11da177e4SLinus Torvalds /* 21da177e4SLinus Torvalds * INET An implementation of the TCP/IP protocol suite for the LINUX 31da177e4SLinus Torvalds * operating system. INET is implemented using the BSD Socket 41da177e4SLinus Torvalds * interface as the means of communication with the user level. 51da177e4SLinus Torvalds * 61da177e4SLinus Torvalds * Implementation of the Transmission Control Protocol(TCP). 71da177e4SLinus Torvalds * 802c30a84SJesper Juhl * Authors: Ross Biro 91da177e4SLinus Torvalds * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> 101da177e4SLinus Torvalds * Mark Evans, <evansmp@uhura.aston.ac.uk> 111da177e4SLinus Torvalds * Corey Minyard <wf-rch!minyard@relay.EU.net> 121da177e4SLinus Torvalds * Florian La Roche, <flla@stud.uni-sb.de> 131da177e4SLinus Torvalds * Charles Hedrick, <hedrick@klinzhai.rutgers.edu> 141da177e4SLinus Torvalds * Linus Torvalds, <torvalds@cs.helsinki.fi> 151da177e4SLinus Torvalds * Alan Cox, <gw4pts@gw4pts.ampr.org> 161da177e4SLinus Torvalds * Matthew Dillon, <dillon@apollo.west.oic.com> 171da177e4SLinus Torvalds * Arnt Gulbrandsen, <agulbra@nvg.unit.no> 181da177e4SLinus Torvalds * Jorge Cwik, <jorge@laser.satlink.net> 191da177e4SLinus Torvalds */ 201da177e4SLinus Torvalds 211da177e4SLinus Torvalds /* 221da177e4SLinus Torvalds * Changes: Pedro Roque : Retransmit queue handled by TCP. 231da177e4SLinus Torvalds * : Fragmentation on mtu decrease 241da177e4SLinus Torvalds * : Segment collapse on retransmit 251da177e4SLinus Torvalds * : AF independence 261da177e4SLinus Torvalds * 271da177e4SLinus Torvalds * Linus Torvalds : send_delayed_ack 281da177e4SLinus Torvalds * David S. Miller : Charge memory using the right skb 291da177e4SLinus Torvalds * during syn/ack processing. 301da177e4SLinus Torvalds * David S. Miller : Output engine completely rewritten. 311da177e4SLinus Torvalds * Andrea Arcangeli: SYNACK carry ts_recent in tsecr. 321da177e4SLinus Torvalds * Cacophonix Gaul : draft-minshall-nagle-01 331da177e4SLinus Torvalds * J Hadi Salim : ECN support 341da177e4SLinus Torvalds * 351da177e4SLinus Torvalds */ 361da177e4SLinus Torvalds 3791df42beSJoe Perches #define pr_fmt(fmt) "TCP: " fmt 3891df42beSJoe Perches 391da177e4SLinus Torvalds #include <net/tcp.h> 401da177e4SLinus Torvalds 411da177e4SLinus Torvalds #include <linux/compiler.h> 425a0e3ad6STejun Heo #include <linux/gfp.h> 431da177e4SLinus Torvalds #include <linux/module.h> 4460e2a778SUrsula Braun #include <linux/static_key.h> 451da177e4SLinus Torvalds 46e086101bSCong Wang #include <trace/events/tcp.h> 4735089bb2SDavid S. Miller 489799ccb0SEric Dumazet /* Refresh clocks of a TCP socket, 499799ccb0SEric Dumazet * ensuring monotically increasing values. 509799ccb0SEric Dumazet */ 519799ccb0SEric Dumazet void tcp_mstamp_refresh(struct tcp_sock *tp) 529799ccb0SEric Dumazet { 539799ccb0SEric Dumazet u64 val = tcp_clock_ns(); 549799ccb0SEric Dumazet 555f6188a8SEric Dumazet if (val > tp->tcp_clock_cache) 565f6188a8SEric Dumazet tp->tcp_clock_cache = val; 579799ccb0SEric Dumazet 589799ccb0SEric Dumazet val = div_u64(val, NSEC_PER_USEC); 599799ccb0SEric Dumazet if (val > tp->tcp_mstamp) 609799ccb0SEric Dumazet tp->tcp_mstamp = val; 619799ccb0SEric Dumazet } 629799ccb0SEric Dumazet 6346d3ceabSEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 6446d3ceabSEric Dumazet int push_one, gfp_t gfp); 65519855c5SWilliam Allen Simpson 6667edfef7SAndi Kleen /* Account for new data that has been sent to the network. */ 6775c119afSEric Dumazet static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb) 686ff03ac3SIlpo Järvinen { 696ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 706ff03ac3SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 7166f5fe62SIlpo Järvinen unsigned int prior_packets = tp->packets_out; 729e412ba7SIlpo Järvinen 731da177e4SLinus Torvalds tp->snd_nxt = TCP_SKB_CB(skb)->end_seq; 748512430eSIlpo Järvinen 7575c119afSEric Dumazet __skb_unlink(skb, &sk->sk_write_queue); 7675c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, skb); 7775c119afSEric Dumazet 7866f5fe62SIlpo Järvinen tp->packets_out += tcp_skb_pcount(skb); 79bec41a11SYuchung Cheng if (!prior_packets || icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) 80750ea2baSYuchung Cheng tcp_rearm_rto(sk); 81f19c29e3SYuchung Cheng 82f7324acdSDavid S. Miller NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT, 83f19c29e3SYuchung Cheng tcp_skb_pcount(skb)); 846a5dc9e5SEric Dumazet } 851da177e4SLinus Torvalds 86a4ecb15aSCui, Cheng /* SND.NXT, if window was not shrunk or the amount of shrunk was less than one 87a4ecb15aSCui, Cheng * window scaling factor due to loss of precision. 881da177e4SLinus Torvalds * If window has been shrunk, what should we make? It is not clear at all. 891da177e4SLinus Torvalds * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-( 901da177e4SLinus Torvalds * Anything in between SND.UNA...SND.UNA+SND.WND also can be already 911da177e4SLinus Torvalds * invalid. OK, let's make this for now: 921da177e4SLinus Torvalds */ 93cf533ea5SEric Dumazet static inline __u32 tcp_acceptable_seq(const struct sock *sk) 941da177e4SLinus Torvalds { 95cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 969e412ba7SIlpo Järvinen 97a4ecb15aSCui, Cheng if (!before(tcp_wnd_end(tp), tp->snd_nxt) || 98a4ecb15aSCui, Cheng (tp->rx_opt.wscale_ok && 99a4ecb15aSCui, Cheng ((tp->snd_nxt - tcp_wnd_end(tp)) < (1 << tp->rx_opt.rcv_wscale)))) 1001da177e4SLinus Torvalds return tp->snd_nxt; 1011da177e4SLinus Torvalds else 10290840defSIlpo Järvinen return tcp_wnd_end(tp); 1031da177e4SLinus Torvalds } 1041da177e4SLinus Torvalds 1051da177e4SLinus Torvalds /* Calculate mss to advertise in SYN segment. 1061da177e4SLinus Torvalds * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that: 1071da177e4SLinus Torvalds * 1081da177e4SLinus Torvalds * 1. It is independent of path mtu. 1091da177e4SLinus Torvalds * 2. Ideally, it is maximal possible segment size i.e. 65535-40. 1101da177e4SLinus Torvalds * 3. For IPv4 it is reasonable to calculate it from maximal MTU of 1111da177e4SLinus Torvalds * attached devices, because some buggy hosts are confused by 1121da177e4SLinus Torvalds * large MSS. 1131da177e4SLinus Torvalds * 4. We do not make 3, we advertise MSS, calculated from first 1141da177e4SLinus Torvalds * hop device mtu, but allow to raise it to ip_rt_min_advmss. 1151da177e4SLinus Torvalds * This may be overridden via information stored in routing table. 1161da177e4SLinus Torvalds * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible, 1171da177e4SLinus Torvalds * probably even Jumbo". 1181da177e4SLinus Torvalds */ 1191da177e4SLinus Torvalds static __u16 tcp_advertise_mss(struct sock *sk) 1201da177e4SLinus Torvalds { 1211da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 122cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1231da177e4SLinus Torvalds int mss = tp->advmss; 1241da177e4SLinus Torvalds 1250dbaee3bSDavid S. Miller if (dst) { 1260dbaee3bSDavid S. Miller unsigned int metric = dst_metric_advmss(dst); 1270dbaee3bSDavid S. Miller 1280dbaee3bSDavid S. Miller if (metric < mss) { 1290dbaee3bSDavid S. Miller mss = metric; 1301da177e4SLinus Torvalds tp->advmss = mss; 1311da177e4SLinus Torvalds } 1320dbaee3bSDavid S. Miller } 1331da177e4SLinus Torvalds 1341da177e4SLinus Torvalds return (__u16)mss; 1351da177e4SLinus Torvalds } 1361da177e4SLinus Torvalds 1371da177e4SLinus Torvalds /* RFC2861. Reset CWND after idle period longer RTO to "restart window". 1386f021c62SEric Dumazet * This is the first part of cwnd validation mechanism. 1396f021c62SEric Dumazet */ 1406f021c62SEric Dumazet void tcp_cwnd_restart(struct sock *sk, s32 delta) 1411da177e4SLinus Torvalds { 142463c84b9SArnaldo Carvalho de Melo struct tcp_sock *tp = tcp_sk(sk); 1436f021c62SEric Dumazet u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk)); 1441da177e4SLinus Torvalds u32 cwnd = tp->snd_cwnd; 1451da177e4SLinus Torvalds 1466687e988SArnaldo Carvalho de Melo tcp_ca_event(sk, CA_EVENT_CWND_RESTART); 1471da177e4SLinus Torvalds 1486687e988SArnaldo Carvalho de Melo tp->snd_ssthresh = tcp_current_ssthresh(sk); 1491da177e4SLinus Torvalds restart_cwnd = min(restart_cwnd, cwnd); 1501da177e4SLinus Torvalds 151463c84b9SArnaldo Carvalho de Melo while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd) 1521da177e4SLinus Torvalds cwnd >>= 1; 1531da177e4SLinus Torvalds tp->snd_cwnd = max(cwnd, restart_cwnd); 154c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 1551da177e4SLinus Torvalds tp->snd_cwnd_used = 0; 1561da177e4SLinus Torvalds } 1571da177e4SLinus Torvalds 15867edfef7SAndi Kleen /* Congestion state accounting after a packet has been sent. */ 15940efc6faSStephen Hemminger static void tcp_event_data_sent(struct tcp_sock *tp, 160cf533ea5SEric Dumazet struct sock *sk) 1611da177e4SLinus Torvalds { 162463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 163d635fbe2SEric Dumazet const u32 now = tcp_jiffies32; 1641da177e4SLinus Torvalds 16505c5a46dSNeal Cardwell if (tcp_packets_in_flight(tp) == 0) 16605c5a46dSNeal Cardwell tcp_ca_event(sk, CA_EVENT_TX_START); 16705c5a46dSNeal Cardwell 1681da177e4SLinus Torvalds tp->lsndtime = now; 1691da177e4SLinus Torvalds 1701da177e4SLinus Torvalds /* If it is a reply for ato after last received 1711da177e4SLinus Torvalds * packet, enter pingpong mode. 1721da177e4SLinus Torvalds */ 1732251ae46SJon Maxwell if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato) 174463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pingpong = 1; 1751da177e4SLinus Torvalds } 1761da177e4SLinus Torvalds 17767edfef7SAndi Kleen /* Account for an ACK we sent. */ 17827cde44aSYuchung Cheng static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts, 17927cde44aSYuchung Cheng u32 rcv_nxt) 1801da177e4SLinus Torvalds { 1815d9f4262SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 1825d9f4262SEric Dumazet 1835d9f4262SEric Dumazet if (unlikely(tp->compressed_ack)) { 184200d95f4SEric Dumazet NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED, 185200d95f4SEric Dumazet tp->compressed_ack); 1865d9f4262SEric Dumazet tp->compressed_ack = 0; 1875d9f4262SEric Dumazet if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1) 1885d9f4262SEric Dumazet __sock_put(sk); 1895d9f4262SEric Dumazet } 19027cde44aSYuchung Cheng 19127cde44aSYuchung Cheng if (unlikely(rcv_nxt != tp->rcv_nxt)) 19227cde44aSYuchung Cheng return; /* Special ACK sent by DCTCP to reflect ECN */ 193463c84b9SArnaldo Carvalho de Melo tcp_dec_quickack_mode(sk, pkts); 194463c84b9SArnaldo Carvalho de Melo inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); 1951da177e4SLinus Torvalds } 1961da177e4SLinus Torvalds 1971da177e4SLinus Torvalds /* Determine a window scaling and initial window to offer. 1981da177e4SLinus Torvalds * Based on the assumption that the given amount of space 1991da177e4SLinus Torvalds * will be offered. Store the results in the tp structure. 2001da177e4SLinus Torvalds * NOTE: for smooth operation initial space offering should 2011da177e4SLinus Torvalds * be a multiple of mss if possible. We assume here that mss >= 1. 2021da177e4SLinus Torvalds * This MUST be enforced by all callers. 2031da177e4SLinus Torvalds */ 204ceef9ab6SEric Dumazet void tcp_select_initial_window(const struct sock *sk, int __space, __u32 mss, 2051da177e4SLinus Torvalds __u32 *rcv_wnd, __u32 *window_clamp, 20631d12926Slaurent chavey int wscale_ok, __u8 *rcv_wscale, 20731d12926Slaurent chavey __u32 init_rcv_wnd) 2081da177e4SLinus Torvalds { 2091da177e4SLinus Torvalds unsigned int space = (__space < 0 ? 0 : __space); 2101da177e4SLinus Torvalds 2111da177e4SLinus Torvalds /* If no clamp set the clamp to the max possible scaled window */ 2121da177e4SLinus Torvalds if (*window_clamp == 0) 213589c49cbSGao Feng (*window_clamp) = (U16_MAX << TCP_MAX_WSCALE); 2141da177e4SLinus Torvalds space = min(*window_clamp, space); 2151da177e4SLinus Torvalds 2161da177e4SLinus Torvalds /* Quantize space offering to a multiple of mss if possible. */ 2171da177e4SLinus Torvalds if (space > mss) 218589c49cbSGao Feng space = rounddown(space, mss); 2191da177e4SLinus Torvalds 2201da177e4SLinus Torvalds /* NOTE: offering an initial window larger than 32767 22115d99e02SRick Jones * will break some buggy TCP stacks. If the admin tells us 22215d99e02SRick Jones * it is likely we could be speaking with such a buggy stack 22315d99e02SRick Jones * we will truncate our initial window offering to 32K-1 22415d99e02SRick Jones * unless the remote has sent us a window scaling option, 22515d99e02SRick Jones * which we interpret as a sign the remote TCP is not 22615d99e02SRick Jones * misinterpreting the window field as a signed quantity. 2271da177e4SLinus Torvalds */ 228ceef9ab6SEric Dumazet if (sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows) 2291da177e4SLinus Torvalds (*rcv_wnd) = min(space, MAX_TCP_WINDOW); 23015d99e02SRick Jones else 231a337531bSYuchung Cheng (*rcv_wnd) = min_t(u32, space, U16_MAX); 232a337531bSYuchung Cheng 233a337531bSYuchung Cheng if (init_rcv_wnd) 234a337531bSYuchung Cheng *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss); 23515d99e02SRick Jones 2361da177e4SLinus Torvalds (*rcv_wscale) = 0; 2371da177e4SLinus Torvalds if (wscale_ok) { 238589c49cbSGao Feng /* Set window scaling on max possible window */ 239356d1833SEric Dumazet space = max_t(u32, space, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]); 240f626300aSSoheil Hassas Yeganeh space = max_t(u32, space, sysctl_rmem_max); 241316c1592SStephen Hemminger space = min_t(u32, space, *window_clamp); 242589c49cbSGao Feng while (space > U16_MAX && (*rcv_wscale) < TCP_MAX_WSCALE) { 2431da177e4SLinus Torvalds space >>= 1; 2441da177e4SLinus Torvalds (*rcv_wscale)++; 2451da177e4SLinus Torvalds } 2461da177e4SLinus Torvalds } 2471da177e4SLinus Torvalds /* Set the clamp no higher than max representable value */ 248589c49cbSGao Feng (*window_clamp) = min_t(__u32, U16_MAX << (*rcv_wscale), *window_clamp); 2491da177e4SLinus Torvalds } 2504bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_select_initial_window); 2511da177e4SLinus Torvalds 2521da177e4SLinus Torvalds /* Chose a new window to advertise, update state in tcp_sock for the 2531da177e4SLinus Torvalds * socket, and return result with RFC1323 scaling applied. The return 2541da177e4SLinus Torvalds * value can be stuffed directly into th->window for an outgoing 2551da177e4SLinus Torvalds * frame. 2561da177e4SLinus Torvalds */ 25740efc6faSStephen Hemminger static u16 tcp_select_window(struct sock *sk) 2581da177e4SLinus Torvalds { 2591da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2608e165e20SFlorian Westphal u32 old_win = tp->rcv_wnd; 2611da177e4SLinus Torvalds u32 cur_win = tcp_receive_window(tp); 2621da177e4SLinus Torvalds u32 new_win = __tcp_select_window(sk); 2631da177e4SLinus Torvalds 2641da177e4SLinus Torvalds /* Never shrink the offered window */ 2651da177e4SLinus Torvalds if (new_win < cur_win) { 2661da177e4SLinus Torvalds /* Danger Will Robinson! 2671da177e4SLinus Torvalds * Don't update rcv_wup/rcv_wnd here or else 2681da177e4SLinus Torvalds * we will not be able to advertise a zero 2691da177e4SLinus Torvalds * window in time. --DaveM 2701da177e4SLinus Torvalds * 2711da177e4SLinus Torvalds * Relax Will Robinson. 2721da177e4SLinus Torvalds */ 2738e165e20SFlorian Westphal if (new_win == 0) 2748e165e20SFlorian Westphal NET_INC_STATS(sock_net(sk), 2758e165e20SFlorian Westphal LINUX_MIB_TCPWANTZEROWINDOWADV); 276607bfbf2SPatrick McHardy new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale); 2771da177e4SLinus Torvalds } 2781da177e4SLinus Torvalds tp->rcv_wnd = new_win; 2791da177e4SLinus Torvalds tp->rcv_wup = tp->rcv_nxt; 2801da177e4SLinus Torvalds 2811da177e4SLinus Torvalds /* Make sure we do not exceed the maximum possible 2821da177e4SLinus Torvalds * scaled window. 2831da177e4SLinus Torvalds */ 284ceef9ab6SEric Dumazet if (!tp->rx_opt.rcv_wscale && 285ceef9ab6SEric Dumazet sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows) 2861da177e4SLinus Torvalds new_win = min(new_win, MAX_TCP_WINDOW); 2871da177e4SLinus Torvalds else 2881da177e4SLinus Torvalds new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); 2891da177e4SLinus Torvalds 2901da177e4SLinus Torvalds /* RFC1323 scaling applied */ 2911da177e4SLinus Torvalds new_win >>= tp->rx_opt.rcv_wscale; 2921da177e4SLinus Torvalds 29331770e34SFlorian Westphal /* If we advertise zero window, disable fast path. */ 2948e165e20SFlorian Westphal if (new_win == 0) { 29531770e34SFlorian Westphal tp->pred_flags = 0; 2968e165e20SFlorian Westphal if (old_win) 2978e165e20SFlorian Westphal NET_INC_STATS(sock_net(sk), 2988e165e20SFlorian Westphal LINUX_MIB_TCPTOZEROWINDOWADV); 2998e165e20SFlorian Westphal } else if (old_win == 0) { 3008e165e20SFlorian Westphal NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV); 3018e165e20SFlorian Westphal } 3021da177e4SLinus Torvalds 3031da177e4SLinus Torvalds return new_win; 3041da177e4SLinus Torvalds } 3051da177e4SLinus Torvalds 30667edfef7SAndi Kleen /* Packet ECN state for a SYN-ACK */ 307735d3831SFlorian Westphal static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb) 308bdf1ee5dSIlpo Järvinen { 30930e502a3SDaniel Borkmann const struct tcp_sock *tp = tcp_sk(sk); 31030e502a3SDaniel Borkmann 3114de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR; 312bdf1ee5dSIlpo Järvinen if (!(tp->ecn_flags & TCP_ECN_OK)) 3134de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE; 31491b5b21cSLawrence Brakmo else if (tcp_ca_needs_ecn(sk) || 31591b5b21cSLawrence Brakmo tcp_bpf_ca_needs_ecn(sk)) 31630e502a3SDaniel Borkmann INET_ECN_xmit(sk); 317bdf1ee5dSIlpo Järvinen } 318bdf1ee5dSIlpo Järvinen 31967edfef7SAndi Kleen /* Packet ECN state for a SYN. */ 320735d3831SFlorian Westphal static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb) 321bdf1ee5dSIlpo Järvinen { 322bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 32391b5b21cSLawrence Brakmo bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk); 324f7b3bec6SFlorian Westphal bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 || 32591b5b21cSLawrence Brakmo tcp_ca_needs_ecn(sk) || bpf_needs_ecn; 326f7b3bec6SFlorian Westphal 327f7b3bec6SFlorian Westphal if (!use_ecn) { 328f7b3bec6SFlorian Westphal const struct dst_entry *dst = __sk_dst_get(sk); 329f7b3bec6SFlorian Westphal 330f7b3bec6SFlorian Westphal if (dst && dst_feature(dst, RTAX_FEATURE_ECN)) 331f7b3bec6SFlorian Westphal use_ecn = true; 332f7b3bec6SFlorian Westphal } 333bdf1ee5dSIlpo Järvinen 334bdf1ee5dSIlpo Järvinen tp->ecn_flags = 0; 335f7b3bec6SFlorian Westphal 336f7b3bec6SFlorian Westphal if (use_ecn) { 3374de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR; 338bdf1ee5dSIlpo Järvinen tp->ecn_flags = TCP_ECN_OK; 33991b5b21cSLawrence Brakmo if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn) 34030e502a3SDaniel Borkmann INET_ECN_xmit(sk); 341bdf1ee5dSIlpo Järvinen } 342bdf1ee5dSIlpo Järvinen } 343bdf1ee5dSIlpo Järvinen 34449213555SDaniel Borkmann static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb) 34549213555SDaniel Borkmann { 34649213555SDaniel Borkmann if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback) 34749213555SDaniel Borkmann /* tp->ecn_flags are cleared at a later point in time when 34849213555SDaniel Borkmann * SYN ACK is ultimatively being received. 34949213555SDaniel Borkmann */ 35049213555SDaniel Borkmann TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR); 35149213555SDaniel Borkmann } 35249213555SDaniel Borkmann 353735d3831SFlorian Westphal static void 3546ac705b1SEric Dumazet tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th) 355bdf1ee5dSIlpo Järvinen { 3566ac705b1SEric Dumazet if (inet_rsk(req)->ecn_ok) 357bdf1ee5dSIlpo Järvinen th->ece = 1; 358bdf1ee5dSIlpo Järvinen } 359bdf1ee5dSIlpo Järvinen 36067edfef7SAndi Kleen /* Set up ECN state for a packet on a ESTABLISHED socket that is about to 36167edfef7SAndi Kleen * be sent. 36267edfef7SAndi Kleen */ 363735d3831SFlorian Westphal static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb, 364ea1627c2SEric Dumazet struct tcphdr *th, int tcp_header_len) 365bdf1ee5dSIlpo Järvinen { 366bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 367bdf1ee5dSIlpo Järvinen 368bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_OK) { 369bdf1ee5dSIlpo Järvinen /* Not-retransmitted data segment: set ECT and inject CWR. */ 370bdf1ee5dSIlpo Järvinen if (skb->len != tcp_header_len && 371bdf1ee5dSIlpo Järvinen !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { 372bdf1ee5dSIlpo Järvinen INET_ECN_xmit(sk); 373bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) { 374bdf1ee5dSIlpo Järvinen tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; 375ea1627c2SEric Dumazet th->cwr = 1; 376bdf1ee5dSIlpo Järvinen skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; 377bdf1ee5dSIlpo Järvinen } 37830e502a3SDaniel Borkmann } else if (!tcp_ca_needs_ecn(sk)) { 379bdf1ee5dSIlpo Järvinen /* ACK or retransmitted segment: clear ECT|CE */ 380bdf1ee5dSIlpo Järvinen INET_ECN_dontxmit(sk); 381bdf1ee5dSIlpo Järvinen } 382bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) 383ea1627c2SEric Dumazet th->ece = 1; 384bdf1ee5dSIlpo Järvinen } 385bdf1ee5dSIlpo Järvinen } 386bdf1ee5dSIlpo Järvinen 387e870a8efSIlpo Järvinen /* Constructs common control bits of non-data skb. If SYN/FIN is present, 388e870a8efSIlpo Järvinen * auto increment end seqno. 389e870a8efSIlpo Järvinen */ 390e870a8efSIlpo Järvinen static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags) 391e870a8efSIlpo Järvinen { 3922e8e18efSDavid S. Miller skb->ip_summed = CHECKSUM_PARTIAL; 393e870a8efSIlpo Järvinen 3944de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags; 395e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->sacked = 0; 396e870a8efSIlpo Järvinen 397cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, 1); 398e870a8efSIlpo Järvinen 399e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->seq = seq; 400a3433f35SChangli Gao if (flags & (TCPHDR_SYN | TCPHDR_FIN)) 401e870a8efSIlpo Järvinen seq++; 402e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->end_seq = seq; 403e870a8efSIlpo Järvinen } 404e870a8efSIlpo Järvinen 405a2a385d6SEric Dumazet static inline bool tcp_urg_mode(const struct tcp_sock *tp) 40633f5f57eSIlpo Järvinen { 40733f5f57eSIlpo Järvinen return tp->snd_una != tp->snd_up; 40833f5f57eSIlpo Järvinen } 40933f5f57eSIlpo Järvinen 41033ad798cSAdam Langley #define OPTION_SACK_ADVERTISE (1 << 0) 41133ad798cSAdam Langley #define OPTION_TS (1 << 1) 41233ad798cSAdam Langley #define OPTION_MD5 (1 << 2) 41389e95a61SOri Finkelman #define OPTION_WSCALE (1 << 3) 4142100c8d2SYuchung Cheng #define OPTION_FAST_OPEN_COOKIE (1 << 8) 41560e2a778SUrsula Braun #define OPTION_SMC (1 << 9) 41660e2a778SUrsula Braun 41760e2a778SUrsula Braun static void smc_options_write(__be32 *ptr, u16 *options) 41860e2a778SUrsula Braun { 41960e2a778SUrsula Braun #if IS_ENABLED(CONFIG_SMC) 42060e2a778SUrsula Braun if (static_branch_unlikely(&tcp_have_smc)) { 42160e2a778SUrsula Braun if (unlikely(OPTION_SMC & *options)) { 42260e2a778SUrsula Braun *ptr++ = htonl((TCPOPT_NOP << 24) | 42360e2a778SUrsula Braun (TCPOPT_NOP << 16) | 42460e2a778SUrsula Braun (TCPOPT_EXP << 8) | 42560e2a778SUrsula Braun (TCPOLEN_EXP_SMC_BASE)); 42660e2a778SUrsula Braun *ptr++ = htonl(TCPOPT_SMC_MAGIC); 42760e2a778SUrsula Braun } 42860e2a778SUrsula Braun } 42960e2a778SUrsula Braun #endif 43060e2a778SUrsula Braun } 43133ad798cSAdam Langley 43233ad798cSAdam Langley struct tcp_out_options { 4332100c8d2SYuchung Cheng u16 options; /* bit field of OPTION_* */ 4342100c8d2SYuchung Cheng u16 mss; /* 0 to disable */ 43533ad798cSAdam Langley u8 ws; /* window scale, 0 to disable */ 43633ad798cSAdam Langley u8 num_sack_blocks; /* number of SACK blocks to include */ 437bd0388aeSWilliam Allen Simpson u8 hash_size; /* bytes in hash_location */ 438bd0388aeSWilliam Allen Simpson __u8 *hash_location; /* temporary pointer, overloaded */ 4392100c8d2SYuchung Cheng __u32 tsval, tsecr; /* need to include OPTION_TS */ 4402100c8d2SYuchung Cheng struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */ 44133ad798cSAdam Langley }; 44233ad798cSAdam Langley 44367edfef7SAndi Kleen /* Write previously computed TCP options to the packet. 44467edfef7SAndi Kleen * 44567edfef7SAndi Kleen * Beware: Something in the Internet is very sensitive to the ordering of 446fd6149d3SIlpo Järvinen * TCP options, we learned this through the hard way, so be careful here. 447fd6149d3SIlpo Järvinen * Luckily we can at least blame others for their non-compliance but from 4488e3bff96Sstephen hemminger * inter-operability perspective it seems that we're somewhat stuck with 449fd6149d3SIlpo Järvinen * the ordering which we have been using if we want to keep working with 450fd6149d3SIlpo Järvinen * those broken things (not that it currently hurts anybody as there isn't 451fd6149d3SIlpo Järvinen * particular reason why the ordering would need to be changed). 452fd6149d3SIlpo Järvinen * 453fd6149d3SIlpo Järvinen * At least SACK_PERM as the first option is known to lead to a disaster 454fd6149d3SIlpo Järvinen * (but it may well be that other scenarios fail similarly). 