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 559799ccb0SEric Dumazet /* departure time for next data packet */ 569799ccb0SEric Dumazet if (val > tp->tcp_wstamp_ns) 579799ccb0SEric Dumazet tp->tcp_wstamp_ns = val; 589799ccb0SEric Dumazet 599799ccb0SEric Dumazet val = div_u64(val, NSEC_PER_USEC); 609799ccb0SEric Dumazet if (val > tp->tcp_mstamp) 619799ccb0SEric Dumazet tp->tcp_mstamp = val; 629799ccb0SEric Dumazet } 639799ccb0SEric Dumazet 6446d3ceabSEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 6546d3ceabSEric Dumazet int push_one, gfp_t gfp); 66519855c5SWilliam Allen Simpson 6767edfef7SAndi Kleen /* Account for new data that has been sent to the network. */ 6875c119afSEric Dumazet static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb) 696ff03ac3SIlpo Järvinen { 706ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 716ff03ac3SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 7266f5fe62SIlpo Järvinen unsigned int prior_packets = tp->packets_out; 739e412ba7SIlpo Järvinen 741da177e4SLinus Torvalds tp->snd_nxt = TCP_SKB_CB(skb)->end_seq; 758512430eSIlpo Järvinen 7675c119afSEric Dumazet __skb_unlink(skb, &sk->sk_write_queue); 7775c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, skb); 7875c119afSEric Dumazet 7966f5fe62SIlpo Järvinen tp->packets_out += tcp_skb_pcount(skb); 80bec41a11SYuchung Cheng if (!prior_packets || icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) 81750ea2baSYuchung Cheng tcp_rearm_rto(sk); 82f19c29e3SYuchung Cheng 83f7324acdSDavid S. Miller NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT, 84f19c29e3SYuchung Cheng tcp_skb_pcount(skb)); 856a5dc9e5SEric Dumazet } 861da177e4SLinus Torvalds 87a4ecb15aSCui, Cheng /* SND.NXT, if window was not shrunk or the amount of shrunk was less than one 88a4ecb15aSCui, Cheng * window scaling factor due to loss of precision. 891da177e4SLinus Torvalds * If window has been shrunk, what should we make? It is not clear at all. 901da177e4SLinus Torvalds * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-( 911da177e4SLinus Torvalds * Anything in between SND.UNA...SND.UNA+SND.WND also can be already 921da177e4SLinus Torvalds * invalid. OK, let's make this for now: 931da177e4SLinus Torvalds */ 94cf533ea5SEric Dumazet static inline __u32 tcp_acceptable_seq(const struct sock *sk) 951da177e4SLinus Torvalds { 96cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 979e412ba7SIlpo Järvinen 98a4ecb15aSCui, Cheng if (!before(tcp_wnd_end(tp), tp->snd_nxt) || 99a4ecb15aSCui, Cheng (tp->rx_opt.wscale_ok && 100a4ecb15aSCui, Cheng ((tp->snd_nxt - tcp_wnd_end(tp)) < (1 << tp->rx_opt.rcv_wscale)))) 1011da177e4SLinus Torvalds return tp->snd_nxt; 1021da177e4SLinus Torvalds else 10390840defSIlpo Järvinen return tcp_wnd_end(tp); 1041da177e4SLinus Torvalds } 1051da177e4SLinus Torvalds 1061da177e4SLinus Torvalds /* Calculate mss to advertise in SYN segment. 1071da177e4SLinus Torvalds * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that: 1081da177e4SLinus Torvalds * 1091da177e4SLinus Torvalds * 1. It is independent of path mtu. 1101da177e4SLinus Torvalds * 2. Ideally, it is maximal possible segment size i.e. 65535-40. 1111da177e4SLinus Torvalds * 3. For IPv4 it is reasonable to calculate it from maximal MTU of 1121da177e4SLinus Torvalds * attached devices, because some buggy hosts are confused by 1131da177e4SLinus Torvalds * large MSS. 1141da177e4SLinus Torvalds * 4. We do not make 3, we advertise MSS, calculated from first 1151da177e4SLinus Torvalds * hop device mtu, but allow to raise it to ip_rt_min_advmss. 1161da177e4SLinus Torvalds * This may be overridden via information stored in routing table. 1171da177e4SLinus Torvalds * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible, 1181da177e4SLinus Torvalds * probably even Jumbo". 1191da177e4SLinus Torvalds */ 1201da177e4SLinus Torvalds static __u16 tcp_advertise_mss(struct sock *sk) 1211da177e4SLinus Torvalds { 1221da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 123cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1241da177e4SLinus Torvalds int mss = tp->advmss; 1251da177e4SLinus Torvalds 1260dbaee3bSDavid S. Miller if (dst) { 1270dbaee3bSDavid S. Miller unsigned int metric = dst_metric_advmss(dst); 1280dbaee3bSDavid S. Miller 1290dbaee3bSDavid S. Miller if (metric < mss) { 1300dbaee3bSDavid S. Miller mss = metric; 1311da177e4SLinus Torvalds tp->advmss = mss; 1321da177e4SLinus Torvalds } 1330dbaee3bSDavid S. Miller } 1341da177e4SLinus Torvalds 1351da177e4SLinus Torvalds return (__u16)mss; 1361da177e4SLinus Torvalds } 1371da177e4SLinus Torvalds 1381da177e4SLinus Torvalds /* RFC2861. Reset CWND after idle period longer RTO to "restart window". 1396f021c62SEric Dumazet * This is the first part of cwnd validation mechanism. 1406f021c62SEric Dumazet */ 1416f021c62SEric Dumazet void tcp_cwnd_restart(struct sock *sk, s32 delta) 1421da177e4SLinus Torvalds { 143463c84b9SArnaldo Carvalho de Melo struct tcp_sock *tp = tcp_sk(sk); 1446f021c62SEric Dumazet u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk)); 1451da177e4SLinus Torvalds u32 cwnd = tp->snd_cwnd; 1461da177e4SLinus Torvalds 1476687e988SArnaldo Carvalho de Melo tcp_ca_event(sk, CA_EVENT_CWND_RESTART); 1481da177e4SLinus Torvalds 1496687e988SArnaldo Carvalho de Melo tp->snd_ssthresh = tcp_current_ssthresh(sk); 1501da177e4SLinus Torvalds restart_cwnd = min(restart_cwnd, cwnd); 1511da177e4SLinus Torvalds 152463c84b9SArnaldo Carvalho de Melo while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd) 1531da177e4SLinus Torvalds cwnd >>= 1; 1541da177e4SLinus Torvalds tp->snd_cwnd = max(cwnd, restart_cwnd); 155c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 1561da177e4SLinus Torvalds tp->snd_cwnd_used = 0; 1571da177e4SLinus Torvalds } 1581da177e4SLinus Torvalds 15967edfef7SAndi Kleen /* Congestion state accounting after a packet has been sent. */ 16040efc6faSStephen Hemminger static void tcp_event_data_sent(struct tcp_sock *tp, 161cf533ea5SEric Dumazet struct sock *sk) 1621da177e4SLinus Torvalds { 163463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 164d635fbe2SEric Dumazet const u32 now = tcp_jiffies32; 1651da177e4SLinus Torvalds 16605c5a46dSNeal Cardwell if (tcp_packets_in_flight(tp) == 0) 16705c5a46dSNeal Cardwell tcp_ca_event(sk, CA_EVENT_TX_START); 16805c5a46dSNeal Cardwell 1691da177e4SLinus Torvalds tp->lsndtime = now; 1701da177e4SLinus Torvalds 1711da177e4SLinus Torvalds /* If it is a reply for ato after last received 1721da177e4SLinus Torvalds * packet, enter pingpong mode. 1731da177e4SLinus Torvalds */ 1742251ae46SJon Maxwell if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato) 175463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pingpong = 1; 1761da177e4SLinus Torvalds } 1771da177e4SLinus Torvalds 17867edfef7SAndi Kleen /* Account for an ACK we sent. */ 17927cde44aSYuchung Cheng static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts, 18027cde44aSYuchung Cheng u32 rcv_nxt) 1811da177e4SLinus Torvalds { 1825d9f4262SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 1835d9f4262SEric Dumazet 1845d9f4262SEric Dumazet if (unlikely(tp->compressed_ack)) { 185200d95f4SEric Dumazet NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED, 186200d95f4SEric Dumazet tp->compressed_ack); 1875d9f4262SEric Dumazet tp->compressed_ack = 0; 1885d9f4262SEric Dumazet if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1) 1895d9f4262SEric Dumazet __sock_put(sk); 1905d9f4262SEric Dumazet } 19127cde44aSYuchung Cheng 19227cde44aSYuchung Cheng if (unlikely(rcv_nxt != tp->rcv_nxt)) 19327cde44aSYuchung Cheng return; /* Special ACK sent by DCTCP to reflect ECN */ 194463c84b9SArnaldo Carvalho de Melo tcp_dec_quickack_mode(sk, pkts); 195463c84b9SArnaldo Carvalho de Melo inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); 1961da177e4SLinus Torvalds } 1971da177e4SLinus Torvalds 19885f16525SYuchung Cheng 19985f16525SYuchung Cheng u32 tcp_default_init_rwnd(u32 mss) 20085f16525SYuchung Cheng { 20185f16525SYuchung Cheng /* Initial receive window should be twice of TCP_INIT_CWND to 2029ef71e0cSWeiping Pan * enable proper sending of new unsent data during fast recovery 20385f16525SYuchung Cheng * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a 20485f16525SYuchung Cheng * limit when mss is larger than 1460. 20585f16525SYuchung Cheng */ 20685f16525SYuchung Cheng u32 init_rwnd = TCP_INIT_CWND * 2; 20785f16525SYuchung Cheng 20885f16525SYuchung Cheng if (mss > 1460) 20985f16525SYuchung Cheng init_rwnd = max((1460 * init_rwnd) / mss, 2U); 21085f16525SYuchung Cheng return init_rwnd; 21185f16525SYuchung Cheng } 21285f16525SYuchung Cheng 2131da177e4SLinus Torvalds /* Determine a window scaling and initial window to offer. 2141da177e4SLinus Torvalds * Based on the assumption that the given amount of space 2151da177e4SLinus Torvalds * will be offered. Store the results in the tp structure. 2161da177e4SLinus Torvalds * NOTE: for smooth operation initial space offering should 2171da177e4SLinus Torvalds * be a multiple of mss if possible. We assume here that mss >= 1. 2181da177e4SLinus Torvalds * This MUST be enforced by all callers. 2191da177e4SLinus Torvalds */ 220ceef9ab6SEric Dumazet void tcp_select_initial_window(const struct sock *sk, int __space, __u32 mss, 2211da177e4SLinus Torvalds __u32 *rcv_wnd, __u32 *window_clamp, 22231d12926Slaurent chavey int wscale_ok, __u8 *rcv_wscale, 22331d12926Slaurent chavey __u32 init_rcv_wnd) 2241da177e4SLinus Torvalds { 2251da177e4SLinus Torvalds unsigned int space = (__space < 0 ? 0 : __space); 2261da177e4SLinus Torvalds 2271da177e4SLinus Torvalds /* If no clamp set the clamp to the max possible scaled window */ 2281da177e4SLinus Torvalds if (*window_clamp == 0) 229589c49cbSGao Feng (*window_clamp) = (U16_MAX << TCP_MAX_WSCALE); 2301da177e4SLinus Torvalds space = min(*window_clamp, space); 2311da177e4SLinus Torvalds 2321da177e4SLinus Torvalds /* Quantize space offering to a multiple of mss if possible. */ 2331da177e4SLinus Torvalds if (space > mss) 234589c49cbSGao Feng space = rounddown(space, mss); 2351da177e4SLinus Torvalds 2361da177e4SLinus Torvalds /* NOTE: offering an initial window larger than 32767 23715d99e02SRick Jones * will break some buggy TCP stacks. If the admin tells us 23815d99e02SRick Jones * it is likely we could be speaking with such a buggy stack 23915d99e02SRick Jones * we will truncate our initial window offering to 32K-1 24015d99e02SRick Jones * unless the remote has sent us a window scaling option, 24115d99e02SRick Jones * which we interpret as a sign the remote TCP is not 24215d99e02SRick Jones * misinterpreting the window field as a signed quantity. 2431da177e4SLinus Torvalds */ 244ceef9ab6SEric Dumazet if (sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows) 2451da177e4SLinus Torvalds (*rcv_wnd) = min(space, MAX_TCP_WINDOW); 24615d99e02SRick Jones else 24715d99e02SRick Jones (*rcv_wnd) = space; 24815d99e02SRick Jones 2491da177e4SLinus Torvalds (*rcv_wscale) = 0; 2501da177e4SLinus Torvalds if (wscale_ok) { 251589c49cbSGao Feng /* Set window scaling on max possible window */ 252356d1833SEric Dumazet space = max_t(u32, space, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]); 253f626300aSSoheil Hassas Yeganeh space = max_t(u32, space, sysctl_rmem_max); 254316c1592SStephen Hemminger space = min_t(u32, space, *window_clamp); 255589c49cbSGao Feng while (space > U16_MAX && (*rcv_wscale) < TCP_MAX_WSCALE) { 2561da177e4SLinus Torvalds space >>= 1; 2571da177e4SLinus Torvalds (*rcv_wscale)++; 2581da177e4SLinus Torvalds } 2591da177e4SLinus Torvalds } 2601da177e4SLinus Torvalds 26185f16525SYuchung Cheng if (!init_rcv_wnd) /* Use default unless specified otherwise */ 26285f16525SYuchung Cheng init_rcv_wnd = tcp_default_init_rwnd(mss); 263b1afde60SNandita Dukkipati *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss); 2641da177e4SLinus Torvalds 2651da177e4SLinus Torvalds /* Set the clamp no higher than max representable value */ 266589c49cbSGao Feng (*window_clamp) = min_t(__u32, U16_MAX << (*rcv_wscale), *window_clamp); 2671da177e4SLinus Torvalds } 2684bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_select_initial_window); 2691da177e4SLinus Torvalds 2701da177e4SLinus Torvalds /* Chose a new window to advertise, update state in tcp_sock for the 2711da177e4SLinus Torvalds * socket, and return result with RFC1323 scaling applied. The return 2721da177e4SLinus Torvalds * value can be stuffed directly into th->window for an outgoing 2731da177e4SLinus Torvalds * frame. 2741da177e4SLinus Torvalds */ 27540efc6faSStephen Hemminger static u16 tcp_select_window(struct sock *sk) 2761da177e4SLinus Torvalds { 2771da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2788e165e20SFlorian Westphal u32 old_win = tp->rcv_wnd; 2791da177e4SLinus Torvalds u32 cur_win = tcp_receive_window(tp); 2801da177e4SLinus Torvalds u32 new_win = __tcp_select_window(sk); 2811da177e4SLinus Torvalds 2821da177e4SLinus Torvalds /* Never shrink the offered window */ 2831da177e4SLinus Torvalds if (new_win < cur_win) { 2841da177e4SLinus Torvalds /* Danger Will Robinson! 2851da177e4SLinus Torvalds * Don't update rcv_wup/rcv_wnd here or else 2861da177e4SLinus Torvalds * we will not be able to advertise a zero 2871da177e4SLinus Torvalds * window in time. --DaveM 2881da177e4SLinus Torvalds * 2891da177e4SLinus Torvalds * Relax Will Robinson. 2901da177e4SLinus Torvalds */ 2918e165e20SFlorian Westphal if (new_win == 0) 2928e165e20SFlorian Westphal NET_INC_STATS(sock_net(sk), 2938e165e20SFlorian Westphal LINUX_MIB_TCPWANTZEROWINDOWADV); 294607bfbf2SPatrick McHardy new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale); 2951da177e4SLinus Torvalds } 2961da177e4SLinus Torvalds tp->rcv_wnd = new_win; 2971da177e4SLinus Torvalds tp->rcv_wup = tp->rcv_nxt; 2981da177e4SLinus Torvalds 2991da177e4SLinus Torvalds /* Make sure we do not exceed the maximum possible 3001da177e4SLinus Torvalds * scaled window. 3011da177e4SLinus Torvalds */ 302ceef9ab6SEric Dumazet if (!tp->rx_opt.rcv_wscale && 303ceef9ab6SEric Dumazet sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows) 3041da177e4SLinus Torvalds new_win = min(new_win, MAX_TCP_WINDOW); 3051da177e4SLinus Torvalds else 3061da177e4SLinus Torvalds new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); 3071da177e4SLinus Torvalds 3081da177e4SLinus Torvalds /* RFC1323 scaling applied */ 3091da177e4SLinus Torvalds new_win >>= tp->rx_opt.rcv_wscale; 3101da177e4SLinus Torvalds 31131770e34SFlorian Westphal /* If we advertise zero window, disable fast path. */ 3128e165e20SFlorian Westphal if (new_win == 0) { 31331770e34SFlorian Westphal tp->pred_flags = 0; 3148e165e20SFlorian Westphal if (old_win) 3158e165e20SFlorian Westphal NET_INC_STATS(sock_net(sk), 3168e165e20SFlorian Westphal LINUX_MIB_TCPTOZEROWINDOWADV); 3178e165e20SFlorian Westphal } else if (old_win == 0) { 3188e165e20SFlorian Westphal NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV); 3198e165e20SFlorian Westphal } 3201da177e4SLinus Torvalds 3211da177e4SLinus Torvalds return new_win; 3221da177e4SLinus Torvalds } 3231da177e4SLinus Torvalds 32467edfef7SAndi Kleen /* Packet ECN state for a SYN-ACK */ 325735d3831SFlorian Westphal static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb) 326bdf1ee5dSIlpo Järvinen { 32730e502a3SDaniel Borkmann const struct tcp_sock *tp = tcp_sk(sk); 32830e502a3SDaniel Borkmann 3294de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR; 330bdf1ee5dSIlpo Järvinen if (!(tp->ecn_flags & TCP_ECN_OK)) 3314de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE; 33291b5b21cSLawrence Brakmo else if (tcp_ca_needs_ecn(sk) || 33391b5b21cSLawrence Brakmo tcp_bpf_ca_needs_ecn(sk)) 33430e502a3SDaniel Borkmann INET_ECN_xmit(sk); 335bdf1ee5dSIlpo Järvinen } 336bdf1ee5dSIlpo Järvinen 33767edfef7SAndi Kleen /* Packet ECN state for a SYN. */ 338735d3831SFlorian Westphal static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb) 339bdf1ee5dSIlpo Järvinen { 340bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 34191b5b21cSLawrence Brakmo bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk); 342f7b3bec6SFlorian Westphal bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 || 34391b5b21cSLawrence Brakmo tcp_ca_needs_ecn(sk) || bpf_needs_ecn; 344f7b3bec6SFlorian Westphal 345f7b3bec6SFlorian Westphal if (!use_ecn) { 346f7b3bec6SFlorian Westphal const struct dst_entry *dst = __sk_dst_get(sk); 347f7b3bec6SFlorian Westphal 348f7b3bec6SFlorian Westphal if (dst && dst_feature(dst, RTAX_FEATURE_ECN)) 349f7b3bec6SFlorian Westphal use_ecn = true; 350f7b3bec6SFlorian Westphal } 351bdf1ee5dSIlpo Järvinen 352bdf1ee5dSIlpo Järvinen tp->ecn_flags = 0; 353f7b3bec6SFlorian Westphal 354f7b3bec6SFlorian Westphal if (use_ecn) { 3554de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR; 356bdf1ee5dSIlpo Järvinen tp->ecn_flags = TCP_ECN_OK; 35791b5b21cSLawrence Brakmo if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn) 35830e502a3SDaniel Borkmann INET_ECN_xmit(sk); 359bdf1ee5dSIlpo Järvinen } 360bdf1ee5dSIlpo Järvinen } 361bdf1ee5dSIlpo Järvinen 36249213555SDaniel Borkmann static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb) 36349213555SDaniel Borkmann { 36449213555SDaniel Borkmann if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback) 36549213555SDaniel Borkmann /* tp->ecn_flags are cleared at a later point in time when 36649213555SDaniel Borkmann * SYN ACK is ultimatively being received. 36749213555SDaniel Borkmann */ 36849213555SDaniel Borkmann TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR); 36949213555SDaniel Borkmann } 37049213555SDaniel Borkmann 371735d3831SFlorian Westphal static void 3726ac705b1SEric Dumazet tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th) 373bdf1ee5dSIlpo Järvinen { 3746ac705b1SEric Dumazet if (inet_rsk(req)->ecn_ok) 375bdf1ee5dSIlpo Järvinen th->ece = 1; 376bdf1ee5dSIlpo Järvinen } 377bdf1ee5dSIlpo Järvinen 37867edfef7SAndi Kleen /* Set up ECN state for a packet on a ESTABLISHED socket that is about to 37967edfef7SAndi Kleen * be sent. 38067edfef7SAndi Kleen */ 381735d3831SFlorian Westphal static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb, 382ea1627c2SEric Dumazet struct tcphdr *th, int tcp_header_len) 383bdf1ee5dSIlpo Järvinen { 384bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 385bdf1ee5dSIlpo Järvinen 386bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_OK) { 387bdf1ee5dSIlpo Järvinen /* Not-retransmitted data segment: set ECT and inject CWR. */ 388bdf1ee5dSIlpo Järvinen if (skb->len != tcp_header_len && 389bdf1ee5dSIlpo Järvinen !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { 390bdf1ee5dSIlpo Järvinen INET_ECN_xmit(sk); 391bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) { 392bdf1ee5dSIlpo Järvinen tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; 393ea1627c2SEric Dumazet th->cwr = 1; 394bdf1ee5dSIlpo Järvinen skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; 395bdf1ee5dSIlpo Järvinen } 39630e502a3SDaniel Borkmann } else if (!tcp_ca_needs_ecn(sk)) { 397bdf1ee5dSIlpo Järvinen /* ACK or retransmitted segment: clear ECT|CE */ 398bdf1ee5dSIlpo Järvinen INET_ECN_dontxmit(sk); 399bdf1ee5dSIlpo Järvinen } 400bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) 401ea1627c2SEric Dumazet th->ece = 1; 402bdf1ee5dSIlpo Järvinen } 403bdf1ee5dSIlpo Järvinen } 404bdf1ee5dSIlpo Järvinen 405e870a8efSIlpo Järvinen /* Constructs common control bits of non-data skb. If SYN/FIN is present, 406e870a8efSIlpo Järvinen * auto increment end seqno. 407e870a8efSIlpo Järvinen */ 408e870a8efSIlpo Järvinen static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags) 409e870a8efSIlpo Järvinen { 4102e8e18efSDavid S. Miller skb->ip_summed = CHECKSUM_PARTIAL; 411e870a8efSIlpo Järvinen 4124de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags; 413e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->sacked = 0; 414e870a8efSIlpo Järvinen 415cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, 1); 416e870a8efSIlpo Järvinen 417e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->seq = seq; 418a3433f35SChangli Gao if (flags & (TCPHDR_SYN | TCPHDR_FIN)) 419e870a8efSIlpo Järvinen seq++; 420e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->end_seq = seq; 421e870a8efSIlpo Järvinen } 422e870a8efSIlpo Järvinen 423a2a385d6SEric Dumazet static inline bool tcp_urg_mode(const struct tcp_sock *tp) 42433f5f57eSIlpo Järvinen { 42533f5f57eSIlpo Järvinen return tp->snd_una != tp->snd_up; 42633f5f57eSIlpo Järvinen } 42733f5f57eSIlpo Järvinen 42833ad798cSAdam Langley #define OPTION_SACK_ADVERTISE (1 << 0) 42933ad798cSAdam Langley #define OPTION_TS (1 << 1) 43033ad798cSAdam Langley #define OPTION_MD5 (1 << 2) 43189e95a61SOri Finkelman #define OPTION_WSCALE (1 << 3) 4322100c8d2SYuchung Cheng #define OPTION_FAST_OPEN_COOKIE (1 << 8) 43360e2a778SUrsula Braun #define OPTION_SMC (1 << 9) 43460e2a778SUrsula Braun 43560e2a778SUrsula Braun static void smc_options_write(__be32 *ptr, u16 *options) 43660e2a778SUrsula Braun { 43760e2a778SUrsula Braun #if IS_ENABLED(CONFIG_SMC) 43860e2a778SUrsula Braun if (static_branch_unlikely(&tcp_have_smc)) { 43960e2a778SUrsula Braun if (unlikely(OPTION_SMC & *options)) { 44060e2a778SUrsula Braun *ptr++ = htonl((TCPOPT_NOP << 24) | 44160e2a778SUrsula Braun (TCPOPT_NOP << 16) | 44260e2a778SUrsula Braun (TCPOPT_EXP << 8) | 44360e2a778SUrsula Braun (TCPOLEN_EXP_SMC_BASE)); 44460e2a778SUrsula Braun *ptr++ = htonl(TCPOPT_SMC_MAGIC); 44560e2a778SUrsula Braun } 44660e2a778SUrsula Braun } 44760e2a778SUrsula Braun #endif 44860e2a778SUrsula Braun } 44933ad798cSAdam Langley 45033ad798cSAdam Langley struct tcp_out_options { 4512100c8d2SYuchung Cheng u16 options; /* bit field of OPTION_* */ 4522100c8d2SYuchung Cheng u16 mss; /* 0 to disable */ 45333ad798cSAdam Langley u8 ws; /* window scale, 0 to disable */ 45433ad798cSAdam Langley u8 num_sack_blocks; /* number of SACK blocks to include */ 455bd0388aeSWilliam Allen Simpson u8 hash_size; /* bytes in hash_location */ 456bd0388aeSWilliam Allen Simpson __u8 *hash_location; /* temporary pointer, overloaded */ 4572100c8d2SYuchung Cheng __u32 tsval, tsecr; /* need to include OPTION_TS */ 4582100c8d2SYuchung Cheng struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */ 45933ad798cSAdam Langley }; 46033ad798cSAdam Langley 46167edfef7SAndi Kleen /* Write previously computed TCP options to the packet. 46267edfef7SAndi Kleen * 46367edfef7SAndi Kleen * Beware: Something in the Internet is very sensitive to the ordering of 464fd6149d3SIlpo Järvinen * TCP options, we learned this through the hard way, so be careful here. 465fd6149d3SIlpo Järvinen * Luckily we can at least blame others for their non-compliance but from 4668e3bff96Sstephen hemminger * inter-operability perspective it seems that we're somewhat stuck with 467fd6149d3SIlpo Järvinen * the ordering which we have been using if we want to keep working with 468fd6149d3SIlpo Järvinen * those broken things (not that it currently hurts anybody as there isn't 469fd6149d3SIlpo Järvinen * particular reason why the ordering would need to be changed). 470fd6149d3SIlpo Järvinen * 471fd6149d3SIlpo Järvinen * At least SACK_PERM as the first option is known to lead to a disaster 472fd6149d3SIlpo Järvinen * (but it may well be that other scenarios fail similarly). 473fd6149d3SIlpo Järvinen */ 47433ad798cSAdam Langley static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp, 475bd0388aeSWilliam Allen Simpson struct tcp_out_options *opts) 476bd0388aeSWilliam Allen Simpson { 4772100c8d2SYuchung Cheng u16 options = opts->options; /* mungable copy */ 478bd0388aeSWilliam Allen Simpson 479bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_MD5 & options)) { 4801a2c6181SChristoph Paasch *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | 4811a2c6181SChristoph Paasch (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG); 482bd0388aeSWilliam Allen Simpson /* overload cookie hash location */ 483bd0388aeSWilliam Allen Simpson opts->hash_location = (__u8 *)ptr; 48433ad798cSAdam Langley ptr += 4; 48533ad798cSAdam Langley } 48633ad798cSAdam Langley 487fd6149d3SIlpo Järvinen if (unlikely(opts->mss)) { 488fd6149d3SIlpo Järvinen *ptr++ = htonl((TCPOPT_MSS << 24) | 489fd6149d3SIlpo Järvinen (TCPOLEN_MSS << 16) | 490fd6149d3SIlpo Järvinen opts->mss); 491fd6149d3SIlpo Järvinen } 492fd6149d3SIlpo Järvinen 493bd0388aeSWilliam Allen Simpson if (likely(OPTION_TS & options)) { 494bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 49533ad798cSAdam Langley *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | 49633ad798cSAdam Langley (TCPOLEN_SACK_PERM << 16) | 49733ad798cSAdam Langley (TCPOPT_TIMESTAMP << 8) | 49833ad798cSAdam Langley TCPOLEN_TIMESTAMP); 499bd0388aeSWilliam Allen Simpson options &= ~OPTION_SACK_ADVERTISE; 50033ad798cSAdam Langley } else { 501496c98dfSYOSHIFUJI Hideaki *ptr++ = htonl((TCPOPT_NOP << 24) | 50240efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 50340efc6faSStephen Hemminger (TCPOPT_TIMESTAMP << 8) | 50440efc6faSStephen Hemminger TCPOLEN_TIMESTAMP); 50540efc6faSStephen Hemminger } 50633ad798cSAdam Langley *ptr++ = htonl(opts->tsval); 50733ad798cSAdam Langley *ptr++ = htonl(opts->tsecr); 50833ad798cSAdam Langley } 50933ad798cSAdam Langley 510bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 51133ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 51233ad798cSAdam Langley (TCPOPT_NOP << 16) | 51333ad798cSAdam Langley (TCPOPT_SACK_PERM << 8) | 51433ad798cSAdam Langley TCPOLEN_SACK_PERM); 51533ad798cSAdam Langley } 51633ad798cSAdam Langley 517bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_WSCALE & options)) { 51833ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 51933ad798cSAdam Langley (TCPOPT_WINDOW << 16) | 52033ad798cSAdam Langley (TCPOLEN_WINDOW << 8) | 52133ad798cSAdam Langley opts->ws); 52233ad798cSAdam Langley } 52333ad798cSAdam Langley 52433ad798cSAdam Langley if (unlikely(opts->num_sack_blocks)) { 52533ad798cSAdam Langley struct tcp_sack_block *sp = tp->rx_opt.dsack ? 