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); 79bec41a11SYuchung Cheng if (!prior_packets || icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) 80750ea2baSYuchung Cheng tcp_rearm_rto(sk); 81f19c29e3SYuchung Cheng 82f7324acdSDavid S. Miller NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT, 83f19c29e3SYuchung Cheng tcp_skb_pcount(skb)); 846a5dc9e5SEric Dumazet } 851da177e4SLinus Torvalds 86a4ecb15aSCui, Cheng /* SND.NXT, if window was not shrunk or the amount of shrunk was less than one 87a4ecb15aSCui, Cheng * window scaling factor due to loss of precision. 881da177e4SLinus Torvalds * If window has been shrunk, what should we make? It is not clear at all. 891da177e4SLinus Torvalds * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-( 901da177e4SLinus Torvalds * Anything in between SND.UNA...SND.UNA+SND.WND also can be already 911da177e4SLinus Torvalds * invalid. OK, let's make this for now: 921da177e4SLinus Torvalds */ 93cf533ea5SEric Dumazet static inline __u32 tcp_acceptable_seq(const struct sock *sk) 941da177e4SLinus Torvalds { 95cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 969e412ba7SIlpo Järvinen 97a4ecb15aSCui, Cheng if (!before(tcp_wnd_end(tp), tp->snd_nxt) || 98a4ecb15aSCui, Cheng (tp->rx_opt.wscale_ok && 99a4ecb15aSCui, Cheng ((tp->snd_nxt - tcp_wnd_end(tp)) < (1 << tp->rx_opt.rcv_wscale)))) 1001da177e4SLinus Torvalds return tp->snd_nxt; 1011da177e4SLinus Torvalds else 10290840defSIlpo Järvinen return tcp_wnd_end(tp); 1031da177e4SLinus Torvalds } 1041da177e4SLinus Torvalds 1051da177e4SLinus Torvalds /* Calculate mss to advertise in SYN segment. 1061da177e4SLinus Torvalds * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that: 1071da177e4SLinus Torvalds * 1081da177e4SLinus Torvalds * 1. It is independent of path mtu. 1091da177e4SLinus Torvalds * 2. Ideally, it is maximal possible segment size i.e. 65535-40. 1101da177e4SLinus Torvalds * 3. For IPv4 it is reasonable to calculate it from maximal MTU of 1111da177e4SLinus Torvalds * attached devices, because some buggy hosts are confused by 1121da177e4SLinus Torvalds * large MSS. 1131da177e4SLinus Torvalds * 4. We do not make 3, we advertise MSS, calculated from first 1141da177e4SLinus Torvalds * hop device mtu, but allow to raise it to ip_rt_min_advmss. 1151da177e4SLinus Torvalds * This may be overridden via information stored in routing table. 1161da177e4SLinus Torvalds * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible, 1171da177e4SLinus Torvalds * probably even Jumbo". 1181da177e4SLinus Torvalds */ 1191da177e4SLinus Torvalds static __u16 tcp_advertise_mss(struct sock *sk) 1201da177e4SLinus Torvalds { 1211da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 122cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1231da177e4SLinus Torvalds int mss = tp->advmss; 1241da177e4SLinus Torvalds 1250dbaee3bSDavid S. Miller if (dst) { 1260dbaee3bSDavid S. Miller unsigned int metric = dst_metric_advmss(dst); 1270dbaee3bSDavid S. Miller 1280dbaee3bSDavid S. Miller if (metric < mss) { 1290dbaee3bSDavid S. Miller mss = metric; 1301da177e4SLinus Torvalds tp->advmss = mss; 1311da177e4SLinus Torvalds } 1320dbaee3bSDavid S. Miller } 1331da177e4SLinus Torvalds 1341da177e4SLinus Torvalds return (__u16)mss; 1351da177e4SLinus Torvalds } 1361da177e4SLinus Torvalds 1371da177e4SLinus Torvalds /* RFC2861. Reset CWND after idle period longer RTO to "restart window". 1386f021c62SEric Dumazet * This is the first part of cwnd validation mechanism. 1396f021c62SEric Dumazet */ 1406f021c62SEric Dumazet void tcp_cwnd_restart(struct sock *sk, s32 delta) 1411da177e4SLinus Torvalds { 142463c84b9SArnaldo Carvalho de Melo struct tcp_sock *tp = tcp_sk(sk); 1436f021c62SEric Dumazet u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk)); 1441da177e4SLinus Torvalds u32 cwnd = tp->snd_cwnd; 1451da177e4SLinus Torvalds 1466687e988SArnaldo Carvalho de Melo tcp_ca_event(sk, CA_EVENT_CWND_RESTART); 1471da177e4SLinus Torvalds 1486687e988SArnaldo Carvalho de Melo tp->snd_ssthresh = tcp_current_ssthresh(sk); 1491da177e4SLinus Torvalds restart_cwnd = min(restart_cwnd, cwnd); 1501da177e4SLinus Torvalds 151463c84b9SArnaldo Carvalho de Melo while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd) 1521da177e4SLinus Torvalds cwnd >>= 1; 1531da177e4SLinus Torvalds tp->snd_cwnd = max(cwnd, restart_cwnd); 1541da177e4SLinus Torvalds tp->snd_cwnd_stamp = tcp_time_stamp; 1551da177e4SLinus Torvalds tp->snd_cwnd_used = 0; 1561da177e4SLinus Torvalds } 1571da177e4SLinus