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 */ 441bd0388aeSWilliam Allen Simpson __u8 *hash_location; /* temporary pointer, overloaded */ 4422100c8d2SYuchung Cheng __u32 tsval, tsecr; /* need to include OPTION_TS */ 4432100c8d2SYuchung Cheng struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */ 444eda7acddSPeter Krystad struct mptcp_out_options mptcp; 44533ad798cSAdam Langley }; 44633ad798cSAdam Langley 447eda7acddSPeter Krystad static void mptcp_options_write(__be32 *ptr, struct tcp_out_options *opts) 448eda7acddSPeter Krystad { 449eda7acddSPeter Krystad #if IS_ENABLED(CONFIG_MPTCP) 450eda7acddSPeter Krystad if (unlikely(OPTION_MPTCP & opts->options)) 451eda7acddSPeter Krystad mptcp_write_options(ptr, &opts->mptcp); 452eda7acddSPeter Krystad #endif 453eda7acddSPeter Krystad } 454eda7acddSPeter Krystad 45567edfef7SAndi Kleen /* Write previously computed TCP options to the packet. 45667edfef7SAndi Kleen * 45767edfef7SAndi Kleen * Beware: Something in the Internet is very sensitive to the ordering of 458fd6149d3SIlpo Järvinen * TCP options, we learned this through the hard way, so be careful here. 459fd6149d3SIlpo Järvinen * Luckily we can at least blame others for their non-compliance but from 4608e3bff96Sstephen hemminger * inter-operability perspective it seems that we're somewhat stuck with 461fd6149d3SIlpo Järvinen * the ordering which we have been using if we want to keep working with 462fd6149d3SIlpo Järvinen * those broken things (not that it currently hurts anybody as there isn't 463fd6149d3SIlpo Järvinen * particular reason why the ordering would need to be changed). 464fd6149d3SIlpo Järvinen * 465fd6149d3SIlpo Järvinen * At least SACK_PERM as the first option is known to lead to a disaster 466fd6149d3SIlpo Järvinen * (but it may well be that other scenarios fail similarly). 467fd6149d3SIlpo Järvinen */ 46833ad798cSAdam Langley static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp, 469bd0388aeSWilliam Allen Simpson struct tcp_out_options *opts) 470bd0388aeSWilliam Allen Simpson { 4712100c8d2SYuchung Cheng u16 options = opts->options; /* mungable copy */ 472bd0388aeSWilliam Allen Simpson 473bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_MD5 & options)) { 4741a2c6181SChristoph Paasch *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | 4751a2c6181SChristoph Paasch (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG); 476bd0388aeSWilliam Allen Simpson /* overload cookie hash location */ 477bd0388aeSWilliam Allen Simpson opts->hash_location = (__u8 *)ptr; 47833ad798cSAdam Langley ptr += 4; 47933ad798cSAdam Langley } 48033ad798cSAdam Langley 481fd6149d3SIlpo Järvinen if (unlikely(opts->mss)) { 482fd6149d3SIlpo Järvinen *ptr++ = htonl((TCPOPT_MSS << 24) | 483fd6149d3SIlpo Järvinen (TCPOLEN_MSS << 16) | 484fd6149d3SIlpo Järvinen opts->mss); 485fd6149d3SIlpo Järvinen } 486fd6149d3SIlpo Järvinen 487bd0388aeSWilliam Allen Simpson if (likely(OPTION_TS & options)) { 488bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 48933ad798cSAdam Langley *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | 49033ad798cSAdam Langley (TCPOLEN_SACK_PERM << 16) | 49133ad798cSAdam Langley (TCPOPT_TIMESTAMP << 8) | 49233ad798cSAdam Langley TCPOLEN_TIMESTAMP); 493bd0388aeSWilliam Allen Simpson options &= ~OPTION_SACK_ADVERTISE; 49433ad798cSAdam Langley } else { 495496c98dfSYOSHIFUJI Hideaki *ptr++ = htonl((TCPOPT_NOP << 24) | 49640efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 49740efc6faSStephen Hemminger (TCPOPT_TIMESTAMP << 8) | 49840efc6faSStephen Hemminger TCPOLEN_TIMESTAMP); 49940efc6faSStephen Hemminger } 50033ad798cSAdam Langley *ptr++ = htonl(opts->tsval); 50133ad798cSAdam Langley *ptr++ = htonl(opts->tsecr); 50233ad798cSAdam Langley } 50333ad798cSAdam Langley 504bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 50533ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 50633ad798cSAdam Langley (TCPOPT_NOP << 16) | 50733ad798cSAdam Langley (TCPOPT_SACK_PERM << 8) | 50833ad798cSAdam Langley TCPOLEN_SACK_PERM); 50933ad798cSAdam Langley } 51033ad798cSAdam Langley 511bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_WSCALE & options)) { 51233ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 51333ad798cSAdam Langley (TCPOPT_WINDOW << 16) | 51433ad798cSAdam Langley (TCPOLEN_WINDOW << 8) | 51533ad798cSAdam Langley opts->ws); 51633ad798cSAdam Langley } 51733ad798cSAdam Langley 51833ad798cSAdam Langley if (unlikely(opts->num_sack_blocks)) { 51933ad798cSAdam Langley struct tcp_sack_block *sp = tp->rx_opt.dsack ? 52033ad798cSAdam Langley tp->duplicate_sack : tp->selective_acks; 52140efc6faSStephen Hemminger int this_sack; 52240efc6faSStephen Hemminger 52340efc6faSStephen Hemminger *ptr++ = htonl((TCPOPT_NOP << 24) | 52440efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 52540efc6faSStephen Hemminger (TCPOPT_SACK << 8) | 52633ad798cSAdam Langley (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * 52740efc6faSStephen Hemminger TCPOLEN_SACK_PERBLOCK))); 5282de979bdSStephen Hemminger 52933ad798cSAdam Langley for (this_sack = 0; this_sack < opts->num_sack_blocks; 53033ad798cSAdam Langley ++this_sack) { 53140efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].start_seq); 53240efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].end_seq); 53340efc6faSStephen Hemminger } 5342de979bdSStephen Hemminger 53540efc6faSStephen Hemminger tp->rx_opt.dsack = 0; 53640efc6faSStephen Hemminger } 5372100c8d2SYuchung Cheng 5382100c8d2SYuchung Cheng if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) { 5392100c8d2SYuchung Cheng struct tcp_fastopen_cookie *foc = opts->fastopen_cookie; 5407f9b838bSDaniel Lee u8 *p = (u8 *)ptr; 5417f9b838bSDaniel Lee u32 len; /* Fast Open option length */ 5422100c8d2SYuchung Cheng 5437f9b838bSDaniel Lee if (foc->exp) { 5447f9b838bSDaniel Lee len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len; 5457f9b838bSDaniel Lee *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) | 5462100c8d2SYuchung Cheng TCPOPT_FASTOPEN_MAGIC); 5477f9b838bSDaniel Lee p += TCPOLEN_EXP_FASTOPEN_BASE; 5487f9b838bSDaniel Lee } else { 5497f9b838bSDaniel Lee len = TCPOLEN_FASTOPEN_BASE + foc->len; 5507f9b838bSDaniel Lee *p++ = TCPOPT_FASTOPEN; 5517f9b838bSDaniel Lee *p++ = len; 5522100c8d2SYuchung Cheng } 5537f9b838bSDaniel Lee 5547f9b838bSDaniel Lee memcpy(p, foc->val, foc->len); 5557f9b838bSDaniel Lee if ((len & 3) == 2) { 5567f9b838bSDaniel Lee p[foc->len] = TCPOPT_NOP; 5577f9b838bSDaniel Lee p[foc->len + 1] = TCPOPT_NOP; 5587f9b838bSDaniel Lee } 5597f9b838bSDaniel Lee ptr += (len + 3) >> 2; 5602100c8d2SYuchung Cheng } 56160e2a778SUrsula Braun 56260e2a778SUrsula Braun smc_options_write(ptr, &options); 563eda7acddSPeter Krystad 564eda7acddSPeter Krystad mptcp_options_write(ptr, opts); 56560e2a778SUrsula Braun } 56660e2a778SUrsula Braun 56760e2a778SUrsula Braun static void smc_set_option(const struct tcp_sock *tp, 56860e2a778SUrsula Braun struct tcp_out_options *opts, 56960e2a778SUrsula Braun unsigned int *remaining) 57060e2a778SUrsula Braun { 57160e2a778SUrsula Braun #if IS_ENABLED(CONFIG_SMC) 57260e2a778SUrsula Braun if (static_branch_unlikely(&tcp_have_smc)) { 57360e2a778SUrsula Braun if (tp->syn_smc) { 57460e2a778SUrsula Braun if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) { 57560e2a778SUrsula Braun opts->options |= OPTION_SMC; 57660e2a778SUrsula Braun *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED; 57760e2a778SUrsula Braun } 57860e2a778SUrsula Braun } 57960e2a778SUrsula Braun } 58060e2a778SUrsula Braun #endif 58160e2a778SUrsula Braun } 58260e2a778SUrsula Braun 58360e2a778SUrsula Braun static void smc_set_option_cond(const struct tcp_sock *tp, 58460e2a778SUrsula Braun const struct inet_request_sock *ireq, 58560e2a778SUrsula Braun struct tcp_out_options *opts, 58660e2a778SUrsula Braun unsigned int *remaining) 58760e2a778SUrsula Braun { 58860e2a778SUrsula Braun #if IS_ENABLED(CONFIG_SMC) 58960e2a778SUrsula Braun if (static_branch_unlikely(&tcp_have_smc)) { 59060e2a778SUrsula Braun if (tp->syn_smc && ireq->smc_ok) { 59160e2a778SUrsula Braun if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) { 59260e2a778SUrsula Braun opts->options |= OPTION_SMC; 59360e2a778SUrsula Braun *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED; 59460e2a778SUrsula Braun } 59560e2a778SUrsula Braun } 59660e2a778SUrsula Braun } 59760e2a778SUrsula Braun #endif 59840efc6faSStephen Hemminger } 59940efc6faSStephen Hemminger 600cec37a6eSPeter Krystad static void mptcp_set_option_cond(const struct request_sock *req, 601cec37a6eSPeter Krystad struct tcp_out_options *opts, 602cec37a6eSPeter Krystad unsigned int *remaining) 603cec37a6eSPeter Krystad { 604cec37a6eSPeter Krystad if (rsk_is_mptcp(req)) { 605cec37a6eSPeter Krystad unsigned int size; 606cec37a6eSPeter Krystad 607cec37a6eSPeter Krystad if (mptcp_synack_options(req, &size, &opts->mptcp)) { 608cec37a6eSPeter Krystad if (*remaining >= size) { 609cec37a6eSPeter Krystad opts->options |= OPTION_MPTCP; 610cec37a6eSPeter Krystad *remaining -= size; 611cec37a6eSPeter Krystad } 612cec37a6eSPeter Krystad } 613cec37a6eSPeter Krystad } 614cec37a6eSPeter Krystad } 615cec37a6eSPeter Krystad 61667edfef7SAndi Kleen /* Compute TCP options for SYN packets. This is not the final 61767edfef7SAndi Kleen * network wire format yet. 61867edfef7SAndi Kleen */ 61995c96174SEric Dumazet static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb, 62033ad798cSAdam Langley struct tcp_out_options *opts, 621cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 622cf533ea5SEric Dumazet { 62333ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 62495c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 625783237e8SYuchung Cheng struct tcp_fastopen_request *fastopen = tp->fastopen_req; 62633ad798cSAdam Langley 6278c2320e8SEric Dumazet *md5 = NULL; 628cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 629921f9a0fSEric Dumazet if (static_branch_unlikely(&tcp_md5_needed) && 6306015c71eSEric Dumazet rcu_access_pointer(tp->md5sig_info)) { 63133ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 63233ad798cSAdam Langley if (*md5) { 63333ad798cSAdam Langley opts->options |= OPTION_MD5; 634bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 635cfb6eeb4SYOSHIFUJI Hideaki } 6368c2320e8SEric Dumazet } 637cfb6eeb4SYOSHIFUJI Hideaki #endif 63833ad798cSAdam Langley 63933ad798cSAdam Langley /* We always get an MSS option. The option bytes which will be seen in 64033ad798cSAdam Langley * normal data packets should timestamps be used, must be in the MSS 64133ad798cSAdam Langley * advertised. But we subtract them from tp->mss_cache so that 64233ad798cSAdam Langley * calculations in tcp_sendmsg are simpler etc. So account for this 64333ad798cSAdam Langley * fact here if necessary. If we don't do this correctly, as a 64433ad798cSAdam Langley * receiver we won't recognize data packets as being full sized when we 64533ad798cSAdam Langley * should, and thus we won't abide by the delayed ACK rules correctly. 64633ad798cSAdam Langley * SACKs don't matter, we never delay an ACK when we have any of those 64733ad798cSAdam Langley * going out. */ 64833ad798cSAdam Langley opts->mss = tcp_advertise_mss(sk); 649bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 65033ad798cSAdam Langley 6515d2ed052SEric Dumazet if (likely(sock_net(sk)->ipv4.sysctl_tcp_timestamps && !*md5)) { 65233ad798cSAdam Langley opts->options |= OPTION_TS; 6537faee5c0SEric Dumazet opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset; 65433ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 655bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 65633ad798cSAdam Langley } 6579bb37ef0SEric Dumazet if (likely(sock_net(sk)->ipv4.sysctl_tcp_window_scaling)) { 65833ad798cSAdam Langley opts->ws = tp->rx_opt.rcv_wscale; 65989e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 660bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 66133ad798cSAdam Langley } 662f9301034SEric Dumazet if (likely(sock_net(sk)->ipv4.sysctl_tcp_sack)) { 66333ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 664b32d1310SDavid S. Miller if (unlikely(!(OPTION_TS & opts->options))) 665bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 66633ad798cSAdam Langley } 66733ad798cSAdam Langley 668783237e8SYuchung Cheng if (fastopen && fastopen->cookie.len >= 0) { 6692646c831SDaniel Lee u32 need = fastopen->cookie.len; 6702646c831SDaniel Lee 6712646c831SDaniel Lee need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE : 6722646c831SDaniel Lee TCPOLEN_FASTOPEN_BASE; 673783237e8SYuchung Cheng need = (need + 3) & ~3U; /* Align to 32 bits */ 674783237e8SYuchung Cheng if (remaining >= need) { 675783237e8SYuchung Cheng opts->options |= OPTION_FAST_OPEN_COOKIE; 676783237e8SYuchung Cheng opts->fastopen_cookie = &fastopen->cookie; 677783237e8SYuchung Cheng remaining -= need; 678783237e8SYuchung Cheng tp->syn_fastopen = 1; 6792646c831SDaniel Lee tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0; 680783237e8SYuchung Cheng } 681783237e8SYuchung Cheng } 682bd0388aeSWilliam Allen Simpson 68360e2a778SUrsula Braun smc_set_option(tp, opts, &remaining); 68460e2a778SUrsula Braun 685cec37a6eSPeter Krystad if (sk_is_mptcp(sk)) { 686cec37a6eSPeter Krystad unsigned int size; 687cec37a6eSPeter Krystad 688cc7972eaSChristoph Paasch if (mptcp_syn_options(sk, skb, &size, &opts->mptcp)) { 689cec37a6eSPeter Krystad opts->options |= OPTION_MPTCP; 690cec37a6eSPeter Krystad remaining -= size; 691cec37a6eSPeter Krystad } 692cec37a6eSPeter Krystad } 693cec37a6eSPeter Krystad 694bd0388aeSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 69533ad798cSAdam Langley } 69633ad798cSAdam Langley 69767edfef7SAndi Kleen /* Set up TCP options for SYN-ACKs. */ 69860e2a778SUrsula Braun static unsigned int tcp_synack_options(const struct sock *sk, 69960e2a778SUrsula Braun struct request_sock *req, 70095c96174SEric Dumazet unsigned int mss, struct sk_buff *skb, 70133ad798cSAdam Langley struct tcp_out_options *opts, 70280f03e27SEric Dumazet const struct tcp_md5sig_key *md5, 7038336886fSJerry Chu struct tcp_fastopen_cookie *foc) 7044957faadSWilliam Allen Simpson { 70533ad798cSAdam Langley struct inet_request_sock *ireq = inet_rsk(req); 70695c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 70733ad798cSAdam Langley 70833ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 70980f03e27SEric Dumazet if (md5) { 71033ad798cSAdam Langley opts->options |= OPTION_MD5; 7114957faadSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 7124957faadSWilliam Allen Simpson 7134957faadSWilliam Allen Simpson /* We can't fit any SACK blocks in a packet with MD5 + TS 7144957faadSWilliam Allen Simpson * options. There was discussion about disabling SACK 7154957faadSWilliam Allen Simpson * rather than TS in order to fit in better with old, 7164957faadSWilliam Allen Simpson * buggy kernels, but that was deemed to be unnecessary. 7174957faadSWilliam Allen Simpson */ 718de213e5eSEric Dumazet ireq->tstamp_ok &= !ireq->sack_ok; 71933ad798cSAdam Langley } 72033ad798cSAdam Langley #endif 72133ad798cSAdam Langley 7224957faadSWilliam Allen Simpson /* We always send an MSS option. */ 72333ad798cSAdam Langley opts->mss = mss; 7244957faadSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 72533ad798cSAdam Langley 72633ad798cSAdam Langley if (likely(ireq->wscale_ok)) { 72733ad798cSAdam Langley opts->ws = ireq->rcv_wscale; 72889e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 7294957faadSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 73033ad798cSAdam Langley } 731de213e5eSEric Dumazet if (likely(ireq->tstamp_ok)) { 73233ad798cSAdam Langley opts->options |= OPTION_TS; 73395a22caeSFlorian Westphal opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off; 73433ad798cSAdam Langley opts->tsecr = req->ts_recent; 7354957faadSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 73633ad798cSAdam Langley } 73733ad798cSAdam Langley if (likely(ireq->sack_ok)) { 73833ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 739de213e5eSEric Dumazet if (unlikely(!ireq->tstamp_ok)) 7404957faadSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 74133ad798cSAdam Langley } 7427f9b838bSDaniel Lee if (foc != NULL && foc->len >= 0) { 7437f9b838bSDaniel Lee u32 need = foc->len; 7447f9b838bSDaniel Lee 7457f9b838bSDaniel Lee need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE : 7467f9b838bSDaniel Lee TCPOLEN_FASTOPEN_BASE; 7478336886fSJerry Chu need = (need + 3) & ~3U; /* Align to 32 bits */ 7488336886fSJerry Chu if (remaining >= need) { 7498336886fSJerry Chu opts->options |= OPTION_FAST_OPEN_COOKIE; 7508336886fSJerry Chu opts->fastopen_cookie = foc; 7518336886fSJerry Chu remaining -= need; 7528336886fSJerry Chu } 7538336886fSJerry Chu } 7544957faadSWilliam Allen Simpson 755cec37a6eSPeter Krystad mptcp_set_option_cond(req, opts, &remaining); 756cec37a6eSPeter Krystad 75760e2a778SUrsula Braun smc_set_option_cond(tcp_sk(sk), ireq, opts, &remaining); 75860e2a778SUrsula Braun 7594957faadSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 76033ad798cSAdam Langley } 76133ad798cSAdam Langley 76267edfef7SAndi Kleen /* Compute TCP options for ESTABLISHED sockets. This is not the 76367edfef7SAndi Kleen * final wire format yet. 76467edfef7SAndi Kleen */ 76595c96174SEric Dumazet static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb, 76633ad798cSAdam Langley struct tcp_out_options *opts, 767cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 768cf533ea5SEric Dumazet { 76933ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 77095c96174SEric Dumazet unsigned int size = 0; 771cabeccbdSIlpo Järvinen unsigned int eff_sacks; 77233ad798cSAdam Langley 7735843ef42SAndi Kleen opts->options = 0; 7745843ef42SAndi Kleen 7758c2320e8SEric Dumazet *md5 = NULL; 77633ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 777921f9a0fSEric Dumazet if (static_branch_unlikely(&tcp_md5_needed) && 7786015c71eSEric Dumazet rcu_access_pointer(tp->md5sig_info)) { 77933ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 7808c2320e8SEric Dumazet if (*md5) { 78133ad798cSAdam Langley opts->options |= OPTION_MD5; 78233ad798cSAdam Langley size += TCPOLEN_MD5SIG_ALIGNED; 78333ad798cSAdam Langley } 7848c2320e8SEric Dumazet } 78533ad798cSAdam Langley #endif 78633ad798cSAdam Langley 78733ad798cSAdam Langley if (likely(tp->rx_opt.tstamp_ok)) { 78833ad798cSAdam Langley opts->options |= OPTION_TS; 7897faee5c0SEric Dumazet opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0; 79033ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 79133ad798cSAdam Langley size += TCPOLEN_TSTAMP_ALIGNED; 79233ad798cSAdam Langley } 79333ad798cSAdam Langley 794cec37a6eSPeter Krystad /* MPTCP options have precedence over SACK for the limited TCP 795cec37a6eSPeter Krystad * option space because a MPTCP connection would be forced to 796cec37a6eSPeter Krystad * fall back to regular TCP if a required multipath option is 797cec37a6eSPeter Krystad * missing. SACK still gets a chance to use whatever space is 798cec37a6eSPeter Krystad * left. 799cec37a6eSPeter Krystad */ 800cec37a6eSPeter Krystad if (sk_is_mptcp(sk)) { 801cec37a6eSPeter Krystad unsigned int remaining = MAX_TCP_OPTION_SPACE - size; 802cec37a6eSPeter Krystad unsigned int opt_size = 0; 803cec37a6eSPeter Krystad 804cec37a6eSPeter Krystad if (mptcp_established_options(sk, skb, &opt_size, remaining, 805cec37a6eSPeter Krystad &opts->mptcp)) { 806cec37a6eSPeter Krystad opts->options |= OPTION_MPTCP; 807cec37a6eSPeter Krystad size += opt_size; 808cec37a6eSPeter Krystad } 809cec37a6eSPeter Krystad } 810cec37a6eSPeter Krystad 811cabeccbdSIlpo Järvinen eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; 812cabeccbdSIlpo Järvinen if (unlikely(eff_sacks)) { 81395c96174SEric Dumazet const unsigned int remaining = MAX_TCP_OPTION_SPACE - size; 8149cfcca23SMat Martineau if (unlikely(remaining < TCPOLEN_SACK_BASE_ALIGNED + 8159cfcca23SMat Martineau TCPOLEN_SACK_PERBLOCK)) 8169cfcca23SMat Martineau return size; 8179cfcca23SMat Martineau 81833ad798cSAdam Langley opts->num_sack_blocks = 81995c96174SEric Dumazet min_t(unsigned int, eff_sacks, 82033ad798cSAdam Langley (remaining - TCPOLEN_SACK_BASE_ALIGNED) / 82133ad798cSAdam Langley TCPOLEN_SACK_PERBLOCK); 8229cfcca23SMat Martineau 82333ad798cSAdam Langley size += TCPOLEN_SACK_BASE_ALIGNED + 82433ad798cSAdam Langley opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; 82533ad798cSAdam Langley } 82633ad798cSAdam Langley 82733ad798cSAdam Langley return size; 82840efc6faSStephen Hemminger } 8291da177e4SLinus Torvalds 83046d3ceabSEric Dumazet 83146d3ceabSEric Dumazet /* TCP SMALL QUEUES (TSQ) 83246d3ceabSEric Dumazet * 83346d3ceabSEric Dumazet * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev) 83446d3ceabSEric Dumazet * to reduce RTT and bufferbloat. 83546d3ceabSEric Dumazet * We do this using a special skb destructor (tcp_wfree). 83646d3ceabSEric Dumazet * 83746d3ceabSEric Dumazet * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb 83846d3ceabSEric Dumazet * needs to be reallocated in a driver. 8398e3bff96Sstephen hemminger * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc 84046d3ceabSEric Dumazet * 84146d3ceabSEric Dumazet * Since transmit from skb destructor is forbidden, we use a tasklet 84246d3ceabSEric Dumazet * to process all sockets that eventually need to send more skbs. 84346d3ceabSEric Dumazet * We use one tasklet per cpu, with its own queue of sockets. 84446d3ceabSEric Dumazet */ 84546d3ceabSEric Dumazet struct tsq_tasklet { 84646d3ceabSEric Dumazet struct tasklet_struct tasklet; 84746d3ceabSEric Dumazet struct list_head head; /* queue of tcp sockets */ 84846d3ceabSEric Dumazet }; 84946d3ceabSEric Dumazet static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet); 85046d3ceabSEric Dumazet 85173a6bab5SEric Dumazet static void tcp_tsq_write(struct sock *sk) 8526f458dfbSEric Dumazet { 8536f458dfbSEric Dumazet if ((1 << sk->sk_state) & 8546f458dfbSEric Dumazet (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING | 855f9616c35SEric Dumazet TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) { 856f9616c35SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 857f9616c35SEric Dumazet 858f9616c35SEric Dumazet if (tp->lost_out > tp->retrans_out && 8593a91d29fSKoichiro Den tp->snd_cwnd > tcp_packets_in_flight(tp)) { 8603a91d29fSKoichiro Den tcp_mstamp_refresh(tp); 861f9616c35SEric Dumazet tcp_xmit_retransmit_queue(sk); 8623a91d29fSKoichiro Den } 863f9616c35SEric Dumazet 864f9616c35SEric Dumazet tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle, 865bf06200eSJohn Ogness 0, GFP_ATOMIC); 8666f458dfbSEric Dumazet } 867f9616c35SEric Dumazet } 86873a6bab5SEric Dumazet 86973a6bab5SEric Dumazet static void tcp_tsq_handler(struct sock *sk) 87073a6bab5SEric Dumazet { 87173a6bab5SEric Dumazet bh_lock_sock(sk); 87273a6bab5SEric Dumazet if (!sock_owned_by_user(sk)) 87373a6bab5SEric Dumazet tcp_tsq_write(sk); 87473a6bab5SEric Dumazet else if (!test_and_set_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags)) 87573a6bab5SEric Dumazet sock_hold(sk); 87673a6bab5SEric Dumazet bh_unlock_sock(sk); 87773a6bab5SEric Dumazet } 87846d3ceabSEric Dumazet /* 8798e3bff96Sstephen hemminger * One tasklet per cpu tries to send more skbs. 88046d3ceabSEric Dumazet * We run in tasklet context but need to disable irqs when 8818e3bff96Sstephen hemminger * transferring tsq->head because tcp_wfree() might 88246d3ceabSEric Dumazet * interrupt us (non NAPI drivers) 88346d3ceabSEric Dumazet */ 88446d3ceabSEric Dumazet static void tcp_tasklet_func(unsigned long data) 88546d3ceabSEric Dumazet { 88646d3ceabSEric Dumazet struct tsq_tasklet *tsq = (struct tsq_tasklet *)data; 88746d3ceabSEric Dumazet LIST_HEAD(list); 88846d3ceabSEric Dumazet unsigned long flags; 88946d3ceabSEric Dumazet struct list_head *q, *n; 89046d3ceabSEric Dumazet struct tcp_sock *tp; 89146d3ceabSEric Dumazet struct sock *sk; 89246d3ceabSEric Dumazet 89346d3ceabSEric Dumazet local_irq_save(flags); 89446d3ceabSEric Dumazet list_splice_init(&tsq->head, &list); 89546d3ceabSEric Dumazet local_irq_restore(flags); 89646d3ceabSEric Dumazet 89746d3ceabSEric Dumazet list_for_each_safe(q, n, &list) { 89846d3ceabSEric Dumazet tp = list_entry(q, struct tcp_sock, tsq_node); 89946d3ceabSEric Dumazet list_del(&tp->tsq_node); 90046d3ceabSEric Dumazet 90146d3ceabSEric Dumazet sk = (struct sock *)tp; 9020a9648f1SEric Dumazet smp_mb__before_atomic(); 9037aa5470cSEric Dumazet clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags); 9047aa5470cSEric Dumazet 9056f458dfbSEric Dumazet tcp_tsq_handler(sk); 90646d3ceabSEric Dumazet sk_free(sk); 90746d3ceabSEric Dumazet } 90846d3ceabSEric Dumazet } 90946d3ceabSEric Dumazet 91040fc3423SEric Dumazet #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \ 91140fc3423SEric Dumazet TCPF_WRITE_TIMER_DEFERRED | \ 91240fc3423SEric Dumazet TCPF_DELACK_TIMER_DEFERRED | \ 91340fc3423SEric Dumazet TCPF_MTU_REDUCED_DEFERRED) 91446d3ceabSEric Dumazet /** 91546d3ceabSEric Dumazet * tcp_release_cb - tcp release_sock() callback 91646d3ceabSEric Dumazet * @sk: socket 91746d3ceabSEric Dumazet * 91846d3ceabSEric Dumazet * called from release_sock() to perform protocol dependent 91946d3ceabSEric Dumazet * actions before socket release. 