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 */ 2391da177e4SLinus Torvalds space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max); 240316c1592SStephen Hemminger space = min_t(u32, space, *window_clamp); 2411da177e4SLinus Torvalds while (space > 65535 && (*rcv_wscale) < 14) { 2421da177e4SLinus Torvalds space >>= 1; 2431da177e4SLinus Torvalds (*rcv_wscale)++; 2441da177e4SLinus Torvalds } 2451da177e4SLinus Torvalds } 2461da177e4SLinus Torvalds 2471da177e4SLinus Torvalds if (mss > (1 << *rcv_wscale)) { 24885f16525SYuchung Cheng if (!init_rcv_wnd) /* Use default unless specified otherwise */ 24985f16525SYuchung Cheng init_rcv_wnd = tcp_default_init_rwnd(mss); 250b1afde60SNandita Dukkipati *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss); 2511da177e4SLinus Torvalds } 2521da177e4SLinus Torvalds 2531da177e4SLinus Torvalds /* Set the clamp no higher than max representable value */ 2541da177e4SLinus Torvalds (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp); 2551da177e4SLinus Torvalds } 2564bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_select_initial_window); 2571da177e4SLinus Torvalds 2581da177e4SLinus Torvalds /* Chose a new window to advertise, update state in tcp_sock for the 2591da177e4SLinus Torvalds * socket, and return result with RFC1323 scaling applied. The return 2601da177e4SLinus Torvalds * value can be stuffed directly into th->window for an outgoing 2611da177e4SLinus Torvalds * frame. 2621da177e4SLinus Torvalds */ 26340efc6faSStephen Hemminger static u16 tcp_select_window(struct sock *sk) 2641da177e4SLinus Torvalds { 2651da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2668e165e20SFlorian Westphal u32 old_win = tp->rcv_wnd; 2671da177e4SLinus Torvalds u32 cur_win = tcp_receive_window(tp); 2681da177e4SLinus Torvalds u32 new_win = __tcp_select_window(sk); 2691da177e4SLinus Torvalds 2701da177e4SLinus Torvalds /* Never shrink the offered window */ 2711da177e4SLinus Torvalds if (new_win < cur_win) { 2721da177e4SLinus Torvalds /* Danger Will Robinson! 2731da177e4SLinus Torvalds * Don't update rcv_wup/rcv_wnd here or else 2741da177e4SLinus Torvalds * we will not be able to advertise a zero 2751da177e4SLinus Torvalds * window in time. --DaveM 2761da177e4SLinus Torvalds * 2771da177e4SLinus Torvalds * Relax Will Robinson. 2781da177e4SLinus Torvalds */ 2798e165e20SFlorian Westphal if (new_win == 0) 2808e165e20SFlorian Westphal NET_INC_STATS(sock_net(sk), 2818e165e20SFlorian Westphal LINUX_MIB_TCPWANTZEROWINDOWADV); 282607bfbf2SPatrick McHardy new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale); 2831da177e4SLinus Torvalds } 2841da177e4SLinus Torvalds tp->rcv_wnd = new_win; 2851da177e4SLinus Torvalds tp->rcv_wup = tp->rcv_nxt; 2861da177e4SLinus Torvalds 2871da177e4SLinus Torvalds /* Make sure we do not exceed the maximum possible 2881da177e4SLinus Torvalds * scaled window. 2891da177e4SLinus Torvalds */ 29015d99e02SRick Jones if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows) 2911da177e4SLinus Torvalds new_win = min(new_win, MAX_TCP_WINDOW); 2921da177e4SLinus Torvalds else 2931da177e4SLinus Torvalds new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); 2941da177e4SLinus Torvalds 2951da177e4SLinus Torvalds /* RFC1323 scaling applied */ 2961da177e4SLinus Torvalds new_win >>= tp->rx_opt.rcv_wscale; 2971da177e4SLinus Torvalds 2981da177e4SLinus Torvalds /* If we advertise zero window, disable fast path. */ 2998e165e20SFlorian Westphal if (new_win == 0) { 3001da177e4SLinus Torvalds tp->pred_flags = 0; 3018e165e20SFlorian Westphal if (old_win) 3028e165e20SFlorian Westphal NET_INC_STATS(sock_net(sk), 3038e165e20SFlorian Westphal LINUX_MIB_TCPTOZEROWINDOWADV); 3048e165e20SFlorian Westphal } else if (old_win == 0) { 3058e165e20SFlorian Westphal NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV); 3068e165e20SFlorian Westphal } 3071da177e4SLinus Torvalds 3081da177e4SLinus Torvalds return new_win; 3091da177e4SLinus Torvalds } 3101da177e4SLinus Torvalds 31167edfef7SAndi Kleen /* Packet ECN state for a SYN-ACK */ 312735d3831SFlorian Westphal static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb) 313bdf1ee5dSIlpo Järvinen { 31430e502a3SDaniel Borkmann const struct tcp_sock *tp = tcp_sk(sk); 31530e502a3SDaniel Borkmann 3164de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR; 317bdf1ee5dSIlpo Järvinen if (!(tp->ecn_flags & TCP_ECN_OK)) 3184de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE; 31930e502a3SDaniel Borkmann else if (tcp_ca_needs_ecn(sk)) 32030e502a3SDaniel Borkmann INET_ECN_xmit(sk); 321bdf1ee5dSIlpo Järvinen } 322bdf1ee5dSIlpo Järvinen 32367edfef7SAndi Kleen /* Packet ECN state for a SYN. */ 324735d3831SFlorian Westphal static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb) 325bdf1ee5dSIlpo Järvinen { 326bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 327f7b3bec6SFlorian Westphal bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 || 328f7b3bec6SFlorian Westphal tcp_ca_needs_ecn(sk); 329f7b3bec6SFlorian Westphal 330f7b3bec6SFlorian Westphal if (!use_ecn) { 331f7b3bec6SFlorian Westphal const struct dst_entry *dst = __sk_dst_get(sk); 332f7b3bec6SFlorian Westphal 333f7b3bec6SFlorian Westphal if (dst && dst_feature(dst, RTAX_FEATURE_ECN)) 334f7b3bec6SFlorian Westphal use_ecn = true; 335f7b3bec6SFlorian Westphal } 336bdf1ee5dSIlpo Järvinen 337bdf1ee5dSIlpo Järvinen tp->ecn_flags = 0; 338f7b3bec6SFlorian Westphal 339f7b3bec6SFlorian Westphal if (use_ecn) { 3404de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR; 341bdf1ee5dSIlpo Järvinen tp->ecn_flags = TCP_ECN_OK; 34230e502a3SDaniel Borkmann if (tcp_ca_needs_ecn(sk)) 34330e502a3SDaniel Borkmann INET_ECN_xmit(sk); 344bdf1ee5dSIlpo Järvinen } 345bdf1ee5dSIlpo Järvinen } 346bdf1ee5dSIlpo Järvinen 34749213555SDaniel Borkmann static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb) 34849213555SDaniel Borkmann { 34949213555SDaniel Borkmann if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback) 35049213555SDaniel Borkmann /* tp->ecn_flags are cleared at a later point in time when 35149213555SDaniel Borkmann * SYN ACK is ultimatively being received. 35249213555SDaniel Borkmann */ 35349213555SDaniel Borkmann TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR); 35449213555SDaniel Borkmann } 35549213555SDaniel Borkmann 356735d3831SFlorian Westphal static void 3576ac705b1SEric Dumazet tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th) 358bdf1ee5dSIlpo Järvinen { 3596ac705b1SEric Dumazet if (inet_rsk(req)->ecn_ok) 360bdf1ee5dSIlpo Järvinen th->ece = 1; 361bdf1ee5dSIlpo Järvinen } 362bdf1ee5dSIlpo Järvinen 36367edfef7SAndi Kleen /* Set up ECN state for a packet on a ESTABLISHED socket that is about to 36467edfef7SAndi Kleen * be sent. 36567edfef7SAndi Kleen */ 366735d3831SFlorian Westphal static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb, 367bdf1ee5dSIlpo Järvinen int tcp_header_len) 368bdf1ee5dSIlpo Järvinen { 369bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 370bdf1ee5dSIlpo Järvinen 371bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_OK) { 372bdf1ee5dSIlpo Järvinen /* Not-retransmitted data segment: set ECT and inject CWR. */ 373bdf1ee5dSIlpo Järvinen if (skb->len != tcp_header_len && 374bdf1ee5dSIlpo Järvinen !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { 375bdf1ee5dSIlpo Järvinen INET_ECN_xmit(sk); 376bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) { 377bdf1ee5dSIlpo Järvinen tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; 378bdf1ee5dSIlpo Järvinen tcp_hdr(skb)->cwr = 1; 379bdf1ee5dSIlpo Järvinen skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; 380bdf1ee5dSIlpo Järvinen } 38130e502a3SDaniel Borkmann } else if (!tcp_ca_needs_ecn(sk)) { 382bdf1ee5dSIlpo Järvinen /* ACK or retransmitted segment: clear ECT|CE */ 383bdf1ee5dSIlpo Järvinen INET_ECN_dontxmit(sk); 384bdf1ee5dSIlpo Järvinen } 385bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) 386bdf1ee5dSIlpo Järvinen tcp_hdr(skb)->ece = 1; 387bdf1ee5dSIlpo Järvinen } 388bdf1ee5dSIlpo Järvinen } 389bdf1ee5dSIlpo Järvinen 390e870a8efSIlpo Järvinen /* Constructs common control bits of non-data skb. If SYN/FIN is present, 391e870a8efSIlpo Järvinen * auto increment end seqno. 392e870a8efSIlpo Järvinen */ 393e870a8efSIlpo Järvinen static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags) 394e870a8efSIlpo Järvinen { 3952e8e18efSDavid S. Miller skb->ip_summed = CHECKSUM_PARTIAL; 396e870a8efSIlpo Järvinen skb->csum = 0; 397e870a8efSIlpo Järvinen 3984de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags; 399e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->sacked = 0; 400e870a8efSIlpo Järvinen 401cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, 1); 402e870a8efSIlpo Järvinen 403e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->seq = seq; 404a3433f35SChangli Gao if (flags & (TCPHDR_SYN | TCPHDR_FIN)) 405e870a8efSIlpo Järvinen seq++; 406e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->end_seq = seq; 407e870a8efSIlpo Järvinen } 408e870a8efSIlpo Järvinen 409a2a385d6SEric Dumazet static inline bool tcp_urg_mode(const struct tcp_sock *tp) 41033f5f57eSIlpo Järvinen { 41133f5f57eSIlpo Järvinen return tp->snd_una != tp->snd_up; 41233f5f57eSIlpo Järvinen } 41333f5f57eSIlpo Järvinen 41433ad798cSAdam Langley #define OPTION_SACK_ADVERTISE (1 << 0) 41533ad798cSAdam Langley #define OPTION_TS (1 << 1) 41633ad798cSAdam Langley #define OPTION_MD5 (1 << 2) 41789e95a61SOri Finkelman #define OPTION_WSCALE (1 << 3) 4182100c8d2SYuchung Cheng #define OPTION_FAST_OPEN_COOKIE (1 << 8) 41933ad798cSAdam Langley 42033ad798cSAdam Langley struct tcp_out_options { 4212100c8d2SYuchung Cheng u16 options; /* bit field of OPTION_* */ 4222100c8d2SYuchung Cheng u16 mss; /* 0 to disable */ 42333ad798cSAdam Langley u8 ws; /* window scale, 0 to disable */ 42433ad798cSAdam Langley u8 num_sack_blocks; /* number of SACK blocks to include */ 425bd0388aeSWilliam Allen Simpson u8 hash_size; /* bytes in hash_location */ 426bd0388aeSWilliam Allen Simpson __u8 *hash_location; /* temporary pointer, overloaded */ 4272100c8d2SYuchung Cheng __u32 tsval, tsecr; /* need to include OPTION_TS */ 4282100c8d2SYuchung Cheng struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */ 42933ad798cSAdam Langley }; 43033ad798cSAdam Langley 43167edfef7SAndi Kleen /* Write previously computed TCP options to the packet. 43267edfef7SAndi Kleen * 43367edfef7SAndi Kleen * Beware: Something in the Internet is very sensitive to the ordering of 434fd6149d3SIlpo Järvinen * TCP options, we learned this through the hard way, so be careful here. 435fd6149d3SIlpo Järvinen * Luckily we can at least blame others for their non-compliance but from 4368e3bff96Sstephen hemminger * inter-operability perspective it seems that we're somewhat stuck with 437fd6149d3SIlpo Järvinen * the ordering which we have been using if we want to keep working with 438fd6149d3SIlpo Järvinen * those broken things (not that it currently hurts anybody as there isn't 439fd6149d3SIlpo Järvinen * particular reason why the ordering would need to be changed). 440fd6149d3SIlpo Järvinen * 441fd6149d3SIlpo Järvinen * At least SACK_PERM as the first option is known to lead to a disaster 442fd6149d3SIlpo Järvinen * (but it may well be that other scenarios fail similarly). 443fd6149d3SIlpo Järvinen */ 44433ad798cSAdam Langley static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp, 445bd0388aeSWilliam Allen Simpson struct tcp_out_options *opts) 446bd0388aeSWilliam Allen Simpson { 4472100c8d2SYuchung Cheng u16 options = opts->options; /* mungable copy */ 448bd0388aeSWilliam Allen Simpson 449bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_MD5 & options)) { 4501a2c6181SChristoph Paasch *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | 4511a2c6181SChristoph Paasch (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG); 452bd0388aeSWilliam Allen Simpson /* overload cookie hash location */ 453bd0388aeSWilliam Allen Simpson opts->hash_location = (__u8 *)ptr; 45433ad798cSAdam Langley ptr += 4; 45533ad798cSAdam Langley } 45633ad798cSAdam Langley 457fd6149d3SIlpo Järvinen if (unlikely(opts->mss)) { 458fd6149d3SIlpo Järvinen *ptr++ = htonl((TCPOPT_MSS << 24) | 459fd6149d3SIlpo Järvinen (TCPOLEN_MSS << 16) | 460fd6149d3SIlpo Järvinen opts->mss); 461fd6149d3SIlpo Järvinen } 462fd6149d3SIlpo Järvinen 463bd0388aeSWilliam Allen Simpson if (likely(OPTION_TS & options)) { 464bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 46533ad798cSAdam Langley *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | 46633ad798cSAdam Langley (TCPOLEN_SACK_PERM << 16) | 46733ad798cSAdam Langley (TCPOPT_TIMESTAMP << 8) | 46833ad798cSAdam Langley TCPOLEN_TIMESTAMP); 469bd0388aeSWilliam Allen Simpson options &= ~OPTION_SACK_ADVERTISE; 47033ad798cSAdam Langley } else { 471496c98dfSYOSHIFUJI Hideaki *ptr++ = htonl((TCPOPT_NOP << 24) | 47240efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 47340efc6faSStephen Hemminger (TCPOPT_TIMESTAMP << 8) | 47440efc6faSStephen Hemminger TCPOLEN_TIMESTAMP); 47540efc6faSStephen Hemminger } 47633ad798cSAdam Langley *ptr++ = htonl(opts->tsval); 47733ad798cSAdam Langley *ptr++ = htonl(opts->tsecr); 47833ad798cSAdam Langley } 47933ad798cSAdam Langley 480bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 48133ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 48233ad798cSAdam Langley (TCPOPT_NOP << 16) | 48333ad798cSAdam Langley (TCPOPT_SACK_PERM << 8) | 48433ad798cSAdam Langley TCPOLEN_SACK_PERM); 48533ad798cSAdam Langley } 48633ad798cSAdam Langley 487bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_WSCALE & options)) { 48833ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 48933ad798cSAdam Langley (TCPOPT_WINDOW << 16) | 49033ad798cSAdam Langley (TCPOLEN_WINDOW << 8) | 49133ad798cSAdam Langley opts->ws); 49233ad798cSAdam Langley } 49333ad798cSAdam Langley 49433ad798cSAdam Langley if (unlikely(opts->num_sack_blocks)) { 49533ad798cSAdam Langley struct tcp_sack_block *sp = tp->rx_opt.dsack ? 49633ad798cSAdam Langley tp->duplicate_sack : tp->selective_acks; 49740efc6faSStephen Hemminger int this_sack; 49840efc6faSStephen Hemminger 49940efc6faSStephen Hemminger *ptr++ = htonl((TCPOPT_NOP << 24) | 50040efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 50140efc6faSStephen Hemminger (TCPOPT_SACK << 8) | 50233ad798cSAdam Langley (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * 50340efc6faSStephen Hemminger TCPOLEN_SACK_PERBLOCK))); 5042de979bdSStephen Hemminger 50533ad798cSAdam Langley for (this_sack = 0; this_sack < opts->num_sack_blocks; 50633ad798cSAdam Langley ++this_sack) { 50740efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].start_seq); 50840efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].end_seq); 50940efc6faSStephen Hemminger } 5102de979bdSStephen Hemminger 51140efc6faSStephen Hemminger tp->rx_opt.dsack = 0; 51240efc6faSStephen Hemminger } 5132100c8d2SYuchung Cheng 5142100c8d2SYuchung Cheng if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) { 5152100c8d2SYuchung Cheng struct tcp_fastopen_cookie *foc = opts->fastopen_cookie; 5167f9b838bSDaniel Lee u8 *p = (u8 *)ptr; 5177f9b838bSDaniel Lee u32 len; /* Fast Open option length */ 5182100c8d2SYuchung Cheng 5197f9b838bSDaniel Lee if (foc->exp) { 5207f9b838bSDaniel Lee len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len; 5217f9b838bSDaniel Lee *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) | 5222100c8d2SYuchung Cheng TCPOPT_FASTOPEN_MAGIC); 5237f9b838bSDaniel Lee p += TCPOLEN_EXP_FASTOPEN_BASE; 5247f9b838bSDaniel Lee } else { 5257f9b838bSDaniel Lee len = TCPOLEN_FASTOPEN_BASE + foc->len; 5267f9b838bSDaniel Lee *p++ = TCPOPT_FASTOPEN; 5277f9b838bSDaniel Lee *p++ = len; 5282100c8d2SYuchung Cheng } 5297f9b838bSDaniel Lee 5307f9b838bSDaniel Lee memcpy(p, foc->val, foc->len); 5317f9b838bSDaniel Lee if ((len & 3) == 2) { 5327f9b838bSDaniel Lee p[foc->len] = TCPOPT_NOP; 5337f9b838bSDaniel Lee p[foc->len + 1] = TCPOPT_NOP; 5347f9b838bSDaniel Lee } 5357f9b838bSDaniel Lee ptr += (len + 3) >> 2; 5362100c8d2SYuchung Cheng } 53740efc6faSStephen Hemminger } 53840efc6faSStephen Hemminger 53967edfef7SAndi Kleen /* Compute TCP options for SYN packets. This is not the final 54067edfef7SAndi Kleen * network wire format yet. 54167edfef7SAndi Kleen */ 54295c96174SEric Dumazet static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb, 54333ad798cSAdam Langley struct tcp_out_options *opts, 544cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 545cf533ea5SEric Dumazet { 54633ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 54795c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 548783237e8SYuchung Cheng struct tcp_fastopen_request *fastopen = tp->fastopen_req; 54933ad798cSAdam Langley 550cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 55133ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 55233ad798cSAdam Langley if (*md5) { 55333ad798cSAdam Langley opts->options |= OPTION_MD5; 554bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 555cfb6eeb4SYOSHIFUJI Hideaki } 55633ad798cSAdam Langley #else 55733ad798cSAdam Langley *md5 = NULL; 558cfb6eeb4SYOSHIFUJI Hideaki #endif 55933ad798cSAdam Langley 56033ad798cSAdam Langley /* We always get an MSS option. The option bytes which will be seen in 56133ad798cSAdam Langley * normal data packets should timestamps be used, must be in the MSS 56233ad798cSAdam Langley * advertised. But we subtract them from tp->mss_cache so that 56333ad798cSAdam Langley * calculations in tcp_sendmsg are simpler etc. So account for this 56433ad798cSAdam Langley * fact here if necessary. If we don't do this correctly, as a 56533ad798cSAdam Langley * receiver we won't recognize data packets as being full sized when we 56633ad798cSAdam Langley * should, and thus we won't abide by the delayed ACK rules correctly. 56733ad798cSAdam Langley * SACKs don't matter, we never delay an ACK when we have any of those 56833ad798cSAdam Langley * going out. */ 56933ad798cSAdam Langley opts->mss = tcp_advertise_mss(sk); 570bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 57133ad798cSAdam Langley 57251456b29SIan Morris if (likely(sysctl_tcp_timestamps && !*md5)) { 57333ad798cSAdam Langley opts->options |= OPTION_TS; 5747faee5c0SEric Dumazet opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset; 57533ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 576bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 57733ad798cSAdam Langley } 578bb5b7c11SDavid S. Miller if (likely(sysctl_tcp_window_scaling)) { 57933ad798cSAdam Langley opts->ws = tp->rx_opt.rcv_wscale; 58089e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 581bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 58233ad798cSAdam Langley } 583bb5b7c11SDavid S. Miller if (likely(sysctl_tcp_sack)) { 58433ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 585b32d1310SDavid S. Miller if (unlikely(!(OPTION_TS & opts->options))) 586bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 58733ad798cSAdam Langley } 58833ad798cSAdam Langley 589783237e8SYuchung Cheng if (fastopen && fastopen->cookie.len >= 0) { 5902646c831SDaniel Lee u32 need = fastopen->cookie.len; 5912646c831SDaniel Lee 5922646c831SDaniel Lee need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE : 5932646c831SDaniel Lee TCPOLEN_FASTOPEN_BASE; 594783237e8SYuchung Cheng need = (need + 3) & ~3U; /* Align to 32 bits */ 595783237e8SYuchung Cheng if (remaining >= need) { 596783237e8SYuchung Cheng opts->options |= OPTION_FAST_OPEN_COOKIE; 597783237e8SYuchung Cheng opts->fastopen_cookie = &fastopen->cookie; 598783237e8SYuchung Cheng remaining -= need; 599783237e8SYuchung Cheng tp->syn_fastopen = 1; 6002646c831SDaniel Lee tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0; 601783237e8SYuchung Cheng } 602783237e8SYuchung Cheng } 603bd0388aeSWilliam Allen Simpson 604bd0388aeSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 60533ad798cSAdam Langley } 60633ad798cSAdam Langley 60767edfef7SAndi Kleen /* Set up TCP options for SYN-ACKs. */ 60837bfbddaSEric Dumazet static unsigned int tcp_synack_options(struct request_sock *req, 60995c96174SEric Dumazet unsigned int mss, struct sk_buff *skb, 61033ad798cSAdam Langley struct tcp_out_options *opts, 61180f03e27SEric Dumazet const struct tcp_md5sig_key *md5, 6128336886fSJerry Chu struct tcp_fastopen_cookie *foc) 6134957faadSWilliam Allen Simpson { 61433ad798cSAdam Langley struct inet_request_sock *ireq = inet_rsk(req); 61595c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 61633ad798cSAdam Langley 61733ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 61880f03e27SEric Dumazet if (md5) { 61933ad798cSAdam Langley opts->options |= OPTION_MD5; 6204957faadSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 6214957faadSWilliam Allen Simpson 6224957faadSWilliam Allen Simpson /* We can't fit any SACK blocks in a packet with MD5 + TS 6234957faadSWilliam Allen Simpson * options. There was discussion about disabling SACK 6244957faadSWilliam Allen Simpson * rather than TS in order to fit in better with old, 6254957faadSWilliam Allen Simpson * buggy kernels, but that was deemed to be unnecessary. 