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 18386de5921SEric Dumazet if (unlikely(tp->compressed_ack > TCP_FASTRETRANS_THRESH)) { 184200d95f4SEric Dumazet NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED, 18586de5921SEric Dumazet tp->compressed_ack - TCP_FASTRETRANS_THRESH); 18686de5921SEric Dumazet tp->compressed_ack = TCP_FASTRETRANS_THRESH; 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 978a7a25630SEric Dumazet static void tcp_update_skb_after_send(struct sock *sk, struct sk_buff *skb, 979a7a25630SEric Dumazet u64 prior_wstamp) 980e2080072SEric Dumazet { 981ab408b6dSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 982ab408b6dSEric Dumazet 983d3edd06eSEric Dumazet skb->skb_mstamp_ns = tp->tcp_wstamp_ns; 984ab408b6dSEric Dumazet if (sk->sk_pacing_status != SK_PACING_NONE) { 98576a9ebe8SEric Dumazet unsigned long rate = sk->sk_pacing_rate; 986ab408b6dSEric Dumazet 987ab408b6dSEric Dumazet /* Original sch_fq does not pace first 10 MSS 988ab408b6dSEric Dumazet * Note that tp->data_segs_out overflows after 2^32 packets, 989ab408b6dSEric Dumazet * this is a minor annoyance. 990ab408b6dSEric Dumazet */ 99176a9ebe8SEric Dumazet if (rate != ~0UL && rate && tp->data_segs_out >= 10) { 992a7a25630SEric Dumazet u64 len_ns = div64_ul((u64)skb->len * NSEC_PER_SEC, rate); 993a7a25630SEric Dumazet u64 credit = tp->tcp_wstamp_ns - prior_wstamp; 994a7a25630SEric Dumazet 995a7a25630SEric Dumazet /* take into account OS jitter */ 996a7a25630SEric Dumazet len_ns -= min_t(u64, len_ns / 2, credit); 997a7a25630SEric Dumazet tp->tcp_wstamp_ns += len_ns; 998ab408b6dSEric Dumazet } 999ab408b6dSEric Dumazet } 1000e2080072SEric Dumazet list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue); 1001e2080072SEric Dumazet } 1002e2080072SEric Dumazet 10031da177e4SLinus Torvalds /* This routine actually transmits TCP packets queued in by 10041da177e4SLinus Torvalds * tcp_do_sendmsg(). This is used by both the initial 10051da177e4SLinus Torvalds * transmission and possible later retransmissions. 10061da177e4SLinus Torvalds * All SKB's seen here are completely headerless. It is our 10071da177e4SLinus Torvalds * job to build the TCP header, and pass the packet down to 10081da177e4SLinus Torvalds * IP so it can do the same plus pass the packet off to the 10091da177e4SLinus Torvalds * device. 10101da177e4SLinus Torvalds * 10111da177e4SLinus Torvalds * We are working here with either a clone of the original 10121da177e4SLinus Torvalds * SKB, or a fresh unique copy made by the retransmit engine. 10131da177e4SLinus Torvalds */ 10142987babbSYuchung Cheng static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, 10152987babbSYuchung Cheng int clone_it, gfp_t gfp_mask, u32 rcv_nxt) 10161da177e4SLinus Torvalds { 10176687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 1018dfb4b9dcSDavid S. Miller struct inet_sock *inet; 1019dfb4b9dcSDavid S. Miller struct tcp_sock *tp; 1020dfb4b9dcSDavid S. Miller struct tcp_skb_cb *tcb; 102133ad798cSAdam Langley struct tcp_out_options opts; 102295c96174SEric Dumazet unsigned int tcp_options_size, tcp_header_size; 10238c72c65bSEric Dumazet struct sk_buff *oskb = NULL; 1024cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 10251da177e4SLinus Torvalds struct tcphdr *th; 1026a7a25630SEric Dumazet u64 prior_wstamp; 10271da177e4SLinus Torvalds int err; 10281da177e4SLinus Torvalds 1029dfb4b9dcSDavid S. Miller BUG_ON(!skb || !tcp_skb_pcount(skb)); 10306f094b9eSLawrence Brakmo tp = tcp_sk(sk); 1031dfb4b9dcSDavid S. Miller 1032ccdbb6e9SEric Dumazet if (clone_it) { 10336f094b9eSLawrence Brakmo TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq 10346f094b9eSLawrence Brakmo - tp->snd_una; 10358c72c65bSEric Dumazet oskb = skb; 1036e2080072SEric Dumazet 1037e2080072SEric Dumazet tcp_skb_tsorted_save(oskb) { 1038e2080072SEric Dumazet if (unlikely(skb_cloned(oskb))) 1039e2080072SEric Dumazet skb = pskb_copy(oskb, gfp_mask); 1040dfb4b9dcSDavid S. Miller else 1041e2080072SEric Dumazet skb = skb_clone(oskb, gfp_mask); 1042e2080072SEric Dumazet } tcp_skb_tsorted_restore(oskb); 1043e2080072SEric Dumazet 1044dfb4b9dcSDavid S. Miller if (unlikely(!skb)) 1045dfb4b9dcSDavid S. Miller return -ENOBUFS; 1046dfb4b9dcSDavid S. Miller } 10475f6188a8SEric Dumazet 1048a7a25630SEric Dumazet prior_wstamp = tp->tcp_wstamp_ns; 10495f6188a8SEric Dumazet tp->tcp_wstamp_ns = max(tp->tcp_wstamp_ns, tp->tcp_clock_cache); 10505f6188a8SEric Dumazet 1051d3edd06eSEric Dumazet skb->skb_mstamp_ns = tp->tcp_wstamp_ns; 1052dfb4b9dcSDavid S. Miller 1053dfb4b9dcSDavid S. Miller inet = inet_sk(sk); 1054dfb4b9dcSDavid S. Miller tcb = TCP_SKB_CB(skb); 105533ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 10561da177e4SLinus Torvalds 10574de075e0SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) 105833ad798cSAdam Langley tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); 105933ad798cSAdam Langley else 106033ad798cSAdam Langley tcp_options_size = tcp_established_options(sk, skb, &opts, 106133ad798cSAdam Langley &md5); 106233ad798cSAdam Langley tcp_header_size = tcp_options_size + sizeof(struct tcphdr); 10631da177e4SLinus Torvalds 1064547669d4SEric Dumazet /* if no packet is in qdisc/device queue, then allow XPS to select 1065b2532eb9SEric Dumazet * another queue. We can be called from tcp_tsq_handler() 106673a6bab5SEric Dumazet * which holds one reference to sk. 1067b2532eb9SEric Dumazet * 1068b2532eb9SEric Dumazet * TODO: Ideally, in-flight pure ACK packets should not matter here. 1069b2532eb9SEric Dumazet * One way to get this would be to set skb->truesize = 2 on them. 1070547669d4SEric Dumazet */ 1071b2532eb9SEric Dumazet skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1); 10721da177e4SLinus Torvalds 107338ab52e8SEric Dumazet /* If we had to use memory reserve to allocate this skb, 107438ab52e8SEric Dumazet * this might cause drops if packet is looped back : 107538ab52e8SEric Dumazet * Other socket might not have SOCK_MEMALLOC. 107638ab52e8SEric Dumazet * Packets not looped back do not care about pfmemalloc. 107738ab52e8SEric Dumazet */ 107838ab52e8SEric Dumazet skb->pfmemalloc = 0; 107938ab52e8SEric Dumazet 1080aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 1081aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 108246d3ceabSEric Dumazet 108346d3ceabSEric Dumazet skb_orphan(skb); 108446d3ceabSEric Dumazet skb->sk = sk; 10851d2077acSEric Dumazet skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree; 1086b73c3d0eSTom Herbert skb_set_hash_from_sk(skb, sk); 108714afee4bSReshetova, Elena refcount_add(skb->truesize, &sk->sk_wmem_alloc); 10881da177e4SLinus Torvalds 1089c3a2e837SJulian Anastasov skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm); 1090c3a2e837SJulian Anastasov 10911da177e4SLinus Torvalds /* Build TCP header and checksum it. */ 1092ea1627c2SEric Dumazet th = (struct tcphdr *)skb->data; 1093c720c7e8SEric Dumazet th->source = inet->inet_sport; 1094c720c7e8SEric Dumazet th->dest = inet->inet_dport; 10951da177e4SLinus Torvalds th->seq = htonl(tcb->seq); 10962987babbSYuchung Cheng th->ack_seq = htonl(rcv_nxt); 1097df7a3b07SAl Viro *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | 10984de075e0SEric Dumazet tcb->tcp_flags); 1099dfb4b9dcSDavid S. Miller 11001da177e4SLinus Torvalds th->check = 0; 11011da177e4SLinus Torvalds th->urg_ptr = 0; 11021da177e4SLinus Torvalds 110333f5f57eSIlpo Järvinen /* The urg_mode check is necessary during a below snd_una win probe */ 11047691367dSHerbert Xu if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { 11057691367dSHerbert Xu if (before(tp->snd_up, tcb->seq + 0x10000)) { 11061da177e4SLinus Torvalds th->urg_ptr = htons(tp->snd_up - tcb->seq); 11071da177e4SLinus Torvalds th->urg = 1; 11087691367dSHerbert Xu } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { 11090eae88f3SEric Dumazet th->urg_ptr = htons(0xFFFF); 11107691367dSHerbert Xu th->urg = 1; 11117691367dSHerbert Xu } 11121da177e4SLinus Torvalds } 11131da177e4SLinus Torvalds 1114bd0388aeSWilliam Allen Simpson tcp_options_write((__be32 *)(th + 1), tp, &opts); 111551466a75SEric Dumazet skb_shinfo(skb)->gso_type = sk->sk_gso_type; 1116ea1627c2SEric Dumazet if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) { 1117ea1627c2SEric Dumazet th->window = htons(tcp_select_window(sk)); 1118ea1627c2SEric Dumazet tcp_ecn_send(sk, skb, th, tcp_header_size); 1119ea1627c2SEric Dumazet } else { 1120ea1627c2SEric Dumazet /* RFC1323: The window in SYN & SYN/ACK segments 1121ea1627c2SEric Dumazet * is never scaled. 1122ea1627c2SEric Dumazet */ 1123ea1627c2SEric Dumazet th->window = htons(min(tp->rcv_wnd, 65535U)); 1124ea1627c2SEric Dumazet } 1125cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 1126cfb6eeb4SYOSHIFUJI Hideaki /* Calculate the MD5 hash, as we have all we need now */ 1127cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 1128a465419bSEric Dumazet sk_nocaps_add(sk, NETIF_F_GSO_MASK); 1129bd0388aeSWilliam Allen Simpson tp->af_specific->calc_md5_hash(opts.hash_location, 113039f8e58eSEric Dumazet md5, sk, skb); 1131cfb6eeb4SYOSHIFUJI Hideaki } 1132cfb6eeb4SYOSHIFUJI Hideaki #endif 1133cfb6eeb4SYOSHIFUJI Hideaki 1134bb296246SHerbert Xu icsk->icsk_af_ops->send_check(sk, skb); 11351da177e4SLinus Torvalds 11364de075e0SEric Dumazet if (likely(tcb->tcp_flags & TCPHDR_ACK)) 113727cde44aSYuchung Cheng tcp_event_ack_sent(sk, tcp_skb_pcount(skb), rcv_nxt); 11381da177e4SLinus Torvalds 1139a44d6eacSMartin KaFai Lau if (skb->len != tcp_header_size) { 1140cf533ea5SEric Dumazet tcp_event_data_sent(tp, sk); 1141a44d6eacSMartin KaFai Lau tp->data_segs_out += tcp_skb_pcount(skb); 1142ba113c3aSWei Wang tp->bytes_sent += skb->len - tcp_header_size; 1143a44d6eacSMartin KaFai Lau } 11441da177e4SLinus Torvalds 1145bd37a088SWei Yongjun if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) 1146aa2ea058STom Herbert TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, 1147aa2ea058STom Herbert tcp_skb_pcount(skb)); 11481da177e4SLinus Torvalds 11492efd055cSMarcelo Ricardo Leitner tp->segs_out += tcp_skb_pcount(skb); 1150f69ad292SEric Dumazet /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */ 1151cd7d8498SEric Dumazet skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb); 1152f69ad292SEric Dumazet skb_shinfo(skb)->gso_size = tcp_skb_mss(skb); 1153cd7d8498SEric Dumazet 1154d3edd06eSEric Dumazet /* Leave earliest departure time in skb->tstamp (skb->skb_mstamp_ns) */ 1155971f10ecSEric Dumazet 1156971f10ecSEric Dumazet /* Cleanup our debris for IP stacks */ 1157971f10ecSEric Dumazet memset(skb->cb, 0, max(sizeof(struct inet_skb_parm), 1158971f10ecSEric Dumazet sizeof(struct inet6_skb_parm))); 1159971f10ecSEric Dumazet 1160b0270e91SEric Dumazet err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl); 11617faee5c0SEric Dumazet 11628c72c65bSEric Dumazet if (unlikely(err > 0)) { 11635ee2c941SChristoph Paasch tcp_enter_cwr(sk); 11648c72c65bSEric Dumazet err = net_xmit_eval(err); 11658c72c65bSEric Dumazet } 1166fc225799SEric Dumazet if (!err && oskb) { 1167a7a25630SEric Dumazet tcp_update_skb_after_send(sk, oskb, prior_wstamp); 1168fc225799SEric Dumazet tcp_rate_skb_sent(sk, oskb); 1169fc225799SEric Dumazet } 11708c72c65bSEric Dumazet return err; 11711da177e4SLinus Torvalds } 11721da177e4SLinus Torvalds 11732987babbSYuchung Cheng static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, 11742987babbSYuchung Cheng gfp_t gfp_mask) 11752987babbSYuchung Cheng { 11762987babbSYuchung Cheng return __tcp_transmit_skb(sk, skb, clone_it, gfp_mask, 11772987babbSYuchung Cheng tcp_sk(sk)->rcv_nxt); 11782987babbSYuchung Cheng } 11792987babbSYuchung Cheng 118067edfef7SAndi Kleen /* This routine just queues the buffer for sending. 11811da177e4SLinus Torvalds * 11821da177e4SLinus Torvalds * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, 11831da177e4SLinus Torvalds * otherwise socket can stall. 11841da177e4SLinus Torvalds */ 11851da177e4SLinus Torvalds static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) 11861da177e4SLinus Torvalds { 11871da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 11881da177e4SLinus Torvalds 11891da177e4SLinus Torvalds /* Advance write_seq and place onto the write_queue. */ 11901da177e4SLinus Torvalds tp->write_seq = TCP_SKB_CB(skb)->end_seq; 1191f4a775d1SEric Dumazet __skb_header_release(skb); 1192fe067e8aSDavid S. Miller tcp_add_write_queue_tail(sk, skb); 11933ab224beSHideo Aoki sk->sk_wmem_queued += skb->truesize; 11943ab224beSHideo Aoki sk_mem_charge(sk, skb->truesize); 11951da177e4SLinus Torvalds } 11961da177e4SLinus Torvalds 119767edfef7SAndi Kleen /* Initialize TSO segments for a packet. */ 11985bbb432cSEric Dumazet static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now) 1199f6302d1dSDavid S. Miller { 12004a64fd6cSEric Dumazet if (skb->len <= mss_now) { 1201f6302d1dSDavid S. Miller /* Avoid the costly divide in the normal 1202f6302d1dSDavid S. Miller * non-TSO case. 1203f6302d1dSDavid S. Miller */ 1204cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, 1); 1205f69ad292SEric Dumazet TCP_SKB_CB(skb)->tcp_gso_size = 0; 1206f6302d1dSDavid S. Miller } else { 1207cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now)); 1208f69ad292SEric Dumazet TCP_SKB_CB(skb)->tcp_gso_size = mss_now; 12091da177e4SLinus Torvalds } 12101da177e4SLinus Torvalds } 12111da177e4SLinus Torvalds 1212797108d1SIlpo Järvinen /* Pcount in the middle of the write queue got changed, we need to do various 1213797108d1SIlpo Järvinen * tweaks to fix counters 1214797108d1SIlpo Järvinen */ 1215cf533ea5SEric Dumazet static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) 1216797108d1SIlpo Järvinen { 1217797108d1SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1218797108d1SIlpo Järvinen 1219797108d1SIlpo Järvinen tp->packets_out -= decr; 1220797108d1SIlpo Järvinen 1221797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 1222797108d1SIlpo Järvinen tp->sacked_out -= decr; 1223797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 1224797108d1SIlpo Järvinen tp->retrans_out -= decr; 1225797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 1226797108d1SIlpo Järvinen tp->lost_out -= decr; 1227797108d1SIlpo Järvinen 1228797108d1SIlpo Järvinen /* Reno case is special. Sigh... */ 1229797108d1SIlpo Järvinen if (tcp_is_reno(tp) && decr > 0) 1230797108d1SIlpo Järvinen tp->sacked_out -= min_t(u32, tp->sacked_out, decr); 1231797108d1SIlpo Järvinen 1232797108d1SIlpo Järvinen if (tp->lost_skb_hint && 1233797108d1SIlpo Järvinen before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && 1234713bafeaSYuchung Cheng (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) 1235797108d1SIlpo Järvinen tp->lost_cnt_hint -= decr; 1236797108d1SIlpo Järvinen 1237797108d1SIlpo Järvinen tcp_verify_left_out(tp); 1238797108d1SIlpo Järvinen } 1239797108d1SIlpo Järvinen 12400a2cf20cSSoheil Hassas Yeganeh static bool tcp_has_tx_tstamp(const struct sk_buff *skb) 12410a2cf20cSSoheil Hassas Yeganeh { 12420a2cf20cSSoheil Hassas Yeganeh return TCP_SKB_CB(skb)->txstamp_ack || 12430a2cf20cSSoheil Hassas Yeganeh (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP); 12440a2cf20cSSoheil Hassas Yeganeh } 12450a2cf20cSSoheil Hassas Yeganeh 1246490cc7d0SWillem de Bruijn static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2) 1247490cc7d0SWillem de Bruijn { 1248490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo = skb_shinfo(skb); 1249490cc7d0SWillem de Bruijn 12500a2cf20cSSoheil Hassas Yeganeh if (unlikely(tcp_has_tx_tstamp(skb)) && 1251490cc7d0SWillem de Bruijn !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) { 1252490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo2 = skb_shinfo(skb2); 1253490cc7d0SWillem de Bruijn u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP; 1254490cc7d0SWillem de Bruijn 1255490cc7d0SWillem de Bruijn shinfo->tx_flags &= ~tsflags; 1256490cc7d0SWillem de Bruijn shinfo2->tx_flags |= tsflags; 1257490cc7d0SWillem de Bruijn swap(shinfo->tskey, shinfo2->tskey); 1258b51e13faSMartin KaFai Lau TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack; 1259b51e13faSMartin KaFai Lau TCP_SKB_CB(skb)->txstamp_ack = 0; 1260490cc7d0SWillem de Bruijn } 1261490cc7d0SWillem de Bruijn } 1262490cc7d0SWillem de Bruijn 1263a166140eSMartin KaFai Lau static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2) 1264a166140eSMartin KaFai Lau { 1265a166140eSMartin KaFai Lau TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor; 1266a166140eSMartin KaFai Lau TCP_SKB_CB(skb)->eor = 0; 1267a166140eSMartin KaFai Lau } 1268a166140eSMartin KaFai Lau 126975c119afSEric Dumazet /* Insert buff after skb on the write or rtx queue of sk. */ 127075c119afSEric Dumazet static void tcp_insert_write_queue_after(struct sk_buff *skb, 127175c119afSEric Dumazet struct sk_buff *buff, 127275c119afSEric Dumazet struct sock *sk, 127375c119afSEric Dumazet enum tcp_queue tcp_queue) 127475c119afSEric Dumazet { 127575c119afSEric Dumazet if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE) 127675c119afSEric Dumazet __skb_queue_after(&sk->sk_write_queue, skb, buff); 127775c119afSEric Dumazet else 127875c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, buff); 127975c119afSEric Dumazet } 128075c119afSEric Dumazet 12811da177e4SLinus Torvalds /* Function to create two new TCP segments. Shrinks the given segment 12821da177e4SLinus Torvalds * to the specified size and appends a new segment with the rest of the 12831da177e4SLinus Torvalds * packet to the list. This won't be called frequently, I hope. 12841da177e4SLinus Torvalds * Remember, these are still headerless SKBs at this point. 12851da177e4SLinus Torvalds */ 128675c119afSEric Dumazet int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue, 128775c119afSEric Dumazet struct sk_buff *skb, u32 len, 12886cc55e09SOctavian Purdila unsigned int mss_now, gfp_t gfp) 12891da177e4SLinus Torvalds { 12901da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 12911da177e4SLinus Torvalds struct sk_buff *buff; 12926475be16SDavid S. Miller int nsize, old_factor; 1293b60b49eaSHerbert Xu int nlen; 12949ce01461SIlpo Järvinen u8 flags; 12951da177e4SLinus Torvalds 12962fceec13SIlpo Järvinen if (WARN_ON(len > skb->len)) 12972fceec13SIlpo Järvinen return -EINVAL; 12986a438bbeSStephen Hemminger 12991da177e4SLinus Torvalds nsize = skb_headlen(skb) - len; 13001da177e4SLinus Torvalds if (nsize < 0) 13011da177e4SLinus Torvalds nsize = 0; 13021da177e4SLinus Torvalds 13036cc55e09SOctavian Purdila if (skb_unclone(skb, gfp)) 13041da177e4SLinus Torvalds return -ENOMEM; 13051da177e4SLinus Torvalds 13061da177e4SLinus Torvalds /* Get a new skb... force flag on. */ 1307eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, nsize, gfp, true); 130851456b29SIan Morris if (!buff) 13091da177e4SLinus Torvalds return -ENOMEM; /* We'll just try again later. */ 1310ef5cb973SHerbert Xu 13113ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 13123ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1313b60b49eaSHerbert Xu nlen = skb->len - len - nsize; 1314b60b49eaSHerbert Xu buff->truesize += nlen; 1315b60b49eaSHerbert Xu skb->truesize -= nlen; 13161da177e4SLinus Torvalds 13171da177e4SLinus Torvalds /* Correct the sequence numbers. */ 13181da177e4SLinus Torvalds TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 13191da177e4SLinus Torvalds TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 13201da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 13211da177e4SLinus Torvalds 13221da177e4SLinus Torvalds /* PSH and FIN should only be set in the second packet. */ 13234de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 13244de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 13254de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1326e14c3cafSHerbert Xu TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; 1327a166140eSMartin KaFai Lau tcp_skb_fragment_eor(skb, buff); 13281da177e4SLinus Torvalds 13291da177e4SLinus Torvalds skb_split(skb, buff, len); 13301da177e4SLinus Torvalds 133198be9b12SEric Dumazet buff->ip_summed = CHECKSUM_PARTIAL; 13321da177e4SLinus Torvalds 1333a61bbcf2SPatrick McHardy buff->tstamp = skb->tstamp; 1334490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 13351da177e4SLinus Torvalds 13366475be16SDavid S. Miller old_factor = tcp_skb_pcount(skb); 13376475be16SDavid S. Miller 13381da177e4SLinus Torvalds /* Fix up tso_factor for both original and new SKB. */ 13395bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 13405bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 13411da177e4SLinus Torvalds 1342b9f64820SYuchung Cheng /* Update delivered info for the new segment */ 1343b9f64820SYuchung Cheng TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx; 1344b9f64820SYuchung Cheng 13456475be16SDavid S. Miller /* If this packet has been sent out already, we must 13466475be16SDavid S. Miller * adjust the various packet counters. 