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 448eda7acddSPeter Krystad static void mptcp_options_write(__be32 *ptr, struct tcp_out_options *opts) 449eda7acddSPeter Krystad { 450eda7acddSPeter Krystad #if IS_ENABLED(CONFIG_MPTCP) 451eda7acddSPeter Krystad if (unlikely(OPTION_MPTCP & opts->options)) 452eda7acddSPeter Krystad mptcp_write_options(ptr, &opts->mptcp); 453eda7acddSPeter Krystad #endif 454eda7acddSPeter Krystad } 455eda7acddSPeter Krystad 456331fca43SMartin KaFai Lau #ifdef CONFIG_CGROUP_BPF 457*0813a841SMartin KaFai Lau static int bpf_skops_write_hdr_opt_arg0(struct sk_buff *skb, 458*0813a841SMartin KaFai Lau enum tcp_synack_type synack_type) 459*0813a841SMartin KaFai Lau { 460*0813a841SMartin KaFai Lau if (unlikely(!skb)) 461*0813a841SMartin KaFai Lau return BPF_WRITE_HDR_TCP_CURRENT_MSS; 462*0813a841SMartin KaFai Lau 463*0813a841SMartin KaFai Lau if (unlikely(synack_type == TCP_SYNACK_COOKIE)) 464*0813a841SMartin KaFai Lau return BPF_WRITE_HDR_TCP_SYNACK_COOKIE; 465*0813a841SMartin KaFai Lau 466*0813a841SMartin KaFai Lau return 0; 467*0813a841SMartin KaFai Lau } 468*0813a841SMartin KaFai Lau 469331fca43SMartin KaFai Lau /* req, syn_skb and synack_type are used when writing synack */ 470331fca43SMartin KaFai Lau static void bpf_skops_hdr_opt_len(struct sock *sk, struct sk_buff *skb, 471331fca43SMartin KaFai Lau struct request_sock *req, 472331fca43SMartin KaFai Lau struct sk_buff *syn_skb, 473331fca43SMartin KaFai Lau enum tcp_synack_type synack_type, 474331fca43SMartin KaFai Lau struct tcp_out_options *opts, 475331fca43SMartin KaFai Lau unsigned int *remaining) 476331fca43SMartin KaFai Lau { 477*0813a841SMartin KaFai Lau struct bpf_sock_ops_kern sock_ops; 478*0813a841SMartin KaFai Lau int err; 479*0813a841SMartin KaFai Lau 480331fca43SMartin KaFai Lau if (likely(!BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), 481331fca43SMartin KaFai Lau BPF_SOCK_OPS_WRITE_HDR_OPT_CB_FLAG)) || 482331fca43SMartin KaFai Lau !*remaining) 483331fca43SMartin KaFai Lau return; 484331fca43SMartin KaFai Lau 485*0813a841SMartin KaFai Lau /* *remaining has already been aligned to 4 bytes, so *remaining >= 4 */ 486*0813a841SMartin KaFai Lau 487*0813a841SMartin KaFai Lau /* init sock_ops */ 488*0813a841SMartin KaFai Lau memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp)); 489*0813a841SMartin KaFai Lau 490*0813a841SMartin KaFai Lau sock_ops.op = BPF_SOCK_OPS_HDR_OPT_LEN_CB; 491*0813a841SMartin KaFai Lau 492*0813a841SMartin KaFai Lau if (req) { 493*0813a841SMartin KaFai Lau /* The listen "sk" cannot be passed here because 494*0813a841SMartin KaFai Lau * it is not locked. It would not make too much 495*0813a841SMartin KaFai Lau * sense to do bpf_setsockopt(listen_sk) based 496*0813a841SMartin KaFai Lau * on individual connection request also. 497*0813a841SMartin KaFai Lau * 498*0813a841SMartin KaFai Lau * Thus, "req" is passed here and the cgroup-bpf-progs 499*0813a841SMartin KaFai Lau * of the listen "sk" will be run. 500*0813a841SMartin KaFai Lau * 501*0813a841SMartin KaFai Lau * "req" is also used here for fastopen even the "sk" here is 502*0813a841SMartin KaFai Lau * a fullsock "child" sk. It is to keep the behavior 503*0813a841SMartin KaFai Lau * consistent between fastopen and non-fastopen on 504*0813a841SMartin KaFai Lau * the bpf programming side. 505331fca43SMartin KaFai Lau */ 506*0813a841SMartin KaFai Lau sock_ops.sk = (struct sock *)req; 507*0813a841SMartin KaFai Lau sock_ops.syn_skb = syn_skb; 508*0813a841SMartin KaFai Lau } else { 509*0813a841SMartin KaFai Lau sock_owned_by_me(sk); 510*0813a841SMartin KaFai Lau 511*0813a841SMartin KaFai Lau sock_ops.is_fullsock = 1; 512*0813a841SMartin KaFai Lau sock_ops.sk = sk; 513*0813a841SMartin KaFai Lau } 514*0813a841SMartin KaFai Lau 515*0813a841SMartin KaFai Lau sock_ops.args[0] = bpf_skops_write_hdr_opt_arg0(skb, synack_type); 516*0813a841SMartin KaFai Lau sock_ops.remaining_opt_len = *remaining; 517*0813a841SMartin KaFai Lau /* tcp_current_mss() does not pass a skb */ 518*0813a841SMartin KaFai Lau if (skb) 519*0813a841SMartin KaFai Lau bpf_skops_init_skb(&sock_ops, skb, 0); 520*0813a841SMartin KaFai Lau 521*0813a841SMartin KaFai Lau err = BPF_CGROUP_RUN_PROG_SOCK_OPS_SK(&sock_ops, sk); 522*0813a841SMartin KaFai Lau 523*0813a841SMartin KaFai Lau if (err || sock_ops.remaining_opt_len == *remaining) 524*0813a841SMartin KaFai Lau return; 525*0813a841SMartin KaFai Lau 526*0813a841SMartin KaFai Lau opts->bpf_opt_len = *remaining - sock_ops.remaining_opt_len; 527*0813a841SMartin KaFai Lau /* round up to 4 bytes */ 528*0813a841SMartin KaFai Lau opts->bpf_opt_len = (opts->bpf_opt_len + 3) & ~3; 529*0813a841SMartin KaFai Lau 530*0813a841SMartin KaFai Lau *remaining -= opts->bpf_opt_len; 531331fca43SMartin KaFai Lau } 532331fca43SMartin KaFai Lau 533331fca43SMartin KaFai Lau static void bpf_skops_write_hdr_opt(struct sock *sk, struct sk_buff *skb, 534331fca43SMartin KaFai Lau struct request_sock *req, 535331fca43SMartin KaFai Lau struct sk_buff *syn_skb, 536331fca43SMartin KaFai Lau enum tcp_synack_type synack_type, 537331fca43SMartin KaFai Lau struct tcp_out_options *opts) 538331fca43SMartin KaFai Lau { 539*0813a841SMartin KaFai Lau u8 first_opt_off, nr_written, max_opt_len = opts->bpf_opt_len; 540*0813a841SMartin KaFai Lau struct bpf_sock_ops_kern sock_ops; 541*0813a841SMartin KaFai Lau int err; 542*0813a841SMartin KaFai Lau 543*0813a841SMartin KaFai Lau if (likely(!max_opt_len)) 544331fca43SMartin KaFai Lau return; 545331fca43SMartin KaFai Lau 546*0813a841SMartin KaFai Lau memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp)); 547*0813a841SMartin KaFai Lau 548*0813a841SMartin KaFai Lau sock_ops.op = BPF_SOCK_OPS_WRITE_HDR_OPT_CB; 549*0813a841SMartin KaFai Lau 550*0813a841SMartin KaFai Lau if (req) { 551*0813a841SMartin KaFai Lau sock_ops.sk = (struct sock *)req; 552*0813a841SMartin KaFai Lau sock_ops.syn_skb = syn_skb; 553*0813a841SMartin KaFai Lau } else { 554*0813a841SMartin KaFai Lau sock_owned_by_me(sk); 555*0813a841SMartin KaFai Lau 556*0813a841SMartin KaFai Lau sock_ops.is_fullsock = 1; 557*0813a841SMartin KaFai Lau sock_ops.sk = sk; 558*0813a841SMartin KaFai Lau } 559*0813a841SMartin KaFai Lau 560*0813a841SMartin KaFai Lau sock_ops.args[0] = bpf_skops_write_hdr_opt_arg0(skb, synack_type); 561*0813a841SMartin KaFai Lau sock_ops.remaining_opt_len = max_opt_len; 562*0813a841SMartin KaFai Lau first_opt_off = tcp_hdrlen(skb) - max_opt_len; 563*0813a841SMartin KaFai Lau bpf_skops_init_skb(&sock_ops, skb, first_opt_off); 564*0813a841SMartin KaFai Lau 565*0813a841SMartin KaFai Lau err = BPF_CGROUP_RUN_PROG_SOCK_OPS_SK(&sock_ops, sk); 566*0813a841SMartin KaFai Lau 567*0813a841SMartin KaFai Lau if (err) 568*0813a841SMartin KaFai Lau nr_written = 0; 569*0813a841SMartin KaFai Lau else 570*0813a841SMartin KaFai Lau nr_written = max_opt_len - sock_ops.remaining_opt_len; 571*0813a841SMartin KaFai Lau 572*0813a841SMartin KaFai Lau if (nr_written < max_opt_len) 573*0813a841SMartin KaFai Lau memset(skb->data + first_opt_off + nr_written, TCPOPT_NOP, 574*0813a841SMartin KaFai Lau max_opt_len - nr_written); 575331fca43SMartin KaFai Lau } 576331fca43SMartin KaFai Lau #else 577331fca43SMartin KaFai Lau static void bpf_skops_hdr_opt_len(struct sock *sk, struct sk_buff *skb, 578331fca43SMartin KaFai Lau struct request_sock *req, 579331fca43SMartin KaFai Lau struct sk_buff *syn_skb, 580331fca43SMartin KaFai Lau enum tcp_synack_type synack_type, 581331fca43SMartin KaFai Lau struct tcp_out_options *opts, 582331fca43SMartin KaFai Lau unsigned int *remaining) 583331fca43SMartin KaFai Lau { 584331fca43SMartin KaFai Lau } 585331fca43SMartin KaFai Lau 586331fca43SMartin KaFai Lau static void bpf_skops_write_hdr_opt(struct sock *sk, struct sk_buff *skb, 587331fca43SMartin KaFai Lau struct request_sock *req, 588331fca43SMartin KaFai Lau struct sk_buff *syn_skb, 589331fca43SMartin KaFai Lau enum tcp_synack_type synack_type, 590331fca43SMartin KaFai Lau struct tcp_out_options *opts) 591331fca43SMartin KaFai Lau { 592331fca43SMartin KaFai Lau } 593331fca43SMartin KaFai Lau #endif 594331fca43SMartin KaFai Lau 59567edfef7SAndi Kleen /* Write previously computed TCP options to the packet. 59667edfef7SAndi Kleen * 59767edfef7SAndi Kleen * Beware: Something in the Internet is very sensitive to the ordering of 598fd6149d3SIlpo Järvinen * TCP options, we learned this through the hard way, so be careful here. 599fd6149d3SIlpo Järvinen * Luckily we can at least blame others for their non-compliance but from 6008e3bff96Sstephen hemminger * inter-operability perspective it seems that we're somewhat stuck with 601fd6149d3SIlpo Järvinen * the ordering which we have been using if we want to keep working with 602fd6149d3SIlpo Järvinen * those broken things (not that it currently hurts anybody as there isn't 603fd6149d3SIlpo Järvinen * particular reason why the ordering would need to be changed). 604fd6149d3SIlpo Järvinen * 605fd6149d3SIlpo Järvinen * At least SACK_PERM as the first option is known to lead to a disaster 606fd6149d3SIlpo Järvinen * (but it may well be that other scenarios fail similarly). 607fd6149d3SIlpo Järvinen */ 60833ad798cSAdam Langley static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp, 609bd0388aeSWilliam Allen Simpson struct tcp_out_options *opts) 610bd0388aeSWilliam Allen Simpson { 6112100c8d2SYuchung Cheng u16 options = opts->options; /* mungable copy */ 612bd0388aeSWilliam Allen Simpson 613bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_MD5 & options)) { 6141a2c6181SChristoph Paasch *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | 6151a2c6181SChristoph Paasch (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG); 616bd0388aeSWilliam Allen Simpson /* overload cookie hash location */ 617bd0388aeSWilliam Allen Simpson opts->hash_location = (__u8 *)ptr; 61833ad798cSAdam Langley ptr += 4; 61933ad798cSAdam Langley } 62033ad798cSAdam Langley 621fd6149d3SIlpo Järvinen if (unlikely(opts->mss)) { 622fd6149d3SIlpo Järvinen *ptr++ = htonl((TCPOPT_MSS << 24) | 623fd6149d3SIlpo Järvinen (TCPOLEN_MSS << 16) | 624fd6149d3SIlpo Järvinen opts->mss); 625fd6149d3SIlpo Järvinen } 626fd6149d3SIlpo Järvinen 627bd0388aeSWilliam Allen Simpson if (likely(OPTION_TS & options)) { 628bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 62933ad798cSAdam Langley *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | 63033ad798cSAdam Langley (TCPOLEN_SACK_PERM << 16) | 63133ad798cSAdam Langley (TCPOPT_TIMESTAMP << 8) | 63233ad798cSAdam Langley TCPOLEN_TIMESTAMP); 633bd0388aeSWilliam Allen Simpson options &= ~OPTION_SACK_ADVERTISE; 63433ad798cSAdam Langley } else { 635496c98dfSYOSHIFUJI Hideaki *ptr++ = htonl((TCPOPT_NOP << 24) | 63640efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 63740efc6faSStephen Hemminger (TCPOPT_TIMESTAMP << 8) | 63840efc6faSStephen Hemminger TCPOLEN_TIMESTAMP); 63940efc6faSStephen Hemminger } 64033ad798cSAdam Langley *ptr++ = htonl(opts->tsval); 64133ad798cSAdam Langley *ptr++ = htonl(opts->tsecr); 64233ad798cSAdam Langley } 64333ad798cSAdam Langley 644bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 64533ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 64633ad798cSAdam Langley (TCPOPT_NOP << 16) | 64733ad798cSAdam Langley (TCPOPT_SACK_PERM << 8) | 64833ad798cSAdam Langley TCPOLEN_SACK_PERM); 64933ad798cSAdam Langley } 65033ad798cSAdam Langley 651bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_WSCALE & options)) { 65233ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 65333ad798cSAdam Langley (TCPOPT_WINDOW << 16) | 65433ad798cSAdam Langley (TCPOLEN_WINDOW << 8) | 65533ad798cSAdam Langley opts->ws); 65633ad798cSAdam Langley } 65733ad798cSAdam Langley 65833ad798cSAdam Langley if (unlikely(opts->num_sack_blocks)) { 65933ad798cSAdam Langley struct tcp_sack_block *sp = tp->rx_opt.dsack ? 66033ad798cSAdam Langley tp->duplicate_sack : tp->selective_acks; 66140efc6faSStephen Hemminger int this_sack; 66240efc6faSStephen Hemminger 66340efc6faSStephen Hemminger *ptr++ = htonl((TCPOPT_NOP << 24) | 66440efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 66540efc6faSStephen Hemminger (TCPOPT_SACK << 8) | 66633ad798cSAdam Langley (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * 66740efc6faSStephen Hemminger TCPOLEN_SACK_PERBLOCK))); 6682de979bdSStephen Hemminger 66933ad798cSAdam Langley for (this_sack = 0; this_sack < opts->num_sack_blocks; 67033ad798cSAdam Langley ++this_sack) { 67140efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].start_seq); 67240efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].end_seq); 67340efc6faSStephen Hemminger } 6742de979bdSStephen Hemminger 67540efc6faSStephen Hemminger tp->rx_opt.dsack = 0; 67640efc6faSStephen Hemminger } 6772100c8d2SYuchung Cheng 6782100c8d2SYuchung Cheng if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) { 6792100c8d2SYuchung Cheng struct tcp_fastopen_cookie *foc = opts->fastopen_cookie; 6807f9b838bSDaniel Lee u8 *p = (u8 *)ptr; 6817f9b838bSDaniel Lee u32 len; /* Fast Open option length */ 6822100c8d2SYuchung Cheng 6837f9b838bSDaniel Lee if (foc->exp) { 6847f9b838bSDaniel Lee len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len; 6857f9b838bSDaniel Lee *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) | 6862100c8d2SYuchung Cheng TCPOPT_FASTOPEN_MAGIC); 6877f9b838bSDaniel Lee p += TCPOLEN_EXP_FASTOPEN_BASE; 6887f9b838bSDaniel Lee } else { 6897f9b838bSDaniel Lee len = TCPOLEN_FASTOPEN_BASE + foc->len; 6907f9b838bSDaniel Lee *p++ = TCPOPT_FASTOPEN; 6917f9b838bSDaniel Lee *p++ = len; 6922100c8d2SYuchung Cheng } 6937f9b838bSDaniel Lee 6947f9b838bSDaniel Lee memcpy(p, foc->val, foc->len); 6957f9b838bSDaniel Lee if ((len & 3) == 2) { 6967f9b838bSDaniel Lee p[foc->len] = TCPOPT_NOP; 6977f9b838bSDaniel Lee p[foc->len + 1] = TCPOPT_NOP; 6987f9b838bSDaniel Lee } 6997f9b838bSDaniel Lee ptr += (len + 3) >> 2; 7002100c8d2SYuchung Cheng } 70160e2a778SUrsula Braun 70260e2a778SUrsula Braun smc_options_write(ptr, &options); 703eda7acddSPeter Krystad 704eda7acddSPeter Krystad mptcp_options_write(ptr, opts); 70560e2a778SUrsula Braun } 70660e2a778SUrsula Braun 70760e2a778SUrsula Braun static void smc_set_option(const struct tcp_sock *tp, 70860e2a778SUrsula Braun struct tcp_out_options *opts, 70960e2a778SUrsula Braun unsigned int *remaining) 71060e2a778SUrsula Braun { 71160e2a778SUrsula Braun #if IS_ENABLED(CONFIG_SMC) 71260e2a778SUrsula Braun if (static_branch_unlikely(&tcp_have_smc)) { 71360e2a778SUrsula Braun if (tp->syn_smc) { 71460e2a778SUrsula Braun if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) { 71560e2a778SUrsula Braun opts->options |= OPTION_SMC; 71660e2a778SUrsula Braun *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED; 71760e2a778SUrsula Braun } 71860e2a778SUrsula Braun } 71960e2a778SUrsula Braun } 72060e2a778SUrsula Braun #endif 72160e2a778SUrsula Braun } 72260e2a778SUrsula Braun 72360e2a778SUrsula Braun static void smc_set_option_cond(const struct tcp_sock *tp, 72460e2a778SUrsula Braun const struct inet_request_sock *ireq, 72560e2a778SUrsula Braun struct tcp_out_options *opts, 72660e2a778SUrsula Braun unsigned int *remaining) 72760e2a778SUrsula Braun { 72860e2a778SUrsula Braun #if IS_ENABLED(CONFIG_SMC) 72960e2a778SUrsula Braun if (static_branch_unlikely(&tcp_have_smc)) { 73060e2a778SUrsula Braun if (tp->syn_smc && ireq->smc_ok) { 73160e2a778SUrsula Braun if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) { 73260e2a778SUrsula Braun opts->options |= OPTION_SMC; 73360e2a778SUrsula Braun *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED; 73460e2a778SUrsula Braun } 73560e2a778SUrsula Braun } 73660e2a778SUrsula Braun } 73760e2a778SUrsula Braun #endif 73840efc6faSStephen Hemminger } 73940efc6faSStephen Hemminger 740cec37a6eSPeter Krystad static void mptcp_set_option_cond(const struct request_sock *req, 741cec37a6eSPeter Krystad struct tcp_out_options *opts, 742cec37a6eSPeter Krystad unsigned int *remaining) 743cec37a6eSPeter Krystad { 744cec37a6eSPeter Krystad if (rsk_is_mptcp(req)) { 745cec37a6eSPeter Krystad unsigned int size; 746cec37a6eSPeter Krystad 747cec37a6eSPeter Krystad if (mptcp_synack_options(req, &size, &opts->mptcp)) { 748cec37a6eSPeter Krystad if (*remaining >= size) { 749cec37a6eSPeter Krystad opts->options |= OPTION_MPTCP; 750cec37a6eSPeter Krystad *remaining -= size; 751cec37a6eSPeter Krystad } 752cec37a6eSPeter Krystad } 753cec37a6eSPeter Krystad } 754cec37a6eSPeter Krystad } 755cec37a6eSPeter Krystad 75667edfef7SAndi Kleen /* Compute TCP options for SYN packets. This is not the final 75767edfef7SAndi Kleen * network wire format yet. 75867edfef7SAndi Kleen */ 75995c96174SEric Dumazet static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb, 76033ad798cSAdam Langley struct tcp_out_options *opts, 761cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 762cf533ea5SEric Dumazet { 76333ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 76495c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 765783237e8SYuchung Cheng struct tcp_fastopen_request *fastopen = tp->fastopen_req; 76633ad798cSAdam Langley 7678c2320e8SEric Dumazet *md5 = NULL; 768cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 769921f9a0fSEric Dumazet if (static_branch_unlikely(&tcp_md5_needed) && 7706015c71eSEric Dumazet rcu_access_pointer(tp->md5sig_info)) { 77133ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 77233ad798cSAdam Langley if (*md5) { 77333ad798cSAdam Langley opts->options |= OPTION_MD5; 774bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 775cfb6eeb4SYOSHIFUJI Hideaki } 7768c2320e8SEric Dumazet } 777cfb6eeb4SYOSHIFUJI Hideaki #endif 77833ad798cSAdam Langley 77933ad798cSAdam Langley /* We always get an MSS option. The option bytes which will be seen in 78033ad798cSAdam Langley * normal data packets should timestamps be used, must be in the MSS 78133ad798cSAdam Langley * advertised. But we subtract them from tp->mss_cache so that 78233ad798cSAdam Langley * calculations in tcp_sendmsg are simpler etc. So account for this 78333ad798cSAdam Langley * fact here if necessary. If we don't do this correctly, as a 78433ad798cSAdam Langley * receiver we won't recognize data packets as being full sized when we 78533ad798cSAdam Langley * should, and thus we won't abide by the delayed ACK rules correctly. 78633ad798cSAdam Langley * SACKs don't matter, we never delay an ACK when we have any of those 78733ad798cSAdam Langley * going out. */ 78833ad798cSAdam Langley opts->mss = tcp_advertise_mss(sk); 789bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 79033ad798cSAdam Langley 7915d2ed052SEric Dumazet if (likely(sock_net(sk)->ipv4.sysctl_tcp_timestamps && !*md5)) { 79233ad798cSAdam Langley opts->options |= OPTION_TS; 7937faee5c0SEric Dumazet opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset; 79433ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 795bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 79633ad798cSAdam Langley } 7979bb37ef0SEric Dumazet if (likely(sock_net(sk)->ipv4.sysctl_tcp_window_scaling)) { 79833ad798cSAdam Langley opts->ws = tp->rx_opt.rcv_wscale; 79989e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 800bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 80133ad798cSAdam Langley } 802f9301034SEric Dumazet if (likely(sock_net(sk)->ipv4.sysctl_tcp_sack)) { 80333ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 804b32d1310SDavid S. Miller if (unlikely(!(OPTION_TS & opts->options))) 805bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 80633ad798cSAdam Langley } 80733ad798cSAdam Langley 808783237e8SYuchung Cheng if (fastopen && fastopen->cookie.len >= 0) { 8092646c831SDaniel Lee u32 need = fastopen->cookie.len; 8102646c831SDaniel Lee 8112646c831SDaniel Lee need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE : 8122646c831SDaniel Lee TCPOLEN_FASTOPEN_BASE; 813783237e8SYuchung Cheng need = (need + 3) & ~3U; /* Align to 32 bits */ 814783237e8SYuchung Cheng if (remaining >= need) { 815783237e8SYuchung Cheng opts->options |= OPTION_FAST_OPEN_COOKIE; 816783237e8SYuchung Cheng opts->fastopen_cookie = &fastopen->cookie; 817783237e8SYuchung Cheng remaining -= need; 818783237e8SYuchung Cheng tp->syn_fastopen = 1; 8192646c831SDaniel Lee tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0; 820783237e8SYuchung Cheng } 821783237e8SYuchung Cheng } 822bd0388aeSWilliam Allen Simpson 82360e2a778SUrsula Braun smc_set_option(tp, opts, &remaining); 82460e2a778SUrsula Braun 825cec37a6eSPeter Krystad if (sk_is_mptcp(sk)) { 826cec37a6eSPeter Krystad unsigned int size; 827cec37a6eSPeter Krystad 828cc7972eaSChristoph Paasch if (mptcp_syn_options(sk, skb, &size, &opts->mptcp)) { 829cec37a6eSPeter Krystad opts->options |= OPTION_MPTCP; 830cec37a6eSPeter Krystad remaining -= size; 831cec37a6eSPeter Krystad } 832cec37a6eSPeter Krystad } 833cec37a6eSPeter Krystad 834331fca43SMartin KaFai Lau bpf_skops_hdr_opt_len(sk, skb, NULL, NULL, 0, opts, &remaining); 835331fca43SMartin KaFai Lau 836bd0388aeSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 83733ad798cSAdam Langley } 83833ad798cSAdam Langley 83967edfef7SAndi Kleen /* Set up TCP options for SYN-ACKs. */ 84060e2a778SUrsula Braun static unsigned int tcp_synack_options(const struct sock *sk, 84160e2a778SUrsula Braun struct request_sock *req, 84295c96174SEric Dumazet unsigned int mss, struct sk_buff *skb, 84333ad798cSAdam Langley struct tcp_out_options *opts, 84480f03e27SEric Dumazet const struct tcp_md5sig_key *md5, 845e114e1e8SEric Dumazet struct tcp_fastopen_cookie *foc, 846331fca43SMartin KaFai Lau enum tcp_synack_type synack_type, 847331fca43SMartin KaFai Lau struct sk_buff *syn_skb) 8484957faadSWilliam Allen Simpson { 84933ad798cSAdam Langley struct inet_request_sock *ireq = inet_rsk(req); 85095c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 85133ad798cSAdam Langley 85233ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 85380f03e27SEric Dumazet if (md5) { 85433ad798cSAdam Langley opts->options |= OPTION_MD5; 8554957faadSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 8564957faadSWilliam Allen Simpson 8574957faadSWilliam Allen Simpson /* We can't fit any SACK blocks in a packet with MD5 + TS 8584957faadSWilliam Allen Simpson * options. There was discussion about disabling SACK 8594957faadSWilliam Allen Simpson * rather than TS in order to fit in better with old, 8604957faadSWilliam Allen Simpson * buggy kernels, but that was deemed to be unnecessary. 