6264957faadSWilliam Allen Simpson */ 627de213e5eSEric Dumazet ireq->tstamp_ok &= !ireq->sack_ok; 62833ad798cSAdam Langley } 62933ad798cSAdam Langley #endif 63033ad798cSAdam Langley 6314957faadSWilliam Allen Simpson /* We always send an MSS option. */ 63233ad798cSAdam Langley opts->mss = mss; 6334957faadSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 63433ad798cSAdam Langley 63533ad798cSAdam Langley if (likely(ireq->wscale_ok)) { 63633ad798cSAdam Langley opts->ws = ireq->rcv_wscale; 63789e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 6384957faadSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 63933ad798cSAdam Langley } 640de213e5eSEric Dumazet if (likely(ireq->tstamp_ok)) { 64133ad798cSAdam Langley opts->options |= OPTION_TS; 6427faee5c0SEric Dumazet opts->tsval = tcp_skb_timestamp(skb); 64333ad798cSAdam Langley opts->tsecr = req->ts_recent; 6444957faadSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 64533ad798cSAdam Langley } 64633ad798cSAdam Langley if (likely(ireq->sack_ok)) { 64733ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 648de213e5eSEric Dumazet if (unlikely(!ireq->tstamp_ok)) 6494957faadSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 65033ad798cSAdam Langley } 6517f9b838bSDaniel Lee if (foc != NULL && foc->len >= 0) { 6527f9b838bSDaniel Lee u32 need = foc->len; 6537f9b838bSDaniel Lee 6547f9b838bSDaniel Lee need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE : 6557f9b838bSDaniel Lee TCPOLEN_FASTOPEN_BASE; 6568336886fSJerry Chu need = (need + 3) & ~3U; /* Align to 32 bits */ 6578336886fSJerry Chu if (remaining >= need) { 6588336886fSJerry Chu opts->options |= OPTION_FAST_OPEN_COOKIE; 6598336886fSJerry Chu opts->fastopen_cookie = foc; 6608336886fSJerry Chu remaining -= need; 6618336886fSJerry Chu } 6628336886fSJerry Chu } 6634957faadSWilliam Allen Simpson 6644957faadSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 66533ad798cSAdam Langley } 66633ad798cSAdam Langley 66767edfef7SAndi Kleen /* Compute TCP options for ESTABLISHED sockets. This is not the 66867edfef7SAndi Kleen * final wire format yet. 66967edfef7SAndi Kleen */ 67095c96174SEric Dumazet static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb, 67133ad798cSAdam Langley struct tcp_out_options *opts, 672cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 673cf533ea5SEric Dumazet { 67433ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 67595c96174SEric Dumazet unsigned int size = 0; 676cabeccbdSIlpo Järvinen unsigned int eff_sacks; 67733ad798cSAdam Langley 6785843ef42SAndi Kleen opts->options = 0; 6795843ef42SAndi Kleen 68033ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 68133ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 68233ad798cSAdam Langley if (unlikely(*md5)) { 68333ad798cSAdam Langley opts->options |= OPTION_MD5; 68433ad798cSAdam Langley size += TCPOLEN_MD5SIG_ALIGNED; 68533ad798cSAdam Langley } 68633ad798cSAdam Langley #else 68733ad798cSAdam Langley *md5 = NULL; 68833ad798cSAdam Langley #endif 68933ad798cSAdam Langley 69033ad798cSAdam Langley if (likely(tp->rx_opt.tstamp_ok)) { 69133ad798cSAdam Langley opts->options |= OPTION_TS; 6927faee5c0SEric Dumazet opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0; 69333ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 69433ad798cSAdam Langley size += TCPOLEN_TSTAMP_ALIGNED; 69533ad798cSAdam Langley } 69633ad798cSAdam Langley 697cabeccbdSIlpo Järvinen eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; 698cabeccbdSIlpo Järvinen if (unlikely(eff_sacks)) { 69995c96174SEric Dumazet const unsigned int remaining = MAX_TCP_OPTION_SPACE - size; 70033ad798cSAdam Langley opts->num_sack_blocks = 70195c96174SEric Dumazet min_t(unsigned int, eff_sacks, 70233ad798cSAdam Langley (remaining - TCPOLEN_SACK_BASE_ALIGNED) / 70333ad798cSAdam Langley TCPOLEN_SACK_PERBLOCK); 70433ad798cSAdam Langley size += TCPOLEN_SACK_BASE_ALIGNED + 70533ad798cSAdam Langley opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; 70633ad798cSAdam Langley } 70733ad798cSAdam Langley 70833ad798cSAdam Langley return size; 70940efc6faSStephen Hemminger } 7101da177e4SLinus Torvalds 71146d3ceabSEric Dumazet 71246d3ceabSEric Dumazet /* TCP SMALL QUEUES (TSQ) 71346d3ceabSEric Dumazet * 71446d3ceabSEric Dumazet * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev) 71546d3ceabSEric Dumazet * to reduce RTT and bufferbloat. 71646d3ceabSEric Dumazet * We do this using a special skb destructor (tcp_wfree). 71746d3ceabSEric Dumazet * 71846d3ceabSEric Dumazet * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb 71946d3ceabSEric Dumazet * needs to be reallocated in a driver. 7208e3bff96Sstephen hemminger * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc 72146d3ceabSEric Dumazet * 72246d3ceabSEric Dumazet * Since transmit from skb destructor is forbidden, we use a tasklet 72346d3ceabSEric Dumazet * to process all sockets that eventually need to send more skbs. 72446d3ceabSEric Dumazet * We use one tasklet per cpu, with its own queue of sockets. 72546d3ceabSEric Dumazet */ 72646d3ceabSEric Dumazet struct tsq_tasklet { 72746d3ceabSEric Dumazet struct tasklet_struct tasklet; 72846d3ceabSEric Dumazet struct list_head head; /* queue of tcp sockets */ 72946d3ceabSEric Dumazet }; 73046d3ceabSEric Dumazet static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet); 73146d3ceabSEric Dumazet 7326f458dfbSEric Dumazet static void tcp_tsq_handler(struct sock *sk) 7336f458dfbSEric Dumazet { 7346f458dfbSEric Dumazet if ((1 << sk->sk_state) & 7356f458dfbSEric Dumazet (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING | 7366f458dfbSEric Dumazet TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) 737bf06200eSJohn Ogness tcp_write_xmit(sk, tcp_current_mss(sk), tcp_sk(sk)->nonagle, 738bf06200eSJohn Ogness 0, GFP_ATOMIC); 7396f458dfbSEric Dumazet } 74046d3ceabSEric Dumazet /* 7418e3bff96Sstephen hemminger * One tasklet per cpu tries to send more skbs. 74246d3ceabSEric Dumazet * We run in tasklet context but need to disable irqs when 7438e3bff96Sstephen hemminger * transferring tsq->head because tcp_wfree() might 74446d3ceabSEric Dumazet * interrupt us (non NAPI drivers) 74546d3ceabSEric Dumazet */ 74646d3ceabSEric Dumazet static void tcp_tasklet_func(unsigned long data) 74746d3ceabSEric Dumazet { 74846d3ceabSEric Dumazet struct tsq_tasklet *tsq = (struct tsq_tasklet *)data; 74946d3ceabSEric Dumazet LIST_HEAD(list); 75046d3ceabSEric Dumazet unsigned long flags; 75146d3ceabSEric Dumazet struct list_head *q, *n; 75246d3ceabSEric Dumazet struct tcp_sock *tp; 75346d3ceabSEric Dumazet struct sock *sk; 75446d3ceabSEric Dumazet 75546d3ceabSEric Dumazet local_irq_save(flags); 75646d3ceabSEric Dumazet list_splice_init(&tsq->head, &list); 75746d3ceabSEric Dumazet local_irq_restore(flags); 75846d3ceabSEric Dumazet 75946d3ceabSEric Dumazet list_for_each_safe(q, n, &list) { 76046d3ceabSEric Dumazet tp = list_entry(q, struct tcp_sock, tsq_node); 76146d3ceabSEric Dumazet list_del(&tp->tsq_node); 76246d3ceabSEric Dumazet 76346d3ceabSEric Dumazet sk = (struct sock *)tp; 76446d3ceabSEric Dumazet bh_lock_sock(sk); 76546d3ceabSEric Dumazet 76646d3ceabSEric Dumazet if (!sock_owned_by_user(sk)) { 7676f458dfbSEric Dumazet tcp_tsq_handler(sk); 76846d3ceabSEric Dumazet } else { 76946d3ceabSEric Dumazet /* defer the work to tcp_release_cb() */ 7706f458dfbSEric Dumazet set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags); 77146d3ceabSEric Dumazet } 77246d3ceabSEric Dumazet bh_unlock_sock(sk); 77346d3ceabSEric Dumazet 77446d3ceabSEric Dumazet clear_bit(TSQ_QUEUED, &tp->tsq_flags); 77546d3ceabSEric Dumazet sk_free(sk); 77646d3ceabSEric Dumazet } 77746d3ceabSEric Dumazet } 77846d3ceabSEric Dumazet 7796f458dfbSEric Dumazet #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \ 7806f458dfbSEric Dumazet (1UL << TCP_WRITE_TIMER_DEFERRED) | \ 781563d34d0SEric Dumazet (1UL << TCP_DELACK_TIMER_DEFERRED) | \ 782563d34d0SEric Dumazet (1UL << TCP_MTU_REDUCED_DEFERRED)) 78346d3ceabSEric Dumazet /** 78446d3ceabSEric Dumazet * tcp_release_cb - tcp release_sock() callback 78546d3ceabSEric Dumazet * @sk: socket 78646d3ceabSEric Dumazet * 78746d3ceabSEric Dumazet * called from release_sock() to perform protocol dependent 78846d3ceabSEric Dumazet * actions before socket release. 78946d3ceabSEric Dumazet */ 79046d3ceabSEric Dumazet void tcp_release_cb(struct sock *sk) 79146d3ceabSEric Dumazet { 79246d3ceabSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 7936f458dfbSEric Dumazet unsigned long flags, nflags; 79446d3ceabSEric Dumazet 7956f458dfbSEric Dumazet /* perform an atomic operation only if at least one flag is set */ 7966f458dfbSEric Dumazet do { 7976f458dfbSEric Dumazet flags = tp->tsq_flags; 7986f458dfbSEric Dumazet if (!(flags & TCP_DEFERRED_ALL)) 7996f458dfbSEric Dumazet return; 8006f458dfbSEric Dumazet nflags = flags & ~TCP_DEFERRED_ALL; 8016f458dfbSEric Dumazet } while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags); 8026f458dfbSEric Dumazet 8036f458dfbSEric Dumazet if (flags & (1UL << TCP_TSQ_DEFERRED)) 8046f458dfbSEric Dumazet tcp_tsq_handler(sk); 8056f458dfbSEric Dumazet 806c3f9b018SEric Dumazet /* Here begins the tricky part : 807c3f9b018SEric Dumazet * We are called from release_sock() with : 808c3f9b018SEric Dumazet * 1) BH disabled 809c3f9b018SEric Dumazet * 2) sk_lock.slock spinlock held 810c3f9b018SEric Dumazet * 3) socket owned by us (sk->sk_lock.owned == 1) 811c3f9b018SEric Dumazet * 812c3f9b018SEric Dumazet * But following code is meant to be called from BH handlers, 813c3f9b018SEric Dumazet * so we should keep BH disabled, but early release socket ownership 814c3f9b018SEric Dumazet */ 815c3f9b018SEric Dumazet sock_release_ownership(sk); 816c3f9b018SEric Dumazet 817144d56e9SEric Dumazet if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) { 8186f458dfbSEric Dumazet tcp_write_timer_handler(sk); 819144d56e9SEric Dumazet __sock_put(sk); 820144d56e9SEric Dumazet } 821144d56e9SEric Dumazet if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) { 8226f458dfbSEric Dumazet tcp_delack_timer_handler(sk); 823144d56e9SEric Dumazet __sock_put(sk); 824144d56e9SEric Dumazet } 825144d56e9SEric Dumazet if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) { 8264fab9071SNeal Cardwell inet_csk(sk)->icsk_af_ops->mtu_reduced(sk); 827144d56e9SEric Dumazet __sock_put(sk); 828144d56e9SEric Dumazet } 82946d3ceabSEric Dumazet } 83046d3ceabSEric Dumazet EXPORT_SYMBOL(tcp_release_cb); 83146d3ceabSEric Dumazet 83246d3ceabSEric Dumazet void __init tcp_tasklet_init(void) 83346d3ceabSEric Dumazet { 83446d3ceabSEric Dumazet int i; 83546d3ceabSEric Dumazet 83646d3ceabSEric Dumazet for_each_possible_cpu(i) { 83746d3ceabSEric Dumazet struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i); 83846d3ceabSEric Dumazet 83946d3ceabSEric Dumazet INIT_LIST_HEAD(&tsq->head); 84046d3ceabSEric Dumazet tasklet_init(&tsq->tasklet, 84146d3ceabSEric Dumazet tcp_tasklet_func, 84246d3ceabSEric Dumazet (unsigned long)tsq); 84346d3ceabSEric Dumazet } 84446d3ceabSEric Dumazet } 84546d3ceabSEric Dumazet 84646d3ceabSEric Dumazet /* 84746d3ceabSEric Dumazet * Write buffer destructor automatically called from kfree_skb. 8488e3bff96Sstephen hemminger * We can't xmit new skbs from this context, as we might already 84946d3ceabSEric Dumazet * hold qdisc lock. 85046d3ceabSEric Dumazet */ 851d6a4a104SEric Dumazet void tcp_wfree(struct sk_buff *skb) 85246d3ceabSEric Dumazet { 85346d3ceabSEric Dumazet struct sock *sk = skb->sk; 85446d3ceabSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 8559b462d02SEric Dumazet int wmem; 8569b462d02SEric Dumazet 8579b462d02SEric Dumazet /* Keep one reference on sk_wmem_alloc. 8589b462d02SEric Dumazet * Will be released by sk_free() from here or tcp_tasklet_func() 8599b462d02SEric Dumazet */ 8609b462d02SEric Dumazet wmem = atomic_sub_return(skb->truesize - 1, &sk->sk_wmem_alloc); 8619b462d02SEric Dumazet 8629b462d02SEric Dumazet /* If this softirq is serviced by ksoftirqd, we are likely under stress. 8639b462d02SEric Dumazet * Wait until our queues (qdisc + devices) are drained. 8649b462d02SEric Dumazet * This gives : 8659b462d02SEric Dumazet * - less callbacks to tcp_write_xmit(), reducing stress (batches) 8669b462d02SEric Dumazet * - chance for incoming ACK (processed by another cpu maybe) 8679b462d02SEric Dumazet * to migrate this flow (skb->ooo_okay will be eventually set) 8689b462d02SEric Dumazet */ 8699b462d02SEric Dumazet if (wmem >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current) 8709b462d02SEric Dumazet goto out; 87146d3ceabSEric Dumazet 87246d3ceabSEric Dumazet if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) && 87346d3ceabSEric Dumazet !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) { 87446d3ceabSEric Dumazet unsigned long flags; 87546d3ceabSEric Dumazet struct tsq_tasklet *tsq; 87646d3ceabSEric Dumazet 87746d3ceabSEric Dumazet /* queue this socket to tasklet queue */ 87846d3ceabSEric Dumazet local_irq_save(flags); 879903ceff7SChristoph Lameter tsq = this_cpu_ptr(&tsq_tasklet); 88046d3ceabSEric Dumazet list_add(&tp->tsq_node, &tsq->head); 88146d3ceabSEric Dumazet tasklet_schedule(&tsq->tasklet); 88246d3ceabSEric Dumazet local_irq_restore(flags); 8839b462d02SEric Dumazet return; 88446d3ceabSEric Dumazet } 8859b462d02SEric Dumazet out: 8869b462d02SEric Dumazet sk_free(sk); 88746d3ceabSEric Dumazet } 88846d3ceabSEric Dumazet 8891da177e4SLinus Torvalds /* This routine actually transmits TCP packets queued in by 8901da177e4SLinus Torvalds * tcp_do_sendmsg(). This is used by both the initial 8911da177e4SLinus Torvalds * transmission and possible later retransmissions. 8921da177e4SLinus Torvalds * All SKB's seen here are completely headerless. It is our 8931da177e4SLinus Torvalds * job to build the TCP header, and pass the packet down to 8941da177e4SLinus Torvalds * IP so it can do the same plus pass the packet off to the 8951da177e4SLinus Torvalds * device. 8961da177e4SLinus Torvalds * 8971da177e4SLinus Torvalds * We are working here with either a clone of the original 8981da177e4SLinus Torvalds * SKB, or a fresh unique copy made by the retransmit engine. 8991da177e4SLinus Torvalds */ 900056834d9SIlpo Järvinen static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, 901056834d9SIlpo Järvinen gfp_t gfp_mask) 9021da177e4SLinus Torvalds { 9036687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 904dfb4b9dcSDavid S. Miller struct inet_sock *inet; 905dfb4b9dcSDavid S. Miller struct tcp_sock *tp; 906dfb4b9dcSDavid S. Miller struct tcp_skb_cb *tcb; 90733ad798cSAdam Langley struct tcp_out_options opts; 90895c96174SEric Dumazet unsigned int tcp_options_size, tcp_header_size; 909cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 9101da177e4SLinus Torvalds struct tcphdr *th; 9111da177e4SLinus Torvalds int err; 9121da177e4SLinus Torvalds 913dfb4b9dcSDavid S. Miller BUG_ON(!skb || !tcp_skb_pcount(skb)); 914dfb4b9dcSDavid S. Miller 915ccdbb6e9SEric Dumazet if (clone_it) { 916740b0f18SEric Dumazet skb_mstamp_get(&skb->skb_mstamp); 917dfb4b9dcSDavid S. Miller 918dfb4b9dcSDavid S. Miller if (unlikely(skb_cloned(skb))) 919dfb4b9dcSDavid S. Miller skb = pskb_copy(skb, gfp_mask); 920dfb4b9dcSDavid S. Miller else 921dfb4b9dcSDavid S. Miller skb = skb_clone(skb, gfp_mask); 922dfb4b9dcSDavid S. Miller if (unlikely(!skb)) 923dfb4b9dcSDavid S. Miller return -ENOBUFS; 924dfb4b9dcSDavid S. Miller } 925dfb4b9dcSDavid S. Miller 926dfb4b9dcSDavid S. Miller inet = inet_sk(sk); 927dfb4b9dcSDavid S. Miller tp = tcp_sk(sk); 928dfb4b9dcSDavid S. Miller tcb = TCP_SKB_CB(skb); 92933ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 9301da177e4SLinus Torvalds 9314de075e0SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) 93233ad798cSAdam Langley tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); 93333ad798cSAdam Langley else 93433ad798cSAdam Langley tcp_options_size = tcp_established_options(sk, skb, &opts, 93533ad798cSAdam Langley &md5); 93633ad798cSAdam Langley tcp_header_size = tcp_options_size + sizeof(struct tcphdr); 9371da177e4SLinus Torvalds 938547669d4SEric Dumazet /* if no packet is in qdisc/device queue, then allow XPS to select 939b2532eb9SEric Dumazet * another queue. We can be called from tcp_tsq_handler() 940b2532eb9SEric Dumazet * which holds one reference to sk_wmem_alloc. 941b2532eb9SEric Dumazet * 942b2532eb9SEric Dumazet * TODO: Ideally, in-flight pure ACK packets should not matter here. 943b2532eb9SEric Dumazet * One way to get this would be to set skb->truesize = 2 on them. 944547669d4SEric Dumazet */ 945b2532eb9SEric Dumazet skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1); 9461da177e4SLinus Torvalds 947aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 948aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 94946d3ceabSEric Dumazet 95046d3ceabSEric Dumazet skb_orphan(skb); 95146d3ceabSEric Dumazet skb->sk = sk; 95298781965SEric Dumazet skb->destructor = skb_is_tcp_pure_ack(skb) ? sock_wfree : tcp_wfree; 953b73c3d0eSTom Herbert skb_set_hash_from_sk(skb, sk); 95446d3ceabSEric Dumazet atomic_add(skb->truesize, &sk->sk_wmem_alloc); 9551da177e4SLinus Torvalds 9561da177e4SLinus Torvalds /* Build TCP header and checksum it. */ 957aa8223c7SArnaldo Carvalho de Melo th = tcp_hdr(skb); 958c720c7e8SEric Dumazet th->source = inet->inet_sport; 959c720c7e8SEric Dumazet th->dest = inet->inet_dport; 9601da177e4SLinus Torvalds th->seq = htonl(tcb->seq); 9611da177e4SLinus Torvalds th->ack_seq = htonl(tp->rcv_nxt); 962df7a3b07SAl Viro *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | 9634de075e0SEric Dumazet tcb->tcp_flags); 964dfb4b9dcSDavid S. Miller 9654de075e0SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) { 9661da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments 9671da177e4SLinus Torvalds * is never scaled. 9681da177e4SLinus Torvalds */ 969600ff0c2SIlpo Järvinen th->window = htons(min(tp->rcv_wnd, 65535U)); 9701da177e4SLinus Torvalds } else { 9711da177e4SLinus Torvalds th->window = htons(tcp_select_window(sk)); 9721da177e4SLinus Torvalds } 9731da177e4SLinus Torvalds th->check = 0; 9741da177e4SLinus Torvalds th->urg_ptr = 0; 9751da177e4SLinus Torvalds 97633f5f57eSIlpo Järvinen /* The urg_mode check is necessary during a below snd_una win probe */ 9777691367dSHerbert Xu if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { 9787691367dSHerbert Xu if (before(tp->snd_up, tcb->seq + 0x10000)) { 9791da177e4SLinus Torvalds th->urg_ptr = htons(tp->snd_up - tcb->seq); 9801da177e4SLinus Torvalds th->urg = 1; 9817691367dSHerbert Xu } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { 9820eae88f3SEric Dumazet th->urg_ptr = htons(0xFFFF); 9837691367dSHerbert Xu th->urg = 1; 9847691367dSHerbert Xu } 9851da177e4SLinus Torvalds } 9861da177e4SLinus Torvalds 987bd0388aeSWilliam Allen Simpson tcp_options_write((__be32 *)(th + 1), tp, &opts); 98851466a75SEric Dumazet skb_shinfo(skb)->gso_type = sk->sk_gso_type; 9894de075e0SEric Dumazet if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0)) 990735d3831SFlorian Westphal tcp_ecn_send(sk, skb, tcp_header_size); 991dfb4b9dcSDavid S. Miller 992cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 993cfb6eeb4SYOSHIFUJI Hideaki /* Calculate the MD5 hash, as we have all we need now */ 994cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 995a465419bSEric Dumazet sk_nocaps_add(sk, NETIF_F_GSO_MASK); 996bd0388aeSWilliam Allen Simpson tp->af_specific->calc_md5_hash(opts.hash_location, 99739f8e58eSEric Dumazet md5, sk, skb); 998cfb6eeb4SYOSHIFUJI Hideaki } 999cfb6eeb4SYOSHIFUJI Hideaki #endif 1000cfb6eeb4SYOSHIFUJI Hideaki 1001bb296246SHerbert Xu icsk->icsk_af_ops->send_check(sk, skb); 10021da177e4SLinus Torvalds 10034de075e0SEric Dumazet if (likely(tcb->tcp_flags & TCPHDR_ACK)) 1004fc6415bcSDavid S. Miller tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); 10051da177e4SLinus Torvalds 1006a44d6eacSMartin KaFai Lau if (skb->len != tcp_header_size) { 1007cf533ea5SEric Dumazet tcp_event_data_sent(tp, sk); 1008a44d6eacSMartin KaFai Lau tp->data_segs_out += tcp_skb_pcount(skb); 1009a44d6eacSMartin KaFai Lau } 10101da177e4SLinus Torvalds 1011bd37a088SWei Yongjun if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) 1012aa2ea058STom Herbert TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, 1013aa2ea058STom Herbert tcp_skb_pcount(skb)); 10141da177e4SLinus Torvalds 10152efd055cSMarcelo Ricardo Leitner tp->segs_out += tcp_skb_pcount(skb); 1016f69ad292SEric Dumazet /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */ 1017cd7d8498SEric Dumazet skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb); 1018f69ad292SEric Dumazet skb_shinfo(skb)->gso_size = tcp_skb_mss(skb); 1019cd7d8498SEric Dumazet 10207faee5c0SEric Dumazet /* Our usage of tstamp should remain private */ 10217faee5c0SEric Dumazet skb->tstamp.tv64 = 0; 1022971f10ecSEric Dumazet 1023971f10ecSEric Dumazet /* Cleanup our debris for IP stacks */ 1024971f10ecSEric Dumazet memset(skb->cb, 0, max(sizeof(struct inet_skb_parm), 1025971f10ecSEric Dumazet sizeof(struct inet6_skb_parm))); 1026971f10ecSEric Dumazet 1027b0270e91SEric Dumazet err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl); 10287faee5c0SEric Dumazet 102983de47cdSHua Zhong if (likely(err <= 0)) 10301da177e4SLinus Torvalds return err; 10311da177e4SLinus Torvalds 10325ee2c941SChristoph Paasch tcp_enter_cwr(sk); 10331da177e4SLinus Torvalds 1034b9df3cb8SGerrit Renker return net_xmit_eval(err); 10351da177e4SLinus Torvalds } 10361da177e4SLinus Torvalds 103767edfef7SAndi Kleen /* This routine just queues the buffer for sending. 10381da177e4SLinus Torvalds * 10391da177e4SLinus Torvalds * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, 10401da177e4SLinus Torvalds * otherwise socket can stall. 