Du s32 delta; 228005cbc0dbSFan Du 228105cbc0dbSFan Du interval = net->ipv4.sysctl_tcp_probe_interval; 2282c74df29aSEric Dumazet delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp; 228305cbc0dbSFan Du if (unlikely(delta >= interval * HZ)) { 228405cbc0dbSFan Du int mss = tcp_current_mss(sk); 228505cbc0dbSFan Du 228605cbc0dbSFan Du /* Update current search range */ 228705cbc0dbSFan Du icsk->icsk_mtup.probe_size = 0; 228805cbc0dbSFan Du icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + 228905cbc0dbSFan Du sizeof(struct tcphdr) + 229005cbc0dbSFan Du icsk->icsk_af_ops->net_header_len; 229105cbc0dbSFan Du icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss); 229205cbc0dbSFan Du 229305cbc0dbSFan Du /* Update probe time stamp */ 2294c74df29aSEric Dumazet icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; 229505cbc0dbSFan Du } 229605cbc0dbSFan Du } 229705cbc0dbSFan Du 2298808cf9e3SIlya Lesokhin static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len) 2299808cf9e3SIlya Lesokhin { 2300808cf9e3SIlya Lesokhin struct sk_buff *skb, *next; 2301808cf9e3SIlya Lesokhin 2302808cf9e3SIlya Lesokhin skb = tcp_send_head(sk); 2303808cf9e3SIlya Lesokhin tcp_for_write_queue_from_safe(skb, next, sk) { 2304808cf9e3SIlya Lesokhin if (len <= skb->len) 2305808cf9e3SIlya Lesokhin break; 2306808cf9e3SIlya Lesokhin 2307888a5c53SWillem de Bruijn if (unlikely(TCP_SKB_CB(skb)->eor) || tcp_has_tx_tstamp(skb)) 2308808cf9e3SIlya Lesokhin return false; 2309808cf9e3SIlya Lesokhin 2310808cf9e3SIlya Lesokhin len -= skb->len; 2311808cf9e3SIlya Lesokhin } 2312808cf9e3SIlya Lesokhin 2313808cf9e3SIlya Lesokhin return true; 2314808cf9e3SIlya Lesokhin } 2315808cf9e3SIlya Lesokhin 23165d424d5aSJohn Heffner /* Create a new MTU probe if we are ready. 231767edfef7SAndi Kleen * MTU probe is regularly attempting to increase the path MTU by 231867edfef7SAndi Kleen * deliberately sending larger packets. This discovers routing 231967edfef7SAndi Kleen * changes resulting in larger path MTUs. 232067edfef7SAndi Kleen * 23215d424d5aSJohn Heffner * Returns 0 if we should wait to probe (no cwnd available), 23225d424d5aSJohn Heffner * 1 if a probe was sent, 2323056834d9SIlpo Järvinen * -1 otherwise 2324056834d9SIlpo Järvinen */ 23255d424d5aSJohn Heffner static int tcp_mtu_probe(struct sock *sk) 23265d424d5aSJohn Heffner { 23275d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 232812a59abcSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 23295d424d5aSJohn Heffner struct sk_buff *skb, *nskb, *next; 23306b58e0a5SFan Du struct net *net = sock_net(sk); 23315d424d5aSJohn Heffner int probe_size; 233291cc17c0SIlpo Järvinen int size_needed; 233312a59abcSEric Dumazet int copy, len; 23345d424d5aSJohn Heffner int mss_now; 23356b58e0a5SFan Du int interval; 23365d424d5aSJohn Heffner 23375d424d5aSJohn Heffner /* Not currently probing/verifying, 23385d424d5aSJohn Heffner * not in recovery, 23395d424d5aSJohn Heffner * have enough cwnd, and 234012a59abcSEric Dumazet * not SACKing (the variable headers throw things off) 234112a59abcSEric Dumazet */ 234212a59abcSEric Dumazet if (likely(!icsk->icsk_mtup.enabled || 23435d424d5aSJohn Heffner icsk->icsk_mtup.probe_size || 23445d424d5aSJohn Heffner inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 23455d424d5aSJohn Heffner tp->snd_cwnd < 11 || 234612a59abcSEric Dumazet tp->rx_opt.num_sacks || tp->rx_opt.dsack)) 23475d424d5aSJohn Heffner return -1; 23485d424d5aSJohn Heffner 23496b58e0a5SFan Du /* Use binary search for probe_size between tcp_mss_base, 23506b58e0a5SFan Du * and current mss_clamp. if (search_high - search_low) 23516b58e0a5SFan Du * smaller than a threshold, backoff from probing. 23526b58e0a5SFan Du */ 23530c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 23546b58e0a5SFan Du probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high + 23556b58e0a5SFan Du icsk->icsk_mtup.search_low) >> 1); 235691cc17c0SIlpo Järvinen size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; 23576b58e0a5SFan Du interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low; 235805cbc0dbSFan Du /* When misfortune happens, we are reprobing actively, 235905cbc0dbSFan Du * and then reprobe timer has expired. We stick with current 236005cbc0dbSFan Du * probing process by not resetting search range to its orignal. 236105cbc0dbSFan Du */ 23626b58e0a5SFan Du if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) || 236305cbc0dbSFan Du interval < net->ipv4.sysctl_tcp_probe_threshold) { 236405cbc0dbSFan Du /* Check whether enough time has elaplased for 236505cbc0dbSFan Du * another round of probing. 236605cbc0dbSFan Du */ 236705cbc0dbSFan Du tcp_mtu_check_reprobe(sk); 23685d424d5aSJohn Heffner return -1; 23695d424d5aSJohn Heffner } 23705d424d5aSJohn Heffner 23715d424d5aSJohn Heffner /* Have enough data in the send queue to probe? */ 23727f9c33e5SIlpo Järvinen if (tp->write_seq - tp->snd_nxt < size_needed) 23735d424d5aSJohn Heffner return -1; 23745d424d5aSJohn Heffner 237591cc17c0SIlpo Järvinen if (tp->snd_wnd < size_needed) 23765d424d5aSJohn Heffner return -1; 237790840defSIlpo Järvinen if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) 23785d424d5aSJohn Heffner return 0; 23795d424d5aSJohn Heffner 2380d67c58e9SIlpo Järvinen /* Do we need to wait to drain cwnd? With none in flight, don't stall */ 2381d67c58e9SIlpo Järvinen if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { 2382d67c58e9SIlpo Järvinen if (!tcp_packets_in_flight(tp)) 23835d424d5aSJohn Heffner return -1; 23845d424d5aSJohn Heffner else 23855d424d5aSJohn Heffner return 0; 23865d424d5aSJohn Heffner } 23875d424d5aSJohn Heffner 2388808cf9e3SIlya Lesokhin if (!tcp_can_coalesce_send_queue_head(sk, probe_size)) 2389808cf9e3SIlya Lesokhin return -1; 2390808cf9e3SIlya Lesokhin 23915d424d5aSJohn Heffner /* We're allowed to probe. Build it now. */ 2392eb934478SEric Dumazet nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false); 239351456b29SIan Morris if (!nskb) 23945d424d5aSJohn Heffner return -1; 2395ab4e846aSEric Dumazet sk_wmem_queued_add(sk, nskb->truesize); 23963ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 23975d424d5aSJohn Heffner 2398fe067e8aSDavid S. Miller skb = tcp_send_head(sk); 239941477662SJakub Kicinski skb_copy_decrypted(nskb, skb); 24005a369ca6SPaolo Abeni mptcp_skb_ext_copy(nskb, skb); 24015d424d5aSJohn Heffner 24025d424d5aSJohn Heffner TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 24035d424d5aSJohn Heffner TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 24044de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; 24055d424d5aSJohn Heffner TCP_SKB_CB(nskb)->sacked = 0; 24065d424d5aSJohn Heffner nskb->csum = 0; 240798be9b12SEric Dumazet nskb->ip_summed = CHECKSUM_PARTIAL; 24085d424d5aSJohn Heffner 240950c4817eSIlpo Järvinen tcp_insert_write_queue_before(nskb, skb, sk); 24102b7cda9cSEric Dumazet tcp_highest_sack_replace(sk, skb, nskb); 241150c4817eSIlpo Järvinen 24125d424d5aSJohn Heffner len = 0; 2413234b6860SIlpo Järvinen tcp_for_write_queue_from_safe(skb, next, sk) { 24145d424d5aSJohn Heffner copy = min_t(int, skb->len, probe_size - len); 24155d424d5aSJohn Heffner skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); 24165d424d5aSJohn Heffner 24175d424d5aSJohn Heffner if (skb->len <= copy) { 24185d424d5aSJohn Heffner /* We've eaten all the data from this skb. 24195d424d5aSJohn Heffner * Throw it away. */ 24204de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 2421808cf9e3SIlya Lesokhin /* If this is the last SKB we copy and eor is set 2422808cf9e3SIlya Lesokhin * we need to propagate it to the new skb. 2423808cf9e3SIlya Lesokhin */ 2424808cf9e3SIlya Lesokhin TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor; 2425888a5c53SWillem de Bruijn tcp_skb_collapse_tstamp(nskb, skb); 2426fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 24273ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 24285d424d5aSJohn Heffner } else { 24294de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & 2430a3433f35SChangli Gao ~(TCPHDR_FIN|TCPHDR_PSH); 24315d424d5aSJohn Heffner if (!skb_shinfo(skb)->nr_frags) { 24325d424d5aSJohn Heffner skb_pull(skb, copy); 24335d424d5aSJohn Heffner } else { 24345d424d5aSJohn Heffner __pskb_trim_head(skb, copy); 24355bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 24365d424d5aSJohn Heffner } 24375d424d5aSJohn Heffner TCP_SKB_CB(skb)->seq += copy; 24385d424d5aSJohn Heffner } 24395d424d5aSJohn Heffner 24405d424d5aSJohn Heffner len += copy; 2441234b6860SIlpo Järvinen 2442234b6860SIlpo Järvinen if (len >= probe_size) 2443234b6860SIlpo Järvinen break; 24445d424d5aSJohn Heffner } 24455bbb432cSEric Dumazet tcp_init_tso_segs(nskb, nskb->len); 24465d424d5aSJohn Heffner 24475d424d5aSJohn Heffner /* We're ready to send. If this fails, the probe will 24487faee5c0SEric Dumazet * be resegmented into mss-sized pieces by tcp_write_xmit(). 24497faee5c0SEric Dumazet */ 24505d424d5aSJohn Heffner if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 24515d424d5aSJohn Heffner /* Decrement cwnd here because we are sending 24525d424d5aSJohn Heffner * effectively two packets. */ 24535d424d5aSJohn Heffner tp->snd_cwnd--; 245466f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, nskb); 24555d424d5aSJohn Heffner 24565d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 24570e7b1368SJohn Heffner tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 24580e7b1368SJohn Heffner tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 24595d424d5aSJohn Heffner 24605d424d5aSJohn Heffner return 1; 24615d424d5aSJohn Heffner } 24625d424d5aSJohn Heffner 24635d424d5aSJohn Heffner return -1; 24645d424d5aSJohn Heffner } 24655d424d5aSJohn Heffner 2466864e5c09SEric Dumazet static bool tcp_pacing_check(struct sock *sk) 2467218af599SEric Dumazet { 2468864e5c09SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 2469864e5c09SEric Dumazet 2470864e5c09SEric Dumazet if (!tcp_needs_internal_pacing(sk)) 2471864e5c09SEric Dumazet return false; 2472864e5c09SEric Dumazet 2473864e5c09SEric Dumazet if (tp->tcp_wstamp_ns <= tp->tcp_clock_cache) 2474864e5c09SEric Dumazet return false; 2475864e5c09SEric Dumazet 2476864e5c09SEric Dumazet if (!hrtimer_is_queued(&tp->pacing_timer)) { 2477864e5c09SEric Dumazet hrtimer_start(&tp->pacing_timer, 2478864e5c09SEric Dumazet ns_to_ktime(tp->tcp_wstamp_ns), 2479864e5c09SEric Dumazet HRTIMER_MODE_ABS_PINNED_SOFT); 2480864e5c09SEric Dumazet sock_hold(sk); 2481864e5c09SEric Dumazet } 2482864e5c09SEric Dumazet return true; 2483218af599SEric Dumazet } 2484218af599SEric Dumazet 2485f9616c35SEric Dumazet /* TCP Small Queues : 2486f9616c35SEric Dumazet * Control number of packets in qdisc/devices to two packets / or ~1 ms. 2487f9616c35SEric Dumazet * (These limits are doubled for retransmits) 2488f9616c35SEric Dumazet * This allows for : 2489f9616c35SEric Dumazet * - better RTT estimation and ACK scheduling 2490f9616c35SEric Dumazet * - faster recovery 2491f9616c35SEric Dumazet * - high rates 2492f9616c35SEric Dumazet * Alas, some drivers / subsystems require a fair amount 2493f9616c35SEric Dumazet * of queued bytes to ensure line rate. 2494f9616c35SEric Dumazet * One example is wifi aggregation (802.11 AMPDU) 2495f9616c35SEric Dumazet */ 2496f9616c35SEric Dumazet static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb, 2497f9616c35SEric Dumazet unsigned int factor) 2498f9616c35SEric Dumazet { 249976a9ebe8SEric Dumazet unsigned long limit; 2500f9616c35SEric Dumazet 250176a9ebe8SEric Dumazet limit = max_t(unsigned long, 250276a9ebe8SEric Dumazet 2 * skb->truesize, 25037c68fa2bSEric Dumazet sk->sk_pacing_rate >> READ_ONCE(sk->sk_pacing_shift)); 2504c73e5807SEric Dumazet if (sk->sk_pacing_status == SK_PACING_NONE) 250576a9ebe8SEric Dumazet limit = min_t(unsigned long, limit, 25069184d8bbSEric Dumazet sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes); 2507f9616c35SEric Dumazet limit <<= factor; 2508f9616c35SEric Dumazet 2509a842fe14SEric Dumazet if (static_branch_unlikely(&tcp_tx_delay_enabled) && 2510a842fe14SEric Dumazet tcp_sk(sk)->tcp_tx_delay) { 2511a842fe14SEric Dumazet u64 extra_bytes = (u64)sk->sk_pacing_rate * tcp_sk(sk)->tcp_tx_delay; 2512a842fe14SEric Dumazet 2513a842fe14SEric Dumazet /* TSQ is based on skb truesize sum (sk_wmem_alloc), so we 2514a842fe14SEric Dumazet * approximate our needs assuming an ~100% skb->truesize overhead. 