52633ad798cSAdam Langley tp->duplicate_sack : tp->selective_acks; 52740efc6faSStephen Hemminger int this_sack; 52840efc6faSStephen Hemminger 52940efc6faSStephen Hemminger *ptr++ = htonl((TCPOPT_NOP << 24) | 53040efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 53140efc6faSStephen Hemminger (TCPOPT_SACK << 8) | 53233ad798cSAdam Langley (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * 53340efc6faSStephen Hemminger TCPOLEN_SACK_PERBLOCK))); 5342de979bdSStephen Hemminger 53533ad798cSAdam Langley for (this_sack = 0; this_sack < opts->num_sack_blocks; 53633ad798cSAdam Langley ++this_sack) { 53740efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].start_seq); 53840efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].end_seq); 53940efc6faSStephen Hemminger } 5402de979bdSStephen Hemminger 54140efc6faSStephen Hemminger tp->rx_opt.dsack = 0; 54240efc6faSStephen Hemminger } 5432100c8d2SYuchung Cheng 5442100c8d2SYuchung Cheng if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) { 5452100c8d2SYuchung Cheng struct tcp_fastopen_cookie *foc = opts->fastopen_cookie; 5467f9b838bSDaniel Lee u8 *p = (u8 *)ptr; 5477f9b838bSDaniel Lee u32 len; /* Fast Open option length */ 5482100c8d2SYuchung Cheng 5497f9b838bSDaniel Lee if (foc->exp) { 5507f9b838bSDaniel Lee len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len; 5517f9b838bSDaniel Lee *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) | 5522100c8d2SYuchung Cheng TCPOPT_FASTOPEN_MAGIC); 5537f9b838bSDaniel Lee p += TCPOLEN_EXP_FASTOPEN_BASE; 5547f9b838bSDaniel Lee } else { 5557f9b838bSDaniel Lee len = TCPOLEN_FASTOPEN_BASE + foc->len; 5567f9b838bSDaniel Lee *p++ = TCPOPT_FASTOPEN; 5577f9b838bSDaniel Lee *p++ = len; 5582100c8d2SYuchung Cheng } 5597f9b838bSDaniel Lee 5607f9b838bSDaniel Lee memcpy(p, foc->val, foc->len); 5617f9b838bSDaniel Lee if ((len & 3) == 2) { 5627f9b838bSDaniel Lee p[foc->len] = TCPOPT_NOP; 5637f9b838bSDaniel Lee p[foc->len + 1] = TCPOPT_NOP; 5647f9b838bSDaniel Lee } 5657f9b838bSDaniel Lee ptr += (len + 3) >> 2; 5662100c8d2SYuchung Cheng } 56760e2a778SUrsula Braun 56860e2a778SUrsula Braun smc_options_write(ptr, &options); 56960e2a778SUrsula Braun } 57060e2a778SUrsula Braun 57160e2a778SUrsula Braun static void smc_set_option(const struct tcp_sock *tp, 57260e2a778SUrsula Braun struct tcp_out_options *opts, 57360e2a778SUrsula Braun unsigned int *remaining) 57460e2a778SUrsula Braun { 57560e2a778SUrsula Braun #if IS_ENABLED(CONFIG_SMC) 57660e2a778SUrsula Braun if (static_branch_unlikely(&tcp_have_smc)) { 57760e2a778SUrsula Braun if (tp->syn_smc) { 57860e2a778SUrsula Braun if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) { 57960e2a778SUrsula Braun opts->options |= OPTION_SMC; 58060e2a778SUrsula Braun *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED; 58160e2a778SUrsula Braun } 58260e2a778SUrsula Braun } 58360e2a778SUrsula Braun } 58460e2a778SUrsula Braun #endif 58560e2a778SUrsula Braun } 58660e2a778SUrsula Braun 58760e2a778SUrsula Braun static void smc_set_option_cond(const struct tcp_sock *tp, 58860e2a778SUrsula Braun const struct inet_request_sock *ireq, 58960e2a778SUrsula Braun struct tcp_out_options *opts, 59060e2a778SUrsula Braun unsigned int *remaining) 59160e2a778SUrsula Braun { 59260e2a778SUrsula Braun #if IS_ENABLED(CONFIG_SMC) 59360e2a778SUrsula Braun if (static_branch_unlikely(&tcp_have_smc)) { 59460e2a778SUrsula Braun if (tp->syn_smc && ireq->smc_ok) { 59560e2a778SUrsula Braun if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) { 59660e2a778SUrsula Braun opts->options |= OPTION_SMC; 59760e2a778SUrsula Braun *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED; 59860e2a778SUrsula Braun } 59960e2a778SUrsula Braun } 60060e2a778SUrsula Braun } 60160e2a778SUrsula Braun #endif 60240efc6faSStephen Hemminger } 60340efc6faSStephen Hemminger 60467edfef7SAndi Kleen /* Compute TCP options for SYN packets. This is not the final 60567edfef7SAndi Kleen * network wire format yet. 60667edfef7SAndi Kleen */ 60795c96174SEric Dumazet static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb, 60833ad798cSAdam Langley struct tcp_out_options *opts, 609cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 610cf533ea5SEric Dumazet { 61133ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 61295c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 613783237e8SYuchung Cheng struct tcp_fastopen_request *fastopen = tp->fastopen_req; 61433ad798cSAdam Langley 6158c2320e8SEric Dumazet *md5 = NULL; 616cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 6178c2320e8SEric Dumazet if (unlikely(rcu_access_pointer(tp->md5sig_info))) { 61833ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 61933ad798cSAdam Langley if (*md5) { 62033ad798cSAdam Langley opts->options |= OPTION_MD5; 621bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 622cfb6eeb4SYOSHIFUJI Hideaki } 6238c2320e8SEric Dumazet } 624cfb6eeb4SYOSHIFUJI Hideaki #endif 62533ad798cSAdam Langley 62633ad798cSAdam Langley /* We always get an MSS option. The option bytes which will be seen in 62733ad798cSAdam Langley * normal data packets should timestamps be used, must be in the MSS 62833ad798cSAdam Langley * advertised. But we subtract them from tp->mss_cache so that 62933ad798cSAdam Langley * calculations in tcp_sendmsg are simpler etc. So account for this 63033ad798cSAdam Langley * fact here if necessary. If we don't do this correctly, as a 63133ad798cSAdam Langley * receiver we won't recognize data packets as being full sized when we 63233ad798cSAdam Langley * should, and thus we won't abide by the delayed ACK rules correctly. 63333ad798cSAdam Langley * SACKs don't matter, we never delay an ACK when we have any of those 63433ad798cSAdam Langley * going out. */ 63533ad798cSAdam Langley opts->mss = tcp_advertise_mss(sk); 636bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 63733ad798cSAdam Langley 6385d2ed052SEric Dumazet if (likely(sock_net(sk)->ipv4.sysctl_tcp_timestamps && !*md5)) { 63933ad798cSAdam Langley opts->options |= OPTION_TS; 6407faee5c0SEric Dumazet opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset; 64133ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 642bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 64333ad798cSAdam Langley } 6449bb37ef0SEric Dumazet if (likely(sock_net(sk)->ipv4.sysctl_tcp_window_scaling)) { 64533ad798cSAdam Langley opts->ws = tp->rx_opt.rcv_wscale; 64689e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 647bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 64833ad798cSAdam Langley } 649f9301034SEric Dumazet if (likely(sock_net(sk)->ipv4.sysctl_tcp_sack)) { 65033ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 651b32d1310SDavid S. Miller if (unlikely(!(OPTION_TS & opts->options))) 652bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 65333ad798cSAdam Langley } 65433ad798cSAdam Langley 655783237e8SYuchung Cheng if (fastopen && fastopen->cookie.len >= 0) { 6562646c831SDaniel Lee u32 need = fastopen->cookie.len; 6572646c831SDaniel Lee 6582646c831SDaniel Lee need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE : 6592646c831SDaniel Lee TCPOLEN_FASTOPEN_BASE; 660783237e8SYuchung Cheng need = (need + 3) & ~3U; /* Align to 32 bits */ 661783237e8SYuchung Cheng if (remaining >= need) { 662783237e8SYuchung Cheng opts->options |= OPTION_FAST_OPEN_COOKIE; 663783237e8SYuchung Cheng opts->fastopen_cookie = &fastopen->cookie; 664783237e8SYuchung Cheng remaining -= need; 665783237e8SYuchung Cheng tp->syn_fastopen = 1; 6662646c831SDaniel Lee tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0; 667783237e8SYuchung Cheng } 668783237e8SYuchung Cheng } 669bd0388aeSWilliam Allen Simpson 67060e2a778SUrsula Braun smc_set_option(tp, opts, &remaining); 67160e2a778SUrsula Braun 672bd0388aeSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 67333ad798cSAdam Langley } 67433ad798cSAdam Langley 67567edfef7SAndi Kleen /* Set up TCP options for SYN-ACKs. */ 67660e2a778SUrsula Braun static unsigned int tcp_synack_options(const struct sock *sk, 67760e2a778SUrsula Braun struct request_sock *req, 67895c96174SEric Dumazet unsigned int mss, struct sk_buff *skb, 67933ad798cSAdam Langley struct tcp_out_options *opts, 68080f03e27SEric Dumazet const struct tcp_md5sig_key *md5, 6818336886fSJerry Chu struct tcp_fastopen_cookie *foc) 6824957faadSWilliam Allen Simpson { 68333ad798cSAdam Langley struct inet_request_sock *ireq = inet_rsk(req); 68495c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 68533ad798cSAdam Langley 68633ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 68780f03e27SEric Dumazet if (md5) { 68833ad798cSAdam Langley opts->options |= OPTION_MD5; 6894957faadSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 6904957faadSWilliam Allen Simpson 6914957faadSWilliam Allen Simpson /* We can't fit any SACK blocks in a packet with MD5 + TS 6924957faadSWilliam Allen Simpson * options. There was discussion about disabling SACK 6934957faadSWilliam Allen Simpson * rather than TS in order to fit in better with old, 6944957faadSWilliam Allen Simpson * buggy kernels, but that was deemed to be unnecessary. 6954957faadSWilliam Allen Simpson */ 696de213e5eSEric Dumazet ireq->tstamp_ok &= !ireq->sack_ok; 69733ad798cSAdam Langley } 69833ad798cSAdam Langley #endif 69933ad798cSAdam Langley 7004957faadSWilliam Allen Simpson /* We always send an MSS option. */ 70133ad798cSAdam Langley opts->mss = mss; 7024957faadSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 70333ad798cSAdam Langley 70433ad798cSAdam Langley if (likely(ireq->wscale_ok)) { 70533ad798cSAdam Langley opts->ws = ireq->rcv_wscale; 70689e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 7074957faadSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 70833ad798cSAdam Langley } 709de213e5eSEric Dumazet if (likely(ireq->tstamp_ok)) { 71033ad798cSAdam Langley opts->options |= OPTION_TS; 71195a22caeSFlorian Westphal opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off; 71233ad798cSAdam Langley opts->tsecr = req->ts_recent; 7134957faadSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 71433ad798cSAdam Langley } 71533ad798cSAdam Langley if (likely(ireq->sack_ok)) { 71633ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 717de213e5eSEric Dumazet if (unlikely(!ireq->tstamp_ok)) 7184957faadSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 71933ad798cSAdam Langley } 7207f9b838bSDaniel Lee if (foc != NULL && foc->len >= 0) { 7217f9b838bSDaniel Lee u32 need = foc->len; 7227f9b838bSDaniel Lee 7237f9b838bSDaniel Lee need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE : 7247f9b838bSDaniel Lee TCPOLEN_FASTOPEN_BASE; 7258336886fSJerry Chu need = (need + 3) & ~3U; /* Align to 32 bits */ 7268336886fSJerry Chu if (remaining >= need) { 7278336886fSJerry Chu opts->options |= OPTION_FAST_OPEN_COOKIE; 7288336886fSJerry Chu opts->fastopen_cookie = foc; 7298336886fSJerry Chu remaining -= need; 7308336886fSJerry Chu } 7318336886fSJerry Chu } 7324957faadSWilliam Allen Simpson 73360e2a778SUrsula Braun smc_set_option_cond(tcp_sk(sk), ireq, opts, &remaining); 73460e2a778SUrsula Braun 7354957faadSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 73633ad798cSAdam Langley } 73733ad798cSAdam Langley 73867edfef7SAndi Kleen /* Compute TCP options for ESTABLISHED sockets. This is not the 73967edfef7SAndi Kleen * final wire format yet. 74067edfef7SAndi Kleen */ 74195c96174SEric Dumazet static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb, 74233ad798cSAdam Langley struct tcp_out_options *opts, 743cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 744cf533ea5SEric Dumazet { 74533ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 74695c96174SEric Dumazet unsigned int size = 0; 747cabeccbdSIlpo Järvinen unsigned int eff_sacks; 74833ad798cSAdam Langley 7495843ef42SAndi Kleen opts->options = 0; 7505843ef42SAndi Kleen 7518c2320e8SEric Dumazet *md5 = NULL; 75233ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 7538c2320e8SEric Dumazet if (unlikely(rcu_access_pointer(tp->md5sig_info))) { 75433ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 7558c2320e8SEric Dumazet if (*md5) { 75633ad798cSAdam Langley opts->options |= OPTION_MD5; 75733ad798cSAdam Langley size += TCPOLEN_MD5SIG_ALIGNED; 75833ad798cSAdam Langley } 7598c2320e8SEric Dumazet } 76033ad798cSAdam Langley #endif 76133ad798cSAdam Langley 76233ad798cSAdam Langley if (likely(tp->rx_opt.tstamp_ok)) { 76333ad798cSAdam Langley opts->options |= OPTION_TS; 7647faee5c0SEric Dumazet opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0; 76533ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 76633ad798cSAdam Langley size += TCPOLEN_TSTAMP_ALIGNED; 76733ad798cSAdam Langley } 76833ad798cSAdam Langley 769cabeccbdSIlpo Järvinen eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; 770cabeccbdSIlpo Järvinen if (unlikely(eff_sacks)) { 77195c96174SEric Dumazet const unsigned int remaining = MAX_TCP_OPTION_SPACE - size; 77233ad798cSAdam Langley opts->num_sack_blocks = 77395c96174SEric Dumazet min_t(unsigned int, eff_sacks, 77433ad798cSAdam Langley (remaining - TCPOLEN_SACK_BASE_ALIGNED) / 77533ad798cSAdam Langley TCPOLEN_SACK_PERBLOCK); 77633ad798cSAdam Langley size += TCPOLEN_SACK_BASE_ALIGNED + 77733ad798cSAdam Langley opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; 77833ad798cSAdam Langley } 77933ad798cSAdam Langley 78033ad798cSAdam Langley return size; 78140efc6faSStephen Hemminger } 7821da177e4SLinus Torvalds 78346d3ceabSEric Dumazet 78446d3ceabSEric Dumazet /* TCP SMALL QUEUES (TSQ) 78546d3ceabSEric Dumazet * 78646d3ceabSEric Dumazet * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev) 78746d3ceabSEric Dumazet * to reduce RTT and bufferbloat. 78846d3ceabSEric Dumazet * We do this using a special skb destructor (tcp_wfree). 78946d3ceabSEric Dumazet * 79046d3ceabSEric Dumazet * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb 79146d3ceabSEric Dumazet * needs to be reallocated in a driver. 7928e3bff96Sstephen hemminger * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc 79346d3ceabSEric Dumazet * 79446d3ceabSEric Dumazet * Since transmit from skb destructor is forbidden, we use a tasklet 79546d3ceabSEric Dumazet * to process all sockets that eventually need to send more skbs. 79646d3ceabSEric Dumazet * We use one tasklet per cpu, with its own queue of sockets. 79746d3ceabSEric Dumazet */ 79846d3ceabSEric Dumazet struct tsq_tasklet { 79946d3ceabSEric Dumazet struct tasklet_struct tasklet; 80046d3ceabSEric Dumazet struct list_head head; /* queue of tcp sockets */ 80146d3ceabSEric Dumazet }; 80246d3ceabSEric Dumazet static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet); 80346d3ceabSEric Dumazet 80473a6bab5SEric Dumazet static void tcp_tsq_write(struct sock *sk) 8056f458dfbSEric Dumazet { 8066f458dfbSEric Dumazet if ((1 << sk->sk_state) & 8076f458dfbSEric Dumazet (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING | 808f9616c35SEric Dumazet TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) { 809f9616c35SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 810f9616c35SEric Dumazet 811f9616c35SEric Dumazet if (tp->lost_out > tp->retrans_out && 8123a91d29fSKoichiro Den tp->snd_cwnd > tcp_packets_in_flight(tp)) { 8133a91d29fSKoichiro Den tcp_mstamp_refresh(tp); 814f9616c35SEric Dumazet tcp_xmit_retransmit_queue(sk); 8153a91d29fSKoichiro Den } 816f9616c35SEric Dumazet 817f9616c35SEric Dumazet tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle, 818bf06200eSJohn Ogness 0, GFP_ATOMIC); 8196f458dfbSEric Dumazet } 820f9616c35SEric Dumazet } 82173a6bab5SEric Dumazet 82273a6bab5SEric Dumazet static void tcp_tsq_handler(struct sock *sk) 82373a6bab5SEric Dumazet { 82473a6bab5SEric Dumazet bh_lock_sock(sk); 82573a6bab5SEric Dumazet if (!sock_owned_by_user(sk)) 82673a6bab5SEric Dumazet tcp_tsq_write(sk); 82773a6bab5SEric Dumazet else if (!test_and_set_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags)) 82873a6bab5SEric Dumazet sock_hold(sk); 82973a6bab5SEric Dumazet bh_unlock_sock(sk); 83073a6bab5SEric Dumazet } 83146d3ceabSEric Dumazet /* 8328e3bff96Sstephen hemminger * One tasklet per cpu tries to send more skbs. 83346d3ceabSEric Dumazet * We run in tasklet context but need to disable irqs when 8348e3bff96Sstephen hemminger * transferring tsq->head because tcp_wfree() might 83546d3ceabSEric Dumazet * interrupt us (non NAPI drivers) 83646d3ceabSEric Dumazet */ 83746d3ceabSEric Dumazet static void tcp_tasklet_func(unsigned long data) 83846d3ceabSEric Dumazet { 83946d3ceabSEric Dumazet struct tsq_tasklet *tsq = (struct tsq_tasklet *)data; 84046d3ceabSEric Dumazet LIST_HEAD(list); 84146d3ceabSEric Dumazet unsigned long flags; 84246d3ceabSEric Dumazet struct list_head *q, *n; 84346d3ceabSEric Dumazet struct tcp_sock *tp; 84446d3ceabSEric Dumazet struct sock *sk; 84546d3ceabSEric Dumazet 84646d3ceabSEric Dumazet local_irq_save(flags); 84746d3ceabSEric Dumazet list_splice_init(&tsq->head, &list); 84846d3ceabSEric Dumazet local_irq_restore(flags); 84946d3ceabSEric Dumazet 85046d3ceabSEric Dumazet list_for_each_safe(q, n, &list) { 85146d3ceabSEric Dumazet tp = list_entry(q, struct tcp_sock, tsq_node); 85246d3ceabSEric Dumazet list_del(&tp->tsq_node); 85346d3ceabSEric Dumazet 85446d3ceabSEric Dumazet sk = (struct sock *)tp; 8550a9648f1SEric Dumazet smp_mb__before_atomic(); 8567aa5470cSEric Dumazet clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags); 8577aa5470cSEric Dumazet 8586f458dfbSEric Dumazet tcp_tsq_handler(sk); 85946d3ceabSEric Dumazet sk_free(sk); 86046d3ceabSEric Dumazet } 86146d3ceabSEric Dumazet } 86246d3ceabSEric Dumazet 86340fc3423SEric Dumazet #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \ 86440fc3423SEric Dumazet TCPF_WRITE_TIMER_DEFERRED | \ 86540fc3423SEric Dumazet TCPF_DELACK_TIMER_DEFERRED | \ 86640fc3423SEric Dumazet TCPF_MTU_REDUCED_DEFERRED) 86746d3ceabSEric Dumazet /** 86846d3ceabSEric Dumazet * tcp_release_cb - tcp release_sock() callback 86946d3ceabSEric Dumazet * @sk: socket 87046d3ceabSEric Dumazet * 87146d3ceabSEric Dumazet * called from release_sock() to perform protocol dependent 87246d3ceabSEric Dumazet * actions before socket release. 87346d3ceabSEric Dumazet */ 87446d3ceabSEric Dumazet void tcp_release_cb(struct sock *sk) 87546d3ceabSEric Dumazet { 8766f458dfbSEric Dumazet unsigned long flags, nflags; 87746d3ceabSEric Dumazet 8786f458dfbSEric Dumazet /* perform an atomic operation only if at least one flag is set */ 8796f458dfbSEric Dumazet do { 8807aa5470cSEric Dumazet flags = sk->sk_tsq_flags; 8816f458dfbSEric Dumazet if (!(flags & TCP_DEFERRED_ALL)) 8826f458dfbSEric Dumazet return; 8836f458dfbSEric Dumazet nflags = flags & ~TCP_DEFERRED_ALL; 8847aa5470cSEric Dumazet } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags); 8856f458dfbSEric Dumazet 88673a6bab5SEric Dumazet if (flags & TCPF_TSQ_DEFERRED) { 88773a6bab5SEric Dumazet tcp_tsq_write(sk); 88873a6bab5SEric Dumazet __sock_put(sk); 88973a6bab5SEric Dumazet } 890c3f9b018SEric Dumazet /* Here begins the tricky part : 891c3f9b018SEric Dumazet * We are called from release_sock() with : 892c3f9b018SEric Dumazet * 1) BH disabled 893c3f9b018SEric Dumazet * 2) sk_lock.slock spinlock held 894c3f9b018SEric Dumazet * 3) socket owned by us (sk->sk_lock.owned == 1) 895c3f9b018SEric Dumazet * 896c3f9b018SEric Dumazet * But following code is meant to be called from BH handlers, 897c3f9b018SEric Dumazet * so we should keep BH disabled, but early release socket ownership 898c3f9b018SEric Dumazet */ 899c3f9b018SEric Dumazet sock_release_ownership(sk); 900c3f9b018SEric Dumazet 90140fc3423SEric Dumazet if (flags & TCPF_WRITE_TIMER_DEFERRED) { 9026f458dfbSEric Dumazet tcp_write_timer_handler(sk); 903144d56e9SEric Dumazet __sock_put(sk); 904144d56e9SEric Dumazet } 90540fc3423SEric Dumazet if (flags & TCPF_DELACK_TIMER_DEFERRED) { 9066f458dfbSEric Dumazet tcp_delack_timer_handler(sk); 907144d56e9SEric Dumazet __sock_put(sk); 908144d56e9SEric Dumazet } 90940fc3423SEric Dumazet if (flags & TCPF_MTU_REDUCED_DEFERRED) { 9104fab9071SNeal Cardwell inet_csk(sk)->icsk_af_ops->mtu_reduced(sk); 911144d56e9SEric Dumazet __sock_put(sk); 912144d56e9SEric Dumazet } 91346d3ceabSEric Dumazet } 91446d3ceabSEric Dumazet EXPORT_SYMBOL(tcp_release_cb); 91546d3ceabSEric Dumazet 91646d3ceabSEric Dumazet void __init tcp_tasklet_init(void) 91746d3ceabSEric Dumazet { 91846d3ceabSEric Dumazet int i; 91946d3ceabSEric Dumazet 92046d3ceabSEric Dumazet for_each_possible_cpu(i) { 92146d3ceabSEric Dumazet struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i); 92246d3ceabSEric Dumazet 92346d3ceabSEric Dumazet INIT_LIST_HEAD(&tsq->head); 92446d3ceabSEric Dumazet tasklet_init(&tsq->tasklet, 92546d3ceabSEric Dumazet tcp_tasklet_func, 92646d3ceabSEric Dumazet (unsigned long)tsq); 92746d3ceabSEric Dumazet } 92846d3ceabSEric Dumazet } 92946d3ceabSEric Dumazet 93046d3ceabSEric Dumazet /* 93146d3ceabSEric Dumazet * Write buffer destructor automatically called from kfree_skb. 9328e3bff96Sstephen hemminger * We can't xmit new skbs from this context, as we might already 93346d3ceabSEric Dumazet * hold qdisc lock. 93446d3ceabSEric Dumazet */ 935d6a4a104SEric Dumazet void tcp_wfree(struct sk_buff *skb) 93646d3ceabSEric Dumazet { 93746d3ceabSEric Dumazet struct sock *sk = skb->sk; 93846d3ceabSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 939408f0a6cSEric Dumazet unsigned long flags, nval, oval; 9409b462d02SEric Dumazet 9419b462d02SEric Dumazet /* Keep one reference on sk_wmem_alloc. 9429b462d02SEric Dumazet * Will be released by sk_free() from here or tcp_tasklet_func() 9439b462d02SEric Dumazet */ 94414afee4bSReshetova, Elena WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc)); 9459b462d02SEric Dumazet 9469b462d02SEric Dumazet /* If this softirq is serviced by ksoftirqd, we are likely under stress. 9479b462d02SEric Dumazet * Wait until our queues (qdisc + devices) are drained. 9489b462d02SEric Dumazet * This gives : 9499b462d02SEric Dumazet * - less callbacks to tcp_write_xmit(), reducing stress (batches) 9509b462d02SEric Dumazet * - chance for incoming ACK (processed by another cpu maybe) 9519b462d02SEric Dumazet * to migrate this flow (skb->ooo_okay will be eventually set) 9529b462d02SEric Dumazet */ 95314afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current) 9549b462d02SEric Dumazet goto out; 95546d3ceabSEric Dumazet 9567aa5470cSEric Dumazet for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) { 95746d3ceabSEric Dumazet struct tsq_tasklet *tsq; 958a9b204d1SEric Dumazet bool empty; 95946d3ceabSEric Dumazet 960408f0a6cSEric Dumazet if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED)) 961408f0a6cSEric Dumazet goto out; 962408f0a6cSEric Dumazet 96373a6bab5SEric Dumazet nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED; 9647aa5470cSEric Dumazet nval = cmpxchg(&sk->sk_tsq_flags, oval, nval); 965408f0a6cSEric Dumazet if (nval != oval) 966408f0a6cSEric Dumazet continue; 967408f0a6cSEric Dumazet 96846d3ceabSEric Dumazet /* queue this socket to tasklet queue */ 96946d3ceabSEric Dumazet local_irq_save(flags); 970903ceff7SChristoph Lameter tsq = this_cpu_ptr(&tsq_tasklet); 971a9b204d1SEric Dumazet empty = list_empty(&tsq->head); 97246d3ceabSEric Dumazet list_add(&tp->tsq_node, &tsq->head); 973a9b204d1SEric Dumazet if (empty) 97446d3ceabSEric Dumazet tasklet_schedule(&tsq->tasklet); 97546d3ceabSEric Dumazet local_irq_restore(flags); 9769b462d02SEric Dumazet return; 97746d3ceabSEric Dumazet } 9789b462d02SEric Dumazet out: 9799b462d02SEric Dumazet sk_free(sk); 98046d3ceabSEric Dumazet } 98146d3ceabSEric Dumazet 98273a6bab5SEric Dumazet /* Note: Called under soft irq. 98373a6bab5SEric Dumazet * We can call TCP stack right away, unless socket is owned by user. 984218af599SEric Dumazet */ 985218af599SEric Dumazet enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer) 986218af599SEric Dumazet { 987218af599SEric Dumazet struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer); 988218af599SEric Dumazet struct sock *sk = (struct sock *)tp; 989218af599SEric Dumazet 99073a6bab5SEric Dumazet tcp_tsq_handler(sk); 99173a6bab5SEric Dumazet sock_put(sk); 992218af599SEric Dumazet 993218af599SEric Dumazet return HRTIMER_NORESTART; 994218af599SEric Dumazet } 995218af599SEric Dumazet 996218af599SEric Dumazet static void tcp_internal_pacing(struct sock *sk, const struct sk_buff *skb) 997218af599SEric Dumazet { 998218af599SEric Dumazet u64 len_ns; 999218af599SEric Dumazet u32 rate; 1000218af599SEric Dumazet 1001218af599SEric Dumazet if (!tcp_needs_internal_pacing(sk)) 1002218af599SEric Dumazet return; 1003218af599SEric Dumazet rate = sk->sk_pacing_rate; 1004218af599SEric Dumazet if (!rate || rate == ~0U) 1005218af599SEric Dumazet return; 1006218af599SEric Dumazet 1007218af599SEric Dumazet len_ns = (u64)skb->len * NSEC_PER_SEC; 1008218af599SEric Dumazet do_div(len_ns, rate); 1009218af599SEric Dumazet hrtimer_start(&tcp_sk(sk)->pacing_timer, 1010218af599SEric Dumazet ktime_add_ns(ktime_get(), len_ns), 101173a6bab5SEric Dumazet HRTIMER_MODE_ABS_PINNED_SOFT); 101273a6bab5SEric Dumazet sock_hold(sk); 1013218af599SEric Dumazet } 1014218af599SEric Dumazet 1015e2080072SEric Dumazet static void tcp_update_skb_after_send(struct tcp_sock *tp, struct sk_buff *skb) 1016e2080072SEric Dumazet { 1017*d3edd06eSEric Dumazet skb->skb_mstamp_ns = tp->tcp_wstamp_ns; 1018e2080072SEric Dumazet list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue); 1019e2080072SEric Dumazet } 1020e2080072SEric Dumazet 10211da177e4SLinus Torvalds /* This routine actually transmits TCP packets queued in by 10221da177e4SLinus Torvalds * tcp_do_sendmsg(). This is used by both the initial 10231da177e4SLinus Torvalds * transmission and possible later retransmissions. 