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> 441da177e4SLinus Torvalds 451da177e4SLinus Torvalds /* People can turn this off for buggy TCP's found in printers etc. */ 46ab32ea5dSBrian Haley int sysctl_tcp_retrans_collapse __read_mostly = 1; 471da177e4SLinus Torvalds 4815d99e02SRick Jones /* People can turn this on to work with those rare, broken TCPs that 4915d99e02SRick Jones * interpret the window field as a signed quantity. 5015d99e02SRick Jones */ 51ab32ea5dSBrian Haley int sysctl_tcp_workaround_signed_windows __read_mostly = 0; 5215d99e02SRick Jones 53c39c4c6aSWei Liu /* Default TSQ limit of four TSO segments */ 54c39c4c6aSWei Liu int sysctl_tcp_limit_output_bytes __read_mostly = 262144; 5546d3ceabSEric Dumazet 561da177e4SLinus Torvalds /* This limits the percentage of the congestion window which we 571da177e4SLinus Torvalds * will allow a single TSO frame to consume. Building TSO frames 581da177e4SLinus Torvalds * which are too large can cause TCP streams to be bursty. 591da177e4SLinus Torvalds */ 60ab32ea5dSBrian Haley int sysctl_tcp_tso_win_divisor __read_mostly = 3; 611da177e4SLinus Torvalds 6235089bb2SDavid S. Miller /* By default, RFC2861 behavior. */ 63ab32ea5dSBrian Haley int sysctl_tcp_slow_start_after_idle __read_mostly = 1; 6435089bb2SDavid S. Miller 6546d3ceabSEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 6646d3ceabSEric Dumazet int push_one, gfp_t gfp); 67519855c5SWilliam Allen Simpson 6867edfef7SAndi Kleen /* Account for new data that has been sent to the network. */ 69*75c119afSEric Dumazet static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb) 706ff03ac3SIlpo Järvinen { 716ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 726ff03ac3SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 7366f5fe62SIlpo Järvinen unsigned int prior_packets = tp->packets_out; 749e412ba7SIlpo Järvinen 751da177e4SLinus Torvalds tp->snd_nxt = TCP_SKB_CB(skb)->end_seq; 768512430eSIlpo Järvinen 77*75c119afSEric Dumazet __skb_unlink(skb, &sk->sk_write_queue); 78*75c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, skb); 79*75c119afSEric Dumazet 8066f5fe62SIlpo Järvinen tp->packets_out += tcp_skb_pcount(skb); 81bec41a11SYuchung Cheng if (!prior_packets || icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) 82750ea2baSYuchung Cheng tcp_rearm_rto(sk); 83f19c29e3SYuchung Cheng 84f7324acdSDavid S. Miller NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT, 85f19c29e3SYuchung Cheng tcp_skb_pcount(skb)); 866a5dc9e5SEric Dumazet } 871da177e4SLinus Torvalds 88a4ecb15aSCui, Cheng /* SND.NXT, if window was not shrunk or the amount of shrunk was less than one 89a4ecb15aSCui, Cheng * window scaling factor due to loss of precision. 901da177e4SLinus Torvalds * If window has been shrunk, what should we make? It is not clear at all. 911da177e4SLinus Torvalds * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-( 921da177e4SLinus Torvalds * Anything in between SND.UNA...SND.UNA+SND.WND also can be already 931da177e4SLinus Torvalds * invalid. OK, let's make this for now: 941da177e4SLinus Torvalds */ 95cf533ea5SEric Dumazet static inline __u32 tcp_acceptable_seq(const struct sock *sk) 961da177e4SLinus Torvalds { 97cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 989e412ba7SIlpo Järvinen 99a4ecb15aSCui, Cheng if (!before(tcp_wnd_end(tp), tp->snd_nxt) || 100a4ecb15aSCui, Cheng (tp->rx_opt.wscale_ok && 101a4ecb15aSCui, Cheng ((tp->snd_nxt - tcp_wnd_end(tp)) < (1 << tp->rx_opt.rcv_wscale)))) 1021da177e4SLinus Torvalds return tp->snd_nxt; 1031da177e4SLinus Torvalds else 10490840defSIlpo Järvinen return tcp_wnd_end(tp); 1051da177e4SLinus Torvalds } 1061da177e4SLinus Torvalds 1071da177e4SLinus Torvalds /* Calculate mss to advertise in SYN segment. 1081da177e4SLinus Torvalds * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that: 1091da177e4SLinus Torvalds * 1101da177e4SLinus Torvalds * 1. It is independent of path mtu. 1111da177e4SLinus Torvalds * 2. Ideally, it is maximal possible segment size i.e. 65535-40. 1121da177e4SLinus Torvalds * 3. For IPv4 it is reasonable to calculate it from maximal MTU of 1131da177e4SLinus Torvalds * attached devices, because some buggy hosts are confused by 1141da177e4SLinus Torvalds * large MSS. 1151da177e4SLinus Torvalds * 4. We do not make 3, we advertise MSS, calculated from first 1161da177e4SLinus Torvalds * hop device mtu, but allow to raise it to ip_rt_min_advmss. 1171da177e4SLinus Torvalds * This may be overridden via information stored in routing table. 1181da177e4SLinus Torvalds * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible, 1191da177e4SLinus Torvalds * probably even Jumbo". 1201da177e4SLinus Torvalds */ 1211da177e4SLinus Torvalds static __u16 tcp_advertise_mss(struct sock *sk) 1221da177e4SLinus Torvalds { 1231da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 124cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1251da177e4SLinus Torvalds int mss = tp->advmss; 1261da177e4SLinus Torvalds 1270dbaee3bSDavid S. Miller if (dst) { 1280dbaee3bSDavid S. Miller unsigned int metric = dst_metric_advmss(dst); 1290dbaee3bSDavid S. Miller 1300dbaee3bSDavid S. Miller if (metric < mss) { 1310dbaee3bSDavid S. Miller mss = metric; 1321da177e4SLinus Torvalds tp->advmss = mss; 1331da177e4SLinus Torvalds } 1340dbaee3bSDavid S. Miller } 1351da177e4SLinus Torvalds 1361da177e4SLinus Torvalds return (__u16)mss; 1371da177e4SLinus Torvalds } 1381da177e4SLinus Torvalds 1391da177e4SLinus Torvalds /* RFC2861. Reset CWND after idle period longer RTO to "restart window". 1406f021c62SEric Dumazet * This is the first part of cwnd validation mechanism. 1416f021c62SEric Dumazet */ 1426f021c62SEric Dumazet void tcp_cwnd_restart(struct sock *sk, s32 delta) 1431da177e4SLinus Torvalds { 144463c84b9SArnaldo Carvalho de Melo struct tcp_sock *tp = tcp_sk(sk); 1456f021c62SEric Dumazet u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk)); 1461da177e4SLinus Torvalds u32 cwnd = tp->snd_cwnd; 1471da177e4SLinus Torvalds 1486687e988SArnaldo Carvalho de Melo tcp_ca_event(sk, CA_EVENT_CWND_RESTART); 1491da177e4SLinus Torvalds 1506687e988SArnaldo Carvalho de Melo tp->snd_ssthresh = tcp_current_ssthresh(sk); 1511da177e4SLinus Torvalds restart_cwnd = min(restart_cwnd, cwnd); 1521da177e4SLinus Torvalds 153463c84b9SArnaldo Carvalho de Melo while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd) 1541da177e4SLinus Torvalds cwnd >>= 1; 1551da177e4SLinus Torvalds tp->snd_cwnd = max(cwnd, restart_cwnd); 156c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 1571da177e4SLinus Torvalds tp->snd_cwnd_used = 0; 1581da177e4SLinus Torvalds } 1591da177e4SLinus Torvalds 16067edfef7SAndi Kleen /* Congestion state accounting after a packet has been sent. */ 16140efc6faSStephen Hemminger static void tcp_event_data_sent(struct tcp_sock *tp, 162cf533ea5SEric Dumazet struct sock *sk) 1631da177e4SLinus Torvalds { 164463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 165d635fbe2SEric Dumazet const u32 now = tcp_jiffies32; 1661da177e4SLinus Torvalds 16705c5a46dSNeal Cardwell if (tcp_packets_in_flight(tp) == 0) 16805c5a46dSNeal Cardwell tcp_ca_event(sk, CA_EVENT_TX_START); 16905c5a46dSNeal Cardwell 1701da177e4SLinus Torvalds tp->lsndtime = now; 1711da177e4SLinus Torvalds 1721da177e4SLinus Torvalds /* If it is a reply for ato after last received 1731da177e4SLinus Torvalds * packet, enter pingpong mode. 1741da177e4SLinus Torvalds */ 1752251ae46SJon Maxwell if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato) 176463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pingpong = 1; 1771da177e4SLinus Torvalds } 1781da177e4SLinus Torvalds 17967edfef7SAndi Kleen /* Account for an ACK we sent. */ 18040efc6faSStephen Hemminger static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts) 1811da177e4SLinus Torvalds { 182463c84b9SArnaldo Carvalho de Melo tcp_dec_quickack_mode(sk, pkts); 183463c84b9SArnaldo Carvalho de Melo inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); 1841da177e4SLinus Torvalds } 1851da177e4SLinus Torvalds 18685f16525SYuchung Cheng 18785f16525SYuchung Cheng u32 tcp_default_init_rwnd(u32 mss) 18885f16525SYuchung Cheng { 18985f16525SYuchung Cheng /* Initial receive window should be twice of TCP_INIT_CWND to 1909ef71e0cSWeiping Pan * enable proper sending of new unsent data during fast recovery 19185f16525SYuchung Cheng * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a 19285f16525SYuchung Cheng * limit when mss is larger than 1460. 19385f16525SYuchung Cheng */ 19485f16525SYuchung Cheng u32 init_rwnd = TCP_INIT_CWND * 2; 19585f16525SYuchung Cheng 19685f16525SYuchung Cheng if (mss > 1460) 19785f16525SYuchung Cheng init_rwnd = max((1460 * init_rwnd) / mss, 2U); 19885f16525SYuchung Cheng return init_rwnd; 19985f16525SYuchung Cheng } 20085f16525SYuchung Cheng 2011da177e4SLinus Torvalds /* Determine a window scaling and initial window to offer. 2021da177e4SLinus Torvalds * Based on the assumption that the given amount of space 2031da177e4SLinus Torvalds * will be offered. Store the results in the tp structure. 2041da177e4SLinus Torvalds * NOTE: for smooth operation initial space offering should 2051da177e4SLinus Torvalds * be a multiple of mss if possible. We assume here that mss >= 1. 2061da177e4SLinus Torvalds * This MUST be enforced by all callers. 2071da177e4SLinus Torvalds */ 2081da177e4SLinus Torvalds void tcp_select_initial_window(int __space, __u32 mss, 2091da177e4SLinus Torvalds __u32 *rcv_wnd, __u32 *window_clamp, 21031d12926Slaurent chavey int wscale_ok, __u8 *rcv_wscale, 21131d12926Slaurent chavey __u32 init_rcv_wnd) 2121da177e4SLinus Torvalds { 2131da177e4SLinus Torvalds unsigned int space = (__space < 0 ? 0 : __space); 2141da177e4SLinus Torvalds 2151da177e4SLinus Torvalds /* If no clamp set the clamp to the max possible scaled window */ 2161da177e4SLinus Torvalds if (*window_clamp == 0) 217589c49cbSGao Feng (*window_clamp) = (U16_MAX << TCP_MAX_WSCALE); 2181da177e4SLinus Torvalds space = min(*window_clamp, space); 2191da177e4SLinus Torvalds 2201da177e4SLinus Torvalds /* Quantize space offering to a multiple of mss if possible. */ 2211da177e4SLinus Torvalds if (space > mss) 222589c49cbSGao Feng space = rounddown(space, mss); 2231da177e4SLinus Torvalds 2241da177e4SLinus Torvalds /* NOTE: offering an initial window larger than 32767 22515d99e02SRick Jones * will break some buggy TCP stacks. If the admin tells us 22615d99e02SRick Jones * it is likely we could be speaking with such a buggy stack 22715d99e02SRick Jones * we will truncate our initial window offering to 32K-1 22815d99e02SRick Jones * unless the remote has sent us a window scaling option, 22915d99e02SRick Jones * which we interpret as a sign the remote TCP is not 23015d99e02SRick Jones * misinterpreting the window field as a signed quantity. 2311da177e4SLinus Torvalds */ 23215d99e02SRick Jones if (sysctl_tcp_workaround_signed_windows) 2331da177e4SLinus Torvalds (*rcv_wnd) = min(space, MAX_TCP_WINDOW); 23415d99e02SRick Jones else 23515d99e02SRick Jones (*rcv_wnd) = space; 23615d99e02SRick Jones 2371da177e4SLinus Torvalds (*rcv_wscale) = 0; 2381da177e4SLinus Torvalds if (wscale_ok) { 239589c49cbSGao Feng /* Set window scaling on max possible window */ 240f626300aSSoheil Hassas Yeganeh space = max_t(u32, space, sysctl_tcp_rmem[2]); 241f626300aSSoheil Hassas Yeganeh space = max_t(u32, space, sysctl_rmem_max); 242316c1592SStephen Hemminger space = min_t(u32, space, *window_clamp); 243589c49cbSGao Feng while (space > U16_MAX && (*rcv_wscale) < TCP_MAX_WSCALE) { 2441da177e4SLinus Torvalds space >>= 1; 2451da177e4SLinus Torvalds (*rcv_wscale)++; 2461da177e4SLinus Torvalds } 2471da177e4SLinus Torvalds } 2481da177e4SLinus Torvalds 2491da177e4SLinus Torvalds if (mss > (1 << *rcv_wscale)) { 25085f16525SYuchung Cheng if (!init_rcv_wnd) /* Use default unless specified otherwise */ 25185f16525SYuchung Cheng init_rcv_wnd = tcp_default_init_rwnd(mss); 252b1afde60SNandita Dukkipati *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss); 2531da177e4SLinus Torvalds } 2541da177e4SLinus Torvalds 2551da177e4SLinus Torvalds /* Set the clamp no higher than max representable value */ 256589c49cbSGao Feng (*window_clamp) = min_t(__u32, U16_MAX << (*rcv_wscale), *window_clamp); 2571da177e4SLinus Torvalds } 2584bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_select_initial_window); 2591da177e4SLinus Torvalds 2601da177e4SLinus Torvalds /* Chose a new window to advertise, update state in tcp_sock for the 2611da177e4SLinus Torvalds * socket, and return result with RFC1323 scaling applied. The return 2621da177e4SLinus Torvalds * value can be stuffed directly into th->window for an outgoing 2631da177e4SLinus Torvalds * frame. 2641da177e4SLinus Torvalds */ 26540efc6faSStephen Hemminger static u16 tcp_select_window(struct sock *sk) 2661da177e4SLinus Torvalds { 2671da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2688e165e20SFlorian Westphal u32 old_win = tp->rcv_wnd; 2691da177e4SLinus Torvalds u32 cur_win = tcp_receive_window(tp); 2701da177e4SLinus Torvalds u32 new_win = __tcp_select_window(sk); 2711da177e4SLinus Torvalds 2721da177e4SLinus Torvalds /* Never shrink the offered window */ 2731da177e4SLinus Torvalds if (new_win < cur_win) { 2741da177e4SLinus Torvalds /* Danger Will Robinson! 2751da177e4SLinus Torvalds * Don't update rcv_wup/rcv_wnd here or else 2761da177e4SLinus Torvalds * we will not be able to advertise a zero 2771da177e4SLinus Torvalds * window in time. --DaveM 2781da177e4SLinus Torvalds * 2791da177e4SLinus Torvalds * Relax Will Robinson. 2801da177e4SLinus Torvalds */ 2818e165e20SFlorian Westphal if (new_win == 0) 2828e165e20SFlorian Westphal NET_INC_STATS(sock_net(sk), 2838e165e20SFlorian Westphal LINUX_MIB_TCPWANTZEROWINDOWADV); 284607bfbf2SPatrick McHardy new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale); 2851da177e4SLinus Torvalds } 2861da177e4SLinus Torvalds tp->rcv_wnd = new_win; 2871da177e4SLinus Torvalds tp->rcv_wup = tp->rcv_nxt; 2881da177e4SLinus Torvalds 2891da177e4SLinus Torvalds /* Make sure we do not exceed the maximum possible 2901da177e4SLinus Torvalds * scaled window. 2911da177e4SLinus Torvalds */ 29215d99e02SRick Jones if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows) 2931da177e4SLinus Torvalds new_win = min(new_win, MAX_TCP_WINDOW); 2941da177e4SLinus Torvalds else 2951da177e4SLinus Torvalds new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); 2961da177e4SLinus Torvalds 2971da177e4SLinus Torvalds /* RFC1323 scaling applied */ 2981da177e4SLinus Torvalds new_win >>= tp->rx_opt.rcv_wscale; 2991da177e4SLinus Torvalds 30031770e34SFlorian Westphal /* If we advertise zero window, disable fast path. */ 3018e165e20SFlorian Westphal if (new_win == 0) { 30231770e34SFlorian Westphal tp->pred_flags = 0; 3038e165e20SFlorian Westphal if (old_win) 3048e165e20SFlorian Westphal NET_INC_STATS(sock_net(sk), 3058e165e20SFlorian Westphal LINUX_MIB_TCPTOZEROWINDOWADV); 3068e165e20SFlorian Westphal } else if (old_win == 0) { 3078e165e20SFlorian Westphal NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV); 3088e165e20SFlorian Westphal } 3091da177e4SLinus Torvalds 3101da177e4SLinus Torvalds return new_win; 3111da177e4SLinus Torvalds } 3121da177e4SLinus Torvalds 31367edfef7SAndi Kleen /* Packet ECN state for a SYN-ACK */ 314735d3831SFlorian Westphal static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb) 315bdf1ee5dSIlpo Järvinen { 31630e502a3SDaniel Borkmann const struct tcp_sock *tp = tcp_sk(sk); 31730e502a3SDaniel Borkmann 3184de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR; 319bdf1ee5dSIlpo Järvinen if (!(tp->ecn_flags & TCP_ECN_OK)) 3204de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE; 32191b5b21cSLawrence Brakmo else if (tcp_ca_needs_ecn(sk) || 32291b5b21cSLawrence Brakmo tcp_bpf_ca_needs_ecn(sk)) 32330e502a3SDaniel Borkmann INET_ECN_xmit(sk); 324bdf1ee5dSIlpo Järvinen } 325bdf1ee5dSIlpo Järvinen 32667edfef7SAndi Kleen /* Packet ECN state for a SYN. */ 327735d3831SFlorian Westphal static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb) 328bdf1ee5dSIlpo Järvinen { 329bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 33091b5b21cSLawrence Brakmo bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk); 331f7b3bec6SFlorian Westphal bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 || 33291b5b21cSLawrence Brakmo tcp_ca_needs_ecn(sk) || bpf_needs_ecn; 333f7b3bec6SFlorian Westphal 334f7b3bec6SFlorian Westphal if (!use_ecn) { 335f7b3bec6SFlorian Westphal const struct dst_entry *dst = __sk_dst_get(sk); 336f7b3bec6SFlorian Westphal 337f7b3bec6SFlorian Westphal if (dst && dst_feature(dst, RTAX_FEATURE_ECN)) 338f7b3bec6SFlorian Westphal use_ecn = true; 339f7b3bec6SFlorian Westphal } 340bdf1ee5dSIlpo Järvinen 341bdf1ee5dSIlpo Järvinen tp->ecn_flags = 0; 342f7b3bec6SFlorian Westphal 343f7b3bec6SFlorian Westphal if (use_ecn) { 3444de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR; 345bdf1ee5dSIlpo Järvinen tp->ecn_flags = TCP_ECN_OK; 34691b5b21cSLawrence Brakmo if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn) 34730e502a3SDaniel Borkmann INET_ECN_xmit(sk); 348bdf1ee5dSIlpo Järvinen } 349bdf1ee5dSIlpo Järvinen } 350bdf1ee5dSIlpo Järvinen 35149213555SDaniel Borkmann static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb) 35249213555SDaniel Borkmann { 35349213555SDaniel Borkmann if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback) 35449213555SDaniel Borkmann /* tp->ecn_flags are cleared at a later point in time when 35549213555SDaniel Borkmann * SYN ACK is ultimatively being received. 35649213555SDaniel Borkmann */ 35749213555SDaniel Borkmann TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR); 35849213555SDaniel Borkmann } 35949213555SDaniel Borkmann 360735d3831SFlorian Westphal static void 3616ac705b1SEric Dumazet tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th) 362bdf1ee5dSIlpo Järvinen { 3636ac705b1SEric Dumazet if (inet_rsk(req)->ecn_ok) 364bdf1ee5dSIlpo Järvinen th->ece = 1; 365bdf1ee5dSIlpo Järvinen } 366bdf1ee5dSIlpo Järvinen 36767edfef7SAndi Kleen /* Set up ECN state for a packet on a ESTABLISHED socket that is about to 36867edfef7SAndi Kleen * be sent. 36967edfef7SAndi Kleen */ 370735d3831SFlorian Westphal static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb, 371ea1627c2SEric Dumazet struct tcphdr *th, int tcp_header_len) 372bdf1ee5dSIlpo Järvinen { 373bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 374bdf1ee5dSIlpo Järvinen 375bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_OK) { 376bdf1ee5dSIlpo Järvinen /* Not-retransmitted data segment: set ECT and inject CWR. */ 377bdf1ee5dSIlpo Järvinen if (skb->len != tcp_header_len && 378bdf1ee5dSIlpo Järvinen !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { 379bdf1ee5dSIlpo Järvinen INET_ECN_xmit(sk); 380bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) { 381bdf1ee5dSIlpo Järvinen tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; 382ea1627c2SEric Dumazet th->cwr = 1; 383bdf1ee5dSIlpo Järvinen skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; 384bdf1ee5dSIlpo Järvinen } 38530e502a3SDaniel Borkmann } else if (!tcp_ca_needs_ecn(sk)) { 386bdf1ee5dSIlpo Järvinen /* ACK or retransmitted segment: clear ECT|CE */ 387bdf1ee5dSIlpo Järvinen INET_ECN_dontxmit(sk); 388bdf1ee5dSIlpo Järvinen } 389bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) 390ea1627c2SEric Dumazet th->ece = 1; 391bdf1ee5dSIlpo Järvinen } 392bdf1ee5dSIlpo Järvinen } 393bdf1ee5dSIlpo Järvinen 394e870a8efSIlpo Järvinen /* Constructs common control bits of non-data skb. If SYN/FIN is present, 395e870a8efSIlpo Järvinen * auto increment end seqno. 396e870a8efSIlpo Järvinen */ 397e870a8efSIlpo Järvinen static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags) 398e870a8efSIlpo Järvinen { 3992e8e18efSDavid S. Miller skb->ip_summed = CHECKSUM_PARTIAL; 400e870a8efSIlpo Järvinen skb->csum = 0; 401e870a8efSIlpo Järvinen 4024de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags; 403e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->sacked = 0; 404e870a8efSIlpo Järvinen 405cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, 1); 406e870a8efSIlpo Järvinen 407e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->seq = seq; 408a3433f35SChangli Gao if (flags & (TCPHDR_SYN | TCPHDR_FIN)) 409e870a8efSIlpo Järvinen seq++; 410e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->end_seq = seq; 411e870a8efSIlpo Järvinen } 412e870a8efSIlpo Järvinen 413a2a385d6SEric Dumazet static inline bool tcp_urg_mode(const struct tcp_sock *tp) 41433f5f57eSIlpo Järvinen { 41533f5f57eSIlpo Järvinen return tp->snd_una != tp->snd_up; 41633f5f57eSIlpo Järvinen } 41733f5f57eSIlpo Järvinen 41833ad798cSAdam Langley #define OPTION_SACK_ADVERTISE (1 << 0) 41933ad798cSAdam Langley #define OPTION_TS (1 << 1) 42033ad798cSAdam Langley #define OPTION_MD5 (1 << 2) 42189e95a61SOri Finkelman #define OPTION_WSCALE (1 << 3) 4222100c8d2SYuchung Cheng #define OPTION_FAST_OPEN_COOKIE (1 << 8) 42333ad798cSAdam Langley 42433ad798cSAdam Langley struct tcp_out_options { 4252100c8d2SYuchung Cheng u16 options; /* bit field of OPTION_* */ 4262100c8d2SYuchung Cheng u16 mss; /* 0 to disable */ 42733ad798cSAdam Langley u8 ws; /* window scale, 0 to disable */ 42833ad798cSAdam Langley u8 num_sack_blocks; /* number of SACK blocks to include */ 429bd0388aeSWilliam Allen Simpson u8 hash_size; /* bytes in hash_location */ 430bd0388aeSWilliam Allen Simpson __u8 *hash_location; /* temporary pointer, overloaded */ 4312100c8d2SYuchung Cheng __u32 tsval, tsecr; /* need to include OPTION_TS */ 4322100c8d2SYuchung Cheng struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */ 43333ad798cSAdam Langley }; 43433ad798cSAdam Langley 43567edfef7SAndi Kleen /* Write previously computed TCP options to the packet. 43667edfef7SAndi Kleen * 43767edfef7SAndi Kleen * Beware: Something in the Internet is very sensitive to the ordering of 438fd6149d3SIlpo Järvinen * TCP options, we learned this through the hard way, so be careful here. 439fd6149d3SIlpo Järvinen * Luckily we can at least blame others for their non-compliance but from 4408e3bff96Sstephen hemminger * inter-operability perspective it seems that we're somewhat stuck with 441fd6149d3SIlpo Järvinen * the ordering which we have been using if we want to keep working with 442fd6149d3SIlpo Järvinen * those broken things (not that it currently hurts anybody as there isn't 443fd6149d3SIlpo Järvinen * particular reason why the ordering would need to be changed). 444fd6149d3SIlpo Järvinen * 445fd6149d3SIlpo Järvinen * At least SACK_PERM as the first option is known to lead to a disaster 446fd6149d3SIlpo Järvinen * (but it may well be that other scenarios fail similarly). 447fd6149d3SIlpo Järvinen */ 44833ad798cSAdam Langley static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp, 449bd0388aeSWilliam Allen Simpson struct tcp_out_options *opts) 450bd0388aeSWilliam Allen Simpson { 4512100c8d2SYuchung Cheng u16 options = opts->options; /* mungable copy */ 452bd0388aeSWilliam Allen Simpson 453bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_MD5 & options)) { 4541a2c6181SChristoph Paasch *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | 4551a2c6181SChristoph Paasch (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG); 456bd0388aeSWilliam Allen Simpson /* overload cookie hash location */ 457bd0388aeSWilliam Allen Simpson opts->hash_location = (__u8 *)ptr; 45833ad798cSAdam Langley ptr += 4; 45933ad798cSAdam Langley } 46033ad798cSAdam Langley 461fd6149d3SIlpo Järvinen if (unlikely(opts->mss)) { 462fd6149d3SIlpo Järvinen *ptr++ = htonl((TCPOPT_MSS << 24) | 463fd6149d3SIlpo Järvinen (TCPOLEN_MSS << 16) | 464fd6149d3SIlpo Järvinen opts->mss); 465fd6149d3SIlpo Järvinen } 466fd6149d3SIlpo Järvinen 467bd0388aeSWilliam Allen Simpson if (likely(OPTION_TS & options)) { 468bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 46933ad798cSAdam Langley *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | 47033ad798cSAdam Langley (TCPOLEN_SACK_PERM << 16) | 47133ad798cSAdam Langley (TCPOPT_TIMESTAMP << 8) | 47233ad798cSAdam Langley TCPOLEN_TIMESTAMP); 473bd0388aeSWilliam Allen Simpson options &= ~OPTION_SACK_ADVERTISE; 47433ad798cSAdam Langley } else { 475496c98dfSYOSHIFUJI Hideaki *ptr++ = htonl((TCPOPT_NOP << 24) | 47640efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 47740efc6faSStephen Hemminger (TCPOPT_TIMESTAMP << 8) | 47840efc6faSStephen Hemminger TCPOLEN_TIMESTAMP); 47940efc6faSStephen Hemminger } 48033ad798cSAdam Langley *ptr++ = htonl(opts->tsval); 48133ad798cSAdam Langley *ptr++ = htonl(opts->tsecr); 48233ad798cSAdam Langley } 48333ad798cSAdam Langley 484bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 48533ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 48633ad798cSAdam Langley (TCPOPT_NOP << 16) | 48733ad798cSAdam Langley (TCPOPT_SACK_PERM << 8) | 48833ad798cSAdam Langley TCPOLEN_SACK_PERM); 48933ad798cSAdam Langley } 49033ad798cSAdam Langley 491bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_WSCALE & options)) { 49233ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 49333ad798cSAdam Langley (TCPOPT_WINDOW << 16) | 49433ad798cSAdam Langley (TCPOLEN_WINDOW << 8) | 49533ad798cSAdam Langley opts->ws); 49633ad798cSAdam Langley } 49733ad798cSAdam Langley 49833ad798cSAdam Langley if (unlikely(opts->num_sack_blocks)) { 49933ad798cSAdam Langley struct tcp_sack_block *sp = tp->rx_opt.dsack ? 