455fd6149d3SIlpo Järvinen */ 45633ad798cSAdam Langley static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp, 457bd0388aeSWilliam Allen Simpson struct tcp_out_options *opts) 458bd0388aeSWilliam Allen Simpson { 4592100c8d2SYuchung Cheng u16 options = opts->options; /* mungable copy */ 460bd0388aeSWilliam Allen Simpson 461bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_MD5 & options)) { 4621a2c6181SChristoph Paasch *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | 4631a2c6181SChristoph Paasch (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG); 464bd0388aeSWilliam Allen Simpson /* overload cookie hash location */ 465bd0388aeSWilliam Allen Simpson opts->hash_location = (__u8 *)ptr; 46633ad798cSAdam Langley ptr += 4; 46733ad798cSAdam Langley } 46833ad798cSAdam Langley 469fd6149d3SIlpo Järvinen if (unlikely(opts->mss)) { 470fd6149d3SIlpo Järvinen *ptr++ = htonl((TCPOPT_MSS << 24) | 471fd6149d3SIlpo Järvinen (TCPOLEN_MSS << 16) | 472fd6149d3SIlpo Järvinen opts->mss); 473fd6149d3SIlpo Järvinen } 474fd6149d3SIlpo Järvinen 475bd0388aeSWilliam Allen Simpson if (likely(OPTION_TS & options)) { 476bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 47733ad798cSAdam Langley *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | 47833ad798cSAdam Langley (TCPOLEN_SACK_PERM << 16) | 47933ad798cSAdam Langley (TCPOPT_TIMESTAMP << 8) | 48033ad798cSAdam Langley TCPOLEN_TIMESTAMP); 481bd0388aeSWilliam Allen Simpson options &= ~OPTION_SACK_ADVERTISE; 48233ad798cSAdam Langley } else { 483496c98dfSYOSHIFUJI Hideaki *ptr++ = htonl((TCPOPT_NOP << 24) | 48440efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 48540efc6faSStephen Hemminger (TCPOPT_TIMESTAMP << 8) | 48640efc6faSStephen Hemminger TCPOLEN_TIMESTAMP); 48740efc6faSStephen Hemminger } 48833ad798cSAdam Langley *ptr++ = htonl(opts->tsval); 48933ad798cSAdam Langley *ptr++ = htonl(opts->tsecr); 49033ad798cSAdam Langley } 49133ad798cSAdam Langley 492bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 49333ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 49433ad798cSAdam Langley (TCPOPT_NOP << 16) | 49533ad798cSAdam Langley (TCPOPT_SACK_PERM << 8) | 49633ad798cSAdam Langley TCPOLEN_SACK_PERM); 49733ad798cSAdam Langley } 49833ad798cSAdam Langley 499bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_WSCALE & options)) { 50033ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 50133ad798cSAdam Langley (TCPOPT_WINDOW << 16) | 50233ad798cSAdam Langley (TCPOLEN_WINDOW << 8) | 50333ad798cSAdam Langley opts->ws); 50433ad798cSAdam Langley } 50533ad798cSAdam Langley 50633ad798cSAdam Langley if (unlikely(opts->num_sack_blocks)) { 50733ad798cSAdam Langley struct tcp_sack_block *sp = tp->rx_opt.dsack ? 50833ad798cSAdam Langley tp->duplicate_sack : tp->selective_acks; 50940efc6faSStephen Hemminger int this_sack; 51040efc6faSStephen Hemminger 51140efc6faSStephen Hemminger *ptr++ = htonl((TCPOPT_NOP << 24) | 51240efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 51340efc6faSStephen Hemminger (TCPOPT_SACK << 8) | 51433ad798cSAdam Langley (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * 51540efc6faSStephen Hemminger TCPOLEN_SACK_PERBLOCK))); 5162de979bdSStephen Hemminger 51733ad798cSAdam Langley for (this_sack = 0; this_sack < opts->num_sack_blocks; 51833ad798cSAdam Langley ++this_sack) { 51940efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].start_seq); 52040efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].end_seq); 52140efc6faSStephen Hemminger } 5222de979bdSStephen Hemminger 52340efc6faSStephen Hemminger tp->rx_opt.dsack = 0; 52440efc6faSStephen Hemminger } 5252100c8d2SYuchung Cheng 5262100c8d2SYuchung Cheng if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) { 5272100c8d2SYuchung Cheng struct tcp_fastopen_cookie *foc = opts->fastopen_cookie; 5287f9b838bSDaniel Lee u8 *p = (u8 *)ptr; 5297f9b838bSDaniel Lee u32 len; /* Fast Open option length */ 5302100c8d2SYuchung Cheng 5317f9b838bSDaniel Lee if (foc->exp) { 5327f9b838bSDaniel Lee len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len; 5337f9b838bSDaniel Lee *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) | 5342100c8d2SYuchung Cheng TCPOPT_FASTOPEN_MAGIC); 5357f9b838bSDaniel Lee p += TCPOLEN_EXP_FASTOPEN_BASE; 5367f9b838bSDaniel Lee } else { 5377f9b838bSDaniel Lee len = TCPOLEN_FASTOPEN_BASE + foc->len; 5387f9b838bSDaniel Lee *p++ = TCPOPT_FASTOPEN; 5397f9b838bSDaniel Lee *p++ = len; 5402100c8d2SYuchung Cheng } 5417f9b838bSDaniel Lee 5427f9b838bSDaniel Lee memcpy(p, foc->val, foc->len); 5437f9b838bSDaniel Lee if ((len & 3) == 2) { 5447f9b838bSDaniel Lee p[foc->len] = TCPOPT_NOP; 5457f9b838bSDaniel Lee p[foc->len + 1] = TCPOPT_NOP; 5467f9b838bSDaniel Lee } 5477f9b838bSDaniel Lee ptr += (len + 3) >> 2; 5482100c8d2SYuchung Cheng } 54960e2a778SUrsula Braun 55060e2a778SUrsula Braun smc_options_write(ptr, &options); 55160e2a778SUrsula Braun } 55260e2a778SUrsula Braun 55360e2a778SUrsula Braun static void smc_set_option(const struct tcp_sock *tp, 55460e2a778SUrsula Braun struct tcp_out_options *opts, 55560e2a778SUrsula Braun unsigned int *remaining) 55660e2a778SUrsula Braun { 55760e2a778SUrsula Braun #if IS_ENABLED(CONFIG_SMC) 55860e2a778SUrsula Braun if (static_branch_unlikely(&tcp_have_smc)) { 55960e2a778SUrsula Braun if (tp->syn_smc) { 56060e2a778SUrsula Braun if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) { 56160e2a778SUrsula Braun opts->options |= OPTION_SMC; 56260e2a778SUrsula Braun *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED; 56360e2a778SUrsula Braun } 56460e2a778SUrsula Braun } 56560e2a778SUrsula Braun } 56660e2a778SUrsula Braun #endif 56760e2a778SUrsula Braun } 56860e2a778SUrsula Braun 56960e2a778SUrsula Braun static void smc_set_option_cond(const struct tcp_sock *tp, 57060e2a778SUrsula Braun const struct inet_request_sock *ireq, 57160e2a778SUrsula Braun struct tcp_out_options *opts, 57260e2a778SUrsula Braun unsigned int *remaining) 57360e2a778SUrsula Braun { 57460e2a778SUrsula Braun #if IS_ENABLED(CONFIG_SMC) 57560e2a778SUrsula Braun if (static_branch_unlikely(&tcp_have_smc)) { 57660e2a778SUrsula Braun if (tp->syn_smc && ireq->smc_ok) { 57760e2a778SUrsula Braun if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) { 57860e2a778SUrsula Braun opts->options |= OPTION_SMC; 57960e2a778SUrsula Braun *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED; 58060e2a778SUrsula Braun } 58160e2a778SUrsula Braun } 58260e2a778SUrsula Braun } 58360e2a778SUrsula Braun #endif 58440efc6faSStephen Hemminger } 58540efc6faSStephen Hemminger 58667edfef7SAndi Kleen /* Compute TCP options for SYN packets. This is not the final 58767edfef7SAndi Kleen * network wire format yet. 58867edfef7SAndi Kleen */ 58995c96174SEric Dumazet static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb, 59033ad798cSAdam Langley struct tcp_out_options *opts, 591cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 592cf533ea5SEric Dumazet { 59333ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 59495c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 595783237e8SYuchung Cheng struct tcp_fastopen_request *fastopen = tp->fastopen_req; 59633ad798cSAdam Langley 5978c2320e8SEric Dumazet *md5 = NULL; 598cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 5998c2320e8SEric Dumazet if (unlikely(rcu_access_pointer(tp->md5sig_info))) { 60033ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 60133ad798cSAdam Langley if (*md5) { 60233ad798cSAdam Langley opts->options |= OPTION_MD5; 603bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 604cfb6eeb4SYOSHIFUJI Hideaki } 6058c2320e8SEric Dumazet } 606cfb6eeb4SYOSHIFUJI Hideaki #endif 60733ad798cSAdam Langley 60833ad798cSAdam Langley /* We always get an MSS option. The option bytes which will be seen in 60933ad798cSAdam Langley * normal data packets should timestamps be used, must be in the MSS 61033ad798cSAdam Langley * advertised. But we subtract them from tp->mss_cache so that 61133ad798cSAdam Langley * calculations in tcp_sendmsg are simpler etc. So account for this 61233ad798cSAdam Langley * fact here if necessary. If we don't do this correctly, as a 61333ad798cSAdam Langley * receiver we won't recognize data packets as being full sized when we 61433ad798cSAdam Langley * should, and thus we won't abide by the delayed ACK rules correctly. 61533ad798cSAdam Langley * SACKs don't matter, we never delay an ACK when we have any of those 61633ad798cSAdam Langley * going out. */ 61733ad798cSAdam Langley opts->mss = tcp_advertise_mss(sk); 618bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 61933ad798cSAdam Langley 6205d2ed052SEric Dumazet if (likely(sock_net(sk)->ipv4.sysctl_tcp_timestamps && !*md5)) { 62133ad798cSAdam Langley opts->options |= OPTION_TS; 6227faee5c0SEric Dumazet opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset; 62333ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 624bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 62533ad798cSAdam Langley } 6269bb37ef0SEric Dumazet if (likely(sock_net(sk)->ipv4.sysctl_tcp_window_scaling)) { 62733ad798cSAdam Langley opts->ws = tp->rx_opt.rcv_wscale; 62889e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 629bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 63033ad798cSAdam Langley } 631f9301034SEric Dumazet if (likely(sock_net(sk)->ipv4.sysctl_tcp_sack)) { 63233ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 633b32d1310SDavid S. Miller if (unlikely(!(OPTION_TS & opts->options))) 634bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 63533ad798cSAdam Langley } 63633ad798cSAdam Langley 637783237e8SYuchung Cheng if (fastopen && fastopen->cookie.len >= 0) { 6382646c831SDaniel Lee u32 need = fastopen->cookie.len; 6392646c831SDaniel Lee 6402646c831SDaniel Lee need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE : 6412646c831SDaniel Lee TCPOLEN_FASTOPEN_BASE; 642783237e8SYuchung Cheng need = (need + 3) & ~3U; /* Align to 32 bits */ 643783237e8SYuchung Cheng if (remaining >= need) { 644783237e8SYuchung Cheng opts->options |= OPTION_FAST_OPEN_COOKIE; 645783237e8SYuchung Cheng opts->fastopen_cookie = &fastopen->cookie; 646783237e8SYuchung Cheng remaining -= need; 647783237e8SYuchung Cheng tp->syn_fastopen = 1; 6482646c831SDaniel Lee tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0; 649783237e8SYuchung Cheng } 650783237e8SYuchung Cheng } 651bd0388aeSWilliam Allen Simpson 65260e2a778SUrsula Braun smc_set_option(tp, opts, &remaining); 65360e2a778SUrsula Braun 654bd0388aeSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 65533ad798cSAdam Langley } 65633ad798cSAdam Langley 65767edfef7SAndi Kleen /* Set up TCP options for SYN-ACKs. */ 65860e2a778SUrsula Braun static unsigned int tcp_synack_options(const struct sock *sk, 65960e2a778SUrsula Braun struct request_sock *req, 66095c96174SEric Dumazet unsigned int mss, struct sk_buff *skb, 66133ad798cSAdam Langley struct tcp_out_options *opts, 66280f03e27SEric Dumazet const struct tcp_md5sig_key *md5, 6638336886fSJerry Chu struct tcp_fastopen_cookie *foc) 6644957faadSWilliam Allen Simpson { 66533ad798cSAdam Langley struct inet_request_sock *ireq = inet_rsk(req); 66695c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 66733ad798cSAdam Langley 66833ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 66980f03e27SEric Dumazet if (md5) { 67033ad798cSAdam Langley opts->options |= OPTION_MD5; 6714957faadSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 6724957faadSWilliam Allen Simpson 6734957faadSWilliam Allen Simpson /* We can't fit any SACK blocks in a packet with MD5 + TS 6744957faadSWilliam Allen Simpson * options. There was discussion about disabling SACK 6754957faadSWilliam Allen Simpson * rather than TS in order to fit in better with old, 6764957faadSWilliam Allen Simpson * buggy kernels, but that was deemed to be unnecessary. 6774957faadSWilliam Allen Simpson */ 678de213e5eSEric Dumazet ireq->tstamp_ok &= !ireq->sack_ok; 67933ad798cSAdam Langley } 68033ad798cSAdam Langley #endif 68133ad798cSAdam Langley 6824957faadSWilliam Allen Simpson /* We always send an MSS option. */ 68333ad798cSAdam Langley opts->mss = mss; 6844957faadSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 68533ad798cSAdam Langley 68633ad798cSAdam Langley if (likely(ireq->wscale_ok)) { 68733ad798cSAdam Langley opts->ws = ireq->rcv_wscale; 68889e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 6894957faadSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 69033ad798cSAdam Langley } 691de213e5eSEric Dumazet if (likely(ireq->tstamp_ok)) { 69233ad798cSAdam Langley opts->options |= OPTION_TS; 69395a22caeSFlorian Westphal opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off; 69433ad798cSAdam Langley opts->tsecr = req->ts_recent; 6954957faadSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 69633ad798cSAdam Langley } 69733ad798cSAdam Langley if (likely(ireq->sack_ok)) { 69833ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 699de213e5eSEric Dumazet if (unlikely(!ireq->tstamp_ok)) 7004957faadSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 70133ad798cSAdam Langley } 7027f9b838bSDaniel Lee if (foc != NULL && foc->len >= 0) { 7037f9b838bSDaniel Lee u32 need = foc->len; 7047f9b838bSDaniel Lee 7057f9b838bSDaniel Lee need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE : 7067f9b838bSDaniel Lee TCPOLEN_FASTOPEN_BASE; 7078336886fSJerry Chu need = (need + 3) & ~3U; /* Align to 32 bits */ 7088336886fSJerry Chu if (remaining >= need) { 7098336886fSJerry Chu opts->options |= OPTION_FAST_OPEN_COOKIE; 7108336886fSJerry Chu opts->fastopen_cookie = foc; 7118336886fSJerry Chu remaining -= need; 7128336886fSJerry Chu } 7138336886fSJerry Chu } 7144957faadSWilliam Allen Simpson 71560e2a778SUrsula Braun smc_set_option_cond(tcp_sk(sk), ireq, opts, &remaining); 71660e2a778SUrsula Braun 7174957faadSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 71833ad798cSAdam Langley } 71933ad798cSAdam Langley 72067edfef7SAndi Kleen /* Compute TCP options for ESTABLISHED sockets. This is not the 72167edfef7SAndi Kleen * final wire format yet. 72267edfef7SAndi Kleen */ 72395c96174SEric Dumazet static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb, 72433ad798cSAdam Langley struct tcp_out_options *opts, 725cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 726cf533ea5SEric Dumazet { 72733ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 72895c96174SEric Dumazet unsigned int size = 0; 729cabeccbdSIlpo Järvinen unsigned int eff_sacks; 73033ad798cSAdam Langley 7315843ef42SAndi Kleen opts->options = 0; 7325843ef42SAndi Kleen 7338c2320e8SEric Dumazet *md5 = NULL; 73433ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 7358c2320e8SEric Dumazet if (unlikely(rcu_access_pointer(tp->md5sig_info))) { 73633ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 7378c2320e8SEric Dumazet if (*md5) { 73833ad798cSAdam Langley opts->options |= OPTION_MD5; 73933ad798cSAdam Langley size += TCPOLEN_MD5SIG_ALIGNED; 74033ad798cSAdam Langley } 7418c2320e8SEric Dumazet } 74233ad798cSAdam Langley #endif 74333ad798cSAdam Langley 74433ad798cSAdam Langley if (likely(tp->rx_opt.tstamp_ok)) { 74533ad798cSAdam Langley opts->options |= OPTION_TS; 7467faee5c0SEric Dumazet opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0; 74733ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 74833ad798cSAdam Langley size += TCPOLEN_TSTAMP_ALIGNED; 74933ad798cSAdam Langley } 75033ad798cSAdam Langley 751cabeccbdSIlpo Järvinen eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; 752cabeccbdSIlpo Järvinen if (unlikely(eff_sacks)) { 75395c96174SEric Dumazet const unsigned int remaining = MAX_TCP_OPTION_SPACE - size; 75433ad798cSAdam Langley opts->num_sack_blocks = 75595c96174SEric Dumazet min_t(unsigned int, eff_sacks, 75633ad798cSAdam Langley (remaining - TCPOLEN_SACK_BASE_ALIGNED) / 75733ad798cSAdam Langley TCPOLEN_SACK_PERBLOCK); 75833ad798cSAdam Langley size += TCPOLEN_SACK_BASE_ALIGNED + 75933ad798cSAdam Langley opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; 76033ad798cSAdam Langley } 76133ad798cSAdam Langley 76233ad798cSAdam Langley return size; 76340efc6faSStephen Hemminger } 7641da177e4SLinus Torvalds 76546d3ceabSEric Dumazet 76646d3ceabSEric Dumazet /* TCP SMALL QUEUES (TSQ) 76746d3ceabSEric Dumazet * 76846d3ceabSEric Dumazet * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev) 76946d3ceabSEric Dumazet * to reduce RTT and bufferbloat. 77046d3ceabSEric Dumazet * We do this using a special skb destructor (tcp_wfree). 77146d3ceabSEric Dumazet * 77246d3ceabSEric Dumazet * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb 77346d3ceabSEric Dumazet * needs to be reallocated in a driver. 7748e3bff96Sstephen hemminger * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc 77546d3ceabSEric Dumazet * 77646d3ceabSEric Dumazet * Since transmit from skb destructor is forbidden, we use a tasklet 77746d3ceabSEric Dumazet * to process all sockets that eventually need to send more skbs. 77846d3ceabSEric Dumazet * We use one tasklet per cpu, with its own queue of sockets. 77946d3ceabSEric Dumazet */ 78046d3ceabSEric Dumazet struct tsq_tasklet { 78146d3ceabSEric Dumazet struct tasklet_struct tasklet; 78246d3ceabSEric Dumazet struct list_head head; /* queue of tcp sockets */ 78346d3ceabSEric Dumazet }; 78446d3ceabSEric Dumazet static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet); 78546d3ceabSEric Dumazet 78673a6bab5SEric Dumazet static void tcp_tsq_write(struct sock *sk) 7876f458dfbSEric Dumazet { 7886f458dfbSEric Dumazet if ((1 << sk->sk_state) & 7896f458dfbSEric Dumazet (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING | 790f9616c35SEric Dumazet TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) { 791f9616c35SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 792f9616c35SEric Dumazet 793f9616c35SEric Dumazet if (tp->lost_out > tp->retrans_out && 7943a91d29fSKoichiro Den tp->snd_cwnd > tcp_packets_in_flight(tp)) { 7953a91d29fSKoichiro Den tcp_mstamp_refresh(tp); 796f9616c35SEric Dumazet tcp_xmit_retransmit_queue(sk); 7973a91d29fSKoichiro Den } 798f9616c35SEric Dumazet 799f9616c35SEric Dumazet tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle, 800bf06200eSJohn Ogness 0, GFP_ATOMIC); 8016f458dfbSEric Dumazet } 802f9616c35SEric Dumazet } 80373a6bab5SEric Dumazet 80473a6bab5SEric Dumazet static void tcp_tsq_handler(struct sock *sk) 80573a6bab5SEric Dumazet { 80673a6bab5SEric Dumazet bh_lock_sock(sk); 80773a6bab5SEric Dumazet if (!sock_owned_by_user(sk)) 80873a6bab5SEric Dumazet tcp_tsq_write(sk); 80973a6bab5SEric Dumazet else if (!test_and_set_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags)) 81073a6bab5SEric Dumazet sock_hold(sk); 81173a6bab5SEric Dumazet bh_unlock_sock(sk); 81273a6bab5SEric Dumazet } 81346d3ceabSEric Dumazet /* 8148e3bff96Sstephen hemminger * One tasklet per cpu tries to send more skbs. 81546d3ceabSEric Dumazet * We run in tasklet context but need to disable irqs when 8168e3bff96Sstephen hemminger * transferring tsq->head because tcp_wfree() might 81746d3ceabSEric Dumazet * interrupt us (non NAPI drivers) 81846d3ceabSEric Dumazet */ 81946d3ceabSEric Dumazet static void tcp_tasklet_func(unsigned long data) 82046d3ceabSEric Dumazet { 82146d3ceabSEric Dumazet struct tsq_tasklet *tsq = (struct tsq_tasklet *)data; 82246d3ceabSEric Dumazet LIST_HEAD(list); 82346d3ceabSEric Dumazet unsigned long flags; 82446d3ceabSEric Dumazet struct list_head *q, *n; 82546d3ceabSEric Dumazet struct tcp_sock *tp; 82646d3ceabSEric Dumazet struct sock *sk; 82746d3ceabSEric Dumazet 82846d3ceabSEric Dumazet local_irq_save(flags); 82946d3ceabSEric Dumazet list_splice_init(&tsq->head, &list); 83046d3ceabSEric Dumazet local_irq_restore(flags); 83146d3ceabSEric Dumazet 83246d3ceabSEric Dumazet list_for_each_safe(q, n, &list) { 83346d3ceabSEric Dumazet tp = list_entry(q, struct tcp_sock, tsq_node); 83446d3ceabSEric Dumazet list_del(&tp->tsq_node); 83546d3ceabSEric Dumazet 83646d3ceabSEric Dumazet sk = (struct sock *)tp; 8370a9648f1SEric Dumazet smp_mb__before_atomic(); 8387aa5470cSEric Dumazet clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags); 8397aa5470cSEric Dumazet 8406f458dfbSEric Dumazet tcp_tsq_handler(sk); 84146d3ceabSEric Dumazet sk_free(sk); 84246d3ceabSEric Dumazet } 84346d3ceabSEric Dumazet } 84446d3ceabSEric Dumazet 84540fc3423SEric Dumazet #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \ 84640fc3423SEric Dumazet TCPF_WRITE_TIMER_DEFERRED | \ 84740fc3423SEric Dumazet TCPF_DELACK_TIMER_DEFERRED | \ 84840fc3423SEric Dumazet TCPF_MTU_REDUCED_DEFERRED) 84946d3ceabSEric Dumazet /** 85046d3ceabSEric Dumazet * tcp_release_cb - tcp release_sock() callback 85146d3ceabSEric Dumazet * @sk: socket 85246d3ceabSEric Dumazet * 85346d3ceabSEric Dumazet * called from release_sock() to perform protocol dependent 85446d3ceabSEric Dumazet * actions before socket release. 85546d3ceabSEric Dumazet */ 85646d3ceabSEric Dumazet void tcp_release_cb(struct sock *sk) 85746d3ceabSEric Dumazet { 8586f458dfbSEric Dumazet unsigned long flags, nflags; 85946d3ceabSEric Dumazet 8606f458dfbSEric Dumazet /* perform an atomic operation only if at least one flag is set */ 8616f458dfbSEric Dumazet do { 8627aa5470cSEric Dumazet flags = sk->sk_tsq_flags; 8636f458dfbSEric Dumazet if (!(flags & TCP_DEFERRED_ALL)) 8646f458dfbSEric Dumazet return; 8656f458dfbSEric Dumazet nflags = flags & ~TCP_DEFERRED_ALL; 8667aa5470cSEric Dumazet } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags); 8676f458dfbSEric Dumazet 86873a6bab5SEric Dumazet if (flags & TCPF_TSQ_DEFERRED) { 86973a6bab5SEric Dumazet tcp_tsq_write(sk); 87073a6bab5SEric Dumazet __sock_put(sk); 87173a6bab5SEric Dumazet } 872c3f9b018SEric Dumazet /* Here begins the tricky part : 873c3f9b018SEric Dumazet * We are called from release_sock() with : 874c3f9b018SEric Dumazet * 1) BH disabled 875c3f9b018SEric Dumazet * 2) sk_lock.slock spinlock held 876c3f9b018SEric Dumazet * 3) socket owned by us (sk->sk_lock.owned == 1) 877c3f9b018SEric Dumazet * 878c3f9b018SEric Dumazet * But following code is meant to be called from BH handlers, 879c3f9b018SEric Dumazet * so we should keep BH disabled, but early release socket ownership 880c3f9b018SEric Dumazet */ 881c3f9b018SEric Dumazet sock_release_ownership(sk); 882c3f9b018SEric Dumazet 88340fc3423SEric Dumazet if (flags & TCPF_WRITE_TIMER_DEFERRED) { 8846f458dfbSEric Dumazet tcp_write_timer_handler(sk); 885144d56e9SEric Dumazet __sock_put(sk); 886144d56e9SEric Dumazet } 88740fc3423SEric Dumazet if (flags & TCPF_DELACK_TIMER_DEFERRED) { 8886f458dfbSEric Dumazet tcp_delack_timer_handler(sk); 889144d56e9SEric Dumazet __sock_put(sk); 890144d56e9SEric Dumazet } 89140fc3423SEric Dumazet if (flags & TCPF_MTU_REDUCED_DEFERRED) { 8924fab9071SNeal Cardwell inet_csk(sk)->icsk_af_ops->mtu_reduced(sk); 893144d56e9SEric Dumazet __sock_put(sk); 894144d56e9SEric Dumazet } 89546d3ceabSEric Dumazet } 89646d3ceabSEric Dumazet EXPORT_SYMBOL(tcp_release_cb); 89746d3ceabSEric Dumazet 89846d3ceabSEric Dumazet void __init tcp_tasklet_init(void) 89946d3ceabSEric Dumazet { 90046d3ceabSEric Dumazet int i; 90146d3ceabSEric Dumazet 90246d3ceabSEric Dumazet for_each_possible_cpu(i) { 90346d3ceabSEric Dumazet struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i); 90446d3ceabSEric Dumazet 90546d3ceabSEric Dumazet INIT_LIST_HEAD(&tsq->head); 90646d3ceabSEric Dumazet tasklet_init(&tsq->tasklet, 90746d3ceabSEric Dumazet tcp_tasklet_func, 90846d3ceabSEric Dumazet (unsigned long)tsq); 90946d3ceabSEric Dumazet } 91046d3ceabSEric Dumazet } 91146d3ceabSEric Dumazet 91246d3ceabSEric Dumazet /* 91346d3ceabSEric Dumazet * Write buffer destructor automatically called from kfree_skb. 9148e3bff96Sstephen hemminger * We can't xmit new skbs from this context, as we might already 91546d3ceabSEric Dumazet * hold qdisc lock. 91646d3ceabSEric Dumazet */ 917d6a4a104SEric Dumazet void tcp_wfree(struct sk_buff *skb) 91846d3ceabSEric Dumazet { 91946d3ceabSEric Dumazet struct sock *sk = skb->sk; 92046d3ceabSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 921408f0a6cSEric Dumazet unsigned long flags, nval, oval; 9229b462d02SEric Dumazet 9239b462d02SEric Dumazet /* Keep one reference on sk_wmem_alloc. 9249b462d02SEric Dumazet * Will be released by sk_free() from here or tcp_tasklet_func() 9259b462d02SEric Dumazet */ 92614afee4bSReshetova, Elena WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc)); 9279b462d02SEric Dumazet 9289b462d02SEric Dumazet /* If this softirq is serviced by ksoftirqd, we are likely under stress. 9299b462d02SEric Dumazet * Wait until our queues (qdisc + devices) are drained. 