10241da177e4SLinus Torvalds * All SKB's seen here are completely headerless. It is our 10251da177e4SLinus Torvalds * job to build the TCP header, and pass the packet down to 10261da177e4SLinus Torvalds * IP so it can do the same plus pass the packet off to the 10271da177e4SLinus Torvalds * device. 10281da177e4SLinus Torvalds * 10291da177e4SLinus Torvalds * We are working here with either a clone of the original 10301da177e4SLinus Torvalds * SKB, or a fresh unique copy made by the retransmit engine. 10311da177e4SLinus Torvalds */ 10322987babbSYuchung Cheng static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, 10332987babbSYuchung Cheng int clone_it, gfp_t gfp_mask, u32 rcv_nxt) 10341da177e4SLinus Torvalds { 10356687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 1036dfb4b9dcSDavid S. Miller struct inet_sock *inet; 1037dfb4b9dcSDavid S. Miller struct tcp_sock *tp; 1038dfb4b9dcSDavid S. Miller struct tcp_skb_cb *tcb; 103933ad798cSAdam Langley struct tcp_out_options opts; 104095c96174SEric Dumazet unsigned int tcp_options_size, tcp_header_size; 10418c72c65bSEric Dumazet struct sk_buff *oskb = NULL; 1042cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 10431da177e4SLinus Torvalds struct tcphdr *th; 10441da177e4SLinus Torvalds int err; 10451da177e4SLinus Torvalds 1046dfb4b9dcSDavid S. Miller BUG_ON(!skb || !tcp_skb_pcount(skb)); 10476f094b9eSLawrence Brakmo tp = tcp_sk(sk); 1048dfb4b9dcSDavid S. Miller 1049ccdbb6e9SEric Dumazet if (clone_it) { 10506f094b9eSLawrence Brakmo TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq 10516f094b9eSLawrence Brakmo - tp->snd_una; 10528c72c65bSEric Dumazet oskb = skb; 1053e2080072SEric Dumazet 1054e2080072SEric Dumazet tcp_skb_tsorted_save(oskb) { 1055e2080072SEric Dumazet if (unlikely(skb_cloned(oskb))) 1056e2080072SEric Dumazet skb = pskb_copy(oskb, gfp_mask); 1057dfb4b9dcSDavid S. Miller else 1058e2080072SEric Dumazet skb = skb_clone(oskb, gfp_mask); 1059e2080072SEric Dumazet } tcp_skb_tsorted_restore(oskb); 1060e2080072SEric Dumazet 1061dfb4b9dcSDavid S. Miller if (unlikely(!skb)) 1062dfb4b9dcSDavid S. Miller return -ENOBUFS; 1063dfb4b9dcSDavid S. Miller } 1064*d3edd06eSEric Dumazet skb->skb_mstamp_ns = tp->tcp_wstamp_ns; 1065dfb4b9dcSDavid S. Miller 1066dfb4b9dcSDavid S. Miller inet = inet_sk(sk); 1067dfb4b9dcSDavid S. Miller tcb = TCP_SKB_CB(skb); 106833ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 10691da177e4SLinus Torvalds 10704de075e0SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) 107133ad798cSAdam Langley tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); 107233ad798cSAdam Langley else 107333ad798cSAdam Langley tcp_options_size = tcp_established_options(sk, skb, &opts, 107433ad798cSAdam Langley &md5); 107533ad798cSAdam Langley tcp_header_size = tcp_options_size + sizeof(struct tcphdr); 10761da177e4SLinus Torvalds 1077547669d4SEric Dumazet /* if no packet is in qdisc/device queue, then allow XPS to select 1078b2532eb9SEric Dumazet * another queue. We can be called from tcp_tsq_handler() 107973a6bab5SEric Dumazet * which holds one reference to sk. 1080b2532eb9SEric Dumazet * 1081b2532eb9SEric Dumazet * TODO: Ideally, in-flight pure ACK packets should not matter here. 1082b2532eb9SEric Dumazet * One way to get this would be to set skb->truesize = 2 on them. 1083547669d4SEric Dumazet */ 1084b2532eb9SEric Dumazet skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1); 10851da177e4SLinus Torvalds 108638ab52e8SEric Dumazet /* If we had to use memory reserve to allocate this skb, 108738ab52e8SEric Dumazet * this might cause drops if packet is looped back : 108838ab52e8SEric Dumazet * Other socket might not have SOCK_MEMALLOC. 108938ab52e8SEric Dumazet * Packets not looped back do not care about pfmemalloc. 109038ab52e8SEric Dumazet */ 109138ab52e8SEric Dumazet skb->pfmemalloc = 0; 109238ab52e8SEric Dumazet 1093aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 1094aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 109546d3ceabSEric Dumazet 109646d3ceabSEric Dumazet skb_orphan(skb); 109746d3ceabSEric Dumazet skb->sk = sk; 10981d2077acSEric Dumazet skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree; 1099b73c3d0eSTom Herbert skb_set_hash_from_sk(skb, sk); 110014afee4bSReshetova, Elena refcount_add(skb->truesize, &sk->sk_wmem_alloc); 11011da177e4SLinus Torvalds 1102c3a2e837SJulian Anastasov skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm); 1103c3a2e837SJulian Anastasov 11041da177e4SLinus Torvalds /* Build TCP header and checksum it. */ 1105ea1627c2SEric Dumazet th = (struct tcphdr *)skb->data; 1106c720c7e8SEric Dumazet th->source = inet->inet_sport; 1107c720c7e8SEric Dumazet th->dest = inet->inet_dport; 11081da177e4SLinus Torvalds th->seq = htonl(tcb->seq); 11092987babbSYuchung Cheng th->ack_seq = htonl(rcv_nxt); 1110df7a3b07SAl Viro *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | 11114de075e0SEric Dumazet tcb->tcp_flags); 1112dfb4b9dcSDavid S. Miller 11131da177e4SLinus Torvalds th->check = 0; 11141da177e4SLinus Torvalds th->urg_ptr = 0; 11151da177e4SLinus Torvalds 111633f5f57eSIlpo Järvinen /* The urg_mode check is necessary during a below snd_una win probe */ 11177691367dSHerbert Xu if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { 11187691367dSHerbert Xu if (before(tp->snd_up, tcb->seq + 0x10000)) { 11191da177e4SLinus Torvalds th->urg_ptr = htons(tp->snd_up - tcb->seq); 11201da177e4SLinus Torvalds th->urg = 1; 11217691367dSHerbert Xu } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { 11220eae88f3SEric Dumazet th->urg_ptr = htons(0xFFFF); 11237691367dSHerbert Xu th->urg = 1; 11247691367dSHerbert Xu } 11251da177e4SLinus Torvalds } 11261da177e4SLinus Torvalds 1127bd0388aeSWilliam Allen Simpson tcp_options_write((__be32 *)(th + 1), tp, &opts); 112851466a75SEric Dumazet skb_shinfo(skb)->gso_type = sk->sk_gso_type; 1129ea1627c2SEric Dumazet if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) { 1130ea1627c2SEric Dumazet th->window = htons(tcp_select_window(sk)); 1131ea1627c2SEric Dumazet tcp_ecn_send(sk, skb, th, tcp_header_size); 1132ea1627c2SEric Dumazet } else { 1133ea1627c2SEric Dumazet /* RFC1323: The window in SYN & SYN/ACK segments 1134ea1627c2SEric Dumazet * is never scaled. 1135ea1627c2SEric Dumazet */ 1136ea1627c2SEric Dumazet th->window = htons(min(tp->rcv_wnd, 65535U)); 1137ea1627c2SEric Dumazet } 1138cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 1139cfb6eeb4SYOSHIFUJI Hideaki /* Calculate the MD5 hash, as we have all we need now */ 1140cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 1141a465419bSEric Dumazet sk_nocaps_add(sk, NETIF_F_GSO_MASK); 1142bd0388aeSWilliam Allen Simpson tp->af_specific->calc_md5_hash(opts.hash_location, 114339f8e58eSEric Dumazet md5, sk, skb); 1144cfb6eeb4SYOSHIFUJI Hideaki } 1145cfb6eeb4SYOSHIFUJI Hideaki #endif 1146cfb6eeb4SYOSHIFUJI Hideaki 1147bb296246SHerbert Xu icsk->icsk_af_ops->send_check(sk, skb); 11481da177e4SLinus Torvalds 11494de075e0SEric Dumazet if (likely(tcb->tcp_flags & TCPHDR_ACK)) 115027cde44aSYuchung Cheng tcp_event_ack_sent(sk, tcp_skb_pcount(skb), rcv_nxt); 11511da177e4SLinus Torvalds 1152a44d6eacSMartin KaFai Lau if (skb->len != tcp_header_size) { 1153cf533ea5SEric Dumazet tcp_event_data_sent(tp, sk); 1154a44d6eacSMartin KaFai Lau tp->data_segs_out += tcp_skb_pcount(skb); 1155ba113c3aSWei Wang tp->bytes_sent += skb->len - tcp_header_size; 1156218af599SEric Dumazet tcp_internal_pacing(sk, skb); 1157a44d6eacSMartin KaFai Lau } 11581da177e4SLinus Torvalds 1159bd37a088SWei Yongjun if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) 1160aa2ea058STom Herbert TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, 1161aa2ea058STom Herbert tcp_skb_pcount(skb)); 11621da177e4SLinus Torvalds 11632efd055cSMarcelo Ricardo Leitner tp->segs_out += tcp_skb_pcount(skb); 1164f69ad292SEric Dumazet /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */ 1165cd7d8498SEric Dumazet skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb); 1166f69ad292SEric Dumazet skb_shinfo(skb)->gso_size = tcp_skb_mss(skb); 1167cd7d8498SEric Dumazet 1168*d3edd06eSEric Dumazet /* Leave earliest departure time in skb->tstamp (skb->skb_mstamp_ns) */ 1169971f10ecSEric Dumazet 1170971f10ecSEric Dumazet /* Cleanup our debris for IP stacks */ 1171971f10ecSEric Dumazet memset(skb->cb, 0, max(sizeof(struct inet_skb_parm), 1172971f10ecSEric Dumazet sizeof(struct inet6_skb_parm))); 1173971f10ecSEric Dumazet 1174b0270e91SEric Dumazet err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl); 11757faee5c0SEric Dumazet 11768c72c65bSEric Dumazet if (unlikely(err > 0)) { 11775ee2c941SChristoph Paasch tcp_enter_cwr(sk); 11788c72c65bSEric Dumazet err = net_xmit_eval(err); 11798c72c65bSEric Dumazet } 1180fc225799SEric Dumazet if (!err && oskb) { 1181e2080072SEric Dumazet tcp_update_skb_after_send(tp, oskb); 1182fc225799SEric Dumazet tcp_rate_skb_sent(sk, oskb); 1183fc225799SEric Dumazet } 11848c72c65bSEric Dumazet return err; 11851da177e4SLinus Torvalds } 11861da177e4SLinus Torvalds 11872987babbSYuchung Cheng static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, 11882987babbSYuchung Cheng gfp_t gfp_mask) 11892987babbSYuchung Cheng { 11902987babbSYuchung Cheng return __tcp_transmit_skb(sk, skb, clone_it, gfp_mask, 11912987babbSYuchung Cheng tcp_sk(sk)->rcv_nxt); 11922987babbSYuchung Cheng } 11932987babbSYuchung Cheng 119467edfef7SAndi Kleen /* This routine just queues the buffer for sending. 11951da177e4SLinus Torvalds * 11961da177e4SLinus Torvalds * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, 11971da177e4SLinus Torvalds * otherwise socket can stall. 11981da177e4SLinus Torvalds */ 11991da177e4SLinus Torvalds static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) 12001da177e4SLinus Torvalds { 12011da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 12021da177e4SLinus Torvalds 12031da177e4SLinus Torvalds /* Advance write_seq and place onto the write_queue. */ 12041da177e4SLinus Torvalds tp->write_seq = TCP_SKB_CB(skb)->end_seq; 1205f4a775d1SEric Dumazet __skb_header_release(skb); 1206fe067e8aSDavid S. Miller tcp_add_write_queue_tail(sk, skb); 12073ab224beSHideo Aoki sk->sk_wmem_queued += skb->truesize; 12083ab224beSHideo Aoki sk_mem_charge(sk, skb->truesize); 12091da177e4SLinus Torvalds } 12101da177e4SLinus Torvalds 121167edfef7SAndi Kleen /* Initialize TSO segments for a packet. */ 12125bbb432cSEric Dumazet static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now) 1213f6302d1dSDavid S. Miller { 12144a64fd6cSEric Dumazet if (skb->len <= mss_now) { 1215f6302d1dSDavid S. Miller /* Avoid the costly divide in the normal 1216f6302d1dSDavid S. Miller * non-TSO case. 1217f6302d1dSDavid S. Miller */ 1218cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, 1); 1219f69ad292SEric Dumazet TCP_SKB_CB(skb)->tcp_gso_size = 0; 1220f6302d1dSDavid S. Miller } else { 1221cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now)); 1222f69ad292SEric Dumazet TCP_SKB_CB(skb)->tcp_gso_size = mss_now; 12231da177e4SLinus Torvalds } 12241da177e4SLinus Torvalds } 12251da177e4SLinus Torvalds 1226797108d1SIlpo Järvinen /* Pcount in the middle of the write queue got changed, we need to do various 1227797108d1SIlpo Järvinen * tweaks to fix counters 1228797108d1SIlpo Järvinen */ 1229cf533ea5SEric Dumazet static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) 1230797108d1SIlpo Järvinen { 1231797108d1SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1232797108d1SIlpo Järvinen 1233797108d1SIlpo Järvinen tp->packets_out -= decr; 1234797108d1SIlpo Järvinen 1235797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 1236797108d1SIlpo Järvinen tp->sacked_out -= decr; 1237797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 1238797108d1SIlpo Järvinen tp->retrans_out -= decr; 1239797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 1240797108d1SIlpo Järvinen tp->lost_out -= decr; 1241797108d1SIlpo Järvinen 1242797108d1SIlpo Järvinen /* Reno case is special. Sigh... */ 1243797108d1SIlpo Järvinen if (tcp_is_reno(tp) && decr > 0) 1244797108d1SIlpo Järvinen tp->sacked_out -= min_t(u32, tp->sacked_out, decr); 1245797108d1SIlpo Järvinen 1246797108d1SIlpo Järvinen if (tp->lost_skb_hint && 1247797108d1SIlpo Järvinen before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && 1248713bafeaSYuchung Cheng (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) 1249797108d1SIlpo Järvinen tp->lost_cnt_hint -= decr; 1250797108d1SIlpo Järvinen 1251797108d1SIlpo Järvinen tcp_verify_left_out(tp); 1252797108d1SIlpo Järvinen } 1253797108d1SIlpo Järvinen 12540a2cf20cSSoheil Hassas Yeganeh static bool tcp_has_tx_tstamp(const struct sk_buff *skb) 12550a2cf20cSSoheil Hassas Yeganeh { 12560a2cf20cSSoheil Hassas Yeganeh return TCP_SKB_CB(skb)->txstamp_ack || 12570a2cf20cSSoheil Hassas Yeganeh (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP); 12580a2cf20cSSoheil Hassas Yeganeh } 12590a2cf20cSSoheil Hassas Yeganeh 1260490cc7d0SWillem de Bruijn static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2) 1261490cc7d0SWillem de Bruijn { 1262490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo = skb_shinfo(skb); 1263490cc7d0SWillem de Bruijn 12640a2cf20cSSoheil Hassas Yeganeh if (unlikely(tcp_has_tx_tstamp(skb)) && 1265490cc7d0SWillem de Bruijn !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) { 1266490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo2 = skb_shinfo(skb2); 1267490cc7d0SWillem de Bruijn u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP; 1268490cc7d0SWillem de Bruijn 1269490cc7d0SWillem de Bruijn shinfo->tx_flags &= ~tsflags; 1270490cc7d0SWillem de Bruijn shinfo2->tx_flags |= tsflags; 1271490cc7d0SWillem de Bruijn swap(shinfo->tskey, shinfo2->tskey); 1272b51e13faSMartin KaFai Lau TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack; 1273b51e13faSMartin KaFai Lau TCP_SKB_CB(skb)->txstamp_ack = 0; 1274490cc7d0SWillem de Bruijn } 1275490cc7d0SWillem de Bruijn } 1276490cc7d0SWillem de Bruijn 1277a166140eSMartin KaFai Lau static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2) 1278a166140eSMartin KaFai Lau { 1279a166140eSMartin KaFai Lau TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor; 1280a166140eSMartin KaFai Lau TCP_SKB_CB(skb)->eor = 0; 1281a166140eSMartin KaFai Lau } 1282a166140eSMartin KaFai Lau 128375c119afSEric Dumazet /* Insert buff after skb on the write or rtx queue of sk. */ 128475c119afSEric Dumazet static void tcp_insert_write_queue_after(struct sk_buff *skb, 128575c119afSEric Dumazet struct sk_buff *buff, 128675c119afSEric Dumazet struct sock *sk, 128775c119afSEric Dumazet enum tcp_queue tcp_queue) 128875c119afSEric Dumazet { 128975c119afSEric Dumazet if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE) 129075c119afSEric Dumazet __skb_queue_after(&sk->sk_write_queue, skb, buff); 129175c119afSEric Dumazet else 129275c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, buff); 129375c119afSEric Dumazet } 129475c119afSEric Dumazet 12951da177e4SLinus Torvalds /* Function to create two new TCP segments. Shrinks the given segment 12961da177e4SLinus Torvalds * to the specified size and appends a new segment with the rest of the 12971da177e4SLinus Torvalds * packet to the list. This won't be called frequently, I hope. 12981da177e4SLinus Torvalds * Remember, these are still headerless SKBs at this point. 12991da177e4SLinus Torvalds */ 130075c119afSEric Dumazet int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue, 130175c119afSEric Dumazet struct sk_buff *skb, u32 len, 13026cc55e09SOctavian Purdila unsigned int mss_now, gfp_t gfp) 13031da177e4SLinus Torvalds { 13041da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 13051da177e4SLinus Torvalds struct sk_buff *buff; 13066475be16SDavid S. Miller int nsize, old_factor; 1307b60b49eaSHerbert Xu int nlen; 13089ce01461SIlpo Järvinen u8 flags; 13091da177e4SLinus Torvalds 13102fceec13SIlpo Järvinen if (WARN_ON(len > skb->len)) 13112fceec13SIlpo Järvinen return -EINVAL; 13126a438bbeSStephen Hemminger 13131da177e4SLinus Torvalds nsize = skb_headlen(skb) - len; 13141da177e4SLinus Torvalds if (nsize < 0) 13151da177e4SLinus Torvalds nsize = 0; 13161da177e4SLinus Torvalds 13176cc55e09SOctavian Purdila if (skb_unclone(skb, gfp)) 13181da177e4SLinus Torvalds return -ENOMEM; 13191da177e4SLinus Torvalds 13201da177e4SLinus Torvalds /* Get a new skb... force flag on. */ 1321eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, nsize, gfp, true); 132251456b29SIan Morris if (!buff) 13231da177e4SLinus Torvalds return -ENOMEM; /* We'll just try again later. */ 1324ef5cb973SHerbert Xu 13253ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 13263ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1327b60b49eaSHerbert Xu nlen = skb->len - len - nsize; 1328b60b49eaSHerbert Xu buff->truesize += nlen; 1329b60b49eaSHerbert Xu skb->truesize -= nlen; 13301da177e4SLinus Torvalds 13311da177e4SLinus Torvalds /* Correct the sequence numbers. */ 13321da177e4SLinus Torvalds TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 13331da177e4SLinus Torvalds TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 13341da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 13351da177e4SLinus Torvalds 13361da177e4SLinus Torvalds /* PSH and FIN should only be set in the second packet. */ 13374de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 13384de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 13394de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1340e14c3cafSHerbert Xu TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; 1341a166140eSMartin KaFai Lau tcp_skb_fragment_eor(skb, buff); 13421da177e4SLinus Torvalds 13431da177e4SLinus Torvalds skb_split(skb, buff, len); 13441da177e4SLinus Torvalds 134598be9b12SEric Dumazet buff->ip_summed = CHECKSUM_PARTIAL; 13461da177e4SLinus Torvalds 1347a61bbcf2SPatrick McHardy buff->tstamp = skb->tstamp; 1348490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 13491da177e4SLinus Torvalds 13506475be16SDavid S. Miller old_factor = tcp_skb_pcount(skb); 13516475be16SDavid S. Miller 13521da177e4SLinus Torvalds /* Fix up tso_factor for both original and new SKB. */ 13535bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 13545bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 13551da177e4SLinus Torvalds 1356b9f64820SYuchung Cheng /* Update delivered info for the new segment */ 1357b9f64820SYuchung Cheng TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx; 1358b9f64820SYuchung Cheng 13596475be16SDavid S. Miller /* If this packet has been sent out already, we must 13606475be16SDavid S. Miller * adjust the various packet counters. 13616475be16SDavid S. Miller */ 1362cf0b450cSHerbert Xu if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 13636475be16SDavid S. Miller int diff = old_factor - tcp_skb_pcount(skb) - 13646475be16SDavid S. Miller tcp_skb_pcount(buff); 13651da177e4SLinus Torvalds 1366797108d1SIlpo Järvinen if (diff) 1367797108d1SIlpo Järvinen tcp_adjust_pcount(sk, skb, diff); 13681da177e4SLinus Torvalds } 13691da177e4SLinus Torvalds 13701da177e4SLinus Torvalds /* Link BUFF into the send queue. */ 1371f4a775d1SEric Dumazet __skb_header_release(buff); 137275c119afSEric Dumazet tcp_insert_write_queue_after(skb, buff, sk, tcp_queue); 1373f67971e6SEric Dumazet if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE) 1374e2080072SEric Dumazet list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor); 13751da177e4SLinus Torvalds 13761da177e4SLinus Torvalds return 0; 13771da177e4SLinus Torvalds } 13781da177e4SLinus Torvalds 1379f4d01666SEric Dumazet /* This is similar to __pskb_pull_tail(). The difference is that pulled 1380f4d01666SEric Dumazet * data is not copied, but immediately discarded. 13811da177e4SLinus Torvalds */ 13827162fb24SEric Dumazet static int __pskb_trim_head(struct sk_buff *skb, int len) 13831da177e4SLinus Torvalds { 13847b7fc97aSEric Dumazet struct skb_shared_info *shinfo; 13851da177e4SLinus Torvalds int i, k, eat; 13861da177e4SLinus Torvalds 13874fa48bf3SEric Dumazet eat = min_t(int, len, skb_headlen(skb)); 13884fa48bf3SEric Dumazet if (eat) { 13894fa48bf3SEric Dumazet __skb_pull(skb, eat); 13904fa48bf3SEric Dumazet len -= eat; 13914fa48bf3SEric Dumazet if (!len) 13927162fb24SEric Dumazet return 0; 13934fa48bf3SEric Dumazet } 13941da177e4SLinus Torvalds eat = len; 13951da177e4SLinus Torvalds k = 0; 13967b7fc97aSEric Dumazet shinfo = skb_shinfo(skb); 13977b7fc97aSEric Dumazet for (i = 0; i < shinfo->nr_frags; i++) { 13987b7fc97aSEric Dumazet int size = skb_frag_size(&shinfo->frags[i]); 13999e903e08SEric Dumazet 14009e903e08SEric Dumazet if (size <= eat) { 1401aff65da0SIan Campbell skb_frag_unref(skb, i); 14029e903e08SEric Dumazet eat -= size; 14031da177e4SLinus Torvalds } else { 14047b7fc97aSEric Dumazet shinfo->frags[k] = shinfo->frags[i]; 14051da177e4SLinus Torvalds if (eat) { 14067b7fc97aSEric Dumazet shinfo->frags[k].page_offset += eat; 14077b7fc97aSEric Dumazet skb_frag_size_sub(&shinfo->frags[k], eat); 14081da177e4SLinus Torvalds eat = 0; 14091da177e4SLinus Torvalds } 14101da177e4SLinus Torvalds k++; 14111da177e4SLinus Torvalds } 14121da177e4SLinus Torvalds } 14137b7fc97aSEric Dumazet shinfo->nr_frags = k; 14141da177e4SLinus Torvalds 14151da177e4SLinus Torvalds skb->data_len -= len; 14161da177e4SLinus Torvalds skb->len = skb->data_len; 14177162fb24SEric Dumazet return len; 14181da177e4SLinus Torvalds } 14191da177e4SLinus Torvalds 142067edfef7SAndi Kleen /* Remove acked data from a packet in the transmit queue. */ 14211da177e4SLinus Torvalds int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 14221da177e4SLinus Torvalds { 14237162fb24SEric Dumazet u32 delta_truesize; 14247162fb24SEric Dumazet 142514bbd6a5SPravin B Shelar if (skb_unclone(skb, GFP_ATOMIC)) 14261da177e4SLinus Torvalds return -ENOMEM; 14271da177e4SLinus Torvalds 14287162fb24SEric Dumazet delta_truesize = __pskb_trim_head(skb, len); 14291da177e4SLinus Torvalds 14301da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq += len; 143184fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 14321da177e4SLinus Torvalds 14337162fb24SEric Dumazet if (delta_truesize) { 14347162fb24SEric Dumazet skb->truesize -= delta_truesize; 14357162fb24SEric Dumazet sk->sk_wmem_queued -= delta_truesize; 14367162fb24SEric Dumazet sk_mem_uncharge(sk, delta_truesize); 14371da177e4SLinus Torvalds sock_set_flag(sk, SOCK_QUEUE_SHRUNK); 14387162fb24SEric Dumazet } 14391da177e4SLinus Torvalds 14405b35e1e6SNeal Cardwell /* Any change of skb->len requires recalculation of tso factor. */ 14411da177e4SLinus Torvalds if (tcp_skb_pcount(skb) > 1) 14425bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb)); 14431da177e4SLinus Torvalds 14441da177e4SLinus Torvalds return 0; 14451da177e4SLinus Torvalds } 14461da177e4SLinus Torvalds 14471b63edd6SYuchung Cheng /* Calculate MSS not accounting any TCP options. */ 14481b63edd6SYuchung Cheng static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu) 14495d424d5aSJohn Heffner { 1450cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1451cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 14525d424d5aSJohn Heffner int mss_now; 14535d424d5aSJohn Heffner 14545d424d5aSJohn Heffner /* Calculate base mss without TCP options: 14555d424d5aSJohn Heffner It is MMS_S - sizeof(tcphdr) of rfc1122 14565d424d5aSJohn Heffner */ 14575d424d5aSJohn Heffner mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 14585d424d5aSJohn Heffner 145967469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 146067469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 146167469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 146267469601SEric Dumazet 146367469601SEric Dumazet if (dst && dst_allfrag(dst)) 146467469601SEric Dumazet mss_now -= icsk->icsk_af_ops->net_frag_header_len; 146567469601SEric Dumazet } 146667469601SEric Dumazet 14675d424d5aSJohn Heffner /* Clamp it (mss_clamp does not include tcp options) */ 14685d424d5aSJohn Heffner if (mss_now > tp->rx_opt.mss_clamp) 14695d424d5aSJohn Heffner mss_now = tp->rx_opt.mss_clamp; 14705d424d5aSJohn Heffner 14715d424d5aSJohn Heffner /* Now subtract optional transport overhead */ 14725d424d5aSJohn Heffner mss_now -= icsk->icsk_ext_hdr_len; 14735d424d5aSJohn Heffner 14745d424d5aSJohn Heffner /* Then reserve room for full set of TCP options and 8 bytes of data */ 14755d424d5aSJohn Heffner if (mss_now < 48) 14765d424d5aSJohn Heffner mss_now = 48; 14775d424d5aSJohn Heffner return mss_now; 14785d424d5aSJohn Heffner } 14795d424d5aSJohn Heffner 14801b63edd6SYuchung Cheng /* Calculate MSS. Not accounting for SACKs here. */ 14811b63edd6SYuchung Cheng int tcp_mtu_to_mss(struct sock *sk, int pmtu) 14821b63edd6SYuchung Cheng { 14831b63edd6SYuchung Cheng /* Subtract TCP options size, not including SACKs */ 14841b63edd6SYuchung Cheng return __tcp_mtu_to_mss(sk, pmtu) - 14851b63edd6SYuchung Cheng (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr)); 14861b63edd6SYuchung Cheng } 14871b63edd6SYuchung Cheng 14885d424d5aSJohn Heffner /* Inverse of above */ 148967469601SEric Dumazet int tcp_mss_to_mtu(struct sock *sk, int mss) 14905d424d5aSJohn Heffner { 1491cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1492cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 14935d424d5aSJohn Heffner int mtu; 14945d424d5aSJohn Heffner 14955d424d5aSJohn Heffner mtu = mss + 14965d424d5aSJohn Heffner tp->tcp_header_len + 14975d424d5aSJohn Heffner icsk->icsk_ext_hdr_len + 14985d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 14995d424d5aSJohn Heffner 150067469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 150167469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 150267469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 150367469601SEric Dumazet 150467469601SEric Dumazet if (dst && dst_allfrag(dst)) 150567469601SEric Dumazet mtu += icsk->icsk_af_ops->net_frag_header_len; 150667469601SEric Dumazet } 15075d424d5aSJohn Heffner return mtu; 15085d424d5aSJohn Heffner } 1509556c6b46SNeal Cardwell EXPORT_SYMBOL(tcp_mss_to_mtu); 15105d424d5aSJohn Heffner 151167edfef7SAndi Kleen /* MTU probing init per socket */ 15125d424d5aSJohn Heffner void tcp_mtup_init(struct sock *sk) 15135d424d5aSJohn Heffner { 15145d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 15155d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 1516b0f9ca53SFan Du struct net *net = sock_net(sk); 15175d424d5aSJohn Heffner 1518b0f9ca53SFan Du icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1; 15195d424d5aSJohn Heffner icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 15205d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 1521b0f9ca53SFan Du icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss); 15225d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 152305cbc0dbSFan Du if (icsk->icsk_mtup.enabled) 1524c74df29aSEric Dumazet icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; 15255d424d5aSJohn Heffner } 15264bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_mtup_init); 15275d424d5aSJohn Heffner 15281da177e4SLinus Torvalds /* This function synchronize snd mss to current pmtu/exthdr set. 15291da177e4SLinus Torvalds 15301da177e4SLinus Torvalds tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 15311da177e4SLinus Torvalds for TCP options, but includes only bare TCP header. 