2515a842fe14SEric Dumazet * USEC_PER_SEC is approximated by 2^20. 2516a842fe14SEric Dumazet * do_div(extra_bytes, USEC_PER_SEC/2) is replaced by a right shift. 2517a842fe14SEric Dumazet */ 2518a842fe14SEric Dumazet extra_bytes >>= (20 - 1); 2519a842fe14SEric Dumazet limit += extra_bytes; 2520a842fe14SEric Dumazet } 252114afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) > limit) { 252275c119afSEric Dumazet /* Always send skb if rtx queue is empty. 252375eefc6cSEric Dumazet * No need to wait for TX completion to call us back, 252475eefc6cSEric Dumazet * after softirq/tasklet schedule. 252575eefc6cSEric Dumazet * This helps when TX completions are delayed too much. 252675eefc6cSEric Dumazet */ 252775c119afSEric Dumazet if (tcp_rtx_queue_empty(sk)) 252875eefc6cSEric Dumazet return false; 252975eefc6cSEric Dumazet 25307aa5470cSEric Dumazet set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags); 2531f9616c35SEric Dumazet /* It is possible TX completion already happened 2532f9616c35SEric Dumazet * before we set TSQ_THROTTLED, so we must 2533f9616c35SEric Dumazet * test again the condition. 2534f9616c35SEric Dumazet */ 2535f9616c35SEric Dumazet smp_mb__after_atomic(); 253614afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) > limit) 2537f9616c35SEric Dumazet return true; 2538f9616c35SEric Dumazet } 2539f9616c35SEric Dumazet return false; 2540f9616c35SEric Dumazet } 2541f9616c35SEric Dumazet 254205b055e8SFrancis Yan static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new) 254305b055e8SFrancis Yan { 2544628174ccSEric Dumazet const u32 now = tcp_jiffies32; 2545efe967cdSArnd Bergmann enum tcp_chrono old = tp->chrono_type; 254605b055e8SFrancis Yan 2547efe967cdSArnd Bergmann if (old > TCP_CHRONO_UNSPEC) 2548efe967cdSArnd Bergmann tp->chrono_stat[old - 1] += now - tp->chrono_start; 254905b055e8SFrancis Yan tp->chrono_start = now; 255005b055e8SFrancis Yan tp->chrono_type = new; 255105b055e8SFrancis Yan } 255205b055e8SFrancis Yan 255305b055e8SFrancis Yan void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type) 255405b055e8SFrancis Yan { 255505b055e8SFrancis Yan struct tcp_sock *tp = tcp_sk(sk); 255605b055e8SFrancis Yan 255705b055e8SFrancis Yan /* If there are multiple conditions worthy of tracking in a 25580f87230dSFrancis Yan * chronograph then the highest priority enum takes precedence 25590f87230dSFrancis Yan * over the other conditions. So that if something "more interesting" 256005b055e8SFrancis Yan * starts happening, stop the previous chrono and start a new one. 256105b055e8SFrancis Yan */ 256205b055e8SFrancis Yan if (type > tp->chrono_type) 256305b055e8SFrancis Yan tcp_chrono_set(tp, type); 256405b055e8SFrancis Yan } 256505b055e8SFrancis Yan 256605b055e8SFrancis Yan void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type) 256705b055e8SFrancis Yan { 256805b055e8SFrancis Yan struct tcp_sock *tp = tcp_sk(sk); 256905b055e8SFrancis Yan 25700f87230dSFrancis Yan 25710f87230dSFrancis Yan /* There are multiple conditions worthy of tracking in a 25720f87230dSFrancis Yan * chronograph, so that the highest priority enum takes 25730f87230dSFrancis Yan * precedence over the other conditions (see tcp_chrono_start). 25740f87230dSFrancis Yan * If a condition stops, we only stop chrono tracking if 25750f87230dSFrancis Yan * it's the "most interesting" or current chrono we are 25760f87230dSFrancis Yan * tracking and starts busy chrono if we have pending data. 25770f87230dSFrancis Yan */ 257875c119afSEric Dumazet if (tcp_rtx_and_write_queues_empty(sk)) 257905b055e8SFrancis Yan tcp_chrono_set(tp, TCP_CHRONO_UNSPEC); 25800f87230dSFrancis Yan else if (type == tp->chrono_type) 25810f87230dSFrancis Yan tcp_chrono_set(tp, TCP_CHRONO_BUSY); 258205b055e8SFrancis Yan } 258305b055e8SFrancis Yan 25841da177e4SLinus Torvalds /* This routine writes packets to the network. It advances the 25851da177e4SLinus Torvalds * send_head. This happens as incoming acks open up the remote 25861da177e4SLinus Torvalds * window for us. 25871da177e4SLinus Torvalds * 2588f8269a49SIlpo Järvinen * LARGESEND note: !tcp_urg_mode is overkill, only frames between 2589f8269a49SIlpo Järvinen * snd_up-64k-mss .. snd_up cannot be large. However, taking into 2590f8269a49SIlpo Järvinen * account rare use of URG, this is not a big flaw. 2591f8269a49SIlpo Järvinen * 25926ba8a3b1SNandita Dukkipati * Send at most one packet when push_one > 0. Temporarily ignore 25936ba8a3b1SNandita Dukkipati * cwnd limit to force at most one packet out when push_one == 2. 25946ba8a3b1SNandita Dukkipati 2595a2a385d6SEric Dumazet * Returns true, if no segments are in flight and we have queued segments, 2596a2a385d6SEric Dumazet * but cannot send anything now because of SWS or another problem. 25971da177e4SLinus Torvalds */ 2598a2a385d6SEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 2599d5dd9175SIlpo Järvinen int push_one, gfp_t gfp) 26001da177e4SLinus Torvalds { 26011da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 260292df7b51SDavid S. Miller struct sk_buff *skb; 2603c1b4a7e6SDavid S. Miller unsigned int tso_segs, sent_pkts; 2604c1b4a7e6SDavid S. Miller int cwnd_quota; 26055d424d5aSJohn Heffner int result; 26065615f886SFrancis Yan bool is_cwnd_limited = false, is_rwnd_limited = false; 2607605ad7f1SEric Dumazet u32 max_segs; 26081da177e4SLinus Torvalds 2609c1b4a7e6SDavid S. Miller sent_pkts = 0; 26105d424d5aSJohn Heffner 2611ee1836aeSEric Dumazet tcp_mstamp_refresh(tp); 2612d5dd9175SIlpo Järvinen if (!push_one) { 26135d424d5aSJohn Heffner /* Do MTU probing. */ 2614d5dd9175SIlpo Järvinen result = tcp_mtu_probe(sk); 2615d5dd9175SIlpo Järvinen if (!result) { 2616a2a385d6SEric Dumazet return false; 26175d424d5aSJohn Heffner } else if (result > 0) { 26185d424d5aSJohn Heffner sent_pkts = 1; 26195d424d5aSJohn Heffner } 2620d5dd9175SIlpo Järvinen } 26215d424d5aSJohn Heffner 2622ed6e7268SNeal Cardwell max_segs = tcp_tso_segs(sk, mss_now); 2623fe067e8aSDavid S. Miller while ((skb = tcp_send_head(sk))) { 2624c8ac3774SHerbert Xu unsigned int limit; 2625c8ac3774SHerbert Xu 262679861919SEric Dumazet if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) { 262779861919SEric Dumazet /* "skb_mstamp_ns" is used as a start point for the retransmit timer */ 262879861919SEric Dumazet skb->skb_mstamp_ns = tp->tcp_wstamp_ns = tp->tcp_clock_cache; 262979861919SEric Dumazet list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue); 2630bf50b606SEric Dumazet tcp_init_tso_segs(skb, mss_now); 263179861919SEric Dumazet goto repair; /* Skip network transmission */ 263279861919SEric Dumazet } 263379861919SEric Dumazet 2634218af599SEric Dumazet if (tcp_pacing_check(sk)) 2635218af599SEric Dumazet break; 2636218af599SEric Dumazet 26375bbb432cSEric Dumazet tso_segs = tcp_init_tso_segs(skb, mss_now); 2638c1b4a7e6SDavid S. Miller BUG_ON(!tso_segs); 2639c1b4a7e6SDavid S. Miller 2640b68e9f85SHerbert Xu cwnd_quota = tcp_cwnd_test(tp, skb); 26416ba8a3b1SNandita Dukkipati if (!cwnd_quota) { 26426ba8a3b1SNandita Dukkipati if (push_one == 2) 26436ba8a3b1SNandita Dukkipati /* Force out a loss probe pkt. */ 26446ba8a3b1SNandita Dukkipati cwnd_quota = 1; 26456ba8a3b1SNandita Dukkipati else 2646b68e9f85SHerbert Xu break; 26476ba8a3b1SNandita Dukkipati } 2648b68e9f85SHerbert Xu 26495615f886SFrancis Yan if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) { 26505615f886SFrancis Yan is_rwnd_limited = true; 2651b68e9f85SHerbert Xu break; 26525615f886SFrancis Yan } 2653b68e9f85SHerbert Xu 2654d6a4e26aSEric Dumazet if (tso_segs == 1) { 2655aa93466bSDavid S. Miller if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 2656aa93466bSDavid S. Miller (tcp_skb_is_last(sk, skb) ? 2657aa93466bSDavid S. Miller nonagle : TCP_NAGLE_PUSH)))) 2658aa93466bSDavid S. Miller break; 2659c1b4a7e6SDavid S. Miller } else { 2660ca8a2263SNeal Cardwell if (!push_one && 2661605ad7f1SEric Dumazet tcp_tso_should_defer(sk, skb, &is_cwnd_limited, 2662f9bfe4e6SEric Dumazet &is_rwnd_limited, max_segs)) 2663aa93466bSDavid S. Miller break; 2664c1b4a7e6SDavid S. Miller } 2665aa93466bSDavid S. Miller 2666605ad7f1SEric Dumazet limit = mss_now; 2667d6a4e26aSEric Dumazet if (tso_segs > 1 && !tcp_urg_mode(tp)) 2668605ad7f1SEric Dumazet limit = tcp_mss_split_point(sk, skb, mss_now, 2669605ad7f1SEric Dumazet min_t(unsigned int, 2670605ad7f1SEric Dumazet cwnd_quota, 2671605ad7f1SEric Dumazet max_segs), 2672605ad7f1SEric Dumazet nonagle); 2673605ad7f1SEric Dumazet 2674605ad7f1SEric Dumazet if (skb->len > limit && 267556483341SEric Dumazet unlikely(tso_fragment(sk, skb, limit, mss_now, gfp))) 2676605ad7f1SEric Dumazet break; 2677605ad7f1SEric Dumazet 2678f9616c35SEric Dumazet if (tcp_small_queue_check(sk, skb, 0)) 267946d3ceabSEric Dumazet break; 2680c9eeec26SEric Dumazet 26811f85e626SEric Dumazet /* Argh, we hit an empty skb(), presumably a thread 26821f85e626SEric Dumazet * is sleeping in sendmsg()/sk_stream_wait_memory(). 26831f85e626SEric Dumazet * We do not want to send a pure-ack packet and have 26841f85e626SEric Dumazet * a strange looking rtx queue with empty packet(s). 26851f85e626SEric Dumazet */ 26861f85e626SEric Dumazet if (TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq) 26871f85e626SEric Dumazet break; 26881f85e626SEric Dumazet 2689d5dd9175SIlpo Järvinen if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) 26901da177e4SLinus Torvalds break; 26911da177e4SLinus Torvalds 2692ec342325SAndrew Vagin repair: 26931da177e4SLinus Torvalds /* Advance the send_head. This one is sent out. 26941da177e4SLinus Torvalds * This call will increment packets_out. 26951da177e4SLinus Torvalds */ 269666f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 26971da177e4SLinus Torvalds 26981da177e4SLinus Torvalds tcp_minshall_update(tp, mss_now, skb); 2699a262f0cdSNandita Dukkipati sent_pkts += tcp_skb_pcount(skb); 2700d5dd9175SIlpo Järvinen 2701d5dd9175SIlpo Järvinen if (push_one) 2702d5dd9175SIlpo Järvinen break; 27031da177e4SLinus Torvalds } 27041da177e4SLinus Torvalds 27055615f886SFrancis Yan if (is_rwnd_limited) 27065615f886SFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED); 27075615f886SFrancis Yan else 27085615f886SFrancis Yan tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED); 27095615f886SFrancis Yan 2710299bcb55SNeal Cardwell is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd); 2711299bcb55SNeal Cardwell if (likely(sent_pkts || is_cwnd_limited)) 2712299bcb55SNeal Cardwell tcp_cwnd_validate(sk, is_cwnd_limited); 2713299bcb55SNeal Cardwell 2714aa93466bSDavid S. Miller if (likely(sent_pkts)) { 2715684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 2716684bad11SYuchung Cheng tp->prr_out += sent_pkts; 27176ba8a3b1SNandita Dukkipati 27186ba8a3b1SNandita Dukkipati /* Send one loss probe per tail loss episode. */ 27196ba8a3b1SNandita Dukkipati if (push_one != 2) 2720ed66dfafSNeal Cardwell tcp_schedule_loss_probe(sk, false); 2721a2a385d6SEric Dumazet return false; 27221da177e4SLinus Torvalds } 272375c119afSEric Dumazet return !tp->packets_out && !tcp_write_queue_empty(sk); 27246ba8a3b1SNandita Dukkipati } 27256ba8a3b1SNandita Dukkipati 2726ed66dfafSNeal Cardwell bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto) 27276ba8a3b1SNandita Dukkipati { 27286ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 27296ba8a3b1SNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 2730a2815817SNeal Cardwell u32 timeout, rto_delta_us; 27312ae21cf5SEric Dumazet int early_retrans; 27326ba8a3b1SNandita Dukkipati 27336ba8a3b1SNandita Dukkipati /* Don't do any loss probe on a Fast Open connection before 3WHS 27346ba8a3b1SNandita Dukkipati * finishes. 