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. */ 69cf533ea5SEric Dumazet static void tcp_event_new_data_sent(struct sock *sk, const 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 75fe067e8aSDavid S. Miller tcp_advance_send_head(sk, skb); 761da177e4SLinus Torvalds tp->snd_nxt = TCP_SKB_CB(skb)->end_seq; 778512430eSIlpo Järvinen 7866f5fe62SIlpo Järvinen tp->packets_out += tcp_skb_pcount(skb); 796ba8a3b1SNandita Dukkipati if (!prior_packets || icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS || 806a5dc9e5SEric Dumazet icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) { 81750ea2baSYuchung Cheng tcp_rearm_rto(sk); 821da177e4SLinus Torvalds } 83f19c29e3SYuchung Cheng 84f7324acdSDavid S. Miller NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT, 85f19c29e3SYuchung Cheng tcp_skb_pcount(skb)); 866a5dc9e5SEric Dumazet } 871da177e4SLinus Torvalds 881da177e4SLinus Torvalds /* SND.NXT, if window was not shrunk. 891da177e4SLinus Torvalds * If window has been shrunk, what should we make? It is not clear at all. 901da177e4SLinus Torvalds * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-( 911da177e4SLinus Torvalds * Anything in between SND.UNA...SND.UNA+SND.WND also can be already 921da177e4SLinus Torvalds * invalid. OK, let's make this for now: 931da177e4SLinus Torvalds */ 94cf533ea5SEric Dumazet static inline __u32 tcp_acceptable_seq(const struct sock *sk) 951da177e4SLinus Torvalds { 96cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 979e412ba7SIlpo Järvinen 9890840defSIlpo Järvinen if (!before(tcp_wnd_end(tp), tp->snd_nxt)) 991da177e4SLinus Torvalds return tp->snd_nxt; 1001da177e4SLinus Torvalds else 10190840defSIlpo Järvinen return tcp_wnd_end(tp); 1021da177e4SLinus Torvalds } 1031da177e4SLinus Torvalds 1041da177e4SLinus Torvalds /* Calculate mss to advertise in SYN segment. 1051da177e4SLinus Torvalds * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that: 1061da177e4SLinus Torvalds * 1071da177e4SLinus Torvalds * 1. It is independent of path mtu. 1081da177e4SLinus Torvalds * 2. Ideally, it is maximal possible segment size i.e. 65535-40. 1091da177e4SLinus Torvalds * 3. For IPv4 it is reasonable to calculate it from maximal MTU of 1101da177e4SLinus Torvalds * attached devices, because some buggy hosts are confused by 1111da177e4SLinus Torvalds * large MSS. 1121da177e4SLinus Torvalds * 4. We do not make 3, we advertise MSS, calculated from first 1131da177e4SLinus Torvalds * hop device mtu, but allow to raise it to ip_rt_min_advmss. 1141da177e4SLinus Torvalds * This may be overridden via information stored in routing table. 1151da177e4SLinus Torvalds * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible, 1161da177e4SLinus Torvalds * probably even Jumbo". 1171da177e4SLinus Torvalds */ 1181da177e4SLinus Torvalds static __u16 tcp_advertise_mss(struct sock *sk) 1191da177e4SLinus Torvalds { 1201da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 121cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1221da177e4SLinus Torvalds int mss = tp->advmss; 1231da177e4SLinus Torvalds 1240dbaee3bSDavid S. Miller if (dst) { 1250dbaee3bSDavid S. Miller unsigned int metric = dst_metric_advmss(dst); 1260dbaee3bSDavid S. Miller 1270dbaee3bSDavid S. Miller if (metric < mss) { 1280dbaee3bSDavid S. Miller mss = metric; 1291da177e4SLinus Torvalds tp->advmss = mss; 1301da177e4SLinus Torvalds } 1310dbaee3bSDavid S. Miller } 1321da177e4SLinus Torvalds 1331da177e4SLinus Torvalds return (__u16)mss; 1341da177e4SLinus Torvalds } 1351da177e4SLinus Torvalds 1361da177e4SLinus Torvalds /* RFC2861. Reset CWND after idle period longer RTO to "restart window". 1376f021c62SEric Dumazet * This is the first part of cwnd validation mechanism. 1386f021c62SEric Dumazet */ 1396f021c62SEric Dumazet void tcp_cwnd_restart(struct sock *sk, s32 delta) 1401da177e4SLinus Torvalds { 141463c84b9SArnaldo Carvalho de Melo struct tcp_sock *tp = tcp_sk(sk); 1426f021c62SEric Dumazet u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk)); 1431da177e4SLinus Torvalds u32 cwnd = tp->snd_cwnd; 1441da177e4SLinus Torvalds 1456687e988SArnaldo Carvalho de Melo tcp_ca_event(sk, CA_EVENT_CWND_RESTART); 1461da177e4SLinus Torvalds 1476687e988SArnaldo Carvalho de Melo tp->snd_ssthresh = tcp_current_ssthresh(sk); 1481da177e4SLinus Torvalds restart_cwnd = min(restart_cwnd, cwnd); 1491da177e4SLinus Torvalds 150463c84b9SArnaldo Carvalho de Melo while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd) 1511da177e4SLinus Torvalds cwnd >>= 1; 1521da177e4SLinus Torvalds tp->snd_cwnd = max(cwnd, restart_cwnd); 1531da177e4SLinus Torvalds tp->snd_cwnd_stamp = tcp_time_stamp; 1541da177e4SLinus Torvalds tp->snd_cwnd_used = 0; 1551da177e4SLinus Torvalds } 1561da177e4SLinus Torvalds 15767edfef7SAndi Kleen /* Congestion state accounting after a packet has been sent. */ 15840efc6faSStephen Hemminger static void tcp_event_data_sent(struct tcp_sock *tp, 159cf533ea5SEric Dumazet struct sock *sk) 1601da177e4SLinus Torvalds { 161463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 162463c84b9SArnaldo Carvalho de Melo const u32 now = tcp_time_stamp; 1631da177e4SLinus Torvalds 16405c5a46dSNeal Cardwell if (tcp_packets_in_flight(tp) == 0) 16505c5a46dSNeal Cardwell tcp_ca_event(sk, CA_EVENT_TX_START); 16605c5a46dSNeal Cardwell 1671da177e4SLinus Torvalds tp->lsndtime = now; 1681da177e4SLinus Torvalds 1691da177e4SLinus Torvalds /* If it is a reply for ato after last received 1701da177e4SLinus Torvalds * packet, enter pingpong mode. 1711da177e4SLinus Torvalds */ 1722251ae46SJon Maxwell if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato) 173463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pingpong = 1; 1741da177e4SLinus Torvalds } 1751da177e4SLinus Torvalds 17667edfef7SAndi Kleen /* Account for an ACK we sent. */ 17740efc6faSStephen Hemminger static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts) 1781da177e4SLinus Torvalds { 179463c84b9SArnaldo Carvalho de Melo tcp_dec_quickack_mode(sk, pkts); 180463c84b9SArnaldo Carvalho de Melo inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); 1811da177e4SLinus Torvalds } 1821da177e4SLinus Torvalds 18385f16525SYuchung Cheng 18485f16525SYuchung Cheng u32 tcp_default_init_rwnd(u32 mss) 18585f16525SYuchung Cheng { 18685f16525SYuchung Cheng /* Initial receive window should be twice of TCP_INIT_CWND to 1879ef71e0cSWeiping Pan * enable proper sending of new unsent data during fast recovery 18885f16525SYuchung Cheng * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a 18985f16525SYuchung Cheng * limit when mss is larger than 1460. 19085f16525SYuchung Cheng */ 19185f16525SYuchung Cheng u32 init_rwnd = TCP_INIT_CWND * 2; 19285f16525SYuchung Cheng 19385f16525SYuchung Cheng if (mss > 1460) 19485f16525SYuchung Cheng init_rwnd = max((1460 * init_rwnd) / mss, 2U); 19585f16525SYuchung Cheng return init_rwnd; 19685f16525SYuchung Cheng } 19785f16525SYuchung Cheng 1981da177e4SLinus Torvalds /* Determine a window scaling and initial window to offer. 1991da177e4SLinus Torvalds * Based on the assumption that the given amount of space 2001da177e4SLinus Torvalds * will be offered. Store the results in the tp structure. 2011da177e4SLinus Torvalds * NOTE: for smooth operation initial space offering should 2021da177e4SLinus Torvalds * be a multiple of mss if possible. We assume here that mss >= 1. 2031da177e4SLinus Torvalds * This MUST be enforced by all callers. 2041da177e4SLinus Torvalds */ 2051da177e4SLinus Torvalds void tcp_select_initial_window(int __space, __u32 mss, 2061da177e4SLinus Torvalds __u32 *rcv_wnd, __u32 *window_clamp, 20731d12926Slaurent chavey int wscale_ok, __u8 *rcv_wscale, 20831d12926Slaurent chavey __u32 init_rcv_wnd) 2091da177e4SLinus Torvalds { 2101da177e4SLinus Torvalds unsigned int space = (__space < 0 ? 0 : __space); 2111da177e4SLinus Torvalds 2121da177e4SLinus Torvalds /* If no clamp set the clamp to the max possible scaled window */ 2131da177e4SLinus Torvalds if (*window_clamp == 0) 2141da177e4SLinus Torvalds (*window_clamp) = (65535 << 14); 2151da177e4SLinus Torvalds space = min(*window_clamp, space); 2161da177e4SLinus Torvalds 2171da177e4SLinus Torvalds /* Quantize space offering to a multiple of mss if possible. */ 2181da177e4SLinus Torvalds if (space > mss) 2191da177e4SLinus Torvalds space = (space / mss) * mss; 2201da177e4SLinus Torvalds 2211da177e4SLinus Torvalds /* NOTE: offering an initial window larger than 32767 22215d99e02SRick Jones * will break some buggy TCP stacks. If the admin tells us 22315d99e02SRick Jones * it is likely we could be speaking with such a buggy stack 22415d99e02SRick Jones * we will truncate our initial window offering to 32K-1 22515d99e02SRick Jones * unless the remote has sent us a window scaling option, 22615d99e02SRick Jones * which we interpret as a sign the remote TCP is not 22715d99e02SRick Jones * misinterpreting the window field as a signed quantity. 2281da177e4SLinus Torvalds */ 22915d99e02SRick Jones if (sysctl_tcp_workaround_signed_windows) 2301da177e4SLinus Torvalds (*rcv_wnd) = min(space, MAX_TCP_WINDOW); 23115d99e02SRick Jones else 23215d99e02SRick Jones (*rcv_wnd) = space; 23315d99e02SRick Jones 2341da177e4SLinus Torvalds (*rcv_wscale) = 0; 2351da177e4SLinus Torvalds if (wscale_ok) { 2361da177e4SLinus Torvalds /* Set window scaling on max possible window 2371da177e4SLinus Torvalds * See RFC1323 for an explanation of the limit to 14 2381da177e4SLinus Torvalds */ 239f626300aSSoheil Hassas Yeganeh space = max_t(u32, space, sysctl_tcp_rmem[2]); 240f626300aSSoheil Hassas Yeganeh space = max_t(u32, space, sysctl_rmem_max); 241316c1592SStephen Hemminger space = min_t(u32, space, *window_clamp); 2421da177e4SLinus Torvalds while (space > 65535 && (*rcv_wscale) < 14) { 2431da177e4SLinus Torvalds space >>= 1; 2441da177e4SLinus Torvalds (*rcv_wscale)++; 2451da177e4SLinus Torvalds } 2461da177e4SLinus Torvalds } 2471da177e4SLinus Torvalds 2481da177e4SLinus Torvalds if (mss > (1 << *rcv_wscale)) { 24985f16525SYuchung Cheng if (!init_rcv_wnd) /* Use default unless specified otherwise */ 25085f16525SYuchung Cheng init_rcv_wnd = tcp_default_init_rwnd(mss); 251b1afde60SNandita Dukkipati *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss); 2521da177e4SLinus Torvalds } 2531da177e4SLinus Torvalds 2541da177e4SLinus Torvalds /* Set the clamp no higher than max representable value */ 2551da177e4SLinus Torvalds (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp); 2561da177e4SLinus Torvalds } 2574bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_select_initial_window); 2581da177e4SLinus Torvalds 2591da177e4SLinus Torvalds /* Chose a new window to advertise, update state in tcp_sock for the 2601da177e4SLinus Torvalds * socket, and return result with RFC1323 scaling applied. The return 2611da177e4SLinus Torvalds * value can be stuffed directly into th->window for an outgoing 2621da177e4SLinus Torvalds * frame. 2631da177e4SLinus Torvalds */ 26440efc6faSStephen Hemminger static u16 tcp_select_window(struct sock *sk) 2651da177e4SLinus Torvalds { 2661da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2678e165e20SFlorian Westphal u32 old_win = tp->rcv_wnd; 2681da177e4SLinus Torvalds u32 cur_win = tcp_receive_window(tp); 2691da177e4SLinus Torvalds u32 new_win = __tcp_select_window(sk); 2701da177e4SLinus Torvalds 2711da177e4SLinus Torvalds /* Never shrink the offered window */ 2721da177e4SLinus Torvalds if (new_win < cur_win) { 2731da177e4SLinus Torvalds /* Danger Will Robinson! 2741da177e4SLinus Torvalds * Don't update rcv_wup/rcv_wnd here or else 2751da177e4SLinus Torvalds * we will not be able to advertise a zero 2761da177e4SLinus Torvalds * window in time. --DaveM 2771da177e4SLinus Torvalds * 2781da177e4SLinus Torvalds * Relax Will Robinson. 2791da177e4SLinus Torvalds */ 2808e165e20SFlorian Westphal if (new_win == 0) 2818e165e20SFlorian Westphal NET_INC_STATS(sock_net(sk), 2828e165e20SFlorian Westphal LINUX_MIB_TCPWANTZEROWINDOWADV); 283607bfbf2SPatrick McHardy new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale); 2841da177e4SLinus Torvalds } 2851da177e4SLinus Torvalds tp->rcv_wnd = new_win; 2861da177e4SLinus Torvalds tp->rcv_wup = tp->rcv_nxt; 2871da177e4SLinus Torvalds 2881da177e4SLinus Torvalds /* Make sure we do not exceed the maximum possible 2891da177e4SLinus Torvalds * scaled window. 2901da177e4SLinus Torvalds */ 29115d99e02SRick Jones if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows) 2921da177e4SLinus Torvalds new_win = min(new_win, MAX_TCP_WINDOW); 2931da177e4SLinus Torvalds else 2941da177e4SLinus Torvalds new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); 2951da177e4SLinus Torvalds 2961da177e4SLinus Torvalds /* RFC1323 scaling applied */ 2971da177e4SLinus Torvalds new_win >>= tp->rx_opt.rcv_wscale; 2981da177e4SLinus Torvalds 2991da177e4SLinus Torvalds /* If we advertise zero window, disable fast path. */ 3008e165e20SFlorian Westphal if (new_win == 0) { 3011da177e4SLinus Torvalds tp->pred_flags = 0; 3028e165e20SFlorian Westphal if (old_win) 3038e165e20SFlorian Westphal NET_INC_STATS(sock_net(sk), 3048e165e20SFlorian Westphal LINUX_MIB_TCPTOZEROWINDOWADV); 3058e165e20SFlorian Westphal } else if (old_win == 0) { 3068e165e20SFlorian Westphal NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV); 3078e165e20SFlorian Westphal } 3081da177e4SLinus Torvalds 3091da177e4SLinus Torvalds return new_win; 3101da177e4SLinus Torvalds } 3111da177e4SLinus Torvalds 31267edfef7SAndi Kleen /* Packet ECN state for a SYN-ACK */ 313735d3831SFlorian Westphal static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb) 314bdf1ee5dSIlpo Järvinen { 31530e502a3SDaniel Borkmann const struct tcp_sock *tp = tcp_sk(sk); 31630e502a3SDaniel Borkmann 3174de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR; 318bdf1ee5dSIlpo Järvinen if (!(tp->ecn_flags & TCP_ECN_OK)) 3194de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE; 32030e502a3SDaniel Borkmann else if (tcp_ca_needs_ecn(sk)) 32130e502a3SDaniel Borkmann INET_ECN_xmit(sk); 322bdf1ee5dSIlpo Järvinen } 323bdf1ee5dSIlpo Järvinen 32467edfef7SAndi Kleen /* Packet ECN state for a SYN. */ 325735d3831SFlorian Westphal static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb) 326bdf1ee5dSIlpo Järvinen { 327bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 328f7b3bec6SFlorian Westphal bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 || 329f7b3bec6SFlorian Westphal tcp_ca_needs_ecn(sk); 330f7b3bec6SFlorian Westphal 331f7b3bec6SFlorian Westphal if (!use_ecn) { 332f7b3bec6SFlorian Westphal const struct dst_entry *dst = __sk_dst_get(sk); 333f7b3bec6SFlorian Westphal 334f7b3bec6SFlorian Westphal if (dst && dst_feature(dst, RTAX_FEATURE_ECN)) 335f7b3bec6SFlorian Westphal use_ecn = true; 336f7b3bec6SFlorian Westphal } 337bdf1ee5dSIlpo Järvinen 338bdf1ee5dSIlpo Järvinen tp->ecn_flags = 0; 339f7b3bec6SFlorian Westphal 340f7b3bec6SFlorian Westphal if (use_ecn) { 3414de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR; 342bdf1ee5dSIlpo Järvinen tp->ecn_flags = TCP_ECN_OK; 34330e502a3SDaniel Borkmann if (tcp_ca_needs_ecn(sk)) 34430e502a3SDaniel Borkmann INET_ECN_xmit(sk); 345bdf1ee5dSIlpo Järvinen } 346bdf1ee5dSIlpo Järvinen } 347bdf1ee5dSIlpo Järvinen 34849213555SDaniel Borkmann static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb) 34949213555SDaniel Borkmann { 35049213555SDaniel Borkmann if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback) 35149213555SDaniel Borkmann /* tp->ecn_flags are cleared at a later point in time when 35249213555SDaniel Borkmann * SYN ACK is ultimatively being received. 35349213555SDaniel Borkmann */ 35449213555SDaniel Borkmann TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR); 35549213555SDaniel Borkmann } 35649213555SDaniel Borkmann 357735d3831SFlorian Westphal static void 3586ac705b1SEric Dumazet tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th) 359bdf1ee5dSIlpo Järvinen { 3606ac705b1SEric Dumazet if (inet_rsk(req)->ecn_ok) 361bdf1ee5dSIlpo Järvinen th->ece = 1; 362bdf1ee5dSIlpo Järvinen } 363bdf1ee5dSIlpo Järvinen 36467edfef7SAndi Kleen /* Set up ECN state for a packet on a ESTABLISHED socket that is about to 36567edfef7SAndi Kleen * be sent. 36667edfef7SAndi Kleen */ 367735d3831SFlorian Westphal static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb, 368ea1627c2SEric Dumazet struct tcphdr *th, int tcp_header_len) 369bdf1ee5dSIlpo Järvinen { 370bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 371bdf1ee5dSIlpo Järvinen 372bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_OK) { 373bdf1ee5dSIlpo Järvinen /* Not-retransmitted data segment: set ECT and inject CWR. */ 374bdf1ee5dSIlpo Järvinen if (skb->len != tcp_header_len && 375bdf1ee5dSIlpo Järvinen !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { 376bdf1ee5dSIlpo Järvinen INET_ECN_xmit(sk); 377bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) { 378bdf1ee5dSIlpo Järvinen tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; 379ea1627c2SEric Dumazet th->cwr = 1; 380bdf1ee5dSIlpo Järvinen skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; 381bdf1ee5dSIlpo Järvinen } 38230e502a3SDaniel Borkmann } else if (!tcp_ca_needs_ecn(sk)) { 383bdf1ee5dSIlpo Järvinen /* ACK or retransmitted segment: clear ECT|CE */ 384bdf1ee5dSIlpo Järvinen INET_ECN_dontxmit(sk); 385bdf1ee5dSIlpo Järvinen } 386bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) 387ea1627c2SEric Dumazet th->ece = 1; 388bdf1ee5dSIlpo Järvinen } 389bdf1ee5dSIlpo Järvinen } 390bdf1ee5dSIlpo Järvinen 391e870a8efSIlpo Järvinen /* Constructs common control bits of non-data skb. If SYN/FIN is present, 392e870a8efSIlpo Järvinen * auto increment end seqno. 393e870a8efSIlpo Järvinen */ 394e870a8efSIlpo Järvinen static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags) 395e870a8efSIlpo Järvinen { 3962e8e18efSDavid S. Miller skb->ip_summed = CHECKSUM_PARTIAL; 397e870a8efSIlpo Järvinen skb->csum = 0; 398e870a8efSIlpo Järvinen 3994de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags; 400e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->sacked = 0; 401e870a8efSIlpo Järvinen 402cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, 1); 403e870a8efSIlpo Järvinen 404e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->seq = seq; 405a3433f35SChangli Gao if (flags & (TCPHDR_SYN | TCPHDR_FIN)) 406e870a8efSIlpo Järvinen seq++; 407e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->end_seq = seq; 408e870a8efSIlpo Järvinen } 409e870a8efSIlpo Järvinen 410a2a385d6SEric Dumazet static inline bool tcp_urg_mode(const struct tcp_sock *tp) 41133f5f57eSIlpo Järvinen { 41233f5f57eSIlpo Järvinen return tp->snd_una != tp->snd_up; 41333f5f57eSIlpo Järvinen } 41433f5f57eSIlpo Järvinen 41533ad798cSAdam Langley #define OPTION_SACK_ADVERTISE (1 << 0) 41633ad798cSAdam Langley #define OPTION_TS (1 << 1) 41733ad798cSAdam Langley #define OPTION_MD5 (1 << 2) 41889e95a61SOri Finkelman #define OPTION_WSCALE (1 << 3) 4192100c8d2SYuchung Cheng #define OPTION_FAST_OPEN_COOKIE (1 << 8) 42033ad798cSAdam Langley 42133ad798cSAdam Langley struct tcp_out_options { 4222100c8d2SYuchung Cheng u16 options; /* bit field of OPTION_* */ 4232100c8d2SYuchung Cheng u16 mss; /* 0 to disable */ 42433ad798cSAdam Langley u8 ws; /* window scale, 0 to disable */ 42533ad798cSAdam Langley u8 num_sack_blocks; /* number of SACK blocks to include */ 426bd0388aeSWilliam Allen Simpson u8 hash_size; /* bytes in hash_location */ 427bd0388aeSWilliam Allen Simpson __u8 *hash_location; /* temporary pointer, overloaded */ 4282100c8d2SYuchung Cheng __u32 tsval, tsecr; /* need to include OPTION_TS */ 4292100c8d2SYuchung Cheng struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */ 43033ad798cSAdam Langley }; 43133ad798cSAdam Langley 43267edfef7SAndi Kleen /* Write previously computed TCP options to the packet. 43367edfef7SAndi Kleen * 43467edfef7SAndi Kleen * Beware: Something in the Internet is very sensitive to the ordering of 435fd6149d3SIlpo Järvinen * TCP options, we learned this through the hard way, so be careful here. 436fd6149d3SIlpo Järvinen * Luckily we can at least blame others for their non-compliance but from 4378e3bff96Sstephen hemminger * inter-operability perspective it seems that we're somewhat stuck with 438fd6149d3SIlpo Järvinen * the ordering which we have been using if we want to keep working with 439fd6149d3SIlpo Järvinen * those broken things (not that it currently hurts anybody as there isn't 440fd6149d3SIlpo Järvinen * particular reason why the ordering would need to be changed). 441fd6149d3SIlpo Järvinen * 442fd6149d3SIlpo Järvinen * At least SACK_PERM as the first option is known to lead to a disaster 443fd6149d3SIlpo Järvinen * (but it may well be that other scenarios fail similarly). 444fd6149d3SIlpo Järvinen */ 44533ad798cSAdam Langley static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp, 446bd0388aeSWilliam Allen Simpson struct tcp_out_options *opts) 447bd0388aeSWilliam Allen Simpson { 4482100c8d2SYuchung Cheng u16 options = opts->options; /* mungable copy */ 449bd0388aeSWilliam Allen Simpson 450bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_MD5 & options)) { 4511a2c6181SChristoph Paasch *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | 4521a2c6181SChristoph Paasch (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG); 453bd0388aeSWilliam Allen Simpson /* overload cookie hash location */ 454bd0388aeSWilliam Allen Simpson opts->hash_location = (__u8 *)ptr; 45533ad798cSAdam Langley ptr += 4; 45633ad798cSAdam Langley } 45733ad798cSAdam Langley 458fd6149d3SIlpo Järvinen if (unlikely(opts->mss)) { 459fd6149d3SIlpo Järvinen *ptr++ = htonl((TCPOPT_MSS << 24) | 460fd6149d3SIlpo Järvinen (TCPOLEN_MSS << 16) | 461fd6149d3SIlpo Järvinen opts->mss); 462fd6149d3SIlpo Järvinen } 463fd6149d3SIlpo Järvinen 464bd0388aeSWilliam Allen Simpson if (likely(OPTION_TS & options)) { 465bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 46633ad798cSAdam Langley *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | 46733ad798cSAdam Langley (TCPOLEN_SACK_PERM << 16) | 46833ad798cSAdam Langley (TCPOPT_TIMESTAMP << 8) | 46933ad798cSAdam Langley TCPOLEN_TIMESTAMP); 470bd0388aeSWilliam Allen Simpson options &= ~OPTION_SACK_ADVERTISE; 47133ad798cSAdam Langley } else { 472496c98dfSYOSHIFUJI Hideaki *ptr++ = htonl((TCPOPT_NOP << 24) | 47340efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 47440efc6faSStephen Hemminger (TCPOPT_TIMESTAMP << 8) | 47540efc6faSStephen Hemminger TCPOLEN_TIMESTAMP); 47640efc6faSStephen Hemminger } 47733ad798cSAdam Langley *ptr++ = htonl(opts->tsval); 47833ad798cSAdam Langley *ptr++ = htonl(opts->tsecr); 47933ad798cSAdam Langley } 48033ad798cSAdam Langley 481bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 48233ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 48333ad798cSAdam Langley (TCPOPT_NOP << 16) | 48433ad798cSAdam Langley (TCPOPT_SACK_PERM << 8) | 48533ad798cSAdam Langley TCPOLEN_SACK_PERM); 48633ad798cSAdam Langley } 48733ad798cSAdam Langley 488bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_WSCALE & options)) { 48933ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 49033ad798cSAdam Langley (TCPOPT_WINDOW << 16) | 49133ad798cSAdam Langley (TCPOLEN_WINDOW << 8) | 49233ad798cSAdam Langley opts->ws); 49333ad798cSAdam Langley } 49433ad798cSAdam Langley 49533ad798cSAdam Langley if (unlikely(opts->num_sack_blocks)) { 49633ad798cSAdam Langley struct tcp_sack_block *sp = tp->rx_opt.dsack ? 