50033ad798cSAdam Langley tp->duplicate_sack : tp->selective_acks; 50140efc6faSStephen Hemminger int this_sack; 50240efc6faSStephen Hemminger 50340efc6faSStephen Hemminger *ptr++ = htonl((TCPOPT_NOP << 24) | 50440efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 50540efc6faSStephen Hemminger (TCPOPT_SACK << 8) | 50633ad798cSAdam Langley (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * 50740efc6faSStephen Hemminger TCPOLEN_SACK_PERBLOCK))); 5082de979bdSStephen Hemminger 50933ad798cSAdam Langley for (this_sack = 0; this_sack < opts->num_sack_blocks; 51033ad798cSAdam Langley ++this_sack) { 51140efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].start_seq); 51240efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].end_seq); 51340efc6faSStephen Hemminger } 5142de979bdSStephen Hemminger 51540efc6faSStephen Hemminger tp->rx_opt.dsack = 0; 51640efc6faSStephen Hemminger } 5172100c8d2SYuchung Cheng 5182100c8d2SYuchung Cheng if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) { 5192100c8d2SYuchung Cheng struct tcp_fastopen_cookie *foc = opts->fastopen_cookie; 5207f9b838bSDaniel Lee u8 *p = (u8 *)ptr; 5217f9b838bSDaniel Lee u32 len; /* Fast Open option length */ 5222100c8d2SYuchung Cheng 5237f9b838bSDaniel Lee if (foc->exp) { 5247f9b838bSDaniel Lee len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len; 5257f9b838bSDaniel Lee *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) | 5262100c8d2SYuchung Cheng TCPOPT_FASTOPEN_MAGIC); 5277f9b838bSDaniel Lee p += TCPOLEN_EXP_FASTOPEN_BASE; 5287f9b838bSDaniel Lee } else { 5297f9b838bSDaniel Lee len = TCPOLEN_FASTOPEN_BASE + foc->len; 5307f9b838bSDaniel Lee *p++ = TCPOPT_FASTOPEN; 5317f9b838bSDaniel Lee *p++ = len; 5322100c8d2SYuchung Cheng } 5337f9b838bSDaniel Lee 5347f9b838bSDaniel Lee memcpy(p, foc->val, foc->len); 5357f9b838bSDaniel Lee if ((len & 3) == 2) { 5367f9b838bSDaniel Lee p[foc->len] = TCPOPT_NOP; 5377f9b838bSDaniel Lee p[foc->len + 1] = TCPOPT_NOP; 5387f9b838bSDaniel Lee } 5397f9b838bSDaniel Lee ptr += (len + 3) >> 2; 5402100c8d2SYuchung Cheng } 54140efc6faSStephen Hemminger } 54240efc6faSStephen Hemminger 54367edfef7SAndi Kleen /* Compute TCP options for SYN packets. This is not the final 54467edfef7SAndi Kleen * network wire format yet. 54567edfef7SAndi Kleen */ 54695c96174SEric Dumazet static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb, 54733ad798cSAdam Langley struct tcp_out_options *opts, 548cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 549cf533ea5SEric Dumazet { 55033ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 55195c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 552783237e8SYuchung Cheng struct tcp_fastopen_request *fastopen = tp->fastopen_req; 55333ad798cSAdam Langley 554cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 55533ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 55633ad798cSAdam Langley if (*md5) { 55733ad798cSAdam Langley opts->options |= OPTION_MD5; 558bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 559cfb6eeb4SYOSHIFUJI Hideaki } 56033ad798cSAdam Langley #else 56133ad798cSAdam Langley *md5 = NULL; 562cfb6eeb4SYOSHIFUJI Hideaki #endif 56333ad798cSAdam Langley 56433ad798cSAdam Langley /* We always get an MSS option. The option bytes which will be seen in 56533ad798cSAdam Langley * normal data packets should timestamps be used, must be in the MSS 56633ad798cSAdam Langley * advertised. But we subtract them from tp->mss_cache so that 56733ad798cSAdam Langley * calculations in tcp_sendmsg are simpler etc. So account for this 56833ad798cSAdam Langley * fact here if necessary. If we don't do this correctly, as a 56933ad798cSAdam Langley * receiver we won't recognize data packets as being full sized when we 57033ad798cSAdam Langley * should, and thus we won't abide by the delayed ACK rules correctly. 57133ad798cSAdam Langley * SACKs don't matter, we never delay an ACK when we have any of those 57233ad798cSAdam Langley * going out. */ 57333ad798cSAdam Langley opts->mss = tcp_advertise_mss(sk); 574bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 57533ad798cSAdam Langley 5765d2ed052SEric Dumazet if (likely(sock_net(sk)->ipv4.sysctl_tcp_timestamps && !*md5)) { 57733ad798cSAdam Langley opts->options |= OPTION_TS; 5787faee5c0SEric Dumazet opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset; 57933ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 580bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 58133ad798cSAdam Langley } 5829bb37ef0SEric Dumazet if (likely(sock_net(sk)->ipv4.sysctl_tcp_window_scaling)) { 58333ad798cSAdam Langley opts->ws = tp->rx_opt.rcv_wscale; 58489e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 585bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 58633ad798cSAdam Langley } 587f9301034SEric Dumazet if (likely(sock_net(sk)->ipv4.sysctl_tcp_sack)) { 58833ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 589b32d1310SDavid S. Miller if (unlikely(!(OPTION_TS & opts->options))) 590bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 59133ad798cSAdam Langley } 59233ad798cSAdam Langley 593783237e8SYuchung Cheng if (fastopen && fastopen->cookie.len >= 0) { 5942646c831SDaniel Lee u32 need = fastopen->cookie.len; 5952646c831SDaniel Lee 5962646c831SDaniel Lee need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE : 5972646c831SDaniel Lee TCPOLEN_FASTOPEN_BASE; 598783237e8SYuchung Cheng need = (need + 3) & ~3U; /* Align to 32 bits */ 599783237e8SYuchung Cheng if (remaining >= need) { 600783237e8SYuchung Cheng opts->options |= OPTION_FAST_OPEN_COOKIE; 601783237e8SYuchung Cheng opts->fastopen_cookie = &fastopen->cookie; 602783237e8SYuchung Cheng remaining -= need; 603783237e8SYuchung Cheng tp->syn_fastopen = 1; 6042646c831SDaniel Lee tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0; 605783237e8SYuchung Cheng } 606783237e8SYuchung Cheng } 607bd0388aeSWilliam Allen Simpson 608bd0388aeSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 60933ad798cSAdam Langley } 61033ad798cSAdam Langley 61167edfef7SAndi Kleen /* Set up TCP options for SYN-ACKs. */ 61237bfbddaSEric Dumazet static unsigned int tcp_synack_options(struct request_sock *req, 61395c96174SEric Dumazet unsigned int mss, struct sk_buff *skb, 61433ad798cSAdam Langley struct tcp_out_options *opts, 61580f03e27SEric Dumazet const struct tcp_md5sig_key *md5, 6168336886fSJerry Chu struct tcp_fastopen_cookie *foc) 6174957faadSWilliam Allen Simpson { 61833ad798cSAdam Langley struct inet_request_sock *ireq = inet_rsk(req); 61995c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 62033ad798cSAdam Langley 62133ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 62280f03e27SEric Dumazet if (md5) { 62333ad798cSAdam Langley opts->options |= OPTION_MD5; 6244957faadSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 6254957faadSWilliam Allen Simpson 6264957faadSWilliam Allen Simpson /* We can't fit any SACK blocks in a packet with MD5 + TS 6274957faadSWilliam Allen Simpson * options. There was discussion about disabling SACK 6284957faadSWilliam Allen Simpson * rather than TS in order to fit in better with old, 6294957faadSWilliam Allen Simpson * buggy kernels, but that was deemed to be unnecessary. 6304957faadSWilliam Allen Simpson */ 631de213e5eSEric Dumazet ireq->tstamp_ok &= !ireq->sack_ok; 63233ad798cSAdam Langley } 63333ad798cSAdam Langley #endif 63433ad798cSAdam Langley 6354957faadSWilliam Allen Simpson /* We always send an MSS option. */ 63633ad798cSAdam Langley opts->mss = mss; 6374957faadSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 63833ad798cSAdam Langley 63933ad798cSAdam Langley if (likely(ireq->wscale_ok)) { 64033ad798cSAdam Langley opts->ws = ireq->rcv_wscale; 64189e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 6424957faadSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 64333ad798cSAdam Langley } 644de213e5eSEric Dumazet if (likely(ireq->tstamp_ok)) { 64533ad798cSAdam Langley opts->options |= OPTION_TS; 64695a22caeSFlorian Westphal opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off; 64733ad798cSAdam Langley opts->tsecr = req->ts_recent; 6484957faadSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 64933ad798cSAdam Langley } 65033ad798cSAdam Langley if (likely(ireq->sack_ok)) { 65133ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 652de213e5eSEric Dumazet if (unlikely(!ireq->tstamp_ok)) 6534957faadSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 65433ad798cSAdam Langley } 6557f9b838bSDaniel Lee if (foc != NULL && foc->len >= 0) { 6567f9b838bSDaniel Lee u32 need = foc->len; 6577f9b838bSDaniel Lee 6587f9b838bSDaniel Lee need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE : 6597f9b838bSDaniel Lee TCPOLEN_FASTOPEN_BASE; 6608336886fSJerry Chu need = (need + 3) & ~3U; /* Align to 32 bits */ 6618336886fSJerry Chu if (remaining >= need) { 6628336886fSJerry Chu opts->options |= OPTION_FAST_OPEN_COOKIE; 6638336886fSJerry Chu opts->fastopen_cookie = foc; 6648336886fSJerry Chu remaining -= need; 6658336886fSJerry Chu } 6668336886fSJerry Chu } 6674957faadSWilliam Allen Simpson 6684957faadSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 66933ad798cSAdam Langley } 67033ad798cSAdam Langley 67167edfef7SAndi Kleen /* Compute TCP options for ESTABLISHED sockets. This is not the 67267edfef7SAndi Kleen * final wire format yet. 67367edfef7SAndi Kleen */ 67495c96174SEric Dumazet static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb, 67533ad798cSAdam Langley struct tcp_out_options *opts, 676cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 677cf533ea5SEric Dumazet { 67833ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 67995c96174SEric Dumazet unsigned int size = 0; 680cabeccbdSIlpo Järvinen unsigned int eff_sacks; 68133ad798cSAdam Langley 6825843ef42SAndi Kleen opts->options = 0; 6835843ef42SAndi Kleen 68433ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 68533ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 68633ad798cSAdam Langley if (unlikely(*md5)) { 68733ad798cSAdam Langley opts->options |= OPTION_MD5; 68833ad798cSAdam Langley size += TCPOLEN_MD5SIG_ALIGNED; 68933ad798cSAdam Langley } 69033ad798cSAdam Langley #else 69133ad798cSAdam Langley *md5 = NULL; 69233ad798cSAdam Langley #endif 69333ad798cSAdam Langley 69433ad798cSAdam Langley if (likely(tp->rx_opt.tstamp_ok)) { 69533ad798cSAdam Langley opts->options |= OPTION_TS; 6967faee5c0SEric Dumazet opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0; 69733ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 69833ad798cSAdam Langley size += TCPOLEN_TSTAMP_ALIGNED; 69933ad798cSAdam Langley } 70033ad798cSAdam Langley 701cabeccbdSIlpo Järvinen eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; 702cabeccbdSIlpo Järvinen if (unlikely(eff_sacks)) { 70395c96174SEric Dumazet const unsigned int remaining = MAX_TCP_OPTION_SPACE - size; 70433ad798cSAdam Langley opts->num_sack_blocks = 70595c96174SEric Dumazet min_t(unsigned int, eff_sacks, 70633ad798cSAdam Langley (remaining - TCPOLEN_SACK_BASE_ALIGNED) / 70733ad798cSAdam Langley TCPOLEN_SACK_PERBLOCK); 70833ad798cSAdam Langley size += TCPOLEN_SACK_BASE_ALIGNED + 70933ad798cSAdam Langley opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; 71033ad798cSAdam Langley } 71133ad798cSAdam Langley 71233ad798cSAdam Langley return size; 71340efc6faSStephen Hemminger } 7141da177e4SLinus Torvalds 71546d3ceabSEric Dumazet 71646d3ceabSEric Dumazet /* TCP SMALL QUEUES (TSQ) 71746d3ceabSEric Dumazet * 71846d3ceabSEric Dumazet * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev) 71946d3ceabSEric Dumazet * to reduce RTT and bufferbloat. 72046d3ceabSEric Dumazet * We do this using a special skb destructor (tcp_wfree). 72146d3ceabSEric Dumazet * 72246d3ceabSEric Dumazet * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb 72346d3ceabSEric Dumazet * needs to be reallocated in a driver. 7248e3bff96Sstephen hemminger * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc 72546d3ceabSEric Dumazet * 72646d3ceabSEric Dumazet * Since transmit from skb destructor is forbidden, we use a tasklet 72746d3ceabSEric Dumazet * to process all sockets that eventually need to send more skbs. 72846d3ceabSEric Dumazet * We use one tasklet per cpu, with its own queue of sockets. 72946d3ceabSEric Dumazet */ 73046d3ceabSEric Dumazet struct tsq_tasklet { 73146d3ceabSEric Dumazet struct tasklet_struct tasklet; 73246d3ceabSEric Dumazet struct list_head head; /* queue of tcp sockets */ 73346d3ceabSEric Dumazet }; 73446d3ceabSEric Dumazet static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet); 73546d3ceabSEric Dumazet 7366f458dfbSEric Dumazet static void tcp_tsq_handler(struct sock *sk) 7376f458dfbSEric Dumazet { 7386f458dfbSEric Dumazet if ((1 << sk->sk_state) & 7396f458dfbSEric Dumazet (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING | 740f9616c35SEric Dumazet TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) { 741f9616c35SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 742f9616c35SEric Dumazet 743f9616c35SEric Dumazet if (tp->lost_out > tp->retrans_out && 744f9616c35SEric Dumazet tp->snd_cwnd > tcp_packets_in_flight(tp)) 745f9616c35SEric Dumazet tcp_xmit_retransmit_queue(sk); 746f9616c35SEric Dumazet 747f9616c35SEric Dumazet tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle, 748bf06200eSJohn Ogness 0, GFP_ATOMIC); 7496f458dfbSEric Dumazet } 750f9616c35SEric Dumazet } 75146d3ceabSEric Dumazet /* 7528e3bff96Sstephen hemminger * One tasklet per cpu tries to send more skbs. 75346d3ceabSEric Dumazet * We run in tasklet context but need to disable irqs when 7548e3bff96Sstephen hemminger * transferring tsq->head because tcp_wfree() might 75546d3ceabSEric Dumazet * interrupt us (non NAPI drivers) 75646d3ceabSEric Dumazet */ 75746d3ceabSEric Dumazet static void tcp_tasklet_func(unsigned long data) 75846d3ceabSEric Dumazet { 75946d3ceabSEric Dumazet struct tsq_tasklet *tsq = (struct tsq_tasklet *)data; 76046d3ceabSEric Dumazet LIST_HEAD(list); 76146d3ceabSEric Dumazet unsigned long flags; 76246d3ceabSEric Dumazet struct list_head *q, *n; 76346d3ceabSEric Dumazet struct tcp_sock *tp; 76446d3ceabSEric Dumazet struct sock *sk; 76546d3ceabSEric Dumazet 76646d3ceabSEric Dumazet local_irq_save(flags); 76746d3ceabSEric Dumazet list_splice_init(&tsq->head, &list); 76846d3ceabSEric Dumazet local_irq_restore(flags); 76946d3ceabSEric Dumazet 77046d3ceabSEric Dumazet list_for_each_safe(q, n, &list) { 77146d3ceabSEric Dumazet tp = list_entry(q, struct tcp_sock, tsq_node); 77246d3ceabSEric Dumazet list_del(&tp->tsq_node); 77346d3ceabSEric Dumazet 77446d3ceabSEric Dumazet sk = (struct sock *)tp; 7750a9648f1SEric Dumazet smp_mb__before_atomic(); 7767aa5470cSEric Dumazet clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags); 7777aa5470cSEric Dumazet 778b223feb9SEric Dumazet if (!sk->sk_lock.owned && 7797aa5470cSEric Dumazet test_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags)) { 78046d3ceabSEric Dumazet bh_lock_sock(sk); 78146d3ceabSEric Dumazet if (!sock_owned_by_user(sk)) { 7827aa5470cSEric Dumazet clear_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags); 7836f458dfbSEric Dumazet tcp_tsq_handler(sk); 78446d3ceabSEric Dumazet } 78546d3ceabSEric Dumazet bh_unlock_sock(sk); 786b223feb9SEric Dumazet } 78746d3ceabSEric Dumazet 78846d3ceabSEric Dumazet sk_free(sk); 78946d3ceabSEric Dumazet } 79046d3ceabSEric Dumazet } 79146d3ceabSEric Dumazet 79240fc3423SEric Dumazet #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \ 79340fc3423SEric Dumazet TCPF_WRITE_TIMER_DEFERRED | \ 79440fc3423SEric Dumazet TCPF_DELACK_TIMER_DEFERRED | \ 79540fc3423SEric Dumazet TCPF_MTU_REDUCED_DEFERRED) 79646d3ceabSEric Dumazet /** 79746d3ceabSEric Dumazet * tcp_release_cb - tcp release_sock() callback 79846d3ceabSEric Dumazet * @sk: socket 79946d3ceabSEric Dumazet * 80046d3ceabSEric Dumazet * called from release_sock() to perform protocol dependent 80146d3ceabSEric Dumazet * actions before socket release. 80246d3ceabSEric Dumazet */ 80346d3ceabSEric Dumazet void tcp_release_cb(struct sock *sk) 80446d3ceabSEric Dumazet { 8056f458dfbSEric Dumazet unsigned long flags, nflags; 80646d3ceabSEric Dumazet 8076f458dfbSEric Dumazet /* perform an atomic operation only if at least one flag is set */ 8086f458dfbSEric Dumazet do { 8097aa5470cSEric Dumazet flags = sk->sk_tsq_flags; 8106f458dfbSEric Dumazet if (!(flags & TCP_DEFERRED_ALL)) 8116f458dfbSEric Dumazet return; 8126f458dfbSEric Dumazet nflags = flags & ~TCP_DEFERRED_ALL; 8137aa5470cSEric Dumazet } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags); 8146f458dfbSEric Dumazet 81540fc3423SEric Dumazet if (flags & TCPF_TSQ_DEFERRED) 8166f458dfbSEric Dumazet tcp_tsq_handler(sk); 8176f458dfbSEric Dumazet 818c3f9b018SEric Dumazet /* Here begins the tricky part : 819c3f9b018SEric Dumazet * We are called from release_sock() with : 820c3f9b018SEric Dumazet * 1) BH disabled 821c3f9b018SEric Dumazet * 2) sk_lock.slock spinlock held 822c3f9b018SEric Dumazet * 3) socket owned by us (sk->sk_lock.owned == 1) 823c3f9b018SEric Dumazet * 824c3f9b018SEric Dumazet * But following code is meant to be called from BH handlers, 825c3f9b018SEric Dumazet * so we should keep BH disabled, but early release socket ownership 826c3f9b018SEric Dumazet */ 827c3f9b018SEric Dumazet sock_release_ownership(sk); 828c3f9b018SEric Dumazet 82940fc3423SEric Dumazet if (flags & TCPF_WRITE_TIMER_DEFERRED) { 8306f458dfbSEric Dumazet tcp_write_timer_handler(sk); 831144d56e9SEric Dumazet __sock_put(sk); 832144d56e9SEric Dumazet } 83340fc3423SEric Dumazet if (flags & TCPF_DELACK_TIMER_DEFERRED) { 8346f458dfbSEric Dumazet tcp_delack_timer_handler(sk); 835144d56e9SEric Dumazet __sock_put(sk); 836144d56e9SEric Dumazet } 83740fc3423SEric Dumazet if (flags & TCPF_MTU_REDUCED_DEFERRED) { 8384fab9071SNeal Cardwell inet_csk(sk)->icsk_af_ops->mtu_reduced(sk); 839144d56e9SEric Dumazet __sock_put(sk); 840144d56e9SEric Dumazet } 84146d3ceabSEric Dumazet } 84246d3ceabSEric Dumazet EXPORT_SYMBOL(tcp_release_cb); 84346d3ceabSEric Dumazet 84446d3ceabSEric Dumazet void __init tcp_tasklet_init(void) 84546d3ceabSEric Dumazet { 84646d3ceabSEric Dumazet int i; 84746d3ceabSEric Dumazet 84846d3ceabSEric Dumazet for_each_possible_cpu(i) { 84946d3ceabSEric Dumazet struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i); 85046d3ceabSEric Dumazet 85146d3ceabSEric Dumazet INIT_LIST_HEAD(&tsq->head); 85246d3ceabSEric Dumazet tasklet_init(&tsq->tasklet, 85346d3ceabSEric Dumazet tcp_tasklet_func, 85446d3ceabSEric Dumazet (unsigned long)tsq); 85546d3ceabSEric Dumazet } 85646d3ceabSEric Dumazet } 85746d3ceabSEric Dumazet 85846d3ceabSEric Dumazet /* 85946d3ceabSEric Dumazet * Write buffer destructor automatically called from kfree_skb. 8608e3bff96Sstephen hemminger * We can't xmit new skbs from this context, as we might already 86146d3ceabSEric Dumazet * hold qdisc lock. 86246d3ceabSEric Dumazet */ 863d6a4a104SEric Dumazet void tcp_wfree(struct sk_buff *skb) 86446d3ceabSEric Dumazet { 86546d3ceabSEric Dumazet struct sock *sk = skb->sk; 86646d3ceabSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 867408f0a6cSEric Dumazet unsigned long flags, nval, oval; 8689b462d02SEric Dumazet 8699b462d02SEric Dumazet /* Keep one reference on sk_wmem_alloc. 8709b462d02SEric Dumazet * Will be released by sk_free() from here or tcp_tasklet_func() 8719b462d02SEric Dumazet */ 87214afee4bSReshetova, Elena WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc)); 8739b462d02SEric Dumazet 8749b462d02SEric Dumazet /* If this softirq is serviced by ksoftirqd, we are likely under stress. 8759b462d02SEric Dumazet * Wait until our queues (qdisc + devices) are drained. 8769b462d02SEric Dumazet * This gives : 8779b462d02SEric Dumazet * - less callbacks to tcp_write_xmit(), reducing stress (batches) 8789b462d02SEric Dumazet * - chance for incoming ACK (processed by another cpu maybe) 8799b462d02SEric Dumazet * to migrate this flow (skb->ooo_okay will be eventually set) 8809b462d02SEric Dumazet */ 88114afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current) 8829b462d02SEric Dumazet goto out; 88346d3ceabSEric Dumazet 8847aa5470cSEric Dumazet for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) { 88546d3ceabSEric Dumazet struct tsq_tasklet *tsq; 886a9b204d1SEric Dumazet bool empty; 88746d3ceabSEric Dumazet 888408f0a6cSEric Dumazet if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED)) 889408f0a6cSEric Dumazet goto out; 890408f0a6cSEric Dumazet 891b223feb9SEric Dumazet nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED | TCPF_TSQ_DEFERRED; 8927aa5470cSEric Dumazet nval = cmpxchg(&sk->sk_tsq_flags, oval, nval); 893408f0a6cSEric Dumazet if (nval != oval) 894408f0a6cSEric Dumazet continue; 895408f0a6cSEric Dumazet 89646d3ceabSEric Dumazet /* queue this socket to tasklet queue */ 89746d3ceabSEric Dumazet local_irq_save(flags); 898903ceff7SChristoph Lameter tsq = this_cpu_ptr(&tsq_tasklet); 899a9b204d1SEric Dumazet empty = list_empty(&tsq->head); 90046d3ceabSEric Dumazet list_add(&tp->tsq_node, &tsq->head); 901a9b204d1SEric Dumazet if (empty) 90246d3ceabSEric Dumazet tasklet_schedule(&tsq->tasklet); 90346d3ceabSEric Dumazet local_irq_restore(flags); 9049b462d02SEric Dumazet return; 90546d3ceabSEric Dumazet } 9069b462d02SEric Dumazet out: 9079b462d02SEric Dumazet sk_free(sk); 90846d3ceabSEric Dumazet } 90946d3ceabSEric Dumazet 910218af599SEric Dumazet /* Note: Called under hard irq. 911218af599SEric Dumazet * We can not call TCP stack right away. 912218af599SEric Dumazet */ 913218af599SEric Dumazet enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer) 914218af599SEric Dumazet { 915218af599SEric Dumazet struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer); 916218af599SEric Dumazet struct sock *sk = (struct sock *)tp; 917218af599SEric Dumazet unsigned long nval, oval; 918218af599SEric Dumazet 919218af599SEric Dumazet for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) { 920218af599SEric Dumazet struct tsq_tasklet *tsq; 921218af599SEric Dumazet bool empty; 922218af599SEric Dumazet 923218af599SEric Dumazet if (oval & TSQF_QUEUED) 924218af599SEric Dumazet break; 925218af599SEric Dumazet 926218af599SEric Dumazet nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED | TCPF_TSQ_DEFERRED; 927218af599SEric Dumazet nval = cmpxchg(&sk->sk_tsq_flags, oval, nval); 928218af599SEric Dumazet if (nval != oval) 929218af599SEric Dumazet continue; 930218af599SEric Dumazet 93114afee4bSReshetova, Elena if (!refcount_inc_not_zero(&sk->sk_wmem_alloc)) 932218af599SEric Dumazet break; 933218af599SEric Dumazet /* queue this socket to tasklet queue */ 934218af599SEric Dumazet tsq = this_cpu_ptr(&tsq_tasklet); 935218af599SEric Dumazet empty = list_empty(&tsq->head); 936218af599SEric Dumazet list_add(&tp->tsq_node, &tsq->head); 937218af599SEric Dumazet if (empty) 938218af599SEric Dumazet tasklet_schedule(&tsq->tasklet); 939218af599SEric Dumazet break; 940218af599SEric Dumazet } 941218af599SEric Dumazet return HRTIMER_NORESTART; 942218af599SEric Dumazet } 943218af599SEric Dumazet 944218af599SEric Dumazet /* BBR congestion control needs pacing. 945218af599SEric Dumazet * Same remark for SO_MAX_PACING_RATE. 946218af599SEric Dumazet * sch_fq packet scheduler is efficiently handling pacing, 947218af599SEric Dumazet * but is not always installed/used. 948218af599SEric Dumazet * Return true if TCP stack should pace packets itself. 949218af599SEric Dumazet */ 950218af599SEric Dumazet static bool tcp_needs_internal_pacing(const struct sock *sk) 951218af599SEric Dumazet { 952218af599SEric Dumazet return smp_load_acquire(&sk->sk_pacing_status) == SK_PACING_NEEDED; 953218af599SEric Dumazet } 954218af599SEric Dumazet 955218af599SEric Dumazet static void tcp_internal_pacing(struct sock *sk, const struct sk_buff *skb) 956218af599SEric Dumazet { 957218af599SEric Dumazet u64 len_ns; 958218af599SEric Dumazet u32 rate; 959218af599SEric Dumazet 960218af599SEric Dumazet if (!tcp_needs_internal_pacing(sk)) 961218af599SEric Dumazet return; 962218af599SEric Dumazet rate = sk->sk_pacing_rate; 963218af599SEric Dumazet if (!rate || rate == ~0U) 964218af599SEric Dumazet return; 965218af599SEric Dumazet 966218af599SEric Dumazet /* Should account for header sizes as sch_fq does, 967218af599SEric Dumazet * but lets make things simple. 968218af599SEric Dumazet */ 969218af599SEric Dumazet len_ns = (u64)skb->len * NSEC_PER_SEC; 970218af599SEric Dumazet do_div(len_ns, rate); 971218af599SEric Dumazet hrtimer_start(&tcp_sk(sk)->pacing_timer, 972218af599SEric Dumazet ktime_add_ns(ktime_get(), len_ns), 973218af599SEric Dumazet HRTIMER_MODE_ABS_PINNED); 974218af599SEric Dumazet } 975218af599SEric Dumazet 976e2080072SEric Dumazet static void tcp_update_skb_after_send(struct tcp_sock *tp, struct sk_buff *skb) 977e2080072SEric Dumazet { 978e2080072SEric Dumazet skb->skb_mstamp = tp->tcp_mstamp; 979e2080072SEric Dumazet list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue); 980e2080072SEric Dumazet } 981e2080072SEric Dumazet 9821da177e4SLinus Torvalds /* This routine actually transmits TCP packets queued in by 9831da177e4SLinus Torvalds * tcp_do_sendmsg(). This is used by both the initial 9841da177e4SLinus Torvalds * transmission and possible later retransmissions. 