9309b462d02SEric Dumazet * This gives : 9319b462d02SEric Dumazet * - less callbacks to tcp_write_xmit(), reducing stress (batches) 9329b462d02SEric Dumazet * - chance for incoming ACK (processed by another cpu maybe) 9339b462d02SEric Dumazet * to migrate this flow (skb->ooo_okay will be eventually set) 9349b462d02SEric Dumazet */ 93514afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current) 9369b462d02SEric Dumazet goto out; 93746d3ceabSEric Dumazet 9387aa5470cSEric Dumazet for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) { 93946d3ceabSEric Dumazet struct tsq_tasklet *tsq; 940a9b204d1SEric Dumazet bool empty; 94146d3ceabSEric Dumazet 942408f0a6cSEric Dumazet if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED)) 943408f0a6cSEric Dumazet goto out; 944408f0a6cSEric Dumazet 94573a6bab5SEric Dumazet nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED; 9467aa5470cSEric Dumazet nval = cmpxchg(&sk->sk_tsq_flags, oval, nval); 947408f0a6cSEric Dumazet if (nval != oval) 948408f0a6cSEric Dumazet continue; 949408f0a6cSEric Dumazet 95046d3ceabSEric Dumazet /* queue this socket to tasklet queue */ 95146d3ceabSEric Dumazet local_irq_save(flags); 952903ceff7SChristoph Lameter tsq = this_cpu_ptr(&tsq_tasklet); 953a9b204d1SEric Dumazet empty = list_empty(&tsq->head); 95446d3ceabSEric Dumazet list_add(&tp->tsq_node, &tsq->head); 955a9b204d1SEric Dumazet if (empty) 95646d3ceabSEric Dumazet tasklet_schedule(&tsq->tasklet); 95746d3ceabSEric Dumazet local_irq_restore(flags); 9589b462d02SEric Dumazet return; 95946d3ceabSEric Dumazet } 9609b462d02SEric Dumazet out: 9619b462d02SEric Dumazet sk_free(sk); 96246d3ceabSEric Dumazet } 96346d3ceabSEric Dumazet 96473a6bab5SEric Dumazet /* Note: Called under soft irq. 96573a6bab5SEric Dumazet * We can call TCP stack right away, unless socket is owned by user. 966218af599SEric Dumazet */ 967218af599SEric Dumazet enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer) 968218af599SEric Dumazet { 969218af599SEric Dumazet struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer); 970218af599SEric Dumazet struct sock *sk = (struct sock *)tp; 971218af599SEric Dumazet 97273a6bab5SEric Dumazet tcp_tsq_handler(sk); 97373a6bab5SEric Dumazet sock_put(sk); 974218af599SEric Dumazet 975218af599SEric Dumazet return HRTIMER_NORESTART; 976218af599SEric Dumazet } 977218af599SEric Dumazet 978c092dd5fSEric Dumazet static void tcp_internal_pacing(struct sock *sk) 979218af599SEric Dumazet { 980218af599SEric Dumazet if (!tcp_needs_internal_pacing(sk)) 981218af599SEric Dumazet return; 982218af599SEric Dumazet hrtimer_start(&tcp_sk(sk)->pacing_timer, 983c092dd5fSEric Dumazet ns_to_ktime(tcp_sk(sk)->tcp_wstamp_ns), 98473a6bab5SEric Dumazet HRTIMER_MODE_ABS_PINNED_SOFT); 98573a6bab5SEric Dumazet sock_hold(sk); 986218af599SEric Dumazet } 987218af599SEric Dumazet 988ab408b6dSEric Dumazet static void tcp_update_skb_after_send(struct sock *sk, struct sk_buff *skb) 989e2080072SEric Dumazet { 990ab408b6dSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 991ab408b6dSEric Dumazet 992d3edd06eSEric Dumazet skb->skb_mstamp_ns = tp->tcp_wstamp_ns; 993ab408b6dSEric Dumazet if (sk->sk_pacing_status != SK_PACING_NONE) { 994*76a9ebe8SEric Dumazet unsigned long rate = sk->sk_pacing_rate; 995ab408b6dSEric Dumazet 996ab408b6dSEric Dumazet /* Original sch_fq does not pace first 10 MSS 997ab408b6dSEric Dumazet * Note that tp->data_segs_out overflows after 2^32 packets, 998ab408b6dSEric Dumazet * this is a minor annoyance. 999ab408b6dSEric Dumazet */ 1000*76a9ebe8SEric Dumazet if (rate != ~0UL && rate && tp->data_segs_out >= 10) { 1001*76a9ebe8SEric Dumazet tp->tcp_wstamp_ns += div64_ul((u64)skb->len * NSEC_PER_SEC, rate); 1002c092dd5fSEric Dumazet 1003c092dd5fSEric Dumazet tcp_internal_pacing(sk); 1004ab408b6dSEric Dumazet } 1005ab408b6dSEric Dumazet } 1006e2080072SEric Dumazet list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue); 1007e2080072SEric Dumazet } 1008e2080072SEric Dumazet 10091da177e4SLinus Torvalds /* This routine actually transmits TCP packets queued in by 10101da177e4SLinus Torvalds * tcp_do_sendmsg(). This is used by both the initial 10111da177e4SLinus Torvalds * transmission and possible later retransmissions. 10121da177e4SLinus Torvalds * All SKB's seen here are completely headerless. It is our 10131da177e4SLinus Torvalds * job to build the TCP header, and pass the packet down to 10141da177e4SLinus Torvalds * IP so it can do the same plus pass the packet off to the 10151da177e4SLinus Torvalds * device. 10161da177e4SLinus Torvalds * 10171da177e4SLinus Torvalds * We are working here with either a clone of the original 10181da177e4SLinus Torvalds * SKB, or a fresh unique copy made by the retransmit engine. 10191da177e4SLinus Torvalds */ 10202987babbSYuchung Cheng static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, 10212987babbSYuchung Cheng int clone_it, gfp_t gfp_mask, u32 rcv_nxt) 10221da177e4SLinus Torvalds { 10236687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 1024dfb4b9dcSDavid S. Miller struct inet_sock *inet; 1025dfb4b9dcSDavid S. Miller struct tcp_sock *tp; 1026dfb4b9dcSDavid S. Miller struct tcp_skb_cb *tcb; 102733ad798cSAdam Langley struct tcp_out_options opts; 102895c96174SEric Dumazet unsigned int tcp_options_size, tcp_header_size; 10298c72c65bSEric Dumazet struct sk_buff *oskb = NULL; 1030cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 10311da177e4SLinus Torvalds struct tcphdr *th; 10321da177e4SLinus Torvalds int err; 10331da177e4SLinus Torvalds 1034dfb4b9dcSDavid S. Miller BUG_ON(!skb || !tcp_skb_pcount(skb)); 10356f094b9eSLawrence Brakmo tp = tcp_sk(sk); 1036dfb4b9dcSDavid S. Miller 1037ccdbb6e9SEric Dumazet if (clone_it) { 10386f094b9eSLawrence Brakmo TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq 10396f094b9eSLawrence Brakmo - tp->snd_una; 10408c72c65bSEric Dumazet oskb = skb; 1041e2080072SEric Dumazet 1042e2080072SEric Dumazet tcp_skb_tsorted_save(oskb) { 1043e2080072SEric Dumazet if (unlikely(skb_cloned(oskb))) 1044e2080072SEric Dumazet skb = pskb_copy(oskb, gfp_mask); 1045dfb4b9dcSDavid S. Miller else 1046e2080072SEric Dumazet skb = skb_clone(oskb, gfp_mask); 1047e2080072SEric Dumazet } tcp_skb_tsorted_restore(oskb); 1048e2080072SEric Dumazet 1049dfb4b9dcSDavid S. Miller if (unlikely(!skb)) 1050dfb4b9dcSDavid S. Miller return -ENOBUFS; 1051dfb4b9dcSDavid S. Miller } 10525f6188a8SEric Dumazet 10535f6188a8SEric Dumazet /* TODO: might take care of jitter here */ 10545f6188a8SEric Dumazet tp->tcp_wstamp_ns = max(tp->tcp_wstamp_ns, tp->tcp_clock_cache); 10555f6188a8SEric Dumazet 1056d3edd06eSEric Dumazet skb->skb_mstamp_ns = tp->tcp_wstamp_ns; 1057dfb4b9dcSDavid S. Miller 1058dfb4b9dcSDavid S. Miller inet = inet_sk(sk); 1059dfb4b9dcSDavid S. Miller tcb = TCP_SKB_CB(skb); 106033ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 10611da177e4SLinus Torvalds 10624de075e0SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) 106333ad798cSAdam Langley tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); 106433ad798cSAdam Langley else 106533ad798cSAdam Langley tcp_options_size = tcp_established_options(sk, skb, &opts, 106633ad798cSAdam Langley &md5); 106733ad798cSAdam Langley tcp_header_size = tcp_options_size + sizeof(struct tcphdr); 10681da177e4SLinus Torvalds 1069547669d4SEric Dumazet /* if no packet is in qdisc/device queue, then allow XPS to select 1070b2532eb9SEric Dumazet * another queue. We can be called from tcp_tsq_handler() 107173a6bab5SEric Dumazet * which holds one reference to sk. 1072b2532eb9SEric Dumazet * 1073b2532eb9SEric Dumazet * TODO: Ideally, in-flight pure ACK packets should not matter here. 1074b2532eb9SEric Dumazet * One way to get this would be to set skb->truesize = 2 on them. 1075547669d4SEric Dumazet */ 1076b2532eb9SEric Dumazet skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1); 10771da177e4SLinus Torvalds 107838ab52e8SEric Dumazet /* If we had to use memory reserve to allocate this skb, 107938ab52e8SEric Dumazet * this might cause drops if packet is looped back : 108038ab52e8SEric Dumazet * Other socket might not have SOCK_MEMALLOC. 108138ab52e8SEric Dumazet * Packets not looped back do not care about pfmemalloc. 108238ab52e8SEric Dumazet */ 108338ab52e8SEric Dumazet skb->pfmemalloc = 0; 108438ab52e8SEric Dumazet 1085aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 1086aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 108746d3ceabSEric Dumazet 108846d3ceabSEric Dumazet skb_orphan(skb); 108946d3ceabSEric Dumazet skb->sk = sk; 10901d2077acSEric Dumazet skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree; 1091b73c3d0eSTom Herbert skb_set_hash_from_sk(skb, sk); 109214afee4bSReshetova, Elena refcount_add(skb->truesize, &sk->sk_wmem_alloc); 10931da177e4SLinus Torvalds 1094c3a2e837SJulian Anastasov skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm); 1095c3a2e837SJulian Anastasov 10961da177e4SLinus Torvalds /* Build TCP header and checksum it. */ 1097ea1627c2SEric Dumazet th = (struct tcphdr *)skb->data; 1098c720c7e8SEric Dumazet th->source = inet->inet_sport; 1099c720c7e8SEric Dumazet th->dest = inet->inet_dport; 11001da177e4SLinus Torvalds th->seq = htonl(tcb->seq); 11012987babbSYuchung Cheng th->ack_seq = htonl(rcv_nxt); 1102df7a3b07SAl Viro *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | 11034de075e0SEric Dumazet tcb->tcp_flags); 1104dfb4b9dcSDavid S. Miller 11051da177e4SLinus Torvalds th->check = 0; 11061da177e4SLinus Torvalds th->urg_ptr = 0; 11071da177e4SLinus Torvalds 110833f5f57eSIlpo Järvinen /* The urg_mode check is necessary during a below snd_una win probe */ 11097691367dSHerbert Xu if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { 11107691367dSHerbert Xu if (before(tp->snd_up, tcb->seq + 0x10000)) { 11111da177e4SLinus Torvalds th->urg_ptr = htons(tp->snd_up - tcb->seq); 11121da177e4SLinus Torvalds th->urg = 1; 11137691367dSHerbert Xu } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { 11140eae88f3SEric Dumazet th->urg_ptr = htons(0xFFFF); 11157691367dSHerbert Xu th->urg = 1; 11167691367dSHerbert Xu } 11171da177e4SLinus Torvalds } 11181da177e4SLinus Torvalds 1119bd0388aeSWilliam Allen Simpson tcp_options_write((__be32 *)(th + 1), tp, &opts); 112051466a75SEric Dumazet skb_shinfo(skb)->gso_type = sk->sk_gso_type; 1121ea1627c2SEric Dumazet if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) { 1122ea1627c2SEric Dumazet th->window = htons(tcp_select_window(sk)); 1123ea1627c2SEric Dumazet tcp_ecn_send(sk, skb, th, tcp_header_size); 1124ea1627c2SEric Dumazet } else { 1125ea1627c2SEric Dumazet /* RFC1323: The window in SYN & SYN/ACK segments 1126ea1627c2SEric Dumazet * is never scaled. 1127ea1627c2SEric Dumazet */ 1128ea1627c2SEric Dumazet th->window = htons(min(tp->rcv_wnd, 65535U)); 1129ea1627c2SEric Dumazet } 1130cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 1131cfb6eeb4SYOSHIFUJI Hideaki /* Calculate the MD5 hash, as we have all we need now */ 1132cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 1133a465419bSEric Dumazet sk_nocaps_add(sk, NETIF_F_GSO_MASK); 1134bd0388aeSWilliam Allen Simpson tp->af_specific->calc_md5_hash(opts.hash_location, 113539f8e58eSEric Dumazet md5, sk, skb); 1136cfb6eeb4SYOSHIFUJI Hideaki } 1137cfb6eeb4SYOSHIFUJI Hideaki #endif 1138cfb6eeb4SYOSHIFUJI Hideaki 1139bb296246SHerbert Xu icsk->icsk_af_ops->send_check(sk, skb); 11401da177e4SLinus Torvalds 11414de075e0SEric Dumazet if (likely(tcb->tcp_flags & TCPHDR_ACK)) 114227cde44aSYuchung Cheng tcp_event_ack_sent(sk, tcp_skb_pcount(skb), rcv_nxt); 11431da177e4SLinus Torvalds 1144a44d6eacSMartin KaFai Lau if (skb->len != tcp_header_size) { 1145cf533ea5SEric Dumazet tcp_event_data_sent(tp, sk); 1146a44d6eacSMartin KaFai Lau tp->data_segs_out += tcp_skb_pcount(skb); 1147ba113c3aSWei Wang tp->bytes_sent += skb->len - tcp_header_size; 1148a44d6eacSMartin KaFai Lau } 11491da177e4SLinus Torvalds 1150bd37a088SWei Yongjun if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) 1151aa2ea058STom Herbert TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, 1152aa2ea058STom Herbert tcp_skb_pcount(skb)); 11531da177e4SLinus Torvalds 11542efd055cSMarcelo Ricardo Leitner tp->segs_out += tcp_skb_pcount(skb); 1155f69ad292SEric Dumazet /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */ 1156cd7d8498SEric Dumazet skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb); 1157f69ad292SEric Dumazet skb_shinfo(skb)->gso_size = tcp_skb_mss(skb); 1158cd7d8498SEric Dumazet 1159d3edd06eSEric Dumazet /* Leave earliest departure time in skb->tstamp (skb->skb_mstamp_ns) */ 1160971f10ecSEric Dumazet 1161971f10ecSEric Dumazet /* Cleanup our debris for IP stacks */ 1162971f10ecSEric Dumazet memset(skb->cb, 0, max(sizeof(struct inet_skb_parm), 1163971f10ecSEric Dumazet sizeof(struct inet6_skb_parm))); 1164971f10ecSEric Dumazet 1165b0270e91SEric Dumazet err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl); 11667faee5c0SEric Dumazet 11678c72c65bSEric Dumazet if (unlikely(err > 0)) { 11685ee2c941SChristoph Paasch tcp_enter_cwr(sk); 11698c72c65bSEric Dumazet err = net_xmit_eval(err); 11708c72c65bSEric Dumazet } 1171fc225799SEric Dumazet if (!err && oskb) { 1172ab408b6dSEric Dumazet tcp_update_skb_after_send(sk, oskb); 1173fc225799SEric Dumazet tcp_rate_skb_sent(sk, oskb); 1174fc225799SEric Dumazet } 11758c72c65bSEric Dumazet return err; 11761da177e4SLinus Torvalds } 11771da177e4SLinus Torvalds 11782987babbSYuchung Cheng static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, 11792987babbSYuchung Cheng gfp_t gfp_mask) 11802987babbSYuchung Cheng { 11812987babbSYuchung Cheng return __tcp_transmit_skb(sk, skb, clone_it, gfp_mask, 11822987babbSYuchung Cheng tcp_sk(sk)->rcv_nxt); 11832987babbSYuchung Cheng } 11842987babbSYuchung Cheng 118567edfef7SAndi Kleen /* This routine just queues the buffer for sending. 11861da177e4SLinus Torvalds * 11871da177e4SLinus Torvalds * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, 11881da177e4SLinus Torvalds * otherwise socket can stall. 11891da177e4SLinus Torvalds */ 11901da177e4SLinus Torvalds static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) 11911da177e4SLinus Torvalds { 11921da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 11931da177e4SLinus Torvalds 11941da177e4SLinus Torvalds /* Advance write_seq and place onto the write_queue. */ 11951da177e4SLinus Torvalds tp->write_seq = TCP_SKB_CB(skb)->end_seq; 1196f4a775d1SEric Dumazet __skb_header_release(skb); 1197fe067e8aSDavid S. Miller tcp_add_write_queue_tail(sk, skb); 11983ab224beSHideo Aoki sk->sk_wmem_queued += skb->truesize; 11993ab224beSHideo Aoki sk_mem_charge(sk, skb->truesize); 12001da177e4SLinus Torvalds } 12011da177e4SLinus Torvalds 120267edfef7SAndi Kleen /* Initialize TSO segments for a packet. */ 12035bbb432cSEric Dumazet static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now) 1204f6302d1dSDavid S. Miller { 12054a64fd6cSEric Dumazet if (skb->len <= mss_now) { 1206f6302d1dSDavid S. Miller /* Avoid the costly divide in the normal 1207f6302d1dSDavid S. Miller * non-TSO case. 1208f6302d1dSDavid S. Miller */ 1209cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, 1); 1210f69ad292SEric Dumazet TCP_SKB_CB(skb)->tcp_gso_size = 0; 1211f6302d1dSDavid S. Miller } else { 1212cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now)); 1213f69ad292SEric Dumazet TCP_SKB_CB(skb)->tcp_gso_size = mss_now; 12141da177e4SLinus Torvalds } 12151da177e4SLinus Torvalds } 12161da177e4SLinus Torvalds 1217797108d1SIlpo Järvinen /* Pcount in the middle of the write queue got changed, we need to do various 1218797108d1SIlpo Järvinen * tweaks to fix counters 1219797108d1SIlpo Järvinen */ 1220cf533ea5SEric Dumazet static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) 1221797108d1SIlpo Järvinen { 1222797108d1SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1223797108d1SIlpo Järvinen 1224797108d1SIlpo Järvinen tp->packets_out -= decr; 1225797108d1SIlpo Järvinen 1226797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 1227797108d1SIlpo Järvinen tp->sacked_out -= decr; 1228797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 1229797108d1SIlpo Järvinen tp->retrans_out -= decr; 1230797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 1231797108d1SIlpo Järvinen tp->lost_out -= decr; 1232797108d1SIlpo Järvinen 1233797108d1SIlpo Järvinen /* Reno case is special. Sigh... */ 1234797108d1SIlpo Järvinen if (tcp_is_reno(tp) && decr > 0) 1235797108d1SIlpo Järvinen tp->sacked_out -= min_t(u32, tp->sacked_out, decr); 1236797108d1SIlpo Järvinen 1237797108d1SIlpo Järvinen if (tp->lost_skb_hint && 1238797108d1SIlpo Järvinen before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && 1239713bafeaSYuchung Cheng (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) 1240797108d1SIlpo Järvinen tp->lost_cnt_hint -= decr; 1241797108d1SIlpo Järvinen 1242797108d1SIlpo Järvinen tcp_verify_left_out(tp); 1243797108d1SIlpo Järvinen } 1244797108d1SIlpo Järvinen 12450a2cf20cSSoheil Hassas Yeganeh static bool tcp_has_tx_tstamp(const struct sk_buff *skb) 12460a2cf20cSSoheil Hassas Yeganeh { 12470a2cf20cSSoheil Hassas Yeganeh return TCP_SKB_CB(skb)->txstamp_ack || 12480a2cf20cSSoheil Hassas Yeganeh (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP); 12490a2cf20cSSoheil Hassas Yeganeh } 12500a2cf20cSSoheil Hassas Yeganeh 1251490cc7d0SWillem de Bruijn static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2) 1252490cc7d0SWillem de Bruijn { 1253490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo = skb_shinfo(skb); 1254490cc7d0SWillem de Bruijn 12550a2cf20cSSoheil Hassas Yeganeh if (unlikely(tcp_has_tx_tstamp(skb)) && 1256490cc7d0SWillem de Bruijn !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) { 1257490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo2 = skb_shinfo(skb2); 1258490cc7d0SWillem de Bruijn u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP; 1259490cc7d0SWillem de Bruijn 1260490cc7d0SWillem de Bruijn shinfo->tx_flags &= ~tsflags; 1261490cc7d0SWillem de Bruijn shinfo2->tx_flags |= tsflags; 1262490cc7d0SWillem de Bruijn swap(shinfo->tskey, shinfo2->tskey); 1263b51e13faSMartin KaFai Lau TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack; 1264b51e13faSMartin KaFai Lau TCP_SKB_CB(skb)->txstamp_ack = 0; 1265490cc7d0SWillem de Bruijn } 1266490cc7d0SWillem de Bruijn } 1267490cc7d0SWillem de Bruijn 1268a166140eSMartin KaFai Lau static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2) 1269a166140eSMartin KaFai Lau { 1270a166140eSMartin KaFai Lau TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor; 1271a166140eSMartin KaFai Lau TCP_SKB_CB(skb)->eor = 0; 1272a166140eSMartin KaFai Lau } 1273a166140eSMartin KaFai Lau 127475c119afSEric Dumazet /* Insert buff after skb on the write or rtx queue of sk. */ 127575c119afSEric Dumazet static void tcp_insert_write_queue_after(struct sk_buff *skb, 127675c119afSEric Dumazet struct sk_buff *buff, 127775c119afSEric Dumazet struct sock *sk, 127875c119afSEric Dumazet enum tcp_queue tcp_queue) 127975c119afSEric Dumazet { 128075c119afSEric Dumazet if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE) 128175c119afSEric Dumazet __skb_queue_after(&sk->sk_write_queue, skb, buff); 128275c119afSEric Dumazet else 128375c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, buff); 128475c119afSEric Dumazet } 128575c119afSEric Dumazet 12861da177e4SLinus Torvalds /* Function to create two new TCP segments. Shrinks the given segment 12871da177e4SLinus Torvalds * to the specified size and appends a new segment with the rest of the 12881da177e4SLinus Torvalds * packet to the list. This won't be called frequently, I hope. 12891da177e4SLinus Torvalds * Remember, these are still headerless SKBs at this point. 12901da177e4SLinus Torvalds */ 129175c119afSEric Dumazet int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue, 129275c119afSEric Dumazet struct sk_buff *skb, u32 len, 12936cc55e09SOctavian Purdila unsigned int mss_now, gfp_t gfp) 12941da177e4SLinus Torvalds { 12951da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 12961da177e4SLinus Torvalds struct sk_buff *buff; 12976475be16SDavid S. Miller int nsize, old_factor; 1298b60b49eaSHerbert Xu int nlen; 12999ce01461SIlpo Järvinen u8 flags; 13001da177e4SLinus Torvalds 13012fceec13SIlpo Järvinen if (WARN_ON(len > skb->len)) 13022fceec13SIlpo Järvinen return -EINVAL; 13036a438bbeSStephen Hemminger 13041da177e4SLinus Torvalds nsize = skb_headlen(skb) - len; 13051da177e4SLinus Torvalds if (nsize < 0) 13061da177e4SLinus Torvalds nsize = 0; 13071da177e4SLinus Torvalds 13086cc55e09SOctavian Purdila if (skb_unclone(skb, gfp)) 13091da177e4SLinus Torvalds return -ENOMEM; 13101da177e4SLinus Torvalds 13111da177e4SLinus Torvalds /* Get a new skb... force flag on. */ 1312eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, nsize, gfp, true); 131351456b29SIan Morris if (!buff) 13141da177e4SLinus Torvalds return -ENOMEM; /* We'll just try again later. */ 1315ef5cb973SHerbert Xu 13163ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 13173ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1318b60b49eaSHerbert Xu nlen = skb->len - len - nsize; 1319b60b49eaSHerbert Xu buff->truesize += nlen; 1320b60b49eaSHerbert Xu skb->truesize -= nlen; 13211da177e4SLinus Torvalds 13221da177e4SLinus Torvalds /* Correct the sequence numbers. */ 13231da177e4SLinus Torvalds TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 13241da177e4SLinus Torvalds TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 13251da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 13261da177e4SLinus Torvalds 13271da177e4SLinus Torvalds /* PSH and FIN should only be set in the second packet. */ 13284de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 13294de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 13304de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1331e14c3cafSHerbert Xu TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; 1332a166140eSMartin KaFai Lau tcp_skb_fragment_eor(skb, buff); 13331da177e4SLinus Torvalds 13341da177e4SLinus Torvalds skb_split(skb, buff, len); 13351da177e4SLinus Torvalds 133698be9b12SEric Dumazet buff->ip_summed = CHECKSUM_PARTIAL; 13371da177e4SLinus Torvalds 1338a61bbcf2SPatrick McHardy buff->tstamp = skb->tstamp; 1339490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 13401da177e4SLinus Torvalds 13416475be16SDavid S. Miller old_factor = tcp_skb_pcount(skb); 13426475be16SDavid S. Miller 13431da177e4SLinus Torvalds /* Fix up tso_factor for both original and new SKB. */ 13445bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 13455bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 13461da177e4SLinus Torvalds 1347b9f64820SYuchung Cheng /* Update delivered info for the new segment */ 1348b9f64820SYuchung Cheng TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx; 1349b9f64820SYuchung Cheng 13506475be16SDavid S. Miller /* If this packet has been sent out already, we must 13516475be16SDavid S. Miller * adjust the various packet counters. 13526475be16SDavid S. Miller */ 1353cf0b450cSHerbert Xu if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 13546475be16SDavid S. Miller int diff = old_factor - tcp_skb_pcount(skb) - 13556475be16SDavid S. Miller tcp_skb_pcount(buff); 13561da177e4SLinus Torvalds 1357797108d1SIlpo Järvinen if (diff) 1358797108d1SIlpo Järvinen tcp_adjust_pcount(sk, skb, diff); 13591da177e4SLinus Torvalds } 13601da177e4SLinus Torvalds 13611da177e4SLinus Torvalds /* Link BUFF into the send queue. */ 1362f4a775d1SEric Dumazet __skb_header_release(buff); 136375c119afSEric Dumazet tcp_insert_write_queue_after(skb, buff, sk, tcp_queue); 1364f67971e6SEric Dumazet if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE) 1365e2080072SEric Dumazet list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor); 13661da177e4SLinus Torvalds 13671da177e4SLinus Torvalds return 0; 13681da177e4SLinus Torvalds } 13691da177e4SLinus Torvalds 1370f4d01666SEric Dumazet /* This is similar to __pskb_pull_tail(). The difference is that pulled 1371f4d01666SEric Dumazet * data is not copied, but immediately discarded. 