27356ba8a3b1SNandita Dukkipati */ 2736d983ea6fSEric Dumazet if (rcu_access_pointer(tp->fastopen_rsk)) 27376ba8a3b1SNandita Dukkipati return false; 27386ba8a3b1SNandita Dukkipati 27392ae21cf5SEric Dumazet early_retrans = sock_net(sk)->ipv4.sysctl_tcp_early_retrans; 27406ba8a3b1SNandita Dukkipati /* Schedule a loss probe in 2*RTT for SACK capable connections 2741b4f70c3dSNeal Cardwell * not in loss recovery, that are either limited by cwnd or application. 27426ba8a3b1SNandita Dukkipati */ 27432ae21cf5SEric Dumazet if ((early_retrans != 3 && early_retrans != 4) || 2744bec41a11SYuchung Cheng !tp->packets_out || !tcp_is_sack(tp) || 2745b4f70c3dSNeal Cardwell (icsk->icsk_ca_state != TCP_CA_Open && 2746b4f70c3dSNeal Cardwell icsk->icsk_ca_state != TCP_CA_CWR)) 27476ba8a3b1SNandita Dukkipati return false; 27486ba8a3b1SNandita Dukkipati 2749bb4d991aSYuchung Cheng /* Probe timeout is 2*rtt. Add minimum RTO to account 2750f9b99582SYuchung Cheng * for delayed ack when there's one outstanding packet. If no RTT 2751f9b99582SYuchung Cheng * sample is available then probe after TCP_TIMEOUT_INIT. 27526ba8a3b1SNandita Dukkipati */ 2753bb4d991aSYuchung Cheng if (tp->srtt_us) { 2754bb4d991aSYuchung Cheng timeout = usecs_to_jiffies(tp->srtt_us >> 2); 27556ba8a3b1SNandita Dukkipati if (tp->packets_out == 1) 2756bb4d991aSYuchung Cheng timeout += TCP_RTO_MIN; 2757bb4d991aSYuchung Cheng else 2758bb4d991aSYuchung Cheng timeout += TCP_TIMEOUT_MIN; 2759bb4d991aSYuchung Cheng } else { 2760bb4d991aSYuchung Cheng timeout = TCP_TIMEOUT_INIT; 2761bb4d991aSYuchung Cheng } 27626ba8a3b1SNandita Dukkipati 2763a2815817SNeal Cardwell /* If the RTO formula yields an earlier time, then use that time. */ 2764ed66dfafSNeal Cardwell rto_delta_us = advancing_rto ? 2765ed66dfafSNeal Cardwell jiffies_to_usecs(inet_csk(sk)->icsk_rto) : 2766ed66dfafSNeal Cardwell tcp_rto_delta_us(sk); /* How far in future is RTO? */ 2767a2815817SNeal Cardwell if (rto_delta_us > 0) 2768a2815817SNeal Cardwell timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us)); 27696ba8a3b1SNandita Dukkipati 27708dc242adSEric Dumazet tcp_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, TCP_RTO_MAX); 27716ba8a3b1SNandita Dukkipati return true; 27726ba8a3b1SNandita Dukkipati } 27736ba8a3b1SNandita Dukkipati 27741f3279aeSEric Dumazet /* Thanks to skb fast clones, we can detect if a prior transmit of 27751f3279aeSEric Dumazet * a packet is still in a qdisc or driver queue. 27761f3279aeSEric Dumazet * In this case, there is very little point doing a retransmit ! 27771f3279aeSEric Dumazet */ 2778*f4dae54eSEric Dumazet static bool skb_still_in_host_queue(struct sock *sk, 27791f3279aeSEric Dumazet const struct sk_buff *skb) 27801f3279aeSEric Dumazet { 278139bb5e62SEric Dumazet if (unlikely(skb_fclone_busy(sk, skb))) { 2782*f4dae54eSEric Dumazet set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags); 2783*f4dae54eSEric Dumazet smp_mb__after_atomic(); 2784*f4dae54eSEric Dumazet if (skb_fclone_busy(sk, skb)) { 2785c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), 27861f3279aeSEric Dumazet LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES); 27871f3279aeSEric Dumazet return true; 27881f3279aeSEric Dumazet } 2789*f4dae54eSEric Dumazet } 27901f3279aeSEric Dumazet return false; 27911f3279aeSEric Dumazet } 27921f3279aeSEric Dumazet 2793b340b264SYuchung Cheng /* When probe timeout (PTO) fires, try send a new segment if possible, else 27946ba8a3b1SNandita Dukkipati * retransmit the last segment. 27956ba8a3b1SNandita Dukkipati */ 27966ba8a3b1SNandita Dukkipati void tcp_send_loss_probe(struct sock *sk) 27976ba8a3b1SNandita Dukkipati { 27989b717a8dSNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 27996ba8a3b1SNandita Dukkipati struct sk_buff *skb; 28006ba8a3b1SNandita Dukkipati int pcount; 28016ba8a3b1SNandita Dukkipati int mss = tcp_current_mss(sk); 28026ba8a3b1SNandita Dukkipati 280376be93fcSYuchung Cheng /* At most one outstanding TLP */ 280476be93fcSYuchung Cheng if (tp->tlp_high_seq) 280576be93fcSYuchung Cheng goto rearm_timer; 280676be93fcSYuchung Cheng 280776be93fcSYuchung Cheng tp->tlp_retrans = 0; 2808b340b264SYuchung Cheng skb = tcp_send_head(sk); 280975c119afSEric Dumazet if (skb && tcp_snd_wnd_test(tp, skb, mss)) { 2810b340b264SYuchung Cheng pcount = tp->packets_out; 2811b340b264SYuchung Cheng tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC); 2812b340b264SYuchung Cheng if (tp->packets_out > pcount) 2813b340b264SYuchung Cheng goto probe_sent; 28146ba8a3b1SNandita Dukkipati goto rearm_timer; 28156ba8a3b1SNandita Dukkipati } 281675c119afSEric Dumazet skb = skb_rb_last(&sk->tcp_rtx_queue); 2817b2b7af86SYuchung Cheng if (unlikely(!skb)) { 2818b2b7af86SYuchung Cheng WARN_ONCE(tp->packets_out, 2819b2b7af86SYuchung Cheng "invalid inflight: %u state %u cwnd %u mss %d\n", 2820b2b7af86SYuchung Cheng tp->packets_out, sk->sk_state, tp->snd_cwnd, mss); 2821b2b7af86SYuchung Cheng inet_csk(sk)->icsk_pending = 0; 2822b2b7af86SYuchung Cheng return; 2823b2b7af86SYuchung Cheng } 28246ba8a3b1SNandita Dukkipati 28251f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 28261f3279aeSEric Dumazet goto rearm_timer; 28271f3279aeSEric Dumazet 28286ba8a3b1SNandita Dukkipati pcount = tcp_skb_pcount(skb); 28296ba8a3b1SNandita Dukkipati if (WARN_ON(!pcount)) 28306ba8a3b1SNandita Dukkipati goto rearm_timer; 28316ba8a3b1SNandita Dukkipati 28326ba8a3b1SNandita Dukkipati if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) { 283375c119afSEric Dumazet if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, 283475c119afSEric Dumazet (pcount - 1) * mss, mss, 28356cc55e09SOctavian Purdila GFP_ATOMIC))) 28366ba8a3b1SNandita Dukkipati goto rearm_timer; 283775c119afSEric Dumazet skb = skb_rb_next(skb); 28386ba8a3b1SNandita Dukkipati } 28396ba8a3b1SNandita Dukkipati 28406ba8a3b1SNandita Dukkipati if (WARN_ON(!skb || !tcp_skb_pcount(skb))) 28416ba8a3b1SNandita Dukkipati goto rearm_timer; 28426ba8a3b1SNandita Dukkipati 284310d3be56SEric Dumazet if (__tcp_retransmit_skb(sk, skb, 1)) 2844b340b264SYuchung Cheng goto rearm_timer; 28456ba8a3b1SNandita Dukkipati 284676be93fcSYuchung Cheng tp->tlp_retrans = 1; 284776be93fcSYuchung Cheng 284876be93fcSYuchung Cheng probe_sent: 28499b717a8dSNandita Dukkipati /* Record snd_nxt for loss detection. */ 28509b717a8dSNandita Dukkipati tp->tlp_high_seq = tp->snd_nxt; 28519b717a8dSNandita Dukkipati 2852c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES); 2853fcd16c0aSYuchung Cheng /* Reset s.t. tcp_rearm_rto will restart timer from now */ 2854fcd16c0aSYuchung Cheng inet_csk(sk)->icsk_pending = 0; 2855b340b264SYuchung Cheng rearm_timer: 2856fcd16c0aSYuchung Cheng tcp_rearm_rto(sk); 28571da177e4SLinus Torvalds } 28581da177e4SLinus Torvalds 2859a762a980SDavid S. Miller /* Push out any pending frames which were held back due to 2860a762a980SDavid S. Miller * TCP_CORK or attempt at coalescing tiny packets. 2861a762a980SDavid S. Miller * The socket must be locked by the caller. 2862a762a980SDavid S. Miller */ 28639e412ba7SIlpo Järvinen void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 28649e412ba7SIlpo Järvinen int nonagle) 2865a762a980SDavid S. Miller { 2866726e07a8SIlpo Järvinen /* If we are closed, the bytes will have to remain here. 2867726e07a8SIlpo Järvinen * In time closedown will finish, we empty the write queue and 2868726e07a8SIlpo Järvinen * all will be happy. 2869726e07a8SIlpo Järvinen */ 2870726e07a8SIlpo Järvinen if (unlikely(sk->sk_state == TCP_CLOSE)) 2871726e07a8SIlpo Järvinen return; 2872726e07a8SIlpo Järvinen 287399a1dec7SMel Gorman if (tcp_write_xmit(sk, cur_mss, nonagle, 0, 28747450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC))) 28759e412ba7SIlpo Järvinen tcp_check_probe_timer(sk); 2876a762a980SDavid S. Miller } 2877a762a980SDavid S. Miller 2878c1b4a7e6SDavid S. Miller /* Send _single_ skb sitting at the send head. This function requires 2879c1b4a7e6SDavid S. Miller * true push pending frames to setup probe timer etc. 2880c1b4a7e6SDavid S. Miller */ 2881c1b4a7e6SDavid S. Miller void tcp_push_one(struct sock *sk, unsigned int mss_now) 2882c1b4a7e6SDavid S. Miller { 2883fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 2884c1b4a7e6SDavid S. Miller 2885c1b4a7e6SDavid S. Miller BUG_ON(!skb || skb->len < mss_now); 2886c1b4a7e6SDavid S. Miller 2887d5dd9175SIlpo Järvinen tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); 2888c1b4a7e6SDavid S. Miller } 2889c1b4a7e6SDavid S. Miller 28901da177e4SLinus Torvalds /* This function returns the amount that we can raise the 28911da177e4SLinus Torvalds * usable window based on the following constraints 28921da177e4SLinus Torvalds * 28931da177e4SLinus Torvalds * 1. The window can never be shrunk once it is offered (RFC 793) 28941da177e4SLinus Torvalds * 2. We limit memory per socket 28951da177e4SLinus Torvalds * 28961da177e4SLinus Torvalds * RFC 1122: 28971da177e4SLinus Torvalds * "the suggested [SWS] avoidance algorithm for the receiver is to keep 28981da177e4SLinus Torvalds * RECV.NEXT + RCV.WIN fixed until: 28991da177e4SLinus Torvalds * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 29001da177e4SLinus Torvalds * 29011da177e4SLinus Torvalds * i.e. don't raise the right edge of the window until you can raise 29021da177e4SLinus Torvalds * it at least MSS bytes. 29031da177e4SLinus Torvalds * 29041da177e4SLinus Torvalds * Unfortunately, the recommended algorithm breaks header prediction, 29051da177e4SLinus Torvalds * since header prediction assumes th->window stays fixed. 29061da177e4SLinus Torvalds * 29071da177e4SLinus Torvalds * Strictly speaking, keeping th->window fixed violates the receiver 29081da177e4SLinus Torvalds * side SWS prevention criteria. The problem is that under this rule 29091da177e4SLinus Torvalds * a stream of single byte packets will cause the right side of the 29101da177e4SLinus Torvalds * window to always advance by a single byte. 29111da177e4SLinus Torvalds * 29121da177e4SLinus Torvalds * Of course, if the sender implements sender side SWS prevention 29131da177e4SLinus Torvalds * then this will not be a problem. 29141da177e4SLinus Torvalds * 29151da177e4SLinus Torvalds * BSD seems to make the following compromise: 29161da177e4SLinus Torvalds * 29171da177e4SLinus Torvalds * If the free space is less than the 1/4 of the maximum 29181da177e4SLinus Torvalds * space available and the free space is less than 1/2 mss, 29191da177e4SLinus Torvalds * then set the window to 0. 29201da177e4SLinus Torvalds * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 29211da177e4SLinus Torvalds * Otherwise, just prevent the window from shrinking 29221da177e4SLinus Torvalds * and from being larger than the largest representable value. 29231da177e4SLinus Torvalds * 29241da177e4SLinus Torvalds * This prevents incremental opening of the window in the regime 29251da177e4SLinus Torvalds * where TCP is limited by the speed of the reader side taking 29261da177e4SLinus Torvalds * data out of the TCP receive queue. It does nothing about 29271da177e4SLinus Torvalds * those cases where the window is constrained on the sender side 29281da177e4SLinus Torvalds * because the pipeline is full. 29291da177e4SLinus Torvalds * 29301da177e4SLinus Torvalds * BSD also seems to "accidentally" limit itself to windows that are a 29311da177e4SLinus Torvalds * multiple of MSS, at least until the free space gets quite small. 29321da177e4SLinus Torvalds * This would appear to be a side effect of the mbuf implementation. 29331da177e4SLinus Torvalds * Combining these two algorithms results in the observed behavior 29341da177e4SLinus Torvalds * of having a fixed window size at almost all times. 29351da177e4SLinus Torvalds * 29361da177e4SLinus Torvalds * Below we obtain similar behavior by forcing the offered window to 29371da177e4SLinus Torvalds * a multiple of the mss when it is feasible to do so. 29381da177e4SLinus Torvalds * 29391da177e4SLinus Torvalds * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 29401da177e4SLinus Torvalds * Regular options like TIMESTAMP are taken into account. 