Torvalds 15867edfef7SAndi Kleen /* Congestion state accounting after a packet has been sent. */ 15940efc6faSStephen Hemminger static void tcp_event_data_sent(struct tcp_sock *tp, 160cf533ea5SEric Dumazet struct sock *sk) 1611da177e4SLinus Torvalds { 162463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 163463c84b9SArnaldo Carvalho de Melo const u32 now = tcp_time_stamp; 1641da177e4SLinus Torvalds 16505c5a46dSNeal Cardwell if (tcp_packets_in_flight(tp) == 0) 16605c5a46dSNeal Cardwell tcp_ca_event(sk, CA_EVENT_TX_START); 16705c5a46dSNeal Cardwell 1681da177e4SLinus Torvalds tp->lsndtime = now; 1691da177e4SLinus Torvalds 1701da177e4SLinus Torvalds /* If it is a reply for ato after last received 1711da177e4SLinus Torvalds * packet, enter pingpong mode. 1721da177e4SLinus Torvalds */ 1732251ae46SJon Maxwell if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato) 174463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pingpong = 1; 1751da177e4SLinus Torvalds } 1761da177e4SLinus Torvalds 17767edfef7SAndi Kleen /* Account for an ACK we sent. */ 17840efc6faSStephen Hemminger static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts) 1791da177e4SLinus Torvalds { 180463c84b9SArnaldo Carvalho de Melo tcp_dec_quickack_mode(sk, pkts); 181463c84b9SArnaldo Carvalho de Melo inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); 1821da177e4SLinus Torvalds } 1831da177e4SLinus Torvalds 18485f16525SYuchung Cheng 18585f16525SYuchung Cheng u32 tcp_default_init_rwnd(u32 mss) 18685f16525SYuchung Cheng { 18785f16525SYuchung Cheng /* Initial receive window should be twice of TCP_INIT_CWND to 1889ef71e0cSWeiping Pan * enable proper sending of new unsent data during fast recovery 18985f16525SYuchung Cheng * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a 19085f16525SYuchung Cheng * limit when mss is larger than 1460. 19185f16525SYuchung Cheng */ 19285f16525SYuchung Cheng u32 init_rwnd = TCP_INIT_CWND * 2; 19385f16525SYuchung Cheng 19485f16525SYuchung Cheng if (mss > 1460) 19585f16525SYuchung Cheng init_rwnd = max((1460 * init_rwnd) / mss, 2U); 19685f16525SYuchung Cheng return init_rwnd; 19785f16525SYuchung Cheng } 19885f16525SYuchung Cheng 1991da177e4SLinus Torvalds /* Determine a window scaling and initial window to offer. 2001da177e4SLinus Torvalds * Based on the assumption that the given amount of space 2011da177e4SLinus Torvalds * will be offered. Store the results in the tp structure. 2021da177e4SLinus Torvalds * NOTE: for smooth operation initial space offering should 2031da177e4SLinus Torvalds * be a multiple of mss if possible. We assume here that mss >= 1. 2041da177e4SLinus Torvalds * This MUST be enforced by all callers. 2051da177e4SLinus Torvalds */ 2061da177e4SLinus Torvalds void tcp_select_initial_window(int __space, __u32 mss, 2071da177e4SLinus Torvalds __u32 *rcv_wnd, __u32 *window_clamp, 20831d12926Slaurent chavey int wscale_ok, __u8 *rcv_wscale, 20931d12926Slaurent chavey __u32 init_rcv_wnd) 2101da177e4SLinus Torvalds { 2111da177e4SLinus Torvalds unsigned int space = (__space < 0 ? 0 : __space); 2121da177e4SLinus Torvalds 2131da177e4SLinus Torvalds /* If no clamp set the clamp to the max possible scaled window */ 2141da177e4SLinus Torvalds if (*window_clamp == 0) 215*589c49cbSGao Feng (*window_clamp) = (U16_MAX << TCP_MAX_WSCALE); 2161da177e4SLinus Torvalds space = min(*window_clamp, space); 2171da177e4SLinus Torvalds 2181da177e4SLinus Torvalds /* Quantize space offering to a multiple of mss if possible. */ 2191da177e4SLinus Torvalds if (space > mss) 220*589c49cbSGao Feng space = rounddown(space, mss); 2211da177e4SLinus Torvalds 2221da177e4SLinus Torvalds /* NOTE: offering an initial window larger than 32767 22315d99e02SRick Jones * will break some buggy TCP stacks. If the admin tells us 22415d99e02SRick Jones * it is likely we could be speaking with such a buggy stack 22515d99e02SRick Jones * we will truncate our initial window offering to 32K-1 22615d99e02SRick Jones * unless the remote has sent us a window scaling option, 22715d99e02SRick Jones * which we interpret as a sign the remote TCP is not 22815d99e02SRick Jones * misinterpreting the window field as a signed quantity. 