13721da177e4SLinus Torvalds */ 13737162fb24SEric Dumazet static int __pskb_trim_head(struct sk_buff *skb, int len) 13741da177e4SLinus Torvalds { 13757b7fc97aSEric Dumazet struct skb_shared_info *shinfo; 13761da177e4SLinus Torvalds int i, k, eat; 13771da177e4SLinus Torvalds 13784fa48bf3SEric Dumazet eat = min_t(int, len, skb_headlen(skb)); 13794fa48bf3SEric Dumazet if (eat) { 13804fa48bf3SEric Dumazet __skb_pull(skb, eat); 13814fa48bf3SEric Dumazet len -= eat; 13824fa48bf3SEric Dumazet if (!len) 13837162fb24SEric Dumazet return 0; 13844fa48bf3SEric Dumazet } 13851da177e4SLinus Torvalds eat = len; 13861da177e4SLinus Torvalds k = 0; 13877b7fc97aSEric Dumazet shinfo = skb_shinfo(skb); 13887b7fc97aSEric Dumazet for (i = 0; i < shinfo->nr_frags; i++) { 13897b7fc97aSEric Dumazet int size = skb_frag_size(&shinfo->frags[i]); 13909e903e08SEric Dumazet 13919e903e08SEric Dumazet if (size <= eat) { 1392aff65da0SIan Campbell skb_frag_unref(skb, i); 13939e903e08SEric Dumazet eat -= size; 13941da177e4SLinus Torvalds } else { 13957b7fc97aSEric Dumazet shinfo->frags[k] = shinfo->frags[i]; 13961da177e4SLinus Torvalds if (eat) { 13977b7fc97aSEric Dumazet shinfo->frags[k].page_offset += eat; 13987b7fc97aSEric Dumazet skb_frag_size_sub(&shinfo->frags[k], eat); 13991da177e4SLinus Torvalds eat = 0; 14001da177e4SLinus Torvalds } 14011da177e4SLinus Torvalds k++; 14021da177e4SLinus Torvalds } 14031da177e4SLinus Torvalds } 14047b7fc97aSEric Dumazet shinfo->nr_frags = k; 14051da177e4SLinus Torvalds 14061da177e4SLinus Torvalds skb->data_len -= len; 14071da177e4SLinus Torvalds skb->len = skb->data_len; 14087162fb24SEric Dumazet return len; 14091da177e4SLinus Torvalds } 14101da177e4SLinus Torvalds 141167edfef7SAndi Kleen /* Remove acked data from a packet in the transmit queue. */ 14121da177e4SLinus Torvalds int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 14131da177e4SLinus Torvalds { 14147162fb24SEric Dumazet u32 delta_truesize; 14157162fb24SEric Dumazet 141614bbd6a5SPravin B Shelar if (skb_unclone(skb, GFP_ATOMIC)) 14171da177e4SLinus Torvalds return -ENOMEM; 14181da177e4SLinus Torvalds 14197162fb24SEric Dumazet delta_truesize = __pskb_trim_head(skb, len); 14201da177e4SLinus Torvalds 14211da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq += len; 142284fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 14231da177e4SLinus Torvalds 14247162fb24SEric Dumazet if (delta_truesize) { 14257162fb24SEric Dumazet skb->truesize -= delta_truesize; 14267162fb24SEric Dumazet sk->sk_wmem_queued -= delta_truesize; 14277162fb24SEric Dumazet sk_mem_uncharge(sk, delta_truesize); 14281da177e4SLinus Torvalds sock_set_flag(sk, SOCK_QUEUE_SHRUNK); 14297162fb24SEric Dumazet } 14301da177e4SLinus Torvalds 14315b35e1e6SNeal Cardwell /* Any change of skb->len requires recalculation of tso factor. */ 14321da177e4SLinus Torvalds if (tcp_skb_pcount(skb) > 1) 14335bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb)); 14341da177e4SLinus Torvalds 14351da177e4SLinus Torvalds return 0; 14361da177e4SLinus Torvalds } 14371da177e4SLinus Torvalds 14381b63edd6SYuchung Cheng /* Calculate MSS not accounting any TCP options. */ 14391b63edd6SYuchung Cheng static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu) 14405d424d5aSJohn Heffner { 1441cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1442cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 14435d424d5aSJohn Heffner int mss_now; 14445d424d5aSJohn Heffner 14455d424d5aSJohn Heffner /* Calculate base mss without TCP options: 14465d424d5aSJohn Heffner It is MMS_S - sizeof(tcphdr) of rfc1122 14475d424d5aSJohn Heffner */ 14485d424d5aSJohn Heffner mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 14495d424d5aSJohn Heffner 145067469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 145167469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 145267469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 145367469601SEric Dumazet 145467469601SEric Dumazet if (dst && dst_allfrag(dst)) 145567469601SEric Dumazet mss_now -= icsk->icsk_af_ops->net_frag_header_len; 145667469601SEric Dumazet } 145767469601SEric Dumazet 14585d424d5aSJohn Heffner /* Clamp it (mss_clamp does not include tcp options) */ 14595d424d5aSJohn Heffner if (mss_now > tp->rx_opt.mss_clamp) 14605d424d5aSJohn Heffner mss_now = tp->rx_opt.mss_clamp; 14615d424d5aSJohn Heffner 14625d424d5aSJohn Heffner /* Now subtract optional transport overhead */ 14635d424d5aSJohn Heffner mss_now -= icsk->icsk_ext_hdr_len; 14645d424d5aSJohn Heffner 14655d424d5aSJohn Heffner /* Then reserve room for full set of TCP options and 8 bytes of data */ 14665d424d5aSJohn Heffner if (mss_now < 48) 14675d424d5aSJohn Heffner mss_now = 48; 14685d424d5aSJohn Heffner return mss_now; 14695d424d5aSJohn Heffner } 14705d424d5aSJohn Heffner 14711b63edd6SYuchung Cheng /* Calculate MSS. Not accounting for SACKs here. */ 14721b63edd6SYuchung Cheng int tcp_mtu_to_mss(struct sock *sk, int pmtu) 14731b63edd6SYuchung Cheng { 14741b63edd6SYuchung Cheng /* Subtract TCP options size, not including SACKs */ 14751b63edd6SYuchung Cheng return __tcp_mtu_to_mss(sk, pmtu) - 14761b63edd6SYuchung Cheng (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr)); 14771b63edd6SYuchung Cheng } 14781b63edd6SYuchung Cheng 14795d424d5aSJohn Heffner /* Inverse of above */ 148067469601SEric Dumazet int tcp_mss_to_mtu(struct sock *sk, int mss) 14815d424d5aSJohn Heffner { 1482cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1483cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 14845d424d5aSJohn Heffner int mtu; 14855d424d5aSJohn Heffner 14865d424d5aSJohn Heffner mtu = mss + 14875d424d5aSJohn Heffner tp->tcp_header_len + 14885d424d5aSJohn Heffner icsk->icsk_ext_hdr_len + 14895d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 14905d424d5aSJohn Heffner 149167469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 149267469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 149367469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 149467469601SEric Dumazet 149567469601SEric Dumazet if (dst && dst_allfrag(dst)) 149667469601SEric Dumazet mtu += icsk->icsk_af_ops->net_frag_header_len; 149767469601SEric Dumazet } 14985d424d5aSJohn Heffner return mtu; 14995d424d5aSJohn Heffner } 1500556c6b46SNeal Cardwell EXPORT_SYMBOL(tcp_mss_to_mtu); 15015d424d5aSJohn Heffner 150267edfef7SAndi Kleen /* MTU probing init per socket */ 15035d424d5aSJohn Heffner void tcp_mtup_init(struct sock *sk) 15045d424d5aSJohn Heffner { 15055d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 15065d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 1507b0f9ca53SFan Du struct net *net = sock_net(sk); 15085d424d5aSJohn Heffner 1509b0f9ca53SFan Du icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1; 15105d424d5aSJohn Heffner icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 15115d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 1512b0f9ca53SFan Du icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss); 15135d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 151405cbc0dbSFan Du if (icsk->icsk_mtup.enabled) 1515c74df29aSEric Dumazet icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; 15165d424d5aSJohn Heffner } 15174bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_mtup_init); 15185d424d5aSJohn Heffner 15191da177e4SLinus Torvalds /* This function synchronize snd mss to current pmtu/exthdr set. 15201da177e4SLinus Torvalds 15211da177e4SLinus Torvalds tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 15221da177e4SLinus Torvalds for TCP options, but includes only bare TCP header. 15231da177e4SLinus Torvalds 15241da177e4SLinus Torvalds tp->rx_opt.mss_clamp is mss negotiated at connection setup. 1525caa20d9aSStephen Hemminger It is minimum of user_mss and mss received with SYN. 15261da177e4SLinus Torvalds It also does not include TCP options. 15271da177e4SLinus Torvalds 1528d83d8461SArnaldo Carvalho de Melo inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. 15291da177e4SLinus Torvalds 15301da177e4SLinus Torvalds tp->mss_cache is current effective sending mss, including 15311da177e4SLinus Torvalds all tcp options except for SACKs. It is evaluated, 15321da177e4SLinus Torvalds taking into account current pmtu, but never exceeds 15331da177e4SLinus Torvalds tp->rx_opt.mss_clamp. 15341da177e4SLinus Torvalds 15351da177e4SLinus Torvalds NOTE1. rfc1122 clearly states that advertised MSS 15361da177e4SLinus Torvalds DOES NOT include either tcp or ip options. 15371da177e4SLinus Torvalds 1538d83d8461SArnaldo Carvalho de Melo NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache 1539d83d8461SArnaldo Carvalho de Melo are READ ONLY outside this function. --ANK (980731) 15401da177e4SLinus Torvalds */ 15411da177e4SLinus Torvalds unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 15421da177e4SLinus Torvalds { 15431da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1544d83d8461SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 15455d424d5aSJohn Heffner int mss_now; 15461da177e4SLinus Torvalds 15475d424d5aSJohn Heffner if (icsk->icsk_mtup.search_high > pmtu) 15485d424d5aSJohn Heffner icsk->icsk_mtup.search_high = pmtu; 15491da177e4SLinus Torvalds 15505d424d5aSJohn Heffner mss_now = tcp_mtu_to_mss(sk, pmtu); 1551409d22b4SIlpo Järvinen mss_now = tcp_bound_to_half_wnd(tp, mss_now); 15521da177e4SLinus Torvalds 15531da177e4SLinus Torvalds /* And store cached results */ 1554d83d8461SArnaldo Carvalho de Melo icsk->icsk_pmtu_cookie = pmtu; 15555d424d5aSJohn Heffner if (icsk->icsk_mtup.enabled) 15565d424d5aSJohn Heffner mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); 1557c1b4a7e6SDavid S. Miller tp->mss_cache = mss_now; 15581da177e4SLinus Torvalds 15591da177e4SLinus Torvalds return mss_now; 15601da177e4SLinus Torvalds } 15614bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_sync_mss); 15621da177e4SLinus Torvalds 15631da177e4SLinus Torvalds /* Compute the current effective MSS, taking SACKs and IP options, 15641da177e4SLinus Torvalds * and even PMTU discovery events into account. 15651da177e4SLinus Torvalds */ 15660c54b85fSIlpo Järvinen unsigned int tcp_current_mss(struct sock *sk) 15671da177e4SLinus Torvalds { 1568cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1569cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1570c1b4a7e6SDavid S. Miller u32 mss_now; 157195c96174SEric Dumazet unsigned int header_len; 157233ad798cSAdam Langley struct tcp_out_options opts; 157333ad798cSAdam Langley struct tcp_md5sig_key *md5; 15741da177e4SLinus Torvalds 1575c1b4a7e6SDavid S. Miller mss_now = tp->mss_cache; 1576c1b4a7e6SDavid S. Miller 15771da177e4SLinus Torvalds if (dst) { 15781da177e4SLinus Torvalds u32 mtu = dst_mtu(dst); 1579d83d8461SArnaldo Carvalho de Melo if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 15801da177e4SLinus Torvalds mss_now = tcp_sync_mss(sk, mtu); 15811da177e4SLinus Torvalds } 15821da177e4SLinus Torvalds 158333ad798cSAdam Langley header_len = tcp_established_options(sk, NULL, &opts, &md5) + 158433ad798cSAdam Langley sizeof(struct tcphdr); 158533ad798cSAdam Langley /* The mss_cache is sized based on tp->tcp_header_len, which assumes 158633ad798cSAdam Langley * some common options. If this is an odd packet (because we have SACK 158733ad798cSAdam Langley * blocks etc) then our calculated header_len will be different, and 158833ad798cSAdam Langley * we have to adjust mss_now correspondingly */ 158933ad798cSAdam Langley if (header_len != tp->tcp_header_len) { 159033ad798cSAdam Langley int delta = (int) header_len - tp->tcp_header_len; 159133ad798cSAdam Langley mss_now -= delta; 159233ad798cSAdam Langley } 1593cfb6eeb4SYOSHIFUJI Hideaki 15941da177e4SLinus Torvalds return mss_now; 15951da177e4SLinus Torvalds } 15961da177e4SLinus Torvalds 159786fd14adSWeiping Pan /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. 159886fd14adSWeiping Pan * As additional protections, we do not touch cwnd in retransmission phases, 159986fd14adSWeiping Pan * and if application hit its sndbuf limit recently. 160086fd14adSWeiping Pan */ 160186fd14adSWeiping Pan static void tcp_cwnd_application_limited(struct sock *sk) 1602a762a980SDavid S. Miller { 16039e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1604a762a980SDavid S. Miller 160586fd14adSWeiping Pan if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && 160686fd14adSWeiping Pan sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 160786fd14adSWeiping Pan /* Limited by application or receiver window. */ 160886fd14adSWeiping Pan u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); 160986fd14adSWeiping Pan u32 win_used = max(tp->snd_cwnd_used, init_win); 161086fd14adSWeiping Pan if (win_used < tp->snd_cwnd) { 161186fd14adSWeiping Pan tp->snd_ssthresh = tcp_current_ssthresh(sk); 161286fd14adSWeiping Pan tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1; 161386fd14adSWeiping Pan } 161486fd14adSWeiping Pan tp->snd_cwnd_used = 0; 161586fd14adSWeiping Pan } 1616c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 161786fd14adSWeiping Pan } 161886fd14adSWeiping Pan 1619ca8a2263SNeal Cardwell static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited) 1620a762a980SDavid S. Miller { 16211b1fc3fdSWei Wang const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 1622a762a980SDavid S. Miller struct tcp_sock *tp = tcp_sk(sk); 1623a762a980SDavid S. Miller 1624ca8a2263SNeal Cardwell /* Track the maximum number of outstanding packets in each 1625ca8a2263SNeal Cardwell * window, and remember whether we were cwnd-limited then. 1626ca8a2263SNeal Cardwell */ 1627ca8a2263SNeal Cardwell if (!before(tp->snd_una, tp->max_packets_seq) || 1628ca8a2263SNeal Cardwell tp->packets_out > tp->max_packets_out) { 1629ca8a2263SNeal Cardwell tp->max_packets_out = tp->packets_out; 1630ca8a2263SNeal Cardwell tp->max_packets_seq = tp->snd_nxt; 1631ca8a2263SNeal Cardwell tp->is_cwnd_limited = is_cwnd_limited; 1632ca8a2263SNeal Cardwell } 1633e114a710SEric Dumazet 163424901551SEric Dumazet if (tcp_is_cwnd_limited(sk)) { 1635a762a980SDavid S. Miller /* Network is feed fully. */ 1636a762a980SDavid S. Miller tp->snd_cwnd_used = 0; 1637c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 1638a762a980SDavid S. Miller } else { 1639a762a980SDavid S. Miller /* Network starves. */ 1640a762a980SDavid S. Miller if (tp->packets_out > tp->snd_cwnd_used) 1641a762a980SDavid S. Miller tp->snd_cwnd_used = tp->packets_out; 1642a762a980SDavid S. Miller 1643b510f0d2SEric Dumazet if (sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle && 1644c2203cf7SEric Dumazet (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto && 16451b1fc3fdSWei Wang !ca_ops->cong_control) 1646a762a980SDavid S. Miller tcp_cwnd_application_limited(sk); 1647b0f71bd3SFrancis Yan 1648b0f71bd3SFrancis Yan /* The following conditions together indicate the starvation 1649b0f71bd3SFrancis Yan * is caused by insufficient sender buffer: 1650b0f71bd3SFrancis Yan * 1) just sent some data (see tcp_write_xmit) 1651b0f71bd3SFrancis Yan * 2) not cwnd limited (this else condition) 165275c119afSEric Dumazet * 3) no more data to send (tcp_write_queue_empty()) 1653b0f71bd3SFrancis Yan * 4) application is hitting buffer limit (SOCK_NOSPACE) 1654b0f71bd3SFrancis Yan */ 165575c119afSEric Dumazet if (tcp_write_queue_empty(sk) && sk->sk_socket && 1656b0f71bd3SFrancis Yan test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) && 1657b0f71bd3SFrancis Yan (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) 1658b0f71bd3SFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED); 1659a762a980SDavid S. Miller } 1660a762a980SDavid S. Miller } 1661a762a980SDavid S. Miller 1662d4589926SEric Dumazet /* Minshall's variant of the Nagle send check. */ 1663d4589926SEric Dumazet static bool tcp_minshall_check(const struct tcp_sock *tp) 1664d4589926SEric Dumazet { 1665d4589926SEric Dumazet return after(tp->snd_sml, tp->snd_una) && 1666d4589926SEric Dumazet !after(tp->snd_sml, tp->snd_nxt); 1667d4589926SEric Dumazet } 1668d4589926SEric Dumazet 1669d4589926SEric Dumazet /* Update snd_sml if this skb is under mss 1670d4589926SEric Dumazet * Note that a TSO packet might end with a sub-mss segment 1671d4589926SEric Dumazet * The test is really : 1672d4589926SEric Dumazet * if ((skb->len % mss) != 0) 1673d4589926SEric Dumazet * tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1674d4589926SEric Dumazet * But we can avoid doing the divide again given we already have 1675d4589926SEric Dumazet * skb_pcount = skb->len / mss_now 16760e3a4803SIlpo Järvinen */ 1677d4589926SEric Dumazet static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now, 1678d4589926SEric Dumazet const struct sk_buff *skb) 1679d4589926SEric Dumazet { 1680d4589926SEric Dumazet if (skb->len < tcp_skb_pcount(skb) * mss_now) 1681d4589926SEric Dumazet tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1682d4589926SEric Dumazet } 1683d4589926SEric Dumazet 1684d4589926SEric Dumazet /* Return false, if packet can be sent now without violation Nagle's rules: 1685d4589926SEric Dumazet * 1. It is full sized. (provided by caller in %partial bool) 1686d4589926SEric Dumazet * 2. Or it contains FIN. (already checked by caller) 1687d4589926SEric Dumazet * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. 1688d4589926SEric Dumazet * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 1689d4589926SEric Dumazet * With Minshall's modification: all sent small packets are ACKed. 1690d4589926SEric Dumazet */ 1691d4589926SEric Dumazet static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp, 1692cc93fc51SPeter Pan(潘卫平) int nonagle) 1693d4589926SEric Dumazet { 1694d4589926SEric Dumazet return partial && 1695d4589926SEric Dumazet ((nonagle & TCP_NAGLE_CORK) || 1696d4589926SEric Dumazet (!nonagle && tp->packets_out && tcp_minshall_check(tp))); 1697d4589926SEric Dumazet } 1698605ad7f1SEric Dumazet 1699605ad7f1SEric Dumazet /* Return how many segs we'd like on a TSO packet, 1700605ad7f1SEric Dumazet * to send one TSO packet per ms 1701605ad7f1SEric Dumazet */ 1702dcb8c9b4SEric Dumazet static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now, 17031b3878caSNeal Cardwell int min_tso_segs) 1704605ad7f1SEric Dumazet { 1705605ad7f1SEric Dumazet u32 bytes, segs; 1706605ad7f1SEric Dumazet 1707*76a9ebe8SEric Dumazet bytes = min_t(unsigned long, 1708*76a9ebe8SEric Dumazet sk->sk_pacing_rate >> sk->sk_pacing_shift, 1709605ad7f1SEric Dumazet sk->sk_gso_max_size - 1 - MAX_TCP_HEADER); 1710605ad7f1SEric Dumazet 1711605ad7f1SEric Dumazet /* Goal is to send at least one packet per ms, 1712605ad7f1SEric Dumazet * not one big TSO packet every 100 ms. 1713605ad7f1SEric Dumazet * This preserves ACK clocking and is consistent 1714605ad7f1SEric Dumazet * with tcp_tso_should_defer() heuristic. 1715605ad7f1SEric Dumazet */ 17161b3878caSNeal Cardwell segs = max_t(u32, bytes / mss_now, min_tso_segs); 1717605ad7f1SEric Dumazet 1718350c9f48SEric Dumazet return segs; 1719605ad7f1SEric Dumazet } 1720605ad7f1SEric Dumazet 1721ed6e7268SNeal Cardwell /* Return the number of segments we want in the skb we are transmitting. 1722ed6e7268SNeal Cardwell * See if congestion control module wants to decide; otherwise, autosize. 1723ed6e7268SNeal Cardwell */ 1724ed6e7268SNeal Cardwell static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now) 1725ed6e7268SNeal Cardwell { 1726ed6e7268SNeal Cardwell const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 1727dcb8c9b4SEric Dumazet u32 min_tso, tso_segs; 1728ed6e7268SNeal Cardwell 1729dcb8c9b4SEric Dumazet min_tso = ca_ops->min_tso_segs ? 1730dcb8c9b4SEric Dumazet ca_ops->min_tso_segs(sk) : 1731dcb8c9b4SEric Dumazet sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs; 1732dcb8c9b4SEric Dumazet 1733dcb8c9b4SEric Dumazet tso_segs = tcp_tso_autosize(sk, mss_now, min_tso); 1734350c9f48SEric Dumazet return min_t(u32, tso_segs, sk->sk_gso_max_segs); 1735ed6e7268SNeal Cardwell } 1736ed6e7268SNeal Cardwell 1737d4589926SEric Dumazet /* Returns the portion of skb which can be sent right away */ 1738d4589926SEric Dumazet static unsigned int tcp_mss_split_point(const struct sock *sk, 1739d4589926SEric Dumazet const struct sk_buff *skb, 1740d4589926SEric Dumazet unsigned int mss_now, 1741d4589926SEric Dumazet unsigned int max_segs, 1742d4589926SEric Dumazet int nonagle) 1743c1b4a7e6SDavid S. Miller { 1744cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1745d4589926SEric Dumazet u32 partial, needed, window, max_len; 1746c1b4a7e6SDavid S. Miller 174790840defSIlpo Järvinen window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 17481485348dSBen Hutchings max_len = mss_now * max_segs; 17490e3a4803SIlpo Järvinen 17501485348dSBen Hutchings if (likely(max_len <= window && skb != tcp_write_queue_tail(sk))) 17511485348dSBen Hutchings return max_len; 17520e3a4803SIlpo Järvinen 17535ea3a748SIlpo Järvinen needed = min(skb->len, window); 17545ea3a748SIlpo Järvinen 17551485348dSBen Hutchings if (max_len <= needed) 17561485348dSBen Hutchings return max_len; 17570e3a4803SIlpo Järvinen 1758d4589926SEric Dumazet partial = needed % mss_now; 1759d4589926SEric Dumazet /* If last segment is not a full MSS, check if Nagle rules allow us 1760d4589926SEric Dumazet * to include this last segment in this skb. 1761d4589926SEric Dumazet * Otherwise, we'll split the skb at last MSS boundary 1762d4589926SEric Dumazet */ 1763cc93fc51SPeter Pan(潘卫平) if (tcp_nagle_check(partial != 0, tp, nonagle)) 1764d4589926SEric Dumazet return needed - partial; 1765d4589926SEric Dumazet 1766d4589926SEric Dumazet return needed; 1767c1b4a7e6SDavid S. Miller } 1768c1b4a7e6SDavid S. Miller 1769c1b4a7e6SDavid S. Miller /* Can at least one segment of SKB be sent right now, according to the 1770c1b4a7e6SDavid S. Miller * congestion window rules? If so, return how many segments are allowed. 1771c1b4a7e6SDavid S. Miller */ 1772cf533ea5SEric Dumazet static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, 1773cf533ea5SEric Dumazet const struct sk_buff *skb) 1774c1b4a7e6SDavid S. Miller { 1775d649a7a8SEric Dumazet u32 in_flight, cwnd, halfcwnd; 1776c1b4a7e6SDavid S. Miller 1777c1b4a7e6SDavid S. Miller /* Don't be strict about the congestion window for the final FIN. */ 17784de075e0SEric Dumazet if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && 17794de075e0SEric Dumazet tcp_skb_pcount(skb) == 1) 1780c1b4a7e6SDavid S. Miller return 1; 1781c1b4a7e6SDavid S. Miller 1782c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1783c1b4a7e6SDavid S. Miller cwnd = tp->snd_cwnd; 1784d649a7a8SEric Dumazet if (in_flight >= cwnd) 1785c1b4a7e6SDavid S. Miller return 0; 1786d649a7a8SEric Dumazet 1787d649a7a8SEric Dumazet /* For better scheduling, ensure we have at least 1788d649a7a8SEric Dumazet * 2 GSO packets in flight. 1789d649a7a8SEric Dumazet */ 1790d649a7a8SEric Dumazet halfcwnd = max(cwnd >> 1, 1U); 1791d649a7a8SEric Dumazet return min(halfcwnd, cwnd - in_flight); 1792c1b4a7e6SDavid S. Miller } 1793c1b4a7e6SDavid S. Miller 1794b595076aSUwe Kleine-König /* Initialize TSO state of a skb. 179567edfef7SAndi Kleen * This must be invoked the first time we consider transmitting 1796c1b4a7e6SDavid S. Miller * SKB onto the wire. 1797c1b4a7e6SDavid S. Miller */ 17985bbb432cSEric Dumazet static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now) 1799c1b4a7e6SDavid S. Miller { 1800c1b4a7e6SDavid S. Miller int tso_segs = tcp_skb_pcount(skb); 1801c1b4a7e6SDavid S. Miller 1802f8269a49SIlpo Järvinen if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { 18035bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 1804c1b4a7e6SDavid S. Miller tso_segs = tcp_skb_pcount(skb); 1805c1b4a7e6SDavid S. Miller } 1806c1b4a7e6SDavid S. Miller return tso_segs; 1807c1b4a7e6SDavid S. Miller } 1808c1b4a7e6SDavid S. Miller 1809c1b4a7e6SDavid S. Miller 1810a2a385d6SEric Dumazet /* Return true if the Nagle test allows this packet to be 1811c1b4a7e6SDavid S. Miller * sent now. 1812c1b4a7e6SDavid S. Miller */ 1813a2a385d6SEric Dumazet static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, 1814c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1815c1b4a7e6SDavid S. Miller { 1816c1b4a7e6SDavid S. Miller /* Nagle rule does not apply to frames, which sit in the middle of the 1817c1b4a7e6SDavid S. Miller * write_queue (they have no chances to get new data). 1818c1b4a7e6SDavid S. Miller * 1819c1b4a7e6SDavid S. Miller * This is implemented in the callers, where they modify the 'nonagle' 1820c1b4a7e6SDavid S. Miller * argument based upon the location of SKB in the send queue. 1821c1b4a7e6SDavid S. Miller */ 1822c1b4a7e6SDavid S. Miller if (nonagle & TCP_NAGLE_PUSH) 1823a2a385d6SEric Dumazet return true; 1824c1b4a7e6SDavid S. Miller 18259b44190dSYuchung Cheng /* Don't use the nagle rule for urgent data (or for the final FIN). */ 18269b44190dSYuchung Cheng if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) 1827a2a385d6SEric Dumazet return true; 1828c1b4a7e6SDavid S. Miller 1829cc93fc51SPeter Pan(潘卫平) if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle)) 1830a2a385d6SEric Dumazet return true; 1831c1b4a7e6SDavid S. Miller 1832a2a385d6SEric Dumazet return false; 1833c1b4a7e6SDavid S. Miller } 1834c1b4a7e6SDavid S. Miller 1835c1b4a7e6SDavid S. Miller /* Does at least the first segment of SKB fit into the send window? */ 1836a2a385d6SEric Dumazet static bool tcp_snd_wnd_test(const struct tcp_sock *tp, 1837a2a385d6SEric Dumazet const struct sk_buff *skb, 1838056834d9SIlpo Järvinen unsigned int cur_mss) 1839c1b4a7e6SDavid S. Miller { 1840c1b4a7e6SDavid S. Miller u32 end_seq = TCP_SKB_CB(skb)->end_seq; 1841c1b4a7e6SDavid S. Miller 1842c1b4a7e6SDavid S. Miller if (skb->len > cur_mss) 1843c1b4a7e6SDavid S. Miller end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 1844c1b4a7e6SDavid S. Miller 184590840defSIlpo Järvinen return !after(end_seq, tcp_wnd_end(tp)); 1846c1b4a7e6SDavid S. Miller } 1847c1b4a7e6SDavid S. Miller 1848c1b4a7e6SDavid S. Miller /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 1849c1b4a7e6SDavid S. Miller * which is put after SKB on the list. It is very much like 1850c1b4a7e6SDavid S. Miller * tcp_fragment() except that it may make several kinds of assumptions 1851c1b4a7e6SDavid S. Miller * in order to speed up the splitting operation. In particular, we 1852c1b4a7e6SDavid S. Miller * know that all the data is in scatter-gather pages, and that the 1853c1b4a7e6SDavid S. Miller * packet has never been sent out before (and thus is not cloned). 1854c1b4a7e6SDavid S. Miller */ 185575c119afSEric Dumazet static int tso_fragment(struct sock *sk, enum tcp_queue tcp_queue, 185675c119afSEric Dumazet struct sk_buff *skb, unsigned int len, 1857c4ead4c5SEric Dumazet unsigned int mss_now, gfp_t gfp) 1858c1b4a7e6SDavid S. Miller { 1859c1b4a7e6SDavid S. Miller struct sk_buff *buff; 1860c1b4a7e6SDavid S. Miller int nlen = skb->len - len; 18619ce01461SIlpo Järvinen u8 flags; 1862c1b4a7e6SDavid S. Miller 1863c1b4a7e6SDavid S. Miller /* All of a TSO frame must be composed of paged data. */ 1864c8ac3774SHerbert Xu if (skb->len != skb->data_len) 186575c119afSEric Dumazet return tcp_fragment(sk, tcp_queue, skb, len, mss_now, gfp); 1866c1b4a7e6SDavid S. Miller 1867eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, gfp, true); 186851456b29SIan Morris if (unlikely(!buff)) 1869c1b4a7e6SDavid S. Miller return -ENOMEM; 1870c1b4a7e6SDavid S. Miller 18713ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 18723ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1873b60b49eaSHerbert Xu buff->truesize += nlen; 1874c1b4a7e6SDavid S. Miller skb->truesize -= nlen; 1875c1b4a7e6SDavid S. Miller 1876c1b4a7e6SDavid S. Miller /* Correct the sequence numbers. */ 1877c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 1878c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 1879c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 1880c1b4a7e6SDavid S. Miller 1881c1b4a7e6SDavid S. Miller /* PSH and FIN should only be set in the second packet. */ 18824de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 18834de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 18844de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1885c1b4a7e6SDavid S. Miller 1886c1b4a7e6SDavid S. Miller /* This packet was never sent out yet, so no SACK bits. */ 1887c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->sacked = 0; 1888c1b4a7e6SDavid S. Miller 1889a166140eSMartin KaFai Lau tcp_skb_fragment_eor(skb, buff); 1890a166140eSMartin KaFai Lau 189198be9b12SEric Dumazet buff->ip_summed = CHECKSUM_PARTIAL; 1892c1b4a7e6SDavid S. Miller skb_split(skb, buff, len); 1893490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 1894c1b4a7e6SDavid S. Miller 1895c1b4a7e6SDavid S. Miller /* Fix up tso_factor for both original and new SKB. */ 18965bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 18975bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 1898c1b4a7e6SDavid S. Miller 1899c1b4a7e6SDavid S. Miller /* Link BUFF into the send queue. */ 1900f4a775d1SEric Dumazet __skb_header_release(buff); 190175c119afSEric Dumazet tcp_insert_write_queue_after(skb, buff, sk, tcp_queue); 1902c1b4a7e6SDavid S. Miller 1903c1b4a7e6SDavid S. Miller return 0; 1904c1b4a7e6SDavid S. Miller } 1905c1b4a7e6SDavid S. Miller 1906c1b4a7e6SDavid S. Miller /* Try to defer sending, if possible, in order to minimize the amount 1907c1b4a7e6SDavid S. Miller * of TSO splitting we do. View it as a kind of TSO Nagle test. 1908c1b4a7e6SDavid S. Miller * 1909c1b4a7e6SDavid S. Miller * This algorithm is from John Heffner. 