13476475be16SDavid S. Miller */ 1348cf0b450cSHerbert Xu if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 13496475be16SDavid S. Miller int diff = old_factor - tcp_skb_pcount(skb) - 13506475be16SDavid S. Miller tcp_skb_pcount(buff); 13511da177e4SLinus Torvalds 1352797108d1SIlpo Järvinen if (diff) 1353797108d1SIlpo Järvinen tcp_adjust_pcount(sk, skb, diff); 13541da177e4SLinus Torvalds } 13551da177e4SLinus Torvalds 13561da177e4SLinus Torvalds /* Link BUFF into the send queue. */ 1357f4a775d1SEric Dumazet __skb_header_release(buff); 135875c119afSEric Dumazet tcp_insert_write_queue_after(skb, buff, sk, tcp_queue); 1359f67971e6SEric Dumazet if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE) 1360e2080072SEric Dumazet list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor); 13611da177e4SLinus Torvalds 13621da177e4SLinus Torvalds return 0; 13631da177e4SLinus Torvalds } 13641da177e4SLinus Torvalds 1365f4d01666SEric Dumazet /* This is similar to __pskb_pull_tail(). The difference is that pulled 1366f4d01666SEric Dumazet * data is not copied, but immediately discarded. 13671da177e4SLinus Torvalds */ 13687162fb24SEric Dumazet static int __pskb_trim_head(struct sk_buff *skb, int len) 13691da177e4SLinus Torvalds { 13707b7fc97aSEric Dumazet struct skb_shared_info *shinfo; 13711da177e4SLinus Torvalds int i, k, eat; 13721da177e4SLinus Torvalds 13734fa48bf3SEric Dumazet eat = min_t(int, len, skb_headlen(skb)); 13744fa48bf3SEric Dumazet if (eat) { 13754fa48bf3SEric Dumazet __skb_pull(skb, eat); 13764fa48bf3SEric Dumazet len -= eat; 13774fa48bf3SEric Dumazet if (!len) 13787162fb24SEric Dumazet return 0; 13794fa48bf3SEric Dumazet } 13801da177e4SLinus Torvalds eat = len; 13811da177e4SLinus Torvalds k = 0; 13827b7fc97aSEric Dumazet shinfo = skb_shinfo(skb); 13837b7fc97aSEric Dumazet for (i = 0; i < shinfo->nr_frags; i++) { 13847b7fc97aSEric Dumazet int size = skb_frag_size(&shinfo->frags[i]); 13859e903e08SEric Dumazet 13869e903e08SEric Dumazet if (size <= eat) { 1387aff65da0SIan Campbell skb_frag_unref(skb, i); 13889e903e08SEric Dumazet eat -= size; 13891da177e4SLinus Torvalds } else { 13907b7fc97aSEric Dumazet shinfo->frags[k] = shinfo->frags[i]; 13911da177e4SLinus Torvalds if (eat) { 13927b7fc97aSEric Dumazet shinfo->frags[k].page_offset += eat; 13937b7fc97aSEric Dumazet skb_frag_size_sub(&shinfo->frags[k], eat); 13941da177e4SLinus Torvalds eat = 0; 13951da177e4SLinus Torvalds } 13961da177e4SLinus Torvalds k++; 13971da177e4SLinus Torvalds } 13981da177e4SLinus Torvalds } 13997b7fc97aSEric Dumazet shinfo->nr_frags = k; 14001da177e4SLinus Torvalds 14011da177e4SLinus Torvalds skb->data_len -= len; 14021da177e4SLinus Torvalds skb->len = skb->data_len; 14037162fb24SEric Dumazet return len; 14041da177e4SLinus Torvalds } 14051da177e4SLinus Torvalds 140667edfef7SAndi Kleen /* Remove acked data from a packet in the transmit queue. */ 14071da177e4SLinus Torvalds int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 14081da177e4SLinus Torvalds { 14097162fb24SEric Dumazet u32 delta_truesize; 14107162fb24SEric Dumazet 141114bbd6a5SPravin B Shelar if (skb_unclone(skb, GFP_ATOMIC)) 14121da177e4SLinus Torvalds return -ENOMEM; 14131da177e4SLinus Torvalds 14147162fb24SEric Dumazet delta_truesize = __pskb_trim_head(skb, len); 14151da177e4SLinus Torvalds 14161da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq += len; 141784fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 14181da177e4SLinus Torvalds 14197162fb24SEric Dumazet if (delta_truesize) { 14207162fb24SEric Dumazet skb->truesize -= delta_truesize; 14217162fb24SEric Dumazet sk->sk_wmem_queued -= delta_truesize; 14227162fb24SEric Dumazet sk_mem_uncharge(sk, delta_truesize); 14231da177e4SLinus Torvalds sock_set_flag(sk, SOCK_QUEUE_SHRUNK); 14247162fb24SEric Dumazet } 14251da177e4SLinus Torvalds 14265b35e1e6SNeal Cardwell /* Any change of skb->len requires recalculation of tso factor. */ 14271da177e4SLinus Torvalds if (tcp_skb_pcount(skb) > 1) 14285bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb)); 14291da177e4SLinus Torvalds 14301da177e4SLinus Torvalds return 0; 14311da177e4SLinus Torvalds } 14321da177e4SLinus Torvalds 14331b63edd6SYuchung Cheng /* Calculate MSS not accounting any TCP options. */ 14341b63edd6SYuchung Cheng static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu) 14355d424d5aSJohn Heffner { 1436cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1437cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 14385d424d5aSJohn Heffner int mss_now; 14395d424d5aSJohn Heffner 14405d424d5aSJohn Heffner /* Calculate base mss without TCP options: 14415d424d5aSJohn Heffner It is MMS_S - sizeof(tcphdr) of rfc1122 14425d424d5aSJohn Heffner */ 14435d424d5aSJohn Heffner mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 14445d424d5aSJohn Heffner 144567469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 144667469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 144767469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 144867469601SEric Dumazet 144967469601SEric Dumazet if (dst && dst_allfrag(dst)) 145067469601SEric Dumazet mss_now -= icsk->icsk_af_ops->net_frag_header_len; 145167469601SEric Dumazet } 145267469601SEric Dumazet 14535d424d5aSJohn Heffner /* Clamp it (mss_clamp does not include tcp options) */ 14545d424d5aSJohn Heffner if (mss_now > tp->rx_opt.mss_clamp) 14555d424d5aSJohn Heffner mss_now = tp->rx_opt.mss_clamp; 14565d424d5aSJohn Heffner 14575d424d5aSJohn Heffner /* Now subtract optional transport overhead */ 14585d424d5aSJohn Heffner mss_now -= icsk->icsk_ext_hdr_len; 14595d424d5aSJohn Heffner 14605d424d5aSJohn Heffner /* Then reserve room for full set of TCP options and 8 bytes of data */ 14615d424d5aSJohn Heffner if (mss_now < 48) 14625d424d5aSJohn Heffner mss_now = 48; 14635d424d5aSJohn Heffner return mss_now; 14645d424d5aSJohn Heffner } 14655d424d5aSJohn Heffner 14661b63edd6SYuchung Cheng /* Calculate MSS. Not accounting for SACKs here. */ 14671b63edd6SYuchung Cheng int tcp_mtu_to_mss(struct sock *sk, int pmtu) 14681b63edd6SYuchung Cheng { 14691b63edd6SYuchung Cheng /* Subtract TCP options size, not including SACKs */ 14701b63edd6SYuchung Cheng return __tcp_mtu_to_mss(sk, pmtu) - 14711b63edd6SYuchung Cheng (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr)); 14721b63edd6SYuchung Cheng } 14731b63edd6SYuchung Cheng 14745d424d5aSJohn Heffner /* Inverse of above */ 147567469601SEric Dumazet int tcp_mss_to_mtu(struct sock *sk, int mss) 14765d424d5aSJohn Heffner { 1477cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1478cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 14795d424d5aSJohn Heffner int mtu; 14805d424d5aSJohn Heffner 14815d424d5aSJohn Heffner mtu = mss + 14825d424d5aSJohn Heffner tp->tcp_header_len + 14835d424d5aSJohn Heffner icsk->icsk_ext_hdr_len + 14845d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 14855d424d5aSJohn Heffner 148667469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 148767469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 148867469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 148967469601SEric Dumazet 149067469601SEric Dumazet if (dst && dst_allfrag(dst)) 149167469601SEric Dumazet mtu += icsk->icsk_af_ops->net_frag_header_len; 149267469601SEric Dumazet } 14935d424d5aSJohn Heffner return mtu; 14945d424d5aSJohn Heffner } 1495556c6b46SNeal Cardwell EXPORT_SYMBOL(tcp_mss_to_mtu); 14965d424d5aSJohn Heffner 149767edfef7SAndi Kleen /* MTU probing init per socket */ 14985d424d5aSJohn Heffner void tcp_mtup_init(struct sock *sk) 14995d424d5aSJohn Heffner { 15005d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 15015d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 1502b0f9ca53SFan Du struct net *net = sock_net(sk); 15035d424d5aSJohn Heffner 1504b0f9ca53SFan Du icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1; 15055d424d5aSJohn Heffner icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 15065d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 1507b0f9ca53SFan Du icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss); 15085d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 150905cbc0dbSFan Du if (icsk->icsk_mtup.enabled) 1510c74df29aSEric Dumazet icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; 15115d424d5aSJohn Heffner } 15124bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_mtup_init); 15135d424d5aSJohn Heffner 15141da177e4SLinus Torvalds /* This function synchronize snd mss to current pmtu/exthdr set. 15151da177e4SLinus Torvalds 15161da177e4SLinus Torvalds tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 15171da177e4SLinus Torvalds for TCP options, but includes only bare TCP header. 15181da177e4SLinus Torvalds 15191da177e4SLinus Torvalds tp->rx_opt.mss_clamp is mss negotiated at connection setup. 1520caa20d9aSStephen Hemminger It is minimum of user_mss and mss received with SYN. 15211da177e4SLinus Torvalds It also does not include TCP options. 15221da177e4SLinus Torvalds 1523d83d8461SArnaldo Carvalho de Melo inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. 15241da177e4SLinus Torvalds 15251da177e4SLinus Torvalds tp->mss_cache is current effective sending mss, including 15261da177e4SLinus Torvalds all tcp options except for SACKs. It is evaluated, 15271da177e4SLinus Torvalds taking into account current pmtu, but never exceeds 15281da177e4SLinus Torvalds tp->rx_opt.mss_clamp. 15291da177e4SLinus Torvalds 15301da177e4SLinus Torvalds NOTE1. rfc1122 clearly states that advertised MSS 15311da177e4SLinus Torvalds DOES NOT include either tcp or ip options. 15321da177e4SLinus Torvalds 1533d83d8461SArnaldo Carvalho de Melo NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache 1534d83d8461SArnaldo Carvalho de Melo are READ ONLY outside this function. --ANK (980731) 15351da177e4SLinus Torvalds */ 15361da177e4SLinus Torvalds unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 15371da177e4SLinus Torvalds { 15381da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1539d83d8461SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 15405d424d5aSJohn Heffner int mss_now; 15411da177e4SLinus Torvalds 15425d424d5aSJohn Heffner if (icsk->icsk_mtup.search_high > pmtu) 15435d424d5aSJohn Heffner icsk->icsk_mtup.search_high = pmtu; 15441da177e4SLinus Torvalds 15455d424d5aSJohn Heffner mss_now = tcp_mtu_to_mss(sk, pmtu); 1546409d22b4SIlpo Järvinen mss_now = tcp_bound_to_half_wnd(tp, mss_now); 15471da177e4SLinus Torvalds 15481da177e4SLinus Torvalds /* And store cached results */ 1549d83d8461SArnaldo Carvalho de Melo icsk->icsk_pmtu_cookie = pmtu; 15505d424d5aSJohn Heffner if (icsk->icsk_mtup.enabled) 15515d424d5aSJohn Heffner mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); 1552c1b4a7e6SDavid S. Miller tp->mss_cache = mss_now; 15531da177e4SLinus Torvalds 15541da177e4SLinus Torvalds return mss_now; 15551da177e4SLinus Torvalds } 15564bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_sync_mss); 15571da177e4SLinus Torvalds 15581da177e4SLinus Torvalds /* Compute the current effective MSS, taking SACKs and IP options, 15591da177e4SLinus Torvalds * and even PMTU discovery events into account. 15601da177e4SLinus Torvalds */ 15610c54b85fSIlpo Järvinen unsigned int tcp_current_mss(struct sock *sk) 15621da177e4SLinus Torvalds { 1563cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1564cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1565c1b4a7e6SDavid S. Miller u32 mss_now; 156695c96174SEric Dumazet unsigned int header_len; 156733ad798cSAdam Langley struct tcp_out_options opts; 156833ad798cSAdam Langley struct tcp_md5sig_key *md5; 15691da177e4SLinus Torvalds 1570c1b4a7e6SDavid S. Miller mss_now = tp->mss_cache; 1571c1b4a7e6SDavid S. Miller 15721da177e4SLinus Torvalds if (dst) { 15731da177e4SLinus Torvalds u32 mtu = dst_mtu(dst); 1574d83d8461SArnaldo Carvalho de Melo if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 15751da177e4SLinus Torvalds mss_now = tcp_sync_mss(sk, mtu); 15761da177e4SLinus Torvalds } 15771da177e4SLinus Torvalds 157833ad798cSAdam Langley header_len = tcp_established_options(sk, NULL, &opts, &md5) + 157933ad798cSAdam Langley sizeof(struct tcphdr); 158033ad798cSAdam Langley /* The mss_cache is sized based on tp->tcp_header_len, which assumes 158133ad798cSAdam Langley * some common options. If this is an odd packet (because we have SACK 158233ad798cSAdam Langley * blocks etc) then our calculated header_len will be different, and 158333ad798cSAdam Langley * we have to adjust mss_now correspondingly */ 158433ad798cSAdam Langley if (header_len != tp->tcp_header_len) { 158533ad798cSAdam Langley int delta = (int) header_len - tp->tcp_header_len; 158633ad798cSAdam Langley mss_now -= delta; 158733ad798cSAdam Langley } 1588cfb6eeb4SYOSHIFUJI Hideaki 15891da177e4SLinus Torvalds return mss_now; 15901da177e4SLinus Torvalds } 15911da177e4SLinus Torvalds 159286fd14adSWeiping Pan /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. 159386fd14adSWeiping Pan * As additional protections, we do not touch cwnd in retransmission phases, 159486fd14adSWeiping Pan * and if application hit its sndbuf limit recently. 159586fd14adSWeiping Pan */ 159686fd14adSWeiping Pan static void tcp_cwnd_application_limited(struct sock *sk) 1597a762a980SDavid S. Miller { 15989e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1599a762a980SDavid S. Miller 160086fd14adSWeiping Pan if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && 160186fd14adSWeiping Pan sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 160286fd14adSWeiping Pan /* Limited by application or receiver window. */ 160386fd14adSWeiping Pan u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); 160486fd14adSWeiping Pan u32 win_used = max(tp->snd_cwnd_used, init_win); 160586fd14adSWeiping Pan if (win_used < tp->snd_cwnd) { 160686fd14adSWeiping Pan tp->snd_ssthresh = tcp_current_ssthresh(sk); 160786fd14adSWeiping Pan tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1; 160886fd14adSWeiping Pan } 160986fd14adSWeiping Pan tp->snd_cwnd_used = 0; 161086fd14adSWeiping Pan } 1611c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 161286fd14adSWeiping Pan } 161386fd14adSWeiping Pan 1614ca8a2263SNeal Cardwell static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited) 1615a762a980SDavid S. Miller { 16161b1fc3fdSWei Wang const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 1617a762a980SDavid S. Miller struct tcp_sock *tp = tcp_sk(sk); 1618a762a980SDavid S. Miller 1619ca8a2263SNeal Cardwell /* Track the maximum number of outstanding packets in each 1620ca8a2263SNeal Cardwell * window, and remember whether we were cwnd-limited then. 1621ca8a2263SNeal Cardwell */ 1622ca8a2263SNeal Cardwell if (!before(tp->snd_una, tp->max_packets_seq) || 1623ca8a2263SNeal Cardwell tp->packets_out > tp->max_packets_out) { 1624ca8a2263SNeal Cardwell tp->max_packets_out = tp->packets_out; 1625ca8a2263SNeal Cardwell tp->max_packets_seq = tp->snd_nxt; 1626ca8a2263SNeal Cardwell tp->is_cwnd_limited = is_cwnd_limited; 1627ca8a2263SNeal Cardwell } 1628e114a710SEric Dumazet 162924901551SEric Dumazet if (tcp_is_cwnd_limited(sk)) { 1630a762a980SDavid S. Miller /* Network is feed fully. */ 1631a762a980SDavid S. Miller tp->snd_cwnd_used = 0; 1632c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 1633a762a980SDavid S. Miller } else { 1634a762a980SDavid S. Miller /* Network starves. */ 1635a762a980SDavid S. Miller if (tp->packets_out > tp->snd_cwnd_used) 1636a762a980SDavid S. Miller tp->snd_cwnd_used = tp->packets_out; 1637a762a980SDavid S. Miller 1638b510f0d2SEric Dumazet if (sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle && 1639c2203cf7SEric Dumazet (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto && 16401b1fc3fdSWei Wang !ca_ops->cong_control) 1641a762a980SDavid S. Miller tcp_cwnd_application_limited(sk); 1642b0f71bd3SFrancis Yan 1643b0f71bd3SFrancis Yan /* The following conditions together indicate the starvation 1644b0f71bd3SFrancis Yan * is caused by insufficient sender buffer: 1645b0f71bd3SFrancis Yan * 1) just sent some data (see tcp_write_xmit) 1646b0f71bd3SFrancis Yan * 2) not cwnd limited (this else condition) 164775c119afSEric Dumazet * 3) no more data to send (tcp_write_queue_empty()) 1648b0f71bd3SFrancis Yan * 4) application is hitting buffer limit (SOCK_NOSPACE) 1649b0f71bd3SFrancis Yan */ 165075c119afSEric Dumazet if (tcp_write_queue_empty(sk) && sk->sk_socket && 1651b0f71bd3SFrancis Yan test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) && 1652b0f71bd3SFrancis Yan (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) 1653b0f71bd3SFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED); 1654a762a980SDavid S. Miller } 1655a762a980SDavid S. Miller } 1656a762a980SDavid S. Miller 1657d4589926SEric Dumazet /* Minshall's variant of the Nagle send check. */ 1658d4589926SEric Dumazet static bool tcp_minshall_check(const struct tcp_sock *tp) 1659d4589926SEric Dumazet { 1660d4589926SEric Dumazet return after(tp->snd_sml, tp->snd_una) && 1661d4589926SEric Dumazet !after(tp->snd_sml, tp->snd_nxt); 1662d4589926SEric Dumazet } 1663d4589926SEric Dumazet 1664d4589926SEric Dumazet /* Update snd_sml if this skb is under mss 1665d4589926SEric Dumazet * Note that a TSO packet might end with a sub-mss segment 1666d4589926SEric Dumazet * The test is really : 1667d4589926SEric Dumazet * if ((skb->len % mss) != 0) 1668d4589926SEric Dumazet * tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1669d4589926SEric Dumazet * But we can avoid doing the divide again given we already have 1670d4589926SEric Dumazet * skb_pcount = skb->len / mss_now 16710e3a4803SIlpo Järvinen */ 1672d4589926SEric Dumazet static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now, 1673d4589926SEric Dumazet const struct sk_buff *skb) 1674d4589926SEric Dumazet { 1675d4589926SEric Dumazet if (skb->len < tcp_skb_pcount(skb) * mss_now) 1676d4589926SEric Dumazet tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1677d4589926SEric Dumazet } 1678d4589926SEric Dumazet 1679d4589926SEric Dumazet /* Return false, if packet can be sent now without violation Nagle's rules: 1680d4589926SEric Dumazet * 1. It is full sized. (provided by caller in %partial bool) 1681d4589926SEric Dumazet * 2. Or it contains FIN. (already checked by caller) 1682d4589926SEric Dumazet * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. 