15321da177e4SLinus Torvalds 15331da177e4SLinus Torvalds tp->rx_opt.mss_clamp is mss negotiated at connection setup. 1534caa20d9aSStephen Hemminger It is minimum of user_mss and mss received with SYN. 15351da177e4SLinus Torvalds It also does not include TCP options. 15361da177e4SLinus Torvalds 1537d83d8461SArnaldo Carvalho de Melo inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. 15381da177e4SLinus Torvalds 15391da177e4SLinus Torvalds tp->mss_cache is current effective sending mss, including 15401da177e4SLinus Torvalds all tcp options except for SACKs. It is evaluated, 15411da177e4SLinus Torvalds taking into account current pmtu, but never exceeds 15421da177e4SLinus Torvalds tp->rx_opt.mss_clamp. 15431da177e4SLinus Torvalds 15441da177e4SLinus Torvalds NOTE1. rfc1122 clearly states that advertised MSS 15451da177e4SLinus Torvalds DOES NOT include either tcp or ip options. 15461da177e4SLinus Torvalds 1547d83d8461SArnaldo Carvalho de Melo NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache 1548d83d8461SArnaldo Carvalho de Melo are READ ONLY outside this function. --ANK (980731) 15491da177e4SLinus Torvalds */ 15501da177e4SLinus Torvalds unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 15511da177e4SLinus Torvalds { 15521da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1553d83d8461SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 15545d424d5aSJohn Heffner int mss_now; 15551da177e4SLinus Torvalds 15565d424d5aSJohn Heffner if (icsk->icsk_mtup.search_high > pmtu) 15575d424d5aSJohn Heffner icsk->icsk_mtup.search_high = pmtu; 15581da177e4SLinus Torvalds 15595d424d5aSJohn Heffner mss_now = tcp_mtu_to_mss(sk, pmtu); 1560409d22b4SIlpo Järvinen mss_now = tcp_bound_to_half_wnd(tp, mss_now); 15611da177e4SLinus Torvalds 15621da177e4SLinus Torvalds /* And store cached results */ 1563d83d8461SArnaldo Carvalho de Melo icsk->icsk_pmtu_cookie = pmtu; 15645d424d5aSJohn Heffner if (icsk->icsk_mtup.enabled) 15655d424d5aSJohn Heffner mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); 1566c1b4a7e6SDavid S. Miller tp->mss_cache = mss_now; 15671da177e4SLinus Torvalds 15681da177e4SLinus Torvalds return mss_now; 15691da177e4SLinus Torvalds } 15704bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_sync_mss); 15711da177e4SLinus Torvalds 15721da177e4SLinus Torvalds /* Compute the current effective MSS, taking SACKs and IP options, 15731da177e4SLinus Torvalds * and even PMTU discovery events into account. 15741da177e4SLinus Torvalds */ 15750c54b85fSIlpo Järvinen unsigned int tcp_current_mss(struct sock *sk) 15761da177e4SLinus Torvalds { 1577cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1578cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1579c1b4a7e6SDavid S. Miller u32 mss_now; 158095c96174SEric Dumazet unsigned int header_len; 158133ad798cSAdam Langley struct tcp_out_options opts; 158233ad798cSAdam Langley struct tcp_md5sig_key *md5; 15831da177e4SLinus Torvalds 1584c1b4a7e6SDavid S. Miller mss_now = tp->mss_cache; 1585c1b4a7e6SDavid S. Miller 15861da177e4SLinus Torvalds if (dst) { 15871da177e4SLinus Torvalds u32 mtu = dst_mtu(dst); 1588d83d8461SArnaldo Carvalho de Melo if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 15891da177e4SLinus Torvalds mss_now = tcp_sync_mss(sk, mtu); 15901da177e4SLinus Torvalds } 15911da177e4SLinus Torvalds 159233ad798cSAdam Langley header_len = tcp_established_options(sk, NULL, &opts, &md5) + 159333ad798cSAdam Langley sizeof(struct tcphdr); 159433ad798cSAdam Langley /* The mss_cache is sized based on tp->tcp_header_len, which assumes 159533ad798cSAdam Langley * some common options. If this is an odd packet (because we have SACK 159633ad798cSAdam Langley * blocks etc) then our calculated header_len will be different, and 159733ad798cSAdam Langley * we have to adjust mss_now correspondingly */ 159833ad798cSAdam Langley if (header_len != tp->tcp_header_len) { 159933ad798cSAdam Langley int delta = (int) header_len - tp->tcp_header_len; 160033ad798cSAdam Langley mss_now -= delta; 160133ad798cSAdam Langley } 1602cfb6eeb4SYOSHIFUJI Hideaki 16031da177e4SLinus Torvalds return mss_now; 16041da177e4SLinus Torvalds } 16051da177e4SLinus Torvalds 160686fd14adSWeiping Pan /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. 160786fd14adSWeiping Pan * As additional protections, we do not touch cwnd in retransmission phases, 160886fd14adSWeiping Pan * and if application hit its sndbuf limit recently. 160986fd14adSWeiping Pan */ 161086fd14adSWeiping Pan static void tcp_cwnd_application_limited(struct sock *sk) 1611a762a980SDavid S. Miller { 16129e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1613a762a980SDavid S. Miller 161486fd14adSWeiping Pan if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && 161586fd14adSWeiping Pan sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 161686fd14adSWeiping Pan /* Limited by application or receiver window. */ 161786fd14adSWeiping Pan u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); 161886fd14adSWeiping Pan u32 win_used = max(tp->snd_cwnd_used, init_win); 161986fd14adSWeiping Pan if (win_used < tp->snd_cwnd) { 162086fd14adSWeiping Pan tp->snd_ssthresh = tcp_current_ssthresh(sk); 162186fd14adSWeiping Pan tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1; 162286fd14adSWeiping Pan } 162386fd14adSWeiping Pan tp->snd_cwnd_used = 0; 162486fd14adSWeiping Pan } 1625c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 162686fd14adSWeiping Pan } 162786fd14adSWeiping Pan 1628ca8a2263SNeal Cardwell static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited) 1629a762a980SDavid S. Miller { 16301b1fc3fdSWei Wang const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 1631a762a980SDavid S. Miller struct tcp_sock *tp = tcp_sk(sk); 1632a762a980SDavid S. Miller 1633ca8a2263SNeal Cardwell /* Track the maximum number of outstanding packets in each 1634ca8a2263SNeal Cardwell * window, and remember whether we were cwnd-limited then. 1635ca8a2263SNeal Cardwell */ 1636ca8a2263SNeal Cardwell if (!before(tp->snd_una, tp->max_packets_seq) || 1637ca8a2263SNeal Cardwell tp->packets_out > tp->max_packets_out) { 1638ca8a2263SNeal Cardwell tp->max_packets_out = tp->packets_out; 1639ca8a2263SNeal Cardwell tp->max_packets_seq = tp->snd_nxt; 1640ca8a2263SNeal Cardwell tp->is_cwnd_limited = is_cwnd_limited; 1641ca8a2263SNeal Cardwell } 1642e114a710SEric Dumazet 164324901551SEric Dumazet if (tcp_is_cwnd_limited(sk)) { 1644a762a980SDavid S. Miller /* Network is feed fully. */ 1645a762a980SDavid S. Miller tp->snd_cwnd_used = 0; 1646c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 1647a762a980SDavid S. Miller } else { 1648a762a980SDavid S. Miller /* Network starves. */ 1649a762a980SDavid S. Miller if (tp->packets_out > tp->snd_cwnd_used) 1650a762a980SDavid S. Miller tp->snd_cwnd_used = tp->packets_out; 1651a762a980SDavid S. Miller 1652b510f0d2SEric Dumazet if (sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle && 1653c2203cf7SEric Dumazet (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto && 16541b1fc3fdSWei Wang !ca_ops->cong_control) 1655a762a980SDavid S. Miller tcp_cwnd_application_limited(sk); 1656b0f71bd3SFrancis Yan 1657b0f71bd3SFrancis Yan /* The following conditions together indicate the starvation 1658b0f71bd3SFrancis Yan * is caused by insufficient sender buffer: 1659b0f71bd3SFrancis Yan * 1) just sent some data (see tcp_write_xmit) 1660b0f71bd3SFrancis Yan * 2) not cwnd limited (this else condition) 166175c119afSEric Dumazet * 3) no more data to send (tcp_write_queue_empty()) 1662b0f71bd3SFrancis Yan * 4) application is hitting buffer limit (SOCK_NOSPACE) 1663b0f71bd3SFrancis Yan */ 166475c119afSEric Dumazet if (tcp_write_queue_empty(sk) && sk->sk_socket && 1665b0f71bd3SFrancis Yan test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) && 1666b0f71bd3SFrancis Yan (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) 1667b0f71bd3SFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED); 1668a762a980SDavid S. Miller } 1669a762a980SDavid S. Miller } 1670a762a980SDavid S. Miller 1671d4589926SEric Dumazet /* Minshall's variant of the Nagle send check. */ 1672d4589926SEric Dumazet static bool tcp_minshall_check(const struct tcp_sock *tp) 1673d4589926SEric Dumazet { 1674d4589926SEric Dumazet return after(tp->snd_sml, tp->snd_una) && 1675d4589926SEric Dumazet !after(tp->snd_sml, tp->snd_nxt); 1676d4589926SEric Dumazet } 1677d4589926SEric Dumazet 1678d4589926SEric Dumazet /* Update snd_sml if this skb is under mss 1679d4589926SEric Dumazet * Note that a TSO packet might end with a sub-mss segment 1680d4589926SEric Dumazet * The test is really : 1681d4589926SEric Dumazet * if ((skb->len % mss) != 0) 1682d4589926SEric Dumazet * tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1683d4589926SEric Dumazet * But we can avoid doing the divide again given we already have 1684d4589926SEric Dumazet * skb_pcount = skb->len / mss_now 16850e3a4803SIlpo Järvinen */ 1686d4589926SEric Dumazet static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now, 1687d4589926SEric Dumazet const struct sk_buff *skb) 1688d4589926SEric Dumazet { 1689d4589926SEric Dumazet if (skb->len < tcp_skb_pcount(skb) * mss_now) 1690d4589926SEric Dumazet tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1691d4589926SEric Dumazet } 1692d4589926SEric Dumazet 1693d4589926SEric Dumazet /* Return false, if packet can be sent now without violation Nagle's rules: 1694d4589926SEric Dumazet * 1. It is full sized. (provided by caller in %partial bool) 1695d4589926SEric Dumazet * 2. Or it contains FIN. (already checked by caller) 1696d4589926SEric Dumazet * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. 1697d4589926SEric Dumazet * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 1698d4589926SEric Dumazet * With Minshall's modification: all sent small packets are ACKed. 1699d4589926SEric Dumazet */ 1700d4589926SEric Dumazet static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp, 1701cc93fc51SPeter Pan(潘卫平) int nonagle) 1702d4589926SEric Dumazet { 1703d4589926SEric Dumazet return partial && 1704d4589926SEric Dumazet ((nonagle & TCP_NAGLE_CORK) || 1705d4589926SEric Dumazet (!nonagle && tp->packets_out && tcp_minshall_check(tp))); 1706d4589926SEric Dumazet } 1707605ad7f1SEric Dumazet 1708605ad7f1SEric Dumazet /* Return how many segs we'd like on a TSO packet, 1709605ad7f1SEric Dumazet * to send one TSO packet per ms 1710605ad7f1SEric Dumazet */ 1711dcb8c9b4SEric Dumazet static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now, 17121b3878caSNeal Cardwell int min_tso_segs) 1713605ad7f1SEric Dumazet { 1714605ad7f1SEric Dumazet u32 bytes, segs; 1715605ad7f1SEric Dumazet 17163a9b76fdSEric Dumazet bytes = min(sk->sk_pacing_rate >> sk->sk_pacing_shift, 1717605ad7f1SEric Dumazet sk->sk_gso_max_size - 1 - MAX_TCP_HEADER); 1718605ad7f1SEric Dumazet 1719605ad7f1SEric Dumazet /* Goal is to send at least one packet per ms, 1720605ad7f1SEric Dumazet * not one big TSO packet every 100 ms. 1721605ad7f1SEric Dumazet * This preserves ACK clocking and is consistent 1722605ad7f1SEric Dumazet * with tcp_tso_should_defer() heuristic. 1723605ad7f1SEric Dumazet */ 17241b3878caSNeal Cardwell segs = max_t(u32, bytes / mss_now, min_tso_segs); 1725605ad7f1SEric Dumazet 1726350c9f48SEric Dumazet return segs; 1727605ad7f1SEric Dumazet } 1728605ad7f1SEric Dumazet 1729ed6e7268SNeal Cardwell /* Return the number of segments we want in the skb we are transmitting. 1730ed6e7268SNeal Cardwell * See if congestion control module wants to decide; otherwise, autosize. 1731ed6e7268SNeal Cardwell */ 1732ed6e7268SNeal Cardwell static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now) 1733ed6e7268SNeal Cardwell { 1734ed6e7268SNeal Cardwell const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 1735dcb8c9b4SEric Dumazet u32 min_tso, tso_segs; 1736ed6e7268SNeal Cardwell 1737dcb8c9b4SEric Dumazet min_tso = ca_ops->min_tso_segs ? 1738dcb8c9b4SEric Dumazet ca_ops->min_tso_segs(sk) : 1739dcb8c9b4SEric Dumazet sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs; 1740dcb8c9b4SEric Dumazet 1741dcb8c9b4SEric Dumazet tso_segs = tcp_tso_autosize(sk, mss_now, min_tso); 1742350c9f48SEric Dumazet return min_t(u32, tso_segs, sk->sk_gso_max_segs); 1743ed6e7268SNeal Cardwell } 1744ed6e7268SNeal Cardwell 1745d4589926SEric Dumazet /* Returns the portion of skb which can be sent right away */ 1746d4589926SEric Dumazet static unsigned int tcp_mss_split_point(const struct sock *sk, 1747d4589926SEric Dumazet const struct sk_buff *skb, 1748d4589926SEric Dumazet unsigned int mss_now, 1749d4589926SEric Dumazet unsigned int max_segs, 1750d4589926SEric Dumazet int nonagle) 1751c1b4a7e6SDavid S. Miller { 1752cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1753d4589926SEric Dumazet u32 partial, needed, window, max_len; 1754c1b4a7e6SDavid S. Miller 175590840defSIlpo Järvinen window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 17561485348dSBen Hutchings max_len = mss_now * max_segs; 17570e3a4803SIlpo Järvinen 17581485348dSBen Hutchings if (likely(max_len <= window && skb != tcp_write_queue_tail(sk))) 17591485348dSBen Hutchings return max_len; 17600e3a4803SIlpo Järvinen 17615ea3a748SIlpo Järvinen needed = min(skb->len, window); 17625ea3a748SIlpo Järvinen 17631485348dSBen Hutchings if (max_len <= needed) 17641485348dSBen Hutchings return max_len; 17650e3a4803SIlpo Järvinen 1766d4589926SEric Dumazet partial = needed % mss_now; 1767d4589926SEric Dumazet /* If last segment is not a full MSS, check if Nagle rules allow us 1768d4589926SEric Dumazet * to include this last segment in this skb. 1769d4589926SEric Dumazet * Otherwise, we'll split the skb at last MSS boundary 1770d4589926SEric Dumazet */ 1771cc93fc51SPeter Pan(潘卫平) if (tcp_nagle_check(partial != 0, tp, nonagle)) 1772d4589926SEric Dumazet return needed - partial; 1773d4589926SEric Dumazet 1774d4589926SEric Dumazet return needed; 1775c1b4a7e6SDavid S. Miller } 1776c1b4a7e6SDavid S. Miller 1777c1b4a7e6SDavid S. Miller /* Can at least one segment of SKB be sent right now, according to the 1778c1b4a7e6SDavid S. Miller * congestion window rules? If so, return how many segments are allowed. 1779c1b4a7e6SDavid S. Miller */ 1780cf533ea5SEric Dumazet static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, 1781cf533ea5SEric Dumazet const struct sk_buff *skb) 1782c1b4a7e6SDavid S. Miller { 1783d649a7a8SEric Dumazet u32 in_flight, cwnd, halfcwnd; 1784c1b4a7e6SDavid S. Miller 1785c1b4a7e6SDavid S. Miller /* Don't be strict about the congestion window for the final FIN. */ 17864de075e0SEric Dumazet if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && 17874de075e0SEric Dumazet tcp_skb_pcount(skb) == 1) 1788c1b4a7e6SDavid S. Miller return 1; 1789c1b4a7e6SDavid S. Miller 1790c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1791c1b4a7e6SDavid S. Miller cwnd = tp->snd_cwnd; 1792d649a7a8SEric Dumazet if (in_flight >= cwnd) 1793c1b4a7e6SDavid S. Miller return 0; 1794d649a7a8SEric Dumazet 1795d649a7a8SEric Dumazet /* For better scheduling, ensure we have at least 1796d649a7a8SEric Dumazet * 2 GSO packets in flight. 1797d649a7a8SEric Dumazet */ 1798d649a7a8SEric Dumazet halfcwnd = max(cwnd >> 1, 1U); 1799d649a7a8SEric Dumazet return min(halfcwnd, cwnd - in_flight); 1800c1b4a7e6SDavid S. Miller } 1801c1b4a7e6SDavid S. Miller 1802b595076aSUwe Kleine-König /* Initialize TSO state of a skb. 180367edfef7SAndi Kleen * This must be invoked the first time we consider transmitting 1804c1b4a7e6SDavid S. Miller * SKB onto the wire. 1805c1b4a7e6SDavid S. Miller */ 18065bbb432cSEric Dumazet static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now) 1807c1b4a7e6SDavid S. Miller { 1808c1b4a7e6SDavid S. Miller int tso_segs = tcp_skb_pcount(skb); 1809c1b4a7e6SDavid S. Miller 1810f8269a49SIlpo Järvinen if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { 18115bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 1812c1b4a7e6SDavid S. Miller tso_segs = tcp_skb_pcount(skb); 1813c1b4a7e6SDavid S. Miller } 1814c1b4a7e6SDavid S. Miller return tso_segs; 1815c1b4a7e6SDavid S. Miller } 1816c1b4a7e6SDavid S. Miller 1817c1b4a7e6SDavid S. Miller 1818a2a385d6SEric Dumazet /* Return true if the Nagle test allows this packet to be 1819c1b4a7e6SDavid S. Miller * sent now. 1820c1b4a7e6SDavid S. Miller */ 1821a2a385d6SEric Dumazet static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, 1822c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1823c1b4a7e6SDavid S. Miller { 1824c1b4a7e6SDavid S. Miller /* Nagle rule does not apply to frames, which sit in the middle of the 1825c1b4a7e6SDavid S. Miller * write_queue (they have no chances to get new data). 1826c1b4a7e6SDavid S. Miller * 1827c1b4a7e6SDavid S. Miller * This is implemented in the callers, where they modify the 'nonagle' 1828c1b4a7e6SDavid S. Miller * argument based upon the location of SKB in the send queue. 1829c1b4a7e6SDavid S. Miller */ 1830c1b4a7e6SDavid S. Miller if (nonagle & TCP_NAGLE_PUSH) 1831a2a385d6SEric Dumazet return true; 1832c1b4a7e6SDavid S. Miller 18339b44190dSYuchung Cheng /* Don't use the nagle rule for urgent data (or for the final FIN). */ 18349b44190dSYuchung Cheng if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) 1835a2a385d6SEric Dumazet return true; 1836c1b4a7e6SDavid S. Miller 1837cc93fc51SPeter Pan(潘卫平) if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle)) 1838a2a385d6SEric Dumazet return true; 1839c1b4a7e6SDavid S. Miller 1840a2a385d6SEric Dumazet return false; 1841c1b4a7e6SDavid S. Miller } 1842c1b4a7e6SDavid S. Miller 1843c1b4a7e6SDavid S. Miller /* Does at least the first segment of SKB fit into the send window? */ 1844a2a385d6SEric Dumazet static bool tcp_snd_wnd_test(const struct tcp_sock *tp, 1845a2a385d6SEric Dumazet const struct sk_buff *skb, 1846056834d9SIlpo Järvinen unsigned int cur_mss) 1847c1b4a7e6SDavid S. Miller { 1848c1b4a7e6SDavid S. Miller u32 end_seq = TCP_SKB_CB(skb)->end_seq; 1849c1b4a7e6SDavid S. Miller 1850c1b4a7e6SDavid S. Miller if (skb->len > cur_mss) 1851c1b4a7e6SDavid S. Miller end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 1852c1b4a7e6SDavid S. Miller 185390840defSIlpo Järvinen return !after(end_seq, tcp_wnd_end(tp)); 1854c1b4a7e6SDavid S. Miller } 1855c1b4a7e6SDavid S. Miller 1856c1b4a7e6SDavid S. Miller /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 1857c1b4a7e6SDavid S. Miller * which is put after SKB on the list. It is very much like 1858c1b4a7e6SDavid S. Miller * tcp_fragment() except that it may make several kinds of assumptions 1859c1b4a7e6SDavid S. Miller * in order to speed up the splitting operation. In particular, we 1860c1b4a7e6SDavid S. Miller * know that all the data is in scatter-gather pages, and that the 1861c1b4a7e6SDavid S. Miller * packet has never been sent out before (and thus is not cloned). 1862c1b4a7e6SDavid S. Miller */ 186375c119afSEric Dumazet static int tso_fragment(struct sock *sk, enum tcp_queue tcp_queue, 186475c119afSEric Dumazet struct sk_buff *skb, unsigned int len, 1865c4ead4c5SEric Dumazet unsigned int mss_now, gfp_t gfp) 1866c1b4a7e6SDavid S. Miller { 1867c1b4a7e6SDavid S. Miller struct sk_buff *buff; 1868c1b4a7e6SDavid S. Miller int nlen = skb->len - len; 18699ce01461SIlpo Järvinen u8 flags; 1870c1b4a7e6SDavid S. Miller 1871c1b4a7e6SDavid S. Miller /* All of a TSO frame must be composed of paged data. */ 1872c8ac3774SHerbert Xu if (skb->len != skb->data_len) 187375c119afSEric Dumazet return tcp_fragment(sk, tcp_queue, skb, len, mss_now, gfp); 1874c1b4a7e6SDavid S. Miller 1875eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, gfp, true); 187651456b29SIan Morris if (unlikely(!buff)) 1877c1b4a7e6SDavid S. Miller return -ENOMEM; 1878c1b4a7e6SDavid S. Miller 18793ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 18803ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1881b60b49eaSHerbert Xu buff->truesize += nlen; 1882c1b4a7e6SDavid S. Miller skb->truesize -= nlen; 1883c1b4a7e6SDavid S. Miller 1884c1b4a7e6SDavid S. Miller /* Correct the sequence numbers. */ 1885c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 1886c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 1887c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 1888c1b4a7e6SDavid S. Miller 1889c1b4a7e6SDavid S. Miller /* PSH and FIN should only be set in the second packet. */ 18904de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 18914de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 18924de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1893c1b4a7e6SDavid S. Miller 1894c1b4a7e6SDavid S. Miller /* This packet was never sent out yet, so no SACK bits. */ 1895c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->sacked = 0; 1896c1b4a7e6SDavid S. Miller 1897a166140eSMartin KaFai Lau tcp_skb_fragment_eor(skb, buff); 1898a166140eSMartin KaFai Lau 189998be9b12SEric Dumazet buff->ip_summed = CHECKSUM_PARTIAL; 1900c1b4a7e6SDavid S. Miller skb_split(skb, buff, len); 1901490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 1902c1b4a7e6SDavid S. Miller 1903c1b4a7e6SDavid S. Miller /* Fix up tso_factor for both original and new SKB. */ 19045bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 19055bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 1906c1b4a7e6SDavid S. Miller 1907c1b4a7e6SDavid S. Miller /* Link BUFF into the send queue. */ 1908f4a775d1SEric Dumazet __skb_header_release(buff); 190975c119afSEric Dumazet tcp_insert_write_queue_after(skb, buff, sk, tcp_queue); 1910c1b4a7e6SDavid S. Miller 1911c1b4a7e6SDavid S. Miller return 0; 1912c1b4a7e6SDavid S. Miller } 1913c1b4a7e6SDavid S. Miller 1914c1b4a7e6SDavid S. Miller /* Try to defer sending, if possible, in order to minimize the amount 1915c1b4a7e6SDavid S. Miller * of TSO splitting we do. View it as a kind of TSO Nagle test. 1916c1b4a7e6SDavid S. Miller * 1917c1b4a7e6SDavid S. Miller * This algorithm is from John Heffner. 