29411da177e4SLinus Torvalds */ 29421da177e4SLinus Torvalds u32 __tcp_select_window(struct sock *sk) 29431da177e4SLinus Torvalds { 2944463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 29451da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2946caa20d9aSStephen Hemminger /* MSS for the peer's data. Previous versions used mss_clamp 29471da177e4SLinus Torvalds * here. I don't know if the value based on our guesses 29481da177e4SLinus Torvalds * of peer's MSS is better for the performance. It's more correct 29491da177e4SLinus Torvalds * but may be worse for the performance because of rcv_mss 29501da177e4SLinus Torvalds * fluctuations. --SAW 1998/11/1 29511da177e4SLinus Torvalds */ 2952463c84b9SArnaldo Carvalho de Melo int mss = icsk->icsk_ack.rcv_mss; 29531da177e4SLinus Torvalds int free_space = tcp_space(sk); 295486c1a045SFlorian Westphal int allowed_space = tcp_full_space(sk); 2955071c8ed6SFlorian Westphal int full_space, window; 2956071c8ed6SFlorian Westphal 2957071c8ed6SFlorian Westphal if (sk_is_mptcp(sk)) 2958071c8ed6SFlorian Westphal mptcp_space(sk, &free_space, &allowed_space); 2959071c8ed6SFlorian Westphal 2960071c8ed6SFlorian Westphal full_space = min_t(int, tp->window_clamp, allowed_space); 29611da177e4SLinus Torvalds 296206425c30SEric Dumazet if (unlikely(mss > full_space)) { 29631da177e4SLinus Torvalds mss = full_space; 296406425c30SEric Dumazet if (mss <= 0) 296506425c30SEric Dumazet return 0; 296606425c30SEric Dumazet } 2967b92edbe0SEric Dumazet if (free_space < (full_space >> 1)) { 2968463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.quick = 0; 29691da177e4SLinus Torvalds 2970b8da51ebSEric Dumazet if (tcp_under_memory_pressure(sk)) 2971056834d9SIlpo Järvinen tp->rcv_ssthresh = min(tp->rcv_ssthresh, 2972056834d9SIlpo Järvinen 4U * tp->advmss); 29731da177e4SLinus Torvalds 297486c1a045SFlorian Westphal /* free_space might become our new window, make sure we don't 297586c1a045SFlorian Westphal * increase it due to wscale. 297686c1a045SFlorian Westphal */ 297786c1a045SFlorian Westphal free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale); 297886c1a045SFlorian Westphal 297986c1a045SFlorian Westphal /* if free space is less than mss estimate, or is below 1/16th 298086c1a045SFlorian Westphal * of the maximum allowed, try to move to zero-window, else 298186c1a045SFlorian Westphal * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and 298286c1a045SFlorian Westphal * new incoming data is dropped due to memory limits. 298386c1a045SFlorian Westphal * With large window, mss test triggers way too late in order 298486c1a045SFlorian Westphal * to announce zero window in time before rmem limit kicks in. 298586c1a045SFlorian Westphal */ 298686c1a045SFlorian Westphal if (free_space < (allowed_space >> 4) || free_space < mss) 29871da177e4SLinus Torvalds return 0; 29881da177e4SLinus Torvalds } 29891da177e4SLinus Torvalds 29901da177e4SLinus Torvalds if (free_space > tp->rcv_ssthresh) 29911da177e4SLinus Torvalds free_space = tp->rcv_ssthresh; 29921da177e4SLinus Torvalds 29931da177e4SLinus Torvalds /* Don't do rounding if we are using window scaling, since the 29941da177e4SLinus Torvalds * scaled window will not line up with the MSS boundary anyway. 29951da177e4SLinus Torvalds */ 29961da177e4SLinus Torvalds if (tp->rx_opt.rcv_wscale) { 29971da177e4SLinus Torvalds window = free_space; 29981da177e4SLinus Torvalds 29991da177e4SLinus Torvalds /* Advertise enough space so that it won't get scaled away. 30001da177e4SLinus Torvalds * Import case: prevent zero window announcement if 30011da177e4SLinus Torvalds * 1<<rcv_wscale > mss. 30021da177e4SLinus Torvalds */ 30031935299dSGao Feng window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale)); 30041da177e4SLinus Torvalds } else { 30051935299dSGao Feng window = tp->rcv_wnd; 30061da177e4SLinus Torvalds /* Get the largest window that is a nice multiple of mss. 30071da177e4SLinus Torvalds * Window clamp already applied above. 30081da177e4SLinus Torvalds * If our current window offering is within 1 mss of the 30091da177e4SLinus Torvalds * free space we just keep it. This prevents the divide 30101da177e4SLinus Torvalds * and multiply from happening most of the time. 30111da177e4SLinus Torvalds * We also don't do any window rounding when the free space 30121da177e4SLinus Torvalds * is too small. 30131da177e4SLinus Torvalds */ 30141da177e4SLinus Torvalds if (window <= free_space - mss || window > free_space) 30151935299dSGao Feng window = rounddown(free_space, mss); 301684565070SJohn Heffner else if (mss == full_space && 3017b92edbe0SEric Dumazet free_space > window + (full_space >> 1)) 301884565070SJohn Heffner window = free_space; 30191da177e4SLinus Torvalds } 30201da177e4SLinus Torvalds 30211da177e4SLinus Torvalds return window; 30221da177e4SLinus Torvalds } 30231da177e4SLinus Torvalds 3024cfea5a68SMartin KaFai Lau void tcp_skb_collapse_tstamp(struct sk_buff *skb, 3025082ac2d5SMartin KaFai Lau const struct sk_buff *next_skb) 3026082ac2d5SMartin KaFai Lau { 30270a2cf20cSSoheil Hassas Yeganeh if (unlikely(tcp_has_tx_tstamp(next_skb))) { 30280a2cf20cSSoheil Hassas Yeganeh const struct skb_shared_info *next_shinfo = 30290a2cf20cSSoheil Hassas Yeganeh skb_shinfo(next_skb); 3030082ac2d5SMartin KaFai Lau struct skb_shared_info *shinfo = skb_shinfo(skb); 3031082ac2d5SMartin KaFai Lau 30320a2cf20cSSoheil Hassas Yeganeh shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP; 3033082ac2d5SMartin KaFai Lau shinfo->tskey = next_shinfo->tskey; 30342de8023eSMartin KaFai Lau TCP_SKB_CB(skb)->txstamp_ack |= 30352de8023eSMartin KaFai Lau TCP_SKB_CB(next_skb)->txstamp_ack; 3036082ac2d5SMartin KaFai Lau } 3037082ac2d5SMartin KaFai Lau } 3038082ac2d5SMartin KaFai Lau 30394a17fc3aSIlpo Järvinen /* Collapses two adjacent SKB's during retransmission. */ 3040f8071cdeSEric Dumazet static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) 30411da177e4SLinus Torvalds { 30421da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 304375c119afSEric Dumazet struct sk_buff *next_skb = skb_rb_next(skb); 304413dde04fSWei Yongjun int next_skb_size; 30451da177e4SLinus Torvalds 3046058dc334SIlpo Järvinen next_skb_size = next_skb->len; 30471da177e4SLinus Torvalds 3048058dc334SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); 30491da177e4SLinus Torvalds 3050f8071cdeSEric Dumazet if (next_skb_size) { 3051f8071cdeSEric Dumazet if (next_skb_size <= skb_availroom(skb)) 3052f8071cdeSEric Dumazet skb_copy_bits(next_skb, 0, skb_put(skb, next_skb_size), 3053f8071cdeSEric Dumazet next_skb_size); 30543b4929f6SEric Dumazet else if (!tcp_skb_shift(skb, next_skb, 1, next_skb_size)) 3055f8071cdeSEric Dumazet return false; 3056f8071cdeSEric 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 31541da177e4SLinus Torvalds /* Do not sent more than we queued. 1/4 is reserved for possible 3155caa20d9aSStephen Hemminger * copying overhead: fragmentation, tunneling, mangling etc. 31561da177e4SLinus Torvalds */ 315714afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) > 3158ffb4d6c8SEric Dumazet min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), 3159ffb4d6c8SEric Dumazet sk->sk_sndbuf)) 31601da177e4SLinus Torvalds return -EAGAIN; 31611da177e4SLinus Torvalds 31621f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 31631f3279aeSEric Dumazet return -EBUSY; 31641f3279aeSEric Dumazet 31651da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 31667f582b24SEric Dumazet if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) { 31677f582b24SEric Dumazet WARN_ON_ONCE(1); 31687f582b24SEric Dumazet return -EINVAL; 31697f582b24SEric Dumazet } 31701da177e4SLinus Torvalds if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 31711da177e4SLinus Torvalds return -ENOMEM; 31721da177e4SLinus Torvalds } 31731da177e4SLinus Torvalds 31747d227cd2SSridhar Samudrala if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 31757d227cd2SSridhar Samudrala return -EHOSTUNREACH; /* Routing failure or similar. */ 31767d227cd2SSridhar Samudrala 31770c54b85fSIlpo Järvinen cur_mss = tcp_current_mss(sk); 31787d227cd2SSridhar Samudrala 31791da177e4SLinus Torvalds /* If receiver has shrunk his window, and skb is out of 31801da177e4SLinus Torvalds * new window, do not retransmit it. The exception is the 31811da177e4SLinus Torvalds * case, when window is shrunk to zero. In this case 31821da177e4SLinus Torvalds * our retransmit serves as a zero window probe. 31831da177e4SLinus Torvalds */ 31849d4fb27dSJoe Perches if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && 31859d4fb27dSJoe Perches TCP_SKB_CB(skb)->seq != tp->snd_una) 31861da177e4SLinus Torvalds return -EAGAIN; 31871da177e4SLinus Torvalds 318810d3be56SEric Dumazet len = cur_mss * segs; 318910d3be56SEric Dumazet if (skb->len > len) { 319075c119afSEric Dumazet if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len, 319175c119afSEric Dumazet cur_mss, GFP_ATOMIC)) 31921da177e4SLinus Torvalds return -ENOMEM; /* We'll try again later. */ 319302276f3cSIlpo Järvinen } else { 3194c52e2421SEric Dumazet if (skb_unclone(skb, GFP_ATOMIC)) 3195c52e2421SEric Dumazet return -ENOMEM; 319610d3be56SEric Dumazet 319710d3be56SEric Dumazet diff = tcp_skb_pcount(skb); 319810d3be56SEric Dumazet tcp_set_skb_tso_segs(skb, cur_mss); 319910d3be56SEric Dumazet diff -= tcp_skb_pcount(skb); 320010d3be56SEric Dumazet if (diff) 320110d3be56SEric Dumazet tcp_adjust_pcount(sk, skb, diff); 320210d3be56SEric Dumazet if (skb->len < cur_mss) 320310d3be56SEric Dumazet tcp_retrans_try_collapse(sk, skb, cur_mss); 32041da177e4SLinus Torvalds } 32051da177e4SLinus Torvalds 320649213555SDaniel Borkmann /* RFC3168, section 6.1.1.1. ECN fallback */ 320749213555SDaniel Borkmann if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN) 320849213555SDaniel Borkmann tcp_ecn_clear_syn(sk, skb); 320949213555SDaniel Borkmann 3210678550c6SYuchung Cheng /* Update global and local TCP statistics. */ 3211678550c6SYuchung Cheng segs = tcp_skb_pcount(skb); 3212678550c6SYuchung Cheng TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs); 3213678550c6SYuchung Cheng if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 3214678550c6SYuchung Cheng __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 3215678550c6SYuchung Cheng tp->total_retrans += segs; 3216fb31c9b9SWei Wang tp->bytes_retrans += skb->len; 3217678550c6SYuchung Cheng 321850bceae9SThomas Graf /* make sure skb->data is aligned on arches that require it 321950bceae9SThomas Graf * and check if ack-trimming & collapsing extended the headroom 322050bceae9SThomas Graf * beyond what csum_start can cover. 