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 988*a7a25630SEric Dumazet static void tcp_update_skb_after_send(struct sock *sk, struct sk_buff *skb, 989*a7a25630SEric Dumazet u64 prior_wstamp) 990e2080072SEric Dumazet { 991ab408b6dSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 992ab408b6dSEric Dumazet 993d3edd06eSEric Dumazet skb->skb_mstamp_ns = tp->tcp_wstamp_ns; 994ab408b6dSEric Dumazet if (sk->sk_pacing_status != SK_PACING_NONE) { 99576a9ebe8SEric Dumazet unsigned long rate = sk->sk_pacing_rate; 996ab408b6dSEric Dumazet 997ab408b6dSEric Dumazet /* Original sch_fq does not pace first 10 MSS 998ab408b6dSEric Dumazet * Note that tp->data_segs_out overflows after 2^32 packets, 999ab408b6dSEric Dumazet * this is a minor annoyance. 1000ab408b6dSEric Dumazet */ 100176a9ebe8SEric Dumazet if (rate != ~0UL && rate && tp->data_segs_out >= 10) { 1002*a7a25630SEric Dumazet u64 len_ns = div64_ul((u64)skb->len * NSEC_PER_SEC, rate); 1003*a7a25630SEric Dumazet u64 credit = tp->tcp_wstamp_ns - prior_wstamp; 1004*a7a25630SEric Dumazet 1005*a7a25630SEric Dumazet /* take into account OS jitter */ 1006*a7a25630SEric Dumazet len_ns -= min_t(u64, len_ns / 2, credit); 1007*a7a25630SEric Dumazet tp->tcp_wstamp_ns += len_ns; 1008c092dd5fSEric Dumazet 1009c092dd5fSEric Dumazet tcp_internal_pacing(sk); 1010ab408b6dSEric Dumazet } 1011ab408b6dSEric Dumazet } 1012e2080072SEric Dumazet list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue); 1013e2080072SEric Dumazet } 1014e2080072SEric Dumazet 10151da177e4SLinus Torvalds /* This routine actually transmits TCP packets queued in by 10161da177e4SLinus Torvalds * tcp_do_sendmsg(). This is used by both the initial 10171da177e4SLinus Torvalds * transmission and possible later retransmissions. 10181da177e4SLinus Torvalds * All SKB's seen here are completely headerless. It is our 10191da177e4SLinus Torvalds * job to build the TCP header, and pass the packet down to 10201da177e4SLinus Torvalds * IP so it can do the same plus pass the packet off to the 10211da177e4SLinus Torvalds * device. 10221da177e4SLinus Torvalds * 10231da177e4SLinus Torvalds * We are working here with either a clone of the original 10241da177e4SLinus Torvalds * SKB, or a fresh unique copy made by the retransmit engine. 10251da177e4SLinus Torvalds */ 10262987babbSYuchung Cheng static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, 10272987babbSYuchung Cheng int clone_it, gfp_t gfp_mask, u32 rcv_nxt) 10281da177e4SLinus Torvalds { 10296687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 1030dfb4b9dcSDavid S. Miller struct inet_sock *inet; 1031dfb4b9dcSDavid S. Miller struct tcp_sock *tp; 1032dfb4b9dcSDavid S. Miller struct tcp_skb_cb *tcb; 103333ad798cSAdam Langley struct tcp_out_options opts; 103495c96174SEric Dumazet unsigned int tcp_options_size, tcp_header_size; 10358c72c65bSEric Dumazet struct sk_buff *oskb = NULL; 1036cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 10371da177e4SLinus Torvalds struct tcphdr *th; 1038*a7a25630SEric Dumazet u64 prior_wstamp; 10391da177e4SLinus Torvalds int err; 10401da177e4SLinus Torvalds 1041dfb4b9dcSDavid S. Miller BUG_ON(!skb || !tcp_skb_pcount(skb)); 10426f094b9eSLawrence Brakmo tp = tcp_sk(sk); 1043dfb4b9dcSDavid S. Miller 1044ccdbb6e9SEric Dumazet if (clone_it) { 10456f094b9eSLawrence Brakmo TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq 10466f094b9eSLawrence Brakmo - tp->snd_una; 10478c72c65bSEric Dumazet oskb = skb; 1048e2080072SEric Dumazet 1049e2080072SEric Dumazet tcp_skb_tsorted_save(oskb) { 1050e2080072SEric Dumazet if (unlikely(skb_cloned(oskb))) 1051e2080072SEric Dumazet skb = pskb_copy(oskb, gfp_mask); 1052dfb4b9dcSDavid S. Miller else 1053e2080072SEric Dumazet skb = skb_clone(oskb, gfp_mask); 1054e2080072SEric Dumazet } tcp_skb_tsorted_restore(oskb); 1055e2080072SEric Dumazet 1056dfb4b9dcSDavid S. Miller if (unlikely(!skb)) 1057dfb4b9dcSDavid S. Miller return -ENOBUFS; 1058dfb4b9dcSDavid S. Miller } 10595f6188a8SEric Dumazet 1060*a7a25630SEric Dumazet prior_wstamp = tp->tcp_wstamp_ns; 10615f6188a8SEric Dumazet tp->tcp_wstamp_ns = max(tp->tcp_wstamp_ns, tp->tcp_clock_cache); 10625f6188a8SEric Dumazet 1063d3edd06eSEric Dumazet skb->skb_mstamp_ns = tp->tcp_wstamp_ns; 1064dfb4b9dcSDavid S. Miller 1065dfb4b9dcSDavid S. Miller inet = inet_sk(sk); 1066dfb4b9dcSDavid S. Miller tcb = TCP_SKB_CB(skb); 106733ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 10681da177e4SLinus Torvalds 10694de075e0SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) 107033ad798cSAdam Langley tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); 107133ad798cSAdam Langley else 107233ad798cSAdam Langley tcp_options_size = tcp_established_options(sk, skb, &opts, 107333ad798cSAdam Langley &md5); 107433ad798cSAdam Langley tcp_header_size = tcp_options_size + sizeof(struct tcphdr); 10751da177e4SLinus Torvalds 1076547669d4SEric Dumazet /* if no packet is in qdisc/device queue, then allow XPS to select 1077b2532eb9SEric Dumazet * another queue. We can be called from tcp_tsq_handler() 107873a6bab5SEric Dumazet * which holds one reference to sk. 1079b2532eb9SEric Dumazet * 1080b2532eb9SEric Dumazet * TODO: Ideally, in-flight pure ACK packets should not matter here. 1081b2532eb9SEric Dumazet * One way to get this would be to set skb->truesize = 2 on them. 1082547669d4SEric Dumazet */ 1083b2532eb9SEric Dumazet skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1); 10841da177e4SLinus Torvalds 108538ab52e8SEric Dumazet /* If we had to use memory reserve to allocate this skb, 108638ab52e8SEric Dumazet * this might cause drops if packet is looped back : 108738ab52e8SEric Dumazet * Other socket might not have SOCK_MEMALLOC. 108838ab52e8SEric Dumazet * Packets not looped back do not care about pfmemalloc. 108938ab52e8SEric Dumazet */ 109038ab52e8SEric Dumazet skb->pfmemalloc = 0; 109138ab52e8SEric Dumazet 1092aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 1093aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 109446d3ceabSEric Dumazet 109546d3ceabSEric Dumazet skb_orphan(skb); 109646d3ceabSEric Dumazet skb->sk = sk; 10971d2077acSEric Dumazet skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree; 1098b73c3d0eSTom Herbert skb_set_hash_from_sk(skb, sk); 109914afee4bSReshetova, Elena refcount_add(skb->truesize, &sk->sk_wmem_alloc); 11001da177e4SLinus Torvalds 1101c3a2e837SJulian Anastasov skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm); 1102c3a2e837SJulian Anastasov 11031da177e4SLinus Torvalds /* Build TCP header and checksum it. */ 1104ea1627c2SEric Dumazet th = (struct tcphdr *)skb->data; 1105c720c7e8SEric Dumazet th->source = inet->inet_sport; 1106c720c7e8SEric Dumazet th->dest = inet->inet_dport; 11071da177e4SLinus Torvalds th->seq = htonl(tcb->seq); 11082987babbSYuchung Cheng th->ack_seq = htonl(rcv_nxt); 1109df7a3b07SAl Viro *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | 11104de075e0SEric Dumazet tcb->tcp_flags); 1111dfb4b9dcSDavid S. Miller 11121da177e4SLinus Torvalds th->check = 0; 11131da177e4SLinus Torvalds th->urg_ptr = 0; 11141da177e4SLinus Torvalds 111533f5f57eSIlpo Järvinen /* The urg_mode check is necessary during a below snd_una win probe */ 11167691367dSHerbert Xu if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { 11177691367dSHerbert Xu if (before(tp->snd_up, tcb->seq + 0x10000)) { 11181da177e4SLinus Torvalds th->urg_ptr = htons(tp->snd_up - tcb->seq); 11191da177e4SLinus Torvalds th->urg = 1; 11207691367dSHerbert Xu } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { 11210eae88f3SEric Dumazet th->urg_ptr = htons(0xFFFF); 11227691367dSHerbert Xu th->urg = 1; 11237691367dSHerbert Xu } 11241da177e4SLinus Torvalds } 11251da177e4SLinus Torvalds 1126bd0388aeSWilliam Allen Simpson tcp_options_write((__be32 *)(th + 1), tp, &opts); 112751466a75SEric Dumazet skb_shinfo(skb)->gso_type = sk->sk_gso_type; 1128ea1627c2SEric Dumazet if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) { 1129ea1627c2SEric Dumazet th->window = htons(tcp_select_window(sk)); 1130ea1627c2SEric Dumazet tcp_ecn_send(sk, skb, th, tcp_header_size); 1131ea1627c2SEric Dumazet } else { 1132ea1627c2SEric Dumazet /* RFC1323: The window in SYN & SYN/ACK segments 1133ea1627c2SEric Dumazet * is never scaled. 1134ea1627c2SEric Dumazet */ 1135ea1627c2SEric Dumazet th->window = htons(min(tp->rcv_wnd, 65535U)); 1136ea1627c2SEric Dumazet } 1137cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 1138cfb6eeb4SYOSHIFUJI Hideaki /* Calculate the MD5 hash, as we have all we need now */ 1139cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 1140a465419bSEric Dumazet sk_nocaps_add(sk, NETIF_F_GSO_MASK); 1141bd0388aeSWilliam Allen Simpson tp->af_specific->calc_md5_hash(opts.hash_location, 114239f8e58eSEric Dumazet md5, sk, skb); 1143cfb6eeb4SYOSHIFUJI Hideaki } 1144cfb6eeb4SYOSHIFUJI Hideaki #endif 1145cfb6eeb4SYOSHIFUJI Hideaki 1146bb296246SHerbert Xu icsk->icsk_af_ops->send_check(sk, skb); 11471da177e4SLinus Torvalds 11484de075e0SEric Dumazet if (likely(tcb->tcp_flags & TCPHDR_ACK)) 114927cde44aSYuchung Cheng tcp_event_ack_sent(sk, tcp_skb_pcount(skb), rcv_nxt); 11501da177e4SLinus Torvalds 1151a44d6eacSMartin KaFai Lau if (skb->len != tcp_header_size) { 1152cf533ea5SEric Dumazet tcp_event_data_sent(tp, sk); 1153a44d6eacSMartin KaFai Lau tp->data_segs_out += tcp_skb_pcount(skb); 1154ba113c3aSWei Wang tp->bytes_sent += skb->len - tcp_header_size; 1155a44d6eacSMartin KaFai Lau } 11561da177e4SLinus Torvalds 1157bd37a088SWei Yongjun if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) 1158aa2ea058STom Herbert TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, 1159aa2ea058STom Herbert tcp_skb_pcount(skb)); 11601da177e4SLinus Torvalds 11612efd055cSMarcelo Ricardo Leitner tp->segs_out += tcp_skb_pcount(skb); 1162f69ad292SEric Dumazet /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */ 1163cd7d8498SEric Dumazet skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb); 1164f69ad292SEric Dumazet skb_shinfo(skb)->gso_size = tcp_skb_mss(skb); 1165cd7d8498SEric Dumazet 1166d3edd06eSEric Dumazet /* Leave earliest departure time in skb->tstamp (skb->skb_mstamp_ns) */ 1167971f10ecSEric Dumazet 1168971f10ecSEric Dumazet /* Cleanup our debris for IP stacks */ 1169971f10ecSEric Dumazet memset(skb->cb, 0, max(sizeof(struct inet_skb_parm), 1170971f10ecSEric Dumazet sizeof(struct inet6_skb_parm))); 1171971f10ecSEric Dumazet 1172b0270e91SEric Dumazet err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl); 11737faee5c0SEric Dumazet 11748c72c65bSEric Dumazet if (unlikely(err > 0)) { 11755ee2c941SChristoph Paasch tcp_enter_cwr(sk); 11768c72c65bSEric Dumazet err = net_xmit_eval(err); 11778c72c65bSEric Dumazet } 1178fc225799SEric Dumazet if (!err && oskb) { 1179*a7a25630SEric Dumazet tcp_update_skb_after_send(sk, oskb, prior_wstamp); 1180fc225799SEric Dumazet tcp_rate_skb_sent(sk, oskb); 1181fc225799SEric Dumazet } 11828c72c65bSEric Dumazet return err; 11831da177e4SLinus Torvalds } 11841da177e4SLinus Torvalds 11852987babbSYuchung Cheng static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, 11862987babbSYuchung Cheng gfp_t gfp_mask) 11872987babbSYuchung Cheng { 11882987babbSYuchung Cheng return __tcp_transmit_skb(sk, skb, clone_it, gfp_mask, 11892987babbSYuchung Cheng tcp_sk(sk)->rcv_nxt); 11902987babbSYuchung Cheng } 11912987babbSYuchung Cheng 119267edfef7SAndi Kleen /* This routine just queues the buffer for sending. 11931da177e4SLinus Torvalds * 11941da177e4SLinus Torvalds * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, 11951da177e4SLinus Torvalds * otherwise socket can stall. 11961da177e4SLinus Torvalds */ 11971da177e4SLinus Torvalds static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) 11981da177e4SLinus Torvalds { 11991da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 12001da177e4SLinus Torvalds 12011da177e4SLinus Torvalds /* Advance write_seq and place onto the write_queue. */ 12021da177e4SLinus Torvalds tp->write_seq = TCP_SKB_CB(skb)->end_seq; 1203f4a775d1SEric Dumazet __skb_header_release(skb); 1204fe067e8aSDavid S. Miller tcp_add_write_queue_tail(sk, skb); 12053ab224beSHideo Aoki sk->sk_wmem_queued += skb->truesize; 12063ab224beSHideo Aoki sk_mem_charge(sk, skb->truesize); 12071da177e4SLinus Torvalds } 12081da177e4SLinus Torvalds 120967edfef7SAndi Kleen /* Initialize TSO segments for a packet. */ 12105bbb432cSEric Dumazet static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now) 1211f6302d1dSDavid S. Miller { 12124a64fd6cSEric Dumazet if (skb->len <= mss_now) { 1213f6302d1dSDavid S. Miller /* Avoid the costly divide in the normal 1214f6302d1dSDavid S. Miller * non-TSO case. 1215f6302d1dSDavid S. Miller */ 1216cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, 1); 1217f69ad292SEric Dumazet TCP_SKB_CB(skb)->tcp_gso_size = 0; 1218f6302d1dSDavid S. Miller } else { 1219cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now)); 1220f69ad292SEric Dumazet TCP_SKB_CB(skb)->tcp_gso_size = mss_now; 12211da177e4SLinus Torvalds } 12221da177e4SLinus Torvalds } 12231da177e4SLinus Torvalds 1224797108d1SIlpo Järvinen /* Pcount in the middle of the write queue got changed, we need to do various 1225797108d1SIlpo Järvinen * tweaks to fix counters 1226797108d1SIlpo Järvinen */ 1227cf533ea5SEric Dumazet static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) 1228797108d1SIlpo Järvinen { 1229797108d1SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1230797108d1SIlpo Järvinen 1231797108d1SIlpo Järvinen tp->packets_out -= decr; 1232797108d1SIlpo Järvinen 1233797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 1234797108d1SIlpo Järvinen tp->sacked_out -= decr; 1235797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 1236797108d1SIlpo Järvinen tp->retrans_out -= decr; 1237797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 1238797108d1SIlpo Järvinen tp->lost_out -= decr; 1239797108d1SIlpo Järvinen 1240797108d1SIlpo Järvinen /* Reno case is special. Sigh... */ 1241797108d1SIlpo Järvinen if (tcp_is_reno(tp) && decr > 0) 1242797108d1SIlpo Järvinen tp->sacked_out -= min_t(u32, tp->sacked_out, decr); 1243797108d1SIlpo Järvinen 1244797108d1SIlpo Järvinen if (tp->lost_skb_hint && 1245797108d1SIlpo Järvinen before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && 1246713bafeaSYuchung Cheng (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) 1247797108d1SIlpo Järvinen tp->lost_cnt_hint -= decr; 1248797108d1SIlpo Järvinen 1249797108d1SIlpo Järvinen tcp_verify_left_out(tp); 1250797108d1SIlpo Järvinen } 1251797108d1SIlpo Järvinen 12520a2cf20cSSoheil Hassas Yeganeh static bool tcp_has_tx_tstamp(const struct sk_buff *skb) 12530a2cf20cSSoheil Hassas Yeganeh { 12540a2cf20cSSoheil Hassas Yeganeh return TCP_SKB_CB(skb)->txstamp_ack || 12550a2cf20cSSoheil Hassas Yeganeh (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP); 12560a2cf20cSSoheil Hassas Yeganeh } 12570a2cf20cSSoheil Hassas Yeganeh 1258490cc7d0SWillem de Bruijn static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2) 1259490cc7d0SWillem de Bruijn { 1260490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo = skb_shinfo(skb); 1261490cc7d0SWillem de Bruijn 12620a2cf20cSSoheil Hassas Yeganeh if (unlikely(tcp_has_tx_tstamp(skb)) && 1263490cc7d0SWillem de Bruijn !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) { 1264490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo2 = skb_shinfo(skb2); 1265490cc7d0SWillem de Bruijn u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP; 1266490cc7d0SWillem de Bruijn 1267490cc7d0SWillem de Bruijn shinfo->tx_flags &= ~tsflags; 1268490cc7d0SWillem de Bruijn shinfo2->tx_flags |= tsflags; 1269490cc7d0SWillem de Bruijn swap(shinfo->tskey, shinfo2->tskey); 1270b51e13faSMartin KaFai Lau TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack; 1271b51e13faSMartin KaFai Lau TCP_SKB_CB(skb)->txstamp_ack = 0; 1272490cc7d0SWillem de Bruijn } 1273490cc7d0SWillem de Bruijn } 1274490cc7d0SWillem de Bruijn 1275a166140eSMartin KaFai Lau static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2) 1276a166140eSMartin KaFai Lau { 1277a166140eSMartin KaFai Lau TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor; 1278a166140eSMartin KaFai Lau TCP_SKB_CB(skb)->eor = 0; 1279a166140eSMartin KaFai Lau } 1280a166140eSMartin KaFai Lau 128175c119afSEric Dumazet /* Insert buff after skb on the write or rtx queue of sk. */ 128275c119afSEric Dumazet static void tcp_insert_write_queue_after(struct sk_buff *skb, 128375c119afSEric Dumazet struct sk_buff *buff, 128475c119afSEric Dumazet struct sock *sk, 128575c119afSEric Dumazet enum tcp_queue tcp_queue) 128675c119afSEric Dumazet { 128775c119afSEric Dumazet if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE) 128875c119afSEric Dumazet __skb_queue_after(&sk->sk_write_queue, skb, buff); 128975c119afSEric Dumazet else 129075c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, buff); 129175c119afSEric Dumazet } 129275c119afSEric Dumazet 12931da177e4SLinus Torvalds /* Function to create two new TCP segments. Shrinks the given segment 12941da177e4SLinus Torvalds * to the specified size and appends a new segment with the rest of the 12951da177e4SLinus Torvalds * packet to the list. This won't be called frequently, I hope. 12961da177e4SLinus Torvalds * Remember, these are still headerless SKBs at this point. 12971da177e4SLinus Torvalds */ 129875c119afSEric Dumazet int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue, 129975c119afSEric Dumazet struct sk_buff *skb, u32 len, 13006cc55e09SOctavian Purdila unsigned int mss_now, gfp_t gfp) 13011da177e4SLinus Torvalds { 13021da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 13031da177e4SLinus Torvalds struct sk_buff *buff; 13046475be16SDavid S. Miller int nsize, old_factor; 1305b60b49eaSHerbert Xu int nlen; 13069ce01461SIlpo Järvinen u8 flags; 13071da177e4SLinus Torvalds 13082fceec13SIlpo Järvinen if (WARN_ON(len > skb->len)) 13092fceec13SIlpo Järvinen return -EINVAL; 13106a438bbeSStephen Hemminger 13111da177e4SLinus Torvalds nsize = skb_headlen(skb) - len; 13121da177e4SLinus Torvalds if (nsize < 0) 13131da177e4SLinus Torvalds nsize = 0; 13141da177e4SLinus Torvalds 13156cc55e09SOctavian Purdila if (skb_unclone(skb, gfp)) 13161da177e4SLinus Torvalds return -ENOMEM; 13171da177e4SLinus Torvalds 13181da177e4SLinus Torvalds /* Get a new skb... force flag on. */ 1319eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, nsize, gfp, true); 132051456b29SIan Morris if (!buff) 13211da177e4SLinus Torvalds return -ENOMEM; /* We'll just try again later. */ 1322ef5cb973SHerbert Xu 13233ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 13243ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1325b60b49eaSHerbert Xu nlen = skb->len - len - nsize; 1326b60b49eaSHerbert Xu buff->truesize += nlen; 1327b60b49eaSHerbert Xu skb->truesize -= nlen; 13281da177e4SLinus Torvalds 13291da177e4SLinus Torvalds /* Correct the sequence numbers. */ 13301da177e4SLinus Torvalds TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 13311da177e4SLinus Torvalds TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 13321da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 13331da177e4SLinus Torvalds 13341da177e4SLinus Torvalds /* PSH and FIN should only be set in the second packet. */ 13354de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 13364de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 13374de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1338e14c3cafSHerbert Xu TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; 1339a166140eSMartin KaFai Lau tcp_skb_fragment_eor(skb, buff); 13401da177e4SLinus Torvalds 13411da177e4SLinus Torvalds skb_split(skb, buff, len); 13421da177e4SLinus Torvalds 134398be9b12SEric Dumazet buff->ip_summed = CHECKSUM_PARTIAL; 13441da177e4SLinus Torvalds 1345a61bbcf2SPatrick McHardy buff->tstamp = skb->tstamp; 1346490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 13471da177e4SLinus Torvalds 13486475be16SDavid S. Miller old_factor = tcp_skb_pcount(skb); 13496475be16SDavid S. Miller 13501da177e4SLinus Torvalds /* Fix up tso_factor for both original and new SKB. */ 13515bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 13525bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 13531da177e4SLinus Torvalds 1354b9f64820SYuchung Cheng /* Update delivered info for the new segment */ 1355b9f64820SYuchung Cheng TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx; 1356b9f64820SYuchung Cheng 13576475be16SDavid S. Miller /* If this packet has been sent out already, we must 13586475be16SDavid S. Miller * adjust the various packet counters. 