2291da177e4SLinus Torvalds */ 23015d99e02SRick Jones if (sysctl_tcp_workaround_signed_windows) 2311da177e4SLinus Torvalds (*rcv_wnd) = min(space, MAX_TCP_WINDOW); 23215d99e02SRick Jones else 23315d99e02SRick Jones (*rcv_wnd) = space; 23415d99e02SRick Jones 2351da177e4SLinus Torvalds (*rcv_wscale) = 0; 2361da177e4SLinus Torvalds if (wscale_ok) { 237*589c49cbSGao Feng /* Set window scaling on max possible window */ 238f626300aSSoheil Hassas Yeganeh space = max_t(u32, space, sysctl_tcp_rmem[2]); 239f626300aSSoheil Hassas Yeganeh space = max_t(u32, space, sysctl_rmem_max); 240316c1592SStephen Hemminger space = min_t(u32, space, *window_clamp); 241*589c49cbSGao Feng while (space > U16_MAX && (*rcv_wscale) < TCP_MAX_WSCALE) { 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 */ 254*589c49cbSGao Feng (*window_clamp) = min_t(__u32, U16_MAX << (*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, 367ea1627c2SEric Dumazet struct tcphdr *th, 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; 378ea1627c2SEric Dumazet th->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) 386ea1627c2SEric Dumazet th->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; 64295a22caeSFlorian Westphal opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off; 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 | 736f9616c35SEric Dumazet TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) { 737f9616c35SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 738f9616c35SEric Dumazet 739f9616c35SEric Dumazet if (tp->lost_out > tp->retrans_out && 740f9616c35SEric Dumazet tp->snd_cwnd > tcp_packets_in_flight(tp)) 741f9616c35SEric Dumazet tcp_xmit_retransmit_queue(sk); 742f9616c35SEric Dumazet 743f9616c35SEric Dumazet tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle, 744bf06200eSJohn Ogness 0, GFP_ATOMIC); 7456f458dfbSEric Dumazet } 746f9616c35SEric Dumazet } 74746d3ceabSEric Dumazet /* 7488e3bff96Sstephen hemminger * One tasklet per cpu tries to send more skbs. 74946d3ceabSEric Dumazet * We run in tasklet context but need to disable irqs when 7508e3bff96Sstephen hemminger * transferring tsq->head because tcp_wfree() might 75146d3ceabSEric Dumazet * interrupt us (non NAPI drivers) 75246d3ceabSEric Dumazet */ 75346d3ceabSEric Dumazet static void tcp_tasklet_func(unsigned long data) 75446d3ceabSEric Dumazet { 75546d3ceabSEric Dumazet struct tsq_tasklet *tsq = (struct tsq_tasklet *)data; 75646d3ceabSEric Dumazet LIST_HEAD(list); 75746d3ceabSEric Dumazet unsigned long flags; 75846d3ceabSEric Dumazet struct list_head *q, *n; 75946d3ceabSEric Dumazet struct tcp_sock *tp; 76046d3ceabSEric Dumazet struct sock *sk; 76146d3ceabSEric Dumazet 76246d3ceabSEric Dumazet local_irq_save(flags); 76346d3ceabSEric Dumazet list_splice_init(&tsq->head, &list); 76446d3ceabSEric Dumazet local_irq_restore(flags); 76546d3ceabSEric Dumazet 76646d3ceabSEric Dumazet list_for_each_safe(q, n, &list) { 76746d3ceabSEric Dumazet tp = list_entry(q, struct tcp_sock, tsq_node); 76846d3ceabSEric Dumazet list_del(&tp->tsq_node); 76946d3ceabSEric Dumazet 77046d3ceabSEric Dumazet sk = (struct sock *)tp; 7710a9648f1SEric Dumazet smp_mb__before_atomic(); 7727aa5470cSEric Dumazet clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags); 7737aa5470cSEric Dumazet 774b223feb9SEric Dumazet if (!sk->sk_lock.owned && 7757aa5470cSEric Dumazet test_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags)) { 77646d3ceabSEric Dumazet bh_lock_sock(sk); 77746d3ceabSEric Dumazet if (!sock_owned_by_user(sk)) { 7787aa5470cSEric Dumazet clear_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags); 7796f458dfbSEric Dumazet tcp_tsq_handler(sk); 78046d3ceabSEric Dumazet } 78146d3ceabSEric Dumazet bh_unlock_sock(sk); 782b223feb9SEric Dumazet } 78346d3ceabSEric Dumazet 78446d3ceabSEric Dumazet sk_free(sk); 78546d3ceabSEric Dumazet } 78646d3ceabSEric Dumazet } 78746d3ceabSEric Dumazet 78840fc3423SEric Dumazet #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \ 78940fc3423SEric Dumazet TCPF_WRITE_TIMER_DEFERRED | \ 79040fc3423SEric Dumazet TCPF_DELACK_TIMER_DEFERRED | \ 79140fc3423SEric Dumazet TCPF_MTU_REDUCED_DEFERRED) 79246d3ceabSEric Dumazet /** 79346d3ceabSEric Dumazet * tcp_release_cb - tcp release_sock() callback 79446d3ceabSEric Dumazet * @sk: socket 79546d3ceabSEric Dumazet * 79646d3ceabSEric Dumazet * called from release_sock() to perform protocol dependent 79746d3ceabSEric Dumazet * actions before socket release. 79846d3ceabSEric Dumazet */ 79946d3ceabSEric Dumazet void tcp_release_cb(struct sock *sk) 80046d3ceabSEric Dumazet { 8016f458dfbSEric Dumazet unsigned long flags, nflags; 80246d3ceabSEric Dumazet 8036f458dfbSEric Dumazet /* perform an atomic operation only if at least one flag is set */ 8046f458dfbSEric Dumazet do { 8057aa5470cSEric Dumazet flags = sk->sk_tsq_flags; 8066f458dfbSEric Dumazet if (!