322150bceae9SThomas Graf */ 322250bceae9SThomas Graf if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) || 322350bceae9SThomas Graf skb_headroom(skb) >= 0xFFFF)) { 322410a81980SEric Dumazet struct sk_buff *nskb; 322510a81980SEric Dumazet 3226e2080072SEric Dumazet tcp_skb_tsorted_save(skb) { 322710a81980SEric Dumazet nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC); 322807f8e4d0SFlorian Westphal if (nskb) { 322907f8e4d0SFlorian Westphal nskb->dev = NULL; 323007f8e4d0SFlorian Westphal err = tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC); 323107f8e4d0SFlorian Westphal } else { 323207f8e4d0SFlorian Westphal err = -ENOBUFS; 323307f8e4d0SFlorian Westphal } 3234e2080072SEric Dumazet } tcp_skb_tsorted_restore(skb); 3235e2080072SEric Dumazet 32365889e2c0SYousuk Seung if (!err) { 3237a7a25630SEric Dumazet tcp_update_skb_after_send(sk, skb, tp->tcp_wstamp_ns); 32385889e2c0SYousuk Seung tcp_rate_skb_sent(sk, skb); 32395889e2c0SYousuk Seung } 3240117632e6SEric Dumazet } else { 3241c84a5711SYuchung Cheng err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 3242117632e6SEric Dumazet } 3243c84a5711SYuchung Cheng 32447f12422cSYuchung Cheng /* To avoid taking spuriously low RTT samples based on a timestamp 32457f12422cSYuchung Cheng * for a transmit that never happened, always mark EVER_RETRANS 32467f12422cSYuchung Cheng */ 32477f12422cSYuchung Cheng TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS; 32487f12422cSYuchung Cheng 3249a31ad29eSLawrence Brakmo if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RETRANS_CB_FLAG)) 3250a31ad29eSLawrence Brakmo tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RETRANS_CB, 3251a31ad29eSLawrence Brakmo TCP_SKB_CB(skb)->seq, segs, err); 3252a31ad29eSLawrence Brakmo 3253fc9f3501SEric Dumazet if (likely(!err)) { 3254e086101bSCong Wang trace_tcp_retransmit_skb(sk, skb); 3255678550c6SYuchung Cheng } else if (err != -EBUSY) { 3256ec641b39SYuchung Cheng NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL, segs); 3257fc9f3501SEric Dumazet } 3258c84a5711SYuchung Cheng return err; 325993b174adSYuchung Cheng } 326093b174adSYuchung Cheng 326110d3be56SEric Dumazet int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) 326293b174adSYuchung Cheng { 326393b174adSYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 326410d3be56SEric Dumazet int err = __tcp_retransmit_skb(sk, skb, segs); 32651da177e4SLinus Torvalds 32661da177e4SLinus Torvalds if (err == 0) { 32671da177e4SLinus Torvalds #if FASTRETRANS_DEBUG > 0 32681da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 3269e87cc472SJoe Perches net_dbg_ratelimited("retrans_out leaked\n"); 32701da177e4SLinus Torvalds } 32711da177e4SLinus Torvalds #endif 32721da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 32731da177e4SLinus Torvalds tp->retrans_out += tcp_skb_pcount(skb); 32747ae18975SYuchung Cheng } 32751da177e4SLinus Torvalds 32767ae18975SYuchung Cheng /* Save stamp of the first (attempted) retransmit. */ 32771da177e4SLinus Torvalds if (!tp->retrans_stamp) 32787faee5c0SEric Dumazet tp->retrans_stamp = tcp_skb_timestamp(skb); 32791da177e4SLinus Torvalds 32806e08d5e3SYuchung Cheng if (tp->undo_retrans < 0) 32816e08d5e3SYuchung Cheng tp->undo_retrans = 0; 32826e08d5e3SYuchung Cheng tp->undo_retrans += tcp_skb_pcount(skb); 32831da177e4SLinus Torvalds return err; 32841da177e4SLinus Torvalds } 32851da177e4SLinus Torvalds 32861da177e4SLinus Torvalds /* This gets called after a retransmit timeout, and the initially 32871da177e4SLinus Torvalds * retransmitted data is acknowledged. It tries to continue 32881da177e4SLinus Torvalds * resending the rest of the retransmit queue, until either 32891da177e4SLinus Torvalds * we've sent it all or the congestion window limit is reached. 32901da177e4SLinus Torvalds */ 32911da177e4SLinus Torvalds void tcp_xmit_retransmit_queue(struct sock *sk) 32921da177e4SLinus Torvalds { 32936687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 3294b9f1f1ceSEric Dumazet struct sk_buff *skb, *rtx_head, *hole = NULL; 32951da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 3296916e6d1aSEric Dumazet bool rearm_timer = false; 3297840a3cbeSYuchung Cheng u32 max_segs; 329861eb55f4SIlpo Järvinen int mib_idx; 32996a438bbeSStephen Hemminger 330045e77d31SIlpo Järvinen if (!tp->packets_out) 330145e77d31SIlpo Järvinen return; 330245e77d31SIlpo Järvinen 330375c119afSEric Dumazet rtx_head = tcp_rtx_queue_head(sk); 3304b9f1f1ceSEric Dumazet skb = tp->retransmit_skb_hint ?: rtx_head; 3305ed6e7268SNeal Cardwell max_segs = tcp_tso_segs(sk, tcp_current_mss(sk)); 330675c119afSEric Dumazet skb_rbtree_walk_from(skb) { 3307dca0aaf8SEric Dumazet __u8 sacked; 330810d3be56SEric Dumazet int segs; 33091da177e4SLinus Torvalds 3310218af599SEric Dumazet if (tcp_pacing_check(sk)) 3311218af599SEric Dumazet break; 3312218af599SEric Dumazet 33136a438bbeSStephen Hemminger /* we could do better than to assign each time */ 331451456b29SIan Morris if (!hole) 33156a438bbeSStephen Hemminger tp->retransmit_skb_hint = skb; 33166a438bbeSStephen Hemminger 331710d3be56SEric Dumazet segs = tp->snd_cwnd - tcp_packets_in_flight(tp); 331810d3be56SEric Dumazet if (segs <= 0) 3319916e6d1aSEric Dumazet break; 3320dca0aaf8SEric Dumazet sacked = TCP_SKB_CB(skb)->sacked; 3321a3d2e9f8SEric Dumazet /* In case tcp_shift_skb_data() have aggregated large skbs, 3322a3d2e9f8SEric Dumazet * we need to make sure not sending too bigs TSO packets 3323a3d2e9f8SEric Dumazet */ 3324a3d2e9f8SEric Dumazet segs = min_t(int, segs, max_segs); 33250e1c54c2SIlpo Järvinen 3326840a3cbeSYuchung Cheng if (tp->retrans_out >= tp->lost_out) { 3327006f582cSIlpo Järvinen break; 33280e1c54c2SIlpo Järvinen } else if (!(sacked & TCPCB_LOST)) { 332951456b29SIan Morris if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) 33300e1c54c2SIlpo Järvinen hole = skb; 333161eb55f4SIlpo Järvinen continue; 33321da177e4SLinus Torvalds 33330e1c54c2SIlpo Järvinen } else { 33340e1c54c2SIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Loss) 33350e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFASTRETRANS; 33360e1c54c2SIlpo Järvinen else 33370e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; 33380e1c54c2SIlpo Järvinen } 33390e1c54c2SIlpo Järvinen 33400e1c54c2SIlpo Järvinen if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) 334161eb55f4SIlpo Järvinen continue; 334240b215e5SPavel Emelyanov 3343f9616c35SEric Dumazet if (tcp_small_queue_check(sk, skb, 1)) 3344916e6d1aSEric Dumazet break; 3345f9616c35SEric Dumazet 334610d3be56SEric Dumazet if (tcp_retransmit_skb(sk, skb, segs)) 3347916e6d1aSEric Dumazet break; 334824ab6becSYuchung Cheng 3349de1d6578SYuchung Cheng NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb)); 33501da177e4SLinus Torvalds 3351684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 3352a262f0cdSNandita Dukkipati tp->prr_out += tcp_skb_pcount(skb); 3353a262f0cdSNandita Dukkipati 335475c119afSEric Dumazet if (skb == rtx_head && 335557dde7f7SYuchung Cheng icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT) 3356916e6d1aSEric Dumazet rearm_timer = true; 3357916e6d1aSEric Dumazet 3358916e6d1aSEric Dumazet } 3359916e6d1aSEric Dumazet if (rearm_timer) 33603f80e08fSEric Dumazet tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 33613f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, 33628dc242adSEric Dumazet TCP_RTO_MAX); 33631da177e4SLinus Torvalds } 33641da177e4SLinus Torvalds 3365d83769a5SEric Dumazet /* We allow to exceed memory limits for FIN packets to expedite 3366d83769a5SEric Dumazet * connection tear down and (memory) recovery. 3367845704a5SEric Dumazet * Otherwise tcp_send_fin() could be tempted to either delay FIN 3368845704a5SEric Dumazet * or even be forced to close flow without any FIN. 3369a6c5ea4cSEric Dumazet * In general, we want to allow one skb per socket to avoid hangs 3370a6c5ea4cSEric Dumazet * with edge trigger epoll() 3371d83769a5SEric Dumazet */ 3372a6c5ea4cSEric Dumazet void sk_forced_mem_schedule(struct sock *sk, int size) 3373d83769a5SEric Dumazet { 3374e805605cSJohannes Weiner int amt; 3375d83769a5SEric Dumazet 3376d83769a5SEric Dumazet if (size <= sk->sk_forward_alloc) 3377d83769a5SEric Dumazet return; 3378d83769a5SEric Dumazet amt = sk_mem_pages(size); 3379d83769a5SEric Dumazet sk->sk_forward_alloc += amt * SK_MEM_QUANTUM; 3380e805605cSJohannes Weiner sk_memory_allocated_add(sk, amt); 3381e805605cSJohannes Weiner 3382baac50bbSJohannes Weiner if (mem_cgroup_sockets_enabled && sk->sk_memcg) 3383baac50bbSJohannes Weiner mem_cgroup_charge_skmem(sk->sk_memcg, amt); 3384d83769a5SEric Dumazet } 3385d83769a5SEric Dumazet 3386845704a5SEric Dumazet /* Send a FIN. The caller locks the socket for us. 3387845704a5SEric Dumazet * We should try to send a FIN packet really hard, but eventually give up. 33881da177e4SLinus Torvalds */ 33891da177e4SLinus Torvalds void tcp_send_fin(struct sock *sk) 33901da177e4SLinus Torvalds { 3391ee2aabd3SEric Dumazet struct sk_buff *skb, *tskb, *tail = tcp_write_queue_tail(sk); 33921da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 33931da177e4SLinus Torvalds 3394845704a5SEric Dumazet /* Optimization, tack on the FIN if we have one skb in write queue and 3395845704a5SEric Dumazet * this skb was not yet sent, or we are under memory pressure. 3396845704a5SEric Dumazet * Note: in the latter case, FIN packet will be sent after a timeout, 3397845704a5SEric Dumazet * as TCP stack thinks it has already been transmitted. 33981da177e4SLinus Torvalds */ 3399ee2aabd3SEric Dumazet tskb = tail; 340075c119afSEric Dumazet if (!tskb && tcp_under_memory_pressure(sk)) 340175c119afSEric Dumazet tskb = skb_rb_last(&sk->tcp_rtx_queue); 340275c119afSEric Dumazet 340375c119afSEric Dumazet if (tskb) { 3404845704a5SEric Dumazet TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN; 3405845704a5SEric Dumazet TCP_SKB_CB(tskb)->end_seq++; 34061da177e4SLinus Torvalds tp->write_seq++; 3407ee2aabd3SEric Dumazet if (!tail) { 3408845704a5SEric Dumazet /* This means tskb was already sent. 3409845704a5SEric Dumazet * Pretend we included the FIN on previous transmit. 3410845704a5SEric Dumazet * We need to set tp->snd_nxt to the value it would have 3411845704a5SEric Dumazet * if FIN had been sent. This is because retransmit path 3412845704a5SEric Dumazet * does not change tp->snd_nxt. 3413845704a5SEric Dumazet */ 3414e0d694d6SEric Dumazet WRITE_ONCE(tp->snd_nxt, tp->snd_nxt + 1); 3415845704a5SEric Dumazet return; 3416845704a5SEric Dumazet } 34171da177e4SLinus Torvalds } else { 3418845704a5SEric Dumazet skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation); 3419d1edc085SColin Ian King if (unlikely(!skb)) 3420845704a5SEric Dumazet return; 3421d1edc085SColin Ian King 3422e2080072SEric Dumazet INIT_LIST_HEAD(&skb->tcp_tsorted_anchor); 3423d83769a5SEric Dumazet skb_reserve(skb, MAX_TCP_HEADER); 3424a6c5ea4cSEric Dumazet sk_forced_mem_schedule(sk, skb->truesize); 34251da177e4SLinus Torvalds /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 3426e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tp->write_seq, 3427a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_FIN); 34281da177e4SLinus Torvalds tcp_queue_skb(sk, skb); 34291da177e4SLinus Torvalds } 3430845704a5SEric Dumazet __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF); 34311da177e4SLinus Torvalds } 34321da177e4SLinus Torvalds 34331da177e4SLinus Torvalds /* We get here when a process closes a file descriptor (either due to 34341da177e4SLinus Torvalds * an explicit close() or as a byproduct of exit()'ing) and there 34351da177e4SLinus Torvalds * was unread data in the receive queue. This behavior is recommended 343665bb723cSGerrit Renker * by RFC 2525, section 2.17. -DaveM 34371da177e4SLinus Torvalds */ 3438dd0fc66fSAl Viro void tcp_send_active_reset(struct sock *sk, gfp_t priority) 34391da177e4SLinus Torvalds { 34401da177e4SLinus Torvalds struct sk_buff *skb; 34411da177e4SLinus Torvalds 34427cc2b043SGao Feng TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); 34437cc2b043SGao Feng 34441da177e4SLinus Torvalds /* NOTE: No TCP options attached and we never retransmit this. */ 34451da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, priority); 34461da177e4SLinus Torvalds if (!skb) { 34474e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 34481da177e4SLinus Torvalds return; 34491da177e4SLinus Torvalds } 34501da177e4SLinus Torvalds 34511da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 34521da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 3453e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), 3454a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_RST); 34559a568de4SEric Dumazet tcp_mstamp_refresh(tcp_sk(sk)); 34561da177e4SLinus Torvalds /* Send it off. */ 3457dfb4b9dcSDavid S. Miller if (tcp_transmit_skb(sk, skb, 0, priority)) 34584e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 3459c24b14c4SSong Liu 3460c24b14c4SSong Liu /* skb of trace_tcp_send_reset() keeps the skb that caused RST, 3461c24b14c4SSong Liu * skb here is different to the troublesome skb, so use NULL 3462c24b14c4SSong Liu */ 3463c24b14c4SSong Liu trace_tcp_send_reset(sk, NULL); 34641da177e4SLinus Torvalds } 34651da177e4SLinus Torvalds 346667edfef7SAndi Kleen /* Send a crossed SYN-ACK during socket establishment. 346767edfef7SAndi Kleen * WARNING: This routine must only be called when we have already sent 34681da177e4SLinus Torvalds * a SYN packet that crossed the incoming SYN that caused this routine 34691da177e4SLinus Torvalds * to get called. If this assumption fails then the initial rcv_wnd 34701da177e4SLinus Torvalds * and rcv_wscale values will not be correct. 