13596475be16SDavid S. Miller */ 1360cf0b450cSHerbert Xu if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 13616475be16SDavid S. Miller int diff = old_factor - tcp_skb_pcount(skb) - 13626475be16SDavid S. Miller tcp_skb_pcount(buff); 13631da177e4SLinus Torvalds 1364797108d1SIlpo Järvinen if (diff) 1365797108d1SIlpo Järvinen tcp_adjust_pcount(sk, skb, diff); 13661da177e4SLinus Torvalds } 13671da177e4SLinus Torvalds 13681da177e4SLinus Torvalds /* Link BUFF into the send queue. */ 1369f4a775d1SEric Dumazet __skb_header_release(buff); 137075c119afSEric Dumazet tcp_insert_write_queue_after(skb, buff, sk, tcp_queue); 1371f67971e6SEric Dumazet if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE) 1372e2080072SEric Dumazet list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor); 13731da177e4SLinus Torvalds 13741da177e4SLinus Torvalds return 0; 13751da177e4SLinus Torvalds } 13761da177e4SLinus Torvalds 1377f4d01666SEric Dumazet /* This is similar to __pskb_pull_tail(). The difference is that pulled 1378f4d01666SEric Dumazet * data is not copied, but immediately discarded. 13791da177e4SLinus Torvalds */ 13807162fb24SEric Dumazet static int __pskb_trim_head(struct sk_buff *skb, int len) 13811da177e4SLinus Torvalds { 13827b7fc97aSEric Dumazet struct skb_shared_info *shinfo; 13831da177e4SLinus Torvalds int i, k, eat; 13841da177e4SLinus Torvalds 13854fa48bf3SEric Dumazet eat = min_t(int, len, skb_headlen(skb)); 13864fa48bf3SEric Dumazet if (eat) { 13874fa48bf3SEric Dumazet __skb_pull(skb, eat); 13884fa48bf3SEric Dumazet len -= eat; 13894fa48bf3SEric Dumazet if (!len) 13907162fb24SEric Dumazet return 0; 13914fa48bf3SEric Dumazet } 13921da177e4SLinus Torvalds eat = len; 13931da177e4SLinus Torvalds k = 0; 13947b7fc97aSEric Dumazet shinfo = skb_shinfo(skb); 13957b7fc97aSEric Dumazet for (i = 0; i < shinfo->nr_frags; i++) { 13967b7fc97aSEric Dumazet int size = skb_frag_size(&shinfo->frags[i]); 13979e903e08SEric Dumazet 13989e903e08SEric Dumazet if (size <= eat) { 1399aff65da0SIan Campbell skb_frag_unref(skb, i); 14009e903e08SEric Dumazet eat -= size; 14011da177e4SLinus Torvalds } else { 14027b7fc97aSEric Dumazet shinfo->frags[k] = shinfo->frags[i]; 14031da177e4SLinus Torvalds if (eat) { 14047b7fc97aSEric Dumazet shinfo->frags[k].page_offset += eat; 14057b7fc97aSEric Dumazet skb_frag_size_sub(&shinfo->frags[k], eat); 14061da177e4SLinus Torvalds eat = 0; 14071da177e4SLinus Torvalds } 14081da177e4SLinus Torvalds k++; 14091da177e4SLinus Torvalds } 14101da177e4SLinus Torvalds } 14117b7fc97aSEric Dumazet shinfo->nr_frags = k; 14121da177e4SLinus Torvalds 14131da177e4SLinus Torvalds skb->data_len -= len; 14141da177e4SLinus Torvalds skb->len = skb->data_len; 14157162fb24SEric Dumazet return len; 14161da177e4SLinus Torvalds } 14171da177e4SLinus Torvalds 141867edfef7SAndi Kleen /* Remove acked data from a packet in the transmit queue. */ 14191da177e4SLinus Torvalds int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 14201da177e4SLinus Torvalds { 14217162fb24SEric Dumazet u32 delta_truesize; 14227162fb24SEric Dumazet 142314bbd6a5SPravin B Shelar if (skb_unclone(skb, GFP_ATOMIC)) 14241da177e4SLinus Torvalds return -ENOMEM; 14251da177e4SLinus Torvalds 14267162fb24SEric Dumazet delta_truesize = __pskb_trim_head(skb, len); 14271da177e4SLinus Torvalds 14281da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq += len; 142984fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 14301da177e4SLinus Torvalds 14317162fb24SEric Dumazet if (delta_truesize) { 14327162fb24SEric Dumazet skb->truesize -= delta_truesize; 14337162fb24SEric Dumazet sk->sk_wmem_queued -= delta_truesize; 14347162fb24SEric Dumazet sk_mem_uncharge(sk, delta_truesize); 14351da177e4SLinus Torvalds sock_set_flag(sk, SOCK_QUEUE_SHRUNK); 14367162fb24SEric Dumazet } 14371da177e4SLinus Torvalds 14385b35e1e6SNeal Cardwell /* Any change of skb->len requires recalculation of tso factor. */ 14391da177e4SLinus Torvalds if (tcp_skb_pcount(skb) > 1) 14405bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb)); 14411da177e4SLinus Torvalds 14421da177e4SLinus Torvalds return 0; 14431da177e4SLinus Torvalds } 14441da177e4SLinus Torvalds 14451b63edd6SYuchung Cheng /* Calculate MSS not accounting any TCP options. */ 14461b63edd6SYuchung Cheng static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu) 14475d424d5aSJohn Heffner { 1448cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1449cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 14505d424d5aSJohn Heffner int mss_now; 14515d424d5aSJohn Heffner 14525d424d5aSJohn Heffner /* Calculate base mss without TCP options: 14535d424d5aSJohn Heffner It is MMS_S - sizeof(tcphdr) of rfc1122 14545d424d5aSJohn Heffner */ 14555d424d5aSJohn Heffner mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 14565d424d5aSJohn Heffner 145767469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 145867469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 145967469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 146067469601SEric Dumazet 146167469601SEric Dumazet if (dst && dst_allfrag(dst)) 146267469601SEric Dumazet mss_now -= icsk->icsk_af_ops->net_frag_header_len; 146367469601SEric Dumazet } 146467469601SEric Dumazet 14655d424d5aSJohn Heffner /* Clamp it (mss_clamp does not include tcp options) */ 14665d424d5aSJohn Heffner if (mss_now > tp->rx_opt.mss_clamp) 14675d424d5aSJohn Heffner mss_now = tp->rx_opt.mss_clamp; 14685d424d5aSJohn Heffner 14695d424d5aSJohn Heffner /* Now subtract optional transport overhead */ 14705d424d5aSJohn Heffner mss_now -= icsk->icsk_ext_hdr_len; 14715d424d5aSJohn Heffner 14725d424d5aSJohn Heffner /* Then reserve room for full set of TCP options and 8 bytes of data */ 14735d424d5aSJohn Heffner if (mss_now < 48) 14745d424d5aSJohn Heffner mss_now = 48; 14755d424d5aSJohn Heffner return mss_now; 14765d424d5aSJohn Heffner } 14775d424d5aSJohn Heffner 14781b63edd6SYuchung Cheng /* Calculate MSS. Not accounting for SACKs here. */ 14791b63edd6SYuchung Cheng int tcp_mtu_to_mss(struct sock *sk, int pmtu) 14801b63edd6SYuchung Cheng { 14811b63edd6SYuchung Cheng /* Subtract TCP options size, not including SACKs */ 14821b63edd6SYuchung Cheng return __tcp_mtu_to_mss(sk, pmtu) - 14831b63edd6SYuchung Cheng (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr)); 14841b63edd6SYuchung Cheng } 14851b63edd6SYuchung Cheng 14865d424d5aSJohn Heffner /* Inverse of above */ 148767469601SEric Dumazet int tcp_mss_to_mtu(struct sock *sk, int mss) 14885d424d5aSJohn Heffner { 1489cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1490cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 14915d424d5aSJohn Heffner int mtu; 14925d424d5aSJohn Heffner 14935d424d5aSJohn Heffner mtu = mss + 14945d424d5aSJohn Heffner tp->tcp_header_len + 14955d424d5aSJohn Heffner icsk->icsk_ext_hdr_len + 14965d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 14975d424d5aSJohn Heffner 149867469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 149967469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 150067469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 150167469601SEric Dumazet 150267469601SEric Dumazet if (dst && dst_allfrag(dst)) 150367469601SEric Dumazet mtu += icsk->icsk_af_ops->net_frag_header_len; 150467469601SEric Dumazet } 15055d424d5aSJohn Heffner return mtu; 15065d424d5aSJohn Heffner } 1507556c6b46SNeal Cardwell EXPORT_SYMBOL(tcp_mss_to_mtu); 15085d424d5aSJohn Heffner 150967edfef7SAndi Kleen /* MTU probing init per socket */ 15105d424d5aSJohn Heffner void tcp_mtup_init(struct sock *sk) 15115d424d5aSJohn Heffner { 15125d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 15135d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 1514b0f9ca53SFan Du struct net *net = sock_net(sk); 15155d424d5aSJohn Heffner 1516b0f9ca53SFan Du icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1; 15175d424d5aSJohn Heffner icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 15185d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 1519b0f9ca53SFan Du icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss); 15205d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 152105cbc0dbSFan Du if (icsk->icsk_mtup.enabled) 1522c74df29aSEric Dumazet icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; 15235d424d5aSJohn Heffner } 15244bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_mtup_init); 15255d424d5aSJohn Heffner 15261da177e4SLinus Torvalds /* This function synchronize snd mss to current pmtu/exthdr set. 15271da177e4SLinus Torvalds 15281da177e4SLinus Torvalds tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 15291da177e4SLinus Torvalds for TCP options, but includes only bare TCP header. 15301da177e4SLinus Torvalds 15311da177e4SLinus Torvalds tp->rx_opt.mss_clamp is mss negotiated at connection setup. 1532caa20d9aSStephen Hemminger It is minimum of user_mss and mss received with SYN. 15331da177e4SLinus Torvalds It also does not include TCP options. 15341da177e4SLinus Torvalds 1535d83d8461SArnaldo Carvalho de Melo inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. 15361da177e4SLinus Torvalds 15371da177e4SLinus Torvalds tp->mss_cache is current effective sending mss, including 15381da177e4SLinus Torvalds all tcp options except for SACKs. It is evaluated, 15391da177e4SLinus Torvalds taking into account current pmtu, but never exceeds 15401da177e4SLinus Torvalds tp->rx_opt.mss_clamp. 15411da177e4SLinus Torvalds 15421da177e4SLinus Torvalds NOTE1. rfc1122 clearly states that advertised MSS 15431da177e4SLinus Torvalds DOES NOT include either tcp or ip options. 15441da177e4SLinus Torvalds 1545d83d8461SArnaldo Carvalho de Melo NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache 1546d83d8461SArnaldo Carvalho de Melo are READ ONLY outside this function. --ANK (980731) 15471da177e4SLinus Torvalds */ 15481da177e4SLinus Torvalds unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 15491da177e4SLinus Torvalds { 15501da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1551d83d8461SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 15525d424d5aSJohn Heffner int mss_now; 15531da177e4SLinus Torvalds 15545d424d5aSJohn Heffner if (icsk->icsk_mtup.search_high > pmtu) 15555d424d5aSJohn Heffner icsk->icsk_mtup.search_high = pmtu; 15561da177e4SLinus Torvalds 15575d424d5aSJohn Heffner mss_now = tcp_mtu_to_mss(sk, pmtu); 1558409d22b4SIlpo Järvinen mss_now = tcp_bound_to_half_wnd(tp, mss_now); 15591da177e4SLinus Torvalds 15601da177e4SLinus Torvalds /* And store cached results */ 1561d83d8461SArnaldo Carvalho de Melo icsk->icsk_pmtu_cookie = pmtu; 15625d424d5aSJohn Heffner if (icsk->icsk_mtup.enabled) 15635d424d5aSJohn Heffner mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); 1564c1b4a7e6SDavid S. Miller tp->mss_cache = mss_now; 15651da177e4SLinus Torvalds 15661da177e4SLinus Torvalds return mss_now; 15671da177e4SLinus Torvalds } 15684bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_sync_mss); 15691da177e4SLinus Torvalds 15701da177e4SLinus Torvalds /* Compute the current effective MSS, taking SACKs and IP options, 15711da177e4SLinus Torvalds * and even PMTU discovery events into account. 15721da177e4SLinus Torvalds */ 15730c54b85fSIlpo Järvinen unsigned int tcp_current_mss(struct sock *sk) 15741da177e4SLinus Torvalds { 1575cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1576cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1577c1b4a7e6SDavid S. Miller u32 mss_now; 157895c96174SEric Dumazet unsigned int header_len; 157933ad798cSAdam Langley struct tcp_out_options opts; 158033ad798cSAdam Langley struct tcp_md5sig_key *md5; 15811da177e4SLinus Torvalds 1582c1b4a7e6SDavid S. Miller mss_now = tp->mss_cache; 1583c1b4a7e6SDavid S. Miller 15841da177e4SLinus Torvalds if (dst) { 15851da177e4SLinus Torvalds u32 mtu = dst_mtu(dst); 1586d83d8461SArnaldo Carvalho de Melo if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 15871da177e4SLinus Torvalds mss_now = tcp_sync_mss(sk, mtu); 15881da177e4SLinus Torvalds } 15891da177e4SLinus Torvalds 159033ad798cSAdam Langley header_len = tcp_established_options(sk, NULL, &opts, &md5) + 159133ad798cSAdam Langley sizeof(struct tcphdr); 159233ad798cSAdam Langley /* The mss_cache is sized based on tp->tcp_header_len, which assumes 159333ad798cSAdam Langley * some common options. If this is an odd packet (because we have SACK 159433ad798cSAdam Langley * blocks etc) then our calculated header_len will be different, and 159533ad798cSAdam Langley * we have to adjust mss_now correspondingly */ 159633ad798cSAdam Langley if (header_len != tp->tcp_header_len) { 159733ad798cSAdam Langley int delta = (int) header_len - tp->tcp_header_len; 159833ad798cSAdam Langley mss_now -= delta; 159933ad798cSAdam Langley } 1600cfb6eeb4SYOSHIFUJI Hideaki 16011da177e4SLinus Torvalds return mss_now; 16021da177e4SLinus Torvalds } 16031da177e4SLinus Torvalds 160486fd14adSWeiping Pan /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. 160586fd14adSWeiping Pan * As additional protections, we do not touch cwnd in retransmission phases, 160686fd14adSWeiping Pan * and if application hit its sndbuf limit recently. 160786fd14adSWeiping Pan */ 160886fd14adSWeiping Pan static void tcp_cwnd_application_limited(struct sock *sk) 1609a762a980SDavid S. Miller { 16109e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1611a762a980SDavid S. Miller 161286fd14adSWeiping Pan if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && 161386fd14adSWeiping Pan sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 161486fd14adSWeiping Pan /* Limited by application or receiver window. */ 161586fd14adSWeiping Pan u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); 161686fd14adSWeiping Pan u32 win_used = max(tp->snd_cwnd_used, init_win); 161786fd14adSWeiping Pan if (win_used < tp->snd_cwnd) { 161886fd14adSWeiping Pan tp->snd_ssthresh = tcp_current_ssthresh(sk); 161986fd14adSWeiping Pan tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1; 162086fd14adSWeiping Pan } 162186fd14adSWeiping Pan tp->snd_cwnd_used = 0; 162286fd14adSWeiping Pan } 1623c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 162486fd14adSWeiping Pan } 162586fd14adSWeiping Pan 1626ca8a2263SNeal Cardwell static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited) 1627a762a980SDavid S. Miller { 16281b1fc3fdSWei Wang const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 1629a762a980SDavid S. Miller struct tcp_sock *tp = tcp_sk(sk); 1630a762a980SDavid S. Miller 1631ca8a2263SNeal Cardwell /* Track the maximum number of outstanding packets in each 1632ca8a2263SNeal Cardwell * window, and remember whether we were cwnd-limited then. 1633ca8a2263SNeal Cardwell */ 1634ca8a2263SNeal Cardwell if (!before(tp->snd_una, tp->max_packets_seq) || 1635ca8a2263SNeal Cardwell tp->packets_out > tp->max_packets_out) { 1636ca8a2263SNeal Cardwell tp->max_packets_out = tp->packets_out; 1637ca8a2263SNeal Cardwell tp->max_packets_seq = tp->snd_nxt; 1638ca8a2263SNeal Cardwell tp->is_cwnd_limited = is_cwnd_limited; 1639ca8a2263SNeal Cardwell } 1640e114a710SEric Dumazet 164124901551SEric Dumazet if (tcp_is_cwnd_limited(sk)) { 1642a762a980SDavid S. Miller /* Network is feed fully. */ 1643a762a980SDavid S. Miller tp->snd_cwnd_used = 0; 1644c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 1645a762a980SDavid S. Miller } else { 1646a762a980SDavid S. Miller /* Network starves. */ 1647a762a980SDavid S. Miller if (tp->packets_out > tp->snd_cwnd_used) 1648a762a980SDavid S. Miller tp->snd_cwnd_used = tp->packets_out; 1649a762a980SDavid S. Miller 1650b510f0d2SEric Dumazet if (sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle && 1651c2203cf7SEric Dumazet (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto && 16521b1fc3fdSWei Wang !ca_ops->cong_control) 1653a762a980SDavid S. Miller tcp_cwnd_application_limited(sk); 1654b0f71bd3SFrancis Yan 1655b0f71bd3SFrancis Yan /* The following conditions together indicate the starvation 1656b0f71bd3SFrancis Yan * is caused by insufficient sender buffer: 1657b0f71bd3SFrancis Yan * 1) just sent some data (see tcp_write_xmit) 1658b0f71bd3SFrancis Yan * 2) not cwnd limited (this else condition) 165975c119afSEric Dumazet * 3) no more data to send (tcp_write_queue_empty()) 1660b0f71bd3SFrancis Yan * 4) application is hitting buffer limit (SOCK_NOSPACE) 1661b0f71bd3SFrancis Yan */ 166275c119afSEric Dumazet if (tcp_write_queue_empty(sk) && sk->sk_socket && 1663b0f71bd3SFrancis Yan test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) && 1664b0f71bd3SFrancis Yan (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) 1665b0f71bd3SFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED); 1666a762a980SDavid S. Miller } 1667a762a980SDavid S. Miller } 1668a762a980SDavid S. Miller 1669d4589926SEric Dumazet /* Minshall's variant of the Nagle send check. */ 1670d4589926SEric Dumazet static bool tcp_minshall_check(const struct tcp_sock *tp) 1671d4589926SEric Dumazet { 1672d4589926SEric Dumazet return after(tp->snd_sml, tp->snd_una) && 1673d4589926SEric Dumazet !after(tp->snd_sml, tp->snd_nxt); 1674d4589926SEric Dumazet } 1675d4589926SEric Dumazet 1676d4589926SEric Dumazet /* Update snd_sml if this skb is under mss 1677d4589926SEric Dumazet * Note that a TSO packet might end with a sub-mss segment 1678d4589926SEric Dumazet * The test is really : 1679d4589926SEric Dumazet * if ((skb->len % mss) != 0) 1680d4589926SEric Dumazet * tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1681d4589926SEric Dumazet * But we can avoid doing the divide again given we already have 1682d4589926SEric Dumazet * skb_pcount = skb->len / mss_now 16830e3a4803SIlpo Järvinen */ 1684d4589926SEric Dumazet static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now, 1685d4589926SEric Dumazet const struct sk_buff *skb) 1686d4589926SEric Dumazet { 1687d4589926SEric Dumazet if (skb->len < tcp_skb_pcount(skb) * mss_now) 1688d4589926SEric Dumazet tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1689d4589926SEric Dumazet } 1690d4589926SEric Dumazet 1691d4589926SEric Dumazet /* Return false, if packet can be sent now without violation Nagle's rules: 1692d4589926SEric Dumazet * 1. It is full sized. (provided by caller in %partial bool) 1693d4589926SEric Dumazet * 2. Or it contains FIN. (already checked by caller) 1694d4589926SEric Dumazet * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. 