92046d3ceabSEric Dumazet */ 92146d3ceabSEric Dumazet void tcp_release_cb(struct sock *sk) 92246d3ceabSEric Dumazet { 9236f458dfbSEric Dumazet unsigned long flags, nflags; 92446d3ceabSEric Dumazet 9256f458dfbSEric Dumazet /* perform an atomic operation only if at least one flag is set */ 9266f458dfbSEric Dumazet do { 9277aa5470cSEric Dumazet flags = sk->sk_tsq_flags; 9286f458dfbSEric Dumazet if (!(flags & TCP_DEFERRED_ALL)) 9296f458dfbSEric Dumazet return; 9306f458dfbSEric Dumazet nflags = flags & ~TCP_DEFERRED_ALL; 9317aa5470cSEric Dumazet } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags); 9326f458dfbSEric Dumazet 93373a6bab5SEric Dumazet if (flags & TCPF_TSQ_DEFERRED) { 93473a6bab5SEric Dumazet tcp_tsq_write(sk); 93573a6bab5SEric Dumazet __sock_put(sk); 93673a6bab5SEric Dumazet } 937c3f9b018SEric Dumazet /* Here begins the tricky part : 938c3f9b018SEric Dumazet * We are called from release_sock() with : 939c3f9b018SEric Dumazet * 1) BH disabled 940c3f9b018SEric Dumazet * 2) sk_lock.slock spinlock held 941c3f9b018SEric Dumazet * 3) socket owned by us (sk->sk_lock.owned == 1) 942c3f9b018SEric Dumazet * 943c3f9b018SEric Dumazet * But following code is meant to be called from BH handlers, 944c3f9b018SEric Dumazet * so we should keep BH disabled, but early release socket ownership 945c3f9b018SEric Dumazet */ 946c3f9b018SEric Dumazet sock_release_ownership(sk); 947c3f9b018SEric Dumazet 94840fc3423SEric Dumazet if (flags & TCPF_WRITE_TIMER_DEFERRED) { 9496f458dfbSEric Dumazet tcp_write_timer_handler(sk); 950144d56e9SEric Dumazet __sock_put(sk); 951144d56e9SEric Dumazet } 95240fc3423SEric Dumazet if (flags & TCPF_DELACK_TIMER_DEFERRED) { 9536f458dfbSEric Dumazet tcp_delack_timer_handler(sk); 954144d56e9SEric Dumazet __sock_put(sk); 955144d56e9SEric Dumazet } 95640fc3423SEric Dumazet if (flags & TCPF_MTU_REDUCED_DEFERRED) { 9574fab9071SNeal Cardwell inet_csk(sk)->icsk_af_ops->mtu_reduced(sk); 958144d56e9SEric Dumazet __sock_put(sk); 959144d56e9SEric Dumazet } 96046d3ceabSEric Dumazet } 96146d3ceabSEric Dumazet EXPORT_SYMBOL(tcp_release_cb); 96246d3ceabSEric Dumazet 96346d3ceabSEric Dumazet void __init tcp_tasklet_init(void) 96446d3ceabSEric Dumazet { 96546d3ceabSEric Dumazet int i; 96646d3ceabSEric Dumazet 96746d3ceabSEric Dumazet for_each_possible_cpu(i) { 96846d3ceabSEric Dumazet struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i); 96946d3ceabSEric Dumazet 97046d3ceabSEric Dumazet INIT_LIST_HEAD(&tsq->head); 97146d3ceabSEric Dumazet tasklet_init(&tsq->tasklet, 97246d3ceabSEric Dumazet tcp_tasklet_func, 97346d3ceabSEric Dumazet (unsigned long)tsq); 97446d3ceabSEric Dumazet } 97546d3ceabSEric Dumazet } 97646d3ceabSEric Dumazet 97746d3ceabSEric Dumazet /* 97846d3ceabSEric Dumazet * Write buffer destructor automatically called from kfree_skb. 9798e3bff96Sstephen hemminger * We can't xmit new skbs from this context, as we might already 98046d3ceabSEric Dumazet * hold qdisc lock. 98146d3ceabSEric Dumazet */ 982d6a4a104SEric Dumazet void tcp_wfree(struct sk_buff *skb) 98346d3ceabSEric Dumazet { 98446d3ceabSEric Dumazet struct sock *sk = skb->sk; 98546d3ceabSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 986408f0a6cSEric Dumazet unsigned long flags, nval, oval; 9879b462d02SEric Dumazet 9889b462d02SEric Dumazet /* Keep one reference on sk_wmem_alloc. 9899b462d02SEric Dumazet * Will be released by sk_free() from here or tcp_tasklet_func() 9909b462d02SEric Dumazet */ 99114afee4bSReshetova, Elena WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc)); 9929b462d02SEric Dumazet 9939b462d02SEric Dumazet /* If this softirq is serviced by ksoftirqd, we are likely under stress. 9949b462d02SEric Dumazet * Wait until our queues (qdisc + devices) are drained. 9959b462d02SEric Dumazet * This gives : 9969b462d02SEric Dumazet * - less callbacks to tcp_write_xmit(), reducing stress (batches) 9979b462d02SEric Dumazet * - chance for incoming ACK (processed by another cpu maybe) 9989b462d02SEric Dumazet * to migrate this flow (skb->ooo_okay will be eventually set) 9999b462d02SEric Dumazet */ 100014afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current) 10019b462d02SEric Dumazet goto out; 100246d3ceabSEric Dumazet 10037aa5470cSEric Dumazet for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) { 100446d3ceabSEric Dumazet struct tsq_tasklet *tsq; 1005a9b204d1SEric Dumazet bool empty; 100646d3ceabSEric Dumazet 1007408f0a6cSEric Dumazet if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED)) 1008408f0a6cSEric Dumazet goto out; 1009408f0a6cSEric Dumazet 101073a6bab5SEric Dumazet nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED; 10117aa5470cSEric Dumazet nval = cmpxchg(&sk->sk_tsq_flags, oval, nval); 1012408f0a6cSEric Dumazet if (nval != oval) 1013408f0a6cSEric Dumazet continue; 1014408f0a6cSEric Dumazet 101546d3ceabSEric Dumazet /* queue this socket to tasklet queue */ 101646d3ceabSEric Dumazet local_irq_save(flags); 1017903ceff7SChristoph Lameter tsq = this_cpu_ptr(&tsq_tasklet); 1018a9b204d1SEric Dumazet empty = list_empty(&tsq->head); 101946d3ceabSEric Dumazet list_add(&tp->tsq_node, &tsq->head); 1020a9b204d1SEric Dumazet if (empty) 102146d3ceabSEric Dumazet tasklet_schedule(&tsq->tasklet); 102246d3ceabSEric Dumazet local_irq_restore(flags); 10239b462d02SEric Dumazet return; 102446d3ceabSEric Dumazet } 10259b462d02SEric Dumazet out: 10269b462d02SEric Dumazet sk_free(sk); 102746d3ceabSEric Dumazet } 102846d3ceabSEric Dumazet 102973a6bab5SEric Dumazet /* Note: Called under soft irq. 103073a6bab5SEric Dumazet * We can call TCP stack right away, unless socket is owned by user. 1031218af599SEric Dumazet */ 1032218af599SEric Dumazet enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer) 1033218af599SEric Dumazet { 1034218af599SEric Dumazet struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer); 1035218af599SEric Dumazet struct sock *sk = (struct sock *)tp; 1036218af599SEric Dumazet 103773a6bab5SEric Dumazet tcp_tsq_handler(sk); 103873a6bab5SEric Dumazet sock_put(sk); 1039218af599SEric Dumazet 1040218af599SEric Dumazet return HRTIMER_NORESTART; 1041218af599SEric Dumazet } 1042218af599SEric Dumazet 1043a7a25630SEric Dumazet static void tcp_update_skb_after_send(struct sock *sk, struct sk_buff *skb, 1044a7a25630SEric Dumazet u64 prior_wstamp) 1045e2080072SEric Dumazet { 1046ab408b6dSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 1047ab408b6dSEric Dumazet 1048ab408b6dSEric Dumazet if (sk->sk_pacing_status != SK_PACING_NONE) { 104976a9ebe8SEric Dumazet unsigned long rate = sk->sk_pacing_rate; 1050ab408b6dSEric Dumazet 1051ab408b6dSEric Dumazet /* Original sch_fq does not pace first 10 MSS 1052ab408b6dSEric Dumazet * Note that tp->data_segs_out overflows after 2^32 packets, 1053ab408b6dSEric Dumazet * this is a minor annoyance. 1054ab408b6dSEric Dumazet */ 105576a9ebe8SEric Dumazet if (rate != ~0UL && rate && tp->data_segs_out >= 10) { 1056a7a25630SEric Dumazet u64 len_ns = div64_ul((u64)skb->len * NSEC_PER_SEC, rate); 1057a7a25630SEric Dumazet u64 credit = tp->tcp_wstamp_ns - prior_wstamp; 1058a7a25630SEric Dumazet 1059a7a25630SEric Dumazet /* take into account OS jitter */ 1060a7a25630SEric Dumazet len_ns -= min_t(u64, len_ns / 2, credit); 1061a7a25630SEric Dumazet tp->tcp_wstamp_ns += len_ns; 1062ab408b6dSEric Dumazet } 1063ab408b6dSEric Dumazet } 1064e2080072SEric Dumazet list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue); 1065e2080072SEric Dumazet } 1066e2080072SEric Dumazet 10671da177e4SLinus Torvalds /* This routine actually transmits TCP packets queued in by 10681da177e4SLinus Torvalds * tcp_do_sendmsg(). This is used by both the initial 10691da177e4SLinus Torvalds * transmission and possible later retransmissions. 10701da177e4SLinus Torvalds * All SKB's seen here are completely headerless. It is our 10711da177e4SLinus Torvalds * job to build the TCP header, and pass the packet down to 10721da177e4SLinus Torvalds * IP so it can do the same plus pass the packet off to the 10731da177e4SLinus Torvalds * device. 10741da177e4SLinus Torvalds * 10751da177e4SLinus Torvalds * We are working here with either a clone of the original 10761da177e4SLinus Torvalds * SKB, or a fresh unique copy made by the retransmit engine. 10771da177e4SLinus Torvalds */ 10782987babbSYuchung Cheng static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, 10792987babbSYuchung Cheng int clone_it, gfp_t gfp_mask, u32 rcv_nxt) 10801da177e4SLinus Torvalds { 10816687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 1082dfb4b9dcSDavid S. Miller struct inet_sock *inet; 1083dfb4b9dcSDavid S. Miller struct tcp_sock *tp; 1084dfb4b9dcSDavid S. Miller struct tcp_skb_cb *tcb; 108533ad798cSAdam Langley struct tcp_out_options opts; 108695c96174SEric Dumazet unsigned int tcp_options_size, tcp_header_size; 10878c72c65bSEric Dumazet struct sk_buff *oskb = NULL; 1088cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 10891da177e4SLinus Torvalds struct tcphdr *th; 1090a7a25630SEric Dumazet u64 prior_wstamp; 10911da177e4SLinus Torvalds int err; 10921da177e4SLinus Torvalds 1093dfb4b9dcSDavid S. Miller BUG_ON(!skb || !tcp_skb_pcount(skb)); 10946f094b9eSLawrence Brakmo tp = tcp_sk(sk); 10957f12422cSYuchung Cheng prior_wstamp = tp->tcp_wstamp_ns; 10967f12422cSYuchung Cheng tp->tcp_wstamp_ns = max(tp->tcp_wstamp_ns, tp->tcp_clock_cache); 10977f12422cSYuchung Cheng skb->skb_mstamp_ns = tp->tcp_wstamp_ns; 1098ccdbb6e9SEric Dumazet if (clone_it) { 10996f094b9eSLawrence Brakmo TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq 11006f094b9eSLawrence Brakmo - tp->snd_una; 11018c72c65bSEric Dumazet oskb = skb; 1102e2080072SEric Dumazet 1103e2080072SEric Dumazet tcp_skb_tsorted_save(oskb) { 1104e2080072SEric Dumazet if (unlikely(skb_cloned(oskb))) 1105e2080072SEric Dumazet skb = pskb_copy(oskb, gfp_mask); 1106dfb4b9dcSDavid S. Miller else 1107e2080072SEric Dumazet skb = skb_clone(oskb, gfp_mask); 1108e2080072SEric Dumazet } tcp_skb_tsorted_restore(oskb); 1109e2080072SEric Dumazet 1110dfb4b9dcSDavid S. Miller if (unlikely(!skb)) 1111dfb4b9dcSDavid S. Miller return -ENOBUFS; 1112b738a185SEric Dumazet /* retransmit skbs might have a non zero value in skb->dev 1113b738a185SEric Dumazet * because skb->dev is aliased with skb->rbnode.rb_left 1114b738a185SEric Dumazet */ 1115b738a185SEric Dumazet skb->dev = NULL; 1116dfb4b9dcSDavid S. Miller } 11175f6188a8SEric Dumazet 1118dfb4b9dcSDavid S. Miller inet = inet_sk(sk); 1119dfb4b9dcSDavid S. Miller tcb = TCP_SKB_CB(skb); 112033ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 11211da177e4SLinus Torvalds 1122051ba674SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) { 112333ad798cSAdam Langley tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); 1124051ba674SEric Dumazet } else { 112533ad798cSAdam Langley tcp_options_size = tcp_established_options(sk, skb, &opts, 112633ad798cSAdam Langley &md5); 1127051ba674SEric Dumazet /* Force a PSH flag on all (GSO) packets to expedite GRO flush 1128051ba674SEric Dumazet * at receiver : This slightly improve GRO performance. 1129051ba674SEric Dumazet * Note that we do not force the PSH flag for non GSO packets, 1130051ba674SEric Dumazet * because they might be sent under high congestion events, 1131051ba674SEric Dumazet * and in this case it is better to delay the delivery of 1-MSS 1132051ba674SEric Dumazet * packets and thus the corresponding ACK packet that would 1133051ba674SEric Dumazet * release the following packet. 1134051ba674SEric Dumazet */ 1135051ba674SEric Dumazet if (tcp_skb_pcount(skb) > 1) 1136051ba674SEric Dumazet tcb->tcp_flags |= TCPHDR_PSH; 1137051ba674SEric Dumazet } 113833ad798cSAdam Langley tcp_header_size = tcp_options_size + sizeof(struct tcphdr); 11391da177e4SLinus Torvalds 1140547669d4SEric Dumazet /* if no packet is in qdisc/device queue, then allow XPS to select 1141b2532eb9SEric Dumazet * another queue. We can be called from tcp_tsq_handler() 114273a6bab5SEric Dumazet * which holds one reference to sk. 1143b2532eb9SEric Dumazet * 1144b2532eb9SEric Dumazet * TODO: Ideally, in-flight pure ACK packets should not matter here. 1145b2532eb9SEric Dumazet * One way to get this would be to set skb->truesize = 2 on them. 1146547669d4SEric Dumazet */ 1147b2532eb9SEric Dumazet skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1); 11481da177e4SLinus Torvalds 114938ab52e8SEric Dumazet /* If we had to use memory reserve to allocate this skb, 115038ab52e8SEric Dumazet * this might cause drops if packet is looped back : 115138ab52e8SEric Dumazet * Other socket might not have SOCK_MEMALLOC. 115238ab52e8SEric Dumazet * Packets not looped back do not care about pfmemalloc. 115338ab52e8SEric Dumazet */ 115438ab52e8SEric Dumazet skb->pfmemalloc = 0; 115538ab52e8SEric Dumazet 1156aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 1157aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 115846d3ceabSEric Dumazet 115946d3ceabSEric Dumazet skb_orphan(skb); 116046d3ceabSEric Dumazet skb->sk = sk; 11611d2077acSEric Dumazet skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree; 1162b73c3d0eSTom Herbert skb_set_hash_from_sk(skb, sk); 116314afee4bSReshetova, Elena refcount_add(skb->truesize, &sk->sk_wmem_alloc); 11641da177e4SLinus Torvalds 1165c3a2e837SJulian Anastasov skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm); 1166c3a2e837SJulian Anastasov 11671da177e4SLinus Torvalds /* Build TCP header and checksum it. */ 1168ea1627c2SEric Dumazet th = (struct tcphdr *)skb->data; 1169c720c7e8SEric Dumazet th->source = inet->inet_sport; 1170c720c7e8SEric Dumazet th->dest = inet->inet_dport; 11711da177e4SLinus Torvalds th->seq = htonl(tcb->seq); 11722987babbSYuchung Cheng th->ack_seq = htonl(rcv_nxt); 1173df7a3b07SAl Viro *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | 11744de075e0SEric Dumazet tcb->tcp_flags); 1175dfb4b9dcSDavid S. Miller 11761da177e4SLinus Torvalds th->check = 0; 11771da177e4SLinus Torvalds th->urg_ptr = 0; 11781da177e4SLinus Torvalds 117933f5f57eSIlpo Järvinen /* The urg_mode check is necessary during a below snd_una win probe */ 11807691367dSHerbert Xu if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { 11817691367dSHerbert Xu if (before(tp->snd_up, tcb->seq + 0x10000)) { 11821da177e4SLinus Torvalds th->urg_ptr = htons(tp->snd_up - tcb->seq); 11831da177e4SLinus Torvalds th->urg = 1; 11847691367dSHerbert Xu } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { 11850eae88f3SEric Dumazet th->urg_ptr = htons(0xFFFF); 11867691367dSHerbert Xu th->urg = 1; 11877691367dSHerbert Xu } 11881da177e4SLinus Torvalds } 11891da177e4SLinus Torvalds 1190bd0388aeSWilliam Allen Simpson tcp_options_write((__be32 *)(th + 1), tp, &opts); 119151466a75SEric Dumazet skb_shinfo(skb)->gso_type = sk->sk_gso_type; 1192ea1627c2SEric Dumazet if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) { 1193ea1627c2SEric Dumazet th->window = htons(tcp_select_window(sk)); 1194ea1627c2SEric Dumazet tcp_ecn_send(sk, skb, th, tcp_header_size); 1195ea1627c2SEric Dumazet } else { 1196ea1627c2SEric Dumazet /* RFC1323: The window in SYN & SYN/ACK segments 1197ea1627c2SEric Dumazet * is never scaled. 1198ea1627c2SEric Dumazet */ 1199ea1627c2SEric Dumazet th->window = htons(min(tp->rcv_wnd, 65535U)); 1200ea1627c2SEric Dumazet } 1201cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 1202cfb6eeb4SYOSHIFUJI Hideaki /* Calculate the MD5 hash, as we have all we need now */ 1203cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 1204a465419bSEric Dumazet sk_nocaps_add(sk, NETIF_F_GSO_MASK); 1205bd0388aeSWilliam Allen Simpson tp->af_specific->calc_md5_hash(opts.hash_location, 120639f8e58eSEric Dumazet md5, sk, skb); 1207cfb6eeb4SYOSHIFUJI Hideaki } 1208cfb6eeb4SYOSHIFUJI Hideaki #endif 1209cfb6eeb4SYOSHIFUJI Hideaki 1210bb296246SHerbert Xu icsk->icsk_af_ops->send_check(sk, skb); 12111da177e4SLinus Torvalds 12124de075e0SEric Dumazet if (likely(tcb->tcp_flags & TCPHDR_ACK)) 121327cde44aSYuchung Cheng tcp_event_ack_sent(sk, tcp_skb_pcount(skb), rcv_nxt); 12141da177e4SLinus Torvalds 1215a44d6eacSMartin KaFai Lau if (skb->len != tcp_header_size) { 1216cf533ea5SEric Dumazet tcp_event_data_sent(tp, sk); 1217a44d6eacSMartin KaFai Lau tp->data_segs_out += tcp_skb_pcount(skb); 1218ba113c3aSWei Wang tp->bytes_sent += skb->len - tcp_header_size; 1219a44d6eacSMartin KaFai Lau } 12201da177e4SLinus Torvalds 1221bd37a088SWei Yongjun if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) 1222aa2ea058STom Herbert TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, 1223aa2ea058STom Herbert tcp_skb_pcount(skb)); 12241da177e4SLinus Torvalds 12252efd055cSMarcelo Ricardo Leitner tp->segs_out += tcp_skb_pcount(skb); 1226f69ad292SEric Dumazet /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */ 1227cd7d8498SEric Dumazet skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb); 1228f69ad292SEric Dumazet skb_shinfo(skb)->gso_size = tcp_skb_mss(skb); 1229cd7d8498SEric Dumazet 1230d3edd06eSEric Dumazet /* Leave earliest departure time in skb->tstamp (skb->skb_mstamp_ns) */ 1231971f10ecSEric Dumazet 1232971f10ecSEric Dumazet /* Cleanup our debris for IP stacks */ 1233971f10ecSEric Dumazet memset(skb->cb, 0, max(sizeof(struct inet_skb_parm), 1234971f10ecSEric Dumazet sizeof(struct inet6_skb_parm))); 1235971f10ecSEric Dumazet 1236a842fe14SEric Dumazet tcp_add_tx_delay(skb, tp); 1237a842fe14SEric Dumazet 1238b0270e91SEric Dumazet err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl); 12397faee5c0SEric Dumazet 12408c72c65bSEric Dumazet if (unlikely(err > 0)) { 12415ee2c941SChristoph Paasch tcp_enter_cwr(sk); 12428c72c65bSEric Dumazet err = net_xmit_eval(err); 12438c72c65bSEric Dumazet } 1244fc225799SEric Dumazet if (!err && oskb) { 1245a7a25630SEric Dumazet tcp_update_skb_after_send(sk, oskb, prior_wstamp); 1246fc225799SEric Dumazet tcp_rate_skb_sent(sk, oskb); 1247fc225799SEric Dumazet } 12488c72c65bSEric Dumazet return err; 12491da177e4SLinus Torvalds } 12501da177e4SLinus Torvalds 12512987babbSYuchung Cheng static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, 12522987babbSYuchung Cheng gfp_t gfp_mask) 12532987babbSYuchung Cheng { 12542987babbSYuchung Cheng return __tcp_transmit_skb(sk, skb, clone_it, gfp_mask, 12552987babbSYuchung Cheng tcp_sk(sk)->rcv_nxt); 12562987babbSYuchung Cheng } 12572987babbSYuchung Cheng 125867edfef7SAndi Kleen /* This routine just queues the buffer for sending. 12591da177e4SLinus Torvalds * 12601da177e4SLinus Torvalds * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, 12611da177e4SLinus Torvalds * otherwise socket can stall. 12621da177e4SLinus Torvalds */ 12631da177e4SLinus Torvalds static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) 12641da177e4SLinus Torvalds { 12651da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 12661da177e4SLinus Torvalds 12671da177e4SLinus Torvalds /* Advance write_seq and place onto the write_queue. */ 12680f317464SEric Dumazet WRITE_ONCE(tp->write_seq, TCP_SKB_CB(skb)->end_seq); 1269f4a775d1SEric Dumazet __skb_header_release(skb); 1270fe067e8aSDavid S. Miller tcp_add_write_queue_tail(sk, skb); 1271ab4e846aSEric Dumazet sk_wmem_queued_add(sk, skb->truesize); 12723ab224beSHideo Aoki sk_mem_charge(sk, skb->truesize); 12731da177e4SLinus Torvalds } 12741da177e4SLinus Torvalds 127567edfef7SAndi Kleen /* Initialize TSO segments for a packet. */ 12765bbb432cSEric Dumazet static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now) 1277f6302d1dSDavid S. Miller { 12784a64fd6cSEric Dumazet if (skb->len <= mss_now) { 1279f6302d1dSDavid S. Miller /* Avoid the costly divide in the normal 1280f6302d1dSDavid S. Miller * non-TSO case. 1281f6302d1dSDavid S. Miller */ 1282cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, 1); 1283f69ad292SEric Dumazet TCP_SKB_CB(skb)->tcp_gso_size = 0; 1284f6302d1dSDavid S. Miller } else { 1285cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now)); 1286f69ad292SEric Dumazet TCP_SKB_CB(skb)->tcp_gso_size = mss_now; 12871da177e4SLinus Torvalds } 12881da177e4SLinus Torvalds } 12891da177e4SLinus Torvalds 1290797108d1SIlpo Järvinen /* Pcount in the middle of the write queue got changed, we need to do various 1291797108d1SIlpo Järvinen * tweaks to fix counters 1292797108d1SIlpo Järvinen */ 1293cf533ea5SEric Dumazet static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) 1294797108d1SIlpo Järvinen { 1295797108d1SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1296797108d1SIlpo Järvinen 1297797108d1SIlpo Järvinen tp->packets_out -= decr; 1298797108d1SIlpo Järvinen 1299797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 1300797108d1SIlpo Järvinen tp->sacked_out -= decr; 1301797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 1302797108d1SIlpo Järvinen tp->retrans_out -= decr; 1303797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 1304797108d1SIlpo Järvinen tp->lost_out -= decr; 1305797108d1SIlpo Järvinen 1306797108d1SIlpo Järvinen /* Reno case is special. Sigh... */ 1307797108d1SIlpo Järvinen if (tcp_is_reno(tp) && decr > 0) 1308797108d1SIlpo Järvinen tp->sacked_out -= min_t(u32, tp->sacked_out, decr); 1309797108d1SIlpo Järvinen 1310797108d1SIlpo Järvinen if (tp->lost_skb_hint && 1311797108d1SIlpo Järvinen before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && 1312713bafeaSYuchung Cheng (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) 1313797108d1SIlpo Järvinen tp->lost_cnt_hint -= decr; 1314797108d1SIlpo Järvinen 1315797108d1SIlpo Järvinen tcp_verify_left_out(tp); 1316797108d1SIlpo Järvinen } 1317797108d1SIlpo Järvinen 13180a2cf20cSSoheil Hassas Yeganeh static bool tcp_has_tx_tstamp(const struct sk_buff *skb) 13190a2cf20cSSoheil Hassas Yeganeh { 13200a2cf20cSSoheil Hassas Yeganeh return TCP_SKB_CB(skb)->txstamp_ack || 13210a2cf20cSSoheil Hassas Yeganeh (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP); 13220a2cf20cSSoheil Hassas Yeganeh } 13230a2cf20cSSoheil Hassas Yeganeh 1324490cc7d0SWillem de Bruijn static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2) 1325490cc7d0SWillem de Bruijn { 1326490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo = skb_shinfo(skb); 1327490cc7d0SWillem de Bruijn 13280a2cf20cSSoheil Hassas Yeganeh if (unlikely(tcp_has_tx_tstamp(skb)) && 1329490cc7d0SWillem de Bruijn !