1918c1b4a7e6SDavid S. Miller */ 1919ca8a2263SNeal Cardwell static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb, 1920605ad7f1SEric Dumazet bool *is_cwnd_limited, u32 max_segs) 1921c1b4a7e6SDavid S. Miller { 19226687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 192350c8339eSEric Dumazet u32 age, send_win, cong_win, limit, in_flight; 192450c8339eSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 192550c8339eSEric Dumazet struct sk_buff *head; 1926ad9f4f50SEric Dumazet int win_divisor; 1927c1b4a7e6SDavid S. Miller 19284de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 1929ae8064acSJohn Heffner goto send_now; 1930c1b4a7e6SDavid S. Miller 193199d7662aSEric Dumazet if (icsk->icsk_ca_state >= TCP_CA_Recovery) 1932ae8064acSJohn Heffner goto send_now; 1933ae8064acSJohn Heffner 19345f852eb5SEric Dumazet /* Avoid bursty behavior by allowing defer 19355f852eb5SEric Dumazet * only if the last write was recent. 19365f852eb5SEric Dumazet */ 1937d635fbe2SEric Dumazet if ((s32)(tcp_jiffies32 - tp->lsndtime) > 0) 1938ae8064acSJohn Heffner goto send_now; 1939908a75c1SDavid S. Miller 1940c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1941c1b4a7e6SDavid S. Miller 1942c8c9aeb5SStefano Brivio BUG_ON(tcp_skb_pcount(skb) <= 1); 1943c8c9aeb5SStefano Brivio BUG_ON(tp->snd_cwnd <= in_flight); 1944c1b4a7e6SDavid S. Miller 194590840defSIlpo Järvinen send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1946c1b4a7e6SDavid S. Miller 1947c1b4a7e6SDavid S. Miller /* From in_flight test above, we know that cwnd > in_flight. */ 1948c1b4a7e6SDavid S. Miller cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; 1949c1b4a7e6SDavid S. Miller 1950c1b4a7e6SDavid S. Miller limit = min(send_win, cong_win); 1951c1b4a7e6SDavid S. Miller 1952ba244fe9SDavid S. Miller /* If a full-sized TSO skb can be sent, do it. */ 1953605ad7f1SEric Dumazet if (limit >= max_segs * tp->mss_cache) 1954ae8064acSJohn Heffner goto send_now; 1955ba244fe9SDavid S. Miller 195662ad2761SIlpo Järvinen /* Middle in queue won't get any more data, full sendable already? */ 195762ad2761SIlpo Järvinen if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) 195862ad2761SIlpo Järvinen goto send_now; 195962ad2761SIlpo Järvinen 19605bbcc0f5SLinus Torvalds win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor); 1961ad9f4f50SEric Dumazet if (win_divisor) { 1962c1b4a7e6SDavid S. Miller u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); 1963c1b4a7e6SDavid S. Miller 1964c1b4a7e6SDavid S. Miller /* If at least some fraction of a window is available, 1965c1b4a7e6SDavid S. Miller * just use it. 1966c1b4a7e6SDavid S. Miller */ 1967ad9f4f50SEric Dumazet chunk /= win_divisor; 1968c1b4a7e6SDavid S. Miller if (limit >= chunk) 1969ae8064acSJohn Heffner goto send_now; 1970c1b4a7e6SDavid S. Miller } else { 1971c1b4a7e6SDavid S. Miller /* Different approach, try not to defer past a single 1972c1b4a7e6SDavid S. Miller * ACK. Receiver should ACK every other full sized 1973c1b4a7e6SDavid S. Miller * frame, so if we have space for more than 3 frames 1974c1b4a7e6SDavid S. Miller * then send now. 1975c1b4a7e6SDavid S. Miller */ 19766b5a5c0dSNeal Cardwell if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) 1977ae8064acSJohn Heffner goto send_now; 1978c1b4a7e6SDavid S. Miller } 1979c1b4a7e6SDavid S. Miller 198075c119afSEric Dumazet /* TODO : use tsorted_sent_queue ? */ 198175c119afSEric Dumazet head = tcp_rtx_queue_head(sk); 198275c119afSEric Dumazet if (!head) 198375c119afSEric Dumazet goto send_now; 19842fd66ffbSEric Dumazet age = tcp_stamp_us_delta(tp->tcp_mstamp, tcp_skb_timestamp_us(head)); 198550c8339eSEric Dumazet /* If next ACK is likely to come too late (half srtt), do not defer */ 198650c8339eSEric Dumazet if (age < (tp->srtt_us >> 4)) 198750c8339eSEric Dumazet goto send_now; 198850c8339eSEric Dumazet 19895f852eb5SEric Dumazet /* Ok, it looks like it is advisable to defer. */ 1990ae8064acSJohn Heffner 1991d2e1339fSBendik Rønning Opstad if (cong_win < send_win && cong_win <= skb->len) 1992ca8a2263SNeal Cardwell *is_cwnd_limited = true; 1993ca8a2263SNeal Cardwell 1994a2a385d6SEric Dumazet return true; 1995ae8064acSJohn Heffner 1996ae8064acSJohn Heffner send_now: 1997a2a385d6SEric Dumazet return false; 1998c1b4a7e6SDavid S. Miller } 1999c1b4a7e6SDavid S. Miller 200005cbc0dbSFan Du static inline void tcp_mtu_check_reprobe(struct sock *sk) 200105cbc0dbSFan Du { 200205cbc0dbSFan Du struct inet_connection_sock *icsk = inet_csk(sk); 200305cbc0dbSFan Du struct tcp_sock *tp = tcp_sk(sk); 200405cbc0dbSFan Du struct net *net = sock_net(sk); 200505cbc0dbSFan Du u32 interval; 200605cbc0dbSFan Du s32 delta; 200705cbc0dbSFan Du 200805cbc0dbSFan Du interval = net->ipv4.sysctl_tcp_probe_interval; 2009c74df29aSEric Dumazet delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp; 201005cbc0dbSFan Du if (unlikely(delta >= interval * HZ)) { 201105cbc0dbSFan Du int mss = tcp_current_mss(sk); 201205cbc0dbSFan Du 201305cbc0dbSFan Du /* Update current search range */ 201405cbc0dbSFan Du icsk->icsk_mtup.probe_size = 0; 201505cbc0dbSFan Du icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + 201605cbc0dbSFan Du sizeof(struct tcphdr) + 201705cbc0dbSFan Du icsk->icsk_af_ops->net_header_len; 201805cbc0dbSFan Du icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss); 201905cbc0dbSFan Du 202005cbc0dbSFan Du /* Update probe time stamp */ 2021c74df29aSEric Dumazet icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; 202205cbc0dbSFan Du } 202305cbc0dbSFan Du } 202405cbc0dbSFan Du 2025808cf9e3SIlya Lesokhin static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len) 2026808cf9e3SIlya Lesokhin { 2027808cf9e3SIlya Lesokhin struct sk_buff *skb, *next; 2028808cf9e3SIlya Lesokhin 2029808cf9e3SIlya Lesokhin skb = tcp_send_head(sk); 2030808cf9e3SIlya Lesokhin tcp_for_write_queue_from_safe(skb, next, sk) { 2031808cf9e3SIlya Lesokhin if (len <= skb->len) 2032808cf9e3SIlya Lesokhin break; 2033808cf9e3SIlya Lesokhin 2034808cf9e3SIlya Lesokhin if (unlikely(TCP_SKB_CB(skb)->eor)) 2035808cf9e3SIlya Lesokhin return false; 2036808cf9e3SIlya Lesokhin 2037808cf9e3SIlya Lesokhin len -= skb->len; 2038808cf9e3SIlya Lesokhin } 2039808cf9e3SIlya Lesokhin 2040808cf9e3SIlya Lesokhin return true; 2041808cf9e3SIlya Lesokhin } 2042808cf9e3SIlya Lesokhin 20435d424d5aSJohn Heffner /* Create a new MTU probe if we are ready. 204467edfef7SAndi Kleen * MTU probe is regularly attempting to increase the path MTU by 204567edfef7SAndi Kleen * deliberately sending larger packets. This discovers routing 204667edfef7SAndi Kleen * changes resulting in larger path MTUs. 204767edfef7SAndi Kleen * 20485d424d5aSJohn Heffner * Returns 0 if we should wait to probe (no cwnd available), 20495d424d5aSJohn Heffner * 1 if a probe was sent, 2050056834d9SIlpo Järvinen * -1 otherwise 2051056834d9SIlpo Järvinen */ 20525d424d5aSJohn Heffner static int tcp_mtu_probe(struct sock *sk) 20535d424d5aSJohn Heffner { 20545d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 205512a59abcSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 20565d424d5aSJohn Heffner struct sk_buff *skb, *nskb, *next; 20576b58e0a5SFan Du struct net *net = sock_net(sk); 20585d424d5aSJohn Heffner int probe_size; 205991cc17c0SIlpo Järvinen int size_needed; 206012a59abcSEric Dumazet int copy, len; 20615d424d5aSJohn Heffner int mss_now; 20626b58e0a5SFan Du int interval; 20635d424d5aSJohn Heffner 20645d424d5aSJohn Heffner /* Not currently probing/verifying, 20655d424d5aSJohn Heffner * not in recovery, 20665d424d5aSJohn Heffner * have enough cwnd, and 206712a59abcSEric Dumazet * not SACKing (the variable headers throw things off) 206812a59abcSEric Dumazet */ 206912a59abcSEric Dumazet if (likely(!icsk->icsk_mtup.enabled || 20705d424d5aSJohn Heffner icsk->icsk_mtup.probe_size || 20715d424d5aSJohn Heffner inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 20725d424d5aSJohn Heffner tp->snd_cwnd < 11 || 207312a59abcSEric Dumazet tp->rx_opt.num_sacks || tp->rx_opt.dsack)) 20745d424d5aSJohn Heffner return -1; 20755d424d5aSJohn Heffner 20766b58e0a5SFan Du /* Use binary search for probe_size between tcp_mss_base, 20776b58e0a5SFan Du * and current mss_clamp. if (search_high - search_low) 20786b58e0a5SFan Du * smaller than a threshold, backoff from probing. 20796b58e0a5SFan Du */ 20800c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 20816b58e0a5SFan Du probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high + 20826b58e0a5SFan Du icsk->icsk_mtup.search_low) >> 1); 208391cc17c0SIlpo Järvinen size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; 20846b58e0a5SFan Du interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low; 208505cbc0dbSFan Du /* When misfortune happens, we are reprobing actively, 208605cbc0dbSFan Du * and then reprobe timer has expired. We stick with current 208705cbc0dbSFan Du * probing process by not resetting search range to its orignal. 208805cbc0dbSFan Du */ 20896b58e0a5SFan Du if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) || 209005cbc0dbSFan Du interval < net->ipv4.sysctl_tcp_probe_threshold) { 209105cbc0dbSFan Du /* Check whether enough time has elaplased for 209205cbc0dbSFan Du * another round of probing. 209305cbc0dbSFan Du */ 209405cbc0dbSFan Du tcp_mtu_check_reprobe(sk); 20955d424d5aSJohn Heffner return -1; 20965d424d5aSJohn Heffner } 20975d424d5aSJohn Heffner 20985d424d5aSJohn Heffner /* Have enough data in the send queue to probe? */ 20997f9c33e5SIlpo Järvinen if (tp->write_seq - tp->snd_nxt < size_needed) 21005d424d5aSJohn Heffner return -1; 21015d424d5aSJohn Heffner 210291cc17c0SIlpo Järvinen if (tp->snd_wnd < size_needed) 21035d424d5aSJohn Heffner return -1; 210490840defSIlpo Järvinen if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) 21055d424d5aSJohn Heffner return 0; 21065d424d5aSJohn Heffner 2107d67c58e9SIlpo Järvinen /* Do we need to wait to drain cwnd? With none in flight, don't stall */ 2108d67c58e9SIlpo Järvinen if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { 2109d67c58e9SIlpo Järvinen if (!tcp_packets_in_flight(tp)) 21105d424d5aSJohn Heffner return -1; 21115d424d5aSJohn Heffner else 21125d424d5aSJohn Heffner return 0; 21135d424d5aSJohn Heffner } 21145d424d5aSJohn Heffner 2115808cf9e3SIlya Lesokhin if (!tcp_can_coalesce_send_queue_head(sk, probe_size)) 2116808cf9e3SIlya Lesokhin return -1; 2117808cf9e3SIlya Lesokhin 21185d424d5aSJohn Heffner /* We're allowed to probe. Build it now. */ 2119eb934478SEric Dumazet nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false); 212051456b29SIan Morris if (!nskb) 21215d424d5aSJohn Heffner return -1; 21223ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 21233ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 21245d424d5aSJohn Heffner 2125fe067e8aSDavid S. Miller skb = tcp_send_head(sk); 21265d424d5aSJohn Heffner 21275d424d5aSJohn Heffner TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 21285d424d5aSJohn Heffner TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 21294de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; 21305d424d5aSJohn Heffner TCP_SKB_CB(nskb)->sacked = 0; 21315d424d5aSJohn Heffner nskb->csum = 0; 213298be9b12SEric Dumazet nskb->ip_summed = CHECKSUM_PARTIAL; 21335d424d5aSJohn Heffner 213450c4817eSIlpo Järvinen tcp_insert_write_queue_before(nskb, skb, sk); 21352b7cda9cSEric Dumazet tcp_highest_sack_replace(sk, skb, nskb); 213650c4817eSIlpo Järvinen 21375d424d5aSJohn Heffner len = 0; 2138234b6860SIlpo Järvinen tcp_for_write_queue_from_safe(skb, next, sk) { 21395d424d5aSJohn Heffner copy = min_t(int, skb->len, probe_size - len); 21405d424d5aSJohn Heffner skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); 21415d424d5aSJohn Heffner 21425d424d5aSJohn Heffner if (skb->len <= copy) { 21435d424d5aSJohn Heffner /* We've eaten all the data from this skb. 21445d424d5aSJohn Heffner * Throw it away. */ 21454de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 2146808cf9e3SIlya Lesokhin /* If this is the last SKB we copy and eor is set 2147808cf9e3SIlya Lesokhin * we need to propagate it to the new skb. 2148808cf9e3SIlya Lesokhin */ 2149808cf9e3SIlya Lesokhin TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor; 2150fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 21513ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 21525d424d5aSJohn Heffner } else { 21534de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & 2154a3433f35SChangli Gao ~(TCPHDR_FIN|TCPHDR_PSH); 21555d424d5aSJohn Heffner if (!skb_shinfo(skb)->nr_frags) { 21565d424d5aSJohn Heffner skb_pull(skb, copy); 21575d424d5aSJohn Heffner } else { 21585d424d5aSJohn Heffner __pskb_trim_head(skb, copy); 21595bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 21605d424d5aSJohn Heffner } 21615d424d5aSJohn Heffner TCP_SKB_CB(skb)->seq += copy; 21625d424d5aSJohn Heffner } 21635d424d5aSJohn Heffner 21645d424d5aSJohn Heffner len += copy; 2165234b6860SIlpo Järvinen 2166234b6860SIlpo Järvinen if (len >= probe_size) 2167234b6860SIlpo Järvinen break; 21685d424d5aSJohn Heffner } 21695bbb432cSEric Dumazet tcp_init_tso_segs(nskb, nskb->len); 21705d424d5aSJohn Heffner 21715d424d5aSJohn Heffner /* We're ready to send. If this fails, the probe will 21727faee5c0SEric Dumazet * be resegmented into mss-sized pieces by tcp_write_xmit(). 21737faee5c0SEric Dumazet */ 21745d424d5aSJohn Heffner if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 21755d424d5aSJohn Heffner /* Decrement cwnd here because we are sending 21765d424d5aSJohn Heffner * effectively two packets. */ 21775d424d5aSJohn Heffner tp->snd_cwnd--; 217866f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, nskb); 21795d424d5aSJohn Heffner 21805d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 21810e7b1368SJohn Heffner tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 21820e7b1368SJohn Heffner tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 21835d424d5aSJohn Heffner 21845d424d5aSJohn Heffner return 1; 21855d424d5aSJohn Heffner } 21865d424d5aSJohn Heffner 21875d424d5aSJohn Heffner return -1; 21885d424d5aSJohn Heffner } 21895d424d5aSJohn Heffner 2190218af599SEric Dumazet static bool tcp_pacing_check(const struct sock *sk) 2191218af599SEric Dumazet { 2192218af599SEric Dumazet return tcp_needs_internal_pacing(sk) && 219373a6bab5SEric Dumazet hrtimer_is_queued(&tcp_sk(sk)->pacing_timer); 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 { 2210f9616c35SEric Dumazet unsigned int limit; 2211f9616c35SEric Dumazet 22123a9b76fdSEric Dumazet limit = max(2 * skb->truesize, sk->sk_pacing_rate >> sk->sk_pacing_shift); 22139184d8bbSEric Dumazet limit = min_t(u32, limit, 22149184d8bbSEric Dumazet sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes); 2215f9616c35SEric Dumazet limit <<= factor; 2216f9616c35SEric Dumazet 221714afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) > limit) { 221875c119afSEric Dumazet /* Always send skb if rtx queue is empty. 221975eefc6cSEric Dumazet * No need to wait for TX completion to call us back, 222075eefc6cSEric Dumazet * after softirq/tasklet schedule. 222175eefc6cSEric Dumazet * This helps when TX completions are delayed too much. 222275eefc6cSEric Dumazet */ 222375c119afSEric Dumazet if (tcp_rtx_queue_empty(sk)) 222475eefc6cSEric Dumazet return false; 222575eefc6cSEric Dumazet 22267aa5470cSEric Dumazet set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags); 2227f9616c35SEric Dumazet /* It is possible TX completion already happened 2228f9616c35SEric Dumazet * before we set TSQ_THROTTLED, so we must 2229f9616c35SEric Dumazet * test again the condition. 2230f9616c35SEric Dumazet */ 2231f9616c35SEric Dumazet smp_mb__after_atomic(); 223214afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) > limit) 2233f9616c35SEric Dumazet return true; 2234f9616c35SEric Dumazet } 2235f9616c35SEric Dumazet return false; 2236f9616c35SEric Dumazet } 2237f9616c35SEric Dumazet 223805b055e8SFrancis Yan static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new) 223905b055e8SFrancis Yan { 2240628174ccSEric Dumazet const u32 now = tcp_jiffies32; 2241efe967cdSArnd Bergmann enum tcp_chrono old = tp->chrono_type; 224205b055e8SFrancis Yan 2243efe967cdSArnd Bergmann if (old > TCP_CHRONO_UNSPEC) 2244efe967cdSArnd Bergmann tp->chrono_stat[old - 1] += now - tp->chrono_start; 224505b055e8SFrancis Yan tp->chrono_start = now; 224605b055e8SFrancis Yan tp->chrono_type = new; 224705b055e8SFrancis Yan } 224805b055e8SFrancis Yan 224905b055e8SFrancis Yan void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type) 225005b055e8SFrancis Yan { 225105b055e8SFrancis Yan struct tcp_sock *tp = tcp_sk(sk); 225205b055e8SFrancis Yan 225305b055e8SFrancis Yan /* If there are multiple conditions worthy of tracking in a 22540f87230dSFrancis Yan * chronograph then the highest priority enum takes precedence 22550f87230dSFrancis Yan * over the other conditions. So that if something "more interesting" 225605b055e8SFrancis Yan * starts happening, stop the previous chrono and start a new one. 225705b055e8SFrancis Yan */ 225805b055e8SFrancis Yan if (type > tp->chrono_type) 225905b055e8SFrancis Yan tcp_chrono_set(tp, type); 226005b055e8SFrancis Yan } 226105b055e8SFrancis Yan 226205b055e8SFrancis Yan void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type) 226305b055e8SFrancis Yan { 226405b055e8SFrancis Yan struct tcp_sock *tp = tcp_sk(sk); 226505b055e8SFrancis Yan 22660f87230dSFrancis Yan 22670f87230dSFrancis Yan /* There are multiple conditions worthy of tracking in a 22680f87230dSFrancis Yan * chronograph, so that the highest priority enum takes 22690f87230dSFrancis Yan * precedence over the other conditions (see tcp_chrono_start). 22700f87230dSFrancis Yan * If a condition stops, we only stop chrono tracking if 22710f87230dSFrancis Yan * it's the "most interesting" or current chrono we are 22720f87230dSFrancis Yan * tracking and starts busy chrono if we have pending data. 22730f87230dSFrancis Yan */ 227475c119afSEric Dumazet if (tcp_rtx_and_write_queues_empty(sk)) 227505b055e8SFrancis Yan tcp_chrono_set(tp, TCP_CHRONO_UNSPEC); 22760f87230dSFrancis Yan else if (type == tp->chrono_type) 22770f87230dSFrancis Yan tcp_chrono_set(tp, TCP_CHRONO_BUSY); 227805b055e8SFrancis Yan } 227905b055e8SFrancis Yan 22801da177e4SLinus Torvalds /* This routine writes packets to the network. It advances the 22811da177e4SLinus Torvalds * send_head. This happens as incoming acks open up the remote 22821da177e4SLinus Torvalds * window for us. 22831da177e4SLinus Torvalds * 2284f8269a49SIlpo Järvinen * LARGESEND note: !tcp_urg_mode is overkill, only frames between 2285f8269a49SIlpo Järvinen * snd_up-64k-mss .. snd_up cannot be large. However, taking into 2286f8269a49SIlpo Järvinen * account rare use of URG, this is not a big flaw. 2287f8269a49SIlpo Järvinen * 22886ba8a3b1SNandita Dukkipati * Send at most one packet when push_one > 0. Temporarily ignore 22896ba8a3b1SNandita Dukkipati * cwnd limit to force at most one packet out when push_one == 2. 22906ba8a3b1SNandita Dukkipati 2291a2a385d6SEric Dumazet * Returns true, if no segments are in flight and we have queued segments, 2292a2a385d6SEric Dumazet * but cannot send anything now because of SWS or another problem. 22931da177e4SLinus Torvalds */ 2294a2a385d6SEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 2295d5dd9175SIlpo Järvinen int push_one, gfp_t gfp) 22961da177e4SLinus Torvalds { 22971da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 229892df7b51SDavid S. Miller struct sk_buff *skb; 2299c1b4a7e6SDavid S. Miller unsigned int tso_segs, sent_pkts; 2300c1b4a7e6SDavid S. Miller int cwnd_quota; 23015d424d5aSJohn Heffner int result; 23025615f886SFrancis Yan bool is_cwnd_limited = false, is_rwnd_limited = false; 2303605ad7f1SEric Dumazet u32 max_segs; 23041da177e4SLinus Torvalds 2305c1b4a7e6SDavid S. Miller sent_pkts = 0; 23065d424d5aSJohn Heffner 2307ee1836aeSEric Dumazet tcp_mstamp_refresh(tp); 2308d5dd9175SIlpo Järvinen if (!push_one) { 23095d424d5aSJohn Heffner /* Do MTU probing. */ 2310d5dd9175SIlpo Järvinen result = tcp_mtu_probe(sk); 2311d5dd9175SIlpo Järvinen if (!result) { 2312a2a385d6SEric Dumazet return false; 23135d424d5aSJohn Heffner } else if (result > 0) { 23145d424d5aSJohn Heffner sent_pkts = 1; 23155d424d5aSJohn Heffner } 2316d5dd9175SIlpo Järvinen } 23175d424d5aSJohn Heffner 2318ed6e7268SNeal Cardwell max_segs = tcp_tso_segs(sk, mss_now); 2319fe067e8aSDavid S. Miller while ((skb = tcp_send_head(sk))) { 2320c8ac3774SHerbert Xu unsigned int limit; 2321c8ac3774SHerbert Xu 2322218af599SEric Dumazet if (tcp_pacing_check(sk)) 2323218af599SEric Dumazet break; 2324218af599SEric Dumazet 23255bbb432cSEric Dumazet tso_segs = tcp_init_tso_segs(skb, mss_now); 2326c1b4a7e6SDavid S. Miller BUG_ON(!tso_segs); 2327c1b4a7e6SDavid S. Miller 23289d186cacSAndrey Vagin if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) { 23297faee5c0SEric Dumazet /* "skb_mstamp" is used as a start point for the retransmit timer */ 2330e2080072SEric Dumazet tcp_update_skb_after_send(tp, skb); 2331ec342325SAndrew Vagin goto repair; /* Skip network transmission */ 23329d186cacSAndrey Vagin } 2333ec342325SAndrew Vagin 2334b68e9f85SHerbert Xu cwnd_quota = tcp_cwnd_test(tp, skb); 23356ba8a3b1SNandita Dukkipati if (!cwnd_quota) { 23366ba8a3b1SNandita Dukkipati if (push_one == 2) 23376ba8a3b1SNandita Dukkipati /* Force out a loss probe pkt. */ 23386ba8a3b1SNandita Dukkipati cwnd_quota = 1; 23396ba8a3b1SNandita Dukkipati else 2340b68e9f85SHerbert Xu break; 23416ba8a3b1SNandita Dukkipati } 2342b68e9f85SHerbert Xu 23435615f886SFrancis Yan if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) { 23445615f886SFrancis Yan is_rwnd_limited = true; 2345b68e9f85SHerbert Xu break; 23465615f886SFrancis Yan } 2347b68e9f85SHerbert Xu 2348d6a4e26aSEric Dumazet if (tso_segs == 1) { 2349aa93466bSDavid S. Miller if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 2350aa93466bSDavid S. Miller (tcp_skb_is_last(sk, skb) ? 2351aa93466bSDavid S. Miller nonagle : TCP_NAGLE_PUSH)))) 2352aa93466bSDavid S. Miller break; 2353c1b4a7e6SDavid S. Miller } else { 2354ca8a2263SNeal Cardwell if (!push_one && 2355605ad7f1SEric Dumazet tcp_tso_should_defer(sk, skb, &is_cwnd_limited, 2356605ad7f1SEric Dumazet max_segs)) 2357aa93466bSDavid S. Miller break; 2358c1b4a7e6SDavid S. Miller } 2359aa93466bSDavid S. Miller 2360605ad7f1SEric Dumazet limit = mss_now; 2361d6a4e26aSEric Dumazet if (tso_segs > 1 && !tcp_urg_mode(tp)) 2362605ad7f1SEric Dumazet limit = tcp_mss_split_point(sk, skb, mss_now, 2363605ad7f1SEric Dumazet min_t(unsigned int, 2364605ad7f1SEric Dumazet cwnd_quota, 2365605ad7f1SEric Dumazet max_segs), 2366605ad7f1SEric Dumazet nonagle); 2367605ad7f1SEric Dumazet 2368605ad7f1SEric Dumazet if (skb->len > limit && 236975c119afSEric Dumazet unlikely(tso_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE, 237075c119afSEric Dumazet skb, limit, mss_now, gfp))) 2371605ad7f1SEric Dumazet break; 2372605ad7f1SEric Dumazet 2373f9616c35SEric Dumazet if (tcp_small_queue_check(sk, skb, 0)) 237446d3ceabSEric Dumazet break; 2375c9eeec26SEric Dumazet 2376d5dd9175SIlpo Järvinen if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) 23771da177e4SLinus Torvalds break; 23781da177e4SLinus Torvalds 2379ec342325SAndrew Vagin repair: 23801da177e4SLinus Torvalds /* Advance the send_head. This one is sent out. 23811da177e4SLinus Torvalds * This call will increment packets_out. 23821da177e4SLinus Torvalds */ 238366f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 23841da177e4SLinus Torvalds 23851da177e4SLinus Torvalds tcp_minshall_update(tp, mss_now, skb); 2386a262f0cdSNandita Dukkipati sent_pkts += tcp_skb_pcount(skb); 2387d5dd9175SIlpo Järvinen 2388d5dd9175SIlpo Järvinen if (push_one) 2389d5dd9175SIlpo Järvinen break; 23901da177e4SLinus Torvalds } 23911da177e4SLinus Torvalds 23925615f886SFrancis Yan if (is_rwnd_limited) 23935615f886SFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED); 23945615f886SFrancis Yan else 23955615f886SFrancis Yan tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED); 23965615f886SFrancis Yan 2397aa93466bSDavid S. Miller if (likely(sent_pkts)) { 2398684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 2399684bad11SYuchung Cheng tp->prr_out += sent_pkts; 24006ba8a3b1SNandita Dukkipati 24016ba8a3b1SNandita Dukkipati /* Send one loss probe per tail loss episode. */ 24026ba8a3b1SNandita Dukkipati if (push_one != 2) 2403ed66dfafSNeal Cardwell tcp_schedule_loss_probe(sk, false); 2404d2e1339fSBendik Rønning Opstad is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd); 2405ca8a2263SNeal Cardwell tcp_cwnd_validate(sk, is_cwnd_limited); 2406a2a385d6SEric Dumazet return false; 24071da177e4SLinus Torvalds } 240875c119afSEric Dumazet return !tp->packets_out && !tcp_write_queue_empty(sk); 24096ba8a3b1SNandita Dukkipati } 24106ba8a3b1SNandita Dukkipati 2411ed66dfafSNeal Cardwell bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto) 24126ba8a3b1SNandita Dukkipati { 24136ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 24146ba8a3b1SNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 2415a2815817SNeal Cardwell u32 timeout, rto_delta_us; 24162ae21cf5SEric Dumazet int early_retrans; 24176ba8a3b1SNandita Dukkipati 24186ba8a3b1SNandita Dukkipati /* Don't do any loss probe on a Fast Open connection before 3WHS 24196ba8a3b1SNandita Dukkipati * finishes. 