1695d4589926SEric Dumazet * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 1696d4589926SEric Dumazet * With Minshall's modification: all sent small packets are ACKed. 1697d4589926SEric Dumazet */ 1698d4589926SEric Dumazet static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp, 1699cc93fc51SPeter Pan(潘卫平) int nonagle) 1700d4589926SEric Dumazet { 1701d4589926SEric Dumazet return partial && 1702d4589926SEric Dumazet ((nonagle & TCP_NAGLE_CORK) || 1703d4589926SEric Dumazet (!nonagle && tp->packets_out && tcp_minshall_check(tp))); 1704d4589926SEric Dumazet } 1705605ad7f1SEric Dumazet 1706605ad7f1SEric Dumazet /* Return how many segs we'd like on a TSO packet, 1707605ad7f1SEric Dumazet * to send one TSO packet per ms 1708605ad7f1SEric Dumazet */ 1709dcb8c9b4SEric Dumazet static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now, 17101b3878caSNeal Cardwell int min_tso_segs) 1711605ad7f1SEric Dumazet { 1712605ad7f1SEric Dumazet u32 bytes, segs; 1713605ad7f1SEric Dumazet 171476a9ebe8SEric Dumazet bytes = min_t(unsigned long, 171576a9ebe8SEric Dumazet sk->sk_pacing_rate >> sk->sk_pacing_shift, 1716605ad7f1SEric Dumazet sk->sk_gso_max_size - 1 - MAX_TCP_HEADER); 1717605ad7f1SEric Dumazet 1718605ad7f1SEric Dumazet /* Goal is to send at least one packet per ms, 1719605ad7f1SEric Dumazet * not one big TSO packet every 100 ms. 1720605ad7f1SEric Dumazet * This preserves ACK clocking and is consistent 1721605ad7f1SEric Dumazet * with tcp_tso_should_defer() heuristic. 1722605ad7f1SEric Dumazet */ 17231b3878caSNeal Cardwell segs = max_t(u32, bytes / mss_now, min_tso_segs); 1724605ad7f1SEric Dumazet 1725350c9f48SEric Dumazet return segs; 1726605ad7f1SEric Dumazet } 1727605ad7f1SEric Dumazet 1728ed6e7268SNeal Cardwell /* Return the number of segments we want in the skb we are transmitting. 1729ed6e7268SNeal Cardwell * See if congestion control module wants to decide; otherwise, autosize. 1730ed6e7268SNeal Cardwell */ 1731ed6e7268SNeal Cardwell static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now) 1732ed6e7268SNeal Cardwell { 1733ed6e7268SNeal Cardwell const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 1734dcb8c9b4SEric Dumazet u32 min_tso, tso_segs; 1735ed6e7268SNeal Cardwell 1736dcb8c9b4SEric Dumazet min_tso = ca_ops->min_tso_segs ? 1737dcb8c9b4SEric Dumazet ca_ops->min_tso_segs(sk) : 1738dcb8c9b4SEric Dumazet sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs; 1739dcb8c9b4SEric Dumazet 1740dcb8c9b4SEric Dumazet tso_segs = tcp_tso_autosize(sk, mss_now, min_tso); 1741350c9f48SEric Dumazet return min_t(u32, tso_segs, sk->sk_gso_max_segs); 1742ed6e7268SNeal Cardwell } 1743ed6e7268SNeal Cardwell 1744d4589926SEric Dumazet /* Returns the portion of skb which can be sent right away */ 1745d4589926SEric Dumazet static unsigned int tcp_mss_split_point(const struct sock *sk, 1746d4589926SEric Dumazet const struct sk_buff *skb, 1747d4589926SEric Dumazet unsigned int mss_now, 1748d4589926SEric Dumazet unsigned int max_segs, 1749d4589926SEric Dumazet int nonagle) 1750c1b4a7e6SDavid S. Miller { 1751cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1752d4589926SEric Dumazet u32 partial, needed, window, max_len; 1753c1b4a7e6SDavid S. Miller 175490840defSIlpo Järvinen window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 17551485348dSBen Hutchings max_len = mss_now * max_segs; 17560e3a4803SIlpo Järvinen 17571485348dSBen Hutchings if (likely(max_len <= window && skb != tcp_write_queue_tail(sk))) 17581485348dSBen Hutchings return max_len; 17590e3a4803SIlpo Järvinen 17605ea3a748SIlpo Järvinen needed = min(skb->len, window); 17615ea3a748SIlpo Järvinen 17621485348dSBen Hutchings if (max_len <= needed) 17631485348dSBen Hutchings return max_len; 17640e3a4803SIlpo Järvinen 1765d4589926SEric Dumazet partial = needed % mss_now; 1766d4589926SEric Dumazet /* If last segment is not a full MSS, check if Nagle rules allow us 1767d4589926SEric Dumazet * to include this last segment in this skb. 1768d4589926SEric Dumazet * Otherwise, we'll split the skb at last MSS boundary 1769d4589926SEric Dumazet */ 1770cc93fc51SPeter Pan(潘卫平) if (tcp_nagle_check(partial != 0, tp, nonagle)) 1771d4589926SEric Dumazet return needed - partial; 1772d4589926SEric Dumazet 1773d4589926SEric Dumazet return needed; 1774c1b4a7e6SDavid S. Miller } 1775c1b4a7e6SDavid S. Miller 1776c1b4a7e6SDavid S. Miller /* Can at least one segment of SKB be sent right now, according to the 1777c1b4a7e6SDavid S. Miller * congestion window rules? If so, return how many segments are allowed. 1778c1b4a7e6SDavid S. Miller */ 1779cf533ea5SEric Dumazet static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, 1780cf533ea5SEric Dumazet const struct sk_buff *skb) 1781c1b4a7e6SDavid S. Miller { 1782d649a7a8SEric Dumazet u32 in_flight, cwnd, halfcwnd; 1783c1b4a7e6SDavid S. Miller 1784c1b4a7e6SDavid S. Miller /* Don't be strict about the congestion window for the final FIN. */ 17854de075e0SEric Dumazet if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && 17864de075e0SEric Dumazet tcp_skb_pcount(skb) == 1) 1787c1b4a7e6SDavid S. Miller return 1; 1788c1b4a7e6SDavid S. Miller 1789c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1790c1b4a7e6SDavid S. Miller cwnd = tp->snd_cwnd; 1791d649a7a8SEric Dumazet if (in_flight >= cwnd) 1792c1b4a7e6SDavid S. Miller return 0; 1793d649a7a8SEric Dumazet 1794d649a7a8SEric Dumazet /* For better scheduling, ensure we have at least 1795d649a7a8SEric Dumazet * 2 GSO packets in flight. 1796d649a7a8SEric Dumazet */ 1797d649a7a8SEric Dumazet halfcwnd = max(cwnd >> 1, 1U); 1798d649a7a8SEric Dumazet return min(halfcwnd, cwnd - in_flight); 1799c1b4a7e6SDavid S. Miller } 1800c1b4a7e6SDavid S. Miller 1801b595076aSUwe Kleine-König /* Initialize TSO state of a skb. 180267edfef7SAndi Kleen * This must be invoked the first time we consider transmitting 1803c1b4a7e6SDavid S. Miller * SKB onto the wire. 1804c1b4a7e6SDavid S. Miller */ 18055bbb432cSEric Dumazet static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now) 1806c1b4a7e6SDavid S. Miller { 1807c1b4a7e6SDavid S. Miller int tso_segs = tcp_skb_pcount(skb); 1808c1b4a7e6SDavid S. Miller 1809f8269a49SIlpo Järvinen if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { 18105bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 1811c1b4a7e6SDavid S. Miller tso_segs = tcp_skb_pcount(skb); 1812c1b4a7e6SDavid S. Miller } 1813c1b4a7e6SDavid S. Miller return tso_segs; 1814c1b4a7e6SDavid S. Miller } 1815c1b4a7e6SDavid S. Miller 1816c1b4a7e6SDavid S. Miller 1817a2a385d6SEric Dumazet /* Return true if the Nagle test allows this packet to be 1818c1b4a7e6SDavid S. Miller * sent now. 1819c1b4a7e6SDavid S. Miller */ 1820a2a385d6SEric Dumazet static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, 1821c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1822c1b4a7e6SDavid S. Miller { 1823c1b4a7e6SDavid S. Miller /* Nagle rule does not apply to frames, which sit in the middle of the 1824c1b4a7e6SDavid S. Miller * write_queue (they have no chances to get new data). 1825c1b4a7e6SDavid S. Miller * 1826c1b4a7e6SDavid S. Miller * This is implemented in the callers, where they modify the 'nonagle' 1827c1b4a7e6SDavid S. Miller * argument based upon the location of SKB in the send queue. 1828c1b4a7e6SDavid S. Miller */ 1829c1b4a7e6SDavid S. Miller if (nonagle & TCP_NAGLE_PUSH) 1830a2a385d6SEric Dumazet return true; 1831c1b4a7e6SDavid S. Miller 18329b44190dSYuchung Cheng /* Don't use the nagle rule for urgent data (or for the final FIN). */ 18339b44190dSYuchung Cheng if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) 1834a2a385d6SEric Dumazet return true; 1835c1b4a7e6SDavid S. Miller 1836cc93fc51SPeter Pan(潘卫平) if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle)) 1837a2a385d6SEric Dumazet return true; 1838c1b4a7e6SDavid S. Miller 1839a2a385d6SEric Dumazet return false; 1840c1b4a7e6SDavid S. Miller } 1841c1b4a7e6SDavid S. Miller 1842c1b4a7e6SDavid S. Miller /* Does at least the first segment of SKB fit into the send window? */ 1843a2a385d6SEric Dumazet static bool tcp_snd_wnd_test(const struct tcp_sock *tp, 1844a2a385d6SEric Dumazet const struct sk_buff *skb, 1845056834d9SIlpo Järvinen unsigned int cur_mss) 1846c1b4a7e6SDavid S. Miller { 1847c1b4a7e6SDavid S. Miller u32 end_seq = TCP_SKB_CB(skb)->end_seq; 1848c1b4a7e6SDavid S. Miller 1849c1b4a7e6SDavid S. Miller if (skb->len > cur_mss) 1850c1b4a7e6SDavid S. Miller end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 1851c1b4a7e6SDavid S. Miller 185290840defSIlpo Järvinen return !after(end_seq, tcp_wnd_end(tp)); 1853c1b4a7e6SDavid S. Miller } 1854c1b4a7e6SDavid S. Miller 1855c1b4a7e6SDavid S. Miller /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 1856c1b4a7e6SDavid S. Miller * which is put after SKB on the list. It is very much like 1857c1b4a7e6SDavid S. Miller * tcp_fragment() except that it may make several kinds of assumptions 1858c1b4a7e6SDavid S. Miller * in order to speed up the splitting operation. In particular, we 1859c1b4a7e6SDavid S. Miller * know that all the data is in scatter-gather pages, and that the 1860c1b4a7e6SDavid S. Miller * packet has never been sent out before (and thus is not cloned). 1861c1b4a7e6SDavid S. Miller */ 186275c119afSEric Dumazet static int tso_fragment(struct sock *sk, enum tcp_queue tcp_queue, 186375c119afSEric Dumazet struct sk_buff *skb, unsigned int len, 1864c4ead4c5SEric Dumazet unsigned int mss_now, gfp_t gfp) 1865c1b4a7e6SDavid S. Miller { 1866c1b4a7e6SDavid S. Miller struct sk_buff *buff; 1867c1b4a7e6SDavid S. Miller int nlen = skb->len - len; 18689ce01461SIlpo Järvinen u8 flags; 1869c1b4a7e6SDavid S. Miller 1870c1b4a7e6SDavid S. Miller /* All of a TSO frame must be composed of paged data. */ 1871c8ac3774SHerbert Xu if (skb->len != skb->data_len) 187275c119afSEric Dumazet return tcp_fragment(sk, tcp_queue, skb, len, mss_now, gfp); 1873c1b4a7e6SDavid S. Miller 1874eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, gfp, true); 187551456b29SIan Morris if (unlikely(!buff)) 1876c1b4a7e6SDavid S. Miller return -ENOMEM; 1877c1b4a7e6SDavid S. Miller 18783ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 18793ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1880b60b49eaSHerbert Xu buff->truesize += nlen; 1881c1b4a7e6SDavid S. Miller skb->truesize -= nlen; 1882c1b4a7e6SDavid S. Miller 1883c1b4a7e6SDavid S. Miller /* Correct the sequence numbers. */ 1884c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 1885c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 1886c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 1887c1b4a7e6SDavid S. Miller 1888c1b4a7e6SDavid S. Miller /* PSH and FIN should only be set in the second packet. */ 18894de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 18904de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 18914de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1892c1b4a7e6SDavid S. Miller 1893c1b4a7e6SDavid S. Miller /* This packet was never sent out yet, so no SACK bits. */ 1894c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->sacked = 0; 1895c1b4a7e6SDavid S. Miller 1896a166140eSMartin KaFai Lau tcp_skb_fragment_eor(skb, buff); 1897a166140eSMartin KaFai Lau 189898be9b12SEric Dumazet buff->ip_summed = CHECKSUM_PARTIAL; 1899c1b4a7e6SDavid S. Miller skb_split(skb, buff, len); 1900490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 1901c1b4a7e6SDavid S. Miller 1902c1b4a7e6SDavid S. Miller /* Fix up tso_factor for both original and new SKB. */ 19035bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 19045bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 1905c1b4a7e6SDavid S. Miller 1906c1b4a7e6SDavid S. Miller /* Link BUFF into the send queue. */ 1907f4a775d1SEric Dumazet __skb_header_release(buff); 190875c119afSEric Dumazet tcp_insert_write_queue_after(skb, buff, sk, tcp_queue); 1909c1b4a7e6SDavid S. Miller 1910c1b4a7e6SDavid S. Miller return 0; 1911c1b4a7e6SDavid S. Miller } 1912c1b4a7e6SDavid S. Miller 1913c1b4a7e6SDavid S. Miller /* Try to defer sending, if possible, in order to minimize the amount 1914c1b4a7e6SDavid S. Miller * of TSO splitting we do. View it as a kind of TSO Nagle test. 1915c1b4a7e6SDavid S. Miller * 1916c1b4a7e6SDavid S. Miller * This algorithm is from John Heffner. 1917c1b4a7e6SDavid S. Miller */ 1918ca8a2263SNeal Cardwell static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb, 1919605ad7f1SEric Dumazet bool *is_cwnd_limited, u32 max_segs) 1920c1b4a7e6SDavid S. Miller { 19216687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 192250c8339eSEric Dumazet u32 age, send_win, cong_win, limit, in_flight; 192350c8339eSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 192450c8339eSEric Dumazet struct sk_buff *head; 1925ad9f4f50SEric Dumazet int win_divisor; 1926c1b4a7e6SDavid S. Miller 19274de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 1928ae8064acSJohn Heffner goto send_now; 1929c1b4a7e6SDavid S. Miller 193099d7662aSEric Dumazet if (icsk->icsk_ca_state >= TCP_CA_Recovery) 1931ae8064acSJohn Heffner goto send_now; 1932ae8064acSJohn Heffner 19335f852eb5SEric Dumazet /* Avoid bursty behavior by allowing defer 19345f852eb5SEric Dumazet * only if the last write was recent. 19355f852eb5SEric Dumazet */ 1936d635fbe2SEric Dumazet if ((s32)(tcp_jiffies32 - tp->lsndtime) > 0) 1937ae8064acSJohn Heffner goto send_now; 1938908a75c1SDavid S. Miller 1939c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1940c1b4a7e6SDavid S. Miller 1941c8c9aeb5SStefano Brivio BUG_ON(tcp_skb_pcount(skb) <= 1); 1942c8c9aeb5SStefano Brivio BUG_ON(tp->snd_cwnd <= in_flight); 1943c1b4a7e6SDavid S. Miller 194490840defSIlpo Järvinen send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1945c1b4a7e6SDavid S. Miller 1946c1b4a7e6SDavid S. Miller /* From in_flight test above, we know that cwnd > in_flight. */ 1947c1b4a7e6SDavid S. Miller cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; 1948c1b4a7e6SDavid S. Miller 1949c1b4a7e6SDavid S. Miller limit = min(send_win, cong_win); 1950c1b4a7e6SDavid S. Miller 1951ba244fe9SDavid S. Miller /* If a full-sized TSO skb can be sent, do it. */ 1952605ad7f1SEric Dumazet if (limit >= max_segs * tp->mss_cache) 1953ae8064acSJohn Heffner goto send_now; 1954ba244fe9SDavid S. Miller 195562ad2761SIlpo Järvinen /* Middle in queue won't get any more data, full sendable already? */ 195662ad2761SIlpo Järvinen if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) 195762ad2761SIlpo Järvinen goto send_now; 195862ad2761SIlpo Järvinen 19595bbcc0f5SLinus Torvalds win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor); 1960ad9f4f50SEric Dumazet if (win_divisor) { 1961c1b4a7e6SDavid S. Miller u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); 1962c1b4a7e6SDavid S. Miller 1963c1b4a7e6SDavid S. Miller /* If at least some fraction of a window is available, 1964c1b4a7e6SDavid S. Miller * just use it. 1965c1b4a7e6SDavid S. Miller */ 1966ad9f4f50SEric Dumazet chunk /= win_divisor; 1967c1b4a7e6SDavid S. Miller if (limit >= chunk) 1968ae8064acSJohn Heffner goto send_now; 1969c1b4a7e6SDavid S. Miller } else { 1970c1b4a7e6SDavid S. Miller /* Different approach, try not to defer past a single 1971c1b4a7e6SDavid S. Miller * ACK. Receiver should ACK every other full sized 1972c1b4a7e6SDavid S. Miller * frame, so if we have space for more than 3 frames 1973c1b4a7e6SDavid S. Miller * then send now. 1974c1b4a7e6SDavid S. Miller */ 19756b5a5c0dSNeal Cardwell if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) 1976ae8064acSJohn Heffner goto send_now; 1977c1b4a7e6SDavid S. Miller } 1978c1b4a7e6SDavid S. Miller 197975c119afSEric Dumazet /* TODO : use tsorted_sent_queue ? */ 198075c119afSEric Dumazet head = tcp_rtx_queue_head(sk); 198175c119afSEric Dumazet if (!head) 198275c119afSEric Dumazet goto send_now; 19832fd66ffbSEric Dumazet age = tcp_stamp_us_delta(tp->tcp_mstamp, tcp_skb_timestamp_us(head)); 198450c8339eSEric Dumazet /* If next ACK is likely to come too late (half srtt), do not defer */ 198550c8339eSEric Dumazet if (age < (tp->srtt_us >> 4)) 198650c8339eSEric Dumazet goto send_now; 198750c8339eSEric Dumazet 19885f852eb5SEric Dumazet /* Ok, it looks like it is advisable to defer. */ 1989ae8064acSJohn Heffner 1990d2e1339fSBendik Rønning Opstad if (cong_win < send_win && cong_win <= skb->len) 1991ca8a2263SNeal Cardwell *is_cwnd_limited = true; 1992ca8a2263SNeal Cardwell 1993a2a385d6SEric Dumazet return true; 1994ae8064acSJohn Heffner 1995ae8064acSJohn Heffner send_now: 1996a2a385d6SEric Dumazet return false; 1997c1b4a7e6SDavid S. Miller } 1998c1b4a7e6SDavid S. Miller 199905cbc0dbSFan Du static inline void tcp_mtu_check_reprobe(struct sock *sk) 200005cbc0dbSFan Du { 200105cbc0dbSFan Du struct inet_connection_sock *icsk = inet_csk(sk); 200205cbc0dbSFan Du struct tcp_sock *tp = tcp_sk(sk); 200305cbc0dbSFan Du struct net *net = sock_net(sk); 200405cbc0dbSFan Du u32 interval; 200505cbc0dbSFan Du s32 delta; 200605cbc0dbSFan Du 200705cbc0dbSFan Du interval = net->ipv4.sysctl_tcp_probe_interval; 2008c74df29aSEric Dumazet delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp; 200905cbc0dbSFan Du if (unlikely(delta >= interval * HZ)) { 201005cbc0dbSFan Du int mss = tcp_current_mss(sk); 201105cbc0dbSFan Du 201205cbc0dbSFan Du /* Update current search range */ 201305cbc0dbSFan Du icsk->icsk_mtup.probe_size = 0; 201405cbc0dbSFan Du icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + 201505cbc0dbSFan Du sizeof(struct tcphdr) + 201605cbc0dbSFan Du icsk->icsk_af_ops->net_header_len; 201705cbc0dbSFan Du icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss); 201805cbc0dbSFan Du 201905cbc0dbSFan Du /* Update probe time stamp */ 2020c74df29aSEric Dumazet icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; 202105cbc0dbSFan Du } 202205cbc0dbSFan Du } 202305cbc0dbSFan Du 2024808cf9e3SIlya Lesokhin static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len) 2025808cf9e3SIlya Lesokhin { 2026808cf9e3SIlya Lesokhin struct sk_buff *skb, *next; 2027808cf9e3SIlya Lesokhin 2028808cf9e3SIlya Lesokhin skb = tcp_send_head(sk); 2029808cf9e3SIlya Lesokhin tcp_for_write_queue_from_safe(skb, next, sk) { 2030808cf9e3SIlya Lesokhin if (len <= skb->len) 2031808cf9e3SIlya Lesokhin break; 2032808cf9e3SIlya Lesokhin 2033808cf9e3SIlya Lesokhin if (unlikely(TCP_SKB_CB(skb)->eor)) 2034808cf9e3SIlya Lesokhin return false; 2035808cf9e3SIlya Lesokhin 2036808cf9e3SIlya Lesokhin len -= skb->len; 2037808cf9e3SIlya Lesokhin } 2038808cf9e3SIlya Lesokhin 2039808cf9e3SIlya Lesokhin return true; 2040808cf9e3SIlya Lesokhin } 2041808cf9e3SIlya Lesokhin 20425d424d5aSJohn Heffner /* Create a new MTU probe if we are ready. 