10411da177e4SLinus Torvalds */ 10421da177e4SLinus Torvalds static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) 10431da177e4SLinus Torvalds { 10441da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 10451da177e4SLinus Torvalds 10461da177e4SLinus Torvalds /* Advance write_seq and place onto the write_queue. */ 10471da177e4SLinus Torvalds tp->write_seq = TCP_SKB_CB(skb)->end_seq; 1048f4a775d1SEric Dumazet __skb_header_release(skb); 1049fe067e8aSDavid S. Miller tcp_add_write_queue_tail(sk, skb); 10503ab224beSHideo Aoki sk->sk_wmem_queued += skb->truesize; 10513ab224beSHideo Aoki sk_mem_charge(sk, skb->truesize); 10521da177e4SLinus Torvalds } 10531da177e4SLinus Torvalds 105467edfef7SAndi Kleen /* Initialize TSO segments for a packet. */ 10555bbb432cSEric Dumazet static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now) 1056f6302d1dSDavid S. Miller { 10578f26fb1cSEric Dumazet if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) { 1058f6302d1dSDavid S. Miller /* Avoid the costly divide in the normal 1059f6302d1dSDavid S. Miller * non-TSO case. 1060f6302d1dSDavid S. Miller */ 1061cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, 1); 1062f69ad292SEric Dumazet TCP_SKB_CB(skb)->tcp_gso_size = 0; 1063f6302d1dSDavid S. Miller } else { 1064cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now)); 1065f69ad292SEric Dumazet TCP_SKB_CB(skb)->tcp_gso_size = mss_now; 10661da177e4SLinus Torvalds } 10671da177e4SLinus Torvalds } 10681da177e4SLinus Torvalds 106991fed7a1SIlpo Järvinen /* When a modification to fackets out becomes necessary, we need to check 107068f8353bSIlpo Järvinen * skb is counted to fackets_out or not. 107191fed7a1SIlpo Järvinen */ 1072cf533ea5SEric Dumazet static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb, 107391fed7a1SIlpo Järvinen int decr) 107491fed7a1SIlpo Järvinen { 1075a47e5a98SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1076a47e5a98SIlpo Järvinen 1077dc86967bSIlpo Järvinen if (!tp->sacked_out || tcp_is_reno(tp)) 107891fed7a1SIlpo Järvinen return; 107991fed7a1SIlpo Järvinen 10806859d494SIlpo Järvinen if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq)) 108191fed7a1SIlpo Järvinen tp->fackets_out -= decr; 108291fed7a1SIlpo Järvinen } 108391fed7a1SIlpo Järvinen 1084797108d1SIlpo Järvinen /* Pcount in the middle of the write queue got changed, we need to do various 1085797108d1SIlpo Järvinen * tweaks to fix counters 1086797108d1SIlpo Järvinen */ 1087cf533ea5SEric Dumazet static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) 1088797108d1SIlpo Järvinen { 1089797108d1SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1090797108d1SIlpo Järvinen 1091797108d1SIlpo Järvinen tp->packets_out -= decr; 1092797108d1SIlpo Järvinen 1093797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 1094797108d1SIlpo Järvinen tp->sacked_out -= decr; 1095797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 1096797108d1SIlpo Järvinen tp->retrans_out -= decr; 1097797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 1098797108d1SIlpo Järvinen tp->lost_out -= decr; 1099797108d1SIlpo Järvinen 1100797108d1SIlpo Järvinen /* Reno case is special. Sigh... */ 1101797108d1SIlpo Järvinen if (tcp_is_reno(tp) && decr > 0) 1102797108d1SIlpo Järvinen tp->sacked_out -= min_t(u32, tp->sacked_out, decr); 1103797108d1SIlpo Järvinen 1104797108d1SIlpo Järvinen tcp_adjust_fackets_out(sk, skb, decr); 1105797108d1SIlpo Järvinen 1106797108d1SIlpo Järvinen if (tp->lost_skb_hint && 1107797108d1SIlpo Järvinen before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && 110852cf3cc8SIlpo Järvinen (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))) 1109797108d1SIlpo Järvinen tp->lost_cnt_hint -= decr; 1110797108d1SIlpo Järvinen 1111797108d1SIlpo Järvinen tcp_verify_left_out(tp); 1112797108d1SIlpo Järvinen } 1113797108d1SIlpo Järvinen 1114490cc7d0SWillem de Bruijn static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2) 1115490cc7d0SWillem de Bruijn { 1116490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo = skb_shinfo(skb); 1117490cc7d0SWillem de Bruijn 1118490cc7d0SWillem de Bruijn if (unlikely(shinfo->tx_flags & SKBTX_ANY_TSTAMP) && 1119490cc7d0SWillem de Bruijn !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) { 1120490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo2 = skb_shinfo(skb2); 1121490cc7d0SWillem de Bruijn u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP; 1122490cc7d0SWillem de Bruijn 1123490cc7d0SWillem de Bruijn shinfo->tx_flags &= ~tsflags; 1124490cc7d0SWillem de Bruijn shinfo2->tx_flags |= tsflags; 1125490cc7d0SWillem de Bruijn swap(shinfo->tskey, shinfo2->tskey); 1126490cc7d0SWillem de Bruijn } 1127490cc7d0SWillem de Bruijn } 1128490cc7d0SWillem de Bruijn 11291da177e4SLinus Torvalds /* Function to create two new TCP segments. Shrinks the given segment 11301da177e4SLinus Torvalds * to the specified size and appends a new segment with the rest of the 11311da177e4SLinus Torvalds * packet to the list. This won't be called frequently, I hope. 11321da177e4SLinus Torvalds * Remember, these are still headerless SKBs at this point. 11331da177e4SLinus Torvalds */ 1134056834d9SIlpo Järvinen int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, 11356cc55e09SOctavian Purdila unsigned int mss_now, gfp_t gfp) 11361da177e4SLinus Torvalds { 11371da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 11381da177e4SLinus Torvalds struct sk_buff *buff; 11396475be16SDavid S. Miller int nsize, old_factor; 1140b60b49eaSHerbert Xu int nlen; 11419ce01461SIlpo Järvinen u8 flags; 11421da177e4SLinus Torvalds 11432fceec13SIlpo Järvinen if (WARN_ON(len > skb->len)) 11442fceec13SIlpo Järvinen return -EINVAL; 11456a438bbeSStephen Hemminger 11461da177e4SLinus Torvalds nsize = skb_headlen(skb) - len; 11471da177e4SLinus Torvalds if (nsize < 0) 11481da177e4SLinus Torvalds nsize = 0; 11491da177e4SLinus Torvalds 11506cc55e09SOctavian Purdila if (skb_unclone(skb, gfp)) 11511da177e4SLinus Torvalds return -ENOMEM; 11521da177e4SLinus Torvalds 11531da177e4SLinus Torvalds /* Get a new skb... force flag on. */ 1154eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, nsize, gfp, true); 115551456b29SIan Morris if (!buff) 11561da177e4SLinus Torvalds return -ENOMEM; /* We'll just try again later. */ 1157ef5cb973SHerbert Xu 11583ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 11593ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1160b60b49eaSHerbert Xu nlen = skb->len - len - nsize; 1161b60b49eaSHerbert Xu buff->truesize += nlen; 1162b60b49eaSHerbert Xu skb->truesize -= nlen; 11631da177e4SLinus Torvalds 11641da177e4SLinus Torvalds /* Correct the sequence numbers. */ 11651da177e4SLinus Torvalds TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 11661da177e4SLinus Torvalds TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 11671da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 11681da177e4SLinus Torvalds 11691da177e4SLinus Torvalds /* PSH and FIN should only be set in the second packet. */ 11704de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 11714de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 11724de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1173e14c3cafSHerbert Xu TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; 11741da177e4SLinus Torvalds 117584fa7933SPatrick McHardy if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { 11761da177e4SLinus Torvalds /* Copy and checksum data tail into the new buffer. */ 1177056834d9SIlpo Järvinen buff->csum = csum_partial_copy_nocheck(skb->data + len, 1178056834d9SIlpo Järvinen skb_put(buff, nsize), 11791da177e4SLinus Torvalds nsize, 0); 11801da177e4SLinus Torvalds 11811da177e4SLinus Torvalds skb_trim(skb, len); 11821da177e4SLinus Torvalds 11831da177e4SLinus Torvalds skb->csum = csum_block_sub(skb->csum, buff->csum, len); 11841da177e4SLinus Torvalds } else { 118584fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 11861da177e4SLinus Torvalds skb_split(skb, buff, len); 11871da177e4SLinus Torvalds } 11881da177e4SLinus Torvalds 11891da177e4SLinus Torvalds buff->ip_summed = skb->ip_summed; 11901da177e4SLinus Torvalds 1191a61bbcf2SPatrick McHardy buff->tstamp = skb->tstamp; 1192490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 11931da177e4SLinus Torvalds 11946475be16SDavid S. Miller old_factor = tcp_skb_pcount(skb); 11956475be16SDavid S. Miller 11961da177e4SLinus Torvalds /* Fix up tso_factor for both original and new SKB. */ 11975bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 11985bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 11991da177e4SLinus Torvalds 12006475be16SDavid S. Miller /* If this packet has been sent out already, we must 12016475be16SDavid S. Miller * adjust the various packet counters. 12026475be16SDavid S. Miller */ 1203cf0b450cSHerbert Xu if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 12046475be16SDavid S. Miller int diff = old_factor - tcp_skb_pcount(skb) - 12056475be16SDavid S. Miller tcp_skb_pcount(buff); 12061da177e4SLinus Torvalds 1207797108d1SIlpo Järvinen if (diff) 1208797108d1SIlpo Järvinen tcp_adjust_pcount(sk, skb, diff); 12091da177e4SLinus Torvalds } 12101da177e4SLinus Torvalds 12111da177e4SLinus Torvalds /* Link BUFF into the send queue. */ 1212f4a775d1SEric Dumazet __skb_header_release(buff); 1213fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 12141da177e4SLinus Torvalds 12151da177e4SLinus Torvalds return 0; 12161da177e4SLinus Torvalds } 12171da177e4SLinus Torvalds 12181da177e4SLinus Torvalds /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c 12191da177e4SLinus Torvalds * eventually). The difference is that pulled data not copied, but 12201da177e4SLinus Torvalds * immediately discarded. 12211da177e4SLinus Torvalds */ 1222f2911969SHerbert Xu ~{PmVHI~} static void __pskb_trim_head(struct sk_buff *skb, int len) 12231da177e4SLinus Torvalds { 12247b7fc97aSEric Dumazet struct skb_shared_info *shinfo; 12251da177e4SLinus Torvalds int i, k, eat; 12261da177e4SLinus Torvalds 12274fa48bf3SEric Dumazet eat = min_t(int, len, skb_headlen(skb)); 12284fa48bf3SEric Dumazet if (eat) { 12294fa48bf3SEric Dumazet __skb_pull(skb, eat); 12304fa48bf3SEric Dumazet len -= eat; 12314fa48bf3SEric Dumazet if (!len) 12324fa48bf3SEric Dumazet return; 12334fa48bf3SEric Dumazet } 12341da177e4SLinus Torvalds eat = len; 12351da177e4SLinus Torvalds k = 0; 12367b7fc97aSEric Dumazet shinfo = skb_shinfo(skb); 12377b7fc97aSEric Dumazet for (i = 0; i < shinfo->nr_frags; i++) { 12387b7fc97aSEric Dumazet int size = skb_frag_size(&shinfo->frags[i]); 12399e903e08SEric Dumazet 12409e903e08SEric Dumazet if (size <= eat) { 1241aff65da0SIan Campbell skb_frag_unref(skb, i); 12429e903e08SEric Dumazet eat -= size; 12431da177e4SLinus Torvalds } else { 12447b7fc97aSEric Dumazet shinfo->frags[k] = shinfo->frags[i]; 12451da177e4SLinus Torvalds if (eat) { 12467b7fc97aSEric Dumazet shinfo->frags[k].page_offset += eat; 12477b7fc97aSEric Dumazet skb_frag_size_sub(&shinfo->frags[k], eat); 12481da177e4SLinus Torvalds eat = 0; 12491da177e4SLinus Torvalds } 12501da177e4SLinus Torvalds k++; 12511da177e4SLinus Torvalds } 12521da177e4SLinus Torvalds } 12537b7fc97aSEric Dumazet shinfo->nr_frags = k; 12541da177e4SLinus Torvalds 125527a884dcSArnaldo Carvalho de Melo skb_reset_tail_pointer(skb); 12561da177e4SLinus Torvalds skb->data_len -= len; 12571da177e4SLinus Torvalds skb->len = skb->data_len; 12581da177e4SLinus Torvalds } 12591da177e4SLinus Torvalds 126067edfef7SAndi Kleen /* Remove acked data from a packet in the transmit queue. */ 12611da177e4SLinus Torvalds int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 12621da177e4SLinus Torvalds { 126314bbd6a5SPravin B Shelar if (skb_unclone(skb, GFP_ATOMIC)) 12641da177e4SLinus Torvalds return -ENOMEM; 12651da177e4SLinus Torvalds 12664fa48bf3SEric Dumazet __pskb_trim_head(skb, len); 12671da177e4SLinus Torvalds 12681da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq += len; 126984fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 12701da177e4SLinus Torvalds 12711da177e4SLinus Torvalds skb->truesize -= len; 12721da177e4SLinus Torvalds sk->sk_wmem_queued -= len; 12733ab224beSHideo Aoki sk_mem_uncharge(sk, len); 12741da177e4SLinus Torvalds sock_set_flag(sk, SOCK_QUEUE_SHRUNK); 12751da177e4SLinus Torvalds 12765b35e1e6SNeal Cardwell /* Any change of skb->len requires recalculation of tso factor. */ 12771da177e4SLinus Torvalds if (tcp_skb_pcount(skb) > 1) 12785bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb)); 12791da177e4SLinus Torvalds 12801da177e4SLinus Torvalds return 0; 12811da177e4SLinus Torvalds } 12821da177e4SLinus Torvalds 12831b63edd6SYuchung Cheng /* Calculate MSS not accounting any TCP options. */ 12841b63edd6SYuchung Cheng static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu) 12855d424d5aSJohn Heffner { 1286cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1287cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 12885d424d5aSJohn Heffner int mss_now; 12895d424d5aSJohn Heffner 12905d424d5aSJohn Heffner /* Calculate base mss without TCP options: 12915d424d5aSJohn Heffner It is MMS_S - sizeof(tcphdr) of rfc1122 12925d424d5aSJohn Heffner */ 12935d424d5aSJohn Heffner mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 12945d424d5aSJohn Heffner 129567469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 129667469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 129767469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 129867469601SEric Dumazet 129967469601SEric Dumazet if (dst && dst_allfrag(dst)) 130067469601SEric Dumazet mss_now -= icsk->icsk_af_ops->net_frag_header_len; 130167469601SEric Dumazet } 130267469601SEric Dumazet 13035d424d5aSJohn Heffner /* Clamp it (mss_clamp does not include tcp options) */ 13045d424d5aSJohn Heffner if (mss_now > tp->rx_opt.mss_clamp) 13055d424d5aSJohn Heffner mss_now = tp->rx_opt.mss_clamp; 13065d424d5aSJohn Heffner 13075d424d5aSJohn Heffner /* Now subtract optional transport overhead */ 13085d424d5aSJohn Heffner mss_now -= icsk->icsk_ext_hdr_len; 13095d424d5aSJohn Heffner 13105d424d5aSJohn Heffner /* Then reserve room for full set of TCP options and 8 bytes of data */ 13115d424d5aSJohn Heffner if (mss_now < 48) 13125d424d5aSJohn Heffner mss_now = 48; 13135d424d5aSJohn Heffner return mss_now; 13145d424d5aSJohn Heffner } 13155d424d5aSJohn Heffner 13161b63edd6SYuchung Cheng /* Calculate MSS. Not accounting for SACKs here. */ 13171b63edd6SYuchung Cheng int tcp_mtu_to_mss(struct sock *sk, int pmtu) 13181b63edd6SYuchung Cheng { 13191b63edd6SYuchung Cheng /* Subtract TCP options size, not including SACKs */ 13201b63edd6SYuchung Cheng return __tcp_mtu_to_mss(sk, pmtu) - 13211b63edd6SYuchung Cheng (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr)); 13221b63edd6SYuchung Cheng } 13231b63edd6SYuchung Cheng 13245d424d5aSJohn Heffner /* Inverse of above */ 132567469601SEric Dumazet int tcp_mss_to_mtu(struct sock *sk, int mss) 13265d424d5aSJohn Heffner { 1327cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1328cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 13295d424d5aSJohn Heffner int mtu; 13305d424d5aSJohn Heffner 13315d424d5aSJohn Heffner mtu = mss + 13325d424d5aSJohn Heffner tp->tcp_header_len + 13335d424d5aSJohn Heffner icsk->icsk_ext_hdr_len + 13345d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 13355d424d5aSJohn Heffner 133667469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 133767469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 133867469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 133967469601SEric Dumazet 134067469601SEric Dumazet if (dst && dst_allfrag(dst)) 134167469601SEric Dumazet mtu += icsk->icsk_af_ops->net_frag_header_len; 134267469601SEric Dumazet } 13435d424d5aSJohn Heffner return mtu; 13445d424d5aSJohn Heffner } 13455d424d5aSJohn Heffner 134667edfef7SAndi Kleen /* MTU probing init per socket */ 13475d424d5aSJohn Heffner void tcp_mtup_init(struct sock *sk) 13485d424d5aSJohn Heffner { 13495d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 13505d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 1351b0f9ca53SFan Du struct net *net = sock_net(sk); 13525d424d5aSJohn Heffner 1353b0f9ca53SFan Du icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1; 13545d424d5aSJohn Heffner icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 13555d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 1356b0f9ca53SFan Du icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss); 13575d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 135805cbc0dbSFan Du if (icsk->icsk_mtup.enabled) 135905cbc0dbSFan Du icsk->icsk_mtup.probe_timestamp = tcp_time_stamp; 13605d424d5aSJohn Heffner } 13614bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_mtup_init); 13625d424d5aSJohn Heffner 13631da177e4SLinus Torvalds /* This function synchronize snd mss to current pmtu/exthdr set. 13641da177e4SLinus Torvalds 13651da177e4SLinus Torvalds tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 13661da177e4SLinus Torvalds for TCP options, but includes only bare TCP header. 13671da177e4SLinus Torvalds 13681da177e4SLinus Torvalds tp->rx_opt.mss_clamp is mss negotiated at connection setup. 1369caa20d9aSStephen Hemminger It is minimum of user_mss and mss received with SYN. 13701da177e4SLinus Torvalds It also does not include TCP options. 13711da177e4SLinus Torvalds 1372d83d8461SArnaldo Carvalho de Melo inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. 13731da177e4SLinus Torvalds 13741da177e4SLinus Torvalds tp->mss_cache is current effective sending mss, including 13751da177e4SLinus Torvalds all tcp options except for SACKs. It is evaluated, 13761da177e4SLinus Torvalds taking into account current pmtu, but never exceeds 13771da177e4SLinus Torvalds tp->rx_opt.mss_clamp. 13781da177e4SLinus Torvalds 13791da177e4SLinus Torvalds NOTE1. rfc1122 clearly states that advertised MSS 13801da177e4SLinus Torvalds DOES NOT include either tcp or ip options. 13811da177e4SLinus Torvalds 1382d83d8461SArnaldo Carvalho de Melo NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache 1383d83d8461SArnaldo Carvalho de Melo are READ ONLY outside this function. --ANK (980731) 13841da177e4SLinus Torvalds */ 13851da177e4SLinus Torvalds unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 13861da177e4SLinus Torvalds { 13871da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1388d83d8461SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 13895d424d5aSJohn Heffner int mss_now; 13901da177e4SLinus Torvalds 13915d424d5aSJohn Heffner if (icsk->icsk_mtup.search_high > pmtu) 13925d424d5aSJohn Heffner icsk->icsk_mtup.search_high = pmtu; 13931da177e4SLinus Torvalds 13945d424d5aSJohn Heffner mss_now = tcp_mtu_to_mss(sk, pmtu); 1395409d22b4SIlpo Järvinen mss_now = tcp_bound_to_half_wnd(tp, mss_now); 13961da177e4SLinus Torvalds 13971da177e4SLinus Torvalds /* And store cached results */ 1398d83d8461SArnaldo Carvalho de Melo icsk->icsk_pmtu_cookie = pmtu; 13995d424d5aSJohn Heffner if (icsk->icsk_mtup.enabled) 14005d424d5aSJohn Heffner mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); 1401c1b4a7e6SDavid S. Miller tp->mss_cache = mss_now; 14021da177e4SLinus Torvalds 14031da177e4SLinus Torvalds return mss_now; 14041da177e4SLinus Torvalds } 14054bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_sync_mss); 14061da177e4SLinus Torvalds 14071da177e4SLinus Torvalds /* Compute the current effective MSS, taking SACKs and IP options, 14081da177e4SLinus Torvalds * and even PMTU discovery events into account. 14091da177e4SLinus Torvalds */ 14100c54b85fSIlpo Järvinen unsigned int tcp_current_mss(struct sock *sk) 14111da177e4SLinus Torvalds { 1412cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1413cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1414c1b4a7e6SDavid S. Miller u32 mss_now; 141595c96174SEric Dumazet unsigned int header_len; 141633ad798cSAdam Langley struct tcp_out_options opts; 141733ad798cSAdam Langley struct tcp_md5sig_key *md5; 14181da177e4SLinus Torvalds 1419c1b4a7e6SDavid S. Miller mss_now = tp->mss_cache; 1420c1b4a7e6SDavid S. Miller 14211da177e4SLinus Torvalds if (dst) { 14221da177e4SLinus Torvalds u32 mtu = dst_mtu(dst); 1423d83d8461SArnaldo Carvalho de Melo if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 14241da177e4SLinus Torvalds mss_now = tcp_sync_mss(sk, mtu); 14251da177e4SLinus Torvalds } 14261da177e4SLinus Torvalds 142733ad798cSAdam Langley header_len = tcp_established_options(sk, NULL, &opts, &md5) + 142833ad798cSAdam Langley sizeof(struct tcphdr); 142933ad798cSAdam Langley /* The mss_cache is sized based on tp->tcp_header_len, which assumes 143033ad798cSAdam Langley * some common options. If this is an odd packet (because we have SACK 143133ad798cSAdam Langley * blocks etc) then our calculated header_len will be different, and 143233ad798cSAdam Langley * we have to adjust mss_now correspondingly */ 143333ad798cSAdam Langley if (header_len != tp->tcp_header_len) { 143433ad798cSAdam Langley int delta = (int) header_len - tp->tcp_header_len; 143533ad798cSAdam Langley mss_now -= delta; 143633ad798cSAdam Langley } 1437cfb6eeb4SYOSHIFUJI Hideaki 14381da177e4SLinus Torvalds return mss_now; 14391da177e4SLinus Torvalds } 14401da177e4SLinus Torvalds 144186fd14adSWeiping Pan /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. 144286fd14adSWeiping Pan * As additional protections, we do not touch cwnd in retransmission phases, 144386fd14adSWeiping Pan * and if application hit its sndbuf limit recently. 144486fd14adSWeiping Pan */ 144586fd14adSWeiping Pan static void tcp_cwnd_application_limited(struct sock *sk) 1446a762a980SDavid S. Miller { 14479e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1448a762a980SDavid S. Miller 144986fd14adSWeiping Pan if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && 145086fd14adSWeiping Pan sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 145186fd14adSWeiping Pan /* Limited by application or receiver window. */ 145286fd14adSWeiping Pan u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); 145386fd14adSWeiping Pan u32 win_used = max(tp->snd_cwnd_used, init_win); 145486fd14adSWeiping Pan if (win_used < tp->snd_cwnd) { 145586fd14adSWeiping Pan tp->snd_ssthresh = tcp_current_ssthresh(sk); 145686fd14adSWeiping Pan tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1; 145786fd14adSWeiping Pan } 145886fd14adSWeiping Pan tp->snd_cwnd_used = 0; 145986fd14adSWeiping Pan } 146086fd14adSWeiping Pan tp->snd_cwnd_stamp = tcp_time_stamp; 146186fd14adSWeiping Pan } 146286fd14adSWeiping Pan 1463ca8a2263SNeal Cardwell static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited) 1464a762a980SDavid S. Miller { 1465a762a980SDavid S. Miller struct tcp_sock *tp = tcp_sk(sk); 1466a762a980SDavid S. Miller 1467ca8a2263SNeal Cardwell /* Track the maximum number of outstanding packets in each 1468ca8a2263SNeal Cardwell * window, and remember whether we were cwnd-limited then. 1469ca8a2263SNeal Cardwell */ 1470ca8a2263SNeal Cardwell if (!before(tp->snd_una, tp->max_packets_seq) || 1471ca8a2263SNeal Cardwell tp->packets_out > tp->max_packets_out) { 1472ca8a2263SNeal Cardwell tp->max_packets_out = tp->packets_out; 1473ca8a2263SNeal Cardwell tp->max_packets_seq = tp->snd_nxt; 1474ca8a2263SNeal Cardwell tp->is_cwnd_limited = is_cwnd_limited; 1475ca8a2263SNeal Cardwell } 1476e114a710SEric Dumazet 147724901551SEric Dumazet if (tcp_is_cwnd_limited(sk)) { 1478a762a980SDavid S. Miller /* Network is feed fully. */ 1479a762a980SDavid S. Miller tp->snd_cwnd_used = 0; 1480a762a980SDavid S. Miller tp->snd_cwnd_stamp = tcp_time_stamp; 1481a762a980SDavid S. Miller } else { 1482a762a980SDavid S. Miller /* Network starves. */ 1483a762a980SDavid S. Miller if (tp->packets_out > tp->snd_cwnd_used) 1484a762a980SDavid S. Miller tp->snd_cwnd_used = tp->packets_out; 1485a762a980SDavid S. Miller 148615d33c07SDavid S. Miller if (sysctl_tcp_slow_start_after_idle && 148715d33c07SDavid S. Miller (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto) 1488a762a980SDavid S. Miller tcp_cwnd_application_limited(sk); 1489a762a980SDavid S. Miller } 1490a762a980SDavid S. Miller } 1491a762a980SDavid S. Miller 1492d4589926SEric Dumazet /* Minshall's variant of the Nagle send check. */ 1493d4589926SEric Dumazet static bool tcp_minshall_check(const struct tcp_sock *tp) 1494d4589926SEric Dumazet { 1495d4589926SEric Dumazet return after(tp->snd_sml, tp->snd_una) && 1496d4589926SEric Dumazet !after(tp->snd_sml, tp->snd_nxt); 1497d4589926SEric Dumazet } 1498d4589926SEric Dumazet 1499d4589926SEric Dumazet /* Update snd_sml if this skb is under mss 1500d4589926SEric Dumazet * Note that a TSO packet might end with a sub-mss segment 1501d4589926SEric Dumazet * The test is really : 1502d4589926SEric Dumazet * if ((skb->len % mss) != 0) 1503d4589926SEric Dumazet * tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1504d4589926SEric Dumazet * But we can avoid doing the divide again given we already have 1505d4589926SEric Dumazet * skb_pcount = skb->len / mss_now 15060e3a4803SIlpo Järvinen */ 1507d4589926SEric Dumazet static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now, 1508d4589926SEric Dumazet const struct sk_buff *skb) 1509d4589926SEric Dumazet { 1510d4589926SEric Dumazet if (skb->len < tcp_skb_pcount(skb) * mss_now) 1511d4589926SEric Dumazet tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1512d4589926SEric Dumazet } 1513d4589926SEric Dumazet 1514d4589926SEric Dumazet /* Return false, if packet can be sent now without violation Nagle's rules: 1515d4589926SEric Dumazet * 1. It is full sized. (provided by caller in %partial bool) 1516d4589926SEric Dumazet * 2. Or it contains FIN. (already checked by caller) 1517d4589926SEric Dumazet * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. 