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) { 1330490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo2 = skb_shinfo(skb2); 1331490cc7d0SWillem de Bruijn u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP; 1332490cc7d0SWillem de Bruijn 1333490cc7d0SWillem de Bruijn shinfo->tx_flags &= ~tsflags; 1334490cc7d0SWillem de Bruijn shinfo2->tx_flags |= tsflags; 1335490cc7d0SWillem de Bruijn swap(shinfo->tskey, shinfo2->tskey); 1336b51e13faSMartin KaFai Lau TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack; 1337b51e13faSMartin KaFai Lau TCP_SKB_CB(skb)->txstamp_ack = 0; 1338490cc7d0SWillem de Bruijn } 1339490cc7d0SWillem de Bruijn } 1340490cc7d0SWillem de Bruijn 1341a166140eSMartin KaFai Lau static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2) 1342a166140eSMartin KaFai Lau { 1343a166140eSMartin KaFai Lau TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor; 1344a166140eSMartin KaFai Lau TCP_SKB_CB(skb)->eor = 0; 1345a166140eSMartin KaFai Lau } 1346a166140eSMartin KaFai Lau 134775c119afSEric Dumazet /* Insert buff after skb on the write or rtx queue of sk. */ 134875c119afSEric Dumazet static void tcp_insert_write_queue_after(struct sk_buff *skb, 134975c119afSEric Dumazet struct sk_buff *buff, 135075c119afSEric Dumazet struct sock *sk, 135175c119afSEric Dumazet enum tcp_queue tcp_queue) 135275c119afSEric Dumazet { 135375c119afSEric Dumazet if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE) 135475c119afSEric Dumazet __skb_queue_after(&sk->sk_write_queue, skb, buff); 135575c119afSEric Dumazet else 135675c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, buff); 135775c119afSEric Dumazet } 135875c119afSEric Dumazet 13591da177e4SLinus Torvalds /* Function to create two new TCP segments. Shrinks the given segment 13601da177e4SLinus Torvalds * to the specified size and appends a new segment with the rest of the 13611da177e4SLinus Torvalds * packet to the list. This won't be called frequently, I hope. 13621da177e4SLinus Torvalds * Remember, these are still headerless SKBs at this point. 13631da177e4SLinus Torvalds */ 136475c119afSEric Dumazet int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue, 136575c119afSEric Dumazet struct sk_buff *skb, u32 len, 13666cc55e09SOctavian Purdila unsigned int mss_now, gfp_t gfp) 13671da177e4SLinus Torvalds { 13681da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 13691da177e4SLinus Torvalds struct sk_buff *buff; 13706475be16SDavid S. Miller int nsize, old_factor; 1371b617158dSEric Dumazet long limit; 1372b60b49eaSHerbert Xu int nlen; 13739ce01461SIlpo Järvinen u8 flags; 13741da177e4SLinus Torvalds 13752fceec13SIlpo Järvinen if (WARN_ON(len > skb->len)) 13762fceec13SIlpo Järvinen return -EINVAL; 13776a438bbeSStephen Hemminger 13781da177e4SLinus Torvalds nsize = skb_headlen(skb) - len; 13791da177e4SLinus Torvalds if (nsize < 0) 13801da177e4SLinus Torvalds nsize = 0; 13811da177e4SLinus Torvalds 1382b617158dSEric Dumazet /* tcp_sendmsg() can overshoot sk_wmem_queued by one full size skb. 1383b617158dSEric Dumazet * We need some allowance to not penalize applications setting small 1384b617158dSEric Dumazet * SO_SNDBUF values. 1385b617158dSEric Dumazet * Also allow first and last skb in retransmit queue to be split. 1386b617158dSEric Dumazet */ 1387b617158dSEric Dumazet limit = sk->sk_sndbuf + 2 * SKB_TRUESIZE(GSO_MAX_SIZE); 1388b617158dSEric Dumazet if (unlikely((sk->sk_wmem_queued >> 1) > limit && 1389b617158dSEric Dumazet tcp_queue != TCP_FRAG_IN_WRITE_QUEUE && 1390b617158dSEric Dumazet skb != tcp_rtx_queue_head(sk) && 1391b617158dSEric Dumazet skb != tcp_rtx_queue_tail(sk))) { 1392f070ef2aSEric Dumazet NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPWQUEUETOOBIG); 1393f070ef2aSEric Dumazet return -ENOMEM; 1394f070ef2aSEric Dumazet } 1395f070ef2aSEric Dumazet 13966cc55e09SOctavian Purdila if (skb_unclone(skb, gfp)) 13971da177e4SLinus Torvalds return -ENOMEM; 13981da177e4SLinus Torvalds 13991da177e4SLinus Torvalds /* Get a new skb... force flag on. */ 1400eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, nsize, gfp, true); 140151456b29SIan Morris if (!buff) 14021da177e4SLinus Torvalds return -ENOMEM; /* We'll just try again later. */ 140341477662SJakub Kicinski skb_copy_decrypted(buff, skb); 1404ef5cb973SHerbert Xu 1405ab4e846aSEric Dumazet sk_wmem_queued_add(sk, buff->truesize); 14063ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1407b60b49eaSHerbert Xu nlen = skb->len - len - nsize; 1408b60b49eaSHerbert Xu buff->truesize += nlen; 1409b60b49eaSHerbert Xu skb->truesize -= nlen; 14101da177e4SLinus Torvalds 14111da177e4SLinus Torvalds /* Correct the sequence numbers. */ 14121da177e4SLinus Torvalds TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 14131da177e4SLinus Torvalds TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 14141da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 14151da177e4SLinus Torvalds 14161da177e4SLinus Torvalds /* PSH and FIN should only be set in the second packet. */ 14174de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 14184de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 14194de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1420e14c3cafSHerbert Xu TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; 1421a166140eSMartin KaFai Lau tcp_skb_fragment_eor(skb, buff); 14221da177e4SLinus Torvalds 14231da177e4SLinus Torvalds skb_split(skb, buff, len); 14241da177e4SLinus Torvalds 142598be9b12SEric Dumazet buff->ip_summed = CHECKSUM_PARTIAL; 14261da177e4SLinus Torvalds 1427a61bbcf2SPatrick McHardy buff->tstamp = skb->tstamp; 1428490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 14291da177e4SLinus Torvalds 14306475be16SDavid S. Miller old_factor = tcp_skb_pcount(skb); 14316475be16SDavid S. Miller 14321da177e4SLinus Torvalds /* Fix up tso_factor for both original and new SKB. */ 14335bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 14345bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 14351da177e4SLinus Torvalds 1436b9f64820SYuchung Cheng /* Update delivered info for the new segment */ 1437b9f64820SYuchung Cheng TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx; 1438b9f64820SYuchung Cheng 14396475be16SDavid S. Miller /* If this packet has been sent out already, we must 14406475be16SDavid S. Miller * adjust the various packet counters. 14416475be16SDavid S. Miller */ 1442cf0b450cSHerbert Xu if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 14436475be16SDavid S. Miller int diff = old_factor - tcp_skb_pcount(skb) - 14446475be16SDavid S. Miller tcp_skb_pcount(buff); 14451da177e4SLinus Torvalds 1446797108d1SIlpo Järvinen if (diff) 1447797108d1SIlpo Järvinen tcp_adjust_pcount(sk, skb, diff); 14481da177e4SLinus Torvalds } 14491da177e4SLinus Torvalds 14501da177e4SLinus Torvalds /* Link BUFF into the send queue. */ 1451f4a775d1SEric Dumazet __skb_header_release(buff); 145275c119afSEric Dumazet tcp_insert_write_queue_after(skb, buff, sk, tcp_queue); 1453f67971e6SEric Dumazet if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE) 1454e2080072SEric Dumazet list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor); 14551da177e4SLinus Torvalds 14561da177e4SLinus Torvalds return 0; 14571da177e4SLinus Torvalds } 14581da177e4SLinus Torvalds 1459f4d01666SEric Dumazet /* This is similar to __pskb_pull_tail(). The difference is that pulled 1460f4d01666SEric Dumazet * data is not copied, but immediately discarded. 14611da177e4SLinus Torvalds */ 14627162fb24SEric Dumazet static int __pskb_trim_head(struct sk_buff *skb, int len) 14631da177e4SLinus Torvalds { 14647b7fc97aSEric Dumazet struct skb_shared_info *shinfo; 14651da177e4SLinus Torvalds int i, k, eat; 14661da177e4SLinus Torvalds 14674fa48bf3SEric Dumazet eat = min_t(int, len, skb_headlen(skb)); 14684fa48bf3SEric Dumazet if (eat) { 14694fa48bf3SEric Dumazet __skb_pull(skb, eat); 14704fa48bf3SEric Dumazet len -= eat; 14714fa48bf3SEric Dumazet if (!len) 14727162fb24SEric Dumazet return 0; 14734fa48bf3SEric Dumazet } 14741da177e4SLinus Torvalds eat = len; 14751da177e4SLinus Torvalds k = 0; 14767b7fc97aSEric Dumazet shinfo = skb_shinfo(skb); 14777b7fc97aSEric Dumazet for (i = 0; i < shinfo->nr_frags; i++) { 14787b7fc97aSEric Dumazet int size = skb_frag_size(&shinfo->frags[i]); 14799e903e08SEric Dumazet 14809e903e08SEric Dumazet if (size <= eat) { 1481aff65da0SIan Campbell skb_frag_unref(skb, i); 14829e903e08SEric Dumazet eat -= size; 14831da177e4SLinus Torvalds } else { 14847b7fc97aSEric Dumazet shinfo->frags[k] = shinfo->frags[i]; 14851da177e4SLinus Torvalds if (eat) { 1486b54c9d5bSJonathan Lemon skb_frag_off_add(&shinfo->frags[k], eat); 14877b7fc97aSEric Dumazet skb_frag_size_sub(&shinfo->frags[k], eat); 14881da177e4SLinus Torvalds eat = 0; 14891da177e4SLinus Torvalds } 14901da177e4SLinus Torvalds k++; 14911da177e4SLinus Torvalds } 14921da177e4SLinus Torvalds } 14937b7fc97aSEric Dumazet shinfo->nr_frags = k; 14941da177e4SLinus Torvalds 14951da177e4SLinus Torvalds skb->data_len -= len; 14961da177e4SLinus Torvalds skb->len = skb->data_len; 14977162fb24SEric Dumazet return len; 14981da177e4SLinus Torvalds } 14991da177e4SLinus Torvalds 150067edfef7SAndi Kleen /* Remove acked data from a packet in the transmit queue. */ 15011da177e4SLinus Torvalds int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 15021da177e4SLinus Torvalds { 15037162fb24SEric Dumazet u32 delta_truesize; 15047162fb24SEric Dumazet 150514bbd6a5SPravin B Shelar if (skb_unclone(skb, GFP_ATOMIC)) 15061da177e4SLinus Torvalds return -ENOMEM; 15071da177e4SLinus Torvalds 15087162fb24SEric Dumazet delta_truesize = __pskb_trim_head(skb, len); 15091da177e4SLinus Torvalds 15101da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq += len; 151184fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 15121da177e4SLinus Torvalds 15137162fb24SEric Dumazet if (delta_truesize) { 15147162fb24SEric Dumazet skb->truesize -= delta_truesize; 1515ab4e846aSEric Dumazet sk_wmem_queued_add(sk, -delta_truesize); 15167162fb24SEric Dumazet sk_mem_uncharge(sk, delta_truesize); 15171da177e4SLinus Torvalds sock_set_flag(sk, SOCK_QUEUE_SHRUNK); 15187162fb24SEric Dumazet } 15191da177e4SLinus Torvalds 15205b35e1e6SNeal Cardwell /* Any change of skb->len requires recalculation of tso factor. */ 15211da177e4SLinus Torvalds if (tcp_skb_pcount(skb) > 1) 15225bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb)); 15231da177e4SLinus Torvalds 15241da177e4SLinus Torvalds return 0; 15251da177e4SLinus Torvalds } 15261da177e4SLinus Torvalds 15271b63edd6SYuchung Cheng /* Calculate MSS not accounting any TCP options. */ 15281b63edd6SYuchung Cheng static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu) 15295d424d5aSJohn Heffner { 1530cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1531cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 15325d424d5aSJohn Heffner int mss_now; 15335d424d5aSJohn Heffner 15345d424d5aSJohn Heffner /* Calculate base mss without TCP options: 15355d424d5aSJohn Heffner It is MMS_S - sizeof(tcphdr) of rfc1122 15365d424d5aSJohn Heffner */ 15375d424d5aSJohn Heffner mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 15385d424d5aSJohn Heffner 153967469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 154067469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 154167469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 154267469601SEric Dumazet 154367469601SEric Dumazet if (dst && dst_allfrag(dst)) 154467469601SEric Dumazet mss_now -= icsk->icsk_af_ops->net_frag_header_len; 154567469601SEric Dumazet } 154667469601SEric Dumazet 15475d424d5aSJohn Heffner /* Clamp it (mss_clamp does not include tcp options) */ 15485d424d5aSJohn Heffner if (mss_now > tp->rx_opt.mss_clamp) 15495d424d5aSJohn Heffner mss_now = tp->rx_opt.mss_clamp; 15505d424d5aSJohn Heffner 15515d424d5aSJohn Heffner /* Now subtract optional transport overhead */ 15525d424d5aSJohn Heffner mss_now -= icsk->icsk_ext_hdr_len; 15535d424d5aSJohn Heffner 15545d424d5aSJohn Heffner /* Then reserve room for full set of TCP options and 8 bytes of data */ 15555f3e2bf0SEric Dumazet mss_now = max(mss_now, sock_net(sk)->ipv4.sysctl_tcp_min_snd_mss); 15565d424d5aSJohn Heffner return mss_now; 15575d424d5aSJohn Heffner } 15585d424d5aSJohn Heffner 15591b63edd6SYuchung Cheng /* Calculate MSS. Not accounting for SACKs here. */ 15601b63edd6SYuchung Cheng int tcp_mtu_to_mss(struct sock *sk, int pmtu) 15611b63edd6SYuchung Cheng { 15621b63edd6SYuchung Cheng /* Subtract TCP options size, not including SACKs */ 15631b63edd6SYuchung Cheng return __tcp_mtu_to_mss(sk, pmtu) - 15641b63edd6SYuchung Cheng (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr)); 15651b63edd6SYuchung Cheng } 15661b63edd6SYuchung Cheng 15675d424d5aSJohn Heffner /* Inverse of above */ 156867469601SEric Dumazet int tcp_mss_to_mtu(struct sock *sk, int mss) 15695d424d5aSJohn Heffner { 1570cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1571cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 15725d424d5aSJohn Heffner int mtu; 15735d424d5aSJohn Heffner 15745d424d5aSJohn Heffner mtu = mss + 15755d424d5aSJohn Heffner tp->tcp_header_len + 15765d424d5aSJohn Heffner icsk->icsk_ext_hdr_len + 15775d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 15785d424d5aSJohn Heffner 157967469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 158067469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 158167469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 158267469601SEric Dumazet 158367469601SEric Dumazet if (dst && dst_allfrag(dst)) 158467469601SEric Dumazet mtu += icsk->icsk_af_ops->net_frag_header_len; 158567469601SEric Dumazet } 15865d424d5aSJohn Heffner return mtu; 15875d424d5aSJohn Heffner } 1588556c6b46SNeal Cardwell EXPORT_SYMBOL(tcp_mss_to_mtu); 15895d424d5aSJohn Heffner 159067edfef7SAndi Kleen /* MTU probing init per socket */ 15915d424d5aSJohn Heffner void tcp_mtup_init(struct sock *sk) 15925d424d5aSJohn Heffner { 15935d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 15945d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 1595b0f9ca53SFan Du struct net *net = sock_net(sk); 15965d424d5aSJohn Heffner 1597b0f9ca53SFan Du icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1; 15985d424d5aSJohn Heffner icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 15995d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 1600b0f9ca53SFan Du icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss); 16015d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 160205cbc0dbSFan Du if (icsk->icsk_mtup.enabled) 1603c74df29aSEric Dumazet icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; 16045d424d5aSJohn Heffner } 16054bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_mtup_init); 16065d424d5aSJohn Heffner 16071da177e4SLinus Torvalds /* This function synchronize snd mss to current pmtu/exthdr set. 16081da177e4SLinus Torvalds 16091da177e4SLinus Torvalds tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 16101da177e4SLinus Torvalds for TCP options, but includes only bare TCP header. 16111da177e4SLinus Torvalds 16121da177e4SLinus Torvalds tp->rx_opt.mss_clamp is mss negotiated at connection setup. 1613caa20d9aSStephen Hemminger It is minimum of user_mss and mss received with SYN. 16141da177e4SLinus Torvalds It also does not include TCP options. 16151da177e4SLinus Torvalds 1616d83d8461SArnaldo Carvalho de Melo inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. 16171da177e4SLinus Torvalds 16181da177e4SLinus Torvalds tp->mss_cache is current effective sending mss, including 16191da177e4SLinus Torvalds all tcp options except for SACKs. It is evaluated, 16201da177e4SLinus Torvalds taking into account current pmtu, but never exceeds 16211da177e4SLinus Torvalds tp->rx_opt.mss_clamp. 16221da177e4SLinus Torvalds 16231da177e4SLinus Torvalds NOTE1. rfc1122 clearly states that advertised MSS 16241da177e4SLinus Torvalds DOES NOT include either tcp or ip options. 16251da177e4SLinus Torvalds 1626d83d8461SArnaldo Carvalho de Melo NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache 1627d83d8461SArnaldo Carvalho de Melo are READ ONLY outside this function. --ANK (980731) 16281da177e4SLinus Torvalds */ 16291da177e4SLinus Torvalds unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 16301da177e4SLinus Torvalds { 16311da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1632d83d8461SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 16335d424d5aSJohn Heffner int mss_now; 16341da177e4SLinus Torvalds 16355d424d5aSJohn Heffner if (icsk->icsk_mtup.search_high > pmtu) 16365d424d5aSJohn Heffner icsk->icsk_mtup.search_high = pmtu; 16371da177e4SLinus Torvalds 16385d424d5aSJohn Heffner mss_now = tcp_mtu_to_mss(sk, pmtu); 1639409d22b4SIlpo Järvinen mss_now = tcp_bound_to_half_wnd(tp, mss_now); 16401da177e4SLinus Torvalds 16411da177e4SLinus Torvalds /* And store cached results */ 1642d83d8461SArnaldo Carvalho de Melo icsk->icsk_pmtu_cookie = pmtu; 16435d424d5aSJohn Heffner if (icsk->icsk_mtup.enabled) 16445d424d5aSJohn Heffner mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); 1645c1b4a7e6SDavid S. Miller tp->mss_cache = mss_now; 16461da177e4SLinus Torvalds 16471da177e4SLinus Torvalds return mss_now; 16481da177e4SLinus Torvalds } 16494bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_sync_mss); 16501da177e4SLinus Torvalds 16511da177e4SLinus Torvalds /* Compute the current effective MSS, taking SACKs and IP options, 16521da177e4SLinus Torvalds * and even PMTU discovery events into account. 16531da177e4SLinus Torvalds */ 16540c54b85fSIlpo Järvinen unsigned int tcp_current_mss(struct sock *sk) 16551da177e4SLinus Torvalds { 1656cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1657cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1658c1b4a7e6SDavid S. Miller u32 mss_now; 165995c96174SEric Dumazet unsigned int header_len; 166033ad798cSAdam Langley struct tcp_out_options opts; 166133ad798cSAdam Langley struct tcp_md5sig_key *md5; 16621da177e4SLinus Torvalds 1663c1b4a7e6SDavid S. Miller mss_now = tp->mss_cache; 1664c1b4a7e6SDavid S. Miller 16651da177e4SLinus Torvalds if (dst) { 16661da177e4SLinus Torvalds u32 mtu = dst_mtu(dst); 1667d83d8461SArnaldo Carvalho de Melo if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 16681da177e4SLinus Torvalds mss_now = tcp_sync_mss(sk, mtu); 16691da177e4SLinus Torvalds } 16701da177e4SLinus Torvalds 167133ad798cSAdam Langley header_len = tcp_established_options(sk, NULL, &opts, &md5) + 167233ad798cSAdam Langley sizeof(struct tcphdr); 167333ad798cSAdam Langley /* The mss_cache is sized based on tp->tcp_header_len, which assumes 167433ad798cSAdam Langley * some common options. If this is an odd packet (because we have SACK 167533ad798cSAdam Langley * blocks etc) then our calculated header_len will be different, and 167633ad798cSAdam Langley * we have to adjust mss_now correspondingly */ 167733ad798cSAdam Langley if (header_len != tp->tcp_header_len) { 167833ad798cSAdam Langley int delta = (int) header_len - tp->tcp_header_len; 167933ad798cSAdam Langley mss_now -= delta; 168033ad798cSAdam Langley } 1681cfb6eeb4SYOSHIFUJI Hideaki 16821da177e4SLinus Torvalds return mss_now; 16831da177e4SLinus Torvalds } 16841da177e4SLinus Torvalds 168586fd14adSWeiping Pan /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. 168686fd14adSWeiping Pan * As additional protections, we do not touch cwnd in retransmission phases, 168786fd14adSWeiping Pan * and if application hit its sndbuf limit recently. 168886fd14adSWeiping Pan */ 168986fd14adSWeiping Pan static void tcp_cwnd_application_limited(struct sock *sk) 1690a762a980SDavid S. Miller { 16919e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1692a762a980SDavid S. Miller 169386fd14adSWeiping Pan if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && 169486fd14adSWeiping Pan sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 169586fd14adSWeiping Pan /* Limited by application or receiver window. */ 169686fd14adSWeiping Pan u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); 169786fd14adSWeiping Pan u32 win_used = max(tp->snd_cwnd_used, init_win); 169886fd14adSWeiping Pan if (win_used < tp->snd_cwnd) { 169986fd14adSWeiping Pan tp->snd_ssthresh = tcp_current_ssthresh(sk); 170086fd14adSWeiping Pan tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1; 170186fd14adSWeiping Pan } 170286fd14adSWeiping Pan tp->snd_cwnd_used = 0; 170386fd14adSWeiping Pan } 1704c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 170586fd14adSWeiping Pan } 170686fd14adSWeiping Pan 1707ca8a2263SNeal Cardwell static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited) 1708a762a980SDavid S. Miller { 17091b1fc3fdSWei Wang const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 1710a762a980SDavid S. Miller struct tcp_sock *tp = tcp_sk(sk); 1711a762a980SDavid S. Miller 1712ca8a2263SNeal Cardwell /* Track the maximum number of outstanding packets in each 1713ca8a2263SNeal Cardwell * window, and remember whether we were cwnd-limited then. 1714ca8a2263SNeal Cardwell */ 1715ca8a2263SNeal Cardwell if (!before(tp->snd_una, tp->max_packets_seq) || 1716ca8a2263SNeal Cardwell tp->packets_out > tp->max_packets_out) { 1717ca8a2263SNeal Cardwell tp->max_packets_out = tp->packets_out; 1718ca8a2263SNeal Cardwell tp->max_packets_seq = tp->snd_nxt; 1719ca8a2263SNeal Cardwell tp->is_cwnd_limited = is_cwnd_limited; 1720ca8a2263SNeal Cardwell } 1721e114a710SEric Dumazet 172224901551SEric Dumazet if (tcp_is_cwnd_limited(sk)) { 1723a762a980SDavid S. Miller /* Network is feed fully. */ 1724a762a980SDavid S. Miller tp->snd_cwnd_used = 0; 1725c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 1726a762a980SDavid S. Miller } else { 1727a762a980SDavid S. Miller /* Network starves. */ 1728a762a980SDavid S. Miller if (tp->packets_out > tp->snd_cwnd_used) 1729a762a980SDavid S. Miller tp->snd_cwnd_used = tp->packets_out; 1730a762a980SDavid S. Miller 1731b510f0d2SEric Dumazet if (sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle && 1732c2203cf7SEric Dumazet (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto && 17331b1fc3fdSWei Wang !ca_ops->cong_control) 1734a762a980SDavid S. Miller tcp_cwnd_application_limited(sk); 1735b0f71bd3SFrancis Yan 1736b0f71bd3SFrancis Yan /* The following conditions together indicate the starvation 1737b0f71bd3SFrancis Yan * is caused by insufficient sender buffer: 1738b0f71bd3SFrancis Yan * 1) just sent some data (see tcp_write_xmit) 1739b0f71bd3SFrancis Yan * 2) not cwnd limited (this else condition) 174075c119afSEric Dumazet * 3) no more data to send (tcp_write_queue_empty()) 1741b0f71bd3SFrancis Yan * 4) application is hitting buffer limit (SOCK_NOSPACE) 1742b0f71bd3SFrancis Yan */ 174375c119afSEric Dumazet if (tcp_write_queue_empty(sk) && sk->sk_socket && 1744b0f71bd3SFrancis Yan test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) && 1745b0f71bd3SFrancis Yan (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) 1746b0f71bd3SFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED); 1747a762a980SDavid S. Miller } 1748a762a980SDavid S. Miller } 1749a762a980SDavid S. Miller 1750d4589926SEric Dumazet /* Minshall's variant of the Nagle send check. */ 1751d4589926SEric Dumazet static bool tcp_minshall_check(const struct tcp_sock *tp) 1752d4589926SEric Dumazet { 1753d4589926SEric Dumazet return after(tp->snd_sml, tp->snd_una) && 1754d4589926SEric Dumazet !after(tp->snd_sml, tp->snd_nxt); 1755d4589926SEric Dumazet } 1756d4589926SEric Dumazet 1757d4589926SEric Dumazet /* Update snd_sml if this skb is under mss 1758d4589926SEric Dumazet * Note that a TSO packet might end with a sub-mss segment 1759d4589926SEric Dumazet * The test is really : 1760d4589926SEric Dumazet * if ((skb->len % mss) != 0) 1761d4589926SEric Dumazet * tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1762d4589926SEric Dumazet * But we can avoid doing the divide again given we already have 1763d4589926SEric Dumazet * skb_pcount = skb->len / mss_now 17640e3a4803SIlpo Järvinen */ 1765d4589926SEric Dumazet static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now, 1766d4589926SEric Dumazet const struct sk_buff *skb) 1767d4589926SEric Dumazet { 1768d4589926SEric Dumazet if (skb->len < tcp_skb_pcount(skb) * mss_now) 1769d4589926SEric Dumazet tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1770d4589926SEric Dumazet } 1771d4589926SEric Dumazet 1772d4589926SEric Dumazet /* Return false, if packet can be sent now without violation Nagle's rules: 1773d4589926SEric Dumazet * 1. It is full sized. (provided by caller in %partial bool) 1774d4589926SEric Dumazet * 2. Or it contains FIN. (already checked by caller) 1775d4589926SEric Dumazet * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. 