204367edfef7SAndi Kleen * MTU probe is regularly attempting to increase the path MTU by 204467edfef7SAndi Kleen * deliberately sending larger packets. This discovers routing 204567edfef7SAndi Kleen * changes resulting in larger path MTUs. 204667edfef7SAndi Kleen * 20475d424d5aSJohn Heffner * Returns 0 if we should wait to probe (no cwnd available), 20485d424d5aSJohn Heffner * 1 if a probe was sent, 2049056834d9SIlpo Järvinen * -1 otherwise 2050056834d9SIlpo Järvinen */ 20515d424d5aSJohn Heffner static int tcp_mtu_probe(struct sock *sk) 20525d424d5aSJohn Heffner { 20535d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 205412a59abcSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 20555d424d5aSJohn Heffner struct sk_buff *skb, *nskb, *next; 20566b58e0a5SFan Du struct net *net = sock_net(sk); 20575d424d5aSJohn Heffner int probe_size; 205891cc17c0SIlpo Järvinen int size_needed; 205912a59abcSEric Dumazet int copy, len; 20605d424d5aSJohn Heffner int mss_now; 20616b58e0a5SFan Du int interval; 20625d424d5aSJohn Heffner 20635d424d5aSJohn Heffner /* Not currently probing/verifying, 20645d424d5aSJohn Heffner * not in recovery, 20655d424d5aSJohn Heffner * have enough cwnd, and 206612a59abcSEric Dumazet * not SACKing (the variable headers throw things off) 206712a59abcSEric Dumazet */ 206812a59abcSEric Dumazet if (likely(!icsk->icsk_mtup.enabled || 20695d424d5aSJohn Heffner icsk->icsk_mtup.probe_size || 20705d424d5aSJohn Heffner inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 20715d424d5aSJohn Heffner tp->snd_cwnd < 11 || 207212a59abcSEric Dumazet tp->rx_opt.num_sacks || tp->rx_opt.dsack)) 20735d424d5aSJohn Heffner return -1; 20745d424d5aSJohn Heffner 20756b58e0a5SFan Du /* Use binary search for probe_size between tcp_mss_base, 20766b58e0a5SFan Du * and current mss_clamp. if (search_high - search_low) 20776b58e0a5SFan Du * smaller than a threshold, backoff from probing. 20786b58e0a5SFan Du */ 20790c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 20806b58e0a5SFan Du probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high + 20816b58e0a5SFan Du icsk->icsk_mtup.search_low) >> 1); 208291cc17c0SIlpo Järvinen size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; 20836b58e0a5SFan Du interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low; 208405cbc0dbSFan Du /* When misfortune happens, we are reprobing actively, 208505cbc0dbSFan Du * and then reprobe timer has expired. We stick with current 208605cbc0dbSFan Du * probing process by not resetting search range to its orignal. 208705cbc0dbSFan Du */ 20886b58e0a5SFan Du if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) || 208905cbc0dbSFan Du interval < net->ipv4.sysctl_tcp_probe_threshold) { 209005cbc0dbSFan Du /* Check whether enough time has elaplased for 209105cbc0dbSFan Du * another round of probing. 209205cbc0dbSFan Du */ 209305cbc0dbSFan Du tcp_mtu_check_reprobe(sk); 20945d424d5aSJohn Heffner return -1; 20955d424d5aSJohn Heffner } 20965d424d5aSJohn Heffner 20975d424d5aSJohn Heffner /* Have enough data in the send queue to probe? */ 20987f9c33e5SIlpo Järvinen if (tp->write_seq - tp->snd_nxt < size_needed) 20995d424d5aSJohn Heffner return -1; 21005d424d5aSJohn Heffner 210191cc17c0SIlpo Järvinen if (tp->snd_wnd < size_needed) 21025d424d5aSJohn Heffner return -1; 210390840defSIlpo Järvinen if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) 21045d424d5aSJohn Heffner return 0; 21055d424d5aSJohn Heffner 2106d67c58e9SIlpo Järvinen /* Do we need to wait to drain cwnd? With none in flight, don't stall */ 2107d67c58e9SIlpo Järvinen if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { 2108d67c58e9SIlpo Järvinen if (!tcp_packets_in_flight(tp)) 21095d424d5aSJohn Heffner return -1; 21105d424d5aSJohn Heffner else 21115d424d5aSJohn Heffner return 0; 21125d424d5aSJohn Heffner } 21135d424d5aSJohn Heffner 2114808cf9e3SIlya Lesokhin if (!tcp_can_coalesce_send_queue_head(sk, probe_size)) 2115808cf9e3SIlya Lesokhin return -1; 2116808cf9e3SIlya Lesokhin 21175d424d5aSJohn Heffner /* We're allowed to probe. Build it now. */ 2118eb934478SEric Dumazet nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false); 211951456b29SIan Morris if (!nskb) 21205d424d5aSJohn Heffner return -1; 21213ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 21223ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 21235d424d5aSJohn Heffner 2124fe067e8aSDavid S. Miller skb = tcp_send_head(sk); 21255d424d5aSJohn Heffner 21265d424d5aSJohn Heffner TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 21275d424d5aSJohn Heffner TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 21284de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; 21295d424d5aSJohn Heffner TCP_SKB_CB(nskb)->sacked = 0; 21305d424d5aSJohn Heffner nskb->csum = 0; 213198be9b12SEric Dumazet nskb->ip_summed = CHECKSUM_PARTIAL; 21325d424d5aSJohn Heffner 213350c4817eSIlpo Järvinen tcp_insert_write_queue_before(nskb, skb, sk); 21342b7cda9cSEric Dumazet tcp_highest_sack_replace(sk, skb, nskb); 213550c4817eSIlpo Järvinen 21365d424d5aSJohn Heffner len = 0; 2137234b6860SIlpo Järvinen tcp_for_write_queue_from_safe(skb, next, sk) { 21385d424d5aSJohn Heffner copy = min_t(int, skb->len, probe_size - len); 21395d424d5aSJohn Heffner skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); 21405d424d5aSJohn Heffner 21415d424d5aSJohn Heffner if (skb->len <= copy) { 21425d424d5aSJohn Heffner /* We've eaten all the data from this skb. 21435d424d5aSJohn Heffner * Throw it away. */ 21444de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 2145808cf9e3SIlya Lesokhin /* If this is the last SKB we copy and eor is set 2146808cf9e3SIlya Lesokhin * we need to propagate it to the new skb. 2147808cf9e3SIlya Lesokhin */ 2148808cf9e3SIlya Lesokhin TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor; 2149fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 21503ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 21515d424d5aSJohn Heffner } else { 21524de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & 2153a3433f35SChangli Gao ~(TCPHDR_FIN|TCPHDR_PSH); 21545d424d5aSJohn Heffner if (!skb_shinfo(skb)->nr_frags) { 21555d424d5aSJohn Heffner skb_pull(skb, copy); 21565d424d5aSJohn Heffner } else { 21575d424d5aSJohn Heffner __pskb_trim_head(skb, copy); 21585bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 21595d424d5aSJohn Heffner } 21605d424d5aSJohn Heffner TCP_SKB_CB(skb)->seq += copy; 21615d424d5aSJohn Heffner } 21625d424d5aSJohn Heffner 21635d424d5aSJohn Heffner len += copy; 2164234b6860SIlpo Järvinen 2165234b6860SIlpo Järvinen if (len >= probe_size) 2166234b6860SIlpo Järvinen break; 21675d424d5aSJohn Heffner } 21685bbb432cSEric Dumazet tcp_init_tso_segs(nskb, nskb->len); 21695d424d5aSJohn Heffner 21705d424d5aSJohn Heffner /* We're ready to send. If this fails, the probe will 21717faee5c0SEric Dumazet * be resegmented into mss-sized pieces by tcp_write_xmit(). 21727faee5c0SEric Dumazet */ 21735d424d5aSJohn Heffner if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 21745d424d5aSJohn Heffner /* Decrement cwnd here because we are sending 21755d424d5aSJohn Heffner * effectively two packets. */ 21765d424d5aSJohn Heffner tp->snd_cwnd--; 217766f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, nskb); 21785d424d5aSJohn Heffner 21795d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 21800e7b1368SJohn Heffner tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 21810e7b1368SJohn Heffner tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 21825d424d5aSJohn Heffner 21835d424d5aSJohn Heffner return 1; 21845d424d5aSJohn Heffner } 21855d424d5aSJohn Heffner 21865d424d5aSJohn Heffner return -1; 21875d424d5aSJohn Heffner } 21885d424d5aSJohn Heffner 2189218af599SEric Dumazet static bool tcp_pacing_check(const struct sock *sk) 2190218af599SEric Dumazet { 2191218af599SEric Dumazet return tcp_needs_internal_pacing(sk) && 219273a6bab5SEric Dumazet hrtimer_is_queued(&tcp_sk(sk)->pacing_timer); 2193218af599SEric Dumazet } 2194218af599SEric Dumazet 2195f9616c35SEric Dumazet /* TCP Small Queues : 2196f9616c35SEric Dumazet * Control number of packets in qdisc/devices to two packets / or ~1 ms. 2197f9616c35SEric Dumazet * (These limits are doubled for retransmits) 2198f9616c35SEric Dumazet * This allows for : 2199f9616c35SEric Dumazet * - better RTT estimation and ACK scheduling 2200f9616c35SEric Dumazet * - faster recovery 2201f9616c35SEric Dumazet * - high rates 2202f9616c35SEric Dumazet * Alas, some drivers / subsystems require a fair amount 2203f9616c35SEric Dumazet * of queued bytes to ensure line rate. 2204f9616c35SEric Dumazet * One example is wifi aggregation (802.11 AMPDU) 2205f9616c35SEric Dumazet */ 2206f9616c35SEric Dumazet static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb, 2207f9616c35SEric Dumazet unsigned int factor) 2208f9616c35SEric Dumazet { 220976a9ebe8SEric Dumazet unsigned long limit; 2210f9616c35SEric Dumazet 221176a9ebe8SEric Dumazet limit = max_t(unsigned long, 221276a9ebe8SEric Dumazet 2 * skb->truesize, 221376a9ebe8SEric Dumazet sk->sk_pacing_rate >> sk->sk_pacing_shift); 221476a9ebe8SEric Dumazet limit = min_t(unsigned long, limit, 22159184d8bbSEric Dumazet sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes); 2216f9616c35SEric Dumazet limit <<= factor; 2217f9616c35SEric Dumazet 221814afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) > limit) { 221975c119afSEric Dumazet /* Always send skb if rtx queue is empty. 222075eefc6cSEric Dumazet * No need to wait for TX completion to call us back, 222175eefc6cSEric Dumazet * after softirq/tasklet schedule. 222275eefc6cSEric Dumazet * This helps when TX completions are delayed too much. 222375eefc6cSEric Dumazet */ 222475c119afSEric Dumazet if (tcp_rtx_queue_empty(sk)) 222575eefc6cSEric Dumazet return false; 222675eefc6cSEric Dumazet 22277aa5470cSEric Dumazet set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags); 2228f9616c35SEric Dumazet /* It is possible TX completion already happened 2229f9616c35SEric Dumazet * before we set TSQ_THROTTLED, so we must 2230f9616c35SEric Dumazet * test again the condition. 2231f9616c35SEric Dumazet */ 2232f9616c35SEric Dumazet smp_mb__after_atomic(); 223314afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) > limit) 2234f9616c35SEric Dumazet return true; 2235f9616c35SEric Dumazet } 2236f9616c35SEric Dumazet return false; 2237f9616c35SEric Dumazet } 2238f9616c35SEric Dumazet 223905b055e8SFrancis Yan static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new) 224005b055e8SFrancis Yan { 2241628174ccSEric Dumazet const u32 now = tcp_jiffies32; 2242efe967cdSArnd Bergmann enum tcp_chrono old = tp->chrono_type; 224305b055e8SFrancis Yan 2244efe967cdSArnd Bergmann if (old > TCP_CHRONO_UNSPEC) 2245efe967cdSArnd Bergmann tp->chrono_stat[old - 1] += now - tp->chrono_start; 224605b055e8SFrancis Yan tp->chrono_start = now; 224705b055e8SFrancis Yan tp->chrono_type = new; 224805b055e8SFrancis Yan } 224905b055e8SFrancis Yan 225005b055e8SFrancis Yan void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type) 225105b055e8SFrancis Yan { 225205b055e8SFrancis Yan struct tcp_sock *tp = tcp_sk(sk); 225305b055e8SFrancis Yan 225405b055e8SFrancis Yan /* If there are multiple conditions worthy of tracking in a 22550f87230dSFrancis Yan * chronograph then the highest priority enum takes precedence 22560f87230dSFrancis Yan * over the other conditions. So that if something "more interesting" 225705b055e8SFrancis Yan * starts happening, stop the previous chrono and start a new one. 225805b055e8SFrancis Yan */ 225905b055e8SFrancis Yan if (type > tp->chrono_type) 226005b055e8SFrancis Yan tcp_chrono_set(tp, type); 226105b055e8SFrancis Yan } 226205b055e8SFrancis Yan 226305b055e8SFrancis Yan void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type) 226405b055e8SFrancis Yan { 226505b055e8SFrancis Yan struct tcp_sock *tp = tcp_sk(sk); 226605b055e8SFrancis Yan 22670f87230dSFrancis Yan 22680f87230dSFrancis Yan /* There are multiple conditions worthy of tracking in a 22690f87230dSFrancis Yan * chronograph, so that the highest priority enum takes 22700f87230dSFrancis Yan * precedence over the other conditions (see tcp_chrono_start). 22710f87230dSFrancis Yan * If a condition stops, we only stop chrono tracking if 22720f87230dSFrancis Yan * it's the "most interesting" or current chrono we are 22730f87230dSFrancis Yan * tracking and starts busy chrono if we have pending data. 22740f87230dSFrancis Yan */ 227575c119afSEric Dumazet if (tcp_rtx_and_write_queues_empty(sk)) 227605b055e8SFrancis Yan tcp_chrono_set(tp, TCP_CHRONO_UNSPEC); 22770f87230dSFrancis Yan else if (type == tp->chrono_type) 22780f87230dSFrancis Yan tcp_chrono_set(tp, TCP_CHRONO_BUSY); 227905b055e8SFrancis Yan } 228005b055e8SFrancis Yan 22811da177e4SLinus Torvalds /* This routine writes packets to the network. It advances the 22821da177e4SLinus Torvalds * send_head. This happens as incoming acks open up the remote 22831da177e4SLinus Torvalds * window for us. 22841da177e4SLinus Torvalds * 2285f8269a49SIlpo Järvinen * LARGESEND note: !tcp_urg_mode is overkill, only frames between 2286f8269a49SIlpo Järvinen * snd_up-64k-mss .. snd_up cannot be large. However, taking into 2287f8269a49SIlpo Järvinen * account rare use of URG, this is not a big flaw. 2288f8269a49SIlpo Järvinen * 22896ba8a3b1SNandita Dukkipati * Send at most one packet when push_one > 0. Temporarily ignore 22906ba8a3b1SNandita Dukkipati * cwnd limit to force at most one packet out when push_one == 2. 22916ba8a3b1SNandita Dukkipati 2292a2a385d6SEric Dumazet * Returns true, if no segments are in flight and we have queued segments, 2293a2a385d6SEric Dumazet * but cannot send anything now because of SWS or another problem. 22941da177e4SLinus Torvalds */ 2295a2a385d6SEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 2296d5dd9175SIlpo Järvinen int push_one, gfp_t gfp) 22971da177e4SLinus Torvalds { 22981da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 229992df7b51SDavid S. Miller struct sk_buff *skb; 2300c1b4a7e6SDavid S. Miller unsigned int tso_segs, sent_pkts; 2301c1b4a7e6SDavid S. Miller int cwnd_quota; 23025d424d5aSJohn Heffner int result; 23035615f886SFrancis Yan bool is_cwnd_limited = false, is_rwnd_limited = false; 2304605ad7f1SEric Dumazet u32 max_segs; 23051da177e4SLinus Torvalds 2306c1b4a7e6SDavid S. Miller sent_pkts = 0; 23075d424d5aSJohn Heffner 2308ee1836aeSEric Dumazet tcp_mstamp_refresh(tp); 2309d5dd9175SIlpo Järvinen if (!push_one) { 23105d424d5aSJohn Heffner /* Do MTU probing. */ 2311d5dd9175SIlpo Järvinen result = tcp_mtu_probe(sk); 2312d5dd9175SIlpo Järvinen if (!result) { 2313a2a385d6SEric Dumazet return false; 23145d424d5aSJohn Heffner } else if (result > 0) { 23155d424d5aSJohn Heffner sent_pkts = 1; 23165d424d5aSJohn Heffner } 2317d5dd9175SIlpo Järvinen } 23185d424d5aSJohn Heffner 2319ed6e7268SNeal Cardwell max_segs = tcp_tso_segs(sk, mss_now); 2320fe067e8aSDavid S. Miller while ((skb = tcp_send_head(sk))) { 2321c8ac3774SHerbert Xu unsigned int limit; 2322c8ac3774SHerbert Xu 2323218af599SEric Dumazet if (tcp_pacing_check(sk)) 2324218af599SEric Dumazet break; 2325218af599SEric Dumazet 23265bbb432cSEric Dumazet tso_segs = tcp_init_tso_segs(skb, mss_now); 2327c1b4a7e6SDavid S. Miller BUG_ON(!tso_segs); 2328c1b4a7e6SDavid S. Miller 23299d186cacSAndrey Vagin if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) { 23307faee5c0SEric Dumazet /* "skb_mstamp" is used as a start point for the retransmit timer */ 2331*a7a25630SEric Dumazet tcp_update_skb_after_send(sk, skb, tp->tcp_wstamp_ns); 2332ec342325SAndrew Vagin goto repair; /* Skip network transmission */ 23339d186cacSAndrey Vagin } 2334ec342325SAndrew Vagin 2335b68e9f85SHerbert Xu cwnd_quota = tcp_cwnd_test(tp, skb); 23366ba8a3b1SNandita Dukkipati if (!cwnd_quota) { 23376ba8a3b1SNandita Dukkipati if (push_one == 2) 23386ba8a3b1SNandita Dukkipati /* Force out a loss probe pkt. */ 23396ba8a3b1SNandita Dukkipati cwnd_quota = 1; 23406ba8a3b1SNandita Dukkipati else 2341b68e9f85SHerbert Xu break; 23426ba8a3b1SNandita Dukkipati } 2343b68e9f85SHerbert Xu 23445615f886SFrancis Yan if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) { 23455615f886SFrancis Yan is_rwnd_limited = true; 2346b68e9f85SHerbert Xu break; 23475615f886SFrancis Yan } 2348b68e9f85SHerbert Xu 2349d6a4e26aSEric Dumazet if (tso_segs == 1) { 2350aa93466bSDavid S. Miller if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 2351aa93466bSDavid S. Miller (tcp_skb_is_last(sk, skb) ? 2352aa93466bSDavid S. Miller nonagle : TCP_NAGLE_PUSH)))) 2353aa93466bSDavid S. Miller break; 2354c1b4a7e6SDavid S. Miller } else { 2355ca8a2263SNeal Cardwell if (!push_one && 2356605ad7f1SEric Dumazet tcp_tso_should_defer(sk, skb, &is_cwnd_limited, 2357605ad7f1SEric Dumazet max_segs)) 2358aa93466bSDavid S. Miller break; 2359c1b4a7e6SDavid S. Miller } 2360aa93466bSDavid S. Miller 2361605ad7f1SEric Dumazet limit = mss_now; 2362d6a4e26aSEric Dumazet if (tso_segs > 1 && !tcp_urg_mode(tp)) 2363605ad7f1SEric Dumazet limit = tcp_mss_split_point(sk, skb, mss_now, 2364605ad7f1SEric Dumazet min_t(unsigned int, 2365605ad7f1SEric Dumazet cwnd_quota, 2366605ad7f1SEric Dumazet max_segs), 2367605ad7f1SEric Dumazet nonagle); 2368605ad7f1SEric Dumazet 2369605ad7f1SEric Dumazet if (skb->len > limit && 237075c119afSEric Dumazet unlikely(tso_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE, 237175c119afSEric Dumazet skb, limit, mss_now, gfp))) 2372605ad7f1SEric Dumazet break; 2373605ad7f1SEric Dumazet 2374f9616c35SEric Dumazet if (tcp_small_queue_check(sk, skb, 0)) 237546d3ceabSEric Dumazet break; 2376c9eeec26SEric Dumazet 2377d5dd9175SIlpo Järvinen if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) 23781da177e4SLinus Torvalds break; 23791da177e4SLinus Torvalds 2380ec342325SAndrew Vagin repair: 23811da177e4SLinus Torvalds /* Advance the send_head. This one is sent out. 23821da177e4SLinus Torvalds * This call will increment packets_out. 23831da177e4SLinus Torvalds */ 238466f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 23851da177e4SLinus Torvalds 23861da177e4SLinus Torvalds tcp_minshall_update(tp, mss_now, skb); 2387a262f0cdSNandita Dukkipati sent_pkts += tcp_skb_pcount(skb); 2388d5dd9175SIlpo Järvinen 2389d5dd9175SIlpo Järvinen if (push_one) 2390d5dd9175SIlpo Järvinen break; 23911da177e4SLinus Torvalds } 23921da177e4SLinus Torvalds 23935615f886SFrancis Yan if (is_rwnd_limited) 23945615f886SFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED); 23955615f886SFrancis Yan else 23965615f886SFrancis Yan tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED); 23975615f886SFrancis Yan 2398aa93466bSDavid S. Miller if (likely(sent_pkts)) { 2399684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 2400684bad11SYuchung Cheng tp->prr_out += sent_pkts; 24016ba8a3b1SNandita Dukkipati 24026ba8a3b1SNandita Dukkipati /* Send one loss probe per tail loss episode. */ 24036ba8a3b1SNandita Dukkipati if (push_one != 2) 2404ed66dfafSNeal Cardwell tcp_schedule_loss_probe(sk, false); 2405d2e1339fSBendik Rønning Opstad is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd); 2406ca8a2263SNeal Cardwell tcp_cwnd_validate(sk, is_cwnd_limited); 2407a2a385d6SEric Dumazet return false; 24081da177e4SLinus Torvalds } 240975c119afSEric Dumazet return !tp->packets_out && !tcp_write_queue_empty(sk); 24106ba8a3b1SNandita Dukkipati } 24116ba8a3b1SNandita Dukkipati 2412ed66dfafSNeal Cardwell bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto) 24136ba8a3b1SNandita Dukkipati { 24146ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 24156ba8a3b1SNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 2416a2815817SNeal Cardwell u32 timeout, rto_delta_us; 24172ae21cf5SEric Dumazet int early_retrans; 24186ba8a3b1SNandita Dukkipati 24196ba8a3b1SNandita Dukkipati /* Don't do any loss probe on a Fast Open connection before 3WHS 24206ba8a3b1SNandita Dukkipati * finishes. 