1518d4589926SEric Dumazet * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 1519d4589926SEric Dumazet * With Minshall's modification: all sent small packets are ACKed. 1520d4589926SEric Dumazet */ 1521d4589926SEric Dumazet static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp, 1522cc93fc51SPeter Pan(潘卫平) int nonagle) 1523d4589926SEric Dumazet { 1524d4589926SEric Dumazet return partial && 1525d4589926SEric Dumazet ((nonagle & TCP_NAGLE_CORK) || 1526d4589926SEric Dumazet (!nonagle && tp->packets_out && tcp_minshall_check(tp))); 1527d4589926SEric Dumazet } 1528605ad7f1SEric Dumazet 1529605ad7f1SEric Dumazet /* Return how many segs we'd like on a TSO packet, 1530605ad7f1SEric Dumazet * to send one TSO packet per ms 1531605ad7f1SEric Dumazet */ 1532605ad7f1SEric Dumazet static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now) 1533605ad7f1SEric Dumazet { 1534605ad7f1SEric Dumazet u32 bytes, segs; 1535605ad7f1SEric Dumazet 1536605ad7f1SEric Dumazet bytes = min(sk->sk_pacing_rate >> 10, 1537605ad7f1SEric Dumazet sk->sk_gso_max_size - 1 - MAX_TCP_HEADER); 1538605ad7f1SEric Dumazet 1539605ad7f1SEric Dumazet /* Goal is to send at least one packet per ms, 1540605ad7f1SEric Dumazet * not one big TSO packet every 100 ms. 1541605ad7f1SEric Dumazet * This preserves ACK clocking and is consistent 1542605ad7f1SEric Dumazet * with tcp_tso_should_defer() heuristic. 1543605ad7f1SEric Dumazet */ 1544605ad7f1SEric Dumazet segs = max_t(u32, bytes / mss_now, sysctl_tcp_min_tso_segs); 1545605ad7f1SEric Dumazet 1546605ad7f1SEric Dumazet return min_t(u32, segs, sk->sk_gso_max_segs); 1547605ad7f1SEric Dumazet } 1548605ad7f1SEric Dumazet 1549d4589926SEric Dumazet /* Returns the portion of skb which can be sent right away */ 1550d4589926SEric Dumazet static unsigned int tcp_mss_split_point(const struct sock *sk, 1551d4589926SEric Dumazet const struct sk_buff *skb, 1552d4589926SEric Dumazet unsigned int mss_now, 1553d4589926SEric Dumazet unsigned int max_segs, 1554d4589926SEric Dumazet int nonagle) 1555c1b4a7e6SDavid S. Miller { 1556cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1557d4589926SEric Dumazet u32 partial, needed, window, max_len; 1558c1b4a7e6SDavid S. Miller 155990840defSIlpo Järvinen window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 15601485348dSBen Hutchings max_len = mss_now * max_segs; 15610e3a4803SIlpo Järvinen 15621485348dSBen Hutchings if (likely(max_len <= window && skb != tcp_write_queue_tail(sk))) 15631485348dSBen Hutchings return max_len; 15640e3a4803SIlpo Järvinen 15655ea3a748SIlpo Järvinen needed = min(skb->len, window); 15665ea3a748SIlpo Järvinen 15671485348dSBen Hutchings if (max_len <= needed) 15681485348dSBen Hutchings return max_len; 15690e3a4803SIlpo Järvinen 1570d4589926SEric Dumazet partial = needed % mss_now; 1571d4589926SEric Dumazet /* If last segment is not a full MSS, check if Nagle rules allow us 1572d4589926SEric Dumazet * to include this last segment in this skb. 1573d4589926SEric Dumazet * Otherwise, we'll split the skb at last MSS boundary 1574d4589926SEric Dumazet */ 1575cc93fc51SPeter Pan(潘卫平) if (tcp_nagle_check(partial != 0, tp, nonagle)) 1576d4589926SEric Dumazet return needed - partial; 1577d4589926SEric Dumazet 1578d4589926SEric Dumazet return needed; 1579c1b4a7e6SDavid S. Miller } 1580c1b4a7e6SDavid S. Miller 1581c1b4a7e6SDavid S. Miller /* Can at least one segment of SKB be sent right now, according to the 1582c1b4a7e6SDavid S. Miller * congestion window rules? If so, return how many segments are allowed. 1583c1b4a7e6SDavid S. Miller */ 1584cf533ea5SEric Dumazet static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, 1585cf533ea5SEric Dumazet const struct sk_buff *skb) 1586c1b4a7e6SDavid S. Miller { 1587d649a7a8SEric Dumazet u32 in_flight, cwnd, halfcwnd; 1588c1b4a7e6SDavid S. Miller 1589c1b4a7e6SDavid S. Miller /* Don't be strict about the congestion window for the final FIN. */ 15904de075e0SEric Dumazet if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && 15914de075e0SEric Dumazet tcp_skb_pcount(skb) == 1) 1592c1b4a7e6SDavid S. Miller return 1; 1593c1b4a7e6SDavid S. Miller 1594c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1595c1b4a7e6SDavid S. Miller cwnd = tp->snd_cwnd; 1596d649a7a8SEric Dumazet if (in_flight >= cwnd) 1597c1b4a7e6SDavid S. Miller return 0; 1598d649a7a8SEric Dumazet 1599d649a7a8SEric Dumazet /* For better scheduling, ensure we have at least 1600d649a7a8SEric Dumazet * 2 GSO packets in flight. 1601d649a7a8SEric Dumazet */ 1602d649a7a8SEric Dumazet halfcwnd = max(cwnd >> 1, 1U); 1603d649a7a8SEric Dumazet return min(halfcwnd, cwnd - in_flight); 1604c1b4a7e6SDavid S. Miller } 1605c1b4a7e6SDavid S. Miller 1606b595076aSUwe Kleine-König /* Initialize TSO state of a skb. 160767edfef7SAndi Kleen * This must be invoked the first time we consider transmitting 1608c1b4a7e6SDavid S. Miller * SKB onto the wire. 1609c1b4a7e6SDavid S. Miller */ 16105bbb432cSEric Dumazet static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now) 1611c1b4a7e6SDavid S. Miller { 1612c1b4a7e6SDavid S. Miller int tso_segs = tcp_skb_pcount(skb); 1613c1b4a7e6SDavid S. Miller 1614f8269a49SIlpo Järvinen if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { 16155bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 1616c1b4a7e6SDavid S. Miller tso_segs = tcp_skb_pcount(skb); 1617c1b4a7e6SDavid S. Miller } 1618c1b4a7e6SDavid S. Miller return tso_segs; 1619c1b4a7e6SDavid S. Miller } 1620c1b4a7e6SDavid S. Miller 1621c1b4a7e6SDavid S. Miller 1622a2a385d6SEric Dumazet /* Return true if the Nagle test allows this packet to be 1623c1b4a7e6SDavid S. Miller * sent now. 1624c1b4a7e6SDavid S. Miller */ 1625a2a385d6SEric Dumazet static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, 1626c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1627c1b4a7e6SDavid S. Miller { 1628c1b4a7e6SDavid S. Miller /* Nagle rule does not apply to frames, which sit in the middle of the 1629c1b4a7e6SDavid S. Miller * write_queue (they have no chances to get new data). 1630c1b4a7e6SDavid S. Miller * 1631c1b4a7e6SDavid S. Miller * This is implemented in the callers, where they modify the 'nonagle' 1632c1b4a7e6SDavid S. Miller * argument based upon the location of SKB in the send queue. 1633c1b4a7e6SDavid S. Miller */ 1634c1b4a7e6SDavid S. Miller if (nonagle & TCP_NAGLE_PUSH) 1635a2a385d6SEric Dumazet return true; 1636c1b4a7e6SDavid S. Miller 16379b44190dSYuchung Cheng /* Don't use the nagle rule for urgent data (or for the final FIN). */ 16389b44190dSYuchung Cheng if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) 1639a2a385d6SEric Dumazet return true; 1640c1b4a7e6SDavid S. Miller 1641cc93fc51SPeter Pan(潘卫平) if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle)) 1642a2a385d6SEric Dumazet return true; 1643c1b4a7e6SDavid S. Miller 1644a2a385d6SEric Dumazet return false; 1645c1b4a7e6SDavid S. Miller } 1646c1b4a7e6SDavid S. Miller 1647c1b4a7e6SDavid S. Miller /* Does at least the first segment of SKB fit into the send window? */ 1648a2a385d6SEric Dumazet static bool tcp_snd_wnd_test(const struct tcp_sock *tp, 1649a2a385d6SEric Dumazet const struct sk_buff *skb, 1650056834d9SIlpo Järvinen unsigned int cur_mss) 1651c1b4a7e6SDavid S. Miller { 1652c1b4a7e6SDavid S. Miller u32 end_seq = TCP_SKB_CB(skb)->end_seq; 1653c1b4a7e6SDavid S. Miller 1654c1b4a7e6SDavid S. Miller if (skb->len > cur_mss) 1655c1b4a7e6SDavid S. Miller end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 1656c1b4a7e6SDavid S. Miller 165790840defSIlpo Järvinen return !after(end_seq, tcp_wnd_end(tp)); 1658c1b4a7e6SDavid S. Miller } 1659c1b4a7e6SDavid S. Miller 1660fe067e8aSDavid S. Miller /* This checks if the data bearing packet SKB (usually tcp_send_head(sk)) 1661c1b4a7e6SDavid S. Miller * should be put on the wire right now. If so, it returns the number of 1662c1b4a7e6SDavid S. Miller * packets allowed by the congestion window. 1663c1b4a7e6SDavid S. Miller */ 1664cf533ea5SEric Dumazet static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb, 1665c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1666c1b4a7e6SDavid S. Miller { 1667cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1668c1b4a7e6SDavid S. Miller unsigned int cwnd_quota; 1669c1b4a7e6SDavid S. Miller 16705bbb432cSEric Dumazet tcp_init_tso_segs(skb, cur_mss); 1671c1b4a7e6SDavid S. Miller 1672c1b4a7e6SDavid S. Miller if (!tcp_nagle_test(tp, skb, cur_mss, nonagle)) 1673c1b4a7e6SDavid S. Miller return 0; 1674c1b4a7e6SDavid S. Miller 1675c1b4a7e6SDavid S. Miller cwnd_quota = tcp_cwnd_test(tp, skb); 1676056834d9SIlpo Järvinen if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss)) 1677c1b4a7e6SDavid S. Miller cwnd_quota = 0; 1678c1b4a7e6SDavid S. Miller 1679c1b4a7e6SDavid S. Miller return cwnd_quota; 1680c1b4a7e6SDavid S. Miller } 1681c1b4a7e6SDavid S. Miller 168267edfef7SAndi Kleen /* Test if sending is allowed right now. */ 1683a2a385d6SEric Dumazet bool tcp_may_send_now(struct sock *sk) 1684c1b4a7e6SDavid S. Miller { 1685cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1686fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 1687c1b4a7e6SDavid S. Miller 1688a02cec21SEric Dumazet return skb && 16890c54b85fSIlpo Järvinen tcp_snd_test(sk, skb, tcp_current_mss(sk), 1690c1b4a7e6SDavid S. Miller (tcp_skb_is_last(sk, skb) ? 1691a02cec21SEric Dumazet tp->nonagle : TCP_NAGLE_PUSH)); 1692c1b4a7e6SDavid S. Miller } 1693c1b4a7e6SDavid S. Miller 1694c1b4a7e6SDavid S. Miller /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 1695c1b4a7e6SDavid S. Miller * which is put after SKB on the list. It is very much like 1696c1b4a7e6SDavid S. Miller * tcp_fragment() except that it may make several kinds of assumptions 1697c1b4a7e6SDavid S. Miller * in order to speed up the splitting operation. In particular, we 1698c1b4a7e6SDavid S. Miller * know that all the data is in scatter-gather pages, and that the 1699c1b4a7e6SDavid S. Miller * packet has never been sent out before (and thus is not cloned). 1700c1b4a7e6SDavid S. Miller */ 1701056834d9SIlpo Järvinen static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, 1702c4ead4c5SEric Dumazet unsigned int mss_now, gfp_t gfp) 1703c1b4a7e6SDavid S. Miller { 1704c1b4a7e6SDavid S. Miller struct sk_buff *buff; 1705c1b4a7e6SDavid S. Miller int nlen = skb->len - len; 17069ce01461SIlpo Järvinen u8 flags; 1707c1b4a7e6SDavid S. Miller 1708c1b4a7e6SDavid S. Miller /* All of a TSO frame must be composed of paged data. */ 1709c8ac3774SHerbert Xu if (skb->len != skb->data_len) 17106cc55e09SOctavian Purdila return tcp_fragment(sk, skb, len, mss_now, gfp); 1711c1b4a7e6SDavid S. Miller 1712eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, gfp, true); 171351456b29SIan Morris if (unlikely(!buff)) 1714c1b4a7e6SDavid S. Miller return -ENOMEM; 1715c1b4a7e6SDavid S. Miller 17163ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 17173ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1718b60b49eaSHerbert Xu buff->truesize += nlen; 1719c1b4a7e6SDavid S. Miller skb->truesize -= nlen; 1720c1b4a7e6SDavid S. Miller 1721c1b4a7e6SDavid S. Miller /* Correct the sequence numbers. */ 1722c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 1723c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 1724c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 1725c1b4a7e6SDavid S. Miller 1726c1b4a7e6SDavid S. Miller /* PSH and FIN should only be set in the second packet. */ 17274de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 17284de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 17294de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1730c1b4a7e6SDavid S. Miller 1731c1b4a7e6SDavid S. Miller /* This packet was never sent out yet, so no SACK bits. */ 1732c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->sacked = 0; 1733c1b4a7e6SDavid S. Miller 173484fa7933SPatrick McHardy buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; 1735c1b4a7e6SDavid S. Miller skb_split(skb, buff, len); 1736490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 1737c1b4a7e6SDavid S. Miller 1738c1b4a7e6SDavid S. Miller /* Fix up tso_factor for both original and new SKB. */ 17395bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 17405bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 1741c1b4a7e6SDavid S. Miller 1742c1b4a7e6SDavid S. Miller /* Link BUFF into the send queue. */ 1743f4a775d1SEric Dumazet __skb_header_release(buff); 1744fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 1745c1b4a7e6SDavid S. Miller 1746c1b4a7e6SDavid S. Miller return 0; 1747c1b4a7e6SDavid S. Miller } 1748c1b4a7e6SDavid S. Miller 1749c1b4a7e6SDavid S. Miller /* Try to defer sending, if possible, in order to minimize the amount 1750c1b4a7e6SDavid S. Miller * of TSO splitting we do. View it as a kind of TSO Nagle test. 1751c1b4a7e6SDavid S. Miller * 1752c1b4a7e6SDavid S. Miller * This algorithm is from John Heffner. 1753c1b4a7e6SDavid S. Miller */ 1754ca8a2263SNeal Cardwell static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb, 1755605ad7f1SEric Dumazet bool *is_cwnd_limited, u32 max_segs) 1756c1b4a7e6SDavid S. Miller { 17576687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 175850c8339eSEric Dumazet u32 age, send_win, cong_win, limit, in_flight; 175950c8339eSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 176050c8339eSEric Dumazet struct skb_mstamp now; 176150c8339eSEric Dumazet struct sk_buff *head; 1762ad9f4f50SEric Dumazet int win_divisor; 1763c1b4a7e6SDavid S. Miller 17644de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 1765ae8064acSJohn Heffner goto send_now; 1766c1b4a7e6SDavid S. Miller 176799d7662aSEric Dumazet if (icsk->icsk_ca_state >= TCP_CA_Recovery) 1768ae8064acSJohn Heffner goto send_now; 1769ae8064acSJohn Heffner 17705f852eb5SEric Dumazet /* Avoid bursty behavior by allowing defer 17715f852eb5SEric Dumazet * only if the last write was recent. 17725f852eb5SEric Dumazet */ 17735f852eb5SEric Dumazet if ((s32)(tcp_time_stamp - tp->lsndtime) > 0) 1774ae8064acSJohn Heffner goto send_now; 1775908a75c1SDavid S. Miller 1776c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1777c1b4a7e6SDavid S. Miller 1778056834d9SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight)); 1779c1b4a7e6SDavid S. Miller 178090840defSIlpo Järvinen send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1781c1b4a7e6SDavid S. Miller 1782c1b4a7e6SDavid S. Miller /* From in_flight test above, we know that cwnd > in_flight. */ 1783c1b4a7e6SDavid S. Miller cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; 1784c1b4a7e6SDavid S. Miller 1785c1b4a7e6SDavid S. Miller limit = min(send_win, cong_win); 1786c1b4a7e6SDavid S. Miller 1787ba244fe9SDavid S. Miller /* If a full-sized TSO skb can be sent, do it. */ 1788605ad7f1SEric Dumazet if (limit >= max_segs * tp->mss_cache) 1789ae8064acSJohn Heffner goto send_now; 1790ba244fe9SDavid S. Miller 179162ad2761SIlpo Järvinen /* Middle in queue won't get any more data, full sendable already? */ 179262ad2761SIlpo Järvinen if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) 179362ad2761SIlpo Järvinen goto send_now; 179462ad2761SIlpo Järvinen 1795ad9f4f50SEric Dumazet win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor); 1796ad9f4f50SEric Dumazet if (win_divisor) { 1797c1b4a7e6SDavid S. Miller u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); 1798c1b4a7e6SDavid S. Miller 1799c1b4a7e6SDavid S. Miller /* If at least some fraction of a window is available, 1800c1b4a7e6SDavid S. Miller * just use it. 1801c1b4a7e6SDavid S. Miller */ 1802ad9f4f50SEric Dumazet chunk /= win_divisor; 1803c1b4a7e6SDavid S. Miller if (limit >= chunk) 1804ae8064acSJohn Heffner goto send_now; 1805c1b4a7e6SDavid S. Miller } else { 1806c1b4a7e6SDavid S. Miller /* Different approach, try not to defer past a single 1807c1b4a7e6SDavid S. Miller * ACK. Receiver should ACK every other full sized 1808c1b4a7e6SDavid S. Miller * frame, so if we have space for more than 3 frames 1809c1b4a7e6SDavid S. Miller * then send now. 1810c1b4a7e6SDavid S. Miller */ 18116b5a5c0dSNeal Cardwell if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) 1812ae8064acSJohn Heffner goto send_now; 1813c1b4a7e6SDavid S. Miller } 1814c1b4a7e6SDavid S. Miller 181550c8339eSEric Dumazet head = tcp_write_queue_head(sk); 181650c8339eSEric Dumazet skb_mstamp_get(&now); 181750c8339eSEric Dumazet age = skb_mstamp_us_delta(&now, &head->skb_mstamp); 181850c8339eSEric Dumazet /* If next ACK is likely to come too late (half srtt), do not defer */ 181950c8339eSEric Dumazet if (age < (tp->srtt_us >> 4)) 182050c8339eSEric Dumazet goto send_now; 182150c8339eSEric Dumazet 18225f852eb5SEric Dumazet /* Ok, it looks like it is advisable to defer. */ 1823ae8064acSJohn Heffner 1824d2e1339fSBendik Rønning Opstad if (cong_win < send_win && cong_win <= skb->len) 1825ca8a2263SNeal Cardwell *is_cwnd_limited = true; 1826ca8a2263SNeal Cardwell 1827a2a385d6SEric Dumazet return true; 1828ae8064acSJohn Heffner 1829ae8064acSJohn Heffner send_now: 1830a2a385d6SEric Dumazet return false; 1831c1b4a7e6SDavid S. Miller } 1832c1b4a7e6SDavid S. Miller 183305cbc0dbSFan Du static inline void tcp_mtu_check_reprobe(struct sock *sk) 183405cbc0dbSFan Du { 183505cbc0dbSFan Du struct inet_connection_sock *icsk = inet_csk(sk); 183605cbc0dbSFan Du struct tcp_sock *tp = tcp_sk(sk); 183705cbc0dbSFan Du struct net *net = sock_net(sk); 183805cbc0dbSFan Du u32 interval; 183905cbc0dbSFan Du s32 delta; 184005cbc0dbSFan Du 184105cbc0dbSFan Du interval = net->ipv4.sysctl_tcp_probe_interval; 184205cbc0dbSFan Du delta = tcp_time_stamp - icsk->icsk_mtup.probe_timestamp; 184305cbc0dbSFan Du if (unlikely(delta >= interval * HZ)) { 184405cbc0dbSFan Du int mss = tcp_current_mss(sk); 184505cbc0dbSFan Du 184605cbc0dbSFan Du /* Update current search range */ 184705cbc0dbSFan Du icsk->icsk_mtup.probe_size = 0; 184805cbc0dbSFan Du icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + 184905cbc0dbSFan Du sizeof(struct tcphdr) + 185005cbc0dbSFan Du icsk->icsk_af_ops->net_header_len; 185105cbc0dbSFan Du icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss); 185205cbc0dbSFan Du 185305cbc0dbSFan Du /* Update probe time stamp */ 185405cbc0dbSFan Du icsk->icsk_mtup.probe_timestamp = tcp_time_stamp; 185505cbc0dbSFan Du } 185605cbc0dbSFan Du } 185705cbc0dbSFan Du 18585d424d5aSJohn Heffner /* Create a new MTU probe if we are ready. 185967edfef7SAndi Kleen * MTU probe is regularly attempting to increase the path MTU by 186067edfef7SAndi Kleen * deliberately sending larger packets. This discovers routing 186167edfef7SAndi Kleen * changes resulting in larger path MTUs. 186267edfef7SAndi Kleen * 18635d424d5aSJohn Heffner * Returns 0 if we should wait to probe (no cwnd available), 18645d424d5aSJohn Heffner * 1 if a probe was sent, 1865056834d9SIlpo Järvinen * -1 otherwise 1866056834d9SIlpo Järvinen */ 18675d424d5aSJohn Heffner static int tcp_mtu_probe(struct sock *sk) 18685d424d5aSJohn Heffner { 18695d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 18705d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 18715d424d5aSJohn Heffner struct sk_buff *skb, *nskb, *next; 18726b58e0a5SFan Du struct net *net = sock_net(sk); 18735d424d5aSJohn Heffner int len; 18745d424d5aSJohn Heffner int probe_size; 187591cc17c0SIlpo Järvinen int size_needed; 18765d424d5aSJohn Heffner int copy; 18775d424d5aSJohn Heffner int mss_now; 18786b58e0a5SFan Du int interval; 18795d424d5aSJohn Heffner 18805d424d5aSJohn Heffner /* Not currently probing/verifying, 18815d424d5aSJohn Heffner * not in recovery, 18825d424d5aSJohn Heffner * have enough cwnd, and 18835d424d5aSJohn Heffner * not SACKing (the variable headers throw things off) */ 18845d424d5aSJohn Heffner if (!icsk->icsk_mtup.enabled || 18855d424d5aSJohn Heffner icsk->icsk_mtup.probe_size || 18865d424d5aSJohn Heffner inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 18875d424d5aSJohn Heffner tp->snd_cwnd < 11 || 1888cabeccbdSIlpo Järvinen tp->rx_opt.num_sacks || tp->rx_opt.dsack) 18895d424d5aSJohn Heffner return -1; 18905d424d5aSJohn Heffner 18916b58e0a5SFan Du /* Use binary search for probe_size between tcp_mss_base, 18926b58e0a5SFan Du * and current mss_clamp. if (search_high - search_low) 18936b58e0a5SFan Du * smaller than a threshold, backoff from probing. 18946b58e0a5SFan Du */ 18950c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 18966b58e0a5SFan Du probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high + 18976b58e0a5SFan Du icsk->icsk_mtup.search_low) >> 1); 189891cc17c0SIlpo Järvinen size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; 18996b58e0a5SFan Du interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low; 190005cbc0dbSFan Du /* When misfortune happens, we are reprobing actively, 190105cbc0dbSFan Du * and then reprobe timer has expired. We stick with current 190205cbc0dbSFan Du * probing process by not resetting search range to its orignal. 190305cbc0dbSFan Du */ 19046b58e0a5SFan Du if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) || 190505cbc0dbSFan Du interval < net->ipv4.sysctl_tcp_probe_threshold) { 190605cbc0dbSFan Du /* Check whether enough time has elaplased for 190705cbc0dbSFan Du * another round of probing. 