1683d4589926SEric Dumazet * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 1684d4589926SEric Dumazet * With Minshall's modification: all sent small packets are ACKed. 1685d4589926SEric Dumazet */ 1686d4589926SEric Dumazet static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp, 1687cc93fc51SPeter Pan(潘卫平) int nonagle) 1688d4589926SEric Dumazet { 1689d4589926SEric Dumazet return partial && 1690d4589926SEric Dumazet ((nonagle & TCP_NAGLE_CORK) || 1691d4589926SEric Dumazet (!nonagle && tp->packets_out && tcp_minshall_check(tp))); 1692d4589926SEric Dumazet } 1693605ad7f1SEric Dumazet 1694605ad7f1SEric Dumazet /* Return how many segs we'd like on a TSO packet, 1695605ad7f1SEric Dumazet * to send one TSO packet per ms 1696605ad7f1SEric Dumazet */ 1697dcb8c9b4SEric Dumazet static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now, 16981b3878caSNeal Cardwell int min_tso_segs) 1699605ad7f1SEric Dumazet { 1700605ad7f1SEric Dumazet u32 bytes, segs; 1701605ad7f1SEric Dumazet 170276a9ebe8SEric Dumazet bytes = min_t(unsigned long, 170376a9ebe8SEric Dumazet sk->sk_pacing_rate >> sk->sk_pacing_shift, 1704605ad7f1SEric Dumazet sk->sk_gso_max_size - 1 - MAX_TCP_HEADER); 1705605ad7f1SEric Dumazet 1706605ad7f1SEric Dumazet /* Goal is to send at least one packet per ms, 1707605ad7f1SEric Dumazet * not one big TSO packet every 100 ms. 1708605ad7f1SEric Dumazet * This preserves ACK clocking and is consistent 1709605ad7f1SEric Dumazet * with tcp_tso_should_defer() heuristic. 1710605ad7f1SEric Dumazet */ 17111b3878caSNeal Cardwell segs = max_t(u32, bytes / mss_now, min_tso_segs); 1712605ad7f1SEric Dumazet 1713350c9f48SEric Dumazet return segs; 1714605ad7f1SEric Dumazet } 1715605ad7f1SEric Dumazet 1716ed6e7268SNeal Cardwell /* Return the number of segments we want in the skb we are transmitting. 1717ed6e7268SNeal Cardwell * See if congestion control module wants to decide; otherwise, autosize. 1718ed6e7268SNeal Cardwell */ 1719ed6e7268SNeal Cardwell static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now) 1720ed6e7268SNeal Cardwell { 1721ed6e7268SNeal Cardwell const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 1722dcb8c9b4SEric Dumazet u32 min_tso, tso_segs; 1723ed6e7268SNeal Cardwell 1724dcb8c9b4SEric Dumazet min_tso = ca_ops->min_tso_segs ? 1725dcb8c9b4SEric Dumazet ca_ops->min_tso_segs(sk) : 1726dcb8c9b4SEric Dumazet sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs; 1727dcb8c9b4SEric Dumazet 1728dcb8c9b4SEric Dumazet tso_segs = tcp_tso_autosize(sk, mss_now, min_tso); 1729350c9f48SEric Dumazet return min_t(u32, tso_segs, sk->sk_gso_max_segs); 1730ed6e7268SNeal Cardwell } 1731ed6e7268SNeal Cardwell 1732d4589926SEric Dumazet /* Returns the portion of skb which can be sent right away */ 1733d4589926SEric Dumazet static unsigned int tcp_mss_split_point(const struct sock *sk, 1734d4589926SEric Dumazet const struct sk_buff *skb, 1735d4589926SEric Dumazet unsigned int mss_now, 1736d4589926SEric Dumazet unsigned int max_segs, 1737d4589926SEric Dumazet int nonagle) 1738c1b4a7e6SDavid S. Miller { 1739cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1740d4589926SEric Dumazet u32 partial, needed, window, max_len; 1741c1b4a7e6SDavid S. Miller 174290840defSIlpo Järvinen window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 17431485348dSBen Hutchings max_len = mss_now * max_segs; 17440e3a4803SIlpo Järvinen 17451485348dSBen Hutchings if (likely(max_len <= window && skb != tcp_write_queue_tail(sk))) 17461485348dSBen Hutchings return max_len; 17470e3a4803SIlpo Järvinen 17485ea3a748SIlpo Järvinen needed = min(skb->len, window); 17495ea3a748SIlpo Järvinen 17501485348dSBen Hutchings if (max_len <= needed) 17511485348dSBen Hutchings return max_len; 17520e3a4803SIlpo Järvinen 1753d4589926SEric Dumazet partial = needed % mss_now; 1754d4589926SEric Dumazet /* If last segment is not a full MSS, check if Nagle rules allow us 1755d4589926SEric Dumazet * to include this last segment in this skb. 1756d4589926SEric Dumazet * Otherwise, we'll split the skb at last MSS boundary 1757d4589926SEric Dumazet */ 1758cc93fc51SPeter Pan(潘卫平) if (tcp_nagle_check(partial != 0, tp, nonagle)) 1759d4589926SEric Dumazet return needed - partial; 1760d4589926SEric Dumazet 1761d4589926SEric Dumazet return needed; 1762c1b4a7e6SDavid S. Miller } 1763c1b4a7e6SDavid S. Miller 1764c1b4a7e6SDavid S. Miller /* Can at least one segment of SKB be sent right now, according to the 1765c1b4a7e6SDavid S. Miller * congestion window rules? If so, return how many segments are allowed. 1766c1b4a7e6SDavid S. Miller */ 1767cf533ea5SEric Dumazet static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, 1768cf533ea5SEric Dumazet const struct sk_buff *skb) 1769c1b4a7e6SDavid S. Miller { 1770d649a7a8SEric Dumazet u32 in_flight, cwnd, halfcwnd; 1771c1b4a7e6SDavid S. Miller 1772c1b4a7e6SDavid S. Miller /* Don't be strict about the congestion window for the final FIN. */ 17734de075e0SEric Dumazet if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && 17744de075e0SEric Dumazet tcp_skb_pcount(skb) == 1) 1775c1b4a7e6SDavid S. Miller return 1; 1776c1b4a7e6SDavid S. Miller 1777c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1778c1b4a7e6SDavid S. Miller cwnd = tp->snd_cwnd; 1779d649a7a8SEric Dumazet if (in_flight >= cwnd) 1780c1b4a7e6SDavid S. Miller return 0; 1781d649a7a8SEric Dumazet 1782d649a7a8SEric Dumazet /* For better scheduling, ensure we have at least 1783d649a7a8SEric Dumazet * 2 GSO packets in flight. 1784d649a7a8SEric Dumazet */ 1785d649a7a8SEric Dumazet halfcwnd = max(cwnd >> 1, 1U); 1786d649a7a8SEric Dumazet return min(halfcwnd, cwnd - in_flight); 1787c1b4a7e6SDavid S. Miller } 1788c1b4a7e6SDavid S. Miller 1789b595076aSUwe Kleine-König /* Initialize TSO state of a skb. 179067edfef7SAndi Kleen * This must be invoked the first time we consider transmitting 1791c1b4a7e6SDavid S. Miller * SKB onto the wire. 1792c1b4a7e6SDavid S. Miller */ 17935bbb432cSEric Dumazet static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now) 1794c1b4a7e6SDavid S. Miller { 1795c1b4a7e6SDavid S. Miller int tso_segs = tcp_skb_pcount(skb); 1796c1b4a7e6SDavid S. Miller 1797f8269a49SIlpo Järvinen if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { 17985bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 1799c1b4a7e6SDavid S. Miller tso_segs = tcp_skb_pcount(skb); 1800c1b4a7e6SDavid S. Miller } 1801c1b4a7e6SDavid S. Miller return tso_segs; 1802c1b4a7e6SDavid S. Miller } 1803c1b4a7e6SDavid S. Miller 1804c1b4a7e6SDavid S. Miller 1805a2a385d6SEric Dumazet /* Return true if the Nagle test allows this packet to be 1806c1b4a7e6SDavid S. Miller * sent now. 1807c1b4a7e6SDavid S. Miller */ 1808a2a385d6SEric Dumazet static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, 1809c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1810c1b4a7e6SDavid S. Miller { 1811c1b4a7e6SDavid S. Miller /* Nagle rule does not apply to frames, which sit in the middle of the 1812c1b4a7e6SDavid S. Miller * write_queue (they have no chances to get new data). 1813c1b4a7e6SDavid S. Miller * 1814c1b4a7e6SDavid S. Miller * This is implemented in the callers, where they modify the 'nonagle' 1815c1b4a7e6SDavid S. Miller * argument based upon the location of SKB in the send queue. 1816c1b4a7e6SDavid S. Miller */ 1817c1b4a7e6SDavid S. Miller if (nonagle & TCP_NAGLE_PUSH) 1818a2a385d6SEric Dumazet return true; 1819c1b4a7e6SDavid S. Miller 18209b44190dSYuchung Cheng /* Don't use the nagle rule for urgent data (or for the final FIN). */ 18219b44190dSYuchung Cheng if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) 1822a2a385d6SEric Dumazet return true; 1823c1b4a7e6SDavid S. Miller 1824cc93fc51SPeter Pan(潘卫平) if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle)) 1825a2a385d6SEric Dumazet return true; 1826c1b4a7e6SDavid S. Miller 1827a2a385d6SEric Dumazet return false; 1828c1b4a7e6SDavid S. Miller } 1829c1b4a7e6SDavid S. Miller 1830c1b4a7e6SDavid S. Miller /* Does at least the first segment of SKB fit into the send window? */ 1831a2a385d6SEric Dumazet static bool tcp_snd_wnd_test(const struct tcp_sock *tp, 1832a2a385d6SEric Dumazet const struct sk_buff *skb, 1833056834d9SIlpo Järvinen unsigned int cur_mss) 1834c1b4a7e6SDavid S. Miller { 1835c1b4a7e6SDavid S. Miller u32 end_seq = TCP_SKB_CB(skb)->end_seq; 1836c1b4a7e6SDavid S. Miller 1837c1b4a7e6SDavid S. Miller if (skb->len > cur_mss) 1838c1b4a7e6SDavid S. Miller end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 1839c1b4a7e6SDavid S. Miller 184090840defSIlpo Järvinen return !after(end_seq, tcp_wnd_end(tp)); 1841c1b4a7e6SDavid S. Miller } 1842c1b4a7e6SDavid S. Miller 1843c1b4a7e6SDavid S. Miller /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 1844c1b4a7e6SDavid S. Miller * which is put after SKB on the list. It is very much like 1845c1b4a7e6SDavid S. Miller * tcp_fragment() except that it may make several kinds of assumptions 1846c1b4a7e6SDavid S. Miller * in order to speed up the splitting operation. In particular, we 1847c1b4a7e6SDavid S. Miller * know that all the data is in scatter-gather pages, and that the 1848c1b4a7e6SDavid S. Miller * packet has never been sent out before (and thus is not cloned). 1849c1b4a7e6SDavid S. Miller */ 185075c119afSEric Dumazet static int tso_fragment(struct sock *sk, enum tcp_queue tcp_queue, 185175c119afSEric Dumazet struct sk_buff *skb, unsigned int len, 1852c4ead4c5SEric Dumazet unsigned int mss_now, gfp_t gfp) 1853c1b4a7e6SDavid S. Miller { 1854c1b4a7e6SDavid S. Miller struct sk_buff *buff; 1855c1b4a7e6SDavid S. Miller int nlen = skb->len - len; 18569ce01461SIlpo Järvinen u8 flags; 1857c1b4a7e6SDavid S. Miller 1858c1b4a7e6SDavid S. Miller /* All of a TSO frame must be composed of paged data. */ 1859c8ac3774SHerbert Xu if (skb->len != skb->data_len) 186075c119afSEric Dumazet return tcp_fragment(sk, tcp_queue, skb, len, mss_now, gfp); 1861c1b4a7e6SDavid S. Miller 1862eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, gfp, true); 186351456b29SIan Morris if (unlikely(!buff)) 1864c1b4a7e6SDavid S. Miller return -ENOMEM; 1865c1b4a7e6SDavid S. Miller 18663ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 18673ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1868b60b49eaSHerbert Xu buff->truesize += nlen; 1869c1b4a7e6SDavid S. Miller skb->truesize -= nlen; 1870c1b4a7e6SDavid S. Miller 1871c1b4a7e6SDavid S. Miller /* Correct the sequence numbers. */ 1872c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 1873c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 1874c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 1875c1b4a7e6SDavid S. Miller 1876c1b4a7e6SDavid S. Miller /* PSH and FIN should only be set in the second packet. */ 18774de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 18784de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 18794de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1880c1b4a7e6SDavid S. Miller 1881c1b4a7e6SDavid S. Miller /* This packet was never sent out yet, so no SACK bits. */ 1882c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->sacked = 0; 1883c1b4a7e6SDavid S. Miller 1884a166140eSMartin KaFai Lau tcp_skb_fragment_eor(skb, buff); 1885a166140eSMartin KaFai Lau 188698be9b12SEric Dumazet buff->ip_summed = CHECKSUM_PARTIAL; 1887c1b4a7e6SDavid S. Miller skb_split(skb, buff, len); 1888490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 1889c1b4a7e6SDavid S. Miller 1890c1b4a7e6SDavid S. Miller /* Fix up tso_factor for both original and new SKB. */ 18915bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 18925bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 1893c1b4a7e6SDavid S. Miller 1894c1b4a7e6SDavid S. Miller /* Link BUFF into the send queue. */ 1895f4a775d1SEric Dumazet __skb_header_release(buff); 189675c119afSEric Dumazet tcp_insert_write_queue_after(skb, buff, sk, tcp_queue); 1897c1b4a7e6SDavid S. Miller 1898c1b4a7e6SDavid S. Miller return 0; 1899c1b4a7e6SDavid S. Miller } 1900c1b4a7e6SDavid S. Miller 1901c1b4a7e6SDavid S. Miller /* Try to defer sending, if possible, in order to minimize the amount 1902c1b4a7e6SDavid S. Miller * of TSO splitting we do. View it as a kind of TSO Nagle test. 1903c1b4a7e6SDavid S. Miller * 1904c1b4a7e6SDavid S. Miller * This algorithm is from John Heffner. 1905c1b4a7e6SDavid S. Miller */ 1906ca8a2263SNeal Cardwell static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb, 1907605ad7f1SEric Dumazet bool *is_cwnd_limited, u32 max_segs) 1908c1b4a7e6SDavid S. Miller { 19096687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 191050c8339eSEric Dumazet u32 age, send_win, cong_win, limit, in_flight; 191150c8339eSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 191250c8339eSEric Dumazet struct sk_buff *head; 1913ad9f4f50SEric Dumazet int win_divisor; 1914c1b4a7e6SDavid S. Miller 19154de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 1916ae8064acSJohn Heffner goto send_now; 1917c1b4a7e6SDavid S. Miller 191899d7662aSEric Dumazet if (icsk->icsk_ca_state >= TCP_CA_Recovery) 1919ae8064acSJohn Heffner goto send_now; 1920ae8064acSJohn Heffner 19215f852eb5SEric Dumazet /* Avoid bursty behavior by allowing defer 19225f852eb5SEric Dumazet * only if the last write was recent. 19235f852eb5SEric Dumazet */ 1924d635fbe2SEric Dumazet if ((s32)(tcp_jiffies32 - tp->lsndtime) > 0) 1925ae8064acSJohn Heffner goto send_now; 1926908a75c1SDavid S. Miller 1927c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1928c1b4a7e6SDavid S. Miller 1929c8c9aeb5SStefano Brivio BUG_ON(tcp_skb_pcount(skb) <= 1); 1930c8c9aeb5SStefano Brivio BUG_ON(tp->snd_cwnd <= in_flight); 1931c1b4a7e6SDavid S. Miller 193290840defSIlpo Järvinen send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1933c1b4a7e6SDavid S. Miller 1934c1b4a7e6SDavid S. Miller /* From in_flight test above, we know that cwnd > in_flight. */ 1935c1b4a7e6SDavid S. Miller cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; 1936c1b4a7e6SDavid S. Miller 1937c1b4a7e6SDavid S. Miller limit = min(send_win, cong_win); 1938c1b4a7e6SDavid S. Miller 1939ba244fe9SDavid S. Miller /* If a full-sized TSO skb can be sent, do it. */ 1940605ad7f1SEric Dumazet if (limit >= max_segs * tp->mss_cache) 1941ae8064acSJohn Heffner goto send_now; 1942ba244fe9SDavid S. Miller 194362ad2761SIlpo Järvinen /* Middle in queue won't get any more data, full sendable already? */ 194462ad2761SIlpo Järvinen if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) 194562ad2761SIlpo Järvinen goto send_now; 194662ad2761SIlpo Järvinen 19475bbcc0f5SLinus Torvalds win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor); 1948ad9f4f50SEric Dumazet if (win_divisor) { 1949c1b4a7e6SDavid S. Miller u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); 1950c1b4a7e6SDavid S. Miller 1951c1b4a7e6SDavid S. Miller /* If at least some fraction of a window is available, 1952c1b4a7e6SDavid S. Miller * just use it. 1953c1b4a7e6SDavid S. Miller */ 1954ad9f4f50SEric Dumazet chunk /= win_divisor; 1955c1b4a7e6SDavid S. Miller if (limit >= chunk) 1956ae8064acSJohn Heffner goto send_now; 1957c1b4a7e6SDavid S. Miller } else { 1958c1b4a7e6SDavid S. Miller /* Different approach, try not to defer past a single 1959c1b4a7e6SDavid S. Miller * ACK. Receiver should ACK every other full sized 1960c1b4a7e6SDavid S. Miller * frame, so if we have space for more than 3 frames 1961c1b4a7e6SDavid S. Miller * then send now. 1962c1b4a7e6SDavid S. Miller */ 19636b5a5c0dSNeal Cardwell if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) 1964ae8064acSJohn Heffner goto send_now; 1965c1b4a7e6SDavid S. Miller } 1966c1b4a7e6SDavid S. Miller 196775c119afSEric Dumazet /* TODO : use tsorted_sent_queue ? */ 196875c119afSEric Dumazet head = tcp_rtx_queue_head(sk); 196975c119afSEric Dumazet if (!head) 197075c119afSEric Dumazet goto send_now; 19712fd66ffbSEric Dumazet age = tcp_stamp_us_delta(tp->tcp_mstamp, tcp_skb_timestamp_us(head)); 197250c8339eSEric Dumazet /* If next ACK is likely to come too late (half srtt), do not defer */ 197350c8339eSEric Dumazet if (age < (tp->srtt_us >> 4)) 197450c8339eSEric Dumazet goto send_now; 197550c8339eSEric Dumazet 19765f852eb5SEric Dumazet /* Ok, it looks like it is advisable to defer. */ 1977ae8064acSJohn Heffner 1978d2e1339fSBendik Rønning Opstad if (cong_win < send_win && cong_win <= skb->len) 1979ca8a2263SNeal Cardwell *is_cwnd_limited = true; 1980ca8a2263SNeal Cardwell 1981a2a385d6SEric Dumazet return true; 1982ae8064acSJohn Heffner 1983ae8064acSJohn Heffner send_now: 1984a2a385d6SEric Dumazet return false; 1985c1b4a7e6SDavid S. Miller } 1986c1b4a7e6SDavid S. Miller 198705cbc0dbSFan Du static inline void tcp_mtu_check_reprobe(struct sock *sk) 198805cbc0dbSFan Du { 198905cbc0dbSFan Du struct inet_connection_sock *icsk = inet_csk(sk); 199005cbc0dbSFan Du struct tcp_sock *tp = tcp_sk(sk); 199105cbc0dbSFan Du struct net *net = sock_net(sk); 199205cbc0dbSFan Du u32 interval; 199305cbc0dbSFan Du s32 delta; 199405cbc0dbSFan Du 199505cbc0dbSFan Du interval = net->ipv4.sysctl_tcp_probe_interval; 1996c74df29aSEric Dumazet delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp; 199705cbc0dbSFan Du if (unlikely(delta >= interval * HZ)) { 199805cbc0dbSFan Du int mss = tcp_current_mss(sk); 199905cbc0dbSFan Du 200005cbc0dbSFan Du /* Update current search range */ 200105cbc0dbSFan Du icsk->icsk_mtup.probe_size = 0; 200205cbc0dbSFan Du icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + 200305cbc0dbSFan Du sizeof(struct tcphdr) + 200405cbc0dbSFan Du icsk->icsk_af_ops->net_header_len; 200505cbc0dbSFan Du icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss); 200605cbc0dbSFan Du 200705cbc0dbSFan Du /* Update probe time stamp */ 2008c74df29aSEric Dumazet icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; 200905cbc0dbSFan Du } 201005cbc0dbSFan Du } 201105cbc0dbSFan Du 2012808cf9e3SIlya Lesokhin static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len) 2013808cf9e3SIlya Lesokhin { 2014808cf9e3SIlya Lesokhin struct sk_buff *skb, *next; 2015808cf9e3SIlya Lesokhin 2016808cf9e3SIlya Lesokhin skb = tcp_send_head(sk); 2017808cf9e3SIlya Lesokhin tcp_for_write_queue_from_safe(skb, next, sk) { 2018808cf9e3SIlya Lesokhin if (len <= skb->len) 2019808cf9e3SIlya Lesokhin break; 2020808cf9e3SIlya Lesokhin 2021808cf9e3SIlya Lesokhin if (unlikely(TCP_SKB_CB(skb)->eor)) 2022808cf9e3SIlya Lesokhin return false; 2023808cf9e3SIlya Lesokhin 2024808cf9e3SIlya Lesokhin len -= skb->len; 2025808cf9e3SIlya Lesokhin } 2026808cf9e3SIlya Lesokhin 2027808cf9e3SIlya Lesokhin return true; 2028808cf9e3SIlya Lesokhin } 2029808cf9e3SIlya Lesokhin 20305d424d5aSJohn Heffner /* Create a new MTU probe if we are ready. 