8614957faadSWilliam Allen Simpson */ 862e114e1e8SEric Dumazet if (synack_type != TCP_SYNACK_COOKIE) 863de213e5eSEric Dumazet ireq->tstamp_ok &= !ireq->sack_ok; 86433ad798cSAdam Langley } 86533ad798cSAdam Langley #endif 86633ad798cSAdam Langley 8674957faadSWilliam Allen Simpson /* We always send an MSS option. */ 86833ad798cSAdam Langley opts->mss = mss; 8694957faadSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 87033ad798cSAdam Langley 87133ad798cSAdam Langley if (likely(ireq->wscale_ok)) { 87233ad798cSAdam Langley opts->ws = ireq->rcv_wscale; 87389e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 8744957faadSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 87533ad798cSAdam Langley } 876de213e5eSEric Dumazet if (likely(ireq->tstamp_ok)) { 87733ad798cSAdam Langley opts->options |= OPTION_TS; 87895a22caeSFlorian Westphal opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off; 87933ad798cSAdam Langley opts->tsecr = req->ts_recent; 8804957faadSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 88133ad798cSAdam Langley } 88233ad798cSAdam Langley if (likely(ireq->sack_ok)) { 88333ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 884de213e5eSEric Dumazet if (unlikely(!ireq->tstamp_ok)) 8854957faadSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 88633ad798cSAdam Langley } 8877f9b838bSDaniel Lee if (foc != NULL && foc->len >= 0) { 8887f9b838bSDaniel Lee u32 need = foc->len; 8897f9b838bSDaniel Lee 8907f9b838bSDaniel Lee need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE : 8917f9b838bSDaniel Lee TCPOLEN_FASTOPEN_BASE; 8928336886fSJerry Chu need = (need + 3) & ~3U; /* Align to 32 bits */ 8938336886fSJerry Chu if (remaining >= need) { 8948336886fSJerry Chu opts->options |= OPTION_FAST_OPEN_COOKIE; 8958336886fSJerry Chu opts->fastopen_cookie = foc; 8968336886fSJerry Chu remaining -= need; 8978336886fSJerry Chu } 8988336886fSJerry Chu } 8994957faadSWilliam Allen Simpson 900cec37a6eSPeter Krystad mptcp_set_option_cond(req, opts, &remaining); 901cec37a6eSPeter Krystad 90260e2a778SUrsula Braun smc_set_option_cond(tcp_sk(sk), ireq, opts, &remaining); 90360e2a778SUrsula Braun 904331fca43SMartin KaFai Lau bpf_skops_hdr_opt_len((struct sock *)sk, skb, req, syn_skb, 905331fca43SMartin KaFai Lau synack_type, opts, &remaining); 906331fca43SMartin KaFai Lau 9074957faadSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 90833ad798cSAdam Langley } 90933ad798cSAdam Langley 91067edfef7SAndi Kleen /* Compute TCP options for ESTABLISHED sockets. This is not the 91167edfef7SAndi Kleen * final wire format yet. 91267edfef7SAndi Kleen */ 91395c96174SEric Dumazet static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb, 91433ad798cSAdam Langley struct tcp_out_options *opts, 915cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 916cf533ea5SEric Dumazet { 91733ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 91895c96174SEric Dumazet unsigned int size = 0; 919cabeccbdSIlpo Järvinen unsigned int eff_sacks; 92033ad798cSAdam Langley 9215843ef42SAndi Kleen opts->options = 0; 9225843ef42SAndi Kleen 9238c2320e8SEric Dumazet *md5 = NULL; 92433ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 925921f9a0fSEric Dumazet if (static_branch_unlikely(&tcp_md5_needed) && 9266015c71eSEric Dumazet rcu_access_pointer(tp->md5sig_info)) { 92733ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 9288c2320e8SEric Dumazet if (*md5) { 92933ad798cSAdam Langley opts->options |= OPTION_MD5; 93033ad798cSAdam Langley size += TCPOLEN_MD5SIG_ALIGNED; 93133ad798cSAdam Langley } 9328c2320e8SEric Dumazet } 93333ad798cSAdam Langley #endif 93433ad798cSAdam Langley 93533ad798cSAdam Langley if (likely(tp->rx_opt.tstamp_ok)) { 93633ad798cSAdam Langley opts->options |= OPTION_TS; 9377faee5c0SEric Dumazet opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0; 93833ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 93933ad798cSAdam Langley size += TCPOLEN_TSTAMP_ALIGNED; 94033ad798cSAdam Langley } 94133ad798cSAdam Langley 942cec37a6eSPeter Krystad /* MPTCP options have precedence over SACK for the limited TCP 943cec37a6eSPeter Krystad * option space because a MPTCP connection would be forced to 944cec37a6eSPeter Krystad * fall back to regular TCP if a required multipath option is 945cec37a6eSPeter Krystad * missing. SACK still gets a chance to use whatever space is 946cec37a6eSPeter Krystad * left. 947cec37a6eSPeter Krystad */ 948cec37a6eSPeter Krystad if (sk_is_mptcp(sk)) { 949cec37a6eSPeter Krystad unsigned int remaining = MAX_TCP_OPTION_SPACE - size; 950cec37a6eSPeter Krystad unsigned int opt_size = 0; 951cec37a6eSPeter Krystad 952cec37a6eSPeter Krystad if (mptcp_established_options(sk, skb, &opt_size, remaining, 953cec37a6eSPeter Krystad &opts->mptcp)) { 954cec37a6eSPeter Krystad opts->options |= OPTION_MPTCP; 955cec37a6eSPeter Krystad size += opt_size; 956cec37a6eSPeter Krystad } 957cec37a6eSPeter Krystad } 958cec37a6eSPeter Krystad 959cabeccbdSIlpo Järvinen eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; 960cabeccbdSIlpo Järvinen if (unlikely(eff_sacks)) { 96195c96174SEric Dumazet const unsigned int remaining = MAX_TCP_OPTION_SPACE - size; 9629cfcca23SMat Martineau if (unlikely(remaining < TCPOLEN_SACK_BASE_ALIGNED + 9639cfcca23SMat Martineau TCPOLEN_SACK_PERBLOCK)) 9649cfcca23SMat Martineau return size; 9659cfcca23SMat Martineau 96633ad798cSAdam Langley opts->num_sack_blocks = 96795c96174SEric Dumazet min_t(unsigned int, eff_sacks, 96833ad798cSAdam Langley (remaining - TCPOLEN_SACK_BASE_ALIGNED) / 96933ad798cSAdam Langley TCPOLEN_SACK_PERBLOCK); 9709cfcca23SMat Martineau 97133ad798cSAdam Langley size += TCPOLEN_SACK_BASE_ALIGNED + 97233ad798cSAdam Langley opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; 97333ad798cSAdam Langley } 97433ad798cSAdam Langley 975331fca43SMartin KaFai Lau if (unlikely(BPF_SOCK_OPS_TEST_FLAG(tp, 976331fca43SMartin KaFai Lau BPF_SOCK_OPS_WRITE_HDR_OPT_CB_FLAG))) { 977331fca43SMartin KaFai Lau unsigned int remaining = MAX_TCP_OPTION_SPACE - size; 978331fca43SMartin KaFai Lau 979331fca43SMartin KaFai Lau bpf_skops_hdr_opt_len(sk, skb, NULL, NULL, 0, opts, &remaining); 980331fca43SMartin KaFai Lau 981331fca43SMartin KaFai Lau size = MAX_TCP_OPTION_SPACE - remaining; 982331fca43SMartin KaFai Lau } 983331fca43SMartin KaFai Lau 98433ad798cSAdam Langley return size; 98540efc6faSStephen Hemminger } 9861da177e4SLinus Torvalds 98746d3ceabSEric Dumazet 98846d3ceabSEric Dumazet /* TCP SMALL QUEUES (TSQ) 98946d3ceabSEric Dumazet * 99046d3ceabSEric Dumazet * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev) 99146d3ceabSEric Dumazet * to reduce RTT and bufferbloat. 99246d3ceabSEric Dumazet * We do this using a special skb destructor (tcp_wfree). 99346d3ceabSEric Dumazet * 99446d3ceabSEric Dumazet * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb 99546d3ceabSEric Dumazet * needs to be reallocated in a driver. 9968e3bff96Sstephen hemminger * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc 99746d3ceabSEric Dumazet * 99846d3ceabSEric Dumazet * Since transmit from skb destructor is forbidden, we use a tasklet 99946d3ceabSEric Dumazet * to process all sockets that eventually need to send more skbs. 100046d3ceabSEric Dumazet * We use one tasklet per cpu, with its own queue of sockets. 100146d3ceabSEric Dumazet */ 100246d3ceabSEric Dumazet struct tsq_tasklet { 100346d3ceabSEric Dumazet struct tasklet_struct tasklet; 100446d3ceabSEric Dumazet struct list_head head; /* queue of tcp sockets */ 100546d3ceabSEric Dumazet }; 100646d3ceabSEric Dumazet static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet); 100746d3ceabSEric Dumazet 100873a6bab5SEric Dumazet static void tcp_tsq_write(struct sock *sk) 10096f458dfbSEric Dumazet { 10106f458dfbSEric Dumazet if ((1 << sk->sk_state) & 10116f458dfbSEric Dumazet (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING | 1012f9616c35SEric Dumazet TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) { 1013f9616c35SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 1014f9616c35SEric Dumazet 1015f9616c35SEric Dumazet if (tp->lost_out > tp->retrans_out && 10163a91d29fSKoichiro Den tp->snd_cwnd > tcp_packets_in_flight(tp)) { 10173a91d29fSKoichiro Den tcp_mstamp_refresh(tp); 1018f9616c35SEric Dumazet tcp_xmit_retransmit_queue(sk); 10193a91d29fSKoichiro Den } 1020f9616c35SEric Dumazet 1021f9616c35SEric Dumazet tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle, 1022bf06200eSJohn Ogness 0, GFP_ATOMIC); 10236f458dfbSEric Dumazet } 1024f9616c35SEric Dumazet } 102573a6bab5SEric Dumazet 102673a6bab5SEric Dumazet static void tcp_tsq_handler(struct sock *sk) 102773a6bab5SEric Dumazet { 102873a6bab5SEric Dumazet bh_lock_sock(sk); 102973a6bab5SEric Dumazet if (!sock_owned_by_user(sk)) 103073a6bab5SEric Dumazet tcp_tsq_write(sk); 103173a6bab5SEric Dumazet else if (!test_and_set_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags)) 103273a6bab5SEric Dumazet sock_hold(sk); 103373a6bab5SEric Dumazet bh_unlock_sock(sk); 103473a6bab5SEric Dumazet } 103546d3ceabSEric Dumazet /* 10368e3bff96Sstephen hemminger * One tasklet per cpu tries to send more skbs. 103746d3ceabSEric Dumazet * We run in tasklet context but need to disable irqs when 10388e3bff96Sstephen hemminger * transferring tsq->head because tcp_wfree() might 103946d3ceabSEric Dumazet * interrupt us (non NAPI drivers) 104046d3ceabSEric Dumazet */ 104146d3ceabSEric Dumazet static void tcp_tasklet_func(unsigned long data) 104246d3ceabSEric Dumazet { 104346d3ceabSEric Dumazet struct tsq_tasklet *tsq = (struct tsq_tasklet *)data; 104446d3ceabSEric Dumazet LIST_HEAD(list); 104546d3ceabSEric Dumazet unsigned long flags; 104646d3ceabSEric Dumazet struct list_head *q, *n; 104746d3ceabSEric Dumazet struct tcp_sock *tp; 104846d3ceabSEric Dumazet struct sock *sk; 104946d3ceabSEric Dumazet 105046d3ceabSEric Dumazet local_irq_save(flags); 105146d3ceabSEric Dumazet list_splice_init(&tsq->head, &list); 105246d3ceabSEric Dumazet local_irq_restore(flags); 105346d3ceabSEric Dumazet 105446d3ceabSEric Dumazet list_for_each_safe(q, n, &list) { 105546d3ceabSEric Dumazet tp = list_entry(q, struct tcp_sock, tsq_node); 105646d3ceabSEric Dumazet list_del(&tp->tsq_node); 105746d3ceabSEric Dumazet 105846d3ceabSEric Dumazet sk = (struct sock *)tp; 10590a9648f1SEric Dumazet smp_mb__before_atomic(); 10607aa5470cSEric Dumazet clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags); 10617aa5470cSEric Dumazet 10626f458dfbSEric Dumazet tcp_tsq_handler(sk); 106346d3ceabSEric Dumazet sk_free(sk); 106446d3ceabSEric Dumazet } 106546d3ceabSEric Dumazet } 106646d3ceabSEric Dumazet 106740fc3423SEric Dumazet #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \ 106840fc3423SEric Dumazet TCPF_WRITE_TIMER_DEFERRED | \ 106940fc3423SEric Dumazet TCPF_DELACK_TIMER_DEFERRED | \ 107040fc3423SEric Dumazet TCPF_MTU_REDUCED_DEFERRED) 107146d3ceabSEric Dumazet /** 107246d3ceabSEric Dumazet * tcp_release_cb - tcp release_sock() callback 107346d3ceabSEric Dumazet * @sk: socket 107446d3ceabSEric Dumazet * 107546d3ceabSEric Dumazet * called from release_sock() to perform protocol dependent 107646d3ceabSEric Dumazet * actions before socket release. 107746d3ceabSEric Dumazet */ 107846d3ceabSEric Dumazet void tcp_release_cb(struct sock *sk) 107946d3ceabSEric Dumazet { 10806f458dfbSEric Dumazet unsigned long flags, nflags; 108146d3ceabSEric Dumazet 10826f458dfbSEric Dumazet /* perform an atomic operation only if at least one flag is set */ 10836f458dfbSEric Dumazet do { 10847aa5470cSEric Dumazet flags = sk->sk_tsq_flags; 10856f458dfbSEric Dumazet if (!(flags & TCP_DEFERRED_ALL)) 10866f458dfbSEric Dumazet return; 10876f458dfbSEric Dumazet nflags = flags & ~TCP_DEFERRED_ALL; 10887aa5470cSEric Dumazet } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags); 10896f458dfbSEric Dumazet 109073a6bab5SEric Dumazet if (flags & TCPF_TSQ_DEFERRED) { 109173a6bab5SEric Dumazet tcp_tsq_write(sk); 109273a6bab5SEric Dumazet __sock_put(sk); 109373a6bab5SEric Dumazet } 1094c3f9b018SEric Dumazet /* Here begins the tricky part : 1095c3f9b018SEric Dumazet * We are called from release_sock() with : 1096c3f9b018SEric Dumazet * 1) BH disabled 1097c3f9b018SEric Dumazet * 2) sk_lock.slock spinlock held 1098c3f9b018SEric Dumazet * 3) socket owned by us (sk->sk_lock.owned == 1) 1099c3f9b018SEric Dumazet * 1100c3f9b018SEric Dumazet * But following code is meant to be called from BH handlers, 1101c3f9b018SEric Dumazet * so we should keep BH disabled, but early release socket ownership 1102c3f9b018SEric Dumazet */ 1103c3f9b018SEric Dumazet sock_release_ownership(sk); 1104c3f9b018SEric Dumazet 110540fc3423SEric Dumazet if (flags & TCPF_WRITE_TIMER_DEFERRED) { 11066f458dfbSEric Dumazet tcp_write_timer_handler(sk); 1107144d56e9SEric Dumazet __sock_put(sk); 1108144d56e9SEric Dumazet } 110940fc3423SEric Dumazet if (flags & TCPF_DELACK_TIMER_DEFERRED) { 11106f458dfbSEric Dumazet tcp_delack_timer_handler(sk); 1111144d56e9SEric Dumazet __sock_put(sk); 1112144d56e9SEric Dumazet } 111340fc3423SEric Dumazet if (flags & TCPF_MTU_REDUCED_DEFERRED) { 11144fab9071SNeal Cardwell inet_csk(sk)->icsk_af_ops->mtu_reduced(sk); 1115144d56e9SEric Dumazet __sock_put(sk); 1116144d56e9SEric Dumazet } 111746d3ceabSEric Dumazet } 111846d3ceabSEric Dumazet EXPORT_SYMBOL(tcp_release_cb); 111946d3ceabSEric Dumazet 112046d3ceabSEric Dumazet void __init tcp_tasklet_init(void) 112146d3ceabSEric Dumazet { 112246d3ceabSEric Dumazet int i; 112346d3ceabSEric Dumazet 112446d3ceabSEric Dumazet for_each_possible_cpu(i) { 112546d3ceabSEric Dumazet struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i); 112646d3ceabSEric Dumazet 112746d3ceabSEric Dumazet INIT_LIST_HEAD(&tsq->head); 112846d3ceabSEric Dumazet tasklet_init(&tsq->tasklet, 112946d3ceabSEric Dumazet tcp_tasklet_func, 113046d3ceabSEric Dumazet (unsigned long)tsq); 113146d3ceabSEric Dumazet } 113246d3ceabSEric Dumazet } 113346d3ceabSEric Dumazet 113446d3ceabSEric Dumazet /* 113546d3ceabSEric Dumazet * Write buffer destructor automatically called from kfree_skb. 11368e3bff96Sstephen hemminger * We can't xmit new skbs from this context, as we might already 113746d3ceabSEric Dumazet * hold qdisc lock. 113846d3ceabSEric Dumazet */ 1139d6a4a104SEric Dumazet void tcp_wfree(struct sk_buff *skb) 114046d3ceabSEric Dumazet { 114146d3ceabSEric Dumazet struct sock *sk = skb->sk; 114246d3ceabSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 1143408f0a6cSEric Dumazet unsigned long flags, nval, oval; 11449b462d02SEric Dumazet 11459b462d02SEric Dumazet /* Keep one reference on sk_wmem_alloc. 11469b462d02SEric Dumazet * Will be released by sk_free() from here or tcp_tasklet_func() 11479b462d02SEric Dumazet */ 114814afee4bSReshetova, Elena WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc)); 11499b462d02SEric Dumazet 11509b462d02SEric Dumazet /* If this softirq is serviced by ksoftirqd, we are likely under stress. 11519b462d02SEric Dumazet * Wait until our queues (qdisc + devices) are drained. 11529b462d02SEric Dumazet * This gives : 11539b462d02SEric Dumazet * - less callbacks to tcp_write_xmit(), reducing stress (batches) 11549b462d02SEric Dumazet * - chance for incoming ACK (processed by another cpu maybe) 11559b462d02SEric Dumazet * to migrate this flow (skb->ooo_okay will be eventually set) 11569b462d02SEric Dumazet */ 115714afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current) 11589b462d02SEric Dumazet goto out; 115946d3ceabSEric Dumazet 11607aa5470cSEric Dumazet for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) { 116146d3ceabSEric Dumazet struct tsq_tasklet *tsq; 1162a9b204d1SEric Dumazet bool empty; 116346d3ceabSEric Dumazet 1164408f0a6cSEric Dumazet if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED)) 1165408f0a6cSEric Dumazet goto out; 1166408f0a6cSEric Dumazet 116773a6bab5SEric Dumazet nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED; 11687aa5470cSEric Dumazet nval = cmpxchg(&sk->sk_tsq_flags, oval, nval); 1169408f0a6cSEric Dumazet if (nval != oval) 1170408f0a6cSEric Dumazet continue; 1171408f0a6cSEric Dumazet 117246d3ceabSEric Dumazet /* queue this socket to tasklet queue */ 117346d3ceabSEric Dumazet local_irq_save(flags); 1174903ceff7SChristoph Lameter tsq = this_cpu_ptr(&tsq_tasklet); 1175a9b204d1SEric Dumazet empty = list_empty(&tsq->head); 117646d3ceabSEric Dumazet list_add(&tp->tsq_node, &tsq->head); 1177a9b204d1SEric Dumazet if (empty) 117846d3ceabSEric Dumazet tasklet_schedule(&tsq->tasklet); 117946d3ceabSEric Dumazet local_irq_restore(flags); 11809b462d02SEric Dumazet return; 118146d3ceabSEric Dumazet } 11829b462d02SEric Dumazet out: 11839b462d02SEric Dumazet sk_free(sk); 118446d3ceabSEric Dumazet } 118546d3ceabSEric Dumazet 118673a6bab5SEric Dumazet /* Note: Called under soft irq. 118773a6bab5SEric Dumazet * We can call TCP stack right away, unless socket is owned by user. 1188218af599SEric Dumazet */ 1189218af599SEric Dumazet enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer) 1190218af599SEric Dumazet { 1191218af599SEric Dumazet struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer); 1192218af599SEric Dumazet struct sock *sk = (struct sock *)tp; 1193218af599SEric Dumazet 119473a6bab5SEric Dumazet tcp_tsq_handler(sk); 119573a6bab5SEric Dumazet sock_put(sk); 1196218af599SEric Dumazet 1197218af599SEric Dumazet return HRTIMER_NORESTART; 1198218af599SEric Dumazet } 1199218af599SEric Dumazet 1200a7a25630SEric Dumazet static void tcp_update_skb_after_send(struct sock *sk, struct sk_buff *skb, 1201a7a25630SEric Dumazet u64 prior_wstamp) 1202e2080072SEric Dumazet { 1203ab408b6dSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 1204ab408b6dSEric Dumazet 1205ab408b6dSEric Dumazet if (sk->sk_pacing_status != SK_PACING_NONE) { 120676a9ebe8SEric Dumazet unsigned long rate = sk->sk_pacing_rate; 1207ab408b6dSEric Dumazet 1208ab408b6dSEric Dumazet /* Original sch_fq does not pace first 10 MSS 1209ab408b6dSEric Dumazet * Note that tp->data_segs_out overflows after 2^32 packets, 1210ab408b6dSEric Dumazet * this is a minor annoyance. 1211ab408b6dSEric Dumazet */ 121276a9ebe8SEric Dumazet if (rate != ~0UL && rate && tp->data_segs_out >= 10) { 1213a7a25630SEric Dumazet u64 len_ns = div64_ul((u64)skb->len * NSEC_PER_SEC, rate); 1214a7a25630SEric Dumazet u64 credit = tp->tcp_wstamp_ns - prior_wstamp; 1215a7a25630SEric Dumazet 1216a7a25630SEric Dumazet /* take into account OS jitter */ 1217a7a25630SEric Dumazet len_ns -= min_t(u64, len_ns / 2, credit); 1218a7a25630SEric Dumazet tp->tcp_wstamp_ns += len_ns; 1219ab408b6dSEric Dumazet } 1220ab408b6dSEric Dumazet } 1221e2080072SEric Dumazet list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue); 1222e2080072SEric Dumazet } 1223e2080072SEric Dumazet 122405e22e83SEric Dumazet INDIRECT_CALLABLE_DECLARE(int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl)); 122505e22e83SEric Dumazet INDIRECT_CALLABLE_DECLARE(int inet6_csk_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl)); 1226dd2e0b86SEric Dumazet INDIRECT_CALLABLE_DECLARE(void tcp_v4_send_check(struct sock *sk, struct sk_buff *skb)); 122705e22e83SEric Dumazet 12281da177e4SLinus Torvalds /* This routine actually transmits TCP packets queued in by 12291da177e4SLinus Torvalds * tcp_do_sendmsg(). This is used by both the initial 12301da177e4SLinus Torvalds * transmission and possible later retransmissions. 12311da177e4SLinus Torvalds * All SKB's seen here are completely headerless. It is our 12321da177e4SLinus Torvalds * job to build the TCP header, and pass the packet down to 12331da177e4SLinus Torvalds * IP so it can do the same plus pass the packet off to the 12341da177e4SLinus Torvalds * device. 12351da177e4SLinus Torvalds * 12361da177e4SLinus Torvalds * We are working here with either a clone of the original 12371da177e4SLinus Torvalds * SKB, or a fresh unique copy made by the retransmit engine. 12381da177e4SLinus Torvalds */ 12392987babbSYuchung Cheng static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, 12402987babbSYuchung Cheng int clone_it, gfp_t gfp_mask, u32 rcv_nxt) 12411da177e4SLinus Torvalds { 12426687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 1243dfb4b9dcSDavid S. Miller struct inet_sock *inet; 1244dfb4b9dcSDavid S. Miller struct tcp_sock *tp; 1245dfb4b9dcSDavid S. Miller struct tcp_skb_cb *tcb; 124633ad798cSAdam Langley struct tcp_out_options opts; 124795c96174SEric Dumazet unsigned int tcp_options_size, tcp_header_size; 12488c72c65bSEric Dumazet struct sk_buff *oskb = NULL; 1249cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 12501da177e4SLinus Torvalds struct tcphdr *th; 1251a7a25630SEric Dumazet u64 prior_wstamp; 12521da177e4SLinus Torvalds int err; 12531da177e4SLinus Torvalds 1254dfb4b9dcSDavid S. Miller BUG_ON(!skb || !tcp_skb_pcount(skb)); 12556f094b9eSLawrence Brakmo tp = tcp_sk(sk); 12567f12422cSYuchung Cheng prior_wstamp = tp->tcp_wstamp_ns; 12577f12422cSYuchung Cheng tp->tcp_wstamp_ns = max(tp->tcp_wstamp_ns, tp->tcp_clock_cache); 12587f12422cSYuchung Cheng skb->skb_mstamp_ns = tp->tcp_wstamp_ns; 1259ccdbb6e9SEric Dumazet if (clone_it) { 12606f094b9eSLawrence Brakmo TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq 12616f094b9eSLawrence Brakmo - tp->snd_una; 12628c72c65bSEric Dumazet oskb = skb; 1263e2080072SEric Dumazet 1264e2080072SEric Dumazet tcp_skb_tsorted_save(oskb) { 1265e2080072SEric Dumazet if (unlikely(skb_cloned(oskb))) 1266e2080072SEric Dumazet skb = pskb_copy(oskb, gfp_mask); 1267dfb4b9dcSDavid S. Miller else 1268e2080072SEric Dumazet skb = skb_clone(oskb, gfp_mask); 1269e2080072SEric Dumazet } tcp_skb_tsorted_restore(oskb); 1270e2080072SEric Dumazet 1271dfb4b9dcSDavid S. Miller if (unlikely(!skb)) 1272dfb4b9dcSDavid S. Miller return -ENOBUFS; 1273b738a185SEric Dumazet /* retransmit skbs might have a non zero value in skb->dev 1274b738a185SEric Dumazet * because skb->dev is aliased with skb->rbnode.rb_left 1275b738a185SEric Dumazet */ 1276b738a185SEric Dumazet skb->dev = NULL; 1277dfb4b9dcSDavid S. Miller } 12785f6188a8SEric Dumazet 1279dfb4b9dcSDavid S. Miller inet = inet_sk(sk); 1280dfb4b9dcSDavid S. Miller tcb = TCP_SKB_CB(skb); 128133ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 12821da177e4SLinus Torvalds 1283051ba674SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) { 128433ad798cSAdam Langley tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); 1285051ba674SEric Dumazet } else { 128633ad798cSAdam Langley tcp_options_size = tcp_established_options(sk, skb, &opts, 128733ad798cSAdam Langley &md5); 1288051ba674SEric Dumazet /* Force a PSH flag on all (GSO) packets to expedite GRO flush 1289051ba674SEric Dumazet * at receiver : This slightly improve GRO performance. 1290051ba674SEric Dumazet * Note that we do not force the PSH flag for non GSO packets, 1291051ba674SEric Dumazet * because they might be sent under high congestion events, 1292051ba674SEric Dumazet * and in this case it is better to delay the delivery of 1-MSS 1293051ba674SEric Dumazet * packets and thus the corresponding ACK packet that would 1294051ba674SEric Dumazet * release the following packet. 1295051ba674SEric Dumazet */ 1296051ba674SEric Dumazet if (tcp_skb_pcount(skb) > 1) 1297051ba674SEric Dumazet tcb->tcp_flags |= TCPHDR_PSH; 1298051ba674SEric Dumazet } 129933ad798cSAdam Langley tcp_header_size = tcp_options_size + sizeof(struct tcphdr); 13001da177e4SLinus Torvalds 1301547669d4SEric Dumazet /* if no packet is in qdisc/device queue, then allow XPS to select 1302b2532eb9SEric Dumazet * another queue. We can be called from tcp_tsq_handler() 130373a6bab5SEric Dumazet * which holds one reference to sk. 1304b2532eb9SEric Dumazet * 1305b2532eb9SEric Dumazet * TODO: Ideally, in-flight pure ACK packets should not matter here. 1306b2532eb9SEric Dumazet * One way to get this would be to set skb->truesize = 2 on them. 1307547669d4SEric Dumazet */ 1308b2532eb9SEric Dumazet skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1); 13091da177e4SLinus Torvalds 131038ab52e8SEric Dumazet /* If we had to use memory reserve to allocate this skb, 131138ab52e8SEric Dumazet * this might cause drops if packet is looped back : 131238ab52e8SEric Dumazet * Other socket might not have SOCK_MEMALLOC. 131338ab52e8SEric Dumazet * Packets not looped back do not care about pfmemalloc. 131438ab52e8SEric Dumazet */ 131538ab52e8SEric Dumazet skb->pfmemalloc = 0; 131638ab52e8SEric Dumazet 1317aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 1318aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 131946d3ceabSEric Dumazet 132046d3ceabSEric Dumazet skb_orphan(skb); 132146d3ceabSEric Dumazet skb->sk = sk; 13221d2077acSEric Dumazet skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree; 1323b73c3d0eSTom Herbert skb_set_hash_from_sk(skb, sk); 132414afee4bSReshetova, Elena refcount_add(skb->truesize, &sk->sk_wmem_alloc); 13251da177e4SLinus Torvalds 1326c3a2e837SJulian Anastasov skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm); 1327c3a2e837SJulian Anastasov 13281da177e4SLinus Torvalds /* Build TCP header and checksum it. */ 1329ea1627c2SEric Dumazet th = (struct tcphdr *)skb->data; 1330c720c7e8SEric Dumazet th->source = inet->inet_sport; 1331c720c7e8SEric Dumazet th->dest = inet->inet_dport; 13321da177e4SLinus Torvalds th->seq = htonl(tcb->seq); 13332987babbSYuchung Cheng th->ack_seq = htonl(rcv_nxt); 1334df7a3b07SAl Viro *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | 13354de075e0SEric Dumazet tcb->tcp_flags); 1336dfb4b9dcSDavid S. Miller 13371da177e4SLinus Torvalds th->check = 0; 13381da177e4SLinus Torvalds th->urg_ptr = 0; 13391da177e4SLinus Torvalds 134033f5f57eSIlpo Järvinen /* The urg_mode check is necessary during a below snd_una win probe */ 13417691367dSHerbert Xu if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { 13427691367dSHerbert Xu if (before(tp->snd_up, tcb->seq + 0x10000)) { 13431da177e4SLinus Torvalds th->urg_ptr = htons(tp->snd_up - tcb->seq); 13441da177e4SLinus Torvalds th->urg = 1; 13457691367dSHerbert Xu } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { 13460eae88f3SEric Dumazet th->urg_ptr = htons(0xFFFF); 13477691367dSHerbert Xu th->urg = 1; 13487691367dSHerbert Xu } 13491da177e4SLinus Torvalds } 13501da177e4SLinus Torvalds 1351bd0388aeSWilliam Allen Simpson tcp_options_write((__be32 *)(th + 1), tp, &opts); 135251466a75SEric Dumazet skb_shinfo(skb)->gso_type = sk->sk_gso_type; 1353ea1627c2SEric Dumazet if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) { 1354ea1627c2SEric Dumazet th->window = htons(tcp_select_window(sk)); 1355ea1627c2SEric Dumazet tcp_ecn_send(sk, skb, th, tcp_header_size); 1356ea1627c2SEric Dumazet } else { 1357ea1627c2SEric Dumazet /* RFC1323: The window in SYN & SYN/ACK segments 1358ea1627c2SEric Dumazet * is never scaled. 