1776d4589926SEric Dumazet * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 1777d4589926SEric Dumazet * With Minshall's modification: all sent small packets are ACKed. 1778d4589926SEric Dumazet */ 1779d4589926SEric Dumazet static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp, 1780cc93fc51SPeter Pan(潘卫平) int nonagle) 1781d4589926SEric Dumazet { 1782d4589926SEric Dumazet return partial && 1783d4589926SEric Dumazet ((nonagle & TCP_NAGLE_CORK) || 1784d4589926SEric Dumazet (!nonagle && tp->packets_out && tcp_minshall_check(tp))); 1785d4589926SEric Dumazet } 1786605ad7f1SEric Dumazet 1787605ad7f1SEric Dumazet /* Return how many segs we'd like on a TSO packet, 1788605ad7f1SEric Dumazet * to send one TSO packet per ms 1789605ad7f1SEric Dumazet */ 1790dcb8c9b4SEric Dumazet static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now, 17911b3878caSNeal Cardwell int min_tso_segs) 1792605ad7f1SEric Dumazet { 1793605ad7f1SEric Dumazet u32 bytes, segs; 1794605ad7f1SEric Dumazet 179576a9ebe8SEric Dumazet bytes = min_t(unsigned long, 17967c68fa2bSEric Dumazet sk->sk_pacing_rate >> READ_ONCE(sk->sk_pacing_shift), 1797605ad7f1SEric Dumazet sk->sk_gso_max_size - 1 - MAX_TCP_HEADER); 1798605ad7f1SEric Dumazet 1799605ad7f1SEric Dumazet /* Goal is to send at least one packet per ms, 1800605ad7f1SEric Dumazet * not one big TSO packet every 100 ms. 1801605ad7f1SEric Dumazet * This preserves ACK clocking and is consistent 1802605ad7f1SEric Dumazet * with tcp_tso_should_defer() heuristic. 1803605ad7f1SEric Dumazet */ 18041b3878caSNeal Cardwell segs = max_t(u32, bytes / mss_now, min_tso_segs); 1805605ad7f1SEric Dumazet 1806350c9f48SEric Dumazet return segs; 1807605ad7f1SEric Dumazet } 1808605ad7f1SEric Dumazet 1809ed6e7268SNeal Cardwell /* Return the number of segments we want in the skb we are transmitting. 1810ed6e7268SNeal Cardwell * See if congestion control module wants to decide; otherwise, autosize. 1811ed6e7268SNeal Cardwell */ 1812ed6e7268SNeal Cardwell static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now) 1813ed6e7268SNeal Cardwell { 1814ed6e7268SNeal Cardwell const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 1815dcb8c9b4SEric Dumazet u32 min_tso, tso_segs; 1816ed6e7268SNeal Cardwell 1817dcb8c9b4SEric Dumazet min_tso = ca_ops->min_tso_segs ? 1818dcb8c9b4SEric Dumazet ca_ops->min_tso_segs(sk) : 1819dcb8c9b4SEric Dumazet sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs; 1820dcb8c9b4SEric Dumazet 1821dcb8c9b4SEric Dumazet tso_segs = tcp_tso_autosize(sk, mss_now, min_tso); 1822350c9f48SEric Dumazet return min_t(u32, tso_segs, sk->sk_gso_max_segs); 1823ed6e7268SNeal Cardwell } 1824ed6e7268SNeal Cardwell 1825d4589926SEric Dumazet /* Returns the portion of skb which can be sent right away */ 1826d4589926SEric Dumazet static unsigned int tcp_mss_split_point(const struct sock *sk, 1827d4589926SEric Dumazet const struct sk_buff *skb, 1828d4589926SEric Dumazet unsigned int mss_now, 1829d4589926SEric Dumazet unsigned int max_segs, 1830d4589926SEric Dumazet int nonagle) 1831c1b4a7e6SDavid S. Miller { 1832cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1833d4589926SEric Dumazet u32 partial, needed, window, max_len; 1834c1b4a7e6SDavid S. Miller 183590840defSIlpo Järvinen window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 18361485348dSBen Hutchings max_len = mss_now * max_segs; 18370e3a4803SIlpo Järvinen 18381485348dSBen Hutchings if (likely(max_len <= window && skb != tcp_write_queue_tail(sk))) 18391485348dSBen Hutchings return max_len; 18400e3a4803SIlpo Järvinen 18415ea3a748SIlpo Järvinen needed = min(skb->len, window); 18425ea3a748SIlpo Järvinen 18431485348dSBen Hutchings if (max_len <= needed) 18441485348dSBen Hutchings return max_len; 18450e3a4803SIlpo Järvinen 1846d4589926SEric Dumazet partial = needed % mss_now; 1847d4589926SEric Dumazet /* If last segment is not a full MSS, check if Nagle rules allow us 1848d4589926SEric Dumazet * to include this last segment in this skb. 1849d4589926SEric Dumazet * Otherwise, we'll split the skb at last MSS boundary 1850d4589926SEric Dumazet */ 1851cc93fc51SPeter Pan(潘卫平) if (tcp_nagle_check(partial != 0, tp, nonagle)) 1852d4589926SEric Dumazet return needed - partial; 1853d4589926SEric Dumazet 1854d4589926SEric Dumazet return needed; 1855c1b4a7e6SDavid S. Miller } 1856c1b4a7e6SDavid S. Miller 1857c1b4a7e6SDavid S. Miller /* Can at least one segment of SKB be sent right now, according to the 1858c1b4a7e6SDavid S. Miller * congestion window rules? If so, return how many segments are allowed. 1859c1b4a7e6SDavid S. Miller */ 1860cf533ea5SEric Dumazet static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, 1861cf533ea5SEric Dumazet const struct sk_buff *skb) 1862c1b4a7e6SDavid S. Miller { 1863d649a7a8SEric Dumazet u32 in_flight, cwnd, halfcwnd; 1864c1b4a7e6SDavid S. Miller 1865c1b4a7e6SDavid S. Miller /* Don't be strict about the congestion window for the final FIN. */ 18664de075e0SEric Dumazet if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && 18674de075e0SEric Dumazet tcp_skb_pcount(skb) == 1) 1868c1b4a7e6SDavid S. Miller return 1; 1869c1b4a7e6SDavid S. Miller 1870c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1871c1b4a7e6SDavid S. Miller cwnd = tp->snd_cwnd; 1872d649a7a8SEric Dumazet if (in_flight >= cwnd) 1873c1b4a7e6SDavid S. Miller return 0; 1874d649a7a8SEric Dumazet 1875d649a7a8SEric Dumazet /* For better scheduling, ensure we have at least 1876d649a7a8SEric Dumazet * 2 GSO packets in flight. 1877d649a7a8SEric Dumazet */ 1878d649a7a8SEric Dumazet halfcwnd = max(cwnd >> 1, 1U); 1879d649a7a8SEric Dumazet return min(halfcwnd, cwnd - in_flight); 1880c1b4a7e6SDavid S. Miller } 1881c1b4a7e6SDavid S. Miller 1882b595076aSUwe Kleine-König /* Initialize TSO state of a skb. 188367edfef7SAndi Kleen * This must be invoked the first time we consider transmitting 1884c1b4a7e6SDavid S. Miller * SKB onto the wire. 1885c1b4a7e6SDavid S. Miller */ 18865bbb432cSEric Dumazet static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now) 1887c1b4a7e6SDavid S. Miller { 1888c1b4a7e6SDavid S. Miller int tso_segs = tcp_skb_pcount(skb); 1889c1b4a7e6SDavid S. Miller 1890f8269a49SIlpo Järvinen if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { 18915bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 1892c1b4a7e6SDavid S. Miller tso_segs = tcp_skb_pcount(skb); 1893c1b4a7e6SDavid S. Miller } 1894c1b4a7e6SDavid S. Miller return tso_segs; 1895c1b4a7e6SDavid S. Miller } 1896c1b4a7e6SDavid S. Miller 1897c1b4a7e6SDavid S. Miller 1898a2a385d6SEric Dumazet /* Return true if the Nagle test allows this packet to be 1899c1b4a7e6SDavid S. Miller * sent now. 1900c1b4a7e6SDavid S. Miller */ 1901a2a385d6SEric Dumazet static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, 1902c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1903c1b4a7e6SDavid S. Miller { 1904c1b4a7e6SDavid S. Miller /* Nagle rule does not apply to frames, which sit in the middle of the 1905c1b4a7e6SDavid S. Miller * write_queue (they have no chances to get new data). 1906c1b4a7e6SDavid S. Miller * 1907c1b4a7e6SDavid S. Miller * This is implemented in the callers, where they modify the 'nonagle' 1908c1b4a7e6SDavid S. Miller * argument based upon the location of SKB in the send queue. 1909c1b4a7e6SDavid S. Miller */ 1910c1b4a7e6SDavid S. Miller if (nonagle & TCP_NAGLE_PUSH) 1911a2a385d6SEric Dumazet return true; 1912c1b4a7e6SDavid S. Miller 19139b44190dSYuchung Cheng /* Don't use the nagle rule for urgent data (or for the final FIN). */ 19149b44190dSYuchung Cheng if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) 1915a2a385d6SEric Dumazet return true; 1916c1b4a7e6SDavid S. Miller 1917cc93fc51SPeter Pan(潘卫平) if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle)) 1918a2a385d6SEric Dumazet return true; 1919c1b4a7e6SDavid S. Miller 1920a2a385d6SEric Dumazet return false; 1921c1b4a7e6SDavid S. Miller } 1922c1b4a7e6SDavid S. Miller 1923c1b4a7e6SDavid S. Miller /* Does at least the first segment of SKB fit into the send window? */ 1924a2a385d6SEric Dumazet static bool tcp_snd_wnd_test(const struct tcp_sock *tp, 1925a2a385d6SEric Dumazet const struct sk_buff *skb, 1926056834d9SIlpo Järvinen unsigned int cur_mss) 1927c1b4a7e6SDavid S. Miller { 1928c1b4a7e6SDavid S. Miller u32 end_seq = TCP_SKB_CB(skb)->end_seq; 1929c1b4a7e6SDavid S. Miller 1930c1b4a7e6SDavid S. Miller if (skb->len > cur_mss) 1931c1b4a7e6SDavid S. Miller end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 1932c1b4a7e6SDavid S. Miller 193390840defSIlpo Järvinen return !after(end_seq, tcp_wnd_end(tp)); 1934c1b4a7e6SDavid S. Miller } 1935c1b4a7e6SDavid S. Miller 1936c1b4a7e6SDavid S. Miller /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 1937c1b4a7e6SDavid S. Miller * which is put after SKB on the list. It is very much like 1938c1b4a7e6SDavid S. Miller * tcp_fragment() except that it may make several kinds of assumptions 1939c1b4a7e6SDavid S. Miller * in order to speed up the splitting operation. In particular, we 1940c1b4a7e6SDavid S. Miller * know that all the data is in scatter-gather pages, and that the 1941c1b4a7e6SDavid S. Miller * packet has never been sent out before (and thus is not cloned). 1942c1b4a7e6SDavid S. Miller */ 194356483341SEric Dumazet static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, 1944c4ead4c5SEric Dumazet unsigned int mss_now, gfp_t gfp) 1945c1b4a7e6SDavid S. Miller { 1946c1b4a7e6SDavid S. Miller int nlen = skb->len - len; 194756483341SEric Dumazet struct sk_buff *buff; 19489ce01461SIlpo Järvinen u8 flags; 1949c1b4a7e6SDavid S. Miller 1950c1b4a7e6SDavid S. Miller /* All of a TSO frame must be composed of paged data. */ 1951c8ac3774SHerbert Xu if (skb->len != skb->data_len) 195256483341SEric Dumazet return tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE, 195356483341SEric Dumazet skb, len, mss_now, gfp); 1954c1b4a7e6SDavid S. Miller 1955eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, gfp, true); 195651456b29SIan Morris if (unlikely(!buff)) 1957c1b4a7e6SDavid S. Miller return -ENOMEM; 195841477662SJakub Kicinski skb_copy_decrypted(buff, skb); 1959c1b4a7e6SDavid S. Miller 1960ab4e846aSEric Dumazet sk_wmem_queued_add(sk, buff->truesize); 19613ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1962b60b49eaSHerbert Xu buff->truesize += nlen; 1963c1b4a7e6SDavid S. Miller skb->truesize -= nlen; 1964c1b4a7e6SDavid S. Miller 1965c1b4a7e6SDavid S. Miller /* Correct the sequence numbers. */ 1966c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 1967c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 1968c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 1969c1b4a7e6SDavid S. Miller 1970c1b4a7e6SDavid S. Miller /* PSH and FIN should only be set in the second packet. */ 19714de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 19724de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 19734de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1974c1b4a7e6SDavid S. Miller 1975c1b4a7e6SDavid S. Miller /* This packet was never sent out yet, so no SACK bits. */ 1976c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->sacked = 0; 1977c1b4a7e6SDavid S. Miller 1978a166140eSMartin KaFai Lau tcp_skb_fragment_eor(skb, buff); 1979a166140eSMartin KaFai Lau 198098be9b12SEric Dumazet buff->ip_summed = CHECKSUM_PARTIAL; 1981c1b4a7e6SDavid S. Miller skb_split(skb, buff, len); 1982490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 1983c1b4a7e6SDavid S. Miller 1984c1b4a7e6SDavid S. Miller /* Fix up tso_factor for both original and new SKB. */ 19855bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 19865bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 1987c1b4a7e6SDavid S. Miller 1988c1b4a7e6SDavid S. Miller /* Link BUFF into the send queue. */ 1989f4a775d1SEric Dumazet __skb_header_release(buff); 199056483341SEric Dumazet tcp_insert_write_queue_after(skb, buff, sk, TCP_FRAG_IN_WRITE_QUEUE); 1991c1b4a7e6SDavid S. Miller 1992c1b4a7e6SDavid S. Miller return 0; 1993c1b4a7e6SDavid S. Miller } 1994c1b4a7e6SDavid S. Miller 1995c1b4a7e6SDavid S. Miller /* Try to defer sending, if possible, in order to minimize the amount 1996c1b4a7e6SDavid S. Miller * of TSO splitting we do. View it as a kind of TSO Nagle test. 1997c1b4a7e6SDavid S. Miller * 1998c1b4a7e6SDavid S. Miller * This algorithm is from John Heffner. 1999c1b4a7e6SDavid S. Miller */ 2000ca8a2263SNeal Cardwell static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb, 2001f9bfe4e6SEric Dumazet bool *is_cwnd_limited, 2002f9bfe4e6SEric Dumazet bool *is_rwnd_limited, 2003f9bfe4e6SEric Dumazet u32 max_segs) 2004c1b4a7e6SDavid S. Miller { 20056687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 2006f1c6ea38SEric Dumazet u32 send_win, cong_win, limit, in_flight; 200750c8339eSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 200850c8339eSEric Dumazet struct sk_buff *head; 2009ad9f4f50SEric Dumazet int win_divisor; 2010f1c6ea38SEric Dumazet s64 delta; 2011c1b4a7e6SDavid S. Miller 201299d7662aSEric Dumazet if (icsk->icsk_ca_state >= TCP_CA_Recovery) 2013ae8064acSJohn Heffner goto send_now; 2014ae8064acSJohn Heffner 20155f852eb5SEric Dumazet /* Avoid bursty behavior by allowing defer 2016a682850aSEric Dumazet * only if the last write was recent (1 ms). 2017a682850aSEric Dumazet * Note that tp->tcp_wstamp_ns can be in the future if we have 2018a682850aSEric Dumazet * packets waiting in a qdisc or device for EDT delivery. 20195f852eb5SEric Dumazet */ 2020a682850aSEric Dumazet delta = tp->tcp_clock_cache - tp->tcp_wstamp_ns - NSEC_PER_MSEC; 2021a682850aSEric Dumazet if (delta > 0) 2022ae8064acSJohn Heffner goto send_now; 2023908a75c1SDavid S. Miller 2024c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 2025c1b4a7e6SDavid S. Miller 2026c8c9aeb5SStefano Brivio BUG_ON(tcp_skb_pcount(skb) <= 1); 2027c8c9aeb5SStefano Brivio BUG_ON(tp->snd_cwnd <= in_flight); 2028c1b4a7e6SDavid S. Miller 202990840defSIlpo Järvinen send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 2030c1b4a7e6SDavid S. Miller 2031c1b4a7e6SDavid S. Miller /* From in_flight test above, we know that cwnd > in_flight. */ 2032c1b4a7e6SDavid S. Miller cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; 2033c1b4a7e6SDavid S. Miller 2034c1b4a7e6SDavid S. Miller limit = min(send_win, cong_win); 2035c1b4a7e6SDavid S. Miller 2036ba244fe9SDavid S. Miller /* If a full-sized TSO skb can be sent, do it. */ 2037605ad7f1SEric Dumazet if (limit >= max_segs * tp->mss_cache) 2038ae8064acSJohn Heffner goto send_now; 2039ba244fe9SDavid S. Miller 204062ad2761SIlpo Järvinen /* Middle in queue won't get any more data, full sendable already? */ 204162ad2761SIlpo Järvinen if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) 204262ad2761SIlpo Järvinen goto send_now; 204362ad2761SIlpo Järvinen 20445bbcc0f5SLinus Torvalds win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor); 2045ad9f4f50SEric Dumazet if (win_divisor) { 2046c1b4a7e6SDavid S. Miller u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); 2047c1b4a7e6SDavid S. Miller 2048c1b4a7e6SDavid S. Miller /* If at least some fraction of a window is available, 2049c1b4a7e6SDavid S. Miller * just use it. 2050c1b4a7e6SDavid S. Miller */ 2051ad9f4f50SEric Dumazet chunk /= win_divisor; 2052c1b4a7e6SDavid S. Miller if (limit >= chunk) 2053ae8064acSJohn Heffner goto send_now; 2054c1b4a7e6SDavid S. Miller } else { 2055c1b4a7e6SDavid S. Miller /* Different approach, try not to defer past a single 2056c1b4a7e6SDavid S. Miller * ACK. Receiver should ACK every other full sized 2057c1b4a7e6SDavid S. Miller * frame, so if we have space for more than 3 frames 2058c1b4a7e6SDavid S. Miller * then send now. 2059c1b4a7e6SDavid S. Miller */ 20606b5a5c0dSNeal Cardwell if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) 2061ae8064acSJohn Heffner goto send_now; 2062c1b4a7e6SDavid S. Miller } 2063c1b4a7e6SDavid S. Miller 206475c119afSEric Dumazet /* TODO : use tsorted_sent_queue ? */ 206575c119afSEric Dumazet head = tcp_rtx_queue_head(sk); 206675c119afSEric Dumazet if (!head) 206775c119afSEric Dumazet goto send_now; 2068f1c6ea38SEric Dumazet delta = tp->tcp_clock_cache - head->tstamp; 206950c8339eSEric Dumazet /* If next ACK is likely to come too late (half srtt), do not defer */ 2070f1c6ea38SEric Dumazet if ((s64)(delta - (u64)NSEC_PER_USEC * (tp->srtt_us >> 4)) < 0) 207150c8339eSEric Dumazet goto send_now; 207250c8339eSEric Dumazet 2073f9bfe4e6SEric Dumazet /* Ok, it looks like it is advisable to defer. 2074f9bfe4e6SEric Dumazet * Three cases are tracked : 2075f9bfe4e6SEric Dumazet * 1) We are cwnd-limited 2076f9bfe4e6SEric Dumazet * 2) We are rwnd-limited 2077f9bfe4e6SEric Dumazet * 3) We are application limited. 2078f9bfe4e6SEric Dumazet */ 2079f9bfe4e6SEric Dumazet if (cong_win < send_win) { 2080f9bfe4e6SEric Dumazet if (cong_win <= skb->len) { 2081ca8a2263SNeal Cardwell *is_cwnd_limited = true; 2082f9bfe4e6SEric Dumazet return true; 2083f9bfe4e6SEric Dumazet } 2084f9bfe4e6SEric Dumazet } else { 2085f9bfe4e6SEric Dumazet if (send_win <= skb->len) { 2086f9bfe4e6SEric Dumazet *is_rwnd_limited = true; 2087f9bfe4e6SEric Dumazet return true; 2088f9bfe4e6SEric Dumazet } 2089f9bfe4e6SEric Dumazet } 2090f9bfe4e6SEric Dumazet 2091f9bfe4e6SEric Dumazet /* If this packet won't get more data, do not wait. */ 2092d8ed257fSEric Dumazet if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) || 2093d8ed257fSEric Dumazet TCP_SKB_CB(skb)->eor) 2094f9bfe4e6SEric Dumazet goto send_now; 2095ca8a2263SNeal Cardwell 2096a2a385d6SEric Dumazet return true; 2097ae8064acSJohn Heffner 2098ae8064acSJohn Heffner send_now: 2099a2a385d6SEric Dumazet return false; 2100c1b4a7e6SDavid S. Miller } 2101c1b4a7e6SDavid S. Miller 210205cbc0dbSFan Du static inline void tcp_mtu_check_reprobe(struct sock *sk) 210305cbc0dbSFan Du { 210405cbc0dbSFan Du struct inet_connection_sock *icsk = inet_csk(sk); 210505cbc0dbSFan Du struct tcp_sock *tp = tcp_sk(sk); 210605cbc0dbSFan Du struct net *net = sock_net(sk); 210705cbc0dbSFan Du u32 interval; 210805cbc0dbSFan Du s32 delta; 210905cbc0dbSFan Du 211005cbc0dbSFan Du interval = net->ipv4.sysctl_tcp_probe_interval; 2111c74df29aSEric Dumazet delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp; 211205cbc0dbSFan Du if (unlikely(delta >= interval * HZ)) { 211305cbc0dbSFan Du int mss = tcp_current_mss(sk); 211405cbc0dbSFan Du 211505cbc0dbSFan Du /* Update current search range */ 211605cbc0dbSFan Du icsk->icsk_mtup.probe_size = 0; 211705cbc0dbSFan Du icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + 211805cbc0dbSFan Du sizeof(struct tcphdr) + 211905cbc0dbSFan Du icsk->icsk_af_ops->net_header_len; 212005cbc0dbSFan Du icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss); 212105cbc0dbSFan Du 212205cbc0dbSFan Du /* Update probe time stamp */ 2123c74df29aSEric Dumazet icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; 212405cbc0dbSFan Du } 212505cbc0dbSFan Du } 212605cbc0dbSFan Du 2127808cf9e3SIlya Lesokhin static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len) 2128808cf9e3SIlya Lesokhin { 2129808cf9e3SIlya Lesokhin struct sk_buff *skb, *next; 2130808cf9e3SIlya Lesokhin 2131808cf9e3SIlya Lesokhin skb = tcp_send_head(sk); 2132808cf9e3SIlya Lesokhin tcp_for_write_queue_from_safe(skb, next, sk) { 2133808cf9e3SIlya Lesokhin if (len <= skb->len) 2134808cf9e3SIlya Lesokhin break; 2135808cf9e3SIlya Lesokhin 2136888a5c53SWillem de Bruijn if (unlikely(TCP_SKB_CB(skb)->eor) || tcp_has_tx_tstamp(skb)) 2137808cf9e3SIlya Lesokhin return false; 2138808cf9e3SIlya Lesokhin 2139808cf9e3SIlya Lesokhin len -= skb->len; 2140808cf9e3SIlya Lesokhin } 2141808cf9e3SIlya Lesokhin 2142808cf9e3SIlya Lesokhin return true; 2143808cf9e3SIlya Lesokhin } 2144808cf9e3SIlya Lesokhin 21455d424d5aSJohn Heffner /* Create a new MTU probe if we are ready. 214667edfef7SAndi Kleen * MTU probe is regularly attempting to increase the path MTU by 214767edfef7SAndi Kleen * deliberately sending larger packets. This discovers routing 214867edfef7SAndi Kleen * changes resulting in larger path MTUs. 214967edfef7SAndi Kleen * 21505d424d5aSJohn Heffner * Returns 0 if we should wait to probe (no cwnd available), 21515d424d5aSJohn Heffner * 1 if a probe was sent, 2152056834d9SIlpo Järvinen * -1 otherwise 2153056834d9SIlpo Järvinen */ 21545d424d5aSJohn Heffner static int tcp_mtu_probe(struct sock *sk) 21555d424d5aSJohn Heffner { 21565d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 215712a59abcSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 21585d424d5aSJohn Heffner struct sk_buff *skb, *nskb, *next; 21596b58e0a5SFan Du struct net *net = sock_net(sk); 21605d424d5aSJohn Heffner int probe_size; 216191cc17c0SIlpo Järvinen int size_needed; 216212a59abcSEric Dumazet int copy, len; 21635d424d5aSJohn Heffner int mss_now; 21646b58e0a5SFan Du int interval; 21655d424d5aSJohn Heffner 21665d424d5aSJohn Heffner /* Not currently probing/verifying, 21675d424d5aSJohn Heffner * not in recovery, 21685d424d5aSJohn Heffner * have enough cwnd, and 216912a59abcSEric Dumazet * not SACKing (the variable headers throw things off) 217012a59abcSEric Dumazet */ 217112a59abcSEric Dumazet if (likely(!icsk->icsk_mtup.enabled || 21725d424d5aSJohn Heffner icsk->icsk_mtup.probe_size || 21735d424d5aSJohn Heffner inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 21745d424d5aSJohn Heffner tp->snd_cwnd < 11 || 217512a59abcSEric Dumazet tp->rx_opt.num_sacks || tp->rx_opt.dsack)) 21765d424d5aSJohn Heffner return -1; 21775d424d5aSJohn Heffner 21786b58e0a5SFan Du /* Use binary search for probe_size between tcp_mss_base, 21796b58e0a5SFan Du * and current mss_clamp. if (search_high - search_low) 21806b58e0a5SFan Du * smaller than a threshold, backoff from probing. 21816b58e0a5SFan Du */ 21820c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 21836b58e0a5SFan Du probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high + 21846b58e0a5SFan Du icsk->icsk_mtup.search_low) >> 1); 218591cc17c0SIlpo Järvinen size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; 21866b58e0a5SFan Du interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low; 218705cbc0dbSFan Du /* When misfortune happens, we are reprobing actively, 218805cbc0dbSFan Du * and then reprobe timer has expired. We stick with current 218905cbc0dbSFan Du * probing process by not resetting search range to its orignal. 219005cbc0dbSFan Du */ 21916b58e0a5SFan Du if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) || 219205cbc0dbSFan Du interval < net->ipv4.sysctl_tcp_probe_threshold) { 219305cbc0dbSFan Du /* Check whether enough time has elaplased for 219405cbc0dbSFan Du * another round of probing. 