49733ad798cSAdam Langley tp->duplicate_sack : tp->selective_acks; 49840efc6faSStephen Hemminger int this_sack; 49940efc6faSStephen Hemminger 50040efc6faSStephen Hemminger *ptr++ = htonl((TCPOPT_NOP << 24) | 50140efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 50240efc6faSStephen Hemminger (TCPOPT_SACK << 8) | 50333ad798cSAdam Langley (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * 50440efc6faSStephen Hemminger TCPOLEN_SACK_PERBLOCK))); 5052de979bdSStephen Hemminger 50633ad798cSAdam Langley for (this_sack = 0; this_sack < opts->num_sack_blocks; 50733ad798cSAdam Langley ++this_sack) { 50840efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].start_seq); 50940efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].end_seq); 51040efc6faSStephen Hemminger } 5112de979bdSStephen Hemminger 51240efc6faSStephen Hemminger tp->rx_opt.dsack = 0; 51340efc6faSStephen Hemminger } 5142100c8d2SYuchung Cheng 5152100c8d2SYuchung Cheng if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) { 5162100c8d2SYuchung Cheng struct tcp_fastopen_cookie *foc = opts->fastopen_cookie; 5177f9b838bSDaniel Lee u8 *p = (u8 *)ptr; 5187f9b838bSDaniel Lee u32 len; /* Fast Open option length */ 5192100c8d2SYuchung Cheng 5207f9b838bSDaniel Lee if (foc->exp) { 5217f9b838bSDaniel Lee len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len; 5227f9b838bSDaniel Lee *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) | 5232100c8d2SYuchung Cheng TCPOPT_FASTOPEN_MAGIC); 5247f9b838bSDaniel Lee p += TCPOLEN_EXP_FASTOPEN_BASE; 5257f9b838bSDaniel Lee } else { 5267f9b838bSDaniel Lee len = TCPOLEN_FASTOPEN_BASE + foc->len; 5277f9b838bSDaniel Lee *p++ = TCPOPT_FASTOPEN; 5287f9b838bSDaniel Lee *p++ = len; 5292100c8d2SYuchung Cheng } 5307f9b838bSDaniel Lee 5317f9b838bSDaniel Lee memcpy(p, foc->val, foc->len); 5327f9b838bSDaniel Lee if ((len & 3) == 2) { 5337f9b838bSDaniel Lee p[foc->len] = TCPOPT_NOP; 5347f9b838bSDaniel Lee p[foc->len + 1] = TCPOPT_NOP; 5357f9b838bSDaniel Lee } 5367f9b838bSDaniel Lee ptr += (len + 3) >> 2; 5372100c8d2SYuchung Cheng } 53840efc6faSStephen Hemminger } 53940efc6faSStephen Hemminger 54067edfef7SAndi Kleen /* Compute TCP options for SYN packets. This is not the final 54167edfef7SAndi Kleen * network wire format yet. 54267edfef7SAndi Kleen */ 54395c96174SEric Dumazet static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb, 54433ad798cSAdam Langley struct tcp_out_options *opts, 545cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 546cf533ea5SEric Dumazet { 54733ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 54895c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 549783237e8SYuchung Cheng struct tcp_fastopen_request *fastopen = tp->fastopen_req; 55033ad798cSAdam Langley 551cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 55233ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 55333ad798cSAdam Langley if (*md5) { 55433ad798cSAdam Langley opts->options |= OPTION_MD5; 555bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 556cfb6eeb4SYOSHIFUJI Hideaki } 55733ad798cSAdam Langley #else 55833ad798cSAdam Langley *md5 = NULL; 559cfb6eeb4SYOSHIFUJI Hideaki #endif 56033ad798cSAdam Langley 56133ad798cSAdam Langley /* We always get an MSS option. The option bytes which will be seen in 56233ad798cSAdam Langley * normal data packets should timestamps be used, must be in the MSS 56333ad798cSAdam Langley * advertised. But we subtract them from tp->mss_cache so that 56433ad798cSAdam Langley * calculations in tcp_sendmsg are simpler etc. So account for this 56533ad798cSAdam Langley * fact here if necessary. If we don't do this correctly, as a 56633ad798cSAdam Langley * receiver we won't recognize data packets as being full sized when we 56733ad798cSAdam Langley * should, and thus we won't abide by the delayed ACK rules correctly. 56833ad798cSAdam Langley * SACKs don't matter, we never delay an ACK when we have any of those 56933ad798cSAdam Langley * going out. */ 57033ad798cSAdam Langley opts->mss = tcp_advertise_mss(sk); 571bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 57233ad798cSAdam Langley 57351456b29SIan Morris if (likely(sysctl_tcp_timestamps && !*md5)) { 57433ad798cSAdam Langley opts->options |= OPTION_TS; 5757faee5c0SEric Dumazet opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset; 57633ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 577bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 57833ad798cSAdam Langley } 579bb5b7c11SDavid S. Miller if (likely(sysctl_tcp_window_scaling)) { 58033ad798cSAdam Langley opts->ws = tp->rx_opt.rcv_wscale; 58189e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 582bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 58333ad798cSAdam Langley } 584bb5b7c11SDavid S. Miller if (likely(sysctl_tcp_sack)) { 58533ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 586b32d1310SDavid S. Miller if (unlikely(!(OPTION_TS & opts->options))) 587bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 58833ad798cSAdam Langley } 58933ad798cSAdam Langley 590783237e8SYuchung Cheng if (fastopen && fastopen->cookie.len >= 0) { 5912646c831SDaniel Lee u32 need = fastopen->cookie.len; 5922646c831SDaniel Lee 5932646c831SDaniel Lee need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE : 5942646c831SDaniel Lee TCPOLEN_FASTOPEN_BASE; 595783237e8SYuchung Cheng need = (need + 3) & ~3U; /* Align to 32 bits */ 596783237e8SYuchung Cheng if (remaining >= need) { 597783237e8SYuchung Cheng opts->options |= OPTION_FAST_OPEN_COOKIE; 598783237e8SYuchung Cheng opts->fastopen_cookie = &fastopen->cookie; 599783237e8SYuchung Cheng remaining -= need; 600783237e8SYuchung Cheng tp->syn_fastopen = 1; 6012646c831SDaniel Lee tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0; 602783237e8SYuchung Cheng } 603783237e8SYuchung Cheng } 604bd0388aeSWilliam Allen Simpson 605bd0388aeSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 60633ad798cSAdam Langley } 60733ad798cSAdam Langley 60867edfef7SAndi Kleen /* Set up TCP options for SYN-ACKs. */ 60937bfbddaSEric Dumazet static unsigned int tcp_synack_options(struct request_sock *req, 61095c96174SEric Dumazet unsigned int mss, struct sk_buff *skb, 61133ad798cSAdam Langley struct tcp_out_options *opts, 61280f03e27SEric Dumazet const struct tcp_md5sig_key *md5, 6138336886fSJerry Chu struct tcp_fastopen_cookie *foc) 6144957faadSWilliam Allen Simpson { 61533ad798cSAdam Langley struct inet_request_sock *ireq = inet_rsk(req); 61695c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 61733ad798cSAdam Langley 61833ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 61980f03e27SEric Dumazet if (md5) { 62033ad798cSAdam Langley opts->options |= OPTION_MD5; 6214957faadSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 6224957faadSWilliam Allen Simpson 6234957faadSWilliam Allen Simpson /* We can't fit any SACK blocks in a packet with MD5 + TS 6244957faadSWilliam Allen Simpson * options. There was discussion about disabling SACK 6254957faadSWilliam Allen Simpson * rather than TS in order to fit in better with old, 6264957faadSWilliam Allen Simpson * buggy kernels, but that was deemed to be unnecessary. 6274957faadSWilliam Allen Simpson */ 628de213e5eSEric Dumazet ireq->tstamp_ok &= !ireq->sack_ok; 62933ad798cSAdam Langley } 63033ad798cSAdam Langley #endif 63133ad798cSAdam Langley 6324957faadSWilliam Allen Simpson /* We always send an MSS option. */ 63333ad798cSAdam Langley opts->mss = mss; 6344957faadSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 63533ad798cSAdam Langley 63633ad798cSAdam Langley if (likely(ireq->wscale_ok)) { 63733ad798cSAdam Langley opts->ws = ireq->rcv_wscale; 63889e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 6394957faadSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 64033ad798cSAdam Langley } 641de213e5eSEric Dumazet if (likely(ireq->tstamp_ok)) { 64233ad798cSAdam Langley opts->options |= OPTION_TS; 6437faee5c0SEric Dumazet opts->tsval = tcp_skb_timestamp(skb); 64433ad798cSAdam Langley opts->tsecr = req->ts_recent; 6454957faadSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 64633ad798cSAdam Langley } 64733ad798cSAdam Langley if (likely(ireq->sack_ok)) { 64833ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 649de213e5eSEric Dumazet if (unlikely(!ireq->tstamp_ok)) 6504957faadSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 65133ad798cSAdam Langley } 6527f9b838bSDaniel Lee if (foc != NULL && foc->len >= 0) { 6537f9b838bSDaniel Lee u32 need = foc->len; 6547f9b838bSDaniel Lee 6557f9b838bSDaniel Lee need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE : 6567f9b838bSDaniel Lee TCPOLEN_FASTOPEN_BASE; 6578336886fSJerry Chu need = (need + 3) & ~3U; /* Align to 32 bits */ 6588336886fSJerry Chu if (remaining >= need) { 6598336886fSJerry Chu opts->options |= OPTION_FAST_OPEN_COOKIE; 6608336886fSJerry Chu opts->fastopen_cookie = foc; 6618336886fSJerry Chu remaining -= need; 6628336886fSJerry Chu } 6638336886fSJerry Chu } 6644957faadSWilliam Allen Simpson 6654957faadSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 66633ad798cSAdam Langley } 66733ad798cSAdam Langley 66867edfef7SAndi Kleen /* Compute TCP options for ESTABLISHED sockets. This is not the 66967edfef7SAndi Kleen * final wire format yet. 67067edfef7SAndi Kleen */ 67195c96174SEric Dumazet static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb, 67233ad798cSAdam Langley struct tcp_out_options *opts, 673cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 674cf533ea5SEric Dumazet { 67533ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 67695c96174SEric Dumazet unsigned int size = 0; 677cabeccbdSIlpo Järvinen unsigned int eff_sacks; 67833ad798cSAdam Langley 6795843ef42SAndi Kleen opts->options = 0; 6805843ef42SAndi Kleen 68133ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 68233ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 68333ad798cSAdam Langley if (unlikely(*md5)) { 68433ad798cSAdam Langley opts->options |= OPTION_MD5; 68533ad798cSAdam Langley size += TCPOLEN_MD5SIG_ALIGNED; 68633ad798cSAdam Langley } 68733ad798cSAdam Langley #else 68833ad798cSAdam Langley *md5 = NULL; 68933ad798cSAdam Langley #endif 69033ad798cSAdam Langley 69133ad798cSAdam Langley if (likely(tp->rx_opt.tstamp_ok)) { 69233ad798cSAdam Langley opts->options |= OPTION_TS; 6937faee5c0SEric Dumazet opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0; 69433ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 69533ad798cSAdam Langley size += TCPOLEN_TSTAMP_ALIGNED; 69633ad798cSAdam Langley } 69733ad798cSAdam Langley 698cabeccbdSIlpo Järvinen eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; 699cabeccbdSIlpo Järvinen if (unlikely(eff_sacks)) { 70095c96174SEric Dumazet const unsigned int remaining = MAX_TCP_OPTION_SPACE - size; 70133ad798cSAdam Langley opts->num_sack_blocks = 70295c96174SEric Dumazet min_t(unsigned int, eff_sacks, 70333ad798cSAdam Langley (remaining - TCPOLEN_SACK_BASE_ALIGNED) / 70433ad798cSAdam Langley TCPOLEN_SACK_PERBLOCK); 70533ad798cSAdam Langley size += TCPOLEN_SACK_BASE_ALIGNED + 70633ad798cSAdam Langley opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; 70733ad798cSAdam Langley } 70833ad798cSAdam Langley 70933ad798cSAdam Langley return size; 71040efc6faSStephen Hemminger } 7111da177e4SLinus Torvalds 71246d3ceabSEric Dumazet 71346d3ceabSEric Dumazet /* TCP SMALL QUEUES (TSQ) 71446d3ceabSEric Dumazet * 71546d3ceabSEric Dumazet * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev) 71646d3ceabSEric Dumazet * to reduce RTT and bufferbloat. 71746d3ceabSEric Dumazet * We do this using a special skb destructor (tcp_wfree). 71846d3ceabSEric Dumazet * 71946d3ceabSEric Dumazet * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb 72046d3ceabSEric Dumazet * needs to be reallocated in a driver. 7218e3bff96Sstephen hemminger * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc 72246d3ceabSEric Dumazet * 72346d3ceabSEric Dumazet * Since transmit from skb destructor is forbidden, we use a tasklet 72446d3ceabSEric Dumazet * to process all sockets that eventually need to send more skbs. 72546d3ceabSEric Dumazet * We use one tasklet per cpu, with its own queue of sockets. 72646d3ceabSEric Dumazet */ 72746d3ceabSEric Dumazet struct tsq_tasklet { 72846d3ceabSEric Dumazet struct tasklet_struct tasklet; 72946d3ceabSEric Dumazet struct list_head head; /* queue of tcp sockets */ 73046d3ceabSEric Dumazet }; 73146d3ceabSEric Dumazet static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet); 73246d3ceabSEric Dumazet 7336f458dfbSEric Dumazet static void tcp_tsq_handler(struct sock *sk) 7346f458dfbSEric Dumazet { 7356f458dfbSEric Dumazet if ((1 << sk->sk_state) & 7366f458dfbSEric Dumazet (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING | 737f9616c35SEric Dumazet TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) { 738f9616c35SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 739f9616c35SEric Dumazet 740f9616c35SEric Dumazet if (tp->lost_out > tp->retrans_out && 741f9616c35SEric Dumazet tp->snd_cwnd > tcp_packets_in_flight(tp)) 742f9616c35SEric Dumazet tcp_xmit_retransmit_queue(sk); 743f9616c35SEric Dumazet 744f9616c35SEric Dumazet tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle, 745bf06200eSJohn Ogness 0, GFP_ATOMIC); 7466f458dfbSEric Dumazet } 747f9616c35SEric Dumazet } 74846d3ceabSEric Dumazet /* 7498e3bff96Sstephen hemminger * One tasklet per cpu tries to send more skbs. 75046d3ceabSEric Dumazet * We run in tasklet context but need to disable irqs when 7518e3bff96Sstephen hemminger * transferring tsq->head because tcp_wfree() might 75246d3ceabSEric Dumazet * interrupt us (non NAPI drivers) 75346d3ceabSEric Dumazet */ 75446d3ceabSEric Dumazet static void tcp_tasklet_func(unsigned long data) 75546d3ceabSEric Dumazet { 75646d3ceabSEric Dumazet struct tsq_tasklet *tsq = (struct tsq_tasklet *)data; 75746d3ceabSEric Dumazet LIST_HEAD(list); 75846d3ceabSEric Dumazet unsigned long flags; 75946d3ceabSEric Dumazet struct list_head *q, *n; 76046d3ceabSEric Dumazet struct tcp_sock *tp; 76146d3ceabSEric Dumazet struct sock *sk; 76246d3ceabSEric Dumazet 76346d3ceabSEric Dumazet local_irq_save(flags); 76446d3ceabSEric Dumazet list_splice_init(&tsq->head, &list); 76546d3ceabSEric Dumazet local_irq_restore(flags); 76646d3ceabSEric Dumazet 76746d3ceabSEric Dumazet list_for_each_safe(q, n, &list) { 76846d3ceabSEric Dumazet tp = list_entry(q, struct tcp_sock, tsq_node); 76946d3ceabSEric Dumazet list_del(&tp->tsq_node); 77046d3ceabSEric Dumazet 77146d3ceabSEric Dumazet sk = (struct sock *)tp; 77246d3ceabSEric Dumazet bh_lock_sock(sk); 77346d3ceabSEric Dumazet 77446d3ceabSEric Dumazet if (!sock_owned_by_user(sk)) { 7756f458dfbSEric Dumazet tcp_tsq_handler(sk); 77646d3ceabSEric Dumazet } else { 77746d3ceabSEric Dumazet /* defer the work to tcp_release_cb() */ 7786f458dfbSEric Dumazet set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags); 77946d3ceabSEric Dumazet } 78046d3ceabSEric Dumazet bh_unlock_sock(sk); 78146d3ceabSEric Dumazet 78246d3ceabSEric Dumazet clear_bit(TSQ_QUEUED, &tp->tsq_flags); 78346d3ceabSEric Dumazet sk_free(sk); 78446d3ceabSEric Dumazet } 78546d3ceabSEric Dumazet } 78646d3ceabSEric Dumazet 7876f458dfbSEric Dumazet #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \ 7886f458dfbSEric Dumazet (1UL << TCP_WRITE_TIMER_DEFERRED) | \ 789563d34d0SEric Dumazet (1UL << TCP_DELACK_TIMER_DEFERRED) | \ 790563d34d0SEric Dumazet (1UL << TCP_MTU_REDUCED_DEFERRED)) 79146d3ceabSEric Dumazet /** 79246d3ceabSEric Dumazet * tcp_release_cb - tcp release_sock() callback 79346d3ceabSEric Dumazet * @sk: socket 79446d3ceabSEric Dumazet * 79546d3ceabSEric Dumazet * called from release_sock() to perform protocol dependent 79646d3ceabSEric Dumazet * actions before socket release. 79746d3ceabSEric Dumazet */ 79846d3ceabSEric Dumazet void tcp_release_cb(struct sock *sk) 79946d3ceabSEric Dumazet { 80046d3ceabSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 8016f458dfbSEric Dumazet unsigned long flags, nflags; 80246d3ceabSEric Dumazet 8036f458dfbSEric Dumazet /* perform an atomic operation only if at least one flag is set */ 8046f458dfbSEric Dumazet do { 8056f458dfbSEric Dumazet flags = tp->tsq_flags; 8066f458dfbSEric Dumazet if (!(flags & TCP_DEFERRED_ALL)) 8076f458dfbSEric Dumazet return; 8086f458dfbSEric Dumazet nflags = flags & ~TCP_DEFERRED_ALL; 8096f458dfbSEric Dumazet } while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags); 8106f458dfbSEric Dumazet 8116f458dfbSEric Dumazet if (flags & (1UL << TCP_TSQ_DEFERRED)) 8126f458dfbSEric Dumazet tcp_tsq_handler(sk); 8136f458dfbSEric Dumazet 814c3f9b018SEric Dumazet /* Here begins the tricky part : 815c3f9b018SEric Dumazet * We are called from release_sock() with : 816c3f9b018SEric Dumazet * 1) BH disabled 817c3f9b018SEric Dumazet * 2) sk_lock.slock spinlock held 818c3f9b018SEric Dumazet * 3) socket owned by us (sk->sk_lock.owned == 1) 819c3f9b018SEric Dumazet * 820c3f9b018SEric Dumazet * But following code is meant to be called from BH handlers, 821c3f9b018SEric Dumazet * so we should keep BH disabled, but early release socket ownership 822c3f9b018SEric Dumazet */ 823c3f9b018SEric Dumazet sock_release_ownership(sk); 824c3f9b018SEric Dumazet 825144d56e9SEric Dumazet if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) { 8266f458dfbSEric Dumazet tcp_write_timer_handler(sk); 827144d56e9SEric Dumazet __sock_put(sk); 828144d56e9SEric Dumazet } 829144d56e9SEric Dumazet if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) { 8306f458dfbSEric Dumazet tcp_delack_timer_handler(sk); 831144d56e9SEric Dumazet __sock_put(sk); 832144d56e9SEric Dumazet } 833144d56e9SEric Dumazet if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) { 8344fab9071SNeal Cardwell inet_csk(sk)->icsk_af_ops->mtu_reduced(sk); 835144d56e9SEric Dumazet __sock_put(sk); 836144d56e9SEric Dumazet } 83746d3ceabSEric Dumazet } 83846d3ceabSEric Dumazet EXPORT_SYMBOL(tcp_release_cb); 83946d3ceabSEric Dumazet 84046d3ceabSEric Dumazet void __init tcp_tasklet_init(void) 84146d3ceabSEric Dumazet { 84246d3ceabSEric Dumazet int i; 84346d3ceabSEric Dumazet 84446d3ceabSEric Dumazet for_each_possible_cpu(i) { 84546d3ceabSEric Dumazet struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i); 84646d3ceabSEric Dumazet 84746d3ceabSEric Dumazet INIT_LIST_HEAD(&tsq->head); 84846d3ceabSEric Dumazet tasklet_init(&tsq->tasklet, 84946d3ceabSEric Dumazet tcp_tasklet_func, 85046d3ceabSEric Dumazet (unsigned long)tsq); 85146d3ceabSEric Dumazet } 85246d3ceabSEric Dumazet } 85346d3ceabSEric Dumazet 85446d3ceabSEric Dumazet /* 85546d3ceabSEric Dumazet * Write buffer destructor automatically called from kfree_skb. 8568e3bff96Sstephen hemminger * We can't xmit new skbs from this context, as we might already 85746d3ceabSEric Dumazet * hold qdisc lock. 85846d3ceabSEric Dumazet */ 859d6a4a104SEric Dumazet void tcp_wfree(struct sk_buff *skb) 86046d3ceabSEric Dumazet { 86146d3ceabSEric Dumazet struct sock *sk = skb->sk; 86246d3ceabSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 8639b462d02SEric Dumazet int wmem; 8649b462d02SEric Dumazet 8659b462d02SEric Dumazet /* Keep one reference on sk_wmem_alloc. 8669b462d02SEric Dumazet * Will be released by sk_free() from here or tcp_tasklet_func() 8679b462d02SEric Dumazet */ 8689b462d02SEric Dumazet wmem = atomic_sub_return(skb->truesize - 1, &sk->sk_wmem_alloc); 8699b462d02SEric Dumazet 8709b462d02SEric Dumazet /* If this softirq is serviced by ksoftirqd, we are likely under stress. 8719b462d02SEric Dumazet * Wait until our queues (qdisc + devices) are drained. 8729b462d02SEric Dumazet * This gives : 8739b462d02SEric Dumazet * - less callbacks to tcp_write_xmit(), reducing stress (batches) 8749b462d02SEric Dumazet * - chance for incoming ACK (processed by another cpu maybe) 8759b462d02SEric Dumazet * to migrate this flow (skb->ooo_okay will be eventually set) 8769b462d02SEric Dumazet */ 8779b462d02SEric Dumazet if (wmem >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current) 8789b462d02SEric Dumazet goto out; 87946d3ceabSEric Dumazet 88046d3ceabSEric Dumazet if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) && 88146d3ceabSEric Dumazet !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) { 88246d3ceabSEric Dumazet unsigned long flags; 88346d3ceabSEric Dumazet struct tsq_tasklet *tsq; 88446d3ceabSEric Dumazet 88546d3ceabSEric Dumazet /* queue this socket to tasklet queue */ 88646d3ceabSEric Dumazet local_irq_save(flags); 887903ceff7SChristoph Lameter tsq = this_cpu_ptr(&tsq_tasklet); 88846d3ceabSEric Dumazet list_add(&tp->tsq_node, &tsq->head); 88946d3ceabSEric Dumazet tasklet_schedule(&tsq->tasklet); 89046d3ceabSEric Dumazet local_irq_restore(flags); 8919b462d02SEric Dumazet return; 89246d3ceabSEric Dumazet } 8939b462d02SEric Dumazet out: 8949b462d02SEric Dumazet sk_free(sk); 89546d3ceabSEric Dumazet } 89646d3ceabSEric Dumazet 8971da177e4SLinus Torvalds /* This routine actually transmits TCP packets queued in by 8981da177e4SLinus Torvalds * tcp_do_sendmsg(). This is used by both the initial 8991da177e4SLinus Torvalds * transmission and possible later retransmissions. 9001da177e4SLinus Torvalds * All SKB's seen here are completely headerless. It is our 9011da177e4SLinus Torvalds * job to build the TCP header, and pass the packet down to 9021da177e4SLinus Torvalds * IP so it can do the same plus pass the packet off to the 9031da177e4SLinus Torvalds * device. 9041da177e4SLinus Torvalds * 9051da177e4SLinus Torvalds * We are working here with either a clone of the original 9061da177e4SLinus Torvalds * SKB, or a fresh unique copy made by the retransmit engine. 9071da177e4SLinus Torvalds */ 908056834d9SIlpo Järvinen static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, 909056834d9SIlpo Järvinen gfp_t gfp_mask) 9101da177e4SLinus Torvalds { 9116687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 912dfb4b9dcSDavid S. Miller struct inet_sock *inet; 913dfb4b9dcSDavid S. Miller struct tcp_sock *tp; 914dfb4b9dcSDavid S. Miller struct tcp_skb_cb *tcb; 91533ad798cSAdam Langley struct tcp_out_options opts; 91695c96174SEric Dumazet unsigned int tcp_options_size, tcp_header_size; 917cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 9181da177e4SLinus Torvalds struct tcphdr *th; 9191da177e4SLinus Torvalds int err; 9201da177e4SLinus Torvalds 921dfb4b9dcSDavid S. Miller BUG_ON(!skb || !tcp_skb_pcount(skb)); 9226f094b9eSLawrence Brakmo tp = tcp_sk(sk); 923dfb4b9dcSDavid S. Miller 924ccdbb6e9SEric Dumazet if (clone_it) { 925740b0f18SEric Dumazet skb_mstamp_get(&skb->skb_mstamp); 9266f094b9eSLawrence Brakmo TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq 9276f094b9eSLawrence Brakmo - tp->snd_una; 928b9f64820SYuchung Cheng tcp_rate_skb_sent(sk, skb); 929dfb4b9dcSDavid S. Miller 930dfb4b9dcSDavid S. Miller if (unlikely(skb_cloned(skb))) 931dfb4b9dcSDavid S. Miller skb = pskb_copy(skb, gfp_mask); 932dfb4b9dcSDavid S. Miller else 933dfb4b9dcSDavid S. Miller skb = skb_clone(skb, gfp_mask); 934dfb4b9dcSDavid S. Miller if (unlikely(!skb)) 935dfb4b9dcSDavid S. Miller return -ENOBUFS; 936dfb4b9dcSDavid S. Miller } 937dfb4b9dcSDavid S. Miller 938dfb4b9dcSDavid S. Miller inet = inet_sk(sk); 939dfb4b9dcSDavid S. Miller tcb = TCP_SKB_CB(skb); 94033ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 9411da177e4SLinus Torvalds 9424de075e0SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) 94333ad798cSAdam Langley tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); 94433ad798cSAdam Langley else 94533ad798cSAdam Langley tcp_options_size = tcp_established_options(sk, skb, &opts, 94633ad798cSAdam Langley &md5); 94733ad798cSAdam Langley tcp_header_size = tcp_options_size + sizeof(struct tcphdr); 9481da177e4SLinus Torvalds 949547669d4SEric Dumazet /* if no packet is in qdisc/device queue, then allow XPS to select 950b2532eb9SEric Dumazet * another queue. We can be called from tcp_tsq_handler() 951b2532eb9SEric Dumazet * which holds one reference to sk_wmem_alloc. 