9851da177e4SLinus Torvalds * All SKB's seen here are completely headerless. It is our 9861da177e4SLinus Torvalds * job to build the TCP header, and pass the packet down to 9871da177e4SLinus Torvalds * IP so it can do the same plus pass the packet off to the 9881da177e4SLinus Torvalds * device. 9891da177e4SLinus Torvalds * 9901da177e4SLinus Torvalds * We are working here with either a clone of the original 9911da177e4SLinus Torvalds * SKB, or a fresh unique copy made by the retransmit engine. 9921da177e4SLinus Torvalds */ 993056834d9SIlpo Järvinen static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, 994056834d9SIlpo Järvinen gfp_t gfp_mask) 9951da177e4SLinus Torvalds { 9966687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 997dfb4b9dcSDavid S. Miller struct inet_sock *inet; 998dfb4b9dcSDavid S. Miller struct tcp_sock *tp; 999dfb4b9dcSDavid S. Miller struct tcp_skb_cb *tcb; 100033ad798cSAdam Langley struct tcp_out_options opts; 100195c96174SEric Dumazet unsigned int tcp_options_size, tcp_header_size; 10028c72c65bSEric Dumazet struct sk_buff *oskb = NULL; 1003cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 10041da177e4SLinus Torvalds struct tcphdr *th; 10051da177e4SLinus Torvalds int err; 10061da177e4SLinus Torvalds 1007dfb4b9dcSDavid S. Miller BUG_ON(!skb || !tcp_skb_pcount(skb)); 10086f094b9eSLawrence Brakmo tp = tcp_sk(sk); 1009dfb4b9dcSDavid S. Miller 1010ccdbb6e9SEric Dumazet if (clone_it) { 10116f094b9eSLawrence Brakmo TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq 10126f094b9eSLawrence Brakmo - tp->snd_una; 10138c72c65bSEric Dumazet oskb = skb; 1014e2080072SEric Dumazet 1015e2080072SEric Dumazet tcp_skb_tsorted_save(oskb) { 1016e2080072SEric Dumazet if (unlikely(skb_cloned(oskb))) 1017e2080072SEric Dumazet skb = pskb_copy(oskb, gfp_mask); 1018dfb4b9dcSDavid S. Miller else 1019e2080072SEric Dumazet skb = skb_clone(oskb, gfp_mask); 1020e2080072SEric Dumazet } tcp_skb_tsorted_restore(oskb); 1021e2080072SEric Dumazet 1022dfb4b9dcSDavid S. Miller if (unlikely(!skb)) 1023dfb4b9dcSDavid S. Miller return -ENOBUFS; 1024dfb4b9dcSDavid S. Miller } 10258c72c65bSEric Dumazet skb->skb_mstamp = tp->tcp_mstamp; 1026dfb4b9dcSDavid S. Miller 1027dfb4b9dcSDavid S. Miller inet = inet_sk(sk); 1028dfb4b9dcSDavid S. Miller tcb = TCP_SKB_CB(skb); 102933ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 10301da177e4SLinus Torvalds 10314de075e0SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) 103233ad798cSAdam Langley tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); 103333ad798cSAdam Langley else 103433ad798cSAdam Langley tcp_options_size = tcp_established_options(sk, skb, &opts, 103533ad798cSAdam Langley &md5); 103633ad798cSAdam Langley tcp_header_size = tcp_options_size + sizeof(struct tcphdr); 10371da177e4SLinus Torvalds 1038547669d4SEric Dumazet /* if no packet is in qdisc/device queue, then allow XPS to select 1039b2532eb9SEric Dumazet * another queue. We can be called from tcp_tsq_handler() 1040b2532eb9SEric Dumazet * which holds one reference to sk_wmem_alloc. 1041b2532eb9SEric Dumazet * 1042b2532eb9SEric Dumazet * TODO: Ideally, in-flight pure ACK packets should not matter here. 1043b2532eb9SEric Dumazet * One way to get this would be to set skb->truesize = 2 on them. 1044547669d4SEric Dumazet */ 1045b2532eb9SEric Dumazet skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1); 10461da177e4SLinus Torvalds 104738ab52e8SEric Dumazet /* If we had to use memory reserve to allocate this skb, 104838ab52e8SEric Dumazet * this might cause drops if packet is looped back : 104938ab52e8SEric Dumazet * Other socket might not have SOCK_MEMALLOC. 105038ab52e8SEric Dumazet * Packets not looped back do not care about pfmemalloc. 105138ab52e8SEric Dumazet */ 105238ab52e8SEric Dumazet skb->pfmemalloc = 0; 105338ab52e8SEric Dumazet 1054aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 1055aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 105646d3ceabSEric Dumazet 105746d3ceabSEric Dumazet skb_orphan(skb); 105846d3ceabSEric Dumazet skb->sk = sk; 10591d2077acSEric Dumazet skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree; 1060b73c3d0eSTom Herbert skb_set_hash_from_sk(skb, sk); 106114afee4bSReshetova, Elena refcount_add(skb->truesize, &sk->sk_wmem_alloc); 10621da177e4SLinus Torvalds 1063c3a2e837SJulian Anastasov skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm); 1064c3a2e837SJulian Anastasov 10651da177e4SLinus Torvalds /* Build TCP header and checksum it. */ 1066ea1627c2SEric Dumazet th = (struct tcphdr *)skb->data; 1067c720c7e8SEric Dumazet th->source = inet->inet_sport; 1068c720c7e8SEric Dumazet th->dest = inet->inet_dport; 10691da177e4SLinus Torvalds th->seq = htonl(tcb->seq); 10701da177e4SLinus Torvalds th->ack_seq = htonl(tp->rcv_nxt); 1071df7a3b07SAl Viro *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | 10724de075e0SEric Dumazet tcb->tcp_flags); 1073dfb4b9dcSDavid S. Miller 10741da177e4SLinus Torvalds th->check = 0; 10751da177e4SLinus Torvalds th->urg_ptr = 0; 10761da177e4SLinus Torvalds 107733f5f57eSIlpo Järvinen /* The urg_mode check is necessary during a below snd_una win probe */ 10787691367dSHerbert Xu if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { 10797691367dSHerbert Xu if (before(tp->snd_up, tcb->seq + 0x10000)) { 10801da177e4SLinus Torvalds th->urg_ptr = htons(tp->snd_up - tcb->seq); 10811da177e4SLinus Torvalds th->urg = 1; 10827691367dSHerbert Xu } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { 10830eae88f3SEric Dumazet th->urg_ptr = htons(0xFFFF); 10847691367dSHerbert Xu th->urg = 1; 10857691367dSHerbert Xu } 10861da177e4SLinus Torvalds } 10871da177e4SLinus Torvalds 1088bd0388aeSWilliam Allen Simpson tcp_options_write((__be32 *)(th + 1), tp, &opts); 108951466a75SEric Dumazet skb_shinfo(skb)->gso_type = sk->sk_gso_type; 1090ea1627c2SEric Dumazet if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) { 1091ea1627c2SEric Dumazet th->window = htons(tcp_select_window(sk)); 1092ea1627c2SEric Dumazet tcp_ecn_send(sk, skb, th, tcp_header_size); 1093ea1627c2SEric Dumazet } else { 1094ea1627c2SEric Dumazet /* RFC1323: The window in SYN & SYN/ACK segments 1095ea1627c2SEric Dumazet * is never scaled. 1096ea1627c2SEric Dumazet */ 1097ea1627c2SEric Dumazet th->window = htons(min(tp->rcv_wnd, 65535U)); 1098ea1627c2SEric Dumazet } 1099cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 1100cfb6eeb4SYOSHIFUJI Hideaki /* Calculate the MD5 hash, as we have all we need now */ 1101cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 1102a465419bSEric Dumazet sk_nocaps_add(sk, NETIF_F_GSO_MASK); 1103bd0388aeSWilliam Allen Simpson tp->af_specific->calc_md5_hash(opts.hash_location, 110439f8e58eSEric Dumazet md5, sk, skb); 1105cfb6eeb4SYOSHIFUJI Hideaki } 1106cfb6eeb4SYOSHIFUJI Hideaki #endif 1107cfb6eeb4SYOSHIFUJI Hideaki 1108bb296246SHerbert Xu icsk->icsk_af_ops->send_check(sk, skb); 11091da177e4SLinus Torvalds 11104de075e0SEric Dumazet if (likely(tcb->tcp_flags & TCPHDR_ACK)) 1111fc6415bcSDavid S. Miller tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); 11121da177e4SLinus Torvalds 1113a44d6eacSMartin KaFai Lau if (skb->len != tcp_header_size) { 1114cf533ea5SEric Dumazet tcp_event_data_sent(tp, sk); 1115a44d6eacSMartin KaFai Lau tp->data_segs_out += tcp_skb_pcount(skb); 1116218af599SEric Dumazet tcp_internal_pacing(sk, skb); 1117a44d6eacSMartin KaFai Lau } 11181da177e4SLinus Torvalds 1119bd37a088SWei Yongjun if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) 1120aa2ea058STom Herbert TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, 1121aa2ea058STom Herbert tcp_skb_pcount(skb)); 11221da177e4SLinus Torvalds 11232efd055cSMarcelo Ricardo Leitner tp->segs_out += tcp_skb_pcount(skb); 1124f69ad292SEric Dumazet /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */ 1125cd7d8498SEric Dumazet skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb); 1126f69ad292SEric Dumazet skb_shinfo(skb)->gso_size = tcp_skb_mss(skb); 1127cd7d8498SEric Dumazet 11287faee5c0SEric Dumazet /* Our usage of tstamp should remain private */ 11292456e855SThomas Gleixner skb->tstamp = 0; 1130971f10ecSEric Dumazet 1131971f10ecSEric Dumazet /* Cleanup our debris for IP stacks */ 1132971f10ecSEric Dumazet memset(skb->cb, 0, max(sizeof(struct inet_skb_parm), 1133971f10ecSEric Dumazet sizeof(struct inet6_skb_parm))); 1134971f10ecSEric Dumazet 1135b0270e91SEric Dumazet err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl); 11367faee5c0SEric Dumazet 11378c72c65bSEric Dumazet if (unlikely(err > 0)) { 11385ee2c941SChristoph Paasch tcp_enter_cwr(sk); 11398c72c65bSEric Dumazet err = net_xmit_eval(err); 11408c72c65bSEric Dumazet } 1141fc225799SEric Dumazet if (!err && oskb) { 1142e2080072SEric Dumazet tcp_update_skb_after_send(tp, oskb); 1143fc225799SEric Dumazet tcp_rate_skb_sent(sk, oskb); 1144fc225799SEric Dumazet } 11458c72c65bSEric Dumazet return err; 11461da177e4SLinus Torvalds } 11471da177e4SLinus Torvalds 114867edfef7SAndi Kleen /* This routine just queues the buffer for sending. 11491da177e4SLinus Torvalds * 11501da177e4SLinus Torvalds * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, 11511da177e4SLinus Torvalds * otherwise socket can stall. 11521da177e4SLinus Torvalds */ 11531da177e4SLinus Torvalds static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) 11541da177e4SLinus Torvalds { 11551da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 11561da177e4SLinus Torvalds 11571da177e4SLinus Torvalds /* Advance write_seq and place onto the write_queue. */ 11581da177e4SLinus Torvalds tp->write_seq = TCP_SKB_CB(skb)->end_seq; 1159f4a775d1SEric Dumazet __skb_header_release(skb); 1160fe067e8aSDavid S. Miller tcp_add_write_queue_tail(sk, skb); 11613ab224beSHideo Aoki sk->sk_wmem_queued += skb->truesize; 11623ab224beSHideo Aoki sk_mem_charge(sk, skb->truesize); 11631da177e4SLinus Torvalds } 11641da177e4SLinus Torvalds 116567edfef7SAndi Kleen /* Initialize TSO segments for a packet. */ 11665bbb432cSEric Dumazet static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now) 1167f6302d1dSDavid S. Miller { 11688f26fb1cSEric Dumazet if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) { 1169f6302d1dSDavid S. Miller /* Avoid the costly divide in the normal 1170f6302d1dSDavid S. Miller * non-TSO case. 1171f6302d1dSDavid S. Miller */ 1172cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, 1); 1173f69ad292SEric Dumazet TCP_SKB_CB(skb)->tcp_gso_size = 0; 1174f6302d1dSDavid S. Miller } else { 1175cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now)); 1176f69ad292SEric Dumazet TCP_SKB_CB(skb)->tcp_gso_size = mss_now; 11771da177e4SLinus Torvalds } 11781da177e4SLinus Torvalds } 11791da177e4SLinus Torvalds 118091fed7a1SIlpo Järvinen /* When a modification to fackets out becomes necessary, we need to check 118168f8353bSIlpo Järvinen * skb is counted to fackets_out or not. 118291fed7a1SIlpo Järvinen */ 1183cf533ea5SEric Dumazet static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb, 118491fed7a1SIlpo Järvinen int decr) 118591fed7a1SIlpo Järvinen { 1186a47e5a98SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1187a47e5a98SIlpo Järvinen 1188dc86967bSIlpo Järvinen if (!tp->sacked_out || tcp_is_reno(tp)) 118991fed7a1SIlpo Järvinen return; 119091fed7a1SIlpo Järvinen 11916859d494SIlpo Järvinen if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq)) 119291fed7a1SIlpo Järvinen tp->fackets_out -= decr; 119391fed7a1SIlpo Järvinen } 119491fed7a1SIlpo Järvinen 1195797108d1SIlpo Järvinen /* Pcount in the middle of the write queue got changed, we need to do various 1196797108d1SIlpo Järvinen * tweaks to fix counters 1197797108d1SIlpo Järvinen */ 1198cf533ea5SEric Dumazet static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) 1199797108d1SIlpo Järvinen { 1200797108d1SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1201797108d1SIlpo Järvinen 1202797108d1SIlpo Järvinen tp->packets_out -= decr; 1203797108d1SIlpo Järvinen 1204797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 1205797108d1SIlpo Järvinen tp->sacked_out -= decr; 1206797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 1207797108d1SIlpo Järvinen tp->retrans_out -= decr; 1208797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 1209797108d1SIlpo Järvinen tp->lost_out -= decr; 1210797108d1SIlpo Järvinen 1211797108d1SIlpo Järvinen /* Reno case is special. Sigh... */ 1212797108d1SIlpo Järvinen if (tcp_is_reno(tp) && decr > 0) 1213797108d1SIlpo Järvinen tp->sacked_out -= min_t(u32, tp->sacked_out, decr); 1214797108d1SIlpo Järvinen 1215797108d1SIlpo Järvinen tcp_adjust_fackets_out(sk, skb, decr); 1216797108d1SIlpo Järvinen 1217797108d1SIlpo Järvinen if (tp->lost_skb_hint && 1218797108d1SIlpo Järvinen before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && 121952cf3cc8SIlpo Järvinen (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))) 1220797108d1SIlpo Järvinen tp->lost_cnt_hint -= decr; 1221797108d1SIlpo Järvinen 1222797108d1SIlpo Järvinen tcp_verify_left_out(tp); 1223797108d1SIlpo Järvinen } 1224797108d1SIlpo Järvinen 12250a2cf20cSSoheil Hassas Yeganeh static bool tcp_has_tx_tstamp(const struct sk_buff *skb) 12260a2cf20cSSoheil Hassas Yeganeh { 12270a2cf20cSSoheil Hassas Yeganeh return TCP_SKB_CB(skb)->txstamp_ack || 12280a2cf20cSSoheil Hassas Yeganeh (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP); 12290a2cf20cSSoheil Hassas Yeganeh } 12300a2cf20cSSoheil Hassas Yeganeh 1231490cc7d0SWillem de Bruijn static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2) 1232490cc7d0SWillem de Bruijn { 1233490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo = skb_shinfo(skb); 1234490cc7d0SWillem de Bruijn 12350a2cf20cSSoheil Hassas Yeganeh if (unlikely(tcp_has_tx_tstamp(skb)) && 1236490cc7d0SWillem de Bruijn !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) { 1237490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo2 = skb_shinfo(skb2); 1238490cc7d0SWillem de Bruijn u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP; 1239490cc7d0SWillem de Bruijn 1240490cc7d0SWillem de Bruijn shinfo->tx_flags &= ~tsflags; 1241490cc7d0SWillem de Bruijn shinfo2->tx_flags |= tsflags; 1242490cc7d0SWillem de Bruijn swap(shinfo->tskey, shinfo2->tskey); 1243b51e13faSMartin KaFai Lau TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack; 1244b51e13faSMartin KaFai Lau TCP_SKB_CB(skb)->txstamp_ack = 0; 1245490cc7d0SWillem de Bruijn } 1246490cc7d0SWillem de Bruijn } 1247490cc7d0SWillem de Bruijn 1248a166140eSMartin KaFai Lau static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2) 1249a166140eSMartin KaFai Lau { 1250a166140eSMartin KaFai Lau TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor; 1251a166140eSMartin KaFai Lau TCP_SKB_CB(skb)->eor = 0; 1252a166140eSMartin KaFai Lau } 1253a166140eSMartin KaFai Lau 1254*75c119afSEric Dumazet /* Insert buff after skb on the write or rtx queue of sk. */ 1255*75c119afSEric Dumazet static void tcp_insert_write_queue_after(struct sk_buff *skb, 1256*75c119afSEric Dumazet struct sk_buff *buff, 1257*75c119afSEric Dumazet struct sock *sk, 1258*75c119afSEric Dumazet enum tcp_queue tcp_queue) 1259*75c119afSEric Dumazet { 1260*75c119afSEric Dumazet if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE) 1261*75c119afSEric Dumazet __skb_queue_after(&sk->sk_write_queue, skb, buff); 1262*75c119afSEric Dumazet else 1263*75c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, buff); 1264*75c119afSEric Dumazet } 1265*75c119afSEric Dumazet 12661da177e4SLinus Torvalds /* Function to create two new TCP segments. Shrinks the given segment 12671da177e4SLinus Torvalds * to the specified size and appends a new segment with the rest of the 12681da177e4SLinus Torvalds * packet to the list. This won't be called frequently, I hope. 12691da177e4SLinus Torvalds * Remember, these are still headerless SKBs at this point. 12701da177e4SLinus Torvalds */ 1271*75c119afSEric Dumazet int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue, 1272*75c119afSEric Dumazet struct sk_buff *skb, u32 len, 12736cc55e09SOctavian Purdila unsigned int mss_now, gfp_t gfp) 12741da177e4SLinus Torvalds { 12751da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 12761da177e4SLinus Torvalds struct sk_buff *buff; 12776475be16SDavid S. Miller int nsize, old_factor; 1278b60b49eaSHerbert Xu int nlen; 12799ce01461SIlpo Järvinen u8 flags; 12801da177e4SLinus Torvalds 12812fceec13SIlpo Järvinen if (WARN_ON(len > skb->len)) 12822fceec13SIlpo Järvinen return -EINVAL; 12836a438bbeSStephen Hemminger 12841da177e4SLinus Torvalds nsize = skb_headlen(skb) - len; 12851da177e4SLinus Torvalds if (nsize < 0) 12861da177e4SLinus Torvalds nsize = 0; 12871da177e4SLinus Torvalds 12886cc55e09SOctavian Purdila if (skb_unclone(skb, gfp)) 12891da177e4SLinus Torvalds return -ENOMEM; 12901da177e4SLinus Torvalds 12911da177e4SLinus Torvalds /* Get a new skb... force flag on. */ 1292eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, nsize, gfp, true); 129351456b29SIan Morris if (!buff) 12941da177e4SLinus Torvalds return -ENOMEM; /* We'll just try again later. */ 1295ef5cb973SHerbert Xu 12963ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 12973ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1298b60b49eaSHerbert Xu nlen = skb->len - len - nsize; 1299b60b49eaSHerbert Xu buff->truesize += nlen; 1300b60b49eaSHerbert Xu skb->truesize -= nlen; 13011da177e4SLinus Torvalds 13021da177e4SLinus Torvalds /* Correct the sequence numbers. */ 13031da177e4SLinus Torvalds TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 13041da177e4SLinus Torvalds TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 13051da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 13061da177e4SLinus Torvalds 13071da177e4SLinus Torvalds /* PSH and FIN should only be set in the second packet. */ 13084de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 13094de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 13104de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1311e14c3cafSHerbert Xu TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; 1312a166140eSMartin KaFai Lau tcp_skb_fragment_eor(skb, buff); 13131da177e4SLinus Torvalds 131484fa7933SPatrick McHardy if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { 13151da177e4SLinus Torvalds /* Copy and checksum data tail into the new buffer. */ 1316056834d9SIlpo Järvinen buff->csum = csum_partial_copy_nocheck(skb->data + len, 1317056834d9SIlpo Järvinen skb_put(buff, nsize), 13181da177e4SLinus Torvalds nsize, 0); 13191da177e4SLinus Torvalds 13201da177e4SLinus Torvalds skb_trim(skb, len); 13211da177e4SLinus Torvalds 13221da177e4SLinus Torvalds skb->csum = csum_block_sub(skb->csum, buff->csum, len); 13231da177e4SLinus Torvalds } else { 132484fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 13251da177e4SLinus Torvalds skb_split(skb, buff, len); 13261da177e4SLinus Torvalds } 13271da177e4SLinus Torvalds 13281da177e4SLinus Torvalds buff->ip_summed = skb->ip_summed; 13291da177e4SLinus Torvalds 1330a61bbcf2SPatrick McHardy buff->tstamp = skb->tstamp; 1331490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 13321da177e4SLinus Torvalds 13336475be16SDavid S. Miller old_factor = tcp_skb_pcount(skb); 13346475be16SDavid S. Miller 13351da177e4SLinus Torvalds /* Fix up tso_factor for both original and new SKB. */ 13365bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 13375bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 13381da177e4SLinus Torvalds 1339b9f64820SYuchung Cheng /* Update delivered info for the new segment */ 1340b9f64820SYuchung Cheng TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx; 1341b9f64820SYuchung Cheng 13426475be16SDavid S. Miller /* If this packet has been sent out already, we must 13436475be16SDavid S. Miller * adjust the various packet counters. 13446475be16SDavid S. Miller */ 1345cf0b450cSHerbert Xu if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 13466475be16SDavid S. Miller int diff = old_factor - tcp_skb_pcount(skb) - 13476475be16SDavid S. Miller tcp_skb_pcount(buff); 13481da177e4SLinus Torvalds 1349797108d1SIlpo Järvinen if (diff) 1350797108d1SIlpo Järvinen tcp_adjust_pcount(sk, skb, diff); 13511da177e4SLinus Torvalds } 13521da177e4SLinus Torvalds 13531da177e4SLinus Torvalds /* Link BUFF into the send queue. */ 1354f4a775d1SEric Dumazet __skb_header_release(buff); 1355*75c119afSEric Dumazet tcp_insert_write_queue_after(skb, buff, sk, tcp_queue); 1356e2080072SEric Dumazet list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor); 13571da177e4SLinus Torvalds 13581da177e4SLinus Torvalds return 0; 13591da177e4SLinus Torvalds } 13601da177e4SLinus Torvalds 1361f4d01666SEric Dumazet /* This is similar to __pskb_pull_tail(). The difference is that pulled 1362f4d01666SEric Dumazet * data is not copied, but immediately discarded. 13631da177e4SLinus Torvalds */ 13647162fb24SEric Dumazet static int __pskb_trim_head(struct sk_buff *skb, int len) 13651da177e4SLinus Torvalds { 13667b7fc97aSEric Dumazet struct skb_shared_info *shinfo; 13671da177e4SLinus Torvalds int i, k, eat; 13681da177e4SLinus Torvalds 13694fa48bf3SEric Dumazet eat = min_t(int, len, skb_headlen(skb)); 13704fa48bf3SEric Dumazet if (eat) { 13714fa48bf3SEric Dumazet __skb_pull(skb, eat); 13724fa48bf3SEric Dumazet len -= eat; 13734fa48bf3SEric Dumazet if (!len) 13747162fb24SEric Dumazet return 0; 13754fa48bf3SEric Dumazet } 13761da177e4SLinus Torvalds eat = len; 13771da177e4SLinus Torvalds k = 0; 13787b7fc97aSEric Dumazet shinfo = skb_shinfo(skb); 13797b7fc97aSEric Dumazet for (i = 0; i < shinfo->nr_frags; i++) { 13807b7fc97aSEric Dumazet int size = skb_frag_size(&shinfo->frags[i]); 13819e903e08SEric Dumazet 13829e903e08SEric Dumazet if (size <= eat) { 1383aff65da0SIan Campbell skb_frag_unref(skb, i); 13849e903e08SEric Dumazet eat -= size; 13851da177e4SLinus Torvalds } else { 13867b7fc97aSEric Dumazet shinfo->frags[k] = shinfo->frags[i]; 13871da177e4SLinus Torvalds if (eat) { 13887b7fc97aSEric Dumazet shinfo->frags[k].page_offset += eat; 13897b7fc97aSEric Dumazet skb_frag_size_sub(&shinfo->frags[k], eat); 13901da177e4SLinus Torvalds eat = 0; 13911da177e4SLinus Torvalds } 13921da177e4SLinus Torvalds k++; 13931da177e4SLinus Torvalds } 13941da177e4SLinus Torvalds } 13957b7fc97aSEric Dumazet shinfo->nr_frags = k; 13961da177e4SLinus Torvalds 13971da177e4SLinus Torvalds skb->data_len -= len; 13981da177e4SLinus Torvalds skb->len = skb->data_len; 13997162fb24SEric Dumazet return len; 14001da177e4SLinus Torvalds } 14011da177e4SLinus Torvalds 140267edfef7SAndi Kleen /* Remove acked data from a packet in the transmit queue. */ 14031da177e4SLinus Torvalds int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 14041da177e4SLinus Torvalds { 14057162fb24SEric Dumazet u32 delta_truesize; 14067162fb24SEric Dumazet 140714bbd6a5SPravin B Shelar if (skb_unclone(skb, GFP_ATOMIC)) 14081da177e4SLinus Torvalds return -ENOMEM; 14091da177e4SLinus Torvalds 14107162fb24SEric Dumazet delta_truesize = __pskb_trim_head(skb, len); 14111da177e4SLinus Torvalds 14121da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq += len; 141384fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 14141da177e4SLinus Torvalds 14157162fb24SEric Dumazet if (delta_truesize) { 14167162fb24SEric Dumazet skb->truesize -= delta_truesize; 14177162fb24SEric Dumazet sk->sk_wmem_queued -= delta_truesize; 14187162fb24SEric Dumazet sk_mem_uncharge(sk, delta_truesize); 14191da177e4SLinus Torvalds sock_set_flag(sk, SOCK_QUEUE_SHRUNK); 14207162fb24SEric Dumazet } 14211da177e4SLinus Torvalds 14225b35e1e6SNeal Cardwell /* Any change of skb->len requires recalculation of tso factor. */ 14231da177e4SLinus Torvalds if (tcp_skb_pcount(skb) > 1) 14245bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb)); 14251da177e4SLinus Torvalds 14261da177e4SLinus Torvalds return 0; 14271da177e4SLinus Torvalds } 14281da177e4SLinus Torvalds 14291b63edd6SYuchung Cheng /* Calculate MSS not accounting any TCP options. */ 14301b63edd6SYuchung Cheng static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu) 14315d424d5aSJohn Heffner { 1432cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1433cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 14345d424d5aSJohn Heffner int mss_now; 14355d424d5aSJohn Heffner 14365d424d5aSJohn Heffner /* Calculate base mss without TCP options: 14375d424d5aSJohn Heffner It is MMS_S - sizeof(tcphdr) of rfc1122 14385d424d5aSJohn Heffner */ 14395d424d5aSJohn Heffner mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 14405d424d5aSJohn Heffner 144167469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 144267469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 144367469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 144467469601SEric Dumazet 144567469601SEric Dumazet if (dst && dst_allfrag(dst)) 144667469601SEric Dumazet mss_now -= icsk->icsk_af_ops->net_frag_header_len; 144767469601SEric Dumazet } 144867469601SEric Dumazet 14495d424d5aSJohn Heffner /* Clamp it (mss_clamp does not include tcp options) */ 14505d424d5aSJohn Heffner if (mss_now > tp->rx_opt.mss_clamp) 14515d424d5aSJohn Heffner mss_now = tp->rx_opt.mss_clamp; 14525d424d5aSJohn Heffner 14535d424d5aSJohn Heffner /* Now subtract optional transport overhead */ 14545d424d5aSJohn Heffner mss_now -= icsk->icsk_ext_hdr_len; 14555d424d5aSJohn Heffner 14565d424d5aSJohn Heffner /* Then reserve room for full set of TCP options and 8 bytes of data */ 14575d424d5aSJohn Heffner if (mss_now < 48) 14585d424d5aSJohn Heffner mss_now = 48; 14595d424d5aSJohn Heffner return mss_now; 14605d424d5aSJohn Heffner } 14615d424d5aSJohn Heffner 14621b63edd6SYuchung Cheng /* Calculate MSS. Not accounting for SACKs here. */ 14631b63edd6SYuchung Cheng int tcp_mtu_to_mss(struct sock *sk, int pmtu) 14641b63edd6SYuchung Cheng { 14651b63edd6SYuchung Cheng /* Subtract TCP options size, not including SACKs */ 14661b63edd6SYuchung Cheng return __tcp_mtu_to_mss(sk, pmtu) - 14671b63edd6SYuchung Cheng (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr)); 14681b63edd6SYuchung Cheng } 14691b63edd6SYuchung Cheng 14705d424d5aSJohn Heffner /* Inverse of above */ 147167469601SEric Dumazet int tcp_mss_to_mtu(struct sock *sk, int mss) 14725d424d5aSJohn Heffner { 1473cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1474cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 14755d424d5aSJohn Heffner int mtu; 14765d424d5aSJohn Heffner 14775d424d5aSJohn Heffner mtu = mss + 14785d424d5aSJohn Heffner tp->tcp_header_len + 14795d424d5aSJohn Heffner icsk->icsk_ext_hdr_len + 14805d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 14815d424d5aSJohn Heffner 148267469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 148367469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 148467469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 148567469601SEric Dumazet 148667469601SEric Dumazet if (dst && dst_allfrag(dst)) 148767469601SEric Dumazet mtu += icsk->icsk_af_ops->net_frag_header_len; 148867469601SEric Dumazet } 14895d424d5aSJohn Heffner return mtu; 14905d424d5aSJohn Heffner } 1491556c6b46SNeal Cardwell EXPORT_SYMBOL(tcp_mss_to_mtu); 14925d424d5aSJohn Heffner 149367edfef7SAndi Kleen /* MTU probing init per socket */ 14945d424d5aSJohn Heffner void tcp_mtup_init(struct sock *sk) 14955d424d5aSJohn Heffner { 14965d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 14975d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 1498b0f9ca53SFan Du struct net *net = sock_net(sk); 14995d424d5aSJohn Heffner 1500b0f9ca53SFan Du icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1; 15015d424d5aSJohn Heffner icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 15025d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 1503b0f9ca53SFan Du icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss); 15045d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 150505cbc0dbSFan Du if (icsk->icsk_mtup.enabled) 1506c74df29aSEric Dumazet icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; 15075d424d5aSJohn Heffner } 15084bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_mtup_init); 15095d424d5aSJohn Heffner 15101da177e4SLinus Torvalds /* This function synchronize snd mss to current pmtu/exthdr set. 15111da177e4SLinus Torvalds 15121da177e4SLinus Torvalds tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 15131da177e4SLinus Torvalds for TCP options, but includes only bare TCP header. 