24206ba8a3b1SNandita Dukkipati */ 2421f9b99582SYuchung Cheng if (tp->fastopen_rsk) 24226ba8a3b1SNandita Dukkipati return false; 24236ba8a3b1SNandita Dukkipati 24242ae21cf5SEric Dumazet early_retrans = sock_net(sk)->ipv4.sysctl_tcp_early_retrans; 24256ba8a3b1SNandita Dukkipati /* Schedule a loss probe in 2*RTT for SACK capable connections 2426b4f70c3dSNeal Cardwell * not in loss recovery, that are either limited by cwnd or application. 24276ba8a3b1SNandita Dukkipati */ 24282ae21cf5SEric Dumazet if ((early_retrans != 3 && early_retrans != 4) || 2429bec41a11SYuchung Cheng !tp->packets_out || !tcp_is_sack(tp) || 2430b4f70c3dSNeal Cardwell (icsk->icsk_ca_state != TCP_CA_Open && 2431b4f70c3dSNeal Cardwell icsk->icsk_ca_state != TCP_CA_CWR)) 24326ba8a3b1SNandita Dukkipati return false; 24336ba8a3b1SNandita Dukkipati 2434bb4d991aSYuchung Cheng /* Probe timeout is 2*rtt. Add minimum RTO to account 2435f9b99582SYuchung Cheng * for delayed ack when there's one outstanding packet. If no RTT 2436f9b99582SYuchung Cheng * sample is available then probe after TCP_TIMEOUT_INIT. 24376ba8a3b1SNandita Dukkipati */ 2438bb4d991aSYuchung Cheng if (tp->srtt_us) { 2439bb4d991aSYuchung Cheng timeout = usecs_to_jiffies(tp->srtt_us >> 2); 24406ba8a3b1SNandita Dukkipati if (tp->packets_out == 1) 2441bb4d991aSYuchung Cheng timeout += TCP_RTO_MIN; 2442bb4d991aSYuchung Cheng else 2443bb4d991aSYuchung Cheng timeout += TCP_TIMEOUT_MIN; 2444bb4d991aSYuchung Cheng } else { 2445bb4d991aSYuchung Cheng timeout = TCP_TIMEOUT_INIT; 2446bb4d991aSYuchung Cheng } 24476ba8a3b1SNandita Dukkipati 2448a2815817SNeal Cardwell /* If the RTO formula yields an earlier time, then use that time. */ 2449ed66dfafSNeal Cardwell rto_delta_us = advancing_rto ? 2450ed66dfafSNeal Cardwell jiffies_to_usecs(inet_csk(sk)->icsk_rto) : 2451ed66dfafSNeal Cardwell tcp_rto_delta_us(sk); /* How far in future is RTO? */ 2452a2815817SNeal Cardwell if (rto_delta_us > 0) 2453a2815817SNeal Cardwell timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us)); 24546ba8a3b1SNandita Dukkipati 24556ba8a3b1SNandita Dukkipati inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, 24566ba8a3b1SNandita Dukkipati TCP_RTO_MAX); 24576ba8a3b1SNandita Dukkipati return true; 24586ba8a3b1SNandita Dukkipati } 24596ba8a3b1SNandita Dukkipati 24601f3279aeSEric Dumazet /* Thanks to skb fast clones, we can detect if a prior transmit of 24611f3279aeSEric Dumazet * a packet is still in a qdisc or driver queue. 24621f3279aeSEric Dumazet * In this case, there is very little point doing a retransmit ! 24631f3279aeSEric Dumazet */ 24641f3279aeSEric Dumazet static bool skb_still_in_host_queue(const struct sock *sk, 24651f3279aeSEric Dumazet const struct sk_buff *skb) 24661f3279aeSEric Dumazet { 246739bb5e62SEric Dumazet if (unlikely(skb_fclone_busy(sk, skb))) { 2468c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), 24691f3279aeSEric Dumazet LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES); 24701f3279aeSEric Dumazet return true; 24711f3279aeSEric Dumazet } 24721f3279aeSEric Dumazet return false; 24731f3279aeSEric Dumazet } 24741f3279aeSEric Dumazet 2475b340b264SYuchung Cheng /* When probe timeout (PTO) fires, try send a new segment if possible, else 24766ba8a3b1SNandita Dukkipati * retransmit the last segment. 24776ba8a3b1SNandita Dukkipati */ 24786ba8a3b1SNandita Dukkipati void tcp_send_loss_probe(struct sock *sk) 24796ba8a3b1SNandita Dukkipati { 24809b717a8dSNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 24816ba8a3b1SNandita Dukkipati struct sk_buff *skb; 24826ba8a3b1SNandita Dukkipati int pcount; 24836ba8a3b1SNandita Dukkipati int mss = tcp_current_mss(sk); 24846ba8a3b1SNandita Dukkipati 2485b340b264SYuchung Cheng skb = tcp_send_head(sk); 248675c119afSEric Dumazet if (skb && tcp_snd_wnd_test(tp, skb, mss)) { 2487b340b264SYuchung Cheng pcount = tp->packets_out; 2488b340b264SYuchung Cheng tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC); 2489b340b264SYuchung Cheng if (tp->packets_out > pcount) 2490b340b264SYuchung Cheng goto probe_sent; 24916ba8a3b1SNandita Dukkipati goto rearm_timer; 24926ba8a3b1SNandita Dukkipati } 249375c119afSEric Dumazet skb = skb_rb_last(&sk->tcp_rtx_queue); 24946ba8a3b1SNandita Dukkipati 24959b717a8dSNandita Dukkipati /* At most one outstanding TLP retransmission. */ 24969b717a8dSNandita Dukkipati if (tp->tlp_high_seq) 24979b717a8dSNandita Dukkipati goto rearm_timer; 24989b717a8dSNandita Dukkipati 24996ba8a3b1SNandita Dukkipati /* Retransmit last segment. */ 25006ba8a3b1SNandita Dukkipati if (WARN_ON(!skb)) 25016ba8a3b1SNandita Dukkipati goto rearm_timer; 25026ba8a3b1SNandita Dukkipati 25031f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 25041f3279aeSEric Dumazet goto rearm_timer; 25051f3279aeSEric Dumazet 25066ba8a3b1SNandita Dukkipati pcount = tcp_skb_pcount(skb); 25076ba8a3b1SNandita Dukkipati if (WARN_ON(!pcount)) 25086ba8a3b1SNandita Dukkipati goto rearm_timer; 25096ba8a3b1SNandita Dukkipati 25106ba8a3b1SNandita Dukkipati if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) { 251175c119afSEric Dumazet if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, 251275c119afSEric Dumazet (pcount - 1) * mss, mss, 25136cc55e09SOctavian Purdila GFP_ATOMIC))) 25146ba8a3b1SNandita Dukkipati goto rearm_timer; 251575c119afSEric Dumazet skb = skb_rb_next(skb); 25166ba8a3b1SNandita Dukkipati } 25176ba8a3b1SNandita Dukkipati 25186ba8a3b1SNandita Dukkipati if (WARN_ON(!skb || !tcp_skb_pcount(skb))) 25196ba8a3b1SNandita Dukkipati goto rearm_timer; 25206ba8a3b1SNandita Dukkipati 252110d3be56SEric Dumazet if (__tcp_retransmit_skb(sk, skb, 1)) 2522b340b264SYuchung Cheng goto rearm_timer; 25236ba8a3b1SNandita Dukkipati 25249b717a8dSNandita Dukkipati /* Record snd_nxt for loss detection. */ 25259b717a8dSNandita Dukkipati tp->tlp_high_seq = tp->snd_nxt; 25269b717a8dSNandita Dukkipati 2527b340b264SYuchung Cheng probe_sent: 2528c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES); 2529fcd16c0aSYuchung Cheng /* Reset s.t. tcp_rearm_rto will restart timer from now */ 2530fcd16c0aSYuchung Cheng inet_csk(sk)->icsk_pending = 0; 2531b340b264SYuchung Cheng rearm_timer: 2532fcd16c0aSYuchung Cheng tcp_rearm_rto(sk); 25331da177e4SLinus Torvalds } 25341da177e4SLinus Torvalds 2535a762a980SDavid S. Miller /* Push out any pending frames which were held back due to 2536a762a980SDavid S. Miller * TCP_CORK or attempt at coalescing tiny packets. 2537a762a980SDavid S. Miller * The socket must be locked by the caller. 2538a762a980SDavid S. Miller */ 25399e412ba7SIlpo Järvinen void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 25409e412ba7SIlpo Järvinen int nonagle) 2541a762a980SDavid S. Miller { 2542726e07a8SIlpo Järvinen /* If we are closed, the bytes will have to remain here. 2543726e07a8SIlpo Järvinen * In time closedown will finish, we empty the write queue and 2544726e07a8SIlpo Järvinen * all will be happy. 2545726e07a8SIlpo Järvinen */ 2546726e07a8SIlpo Järvinen if (unlikely(sk->sk_state == TCP_CLOSE)) 2547726e07a8SIlpo Järvinen return; 2548726e07a8SIlpo Järvinen 254999a1dec7SMel Gorman if (tcp_write_xmit(sk, cur_mss, nonagle, 0, 25507450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC))) 25519e412ba7SIlpo Järvinen tcp_check_probe_timer(sk); 2552a762a980SDavid S. Miller } 2553a762a980SDavid S. Miller 2554c1b4a7e6SDavid S. Miller /* Send _single_ skb sitting at the send head. This function requires 2555c1b4a7e6SDavid S. Miller * true push pending frames to setup probe timer etc. 2556c1b4a7e6SDavid S. Miller */ 2557c1b4a7e6SDavid S. Miller void tcp_push_one(struct sock *sk, unsigned int mss_now) 2558c1b4a7e6SDavid S. Miller { 2559fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 2560c1b4a7e6SDavid S. Miller 2561c1b4a7e6SDavid S. Miller BUG_ON(!skb || skb->len < mss_now); 2562c1b4a7e6SDavid S. Miller 2563d5dd9175SIlpo Järvinen tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); 2564c1b4a7e6SDavid S. Miller } 2565c1b4a7e6SDavid S. Miller 25661da177e4SLinus Torvalds /* This function returns the amount that we can raise the 25671da177e4SLinus Torvalds * usable window based on the following constraints 25681da177e4SLinus Torvalds * 25691da177e4SLinus Torvalds * 1. The window can never be shrunk once it is offered (RFC 793) 25701da177e4SLinus Torvalds * 2. We limit memory per socket 25711da177e4SLinus Torvalds * 25721da177e4SLinus Torvalds * RFC 1122: 25731da177e4SLinus Torvalds * "the suggested [SWS] avoidance algorithm for the receiver is to keep 25741da177e4SLinus Torvalds * RECV.NEXT + RCV.WIN fixed until: 25751da177e4SLinus Torvalds * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 25761da177e4SLinus Torvalds * 25771da177e4SLinus Torvalds * i.e. don't raise the right edge of the window until you can raise 25781da177e4SLinus Torvalds * it at least MSS bytes. 25791da177e4SLinus Torvalds * 25801da177e4SLinus Torvalds * Unfortunately, the recommended algorithm breaks header prediction, 25811da177e4SLinus Torvalds * since header prediction assumes th->window stays fixed. 25821da177e4SLinus Torvalds * 25831da177e4SLinus Torvalds * Strictly speaking, keeping th->window fixed violates the receiver 25841da177e4SLinus Torvalds * side SWS prevention criteria. The problem is that under this rule 25851da177e4SLinus Torvalds * a stream of single byte packets will cause the right side of the 25861da177e4SLinus Torvalds * window to always advance by a single byte. 25871da177e4SLinus Torvalds * 25881da177e4SLinus Torvalds * Of course, if the sender implements sender side SWS prevention 25891da177e4SLinus Torvalds * then this will not be a problem. 25901da177e4SLinus Torvalds * 25911da177e4SLinus Torvalds * BSD seems to make the following compromise: 25921da177e4SLinus Torvalds * 25931da177e4SLinus Torvalds * If the free space is less than the 1/4 of the maximum 25941da177e4SLinus Torvalds * space available and the free space is less than 1/2 mss, 25951da177e4SLinus Torvalds * then set the window to 0. 25961da177e4SLinus Torvalds * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 25971da177e4SLinus Torvalds * Otherwise, just prevent the window from shrinking 25981da177e4SLinus Torvalds * and from being larger than the largest representable value. 25991da177e4SLinus Torvalds * 26001da177e4SLinus Torvalds * This prevents incremental opening of the window in the regime 26011da177e4SLinus Torvalds * where TCP is limited by the speed of the reader side taking 26021da177e4SLinus Torvalds * data out of the TCP receive queue. It does nothing about 26031da177e4SLinus Torvalds * those cases where the window is constrained on the sender side 26041da177e4SLinus Torvalds * because the pipeline is full. 26051da177e4SLinus Torvalds * 26061da177e4SLinus Torvalds * BSD also seems to "accidentally" limit itself to windows that are a 26071da177e4SLinus Torvalds * multiple of MSS, at least until the free space gets quite small. 26081da177e4SLinus Torvalds * This would appear to be a side effect of the mbuf implementation. 26091da177e4SLinus Torvalds * Combining these two algorithms results in the observed behavior 26101da177e4SLinus Torvalds * of having a fixed window size at almost all times. 26111da177e4SLinus Torvalds * 26121da177e4SLinus Torvalds * Below we obtain similar behavior by forcing the offered window to 26131da177e4SLinus Torvalds * a multiple of the mss when it is feasible to do so. 26141da177e4SLinus Torvalds * 26151da177e4SLinus Torvalds * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 26161da177e4SLinus Torvalds * Regular options like TIMESTAMP are taken into account. 26171da177e4SLinus Torvalds */ 26181da177e4SLinus Torvalds u32 __tcp_select_window(struct sock *sk) 26191da177e4SLinus Torvalds { 2620463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 26211da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2622caa20d9aSStephen Hemminger /* MSS for the peer's data. Previous versions used mss_clamp 26231da177e4SLinus Torvalds * here. I don't know if the value based on our guesses 26241da177e4SLinus Torvalds * of peer's MSS is better for the performance. It's more correct 26251da177e4SLinus Torvalds * but may be worse for the performance because of rcv_mss 26261da177e4SLinus Torvalds * fluctuations. --SAW 1998/11/1 26271da177e4SLinus Torvalds */ 2628463c84b9SArnaldo Carvalho de Melo int mss = icsk->icsk_ack.rcv_mss; 26291da177e4SLinus Torvalds int free_space = tcp_space(sk); 263086c1a045SFlorian Westphal int allowed_space = tcp_full_space(sk); 263186c1a045SFlorian Westphal int full_space = min_t(int, tp->window_clamp, allowed_space); 26321da177e4SLinus Torvalds int window; 26331da177e4SLinus Torvalds 263406425c30SEric Dumazet if (unlikely(mss > full_space)) { 26351da177e4SLinus Torvalds mss = full_space; 263606425c30SEric Dumazet if (mss <= 0) 263706425c30SEric Dumazet return 0; 263806425c30SEric Dumazet } 2639b92edbe0SEric Dumazet if (free_space < (full_space >> 1)) { 2640463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.quick = 0; 26411da177e4SLinus Torvalds 2642b8da51ebSEric Dumazet if (tcp_under_memory_pressure(sk)) 2643056834d9SIlpo Järvinen tp->rcv_ssthresh = min(tp->rcv_ssthresh, 2644056834d9SIlpo Järvinen 4U * tp->advmss); 26451da177e4SLinus Torvalds 264686c1a045SFlorian Westphal /* free_space might become our new window, make sure we don't 264786c1a045SFlorian Westphal * increase it due to wscale. 264886c1a045SFlorian Westphal */ 264986c1a045SFlorian Westphal free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale); 265086c1a045SFlorian Westphal 265186c1a045SFlorian Westphal /* if free space is less than mss estimate, or is below 1/16th 265286c1a045SFlorian Westphal * of the maximum allowed, try to move to zero-window, else 265386c1a045SFlorian Westphal * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and 265486c1a045SFlorian Westphal * new incoming data is dropped due to memory limits. 265586c1a045SFlorian Westphal * With large window, mss test triggers way too late in order 265686c1a045SFlorian Westphal * to announce zero window in time before rmem limit kicks in. 265786c1a045SFlorian Westphal */ 265886c1a045SFlorian Westphal if (free_space < (allowed_space >> 4) || free_space < mss) 26591da177e4SLinus Torvalds return 0; 26601da177e4SLinus Torvalds } 26611da177e4SLinus Torvalds 26621da177e4SLinus Torvalds if (free_space > tp->rcv_ssthresh) 26631da177e4SLinus Torvalds free_space = tp->rcv_ssthresh; 26641da177e4SLinus Torvalds 26651da177e4SLinus Torvalds /* Don't do rounding if we are using window scaling, since the 26661da177e4SLinus Torvalds * scaled window will not line up with the MSS boundary anyway. 26671da177e4SLinus Torvalds */ 26681da177e4SLinus Torvalds if (tp->rx_opt.rcv_wscale) { 26691da177e4SLinus Torvalds window = free_space; 26701da177e4SLinus Torvalds 26711da177e4SLinus Torvalds /* Advertise enough space so that it won't get scaled away. 26721da177e4SLinus Torvalds * Import case: prevent zero window announcement if 26731da177e4SLinus Torvalds * 1<<rcv_wscale > mss. 26741da177e4SLinus Torvalds */ 26751935299dSGao Feng window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale)); 26761da177e4SLinus Torvalds } else { 26771935299dSGao Feng window = tp->rcv_wnd; 26781da177e4SLinus Torvalds /* Get the largest window that is a nice multiple of mss. 26791da177e4SLinus Torvalds * Window clamp already applied above. 26801da177e4SLinus Torvalds * If our current window offering is within 1 mss of the 26811da177e4SLinus Torvalds * free space we just keep it. This prevents the divide 26821da177e4SLinus Torvalds * and multiply from happening most of the time. 26831da177e4SLinus Torvalds * We also don't do any window rounding when the free space 26841da177e4SLinus Torvalds * is too small. 26851da177e4SLinus Torvalds */ 26861da177e4SLinus Torvalds if (window <= free_space - mss || window > free_space) 26871935299dSGao Feng window = rounddown(free_space, mss); 268884565070SJohn Heffner else if (mss == full_space && 2689b92edbe0SEric Dumazet free_space > window + (full_space >> 1)) 269084565070SJohn Heffner window = free_space; 26911da177e4SLinus Torvalds } 26921da177e4SLinus Torvalds 26931da177e4SLinus Torvalds return window; 26941da177e4SLinus Torvalds } 26951da177e4SLinus Torvalds 2696cfea5a68SMartin KaFai Lau void tcp_skb_collapse_tstamp(struct sk_buff *skb, 2697082ac2d5SMartin KaFai Lau const struct sk_buff *next_skb) 2698082ac2d5SMartin KaFai Lau { 26990a2cf20cSSoheil Hassas Yeganeh if (unlikely(tcp_has_tx_tstamp(next_skb))) { 27000a2cf20cSSoheil Hassas Yeganeh const struct skb_shared_info *next_shinfo = 27010a2cf20cSSoheil Hassas Yeganeh skb_shinfo(next_skb); 2702082ac2d5SMartin KaFai Lau struct skb_shared_info *shinfo = skb_shinfo(skb); 2703082ac2d5SMartin KaFai Lau 27040a2cf20cSSoheil Hassas Yeganeh shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP; 2705082ac2d5SMartin KaFai Lau shinfo->tskey = next_shinfo->tskey; 27062de8023eSMartin KaFai Lau TCP_SKB_CB(skb)->txstamp_ack |= 27072de8023eSMartin KaFai Lau TCP_SKB_CB(next_skb)->txstamp_ack; 2708082ac2d5SMartin KaFai Lau } 2709082ac2d5SMartin KaFai Lau } 2710082ac2d5SMartin KaFai Lau 27114a17fc3aSIlpo Järvinen /* Collapses two adjacent SKB's during retransmission. */ 2712f8071cdeSEric Dumazet static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) 27131da177e4SLinus Torvalds { 27141da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 271575c119afSEric Dumazet struct sk_buff *next_skb = skb_rb_next(skb); 271613dde04fSWei Yongjun int next_skb_size; 27171da177e4SLinus Torvalds 2718058dc334SIlpo Järvinen next_skb_size = next_skb->len; 27191da177e4SLinus Torvalds 2720058dc334SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); 27211da177e4SLinus Torvalds 2722f8071cdeSEric Dumazet if (next_skb_size) { 2723f8071cdeSEric Dumazet if (next_skb_size <= skb_availroom(skb)) 2724f8071cdeSEric Dumazet skb_copy_bits(next_skb, 0, skb_put(skb, next_skb_size), 2725f8071cdeSEric Dumazet next_skb_size); 2726f8071cdeSEric Dumazet else if (!skb_shift(skb, next_skb, next_skb_size)) 2727f8071cdeSEric Dumazet return false; 2728f8071cdeSEric Dumazet } 27292b7cda9cSEric Dumazet tcp_highest_sack_replace(sk, next_skb, skb); 2730a6963a6bSIlpo Järvinen 27311da177e4SLinus Torvalds /* Update sequence range on original skb. */ 27321da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 27331da177e4SLinus Torvalds 2734e6c7d085SIlpo Järvinen /* Merge over control information. This moves PSH/FIN etc. over */ 27354de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; 27361da177e4SLinus Torvalds 27371da177e4SLinus Torvalds /* All done, get rid of second SKB and account for it so 27381da177e4SLinus Torvalds * packet counting does not break. 27391da177e4SLinus Torvalds */ 27404828e7f4SIlpo Järvinen TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; 2741a643b5d4SMartin KaFai Lau TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor; 2742b7689205SIlpo Järvinen 2743b7689205SIlpo Järvinen /* changed transmit queue under us so clear hints */ 2744ef9da47cSIlpo Järvinen tcp_clear_retrans_hints_partial(tp); 2745ef9da47cSIlpo Järvinen if (next_skb == tp->retransmit_skb_hint) 2746ef9da47cSIlpo Järvinen tp->retransmit_skb_hint = skb; 2747b7689205SIlpo Järvinen 2748797108d1SIlpo Järvinen tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); 2749797108d1SIlpo Järvinen 2750082ac2d5SMartin KaFai Lau tcp_skb_collapse_tstamp(skb, next_skb); 2751082ac2d5SMartin KaFai Lau 275275c119afSEric Dumazet tcp_rtx_queue_unlink_and_free(next_skb, sk); 2753f8071cdeSEric Dumazet return true; 27541da177e4SLinus Torvalds } 27551da177e4SLinus Torvalds 275667edfef7SAndi Kleen /* Check if coalescing SKBs is legal. */ 2757a2a385d6SEric Dumazet static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) 27584a17fc3aSIlpo Järvinen { 27594a17fc3aSIlpo Järvinen if (tcp_skb_pcount(skb) > 1) 2760a2a385d6SEric Dumazet return false; 27614a17fc3aSIlpo Järvinen if (skb_cloned(skb)) 2762a2a385d6SEric Dumazet return false; 27632331ccc5SEric Dumazet /* Some heuristics for collapsing over SACK'd could be invented */ 27644a17fc3aSIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 2765a2a385d6SEric Dumazet return false; 27664a17fc3aSIlpo Järvinen 2767a2a385d6SEric Dumazet return true; 27684a17fc3aSIlpo Järvinen } 27694a17fc3aSIlpo Järvinen 277067edfef7SAndi Kleen /* Collapse packets in the retransmit queue to make to create 277167edfef7SAndi Kleen * less packets on the wire. This is only done on retransmission. 277267edfef7SAndi Kleen */ 27734a17fc3aSIlpo Järvinen static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, 27744a17fc3aSIlpo Järvinen int space) 27754a17fc3aSIlpo Järvinen { 27764a17fc3aSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 27774a17fc3aSIlpo Järvinen struct sk_buff *skb = to, *tmp; 2778a2a385d6SEric Dumazet bool first = true; 27794a17fc3aSIlpo Järvinen 2780e0a1e5b5SEric Dumazet if (!sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse) 27814a17fc3aSIlpo Järvinen return; 27824de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 27834a17fc3aSIlpo Järvinen return; 27844a17fc3aSIlpo Järvinen 278575c119afSEric Dumazet skb_rbtree_walk_from_safe(skb, tmp) { 27864a17fc3aSIlpo Järvinen if (!tcp_can_collapse(sk, skb)) 27874a17fc3aSIlpo Järvinen break; 27884a17fc3aSIlpo Järvinen 2789a643b5d4SMartin KaFai Lau if (!tcp_skb_can_collapse_to(to)) 2790a643b5d4SMartin KaFai Lau break; 2791a643b5d4SMartin KaFai Lau 27924a17fc3aSIlpo Järvinen space -= skb->len; 27934a17fc3aSIlpo Järvinen 27944a17fc3aSIlpo Järvinen if (first) { 2795a2a385d6SEric Dumazet first = false; 27964a17fc3aSIlpo Järvinen continue; 27974a17fc3aSIlpo Järvinen } 27984a17fc3aSIlpo Järvinen 27994a17fc3aSIlpo Järvinen if (space < 0) 28004a17fc3aSIlpo Järvinen break; 28014a17fc3aSIlpo Järvinen 28024a17fc3aSIlpo Järvinen if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) 28034a17fc3aSIlpo Järvinen break; 28044a17fc3aSIlpo Järvinen 2805f8071cdeSEric Dumazet if (!tcp_collapse_retrans(sk, to)) 2806f8071cdeSEric Dumazet break; 28074a17fc3aSIlpo Järvinen } 28084a17fc3aSIlpo Järvinen } 28094a17fc3aSIlpo Järvinen 28101da177e4SLinus Torvalds /* This retransmits one SKB. Policy decisions and retransmit queue 28111da177e4SLinus Torvalds * state updates are done by the caller. Returns non-zero if an 28121da177e4SLinus Torvalds * error occurred which prevented the send. 28131da177e4SLinus Torvalds */ 281410d3be56SEric Dumazet int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) 28151da177e4SLinus Torvalds { 28165d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 281710d3be56SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 28187d227cd2SSridhar Samudrala unsigned int cur_mss; 281910d3be56SEric Dumazet int diff, len, err; 28201da177e4SLinus Torvalds 282110d3be56SEric Dumazet 282210d3be56SEric Dumazet /* Inconclusive MTU probe */ 282310d3be56SEric Dumazet if (icsk->icsk_mtup.probe_size) 28245d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 28255d424d5aSJohn Heffner 28261da177e4SLinus Torvalds /* Do not sent more than we queued. 1/4 is reserved for possible 2827caa20d9aSStephen Hemminger * copying overhead: fragmentation, tunneling, mangling etc. 28281da177e4SLinus Torvalds */ 282914afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) > 2830ffb4d6c8SEric Dumazet min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), 2831ffb4d6c8SEric Dumazet sk->sk_sndbuf)) 28321da177e4SLinus Torvalds return -EAGAIN; 28331da177e4SLinus Torvalds 28341f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 28351f3279aeSEric Dumazet return -EBUSY; 28361f3279aeSEric Dumazet 28371da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 28387f582b24SEric Dumazet if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) { 28397f582b24SEric Dumazet WARN_ON_ONCE(1); 28407f582b24SEric Dumazet return -EINVAL; 28417f582b24SEric Dumazet } 28421da177e4SLinus Torvalds if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 28431da177e4SLinus Torvalds return -ENOMEM; 28441da177e4SLinus Torvalds } 28451da177e4SLinus Torvalds 28467d227cd2SSridhar Samudrala if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 28477d227cd2SSridhar Samudrala return -EHOSTUNREACH; /* Routing failure or similar. */ 28487d227cd2SSridhar Samudrala 28490c54b85fSIlpo Järvinen cur_mss = tcp_current_mss(sk); 28507d227cd2SSridhar Samudrala 28511da177e4SLinus Torvalds /* If receiver has shrunk his window, and skb is out of 28521da177e4SLinus Torvalds * new window, do not retransmit it. The exception is the 28531da177e4SLinus Torvalds * case, when window is shrunk to zero. In this case 28541da177e4SLinus Torvalds * our retransmit serves as a zero window probe. 