1359ea1627c2SEric Dumazet */ 1360ea1627c2SEric Dumazet th->window = htons(min(tp->rcv_wnd, 65535U)); 1361ea1627c2SEric Dumazet } 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); 1392f69ad292SEric Dumazet /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */ 1393cd7d8498SEric Dumazet skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb); 1394f69ad292SEric Dumazet skb_shinfo(skb)->gso_size = tcp_skb_mss(skb); 1395cd7d8498SEric Dumazet 1396d3edd06eSEric Dumazet /* Leave earliest departure time in skb->tstamp (skb->skb_mstamp_ns) */ 1397971f10ecSEric Dumazet 1398971f10ecSEric Dumazet /* Cleanup our debris for IP stacks */ 1399971f10ecSEric Dumazet memset(skb->cb, 0, max(sizeof(struct inet_skb_parm), 1400971f10ecSEric Dumazet sizeof(struct inet6_skb_parm))); 1401971f10ecSEric Dumazet 1402a842fe14SEric Dumazet tcp_add_tx_delay(skb, tp); 1403a842fe14SEric Dumazet 140405e22e83SEric Dumazet err = INDIRECT_CALL_INET(icsk->icsk_af_ops->queue_xmit, 140505e22e83SEric Dumazet inet6_csk_xmit, ip_queue_xmit, 140605e22e83SEric Dumazet sk, skb, &inet->cork.fl); 14077faee5c0SEric Dumazet 14088c72c65bSEric Dumazet if (unlikely(err > 0)) { 14095ee2c941SChristoph Paasch tcp_enter_cwr(sk); 14108c72c65bSEric Dumazet err = net_xmit_eval(err); 14118c72c65bSEric Dumazet } 1412fc225799SEric Dumazet if (!err && oskb) { 1413a7a25630SEric Dumazet tcp_update_skb_after_send(sk, oskb, prior_wstamp); 1414fc225799SEric Dumazet tcp_rate_skb_sent(sk, oskb); 1415fc225799SEric Dumazet } 14168c72c65bSEric Dumazet return err; 14171da177e4SLinus Torvalds } 14181da177e4SLinus Torvalds 14192987babbSYuchung Cheng static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, 14202987babbSYuchung Cheng gfp_t gfp_mask) 14212987babbSYuchung Cheng { 14222987babbSYuchung Cheng return __tcp_transmit_skb(sk, skb, clone_it, gfp_mask, 14232987babbSYuchung Cheng tcp_sk(sk)->rcv_nxt); 14242987babbSYuchung Cheng } 14252987babbSYuchung Cheng 142667edfef7SAndi Kleen /* This routine just queues the buffer for sending. 14271da177e4SLinus Torvalds * 14281da177e4SLinus Torvalds * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, 14291da177e4SLinus Torvalds * otherwise socket can stall. 14301da177e4SLinus Torvalds */ 14311da177e4SLinus Torvalds static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) 14321da177e4SLinus Torvalds { 14331da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 14341da177e4SLinus Torvalds 14351da177e4SLinus Torvalds /* Advance write_seq and place onto the write_queue. */ 14360f317464SEric Dumazet WRITE_ONCE(tp->write_seq, TCP_SKB_CB(skb)->end_seq); 1437f4a775d1SEric Dumazet __skb_header_release(skb); 1438fe067e8aSDavid S. Miller tcp_add_write_queue_tail(sk, skb); 1439ab4e846aSEric Dumazet sk_wmem_queued_add(sk, skb->truesize); 14403ab224beSHideo Aoki sk_mem_charge(sk, skb->truesize); 14411da177e4SLinus Torvalds } 14421da177e4SLinus Torvalds 144367edfef7SAndi Kleen /* Initialize TSO segments for a packet. */ 14445bbb432cSEric Dumazet static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now) 1445f6302d1dSDavid S. Miller { 14464a64fd6cSEric Dumazet if (skb->len <= mss_now) { 1447f6302d1dSDavid S. Miller /* Avoid the costly divide in the normal 1448f6302d1dSDavid S. Miller * non-TSO case. 1449f6302d1dSDavid S. Miller */ 1450cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, 1); 1451f69ad292SEric Dumazet TCP_SKB_CB(skb)->tcp_gso_size = 0; 1452f6302d1dSDavid S. Miller } else { 1453cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now)); 1454f69ad292SEric Dumazet TCP_SKB_CB(skb)->tcp_gso_size = mss_now; 14551da177e4SLinus Torvalds } 14561da177e4SLinus Torvalds } 14571da177e4SLinus Torvalds 1458797108d1SIlpo Järvinen /* Pcount in the middle of the write queue got changed, we need to do various 1459797108d1SIlpo Järvinen * tweaks to fix counters 1460797108d1SIlpo Järvinen */ 1461cf533ea5SEric Dumazet static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) 1462797108d1SIlpo Järvinen { 1463797108d1SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1464797108d1SIlpo Järvinen 1465797108d1SIlpo Järvinen tp->packets_out -= decr; 1466797108d1SIlpo Järvinen 1467797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 1468797108d1SIlpo Järvinen tp->sacked_out -= decr; 1469797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 1470797108d1SIlpo Järvinen tp->retrans_out -= decr; 1471797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 1472797108d1SIlpo Järvinen tp->lost_out -= decr; 1473797108d1SIlpo Järvinen 1474797108d1SIlpo Järvinen /* Reno case is special. Sigh... */ 1475797108d1SIlpo Järvinen if (tcp_is_reno(tp) && decr > 0) 1476797108d1SIlpo Järvinen tp->sacked_out -= min_t(u32, tp->sacked_out, decr); 1477797108d1SIlpo Järvinen 1478797108d1SIlpo Järvinen if (tp->lost_skb_hint && 1479797108d1SIlpo Järvinen before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && 1480713bafeaSYuchung Cheng (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) 1481797108d1SIlpo Järvinen tp->lost_cnt_hint -= decr; 1482797108d1SIlpo Järvinen 1483797108d1SIlpo Järvinen tcp_verify_left_out(tp); 1484797108d1SIlpo Järvinen } 1485797108d1SIlpo Järvinen 14860a2cf20cSSoheil Hassas Yeganeh static bool tcp_has_tx_tstamp(const struct sk_buff *skb) 14870a2cf20cSSoheil Hassas Yeganeh { 14880a2cf20cSSoheil Hassas Yeganeh return TCP_SKB_CB(skb)->txstamp_ack || 14890a2cf20cSSoheil Hassas Yeganeh (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP); 14900a2cf20cSSoheil Hassas Yeganeh } 14910a2cf20cSSoheil Hassas Yeganeh 1492490cc7d0SWillem de Bruijn static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2) 1493490cc7d0SWillem de Bruijn { 1494490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo = skb_shinfo(skb); 1495490cc7d0SWillem de Bruijn 14960a2cf20cSSoheil Hassas Yeganeh if (unlikely(tcp_has_tx_tstamp(skb)) && 1497490cc7d0SWillem de Bruijn !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) { 1498490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo2 = skb_shinfo(skb2); 1499490cc7d0SWillem de Bruijn u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP; 1500490cc7d0SWillem de Bruijn 1501490cc7d0SWillem de Bruijn shinfo->tx_flags &= ~tsflags; 1502490cc7d0SWillem de Bruijn shinfo2->tx_flags |= tsflags; 1503490cc7d0SWillem de Bruijn swap(shinfo->tskey, shinfo2->tskey); 1504b51e13faSMartin KaFai Lau TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack; 1505b51e13faSMartin KaFai Lau TCP_SKB_CB(skb)->txstamp_ack = 0; 1506490cc7d0SWillem de Bruijn } 1507490cc7d0SWillem de Bruijn } 1508490cc7d0SWillem de Bruijn 1509a166140eSMartin KaFai Lau static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2) 1510a166140eSMartin KaFai Lau { 1511a166140eSMartin KaFai Lau TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor; 1512a166140eSMartin KaFai Lau TCP_SKB_CB(skb)->eor = 0; 1513a166140eSMartin KaFai Lau } 1514a166140eSMartin KaFai Lau 151575c119afSEric Dumazet /* Insert buff after skb on the write or rtx queue of sk. */ 151675c119afSEric Dumazet static void tcp_insert_write_queue_after(struct sk_buff *skb, 151775c119afSEric Dumazet struct sk_buff *buff, 151875c119afSEric Dumazet struct sock *sk, 151975c119afSEric Dumazet enum tcp_queue tcp_queue) 152075c119afSEric Dumazet { 152175c119afSEric Dumazet if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE) 152275c119afSEric Dumazet __skb_queue_after(&sk->sk_write_queue, skb, buff); 152375c119afSEric Dumazet else 152475c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, buff); 152575c119afSEric Dumazet } 152675c119afSEric Dumazet 15271da177e4SLinus Torvalds /* Function to create two new TCP segments. Shrinks the given segment 15281da177e4SLinus Torvalds * to the specified size and appends a new segment with the rest of the 15291da177e4SLinus Torvalds * packet to the list. This won't be called frequently, I hope. 15301da177e4SLinus Torvalds * Remember, these are still headerless SKBs at this point. 15311da177e4SLinus Torvalds */ 153275c119afSEric Dumazet int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue, 153375c119afSEric Dumazet struct sk_buff *skb, u32 len, 15346cc55e09SOctavian Purdila unsigned int mss_now, gfp_t gfp) 15351da177e4SLinus Torvalds { 15361da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 15371da177e4SLinus Torvalds struct sk_buff *buff; 15386475be16SDavid S. Miller int nsize, old_factor; 1539b617158dSEric Dumazet long limit; 1540b60b49eaSHerbert Xu int nlen; 15419ce01461SIlpo Järvinen u8 flags; 15421da177e4SLinus Torvalds 15432fceec13SIlpo Järvinen if (WARN_ON(len > skb->len)) 15442fceec13SIlpo Järvinen return -EINVAL; 15456a438bbeSStephen Hemminger 15461da177e4SLinus Torvalds nsize = skb_headlen(skb) - len; 15471da177e4SLinus Torvalds if (nsize < 0) 15481da177e4SLinus Torvalds nsize = 0; 15491da177e4SLinus Torvalds 1550b617158dSEric Dumazet /* tcp_sendmsg() can overshoot sk_wmem_queued by one full size skb. 1551b617158dSEric Dumazet * We need some allowance to not penalize applications setting small 1552b617158dSEric Dumazet * SO_SNDBUF values. 1553b617158dSEric Dumazet * Also allow first and last skb in retransmit queue to be split. 1554b617158dSEric Dumazet */ 1555b617158dSEric Dumazet limit = sk->sk_sndbuf + 2 * SKB_TRUESIZE(GSO_MAX_SIZE); 1556b617158dSEric Dumazet if (unlikely((sk->sk_wmem_queued >> 1) > limit && 1557b617158dSEric Dumazet tcp_queue != TCP_FRAG_IN_WRITE_QUEUE && 1558b617158dSEric Dumazet skb != tcp_rtx_queue_head(sk) && 1559b617158dSEric Dumazet skb != tcp_rtx_queue_tail(sk))) { 1560f070ef2aSEric Dumazet NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPWQUEUETOOBIG); 1561f070ef2aSEric Dumazet return -ENOMEM; 1562f070ef2aSEric Dumazet } 1563f070ef2aSEric Dumazet 15646cc55e09SOctavian Purdila if (skb_unclone(skb, gfp)) 15651da177e4SLinus Torvalds return -ENOMEM; 15661da177e4SLinus Torvalds 15671da177e4SLinus Torvalds /* Get a new skb... force flag on. */ 1568eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, nsize, gfp, true); 156951456b29SIan Morris if (!buff) 15701da177e4SLinus Torvalds return -ENOMEM; /* We'll just try again later. */ 157141477662SJakub Kicinski skb_copy_decrypted(buff, skb); 1572ef5cb973SHerbert Xu 1573ab4e846aSEric Dumazet sk_wmem_queued_add(sk, buff->truesize); 15743ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1575b60b49eaSHerbert Xu nlen = skb->len - len - nsize; 1576b60b49eaSHerbert Xu buff->truesize += nlen; 1577b60b49eaSHerbert Xu skb->truesize -= nlen; 15781da177e4SLinus Torvalds 15791da177e4SLinus Torvalds /* Correct the sequence numbers. */ 15801da177e4SLinus Torvalds TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 15811da177e4SLinus Torvalds TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 15821da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 15831da177e4SLinus Torvalds 15841da177e4SLinus Torvalds /* PSH and FIN should only be set in the second packet. */ 15854de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 15864de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 15874de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1588e14c3cafSHerbert Xu TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; 1589a166140eSMartin KaFai Lau tcp_skb_fragment_eor(skb, buff); 15901da177e4SLinus Torvalds 15911da177e4SLinus Torvalds skb_split(skb, buff, len); 15921da177e4SLinus Torvalds 159398be9b12SEric Dumazet buff->ip_summed = CHECKSUM_PARTIAL; 15941da177e4SLinus Torvalds 1595a61bbcf2SPatrick McHardy buff->tstamp = skb->tstamp; 1596490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 15971da177e4SLinus Torvalds 15986475be16SDavid S. Miller old_factor = tcp_skb_pcount(skb); 15996475be16SDavid S. Miller 16001da177e4SLinus Torvalds /* Fix up tso_factor for both original and new SKB. */ 16015bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 16025bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 16031da177e4SLinus Torvalds 1604b9f64820SYuchung Cheng /* Update delivered info for the new segment */ 1605b9f64820SYuchung Cheng TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx; 1606b9f64820SYuchung Cheng 16076475be16SDavid S. Miller /* If this packet has been sent out already, we must 16086475be16SDavid S. Miller * adjust the various packet counters. 16096475be16SDavid S. Miller */ 1610cf0b450cSHerbert Xu if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 16116475be16SDavid S. Miller int diff = old_factor - tcp_skb_pcount(skb) - 16126475be16SDavid S. Miller tcp_skb_pcount(buff); 16131da177e4SLinus Torvalds 1614797108d1SIlpo Järvinen if (diff) 1615797108d1SIlpo Järvinen tcp_adjust_pcount(sk, skb, diff); 16161da177e4SLinus Torvalds } 16171da177e4SLinus Torvalds 16181da177e4SLinus Torvalds /* Link BUFF into the send queue. */ 1619f4a775d1SEric Dumazet __skb_header_release(buff); 162075c119afSEric Dumazet tcp_insert_write_queue_after(skb, buff, sk, tcp_queue); 1621f67971e6SEric Dumazet if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE) 1622e2080072SEric Dumazet list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor); 16231da177e4SLinus Torvalds 16241da177e4SLinus Torvalds return 0; 16251da177e4SLinus Torvalds } 16261da177e4SLinus Torvalds 1627f4d01666SEric Dumazet /* This is similar to __pskb_pull_tail(). The difference is that pulled 1628f4d01666SEric Dumazet * data is not copied, but immediately discarded. 16291da177e4SLinus Torvalds */ 16307162fb24SEric Dumazet static int __pskb_trim_head(struct sk_buff *skb, int len) 16311da177e4SLinus Torvalds { 16327b7fc97aSEric Dumazet struct skb_shared_info *shinfo; 16331da177e4SLinus Torvalds int i, k, eat; 16341da177e4SLinus Torvalds 16354fa48bf3SEric Dumazet eat = min_t(int, len, skb_headlen(skb)); 16364fa48bf3SEric Dumazet if (eat) { 16374fa48bf3SEric Dumazet __skb_pull(skb, eat); 16384fa48bf3SEric Dumazet len -= eat; 16394fa48bf3SEric Dumazet if (!len) 16407162fb24SEric Dumazet return 0; 16414fa48bf3SEric Dumazet } 16421da177e4SLinus Torvalds eat = len; 16431da177e4SLinus Torvalds k = 0; 16447b7fc97aSEric Dumazet shinfo = skb_shinfo(skb); 16457b7fc97aSEric Dumazet for (i = 0; i < shinfo->nr_frags; i++) { 16467b7fc97aSEric Dumazet int size = skb_frag_size(&shinfo->frags[i]); 16479e903e08SEric Dumazet 16489e903e08SEric Dumazet if (size <= eat) { 1649aff65da0SIan Campbell skb_frag_unref(skb, i); 16509e903e08SEric Dumazet eat -= size; 16511da177e4SLinus Torvalds } else { 16527b7fc97aSEric Dumazet shinfo->frags[k] = shinfo->frags[i]; 16531da177e4SLinus Torvalds if (eat) { 1654b54c9d5bSJonathan Lemon skb_frag_off_add(&shinfo->frags[k], eat); 16557b7fc97aSEric Dumazet skb_frag_size_sub(&shinfo->frags[k], eat); 16561da177e4SLinus Torvalds eat = 0; 16571da177e4SLinus Torvalds } 16581da177e4SLinus Torvalds k++; 16591da177e4SLinus Torvalds } 16601da177e4SLinus Torvalds } 16617b7fc97aSEric Dumazet shinfo->nr_frags = k; 16621da177e4SLinus Torvalds 16631da177e4SLinus Torvalds skb->data_len -= len; 16641da177e4SLinus Torvalds skb->len = skb->data_len; 16657162fb24SEric Dumazet return len; 16661da177e4SLinus Torvalds } 16671da177e4SLinus Torvalds 166867edfef7SAndi Kleen /* Remove acked data from a packet in the transmit queue. */ 16691da177e4SLinus Torvalds int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 16701da177e4SLinus Torvalds { 16717162fb24SEric Dumazet u32 delta_truesize; 16727162fb24SEric Dumazet 167314bbd6a5SPravin B Shelar if (skb_unclone(skb, GFP_ATOMIC)) 16741da177e4SLinus Torvalds return -ENOMEM; 16751da177e4SLinus Torvalds 16767162fb24SEric Dumazet delta_truesize = __pskb_trim_head(skb, len); 16771da177e4SLinus Torvalds 16781da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq += len; 167984fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 16801da177e4SLinus Torvalds 16817162fb24SEric Dumazet if (delta_truesize) { 16827162fb24SEric Dumazet skb->truesize -= delta_truesize; 1683ab4e846aSEric Dumazet sk_wmem_queued_add(sk, -delta_truesize); 16847162fb24SEric Dumazet sk_mem_uncharge(sk, delta_truesize); 16851da177e4SLinus Torvalds sock_set_flag(sk, SOCK_QUEUE_SHRUNK); 16867162fb24SEric Dumazet } 16871da177e4SLinus Torvalds 16885b35e1e6SNeal Cardwell /* Any change of skb->len requires recalculation of tso factor. */ 16891da177e4SLinus Torvalds if (tcp_skb_pcount(skb) > 1) 16905bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb)); 16911da177e4SLinus Torvalds 16921da177e4SLinus Torvalds return 0; 16931da177e4SLinus Torvalds } 16941da177e4SLinus Torvalds 16951b63edd6SYuchung Cheng /* Calculate MSS not accounting any TCP options. */ 16961b63edd6SYuchung Cheng static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu) 16975d424d5aSJohn Heffner { 1698cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1699cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 17005d424d5aSJohn Heffner int mss_now; 17015d424d5aSJohn Heffner 17025d424d5aSJohn Heffner /* Calculate base mss without TCP options: 17035d424d5aSJohn Heffner It is MMS_S - sizeof(tcphdr) of rfc1122 17045d424d5aSJohn Heffner */ 17055d424d5aSJohn Heffner mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 17065d424d5aSJohn Heffner 170767469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 170867469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 170967469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 171067469601SEric Dumazet 171167469601SEric Dumazet if (dst && dst_allfrag(dst)) 171267469601SEric Dumazet mss_now -= icsk->icsk_af_ops->net_frag_header_len; 171367469601SEric Dumazet } 171467469601SEric Dumazet 17155d424d5aSJohn Heffner /* Clamp it (mss_clamp does not include tcp options) */ 17165d424d5aSJohn Heffner if (mss_now > tp->rx_opt.mss_clamp) 17175d424d5aSJohn Heffner mss_now = tp->rx_opt.mss_clamp; 17185d424d5aSJohn Heffner 17195d424d5aSJohn Heffner /* Now subtract optional transport overhead */ 17205d424d5aSJohn Heffner mss_now -= icsk->icsk_ext_hdr_len; 17215d424d5aSJohn Heffner 17225d424d5aSJohn Heffner /* Then reserve room for full set of TCP options and 8 bytes of data */ 17235f3e2bf0SEric Dumazet mss_now = max(mss_now, sock_net(sk)->ipv4.sysctl_tcp_min_snd_mss); 17245d424d5aSJohn Heffner return mss_now; 17255d424d5aSJohn Heffner } 17265d424d5aSJohn Heffner 17271b63edd6SYuchung Cheng /* Calculate MSS. Not accounting for SACKs here. */ 17281b63edd6SYuchung Cheng int tcp_mtu_to_mss(struct sock *sk, int pmtu) 17291b63edd6SYuchung Cheng { 17301b63edd6SYuchung Cheng /* Subtract TCP options size, not including SACKs */ 17311b63edd6SYuchung Cheng return __tcp_mtu_to_mss(sk, pmtu) - 17321b63edd6SYuchung Cheng (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr)); 17331b63edd6SYuchung Cheng } 17341b63edd6SYuchung Cheng 17355d424d5aSJohn Heffner /* Inverse of above */ 173667469601SEric Dumazet int tcp_mss_to_mtu(struct sock *sk, int mss) 17375d424d5aSJohn Heffner { 1738cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1739cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 17405d424d5aSJohn Heffner int mtu; 17415d424d5aSJohn Heffner 17425d424d5aSJohn Heffner mtu = mss + 17435d424d5aSJohn Heffner tp->tcp_header_len + 17445d424d5aSJohn Heffner icsk->icsk_ext_hdr_len + 17455d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 17465d424d5aSJohn Heffner 174767469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 174867469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 174967469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 175067469601SEric Dumazet 175167469601SEric Dumazet if (dst && dst_allfrag(dst)) 175267469601SEric Dumazet mtu += icsk->icsk_af_ops->net_frag_header_len; 175367469601SEric Dumazet } 17545d424d5aSJohn Heffner return mtu; 17555d424d5aSJohn Heffner } 1756556c6b46SNeal Cardwell EXPORT_SYMBOL(tcp_mss_to_mtu); 17575d424d5aSJohn Heffner 175867edfef7SAndi Kleen /* MTU probing init per socket */ 17595d424d5aSJohn Heffner void tcp_mtup_init(struct sock *sk) 17605d424d5aSJohn Heffner { 17615d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 17625d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 1763b0f9ca53SFan Du struct net *net = sock_net(sk); 17645d424d5aSJohn Heffner 1765b0f9ca53SFan Du icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1; 17665d424d5aSJohn Heffner icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 17675d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 1768b0f9ca53SFan Du icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss); 17695d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 177005cbc0dbSFan Du if (icsk->icsk_mtup.enabled) 1771c74df29aSEric Dumazet icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; 17725d424d5aSJohn Heffner } 17734bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_mtup_init); 17745d424d5aSJohn Heffner 17751da177e4SLinus Torvalds /* This function synchronize snd mss to current pmtu/exthdr set. 17761da177e4SLinus Torvalds 17771da177e4SLinus Torvalds tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 17781da177e4SLinus Torvalds for TCP options, but includes only bare TCP header. 17791da177e4SLinus Torvalds 17801da177e4SLinus Torvalds tp->rx_opt.mss_clamp is mss negotiated at connection setup. 1781caa20d9aSStephen Hemminger It is minimum of user_mss and mss received with SYN. 17821da177e4SLinus Torvalds It also does not include TCP options. 17831da177e4SLinus Torvalds 1784d83d8461SArnaldo Carvalho de Melo inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. 17851da177e4SLinus Torvalds 17861da177e4SLinus Torvalds tp->mss_cache is current effective sending mss, including 17871da177e4SLinus Torvalds all tcp options except for SACKs. It is evaluated, 17881da177e4SLinus Torvalds taking into account current pmtu, but never exceeds 17891da177e4SLinus Torvalds tp->rx_opt.mss_clamp. 17901da177e4SLinus Torvalds 17911da177e4SLinus Torvalds NOTE1. rfc1122 clearly states that advertised MSS 17921da177e4SLinus Torvalds DOES NOT include either tcp or ip options. 17931da177e4SLinus Torvalds 1794d83d8461SArnaldo Carvalho de Melo NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache 1795d83d8461SArnaldo Carvalho de Melo are READ ONLY outside this function. --ANK (980731) 17961da177e4SLinus Torvalds */ 17971da177e4SLinus Torvalds unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 17981da177e4SLinus Torvalds { 17991da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1800d83d8461SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 18015d424d5aSJohn Heffner int mss_now; 18021da177e4SLinus Torvalds 18035d424d5aSJohn Heffner if (icsk->icsk_mtup.search_high > pmtu) 18045d424d5aSJohn Heffner icsk->icsk_mtup.search_high = pmtu; 18051da177e4SLinus Torvalds 18065d424d5aSJohn Heffner mss_now = tcp_mtu_to_mss(sk, pmtu); 1807409d22b4SIlpo Järvinen mss_now = tcp_bound_to_half_wnd(tp, mss_now); 18081da177e4SLinus Torvalds 18091da177e4SLinus Torvalds /* And store cached results */ 1810d83d8461SArnaldo Carvalho de Melo icsk->icsk_pmtu_cookie = pmtu; 18115d424d5aSJohn Heffner if (icsk->icsk_mtup.enabled) 18125d424d5aSJohn Heffner mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); 1813c1b4a7e6SDavid S. Miller tp->mss_cache = mss_now; 18141da177e4SLinus Torvalds 18151da177e4SLinus Torvalds return mss_now; 18161da177e4SLinus Torvalds } 18174bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_sync_mss); 18181da177e4SLinus Torvalds 18191da177e4SLinus Torvalds /* Compute the current effective MSS, taking SACKs and IP options, 18201da177e4SLinus Torvalds * and even PMTU discovery events into account. 18211da177e4SLinus Torvalds */ 18220c54b85fSIlpo Järvinen unsigned int tcp_current_mss(struct sock *sk) 18231da177e4SLinus Torvalds { 1824cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1825cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1826c1b4a7e6SDavid S. Miller u32 mss_now; 182795c96174SEric Dumazet unsigned int header_len; 182833ad798cSAdam Langley struct tcp_out_options opts; 182933ad798cSAdam Langley struct tcp_md5sig_key *md5; 18301da177e4SLinus Torvalds 1831c1b4a7e6SDavid S. Miller mss_now = tp->mss_cache; 1832c1b4a7e6SDavid S. Miller 18331da177e4SLinus Torvalds if (dst) { 18341da177e4SLinus Torvalds u32 mtu = dst_mtu(dst); 1835d83d8461SArnaldo Carvalho de Melo if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 18361da177e4SLinus Torvalds mss_now = tcp_sync_mss(sk, mtu); 18371da177e4SLinus Torvalds } 18381da177e4SLinus Torvalds 183933ad798cSAdam Langley header_len = tcp_established_options(sk, NULL, &opts, &md5) + 184033ad798cSAdam Langley sizeof(struct tcphdr); 184133ad798cSAdam Langley /* The mss_cache is sized based on tp->tcp_header_len, which assumes 184233ad798cSAdam Langley * some common options. If this is an odd packet (because we have SACK 184333ad798cSAdam Langley * blocks etc) then our calculated header_len will be different, and 184433ad798cSAdam Langley * we have to adjust mss_now correspondingly */ 184533ad798cSAdam Langley if (header_len != tp->tcp_header_len) { 184633ad798cSAdam Langley int delta = (int) header_len - tp->tcp_header_len; 184733ad798cSAdam Langley mss_now -= delta; 184833ad798cSAdam Langley } 1849cfb6eeb4SYOSHIFUJI Hideaki 18501da177e4SLinus Torvalds return mss_now; 18511da177e4SLinus Torvalds } 18521da177e4SLinus Torvalds 185386fd14adSWeiping Pan /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. 185486fd14adSWeiping Pan * As additional protections, we do not touch cwnd in retransmission phases, 185586fd14adSWeiping Pan * and if application hit its sndbuf limit recently. 