190805cbc0dbSFan Du */ 190905cbc0dbSFan Du tcp_mtu_check_reprobe(sk); 19105d424d5aSJohn Heffner return -1; 19115d424d5aSJohn Heffner } 19125d424d5aSJohn Heffner 19135d424d5aSJohn Heffner /* Have enough data in the send queue to probe? */ 19147f9c33e5SIlpo Järvinen if (tp->write_seq - tp->snd_nxt < size_needed) 19155d424d5aSJohn Heffner return -1; 19165d424d5aSJohn Heffner 191791cc17c0SIlpo Järvinen if (tp->snd_wnd < size_needed) 19185d424d5aSJohn Heffner return -1; 191990840defSIlpo Järvinen if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) 19205d424d5aSJohn Heffner return 0; 19215d424d5aSJohn Heffner 1922d67c58e9SIlpo Järvinen /* Do we need to wait to drain cwnd? With none in flight, don't stall */ 1923d67c58e9SIlpo Järvinen if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { 1924d67c58e9SIlpo Järvinen if (!tcp_packets_in_flight(tp)) 19255d424d5aSJohn Heffner return -1; 19265d424d5aSJohn Heffner else 19275d424d5aSJohn Heffner return 0; 19285d424d5aSJohn Heffner } 19295d424d5aSJohn Heffner 19305d424d5aSJohn Heffner /* We're allowed to probe. Build it now. */ 1931eb934478SEric Dumazet nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false); 193251456b29SIan Morris if (!nskb) 19335d424d5aSJohn Heffner return -1; 19343ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 19353ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 19365d424d5aSJohn Heffner 1937fe067e8aSDavid S. Miller skb = tcp_send_head(sk); 19385d424d5aSJohn Heffner 19395d424d5aSJohn Heffner TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 19405d424d5aSJohn Heffner TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 19414de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; 19425d424d5aSJohn Heffner TCP_SKB_CB(nskb)->sacked = 0; 19435d424d5aSJohn Heffner nskb->csum = 0; 194484fa7933SPatrick McHardy nskb->ip_summed = skb->ip_summed; 19455d424d5aSJohn Heffner 194650c4817eSIlpo Järvinen tcp_insert_write_queue_before(nskb, skb, sk); 194750c4817eSIlpo Järvinen 19485d424d5aSJohn Heffner len = 0; 1949234b6860SIlpo Järvinen tcp_for_write_queue_from_safe(skb, next, sk) { 19505d424d5aSJohn Heffner copy = min_t(int, skb->len, probe_size - len); 19515d424d5aSJohn Heffner if (nskb->ip_summed) 19525d424d5aSJohn Heffner skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); 19535d424d5aSJohn Heffner else 19545d424d5aSJohn Heffner nskb->csum = skb_copy_and_csum_bits(skb, 0, 1955056834d9SIlpo Järvinen skb_put(nskb, copy), 1956056834d9SIlpo Järvinen copy, nskb->csum); 19575d424d5aSJohn Heffner 19585d424d5aSJohn Heffner if (skb->len <= copy) { 19595d424d5aSJohn Heffner /* We've eaten all the data from this skb. 19605d424d5aSJohn Heffner * Throw it away. */ 19614de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 1962fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 19633ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 19645d424d5aSJohn Heffner } else { 19654de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & 1966a3433f35SChangli Gao ~(TCPHDR_FIN|TCPHDR_PSH); 19675d424d5aSJohn Heffner if (!skb_shinfo(skb)->nr_frags) { 19685d424d5aSJohn Heffner skb_pull(skb, copy); 196984fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 1970056834d9SIlpo Järvinen skb->csum = csum_partial(skb->data, 1971056834d9SIlpo Järvinen skb->len, 0); 19725d424d5aSJohn Heffner } else { 19735d424d5aSJohn Heffner __pskb_trim_head(skb, copy); 19745bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 19755d424d5aSJohn Heffner } 19765d424d5aSJohn Heffner TCP_SKB_CB(skb)->seq += copy; 19775d424d5aSJohn Heffner } 19785d424d5aSJohn Heffner 19795d424d5aSJohn Heffner len += copy; 1980234b6860SIlpo Järvinen 1981234b6860SIlpo Järvinen if (len >= probe_size) 1982234b6860SIlpo Järvinen break; 19835d424d5aSJohn Heffner } 19845bbb432cSEric Dumazet tcp_init_tso_segs(nskb, nskb->len); 19855d424d5aSJohn Heffner 19865d424d5aSJohn Heffner /* We're ready to send. If this fails, the probe will 19877faee5c0SEric Dumazet * be resegmented into mss-sized pieces by tcp_write_xmit(). 19887faee5c0SEric Dumazet */ 19895d424d5aSJohn Heffner if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 19905d424d5aSJohn Heffner /* Decrement cwnd here because we are sending 19915d424d5aSJohn Heffner * effectively two packets. */ 19925d424d5aSJohn Heffner tp->snd_cwnd--; 199366f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, nskb); 19945d424d5aSJohn Heffner 19955d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 19960e7b1368SJohn Heffner tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 19970e7b1368SJohn Heffner tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 19985d424d5aSJohn Heffner 19995d424d5aSJohn Heffner return 1; 20005d424d5aSJohn Heffner } 20015d424d5aSJohn Heffner 20025d424d5aSJohn Heffner return -1; 20035d424d5aSJohn Heffner } 20045d424d5aSJohn Heffner 20051da177e4SLinus Torvalds /* This routine writes packets to the network. It advances the 20061da177e4SLinus Torvalds * send_head. This happens as incoming acks open up the remote 20071da177e4SLinus Torvalds * window for us. 20081da177e4SLinus Torvalds * 2009f8269a49SIlpo Järvinen * LARGESEND note: !tcp_urg_mode is overkill, only frames between 2010f8269a49SIlpo Järvinen * snd_up-64k-mss .. snd_up cannot be large. However, taking into 2011f8269a49SIlpo Järvinen * account rare use of URG, this is not a big flaw. 2012f8269a49SIlpo Järvinen * 20136ba8a3b1SNandita Dukkipati * Send at most one packet when push_one > 0. Temporarily ignore 20146ba8a3b1SNandita Dukkipati * cwnd limit to force at most one packet out when push_one == 2. 20156ba8a3b1SNandita Dukkipati 2016a2a385d6SEric Dumazet * Returns true, if no segments are in flight and we have queued segments, 2017a2a385d6SEric Dumazet * but cannot send anything now because of SWS or another problem. 20181da177e4SLinus Torvalds */ 2019a2a385d6SEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 2020d5dd9175SIlpo Järvinen int push_one, gfp_t gfp) 20211da177e4SLinus Torvalds { 20221da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 202392df7b51SDavid S. Miller struct sk_buff *skb; 2024c1b4a7e6SDavid S. Miller unsigned int tso_segs, sent_pkts; 2025c1b4a7e6SDavid S. Miller int cwnd_quota; 20265d424d5aSJohn Heffner int result; 2027ca8a2263SNeal Cardwell bool is_cwnd_limited = false; 2028605ad7f1SEric Dumazet u32 max_segs; 20291da177e4SLinus Torvalds 2030c1b4a7e6SDavid S. Miller sent_pkts = 0; 20315d424d5aSJohn Heffner 2032d5dd9175SIlpo Järvinen if (!push_one) { 20335d424d5aSJohn Heffner /* Do MTU probing. */ 2034d5dd9175SIlpo Järvinen result = tcp_mtu_probe(sk); 2035d5dd9175SIlpo Järvinen if (!result) { 2036a2a385d6SEric Dumazet return false; 20375d424d5aSJohn Heffner } else if (result > 0) { 20385d424d5aSJohn Heffner sent_pkts = 1; 20395d424d5aSJohn Heffner } 2040d5dd9175SIlpo Järvinen } 20415d424d5aSJohn Heffner 2042605ad7f1SEric Dumazet max_segs = tcp_tso_autosize(sk, mss_now); 2043fe067e8aSDavid S. Miller while ((skb = tcp_send_head(sk))) { 2044c8ac3774SHerbert Xu unsigned int limit; 2045c8ac3774SHerbert Xu 20465bbb432cSEric Dumazet tso_segs = tcp_init_tso_segs(skb, mss_now); 2047c1b4a7e6SDavid S. Miller BUG_ON(!tso_segs); 2048c1b4a7e6SDavid S. Miller 20499d186cacSAndrey Vagin if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) { 20507faee5c0SEric Dumazet /* "skb_mstamp" is used as a start point for the retransmit timer */ 20517faee5c0SEric Dumazet skb_mstamp_get(&skb->skb_mstamp); 2052ec342325SAndrew Vagin goto repair; /* Skip network transmission */ 20539d186cacSAndrey Vagin } 2054ec342325SAndrew Vagin 2055b68e9f85SHerbert Xu cwnd_quota = tcp_cwnd_test(tp, skb); 20566ba8a3b1SNandita Dukkipati if (!cwnd_quota) { 20576ba8a3b1SNandita Dukkipati if (push_one == 2) 20586ba8a3b1SNandita Dukkipati /* Force out a loss probe pkt. */ 20596ba8a3b1SNandita Dukkipati cwnd_quota = 1; 20606ba8a3b1SNandita Dukkipati else 2061b68e9f85SHerbert Xu break; 20626ba8a3b1SNandita Dukkipati } 2063b68e9f85SHerbert Xu 2064b68e9f85SHerbert Xu if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) 2065b68e9f85SHerbert Xu break; 2066b68e9f85SHerbert Xu 2067d6a4e26aSEric Dumazet if (tso_segs == 1) { 2068aa93466bSDavid S. Miller if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 2069aa93466bSDavid S. Miller (tcp_skb_is_last(sk, skb) ? 2070aa93466bSDavid S. Miller nonagle : TCP_NAGLE_PUSH)))) 2071aa93466bSDavid S. Miller break; 2072c1b4a7e6SDavid S. Miller } else { 2073ca8a2263SNeal Cardwell if (!push_one && 2074605ad7f1SEric Dumazet tcp_tso_should_defer(sk, skb, &is_cwnd_limited, 2075605ad7f1SEric Dumazet max_segs)) 2076aa93466bSDavid S. Miller break; 2077c1b4a7e6SDavid S. Miller } 2078aa93466bSDavid S. Miller 2079605ad7f1SEric Dumazet limit = mss_now; 2080d6a4e26aSEric Dumazet if (tso_segs > 1 && !tcp_urg_mode(tp)) 2081605ad7f1SEric Dumazet limit = tcp_mss_split_point(sk, skb, mss_now, 2082605ad7f1SEric Dumazet min_t(unsigned int, 2083605ad7f1SEric Dumazet cwnd_quota, 2084605ad7f1SEric Dumazet max_segs), 2085605ad7f1SEric Dumazet nonagle); 2086605ad7f1SEric Dumazet 2087605ad7f1SEric Dumazet if (skb->len > limit && 2088605ad7f1SEric Dumazet unlikely(tso_fragment(sk, skb, limit, mss_now, gfp))) 2089605ad7f1SEric Dumazet break; 2090605ad7f1SEric Dumazet 2091c9eeec26SEric Dumazet /* TCP Small Queues : 2092c9eeec26SEric Dumazet * Control number of packets in qdisc/devices to two packets / or ~1 ms. 2093c9eeec26SEric Dumazet * This allows for : 2094c9eeec26SEric Dumazet * - better RTT estimation and ACK scheduling 2095c9eeec26SEric Dumazet * - faster recovery 2096c9eeec26SEric Dumazet * - high rates 209798e09386SEric Dumazet * Alas, some drivers / subsystems require a fair amount 209898e09386SEric Dumazet * of queued bytes to ensure line rate. 209998e09386SEric Dumazet * One example is wifi aggregation (802.11 AMPDU) 210046d3ceabSEric Dumazet */ 2101605ad7f1SEric Dumazet limit = max(2 * skb->truesize, sk->sk_pacing_rate >> 10); 2102605ad7f1SEric Dumazet limit = min_t(u32, limit, sysctl_tcp_limit_output_bytes); 2103c9eeec26SEric Dumazet 2104c9eeec26SEric Dumazet if (atomic_read(&sk->sk_wmem_alloc) > limit) { 210546d3ceabSEric Dumazet set_bit(TSQ_THROTTLED, &tp->tsq_flags); 2106bf06200eSJohn Ogness /* It is possible TX completion already happened 2107bf06200eSJohn Ogness * before we set TSQ_THROTTLED, so we must 2108bf06200eSJohn Ogness * test again the condition. 2109bf06200eSJohn Ogness */ 21104e857c58SPeter Zijlstra smp_mb__after_atomic(); 2111bf06200eSJohn Ogness if (atomic_read(&sk->sk_wmem_alloc) > limit) 211246d3ceabSEric Dumazet break; 211346d3ceabSEric Dumazet } 2114c9eeec26SEric Dumazet 2115d5dd9175SIlpo Järvinen if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) 21161da177e4SLinus Torvalds break; 21171da177e4SLinus Torvalds 2118ec342325SAndrew Vagin repair: 21191da177e4SLinus Torvalds /* Advance the send_head. This one is sent out. 21201da177e4SLinus Torvalds * This call will increment packets_out. 21211da177e4SLinus Torvalds */ 212266f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 21231da177e4SLinus Torvalds 21241da177e4SLinus Torvalds tcp_minshall_update(tp, mss_now, skb); 2125a262f0cdSNandita Dukkipati sent_pkts += tcp_skb_pcount(skb); 2126d5dd9175SIlpo Järvinen 2127d5dd9175SIlpo Järvinen if (push_one) 2128d5dd9175SIlpo Järvinen break; 21291da177e4SLinus Torvalds } 21301da177e4SLinus Torvalds 2131aa93466bSDavid S. Miller if (likely(sent_pkts)) { 2132684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 2133684bad11SYuchung Cheng tp->prr_out += sent_pkts; 21346ba8a3b1SNandita Dukkipati 21356ba8a3b1SNandita Dukkipati /* Send one loss probe per tail loss episode. */ 21366ba8a3b1SNandita Dukkipati if (push_one != 2) 21376ba8a3b1SNandita Dukkipati tcp_schedule_loss_probe(sk); 2138d2e1339fSBendik Rønning Opstad is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd); 2139ca8a2263SNeal Cardwell tcp_cwnd_validate(sk, is_cwnd_limited); 2140a2a385d6SEric Dumazet return false; 21411da177e4SLinus Torvalds } 2142b340b264SYuchung Cheng return !tp->packets_out && tcp_send_head(sk); 21436ba8a3b1SNandita Dukkipati } 21446ba8a3b1SNandita Dukkipati 21456ba8a3b1SNandita Dukkipati bool tcp_schedule_loss_probe(struct sock *sk) 21466ba8a3b1SNandita Dukkipati { 21476ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 21486ba8a3b1SNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 21496ba8a3b1SNandita Dukkipati u32 timeout, tlp_time_stamp, rto_time_stamp; 2150740b0f18SEric Dumazet u32 rtt = usecs_to_jiffies(tp->srtt_us >> 3); 21516ba8a3b1SNandita Dukkipati 21526ba8a3b1SNandita Dukkipati if (WARN_ON(icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS)) 21536ba8a3b1SNandita Dukkipati return false; 21546ba8a3b1SNandita Dukkipati /* No consecutive loss probes. */ 21556ba8a3b1SNandita Dukkipati if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) { 21566ba8a3b1SNandita Dukkipati tcp_rearm_rto(sk); 21576ba8a3b1SNandita Dukkipati return false; 21586ba8a3b1SNandita Dukkipati } 21596ba8a3b1SNandita Dukkipati /* Don't do any loss probe on a Fast Open connection before 3WHS 21606ba8a3b1SNandita Dukkipati * finishes. 21616ba8a3b1SNandita Dukkipati */ 2162f9b99582SYuchung Cheng if (tp->fastopen_rsk) 21636ba8a3b1SNandita Dukkipati return false; 21646ba8a3b1SNandita Dukkipati 21656ba8a3b1SNandita Dukkipati /* TLP is only scheduled when next timer event is RTO. */ 21666ba8a3b1SNandita Dukkipati if (icsk->icsk_pending != ICSK_TIME_RETRANS) 21676ba8a3b1SNandita Dukkipati return false; 21686ba8a3b1SNandita Dukkipati 21696ba8a3b1SNandita Dukkipati /* Schedule a loss probe in 2*RTT for SACK capable connections 21706ba8a3b1SNandita Dukkipati * in Open state, that are either limited by cwnd or application. 21716ba8a3b1SNandita Dukkipati */ 2172f9b99582SYuchung Cheng if (sysctl_tcp_early_retrans < 3 || !tp->packets_out || 21736ba8a3b1SNandita Dukkipati !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open) 21746ba8a3b1SNandita Dukkipati return false; 21756ba8a3b1SNandita Dukkipati 21766ba8a3b1SNandita Dukkipati if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) && 21776ba8a3b1SNandita Dukkipati tcp_send_head(sk)) 21786ba8a3b1SNandita Dukkipati return false; 21796ba8a3b1SNandita Dukkipati 21806ba8a3b1SNandita Dukkipati /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account 2181f9b99582SYuchung Cheng * for delayed ack when there's one outstanding packet. If no RTT 2182f9b99582SYuchung Cheng * sample is available then probe after TCP_TIMEOUT_INIT. 21836ba8a3b1SNandita Dukkipati */ 2184f9b99582SYuchung Cheng timeout = rtt << 1 ? : TCP_TIMEOUT_INIT; 21856ba8a3b1SNandita Dukkipati if (tp->packets_out == 1) 21866ba8a3b1SNandita Dukkipati timeout = max_t(u32, timeout, 21876ba8a3b1SNandita Dukkipati (rtt + (rtt >> 1) + TCP_DELACK_MAX)); 21886ba8a3b1SNandita Dukkipati timeout = max_t(u32, timeout, msecs_to_jiffies(10)); 21896ba8a3b1SNandita Dukkipati 21906ba8a3b1SNandita Dukkipati /* If RTO is shorter, just schedule TLP in its place. */ 21916ba8a3b1SNandita Dukkipati tlp_time_stamp = tcp_time_stamp + timeout; 21926ba8a3b1SNandita Dukkipati rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout; 21936ba8a3b1SNandita Dukkipati if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) { 21946ba8a3b1SNandita Dukkipati s32 delta = rto_time_stamp - tcp_time_stamp; 21956ba8a3b1SNandita Dukkipati if (delta > 0) 21966ba8a3b1SNandita Dukkipati timeout = delta; 21976ba8a3b1SNandita Dukkipati } 21986ba8a3b1SNandita Dukkipati 21996ba8a3b1SNandita Dukkipati inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, 22006ba8a3b1SNandita Dukkipati TCP_RTO_MAX); 22016ba8a3b1SNandita Dukkipati return true; 22026ba8a3b1SNandita Dukkipati } 22036ba8a3b1SNandita Dukkipati 22041f3279aeSEric Dumazet /* Thanks to skb fast clones, we can detect if a prior transmit of 22051f3279aeSEric Dumazet * a packet is still in a qdisc or driver queue. 22061f3279aeSEric Dumazet * In this case, there is very little point doing a retransmit ! 22071f3279aeSEric Dumazet * Note: This is called from BH context only. 22081f3279aeSEric Dumazet */ 22091f3279aeSEric Dumazet static bool skb_still_in_host_queue(const struct sock *sk, 22101f3279aeSEric Dumazet const struct sk_buff *skb) 22111f3279aeSEric Dumazet { 221239bb5e62SEric Dumazet if (unlikely(skb_fclone_busy(sk, skb))) { 22131f3279aeSEric Dumazet NET_INC_STATS_BH(sock_net(sk), 22141f3279aeSEric Dumazet LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES); 22151f3279aeSEric Dumazet return true; 22161f3279aeSEric Dumazet } 22171f3279aeSEric Dumazet return false; 22181f3279aeSEric Dumazet } 22191f3279aeSEric Dumazet 2220b340b264SYuchung Cheng /* When probe timeout (PTO) fires, try send a new segment if possible, else 22216ba8a3b1SNandita Dukkipati * retransmit the last segment. 22226ba8a3b1SNandita Dukkipati */ 22236ba8a3b1SNandita Dukkipati void tcp_send_loss_probe(struct sock *sk) 22246ba8a3b1SNandita Dukkipati { 22259b717a8dSNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 22266ba8a3b1SNandita Dukkipati struct sk_buff *skb; 22276ba8a3b1SNandita Dukkipati int pcount; 22286ba8a3b1SNandita Dukkipati int mss = tcp_current_mss(sk); 22296ba8a3b1SNandita Dukkipati 2230b340b264SYuchung Cheng skb = tcp_send_head(sk); 2231b340b264SYuchung Cheng if (skb) { 2232b340b264SYuchung Cheng if (tcp_snd_wnd_test(tp, skb, mss)) { 2233b340b264SYuchung Cheng pcount = tp->packets_out; 2234b340b264SYuchung Cheng tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC); 2235b340b264SYuchung Cheng if (tp->packets_out > pcount) 2236b340b264SYuchung Cheng goto probe_sent; 22376ba8a3b1SNandita Dukkipati goto rearm_timer; 22386ba8a3b1SNandita Dukkipati } 2239b340b264SYuchung Cheng skb = tcp_write_queue_prev(sk, skb); 2240b340b264SYuchung Cheng } else { 2241b340b264SYuchung Cheng skb = tcp_write_queue_tail(sk); 2242b340b264SYuchung Cheng } 22436ba8a3b1SNandita Dukkipati 22449b717a8dSNandita Dukkipati /* At most one outstanding TLP retransmission. */ 22459b717a8dSNandita Dukkipati if (tp->tlp_high_seq) 22469b717a8dSNandita Dukkipati goto rearm_timer; 22479b717a8dSNandita Dukkipati 22486ba8a3b1SNandita Dukkipati /* Retransmit last segment. */ 22496ba8a3b1SNandita Dukkipati if (WARN_ON(!skb)) 22506ba8a3b1SNandita Dukkipati goto rearm_timer; 22516ba8a3b1SNandita Dukkipati 22521f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 22531f3279aeSEric Dumazet goto rearm_timer; 22541f3279aeSEric Dumazet 22556ba8a3b1SNandita Dukkipati pcount = tcp_skb_pcount(skb); 22566ba8a3b1SNandita Dukkipati if (WARN_ON(!pcount)) 22576ba8a3b1SNandita Dukkipati goto rearm_timer; 22586ba8a3b1SNandita Dukkipati 22596ba8a3b1SNandita Dukkipati if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) { 22606cc55e09SOctavian Purdila if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss, 22616cc55e09SOctavian Purdila GFP_ATOMIC))) 22626ba8a3b1SNandita Dukkipati goto rearm_timer; 2263b340b264SYuchung Cheng skb = tcp_write_queue_next(sk, skb); 22646ba8a3b1SNandita Dukkipati } 22656ba8a3b1SNandita Dukkipati 22666ba8a3b1SNandita Dukkipati if (WARN_ON(!skb || !tcp_skb_pcount(skb))) 22676ba8a3b1SNandita Dukkipati goto rearm_timer; 22686ba8a3b1SNandita Dukkipati 2269b340b264SYuchung Cheng if (__tcp_retransmit_skb(sk, skb)) 2270b340b264SYuchung Cheng goto rearm_timer; 22716ba8a3b1SNandita Dukkipati 22729b717a8dSNandita Dukkipati /* Record snd_nxt for loss detection. */ 22739b717a8dSNandita Dukkipati tp->tlp_high_seq = tp->snd_nxt; 22749b717a8dSNandita Dukkipati 2275b340b264SYuchung Cheng probe_sent: 2276fcd16c0aSYuchung Cheng NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSPROBES); 2277fcd16c0aSYuchung Cheng /* Reset s.t. tcp_rearm_rto will restart timer from now */ 2278fcd16c0aSYuchung Cheng inet_csk(sk)->icsk_pending = 0; 2279b340b264SYuchung Cheng rearm_timer: 2280fcd16c0aSYuchung Cheng tcp_rearm_rto(sk); 22811da177e4SLinus Torvalds } 22821da177e4SLinus Torvalds 2283a762a980SDavid S. Miller /* Push out any pending frames which were held back due to 2284a762a980SDavid S. Miller * TCP_CORK or attempt at coalescing tiny packets. 2285a762a980SDavid S. Miller * The socket must be locked by the caller. 2286a762a980SDavid S. Miller */ 22879e412ba7SIlpo Järvinen void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 22889e412ba7SIlpo Järvinen int nonagle) 2289a762a980SDavid S. Miller { 2290726e07a8SIlpo Järvinen /* If we are closed, the bytes will have to remain here. 2291726e07a8SIlpo Järvinen * In time closedown will finish, we empty the write queue and 2292726e07a8SIlpo Järvinen * all will be happy. 2293726e07a8SIlpo Järvinen */ 2294726e07a8SIlpo Järvinen if (unlikely(sk->sk_state == TCP_CLOSE)) 2295726e07a8SIlpo Järvinen return; 2296726e07a8SIlpo Järvinen 229799a1dec7SMel Gorman if (tcp_write_xmit(sk, cur_mss, nonagle, 0, 22987450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC))) 22999e412ba7SIlpo Järvinen tcp_check_probe_timer(sk); 2300a762a980SDavid S. Miller } 2301a762a980SDavid S. Miller 2302c1b4a7e6SDavid S. Miller /* Send _single_ skb sitting at the send head. This function requires 2303c1b4a7e6SDavid S. Miller * true push pending frames to setup probe timer etc. 2304c1b4a7e6SDavid S. Miller */ 2305c1b4a7e6SDavid S. Miller void tcp_push_one(struct sock *sk, unsigned int mss_now) 2306c1b4a7e6SDavid S. Miller { 2307fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 2308c1b4a7e6SDavid S. Miller 2309c1b4a7e6SDavid S. Miller BUG_ON(!skb || skb->len < mss_now); 2310c1b4a7e6SDavid S. Miller 2311d5dd9175SIlpo Järvinen tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); 2312c1b4a7e6SDavid S. Miller } 2313c1b4a7e6SDavid S. Miller 23141da177e4SLinus Torvalds /* This function returns the amount that we can raise the 23151da177e4SLinus Torvalds * usable window based on the following constraints 23161da177e4SLinus Torvalds * 23171da177e4SLinus Torvalds * 1. The window can never be shrunk once it is offered (RFC 793) 23181da177e4SLinus Torvalds * 2. We limit memory per socket 23191da177e4SLinus Torvalds * 23201da177e4SLinus Torvalds * RFC 1122: 23211da177e4SLinus Torvalds * "the suggested [SWS] avoidance algorithm for the receiver is to keep 23221da177e4SLinus Torvalds * RECV.NEXT + RCV.WIN fixed until: 23231da177e4SLinus Torvalds * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 23241da177e4SLinus Torvalds * 23251da177e4SLinus Torvalds * i.e. don't raise the right edge of the window until you can raise 23261da177e4SLinus Torvalds * it at least MSS bytes. 23271da177e4SLinus Torvalds * 23281da177e4SLinus Torvalds * Unfortunately, the recommended algorithm breaks header prediction, 23291da177e4SLinus Torvalds * since header prediction assumes th->window stays fixed. 