(flags & TCP_DEFERRED_ALL)) 8076f458dfbSEric Dumazet return; 8086f458dfbSEric Dumazet nflags = flags & ~TCP_DEFERRED_ALL; 8097aa5470cSEric Dumazet } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags); 8106f458dfbSEric Dumazet 81140fc3423SEric Dumazet if (flags & TCPF_TSQ_DEFERRED) 8126f458dfbSEric Dumazet tcp_tsq_handler(sk); 8136f458dfbSEric Dumazet 814c3f9b018SEric Dumazet /* Here begins the tricky part : 815c3f9b018SEric Dumazet * We are called from release_sock() with : 816c3f9b018SEric Dumazet * 1) BH disabled 817c3f9b018SEric Dumazet * 2) sk_lock.slock spinlock held 818c3f9b018SEric Dumazet * 3) socket owned by us (sk->sk_lock.owned == 1) 819c3f9b018SEric Dumazet * 820c3f9b018SEric Dumazet * But following code is meant to be called from BH handlers, 821c3f9b018SEric Dumazet * so we should keep BH disabled, but early release socket ownership 822c3f9b018SEric Dumazet */ 823c3f9b018SEric Dumazet sock_release_ownership(sk); 824c3f9b018SEric Dumazet 82540fc3423SEric Dumazet if (flags & TCPF_WRITE_TIMER_DEFERRED) { 8266f458dfbSEric Dumazet tcp_write_timer_handler(sk); 827144d56e9SEric Dumazet __sock_put(sk); 828144d56e9SEric Dumazet } 82940fc3423SEric Dumazet if (flags & TCPF_DELACK_TIMER_DEFERRED) { 8306f458dfbSEric Dumazet tcp_delack_timer_handler(sk); 831144d56e9SEric Dumazet __sock_put(sk); 832144d56e9SEric Dumazet } 83340fc3423SEric Dumazet if (flags & TCPF_MTU_REDUCED_DEFERRED) { 8344fab9071SNeal Cardwell inet_csk(sk)->icsk_af_ops->mtu_reduced(sk); 835144d56e9SEric Dumazet __sock_put(sk); 836144d56e9SEric Dumazet } 83746d3ceabSEric Dumazet } 83846d3ceabSEric Dumazet EXPORT_SYMBOL(tcp_release_cb); 83946d3ceabSEric Dumazet 84046d3ceabSEric Dumazet void __init tcp_tasklet_init(void) 84146d3ceabSEric Dumazet { 84246d3ceabSEric Dumazet int i; 84346d3ceabSEric Dumazet 84446d3ceabSEric Dumazet for_each_possible_cpu(i) { 84546d3ceabSEric Dumazet struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i); 84646d3ceabSEric Dumazet 84746d3ceabSEric Dumazet INIT_LIST_HEAD(&tsq->head); 84846d3ceabSEric Dumazet tasklet_init(&tsq->tasklet, 84946d3ceabSEric Dumazet tcp_tasklet_func, 85046d3ceabSEric Dumazet (unsigned long)tsq); 85146d3ceabSEric Dumazet } 85246d3ceabSEric Dumazet } 85346d3ceabSEric Dumazet 85446d3ceabSEric Dumazet /* 85546d3ceabSEric Dumazet * Write buffer destructor automatically called from kfree_skb. 8568e3bff96Sstephen hemminger * We can't xmit new skbs from this context, as we might already 85746d3ceabSEric Dumazet * hold qdisc lock. 85846d3ceabSEric Dumazet */ 859d6a4a104SEric Dumazet void tcp_wfree(struct sk_buff *skb) 86046d3ceabSEric Dumazet { 86146d3ceabSEric Dumazet struct sock *sk = skb->sk; 86246d3ceabSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 863408f0a6cSEric Dumazet unsigned long flags, nval, oval; 8649b462d02SEric Dumazet int wmem; 8659b462d02SEric Dumazet 8669b462d02SEric Dumazet /* Keep one reference on sk_wmem_alloc. 8679b462d02SEric Dumazet * Will be released by sk_free() from here or tcp_tasklet_func() 8689b462d02SEric Dumazet */ 8699b462d02SEric Dumazet wmem = atomic_sub_return(skb->truesize - 1, &sk->sk_wmem_alloc); 8709b462d02SEric Dumazet 8719b462d02SEric Dumazet /* If this softirq is serviced by ksoftirqd, we are likely under stress. 8729b462d02SEric Dumazet * Wait until our queues (qdisc + devices) are drained. 8739b462d02SEric Dumazet * This gives : 8749b462d02SEric Dumazet * - less callbacks to tcp_write_xmit(), reducing stress (batches) 8759b462d02SEric Dumazet * - chance for incoming ACK (processed by another cpu maybe) 8769b462d02SEric Dumazet * to migrate this flow (skb->ooo_okay will be eventually set) 8779b462d02SEric Dumazet */ 8789b462d02SEric Dumazet if (wmem >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current) 8799b462d02SEric Dumazet goto out; 88046d3ceabSEric Dumazet 8817aa5470cSEric Dumazet for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) { 88246d3ceabSEric Dumazet struct tsq_tasklet *tsq; 883a9b204d1SEric Dumazet bool empty; 88446d3ceabSEric Dumazet 885408f0a6cSEric Dumazet if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED)) 886408f0a6cSEric Dumazet goto out; 887408f0a6cSEric Dumazet 888b223feb9SEric Dumazet nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED | TCPF_TSQ_DEFERRED; 8897aa5470cSEric Dumazet nval = cmpxchg(&sk->sk_tsq_flags, oval, nval); 890408f0a6cSEric Dumazet if (nval != oval) 891408f0a6cSEric Dumazet continue; 892408f0a6cSEric Dumazet 89346d3ceabSEric Dumazet /* queue this socket to tasklet queue */ 89446d3ceabSEric Dumazet local_irq_save(flags); 895903ceff7SChristoph Lameter tsq = this_cpu_ptr(&tsq_tasklet); 896a9b204d1SEric Dumazet empty = list_empty(&tsq->head); 89746d3ceabSEric Dumazet list_add(&tp->tsq_node, &tsq->head); 898a9b204d1SEric Dumazet if (empty) 89946d3ceabSEric Dumazet tasklet_schedule(&tsq->tasklet); 90046d3ceabSEric Dumazet local_irq_restore(flags); 9019b462d02SEric Dumazet return; 90246d3ceabSEric Dumazet } 9039b462d02SEric Dumazet out: 9049b462d02SEric Dumazet sk_free(sk); 90546d3ceabSEric Dumazet } 90646d3ceabSEric Dumazet 9071da177e4SLinus Torvalds /* This routine actually transmits TCP packets queued in by 9081da177e4SLinus Torvalds * tcp_do_sendmsg(). This is used by both the initial 9091da177e4SLinus Torvalds * transmission and possible later retransmissions. 9101da177e4SLinus Torvalds * All SKB's seen here are completely headerless. It is our 9111da177e4SLinus Torvalds * job to build the TCP header, and pass the packet down to 9121da177e4SLinus Torvalds * IP so it can do the same plus pass the packet off to the 9131da177e4SLinus Torvalds * device. 9141da177e4SLinus Torvalds * 9151da177e4SLinus Torvalds * We are working here with either a clone of the original 9161da177e4SLinus Torvalds * SKB, or a fresh unique copy made by the retransmit engine. 9171da177e4SLinus Torvalds */ 918056834d9SIlpo Järvinen static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, 919056834d9SIlpo Järvinen gfp_t gfp_mask) 9201da177e4SLinus Torvalds { 9216687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 922dfb4b9dcSDavid S. Miller struct inet_sock *inet; 923dfb4b9dcSDavid S. Miller struct tcp_sock *tp; 924dfb4b9dcSDavid S. Miller struct tcp_skb_cb *tcb; 92533ad798cSAdam Langley struct tcp_out_options opts; 92695c96174SEric Dumazet unsigned int tcp_options_size, tcp_header_size; 927cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 9281da177e4SLinus Torvalds struct tcphdr *th; 9291da177e4SLinus Torvalds int err; 9301da177e4SLinus Torvalds 931dfb4b9dcSDavid S. Miller BUG_ON(!skb || !tcp_skb_pcount(skb)); 9326f094b9eSLawrence Brakmo tp = tcp_sk(sk); 933dfb4b9dcSDavid S. Miller 934ccdbb6e9SEric Dumazet if (clone_it) { 935740b0f18SEric Dumazet skb_mstamp_get(&skb->skb_mstamp); 9366f094b9eSLawrence Brakmo TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq 9376f094b9eSLawrence Brakmo - tp->snd_una; 938b9f64820SYuchung Cheng tcp_rate_skb_sent(sk, skb); 939dfb4b9dcSDavid S. Miller 940dfb4b9dcSDavid S. Miller if (unlikely(skb_cloned(skb))) 941dfb4b9dcSDavid S. Miller skb = pskb_copy(skb, gfp_mask); 942dfb4b9dcSDavid S. Miller else 943dfb4b9dcSDavid S. Miller skb = skb_clone(skb, gfp_mask); 944dfb4b9dcSDavid S. Miller if (unlikely(!skb)) 945dfb4b9dcSDavid S. Miller return -ENOBUFS; 946dfb4b9dcSDavid S. Miller } 947dfb4b9dcSDavid S. Miller 948dfb4b9dcSDavid S. Miller inet = inet_sk(sk); 949dfb4b9dcSDavid S. Miller tcb = TCP_SKB_CB(skb); 95033ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 9511da177e4SLinus Torvalds 9524de075e0SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) 95333ad798cSAdam Langley tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); 95433ad798cSAdam Langley else 95533ad798cSAdam Langley tcp_options_size = tcp_established_options(sk, skb, &opts, 95633ad798cSAdam Langley &md5); 95733ad798cSAdam Langley tcp_header_size = tcp_options_size + sizeof(struct tcphdr); 9581da177e4SLinus Torvalds 959547669d4SEric Dumazet /* if no packet is in qdisc/device queue, then allow XPS to select 960b2532eb9SEric Dumazet * another queue. We can be called from tcp_tsq_handler() 961b2532eb9SEric Dumazet * which holds one reference to sk_wmem_alloc. 962b2532eb9SEric Dumazet * 963b2532eb9SEric Dumazet * TODO: Ideally, in-flight pure ACK packets should not matter here. 964b2532eb9SEric Dumazet * One way to get this would be to set skb->truesize = 2 on them. 965547669d4SEric Dumazet */ 966b2532eb9SEric Dumazet skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1); 9671da177e4SLinus Torvalds 96838ab52e8SEric Dumazet /* If we had to use memory reserve to allocate this skb, 96938ab52e8SEric Dumazet * this might cause drops if packet is looped back : 97038ab52e8SEric Dumazet * Other socket might not have SOCK_MEMALLOC. 97138ab52e8SEric Dumazet * Packets not looped back do not care about pfmemalloc. 