34711da177e4SLinus Torvalds */ 34721da177e4SLinus Torvalds int tcp_send_synack(struct sock *sk) 34731da177e4SLinus Torvalds { 34741da177e4SLinus Torvalds struct sk_buff *skb; 34751da177e4SLinus Torvalds 347675c119afSEric Dumazet skb = tcp_rtx_queue_head(sk); 347751456b29SIan Morris if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 347875c119afSEric Dumazet pr_err("%s: wrong queue state\n", __func__); 34791da177e4SLinus Torvalds return -EFAULT; 34801da177e4SLinus Torvalds } 34814de075e0SEric Dumazet if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { 34821da177e4SLinus Torvalds if (skb_cloned(skb)) { 3483e2080072SEric Dumazet struct sk_buff *nskb; 3484e2080072SEric Dumazet 3485e2080072SEric Dumazet tcp_skb_tsorted_save(skb) { 3486e2080072SEric Dumazet nskb = skb_copy(skb, GFP_ATOMIC); 3487e2080072SEric Dumazet } tcp_skb_tsorted_restore(skb); 348851456b29SIan Morris if (!nskb) 34891da177e4SLinus Torvalds return -ENOMEM; 3490e2080072SEric Dumazet INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor); 34912bec445fSEric Dumazet tcp_highest_sack_replace(sk, skb, nskb); 349275c119afSEric Dumazet tcp_rtx_queue_unlink_and_free(skb, sk); 3493f4a775d1SEric Dumazet __skb_header_release(nskb); 349475c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb); 3495ab4e846aSEric Dumazet sk_wmem_queued_add(sk, nskb->truesize); 34963ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 34971da177e4SLinus Torvalds skb = nskb; 34981da177e4SLinus Torvalds } 34991da177e4SLinus Torvalds 35004de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; 3501735d3831SFlorian Westphal tcp_ecn_send_synack(sk, skb); 35021da177e4SLinus Torvalds } 3503dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 35041da177e4SLinus Torvalds } 35051da177e4SLinus Torvalds 35064aea39c1SEric Dumazet /** 3507331fca43SMartin KaFai Lau * tcp_make_synack - Allocate one skb and build a SYNACK packet. 3508331fca43SMartin KaFai Lau * @sk: listener socket 3509331fca43SMartin KaFai Lau * @dst: dst entry attached to the SYNACK. It is consumed and caller 3510331fca43SMartin KaFai Lau * should not use it again. 3511331fca43SMartin KaFai Lau * @req: request_sock pointer 3512331fca43SMartin KaFai Lau * @foc: cookie for tcp fast open 3513331fca43SMartin KaFai Lau * @synack_type: Type of synack to prepare 3514331fca43SMartin KaFai Lau * @syn_skb: SYN packet just received. It could be NULL for rtx case. 35154aea39c1SEric Dumazet */ 35165d062de7SEric Dumazet struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst, 3517e6b4d113SWilliam Allen Simpson struct request_sock *req, 3518ca6fb065SEric Dumazet struct tcp_fastopen_cookie *foc, 3519331fca43SMartin KaFai Lau enum tcp_synack_type synack_type, 3520331fca43SMartin KaFai Lau struct sk_buff *syn_skb) 35211da177e4SLinus Torvalds { 35222e6599cbSArnaldo Carvalho de Melo struct inet_request_sock *ireq = inet_rsk(req); 35235d062de7SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 352480f03e27SEric Dumazet struct tcp_md5sig_key *md5 = NULL; 35255d062de7SEric Dumazet struct tcp_out_options opts; 35265d062de7SEric Dumazet struct sk_buff *skb; 3527bd0388aeSWilliam Allen Simpson int tcp_header_size; 35285d062de7SEric Dumazet struct tcphdr *th; 3529f5fff5dcSTom Quetchenbach int mss; 3530a842fe14SEric Dumazet u64 now; 35311da177e4SLinus Torvalds 3532ca6fb065SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 35334aea39c1SEric Dumazet if (unlikely(!skb)) { 35344aea39c1SEric Dumazet dst_release(dst); 35351da177e4SLinus Torvalds return NULL; 35364aea39c1SEric Dumazet } 35371da177e4SLinus Torvalds /* Reserve space for headers. */ 35381da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 35391da177e4SLinus Torvalds 3540b3d05147SEric Dumazet switch (synack_type) { 3541b3d05147SEric Dumazet case TCP_SYNACK_NORMAL: 35429e17f8a4SEric Dumazet skb_set_owner_w(skb, req_to_sk(req)); 3543b3d05147SEric Dumazet break; 3544b3d05147SEric Dumazet case TCP_SYNACK_COOKIE: 3545b3d05147SEric Dumazet /* Under synflood, we do not attach skb to a socket, 3546b3d05147SEric Dumazet * to avoid false sharing. 3547b3d05147SEric Dumazet */ 3548b3d05147SEric Dumazet break; 3549b3d05147SEric Dumazet case TCP_SYNACK_FASTOPEN: 3550ca6fb065SEric Dumazet /* sk is a const pointer, because we want to express multiple 3551ca6fb065SEric Dumazet * cpu might call us concurrently. 3552ca6fb065SEric Dumazet * sk->sk_wmem_alloc in an atomic, we can promote to rw. 3553ca6fb065SEric Dumazet */ 3554ca6fb065SEric Dumazet skb_set_owner_w(skb, (struct sock *)sk); 3555b3d05147SEric Dumazet break; 3556ca6fb065SEric Dumazet } 35574aea39c1SEric Dumazet skb_dst_set(skb, dst); 35581da177e4SLinus Torvalds 35593541f9e8SEric Dumazet mss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 3560f5fff5dcSTom Quetchenbach 356133ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 3562a842fe14SEric Dumazet now = tcp_clock_ns(); 35638b5f12d0SFlorian Westphal #ifdef CONFIG_SYN_COOKIES 3564f8ace8d9SFlorian Westphal if (unlikely(synack_type == TCP_SYNACK_COOKIE && ireq->tstamp_ok)) 3565200ecef6SEric Dumazet skb->skb_mstamp_ns = cookie_init_timestamp(req, now); 35668b5f12d0SFlorian Westphal else 35678b5f12d0SFlorian Westphal #endif 35689e450c1eSYuchung Cheng { 3569a842fe14SEric Dumazet skb->skb_mstamp_ns = now; 35709e450c1eSYuchung Cheng if (!tcp_rsk(req)->snt_synack) /* Timestamp first SYNACK */ 35719e450c1eSYuchung Cheng tcp_rsk(req)->snt_synack = tcp_skb_timestamp_us(skb); 35729e450c1eSYuchung Cheng } 357380f03e27SEric Dumazet 357480f03e27SEric Dumazet #ifdef CONFIG_TCP_MD5SIG 357580f03e27SEric Dumazet rcu_read_lock(); 3576fd3a154aSEric Dumazet md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req)); 357780f03e27SEric Dumazet #endif 357858d607d3SEric Dumazet skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4); 3579331fca43SMartin KaFai Lau /* bpf program will be interested in the tcp_flags */ 3580331fca43SMartin KaFai Lau TCP_SKB_CB(skb)->tcp_flags = TCPHDR_SYN | TCPHDR_ACK; 358160e2a778SUrsula Braun tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5, 3582331fca43SMartin KaFai Lau foc, synack_type, 3583331fca43SMartin KaFai Lau syn_skb) + sizeof(*th); 358433ad798cSAdam Langley 3585aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 3586aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 35871da177e4SLinus Torvalds 3588ea1627c2SEric Dumazet th = (struct tcphdr *)skb->data; 35891da177e4SLinus Torvalds memset(th, 0, sizeof(struct tcphdr)); 35901da177e4SLinus Torvalds th->syn = 1; 35911da177e4SLinus Torvalds th->ack = 1; 35926ac705b1SEric Dumazet tcp_ecn_make_synack(req, th); 3593b44084c2SEric Dumazet th->source = htons(ireq->ir_num); 3594634fb979SEric Dumazet th->dest = ireq->ir_rmt_port; 3595e05a90ecSJamal Hadi Salim skb->mark = ireq->ir_mark; 35963b117750SEric Dumazet skb->ip_summed = CHECKSUM_PARTIAL; 35973b117750SEric Dumazet th->seq = htonl(tcp_rsk(req)->snt_isn); 35988336886fSJerry Chu /* XXX data is queued and acked as is. No buffer/window check */ 35998336886fSJerry Chu th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt); 36001da177e4SLinus Torvalds 36011da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 3602ed53d0abSEric Dumazet th->window = htons(min(req->rsk_rcv_wnd, 65535U)); 36035d062de7SEric Dumazet tcp_options_write((__be32 *)(th + 1), NULL, &opts); 36041da177e4SLinus Torvalds th->doff = (tcp_header_size >> 2); 360590bbcc60SEric Dumazet __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS); 3606cfb6eeb4SYOSHIFUJI Hideaki 3607cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 3608cfb6eeb4SYOSHIFUJI Hideaki /* Okay, we have all we need - do the md5 hash if needed */ 360980f03e27SEric Dumazet if (md5) 3610bd0388aeSWilliam Allen Simpson tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, 361139f8e58eSEric Dumazet md5, req_to_sk(req), skb); 361280f03e27SEric Dumazet rcu_read_unlock(); 3613cfb6eeb4SYOSHIFUJI Hideaki #endif 3614cfb6eeb4SYOSHIFUJI Hideaki 3615331fca43SMartin KaFai Lau bpf_skops_write_hdr_opt((struct sock *)sk, skb, req, syn_skb, 3616331fca43SMartin KaFai Lau synack_type, &opts); 3617331fca43SMartin KaFai Lau 3618a842fe14SEric Dumazet skb->skb_mstamp_ns = now; 3619a842fe14SEric Dumazet tcp_add_tx_delay(skb, tp); 3620a842fe14SEric Dumazet 36211da177e4SLinus Torvalds return skb; 36221da177e4SLinus Torvalds } 36234bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_make_synack); 36241da177e4SLinus Torvalds 362581164413SDaniel Borkmann static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst) 362681164413SDaniel Borkmann { 362781164413SDaniel Borkmann struct inet_connection_sock *icsk = inet_csk(sk); 362881164413SDaniel Borkmann const struct tcp_congestion_ops *ca; 362981164413SDaniel Borkmann u32 ca_key = dst_metric(dst, RTAX_CC_ALGO); 363081164413SDaniel Borkmann 363181164413SDaniel Borkmann if (ca_key == TCP_CA_UNSPEC) 363281164413SDaniel Borkmann return; 363381164413SDaniel Borkmann 363481164413SDaniel Borkmann rcu_read_lock(); 363581164413SDaniel Borkmann ca = tcp_ca_find_key(ca_key); 36360baf26b0SMartin KaFai Lau if (likely(ca && bpf_try_module_get(ca, ca->owner))) { 36370baf26b0SMartin KaFai Lau bpf_module_put(icsk->icsk_ca_ops, icsk->icsk_ca_ops->owner); 363881164413SDaniel Borkmann icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst); 363981164413SDaniel Borkmann icsk->icsk_ca_ops = ca; 364081164413SDaniel Borkmann } 364181164413SDaniel Borkmann rcu_read_unlock(); 364281164413SDaniel Borkmann } 364381164413SDaniel Borkmann 364467edfef7SAndi Kleen /* Do all connect socket setups that can be done AF independent. */ 3645f7e56a76Sstephen hemminger static void tcp_connect_init(struct sock *sk) 36461da177e4SLinus Torvalds { 3647cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 36481da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 36491da177e4SLinus Torvalds __u8 rcv_wscale; 365013d3b1ebSLawrence Brakmo u32 rcv_wnd; 36511da177e4SLinus Torvalds 36521da177e4SLinus Torvalds /* We'll fix this up when we get a response from the other end. 36531da177e4SLinus Torvalds * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 36541da177e4SLinus Torvalds */ 36555d2ed052SEric Dumazet tp->tcp_header_len = sizeof(struct tcphdr); 36565d2ed052SEric Dumazet if (sock_net(sk)->ipv4.sysctl_tcp_timestamps) 36575d2ed052SEric Dumazet tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED; 36581da177e4SLinus Torvalds 3659cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 366000db4124SIan Morris if (tp->af_specific->md5_lookup(sk, sk)) 3661cfb6eeb4SYOSHIFUJI Hideaki tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 3662cfb6eeb4SYOSHIFUJI Hideaki #endif 3663cfb6eeb4SYOSHIFUJI Hideaki 36641da177e4SLinus Torvalds /* If user gave his TCP_MAXSEG, record it to clamp */ 36651da177e4SLinus Torvalds if (tp->rx_opt.user_mss) 36661da177e4SLinus Torvalds tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 36671da177e4SLinus Torvalds tp->max_window = 0; 36685d424d5aSJohn Heffner tcp_mtup_init(sk); 36691da177e4SLinus Torvalds tcp_sync_mss(sk, dst_mtu(dst)); 36701da177e4SLinus Torvalds 367181164413SDaniel Borkmann tcp_ca_dst_init(sk, dst); 367281164413SDaniel Borkmann 36731da177e4SLinus Torvalds if (!tp->window_clamp) 36741da177e4SLinus Torvalds tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 36753541f9e8SEric Dumazet tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 3676f5fff5dcSTom Quetchenbach 36771da177e4SLinus Torvalds tcp_initialize_rcv_mss(sk); 36781da177e4SLinus Torvalds 3679e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 3680e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 3681e88c64f0SHagen Paul Pfeifer (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) 3682e88c64f0SHagen Paul Pfeifer tp->window_clamp = tcp_full_space(sk); 3683e88c64f0SHagen Paul Pfeifer 368413d3b1ebSLawrence Brakmo rcv_wnd = tcp_rwnd_init_bpf(sk); 368513d3b1ebSLawrence Brakmo if (rcv_wnd == 0) 368613d3b1ebSLawrence Brakmo rcv_wnd = dst_metric(dst, RTAX_INITRWND); 368713d3b1ebSLawrence Brakmo 3688ceef9ab6SEric Dumazet tcp_select_initial_window(sk, tcp_full_space(sk), 36891da177e4SLinus Torvalds tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 36901da177e4SLinus Torvalds &tp->rcv_wnd, 36911da177e4SLinus Torvalds &tp->window_clamp, 36929bb37ef0SEric Dumazet sock_net(sk)->ipv4.sysctl_tcp_window_scaling, 