203167edfef7SAndi Kleen * MTU probe is regularly attempting to increase the path MTU by 203267edfef7SAndi Kleen * deliberately sending larger packets. This discovers routing 203367edfef7SAndi Kleen * changes resulting in larger path MTUs. 203467edfef7SAndi Kleen * 20355d424d5aSJohn Heffner * Returns 0 if we should wait to probe (no cwnd available), 20365d424d5aSJohn Heffner * 1 if a probe was sent, 2037056834d9SIlpo Järvinen * -1 otherwise 2038056834d9SIlpo Järvinen */ 20395d424d5aSJohn Heffner static int tcp_mtu_probe(struct sock *sk) 20405d424d5aSJohn Heffner { 20415d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 204212a59abcSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 20435d424d5aSJohn Heffner struct sk_buff *skb, *nskb, *next; 20446b58e0a5SFan Du struct net *net = sock_net(sk); 20455d424d5aSJohn Heffner int probe_size; 204691cc17c0SIlpo Järvinen int size_needed; 204712a59abcSEric Dumazet int copy, len; 20485d424d5aSJohn Heffner int mss_now; 20496b58e0a5SFan Du int interval; 20505d424d5aSJohn Heffner 20515d424d5aSJohn Heffner /* Not currently probing/verifying, 20525d424d5aSJohn Heffner * not in recovery, 20535d424d5aSJohn Heffner * have enough cwnd, and 205412a59abcSEric Dumazet * not SACKing (the variable headers throw things off) 205512a59abcSEric Dumazet */ 205612a59abcSEric Dumazet if (likely(!icsk->icsk_mtup.enabled || 20575d424d5aSJohn Heffner icsk->icsk_mtup.probe_size || 20585d424d5aSJohn Heffner inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 20595d424d5aSJohn Heffner tp->snd_cwnd < 11 || 206012a59abcSEric Dumazet tp->rx_opt.num_sacks || tp->rx_opt.dsack)) 20615d424d5aSJohn Heffner return -1; 20625d424d5aSJohn Heffner 20636b58e0a5SFan Du /* Use binary search for probe_size between tcp_mss_base, 20646b58e0a5SFan Du * and current mss_clamp. if (search_high - search_low) 20656b58e0a5SFan Du * smaller than a threshold, backoff from probing. 20666b58e0a5SFan Du */ 20670c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 20686b58e0a5SFan Du probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high + 20696b58e0a5SFan Du icsk->icsk_mtup.search_low) >> 1); 207091cc17c0SIlpo Järvinen size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; 20716b58e0a5SFan Du interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low; 207205cbc0dbSFan Du /* When misfortune happens, we are reprobing actively, 207305cbc0dbSFan Du * and then reprobe timer has expired. We stick with current 207405cbc0dbSFan Du * probing process by not resetting search range to its orignal. 207505cbc0dbSFan Du */ 20766b58e0a5SFan Du if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) || 207705cbc0dbSFan Du interval < net->ipv4.sysctl_tcp_probe_threshold) { 207805cbc0dbSFan Du /* Check whether enough time has elaplased for 207905cbc0dbSFan Du * another round of probing. 208005cbc0dbSFan Du */ 208105cbc0dbSFan Du tcp_mtu_check_reprobe(sk); 20825d424d5aSJohn Heffner return -1; 20835d424d5aSJohn Heffner } 20845d424d5aSJohn Heffner 20855d424d5aSJohn Heffner /* Have enough data in the send queue to probe? */ 20867f9c33e5SIlpo Järvinen if (tp->write_seq - tp->snd_nxt < size_needed) 20875d424d5aSJohn Heffner return -1; 20885d424d5aSJohn Heffner 208991cc17c0SIlpo Järvinen if (tp->snd_wnd < size_needed) 20905d424d5aSJohn Heffner return -1; 209190840defSIlpo Järvinen if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) 20925d424d5aSJohn Heffner return 0; 20935d424d5aSJohn Heffner 2094d67c58e9SIlpo Järvinen /* Do we need to wait to drain cwnd? With none in flight, don't stall */ 2095d67c58e9SIlpo Järvinen if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { 2096d67c58e9SIlpo Järvinen if (!tcp_packets_in_flight(tp)) 20975d424d5aSJohn Heffner return -1; 20985d424d5aSJohn Heffner else 20995d424d5aSJohn Heffner return 0; 21005d424d5aSJohn Heffner } 21015d424d5aSJohn Heffner 2102808cf9e3SIlya Lesokhin if (!tcp_can_coalesce_send_queue_head(sk, probe_size)) 2103808cf9e3SIlya Lesokhin return -1; 2104808cf9e3SIlya Lesokhin 21055d424d5aSJohn Heffner /* We're allowed to probe. Build it now. */ 2106eb934478SEric Dumazet nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false); 210751456b29SIan Morris if (!nskb) 21085d424d5aSJohn Heffner return -1; 21093ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 21103ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 21115d424d5aSJohn Heffner 2112fe067e8aSDavid S. Miller skb = tcp_send_head(sk); 21135d424d5aSJohn Heffner 21145d424d5aSJohn Heffner TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 21155d424d5aSJohn Heffner TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 21164de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; 21175d424d5aSJohn Heffner TCP_SKB_CB(nskb)->sacked = 0; 21185d424d5aSJohn Heffner nskb->csum = 0; 211998be9b12SEric Dumazet nskb->ip_summed = CHECKSUM_PARTIAL; 21205d424d5aSJohn Heffner 212150c4817eSIlpo Järvinen tcp_insert_write_queue_before(nskb, skb, sk); 21222b7cda9cSEric Dumazet tcp_highest_sack_replace(sk, skb, nskb); 212350c4817eSIlpo Järvinen 21245d424d5aSJohn Heffner len = 0; 2125234b6860SIlpo Järvinen tcp_for_write_queue_from_safe(skb, next, sk) { 21265d424d5aSJohn Heffner copy = min_t(int, skb->len, probe_size - len); 21275d424d5aSJohn Heffner skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); 21285d424d5aSJohn Heffner 21295d424d5aSJohn Heffner if (skb->len <= copy) { 21305d424d5aSJohn Heffner /* We've eaten all the data from this skb. 21315d424d5aSJohn Heffner * Throw it away. */ 21324de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 2133808cf9e3SIlya Lesokhin /* If this is the last SKB we copy and eor is set 2134808cf9e3SIlya Lesokhin * we need to propagate it to the new skb. 2135808cf9e3SIlya Lesokhin */ 2136808cf9e3SIlya Lesokhin TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor; 2137fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 21383ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 21395d424d5aSJohn Heffner } else { 21404de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & 2141a3433f35SChangli Gao ~(TCPHDR_FIN|TCPHDR_PSH); 21425d424d5aSJohn Heffner if (!skb_shinfo(skb)->nr_frags) { 21435d424d5aSJohn Heffner skb_pull(skb, copy); 21445d424d5aSJohn Heffner } else { 21455d424d5aSJohn Heffner __pskb_trim_head(skb, copy); 21465bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 21475d424d5aSJohn Heffner } 21485d424d5aSJohn Heffner TCP_SKB_CB(skb)->seq += copy; 21495d424d5aSJohn Heffner } 21505d424d5aSJohn Heffner 21515d424d5aSJohn Heffner len += copy; 2152234b6860SIlpo Järvinen 2153234b6860SIlpo Järvinen if (len >= probe_size) 2154234b6860SIlpo Järvinen break; 21555d424d5aSJohn Heffner } 21565bbb432cSEric Dumazet tcp_init_tso_segs(nskb, nskb->len); 21575d424d5aSJohn Heffner 21585d424d5aSJohn Heffner /* We're ready to send. If this fails, the probe will 21597faee5c0SEric Dumazet * be resegmented into mss-sized pieces by tcp_write_xmit(). 21607faee5c0SEric Dumazet */ 21615d424d5aSJohn Heffner if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 21625d424d5aSJohn Heffner /* Decrement cwnd here because we are sending 21635d424d5aSJohn Heffner * effectively two packets. */ 21645d424d5aSJohn Heffner tp->snd_cwnd--; 216566f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, nskb); 21665d424d5aSJohn Heffner 21675d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 21680e7b1368SJohn Heffner tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 21690e7b1368SJohn Heffner tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 21705d424d5aSJohn Heffner 21715d424d5aSJohn Heffner return 1; 21725d424d5aSJohn Heffner } 21735d424d5aSJohn Heffner 21745d424d5aSJohn Heffner return -1; 21755d424d5aSJohn Heffner } 21765d424d5aSJohn Heffner 2177864e5c09SEric Dumazet static bool tcp_pacing_check(struct sock *sk) 2178218af599SEric Dumazet { 2179864e5c09SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 2180864e5c09SEric Dumazet 2181864e5c09SEric Dumazet if (!tcp_needs_internal_pacing(sk)) 2182864e5c09SEric Dumazet return false; 2183864e5c09SEric Dumazet 2184864e5c09SEric Dumazet if (tp->tcp_wstamp_ns <= tp->tcp_clock_cache) 2185864e5c09SEric Dumazet return false; 2186864e5c09SEric Dumazet 2187864e5c09SEric Dumazet if (!hrtimer_is_queued(&tp->pacing_timer)) { 2188864e5c09SEric Dumazet hrtimer_start(&tp->pacing_timer, 2189864e5c09SEric Dumazet ns_to_ktime(tp->tcp_wstamp_ns), 2190864e5c09SEric Dumazet HRTIMER_MODE_ABS_PINNED_SOFT); 2191864e5c09SEric Dumazet sock_hold(sk); 2192864e5c09SEric Dumazet } 2193864e5c09SEric Dumazet return true; 2194218af599SEric Dumazet } 2195218af599SEric Dumazet 2196f9616c35SEric Dumazet /* TCP Small Queues : 2197f9616c35SEric Dumazet * Control number of packets in qdisc/devices to two packets / or ~1 ms. 2198f9616c35SEric Dumazet * (These limits are doubled for retransmits) 2199f9616c35SEric Dumazet * This allows for : 2200f9616c35SEric Dumazet * - better RTT estimation and ACK scheduling 2201f9616c35SEric Dumazet * - faster recovery 2202f9616c35SEric Dumazet * - high rates 2203f9616c35SEric Dumazet * Alas, some drivers / subsystems require a fair amount 2204f9616c35SEric Dumazet * of queued bytes to ensure line rate. 2205f9616c35SEric Dumazet * One example is wifi aggregation (802.11 AMPDU) 2206f9616c35SEric Dumazet */ 2207f9616c35SEric Dumazet static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb, 2208f9616c35SEric Dumazet unsigned int factor) 2209f9616c35SEric Dumazet { 221076a9ebe8SEric Dumazet unsigned long limit; 2211f9616c35SEric Dumazet 221276a9ebe8SEric Dumazet limit = max_t(unsigned long, 221376a9ebe8SEric Dumazet 2 * skb->truesize, 221476a9ebe8SEric Dumazet sk->sk_pacing_rate >> sk->sk_pacing_shift); 221576a9ebe8SEric Dumazet limit = min_t(unsigned long, limit, 22169184d8bbSEric Dumazet sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes); 2217f9616c35SEric Dumazet limit <<= factor; 2218f9616c35SEric Dumazet 221914afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) > limit) { 222075c119afSEric Dumazet /* Always send skb if rtx queue is empty. 222175eefc6cSEric Dumazet * No need to wait for TX completion to call us back, 222275eefc6cSEric Dumazet * after softirq/tasklet schedule. 222375eefc6cSEric Dumazet * This helps when TX completions are delayed too much. 222475eefc6cSEric Dumazet */ 222575c119afSEric Dumazet if (tcp_rtx_queue_empty(sk)) 222675eefc6cSEric Dumazet return false; 222775eefc6cSEric Dumazet 22287aa5470cSEric Dumazet set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags); 2229f9616c35SEric Dumazet /* It is possible TX completion already happened 2230f9616c35SEric Dumazet * before we set TSQ_THROTTLED, so we must 2231f9616c35SEric Dumazet * test again the condition. 2232f9616c35SEric Dumazet */ 2233f9616c35SEric Dumazet smp_mb__after_atomic(); 223414afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) > limit) 2235f9616c35SEric Dumazet return true; 2236f9616c35SEric Dumazet } 2237f9616c35SEric Dumazet return false; 2238f9616c35SEric Dumazet } 2239f9616c35SEric Dumazet 224005b055e8SFrancis Yan static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new) 224105b055e8SFrancis Yan { 2242628174ccSEric Dumazet const u32 now = tcp_jiffies32; 2243efe967cdSArnd Bergmann enum tcp_chrono old = tp->chrono_type; 224405b055e8SFrancis Yan 2245efe967cdSArnd Bergmann if (old > TCP_CHRONO_UNSPEC) 2246efe967cdSArnd Bergmann tp->chrono_stat[old - 1] += now - tp->chrono_start; 224705b055e8SFrancis Yan tp->chrono_start = now; 224805b055e8SFrancis Yan tp->chrono_type = new; 224905b055e8SFrancis Yan } 225005b055e8SFrancis Yan 225105b055e8SFrancis Yan void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type) 225205b055e8SFrancis Yan { 225305b055e8SFrancis Yan struct tcp_sock *tp = tcp_sk(sk); 225405b055e8SFrancis Yan 225505b055e8SFrancis Yan /* If there are multiple conditions worthy of tracking in a 22560f87230dSFrancis Yan * chronograph then the highest priority enum takes precedence 22570f87230dSFrancis Yan * over the other conditions. So that if something "more interesting" 225805b055e8SFrancis Yan * starts happening, stop the previous chrono and start a new one. 225905b055e8SFrancis Yan */ 226005b055e8SFrancis Yan if (type > tp->chrono_type) 226105b055e8SFrancis Yan tcp_chrono_set(tp, type); 226205b055e8SFrancis Yan } 226305b055e8SFrancis Yan 226405b055e8SFrancis Yan void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type) 226505b055e8SFrancis Yan { 226605b055e8SFrancis Yan struct tcp_sock *tp = tcp_sk(sk); 226705b055e8SFrancis Yan 22680f87230dSFrancis Yan 22690f87230dSFrancis Yan /* There are multiple conditions worthy of tracking in a 22700f87230dSFrancis Yan * chronograph, so that the highest priority enum takes 22710f87230dSFrancis Yan * precedence over the other conditions (see tcp_chrono_start). 22720f87230dSFrancis Yan * If a condition stops, we only stop chrono tracking if 22730f87230dSFrancis Yan * it's the "most interesting" or current chrono we are 22740f87230dSFrancis Yan * tracking and starts busy chrono if we have pending data. 22750f87230dSFrancis Yan */ 227675c119afSEric Dumazet if (tcp_rtx_and_write_queues_empty(sk)) 227705b055e8SFrancis Yan tcp_chrono_set(tp, TCP_CHRONO_UNSPEC); 22780f87230dSFrancis Yan else if (type == tp->chrono_type) 22790f87230dSFrancis Yan tcp_chrono_set(tp, TCP_CHRONO_BUSY); 228005b055e8SFrancis Yan } 228105b055e8SFrancis Yan 22821da177e4SLinus Torvalds /* This routine writes packets to the network. It advances the 22831da177e4SLinus Torvalds * send_head. This happens as incoming acks open up the remote 22841da177e4SLinus Torvalds * window for us. 22851da177e4SLinus Torvalds * 2286f8269a49SIlpo Järvinen * LARGESEND note: !tcp_urg_mode is overkill, only frames between 2287f8269a49SIlpo Järvinen * snd_up-64k-mss .. snd_up cannot be large. However, taking into 2288f8269a49SIlpo Järvinen * account rare use of URG, this is not a big flaw. 2289f8269a49SIlpo Järvinen * 22906ba8a3b1SNandita Dukkipati * Send at most one packet when push_one > 0. Temporarily ignore 22916ba8a3b1SNandita Dukkipati * cwnd limit to force at most one packet out when push_one == 2. 22926ba8a3b1SNandita Dukkipati 2293a2a385d6SEric Dumazet * Returns true, if no segments are in flight and we have queued segments, 2294a2a385d6SEric Dumazet * but cannot send anything now because of SWS or another problem. 22951da177e4SLinus Torvalds */ 2296a2a385d6SEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 2297d5dd9175SIlpo Järvinen int push_one, gfp_t gfp) 22981da177e4SLinus Torvalds { 22991da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 230092df7b51SDavid S. Miller struct sk_buff *skb; 2301c1b4a7e6SDavid S. Miller unsigned int tso_segs, sent_pkts; 2302c1b4a7e6SDavid S. Miller int cwnd_quota; 23035d424d5aSJohn Heffner int result; 23045615f886SFrancis Yan bool is_cwnd_limited = false, is_rwnd_limited = false; 2305605ad7f1SEric Dumazet u32 max_segs; 23061da177e4SLinus Torvalds 2307c1b4a7e6SDavid S. Miller sent_pkts = 0; 23085d424d5aSJohn Heffner 2309ee1836aeSEric Dumazet tcp_mstamp_refresh(tp); 2310d5dd9175SIlpo Järvinen if (!push_one) { 23115d424d5aSJohn Heffner /* Do MTU probing. */ 2312d5dd9175SIlpo Järvinen result = tcp_mtu_probe(sk); 2313d5dd9175SIlpo Järvinen if (!result) { 2314a2a385d6SEric Dumazet return false; 23155d424d5aSJohn Heffner } else if (result > 0) { 23165d424d5aSJohn Heffner sent_pkts = 1; 23175d424d5aSJohn Heffner } 2318d5dd9175SIlpo Järvinen } 23195d424d5aSJohn Heffner 2320ed6e7268SNeal Cardwell max_segs = tcp_tso_segs(sk, mss_now); 2321fe067e8aSDavid S. Miller while ((skb = tcp_send_head(sk))) { 2322c8ac3774SHerbert Xu unsigned int limit; 2323c8ac3774SHerbert Xu 232479861919SEric Dumazet if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) { 232579861919SEric Dumazet /* "skb_mstamp_ns" is used as a start point for the retransmit timer */ 232679861919SEric Dumazet skb->skb_mstamp_ns = tp->tcp_wstamp_ns = tp->tcp_clock_cache; 232779861919SEric Dumazet list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue); 232879861919SEric Dumazet goto repair; /* Skip network transmission */ 232979861919SEric Dumazet } 233079861919SEric Dumazet 2331218af599SEric Dumazet if (tcp_pacing_check(sk)) 2332218af599SEric Dumazet break; 2333218af599SEric Dumazet 23345bbb432cSEric Dumazet tso_segs = tcp_init_tso_segs(skb, mss_now); 2335c1b4a7e6SDavid S. Miller BUG_ON(!tso_segs); 2336c1b4a7e6SDavid S. Miller 2337b68e9f85SHerbert Xu cwnd_quota = tcp_cwnd_test(tp, skb); 23386ba8a3b1SNandita Dukkipati if (!cwnd_quota) { 23396ba8a3b1SNandita Dukkipati if (push_one == 2) 23406ba8a3b1SNandita Dukkipati /* Force out a loss probe pkt. */ 23416ba8a3b1SNandita Dukkipati cwnd_quota = 1; 23426ba8a3b1SNandita Dukkipati else 2343b68e9f85SHerbert Xu break; 23446ba8a3b1SNandita Dukkipati } 2345b68e9f85SHerbert Xu 23465615f886SFrancis Yan if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) { 23475615f886SFrancis Yan is_rwnd_limited = true; 2348b68e9f85SHerbert Xu break; 23495615f886SFrancis Yan } 2350b68e9f85SHerbert Xu 2351d6a4e26aSEric Dumazet if (tso_segs == 1) { 2352aa93466bSDavid S. Miller if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 2353aa93466bSDavid S. Miller (tcp_skb_is_last(sk, skb) ? 2354aa93466bSDavid S. Miller nonagle : TCP_NAGLE_PUSH)))) 2355aa93466bSDavid S. Miller break; 2356c1b4a7e6SDavid S. Miller } else { 2357ca8a2263SNeal Cardwell if (!push_one && 2358605ad7f1SEric Dumazet tcp_tso_should_defer(sk, skb, &is_cwnd_limited, 235941727549SEric Dumazet max_segs)) { 236041727549SEric Dumazet if (!is_cwnd_limited) 236141727549SEric Dumazet is_rwnd_limited = true; 2362aa93466bSDavid S. Miller break; 2363c1b4a7e6SDavid S. Miller } 236441727549SEric Dumazet } 2365aa93466bSDavid S. Miller 2366605ad7f1SEric Dumazet limit = mss_now; 2367d6a4e26aSEric Dumazet if (tso_segs > 1 && !tcp_urg_mode(tp)) 2368605ad7f1SEric Dumazet limit = tcp_mss_split_point(sk, skb, mss_now, 2369605ad7f1SEric Dumazet min_t(unsigned int, 2370605ad7f1SEric Dumazet cwnd_quota, 2371605ad7f1SEric Dumazet max_segs), 2372605ad7f1SEric Dumazet nonagle); 2373605ad7f1SEric Dumazet 2374605ad7f1SEric Dumazet if (skb->len > limit && 237575c119afSEric Dumazet unlikely(tso_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE, 237675c119afSEric Dumazet skb, limit, mss_now, gfp))) 2377605ad7f1SEric Dumazet break; 2378605ad7f1SEric Dumazet 2379f9616c35SEric Dumazet if (tcp_small_queue_check(sk, skb, 0)) 238046d3ceabSEric Dumazet break; 2381c9eeec26SEric Dumazet 2382d5dd9175SIlpo Järvinen if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) 23831da177e4SLinus Torvalds break; 23841da177e4SLinus Torvalds 2385ec342325SAndrew Vagin repair: 23861da177e4SLinus Torvalds /* Advance the send_head. This one is sent out. 23871da177e4SLinus Torvalds * This call will increment packets_out. 23881da177e4SLinus Torvalds */ 238966f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 23901da177e4SLinus Torvalds 23911da177e4SLinus Torvalds tcp_minshall_update(tp, mss_now, skb); 2392a262f0cdSNandita Dukkipati sent_pkts += tcp_skb_pcount(skb); 2393d5dd9175SIlpo Järvinen 2394d5dd9175SIlpo Järvinen if (push_one) 2395d5dd9175SIlpo Järvinen break; 23961da177e4SLinus Torvalds } 23971da177e4SLinus Torvalds 23985615f886SFrancis Yan if (is_rwnd_limited) 23995615f886SFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED); 24005615f886SFrancis Yan else 24015615f886SFrancis Yan tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED); 24025615f886SFrancis Yan 2403aa93466bSDavid S. Miller if (likely(sent_pkts)) { 2404684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 2405684bad11SYuchung Cheng tp->prr_out += sent_pkts; 24066ba8a3b1SNandita Dukkipati 24076ba8a3b1SNandita Dukkipati /* Send one loss probe per tail loss episode. */ 24086ba8a3b1SNandita Dukkipati if (push_one != 2) 2409ed66dfafSNeal Cardwell tcp_schedule_loss_probe(sk, false); 2410d2e1339fSBendik Rønning Opstad is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd); 2411ca8a2263SNeal Cardwell tcp_cwnd_validate(sk, is_cwnd_limited); 2412a2a385d6SEric Dumazet return false; 24131da177e4SLinus Torvalds } 241475c119afSEric Dumazet return !tp->packets_out && !tcp_write_queue_empty(sk); 24156ba8a3b1SNandita Dukkipati } 24166ba8a3b1SNandita Dukkipati 2417ed66dfafSNeal Cardwell bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto) 24186ba8a3b1SNandita Dukkipati { 24196ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 24206ba8a3b1SNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 2421a2815817SNeal Cardwell u32 timeout, rto_delta_us; 24222ae21cf5SEric Dumazet int early_retrans; 24236ba8a3b1SNandita Dukkipati 24246ba8a3b1SNandita Dukkipati /* Don't do any loss probe on a Fast Open connection before 3WHS 24256ba8a3b1SNandita Dukkipati * finishes. 