185686fd14adSWeiping Pan */ 185786fd14adSWeiping Pan static void tcp_cwnd_application_limited(struct sock *sk) 1858a762a980SDavid S. Miller { 18599e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1860a762a980SDavid S. Miller 186186fd14adSWeiping Pan if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && 186286fd14adSWeiping Pan sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 186386fd14adSWeiping Pan /* Limited by application or receiver window. */ 186486fd14adSWeiping Pan u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); 186586fd14adSWeiping Pan u32 win_used = max(tp->snd_cwnd_used, init_win); 186686fd14adSWeiping Pan if (win_used < tp->snd_cwnd) { 186786fd14adSWeiping Pan tp->snd_ssthresh = tcp_current_ssthresh(sk); 186886fd14adSWeiping Pan tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1; 186986fd14adSWeiping Pan } 187086fd14adSWeiping Pan tp->snd_cwnd_used = 0; 187186fd14adSWeiping Pan } 1872c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 187386fd14adSWeiping Pan } 187486fd14adSWeiping Pan 1875ca8a2263SNeal Cardwell static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited) 1876a762a980SDavid S. Miller { 18771b1fc3fdSWei Wang const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 1878a762a980SDavid S. Miller struct tcp_sock *tp = tcp_sk(sk); 1879a762a980SDavid S. Miller 1880ca8a2263SNeal Cardwell /* Track the maximum number of outstanding packets in each 1881ca8a2263SNeal Cardwell * window, and remember whether we were cwnd-limited then. 1882ca8a2263SNeal Cardwell */ 1883ca8a2263SNeal Cardwell if (!before(tp->snd_una, tp->max_packets_seq) || 1884ca8a2263SNeal Cardwell tp->packets_out > tp->max_packets_out) { 1885ca8a2263SNeal Cardwell tp->max_packets_out = tp->packets_out; 1886ca8a2263SNeal Cardwell tp->max_packets_seq = tp->snd_nxt; 1887ca8a2263SNeal Cardwell tp->is_cwnd_limited = is_cwnd_limited; 1888ca8a2263SNeal Cardwell } 1889e114a710SEric Dumazet 189024901551SEric Dumazet if (tcp_is_cwnd_limited(sk)) { 1891a762a980SDavid S. Miller /* Network is feed fully. */ 1892a762a980SDavid S. Miller tp->snd_cwnd_used = 0; 1893c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 1894a762a980SDavid S. Miller } else { 1895a762a980SDavid S. Miller /* Network starves. */ 1896a762a980SDavid S. Miller if (tp->packets_out > tp->snd_cwnd_used) 1897a762a980SDavid S. Miller tp->snd_cwnd_used = tp->packets_out; 1898a762a980SDavid S. Miller 1899b510f0d2SEric Dumazet if (sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle && 1900c2203cf7SEric Dumazet (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto && 19011b1fc3fdSWei Wang !ca_ops->cong_control) 1902a762a980SDavid S. Miller tcp_cwnd_application_limited(sk); 1903b0f71bd3SFrancis Yan 1904b0f71bd3SFrancis Yan /* The following conditions together indicate the starvation 1905b0f71bd3SFrancis Yan * is caused by insufficient sender buffer: 1906b0f71bd3SFrancis Yan * 1) just sent some data (see tcp_write_xmit) 1907b0f71bd3SFrancis Yan * 2) not cwnd limited (this else condition) 190875c119afSEric Dumazet * 3) no more data to send (tcp_write_queue_empty()) 1909b0f71bd3SFrancis Yan * 4) application is hitting buffer limit (SOCK_NOSPACE) 1910b0f71bd3SFrancis Yan */ 191175c119afSEric Dumazet if (tcp_write_queue_empty(sk) && sk->sk_socket && 1912b0f71bd3SFrancis Yan test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) && 1913b0f71bd3SFrancis Yan (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) 1914b0f71bd3SFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED); 1915a762a980SDavid S. Miller } 1916a762a980SDavid S. Miller } 1917a762a980SDavid S. Miller 1918d4589926SEric Dumazet /* Minshall's variant of the Nagle send check. */ 1919d4589926SEric Dumazet static bool tcp_minshall_check(const struct tcp_sock *tp) 1920d4589926SEric Dumazet { 1921d4589926SEric Dumazet return after(tp->snd_sml, tp->snd_una) && 1922d4589926SEric Dumazet !after(tp->snd_sml, tp->snd_nxt); 1923d4589926SEric Dumazet } 1924d4589926SEric Dumazet 1925d4589926SEric Dumazet /* Update snd_sml if this skb is under mss 1926d4589926SEric Dumazet * Note that a TSO packet might end with a sub-mss segment 1927d4589926SEric Dumazet * The test is really : 1928d4589926SEric Dumazet * if ((skb->len % mss) != 0) 1929d4589926SEric Dumazet * tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1930d4589926SEric Dumazet * But we can avoid doing the divide again given we already have 1931d4589926SEric Dumazet * skb_pcount = skb->len / mss_now 19320e3a4803SIlpo Järvinen */ 1933d4589926SEric Dumazet static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now, 1934d4589926SEric Dumazet const struct sk_buff *skb) 1935d4589926SEric Dumazet { 1936d4589926SEric Dumazet if (skb->len < tcp_skb_pcount(skb) * mss_now) 1937d4589926SEric Dumazet tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1938d4589926SEric Dumazet } 1939d4589926SEric Dumazet 1940d4589926SEric Dumazet /* Return false, if packet can be sent now without violation Nagle's rules: 1941d4589926SEric Dumazet * 1. It is full sized. (provided by caller in %partial bool) 1942d4589926SEric Dumazet * 2. Or it contains FIN. (already checked by caller) 1943d4589926SEric Dumazet * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. 1944d4589926SEric Dumazet * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 1945d4589926SEric Dumazet * With Minshall's modification: all sent small packets are ACKed. 1946d4589926SEric Dumazet */ 1947d4589926SEric Dumazet static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp, 1948cc93fc51SPeter Pan(潘卫平) int nonagle) 1949d4589926SEric Dumazet { 1950d4589926SEric Dumazet return partial && 1951d4589926SEric Dumazet ((nonagle & TCP_NAGLE_CORK) || 1952d4589926SEric Dumazet (!nonagle && tp->packets_out && tcp_minshall_check(tp))); 1953d4589926SEric Dumazet } 1954605ad7f1SEric Dumazet 1955605ad7f1SEric Dumazet /* Return how many segs we'd like on a TSO packet, 1956605ad7f1SEric Dumazet * to send one TSO packet per ms 1957605ad7f1SEric Dumazet */ 1958dcb8c9b4SEric Dumazet static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now, 19591b3878caSNeal Cardwell int min_tso_segs) 1960605ad7f1SEric Dumazet { 1961605ad7f1SEric Dumazet u32 bytes, segs; 1962605ad7f1SEric Dumazet 196376a9ebe8SEric Dumazet bytes = min_t(unsigned long, 19647c68fa2bSEric Dumazet sk->sk_pacing_rate >> READ_ONCE(sk->sk_pacing_shift), 1965605ad7f1SEric Dumazet sk->sk_gso_max_size - 1 - MAX_TCP_HEADER); 1966605ad7f1SEric Dumazet 1967605ad7f1SEric Dumazet /* Goal is to send at least one packet per ms, 1968605ad7f1SEric Dumazet * not one big TSO packet every 100 ms. 1969605ad7f1SEric Dumazet * This preserves ACK clocking and is consistent 1970605ad7f1SEric Dumazet * with tcp_tso_should_defer() heuristic. 1971605ad7f1SEric Dumazet */ 19721b3878caSNeal Cardwell segs = max_t(u32, bytes / mss_now, min_tso_segs); 1973605ad7f1SEric Dumazet 1974350c9f48SEric Dumazet return segs; 1975605ad7f1SEric Dumazet } 1976605ad7f1SEric Dumazet 1977ed6e7268SNeal Cardwell /* Return the number of segments we want in the skb we are transmitting. 1978ed6e7268SNeal Cardwell * See if congestion control module wants to decide; otherwise, autosize. 1979ed6e7268SNeal Cardwell */ 1980ed6e7268SNeal Cardwell static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now) 1981ed6e7268SNeal Cardwell { 1982ed6e7268SNeal Cardwell const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 1983dcb8c9b4SEric Dumazet u32 min_tso, tso_segs; 1984ed6e7268SNeal Cardwell 1985dcb8c9b4SEric Dumazet min_tso = ca_ops->min_tso_segs ? 1986dcb8c9b4SEric Dumazet ca_ops->min_tso_segs(sk) : 1987dcb8c9b4SEric Dumazet sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs; 1988dcb8c9b4SEric Dumazet 1989dcb8c9b4SEric Dumazet tso_segs = tcp_tso_autosize(sk, mss_now, min_tso); 1990350c9f48SEric Dumazet return min_t(u32, tso_segs, sk->sk_gso_max_segs); 1991ed6e7268SNeal Cardwell } 1992ed6e7268SNeal Cardwell 1993d4589926SEric Dumazet /* Returns the portion of skb which can be sent right away */ 1994d4589926SEric Dumazet static unsigned int tcp_mss_split_point(const struct sock *sk, 1995d4589926SEric Dumazet const struct sk_buff *skb, 1996d4589926SEric Dumazet unsigned int mss_now, 1997d4589926SEric Dumazet unsigned int max_segs, 1998d4589926SEric Dumazet int nonagle) 1999c1b4a7e6SDavid S. Miller { 2000cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 2001d4589926SEric Dumazet u32 partial, needed, window, max_len; 2002c1b4a7e6SDavid S. Miller 200390840defSIlpo Järvinen window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 20041485348dSBen Hutchings max_len = mss_now * max_segs; 20050e3a4803SIlpo Järvinen 20061485348dSBen Hutchings if (likely(max_len <= window && skb != tcp_write_queue_tail(sk))) 20071485348dSBen Hutchings return max_len; 20080e3a4803SIlpo Järvinen 20095ea3a748SIlpo Järvinen needed = min(skb->len, window); 20105ea3a748SIlpo Järvinen 20111485348dSBen Hutchings if (max_len <= needed) 20121485348dSBen Hutchings return max_len; 20130e3a4803SIlpo Järvinen 2014d4589926SEric Dumazet partial = needed % mss_now; 2015d4589926SEric Dumazet /* If last segment is not a full MSS, check if Nagle rules allow us 2016d4589926SEric Dumazet * to include this last segment in this skb. 2017d4589926SEric Dumazet * Otherwise, we'll split the skb at last MSS boundary 2018d4589926SEric Dumazet */ 2019cc93fc51SPeter Pan(潘卫平) if (tcp_nagle_check(partial != 0, tp, nonagle)) 2020d4589926SEric Dumazet return needed - partial; 2021d4589926SEric Dumazet 2022d4589926SEric Dumazet return needed; 2023c1b4a7e6SDavid S. Miller } 2024c1b4a7e6SDavid S. Miller 2025c1b4a7e6SDavid S. Miller /* Can at least one segment of SKB be sent right now, according to the 2026c1b4a7e6SDavid S. Miller * congestion window rules? If so, return how many segments are allowed. 2027c1b4a7e6SDavid S. Miller */ 2028cf533ea5SEric Dumazet static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, 2029cf533ea5SEric Dumazet const struct sk_buff *skb) 2030c1b4a7e6SDavid S. Miller { 2031d649a7a8SEric Dumazet u32 in_flight, cwnd, halfcwnd; 2032c1b4a7e6SDavid S. Miller 2033c1b4a7e6SDavid S. Miller /* Don't be strict about the congestion window for the final FIN. */ 20344de075e0SEric Dumazet if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && 20354de075e0SEric Dumazet tcp_skb_pcount(skb) == 1) 2036c1b4a7e6SDavid S. Miller return 1; 2037c1b4a7e6SDavid S. Miller 2038c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 2039c1b4a7e6SDavid S. Miller cwnd = tp->snd_cwnd; 2040d649a7a8SEric Dumazet if (in_flight >= cwnd) 2041c1b4a7e6SDavid S. Miller return 0; 2042d649a7a8SEric Dumazet 2043d649a7a8SEric Dumazet /* For better scheduling, ensure we have at least 2044d649a7a8SEric Dumazet * 2 GSO packets in flight. 2045d649a7a8SEric Dumazet */ 2046d649a7a8SEric Dumazet halfcwnd = max(cwnd >> 1, 1U); 2047d649a7a8SEric Dumazet return min(halfcwnd, cwnd - in_flight); 2048c1b4a7e6SDavid S. Miller } 2049c1b4a7e6SDavid S. Miller 2050b595076aSUwe Kleine-König /* Initialize TSO state of a skb. 205167edfef7SAndi Kleen * This must be invoked the first time we consider transmitting 2052c1b4a7e6SDavid S. Miller * SKB onto the wire. 2053c1b4a7e6SDavid S. Miller */ 20545bbb432cSEric Dumazet static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now) 2055c1b4a7e6SDavid S. Miller { 2056c1b4a7e6SDavid S. Miller int tso_segs = tcp_skb_pcount(skb); 2057c1b4a7e6SDavid S. Miller 2058f8269a49SIlpo Järvinen if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { 20595bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 2060c1b4a7e6SDavid S. Miller tso_segs = tcp_skb_pcount(skb); 2061c1b4a7e6SDavid S. Miller } 2062c1b4a7e6SDavid S. Miller return tso_segs; 2063c1b4a7e6SDavid S. Miller } 2064c1b4a7e6SDavid S. Miller 2065c1b4a7e6SDavid S. Miller 2066a2a385d6SEric Dumazet /* Return true if the Nagle test allows this packet to be 2067c1b4a7e6SDavid S. Miller * sent now. 2068c1b4a7e6SDavid S. Miller */ 2069a2a385d6SEric Dumazet static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, 2070c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 2071c1b4a7e6SDavid S. Miller { 2072c1b4a7e6SDavid S. Miller /* Nagle rule does not apply to frames, which sit in the middle of the 2073c1b4a7e6SDavid S. Miller * write_queue (they have no chances to get new data). 2074c1b4a7e6SDavid S. Miller * 2075c1b4a7e6SDavid S. Miller * This is implemented in the callers, where they modify the 'nonagle' 2076c1b4a7e6SDavid S. Miller * argument based upon the location of SKB in the send queue. 2077c1b4a7e6SDavid S. Miller */ 2078c1b4a7e6SDavid S. Miller if (nonagle & TCP_NAGLE_PUSH) 2079a2a385d6SEric Dumazet return true; 2080c1b4a7e6SDavid S. Miller 20819b44190dSYuchung Cheng /* Don't use the nagle rule for urgent data (or for the final FIN). */ 20829b44190dSYuchung Cheng if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) 2083a2a385d6SEric Dumazet return true; 2084c1b4a7e6SDavid S. Miller 2085cc93fc51SPeter Pan(潘卫平) if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle)) 2086a2a385d6SEric Dumazet return true; 2087c1b4a7e6SDavid S. Miller 2088a2a385d6SEric Dumazet return false; 2089c1b4a7e6SDavid S. Miller } 2090c1b4a7e6SDavid S. Miller 2091c1b4a7e6SDavid S. Miller /* Does at least the first segment of SKB fit into the send window? */ 2092a2a385d6SEric Dumazet static bool tcp_snd_wnd_test(const struct tcp_sock *tp, 2093a2a385d6SEric Dumazet const struct sk_buff *skb, 2094056834d9SIlpo Järvinen unsigned int cur_mss) 2095c1b4a7e6SDavid S. Miller { 2096c1b4a7e6SDavid S. Miller u32 end_seq = TCP_SKB_CB(skb)->end_seq; 2097c1b4a7e6SDavid S. Miller 2098c1b4a7e6SDavid S. Miller if (skb->len > cur_mss) 2099c1b4a7e6SDavid S. Miller end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 2100c1b4a7e6SDavid S. Miller 210190840defSIlpo Järvinen return !after(end_seq, tcp_wnd_end(tp)); 2102c1b4a7e6SDavid S. Miller } 2103c1b4a7e6SDavid S. Miller 2104c1b4a7e6SDavid S. Miller /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 2105c1b4a7e6SDavid S. Miller * which is put after SKB on the list. It is very much like 2106c1b4a7e6SDavid S. Miller * tcp_fragment() except that it may make several kinds of assumptions 2107c1b4a7e6SDavid S. Miller * in order to speed up the splitting operation. In particular, we 2108c1b4a7e6SDavid S. Miller * know that all the data is in scatter-gather pages, and that the 2109c1b4a7e6SDavid S. Miller * packet has never been sent out before (and thus is not cloned). 2110c1b4a7e6SDavid S. Miller */ 211156483341SEric Dumazet static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, 2112c4ead4c5SEric Dumazet unsigned int mss_now, gfp_t gfp) 2113c1b4a7e6SDavid S. Miller { 2114c1b4a7e6SDavid S. Miller int nlen = skb->len - len; 211556483341SEric Dumazet struct sk_buff *buff; 21169ce01461SIlpo Järvinen u8 flags; 2117c1b4a7e6SDavid S. Miller 2118c1b4a7e6SDavid S. Miller /* All of a TSO frame must be composed of paged data. */ 2119c8ac3774SHerbert Xu if (skb->len != skb->data_len) 212056483341SEric Dumazet return tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE, 212156483341SEric Dumazet skb, len, mss_now, gfp); 2122c1b4a7e6SDavid S. Miller 2123eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, gfp, true); 212451456b29SIan Morris if (unlikely(!buff)) 2125c1b4a7e6SDavid S. Miller return -ENOMEM; 212641477662SJakub Kicinski skb_copy_decrypted(buff, skb); 2127c1b4a7e6SDavid S. Miller 2128ab4e846aSEric Dumazet sk_wmem_queued_add(sk, buff->truesize); 21293ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 2130b60b49eaSHerbert Xu buff->truesize += nlen; 2131c1b4a7e6SDavid S. Miller skb->truesize -= nlen; 2132c1b4a7e6SDavid S. Miller 2133c1b4a7e6SDavid S. Miller /* Correct the sequence numbers. */ 2134c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 2135c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 2136c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 2137c1b4a7e6SDavid S. Miller 2138c1b4a7e6SDavid S. Miller /* PSH and FIN should only be set in the second packet. */ 21394de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 21404de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 21414de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 2142c1b4a7e6SDavid S. Miller 2143c1b4a7e6SDavid S. Miller /* This packet was never sent out yet, so no SACK bits. */ 2144c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->sacked = 0; 2145c1b4a7e6SDavid S. Miller 2146a166140eSMartin KaFai Lau tcp_skb_fragment_eor(skb, buff); 2147a166140eSMartin KaFai Lau 214898be9b12SEric Dumazet buff->ip_summed = CHECKSUM_PARTIAL; 2149c1b4a7e6SDavid S. Miller skb_split(skb, buff, len); 2150490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 2151c1b4a7e6SDavid S. Miller 2152c1b4a7e6SDavid S. Miller /* Fix up tso_factor for both original and new SKB. */ 21535bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 21545bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 2155c1b4a7e6SDavid S. Miller 2156c1b4a7e6SDavid S. Miller /* Link BUFF into the send queue. */ 2157f4a775d1SEric Dumazet __skb_header_release(buff); 215856483341SEric Dumazet tcp_insert_write_queue_after(skb, buff, sk, TCP_FRAG_IN_WRITE_QUEUE); 2159c1b4a7e6SDavid S. Miller 2160c1b4a7e6SDavid S. Miller return 0; 2161c1b4a7e6SDavid S. Miller } 2162c1b4a7e6SDavid S. Miller 2163c1b4a7e6SDavid S. Miller /* Try to defer sending, if possible, in order to minimize the amount 2164c1b4a7e6SDavid S. Miller * of TSO splitting we do. View it as a kind of TSO Nagle test. 2165c1b4a7e6SDavid S. Miller * 2166c1b4a7e6SDavid S. Miller * This algorithm is from John Heffner. 2167c1b4a7e6SDavid S. Miller */ 2168ca8a2263SNeal Cardwell static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb, 2169f9bfe4e6SEric Dumazet bool *is_cwnd_limited, 2170f9bfe4e6SEric Dumazet bool *is_rwnd_limited, 2171f9bfe4e6SEric Dumazet u32 max_segs) 2172c1b4a7e6SDavid S. Miller { 21736687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 2174f1c6ea38SEric Dumazet u32 send_win, cong_win, limit, in_flight; 217550c8339eSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 217650c8339eSEric Dumazet struct sk_buff *head; 2177ad9f4f50SEric Dumazet int win_divisor; 2178f1c6ea38SEric Dumazet s64 delta; 2179c1b4a7e6SDavid S. Miller 218099d7662aSEric Dumazet if (icsk->icsk_ca_state >= TCP_CA_Recovery) 2181ae8064acSJohn Heffner goto send_now; 2182ae8064acSJohn Heffner 21835f852eb5SEric Dumazet /* Avoid bursty behavior by allowing defer 2184a682850aSEric Dumazet * only if the last write was recent (1 ms). 2185a682850aSEric Dumazet * Note that tp->tcp_wstamp_ns can be in the future if we have 2186a682850aSEric Dumazet * packets waiting in a qdisc or device for EDT delivery. 21875f852eb5SEric Dumazet */ 2188a682850aSEric Dumazet delta = tp->tcp_clock_cache - tp->tcp_wstamp_ns - NSEC_PER_MSEC; 2189a682850aSEric Dumazet if (delta > 0) 2190ae8064acSJohn Heffner goto send_now; 2191908a75c1SDavid S. Miller 2192c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 2193c1b4a7e6SDavid S. Miller 2194c8c9aeb5SStefano Brivio BUG_ON(tcp_skb_pcount(skb) <= 1); 2195c8c9aeb5SStefano Brivio BUG_ON(tp->snd_cwnd <= in_flight); 2196c1b4a7e6SDavid S. Miller 219790840defSIlpo Järvinen send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 2198c1b4a7e6SDavid S. Miller 2199c1b4a7e6SDavid S. Miller /* From in_flight test above, we know that cwnd > in_flight. */ 2200c1b4a7e6SDavid S. Miller cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; 2201c1b4a7e6SDavid S. Miller 2202c1b4a7e6SDavid S. Miller limit = min(send_win, cong_win); 2203c1b4a7e6SDavid S. Miller 2204ba244fe9SDavid S. Miller /* If a full-sized TSO skb can be sent, do it. */ 2205605ad7f1SEric Dumazet if (limit >= max_segs * tp->mss_cache) 2206ae8064acSJohn Heffner goto send_now; 2207ba244fe9SDavid S. Miller 220862ad2761SIlpo Järvinen /* Middle in queue won't get any more data, full sendable already? */ 220962ad2761SIlpo Järvinen if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) 221062ad2761SIlpo Järvinen goto send_now; 221162ad2761SIlpo Järvinen 22125bbcc0f5SLinus Torvalds win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor); 2213ad9f4f50SEric Dumazet if (win_divisor) { 2214c1b4a7e6SDavid S. Miller u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); 2215c1b4a7e6SDavid S. Miller 2216c1b4a7e6SDavid S. Miller /* If at least some fraction of a window is available, 2217c1b4a7e6SDavid S. Miller * just use it. 2218c1b4a7e6SDavid S. Miller */ 2219ad9f4f50SEric Dumazet chunk /= win_divisor; 2220c1b4a7e6SDavid S. Miller if (limit >= chunk) 2221ae8064acSJohn Heffner goto send_now; 2222c1b4a7e6SDavid S. Miller } else { 2223c1b4a7e6SDavid S. Miller /* Different approach, try not to defer past a single 2224c1b4a7e6SDavid S. Miller * ACK. Receiver should ACK every other full sized 2225c1b4a7e6SDavid S. Miller * frame, so if we have space for more than 3 frames 2226c1b4a7e6SDavid S. Miller * then send now. 2227c1b4a7e6SDavid S. Miller */ 22286b5a5c0dSNeal Cardwell if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) 2229ae8064acSJohn Heffner goto send_now; 2230c1b4a7e6SDavid S. Miller } 2231c1b4a7e6SDavid S. Miller 223275c119afSEric Dumazet /* TODO : use tsorted_sent_queue ? */ 223375c119afSEric Dumazet head = tcp_rtx_queue_head(sk); 223475c119afSEric Dumazet if (!head) 223575c119afSEric Dumazet goto send_now; 2236f1c6ea38SEric Dumazet delta = tp->tcp_clock_cache - head->tstamp; 223750c8339eSEric Dumazet /* If next ACK is likely to come too late (half srtt), do not defer */ 2238f1c6ea38SEric Dumazet if ((s64)(delta - (u64)NSEC_PER_USEC * (tp->srtt_us >> 4)) < 0) 223950c8339eSEric Dumazet goto send_now; 224050c8339eSEric Dumazet 2241f9bfe4e6SEric Dumazet /* Ok, it looks like it is advisable to defer. 2242f9bfe4e6SEric Dumazet * Three cases are tracked : 2243f9bfe4e6SEric Dumazet * 1) We are cwnd-limited 2244f9bfe4e6SEric Dumazet * 2) We are rwnd-limited 2245f9bfe4e6SEric Dumazet * 3) We are application limited. 2246f9bfe4e6SEric Dumazet */ 2247f9bfe4e6SEric Dumazet if (cong_win < send_win) { 2248f9bfe4e6SEric Dumazet if (cong_win <= skb->len) { 2249ca8a2263SNeal Cardwell *is_cwnd_limited = true; 2250f9bfe4e6SEric Dumazet return true; 2251f9bfe4e6SEric Dumazet } 2252f9bfe4e6SEric Dumazet } else { 2253f9bfe4e6SEric Dumazet if (send_win <= skb->len) { 2254f9bfe4e6SEric Dumazet *is_rwnd_limited = true; 2255f9bfe4e6SEric Dumazet return true; 2256f9bfe4e6SEric Dumazet } 2257f9bfe4e6SEric Dumazet } 2258f9bfe4e6SEric Dumazet 2259f9bfe4e6SEric Dumazet /* If this packet won't get more data, do not wait. */ 2260d8ed257fSEric Dumazet if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) || 2261d8ed257fSEric Dumazet TCP_SKB_CB(skb)->eor) 2262f9bfe4e6SEric Dumazet goto send_now; 2263ca8a2263SNeal Cardwell 2264a2a385d6SEric Dumazet return true; 2265ae8064acSJohn Heffner 2266ae8064acSJohn Heffner send_now: 2267a2a385d6SEric Dumazet return false; 2268c1b4a7e6SDavid S. Miller } 2269c1b4a7e6SDavid S. Miller 227005cbc0dbSFan Du static inline void tcp_mtu_check_reprobe(struct sock *sk) 227105cbc0dbSFan Du { 227205cbc0dbSFan Du struct inet_connection_sock *icsk = inet_csk(sk); 227305cbc0dbSFan Du struct tcp_sock *tp = tcp_sk(sk); 227405cbc0dbSFan Du struct net *net = sock_net(sk); 227505cbc0dbSFan Du u32 interval; 227605cbc0dbSFan Du s32 delta; 227705cbc0dbSFan Du 227805cbc0dbSFan Du interval = net->ipv4.sysctl_tcp_probe_interval; 2279c74df29aSEric Dumazet delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp; 228005cbc0dbSFan Du if (unlikely(delta >= interval * HZ)) { 228105cbc0dbSFan Du int mss = tcp_current_mss(sk); 228205cbc0dbSFan Du 228305cbc0dbSFan Du /* Update current search range */ 228405cbc0dbSFan Du icsk->icsk_mtup.probe_size = 0; 228505cbc0dbSFan Du icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + 228605cbc0dbSFan Du sizeof(struct tcphdr) + 228705cbc0dbSFan Du icsk->icsk_af_ops->net_header_len; 228805cbc0dbSFan Du icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss); 228905cbc0dbSFan Du 229005cbc0dbSFan Du /* Update probe time stamp */ 2291c74df29aSEric Dumazet icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; 229205cbc0dbSFan Du } 229305cbc0dbSFan Du } 229405cbc0dbSFan Du 2295808cf9e3SIlya Lesokhin static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len) 2296808cf9e3SIlya Lesokhin { 2297808cf9e3SIlya Lesokhin struct sk_buff *skb, *next; 2298808cf9e3SIlya Lesokhin 2299808cf9e3SIlya Lesokhin skb = tcp_send_head(sk); 2300808cf9e3SIlya Lesokhin tcp_for_write_queue_from_safe(skb, next, sk) { 2301808cf9e3SIlya Lesokhin if (len <= skb->len) 2302808cf9e3SIlya Lesokhin break; 2303808cf9e3SIlya Lesokhin 2304888a5c53SWillem de Bruijn if (unlikely(TCP_SKB_CB(skb)->eor) || tcp_has_tx_tstamp(skb)) 2305808cf9e3SIlya Lesokhin return false; 2306808cf9e3SIlya Lesokhin 2307808cf9e3SIlya Lesokhin len -= skb->len; 2308808cf9e3SIlya Lesokhin } 2309808cf9e3SIlya Lesokhin 2310808cf9e3SIlya Lesokhin return true; 2311808cf9e3SIlya Lesokhin } 2312808cf9e3SIlya Lesokhin 23135d424d5aSJohn Heffner /* Create a new MTU probe if we are ready. 