1910c1b4a7e6SDavid S. Miller */ 1911ca8a2263SNeal Cardwell static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb, 1912605ad7f1SEric Dumazet bool *is_cwnd_limited, u32 max_segs) 1913c1b4a7e6SDavid S. Miller { 19146687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 191550c8339eSEric Dumazet u32 age, send_win, cong_win, limit, in_flight; 191650c8339eSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 191750c8339eSEric Dumazet struct sk_buff *head; 1918ad9f4f50SEric Dumazet int win_divisor; 1919c1b4a7e6SDavid S. Miller 19204de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 1921ae8064acSJohn Heffner goto send_now; 1922c1b4a7e6SDavid S. Miller 192399d7662aSEric Dumazet if (icsk->icsk_ca_state >= TCP_CA_Recovery) 1924ae8064acSJohn Heffner goto send_now; 1925ae8064acSJohn Heffner 19265f852eb5SEric Dumazet /* Avoid bursty behavior by allowing defer 19275f852eb5SEric Dumazet * only if the last write was recent. 19285f852eb5SEric Dumazet */ 1929d635fbe2SEric Dumazet if ((s32)(tcp_jiffies32 - tp->lsndtime) > 0) 1930ae8064acSJohn Heffner goto send_now; 1931908a75c1SDavid S. Miller 1932c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1933c1b4a7e6SDavid S. Miller 1934c8c9aeb5SStefano Brivio BUG_ON(tcp_skb_pcount(skb) <= 1); 1935c8c9aeb5SStefano Brivio BUG_ON(tp->snd_cwnd <= in_flight); 1936c1b4a7e6SDavid S. Miller 193790840defSIlpo Järvinen send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1938c1b4a7e6SDavid S. Miller 1939c1b4a7e6SDavid S. Miller /* From in_flight test above, we know that cwnd > in_flight. */ 1940c1b4a7e6SDavid S. Miller cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; 1941c1b4a7e6SDavid S. Miller 1942c1b4a7e6SDavid S. Miller limit = min(send_win, cong_win); 1943c1b4a7e6SDavid S. Miller 1944ba244fe9SDavid S. Miller /* If a full-sized TSO skb can be sent, do it. */ 1945605ad7f1SEric Dumazet if (limit >= max_segs * tp->mss_cache) 1946ae8064acSJohn Heffner goto send_now; 1947ba244fe9SDavid S. Miller 194862ad2761SIlpo Järvinen /* Middle in queue won't get any more data, full sendable already? */ 194962ad2761SIlpo Järvinen if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) 195062ad2761SIlpo Järvinen goto send_now; 195162ad2761SIlpo Järvinen 19525bbcc0f5SLinus Torvalds win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor); 1953ad9f4f50SEric Dumazet if (win_divisor) { 1954c1b4a7e6SDavid S. Miller u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); 1955c1b4a7e6SDavid S. Miller 1956c1b4a7e6SDavid S. Miller /* If at least some fraction of a window is available, 1957c1b4a7e6SDavid S. Miller * just use it. 1958c1b4a7e6SDavid S. Miller */ 1959ad9f4f50SEric Dumazet chunk /= win_divisor; 1960c1b4a7e6SDavid S. Miller if (limit >= chunk) 1961ae8064acSJohn Heffner goto send_now; 1962c1b4a7e6SDavid S. Miller } else { 1963c1b4a7e6SDavid S. Miller /* Different approach, try not to defer past a single 1964c1b4a7e6SDavid S. Miller * ACK. Receiver should ACK every other full sized 1965c1b4a7e6SDavid S. Miller * frame, so if we have space for more than 3 frames 1966c1b4a7e6SDavid S. Miller * then send now. 1967c1b4a7e6SDavid S. Miller */ 19686b5a5c0dSNeal Cardwell if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) 1969ae8064acSJohn Heffner goto send_now; 1970c1b4a7e6SDavid S. Miller } 1971c1b4a7e6SDavid S. Miller 197275c119afSEric Dumazet /* TODO : use tsorted_sent_queue ? */ 197375c119afSEric Dumazet head = tcp_rtx_queue_head(sk); 197475c119afSEric Dumazet if (!head) 197575c119afSEric Dumazet goto send_now; 19762fd66ffbSEric Dumazet age = tcp_stamp_us_delta(tp->tcp_mstamp, tcp_skb_timestamp_us(head)); 197750c8339eSEric Dumazet /* If next ACK is likely to come too late (half srtt), do not defer */ 197850c8339eSEric Dumazet if (age < (tp->srtt_us >> 4)) 197950c8339eSEric Dumazet goto send_now; 198050c8339eSEric Dumazet 19815f852eb5SEric Dumazet /* Ok, it looks like it is advisable to defer. */ 1982ae8064acSJohn Heffner 1983d2e1339fSBendik Rønning Opstad if (cong_win < send_win && cong_win <= skb->len) 1984ca8a2263SNeal Cardwell *is_cwnd_limited = true; 1985ca8a2263SNeal Cardwell 1986a2a385d6SEric Dumazet return true; 1987ae8064acSJohn Heffner 1988ae8064acSJohn Heffner send_now: 1989a2a385d6SEric Dumazet return false; 1990c1b4a7e6SDavid S. Miller } 1991c1b4a7e6SDavid S. Miller 199205cbc0dbSFan Du static inline void tcp_mtu_check_reprobe(struct sock *sk) 199305cbc0dbSFan Du { 199405cbc0dbSFan Du struct inet_connection_sock *icsk = inet_csk(sk); 199505cbc0dbSFan Du struct tcp_sock *tp = tcp_sk(sk); 199605cbc0dbSFan Du struct net *net = sock_net(sk); 199705cbc0dbSFan Du u32 interval; 199805cbc0dbSFan Du s32 delta; 199905cbc0dbSFan Du 200005cbc0dbSFan Du interval = net->ipv4.sysctl_tcp_probe_interval; 2001c74df29aSEric Dumazet delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp; 200205cbc0dbSFan Du if (unlikely(delta >= interval * HZ)) { 200305cbc0dbSFan Du int mss = tcp_current_mss(sk); 200405cbc0dbSFan Du 200505cbc0dbSFan Du /* Update current search range */ 200605cbc0dbSFan Du icsk->icsk_mtup.probe_size = 0; 200705cbc0dbSFan Du icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + 200805cbc0dbSFan Du sizeof(struct tcphdr) + 200905cbc0dbSFan Du icsk->icsk_af_ops->net_header_len; 201005cbc0dbSFan Du icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss); 201105cbc0dbSFan Du 201205cbc0dbSFan Du /* Update probe time stamp */ 2013c74df29aSEric Dumazet icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; 201405cbc0dbSFan Du } 201505cbc0dbSFan Du } 201605cbc0dbSFan Du 2017808cf9e3SIlya Lesokhin static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len) 2018808cf9e3SIlya Lesokhin { 2019808cf9e3SIlya Lesokhin struct sk_buff *skb, *next; 2020808cf9e3SIlya Lesokhin 2021808cf9e3SIlya Lesokhin skb = tcp_send_head(sk); 2022808cf9e3SIlya Lesokhin tcp_for_write_queue_from_safe(skb, next, sk) { 2023808cf9e3SIlya Lesokhin if (len <= skb->len) 2024808cf9e3SIlya Lesokhin break; 2025808cf9e3SIlya Lesokhin 2026808cf9e3SIlya Lesokhin if (unlikely(TCP_SKB_CB(skb)->eor)) 2027808cf9e3SIlya Lesokhin return false; 2028808cf9e3SIlya Lesokhin 2029808cf9e3SIlya Lesokhin len -= skb->len; 2030808cf9e3SIlya Lesokhin } 2031808cf9e3SIlya Lesokhin 2032808cf9e3SIlya Lesokhin return true; 2033808cf9e3SIlya Lesokhin } 2034808cf9e3SIlya Lesokhin 20355d424d5aSJohn Heffner /* Create a new MTU probe if we are ready. 203667edfef7SAndi Kleen * MTU probe is regularly attempting to increase the path MTU by 203767edfef7SAndi Kleen * deliberately sending larger packets. This discovers routing 203867edfef7SAndi Kleen * changes resulting in larger path MTUs. 203967edfef7SAndi Kleen * 20405d424d5aSJohn Heffner * Returns 0 if we should wait to probe (no cwnd available), 20415d424d5aSJohn Heffner * 1 if a probe was sent, 2042056834d9SIlpo Järvinen * -1 otherwise 2043056834d9SIlpo Järvinen */ 20445d424d5aSJohn Heffner static int tcp_mtu_probe(struct sock *sk) 20455d424d5aSJohn Heffner { 20465d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 204712a59abcSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 20485d424d5aSJohn Heffner struct sk_buff *skb, *nskb, *next; 20496b58e0a5SFan Du struct net *net = sock_net(sk); 20505d424d5aSJohn Heffner int probe_size; 205191cc17c0SIlpo Järvinen int size_needed; 205212a59abcSEric Dumazet int copy, len; 20535d424d5aSJohn Heffner int mss_now; 20546b58e0a5SFan Du int interval; 20555d424d5aSJohn Heffner 20565d424d5aSJohn Heffner /* Not currently probing/verifying, 20575d424d5aSJohn Heffner * not in recovery, 20585d424d5aSJohn Heffner * have enough cwnd, and 205912a59abcSEric Dumazet * not SACKing (the variable headers throw things off) 206012a59abcSEric Dumazet */ 206112a59abcSEric Dumazet if (likely(!icsk->icsk_mtup.enabled || 20625d424d5aSJohn Heffner icsk->icsk_mtup.probe_size || 20635d424d5aSJohn Heffner inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 20645d424d5aSJohn Heffner tp->snd_cwnd < 11 || 206512a59abcSEric Dumazet tp->rx_opt.num_sacks || tp->rx_opt.dsack)) 20665d424d5aSJohn Heffner return -1; 20675d424d5aSJohn Heffner 20686b58e0a5SFan Du /* Use binary search for probe_size between tcp_mss_base, 20696b58e0a5SFan Du * and current mss_clamp. if (search_high - search_low) 20706b58e0a5SFan Du * smaller than a threshold, backoff from probing. 20716b58e0a5SFan Du */ 20720c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 20736b58e0a5SFan Du probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high + 20746b58e0a5SFan Du icsk->icsk_mtup.search_low) >> 1); 207591cc17c0SIlpo Järvinen size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; 20766b58e0a5SFan Du interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low; 207705cbc0dbSFan Du /* When misfortune happens, we are reprobing actively, 207805cbc0dbSFan Du * and then reprobe timer has expired. We stick with current 207905cbc0dbSFan Du * probing process by not resetting search range to its orignal. 208005cbc0dbSFan Du */ 20816b58e0a5SFan Du if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) || 208205cbc0dbSFan Du interval < net->ipv4.sysctl_tcp_probe_threshold) { 208305cbc0dbSFan Du /* Check whether enough time has elaplased for 208405cbc0dbSFan Du * another round of probing. 208505cbc0dbSFan Du */ 208605cbc0dbSFan Du tcp_mtu_check_reprobe(sk); 20875d424d5aSJohn Heffner return -1; 20885d424d5aSJohn Heffner } 20895d424d5aSJohn Heffner 20905d424d5aSJohn Heffner /* Have enough data in the send queue to probe? */ 20917f9c33e5SIlpo Järvinen if (tp->write_seq - tp->snd_nxt < size_needed) 20925d424d5aSJohn Heffner return -1; 20935d424d5aSJohn Heffner 209491cc17c0SIlpo Järvinen if (tp->snd_wnd < size_needed) 20955d424d5aSJohn Heffner return -1; 209690840defSIlpo Järvinen if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) 20975d424d5aSJohn Heffner return 0; 20985d424d5aSJohn Heffner 2099d67c58e9SIlpo Järvinen /* Do we need to wait to drain cwnd? With none in flight, don't stall */ 2100d67c58e9SIlpo Järvinen if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { 2101d67c58e9SIlpo Järvinen if (!tcp_packets_in_flight(tp)) 21025d424d5aSJohn Heffner return -1; 21035d424d5aSJohn Heffner else 21045d424d5aSJohn Heffner return 0; 21055d424d5aSJohn Heffner } 21065d424d5aSJohn Heffner 2107808cf9e3SIlya Lesokhin if (!tcp_can_coalesce_send_queue_head(sk, probe_size)) 2108808cf9e3SIlya Lesokhin return -1; 2109808cf9e3SIlya Lesokhin 21105d424d5aSJohn Heffner /* We're allowed to probe. Build it now. */ 2111eb934478SEric Dumazet nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false); 211251456b29SIan Morris if (!nskb) 21135d424d5aSJohn Heffner return -1; 21143ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 21153ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 21165d424d5aSJohn Heffner 2117fe067e8aSDavid S. Miller skb = tcp_send_head(sk); 21185d424d5aSJohn Heffner 21195d424d5aSJohn Heffner TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 21205d424d5aSJohn Heffner TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 21214de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; 21225d424d5aSJohn Heffner TCP_SKB_CB(nskb)->sacked = 0; 21235d424d5aSJohn Heffner nskb->csum = 0; 212498be9b12SEric Dumazet nskb->ip_summed = CHECKSUM_PARTIAL; 21255d424d5aSJohn Heffner 212650c4817eSIlpo Järvinen tcp_insert_write_queue_before(nskb, skb, sk); 21272b7cda9cSEric Dumazet tcp_highest_sack_replace(sk, skb, nskb); 212850c4817eSIlpo Järvinen 21295d424d5aSJohn Heffner len = 0; 2130234b6860SIlpo Järvinen tcp_for_write_queue_from_safe(skb, next, sk) { 21315d424d5aSJohn Heffner copy = min_t(int, skb->len, probe_size - len); 21325d424d5aSJohn Heffner skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); 21335d424d5aSJohn Heffner 21345d424d5aSJohn Heffner if (skb->len <= copy) { 21355d424d5aSJohn Heffner /* We've eaten all the data from this skb. 21365d424d5aSJohn Heffner * Throw it away. */ 21374de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 2138808cf9e3SIlya Lesokhin /* If this is the last SKB we copy and eor is set 2139808cf9e3SIlya Lesokhin * we need to propagate it to the new skb. 2140808cf9e3SIlya Lesokhin */ 2141808cf9e3SIlya Lesokhin TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor; 2142fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 21433ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 21445d424d5aSJohn Heffner } else { 21454de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & 2146a3433f35SChangli Gao ~(TCPHDR_FIN|TCPHDR_PSH); 21475d424d5aSJohn Heffner if (!skb_shinfo(skb)->nr_frags) { 21485d424d5aSJohn Heffner skb_pull(skb, copy); 21495d424d5aSJohn Heffner } else { 21505d424d5aSJohn Heffner __pskb_trim_head(skb, copy); 21515bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 21525d424d5aSJohn Heffner } 21535d424d5aSJohn Heffner TCP_SKB_CB(skb)->seq += copy; 21545d424d5aSJohn Heffner } 21555d424d5aSJohn Heffner 21565d424d5aSJohn Heffner len += copy; 2157234b6860SIlpo Järvinen 2158234b6860SIlpo Järvinen if (len >= probe_size) 2159234b6860SIlpo Järvinen break; 21605d424d5aSJohn Heffner } 21615bbb432cSEric Dumazet tcp_init_tso_segs(nskb, nskb->len); 21625d424d5aSJohn Heffner 21635d424d5aSJohn Heffner /* We're ready to send. If this fails, the probe will 21647faee5c0SEric Dumazet * be resegmented into mss-sized pieces by tcp_write_xmit(). 21657faee5c0SEric Dumazet */ 21665d424d5aSJohn Heffner if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 21675d424d5aSJohn Heffner /* Decrement cwnd here because we are sending 21685d424d5aSJohn Heffner * effectively two packets. */ 21695d424d5aSJohn Heffner tp->snd_cwnd--; 217066f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, nskb); 21715d424d5aSJohn Heffner 21725d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 21730e7b1368SJohn Heffner tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 21740e7b1368SJohn Heffner tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 21755d424d5aSJohn Heffner 21765d424d5aSJohn Heffner return 1; 21775d424d5aSJohn Heffner } 21785d424d5aSJohn Heffner 21795d424d5aSJohn Heffner return -1; 21805d424d5aSJohn Heffner } 21815d424d5aSJohn Heffner 2182218af599SEric Dumazet static bool tcp_pacing_check(const struct sock *sk) 2183218af599SEric Dumazet { 2184218af599SEric Dumazet return tcp_needs_internal_pacing(sk) && 218573a6bab5SEric Dumazet hrtimer_is_queued(&tcp_sk(sk)->pacing_timer); 2186218af599SEric Dumazet } 2187218af599SEric Dumazet 2188f9616c35SEric Dumazet /* TCP Small Queues : 2189f9616c35SEric Dumazet * Control number of packets in qdisc/devices to two packets / or ~1 ms. 2190f9616c35SEric Dumazet * (These limits are doubled for retransmits) 2191f9616c35SEric Dumazet * This allows for : 2192f9616c35SEric Dumazet * - better RTT estimation and ACK scheduling 2193f9616c35SEric Dumazet * - faster recovery 2194f9616c35SEric Dumazet * - high rates 2195f9616c35SEric Dumazet * Alas, some drivers / subsystems require a fair amount 2196f9616c35SEric Dumazet * of queued bytes to ensure line rate. 2197f9616c35SEric Dumazet * One example is wifi aggregation (802.11 AMPDU) 2198f9616c35SEric Dumazet */ 2199f9616c35SEric Dumazet static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb, 2200f9616c35SEric Dumazet unsigned int factor) 2201f9616c35SEric Dumazet { 2202*76a9ebe8SEric Dumazet unsigned long limit; 2203f9616c35SEric Dumazet 2204*76a9ebe8SEric Dumazet limit = max_t(unsigned long, 2205*76a9ebe8SEric Dumazet 2 * skb->truesize, 2206*76a9ebe8SEric Dumazet sk->sk_pacing_rate >> sk->sk_pacing_shift); 2207*76a9ebe8SEric Dumazet limit = min_t(unsigned long, limit, 22089184d8bbSEric Dumazet sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes); 2209f9616c35SEric Dumazet limit <<= factor; 2210f9616c35SEric Dumazet 221114afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) > limit) { 221275c119afSEric Dumazet /* Always send skb if rtx queue is empty. 221375eefc6cSEric Dumazet * No need to wait for TX completion to call us back, 221475eefc6cSEric Dumazet * after softirq/tasklet schedule. 221575eefc6cSEric Dumazet * This helps when TX completions are delayed too much. 221675eefc6cSEric Dumazet */ 221775c119afSEric Dumazet if (tcp_rtx_queue_empty(sk)) 221875eefc6cSEric Dumazet return false; 221975eefc6cSEric Dumazet 22207aa5470cSEric Dumazet set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags); 2221f9616c35SEric Dumazet /* It is possible TX completion already happened 2222f9616c35SEric Dumazet * before we set TSQ_THROTTLED, so we must 2223f9616c35SEric Dumazet * test again the condition. 2224f9616c35SEric Dumazet */ 2225f9616c35SEric Dumazet smp_mb__after_atomic(); 222614afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) > limit) 2227f9616c35SEric Dumazet return true; 2228f9616c35SEric Dumazet } 2229f9616c35SEric Dumazet return false; 2230f9616c35SEric Dumazet } 2231f9616c35SEric Dumazet 223205b055e8SFrancis Yan static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new) 223305b055e8SFrancis Yan { 2234628174ccSEric Dumazet const u32 now = tcp_jiffies32; 2235efe967cdSArnd Bergmann enum tcp_chrono old = tp->chrono_type; 223605b055e8SFrancis Yan 2237efe967cdSArnd Bergmann if (old > TCP_CHRONO_UNSPEC) 2238efe967cdSArnd Bergmann tp->chrono_stat[old - 1] += now - tp->chrono_start; 223905b055e8SFrancis Yan tp->chrono_start = now; 224005b055e8SFrancis Yan tp->chrono_type = new; 224105b055e8SFrancis Yan } 224205b055e8SFrancis Yan 224305b055e8SFrancis Yan void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type) 224405b055e8SFrancis Yan { 224505b055e8SFrancis Yan struct tcp_sock *tp = tcp_sk(sk); 224605b055e8SFrancis Yan 224705b055e8SFrancis Yan /* If there are multiple conditions worthy of tracking in a 22480f87230dSFrancis Yan * chronograph then the highest priority enum takes precedence 22490f87230dSFrancis Yan * over the other conditions. So that if something "more interesting" 225005b055e8SFrancis Yan * starts happening, stop the previous chrono and start a new one. 225105b055e8SFrancis Yan */ 225205b055e8SFrancis Yan if (type > tp->chrono_type) 225305b055e8SFrancis Yan tcp_chrono_set(tp, type); 225405b055e8SFrancis Yan } 225505b055e8SFrancis Yan 225605b055e8SFrancis Yan void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type) 225705b055e8SFrancis Yan { 225805b055e8SFrancis Yan struct tcp_sock *tp = tcp_sk(sk); 225905b055e8SFrancis Yan 22600f87230dSFrancis Yan 22610f87230dSFrancis Yan /* There are multiple conditions worthy of tracking in a 22620f87230dSFrancis Yan * chronograph, so that the highest priority enum takes 22630f87230dSFrancis Yan * precedence over the other conditions (see tcp_chrono_start). 22640f87230dSFrancis Yan * If a condition stops, we only stop chrono tracking if 22650f87230dSFrancis Yan * it's the "most interesting" or current chrono we are 22660f87230dSFrancis Yan * tracking and starts busy chrono if we have pending data. 22670f87230dSFrancis Yan */ 226875c119afSEric Dumazet if (tcp_rtx_and_write_queues_empty(sk)) 226905b055e8SFrancis Yan tcp_chrono_set(tp, TCP_CHRONO_UNSPEC); 22700f87230dSFrancis Yan else if (type == tp->chrono_type) 22710f87230dSFrancis Yan tcp_chrono_set(tp, TCP_CHRONO_BUSY); 227205b055e8SFrancis Yan } 227305b055e8SFrancis Yan 22741da177e4SLinus Torvalds /* This routine writes packets to the network. It advances the 22751da177e4SLinus Torvalds * send_head. This happens as incoming acks open up the remote 22761da177e4SLinus Torvalds * window for us. 22771da177e4SLinus Torvalds * 2278f8269a49SIlpo Järvinen * LARGESEND note: !tcp_urg_mode is overkill, only frames between 2279f8269a49SIlpo Järvinen * snd_up-64k-mss .. snd_up cannot be large. However, taking into 2280f8269a49SIlpo Järvinen * account rare use of URG, this is not a big flaw. 2281f8269a49SIlpo Järvinen * 22826ba8a3b1SNandita Dukkipati * Send at most one packet when push_one > 0. Temporarily ignore 22836ba8a3b1SNandita Dukkipati * cwnd limit to force at most one packet out when push_one == 2. 22846ba8a3b1SNandita Dukkipati 2285a2a385d6SEric Dumazet * Returns true, if no segments are in flight and we have queued segments, 2286a2a385d6SEric Dumazet * but cannot send anything now because of SWS or another problem. 22871da177e4SLinus Torvalds */ 2288a2a385d6SEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 2289d5dd9175SIlpo Järvinen int push_one, gfp_t gfp) 22901da177e4SLinus Torvalds { 22911da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 229292df7b51SDavid S. Miller struct sk_buff *skb; 2293c1b4a7e6SDavid S. Miller unsigned int tso_segs, sent_pkts; 2294c1b4a7e6SDavid S. Miller int cwnd_quota; 22955d424d5aSJohn Heffner int result; 22965615f886SFrancis Yan bool is_cwnd_limited = false, is_rwnd_limited = false; 2297605ad7f1SEric Dumazet u32 max_segs; 22981da177e4SLinus Torvalds 2299c1b4a7e6SDavid S. Miller sent_pkts = 0; 23005d424d5aSJohn Heffner 2301ee1836aeSEric Dumazet tcp_mstamp_refresh(tp); 2302d5dd9175SIlpo Järvinen if (!push_one) { 23035d424d5aSJohn Heffner /* Do MTU probing. */ 2304d5dd9175SIlpo Järvinen result = tcp_mtu_probe(sk); 2305d5dd9175SIlpo Järvinen if (!result) { 2306a2a385d6SEric Dumazet return false; 23075d424d5aSJohn Heffner } else if (result > 0) { 23085d424d5aSJohn Heffner sent_pkts = 1; 23095d424d5aSJohn Heffner } 2310d5dd9175SIlpo Järvinen } 23115d424d5aSJohn Heffner 2312ed6e7268SNeal Cardwell max_segs = tcp_tso_segs(sk, mss_now); 2313fe067e8aSDavid S. Miller while ((skb = tcp_send_head(sk))) { 2314c8ac3774SHerbert Xu unsigned int limit; 2315c8ac3774SHerbert Xu 2316218af599SEric Dumazet if (tcp_pacing_check(sk)) 2317218af599SEric Dumazet break; 2318218af599SEric Dumazet 23195bbb432cSEric Dumazet tso_segs = tcp_init_tso_segs(skb, mss_now); 2320c1b4a7e6SDavid S. Miller BUG_ON(!tso_segs); 2321c1b4a7e6SDavid S. Miller 23229d186cacSAndrey Vagin if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) { 23237faee5c0SEric Dumazet /* "skb_mstamp" is used as a start point for the retransmit timer */ 2324ab408b6dSEric Dumazet tcp_update_skb_after_send(sk, skb); 2325ec342325SAndrew Vagin goto repair; /* Skip network transmission */ 23269d186cacSAndrey Vagin } 2327ec342325SAndrew Vagin 2328b68e9f85SHerbert Xu cwnd_quota = tcp_cwnd_test(tp, skb); 23296ba8a3b1SNandita Dukkipati if (!cwnd_quota) { 23306ba8a3b1SNandita Dukkipati if (push_one == 2) 23316ba8a3b1SNandita Dukkipati /* Force out a loss probe pkt. */ 23326ba8a3b1SNandita Dukkipati cwnd_quota = 1; 23336ba8a3b1SNandita Dukkipati else 2334b68e9f85SHerbert Xu break; 23356ba8a3b1SNandita Dukkipati } 2336b68e9f85SHerbert Xu 23375615f886SFrancis Yan if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) { 23385615f886SFrancis Yan is_rwnd_limited = true; 2339b68e9f85SHerbert Xu break; 23405615f886SFrancis Yan } 2341b68e9f85SHerbert Xu 2342d6a4e26aSEric Dumazet if (tso_segs == 1) { 2343aa93466bSDavid S. Miller if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 2344aa93466bSDavid S. Miller (tcp_skb_is_last(sk, skb) ? 