15141da177e4SLinus Torvalds 15151da177e4SLinus Torvalds tp->rx_opt.mss_clamp is mss negotiated at connection setup. 1516caa20d9aSStephen Hemminger It is minimum of user_mss and mss received with SYN. 15171da177e4SLinus Torvalds It also does not include TCP options. 15181da177e4SLinus Torvalds 1519d83d8461SArnaldo Carvalho de Melo inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. 15201da177e4SLinus Torvalds 15211da177e4SLinus Torvalds tp->mss_cache is current effective sending mss, including 15221da177e4SLinus Torvalds all tcp options except for SACKs. It is evaluated, 15231da177e4SLinus Torvalds taking into account current pmtu, but never exceeds 15241da177e4SLinus Torvalds tp->rx_opt.mss_clamp. 15251da177e4SLinus Torvalds 15261da177e4SLinus Torvalds NOTE1. rfc1122 clearly states that advertised MSS 15271da177e4SLinus Torvalds DOES NOT include either tcp or ip options. 15281da177e4SLinus Torvalds 1529d83d8461SArnaldo Carvalho de Melo NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache 1530d83d8461SArnaldo Carvalho de Melo are READ ONLY outside this function. --ANK (980731) 15311da177e4SLinus Torvalds */ 15321da177e4SLinus Torvalds unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 15331da177e4SLinus Torvalds { 15341da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1535d83d8461SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 15365d424d5aSJohn Heffner int mss_now; 15371da177e4SLinus Torvalds 15385d424d5aSJohn Heffner if (icsk->icsk_mtup.search_high > pmtu) 15395d424d5aSJohn Heffner icsk->icsk_mtup.search_high = pmtu; 15401da177e4SLinus Torvalds 15415d424d5aSJohn Heffner mss_now = tcp_mtu_to_mss(sk, pmtu); 1542409d22b4SIlpo Järvinen mss_now = tcp_bound_to_half_wnd(tp, mss_now); 15431da177e4SLinus Torvalds 15441da177e4SLinus Torvalds /* And store cached results */ 1545d83d8461SArnaldo Carvalho de Melo icsk->icsk_pmtu_cookie = pmtu; 15465d424d5aSJohn Heffner if (icsk->icsk_mtup.enabled) 15475d424d5aSJohn Heffner mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); 1548c1b4a7e6SDavid S. Miller tp->mss_cache = mss_now; 15491da177e4SLinus Torvalds 15501da177e4SLinus Torvalds return mss_now; 15511da177e4SLinus Torvalds } 15524bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_sync_mss); 15531da177e4SLinus Torvalds 15541da177e4SLinus Torvalds /* Compute the current effective MSS, taking SACKs and IP options, 15551da177e4SLinus Torvalds * and even PMTU discovery events into account. 15561da177e4SLinus Torvalds */ 15570c54b85fSIlpo Järvinen unsigned int tcp_current_mss(struct sock *sk) 15581da177e4SLinus Torvalds { 1559cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1560cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1561c1b4a7e6SDavid S. Miller u32 mss_now; 156295c96174SEric Dumazet unsigned int header_len; 156333ad798cSAdam Langley struct tcp_out_options opts; 156433ad798cSAdam Langley struct tcp_md5sig_key *md5; 15651da177e4SLinus Torvalds 1566c1b4a7e6SDavid S. Miller mss_now = tp->mss_cache; 1567c1b4a7e6SDavid S. Miller 15681da177e4SLinus Torvalds if (dst) { 15691da177e4SLinus Torvalds u32 mtu = dst_mtu(dst); 1570d83d8461SArnaldo Carvalho de Melo if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 15711da177e4SLinus Torvalds mss_now = tcp_sync_mss(sk, mtu); 15721da177e4SLinus Torvalds } 15731da177e4SLinus Torvalds 157433ad798cSAdam Langley header_len = tcp_established_options(sk, NULL, &opts, &md5) + 157533ad798cSAdam Langley sizeof(struct tcphdr); 157633ad798cSAdam Langley /* The mss_cache is sized based on tp->tcp_header_len, which assumes 157733ad798cSAdam Langley * some common options. If this is an odd packet (because we have SACK 157833ad798cSAdam Langley * blocks etc) then our calculated header_len will be different, and 157933ad798cSAdam Langley * we have to adjust mss_now correspondingly */ 158033ad798cSAdam Langley if (header_len != tp->tcp_header_len) { 158133ad798cSAdam Langley int delta = (int) header_len - tp->tcp_header_len; 158233ad798cSAdam Langley mss_now -= delta; 158333ad798cSAdam Langley } 1584cfb6eeb4SYOSHIFUJI Hideaki 15851da177e4SLinus Torvalds return mss_now; 15861da177e4SLinus Torvalds } 15871da177e4SLinus Torvalds 158886fd14adSWeiping Pan /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. 158986fd14adSWeiping Pan * As additional protections, we do not touch cwnd in retransmission phases, 159086fd14adSWeiping Pan * and if application hit its sndbuf limit recently. 159186fd14adSWeiping Pan */ 159286fd14adSWeiping Pan static void tcp_cwnd_application_limited(struct sock *sk) 1593a762a980SDavid S. Miller { 15949e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1595a762a980SDavid S. Miller 159686fd14adSWeiping Pan if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && 159786fd14adSWeiping Pan sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 159886fd14adSWeiping Pan /* Limited by application or receiver window. */ 159986fd14adSWeiping Pan u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); 160086fd14adSWeiping Pan u32 win_used = max(tp->snd_cwnd_used, init_win); 160186fd14adSWeiping Pan if (win_used < tp->snd_cwnd) { 160286fd14adSWeiping Pan tp->snd_ssthresh = tcp_current_ssthresh(sk); 160386fd14adSWeiping Pan tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1; 160486fd14adSWeiping Pan } 160586fd14adSWeiping Pan tp->snd_cwnd_used = 0; 160686fd14adSWeiping Pan } 1607c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 160886fd14adSWeiping Pan } 160986fd14adSWeiping Pan 1610ca8a2263SNeal Cardwell static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited) 1611a762a980SDavid S. Miller { 16121b1fc3fdSWei Wang const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 1613a762a980SDavid S. Miller struct tcp_sock *tp = tcp_sk(sk); 1614a762a980SDavid S. Miller 1615ca8a2263SNeal Cardwell /* Track the maximum number of outstanding packets in each 1616ca8a2263SNeal Cardwell * window, and remember whether we were cwnd-limited then. 1617ca8a2263SNeal Cardwell */ 1618ca8a2263SNeal Cardwell if (!before(tp->snd_una, tp->max_packets_seq) || 1619ca8a2263SNeal Cardwell tp->packets_out > tp->max_packets_out) { 1620ca8a2263SNeal Cardwell tp->max_packets_out = tp->packets_out; 1621ca8a2263SNeal Cardwell tp->max_packets_seq = tp->snd_nxt; 1622ca8a2263SNeal Cardwell tp->is_cwnd_limited = is_cwnd_limited; 1623ca8a2263SNeal Cardwell } 1624e114a710SEric Dumazet 162524901551SEric Dumazet if (tcp_is_cwnd_limited(sk)) { 1626a762a980SDavid S. Miller /* Network is feed fully. */ 1627a762a980SDavid S. Miller tp->snd_cwnd_used = 0; 1628c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 1629a762a980SDavid S. Miller } else { 1630a762a980SDavid S. Miller /* Network starves. */ 1631a762a980SDavid S. Miller if (tp->packets_out > tp->snd_cwnd_used) 1632a762a980SDavid S. Miller tp->snd_cwnd_used = tp->packets_out; 1633a762a980SDavid S. Miller 163415d33c07SDavid S. Miller if (sysctl_tcp_slow_start_after_idle && 1635c2203cf7SEric Dumazet (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto && 16361b1fc3fdSWei Wang !ca_ops->cong_control) 1637a762a980SDavid S. Miller tcp_cwnd_application_limited(sk); 1638b0f71bd3SFrancis Yan 1639b0f71bd3SFrancis Yan /* The following conditions together indicate the starvation 1640b0f71bd3SFrancis Yan * is caused by insufficient sender buffer: 1641b0f71bd3SFrancis Yan * 1) just sent some data (see tcp_write_xmit) 1642b0f71bd3SFrancis Yan * 2) not cwnd limited (this else condition) 1643*75c119afSEric Dumazet * 3) no more data to send (tcp_write_queue_empty()) 1644b0f71bd3SFrancis Yan * 4) application is hitting buffer limit (SOCK_NOSPACE) 1645b0f71bd3SFrancis Yan */ 1646*75c119afSEric Dumazet if (tcp_write_queue_empty(sk) && sk->sk_socket && 1647b0f71bd3SFrancis Yan test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) && 1648b0f71bd3SFrancis Yan (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) 1649b0f71bd3SFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED); 1650a762a980SDavid S. Miller } 1651a762a980SDavid S. Miller } 1652a762a980SDavid S. Miller 1653d4589926SEric Dumazet /* Minshall's variant of the Nagle send check. */ 1654d4589926SEric Dumazet static bool tcp_minshall_check(const struct tcp_sock *tp) 1655d4589926SEric Dumazet { 1656d4589926SEric Dumazet return after(tp->snd_sml, tp->snd_una) && 1657d4589926SEric Dumazet !after(tp->snd_sml, tp->snd_nxt); 1658d4589926SEric Dumazet } 1659d4589926SEric Dumazet 1660d4589926SEric Dumazet /* Update snd_sml if this skb is under mss 1661d4589926SEric Dumazet * Note that a TSO packet might end with a sub-mss segment 1662d4589926SEric Dumazet * The test is really : 1663d4589926SEric Dumazet * if ((skb->len % mss) != 0) 1664d4589926SEric Dumazet * tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1665d4589926SEric Dumazet * But we can avoid doing the divide again given we already have 1666d4589926SEric Dumazet * skb_pcount = skb->len / mss_now 16670e3a4803SIlpo Järvinen */ 1668d4589926SEric Dumazet static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now, 1669d4589926SEric Dumazet const struct sk_buff *skb) 1670d4589926SEric Dumazet { 1671d4589926SEric Dumazet if (skb->len < tcp_skb_pcount(skb) * mss_now) 1672d4589926SEric Dumazet tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1673d4589926SEric Dumazet } 1674d4589926SEric Dumazet 1675d4589926SEric Dumazet /* Return false, if packet can be sent now without violation Nagle's rules: 1676d4589926SEric Dumazet * 1. It is full sized. (provided by caller in %partial bool) 1677d4589926SEric Dumazet * 2. Or it contains FIN. (already checked by caller) 1678d4589926SEric Dumazet * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. 1679d4589926SEric Dumazet * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 1680d4589926SEric Dumazet * With Minshall's modification: all sent small packets are ACKed. 1681d4589926SEric Dumazet */ 1682d4589926SEric Dumazet static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp, 1683cc93fc51SPeter Pan(潘卫平) int nonagle) 1684d4589926SEric Dumazet { 1685d4589926SEric Dumazet return partial && 1686d4589926SEric Dumazet ((nonagle & TCP_NAGLE_CORK) || 1687d4589926SEric Dumazet (!nonagle && tp->packets_out && tcp_minshall_check(tp))); 1688d4589926SEric Dumazet } 1689605ad7f1SEric Dumazet 1690605ad7f1SEric Dumazet /* Return how many segs we'd like on a TSO packet, 1691605ad7f1SEric Dumazet * to send one TSO packet per ms 1692605ad7f1SEric Dumazet */ 16931b3878caSNeal Cardwell u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now, 16941b3878caSNeal Cardwell int min_tso_segs) 1695605ad7f1SEric Dumazet { 1696605ad7f1SEric Dumazet u32 bytes, segs; 1697605ad7f1SEric Dumazet 1698605ad7f1SEric Dumazet bytes = min(sk->sk_pacing_rate >> 10, 1699605ad7f1SEric Dumazet sk->sk_gso_max_size - 1 - MAX_TCP_HEADER); 1700605ad7f1SEric Dumazet 1701605ad7f1SEric Dumazet /* Goal is to send at least one packet per ms, 1702605ad7f1SEric Dumazet * not one big TSO packet every 100 ms. 1703605ad7f1SEric Dumazet * This preserves ACK clocking and is consistent 1704605ad7f1SEric Dumazet * with tcp_tso_should_defer() heuristic. 1705605ad7f1SEric Dumazet */ 17061b3878caSNeal Cardwell segs = max_t(u32, bytes / mss_now, min_tso_segs); 1707605ad7f1SEric Dumazet 1708605ad7f1SEric Dumazet return min_t(u32, segs, sk->sk_gso_max_segs); 1709605ad7f1SEric Dumazet } 17101b3878caSNeal Cardwell EXPORT_SYMBOL(tcp_tso_autosize); 1711605ad7f1SEric Dumazet 1712ed6e7268SNeal Cardwell /* Return the number of segments we want in the skb we are transmitting. 1713ed6e7268SNeal Cardwell * See if congestion control module wants to decide; otherwise, autosize. 1714ed6e7268SNeal Cardwell */ 1715ed6e7268SNeal Cardwell static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now) 1716ed6e7268SNeal Cardwell { 1717ed6e7268SNeal Cardwell const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 1718ed6e7268SNeal Cardwell u32 tso_segs = ca_ops->tso_segs_goal ? ca_ops->tso_segs_goal(sk) : 0; 1719ed6e7268SNeal Cardwell 17201b3878caSNeal Cardwell return tso_segs ? : 17211b3878caSNeal Cardwell tcp_tso_autosize(sk, mss_now, sysctl_tcp_min_tso_segs); 1722ed6e7268SNeal Cardwell } 1723ed6e7268SNeal Cardwell 1724d4589926SEric Dumazet /* Returns the portion of skb which can be sent right away */ 1725d4589926SEric Dumazet static unsigned int tcp_mss_split_point(const struct sock *sk, 1726d4589926SEric Dumazet const struct sk_buff *skb, 1727d4589926SEric Dumazet unsigned int mss_now, 1728d4589926SEric Dumazet unsigned int max_segs, 1729d4589926SEric Dumazet int nonagle) 1730c1b4a7e6SDavid S. Miller { 1731cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1732d4589926SEric Dumazet u32 partial, needed, window, max_len; 1733c1b4a7e6SDavid S. Miller 173490840defSIlpo Järvinen window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 17351485348dSBen Hutchings max_len = mss_now * max_segs; 17360e3a4803SIlpo Järvinen 17371485348dSBen Hutchings if (likely(max_len <= window && skb != tcp_write_queue_tail(sk))) 17381485348dSBen Hutchings return max_len; 17390e3a4803SIlpo Järvinen 17405ea3a748SIlpo Järvinen needed = min(skb->len, window); 17415ea3a748SIlpo Järvinen 17421485348dSBen Hutchings if (max_len <= needed) 17431485348dSBen Hutchings return max_len; 17440e3a4803SIlpo Järvinen 1745d4589926SEric Dumazet partial = needed % mss_now; 1746d4589926SEric Dumazet /* If last segment is not a full MSS, check if Nagle rules allow us 1747d4589926SEric Dumazet * to include this last segment in this skb. 1748d4589926SEric Dumazet * Otherwise, we'll split the skb at last MSS boundary 1749d4589926SEric Dumazet */ 1750cc93fc51SPeter Pan(潘卫平) if (tcp_nagle_check(partial != 0, tp, nonagle)) 1751d4589926SEric Dumazet return needed - partial; 1752d4589926SEric Dumazet 1753d4589926SEric Dumazet return needed; 1754c1b4a7e6SDavid S. Miller } 1755c1b4a7e6SDavid S. Miller 1756c1b4a7e6SDavid S. Miller /* Can at least one segment of SKB be sent right now, according to the 1757c1b4a7e6SDavid S. Miller * congestion window rules? If so, return how many segments are allowed. 1758c1b4a7e6SDavid S. Miller */ 1759cf533ea5SEric Dumazet static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, 1760cf533ea5SEric Dumazet const struct sk_buff *skb) 1761c1b4a7e6SDavid S. Miller { 1762d649a7a8SEric Dumazet u32 in_flight, cwnd, halfcwnd; 1763c1b4a7e6SDavid S. Miller 1764c1b4a7e6SDavid S. Miller /* Don't be strict about the congestion window for the final FIN. */ 17654de075e0SEric Dumazet if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && 17664de075e0SEric Dumazet tcp_skb_pcount(skb) == 1) 1767c1b4a7e6SDavid S. Miller return 1; 1768c1b4a7e6SDavid S. Miller 1769c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1770c1b4a7e6SDavid S. Miller cwnd = tp->snd_cwnd; 1771d649a7a8SEric Dumazet if (in_flight >= cwnd) 1772c1b4a7e6SDavid S. Miller return 0; 1773d649a7a8SEric Dumazet 1774d649a7a8SEric Dumazet /* For better scheduling, ensure we have at least 1775d649a7a8SEric Dumazet * 2 GSO packets in flight. 1776d649a7a8SEric Dumazet */ 1777d649a7a8SEric Dumazet halfcwnd = max(cwnd >> 1, 1U); 1778d649a7a8SEric Dumazet return min(halfcwnd, cwnd - in_flight); 1779c1b4a7e6SDavid S. Miller } 1780c1b4a7e6SDavid S. Miller 1781b595076aSUwe Kleine-König /* Initialize TSO state of a skb. 178267edfef7SAndi Kleen * This must be invoked the first time we consider transmitting 1783c1b4a7e6SDavid S. Miller * SKB onto the wire. 1784c1b4a7e6SDavid S. Miller */ 17855bbb432cSEric Dumazet static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now) 1786c1b4a7e6SDavid S. Miller { 1787c1b4a7e6SDavid S. Miller int tso_segs = tcp_skb_pcount(skb); 1788c1b4a7e6SDavid S. Miller 1789f8269a49SIlpo Järvinen if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { 17905bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 1791c1b4a7e6SDavid S. Miller tso_segs = tcp_skb_pcount(skb); 1792c1b4a7e6SDavid S. Miller } 1793c1b4a7e6SDavid S. Miller return tso_segs; 1794c1b4a7e6SDavid S. Miller } 1795c1b4a7e6SDavid S. Miller 1796c1b4a7e6SDavid S. Miller 1797a2a385d6SEric Dumazet /* Return true if the Nagle test allows this packet to be 1798c1b4a7e6SDavid S. Miller * sent now. 1799c1b4a7e6SDavid S. Miller */ 1800a2a385d6SEric Dumazet static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, 1801c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1802c1b4a7e6SDavid S. Miller { 1803c1b4a7e6SDavid S. Miller /* Nagle rule does not apply to frames, which sit in the middle of the 1804c1b4a7e6SDavid S. Miller * write_queue (they have no chances to get new data). 1805c1b4a7e6SDavid S. Miller * 1806c1b4a7e6SDavid S. Miller * This is implemented in the callers, where they modify the 'nonagle' 1807c1b4a7e6SDavid S. Miller * argument based upon the location of SKB in the send queue. 1808c1b4a7e6SDavid S. Miller */ 1809c1b4a7e6SDavid S. Miller if (nonagle & TCP_NAGLE_PUSH) 1810a2a385d6SEric Dumazet return true; 1811c1b4a7e6SDavid S. Miller 18129b44190dSYuchung Cheng /* Don't use the nagle rule for urgent data (or for the final FIN). */ 18139b44190dSYuchung Cheng if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) 1814a2a385d6SEric Dumazet return true; 1815c1b4a7e6SDavid S. Miller 1816cc93fc51SPeter Pan(潘卫平) if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle)) 1817a2a385d6SEric Dumazet return true; 1818c1b4a7e6SDavid S. Miller 1819a2a385d6SEric Dumazet return false; 1820c1b4a7e6SDavid S. Miller } 1821c1b4a7e6SDavid S. Miller 1822c1b4a7e6SDavid S. Miller /* Does at least the first segment of SKB fit into the send window? */ 1823a2a385d6SEric Dumazet static bool tcp_snd_wnd_test(const struct tcp_sock *tp, 1824a2a385d6SEric Dumazet const struct sk_buff *skb, 1825056834d9SIlpo Järvinen unsigned int cur_mss) 1826c1b4a7e6SDavid S. Miller { 1827c1b4a7e6SDavid S. Miller u32 end_seq = TCP_SKB_CB(skb)->end_seq; 1828c1b4a7e6SDavid S. Miller 1829c1b4a7e6SDavid S. Miller if (skb->len > cur_mss) 1830c1b4a7e6SDavid S. Miller end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 1831c1b4a7e6SDavid S. Miller 183290840defSIlpo Järvinen return !after(end_seq, tcp_wnd_end(tp)); 1833c1b4a7e6SDavid S. Miller } 1834c1b4a7e6SDavid S. Miller 1835c1b4a7e6SDavid S. Miller /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 1836c1b4a7e6SDavid S. Miller * which is put after SKB on the list. It is very much like 1837c1b4a7e6SDavid S. Miller * tcp_fragment() except that it may make several kinds of assumptions 1838c1b4a7e6SDavid S. Miller * in order to speed up the splitting operation. In particular, we 1839c1b4a7e6SDavid S. Miller * know that all the data is in scatter-gather pages, and that the 1840c1b4a7e6SDavid S. Miller * packet has never been sent out before (and thus is not cloned). 1841c1b4a7e6SDavid S. Miller */ 1842*75c119afSEric Dumazet static int tso_fragment(struct sock *sk, enum tcp_queue tcp_queue, 1843*75c119afSEric Dumazet struct sk_buff *skb, unsigned int len, 1844c4ead4c5SEric Dumazet unsigned int mss_now, gfp_t gfp) 1845c1b4a7e6SDavid S. Miller { 1846c1b4a7e6SDavid S. Miller struct sk_buff *buff; 1847c1b4a7e6SDavid S. Miller int nlen = skb->len - len; 18489ce01461SIlpo Järvinen u8 flags; 1849c1b4a7e6SDavid S. Miller 1850c1b4a7e6SDavid S. Miller /* All of a TSO frame must be composed of paged data. */ 1851c8ac3774SHerbert Xu if (skb->len != skb->data_len) 1852*75c119afSEric Dumazet return tcp_fragment(sk, tcp_queue, skb, len, mss_now, gfp); 1853c1b4a7e6SDavid S. Miller 1854eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, gfp, true); 185551456b29SIan Morris if (unlikely(!buff)) 1856c1b4a7e6SDavid S. Miller return -ENOMEM; 1857c1b4a7e6SDavid S. Miller 18583ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 18593ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1860b60b49eaSHerbert Xu buff->truesize += nlen; 1861c1b4a7e6SDavid S. Miller skb->truesize -= nlen; 1862c1b4a7e6SDavid S. Miller 1863c1b4a7e6SDavid S. Miller /* Correct the sequence numbers. */ 1864c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 1865c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 1866c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 1867c1b4a7e6SDavid S. Miller 1868c1b4a7e6SDavid S. Miller /* PSH and FIN should only be set in the second packet. */ 18694de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 18704de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 18714de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1872c1b4a7e6SDavid S. Miller 1873c1b4a7e6SDavid S. Miller /* This packet was never sent out yet, so no SACK bits. */ 1874c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->sacked = 0; 1875c1b4a7e6SDavid S. Miller 1876a166140eSMartin KaFai Lau tcp_skb_fragment_eor(skb, buff); 1877a166140eSMartin KaFai Lau 187884fa7933SPatrick McHardy buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; 1879c1b4a7e6SDavid S. Miller skb_split(skb, buff, len); 1880490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 1881c1b4a7e6SDavid S. Miller 1882c1b4a7e6SDavid S. Miller /* Fix up tso_factor for both original and new SKB. */ 18835bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 18845bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 1885c1b4a7e6SDavid S. Miller 1886c1b4a7e6SDavid S. Miller /* Link BUFF into the send queue. */ 1887f4a775d1SEric Dumazet __skb_header_release(buff); 1888*75c119afSEric Dumazet tcp_insert_write_queue_after(skb, buff, sk, tcp_queue); 1889c1b4a7e6SDavid S. Miller 1890c1b4a7e6SDavid S. Miller return 0; 1891c1b4a7e6SDavid S. Miller } 1892c1b4a7e6SDavid S. Miller 1893c1b4a7e6SDavid S. Miller /* Try to defer sending, if possible, in order to minimize the amount 1894c1b4a7e6SDavid S. Miller * of TSO splitting we do. View it as a kind of TSO Nagle test. 1895c1b4a7e6SDavid S. Miller * 1896c1b4a7e6SDavid S. Miller * This algorithm is from John Heffner. 