231467edfef7SAndi Kleen * MTU probe is regularly attempting to increase the path MTU by 231567edfef7SAndi Kleen * deliberately sending larger packets. This discovers routing 231667edfef7SAndi Kleen * changes resulting in larger path MTUs. 231767edfef7SAndi Kleen * 23185d424d5aSJohn Heffner * Returns 0 if we should wait to probe (no cwnd available), 23195d424d5aSJohn Heffner * 1 if a probe was sent, 2320056834d9SIlpo Järvinen * -1 otherwise 2321056834d9SIlpo Järvinen */ 23225d424d5aSJohn Heffner static int tcp_mtu_probe(struct sock *sk) 23235d424d5aSJohn Heffner { 23245d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 232512a59abcSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 23265d424d5aSJohn Heffner struct sk_buff *skb, *nskb, *next; 23276b58e0a5SFan Du struct net *net = sock_net(sk); 23285d424d5aSJohn Heffner int probe_size; 232991cc17c0SIlpo Järvinen int size_needed; 233012a59abcSEric Dumazet int copy, len; 23315d424d5aSJohn Heffner int mss_now; 23326b58e0a5SFan Du int interval; 23335d424d5aSJohn Heffner 23345d424d5aSJohn Heffner /* Not currently probing/verifying, 23355d424d5aSJohn Heffner * not in recovery, 23365d424d5aSJohn Heffner * have enough cwnd, and 233712a59abcSEric Dumazet * not SACKing (the variable headers throw things off) 233812a59abcSEric Dumazet */ 233912a59abcSEric Dumazet if (likely(!icsk->icsk_mtup.enabled || 23405d424d5aSJohn Heffner icsk->icsk_mtup.probe_size || 23415d424d5aSJohn Heffner inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 23425d424d5aSJohn Heffner tp->snd_cwnd < 11 || 234312a59abcSEric Dumazet tp->rx_opt.num_sacks || tp->rx_opt.dsack)) 23445d424d5aSJohn Heffner return -1; 23455d424d5aSJohn Heffner 23466b58e0a5SFan Du /* Use binary search for probe_size between tcp_mss_base, 23476b58e0a5SFan Du * and current mss_clamp. if (search_high - search_low) 23486b58e0a5SFan Du * smaller than a threshold, backoff from probing. 23496b58e0a5SFan Du */ 23500c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 23516b58e0a5SFan Du probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high + 23526b58e0a5SFan Du icsk->icsk_mtup.search_low) >> 1); 235391cc17c0SIlpo Järvinen size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; 23546b58e0a5SFan Du interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low; 235505cbc0dbSFan Du /* When misfortune happens, we are reprobing actively, 235605cbc0dbSFan Du * and then reprobe timer has expired. We stick with current 235705cbc0dbSFan Du * probing process by not resetting search range to its orignal. 235805cbc0dbSFan Du */ 23596b58e0a5SFan Du if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) || 236005cbc0dbSFan Du interval < net->ipv4.sysctl_tcp_probe_threshold) { 236105cbc0dbSFan Du /* Check whether enough time has elaplased for 236205cbc0dbSFan Du * another round of probing. 236305cbc0dbSFan Du */ 236405cbc0dbSFan Du tcp_mtu_check_reprobe(sk); 23655d424d5aSJohn Heffner return -1; 23665d424d5aSJohn Heffner } 23675d424d5aSJohn Heffner 23685d424d5aSJohn Heffner /* Have enough data in the send queue to probe? */ 23697f9c33e5SIlpo Järvinen if (tp->write_seq - tp->snd_nxt < size_needed) 23705d424d5aSJohn Heffner return -1; 23715d424d5aSJohn Heffner 237291cc17c0SIlpo Järvinen if (tp->snd_wnd < size_needed) 23735d424d5aSJohn Heffner return -1; 237490840defSIlpo Järvinen if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) 23755d424d5aSJohn Heffner return 0; 23765d424d5aSJohn Heffner 2377d67c58e9SIlpo Järvinen /* Do we need to wait to drain cwnd? With none in flight, don't stall */ 2378d67c58e9SIlpo Järvinen if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { 2379d67c58e9SIlpo Järvinen if (!tcp_packets_in_flight(tp)) 23805d424d5aSJohn Heffner return -1; 23815d424d5aSJohn Heffner else 23825d424d5aSJohn Heffner return 0; 23835d424d5aSJohn Heffner } 23845d424d5aSJohn Heffner 2385808cf9e3SIlya Lesokhin if (!tcp_can_coalesce_send_queue_head(sk, probe_size)) 2386808cf9e3SIlya Lesokhin return -1; 2387808cf9e3SIlya Lesokhin 23885d424d5aSJohn Heffner /* We're allowed to probe. Build it now. */ 2389eb934478SEric Dumazet nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false); 239051456b29SIan Morris if (!nskb) 23915d424d5aSJohn Heffner return -1; 2392ab4e846aSEric Dumazet sk_wmem_queued_add(sk, nskb->truesize); 23933ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 23945d424d5aSJohn Heffner 2395fe067e8aSDavid S. Miller skb = tcp_send_head(sk); 239641477662SJakub Kicinski skb_copy_decrypted(nskb, skb); 23975d424d5aSJohn Heffner 23985d424d5aSJohn Heffner TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 23995d424d5aSJohn Heffner TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 24004de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; 24015d424d5aSJohn Heffner TCP_SKB_CB(nskb)->sacked = 0; 24025d424d5aSJohn Heffner nskb->csum = 0; 240398be9b12SEric Dumazet nskb->ip_summed = CHECKSUM_PARTIAL; 24045d424d5aSJohn Heffner 240550c4817eSIlpo Järvinen tcp_insert_write_queue_before(nskb, skb, sk); 24062b7cda9cSEric Dumazet tcp_highest_sack_replace(sk, skb, nskb); 240750c4817eSIlpo Järvinen 24085d424d5aSJohn Heffner len = 0; 2409234b6860SIlpo Järvinen tcp_for_write_queue_from_safe(skb, next, sk) { 24105d424d5aSJohn Heffner copy = min_t(int, skb->len, probe_size - len); 24115d424d5aSJohn Heffner skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); 24125d424d5aSJohn Heffner 24135d424d5aSJohn Heffner if (skb->len <= copy) { 24145d424d5aSJohn Heffner /* We've eaten all the data from this skb. 24155d424d5aSJohn Heffner * Throw it away. */ 24164de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 2417808cf9e3SIlya Lesokhin /* If this is the last SKB we copy and eor is set 2418808cf9e3SIlya Lesokhin * we need to propagate it to the new skb. 2419808cf9e3SIlya Lesokhin */ 2420808cf9e3SIlya Lesokhin TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor; 2421888a5c53SWillem de Bruijn tcp_skb_collapse_tstamp(nskb, skb); 2422fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 24233ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 24245d424d5aSJohn Heffner } else { 24254de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & 2426a3433f35SChangli Gao ~(TCPHDR_FIN|TCPHDR_PSH); 24275d424d5aSJohn Heffner if (!skb_shinfo(skb)->nr_frags) { 24285d424d5aSJohn Heffner skb_pull(skb, copy); 24295d424d5aSJohn Heffner } else { 24305d424d5aSJohn Heffner __pskb_trim_head(skb, copy); 24315bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 24325d424d5aSJohn Heffner } 24335d424d5aSJohn Heffner TCP_SKB_CB(skb)->seq += copy; 24345d424d5aSJohn Heffner } 24355d424d5aSJohn Heffner 24365d424d5aSJohn Heffner len += copy; 2437234b6860SIlpo Järvinen 2438234b6860SIlpo Järvinen if (len >= probe_size) 2439234b6860SIlpo Järvinen break; 24405d424d5aSJohn Heffner } 24415bbb432cSEric Dumazet tcp_init_tso_segs(nskb, nskb->len); 24425d424d5aSJohn Heffner 24435d424d5aSJohn Heffner /* We're ready to send. If this fails, the probe will 24447faee5c0SEric Dumazet * be resegmented into mss-sized pieces by tcp_write_xmit(). 24457faee5c0SEric Dumazet */ 24465d424d5aSJohn Heffner if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 24475d424d5aSJohn Heffner /* Decrement cwnd here because we are sending 24485d424d5aSJohn Heffner * effectively two packets. */ 24495d424d5aSJohn Heffner tp->snd_cwnd--; 245066f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, nskb); 24515d424d5aSJohn Heffner 24525d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 24530e7b1368SJohn Heffner tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 24540e7b1368SJohn Heffner tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 24555d424d5aSJohn Heffner 24565d424d5aSJohn Heffner return 1; 24575d424d5aSJohn Heffner } 24585d424d5aSJohn Heffner 24595d424d5aSJohn Heffner return -1; 24605d424d5aSJohn Heffner } 24615d424d5aSJohn Heffner 2462864e5c09SEric Dumazet static bool tcp_pacing_check(struct sock *sk) 2463218af599SEric Dumazet { 2464864e5c09SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 2465864e5c09SEric Dumazet 2466864e5c09SEric Dumazet if (!tcp_needs_internal_pacing(sk)) 2467864e5c09SEric Dumazet return false; 2468864e5c09SEric Dumazet 2469864e5c09SEric Dumazet if (tp->tcp_wstamp_ns <= tp->tcp_clock_cache) 2470864e5c09SEric Dumazet return false; 2471864e5c09SEric Dumazet 2472864e5c09SEric Dumazet if (!hrtimer_is_queued(&tp->pacing_timer)) { 2473864e5c09SEric Dumazet hrtimer_start(&tp->pacing_timer, 2474864e5c09SEric Dumazet ns_to_ktime(tp->tcp_wstamp_ns), 2475864e5c09SEric Dumazet HRTIMER_MODE_ABS_PINNED_SOFT); 2476864e5c09SEric Dumazet sock_hold(sk); 2477864e5c09SEric Dumazet } 2478864e5c09SEric Dumazet return true; 2479218af599SEric Dumazet } 2480218af599SEric Dumazet 2481f9616c35SEric Dumazet /* TCP Small Queues : 2482f9616c35SEric Dumazet * Control number of packets in qdisc/devices to two packets / or ~1 ms. 2483f9616c35SEric Dumazet * (These limits are doubled for retransmits) 2484f9616c35SEric Dumazet * This allows for : 2485f9616c35SEric Dumazet * - better RTT estimation and ACK scheduling 2486f9616c35SEric Dumazet * - faster recovery 2487f9616c35SEric Dumazet * - high rates 2488f9616c35SEric Dumazet * Alas, some drivers / subsystems require a fair amount 2489f9616c35SEric Dumazet * of queued bytes to ensure line rate. 2490f9616c35SEric Dumazet * One example is wifi aggregation (802.11 AMPDU) 2491f9616c35SEric Dumazet */ 2492f9616c35SEric Dumazet static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb, 2493f9616c35SEric Dumazet unsigned int factor) 2494f9616c35SEric Dumazet { 249576a9ebe8SEric Dumazet unsigned long limit; 2496f9616c35SEric Dumazet 249776a9ebe8SEric Dumazet limit = max_t(unsigned long, 249876a9ebe8SEric Dumazet 2 * skb->truesize, 24997c68fa2bSEric Dumazet sk->sk_pacing_rate >> READ_ONCE(sk->sk_pacing_shift)); 2500c73e5807SEric Dumazet if (sk->sk_pacing_status == SK_PACING_NONE) 250176a9ebe8SEric Dumazet limit = min_t(unsigned long, limit, 25029184d8bbSEric Dumazet sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes); 2503f9616c35SEric Dumazet limit <<= factor; 2504f9616c35SEric Dumazet 2505a842fe14SEric Dumazet if (static_branch_unlikely(&tcp_tx_delay_enabled) && 2506a842fe14SEric Dumazet tcp_sk(sk)->tcp_tx_delay) { 2507a842fe14SEric Dumazet u64 extra_bytes = (u64)sk->sk_pacing_rate * tcp_sk(sk)->tcp_tx_delay; 2508a842fe14SEric Dumazet 2509a842fe14SEric Dumazet /* TSQ is based on skb truesize sum (sk_wmem_alloc), so we 2510a842fe14SEric Dumazet * approximate our needs assuming an ~100% skb->truesize overhead. 2511a842fe14SEric Dumazet * USEC_PER_SEC is approximated by 2^20. 2512a842fe14SEric Dumazet * do_div(extra_bytes, USEC_PER_SEC/2) is replaced by a right shift. 2513a842fe14SEric Dumazet */ 2514a842fe14SEric Dumazet extra_bytes >>= (20 - 1); 2515a842fe14SEric Dumazet limit += extra_bytes; 2516a842fe14SEric Dumazet } 251714afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) > limit) { 251875c119afSEric Dumazet /* Always send skb if rtx queue is empty. 251975eefc6cSEric Dumazet * No need to wait for TX completion to call us back, 252075eefc6cSEric Dumazet * after softirq/tasklet schedule. 252175eefc6cSEric Dumazet * This helps when TX completions are delayed too much. 252275eefc6cSEric Dumazet */ 252375c119afSEric Dumazet if (tcp_rtx_queue_empty(sk)) 252475eefc6cSEric Dumazet return false; 252575eefc6cSEric Dumazet 25267aa5470cSEric Dumazet set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags); 2527f9616c35SEric Dumazet /* It is possible TX completion already happened 2528f9616c35SEric Dumazet * before we set TSQ_THROTTLED, so we must 2529f9616c35SEric Dumazet * test again the condition. 2530f9616c35SEric Dumazet */ 2531f9616c35SEric Dumazet smp_mb__after_atomic(); 253214afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) > limit) 2533f9616c35SEric Dumazet return true; 2534f9616c35SEric Dumazet } 2535f9616c35SEric Dumazet return false; 2536f9616c35SEric Dumazet } 2537f9616c35SEric Dumazet 253805b055e8SFrancis Yan static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new) 253905b055e8SFrancis Yan { 2540628174ccSEric Dumazet const u32 now = tcp_jiffies32; 2541efe967cdSArnd Bergmann enum tcp_chrono old = tp->chrono_type; 254205b055e8SFrancis Yan 2543efe967cdSArnd Bergmann if (old > TCP_CHRONO_UNSPEC) 2544efe967cdSArnd Bergmann tp->chrono_stat[old - 1] += now - tp->chrono_start; 254505b055e8SFrancis Yan tp->chrono_start = now; 254605b055e8SFrancis Yan tp->chrono_type = new; 254705b055e8SFrancis Yan } 254805b055e8SFrancis Yan 254905b055e8SFrancis Yan void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type) 255005b055e8SFrancis Yan { 255105b055e8SFrancis Yan struct tcp_sock *tp = tcp_sk(sk); 255205b055e8SFrancis Yan 255305b055e8SFrancis Yan /* If there are multiple conditions worthy of tracking in a 25540f87230dSFrancis Yan * chronograph then the highest priority enum takes precedence 25550f87230dSFrancis Yan * over the other conditions. So that if something "more interesting" 255605b055e8SFrancis Yan * starts happening, stop the previous chrono and start a new one. 255705b055e8SFrancis Yan */ 255805b055e8SFrancis Yan if (type > tp->chrono_type) 255905b055e8SFrancis Yan tcp_chrono_set(tp, type); 256005b055e8SFrancis Yan } 256105b055e8SFrancis Yan 256205b055e8SFrancis Yan void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type) 256305b055e8SFrancis Yan { 256405b055e8SFrancis Yan struct tcp_sock *tp = tcp_sk(sk); 256505b055e8SFrancis Yan 25660f87230dSFrancis Yan 25670f87230dSFrancis Yan /* There are multiple conditions worthy of tracking in a 25680f87230dSFrancis Yan * chronograph, so that the highest priority enum takes 25690f87230dSFrancis Yan * precedence over the other conditions (see tcp_chrono_start). 25700f87230dSFrancis Yan * If a condition stops, we only stop chrono tracking if 25710f87230dSFrancis Yan * it's the "most interesting" or current chrono we are 25720f87230dSFrancis Yan * tracking and starts busy chrono if we have pending data. 25730f87230dSFrancis Yan */ 257475c119afSEric Dumazet if (tcp_rtx_and_write_queues_empty(sk)) 257505b055e8SFrancis Yan tcp_chrono_set(tp, TCP_CHRONO_UNSPEC); 25760f87230dSFrancis Yan else if (type == tp->chrono_type) 25770f87230dSFrancis Yan tcp_chrono_set(tp, TCP_CHRONO_BUSY); 257805b055e8SFrancis Yan } 257905b055e8SFrancis Yan 25801da177e4SLinus Torvalds /* This routine writes packets to the network. It advances the 25811da177e4SLinus Torvalds * send_head. This happens as incoming acks open up the remote 25821da177e4SLinus Torvalds * window for us. 25831da177e4SLinus Torvalds * 2584f8269a49SIlpo Järvinen * LARGESEND note: !tcp_urg_mode is overkill, only frames between 2585f8269a49SIlpo Järvinen * snd_up-64k-mss .. snd_up cannot be large. However, taking into 2586f8269a49SIlpo Järvinen * account rare use of URG, this is not a big flaw. 2587f8269a49SIlpo Järvinen * 25886ba8a3b1SNandita Dukkipati * Send at most one packet when push_one > 0. Temporarily ignore 25896ba8a3b1SNandita Dukkipati * cwnd limit to force at most one packet out when push_one == 2. 25906ba8a3b1SNandita Dukkipati 2591a2a385d6SEric Dumazet * Returns true, if no segments are in flight and we have queued segments, 2592a2a385d6SEric Dumazet * but cannot send anything now because of SWS or another problem. 25931da177e4SLinus Torvalds */ 2594a2a385d6SEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 2595d5dd9175SIlpo Järvinen int push_one, gfp_t gfp) 25961da177e4SLinus Torvalds { 25971da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 259892df7b51SDavid S. Miller struct sk_buff *skb; 2599c1b4a7e6SDavid S. Miller unsigned int tso_segs, sent_pkts; 2600c1b4a7e6SDavid S. Miller int cwnd_quota; 26015d424d5aSJohn Heffner int result; 26025615f886SFrancis Yan bool is_cwnd_limited = false, is_rwnd_limited = false; 2603605ad7f1SEric Dumazet u32 max_segs; 26041da177e4SLinus Torvalds 2605c1b4a7e6SDavid S. Miller sent_pkts = 0; 26065d424d5aSJohn Heffner 2607ee1836aeSEric Dumazet tcp_mstamp_refresh(tp); 2608d5dd9175SIlpo Järvinen if (!push_one) { 26095d424d5aSJohn Heffner /* Do MTU probing. */ 2610d5dd9175SIlpo Järvinen result = tcp_mtu_probe(sk); 2611d5dd9175SIlpo Järvinen if (!result) { 2612a2a385d6SEric Dumazet return false; 26135d424d5aSJohn Heffner } else if (result > 0) { 26145d424d5aSJohn Heffner sent_pkts = 1; 26155d424d5aSJohn Heffner } 2616d5dd9175SIlpo Järvinen } 26175d424d5aSJohn Heffner 2618ed6e7268SNeal Cardwell max_segs = tcp_tso_segs(sk, mss_now); 2619fe067e8aSDavid S. Miller while ((skb = tcp_send_head(sk))) { 2620c8ac3774SHerbert Xu unsigned int limit; 2621c8ac3774SHerbert Xu 262279861919SEric Dumazet if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) { 262379861919SEric Dumazet /* "skb_mstamp_ns" is used as a start point for the retransmit timer */ 262479861919SEric Dumazet skb->skb_mstamp_ns = tp->tcp_wstamp_ns = tp->tcp_clock_cache; 262579861919SEric Dumazet list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue); 2626bf50b606SEric Dumazet tcp_init_tso_segs(skb, mss_now); 262779861919SEric Dumazet goto repair; /* Skip network transmission */ 262879861919SEric Dumazet } 262979861919SEric Dumazet 2630218af599SEric Dumazet if (tcp_pacing_check(sk)) 2631218af599SEric Dumazet break; 2632218af599SEric Dumazet 26335bbb432cSEric Dumazet tso_segs = tcp_init_tso_segs(skb, mss_now); 2634c1b4a7e6SDavid S. Miller BUG_ON(!tso_segs); 2635c1b4a7e6SDavid S. Miller 2636b68e9f85SHerbert Xu cwnd_quota = tcp_cwnd_test(tp, skb); 26376ba8a3b1SNandita Dukkipati if (!cwnd_quota) { 26386ba8a3b1SNandita Dukkipati if (push_one == 2) 26396ba8a3b1SNandita Dukkipati /* Force out a loss probe pkt. */ 26406ba8a3b1SNandita Dukkipati cwnd_quota = 1; 26416ba8a3b1SNandita Dukkipati else 2642b68e9f85SHerbert Xu break; 26436ba8a3b1SNandita Dukkipati } 2644b68e9f85SHerbert Xu 26455615f886SFrancis Yan if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) { 26465615f886SFrancis Yan is_rwnd_limited = true; 2647b68e9f85SHerbert Xu break; 26485615f886SFrancis Yan } 2649b68e9f85SHerbert Xu 2650d6a4e26aSEric Dumazet if (tso_segs == 1) { 2651aa93466bSDavid S. Miller if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 2652aa93466bSDavid S. Miller (tcp_skb_is_last(sk, skb) ? 2653aa93466bSDavid S. Miller nonagle : TCP_NAGLE_PUSH)))) 2654aa93466bSDavid S. Miller break; 2655c1b4a7e6SDavid S. Miller } else { 2656ca8a2263SNeal Cardwell if (!push_one && 2657605ad7f1SEric Dumazet tcp_tso_should_defer(sk, skb, &is_cwnd_limited, 2658f9bfe4e6SEric Dumazet &is_rwnd_limited, max_segs)) 2659aa93466bSDavid S. Miller break; 2660c1b4a7e6SDavid S. Miller } 2661aa93466bSDavid S. Miller 2662605ad7f1SEric Dumazet limit = mss_now; 2663d6a4e26aSEric Dumazet if (tso_segs > 1 && !tcp_urg_mode(tp)) 2664605ad7f1SEric Dumazet limit = tcp_mss_split_point(sk, skb, mss_now, 2665605ad7f1SEric Dumazet min_t(unsigned int, 2666605ad7f1SEric Dumazet cwnd_quota, 2667605ad7f1SEric Dumazet max_segs), 2668605ad7f1SEric Dumazet nonagle); 2669605ad7f1SEric Dumazet 2670605ad7f1SEric Dumazet if (skb->len > limit && 267156483341SEric Dumazet unlikely(tso_fragment(sk, skb, limit, mss_now, gfp))) 2672605ad7f1SEric Dumazet break; 2673605ad7f1SEric Dumazet 2674f9616c35SEric Dumazet if (tcp_small_queue_check(sk, skb, 0)) 267546d3ceabSEric Dumazet break; 2676c9eeec26SEric Dumazet 26771f85e626SEric Dumazet /* Argh, we hit an empty skb(), presumably a thread 26781f85e626SEric Dumazet * is sleeping in sendmsg()/sk_stream_wait_memory(). 26791f85e626SEric Dumazet * We do not want to send a pure-ack packet and have 26801f85e626SEric Dumazet * a strange looking rtx queue with empty packet(s). 26811f85e626SEric Dumazet */ 26821f85e626SEric Dumazet if (TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq) 26831f85e626SEric Dumazet break; 26841f85e626SEric Dumazet 2685d5dd9175SIlpo Järvinen if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) 26861da177e4SLinus Torvalds break; 26871da177e4SLinus Torvalds 2688ec342325SAndrew Vagin repair: 26891da177e4SLinus Torvalds /* Advance the send_head. This one is sent out. 26901da177e4SLinus Torvalds * This call will increment packets_out. 26911da177e4SLinus Torvalds */ 269266f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 26931da177e4SLinus Torvalds 26941da177e4SLinus Torvalds tcp_minshall_update(tp, mss_now, skb); 2695a262f0cdSNandita Dukkipati sent_pkts += tcp_skb_pcount(skb); 2696d5dd9175SIlpo Järvinen 2697d5dd9175SIlpo Järvinen if (push_one) 2698d5dd9175SIlpo Järvinen break; 26991da177e4SLinus Torvalds } 27001da177e4SLinus Torvalds 27015615f886SFrancis Yan if (is_rwnd_limited) 27025615f886SFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED); 27035615f886SFrancis Yan else 27045615f886SFrancis Yan tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED); 27055615f886SFrancis Yan 2706aa93466bSDavid S. Miller if (likely(sent_pkts)) { 2707684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 2708684bad11SYuchung Cheng tp->prr_out += sent_pkts; 27096ba8a3b1SNandita Dukkipati 27106ba8a3b1SNandita Dukkipati /* Send one loss probe per tail loss episode. */ 27116ba8a3b1SNandita Dukkipati if (push_one != 2) 2712ed66dfafSNeal Cardwell tcp_schedule_loss_probe(sk, false); 2713d2e1339fSBendik Rønning Opstad is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd); 2714ca8a2263SNeal Cardwell tcp_cwnd_validate(sk, is_cwnd_limited); 2715a2a385d6SEric Dumazet return false; 27161da177e4SLinus Torvalds } 271775c119afSEric Dumazet return !tp->packets_out && !tcp_write_queue_empty(sk); 27186ba8a3b1SNandita Dukkipati } 27196ba8a3b1SNandita Dukkipati 2720ed66dfafSNeal Cardwell bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto) 27216ba8a3b1SNandita Dukkipati { 27226ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 27236ba8a3b1SNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 2724a2815817SNeal Cardwell u32 timeout, rto_delta_us; 27252ae21cf5SEric Dumazet int early_retrans; 27266ba8a3b1SNandita Dukkipati 27276ba8a3b1SNandita Dukkipati /* Don't do any loss probe on a Fast Open connection before 3WHS 27286ba8a3b1SNandita Dukkipati * finishes. 27296ba8a3b1SNandita Dukkipati */ 2730d983ea6fSEric Dumazet if (rcu_access_pointer(tp->fastopen_rsk)) 27316ba8a3b1SNandita Dukkipati return false; 27326ba8a3b1SNandita Dukkipati 27332ae21cf5SEric Dumazet early_retrans = sock_net(sk)->ipv4.sysctl_tcp_early_retrans; 27346ba8a3b1SNandita Dukkipati /* Schedule a loss probe in 2*RTT for SACK capable connections 2735b4f70c3dSNeal Cardwell * not in loss recovery, that are either limited by cwnd or application. 27366ba8a3b1SNandita Dukkipati */ 27372ae21cf5SEric Dumazet if ((early_retrans != 3 && early_retrans != 4) || 2738bec41a11SYuchung Cheng !tp->packets_out || !tcp_is_sack(tp) || 2739b4f70c3dSNeal Cardwell (icsk->icsk_ca_state != TCP_CA_Open && 2740b4f70c3dSNeal Cardwell icsk->icsk_ca_state != TCP_CA_CWR)) 27416ba8a3b1SNandita Dukkipati return false; 27426ba8a3b1SNandita Dukkipati 2743bb4d991aSYuchung Cheng /* Probe timeout is 2*rtt. Add minimum RTO to account 2744f9b99582SYuchung Cheng * for delayed ack when there's one outstanding packet. If no RTT 2745f9b99582SYuchung Cheng * sample is available then probe after TCP_TIMEOUT_INIT. 27466ba8a3b1SNandita Dukkipati */ 2747bb4d991aSYuchung Cheng if (tp->srtt_us) { 2748bb4d991aSYuchung Cheng timeout = usecs_to_jiffies(tp->srtt_us >> 2); 27496ba8a3b1SNandita Dukkipati if (tp->packets_out == 1) 2750bb4d991aSYuchung Cheng timeout += TCP_RTO_MIN; 2751bb4d991aSYuchung Cheng else 2752bb4d991aSYuchung Cheng timeout += TCP_TIMEOUT_MIN; 2753bb4d991aSYuchung Cheng } else { 2754bb4d991aSYuchung Cheng timeout = TCP_TIMEOUT_INIT; 2755bb4d991aSYuchung Cheng } 27566ba8a3b1SNandita Dukkipati 2757a2815817SNeal Cardwell /* If the RTO formula yields an earlier time, then use that time. */ 2758ed66dfafSNeal Cardwell rto_delta_us = advancing_rto ? 2759ed66dfafSNeal Cardwell jiffies_to_usecs(inet_csk(sk)->icsk_rto) : 2760ed66dfafSNeal Cardwell tcp_rto_delta_us(sk); /* How far in future is RTO? */ 2761a2815817SNeal Cardwell if (rto_delta_us > 0) 2762a2815817SNeal Cardwell timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us)); 27636ba8a3b1SNandita Dukkipati 27648dc242adSEric Dumazet tcp_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, TCP_RTO_MAX); 27656ba8a3b1SNandita Dukkipati return true; 27666ba8a3b1SNandita Dukkipati } 27676ba8a3b1SNandita Dukkipati 27681f3279aeSEric Dumazet /* Thanks to skb fast clones, we can detect if a prior transmit of 27691f3279aeSEric Dumazet * a packet is still in a qdisc or driver queue. 27701f3279aeSEric Dumazet * In this case, there is very little point doing a retransmit ! 