97238ab52e8SEric Dumazet */ 97338ab52e8SEric Dumazet skb->pfmemalloc = 0; 97438ab52e8SEric Dumazet 975aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 976aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 97746d3ceabSEric Dumazet 97846d3ceabSEric Dumazet skb_orphan(skb); 97946d3ceabSEric Dumazet skb->sk = sk; 9801d2077acSEric Dumazet skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree; 981b73c3d0eSTom Herbert skb_set_hash_from_sk(skb, sk); 98246d3ceabSEric Dumazet atomic_add(skb->truesize, &sk->sk_wmem_alloc); 9831da177e4SLinus Torvalds 984c3a2e837SJulian Anastasov skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm); 985c3a2e837SJulian Anastasov 9861da177e4SLinus Torvalds /* Build TCP header and checksum it. */ 987ea1627c2SEric Dumazet th = (struct tcphdr *)skb->data; 988c720c7e8SEric Dumazet th->source = inet->inet_sport; 989c720c7e8SEric Dumazet th->dest = inet->inet_dport; 9901da177e4SLinus Torvalds th->seq = htonl(tcb->seq); 9911da177e4SLinus Torvalds th->ack_seq = htonl(tp->rcv_nxt); 992df7a3b07SAl Viro *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | 9934de075e0SEric Dumazet tcb->tcp_flags); 994dfb4b9dcSDavid S. Miller 9951da177e4SLinus Torvalds th->check = 0; 9961da177e4SLinus Torvalds th->urg_ptr = 0; 9971da177e4SLinus Torvalds 99833f5f57eSIlpo Järvinen /* The urg_mode check is necessary during a below snd_una win probe */ 9997691367dSHerbert Xu if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { 10007691367dSHerbert Xu if (before(tp->snd_up, tcb->seq + 0x10000)) { 10011da177e4SLinus Torvalds th->urg_ptr = htons(tp->snd_up - tcb->seq); 10021da177e4SLinus Torvalds th->urg = 1; 10037691367dSHerbert Xu } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { 10040eae88f3SEric Dumazet th->urg_ptr = htons(0xFFFF); 10057691367dSHerbert Xu th->urg = 1; 10067691367dSHerbert Xu } 10071da177e4SLinus Torvalds } 10081da177e4SLinus Torvalds 1009bd0388aeSWilliam Allen Simpson tcp_options_write((__be32 *)(th + 1), tp, &opts); 101051466a75SEric Dumazet skb_shinfo(skb)->gso_type = sk->sk_gso_type; 1011ea1627c2SEric Dumazet if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) { 1012ea1627c2SEric Dumazet th->window = htons(tcp_select_window(sk)); 1013ea1627c2SEric Dumazet tcp_ecn_send(sk, skb, th, tcp_header_size); 1014ea1627c2SEric Dumazet } else { 1015ea1627c2SEric Dumazet /* RFC1323: The window in SYN & SYN/ACK segments 1016ea1627c2SEric Dumazet * is never scaled. 1017ea1627c2SEric Dumazet */ 1018ea1627c2SEric Dumazet th->window = htons(min(tp->rcv_wnd, 65535U)); 1019ea1627c2SEric Dumazet } 1020cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 1021cfb6eeb4SYOSHIFUJI Hideaki /* Calculate the MD5 hash, as we have all we need now */ 1022cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 1023a465419bSEric Dumazet sk_nocaps_add(sk, NETIF_F_GSO_MASK); 1024bd0388aeSWilliam Allen Simpson tp->af_specific->calc_md5_hash(opts.hash_location, 102539f8e58eSEric Dumazet md5, sk, skb); 1026cfb6eeb4SYOSHIFUJI Hideaki } 1027cfb6eeb4SYOSHIFUJI Hideaki #endif 1028cfb6eeb4SYOSHIFUJI Hideaki 1029bb296246SHerbert Xu icsk->icsk_af_ops->send_check(sk, skb); 10301da177e4SLinus Torvalds 10314de075e0SEric Dumazet if (likely(tcb->tcp_flags & TCPHDR_ACK)) 1032fc6415bcSDavid S. Miller tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); 10331da177e4SLinus Torvalds 1034a44d6eacSMartin KaFai Lau if (skb->len != tcp_header_size) { 1035cf533ea5SEric Dumazet tcp_event_data_sent(tp, sk); 1036a44d6eacSMartin KaFai Lau tp->data_segs_out += tcp_skb_pcount(skb); 1037a44d6eacSMartin KaFai Lau } 10381da177e4SLinus Torvalds 1039bd37a088SWei Yongjun if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) 1040aa2ea058STom Herbert TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, 1041aa2ea058STom Herbert tcp_skb_pcount(skb)); 10421da177e4SLinus Torvalds 10432efd055cSMarcelo Ricardo Leitner tp->segs_out += tcp_skb_pcount(skb); 1044f69ad292SEric Dumazet /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */ 1045cd7d8498SEric Dumazet skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb); 1046f69ad292SEric Dumazet skb_shinfo(skb)->gso_size = tcp_skb_mss(skb); 1047cd7d8498SEric Dumazet 10487faee5c0SEric Dumazet /* Our usage of tstamp should remain private */ 10492456e855SThomas Gleixner skb->tstamp = 0; 1050971f10ecSEric Dumazet 1051971f10ecSEric Dumazet /* Cleanup our debris for IP stacks */ 1052971f10ecSEric Dumazet memset(skb->cb, 0, max(sizeof(struct inet_skb_parm), 1053971f10ecSEric Dumazet sizeof(struct inet6_skb_parm))); 1054971f10ecSEric Dumazet 1055b0270e91SEric Dumazet err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl); 10567faee5c0SEric Dumazet 105783de47cdSHua Zhong if (likely(err <= 0)) 10581da177e4SLinus Torvalds return err; 10591da177e4SLinus Torvalds 10605ee2c941SChristoph Paasch tcp_enter_cwr(sk); 10611da177e4SLinus Torvalds 1062b9df3cb8SGerrit Renker return net_xmit_eval(err); 10631da177e4SLinus Torvalds } 10641da177e4SLinus Torvalds 106567edfef7SAndi Kleen /* This routine just queues the buffer for sending. 