24266ba8a3b1SNandita Dukkipati */ 2427f9b99582SYuchung Cheng if (tp->fastopen_rsk) 24286ba8a3b1SNandita Dukkipati return false; 24296ba8a3b1SNandita Dukkipati 24302ae21cf5SEric Dumazet early_retrans = sock_net(sk)->ipv4.sysctl_tcp_early_retrans; 24316ba8a3b1SNandita Dukkipati /* Schedule a loss probe in 2*RTT for SACK capable connections 2432b4f70c3dSNeal Cardwell * not in loss recovery, that are either limited by cwnd or application. 24336ba8a3b1SNandita Dukkipati */ 24342ae21cf5SEric Dumazet if ((early_retrans != 3 && early_retrans != 4) || 2435bec41a11SYuchung Cheng !tp->packets_out || !tcp_is_sack(tp) || 2436b4f70c3dSNeal Cardwell (icsk->icsk_ca_state != TCP_CA_Open && 2437b4f70c3dSNeal Cardwell icsk->icsk_ca_state != TCP_CA_CWR)) 24386ba8a3b1SNandita Dukkipati return false; 24396ba8a3b1SNandita Dukkipati 2440bb4d991aSYuchung Cheng /* Probe timeout is 2*rtt. Add minimum RTO to account 2441f9b99582SYuchung Cheng * for delayed ack when there's one outstanding packet. If no RTT 2442f9b99582SYuchung Cheng * sample is available then probe after TCP_TIMEOUT_INIT. 24436ba8a3b1SNandita Dukkipati */ 2444bb4d991aSYuchung Cheng if (tp->srtt_us) { 2445bb4d991aSYuchung Cheng timeout = usecs_to_jiffies(tp->srtt_us >> 2); 24466ba8a3b1SNandita Dukkipati if (tp->packets_out == 1) 2447bb4d991aSYuchung Cheng timeout += TCP_RTO_MIN; 2448bb4d991aSYuchung Cheng else 2449bb4d991aSYuchung Cheng timeout += TCP_TIMEOUT_MIN; 2450bb4d991aSYuchung Cheng } else { 2451bb4d991aSYuchung Cheng timeout = TCP_TIMEOUT_INIT; 2452bb4d991aSYuchung Cheng } 24536ba8a3b1SNandita Dukkipati 2454a2815817SNeal Cardwell /* If the RTO formula yields an earlier time, then use that time. */ 2455ed66dfafSNeal Cardwell rto_delta_us = advancing_rto ? 2456ed66dfafSNeal Cardwell jiffies_to_usecs(inet_csk(sk)->icsk_rto) : 2457ed66dfafSNeal Cardwell tcp_rto_delta_us(sk); /* How far in future is RTO? */ 2458a2815817SNeal Cardwell if (rto_delta_us > 0) 2459a2815817SNeal Cardwell timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us)); 24606ba8a3b1SNandita Dukkipati 24613f80e08fSEric Dumazet tcp_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, 24623f80e08fSEric Dumazet TCP_RTO_MAX, NULL); 24636ba8a3b1SNandita Dukkipati return true; 24646ba8a3b1SNandita Dukkipati } 24656ba8a3b1SNandita Dukkipati 24661f3279aeSEric Dumazet /* Thanks to skb fast clones, we can detect if a prior transmit of 24671f3279aeSEric Dumazet * a packet is still in a qdisc or driver queue. 24681f3279aeSEric Dumazet * In this case, there is very little point doing a retransmit ! 24691f3279aeSEric Dumazet */ 24701f3279aeSEric Dumazet static bool skb_still_in_host_queue(const struct sock *sk, 24711f3279aeSEric Dumazet const struct sk_buff *skb) 24721f3279aeSEric Dumazet { 247339bb5e62SEric Dumazet if (unlikely(skb_fclone_busy(sk, skb))) { 2474c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), 24751f3279aeSEric Dumazet LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES); 24761f3279aeSEric Dumazet return true; 24771f3279aeSEric Dumazet } 24781f3279aeSEric Dumazet return false; 24791f3279aeSEric Dumazet } 24801f3279aeSEric Dumazet 2481b340b264SYuchung Cheng /* When probe timeout (PTO) fires, try send a new segment if possible, else 24826ba8a3b1SNandita Dukkipati * retransmit the last segment. 24836ba8a3b1SNandita Dukkipati */ 24846ba8a3b1SNandita Dukkipati void tcp_send_loss_probe(struct sock *sk) 24856ba8a3b1SNandita Dukkipati { 24869b717a8dSNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 24876ba8a3b1SNandita Dukkipati struct sk_buff *skb; 24886ba8a3b1SNandita Dukkipati int pcount; 24896ba8a3b1SNandita Dukkipati int mss = tcp_current_mss(sk); 24906ba8a3b1SNandita Dukkipati 2491b340b264SYuchung Cheng skb = tcp_send_head(sk); 249275c119afSEric Dumazet if (skb && tcp_snd_wnd_test(tp, skb, mss)) { 2493b340b264SYuchung Cheng pcount = tp->packets_out; 2494b340b264SYuchung Cheng tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC); 2495b340b264SYuchung Cheng if (tp->packets_out > pcount) 2496b340b264SYuchung Cheng goto probe_sent; 24976ba8a3b1SNandita Dukkipati goto rearm_timer; 24986ba8a3b1SNandita Dukkipati } 249975c119afSEric Dumazet skb = skb_rb_last(&sk->tcp_rtx_queue); 2500*b2b7af86SYuchung Cheng if (unlikely(!skb)) { 2501*b2b7af86SYuchung Cheng WARN_ONCE(tp->packets_out, 2502*b2b7af86SYuchung Cheng "invalid inflight: %u state %u cwnd %u mss %d\n", 2503*b2b7af86SYuchung Cheng tp->packets_out, sk->sk_state, tp->snd_cwnd, mss); 2504*b2b7af86SYuchung Cheng inet_csk(sk)->icsk_pending = 0; 2505*b2b7af86SYuchung Cheng return; 2506*b2b7af86SYuchung Cheng } 25076ba8a3b1SNandita Dukkipati 25089b717a8dSNandita Dukkipati /* At most one outstanding TLP retransmission. */ 25099b717a8dSNandita Dukkipati if (tp->tlp_high_seq) 25109b717a8dSNandita Dukkipati goto rearm_timer; 25119b717a8dSNandita Dukkipati 25121f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 25131f3279aeSEric Dumazet goto rearm_timer; 25141f3279aeSEric Dumazet 25156ba8a3b1SNandita Dukkipati pcount = tcp_skb_pcount(skb); 25166ba8a3b1SNandita Dukkipati if (WARN_ON(!pcount)) 25176ba8a3b1SNandita Dukkipati goto rearm_timer; 25186ba8a3b1SNandita Dukkipati 25196ba8a3b1SNandita Dukkipati if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) { 252075c119afSEric Dumazet if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, 252175c119afSEric Dumazet (pcount - 1) * mss, mss, 25226cc55e09SOctavian Purdila GFP_ATOMIC))) 25236ba8a3b1SNandita Dukkipati goto rearm_timer; 252475c119afSEric Dumazet skb = skb_rb_next(skb); 25256ba8a3b1SNandita Dukkipati } 25266ba8a3b1SNandita Dukkipati 25276ba8a3b1SNandita Dukkipati if (WARN_ON(!skb || !tcp_skb_pcount(skb))) 25286ba8a3b1SNandita Dukkipati goto rearm_timer; 25296ba8a3b1SNandita Dukkipati 253010d3be56SEric Dumazet if (__tcp_retransmit_skb(sk, skb, 1)) 2531b340b264SYuchung Cheng goto rearm_timer; 25326ba8a3b1SNandita Dukkipati 25339b717a8dSNandita Dukkipati /* Record snd_nxt for loss detection. */ 25349b717a8dSNandita Dukkipati tp->tlp_high_seq = tp->snd_nxt; 25359b717a8dSNandita Dukkipati 2536b340b264SYuchung Cheng probe_sent: 2537c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES); 2538fcd16c0aSYuchung Cheng /* Reset s.t. tcp_rearm_rto will restart timer from now */ 2539fcd16c0aSYuchung Cheng inet_csk(sk)->icsk_pending = 0; 2540b340b264SYuchung Cheng rearm_timer: 2541fcd16c0aSYuchung Cheng tcp_rearm_rto(sk); 25421da177e4SLinus Torvalds } 25431da177e4SLinus Torvalds 2544a762a980SDavid S. Miller /* Push out any pending frames which were held back due to 2545a762a980SDavid S. Miller * TCP_CORK or attempt at coalescing tiny packets. 2546a762a980SDavid S. Miller * The socket must be locked by the caller. 2547a762a980SDavid S. Miller */ 25489e412ba7SIlpo Järvinen void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 25499e412ba7SIlpo Järvinen int nonagle) 2550a762a980SDavid S. Miller { 2551726e07a8SIlpo Järvinen /* If we are closed, the bytes will have to remain here. 2552726e07a8SIlpo Järvinen * In time closedown will finish, we empty the write queue and 2553726e07a8SIlpo Järvinen * all will be happy. 2554726e07a8SIlpo Järvinen */ 2555726e07a8SIlpo Järvinen if (unlikely(sk->sk_state == TCP_CLOSE)) 2556726e07a8SIlpo Järvinen return; 2557726e07a8SIlpo Järvinen 255899a1dec7SMel Gorman if (tcp_write_xmit(sk, cur_mss, nonagle, 0, 25597450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC))) 25609e412ba7SIlpo Järvinen tcp_check_probe_timer(sk); 2561a762a980SDavid S. Miller } 2562a762a980SDavid S. Miller 2563c1b4a7e6SDavid S. Miller /* Send _single_ skb sitting at the send head. This function requires 2564c1b4a7e6SDavid S. Miller * true push pending frames to setup probe timer etc. 2565c1b4a7e6SDavid S. Miller */ 2566c1b4a7e6SDavid S. Miller void tcp_push_one(struct sock *sk, unsigned int mss_now) 2567c1b4a7e6SDavid S. Miller { 2568fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 2569c1b4a7e6SDavid S. Miller 2570c1b4a7e6SDavid S. Miller BUG_ON(!skb || skb->len < mss_now); 2571c1b4a7e6SDavid S. Miller 2572d5dd9175SIlpo Järvinen tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); 2573c1b4a7e6SDavid S. Miller } 2574c1b4a7e6SDavid S. Miller 25751da177e4SLinus Torvalds /* This function returns the amount that we can raise the 25761da177e4SLinus Torvalds * usable window based on the following constraints 25771da177e4SLinus Torvalds * 25781da177e4SLinus Torvalds * 1. The window can never be shrunk once it is offered (RFC 793) 25791da177e4SLinus Torvalds * 2. We limit memory per socket 25801da177e4SLinus Torvalds * 25811da177e4SLinus Torvalds * RFC 1122: 25821da177e4SLinus Torvalds * "the suggested [SWS] avoidance algorithm for the receiver is to keep 25831da177e4SLinus Torvalds * RECV.NEXT + RCV.WIN fixed until: 25841da177e4SLinus Torvalds * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 25851da177e4SLinus Torvalds * 25861da177e4SLinus Torvalds * i.e. don't raise the right edge of the window until you can raise 25871da177e4SLinus Torvalds * it at least MSS bytes. 25881da177e4SLinus Torvalds * 25891da177e4SLinus Torvalds * Unfortunately, the recommended algorithm breaks header prediction, 25901da177e4SLinus Torvalds * since header prediction assumes th->window stays fixed. 25911da177e4SLinus Torvalds * 25921da177e4SLinus Torvalds * Strictly speaking, keeping th->window fixed violates the receiver 25931da177e4SLinus Torvalds * side SWS prevention criteria. The problem is that under this rule 25941da177e4SLinus Torvalds * a stream of single byte packets will cause the right side of the 25951da177e4SLinus Torvalds * window to always advance by a single byte. 25961da177e4SLinus Torvalds * 25971da177e4SLinus Torvalds * Of course, if the sender implements sender side SWS prevention 25981da177e4SLinus Torvalds * then this will not be a problem. 25991da177e4SLinus Torvalds * 26001da177e4SLinus Torvalds * BSD seems to make the following compromise: 26011da177e4SLinus Torvalds * 26021da177e4SLinus Torvalds * If the free space is less than the 1/4 of the maximum 26031da177e4SLinus Torvalds * space available and the free space is less than 1/2 mss, 26041da177e4SLinus Torvalds * then set the window to 0. 26051da177e4SLinus Torvalds * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 26061da177e4SLinus Torvalds * Otherwise, just prevent the window from shrinking 26071da177e4SLinus Torvalds * and from being larger than the largest representable value. 26081da177e4SLinus Torvalds * 26091da177e4SLinus Torvalds * This prevents incremental opening of the window in the regime 26101da177e4SLinus Torvalds * where TCP is limited by the speed of the reader side taking 26111da177e4SLinus Torvalds * data out of the TCP receive queue. It does nothing about 26121da177e4SLinus Torvalds * those cases where the window is constrained on the sender side 26131da177e4SLinus Torvalds * because the pipeline is full. 26141da177e4SLinus Torvalds * 26151da177e4SLinus Torvalds * BSD also seems to "accidentally" limit itself to windows that are a 26161da177e4SLinus Torvalds * multiple of MSS, at least until the free space gets quite small. 26171da177e4SLinus Torvalds * This would appear to be a side effect of the mbuf implementation. 26181da177e4SLinus Torvalds * Combining these two algorithms results in the observed behavior 26191da177e4SLinus Torvalds * of having a fixed window size at almost all times. 26201da177e4SLinus Torvalds * 26211da177e4SLinus Torvalds * Below we obtain similar behavior by forcing the offered window to 26221da177e4SLinus Torvalds * a multiple of the mss when it is feasible to do so. 26231da177e4SLinus Torvalds * 26241da177e4SLinus Torvalds * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 26251da177e4SLinus Torvalds * Regular options like TIMESTAMP are taken into account. 26261da177e4SLinus Torvalds */ 26271da177e4SLinus Torvalds u32 __tcp_select_window(struct sock *sk) 26281da177e4SLinus Torvalds { 2629463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 26301da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2631caa20d9aSStephen Hemminger /* MSS for the peer's data. Previous versions used mss_clamp 26321da177e4SLinus Torvalds * here. I don't know if the value based on our guesses 26331da177e4SLinus Torvalds * of peer's MSS is better for the performance. It's more correct 26341da177e4SLinus Torvalds * but may be worse for the performance because of rcv_mss 26351da177e4SLinus Torvalds * fluctuations. --SAW 1998/11/1 26361da177e4SLinus Torvalds */ 2637463c84b9SArnaldo Carvalho de Melo int mss = icsk->icsk_ack.rcv_mss; 26381da177e4SLinus Torvalds int free_space = tcp_space(sk); 263986c1a045SFlorian Westphal int allowed_space = tcp_full_space(sk); 264086c1a045SFlorian Westphal int full_space = min_t(int, tp->window_clamp, allowed_space); 26411da177e4SLinus Torvalds int window; 26421da177e4SLinus Torvalds 264306425c30SEric Dumazet if (unlikely(mss > full_space)) { 26441da177e4SLinus Torvalds mss = full_space; 264506425c30SEric Dumazet if (mss <= 0) 264606425c30SEric Dumazet return 0; 264706425c30SEric Dumazet } 2648b92edbe0SEric Dumazet if (free_space < (full_space >> 1)) { 2649463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.quick = 0; 26501da177e4SLinus Torvalds 2651b8da51ebSEric Dumazet if (tcp_under_memory_pressure(sk)) 2652056834d9SIlpo Järvinen tp->rcv_ssthresh = min(tp->rcv_ssthresh, 2653056834d9SIlpo Järvinen 4U * tp->advmss); 26541da177e4SLinus Torvalds 265586c1a045SFlorian Westphal /* free_space might become our new window, make sure we don't 265686c1a045SFlorian Westphal * increase it due to wscale. 265786c1a045SFlorian Westphal */ 265886c1a045SFlorian Westphal free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale); 265986c1a045SFlorian Westphal 266086c1a045SFlorian Westphal /* if free space is less than mss estimate, or is below 1/16th 266186c1a045SFlorian Westphal * of the maximum allowed, try to move to zero-window, else 266286c1a045SFlorian Westphal * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and 266386c1a045SFlorian Westphal * new incoming data is dropped due to memory limits. 266486c1a045SFlorian Westphal * With large window, mss test triggers way too late in order 266586c1a045SFlorian Westphal * to announce zero window in time before rmem limit kicks in. 266686c1a045SFlorian Westphal */ 266786c1a045SFlorian Westphal if (free_space < (allowed_space >> 4) || free_space < mss) 26681da177e4SLinus Torvalds return 0; 26691da177e4SLinus Torvalds } 26701da177e4SLinus Torvalds 26711da177e4SLinus Torvalds if (free_space > tp->rcv_ssthresh) 26721da177e4SLinus Torvalds free_space = tp->rcv_ssthresh; 26731da177e4SLinus Torvalds 26741da177e4SLinus Torvalds /* Don't do rounding if we are using window scaling, since the 26751da177e4SLinus Torvalds * scaled window will not line up with the MSS boundary anyway. 26761da177e4SLinus Torvalds */ 26771da177e4SLinus Torvalds if (tp->rx_opt.rcv_wscale) { 26781da177e4SLinus Torvalds window = free_space; 26791da177e4SLinus Torvalds 26801da177e4SLinus Torvalds /* Advertise enough space so that it won't get scaled away. 26811da177e4SLinus Torvalds * Import case: prevent zero window announcement if 26821da177e4SLinus Torvalds * 1<<rcv_wscale > mss. 26831da177e4SLinus Torvalds */ 26841935299dSGao Feng window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale)); 26851da177e4SLinus Torvalds } else { 26861935299dSGao Feng window = tp->rcv_wnd; 26871da177e4SLinus Torvalds /* Get the largest window that is a nice multiple of mss. 26881da177e4SLinus Torvalds * Window clamp already applied above. 26891da177e4SLinus Torvalds * If our current window offering is within 1 mss of the 26901da177e4SLinus Torvalds * free space we just keep it. This prevents the divide 26911da177e4SLinus Torvalds * and multiply from happening most of the time. 26921da177e4SLinus Torvalds * We also don't do any window rounding when the free space 26931da177e4SLinus Torvalds * is too small. 26941da177e4SLinus Torvalds */ 26951da177e4SLinus Torvalds if (window <= free_space - mss || window > free_space) 26961935299dSGao Feng window = rounddown(free_space, mss); 269784565070SJohn Heffner else if (mss == full_space && 2698b92edbe0SEric Dumazet free_space > window + (full_space >> 1)) 269984565070SJohn Heffner window = free_space; 27001da177e4SLinus Torvalds } 27011da177e4SLinus Torvalds 27021da177e4SLinus Torvalds return window; 27031da177e4SLinus Torvalds } 27041da177e4SLinus Torvalds 2705cfea5a68SMartin KaFai Lau void tcp_skb_collapse_tstamp(struct sk_buff *skb, 2706082ac2d5SMartin KaFai Lau const struct sk_buff *next_skb) 2707082ac2d5SMartin KaFai Lau { 27080a2cf20cSSoheil Hassas Yeganeh if (unlikely(tcp_has_tx_tstamp(next_skb))) { 27090a2cf20cSSoheil Hassas Yeganeh const struct skb_shared_info *next_shinfo = 27100a2cf20cSSoheil Hassas Yeganeh skb_shinfo(next_skb); 2711082ac2d5SMartin KaFai Lau struct skb_shared_info *shinfo = skb_shinfo(skb); 2712082ac2d5SMartin KaFai Lau 27130a2cf20cSSoheil Hassas Yeganeh shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP; 2714082ac2d5SMartin KaFai Lau shinfo->tskey = next_shinfo->tskey; 27152de8023eSMartin KaFai Lau TCP_SKB_CB(skb)->txstamp_ack |= 27162de8023eSMartin KaFai Lau TCP_SKB_CB(next_skb)->txstamp_ack; 2717082ac2d5SMartin KaFai Lau } 2718082ac2d5SMartin KaFai Lau } 2719082ac2d5SMartin KaFai Lau 27204a17fc3aSIlpo Järvinen /* Collapses two adjacent SKB's during retransmission. */ 2721f8071cdeSEric Dumazet static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) 27221da177e4SLinus Torvalds { 27231da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 272475c119afSEric Dumazet struct sk_buff *next_skb = skb_rb_next(skb); 272513dde04fSWei Yongjun int next_skb_size; 27261da177e4SLinus Torvalds 2727058dc334SIlpo Järvinen next_skb_size = next_skb->len; 27281da177e4SLinus Torvalds 2729058dc334SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); 27301da177e4SLinus Torvalds 2731f8071cdeSEric Dumazet if (next_skb_size) { 2732f8071cdeSEric Dumazet if (next_skb_size <= skb_availroom(skb)) 2733f8071cdeSEric Dumazet skb_copy_bits(next_skb, 0, skb_put(skb, next_skb_size), 2734f8071cdeSEric Dumazet next_skb_size); 2735f8071cdeSEric Dumazet else if (!skb_shift(skb, next_skb, next_skb_size)) 2736f8071cdeSEric Dumazet return false; 2737f8071cdeSEric Dumazet } 27382b7cda9cSEric Dumazet tcp_highest_sack_replace(sk, next_skb, skb); 2739a6963a6bSIlpo Järvinen 27401da177e4SLinus Torvalds /* Update sequence range on original skb. */ 27411da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 27421da177e4SLinus Torvalds 2743e6c7d085SIlpo Järvinen /* Merge over control information. This moves PSH/FIN etc. over */ 27444de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; 27451da177e4SLinus Torvalds 27461da177e4SLinus Torvalds /* All done, get rid of second SKB and account for it so 27471da177e4SLinus Torvalds * packet counting does not break. 