27711f3279aeSEric Dumazet */ 27721f3279aeSEric Dumazet static bool skb_still_in_host_queue(const struct sock *sk, 27731f3279aeSEric Dumazet const struct sk_buff *skb) 27741f3279aeSEric Dumazet { 277539bb5e62SEric Dumazet if (unlikely(skb_fclone_busy(sk, skb))) { 2776c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), 27771f3279aeSEric Dumazet LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES); 27781f3279aeSEric Dumazet return true; 27791f3279aeSEric Dumazet } 27801f3279aeSEric Dumazet return false; 27811f3279aeSEric Dumazet } 27821f3279aeSEric Dumazet 2783b340b264SYuchung Cheng /* When probe timeout (PTO) fires, try send a new segment if possible, else 27846ba8a3b1SNandita Dukkipati * retransmit the last segment. 27856ba8a3b1SNandita Dukkipati */ 27866ba8a3b1SNandita Dukkipati void tcp_send_loss_probe(struct sock *sk) 27876ba8a3b1SNandita Dukkipati { 27889b717a8dSNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 27896ba8a3b1SNandita Dukkipati struct sk_buff *skb; 27906ba8a3b1SNandita Dukkipati int pcount; 27916ba8a3b1SNandita Dukkipati int mss = tcp_current_mss(sk); 27926ba8a3b1SNandita Dukkipati 279376be93fcSYuchung Cheng /* At most one outstanding TLP */ 279476be93fcSYuchung Cheng if (tp->tlp_high_seq) 279576be93fcSYuchung Cheng goto rearm_timer; 279676be93fcSYuchung Cheng 279776be93fcSYuchung Cheng tp->tlp_retrans = 0; 2798b340b264SYuchung Cheng skb = tcp_send_head(sk); 279975c119afSEric Dumazet if (skb && tcp_snd_wnd_test(tp, skb, mss)) { 2800b340b264SYuchung Cheng pcount = tp->packets_out; 2801b340b264SYuchung Cheng tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC); 2802b340b264SYuchung Cheng if (tp->packets_out > pcount) 2803b340b264SYuchung Cheng goto probe_sent; 28046ba8a3b1SNandita Dukkipati goto rearm_timer; 28056ba8a3b1SNandita Dukkipati } 280675c119afSEric Dumazet skb = skb_rb_last(&sk->tcp_rtx_queue); 2807b2b7af86SYuchung Cheng if (unlikely(!skb)) { 2808b2b7af86SYuchung Cheng WARN_ONCE(tp->packets_out, 2809b2b7af86SYuchung Cheng "invalid inflight: %u state %u cwnd %u mss %d\n", 2810b2b7af86SYuchung Cheng tp->packets_out, sk->sk_state, tp->snd_cwnd, mss); 2811b2b7af86SYuchung Cheng inet_csk(sk)->icsk_pending = 0; 2812b2b7af86SYuchung Cheng return; 2813b2b7af86SYuchung Cheng } 28146ba8a3b1SNandita Dukkipati 28151f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 28161f3279aeSEric Dumazet goto rearm_timer; 28171f3279aeSEric Dumazet 28186ba8a3b1SNandita Dukkipati pcount = tcp_skb_pcount(skb); 28196ba8a3b1SNandita Dukkipati if (WARN_ON(!pcount)) 28206ba8a3b1SNandita Dukkipati goto rearm_timer; 28216ba8a3b1SNandita Dukkipati 28226ba8a3b1SNandita Dukkipati if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) { 282375c119afSEric Dumazet if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, 282475c119afSEric Dumazet (pcount - 1) * mss, mss, 28256cc55e09SOctavian Purdila GFP_ATOMIC))) 28266ba8a3b1SNandita Dukkipati goto rearm_timer; 282775c119afSEric Dumazet skb = skb_rb_next(skb); 28286ba8a3b1SNandita Dukkipati } 28296ba8a3b1SNandita Dukkipati 28306ba8a3b1SNandita Dukkipati if (WARN_ON(!skb || !tcp_skb_pcount(skb))) 28316ba8a3b1SNandita Dukkipati goto rearm_timer; 28326ba8a3b1SNandita Dukkipati 283310d3be56SEric Dumazet if (__tcp_retransmit_skb(sk, skb, 1)) 2834b340b264SYuchung Cheng goto rearm_timer; 28356ba8a3b1SNandita Dukkipati 283676be93fcSYuchung Cheng tp->tlp_retrans = 1; 283776be93fcSYuchung Cheng 283876be93fcSYuchung Cheng probe_sent: 28399b717a8dSNandita Dukkipati /* Record snd_nxt for loss detection. */ 28409b717a8dSNandita Dukkipati tp->tlp_high_seq = tp->snd_nxt; 28419b717a8dSNandita Dukkipati 2842c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES); 2843fcd16c0aSYuchung Cheng /* Reset s.t. tcp_rearm_rto will restart timer from now */ 2844fcd16c0aSYuchung Cheng inet_csk(sk)->icsk_pending = 0; 2845b340b264SYuchung Cheng rearm_timer: 2846fcd16c0aSYuchung Cheng tcp_rearm_rto(sk); 28471da177e4SLinus Torvalds } 28481da177e4SLinus Torvalds 2849a762a980SDavid S. Miller /* Push out any pending frames which were held back due to 2850a762a980SDavid S. Miller * TCP_CORK or attempt at coalescing tiny packets. 2851a762a980SDavid S. Miller * The socket must be locked by the caller. 2852a762a980SDavid S. Miller */ 28539e412ba7SIlpo Järvinen void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 28549e412ba7SIlpo Järvinen int nonagle) 2855a762a980SDavid S. Miller { 2856726e07a8SIlpo Järvinen /* If we are closed, the bytes will have to remain here. 2857726e07a8SIlpo Järvinen * In time closedown will finish, we empty the write queue and 2858726e07a8SIlpo Järvinen * all will be happy. 2859726e07a8SIlpo Järvinen */ 2860726e07a8SIlpo Järvinen if (unlikely(sk->sk_state == TCP_CLOSE)) 2861726e07a8SIlpo Järvinen return; 2862726e07a8SIlpo Järvinen 286399a1dec7SMel Gorman if (tcp_write_xmit(sk, cur_mss, nonagle, 0, 28647450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC))) 28659e412ba7SIlpo Järvinen tcp_check_probe_timer(sk); 2866a762a980SDavid S. Miller } 2867a762a980SDavid S. Miller 2868c1b4a7e6SDavid S. Miller /* Send _single_ skb sitting at the send head. This function requires 2869c1b4a7e6SDavid S. Miller * true push pending frames to setup probe timer etc. 2870c1b4a7e6SDavid S. Miller */ 2871c1b4a7e6SDavid S. Miller void tcp_push_one(struct sock *sk, unsigned int mss_now) 2872c1b4a7e6SDavid S. Miller { 2873fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 2874c1b4a7e6SDavid S. Miller 2875c1b4a7e6SDavid S. Miller BUG_ON(!skb || skb->len < mss_now); 2876c1b4a7e6SDavid S. Miller 2877d5dd9175SIlpo Järvinen tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); 2878c1b4a7e6SDavid S. Miller } 2879c1b4a7e6SDavid S. Miller 28801da177e4SLinus Torvalds /* This function returns the amount that we can raise the 28811da177e4SLinus Torvalds * usable window based on the following constraints 28821da177e4SLinus Torvalds * 28831da177e4SLinus Torvalds * 1. The window can never be shrunk once it is offered (RFC 793) 28841da177e4SLinus Torvalds * 2. We limit memory per socket 28851da177e4SLinus Torvalds * 28861da177e4SLinus Torvalds * RFC 1122: 28871da177e4SLinus Torvalds * "the suggested [SWS] avoidance algorithm for the receiver is to keep 28881da177e4SLinus Torvalds * RECV.NEXT + RCV.WIN fixed until: 28891da177e4SLinus Torvalds * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 28901da177e4SLinus Torvalds * 28911da177e4SLinus Torvalds * i.e. don't raise the right edge of the window until you can raise 28921da177e4SLinus Torvalds * it at least MSS bytes. 28931da177e4SLinus Torvalds * 28941da177e4SLinus Torvalds * Unfortunately, the recommended algorithm breaks header prediction, 28951da177e4SLinus Torvalds * since header prediction assumes th->window stays fixed. 28961da177e4SLinus Torvalds * 28971da177e4SLinus Torvalds * Strictly speaking, keeping th->window fixed violates the receiver 28981da177e4SLinus Torvalds * side SWS prevention criteria. The problem is that under this rule 28991da177e4SLinus Torvalds * a stream of single byte packets will cause the right side of the 29001da177e4SLinus Torvalds * window to always advance by a single byte. 29011da177e4SLinus Torvalds * 29021da177e4SLinus Torvalds * Of course, if the sender implements sender side SWS prevention 29031da177e4SLinus Torvalds * then this will not be a problem. 29041da177e4SLinus Torvalds * 29051da177e4SLinus Torvalds * BSD seems to make the following compromise: 29061da177e4SLinus Torvalds * 29071da177e4SLinus Torvalds * If the free space is less than the 1/4 of the maximum 29081da177e4SLinus Torvalds * space available and the free space is less than 1/2 mss, 29091da177e4SLinus Torvalds * then set the window to 0. 29101da177e4SLinus Torvalds * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 29111da177e4SLinus Torvalds * Otherwise, just prevent the window from shrinking 29121da177e4SLinus Torvalds * and from being larger than the largest representable value. 29131da177e4SLinus Torvalds * 29141da177e4SLinus Torvalds * This prevents incremental opening of the window in the regime 29151da177e4SLinus Torvalds * where TCP is limited by the speed of the reader side taking 29161da177e4SLinus Torvalds * data out of the TCP receive queue. It does nothing about 29171da177e4SLinus Torvalds * those cases where the window is constrained on the sender side 29181da177e4SLinus Torvalds * because the pipeline is full. 29191da177e4SLinus Torvalds * 29201da177e4SLinus Torvalds * BSD also seems to "accidentally" limit itself to windows that are a 29211da177e4SLinus Torvalds * multiple of MSS, at least until the free space gets quite small. 29221da177e4SLinus Torvalds * This would appear to be a side effect of the mbuf implementation. 29231da177e4SLinus Torvalds * Combining these two algorithms results in the observed behavior 29241da177e4SLinus Torvalds * of having a fixed window size at almost all times. 29251da177e4SLinus Torvalds * 29261da177e4SLinus Torvalds * Below we obtain similar behavior by forcing the offered window to 29271da177e4SLinus Torvalds * a multiple of the mss when it is feasible to do so. 29281da177e4SLinus Torvalds * 29291da177e4SLinus Torvalds * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 29301da177e4SLinus Torvalds * Regular options like TIMESTAMP are taken into account. 29311da177e4SLinus Torvalds */ 29321da177e4SLinus Torvalds u32 __tcp_select_window(struct sock *sk) 29331da177e4SLinus Torvalds { 2934463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 29351da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2936caa20d9aSStephen Hemminger /* MSS for the peer's data. Previous versions used mss_clamp 29371da177e4SLinus Torvalds * here. I don't know if the value based on our guesses 29381da177e4SLinus Torvalds * of peer's MSS is better for the performance. It's more correct 29391da177e4SLinus Torvalds * but may be worse for the performance because of rcv_mss 29401da177e4SLinus Torvalds * fluctuations. --SAW 1998/11/1 29411da177e4SLinus Torvalds */ 2942463c84b9SArnaldo Carvalho de Melo int mss = icsk->icsk_ack.rcv_mss; 29431da177e4SLinus Torvalds int free_space = tcp_space(sk); 294486c1a045SFlorian Westphal int allowed_space = tcp_full_space(sk); 2945071c8ed6SFlorian Westphal int full_space, window; 2946071c8ed6SFlorian Westphal 2947071c8ed6SFlorian Westphal if (sk_is_mptcp(sk)) 2948071c8ed6SFlorian Westphal mptcp_space(sk, &free_space, &allowed_space); 2949071c8ed6SFlorian Westphal 2950071c8ed6SFlorian Westphal full_space = min_t(int, tp->window_clamp, allowed_space); 29511da177e4SLinus Torvalds 295206425c30SEric Dumazet if (unlikely(mss > full_space)) { 29531da177e4SLinus Torvalds mss = full_space; 295406425c30SEric Dumazet if (mss <= 0) 295506425c30SEric Dumazet return 0; 295606425c30SEric Dumazet } 2957b92edbe0SEric Dumazet if (free_space < (full_space >> 1)) { 2958463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.quick = 0; 29591da177e4SLinus Torvalds 2960b8da51ebSEric Dumazet if (tcp_under_memory_pressure(sk)) 2961056834d9SIlpo Järvinen tp->rcv_ssthresh = min(tp->rcv_ssthresh, 2962056834d9SIlpo Järvinen 4U * tp->advmss); 29631da177e4SLinus Torvalds 296486c1a045SFlorian Westphal /* free_space might become our new window, make sure we don't 296586c1a045SFlorian Westphal * increase it due to wscale. 296686c1a045SFlorian Westphal */ 296786c1a045SFlorian Westphal free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale); 296886c1a045SFlorian Westphal 296986c1a045SFlorian Westphal /* if free space is less than mss estimate, or is below 1/16th 297086c1a045SFlorian Westphal * of the maximum allowed, try to move to zero-window, else 297186c1a045SFlorian Westphal * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and 297286c1a045SFlorian Westphal * new incoming data is dropped due to memory limits. 297386c1a045SFlorian Westphal * With large window, mss test triggers way too late in order 297486c1a045SFlorian Westphal * to announce zero window in time before rmem limit kicks in. 297586c1a045SFlorian Westphal */ 297686c1a045SFlorian Westphal if (free_space < (allowed_space >> 4) || free_space < mss) 29771da177e4SLinus Torvalds return 0; 29781da177e4SLinus Torvalds } 29791da177e4SLinus Torvalds 29801da177e4SLinus Torvalds if (free_space > tp->rcv_ssthresh) 29811da177e4SLinus Torvalds free_space = tp->rcv_ssthresh; 29821da177e4SLinus Torvalds 29831da177e4SLinus Torvalds /* Don't do rounding if we are using window scaling, since the 29841da177e4SLinus Torvalds * scaled window will not line up with the MSS boundary anyway. 29851da177e4SLinus Torvalds */ 29861da177e4SLinus Torvalds if (tp->rx_opt.rcv_wscale) { 29871da177e4SLinus Torvalds window = free_space; 29881da177e4SLinus Torvalds 29891da177e4SLinus Torvalds /* Advertise enough space so that it won't get scaled away. 29901da177e4SLinus Torvalds * Import case: prevent zero window announcement if 29911da177e4SLinus Torvalds * 1<<rcv_wscale > mss. 29921da177e4SLinus Torvalds */ 29931935299dSGao Feng window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale)); 29941da177e4SLinus Torvalds } else { 29951935299dSGao Feng window = tp->rcv_wnd; 29961da177e4SLinus Torvalds /* Get the largest window that is a nice multiple of mss. 29971da177e4SLinus Torvalds * Window clamp already applied above. 29981da177e4SLinus Torvalds * If our current window offering is within 1 mss of the 29991da177e4SLinus Torvalds * free space we just keep it. This prevents the divide 30001da177e4SLinus Torvalds * and multiply from happening most of the time. 30011da177e4SLinus Torvalds * We also don't do any window rounding when the free space 30021da177e4SLinus Torvalds * is too small. 30031da177e4SLinus Torvalds */ 30041da177e4SLinus Torvalds if (window <= free_space - mss || window > free_space) 30051935299dSGao Feng window = rounddown(free_space, mss); 300684565070SJohn Heffner else if (mss == full_space && 3007b92edbe0SEric Dumazet free_space > window + (full_space >> 1)) 300884565070SJohn Heffner window = free_space; 30091da177e4SLinus Torvalds } 30101da177e4SLinus Torvalds 30111da177e4SLinus Torvalds return window; 30121da177e4SLinus Torvalds } 30131da177e4SLinus Torvalds 3014cfea5a68SMartin KaFai Lau void tcp_skb_collapse_tstamp(struct sk_buff *skb, 3015082ac2d5SMartin KaFai Lau const struct sk_buff *next_skb) 3016082ac2d5SMartin KaFai Lau { 30170a2cf20cSSoheil Hassas Yeganeh if (unlikely(tcp_has_tx_tstamp(next_skb))) { 30180a2cf20cSSoheil Hassas Yeganeh const struct skb_shared_info *next_shinfo = 30190a2cf20cSSoheil Hassas Yeganeh skb_shinfo(next_skb); 3020082ac2d5SMartin KaFai Lau struct skb_shared_info *shinfo = skb_shinfo(skb); 3021082ac2d5SMartin KaFai Lau 30220a2cf20cSSoheil Hassas Yeganeh shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP; 3023082ac2d5SMartin KaFai Lau shinfo->tskey = next_shinfo->tskey; 30242de8023eSMartin KaFai Lau TCP_SKB_CB(skb)->txstamp_ack |= 30252de8023eSMartin KaFai Lau TCP_SKB_CB(next_skb)->txstamp_ack; 3026082ac2d5SMartin KaFai Lau } 3027082ac2d5SMartin KaFai Lau } 3028082ac2d5SMartin KaFai Lau 30294a17fc3aSIlpo Järvinen /* Collapses two adjacent SKB's during retransmission. */ 3030f8071cdeSEric Dumazet static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) 30311da177e4SLinus Torvalds { 30321da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 303375c119afSEric Dumazet struct sk_buff *next_skb = skb_rb_next(skb); 303413dde04fSWei Yongjun int next_skb_size; 30351da177e4SLinus Torvalds 3036058dc334SIlpo Järvinen next_skb_size = next_skb->len; 30371da177e4SLinus Torvalds 3038058dc334SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); 30391da177e4SLinus Torvalds 3040f8071cdeSEric Dumazet if (next_skb_size) { 3041f8071cdeSEric Dumazet if (next_skb_size <= skb_availroom(skb)) 3042f8071cdeSEric Dumazet skb_copy_bits(next_skb, 0, skb_put(skb, next_skb_size), 3043f8071cdeSEric Dumazet next_skb_size); 30443b4929f6SEric Dumazet else if (!tcp_skb_shift(skb, next_skb, 1, next_skb_size)) 3045f8071cdeSEric Dumazet return false; 3046f8071cdeSEric Dumazet } 30472b7cda9cSEric Dumazet tcp_highest_sack_replace(sk, next_skb, skb); 3048a6963a6bSIlpo Järvinen 30491da177e4SLinus Torvalds /* Update sequence range on original skb. */ 30501da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 30511da177e4SLinus Torvalds 3052e6c7d085SIlpo Järvinen /* Merge over control information. This moves PSH/FIN etc. over */ 30534de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; 30541da177e4SLinus Torvalds 30551da177e4SLinus Torvalds /* All done, get rid of second SKB and account for it so 30561da177e4SLinus Torvalds * packet counting does not break. 30571da177e4SLinus Torvalds */ 30584828e7f4SIlpo Järvinen TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; 3059a643b5d4SMartin KaFai Lau TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor; 3060b7689205SIlpo Järvinen 3061b7689205SIlpo Järvinen /* changed transmit queue under us so clear hints */ 3062ef9da47cSIlpo Järvinen tcp_clear_retrans_hints_partial(tp); 3063ef9da47cSIlpo Järvinen if (next_skb == tp->retransmit_skb_hint) 3064ef9da47cSIlpo Järvinen tp->retransmit_skb_hint = skb; 3065b7689205SIlpo Järvinen 3066797108d1SIlpo Järvinen tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); 3067797108d1SIlpo Järvinen 3068082ac2d5SMartin KaFai Lau tcp_skb_collapse_tstamp(skb, next_skb); 3069082ac2d5SMartin KaFai Lau 307075c119afSEric Dumazet tcp_rtx_queue_unlink_and_free(next_skb, sk); 3071f8071cdeSEric Dumazet return true; 30721da177e4SLinus Torvalds } 30731da177e4SLinus Torvalds 307467edfef7SAndi Kleen /* Check if coalescing SKBs is legal. */ 3075a2a385d6SEric Dumazet static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) 30764a17fc3aSIlpo Järvinen { 30774a17fc3aSIlpo Järvinen if (tcp_skb_pcount(skb) > 1) 3078a2a385d6SEric Dumazet return false; 30794a17fc3aSIlpo Järvinen if (skb_cloned(skb)) 3080a2a385d6SEric Dumazet return false; 30812331ccc5SEric Dumazet /* Some heuristics for collapsing over SACK'd could be invented */ 30824a17fc3aSIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 3083a2a385d6SEric Dumazet return false; 30844a17fc3aSIlpo Järvinen 3085a2a385d6SEric Dumazet return true; 30864a17fc3aSIlpo Järvinen } 30874a17fc3aSIlpo Järvinen 308867edfef7SAndi Kleen /* Collapse packets in the retransmit queue to make to create 308967edfef7SAndi Kleen * less packets on the wire. This is only done on retransmission. 309067edfef7SAndi Kleen */ 30914a17fc3aSIlpo Järvinen static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, 30924a17fc3aSIlpo Järvinen int space) 30934a17fc3aSIlpo Järvinen { 30944a17fc3aSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 30954a17fc3aSIlpo Järvinen struct sk_buff *skb = to, *tmp; 3096a2a385d6SEric Dumazet bool first = true; 30974a17fc3aSIlpo Järvinen 3098e0a1e5b5SEric Dumazet if (!sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse) 30994a17fc3aSIlpo Järvinen return; 31004de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 31014a17fc3aSIlpo Järvinen return; 31024a17fc3aSIlpo Järvinen 310375c119afSEric Dumazet skb_rbtree_walk_from_safe(skb, tmp) { 31044a17fc3aSIlpo Järvinen if (!tcp_can_collapse(sk, skb)) 31054a17fc3aSIlpo Järvinen break; 31064a17fc3aSIlpo Järvinen 310785712484SMat Martineau if (!tcp_skb_can_collapse(to, skb)) 3108a643b5d4SMartin KaFai Lau break; 3109a643b5d4SMartin KaFai Lau 31104a17fc3aSIlpo Järvinen space -= skb->len; 31114a17fc3aSIlpo Järvinen 31124a17fc3aSIlpo Järvinen if (first) { 3113a2a385d6SEric Dumazet first = false; 31144a17fc3aSIlpo Järvinen continue; 31154a17fc3aSIlpo Järvinen } 31164a17fc3aSIlpo Järvinen 31174a17fc3aSIlpo Järvinen if (space < 0) 31184a17fc3aSIlpo Järvinen break; 31194a17fc3aSIlpo Järvinen 31204a17fc3aSIlpo Järvinen if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) 31214a17fc3aSIlpo Järvinen break; 31224a17fc3aSIlpo Järvinen 3123f8071cdeSEric Dumazet if (!tcp_collapse_retrans(sk, to)) 3124f8071cdeSEric Dumazet break; 31254a17fc3aSIlpo Järvinen } 31264a17fc3aSIlpo Järvinen } 31274a17fc3aSIlpo Järvinen 31281da177e4SLinus Torvalds /* This retransmits one SKB. Policy decisions and retransmit queue 31291da177e4SLinus Torvalds * state updates are done by the caller. Returns non-zero if an 31301da177e4SLinus Torvalds * error occurred which prevented the send. 31311da177e4SLinus Torvalds */ 313210d3be56SEric Dumazet int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) 31331da177e4SLinus Torvalds { 31345d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 313510d3be56SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 31367d227cd2SSridhar Samudrala unsigned int cur_mss; 313710d3be56SEric Dumazet int diff, len, err; 31381da177e4SLinus Torvalds 313910d3be56SEric Dumazet 314010d3be56SEric Dumazet /* Inconclusive MTU probe */ 314110d3be56SEric Dumazet if (icsk->icsk_mtup.probe_size) 31425d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 31435d424d5aSJohn Heffner 31441da177e4SLinus Torvalds /* Do not sent more than we queued. 1/4 is reserved for possible 3145caa20d9aSStephen Hemminger * copying overhead: fragmentation, tunneling, mangling etc. 31461da177e4SLinus Torvalds */ 314714afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) > 3148ffb4d6c8SEric Dumazet min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), 3149ffb4d6c8SEric Dumazet sk->sk_sndbuf)) 31501da177e4SLinus Torvalds return -EAGAIN; 31511da177e4SLinus Torvalds 31521f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 31531f3279aeSEric Dumazet return -EBUSY; 31541f3279aeSEric Dumazet 31551da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 31567f582b24SEric Dumazet if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) { 31577f582b24SEric Dumazet WARN_ON_ONCE(1); 31587f582b24SEric Dumazet return -EINVAL; 31597f582b24SEric Dumazet } 31601da177e4SLinus Torvalds if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 31611da177e4SLinus Torvalds return -ENOMEM; 31621da177e4SLinus Torvalds } 31631da177e4SLinus Torvalds 31647d227cd2SSridhar Samudrala if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 31657d227cd2SSridhar Samudrala return -EHOSTUNREACH; /* Routing failure or similar. */ 31667d227cd2SSridhar Samudrala 31670c54b85fSIlpo Järvinen cur_mss = tcp_current_mss(sk); 31687d227cd2SSridhar Samudrala 31691da177e4SLinus Torvalds /* If receiver has shrunk his window, and skb is out of 31701da177e4SLinus Torvalds * new window, do not retransmit it. The exception is the 31711da177e4SLinus Torvalds * case, when window is shrunk to zero. In this case 31721da177e4SLinus Torvalds * our retransmit serves as a zero window probe. 31731da177e4SLinus Torvalds */ 31749d4fb27dSJoe Perches if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && 31759d4fb27dSJoe Perches TCP_SKB_CB(skb)->seq != tp->snd_una) 31761da177e4SLinus Torvalds return -EAGAIN; 31771da177e4SLinus Torvalds 317810d3be56SEric Dumazet len = cur_mss * segs; 317910d3be56SEric Dumazet if (skb->len > len) { 318075c119afSEric Dumazet if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len, 318175c119afSEric Dumazet cur_mss, GFP_ATOMIC)) 31821da177e4SLinus Torvalds return -ENOMEM; /* We'll try again later. */ 318302276f3cSIlpo Järvinen } else { 3184c52e2421SEric Dumazet if (skb_unclone(skb, GFP_ATOMIC)) 3185c52e2421SEric Dumazet return -ENOMEM; 318610d3be56SEric Dumazet 318710d3be56SEric Dumazet diff = tcp_skb_pcount(skb); 318810d3be56SEric Dumazet tcp_set_skb_tso_segs(skb, cur_mss); 318910d3be56SEric Dumazet diff -= tcp_skb_pcount(skb); 319010d3be56SEric Dumazet if (diff) 319110d3be56SEric Dumazet tcp_adjust_pcount(sk, skb, diff); 319210d3be56SEric Dumazet if (skb->len < cur_mss) 319310d3be56SEric Dumazet tcp_retrans_try_collapse(sk, skb, cur_mss); 31941da177e4SLinus Torvalds } 31951da177e4SLinus Torvalds 319649213555SDaniel Borkmann /* RFC3168, section 6.1.1.1. ECN fallback */ 319749213555SDaniel Borkmann if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN) 319849213555SDaniel Borkmann tcp_ecn_clear_syn(sk, skb); 319949213555SDaniel Borkmann 3200678550c6SYuchung Cheng /* Update global and local TCP statistics. */ 3201678550c6SYuchung Cheng segs = tcp_skb_pcount(skb); 3202678550c6SYuchung Cheng TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs); 3203678550c6SYuchung Cheng if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 3204678550c6SYuchung Cheng __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 3205678550c6SYuchung Cheng tp->total_retrans += segs; 3206fb31c9b9SWei Wang tp->bytes_retrans += skb->len; 3207678550c6SYuchung Cheng 320850bceae9SThomas Graf /* make sure skb->data is aligned on arches that require it 320950bceae9SThomas Graf * and check if ack-trimming & collapsing extended the headroom 321050bceae9SThomas Graf * beyond what csum_start can cover. 