369331d12926Slaurent chavey &rcv_wscale, 369413d3b1ebSLawrence Brakmo rcv_wnd); 36951da177e4SLinus Torvalds 36961da177e4SLinus Torvalds tp->rx_opt.rcv_wscale = rcv_wscale; 36971da177e4SLinus Torvalds tp->rcv_ssthresh = tp->rcv_wnd; 36981da177e4SLinus Torvalds 36991da177e4SLinus Torvalds sk->sk_err = 0; 37001da177e4SLinus Torvalds sock_reset_flag(sk, SOCK_DONE); 37011da177e4SLinus Torvalds tp->snd_wnd = 0; 3702ee7537b6SHantzis Fotis tcp_init_wl(tp, 0); 37037f582b24SEric Dumazet tcp_write_queue_purge(sk); 37041da177e4SLinus Torvalds tp->snd_una = tp->write_seq; 37051da177e4SLinus Torvalds tp->snd_sml = tp->write_seq; 370633f5f57eSIlpo Järvinen tp->snd_up = tp->write_seq; 3707e0d694d6SEric Dumazet WRITE_ONCE(tp->snd_nxt, tp->write_seq); 3708ee995283SPavel Emelyanov 3709ee995283SPavel Emelyanov if (likely(!tp->repair)) 37101da177e4SLinus Torvalds tp->rcv_nxt = 0; 3711c7781a6eSAndrew Vagin else 371270eabf0eSEric Dumazet tp->rcv_tstamp = tcp_jiffies32; 3713ee995283SPavel Emelyanov tp->rcv_wup = tp->rcv_nxt; 37147db48e98SEric Dumazet WRITE_ONCE(tp->copied_seq, tp->rcv_nxt); 37151da177e4SLinus Torvalds 37168550f328SLawrence Brakmo inet_csk(sk)->icsk_rto = tcp_timeout_init(sk); 3717463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_retransmits = 0; 37181da177e4SLinus Torvalds tcp_clear_retrans(tp); 37191da177e4SLinus Torvalds } 37201da177e4SLinus Torvalds 3721783237e8SYuchung Cheng static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) 3722783237e8SYuchung Cheng { 3723783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3724783237e8SYuchung Cheng struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 3725783237e8SYuchung Cheng 3726783237e8SYuchung Cheng tcb->end_seq += skb->len; 3727f4a775d1SEric Dumazet __skb_header_release(skb); 3728ab4e846aSEric Dumazet sk_wmem_queued_add(sk, skb->truesize); 3729783237e8SYuchung Cheng sk_mem_charge(sk, skb->truesize); 37300f317464SEric Dumazet WRITE_ONCE(tp->write_seq, tcb->end_seq); 3731783237e8SYuchung Cheng tp->packets_out += tcp_skb_pcount(skb); 3732783237e8SYuchung Cheng } 3733783237e8SYuchung Cheng 3734783237e8SYuchung Cheng /* Build and send a SYN with data and (cached) Fast Open cookie. However, 3735783237e8SYuchung Cheng * queue a data-only packet after the regular SYN, such that regular SYNs 3736783237e8SYuchung Cheng * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges 3737783237e8SYuchung Cheng * only the SYN sequence, the data are retransmitted in the first ACK. 3738783237e8SYuchung Cheng * If cookie is not cached or other error occurs, falls back to send a 3739783237e8SYuchung Cheng * regular SYN with Fast Open cookie request option. 3740783237e8SYuchung Cheng */ 3741783237e8SYuchung Cheng static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) 3742783237e8SYuchung Cheng { 3743783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3744783237e8SYuchung Cheng struct tcp_fastopen_request *fo = tp->fastopen_req; 3745065263f4SWei Wang int space, err = 0; 3746355a901eSEric Dumazet struct sk_buff *syn_data; 3747783237e8SYuchung Cheng 374867da22d2SYuchung Cheng tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ 3749065263f4SWei Wang if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie)) 3750783237e8SYuchung Cheng goto fallback; 3751783237e8SYuchung Cheng 3752783237e8SYuchung Cheng /* MSS for SYN-data is based on cached MSS and bounded by PMTU and 3753783237e8SYuchung Cheng * user-MSS. Reserve maximum option space for middleboxes that add 3754783237e8SYuchung Cheng * private TCP options. The cost is reduced data space in SYN :( 3755783237e8SYuchung Cheng */ 37563541f9e8SEric Dumazet tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp); 37573541f9e8SEric Dumazet 37581b63edd6SYuchung Cheng space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) - 3759783237e8SYuchung Cheng MAX_TCP_OPTION_SPACE; 3760783237e8SYuchung Cheng 3761f5ddcbbbSEric Dumazet space = min_t(size_t, space, fo->size); 3762f5ddcbbbSEric Dumazet 3763f5ddcbbbSEric Dumazet /* limit to order-0 allocations */ 3764f5ddcbbbSEric Dumazet space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER)); 3765f5ddcbbbSEric Dumazet 3766eb934478SEric Dumazet syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false); 3767355a901eSEric Dumazet if (!syn_data) 3768783237e8SYuchung Cheng goto fallback; 3769355a901eSEric Dumazet syn_data->ip_summed = CHECKSUM_PARTIAL; 3770355a901eSEric Dumazet memcpy(syn_data->cb, syn->cb, sizeof(syn->cb)); 377107e100f9SEric Dumazet if (space) { 377207e100f9SEric Dumazet int copied = copy_from_iter(skb_put(syn_data, space), space, 377357be5bdaSAl Viro &fo->data->msg_iter); 377457be5bdaSAl Viro if (unlikely(!copied)) { 3775ba233b34SEric Dumazet tcp_skb_tsorted_anchor_cleanup(syn_data); 3776355a901eSEric Dumazet kfree_skb(syn_data); 3777783237e8SYuchung Cheng goto fallback; 3778783237e8SYuchung Cheng } 377957be5bdaSAl Viro if (copied != space) { 378057be5bdaSAl Viro skb_trim(syn_data, copied); 378157be5bdaSAl Viro space = copied; 378257be5bdaSAl Viro } 3783f859a448SWillem de Bruijn skb_zcopy_set(syn_data, fo->uarg, NULL); 378407e100f9SEric Dumazet } 3785355a901eSEric Dumazet /* No more data pending in inet_wait_for_connect() */ 3786355a901eSEric Dumazet if (space == fo->size) 3787355a901eSEric Dumazet fo->data = NULL; 3788355a901eSEric Dumazet fo->copied = space; 3789783237e8SYuchung Cheng 3790355a901eSEric Dumazet tcp_connect_queue_skb(sk, syn_data); 37910f87230dSFrancis Yan if (syn_data->len) 37920f87230dSFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_BUSY); 3793355a901eSEric Dumazet 3794355a901eSEric Dumazet err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation); 3795355a901eSEric Dumazet 3796d3edd06eSEric Dumazet syn->skb_mstamp_ns = syn_data->skb_mstamp_ns; 3797355a901eSEric Dumazet 3798355a901eSEric Dumazet /* Now full SYN+DATA was cloned and sent (or not), 3799355a901eSEric Dumazet * remove the SYN from the original skb (syn_data) 3800355a901eSEric Dumazet * we keep in write queue in case of a retransmit, as we 3801355a901eSEric Dumazet * also have the SYN packet (with no data) in the same queue. 3802431a9124SEric Dumazet */ 3803355a901eSEric Dumazet TCP_SKB_CB(syn_data)->seq++; 3804355a901eSEric Dumazet TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH; 3805355a901eSEric Dumazet if (!err) { 380667da22d2SYuchung Cheng tp->syn_data = (fo->copied > 0); 380775c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data); 3808f19c29e3SYuchung Cheng NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT); 3809783237e8SYuchung Cheng goto done; 3810783237e8SYuchung Cheng } 3811783237e8SYuchung Cheng 381275c119afSEric Dumazet /* data was not sent, put it in write_queue */ 381375c119afSEric Dumazet __skb_queue_tail(&sk->sk_write_queue, syn_data); 3814b5b7db8dSEric Dumazet tp->packets_out -= tcp_skb_pcount(syn_data); 3815b5b7db8dSEric Dumazet 3816783237e8SYuchung Cheng fallback: 3817783237e8SYuchung Cheng /* Send a regular SYN with Fast Open cookie request option */ 3818783237e8SYuchung Cheng if (fo->cookie.len > 0) 3819783237e8SYuchung Cheng fo->cookie.len = 0; 3820783237e8SYuchung Cheng err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); 3821783237e8SYuchung Cheng if (err) 3822783237e8SYuchung Cheng tp->syn_fastopen = 0; 3823783237e8SYuchung Cheng done: 3824783237e8SYuchung Cheng fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ 3825783237e8SYuchung Cheng return err; 3826783237e8SYuchung Cheng } 3827783237e8SYuchung Cheng 382867edfef7SAndi Kleen /* Build a SYN and send it off. */ 38291da177e4SLinus Torvalds int tcp_connect(struct sock *sk) 38301da177e4SLinus Torvalds { 38311da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 38321da177e4SLinus Torvalds struct sk_buff *buff; 3833ee586811SEric Paris int err; 38341da177e4SLinus Torvalds 3835de525be2SLawrence Brakmo tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB, 0, NULL); 38368ba60924SEric Dumazet 38378ba60924SEric Dumazet if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 38388ba60924SEric Dumazet return -EHOSTUNREACH; /* Routing failure or similar. */ 38398ba60924SEric Dumazet 38401da177e4SLinus Torvalds tcp_connect_init(sk); 38411da177e4SLinus Torvalds 38422b916477SAndrey Vagin if (unlikely(tp->repair)) { 38432b916477SAndrey Vagin tcp_finish_connect(sk, NULL); 38442b916477SAndrey Vagin return 0; 38452b916477SAndrey Vagin } 38462b916477SAndrey Vagin 3847eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true); 3848355a901eSEric Dumazet if (unlikely(!buff)) 38491da177e4SLinus Torvalds return -ENOBUFS; 38501da177e4SLinus Torvalds 3851a3433f35SChangli Gao tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); 38529a568de4SEric Dumazet tcp_mstamp_refresh(tp); 38539a568de4SEric Dumazet tp->retrans_stamp = tcp_time_stamp(tp); 3854783237e8SYuchung Cheng tcp_connect_queue_skb(sk, buff); 3855735d3831SFlorian Westphal tcp_ecn_send_syn(sk, buff); 385675c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, buff); 38571da177e4SLinus Torvalds 3858783237e8SYuchung Cheng /* Send off SYN; include data in Fast Open. */ 3859783237e8SYuchung Cheng err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : 3860783237e8SYuchung Cheng tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); 3861ee586811SEric Paris if (err == -ECONNREFUSED) 3862ee586811SEric Paris return err; 3863bd37a088SWei Yongjun 3864bd37a088SWei Yongjun /* We change tp->snd_nxt after the tcp_transmit_skb() call 3865bd37a088SWei Yongjun * in order to make this packet get counted in tcpOutSegs. 3866bd37a088SWei Yongjun */ 3867e0d694d6SEric Dumazet WRITE_ONCE(tp->snd_nxt, tp->write_seq); 3868bd37a088SWei Yongjun tp->pushed_seq = tp->write_seq; 3869b5b7db8dSEric Dumazet buff = tcp_send_head(sk); 3870b5b7db8dSEric Dumazet if (unlikely(buff)) { 3871e0d694d6SEric Dumazet WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(buff)->seq); 3872b5b7db8dSEric Dumazet tp->pushed_seq = TCP_SKB_CB(buff)->seq; 3873b5b7db8dSEric Dumazet } 387481cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); 38751da177e4SLinus Torvalds 38761da177e4SLinus Torvalds /* Timer for repeating the SYN until an answer. */ 38773f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 38783f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 38791da177e4SLinus Torvalds return 0; 38801da177e4SLinus Torvalds } 38814bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_connect); 38821da177e4SLinus Torvalds 38831da177e4SLinus Torvalds /* Send out a delayed ack, the caller does the policy checking 38841da177e4SLinus Torvalds * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 38851da177e4SLinus Torvalds * for details. 38861da177e4SLinus Torvalds */ 38871da177e4SLinus Torvalds void tcp_send_delayed_ack(struct sock *sk) 38881da177e4SLinus Torvalds { 3889463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 3890463c84b9SArnaldo Carvalho de Melo int ato = icsk->icsk_ack.ato; 38911da177e4SLinus Torvalds unsigned long timeout; 38921da177e4SLinus Torvalds 38931da177e4SLinus Torvalds if (ato > TCP_DELACK_MIN) { 3894463c84b9SArnaldo Carvalho de Melo const struct tcp_sock *tp = tcp_sk(sk); 38951da177e4SLinus Torvalds int max_ato = HZ / 2; 38961da177e4SLinus Torvalds 389731954cd8SWei Wang if (inet_csk_in_pingpong_mode(sk) || 3898056834d9SIlpo Järvinen (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 38991da177e4SLinus Torvalds max_ato = TCP_DELACK_MAX; 39001da177e4SLinus Torvalds 39011da177e4SLinus Torvalds /* Slow path, intersegment interval is "high". */ 39021da177e4SLinus Torvalds 39031da177e4SLinus Torvalds /* If some rtt estimate is known, use it to bound delayed ack. 3904463c84b9SArnaldo Carvalho de Melo * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 39051da177e4SLinus Torvalds * directly. 