23301da177e4SLinus Torvalds * 23311da177e4SLinus Torvalds * Strictly speaking, keeping th->window fixed violates the receiver 23321da177e4SLinus Torvalds * side SWS prevention criteria. The problem is that under this rule 23331da177e4SLinus Torvalds * a stream of single byte packets will cause the right side of the 23341da177e4SLinus Torvalds * window to always advance by a single byte. 23351da177e4SLinus Torvalds * 23361da177e4SLinus Torvalds * Of course, if the sender implements sender side SWS prevention 23371da177e4SLinus Torvalds * then this will not be a problem. 23381da177e4SLinus Torvalds * 23391da177e4SLinus Torvalds * BSD seems to make the following compromise: 23401da177e4SLinus Torvalds * 23411da177e4SLinus Torvalds * If the free space is less than the 1/4 of the maximum 23421da177e4SLinus Torvalds * space available and the free space is less than 1/2 mss, 23431da177e4SLinus Torvalds * then set the window to 0. 23441da177e4SLinus Torvalds * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 23451da177e4SLinus Torvalds * Otherwise, just prevent the window from shrinking 23461da177e4SLinus Torvalds * and from being larger than the largest representable value. 23471da177e4SLinus Torvalds * 23481da177e4SLinus Torvalds * This prevents incremental opening of the window in the regime 23491da177e4SLinus Torvalds * where TCP is limited by the speed of the reader side taking 23501da177e4SLinus Torvalds * data out of the TCP receive queue. It does nothing about 23511da177e4SLinus Torvalds * those cases where the window is constrained on the sender side 23521da177e4SLinus Torvalds * because the pipeline is full. 23531da177e4SLinus Torvalds * 23541da177e4SLinus Torvalds * BSD also seems to "accidentally" limit itself to windows that are a 23551da177e4SLinus Torvalds * multiple of MSS, at least until the free space gets quite small. 23561da177e4SLinus Torvalds * This would appear to be a side effect of the mbuf implementation. 23571da177e4SLinus Torvalds * Combining these two algorithms results in the observed behavior 23581da177e4SLinus Torvalds * of having a fixed window size at almost all times. 23591da177e4SLinus Torvalds * 23601da177e4SLinus Torvalds * Below we obtain similar behavior by forcing the offered window to 23611da177e4SLinus Torvalds * a multiple of the mss when it is feasible to do so. 23621da177e4SLinus Torvalds * 23631da177e4SLinus Torvalds * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 23641da177e4SLinus Torvalds * Regular options like TIMESTAMP are taken into account. 23651da177e4SLinus Torvalds */ 23661da177e4SLinus Torvalds u32 __tcp_select_window(struct sock *sk) 23671da177e4SLinus Torvalds { 2368463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 23691da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2370caa20d9aSStephen Hemminger /* MSS for the peer's data. Previous versions used mss_clamp 23711da177e4SLinus Torvalds * here. I don't know if the value based on our guesses 23721da177e4SLinus Torvalds * of peer's MSS is better for the performance. It's more correct 23731da177e4SLinus Torvalds * but may be worse for the performance because of rcv_mss 23741da177e4SLinus Torvalds * fluctuations. --SAW 1998/11/1 23751da177e4SLinus Torvalds */ 2376463c84b9SArnaldo Carvalho de Melo int mss = icsk->icsk_ack.rcv_mss; 23771da177e4SLinus Torvalds int free_space = tcp_space(sk); 237886c1a045SFlorian Westphal int allowed_space = tcp_full_space(sk); 237986c1a045SFlorian Westphal int full_space = min_t(int, tp->window_clamp, allowed_space); 23801da177e4SLinus Torvalds int window; 23811da177e4SLinus Torvalds 23821da177e4SLinus Torvalds if (mss > full_space) 23831da177e4SLinus Torvalds mss = full_space; 23841da177e4SLinus Torvalds 2385b92edbe0SEric Dumazet if (free_space < (full_space >> 1)) { 2386463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.quick = 0; 23871da177e4SLinus Torvalds 2388b8da51ebSEric Dumazet if (tcp_under_memory_pressure(sk)) 2389056834d9SIlpo Järvinen tp->rcv_ssthresh = min(tp->rcv_ssthresh, 2390056834d9SIlpo Järvinen 4U * tp->advmss); 23911da177e4SLinus Torvalds 239286c1a045SFlorian Westphal /* free_space might become our new window, make sure we don't 239386c1a045SFlorian Westphal * increase it due to wscale. 239486c1a045SFlorian Westphal */ 239586c1a045SFlorian Westphal free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale); 239686c1a045SFlorian Westphal 239786c1a045SFlorian Westphal /* if free space is less than mss estimate, or is below 1/16th 239886c1a045SFlorian Westphal * of the maximum allowed, try to move to zero-window, else 239986c1a045SFlorian Westphal * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and 240086c1a045SFlorian Westphal * new incoming data is dropped due to memory limits. 240186c1a045SFlorian Westphal * With large window, mss test triggers way too late in order 240286c1a045SFlorian Westphal * to announce zero window in time before rmem limit kicks in. 240386c1a045SFlorian Westphal */ 240486c1a045SFlorian Westphal if (free_space < (allowed_space >> 4) || free_space < mss) 24051da177e4SLinus Torvalds return 0; 24061da177e4SLinus Torvalds } 24071da177e4SLinus Torvalds 24081da177e4SLinus Torvalds if (free_space > tp->rcv_ssthresh) 24091da177e4SLinus Torvalds free_space = tp->rcv_ssthresh; 24101da177e4SLinus Torvalds 24111da177e4SLinus Torvalds /* Don't do rounding if we are using window scaling, since the 24121da177e4SLinus Torvalds * scaled window will not line up with the MSS boundary anyway. 24131da177e4SLinus Torvalds */ 24141da177e4SLinus Torvalds window = tp->rcv_wnd; 24151da177e4SLinus Torvalds if (tp->rx_opt.rcv_wscale) { 24161da177e4SLinus Torvalds window = free_space; 24171da177e4SLinus Torvalds 24181da177e4SLinus Torvalds /* Advertise enough space so that it won't get scaled away. 24191da177e4SLinus Torvalds * Import case: prevent zero window announcement if 24201da177e4SLinus Torvalds * 1<<rcv_wscale > mss. 24211da177e4SLinus Torvalds */ 24221da177e4SLinus Torvalds if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window) 24231da177e4SLinus Torvalds window = (((window >> tp->rx_opt.rcv_wscale) + 1) 24241da177e4SLinus Torvalds << tp->rx_opt.rcv_wscale); 24251da177e4SLinus Torvalds } else { 24261da177e4SLinus Torvalds /* Get the largest window that is a nice multiple of mss. 24271da177e4SLinus Torvalds * Window clamp already applied above. 24281da177e4SLinus Torvalds * If our current window offering is within 1 mss of the 24291da177e4SLinus Torvalds * free space we just keep it. This prevents the divide 24301da177e4SLinus Torvalds * and multiply from happening most of the time. 24311da177e4SLinus Torvalds * We also don't do any window rounding when the free space 24321da177e4SLinus Torvalds * is too small. 24331da177e4SLinus Torvalds */ 24341da177e4SLinus Torvalds if (window <= free_space - mss || window > free_space) 24351da177e4SLinus Torvalds window = (free_space / mss) * mss; 243684565070SJohn Heffner else if (mss == full_space && 2437b92edbe0SEric Dumazet free_space > window + (full_space >> 1)) 243884565070SJohn Heffner window = free_space; 24391da177e4SLinus Torvalds } 24401da177e4SLinus Torvalds 24411da177e4SLinus Torvalds return window; 24421da177e4SLinus Torvalds } 24431da177e4SLinus Torvalds 2444*082ac2d5SMartin KaFai Lau static void tcp_skb_collapse_tstamp(struct sk_buff *skb, 2445*082ac2d5SMartin KaFai Lau const struct sk_buff *next_skb) 2446*082ac2d5SMartin KaFai Lau { 2447*082ac2d5SMartin KaFai Lau const struct skb_shared_info *next_shinfo = skb_shinfo(next_skb); 2448*082ac2d5SMartin KaFai Lau u8 tsflags = next_shinfo->tx_flags & SKBTX_ANY_TSTAMP; 2449*082ac2d5SMartin KaFai Lau 2450*082ac2d5SMartin KaFai Lau if (unlikely(tsflags)) { 2451*082ac2d5SMartin KaFai Lau struct skb_shared_info *shinfo = skb_shinfo(skb); 2452*082ac2d5SMartin KaFai Lau 2453*082ac2d5SMartin KaFai Lau shinfo->tx_flags |= tsflags; 2454*082ac2d5SMartin KaFai Lau shinfo->tskey = next_shinfo->tskey; 2455*082ac2d5SMartin KaFai Lau } 2456*082ac2d5SMartin KaFai Lau } 2457*082ac2d5SMartin KaFai Lau 24584a17fc3aSIlpo Järvinen /* Collapses two adjacent SKB's during retransmission. */ 24594a17fc3aSIlpo Järvinen static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) 24601da177e4SLinus Torvalds { 24611da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2462fe067e8aSDavid S. Miller struct sk_buff *next_skb = tcp_write_queue_next(sk, skb); 2463058dc334SIlpo Järvinen int skb_size, next_skb_size; 24641da177e4SLinus Torvalds 2465058dc334SIlpo Järvinen skb_size = skb->len; 2466058dc334SIlpo Järvinen next_skb_size = next_skb->len; 24671da177e4SLinus Torvalds 2468058dc334SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); 24691da177e4SLinus Torvalds 24706859d494SIlpo Järvinen tcp_highest_sack_combine(sk, next_skb, skb); 2471a6963a6bSIlpo Järvinen 2472fe067e8aSDavid S. Miller tcp_unlink_write_queue(next_skb, sk); 24731da177e4SLinus Torvalds 2474058dc334SIlpo Järvinen skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size), 24751a4e2d09SArnaldo Carvalho de Melo next_skb_size); 24761da177e4SLinus Torvalds 247752d570aaSJarek Poplawski if (next_skb->ip_summed == CHECKSUM_PARTIAL) 247852d570aaSJarek Poplawski skb->ip_summed = CHECKSUM_PARTIAL; 24791da177e4SLinus Torvalds 248084fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 24811da177e4SLinus Torvalds skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); 24821da177e4SLinus Torvalds 24831da177e4SLinus Torvalds /* Update sequence range on original skb. */ 24841da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 24851da177e4SLinus Torvalds 2486e6c7d085SIlpo Järvinen /* Merge over control information. This moves PSH/FIN etc. over */ 24874de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; 24881da177e4SLinus Torvalds 24891da177e4SLinus Torvalds /* All done, get rid of second SKB and account for it so 24901da177e4SLinus Torvalds * packet counting does not break. 24911da177e4SLinus Torvalds */ 24924828e7f4SIlpo Järvinen TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; 2493b7689205SIlpo Järvinen 2494b7689205SIlpo Järvinen /* changed transmit queue under us so clear hints */ 2495ef9da47cSIlpo Järvinen tcp_clear_retrans_hints_partial(tp); 2496ef9da47cSIlpo Järvinen if (next_skb == tp->retransmit_skb_hint) 2497ef9da47cSIlpo Järvinen tp->retransmit_skb_hint = skb; 2498b7689205SIlpo Järvinen 2499797108d1SIlpo Järvinen tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); 2500797108d1SIlpo Järvinen 2501*082ac2d5SMartin KaFai Lau tcp_skb_collapse_tstamp(skb, next_skb); 2502*082ac2d5SMartin KaFai Lau 25033ab224beSHideo Aoki sk_wmem_free_skb(sk, next_skb); 25041da177e4SLinus Torvalds } 25051da177e4SLinus Torvalds 250667edfef7SAndi Kleen /* Check if coalescing SKBs is legal. */ 2507a2a385d6SEric Dumazet static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) 25084a17fc3aSIlpo Järvinen { 25094a17fc3aSIlpo Järvinen if (tcp_skb_pcount(skb) > 1) 2510a2a385d6SEric Dumazet return false; 25114a17fc3aSIlpo Järvinen /* TODO: SACK collapsing could be used to remove this condition */ 25124a17fc3aSIlpo Järvinen if (skb_shinfo(skb)->nr_frags != 0) 2513a2a385d6SEric Dumazet return false; 25144a17fc3aSIlpo Järvinen if (skb_cloned(skb)) 2515a2a385d6SEric Dumazet return false; 25164a17fc3aSIlpo Järvinen if (skb == tcp_send_head(sk)) 2517a2a385d6SEric Dumazet return false; 25184a17fc3aSIlpo Järvinen /* Some heurestics for collapsing over SACK'd could be invented */ 25194a17fc3aSIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 2520a2a385d6SEric Dumazet return false; 25214a17fc3aSIlpo Järvinen 2522a2a385d6SEric Dumazet return true; 25234a17fc3aSIlpo Järvinen } 25244a17fc3aSIlpo Järvinen 252567edfef7SAndi Kleen /* Collapse packets in the retransmit queue to make to create 252667edfef7SAndi Kleen * less packets on the wire. This is only done on retransmission. 252767edfef7SAndi Kleen */ 25284a17fc3aSIlpo Järvinen static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, 25294a17fc3aSIlpo Järvinen int space) 25304a17fc3aSIlpo Järvinen { 25314a17fc3aSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 25324a17fc3aSIlpo Järvinen struct sk_buff *skb = to, *tmp; 2533a2a385d6SEric Dumazet bool first = true; 25344a17fc3aSIlpo Järvinen 25354a17fc3aSIlpo Järvinen if (!sysctl_tcp_retrans_collapse) 25364a17fc3aSIlpo Järvinen return; 25374de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 25384a17fc3aSIlpo Järvinen return; 25394a17fc3aSIlpo Järvinen 25404a17fc3aSIlpo Järvinen tcp_for_write_queue_from_safe(skb, tmp, sk) { 25414a17fc3aSIlpo Järvinen if (!tcp_can_collapse(sk, skb)) 25424a17fc3aSIlpo Järvinen break; 25434a17fc3aSIlpo Järvinen 25444a17fc3aSIlpo Järvinen space -= skb->len; 25454a17fc3aSIlpo Järvinen 25464a17fc3aSIlpo Järvinen if (first) { 2547a2a385d6SEric Dumazet first = false; 25484a17fc3aSIlpo Järvinen continue; 25494a17fc3aSIlpo Järvinen } 25504a17fc3aSIlpo Järvinen 25514a17fc3aSIlpo Järvinen if (space < 0) 25524a17fc3aSIlpo Järvinen break; 25534a17fc3aSIlpo Järvinen /* Punt if not enough space exists in the first SKB for 25544a17fc3aSIlpo Järvinen * the data in the second 25554a17fc3aSIlpo Järvinen */ 2556a21d4572SEric Dumazet if (skb->len > skb_availroom(to)) 25574a17fc3aSIlpo Järvinen break; 25584a17fc3aSIlpo Järvinen 25594a17fc3aSIlpo Järvinen if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) 25604a17fc3aSIlpo Järvinen break; 25614a17fc3aSIlpo Järvinen 25624a17fc3aSIlpo Järvinen tcp_collapse_retrans(sk, to); 25634a17fc3aSIlpo Järvinen } 25644a17fc3aSIlpo Järvinen } 25654a17fc3aSIlpo Järvinen 25661da177e4SLinus Torvalds /* This retransmits one SKB. Policy decisions and retransmit queue 25671da177e4SLinus Torvalds * state updates are done by the caller. Returns non-zero if an 25681da177e4SLinus Torvalds * error occurred which prevented the send. 25691da177e4SLinus Torvalds */ 257093b174adSYuchung Cheng int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) 25711da177e4SLinus Torvalds { 25721da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 25735d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 25747d227cd2SSridhar Samudrala unsigned int cur_mss; 2575c84a5711SYuchung Cheng int err; 25761da177e4SLinus Torvalds 25775d424d5aSJohn Heffner /* Inconslusive MTU probe */ 25785d424d5aSJohn Heffner if (icsk->icsk_mtup.probe_size) { 25795d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 25805d424d5aSJohn Heffner } 25815d424d5aSJohn Heffner 25821da177e4SLinus Torvalds /* Do not sent more than we queued. 1/4 is reserved for possible 2583caa20d9aSStephen Hemminger * copying overhead: fragmentation, tunneling, mangling etc. 25841da177e4SLinus Torvalds */ 25851da177e4SLinus Torvalds if (atomic_read(&sk->sk_wmem_alloc) > 25861da177e4SLinus Torvalds min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf)) 25871da177e4SLinus Torvalds return -EAGAIN; 25881da177e4SLinus Torvalds 25891f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 25901f3279aeSEric Dumazet return -EBUSY; 25911f3279aeSEric Dumazet 25921da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 25931da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 25941da177e4SLinus Torvalds BUG(); 25951da177e4SLinus Torvalds if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 25961da177e4SLinus Torvalds return -ENOMEM; 25971da177e4SLinus Torvalds } 25981da177e4SLinus Torvalds 25997d227cd2SSridhar Samudrala if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 26007d227cd2SSridhar Samudrala return -EHOSTUNREACH; /* Routing failure or similar. */ 26017d227cd2SSridhar Samudrala 26020c54b85fSIlpo Järvinen cur_mss = tcp_current_mss(sk); 26037d227cd2SSridhar Samudrala 26041da177e4SLinus Torvalds /* If receiver has shrunk his window, and skb is out of 26051da177e4SLinus Torvalds * new window, do not retransmit it. The exception is the 26061da177e4SLinus Torvalds * case, when window is shrunk to zero. In this case 26071da177e4SLinus Torvalds * our retransmit serves as a zero window probe. 26081da177e4SLinus Torvalds */ 26099d4fb27dSJoe Perches if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && 26109d4fb27dSJoe Perches TCP_SKB_CB(skb)->seq != tp->snd_una) 26111da177e4SLinus Torvalds return -EAGAIN; 26121da177e4SLinus Torvalds 26131da177e4SLinus Torvalds if (skb->len > cur_mss) { 26146cc55e09SOctavian Purdila if (tcp_fragment(sk, skb, cur_mss, cur_mss, GFP_ATOMIC)) 26151da177e4SLinus Torvalds return -ENOMEM; /* We'll try again later. */ 261602276f3cSIlpo Järvinen } else { 26179eb9362eSIlpo Järvinen int oldpcount = tcp_skb_pcount(skb); 26189eb9362eSIlpo Järvinen 26199eb9362eSIlpo Järvinen if (unlikely(oldpcount > 1)) { 2620c52e2421SEric Dumazet if (skb_unclone(skb, GFP_ATOMIC)) 2621c52e2421SEric Dumazet return -ENOMEM; 26225bbb432cSEric Dumazet tcp_init_tso_segs(skb, cur_mss); 26239eb9362eSIlpo Järvinen tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb)); 26249eb9362eSIlpo Järvinen } 26251da177e4SLinus Torvalds } 26261da177e4SLinus Torvalds 262749213555SDaniel Borkmann /* RFC3168, section 6.1.1.1. ECN fallback */ 262849213555SDaniel Borkmann if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN) 262949213555SDaniel Borkmann tcp_ecn_clear_syn(sk, skb); 263049213555SDaniel Borkmann 26311da177e4SLinus Torvalds tcp_retrans_try_collapse(sk, skb, cur_mss); 26321da177e4SLinus Torvalds 26331da177e4SLinus Torvalds /* Make a copy, if the first transmission SKB clone we made 26341da177e4SLinus Torvalds * is still in somebody's hands, else make a clone. 26351da177e4SLinus Torvalds */ 26361da177e4SLinus Torvalds 263750bceae9SThomas Graf /* make sure skb->data is aligned on arches that require it 263850bceae9SThomas Graf * and check if ack-trimming & collapsing extended the headroom 263950bceae9SThomas Graf * beyond what csum_start can cover. 264050bceae9SThomas Graf */ 264150bceae9SThomas Graf if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) || 264250bceae9SThomas Graf skb_headroom(skb) >= 0xFFFF)) { 2643117632e6SEric Dumazet struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER, 2644117632e6SEric Dumazet GFP_ATOMIC); 2645c84a5711SYuchung Cheng err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) : 2646117632e6SEric Dumazet -ENOBUFS; 2647117632e6SEric Dumazet } else { 2648c84a5711SYuchung Cheng err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 2649117632e6SEric Dumazet } 2650c84a5711SYuchung Cheng 2651fc9f3501SEric Dumazet if (likely(!err)) { 2652c84a5711SYuchung Cheng TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS; 2653fc9f3501SEric Dumazet /* Update global TCP statistics. */ 2654fc9f3501SEric Dumazet TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); 2655fc9f3501SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 2656fc9f3501SEric Dumazet NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 2657fc9f3501SEric Dumazet tp->total_retrans++; 2658fc9f3501SEric Dumazet } 2659c84a5711SYuchung Cheng return err; 266093b174adSYuchung Cheng } 266193b174adSYuchung Cheng 266293b174adSYuchung Cheng int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) 266393b174adSYuchung Cheng { 266493b174adSYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 266593b174adSYuchung Cheng int err = __tcp_retransmit_skb(sk, skb); 26661da177e4SLinus Torvalds 26671da177e4SLinus Torvalds if (err == 0) { 26681da177e4SLinus Torvalds #if FASTRETRANS_DEBUG > 0 26691da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 2670e87cc472SJoe Perches net_dbg_ratelimited("retrans_out leaked\n"); 26711da177e4SLinus Torvalds } 26721da177e4SLinus Torvalds #endif 26731da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 26741da177e4SLinus Torvalds tp->retrans_out += tcp_skb_pcount(skb); 26751da177e4SLinus Torvalds 26761da177e4SLinus Torvalds /* Save stamp of the first retransmit. */ 26771da177e4SLinus Torvalds if (!tp->retrans_stamp) 26787faee5c0SEric Dumazet tp->retrans_stamp = tcp_skb_timestamp(skb); 26791da177e4SLinus Torvalds 26801f3279aeSEric Dumazet } else if (err != -EBUSY) { 268124ab6becSYuchung Cheng NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL); 26821da177e4SLinus Torvalds } 26836e08d5e3SYuchung Cheng 26846e08d5e3SYuchung Cheng if (tp->undo_retrans < 0) 26856e08d5e3SYuchung Cheng tp->undo_retrans = 0; 26866e08d5e3SYuchung Cheng tp->undo_retrans += tcp_skb_pcount(skb); 26871da177e4SLinus Torvalds return err; 26881da177e4SLinus Torvalds } 26891da177e4SLinus Torvalds 269067edfef7SAndi Kleen /* Check if we forward retransmits are possible in the current 269167edfef7SAndi Kleen * window/congestion state. 269267edfef7SAndi Kleen */ 2693a2a385d6SEric Dumazet static bool tcp_can_forward_retransmit(struct sock *sk) 2694b5afe7bcSIlpo Järvinen { 2695b5afe7bcSIlpo Järvinen const struct inet_connection_sock *icsk = inet_csk(sk); 2696cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 2697b5afe7bcSIlpo Järvinen 2698b5afe7bcSIlpo Järvinen /* Forward retransmissions are possible only during Recovery. */ 2699b5afe7bcSIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Recovery) 2700a2a385d6SEric Dumazet return false; 2701b5afe7bcSIlpo Järvinen 2702b5afe7bcSIlpo Järvinen /* No forward retransmissions in Reno are possible. */ 2703b5afe7bcSIlpo Järvinen if (tcp_is_reno(tp)) 2704a2a385d6SEric Dumazet return false; 2705b5afe7bcSIlpo Järvinen 2706b5afe7bcSIlpo Järvinen /* Yeah, we have to make difficult choice between forward transmission 2707b5afe7bcSIlpo Järvinen * and retransmission... Both ways have their merits... 2708b5afe7bcSIlpo Järvinen * 2709b5afe7bcSIlpo Järvinen * For now we do not retransmit anything, while we have some new 2710b5afe7bcSIlpo Järvinen * segments to send. In the other cases, follow rule 3 for 2711b5afe7bcSIlpo Järvinen * NextSeg() specified in RFC3517. 2712b5afe7bcSIlpo Järvinen */ 2713b5afe7bcSIlpo Järvinen 2714b5afe7bcSIlpo Järvinen if (tcp_may_send_now(sk)) 2715a2a385d6SEric Dumazet return false; 2716b5afe7bcSIlpo Järvinen 2717a2a385d6SEric Dumazet return true; 2718b5afe7bcSIlpo Järvinen } 2719b5afe7bcSIlpo Järvinen 27201da177e4SLinus Torvalds /* This gets called after a retransmit timeout, and the initially 27211da177e4SLinus Torvalds * retransmitted data is acknowledged. It tries to continue 27221da177e4SLinus Torvalds * resending the rest of the retransmit queue, until either 27231da177e4SLinus Torvalds * we've sent it all or the congestion window limit is reached. 27241da177e4SLinus Torvalds * If doing SACK, the first ACK which comes back for a timeout 27251da177e4SLinus Torvalds * based retransmit packet might feed us FACK information again. 