321150bceae9SThomas Graf */ 321250bceae9SThomas Graf if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) || 321350bceae9SThomas Graf skb_headroom(skb) >= 0xFFFF)) { 321410a81980SEric Dumazet struct sk_buff *nskb; 321510a81980SEric Dumazet 3216e2080072SEric Dumazet tcp_skb_tsorted_save(skb) { 321710a81980SEric Dumazet nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC); 321807f8e4d0SFlorian Westphal if (nskb) { 321907f8e4d0SFlorian Westphal nskb->dev = NULL; 322007f8e4d0SFlorian Westphal err = tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC); 322107f8e4d0SFlorian Westphal } else { 322207f8e4d0SFlorian Westphal err = -ENOBUFS; 322307f8e4d0SFlorian Westphal } 3224e2080072SEric Dumazet } tcp_skb_tsorted_restore(skb); 3225e2080072SEric Dumazet 32265889e2c0SYousuk Seung if (!err) { 3227a7a25630SEric Dumazet tcp_update_skb_after_send(sk, skb, tp->tcp_wstamp_ns); 32285889e2c0SYousuk Seung tcp_rate_skb_sent(sk, skb); 32295889e2c0SYousuk Seung } 3230117632e6SEric Dumazet } else { 3231c84a5711SYuchung Cheng err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 3232117632e6SEric Dumazet } 3233c84a5711SYuchung Cheng 32347f12422cSYuchung Cheng /* To avoid taking spuriously low RTT samples based on a timestamp 32357f12422cSYuchung Cheng * for a transmit that never happened, always mark EVER_RETRANS 32367f12422cSYuchung Cheng */ 32377f12422cSYuchung Cheng TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS; 32387f12422cSYuchung Cheng 3239a31ad29eSLawrence Brakmo if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RETRANS_CB_FLAG)) 3240a31ad29eSLawrence Brakmo tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RETRANS_CB, 3241a31ad29eSLawrence Brakmo TCP_SKB_CB(skb)->seq, segs, err); 3242a31ad29eSLawrence Brakmo 3243fc9f3501SEric Dumazet if (likely(!err)) { 3244e086101bSCong Wang trace_tcp_retransmit_skb(sk, skb); 3245678550c6SYuchung Cheng } else if (err != -EBUSY) { 3246ec641b39SYuchung Cheng NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL, segs); 3247fc9f3501SEric Dumazet } 3248c84a5711SYuchung Cheng return err; 324993b174adSYuchung Cheng } 325093b174adSYuchung Cheng 325110d3be56SEric Dumazet int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) 325293b174adSYuchung Cheng { 325393b174adSYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 325410d3be56SEric Dumazet int err = __tcp_retransmit_skb(sk, skb, segs); 32551da177e4SLinus Torvalds 32561da177e4SLinus Torvalds if (err == 0) { 32571da177e4SLinus Torvalds #if FASTRETRANS_DEBUG > 0 32581da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 3259e87cc472SJoe Perches net_dbg_ratelimited("retrans_out leaked\n"); 32601da177e4SLinus Torvalds } 32611da177e4SLinus Torvalds #endif 32621da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 32631da177e4SLinus Torvalds tp->retrans_out += tcp_skb_pcount(skb); 32647ae18975SYuchung Cheng } 32651da177e4SLinus Torvalds 32667ae18975SYuchung Cheng /* Save stamp of the first (attempted) retransmit. */ 32671da177e4SLinus Torvalds if (!tp->retrans_stamp) 32687faee5c0SEric Dumazet tp->retrans_stamp = tcp_skb_timestamp(skb); 32691da177e4SLinus Torvalds 32706e08d5e3SYuchung Cheng if (tp->undo_retrans < 0) 32716e08d5e3SYuchung Cheng tp->undo_retrans = 0; 32726e08d5e3SYuchung Cheng tp->undo_retrans += tcp_skb_pcount(skb); 32731da177e4SLinus Torvalds return err; 32741da177e4SLinus Torvalds } 32751da177e4SLinus Torvalds 32761da177e4SLinus Torvalds /* This gets called after a retransmit timeout, and the initially 32771da177e4SLinus Torvalds * retransmitted data is acknowledged. It tries to continue 32781da177e4SLinus Torvalds * resending the rest of the retransmit queue, until either 32791da177e4SLinus Torvalds * we've sent it all or the congestion window limit is reached. 32801da177e4SLinus Torvalds */ 32811da177e4SLinus Torvalds void tcp_xmit_retransmit_queue(struct sock *sk) 32821da177e4SLinus Torvalds { 32836687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 3284b9f1f1ceSEric Dumazet struct sk_buff *skb, *rtx_head, *hole = NULL; 32851da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 3286916e6d1aSEric Dumazet bool rearm_timer = false; 3287840a3cbeSYuchung Cheng u32 max_segs; 328861eb55f4SIlpo Järvinen int mib_idx; 32896a438bbeSStephen Hemminger 329045e77d31SIlpo Järvinen if (!tp->packets_out) 329145e77d31SIlpo Järvinen return; 329245e77d31SIlpo Järvinen 329375c119afSEric Dumazet rtx_head = tcp_rtx_queue_head(sk); 3294b9f1f1ceSEric Dumazet skb = tp->retransmit_skb_hint ?: rtx_head; 3295ed6e7268SNeal Cardwell max_segs = tcp_tso_segs(sk, tcp_current_mss(sk)); 329675c119afSEric Dumazet skb_rbtree_walk_from(skb) { 3297dca0aaf8SEric Dumazet __u8 sacked; 329810d3be56SEric Dumazet int segs; 32991da177e4SLinus Torvalds 3300218af599SEric Dumazet if (tcp_pacing_check(sk)) 3301218af599SEric Dumazet break; 3302218af599SEric Dumazet 33036a438bbeSStephen Hemminger /* we could do better than to assign each time */ 330451456b29SIan Morris if (!hole) 33056a438bbeSStephen Hemminger tp->retransmit_skb_hint = skb; 33066a438bbeSStephen Hemminger 330710d3be56SEric Dumazet segs = tp->snd_cwnd - tcp_packets_in_flight(tp); 330810d3be56SEric Dumazet if (segs <= 0) 3309916e6d1aSEric Dumazet break; 3310dca0aaf8SEric Dumazet sacked = TCP_SKB_CB(skb)->sacked; 3311a3d2e9f8SEric Dumazet /* In case tcp_shift_skb_data() have aggregated large skbs, 3312a3d2e9f8SEric Dumazet * we need to make sure not sending too bigs TSO packets 3313a3d2e9f8SEric Dumazet */ 3314a3d2e9f8SEric Dumazet segs = min_t(int, segs, max_segs); 33150e1c54c2SIlpo Järvinen 3316840a3cbeSYuchung Cheng if (tp->retrans_out >= tp->lost_out) { 3317006f582cSIlpo Järvinen break; 33180e1c54c2SIlpo Järvinen } else if (!(sacked & TCPCB_LOST)) { 331951456b29SIan Morris if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) 33200e1c54c2SIlpo Järvinen hole = skb; 332161eb55f4SIlpo Järvinen continue; 33221da177e4SLinus Torvalds 33230e1c54c2SIlpo Järvinen } else { 33240e1c54c2SIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Loss) 33250e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFASTRETRANS; 33260e1c54c2SIlpo Järvinen else 33270e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; 33280e1c54c2SIlpo Järvinen } 33290e1c54c2SIlpo Järvinen 33300e1c54c2SIlpo Järvinen if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) 333161eb55f4SIlpo Järvinen continue; 333240b215e5SPavel Emelyanov 3333f9616c35SEric Dumazet if (tcp_small_queue_check(sk, skb, 1)) 3334916e6d1aSEric Dumazet break; 3335f9616c35SEric Dumazet 333610d3be56SEric Dumazet if (tcp_retransmit_skb(sk, skb, segs)) 3337916e6d1aSEric Dumazet break; 333824ab6becSYuchung Cheng 3339de1d6578SYuchung Cheng NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb)); 33401da177e4SLinus Torvalds 3341684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 3342a262f0cdSNandita Dukkipati tp->prr_out += tcp_skb_pcount(skb); 3343a262f0cdSNandita Dukkipati 334475c119afSEric Dumazet if (skb == rtx_head && 334557dde7f7SYuchung Cheng icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT) 3346916e6d1aSEric Dumazet rearm_timer = true; 3347916e6d1aSEric Dumazet 3348916e6d1aSEric Dumazet } 3349916e6d1aSEric Dumazet if (rearm_timer) 33503f80e08fSEric Dumazet tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 33513f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, 33528dc242adSEric Dumazet TCP_RTO_MAX); 33531da177e4SLinus Torvalds } 33541da177e4SLinus Torvalds 3355d83769a5SEric Dumazet /* We allow to exceed memory limits for FIN packets to expedite 3356d83769a5SEric Dumazet * connection tear down and (memory) recovery. 3357845704a5SEric Dumazet * Otherwise tcp_send_fin() could be tempted to either delay FIN 3358845704a5SEric Dumazet * or even be forced to close flow without any FIN. 3359a6c5ea4cSEric Dumazet * In general, we want to allow one skb per socket to avoid hangs 3360a6c5ea4cSEric Dumazet * with edge trigger epoll() 3361d83769a5SEric Dumazet */ 3362a6c5ea4cSEric Dumazet void sk_forced_mem_schedule(struct sock *sk, int size) 3363d83769a5SEric Dumazet { 3364e805605cSJohannes Weiner int amt; 3365d83769a5SEric Dumazet 3366d83769a5SEric Dumazet if (size <= sk->sk_forward_alloc) 3367d83769a5SEric Dumazet return; 3368d83769a5SEric Dumazet amt = sk_mem_pages(size); 3369d83769a5SEric Dumazet sk->sk_forward_alloc += amt * SK_MEM_QUANTUM; 3370e805605cSJohannes Weiner sk_memory_allocated_add(sk, amt); 3371e805605cSJohannes Weiner 3372baac50bbSJohannes Weiner if (mem_cgroup_sockets_enabled && sk->sk_memcg) 3373baac50bbSJohannes Weiner mem_cgroup_charge_skmem(sk->sk_memcg, amt); 3374d83769a5SEric Dumazet } 3375d83769a5SEric Dumazet 3376845704a5SEric Dumazet /* Send a FIN. The caller locks the socket for us. 3377845704a5SEric Dumazet * We should try to send a FIN packet really hard, but eventually give up. 33781da177e4SLinus Torvalds */ 33791da177e4SLinus Torvalds void tcp_send_fin(struct sock *sk) 33801da177e4SLinus Torvalds { 3381ee2aabd3SEric Dumazet struct sk_buff *skb, *tskb, *tail = tcp_write_queue_tail(sk); 33821da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 33831da177e4SLinus Torvalds 3384845704a5SEric Dumazet /* Optimization, tack on the FIN if we have one skb in write queue and 3385845704a5SEric Dumazet * this skb was not yet sent, or we are under memory pressure. 3386845704a5SEric Dumazet * Note: in the latter case, FIN packet will be sent after a timeout, 3387845704a5SEric Dumazet * as TCP stack thinks it has already been transmitted. 33881da177e4SLinus Torvalds */ 3389ee2aabd3SEric Dumazet tskb = tail; 339075c119afSEric Dumazet if (!tskb && tcp_under_memory_pressure(sk)) 339175c119afSEric Dumazet tskb = skb_rb_last(&sk->tcp_rtx_queue); 339275c119afSEric Dumazet 339375c119afSEric Dumazet if (tskb) { 3394845704a5SEric Dumazet TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN; 3395845704a5SEric Dumazet TCP_SKB_CB(tskb)->end_seq++; 33961da177e4SLinus Torvalds tp->write_seq++; 3397ee2aabd3SEric Dumazet if (!tail) { 3398845704a5SEric Dumazet /* This means tskb was already sent. 3399845704a5SEric Dumazet * Pretend we included the FIN on previous transmit. 3400845704a5SEric Dumazet * We need to set tp->snd_nxt to the value it would have 3401845704a5SEric Dumazet * if FIN had been sent. This is because retransmit path 3402845704a5SEric Dumazet * does not change tp->snd_nxt. 3403845704a5SEric Dumazet */ 3404e0d694d6SEric Dumazet WRITE_ONCE(tp->snd_nxt, tp->snd_nxt + 1); 3405845704a5SEric Dumazet return; 3406845704a5SEric Dumazet } 34071da177e4SLinus Torvalds } else { 3408845704a5SEric Dumazet skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation); 3409d1edc085SColin Ian King if (unlikely(!skb)) 3410845704a5SEric Dumazet return; 3411d1edc085SColin Ian King 3412e2080072SEric Dumazet INIT_LIST_HEAD(&skb->tcp_tsorted_anchor); 3413d83769a5SEric Dumazet skb_reserve(skb, MAX_TCP_HEADER); 3414a6c5ea4cSEric Dumazet sk_forced_mem_schedule(sk, skb->truesize); 34151da177e4SLinus Torvalds /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 3416e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tp->write_seq, 3417a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_FIN); 34181da177e4SLinus Torvalds tcp_queue_skb(sk, skb); 34191da177e4SLinus Torvalds } 3420845704a5SEric Dumazet __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF); 34211da177e4SLinus Torvalds } 34221da177e4SLinus Torvalds 34231da177e4SLinus Torvalds /* We get here when a process closes a file descriptor (either due to 34241da177e4SLinus Torvalds * an explicit close() or as a byproduct of exit()'ing) and there 34251da177e4SLinus Torvalds * was unread data in the receive queue. This behavior is recommended 342665bb723cSGerrit Renker * by RFC 2525, section 2.17. -DaveM 34271da177e4SLinus Torvalds */ 3428dd0fc66fSAl Viro void tcp_send_active_reset(struct sock *sk, gfp_t priority) 34291da177e4SLinus Torvalds { 34301da177e4SLinus Torvalds struct sk_buff *skb; 34311da177e4SLinus Torvalds 34327cc2b043SGao Feng TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); 34337cc2b043SGao Feng 34341da177e4SLinus Torvalds /* NOTE: No TCP options attached and we never retransmit this. */ 34351da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, priority); 34361da177e4SLinus Torvalds if (!skb) { 34374e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 34381da177e4SLinus Torvalds return; 34391da177e4SLinus Torvalds } 34401da177e4SLinus Torvalds 34411da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 34421da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 3443e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), 3444a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_RST); 34459a568de4SEric Dumazet tcp_mstamp_refresh(tcp_sk(sk)); 34461da177e4SLinus Torvalds /* Send it off. */ 3447dfb4b9dcSDavid S. Miller if (tcp_transmit_skb(sk, skb, 0, priority)) 34484e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 3449c24b14c4SSong Liu 3450c24b14c4SSong Liu /* skb of trace_tcp_send_reset() keeps the skb that caused RST, 3451c24b14c4SSong Liu * skb here is different to the troublesome skb, so use NULL 3452c24b14c4SSong Liu */ 3453c24b14c4SSong Liu trace_tcp_send_reset(sk, NULL); 34541da177e4SLinus Torvalds } 34551da177e4SLinus Torvalds 345667edfef7SAndi Kleen /* Send a crossed SYN-ACK during socket establishment. 345767edfef7SAndi Kleen * WARNING: This routine must only be called when we have already sent 34581da177e4SLinus Torvalds * a SYN packet that crossed the incoming SYN that caused this routine 34591da177e4SLinus Torvalds * to get called. If this assumption fails then the initial rcv_wnd 34601da177e4SLinus Torvalds * and rcv_wscale values will not be correct. 34611da177e4SLinus Torvalds */ 34621da177e4SLinus Torvalds int tcp_send_synack(struct sock *sk) 34631da177e4SLinus Torvalds { 34641da177e4SLinus Torvalds struct sk_buff *skb; 34651da177e4SLinus Torvalds 346675c119afSEric Dumazet skb = tcp_rtx_queue_head(sk); 346751456b29SIan Morris if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 346875c119afSEric Dumazet pr_err("%s: wrong queue state\n", __func__); 34691da177e4SLinus Torvalds return -EFAULT; 34701da177e4SLinus Torvalds } 34714de075e0SEric Dumazet if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { 34721da177e4SLinus Torvalds if (skb_cloned(skb)) { 3473e2080072SEric Dumazet struct sk_buff *nskb; 3474e2080072SEric Dumazet 3475e2080072SEric Dumazet tcp_skb_tsorted_save(skb) { 3476e2080072SEric Dumazet nskb = skb_copy(skb, GFP_ATOMIC); 3477e2080072SEric Dumazet } tcp_skb_tsorted_restore(skb); 347851456b29SIan Morris if (!nskb) 34791da177e4SLinus Torvalds return -ENOMEM; 3480e2080072SEric Dumazet INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor); 34812bec445fSEric Dumazet tcp_highest_sack_replace(sk, skb, nskb); 348275c119afSEric Dumazet tcp_rtx_queue_unlink_and_free(skb, sk); 3483f4a775d1SEric Dumazet __skb_header_release(nskb); 348475c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb); 3485ab4e846aSEric Dumazet sk_wmem_queued_add(sk, nskb->truesize); 34863ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 34871da177e4SLinus Torvalds skb = nskb; 34881da177e4SLinus Torvalds } 34891da177e4SLinus Torvalds 34904de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; 3491735d3831SFlorian Westphal tcp_ecn_send_synack(sk, skb); 34921da177e4SLinus Torvalds } 3493dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 34941da177e4SLinus Torvalds } 34951da177e4SLinus Torvalds 34964aea39c1SEric Dumazet /** 3497331fca43SMartin KaFai Lau * tcp_make_synack - Allocate one skb and build a SYNACK packet. 3498331fca43SMartin KaFai Lau * @sk: listener socket 3499331fca43SMartin KaFai Lau * @dst: dst entry attached to the SYNACK. It is consumed and caller 3500331fca43SMartin KaFai Lau * should not use it again. 3501331fca43SMartin KaFai Lau * @req: request_sock pointer 3502331fca43SMartin KaFai Lau * @foc: cookie for tcp fast open 3503331fca43SMartin KaFai Lau * @synack_type: Type of synack to prepare 3504331fca43SMartin KaFai Lau * @syn_skb: SYN packet just received. It could be NULL for rtx case. 35054aea39c1SEric Dumazet */ 35065d062de7SEric Dumazet struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst, 3507e6b4d113SWilliam Allen Simpson struct request_sock *req, 3508ca6fb065SEric Dumazet struct tcp_fastopen_cookie *foc, 3509331fca43SMartin KaFai Lau enum tcp_synack_type synack_type, 3510331fca43SMartin KaFai Lau struct sk_buff *syn_skb) 35111da177e4SLinus Torvalds { 35122e6599cbSArnaldo Carvalho de Melo struct inet_request_sock *ireq = inet_rsk(req); 35135d062de7SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 351480f03e27SEric Dumazet struct tcp_md5sig_key *md5 = NULL; 35155d062de7SEric Dumazet struct tcp_out_options opts; 35165d062de7SEric Dumazet struct sk_buff *skb; 3517bd0388aeSWilliam Allen Simpson int tcp_header_size; 35185d062de7SEric Dumazet struct tcphdr *th; 3519f5fff5dcSTom Quetchenbach int mss; 3520a842fe14SEric Dumazet u64 now; 35211da177e4SLinus Torvalds 3522ca6fb065SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 35234aea39c1SEric Dumazet if (unlikely(!skb)) { 35244aea39c1SEric Dumazet dst_release(dst); 35251da177e4SLinus Torvalds return NULL; 35264aea39c1SEric Dumazet } 35271da177e4SLinus Torvalds /* Reserve space for headers. */ 35281da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 35291da177e4SLinus Torvalds 3530b3d05147SEric Dumazet switch (synack_type) { 3531b3d05147SEric Dumazet case TCP_SYNACK_NORMAL: 35329e17f8a4SEric Dumazet skb_set_owner_w(skb, req_to_sk(req)); 3533b3d05147SEric Dumazet break; 3534b3d05147SEric Dumazet case TCP_SYNACK_COOKIE: 3535b3d05147SEric Dumazet /* Under synflood, we do not attach skb to a socket, 3536b3d05147SEric Dumazet * to avoid false sharing. 3537b3d05147SEric Dumazet */ 3538b3d05147SEric Dumazet break; 3539b3d05147SEric Dumazet case TCP_SYNACK_FASTOPEN: 3540ca6fb065SEric Dumazet /* sk is a const pointer, because we want to express multiple 3541ca6fb065SEric Dumazet * cpu might call us concurrently. 3542ca6fb065SEric Dumazet * sk->sk_wmem_alloc in an atomic, we can promote to rw. 3543ca6fb065SEric Dumazet */ 3544ca6fb065SEric Dumazet skb_set_owner_w(skb, (struct sock *)sk); 3545b3d05147SEric Dumazet break; 3546ca6fb065SEric Dumazet } 35474aea39c1SEric Dumazet skb_dst_set(skb, dst); 35481da177e4SLinus Torvalds 35493541f9e8SEric Dumazet mss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 3550f5fff5dcSTom Quetchenbach 355133ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 3552a842fe14SEric Dumazet now = tcp_clock_ns(); 35538b5f12d0SFlorian Westphal #ifdef CONFIG_SYN_COOKIES 3554f8ace8d9SFlorian Westphal if (unlikely(synack_type == TCP_SYNACK_COOKIE && ireq->tstamp_ok)) 3555200ecef6SEric Dumazet skb->skb_mstamp_ns = cookie_init_timestamp(req, now); 35568b5f12d0SFlorian Westphal else 35578b5f12d0SFlorian Westphal #endif 35589e450c1eSYuchung Cheng { 3559a842fe14SEric Dumazet skb->skb_mstamp_ns = now; 35609e450c1eSYuchung Cheng if (!tcp_rsk(req)->snt_synack) /* Timestamp first SYNACK */ 35619e450c1eSYuchung Cheng tcp_rsk(req)->snt_synack = tcp_skb_timestamp_us(skb); 35629e450c1eSYuchung Cheng } 356380f03e27SEric Dumazet 356480f03e27SEric Dumazet #ifdef CONFIG_TCP_MD5SIG 356580f03e27SEric Dumazet rcu_read_lock(); 3566fd3a154aSEric Dumazet md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req)); 356780f03e27SEric Dumazet #endif 356858d607d3SEric Dumazet skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4); 3569331fca43SMartin KaFai Lau /* bpf program will be interested in the tcp_flags */ 3570331fca43SMartin KaFai Lau TCP_SKB_CB(skb)->tcp_flags = TCPHDR_SYN | TCPHDR_ACK; 357160e2a778SUrsula Braun tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5, 3572331fca43SMartin KaFai Lau foc, synack_type, 3573331fca43SMartin KaFai Lau syn_skb) + sizeof(*th); 357433ad798cSAdam Langley 3575aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 3576aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 35771da177e4SLinus Torvalds 3578ea1627c2SEric Dumazet th = (struct tcphdr *)skb->data; 35791da177e4SLinus Torvalds memset(th, 0, sizeof(struct tcphdr)); 35801da177e4SLinus Torvalds th->syn = 1; 35811da177e4SLinus Torvalds th->ack = 1; 35826ac705b1SEric Dumazet tcp_ecn_make_synack(req, th); 3583b44084c2SEric Dumazet th->source = htons(ireq->ir_num); 3584634fb979SEric Dumazet th->dest = ireq->ir_rmt_port; 3585e05a90ecSJamal Hadi Salim skb->mark = ireq->ir_mark; 35863b117750SEric Dumazet skb->ip_summed = CHECKSUM_PARTIAL; 35873b117750SEric Dumazet th->seq = htonl(tcp_rsk(req)->snt_isn); 35888336886fSJerry Chu /* XXX data is queued and acked as is. No buffer/window check */ 35898336886fSJerry Chu th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt); 35901da177e4SLinus Torvalds 35911da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 3592ed53d0abSEric Dumazet th->window = htons(min(req->rsk_rcv_wnd, 65535U)); 35935d062de7SEric Dumazet tcp_options_write((__be32 *)(th + 1), NULL, &opts); 35941da177e4SLinus Torvalds th->doff = (tcp_header_size >> 2); 359590bbcc60SEric Dumazet __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS); 3596cfb6eeb4SYOSHIFUJI Hideaki 3597cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 3598cfb6eeb4SYOSHIFUJI Hideaki /* Okay, we have all we need - do the md5 hash if needed */ 359980f03e27SEric Dumazet if (md5) 3600bd0388aeSWilliam Allen Simpson tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, 360139f8e58eSEric Dumazet md5, req_to_sk(req), skb); 360280f03e27SEric Dumazet rcu_read_unlock(); 3603cfb6eeb4SYOSHIFUJI Hideaki #endif 3604cfb6eeb4SYOSHIFUJI Hideaki 3605331fca43SMartin KaFai Lau bpf_skops_write_hdr_opt((struct sock *)sk, skb, req, syn_skb, 3606331fca43SMartin KaFai Lau synack_type, &opts); 3607331fca43SMartin KaFai Lau 3608a842fe14SEric Dumazet skb->skb_mstamp_ns = now; 3609a842fe14SEric Dumazet tcp_add_tx_delay(skb, tp); 3610a842fe14SEric Dumazet 36111da177e4SLinus Torvalds return skb; 36121da177e4SLinus Torvalds } 36134bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_make_synack); 36141da177e4SLinus Torvalds 361581164413SDaniel Borkmann static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst) 361681164413SDaniel Borkmann { 361781164413SDaniel Borkmann struct inet_connection_sock *icsk = inet_csk(sk); 361881164413SDaniel Borkmann const struct tcp_congestion_ops *ca; 361981164413SDaniel Borkmann u32 ca_key = dst_metric(dst, RTAX_CC_ALGO); 362081164413SDaniel Borkmann 362181164413SDaniel Borkmann if (ca_key == TCP_CA_UNSPEC) 362281164413SDaniel Borkmann return; 362381164413SDaniel Borkmann 362481164413SDaniel Borkmann rcu_read_lock(); 362581164413SDaniel Borkmann ca = tcp_ca_find_key(ca_key); 36260baf26b0SMartin KaFai Lau if (likely(ca && bpf_try_module_get(ca, ca->owner))) { 36270baf26b0SMartin KaFai Lau bpf_module_put(icsk->icsk_ca_ops, icsk->icsk_ca_ops->owner); 362881164413SDaniel Borkmann icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst); 362981164413SDaniel Borkmann icsk->icsk_ca_ops = ca; 363081164413SDaniel Borkmann } 363181164413SDaniel Borkmann rcu_read_unlock(); 363281164413SDaniel Borkmann } 363381164413SDaniel Borkmann 363467edfef7SAndi Kleen /* Do all connect socket setups that can be done AF independent. */ 3635f7e56a76Sstephen hemminger static void tcp_connect_init(struct sock *sk) 36361da177e4SLinus Torvalds { 3637cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 36381da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 36391da177e4SLinus Torvalds __u8 rcv_wscale; 364013d3b1ebSLawrence Brakmo u32 rcv_wnd; 36411da177e4SLinus Torvalds 36421da177e4SLinus Torvalds /* We'll fix this up when we get a response from the other end. 36431da177e4SLinus Torvalds * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 36441da177e4SLinus Torvalds */ 36455d2ed052SEric Dumazet tp->tcp_header_len = sizeof(struct tcphdr); 36465d2ed052SEric Dumazet if (sock_net(sk)->ipv4.sysctl_tcp_timestamps) 36475d2ed052SEric Dumazet tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED; 36481da177e4SLinus Torvalds 3649cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 365000db4124SIan Morris if (tp->af_specific->md5_lookup(sk, sk)) 3651cfb6eeb4SYOSHIFUJI Hideaki tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 3652cfb6eeb4SYOSHIFUJI Hideaki #endif 3653cfb6eeb4SYOSHIFUJI Hideaki 36541da177e4SLinus Torvalds /* If user gave his TCP_MAXSEG, record it to clamp */ 36551da177e4SLinus Torvalds if (tp->rx_opt.user_mss) 36561da177e4SLinus Torvalds tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 36571da177e4SLinus Torvalds tp->max_window = 0; 36585d424d5aSJohn Heffner tcp_mtup_init(sk); 36591da177e4SLinus Torvalds tcp_sync_mss(sk, dst_mtu(dst)); 36601da177e4SLinus Torvalds 366181164413SDaniel Borkmann tcp_ca_dst_init(sk, dst); 366281164413SDaniel Borkmann 36631da177e4SLinus Torvalds if (!tp->window_clamp) 36641da177e4SLinus Torvalds tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 36653541f9e8SEric Dumazet tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 3666f5fff5dcSTom Quetchenbach 36671da177e4SLinus Torvalds tcp_initialize_rcv_mss(sk); 36681da177e4SLinus Torvalds 3669e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 3670e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 3671e88c64f0SHagen Paul Pfeifer (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) 3672e88c64f0SHagen Paul Pfeifer tp->window_clamp = tcp_full_space(sk); 3673e88c64f0SHagen Paul Pfeifer 367413d3b1ebSLawrence Brakmo rcv_wnd = tcp_rwnd_init_bpf(sk); 367513d3b1ebSLawrence Brakmo if (rcv_wnd == 0) 367613d3b1ebSLawrence Brakmo rcv_wnd = dst_metric(dst, RTAX_INITRWND); 367713d3b1ebSLawrence Brakmo 3678ceef9ab6SEric Dumazet tcp_select_initial_window(sk, tcp_full_space(sk), 36791da177e4SLinus Torvalds tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 36801da177e4SLinus Torvalds &tp->rcv_wnd, 36811da177e4SLinus Torvalds &tp->window_clamp, 36829bb37ef0SEric Dumazet sock_net(sk)->ipv4.sysctl_tcp_window_scaling, 368331d12926Slaurent chavey &rcv_wscale, 368413d3b1ebSLawrence Brakmo rcv_wnd); 36851da177e4SLinus Torvalds 36861da177e4SLinus Torvalds tp->rx_opt.rcv_wscale = rcv_wscale; 36871da177e4SLinus Torvalds tp->rcv_ssthresh = tp->rcv_wnd; 36881da177e4SLinus Torvalds 36891da177e4SLinus Torvalds sk->sk_err = 0; 36901da177e4SLinus Torvalds sock_reset_flag(sk, SOCK_DONE); 36911da177e4SLinus Torvalds tp->snd_wnd = 0; 3692ee7537b6SHantzis Fotis tcp_init_wl(tp, 0); 36937f582b24SEric Dumazet tcp_write_queue_purge(sk); 36941da177e4SLinus Torvalds tp->snd_una = tp->write_seq; 36951da177e4SLinus Torvalds tp->snd_sml = tp->write_seq; 369633f5f57eSIlpo Järvinen tp->snd_up = tp->write_seq; 3697e0d694d6SEric Dumazet WRITE_ONCE(tp->snd_nxt, tp->write_seq); 3698ee995283SPavel Emelyanov 3699ee995283SPavel Emelyanov if (likely(!tp->repair)) 37001da177e4SLinus Torvalds tp->rcv_nxt = 0; 3701c7781a6eSAndrew Vagin else 370270eabf0eSEric Dumazet tp->rcv_tstamp = tcp_jiffies32; 3703ee995283SPavel Emelyanov tp->rcv_wup = tp->rcv_nxt; 37047db48e98SEric Dumazet WRITE_ONCE(tp->copied_seq, tp->rcv_nxt); 37051da177e4SLinus Torvalds 37068550f328SLawrence Brakmo inet_csk(sk)->icsk_rto = tcp_timeout_init(sk); 3707463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_retransmits = 0; 37081da177e4SLinus Torvalds tcp_clear_retrans(tp); 37091da177e4SLinus Torvalds } 37101da177e4SLinus Torvalds 3711783237e8SYuchung Cheng static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) 3712783237e8SYuchung Cheng { 3713783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3714783237e8SYuchung Cheng struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 3715783237e8SYuchung Cheng 3716783237e8SYuchung Cheng tcb->end_seq += skb->len; 3717f4a775d1SEric Dumazet __skb_header_release(skb); 3718ab4e846aSEric Dumazet sk_wmem_queued_add(sk, skb->truesize); 3719783237e8SYuchung Cheng sk_mem_charge(sk, skb->truesize); 37200f317464SEric Dumazet WRITE_ONCE(tp->write_seq, tcb->end_seq); 3721783237e8SYuchung Cheng tp->packets_out += tcp_skb_pcount(skb); 3722783237e8SYuchung Cheng } 3723783237e8SYuchung Cheng 3724783237e8SYuchung Cheng /* Build and send a SYN with data and (cached) Fast Open cookie. However, 3725783237e8SYuchung Cheng * queue a data-only packet after the regular SYN, such that regular SYNs 3726783237e8SYuchung Cheng * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges 3727783237e8SYuchung Cheng * only the SYN sequence, the data are retransmitted in the first ACK. 3728783237e8SYuchung Cheng * If cookie is not cached or other error occurs, falls back to send a 3729783237e8SYuchung Cheng * regular SYN with Fast Open cookie request option. 