24216ba8a3b1SNandita Dukkipati */ 2422f9b99582SYuchung Cheng if (tp->fastopen_rsk) 24236ba8a3b1SNandita Dukkipati return false; 24246ba8a3b1SNandita Dukkipati 24252ae21cf5SEric Dumazet early_retrans = sock_net(sk)->ipv4.sysctl_tcp_early_retrans; 24266ba8a3b1SNandita Dukkipati /* Schedule a loss probe in 2*RTT for SACK capable connections 2427b4f70c3dSNeal Cardwell * not in loss recovery, that are either limited by cwnd or application. 24286ba8a3b1SNandita Dukkipati */ 24292ae21cf5SEric Dumazet if ((early_retrans != 3 && early_retrans != 4) || 2430bec41a11SYuchung Cheng !tp->packets_out || !tcp_is_sack(tp) || 2431b4f70c3dSNeal Cardwell (icsk->icsk_ca_state != TCP_CA_Open && 2432b4f70c3dSNeal Cardwell icsk->icsk_ca_state != TCP_CA_CWR)) 24336ba8a3b1SNandita Dukkipati return false; 24346ba8a3b1SNandita Dukkipati 2435bb4d991aSYuchung Cheng /* Probe timeout is 2*rtt. Add minimum RTO to account 2436f9b99582SYuchung Cheng * for delayed ack when there's one outstanding packet. If no RTT 2437f9b99582SYuchung Cheng * sample is available then probe after TCP_TIMEOUT_INIT. 24386ba8a3b1SNandita Dukkipati */ 2439bb4d991aSYuchung Cheng if (tp->srtt_us) { 2440bb4d991aSYuchung Cheng timeout = usecs_to_jiffies(tp->srtt_us >> 2); 24416ba8a3b1SNandita Dukkipati if (tp->packets_out == 1) 2442bb4d991aSYuchung Cheng timeout += TCP_RTO_MIN; 2443bb4d991aSYuchung Cheng else 2444bb4d991aSYuchung Cheng timeout += TCP_TIMEOUT_MIN; 2445bb4d991aSYuchung Cheng } else { 2446bb4d991aSYuchung Cheng timeout = TCP_TIMEOUT_INIT; 2447bb4d991aSYuchung Cheng } 24486ba8a3b1SNandita Dukkipati 2449a2815817SNeal Cardwell /* If the RTO formula yields an earlier time, then use that time. */ 2450ed66dfafSNeal Cardwell rto_delta_us = advancing_rto ? 2451ed66dfafSNeal Cardwell jiffies_to_usecs(inet_csk(sk)->icsk_rto) : 2452ed66dfafSNeal Cardwell tcp_rto_delta_us(sk); /* How far in future is RTO? */ 2453a2815817SNeal Cardwell if (rto_delta_us > 0) 2454a2815817SNeal Cardwell timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us)); 24556ba8a3b1SNandita Dukkipati 24566ba8a3b1SNandita Dukkipati inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, 24576ba8a3b1SNandita Dukkipati TCP_RTO_MAX); 24586ba8a3b1SNandita Dukkipati return true; 24596ba8a3b1SNandita Dukkipati } 24606ba8a3b1SNandita Dukkipati 24611f3279aeSEric Dumazet /* Thanks to skb fast clones, we can detect if a prior transmit of 24621f3279aeSEric Dumazet * a packet is still in a qdisc or driver queue. 24631f3279aeSEric Dumazet * In this case, there is very little point doing a retransmit ! 24641f3279aeSEric Dumazet */ 24651f3279aeSEric Dumazet static bool skb_still_in_host_queue(const struct sock *sk, 24661f3279aeSEric Dumazet const struct sk_buff *skb) 24671f3279aeSEric Dumazet { 246839bb5e62SEric Dumazet if (unlikely(skb_fclone_busy(sk, skb))) { 2469c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), 24701f3279aeSEric Dumazet LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES); 24711f3279aeSEric Dumazet return true; 24721f3279aeSEric Dumazet } 24731f3279aeSEric Dumazet return false; 24741f3279aeSEric Dumazet } 24751f3279aeSEric Dumazet 2476b340b264SYuchung Cheng /* When probe timeout (PTO) fires, try send a new segment if possible, else 24776ba8a3b1SNandita Dukkipati * retransmit the last segment. 24786ba8a3b1SNandita Dukkipati */ 24796ba8a3b1SNandita Dukkipati void tcp_send_loss_probe(struct sock *sk) 24806ba8a3b1SNandita Dukkipati { 24819b717a8dSNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 24826ba8a3b1SNandita Dukkipati struct sk_buff *skb; 24836ba8a3b1SNandita Dukkipati int pcount; 24846ba8a3b1SNandita Dukkipati int mss = tcp_current_mss(sk); 24856ba8a3b1SNandita Dukkipati 2486b340b264SYuchung Cheng skb = tcp_send_head(sk); 248775c119afSEric Dumazet if (skb && tcp_snd_wnd_test(tp, skb, mss)) { 2488b340b264SYuchung Cheng pcount = tp->packets_out; 2489b340b264SYuchung Cheng tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC); 2490b340b264SYuchung Cheng if (tp->packets_out > pcount) 2491b340b264SYuchung Cheng goto probe_sent; 24926ba8a3b1SNandita Dukkipati goto rearm_timer; 24936ba8a3b1SNandita Dukkipati } 249475c119afSEric Dumazet skb = skb_rb_last(&sk->tcp_rtx_queue); 24956ba8a3b1SNandita Dukkipati 24969b717a8dSNandita Dukkipati /* At most one outstanding TLP retransmission. */ 24979b717a8dSNandita Dukkipati if (tp->tlp_high_seq) 24989b717a8dSNandita Dukkipati goto rearm_timer; 24999b717a8dSNandita Dukkipati 25006ba8a3b1SNandita Dukkipati /* Retransmit last segment. */ 25016ba8a3b1SNandita Dukkipati if (WARN_ON(!skb)) 25026ba8a3b1SNandita Dukkipati goto rearm_timer; 25036ba8a3b1SNandita Dukkipati 25041f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 25051f3279aeSEric Dumazet goto rearm_timer; 25061f3279aeSEric Dumazet 25076ba8a3b1SNandita Dukkipati pcount = tcp_skb_pcount(skb); 25086ba8a3b1SNandita Dukkipati if (WARN_ON(!pcount)) 25096ba8a3b1SNandita Dukkipati goto rearm_timer; 25106ba8a3b1SNandita Dukkipati 25116ba8a3b1SNandita Dukkipati if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) { 251275c119afSEric Dumazet if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, 251375c119afSEric Dumazet (pcount - 1) * mss, mss, 25146cc55e09SOctavian Purdila GFP_ATOMIC))) 25156ba8a3b1SNandita Dukkipati goto rearm_timer; 251675c119afSEric Dumazet skb = skb_rb_next(skb); 25176ba8a3b1SNandita Dukkipati } 25186ba8a3b1SNandita Dukkipati 25196ba8a3b1SNandita Dukkipati if (WARN_ON(!skb || !tcp_skb_pcount(skb))) 25206ba8a3b1SNandita Dukkipati goto rearm_timer; 25216ba8a3b1SNandita Dukkipati 252210d3be56SEric Dumazet if (__tcp_retransmit_skb(sk, skb, 1)) 2523b340b264SYuchung Cheng goto rearm_timer; 25246ba8a3b1SNandita Dukkipati 25259b717a8dSNandita Dukkipati /* Record snd_nxt for loss detection. */ 25269b717a8dSNandita Dukkipati tp->tlp_high_seq = tp->snd_nxt; 25279b717a8dSNandita Dukkipati 2528b340b264SYuchung Cheng probe_sent: 2529c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES); 2530fcd16c0aSYuchung Cheng /* Reset s.t. tcp_rearm_rto will restart timer from now */ 2531fcd16c0aSYuchung Cheng inet_csk(sk)->icsk_pending = 0; 2532b340b264SYuchung Cheng rearm_timer: 2533fcd16c0aSYuchung Cheng tcp_rearm_rto(sk); 25341da177e4SLinus Torvalds } 25351da177e4SLinus Torvalds 2536a762a980SDavid S. Miller /* Push out any pending frames which were held back due to 2537a762a980SDavid S. Miller * TCP_CORK or attempt at coalescing tiny packets. 2538a762a980SDavid S. Miller * The socket must be locked by the caller. 2539a762a980SDavid S. Miller */ 25409e412ba7SIlpo Järvinen void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 25419e412ba7SIlpo Järvinen int nonagle) 2542a762a980SDavid S. Miller { 2543726e07a8SIlpo Järvinen /* If we are closed, the bytes will have to remain here. 2544726e07a8SIlpo Järvinen * In time closedown will finish, we empty the write queue and 2545726e07a8SIlpo Järvinen * all will be happy. 2546726e07a8SIlpo Järvinen */ 2547726e07a8SIlpo Järvinen if (unlikely(sk->sk_state == TCP_CLOSE)) 2548726e07a8SIlpo Järvinen return; 2549726e07a8SIlpo Järvinen 255099a1dec7SMel Gorman if (tcp_write_xmit(sk, cur_mss, nonagle, 0, 25517450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC))) 25529e412ba7SIlpo Järvinen tcp_check_probe_timer(sk); 2553a762a980SDavid S. Miller } 2554a762a980SDavid S. Miller 2555c1b4a7e6SDavid S. Miller /* Send _single_ skb sitting at the send head. This function requires 2556c1b4a7e6SDavid S. Miller * true push pending frames to setup probe timer etc. 2557c1b4a7e6SDavid S. Miller */ 2558c1b4a7e6SDavid S. Miller void tcp_push_one(struct sock *sk, unsigned int mss_now) 2559c1b4a7e6SDavid S. Miller { 2560fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 2561c1b4a7e6SDavid S. Miller 2562c1b4a7e6SDavid S. Miller BUG_ON(!skb || skb->len < mss_now); 2563c1b4a7e6SDavid S. Miller 2564d5dd9175SIlpo Järvinen tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); 2565c1b4a7e6SDavid S. Miller } 2566c1b4a7e6SDavid S. Miller 25671da177e4SLinus Torvalds /* This function returns the amount that we can raise the 25681da177e4SLinus Torvalds * usable window based on the following constraints 25691da177e4SLinus Torvalds * 25701da177e4SLinus Torvalds * 1. The window can never be shrunk once it is offered (RFC 793) 25711da177e4SLinus Torvalds * 2. We limit memory per socket 25721da177e4SLinus Torvalds * 25731da177e4SLinus Torvalds * RFC 1122: 25741da177e4SLinus Torvalds * "the suggested [SWS] avoidance algorithm for the receiver is to keep 25751da177e4SLinus Torvalds * RECV.NEXT + RCV.WIN fixed until: 25761da177e4SLinus Torvalds * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 25771da177e4SLinus Torvalds * 25781da177e4SLinus Torvalds * i.e. don't raise the right edge of the window until you can raise 25791da177e4SLinus Torvalds * it at least MSS bytes. 25801da177e4SLinus Torvalds * 25811da177e4SLinus Torvalds * Unfortunately, the recommended algorithm breaks header prediction, 25821da177e4SLinus Torvalds * since header prediction assumes th->window stays fixed. 25831da177e4SLinus Torvalds * 25841da177e4SLinus Torvalds * Strictly speaking, keeping th->window fixed violates the receiver 25851da177e4SLinus Torvalds * side SWS prevention criteria. The problem is that under this rule 25861da177e4SLinus Torvalds * a stream of single byte packets will cause the right side of the 25871da177e4SLinus Torvalds * window to always advance by a single byte. 25881da177e4SLinus Torvalds * 25891da177e4SLinus Torvalds * Of course, if the sender implements sender side SWS prevention 25901da177e4SLinus Torvalds * then this will not be a problem. 25911da177e4SLinus Torvalds * 25921da177e4SLinus Torvalds * BSD seems to make the following compromise: 25931da177e4SLinus Torvalds * 25941da177e4SLinus Torvalds * If the free space is less than the 1/4 of the maximum 25951da177e4SLinus Torvalds * space available and the free space is less than 1/2 mss, 25961da177e4SLinus Torvalds * then set the window to 0. 25971da177e4SLinus Torvalds * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 25981da177e4SLinus Torvalds * Otherwise, just prevent the window from shrinking 25991da177e4SLinus Torvalds * and from being larger than the largest representable value. 26001da177e4SLinus Torvalds * 26011da177e4SLinus Torvalds * This prevents incremental opening of the window in the regime 26021da177e4SLinus Torvalds * where TCP is limited by the speed of the reader side taking 26031da177e4SLinus Torvalds * data out of the TCP receive queue. It does nothing about 26041da177e4SLinus Torvalds * those cases where the window is constrained on the sender side 26051da177e4SLinus Torvalds * because the pipeline is full. 26061da177e4SLinus Torvalds * 26071da177e4SLinus Torvalds * BSD also seems to "accidentally" limit itself to windows that are a 26081da177e4SLinus Torvalds * multiple of MSS, at least until the free space gets quite small. 26091da177e4SLinus Torvalds * This would appear to be a side effect of the mbuf implementation. 26101da177e4SLinus Torvalds * Combining these two algorithms results in the observed behavior 26111da177e4SLinus Torvalds * of having a fixed window size at almost all times. 26121da177e4SLinus Torvalds * 26131da177e4SLinus Torvalds * Below we obtain similar behavior by forcing the offered window to 26141da177e4SLinus Torvalds * a multiple of the mss when it is feasible to do so. 26151da177e4SLinus Torvalds * 26161da177e4SLinus Torvalds * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 26171da177e4SLinus Torvalds * Regular options like TIMESTAMP are taken into account. 26181da177e4SLinus Torvalds */ 26191da177e4SLinus Torvalds u32 __tcp_select_window(struct sock *sk) 26201da177e4SLinus Torvalds { 2621463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 26221da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2623caa20d9aSStephen Hemminger /* MSS for the peer's data. Previous versions used mss_clamp 26241da177e4SLinus Torvalds * here. I don't know if the value based on our guesses 26251da177e4SLinus Torvalds * of peer's MSS is better for the performance. It's more correct 26261da177e4SLinus Torvalds * but may be worse for the performance because of rcv_mss 26271da177e4SLinus Torvalds * fluctuations. --SAW 1998/11/1 26281da177e4SLinus Torvalds */ 2629463c84b9SArnaldo Carvalho de Melo int mss = icsk->icsk_ack.rcv_mss; 26301da177e4SLinus Torvalds int free_space = tcp_space(sk); 263186c1a045SFlorian Westphal int allowed_space = tcp_full_space(sk); 263286c1a045SFlorian Westphal int full_space = min_t(int, tp->window_clamp, allowed_space); 26331da177e4SLinus Torvalds int window; 26341da177e4SLinus Torvalds 263506425c30SEric Dumazet if (unlikely(mss > full_space)) { 26361da177e4SLinus Torvalds mss = full_space; 263706425c30SEric Dumazet if (mss <= 0) 263806425c30SEric Dumazet return 0; 263906425c30SEric Dumazet } 2640b92edbe0SEric Dumazet if (free_space < (full_space >> 1)) { 2641463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.quick = 0; 26421da177e4SLinus Torvalds 2643b8da51ebSEric Dumazet if (tcp_under_memory_pressure(sk)) 2644056834d9SIlpo Järvinen tp->rcv_ssthresh = min(tp->rcv_ssthresh, 2645056834d9SIlpo Järvinen 4U * tp->advmss); 26461da177e4SLinus Torvalds 264786c1a045SFlorian Westphal /* free_space might become our new window, make sure we don't 264886c1a045SFlorian Westphal * increase it due to wscale. 264986c1a045SFlorian Westphal */ 265086c1a045SFlorian Westphal free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale); 265186c1a045SFlorian Westphal 265286c1a045SFlorian Westphal /* if free space is less than mss estimate, or is below 1/16th 265386c1a045SFlorian Westphal * of the maximum allowed, try to move to zero-window, else 265486c1a045SFlorian Westphal * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and 265586c1a045SFlorian Westphal * new incoming data is dropped due to memory limits. 265686c1a045SFlorian Westphal * With large window, mss test triggers way too late in order 265786c1a045SFlorian Westphal * to announce zero window in time before rmem limit kicks in. 265886c1a045SFlorian Westphal */ 265986c1a045SFlorian Westphal if (free_space < (allowed_space >> 4) || free_space < mss) 26601da177e4SLinus Torvalds return 0; 26611da177e4SLinus Torvalds } 26621da177e4SLinus Torvalds 26631da177e4SLinus Torvalds if (free_space > tp->rcv_ssthresh) 26641da177e4SLinus Torvalds free_space = tp->rcv_ssthresh; 26651da177e4SLinus Torvalds 26661da177e4SLinus Torvalds /* Don't do rounding if we are using window scaling, since the 26671da177e4SLinus Torvalds * scaled window will not line up with the MSS boundary anyway. 26681da177e4SLinus Torvalds */ 26691da177e4SLinus Torvalds if (tp->rx_opt.rcv_wscale) { 26701da177e4SLinus Torvalds window = free_space; 26711da177e4SLinus Torvalds 26721da177e4SLinus Torvalds /* Advertise enough space so that it won't get scaled away. 26731da177e4SLinus Torvalds * Import case: prevent zero window announcement if 26741da177e4SLinus Torvalds * 1<<rcv_wscale > mss. 26751da177e4SLinus Torvalds */ 26761935299dSGao Feng window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale)); 26771da177e4SLinus Torvalds } else { 26781935299dSGao Feng window = tp->rcv_wnd; 26791da177e4SLinus Torvalds /* Get the largest window that is a nice multiple of mss. 26801da177e4SLinus Torvalds * Window clamp already applied above. 26811da177e4SLinus Torvalds * If our current window offering is within 1 mss of the 26821da177e4SLinus Torvalds * free space we just keep it. This prevents the divide 26831da177e4SLinus Torvalds * and multiply from happening most of the time. 26841da177e4SLinus Torvalds * We also don't do any window rounding when the free space 26851da177e4SLinus Torvalds * is too small. 26861da177e4SLinus Torvalds */ 26871da177e4SLinus Torvalds if (window <= free_space - mss || window > free_space) 26881935299dSGao Feng window = rounddown(free_space, mss); 268984565070SJohn Heffner else if (mss == full_space && 2690b92edbe0SEric Dumazet free_space > window + (full_space >> 1)) 269184565070SJohn Heffner window = free_space; 26921da177e4SLinus Torvalds } 26931da177e4SLinus Torvalds 26941da177e4SLinus Torvalds return window; 26951da177e4SLinus Torvalds } 26961da177e4SLinus Torvalds 2697cfea5a68SMartin KaFai Lau void tcp_skb_collapse_tstamp(struct sk_buff *skb, 2698082ac2d5SMartin KaFai Lau const struct sk_buff *next_skb) 2699082ac2d5SMartin KaFai Lau { 27000a2cf20cSSoheil Hassas Yeganeh if (unlikely(tcp_has_tx_tstamp(next_skb))) { 27010a2cf20cSSoheil Hassas Yeganeh const struct skb_shared_info *next_shinfo = 27020a2cf20cSSoheil Hassas Yeganeh skb_shinfo(next_skb); 2703082ac2d5SMartin KaFai Lau struct skb_shared_info *shinfo = skb_shinfo(skb); 2704082ac2d5SMartin KaFai Lau 27050a2cf20cSSoheil Hassas Yeganeh shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP; 2706082ac2d5SMartin KaFai Lau shinfo->tskey = next_shinfo->tskey; 27072de8023eSMartin KaFai Lau TCP_SKB_CB(skb)->txstamp_ack |= 27082de8023eSMartin KaFai Lau TCP_SKB_CB(next_skb)->txstamp_ack; 2709082ac2d5SMartin KaFai Lau } 2710082ac2d5SMartin KaFai Lau } 2711082ac2d5SMartin KaFai Lau 27124a17fc3aSIlpo Järvinen /* Collapses two adjacent SKB's during retransmission. */ 2713f8071cdeSEric Dumazet static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) 27141da177e4SLinus Torvalds { 27151da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 271675c119afSEric Dumazet struct sk_buff *next_skb = skb_rb_next(skb); 271713dde04fSWei Yongjun int next_skb_size; 27181da177e4SLinus Torvalds 2719058dc334SIlpo Järvinen next_skb_size = next_skb->len; 27201da177e4SLinus Torvalds 2721058dc334SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); 27221da177e4SLinus Torvalds 2723f8071cdeSEric Dumazet if (next_skb_size) { 2724f8071cdeSEric Dumazet if (next_skb_size <= skb_availroom(skb)) 2725f8071cdeSEric Dumazet skb_copy_bits(next_skb, 0, skb_put(skb, next_skb_size), 2726f8071cdeSEric Dumazet next_skb_size); 2727f8071cdeSEric Dumazet else if (!skb_shift(skb, next_skb, next_skb_size)) 2728f8071cdeSEric Dumazet return false; 2729f8071cdeSEric Dumazet } 27302b7cda9cSEric Dumazet tcp_highest_sack_replace(sk, next_skb, skb); 2731a6963a6bSIlpo Järvinen 27321da177e4SLinus Torvalds /* Update sequence range on original skb. */ 27331da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 27341da177e4SLinus Torvalds 2735e6c7d085SIlpo Järvinen /* Merge over control information. This moves PSH/FIN etc. over */ 27364de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; 27371da177e4SLinus Torvalds 27381da177e4SLinus Torvalds /* All done, get rid of second SKB and account for it so 27391da177e4SLinus Torvalds * packet counting does not break. 