28551da177e4SLinus Torvalds */ 28569d4fb27dSJoe Perches if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && 28579d4fb27dSJoe Perches TCP_SKB_CB(skb)->seq != tp->snd_una) 28581da177e4SLinus Torvalds return -EAGAIN; 28591da177e4SLinus Torvalds 286010d3be56SEric Dumazet len = cur_mss * segs; 286110d3be56SEric Dumazet if (skb->len > len) { 286275c119afSEric Dumazet if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len, 286375c119afSEric Dumazet cur_mss, GFP_ATOMIC)) 28641da177e4SLinus Torvalds return -ENOMEM; /* We'll try again later. */ 286502276f3cSIlpo Järvinen } else { 2866c52e2421SEric Dumazet if (skb_unclone(skb, GFP_ATOMIC)) 2867c52e2421SEric Dumazet return -ENOMEM; 286810d3be56SEric Dumazet 286910d3be56SEric Dumazet diff = tcp_skb_pcount(skb); 287010d3be56SEric Dumazet tcp_set_skb_tso_segs(skb, cur_mss); 287110d3be56SEric Dumazet diff -= tcp_skb_pcount(skb); 287210d3be56SEric Dumazet if (diff) 287310d3be56SEric Dumazet tcp_adjust_pcount(sk, skb, diff); 287410d3be56SEric Dumazet if (skb->len < cur_mss) 287510d3be56SEric Dumazet tcp_retrans_try_collapse(sk, skb, cur_mss); 28761da177e4SLinus Torvalds } 28771da177e4SLinus Torvalds 287849213555SDaniel Borkmann /* RFC3168, section 6.1.1.1. ECN fallback */ 287949213555SDaniel Borkmann if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN) 288049213555SDaniel Borkmann tcp_ecn_clear_syn(sk, skb); 288149213555SDaniel Borkmann 2882678550c6SYuchung Cheng /* Update global and local TCP statistics. */ 2883678550c6SYuchung Cheng segs = tcp_skb_pcount(skb); 2884678550c6SYuchung Cheng TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs); 2885678550c6SYuchung Cheng if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 2886678550c6SYuchung Cheng __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 2887678550c6SYuchung Cheng tp->total_retrans += segs; 2888fb31c9b9SWei Wang tp->bytes_retrans += skb->len; 2889678550c6SYuchung Cheng 289050bceae9SThomas Graf /* make sure skb->data is aligned on arches that require it 289150bceae9SThomas Graf * and check if ack-trimming & collapsing extended the headroom 289250bceae9SThomas Graf * beyond what csum_start can cover. 289350bceae9SThomas Graf */ 289450bceae9SThomas Graf if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) || 289550bceae9SThomas Graf skb_headroom(skb) >= 0xFFFF)) { 289610a81980SEric Dumazet struct sk_buff *nskb; 289710a81980SEric Dumazet 2898e2080072SEric Dumazet tcp_skb_tsorted_save(skb) { 289910a81980SEric Dumazet nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC); 2900c84a5711SYuchung Cheng err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) : 2901117632e6SEric Dumazet -ENOBUFS; 2902e2080072SEric Dumazet } tcp_skb_tsorted_restore(skb); 2903e2080072SEric Dumazet 29045889e2c0SYousuk Seung if (!err) { 2905e2080072SEric Dumazet tcp_update_skb_after_send(tp, skb); 29065889e2c0SYousuk Seung tcp_rate_skb_sent(sk, skb); 29075889e2c0SYousuk Seung } 2908117632e6SEric Dumazet } else { 2909c84a5711SYuchung Cheng err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 2910117632e6SEric Dumazet } 2911c84a5711SYuchung Cheng 2912a31ad29eSLawrence Brakmo if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RETRANS_CB_FLAG)) 2913a31ad29eSLawrence Brakmo tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RETRANS_CB, 2914a31ad29eSLawrence Brakmo TCP_SKB_CB(skb)->seq, segs, err); 2915a31ad29eSLawrence Brakmo 2916fc9f3501SEric Dumazet if (likely(!err)) { 2917c84a5711SYuchung Cheng TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS; 2918e086101bSCong Wang trace_tcp_retransmit_skb(sk, skb); 2919678550c6SYuchung Cheng } else if (err != -EBUSY) { 2920678550c6SYuchung Cheng NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL); 2921fc9f3501SEric Dumazet } 2922c84a5711SYuchung Cheng return err; 292393b174adSYuchung Cheng } 292493b174adSYuchung Cheng 292510d3be56SEric Dumazet int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) 292693b174adSYuchung Cheng { 292793b174adSYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 292810d3be56SEric Dumazet int err = __tcp_retransmit_skb(sk, skb, segs); 29291da177e4SLinus Torvalds 29301da177e4SLinus Torvalds if (err == 0) { 29311da177e4SLinus Torvalds #if FASTRETRANS_DEBUG > 0 29321da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 2933e87cc472SJoe Perches net_dbg_ratelimited("retrans_out leaked\n"); 29341da177e4SLinus Torvalds } 29351da177e4SLinus Torvalds #endif 29361da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 29371da177e4SLinus Torvalds tp->retrans_out += tcp_skb_pcount(skb); 29381da177e4SLinus Torvalds 29391da177e4SLinus Torvalds /* Save stamp of the first retransmit. */ 29401da177e4SLinus Torvalds if (!tp->retrans_stamp) 29417faee5c0SEric Dumazet tp->retrans_stamp = tcp_skb_timestamp(skb); 29421da177e4SLinus Torvalds 29431da177e4SLinus Torvalds } 29446e08d5e3SYuchung Cheng 29456e08d5e3SYuchung Cheng if (tp->undo_retrans < 0) 29466e08d5e3SYuchung Cheng tp->undo_retrans = 0; 29476e08d5e3SYuchung Cheng tp->undo_retrans += tcp_skb_pcount(skb); 29481da177e4SLinus Torvalds return err; 29491da177e4SLinus Torvalds } 29501da177e4SLinus Torvalds 29511da177e4SLinus Torvalds /* This gets called after a retransmit timeout, and the initially 29521da177e4SLinus Torvalds * retransmitted data is acknowledged. It tries to continue 29531da177e4SLinus Torvalds * resending the rest of the retransmit queue, until either 29541da177e4SLinus Torvalds * we've sent it all or the congestion window limit is reached. 29551da177e4SLinus Torvalds */ 29561da177e4SLinus Torvalds void tcp_xmit_retransmit_queue(struct sock *sk) 29571da177e4SLinus Torvalds { 29586687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 2959b9f1f1ceSEric Dumazet struct sk_buff *skb, *rtx_head, *hole = NULL; 29601da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2961840a3cbeSYuchung Cheng u32 max_segs; 296261eb55f4SIlpo Järvinen int mib_idx; 29636a438bbeSStephen Hemminger 296445e77d31SIlpo Järvinen if (!tp->packets_out) 296545e77d31SIlpo Järvinen return; 296645e77d31SIlpo Järvinen 296775c119afSEric Dumazet rtx_head = tcp_rtx_queue_head(sk); 2968b9f1f1ceSEric Dumazet skb = tp->retransmit_skb_hint ?: rtx_head; 2969ed6e7268SNeal Cardwell max_segs = tcp_tso_segs(sk, tcp_current_mss(sk)); 297075c119afSEric Dumazet skb_rbtree_walk_from(skb) { 2971dca0aaf8SEric Dumazet __u8 sacked; 297210d3be56SEric Dumazet int segs; 29731da177e4SLinus Torvalds 2974218af599SEric Dumazet if (tcp_pacing_check(sk)) 2975218af599SEric Dumazet break; 2976218af599SEric Dumazet 29776a438bbeSStephen Hemminger /* we could do better than to assign each time */ 297851456b29SIan Morris if (!hole) 29796a438bbeSStephen Hemminger tp->retransmit_skb_hint = skb; 29806a438bbeSStephen Hemminger 298110d3be56SEric Dumazet segs = tp->snd_cwnd - tcp_packets_in_flight(tp); 298210d3be56SEric Dumazet if (segs <= 0) 29831da177e4SLinus Torvalds return; 2984dca0aaf8SEric Dumazet sacked = TCP_SKB_CB(skb)->sacked; 2985a3d2e9f8SEric Dumazet /* In case tcp_shift_skb_data() have aggregated large skbs, 2986a3d2e9f8SEric Dumazet * we need to make sure not sending too bigs TSO packets 2987a3d2e9f8SEric Dumazet */ 2988a3d2e9f8SEric Dumazet segs = min_t(int, segs, max_segs); 29890e1c54c2SIlpo Järvinen 2990840a3cbeSYuchung Cheng if (tp->retrans_out >= tp->lost_out) { 2991006f582cSIlpo Järvinen break; 29920e1c54c2SIlpo Järvinen } else if (!(sacked & TCPCB_LOST)) { 299351456b29SIan Morris if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) 29940e1c54c2SIlpo Järvinen hole = skb; 299561eb55f4SIlpo Järvinen continue; 29961da177e4SLinus Torvalds 29970e1c54c2SIlpo Järvinen } else { 29980e1c54c2SIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Loss) 29990e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFASTRETRANS; 30000e1c54c2SIlpo Järvinen else 30010e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; 30020e1c54c2SIlpo Järvinen } 30030e1c54c2SIlpo Järvinen 30040e1c54c2SIlpo Järvinen if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) 300561eb55f4SIlpo Järvinen continue; 300640b215e5SPavel Emelyanov 3007f9616c35SEric Dumazet if (tcp_small_queue_check(sk, skb, 1)) 3008f9616c35SEric Dumazet return; 3009f9616c35SEric Dumazet 301010d3be56SEric Dumazet if (tcp_retransmit_skb(sk, skb, segs)) 30111da177e4SLinus Torvalds return; 301224ab6becSYuchung Cheng 3013de1d6578SYuchung Cheng NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb)); 30141da177e4SLinus Torvalds 3015684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 3016a262f0cdSNandita Dukkipati tp->prr_out += tcp_skb_pcount(skb); 3017a262f0cdSNandita Dukkipati 301875c119afSEric Dumazet if (skb == rtx_head && 301957dde7f7SYuchung Cheng icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT) 3020463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 30213f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, 30223f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 30231da177e4SLinus Torvalds } 30241da177e4SLinus Torvalds } 30251da177e4SLinus Torvalds 3026d83769a5SEric Dumazet /* We allow to exceed memory limits for FIN packets to expedite 3027d83769a5SEric Dumazet * connection tear down and (memory) recovery. 3028845704a5SEric Dumazet * Otherwise tcp_send_fin() could be tempted to either delay FIN 3029845704a5SEric Dumazet * or even be forced to close flow without any FIN. 3030a6c5ea4cSEric Dumazet * In general, we want to allow one skb per socket to avoid hangs 3031a6c5ea4cSEric Dumazet * with edge trigger epoll() 3032d83769a5SEric Dumazet */ 3033a6c5ea4cSEric Dumazet void sk_forced_mem_schedule(struct sock *sk, int size) 3034d83769a5SEric Dumazet { 3035e805605cSJohannes Weiner int amt; 3036d83769a5SEric Dumazet 3037d83769a5SEric Dumazet if (size <= sk->sk_forward_alloc) 3038d83769a5SEric Dumazet return; 3039d83769a5SEric Dumazet amt = sk_mem_pages(size); 3040d83769a5SEric Dumazet sk->sk_forward_alloc += amt * SK_MEM_QUANTUM; 3041e805605cSJohannes Weiner sk_memory_allocated_add(sk, amt); 3042e805605cSJohannes Weiner 3043baac50bbSJohannes Weiner if (mem_cgroup_sockets_enabled && sk->sk_memcg) 3044baac50bbSJohannes Weiner mem_cgroup_charge_skmem(sk->sk_memcg, amt); 3045d83769a5SEric Dumazet } 3046d83769a5SEric Dumazet 3047845704a5SEric Dumazet /* Send a FIN. The caller locks the socket for us. 3048845704a5SEric Dumazet * We should try to send a FIN packet really hard, but eventually give up. 30491da177e4SLinus Torvalds */ 30501da177e4SLinus Torvalds void tcp_send_fin(struct sock *sk) 30511da177e4SLinus Torvalds { 3052845704a5SEric Dumazet struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk); 30531da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 30541da177e4SLinus Torvalds 3055845704a5SEric Dumazet /* Optimization, tack on the FIN if we have one skb in write queue and 3056845704a5SEric Dumazet * this skb was not yet sent, or we are under memory pressure. 3057845704a5SEric Dumazet * Note: in the latter case, FIN packet will be sent after a timeout, 3058845704a5SEric Dumazet * as TCP stack thinks it has already been transmitted. 30591da177e4SLinus Torvalds */ 306075c119afSEric Dumazet if (!tskb && tcp_under_memory_pressure(sk)) 306175c119afSEric Dumazet tskb = skb_rb_last(&sk->tcp_rtx_queue); 306275c119afSEric Dumazet 306375c119afSEric Dumazet if (tskb) { 3064845704a5SEric Dumazet coalesce: 3065845704a5SEric Dumazet TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN; 3066845704a5SEric Dumazet TCP_SKB_CB(tskb)->end_seq++; 30671da177e4SLinus Torvalds tp->write_seq++; 306875c119afSEric Dumazet if (tcp_write_queue_empty(sk)) { 3069845704a5SEric Dumazet /* This means tskb was already sent. 3070845704a5SEric Dumazet * Pretend we included the FIN on previous transmit. 3071845704a5SEric Dumazet * We need to set tp->snd_nxt to the value it would have 3072845704a5SEric Dumazet * if FIN had been sent. This is because retransmit path 3073845704a5SEric Dumazet * does not change tp->snd_nxt. 3074845704a5SEric Dumazet */ 3075845704a5SEric Dumazet tp->snd_nxt++; 3076845704a5SEric Dumazet return; 3077845704a5SEric Dumazet } 30781da177e4SLinus Torvalds } else { 3079845704a5SEric Dumazet skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation); 3080845704a5SEric Dumazet if (unlikely(!skb)) { 3081845704a5SEric Dumazet if (tskb) 3082845704a5SEric Dumazet goto coalesce; 3083845704a5SEric Dumazet return; 30841da177e4SLinus Torvalds } 3085e2080072SEric Dumazet INIT_LIST_HEAD(&skb->tcp_tsorted_anchor); 3086d83769a5SEric Dumazet skb_reserve(skb, MAX_TCP_HEADER); 3087a6c5ea4cSEric Dumazet sk_forced_mem_schedule(sk, skb->truesize); 30881da177e4SLinus Torvalds /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 3089e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tp->write_seq, 3090a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_FIN); 30911da177e4SLinus Torvalds tcp_queue_skb(sk, skb); 30921da177e4SLinus Torvalds } 3093845704a5SEric Dumazet __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF); 30941da177e4SLinus Torvalds } 30951da177e4SLinus Torvalds 30961da177e4SLinus Torvalds /* We get here when a process closes a file descriptor (either due to 30971da177e4SLinus Torvalds * an explicit close() or as a byproduct of exit()'ing) and there 30981da177e4SLinus Torvalds * was unread data in the receive queue. This behavior is recommended 309965bb723cSGerrit Renker * by RFC 2525, section 2.17. -DaveM 31001da177e4SLinus Torvalds */ 3101dd0fc66fSAl Viro void tcp_send_active_reset(struct sock *sk, gfp_t priority) 31021da177e4SLinus Torvalds { 31031da177e4SLinus Torvalds struct sk_buff *skb; 31041da177e4SLinus Torvalds 31057cc2b043SGao Feng TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); 31067cc2b043SGao Feng 31071da177e4SLinus Torvalds /* NOTE: No TCP options attached and we never retransmit this. */ 31081da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, priority); 31091da177e4SLinus Torvalds if (!skb) { 31104e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 31111da177e4SLinus Torvalds return; 31121da177e4SLinus Torvalds } 31131da177e4SLinus Torvalds 31141da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 31151da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 3116e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), 3117a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_RST); 31189a568de4SEric Dumazet tcp_mstamp_refresh(tcp_sk(sk)); 31191da177e4SLinus Torvalds /* Send it off. */ 3120dfb4b9dcSDavid S. Miller if (tcp_transmit_skb(sk, skb, 0, priority)) 31214e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 3122c24b14c4SSong Liu 3123c24b14c4SSong Liu /* skb of trace_tcp_send_reset() keeps the skb that caused RST, 3124c24b14c4SSong Liu * skb here is different to the troublesome skb, so use NULL 3125c24b14c4SSong Liu */ 3126c24b14c4SSong Liu trace_tcp_send_reset(sk, NULL); 31271da177e4SLinus Torvalds } 31281da177e4SLinus Torvalds 312967edfef7SAndi Kleen /* Send a crossed SYN-ACK during socket establishment. 313067edfef7SAndi Kleen * WARNING: This routine must only be called when we have already sent 31311da177e4SLinus Torvalds * a SYN packet that crossed the incoming SYN that caused this routine 31321da177e4SLinus Torvalds * to get called. If this assumption fails then the initial rcv_wnd 31331da177e4SLinus Torvalds * and rcv_wscale values will not be correct. 31341da177e4SLinus Torvalds */ 31351da177e4SLinus Torvalds int tcp_send_synack(struct sock *sk) 31361da177e4SLinus Torvalds { 31371da177e4SLinus Torvalds struct sk_buff *skb; 31381da177e4SLinus Torvalds 313975c119afSEric Dumazet skb = tcp_rtx_queue_head(sk); 314051456b29SIan Morris if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 314175c119afSEric Dumazet pr_err("%s: wrong queue state\n", __func__); 31421da177e4SLinus Torvalds return -EFAULT; 31431da177e4SLinus Torvalds } 31444de075e0SEric Dumazet if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { 31451da177e4SLinus Torvalds if (skb_cloned(skb)) { 3146e2080072SEric Dumazet struct sk_buff *nskb; 3147e2080072SEric Dumazet 3148e2080072SEric Dumazet tcp_skb_tsorted_save(skb) { 3149e2080072SEric Dumazet nskb = skb_copy(skb, GFP_ATOMIC); 3150e2080072SEric Dumazet } tcp_skb_tsorted_restore(skb); 315151456b29SIan Morris if (!nskb) 31521da177e4SLinus Torvalds return -ENOMEM; 3153e2080072SEric Dumazet INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor); 315475c119afSEric Dumazet tcp_rtx_queue_unlink_and_free(skb, sk); 3155f4a775d1SEric Dumazet __skb_header_release(nskb); 315675c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb); 31573ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 31583ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 31591da177e4SLinus Torvalds skb = nskb; 31601da177e4SLinus Torvalds } 31611da177e4SLinus Torvalds 31624de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; 3163735d3831SFlorian Westphal tcp_ecn_send_synack(sk, skb); 31641da177e4SLinus Torvalds } 3165dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 31661da177e4SLinus Torvalds } 31671da177e4SLinus Torvalds 31684aea39c1SEric Dumazet /** 31694aea39c1SEric Dumazet * tcp_make_synack - Prepare a SYN-ACK. 31704aea39c1SEric Dumazet * sk: listener socket 31714aea39c1SEric Dumazet * dst: dst entry attached to the SYNACK 31724aea39c1SEric Dumazet * req: request_sock pointer 31734aea39c1SEric Dumazet * 31744aea39c1SEric Dumazet * Allocate one skb and build a SYNACK packet. 31754aea39c1SEric Dumazet * @dst is consumed : Caller should not use it again. 31764aea39c1SEric Dumazet */ 31775d062de7SEric Dumazet struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst, 3178e6b4d113SWilliam Allen Simpson struct request_sock *req, 3179ca6fb065SEric Dumazet struct tcp_fastopen_cookie *foc, 3180b3d05147SEric Dumazet enum tcp_synack_type synack_type) 31811da177e4SLinus Torvalds { 31822e6599cbSArnaldo Carvalho de Melo struct inet_request_sock *ireq = inet_rsk(req); 31835d062de7SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 318480f03e27SEric Dumazet struct tcp_md5sig_key *md5 = NULL; 31855d062de7SEric Dumazet struct tcp_out_options opts; 31865d062de7SEric Dumazet struct sk_buff *skb; 3187bd0388aeSWilliam Allen Simpson int tcp_header_size; 31885d062de7SEric Dumazet struct tcphdr *th; 3189f5fff5dcSTom Quetchenbach int mss; 31901da177e4SLinus Torvalds 3191ca6fb065SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 31924aea39c1SEric Dumazet if (unlikely(!skb)) { 31934aea39c1SEric Dumazet dst_release(dst); 31941da177e4SLinus Torvalds return NULL; 31954aea39c1SEric Dumazet } 31961da177e4SLinus Torvalds /* Reserve space for headers. */ 31971da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 31981da177e4SLinus Torvalds 3199b3d05147SEric Dumazet switch (synack_type) { 3200b3d05147SEric Dumazet case TCP_SYNACK_NORMAL: 32019e17f8a4SEric Dumazet skb_set_owner_w(skb, req_to_sk(req)); 3202b3d05147SEric Dumazet break; 3203b3d05147SEric Dumazet case TCP_SYNACK_COOKIE: 3204b3d05147SEric Dumazet /* Under synflood, we do not attach skb to a socket, 3205b3d05147SEric Dumazet * to avoid false sharing. 3206b3d05147SEric Dumazet */ 3207b3d05147SEric Dumazet break; 3208b3d05147SEric Dumazet case TCP_SYNACK_FASTOPEN: 3209ca6fb065SEric Dumazet /* sk is a const pointer, because we want to express multiple 3210ca6fb065SEric Dumazet * cpu might call us concurrently. 3211ca6fb065SEric Dumazet * sk->sk_wmem_alloc in an atomic, we can promote to rw. 3212ca6fb065SEric Dumazet */ 3213ca6fb065SEric Dumazet skb_set_owner_w(skb, (struct sock *)sk); 3214b3d05147SEric Dumazet break; 3215ca6fb065SEric Dumazet } 32164aea39c1SEric Dumazet skb_dst_set(skb, dst); 32171da177e4SLinus Torvalds 32183541f9e8SEric Dumazet mss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 3219f5fff5dcSTom Quetchenbach 322033ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 32218b5f12d0SFlorian Westphal #ifdef CONFIG_SYN_COOKIES 32228b5f12d0SFlorian Westphal if (unlikely(req->cookie_ts)) 3223*d3edd06eSEric Dumazet skb->skb_mstamp_ns = cookie_init_timestamp(req); 32248b5f12d0SFlorian Westphal else 32258b5f12d0SFlorian Westphal #endif 3226*d3edd06eSEric Dumazet skb->skb_mstamp_ns = tcp_clock_ns(); 322780f03e27SEric Dumazet 322880f03e27SEric Dumazet #ifdef CONFIG_TCP_MD5SIG 322980f03e27SEric Dumazet rcu_read_lock(); 3230fd3a154aSEric Dumazet md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req)); 323180f03e27SEric Dumazet #endif 323258d607d3SEric Dumazet skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4); 323360e2a778SUrsula Braun tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5, 323460e2a778SUrsula Braun foc) + sizeof(*th); 323533ad798cSAdam Langley 3236aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 3237aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 32381da177e4SLinus Torvalds 3239ea1627c2SEric Dumazet th = (struct tcphdr *)skb->data; 32401da177e4SLinus Torvalds memset(th, 0, sizeof(struct tcphdr)); 32411da177e4SLinus Torvalds th->syn = 1; 32421da177e4SLinus Torvalds th->ack = 1; 32436ac705b1SEric Dumazet tcp_ecn_make_synack(req, th); 3244b44084c2SEric Dumazet th->source = htons(ireq->ir_num); 3245634fb979SEric Dumazet th->dest = ireq->ir_rmt_port; 3246e05a90ecSJamal Hadi Salim skb->mark = ireq->ir_mark; 32473b117750SEric Dumazet skb->ip_summed = CHECKSUM_PARTIAL; 32483b117750SEric Dumazet th->seq = htonl(tcp_rsk(req)->snt_isn); 32498336886fSJerry Chu /* XXX data is queued and acked as is. No buffer/window check */ 32508336886fSJerry Chu th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt); 32511da177e4SLinus Torvalds 32521da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 3253ed53d0abSEric Dumazet th->window = htons(min(req->rsk_rcv_wnd, 65535U)); 32545d062de7SEric Dumazet tcp_options_write((__be32 *)(th + 1), NULL, &opts); 32551da177e4SLinus Torvalds th->doff = (tcp_header_size >> 2); 325690bbcc60SEric Dumazet __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS); 3257cfb6eeb4SYOSHIFUJI Hideaki 3258cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 3259cfb6eeb4SYOSHIFUJI Hideaki /* Okay, we have all we need - do the md5 hash if needed */ 326080f03e27SEric Dumazet if (md5) 3261bd0388aeSWilliam Allen Simpson tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, 326239f8e58eSEric Dumazet md5, req_to_sk(req), skb); 326380f03e27SEric Dumazet rcu_read_unlock(); 3264cfb6eeb4SYOSHIFUJI Hideaki #endif 3265cfb6eeb4SYOSHIFUJI Hideaki 3266b50edd78SEric Dumazet /* Do not fool tcpdump (if any), clean our debris */ 32672456e855SThomas Gleixner skb->tstamp = 0; 32681da177e4SLinus Torvalds return skb; 32691da177e4SLinus Torvalds } 32704bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_make_synack); 32711da177e4SLinus Torvalds 327281164413SDaniel Borkmann static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst) 327381164413SDaniel Borkmann { 327481164413SDaniel Borkmann struct inet_connection_sock *icsk = inet_csk(sk); 327581164413SDaniel Borkmann const struct tcp_congestion_ops *ca; 327681164413SDaniel Borkmann u32 ca_key = dst_metric(dst, RTAX_CC_ALGO); 327781164413SDaniel Borkmann 327881164413SDaniel Borkmann if (ca_key == TCP_CA_UNSPEC) 327981164413SDaniel Borkmann return; 328081164413SDaniel Borkmann 328181164413SDaniel Borkmann rcu_read_lock(); 328281164413SDaniel Borkmann ca = tcp_ca_find_key(ca_key); 328381164413SDaniel Borkmann if (likely(ca && try_module_get(ca->owner))) { 328481164413SDaniel Borkmann module_put(icsk->icsk_ca_ops->owner); 328581164413SDaniel Borkmann icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst); 328681164413SDaniel Borkmann icsk->icsk_ca_ops = ca; 328781164413SDaniel Borkmann } 328881164413SDaniel Borkmann rcu_read_unlock(); 328981164413SDaniel Borkmann } 329081164413SDaniel Borkmann 329167edfef7SAndi Kleen /* Do all connect socket setups that can be done AF independent. */ 3292f7e56a76Sstephen hemminger static void tcp_connect_init(struct sock *sk) 32931da177e4SLinus Torvalds { 3294cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 32951da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 32961da177e4SLinus Torvalds __u8 rcv_wscale; 329713d3b1ebSLawrence Brakmo u32 rcv_wnd; 32981da177e4SLinus Torvalds 32991da177e4SLinus Torvalds /* We'll fix this up when we get a response from the other end. 33001da177e4SLinus Torvalds * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 33011da177e4SLinus Torvalds */ 33025d2ed052SEric Dumazet tp->tcp_header_len = sizeof(struct tcphdr); 33035d2ed052SEric Dumazet if (sock_net(sk)->ipv4.sysctl_tcp_timestamps) 33045d2ed052SEric Dumazet tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED; 33051da177e4SLinus Torvalds 3306cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 330700db4124SIan Morris if (tp->af_specific->md5_lookup(sk, sk)) 3308cfb6eeb4SYOSHIFUJI Hideaki tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 3309cfb6eeb4SYOSHIFUJI Hideaki #endif 3310cfb6eeb4SYOSHIFUJI Hideaki 33111da177e4SLinus Torvalds /* If user gave his TCP_MAXSEG, record it to clamp */ 33121da177e4SLinus Torvalds if (tp->rx_opt.user_mss) 33131da177e4SLinus Torvalds tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 33141da177e4SLinus Torvalds tp->max_window = 0; 33155d424d5aSJohn Heffner tcp_mtup_init(sk); 33161da177e4SLinus Torvalds tcp_sync_mss(sk, dst_mtu(dst)); 33171da177e4SLinus Torvalds 331881164413SDaniel Borkmann