952b2532eb9SEric Dumazet * 953b2532eb9SEric Dumazet * TODO: Ideally, in-flight pure ACK packets should not matter here. 954b2532eb9SEric Dumazet * One way to get this would be to set skb->truesize = 2 on them. 955547669d4SEric Dumazet */ 956b2532eb9SEric Dumazet skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1); 9571da177e4SLinus Torvalds 958aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 959aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 96046d3ceabSEric Dumazet 96146d3ceabSEric Dumazet skb_orphan(skb); 96246d3ceabSEric Dumazet skb->sk = sk; 9631d2077acSEric Dumazet skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree; 964b73c3d0eSTom Herbert skb_set_hash_from_sk(skb, sk); 96546d3ceabSEric Dumazet atomic_add(skb->truesize, &sk->sk_wmem_alloc); 9661da177e4SLinus Torvalds 9671da177e4SLinus Torvalds /* Build TCP header and checksum it. */ 968ea1627c2SEric Dumazet th = (struct tcphdr *)skb->data; 969c720c7e8SEric Dumazet th->source = inet->inet_sport; 970c720c7e8SEric Dumazet th->dest = inet->inet_dport; 9711da177e4SLinus Torvalds th->seq = htonl(tcb->seq); 9721da177e4SLinus Torvalds th->ack_seq = htonl(tp->rcv_nxt); 973df7a3b07SAl Viro *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | 9744de075e0SEric Dumazet tcb->tcp_flags); 975dfb4b9dcSDavid S. Miller 9761da177e4SLinus Torvalds th->check = 0; 9771da177e4SLinus Torvalds th->urg_ptr = 0; 9781da177e4SLinus Torvalds 97933f5f57eSIlpo Järvinen /* The urg_mode check is necessary during a below snd_una win probe */ 9807691367dSHerbert Xu if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { 9817691367dSHerbert Xu if (before(tp->snd_up, tcb->seq + 0x10000)) { 9821da177e4SLinus Torvalds th->urg_ptr = htons(tp->snd_up - tcb->seq); 9831da177e4SLinus Torvalds th->urg = 1; 9847691367dSHerbert Xu } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { 9850eae88f3SEric Dumazet th->urg_ptr = htons(0xFFFF); 9867691367dSHerbert Xu th->urg = 1; 9877691367dSHerbert Xu } 9881da177e4SLinus Torvalds } 9891da177e4SLinus Torvalds 990bd0388aeSWilliam Allen Simpson tcp_options_write((__be32 *)(th + 1), tp, &opts); 99151466a75SEric Dumazet skb_shinfo(skb)->gso_type = sk->sk_gso_type; 992ea1627c2SEric Dumazet if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) { 993ea1627c2SEric Dumazet th->window = htons(tcp_select_window(sk)); 994ea1627c2SEric Dumazet tcp_ecn_send(sk, skb, th, tcp_header_size); 995ea1627c2SEric Dumazet } else { 996ea1627c2SEric Dumazet /* RFC1323: The window in SYN & SYN/ACK segments 997ea1627c2SEric Dumazet * is never scaled. 998ea1627c2SEric Dumazet */ 999ea1627c2SEric Dumazet th->window = htons(min(tp->rcv_wnd, 65535U)); 1000ea1627c2SEric Dumazet } 1001cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 1002cfb6eeb4SYOSHIFUJI Hideaki /* Calculate the MD5 hash, as we have all we need now */ 1003cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 1004a465419bSEric Dumazet sk_nocaps_add(sk, NETIF_F_GSO_MASK); 1005bd0388aeSWilliam Allen Simpson tp->af_specific->calc_md5_hash(opts.hash_location, 100639f8e58eSEric Dumazet md5, sk, skb); 1007cfb6eeb4SYOSHIFUJI Hideaki } 1008cfb6eeb4SYOSHIFUJI Hideaki #endif 1009cfb6eeb4SYOSHIFUJI Hideaki 1010bb296246SHerbert Xu icsk->icsk_af_ops->send_check(sk, skb); 10111da177e4SLinus Torvalds 10124de075e0SEric Dumazet if (likely(tcb->tcp_flags & TCPHDR_ACK)) 1013fc6415bcSDavid S. Miller tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); 10141da177e4SLinus Torvalds 1015a44d6eacSMartin KaFai Lau if (skb->len != tcp_header_size) { 1016cf533ea5SEric Dumazet tcp_event_data_sent(tp, sk); 1017a44d6eacSMartin KaFai Lau tp->data_segs_out += tcp_skb_pcount(skb); 1018a44d6eacSMartin KaFai Lau } 10191da177e4SLinus Torvalds 1020bd37a088SWei Yongjun if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) 1021aa2ea058STom Herbert TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, 1022aa2ea058STom Herbert tcp_skb_pcount(skb)); 10231da177e4SLinus Torvalds 10242efd055cSMarcelo Ricardo Leitner tp->segs_out += tcp_skb_pcount(skb); 1025f69ad292SEric Dumazet /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */ 1026cd7d8498SEric Dumazet skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb); 1027f69ad292SEric Dumazet skb_shinfo(skb)->gso_size = tcp_skb_mss(skb); 1028cd7d8498SEric Dumazet 10297faee5c0SEric Dumazet /* Our usage of tstamp should remain private */ 10307faee5c0SEric Dumazet skb->tstamp.tv64 = 0; 1031971f10ecSEric Dumazet 1032971f10ecSEric Dumazet /* Cleanup our debris for IP stacks */ 1033971f10ecSEric Dumazet memset(skb->cb, 0, max(sizeof(struct inet_skb_parm), 1034971f10ecSEric Dumazet sizeof(struct inet6_skb_parm))); 1035971f10ecSEric Dumazet 1036b0270e91SEric Dumazet err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl); 10377faee5c0SEric Dumazet 103883de47cdSHua Zhong if (likely(err <= 0)) 10391da177e4SLinus Torvalds return err; 10401da177e4SLinus Torvalds 10415ee2c941SChristoph Paasch tcp_enter_cwr(sk); 10421da177e4SLinus Torvalds 1043b9df3cb8SGerrit Renker return net_xmit_eval(err); 10441da177e4SLinus Torvalds } 10451da177e4SLinus Torvalds 104667edfef7SAndi Kleen /* This routine just queues the buffer for sending. 10471da177e4SLinus Torvalds * 10481da177e4SLinus Torvalds * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, 10491da177e4SLinus Torvalds * otherwise socket can stall. 10501da177e4SLinus Torvalds */ 10511da177e4SLinus Torvalds static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) 10521da177e4SLinus Torvalds { 10531da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 10541da177e4SLinus Torvalds 10551da177e4SLinus Torvalds /* Advance write_seq and place onto the write_queue. */ 10561da177e4SLinus Torvalds tp->write_seq = TCP_SKB_CB(skb)->end_seq; 1057f4a775d1SEric Dumazet __skb_header_release(skb); 1058fe067e8aSDavid S. Miller tcp_add_write_queue_tail(sk, skb); 10593ab224beSHideo Aoki sk->sk_wmem_queued += skb->truesize; 10603ab224beSHideo Aoki sk_mem_charge(sk, skb->truesize); 10611da177e4SLinus Torvalds } 10621da177e4SLinus Torvalds 106367edfef7SAndi Kleen /* Initialize TSO segments for a packet. */ 10645bbb432cSEric Dumazet static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now) 1065f6302d1dSDavid S. Miller { 10668f26fb1cSEric Dumazet if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) { 1067f6302d1dSDavid S. Miller /* Avoid the costly divide in the normal 1068f6302d1dSDavid S. Miller * non-TSO case. 1069f6302d1dSDavid S. Miller */ 1070cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, 1); 1071f69ad292SEric Dumazet TCP_SKB_CB(skb)->tcp_gso_size = 0; 1072f6302d1dSDavid S. Miller } else { 1073cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now)); 1074f69ad292SEric Dumazet TCP_SKB_CB(skb)->tcp_gso_size = mss_now; 10751da177e4SLinus Torvalds } 10761da177e4SLinus Torvalds } 10771da177e4SLinus Torvalds 107891fed7a1SIlpo Järvinen /* When a modification to fackets out becomes necessary, we need to check 107968f8353bSIlpo Järvinen * skb is counted to fackets_out or not. 108091fed7a1SIlpo Järvinen */ 1081cf533ea5SEric Dumazet static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb, 108291fed7a1SIlpo Järvinen int decr) 108391fed7a1SIlpo Järvinen { 1084a47e5a98SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1085a47e5a98SIlpo Järvinen 1086dc86967bSIlpo Järvinen if (!tp->sacked_out || tcp_is_reno(tp)) 108791fed7a1SIlpo Järvinen return; 108891fed7a1SIlpo Järvinen 10896859d494SIlpo Järvinen if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq)) 109091fed7a1SIlpo Järvinen tp->fackets_out -= decr; 109191fed7a1SIlpo Järvinen } 109291fed7a1SIlpo Järvinen 1093797108d1SIlpo Järvinen /* Pcount in the middle of the write queue got changed, we need to do various 1094797108d1SIlpo Järvinen * tweaks to fix counters 1095797108d1SIlpo Järvinen */ 1096cf533ea5SEric Dumazet static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) 1097797108d1SIlpo Järvinen { 1098797108d1SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1099797108d1SIlpo Järvinen 1100797108d1SIlpo Järvinen tp->packets_out -= decr; 1101797108d1SIlpo Järvinen 1102797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 1103797108d1SIlpo Järvinen tp->sacked_out -= decr; 1104797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 1105797108d1SIlpo Järvinen tp->retrans_out -= decr; 1106797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 1107797108d1SIlpo Järvinen tp->lost_out -= decr; 1108797108d1SIlpo Järvinen 1109797108d1SIlpo Järvinen /* Reno case is special. Sigh... */ 1110797108d1SIlpo Järvinen if (tcp_is_reno(tp) && decr > 0) 1111797108d1SIlpo Järvinen tp->sacked_out -= min_t(u32, tp->sacked_out, decr); 1112797108d1SIlpo Järvinen 1113797108d1SIlpo Järvinen tcp_adjust_fackets_out(sk, skb, decr); 1114797108d1SIlpo Järvinen 1115797108d1SIlpo Järvinen if (tp->lost_skb_hint && 1116797108d1SIlpo Järvinen before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && 111752cf3cc8SIlpo Järvinen (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))) 1118797108d1SIlpo Järvinen tp->lost_cnt_hint -= decr; 1119797108d1SIlpo Järvinen 1120797108d1SIlpo Järvinen tcp_verify_left_out(tp); 1121797108d1SIlpo Järvinen } 1122797108d1SIlpo Järvinen 11230a2cf20cSSoheil Hassas Yeganeh static bool tcp_has_tx_tstamp(const struct sk_buff *skb) 11240a2cf20cSSoheil Hassas Yeganeh { 11250a2cf20cSSoheil Hassas Yeganeh return TCP_SKB_CB(skb)->txstamp_ack || 11260a2cf20cSSoheil Hassas Yeganeh (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP); 11270a2cf20cSSoheil Hassas Yeganeh } 11280a2cf20cSSoheil Hassas Yeganeh 1129490cc7d0SWillem de Bruijn static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2) 1130490cc7d0SWillem de Bruijn { 1131490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo = skb_shinfo(skb); 1132490cc7d0SWillem de Bruijn 11330a2cf20cSSoheil Hassas Yeganeh if (unlikely(tcp_has_tx_tstamp(skb)) && 1134490cc7d0SWillem de Bruijn !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) { 1135490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo2 = skb_shinfo(skb2); 1136490cc7d0SWillem de Bruijn u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP; 1137490cc7d0SWillem de Bruijn 1138490cc7d0SWillem de Bruijn shinfo->tx_flags &= ~tsflags; 1139490cc7d0SWillem de Bruijn shinfo2->tx_flags |= tsflags; 1140490cc7d0SWillem de Bruijn swap(shinfo->tskey, shinfo2->tskey); 1141b51e13faSMartin KaFai Lau TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack; 1142b51e13faSMartin KaFai Lau TCP_SKB_CB(skb)->txstamp_ack = 0; 1143490cc7d0SWillem de Bruijn } 1144490cc7d0SWillem de Bruijn } 1145490cc7d0SWillem de Bruijn 1146a166140eSMartin KaFai Lau static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2) 1147a166140eSMartin KaFai Lau { 1148a166140eSMartin KaFai Lau TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor; 1149a166140eSMartin KaFai Lau TCP_SKB_CB(skb)->eor = 0; 1150a166140eSMartin KaFai Lau } 1151a166140eSMartin KaFai Lau 11521da177e4SLinus Torvalds /* Function to create two new TCP segments. Shrinks the given segment 11531da177e4SLinus Torvalds * to the specified size and appends a new segment with the rest of the 11541da177e4SLinus Torvalds * packet to the list. This won't be called frequently, I hope. 11551da177e4SLinus Torvalds * Remember, these are still headerless SKBs at this point. 11561da177e4SLinus Torvalds */ 1157056834d9SIlpo Järvinen int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, 11586cc55e09SOctavian Purdila unsigned int mss_now, gfp_t gfp) 11591da177e4SLinus Torvalds { 11601da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 11611da177e4SLinus Torvalds struct sk_buff *buff; 11626475be16SDavid S. Miller int nsize, old_factor; 1163b60b49eaSHerbert Xu int nlen; 11649ce01461SIlpo Järvinen u8 flags; 11651da177e4SLinus Torvalds 11662fceec13SIlpo Järvinen if (WARN_ON(len > skb->len)) 11672fceec13SIlpo Järvinen return -EINVAL; 11686a438bbeSStephen Hemminger 11691da177e4SLinus Torvalds nsize = skb_headlen(skb) - len; 11701da177e4SLinus Torvalds if (nsize < 0) 11711da177e4SLinus Torvalds nsize = 0; 11721da177e4SLinus Torvalds 11736cc55e09SOctavian Purdila if (skb_unclone(skb, gfp)) 11741da177e4SLinus Torvalds return -ENOMEM; 11751da177e4SLinus Torvalds 11761da177e4SLinus Torvalds /* Get a new skb... force flag on. */ 1177eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, nsize, gfp, true); 117851456b29SIan Morris if (!buff) 11791da177e4SLinus Torvalds return -ENOMEM; /* We'll just try again later. */ 1180ef5cb973SHerbert Xu 11813ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 11823ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1183b60b49eaSHerbert Xu nlen = skb->len - len - nsize; 1184b60b49eaSHerbert Xu buff->truesize += nlen; 1185b60b49eaSHerbert Xu skb->truesize -= nlen; 11861da177e4SLinus Torvalds 11871da177e4SLinus Torvalds /* Correct the sequence numbers. */ 11881da177e4SLinus Torvalds TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 11891da177e4SLinus Torvalds TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 11901da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 11911da177e4SLinus Torvalds 11921da177e4SLinus Torvalds /* PSH and FIN should only be set in the second packet. */ 11934de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 11944de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 11954de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1196e14c3cafSHerbert Xu TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; 1197a166140eSMartin KaFai Lau tcp_skb_fragment_eor(skb, buff); 11981da177e4SLinus Torvalds 119984fa7933SPatrick McHardy if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { 12001da177e4SLinus Torvalds /* Copy and checksum data tail into the new buffer. */ 1201056834d9SIlpo Järvinen buff->csum = csum_partial_copy_nocheck(skb->data + len, 1202056834d9SIlpo Järvinen skb_put(buff, nsize), 12031da177e4SLinus Torvalds nsize, 0); 12041da177e4SLinus Torvalds 12051da177e4SLinus Torvalds skb_trim(skb, len); 12061da177e4SLinus Torvalds 12071da177e4SLinus Torvalds skb->csum = csum_block_sub(skb->csum, buff->csum, len); 12081da177e4SLinus Torvalds } else { 120984fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 12101da177e4SLinus Torvalds skb_split(skb, buff, len); 12111da177e4SLinus Torvalds } 12121da177e4SLinus Torvalds 12131da177e4SLinus Torvalds buff->ip_summed = skb->ip_summed; 12141da177e4SLinus Torvalds 1215a61bbcf2SPatrick McHardy buff->tstamp = skb->tstamp; 1216490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 12171da177e4SLinus Torvalds 12186475be16SDavid S. Miller old_factor = tcp_skb_pcount(skb); 12196475be16SDavid S. Miller 12201da177e4SLinus Torvalds /* Fix up tso_factor for both original and new SKB. */ 12215bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 12225bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 12231da177e4SLinus Torvalds 1224b9f64820SYuchung Cheng /* Update delivered info for the new segment */ 1225b9f64820SYuchung Cheng TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx; 1226b9f64820SYuchung Cheng 12276475be16SDavid S. Miller /* If this packet has been sent out already, we must 12286475be16SDavid S. Miller * adjust the various packet counters. 12296475be16SDavid S. Miller */ 1230cf0b450cSHerbert Xu if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 12316475be16SDavid S. Miller int diff = old_factor - tcp_skb_pcount(skb) - 12326475be16SDavid S. Miller tcp_skb_pcount(buff); 12331da177e4SLinus Torvalds 1234797108d1SIlpo Järvinen if (diff) 1235797108d1SIlpo Järvinen tcp_adjust_pcount(sk, skb, diff); 12361da177e4SLinus Torvalds } 12371da177e4SLinus Torvalds 12381da177e4SLinus Torvalds /* Link BUFF into the send queue. */ 1239f4a775d1SEric Dumazet __skb_header_release(buff); 1240fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 12411da177e4SLinus Torvalds 12421da177e4SLinus Torvalds return 0; 12431da177e4SLinus Torvalds } 12441da177e4SLinus Torvalds 12451da177e4SLinus Torvalds /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c 12461da177e4SLinus Torvalds * eventually). The difference is that pulled data not copied, but 12471da177e4SLinus Torvalds * immediately discarded. 12481da177e4SLinus Torvalds */ 1249f2911969SHerbert Xu ~{PmVHI~} static void __pskb_trim_head(struct sk_buff *skb, int len) 12501da177e4SLinus Torvalds { 12517b7fc97aSEric Dumazet struct skb_shared_info *shinfo; 12521da177e4SLinus Torvalds int i, k, eat; 12531da177e4SLinus Torvalds 12544fa48bf3SEric Dumazet eat = min_t(int, len, skb_headlen(skb)); 12554fa48bf3SEric Dumazet if (eat) { 12564fa48bf3SEric Dumazet __skb_pull(skb, eat); 12574fa48bf3SEric Dumazet len -= eat; 12584fa48bf3SEric Dumazet if (!len) 12594fa48bf3SEric Dumazet return; 12604fa48bf3SEric Dumazet } 12611da177e4SLinus Torvalds eat = len; 12621da177e4SLinus Torvalds k = 0; 12637b7fc97aSEric Dumazet shinfo = skb_shinfo(skb); 12647b7fc97aSEric Dumazet for (i = 0; i < shinfo->nr_frags; i++) { 12657b7fc97aSEric Dumazet int size = skb_frag_size(&shinfo->frags[i]); 12669e903e08SEric Dumazet 12679e903e08SEric Dumazet if (size <= eat) { 1268aff65da0SIan Campbell skb_frag_unref(skb, i); 12699e903e08SEric Dumazet eat -= size; 12701da177e4SLinus Torvalds } else { 12717b7fc97aSEric Dumazet shinfo->frags[k] = shinfo->frags[i]; 12721da177e4SLinus Torvalds if (eat) { 12737b7fc97aSEric Dumazet shinfo->frags[k].page_offset += eat; 12747b7fc97aSEric Dumazet skb_frag_size_sub(&shinfo->frags[k], eat); 12751da177e4SLinus Torvalds eat = 0; 12761da177e4SLinus Torvalds } 12771da177e4SLinus Torvalds k++; 12781da177e4SLinus Torvalds } 12791da177e4SLinus Torvalds } 12807b7fc97aSEric Dumazet shinfo->nr_frags = k; 12811da177e4SLinus Torvalds 128227a884dcSArnaldo Carvalho de Melo skb_reset_tail_pointer(skb); 12831da177e4SLinus Torvalds skb->data_len -= len; 12841da177e4SLinus Torvalds skb->len = skb->data_len; 12851da177e4SLinus Torvalds } 12861da177e4SLinus Torvalds 128767edfef7SAndi Kleen /* Remove acked data from a packet in the transmit queue. */ 12881da177e4SLinus Torvalds int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 12891da177e4SLinus Torvalds { 129014bbd6a5SPravin B Shelar if (skb_unclone(skb, GFP_ATOMIC)) 12911da177e4SLinus Torvalds return -ENOMEM; 12921da177e4SLinus Torvalds 12934fa48bf3SEric Dumazet __pskb_trim_head(skb, len); 12941da177e4SLinus Torvalds 12951da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq += len; 129684fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 12971da177e4SLinus Torvalds 12981da177e4SLinus Torvalds skb->truesize -= len; 12991da177e4SLinus Torvalds sk->sk_wmem_queued -= len; 13003ab224beSHideo Aoki sk_mem_uncharge(sk, len); 13011da177e4SLinus Torvalds sock_set_flag(sk, SOCK_QUEUE_SHRUNK); 13021da177e4SLinus Torvalds 13035b35e1e6SNeal Cardwell /* Any change of skb->len requires recalculation of tso factor. */ 13041da177e4SLinus Torvalds if (tcp_skb_pcount(skb) > 1) 13055bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb)); 13061da177e4SLinus Torvalds 13071da177e4SLinus Torvalds return 0; 13081da177e4SLinus Torvalds } 13091da177e4SLinus Torvalds 13101b63edd6SYuchung Cheng /* Calculate MSS not accounting any TCP options. */ 13111b63edd6SYuchung Cheng static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu) 13125d424d5aSJohn Heffner { 1313cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1314cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 13155d424d5aSJohn Heffner int mss_now; 13165d424d5aSJohn Heffner 13175d424d5aSJohn Heffner /* Calculate base mss without TCP options: 13185d424d5aSJohn Heffner It is MMS_S - sizeof(tcphdr) of rfc1122 13195d424d5aSJohn Heffner */ 13205d424d5aSJohn Heffner mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 13215d424d5aSJohn Heffner 132267469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 132367469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 132467469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 132567469601SEric Dumazet 132667469601SEric Dumazet if (dst && dst_allfrag(dst)) 132767469601SEric Dumazet mss_now -= icsk->icsk_af_ops->net_frag_header_len; 132867469601SEric Dumazet } 132967469601SEric Dumazet 13305d424d5aSJohn Heffner /* Clamp it (mss_clamp does not include tcp options) */ 13315d424d5aSJohn Heffner if (mss_now > tp->rx_opt.mss_clamp) 13325d424d5aSJohn Heffner mss_now = tp->rx_opt.mss_clamp; 13335d424d5aSJohn Heffner 13345d424d5aSJohn Heffner /* Now subtract optional transport overhead */ 13355d424d5aSJohn Heffner mss_now -= icsk->icsk_ext_hdr_len; 13365d424d5aSJohn Heffner 13375d424d5aSJohn Heffner /* Then reserve room for full set of TCP options and 8 bytes of data */ 13385d424d5aSJohn Heffner if (mss_now < 48) 13395d424d5aSJohn Heffner mss_now = 48; 13405d424d5aSJohn Heffner return mss_now; 13415d424d5aSJohn Heffner } 13425d424d5aSJohn Heffner 13431b63edd6SYuchung Cheng /* Calculate MSS. Not accounting for SACKs here. */ 13441b63edd6SYuchung Cheng int tcp_mtu_to_mss(struct sock *sk, int pmtu) 13451b63edd6SYuchung Cheng { 13461b63edd6SYuchung Cheng /* Subtract TCP options size, not including SACKs */ 13471b63edd6SYuchung Cheng return __tcp_mtu_to_mss(sk, pmtu) - 13481b63edd6SYuchung Cheng (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr)); 13491b63edd6SYuchung Cheng } 13501b63edd6SYuchung Cheng 13515d424d5aSJohn Heffner /* Inverse of above */ 135267469601SEric Dumazet int tcp_mss_to_mtu(struct sock *sk, int mss) 13535d424d5aSJohn Heffner { 1354cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1355cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 13565d424d5aSJohn Heffner int mtu; 13575d424d5aSJohn Heffner 13585d424d5aSJohn Heffner mtu = mss + 13595d424d5aSJohn Heffner tp->tcp_header_len + 13605d424d5aSJohn Heffner icsk->icsk_ext_hdr_len + 13615d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 13625d424d5aSJohn Heffner 136367469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 136467469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 136567469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 136667469601SEric Dumazet 136767469601SEric Dumazet if (dst && dst_allfrag(dst)) 136867469601SEric Dumazet mtu += icsk->icsk_af_ops->net_frag_header_len; 136967469601SEric Dumazet } 13705d424d5aSJohn Heffner return mtu; 13715d424d5aSJohn Heffner } 1372556c6b46SNeal Cardwell EXPORT_SYMBOL(tcp_mss_to_mtu); 13735d424d5aSJohn Heffner 137467edfef7SAndi Kleen /* MTU probing init per socket */ 13755d424d5aSJohn Heffner void tcp_mtup_init(struct sock *sk) 13765d424d5aSJohn Heffner { 13775d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 13785d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 1379b0f9ca53SFan Du struct net *net = sock_net(sk); 13805d424d5aSJohn Heffner 1381b0f9ca53SFan Du icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1; 13825d424d5aSJohn Heffner icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 13835d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 1384b0f9ca53SFan Du icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss); 13855d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 138605cbc0dbSFan Du if (icsk->icsk_mtup.enabled) 138705cbc0dbSFan Du icsk->icsk_mtup.probe_timestamp = tcp_time_stamp; 13885d424d5aSJohn Heffner } 13894bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_mtup_init); 13905d424d5aSJohn Heffner 13911da177e4SLinus Torvalds /* This function synchronize snd mss to current pmtu/exthdr set. 13921da177e4SLinus Torvalds 13931da177e4SLinus Torvalds tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 13941da177e4SLinus Torvalds for TCP options, but includes only bare TCP header. 13951da177e4SLinus Torvalds 13961da177e4SLinus Torvalds tp->rx_opt.mss_clamp is mss negotiated at connection setup. 1397caa20d9aSStephen Hemminger It is minimum of user_mss and mss received with SYN. 13981da177e4SLinus Torvalds It also does not include TCP options. 13991da177e4SLinus Torvalds 1400d83d8461SArnaldo Carvalho de Melo inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. 14011da177e4SLinus Torvalds 14021da177e4SLinus Torvalds tp->mss_cache is current effective sending mss, including 14031da177e4SLinus Torvalds all tcp options except for SACKs. It is evaluated, 14041da177e4SLinus Torvalds taking into account current pmtu, but never exceeds 14051da177e4SLinus Torvalds tp->rx_opt.mss_clamp. 14061da177e4SLinus Torvalds 14071da177e4SLinus Torvalds NOTE1. rfc1122 clearly states that advertised MSS 14081da177e4SLinus Torvalds DOES NOT include either tcp or ip options. 