27401da177e4SLinus Torvalds */ 27414828e7f4SIlpo Järvinen TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; 2742a643b5d4SMartin KaFai Lau TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor; 2743b7689205SIlpo Järvinen 2744b7689205SIlpo Järvinen /* changed transmit queue under us so clear hints */ 2745ef9da47cSIlpo Järvinen tcp_clear_retrans_hints_partial(tp); 2746ef9da47cSIlpo Järvinen if (next_skb == tp->retransmit_skb_hint) 2747ef9da47cSIlpo Järvinen tp->retransmit_skb_hint = skb; 2748b7689205SIlpo Järvinen 2749797108d1SIlpo Järvinen tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); 2750797108d1SIlpo Järvinen 2751082ac2d5SMartin KaFai Lau tcp_skb_collapse_tstamp(skb, next_skb); 2752082ac2d5SMartin KaFai Lau 275375c119afSEric Dumazet tcp_rtx_queue_unlink_and_free(next_skb, sk); 2754f8071cdeSEric Dumazet return true; 27551da177e4SLinus Torvalds } 27561da177e4SLinus Torvalds 275767edfef7SAndi Kleen /* Check if coalescing SKBs is legal. */ 2758a2a385d6SEric Dumazet static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) 27594a17fc3aSIlpo Järvinen { 27604a17fc3aSIlpo Järvinen if (tcp_skb_pcount(skb) > 1) 2761a2a385d6SEric Dumazet return false; 27624a17fc3aSIlpo Järvinen if (skb_cloned(skb)) 2763a2a385d6SEric Dumazet return false; 27642331ccc5SEric Dumazet /* Some heuristics for collapsing over SACK'd could be invented */ 27654a17fc3aSIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 2766a2a385d6SEric Dumazet return false; 27674a17fc3aSIlpo Järvinen 2768a2a385d6SEric Dumazet return true; 27694a17fc3aSIlpo Järvinen } 27704a17fc3aSIlpo Järvinen 277167edfef7SAndi Kleen /* Collapse packets in the retransmit queue to make to create 277267edfef7SAndi Kleen * less packets on the wire. This is only done on retransmission. 277367edfef7SAndi Kleen */ 27744a17fc3aSIlpo Järvinen static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, 27754a17fc3aSIlpo Järvinen int space) 27764a17fc3aSIlpo Järvinen { 27774a17fc3aSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 27784a17fc3aSIlpo Järvinen struct sk_buff *skb = to, *tmp; 2779a2a385d6SEric Dumazet bool first = true; 27804a17fc3aSIlpo Järvinen 2781e0a1e5b5SEric Dumazet if (!sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse) 27824a17fc3aSIlpo Järvinen return; 27834de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 27844a17fc3aSIlpo Järvinen return; 27854a17fc3aSIlpo Järvinen 278675c119afSEric Dumazet skb_rbtree_walk_from_safe(skb, tmp) { 27874a17fc3aSIlpo Järvinen if (!tcp_can_collapse(sk, skb)) 27884a17fc3aSIlpo Järvinen break; 27894a17fc3aSIlpo Järvinen 2790a643b5d4SMartin KaFai Lau if (!tcp_skb_can_collapse_to(to)) 2791a643b5d4SMartin KaFai Lau break; 2792a643b5d4SMartin KaFai Lau 27934a17fc3aSIlpo Järvinen space -= skb->len; 27944a17fc3aSIlpo Järvinen 27954a17fc3aSIlpo Järvinen if (first) { 2796a2a385d6SEric Dumazet first = false; 27974a17fc3aSIlpo Järvinen continue; 27984a17fc3aSIlpo Järvinen } 27994a17fc3aSIlpo Järvinen 28004a17fc3aSIlpo Järvinen if (space < 0) 28014a17fc3aSIlpo Järvinen break; 28024a17fc3aSIlpo Järvinen 28034a17fc3aSIlpo Järvinen if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) 28044a17fc3aSIlpo Järvinen break; 28054a17fc3aSIlpo Järvinen 2806f8071cdeSEric Dumazet if (!tcp_collapse_retrans(sk, to)) 2807f8071cdeSEric Dumazet break; 28084a17fc3aSIlpo Järvinen } 28094a17fc3aSIlpo Järvinen } 28104a17fc3aSIlpo Järvinen 28111da177e4SLinus Torvalds /* This retransmits one SKB. Policy decisions and retransmit queue 28121da177e4SLinus Torvalds * state updates are done by the caller. Returns non-zero if an 28131da177e4SLinus Torvalds * error occurred which prevented the send. 28141da177e4SLinus Torvalds */ 281510d3be56SEric Dumazet int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) 28161da177e4SLinus Torvalds { 28175d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 281810d3be56SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 28197d227cd2SSridhar Samudrala unsigned int cur_mss; 282010d3be56SEric Dumazet int diff, len, err; 28211da177e4SLinus Torvalds 282210d3be56SEric Dumazet 282310d3be56SEric Dumazet /* Inconclusive MTU probe */ 282410d3be56SEric Dumazet if (icsk->icsk_mtup.probe_size) 28255d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 28265d424d5aSJohn Heffner 28271da177e4SLinus Torvalds /* Do not sent more than we queued. 1/4 is reserved for possible 2828caa20d9aSStephen Hemminger * copying overhead: fragmentation, tunneling, mangling etc. 28291da177e4SLinus Torvalds */ 283014afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) > 2831ffb4d6c8SEric Dumazet min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), 2832ffb4d6c8SEric Dumazet sk->sk_sndbuf)) 28331da177e4SLinus Torvalds return -EAGAIN; 28341da177e4SLinus Torvalds 28351f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 28361f3279aeSEric Dumazet return -EBUSY; 28371f3279aeSEric Dumazet 28381da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 28397f582b24SEric Dumazet if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) { 28407f582b24SEric Dumazet WARN_ON_ONCE(1); 28417f582b24SEric Dumazet return -EINVAL; 28427f582b24SEric Dumazet } 28431da177e4SLinus Torvalds if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 28441da177e4SLinus Torvalds return -ENOMEM; 28451da177e4SLinus Torvalds } 28461da177e4SLinus Torvalds 28477d227cd2SSridhar Samudrala if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 28487d227cd2SSridhar Samudrala return -EHOSTUNREACH; /* Routing failure or similar. */ 28497d227cd2SSridhar Samudrala 28500c54b85fSIlpo Järvinen cur_mss = tcp_current_mss(sk); 28517d227cd2SSridhar Samudrala 28521da177e4SLinus Torvalds /* If receiver has shrunk his window, and skb is out of 28531da177e4SLinus Torvalds * new window, do not retransmit it. The exception is the 28541da177e4SLinus Torvalds * case, when window is shrunk to zero. In this case 28551da177e4SLinus Torvalds * our retransmit serves as a zero window probe. 28561da177e4SLinus Torvalds */ 28579d4fb27dSJoe Perches if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && 28589d4fb27dSJoe Perches TCP_SKB_CB(skb)->seq != tp->snd_una) 28591da177e4SLinus Torvalds return -EAGAIN; 28601da177e4SLinus Torvalds 286110d3be56SEric Dumazet len = cur_mss * segs; 286210d3be56SEric Dumazet if (skb->len > len) { 286375c119afSEric Dumazet if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len, 286475c119afSEric Dumazet cur_mss, GFP_ATOMIC)) 28651da177e4SLinus Torvalds return -ENOMEM; /* We'll try again later. */ 286602276f3cSIlpo Järvinen } else { 2867c52e2421SEric Dumazet if (skb_unclone(skb, GFP_ATOMIC)) 2868c52e2421SEric Dumazet return -ENOMEM; 286910d3be56SEric Dumazet 287010d3be56SEric Dumazet diff = tcp_skb_pcount(skb); 287110d3be56SEric Dumazet tcp_set_skb_tso_segs(skb, cur_mss); 287210d3be56SEric Dumazet diff -= tcp_skb_pcount(skb); 287310d3be56SEric Dumazet if (diff) 287410d3be56SEric Dumazet tcp_adjust_pcount(sk, skb, diff); 287510d3be56SEric Dumazet if (skb->len < cur_mss) 287610d3be56SEric Dumazet tcp_retrans_try_collapse(sk, skb, cur_mss); 28771da177e4SLinus Torvalds } 28781da177e4SLinus Torvalds 287949213555SDaniel Borkmann /* RFC3168, section 6.1.1.1. ECN fallback */ 288049213555SDaniel Borkmann if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN) 288149213555SDaniel Borkmann tcp_ecn_clear_syn(sk, skb); 288249213555SDaniel Borkmann 2883678550c6SYuchung Cheng /* Update global and local TCP statistics. */ 2884678550c6SYuchung Cheng segs = tcp_skb_pcount(skb); 2885678550c6SYuchung Cheng TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs); 2886678550c6SYuchung Cheng if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 2887678550c6SYuchung Cheng __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 2888678550c6SYuchung Cheng tp->total_retrans += segs; 2889fb31c9b9SWei Wang tp->bytes_retrans += skb->len; 2890678550c6SYuchung Cheng 289150bceae9SThomas Graf /* make sure skb->data is aligned on arches that require it 289250bceae9SThomas Graf * and check if ack-trimming & collapsing extended the headroom 289350bceae9SThomas Graf * beyond what csum_start can cover. 289450bceae9SThomas Graf */ 289550bceae9SThomas Graf if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) || 289650bceae9SThomas Graf skb_headroom(skb) >= 0xFFFF)) { 289710a81980SEric Dumazet struct sk_buff *nskb; 289810a81980SEric Dumazet 2899e2080072SEric Dumazet tcp_skb_tsorted_save(skb) { 290010a81980SEric Dumazet nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC); 2901c84a5711SYuchung Cheng err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) : 2902117632e6SEric Dumazet -ENOBUFS; 2903e2080072SEric Dumazet } tcp_skb_tsorted_restore(skb); 2904e2080072SEric Dumazet 29055889e2c0SYousuk Seung if (!err) { 2906*a7a25630SEric Dumazet tcp_update_skb_after_send(sk, skb, tp->tcp_wstamp_ns); 29075889e2c0SYousuk Seung tcp_rate_skb_sent(sk, skb); 29085889e2c0SYousuk Seung } 2909117632e6SEric Dumazet } else { 2910c84a5711SYuchung Cheng err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 2911117632e6SEric Dumazet } 2912c84a5711SYuchung Cheng 2913a31ad29eSLawrence Brakmo if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RETRANS_CB_FLAG)) 2914a31ad29eSLawrence Brakmo tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RETRANS_CB, 2915a31ad29eSLawrence Brakmo TCP_SKB_CB(skb)->seq, segs, err); 2916a31ad29eSLawrence Brakmo 2917fc9f3501SEric Dumazet if (likely(!err)) { 2918c84a5711SYuchung Cheng TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS; 2919e086101bSCong Wang trace_tcp_retransmit_skb(sk, skb); 2920678550c6SYuchung Cheng } else if (err != -EBUSY) { 2921678550c6SYuchung Cheng NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL); 2922fc9f3501SEric Dumazet } 2923c84a5711SYuchung Cheng return err; 292493b174adSYuchung Cheng } 292593b174adSYuchung Cheng 292610d3be56SEric Dumazet int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) 292793b174adSYuchung Cheng { 292893b174adSYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 292910d3be56SEric Dumazet int err = __tcp_retransmit_skb(sk, skb, segs); 29301da177e4SLinus Torvalds 29311da177e4SLinus Torvalds if (err == 0) { 29321da177e4SLinus Torvalds #if FASTRETRANS_DEBUG > 0 29331da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 2934e87cc472SJoe Perches net_dbg_ratelimited("retrans_out leaked\n"); 29351da177e4SLinus Torvalds } 29361da177e4SLinus Torvalds #endif 29371da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 29381da177e4SLinus Torvalds tp->retrans_out += tcp_skb_pcount(skb); 29391da177e4SLinus Torvalds 29401da177e4SLinus Torvalds /* Save stamp of the first retransmit. */ 29411da177e4SLinus Torvalds if (!tp->retrans_stamp) 29427faee5c0SEric Dumazet tp->retrans_stamp = tcp_skb_timestamp(skb); 29431da177e4SLinus Torvalds 29441da177e4SLinus Torvalds } 29456e08d5e3SYuchung Cheng 29466e08d5e3SYuchung Cheng if (tp->undo_retrans < 0) 29476e08d5e3SYuchung Cheng tp->undo_retrans = 0; 29486e08d5e3SYuchung Cheng tp->undo_retrans += tcp_skb_pcount(skb); 29491da177e4SLinus Torvalds return err; 29501da177e4SLinus Torvalds } 29511da177e4SLinus Torvalds 29521da177e4SLinus Torvalds /* This gets called after a retransmit timeout, and the initially 29531da177e4SLinus Torvalds * retransmitted data is acknowledged. It tries to continue 29541da177e4SLinus Torvalds * resending the rest of the retransmit queue, until either 29551da177e4SLinus Torvalds * we've sent it all or the congestion window limit is reached. 29561da177e4SLinus Torvalds */ 29571da177e4SLinus Torvalds void tcp_xmit_retransmit_queue(struct sock *sk) 29581da177e4SLinus Torvalds { 29596687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 2960b9f1f1ceSEric Dumazet struct sk_buff *skb, *rtx_head, *hole = NULL; 29611da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2962840a3cbeSYuchung Cheng u32 max_segs; 296361eb55f4SIlpo Järvinen int mib_idx; 29646a438bbeSStephen Hemminger 296545e77d31SIlpo Järvinen if (!tp->packets_out) 296645e77d31SIlpo Järvinen return; 296745e77d31SIlpo Järvinen 296875c119afSEric Dumazet rtx_head = tcp_rtx_queue_head(sk); 2969b9f1f1ceSEric Dumazet skb = tp->retransmit_skb_hint ?: rtx_head; 2970ed6e7268SNeal Cardwell max_segs = tcp_tso_segs(sk, tcp_current_mss(sk)); 297175c119afSEric Dumazet skb_rbtree_walk_from(skb) { 2972dca0aaf8SEric Dumazet __u8 sacked; 297310d3be56SEric Dumazet int segs; 29741da177e4SLinus Torvalds 2975218af599SEric Dumazet if (tcp_pacing_check(sk)) 2976218af599SEric Dumazet break; 2977218af599SEric Dumazet 29786a438bbeSStephen Hemminger /* we could do better than to assign each time */ 297951456b29SIan Morris if (!hole) 29806a438bbeSStephen Hemminger tp->retransmit_skb_hint = skb; 29816a438bbeSStephen Hemminger 298210d3be56SEric Dumazet segs = tp->snd_cwnd - tcp_packets_in_flight(tp); 298310d3be56SEric Dumazet if (segs <= 0) 29841da177e4SLinus Torvalds return; 2985dca0aaf8SEric Dumazet sacked = TCP_SKB_CB(skb)->sacked; 2986a3d2e9f8SEric Dumazet /* In case tcp_shift_skb_data() have aggregated large skbs, 2987a3d2e9f8SEric Dumazet * we need to make sure not sending too bigs TSO packets 2988a3d2e9f8SEric Dumazet */ 2989a3d2e9f8SEric Dumazet segs = min_t(int, segs, max_segs); 29900e1c54c2SIlpo Järvinen 2991840a3cbeSYuchung Cheng if (tp->retrans_out >= tp->lost_out) { 2992006f582cSIlpo Järvinen break; 29930e1c54c2SIlpo Järvinen } else if (!(sacked & TCPCB_LOST)) { 299451456b29SIan Morris if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) 29950e1c54c2SIlpo Järvinen hole = skb; 299661eb55f4SIlpo Järvinen continue; 29971da177e4SLinus Torvalds 29980e1c54c2SIlpo Järvinen } else { 29990e1c54c2SIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Loss) 30000e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFASTRETRANS; 30010e1c54c2SIlpo Järvinen else 30020e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; 30030e1c54c2SIlpo Järvinen } 30040e1c54c2SIlpo Järvinen 30050e1c54c2SIlpo Järvinen if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) 300661eb55f4SIlpo Järvinen continue; 300740b215e5SPavel Emelyanov 3008f9616c35SEric Dumazet if (tcp_small_queue_check(sk, skb, 1)) 3009f9616c35SEric Dumazet return; 3010f9616c35SEric Dumazet 301110d3be56SEric Dumazet if (tcp_retransmit_skb(sk, skb, segs)) 30121da177e4SLinus Torvalds return; 301324ab6becSYuchung Cheng 3014de1d6578SYuchung Cheng NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb)); 30151da177e4SLinus Torvalds 3016684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 3017a262f0cdSNandita Dukkipati tp->prr_out += tcp_skb_pcount(skb); 3018a262f0cdSNandita Dukkipati 301975c119afSEric Dumazet if (skb == rtx_head && 302057dde7f7SYuchung Cheng icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT) 3021463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 30223f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, 30233f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 30241da177e4SLinus Torvalds } 30251da177e4SLinus Torvalds } 30261da177e4SLinus Torvalds 3027d83769a5SEric Dumazet /* We allow to exceed memory limits for FIN packets to expedite 3028d83769a5SEric Dumazet * connection tear down and (memory) recovery. 3029845704a5SEric Dumazet * Otherwise tcp_send_fin() could be tempted to either delay FIN 3030845704a5SEric Dumazet * or even be forced to close flow without any FIN. 3031a6c5ea4cSEric Dumazet * In general, we want to allow one skb per socket to avoid hangs 3032a6c5ea4cSEric Dumazet * with edge trigger epoll() 3033d83769a5SEric Dumazet */ 3034a6c5ea4cSEric Dumazet void sk_forced_mem_schedule(struct sock *sk, int size) 3035d83769a5SEric Dumazet { 3036e805605cSJohannes Weiner int amt; 3037d83769a5SEric Dumazet 3038d83769a5SEric Dumazet if (size <= sk->sk_forward_alloc) 3039d83769a5SEric Dumazet return; 3040d83769a5SEric Dumazet amt = sk_mem_pages(size); 3041d83769a5SEric Dumazet sk->sk_forward_alloc += amt * SK_MEM_QUANTUM; 3042e805605cSJohannes Weiner sk_memory_allocated_add(sk, amt); 3043e805605cSJohannes Weiner 3044baac50bbSJohannes Weiner if (mem_cgroup_sockets_enabled && sk->sk_memcg) 3045baac50bbSJohannes Weiner mem_cgroup_charge_skmem(sk->sk_memcg, amt); 3046d83769a5SEric Dumazet } 3047d83769a5SEric Dumazet 3048845704a5SEric Dumazet /* Send a FIN. The caller locks the socket for us. 3049845704a5SEric Dumazet * We should try to send a FIN packet really hard, but eventually give up. 30501da177e4SLinus Torvalds */ 30511da177e4SLinus Torvalds void tcp_send_fin(struct sock *sk) 30521da177e4SLinus Torvalds { 3053845704a5SEric Dumazet struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk); 30541da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 30551da177e4SLinus Torvalds 3056845704a5SEric Dumazet /* Optimization, tack on the FIN if we have one skb in write queue and 3057845704a5SEric Dumazet * this skb was not yet sent, or we are under memory pressure. 3058845704a5SEric Dumazet * Note: in the latter case, FIN packet will be sent after a timeout, 3059845704a5SEric Dumazet * as TCP stack thinks it has already been transmitted. 30601da177e4SLinus Torvalds */ 306175c119afSEric Dumazet if (!tskb && tcp_under_memory_pressure(sk)) 306275c119afSEric Dumazet tskb = skb_rb_last(&sk->tcp_rtx_queue); 306375c119afSEric Dumazet 306475c119afSEric Dumazet if (tskb) { 3065845704a5SEric Dumazet coalesce: 3066845704a5SEric Dumazet TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN; 3067845704a5SEric Dumazet TCP_SKB_CB(tskb)->end_seq++; 30681da177e4SLinus Torvalds tp->write_seq++; 306975c119afSEric Dumazet if (tcp_write_queue_empty(sk)) { 3070845704a5SEric Dumazet /* This means tskb was already sent. 3071845704a5SEric Dumazet * Pretend we included the FIN on previous transmit. 