27261da177e4SLinus Torvalds * If so, we use it to avoid unnecessarily retransmissions. 27271da177e4SLinus Torvalds */ 27281da177e4SLinus Torvalds void tcp_xmit_retransmit_queue(struct sock *sk) 27291da177e4SLinus Torvalds { 27306687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 27311da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 27321da177e4SLinus Torvalds struct sk_buff *skb; 27330e1c54c2SIlpo Järvinen struct sk_buff *hole = NULL; 2734618d9f25SIlpo Järvinen u32 last_lost; 273561eb55f4SIlpo Järvinen int mib_idx; 27360e1c54c2SIlpo Järvinen int fwd_rexmitting = 0; 27376a438bbeSStephen Hemminger 273845e77d31SIlpo Järvinen if (!tp->packets_out) 273945e77d31SIlpo Järvinen return; 274045e77d31SIlpo Järvinen 274108ebd172SIlpo Järvinen if (!tp->lost_out) 274208ebd172SIlpo Järvinen tp->retransmit_high = tp->snd_una; 274308ebd172SIlpo Järvinen 2744618d9f25SIlpo Järvinen if (tp->retransmit_skb_hint) { 27456a438bbeSStephen Hemminger skb = tp->retransmit_skb_hint; 2746618d9f25SIlpo Järvinen last_lost = TCP_SKB_CB(skb)->end_seq; 2747618d9f25SIlpo Järvinen if (after(last_lost, tp->retransmit_high)) 2748618d9f25SIlpo Järvinen last_lost = tp->retransmit_high; 2749618d9f25SIlpo Järvinen } else { 2750fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 2751618d9f25SIlpo Järvinen last_lost = tp->snd_una; 2752618d9f25SIlpo Järvinen } 27531da177e4SLinus Torvalds 2754fe067e8aSDavid S. Miller tcp_for_write_queue_from(skb, sk) { 27551da177e4SLinus Torvalds __u8 sacked = TCP_SKB_CB(skb)->sacked; 27561da177e4SLinus Torvalds 2757fe067e8aSDavid S. Miller if (skb == tcp_send_head(sk)) 2758fe067e8aSDavid S. Miller break; 27596a438bbeSStephen Hemminger /* we could do better than to assign each time */ 276051456b29SIan Morris if (!hole) 27616a438bbeSStephen Hemminger tp->retransmit_skb_hint = skb; 27626a438bbeSStephen Hemminger 27631da177e4SLinus Torvalds /* Assume this retransmit will generate 27641da177e4SLinus Torvalds * only one packet for congestion window 27651da177e4SLinus Torvalds * calculation purposes. This works because 27661da177e4SLinus Torvalds * tcp_retransmit_skb() will chop up the 27671da177e4SLinus Torvalds * packet to be MSS sized and all the 27681da177e4SLinus Torvalds * packet counting works out. 27691da177e4SLinus Torvalds */ 27701da177e4SLinus Torvalds if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) 27711da177e4SLinus Torvalds return; 27720e1c54c2SIlpo Järvinen 27730e1c54c2SIlpo Järvinen if (fwd_rexmitting) { 27740e1c54c2SIlpo Järvinen begin_fwd: 27750e1c54c2SIlpo Järvinen if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp))) 2776006f582cSIlpo Järvinen break; 27770e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFORWARDRETRANS; 27780e1c54c2SIlpo Järvinen 27790e1c54c2SIlpo Järvinen } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) { 2780618d9f25SIlpo Järvinen tp->retransmit_high = last_lost; 27810e1c54c2SIlpo Järvinen if (!tcp_can_forward_retransmit(sk)) 27820e1c54c2SIlpo Järvinen break; 27830e1c54c2SIlpo Järvinen /* Backtrack if necessary to non-L'ed skb */ 278400db4124SIan Morris if (hole) { 27850e1c54c2SIlpo Järvinen skb = hole; 27860e1c54c2SIlpo Järvinen hole = NULL; 27870e1c54c2SIlpo Järvinen } 27880e1c54c2SIlpo Järvinen fwd_rexmitting = 1; 27890e1c54c2SIlpo Järvinen goto begin_fwd; 27900e1c54c2SIlpo Järvinen 27910e1c54c2SIlpo Järvinen } else if (!(sacked & TCPCB_LOST)) { 279251456b29SIan Morris if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) 27930e1c54c2SIlpo Järvinen hole = skb; 279461eb55f4SIlpo Järvinen continue; 27951da177e4SLinus Torvalds 27960e1c54c2SIlpo Järvinen } else { 2797618d9f25SIlpo Järvinen last_lost = TCP_SKB_CB(skb)->end_seq; 27980e1c54c2SIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Loss) 27990e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFASTRETRANS; 28000e1c54c2SIlpo Järvinen else 28010e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; 28020e1c54c2SIlpo Järvinen } 28030e1c54c2SIlpo Järvinen 28040e1c54c2SIlpo Järvinen if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) 280561eb55f4SIlpo Järvinen continue; 280640b215e5SPavel Emelyanov 280724ab6becSYuchung Cheng if (tcp_retransmit_skb(sk, skb)) 28081da177e4SLinus Torvalds return; 280924ab6becSYuchung Cheng 2810de0744afSPavel Emelyanov NET_INC_STATS_BH(sock_net(sk), mib_idx); 28111da177e4SLinus Torvalds 2812684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 2813a262f0cdSNandita Dukkipati tp->prr_out += tcp_skb_pcount(skb); 2814a262f0cdSNandita Dukkipati 2815fe067e8aSDavid S. Miller if (skb == tcp_write_queue_head(sk)) 2816463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 28173f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, 28183f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 28191da177e4SLinus Torvalds } 28201da177e4SLinus Torvalds } 28211da177e4SLinus Torvalds 2822d83769a5SEric Dumazet /* We allow to exceed memory limits for FIN packets to expedite 2823d83769a5SEric Dumazet * connection tear down and (memory) recovery. 2824845704a5SEric Dumazet * Otherwise tcp_send_fin() could be tempted to either delay FIN 2825845704a5SEric Dumazet * or even be forced to close flow without any FIN. 2826a6c5ea4cSEric Dumazet * In general, we want to allow one skb per socket to avoid hangs 2827a6c5ea4cSEric Dumazet * with edge trigger epoll() 2828d83769a5SEric Dumazet */ 2829a6c5ea4cSEric Dumazet void sk_forced_mem_schedule(struct sock *sk, int size) 2830d83769a5SEric Dumazet { 2831e805605cSJohannes Weiner int amt; 2832d83769a5SEric Dumazet 2833d83769a5SEric Dumazet if (size <= sk->sk_forward_alloc) 2834d83769a5SEric Dumazet return; 2835d83769a5SEric Dumazet amt = sk_mem_pages(size); 2836d83769a5SEric Dumazet sk->sk_forward_alloc += amt * SK_MEM_QUANTUM; 2837e805605cSJohannes Weiner sk_memory_allocated_add(sk, amt); 2838e805605cSJohannes Weiner 2839baac50bbSJohannes Weiner if (mem_cgroup_sockets_enabled && sk->sk_memcg) 2840baac50bbSJohannes Weiner mem_cgroup_charge_skmem(sk->sk_memcg, amt); 2841d83769a5SEric Dumazet } 2842d83769a5SEric Dumazet 2843845704a5SEric Dumazet /* Send a FIN. The caller locks the socket for us. 2844845704a5SEric Dumazet * We should try to send a FIN packet really hard, but eventually give up. 28451da177e4SLinus Torvalds */ 28461da177e4SLinus Torvalds void tcp_send_fin(struct sock *sk) 28471da177e4SLinus Torvalds { 2848845704a5SEric Dumazet struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk); 28491da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 28501da177e4SLinus Torvalds 2851845704a5SEric Dumazet /* Optimization, tack on the FIN if we have one skb in write queue and 2852845704a5SEric Dumazet * this skb was not yet sent, or we are under memory pressure. 2853845704a5SEric Dumazet * Note: in the latter case, FIN packet will be sent after a timeout, 2854845704a5SEric Dumazet * as TCP stack thinks it has already been transmitted. 28551da177e4SLinus Torvalds */ 2856b8da51ebSEric Dumazet if (tskb && (tcp_send_head(sk) || tcp_under_memory_pressure(sk))) { 2857845704a5SEric Dumazet coalesce: 2858845704a5SEric Dumazet TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN; 2859845704a5SEric Dumazet TCP_SKB_CB(tskb)->end_seq++; 28601da177e4SLinus Torvalds tp->write_seq++; 2861845704a5SEric Dumazet if (!tcp_send_head(sk)) { 2862845704a5SEric Dumazet /* This means tskb was already sent. 2863845704a5SEric Dumazet * Pretend we included the FIN on previous transmit. 2864845704a5SEric Dumazet * We need to set tp->snd_nxt to the value it would have 2865845704a5SEric Dumazet * if FIN had been sent. This is because retransmit path 2866845704a5SEric Dumazet * does not change tp->snd_nxt. 2867845704a5SEric Dumazet */ 2868845704a5SEric Dumazet tp->snd_nxt++; 2869845704a5SEric Dumazet return; 2870845704a5SEric Dumazet } 28711da177e4SLinus Torvalds } else { 2872845704a5SEric Dumazet skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation); 2873845704a5SEric Dumazet if (unlikely(!skb)) { 2874845704a5SEric Dumazet if (tskb) 2875845704a5SEric Dumazet goto coalesce; 2876845704a5SEric Dumazet return; 28771da177e4SLinus Torvalds } 2878d83769a5SEric Dumazet skb_reserve(skb, MAX_TCP_HEADER); 2879a6c5ea4cSEric Dumazet sk_forced_mem_schedule(sk, skb->truesize); 28801da177e4SLinus Torvalds /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 2881e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tp->write_seq, 2882a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_FIN); 28831da177e4SLinus Torvalds tcp_queue_skb(sk, skb); 28841da177e4SLinus Torvalds } 2885845704a5SEric Dumazet __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF); 28861da177e4SLinus Torvalds } 28871da177e4SLinus Torvalds 28881da177e4SLinus Torvalds /* We get here when a process closes a file descriptor (either due to 28891da177e4SLinus Torvalds * an explicit close() or as a byproduct of exit()'ing) and there 28901da177e4SLinus Torvalds * was unread data in the receive queue. This behavior is recommended 289165bb723cSGerrit Renker * by RFC 2525, section 2.17. -DaveM 28921da177e4SLinus Torvalds */ 2893dd0fc66fSAl Viro void tcp_send_active_reset(struct sock *sk, gfp_t priority) 28941da177e4SLinus Torvalds { 28951da177e4SLinus Torvalds struct sk_buff *skb; 28961da177e4SLinus Torvalds 28971da177e4SLinus Torvalds /* NOTE: No TCP options attached and we never retransmit this. */ 28981da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, priority); 28991da177e4SLinus Torvalds if (!skb) { 29004e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 29011da177e4SLinus Torvalds return; 29021da177e4SLinus Torvalds } 29031da177e4SLinus Torvalds 29041da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 29051da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 2906e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), 2907a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_RST); 2908675ee231SEric Dumazet skb_mstamp_get(&skb->skb_mstamp); 29091da177e4SLinus Torvalds /* Send it off. */ 2910dfb4b9dcSDavid S. Miller if (tcp_transmit_skb(sk, skb, 0, priority)) 29114e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 291226af65cbSSridhar Samudrala 291381cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); 29141da177e4SLinus Torvalds } 29151da177e4SLinus Torvalds 291667edfef7SAndi Kleen /* Send a crossed SYN-ACK during socket establishment. 291767edfef7SAndi Kleen * WARNING: This routine must only be called when we have already sent 29181da177e4SLinus Torvalds * a SYN packet that crossed the incoming SYN that caused this routine 29191da177e4SLinus Torvalds * to get called. If this assumption fails then the initial rcv_wnd 29201da177e4SLinus Torvalds * and rcv_wscale values will not be correct. 29211da177e4SLinus Torvalds */ 29221da177e4SLinus Torvalds int tcp_send_synack(struct sock *sk) 29231da177e4SLinus Torvalds { 29241da177e4SLinus Torvalds struct sk_buff *skb; 29251da177e4SLinus Torvalds 2926fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 292751456b29SIan Morris if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 292891df42beSJoe Perches pr_debug("%s: wrong queue state\n", __func__); 29291da177e4SLinus Torvalds return -EFAULT; 29301da177e4SLinus Torvalds } 29314de075e0SEric Dumazet if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { 29321da177e4SLinus Torvalds if (skb_cloned(skb)) { 29331da177e4SLinus Torvalds struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC); 293451456b29SIan Morris if (!nskb) 29351da177e4SLinus Torvalds return -ENOMEM; 2936fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 2937f4a775d1SEric Dumazet __skb_header_release(nskb); 2938fe067e8aSDavid S. Miller __tcp_add_write_queue_head(sk, nskb); 29393ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 29403ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 29413ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 29421da177e4SLinus Torvalds skb = nskb; 29431da177e4SLinus Torvalds } 29441da177e4SLinus Torvalds 29454de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; 2946735d3831SFlorian Westphal tcp_ecn_send_synack(sk, skb); 29471da177e4SLinus Torvalds } 2948dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 29491da177e4SLinus Torvalds } 29501da177e4SLinus Torvalds 29514aea39c1SEric Dumazet /** 29524aea39c1SEric Dumazet * tcp_make_synack - Prepare a SYN-ACK. 29534aea39c1SEric Dumazet * sk: listener socket 29544aea39c1SEric Dumazet * dst: dst entry attached to the SYNACK 29554aea39c1SEric Dumazet * req: request_sock pointer 29564aea39c1SEric Dumazet * 29574aea39c1SEric Dumazet * Allocate one skb and build a SYNACK packet. 29584aea39c1SEric Dumazet * @dst is consumed : Caller should not use it again. 29594aea39c1SEric Dumazet */ 29605d062de7SEric Dumazet struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst, 2961e6b4d113SWilliam Allen Simpson struct request_sock *req, 2962ca6fb065SEric Dumazet struct tcp_fastopen_cookie *foc, 2963ca6fb065SEric Dumazet bool attach_req) 29641da177e4SLinus Torvalds { 29652e6599cbSArnaldo Carvalho de Melo struct inet_request_sock *ireq = inet_rsk(req); 29665d062de7SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 296780f03e27SEric Dumazet struct tcp_md5sig_key *md5 = NULL; 29685d062de7SEric Dumazet struct tcp_out_options opts; 29695d062de7SEric Dumazet struct sk_buff *skb; 2970bd0388aeSWilliam Allen Simpson int tcp_header_size; 29715d062de7SEric Dumazet struct tcphdr *th; 29725d062de7SEric Dumazet u16 user_mss; 2973f5fff5dcSTom Quetchenbach int mss; 29741da177e4SLinus Torvalds 2975ca6fb065SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 29764aea39c1SEric Dumazet if (unlikely(!skb)) { 29774aea39c1SEric Dumazet dst_release(dst); 29781da177e4SLinus Torvalds return NULL; 29794aea39c1SEric Dumazet } 29801da177e4SLinus Torvalds /* Reserve space for headers. */ 29811da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 29821da177e4SLinus Torvalds 2983ca6fb065SEric Dumazet if (attach_req) { 29849e17f8a4SEric Dumazet skb_set_owner_w(skb, req_to_sk(req)); 2985ca6fb065SEric Dumazet } else { 2986ca6fb065SEric Dumazet /* sk is a const pointer, because we want to express multiple 2987ca6fb065SEric Dumazet * cpu might call us concurrently. 2988ca6fb065SEric Dumazet * sk->sk_wmem_alloc in an atomic, we can promote to rw. 2989ca6fb065SEric Dumazet */ 2990ca6fb065SEric Dumazet skb_set_owner_w(skb, (struct sock *)sk); 2991ca6fb065SEric Dumazet } 29924aea39c1SEric Dumazet skb_dst_set(skb, dst); 29931da177e4SLinus Torvalds 29940dbaee3bSDavid S. Miller mss = dst_metric_advmss(dst); 29955d062de7SEric Dumazet user_mss = READ_ONCE(tp->rx_opt.user_mss); 29965d062de7SEric Dumazet if (user_mss && user_mss < mss) 29975d062de7SEric Dumazet mss = user_mss; 2998f5fff5dcSTom Quetchenbach 299933ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 30008b5f12d0SFlorian Westphal #ifdef CONFIG_SYN_COOKIES 30018b5f12d0SFlorian Westphal if (unlikely(req->cookie_ts)) 30027faee5c0SEric Dumazet skb->skb_mstamp.stamp_jiffies = cookie_init_timestamp(req); 30038b5f12d0SFlorian Westphal else 30048b5f12d0SFlorian Westphal #endif 30057faee5c0SEric Dumazet skb_mstamp_get(&skb->skb_mstamp); 300680f03e27SEric Dumazet 300780f03e27SEric Dumazet #ifdef CONFIG_TCP_MD5SIG 300880f03e27SEric Dumazet rcu_read_lock(); 3009fd3a154aSEric Dumazet md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req)); 301080f03e27SEric Dumazet #endif 301158d607d3SEric Dumazet skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4); 301237bfbddaSEric Dumazet tcp_header_size = tcp_synack_options(req, mss, skb, &opts, md5, foc) + 301337bfbddaSEric Dumazet sizeof(*th); 301433ad798cSAdam Langley 3015aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 3016aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 30171da177e4SLinus Torvalds 3018aa8223c7SArnaldo Carvalho de Melo th = tcp_hdr(skb); 30191da177e4SLinus Torvalds memset(th, 0, sizeof(struct tcphdr)); 30201da177e4SLinus Torvalds th->syn = 1; 30211da177e4SLinus Torvalds th->ack = 1; 30226ac705b1SEric Dumazet tcp_ecn_make_synack(req, th); 3023b44084c2SEric Dumazet th->source = htons(ireq->ir_num); 3024634fb979SEric Dumazet th->dest = ireq->ir_rmt_port; 3025e870a8efSIlpo Järvinen /* Setting of flags are superfluous here for callers (and ECE is 3026e870a8efSIlpo Järvinen * not even correctly set) 3027e870a8efSIlpo Järvinen */ 3028e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn, 3029a3433f35SChangli Gao TCPHDR_SYN | TCPHDR_ACK); 30304957faadSWilliam Allen Simpson 30311da177e4SLinus Torvalds th->seq = htonl(TCP_SKB_CB(skb)->seq); 30328336886fSJerry Chu /* XXX data is queued and acked as is. No buffer/window check */ 30338336886fSJerry Chu th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt); 30341da177e4SLinus Torvalds 30351da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 3036ed53d0abSEric Dumazet th->window = htons(min(req->rsk_rcv_wnd, 65535U)); 30375d062de7SEric Dumazet tcp_options_write((__be32 *)(th + 1), NULL, &opts); 30381da177e4SLinus Torvalds th->doff = (tcp_header_size >> 2); 3039a0b8486cSEric Dumazet TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_OUTSEGS); 3040cfb6eeb4SYOSHIFUJI Hideaki 3041cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 3042cfb6eeb4SYOSHIFUJI Hideaki /* Okay, we have all we need - do the md5 hash if needed */ 304380f03e27SEric Dumazet if (md5) 3044bd0388aeSWilliam Allen Simpson tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, 304539f8e58eSEric Dumazet md5, req_to_sk(req), skb); 304680f03e27SEric Dumazet rcu_read_unlock(); 3047cfb6eeb4SYOSHIFUJI Hideaki #endif 3048cfb6eeb4SYOSHIFUJI Hideaki 3049b50edd78SEric Dumazet /* Do not fool tcpdump (if any), clean our debris */ 3050b50edd78SEric Dumazet skb->tstamp.tv64 = 0; 30511da177e4SLinus Torvalds return skb; 30521da177e4SLinus Torvalds } 30534bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_make_synack); 30541da177e4SLinus Torvalds 305581164413SDaniel Borkmann static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst) 305681164413SDaniel Borkmann { 305781164413SDaniel Borkmann struct inet_connection_sock *icsk = inet_csk(sk); 305881164413SDaniel Borkmann const struct tcp_congestion_ops *ca; 305981164413SDaniel Borkmann u32 ca_key = dst_metric(dst, RTAX_CC_ALGO); 306081164413SDaniel Borkmann 306181164413SDaniel Borkmann if (ca_key == TCP_CA_UNSPEC) 306281164413SDaniel Borkmann return; 306381164413SDaniel Borkmann 306481164413SDaniel Borkmann rcu_read_lock(); 306581164413SDaniel Borkmann ca = tcp_ca_find_key(ca_key); 306681164413SDaniel Borkmann if (likely(ca && try_module_get(ca->owner))) { 306781164413SDaniel Borkmann module_put(icsk->icsk_ca_ops->owner); 306881164413SDaniel Borkmann icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst); 306981164413SDaniel Borkmann icsk->icsk_ca_ops = ca; 307081164413SDaniel Borkmann } 307181164413SDaniel Borkmann rcu_read_unlock(); 307281164413SDaniel Borkmann } 307381164413SDaniel Borkmann 307467edfef7SAndi Kleen /* Do all connect socket setups that can be done AF independent. */ 3075f7e56a76Sstephen hemminger static void tcp_connect_init(struct sock *sk) 30761da177e4SLinus Torvalds { 3077cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 30781da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 30791da177e4SLinus Torvalds __u8 rcv_wscale; 30801da177e4SLinus Torvalds 30811da177e4SLinus Torvalds /* We'll fix this up when we get a response from the other end. 30821da177e4SLinus Torvalds * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 30831da177e4SLinus Torvalds */ 30841da177e4SLinus Torvalds tp->tcp_header_len = sizeof(struct tcphdr) + 3085bb5b7c11SDavid S. Miller (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0); 30861da177e4SLinus Torvalds 3087cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 308800db4124SIan Morris if (tp->af_specific->md5_lookup(sk, sk)) 3089cfb6eeb4SYOSHIFUJI Hideaki tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 3090cfb6eeb4SYOSHIFUJI Hideaki #endif 3091cfb6eeb4SYOSHIFUJI Hideaki 30921da177e4SLinus Torvalds /* If user gave his TCP_MAXSEG, record it to clamp */ 30931da177e4SLinus Torvalds if (tp->rx_opt.user_mss) 30941da177e4SLinus Torvalds tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 30951da177e4SLinus Torvalds tp->max_window = 0; 30965d424d5aSJohn Heffner tcp_mtup_init(sk); 30971da177e4SLinus Torvalds tcp_sync_mss(sk, dst_mtu(dst)); 30981da177e4SLinus Torvalds 309981164413SDaniel Borkmann tcp_ca_dst_init(sk, dst); 310081164413SDaniel Borkmann 31011da177e4SLinus Torvalds if (!tp->window_clamp) 31021da177e4SLinus Torvalds tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 31030dbaee3bSDavid S. Miller tp->advmss = dst_metric_advmss(dst); 3104f5fff5dcSTom Quetchenbach if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss) 3105f5fff5dcSTom Quetchenbach tp->advmss = tp->rx_opt.user_mss; 3106f5fff5dcSTom Quetchenbach 31071da177e4SLinus Torvalds tcp_initialize_rcv_mss(sk); 31081da177e4SLinus Torvalds 3109e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 3110e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 3111e88c64f0SHagen Paul Pfeifer (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) 3112e88c64f0SHagen Paul Pfeifer tp->window_clamp = tcp_full_space(sk); 3113e88c64f0SHagen Paul Pfeifer 31141da177e4SLinus Torvalds tcp_select_initial_window(tcp_full_space(sk), 31151da177e4SLinus Torvalds tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 31161da177e4SLinus Torvalds &tp->rcv_wnd, 31171da177e4SLinus Torvalds &tp->window_clamp, 3118bb5b7c11SDavid S. Miller sysctl_tcp_window_scaling, 311931d12926Slaurent chavey &rcv_wscale, 312031d12926Slaurent chavey dst_metric(dst, RTAX_INITRWND)); 31211da177e4SLinus Torvalds 31221da177e4SLinus Torvalds tp->rx_opt.rcv_wscale = rcv_wscale; 31231da177e4SLinus Torvalds tp->rcv_ssthresh = tp->rcv_wnd; 31241da177e4SLinus Torvalds 31251da177e4SLinus Torvalds sk->sk_err = 0; 31261da177e4SLinus Torvalds sock_reset_flag(sk, SOCK_DONE); 31271da177e4SLinus Torvalds tp->snd_wnd = 0; 3128ee7537b6SHantzis Fotis tcp_init_wl(tp, 0); 31291da177e4SLinus Torvalds tp->snd_una = tp->write_seq; 31301da177e4SLinus Torvalds tp->snd_sml = tp->write_seq; 313133f5f57eSIlpo Järvinen tp->snd_up = tp->write_seq; 3132370816aeSPavel Emelyanov tp->snd_nxt = tp->write_seq; 3133ee995283SPavel Emelyanov 3134ee995283SPavel Emelyanov if (likely(!tp->repair)) 31351da177e4SLinus Torvalds tp->rcv_nxt = 0; 3136c7781a6eSAndrew Vagin else 3137c7781a6eSAndrew Vagin tp->rcv_tstamp = tcp_time_stamp; 3138ee995283SPavel Emelyanov tp->rcv_wup = tp->rcv_nxt; 3139ee995283SPavel Emelyanov tp->copied_seq = tp->rcv_nxt; 31401da177e4SLinus Torvalds 3141463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; 3142463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_retransmits = 0; 31431da177e4SLinus Torvalds tcp_clear_retrans(tp); 31441da177e4SLinus Torvalds } 31451da177e4SLinus Torvalds 3146783237e8SYuchung Cheng static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) 3147783237e8SYuchung Cheng { 3148783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3149783237e8SYuchung Cheng struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 3150783237e8SYuchung Cheng 3151783237e8SYuchung Cheng tcb->end_seq += skb->len; 3152f4a775d1SEric Dumazet __skb_header_release(skb); 3153783237e8SYuchung Cheng __tcp_add_write_queue_tail(sk, skb); 3154783237e8SYuchung Cheng sk->sk_wmem_queued += skb->truesize; 3155783237e8SYuchung Cheng sk_mem_charge(sk, skb->truesize); 3156783237e8SYuchung Cheng tp->write_seq = tcb->end_seq; 3157783237e8SYuchung Cheng tp->packets_out += tcp_skb_pcount(skb); 3158783237e8SYuchung Cheng } 3159783237e8SYuchung Cheng 3160783237e8SYuchung Cheng /* Build and send a SYN with data and (cached) Fast Open cookie. However, 3161783237e8SYuchung Cheng * queue a data-only packet after the regular SYN, such that regular SYNs 3162783237e8SYuchung Cheng * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges 3163783237e8SYuchung Cheng * only the SYN sequence, the data are retransmitted in the first ACK. 3164783237e8SYuchung Cheng * If cookie is not cached or other error occurs, falls back to send a 3165783237e8SYuchung Cheng * regular SYN with Fast Open cookie request option. 