1897c1b4a7e6SDavid S. Miller */ 1898ca8a2263SNeal Cardwell static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb, 1899605ad7f1SEric Dumazet bool *is_cwnd_limited, u32 max_segs) 1900c1b4a7e6SDavid S. Miller { 19016687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 190250c8339eSEric Dumazet u32 age, send_win, cong_win, limit, in_flight; 190350c8339eSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 190450c8339eSEric Dumazet struct sk_buff *head; 1905ad9f4f50SEric Dumazet int win_divisor; 1906c1b4a7e6SDavid S. Miller 19074de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 1908ae8064acSJohn Heffner goto send_now; 1909c1b4a7e6SDavid S. Miller 191099d7662aSEric Dumazet if (icsk->icsk_ca_state >= TCP_CA_Recovery) 1911ae8064acSJohn Heffner goto send_now; 1912ae8064acSJohn Heffner 19135f852eb5SEric Dumazet /* Avoid bursty behavior by allowing defer 19145f852eb5SEric Dumazet * only if the last write was recent. 19155f852eb5SEric Dumazet */ 1916d635fbe2SEric Dumazet if ((s32)(tcp_jiffies32 - tp->lsndtime) > 0) 1917ae8064acSJohn Heffner goto send_now; 1918908a75c1SDavid S. Miller 1919c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1920c1b4a7e6SDavid S. Miller 1921056834d9SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight)); 1922c1b4a7e6SDavid S. Miller 192390840defSIlpo Järvinen send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1924c1b4a7e6SDavid S. Miller 1925c1b4a7e6SDavid S. Miller /* From in_flight test above, we know that cwnd > in_flight. */ 1926c1b4a7e6SDavid S. Miller cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; 1927c1b4a7e6SDavid S. Miller 1928c1b4a7e6SDavid S. Miller limit = min(send_win, cong_win); 1929c1b4a7e6SDavid S. Miller 1930ba244fe9SDavid S. Miller /* If a full-sized TSO skb can be sent, do it. */ 1931605ad7f1SEric Dumazet if (limit >= max_segs * tp->mss_cache) 1932ae8064acSJohn Heffner goto send_now; 1933ba244fe9SDavid S. Miller 193462ad2761SIlpo Järvinen /* Middle in queue won't get any more data, full sendable already? */ 193562ad2761SIlpo Järvinen if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) 193662ad2761SIlpo Järvinen goto send_now; 193762ad2761SIlpo Järvinen 1938ad9f4f50SEric Dumazet win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor); 1939ad9f4f50SEric Dumazet if (win_divisor) { 1940c1b4a7e6SDavid S. Miller u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); 1941c1b4a7e6SDavid S. Miller 1942c1b4a7e6SDavid S. Miller /* If at least some fraction of a window is available, 1943c1b4a7e6SDavid S. Miller * just use it. 1944c1b4a7e6SDavid S. Miller */ 1945ad9f4f50SEric Dumazet chunk /= win_divisor; 1946c1b4a7e6SDavid S. Miller if (limit >= chunk) 1947ae8064acSJohn Heffner goto send_now; 1948c1b4a7e6SDavid S. Miller } else { 1949c1b4a7e6SDavid S. Miller /* Different approach, try not to defer past a single 1950c1b4a7e6SDavid S. Miller * ACK. Receiver should ACK every other full sized 1951c1b4a7e6SDavid S. Miller * frame, so if we have space for more than 3 frames 1952c1b4a7e6SDavid S. Miller * then send now. 1953c1b4a7e6SDavid S. Miller */ 19546b5a5c0dSNeal Cardwell if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) 1955ae8064acSJohn Heffner goto send_now; 1956c1b4a7e6SDavid S. Miller } 1957c1b4a7e6SDavid S. Miller 1958*75c119afSEric Dumazet /* TODO : use tsorted_sent_queue ? */ 1959*75c119afSEric Dumazet head = tcp_rtx_queue_head(sk); 1960*75c119afSEric Dumazet if (!head) 1961*75c119afSEric Dumazet goto send_now; 19629a568de4SEric Dumazet age = tcp_stamp_us_delta(tp->tcp_mstamp, head->skb_mstamp); 196350c8339eSEric Dumazet /* If next ACK is likely to come too late (half srtt), do not defer */ 196450c8339eSEric Dumazet if (age < (tp->srtt_us >> 4)) 196550c8339eSEric Dumazet goto send_now; 196650c8339eSEric Dumazet 19675f852eb5SEric Dumazet /* Ok, it looks like it is advisable to defer. */ 1968ae8064acSJohn Heffner 1969d2e1339fSBendik Rønning Opstad if (cong_win < send_win && cong_win <= skb->len) 1970ca8a2263SNeal Cardwell *is_cwnd_limited = true; 1971ca8a2263SNeal Cardwell 1972a2a385d6SEric Dumazet return true; 1973ae8064acSJohn Heffner 1974ae8064acSJohn Heffner send_now: 1975a2a385d6SEric Dumazet return false; 1976c1b4a7e6SDavid S. Miller } 1977c1b4a7e6SDavid S. Miller 197805cbc0dbSFan Du static inline void tcp_mtu_check_reprobe(struct sock *sk) 197905cbc0dbSFan Du { 198005cbc0dbSFan Du struct inet_connection_sock *icsk = inet_csk(sk); 198105cbc0dbSFan Du struct tcp_sock *tp = tcp_sk(sk); 198205cbc0dbSFan Du struct net *net = sock_net(sk); 198305cbc0dbSFan Du u32 interval; 198405cbc0dbSFan Du s32 delta; 198505cbc0dbSFan Du 198605cbc0dbSFan Du interval = net->ipv4.sysctl_tcp_probe_interval; 1987c74df29aSEric Dumazet delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp; 198805cbc0dbSFan Du if (unlikely(delta >= interval * HZ)) { 198905cbc0dbSFan Du int mss = tcp_current_mss(sk); 199005cbc0dbSFan Du 199105cbc0dbSFan Du /* Update current search range */ 199205cbc0dbSFan Du icsk->icsk_mtup.probe_size = 0; 199305cbc0dbSFan Du icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + 199405cbc0dbSFan Du sizeof(struct tcphdr) + 199505cbc0dbSFan Du icsk->icsk_af_ops->net_header_len; 199605cbc0dbSFan Du icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss); 199705cbc0dbSFan Du 199805cbc0dbSFan Du /* Update probe time stamp */ 1999c74df29aSEric Dumazet icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; 200005cbc0dbSFan Du } 200105cbc0dbSFan Du } 200205cbc0dbSFan Du 20035d424d5aSJohn Heffner /* Create a new MTU probe if we are ready. 200467edfef7SAndi Kleen * MTU probe is regularly attempting to increase the path MTU by 200567edfef7SAndi Kleen * deliberately sending larger packets. This discovers routing 200667edfef7SAndi Kleen * changes resulting in larger path MTUs. 200767edfef7SAndi Kleen * 20085d424d5aSJohn Heffner * Returns 0 if we should wait to probe (no cwnd available), 20095d424d5aSJohn Heffner * 1 if a probe was sent, 2010056834d9SIlpo Järvinen * -1 otherwise 2011056834d9SIlpo Järvinen */ 20125d424d5aSJohn Heffner static int tcp_mtu_probe(struct sock *sk) 20135d424d5aSJohn Heffner { 20145d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 201512a59abcSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 20165d424d5aSJohn Heffner struct sk_buff *skb, *nskb, *next; 20176b58e0a5SFan Du struct net *net = sock_net(sk); 20185d424d5aSJohn Heffner int probe_size; 201991cc17c0SIlpo Järvinen int size_needed; 202012a59abcSEric Dumazet int copy, len; 20215d424d5aSJohn Heffner int mss_now; 20226b58e0a5SFan Du int interval; 20235d424d5aSJohn Heffner 20245d424d5aSJohn Heffner /* Not currently probing/verifying, 20255d424d5aSJohn Heffner * not in recovery, 20265d424d5aSJohn Heffner * have enough cwnd, and 202712a59abcSEric Dumazet * not SACKing (the variable headers throw things off) 202812a59abcSEric Dumazet */ 202912a59abcSEric Dumazet if (likely(!icsk->icsk_mtup.enabled || 20305d424d5aSJohn Heffner icsk->icsk_mtup.probe_size || 20315d424d5aSJohn Heffner inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 20325d424d5aSJohn Heffner tp->snd_cwnd < 11 || 203312a59abcSEric Dumazet tp->rx_opt.num_sacks || tp->rx_opt.dsack)) 20345d424d5aSJohn Heffner return -1; 20355d424d5aSJohn Heffner 20366b58e0a5SFan Du /* Use binary search for probe_size between tcp_mss_base, 20376b58e0a5SFan Du * and current mss_clamp. if (search_high - search_low) 20386b58e0a5SFan Du * smaller than a threshold, backoff from probing. 20396b58e0a5SFan Du */ 20400c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 20416b58e0a5SFan Du probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high + 20426b58e0a5SFan Du icsk->icsk_mtup.search_low) >> 1); 204391cc17c0SIlpo Järvinen size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; 20446b58e0a5SFan Du interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low; 204505cbc0dbSFan Du /* When misfortune happens, we are reprobing actively, 204605cbc0dbSFan Du * and then reprobe timer has expired. We stick with current 204705cbc0dbSFan Du * probing process by not resetting search range to its orignal. 204805cbc0dbSFan Du */ 20496b58e0a5SFan Du if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) || 205005cbc0dbSFan Du interval < net->ipv4.sysctl_tcp_probe_threshold) { 205105cbc0dbSFan Du /* Check whether enough time has elaplased for 205205cbc0dbSFan Du * another round of probing. 205305cbc0dbSFan Du */ 205405cbc0dbSFan Du tcp_mtu_check_reprobe(sk); 20555d424d5aSJohn Heffner return -1; 20565d424d5aSJohn Heffner } 20575d424d5aSJohn Heffner 20585d424d5aSJohn Heffner /* Have enough data in the send queue to probe? */ 20597f9c33e5SIlpo Järvinen if (tp->write_seq - tp->snd_nxt < size_needed) 20605d424d5aSJohn Heffner return -1; 20615d424d5aSJohn Heffner 206291cc17c0SIlpo Järvinen if (tp->snd_wnd < size_needed) 20635d424d5aSJohn Heffner return -1; 206490840defSIlpo Järvinen if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) 20655d424d5aSJohn Heffner return 0; 20665d424d5aSJohn Heffner 2067d67c58e9SIlpo Järvinen /* Do we need to wait to drain cwnd? With none in flight, don't stall */ 2068d67c58e9SIlpo Järvinen if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { 2069d67c58e9SIlpo Järvinen if (!tcp_packets_in_flight(tp)) 20705d424d5aSJohn Heffner return -1; 20715d424d5aSJohn Heffner else 20725d424d5aSJohn Heffner return 0; 20735d424d5aSJohn Heffner } 20745d424d5aSJohn Heffner 20755d424d5aSJohn Heffner /* We're allowed to probe. Build it now. */ 2076eb934478SEric Dumazet nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false); 207751456b29SIan Morris if (!nskb) 20785d424d5aSJohn Heffner return -1; 20793ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 20803ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 20815d424d5aSJohn Heffner 2082fe067e8aSDavid S. Miller skb = tcp_send_head(sk); 20835d424d5aSJohn Heffner 20845d424d5aSJohn Heffner TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 20855d424d5aSJohn Heffner TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 20864de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; 20875d424d5aSJohn Heffner TCP_SKB_CB(nskb)->sacked = 0; 20885d424d5aSJohn Heffner nskb->csum = 0; 208984fa7933SPatrick McHardy nskb->ip_summed = skb->ip_summed; 20905d424d5aSJohn Heffner 209150c4817eSIlpo Järvinen tcp_insert_write_queue_before(nskb, skb, sk); 209250c4817eSIlpo Järvinen 20935d424d5aSJohn Heffner len = 0; 2094234b6860SIlpo Järvinen tcp_for_write_queue_from_safe(skb, next, sk) { 20955d424d5aSJohn Heffner copy = min_t(int, skb->len, probe_size - len); 20962fe664f1SDouglas Caetano dos Santos if (nskb->ip_summed) { 20975d424d5aSJohn Heffner skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); 20982fe664f1SDouglas Caetano dos Santos } else { 20992fe664f1SDouglas Caetano dos Santos __wsum csum = skb_copy_and_csum_bits(skb, 0, 2100056834d9SIlpo Järvinen skb_put(nskb, copy), 21012fe664f1SDouglas Caetano dos Santos copy, 0); 21022fe664f1SDouglas Caetano dos Santos nskb->csum = csum_block_add(nskb->csum, csum, len); 21032fe664f1SDouglas Caetano dos Santos } 21045d424d5aSJohn Heffner 21055d424d5aSJohn Heffner if (skb->len <= copy) { 21065d424d5aSJohn Heffner /* We've eaten all the data from this skb. 21075d424d5aSJohn Heffner * Throw it away. */ 21084de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 2109fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 21103ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 21115d424d5aSJohn Heffner } else { 21124de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & 2113a3433f35SChangli Gao ~(TCPHDR_FIN|TCPHDR_PSH); 21145d424d5aSJohn Heffner if (!skb_shinfo(skb)->nr_frags) { 21155d424d5aSJohn Heffner skb_pull(skb, copy); 211684fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 2117056834d9SIlpo Järvinen skb->csum = csum_partial(skb->data, 2118056834d9SIlpo Järvinen skb->len, 0); 21195d424d5aSJohn Heffner } else { 21205d424d5aSJohn Heffner __pskb_trim_head(skb, copy); 21215bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 21225d424d5aSJohn Heffner } 21235d424d5aSJohn Heffner TCP_SKB_CB(skb)->seq += copy; 21245d424d5aSJohn Heffner } 21255d424d5aSJohn Heffner 21265d424d5aSJohn Heffner len += copy; 2127234b6860SIlpo Järvinen 2128234b6860SIlpo Järvinen if (len >= probe_size) 2129234b6860SIlpo Järvinen break; 21305d424d5aSJohn Heffner } 21315bbb432cSEric Dumazet tcp_init_tso_segs(nskb, nskb->len); 21325d424d5aSJohn Heffner 21335d424d5aSJohn Heffner /* We're ready to send. If this fails, the probe will 21347faee5c0SEric Dumazet * be resegmented into mss-sized pieces by tcp_write_xmit(). 21357faee5c0SEric Dumazet */ 21365d424d5aSJohn Heffner if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 21375d424d5aSJohn Heffner /* Decrement cwnd here because we are sending 21385d424d5aSJohn Heffner * effectively two packets. */ 21395d424d5aSJohn Heffner tp->snd_cwnd--; 214066f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, nskb); 21415d424d5aSJohn Heffner 21425d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 21430e7b1368SJohn Heffner tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 21440e7b1368SJohn Heffner tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 21455d424d5aSJohn Heffner 21465d424d5aSJohn Heffner return 1; 21475d424d5aSJohn Heffner } 21485d424d5aSJohn Heffner 21495d424d5aSJohn Heffner return -1; 21505d424d5aSJohn Heffner } 21515d424d5aSJohn Heffner 2152218af599SEric Dumazet static bool tcp_pacing_check(const struct sock *sk) 2153218af599SEric Dumazet { 2154218af599SEric Dumazet return tcp_needs_internal_pacing(sk) && 2155218af599SEric Dumazet hrtimer_active(&tcp_sk(sk)->pacing_timer); 2156218af599SEric Dumazet } 2157218af599SEric Dumazet 2158f9616c35SEric Dumazet /* TCP Small Queues : 2159f9616c35SEric Dumazet * Control number of packets in qdisc/devices to two packets / or ~1 ms. 2160f9616c35SEric Dumazet * (These limits are doubled for retransmits) 2161f9616c35SEric Dumazet * This allows for : 2162f9616c35SEric Dumazet * - better RTT estimation and ACK scheduling 2163f9616c35SEric Dumazet * - faster recovery 2164f9616c35SEric Dumazet * - high rates 2165f9616c35SEric Dumazet * Alas, some drivers / subsystems require a fair amount 2166f9616c35SEric Dumazet * of queued bytes to ensure line rate. 2167f9616c35SEric Dumazet * One example is wifi aggregation (802.11 AMPDU) 2168f9616c35SEric Dumazet */ 2169f9616c35SEric Dumazet static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb, 2170f9616c35SEric Dumazet unsigned int factor) 2171f9616c35SEric Dumazet { 2172f9616c35SEric Dumazet unsigned int limit; 2173f9616c35SEric Dumazet 2174f9616c35SEric Dumazet limit = max(2 * skb->truesize, sk->sk_pacing_rate >> 10); 2175f9616c35SEric Dumazet limit = min_t(u32, limit, sysctl_tcp_limit_output_bytes); 2176f9616c35SEric Dumazet limit <<= factor; 2177f9616c35SEric Dumazet 217814afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) > limit) { 2179*75c119afSEric Dumazet /* Always send skb if rtx queue is empty. 218075eefc6cSEric Dumazet * No need to wait for TX completion to call us back, 218175eefc6cSEric Dumazet * after softirq/tasklet schedule. 218275eefc6cSEric Dumazet * This helps when TX completions are delayed too much. 218375eefc6cSEric Dumazet */ 2184*75c119afSEric Dumazet if (tcp_rtx_queue_empty(sk)) 218575eefc6cSEric Dumazet return false; 218675eefc6cSEric Dumazet 21877aa5470cSEric Dumazet set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags); 2188f9616c35SEric Dumazet /* It is possible TX completion already happened 2189f9616c35SEric Dumazet * before we set TSQ_THROTTLED, so we must 2190f9616c35SEric Dumazet * test again the condition. 2191f9616c35SEric Dumazet */ 2192f9616c35SEric Dumazet smp_mb__after_atomic(); 219314afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) > limit) 2194f9616c35SEric Dumazet return true; 2195f9616c35SEric Dumazet } 2196f9616c35SEric Dumazet return false; 2197f9616c35SEric Dumazet } 2198f9616c35SEric Dumazet 219905b055e8SFrancis Yan static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new) 220005b055e8SFrancis Yan { 2201628174ccSEric Dumazet const u32 now = tcp_jiffies32; 2202efe967cdSArnd Bergmann enum tcp_chrono old = tp->chrono_type; 220305b055e8SFrancis Yan 2204efe967cdSArnd Bergmann if (old > TCP_CHRONO_UNSPEC) 2205efe967cdSArnd Bergmann tp->chrono_stat[old - 1] += now - tp->chrono_start; 220605b055e8SFrancis Yan tp->chrono_start = now; 220705b055e8SFrancis Yan tp->chrono_type = new; 220805b055e8SFrancis Yan } 220905b055e8SFrancis Yan 221005b055e8SFrancis Yan void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type) 221105b055e8SFrancis Yan { 221205b055e8SFrancis Yan struct tcp_sock *tp = tcp_sk(sk); 221305b055e8SFrancis Yan 221405b055e8SFrancis Yan /* If there are multiple conditions worthy of tracking in a 22150f87230dSFrancis Yan * chronograph then the highest priority enum takes precedence 22160f87230dSFrancis Yan * over the other conditions. So that if something "more interesting" 221705b055e8SFrancis Yan * starts happening, stop the previous chrono and start a new one. 221805b055e8SFrancis Yan */ 221905b055e8SFrancis Yan if (type > tp->chrono_type) 222005b055e8SFrancis Yan tcp_chrono_set(tp, type); 222105b055e8SFrancis Yan } 222205b055e8SFrancis Yan 222305b055e8SFrancis Yan void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type) 222405b055e8SFrancis Yan { 222505b055e8SFrancis Yan struct tcp_sock *tp = tcp_sk(sk); 222605b055e8SFrancis Yan 22270f87230dSFrancis Yan 22280f87230dSFrancis Yan /* There are multiple conditions worthy of tracking in a 22290f87230dSFrancis Yan * chronograph, so that the highest priority enum takes 22300f87230dSFrancis Yan * precedence over the other conditions (see tcp_chrono_start). 22310f87230dSFrancis Yan * If a condition stops, we only stop chrono tracking if 22320f87230dSFrancis Yan * it's the "most interesting" or current chrono we are 22330f87230dSFrancis Yan * tracking and starts busy chrono if we have pending data. 22340f87230dSFrancis Yan */ 2235*75c119afSEric Dumazet if (tcp_rtx_and_write_queues_empty(sk)) 223605b055e8SFrancis Yan tcp_chrono_set(tp, TCP_CHRONO_UNSPEC); 22370f87230dSFrancis Yan else if (type == tp->chrono_type) 22380f87230dSFrancis Yan tcp_chrono_set(tp, TCP_CHRONO_BUSY); 223905b055e8SFrancis Yan } 224005b055e8SFrancis Yan 22411da177e4SLinus Torvalds /* This routine writes packets to the network. It advances the 22421da177e4SLinus Torvalds * send_head. This happens as incoming acks open up the remote 22431da177e4SLinus Torvalds * window for us. 22441da177e4SLinus Torvalds * 2245f8269a49SIlpo Järvinen * LARGESEND note: !tcp_urg_mode is overkill, only frames between 2246f8269a49SIlpo Järvinen * snd_up-64k-mss .. snd_up cannot be large. However, taking into 2247f8269a49SIlpo Järvinen * account rare use of URG, this is not a big flaw. 2248f8269a49SIlpo Järvinen * 22496ba8a3b1SNandita Dukkipati * Send at most one packet when push_one > 0. Temporarily ignore 22506ba8a3b1SNandita Dukkipati * cwnd limit to force at most one packet out when push_one == 2. 22516ba8a3b1SNandita Dukkipati 2252a2a385d6SEric Dumazet * Returns true, if no segments are in flight and we have queued segments, 2253a2a385d6SEric Dumazet * but cannot send anything now because of SWS or another problem. 22541da177e4SLinus Torvalds */ 2255a2a385d6SEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 2256d5dd9175SIlpo Järvinen int push_one, gfp_t gfp) 22571da177e4SLinus Torvalds { 22581da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 225992df7b51SDavid S. Miller struct sk_buff *skb; 2260c1b4a7e6SDavid S. Miller unsigned int tso_segs, sent_pkts; 2261c1b4a7e6SDavid S. Miller int cwnd_quota; 22625d424d5aSJohn Heffner int result; 22635615f886SFrancis Yan bool is_cwnd_limited = false, is_rwnd_limited = false; 2264605ad7f1SEric Dumazet u32 max_segs; 22651da177e4SLinus Torvalds 2266c1b4a7e6SDavid S. Miller sent_pkts = 0; 22675d424d5aSJohn Heffner 2268d5dd9175SIlpo Järvinen if (!push_one) { 22695d424d5aSJohn Heffner /* Do MTU probing. */ 2270d5dd9175SIlpo Järvinen result = tcp_mtu_probe(sk); 2271d5dd9175SIlpo Järvinen if (!result) { 2272a2a385d6SEric Dumazet return false; 22735d424d5aSJohn Heffner } else if (result > 0) { 22745d424d5aSJohn Heffner sent_pkts = 1; 22755d424d5aSJohn Heffner } 2276d5dd9175SIlpo Järvinen } 22775d424d5aSJohn Heffner 2278ed6e7268SNeal Cardwell max_segs = tcp_tso_segs(sk, mss_now); 22799a568de4SEric Dumazet tcp_mstamp_refresh(tp); 2280fe067e8aSDavid S. Miller while ((skb = tcp_send_head(sk))) { 2281c8ac3774SHerbert Xu unsigned int limit; 2282c8ac3774SHerbert Xu 2283218af599SEric Dumazet if (tcp_pacing_check(sk)) 2284218af599SEric Dumazet break; 2285218af599SEric Dumazet 22865bbb432cSEric Dumazet tso_segs = tcp_init_tso_segs(skb, mss_now); 2287c1b4a7e6SDavid S. Miller BUG_ON(!tso_segs); 2288c1b4a7e6SDavid S. Miller 22899d186cacSAndrey Vagin if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) { 22907faee5c0SEric Dumazet /* "skb_mstamp" is used as a start point for the retransmit timer */ 2291e2080072SEric Dumazet tcp_update_skb_after_send(tp, skb); 2292ec342325SAndrew Vagin goto repair; /* Skip network transmission */ 22939d186cacSAndrey Vagin } 2294ec342325SAndrew Vagin 2295b68e9f85SHerbert Xu cwnd_quota = tcp_cwnd_test(tp, skb); 22966ba8a3b1SNandita Dukkipati if (!cwnd_quota) { 22976ba8a3b1SNandita Dukkipati if (push_one == 2) 22986ba8a3b1SNandita Dukkipati /* Force out a loss probe pkt. */ 22996ba8a3b1SNandita Dukkipati cwnd_quota = 1; 23006ba8a3b1SNandita Dukkipati else 2301b68e9f85SHerbert Xu break; 23026ba8a3b1SNandita Dukkipati } 2303b68e9f85SHerbert Xu 23045615f886SFrancis Yan if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) { 23055615f886SFrancis Yan is_rwnd_limited = true; 2306b68e9f85SHerbert Xu break; 23075615f886SFrancis Yan } 2308b68e9f85SHerbert Xu 2309d6a4e26aSEric Dumazet if (tso_segs == 1) { 2310aa93466bSDavid S. Miller if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 2311aa93466bSDavid S. Miller (tcp_skb_is_last(sk, skb) ? 2312aa93466bSDavid S. Miller nonagle : TCP_NAGLE_PUSH)))) 2313aa93466bSDavid S. Miller break; 2314c1b4a7e6SDavid S. Miller } else { 2315ca8a2263SNeal Cardwell if (!push_one && 2316605ad7f1SEric Dumazet tcp_tso_should_defer(sk, skb, &is_cwnd_limited, 2317605ad7f1SEric Dumazet max_segs)) 2318aa93466bSDavid S. Miller break; 2319c1b4a7e6SDavid S. Miller } 2320aa93466bSDavid S. Miller 2321605ad7f1SEric Dumazet limit = mss_now; 2322d6a4e26aSEric Dumazet if (tso_segs > 1 && !tcp_urg_mode(tp)) 2323605ad7f1SEric Dumazet limit = tcp_mss_split_point(sk, skb, mss_now, 2324605ad7f1SEric Dumazet min_t(unsigned int, 2325605ad7f1SEric Dumazet cwnd_quota, 2326605ad7f1SEric Dumazet max_segs), 2327605ad7f1SEric Dumazet nonagle); 2328605ad7f1SEric Dumazet 2329605ad7f1SEric Dumazet if (skb->len > limit && 2330*75c119afSEric Dumazet unlikely(tso_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE, 2331*75c119afSEric Dumazet skb, limit, mss_now, gfp))) 2332605ad7f1SEric Dumazet break; 2333605ad7f1SEric Dumazet 23347aa5470cSEric Dumazet if (test_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags)) 23357aa5470cSEric Dumazet clear_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags); 2336f9616c35SEric Dumazet if (tcp_small_queue_check(sk, skb, 0)) 233746d3ceabSEric Dumazet break; 2338c9eeec26SEric Dumazet 2339d5dd9175SIlpo Järvinen if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) 23401da177e4SLinus Torvalds break; 23411da177e4SLinus Torvalds 2342ec342325SAndrew Vagin repair: 23431da177e4SLinus Torvalds /* Advance the send_head. This one is sent out. 23441da177e4SLinus Torvalds * This call will increment packets_out. 23451da177e4SLinus Torvalds */ 234666f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 23471da177e4SLinus Torvalds 23481da177e4SLinus Torvalds tcp_minshall_update(tp, mss_now, skb); 2349a262f0cdSNandita Dukkipati sent_pkts += tcp_skb_pcount(skb); 2350d5dd9175SIlpo Järvinen 2351d5dd9175SIlpo Järvinen if (push_one) 2352d5dd9175SIlpo Järvinen break; 23531da177e4SLinus Torvalds } 23541da177e4SLinus Torvalds 23555615f886SFrancis Yan if (is_rwnd_limited) 23565615f886SFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED); 23575615f886SFrancis Yan else 23585615f886SFrancis Yan tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED); 23595615f886SFrancis Yan 2360aa93466bSDavid S. Miller if (likely(sent_pkts)) { 2361684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 2362684bad11SYuchung Cheng tp->prr_out += sent_pkts; 23636ba8a3b1SNandita Dukkipati 23646ba8a3b1SNandita Dukkipati /* Send one loss probe per tail loss episode. */ 23656ba8a3b1SNandita Dukkipati if (push_one != 2) 23666ba8a3b1SNandita Dukkipati tcp_schedule_loss_probe(sk); 2367d2e1339fSBendik Rønning Opstad is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd); 2368ca8a2263SNeal Cardwell tcp_cwnd_validate(sk, is_cwnd_limited); 2369a2a385d6SEric Dumazet return false; 23701da177e4SLinus Torvalds } 2371*75c119afSEric Dumazet return !tp->packets_out && !tcp_write_queue_empty(sk); 23726ba8a3b1SNandita Dukkipati } 23736ba8a3b1SNandita Dukkipati 23746ba8a3b1SNandita Dukkipati bool tcp_schedule_loss_probe(struct sock *sk) 23756ba8a3b1SNandita Dukkipati { 23766ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 23776ba8a3b1SNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 2378a2815817SNeal Cardwell u32 timeout, rto_delta_us; 23796ba8a3b1SNandita Dukkipati 23806ba8a3b1SNandita Dukkipati /* Don't do any loss probe on a Fast Open connection before 3WHS 23816ba8a3b1SNandita Dukkipati * finishes. 