3730783237e8SYuchung Cheng */ 3731783237e8SYuchung Cheng static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) 3732783237e8SYuchung Cheng { 3733783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3734783237e8SYuchung Cheng struct tcp_fastopen_request *fo = tp->fastopen_req; 3735065263f4SWei Wang int space, err = 0; 3736355a901eSEric Dumazet struct sk_buff *syn_data; 3737783237e8SYuchung Cheng 373867da22d2SYuchung Cheng tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ 3739065263f4SWei Wang if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie)) 3740783237e8SYuchung Cheng goto fallback; 3741783237e8SYuchung Cheng 3742783237e8SYuchung Cheng /* MSS for SYN-data is based on cached MSS and bounded by PMTU and 3743783237e8SYuchung Cheng * user-MSS. Reserve maximum option space for middleboxes that add 3744783237e8SYuchung Cheng * private TCP options. The cost is reduced data space in SYN :( 3745783237e8SYuchung Cheng */ 37463541f9e8SEric Dumazet tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp); 37473541f9e8SEric Dumazet 37481b63edd6SYuchung Cheng space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) - 3749783237e8SYuchung Cheng MAX_TCP_OPTION_SPACE; 3750783237e8SYuchung Cheng 3751f5ddcbbbSEric Dumazet space = min_t(size_t, space, fo->size); 3752f5ddcbbbSEric Dumazet 3753f5ddcbbbSEric Dumazet /* limit to order-0 allocations */ 3754f5ddcbbbSEric Dumazet space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER)); 3755f5ddcbbbSEric Dumazet 3756eb934478SEric Dumazet syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false); 3757355a901eSEric Dumazet if (!syn_data) 3758783237e8SYuchung Cheng goto fallback; 3759355a901eSEric Dumazet syn_data->ip_summed = CHECKSUM_PARTIAL; 3760355a901eSEric Dumazet memcpy(syn_data->cb, syn->cb, sizeof(syn->cb)); 376107e100f9SEric Dumazet if (space) { 376207e100f9SEric Dumazet int copied = copy_from_iter(skb_put(syn_data, space), space, 376357be5bdaSAl Viro &fo->data->msg_iter); 376457be5bdaSAl Viro if (unlikely(!copied)) { 3765ba233b34SEric Dumazet tcp_skb_tsorted_anchor_cleanup(syn_data); 3766355a901eSEric Dumazet kfree_skb(syn_data); 3767783237e8SYuchung Cheng goto fallback; 3768783237e8SYuchung Cheng } 376957be5bdaSAl Viro if (copied != space) { 377057be5bdaSAl Viro skb_trim(syn_data, copied); 377157be5bdaSAl Viro space = copied; 377257be5bdaSAl Viro } 3773f859a448SWillem de Bruijn skb_zcopy_set(syn_data, fo->uarg, NULL); 377407e100f9SEric Dumazet } 3775355a901eSEric Dumazet /* No more data pending in inet_wait_for_connect() */ 3776355a901eSEric Dumazet if (space == fo->size) 3777355a901eSEric Dumazet fo->data = NULL; 3778355a901eSEric Dumazet fo->copied = space; 3779783237e8SYuchung Cheng 3780355a901eSEric Dumazet tcp_connect_queue_skb(sk, syn_data); 37810f87230dSFrancis Yan if (syn_data->len) 37820f87230dSFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_BUSY); 3783355a901eSEric Dumazet 3784355a901eSEric Dumazet err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation); 3785355a901eSEric Dumazet 3786d3edd06eSEric Dumazet syn->skb_mstamp_ns = syn_data->skb_mstamp_ns; 3787355a901eSEric Dumazet 3788355a901eSEric Dumazet /* Now full SYN+DATA was cloned and sent (or not), 3789355a901eSEric Dumazet * remove the SYN from the original skb (syn_data) 3790355a901eSEric Dumazet * we keep in write queue in case of a retransmit, as we 3791355a901eSEric Dumazet * also have the SYN packet (with no data) in the same queue. 3792431a9124SEric Dumazet */ 3793355a901eSEric Dumazet TCP_SKB_CB(syn_data)->seq++; 3794355a901eSEric Dumazet TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH; 3795355a901eSEric Dumazet if (!err) { 379667da22d2SYuchung Cheng tp->syn_data = (fo->copied > 0); 379775c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data); 3798f19c29e3SYuchung Cheng NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT); 3799783237e8SYuchung Cheng goto done; 3800783237e8SYuchung Cheng } 3801783237e8SYuchung Cheng 380275c119afSEric Dumazet /* data was not sent, put it in write_queue */ 380375c119afSEric Dumazet __skb_queue_tail(&sk->sk_write_queue, syn_data); 3804b5b7db8dSEric Dumazet tp->packets_out -= tcp_skb_pcount(syn_data); 3805b5b7db8dSEric Dumazet 3806783237e8SYuchung Cheng fallback: 3807783237e8SYuchung Cheng /* Send a regular SYN with Fast Open cookie request option */ 3808783237e8SYuchung Cheng if (fo->cookie.len > 0) 3809783237e8SYuchung Cheng fo->cookie.len = 0; 3810783237e8SYuchung Cheng err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); 3811783237e8SYuchung Cheng if (err) 3812783237e8SYuchung Cheng tp->syn_fastopen = 0; 3813783237e8SYuchung Cheng done: 3814783237e8SYuchung Cheng fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ 3815783237e8SYuchung Cheng return err; 3816783237e8SYuchung Cheng } 3817783237e8SYuchung Cheng 381867edfef7SAndi Kleen /* Build a SYN and send it off. */ 38191da177e4SLinus Torvalds int tcp_connect(struct sock *sk) 38201da177e4SLinus Torvalds { 38211da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 38221da177e4SLinus Torvalds struct sk_buff *buff; 3823ee586811SEric Paris int err; 38241da177e4SLinus Torvalds 3825de525be2SLawrence Brakmo tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB, 0, NULL); 38268ba60924SEric Dumazet 38278ba60924SEric Dumazet if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 38288ba60924SEric Dumazet return -EHOSTUNREACH; /* Routing failure or similar. */ 38298ba60924SEric Dumazet 38301da177e4SLinus Torvalds tcp_connect_init(sk); 38311da177e4SLinus Torvalds 38322b916477SAndrey Vagin if (unlikely(tp->repair)) { 38332b916477SAndrey Vagin tcp_finish_connect(sk, NULL); 38342b916477SAndrey Vagin return 0; 38352b916477SAndrey Vagin } 38362b916477SAndrey Vagin 3837eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true); 3838355a901eSEric Dumazet if (unlikely(!buff)) 38391da177e4SLinus Torvalds return -ENOBUFS; 38401da177e4SLinus Torvalds 3841a3433f35SChangli Gao tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); 38429a568de4SEric Dumazet tcp_mstamp_refresh(tp); 38439a568de4SEric Dumazet tp->retrans_stamp = tcp_time_stamp(tp); 3844783237e8SYuchung Cheng tcp_connect_queue_skb(sk, buff); 3845735d3831SFlorian Westphal tcp_ecn_send_syn(sk, buff); 384675c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, buff); 38471da177e4SLinus Torvalds 3848783237e8SYuchung Cheng /* Send off SYN; include data in Fast Open. */ 3849783237e8SYuchung Cheng err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : 3850783237e8SYuchung Cheng tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); 3851ee586811SEric Paris if (err == -ECONNREFUSED) 3852ee586811SEric Paris return err; 3853bd37a088SWei Yongjun 3854bd37a088SWei Yongjun /* We change tp->snd_nxt after the tcp_transmit_skb() call 3855bd37a088SWei Yongjun * in order to make this packet get counted in tcpOutSegs. 3856bd37a088SWei Yongjun */ 3857e0d694d6SEric Dumazet WRITE_ONCE(tp->snd_nxt, tp->write_seq); 3858bd37a088SWei Yongjun tp->pushed_seq = tp->write_seq; 3859b5b7db8dSEric Dumazet buff = tcp_send_head(sk); 3860b5b7db8dSEric Dumazet if (unlikely(buff)) { 3861e0d694d6SEric Dumazet WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(buff)->seq); 3862b5b7db8dSEric Dumazet tp->pushed_seq = TCP_SKB_CB(buff)->seq; 3863b5b7db8dSEric Dumazet } 386481cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); 38651da177e4SLinus Torvalds 38661da177e4SLinus Torvalds /* Timer for repeating the SYN until an answer. */ 38673f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 38683f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 38691da177e4SLinus Torvalds return 0; 38701da177e4SLinus Torvalds } 38714bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_connect); 38721da177e4SLinus Torvalds 38731da177e4SLinus Torvalds /* Send out a delayed ack, the caller does the policy checking 38741da177e4SLinus Torvalds * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 38751da177e4SLinus Torvalds * for details. 38761da177e4SLinus Torvalds */ 38771da177e4SLinus Torvalds void tcp_send_delayed_ack(struct sock *sk) 38781da177e4SLinus Torvalds { 3879463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 3880463c84b9SArnaldo Carvalho de Melo int ato = icsk->icsk_ack.ato; 38811da177e4SLinus Torvalds unsigned long timeout; 38821da177e4SLinus Torvalds 38831da177e4SLinus Torvalds if (ato > TCP_DELACK_MIN) { 3884463c84b9SArnaldo Carvalho de Melo const struct tcp_sock *tp = tcp_sk(sk); 38851da177e4SLinus Torvalds int max_ato = HZ / 2; 38861da177e4SLinus Torvalds 388731954cd8SWei Wang if (inet_csk_in_pingpong_mode(sk) || 3888056834d9SIlpo Järvinen (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 38891da177e4SLinus Torvalds max_ato = TCP_DELACK_MAX; 38901da177e4SLinus Torvalds 38911da177e4SLinus Torvalds /* Slow path, intersegment interval is "high". */ 38921da177e4SLinus Torvalds 38931da177e4SLinus Torvalds /* If some rtt estimate is known, use it to bound delayed ack. 3894463c84b9SArnaldo Carvalho de Melo * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 38951da177e4SLinus Torvalds * directly. 38961da177e4SLinus Torvalds */ 3897740b0f18SEric Dumazet if (tp->srtt_us) { 3898740b0f18SEric Dumazet int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3), 3899740b0f18SEric Dumazet TCP_DELACK_MIN); 39001da177e4SLinus Torvalds 39011da177e4SLinus Torvalds if (rtt < max_ato) 39021da177e4SLinus Torvalds max_ato = rtt; 39031da177e4SLinus Torvalds } 39041da177e4SLinus Torvalds 39051da177e4SLinus Torvalds ato = min(ato, max_ato); 39061da177e4SLinus Torvalds } 39071da177e4SLinus Torvalds 39082b8ee4f0SMartin KaFai Lau ato = min_t(u32, ato, inet_csk(sk)->icsk_delack_max); 39092b8ee4f0SMartin KaFai Lau 39101da177e4SLinus Torvalds /* Stay within the limit we were given */ 39111da177e4SLinus Torvalds timeout = jiffies + ato; 39121da177e4SLinus Torvalds 39131da177e4SLinus Torvalds /* Use new timeout only if there wasn't a older one earlier. */ 3914463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 39151da177e4SLinus Torvalds /* If delack timer was blocked or is about to expire, 39161da177e4SLinus Torvalds * send ACK now. 39171da177e4SLinus Torvalds */ 3918463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.blocked || 3919463c84b9SArnaldo Carvalho de Melo time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 39201da177e4SLinus Torvalds tcp_send_ack(sk); 39211da177e4SLinus Torvalds return; 39221da177e4SLinus Torvalds } 39231da177e4SLinus Torvalds 3924463c84b9SArnaldo Carvalho de Melo if (!time_before(timeout, icsk->icsk_ack.timeout)) 3925463c84b9SArnaldo Carvalho de Melo timeout = icsk->icsk_ack.timeout; 39261da177e4SLinus Torvalds } 3927463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 3928463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.timeout = timeout; 3929463c84b9SArnaldo Carvalho de Melo sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 39301da177e4SLinus Torvalds } 39311da177e4SLinus Torvalds 39321da177e4SLinus Torvalds /* This routine sends an ack and also updates the window. */ 39332987babbSYuchung Cheng void __tcp_send_ack(struct sock *sk, u32 rcv_nxt) 39341da177e4SLinus Torvalds { 39351da177e4SLinus Torvalds struct sk_buff *buff; 39361da177e4SLinus Torvalds 3937058dc334SIlpo Järvinen /* If we have been reset, we may not send again. */ 3938058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3939058dc334SIlpo Järvinen return; 3940058dc334SIlpo Järvinen 39411da177e4SLinus Torvalds /* We are not putting this on the write queue, so 39421da177e4SLinus Torvalds * tcp_transmit_skb() will set the ownership to this 39431da177e4SLinus Torvalds * sock. 39441da177e4SLinus Torvalds */ 39457450aaf6SEric Dumazet buff = alloc_skb(MAX_TCP_HEADER, 39467450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 39477450aaf6SEric Dumazet if (unlikely(!buff)) { 3948463c84b9SArnaldo Carvalho de Melo inet_csk_schedule_ack(sk); 3949463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; 39503f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 39513f421baaSArnaldo Carvalho de Melo TCP_DELACK_MAX, TCP_RTO_MAX); 39521da177e4SLinus Torvalds return; 39531da177e4SLinus Torvalds } 39541da177e4SLinus Torvalds 39551da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 39561da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 3957a3433f35SChangli Gao tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); 39581da177e4SLinus Torvalds 395998781965SEric Dumazet /* We do not want pure acks influencing TCP Small Queues or fq/pacing 396098781965SEric Dumazet * too much. 396198781965SEric Dumazet * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784 396298781965SEric Dumazet */ 396398781965SEric Dumazet skb_set_tcp_pure_ack(buff); 396498781965SEric Dumazet 39651da177e4SLinus Torvalds /* Send it off, this clears delayed acks for us. */ 39662987babbSYuchung Cheng __tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0, rcv_nxt); 39671da177e4SLinus Torvalds } 396827cde44aSYuchung Cheng EXPORT_SYMBOL_GPL(__tcp_send_ack); 39692987babbSYuchung Cheng 39702987babbSYuchung Cheng void tcp_send_ack(struct sock *sk) 39712987babbSYuchung Cheng { 39722987babbSYuchung Cheng __tcp_send_ack(sk, tcp_sk(sk)->rcv_nxt); 39731da177e4SLinus Torvalds } 39741da177e4SLinus Torvalds 39751da177e4SLinus Torvalds /* This routine sends a packet with an out of date sequence 39761da177e4SLinus Torvalds * number. It assumes the other end will try to ack it. 39771da177e4SLinus Torvalds * 39781da177e4SLinus Torvalds * Question: what should we make while urgent mode? 39791da177e4SLinus Torvalds * 4.4BSD forces sending single byte of data. We cannot send 39801da177e4SLinus Torvalds * out of window data, because we have SND.NXT==SND.MAX... 39811da177e4SLinus Torvalds * 39821da177e4SLinus Torvalds * Current solution: to send TWO zero-length segments in urgent mode: 39831da177e4SLinus Torvalds * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 39841da177e4SLinus Torvalds * out-of-date with SND.UNA-1 to probe window. 39851da177e4SLinus Torvalds */ 3986e520af48SEric Dumazet static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib) 39871da177e4SLinus Torvalds { 39881da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 39891da177e4SLinus Torvalds struct sk_buff *skb; 39901da177e4SLinus Torvalds 39911da177e4SLinus Torvalds /* We don't queue it, tcp_transmit_skb() sets ownership. */ 39927450aaf6SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, 39937450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 399451456b29SIan Morris if (!skb) 39951da177e4SLinus Torvalds return -1; 39961da177e4SLinus Torvalds 39971da177e4SLinus Torvalds /* Reserve space for headers and set control bits. */ 39981da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 39991da177e4SLinus Torvalds /* Use a previous sequence. This should cause the other 40001da177e4SLinus Torvalds * end to send an ack. Don't queue or clone SKB, just 40011da177e4SLinus Torvalds * send it. 40021da177e4SLinus Torvalds */ 4003a3433f35SChangli Gao tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); 4004e2e8009fSRenato Westphal NET_INC_STATS(sock_net(sk), mib); 40057450aaf6SEric Dumazet return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0); 40061da177e4SLinus Torvalds } 40071da177e4SLinus Torvalds 4008385e2070SEric Dumazet /* Called from setsockopt( ... TCP_REPAIR ) */ 4009ee995283SPavel Emelyanov void tcp_send_window_probe(struct sock *sk) 4010ee995283SPavel Emelyanov { 4011ee995283SPavel Emelyanov if (sk->sk_state == TCP_ESTABLISHED) { 4012ee995283SPavel Emelyanov tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; 40139a568de4SEric Dumazet tcp_mstamp_refresh(tcp_sk(sk)); 4014e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE); 4015ee995283SPavel Emelyanov } 4016ee995283SPavel Emelyanov } 4017ee995283SPavel Emelyanov 401867edfef7SAndi Kleen /* Initiate keepalive or window probe from timer. */ 4019e520af48SEric Dumazet int tcp_write_wakeup(struct sock *sk, int mib) 40201da177e4SLinus Torvalds { 40211da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 40221da177e4SLinus Torvalds struct sk_buff *skb; 40231da177e4SLinus Torvalds 4024058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 4025058dc334SIlpo Järvinen return -1; 4026058dc334SIlpo Järvinen 402700db4124SIan Morris skb = tcp_send_head(sk); 402800db4124SIan Morris if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { 40291da177e4SLinus Torvalds int err; 40300c54b85fSIlpo Järvinen unsigned int mss = tcp_current_mss(sk); 403190840defSIlpo Järvinen unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 40321da177e4SLinus Torvalds 40331da177e4SLinus Torvalds if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 40341da177e4SLinus Torvalds tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 40351da177e4SLinus Torvalds 40361da177e4SLinus Torvalds /* We are probing the opening of a window 40371da177e4SLinus Torvalds * but the window size is != 0 40381da177e4SLinus Torvalds * must have been a result SWS avoidance ( sender ) 40391da177e4SLinus Torvalds */ 40401da177e4SLinus Torvalds if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 40411da177e4SLinus Torvalds skb->len > mss) { 40421da177e4SLinus Torvalds seg_size = min(seg_size, mss); 40434de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 404475c119afSEric Dumazet if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE, 404575c119afSEric Dumazet skb, seg_size, mss, GFP_ATOMIC)) 40461da177e4SLinus Torvalds return -1; 40471da177e4SLinus Torvalds } else if (!tcp_skb_pcount(skb)) 40485bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss); 40491da177e4SLinus Torvalds 40504de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 4051dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 405266f5fe62SIlpo Järvinen if (!err) 405366f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 40541da177e4SLinus Torvalds return err; 40551da177e4SLinus Torvalds } else { 405633f5f57eSIlpo Järvinen if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) 4057e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 1, mib); 4058e520af48SEric Dumazet return tcp_xmit_probe_skb(sk, 0, mib); 40591da177e4SLinus Torvalds } 40601da177e4SLinus Torvalds } 40611da177e4SLinus Torvalds 40621da177e4SLinus Torvalds /* A window probe timeout has occurred. If window is not closed send 40631da177e4SLinus Torvalds * a partial packet else a zero probe. 40641da177e4SLinus Torvalds */ 40651da177e4SLinus Torvalds void tcp_send_probe0(struct sock *sk) 40661da177e4SLinus Torvalds { 4067463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 40681da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 4069c6214a97SNikolay Borisov struct net *net = sock_net(sk); 4070c1d5674fSYuchung Cheng unsigned long timeout; 40711da177e4SLinus Torvalds int err; 40721da177e4SLinus Torvalds 4073e520af48SEric Dumazet err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE); 40741da177e4SLinus Torvalds 407575c119afSEric Dumazet if (tp->packets_out || tcp_write_queue_empty(sk)) { 40761da177e4SLinus Torvalds /* Cancel probe timer, if it is not required. */ 40776687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 0; 4078463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff = 0; 40791da177e4SLinus Torvalds return; 40801da177e4SLinus Torvalds } 40811da177e4SLinus Torvalds 4082c1d5674fSYuchung Cheng icsk->icsk_probes_out++; 40831da177e4SLinus Torvalds if (err <= 0) { 4084c6214a97SNikolay Borisov if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2) 4085463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff++; 4086c1d5674fSYuchung Cheng timeout = tcp_probe0_when(sk, TCP_RTO_MAX); 40871da177e4SLinus Torvalds } else { 40881da177e4SLinus Torvalds /* If packet was not sent due to local congestion, 4089c1d5674fSYuchung Cheng * Let senders fight for local resources conservatively. 40901da177e4SLinus Torvalds */ 4091c1d5674fSYuchung Cheng timeout = TCP_RESOURCE_PROBE_INTERVAL; 40921da177e4SLinus Torvalds } 40938dc242adSEric Dumazet tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, timeout, TCP_RTO_MAX); 40941da177e4SLinus Torvalds } 40955db92c99SOctavian Purdila 4096ea3bea3aSEric Dumazet int tcp_rtx_synack(const struct sock *sk, struct request_sock *req) 40975db92c99SOctavian Purdila { 40985db92c99SOctavian Purdila const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific; 40995db92c99SOctavian Purdila struct flowi fl; 41005db92c99SOctavian Purdila int res; 41015db92c99SOctavian Purdila 410258d607d3SEric Dumazet tcp_rsk(req)->txhash = net_tx_rndhash(); 4103331fca43SMartin KaFai Lau res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL, 4104331fca43SMartin KaFai Lau NULL); 41055db92c99SOctavian Purdila if (!res) { 410690bbcc60SEric Dumazet __TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); 410702a1d6e7SEric Dumazet __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 41087e32b443SYuchung Cheng if (unlikely(tcp_passive_fastopen(sk))) 41097e32b443SYuchung Cheng tcp_sk(sk)->total_retrans++; 4110cf34ce3dSSong Liu trace_tcp_retransmit_synack(sk, req); 41115db92c99SOctavian Purdila } 41125db92c99SOctavian Purdila return res; 41135db92c99SOctavian Purdila } 41145db92c99SOctavian Purdila EXPORT_SYMBOL(tcp_rtx_synack); 4115