219505cbc0dbSFan Du */ 219605cbc0dbSFan Du tcp_mtu_check_reprobe(sk); 21975d424d5aSJohn Heffner return -1; 21985d424d5aSJohn Heffner } 21995d424d5aSJohn Heffner 22005d424d5aSJohn Heffner /* Have enough data in the send queue to probe? */ 22017f9c33e5SIlpo Järvinen if (tp->write_seq - tp->snd_nxt < size_needed) 22025d424d5aSJohn Heffner return -1; 22035d424d5aSJohn Heffner 220491cc17c0SIlpo Järvinen if (tp->snd_wnd < size_needed) 22055d424d5aSJohn Heffner return -1; 220690840defSIlpo Järvinen if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) 22075d424d5aSJohn Heffner return 0; 22085d424d5aSJohn Heffner 2209d67c58e9SIlpo Järvinen /* Do we need to wait to drain cwnd? With none in flight, don't stall */ 2210d67c58e9SIlpo Järvinen if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { 2211d67c58e9SIlpo Järvinen if (!tcp_packets_in_flight(tp)) 22125d424d5aSJohn Heffner return -1; 22135d424d5aSJohn Heffner else 22145d424d5aSJohn Heffner return 0; 22155d424d5aSJohn Heffner } 22165d424d5aSJohn Heffner 2217808cf9e3SIlya Lesokhin if (!tcp_can_coalesce_send_queue_head(sk, probe_size)) 2218808cf9e3SIlya Lesokhin return -1; 2219808cf9e3SIlya Lesokhin 22205d424d5aSJohn Heffner /* We're allowed to probe. Build it now. */ 2221eb934478SEric Dumazet nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false); 222251456b29SIan Morris if (!nskb) 22235d424d5aSJohn Heffner return -1; 2224ab4e846aSEric Dumazet sk_wmem_queued_add(sk, nskb->truesize); 22253ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 22265d424d5aSJohn Heffner 2227fe067e8aSDavid S. Miller skb = tcp_send_head(sk); 222841477662SJakub Kicinski skb_copy_decrypted(nskb, skb); 22295d424d5aSJohn Heffner 22305d424d5aSJohn Heffner TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 22315d424d5aSJohn Heffner TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 22324de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; 22335d424d5aSJohn Heffner TCP_SKB_CB(nskb)->sacked = 0; 22345d424d5aSJohn Heffner nskb->csum = 0; 223598be9b12SEric Dumazet nskb->ip_summed = CHECKSUM_PARTIAL; 22365d424d5aSJohn Heffner 223750c4817eSIlpo Järvinen tcp_insert_write_queue_before(nskb, skb, sk); 22382b7cda9cSEric Dumazet tcp_highest_sack_replace(sk, skb, nskb); 223950c4817eSIlpo Järvinen 22405d424d5aSJohn Heffner len = 0; 2241234b6860SIlpo Järvinen tcp_for_write_queue_from_safe(skb, next, sk) { 22425d424d5aSJohn Heffner copy = min_t(int, skb->len, probe_size - len); 22435d424d5aSJohn Heffner skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); 22445d424d5aSJohn Heffner 22455d424d5aSJohn Heffner if (skb->len <= copy) { 22465d424d5aSJohn Heffner /* We've eaten all the data from this skb. 22475d424d5aSJohn Heffner * Throw it away. */ 22484de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 2249808cf9e3SIlya Lesokhin /* If this is the last SKB we copy and eor is set 2250808cf9e3SIlya Lesokhin * we need to propagate it to the new skb. 2251808cf9e3SIlya Lesokhin */ 2252808cf9e3SIlya Lesokhin TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor; 2253888a5c53SWillem de Bruijn tcp_skb_collapse_tstamp(nskb, skb); 2254fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 22553ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 22565d424d5aSJohn Heffner } else { 22574de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & 2258a3433f35SChangli Gao ~(TCPHDR_FIN|TCPHDR_PSH); 22595d424d5aSJohn Heffner if (!skb_shinfo(skb)->nr_frags) { 22605d424d5aSJohn Heffner skb_pull(skb, copy); 22615d424d5aSJohn Heffner } else { 22625d424d5aSJohn Heffner __pskb_trim_head(skb, copy); 22635bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 22645d424d5aSJohn Heffner } 22655d424d5aSJohn Heffner TCP_SKB_CB(skb)->seq += copy; 22665d424d5aSJohn Heffner } 22675d424d5aSJohn Heffner 22685d424d5aSJohn Heffner len += copy; 2269234b6860SIlpo Järvinen 2270234b6860SIlpo Järvinen if (len >= probe_size) 2271234b6860SIlpo Järvinen break; 22725d424d5aSJohn Heffner } 22735bbb432cSEric Dumazet tcp_init_tso_segs(nskb, nskb->len); 22745d424d5aSJohn Heffner 22755d424d5aSJohn Heffner /* We're ready to send. If this fails, the probe will 22767faee5c0SEric Dumazet * be resegmented into mss-sized pieces by tcp_write_xmit(). 22777faee5c0SEric Dumazet */ 22785d424d5aSJohn Heffner if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 22795d424d5aSJohn Heffner /* Decrement cwnd here because we are sending 22805d424d5aSJohn Heffner * effectively two packets. */ 22815d424d5aSJohn Heffner tp->snd_cwnd--; 228266f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, nskb); 22835d424d5aSJohn Heffner 22845d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 22850e7b1368SJohn Heffner tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 22860e7b1368SJohn Heffner tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 22875d424d5aSJohn Heffner 22885d424d5aSJohn Heffner return 1; 22895d424d5aSJohn Heffner } 22905d424d5aSJohn Heffner 22915d424d5aSJohn Heffner return -1; 22925d424d5aSJohn Heffner } 22935d424d5aSJohn Heffner 2294864e5c09SEric Dumazet static bool tcp_pacing_check(struct sock *sk) 2295218af599SEric Dumazet { 2296864e5c09SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 2297864e5c09SEric Dumazet 2298864e5c09SEric Dumazet if (!tcp_needs_internal_pacing(sk)) 2299864e5c09SEric Dumazet return false; 2300864e5c09SEric Dumazet 2301864e5c09SEric Dumazet if (tp->tcp_wstamp_ns <= tp->tcp_clock_cache) 2302864e5c09SEric Dumazet return false; 2303864e5c09SEric Dumazet 2304864e5c09SEric Dumazet if (!hrtimer_is_queued(&tp->pacing_timer)) { 2305864e5c09SEric Dumazet hrtimer_start(&tp->pacing_timer, 2306864e5c09SEric Dumazet ns_to_ktime(tp->tcp_wstamp_ns), 2307864e5c09SEric Dumazet HRTIMER_MODE_ABS_PINNED_SOFT); 2308864e5c09SEric Dumazet sock_hold(sk); 2309864e5c09SEric Dumazet } 2310864e5c09SEric Dumazet return true; 2311218af599SEric Dumazet } 2312218af599SEric Dumazet 2313f9616c35SEric Dumazet /* TCP Small Queues : 2314f9616c35SEric Dumazet * Control number of packets in qdisc/devices to two packets / or ~1 ms. 2315f9616c35SEric Dumazet * (These limits are doubled for retransmits) 2316f9616c35SEric Dumazet * This allows for : 2317f9616c35SEric Dumazet * - better RTT estimation and ACK scheduling 2318f9616c35SEric Dumazet * - faster recovery 2319f9616c35SEric Dumazet * - high rates 2320f9616c35SEric Dumazet * Alas, some drivers / subsystems require a fair amount 2321f9616c35SEric Dumazet * of queued bytes to ensure line rate. 2322f9616c35SEric Dumazet * One example is wifi aggregation (802.11 AMPDU) 2323f9616c35SEric Dumazet */ 2324f9616c35SEric Dumazet static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb, 2325f9616c35SEric Dumazet unsigned int factor) 2326f9616c35SEric Dumazet { 232776a9ebe8SEric Dumazet unsigned long limit; 2328f9616c35SEric Dumazet 232976a9ebe8SEric Dumazet limit = max_t(unsigned long, 233076a9ebe8SEric Dumazet 2 * skb->truesize, 23317c68fa2bSEric Dumazet sk->sk_pacing_rate >> READ_ONCE(sk->sk_pacing_shift)); 2332c73e5807SEric Dumazet if (sk->sk_pacing_status == SK_PACING_NONE) 233376a9ebe8SEric Dumazet limit = min_t(unsigned long, limit, 23349184d8bbSEric Dumazet sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes); 2335f9616c35SEric Dumazet limit <<= factor; 2336f9616c35SEric Dumazet 2337a842fe14SEric Dumazet if (static_branch_unlikely(&tcp_tx_delay_enabled) && 2338a842fe14SEric Dumazet tcp_sk(sk)->tcp_tx_delay) { 2339a842fe14SEric Dumazet u64 extra_bytes = (u64)sk->sk_pacing_rate * tcp_sk(sk)->tcp_tx_delay; 2340a842fe14SEric Dumazet 2341a842fe14SEric Dumazet /* TSQ is based on skb truesize sum (sk_wmem_alloc), so we 2342a842fe14SEric Dumazet * approximate our needs assuming an ~100% skb->truesize overhead. 2343a842fe14SEric Dumazet * USEC_PER_SEC is approximated by 2^20. 2344a842fe14SEric Dumazet * do_div(extra_bytes, USEC_PER_SEC/2) is replaced by a right shift. 2345a842fe14SEric Dumazet */ 2346a842fe14SEric Dumazet extra_bytes >>= (20 - 1); 2347a842fe14SEric Dumazet limit += extra_bytes; 2348a842fe14SEric Dumazet } 234914afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) > limit) { 235075c119afSEric Dumazet /* Always send skb if rtx queue is empty. 235175eefc6cSEric Dumazet * No need to wait for TX completion to call us back, 235275eefc6cSEric Dumazet * after softirq/tasklet schedule. 235375eefc6cSEric Dumazet * This helps when TX completions are delayed too much. 235475eefc6cSEric Dumazet */ 235575c119afSEric Dumazet if (tcp_rtx_queue_empty(sk)) 235675eefc6cSEric Dumazet return false; 235775eefc6cSEric Dumazet 23587aa5470cSEric Dumazet set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags); 2359f9616c35SEric Dumazet /* It is possible TX completion already happened 2360f9616c35SEric Dumazet * before we set TSQ_THROTTLED, so we must 2361f9616c35SEric Dumazet * test again the condition. 2362f9616c35SEric Dumazet */ 2363f9616c35SEric Dumazet smp_mb__after_atomic(); 236414afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) > limit) 2365f9616c35SEric Dumazet return true; 2366f9616c35SEric Dumazet } 2367f9616c35SEric Dumazet return false; 2368f9616c35SEric Dumazet } 2369f9616c35SEric Dumazet 237005b055e8SFrancis Yan static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new) 237105b055e8SFrancis Yan { 2372628174ccSEric Dumazet const u32 now = tcp_jiffies32; 2373efe967cdSArnd Bergmann enum tcp_chrono old = tp->chrono_type; 237405b055e8SFrancis Yan 2375efe967cdSArnd Bergmann if (old > TCP_CHRONO_UNSPEC) 2376efe967cdSArnd Bergmann tp->chrono_stat[old - 1] += now - tp->chrono_start; 237705b055e8SFrancis Yan tp->chrono_start = now; 237805b055e8SFrancis Yan tp->chrono_type = new; 237905b055e8SFrancis Yan } 238005b055e8SFrancis Yan 238105b055e8SFrancis Yan void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type) 238205b055e8SFrancis Yan { 238305b055e8SFrancis Yan struct tcp_sock *tp = tcp_sk(sk); 238405b055e8SFrancis Yan 238505b055e8SFrancis Yan /* If there are multiple conditions worthy of tracking in a 23860f87230dSFrancis Yan * chronograph then the highest priority enum takes precedence 23870f87230dSFrancis Yan * over the other conditions. So that if something "more interesting" 238805b055e8SFrancis Yan * starts happening, stop the previous chrono and start a new one. 238905b055e8SFrancis Yan */ 239005b055e8SFrancis Yan if (type > tp->chrono_type) 239105b055e8SFrancis Yan tcp_chrono_set(tp, type); 239205b055e8SFrancis Yan } 239305b055e8SFrancis Yan 239405b055e8SFrancis Yan void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type) 239505b055e8SFrancis Yan { 239605b055e8SFrancis Yan struct tcp_sock *tp = tcp_sk(sk); 239705b055e8SFrancis Yan 23980f87230dSFrancis Yan 23990f87230dSFrancis Yan /* There are multiple conditions worthy of tracking in a 24000f87230dSFrancis Yan * chronograph, so that the highest priority enum takes 24010f87230dSFrancis Yan * precedence over the other conditions (see tcp_chrono_start). 24020f87230dSFrancis Yan * If a condition stops, we only stop chrono tracking if 24030f87230dSFrancis Yan * it's the "most interesting" or current chrono we are 24040f87230dSFrancis Yan * tracking and starts busy chrono if we have pending data. 24050f87230dSFrancis Yan */ 240675c119afSEric Dumazet if (tcp_rtx_and_write_queues_empty(sk)) 240705b055e8SFrancis Yan tcp_chrono_set(tp, TCP_CHRONO_UNSPEC); 24080f87230dSFrancis Yan else if (type == tp->chrono_type) 24090f87230dSFrancis Yan tcp_chrono_set(tp, TCP_CHRONO_BUSY); 241005b055e8SFrancis Yan } 241105b055e8SFrancis Yan 24121da177e4SLinus Torvalds /* This routine writes packets to the network. It advances the 24131da177e4SLinus Torvalds * send_head. This happens as incoming acks open up the remote 24141da177e4SLinus Torvalds * window for us. 24151da177e4SLinus Torvalds * 2416f8269a49SIlpo Järvinen * LARGESEND note: !tcp_urg_mode is overkill, only frames between 2417f8269a49SIlpo Järvinen * snd_up-64k-mss .. snd_up cannot be large. However, taking into 2418f8269a49SIlpo Järvinen * account rare use of URG, this is not a big flaw. 2419f8269a49SIlpo Järvinen * 24206ba8a3b1SNandita Dukkipati * Send at most one packet when push_one > 0. Temporarily ignore 24216ba8a3b1SNandita Dukkipati * cwnd limit to force at most one packet out when push_one == 2. 24226ba8a3b1SNandita Dukkipati 2423a2a385d6SEric Dumazet * Returns true, if no segments are in flight and we have queued segments, 2424a2a385d6SEric Dumazet * but cannot send anything now because of SWS or another problem. 24251da177e4SLinus Torvalds */ 2426a2a385d6SEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 2427d5dd9175SIlpo Järvinen int push_one, gfp_t gfp) 24281da177e4SLinus Torvalds { 24291da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 243092df7b51SDavid S. Miller struct sk_buff *skb; 2431c1b4a7e6SDavid S. Miller unsigned int tso_segs, sent_pkts; 2432c1b4a7e6SDavid S. Miller int cwnd_quota; 24335d424d5aSJohn Heffner int result; 24345615f886SFrancis Yan bool is_cwnd_limited = false, is_rwnd_limited = false; 2435605ad7f1SEric Dumazet u32 max_segs; 24361da177e4SLinus Torvalds 2437c1b4a7e6SDavid S. Miller sent_pkts = 0; 24385d424d5aSJohn Heffner 2439ee1836aeSEric Dumazet tcp_mstamp_refresh(tp); 2440d5dd9175SIlpo Järvinen if (!push_one) { 24415d424d5aSJohn Heffner /* Do MTU probing. */ 2442d5dd9175SIlpo Järvinen result = tcp_mtu_probe(sk); 2443d5dd9175SIlpo Järvinen if (!result) { 2444a2a385d6SEric Dumazet return false; 24455d424d5aSJohn Heffner } else if (result > 0) { 24465d424d5aSJohn Heffner sent_pkts = 1; 24475d424d5aSJohn Heffner } 2448d5dd9175SIlpo Järvinen } 24495d424d5aSJohn Heffner 2450ed6e7268SNeal Cardwell max_segs = tcp_tso_segs(sk, mss_now); 2451fe067e8aSDavid S. Miller while ((skb = tcp_send_head(sk))) { 2452c8ac3774SHerbert Xu unsigned int limit; 2453c8ac3774SHerbert Xu 245479861919SEric Dumazet if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) { 245579861919SEric Dumazet /* "skb_mstamp_ns" is used as a start point for the retransmit timer */ 245679861919SEric Dumazet skb->skb_mstamp_ns = tp->tcp_wstamp_ns = tp->tcp_clock_cache; 245779861919SEric Dumazet list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue); 2458bf50b606SEric Dumazet tcp_init_tso_segs(skb, mss_now); 245979861919SEric Dumazet goto repair; /* Skip network transmission */ 246079861919SEric Dumazet } 246179861919SEric Dumazet 2462218af599SEric Dumazet if (tcp_pacing_check(sk)) 2463218af599SEric Dumazet break; 2464218af599SEric Dumazet 24655bbb432cSEric Dumazet tso_segs = tcp_init_tso_segs(skb, mss_now); 2466c1b4a7e6SDavid S. Miller BUG_ON(!tso_segs); 2467c1b4a7e6SDavid S. Miller 2468b68e9f85SHerbert Xu cwnd_quota = tcp_cwnd_test(tp, skb); 24696ba8a3b1SNandita Dukkipati if (!cwnd_quota) { 24706ba8a3b1SNandita Dukkipati if (push_one == 2) 24716ba8a3b1SNandita Dukkipati /* Force out a loss probe pkt. */ 24726ba8a3b1SNandita Dukkipati cwnd_quota = 1; 24736ba8a3b1SNandita Dukkipati else 2474b68e9f85SHerbert Xu break; 24756ba8a3b1SNandita Dukkipati } 2476b68e9f85SHerbert Xu 24775615f886SFrancis Yan if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) { 24785615f886SFrancis Yan is_rwnd_limited = true; 2479b68e9f85SHerbert Xu break; 24805615f886SFrancis Yan } 2481b68e9f85SHerbert Xu 2482d6a4e26aSEric Dumazet if (tso_segs == 1) { 2483aa93466bSDavid S. Miller if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 2484aa93466bSDavid S. Miller (tcp_skb_is_last(sk, skb) ? 2485aa93466bSDavid S. Miller nonagle : TCP_NAGLE_PUSH)))) 2486aa93466bSDavid S. Miller break; 2487c1b4a7e6SDavid S. Miller } else { 2488ca8a2263SNeal Cardwell if (!push_one && 2489605ad7f1SEric Dumazet tcp_tso_should_defer(sk, skb, &is_cwnd_limited, 2490f9bfe4e6SEric Dumazet &is_rwnd_limited, max_segs)) 2491aa93466bSDavid S. Miller break; 2492c1b4a7e6SDavid S. Miller } 2493aa93466bSDavid S. Miller 2494605ad7f1SEric Dumazet limit = mss_now; 2495d6a4e26aSEric Dumazet if (tso_segs > 1 && !tcp_urg_mode(tp)) 2496605ad7f1SEric Dumazet limit = tcp_mss_split_point(sk, skb, mss_now, 2497605ad7f1SEric Dumazet min_t(unsigned int, 2498605ad7f1SEric Dumazet cwnd_quota, 2499605ad7f1SEric Dumazet max_segs), 2500605ad7f1SEric Dumazet nonagle); 2501605ad7f1SEric Dumazet 2502605ad7f1SEric Dumazet if (skb->len > limit && 250356483341SEric Dumazet unlikely(tso_fragment(sk, skb, limit, mss_now, gfp))) 2504605ad7f1SEric Dumazet break; 2505605ad7f1SEric Dumazet 2506f9616c35SEric Dumazet if (tcp_small_queue_check(sk, skb, 0)) 250746d3ceabSEric Dumazet break; 2508c9eeec26SEric Dumazet 25091f85e626SEric Dumazet /* Argh, we hit an empty skb(), presumably a thread 25101f85e626SEric Dumazet * is sleeping in sendmsg()/sk_stream_wait_memory(). 25111f85e626SEric Dumazet * We do not want to send a pure-ack packet and have 25121f85e626SEric Dumazet * a strange looking rtx queue with empty packet(s). 25131f85e626SEric Dumazet */ 25141f85e626SEric Dumazet if (TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq) 25151f85e626SEric Dumazet break; 25161f85e626SEric Dumazet 2517d5dd9175SIlpo Järvinen if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) 25181da177e4SLinus Torvalds break; 25191da177e4SLinus Torvalds 2520ec342325SAndrew Vagin repair: 25211da177e4SLinus Torvalds /* Advance the send_head. This one is sent out. 25221da177e4SLinus Torvalds * This call will increment packets_out. 25231da177e4SLinus Torvalds */ 252466f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 25251da177e4SLinus Torvalds 25261da177e4SLinus Torvalds tcp_minshall_update(tp, mss_now, skb); 2527a262f0cdSNandita Dukkipati sent_pkts += tcp_skb_pcount(skb); 2528d5dd9175SIlpo Järvinen 2529d5dd9175SIlpo Järvinen if (push_one) 2530d5dd9175SIlpo Järvinen break; 25311da177e4SLinus Torvalds } 25321da177e4SLinus Torvalds 25335615f886SFrancis Yan if (is_rwnd_limited) 25345615f886SFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED); 25355615f886SFrancis Yan else 25365615f886SFrancis Yan tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED); 25375615f886SFrancis Yan 2538aa93466bSDavid S. Miller if (likely(sent_pkts)) { 2539684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 2540684bad11SYuchung Cheng tp->prr_out += sent_pkts; 25416ba8a3b1SNandita Dukkipati 25426ba8a3b1SNandita Dukkipati /* Send one loss probe per tail loss episode. */ 25436ba8a3b1SNandita Dukkipati if (push_one != 2) 2544ed66dfafSNeal Cardwell tcp_schedule_loss_probe(sk, false); 2545d2e1339fSBendik Rønning Opstad is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd); 2546ca8a2263SNeal Cardwell tcp_cwnd_validate(sk, is_cwnd_limited); 2547a2a385d6SEric Dumazet return false; 25481da177e4SLinus Torvalds } 254975c119afSEric Dumazet return !tp->packets_out && !tcp_write_queue_empty(sk); 25506ba8a3b1SNandita Dukkipati } 25516ba8a3b1SNandita Dukkipati 2552ed66dfafSNeal Cardwell bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto) 25536ba8a3b1SNandita Dukkipati { 25546ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 25556ba8a3b1SNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 2556a2815817SNeal Cardwell u32 timeout, rto_delta_us; 25572ae21cf5SEric Dumazet int early_retrans; 25586ba8a3b1SNandita Dukkipati 25596ba8a3b1SNandita Dukkipati /* Don't do any loss probe on a Fast Open connection before 3WHS 25606ba8a3b1SNandita Dukkipati * finishes. 25616ba8a3b1SNandita Dukkipati */ 2562d983ea6fSEric Dumazet if (rcu_access_pointer(tp->fastopen_rsk)) 25636ba8a3b1SNandita Dukkipati return false; 25646ba8a3b1SNandita Dukkipati 25652ae21cf5SEric Dumazet early_retrans = sock_net(sk)->ipv4.sysctl_tcp_early_retrans; 25666ba8a3b1SNandita Dukkipati /* Schedule a loss probe in 2*RTT for SACK capable connections 2567b4f70c3dSNeal Cardwell * not in loss recovery, that are either limited by cwnd or application. 25686ba8a3b1SNandita Dukkipati */ 25692ae21cf5SEric Dumazet if ((early_retrans != 3 && early_retrans != 4) || 2570bec41a11SYuchung Cheng !tp->packets_out || !tcp_is_sack(tp) || 2571b4f70c3dSNeal Cardwell (icsk->icsk_ca_state != TCP_CA_Open && 2572b4f70c3dSNeal Cardwell icsk->icsk_ca_state != TCP_CA_CWR)) 25736ba8a3b1SNandita Dukkipati return false; 25746ba8a3b1SNandita Dukkipati 2575bb4d991aSYuchung Cheng /* Probe timeout is 2*rtt. Add minimum RTO to account 2576f9b99582SYuchung Cheng * for delayed ack when there's one outstanding packet. If no RTT 2577f9b99582SYuchung Cheng * sample is available then probe after TCP_TIMEOUT_INIT. 25786ba8a3b1SNandita Dukkipati */ 2579bb4d991aSYuchung Cheng if (tp->srtt_us) { 2580bb4d991aSYuchung Cheng timeout = usecs_to_jiffies(tp->srtt_us >> 2); 25816ba8a3b1SNandita Dukkipati if (tp->packets_out == 1) 2582bb4d991aSYuchung Cheng timeout += TCP_RTO_MIN; 2583bb4d991aSYuchung Cheng else 2584bb4d991aSYuchung Cheng timeout += TCP_TIMEOUT_MIN; 2585bb4d991aSYuchung Cheng } else { 2586bb4d991aSYuchung Cheng timeout = TCP_TIMEOUT_INIT; 2587bb4d991aSYuchung Cheng } 25886ba8a3b1SNandita Dukkipati 2589a2815817SNeal Cardwell /* If the RTO formula yields an earlier time, then use that time. */ 2590ed66dfafSNeal Cardwell rto_delta_us = advancing_rto ? 2591ed66dfafSNeal Cardwell jiffies_to_usecs(inet_csk(sk)->icsk_rto) : 2592ed66dfafSNeal Cardwell tcp_rto_delta_us(sk); /* How far in future is RTO? */ 2593a2815817SNeal Cardwell if (rto_delta_us > 0) 2594a2815817SNeal Cardwell timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us)); 25956ba8a3b1SNandita Dukkipati 25968dc242adSEric Dumazet tcp_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, TCP_RTO_MAX); 25976ba8a3b1SNandita Dukkipati return true; 25986ba8a3b1SNandita Dukkipati } 25996ba8a3b1SNandita Dukkipati 26001f3279aeSEric Dumazet /* Thanks to skb fast clones, we can detect if a prior transmit of 26011f3279aeSEric Dumazet * a packet is still in a qdisc or driver queue. 26021f3279aeSEric Dumazet * In this case, there is very little point doing a retransmit ! 26031f3279aeSEric Dumazet */ 26041f3279aeSEric Dumazet static bool skb_still_in_host_queue(const struct sock *sk, 26051f3279aeSEric Dumazet const struct sk_buff *skb) 26061f3279aeSEric Dumazet { 260739bb5e62SEric Dumazet if (unlikely(skb_fclone_busy(sk, skb))) { 2608c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), 26091f3279aeSEric Dumazet LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES); 26101f3279aeSEric Dumazet return true; 26111f3279aeSEric Dumazet } 26121f3279aeSEric Dumazet return false; 26131f3279aeSEric Dumazet } 26141f3279aeSEric Dumazet 2615b340b264SYuchung Cheng /* When probe timeout (PTO) fires, try send a new segment if possible, else 26166ba8a3b1SNandita Dukkipati * retransmit the last segment. 