23826ba8a3b1SNandita Dukkipati */ 2383f9b99582SYuchung Cheng if (tp->fastopen_rsk) 23846ba8a3b1SNandita Dukkipati return false; 23856ba8a3b1SNandita Dukkipati 23866ba8a3b1SNandita Dukkipati /* Schedule a loss probe in 2*RTT for SACK capable connections 23876ba8a3b1SNandita Dukkipati * in Open state, that are either limited by cwnd or application. 23886ba8a3b1SNandita Dukkipati */ 2389bec41a11SYuchung Cheng if ((sysctl_tcp_early_retrans != 3 && sysctl_tcp_early_retrans != 4) || 2390bec41a11SYuchung Cheng !tp->packets_out || !tcp_is_sack(tp) || 2391bec41a11SYuchung Cheng icsk->icsk_ca_state != TCP_CA_Open) 23926ba8a3b1SNandita Dukkipati return false; 23936ba8a3b1SNandita Dukkipati 23946ba8a3b1SNandita Dukkipati if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) && 2395*75c119afSEric Dumazet !tcp_write_queue_empty(sk)) 23966ba8a3b1SNandita Dukkipati return false; 23976ba8a3b1SNandita Dukkipati 2398bb4d991aSYuchung Cheng /* Probe timeout is 2*rtt. Add minimum RTO to account 2399f9b99582SYuchung Cheng * for delayed ack when there's one outstanding packet. If no RTT 2400f9b99582SYuchung Cheng * sample is available then probe after TCP_TIMEOUT_INIT. 24016ba8a3b1SNandita Dukkipati */ 2402bb4d991aSYuchung Cheng if (tp->srtt_us) { 2403bb4d991aSYuchung Cheng timeout = usecs_to_jiffies(tp->srtt_us >> 2); 24046ba8a3b1SNandita Dukkipati if (tp->packets_out == 1) 2405bb4d991aSYuchung Cheng timeout += TCP_RTO_MIN; 2406bb4d991aSYuchung Cheng else 2407bb4d991aSYuchung Cheng timeout += TCP_TIMEOUT_MIN; 2408bb4d991aSYuchung Cheng } else { 2409bb4d991aSYuchung Cheng timeout = TCP_TIMEOUT_INIT; 2410bb4d991aSYuchung Cheng } 24116ba8a3b1SNandita Dukkipati 2412a2815817SNeal Cardwell /* If the RTO formula yields an earlier time, then use that time. */ 2413a2815817SNeal Cardwell rto_delta_us = tcp_rto_delta_us(sk); /* How far in future is RTO? */ 2414a2815817SNeal Cardwell if (rto_delta_us > 0) 2415a2815817SNeal Cardwell timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us)); 24166ba8a3b1SNandita Dukkipati 24176ba8a3b1SNandita Dukkipati inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, 24186ba8a3b1SNandita Dukkipati TCP_RTO_MAX); 24196ba8a3b1SNandita Dukkipati return true; 24206ba8a3b1SNandita Dukkipati } 24216ba8a3b1SNandita Dukkipati 24221f3279aeSEric Dumazet /* Thanks to skb fast clones, we can detect if a prior transmit of 24231f3279aeSEric Dumazet * a packet is still in a qdisc or driver queue. 24241f3279aeSEric Dumazet * In this case, there is very little point doing a retransmit ! 24251f3279aeSEric Dumazet */ 24261f3279aeSEric Dumazet static bool skb_still_in_host_queue(const struct sock *sk, 24271f3279aeSEric Dumazet const struct sk_buff *skb) 24281f3279aeSEric Dumazet { 242939bb5e62SEric Dumazet if (unlikely(skb_fclone_busy(sk, skb))) { 2430c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), 24311f3279aeSEric Dumazet LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES); 24321f3279aeSEric Dumazet return true; 24331f3279aeSEric Dumazet } 24341f3279aeSEric Dumazet return false; 24351f3279aeSEric Dumazet } 24361f3279aeSEric Dumazet 2437b340b264SYuchung Cheng /* When probe timeout (PTO) fires, try send a new segment if possible, else 24386ba8a3b1SNandita Dukkipati * retransmit the last segment. 24396ba8a3b1SNandita Dukkipati */ 24406ba8a3b1SNandita Dukkipati void tcp_send_loss_probe(struct sock *sk) 24416ba8a3b1SNandita Dukkipati { 24429b717a8dSNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 24436ba8a3b1SNandita Dukkipati struct sk_buff *skb; 24446ba8a3b1SNandita Dukkipati int pcount; 24456ba8a3b1SNandita Dukkipati int mss = tcp_current_mss(sk); 24466ba8a3b1SNandita Dukkipati 2447b340b264SYuchung Cheng skb = tcp_send_head(sk); 2448*75c119afSEric Dumazet if (skb && tcp_snd_wnd_test(tp, skb, mss)) { 2449b340b264SYuchung Cheng pcount = tp->packets_out; 2450b340b264SYuchung Cheng tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC); 2451b340b264SYuchung Cheng if (tp->packets_out > pcount) 2452b340b264SYuchung Cheng goto probe_sent; 24536ba8a3b1SNandita Dukkipati goto rearm_timer; 24546ba8a3b1SNandita Dukkipati } 2455*75c119afSEric Dumazet skb = skb_rb_last(&sk->tcp_rtx_queue); 24566ba8a3b1SNandita Dukkipati 24579b717a8dSNandita Dukkipati /* At most one outstanding TLP retransmission. */ 24589b717a8dSNandita Dukkipati if (tp->tlp_high_seq) 24599b717a8dSNandita Dukkipati goto rearm_timer; 24609b717a8dSNandita Dukkipati 24616ba8a3b1SNandita Dukkipati /* Retransmit last segment. */ 24626ba8a3b1SNandita Dukkipati if (WARN_ON(!skb)) 24636ba8a3b1SNandita Dukkipati goto rearm_timer; 24646ba8a3b1SNandita Dukkipati 24651f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 24661f3279aeSEric Dumazet goto rearm_timer; 24671f3279aeSEric Dumazet 24686ba8a3b1SNandita Dukkipati pcount = tcp_skb_pcount(skb); 24696ba8a3b1SNandita Dukkipati if (WARN_ON(!pcount)) 24706ba8a3b1SNandita Dukkipati goto rearm_timer; 24716ba8a3b1SNandita Dukkipati 24726ba8a3b1SNandita Dukkipati if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) { 2473*75c119afSEric Dumazet if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, 2474*75c119afSEric Dumazet (pcount - 1) * mss, mss, 24756cc55e09SOctavian Purdila GFP_ATOMIC))) 24766ba8a3b1SNandita Dukkipati goto rearm_timer; 2477*75c119afSEric Dumazet skb = skb_rb_next(skb); 24786ba8a3b1SNandita Dukkipati } 24796ba8a3b1SNandita Dukkipati 24806ba8a3b1SNandita Dukkipati if (WARN_ON(!skb || !tcp_skb_pcount(skb))) 24816ba8a3b1SNandita Dukkipati goto rearm_timer; 24826ba8a3b1SNandita Dukkipati 248310d3be56SEric Dumazet if (__tcp_retransmit_skb(sk, skb, 1)) 2484b340b264SYuchung Cheng goto rearm_timer; 24856ba8a3b1SNandita Dukkipati 24869b717a8dSNandita Dukkipati /* Record snd_nxt for loss detection. */ 24879b717a8dSNandita Dukkipati tp->tlp_high_seq = tp->snd_nxt; 24889b717a8dSNandita Dukkipati 2489b340b264SYuchung Cheng probe_sent: 2490c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES); 2491fcd16c0aSYuchung Cheng /* Reset s.t. tcp_rearm_rto will restart timer from now */ 2492fcd16c0aSYuchung Cheng inet_csk(sk)->icsk_pending = 0; 2493b340b264SYuchung Cheng rearm_timer: 2494fcd16c0aSYuchung Cheng tcp_rearm_rto(sk); 24951da177e4SLinus Torvalds } 24961da177e4SLinus Torvalds 2497a762a980SDavid S. Miller /* Push out any pending frames which were held back due to 2498a762a980SDavid S. Miller * TCP_CORK or attempt at coalescing tiny packets. 2499a762a980SDavid S. Miller * The socket must be locked by the caller. 2500a762a980SDavid S. Miller */ 25019e412ba7SIlpo Järvinen void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 25029e412ba7SIlpo Järvinen int nonagle) 2503a762a980SDavid S. Miller { 2504726e07a8SIlpo Järvinen /* If we are closed, the bytes will have to remain here. 2505726e07a8SIlpo Järvinen * In time closedown will finish, we empty the write queue and 2506726e07a8SIlpo Järvinen * all will be happy. 2507726e07a8SIlpo Järvinen */ 2508726e07a8SIlpo Järvinen if (unlikely(sk->sk_state == TCP_CLOSE)) 2509726e07a8SIlpo Järvinen return; 2510726e07a8SIlpo Järvinen 251199a1dec7SMel Gorman if (tcp_write_xmit(sk, cur_mss, nonagle, 0, 25127450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC))) 25139e412ba7SIlpo Järvinen tcp_check_probe_timer(sk); 2514a762a980SDavid S. Miller } 2515a762a980SDavid S. Miller 2516c1b4a7e6SDavid S. Miller /* Send _single_ skb sitting at the send head. This function requires 2517c1b4a7e6SDavid S. Miller * true push pending frames to setup probe timer etc. 2518c1b4a7e6SDavid S. Miller */ 2519c1b4a7e6SDavid S. Miller void tcp_push_one(struct sock *sk, unsigned int mss_now) 2520c1b4a7e6SDavid S. Miller { 2521fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 2522c1b4a7e6SDavid S. Miller 2523c1b4a7e6SDavid S. Miller BUG_ON(!skb || skb->len < mss_now); 2524c1b4a7e6SDavid S. Miller 2525d5dd9175SIlpo Järvinen tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); 2526c1b4a7e6SDavid S. Miller } 2527c1b4a7e6SDavid S. Miller 25281da177e4SLinus Torvalds /* This function returns the amount that we can raise the 25291da177e4SLinus Torvalds * usable window based on the following constraints 25301da177e4SLinus Torvalds * 25311da177e4SLinus Torvalds * 1. The window can never be shrunk once it is offered (RFC 793) 25321da177e4SLinus Torvalds * 2. We limit memory per socket 25331da177e4SLinus Torvalds * 25341da177e4SLinus Torvalds * RFC 1122: 25351da177e4SLinus Torvalds * "the suggested [SWS] avoidance algorithm for the receiver is to keep 25361da177e4SLinus Torvalds * RECV.NEXT + RCV.WIN fixed until: 25371da177e4SLinus Torvalds * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 25381da177e4SLinus Torvalds * 25391da177e4SLinus Torvalds * i.e. don't raise the right edge of the window until you can raise 25401da177e4SLinus Torvalds * it at least MSS bytes. 25411da177e4SLinus Torvalds * 25421da177e4SLinus Torvalds * Unfortunately, the recommended algorithm breaks header prediction, 25431da177e4SLinus Torvalds * since header prediction assumes th->window stays fixed. 25441da177e4SLinus Torvalds * 25451da177e4SLinus Torvalds * Strictly speaking, keeping th->window fixed violates the receiver 25461da177e4SLinus Torvalds * side SWS prevention criteria. The problem is that under this rule 25471da177e4SLinus Torvalds * a stream of single byte packets will cause the right side of the 25481da177e4SLinus Torvalds * window to always advance by a single byte. 25491da177e4SLinus Torvalds * 25501da177e4SLinus Torvalds * Of course, if the sender implements sender side SWS prevention 25511da177e4SLinus Torvalds * then this will not be a problem. 25521da177e4SLinus Torvalds * 25531da177e4SLinus Torvalds * BSD seems to make the following compromise: 25541da177e4SLinus Torvalds * 25551da177e4SLinus Torvalds * If the free space is less than the 1/4 of the maximum 25561da177e4SLinus Torvalds * space available and the free space is less than 1/2 mss, 25571da177e4SLinus Torvalds * then set the window to 0. 25581da177e4SLinus Torvalds * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 25591da177e4SLinus Torvalds * Otherwise, just prevent the window from shrinking 25601da177e4SLinus Torvalds * and from being larger than the largest representable value. 25611da177e4SLinus Torvalds * 25621da177e4SLinus Torvalds * This prevents incremental opening of the window in the regime 25631da177e4SLinus Torvalds * where TCP is limited by the speed of the reader side taking 25641da177e4SLinus Torvalds * data out of the TCP receive queue. It does nothing about 25651da177e4SLinus Torvalds * those cases where the window is constrained on the sender side 25661da177e4SLinus Torvalds * because the pipeline is full. 25671da177e4SLinus Torvalds * 25681da177e4SLinus Torvalds * BSD also seems to "accidentally" limit itself to windows that are a 25691da177e4SLinus Torvalds * multiple of MSS, at least until the free space gets quite small. 25701da177e4SLinus Torvalds * This would appear to be a side effect of the mbuf implementation. 25711da177e4SLinus Torvalds * Combining these two algorithms results in the observed behavior 25721da177e4SLinus Torvalds * of having a fixed window size at almost all times. 25731da177e4SLinus Torvalds * 25741da177e4SLinus Torvalds * Below we obtain similar behavior by forcing the offered window to 25751da177e4SLinus Torvalds * a multiple of the mss when it is feasible to do so. 25761da177e4SLinus Torvalds * 25771da177e4SLinus Torvalds * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 25781da177e4SLinus Torvalds * Regular options like TIMESTAMP are taken into account. 25791da177e4SLinus Torvalds */ 25801da177e4SLinus Torvalds u32 __tcp_select_window(struct sock *sk) 25811da177e4SLinus Torvalds { 2582463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 25831da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2584caa20d9aSStephen Hemminger /* MSS for the peer's data. Previous versions used mss_clamp 25851da177e4SLinus Torvalds * here. I don't know if the value based on our guesses 25861da177e4SLinus Torvalds * of peer's MSS is better for the performance. It's more correct 25871da177e4SLinus Torvalds * but may be worse for the performance because of rcv_mss 25881da177e4SLinus Torvalds * fluctuations. --SAW 1998/11/1 25891da177e4SLinus Torvalds */ 2590463c84b9SArnaldo Carvalho de Melo int mss = icsk->icsk_ack.rcv_mss; 25911da177e4SLinus Torvalds int free_space = tcp_space(sk); 259286c1a045SFlorian Westphal int allowed_space = tcp_full_space(sk); 259386c1a045SFlorian Westphal int full_space = min_t(int, tp->window_clamp, allowed_space); 25941da177e4SLinus Torvalds int window; 25951da177e4SLinus Torvalds 259606425c30SEric Dumazet if (unlikely(mss > full_space)) { 25971da177e4SLinus Torvalds mss = full_space; 259806425c30SEric Dumazet if (mss <= 0) 259906425c30SEric Dumazet return 0; 260006425c30SEric Dumazet } 2601b92edbe0SEric Dumazet if (free_space < (full_space >> 1)) { 2602463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.quick = 0; 26031da177e4SLinus Torvalds 2604b8da51ebSEric Dumazet if (tcp_under_memory_pressure(sk)) 2605056834d9SIlpo Järvinen tp->rcv_ssthresh = min(tp->rcv_ssthresh, 2606056834d9SIlpo Järvinen 4U * tp->advmss); 26071da177e4SLinus Torvalds 260886c1a045SFlorian Westphal /* free_space might become our new window, make sure we don't 260986c1a045SFlorian Westphal * increase it due to wscale. 261086c1a045SFlorian Westphal */ 261186c1a045SFlorian Westphal free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale); 261286c1a045SFlorian Westphal 261386c1a045SFlorian Westphal /* if free space is less than mss estimate, or is below 1/16th 261486c1a045SFlorian Westphal * of the maximum allowed, try to move to zero-window, else 261586c1a045SFlorian Westphal * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and 261686c1a045SFlorian Westphal * new incoming data is dropped due to memory limits. 261786c1a045SFlorian Westphal * With large window, mss test triggers way too late in order 261886c1a045SFlorian Westphal * to announce zero window in time before rmem limit kicks in. 261986c1a045SFlorian Westphal */ 262086c1a045SFlorian Westphal if (free_space < (allowed_space >> 4) || free_space < mss) 26211da177e4SLinus Torvalds return 0; 26221da177e4SLinus Torvalds } 26231da177e4SLinus Torvalds 26241da177e4SLinus Torvalds if (free_space > tp->rcv_ssthresh) 26251da177e4SLinus Torvalds free_space = tp->rcv_ssthresh; 26261da177e4SLinus Torvalds 26271da177e4SLinus Torvalds /* Don't do rounding if we are using window scaling, since the 26281da177e4SLinus Torvalds * scaled window will not line up with the MSS boundary anyway. 26291da177e4SLinus Torvalds */ 26301da177e4SLinus Torvalds if (tp->rx_opt.rcv_wscale) { 26311da177e4SLinus Torvalds window = free_space; 26321da177e4SLinus Torvalds 26331da177e4SLinus Torvalds /* Advertise enough space so that it won't get scaled away. 26341da177e4SLinus Torvalds * Import case: prevent zero window announcement if 26351da177e4SLinus Torvalds * 1<<rcv_wscale > mss. 26361da177e4SLinus Torvalds */ 26371935299dSGao Feng window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale)); 26381da177e4SLinus Torvalds } else { 26391935299dSGao Feng window = tp->rcv_wnd; 26401da177e4SLinus Torvalds /* Get the largest window that is a nice multiple of mss. 26411da177e4SLinus Torvalds * Window clamp already applied above. 26421da177e4SLinus Torvalds * If our current window offering is within 1 mss of the 26431da177e4SLinus Torvalds * free space we just keep it. This prevents the divide 26441da177e4SLinus Torvalds * and multiply from happening most of the time. 26451da177e4SLinus Torvalds * We also don't do any window rounding when the free space 26461da177e4SLinus Torvalds * is too small. 26471da177e4SLinus Torvalds */ 26481da177e4SLinus Torvalds if (window <= free_space - mss || window > free_space) 26491935299dSGao Feng window = rounddown(free_space, mss); 265084565070SJohn Heffner else if (mss == full_space && 2651b92edbe0SEric Dumazet free_space > window + (full_space >> 1)) 265284565070SJohn Heffner window = free_space; 26531da177e4SLinus Torvalds } 26541da177e4SLinus Torvalds 26551da177e4SLinus Torvalds return window; 26561da177e4SLinus Torvalds } 26571da177e4SLinus Torvalds 2658cfea5a68SMartin KaFai Lau void tcp_skb_collapse_tstamp(struct sk_buff *skb, 2659082ac2d5SMartin KaFai Lau const struct sk_buff *next_skb) 2660082ac2d5SMartin KaFai Lau { 26610a2cf20cSSoheil Hassas Yeganeh if (unlikely(tcp_has_tx_tstamp(next_skb))) { 26620a2cf20cSSoheil Hassas Yeganeh const struct skb_shared_info *next_shinfo = 26630a2cf20cSSoheil Hassas Yeganeh skb_shinfo(next_skb); 2664082ac2d5SMartin KaFai Lau struct skb_shared_info *shinfo = skb_shinfo(skb); 2665082ac2d5SMartin KaFai Lau 26660a2cf20cSSoheil Hassas Yeganeh shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP; 2667082ac2d5SMartin KaFai Lau shinfo->tskey = next_shinfo->tskey; 26682de8023eSMartin KaFai Lau TCP_SKB_CB(skb)->txstamp_ack |= 26692de8023eSMartin KaFai Lau TCP_SKB_CB(next_skb)->txstamp_ack; 2670082ac2d5SMartin KaFai Lau } 2671082ac2d5SMartin KaFai Lau } 2672082ac2d5SMartin KaFai Lau 26734a17fc3aSIlpo Järvinen /* Collapses two adjacent SKB's during retransmission. */ 2674f8071cdeSEric Dumazet static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) 26751da177e4SLinus Torvalds { 26761da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2677*75c119afSEric Dumazet struct sk_buff *next_skb = skb_rb_next(skb); 2678058dc334SIlpo Järvinen int skb_size, next_skb_size; 26791da177e4SLinus Torvalds 2680058dc334SIlpo Järvinen skb_size = skb->len; 2681058dc334SIlpo Järvinen next_skb_size = next_skb->len; 26821da177e4SLinus Torvalds 2683058dc334SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); 26841da177e4SLinus Torvalds 2685f8071cdeSEric Dumazet if (next_skb_size) { 2686f8071cdeSEric Dumazet if (next_skb_size <= skb_availroom(skb)) 2687f8071cdeSEric Dumazet skb_copy_bits(next_skb, 0, skb_put(skb, next_skb_size), 2688f8071cdeSEric Dumazet next_skb_size); 2689f8071cdeSEric Dumazet else if (!skb_shift(skb, next_skb, next_skb_size)) 2690f8071cdeSEric Dumazet return false; 2691f8071cdeSEric Dumazet } 26926859d494SIlpo Järvinen tcp_highest_sack_combine(sk, next_skb, skb); 2693a6963a6bSIlpo Järvinen 269452d570aaSJarek Poplawski if (next_skb->ip_summed == CHECKSUM_PARTIAL) 269552d570aaSJarek Poplawski skb->ip_summed = CHECKSUM_PARTIAL; 26961da177e4SLinus Torvalds 269784fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 26981da177e4SLinus Torvalds skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); 26991da177e4SLinus Torvalds 27001da177e4SLinus Torvalds /* Update sequence range on original skb. */ 27011da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 27021da177e4SLinus Torvalds 2703e6c7d085SIlpo Järvinen /* Merge over control information. This moves PSH/FIN etc. over */ 27044de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; 27051da177e4SLinus Torvalds 27061da177e4SLinus Torvalds /* All done, get rid of second SKB and account for it so 27071da177e4SLinus Torvalds * packet counting does not break. 27081da177e4SLinus Torvalds */ 27094828e7f4SIlpo Järvinen TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; 2710a643b5d4SMartin KaFai Lau TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor; 2711b7689205SIlpo Järvinen 2712b7689205SIlpo Järvinen /* changed transmit queue under us so clear hints */ 2713ef9da47cSIlpo Järvinen tcp_clear_retrans_hints_partial(tp); 2714ef9da47cSIlpo Järvinen if (next_skb == tp->retransmit_skb_hint) 2715ef9da47cSIlpo Järvinen tp->retransmit_skb_hint = skb; 2716b7689205SIlpo Järvinen 2717797108d1SIlpo Järvinen tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); 2718797108d1SIlpo Järvinen 2719082ac2d5SMartin KaFai Lau tcp_skb_collapse_tstamp(skb, next_skb); 2720082ac2d5SMartin KaFai Lau 2721*75c119afSEric Dumazet tcp_rtx_queue_unlink_and_free(next_skb, sk); 2722f8071cdeSEric Dumazet return true; 27231da177e4SLinus Torvalds } 27241da177e4SLinus Torvalds 272567edfef7SAndi Kleen /* Check if coalescing SKBs is legal. */ 2726a2a385d6SEric Dumazet static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) 27274a17fc3aSIlpo Järvinen { 27284a17fc3aSIlpo Järvinen if (tcp_skb_pcount(skb) > 1) 2729a2a385d6SEric Dumazet return false; 27304a17fc3aSIlpo Järvinen if (skb_cloned(skb)) 2731a2a385d6SEric Dumazet return false; 27322331ccc5SEric Dumazet /* Some heuristics for collapsing over SACK'd could be invented */ 27334a17fc3aSIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 2734a2a385d6SEric Dumazet return false; 27354a17fc3aSIlpo Järvinen 2736a2a385d6SEric Dumazet return true; 27374a17fc3aSIlpo Järvinen } 27384a17fc3aSIlpo Järvinen 273967edfef7SAndi Kleen /* Collapse packets in the retransmit queue to make to create 274067edfef7SAndi Kleen * less packets on the wire. This is only done on retransmission. 274167edfef7SAndi Kleen */ 27424a17fc3aSIlpo Järvinen static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, 27434a17fc3aSIlpo Järvinen int space) 27444a17fc3aSIlpo Järvinen { 27454a17fc3aSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 27464a17fc3aSIlpo Järvinen struct sk_buff *skb = to, *tmp; 2747a2a385d6SEric Dumazet bool first = true; 27484a17fc3aSIlpo Järvinen 27494a17fc3aSIlpo Järvinen if (!sysctl_tcp_retrans_collapse) 27504a17fc3aSIlpo Järvinen return; 27514de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 27524a17fc3aSIlpo Järvinen return; 27534a17fc3aSIlpo Järvinen 2754*75c119afSEric Dumazet skb_rbtree_walk_from_safe(skb, tmp) { 27554a17fc3aSIlpo Järvinen if (!tcp_can_collapse(sk, skb)) 27564a17fc3aSIlpo Järvinen break; 27574a17fc3aSIlpo Järvinen 2758a643b5d4SMartin KaFai Lau if (!tcp_skb_can_collapse_to(to)) 2759a643b5d4SMartin KaFai Lau break; 2760a643b5d4SMartin KaFai Lau 27614a17fc3aSIlpo Järvinen space -= skb->len; 27624a17fc3aSIlpo Järvinen 27634a17fc3aSIlpo Järvinen if (first) { 2764a2a385d6SEric Dumazet first = false; 27654a17fc3aSIlpo Järvinen continue; 27664a17fc3aSIlpo Järvinen } 27674a17fc3aSIlpo Järvinen 27684a17fc3aSIlpo Järvinen if (space < 0) 27694a17fc3aSIlpo Järvinen break; 27704a17fc3aSIlpo Järvinen 27714a17fc3aSIlpo Järvinen if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) 27724a17fc3aSIlpo Järvinen break; 27734a17fc3aSIlpo Järvinen 2774f8071cdeSEric Dumazet if (!tcp_collapse_retrans(sk, to)) 2775f8071cdeSEric Dumazet break; 27764a17fc3aSIlpo Järvinen } 27774a17fc3aSIlpo Järvinen } 27784a17fc3aSIlpo Järvinen 27791da177e4SLinus Torvalds /* This retransmits one SKB. Policy decisions and retransmit queue 27801da177e4SLinus Torvalds * state updates are done by the caller. Returns non-zero if an 27811da177e4SLinus Torvalds * error occurred which prevented the send. 27821da177e4SLinus Torvalds */ 278310d3be56SEric Dumazet int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) 27841da177e4SLinus Torvalds { 27855d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 278610d3be56SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 27877d227cd2SSridhar Samudrala unsigned int cur_mss; 278810d3be56SEric Dumazet int diff, len, err; 27891da177e4SLinus Torvalds 279010d3be56SEric Dumazet 279110d3be56SEric Dumazet /* Inconclusive MTU probe */ 279210d3be56SEric Dumazet if (icsk->icsk_mtup.probe_size) 27935d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 27945d424d5aSJohn Heffner 27951da177e4SLinus Torvalds /* Do not sent more than we queued. 1/4 is reserved for possible 2796caa20d9aSStephen Hemminger * copying overhead: fragmentation, tunneling, mangling etc. 27971da177e4SLinus Torvalds */ 279814afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) > 2799ffb4d6c8SEric Dumazet min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), 2800ffb4d6c8SEric Dumazet sk->sk_sndbuf)) 28011da177e4SLinus Torvalds return -EAGAIN; 28021da177e4SLinus Torvalds 28031f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 28041f3279aeSEric Dumazet return -EBUSY; 28051f3279aeSEric Dumazet 28061da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 28071da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 28081da177e4SLinus Torvalds BUG(); 28091da177e4SLinus Torvalds if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 28101da177e4SLinus Torvalds return -ENOMEM; 28111da177e4SLinus Torvalds } 28121da177e4SLinus Torvalds 28137d227cd2SSridhar Samudrala if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 28147d227cd2SSridhar Samudrala return -EHOSTUNREACH; /* Routing failure or similar. */ 28157d227cd2SSridhar Samudrala 28160c54b85fSIlpo Järvinen cur_mss = tcp_current_mss(sk); 28177d227cd2SSridhar Samudrala 28181da177e4SLinus Torvalds /* If receiver has shrunk his window, and skb is out of 28191da177e4SLinus Torvalds * new window, do not retransmit it. The exception is the 28201da177e4SLinus Torvalds * case, when window is shrunk to zero. In this case 28211da177e4SLinus Torvalds * our retransmit serves as a zero window probe. 