2345aa93466bSDavid S. Miller nonagle : TCP_NAGLE_PUSH)))) 2346aa93466bSDavid S. Miller break; 2347c1b4a7e6SDavid S. Miller } else { 2348ca8a2263SNeal Cardwell if (!push_one && 2349605ad7f1SEric Dumazet tcp_tso_should_defer(sk, skb, &is_cwnd_limited, 2350605ad7f1SEric Dumazet max_segs)) 2351aa93466bSDavid S. Miller break; 2352c1b4a7e6SDavid S. Miller } 2353aa93466bSDavid S. Miller 2354605ad7f1SEric Dumazet limit = mss_now; 2355d6a4e26aSEric Dumazet if (tso_segs > 1 && !tcp_urg_mode(tp)) 2356605ad7f1SEric Dumazet limit = tcp_mss_split_point(sk, skb, mss_now, 2357605ad7f1SEric Dumazet min_t(unsigned int, 2358605ad7f1SEric Dumazet cwnd_quota, 2359605ad7f1SEric Dumazet max_segs), 2360605ad7f1SEric Dumazet nonagle); 2361605ad7f1SEric Dumazet 2362605ad7f1SEric Dumazet if (skb->len > limit && 236375c119afSEric Dumazet unlikely(tso_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE, 236475c119afSEric Dumazet skb, limit, mss_now, gfp))) 2365605ad7f1SEric Dumazet break; 2366605ad7f1SEric Dumazet 2367f9616c35SEric Dumazet if (tcp_small_queue_check(sk, skb, 0)) 236846d3ceabSEric Dumazet break; 2369c9eeec26SEric Dumazet 2370d5dd9175SIlpo Järvinen if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) 23711da177e4SLinus Torvalds break; 23721da177e4SLinus Torvalds 2373ec342325SAndrew Vagin repair: 23741da177e4SLinus Torvalds /* Advance the send_head. This one is sent out. 23751da177e4SLinus Torvalds * This call will increment packets_out. 23761da177e4SLinus Torvalds */ 237766f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 23781da177e4SLinus Torvalds 23791da177e4SLinus Torvalds tcp_minshall_update(tp, mss_now, skb); 2380a262f0cdSNandita Dukkipati sent_pkts += tcp_skb_pcount(skb); 2381d5dd9175SIlpo Järvinen 2382d5dd9175SIlpo Järvinen if (push_one) 2383d5dd9175SIlpo Järvinen break; 23841da177e4SLinus Torvalds } 23851da177e4SLinus Torvalds 23865615f886SFrancis Yan if (is_rwnd_limited) 23875615f886SFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED); 23885615f886SFrancis Yan else 23895615f886SFrancis Yan tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED); 23905615f886SFrancis Yan 2391aa93466bSDavid S. Miller if (likely(sent_pkts)) { 2392684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 2393684bad11SYuchung Cheng tp->prr_out += sent_pkts; 23946ba8a3b1SNandita Dukkipati 23956ba8a3b1SNandita Dukkipati /* Send one loss probe per tail loss episode. */ 23966ba8a3b1SNandita Dukkipati if (push_one != 2) 2397ed66dfafSNeal Cardwell tcp_schedule_loss_probe(sk, false); 2398d2e1339fSBendik Rønning Opstad is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd); 2399ca8a2263SNeal Cardwell tcp_cwnd_validate(sk, is_cwnd_limited); 2400a2a385d6SEric Dumazet return false; 24011da177e4SLinus Torvalds } 240275c119afSEric Dumazet return !tp->packets_out && !tcp_write_queue_empty(sk); 24036ba8a3b1SNandita Dukkipati } 24046ba8a3b1SNandita Dukkipati 2405ed66dfafSNeal Cardwell bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto) 24066ba8a3b1SNandita Dukkipati { 24076ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 24086ba8a3b1SNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 2409a2815817SNeal Cardwell u32 timeout, rto_delta_us; 24102ae21cf5SEric Dumazet int early_retrans; 24116ba8a3b1SNandita Dukkipati 24126ba8a3b1SNandita Dukkipati /* Don't do any loss probe on a Fast Open connection before 3WHS 24136ba8a3b1SNandita Dukkipati * finishes. 24146ba8a3b1SNandita Dukkipati */ 2415f9b99582SYuchung Cheng if (tp->fastopen_rsk) 24166ba8a3b1SNandita Dukkipati return false; 24176ba8a3b1SNandita Dukkipati 24182ae21cf5SEric Dumazet early_retrans = sock_net(sk)->ipv4.sysctl_tcp_early_retrans; 24196ba8a3b1SNandita Dukkipati /* Schedule a loss probe in 2*RTT for SACK capable connections 2420b4f70c3dSNeal Cardwell * not in loss recovery, that are either limited by cwnd or application. 24216ba8a3b1SNandita Dukkipati */ 24222ae21cf5SEric Dumazet if ((early_retrans != 3 && early_retrans != 4) || 2423bec41a11SYuchung Cheng !tp->packets_out || !tcp_is_sack(tp) || 2424b4f70c3dSNeal Cardwell (icsk->icsk_ca_state != TCP_CA_Open && 2425b4f70c3dSNeal Cardwell icsk->icsk_ca_state != TCP_CA_CWR)) 24266ba8a3b1SNandita Dukkipati return false; 24276ba8a3b1SNandita Dukkipati 2428bb4d991aSYuchung Cheng /* Probe timeout is 2*rtt. Add minimum RTO to account 2429f9b99582SYuchung Cheng * for delayed ack when there's one outstanding packet. If no RTT 2430f9b99582SYuchung Cheng * sample is available then probe after TCP_TIMEOUT_INIT. 24316ba8a3b1SNandita Dukkipati */ 2432bb4d991aSYuchung Cheng if (tp->srtt_us) { 2433bb4d991aSYuchung Cheng timeout = usecs_to_jiffies(tp->srtt_us >> 2); 24346ba8a3b1SNandita Dukkipati if (tp->packets_out == 1) 2435bb4d991aSYuchung Cheng timeout += TCP_RTO_MIN; 2436bb4d991aSYuchung Cheng else 2437bb4d991aSYuchung Cheng timeout += TCP_TIMEOUT_MIN; 2438bb4d991aSYuchung Cheng } else { 2439bb4d991aSYuchung Cheng timeout = TCP_TIMEOUT_INIT; 2440bb4d991aSYuchung Cheng } 24416ba8a3b1SNandita Dukkipati 2442a2815817SNeal Cardwell /* If the RTO formula yields an earlier time, then use that time. */ 2443ed66dfafSNeal Cardwell rto_delta_us = advancing_rto ? 2444ed66dfafSNeal Cardwell jiffies_to_usecs(inet_csk(sk)->icsk_rto) : 2445ed66dfafSNeal Cardwell tcp_rto_delta_us(sk); /* How far in future is RTO? */ 2446a2815817SNeal Cardwell if (rto_delta_us > 0) 2447a2815817SNeal Cardwell timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us)); 24486ba8a3b1SNandita Dukkipati 24496ba8a3b1SNandita Dukkipati inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, 24506ba8a3b1SNandita Dukkipati TCP_RTO_MAX); 24516ba8a3b1SNandita Dukkipati return true; 24526ba8a3b1SNandita Dukkipati } 24536ba8a3b1SNandita Dukkipati 24541f3279aeSEric Dumazet /* Thanks to skb fast clones, we can detect if a prior transmit of 24551f3279aeSEric Dumazet * a packet is still in a qdisc or driver queue. 24561f3279aeSEric Dumazet * In this case, there is very little point doing a retransmit ! 24571f3279aeSEric Dumazet */ 24581f3279aeSEric Dumazet static bool skb_still_in_host_queue(const struct sock *sk, 24591f3279aeSEric Dumazet const struct sk_buff *skb) 24601f3279aeSEric Dumazet { 246139bb5e62SEric Dumazet if (unlikely(skb_fclone_busy(sk, skb))) { 2462c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), 24631f3279aeSEric Dumazet LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES); 24641f3279aeSEric Dumazet return true; 24651f3279aeSEric Dumazet } 24661f3279aeSEric Dumazet return false; 24671f3279aeSEric Dumazet } 24681f3279aeSEric Dumazet 2469b340b264SYuchung Cheng /* When probe timeout (PTO) fires, try send a new segment if possible, else 24706ba8a3b1SNandita Dukkipati * retransmit the last segment. 24716ba8a3b1SNandita Dukkipati */ 24726ba8a3b1SNandita Dukkipati void tcp_send_loss_probe(struct sock *sk) 24736ba8a3b1SNandita Dukkipati { 24749b717a8dSNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 24756ba8a3b1SNandita Dukkipati struct sk_buff *skb; 24766ba8a3b1SNandita Dukkipati int pcount; 24776ba8a3b1SNandita Dukkipati int mss = tcp_current_mss(sk); 24786ba8a3b1SNandita Dukkipati 2479b340b264SYuchung Cheng skb = tcp_send_head(sk); 248075c119afSEric Dumazet if (skb && tcp_snd_wnd_test(tp, skb, mss)) { 2481b340b264SYuchung Cheng pcount = tp->packets_out; 2482b340b264SYuchung Cheng tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC); 2483b340b264SYuchung Cheng if (tp->packets_out > pcount) 2484b340b264SYuchung Cheng goto probe_sent; 24856ba8a3b1SNandita Dukkipati goto rearm_timer; 24866ba8a3b1SNandita Dukkipati } 248775c119afSEric Dumazet skb = skb_rb_last(&sk->tcp_rtx_queue); 24886ba8a3b1SNandita Dukkipati 24899b717a8dSNandita Dukkipati /* At most one outstanding TLP retransmission. */ 24909b717a8dSNandita Dukkipati if (tp->tlp_high_seq) 24919b717a8dSNandita Dukkipati goto rearm_timer; 24929b717a8dSNandita Dukkipati 24936ba8a3b1SNandita Dukkipati /* Retransmit last segment. */ 24946ba8a3b1SNandita Dukkipati if (WARN_ON(!skb)) 24956ba8a3b1SNandita Dukkipati goto rearm_timer; 24966ba8a3b1SNandita Dukkipati 24971f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 24981f3279aeSEric Dumazet goto rearm_timer; 24991f3279aeSEric Dumazet 25006ba8a3b1SNandita Dukkipati pcount = tcp_skb_pcount(skb); 25016ba8a3b1SNandita Dukkipati if (WARN_ON(!pcount)) 25026ba8a3b1SNandita Dukkipati goto rearm_timer; 25036ba8a3b1SNandita Dukkipati 25046ba8a3b1SNandita Dukkipati if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) { 250575c119afSEric Dumazet if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, 250675c119afSEric Dumazet (pcount - 1) * mss, mss, 25076cc55e09SOctavian Purdila GFP_ATOMIC))) 25086ba8a3b1SNandita Dukkipati goto rearm_timer; 250975c119afSEric Dumazet skb = skb_rb_next(skb); 25106ba8a3b1SNandita Dukkipati } 25116ba8a3b1SNandita Dukkipati 25126ba8a3b1SNandita Dukkipati if (WARN_ON(!skb || !tcp_skb_pcount(skb))) 25136ba8a3b1SNandita Dukkipati goto rearm_timer; 25146ba8a3b1SNandita Dukkipati 251510d3be56SEric Dumazet if (__tcp_retransmit_skb(sk, skb, 1)) 2516b340b264SYuchung Cheng goto rearm_timer; 25176ba8a3b1SNandita Dukkipati 25189b717a8dSNandita Dukkipati /* Record snd_nxt for loss detection. */ 25199b717a8dSNandita Dukkipati tp->tlp_high_seq = tp->snd_nxt; 25209b717a8dSNandita Dukkipati 2521b340b264SYuchung Cheng probe_sent: 2522c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES); 2523fcd16c0aSYuchung Cheng /* Reset s.t. tcp_rearm_rto will restart timer from now */ 2524fcd16c0aSYuchung Cheng inet_csk(sk)->icsk_pending = 0; 2525b340b264SYuchung Cheng rearm_timer: 2526fcd16c0aSYuchung Cheng tcp_rearm_rto(sk); 25271da177e4SLinus Torvalds } 25281da177e4SLinus Torvalds 2529a762a980SDavid S. Miller /* Push out any pending frames which were held back due to 2530a762a980SDavid S. Miller * TCP_CORK or attempt at coalescing tiny packets. 2531a762a980SDavid S. Miller * The socket must be locked by the caller. 2532a762a980SDavid S. Miller */ 25339e412ba7SIlpo Järvinen void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 25349e412ba7SIlpo Järvinen int nonagle) 2535a762a980SDavid S. Miller { 2536726e07a8SIlpo Järvinen /* If we are closed, the bytes will have to remain here. 2537726e07a8SIlpo Järvinen * In time closedown will finish, we empty the write queue and 2538726e07a8SIlpo Järvinen * all will be happy. 2539726e07a8SIlpo Järvinen */ 2540726e07a8SIlpo Järvinen if (unlikely(sk->sk_state == TCP_CLOSE)) 2541726e07a8SIlpo Järvinen return; 2542726e07a8SIlpo Järvinen 254399a1dec7SMel Gorman if (tcp_write_xmit(sk, cur_mss, nonagle, 0, 25447450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC))) 25459e412ba7SIlpo Järvinen tcp_check_probe_timer(sk); 2546a762a980SDavid S. Miller } 2547a762a980SDavid S. Miller 2548c1b4a7e6SDavid S. Miller /* Send _single_ skb sitting at the send head. This function requires 2549c1b4a7e6SDavid S. Miller * true push pending frames to setup probe timer etc. 2550c1b4a7e6SDavid S. Miller */ 2551c1b4a7e6SDavid S. Miller void tcp_push_one(struct sock *sk, unsigned int mss_now) 2552c1b4a7e6SDavid S. Miller { 2553fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 2554c1b4a7e6SDavid S. Miller 2555c1b4a7e6SDavid S. Miller BUG_ON(!skb || skb->len < mss_now); 2556c1b4a7e6SDavid S. Miller 2557d5dd9175SIlpo Järvinen tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); 2558c1b4a7e6SDavid S. Miller } 2559c1b4a7e6SDavid S. Miller 25601da177e4SLinus Torvalds /* This function returns the amount that we can raise the 25611da177e4SLinus Torvalds * usable window based on the following constraints 25621da177e4SLinus Torvalds * 25631da177e4SLinus Torvalds * 1. The window can never be shrunk once it is offered (RFC 793) 25641da177e4SLinus Torvalds * 2. We limit memory per socket 25651da177e4SLinus Torvalds * 25661da177e4SLinus Torvalds * RFC 1122: 25671da177e4SLinus Torvalds * "the suggested [SWS] avoidance algorithm for the receiver is to keep 25681da177e4SLinus Torvalds * RECV.NEXT + RCV.WIN fixed until: 25691da177e4SLinus Torvalds * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 25701da177e4SLinus Torvalds * 25711da177e4SLinus Torvalds * i.e. don't raise the right edge of the window until you can raise 25721da177e4SLinus Torvalds * it at least MSS bytes. 25731da177e4SLinus Torvalds * 25741da177e4SLinus Torvalds * Unfortunately, the recommended algorithm breaks header prediction, 25751da177e4SLinus Torvalds * since header prediction assumes th->window stays fixed. 25761da177e4SLinus Torvalds * 25771da177e4SLinus Torvalds * Strictly speaking, keeping th->window fixed violates the receiver 25781da177e4SLinus Torvalds * side SWS prevention criteria. The problem is that under this rule 25791da177e4SLinus Torvalds * a stream of single byte packets will cause the right side of the 25801da177e4SLinus Torvalds * window to always advance by a single byte. 25811da177e4SLinus Torvalds * 25821da177e4SLinus Torvalds * Of course, if the sender implements sender side SWS prevention 25831da177e4SLinus Torvalds * then this will not be a problem. 25841da177e4SLinus Torvalds * 25851da177e4SLinus Torvalds * BSD seems to make the following compromise: 25861da177e4SLinus Torvalds * 25871da177e4SLinus Torvalds * If the free space is less than the 1/4 of the maximum 25881da177e4SLinus Torvalds * space available and the free space is less than 1/2 mss, 25891da177e4SLinus Torvalds * then set the window to 0. 25901da177e4SLinus Torvalds * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 25911da177e4SLinus Torvalds * Otherwise, just prevent the window from shrinking 25921da177e4SLinus Torvalds * and from being larger than the largest representable value. 25931da177e4SLinus Torvalds * 25941da177e4SLinus Torvalds * This prevents incremental opening of the window in the regime 25951da177e4SLinus Torvalds * where TCP is limited by the speed of the reader side taking 25961da177e4SLinus Torvalds * data out of the TCP receive queue. It does nothing about 25971da177e4SLinus Torvalds * those cases where the window is constrained on the sender side 25981da177e4SLinus Torvalds * because the pipeline is full. 25991da177e4SLinus Torvalds * 26001da177e4SLinus Torvalds * BSD also seems to "accidentally" limit itself to windows that are a 26011da177e4SLinus Torvalds * multiple of MSS, at least until the free space gets quite small. 26021da177e4SLinus Torvalds * This would appear to be a side effect of the mbuf implementation. 26031da177e4SLinus Torvalds * Combining these two algorithms results in the observed behavior 26041da177e4SLinus Torvalds * of having a fixed window size at almost all times. 26051da177e4SLinus Torvalds * 26061da177e4SLinus Torvalds * Below we obtain similar behavior by forcing the offered window to 26071da177e4SLinus Torvalds * a multiple of the mss when it is feasible to do so. 26081da177e4SLinus Torvalds * 26091da177e4SLinus Torvalds * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 26101da177e4SLinus Torvalds * Regular options like TIMESTAMP are taken into account. 26111da177e4SLinus Torvalds */ 26121da177e4SLinus Torvalds u32 __tcp_select_window(struct sock *sk) 26131da177e4SLinus Torvalds { 2614463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 26151da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2616caa20d9aSStephen Hemminger /* MSS for the peer's data. Previous versions used mss_clamp 26171da177e4SLinus Torvalds * here. I don't know if the value based on our guesses 26181da177e4SLinus Torvalds * of peer's MSS is better for the performance. It's more correct 26191da177e4SLinus Torvalds * but may be worse for the performance because of rcv_mss 26201da177e4SLinus Torvalds * fluctuations. --SAW 1998/11/1 26211da177e4SLinus Torvalds */ 2622463c84b9SArnaldo Carvalho de Melo int mss = icsk->icsk_ack.rcv_mss; 26231da177e4SLinus Torvalds int free_space = tcp_space(sk); 262486c1a045SFlorian Westphal int allowed_space = tcp_full_space(sk); 262586c1a045SFlorian Westphal int full_space = min_t(int, tp->window_clamp, allowed_space); 26261da177e4SLinus Torvalds int window; 26271da177e4SLinus Torvalds 262806425c30SEric Dumazet if (unlikely(mss > full_space)) { 26291da177e4SLinus Torvalds mss = full_space; 263006425c30SEric Dumazet if (mss <= 0) 263106425c30SEric Dumazet return 0; 263206425c30SEric Dumazet } 2633b92edbe0SEric Dumazet if (free_space < (full_space >> 1)) { 2634463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.quick = 0; 26351da177e4SLinus Torvalds 2636b8da51ebSEric Dumazet if (tcp_under_memory_pressure(sk)) 2637056834d9SIlpo Järvinen tp->rcv_ssthresh = min(tp->rcv_ssthresh, 2638056834d9SIlpo Järvinen 4U * tp->advmss); 26391da177e4SLinus Torvalds 264086c1a045SFlorian Westphal /* free_space might become our new window, make sure we don't 264186c1a045SFlorian Westphal * increase it due to wscale. 264286c1a045SFlorian Westphal */ 264386c1a045SFlorian Westphal free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale); 264486c1a045SFlorian Westphal 264586c1a045SFlorian Westphal /* if free space is less than mss estimate, or is below 1/16th 264686c1a045SFlorian Westphal * of the maximum allowed, try to move to zero-window, else 264786c1a045SFlorian Westphal * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and 264886c1a045SFlorian Westphal * new incoming data is dropped due to memory limits. 264986c1a045SFlorian Westphal * With large window, mss test triggers way too late in order 265086c1a045SFlorian Westphal * to announce zero window in time before rmem limit kicks in. 265186c1a045SFlorian Westphal */ 265286c1a045SFlorian Westphal if (free_space < (allowed_space >> 4) || free_space < mss) 26531da177e4SLinus Torvalds return 0; 26541da177e4SLinus Torvalds } 26551da177e4SLinus Torvalds 26561da177e4SLinus Torvalds if (free_space > tp->rcv_ssthresh) 26571da177e4SLinus Torvalds free_space = tp->rcv_ssthresh; 26581da177e4SLinus Torvalds 26591da177e4SLinus Torvalds /* Don't do rounding if we are using window scaling, since the 26601da177e4SLinus Torvalds * scaled window will not line up with the MSS boundary anyway. 26611da177e4SLinus Torvalds */ 26621da177e4SLinus Torvalds if (tp->rx_opt.rcv_wscale) { 26631da177e4SLinus Torvalds window = free_space; 26641da177e4SLinus Torvalds 26651da177e4SLinus Torvalds /* Advertise enough space so that it won't get scaled away. 26661da177e4SLinus Torvalds * Import case: prevent zero window announcement if 26671da177e4SLinus Torvalds * 1<<rcv_wscale > mss. 26681da177e4SLinus Torvalds */ 26691935299dSGao Feng window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale)); 26701da177e4SLinus Torvalds } else { 26711935299dSGao Feng window = tp->rcv_wnd; 26721da177e4SLinus Torvalds /* Get the largest window that is a nice multiple of mss. 26731da177e4SLinus Torvalds * Window clamp already applied above. 26741da177e4SLinus Torvalds * If our current window offering is within 1 mss of the 26751da177e4SLinus Torvalds * free space we just keep it. This prevents the divide 26761da177e4SLinus Torvalds * and multiply from happening most of the time. 26771da177e4SLinus Torvalds * We also don't do any window rounding when the free space 26781da177e4SLinus Torvalds * is too small. 26791da177e4SLinus Torvalds */ 26801da177e4SLinus Torvalds if (window <= free_space - mss || window > free_space) 26811935299dSGao Feng window = rounddown(free_space, mss); 268284565070SJohn Heffner else if (mss == full_space && 2683b92edbe0SEric Dumazet free_space > window + (full_space >> 1)) 268484565070SJohn Heffner window = free_space; 26851da177e4SLinus Torvalds } 26861da177e4SLinus Torvalds 26871da177e4SLinus Torvalds return window; 26881da177e4SLinus Torvalds } 26891da177e4SLinus Torvalds 2690cfea5a68SMartin KaFai Lau void tcp_skb_collapse_tstamp(struct sk_buff *skb, 2691082ac2d5SMartin KaFai Lau const struct sk_buff *next_skb) 2692082ac2d5SMartin KaFai Lau { 26930a2cf20cSSoheil Hassas Yeganeh if (unlikely(tcp_has_tx_tstamp(next_skb))) { 26940a2cf20cSSoheil Hassas Yeganeh const struct skb_shared_info *next_shinfo = 26950a2cf20cSSoheil Hassas Yeganeh skb_shinfo(next_skb); 2696082ac2d5SMartin KaFai Lau struct skb_shared_info *shinfo = skb_shinfo(skb); 2697082ac2d5SMartin KaFai Lau 26980a2cf20cSSoheil Hassas Yeganeh shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP; 2699082ac2d5SMartin KaFai Lau shinfo->tskey = next_shinfo->tskey; 27002de8023eSMartin KaFai Lau TCP_SKB_CB(skb)->txstamp_ack |= 27012de8023eSMartin KaFai Lau TCP_SKB_CB(next_skb)->txstamp_ack; 2702082ac2d5SMartin KaFai Lau } 2703082ac2d5SMartin KaFai Lau } 2704082ac2d5SMartin KaFai Lau 27054a17fc3aSIlpo Järvinen /* Collapses two adjacent SKB's during retransmission. */ 2706f8071cdeSEric Dumazet static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) 27071da177e4SLinus Torvalds { 27081da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 270975c119afSEric Dumazet struct sk_buff *next_skb = skb_rb_next(skb); 271013dde04fSWei Yongjun int next_skb_size; 27111da177e4SLinus Torvalds 2712058dc334SIlpo Järvinen next_skb_size = next_skb->len; 27131da177e4SLinus Torvalds 2714058dc334SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); 27151da177e4SLinus Torvalds 2716f8071cdeSEric Dumazet if (next_skb_size) { 2717f8071cdeSEric Dumazet if (next_skb_size <= skb_availroom(skb)) 2718f8071cdeSEric Dumazet skb_copy_bits(next_skb, 0, skb_put(skb, next_skb_size), 2719f8071cdeSEric Dumazet next_skb_size); 2720f8071cdeSEric Dumazet else if (!skb_shift(skb, next_skb, next_skb_size)) 2721f8071cdeSEric Dumazet return false; 2722f8071cdeSEric Dumazet } 27232b7cda9cSEric Dumazet tcp_highest_sack_replace(sk, next_skb, skb); 2724a6963a6bSIlpo Järvinen 27251da177e4SLinus Torvalds /* Update sequence range on original skb. */ 27261da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 27271da177e4SLinus Torvalds 2728e6c7d085SIlpo Järvinen /* Merge over control information. This moves PSH/FIN etc. over */ 27294de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; 27301da177e4SLinus Torvalds 27311da177e4SLinus Torvalds /* All done, get rid of second SKB and account for it so 27321da177e4SLinus Torvalds * packet counting does not break. 27331da177e4SLinus Torvalds */ 27344828e7f4SIlpo Järvinen TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; 2735a643b5d4SMartin KaFai Lau TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor; 2736b7689205SIlpo Järvinen 2737b7689205SIlpo Järvinen /* changed transmit queue under us so clear hints */ 2738ef9da47cSIlpo Järvinen tcp_clear_retrans_hints_partial(tp); 2739ef9da47cSIlpo Järvinen if (next_skb == tp->retransmit_skb_hint) 2740ef9da47cSIlpo Järvinen tp->retransmit_skb_hint = skb; 2741b7689205SIlpo Järvinen 2742797108d1SIlpo Järvinen tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); 2743797108d1SIlpo Järvinen 2744082ac2d5SMartin KaFai Lau tcp_skb_collapse_tstamp(skb, next_skb); 2745082ac2d5SMartin KaFai Lau 274675c119afSEric Dumazet tcp_rtx_queue_unlink_and_free(next_skb, sk); 2747f8071cdeSEric Dumazet return true; 27481da177e4SLinus Torvalds } 27491da177e4SLinus Torvalds 275067edfef7SAndi Kleen /* Check if coalescing SKBs is legal. */ 2751a2a385d6SEric Dumazet static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) 27524a17fc3aSIlpo Järvinen { 27534a17fc3aSIlpo Järvinen if (tcp_skb_pcount(skb) > 1) 2754a2a385d6SEric Dumazet return false; 27554a17fc3aSIlpo Järvinen if (skb_cloned(skb)) 2756a2a385d6SEric Dumazet return false; 27572331ccc5SEric Dumazet /* Some heuristics for collapsing over SACK'd could be invented */ 27584a17fc3aSIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 2759a2a385d6SEric Dumazet return false; 27604a17fc3aSIlpo Järvinen 2761a2a385d6SEric Dumazet return true; 27624a17fc3aSIlpo Järvinen } 27634a17fc3aSIlpo Järvinen 276467edfef7SAndi Kleen /* Collapse packets in the retransmit queue to make to create 276567edfef7SAndi Kleen * less packets on the wire. This is only done on retransmission. 276667edfef7SAndi Kleen */ 27674a17fc3aSIlpo Järvinen static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, 27684a17fc3aSIlpo Järvinen int space) 27694a17fc3aSIlpo Järvinen { 27704a17fc3aSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 27714a17fc3aSIlpo Järvinen struct sk_buff *skb = to, *tmp; 2772a2a385d6SEric Dumazet bool first = true; 27734a17fc3aSIlpo Järvinen 2774e0a1e5b5SEric Dumazet if (!sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse) 27754a17fc3aSIlpo Järvinen return; 27764de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 27774a17fc3aSIlpo Järvinen return; 27784a17fc3aSIlpo Järvinen 277975c119afSEric Dumazet skb_rbtree_walk_from_safe(skb, tmp) { 27804a17fc3aSIlpo Järvinen if (!tcp_can_collapse(sk, skb)) 27814a17fc3aSIlpo Järvinen break; 27824a17fc3aSIlpo Järvinen 2783a643b5d4SMartin KaFai Lau if (!tcp_skb_can_collapse_to(to)) 2784a643b5d4SMartin KaFai Lau break; 2785a643b5d4SMartin KaFai Lau 27864a17fc3aSIlpo Järvinen space -= skb->len; 27874a17fc3aSIlpo Järvinen 27884a17fc3aSIlpo Järvinen if (first) { 2789a2a385d6SEric Dumazet first = false; 27904a17fc3aSIlpo Järvinen continue; 27914a17fc3aSIlpo Järvinen } 27924a17fc3aSIlpo Järvinen 27934a17fc3aSIlpo Järvinen if (space < 0) 27944a17fc3aSIlpo Järvinen break; 27954a17fc3aSIlpo Järvinen 27964a17fc3aSIlpo Järvinen if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) 27974a17fc3aSIlpo Järvinen break; 27984a17fc3aSIlpo Järvinen 2799f8071cdeSEric Dumazet if (!tcp_collapse_retrans(sk, to)) 2800f8071cdeSEric Dumazet break; 28014a17fc3aSIlpo Järvinen } 28024a17fc3aSIlpo Järvinen } 28034a17fc3aSIlpo Järvinen 28041da177e4SLinus Torvalds /* This retransmits one SKB. Policy decisions and retransmit queue 28051da177e4SLinus Torvalds * state updates are done by the caller. Returns non-zero if an 28061da177e4SLinus Torvalds * error occurred which prevented the send. 28071da177e4SLinus Torvalds */ 280810d3be56SEric Dumazet int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) 28091da177e4SLinus Torvalds { 28105d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 281110d3be56SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 28127d227cd2SSridhar Samudrala unsigned int cur_mss; 281310d3be56SEric Dumazet int diff, len, err; 28141da177e4SLinus Torvalds 281510d3be56SEric Dumazet 281610d3be56SEric Dumazet /* Inconclusive MTU probe */ 281710d3be56SEric Dumazet if (icsk->icsk_mtup.probe_size) 28185d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 28195d424d5aSJohn Heffner 28201da177e4SLinus Torvalds /* Do not sent more than we queued. 1/4 is reserved for possible 2821caa20d9aSStephen Hemminger * copying overhead: fragmentation, tunneling, mangling etc. 28221da177e4SLinus Torvalds */ 282314afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) > 2824ffb4d6c8SEric Dumazet min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), 2825ffb4d6c8SEric Dumazet sk->sk_sndbuf)) 28261da177e4SLinus Torvalds return -EAGAIN; 28271da177e4SLinus Torvalds 28281f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 28291f3279aeSEric Dumazet return -EBUSY; 28301f3279aeSEric Dumazet 28311da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 28327f582b24SEric Dumazet if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) { 28337f582b24SEric Dumazet WARN_ON_ONCE(1); 28347f582b24SEric Dumazet return -EINVAL; 28357f582b24SEric Dumazet } 28361da177e4SLinus Torvalds if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 28371da177e4SLinus Torvalds return -ENOMEM; 28381da177e4SLinus Torvalds } 28391da177e4SLinus Torvalds 28407d227cd2SSridhar Samudrala if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 28417d227cd2SSridhar Samudrala return -EHOSTUNREACH; /* Routing failure or similar. */ 28427d227cd2SSridhar Samudrala 28430c54b85fSIlpo Järvinen cur_mss = tcp_current_mss(sk); 28447d227cd2SSridhar Samudrala 28451da177e4SLinus Torvalds /* If receiver has shrunk his window, and skb is out of 28461da177e4SLinus Torvalds * new window, do not retransmit it. The exception is the 28471da177e4SLinus Torvalds * case, when window is shrunk to zero. In this case 28481da177e4SLinus Torvalds * our retransmit serves as a zero window probe. 