26176ba8a3b1SNandita Dukkipati */ 26186ba8a3b1SNandita Dukkipati void tcp_send_loss_probe(struct sock *sk) 26196ba8a3b1SNandita Dukkipati { 26209b717a8dSNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 26216ba8a3b1SNandita Dukkipati struct sk_buff *skb; 26226ba8a3b1SNandita Dukkipati int pcount; 26236ba8a3b1SNandita Dukkipati int mss = tcp_current_mss(sk); 26246ba8a3b1SNandita Dukkipati 2625b340b264SYuchung Cheng skb = tcp_send_head(sk); 262675c119afSEric Dumazet if (skb && tcp_snd_wnd_test(tp, skb, mss)) { 2627b340b264SYuchung Cheng pcount = tp->packets_out; 2628b340b264SYuchung Cheng tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC); 2629b340b264SYuchung Cheng if (tp->packets_out > pcount) 2630b340b264SYuchung Cheng goto probe_sent; 26316ba8a3b1SNandita Dukkipati goto rearm_timer; 26326ba8a3b1SNandita Dukkipati } 263375c119afSEric Dumazet skb = skb_rb_last(&sk->tcp_rtx_queue); 2634b2b7af86SYuchung Cheng if (unlikely(!skb)) { 2635b2b7af86SYuchung Cheng WARN_ONCE(tp->packets_out, 2636b2b7af86SYuchung Cheng "invalid inflight: %u state %u cwnd %u mss %d\n", 2637b2b7af86SYuchung Cheng tp->packets_out, sk->sk_state, tp->snd_cwnd, mss); 2638b2b7af86SYuchung Cheng inet_csk(sk)->icsk_pending = 0; 2639b2b7af86SYuchung Cheng return; 2640b2b7af86SYuchung Cheng } 26416ba8a3b1SNandita Dukkipati 26429b717a8dSNandita Dukkipati /* At most one outstanding TLP retransmission. */ 26439b717a8dSNandita Dukkipati if (tp->tlp_high_seq) 26449b717a8dSNandita Dukkipati goto rearm_timer; 26459b717a8dSNandita Dukkipati 26461f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 26471f3279aeSEric Dumazet goto rearm_timer; 26481f3279aeSEric Dumazet 26496ba8a3b1SNandita Dukkipati pcount = tcp_skb_pcount(skb); 26506ba8a3b1SNandita Dukkipati if (WARN_ON(!pcount)) 26516ba8a3b1SNandita Dukkipati goto rearm_timer; 26526ba8a3b1SNandita Dukkipati 26536ba8a3b1SNandita Dukkipati if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) { 265475c119afSEric Dumazet if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, 265575c119afSEric Dumazet (pcount - 1) * mss, mss, 26566cc55e09SOctavian Purdila GFP_ATOMIC))) 26576ba8a3b1SNandita Dukkipati goto rearm_timer; 265875c119afSEric Dumazet skb = skb_rb_next(skb); 26596ba8a3b1SNandita Dukkipati } 26606ba8a3b1SNandita Dukkipati 26616ba8a3b1SNandita Dukkipati if (WARN_ON(!skb || !tcp_skb_pcount(skb))) 26626ba8a3b1SNandita Dukkipati goto rearm_timer; 26636ba8a3b1SNandita Dukkipati 266410d3be56SEric Dumazet if (__tcp_retransmit_skb(sk, skb, 1)) 2665b340b264SYuchung Cheng goto rearm_timer; 26666ba8a3b1SNandita Dukkipati 26679b717a8dSNandita Dukkipati /* Record snd_nxt for loss detection. */ 26689b717a8dSNandita Dukkipati tp->tlp_high_seq = tp->snd_nxt; 26699b717a8dSNandita Dukkipati 2670b340b264SYuchung Cheng probe_sent: 2671c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES); 2672fcd16c0aSYuchung Cheng /* Reset s.t. tcp_rearm_rto will restart timer from now */ 2673fcd16c0aSYuchung Cheng inet_csk(sk)->icsk_pending = 0; 2674b340b264SYuchung Cheng rearm_timer: 2675fcd16c0aSYuchung Cheng tcp_rearm_rto(sk); 26761da177e4SLinus Torvalds } 26771da177e4SLinus Torvalds 2678a762a980SDavid S. Miller /* Push out any pending frames which were held back due to 2679a762a980SDavid S. Miller * TCP_CORK or attempt at coalescing tiny packets. 2680a762a980SDavid S. Miller * The socket must be locked by the caller. 2681a762a980SDavid S. Miller */ 26829e412ba7SIlpo Järvinen void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 26839e412ba7SIlpo Järvinen int nonagle) 2684a762a980SDavid S. Miller { 2685726e07a8SIlpo Järvinen /* If we are closed, the bytes will have to remain here. 2686726e07a8SIlpo Järvinen * In time closedown will finish, we empty the write queue and 2687726e07a8SIlpo Järvinen * all will be happy. 2688726e07a8SIlpo Järvinen */ 2689726e07a8SIlpo Järvinen if (unlikely(sk->sk_state == TCP_CLOSE)) 2690726e07a8SIlpo Järvinen return; 2691726e07a8SIlpo Järvinen 269299a1dec7SMel Gorman if (tcp_write_xmit(sk, cur_mss, nonagle, 0, 26937450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC))) 26949e412ba7SIlpo Järvinen tcp_check_probe_timer(sk); 2695a762a980SDavid S. Miller } 2696a762a980SDavid S. Miller 2697c1b4a7e6SDavid S. Miller /* Send _single_ skb sitting at the send head. This function requires 2698c1b4a7e6SDavid S. Miller * true push pending frames to setup probe timer etc. 2699c1b4a7e6SDavid S. Miller */ 2700c1b4a7e6SDavid S. Miller void tcp_push_one(struct sock *sk, unsigned int mss_now) 2701c1b4a7e6SDavid S. Miller { 2702fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 2703c1b4a7e6SDavid S. Miller 2704c1b4a7e6SDavid S. Miller BUG_ON(!skb || skb->len < mss_now); 2705c1b4a7e6SDavid S. Miller 2706d5dd9175SIlpo Järvinen tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); 2707c1b4a7e6SDavid S. Miller } 2708c1b4a7e6SDavid S. Miller 27091da177e4SLinus Torvalds /* This function returns the amount that we can raise the 27101da177e4SLinus Torvalds * usable window based on the following constraints 27111da177e4SLinus Torvalds * 27121da177e4SLinus Torvalds * 1. The window can never be shrunk once it is offered (RFC 793) 27131da177e4SLinus Torvalds * 2. We limit memory per socket 27141da177e4SLinus Torvalds * 27151da177e4SLinus Torvalds * RFC 1122: 27161da177e4SLinus Torvalds * "the suggested [SWS] avoidance algorithm for the receiver is to keep 27171da177e4SLinus Torvalds * RECV.NEXT + RCV.WIN fixed until: 27181da177e4SLinus Torvalds * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 27191da177e4SLinus Torvalds * 27201da177e4SLinus Torvalds * i.e. don't raise the right edge of the window until you can raise 27211da177e4SLinus Torvalds * it at least MSS bytes. 27221da177e4SLinus Torvalds * 27231da177e4SLinus Torvalds * Unfortunately, the recommended algorithm breaks header prediction, 27241da177e4SLinus Torvalds * since header prediction assumes th->window stays fixed. 27251da177e4SLinus Torvalds * 27261da177e4SLinus Torvalds * Strictly speaking, keeping th->window fixed violates the receiver 27271da177e4SLinus Torvalds * side SWS prevention criteria. The problem is that under this rule 27281da177e4SLinus Torvalds * a stream of single byte packets will cause the right side of the 27291da177e4SLinus Torvalds * window to always advance by a single byte. 27301da177e4SLinus Torvalds * 27311da177e4SLinus Torvalds * Of course, if the sender implements sender side SWS prevention 27321da177e4SLinus Torvalds * then this will not be a problem. 27331da177e4SLinus Torvalds * 27341da177e4SLinus Torvalds * BSD seems to make the following compromise: 27351da177e4SLinus Torvalds * 27361da177e4SLinus Torvalds * If the free space is less than the 1/4 of the maximum 27371da177e4SLinus Torvalds * space available and the free space is less than 1/2 mss, 27381da177e4SLinus Torvalds * then set the window to 0. 27391da177e4SLinus Torvalds * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 27401da177e4SLinus Torvalds * Otherwise, just prevent the window from shrinking 27411da177e4SLinus Torvalds * and from being larger than the largest representable value. 27421da177e4SLinus Torvalds * 27431da177e4SLinus Torvalds * This prevents incremental opening of the window in the regime 27441da177e4SLinus Torvalds * where TCP is limited by the speed of the reader side taking 27451da177e4SLinus Torvalds * data out of the TCP receive queue. It does nothing about 27461da177e4SLinus Torvalds * those cases where the window is constrained on the sender side 27471da177e4SLinus Torvalds * because the pipeline is full. 27481da177e4SLinus Torvalds * 27491da177e4SLinus Torvalds * BSD also seems to "accidentally" limit itself to windows that are a 27501da177e4SLinus Torvalds * multiple of MSS, at least until the free space gets quite small. 27511da177e4SLinus Torvalds * This would appear to be a side effect of the mbuf implementation. 27521da177e4SLinus Torvalds * Combining these two algorithms results in the observed behavior 27531da177e4SLinus Torvalds * of having a fixed window size at almost all times. 27541da177e4SLinus Torvalds * 27551da177e4SLinus Torvalds * Below we obtain similar behavior by forcing the offered window to 27561da177e4SLinus Torvalds * a multiple of the mss when it is feasible to do so. 27571da177e4SLinus Torvalds * 27581da177e4SLinus Torvalds * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 27591da177e4SLinus Torvalds * Regular options like TIMESTAMP are taken into account. 27601da177e4SLinus Torvalds */ 27611da177e4SLinus Torvalds u32 __tcp_select_window(struct sock *sk) 27621da177e4SLinus Torvalds { 2763463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 27641da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2765caa20d9aSStephen Hemminger /* MSS for the peer's data. Previous versions used mss_clamp 27661da177e4SLinus Torvalds * here. I don't know if the value based on our guesses 27671da177e4SLinus Torvalds * of peer's MSS is better for the performance. It's more correct 27681da177e4SLinus Torvalds * but may be worse for the performance because of rcv_mss 27691da177e4SLinus Torvalds * fluctuations. --SAW 1998/11/1 27701da177e4SLinus Torvalds */ 2771463c84b9SArnaldo Carvalho de Melo int mss = icsk->icsk_ack.rcv_mss; 27721da177e4SLinus Torvalds int free_space = tcp_space(sk); 277386c1a045SFlorian Westphal int allowed_space = tcp_full_space(sk); 2774071c8ed6SFlorian Westphal int full_space, window; 2775071c8ed6SFlorian Westphal 2776071c8ed6SFlorian Westphal if (sk_is_mptcp(sk)) 2777071c8ed6SFlorian Westphal mptcp_space(sk, &free_space, &allowed_space); 2778071c8ed6SFlorian Westphal 2779071c8ed6SFlorian Westphal full_space = min_t(int, tp->window_clamp, allowed_space); 27801da177e4SLinus Torvalds 278106425c30SEric Dumazet if (unlikely(mss > full_space)) { 27821da177e4SLinus Torvalds mss = full_space; 278306425c30SEric Dumazet if (mss <= 0) 278406425c30SEric Dumazet return 0; 278506425c30SEric Dumazet } 2786b92edbe0SEric Dumazet if (free_space < (full_space >> 1)) { 2787463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.quick = 0; 27881da177e4SLinus Torvalds 2789b8da51ebSEric Dumazet if (tcp_under_memory_pressure(sk)) 2790056834d9SIlpo Järvinen tp->rcv_ssthresh = min(tp->rcv_ssthresh, 2791056834d9SIlpo Järvinen 4U * tp->advmss); 27921da177e4SLinus Torvalds 279386c1a045SFlorian Westphal /* free_space might become our new window, make sure we don't 279486c1a045SFlorian Westphal * increase it due to wscale. 279586c1a045SFlorian Westphal */ 279686c1a045SFlorian Westphal free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale); 279786c1a045SFlorian Westphal 279886c1a045SFlorian Westphal /* if free space is less than mss estimate, or is below 1/16th 279986c1a045SFlorian Westphal * of the maximum allowed, try to move to zero-window, else 280086c1a045SFlorian Westphal * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and 280186c1a045SFlorian Westphal * new incoming data is dropped due to memory limits. 280286c1a045SFlorian Westphal * With large window, mss test triggers way too late in order 280386c1a045SFlorian Westphal * to announce zero window in time before rmem limit kicks in. 280486c1a045SFlorian Westphal */ 280586c1a045SFlorian Westphal if (free_space < (allowed_space >> 4) || free_space < mss) 28061da177e4SLinus Torvalds return 0; 28071da177e4SLinus Torvalds } 28081da177e4SLinus Torvalds 28091da177e4SLinus Torvalds if (free_space > tp->rcv_ssthresh) 28101da177e4SLinus Torvalds free_space = tp->rcv_ssthresh; 28111da177e4SLinus Torvalds 28121da177e4SLinus Torvalds /* Don't do rounding if we are using window scaling, since the 28131da177e4SLinus Torvalds * scaled window will not line up with the MSS boundary anyway. 28141da177e4SLinus Torvalds */ 28151da177e4SLinus Torvalds if (tp->rx_opt.rcv_wscale) { 28161da177e4SLinus Torvalds window = free_space; 28171da177e4SLinus Torvalds 28181da177e4SLinus Torvalds /* Advertise enough space so that it won't get scaled away. 28191da177e4SLinus Torvalds * Import case: prevent zero window announcement if 28201da177e4SLinus Torvalds * 1<<rcv_wscale > mss. 28211da177e4SLinus Torvalds */ 28221935299dSGao Feng window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale)); 28231da177e4SLinus Torvalds } else { 28241935299dSGao Feng window = tp->rcv_wnd; 28251da177e4SLinus Torvalds /* Get the largest window that is a nice multiple of mss. 28261da177e4SLinus Torvalds * Window clamp already applied above. 28271da177e4SLinus Torvalds * If our current window offering is within 1 mss of the 28281da177e4SLinus Torvalds * free space we just keep it. This prevents the divide 28291da177e4SLinus Torvalds * and multiply from happening most of the time. 28301da177e4SLinus Torvalds * We also don't do any window rounding when the free space 28311da177e4SLinus Torvalds * is too small. 28321da177e4SLinus Torvalds */ 28331da177e4SLinus Torvalds if (window <= free_space - mss || window > free_space) 28341935299dSGao Feng window = rounddown(free_space, mss); 283584565070SJohn Heffner else if (mss == full_space && 2836b92edbe0SEric Dumazet free_space > window + (full_space >> 1)) 283784565070SJohn Heffner window = free_space; 28381da177e4SLinus Torvalds } 28391da177e4SLinus Torvalds 28401da177e4SLinus Torvalds return window; 28411da177e4SLinus Torvalds } 28421da177e4SLinus Torvalds 2843cfea5a68SMartin KaFai Lau void tcp_skb_collapse_tstamp(struct sk_buff *skb, 2844082ac2d5SMartin KaFai Lau const struct sk_buff *next_skb) 2845082ac2d5SMartin KaFai Lau { 28460a2cf20cSSoheil Hassas Yeganeh if (unlikely(tcp_has_tx_tstamp(next_skb))) { 28470a2cf20cSSoheil Hassas Yeganeh const struct skb_shared_info *next_shinfo = 28480a2cf20cSSoheil Hassas Yeganeh skb_shinfo(next_skb); 2849082ac2d5SMartin KaFai Lau struct skb_shared_info *shinfo = skb_shinfo(skb); 2850082ac2d5SMartin KaFai Lau 28510a2cf20cSSoheil Hassas Yeganeh shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP; 2852082ac2d5SMartin KaFai Lau shinfo->tskey = next_shinfo->tskey; 28532de8023eSMartin KaFai Lau TCP_SKB_CB(skb)->txstamp_ack |= 28542de8023eSMartin KaFai Lau TCP_SKB_CB(next_skb)->txstamp_ack; 2855082ac2d5SMartin KaFai Lau } 2856082ac2d5SMartin KaFai Lau } 2857082ac2d5SMartin KaFai Lau 28584a17fc3aSIlpo Järvinen /* Collapses two adjacent SKB's during retransmission. */ 2859f8071cdeSEric Dumazet static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) 28601da177e4SLinus Torvalds { 28611da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 286275c119afSEric Dumazet struct sk_buff *next_skb = skb_rb_next(skb); 286313dde04fSWei Yongjun int next_skb_size; 28641da177e4SLinus Torvalds 2865058dc334SIlpo Järvinen next_skb_size = next_skb->len; 28661da177e4SLinus Torvalds 2867058dc334SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); 28681da177e4SLinus Torvalds 2869f8071cdeSEric Dumazet if (next_skb_size) { 2870f8071cdeSEric Dumazet if (next_skb_size <= skb_availroom(skb)) 2871f8071cdeSEric Dumazet skb_copy_bits(next_skb, 0, skb_put(skb, next_skb_size), 2872f8071cdeSEric Dumazet next_skb_size); 28733b4929f6SEric Dumazet else if (!tcp_skb_shift(skb, next_skb, 1, next_skb_size)) 2874f8071cdeSEric Dumazet return false; 2875f8071cdeSEric Dumazet } 28762b7cda9cSEric Dumazet tcp_highest_sack_replace(sk, next_skb, skb); 2877a6963a6bSIlpo Järvinen 28781da177e4SLinus Torvalds /* Update sequence range on original skb. */ 28791da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 28801da177e4SLinus Torvalds 2881e6c7d085SIlpo Järvinen /* Merge over control information. This moves PSH/FIN etc. over */ 28824de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; 28831da177e4SLinus Torvalds 28841da177e4SLinus Torvalds /* All done, get rid of second SKB and account for it so 28851da177e4SLinus Torvalds * packet counting does not break. 28861da177e4SLinus Torvalds */ 28874828e7f4SIlpo Järvinen TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; 2888a643b5d4SMartin KaFai Lau TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor; 2889b7689205SIlpo Järvinen 2890b7689205SIlpo Järvinen /* changed transmit queue under us so clear hints */ 2891ef9da47cSIlpo Järvinen tcp_clear_retrans_hints_partial(tp); 2892ef9da47cSIlpo Järvinen if (next_skb == tp->retransmit_skb_hint) 2893ef9da47cSIlpo Järvinen tp->retransmit_skb_hint = skb; 2894b7689205SIlpo Järvinen 2895797108d1SIlpo Järvinen tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); 2896797108d1SIlpo Järvinen 2897082ac2d5SMartin KaFai Lau tcp_skb_collapse_tstamp(skb, next_skb); 2898082ac2d5SMartin KaFai Lau 289975c119afSEric Dumazet tcp_rtx_queue_unlink_and_free(next_skb, sk); 2900f8071cdeSEric Dumazet return true; 29011da177e4SLinus Torvalds } 29021da177e4SLinus Torvalds 290367edfef7SAndi Kleen /* Check if coalescing SKBs is legal. */ 2904a2a385d6SEric Dumazet static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) 29054a17fc3aSIlpo Järvinen { 29064a17fc3aSIlpo Järvinen if (tcp_skb_pcount(skb) > 1) 2907a2a385d6SEric Dumazet return false; 29084a17fc3aSIlpo Järvinen if (skb_cloned(skb)) 2909a2a385d6SEric Dumazet return false; 29102331ccc5SEric Dumazet /* Some heuristics for collapsing over SACK'd could be invented */ 29114a17fc3aSIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 2912a2a385d6SEric Dumazet return false; 29134a17fc3aSIlpo Järvinen 2914a2a385d6SEric Dumazet return true; 29154a17fc3aSIlpo Järvinen } 29164a17fc3aSIlpo Järvinen 291767edfef7SAndi Kleen /* Collapse packets in the retransmit queue to make to create 291867edfef7SAndi Kleen * less packets on the wire. This is only done on retransmission. 291967edfef7SAndi Kleen */ 29204a17fc3aSIlpo Järvinen static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, 29214a17fc3aSIlpo Järvinen int space) 29224a17fc3aSIlpo Järvinen { 29234a17fc3aSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 29244a17fc3aSIlpo Järvinen struct sk_buff *skb = to, *tmp; 2925a2a385d6SEric Dumazet bool first = true; 29264a17fc3aSIlpo Järvinen 2927e0a1e5b5SEric Dumazet if (!sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse) 29284a17fc3aSIlpo Järvinen return; 29294de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 29304a17fc3aSIlpo Järvinen return; 29314a17fc3aSIlpo Järvinen 293275c119afSEric Dumazet skb_rbtree_walk_from_safe(skb, tmp) { 29334a17fc3aSIlpo Järvinen if (!tcp_can_collapse(sk, skb)) 29344a17fc3aSIlpo Järvinen break; 29354a17fc3aSIlpo Järvinen 293685712484SMat Martineau if (!tcp_skb_can_collapse(to, skb)) 2937a643b5d4SMartin KaFai Lau break; 2938a643b5d4SMartin KaFai Lau 29394a17fc3aSIlpo Järvinen space -= skb->len; 29404a17fc3aSIlpo Järvinen 29414a17fc3aSIlpo Järvinen if (first) { 2942a2a385d6SEric Dumazet first = false; 29434a17fc3aSIlpo Järvinen continue; 29444a17fc3aSIlpo Järvinen } 29454a17fc3aSIlpo Järvinen 29464a17fc3aSIlpo Järvinen if (space < 0) 29474a17fc3aSIlpo Järvinen break; 29484a17fc3aSIlpo Järvinen 29494a17fc3aSIlpo Järvinen if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) 29504a17fc3aSIlpo Järvinen break; 29514a17fc3aSIlpo Järvinen 2952f8071cdeSEric Dumazet if (!tcp_collapse_retrans(sk, to)) 2953f8071cdeSEric Dumazet break; 29544a17fc3aSIlpo Järvinen } 29554a17fc3aSIlpo Järvinen } 29564a17fc3aSIlpo Järvinen 29571da177e4SLinus Torvalds /* This retransmits one SKB. Policy decisions and retransmit queue 29581da177e4SLinus Torvalds * state updates are done by the caller. Returns non-zero if an 29591da177e4SLinus Torvalds * error occurred which prevented the send. 29601da177e4SLinus Torvalds */ 296110d3be56SEric Dumazet int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) 29621da177e4SLinus Torvalds { 29635d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 296410d3be56SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 29657d227cd2SSridhar Samudrala unsigned int cur_mss; 296610d3be56SEric Dumazet int diff, len, err; 29671da177e4SLinus Torvalds 296810d3be56SEric Dumazet 296910d3be56SEric Dumazet /* Inconclusive MTU probe */ 297010d3be56SEric Dumazet if (icsk->icsk_mtup.probe_size) 29715d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 29725d424d5aSJohn Heffner 29731da177e4SLinus Torvalds /* Do not sent more than we queued. 1/4 is reserved for possible 2974caa20d9aSStephen Hemminger * copying overhead: fragmentation, tunneling, mangling etc. 29751da177e4SLinus Torvalds */ 297614afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) > 2977ffb4d6c8SEric Dumazet min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), 2978ffb4d6c8SEric Dumazet sk->sk_sndbuf)) 29791da177e4SLinus Torvalds return -EAGAIN; 29801da177e4SLinus Torvalds 29811f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 29821f3279aeSEric Dumazet return -EBUSY; 29831f3279aeSEric Dumazet 29841da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 29857f582b24SEric Dumazet if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) { 29867f582b24SEric Dumazet WARN_ON_ONCE(1); 29877f582b24SEric Dumazet return -EINVAL; 29887f582b24SEric Dumazet } 29891da177e4SLinus Torvalds if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 29901da177e4SLinus Torvalds return -ENOMEM; 29911da177e4SLinus Torvalds } 29921da177e4SLinus Torvalds 29937d227cd2SSridhar Samudrala if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 29947d227cd2SSridhar Samudrala return -EHOSTUNREACH; /* Routing failure or similar. */ 29957d227cd2SSridhar Samudrala 29960c54b85fSIlpo Järvinen cur_mss = tcp_current_mss(sk); 29977d227cd2SSridhar Samudrala 29981da177e4SLinus Torvalds /* If receiver has shrunk his window, and skb is out of 29991da177e4SLinus Torvalds * new window, do not retransmit it. The exception is the 30001da177e4SLinus Torvalds * case, when window is shrunk to zero. In this case 30011da177e4SLinus Torvalds * our retransmit serves as a zero window probe. 30021da177e4SLinus Torvalds */ 30039d4fb27dSJoe Perches if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && 30049d4fb27dSJoe Perches TCP_SKB_CB(skb)->seq != tp->snd_una) 30051da177e4SLinus Torvalds return -EAGAIN; 30061da177e4SLinus Torvalds 300710d3be56SEric Dumazet len = cur_mss * segs; 300810d3be56SEric Dumazet if (skb->len > len) { 300975c119afSEric Dumazet if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len, 301075c119afSEric Dumazet cur_mss, GFP_ATOMIC)) 30111da177e4SLinus Torvalds return -ENOMEM; /* We'll try again later. */ 301202276f3cSIlpo Järvinen } else { 3013c52e2421SEric Dumazet if (skb_unclone(skb, GFP_ATOMIC)) 3014c52e2421SEric Dumazet return -ENOMEM; 301510d3be56SEric Dumazet 301610d3be56SEric Dumazet diff = tcp_skb_pcount(skb); 301710d3be56SEric Dumazet tcp_set_skb_tso_segs(skb, cur_mss); 301810d3be56SEric Dumazet diff -= tcp_skb_pcount(skb); 301910d3be56SEric Dumazet if (diff) 302010d3be56SEric Dumazet tcp_adjust_pcount(sk, skb, diff); 302110d3be56SEric Dumazet if (skb->len < cur_mss) 302210d3be56SEric Dumazet tcp_retrans_try_collapse(sk, skb, cur_mss); 30231da177e4SLinus Torvalds } 30241da177e4SLinus Torvalds 302549213555SDaniel Borkmann /* RFC3168, section 6.1.1.1. ECN fallback */ 302649213555SDaniel Borkmann if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN) 302749213555SDaniel Borkmann tcp_ecn_clear_syn(sk, skb); 302849213555SDaniel Borkmann 3029678550c6SYuchung Cheng /* Update global and local TCP statistics. */ 3030678550c6SYuchung Cheng segs = tcp_skb_pcount(skb); 3031678550c6SYuchung Cheng TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs); 3032678550c6SYuchung Cheng if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 3033678550c6SYuchung Cheng __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 3034678550c6SYuchung Cheng tp->total_retrans += segs; 3035fb31c9b9SWei Wang tp->bytes_retrans += skb->len; 3036678550c6SYuchung Cheng 303750bceae9SThomas Graf /* make sure skb->data is aligned on arches that require it 303850bceae9SThomas Graf * and check if ack-trimming & collapsing extended the headroom 303950bceae9SThomas Graf * beyond what csum_start can cover. 