28221da177e4SLinus Torvalds */ 28239d4fb27dSJoe Perches if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && 28249d4fb27dSJoe Perches TCP_SKB_CB(skb)->seq != tp->snd_una) 28251da177e4SLinus Torvalds return -EAGAIN; 28261da177e4SLinus Torvalds 282710d3be56SEric Dumazet len = cur_mss * segs; 282810d3be56SEric Dumazet if (skb->len > len) { 2829*75c119afSEric Dumazet if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len, 2830*75c119afSEric Dumazet cur_mss, GFP_ATOMIC)) 28311da177e4SLinus Torvalds return -ENOMEM; /* We'll try again later. */ 283202276f3cSIlpo Järvinen } else { 2833c52e2421SEric Dumazet if (skb_unclone(skb, GFP_ATOMIC)) 2834c52e2421SEric Dumazet return -ENOMEM; 283510d3be56SEric Dumazet 283610d3be56SEric Dumazet diff = tcp_skb_pcount(skb); 283710d3be56SEric Dumazet tcp_set_skb_tso_segs(skb, cur_mss); 283810d3be56SEric Dumazet diff -= tcp_skb_pcount(skb); 283910d3be56SEric Dumazet if (diff) 284010d3be56SEric Dumazet tcp_adjust_pcount(sk, skb, diff); 284110d3be56SEric Dumazet if (skb->len < cur_mss) 284210d3be56SEric Dumazet tcp_retrans_try_collapse(sk, skb, cur_mss); 28431da177e4SLinus Torvalds } 28441da177e4SLinus Torvalds 284549213555SDaniel Borkmann /* RFC3168, section 6.1.1.1. ECN fallback */ 284649213555SDaniel Borkmann if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN) 284749213555SDaniel Borkmann tcp_ecn_clear_syn(sk, skb); 284849213555SDaniel Borkmann 2849678550c6SYuchung Cheng /* Update global and local TCP statistics. */ 2850678550c6SYuchung Cheng segs = tcp_skb_pcount(skb); 2851678550c6SYuchung Cheng TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs); 2852678550c6SYuchung Cheng if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 2853678550c6SYuchung Cheng __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 2854678550c6SYuchung Cheng tp->total_retrans += segs; 2855678550c6SYuchung Cheng 285650bceae9SThomas Graf /* make sure skb->data is aligned on arches that require it 285750bceae9SThomas Graf * and check if ack-trimming & collapsing extended the headroom 285850bceae9SThomas Graf * beyond what csum_start can cover. 285950bceae9SThomas Graf */ 286050bceae9SThomas Graf if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) || 286150bceae9SThomas Graf skb_headroom(skb) >= 0xFFFF)) { 286210a81980SEric Dumazet struct sk_buff *nskb; 286310a81980SEric Dumazet 2864e2080072SEric Dumazet tcp_skb_tsorted_save(skb) { 286510a81980SEric Dumazet nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC); 2866c84a5711SYuchung Cheng err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) : 2867117632e6SEric Dumazet -ENOBUFS; 2868e2080072SEric Dumazet } tcp_skb_tsorted_restore(skb); 2869e2080072SEric Dumazet 28708c72c65bSEric Dumazet if (!err) 2871e2080072SEric Dumazet tcp_update_skb_after_send(tp, skb); 2872117632e6SEric Dumazet } else { 2873c84a5711SYuchung Cheng err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 2874117632e6SEric Dumazet } 2875c84a5711SYuchung Cheng 2876fc9f3501SEric Dumazet if (likely(!err)) { 2877c84a5711SYuchung Cheng TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS; 2878678550c6SYuchung Cheng } else if (err != -EBUSY) { 2879678550c6SYuchung Cheng NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL); 2880fc9f3501SEric Dumazet } 2881c84a5711SYuchung Cheng return err; 288293b174adSYuchung Cheng } 288393b174adSYuchung Cheng 288410d3be56SEric Dumazet int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) 288593b174adSYuchung Cheng { 288693b174adSYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 288710d3be56SEric Dumazet int err = __tcp_retransmit_skb(sk, skb, segs); 28881da177e4SLinus Torvalds 28891da177e4SLinus Torvalds if (err == 0) { 28901da177e4SLinus Torvalds #if FASTRETRANS_DEBUG > 0 28911da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 2892e87cc472SJoe Perches net_dbg_ratelimited("retrans_out leaked\n"); 28931da177e4SLinus Torvalds } 28941da177e4SLinus Torvalds #endif 28951da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 28961da177e4SLinus Torvalds tp->retrans_out += tcp_skb_pcount(skb); 28971da177e4SLinus Torvalds 28981da177e4SLinus Torvalds /* Save stamp of the first retransmit. */ 28991da177e4SLinus Torvalds if (!tp->retrans_stamp) 29007faee5c0SEric Dumazet tp->retrans_stamp = tcp_skb_timestamp(skb); 29011da177e4SLinus Torvalds 29021da177e4SLinus Torvalds } 29036e08d5e3SYuchung Cheng 29046e08d5e3SYuchung Cheng if (tp->undo_retrans < 0) 29056e08d5e3SYuchung Cheng tp->undo_retrans = 0; 29066e08d5e3SYuchung Cheng tp->undo_retrans += tcp_skb_pcount(skb); 29071da177e4SLinus Torvalds return err; 29081da177e4SLinus Torvalds } 29091da177e4SLinus Torvalds 29101da177e4SLinus Torvalds /* This gets called after a retransmit timeout, and the initially 29111da177e4SLinus Torvalds * retransmitted data is acknowledged. It tries to continue 29121da177e4SLinus Torvalds * resending the rest of the retransmit queue, until either 29131da177e4SLinus Torvalds * we've sent it all or the congestion window limit is reached. 29141da177e4SLinus Torvalds * If doing SACK, the first ACK which comes back for a timeout 29151da177e4SLinus Torvalds * based retransmit packet might feed us FACK information again. 29161da177e4SLinus Torvalds * If so, we use it to avoid unnecessarily retransmissions. 29171da177e4SLinus Torvalds */ 29181da177e4SLinus Torvalds void tcp_xmit_retransmit_queue(struct sock *sk) 29191da177e4SLinus Torvalds { 29206687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 2921*75c119afSEric Dumazet struct sk_buff *skb, *rtx_head = NULL, *hole = NULL; 29221da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2923840a3cbeSYuchung Cheng u32 max_segs; 292461eb55f4SIlpo Järvinen int mib_idx; 29256a438bbeSStephen Hemminger 292645e77d31SIlpo Järvinen if (!tp->packets_out) 292745e77d31SIlpo Järvinen return; 292845e77d31SIlpo Järvinen 29296a438bbeSStephen Hemminger skb = tp->retransmit_skb_hint; 2930*75c119afSEric Dumazet if (!skb) { 2931*75c119afSEric Dumazet rtx_head = tcp_rtx_queue_head(sk); 2932*75c119afSEric Dumazet skb = rtx_head; 2933618d9f25SIlpo Järvinen } 2934ed6e7268SNeal Cardwell max_segs = tcp_tso_segs(sk, tcp_current_mss(sk)); 2935*75c119afSEric Dumazet skb_rbtree_walk_from(skb) { 2936dca0aaf8SEric Dumazet __u8 sacked; 293710d3be56SEric Dumazet int segs; 29381da177e4SLinus Torvalds 2939218af599SEric Dumazet if (tcp_pacing_check(sk)) 2940218af599SEric Dumazet break; 2941218af599SEric Dumazet 29426a438bbeSStephen Hemminger /* we could do better than to assign each time */ 294351456b29SIan Morris if (!hole) 29446a438bbeSStephen Hemminger tp->retransmit_skb_hint = skb; 29456a438bbeSStephen Hemminger 294610d3be56SEric Dumazet segs = tp->snd_cwnd - tcp_packets_in_flight(tp); 294710d3be56SEric Dumazet if (segs <= 0) 29481da177e4SLinus Torvalds return; 2949dca0aaf8SEric Dumazet sacked = TCP_SKB_CB(skb)->sacked; 2950a3d2e9f8SEric Dumazet /* In case tcp_shift_skb_data() have aggregated large skbs, 2951a3d2e9f8SEric Dumazet * we need to make sure not sending too bigs TSO packets 2952a3d2e9f8SEric Dumazet */ 2953a3d2e9f8SEric Dumazet segs = min_t(int, segs, max_segs); 29540e1c54c2SIlpo Järvinen 2955840a3cbeSYuchung Cheng if (tp->retrans_out >= tp->lost_out) { 2956006f582cSIlpo Järvinen break; 29570e1c54c2SIlpo Järvinen } else if (!(sacked & TCPCB_LOST)) { 295851456b29SIan Morris if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) 29590e1c54c2SIlpo Järvinen hole = skb; 296061eb55f4SIlpo Järvinen continue; 29611da177e4SLinus Torvalds 29620e1c54c2SIlpo Järvinen } else { 29630e1c54c2SIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Loss) 29640e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFASTRETRANS; 29650e1c54c2SIlpo Järvinen else 29660e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; 29670e1c54c2SIlpo Järvinen } 29680e1c54c2SIlpo Järvinen 29690e1c54c2SIlpo Järvinen if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) 297061eb55f4SIlpo Järvinen continue; 297140b215e5SPavel Emelyanov 2972f9616c35SEric Dumazet if (tcp_small_queue_check(sk, skb, 1)) 2973f9616c35SEric Dumazet return; 2974f9616c35SEric Dumazet 297510d3be56SEric Dumazet if (tcp_retransmit_skb(sk, skb, segs)) 29761da177e4SLinus Torvalds return; 297724ab6becSYuchung Cheng 2978de1d6578SYuchung Cheng NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb)); 29791da177e4SLinus Torvalds 2980684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 2981a262f0cdSNandita Dukkipati tp->prr_out += tcp_skb_pcount(skb); 2982a262f0cdSNandita Dukkipati 2983*75c119afSEric Dumazet if (skb == rtx_head && 298457dde7f7SYuchung Cheng icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT) 2985463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 29863f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, 29873f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 29881da177e4SLinus Torvalds } 29891da177e4SLinus Torvalds } 29901da177e4SLinus Torvalds 2991d83769a5SEric Dumazet /* We allow to exceed memory limits for FIN packets to expedite 2992d83769a5SEric Dumazet * connection tear down and (memory) recovery. 2993845704a5SEric Dumazet * Otherwise tcp_send_fin() could be tempted to either delay FIN 2994845704a5SEric Dumazet * or even be forced to close flow without any FIN. 2995a6c5ea4cSEric Dumazet * In general, we want to allow one skb per socket to avoid hangs 2996a6c5ea4cSEric Dumazet * with edge trigger epoll() 2997d83769a5SEric Dumazet */ 2998a6c5ea4cSEric Dumazet void sk_forced_mem_schedule(struct sock *sk, int size) 2999d83769a5SEric Dumazet { 3000e805605cSJohannes Weiner int amt; 3001d83769a5SEric Dumazet 3002d83769a5SEric Dumazet if (size <= sk->sk_forward_alloc) 3003d83769a5SEric Dumazet return; 3004d83769a5SEric Dumazet amt = sk_mem_pages(size); 3005d83769a5SEric Dumazet sk->sk_forward_alloc += amt * SK_MEM_QUANTUM; 3006e805605cSJohannes Weiner sk_memory_allocated_add(sk, amt); 3007e805605cSJohannes Weiner 3008baac50bbSJohannes Weiner if (mem_cgroup_sockets_enabled && sk->sk_memcg) 3009baac50bbSJohannes Weiner mem_cgroup_charge_skmem(sk->sk_memcg, amt); 3010d83769a5SEric Dumazet } 3011d83769a5SEric Dumazet 3012845704a5SEric Dumazet /* Send a FIN. The caller locks the socket for us. 3013845704a5SEric Dumazet * We should try to send a FIN packet really hard, but eventually give up. 30141da177e4SLinus Torvalds */ 30151da177e4SLinus Torvalds void tcp_send_fin(struct sock *sk) 30161da177e4SLinus Torvalds { 3017845704a5SEric Dumazet struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk); 30181da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 30191da177e4SLinus Torvalds 3020845704a5SEric Dumazet /* Optimization, tack on the FIN if we have one skb in write queue and 3021845704a5SEric Dumazet * this skb was not yet sent, or we are under memory pressure. 3022845704a5SEric Dumazet * Note: in the latter case, FIN packet will be sent after a timeout, 3023845704a5SEric Dumazet * as TCP stack thinks it has already been transmitted. 30241da177e4SLinus Torvalds */ 3025*75c119afSEric Dumazet if (!tskb && tcp_under_memory_pressure(sk)) 3026*75c119afSEric Dumazet tskb = skb_rb_last(&sk->tcp_rtx_queue); 3027*75c119afSEric Dumazet 3028*75c119afSEric Dumazet if (tskb) { 3029845704a5SEric Dumazet coalesce: 3030845704a5SEric Dumazet TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN; 3031845704a5SEric Dumazet TCP_SKB_CB(tskb)->end_seq++; 30321da177e4SLinus Torvalds tp->write_seq++; 3033*75c119afSEric Dumazet if (tcp_write_queue_empty(sk)) { 3034845704a5SEric Dumazet /* This means tskb was already sent. 3035845704a5SEric Dumazet * Pretend we included the FIN on previous transmit. 3036845704a5SEric Dumazet * We need to set tp->snd_nxt to the value it would have 3037845704a5SEric Dumazet * if FIN had been sent. This is because retransmit path 3038845704a5SEric Dumazet * does not change tp->snd_nxt. 3039845704a5SEric Dumazet */ 3040845704a5SEric Dumazet tp->snd_nxt++; 3041845704a5SEric Dumazet return; 3042845704a5SEric Dumazet } 30431da177e4SLinus Torvalds } else { 3044845704a5SEric Dumazet skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation); 3045845704a5SEric Dumazet if (unlikely(!skb)) { 3046845704a5SEric Dumazet if (tskb) 3047845704a5SEric Dumazet goto coalesce; 3048845704a5SEric Dumazet return; 30491da177e4SLinus Torvalds } 3050e2080072SEric Dumazet INIT_LIST_HEAD(&skb->tcp_tsorted_anchor); 3051d83769a5SEric Dumazet skb_reserve(skb, MAX_TCP_HEADER); 3052a6c5ea4cSEric Dumazet sk_forced_mem_schedule(sk, skb->truesize); 30531da177e4SLinus Torvalds /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 3054e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tp->write_seq, 3055a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_FIN); 30561da177e4SLinus Torvalds tcp_queue_skb(sk, skb); 30571da177e4SLinus Torvalds } 3058845704a5SEric Dumazet __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF); 30591da177e4SLinus Torvalds } 30601da177e4SLinus Torvalds 30611da177e4SLinus Torvalds /* We get here when a process closes a file descriptor (either due to 30621da177e4SLinus Torvalds * an explicit close() or as a byproduct of exit()'ing) and there 30631da177e4SLinus Torvalds * was unread data in the receive queue. This behavior is recommended 306465bb723cSGerrit Renker * by RFC 2525, section 2.17. -DaveM 30651da177e4SLinus Torvalds */ 3066dd0fc66fSAl Viro void tcp_send_active_reset(struct sock *sk, gfp_t priority) 30671da177e4SLinus Torvalds { 30681da177e4SLinus Torvalds struct sk_buff *skb; 30691da177e4SLinus Torvalds 30707cc2b043SGao Feng TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); 30717cc2b043SGao Feng 30721da177e4SLinus Torvalds /* NOTE: No TCP options attached and we never retransmit this. */ 30731da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, priority); 30741da177e4SLinus Torvalds if (!skb) { 30754e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 30761da177e4SLinus Torvalds return; 30771da177e4SLinus Torvalds } 30781da177e4SLinus Torvalds 30791da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 30801da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 3081e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), 3082a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_RST); 30839a568de4SEric Dumazet tcp_mstamp_refresh(tcp_sk(sk)); 30841da177e4SLinus Torvalds /* Send it off. */ 3085dfb4b9dcSDavid S. Miller if (tcp_transmit_skb(sk, skb, 0, priority)) 30864e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 30871da177e4SLinus Torvalds } 30881da177e4SLinus Torvalds 308967edfef7SAndi Kleen /* Send a crossed SYN-ACK during socket establishment. 309067edfef7SAndi Kleen * WARNING: This routine must only be called when we have already sent 30911da177e4SLinus Torvalds * a SYN packet that crossed the incoming SYN that caused this routine 30921da177e4SLinus Torvalds * to get called. If this assumption fails then the initial rcv_wnd 30931da177e4SLinus Torvalds * and rcv_wscale values will not be correct. 30941da177e4SLinus Torvalds */ 30951da177e4SLinus Torvalds int tcp_send_synack(struct sock *sk) 30961da177e4SLinus Torvalds { 30971da177e4SLinus Torvalds struct sk_buff *skb; 30981da177e4SLinus Torvalds 3099*75c119afSEric Dumazet skb = tcp_rtx_queue_head(sk); 310051456b29SIan Morris if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 3101*75c119afSEric Dumazet pr_err("%s: wrong queue state\n", __func__); 31021da177e4SLinus Torvalds return -EFAULT; 31031da177e4SLinus Torvalds } 31044de075e0SEric Dumazet if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { 31051da177e4SLinus Torvalds if (skb_cloned(skb)) { 3106e2080072SEric Dumazet struct sk_buff *nskb; 3107e2080072SEric Dumazet 3108e2080072SEric Dumazet tcp_skb_tsorted_save(skb) { 3109e2080072SEric Dumazet nskb = skb_copy(skb, GFP_ATOMIC); 3110e2080072SEric Dumazet } tcp_skb_tsorted_restore(skb); 311151456b29SIan Morris if (!nskb) 31121da177e4SLinus Torvalds return -ENOMEM; 3113e2080072SEric Dumazet INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor); 3114*75c119afSEric Dumazet tcp_rtx_queue_unlink_and_free(skb, sk); 3115f4a775d1SEric Dumazet __skb_header_release(nskb); 3116*75c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb); 31173ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 31183ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 31191da177e4SLinus Torvalds skb = nskb; 31201da177e4SLinus Torvalds } 31211da177e4SLinus Torvalds 31224de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; 3123735d3831SFlorian Westphal tcp_ecn_send_synack(sk, skb); 31241da177e4SLinus Torvalds } 3125dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 31261da177e4SLinus Torvalds } 31271da177e4SLinus Torvalds 31284aea39c1SEric Dumazet /** 31294aea39c1SEric Dumazet * tcp_make_synack - Prepare a SYN-ACK. 31304aea39c1SEric Dumazet * sk: listener socket 31314aea39c1SEric Dumazet * dst: dst entry attached to the SYNACK 31324aea39c1SEric Dumazet * req: request_sock pointer 31334aea39c1SEric Dumazet * 31344aea39c1SEric Dumazet * Allocate one skb and build a SYNACK packet. 31354aea39c1SEric Dumazet * @dst is consumed : Caller should not use it again. 31364aea39c1SEric Dumazet */ 31375d062de7SEric Dumazet struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst, 3138e6b4d113SWilliam Allen Simpson struct request_sock *req, 3139ca6fb065SEric Dumazet struct tcp_fastopen_cookie *foc, 3140b3d05147SEric Dumazet enum tcp_synack_type synack_type) 31411da177e4SLinus Torvalds { 31422e6599cbSArnaldo Carvalho de Melo struct inet_request_sock *ireq = inet_rsk(req); 31435d062de7SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 314480f03e27SEric Dumazet struct tcp_md5sig_key *md5 = NULL; 31455d062de7SEric Dumazet struct tcp_out_options opts; 31465d062de7SEric Dumazet struct sk_buff *skb; 3147bd0388aeSWilliam Allen Simpson int tcp_header_size; 31485d062de7SEric Dumazet struct tcphdr *th; 3149f5fff5dcSTom Quetchenbach int mss; 31501da177e4SLinus Torvalds 3151ca6fb065SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 31524aea39c1SEric Dumazet if (unlikely(!skb)) { 31534aea39c1SEric Dumazet dst_release(dst); 31541da177e4SLinus Torvalds return NULL; 31554aea39c1SEric Dumazet } 31561da177e4SLinus Torvalds /* Reserve space for headers. */ 31571da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 31581da177e4SLinus Torvalds 3159b3d05147SEric Dumazet switch (synack_type) { 3160b3d05147SEric Dumazet case TCP_SYNACK_NORMAL: 31619e17f8a4SEric Dumazet skb_set_owner_w(skb, req_to_sk(req)); 3162b3d05147SEric Dumazet break; 3163b3d05147SEric Dumazet case TCP_SYNACK_COOKIE: 3164b3d05147SEric Dumazet /* Under synflood, we do not attach skb to a socket, 3165b3d05147SEric Dumazet * to avoid false sharing. 3166b3d05147SEric Dumazet */ 3167b3d05147SEric Dumazet break; 3168b3d05147SEric Dumazet case TCP_SYNACK_FASTOPEN: 3169ca6fb065SEric Dumazet /* sk is a const pointer, because we want to express multiple 3170ca6fb065SEric Dumazet * cpu might call us concurrently. 3171ca6fb065SEric Dumazet * sk->sk_wmem_alloc in an atomic, we can promote to rw. 3172ca6fb065SEric Dumazet */ 3173ca6fb065SEric Dumazet skb_set_owner_w(skb, (struct sock *)sk); 3174b3d05147SEric Dumazet break; 3175ca6fb065SEric Dumazet } 31764aea39c1SEric Dumazet skb_dst_set(skb, dst); 31771da177e4SLinus Torvalds 31783541f9e8SEric Dumazet mss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 3179f5fff5dcSTom Quetchenbach 318033ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 31818b5f12d0SFlorian Westphal #ifdef CONFIG_SYN_COOKIES 31828b5f12d0SFlorian Westphal if (unlikely(req->cookie_ts)) 31839a568de4SEric Dumazet skb->skb_mstamp = cookie_init_timestamp(req); 31848b5f12d0SFlorian Westphal else 31858b5f12d0SFlorian Westphal #endif 31869a568de4SEric Dumazet skb->skb_mstamp = tcp_clock_us(); 318780f03e27SEric Dumazet 318880f03e27SEric Dumazet #ifdef CONFIG_TCP_MD5SIG 318980f03e27SEric Dumazet rcu_read_lock(); 3190fd3a154aSEric Dumazet md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req)); 319180f03e27SEric Dumazet #endif 319258d607d3SEric Dumazet skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4); 319337bfbddaSEric Dumazet tcp_header_size = tcp_synack_options(req, mss, skb, &opts, md5, foc) + 319437bfbddaSEric Dumazet sizeof(*th); 319533ad798cSAdam Langley 3196aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 3197aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 31981da177e4SLinus Torvalds 3199ea1627c2SEric Dumazet th = (struct tcphdr *)skb->data; 32001da177e4SLinus Torvalds memset(th, 0, sizeof(struct tcphdr)); 32011da177e4SLinus Torvalds th->syn = 1; 32021da177e4SLinus Torvalds th->ack = 1; 32036ac705b1SEric Dumazet tcp_ecn_make_synack(req, th); 3204b44084c2SEric Dumazet th->source = htons(ireq->ir_num); 3205634fb979SEric Dumazet th->dest = ireq->ir_rmt_port; 3206e05a90ecSJamal Hadi Salim skb->mark = ireq->ir_mark; 3207e870a8efSIlpo Järvinen /* Setting of flags are superfluous here for callers (and ECE is 3208e870a8efSIlpo Järvinen * not even correctly set) 3209e870a8efSIlpo Järvinen */ 3210e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn, 3211a3433f35SChangli Gao TCPHDR_SYN | TCPHDR_ACK); 32124957faadSWilliam Allen Simpson 32131da177e4SLinus Torvalds th->seq = htonl(TCP_SKB_CB(skb)->seq); 32148336886fSJerry Chu /* XXX data is queued and acked as is. No buffer/window check */ 32158336886fSJerry Chu th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt); 32161da177e4SLinus Torvalds 32171da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 3218ed53d0abSEric Dumazet th->window = htons(min(req->rsk_rcv_wnd, 65535U)); 32195d062de7SEric Dumazet tcp_options_write((__be32 *)(th + 1), NULL, &opts); 32201da177e4SLinus Torvalds th->doff = (tcp_header_size >> 2); 322190bbcc60SEric Dumazet __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS); 3222cfb6eeb4SYOSHIFUJI Hideaki 3223cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 3224cfb6eeb4SYOSHIFUJI Hideaki /* Okay, we have all we need - do the md5 hash if needed */ 322580f03e27SEric Dumazet if (md5) 3226bd0388aeSWilliam Allen Simpson tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, 322739f8e58eSEric Dumazet md5, req_to_sk(req), skb); 322880f03e27SEric Dumazet rcu_read_unlock(); 3229cfb6eeb4SYOSHIFUJI Hideaki #endif 3230cfb6eeb4SYOSHIFUJI Hideaki 3231b50edd78SEric Dumazet /* Do not fool tcpdump (if any), clean our debris */ 32322456e855SThomas Gleixner skb->tstamp = 0; 32331da177e4SLinus Torvalds return skb; 32341da177e4SLinus Torvalds } 32354bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_make_synack); 32361da177e4SLinus Torvalds 323781164413SDaniel Borkmann static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst) 323881164413SDaniel Borkmann { 323981164413SDaniel Borkmann struct inet_connection_sock *icsk = inet_csk(sk); 324081164413SDaniel Borkmann const struct tcp_congestion_ops *ca; 324181164413SDaniel Borkmann u32 ca_key = dst_metric(dst, RTAX_CC_ALGO); 324281164413SDaniel Borkmann 324381164413SDaniel Borkmann if (ca_key == TCP_CA_UNSPEC) 324481164413SDaniel Borkmann return; 324581164413SDaniel Borkmann 324681164413SDaniel Borkmann rcu_read_lock(); 324781164413SDaniel Borkmann ca = tcp_ca_find_key(ca_key); 324881164413SDaniel Borkmann if (likely(ca && try_module_get(ca->owner))) { 324981164413SDaniel Borkmann module_put(icsk->icsk_ca_ops->owner); 325081164413SDaniel Borkmann icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst); 325181164413SDaniel Borkmann icsk->icsk_ca_ops = ca; 325281164413SDaniel Borkmann } 325381164413SDaniel Borkmann rcu_read_unlock(); 325481164413SDaniel Borkmann } 325581164413SDaniel Borkmann 325667edfef7SAndi Kleen /* Do all connect socket setups that can be done AF independent. */ 3257f7e56a76Sstephen hemminger static void tcp_connect_init(struct sock *sk) 32581da177e4SLinus Torvalds { 3259cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 32601da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 32611da177e4SLinus Torvalds __u8 rcv_wscale; 326213d3b1ebSLawrence Brakmo u32 rcv_wnd; 32631da177e4SLinus Torvalds 32641da177e4SLinus Torvalds /* We'll fix this up when we get a response from the other end. 32651da177e4SLinus Torvalds * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 32661da177e4SLinus Torvalds */ 32675d2ed052SEric Dumazet tp->tcp_header_len = sizeof(struct tcphdr); 32685d2ed052SEric Dumazet if (sock_net(sk)->ipv4.sysctl_tcp_timestamps) 32695d2ed052SEric Dumazet tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED; 32701da177e4SLinus Torvalds 3271cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 327200db4124SIan Morris if (tp->af_specific->md5_lookup(sk, sk)) 3273cfb6eeb4SYOSHIFUJI Hideaki tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 3274cfb6eeb4SYOSHIFUJI Hideaki #endif 3275cfb6eeb4SYOSHIFUJI Hideaki 32761da177e4SLinus Torvalds /* If user gave his TCP_MAXSEG, record it to clamp */ 32771da177e4SLinus Torvalds if (tp->rx_opt.user_mss) 32781da177e4SLinus Torvalds tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 