10661da177e4SLinus Torvalds * 10671da177e4SLinus Torvalds * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, 10681da177e4SLinus Torvalds * otherwise socket can stall. 10691da177e4SLinus Torvalds */ 10701da177e4SLinus Torvalds static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) 10711da177e4SLinus Torvalds { 10721da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 10731da177e4SLinus Torvalds 10741da177e4SLinus Torvalds /* Advance write_seq and place onto the write_queue. */ 10751da177e4SLinus Torvalds tp->write_seq = TCP_SKB_CB(skb)->end_seq; 1076f4a775d1SEric Dumazet __skb_header_release(skb); 1077fe067e8aSDavid S. Miller tcp_add_write_queue_tail(sk, skb); 10783ab224beSHideo Aoki sk->sk_wmem_queued += skb->truesize; 10793ab224beSHideo Aoki sk_mem_charge(sk, skb->truesize); 10801da177e4SLinus Torvalds } 10811da177e4SLinus Torvalds 108267edfef7SAndi Kleen /* Initialize TSO segments for a packet. */ 10835bbb432cSEric Dumazet static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now) 1084f6302d1dSDavid S. Miller { 10858f26fb1cSEric Dumazet if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) { 1086f6302d1dSDavid S. Miller /* Avoid the costly divide in the normal 1087f6302d1dSDavid S. Miller * non-TSO case. 1088f6302d1dSDavid S. Miller */ 1089cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, 1); 1090f69ad292SEric Dumazet TCP_SKB_CB(skb)->tcp_gso_size = 0; 1091f6302d1dSDavid S. Miller } else { 1092cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now)); 1093f69ad292SEric Dumazet TCP_SKB_CB(skb)->tcp_gso_size = mss_now; 10941da177e4SLinus Torvalds } 10951da177e4SLinus Torvalds } 10961da177e4SLinus Torvalds 109791fed7a1SIlpo Järvinen /* When a modification to fackets out becomes necessary, we need to check 109868f8353bSIlpo Järvinen * skb is counted to fackets_out or not. 109991fed7a1SIlpo Järvinen */ 1100cf533ea5SEric Dumazet static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb, 110191fed7a1SIlpo Järvinen int decr) 110291fed7a1SIlpo Järvinen { 1103a47e5a98SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1104a47e5a98SIlpo Järvinen 1105dc86967bSIlpo Järvinen if (!tp->sacked_out || tcp_is_reno(tp)) 110691fed7a1SIlpo Järvinen return; 110791fed7a1SIlpo Järvinen 11086859d494SIlpo Järvinen if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq)) 110991fed7a1SIlpo Järvinen tp->fackets_out -= decr; 111091fed7a1SIlpo Järvinen } 111191fed7a1SIlpo Järvinen 1112797108d1SIlpo Järvinen /* Pcount in the middle of the write queue got changed, we need to do various 1113797108d1SIlpo Järvinen * tweaks to fix counters 1114797108d1SIlpo Järvinen */ 1115cf533ea5SEric Dumazet static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) 1116797108d1SIlpo Järvinen { 1117797108d1SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1118797108d1SIlpo Järvinen 1119797108d1SIlpo Järvinen tp->packets_out -= decr; 1120797108d1SIlpo Järvinen 1121797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 1122797108d1SIlpo Järvinen tp->sacked_out -= decr; 1123797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 1124797108d1SIlpo Järvinen tp->retrans_out -= decr; 1125797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 1126797108d1SIlpo Järvinen tp->lost_out -= decr; 1127797108d1SIlpo Järvinen 1128797108d1SIlpo Järvinen /* Reno case is special. Sigh... */ 1129797108d1SIlpo Järvinen if (tcp_is_reno(tp) && decr > 0) 1130797108d1SIlpo Järvinen tp->sacked_out -= min_t(u32, tp->sacked_out, decr); 1131797108d1SIlpo Järvinen 1132797108d1SIlpo Järvinen tcp_adjust_fackets_out(sk, skb, decr); 1133797108d1SIlpo Järvinen 1134797108d1SIlpo Järvinen if (tp->lost_skb_hint && 1135797108d1SIlpo Järvinen before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && 113652cf3cc8SIlpo Järvinen (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))) 1137797108d1SIlpo Järvinen tp->lost_cnt_hint -= decr; 1138797108d1SIlpo Järvinen 1139797108d1SIlpo Järvinen tcp_verify_left_out(tp); 1140797108d1SIlpo Järvinen } 1141797108d1SIlpo Järvinen 11420a2cf20cSSoheil Hassas Yeganeh static bool tcp_has_tx_tstamp(const struct sk_buff *skb) 11430a2cf20cSSoheil Hassas Yeganeh { 11440a2cf20cSSoheil Hassas Yeganeh return TCP_SKB_CB(skb)->txstamp_ack || 11450a2cf20cSSoheil Hassas Yeganeh (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP); 11460a2cf20cSSoheil Hassas Yeganeh } 11470a2cf20cSSoheil Hassas Yeganeh 1148490cc7d0SWillem de Bruijn static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2) 1149490cc7d0SWillem de Bruijn { 1150490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo = skb_shinfo(skb); 1151490cc7d0SWillem de Bruijn 11520a2cf20cSSoheil Hassas Yeganeh if (unlikely(tcp_has_tx_tstamp(skb)) && 1153490cc7d0SWillem de Bruijn !