39061da177e4SLinus Torvalds */ 3907740b0f18SEric Dumazet if (tp->srtt_us) { 3908740b0f18SEric Dumazet int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3), 3909740b0f18SEric Dumazet TCP_DELACK_MIN); 39101da177e4SLinus Torvalds 39111da177e4SLinus Torvalds if (rtt < max_ato) 39121da177e4SLinus Torvalds max_ato = rtt; 39131da177e4SLinus Torvalds } 39141da177e4SLinus Torvalds 39151da177e4SLinus Torvalds ato = min(ato, max_ato); 39161da177e4SLinus Torvalds } 39171da177e4SLinus Torvalds 39182b8ee4f0SMartin KaFai Lau ato = min_t(u32, ato, inet_csk(sk)->icsk_delack_max); 39192b8ee4f0SMartin KaFai Lau 39201da177e4SLinus Torvalds /* Stay within the limit we were given */ 39211da177e4SLinus Torvalds timeout = jiffies + ato; 39221da177e4SLinus Torvalds 39231da177e4SLinus Torvalds /* Use new timeout only if there wasn't a older one earlier. */ 3924463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 3925b6b6d653SEric Dumazet /* If delack timer is about to expire, send ACK now. */ 3926b6b6d653SEric Dumazet if (time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 39271da177e4SLinus Torvalds tcp_send_ack(sk); 39281da177e4SLinus Torvalds return; 39291da177e4SLinus Torvalds } 39301da177e4SLinus Torvalds 3931463c84b9SArnaldo Carvalho de Melo if (!time_before(timeout, icsk->icsk_ack.timeout)) 3932463c84b9SArnaldo Carvalho de Melo timeout = icsk->icsk_ack.timeout; 39331da177e4SLinus Torvalds } 3934463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 3935463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.timeout = timeout; 3936463c84b9SArnaldo Carvalho de Melo sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 39371da177e4SLinus Torvalds } 39381da177e4SLinus Torvalds 39391da177e4SLinus Torvalds /* This routine sends an ack and also updates the window. */ 39402987babbSYuchung Cheng void __tcp_send_ack(struct sock *sk, u32 rcv_nxt) 39411da177e4SLinus Torvalds { 39421da177e4SLinus Torvalds struct sk_buff *buff; 39431da177e4SLinus Torvalds 3944058dc334SIlpo Järvinen /* If we have been reset, we may not send again. */ 3945058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3946058dc334SIlpo Järvinen return; 3947058dc334SIlpo Järvinen 39481da177e4SLinus Torvalds /* We are not putting this on the write queue, so 39491da177e4SLinus Torvalds * tcp_transmit_skb() will set the ownership to this 39501da177e4SLinus Torvalds * sock. 39511da177e4SLinus Torvalds */ 39527450aaf6SEric Dumazet buff = alloc_skb(MAX_TCP_HEADER, 39537450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 39547450aaf6SEric Dumazet if (unlikely(!buff)) { 3955a37c2134SEric Dumazet struct inet_connection_sock *icsk = inet_csk(sk); 3956a37c2134SEric Dumazet unsigned long delay; 3957a37c2134SEric Dumazet 3958a37c2134SEric Dumazet delay = TCP_DELACK_MAX << icsk->icsk_ack.retry; 3959a37c2134SEric Dumazet if (delay < TCP_RTO_MAX) 3960a37c2134SEric Dumazet icsk->icsk_ack.retry++; 3961463c84b9SArnaldo Carvalho de Melo inet_csk_schedule_ack(sk); 3962a37c2134SEric Dumazet icsk->icsk_ack.ato = TCP_ATO_MIN; 3963a37c2134SEric Dumazet inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, delay, TCP_RTO_MAX); 39641da177e4SLinus Torvalds return; 39651da177e4SLinus Torvalds } 39661da177e4SLinus Torvalds 39671da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 39681da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 3969a3433f35SChangli Gao tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); 39701da177e4SLinus Torvalds 397198781965SEric Dumazet /* We do not want pure acks influencing TCP Small Queues or fq/pacing 397298781965SEric Dumazet * too much. 397398781965SEric Dumazet * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784 397498781965SEric Dumazet */ 397598781965SEric Dumazet skb_set_tcp_pure_ack(buff); 397698781965SEric Dumazet 39771da177e4SLinus Torvalds /* Send it off, this clears delayed acks for us. */ 39782987babbSYuchung Cheng __tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0, rcv_nxt); 39791da177e4SLinus Torvalds } 398027cde44aSYuchung Cheng EXPORT_SYMBOL_GPL(__tcp_send_ack); 39812987babbSYuchung Cheng 39822987babbSYuchung Cheng void tcp_send_ack(struct sock *sk) 39832987babbSYuchung Cheng { 39842987babbSYuchung Cheng __tcp_send_ack(sk, tcp_sk(sk)->rcv_nxt); 39851da177e4SLinus Torvalds } 39861da177e4SLinus Torvalds 39871da177e4SLinus Torvalds /* This routine sends a packet with an out of date sequence 39881da177e4SLinus Torvalds * number. It assumes the other end will try to ack it. 39891da177e4SLinus Torvalds * 39901da177e4SLinus Torvalds * Question: what should we make while urgent mode? 39911da177e4SLinus Torvalds * 4.4BSD forces sending single byte of data. We cannot send 39921da177e4SLinus Torvalds * out of window data, because we have SND.NXT==SND.MAX... 39931da177e4SLinus Torvalds * 39941da177e4SLinus Torvalds * Current solution: to send TWO zero-length segments in urgent mode: 39951da177e4SLinus Torvalds * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 39961da177e4SLinus Torvalds * out-of-date with SND.UNA-1 to probe window. 39971da177e4SLinus Torvalds */ 3998e520af48SEric Dumazet static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib) 39991da177e4SLinus Torvalds { 40001da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 40011da177e4SLinus Torvalds struct sk_buff *skb; 40021da177e4SLinus Torvalds 40031da177e4SLinus Torvalds /* We don't queue it, tcp_transmit_skb() sets ownership. */ 40047450aaf6SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, 40057450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 400651456b29SIan Morris if (!skb) 40071da177e4SLinus Torvalds return -1; 40081da177e4SLinus Torvalds 40091da177e4SLinus Torvalds /* Reserve space for headers and set control bits. */ 40101da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 40111da177e4SLinus Torvalds /* Use a previous sequence. This should cause the other 40121da177e4SLinus Torvalds * end to send an ack. Don't queue or clone SKB, just 40131da177e4SLinus Torvalds * send it. 40141da177e4SLinus Torvalds */ 4015a3433f35SChangli Gao tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); 4016e2e8009fSRenato Westphal NET_INC_STATS(sock_net(sk), mib); 40177450aaf6SEric Dumazet return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0); 40181da177e4SLinus Torvalds } 40191da177e4SLinus Torvalds 4020385e2070SEric Dumazet /* Called from setsockopt( ... TCP_REPAIR ) */ 4021ee995283SPavel Emelyanov void tcp_send_window_probe(struct sock *sk) 4022ee995283SPavel Emelyanov { 4023ee995283SPavel Emelyanov if (sk->sk_state == TCP_ESTABLISHED) { 4024ee995283SPavel Emelyanov tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; 40259a568de4SEric Dumazet tcp_mstamp_refresh(tcp_sk(sk)); 4026e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE); 4027ee995283SPavel Emelyanov } 4028ee995283SPavel Emelyanov } 4029ee995283SPavel Emelyanov 403067edfef7SAndi Kleen /* Initiate keepalive or window probe from timer. */ 4031e520af48SEric Dumazet int tcp_write_wakeup(struct sock *sk, int mib) 40321da177e4SLinus Torvalds { 40331da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 40341da177e4SLinus Torvalds struct sk_buff *skb; 40351da177e4SLinus Torvalds 4036058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 4037058dc334SIlpo Järvinen return -1; 4038058dc334SIlpo Järvinen 403900db4124SIan Morris skb = tcp_send_head(sk); 404000db4124SIan Morris if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { 40411da177e4SLinus Torvalds int err; 40420c54b85fSIlpo Järvinen unsigned int mss = tcp_current_mss(sk); 404390840defSIlpo Järvinen unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 40441da177e4SLinus Torvalds 40451da177e4SLinus Torvalds if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 40461da177e4SLinus Torvalds tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 40471da177e4SLinus Torvalds 40481da177e4SLinus Torvalds /* We are probing the opening of a window 40491da177e4SLinus Torvalds * but the window size is != 0 40501da177e4SLinus Torvalds * must have been a result SWS avoidance ( sender ) 40511da177e4SLinus Torvalds */ 40521da177e4SLinus Torvalds if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 40531da177e4SLinus Torvalds skb->len > mss) { 40541da177e4SLinus Torvalds seg_size = min(seg_size, mss); 40554de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 405675c119afSEric Dumazet if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE, 405775c119afSEric Dumazet skb, seg_size, mss, GFP_ATOMIC)) 40581da177e4SLinus Torvalds return -1; 40591da177e4SLinus Torvalds } else if (!tcp_skb_pcount(skb)) 40605bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss); 40611da177e4SLinus Torvalds 40624de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 4063dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 406466f5fe62SIlpo Järvinen if (!err) 406566f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 40661da177e4SLinus Torvalds return err; 40671da177e4SLinus Torvalds } else { 406833f5f57eSIlpo Järvinen if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) 4069e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 1, mib); 4070e520af48SEric Dumazet return tcp_xmit_probe_skb(sk, 0, mib); 40711da177e4SLinus Torvalds } 40721da177e4SLinus Torvalds } 40731da177e4SLinus Torvalds 40741da177e4SLinus Torvalds /* A window probe timeout has occurred. If window is not closed send 40751da177e4SLinus Torvalds * a partial packet else a zero probe. 40761da177e4SLinus Torvalds */ 40771da177e4SLinus Torvalds void tcp_send_probe0(struct sock *sk) 40781da177e4SLinus Torvalds { 4079463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 40801da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 4081c6214a97SNikolay Borisov struct net *net = sock_net(sk); 4082c1d5674fSYuchung Cheng unsigned long timeout; 40831da177e4SLinus Torvalds int err; 40841da177e4SLinus Torvalds 4085e520af48SEric Dumazet err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE); 40861da177e4SLinus Torvalds 408775c119afSEric Dumazet if (tp->packets_out || tcp_write_queue_empty(sk)) { 40881da177e4SLinus Torvalds /* Cancel probe timer, if it is not required. */ 40896687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 0; 4090463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff = 0; 40919d9b1ee0SEnke Chen icsk->icsk_probes_tstamp = 0; 40921da177e4SLinus Torvalds return; 40931da177e4SLinus Torvalds } 40941da177e4SLinus Torvalds 4095c1d5674fSYuchung Cheng icsk->icsk_probes_out++; 40961da177e4SLinus Torvalds if (err <= 0) { 4097c6214a97SNikolay Borisov if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2) 4098463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff++; 4099c1d5674fSYuchung Cheng timeout = tcp_probe0_when(sk, TCP_RTO_MAX); 41001da177e4SLinus Torvalds } else { 41011da177e4SLinus Torvalds /* If packet was not sent due to local congestion, 4102c1d5674fSYuchung Cheng * Let senders fight for local resources conservatively. 41031da177e4SLinus Torvalds */ 4104c1d5674fSYuchung Cheng timeout = TCP_RESOURCE_PROBE_INTERVAL; 41051da177e4SLinus Torvalds } 4106344db93aSEnke Chen 4107344db93aSEnke Chen timeout = tcp_clamp_probe0_to_user_timeout(sk, timeout); 41088dc242adSEric Dumazet tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, timeout, TCP_RTO_MAX); 41091da177e4SLinus Torvalds } 41105db92c99SOctavian Purdila 4111ea3bea3aSEric Dumazet int tcp_rtx_synack(const struct sock *sk, struct request_sock *req) 41125db92c99SOctavian Purdila { 41135db92c99SOctavian Purdila const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific; 41145db92c99SOctavian Purdila struct flowi fl; 41155db92c99SOctavian Purdila int res; 41165db92c99SOctavian Purdila 411758d607d3SEric Dumazet tcp_rsk(req)->txhash = net_tx_rndhash(); 4118331fca43SMartin KaFai Lau res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL, 4119331fca43SMartin KaFai Lau NULL); 41205db92c99SOctavian Purdila if (!res) { 412190bbcc60SEric Dumazet __TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); 412202a1d6e7SEric Dumazet __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 41237e32b443SYuchung Cheng if (unlikely(tcp_passive_fastopen(sk))) 41247e32b443SYuchung Cheng tcp_sk(sk)->total_retrans++; 4125cf34ce3dSSong Liu trace_tcp_retransmit_synack(sk, req); 41265db92c99SOctavian Purdila } 41275db92c99SOctavian Purdila return res; 41285db92c99SOctavian Purdila } 41295db92c99SOctavian Purdila EXPORT_SYMBOL(tcp_rtx_synack); 4130