14091da177e4SLinus Torvalds 1410d83d8461SArnaldo Carvalho de Melo NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache 1411d83d8461SArnaldo Carvalho de Melo are READ ONLY outside this function. --ANK (980731) 14121da177e4SLinus Torvalds */ 14131da177e4SLinus Torvalds unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 14141da177e4SLinus Torvalds { 14151da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1416d83d8461SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 14175d424d5aSJohn Heffner int mss_now; 14181da177e4SLinus Torvalds 14195d424d5aSJohn Heffner if (icsk->icsk_mtup.search_high > pmtu) 14205d424d5aSJohn Heffner icsk->icsk_mtup.search_high = pmtu; 14211da177e4SLinus Torvalds 14225d424d5aSJohn Heffner mss_now = tcp_mtu_to_mss(sk, pmtu); 1423409d22b4SIlpo Järvinen mss_now = tcp_bound_to_half_wnd(tp, mss_now); 14241da177e4SLinus Torvalds 14251da177e4SLinus Torvalds /* And store cached results */ 1426d83d8461SArnaldo Carvalho de Melo icsk->icsk_pmtu_cookie = pmtu; 14275d424d5aSJohn Heffner if (icsk->icsk_mtup.enabled) 14285d424d5aSJohn Heffner mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); 1429c1b4a7e6SDavid S. Miller tp->mss_cache = mss_now; 14301da177e4SLinus Torvalds 14311da177e4SLinus Torvalds return mss_now; 14321da177e4SLinus Torvalds } 14334bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_sync_mss); 14341da177e4SLinus Torvalds 14351da177e4SLinus Torvalds /* Compute the current effective MSS, taking SACKs and IP options, 14361da177e4SLinus Torvalds * and even PMTU discovery events into account. 14371da177e4SLinus Torvalds */ 14380c54b85fSIlpo Järvinen unsigned int tcp_current_mss(struct sock *sk) 14391da177e4SLinus Torvalds { 1440cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1441cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1442c1b4a7e6SDavid S. Miller u32 mss_now; 144395c96174SEric Dumazet unsigned int header_len; 144433ad798cSAdam Langley struct tcp_out_options opts; 144533ad798cSAdam Langley struct tcp_md5sig_key *md5; 14461da177e4SLinus Torvalds 1447c1b4a7e6SDavid S. Miller mss_now = tp->mss_cache; 1448c1b4a7e6SDavid S. Miller 14491da177e4SLinus Torvalds if (dst) { 14501da177e4SLinus Torvalds u32 mtu = dst_mtu(dst); 1451d83d8461SArnaldo Carvalho de Melo if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 14521da177e4SLinus Torvalds mss_now = tcp_sync_mss(sk, mtu); 14531da177e4SLinus Torvalds } 14541da177e4SLinus Torvalds 145533ad798cSAdam Langley header_len = tcp_established_options(sk, NULL, &opts, &md5) + 145633ad798cSAdam Langley sizeof(struct tcphdr); 145733ad798cSAdam Langley /* The mss_cache is sized based on tp->tcp_header_len, which assumes 145833ad798cSAdam Langley * some common options. If this is an odd packet (because we have SACK 145933ad798cSAdam Langley * blocks etc) then our calculated header_len will be different, and 146033ad798cSAdam Langley * we have to adjust mss_now correspondingly */ 146133ad798cSAdam Langley if (header_len != tp->tcp_header_len) { 146233ad798cSAdam Langley int delta = (int) header_len - tp->tcp_header_len; 146333ad798cSAdam Langley mss_now -= delta; 146433ad798cSAdam Langley } 1465cfb6eeb4SYOSHIFUJI Hideaki 14661da177e4SLinus Torvalds return mss_now; 14671da177e4SLinus Torvalds } 14681da177e4SLinus Torvalds 146986fd14adSWeiping Pan /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. 147086fd14adSWeiping Pan * As additional protections, we do not touch cwnd in retransmission phases, 147186fd14adSWeiping Pan * and if application hit its sndbuf limit recently. 147286fd14adSWeiping Pan */ 147386fd14adSWeiping Pan static void tcp_cwnd_application_limited(struct sock *sk) 1474a762a980SDavid S. Miller { 14759e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1476a762a980SDavid S. Miller 147786fd14adSWeiping Pan if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && 147886fd14adSWeiping Pan sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 147986fd14adSWeiping Pan /* Limited by application or receiver window. */ 148086fd14adSWeiping Pan u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); 148186fd14adSWeiping Pan u32 win_used = max(tp->snd_cwnd_used, init_win); 148286fd14adSWeiping Pan if (win_used < tp->snd_cwnd) { 148386fd14adSWeiping Pan tp->snd_ssthresh = tcp_current_ssthresh(sk); 148486fd14adSWeiping Pan tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1; 148586fd14adSWeiping Pan } 148686fd14adSWeiping Pan tp->snd_cwnd_used = 0; 148786fd14adSWeiping Pan } 148886fd14adSWeiping Pan tp->snd_cwnd_stamp = tcp_time_stamp; 148986fd14adSWeiping Pan } 149086fd14adSWeiping Pan 1491ca8a2263SNeal Cardwell static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited) 1492a762a980SDavid S. Miller { 1493a762a980SDavid S. Miller struct tcp_sock *tp = tcp_sk(sk); 1494a762a980SDavid S. Miller 1495ca8a2263SNeal Cardwell /* Track the maximum number of outstanding packets in each 1496ca8a2263SNeal Cardwell * window, and remember whether we were cwnd-limited then. 1497ca8a2263SNeal Cardwell */ 1498ca8a2263SNeal Cardwell if (!before(tp->snd_una, tp->max_packets_seq) || 1499ca8a2263SNeal Cardwell tp->packets_out > tp->max_packets_out) { 1500ca8a2263SNeal Cardwell tp->max_packets_out = tp->packets_out; 1501ca8a2263SNeal Cardwell tp->max_packets_seq = tp->snd_nxt; 1502ca8a2263SNeal Cardwell tp->is_cwnd_limited = is_cwnd_limited; 1503ca8a2263SNeal Cardwell } 1504e114a710SEric Dumazet 150524901551SEric Dumazet if (tcp_is_cwnd_limited(sk)) { 1506a762a980SDavid S. Miller /* Network is feed fully. */ 1507a762a980SDavid S. Miller tp->snd_cwnd_used = 0; 1508a762a980SDavid S. Miller tp->snd_cwnd_stamp = tcp_time_stamp; 1509a762a980SDavid S. Miller } else { 1510a762a980SDavid S. Miller /* Network starves. */ 1511a762a980SDavid S. Miller if (tp->packets_out > tp->snd_cwnd_used) 1512a762a980SDavid S. Miller tp->snd_cwnd_used = tp->packets_out; 1513a762a980SDavid S. Miller 151415d33c07SDavid S. Miller if (sysctl_tcp_slow_start_after_idle && 151515d33c07SDavid S. Miller (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto) 1516a762a980SDavid S. Miller tcp_cwnd_application_limited(sk); 1517a762a980SDavid S. Miller } 1518a762a980SDavid S. Miller } 1519a762a980SDavid S. Miller 1520d4589926SEric Dumazet /* Minshall's variant of the Nagle send check. */ 1521d4589926SEric Dumazet static bool tcp_minshall_check(const struct tcp_sock *tp) 1522d4589926SEric Dumazet { 1523d4589926SEric Dumazet return after(tp->snd_sml, tp->snd_una) && 1524d4589926SEric Dumazet !after(tp->snd_sml, tp->snd_nxt); 1525d4589926SEric Dumazet } 1526d4589926SEric Dumazet 1527d4589926SEric Dumazet /* Update snd_sml if this skb is under mss 1528d4589926SEric Dumazet * Note that a TSO packet might end with a sub-mss segment 1529d4589926SEric Dumazet * The test is really : 1530d4589926SEric Dumazet * if ((skb->len % mss) != 0) 1531d4589926SEric Dumazet * tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1532d4589926SEric Dumazet * But we can avoid doing the divide again given we already have 1533d4589926SEric Dumazet * skb_pcount = skb->len / mss_now 15340e3a4803SIlpo Järvinen */ 1535d4589926SEric Dumazet static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now, 1536d4589926SEric Dumazet const struct sk_buff *skb) 1537d4589926SEric Dumazet { 1538d4589926SEric Dumazet if (skb->len < tcp_skb_pcount(skb) * mss_now) 1539d4589926SEric Dumazet tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1540d4589926SEric Dumazet } 1541d4589926SEric Dumazet 1542d4589926SEric Dumazet /* Return false, if packet can be sent now without violation Nagle's rules: 1543d4589926SEric Dumazet * 1. It is full sized. (provided by caller in %partial bool) 1544d4589926SEric Dumazet * 2. Or it contains FIN. (already checked by caller) 1545d4589926SEric Dumazet * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. 1546d4589926SEric Dumazet * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 1547d4589926SEric Dumazet * With Minshall's modification: all sent small packets are ACKed. 1548d4589926SEric Dumazet */ 1549d4589926SEric Dumazet static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp, 1550cc93fc51SPeter Pan(潘卫平) int nonagle) 1551d4589926SEric Dumazet { 1552d4589926SEric Dumazet return partial && 1553d4589926SEric Dumazet ((nonagle & TCP_NAGLE_CORK) || 1554d4589926SEric Dumazet (!nonagle && tp->packets_out && tcp_minshall_check(tp))); 1555d4589926SEric Dumazet } 1556605ad7f1SEric Dumazet 1557605ad7f1SEric Dumazet /* Return how many segs we'd like on a TSO packet, 1558605ad7f1SEric Dumazet * to send one TSO packet per ms 1559605ad7f1SEric Dumazet */ 15601b3878caSNeal Cardwell u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now, 15611b3878caSNeal Cardwell int min_tso_segs) 1562605ad7f1SEric Dumazet { 1563605ad7f1SEric Dumazet u32 bytes, segs; 1564605ad7f1SEric Dumazet 1565605ad7f1SEric Dumazet bytes = min(sk->sk_pacing_rate >> 10, 1566605ad7f1SEric Dumazet sk->sk_gso_max_size - 1 - MAX_TCP_HEADER); 1567605ad7f1SEric Dumazet 1568605ad7f1SEric Dumazet /* Goal is to send at least one packet per ms, 1569605ad7f1SEric Dumazet * not one big TSO packet every 100 ms. 1570605ad7f1SEric Dumazet * This preserves ACK clocking and is consistent 1571605ad7f1SEric Dumazet * with tcp_tso_should_defer() heuristic. 1572605ad7f1SEric Dumazet */ 15731b3878caSNeal Cardwell segs = max_t(u32, bytes / mss_now, min_tso_segs); 1574605ad7f1SEric Dumazet 1575605ad7f1SEric Dumazet return min_t(u32, segs, sk->sk_gso_max_segs); 1576605ad7f1SEric Dumazet } 15771b3878caSNeal Cardwell EXPORT_SYMBOL(tcp_tso_autosize); 1578605ad7f1SEric Dumazet 1579ed6e7268SNeal Cardwell /* Return the number of segments we want in the skb we are transmitting. 1580ed6e7268SNeal Cardwell * See if congestion control module wants to decide; otherwise, autosize. 1581ed6e7268SNeal Cardwell */ 1582ed6e7268SNeal Cardwell static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now) 1583ed6e7268SNeal Cardwell { 1584ed6e7268SNeal Cardwell const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 1585ed6e7268SNeal Cardwell u32 tso_segs = ca_ops->tso_segs_goal ? ca_ops->tso_segs_goal(sk) : 0; 1586ed6e7268SNeal Cardwell 15871b3878caSNeal Cardwell return tso_segs ? : 15881b3878caSNeal Cardwell tcp_tso_autosize(sk, mss_now, sysctl_tcp_min_tso_segs); 1589ed6e7268SNeal Cardwell } 1590ed6e7268SNeal Cardwell 1591d4589926SEric Dumazet /* Returns the portion of skb which can be sent right away */ 1592d4589926SEric Dumazet static unsigned int tcp_mss_split_point(const struct sock *sk, 1593d4589926SEric Dumazet const struct sk_buff *skb, 1594d4589926SEric Dumazet unsigned int mss_now, 1595d4589926SEric Dumazet unsigned int max_segs, 1596d4589926SEric Dumazet int nonagle) 1597c1b4a7e6SDavid S. Miller { 1598cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1599d4589926SEric Dumazet u32 partial, needed, window, max_len; 1600c1b4a7e6SDavid S. Miller 160190840defSIlpo Järvinen window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 16021485348dSBen Hutchings max_len = mss_now * max_segs; 16030e3a4803SIlpo Järvinen 16041485348dSBen Hutchings if (likely(max_len <= window && skb != tcp_write_queue_tail(sk))) 16051485348dSBen Hutchings return max_len; 16060e3a4803SIlpo Järvinen 16075ea3a748SIlpo Järvinen needed = min(skb->len, window); 16085ea3a748SIlpo Järvinen 16091485348dSBen Hutchings if (max_len <= needed) 16101485348dSBen Hutchings return max_len; 16110e3a4803SIlpo Järvinen 1612d4589926SEric Dumazet partial = needed % mss_now; 1613d4589926SEric Dumazet /* If last segment is not a full MSS, check if Nagle rules allow us 1614d4589926SEric Dumazet * to include this last segment in this skb. 1615d4589926SEric Dumazet * Otherwise, we'll split the skb at last MSS boundary 1616d4589926SEric Dumazet */ 1617cc93fc51SPeter Pan(潘卫平) if (tcp_nagle_check(partial != 0, tp, nonagle)) 1618d4589926SEric Dumazet return needed - partial; 1619d4589926SEric Dumazet 1620d4589926SEric Dumazet return needed; 1621c1b4a7e6SDavid S. Miller } 1622c1b4a7e6SDavid S. Miller 1623c1b4a7e6SDavid S. Miller /* Can at least one segment of SKB be sent right now, according to the 1624c1b4a7e6SDavid S. Miller * congestion window rules? If so, return how many segments are allowed. 1625c1b4a7e6SDavid S. Miller */ 1626cf533ea5SEric Dumazet static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, 1627cf533ea5SEric Dumazet const struct sk_buff *skb) 1628c1b4a7e6SDavid S. Miller { 1629d649a7a8SEric Dumazet u32 in_flight, cwnd, halfcwnd; 1630c1b4a7e6SDavid S. Miller 1631c1b4a7e6SDavid S. Miller /* Don't be strict about the congestion window for the final FIN. */ 16324de075e0SEric Dumazet if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && 16334de075e0SEric Dumazet tcp_skb_pcount(skb) == 1) 1634c1b4a7e6SDavid S. Miller return 1; 1635c1b4a7e6SDavid S. Miller 1636c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1637c1b4a7e6SDavid S. Miller cwnd = tp->snd_cwnd; 1638d649a7a8SEric Dumazet if (in_flight >= cwnd) 1639c1b4a7e6SDavid S. Miller return 0; 1640d649a7a8SEric Dumazet 1641d649a7a8SEric Dumazet /* For better scheduling, ensure we have at least 1642d649a7a8SEric Dumazet * 2 GSO packets in flight. 1643d649a7a8SEric Dumazet */ 1644d649a7a8SEric Dumazet halfcwnd = max(cwnd >> 1, 1U); 1645d649a7a8SEric Dumazet return min(halfcwnd, cwnd - in_flight); 1646c1b4a7e6SDavid S. Miller } 1647c1b4a7e6SDavid S. Miller 1648b595076aSUwe Kleine-König /* Initialize TSO state of a skb. 164967edfef7SAndi Kleen * This must be invoked the first time we consider transmitting 1650c1b4a7e6SDavid S. Miller * SKB onto the wire. 1651c1b4a7e6SDavid S. Miller */ 16525bbb432cSEric Dumazet static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now) 1653c1b4a7e6SDavid S. Miller { 1654c1b4a7e6SDavid S. Miller int tso_segs = tcp_skb_pcount(skb); 1655c1b4a7e6SDavid S. Miller 1656f8269a49SIlpo Järvinen if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { 16575bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 1658c1b4a7e6SDavid S. Miller tso_segs = tcp_skb_pcount(skb); 1659c1b4a7e6SDavid S. Miller } 1660c1b4a7e6SDavid S. Miller return tso_segs; 1661c1b4a7e6SDavid S. Miller } 1662c1b4a7e6SDavid S. Miller 1663c1b4a7e6SDavid S. Miller 1664a2a385d6SEric Dumazet /* Return true if the Nagle test allows this packet to be 1665c1b4a7e6SDavid S. Miller * sent now. 1666c1b4a7e6SDavid S. Miller */ 1667a2a385d6SEric Dumazet static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, 1668c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1669c1b4a7e6SDavid S. Miller { 1670c1b4a7e6SDavid S. Miller /* Nagle rule does not apply to frames, which sit in the middle of the 1671c1b4a7e6SDavid S. Miller * write_queue (they have no chances to get new data). 1672c1b4a7e6SDavid S. Miller * 1673c1b4a7e6SDavid S. Miller * This is implemented in the callers, where they modify the 'nonagle' 1674c1b4a7e6SDavid S. Miller * argument based upon the location of SKB in the send queue. 1675c1b4a7e6SDavid S. Miller */ 1676c1b4a7e6SDavid S. Miller if (nonagle & TCP_NAGLE_PUSH) 1677a2a385d6SEric Dumazet return true; 1678c1b4a7e6SDavid S. Miller 16799b44190dSYuchung Cheng /* Don't use the nagle rule for urgent data (or for the final FIN). */ 16809b44190dSYuchung Cheng if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) 1681a2a385d6SEric Dumazet return true; 1682c1b4a7e6SDavid S. Miller 1683cc93fc51SPeter Pan(潘卫平) if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle)) 1684a2a385d6SEric Dumazet return true; 1685c1b4a7e6SDavid S. Miller 1686a2a385d6SEric Dumazet return false; 1687c1b4a7e6SDavid S. Miller } 1688c1b4a7e6SDavid S. Miller 1689c1b4a7e6SDavid S. Miller /* Does at least the first segment of SKB fit into the send window? */ 1690a2a385d6SEric Dumazet static bool tcp_snd_wnd_test(const struct tcp_sock *tp, 1691a2a385d6SEric Dumazet const struct sk_buff *skb, 1692056834d9SIlpo Järvinen unsigned int cur_mss) 1693c1b4a7e6SDavid S. Miller { 1694c1b4a7e6SDavid S. Miller u32 end_seq = TCP_SKB_CB(skb)->end_seq; 1695c1b4a7e6SDavid S. Miller 1696c1b4a7e6SDavid S. Miller if (skb->len > cur_mss) 1697c1b4a7e6SDavid S. Miller end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 1698c1b4a7e6SDavid S. Miller 169990840defSIlpo Järvinen return !after(end_seq, tcp_wnd_end(tp)); 1700c1b4a7e6SDavid S. Miller } 1701c1b4a7e6SDavid S. Miller 1702fe067e8aSDavid S. Miller /* This checks if the data bearing packet SKB (usually tcp_send_head(sk)) 1703c1b4a7e6SDavid S. Miller * should be put on the wire right now. If so, it returns the number of 1704c1b4a7e6SDavid S. Miller * packets allowed by the congestion window. 1705c1b4a7e6SDavid S. Miller */ 1706cf533ea5SEric Dumazet static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb, 1707c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1708c1b4a7e6SDavid S. Miller { 1709cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1710c1b4a7e6SDavid S. Miller unsigned int cwnd_quota; 1711c1b4a7e6SDavid S. Miller 17125bbb432cSEric Dumazet tcp_init_tso_segs(skb, cur_mss); 1713c1b4a7e6SDavid S. Miller 1714c1b4a7e6SDavid S. Miller if (!tcp_nagle_test(tp, skb, cur_mss, nonagle)) 1715c1b4a7e6SDavid S. Miller return 0; 1716c1b4a7e6SDavid S. Miller 1717c1b4a7e6SDavid S. Miller cwnd_quota = tcp_cwnd_test(tp, skb); 1718056834d9SIlpo Järvinen if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss)) 1719c1b4a7e6SDavid S. Miller cwnd_quota = 0; 1720c1b4a7e6SDavid S. Miller 1721c1b4a7e6SDavid S. Miller return cwnd_quota; 1722c1b4a7e6SDavid S. Miller } 1723c1b4a7e6SDavid S. Miller 172467edfef7SAndi Kleen /* Test if sending is allowed right now. */ 1725a2a385d6SEric Dumazet bool tcp_may_send_now(struct sock *sk) 1726c1b4a7e6SDavid S. Miller { 1727cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1728fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 1729c1b4a7e6SDavid S. Miller 1730a02cec21SEric Dumazet return skb && 17310c54b85fSIlpo Järvinen tcp_snd_test(sk, skb, tcp_current_mss(sk), 1732c1b4a7e6SDavid S. Miller (tcp_skb_is_last(sk, skb) ? 1733a02cec21SEric Dumazet tp->nonagle : TCP_NAGLE_PUSH)); 1734c1b4a7e6SDavid S. Miller } 1735c1b4a7e6SDavid S. Miller 1736c1b4a7e6SDavid S. Miller /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 1737c1b4a7e6SDavid S. Miller * which is put after SKB on the list. It is very much like 1738c1b4a7e6SDavid S. Miller * tcp_fragment() except that it may make several kinds of assumptions 1739c1b4a7e6SDavid S. Miller * in order to speed up the splitting operation. In particular, we 1740c1b4a7e6SDavid S. Miller * know that all the data is in scatter-gather pages, and that the 1741c1b4a7e6SDavid S. Miller * packet has never been sent out before (and thus is not cloned). 1742c1b4a7e6SDavid S. Miller */ 1743056834d9SIlpo Järvinen static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, 1744c4ead4c5SEric Dumazet unsigned int mss_now, gfp_t gfp) 1745c1b4a7e6SDavid S. Miller { 1746c1b4a7e6SDavid S. Miller struct sk_buff *buff; 1747c1b4a7e6SDavid S. Miller int nlen = skb->len - len; 17489ce01461SIlpo Järvinen u8 flags; 1749c1b4a7e6SDavid S. Miller 1750c1b4a7e6SDavid S. Miller /* All of a TSO frame must be composed of paged data. */ 1751c8ac3774SHerbert Xu if (skb->len != skb->data_len) 17526cc55e09SOctavian Purdila return tcp_fragment(sk, skb, len, mss_now, gfp); 1753c1b4a7e6SDavid S. Miller 1754eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, gfp, true); 175551456b29SIan Morris if (unlikely(!buff)) 1756c1b4a7e6SDavid S. Miller return -ENOMEM; 1757c1b4a7e6SDavid S. Miller 17583ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 17593ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1760b60b49eaSHerbert Xu buff->truesize += nlen; 1761c1b4a7e6SDavid S. Miller skb->truesize -= nlen; 1762c1b4a7e6SDavid S. Miller 1763c1b4a7e6SDavid S. Miller /* Correct the sequence numbers. */ 1764c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 1765c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 1766c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 1767c1b4a7e6SDavid S. Miller 1768c1b4a7e6SDavid S. Miller /* PSH and FIN should only be set in the second packet. */ 17694de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 17704de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 17714de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1772c1b4a7e6SDavid S. Miller 1773c1b4a7e6SDavid S. Miller /* This packet was never sent out yet, so no SACK bits. */ 1774c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->sacked = 0; 1775c1b4a7e6SDavid S. Miller 1776a166140eSMartin KaFai Lau tcp_skb_fragment_eor(skb, buff); 1777a166140eSMartin KaFai Lau 177884fa7933SPatrick McHardy buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; 1779c1b4a7e6SDavid S. Miller skb_split(skb, buff, len); 1780490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 1781c1b4a7e6SDavid S. Miller 1782c1b4a7e6SDavid S. Miller /* Fix up tso_factor for both original and new SKB. */ 17835bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 17845bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 1785c1b4a7e6SDavid S. Miller 1786c1b4a7e6SDavid S. Miller /* Link BUFF into the send queue. */ 1787f4a775d1SEric Dumazet __skb_header_release(buff); 1788fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 1789c1b4a7e6SDavid S. Miller 1790c1b4a7e6SDavid S. Miller return 0; 1791c1b4a7e6SDavid S. Miller } 1792c1b4a7e6SDavid S. Miller 1793c1b4a7e6SDavid S. Miller /* Try to defer sending, if possible, in order to minimize the amount 1794c1b4a7e6SDavid S. Miller * of TSO splitting we do. View it as a kind of TSO Nagle test. 1795c1b4a7e6SDavid S. Miller * 1796c1b4a7e6SDavid S. Miller * This algorithm is from John Heffner. 1797c1b4a7e6SDavid S. Miller */ 1798ca8a2263SNeal Cardwell static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb, 1799605ad7f1SEric Dumazet bool *is_cwnd_limited, u32 max_segs) 1800c1b4a7e6SDavid S. Miller { 18016687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 180250c8339eSEric Dumazet u32 age, send_win, cong_win, limit, in_flight; 180350c8339eSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 180450c8339eSEric Dumazet struct skb_mstamp now; 180550c8339eSEric Dumazet struct sk_buff *head; 1806ad9f4f50SEric Dumazet int win_divisor; 1807c1b4a7e6SDavid S. Miller 18084de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 1809ae8064acSJohn Heffner goto send_now; 1810c1b4a7e6SDavid S. Miller 181199d7662aSEric Dumazet if (icsk->icsk_ca_state >= TCP_CA_Recovery) 1812ae8064acSJohn Heffner goto send_now; 1813ae8064acSJohn Heffner 18145f852eb5SEric Dumazet /* Avoid bursty behavior by allowing defer 18155f852eb5SEric Dumazet * only if the last write was recent. 