304050bceae9SThomas Graf */ 304150bceae9SThomas Graf if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) || 304250bceae9SThomas Graf skb_headroom(skb) >= 0xFFFF)) { 304310a81980SEric Dumazet struct sk_buff *nskb; 304410a81980SEric Dumazet 3045e2080072SEric Dumazet tcp_skb_tsorted_save(skb) { 304610a81980SEric Dumazet nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC); 304707f8e4d0SFlorian Westphal if (nskb) { 304807f8e4d0SFlorian Westphal nskb->dev = NULL; 304907f8e4d0SFlorian Westphal err = tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC); 305007f8e4d0SFlorian Westphal } else { 305107f8e4d0SFlorian Westphal err = -ENOBUFS; 305207f8e4d0SFlorian Westphal } 3053e2080072SEric Dumazet } tcp_skb_tsorted_restore(skb); 3054e2080072SEric Dumazet 30555889e2c0SYousuk Seung if (!err) { 3056a7a25630SEric Dumazet tcp_update_skb_after_send(sk, skb, tp->tcp_wstamp_ns); 30575889e2c0SYousuk Seung tcp_rate_skb_sent(sk, skb); 30585889e2c0SYousuk Seung } 3059117632e6SEric Dumazet } else { 3060c84a5711SYuchung Cheng err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 3061117632e6SEric Dumazet } 3062c84a5711SYuchung Cheng 30637f12422cSYuchung Cheng /* To avoid taking spuriously low RTT samples based on a timestamp 30647f12422cSYuchung Cheng * for a transmit that never happened, always mark EVER_RETRANS 30657f12422cSYuchung Cheng */ 30667f12422cSYuchung Cheng TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS; 30677f12422cSYuchung Cheng 3068a31ad29eSLawrence Brakmo if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RETRANS_CB_FLAG)) 3069a31ad29eSLawrence Brakmo tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RETRANS_CB, 3070a31ad29eSLawrence Brakmo TCP_SKB_CB(skb)->seq, segs, err); 3071a31ad29eSLawrence Brakmo 3072fc9f3501SEric Dumazet if (likely(!err)) { 3073e086101bSCong Wang trace_tcp_retransmit_skb(sk, skb); 3074678550c6SYuchung Cheng } else if (err != -EBUSY) { 3075ec641b39SYuchung Cheng NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL, segs); 3076fc9f3501SEric Dumazet } 3077c84a5711SYuchung Cheng return err; 307893b174adSYuchung Cheng } 307993b174adSYuchung Cheng 308010d3be56SEric Dumazet int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) 308193b174adSYuchung Cheng { 308293b174adSYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 308310d3be56SEric Dumazet int err = __tcp_retransmit_skb(sk, skb, segs); 30841da177e4SLinus Torvalds 30851da177e4SLinus Torvalds if (err == 0) { 30861da177e4SLinus Torvalds #if FASTRETRANS_DEBUG > 0 30871da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 3088e87cc472SJoe Perches net_dbg_ratelimited("retrans_out leaked\n"); 30891da177e4SLinus Torvalds } 30901da177e4SLinus Torvalds #endif 30911da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 30921da177e4SLinus Torvalds tp->retrans_out += tcp_skb_pcount(skb); 30937ae18975SYuchung Cheng } 30941da177e4SLinus Torvalds 30957ae18975SYuchung Cheng /* Save stamp of the first (attempted) retransmit. */ 30961da177e4SLinus Torvalds if (!tp->retrans_stamp) 30977faee5c0SEric Dumazet tp->retrans_stamp = tcp_skb_timestamp(skb); 30981da177e4SLinus Torvalds 30996e08d5e3SYuchung Cheng if (tp->undo_retrans < 0) 31006e08d5e3SYuchung Cheng tp->undo_retrans = 0; 31016e08d5e3SYuchung Cheng tp->undo_retrans += tcp_skb_pcount(skb); 31021da177e4SLinus Torvalds return err; 31031da177e4SLinus Torvalds } 31041da177e4SLinus Torvalds 31051da177e4SLinus Torvalds /* This gets called after a retransmit timeout, and the initially 31061da177e4SLinus Torvalds * retransmitted data is acknowledged. It tries to continue 31071da177e4SLinus Torvalds * resending the rest of the retransmit queue, until either 31081da177e4SLinus Torvalds * we've sent it all or the congestion window limit is reached. 31091da177e4SLinus Torvalds */ 31101da177e4SLinus Torvalds void tcp_xmit_retransmit_queue(struct sock *sk) 31111da177e4SLinus Torvalds { 31126687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 3113b9f1f1ceSEric Dumazet struct sk_buff *skb, *rtx_head, *hole = NULL; 31141da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 3115*916e6d1aSEric Dumazet bool rearm_timer = false; 3116840a3cbeSYuchung Cheng u32 max_segs; 311761eb55f4SIlpo Järvinen int mib_idx; 31186a438bbeSStephen Hemminger 311945e77d31SIlpo Järvinen if (!tp->packets_out) 312045e77d31SIlpo Järvinen return; 312145e77d31SIlpo Järvinen 312275c119afSEric Dumazet rtx_head = tcp_rtx_queue_head(sk); 3123b9f1f1ceSEric Dumazet skb = tp->retransmit_skb_hint ?: rtx_head; 3124ed6e7268SNeal Cardwell max_segs = tcp_tso_segs(sk, tcp_current_mss(sk)); 312575c119afSEric Dumazet skb_rbtree_walk_from(skb) { 3126dca0aaf8SEric Dumazet __u8 sacked; 312710d3be56SEric Dumazet int segs; 31281da177e4SLinus Torvalds 3129218af599SEric Dumazet if (tcp_pacing_check(sk)) 3130218af599SEric Dumazet break; 3131218af599SEric Dumazet 31326a438bbeSStephen Hemminger /* we could do better than to assign each time */ 313351456b29SIan Morris if (!hole) 31346a438bbeSStephen Hemminger tp->retransmit_skb_hint = skb; 31356a438bbeSStephen Hemminger 313610d3be56SEric Dumazet segs = tp->snd_cwnd - tcp_packets_in_flight(tp); 313710d3be56SEric Dumazet if (segs <= 0) 3138*916e6d1aSEric Dumazet break; 3139dca0aaf8SEric Dumazet sacked = TCP_SKB_CB(skb)->sacked; 3140a3d2e9f8SEric Dumazet /* In case tcp_shift_skb_data() have aggregated large skbs, 3141a3d2e9f8SEric Dumazet * we need to make sure not sending too bigs TSO packets 3142a3d2e9f8SEric Dumazet */ 3143a3d2e9f8SEric Dumazet segs = min_t(int, segs, max_segs); 31440e1c54c2SIlpo Järvinen 3145840a3cbeSYuchung Cheng if (tp->retrans_out >= tp->lost_out) { 3146006f582cSIlpo Järvinen break; 31470e1c54c2SIlpo Järvinen } else if (!(sacked & TCPCB_LOST)) { 314851456b29SIan Morris if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) 31490e1c54c2SIlpo Järvinen hole = skb; 315061eb55f4SIlpo Järvinen continue; 31511da177e4SLinus Torvalds 31520e1c54c2SIlpo Järvinen } else { 31530e1c54c2SIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Loss) 31540e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFASTRETRANS; 31550e1c54c2SIlpo Järvinen else 31560e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; 31570e1c54c2SIlpo Järvinen } 31580e1c54c2SIlpo Järvinen 31590e1c54c2SIlpo Järvinen if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) 316061eb55f4SIlpo Järvinen continue; 316140b215e5SPavel Emelyanov 3162f9616c35SEric Dumazet if (tcp_small_queue_check(sk, skb, 1)) 3163*916e6d1aSEric Dumazet break; 3164f9616c35SEric Dumazet 316510d3be56SEric Dumazet if (tcp_retransmit_skb(sk, skb, segs)) 3166*916e6d1aSEric Dumazet break; 316724ab6becSYuchung Cheng 3168de1d6578SYuchung Cheng NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb)); 31691da177e4SLinus Torvalds 3170684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 3171a262f0cdSNandita Dukkipati tp->prr_out += tcp_skb_pcount(skb); 3172a262f0cdSNandita Dukkipati 317375c119afSEric Dumazet if (skb == rtx_head && 317457dde7f7SYuchung Cheng icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT) 3175*916e6d1aSEric Dumazet rearm_timer = true; 3176*916e6d1aSEric Dumazet 3177*916e6d1aSEric Dumazet } 3178*916e6d1aSEric Dumazet if (rearm_timer) 31793f80e08fSEric Dumazet tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 31803f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, 31818dc242adSEric Dumazet TCP_RTO_MAX); 31821da177e4SLinus Torvalds } 31831da177e4SLinus Torvalds 3184d83769a5SEric Dumazet /* We allow to exceed memory limits for FIN packets to expedite 3185d83769a5SEric Dumazet * connection tear down and (memory) recovery. 3186845704a5SEric Dumazet * Otherwise tcp_send_fin() could be tempted to either delay FIN 3187845704a5SEric Dumazet * or even be forced to close flow without any FIN. 3188a6c5ea4cSEric Dumazet * In general, we want to allow one skb per socket to avoid hangs 3189a6c5ea4cSEric Dumazet * with edge trigger epoll() 3190d83769a5SEric Dumazet */ 3191a6c5ea4cSEric Dumazet void sk_forced_mem_schedule(struct sock *sk, int size) 3192d83769a5SEric Dumazet { 3193e805605cSJohannes Weiner int amt; 3194d83769a5SEric Dumazet 3195d83769a5SEric Dumazet if (size <= sk->sk_forward_alloc) 3196d83769a5SEric Dumazet return; 3197d83769a5SEric Dumazet amt = sk_mem_pages(size); 3198d83769a5SEric Dumazet sk->sk_forward_alloc += amt * SK_MEM_QUANTUM; 3199e805605cSJohannes Weiner sk_memory_allocated_add(sk, amt); 3200e805605cSJohannes Weiner 3201baac50bbSJohannes Weiner if (mem_cgroup_sockets_enabled && sk->sk_memcg) 3202baac50bbSJohannes Weiner mem_cgroup_charge_skmem(sk->sk_memcg, amt); 3203d83769a5SEric Dumazet } 3204d83769a5SEric Dumazet 3205845704a5SEric Dumazet /* Send a FIN. The caller locks the socket for us. 3206845704a5SEric Dumazet * We should try to send a FIN packet really hard, but eventually give up. 32071da177e4SLinus Torvalds */ 32081da177e4SLinus Torvalds void tcp_send_fin(struct sock *sk) 32091da177e4SLinus Torvalds { 3210ee2aabd3SEric Dumazet struct sk_buff *skb, *tskb, *tail = tcp_write_queue_tail(sk); 32111da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 32121da177e4SLinus Torvalds 3213845704a5SEric Dumazet /* Optimization, tack on the FIN if we have one skb in write queue and 3214845704a5SEric Dumazet * this skb was not yet sent, or we are under memory pressure. 3215845704a5SEric Dumazet * Note: in the latter case, FIN packet will be sent after a timeout, 3216845704a5SEric Dumazet * as TCP stack thinks it has already been transmitted. 32171da177e4SLinus Torvalds */ 3218ee2aabd3SEric Dumazet tskb = tail; 321975c119afSEric Dumazet if (!tskb && tcp_under_memory_pressure(sk)) 322075c119afSEric Dumazet tskb = skb_rb_last(&sk->tcp_rtx_queue); 322175c119afSEric Dumazet 322275c119afSEric Dumazet if (tskb) { 3223845704a5SEric Dumazet TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN; 3224845704a5SEric Dumazet TCP_SKB_CB(tskb)->end_seq++; 32251da177e4SLinus Torvalds tp->write_seq++; 3226ee2aabd3SEric Dumazet if (!tail) { 3227845704a5SEric Dumazet /* This means tskb was already sent. 3228845704a5SEric Dumazet * Pretend we included the FIN on previous transmit. 3229845704a5SEric Dumazet * We need to set tp->snd_nxt to the value it would have 3230845704a5SEric Dumazet * if FIN had been sent. This is because retransmit path 3231845704a5SEric Dumazet * does not change tp->snd_nxt. 3232845704a5SEric Dumazet */ 3233e0d694d6SEric Dumazet WRITE_ONCE(tp->snd_nxt, tp->snd_nxt + 1); 3234845704a5SEric Dumazet return; 3235845704a5SEric Dumazet } 32361da177e4SLinus Torvalds } else { 3237845704a5SEric Dumazet skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation); 3238d1edc085SColin Ian King if (unlikely(!skb)) 3239845704a5SEric Dumazet return; 3240d1edc085SColin Ian King 3241e2080072SEric Dumazet INIT_LIST_HEAD(&skb->tcp_tsorted_anchor); 3242d83769a5SEric Dumazet skb_reserve(skb, MAX_TCP_HEADER); 3243a6c5ea4cSEric Dumazet sk_forced_mem_schedule(sk, skb->truesize); 32441da177e4SLinus Torvalds /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 3245e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tp->write_seq, 3246a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_FIN); 32471da177e4SLinus Torvalds tcp_queue_skb(sk, skb); 32481da177e4SLinus Torvalds } 3249845704a5SEric Dumazet __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF); 32501da177e4SLinus Torvalds } 32511da177e4SLinus Torvalds 32521da177e4SLinus Torvalds /* We get here when a process closes a file descriptor (either due to 32531da177e4SLinus Torvalds * an explicit close() or as a byproduct of exit()'ing) and there 32541da177e4SLinus Torvalds * was unread data in the receive queue. This behavior is recommended 325565bb723cSGerrit Renker * by RFC 2525, section 2.17. -DaveM 32561da177e4SLinus Torvalds */ 3257dd0fc66fSAl Viro void tcp_send_active_reset(struct sock *sk, gfp_t priority) 32581da177e4SLinus Torvalds { 32591da177e4SLinus Torvalds struct sk_buff *skb; 32601da177e4SLinus Torvalds 32617cc2b043SGao Feng TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); 32627cc2b043SGao Feng 32631da177e4SLinus Torvalds /* NOTE: No TCP options attached and we never retransmit this. */ 32641da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, priority); 32651da177e4SLinus Torvalds if (!skb) { 32664e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 32671da177e4SLinus Torvalds return; 32681da177e4SLinus Torvalds } 32691da177e4SLinus Torvalds 32701da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 32711da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 3272e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), 3273a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_RST); 32749a568de4SEric Dumazet tcp_mstamp_refresh(tcp_sk(sk)); 32751da177e4SLinus Torvalds /* Send it off. */ 3276dfb4b9dcSDavid S. Miller if (tcp_transmit_skb(sk, skb, 0, priority)) 32774e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 3278c24b14c4SSong Liu 3279c24b14c4SSong Liu /* skb of trace_tcp_send_reset() keeps the skb that caused RST, 3280c24b14c4SSong Liu * skb here is different to the troublesome skb, so use NULL 3281c24b14c4SSong Liu */ 3282c24b14c4SSong Liu trace_tcp_send_reset(sk, NULL); 32831da177e4SLinus Torvalds } 32841da177e4SLinus Torvalds 328567edfef7SAndi Kleen /* Send a crossed SYN-ACK during socket establishment. 328667edfef7SAndi Kleen * WARNING: This routine must only be called when we have already sent 32871da177e4SLinus Torvalds * a SYN packet that crossed the incoming SYN that caused this routine 32881da177e4SLinus Torvalds * to get called. If this assumption fails then the initial rcv_wnd 32891da177e4SLinus Torvalds * and rcv_wscale values will not be correct. 32901da177e4SLinus Torvalds */ 32911da177e4SLinus Torvalds int tcp_send_synack(struct sock *sk) 32921da177e4SLinus Torvalds { 32931da177e4SLinus Torvalds struct sk_buff *skb; 32941da177e4SLinus Torvalds 329575c119afSEric Dumazet skb = tcp_rtx_queue_head(sk); 329651456b29SIan Morris if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 329775c119afSEric Dumazet pr_err("%s: wrong queue state\n", __func__); 32981da177e4SLinus Torvalds return -EFAULT; 32991da177e4SLinus Torvalds } 33004de075e0SEric Dumazet if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { 33011da177e4SLinus Torvalds if (skb_cloned(skb)) { 3302e2080072SEric Dumazet struct sk_buff *nskb; 3303e2080072SEric Dumazet 3304e2080072SEric Dumazet tcp_skb_tsorted_save(skb) { 3305e2080072SEric Dumazet nskb = skb_copy(skb, GFP_ATOMIC); 3306e2080072SEric Dumazet } tcp_skb_tsorted_restore(skb); 330751456b29SIan Morris if (!nskb) 33081da177e4SLinus Torvalds return -ENOMEM; 3309e2080072SEric Dumazet INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor); 33102bec445fSEric Dumazet tcp_highest_sack_replace(sk, skb, nskb); 331175c119afSEric Dumazet tcp_rtx_queue_unlink_and_free(skb, sk); 3312f4a775d1SEric Dumazet __skb_header_release(nskb); 331375c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb); 3314ab4e846aSEric Dumazet sk_wmem_queued_add(sk, nskb->truesize); 33153ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 33161da177e4SLinus Torvalds skb = nskb; 33171da177e4SLinus Torvalds } 33181da177e4SLinus Torvalds 33194de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; 3320735d3831SFlorian Westphal tcp_ecn_send_synack(sk, skb); 33211da177e4SLinus Torvalds } 3322dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 33231da177e4SLinus Torvalds } 33241da177e4SLinus Torvalds 33254aea39c1SEric Dumazet /** 33264aea39c1SEric Dumazet * tcp_make_synack - Prepare a SYN-ACK. 33274aea39c1SEric Dumazet * sk: listener socket 33284aea39c1SEric Dumazet * dst: dst entry attached to the SYNACK 33294aea39c1SEric Dumazet * req: request_sock pointer 33304aea39c1SEric Dumazet * 33314aea39c1SEric Dumazet * Allocate one skb and build a SYNACK packet. 33324aea39c1SEric Dumazet * @dst is consumed : Caller should not use it again. 33334aea39c1SEric Dumazet */ 33345d062de7SEric Dumazet struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst, 3335e6b4d113SWilliam Allen Simpson struct request_sock *req, 3336ca6fb065SEric Dumazet struct tcp_fastopen_cookie *foc, 3337b3d05147SEric Dumazet enum tcp_synack_type synack_type) 33381da177e4SLinus Torvalds { 33392e6599cbSArnaldo Carvalho de Melo struct inet_request_sock *ireq = inet_rsk(req); 33405d062de7SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 334180f03e27SEric Dumazet struct tcp_md5sig_key *md5 = NULL; 33425d062de7SEric Dumazet struct tcp_out_options opts; 33435d062de7SEric Dumazet struct sk_buff *skb; 3344bd0388aeSWilliam Allen Simpson int tcp_header_size; 33455d062de7SEric Dumazet struct tcphdr *th; 3346f5fff5dcSTom Quetchenbach int mss; 3347a842fe14SEric Dumazet u64 now; 33481da177e4SLinus Torvalds 3349ca6fb065SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 33504aea39c1SEric Dumazet if (unlikely(!skb)) { 33514aea39c1SEric Dumazet dst_release(dst); 33521da177e4SLinus Torvalds return NULL; 33534aea39c1SEric Dumazet } 33541da177e4SLinus Torvalds /* Reserve space for headers. */ 33551da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 33561da177e4SLinus Torvalds 3357b3d05147SEric Dumazet switch (synack_type) { 3358b3d05147SEric Dumazet case TCP_SYNACK_NORMAL: 33599e17f8a4SEric Dumazet skb_set_owner_w(skb, req_to_sk(req)); 3360b3d05147SEric Dumazet break; 3361b3d05147SEric Dumazet case TCP_SYNACK_COOKIE: 3362b3d05147SEric Dumazet /* Under synflood, we do not attach skb to a socket, 3363b3d05147SEric Dumazet * to avoid false sharing. 3364b3d05147SEric Dumazet */ 3365b3d05147SEric Dumazet break; 3366b3d05147SEric Dumazet case TCP_SYNACK_FASTOPEN: 3367ca6fb065SEric Dumazet /* sk is a const pointer, because we want to express multiple 3368ca6fb065SEric Dumazet * cpu might call us concurrently. 3369ca6fb065SEric Dumazet * sk->sk_wmem_alloc in an atomic, we can promote to rw. 3370ca6fb065SEric Dumazet */ 3371ca6fb065SEric Dumazet skb_set_owner_w(skb, (struct sock *)sk); 3372b3d05147SEric Dumazet break; 3373ca6fb065SEric Dumazet } 33744aea39c1SEric Dumazet skb_dst_set(skb, dst); 33751da177e4SLinus Torvalds 33763541f9e8SEric Dumazet mss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 3377f5fff5dcSTom Quetchenbach 337833ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 3379a842fe14SEric Dumazet now = tcp_clock_ns(); 33808b5f12d0SFlorian Westphal #ifdef CONFIG_SYN_COOKIES 33818b5f12d0SFlorian Westphal if (unlikely(req->cookie_ts)) 3382200ecef6SEric Dumazet skb->skb_mstamp_ns = cookie_init_timestamp(req, now); 33838b5f12d0SFlorian Westphal else 33848b5f12d0SFlorian Westphal #endif 33859e450c1eSYuchung Cheng { 3386a842fe14SEric Dumazet skb->skb_mstamp_ns = now; 33879e450c1eSYuchung Cheng if (!tcp_rsk(req)->snt_synack) /* Timestamp first SYNACK */ 33889e450c1eSYuchung Cheng tcp_rsk(req)->snt_synack = tcp_skb_timestamp_us(skb); 33899e450c1eSYuchung Cheng } 339080f03e27SEric Dumazet 339180f03e27SEric Dumazet #ifdef CONFIG_TCP_MD5SIG 339280f03e27SEric Dumazet rcu_read_lock(); 3393fd3a154aSEric Dumazet md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req)); 339480f03e27SEric Dumazet #endif 339558d607d3SEric Dumazet skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4); 339660e2a778SUrsula Braun tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5, 339760e2a778SUrsula Braun foc) + sizeof(*th); 339833ad798cSAdam Langley 3399aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 3400aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 34011da177e4SLinus Torvalds 3402ea1627c2SEric Dumazet th = (struct tcphdr *)skb->data; 34031da177e4SLinus Torvalds memset(th, 0, sizeof(struct tcphdr)); 34041da177e4SLinus Torvalds th->syn = 1; 34051da177e4SLinus Torvalds th->ack = 1; 34066ac705b1SEric Dumazet tcp_ecn_make_synack(req, th); 3407b44084c2SEric Dumazet th->source = htons(ireq->ir_num); 3408634fb979SEric Dumazet th->dest = ireq->ir_rmt_port; 3409e05a90ecSJamal Hadi Salim skb->mark = ireq->ir_mark; 34103b117750SEric Dumazet skb->ip_summed = CHECKSUM_PARTIAL; 34113b117750SEric Dumazet th->seq = htonl(tcp_rsk(req)->snt_isn); 34128336886fSJerry Chu /* XXX data is queued and acked as is. No buffer/window check */ 34138336886fSJerry Chu th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt); 34141da177e4SLinus Torvalds 34151da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 3416ed53d0abSEric Dumazet th->window = htons(min(req->rsk_rcv_wnd, 65535U)); 34175d062de7SEric Dumazet tcp_options_write((__be32 *)(th + 1), NULL, &opts); 34181da177e4SLinus Torvalds th->doff = (tcp_header_size >> 2); 341990bbcc60SEric Dumazet __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS); 3420cfb6eeb4SYOSHIFUJI Hideaki 3421cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 3422cfb6eeb4SYOSHIFUJI Hideaki /* Okay, we have all we need - do the md5 hash if needed */ 342380f03e27SEric Dumazet if (md5) 3424bd0388aeSWilliam Allen Simpson tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, 342539f8e58eSEric Dumazet md5, req_to_sk(req), skb); 342680f03e27SEric Dumazet rcu_read_unlock(); 3427cfb6eeb4SYOSHIFUJI Hideaki #endif 3428cfb6eeb4SYOSHIFUJI Hideaki 3429a842fe14SEric Dumazet skb->skb_mstamp_ns = now; 3430a842fe14SEric Dumazet tcp_add_tx_delay(skb, tp); 3431a842fe14SEric Dumazet 34321da177e4SLinus Torvalds return skb; 34331da177e4SLinus Torvalds } 34344bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_make_synack); 34351da177e4SLinus Torvalds 343681164413SDaniel Borkmann static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst) 343781164413SDaniel Borkmann { 343881164413SDaniel Borkmann struct inet_connection_sock *icsk = inet_csk(sk); 343981164413SDaniel Borkmann const struct tcp_congestion_ops *ca; 344081164413SDaniel Borkmann u32 ca_key = dst_metric(dst, RTAX_CC_ALGO); 344181164413SDaniel Borkmann 344281164413SDaniel Borkmann if (ca_key == TCP_CA_UNSPEC) 344381164413SDaniel Borkmann return; 344481164413SDaniel Borkmann 344581164413SDaniel Borkmann rcu_read_lock(); 344681164413SDaniel Borkmann ca = tcp_ca_find_key(ca_key); 34470baf26b0SMartin KaFai Lau if (likely(ca && bpf_try_module_get(ca, ca->owner))) { 34480baf26b0SMartin KaFai Lau bpf_module_put(icsk->icsk_ca_ops, icsk->icsk_ca_ops->owner); 344981164413SDaniel Borkmann icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst); 345081164413SDaniel Borkmann icsk->icsk_ca_ops = ca; 345181164413SDaniel Borkmann } 345281164413SDaniel Borkmann rcu_read_unlock(); 345381164413SDaniel Borkmann } 345481164413SDaniel Borkmann 345567edfef7SAndi Kleen /* Do all connect socket setups that can be done AF independent. */ 3456f7e56a76Sstephen hemminger static void tcp_connect_init(struct sock *sk) 34571da177e4SLinus Torvalds { 3458cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 34591da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 34601da177e4SLinus Torvalds __u8 rcv_wscale; 346113d3b1ebSLawrence Brakmo u32 rcv_wnd; 34621da177e4SLinus Torvalds 34631da177e4SLinus Torvalds /* We'll fix this up when we get a response from the other end. 34641da177e4SLinus Torvalds * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 34651da177e4SLinus Torvalds */ 34665d2ed052SEric Dumazet tp->tcp_header_len = sizeof(struct tcphdr); 34675d2ed052SEric Dumazet if (sock_net(sk)->ipv4.sysctl_tcp_timestamps) 34685d2ed052SEric Dumazet tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED; 34691da177e4SLinus Torvalds 3470cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 347100db4124SIan Morris if (tp->af_specific->md5_lookup(sk, sk)) 3472cfb6eeb4SYOSHIFUJI Hideaki tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 3473cfb6eeb4SYOSHIFUJI Hideaki #endif 3474cfb6eeb4SYOSHIFUJI Hideaki 34751da177e4SLinus Torvalds /* If user gave his TCP_MAXSEG, record it to clamp */ 34761da177e4SLinus Torvalds if (tp->rx_opt.user_mss) 34771da177e4SLinus Torvalds tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 34781da177e4SLinus Torvalds tp->max_window = 0; 34795d424d5aSJohn Heffner tcp_mtup_init(sk); 34801da177e4SLinus Torvalds tcp_sync_mss(sk, dst_mtu(dst)); 34811da177e4SLinus Torvalds 348281164413SDaniel Borkmann tcp_ca_dst_init(sk, dst); 348381164413SDaniel Borkmann 34841da177e4SLinus Torvalds if (!tp->window_clamp) 34851da177e4SLinus Torvalds tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 34863541f9e8SEric Dumazet tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 3487f5fff5dcSTom Quetchenbach 34881da177e4SLinus Torvalds tcp_initialize_rcv_mss(sk); 34891da177e4SLinus Torvalds 3490e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 3491e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 3492e88c64f0SHagen Paul Pfeifer (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) 3493e88c64f0SHagen Paul Pfeifer tp->window_clamp = tcp_full_space(sk); 3494e88c64f0SHagen Paul Pfeifer 349513d3b1ebSLawrence Brakmo rcv_wnd = tcp_rwnd_init_bpf(sk); 349613d3b1ebSLawrence Brakmo if (rcv_wnd == 0) 349713d3b1ebSLawrence Brakmo rcv_wnd = dst_metric(dst, RTAX_INITRWND); 349813d3b1ebSLawrence Brakmo 3499ceef9ab6SEric Dumazet tcp_select_initial_window(sk, tcp_full_space(sk), 35001da177e4SLinus Torvalds tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 35011da177e4SLinus Torvalds &tp->rcv_wnd, 35021da177e4SLinus Torvalds &tp->window_clamp, 35039bb37ef0SEric Dumazet sock_net(sk)->ipv4.sysctl_tcp_window_scaling, 350431d12926Slaurent chavey &rcv_wscale, 350513d3b1ebSLawrence Brakmo rcv_wnd); 35061da177e4SLinus Torvalds 35071da177e4SLinus Torvalds tp->rx_opt.rcv_wscale = rcv_wscale; 35081da177e4SLinus Torvalds tp->rcv_ssthresh = tp->rcv_wnd; 35091da177e4SLinus Torvalds 35101da177e4SLinus Torvalds sk->sk_err = 0; 35111da177e4SLinus Torvalds sock_reset_flag(sk, SOCK_DONE); 35121da177e4SLinus Torvalds tp->snd_wnd = 0; 3513ee7537b6SHantzis Fotis tcp_init_wl(tp, 0); 35147f582b24SEric Dumazet tcp_write_queue_purge(sk); 35151da177e4SLinus Torvalds tp->snd_una = tp->write_seq; 35161da177e4SLinus Torvalds tp->snd_sml = tp->write_seq; 351733f5f57eSIlpo Järvinen tp->snd_up = tp->write_seq; 3518e0d694d6SEric Dumazet WRITE_ONCE(tp->snd_nxt, tp->write_seq); 3519ee995283SPavel Emelyanov 3520ee995283SPavel Emelyanov if (likely(!tp->repair)) 35211da177e4SLinus Torvalds tp->rcv_nxt = 0; 3522c7781a6eSAndrew Vagin else 352370eabf0eSEric Dumazet tp->rcv_tstamp = tcp_jiffies32; 3524ee995283SPavel Emelyanov tp->rcv_wup = tp->rcv_nxt; 35257db48e98SEric Dumazet WRITE_ONCE(tp->copied_seq, tp->rcv_nxt); 35261da177e4SLinus Torvalds 35278550f328SLawrence Brakmo inet_csk(sk)->icsk_rto = tcp_timeout_init(sk); 3528463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_retransmits = 0; 35291da177e4SLinus Torvalds tcp_clear_retrans(tp); 35301da177e4SLinus Torvalds } 35311da177e4SLinus Torvalds 3532783237e8SYuchung Cheng static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) 3533783237e8SYuchung Cheng { 3534783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3535783237e8SYuchung Cheng struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 3536783237e8SYuchung Cheng 3537783237e8SYuchung Cheng tcb->end_seq += skb->len; 3538f4a775d1SEric Dumazet __skb_header_release(skb); 3539ab4e846aSEric Dumazet sk_wmem_queued_add(sk, skb->truesize); 3540783237e8SYuchung Cheng sk_mem_charge(sk, skb->truesize); 35410f317464SEric Dumazet WRITE_ONCE(tp->write_seq, tcb->end_seq); 3542783237e8SYuchung Cheng tp->packets_out += tcp_skb_pcount(skb); 3543783237e8SYuchung Cheng } 3544783237e8SYuchung Cheng 3545783237e8SYuchung Cheng /* Build and send a SYN with data and (cached) Fast Open cookie. However, 3546783237e8SYuchung Cheng * queue a data-only packet after the regular SYN, such that regular SYNs 3547783237e8SYuchung Cheng * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges 3548783237e8SYuchung Cheng * only the SYN sequence, the data are retransmitted in the first ACK. 3549783237e8SYuchung Cheng * If cookie is not cached or other error occurs, falls back to send a 3550783237e8SYuchung Cheng * regular SYN with Fast Open cookie request option. 