3072845704a5SEric Dumazet * We need to set tp->snd_nxt to the value it would have 3073845704a5SEric Dumazet * if FIN had been sent. This is because retransmit path 3074845704a5SEric Dumazet * does not change tp->snd_nxt. 3075845704a5SEric Dumazet */ 3076845704a5SEric Dumazet tp->snd_nxt++; 3077845704a5SEric Dumazet return; 3078845704a5SEric Dumazet } 30791da177e4SLinus Torvalds } else { 3080845704a5SEric Dumazet skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation); 3081845704a5SEric Dumazet if (unlikely(!skb)) { 3082845704a5SEric Dumazet if (tskb) 3083845704a5SEric Dumazet goto coalesce; 3084845704a5SEric Dumazet return; 30851da177e4SLinus Torvalds } 3086e2080072SEric Dumazet INIT_LIST_HEAD(&skb->tcp_tsorted_anchor); 3087d83769a5SEric Dumazet skb_reserve(skb, MAX_TCP_HEADER); 3088a6c5ea4cSEric Dumazet sk_forced_mem_schedule(sk, skb->truesize); 30891da177e4SLinus Torvalds /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 3090e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tp->write_seq, 3091a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_FIN); 30921da177e4SLinus Torvalds tcp_queue_skb(sk, skb); 30931da177e4SLinus Torvalds } 3094845704a5SEric Dumazet __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF); 30951da177e4SLinus Torvalds } 30961da177e4SLinus Torvalds 30971da177e4SLinus Torvalds /* We get here when a process closes a file descriptor (either due to 30981da177e4SLinus Torvalds * an explicit close() or as a byproduct of exit()'ing) and there 30991da177e4SLinus Torvalds * was unread data in the receive queue. This behavior is recommended 310065bb723cSGerrit Renker * by RFC 2525, section 2.17. -DaveM 31011da177e4SLinus Torvalds */ 3102dd0fc66fSAl Viro void tcp_send_active_reset(struct sock *sk, gfp_t priority) 31031da177e4SLinus Torvalds { 31041da177e4SLinus Torvalds struct sk_buff *skb; 31051da177e4SLinus Torvalds 31067cc2b043SGao Feng TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); 31077cc2b043SGao Feng 31081da177e4SLinus Torvalds /* NOTE: No TCP options attached and we never retransmit this. */ 31091da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, priority); 31101da177e4SLinus Torvalds if (!skb) { 31114e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 31121da177e4SLinus Torvalds return; 31131da177e4SLinus Torvalds } 31141da177e4SLinus Torvalds 31151da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 31161da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 3117e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), 3118a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_RST); 31199a568de4SEric Dumazet tcp_mstamp_refresh(tcp_sk(sk)); 31201da177e4SLinus Torvalds /* Send it off. */ 3121dfb4b9dcSDavid S. Miller if (tcp_transmit_skb(sk, skb, 0, priority)) 31224e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 3123c24b14c4SSong Liu 3124c24b14c4SSong Liu /* skb of trace_tcp_send_reset() keeps the skb that caused RST, 3125c24b14c4SSong Liu * skb here is different to the troublesome skb, so use NULL 3126c24b14c4SSong Liu */ 3127c24b14c4SSong Liu trace_tcp_send_reset(sk, NULL); 31281da177e4SLinus Torvalds } 31291da177e4SLinus Torvalds 313067edfef7SAndi Kleen /* Send a crossed SYN-ACK during socket establishment. 313167edfef7SAndi Kleen * WARNING: This routine must only be called when we have already sent 31321da177e4SLinus Torvalds * a SYN packet that crossed the incoming SYN that caused this routine 31331da177e4SLinus Torvalds * to get called. If this assumption fails then the initial rcv_wnd 31341da177e4SLinus Torvalds * and rcv_wscale values will not be correct. 31351da177e4SLinus Torvalds */ 31361da177e4SLinus Torvalds int tcp_send_synack(struct sock *sk) 31371da177e4SLinus Torvalds { 31381da177e4SLinus Torvalds struct sk_buff *skb; 31391da177e4SLinus Torvalds 314075c119afSEric Dumazet skb = tcp_rtx_queue_head(sk); 314151456b29SIan Morris if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 314275c119afSEric Dumazet pr_err("%s: wrong queue state\n", __func__); 31431da177e4SLinus Torvalds return -EFAULT; 31441da177e4SLinus Torvalds } 31454de075e0SEric Dumazet if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { 31461da177e4SLinus Torvalds if (skb_cloned(skb)) { 3147e2080072SEric Dumazet struct sk_buff *nskb; 3148e2080072SEric Dumazet 3149e2080072SEric Dumazet tcp_skb_tsorted_save(skb) { 3150e2080072SEric Dumazet nskb = skb_copy(skb, GFP_ATOMIC); 3151e2080072SEric Dumazet } tcp_skb_tsorted_restore(skb); 315251456b29SIan Morris if (!nskb) 31531da177e4SLinus Torvalds return -ENOMEM; 3154e2080072SEric Dumazet INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor); 315575c119afSEric Dumazet tcp_rtx_queue_unlink_and_free(skb, sk); 3156f4a775d1SEric Dumazet __skb_header_release(nskb); 315775c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb); 31583ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 31593ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 31601da177e4SLinus Torvalds skb = nskb; 31611da177e4SLinus Torvalds } 31621da177e4SLinus Torvalds 31634de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; 3164735d3831SFlorian Westphal tcp_ecn_send_synack(sk, skb); 31651da177e4SLinus Torvalds } 3166dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 31671da177e4SLinus Torvalds } 31681da177e4SLinus Torvalds 31694aea39c1SEric Dumazet /** 31704aea39c1SEric Dumazet * tcp_make_synack - Prepare a SYN-ACK. 31714aea39c1SEric Dumazet * sk: listener socket 31724aea39c1SEric Dumazet * dst: dst entry attached to the SYNACK 31734aea39c1SEric Dumazet * req: request_sock pointer 31744aea39c1SEric Dumazet * 31754aea39c1SEric Dumazet * Allocate one skb and build a SYNACK packet. 31764aea39c1SEric Dumazet * @dst is consumed : Caller should not use it again. 31774aea39c1SEric Dumazet */ 31785d062de7SEric Dumazet struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst, 3179e6b4d113SWilliam Allen Simpson struct request_sock *req, 3180ca6fb065SEric Dumazet struct tcp_fastopen_cookie *foc, 3181b3d05147SEric Dumazet enum tcp_synack_type synack_type) 31821da177e4SLinus Torvalds { 31832e6599cbSArnaldo Carvalho de Melo struct inet_request_sock *ireq = inet_rsk(req); 31845d062de7SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 318580f03e27SEric Dumazet struct tcp_md5sig_key *md5 = NULL; 31865d062de7SEric Dumazet struct tcp_out_options opts; 31875d062de7SEric Dumazet struct sk_buff *skb; 3188bd0388aeSWilliam Allen Simpson int tcp_header_size; 31895d062de7SEric Dumazet struct tcphdr *th; 3190f5fff5dcSTom Quetchenbach int mss; 31911da177e4SLinus Torvalds 3192ca6fb065SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 31934aea39c1SEric Dumazet if (unlikely(!skb)) { 31944aea39c1SEric Dumazet dst_release(dst); 31951da177e4SLinus Torvalds return NULL; 31964aea39c1SEric Dumazet } 31971da177e4SLinus Torvalds /* Reserve space for headers. */ 31981da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 31991da177e4SLinus Torvalds 3200b3d05147SEric Dumazet switch (synack_type) { 3201b3d05147SEric Dumazet case TCP_SYNACK_NORMAL: 32029e17f8a4SEric Dumazet skb_set_owner_w(skb, req_to_sk(req)); 3203b3d05147SEric Dumazet break; 3204b3d05147SEric Dumazet case TCP_SYNACK_COOKIE: 3205b3d05147SEric Dumazet /* Under synflood, we do not attach skb to a socket, 3206b3d05147SEric Dumazet * to avoid false sharing. 3207b3d05147SEric Dumazet */ 3208b3d05147SEric Dumazet break; 3209b3d05147SEric Dumazet case TCP_SYNACK_FASTOPEN: 3210ca6fb065SEric Dumazet /* sk is a const pointer, because we want to express multiple 3211ca6fb065SEric Dumazet * cpu might call us concurrently. 3212ca6fb065SEric Dumazet * sk->sk_wmem_alloc in an atomic, we can promote to rw. 3213ca6fb065SEric Dumazet */ 3214ca6fb065SEric Dumazet skb_set_owner_w(skb, (struct sock *)sk); 3215b3d05147SEric Dumazet break; 3216ca6fb065SEric Dumazet } 32174aea39c1SEric Dumazet skb_dst_set(skb, dst); 32181da177e4SLinus Torvalds 32193541f9e8SEric Dumazet mss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 3220f5fff5dcSTom Quetchenbach 322133ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 32228b5f12d0SFlorian Westphal #ifdef CONFIG_SYN_COOKIES 32238b5f12d0SFlorian Westphal if (unlikely(req->cookie_ts)) 3224d3edd06eSEric Dumazet skb->skb_mstamp_ns = cookie_init_timestamp(req); 32258b5f12d0SFlorian Westphal else 32268b5f12d0SFlorian Westphal #endif 3227d3edd06eSEric Dumazet skb->skb_mstamp_ns = tcp_clock_ns(); 322880f03e27SEric Dumazet 322980f03e27SEric Dumazet #ifdef CONFIG_TCP_MD5SIG 323080f03e27SEric Dumazet rcu_read_lock(); 3231fd3a154aSEric Dumazet md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req)); 323280f03e27SEric Dumazet #endif 323358d607d3SEric Dumazet skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4); 323460e2a778SUrsula Braun tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5, 323560e2a778SUrsula Braun foc) + sizeof(*th); 323633ad798cSAdam Langley 3237aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 3238aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 32391da177e4SLinus Torvalds 3240ea1627c2SEric Dumazet th = (struct tcphdr *)skb->data; 32411da177e4SLinus Torvalds memset(th, 0, sizeof(struct tcphdr)); 32421da177e4SLinus Torvalds th->syn = 1; 32431da177e4SLinus Torvalds th->ack = 1; 32446ac705b1SEric Dumazet tcp_ecn_make_synack(req, th); 3245b44084c2SEric Dumazet th->source = htons(ireq->ir_num); 3246634fb979SEric Dumazet th->dest = ireq->ir_rmt_port; 3247e05a90ecSJamal Hadi Salim skb->mark = ireq->ir_mark; 32483b117750SEric Dumazet skb->ip_summed = CHECKSUM_PARTIAL; 32493b117750SEric Dumazet th->seq = htonl(tcp_rsk(req)->snt_isn); 32508336886fSJerry Chu /* XXX data is queued and acked as is. No buffer/window check */ 32518336886fSJerry Chu th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt); 32521da177e4SLinus Torvalds 32531da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 3254ed53d0abSEric Dumazet th->window = htons(min(req->rsk_rcv_wnd, 65535U)); 32555d062de7SEric Dumazet tcp_options_write((__be32 *)(th + 1), NULL, &opts); 32561da177e4SLinus Torvalds th->doff = (tcp_header_size >> 2); 325790bbcc60SEric Dumazet __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS); 3258cfb6eeb4SYOSHIFUJI Hideaki 3259cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 3260cfb6eeb4SYOSHIFUJI Hideaki /* Okay, we have all we need - do the md5 hash if needed */ 326180f03e27SEric Dumazet if (md5) 3262bd0388aeSWilliam Allen Simpson tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, 326339f8e58eSEric Dumazet md5, req_to_sk(req), skb); 326480f03e27SEric Dumazet rcu_read_unlock(); 3265cfb6eeb4SYOSHIFUJI Hideaki #endif 3266cfb6eeb4SYOSHIFUJI Hideaki 3267b50edd78SEric Dumazet /* Do not fool tcpdump (if any), clean our debris */ 32682456e855SThomas Gleixner skb->tstamp = 0; 32691da177e4SLinus Torvalds return skb; 32701da177e4SLinus Torvalds } 32714bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_make_synack); 32721da177e4SLinus Torvalds 327381164413SDaniel Borkmann static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst) 327481164413SDaniel Borkmann { 327581164413SDaniel Borkmann struct inet_connection_sock *icsk = inet_csk(sk); 327681164413SDaniel Borkmann const struct tcp_congestion_ops *ca; 327781164413SDaniel Borkmann u32 ca_key = dst_metric(dst, RTAX_CC_ALGO); 327881164413SDaniel Borkmann 327981164413SDaniel Borkmann if (ca_key == TCP_CA_UNSPEC) 328081164413SDaniel Borkmann return; 328181164413SDaniel Borkmann 328281164413SDaniel Borkmann rcu_read_lock(); 328381164413SDaniel Borkmann ca = tcp_ca_find_key(ca_key); 328481164413SDaniel Borkmann if (likely(ca && try_module_get(ca->owner))) { 328581164413SDaniel Borkmann module_put(icsk->icsk_ca_ops->owner); 328681164413SDaniel Borkmann icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst); 328781164413SDaniel Borkmann icsk->icsk_ca_ops = ca; 328881164413SDaniel Borkmann } 328981164413SDaniel Borkmann rcu_read_unlock(); 329081164413SDaniel Borkmann } 329181164413SDaniel Borkmann 329267edfef7SAndi Kleen /* Do all connect socket setups that can be done AF independent. */ 3293f7e56a76Sstephen hemminger static void tcp_connect_init(struct sock *sk) 32941da177e4SLinus Torvalds { 3295cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 32961da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 32971da177e4SLinus Torvalds __u8 rcv_wscale; 329813d3b1ebSLawrence Brakmo u32 rcv_wnd; 32991da177e4SLinus Torvalds 33001da177e4SLinus Torvalds /* We'll fix this up when we get a response from the other end. 33011da177e4SLinus Torvalds * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 33021da177e4SLinus Torvalds */ 33035d2ed052SEric Dumazet tp->tcp_header_len = sizeof(struct tcphdr); 33045d2ed052SEric Dumazet if (sock_net(sk)->ipv4.sysctl_tcp_timestamps) 33055d2ed052SEric Dumazet tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED; 33061da177e4SLinus Torvalds 3307cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 330800db4124SIan Morris if (tp->af_specific->md5_lookup(sk, sk)) 3309cfb6eeb4SYOSHIFUJI Hideaki tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 3310cfb6eeb4SYOSHIFUJI Hideaki #endif 3311cfb6eeb4SYOSHIFUJI Hideaki 33121da177e4SLinus Torvalds /* If user gave his TCP_MAXSEG, record it to clamp */ 33131da177e4SLinus Torvalds if (tp->rx_opt.user_mss) 33141da177e4SLinus Torvalds tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 33151da177e4SLinus Torvalds tp->max_window = 0; 33165d424d5aSJohn Heffner tcp_mtup_init(sk); 33171da177e4SLinus Torvalds tcp_sync_mss(sk, dst_mtu(dst)); 33181da177e4SLinus Torvalds 331981164413SDaniel Borkmann tcp_ca_dst_init(sk, dst); 332081164413SDaniel Borkmann 33211da177e4SLinus Torvalds if (!tp->window_clamp) 33221da177e4SLinus Torvalds tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 33233541f9e8SEric Dumazet tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 3324f5fff5dcSTom Quetchenbach 33251da177e4SLinus Torvalds tcp_initialize_rcv_mss(sk); 33261da177e4SLinus Torvalds 3327e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 3328e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 3329e88c64f0SHagen Paul Pfeifer (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) 3330e88c64f0SHagen Paul Pfeifer tp->window_clamp = tcp_full_space(sk); 3331e88c64f0SHagen Paul Pfeifer 333213d3b1ebSLawrence Brakmo rcv_wnd = tcp_rwnd_init_bpf(sk); 333313d3b1ebSLawrence Brakmo if (rcv_wnd == 0) 333413d3b1ebSLawrence Brakmo rcv_wnd = dst_metric(dst, RTAX_INITRWND); 333513d3b1ebSLawrence Brakmo 3336ceef9ab6SEric Dumazet tcp_select_initial_window(sk, tcp_full_space(sk), 33371da177e4SLinus Torvalds tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 33381da177e4SLinus Torvalds &tp->rcv_wnd, 33391da177e4SLinus Torvalds &tp->window_clamp, 33409bb37ef0SEric Dumazet sock_net(sk)->ipv4.sysctl_tcp_window_scaling, 334131d12926Slaurent chavey &rcv_wscale, 334213d3b1ebSLawrence Brakmo rcv_wnd); 33431da177e4SLinus Torvalds 33441da177e4SLinus Torvalds tp->rx_opt.rcv_wscale = rcv_wscale; 33451da177e4SLinus Torvalds tp->rcv_ssthresh = tp->rcv_wnd; 33461da177e4SLinus Torvalds 33471da177e4SLinus Torvalds sk->sk_err = 0; 33481da177e4SLinus Torvalds sock_reset_flag(sk, SOCK_DONE); 33491da177e4SLinus Torvalds tp->snd_wnd = 0; 3350ee7537b6SHantzis Fotis tcp_init_wl(tp, 0); 33517f582b24SEric Dumazet tcp_write_queue_purge(sk); 33521da177e4SLinus Torvalds tp->snd_una = tp->write_seq; 33531da177e4SLinus Torvalds tp->snd_sml = tp->write_seq; 335433f5f57eSIlpo Järvinen tp->snd_up = tp->write_seq; 3355370816aeSPavel Emelyanov tp->snd_nxt = tp->write_seq; 3356ee995283SPavel Emelyanov 3357ee995283SPavel Emelyanov if (likely(!tp->repair)) 33581da177e4SLinus Torvalds tp->rcv_nxt = 0; 3359c7781a6eSAndrew Vagin else 336070eabf0eSEric Dumazet tp->rcv_tstamp = tcp_jiffies32; 3361ee995283SPavel Emelyanov tp->rcv_wup = tp->rcv_nxt; 3362ee995283SPavel Emelyanov tp->copied_seq = tp->rcv_nxt; 33631da177e4SLinus Torvalds 33648550f328SLawrence Brakmo inet_csk(sk)->icsk_rto = tcp_timeout_init(sk); 3365463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_retransmits = 0; 33661da177e4SLinus Torvalds tcp_clear_retrans(tp); 33671da177e4SLinus Torvalds } 33681da177e4SLinus Torvalds 3369783237e8SYuchung Cheng static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) 3370783237e8SYuchung Cheng { 3371783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3372783237e8SYuchung Cheng struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 3373783237e8SYuchung Cheng 3374783237e8SYuchung Cheng tcb->end_seq += skb->len; 3375f4a775d1SEric Dumazet __skb_header_release(skb); 3376783237e8SYuchung Cheng sk->sk_wmem_queued += skb->truesize; 3377783237e8SYuchung Cheng sk_mem_charge(sk, skb->truesize); 3378783237e8SYuchung Cheng tp->write_seq = tcb->end_seq; 3379783237e8SYuchung Cheng tp->packets_out += tcp_skb_pcount(skb); 3380783237e8SYuchung Cheng } 3381783237e8SYuchung Cheng 3382783237e8SYuchung Cheng /* Build and send a SYN with data and (cached) Fast Open cookie. However, 3383783237e8SYuchung Cheng * queue a data-only packet after the regular SYN, such that regular SYNs 3384783237e8SYuchung Cheng * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges 3385783237e8SYuchung Cheng * only the SYN sequence, the data are retransmitted in the first ACK. 3386783237e8SYuchung Cheng * If cookie is not cached or other error occurs, falls back to send a 3387783237e8SYuchung Cheng * regular SYN with Fast Open cookie request option. 3388783237e8SYuchung Cheng */ 3389783237e8SYuchung Cheng static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) 3390783237e8SYuchung Cheng { 3391783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3392783237e8SYuchung Cheng struct tcp_fastopen_request *fo = tp->fastopen_req; 3393065263f4SWei Wang int space, err = 0; 3394355a901eSEric Dumazet struct sk_buff *syn_data; 3395783237e8SYuchung Cheng 339667da22d2SYuchung Cheng tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ 3397065263f4SWei Wang if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie)) 3398783237e8SYuchung Cheng goto fallback; 3399783237e8SYuchung Cheng 3400783237e8SYuchung Cheng /* MSS for SYN-data is based on cached MSS and bounded by PMTU and 3401783237e8SYuchung Cheng * user-MSS. Reserve maximum option space for middleboxes that add 3402783237e8SYuchung Cheng * private TCP options. The cost is reduced data space in SYN :( 3403783237e8SYuchung Cheng */ 34043541f9e8SEric Dumazet tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp); 34053541f9e8SEric Dumazet 34061b63edd6SYuchung Cheng space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) - 3407783237e8SYuchung Cheng MAX_TCP_OPTION_SPACE; 3408783237e8SYuchung Cheng 3409f5ddcbbbSEric Dumazet space = min_t(size_t, space, fo->size); 3410f5ddcbbbSEric Dumazet 3411f5ddcbbbSEric Dumazet /* limit to order-0 allocations */ 3412f5ddcbbbSEric Dumazet space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER)); 3413f5ddcbbbSEric Dumazet 3414eb934478SEric Dumazet syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false); 3415355a901eSEric Dumazet if (!syn_data) 3416783237e8SYuchung Cheng goto fallback; 3417355a901eSEric Dumazet syn_data->ip_summed = CHECKSUM_PARTIAL; 3418355a901eSEric Dumazet memcpy(syn_data->cb, syn->cb, sizeof(syn->cb)); 341907e100f9SEric Dumazet if (space) { 342007e100f9SEric Dumazet int copied = copy_from_iter(skb_put(syn_data, space), space, 342157be5bdaSAl Viro &fo->data->msg_iter); 342257be5bdaSAl Viro if (unlikely(!copied)) { 3423ba233b34SEric Dumazet tcp_skb_tsorted_anchor_cleanup(syn_data); 3424355a901eSEric Dumazet kfree_skb(syn_data); 3425783237e8SYuchung Cheng goto fallback; 3426783237e8SYuchung Cheng } 342757be5bdaSAl Viro if (copied != space) { 342857be5bdaSAl Viro skb_trim(syn_data, copied); 342957be5bdaSAl Viro space = copied; 343057be5bdaSAl Viro } 343107e100f9SEric Dumazet } 3432355a901eSEric Dumazet /* No more data pending in inet_wait_for_connect() */ 3433355a901eSEric Dumazet if (space == fo->size) 3434355a901eSEric Dumazet fo->data = NULL; 3435355a901eSEric Dumazet fo->copied = space; 3436783237e8SYuchung Cheng 3437355a901eSEric Dumazet tcp_connect_queue_skb(sk, syn_data); 34380f87230dSFrancis Yan if (syn_data->len) 34390f87230dSFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_BUSY); 3440355a901eSEric Dumazet 3441355a901eSEric Dumazet err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation); 3442355a901eSEric Dumazet 3443d3edd06eSEric Dumazet syn->skb_mstamp_ns = syn_data->skb_mstamp_ns; 3444355a901eSEric Dumazet 3445355a901eSEric Dumazet /* Now full SYN+DATA was cloned and sent (or not), 3446355a901eSEric Dumazet * remove the SYN from the original skb (syn_data) 3447355a901eSEric Dumazet * we keep in write queue in case of a retransmit, as we 3448355a901eSEric Dumazet * also have the SYN packet (with no data) in the same queue. 