18165f852eb5SEric Dumazet */ 18175f852eb5SEric Dumazet if ((s32)(tcp_time_stamp - tp->lsndtime) > 0) 1818ae8064acSJohn Heffner goto send_now; 1819908a75c1SDavid S. Miller 1820c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1821c1b4a7e6SDavid S. Miller 1822056834d9SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight)); 1823c1b4a7e6SDavid S. Miller 182490840defSIlpo Järvinen send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1825c1b4a7e6SDavid S. Miller 1826c1b4a7e6SDavid S. Miller /* From in_flight test above, we know that cwnd > in_flight. */ 1827c1b4a7e6SDavid S. Miller cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; 1828c1b4a7e6SDavid S. Miller 1829c1b4a7e6SDavid S. Miller limit = min(send_win, cong_win); 1830c1b4a7e6SDavid S. Miller 1831ba244fe9SDavid S. Miller /* If a full-sized TSO skb can be sent, do it. */ 1832605ad7f1SEric Dumazet if (limit >= max_segs * tp->mss_cache) 1833ae8064acSJohn Heffner goto send_now; 1834ba244fe9SDavid S. Miller 183562ad2761SIlpo Järvinen /* Middle in queue won't get any more data, full sendable already? */ 183662ad2761SIlpo Järvinen if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) 183762ad2761SIlpo Järvinen goto send_now; 183862ad2761SIlpo Järvinen 1839ad9f4f50SEric Dumazet win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor); 1840ad9f4f50SEric Dumazet if (win_divisor) { 1841c1b4a7e6SDavid S. Miller u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); 1842c1b4a7e6SDavid S. Miller 1843c1b4a7e6SDavid S. Miller /* If at least some fraction of a window is available, 1844c1b4a7e6SDavid S. Miller * just use it. 1845c1b4a7e6SDavid S. Miller */ 1846ad9f4f50SEric Dumazet chunk /= win_divisor; 1847c1b4a7e6SDavid S. Miller if (limit >= chunk) 1848ae8064acSJohn Heffner goto send_now; 1849c1b4a7e6SDavid S. Miller } else { 1850c1b4a7e6SDavid S. Miller /* Different approach, try not to defer past a single 1851c1b4a7e6SDavid S. Miller * ACK. Receiver should ACK every other full sized 1852c1b4a7e6SDavid S. Miller * frame, so if we have space for more than 3 frames 1853c1b4a7e6SDavid S. Miller * then send now. 1854c1b4a7e6SDavid S. Miller */ 18556b5a5c0dSNeal Cardwell if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) 1856ae8064acSJohn Heffner goto send_now; 1857c1b4a7e6SDavid S. Miller } 1858c1b4a7e6SDavid S. Miller 185950c8339eSEric Dumazet head = tcp_write_queue_head(sk); 186050c8339eSEric Dumazet skb_mstamp_get(&now); 186150c8339eSEric Dumazet age = skb_mstamp_us_delta(&now, &head->skb_mstamp); 186250c8339eSEric Dumazet /* If next ACK is likely to come too late (half srtt), do not defer */ 186350c8339eSEric Dumazet if (age < (tp->srtt_us >> 4)) 186450c8339eSEric Dumazet goto send_now; 186550c8339eSEric Dumazet 18665f852eb5SEric Dumazet /* Ok, it looks like it is advisable to defer. */ 1867ae8064acSJohn Heffner 1868d2e1339fSBendik Rønning Opstad if (cong_win < send_win && cong_win <= skb->len) 1869ca8a2263SNeal Cardwell *is_cwnd_limited = true; 1870ca8a2263SNeal Cardwell 1871a2a385d6SEric Dumazet return true; 1872ae8064acSJohn Heffner 1873ae8064acSJohn Heffner send_now: 1874a2a385d6SEric Dumazet return false; 1875c1b4a7e6SDavid S. Miller } 1876c1b4a7e6SDavid S. Miller 187705cbc0dbSFan Du static inline void tcp_mtu_check_reprobe(struct sock *sk) 187805cbc0dbSFan Du { 187905cbc0dbSFan Du struct inet_connection_sock *icsk = inet_csk(sk); 188005cbc0dbSFan Du struct tcp_sock *tp = tcp_sk(sk); 188105cbc0dbSFan Du struct net *net = sock_net(sk); 188205cbc0dbSFan Du u32 interval; 188305cbc0dbSFan Du s32 delta; 188405cbc0dbSFan Du 188505cbc0dbSFan Du interval = net->ipv4.sysctl_tcp_probe_interval; 188605cbc0dbSFan Du delta = tcp_time_stamp - icsk->icsk_mtup.probe_timestamp; 188705cbc0dbSFan Du if (unlikely(delta >= interval * HZ)) { 188805cbc0dbSFan Du int mss = tcp_current_mss(sk); 188905cbc0dbSFan Du 189005cbc0dbSFan Du /* Update current search range */ 189105cbc0dbSFan Du icsk->icsk_mtup.probe_size = 0; 189205cbc0dbSFan Du icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + 189305cbc0dbSFan Du sizeof(struct tcphdr) + 189405cbc0dbSFan Du icsk->icsk_af_ops->net_header_len; 189505cbc0dbSFan Du icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss); 189605cbc0dbSFan Du 189705cbc0dbSFan Du /* Update probe time stamp */ 189805cbc0dbSFan Du icsk->icsk_mtup.probe_timestamp = tcp_time_stamp; 189905cbc0dbSFan Du } 190005cbc0dbSFan Du } 190105cbc0dbSFan Du 19025d424d5aSJohn Heffner /* Create a new MTU probe if we are ready. 190367edfef7SAndi Kleen * MTU probe is regularly attempting to increase the path MTU by 190467edfef7SAndi Kleen * deliberately sending larger packets. This discovers routing 190567edfef7SAndi Kleen * changes resulting in larger path MTUs. 190667edfef7SAndi Kleen * 19075d424d5aSJohn Heffner * Returns 0 if we should wait to probe (no cwnd available), 19085d424d5aSJohn Heffner * 1 if a probe was sent, 1909056834d9SIlpo Järvinen * -1 otherwise 1910056834d9SIlpo Järvinen */ 19115d424d5aSJohn Heffner static int tcp_mtu_probe(struct sock *sk) 19125d424d5aSJohn Heffner { 19135d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 19145d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 19155d424d5aSJohn Heffner struct sk_buff *skb, *nskb, *next; 19166b58e0a5SFan Du struct net *net = sock_net(sk); 19175d424d5aSJohn Heffner int len; 19185d424d5aSJohn Heffner int probe_size; 191991cc17c0SIlpo Järvinen int size_needed; 19205d424d5aSJohn Heffner int copy; 19215d424d5aSJohn Heffner int mss_now; 19226b58e0a5SFan Du int interval; 19235d424d5aSJohn Heffner 19245d424d5aSJohn Heffner /* Not currently probing/verifying, 19255d424d5aSJohn Heffner * not in recovery, 19265d424d5aSJohn Heffner * have enough cwnd, and 19275d424d5aSJohn Heffner * not SACKing (the variable headers throw things off) */ 19285d424d5aSJohn Heffner if (!icsk->icsk_mtup.enabled || 19295d424d5aSJohn Heffner icsk->icsk_mtup.probe_size || 19305d424d5aSJohn Heffner inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 19315d424d5aSJohn Heffner tp->snd_cwnd < 11 || 1932cabeccbdSIlpo Järvinen tp->rx_opt.num_sacks || tp->rx_opt.dsack) 19335d424d5aSJohn Heffner return -1; 19345d424d5aSJohn Heffner 19356b58e0a5SFan Du /* Use binary search for probe_size between tcp_mss_base, 19366b58e0a5SFan Du * and current mss_clamp. if (search_high - search_low) 19376b58e0a5SFan Du * smaller than a threshold, backoff from probing. 19386b58e0a5SFan Du */ 19390c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 19406b58e0a5SFan Du probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high + 19416b58e0a5SFan Du icsk->icsk_mtup.search_low) >> 1); 194291cc17c0SIlpo Järvinen size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; 19436b58e0a5SFan Du interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low; 194405cbc0dbSFan Du /* When misfortune happens, we are reprobing actively, 194505cbc0dbSFan Du * and then reprobe timer has expired. We stick with current 194605cbc0dbSFan Du * probing process by not resetting search range to its orignal. 194705cbc0dbSFan Du */ 19486b58e0a5SFan Du if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) || 194905cbc0dbSFan Du interval < net->ipv4.sysctl_tcp_probe_threshold) { 195005cbc0dbSFan Du /* Check whether enough time has elaplased for 195105cbc0dbSFan Du * another round of probing. 195205cbc0dbSFan Du */ 195305cbc0dbSFan Du tcp_mtu_check_reprobe(sk); 19545d424d5aSJohn Heffner return -1; 19555d424d5aSJohn Heffner } 19565d424d5aSJohn Heffner 19575d424d5aSJohn Heffner /* Have enough data in the send queue to probe? */ 19587f9c33e5SIlpo Järvinen if (tp->write_seq - tp->snd_nxt < size_needed) 19595d424d5aSJohn Heffner return -1; 19605d424d5aSJohn Heffner 196191cc17c0SIlpo Järvinen if (tp->snd_wnd < size_needed) 19625d424d5aSJohn Heffner return -1; 196390840defSIlpo Järvinen if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) 19645d424d5aSJohn Heffner return 0; 19655d424d5aSJohn Heffner 1966d67c58e9SIlpo Järvinen /* Do we need to wait to drain cwnd? With none in flight, don't stall */ 1967d67c58e9SIlpo Järvinen if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { 1968d67c58e9SIlpo Järvinen if (!tcp_packets_in_flight(tp)) 19695d424d5aSJohn Heffner return -1; 19705d424d5aSJohn Heffner else 19715d424d5aSJohn Heffner return 0; 19725d424d5aSJohn Heffner } 19735d424d5aSJohn Heffner 19745d424d5aSJohn Heffner /* We're allowed to probe. Build it now. */ 1975eb934478SEric Dumazet nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false); 197651456b29SIan Morris if (!nskb) 19775d424d5aSJohn Heffner return -1; 19783ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 19793ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 19805d424d5aSJohn Heffner 1981fe067e8aSDavid S. Miller skb = tcp_send_head(sk); 19825d424d5aSJohn Heffner 19835d424d5aSJohn Heffner TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 19845d424d5aSJohn Heffner TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 19854de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; 19865d424d5aSJohn Heffner TCP_SKB_CB(nskb)->sacked = 0; 19875d424d5aSJohn Heffner nskb->csum = 0; 198884fa7933SPatrick McHardy nskb->ip_summed = skb->ip_summed; 19895d424d5aSJohn Heffner 199050c4817eSIlpo Järvinen tcp_insert_write_queue_before(nskb, skb, sk); 199150c4817eSIlpo Järvinen 19925d424d5aSJohn Heffner len = 0; 1993234b6860SIlpo Järvinen tcp_for_write_queue_from_safe(skb, next, sk) { 19945d424d5aSJohn Heffner copy = min_t(int, skb->len, probe_size - len); 19952fe664f1SDouglas Caetano dos Santos if (nskb->ip_summed) { 19965d424d5aSJohn Heffner skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); 19972fe664f1SDouglas Caetano dos Santos } else { 19982fe664f1SDouglas Caetano dos Santos __wsum csum = skb_copy_and_csum_bits(skb, 0, 1999056834d9SIlpo Järvinen skb_put(nskb, copy), 20002fe664f1SDouglas Caetano dos Santos copy, 0); 20012fe664f1SDouglas Caetano dos Santos nskb->csum = csum_block_add(nskb->csum, csum, len); 20022fe664f1SDouglas Caetano dos Santos } 20035d424d5aSJohn Heffner 20045d424d5aSJohn Heffner if (skb->len <= copy) { 20055d424d5aSJohn Heffner /* We've eaten all the data from this skb. 20065d424d5aSJohn Heffner * Throw it away. */ 20074de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 2008fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 20093ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 20105d424d5aSJohn Heffner } else { 20114de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & 2012a3433f35SChangli Gao ~(TCPHDR_FIN|TCPHDR_PSH); 20135d424d5aSJohn Heffner if (!skb_shinfo(skb)->nr_frags) { 20145d424d5aSJohn Heffner skb_pull(skb, copy); 201584fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 2016056834d9SIlpo Järvinen skb->csum = csum_partial(skb->data, 2017056834d9SIlpo Järvinen skb->len, 0); 20185d424d5aSJohn Heffner } else { 20195d424d5aSJohn Heffner __pskb_trim_head(skb, copy); 20205bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 20215d424d5aSJohn Heffner } 20225d424d5aSJohn Heffner TCP_SKB_CB(skb)->seq += copy; 20235d424d5aSJohn Heffner } 20245d424d5aSJohn Heffner 20255d424d5aSJohn Heffner len += copy; 2026234b6860SIlpo Järvinen 2027234b6860SIlpo Järvinen if (len >= probe_size) 2028234b6860SIlpo Järvinen break; 20295d424d5aSJohn Heffner } 20305bbb432cSEric Dumazet tcp_init_tso_segs(nskb, nskb->len); 20315d424d5aSJohn Heffner 20325d424d5aSJohn Heffner /* We're ready to send. If this fails, the probe will 20337faee5c0SEric Dumazet * be resegmented into mss-sized pieces by tcp_write_xmit(). 20347faee5c0SEric Dumazet */ 20355d424d5aSJohn Heffner if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 20365d424d5aSJohn Heffner /* Decrement cwnd here because we are sending 20375d424d5aSJohn Heffner * effectively two packets. */ 20385d424d5aSJohn Heffner tp->snd_cwnd--; 203966f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, nskb); 20405d424d5aSJohn Heffner 20415d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 20420e7b1368SJohn Heffner tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 20430e7b1368SJohn Heffner tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 20445d424d5aSJohn Heffner 20455d424d5aSJohn Heffner return 1; 20465d424d5aSJohn Heffner } 20475d424d5aSJohn Heffner 20485d424d5aSJohn Heffner return -1; 20495d424d5aSJohn Heffner } 20505d424d5aSJohn Heffner 2051f9616c35SEric Dumazet /* TCP Small Queues : 2052f9616c35SEric Dumazet * Control number of packets in qdisc/devices to two packets / or ~1 ms. 2053f9616c35SEric Dumazet * (These limits are doubled for retransmits) 2054f9616c35SEric Dumazet * This allows for : 2055f9616c35SEric Dumazet * - better RTT estimation and ACK scheduling 2056f9616c35SEric Dumazet * - faster recovery 2057f9616c35SEric Dumazet * - high rates 2058f9616c35SEric Dumazet * Alas, some drivers / subsystems require a fair amount 2059f9616c35SEric Dumazet * of queued bytes to ensure line rate. 2060f9616c35SEric Dumazet * One example is wifi aggregation (802.11 AMPDU) 2061f9616c35SEric Dumazet */ 2062f9616c35SEric Dumazet static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb, 2063f9616c35SEric Dumazet unsigned int factor) 2064f9616c35SEric Dumazet { 2065f9616c35SEric Dumazet unsigned int limit; 2066f9616c35SEric Dumazet 2067f9616c35SEric Dumazet limit = max(2 * skb->truesize, sk->sk_pacing_rate >> 10); 2068f9616c35SEric Dumazet limit = min_t(u32, limit, sysctl_tcp_limit_output_bytes); 2069f9616c35SEric Dumazet limit <<= factor; 2070f9616c35SEric Dumazet 2071f9616c35SEric Dumazet if (atomic_read(&sk->sk_wmem_alloc) > limit) { 2072f9616c35SEric Dumazet set_bit(TSQ_THROTTLED, &tcp_sk(sk)->tsq_flags); 2073f9616c35SEric Dumazet /* It is possible TX completion already happened 2074f9616c35SEric Dumazet * before we set TSQ_THROTTLED, so we must 2075f9616c35SEric Dumazet * test again the condition. 2076f9616c35SEric Dumazet */ 2077f9616c35SEric Dumazet smp_mb__after_atomic(); 2078f9616c35SEric Dumazet if (atomic_read(&sk->sk_wmem_alloc) > limit) 2079f9616c35SEric Dumazet return true; 2080f9616c35SEric Dumazet } 2081f9616c35SEric Dumazet return false; 2082f9616c35SEric Dumazet } 2083f9616c35SEric Dumazet 20841da177e4SLinus Torvalds /* This routine writes packets to the network. It advances the 20851da177e4SLinus Torvalds * send_head. This happens as incoming acks open up the remote 20861da177e4SLinus Torvalds * window for us. 20871da177e4SLinus Torvalds * 2088f8269a49SIlpo Järvinen * LARGESEND note: !tcp_urg_mode is overkill, only frames between 2089f8269a49SIlpo Järvinen * snd_up-64k-mss .. snd_up cannot be large. However, taking into 2090f8269a49SIlpo Järvinen * account rare use of URG, this is not a big flaw. 2091f8269a49SIlpo Järvinen * 20926ba8a3b1SNandita Dukkipati * Send at most one packet when push_one > 0. Temporarily ignore 20936ba8a3b1SNandita Dukkipati * cwnd limit to force at most one packet out when push_one == 2. 20946ba8a3b1SNandita Dukkipati 2095a2a385d6SEric Dumazet * Returns true, if no segments are in flight and we have queued segments, 2096a2a385d6SEric Dumazet * but cannot send anything now because of SWS or another problem. 20971da177e4SLinus Torvalds */ 2098a2a385d6SEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 2099d5dd9175SIlpo Järvinen int push_one, gfp_t gfp) 21001da177e4SLinus Torvalds { 21011da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 210292df7b51SDavid S. Miller struct sk_buff *skb; 2103c1b4a7e6SDavid S. Miller unsigned int tso_segs, sent_pkts; 2104c1b4a7e6SDavid S. Miller int cwnd_quota; 21055d424d5aSJohn Heffner int result; 2106ca8a2263SNeal Cardwell bool is_cwnd_limited = false; 2107605ad7f1SEric Dumazet u32 max_segs; 21081da177e4SLinus Torvalds 2109c1b4a7e6SDavid S. Miller sent_pkts = 0; 21105d424d5aSJohn Heffner 2111d5dd9175SIlpo Järvinen if (!push_one) { 21125d424d5aSJohn Heffner /* Do MTU probing. */ 2113d5dd9175SIlpo Järvinen result = tcp_mtu_probe(sk); 2114d5dd9175SIlpo Järvinen if (!result) { 2115a2a385d6SEric Dumazet return false; 21165d424d5aSJohn Heffner } else if (result > 0) { 21175d424d5aSJohn Heffner sent_pkts = 1; 21185d424d5aSJohn Heffner } 2119d5dd9175SIlpo Järvinen } 21205d424d5aSJohn Heffner 2121ed6e7268SNeal Cardwell max_segs = tcp_tso_segs(sk, mss_now); 2122fe067e8aSDavid S. Miller while ((skb = tcp_send_head(sk))) { 2123c8ac3774SHerbert Xu unsigned int limit; 2124c8ac3774SHerbert Xu 21255bbb432cSEric Dumazet tso_segs = tcp_init_tso_segs(skb, mss_now); 2126c1b4a7e6SDavid S. Miller BUG_ON(!tso_segs); 2127c1b4a7e6SDavid S. Miller 21289d186cacSAndrey Vagin if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) { 21297faee5c0SEric Dumazet /* "skb_mstamp" is used as a start point for the retransmit timer */ 21307faee5c0SEric Dumazet skb_mstamp_get(&skb->skb_mstamp); 2131ec342325SAndrew Vagin goto repair; /* Skip network transmission */ 21329d186cacSAndrey Vagin } 2133ec342325SAndrew Vagin 2134b68e9f85SHerbert Xu cwnd_quota = tcp_cwnd_test(tp, skb); 21356ba8a3b1SNandita Dukkipati if (!cwnd_quota) { 21366ba8a3b1SNandita Dukkipati if (push_one == 2) 21376ba8a3b1SNandita Dukkipati /* Force out a loss probe pkt. */ 21386ba8a3b1SNandita Dukkipati cwnd_quota = 1; 21396ba8a3b1SNandita Dukkipati else 2140b68e9f85SHerbert Xu break; 21416ba8a3b1SNandita Dukkipati } 2142b68e9f85SHerbert Xu 2143b68e9f85SHerbert Xu if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) 2144b68e9f85SHerbert Xu break; 2145b68e9f85SHerbert Xu 2146d6a4e26aSEric Dumazet if (tso_segs == 1) { 2147aa93466bSDavid S. Miller if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 2148aa93466bSDavid S. Miller (tcp_skb_is_last(sk, skb) ? 2149aa93466bSDavid S. Miller nonagle : TCP_NAGLE_PUSH)))) 2150aa93466bSDavid S. Miller break; 2151c1b4a7e6SDavid S. Miller } else { 2152ca8a2263SNeal Cardwell if (!push_one && 2153605ad7f1SEric Dumazet tcp_tso_should_defer(sk, skb, &is_cwnd_limited, 2154605ad7f1SEric Dumazet max_segs)) 2155aa93466bSDavid S. Miller break; 2156c1b4a7e6SDavid S. Miller } 2157aa93466bSDavid S. Miller 2158605ad7f1SEric Dumazet limit = mss_now; 2159d6a4e26aSEric Dumazet if (tso_segs > 1 && !tcp_urg_mode(tp)) 2160605ad7f1SEric Dumazet limit = tcp_mss_split_point(sk, skb, mss_now, 2161605ad7f1SEric Dumazet min_t(unsigned int, 2162605ad7f1SEric Dumazet cwnd_quota, 2163605ad7f1SEric Dumazet max_segs), 2164605ad7f1SEric Dumazet nonagle); 2165605ad7f1SEric Dumazet 2166605ad7f1SEric Dumazet if (skb->len > limit && 2167605ad7f1SEric Dumazet unlikely(tso_fragment(sk, skb, limit, mss_now, gfp))) 2168605ad7f1SEric Dumazet break; 2169605ad7f1SEric Dumazet 2170f9616c35SEric Dumazet if (tcp_small_queue_check(sk, skb, 0)) 217146d3ceabSEric Dumazet break; 2172c9eeec26SEric Dumazet 2173d5dd9175SIlpo Järvinen if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) 21741da177e4SLinus Torvalds break; 21751da177e4SLinus Torvalds 2176ec342325SAndrew Vagin repair: 21771da177e4SLinus Torvalds /* Advance the send_head. This one is sent out. 21781da177e4SLinus Torvalds * This call will increment packets_out. 21791da177e4SLinus Torvalds */ 218066f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 21811da177e4SLinus Torvalds 21821da177e4SLinus Torvalds tcp_minshall_update(tp, mss_now, skb); 2183a262f0cdSNandita Dukkipati sent_pkts += tcp_skb_pcount(skb); 2184d5dd9175SIlpo Järvinen 2185d5dd9175SIlpo Järvinen if (push_one) 2186d5dd9175SIlpo Järvinen break; 21871da177e4SLinus Torvalds } 21881da177e4SLinus Torvalds 2189aa93466bSDavid S. Miller if (likely(sent_pkts)) { 2190684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 2191684bad11SYuchung Cheng tp->prr_out += sent_pkts; 21926ba8a3b1SNandita Dukkipati 21936ba8a3b1SNandita Dukkipati /* Send one loss probe per tail loss episode. */ 21946ba8a3b1SNandita Dukkipati if (push_one != 2) 21956ba8a3b1SNandita Dukkipati tcp_schedule_loss_probe(sk); 2196d2e1339fSBendik Rønning Opstad is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd); 2197ca8a2263SNeal Cardwell tcp_cwnd_validate(sk, is_cwnd_limited); 2198a2a385d6SEric Dumazet return false; 21991da177e4SLinus Torvalds } 2200b340b264SYuchung Cheng return !tp->packets_out && tcp_send_head(sk); 22016ba8a3b1SNandita Dukkipati } 22026ba8a3b1SNandita Dukkipati 22036ba8a3b1SNandita Dukkipati bool tcp_schedule_loss_probe(struct sock *sk) 22046ba8a3b1SNandita Dukkipati { 22056ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 22066ba8a3b1SNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 22076ba8a3b1SNandita Dukkipati u32 timeout, tlp_time_stamp, rto_time_stamp; 2208740b0f18SEric Dumazet u32 rtt = usecs_to_jiffies(tp->srtt_us >> 3); 22096ba8a3b1SNandita Dukkipati 22106ba8a3b1SNandita Dukkipati if (WARN_ON(icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS)) 22116ba8a3b1SNandita Dukkipati return false; 22126ba8a3b1SNandita Dukkipati /* No consecutive loss probes. */ 22136ba8a3b1SNandita Dukkipati if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) { 22146ba8a3b1SNandita Dukkipati tcp_rearm_rto(sk); 22156ba8a3b1SNandita Dukkipati return false; 22166ba8a3b1SNandita Dukkipati } 22176ba8a3b1SNandita Dukkipati /* Don't do any loss probe on a Fast Open connection before 3WHS 22186ba8a3b1SNandita Dukkipati * finishes. 22196ba8a3b1SNandita Dukkipati */ 2220f9b99582SYuchung Cheng if (tp->fastopen_rsk) 22216ba8a3b1SNandita Dukkipati return false; 22226ba8a3b1SNandita Dukkipati 22236ba8a3b1SNandita Dukkipati /* TLP is only scheduled when next timer event is RTO. */ 22246ba8a3b1SNandita Dukkipati if (icsk->icsk_pending != ICSK_TIME_RETRANS) 22256ba8a3b1SNandita Dukkipati return false; 22266ba8a3b1SNandita Dukkipati 22276ba8a3b1SNandita Dukkipati /* Schedule a loss probe in 2*RTT for SACK capable connections 22286ba8a3b1SNandita Dukkipati * in Open state, that are either limited by cwnd or application. 22296ba8a3b1SNandita Dukkipati */ 2230f9b99582SYuchung Cheng if (sysctl_tcp_early_retrans < 3 || !tp->packets_out || 22316ba8a3b1SNandita Dukkipati !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open) 22326ba8a3b1SNandita Dukkipati return false; 22336ba8a3b1SNandita Dukkipati 22346ba8a3b1SNandita Dukkipati if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) && 22356ba8a3b1SNandita Dukkipati tcp_send_head(sk)) 22366ba8a3b1SNandita Dukkipati return false; 22376ba8a3b1SNandita Dukkipati 22386ba8a3b1SNandita Dukkipati /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account 2239f9b99582SYuchung Cheng * for delayed ack when there's one outstanding packet. If no RTT 2240f9b99582SYuchung Cheng * sample is available then probe after TCP_TIMEOUT_INIT. 22416ba8a3b1SNandita Dukkipati */ 2242f9b99582SYuchung Cheng timeout = rtt << 1 ? : TCP_TIMEOUT_INIT; 22436ba8a3b1SNandita Dukkipati if (tp->packets_out == 1) 22446ba8a3b1SNandita Dukkipati timeout = max_t(u32, timeout, 22456ba8a3b1SNandita Dukkipati (rtt + (rtt >> 1) + TCP_DELACK_MAX)); 22466ba8a3b1SNandita Dukkipati timeout = max_t(u32, timeout, msecs_to_jiffies(10)); 22476ba8a3b1SNandita Dukkipati 22486ba8a3b1SNandita Dukkipati /* If RTO is shorter, just schedule TLP in its place. */ 22496ba8a3b1SNandita Dukkipati tlp_time_stamp = tcp_time_stamp + timeout; 22506ba8a3b1SNandita Dukkipati rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout; 22516ba8a3b1SNandita Dukkipati if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) { 22526ba8a3b1SNandita Dukkipati s32 delta = rto_time_stamp - tcp_time_stamp; 22536ba8a3b1SNandita Dukkipati if (delta > 0) 22546ba8a3b1SNandita Dukkipati timeout = delta; 22556ba8a3b1SNandita Dukkipati } 22566ba8a3b1SNandita Dukkipati 22576ba8a3b1SNandita Dukkipati inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, 22586ba8a3b1SNandita Dukkipati TCP_RTO_MAX); 22596ba8a3b1SNandita Dukkipati return true; 22606ba8a3b1SNandita Dukkipati } 22616ba8a3b1SNandita Dukkipati 22621f3279aeSEric Dumazet /* Thanks to skb fast clones, we can detect if a prior transmit of 22631f3279aeSEric Dumazet * a packet is still in a qdisc or driver queue. 22641f3279aeSEric Dumazet * In this case, there is very little point doing a retransmit ! 22651f3279aeSEric Dumazet */ 22661f3279aeSEric Dumazet static bool skb_still_in_host_queue(const struct sock *sk, 22671f3279aeSEric Dumazet const struct sk_buff *skb) 22681f3279aeSEric Dumazet { 226939bb5e62SEric Dumazet if (unlikely(skb_fclone_busy(sk, skb))) { 2270c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), 22711f3279aeSEric Dumazet LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES); 22721f3279aeSEric Dumazet return true; 22731f3279aeSEric Dumazet } 22741f3279aeSEric Dumazet return false; 22751f3279aeSEric Dumazet } 22761f3279aeSEric Dumazet 2277b340b264SYuchung Cheng /* When probe timeout (PTO) fires, try send a new segment if possible, else 22786ba8a3b1SNandita Dukkipati * retransmit the last segment. 