27481da177e4SLinus Torvalds */ 27494828e7f4SIlpo Järvinen TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; 2750a643b5d4SMartin KaFai Lau TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor; 2751b7689205SIlpo Järvinen 2752b7689205SIlpo Järvinen /* changed transmit queue under us so clear hints */ 2753ef9da47cSIlpo Järvinen tcp_clear_retrans_hints_partial(tp); 2754ef9da47cSIlpo Järvinen if (next_skb == tp->retransmit_skb_hint) 2755ef9da47cSIlpo Järvinen tp->retransmit_skb_hint = skb; 2756b7689205SIlpo Järvinen 2757797108d1SIlpo Järvinen tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); 2758797108d1SIlpo Järvinen 2759082ac2d5SMartin KaFai Lau tcp_skb_collapse_tstamp(skb, next_skb); 2760082ac2d5SMartin KaFai Lau 276175c119afSEric Dumazet tcp_rtx_queue_unlink_and_free(next_skb, sk); 2762f8071cdeSEric Dumazet return true; 27631da177e4SLinus Torvalds } 27641da177e4SLinus Torvalds 276567edfef7SAndi Kleen /* Check if coalescing SKBs is legal. */ 2766a2a385d6SEric Dumazet static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) 27674a17fc3aSIlpo Järvinen { 27684a17fc3aSIlpo Järvinen if (tcp_skb_pcount(skb) > 1) 2769a2a385d6SEric Dumazet return false; 27704a17fc3aSIlpo Järvinen if (skb_cloned(skb)) 2771a2a385d6SEric Dumazet return false; 27722331ccc5SEric Dumazet /* Some heuristics for collapsing over SACK'd could be invented */ 27734a17fc3aSIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 2774a2a385d6SEric Dumazet return false; 27754a17fc3aSIlpo Järvinen 2776a2a385d6SEric Dumazet return true; 27774a17fc3aSIlpo Järvinen } 27784a17fc3aSIlpo Järvinen 277967edfef7SAndi Kleen /* Collapse packets in the retransmit queue to make to create 278067edfef7SAndi Kleen * less packets on the wire. This is only done on retransmission. 278167edfef7SAndi Kleen */ 27824a17fc3aSIlpo Järvinen static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, 27834a17fc3aSIlpo Järvinen int space) 27844a17fc3aSIlpo Järvinen { 27854a17fc3aSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 27864a17fc3aSIlpo Järvinen struct sk_buff *skb = to, *tmp; 2787a2a385d6SEric Dumazet bool first = true; 27884a17fc3aSIlpo Järvinen 2789e0a1e5b5SEric Dumazet if (!sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse) 27904a17fc3aSIlpo Järvinen return; 27914de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 27924a17fc3aSIlpo Järvinen return; 27934a17fc3aSIlpo Järvinen 279475c119afSEric Dumazet skb_rbtree_walk_from_safe(skb, tmp) { 27954a17fc3aSIlpo Järvinen if (!tcp_can_collapse(sk, skb)) 27964a17fc3aSIlpo Järvinen break; 27974a17fc3aSIlpo Järvinen 2798a643b5d4SMartin KaFai Lau if (!tcp_skb_can_collapse_to(to)) 2799a643b5d4SMartin KaFai Lau break; 2800a643b5d4SMartin KaFai Lau 28014a17fc3aSIlpo Järvinen space -= skb->len; 28024a17fc3aSIlpo Järvinen 28034a17fc3aSIlpo Järvinen if (first) { 2804a2a385d6SEric Dumazet first = false; 28054a17fc3aSIlpo Järvinen continue; 28064a17fc3aSIlpo Järvinen } 28074a17fc3aSIlpo Järvinen 28084a17fc3aSIlpo Järvinen if (space < 0) 28094a17fc3aSIlpo Järvinen break; 28104a17fc3aSIlpo Järvinen 28114a17fc3aSIlpo Järvinen if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) 28124a17fc3aSIlpo Järvinen break; 28134a17fc3aSIlpo Järvinen 2814f8071cdeSEric Dumazet if (!tcp_collapse_retrans(sk, to)) 2815f8071cdeSEric Dumazet break; 28164a17fc3aSIlpo Järvinen } 28174a17fc3aSIlpo Järvinen } 28184a17fc3aSIlpo Järvinen 28191da177e4SLinus Torvalds /* This retransmits one SKB. Policy decisions and retransmit queue 28201da177e4SLinus Torvalds * state updates are done by the caller. Returns non-zero if an 28211da177e4SLinus Torvalds * error occurred which prevented the send. 28221da177e4SLinus Torvalds */ 282310d3be56SEric Dumazet int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) 28241da177e4SLinus Torvalds { 28255d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 282610d3be56SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 28277d227cd2SSridhar Samudrala unsigned int cur_mss; 282810d3be56SEric Dumazet int diff, len, err; 28291da177e4SLinus Torvalds 283010d3be56SEric Dumazet 283110d3be56SEric Dumazet /* Inconclusive MTU probe */ 283210d3be56SEric Dumazet if (icsk->icsk_mtup.probe_size) 28335d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 28345d424d5aSJohn Heffner 28351da177e4SLinus Torvalds /* Do not sent more than we queued. 1/4 is reserved for possible 2836caa20d9aSStephen Hemminger * copying overhead: fragmentation, tunneling, mangling etc. 28371da177e4SLinus Torvalds */ 283814afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) > 2839ffb4d6c8SEric Dumazet min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), 2840ffb4d6c8SEric Dumazet sk->sk_sndbuf)) 28411da177e4SLinus Torvalds return -EAGAIN; 28421da177e4SLinus Torvalds 28431f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 28441f3279aeSEric Dumazet return -EBUSY; 28451f3279aeSEric Dumazet 28461da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 28477f582b24SEric Dumazet if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) { 28487f582b24SEric Dumazet WARN_ON_ONCE(1); 28497f582b24SEric Dumazet return -EINVAL; 28507f582b24SEric Dumazet } 28511da177e4SLinus Torvalds if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 28521da177e4SLinus Torvalds return -ENOMEM; 28531da177e4SLinus Torvalds } 28541da177e4SLinus Torvalds 28557d227cd2SSridhar Samudrala if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 28567d227cd2SSridhar Samudrala return -EHOSTUNREACH; /* Routing failure or similar. */ 28577d227cd2SSridhar Samudrala 28580c54b85fSIlpo Järvinen cur_mss = tcp_current_mss(sk); 28597d227cd2SSridhar Samudrala 28601da177e4SLinus Torvalds /* If receiver has shrunk his window, and skb is out of 28611da177e4SLinus Torvalds * new window, do not retransmit it. The exception is the 28621da177e4SLinus Torvalds * case, when window is shrunk to zero. In this case 28631da177e4SLinus Torvalds * our retransmit serves as a zero window probe. 28641da177e4SLinus Torvalds */ 28659d4fb27dSJoe Perches if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && 28669d4fb27dSJoe Perches TCP_SKB_CB(skb)->seq != tp->snd_una) 28671da177e4SLinus Torvalds return -EAGAIN; 28681da177e4SLinus Torvalds 286910d3be56SEric Dumazet len = cur_mss * segs; 287010d3be56SEric Dumazet if (skb->len > len) { 287175c119afSEric Dumazet if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len, 287275c119afSEric Dumazet cur_mss, GFP_ATOMIC)) 28731da177e4SLinus Torvalds return -ENOMEM; /* We'll try again later. */ 287402276f3cSIlpo Järvinen } else { 2875c52e2421SEric Dumazet if (skb_unclone(skb, GFP_ATOMIC)) 2876c52e2421SEric Dumazet return -ENOMEM; 287710d3be56SEric Dumazet 287810d3be56SEric Dumazet diff = tcp_skb_pcount(skb); 287910d3be56SEric Dumazet tcp_set_skb_tso_segs(skb, cur_mss); 288010d3be56SEric Dumazet diff -= tcp_skb_pcount(skb); 288110d3be56SEric Dumazet if (diff) 288210d3be56SEric Dumazet tcp_adjust_pcount(sk, skb, diff); 288310d3be56SEric Dumazet if (skb->len < cur_mss) 288410d3be56SEric Dumazet tcp_retrans_try_collapse(sk, skb, cur_mss); 28851da177e4SLinus Torvalds } 28861da177e4SLinus Torvalds 288749213555SDaniel Borkmann /* RFC3168, section 6.1.1.1. ECN fallback */ 288849213555SDaniel Borkmann if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN) 288949213555SDaniel Borkmann tcp_ecn_clear_syn(sk, skb); 289049213555SDaniel Borkmann 2891678550c6SYuchung Cheng /* Update global and local TCP statistics. */ 2892678550c6SYuchung Cheng segs = tcp_skb_pcount(skb); 2893678550c6SYuchung Cheng TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs); 2894678550c6SYuchung Cheng if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 2895678550c6SYuchung Cheng __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 2896678550c6SYuchung Cheng tp->total_retrans += segs; 2897fb31c9b9SWei Wang tp->bytes_retrans += skb->len; 2898678550c6SYuchung Cheng 289950bceae9SThomas Graf /* make sure skb->data is aligned on arches that require it 290050bceae9SThomas Graf * and check if ack-trimming & collapsing extended the headroom 290150bceae9SThomas Graf * beyond what csum_start can cover. 290250bceae9SThomas Graf */ 290350bceae9SThomas Graf if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) || 290450bceae9SThomas Graf skb_headroom(skb) >= 0xFFFF)) { 290510a81980SEric Dumazet struct sk_buff *nskb; 290610a81980SEric Dumazet 2907e2080072SEric Dumazet tcp_skb_tsorted_save(skb) { 290810a81980SEric Dumazet nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC); 2909c84a5711SYuchung Cheng err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) : 2910117632e6SEric Dumazet -ENOBUFS; 2911e2080072SEric Dumazet } tcp_skb_tsorted_restore(skb); 2912e2080072SEric Dumazet 29135889e2c0SYousuk Seung if (!err) { 2914a7a25630SEric Dumazet tcp_update_skb_after_send(sk, skb, tp->tcp_wstamp_ns); 29155889e2c0SYousuk Seung tcp_rate_skb_sent(sk, skb); 29165889e2c0SYousuk Seung } 2917117632e6SEric Dumazet } else { 2918c84a5711SYuchung Cheng err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 2919117632e6SEric Dumazet } 2920c84a5711SYuchung Cheng 2921a31ad29eSLawrence Brakmo if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RETRANS_CB_FLAG)) 2922a31ad29eSLawrence Brakmo tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RETRANS_CB, 2923a31ad29eSLawrence Brakmo TCP_SKB_CB(skb)->seq, segs, err); 2924a31ad29eSLawrence Brakmo 2925fc9f3501SEric Dumazet if (likely(!err)) { 2926c84a5711SYuchung Cheng TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS; 2927e086101bSCong Wang trace_tcp_retransmit_skb(sk, skb); 2928678550c6SYuchung Cheng } else if (err != -EBUSY) { 2929ec641b39SYuchung Cheng NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL, segs); 2930fc9f3501SEric Dumazet } 2931c84a5711SYuchung Cheng return err; 293293b174adSYuchung Cheng } 293393b174adSYuchung Cheng 293410d3be56SEric Dumazet int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) 293593b174adSYuchung Cheng { 293693b174adSYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 293710d3be56SEric Dumazet int err = __tcp_retransmit_skb(sk, skb, segs); 29381da177e4SLinus Torvalds 29391da177e4SLinus Torvalds if (err == 0) { 29401da177e4SLinus Torvalds #if FASTRETRANS_DEBUG > 0 29411da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 2942e87cc472SJoe Perches net_dbg_ratelimited("retrans_out leaked\n"); 29431da177e4SLinus Torvalds } 29441da177e4SLinus Torvalds #endif 29451da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 29461da177e4SLinus Torvalds tp->retrans_out += tcp_skb_pcount(skb); 29471da177e4SLinus Torvalds 29481da177e4SLinus Torvalds /* Save stamp of the first retransmit. */ 29491da177e4SLinus Torvalds if (!tp->retrans_stamp) 29507faee5c0SEric Dumazet tp->retrans_stamp = tcp_skb_timestamp(skb); 29511da177e4SLinus Torvalds 29521da177e4SLinus Torvalds } 29536e08d5e3SYuchung Cheng 29546e08d5e3SYuchung Cheng if (tp->undo_retrans < 0) 29556e08d5e3SYuchung Cheng tp->undo_retrans = 0; 29566e08d5e3SYuchung Cheng tp->undo_retrans += tcp_skb_pcount(skb); 29571da177e4SLinus Torvalds return err; 29581da177e4SLinus Torvalds } 29591da177e4SLinus Torvalds 29601da177e4SLinus Torvalds /* This gets called after a retransmit timeout, and the initially 29611da177e4SLinus Torvalds * retransmitted data is acknowledged. It tries to continue 29621da177e4SLinus Torvalds * resending the rest of the retransmit queue, until either 29631da177e4SLinus Torvalds * we've sent it all or the congestion window limit is reached. 29641da177e4SLinus Torvalds */ 29651da177e4SLinus Torvalds void tcp_xmit_retransmit_queue(struct sock *sk) 29661da177e4SLinus Torvalds { 29676687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 2968b9f1f1ceSEric Dumazet struct sk_buff *skb, *rtx_head, *hole = NULL; 29691da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2970840a3cbeSYuchung Cheng u32 max_segs; 297161eb55f4SIlpo Järvinen int mib_idx; 29726a438bbeSStephen Hemminger 297345e77d31SIlpo Järvinen if (!tp->packets_out) 297445e77d31SIlpo Järvinen return; 297545e77d31SIlpo Järvinen 297675c119afSEric Dumazet rtx_head = tcp_rtx_queue_head(sk); 2977b9f1f1ceSEric Dumazet skb = tp->retransmit_skb_hint ?: rtx_head; 2978ed6e7268SNeal Cardwell max_segs = tcp_tso_segs(sk, tcp_current_mss(sk)); 297975c119afSEric Dumazet skb_rbtree_walk_from(skb) { 2980dca0aaf8SEric Dumazet __u8 sacked; 298110d3be56SEric Dumazet int segs; 29821da177e4SLinus Torvalds 2983218af599SEric Dumazet if (tcp_pacing_check(sk)) 2984218af599SEric Dumazet break; 2985218af599SEric Dumazet 29866a438bbeSStephen Hemminger /* we could do better than to assign each time */ 298751456b29SIan Morris if (!hole) 29886a438bbeSStephen Hemminger tp->retransmit_skb_hint = skb; 29896a438bbeSStephen Hemminger 299010d3be56SEric Dumazet segs = tp->snd_cwnd - tcp_packets_in_flight(tp); 299110d3be56SEric Dumazet if (segs <= 0) 29921da177e4SLinus Torvalds return; 2993dca0aaf8SEric Dumazet sacked = TCP_SKB_CB(skb)->sacked; 2994a3d2e9f8SEric Dumazet /* In case tcp_shift_skb_data() have aggregated large skbs, 2995a3d2e9f8SEric Dumazet * we need to make sure not sending too bigs TSO packets 2996a3d2e9f8SEric Dumazet */ 2997a3d2e9f8SEric Dumazet segs = min_t(int, segs, max_segs); 29980e1c54c2SIlpo Järvinen 2999840a3cbeSYuchung Cheng if (tp->retrans_out >= tp->lost_out) { 3000006f582cSIlpo Järvinen break; 30010e1c54c2SIlpo Järvinen } else if (!(sacked & TCPCB_LOST)) { 300251456b29SIan Morris if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) 30030e1c54c2SIlpo Järvinen hole = skb; 300461eb55f4SIlpo Järvinen continue; 30051da177e4SLinus Torvalds 30060e1c54c2SIlpo Järvinen } else { 30070e1c54c2SIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Loss) 30080e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFASTRETRANS; 30090e1c54c2SIlpo Järvinen else 30100e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; 30110e1c54c2SIlpo Järvinen } 30120e1c54c2SIlpo Järvinen 30130e1c54c2SIlpo Järvinen if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) 301461eb55f4SIlpo Järvinen continue; 301540b215e5SPavel Emelyanov 3016f9616c35SEric Dumazet if (tcp_small_queue_check(sk, skb, 1)) 3017f9616c35SEric Dumazet return; 3018f9616c35SEric Dumazet 301910d3be56SEric Dumazet if (tcp_retransmit_skb(sk, skb, segs)) 30201da177e4SLinus Torvalds return; 302124ab6becSYuchung Cheng 3022de1d6578SYuchung Cheng NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb)); 30231da177e4SLinus Torvalds 3024684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 3025a262f0cdSNandita Dukkipati tp->prr_out += tcp_skb_pcount(skb); 3026a262f0cdSNandita Dukkipati 302775c119afSEric Dumazet if (skb == rtx_head && 302857dde7f7SYuchung Cheng icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT) 30293f80e08fSEric Dumazet tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 30303f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, 30313f80e08fSEric Dumazet TCP_RTO_MAX, 30323f80e08fSEric Dumazet skb); 30331da177e4SLinus Torvalds } 30341da177e4SLinus Torvalds } 30351da177e4SLinus Torvalds 3036d83769a5SEric Dumazet /* We allow to exceed memory limits for FIN packets to expedite 3037d83769a5SEric Dumazet * connection tear down and (memory) recovery. 3038845704a5SEric Dumazet * Otherwise tcp_send_fin() could be tempted to either delay FIN 3039845704a5SEric Dumazet * or even be forced to close flow without any FIN. 3040a6c5ea4cSEric Dumazet * In general, we want to allow one skb per socket to avoid hangs 3041a6c5ea4cSEric Dumazet * with edge trigger epoll() 3042d83769a5SEric Dumazet */ 3043a6c5ea4cSEric Dumazet void sk_forced_mem_schedule(struct sock *sk, int size) 3044d83769a5SEric Dumazet { 3045e805605cSJohannes Weiner int amt; 3046d83769a5SEric Dumazet 3047d83769a5SEric Dumazet if (size <= sk->sk_forward_alloc) 3048d83769a5SEric Dumazet return; 3049d83769a5SEric Dumazet amt = sk_mem_pages(size); 3050d83769a5SEric Dumazet sk->sk_forward_alloc += amt * SK_MEM_QUANTUM; 3051e805605cSJohannes Weiner sk_memory_allocated_add(sk, amt); 3052e805605cSJohannes Weiner 3053baac50bbSJohannes Weiner if (mem_cgroup_sockets_enabled && sk->sk_memcg) 3054baac50bbSJohannes Weiner mem_cgroup_charge_skmem(sk->sk_memcg, amt); 3055d83769a5SEric Dumazet } 3056d83769a5SEric Dumazet 3057845704a5SEric Dumazet /* Send a FIN. The caller locks the socket for us. 3058845704a5SEric Dumazet * We should try to send a FIN packet really hard, but eventually give up. 30591da177e4SLinus Torvalds */ 30601da177e4SLinus Torvalds void tcp_send_fin(struct sock *sk) 30611da177e4SLinus Torvalds { 3062845704a5SEric Dumazet struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk); 30631da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 30641da177e4SLinus Torvalds 3065845704a5SEric Dumazet /* Optimization, tack on the FIN if we have one skb in write queue and 3066845704a5SEric Dumazet * this skb was not yet sent, or we are under memory pressure. 3067845704a5SEric Dumazet * Note: in the latter case, FIN packet will be sent after a timeout, 3068845704a5SEric Dumazet * as TCP stack thinks it has already been transmitted. 30691da177e4SLinus Torvalds */ 307075c119afSEric Dumazet if (!tskb && tcp_under_memory_pressure(sk)) 307175c119afSEric Dumazet tskb = skb_rb_last(&sk->tcp_rtx_queue); 307275c119afSEric Dumazet 307375c119afSEric Dumazet if (tskb) { 3074845704a5SEric Dumazet coalesce: 3075845704a5SEric Dumazet TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN; 3076845704a5SEric Dumazet TCP_SKB_CB(tskb)->end_seq++; 30771da177e4SLinus Torvalds tp->write_seq++; 307875c119afSEric Dumazet if (tcp_write_queue_empty(sk)) { 3079845704a5SEric Dumazet /* This means tskb was already sent. 3080845704a5SEric Dumazet * Pretend we included the FIN on previous transmit. 