28491da177e4SLinus Torvalds */ 28509d4fb27dSJoe Perches if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && 28519d4fb27dSJoe Perches TCP_SKB_CB(skb)->seq != tp->snd_una) 28521da177e4SLinus Torvalds return -EAGAIN; 28531da177e4SLinus Torvalds 285410d3be56SEric Dumazet len = cur_mss * segs; 285510d3be56SEric Dumazet if (skb->len > len) { 285675c119afSEric Dumazet if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len, 285775c119afSEric Dumazet cur_mss, GFP_ATOMIC)) 28581da177e4SLinus Torvalds return -ENOMEM; /* We'll try again later. */ 285902276f3cSIlpo Järvinen } else { 2860c52e2421SEric Dumazet if (skb_unclone(skb, GFP_ATOMIC)) 2861c52e2421SEric Dumazet return -ENOMEM; 286210d3be56SEric Dumazet 286310d3be56SEric Dumazet diff = tcp_skb_pcount(skb); 286410d3be56SEric Dumazet tcp_set_skb_tso_segs(skb, cur_mss); 286510d3be56SEric Dumazet diff -= tcp_skb_pcount(skb); 286610d3be56SEric Dumazet if (diff) 286710d3be56SEric Dumazet tcp_adjust_pcount(sk, skb, diff); 286810d3be56SEric Dumazet if (skb->len < cur_mss) 286910d3be56SEric Dumazet tcp_retrans_try_collapse(sk, skb, cur_mss); 28701da177e4SLinus Torvalds } 28711da177e4SLinus Torvalds 287249213555SDaniel Borkmann /* RFC3168, section 6.1.1.1. ECN fallback */ 287349213555SDaniel Borkmann if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN) 287449213555SDaniel Borkmann tcp_ecn_clear_syn(sk, skb); 287549213555SDaniel Borkmann 2876678550c6SYuchung Cheng /* Update global and local TCP statistics. */ 2877678550c6SYuchung Cheng segs = tcp_skb_pcount(skb); 2878678550c6SYuchung Cheng TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs); 2879678550c6SYuchung Cheng if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 2880678550c6SYuchung Cheng __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 2881678550c6SYuchung Cheng tp->total_retrans += segs; 2882fb31c9b9SWei Wang tp->bytes_retrans += skb->len; 2883678550c6SYuchung Cheng 288450bceae9SThomas Graf /* make sure skb->data is aligned on arches that require it 288550bceae9SThomas Graf * and check if ack-trimming & collapsing extended the headroom 288650bceae9SThomas Graf * beyond what csum_start can cover. 288750bceae9SThomas Graf */ 288850bceae9SThomas Graf if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) || 288950bceae9SThomas Graf skb_headroom(skb) >= 0xFFFF)) { 289010a81980SEric Dumazet struct sk_buff *nskb; 289110a81980SEric Dumazet 2892e2080072SEric Dumazet tcp_skb_tsorted_save(skb) { 289310a81980SEric Dumazet nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC); 2894c84a5711SYuchung Cheng err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) : 2895117632e6SEric Dumazet -ENOBUFS; 2896e2080072SEric Dumazet } tcp_skb_tsorted_restore(skb); 2897e2080072SEric Dumazet 28985889e2c0SYousuk Seung if (!err) { 2899ab408b6dSEric Dumazet tcp_update_skb_after_send(sk, skb); 29005889e2c0SYousuk Seung tcp_rate_skb_sent(sk, skb); 29015889e2c0SYousuk Seung } 2902117632e6SEric Dumazet } else { 2903c84a5711SYuchung Cheng err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 2904117632e6SEric Dumazet } 2905c84a5711SYuchung Cheng 2906a31ad29eSLawrence Brakmo if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RETRANS_CB_FLAG)) 2907a31ad29eSLawrence Brakmo tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RETRANS_CB, 2908a31ad29eSLawrence Brakmo TCP_SKB_CB(skb)->seq, segs, err); 2909a31ad29eSLawrence Brakmo 2910fc9f3501SEric Dumazet if (likely(!err)) { 2911c84a5711SYuchung Cheng TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS; 2912e086101bSCong Wang trace_tcp_retransmit_skb(sk, skb); 2913678550c6SYuchung Cheng } else if (err != -EBUSY) { 2914678550c6SYuchung Cheng NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL); 2915fc9f3501SEric Dumazet } 2916c84a5711SYuchung Cheng return err; 291793b174adSYuchung Cheng } 291893b174adSYuchung Cheng 291910d3be56SEric Dumazet int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) 292093b174adSYuchung Cheng { 292193b174adSYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 292210d3be56SEric Dumazet int err = __tcp_retransmit_skb(sk, skb, segs); 29231da177e4SLinus Torvalds 29241da177e4SLinus Torvalds if (err == 0) { 29251da177e4SLinus Torvalds #if FASTRETRANS_DEBUG > 0 29261da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 2927e87cc472SJoe Perches net_dbg_ratelimited("retrans_out leaked\n"); 29281da177e4SLinus Torvalds } 29291da177e4SLinus Torvalds #endif 29301da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 29311da177e4SLinus Torvalds tp->retrans_out += tcp_skb_pcount(skb); 29321da177e4SLinus Torvalds 29331da177e4SLinus Torvalds /* Save stamp of the first retransmit. */ 29341da177e4SLinus Torvalds if (!tp->retrans_stamp) 29357faee5c0SEric Dumazet tp->retrans_stamp = tcp_skb_timestamp(skb); 29361da177e4SLinus Torvalds 29371da177e4SLinus Torvalds } 29386e08d5e3SYuchung Cheng 29396e08d5e3SYuchung Cheng if (tp->undo_retrans < 0) 29406e08d5e3SYuchung Cheng tp->undo_retrans = 0; 29416e08d5e3SYuchung Cheng tp->undo_retrans += tcp_skb_pcount(skb); 29421da177e4SLinus Torvalds return err; 29431da177e4SLinus Torvalds } 29441da177e4SLinus Torvalds 29451da177e4SLinus Torvalds /* This gets called after a retransmit timeout, and the initially 29461da177e4SLinus Torvalds * retransmitted data is acknowledged. It tries to continue 29471da177e4SLinus Torvalds * resending the rest of the retransmit queue, until either 29481da177e4SLinus Torvalds * we've sent it all or the congestion window limit is reached. 29491da177e4SLinus Torvalds */ 29501da177e4SLinus Torvalds void tcp_xmit_retransmit_queue(struct sock *sk) 29511da177e4SLinus Torvalds { 29526687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 2953b9f1f1ceSEric Dumazet struct sk_buff *skb, *rtx_head, *hole = NULL; 29541da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2955840a3cbeSYuchung Cheng u32 max_segs; 295661eb55f4SIlpo Järvinen int mib_idx; 29576a438bbeSStephen Hemminger 295845e77d31SIlpo Järvinen if (!tp->packets_out) 295945e77d31SIlpo Järvinen return; 296045e77d31SIlpo Järvinen 296175c119afSEric Dumazet rtx_head = tcp_rtx_queue_head(sk); 2962b9f1f1ceSEric Dumazet skb = tp->retransmit_skb_hint ?: rtx_head; 2963ed6e7268SNeal Cardwell max_segs = tcp_tso_segs(sk, tcp_current_mss(sk)); 296475c119afSEric Dumazet skb_rbtree_walk_from(skb) { 2965dca0aaf8SEric Dumazet __u8 sacked; 296610d3be56SEric Dumazet int segs; 29671da177e4SLinus Torvalds 2968218af599SEric Dumazet if (tcp_pacing_check(sk)) 2969218af599SEric Dumazet break; 2970218af599SEric Dumazet 29716a438bbeSStephen Hemminger /* we could do better than to assign each time */ 297251456b29SIan Morris if (!hole) 29736a438bbeSStephen Hemminger tp->retransmit_skb_hint = skb; 29746a438bbeSStephen Hemminger 297510d3be56SEric Dumazet segs = tp->snd_cwnd - tcp_packets_in_flight(tp); 297610d3be56SEric Dumazet if (segs <= 0) 29771da177e4SLinus Torvalds return; 2978dca0aaf8SEric Dumazet sacked = TCP_SKB_CB(skb)->sacked; 2979a3d2e9f8SEric Dumazet /* In case tcp_shift_skb_data() have aggregated large skbs, 2980a3d2e9f8SEric Dumazet * we need to make sure not sending too bigs TSO packets 2981a3d2e9f8SEric Dumazet */ 2982a3d2e9f8SEric Dumazet segs = min_t(int, segs, max_segs); 29830e1c54c2SIlpo Järvinen 2984840a3cbeSYuchung Cheng if (tp->retrans_out >= tp->lost_out) { 2985006f582cSIlpo Järvinen break; 29860e1c54c2SIlpo Järvinen } else if (!(sacked & TCPCB_LOST)) { 298751456b29SIan Morris if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) 29880e1c54c2SIlpo Järvinen hole = skb; 298961eb55f4SIlpo Järvinen continue; 29901da177e4SLinus Torvalds 29910e1c54c2SIlpo Järvinen } else { 29920e1c54c2SIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Loss) 29930e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFASTRETRANS; 29940e1c54c2SIlpo Järvinen else 29950e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; 29960e1c54c2SIlpo Järvinen } 29970e1c54c2SIlpo Järvinen 29980e1c54c2SIlpo Järvinen if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) 299961eb55f4SIlpo Järvinen continue; 300040b215e5SPavel Emelyanov 3001f9616c35SEric Dumazet if (tcp_small_queue_check(sk, skb, 1)) 3002f9616c35SEric Dumazet return; 3003f9616c35SEric Dumazet 300410d3be56SEric Dumazet if (tcp_retransmit_skb(sk, skb, segs)) 30051da177e4SLinus Torvalds return; 300624ab6becSYuchung Cheng 3007de1d6578SYuchung Cheng NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb)); 30081da177e4SLinus Torvalds 3009684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 3010a262f0cdSNandita Dukkipati tp->prr_out += tcp_skb_pcount(skb); 3011a262f0cdSNandita Dukkipati 301275c119afSEric Dumazet if (skb == rtx_head && 301357dde7f7SYuchung Cheng icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT) 3014463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 30153f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, 30163f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 30171da177e4SLinus Torvalds } 30181da177e4SLinus Torvalds } 30191da177e4SLinus Torvalds 3020d83769a5SEric Dumazet /* We allow to exceed memory limits for FIN packets to expedite 3021d83769a5SEric Dumazet * connection tear down and (memory) recovery. 3022845704a5SEric Dumazet * Otherwise tcp_send_fin() could be tempted to either delay FIN 3023845704a5SEric Dumazet * or even be forced to close flow without any FIN. 3024a6c5ea4cSEric Dumazet * In general, we want to allow one skb per socket to avoid hangs 3025a6c5ea4cSEric Dumazet * with edge trigger epoll() 3026d83769a5SEric Dumazet */ 3027a6c5ea4cSEric Dumazet void sk_forced_mem_schedule(struct sock *sk, int size) 3028d83769a5SEric Dumazet { 3029e805605cSJohannes Weiner int amt; 3030d83769a5SEric Dumazet 3031d83769a5SEric Dumazet if (size <= sk->sk_forward_alloc) 3032d83769a5SEric Dumazet return; 3033d83769a5SEric Dumazet amt = sk_mem_pages(size); 3034d83769a5SEric Dumazet sk->sk_forward_alloc += amt * SK_MEM_QUANTUM; 3035e805605cSJohannes Weiner sk_memory_allocated_add(sk, amt); 3036e805605cSJohannes Weiner 3037baac50bbSJohannes Weiner if (mem_cgroup_sockets_enabled && sk->sk_memcg) 3038baac50bbSJohannes Weiner mem_cgroup_charge_skmem(sk->sk_memcg, amt); 3039d83769a5SEric Dumazet } 3040d83769a5SEric Dumazet 3041845704a5SEric Dumazet /* Send a FIN. The caller locks the socket for us. 3042845704a5SEric Dumazet * We should try to send a FIN packet really hard, but eventually give up. 30431da177e4SLinus Torvalds */ 30441da177e4SLinus Torvalds void tcp_send_fin(struct sock *sk) 30451da177e4SLinus Torvalds { 3046845704a5SEric Dumazet struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk); 30471da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 30481da177e4SLinus Torvalds 3049845704a5SEric Dumazet /* Optimization, tack on the FIN if we have one skb in write queue and 3050845704a5SEric Dumazet * this skb was not yet sent, or we are under memory pressure. 3051845704a5SEric Dumazet * Note: in the latter case, FIN packet will be sent after a timeout, 3052845704a5SEric Dumazet * as TCP stack thinks it has already been transmitted. 30531da177e4SLinus Torvalds */ 305475c119afSEric Dumazet if (!tskb && tcp_under_memory_pressure(sk)) 305575c119afSEric Dumazet tskb = skb_rb_last(&sk->tcp_rtx_queue); 305675c119afSEric Dumazet 305775c119afSEric Dumazet if (tskb) { 3058845704a5SEric Dumazet coalesce: 3059845704a5SEric Dumazet TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN; 3060845704a5SEric Dumazet TCP_SKB_CB(tskb)->end_seq++; 30611da177e4SLinus Torvalds tp->write_seq++; 306275c119afSEric Dumazet if (tcp_write_queue_empty(sk)) { 3063845704a5SEric Dumazet /* This means tskb was already sent. 3064845704a5SEric Dumazet * Pretend we included the FIN on previous transmit. 3065845704a5SEric Dumazet * We need to set tp->snd_nxt to the value it would have 3066845704a5SEric Dumazet * if FIN had been sent. This is because retransmit path 3067845704a5SEric Dumazet * does not change tp->snd_nxt. 3068845704a5SEric Dumazet */ 3069845704a5SEric Dumazet tp->snd_nxt++; 3070845704a5SEric Dumazet return; 3071845704a5SEric Dumazet } 30721da177e4SLinus Torvalds } else { 3073845704a5SEric Dumazet skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation); 3074845704a5SEric Dumazet if (unlikely(!skb)) { 3075845704a5SEric Dumazet if (tskb) 3076845704a5SEric Dumazet goto coalesce; 3077845704a5SEric Dumazet return; 30781da177e4SLinus Torvalds } 3079e2080072SEric Dumazet INIT_LIST_HEAD(&skb->tcp_tsorted_anchor); 3080d83769a5SEric Dumazet skb_reserve(skb, MAX_TCP_HEADER); 3081a6c5ea4cSEric Dumazet sk_forced_mem_schedule(sk, skb->truesize); 30821da177e4SLinus Torvalds /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 3083e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tp->write_seq, 3084a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_FIN); 30851da177e4SLinus Torvalds tcp_queue_skb(sk, skb); 30861da177e4SLinus Torvalds } 3087845704a5SEric Dumazet __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF); 30881da177e4SLinus Torvalds } 30891da177e4SLinus Torvalds 30901da177e4SLinus Torvalds /* We get here when a process closes a file descriptor (either due to 30911da177e4SLinus Torvalds * an explicit close() or as a byproduct of exit()'ing) and there 30921da177e4SLinus Torvalds * was unread data in the receive queue. This behavior is recommended 309365bb723cSGerrit Renker * by RFC 2525, section 2.17. -DaveM 30941da177e4SLinus Torvalds */ 3095dd0fc66fSAl Viro void tcp_send_active_reset(struct sock *sk, gfp_t priority) 30961da177e4SLinus Torvalds { 30971da177e4SLinus Torvalds struct sk_buff *skb; 30981da177e4SLinus Torvalds 30997cc2b043SGao Feng TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); 31007cc2b043SGao Feng 31011da177e4SLinus Torvalds /* NOTE: No TCP options attached and we never retransmit this. */ 31021da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, priority); 31031da177e4SLinus Torvalds if (!skb) { 31044e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 31051da177e4SLinus Torvalds return; 31061da177e4SLinus Torvalds } 31071da177e4SLinus Torvalds 31081da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 31091da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 3110e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), 3111a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_RST); 31129a568de4SEric Dumazet tcp_mstamp_refresh(tcp_sk(sk)); 31131da177e4SLinus Torvalds /* Send it off. */ 3114dfb4b9dcSDavid S. Miller if (tcp_transmit_skb(sk, skb, 0, priority)) 31154e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 3116c24b14c4SSong Liu 3117c24b14c4SSong Liu /* skb of trace_tcp_send_reset() keeps the skb that caused RST, 3118c24b14c4SSong Liu * skb here is different to the troublesome skb, so use NULL 3119c24b14c4SSong Liu */ 3120c24b14c4SSong Liu trace_tcp_send_reset(sk, NULL); 31211da177e4SLinus Torvalds } 31221da177e4SLinus Torvalds 312367edfef7SAndi Kleen /* Send a crossed SYN-ACK during socket establishment. 312467edfef7SAndi Kleen * WARNING: This routine must only be called when we have already sent 31251da177e4SLinus Torvalds * a SYN packet that crossed the incoming SYN that caused this routine 31261da177e4SLinus Torvalds * to get called. If this assumption fails then the initial rcv_wnd 31271da177e4SLinus Torvalds * and rcv_wscale values will not be correct. 31281da177e4SLinus Torvalds */ 31291da177e4SLinus Torvalds int tcp_send_synack(struct sock *sk) 31301da177e4SLinus Torvalds { 31311da177e4SLinus Torvalds struct sk_buff *skb; 31321da177e4SLinus Torvalds 313375c119afSEric Dumazet skb = tcp_rtx_queue_head(sk); 313451456b29SIan Morris if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 313575c119afSEric Dumazet pr_err("%s: wrong queue state\n", __func__); 31361da177e4SLinus Torvalds return -EFAULT; 31371da177e4SLinus Torvalds } 31384de075e0SEric Dumazet if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { 31391da177e4SLinus Torvalds if (skb_cloned(skb)) { 3140e2080072SEric Dumazet struct sk_buff *nskb; 3141e2080072SEric Dumazet 3142e2080072SEric Dumazet tcp_skb_tsorted_save(skb) { 3143e2080072SEric Dumazet nskb = skb_copy(skb, GFP_ATOMIC); 3144e2080072SEric Dumazet } tcp_skb_tsorted_restore(skb); 314551456b29SIan Morris if (!nskb) 31461da177e4SLinus Torvalds return -ENOMEM; 3147e2080072SEric Dumazet INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor); 314875c119afSEric Dumazet tcp_rtx_queue_unlink_and_free(skb, sk); 3149f4a775d1SEric Dumazet __skb_header_release(nskb); 315075c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb); 31513ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 31523ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 31531da177e4SLinus Torvalds skb = nskb; 31541da177e4SLinus Torvalds } 31551da177e4SLinus Torvalds 31564de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; 3157735d3831SFlorian Westphal tcp_ecn_send_synack(sk, skb); 31581da177e4SLinus Torvalds } 3159dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 31601da177e4SLinus Torvalds } 31611da177e4SLinus Torvalds 31624aea39c1SEric Dumazet /** 31634aea39c1SEric Dumazet * tcp_make_synack - Prepare a SYN-ACK. 31644aea39c1SEric Dumazet * sk: listener socket 31654aea39c1SEric Dumazet * dst: dst entry attached to the SYNACK 31664aea39c1SEric Dumazet * req: request_sock pointer 31674aea39c1SEric Dumazet * 31684aea39c1SEric Dumazet * Allocate one skb and build a SYNACK packet. 31694aea39c1SEric Dumazet * @dst is consumed : Caller should not use it again. 31704aea39c1SEric Dumazet */ 31715d062de7SEric Dumazet struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst, 3172e6b4d113SWilliam Allen Simpson struct request_sock *req, 3173ca6fb065SEric Dumazet struct tcp_fastopen_cookie *foc, 3174b3d05147SEric Dumazet enum tcp_synack_type synack_type) 31751da177e4SLinus Torvalds { 31762e6599cbSArnaldo Carvalho de Melo struct inet_request_sock *ireq = inet_rsk(req); 31775d062de7SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 317880f03e27SEric Dumazet struct tcp_md5sig_key *md5 = NULL; 31795d062de7SEric Dumazet struct tcp_out_options opts; 31805d062de7SEric Dumazet struct sk_buff *skb; 3181bd0388aeSWilliam Allen Simpson int tcp_header_size; 31825d062de7SEric Dumazet struct tcphdr *th; 3183f5fff5dcSTom Quetchenbach int mss; 31841da177e4SLinus Torvalds 3185ca6fb065SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 31864aea39c1SEric Dumazet if (unlikely(!skb)) { 31874aea39c1SEric Dumazet dst_release(dst); 31881da177e4SLinus Torvalds return NULL; 31894aea39c1SEric Dumazet } 31901da177e4SLinus Torvalds /* Reserve space for headers. */ 31911da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 31921da177e4SLinus Torvalds 3193b3d05147SEric Dumazet switch (synack_type) { 3194b3d05147SEric Dumazet case TCP_SYNACK_NORMAL: 31959e17f8a4SEric Dumazet skb_set_owner_w(skb, req_to_sk(req)); 3196b3d05147SEric Dumazet break; 3197b3d05147SEric Dumazet case TCP_SYNACK_COOKIE: 3198b3d05147SEric Dumazet /* Under synflood, we do not attach skb to a socket, 3199b3d05147SEric Dumazet * to avoid false sharing. 3200b3d05147SEric Dumazet */ 3201b3d05147SEric Dumazet break; 3202b3d05147SEric Dumazet case TCP_SYNACK_FASTOPEN: 3203ca6fb065SEric Dumazet /* sk is a const pointer, because we want to express multiple 3204ca6fb065SEric Dumazet * cpu might call us concurrently. 3205ca6fb065SEric Dumazet * sk->sk_wmem_alloc in an atomic, we can promote to rw. 3206ca6fb065SEric Dumazet */ 3207ca6fb065SEric Dumazet skb_set_owner_w(skb, (struct sock *)sk); 3208b3d05147SEric Dumazet break; 3209ca6fb065SEric Dumazet } 32104aea39c1SEric Dumazet skb_dst_set(skb, dst); 32111da177e4SLinus Torvalds 32123541f9e8SEric Dumazet mss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 3213f5fff5dcSTom Quetchenbach 321433ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 32158b5f12d0SFlorian Westphal #ifdef CONFIG_SYN_COOKIES 32168b5f12d0SFlorian Westphal if (unlikely(req->cookie_ts)) 3217d3edd06eSEric Dumazet skb->skb_mstamp_ns = cookie_init_timestamp(req); 32188b5f12d0SFlorian Westphal else 32198b5f12d0SFlorian Westphal #endif 3220d3edd06eSEric Dumazet skb->skb_mstamp_ns = tcp_clock_ns(); 322180f03e27SEric Dumazet 322280f03e27SEric Dumazet #ifdef CONFIG_TCP_MD5SIG 322380f03e27SEric Dumazet rcu_read_lock(); 3224fd3a154aSEric Dumazet md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req)); 322580f03e27SEric Dumazet #endif 322658d607d3SEric Dumazet skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4); 322760e2a778SUrsula Braun tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5, 322860e2a778SUrsula Braun foc) + sizeof(*th); 322933ad798cSAdam Langley 3230aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 3231aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 32321da177e4SLinus Torvalds 3233ea1627c2SEric Dumazet th = (struct tcphdr *)skb->data; 32341da177e4SLinus Torvalds memset(th, 0, sizeof(struct tcphdr)); 32351da177e4SLinus Torvalds th->syn = 1; 32361da177e4SLinus Torvalds th->ack = 1; 32376ac705b1SEric Dumazet tcp_ecn_make_synack(req, th); 3238b44084c2SEric Dumazet th->source = htons(ireq->ir_num); 3239634fb979SEric Dumazet th->dest = ireq->ir_rmt_port; 3240e05a90ecSJamal Hadi Salim skb->mark = ireq->ir_mark; 32413b117750SEric Dumazet skb->ip_summed = CHECKSUM_PARTIAL; 32423b117750SEric Dumazet th->seq = htonl(tcp_rsk(req)->snt_isn); 32438336886fSJerry Chu /* XXX data is queued and acked as is. No buffer/window check */ 32448336886fSJerry Chu th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt); 32451da177e4SLinus Torvalds 32461da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 3247ed53d0abSEric Dumazet th->window = htons(min(req->rsk_rcv_wnd, 65535U)); 32485d062de7SEric Dumazet tcp_options_write((__be32 *)(th + 1), NULL, &opts); 32491da177e4SLinus Torvalds th->doff = (tcp_header_size >> 2); 325090bbcc60SEric Dumazet __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS); 3251cfb6eeb4SYOSHIFUJI Hideaki 3252cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 3253cfb6eeb4SYOSHIFUJI Hideaki /* Okay, we have all we need - do the md5 hash if needed */ 325480f03e27SEric Dumazet if (md5) 3255bd0388aeSWilliam Allen Simpson tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, 325639f8e58eSEric Dumazet md5, req_to_sk(req), skb); 325780f03e27SEric Dumazet rcu_read_unlock(); 3258cfb6eeb4SYOSHIFUJI Hideaki #endif 3259cfb6eeb4SYOSHIFUJI Hideaki 3260b50edd78SEric Dumazet /* Do not fool tcpdump (if any), clean our debris */ 32612456e855SThomas Gleixner skb->tstamp = 0; 32621da177e4SLinus Torvalds return skb; 32631da177e4SLinus Torvalds } 32644bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_make_synack); 32651da177e4SLinus Torvalds 326681164413SDaniel Borkmann static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst) 326781164413SDaniel Borkmann { 326881164413SDaniel Borkmann struct inet_connection_sock *icsk = inet_csk(sk); 326981164413SDaniel Borkmann const struct tcp_congestion_ops *ca; 327081164413SDaniel Borkmann u32 ca_key = dst_metric(dst, RTAX_CC_ALGO); 327181164413SDaniel Borkmann 327281164413SDaniel Borkmann if (ca_key == TCP_CA_UNSPEC) 327381164413SDaniel Borkmann return; 327481164413SDaniel Borkmann 327581164413SDaniel Borkmann rcu_read_lock(); 327681164413SDaniel Borkmann ca = tcp_ca_find_key(ca_key); 327781164413SDaniel Borkmann if (likely(ca && try_module_get(ca->owner))) { 327881164413SDaniel Borkmann module_put(icsk->icsk_ca_ops->owner); 327981164413SDaniel Borkmann icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst); 328081164413SDaniel Borkmann icsk->icsk_ca_ops = ca; 328181164413SDaniel Borkmann } 328281164413SDaniel Borkmann rcu_read_unlock(); 328381164413SDaniel Borkmann } 328481164413SDaniel Borkmann 328567edfef7SAndi Kleen /* Do all connect socket setups that can be done AF independent. */ 3286f7e56a76Sstephen hemminger static void tcp_connect_init(struct sock *sk) 32871da177e4SLinus Torvalds { 3288cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 32891da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 32901da177e4SLinus Torvalds __u8 rcv_wscale; 329113d3b1ebSLawrence Brakmo u32 rcv_wnd; 32921da177e4SLinus Torvalds 32931da177e4SLinus Torvalds /* We'll fix this up when we get a response from the other end. 32941da177e4SLinus Torvalds * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 32951da177e4SLinus Torvalds */ 32965d2ed052SEric Dumazet tp->tcp_header_len = sizeof(struct tcphdr); 32975d2ed052SEric Dumazet if (sock_net(sk)->ipv4.sysctl_tcp_timestamps) 32985d2ed052SEric Dumazet tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED; 32991da177e4SLinus Torvalds 3300cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 330100db4124SIan Morris if (tp->af_specific->md5_lookup(sk, sk)) 3302cfb6eeb4SYOSHIFUJI Hideaki tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 3303cfb6eeb4SYOSHIFUJI Hideaki #endif 3304cfb6eeb4SYOSHIFUJI Hideaki 33051da177e4SLinus