22796ba8a3b1SNandita Dukkipati */ 22806ba8a3b1SNandita Dukkipati void tcp_send_loss_probe(struct sock *sk) 22816ba8a3b1SNandita Dukkipati { 22829b717a8dSNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 22836ba8a3b1SNandita Dukkipati struct sk_buff *skb; 22846ba8a3b1SNandita Dukkipati int pcount; 22856ba8a3b1SNandita Dukkipati int mss = tcp_current_mss(sk); 22866ba8a3b1SNandita Dukkipati 2287b340b264SYuchung Cheng skb = tcp_send_head(sk); 2288b340b264SYuchung Cheng if (skb) { 2289b340b264SYuchung Cheng if (tcp_snd_wnd_test(tp, skb, mss)) { 2290b340b264SYuchung Cheng pcount = tp->packets_out; 2291b340b264SYuchung Cheng tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC); 2292b340b264SYuchung Cheng if (tp->packets_out > pcount) 2293b340b264SYuchung Cheng goto probe_sent; 22946ba8a3b1SNandita Dukkipati goto rearm_timer; 22956ba8a3b1SNandita Dukkipati } 2296b340b264SYuchung Cheng skb = tcp_write_queue_prev(sk, skb); 2297b340b264SYuchung Cheng } else { 2298b340b264SYuchung Cheng skb = tcp_write_queue_tail(sk); 2299b340b264SYuchung Cheng } 23006ba8a3b1SNandita Dukkipati 23019b717a8dSNandita Dukkipati /* At most one outstanding TLP retransmission. */ 23029b717a8dSNandita Dukkipati if (tp->tlp_high_seq) 23039b717a8dSNandita Dukkipati goto rearm_timer; 23049b717a8dSNandita Dukkipati 23056ba8a3b1SNandita Dukkipati /* Retransmit last segment. */ 23066ba8a3b1SNandita Dukkipati if (WARN_ON(!skb)) 23076ba8a3b1SNandita Dukkipati goto rearm_timer; 23086ba8a3b1SNandita Dukkipati 23091f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 23101f3279aeSEric Dumazet goto rearm_timer; 23111f3279aeSEric Dumazet 23126ba8a3b1SNandita Dukkipati pcount = tcp_skb_pcount(skb); 23136ba8a3b1SNandita Dukkipati if (WARN_ON(!pcount)) 23146ba8a3b1SNandita Dukkipati goto rearm_timer; 23156ba8a3b1SNandita Dukkipati 23166ba8a3b1SNandita Dukkipati if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) { 23176cc55e09SOctavian Purdila if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss, 23186cc55e09SOctavian Purdila GFP_ATOMIC))) 23196ba8a3b1SNandita Dukkipati goto rearm_timer; 2320b340b264SYuchung Cheng skb = tcp_write_queue_next(sk, skb); 23216ba8a3b1SNandita Dukkipati } 23226ba8a3b1SNandita Dukkipati 23236ba8a3b1SNandita Dukkipati if (WARN_ON(!skb || !tcp_skb_pcount(skb))) 23246ba8a3b1SNandita Dukkipati goto rearm_timer; 23256ba8a3b1SNandita Dukkipati 232610d3be56SEric Dumazet if (__tcp_retransmit_skb(sk, skb, 1)) 2327b340b264SYuchung Cheng goto rearm_timer; 23286ba8a3b1SNandita Dukkipati 23299b717a8dSNandita Dukkipati /* Record snd_nxt for loss detection. */ 23309b717a8dSNandita Dukkipati tp->tlp_high_seq = tp->snd_nxt; 23319b717a8dSNandita Dukkipati 2332b340b264SYuchung Cheng probe_sent: 2333c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES); 2334fcd16c0aSYuchung Cheng /* Reset s.t. tcp_rearm_rto will restart timer from now */ 2335fcd16c0aSYuchung Cheng inet_csk(sk)->icsk_pending = 0; 2336b340b264SYuchung Cheng rearm_timer: 2337fcd16c0aSYuchung Cheng tcp_rearm_rto(sk); 23381da177e4SLinus Torvalds } 23391da177e4SLinus Torvalds 2340a762a980SDavid S. Miller /* Push out any pending frames which were held back due to 2341a762a980SDavid S. Miller * TCP_CORK or attempt at coalescing tiny packets. 2342a762a980SDavid S. Miller * The socket must be locked by the caller. 2343a762a980SDavid S. Miller */ 23449e412ba7SIlpo Järvinen void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 23459e412ba7SIlpo Järvinen int nonagle) 2346a762a980SDavid S. Miller { 2347726e07a8SIlpo Järvinen /* If we are closed, the bytes will have to remain here. 2348726e07a8SIlpo Järvinen * In time closedown will finish, we empty the write queue and 2349726e07a8SIlpo Järvinen * all will be happy. 2350726e07a8SIlpo Järvinen */ 2351726e07a8SIlpo Järvinen if (unlikely(sk->sk_state == TCP_CLOSE)) 2352726e07a8SIlpo Järvinen return; 2353726e07a8SIlpo Järvinen 235499a1dec7SMel Gorman if (tcp_write_xmit(sk, cur_mss, nonagle, 0, 23557450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC))) 23569e412ba7SIlpo Järvinen tcp_check_probe_timer(sk); 2357a762a980SDavid S. Miller } 2358a762a980SDavid S. Miller 2359c1b4a7e6SDavid S. Miller /* Send _single_ skb sitting at the send head. This function requires 2360c1b4a7e6SDavid S. Miller * true push pending frames to setup probe timer etc. 2361c1b4a7e6SDavid S. Miller */ 2362c1b4a7e6SDavid S. Miller void tcp_push_one(struct sock *sk, unsigned int mss_now) 2363c1b4a7e6SDavid S. Miller { 2364fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 2365c1b4a7e6SDavid S. Miller 2366c1b4a7e6SDavid S. Miller BUG_ON(!skb || skb->len < mss_now); 2367c1b4a7e6SDavid S. Miller 2368d5dd9175SIlpo Järvinen tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); 2369c1b4a7e6SDavid S. Miller } 2370c1b4a7e6SDavid S. Miller 23711da177e4SLinus Torvalds /* This function returns the amount that we can raise the 23721da177e4SLinus Torvalds * usable window based on the following constraints 23731da177e4SLinus Torvalds * 23741da177e4SLinus Torvalds * 1. The window can never be shrunk once it is offered (RFC 793) 23751da177e4SLinus Torvalds * 2. We limit memory per socket 23761da177e4SLinus Torvalds * 23771da177e4SLinus Torvalds * RFC 1122: 23781da177e4SLinus Torvalds * "the suggested [SWS] avoidance algorithm for the receiver is to keep 23791da177e4SLinus Torvalds * RECV.NEXT + RCV.WIN fixed until: 23801da177e4SLinus Torvalds * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 23811da177e4SLinus Torvalds * 23821da177e4SLinus Torvalds * i.e. don't raise the right edge of the window until you can raise 23831da177e4SLinus Torvalds * it at least MSS bytes. 23841da177e4SLinus Torvalds * 23851da177e4SLinus Torvalds * Unfortunately, the recommended algorithm breaks header prediction, 23861da177e4SLinus Torvalds * since header prediction assumes th->window stays fixed. 23871da177e4SLinus Torvalds * 23881da177e4SLinus Torvalds * Strictly speaking, keeping th->window fixed violates the receiver 23891da177e4SLinus Torvalds * side SWS prevention criteria. The problem is that under this rule 23901da177e4SLinus Torvalds * a stream of single byte packets will cause the right side of the 23911da177e4SLinus Torvalds * window to always advance by a single byte. 23921da177e4SLinus Torvalds * 23931da177e4SLinus Torvalds * Of course, if the sender implements sender side SWS prevention 23941da177e4SLinus Torvalds * then this will not be a problem. 23951da177e4SLinus Torvalds * 23961da177e4SLinus Torvalds * BSD seems to make the following compromise: 23971da177e4SLinus Torvalds * 23981da177e4SLinus Torvalds * If the free space is less than the 1/4 of the maximum 23991da177e4SLinus Torvalds * space available and the free space is less than 1/2 mss, 24001da177e4SLinus Torvalds * then set the window to 0. 24011da177e4SLinus Torvalds * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 24021da177e4SLinus Torvalds * Otherwise, just prevent the window from shrinking 24031da177e4SLinus Torvalds * and from being larger than the largest representable value. 24041da177e4SLinus Torvalds * 24051da177e4SLinus Torvalds * This prevents incremental opening of the window in the regime 24061da177e4SLinus Torvalds * where TCP is limited by the speed of the reader side taking 24071da177e4SLinus Torvalds * data out of the TCP receive queue. It does nothing about 24081da177e4SLinus Torvalds * those cases where the window is constrained on the sender side 24091da177e4SLinus Torvalds * because the pipeline is full. 24101da177e4SLinus Torvalds * 24111da177e4SLinus Torvalds * BSD also seems to "accidentally" limit itself to windows that are a 24121da177e4SLinus Torvalds * multiple of MSS, at least until the free space gets quite small. 24131da177e4SLinus Torvalds * This would appear to be a side effect of the mbuf implementation. 24141da177e4SLinus Torvalds * Combining these two algorithms results in the observed behavior 24151da177e4SLinus Torvalds * of having a fixed window size at almost all times. 24161da177e4SLinus Torvalds * 24171da177e4SLinus Torvalds * Below we obtain similar behavior by forcing the offered window to 24181da177e4SLinus Torvalds * a multiple of the mss when it is feasible to do so. 24191da177e4SLinus Torvalds * 24201da177e4SLinus Torvalds * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 24211da177e4SLinus Torvalds * Regular options like TIMESTAMP are taken into account. 24221da177e4SLinus Torvalds */ 24231da177e4SLinus Torvalds u32 __tcp_select_window(struct sock *sk) 24241da177e4SLinus Torvalds { 2425463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 24261da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2427caa20d9aSStephen Hemminger /* MSS for the peer's data. Previous versions used mss_clamp 24281da177e4SLinus Torvalds * here. I don't know if the value based on our guesses 24291da177e4SLinus Torvalds * of peer's MSS is better for the performance. It's more correct 24301da177e4SLinus Torvalds * but may be worse for the performance because of rcv_mss 24311da177e4SLinus Torvalds * fluctuations. --SAW 1998/11/1 24321da177e4SLinus Torvalds */ 2433463c84b9SArnaldo Carvalho de Melo int mss = icsk->icsk_ack.rcv_mss; 24341da177e4SLinus Torvalds int free_space = tcp_space(sk); 243586c1a045SFlorian Westphal int allowed_space = tcp_full_space(sk); 243686c1a045SFlorian Westphal int full_space = min_t(int, tp->window_clamp, allowed_space); 24371da177e4SLinus Torvalds int window; 24381da177e4SLinus Torvalds 24391da177e4SLinus Torvalds if (mss > full_space) 24401da177e4SLinus Torvalds mss = full_space; 24411da177e4SLinus Torvalds 2442b92edbe0SEric Dumazet if (free_space < (full_space >> 1)) { 2443463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.quick = 0; 24441da177e4SLinus Torvalds 2445b8da51ebSEric Dumazet if (tcp_under_memory_pressure(sk)) 2446056834d9SIlpo Järvinen tp->rcv_ssthresh = min(tp->rcv_ssthresh, 2447056834d9SIlpo Järvinen 4U * tp->advmss); 24481da177e4SLinus Torvalds 244986c1a045SFlorian Westphal /* free_space might become our new window, make sure we don't 245086c1a045SFlorian Westphal * increase it due to wscale. 245186c1a045SFlorian Westphal */ 245286c1a045SFlorian Westphal free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale); 245386c1a045SFlorian Westphal 245486c1a045SFlorian Westphal /* if free space is less than mss estimate, or is below 1/16th 245586c1a045SFlorian Westphal * of the maximum allowed, try to move to zero-window, else 245686c1a045SFlorian Westphal * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and 245786c1a045SFlorian Westphal * new incoming data is dropped due to memory limits. 245886c1a045SFlorian Westphal * With large window, mss test triggers way too late in order 245986c1a045SFlorian Westphal * to announce zero window in time before rmem limit kicks in. 246086c1a045SFlorian Westphal */ 246186c1a045SFlorian Westphal if (free_space < (allowed_space >> 4) || free_space < mss) 24621da177e4SLinus Torvalds return 0; 24631da177e4SLinus Torvalds } 24641da177e4SLinus Torvalds 24651da177e4SLinus Torvalds if (free_space > tp->rcv_ssthresh) 24661da177e4SLinus Torvalds free_space = tp->rcv_ssthresh; 24671da177e4SLinus Torvalds 24681da177e4SLinus Torvalds /* Don't do rounding if we are using window scaling, since the 24691da177e4SLinus Torvalds * scaled window will not line up with the MSS boundary anyway. 24701da177e4SLinus Torvalds */ 24711da177e4SLinus Torvalds window = tp->rcv_wnd; 24721da177e4SLinus Torvalds if (tp->rx_opt.rcv_wscale) { 24731da177e4SLinus Torvalds window = free_space; 24741da177e4SLinus Torvalds 24751da177e4SLinus Torvalds /* Advertise enough space so that it won't get scaled away. 24761da177e4SLinus Torvalds * Import case: prevent zero window announcement if 24771da177e4SLinus Torvalds * 1<<rcv_wscale > mss. 24781da177e4SLinus Torvalds */ 24791da177e4SLinus Torvalds if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window) 24801da177e4SLinus Torvalds window = (((window >> tp->rx_opt.rcv_wscale) + 1) 24811da177e4SLinus Torvalds << tp->rx_opt.rcv_wscale); 24821da177e4SLinus Torvalds } else { 24831da177e4SLinus Torvalds /* Get the largest window that is a nice multiple of mss. 24841da177e4SLinus Torvalds * Window clamp already applied above. 24851da177e4SLinus Torvalds * If our current window offering is within 1 mss of the 24861da177e4SLinus Torvalds * free space we just keep it. This prevents the divide 24871da177e4SLinus Torvalds * and multiply from happening most of the time. 24881da177e4SLinus Torvalds * We also don't do any window rounding when the free space 24891da177e4SLinus Torvalds * is too small. 24901da177e4SLinus Torvalds */ 24911da177e4SLinus Torvalds if (window <= free_space - mss || window > free_space) 24921da177e4SLinus Torvalds window = (free_space / mss) * mss; 249384565070SJohn Heffner else if (mss == full_space && 2494b92edbe0SEric Dumazet free_space > window + (full_space >> 1)) 249584565070SJohn Heffner window = free_space; 24961da177e4SLinus Torvalds } 24971da177e4SLinus Torvalds 24981da177e4SLinus Torvalds return window; 24991da177e4SLinus Torvalds } 25001da177e4SLinus Torvalds 2501cfea5a68SMartin KaFai Lau void tcp_skb_collapse_tstamp(struct sk_buff *skb, 2502082ac2d5SMartin KaFai Lau const struct sk_buff *next_skb) 2503082ac2d5SMartin KaFai Lau { 25040a2cf20cSSoheil Hassas Yeganeh if (unlikely(tcp_has_tx_tstamp(next_skb))) { 25050a2cf20cSSoheil Hassas Yeganeh const struct skb_shared_info *next_shinfo = 25060a2cf20cSSoheil Hassas Yeganeh skb_shinfo(next_skb); 2507082ac2d5SMartin KaFai Lau struct skb_shared_info *shinfo = skb_shinfo(skb); 2508082ac2d5SMartin KaFai Lau 25090a2cf20cSSoheil Hassas Yeganeh shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP; 2510082ac2d5SMartin KaFai Lau shinfo->tskey = next_shinfo->tskey; 25112de8023eSMartin KaFai Lau TCP_SKB_CB(skb)->txstamp_ack |= 25122de8023eSMartin KaFai Lau TCP_SKB_CB(next_skb)->txstamp_ack; 2513082ac2d5SMartin KaFai Lau } 2514082ac2d5SMartin KaFai Lau } 2515082ac2d5SMartin KaFai Lau 25164a17fc3aSIlpo Järvinen /* Collapses two adjacent SKB's during retransmission. */ 25174a17fc3aSIlpo Järvinen static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) 25181da177e4SLinus Torvalds { 25191da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2520fe067e8aSDavid S. Miller struct sk_buff *next_skb = tcp_write_queue_next(sk, skb); 2521058dc334SIlpo Järvinen int skb_size, next_skb_size; 25221da177e4SLinus Torvalds 2523058dc334SIlpo Järvinen skb_size = skb->len; 2524058dc334SIlpo Järvinen next_skb_size = next_skb->len; 25251da177e4SLinus Torvalds 2526058dc334SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); 25271da177e4SLinus Torvalds 25286859d494SIlpo Järvinen tcp_highest_sack_combine(sk, next_skb, skb); 2529a6963a6bSIlpo Järvinen 2530fe067e8aSDavid S. Miller tcp_unlink_write_queue(next_skb, sk); 25311da177e4SLinus Torvalds 2532*2331ccc5SEric Dumazet if (next_skb_size) 2533*2331ccc5SEric Dumazet skb_copy_bits(next_skb, 0, skb_put(skb, next_skb_size), 25341a4e2d09SArnaldo Carvalho de Melo next_skb_size); 25351da177e4SLinus Torvalds 253652d570aaSJarek Poplawski if (next_skb->ip_summed == CHECKSUM_PARTIAL) 253752d570aaSJarek Poplawski skb->ip_summed = CHECKSUM_PARTIAL; 25381da177e4SLinus Torvalds 253984fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 25401da177e4SLinus Torvalds skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); 25411da177e4SLinus Torvalds 25421da177e4SLinus Torvalds /* Update sequence range on original skb. */ 25431da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 25441da177e4SLinus Torvalds 2545e6c7d085SIlpo Järvinen /* Merge over control information. This moves PSH/FIN etc. over */ 25464de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; 25471da177e4SLinus Torvalds 25481da177e4SLinus Torvalds /* All done, get rid of second SKB and account for it so 25491da177e4SLinus Torvalds * packet counting does not break. 25501da177e4SLinus Torvalds */ 25514828e7f4SIlpo Järvinen TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; 2552a643b5d4SMartin KaFai Lau TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor; 2553b7689205SIlpo Järvinen 2554b7689205SIlpo Järvinen /* changed transmit queue under us so clear hints */ 2555ef9da47cSIlpo Järvinen tcp_clear_retrans_hints_partial(tp); 2556ef9da47cSIlpo Järvinen if (next_skb == tp->retransmit_skb_hint) 2557ef9da47cSIlpo Järvinen tp->retransmit_skb_hint = skb; 2558b7689205SIlpo Järvinen 2559797108d1SIlpo Järvinen tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); 2560797108d1SIlpo Järvinen 2561082ac2d5SMartin KaFai Lau tcp_skb_collapse_tstamp(skb, next_skb); 2562082ac2d5SMartin KaFai Lau 25633ab224beSHideo Aoki sk_wmem_free_skb(sk, next_skb); 25641da177e4SLinus Torvalds } 25651da177e4SLinus Torvalds 256667edfef7SAndi Kleen /* Check if coalescing SKBs is legal. */ 2567a2a385d6SEric Dumazet static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) 25684a17fc3aSIlpo Järvinen { 25694a17fc3aSIlpo Järvinen if (tcp_skb_pcount(skb) > 1) 2570a2a385d6SEric Dumazet return false; 25714a17fc3aSIlpo Järvinen if (skb_cloned(skb)) 2572a2a385d6SEric Dumazet return false; 25734a17fc3aSIlpo Järvinen if (skb == tcp_send_head(sk)) 2574a2a385d6SEric Dumazet return false; 2575*2331ccc5SEric Dumazet /* Some heuristics for collapsing over SACK'd could be invented */ 25764a17fc3aSIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 2577a2a385d6SEric Dumazet return false; 25784a17fc3aSIlpo Järvinen 2579a2a385d6SEric Dumazet return true; 25804a17fc3aSIlpo Järvinen } 25814a17fc3aSIlpo Järvinen 258267edfef7SAndi Kleen /* Collapse packets in the retransmit queue to make to create 258367edfef7SAndi Kleen * less packets on the wire. This is only done on retransmission. 258467edfef7SAndi Kleen */ 25854a17fc3aSIlpo Järvinen static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, 25864a17fc3aSIlpo Järvinen int space) 25874a17fc3aSIlpo Järvinen { 25884a17fc3aSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 25894a17fc3aSIlpo Järvinen struct sk_buff *skb = to, *tmp; 2590a2a385d6SEric Dumazet bool first = true; 25914a17fc3aSIlpo Järvinen 25924a17fc3aSIlpo Järvinen if (!sysctl_tcp_retrans_collapse) 25934a17fc3aSIlpo Järvinen return; 25944de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 25954a17fc3aSIlpo Järvinen return; 25964a17fc3aSIlpo Järvinen 25974a17fc3aSIlpo Järvinen tcp_for_write_queue_from_safe(skb, tmp, sk) { 25984a17fc3aSIlpo Järvinen if (!tcp_can_collapse(sk, skb)) 25994a17fc3aSIlpo Järvinen break; 26004a17fc3aSIlpo Järvinen 2601a643b5d4SMartin KaFai Lau if (!tcp_skb_can_collapse_to(to)) 2602a643b5d4SMartin KaFai Lau break; 2603a643b5d4SMartin KaFai Lau 26044a17fc3aSIlpo Järvinen space -= skb->len; 26054a17fc3aSIlpo Järvinen 26064a17fc3aSIlpo Järvinen if (first) { 2607a2a385d6SEric Dumazet first = false; 26084a17fc3aSIlpo Järvinen continue; 26094a17fc3aSIlpo Järvinen } 26104a17fc3aSIlpo Järvinen 26114a17fc3aSIlpo Järvinen if (space < 0) 26124a17fc3aSIlpo Järvinen break; 26134a17fc3aSIlpo Järvinen /* Punt if not enough space exists in the first SKB for 26144a17fc3aSIlpo Järvinen * the data in the second 26154a17fc3aSIlpo Järvinen */ 2616a21d4572SEric Dumazet if (skb->len > skb_availroom(to)) 26174a17fc3aSIlpo Järvinen break; 26184a17fc3aSIlpo Järvinen 26194a17fc3aSIlpo Järvinen if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) 26204a17fc3aSIlpo Järvinen break; 26214a17fc3aSIlpo Järvinen 26224a17fc3aSIlpo Järvinen tcp_collapse_retrans(sk, to); 26234a17fc3aSIlpo Järvinen } 26244a17fc3aSIlpo Järvinen } 26254a17fc3aSIlpo Järvinen 26261da177e4SLinus Torvalds /* This retransmits one SKB. Policy decisions and retransmit queue 26271da177e4SLinus Torvalds * state updates are done by the caller. Returns non-zero if an 26281da177e4SLinus Torvalds * error occurred which prevented the send. 26291da177e4SLinus Torvalds */ 263010d3be56SEric Dumazet int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) 26311da177e4SLinus Torvalds { 26325d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 263310d3be56SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 26347d227cd2SSridhar Samudrala unsigned int cur_mss; 263510d3be56SEric Dumazet int diff, len, err; 26361da177e4SLinus Torvalds 263710d3be56SEric Dumazet 263810d3be56SEric Dumazet /* Inconclusive MTU probe */ 263910d3be56SEric Dumazet if (icsk->icsk_mtup.probe_size) 26405d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 26415d424d5aSJohn Heffner 26421da177e4SLinus Torvalds /* Do not sent more than we queued. 1/4 is reserved for possible 2643caa20d9aSStephen Hemminger * copying overhead: fragmentation, tunneling, mangling etc. 26441da177e4SLinus Torvalds */ 26451da177e4SLinus Torvalds if (atomic_read(&sk->sk_wmem_alloc) > 2646ffb4d6c8SEric Dumazet min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), 2647ffb4d6c8SEric Dumazet sk->sk_sndbuf)) 26481da177e4SLinus Torvalds return -EAGAIN; 26491da177e4SLinus Torvalds 26501f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 26511f3279aeSEric Dumazet return -EBUSY; 26521f3279aeSEric Dumazet 26531da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 26541da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 26551da177e4SLinus Torvalds BUG(); 26561da177e4SLinus Torvalds if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 26571da177e4SLinus Torvalds return -ENOMEM; 26581da177e4SLinus Torvalds } 26591da177e4SLinus Torvalds 26607d227cd2SSridhar Samudrala if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 26617d227cd2SSridhar Samudrala return -EHOSTUNREACH; /* Routing failure or similar. */ 26627d227cd2SSridhar Samudrala 26630c54b85fSIlpo Järvinen cur_mss = tcp_current_mss(sk); 26647d227cd2SSridhar Samudrala 26651da177e4SLinus Torvalds /* If receiver has shrunk his window, and skb is out of 26661da177e4SLinus Torvalds * new window, do not retransmit it. The exception is the 26671da177e4SLinus Torvalds * case, when window is shrunk to zero. In this case 26681da177e4SLinus Torvalds * our retransmit serves as a zero window probe. 26691da177e4SLinus Torvalds */ 26709d4fb27dSJoe Perches if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && 26719d4fb27dSJoe Perches TCP_SKB_CB(skb)->seq != tp->snd_una) 26721da177e4SLinus Torvalds return -EAGAIN; 26731da177e4SLinus Torvalds 267410d3be56SEric Dumazet len = cur_mss * segs; 267510d3be56SEric Dumazet if (skb->len > len) { 267610d3be56SEric Dumazet if (tcp_fragment(sk, skb, len, cur_mss, GFP_ATOMIC)) 26771da177e4SLinus Torvalds return -ENOMEM; /* We'll try again later. */ 267802276f3cSIlpo Järvinen } else { 2679c52e2421SEric Dumazet if (skb_unclone(skb, GFP_ATOMIC)) 2680c52e2421SEric Dumazet return -ENOMEM; 268110d3be56SEric Dumazet 268210d3be56SEric Dumazet diff = tcp_skb_pcount(skb); 268310d3be56SEric Dumazet tcp_set_skb_tso_segs(skb, cur_mss); 268410d3be56SEric Dumazet diff -= tcp_skb_pcount(skb); 268510d3be56SEric Dumazet if (diff) 268610d3be56SEric Dumazet tcp_adjust_pcount(sk, skb, diff); 268710d3be56SEric Dumazet if (skb->len < cur_mss) 268810d3be56SEric Dumazet tcp_retrans_try_collapse(sk, skb, cur_mss); 26891da177e4SLinus Torvalds } 26901da177e4SLinus Torvalds 269149213555SDaniel Borkmann /* RFC3168, section 6.1.1.1. ECN fallback */ 269249213555SDaniel Borkmann if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN) 269349213555SDaniel Borkmann tcp_ecn_clear_syn(sk, skb); 269449213555SDaniel Borkmann 269550bceae9SThomas Graf /* make sure skb->data is aligned on arches that require it 269650bceae9SThomas Graf * and check if ack-trimming & collapsing extended the headroom 269750bceae9SThomas Graf * beyond what csum_start can cover. 269850bceae9SThomas Graf */ 269950bceae9SThomas Graf if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) || 270050bceae9SThomas Graf skb_headroom(skb) >= 0xFFFF)) { 270110a81980SEric Dumazet struct sk_buff *nskb; 270210a81980SEric Dumazet 270310a81980SEric Dumazet skb_mstamp_get(&skb->skb_mstamp); 270410a81980SEric Dumazet nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC); 2705c84a5711SYuchung Cheng err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) : 2706117632e6SEric Dumazet -ENOBUFS; 2707117632e6SEric Dumazet } else { 2708c84a5711SYuchung Cheng err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 2709117632e6SEric Dumazet } 2710c84a5711SYuchung Cheng 2711fc9f3501SEric Dumazet if (likely(!err)) { 271210d3be56SEric Dumazet segs = tcp_skb_pcount(skb); 271310d3be56SEric Dumazet 2714c84a5711SYuchung Cheng TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS; 2715fc9f3501SEric Dumazet /* Update global TCP statistics. */ 271610d3be56SEric Dumazet TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs); 2717fc9f3501SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 271802a1d6e7SEric Dumazet __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 271910d3be56SEric Dumazet tp->total_retrans += segs; 2720fc9f3501SEric Dumazet } 2721c84a5711SYuchung Cheng return err; 272293b174adSYuchung Cheng } 272393b174adSYuchung Cheng 272410d3be56SEric Dumazet int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) 272593b174adSYuchung Cheng { 272693b174adSYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 272710d3be56SEric Dumazet int err = __tcp_retransmit_skb(sk, skb, segs); 27281da177e4SLinus Torvalds 27291da177e4SLinus Torvalds if (err == 0) { 27301da177e4SLinus Torvalds #if FASTRETRANS_DEBUG > 0 27311da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 2732e87cc472SJoe Perches net_dbg_ratelimited("retrans_out leaked\n"); 27331da177e4SLinus Torvalds } 27341da177e4SLinus Torvalds #endif 27351da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 27361da177e4SLinus Torvalds tp->retrans_out += tcp_skb_pcount(skb); 27371da177e4SLinus Torvalds 27381da177e4SLinus Torvalds /* Save stamp of the first retransmit. */ 27391da177e4SLinus Torvalds if (!tp->retrans_stamp) 27407faee5c0SEric Dumazet tp->retrans_stamp = tcp_skb_timestamp(skb); 27411da177e4SLinus Torvalds 27421f3279aeSEric Dumazet } else if (err != -EBUSY) { 2743c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL); 27441da177e4SLinus Torvalds } 27456e08d5e3SYuchung Cheng 27466e08d5e3SYuchung Cheng if (tp->undo_retrans < 0) 27476e08d5e3SYuchung Cheng tp->undo_retrans = 0; 27486e08d5e3SYuchung Cheng tp->undo_retrans += tcp_skb_pcount(skb); 27491da177e4SLinus Torvalds return err; 27501da177e4SLinus Torvalds } 27511da177e4SLinus Torvalds 275267edfef7SAndi Kleen /* Check if we forward retransmits are possible in the current 275367edfef7SAndi Kleen * window/congestion state. 