3449431a9124SEric Dumazet */ 3450355a901eSEric Dumazet TCP_SKB_CB(syn_data)->seq++; 3451355a901eSEric Dumazet TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH; 3452355a901eSEric Dumazet if (!err) { 345367da22d2SYuchung Cheng tp->syn_data = (fo->copied > 0); 345475c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data); 3455f19c29e3SYuchung Cheng NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT); 3456783237e8SYuchung Cheng goto done; 3457783237e8SYuchung Cheng } 3458783237e8SYuchung Cheng 345975c119afSEric Dumazet /* data was not sent, put it in write_queue */ 346075c119afSEric Dumazet __skb_queue_tail(&sk->sk_write_queue, syn_data); 3461b5b7db8dSEric Dumazet tp->packets_out -= tcp_skb_pcount(syn_data); 3462b5b7db8dSEric Dumazet 3463783237e8SYuchung Cheng fallback: 3464783237e8SYuchung Cheng /* Send a regular SYN with Fast Open cookie request option */ 3465783237e8SYuchung Cheng if (fo->cookie.len > 0) 3466783237e8SYuchung Cheng fo->cookie.len = 0; 3467783237e8SYuchung Cheng err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); 3468783237e8SYuchung Cheng if (err) 3469783237e8SYuchung Cheng tp->syn_fastopen = 0; 3470783237e8SYuchung Cheng done: 3471783237e8SYuchung Cheng fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ 3472783237e8SYuchung Cheng return err; 3473783237e8SYuchung Cheng } 3474783237e8SYuchung Cheng 347567edfef7SAndi Kleen /* Build a SYN and send it off. */ 34761da177e4SLinus Torvalds int tcp_connect(struct sock *sk) 34771da177e4SLinus Torvalds { 34781da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 34791da177e4SLinus Torvalds struct sk_buff *buff; 3480ee586811SEric Paris int err; 34811da177e4SLinus Torvalds 3482de525be2SLawrence Brakmo tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB, 0, NULL); 34838ba60924SEric Dumazet 34848ba60924SEric Dumazet if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 34858ba60924SEric Dumazet return -EHOSTUNREACH; /* Routing failure or similar. */ 34868ba60924SEric Dumazet 34871da177e4SLinus Torvalds tcp_connect_init(sk); 34881da177e4SLinus Torvalds 34892b916477SAndrey Vagin if (unlikely(tp->repair)) { 34902b916477SAndrey Vagin tcp_finish_connect(sk, NULL); 34912b916477SAndrey Vagin return 0; 34922b916477SAndrey Vagin } 34932b916477SAndrey Vagin 3494eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true); 3495355a901eSEric Dumazet if (unlikely(!buff)) 34961da177e4SLinus Torvalds return -ENOBUFS; 34971da177e4SLinus Torvalds 3498a3433f35SChangli Gao tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); 34999a568de4SEric Dumazet tcp_mstamp_refresh(tp); 35009a568de4SEric Dumazet tp->retrans_stamp = tcp_time_stamp(tp); 3501783237e8SYuchung Cheng tcp_connect_queue_skb(sk, buff); 3502735d3831SFlorian Westphal tcp_ecn_send_syn(sk, buff); 350375c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, buff); 35041da177e4SLinus Torvalds 3505783237e8SYuchung Cheng /* Send off SYN; include data in Fast Open. */ 3506783237e8SYuchung Cheng err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : 3507783237e8SYuchung Cheng tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); 3508ee586811SEric Paris if (err == -ECONNREFUSED) 3509ee586811SEric Paris return err; 3510bd37a088SWei Yongjun 3511bd37a088SWei Yongjun /* We change tp->snd_nxt after the tcp_transmit_skb() call 3512bd37a088SWei Yongjun * in order to make this packet get counted in tcpOutSegs. 3513bd37a088SWei Yongjun */ 3514bd37a088SWei Yongjun tp->snd_nxt = tp->write_seq; 3515bd37a088SWei Yongjun tp->pushed_seq = tp->write_seq; 3516b5b7db8dSEric Dumazet buff = tcp_send_head(sk); 3517b5b7db8dSEric Dumazet if (unlikely(buff)) { 3518b5b7db8dSEric Dumazet tp->snd_nxt = TCP_SKB_CB(buff)->seq; 3519b5b7db8dSEric Dumazet tp->pushed_seq = TCP_SKB_CB(buff)->seq; 3520b5b7db8dSEric Dumazet } 352181cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); 35221da177e4SLinus Torvalds 35231da177e4SLinus Torvalds /* Timer for repeating the SYN until an answer. */ 35243f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 35253f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 35261da177e4SLinus Torvalds return 0; 35271da177e4SLinus Torvalds } 35284bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_connect); 35291da177e4SLinus Torvalds 35301da177e4SLinus Torvalds /* Send out a delayed ack, the caller does the policy checking 35311da177e4SLinus Torvalds * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 35321da177e4SLinus Torvalds * for details. 35331da177e4SLinus Torvalds */ 35341da177e4SLinus Torvalds void tcp_send_delayed_ack(struct sock *sk) 35351da177e4SLinus Torvalds { 3536463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 3537463c84b9SArnaldo Carvalho de Melo int ato = icsk->icsk_ack.ato; 35381da177e4SLinus Torvalds unsigned long timeout; 35391da177e4SLinus Torvalds 35401da177e4SLinus Torvalds if (ato > TCP_DELACK_MIN) { 3541463c84b9SArnaldo Carvalho de Melo const struct tcp_sock *tp = tcp_sk(sk); 35421da177e4SLinus Torvalds int max_ato = HZ / 2; 35431da177e4SLinus Torvalds 3544056834d9SIlpo Järvinen if (icsk->icsk_ack.pingpong || 3545056834d9SIlpo Järvinen (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 35461da177e4SLinus Torvalds max_ato = TCP_DELACK_MAX; 35471da177e4SLinus Torvalds 35481da177e4SLinus Torvalds /* Slow path, intersegment interval is "high". */ 35491da177e4SLinus Torvalds 35501da177e4SLinus Torvalds /* If some rtt estimate is known, use it to bound delayed ack. 3551463c84b9SArnaldo Carvalho de Melo * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 35521da177e4SLinus Torvalds * directly. 35531da177e4SLinus Torvalds */ 3554740b0f18SEric Dumazet if (tp->srtt_us) { 3555740b0f18SEric Dumazet int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3), 3556740b0f18SEric Dumazet TCP_DELACK_MIN); 35571da177e4SLinus Torvalds 35581da177e4SLinus Torvalds if (rtt < max_ato) 35591da177e4SLinus Torvalds max_ato = rtt; 35601da177e4SLinus Torvalds } 35611da177e4SLinus Torvalds 35621da177e4SLinus Torvalds ato = min(ato, max_ato); 35631da177e4SLinus Torvalds } 35641da177e4SLinus Torvalds 35651da177e4SLinus Torvalds /* Stay within the limit we were given */ 35661da177e4SLinus Torvalds timeout = jiffies + ato; 35671da177e4SLinus Torvalds 35681da177e4SLinus Torvalds /* Use new timeout only if there wasn't a older one earlier. */ 3569463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 35701da177e4SLinus Torvalds /* If delack timer was blocked or is about to expire, 35711da177e4SLinus Torvalds * send ACK now. 35721da177e4SLinus Torvalds */ 3573463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.blocked || 3574463c84b9SArnaldo Carvalho de Melo time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 35751da177e4SLinus Torvalds tcp_send_ack(sk); 35761da177e4SLinus Torvalds return; 35771da177e4SLinus Torvalds } 35781da177e4SLinus Torvalds 3579463c84b9SArnaldo Carvalho de Melo if (!time_before(timeout, icsk->icsk_ack.timeout)) 3580463c84b9SArnaldo Carvalho de Melo timeout = icsk->icsk_ack.timeout; 35811da177e4SLinus Torvalds } 3582463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 3583463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.timeout = timeout; 3584463c84b9SArnaldo Carvalho de Melo sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 35851da177e4SLinus Torvalds } 35861da177e4SLinus Torvalds 35871da177e4SLinus Torvalds /* This routine sends an ack and also updates the window. */ 35882987babbSYuchung Cheng void __tcp_send_ack(struct sock *sk, u32 rcv_nxt) 35891da177e4SLinus Torvalds { 35901da177e4SLinus Torvalds struct sk_buff *buff; 35911da177e4SLinus Torvalds 3592058dc334SIlpo Järvinen /* If we have been reset, we may not send again. */ 3593058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3594058dc334SIlpo Järvinen return; 3595058dc334SIlpo Järvinen 35961da177e4SLinus Torvalds /* We are not putting this on the write queue, so 35971da177e4SLinus Torvalds * tcp_transmit_skb() will set the ownership to this 35981da177e4SLinus Torvalds * sock. 35991da177e4SLinus Torvalds */ 36007450aaf6SEric Dumazet buff = alloc_skb(MAX_TCP_HEADER, 36017450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 36027450aaf6SEric Dumazet if (unlikely(!buff)) { 3603463c84b9SArnaldo Carvalho de Melo inet_csk_schedule_ack(sk); 3604463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; 36053f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 36063f421baaSArnaldo Carvalho de Melo TCP_DELACK_MAX, TCP_RTO_MAX); 36071da177e4SLinus Torvalds return; 36081da177e4SLinus Torvalds } 36091da177e4SLinus Torvalds 36101da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 36111da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 3612a3433f35SChangli Gao tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); 36131da177e4SLinus Torvalds 361498781965SEric Dumazet /* We do not want pure acks influencing TCP Small Queues or fq/pacing 361598781965SEric Dumazet * too much. 361698781965SEric Dumazet * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784 361798781965SEric Dumazet */ 361898781965SEric Dumazet skb_set_tcp_pure_ack(buff); 361998781965SEric Dumazet 36201da177e4SLinus Torvalds /* Send it off, this clears delayed acks for us. */ 36212987babbSYuchung Cheng __tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0, rcv_nxt); 36221da177e4SLinus Torvalds } 362327cde44aSYuchung Cheng EXPORT_SYMBOL_GPL(__tcp_send_ack); 36242987babbSYuchung Cheng 36252987babbSYuchung Cheng void tcp_send_ack(struct sock *sk) 36262987babbSYuchung Cheng { 36272987babbSYuchung Cheng __tcp_send_ack(sk, tcp_sk(sk)->rcv_nxt); 36281da177e4SLinus Torvalds } 36291da177e4SLinus Torvalds 36301da177e4SLinus Torvalds /* This routine sends a packet with an out of date sequence 36311da177e4SLinus Torvalds * number. It assumes the other end will try to ack it. 36321da177e4SLinus Torvalds * 36331da177e4SLinus Torvalds * Question: what should we make while urgent mode? 36341da177e4SLinus Torvalds * 4.4BSD forces sending single byte of data. We cannot send 36351da177e4SLinus Torvalds * out of window data, because we have SND.NXT==SND.MAX... 36361da177e4SLinus Torvalds * 36371da177e4SLinus Torvalds * Current solution: to send TWO zero-length segments in urgent mode: 36381da177e4SLinus Torvalds * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 36391da177e4SLinus Torvalds * out-of-date with SND.UNA-1 to probe window. 36401da177e4SLinus Torvalds */ 3641e520af48SEric Dumazet static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib) 36421da177e4SLinus Torvalds { 36431da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 36441da177e4SLinus Torvalds struct sk_buff *skb; 36451da177e4SLinus Torvalds 36461da177e4SLinus Torvalds /* We don't queue it, tcp_transmit_skb() sets ownership. */ 36477450aaf6SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, 36487450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 364951456b29SIan Morris if (!skb) 36501da177e4SLinus Torvalds return -1; 36511da177e4SLinus Torvalds 36521da177e4SLinus Torvalds /* Reserve space for headers and set control bits. */ 36531da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 36541da177e4SLinus Torvalds /* Use a previous sequence. This should cause the other 36551da177e4SLinus Torvalds * end to send an ack. Don't queue or clone SKB, just 36561da177e4SLinus Torvalds * send it. 36571da177e4SLinus Torvalds */ 3658a3433f35SChangli Gao tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); 3659e2e8009fSRenato Westphal NET_INC_STATS(sock_net(sk), mib); 36607450aaf6SEric Dumazet return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0); 36611da177e4SLinus Torvalds } 36621da177e4SLinus Torvalds 3663385e2070SEric Dumazet /* Called from setsockopt( ... TCP_REPAIR ) */ 3664ee995283SPavel Emelyanov void tcp_send_window_probe(struct sock *sk) 3665ee995283SPavel Emelyanov { 3666ee995283SPavel Emelyanov if (sk->sk_state == TCP_ESTABLISHED) { 3667ee995283SPavel Emelyanov tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; 36689a568de4SEric Dumazet tcp_mstamp_refresh(tcp_sk(sk)); 3669e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE); 3670ee995283SPavel Emelyanov } 3671ee995283SPavel Emelyanov } 3672ee995283SPavel Emelyanov 367367edfef7SAndi Kleen /* Initiate keepalive or window probe from timer. */ 3674e520af48SEric Dumazet int tcp_write_wakeup(struct sock *sk, int mib) 36751da177e4SLinus Torvalds { 36761da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 36771da177e4SLinus Torvalds struct sk_buff *skb; 36781da177e4SLinus Torvalds 3679058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3680058dc334SIlpo Järvinen return -1; 3681058dc334SIlpo Järvinen 368200db4124SIan Morris skb = tcp_send_head(sk); 368300db4124SIan Morris if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { 36841da177e4SLinus Torvalds int err; 36850c54b85fSIlpo Järvinen unsigned int mss = tcp_current_mss(sk); 368690840defSIlpo Järvinen unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 36871da177e4SLinus Torvalds 36881da177e4SLinus Torvalds if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 36891da177e4SLinus Torvalds tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 36901da177e4SLinus Torvalds 36911da177e4SLinus Torvalds /* We are probing the opening of a window 36921da177e4SLinus Torvalds * but the window size is != 0 36931da177e4SLinus Torvalds * must have been a result SWS avoidance ( sender ) 36941da177e4SLinus Torvalds */ 36951da177e4SLinus Torvalds if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 36961da177e4SLinus Torvalds skb->len > mss) { 36971da177e4SLinus Torvalds seg_size = min(seg_size, mss); 36984de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 369975c119afSEric Dumazet if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE, 370075c119afSEric Dumazet skb, seg_size, mss, GFP_ATOMIC)) 37011da177e4SLinus Torvalds return -1; 37021da177e4SLinus Torvalds } else if (!tcp_skb_pcount(skb)) 37035bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss); 37041da177e4SLinus Torvalds 37054de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 3706dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 370766f5fe62SIlpo Järvinen if (!err) 370866f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 37091da177e4SLinus Torvalds return err; 37101da177e4SLinus Torvalds } else { 371133f5f57eSIlpo Järvinen if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) 3712e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 1, mib); 3713e520af48SEric Dumazet return tcp_xmit_probe_skb(sk, 0, mib); 37141da177e4SLinus Torvalds } 37151da177e4SLinus Torvalds } 37161da177e4SLinus Torvalds 37171da177e4SLinus Torvalds /* A window probe timeout has occurred. If window is not closed send 37181da177e4SLinus Torvalds * a partial packet else a zero probe. 37191da177e4SLinus Torvalds */ 37201da177e4SLinus Torvalds void tcp_send_probe0(struct sock *sk) 37211da177e4SLinus Torvalds { 3722463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 37231da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 3724c6214a97SNikolay Borisov struct net *net = sock_net(sk); 3725fcdd1cf4SEric Dumazet unsigned long probe_max; 37261da177e4SLinus Torvalds int err; 37271da177e4SLinus Torvalds 3728e520af48SEric Dumazet err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE); 37291da177e4SLinus Torvalds 373075c119afSEric Dumazet if (tp->packets_out || tcp_write_queue_empty(sk)) { 37311da177e4SLinus Torvalds /* Cancel probe timer, if it is not required. */ 37326687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 0; 3733463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff = 0; 37341da177e4SLinus Torvalds return; 37351da177e4SLinus Torvalds } 37361da177e4SLinus Torvalds 37371da177e4SLinus Torvalds if (err <= 0) { 3738c6214a97SNikolay Borisov if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2) 3739463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff++; 37406687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out++; 3741fcdd1cf4SEric Dumazet probe_max = TCP_RTO_MAX; 37421da177e4SLinus Torvalds } else { 37431da177e4SLinus Torvalds /* If packet was not sent due to local congestion, 37446687e988SArnaldo Carvalho de Melo * do not backoff and do not remember icsk_probes_out. 37451da177e4SLinus Torvalds * Let local senders to fight for local resources. 37461da177e4SLinus Torvalds * 37471da177e4SLinus Torvalds * Use accumulated backoff yet. 37481da177e4SLinus Torvalds */ 37496687e988SArnaldo Carvalho de Melo if (!icsk->icsk_probes_out) 37506687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 1; 3751fcdd1cf4SEric Dumazet probe_max = TCP_RESOURCE_PROBE_INTERVAL; 37521da177e4SLinus Torvalds } 3753fcdd1cf4SEric Dumazet inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 375421c8fe99SEric Dumazet tcp_probe0_when(sk, probe_max), 3755fcdd1cf4SEric Dumazet TCP_RTO_MAX); 37561da177e4SLinus Torvalds } 37575db92c99SOctavian Purdila 3758ea3bea3aSEric Dumazet int tcp_rtx_synack(const struct sock *sk, struct request_sock *req) 37595db92c99SOctavian Purdila { 37605db92c99SOctavian Purdila const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific; 37615db92c99SOctavian Purdila struct flowi fl; 37625db92c99SOctavian Purdila int res; 37635db92c99SOctavian Purdila 376458d607d3SEric Dumazet tcp_rsk(req)->txhash = net_tx_rndhash(); 3765b3d05147SEric Dumazet res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL); 37665db92c99SOctavian Purdila if (!res) { 376790bbcc60SEric Dumazet __TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); 376802a1d6e7SEric Dumazet __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 37697e32b443SYuchung Cheng if (unlikely(tcp_passive_fastopen(sk))) 37707e32b443SYuchung Cheng tcp_sk(sk)->total_retrans++; 3771cf34ce3dSSong Liu trace_tcp_retransmit_synack(sk, req); 37725db92c99SOctavian Purdila } 37735db92c99SOctavian Purdila return res; 37745db92c99SOctavian Purdila } 37755db92c99SOctavian Purdila EXPORT_SYMBOL(tcp_rtx_synack); 3776