32791da177e4SLinus Torvalds tp->max_window = 0; 32805d424d5aSJohn Heffner tcp_mtup_init(sk); 32811da177e4SLinus Torvalds tcp_sync_mss(sk, dst_mtu(dst)); 32821da177e4SLinus Torvalds 328381164413SDaniel Borkmann tcp_ca_dst_init(sk, dst); 328481164413SDaniel Borkmann 32851da177e4SLinus Torvalds if (!tp->window_clamp) 32861da177e4SLinus Torvalds tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 32873541f9e8SEric Dumazet tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 3288f5fff5dcSTom Quetchenbach 32891da177e4SLinus Torvalds tcp_initialize_rcv_mss(sk); 32901da177e4SLinus Torvalds 3291e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 3292e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 3293e88c64f0SHagen Paul Pfeifer (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) 3294e88c64f0SHagen Paul Pfeifer tp->window_clamp = tcp_full_space(sk); 3295e88c64f0SHagen Paul Pfeifer 329613d3b1ebSLawrence Brakmo rcv_wnd = tcp_rwnd_init_bpf(sk); 329713d3b1ebSLawrence Brakmo if (rcv_wnd == 0) 329813d3b1ebSLawrence Brakmo rcv_wnd = dst_metric(dst, RTAX_INITRWND); 329913d3b1ebSLawrence Brakmo 33001da177e4SLinus Torvalds tcp_select_initial_window(tcp_full_space(sk), 33011da177e4SLinus Torvalds tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 33021da177e4SLinus Torvalds &tp->rcv_wnd, 33031da177e4SLinus Torvalds &tp->window_clamp, 33049bb37ef0SEric Dumazet sock_net(sk)->ipv4.sysctl_tcp_window_scaling, 330531d12926Slaurent chavey &rcv_wscale, 330613d3b1ebSLawrence Brakmo rcv_wnd); 33071da177e4SLinus Torvalds 33081da177e4SLinus Torvalds tp->rx_opt.rcv_wscale = rcv_wscale; 33091da177e4SLinus Torvalds tp->rcv_ssthresh = tp->rcv_wnd; 33101da177e4SLinus Torvalds 33111da177e4SLinus Torvalds sk->sk_err = 0; 33121da177e4SLinus Torvalds sock_reset_flag(sk, SOCK_DONE); 33131da177e4SLinus Torvalds tp->snd_wnd = 0; 3314ee7537b6SHantzis Fotis tcp_init_wl(tp, 0); 33151da177e4SLinus Torvalds tp->snd_una = tp->write_seq; 33161da177e4SLinus Torvalds tp->snd_sml = tp->write_seq; 331733f5f57eSIlpo Järvinen tp->snd_up = tp->write_seq; 3318370816aeSPavel Emelyanov tp->snd_nxt = tp->write_seq; 3319ee995283SPavel Emelyanov 3320ee995283SPavel Emelyanov if (likely(!tp->repair)) 33211da177e4SLinus Torvalds tp->rcv_nxt = 0; 3322c7781a6eSAndrew Vagin else 332370eabf0eSEric Dumazet tp->rcv_tstamp = tcp_jiffies32; 3324ee995283SPavel Emelyanov tp->rcv_wup = tp->rcv_nxt; 3325ee995283SPavel Emelyanov tp->copied_seq = tp->rcv_nxt; 33261da177e4SLinus Torvalds 33278550f328SLawrence Brakmo inet_csk(sk)->icsk_rto = tcp_timeout_init(sk); 3328463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_retransmits = 0; 33291da177e4SLinus Torvalds tcp_clear_retrans(tp); 33301da177e4SLinus Torvalds } 33311da177e4SLinus Torvalds 3332783237e8SYuchung Cheng static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) 3333783237e8SYuchung Cheng { 3334783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3335783237e8SYuchung Cheng struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 3336783237e8SYuchung Cheng 3337783237e8SYuchung Cheng tcb->end_seq += skb->len; 3338f4a775d1SEric Dumazet __skb_header_release(skb); 3339783237e8SYuchung Cheng sk->sk_wmem_queued += skb->truesize; 3340783237e8SYuchung Cheng sk_mem_charge(sk, skb->truesize); 3341783237e8SYuchung Cheng tp->write_seq = tcb->end_seq; 3342783237e8SYuchung Cheng tp->packets_out += tcp_skb_pcount(skb); 3343783237e8SYuchung Cheng } 3344783237e8SYuchung Cheng 3345783237e8SYuchung Cheng /* Build and send a SYN with data and (cached) Fast Open cookie. However, 3346783237e8SYuchung Cheng * queue a data-only packet after the regular SYN, such that regular SYNs 3347783237e8SYuchung Cheng * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges 3348783237e8SYuchung Cheng * only the SYN sequence, the data are retransmitted in the first ACK. 3349783237e8SYuchung Cheng * If cookie is not cached or other error occurs, falls back to send a 3350783237e8SYuchung Cheng * regular SYN with Fast Open cookie request option. 3351783237e8SYuchung Cheng */ 3352783237e8SYuchung Cheng static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) 3353783237e8SYuchung Cheng { 3354783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3355783237e8SYuchung Cheng struct tcp_fastopen_request *fo = tp->fastopen_req; 3356065263f4SWei Wang int space, err = 0; 3357355a901eSEric Dumazet struct sk_buff *syn_data; 3358783237e8SYuchung Cheng 335967da22d2SYuchung Cheng tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ 3360065263f4SWei Wang if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie)) 3361783237e8SYuchung Cheng goto fallback; 3362783237e8SYuchung Cheng 3363783237e8SYuchung Cheng /* MSS for SYN-data is based on cached MSS and bounded by PMTU and 3364783237e8SYuchung Cheng * user-MSS. Reserve maximum option space for middleboxes that add 3365783237e8SYuchung Cheng * private TCP options. The cost is reduced data space in SYN :( 3366783237e8SYuchung Cheng */ 33673541f9e8SEric Dumazet tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp); 33683541f9e8SEric Dumazet 33691b63edd6SYuchung Cheng space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) - 3370783237e8SYuchung Cheng MAX_TCP_OPTION_SPACE; 3371783237e8SYuchung Cheng 3372f5ddcbbbSEric Dumazet space = min_t(size_t, space, fo->size); 3373f5ddcbbbSEric Dumazet 3374f5ddcbbbSEric Dumazet /* limit to order-0 allocations */ 3375f5ddcbbbSEric Dumazet space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER)); 3376f5ddcbbbSEric Dumazet 3377eb934478SEric Dumazet syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false); 3378355a901eSEric Dumazet if (!syn_data) 3379783237e8SYuchung Cheng goto fallback; 3380355a901eSEric Dumazet syn_data->ip_summed = CHECKSUM_PARTIAL; 3381355a901eSEric Dumazet memcpy(syn_data->cb, syn->cb, sizeof(syn->cb)); 338207e100f9SEric Dumazet if (space) { 338307e100f9SEric Dumazet int copied = copy_from_iter(skb_put(syn_data, space), space, 338457be5bdaSAl Viro &fo->data->msg_iter); 338557be5bdaSAl Viro if (unlikely(!copied)) { 3386355a901eSEric Dumazet kfree_skb(syn_data); 3387783237e8SYuchung Cheng goto fallback; 3388783237e8SYuchung Cheng } 338957be5bdaSAl Viro if (copied != space) { 339057be5bdaSAl Viro skb_trim(syn_data, copied); 339157be5bdaSAl Viro space = copied; 339257be5bdaSAl Viro } 339307e100f9SEric Dumazet } 3394355a901eSEric Dumazet /* No more data pending in inet_wait_for_connect() */ 3395355a901eSEric Dumazet if (space == fo->size) 3396355a901eSEric Dumazet fo->data = NULL; 3397355a901eSEric Dumazet fo->copied = space; 3398783237e8SYuchung Cheng 3399355a901eSEric Dumazet tcp_connect_queue_skb(sk, syn_data); 34000f87230dSFrancis Yan if (syn_data->len) 34010f87230dSFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_BUSY); 3402355a901eSEric Dumazet 3403355a901eSEric Dumazet err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation); 3404355a901eSEric Dumazet 3405355a901eSEric Dumazet syn->skb_mstamp = syn_data->skb_mstamp; 3406355a901eSEric Dumazet 3407355a901eSEric Dumazet /* Now full SYN+DATA was cloned and sent (or not), 3408355a901eSEric Dumazet * remove the SYN from the original skb (syn_data) 3409355a901eSEric Dumazet * we keep in write queue in case of a retransmit, as we 3410355a901eSEric Dumazet * also have the SYN packet (with no data) in the same queue. 3411431a9124SEric Dumazet */ 3412355a901eSEric Dumazet TCP_SKB_CB(syn_data)->seq++; 3413355a901eSEric Dumazet TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH; 3414355a901eSEric Dumazet if (!err) { 341567da22d2SYuchung Cheng tp->syn_data = (fo->copied > 0); 3416*75c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data); 3417f19c29e3SYuchung Cheng NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT); 3418783237e8SYuchung Cheng goto done; 3419783237e8SYuchung Cheng } 3420783237e8SYuchung Cheng 3421*75c119afSEric Dumazet /* data was not sent, put it in write_queue */ 3422*75c119afSEric Dumazet __skb_queue_tail(&sk->sk_write_queue, syn_data); 3423b5b7db8dSEric Dumazet tp->packets_out -= tcp_skb_pcount(syn_data); 3424b5b7db8dSEric Dumazet 3425783237e8SYuchung Cheng fallback: 3426783237e8SYuchung Cheng /* Send a regular SYN with Fast Open cookie request option */ 3427783237e8SYuchung Cheng if (fo->cookie.len > 0) 3428783237e8SYuchung Cheng fo->cookie.len = 0; 3429783237e8SYuchung Cheng err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); 3430783237e8SYuchung Cheng if (err) 3431783237e8SYuchung Cheng tp->syn_fastopen = 0; 3432783237e8SYuchung Cheng done: 3433783237e8SYuchung Cheng fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ 3434783237e8SYuchung Cheng return err; 3435783237e8SYuchung Cheng } 3436783237e8SYuchung Cheng 343767edfef7SAndi Kleen /* Build a SYN and send it off. */ 34381da177e4SLinus Torvalds int tcp_connect(struct sock *sk) 34391da177e4SLinus Torvalds { 34401da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 34411da177e4SLinus Torvalds struct sk_buff *buff; 3442ee586811SEric Paris int err; 34431da177e4SLinus Torvalds 34449872a4bdSLawrence Brakmo tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB); 34458ba60924SEric Dumazet 34468ba60924SEric Dumazet if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 34478ba60924SEric Dumazet return -EHOSTUNREACH; /* Routing failure or similar. */ 34488ba60924SEric Dumazet 34491da177e4SLinus Torvalds tcp_connect_init(sk); 34501da177e4SLinus Torvalds 34512b916477SAndrey Vagin if (unlikely(tp->repair)) { 34522b916477SAndrey Vagin tcp_finish_connect(sk, NULL); 34532b916477SAndrey Vagin return 0; 34542b916477SAndrey Vagin } 34552b916477SAndrey Vagin 3456eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true); 3457355a901eSEric Dumazet if (unlikely(!buff)) 34581da177e4SLinus Torvalds return -ENOBUFS; 34591da177e4SLinus Torvalds 3460a3433f35SChangli Gao tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); 34619a568de4SEric Dumazet tcp_mstamp_refresh(tp); 34629a568de4SEric Dumazet tp->retrans_stamp = tcp_time_stamp(tp); 3463783237e8SYuchung Cheng tcp_connect_queue_skb(sk, buff); 3464735d3831SFlorian Westphal tcp_ecn_send_syn(sk, buff); 3465*75c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, buff); 34661da177e4SLinus Torvalds 3467783237e8SYuchung Cheng /* Send off SYN; include data in Fast Open. */ 3468783237e8SYuchung Cheng err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : 3469783237e8SYuchung Cheng tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); 3470ee586811SEric Paris if (err == -ECONNREFUSED) 3471ee586811SEric Paris return err; 3472bd37a088SWei Yongjun 3473bd37a088SWei Yongjun /* We change tp->snd_nxt after the tcp_transmit_skb() call 3474bd37a088SWei Yongjun * in order to make this packet get counted in tcpOutSegs. 3475bd37a088SWei Yongjun */ 3476bd37a088SWei Yongjun tp->snd_nxt = tp->write_seq; 3477bd37a088SWei Yongjun tp->pushed_seq = tp->write_seq; 3478b5b7db8dSEric Dumazet buff = tcp_send_head(sk); 3479b5b7db8dSEric Dumazet if (unlikely(buff)) { 3480b5b7db8dSEric Dumazet tp->snd_nxt = TCP_SKB_CB(buff)->seq; 3481b5b7db8dSEric Dumazet tp->pushed_seq = TCP_SKB_CB(buff)->seq; 3482b5b7db8dSEric Dumazet } 348381cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); 34841da177e4SLinus Torvalds 34851da177e4SLinus Torvalds /* Timer for repeating the SYN until an answer. */ 34863f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 34873f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 34881da177e4SLinus Torvalds return 0; 34891da177e4SLinus Torvalds } 34904bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_connect); 34911da177e4SLinus Torvalds 34921da177e4SLinus Torvalds /* Send out a delayed ack, the caller does the policy checking 34931da177e4SLinus Torvalds * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 34941da177e4SLinus Torvalds * for details. 34951da177e4SLinus Torvalds */ 34961da177e4SLinus Torvalds void tcp_send_delayed_ack(struct sock *sk) 34971da177e4SLinus Torvalds { 3498463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 3499463c84b9SArnaldo Carvalho de Melo int ato = icsk->icsk_ack.ato; 35001da177e4SLinus Torvalds unsigned long timeout; 35011da177e4SLinus Torvalds 35029890092eSFlorian Westphal tcp_ca_event(sk, CA_EVENT_DELAYED_ACK); 35039890092eSFlorian Westphal 35041da177e4SLinus Torvalds if (ato > TCP_DELACK_MIN) { 3505463c84b9SArnaldo Carvalho de Melo const struct tcp_sock *tp = tcp_sk(sk); 35061da177e4SLinus Torvalds int max_ato = HZ / 2; 35071da177e4SLinus Torvalds 3508056834d9SIlpo Järvinen if (icsk->icsk_ack.pingpong || 3509056834d9SIlpo Järvinen (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 35101da177e4SLinus Torvalds max_ato = TCP_DELACK_MAX; 35111da177e4SLinus Torvalds 35121da177e4SLinus Torvalds /* Slow path, intersegment interval is "high". */ 35131da177e4SLinus Torvalds 35141da177e4SLinus Torvalds /* If some rtt estimate is known, use it to bound delayed ack. 3515463c84b9SArnaldo Carvalho de Melo * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 35161da177e4SLinus Torvalds * directly. 35171da177e4SLinus Torvalds */ 3518740b0f18SEric Dumazet if (tp->srtt_us) { 3519740b0f18SEric Dumazet int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3), 3520740b0f18SEric Dumazet TCP_DELACK_MIN); 35211da177e4SLinus Torvalds 35221da177e4SLinus Torvalds if (rtt < max_ato) 35231da177e4SLinus Torvalds max_ato = rtt; 35241da177e4SLinus Torvalds } 35251da177e4SLinus Torvalds 35261da177e4SLinus Torvalds ato = min(ato, max_ato); 35271da177e4SLinus Torvalds } 35281da177e4SLinus Torvalds 35291da177e4SLinus Torvalds /* Stay within the limit we were given */ 35301da177e4SLinus Torvalds timeout = jiffies + ato; 35311da177e4SLinus Torvalds 35321da177e4SLinus Torvalds /* Use new timeout only if there wasn't a older one earlier. */ 3533463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 35341da177e4SLinus Torvalds /* If delack timer was blocked or is about to expire, 35351da177e4SLinus Torvalds * send ACK now. 35361da177e4SLinus Torvalds */ 3537463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.blocked || 3538463c84b9SArnaldo Carvalho de Melo time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 35391da177e4SLinus Torvalds tcp_send_ack(sk); 35401da177e4SLinus Torvalds return; 35411da177e4SLinus Torvalds } 35421da177e4SLinus Torvalds 3543463c84b9SArnaldo Carvalho de Melo if (!time_before(timeout, icsk->icsk_ack.timeout)) 3544463c84b9SArnaldo Carvalho de Melo timeout = icsk->icsk_ack.timeout; 35451da177e4SLinus Torvalds } 3546463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 3547463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.timeout = timeout; 3548463c84b9SArnaldo Carvalho de Melo sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 35491da177e4SLinus Torvalds } 35501da177e4SLinus Torvalds 35511da177e4SLinus Torvalds /* This routine sends an ack and also updates the window. */ 35521da177e4SLinus Torvalds void tcp_send_ack(struct sock *sk) 35531da177e4SLinus Torvalds { 35541da177e4SLinus Torvalds struct sk_buff *buff; 35551da177e4SLinus Torvalds 3556058dc334SIlpo Järvinen /* If we have been reset, we may not send again. */ 3557058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3558058dc334SIlpo Järvinen return; 3559058dc334SIlpo Järvinen 35609890092eSFlorian Westphal tcp_ca_event(sk, CA_EVENT_NON_DELAYED_ACK); 35619890092eSFlorian Westphal 35621da177e4SLinus Torvalds /* We are not putting this on the write queue, so 35631da177e4SLinus Torvalds * tcp_transmit_skb() will set the ownership to this 35641da177e4SLinus Torvalds * sock. 35651da177e4SLinus Torvalds */ 35667450aaf6SEric Dumazet buff = alloc_skb(MAX_TCP_HEADER, 35677450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 35687450aaf6SEric Dumazet if (unlikely(!buff)) { 3569463c84b9SArnaldo Carvalho de Melo inet_csk_schedule_ack(sk); 3570463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; 35713f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 35723f421baaSArnaldo Carvalho de Melo TCP_DELACK_MAX, TCP_RTO_MAX); 35731da177e4SLinus Torvalds return; 35741da177e4SLinus Torvalds } 35751da177e4SLinus Torvalds 35761da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 35771da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 3578a3433f35SChangli Gao tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); 35791da177e4SLinus Torvalds 358098781965SEric Dumazet /* We do not want pure acks influencing TCP Small Queues or fq/pacing 358198781965SEric Dumazet * too much. 358298781965SEric Dumazet * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784 358398781965SEric Dumazet */ 358498781965SEric Dumazet skb_set_tcp_pure_ack(buff); 358598781965SEric Dumazet 35861da177e4SLinus Torvalds /* Send it off, this clears delayed acks for us. */ 35877450aaf6SEric Dumazet tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0); 35881da177e4SLinus Torvalds } 3589e3118e83SDaniel Borkmann EXPORT_SYMBOL_GPL(tcp_send_ack); 35901da177e4SLinus Torvalds 35911da177e4SLinus Torvalds /* This routine sends a packet with an out of date sequence 35921da177e4SLinus Torvalds * number. It assumes the other end will try to ack it. 35931da177e4SLinus Torvalds * 35941da177e4SLinus Torvalds * Question: what should we make while urgent mode? 35951da177e4SLinus Torvalds * 4.4BSD forces sending single byte of data. We cannot send 35961da177e4SLinus Torvalds * out of window data, because we have SND.NXT==SND.MAX... 35971da177e4SLinus Torvalds * 35981da177e4SLinus Torvalds * Current solution: to send TWO zero-length segments in urgent mode: 35991da177e4SLinus Torvalds * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 36001da177e4SLinus Torvalds * out-of-date with SND.UNA-1 to probe window. 36011da177e4SLinus Torvalds */ 3602e520af48SEric Dumazet static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib) 36031da177e4SLinus Torvalds { 36041da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 36051da177e4SLinus Torvalds struct sk_buff *skb; 36061da177e4SLinus Torvalds 36071da177e4SLinus Torvalds /* We don't queue it, tcp_transmit_skb() sets ownership. */ 36087450aaf6SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, 36097450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 361051456b29SIan Morris if (!skb) 36111da177e4SLinus Torvalds return -1; 36121da177e4SLinus Torvalds 36131da177e4SLinus Torvalds /* Reserve space for headers and set control bits. */ 36141da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 36151da177e4SLinus Torvalds /* Use a previous sequence. This should cause the other 36161da177e4SLinus Torvalds * end to send an ack. Don't queue or clone SKB, just 36171da177e4SLinus Torvalds * send it. 36181da177e4SLinus Torvalds */ 3619a3433f35SChangli Gao tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); 3620e2e8009fSRenato Westphal NET_INC_STATS(sock_net(sk), mib); 36217450aaf6SEric Dumazet return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0); 36221da177e4SLinus Torvalds } 36231da177e4SLinus Torvalds 3624385e2070SEric Dumazet /* Called from setsockopt( ... TCP_REPAIR ) */ 3625ee995283SPavel Emelyanov void tcp_send_window_probe(struct sock *sk) 3626ee995283SPavel Emelyanov { 3627ee995283SPavel Emelyanov if (sk->sk_state == TCP_ESTABLISHED) { 3628ee995283SPavel Emelyanov tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; 36299a568de4SEric Dumazet tcp_mstamp_refresh(tcp_sk(sk)); 3630e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE); 3631ee995283SPavel Emelyanov } 3632ee995283SPavel Emelyanov } 3633ee995283SPavel Emelyanov 363467edfef7SAndi Kleen /* Initiate keepalive or window probe from timer. */ 3635e520af48SEric Dumazet int tcp_write_wakeup(struct sock *sk, int mib) 36361da177e4SLinus Torvalds { 36371da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 36381da177e4SLinus Torvalds struct sk_buff *skb; 36391da177e4SLinus Torvalds 3640058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3641058dc334SIlpo Järvinen return -1; 3642058dc334SIlpo Järvinen 364300db4124SIan Morris skb = tcp_send_head(sk); 364400db4124SIan Morris if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { 36451da177e4SLinus Torvalds int err; 36460c54b85fSIlpo Järvinen unsigned int mss = tcp_current_mss(sk); 364790840defSIlpo Järvinen unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 36481da177e4SLinus Torvalds 36491da177e4SLinus Torvalds if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 36501da177e4SLinus Torvalds tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 36511da177e4SLinus Torvalds 36521da177e4SLinus Torvalds /* We are probing the opening of a window 36531da177e4SLinus Torvalds * but the window size is != 0 36541da177e4SLinus Torvalds * must have been a result SWS avoidance ( sender ) 36551da177e4SLinus Torvalds */ 36561da177e4SLinus Torvalds if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 36571da177e4SLinus Torvalds skb->len > mss) { 36581da177e4SLinus Torvalds seg_size = min(seg_size, mss); 36594de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 3660*75c119afSEric Dumazet if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE, 3661*75c119afSEric Dumazet skb, seg_size, mss, GFP_ATOMIC)) 36621da177e4SLinus Torvalds return -1; 36631da177e4SLinus Torvalds } else if (!tcp_skb_pcount(skb)) 36645bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss); 36651da177e4SLinus Torvalds 36664de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 3667dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 366866f5fe62SIlpo Järvinen if (!err) 366966f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 36701da177e4SLinus Torvalds return err; 36711da177e4SLinus Torvalds } else { 367233f5f57eSIlpo Järvinen if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) 3673e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 1, mib); 3674e520af48SEric Dumazet return tcp_xmit_probe_skb(sk, 0, mib); 36751da177e4SLinus Torvalds } 36761da177e4SLinus Torvalds } 36771da177e4SLinus Torvalds 36781da177e4SLinus Torvalds /* A window probe timeout has occurred. If window is not closed send 36791da177e4SLinus Torvalds * a partial packet else a zero probe. 36801da177e4SLinus Torvalds */ 36811da177e4SLinus Torvalds void tcp_send_probe0(struct sock *sk) 36821da177e4SLinus Torvalds { 3683463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 36841da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 3685c6214a97SNikolay Borisov struct net *net = sock_net(sk); 3686fcdd1cf4SEric Dumazet unsigned long probe_max; 36871da177e4SLinus Torvalds int err; 36881da177e4SLinus Torvalds 3689e520af48SEric Dumazet err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE); 36901da177e4SLinus Torvalds 3691*75c119afSEric Dumazet if (tp->packets_out || tcp_write_queue_empty(sk)) { 36921da177e4SLinus Torvalds /* Cancel probe timer, if it is not required. */ 36936687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 0; 3694463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff = 0; 36951da177e4SLinus Torvalds return; 36961da177e4SLinus Torvalds } 36971da177e4SLinus Torvalds 36981da177e4SLinus Torvalds if (err <= 0) { 3699c6214a97SNikolay Borisov if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2) 3700463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff++; 37016687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out++; 3702fcdd1cf4SEric Dumazet probe_max = TCP_RTO_MAX; 37031da177e4SLinus Torvalds } else { 37041da177e4SLinus Torvalds /* If packet was not sent due to local congestion, 37056687e988SArnaldo Carvalho de Melo * do not backoff and do not remember icsk_probes_out. 37061da177e4SLinus Torvalds * Let local senders to fight for local resources. 37071da177e4SLinus Torvalds * 37081da177e4SLinus Torvalds * Use accumulated backoff yet. 37091da177e4SLinus Torvalds */ 37106687e988SArnaldo Carvalho de Melo if (!icsk->icsk_probes_out) 37116687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 1; 3712fcdd1cf4SEric Dumazet probe_max = TCP_RESOURCE_PROBE_INTERVAL; 37131da177e4SLinus Torvalds } 3714fcdd1cf4SEric Dumazet inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 371521c8fe99SEric Dumazet tcp_probe0_when(sk, probe_max), 3716fcdd1cf4SEric Dumazet TCP_RTO_MAX); 37171da177e4SLinus Torvalds } 37185db92c99SOctavian Purdila 3719ea3bea3aSEric Dumazet int tcp_rtx_synack(const struct sock *sk, struct request_sock *req) 37205db92c99SOctavian Purdila { 37215db92c99SOctavian Purdila const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific; 37225db92c99SOctavian Purdila struct flowi fl; 37235db92c99SOctavian Purdila int res; 37245db92c99SOctavian Purdila 372558d607d3SEric Dumazet tcp_rsk(req)->txhash = net_tx_rndhash(); 3726b3d05147SEric Dumazet res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL); 37275db92c99SOctavian Purdila if (!res) { 372890bbcc60SEric Dumazet __TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); 372902a1d6e7SEric Dumazet __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 37307e32b443SYuchung Cheng if (unlikely(tcp_passive_fastopen(sk))) 37317e32b443SYuchung Cheng tcp_sk(sk)->total_retrans++; 37325db92c99SOctavian Purdila } 37335db92c99SOctavian Purdila return res; 37345db92c99SOctavian Purdila } 37355db92c99SOctavian Purdila EXPORT_SYMBOL(tcp_rtx_synack); 3736