275467edfef7SAndi Kleen */ 2755a2a385d6SEric Dumazet static bool tcp_can_forward_retransmit(struct sock *sk) 2756b5afe7bcSIlpo Järvinen { 2757b5afe7bcSIlpo Järvinen const struct inet_connection_sock *icsk = inet_csk(sk); 2758cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 2759b5afe7bcSIlpo Järvinen 2760b5afe7bcSIlpo Järvinen /* Forward retransmissions are possible only during Recovery. */ 2761b5afe7bcSIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Recovery) 2762a2a385d6SEric Dumazet return false; 2763b5afe7bcSIlpo Järvinen 2764b5afe7bcSIlpo Järvinen /* No forward retransmissions in Reno are possible. */ 2765b5afe7bcSIlpo Järvinen if (tcp_is_reno(tp)) 2766a2a385d6SEric Dumazet return false; 2767b5afe7bcSIlpo Järvinen 2768b5afe7bcSIlpo Järvinen /* Yeah, we have to make difficult choice between forward transmission 2769b5afe7bcSIlpo Järvinen * and retransmission... Both ways have their merits... 2770b5afe7bcSIlpo Järvinen * 2771b5afe7bcSIlpo Järvinen * For now we do not retransmit anything, while we have some new 2772b5afe7bcSIlpo Järvinen * segments to send. In the other cases, follow rule 3 for 2773b5afe7bcSIlpo Järvinen * NextSeg() specified in RFC3517. 2774b5afe7bcSIlpo Järvinen */ 2775b5afe7bcSIlpo Järvinen 2776b5afe7bcSIlpo Järvinen if (tcp_may_send_now(sk)) 2777a2a385d6SEric Dumazet return false; 2778b5afe7bcSIlpo Järvinen 2779a2a385d6SEric Dumazet return true; 2780b5afe7bcSIlpo Järvinen } 2781b5afe7bcSIlpo Järvinen 27821da177e4SLinus Torvalds /* This gets called after a retransmit timeout, and the initially 27831da177e4SLinus Torvalds * retransmitted data is acknowledged. It tries to continue 27841da177e4SLinus Torvalds * resending the rest of the retransmit queue, until either 27851da177e4SLinus Torvalds * we've sent it all or the congestion window limit is reached. 27861da177e4SLinus Torvalds * If doing SACK, the first ACK which comes back for a timeout 27871da177e4SLinus Torvalds * based retransmit packet might feed us FACK information again. 27881da177e4SLinus Torvalds * If so, we use it to avoid unnecessarily retransmissions. 27891da177e4SLinus Torvalds */ 27901da177e4SLinus Torvalds void tcp_xmit_retransmit_queue(struct sock *sk) 27911da177e4SLinus Torvalds { 27926687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 27931da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 27941da177e4SLinus Torvalds struct sk_buff *skb; 27950e1c54c2SIlpo Järvinen struct sk_buff *hole = NULL; 2796a3d2e9f8SEric Dumazet u32 max_segs, last_lost; 279761eb55f4SIlpo Järvinen int mib_idx; 27980e1c54c2SIlpo Järvinen int fwd_rexmitting = 0; 27996a438bbeSStephen Hemminger 280045e77d31SIlpo Järvinen if (!tp->packets_out) 280145e77d31SIlpo Järvinen return; 280245e77d31SIlpo Järvinen 280308ebd172SIlpo Järvinen if (!tp->lost_out) 280408ebd172SIlpo Järvinen tp->retransmit_high = tp->snd_una; 280508ebd172SIlpo Järvinen 2806618d9f25SIlpo Järvinen if (tp->retransmit_skb_hint) { 28076a438bbeSStephen Hemminger skb = tp->retransmit_skb_hint; 2808618d9f25SIlpo Järvinen last_lost = TCP_SKB_CB(skb)->end_seq; 2809618d9f25SIlpo Järvinen if (after(last_lost, tp->retransmit_high)) 2810618d9f25SIlpo Järvinen last_lost = tp->retransmit_high; 2811618d9f25SIlpo Järvinen } else { 2812fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 2813618d9f25SIlpo Järvinen last_lost = tp->snd_una; 2814618d9f25SIlpo Järvinen } 28151da177e4SLinus Torvalds 2816ed6e7268SNeal Cardwell max_segs = tcp_tso_segs(sk, tcp_current_mss(sk)); 2817fe067e8aSDavid S. Miller tcp_for_write_queue_from(skb, sk) { 2818dca0aaf8SEric Dumazet __u8 sacked; 281910d3be56SEric Dumazet int segs; 28201da177e4SLinus Torvalds 2821fe067e8aSDavid S. Miller if (skb == tcp_send_head(sk)) 2822fe067e8aSDavid S. Miller break; 28236a438bbeSStephen Hemminger /* we could do better than to assign each time */ 282451456b29SIan Morris if (!hole) 28256a438bbeSStephen Hemminger tp->retransmit_skb_hint = skb; 28266a438bbeSStephen Hemminger 282710d3be56SEric Dumazet segs = tp->snd_cwnd - tcp_packets_in_flight(tp); 282810d3be56SEric Dumazet if (segs <= 0) 28291da177e4SLinus Torvalds return; 2830dca0aaf8SEric Dumazet sacked = TCP_SKB_CB(skb)->sacked; 2831a3d2e9f8SEric Dumazet /* In case tcp_shift_skb_data() have aggregated large skbs, 2832a3d2e9f8SEric Dumazet * we need to make sure not sending too bigs TSO packets 2833a3d2e9f8SEric Dumazet */ 2834a3d2e9f8SEric Dumazet segs = min_t(int, segs, max_segs); 28350e1c54c2SIlpo Järvinen 28360e1c54c2SIlpo Järvinen if (fwd_rexmitting) { 28370e1c54c2SIlpo Järvinen begin_fwd: 28380e1c54c2SIlpo Järvinen if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp))) 2839006f582cSIlpo Järvinen break; 28400e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFORWARDRETRANS; 28410e1c54c2SIlpo Järvinen 28420e1c54c2SIlpo Järvinen } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) { 2843618d9f25SIlpo Järvinen tp->retransmit_high = last_lost; 28440e1c54c2SIlpo Järvinen if (!tcp_can_forward_retransmit(sk)) 28450e1c54c2SIlpo Järvinen break; 28460e1c54c2SIlpo Järvinen /* Backtrack if necessary to non-L'ed skb */ 284700db4124SIan Morris if (hole) { 28480e1c54c2SIlpo Järvinen skb = hole; 28490e1c54c2SIlpo Järvinen hole = NULL; 28500e1c54c2SIlpo Järvinen } 28510e1c54c2SIlpo Järvinen fwd_rexmitting = 1; 28520e1c54c2SIlpo Järvinen goto begin_fwd; 28530e1c54c2SIlpo Järvinen 28540e1c54c2SIlpo Järvinen } else if (!(sacked & TCPCB_LOST)) { 285551456b29SIan Morris if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) 28560e1c54c2SIlpo Järvinen hole = skb; 285761eb55f4SIlpo Järvinen continue; 28581da177e4SLinus Torvalds 28590e1c54c2SIlpo Järvinen } else { 2860618d9f25SIlpo Järvinen last_lost = TCP_SKB_CB(skb)->end_seq; 28610e1c54c2SIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Loss) 28620e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFASTRETRANS; 28630e1c54c2SIlpo Järvinen else 28640e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; 28650e1c54c2SIlpo Järvinen } 28660e1c54c2SIlpo Järvinen 28670e1c54c2SIlpo Järvinen if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) 286861eb55f4SIlpo Järvinen continue; 286940b215e5SPavel Emelyanov 2870f9616c35SEric Dumazet if (tcp_small_queue_check(sk, skb, 1)) 2871f9616c35SEric Dumazet return; 2872f9616c35SEric Dumazet 287310d3be56SEric Dumazet if (tcp_retransmit_skb(sk, skb, segs)) 28741da177e4SLinus Torvalds return; 287524ab6becSYuchung Cheng 2876de1d6578SYuchung Cheng NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb)); 28771da177e4SLinus Torvalds 2878684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 2879a262f0cdSNandita Dukkipati tp->prr_out += tcp_skb_pcount(skb); 2880a262f0cdSNandita Dukkipati 2881fe067e8aSDavid S. Miller if (skb == tcp_write_queue_head(sk)) 2882463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 28833f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, 28843f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 28851da177e4SLinus Torvalds } 28861da177e4SLinus Torvalds } 28871da177e4SLinus Torvalds 2888d83769a5SEric Dumazet /* We allow to exceed memory limits for FIN packets to expedite 2889d83769a5SEric Dumazet * connection tear down and (memory) recovery. 2890845704a5SEric Dumazet * Otherwise tcp_send_fin() could be tempted to either delay FIN 2891845704a5SEric Dumazet * or even be forced to close flow without any FIN. 2892a6c5ea4cSEric Dumazet * In general, we want to allow one skb per socket to avoid hangs 2893a6c5ea4cSEric Dumazet * with edge trigger epoll() 2894d83769a5SEric Dumazet */ 2895a6c5ea4cSEric Dumazet void sk_forced_mem_schedule(struct sock *sk, int size) 2896d83769a5SEric Dumazet { 2897e805605cSJohannes Weiner int amt; 2898d83769a5SEric Dumazet 2899d83769a5SEric Dumazet if (size <= sk->sk_forward_alloc) 2900d83769a5SEric Dumazet return; 2901d83769a5SEric Dumazet amt = sk_mem_pages(size); 2902d83769a5SEric Dumazet sk->sk_forward_alloc += amt * SK_MEM_QUANTUM; 2903e805605cSJohannes Weiner sk_memory_allocated_add(sk, amt); 2904e805605cSJohannes Weiner 2905baac50bbSJohannes Weiner if (mem_cgroup_sockets_enabled && sk->sk_memcg) 2906baac50bbSJohannes Weiner mem_cgroup_charge_skmem(sk->sk_memcg, amt); 2907d83769a5SEric Dumazet } 2908d83769a5SEric Dumazet 2909845704a5SEric Dumazet /* Send a FIN. The caller locks the socket for us. 2910845704a5SEric Dumazet * We should try to send a FIN packet really hard, but eventually give up. 29111da177e4SLinus Torvalds */ 29121da177e4SLinus Torvalds void tcp_send_fin(struct sock *sk) 29131da177e4SLinus Torvalds { 2914845704a5SEric Dumazet struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk); 29151da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 29161da177e4SLinus Torvalds 2917845704a5SEric Dumazet /* Optimization, tack on the FIN if we have one skb in write queue and 2918845704a5SEric Dumazet * this skb was not yet sent, or we are under memory pressure. 2919845704a5SEric Dumazet * Note: in the latter case, FIN packet will be sent after a timeout, 2920845704a5SEric Dumazet * as TCP stack thinks it has already been transmitted. 29211da177e4SLinus Torvalds */ 2922b8da51ebSEric Dumazet if (tskb && (tcp_send_head(sk) || tcp_under_memory_pressure(sk))) { 2923845704a5SEric Dumazet coalesce: 2924845704a5SEric Dumazet TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN; 2925845704a5SEric Dumazet TCP_SKB_CB(tskb)->end_seq++; 29261da177e4SLinus Torvalds tp->write_seq++; 2927845704a5SEric Dumazet if (!tcp_send_head(sk)) { 2928845704a5SEric Dumazet /* This means tskb was already sent. 2929845704a5SEric Dumazet * Pretend we included the FIN on previous transmit. 2930845704a5SEric Dumazet * We need to set tp->snd_nxt to the value it would have 2931845704a5SEric Dumazet * if FIN had been sent. This is because retransmit path 2932845704a5SEric Dumazet * does not change tp->snd_nxt. 2933845704a5SEric Dumazet */ 2934845704a5SEric Dumazet tp->snd_nxt++; 2935845704a5SEric Dumazet return; 2936845704a5SEric Dumazet } 29371da177e4SLinus Torvalds } else { 2938845704a5SEric Dumazet skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation); 2939845704a5SEric Dumazet if (unlikely(!skb)) { 2940845704a5SEric Dumazet if (tskb) 2941845704a5SEric Dumazet goto coalesce; 2942845704a5SEric Dumazet return; 29431da177e4SLinus Torvalds } 2944d83769a5SEric Dumazet skb_reserve(skb, MAX_TCP_HEADER); 2945a6c5ea4cSEric Dumazet sk_forced_mem_schedule(sk, skb->truesize); 29461da177e4SLinus Torvalds /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 2947e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tp->write_seq, 2948a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_FIN); 29491da177e4SLinus Torvalds tcp_queue_skb(sk, skb); 29501da177e4SLinus Torvalds } 2951845704a5SEric Dumazet __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF); 29521da177e4SLinus Torvalds } 29531da177e4SLinus Torvalds 29541da177e4SLinus Torvalds /* We get here when a process closes a file descriptor (either due to 29551da177e4SLinus Torvalds * an explicit close() or as a byproduct of exit()'ing) and there 29561da177e4SLinus Torvalds * was unread data in the receive queue. This behavior is recommended 295765bb723cSGerrit Renker * by RFC 2525, section 2.17. -DaveM 29581da177e4SLinus Torvalds */ 2959dd0fc66fSAl Viro void tcp_send_active_reset(struct sock *sk, gfp_t priority) 29601da177e4SLinus Torvalds { 29611da177e4SLinus Torvalds struct sk_buff *skb; 29621da177e4SLinus Torvalds 29631da177e4SLinus Torvalds /* NOTE: No TCP options attached and we never retransmit this. */ 29641da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, priority); 29651da177e4SLinus Torvalds if (!skb) { 29664e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 29671da177e4SLinus Torvalds return; 29681da177e4SLinus Torvalds } 29691da177e4SLinus Torvalds 29701da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 29711da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 2972e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), 2973a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_RST); 2974675ee231SEric Dumazet skb_mstamp_get(&skb->skb_mstamp); 29751da177e4SLinus Torvalds /* Send it off. */ 2976dfb4b9dcSDavid S. Miller if (tcp_transmit_skb(sk, skb, 0, priority)) 29774e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 297826af65cbSSridhar Samudrala 297981cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); 29801da177e4SLinus Torvalds } 29811da177e4SLinus Torvalds 298267edfef7SAndi Kleen /* Send a crossed SYN-ACK during socket establishment. 298367edfef7SAndi Kleen * WARNING: This routine must only be called when we have already sent 29841da177e4SLinus Torvalds * a SYN packet that crossed the incoming SYN that caused this routine 29851da177e4SLinus Torvalds * to get called. If this assumption fails then the initial rcv_wnd 29861da177e4SLinus Torvalds * and rcv_wscale values will not be correct. 29871da177e4SLinus Torvalds */ 29881da177e4SLinus Torvalds int tcp_send_synack(struct sock *sk) 29891da177e4SLinus Torvalds { 29901da177e4SLinus Torvalds struct sk_buff *skb; 29911da177e4SLinus Torvalds 2992fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 299351456b29SIan Morris if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 299491df42beSJoe Perches pr_debug("%s: wrong queue state\n", __func__); 29951da177e4SLinus Torvalds return -EFAULT; 29961da177e4SLinus Torvalds } 29974de075e0SEric Dumazet if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { 29981da177e4SLinus Torvalds if (skb_cloned(skb)) { 29991da177e4SLinus Torvalds struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC); 300051456b29SIan Morris if (!nskb) 30011da177e4SLinus Torvalds return -ENOMEM; 3002fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 3003f4a775d1SEric Dumazet __skb_header_release(nskb); 3004fe067e8aSDavid S. Miller __tcp_add_write_queue_head(sk, nskb); 30053ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 30063ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 30073ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 30081da177e4SLinus Torvalds skb = nskb; 30091da177e4SLinus Torvalds } 30101da177e4SLinus Torvalds 30114de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; 3012735d3831SFlorian Westphal tcp_ecn_send_synack(sk, skb); 30131da177e4SLinus Torvalds } 3014dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 30151da177e4SLinus Torvalds } 30161da177e4SLinus Torvalds 30174aea39c1SEric Dumazet /** 30184aea39c1SEric Dumazet * tcp_make_synack - Prepare a SYN-ACK. 30194aea39c1SEric Dumazet * sk: listener socket 30204aea39c1SEric Dumazet * dst: dst entry attached to the SYNACK 30214aea39c1SEric Dumazet * req: request_sock pointer 30224aea39c1SEric Dumazet * 30234aea39c1SEric Dumazet * Allocate one skb and build a SYNACK packet. 30244aea39c1SEric Dumazet * @dst is consumed : Caller should not use it again. 30254aea39c1SEric Dumazet */ 30265d062de7SEric Dumazet struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst, 3027e6b4d113SWilliam Allen Simpson struct request_sock *req, 3028ca6fb065SEric Dumazet struct tcp_fastopen_cookie *foc, 3029b3d05147SEric Dumazet enum tcp_synack_type synack_type) 30301da177e4SLinus Torvalds { 30312e6599cbSArnaldo Carvalho de Melo struct inet_request_sock *ireq = inet_rsk(req); 30325d062de7SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 303380f03e27SEric Dumazet struct tcp_md5sig_key *md5 = NULL; 30345d062de7SEric Dumazet struct tcp_out_options opts; 30355d062de7SEric Dumazet struct sk_buff *skb; 3036bd0388aeSWilliam Allen Simpson int tcp_header_size; 30375d062de7SEric Dumazet struct tcphdr *th; 30385d062de7SEric Dumazet u16 user_mss; 3039f5fff5dcSTom Quetchenbach int mss; 30401da177e4SLinus Torvalds 3041ca6fb065SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 30424aea39c1SEric Dumazet if (unlikely(!skb)) { 30434aea39c1SEric Dumazet dst_release(dst); 30441da177e4SLinus Torvalds return NULL; 30454aea39c1SEric Dumazet } 30461da177e4SLinus Torvalds /* Reserve space for headers. */ 30471da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 30481da177e4SLinus Torvalds 3049b3d05147SEric Dumazet switch (synack_type) { 3050b3d05147SEric Dumazet case TCP_SYNACK_NORMAL: 30519e17f8a4SEric Dumazet skb_set_owner_w(skb, req_to_sk(req)); 3052b3d05147SEric Dumazet break; 3053b3d05147SEric Dumazet case TCP_SYNACK_COOKIE: 3054b3d05147SEric Dumazet /* Under synflood, we do not attach skb to a socket, 3055b3d05147SEric Dumazet * to avoid false sharing. 3056b3d05147SEric Dumazet */ 3057b3d05147SEric Dumazet break; 3058b3d05147SEric Dumazet case TCP_SYNACK_FASTOPEN: 3059ca6fb065SEric Dumazet /* sk is a const pointer, because we want to express multiple 3060ca6fb065SEric Dumazet * cpu might call us concurrently. 3061ca6fb065SEric Dumazet * sk->sk_wmem_alloc in an atomic, we can promote to rw. 3062ca6fb065SEric Dumazet */ 3063ca6fb065SEric Dumazet skb_set_owner_w(skb, (struct sock *)sk); 3064b3d05147SEric Dumazet break; 3065ca6fb065SEric Dumazet } 30664aea39c1SEric Dumazet skb_dst_set(skb, dst); 30671da177e4SLinus Torvalds 30680dbaee3bSDavid S. Miller mss = dst_metric_advmss(dst); 30695d062de7SEric Dumazet user_mss = READ_ONCE(tp->rx_opt.user_mss); 30705d062de7SEric Dumazet if (user_mss && user_mss < mss) 30715d062de7SEric Dumazet mss = user_mss; 3072f5fff5dcSTom Quetchenbach 307333ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 30748b5f12d0SFlorian Westphal #ifdef CONFIG_SYN_COOKIES 30758b5f12d0SFlorian Westphal if (unlikely(req->cookie_ts)) 30767faee5c0SEric Dumazet skb->skb_mstamp.stamp_jiffies = cookie_init_timestamp(req); 30778b5f12d0SFlorian Westphal else 30788b5f12d0SFlorian Westphal #endif 30797faee5c0SEric Dumazet skb_mstamp_get(&skb->skb_mstamp); 308080f03e27SEric Dumazet 308180f03e27SEric Dumazet #ifdef CONFIG_TCP_MD5SIG 308280f03e27SEric Dumazet rcu_read_lock(); 3083fd3a154aSEric Dumazet md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req)); 308480f03e27SEric Dumazet #endif 308558d607d3SEric Dumazet skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4); 308637bfbddaSEric Dumazet tcp_header_size = tcp_synack_options(req, mss, skb, &opts, md5, foc) + 308737bfbddaSEric Dumazet sizeof(*th); 308833ad798cSAdam Langley 3089aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 3090aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 30911da177e4SLinus Torvalds 3092ea1627c2SEric Dumazet th = (struct tcphdr *)skb->data; 30931da177e4SLinus Torvalds memset(th, 0, sizeof(struct tcphdr)); 30941da177e4SLinus Torvalds th->syn = 1; 30951da177e4SLinus Torvalds th->ack = 1; 30966ac705b1SEric Dumazet tcp_ecn_make_synack(req, th); 3097b44084c2SEric Dumazet th->source = htons(ireq->ir_num); 3098634fb979SEric Dumazet th->dest = ireq->ir_rmt_port; 3099e870a8efSIlpo Järvinen /* Setting of flags are superfluous here for callers (and ECE is 3100e870a8efSIlpo Järvinen * not even correctly set) 3101e870a8efSIlpo Järvinen */ 3102e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn, 3103a3433f35SChangli Gao TCPHDR_SYN | TCPHDR_ACK); 31044957faadSWilliam Allen Simpson 31051da177e4SLinus Torvalds th->seq = htonl(TCP_SKB_CB(skb)->seq); 31068336886fSJerry Chu /* XXX data is queued and acked as is. No buffer/window check */ 31078336886fSJerry Chu th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt); 31081da177e4SLinus Torvalds 31091da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 3110ed53d0abSEric Dumazet th->window = htons(min(req->rsk_rcv_wnd, 65535U)); 31115d062de7SEric Dumazet tcp_options_write((__be32 *)(th + 1), NULL, &opts); 31121da177e4SLinus Torvalds th->doff = (tcp_header_size >> 2); 311390bbcc60SEric Dumazet __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS); 3114cfb6eeb4SYOSHIFUJI Hideaki 3115cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 3116cfb6eeb4SYOSHIFUJI Hideaki /* Okay, we have all we need - do the md5 hash if needed */ 311780f03e27SEric Dumazet if (md5) 3118bd0388aeSWilliam Allen Simpson tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, 311939f8e58eSEric Dumazet md5, req_to_sk(req), skb); 312080f03e27SEric Dumazet rcu_read_unlock(); 3121cfb6eeb4SYOSHIFUJI Hideaki #endif 3122cfb6eeb4SYOSHIFUJI Hideaki 3123b50edd78SEric Dumazet /* Do not fool tcpdump (if any), clean our debris */ 3124b50edd78SEric Dumazet skb->tstamp.tv64 = 0; 31251da177e4SLinus Torvalds return skb; 31261da177e4SLinus Torvalds } 31274bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_make_synack); 31281da177e4SLinus Torvalds 312981164413SDaniel Borkmann static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst) 313081164413SDaniel Borkmann { 313181164413SDaniel Borkmann struct inet_connection_sock *icsk = inet_csk(sk); 313281164413SDaniel Borkmann const struct tcp_congestion_ops *ca; 313381164413SDaniel Borkmann u32 ca_key = dst_metric(dst, RTAX_CC_ALGO); 313481164413SDaniel Borkmann 313581164413SDaniel Borkmann if (ca_key == TCP_CA_UNSPEC) 313681164413SDaniel Borkmann return; 313781164413SDaniel Borkmann 313881164413SDaniel Borkmann rcu_read_lock(); 313981164413SDaniel Borkmann ca = tcp_ca_find_key(ca_key); 314081164413SDaniel Borkmann if (likely(ca && try_module_get(ca->owner))) { 314181164413SDaniel Borkmann module_put(icsk->icsk_ca_ops->owner); 314281164413SDaniel Borkmann icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst); 314381164413SDaniel Borkmann icsk->icsk_ca_ops = ca; 314481164413SDaniel Borkmann } 314581164413SDaniel Borkmann rcu_read_unlock(); 314681164413SDaniel Borkmann } 314781164413SDaniel Borkmann 314867edfef7SAndi Kleen /* Do all connect socket setups that can be done AF independent. */ 3149f7e56a76Sstephen hemminger static void tcp_connect_init(struct sock *sk) 31501da177e4SLinus Torvalds { 3151cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 31521da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 31531da177e4SLinus Torvalds __u8 rcv_wscale; 31541da177e4SLinus Torvalds 31551da177e4SLinus Torvalds /* We'll fix this up when we get a response from the other end. 31561da177e4SLinus Torvalds * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 31571da177e4SLinus Torvalds */ 31581da177e4SLinus Torvalds tp->tcp_header_len = sizeof(struct tcphdr) + 3159bb5b7c11SDavid S. Miller (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0); 31601da177e4SLinus Torvalds 3161cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 316200db4124SIan Morris if (tp->af_specific->md5_lookup(sk, sk)) 3163cfb6eeb4SYOSHIFUJI Hideaki tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 3164cfb6eeb4SYOSHIFUJI Hideaki #endif 3165cfb6eeb4SYOSHIFUJI Hideaki 31661da177e4SLinus Torvalds /* If user gave his TCP_MAXSEG, record it to clamp */ 31671da177e4SLinus Torvalds if (tp->rx_opt.user_mss) 31681da177e4SLinus Torvalds tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 31691da177e4SLinus Torvalds tp->max_window = 0; 31705d424d5aSJohn Heffner tcp_mtup_init(sk); 31711da177e4SLinus Torvalds tcp_sync_mss(sk, dst_mtu(dst)); 31721da177e4SLinus Torvalds 317381164413SDaniel Borkmann tcp_ca_dst_init(sk, dst); 317481164413SDaniel Borkmann 31751da177e4SLinus Torvalds if (!tp->window_clamp) 31761da177e4SLinus Torvalds tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 31770dbaee3bSDavid S. Miller tp->advmss = dst_metric_advmss(dst); 3178f5fff5dcSTom Quetchenbach if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss) 3179f5fff5dcSTom Quetchenbach tp->advmss = tp->rx_opt.user_mss; 3180f5fff5dcSTom Quetchenbach 31811da177e4SLinus Torvalds tcp_initialize_rcv_mss(sk); 31821da177e4SLinus Torvalds 3183e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 3184e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 3185e88c64f0SHagen Paul Pfeifer (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) 3186e88c64f0SHagen Paul Pfeifer tp->window_clamp = tcp_full_space(sk); 3187e88c64f0SHagen Paul Pfeifer 31881da177e4SLinus Torvalds tcp_select_initial_window(tcp_full_space(sk), 31891da177e4SLinus Torvalds tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 31901da177e4SLinus Torvalds &tp->rcv_wnd, 31911da177e4SLinus Torvalds &tp->window_clamp, 3192bb5b7c11SDavid S. Miller sysctl_tcp_window_scaling, 319331d12926Slaurent chavey &rcv_wscale, 319431d12926Slaurent chavey dst_metric(dst, RTAX_INITRWND)); 31951da177e4SLinus Torvalds 31961da177e4SLinus Torvalds tp->rx_opt.rcv_wscale = rcv_wscale; 31971da177e4SLinus Torvalds tp->rcv_ssthresh = tp->rcv_wnd; 31981da177e4SLinus Torvalds 31991da177e4SLinus Torvalds sk->sk_err = 0; 32001da177e4SLinus Torvalds sock_reset_flag(sk, SOCK_DONE); 32011da177e4SLinus Torvalds tp->snd_wnd = 0; 3202ee7537b6SHantzis Fotis tcp_init_wl(tp, 0); 32031da177e4SLinus Torvalds tp->snd_una = tp->write_seq; 32041da177e4SLinus Torvalds tp->snd_sml = tp->write_seq; 320533f5f57eSIlpo Järvinen tp->snd_up = tp->write_seq; 3206370816aeSPavel Emelyanov tp->snd_nxt = tp->write_seq; 3207ee995283SPavel Emelyanov 3208ee995283SPavel Emelyanov if (likely(!tp->repair)) 32091da177e4SLinus Torvalds tp->rcv_nxt = 0; 3210c7781a6eSAndrew Vagin else 3211c7781a6eSAndrew Vagin tp->rcv_tstamp = tcp_time_stamp; 3212ee995283SPavel Emelyanov tp->rcv_wup = tp->rcv_nxt; 3213ee995283SPavel Emelyanov tp->copied_seq = tp->rcv_nxt; 32141da177e4SLinus Torvalds 3215463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; 3216463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_retransmits = 0; 32171da177e4SLinus Torvalds tcp_clear_retrans(tp); 32181da177e4SLinus Torvalds } 32191da177e4SLinus Torvalds 3220783237e8SYuchung Cheng static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) 3221783237e8SYuchung Cheng { 3222783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3223783237e8SYuchung Cheng struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 3224783237e8SYuchung Cheng 3225783237e8SYuchung Cheng tcb->end_seq += skb->len; 3226f4a775d1SEric Dumazet __skb_header_release(skb); 3227783237e8SYuchung Cheng __tcp_add_write_queue_tail(sk, skb); 3228783237e8SYuchung Cheng sk->sk_wmem_queued += skb->truesize; 3229783237e8SYuchung Cheng sk_mem_charge(sk, skb->truesize); 3230783237e8SYuchung Cheng tp->write_seq = tcb->end_seq; 3231783237e8SYuchung Cheng tp->packets_out += tcp_skb_pcount(skb); 3232783237e8SYuchung Cheng } 3233783237e8SYuchung Cheng 3234783237e8SYuchung Cheng /* Build and send a SYN with data and (cached) Fast Open cookie. However, 3235783237e8SYuchung Cheng * queue a data-only packet after the regular SYN, such that regular SYNs 3236783237e8SYuchung Cheng * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges 3237783237e8SYuchung Cheng * only the SYN sequence, the data are retransmitted in the first ACK. 3238783237e8SYuchung Cheng * If cookie is not cached or other error occurs, falls back to send a 3239783237e8SYuchung Cheng * regular SYN with Fast Open cookie request option. 