3081845704a5SEric Dumazet * We need to set tp->snd_nxt to the value it would have 3082845704a5SEric Dumazet * if FIN had been sent. This is because retransmit path 3083845704a5SEric Dumazet * does not change tp->snd_nxt. 3084845704a5SEric Dumazet */ 3085845704a5SEric Dumazet tp->snd_nxt++; 3086845704a5SEric Dumazet return; 3087845704a5SEric Dumazet } 30881da177e4SLinus Torvalds } else { 3089845704a5SEric Dumazet skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation); 3090845704a5SEric Dumazet if (unlikely(!skb)) { 3091845704a5SEric Dumazet if (tskb) 3092845704a5SEric Dumazet goto coalesce; 3093845704a5SEric Dumazet return; 30941da177e4SLinus Torvalds } 3095e2080072SEric Dumazet INIT_LIST_HEAD(&skb->tcp_tsorted_anchor); 3096d83769a5SEric Dumazet skb_reserve(skb, MAX_TCP_HEADER); 3097a6c5ea4cSEric Dumazet sk_forced_mem_schedule(sk, skb->truesize); 30981da177e4SLinus Torvalds /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 3099e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tp->write_seq, 3100a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_FIN); 31011da177e4SLinus Torvalds tcp_queue_skb(sk, skb); 31021da177e4SLinus Torvalds } 3103845704a5SEric Dumazet __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF); 31041da177e4SLinus Torvalds } 31051da177e4SLinus Torvalds 31061da177e4SLinus Torvalds /* We get here when a process closes a file descriptor (either due to 31071da177e4SLinus Torvalds * an explicit close() or as a byproduct of exit()'ing) and there 31081da177e4SLinus Torvalds * was unread data in the receive queue. This behavior is recommended 310965bb723cSGerrit Renker * by RFC 2525, section 2.17. -DaveM 31101da177e4SLinus Torvalds */ 3111dd0fc66fSAl Viro void tcp_send_active_reset(struct sock *sk, gfp_t priority) 31121da177e4SLinus Torvalds { 31131da177e4SLinus Torvalds struct sk_buff *skb; 31141da177e4SLinus Torvalds 31157cc2b043SGao Feng TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); 31167cc2b043SGao Feng 31171da177e4SLinus Torvalds /* NOTE: No TCP options attached and we never retransmit this. */ 31181da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, priority); 31191da177e4SLinus Torvalds if (!skb) { 31204e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 31211da177e4SLinus Torvalds return; 31221da177e4SLinus Torvalds } 31231da177e4SLinus Torvalds 31241da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 31251da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 3126e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), 3127a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_RST); 31289a568de4SEric Dumazet tcp_mstamp_refresh(tcp_sk(sk)); 31291da177e4SLinus Torvalds /* Send it off. */ 3130dfb4b9dcSDavid S. Miller if (tcp_transmit_skb(sk, skb, 0, priority)) 31314e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 3132c24b14c4SSong Liu 3133c24b14c4SSong Liu /* skb of trace_tcp_send_reset() keeps the skb that caused RST, 3134c24b14c4SSong Liu * skb here is different to the troublesome skb, so use NULL 3135c24b14c4SSong Liu */ 3136c24b14c4SSong Liu trace_tcp_send_reset(sk, NULL); 31371da177e4SLinus Torvalds } 31381da177e4SLinus Torvalds 313967edfef7SAndi Kleen /* Send a crossed SYN-ACK during socket establishment. 314067edfef7SAndi Kleen * WARNING: This routine must only be called when we have already sent 31411da177e4SLinus Torvalds * a SYN packet that crossed the incoming SYN that caused this routine 31421da177e4SLinus Torvalds * to get called. If this assumption fails then the initial rcv_wnd 31431da177e4SLinus Torvalds * and rcv_wscale values will not be correct. 31441da177e4SLinus Torvalds */ 31451da177e4SLinus Torvalds int tcp_send_synack(struct sock *sk) 31461da177e4SLinus Torvalds { 31471da177e4SLinus Torvalds struct sk_buff *skb; 31481da177e4SLinus Torvalds 314975c119afSEric Dumazet skb = tcp_rtx_queue_head(sk); 315051456b29SIan Morris if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 315175c119afSEric Dumazet pr_err("%s: wrong queue state\n", __func__); 31521da177e4SLinus Torvalds return -EFAULT; 31531da177e4SLinus Torvalds } 31544de075e0SEric Dumazet if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { 31551da177e4SLinus Torvalds if (skb_cloned(skb)) { 3156e2080072SEric Dumazet struct sk_buff *nskb; 3157e2080072SEric Dumazet 3158e2080072SEric Dumazet tcp_skb_tsorted_save(skb) { 3159e2080072SEric Dumazet nskb = skb_copy(skb, GFP_ATOMIC); 3160e2080072SEric Dumazet } tcp_skb_tsorted_restore(skb); 316151456b29SIan Morris if (!nskb) 31621da177e4SLinus Torvalds return -ENOMEM; 3163e2080072SEric Dumazet INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor); 316475c119afSEric Dumazet tcp_rtx_queue_unlink_and_free(skb, sk); 3165f4a775d1SEric Dumazet __skb_header_release(nskb); 316675c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb); 31673ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 31683ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 31691da177e4SLinus Torvalds skb = nskb; 31701da177e4SLinus Torvalds } 31711da177e4SLinus Torvalds 31724de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; 3173735d3831SFlorian Westphal tcp_ecn_send_synack(sk, skb); 31741da177e4SLinus Torvalds } 3175dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 31761da177e4SLinus Torvalds } 31771da177e4SLinus Torvalds 31784aea39c1SEric Dumazet /** 31794aea39c1SEric Dumazet * tcp_make_synack - Prepare a SYN-ACK. 31804aea39c1SEric Dumazet * sk: listener socket 31814aea39c1SEric Dumazet * dst: dst entry attached to the SYNACK 31824aea39c1SEric Dumazet * req: request_sock pointer 31834aea39c1SEric Dumazet * 31844aea39c1SEric Dumazet * Allocate one skb and build a SYNACK packet. 31854aea39c1SEric Dumazet * @dst is consumed : Caller should not use it again. 31864aea39c1SEric Dumazet */ 31875d062de7SEric Dumazet struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst, 3188e6b4d113SWilliam Allen Simpson struct request_sock *req, 3189ca6fb065SEric Dumazet struct tcp_fastopen_cookie *foc, 3190b3d05147SEric Dumazet enum tcp_synack_type synack_type) 31911da177e4SLinus Torvalds { 31922e6599cbSArnaldo Carvalho de Melo struct inet_request_sock *ireq = inet_rsk(req); 31935d062de7SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 319480f03e27SEric Dumazet struct tcp_md5sig_key *md5 = NULL; 31955d062de7SEric Dumazet struct tcp_out_options opts; 31965d062de7SEric Dumazet struct sk_buff *skb; 3197bd0388aeSWilliam Allen Simpson int tcp_header_size; 31985d062de7SEric Dumazet struct tcphdr *th; 3199f5fff5dcSTom Quetchenbach int mss; 32001da177e4SLinus Torvalds 3201ca6fb065SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 32024aea39c1SEric Dumazet if (unlikely(!skb)) { 32034aea39c1SEric Dumazet dst_release(dst); 32041da177e4SLinus Torvalds return NULL; 32054aea39c1SEric Dumazet } 32061da177e4SLinus Torvalds /* Reserve space for headers. */ 32071da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 32081da177e4SLinus Torvalds 3209b3d05147SEric Dumazet switch (synack_type) { 3210b3d05147SEric Dumazet case TCP_SYNACK_NORMAL: 32119e17f8a4SEric Dumazet skb_set_owner_w(skb, req_to_sk(req)); 3212b3d05147SEric Dumazet break; 3213b3d05147SEric Dumazet case TCP_SYNACK_COOKIE: 3214b3d05147SEric Dumazet /* Under synflood, we do not attach skb to a socket, 3215b3d05147SEric Dumazet * to avoid false sharing. 3216b3d05147SEric Dumazet */ 3217b3d05147SEric Dumazet break; 3218b3d05147SEric Dumazet case TCP_SYNACK_FASTOPEN: 3219ca6fb065SEric Dumazet /* sk is a const pointer, because we want to express multiple 3220ca6fb065SEric Dumazet * cpu might call us concurrently. 3221ca6fb065SEric Dumazet * sk->sk_wmem_alloc in an atomic, we can promote to rw. 3222ca6fb065SEric Dumazet */ 3223ca6fb065SEric Dumazet skb_set_owner_w(skb, (struct sock *)sk); 3224b3d05147SEric Dumazet break; 3225ca6fb065SEric Dumazet } 32264aea39c1SEric Dumazet skb_dst_set(skb, dst); 32271da177e4SLinus Torvalds 32283541f9e8SEric Dumazet mss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 3229f5fff5dcSTom Quetchenbach 323033ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 32318b5f12d0SFlorian Westphal #ifdef CONFIG_SYN_COOKIES 32328b5f12d0SFlorian Westphal if (unlikely(req->cookie_ts)) 3233d3edd06eSEric Dumazet skb->skb_mstamp_ns = cookie_init_timestamp(req); 32348b5f12d0SFlorian Westphal else 32358b5f12d0SFlorian Westphal #endif 3236d3edd06eSEric Dumazet skb->skb_mstamp_ns = tcp_clock_ns(); 323780f03e27SEric Dumazet 323880f03e27SEric Dumazet #ifdef CONFIG_TCP_MD5SIG 323980f03e27SEric Dumazet rcu_read_lock(); 3240fd3a154aSEric Dumazet md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req)); 324180f03e27SEric Dumazet #endif 324258d607d3SEric Dumazet skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4); 324360e2a778SUrsula Braun tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5, 324460e2a778SUrsula Braun foc) + sizeof(*th); 324533ad798cSAdam Langley 3246aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 3247aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 32481da177e4SLinus Torvalds 3249ea1627c2SEric Dumazet th = (struct tcphdr *)skb->data; 32501da177e4SLinus Torvalds memset(th, 0, sizeof(struct tcphdr)); 32511da177e4SLinus Torvalds th->syn = 1; 32521da177e4SLinus Torvalds th->ack = 1; 32536ac705b1SEric Dumazet tcp_ecn_make_synack(req, th); 3254b44084c2SEric Dumazet th->source = htons(ireq->ir_num); 3255634fb979SEric Dumazet th->dest = ireq->ir_rmt_port; 3256e05a90ecSJamal Hadi Salim skb->mark = ireq->ir_mark; 32573b117750SEric Dumazet skb->ip_summed = CHECKSUM_PARTIAL; 32583b117750SEric Dumazet th->seq = htonl(tcp_rsk(req)->snt_isn); 32598336886fSJerry Chu /* XXX data is queued and acked as is. No buffer/window check */ 32608336886fSJerry Chu th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt); 32611da177e4SLinus Torvalds 32621da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 3263ed53d0abSEric Dumazet th->window = htons(min(req->rsk_rcv_wnd, 65535U)); 32645d062de7SEric Dumazet tcp_options_write((__be32 *)(th + 1), NULL, &opts); 32651da177e4SLinus Torvalds th->doff = (tcp_header_size >> 2); 326690bbcc60SEric Dumazet __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS); 3267cfb6eeb4SYOSHIFUJI Hideaki 3268cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 3269cfb6eeb4SYOSHIFUJI Hideaki /* Okay, we have all we need - do the md5 hash if needed */ 327080f03e27SEric Dumazet if (md5) 3271bd0388aeSWilliam Allen Simpson tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, 327239f8e58eSEric Dumazet md5, req_to_sk(req), skb); 327380f03e27SEric Dumazet rcu_read_unlock(); 3274cfb6eeb4SYOSHIFUJI Hideaki #endif 3275cfb6eeb4SYOSHIFUJI Hideaki 3276b50edd78SEric Dumazet /* Do not fool tcpdump (if any), clean our debris */ 32772456e855SThomas Gleixner skb->tstamp = 0; 32781da177e4SLinus Torvalds return skb; 32791da177e4SLinus Torvalds } 32804bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_make_synack); 32811da177e4SLinus Torvalds 328281164413SDaniel Borkmann static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst) 328381164413SDaniel Borkmann { 328481164413SDaniel Borkmann struct inet_connection_sock *icsk = inet_csk(sk); 328581164413SDaniel Borkmann const struct tcp_congestion_ops *ca; 328681164413SDaniel Borkmann u32 ca_key = dst_metric(dst, RTAX_CC_ALGO); 328781164413SDaniel Borkmann 328881164413SDaniel Borkmann if (ca_key == TCP_CA_UNSPEC) 328981164413SDaniel Borkmann return; 329081164413SDaniel Borkmann 329181164413SDaniel Borkmann rcu_read_lock(); 329281164413SDaniel Borkmann ca = tcp_ca_find_key(ca_key); 329381164413SDaniel Borkmann if (likely(ca && try_module_get(ca->owner))) { 329481164413SDaniel Borkmann module_put(icsk->icsk_ca_ops->owner); 329581164413SDaniel Borkmann icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst); 329681164413SDaniel Borkmann icsk->icsk_ca_ops = ca; 329781164413SDaniel Borkmann } 329881164413SDaniel Borkmann rcu_read_unlock(); 329981164413SDaniel Borkmann } 330081164413SDaniel Borkmann 330167edfef7SAndi Kleen /* Do all connect socket setups that can be done AF independent. */ 3302f7e56a76Sstephen hemminger static void tcp_connect_init(struct sock *sk) 33031da177e4SLinus Torvalds { 3304cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 33051da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 33061da177e4SLinus Torvalds __u8 rcv_wscale; 330713d3b1ebSLawrence Brakmo u32 rcv_wnd; 33081da177e4SLinus Torvalds 33091da177e4SLinus Torvalds /* We'll fix this up when we get a response from the other end. 33101da177e4SLinus Torvalds * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 33111da177e4SLinus Torvalds */ 33125d2ed052SEric Dumazet tp->tcp_header_len = sizeof(struct tcphdr); 33135d2ed052SEric Dumazet if (sock_net(sk)->ipv4.sysctl_tcp_timestamps) 33145d2ed052SEric Dumazet tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED; 33151da177e4SLinus Torvalds 3316cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 331700db4124SIan Morris if (tp->af_specific->md5_lookup(sk, sk)) 3318cfb6eeb4SYOSHIFUJI Hideaki tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 3319cfb6eeb4SYOSHIFUJI Hideaki #endif 3320cfb6eeb4SYOSHIFUJI Hideaki 33211da177e4SLinus Torvalds /* If user gave his TCP_MAXSEG, record it to clamp */ 33221da177e4SLinus Torvalds if (tp->rx_opt.user_mss) 33231da177e4SLinus Torvalds tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 33241da177e4SLinus Torvalds tp->max_window = 0; 33255d424d5aSJohn Heffner tcp_mtup_init(sk); 33261da177e4SLinus Torvalds tcp_sync_mss(sk, dst_mtu(dst)); 33271da177e4SLinus Torvalds 332881164413SDaniel Borkmann tcp_ca_dst_init(sk, dst); 332981164413SDaniel Borkmann 33301da177e4SLinus Torvalds if (!tp->window_clamp) 33311da177e4SLinus Torvalds tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 33323541f9e8SEric Dumazet tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 3333f5fff5dcSTom Quetchenbach 33341da177e4SLinus Torvalds tcp_initialize_rcv_mss(sk); 33351da177e4SLinus Torvalds 3336e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 3337e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 3338e88c64f0SHagen Paul Pfeifer (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) 3339e88c64f0SHagen Paul Pfeifer tp->window_clamp = tcp_full_space(sk); 3340e88c64f0SHagen Paul Pfeifer 334113d3b1ebSLawrence Brakmo rcv_wnd = tcp_rwnd_init_bpf(sk); 334213d3b1ebSLawrence Brakmo if (rcv_wnd == 0) 334313d3b1ebSLawrence Brakmo rcv_wnd = dst_metric(dst, RTAX_INITRWND); 334413d3b1ebSLawrence Brakmo 3345ceef9ab6SEric Dumazet tcp_select_initial_window(sk, tcp_full_space(sk), 33461da177e4SLinus Torvalds tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 33471da177e4SLinus Torvalds &tp->rcv_wnd, 33481da177e4SLinus Torvalds &tp->window_clamp, 33499bb37ef0SEric Dumazet sock_net(sk)->ipv4.sysctl_tcp_window_scaling, 335031d12926Slaurent chavey &rcv_wscale, 335113d3b1ebSLawrence Brakmo rcv_wnd); 33521da177e4SLinus Torvalds 33531da177e4SLinus Torvalds tp->rx_opt.rcv_wscale = rcv_wscale; 33541da177e4SLinus Torvalds tp->rcv_ssthresh = tp->rcv_wnd; 33551da177e4SLinus Torvalds 33561da177e4SLinus Torvalds sk->sk_err = 0; 33571da177e4SLinus Torvalds sock_reset_flag(sk, SOCK_DONE); 33581da177e4SLinus Torvalds tp->snd_wnd = 0; 3359ee7537b6SHantzis Fotis tcp_init_wl(tp, 0); 33607f582b24SEric Dumazet tcp_write_queue_purge(sk); 33611da177e4SLinus Torvalds tp->snd_una = tp->write_seq; 33621da177e4SLinus Torvalds tp->snd_sml = tp->write_seq; 336333f5f57eSIlpo Järvinen tp->snd_up = tp->write_seq; 3364370816aeSPavel Emelyanov tp->snd_nxt = tp->write_seq; 3365ee995283SPavel Emelyanov 3366ee995283SPavel Emelyanov if (likely(!tp->repair)) 33671da177e4SLinus Torvalds tp->rcv_nxt = 0; 3368c7781a6eSAndrew Vagin else 336970eabf0eSEric Dumazet tp->rcv_tstamp = tcp_jiffies32; 3370ee995283SPavel Emelyanov tp->rcv_wup = tp->rcv_nxt; 3371ee995283SPavel Emelyanov tp->copied_seq = tp->rcv_nxt; 33721da177e4SLinus Torvalds 33738550f328SLawrence Brakmo inet_csk(sk)->icsk_rto = tcp_timeout_init(sk); 3374463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_retransmits = 0; 33751da177e4SLinus Torvalds tcp_clear_retrans(tp); 33761da177e4SLinus Torvalds } 33771da177e4SLinus Torvalds 3378783237e8SYuchung Cheng static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) 3379783237e8SYuchung Cheng { 3380783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3381783237e8SYuchung Cheng struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 3382783237e8SYuchung Cheng 3383783237e8SYuchung Cheng tcb->end_seq += skb->len; 3384f4a775d1SEric Dumazet __skb_header_release(skb); 3385783237e8SYuchung Cheng sk->sk_wmem_queued += skb->truesize; 3386783237e8SYuchung Cheng sk_mem_charge(sk, skb->truesize); 3387783237e8SYuchung Cheng tp->write_seq = tcb->end_seq; 3388783237e8SYuchung Cheng tp->packets_out += tcp_skb_pcount(skb); 3389783237e8SYuchung Cheng } 3390783237e8SYuchung Cheng 3391783237e8SYuchung Cheng /* Build and send a SYN with data and (cached) Fast Open cookie. However, 3392783237e8SYuchung Cheng * queue a data-only packet after the regular SYN, such that regular SYNs 3393783237e8SYuchung Cheng * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges 3394783237e8SYuchung Cheng * only the SYN sequence, the data are retransmitted in the first ACK. 3395783237e8SYuchung Cheng * If cookie is not cached or other error occurs, falls back to send a 3396783237e8SYuchung Cheng * regular SYN with Fast Open cookie request option. 3397783237e8SYuchung Cheng */ 3398783237e8SYuchung Cheng static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) 3399783237e8SYuchung Cheng { 3400783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3401783237e8SYuchung Cheng struct tcp_fastopen_request *fo = tp->fastopen_req; 3402065263f4SWei Wang int space, err = 0; 3403355a901eSEric Dumazet struct sk_buff *syn_data; 3404783237e8SYuchung Cheng 340567da22d2SYuchung Cheng tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ 3406065263f4SWei Wang if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie)) 3407783237e8SYuchung Cheng goto fallback; 3408783237e8SYuchung Cheng 3409783237e8SYuchung Cheng /* MSS for SYN-data is based on cached MSS and bounded by PMTU and 3410783237e8SYuchung Cheng * user-MSS. Reserve maximum option space for middleboxes that add 3411783237e8SYuchung Cheng * private TCP options. The cost is reduced data space in SYN :( 3412783237e8SYuchung Cheng */ 34133541f9e8SEric Dumazet tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp); 34143541f9e8SEric Dumazet 34151b63edd6SYuchung Cheng space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) - 3416783237e8SYuchung Cheng MAX_TCP_OPTION_SPACE; 3417783237e8SYuchung Cheng 3418f5ddcbbbSEric Dumazet space = min_t(size_t, space, fo->size); 3419f5ddcbbbSEric Dumazet 3420f5ddcbbbSEric Dumazet /* limit to order-0 allocations */ 3421f5ddcbbbSEric Dumazet space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER)); 3422f5ddcbbbSEric Dumazet 3423eb934478SEric Dumazet syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false); 3424355a901eSEric Dumazet if (!syn_data) 3425783237e8SYuchung Cheng goto fallback; 3426355a901eSEric Dumazet syn_data->ip_summed = CHECKSUM_PARTIAL; 3427355a901eSEric Dumazet memcpy(syn_data->cb, syn->cb, sizeof(syn->cb)); 342807e100f9SEric Dumazet if (space) { 342907e100f9SEric Dumazet int copied = copy_from_iter(skb_put(syn_data, space), space, 343057be5bdaSAl Viro &fo->data->msg_iter); 343157be5bdaSAl Viro if (unlikely(!copied)) { 3432ba233b34SEric Dumazet tcp_skb_tsorted_anchor_cleanup(syn_data); 3433355a901eSEric Dumazet kfree_skb(syn_data); 3434783237e8SYuchung Cheng goto fallback; 3435783237e8SYuchung Cheng } 343657be5bdaSAl Viro if (copied != space) { 343757be5bdaSAl Viro skb_trim(syn_data, copied); 343857be5bdaSAl Viro space = copied; 343957be5bdaSAl Viro } 344007e100f9SEric Dumazet } 3441355a901eSEric Dumazet /* No more data pending in inet_wait_for_connect() */ 3442355a901eSEric Dumazet if (space == fo->size) 3443355a901eSEric Dumazet fo->data = NULL; 3444355a901eSEric Dumazet fo->copied = space; 3445783237e8SYuchung Cheng 3446355a901eSEric Dumazet tcp_connect_queue_skb(sk, syn_data); 34470f87230dSFrancis Yan if (syn_data->len) 34480f87230dSFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_BUSY); 3449355a901eSEric Dumazet 3450355a901eSEric Dumazet err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation); 3451355a901eSEric Dumazet 3452d3edd06eSEric Dumazet syn->skb_mstamp_ns = syn_data->skb_mstamp_ns; 3453355a901eSEric Dumazet 3454355a901eSEric Dumazet /* Now full SYN+DATA was cloned and sent (or not), 3455355a901eSEric Dumazet * remove the SYN from the original skb (syn_data) 3456355a901eSEric Dumazet * we keep in write queue in case of a retransmit, as we 3457355a901eSEric Dumazet * also have the SYN packet (with no data) in the same queue. 