tcp_ca_dst_init(sk, dst); 331981164413SDaniel Borkmann 33201da177e4SLinus Torvalds if (!tp->window_clamp) 33211da177e4SLinus Torvalds tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 33223541f9e8SEric Dumazet tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 3323f5fff5dcSTom Quetchenbach 33241da177e4SLinus Torvalds tcp_initialize_rcv_mss(sk); 33251da177e4SLinus Torvalds 3326e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 3327e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 3328e88c64f0SHagen Paul Pfeifer (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) 3329e88c64f0SHagen Paul Pfeifer tp->window_clamp = tcp_full_space(sk); 3330e88c64f0SHagen Paul Pfeifer 333113d3b1ebSLawrence Brakmo rcv_wnd = tcp_rwnd_init_bpf(sk); 333213d3b1ebSLawrence Brakmo if (rcv_wnd == 0) 333313d3b1ebSLawrence Brakmo rcv_wnd = dst_metric(dst, RTAX_INITRWND); 333413d3b1ebSLawrence Brakmo 3335ceef9ab6SEric Dumazet tcp_select_initial_window(sk, tcp_full_space(sk), 33361da177e4SLinus Torvalds tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 33371da177e4SLinus Torvalds &tp->rcv_wnd, 33381da177e4SLinus Torvalds &tp->window_clamp, 33399bb37ef0SEric Dumazet sock_net(sk)->ipv4.sysctl_tcp_window_scaling, 334031d12926Slaurent chavey &rcv_wscale, 334113d3b1ebSLawrence Brakmo rcv_wnd); 33421da177e4SLinus Torvalds 33431da177e4SLinus Torvalds tp->rx_opt.rcv_wscale = rcv_wscale; 33441da177e4SLinus Torvalds tp->rcv_ssthresh = tp->rcv_wnd; 33451da177e4SLinus Torvalds 33461da177e4SLinus Torvalds sk->sk_err = 0; 33471da177e4SLinus Torvalds sock_reset_flag(sk, SOCK_DONE); 33481da177e4SLinus Torvalds tp->snd_wnd = 0; 3349ee7537b6SHantzis Fotis tcp_init_wl(tp, 0); 33507f582b24SEric Dumazet tcp_write_queue_purge(sk); 33511da177e4SLinus Torvalds tp->snd_una = tp->write_seq; 33521da177e4SLinus Torvalds tp->snd_sml = tp->write_seq; 335333f5f57eSIlpo Järvinen tp->snd_up = tp->write_seq; 3354370816aeSPavel Emelyanov tp->snd_nxt = tp->write_seq; 3355ee995283SPavel Emelyanov 3356ee995283SPavel Emelyanov if (likely(!tp->repair)) 33571da177e4SLinus Torvalds tp->rcv_nxt = 0; 3358c7781a6eSAndrew Vagin else 335970eabf0eSEric Dumazet tp->rcv_tstamp = tcp_jiffies32; 3360ee995283SPavel Emelyanov tp->rcv_wup = tp->rcv_nxt; 3361ee995283SPavel Emelyanov tp->copied_seq = tp->rcv_nxt; 33621da177e4SLinus Torvalds 33638550f328SLawrence Brakmo inet_csk(sk)->icsk_rto = tcp_timeout_init(sk); 3364463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_retransmits = 0; 33651da177e4SLinus Torvalds tcp_clear_retrans(tp); 33661da177e4SLinus Torvalds } 33671da177e4SLinus Torvalds 3368783237e8SYuchung Cheng static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) 3369783237e8SYuchung Cheng { 3370783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3371783237e8SYuchung Cheng struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 3372783237e8SYuchung Cheng 3373783237e8SYuchung Cheng tcb->end_seq += skb->len; 3374f4a775d1SEric Dumazet __skb_header_release(skb); 3375783237e8SYuchung Cheng sk->sk_wmem_queued += skb->truesize; 3376783237e8SYuchung Cheng sk_mem_charge(sk, skb->truesize); 3377783237e8SYuchung Cheng tp->write_seq = tcb->end_seq; 3378783237e8SYuchung Cheng tp->packets_out += tcp_skb_pcount(skb); 3379783237e8SYuchung Cheng } 3380783237e8SYuchung Cheng 3381783237e8SYuchung Cheng /* Build and send a SYN with data and (cached) Fast Open cookie. However, 3382783237e8SYuchung Cheng * queue a data-only packet after the regular SYN, such that regular SYNs 3383783237e8SYuchung Cheng * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges 3384783237e8SYuchung Cheng * only the SYN sequence, the data are retransmitted in the first ACK. 3385783237e8SYuchung Cheng * If cookie is not cached or other error occurs, falls back to send a 3386783237e8SYuchung Cheng * regular SYN with Fast Open cookie request option. 3387783237e8SYuchung Cheng */ 3388783237e8SYuchung Cheng static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) 3389783237e8SYuchung Cheng { 3390783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3391783237e8SYuchung Cheng struct tcp_fastopen_request *fo = tp->fastopen_req; 3392065263f4SWei Wang int space, err = 0; 3393355a901eSEric Dumazet struct sk_buff *syn_data; 3394783237e8SYuchung Cheng 339567da22d2SYuchung Cheng tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ 3396065263f4SWei Wang if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie)) 3397783237e8SYuchung Cheng goto fallback; 3398783237e8SYuchung Cheng 3399783237e8SYuchung Cheng /* MSS for SYN-data is based on cached MSS and bounded by PMTU and 3400783237e8SYuchung Cheng * user-MSS. Reserve maximum option space for middleboxes that add 3401783237e8SYuchung Cheng * private TCP options. The cost is reduced data space in SYN :( 3402783237e8SYuchung Cheng */ 34033541f9e8SEric Dumazet tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp); 34043541f9e8SEric Dumazet 34051b63edd6SYuchung Cheng space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) - 3406783237e8SYuchung Cheng MAX_TCP_OPTION_SPACE; 3407783237e8SYuchung Cheng 3408f5ddcbbbSEric Dumazet space = min_t(size_t, space, fo->size); 3409f5ddcbbbSEric Dumazet 3410f5ddcbbbSEric Dumazet /* limit to order-0 allocations */ 3411f5ddcbbbSEric Dumazet space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER)); 3412f5ddcbbbSEric Dumazet 3413eb934478SEric Dumazet syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false); 3414355a901eSEric Dumazet if (!syn_data) 3415783237e8SYuchung Cheng goto fallback; 3416355a901eSEric Dumazet syn_data->ip_summed = CHECKSUM_PARTIAL; 3417355a901eSEric Dumazet memcpy(syn_data->cb, syn->cb, sizeof(syn->cb)); 341807e100f9SEric Dumazet if (space) { 341907e100f9SEric Dumazet int copied = copy_from_iter(skb_put(syn_data, space), space, 342057be5bdaSAl Viro &fo->data->msg_iter); 342157be5bdaSAl Viro if (unlikely(!copied)) { 3422ba233b34SEric Dumazet tcp_skb_tsorted_anchor_cleanup(syn_data); 3423355a901eSEric Dumazet kfree_skb(syn_data); 3424783237e8SYuchung Cheng goto fallback; 3425783237e8SYuchung Cheng } 342657be5bdaSAl Viro if (copied != space) { 342757be5bdaSAl Viro skb_trim(syn_data, copied); 342857be5bdaSAl Viro space = copied; 342957be5bdaSAl Viro } 343007e100f9SEric Dumazet } 3431355a901eSEric Dumazet /* No more data pending in inet_wait_for_connect() */ 3432355a901eSEric Dumazet if (space == fo->size) 3433355a901eSEric Dumazet fo->data = NULL; 3434355a901eSEric Dumazet fo->copied = space; 3435783237e8SYuchung Cheng 3436355a901eSEric Dumazet tcp_connect_queue_skb(sk, syn_data); 34370f87230dSFrancis Yan if (syn_data->len) 34380f87230dSFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_BUSY); 3439355a901eSEric Dumazet 3440355a901eSEric Dumazet err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation); 3441355a901eSEric Dumazet 3442*d3edd06eSEric Dumazet syn->skb_mstamp_ns = syn_data->skb_mstamp_ns; 3443355a901eSEric Dumazet 3444355a901eSEric Dumazet /* Now full SYN+DATA was cloned and sent (or not), 3445355a901eSEric Dumazet * remove the SYN from the original skb (syn_data) 3446355a901eSEric Dumazet * we keep in write queue in case of a retransmit, as we 3447355a901eSEric Dumazet * also have the SYN packet (with no data) in the same queue. 3448431a9124SEric Dumazet */ 3449355a901eSEric Dumazet TCP_SKB_CB(syn_data)->seq++; 3450355a901eSEric Dumazet TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH; 3451355a901eSEric Dumazet if (!err) { 345267da22d2SYuchung Cheng tp->syn_data = (fo->copied > 0); 345375c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data); 3454f19c29e3SYuchung Cheng NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT); 3455783237e8SYuchung Cheng goto done; 3456783237e8SYuchung Cheng } 3457783237e8SYuchung Cheng 345875c119afSEric Dumazet /* data was not sent, put it in write_queue */ 345975c119afSEric Dumazet __skb_queue_tail(&sk->sk_write_queue, syn_data); 3460b5b7db8dSEric Dumazet tp->packets_out -= tcp_skb_pcount(syn_data); 3461b5b7db8dSEric Dumazet 3462783237e8SYuchung Cheng fallback: 3463783237e8SYuchung Cheng /* Send a regular SYN with Fast Open cookie request option */ 3464783237e8SYuchung Cheng if (fo->cookie.len > 0) 3465783237e8SYuchung Cheng fo->cookie.len = 0; 3466783237e8SYuchung Cheng err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); 3467783237e8SYuchung Cheng if (err) 3468783237e8SYuchung Cheng tp->syn_fastopen = 0; 3469783237e8SYuchung Cheng done: 3470783237e8SYuchung Cheng fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ 3471783237e8SYuchung Cheng return err; 3472783237e8SYuchung Cheng } 3473783237e8SYuchung Cheng 347467edfef7SAndi Kleen /* Build a SYN and send it off. */ 34751da177e4SLinus Torvalds int tcp_connect(struct sock *sk) 34761da177e4SLinus Torvalds { 34771da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 34781da177e4SLinus Torvalds struct sk_buff *buff; 3479ee586811SEric Paris int err; 34801da177e4SLinus Torvalds 3481de525be2SLawrence Brakmo tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB, 0, NULL); 34828ba60924SEric Dumazet 34838ba60924SEric Dumazet if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 34848ba60924SEric Dumazet return -EHOSTUNREACH; /* Routing failure or similar. */ 34858ba60924SEric Dumazet 34861da177e4SLinus Torvalds tcp_connect_init(sk); 34871da177e4SLinus Torvalds 34882b916477SAndrey Vagin if (unlikely(tp->repair)) { 34892b916477SAndrey Vagin tcp_finish_connect(sk, NULL); 34902b916477SAndrey Vagin return 0; 34912b916477SAndrey Vagin } 34922b916477SAndrey Vagin 3493eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true); 3494355a901eSEric Dumazet if (unlikely(!buff)) 34951da177e4SLinus Torvalds return -ENOBUFS; 34961da177e4SLinus Torvalds 3497a3433f35SChangli Gao tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); 34989a568de4SEric Dumazet tcp_mstamp_refresh(tp); 34999a568de4SEric Dumazet tp->retrans_stamp = tcp_time_stamp(tp); 3500783237e8SYuchung Cheng tcp_connect_queue_skb(sk, buff); 3501735d3831SFlorian Westphal tcp_ecn_send_syn(sk, buff); 350275c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, buff); 35031da177e4SLinus Torvalds 3504783237e8SYuchung Cheng /* Send off SYN; include data in Fast Open. */ 3505783237e8SYuchung Cheng err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : 3506783237e8SYuchung Cheng tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); 3507ee586811SEric Paris if (err == -ECONNREFUSED) 3508ee586811SEric Paris return err; 3509bd37a088SWei Yongjun 3510bd37a088SWei Yongjun /* We change tp->snd_nxt after the tcp_transmit_skb() call 3511bd37a088SWei Yongjun * in order to make this packet get counted in tcpOutSegs. 3512bd37a088SWei Yongjun */ 3513bd37a088SWei Yongjun tp->snd_nxt = tp->write_seq; 3514bd37a088SWei Yongjun tp->pushed_seq = tp->write_seq; 3515b5b7db8dSEric Dumazet buff = tcp_send_head(sk); 3516b5b7db8dSEric Dumazet if (unlikely(buff)) { 3517b5b7db8dSEric Dumazet tp->snd_nxt = TCP_SKB_CB(buff)->seq; 3518b5b7db8dSEric Dumazet tp->pushed_seq = TCP_SKB_CB(buff)->seq; 3519b5b7db8dSEric Dumazet } 352081cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); 35211da177e4SLinus Torvalds 35221da177e4SLinus Torvalds /* Timer for repeating the SYN until an answer. */ 35233f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 35243f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 35251da177e4SLinus Torvalds return 0; 35261da177e4SLinus Torvalds } 35274bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_connect); 35281da177e4SLinus Torvalds 35291da177e4SLinus Torvalds /* Send out a delayed ack, the caller does the policy checking 35301da177e4SLinus Torvalds * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 35311da177e4SLinus Torvalds * for details. 35321da177e4SLinus Torvalds */ 35331da177e4SLinus Torvalds void tcp_send_delayed_ack(struct sock *sk) 35341da177e4SLinus Torvalds { 3535463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 3536463c84b9SArnaldo Carvalho de Melo int ato = icsk->icsk_ack.ato; 35371da177e4SLinus Torvalds unsigned long timeout; 35381da177e4SLinus Torvalds 35391da177e4SLinus Torvalds if (ato > TCP_DELACK_MIN) { 3540463c84b9SArnaldo Carvalho de Melo const struct tcp_sock *tp = tcp_sk(sk); 35411da177e4SLinus Torvalds int max_ato = HZ / 2; 35421da177e4SLinus Torvalds 3543056834d9SIlpo Järvinen if (icsk->icsk_ack.pingpong || 3544056834d9SIlpo Järvinen (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 35451da177e4SLinus Torvalds max_ato = TCP_DELACK_MAX; 35461da177e4SLinus Torvalds 35471da177e4SLinus Torvalds /* Slow path, intersegment interval is "high". */ 35481da177e4SLinus Torvalds 35491da177e4SLinus Torvalds /* If some rtt estimate is known, use it to bound delayed ack. 3550463c84b9SArnaldo Carvalho de Melo * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 35511da177e4SLinus Torvalds * directly. 35521da177e4SLinus Torvalds */ 3553740b0f18SEric Dumazet if (tp->srtt_us) { 3554740b0f18SEric Dumazet int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3), 3555740b0f18SEric Dumazet TCP_DELACK_MIN); 35561da177e4SLinus Torvalds 35571da177e4SLinus Torvalds if (rtt < max_ato) 35581da177e4SLinus Torvalds max_ato = rtt; 35591da177e4SLinus Torvalds } 35601da177e4SLinus Torvalds 35611da177e4SLinus Torvalds ato = min(ato, max_ato); 35621da177e4SLinus Torvalds } 35631da177e4SLinus Torvalds 35641da177e4SLinus Torvalds /* Stay within the limit we were given */ 35651da177e4SLinus Torvalds timeout = jiffies + ato; 35661da177e4SLinus Torvalds 35671da177e4SLinus Torvalds /* Use new timeout only if there wasn't a older one earlier. */ 3568463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 35691da177e4SLinus Torvalds /* If delack timer was blocked or is about to expire, 35701da177e4SLinus Torvalds * send ACK now. 35711da177e4SLinus Torvalds */ 3572463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.blocked || 3573463c84b9SArnaldo Carvalho de Melo time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 35741da177e4SLinus Torvalds tcp_send_ack(sk); 35751da177e4SLinus Torvalds return; 35761da177e4SLinus Torvalds } 35771da177e4SLinus Torvalds 3578463c84b9SArnaldo Carvalho de Melo if (!time_before(timeout, icsk->icsk_ack.timeout)) 3579463c84b9SArnaldo Carvalho de Melo timeout = icsk->icsk_ack.timeout; 35801da177e4SLinus Torvalds } 3581463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 3582463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.timeout = timeout; 3583463c84b9SArnaldo Carvalho de Melo sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 35841da177e4SLinus Torvalds } 35851da177e4SLinus Torvalds 35861da177e4SLinus Torvalds /* This routine sends an ack and also updates the window. */ 35872987babbSYuchung Cheng void __tcp_send_ack(struct sock *sk, u32 rcv_nxt) 35881da177e4SLinus Torvalds { 35891da177e4SLinus Torvalds struct sk_buff *buff; 35901da177e4SLinus Torvalds 3591058dc334SIlpo Järvinen /* If we have been reset, we may not send again. */ 3592058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3593058dc334SIlpo Järvinen return; 3594058dc334SIlpo Järvinen 35951da177e4SLinus Torvalds /* We are not putting this on the write queue, so 35961da177e4SLinus Torvalds * tcp_transmit_skb() will set the ownership to this 35971da177e4SLinus Torvalds * sock. 35981da177e4SLinus Torvalds */ 35997450aaf6SEric Dumazet buff = alloc_skb(MAX_TCP_HEADER, 36007450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 36017450aaf6SEric Dumazet if (unlikely(!buff)) { 3602463c84b9SArnaldo Carvalho de Melo inet_csk_schedule_ack(sk); 3603463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; 36043f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 36053f421baaSArnaldo Carvalho de Melo TCP_DELACK_MAX, TCP_RTO_MAX); 36061da177e4SLinus Torvalds return; 36071da177e4SLinus Torvalds } 36081da177e4SLinus Torvalds 36091da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 36101da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 3611a3433f35SChangli Gao tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); 36121da177e4SLinus Torvalds 361398781965SEric Dumazet /* We do not want pure acks influencing TCP Small Queues or fq/pacing 361498781965SEric Dumazet * too much. 361598781965SEric Dumazet * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784 361698781965SEric Dumazet */ 361798781965SEric Dumazet skb_set_tcp_pure_ack(buff); 361898781965SEric Dumazet 36191da177e4SLinus Torvalds /* Send it off, this clears delayed acks for us. */ 36202987babbSYuchung Cheng __tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0, rcv_nxt); 36211da177e4SLinus Torvalds } 362227cde44aSYuchung Cheng EXPORT_SYMBOL_GPL(__tcp_send_ack); 36232987babbSYuchung Cheng 36242987babbSYuchung Cheng void tcp_send_ack(struct sock *sk) 36252987babbSYuchung Cheng { 36262987babbSYuchung Cheng __tcp_send_ack(sk, tcp_sk(sk)->rcv_nxt); 36271da177e4SLinus Torvalds } 36281da177e4SLinus Torvalds 36291da177e4SLinus Torvalds /* This routine sends a packet with an out of date sequence 36301da177e4SLinus Torvalds * number. It assumes the other end will try to ack it. 36311da177e4SLinus Torvalds * 36321da177e4SLinus Torvalds * Question: what should we make while urgent mode? 36331da177e4SLinus Torvalds * 4.4BSD forces sending single byte of data. We cannot send 36341da177e4SLinus Torvalds * out of window data, because we have SND.NXT==SND.MAX... 36351da177e4SLinus Torvalds * 36361da177e4SLinus Torvalds * Current solution: to send TWO zero-length segments in urgent mode: 36371da177e4SLinus Torvalds * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 36381da177e4SLinus Torvalds * out-of-date with SND.UNA-1 to probe window. 36391da177e4SLinus Torvalds */ 3640e520af48SEric Dumazet static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib) 36411da177e4SLinus Torvalds { 36421da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 36431da177e4SLinus Torvalds struct sk_buff *skb; 36441da177e4SLinus Torvalds 36451da177e4SLinus Torvalds /* We don't queue it, tcp_transmit_skb() sets ownership. */ 36467450aaf6SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, 36477450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 364851456b29SIan Morris if (!skb) 36491da177e4SLinus Torvalds return -1; 36501da177e4SLinus Torvalds 36511da177e4SLinus Torvalds /* Reserve space for headers and set control bits. */ 36521da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 36531da177e4SLinus Torvalds /* Use a previous sequence. This should cause the other 36541da177e4SLinus Torvalds * end to send an ack. Don't queue or clone SKB, just 36551da177e4SLinus Torvalds * send it. 36561da177e4SLinus Torvalds */ 3657a3433f35SChangli Gao tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); 3658e2e8009fSRenato Westphal NET_INC_STATS(sock_net(sk), mib); 36597450aaf6SEric Dumazet return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0); 36601da177e4SLinus Torvalds } 36611da177e4SLinus Torvalds 3662385e2070SEric Dumazet /* Called from setsockopt( ... TCP_REPAIR ) */ 3663ee995283SPavel Emelyanov void tcp_send_window_probe(struct sock *sk) 3664ee995283SPavel Emelyanov { 3665ee995283SPavel Emelyanov if (sk->sk_state == TCP_ESTABLISHED) { 3666ee995283SPavel Emelyanov tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; 36679a568de4SEric Dumazet tcp_mstamp_refresh(tcp_sk(sk)); 3668e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE); 3669ee995283SPavel Emelyanov } 3670ee995283SPavel Emelyanov } 3671ee995283SPavel Emelyanov 367267edfef7SAndi Kleen /* Initiate keepalive or window probe from timer. */ 3673e520af48SEric Dumazet int tcp_write_wakeup(struct sock *sk, int mib) 36741da177e4SLinus Torvalds { 36751da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 36761da177e4SLinus Torvalds struct sk_buff *skb; 36771da177e4SLinus Torvalds 3678058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3679058dc334SIlpo Järvinen return -1; 3680058dc334SIlpo Järvinen 368100db4124SIan Morris skb = tcp_send_head(sk); 368200db4124SIan Morris if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { 36831da177e4SLinus Torvalds int err; 36840c54b85fSIlpo Järvinen unsigned int mss = tcp_current_mss(sk); 368590840defSIlpo Järvinen unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 36861da177e4SLinus Torvalds 36871da177e4SLinus Torvalds if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 36881da177e4SLinus Torvalds tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 36891da177e4SLinus Torvalds 36901da177e4SLinus Torvalds /* We are probing the opening of a window 36911da177e4SLinus Torvalds * but the window size is != 0 36921da177e4SLinus Torvalds * must have been a result SWS avoidance ( sender ) 36931da177e4SLinus Torvalds */ 36941da177e4SLinus Torvalds if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 36951da177e4SLinus Torvalds skb->len > mss) { 36961da177e4SLinus Torvalds seg_size = min(seg_size, mss); 36974de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 369875c119afSEric Dumazet if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE, 369975c119afSEric Dumazet skb, seg_size, mss, GFP_ATOMIC)) 37001da177e4SLinus Torvalds return -1; 37011da177e4SLinus Torvalds } else if (!tcp_skb_pcount(skb)) 37025bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss); 37031da177e4SLinus Torvalds 37044de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 3705dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 370666f5fe62SIlpo Järvinen if (!err) 370766f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 37081da177e4SLinus Torvalds return err; 37091da177e4SLinus Torvalds } else { 371033f5f57eSIlpo Järvinen if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) 3711e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 1, mib); 3712e520af48SEric Dumazet return tcp_xmit_probe_skb(sk, 0, mib); 37131da177e4SLinus Torvalds } 37141da177e4SLinus Torvalds } 37151da177e4SLinus Torvalds 37161da177e4SLinus Torvalds /* A window probe timeout has occurred. If window is not closed send 37171da177e4SLinus Torvalds * a partial packet else a zero probe. 37181da177e4SLinus Torvalds */ 37191da177e4SLinus Torvalds void tcp_send_probe0(struct sock *sk) 37201da177e4SLinus Torvalds { 3721463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 37221da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 3723c6214a97SNikolay Borisov struct net *net = sock_net(sk); 3724fcdd1cf4SEric Dumazet unsigned long probe_max; 37251da177e4SLinus Torvalds int err; 37261da177e4SLinus Torvalds 3727e520af48SEric Dumazet err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE); 37281da177e4SLinus Torvalds 372975c119afSEric Dumazet if (tp->packets_out || tcp_write_queue_empty(sk)) { 37301da177e4SLinus Torvalds /* Cancel probe timer, if it is not required. */ 37316687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 0; 3732463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff = 0; 37331da177e4SLinus Torvalds return; 37341da177e4SLinus Torvalds } 37351da177e4SLinus Torvalds 37361da177e4SLinus Torvalds if (err <= 0) { 3737c6214a97SNikolay Borisov if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2) 3738463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff++; 37396687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out++; 3740fcdd1cf4SEric Dumazet probe_max = TCP_RTO_MAX; 37411da177e4SLinus Torvalds } else { 37421da177e4SLinus Torvalds /* If packet was not sent due to local congestion, 37436687e988SArnaldo Carvalho de Melo * do not backoff and do not remember icsk_probes_out. 37441da177e4SLinus Torvalds * Let local senders to fight for local resources. 37451da177e4SLinus Torvalds * 37461da177e4SLinus Torvalds * Use accumulated backoff yet. 37471da177e4SLinus Torvalds */ 37486687e988SArnaldo Carvalho de Melo if (!icsk->icsk_probes_out) 37496687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 1; 3750fcdd1cf4SEric Dumazet probe_max = TCP_RESOURCE_PROBE_INTERVAL; 37511da177e4SLinus Torvalds } 3752fcdd1cf4SEric Dumazet inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 375321c8fe99SEric Dumazet tcp_probe0_when(sk, probe_max), 3754fcdd1cf4SEric Dumazet TCP_RTO_MAX); 37551da177e4SLinus Torvalds } 37565db92c99SOctavian Purdila 3757ea3bea3aSEric Dumazet int tcp_rtx_synack(const struct sock *sk, struct request_sock *req) 37585db92c99SOctavian Purdila { 37595db92c99SOctavian Purdila const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific; 37605db92c99SOctavian Purdila struct flowi fl; 37615db92c99SOctavian Purdila int res; 37625db92c99SOctavian Purdila 376358d607d3SEric Dumazet tcp_rsk(req)->txhash = net_tx_rndhash(); 3764b3d05147SEric Dumazet res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL); 37655db92c99SOctavian Purdila if (!res) { 376690bbcc60SEric Dumazet __TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); 376702a1d6e7SEric Dumazet __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 37687e32b443SYuchung Cheng if (unlikely(tcp_passive_fastopen(sk))) 37697e32b443SYuchung Cheng tcp_sk(sk)->total_retrans++; 3770cf34ce3dSSong Liu trace_tcp_retransmit_synack(sk, req); 37715db92c99SOctavian Purdila } 37725db92c99SOctavian Purdila return res; 37735db92c99SOctavian Purdila } 37745db92c99SOctavian Purdila EXPORT_SYMBOL(tcp_rtx_synack); 3775