3240783237e8SYuchung Cheng */ 3241783237e8SYuchung Cheng static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) 3242783237e8SYuchung Cheng { 3243783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3244783237e8SYuchung Cheng struct tcp_fastopen_request *fo = tp->fastopen_req; 324507e100f9SEric Dumazet int syn_loss = 0, space, err = 0; 3246aab48743SYuchung Cheng unsigned long last_syn_loss = 0; 3247355a901eSEric Dumazet struct sk_buff *syn_data; 3248783237e8SYuchung Cheng 324967da22d2SYuchung Cheng tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ 3250aab48743SYuchung Cheng tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie, 3251aab48743SYuchung Cheng &syn_loss, &last_syn_loss); 3252aab48743SYuchung Cheng /* Recurring FO SYN losses: revert to regular handshake temporarily */ 3253aab48743SYuchung Cheng if (syn_loss > 1 && 3254aab48743SYuchung Cheng time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) { 3255aab48743SYuchung Cheng fo->cookie.len = -1; 3256aab48743SYuchung Cheng goto fallback; 3257aab48743SYuchung Cheng } 3258aab48743SYuchung Cheng 325967da22d2SYuchung Cheng if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE) 326067da22d2SYuchung Cheng fo->cookie.len = -1; 326167da22d2SYuchung Cheng else if (fo->cookie.len <= 0) 3262783237e8SYuchung Cheng goto fallback; 3263783237e8SYuchung Cheng 3264783237e8SYuchung Cheng /* MSS for SYN-data is based on cached MSS and bounded by PMTU and 3265783237e8SYuchung Cheng * user-MSS. Reserve maximum option space for middleboxes that add 3266783237e8SYuchung Cheng * private TCP options. The cost is reduced data space in SYN :( 3267783237e8SYuchung Cheng */ 3268783237e8SYuchung Cheng if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp) 3269783237e8SYuchung Cheng tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 32701b63edd6SYuchung Cheng space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) - 3271783237e8SYuchung Cheng MAX_TCP_OPTION_SPACE; 3272783237e8SYuchung Cheng 3273f5ddcbbbSEric Dumazet space = min_t(size_t, space, fo->size); 3274f5ddcbbbSEric Dumazet 3275f5ddcbbbSEric Dumazet /* limit to order-0 allocations */ 3276f5ddcbbbSEric Dumazet space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER)); 3277f5ddcbbbSEric Dumazet 3278eb934478SEric Dumazet syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false); 3279355a901eSEric Dumazet if (!syn_data) 3280783237e8SYuchung Cheng goto fallback; 3281355a901eSEric Dumazet syn_data->ip_summed = CHECKSUM_PARTIAL; 3282355a901eSEric Dumazet memcpy(syn_data->cb, syn->cb, sizeof(syn->cb)); 328307e100f9SEric Dumazet if (space) { 328407e100f9SEric Dumazet int copied = copy_from_iter(skb_put(syn_data, space), space, 328557be5bdaSAl Viro &fo->data->msg_iter); 328657be5bdaSAl Viro if (unlikely(!copied)) { 3287355a901eSEric Dumazet kfree_skb(syn_data); 3288783237e8SYuchung Cheng goto fallback; 3289783237e8SYuchung Cheng } 329057be5bdaSAl Viro if (copied != space) { 329157be5bdaSAl Viro skb_trim(syn_data, copied); 329257be5bdaSAl Viro space = copied; 329357be5bdaSAl Viro } 329407e100f9SEric Dumazet } 3295355a901eSEric Dumazet /* No more data pending in inet_wait_for_connect() */ 3296355a901eSEric Dumazet if (space == fo->size) 3297355a901eSEric Dumazet fo->data = NULL; 3298355a901eSEric Dumazet fo->copied = space; 3299783237e8SYuchung Cheng 3300355a901eSEric Dumazet tcp_connect_queue_skb(sk, syn_data); 3301355a901eSEric Dumazet 3302355a901eSEric Dumazet err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation); 3303355a901eSEric Dumazet 3304355a901eSEric Dumazet syn->skb_mstamp = syn_data->skb_mstamp; 3305355a901eSEric Dumazet 3306355a901eSEric Dumazet /* Now full SYN+DATA was cloned and sent (or not), 3307355a901eSEric Dumazet * remove the SYN from the original skb (syn_data) 3308355a901eSEric Dumazet * we keep in write queue in case of a retransmit, as we 3309355a901eSEric Dumazet * also have the SYN packet (with no data) in the same queue. 3310431a9124SEric Dumazet */ 3311355a901eSEric Dumazet TCP_SKB_CB(syn_data)->seq++; 3312355a901eSEric Dumazet TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH; 3313355a901eSEric Dumazet if (!err) { 331467da22d2SYuchung Cheng tp->syn_data = (fo->copied > 0); 3315f19c29e3SYuchung Cheng NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT); 3316783237e8SYuchung Cheng goto done; 3317783237e8SYuchung Cheng } 3318783237e8SYuchung Cheng 3319783237e8SYuchung Cheng fallback: 3320783237e8SYuchung Cheng /* Send a regular SYN with Fast Open cookie request option */ 3321783237e8SYuchung Cheng if (fo->cookie.len > 0) 3322783237e8SYuchung Cheng fo->cookie.len = 0; 3323783237e8SYuchung Cheng err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); 3324783237e8SYuchung Cheng if (err) 3325783237e8SYuchung Cheng tp->syn_fastopen = 0; 3326783237e8SYuchung Cheng done: 3327783237e8SYuchung Cheng fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ 3328783237e8SYuchung Cheng return err; 3329783237e8SYuchung Cheng } 3330783237e8SYuchung Cheng 333167edfef7SAndi Kleen /* Build a SYN and send it off. */ 33321da177e4SLinus Torvalds int tcp_connect(struct sock *sk) 33331da177e4SLinus Torvalds { 33341da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 33351da177e4SLinus Torvalds struct sk_buff *buff; 3336ee586811SEric Paris int err; 33371da177e4SLinus Torvalds 33381da177e4SLinus Torvalds tcp_connect_init(sk); 33391da177e4SLinus Torvalds 33402b916477SAndrey Vagin if (unlikely(tp->repair)) { 33412b916477SAndrey Vagin tcp_finish_connect(sk, NULL); 33422b916477SAndrey Vagin return 0; 33432b916477SAndrey Vagin } 33442b916477SAndrey Vagin 3345eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true); 3346355a901eSEric Dumazet if (unlikely(!buff)) 33471da177e4SLinus Torvalds return -ENOBUFS; 33481da177e4SLinus Torvalds 3349a3433f35SChangli Gao tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); 33507faee5c0SEric Dumazet tp->retrans_stamp = tcp_time_stamp; 3351783237e8SYuchung Cheng tcp_connect_queue_skb(sk, buff); 3352735d3831SFlorian Westphal tcp_ecn_send_syn(sk, buff); 33531da177e4SLinus Torvalds 3354783237e8SYuchung Cheng /* Send off SYN; include data in Fast Open. */ 3355783237e8SYuchung Cheng err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : 3356783237e8SYuchung Cheng tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); 3357ee586811SEric Paris if (err == -ECONNREFUSED) 3358ee586811SEric Paris return err; 3359bd37a088SWei Yongjun 3360bd37a088SWei Yongjun /* We change tp->snd_nxt after the tcp_transmit_skb() call 3361bd37a088SWei Yongjun * in order to make this packet get counted in tcpOutSegs. 3362bd37a088SWei Yongjun */ 3363bd37a088SWei Yongjun tp->snd_nxt = tp->write_seq; 3364bd37a088SWei Yongjun tp->pushed_seq = tp->write_seq; 336581cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); 33661da177e4SLinus Torvalds 33671da177e4SLinus Torvalds /* Timer for repeating the SYN until an answer. */ 33683f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 33693f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 33701da177e4SLinus Torvalds return 0; 33711da177e4SLinus Torvalds } 33724bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_connect); 33731da177e4SLinus Torvalds 33741da177e4SLinus Torvalds /* Send out a delayed ack, the caller does the policy checking 33751da177e4SLinus Torvalds * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 33761da177e4SLinus Torvalds * for details. 33771da177e4SLinus Torvalds */ 33781da177e4SLinus Torvalds void tcp_send_delayed_ack(struct sock *sk) 33791da177e4SLinus Torvalds { 3380463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 3381463c84b9SArnaldo Carvalho de Melo int ato = icsk->icsk_ack.ato; 33821da177e4SLinus Torvalds unsigned long timeout; 33831da177e4SLinus Torvalds 33849890092eSFlorian Westphal tcp_ca_event(sk, CA_EVENT_DELAYED_ACK); 33859890092eSFlorian Westphal 33861da177e4SLinus Torvalds if (ato > TCP_DELACK_MIN) { 3387463c84b9SArnaldo Carvalho de Melo const struct tcp_sock *tp = tcp_sk(sk); 33881da177e4SLinus Torvalds int max_ato = HZ / 2; 33891da177e4SLinus Torvalds 3390056834d9SIlpo Järvinen if (icsk->icsk_ack.pingpong || 3391056834d9SIlpo Järvinen (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 33921da177e4SLinus Torvalds max_ato = TCP_DELACK_MAX; 33931da177e4SLinus Torvalds 33941da177e4SLinus Torvalds /* Slow path, intersegment interval is "high". */ 33951da177e4SLinus Torvalds 33961da177e4SLinus Torvalds /* If some rtt estimate is known, use it to bound delayed ack. 3397463c84b9SArnaldo Carvalho de Melo * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 33981da177e4SLinus Torvalds * directly. 33991da177e4SLinus Torvalds */ 3400740b0f18SEric Dumazet if (tp->srtt_us) { 3401740b0f18SEric Dumazet int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3), 3402740b0f18SEric Dumazet TCP_DELACK_MIN); 34031da177e4SLinus Torvalds 34041da177e4SLinus Torvalds if (rtt < max_ato) 34051da177e4SLinus Torvalds max_ato = rtt; 34061da177e4SLinus Torvalds } 34071da177e4SLinus Torvalds 34081da177e4SLinus Torvalds ato = min(ato, max_ato); 34091da177e4SLinus Torvalds } 34101da177e4SLinus Torvalds 34111da177e4SLinus Torvalds /* Stay within the limit we were given */ 34121da177e4SLinus Torvalds timeout = jiffies + ato; 34131da177e4SLinus Torvalds 34141da177e4SLinus Torvalds /* Use new timeout only if there wasn't a older one earlier. */ 3415463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 34161da177e4SLinus Torvalds /* If delack timer was blocked or is about to expire, 34171da177e4SLinus Torvalds * send ACK now. 34181da177e4SLinus Torvalds */ 3419463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.blocked || 3420463c84b9SArnaldo Carvalho de Melo time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 34211da177e4SLinus Torvalds tcp_send_ack(sk); 34221da177e4SLinus Torvalds return; 34231da177e4SLinus Torvalds } 34241da177e4SLinus Torvalds 3425463c84b9SArnaldo Carvalho de Melo if (!time_before(timeout, icsk->icsk_ack.timeout)) 3426463c84b9SArnaldo Carvalho de Melo timeout = icsk->icsk_ack.timeout; 34271da177e4SLinus Torvalds } 3428463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 3429463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.timeout = timeout; 3430463c84b9SArnaldo Carvalho de Melo sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 34311da177e4SLinus Torvalds } 34321da177e4SLinus Torvalds 34331da177e4SLinus Torvalds /* This routine sends an ack and also updates the window. */ 34341da177e4SLinus Torvalds void tcp_send_ack(struct sock *sk) 34351da177e4SLinus Torvalds { 34361da177e4SLinus Torvalds struct sk_buff *buff; 34371da177e4SLinus Torvalds 3438058dc334SIlpo Järvinen /* If we have been reset, we may not send again. */ 3439058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3440058dc334SIlpo Järvinen return; 3441058dc334SIlpo Järvinen 34429890092eSFlorian Westphal tcp_ca_event(sk, CA_EVENT_NON_DELAYED_ACK); 34439890092eSFlorian Westphal 34441da177e4SLinus Torvalds /* We are not putting this on the write queue, so 34451da177e4SLinus Torvalds * tcp_transmit_skb() will set the ownership to this 34461da177e4SLinus Torvalds * sock. 34471da177e4SLinus Torvalds */ 34487450aaf6SEric Dumazet buff = alloc_skb(MAX_TCP_HEADER, 34497450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 34507450aaf6SEric Dumazet if (unlikely(!buff)) { 3451463c84b9SArnaldo Carvalho de Melo inet_csk_schedule_ack(sk); 3452463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; 34533f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 34543f421baaSArnaldo Carvalho de Melo TCP_DELACK_MAX, TCP_RTO_MAX); 34551da177e4SLinus Torvalds return; 34561da177e4SLinus Torvalds } 34571da177e4SLinus Torvalds 34581da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 34591da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 3460a3433f35SChangli Gao tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); 34611da177e4SLinus Torvalds 346298781965SEric Dumazet /* We do not want pure acks influencing TCP Small Queues or fq/pacing 346398781965SEric Dumazet * too much. 346498781965SEric Dumazet * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784 346598781965SEric Dumazet * We also avoid tcp_wfree() overhead (cache line miss accessing 346698781965SEric Dumazet * tp->tsq_flags) by using regular sock_wfree() 346798781965SEric Dumazet */ 346898781965SEric Dumazet skb_set_tcp_pure_ack(buff); 346998781965SEric Dumazet 34701da177e4SLinus Torvalds /* Send it off, this clears delayed acks for us. */ 34717faee5c0SEric Dumazet skb_mstamp_get(&buff->skb_mstamp); 34727450aaf6SEric Dumazet tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0); 34731da177e4SLinus Torvalds } 3474e3118e83SDaniel Borkmann EXPORT_SYMBOL_GPL(tcp_send_ack); 34751da177e4SLinus Torvalds 34761da177e4SLinus Torvalds /* This routine sends a packet with an out of date sequence 34771da177e4SLinus Torvalds * number. It assumes the other end will try to ack it. 34781da177e4SLinus Torvalds * 34791da177e4SLinus Torvalds * Question: what should we make while urgent mode? 34801da177e4SLinus Torvalds * 4.4BSD forces sending single byte of data. We cannot send 34811da177e4SLinus Torvalds * out of window data, because we have SND.NXT==SND.MAX... 34821da177e4SLinus Torvalds * 34831da177e4SLinus Torvalds * Current solution: to send TWO zero-length segments in urgent mode: 34841da177e4SLinus Torvalds * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 34851da177e4SLinus Torvalds * out-of-date with SND.UNA-1 to probe window. 34861da177e4SLinus Torvalds */ 3487e520af48SEric Dumazet static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib) 34881da177e4SLinus Torvalds { 34891da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 34901da177e4SLinus Torvalds struct sk_buff *skb; 34911da177e4SLinus Torvalds 34921da177e4SLinus Torvalds /* We don't queue it, tcp_transmit_skb() sets ownership. */ 34937450aaf6SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, 34947450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 349551456b29SIan Morris if (!skb) 34961da177e4SLinus Torvalds return -1; 34971da177e4SLinus Torvalds 34981da177e4SLinus Torvalds /* Reserve space for headers and set control bits. */ 34991da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 35001da177e4SLinus Torvalds /* Use a previous sequence. This should cause the other 35011da177e4SLinus Torvalds * end to send an ack. Don't queue or clone SKB, just 35021da177e4SLinus Torvalds * send it. 35031da177e4SLinus Torvalds */ 3504a3433f35SChangli Gao tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); 35057faee5c0SEric Dumazet skb_mstamp_get(&skb->skb_mstamp); 3506e2e8009fSRenato Westphal NET_INC_STATS(sock_net(sk), mib); 35077450aaf6SEric Dumazet return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0); 35081da177e4SLinus Torvalds } 35091da177e4SLinus Torvalds 3510ee995283SPavel Emelyanov void tcp_send_window_probe(struct sock *sk) 3511ee995283SPavel Emelyanov { 3512ee995283SPavel Emelyanov if (sk->sk_state == TCP_ESTABLISHED) { 3513ee995283SPavel Emelyanov tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; 3514e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE); 3515ee995283SPavel Emelyanov } 3516ee995283SPavel Emelyanov } 3517ee995283SPavel Emelyanov 351867edfef7SAndi Kleen /* Initiate keepalive or window probe from timer. */ 3519e520af48SEric Dumazet int tcp_write_wakeup(struct sock *sk, int mib) 35201da177e4SLinus Torvalds { 35211da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 35221da177e4SLinus Torvalds struct sk_buff *skb; 35231da177e4SLinus Torvalds 3524058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3525058dc334SIlpo Järvinen return -1; 3526058dc334SIlpo Järvinen 352700db4124SIan Morris skb = tcp_send_head(sk); 352800db4124SIan Morris if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { 35291da177e4SLinus Torvalds int err; 35300c54b85fSIlpo Järvinen unsigned int mss = tcp_current_mss(sk); 353190840defSIlpo Järvinen unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 35321da177e4SLinus Torvalds 35331da177e4SLinus Torvalds if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 35341da177e4SLinus Torvalds tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 35351da177e4SLinus Torvalds 35361da177e4SLinus Torvalds /* We are probing the opening of a window 35371da177e4SLinus Torvalds * but the window size is != 0 35381da177e4SLinus Torvalds * must have been a result SWS avoidance ( sender ) 35391da177e4SLinus Torvalds */ 35401da177e4SLinus Torvalds if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 35411da177e4SLinus Torvalds skb->len > mss) { 35421da177e4SLinus Torvalds seg_size = min(seg_size, mss); 35434de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 35446cc55e09SOctavian Purdila if (tcp_fragment(sk, skb, seg_size, mss, GFP_ATOMIC)) 35451da177e4SLinus Torvalds return -1; 35461da177e4SLinus Torvalds } else if (!tcp_skb_pcount(skb)) 35475bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss); 35481da177e4SLinus Torvalds 35494de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 3550dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 355166f5fe62SIlpo Järvinen if (!err) 355266f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 35531da177e4SLinus Torvalds return err; 35541da177e4SLinus Torvalds } else { 355533f5f57eSIlpo Järvinen if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) 3556e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 1, mib); 3557e520af48SEric Dumazet return tcp_xmit_probe_skb(sk, 0, mib); 35581da177e4SLinus Torvalds } 35591da177e4SLinus Torvalds } 35601da177e4SLinus Torvalds 35611da177e4SLinus Torvalds /* A window probe timeout has occurred. If window is not closed send 35621da177e4SLinus Torvalds * a partial packet else a zero probe. 35631da177e4SLinus Torvalds */ 35641da177e4SLinus Torvalds void tcp_send_probe0(struct sock *sk) 35651da177e4SLinus Torvalds { 3566463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 35671da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 3568c6214a97SNikolay Borisov struct net *net = sock_net(sk); 3569fcdd1cf4SEric Dumazet unsigned long probe_max; 35701da177e4SLinus Torvalds int err; 35711da177e4SLinus Torvalds 3572e520af48SEric Dumazet err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE); 35731da177e4SLinus Torvalds 3574fe067e8aSDavid S. Miller if (tp->packets_out || !tcp_send_head(sk)) { 35751da177e4SLinus Torvalds /* Cancel probe timer, if it is not required. */ 35766687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 0; 3577463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff = 0; 35781da177e4SLinus Torvalds return; 35791da177e4SLinus Torvalds } 35801da177e4SLinus Torvalds 35811da177e4SLinus Torvalds if (err <= 0) { 3582c6214a97SNikolay Borisov if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2) 3583463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff++; 35846687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out++; 3585fcdd1cf4SEric Dumazet probe_max = TCP_RTO_MAX; 35861da177e4SLinus Torvalds } else { 35871da177e4SLinus Torvalds /* If packet was not sent due to local congestion, 35886687e988SArnaldo Carvalho de Melo * do not backoff and do not remember icsk_probes_out. 35891da177e4SLinus Torvalds * Let local senders to fight for local resources. 35901da177e4SLinus Torvalds * 35911da177e4SLinus Torvalds * Use accumulated backoff yet. 35921da177e4SLinus Torvalds */ 35936687e988SArnaldo Carvalho de Melo if (!icsk->icsk_probes_out) 35946687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 1; 3595fcdd1cf4SEric Dumazet probe_max = TCP_RESOURCE_PROBE_INTERVAL; 35961da177e4SLinus Torvalds } 3597fcdd1cf4SEric Dumazet inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 359821c8fe99SEric Dumazet tcp_probe0_when(sk, probe_max), 3599fcdd1cf4SEric Dumazet TCP_RTO_MAX); 36001da177e4SLinus Torvalds } 36015db92c99SOctavian Purdila 3602ea3bea3aSEric Dumazet int tcp_rtx_synack(const struct sock *sk, struct request_sock *req) 36035db92c99SOctavian Purdila { 36045db92c99SOctavian Purdila const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific; 36055db92c99SOctavian Purdila struct flowi fl; 36065db92c99SOctavian Purdila int res; 36075db92c99SOctavian Purdila 360858d607d3SEric Dumazet tcp_rsk(req)->txhash = net_tx_rndhash(); 3609b3d05147SEric Dumazet res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL); 36105db92c99SOctavian Purdila if (!res) { 361190bbcc60SEric Dumazet __TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); 361202a1d6e7SEric Dumazet __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 36137e32b443SYuchung Cheng if (unlikely(tcp_passive_fastopen(sk))) 36147e32b443SYuchung Cheng tcp_sk(sk)->total_retrans++; 36155db92c99SOctavian Purdila } 36165db92c99SOctavian Purdila return res; 36175db92c99SOctavian Purdila } 36185db92c99SOctavian Purdila EXPORT_SYMBOL(tcp_rtx_synack); 3619