3551783237e8SYuchung Cheng */ 3552783237e8SYuchung Cheng static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) 3553783237e8SYuchung Cheng { 3554783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3555783237e8SYuchung Cheng struct tcp_fastopen_request *fo = tp->fastopen_req; 3556065263f4SWei Wang int space, err = 0; 3557355a901eSEric Dumazet struct sk_buff *syn_data; 3558783237e8SYuchung Cheng 355967da22d2SYuchung Cheng tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ 3560065263f4SWei Wang if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie)) 3561783237e8SYuchung Cheng goto fallback; 3562783237e8SYuchung Cheng 3563783237e8SYuchung Cheng /* MSS for SYN-data is based on cached MSS and bounded by PMTU and 3564783237e8SYuchung Cheng * user-MSS. Reserve maximum option space for middleboxes that add 3565783237e8SYuchung Cheng * private TCP options. The cost is reduced data space in SYN :( 3566783237e8SYuchung Cheng */ 35673541f9e8SEric Dumazet tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp); 35683541f9e8SEric Dumazet 35691b63edd6SYuchung Cheng space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) - 3570783237e8SYuchung Cheng MAX_TCP_OPTION_SPACE; 3571783237e8SYuchung Cheng 3572f5ddcbbbSEric Dumazet space = min_t(size_t, space, fo->size); 3573f5ddcbbbSEric Dumazet 3574f5ddcbbbSEric Dumazet /* limit to order-0 allocations */ 3575f5ddcbbbSEric Dumazet space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER)); 3576f5ddcbbbSEric Dumazet 3577eb934478SEric Dumazet syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false); 3578355a901eSEric Dumazet if (!syn_data) 3579783237e8SYuchung Cheng goto fallback; 3580355a901eSEric Dumazet syn_data->ip_summed = CHECKSUM_PARTIAL; 3581355a901eSEric Dumazet memcpy(syn_data->cb, syn->cb, sizeof(syn->cb)); 358207e100f9SEric Dumazet if (space) { 358307e100f9SEric Dumazet int copied = copy_from_iter(skb_put(syn_data, space), space, 358457be5bdaSAl Viro &fo->data->msg_iter); 358557be5bdaSAl Viro if (unlikely(!copied)) { 3586ba233b34SEric Dumazet tcp_skb_tsorted_anchor_cleanup(syn_data); 3587355a901eSEric Dumazet kfree_skb(syn_data); 3588783237e8SYuchung Cheng goto fallback; 3589783237e8SYuchung Cheng } 359057be5bdaSAl Viro if (copied != space) { 359157be5bdaSAl Viro skb_trim(syn_data, copied); 359257be5bdaSAl Viro space = copied; 359357be5bdaSAl Viro } 3594f859a448SWillem de Bruijn skb_zcopy_set(syn_data, fo->uarg, NULL); 359507e100f9SEric Dumazet } 3596355a901eSEric Dumazet /* No more data pending in inet_wait_for_connect() */ 3597355a901eSEric Dumazet if (space == fo->size) 3598355a901eSEric Dumazet fo->data = NULL; 3599355a901eSEric Dumazet fo->copied = space; 3600783237e8SYuchung Cheng 3601355a901eSEric Dumazet tcp_connect_queue_skb(sk, syn_data); 36020f87230dSFrancis Yan if (syn_data->len) 36030f87230dSFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_BUSY); 3604355a901eSEric Dumazet 3605355a901eSEric Dumazet err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation); 3606355a901eSEric Dumazet 3607d3edd06eSEric Dumazet syn->skb_mstamp_ns = syn_data->skb_mstamp_ns; 3608355a901eSEric Dumazet 3609355a901eSEric Dumazet /* Now full SYN+DATA was cloned and sent (or not), 3610355a901eSEric Dumazet * remove the SYN from the original skb (syn_data) 3611355a901eSEric Dumazet * we keep in write queue in case of a retransmit, as we 3612355a901eSEric Dumazet * also have the SYN packet (with no data) in the same queue. 3613431a9124SEric Dumazet */ 3614355a901eSEric Dumazet TCP_SKB_CB(syn_data)->seq++; 3615355a901eSEric Dumazet TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH; 3616355a901eSEric Dumazet if (!err) { 361767da22d2SYuchung Cheng tp->syn_data = (fo->copied > 0); 361875c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data); 3619f19c29e3SYuchung Cheng NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT); 3620783237e8SYuchung Cheng goto done; 3621783237e8SYuchung Cheng } 3622783237e8SYuchung Cheng 362375c119afSEric Dumazet /* data was not sent, put it in write_queue */ 362475c119afSEric Dumazet __skb_queue_tail(&sk->sk_write_queue, syn_data); 3625b5b7db8dSEric Dumazet tp->packets_out -= tcp_skb_pcount(syn_data); 3626b5b7db8dSEric Dumazet 3627783237e8SYuchung Cheng fallback: 3628783237e8SYuchung Cheng /* Send a regular SYN with Fast Open cookie request option */ 3629783237e8SYuchung Cheng if (fo->cookie.len > 0) 3630783237e8SYuchung Cheng fo->cookie.len = 0; 3631783237e8SYuchung Cheng err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); 3632783237e8SYuchung Cheng if (err) 3633783237e8SYuchung Cheng tp->syn_fastopen = 0; 3634783237e8SYuchung Cheng done: 3635783237e8SYuchung Cheng fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ 3636783237e8SYuchung Cheng return err; 3637783237e8SYuchung Cheng } 3638783237e8SYuchung Cheng 363967edfef7SAndi Kleen /* Build a SYN and send it off. */ 36401da177e4SLinus Torvalds int tcp_connect(struct sock *sk) 36411da177e4SLinus Torvalds { 36421da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 36431da177e4SLinus Torvalds struct sk_buff *buff; 3644ee586811SEric Paris int err; 36451da177e4SLinus Torvalds 3646de525be2SLawrence Brakmo tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB, 0, NULL); 36478ba60924SEric Dumazet 36488ba60924SEric Dumazet if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 36498ba60924SEric Dumazet return -EHOSTUNREACH; /* Routing failure or similar. */ 36508ba60924SEric Dumazet 36511da177e4SLinus Torvalds tcp_connect_init(sk); 36521da177e4SLinus Torvalds 36532b916477SAndrey Vagin if (unlikely(tp->repair)) { 36542b916477SAndrey Vagin tcp_finish_connect(sk, NULL); 36552b916477SAndrey Vagin return 0; 36562b916477SAndrey Vagin } 36572b916477SAndrey Vagin 3658eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true); 3659355a901eSEric Dumazet if (unlikely(!buff)) 36601da177e4SLinus Torvalds return -ENOBUFS; 36611da177e4SLinus Torvalds 3662a3433f35SChangli Gao tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); 36639a568de4SEric Dumazet tcp_mstamp_refresh(tp); 36649a568de4SEric Dumazet tp->retrans_stamp = tcp_time_stamp(tp); 3665783237e8SYuchung Cheng tcp_connect_queue_skb(sk, buff); 3666735d3831SFlorian Westphal tcp_ecn_send_syn(sk, buff); 366775c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, buff); 36681da177e4SLinus Torvalds 3669783237e8SYuchung Cheng /* Send off SYN; include data in Fast Open. */ 3670783237e8SYuchung Cheng err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : 3671783237e8SYuchung Cheng tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); 3672ee586811SEric Paris if (err == -ECONNREFUSED) 3673ee586811SEric Paris return err; 3674bd37a088SWei Yongjun 3675bd37a088SWei Yongjun /* We change tp->snd_nxt after the tcp_transmit_skb() call 3676bd37a088SWei Yongjun * in order to make this packet get counted in tcpOutSegs. 3677bd37a088SWei Yongjun */ 3678e0d694d6SEric Dumazet WRITE_ONCE(tp->snd_nxt, tp->write_seq); 3679bd37a088SWei Yongjun tp->pushed_seq = tp->write_seq; 3680b5b7db8dSEric Dumazet buff = tcp_send_head(sk); 3681b5b7db8dSEric Dumazet if (unlikely(buff)) { 3682e0d694d6SEric Dumazet WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(buff)->seq); 3683b5b7db8dSEric Dumazet tp->pushed_seq = TCP_SKB_CB(buff)->seq; 3684b5b7db8dSEric Dumazet } 368581cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); 36861da177e4SLinus Torvalds 36871da177e4SLinus Torvalds /* Timer for repeating the SYN until an answer. */ 36883f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 36893f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 36901da177e4SLinus Torvalds return 0; 36911da177e4SLinus Torvalds } 36924bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_connect); 36931da177e4SLinus Torvalds 36941da177e4SLinus Torvalds /* Send out a delayed ack, the caller does the policy checking 36951da177e4SLinus Torvalds * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 36961da177e4SLinus Torvalds * for details. 36971da177e4SLinus Torvalds */ 36981da177e4SLinus Torvalds void tcp_send_delayed_ack(struct sock *sk) 36991da177e4SLinus Torvalds { 3700463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 3701463c84b9SArnaldo Carvalho de Melo int ato = icsk->icsk_ack.ato; 37021da177e4SLinus Torvalds unsigned long timeout; 37031da177e4SLinus Torvalds 37041da177e4SLinus Torvalds if (ato > TCP_DELACK_MIN) { 3705463c84b9SArnaldo Carvalho de Melo const struct tcp_sock *tp = tcp_sk(sk); 37061da177e4SLinus Torvalds int max_ato = HZ / 2; 37071da177e4SLinus Torvalds 370831954cd8SWei Wang if (inet_csk_in_pingpong_mode(sk) || 3709056834d9SIlpo Järvinen (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 37101da177e4SLinus Torvalds max_ato = TCP_DELACK_MAX; 37111da177e4SLinus Torvalds 37121da177e4SLinus Torvalds /* Slow path, intersegment interval is "high". */ 37131da177e4SLinus Torvalds 37141da177e4SLinus Torvalds /* If some rtt estimate is known, use it to bound delayed ack. 3715463c84b9SArnaldo Carvalho de Melo * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 37161da177e4SLinus Torvalds * directly. 37171da177e4SLinus Torvalds */ 3718740b0f18SEric Dumazet if (tp->srtt_us) { 3719740b0f18SEric Dumazet int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3), 3720740b0f18SEric Dumazet TCP_DELACK_MIN); 37211da177e4SLinus Torvalds 37221da177e4SLinus Torvalds if (rtt < max_ato) 37231da177e4SLinus Torvalds max_ato = rtt; 37241da177e4SLinus Torvalds } 37251da177e4SLinus Torvalds 37261da177e4SLinus Torvalds ato = min(ato, max_ato); 37271da177e4SLinus Torvalds } 37281da177e4SLinus Torvalds 37291da177e4SLinus Torvalds /* Stay within the limit we were given */ 37301da177e4SLinus Torvalds timeout = jiffies + ato; 37311da177e4SLinus Torvalds 37321da177e4SLinus Torvalds /* Use new timeout only if there wasn't a older one earlier. */ 3733463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 37341da177e4SLinus Torvalds /* If delack timer was blocked or is about to expire, 37351da177e4SLinus Torvalds * send ACK now. 37361da177e4SLinus Torvalds */ 3737463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.blocked || 3738463c84b9SArnaldo Carvalho de Melo time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 37391da177e4SLinus Torvalds tcp_send_ack(sk); 37401da177e4SLinus Torvalds return; 37411da177e4SLinus Torvalds } 37421da177e4SLinus Torvalds 3743463c84b9SArnaldo Carvalho de Melo if (!time_before(timeout, icsk->icsk_ack.timeout)) 3744463c84b9SArnaldo Carvalho de Melo timeout = icsk->icsk_ack.timeout; 37451da177e4SLinus Torvalds } 3746463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 3747463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.timeout = timeout; 3748463c84b9SArnaldo Carvalho de Melo sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 37491da177e4SLinus Torvalds } 37501da177e4SLinus Torvalds 37511da177e4SLinus Torvalds /* This routine sends an ack and also updates the window. */ 37522987babbSYuchung Cheng void __tcp_send_ack(struct sock *sk, u32 rcv_nxt) 37531da177e4SLinus Torvalds { 37541da177e4SLinus Torvalds struct sk_buff *buff; 37551da177e4SLinus Torvalds 3756058dc334SIlpo Järvinen /* If we have been reset, we may not send again. */ 3757058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3758058dc334SIlpo Järvinen return; 3759058dc334SIlpo Järvinen 37601da177e4SLinus Torvalds /* We are not putting this on the write queue, so 37611da177e4SLinus Torvalds * tcp_transmit_skb() will set the ownership to this 37621da177e4SLinus Torvalds * sock. 37631da177e4SLinus Torvalds */ 37647450aaf6SEric Dumazet buff = alloc_skb(MAX_TCP_HEADER, 37657450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 37667450aaf6SEric Dumazet if (unlikely(!buff)) { 3767463c84b9SArnaldo Carvalho de Melo inet_csk_schedule_ack(sk); 3768463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; 37693f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 37703f421baaSArnaldo Carvalho de Melo TCP_DELACK_MAX, TCP_RTO_MAX); 37711da177e4SLinus Torvalds return; 37721da177e4SLinus Torvalds } 37731da177e4SLinus Torvalds 37741da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 37751da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 3776a3433f35SChangli Gao tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); 37771da177e4SLinus Torvalds 377898781965SEric Dumazet /* We do not want pure acks influencing TCP Small Queues or fq/pacing 377998781965SEric Dumazet * too much. 378098781965SEric Dumazet * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784 378198781965SEric Dumazet */ 378298781965SEric Dumazet skb_set_tcp_pure_ack(buff); 378398781965SEric Dumazet 37841da177e4SLinus Torvalds /* Send it off, this clears delayed acks for us. */ 37852987babbSYuchung Cheng __tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0, rcv_nxt); 37861da177e4SLinus Torvalds } 378727cde44aSYuchung Cheng EXPORT_SYMBOL_GPL(__tcp_send_ack); 37882987babbSYuchung Cheng 37892987babbSYuchung Cheng void tcp_send_ack(struct sock *sk) 37902987babbSYuchung Cheng { 37912987babbSYuchung Cheng __tcp_send_ack(sk, tcp_sk(sk)->rcv_nxt); 37921da177e4SLinus Torvalds } 37931da177e4SLinus Torvalds 37941da177e4SLinus Torvalds /* This routine sends a packet with an out of date sequence 37951da177e4SLinus Torvalds * number. It assumes the other end will try to ack it. 37961da177e4SLinus Torvalds * 37971da177e4SLinus Torvalds * Question: what should we make while urgent mode? 37981da177e4SLinus Torvalds * 4.4BSD forces sending single byte of data. We cannot send 37991da177e4SLinus Torvalds * out of window data, because we have SND.NXT==SND.MAX... 38001da177e4SLinus Torvalds * 38011da177e4SLinus Torvalds * Current solution: to send TWO zero-length segments in urgent mode: 38021da177e4SLinus Torvalds * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 38031da177e4SLinus Torvalds * out-of-date with SND.UNA-1 to probe window. 38041da177e4SLinus Torvalds */ 3805e520af48SEric Dumazet static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib) 38061da177e4SLinus Torvalds { 38071da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 38081da177e4SLinus Torvalds struct sk_buff *skb; 38091da177e4SLinus Torvalds 38101da177e4SLinus Torvalds /* We don't queue it, tcp_transmit_skb() sets ownership. */ 38117450aaf6SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, 38127450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 381351456b29SIan Morris if (!skb) 38141da177e4SLinus Torvalds return -1; 38151da177e4SLinus Torvalds 38161da177e4SLinus Torvalds /* Reserve space for headers and set control bits. */ 38171da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 38181da177e4SLinus Torvalds /* Use a previous sequence. This should cause the other 38191da177e4SLinus Torvalds * end to send an ack. Don't queue or clone SKB, just 38201da177e4SLinus Torvalds * send it. 38211da177e4SLinus Torvalds */ 3822a3433f35SChangli Gao tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); 3823e2e8009fSRenato Westphal NET_INC_STATS(sock_net(sk), mib); 38247450aaf6SEric Dumazet return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0); 38251da177e4SLinus Torvalds } 38261da177e4SLinus Torvalds 3827385e2070SEric Dumazet /* Called from setsockopt( ... TCP_REPAIR ) */ 3828ee995283SPavel Emelyanov void tcp_send_window_probe(struct sock *sk) 3829ee995283SPavel Emelyanov { 3830ee995283SPavel Emelyanov if (sk->sk_state == TCP_ESTABLISHED) { 3831ee995283SPavel Emelyanov tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; 38329a568de4SEric Dumazet tcp_mstamp_refresh(tcp_sk(sk)); 3833e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE); 3834ee995283SPavel Emelyanov } 3835ee995283SPavel Emelyanov } 3836ee995283SPavel Emelyanov 383767edfef7SAndi Kleen /* Initiate keepalive or window probe from timer. */ 3838e520af48SEric Dumazet int tcp_write_wakeup(struct sock *sk, int mib) 38391da177e4SLinus Torvalds { 38401da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 38411da177e4SLinus Torvalds struct sk_buff *skb; 38421da177e4SLinus Torvalds 3843058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3844058dc334SIlpo Järvinen return -1; 3845058dc334SIlpo Järvinen 384600db4124SIan Morris skb = tcp_send_head(sk); 384700db4124SIan Morris if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { 38481da177e4SLinus Torvalds int err; 38490c54b85fSIlpo Järvinen unsigned int mss = tcp_current_mss(sk); 385090840defSIlpo Järvinen unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 38511da177e4SLinus Torvalds 38521da177e4SLinus Torvalds if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 38531da177e4SLinus Torvalds tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 38541da177e4SLinus Torvalds 38551da177e4SLinus Torvalds /* We are probing the opening of a window 38561da177e4SLinus Torvalds * but the window size is != 0 38571da177e4SLinus Torvalds * must have been a result SWS avoidance ( sender ) 38581da177e4SLinus Torvalds */ 38591da177e4SLinus Torvalds if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 38601da177e4SLinus Torvalds skb->len > mss) { 38611da177e4SLinus Torvalds seg_size = min(seg_size, mss); 38624de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 386375c119afSEric Dumazet if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE, 386475c119afSEric Dumazet skb, seg_size, mss, GFP_ATOMIC)) 38651da177e4SLinus Torvalds return -1; 38661da177e4SLinus Torvalds } else if (!tcp_skb_pcount(skb)) 38675bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss); 38681da177e4SLinus Torvalds 38694de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 3870dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 387166f5fe62SIlpo Järvinen if (!err) 387266f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 38731da177e4SLinus Torvalds return err; 38741da177e4SLinus Torvalds } else { 387533f5f57eSIlpo Järvinen if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) 3876e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 1, mib); 3877e520af48SEric Dumazet return tcp_xmit_probe_skb(sk, 0, mib); 38781da177e4SLinus Torvalds } 38791da177e4SLinus Torvalds } 38801da177e4SLinus Torvalds 38811da177e4SLinus Torvalds /* A window probe timeout has occurred. If window is not closed send 38821da177e4SLinus Torvalds * a partial packet else a zero probe. 38831da177e4SLinus Torvalds */ 38841da177e4SLinus Torvalds void tcp_send_probe0(struct sock *sk) 38851da177e4SLinus Torvalds { 3886463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 38871da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 3888c6214a97SNikolay Borisov struct net *net = sock_net(sk); 3889c1d5674fSYuchung Cheng unsigned long timeout; 38901da177e4SLinus Torvalds int err; 38911da177e4SLinus Torvalds 3892e520af48SEric Dumazet err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE); 38931da177e4SLinus Torvalds 389475c119afSEric Dumazet if (tp->packets_out || tcp_write_queue_empty(sk)) { 38951da177e4SLinus Torvalds /* Cancel probe timer, if it is not required. */ 38966687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 0; 3897463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff = 0; 38981da177e4SLinus Torvalds return; 38991da177e4SLinus Torvalds } 39001da177e4SLinus Torvalds 3901c1d5674fSYuchung Cheng icsk->icsk_probes_out++; 39021da177e4SLinus Torvalds if (err <= 0) { 3903c6214a97SNikolay Borisov if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2) 3904463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff++; 3905c1d5674fSYuchung Cheng timeout = tcp_probe0_when(sk, TCP_RTO_MAX); 39061da177e4SLinus Torvalds } else { 39071da177e4SLinus Torvalds /* If packet was not sent due to local congestion, 3908c1d5674fSYuchung Cheng * Let senders fight for local resources conservatively. 39091da177e4SLinus Torvalds */ 3910c1d5674fSYuchung Cheng timeout = TCP_RESOURCE_PROBE_INTERVAL; 39111da177e4SLinus Torvalds } 39128dc242adSEric Dumazet tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, timeout, TCP_RTO_MAX); 39131da177e4SLinus Torvalds } 39145db92c99SOctavian Purdila 3915ea3bea3aSEric Dumazet int tcp_rtx_synack(const struct sock *sk, struct request_sock *req) 39165db92c99SOctavian Purdila { 39175db92c99SOctavian Purdila const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific; 39185db92c99SOctavian Purdila struct flowi fl; 39195db92c99SOctavian Purdila int res; 39205db92c99SOctavian Purdila 392158d607d3SEric Dumazet tcp_rsk(req)->txhash = net_tx_rndhash(); 3922b3d05147SEric Dumazet res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL); 39235db92c99SOctavian Purdila if (!res) { 392490bbcc60SEric Dumazet __TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); 392502a1d6e7SEric Dumazet __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 39267e32b443SYuchung Cheng if (unlikely(tcp_passive_fastopen(sk))) 39277e32b443SYuchung Cheng tcp_sk(sk)->total_retrans++; 3928cf34ce3dSSong Liu trace_tcp_retransmit_synack(sk, req); 39295db92c99SOctavian Purdila } 39305db92c99SOctavian Purdila return res; 39315db92c99SOctavian Purdila } 39325db92c99SOctavian Purdila EXPORT_SYMBOL(tcp_rtx_synack); 3933