Torvalds /* If user gave his TCP_MAXSEG, record it to clamp */ 33061da177e4SLinus Torvalds if (tp->rx_opt.user_mss) 33071da177e4SLinus Torvalds tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 33081da177e4SLinus Torvalds tp->max_window = 0; 33095d424d5aSJohn Heffner tcp_mtup_init(sk); 33101da177e4SLinus Torvalds tcp_sync_mss(sk, dst_mtu(dst)); 33111da177e4SLinus Torvalds 331281164413SDaniel Borkmann tcp_ca_dst_init(sk, dst); 331381164413SDaniel Borkmann 33141da177e4SLinus Torvalds if (!tp->window_clamp) 33151da177e4SLinus Torvalds tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 33163541f9e8SEric Dumazet tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 3317f5fff5dcSTom Quetchenbach 33181da177e4SLinus Torvalds tcp_initialize_rcv_mss(sk); 33191da177e4SLinus Torvalds 3320e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 3321e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 3322e88c64f0SHagen Paul Pfeifer (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) 3323e88c64f0SHagen Paul Pfeifer tp->window_clamp = tcp_full_space(sk); 3324e88c64f0SHagen Paul Pfeifer 332513d3b1ebSLawrence Brakmo rcv_wnd = tcp_rwnd_init_bpf(sk); 332613d3b1ebSLawrence Brakmo if (rcv_wnd == 0) 332713d3b1ebSLawrence Brakmo rcv_wnd = dst_metric(dst, RTAX_INITRWND); 332813d3b1ebSLawrence Brakmo 3329ceef9ab6SEric Dumazet tcp_select_initial_window(sk, tcp_full_space(sk), 33301da177e4SLinus Torvalds tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 33311da177e4SLinus Torvalds &tp->rcv_wnd, 33321da177e4SLinus Torvalds &tp->window_clamp, 33339bb37ef0SEric Dumazet sock_net(sk)->ipv4.sysctl_tcp_window_scaling, 333431d12926Slaurent chavey &rcv_wscale, 333513d3b1ebSLawrence Brakmo rcv_wnd); 33361da177e4SLinus Torvalds 33371da177e4SLinus Torvalds tp->rx_opt.rcv_wscale = rcv_wscale; 33381da177e4SLinus Torvalds tp->rcv_ssthresh = tp->rcv_wnd; 33391da177e4SLinus Torvalds 33401da177e4SLinus Torvalds sk->sk_err = 0; 33411da177e4SLinus Torvalds sock_reset_flag(sk, SOCK_DONE); 33421da177e4SLinus Torvalds tp->snd_wnd = 0; 3343ee7537b6SHantzis Fotis tcp_init_wl(tp, 0); 33447f582b24SEric Dumazet tcp_write_queue_purge(sk); 33451da177e4SLinus Torvalds tp->snd_una = tp->write_seq; 33461da177e4SLinus Torvalds tp->snd_sml = tp->write_seq; 334733f5f57eSIlpo Järvinen tp->snd_up = tp->write_seq; 3348370816aeSPavel Emelyanov tp->snd_nxt = tp->write_seq; 3349ee995283SPavel Emelyanov 3350ee995283SPavel Emelyanov if (likely(!tp->repair)) 33511da177e4SLinus Torvalds tp->rcv_nxt = 0; 3352c7781a6eSAndrew Vagin else 335370eabf0eSEric Dumazet tp->rcv_tstamp = tcp_jiffies32; 3354ee995283SPavel Emelyanov tp->rcv_wup = tp->rcv_nxt; 3355ee995283SPavel Emelyanov tp->copied_seq = tp->rcv_nxt; 33561da177e4SLinus Torvalds 33578550f328SLawrence Brakmo inet_csk(sk)->icsk_rto = tcp_timeout_init(sk); 3358463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_retransmits = 0; 33591da177e4SLinus Torvalds tcp_clear_retrans(tp); 33601da177e4SLinus Torvalds } 33611da177e4SLinus Torvalds 3362783237e8SYuchung Cheng static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) 3363783237e8SYuchung Cheng { 3364783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3365783237e8SYuchung Cheng struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 3366783237e8SYuchung Cheng 3367783237e8SYuchung Cheng tcb->end_seq += skb->len; 3368f4a775d1SEric Dumazet __skb_header_release(skb); 3369783237e8SYuchung Cheng sk->sk_wmem_queued += skb->truesize; 3370783237e8SYuchung Cheng sk_mem_charge(sk, skb->truesize); 3371783237e8SYuchung Cheng tp->write_seq = tcb->end_seq; 3372783237e8SYuchung Cheng tp->packets_out += tcp_skb_pcount(skb); 3373783237e8SYuchung Cheng } 3374783237e8SYuchung Cheng 3375783237e8SYuchung Cheng /* Build and send a SYN with data and (cached) Fast Open cookie. However, 3376783237e8SYuchung Cheng * queue a data-only packet after the regular SYN, such that regular SYNs 3377783237e8SYuchung Cheng * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges 3378783237e8SYuchung Cheng * only the SYN sequence, the data are retransmitted in the first ACK. 3379783237e8SYuchung Cheng * If cookie is not cached or other error occurs, falls back to send a 3380783237e8SYuchung Cheng * regular SYN with Fast Open cookie request option. 3381783237e8SYuchung Cheng */ 3382783237e8SYuchung Cheng static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) 3383783237e8SYuchung Cheng { 3384783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3385783237e8SYuchung Cheng struct tcp_fastopen_request *fo = tp->fastopen_req; 3386065263f4SWei Wang int space, err = 0; 3387355a901eSEric Dumazet struct sk_buff *syn_data; 3388783237e8SYuchung Cheng 338967da22d2SYuchung Cheng tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ 3390065263f4SWei Wang if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie)) 3391783237e8SYuchung Cheng goto fallback; 3392783237e8SYuchung Cheng 3393783237e8SYuchung Cheng /* MSS for SYN-data is based on cached MSS and bounded by PMTU and 3394783237e8SYuchung Cheng * user-MSS. Reserve maximum option space for middleboxes that add 3395783237e8SYuchung Cheng * private TCP options. The cost is reduced data space in SYN :( 3396783237e8SYuchung Cheng */ 33973541f9e8SEric Dumazet tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp); 33983541f9e8SEric Dumazet 33991b63edd6SYuchung Cheng space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) - 3400783237e8SYuchung Cheng MAX_TCP_OPTION_SPACE; 3401783237e8SYuchung Cheng 3402f5ddcbbbSEric Dumazet space = min_t(size_t, space, fo->size); 3403f5ddcbbbSEric Dumazet 3404f5ddcbbbSEric Dumazet /* limit to order-0 allocations */ 3405f5ddcbbbSEric Dumazet space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER)); 3406f5ddcbbbSEric Dumazet 3407eb934478SEric Dumazet syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false); 3408355a901eSEric Dumazet if (!syn_data) 3409783237e8SYuchung Cheng goto fallback; 3410355a901eSEric Dumazet syn_data->ip_summed = CHECKSUM_PARTIAL; 3411355a901eSEric Dumazet memcpy(syn_data->cb, syn->cb, sizeof(syn->cb)); 341207e100f9SEric Dumazet if (space) { 341307e100f9SEric Dumazet int copied = copy_from_iter(skb_put(syn_data, space), space, 341457be5bdaSAl Viro &fo->data->msg_iter); 341557be5bdaSAl Viro if (unlikely(!copied)) { 3416ba233b34SEric Dumazet tcp_skb_tsorted_anchor_cleanup(syn_data); 3417355a901eSEric Dumazet kfree_skb(syn_data); 3418783237e8SYuchung Cheng goto fallback; 3419783237e8SYuchung Cheng } 342057be5bdaSAl Viro if (copied != space) { 342157be5bdaSAl Viro skb_trim(syn_data, copied); 342257be5bdaSAl Viro space = copied; 342357be5bdaSAl Viro } 342407e100f9SEric Dumazet } 3425355a901eSEric Dumazet /* No more data pending in inet_wait_for_connect() */ 3426355a901eSEric Dumazet if (space == fo->size) 3427355a901eSEric Dumazet fo->data = NULL; 3428355a901eSEric Dumazet fo->copied = space; 3429783237e8SYuchung Cheng 3430355a901eSEric Dumazet tcp_connect_queue_skb(sk, syn_data); 34310f87230dSFrancis Yan if (syn_data->len) 34320f87230dSFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_BUSY); 3433355a901eSEric Dumazet 3434355a901eSEric Dumazet err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation); 3435355a901eSEric Dumazet 3436d3edd06eSEric Dumazet syn->skb_mstamp_ns = syn_data->skb_mstamp_ns; 3437355a901eSEric Dumazet 3438355a901eSEric Dumazet /* Now full SYN+DATA was cloned and sent (or not), 3439355a901eSEric Dumazet * remove the SYN from the original skb (syn_data) 3440355a901eSEric Dumazet * we keep in write queue in case of a retransmit, as we 3441355a901eSEric Dumazet * also have the SYN packet (with no data) in the same queue. 3442431a9124SEric Dumazet */ 3443355a901eSEric Dumazet TCP_SKB_CB(syn_data)->seq++; 3444355a901eSEric Dumazet TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH; 3445355a901eSEric Dumazet if (!err) { 344667da22d2SYuchung Cheng tp->syn_data = (fo->copied > 0); 344775c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data); 3448f19c29e3SYuchung Cheng NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT); 3449783237e8SYuchung Cheng goto done; 3450783237e8SYuchung Cheng } 3451783237e8SYuchung Cheng 345275c119afSEric Dumazet /* data was not sent, put it in write_queue */ 345375c119afSEric Dumazet __skb_queue_tail(&sk->sk_write_queue, syn_data); 3454b5b7db8dSEric Dumazet tp->packets_out -= tcp_skb_pcount(syn_data); 3455b5b7db8dSEric Dumazet 3456783237e8SYuchung Cheng fallback: 3457783237e8SYuchung Cheng /* Send a regular SYN with Fast Open cookie request option */ 3458783237e8SYuchung Cheng if (fo->cookie.len > 0) 3459783237e8SYuchung Cheng fo->cookie.len = 0; 3460783237e8SYuchung Cheng err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); 3461783237e8SYuchung Cheng if (err) 3462783237e8SYuchung Cheng tp->syn_fastopen = 0; 3463783237e8SYuchung Cheng done: 3464783237e8SYuchung Cheng fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ 3465783237e8SYuchung Cheng return err; 3466783237e8SYuchung Cheng } 3467783237e8SYuchung Cheng 346867edfef7SAndi Kleen /* Build a SYN and send it off. */ 34691da177e4SLinus Torvalds int tcp_connect(struct sock *sk) 34701da177e4SLinus Torvalds { 34711da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 34721da177e4SLinus Torvalds struct sk_buff *buff; 3473ee586811SEric Paris int err; 34741da177e4SLinus Torvalds 3475de525be2SLawrence Brakmo tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB, 0, NULL); 34768ba60924SEric Dumazet 34778ba60924SEric Dumazet if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 34788ba60924SEric Dumazet return -EHOSTUNREACH; /* Routing failure or similar. */ 34798ba60924SEric Dumazet 34801da177e4SLinus Torvalds tcp_connect_init(sk); 34811da177e4SLinus Torvalds 34822b916477SAndrey Vagin if (unlikely(tp->repair)) { 34832b916477SAndrey Vagin tcp_finish_connect(sk, NULL); 34842b916477SAndrey Vagin return 0; 34852b916477SAndrey Vagin } 34862b916477SAndrey Vagin 3487eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true); 3488355a901eSEric Dumazet if (unlikely(!buff)) 34891da177e4SLinus Torvalds return -ENOBUFS; 34901da177e4SLinus Torvalds 3491a3433f35SChangli Gao tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); 34929a568de4SEric Dumazet tcp_mstamp_refresh(tp); 34939a568de4SEric Dumazet tp->retrans_stamp = tcp_time_stamp(tp); 3494783237e8SYuchung Cheng tcp_connect_queue_skb(sk, buff); 3495735d3831SFlorian Westphal tcp_ecn_send_syn(sk, buff); 349675c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, buff); 34971da177e4SLinus Torvalds 3498783237e8SYuchung Cheng /* Send off SYN; include data in Fast Open. */ 3499783237e8SYuchung Cheng err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : 3500783237e8SYuchung Cheng tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); 3501ee586811SEric Paris if (err == -ECONNREFUSED) 3502ee586811SEric Paris return err; 3503bd37a088SWei Yongjun 3504bd37a088SWei Yongjun /* We change tp->snd_nxt after the tcp_transmit_skb() call 3505bd37a088SWei Yongjun * in order to make this packet get counted in tcpOutSegs. 3506bd37a088SWei Yongjun */ 3507bd37a088SWei Yongjun tp->snd_nxt = tp->write_seq; 3508bd37a088SWei Yongjun tp->pushed_seq = tp->write_seq; 3509b5b7db8dSEric Dumazet buff = tcp_send_head(sk); 3510b5b7db8dSEric Dumazet if (unlikely(buff)) { 3511b5b7db8dSEric Dumazet tp->snd_nxt = TCP_SKB_CB(buff)->seq; 3512b5b7db8dSEric Dumazet tp->pushed_seq = TCP_SKB_CB(buff)->seq; 3513b5b7db8dSEric Dumazet } 351481cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); 35151da177e4SLinus Torvalds 35161da177e4SLinus Torvalds /* Timer for repeating the SYN until an answer. */ 35173f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 35183f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 35191da177e4SLinus Torvalds return 0; 35201da177e4SLinus Torvalds } 35214bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_connect); 35221da177e4SLinus Torvalds 35231da177e4SLinus Torvalds /* Send out a delayed ack, the caller does the policy checking 35241da177e4SLinus Torvalds * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 35251da177e4SLinus Torvalds * for details. 35261da177e4SLinus Torvalds */ 35271da177e4SLinus Torvalds void tcp_send_delayed_ack(struct sock *sk) 35281da177e4SLinus Torvalds { 3529463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 3530463c84b9SArnaldo Carvalho de Melo int ato = icsk->icsk_ack.ato; 35311da177e4SLinus Torvalds unsigned long timeout; 35321da177e4SLinus Torvalds 35331da177e4SLinus Torvalds if (ato > TCP_DELACK_MIN) { 3534463c84b9SArnaldo Carvalho de Melo const struct tcp_sock *tp = tcp_sk(sk); 35351da177e4SLinus Torvalds int max_ato = HZ / 2; 35361da177e4SLinus Torvalds 3537056834d9SIlpo Järvinen if (icsk->icsk_ack.pingpong || 3538056834d9SIlpo Järvinen (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 35391da177e4SLinus Torvalds max_ato = TCP_DELACK_MAX; 35401da177e4SLinus Torvalds 35411da177e4SLinus Torvalds /* Slow path, intersegment interval is "high". */ 35421da177e4SLinus Torvalds 35431da177e4SLinus Torvalds /* If some rtt estimate is known, use it to bound delayed ack. 3544463c84b9SArnaldo Carvalho de Melo * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 35451da177e4SLinus Torvalds * directly. 35461da177e4SLinus Torvalds */ 3547740b0f18SEric Dumazet if (tp->srtt_us) { 3548740b0f18SEric Dumazet int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3), 3549740b0f18SEric Dumazet TCP_DELACK_MIN); 35501da177e4SLinus Torvalds 35511da177e4SLinus Torvalds if (rtt < max_ato) 35521da177e4SLinus Torvalds max_ato = rtt; 35531da177e4SLinus Torvalds } 35541da177e4SLinus Torvalds 35551da177e4SLinus Torvalds ato = min(ato, max_ato); 35561da177e4SLinus Torvalds } 35571da177e4SLinus Torvalds 35581da177e4SLinus Torvalds /* Stay within the limit we were given */ 35591da177e4SLinus Torvalds timeout = jiffies + ato; 35601da177e4SLinus Torvalds 35611da177e4SLinus Torvalds /* Use new timeout only if there wasn't a older one earlier. */ 3562463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 35631da177e4SLinus Torvalds /* If delack timer was blocked or is about to expire, 35641da177e4SLinus Torvalds * send ACK now. 35651da177e4SLinus Torvalds */ 3566463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.blocked || 3567463c84b9SArnaldo Carvalho de Melo time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 35681da177e4SLinus Torvalds tcp_send_ack(sk); 35691da177e4SLinus Torvalds return; 35701da177e4SLinus Torvalds } 35711da177e4SLinus Torvalds 3572463c84b9SArnaldo Carvalho de Melo if (!time_before(timeout, icsk->icsk_ack.timeout)) 3573463c84b9SArnaldo Carvalho de Melo timeout = icsk->icsk_ack.timeout; 35741da177e4SLinus Torvalds } 3575463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 3576463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.timeout = timeout; 3577463c84b9SArnaldo Carvalho de Melo sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 35781da177e4SLinus Torvalds } 35791da177e4SLinus Torvalds 35801da177e4SLinus Torvalds /* This routine sends an ack and also updates the window. */ 35812987babbSYuchung Cheng void __tcp_send_ack(struct sock *sk, u32 rcv_nxt) 35821da177e4SLinus Torvalds { 35831da177e4SLinus Torvalds struct sk_buff *buff; 35841da177e4SLinus Torvalds 3585058dc334SIlpo Järvinen /* If we have been reset, we may not send again. */ 3586058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3587058dc334SIlpo Järvinen return; 3588058dc334SIlpo Järvinen 35891da177e4SLinus Torvalds /* We are not putting this on the write queue, so 35901da177e4SLinus Torvalds * tcp_transmit_skb() will set the ownership to this 35911da177e4SLinus Torvalds * sock. 35921da177e4SLinus Torvalds */ 35937450aaf6SEric Dumazet buff = alloc_skb(MAX_TCP_HEADER, 35947450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 35957450aaf6SEric Dumazet if (unlikely(!buff)) { 3596463c84b9SArnaldo Carvalho de Melo inet_csk_schedule_ack(sk); 3597463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; 35983f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 35993f421baaSArnaldo Carvalho de Melo TCP_DELACK_MAX, TCP_RTO_MAX); 36001da177e4SLinus Torvalds return; 36011da177e4SLinus Torvalds } 36021da177e4SLinus Torvalds 36031da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 36041da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 3605a3433f35SChangli Gao tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); 36061da177e4SLinus Torvalds 360798781965SEric Dumazet /* We do not want pure acks influencing TCP Small Queues or fq/pacing 360898781965SEric Dumazet * too much. 360998781965SEric Dumazet * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784 361098781965SEric Dumazet */ 361198781965SEric Dumazet skb_set_tcp_pure_ack(buff); 361298781965SEric Dumazet 36131da177e4SLinus Torvalds /* Send it off, this clears delayed acks for us. */ 36142987babbSYuchung Cheng __tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0, rcv_nxt); 36151da177e4SLinus Torvalds } 361627cde44aSYuchung Cheng EXPORT_SYMBOL_GPL(__tcp_send_ack); 36172987babbSYuchung Cheng 36182987babbSYuchung Cheng void tcp_send_ack(struct sock *sk) 36192987babbSYuchung Cheng { 36202987babbSYuchung Cheng __tcp_send_ack(sk, tcp_sk(sk)->rcv_nxt); 36211da177e4SLinus Torvalds } 36221da177e4SLinus Torvalds 36231da177e4SLinus Torvalds /* This routine sends a packet with an out of date sequence 36241da177e4SLinus Torvalds * number. It assumes the other end will try to ack it. 36251da177e4SLinus Torvalds * 36261da177e4SLinus Torvalds * Question: what should we make while urgent mode? 36271da177e4SLinus Torvalds * 4.4BSD forces sending single byte of data. We cannot send 36281da177e4SLinus Torvalds * out of window data, because we have SND.NXT==SND.MAX... 36291da177e4SLinus Torvalds * 36301da177e4SLinus Torvalds * Current solution: to send TWO zero-length segments in urgent mode: 36311da177e4SLinus Torvalds * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 36321da177e4SLinus Torvalds * out-of-date with SND.UNA-1 to probe window. 36331da177e4SLinus Torvalds */ 3634e520af48SEric Dumazet static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib) 36351da177e4SLinus Torvalds { 36361da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 36371da177e4SLinus Torvalds struct sk_buff *skb; 36381da177e4SLinus Torvalds 36391da177e4SLinus Torvalds /* We don't queue it, tcp_transmit_skb() sets ownership. */ 36407450aaf6SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, 36417450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 364251456b29SIan Morris if (!skb) 36431da177e4SLinus Torvalds return -1; 36441da177e4SLinus Torvalds 36451da177e4SLinus Torvalds /* Reserve space for headers and set control bits. */ 36461da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 36471da177e4SLinus Torvalds /* Use a previous sequence. This should cause the other 36481da177e4SLinus Torvalds * end to send an ack. Don't queue or clone SKB, just 36491da177e4SLinus Torvalds * send it. 36501da177e4SLinus Torvalds */ 3651a3433f35SChangli Gao tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); 3652e2e8009fSRenato Westphal NET_INC_STATS(sock_net(sk), mib); 36537450aaf6SEric Dumazet return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0); 36541da177e4SLinus Torvalds } 36551da177e4SLinus Torvalds 3656385e2070SEric Dumazet /* Called from setsockopt( ... TCP_REPAIR ) */ 3657ee995283SPavel Emelyanov void tcp_send_window_probe(struct sock *sk) 3658ee995283SPavel Emelyanov { 3659ee995283SPavel Emelyanov if (sk->sk_state == TCP_ESTABLISHED) { 3660ee995283SPavel Emelyanov tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; 36619a568de4SEric Dumazet tcp_mstamp_refresh(tcp_sk(sk)); 3662e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE); 3663ee995283SPavel Emelyanov } 3664ee995283SPavel Emelyanov } 3665ee995283SPavel Emelyanov 366667edfef7SAndi Kleen /* Initiate keepalive or window probe from timer. */ 3667e520af48SEric Dumazet int tcp_write_wakeup(struct sock *sk, int mib) 36681da177e4SLinus Torvalds { 36691da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 36701da177e4SLinus Torvalds struct sk_buff *skb; 36711da177e4SLinus Torvalds 3672058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3673058dc334SIlpo Järvinen return -1; 3674058dc334SIlpo Järvinen 367500db4124SIan Morris skb = tcp_send_head(sk); 367600db4124SIan Morris if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { 36771da177e4SLinus Torvalds int err; 36780c54b85fSIlpo Järvinen unsigned int mss = tcp_current_mss(sk); 367990840defSIlpo Järvinen unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 36801da177e4SLinus Torvalds 36811da177e4SLinus Torvalds if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 36821da177e4SLinus Torvalds tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 36831da177e4SLinus Torvalds 36841da177e4SLinus Torvalds /* We are probing the opening of a window 36851da177e4SLinus Torvalds * but the window size is != 0 36861da177e4SLinus Torvalds * must have been a result SWS avoidance ( sender ) 36871da177e4SLinus Torvalds */ 36881da177e4SLinus Torvalds if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 36891da177e4SLinus Torvalds skb->len > mss) { 36901da177e4SLinus Torvalds seg_size = min(seg_size, mss); 36914de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 369275c119afSEric Dumazet if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE, 369375c119afSEric Dumazet skb, seg_size, mss, GFP_ATOMIC)) 36941da177e4SLinus Torvalds return -1; 36951da177e4SLinus Torvalds } else if (!tcp_skb_pcount(skb)) 36965bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss); 36971da177e4SLinus Torvalds 36984de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 3699dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 370066f5fe62SIlpo Järvinen if (!err) 370166f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 37021da177e4SLinus Torvalds return err; 37031da177e4SLinus Torvalds } else { 370433f5f57eSIlpo Järvinen if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) 3705e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 1, mib); 3706e520af48SEric Dumazet return tcp_xmit_probe_skb(sk, 0, mib); 37071da177e4SLinus Torvalds } 37081da177e4SLinus Torvalds } 37091da177e4SLinus Torvalds 37101da177e4SLinus Torvalds /* A window probe timeout has occurred. If window is not closed send 37111da177e4SLinus Torvalds * a partial packet else a zero probe. 37121da177e4SLinus Torvalds */ 37131da177e4SLinus Torvalds void tcp_send_probe0(struct sock *sk) 37141da177e4SLinus Torvalds { 3715463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 37161da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 3717c6214a97SNikolay Borisov struct net *net = sock_net(sk); 3718fcdd1cf4SEric Dumazet unsigned long probe_max; 37191da177e4SLinus Torvalds int err; 37201da177e4SLinus Torvalds 3721e520af48SEric Dumazet err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE); 37221da177e4SLinus Torvalds 372375c119afSEric Dumazet if (tp->packets_out || tcp_write_queue_empty(sk)) { 37241da177e4SLinus Torvalds /* Cancel probe timer, if it is not required. */ 37256687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 0; 3726463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff = 0; 37271da177e4SLinus Torvalds return; 37281da177e4SLinus Torvalds } 37291da177e4SLinus Torvalds 37301da177e4SLinus Torvalds if (err <= 0) { 3731c6214a97SNikolay Borisov if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2) 3732463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff++; 37336687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out++; 3734fcdd1cf4SEric Dumazet probe_max = TCP_RTO_MAX; 37351da177e4SLinus Torvalds } else { 37361da177e4SLinus Torvalds /* If packet was not sent due to local congestion, 37376687e988SArnaldo Carvalho de Melo * do not backoff and do not remember icsk_probes_out. 37381da177e4SLinus Torvalds * Let local senders to fight for local resources. 37391da177e4SLinus Torvalds * 37401da177e4SLinus Torvalds * Use accumulated backoff yet. 37411da177e4SLinus Torvalds */ 37426687e988SArnaldo Carvalho de Melo if (!icsk->icsk_probes_out) 37436687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 1; 3744fcdd1cf4SEric Dumazet probe_max = TCP_RESOURCE_PROBE_INTERVAL; 37451da177e4SLinus Torvalds } 3746fcdd1cf4SEric Dumazet inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 374721c8fe99SEric Dumazet tcp_probe0_when(sk, probe_max), 3748fcdd1cf4SEric Dumazet TCP_RTO_MAX); 37491da177e4SLinus Torvalds } 37505db92c99SOctavian Purdila 3751ea3bea3aSEric Dumazet int tcp_rtx_synack(const struct sock *sk, struct request_sock *req) 37525db92c99SOctavian Purdila { 37535db92c99SOctavian Purdila const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific; 37545db92c99SOctavian Purdila struct flowi fl; 37555db92c99SOctavian Purdila int res; 37565db92c99SOctavian Purdila 375758d607d3SEric Dumazet tcp_rsk(req)->txhash = net_tx_rndhash(); 3758b3d05147SEric Dumazet res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL); 37595db92c99SOctavian Purdila if (!res) { 376090bbcc60SEric Dumazet __TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); 376102a1d6e7SEric Dumazet __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 37627e32b443SYuchung Cheng if (unlikely(tcp_passive_fastopen(sk))) 37637e32b443SYuchung Cheng tcp_sk(sk)->total_retrans++; 3764cf34ce3dSSong Liu trace_tcp_retransmit_synack(sk, req); 37655db92c99SOctavian Purdila } 37665db92c99SOctavian Purdila return res; 37675db92c99SOctavian Purdila } 37685db92c99SOctavian Purdila EXPORT_SYMBOL(tcp_rtx_synack); 3769