3166783237e8SYuchung Cheng */ 3167783237e8SYuchung Cheng static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) 3168783237e8SYuchung Cheng { 3169783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3170783237e8SYuchung Cheng struct tcp_fastopen_request *fo = tp->fastopen_req; 317107e100f9SEric Dumazet int syn_loss = 0, space, err = 0; 3172aab48743SYuchung Cheng unsigned long last_syn_loss = 0; 3173355a901eSEric Dumazet struct sk_buff *syn_data; 3174783237e8SYuchung Cheng 317567da22d2SYuchung Cheng tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ 3176aab48743SYuchung Cheng tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie, 3177aab48743SYuchung Cheng &syn_loss, &last_syn_loss); 3178aab48743SYuchung Cheng /* Recurring FO SYN losses: revert to regular handshake temporarily */ 3179aab48743SYuchung Cheng if (syn_loss > 1 && 3180aab48743SYuchung Cheng time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) { 3181aab48743SYuchung Cheng fo->cookie.len = -1; 3182aab48743SYuchung Cheng goto fallback; 3183aab48743SYuchung Cheng } 3184aab48743SYuchung Cheng 318567da22d2SYuchung Cheng if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE) 318667da22d2SYuchung Cheng fo->cookie.len = -1; 318767da22d2SYuchung Cheng else if (fo->cookie.len <= 0) 3188783237e8SYuchung Cheng goto fallback; 3189783237e8SYuchung Cheng 3190783237e8SYuchung Cheng /* MSS for SYN-data is based on cached MSS and bounded by PMTU and 3191783237e8SYuchung Cheng * user-MSS. Reserve maximum option space for middleboxes that add 3192783237e8SYuchung Cheng * private TCP options. The cost is reduced data space in SYN :( 3193783237e8SYuchung Cheng */ 3194783237e8SYuchung Cheng if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp) 3195783237e8SYuchung Cheng tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 31961b63edd6SYuchung Cheng space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) - 3197783237e8SYuchung Cheng MAX_TCP_OPTION_SPACE; 3198783237e8SYuchung Cheng 3199f5ddcbbbSEric Dumazet space = min_t(size_t, space, fo->size); 3200f5ddcbbbSEric Dumazet 3201f5ddcbbbSEric Dumazet /* limit to order-0 allocations */ 3202f5ddcbbbSEric Dumazet space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER)); 3203f5ddcbbbSEric Dumazet 3204eb934478SEric Dumazet syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false); 3205355a901eSEric Dumazet if (!syn_data) 3206783237e8SYuchung Cheng goto fallback; 3207355a901eSEric Dumazet syn_data->ip_summed = CHECKSUM_PARTIAL; 3208355a901eSEric Dumazet memcpy(syn_data->cb, syn->cb, sizeof(syn->cb)); 320907e100f9SEric Dumazet if (space) { 321007e100f9SEric Dumazet int copied = copy_from_iter(skb_put(syn_data, space), space, 321157be5bdaSAl Viro &fo->data->msg_iter); 321257be5bdaSAl Viro if (unlikely(!copied)) { 3213355a901eSEric Dumazet kfree_skb(syn_data); 3214783237e8SYuchung Cheng goto fallback; 3215783237e8SYuchung Cheng } 321657be5bdaSAl Viro if (copied != space) { 321757be5bdaSAl Viro skb_trim(syn_data, copied); 321857be5bdaSAl Viro space = copied; 321957be5bdaSAl Viro } 322007e100f9SEric Dumazet } 3221355a901eSEric Dumazet /* No more data pending in inet_wait_for_connect() */ 3222355a901eSEric Dumazet if (space == fo->size) 3223355a901eSEric Dumazet fo->data = NULL; 3224355a901eSEric Dumazet fo->copied = space; 3225783237e8SYuchung Cheng 3226355a901eSEric Dumazet tcp_connect_queue_skb(sk, syn_data); 3227355a901eSEric Dumazet 3228355a901eSEric Dumazet err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation); 3229355a901eSEric Dumazet 3230355a901eSEric Dumazet syn->skb_mstamp = syn_data->skb_mstamp; 3231355a901eSEric Dumazet 3232355a901eSEric Dumazet /* Now full SYN+DATA was cloned and sent (or not), 3233355a901eSEric Dumazet * remove the SYN from the original skb (syn_data) 3234355a901eSEric Dumazet * we keep in write queue in case of a retransmit, as we 3235355a901eSEric Dumazet * also have the SYN packet (with no data) in the same queue. 3236431a9124SEric Dumazet */ 3237355a901eSEric Dumazet TCP_SKB_CB(syn_data)->seq++; 3238355a901eSEric Dumazet TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH; 3239355a901eSEric Dumazet if (!err) { 324067da22d2SYuchung Cheng tp->syn_data = (fo->copied > 0); 3241f19c29e3SYuchung Cheng NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT); 3242783237e8SYuchung Cheng goto done; 3243783237e8SYuchung Cheng } 3244783237e8SYuchung Cheng 3245783237e8SYuchung Cheng fallback: 3246783237e8SYuchung Cheng /* Send a regular SYN with Fast Open cookie request option */ 3247783237e8SYuchung Cheng if (fo->cookie.len > 0) 3248783237e8SYuchung Cheng fo->cookie.len = 0; 3249783237e8SYuchung Cheng err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); 3250783237e8SYuchung Cheng if (err) 3251783237e8SYuchung Cheng tp->syn_fastopen = 0; 3252783237e8SYuchung Cheng done: 3253783237e8SYuchung Cheng fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ 3254783237e8SYuchung Cheng return err; 3255783237e8SYuchung Cheng } 3256783237e8SYuchung Cheng 325767edfef7SAndi Kleen /* Build a SYN and send it off. */ 32581da177e4SLinus Torvalds int tcp_connect(struct sock *sk) 32591da177e4SLinus Torvalds { 32601da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 32611da177e4SLinus Torvalds struct sk_buff *buff; 3262ee586811SEric Paris int err; 32631da177e4SLinus Torvalds 32641da177e4SLinus Torvalds tcp_connect_init(sk); 32651da177e4SLinus Torvalds 32662b916477SAndrey Vagin if (unlikely(tp->repair)) { 32672b916477SAndrey Vagin tcp_finish_connect(sk, NULL); 32682b916477SAndrey Vagin return 0; 32692b916477SAndrey Vagin } 32702b916477SAndrey Vagin 3271eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true); 3272355a901eSEric Dumazet if (unlikely(!buff)) 32731da177e4SLinus Torvalds return -ENOBUFS; 32741da177e4SLinus Torvalds 3275a3433f35SChangli Gao tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); 32767faee5c0SEric Dumazet tp->retrans_stamp = tcp_time_stamp; 3277783237e8SYuchung Cheng tcp_connect_queue_skb(sk, buff); 3278735d3831SFlorian Westphal tcp_ecn_send_syn(sk, buff); 32791da177e4SLinus Torvalds 3280783237e8SYuchung Cheng /* Send off SYN; include data in Fast Open. */ 3281783237e8SYuchung Cheng err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : 3282783237e8SYuchung Cheng tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); 3283ee586811SEric Paris if (err == -ECONNREFUSED) 3284ee586811SEric Paris return err; 3285bd37a088SWei Yongjun 3286bd37a088SWei Yongjun /* We change tp->snd_nxt after the tcp_transmit_skb() call 3287bd37a088SWei Yongjun * in order to make this packet get counted in tcpOutSegs. 3288bd37a088SWei Yongjun */ 3289bd37a088SWei Yongjun tp->snd_nxt = tp->write_seq; 3290bd37a088SWei Yongjun tp->pushed_seq = tp->write_seq; 329181cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); 32921da177e4SLinus Torvalds 32931da177e4SLinus Torvalds /* Timer for repeating the SYN until an answer. */ 32943f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 32953f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 32961da177e4SLinus Torvalds return 0; 32971da177e4SLinus Torvalds } 32984bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_connect); 32991da177e4SLinus Torvalds 33001da177e4SLinus Torvalds /* Send out a delayed ack, the caller does the policy checking 33011da177e4SLinus Torvalds * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 33021da177e4SLinus Torvalds * for details. 33031da177e4SLinus Torvalds */ 33041da177e4SLinus Torvalds void tcp_send_delayed_ack(struct sock *sk) 33051da177e4SLinus Torvalds { 3306463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 3307463c84b9SArnaldo Carvalho de Melo int ato = icsk->icsk_ack.ato; 33081da177e4SLinus Torvalds unsigned long timeout; 33091da177e4SLinus Torvalds 33109890092eSFlorian Westphal tcp_ca_event(sk, CA_EVENT_DELAYED_ACK); 33119890092eSFlorian Westphal 33121da177e4SLinus Torvalds if (ato > TCP_DELACK_MIN) { 3313463c84b9SArnaldo Carvalho de Melo const struct tcp_sock *tp = tcp_sk(sk); 33141da177e4SLinus Torvalds int max_ato = HZ / 2; 33151da177e4SLinus Torvalds 3316056834d9SIlpo Järvinen if (icsk->icsk_ack.pingpong || 3317056834d9SIlpo Järvinen (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 33181da177e4SLinus Torvalds max_ato = TCP_DELACK_MAX; 33191da177e4SLinus Torvalds 33201da177e4SLinus Torvalds /* Slow path, intersegment interval is "high". */ 33211da177e4SLinus Torvalds 33221da177e4SLinus Torvalds /* If some rtt estimate is known, use it to bound delayed ack. 3323463c84b9SArnaldo Carvalho de Melo * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 33241da177e4SLinus Torvalds * directly. 33251da177e4SLinus Torvalds */ 3326740b0f18SEric Dumazet if (tp->srtt_us) { 3327740b0f18SEric Dumazet int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3), 3328740b0f18SEric Dumazet TCP_DELACK_MIN); 33291da177e4SLinus Torvalds 33301da177e4SLinus Torvalds if (rtt < max_ato) 33311da177e4SLinus Torvalds max_ato = rtt; 33321da177e4SLinus Torvalds } 33331da177e4SLinus Torvalds 33341da177e4SLinus Torvalds ato = min(ato, max_ato); 33351da177e4SLinus Torvalds } 33361da177e4SLinus Torvalds 33371da177e4SLinus Torvalds /* Stay within the limit we were given */ 33381da177e4SLinus Torvalds timeout = jiffies + ato; 33391da177e4SLinus Torvalds 33401da177e4SLinus Torvalds /* Use new timeout only if there wasn't a older one earlier. */ 3341463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 33421da177e4SLinus Torvalds /* If delack timer was blocked or is about to expire, 33431da177e4SLinus Torvalds * send ACK now. 33441da177e4SLinus Torvalds */ 3345463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.blocked || 3346463c84b9SArnaldo Carvalho de Melo time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 33471da177e4SLinus Torvalds tcp_send_ack(sk); 33481da177e4SLinus Torvalds return; 33491da177e4SLinus Torvalds } 33501da177e4SLinus Torvalds 3351463c84b9SArnaldo Carvalho de Melo if (!time_before(timeout, icsk->icsk_ack.timeout)) 3352463c84b9SArnaldo Carvalho de Melo timeout = icsk->icsk_ack.timeout; 33531da177e4SLinus Torvalds } 3354463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 3355463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.timeout = timeout; 3356463c84b9SArnaldo Carvalho de Melo sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 33571da177e4SLinus Torvalds } 33581da177e4SLinus Torvalds 33591da177e4SLinus Torvalds /* This routine sends an ack and also updates the window. */ 33601da177e4SLinus Torvalds void tcp_send_ack(struct sock *sk) 33611da177e4SLinus Torvalds { 33621da177e4SLinus Torvalds struct sk_buff *buff; 33631da177e4SLinus Torvalds 3364058dc334SIlpo Järvinen /* If we have been reset, we may not send again. */ 3365058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3366058dc334SIlpo Järvinen return; 3367058dc334SIlpo Järvinen 33689890092eSFlorian Westphal tcp_ca_event(sk, CA_EVENT_NON_DELAYED_ACK); 33699890092eSFlorian Westphal 33701da177e4SLinus Torvalds /* We are not putting this on the write queue, so 33711da177e4SLinus Torvalds * tcp_transmit_skb() will set the ownership to this 33721da177e4SLinus Torvalds * sock. 33731da177e4SLinus Torvalds */ 33747450aaf6SEric Dumazet buff = alloc_skb(MAX_TCP_HEADER, 33757450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 33767450aaf6SEric Dumazet if (unlikely(!buff)) { 3377463c84b9SArnaldo Carvalho de Melo inet_csk_schedule_ack(sk); 3378463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; 33793f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 33803f421baaSArnaldo Carvalho de Melo TCP_DELACK_MAX, TCP_RTO_MAX); 33811da177e4SLinus Torvalds return; 33821da177e4SLinus Torvalds } 33831da177e4SLinus Torvalds 33841da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 33851da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 3386a3433f35SChangli Gao tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); 33871da177e4SLinus Torvalds 338898781965SEric Dumazet /* We do not want pure acks influencing TCP Small Queues or fq/pacing 338998781965SEric Dumazet * too much. 339098781965SEric Dumazet * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784 339198781965SEric Dumazet * We also avoid tcp_wfree() overhead (cache line miss accessing 339298781965SEric Dumazet * tp->tsq_flags) by using regular sock_wfree() 339398781965SEric Dumazet */ 339498781965SEric Dumazet skb_set_tcp_pure_ack(buff); 339598781965SEric Dumazet 33961da177e4SLinus Torvalds /* Send it off, this clears delayed acks for us. */ 33977faee5c0SEric Dumazet skb_mstamp_get(&buff->skb_mstamp); 33987450aaf6SEric Dumazet tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0); 33991da177e4SLinus Torvalds } 3400e3118e83SDaniel Borkmann EXPORT_SYMBOL_GPL(tcp_send_ack); 34011da177e4SLinus Torvalds 34021da177e4SLinus Torvalds /* This routine sends a packet with an out of date sequence 34031da177e4SLinus Torvalds * number. It assumes the other end will try to ack it. 34041da177e4SLinus Torvalds * 34051da177e4SLinus Torvalds * Question: what should we make while urgent mode? 34061da177e4SLinus Torvalds * 4.4BSD forces sending single byte of data. We cannot send 34071da177e4SLinus Torvalds * out of window data, because we have SND.NXT==SND.MAX... 34081da177e4SLinus Torvalds * 34091da177e4SLinus Torvalds * Current solution: to send TWO zero-length segments in urgent mode: 34101da177e4SLinus Torvalds * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 34111da177e4SLinus Torvalds * out-of-date with SND.UNA-1 to probe window. 34121da177e4SLinus Torvalds */ 3413e520af48SEric Dumazet static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib) 34141da177e4SLinus Torvalds { 34151da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 34161da177e4SLinus Torvalds struct sk_buff *skb; 34171da177e4SLinus Torvalds 34181da177e4SLinus Torvalds /* We don't queue it, tcp_transmit_skb() sets ownership. */ 34197450aaf6SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, 34207450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 342151456b29SIan Morris if (!skb) 34221da177e4SLinus Torvalds return -1; 34231da177e4SLinus Torvalds 34241da177e4SLinus Torvalds /* Reserve space for headers and set control bits. */ 34251da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 34261da177e4SLinus Torvalds /* Use a previous sequence. This should cause the other 34271da177e4SLinus Torvalds * end to send an ack. Don't queue or clone SKB, just 34281da177e4SLinus Torvalds * send it. 34291da177e4SLinus Torvalds */ 3430a3433f35SChangli Gao tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); 34317faee5c0SEric Dumazet skb_mstamp_get(&skb->skb_mstamp); 3432e2e8009fSRenato Westphal NET_INC_STATS(sock_net(sk), mib); 34337450aaf6SEric Dumazet return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0); 34341da177e4SLinus Torvalds } 34351da177e4SLinus Torvalds 3436ee995283SPavel Emelyanov void tcp_send_window_probe(struct sock *sk) 3437ee995283SPavel Emelyanov { 3438ee995283SPavel Emelyanov if (sk->sk_state == TCP_ESTABLISHED) { 3439ee995283SPavel Emelyanov tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; 3440e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE); 3441ee995283SPavel Emelyanov } 3442ee995283SPavel Emelyanov } 3443ee995283SPavel Emelyanov 344467edfef7SAndi Kleen /* Initiate keepalive or window probe from timer. */ 3445e520af48SEric Dumazet int tcp_write_wakeup(struct sock *sk, int mib) 34461da177e4SLinus Torvalds { 34471da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 34481da177e4SLinus Torvalds struct sk_buff *skb; 34491da177e4SLinus Torvalds 3450058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3451058dc334SIlpo Järvinen return -1; 3452058dc334SIlpo Järvinen 345300db4124SIan Morris skb = tcp_send_head(sk); 345400db4124SIan Morris if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { 34551da177e4SLinus Torvalds int err; 34560c54b85fSIlpo Järvinen unsigned int mss = tcp_current_mss(sk); 345790840defSIlpo Järvinen unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 34581da177e4SLinus Torvalds 34591da177e4SLinus Torvalds if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 34601da177e4SLinus Torvalds tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 34611da177e4SLinus Torvalds 34621da177e4SLinus Torvalds /* We are probing the opening of a window 34631da177e4SLinus Torvalds * but the window size is != 0 34641da177e4SLinus Torvalds * must have been a result SWS avoidance ( sender ) 34651da177e4SLinus Torvalds */ 34661da177e4SLinus Torvalds if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 34671da177e4SLinus Torvalds skb->len > mss) { 34681da177e4SLinus Torvalds seg_size = min(seg_size, mss); 34694de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 34706cc55e09SOctavian Purdila if (tcp_fragment(sk, skb, seg_size, mss, GFP_ATOMIC)) 34711da177e4SLinus Torvalds return -1; 34721da177e4SLinus Torvalds } else if (!tcp_skb_pcount(skb)) 34735bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss); 34741da177e4SLinus Torvalds 34754de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 3476dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 347766f5fe62SIlpo Järvinen if (!err) 347866f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 34791da177e4SLinus Torvalds return err; 34801da177e4SLinus Torvalds } else { 348133f5f57eSIlpo Järvinen if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) 3482e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 1, mib); 3483e520af48SEric Dumazet return tcp_xmit_probe_skb(sk, 0, mib); 34841da177e4SLinus Torvalds } 34851da177e4SLinus Torvalds } 34861da177e4SLinus Torvalds 34871da177e4SLinus Torvalds /* A window probe timeout has occurred. If window is not closed send 34881da177e4SLinus Torvalds * a partial packet else a zero probe. 34891da177e4SLinus Torvalds */ 34901da177e4SLinus Torvalds void tcp_send_probe0(struct sock *sk) 34911da177e4SLinus Torvalds { 3492463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 34931da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 3494c6214a97SNikolay Borisov struct net *net = sock_net(sk); 3495fcdd1cf4SEric Dumazet unsigned long probe_max; 34961da177e4SLinus Torvalds int err; 34971da177e4SLinus Torvalds 3498e520af48SEric Dumazet err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE); 34991da177e4SLinus Torvalds 3500fe067e8aSDavid S. Miller if (tp->packets_out || !tcp_send_head(sk)) { 35011da177e4SLinus Torvalds /* Cancel probe timer, if it is not required. */ 35026687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 0; 3503463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff = 0; 35041da177e4SLinus Torvalds return; 35051da177e4SLinus Torvalds } 35061da177e4SLinus Torvalds 35071da177e4SLinus Torvalds if (err <= 0) { 3508c6214a97SNikolay Borisov if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2) 3509463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff++; 35106687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out++; 3511fcdd1cf4SEric Dumazet probe_max = TCP_RTO_MAX; 35121da177e4SLinus Torvalds } else { 35131da177e4SLinus Torvalds /* If packet was not sent due to local congestion, 35146687e988SArnaldo Carvalho de Melo * do not backoff and do not remember icsk_probes_out. 35151da177e4SLinus Torvalds * Let local senders to fight for local resources. 35161da177e4SLinus Torvalds * 35171da177e4SLinus Torvalds * Use accumulated backoff yet. 35181da177e4SLinus Torvalds */ 35196687e988SArnaldo Carvalho de Melo if (!icsk->icsk_probes_out) 35206687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 1; 3521fcdd1cf4SEric Dumazet probe_max = TCP_RESOURCE_PROBE_INTERVAL; 35221da177e4SLinus Torvalds } 3523fcdd1cf4SEric Dumazet inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 352421c8fe99SEric Dumazet tcp_probe0_when(sk, probe_max), 3525fcdd1cf4SEric Dumazet TCP_RTO_MAX); 35261da177e4SLinus Torvalds } 35275db92c99SOctavian Purdila 3528ea3bea3aSEric Dumazet int tcp_rtx_synack(const struct sock *sk, struct request_sock *req) 35295db92c99SOctavian Purdila { 35305db92c99SOctavian Purdila const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific; 35315db92c99SOctavian Purdila struct flowi fl; 35325db92c99SOctavian Purdila int res; 35335db92c99SOctavian Purdila 353458d607d3SEric Dumazet tcp_rsk(req)->txhash = net_tx_rndhash(); 3535dc6ef6beSEric Dumazet res = af_ops->send_synack(sk, NULL, &fl, req, NULL, true); 35365db92c99SOctavian Purdila if (!res) { 35375db92c99SOctavian Purdila TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_RETRANSSEGS); 35385db92c99SOctavian Purdila NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 35395db92c99SOctavian Purdila } 35405db92c99SOctavian Purdila return res; 35415db92c99SOctavian Purdila } 35425db92c99SOctavian Purdila EXPORT_SYMBOL(tcp_rtx_synack); 3543