3458431a9124SEric Dumazet */ 3459355a901eSEric Dumazet TCP_SKB_CB(syn_data)->seq++; 3460355a901eSEric Dumazet TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH; 3461355a901eSEric Dumazet if (!err) { 346267da22d2SYuchung Cheng tp->syn_data = (fo->copied > 0); 346375c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data); 3464f19c29e3SYuchung Cheng NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT); 3465783237e8SYuchung Cheng goto done; 3466783237e8SYuchung Cheng } 3467783237e8SYuchung Cheng 346875c119afSEric Dumazet /* data was not sent, put it in write_queue */ 346975c119afSEric Dumazet __skb_queue_tail(&sk->sk_write_queue, syn_data); 3470b5b7db8dSEric Dumazet tp->packets_out -= tcp_skb_pcount(syn_data); 3471b5b7db8dSEric Dumazet 3472783237e8SYuchung Cheng fallback: 3473783237e8SYuchung Cheng /* Send a regular SYN with Fast Open cookie request option */ 3474783237e8SYuchung Cheng if (fo->cookie.len > 0) 3475783237e8SYuchung Cheng fo->cookie.len = 0; 3476783237e8SYuchung Cheng err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); 3477783237e8SYuchung Cheng if (err) 3478783237e8SYuchung Cheng tp->syn_fastopen = 0; 3479783237e8SYuchung Cheng done: 3480783237e8SYuchung Cheng fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ 3481783237e8SYuchung Cheng return err; 3482783237e8SYuchung Cheng } 3483783237e8SYuchung Cheng 348467edfef7SAndi Kleen /* Build a SYN and send it off. */ 34851da177e4SLinus Torvalds int tcp_connect(struct sock *sk) 34861da177e4SLinus Torvalds { 34871da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 34881da177e4SLinus Torvalds struct sk_buff *buff; 3489ee586811SEric Paris int err; 34901da177e4SLinus Torvalds 3491de525be2SLawrence Brakmo tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB, 0, NULL); 34928ba60924SEric Dumazet 34938ba60924SEric Dumazet if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 34948ba60924SEric Dumazet return -EHOSTUNREACH; /* Routing failure or similar. */ 34958ba60924SEric Dumazet 34961da177e4SLinus Torvalds tcp_connect_init(sk); 34971da177e4SLinus Torvalds 34982b916477SAndrey Vagin if (unlikely(tp->repair)) { 34992b916477SAndrey Vagin tcp_finish_connect(sk, NULL); 35002b916477SAndrey Vagin return 0; 35012b916477SAndrey Vagin } 35022b916477SAndrey Vagin 3503eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true); 3504355a901eSEric Dumazet if (unlikely(!buff)) 35051da177e4SLinus Torvalds return -ENOBUFS; 35061da177e4SLinus Torvalds 3507a3433f35SChangli Gao tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); 35089a568de4SEric Dumazet tcp_mstamp_refresh(tp); 35099a568de4SEric Dumazet tp->retrans_stamp = tcp_time_stamp(tp); 3510783237e8SYuchung Cheng tcp_connect_queue_skb(sk, buff); 3511735d3831SFlorian Westphal tcp_ecn_send_syn(sk, buff); 351275c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, buff); 35131da177e4SLinus Torvalds 3514783237e8SYuchung Cheng /* Send off SYN; include data in Fast Open. */ 3515783237e8SYuchung Cheng err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : 3516783237e8SYuchung Cheng tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); 3517ee586811SEric Paris if (err == -ECONNREFUSED) 3518ee586811SEric Paris return err; 3519bd37a088SWei Yongjun 3520bd37a088SWei Yongjun /* We change tp->snd_nxt after the tcp_transmit_skb() call 3521bd37a088SWei Yongjun * in order to make this packet get counted in tcpOutSegs. 3522bd37a088SWei Yongjun */ 3523bd37a088SWei Yongjun tp->snd_nxt = tp->write_seq; 3524bd37a088SWei Yongjun tp->pushed_seq = tp->write_seq; 3525b5b7db8dSEric Dumazet buff = tcp_send_head(sk); 3526b5b7db8dSEric Dumazet if (unlikely(buff)) { 3527b5b7db8dSEric Dumazet tp->snd_nxt = TCP_SKB_CB(buff)->seq; 3528b5b7db8dSEric Dumazet tp->pushed_seq = TCP_SKB_CB(buff)->seq; 3529b5b7db8dSEric Dumazet } 353081cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); 35311da177e4SLinus Torvalds 35321da177e4SLinus Torvalds /* Timer for repeating the SYN until an answer. */ 35333f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 35343f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 35351da177e4SLinus Torvalds return 0; 35361da177e4SLinus Torvalds } 35374bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_connect); 35381da177e4SLinus Torvalds 35391da177e4SLinus Torvalds /* Send out a delayed ack, the caller does the policy checking 35401da177e4SLinus Torvalds * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 35411da177e4SLinus Torvalds * for details. 35421da177e4SLinus Torvalds */ 35431da177e4SLinus Torvalds void tcp_send_delayed_ack(struct sock *sk) 35441da177e4SLinus Torvalds { 3545463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 3546463c84b9SArnaldo Carvalho de Melo int ato = icsk->icsk_ack.ato; 35471da177e4SLinus Torvalds unsigned long timeout; 35481da177e4SLinus Torvalds 35491da177e4SLinus Torvalds if (ato > TCP_DELACK_MIN) { 3550463c84b9SArnaldo Carvalho de Melo const struct tcp_sock *tp = tcp_sk(sk); 35511da177e4SLinus Torvalds int max_ato = HZ / 2; 35521da177e4SLinus Torvalds 3553056834d9SIlpo Järvinen if (icsk->icsk_ack.pingpong || 3554056834d9SIlpo Järvinen (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 35551da177e4SLinus Torvalds max_ato = TCP_DELACK_MAX; 35561da177e4SLinus Torvalds 35571da177e4SLinus Torvalds /* Slow path, intersegment interval is "high". */ 35581da177e4SLinus Torvalds 35591da177e4SLinus Torvalds /* If some rtt estimate is known, use it to bound delayed ack. 3560463c84b9SArnaldo Carvalho de Melo * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 35611da177e4SLinus Torvalds * directly. 35621da177e4SLinus Torvalds */ 3563740b0f18SEric Dumazet if (tp->srtt_us) { 3564740b0f18SEric Dumazet int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3), 3565740b0f18SEric Dumazet TCP_DELACK_MIN); 35661da177e4SLinus Torvalds 35671da177e4SLinus Torvalds if (rtt < max_ato) 35681da177e4SLinus Torvalds max_ato = rtt; 35691da177e4SLinus Torvalds } 35701da177e4SLinus Torvalds 35711da177e4SLinus Torvalds ato = min(ato, max_ato); 35721da177e4SLinus Torvalds } 35731da177e4SLinus Torvalds 35741da177e4SLinus Torvalds /* Stay within the limit we were given */ 35751da177e4SLinus Torvalds timeout = jiffies + ato; 35761da177e4SLinus Torvalds 35771da177e4SLinus Torvalds /* Use new timeout only if there wasn't a older one earlier. */ 3578463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 35791da177e4SLinus Torvalds /* If delack timer was blocked or is about to expire, 35801da177e4SLinus Torvalds * send ACK now. 35811da177e4SLinus Torvalds */ 3582463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.blocked || 3583463c84b9SArnaldo Carvalho de Melo time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 35841da177e4SLinus Torvalds tcp_send_ack(sk); 35851da177e4SLinus Torvalds return; 35861da177e4SLinus Torvalds } 35871da177e4SLinus Torvalds 3588463c84b9SArnaldo Carvalho de Melo if (!time_before(timeout, icsk->icsk_ack.timeout)) 3589463c84b9SArnaldo Carvalho de Melo timeout = icsk->icsk_ack.timeout; 35901da177e4SLinus Torvalds } 3591463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 3592463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.timeout = timeout; 3593463c84b9SArnaldo Carvalho de Melo sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 35941da177e4SLinus Torvalds } 35951da177e4SLinus Torvalds 35961da177e4SLinus Torvalds /* This routine sends an ack and also updates the window. */ 35972987babbSYuchung Cheng void __tcp_send_ack(struct sock *sk, u32 rcv_nxt) 35981da177e4SLinus Torvalds { 35991da177e4SLinus Torvalds struct sk_buff *buff; 36001da177e4SLinus Torvalds 3601058dc334SIlpo Järvinen /* If we have been reset, we may not send again. */ 3602058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3603058dc334SIlpo Järvinen return; 3604058dc334SIlpo Järvinen 36051da177e4SLinus Torvalds /* We are not putting this on the write queue, so 36061da177e4SLinus Torvalds * tcp_transmit_skb() will set the ownership to this 36071da177e4SLinus Torvalds * sock. 36081da177e4SLinus Torvalds */ 36097450aaf6SEric Dumazet buff = alloc_skb(MAX_TCP_HEADER, 36107450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 36117450aaf6SEric Dumazet if (unlikely(!buff)) { 3612463c84b9SArnaldo Carvalho de Melo inet_csk_schedule_ack(sk); 3613463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; 36143f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 36153f421baaSArnaldo Carvalho de Melo TCP_DELACK_MAX, TCP_RTO_MAX); 36161da177e4SLinus Torvalds return; 36171da177e4SLinus Torvalds } 36181da177e4SLinus Torvalds 36191da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 36201da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 3621a3433f35SChangli Gao tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); 36221da177e4SLinus Torvalds 362398781965SEric Dumazet /* We do not want pure acks influencing TCP Small Queues or fq/pacing 362498781965SEric Dumazet * too much. 362598781965SEric Dumazet * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784 362698781965SEric Dumazet */ 362798781965SEric Dumazet skb_set_tcp_pure_ack(buff); 362898781965SEric Dumazet 36291da177e4SLinus Torvalds /* Send it off, this clears delayed acks for us. */ 36302987babbSYuchung Cheng __tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0, rcv_nxt); 36311da177e4SLinus Torvalds } 363227cde44aSYuchung Cheng EXPORT_SYMBOL_GPL(__tcp_send_ack); 36332987babbSYuchung Cheng 36342987babbSYuchung Cheng void tcp_send_ack(struct sock *sk) 36352987babbSYuchung Cheng { 36362987babbSYuchung Cheng __tcp_send_ack(sk, tcp_sk(sk)->rcv_nxt); 36371da177e4SLinus Torvalds } 36381da177e4SLinus Torvalds 36391da177e4SLinus Torvalds /* This routine sends a packet with an out of date sequence 36401da177e4SLinus Torvalds * number. It assumes the other end will try to ack it. 36411da177e4SLinus Torvalds * 36421da177e4SLinus Torvalds * Question: what should we make while urgent mode? 36431da177e4SLinus Torvalds * 4.4BSD forces sending single byte of data. We cannot send 36441da177e4SLinus Torvalds * out of window data, because we have SND.NXT==SND.MAX... 36451da177e4SLinus Torvalds * 36461da177e4SLinus Torvalds * Current solution: to send TWO zero-length segments in urgent mode: 36471da177e4SLinus Torvalds * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 36481da177e4SLinus Torvalds * out-of-date with SND.UNA-1 to probe window. 36491da177e4SLinus Torvalds */ 3650e520af48SEric Dumazet static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib) 36511da177e4SLinus Torvalds { 36521da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 36531da177e4SLinus Torvalds struct sk_buff *skb; 36541da177e4SLinus Torvalds 36551da177e4SLinus Torvalds /* We don't queue it, tcp_transmit_skb() sets ownership. */ 36567450aaf6SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, 36577450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 365851456b29SIan Morris if (!skb) 36591da177e4SLinus Torvalds return -1; 36601da177e4SLinus Torvalds 36611da177e4SLinus Torvalds /* Reserve space for headers and set control bits. */ 36621da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 36631da177e4SLinus Torvalds /* Use a previous sequence. This should cause the other 36641da177e4SLinus Torvalds * end to send an ack. Don't queue or clone SKB, just 36651da177e4SLinus Torvalds * send it. 36661da177e4SLinus Torvalds */ 3667a3433f35SChangli Gao tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); 3668e2e8009fSRenato Westphal NET_INC_STATS(sock_net(sk), mib); 36697450aaf6SEric Dumazet return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0); 36701da177e4SLinus Torvalds } 36711da177e4SLinus Torvalds 3672385e2070SEric Dumazet /* Called from setsockopt( ... TCP_REPAIR ) */ 3673ee995283SPavel Emelyanov void tcp_send_window_probe(struct sock *sk) 3674ee995283SPavel Emelyanov { 3675ee995283SPavel Emelyanov if (sk->sk_state == TCP_ESTABLISHED) { 3676ee995283SPavel Emelyanov tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; 36779a568de4SEric Dumazet tcp_mstamp_refresh(tcp_sk(sk)); 3678e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE); 3679ee995283SPavel Emelyanov } 3680ee995283SPavel Emelyanov } 3681ee995283SPavel Emelyanov 368267edfef7SAndi Kleen /* Initiate keepalive or window probe from timer. */ 3683e520af48SEric Dumazet int tcp_write_wakeup(struct sock *sk, int mib) 36841da177e4SLinus Torvalds { 36851da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 36861da177e4SLinus Torvalds struct sk_buff *skb; 36871da177e4SLinus Torvalds 3688058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3689058dc334SIlpo Järvinen return -1; 3690058dc334SIlpo Järvinen 369100db4124SIan Morris skb = tcp_send_head(sk); 369200db4124SIan Morris if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { 36931da177e4SLinus Torvalds int err; 36940c54b85fSIlpo Järvinen unsigned int mss = tcp_current_mss(sk); 369590840defSIlpo Järvinen unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 36961da177e4SLinus Torvalds 36971da177e4SLinus Torvalds if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 36981da177e4SLinus Torvalds tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 36991da177e4SLinus Torvalds 37001da177e4SLinus Torvalds /* We are probing the opening of a window 37011da177e4SLinus Torvalds * but the window size is != 0 37021da177e4SLinus Torvalds * must have been a result SWS avoidance ( sender ) 37031da177e4SLinus Torvalds */ 37041da177e4SLinus Torvalds if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 37051da177e4SLinus Torvalds skb->len > mss) { 37061da177e4SLinus Torvalds seg_size = min(seg_size, mss); 37074de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 370875c119afSEric Dumazet if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE, 370975c119afSEric Dumazet skb, seg_size, mss, GFP_ATOMIC)) 37101da177e4SLinus Torvalds return -1; 37111da177e4SLinus Torvalds } else if (!tcp_skb_pcount(skb)) 37125bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss); 37131da177e4SLinus Torvalds 37144de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 3715dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 371666f5fe62SIlpo Järvinen if (!err) 371766f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 37181da177e4SLinus Torvalds return err; 37191da177e4SLinus Torvalds } else { 372033f5f57eSIlpo Järvinen if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) 3721e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 1, mib); 3722e520af48SEric Dumazet return tcp_xmit_probe_skb(sk, 0, mib); 37231da177e4SLinus Torvalds } 37241da177e4SLinus Torvalds } 37251da177e4SLinus Torvalds 37261da177e4SLinus Torvalds /* A window probe timeout has occurred. If window is not closed send 37271da177e4SLinus Torvalds * a partial packet else a zero probe. 37281da177e4SLinus Torvalds */ 37291da177e4SLinus Torvalds void tcp_send_probe0(struct sock *sk) 37301da177e4SLinus Torvalds { 3731463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 37321da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 3733c6214a97SNikolay Borisov struct net *net = sock_net(sk); 3734fcdd1cf4SEric Dumazet unsigned long probe_max; 37351da177e4SLinus Torvalds int err; 37361da177e4SLinus Torvalds 3737e520af48SEric Dumazet err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE); 37381da177e4SLinus Torvalds 373975c119afSEric Dumazet if (tp->packets_out || tcp_write_queue_empty(sk)) { 37401da177e4SLinus Torvalds /* Cancel probe timer, if it is not required. */ 37416687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 0; 3742463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff = 0; 37431da177e4SLinus Torvalds return; 37441da177e4SLinus Torvalds } 37451da177e4SLinus Torvalds 37461da177e4SLinus Torvalds if (err <= 0) { 3747c6214a97SNikolay Borisov if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2) 3748463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff++; 37496687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out++; 3750fcdd1cf4SEric Dumazet probe_max = TCP_RTO_MAX; 37511da177e4SLinus Torvalds } else { 37521da177e4SLinus Torvalds /* If packet was not sent due to local congestion, 37536687e988SArnaldo Carvalho de Melo * do not backoff and do not remember icsk_probes_out. 37541da177e4SLinus Torvalds * Let local senders to fight for local resources. 37551da177e4SLinus Torvalds * 37561da177e4SLinus Torvalds * Use accumulated backoff yet. 37571da177e4SLinus Torvalds */ 37586687e988SArnaldo Carvalho de Melo if (!icsk->icsk_probes_out) 37596687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 1; 3760fcdd1cf4SEric Dumazet probe_max = TCP_RESOURCE_PROBE_INTERVAL; 37611da177e4SLinus Torvalds } 37623f80e08fSEric Dumazet tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 376321c8fe99SEric Dumazet tcp_probe0_when(sk, probe_max), 37643f80e08fSEric Dumazet TCP_RTO_MAX, 37653f80e08fSEric Dumazet NULL); 37661da177e4SLinus Torvalds } 37675db92c99SOctavian Purdila 3768ea3bea3aSEric Dumazet int tcp_rtx_synack(const struct sock *sk, struct request_sock *req) 37695db92c99SOctavian Purdila { 37705db92c99SOctavian Purdila const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific; 37715db92c99SOctavian Purdila struct flowi fl; 37725db92c99SOctavian Purdila int res; 37735db92c99SOctavian Purdila 377458d607d3SEric Dumazet tcp_rsk(req)->txhash = net_tx_rndhash(); 3775b3d05147SEric Dumazet res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL); 37765db92c99SOctavian Purdila if (!res) { 377790bbcc60SEric Dumazet __TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); 377802a1d6e7SEric Dumazet __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 37797e32b443SYuchung Cheng if (unlikely(tcp_passive_fastopen(sk))) 37807e32b443SYuchung Cheng tcp_sk(sk)->total_retrans++; 3781cf34ce3dSSong Liu trace_tcp_retransmit_synack(sk, req); 37825db92c99SOctavian Purdila } 37835db92c99SOctavian Purdila return res; 37845db92c99SOctavian Purdila } 37855db92c99SOctavian Purdila EXPORT_SYMBOL(tcp_rtx_synack); 3786