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) { 1154490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo2 = skb_shinfo(skb2); 1155490cc7d0SWillem de Bruijn u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP; 1156490cc7d0SWillem de Bruijn 1157490cc7d0SWillem de Bruijn shinfo->tx_flags &= ~tsflags; 1158490cc7d0SWillem de Bruijn shinfo2->tx_flags |= tsflags; 1159490cc7d0SWillem de Bruijn swap(shinfo->tskey, shinfo2->tskey); 1160b51e13faSMartin KaFai Lau TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack; 1161b51e13faSMartin KaFai Lau TCP_SKB_CB(skb)->txstamp_ack = 0; 1162490cc7d0SWillem de Bruijn } 1163490cc7d0SWillem de Bruijn } 1164490cc7d0SWillem de Bruijn 1165a166140eSMartin KaFai Lau static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2) 1166a166140eSMartin KaFai Lau { 1167a166140eSMartin KaFai Lau TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor; 1168a166140eSMartin KaFai Lau TCP_SKB_CB(skb)->eor = 0; 1169a166140eSMartin KaFai Lau } 1170a166140eSMartin KaFai Lau 11711da177e4SLinus Torvalds /* Function to create two new TCP segments. Shrinks the given segment 11721da177e4SLinus Torvalds * to the specified size and appends a new segment with the rest of the 11731da177e4SLinus Torvalds * packet to the list. This won't be called frequently, I hope. 11741da177e4SLinus Torvalds * Remember, these are still headerless SKBs at this point. 11751da177e4SLinus Torvalds */ 1176056834d9SIlpo Järvinen int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, 11776cc55e09SOctavian Purdila unsigned int mss_now, gfp_t gfp) 11781da177e4SLinus Torvalds { 11791da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 11801da177e4SLinus Torvalds struct sk_buff *buff; 11816475be16SDavid S. Miller int nsize, old_factor; 1182b60b49eaSHerbert Xu int nlen; 11839ce01461SIlpo Järvinen u8 flags; 11841da177e4SLinus Torvalds 11852fceec13SIlpo Järvinen if (WARN_ON(len > skb->len)) 11862fceec13SIlpo Järvinen return -EINVAL; 11876a438bbeSStephen Hemminger 11881da177e4SLinus Torvalds nsize = skb_headlen(skb) - len; 11891da177e4SLinus Torvalds if (nsize < 0) 11901da177e4SLinus Torvalds nsize = 0; 11911da177e4SLinus Torvalds 11926cc55e09SOctavian Purdila if (skb_unclone(skb, gfp)) 11931da177e4SLinus Torvalds return -ENOMEM; 11941da177e4SLinus Torvalds 11951da177e4SLinus Torvalds /* Get a new skb... force flag on. */ 1196eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, nsize, gfp, true); 119751456b29SIan Morris if (!buff) 11981da177e4SLinus Torvalds return -ENOMEM; /* We'll just try again later. */ 1199ef5cb973SHerbert Xu 12003ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 12013ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1202b60b49eaSHerbert Xu nlen = skb->len - len - nsize; 1203b60b49eaSHerbert Xu buff->truesize += nlen; 1204b60b49eaSHerbert Xu skb->truesize -= nlen; 12051da177e4SLinus Torvalds 12061da177e4SLinus Torvalds /* Correct the sequence numbers. */ 12071da177e4SLinus Torvalds TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 12081da177e4SLinus Torvalds TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 12091da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 12101da177e4SLinus Torvalds 12111da177e4SLinus Torvalds /* PSH and FIN should only be set in the second packet. */ 12124de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 12134de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 12144de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1215e14c3cafSHerbert Xu TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; 1216a166140eSMartin KaFai Lau tcp_skb_fragment_eor(skb, buff); 12171da177e4SLinus Torvalds 121884fa7933SPatrick McHardy if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { 12191da177e4SLinus Torvalds /* Copy and checksum data tail into the new buffer. */ 1220056834d9SIlpo Järvinen buff->csum = csum_partial_copy_nocheck(skb->data + len, 1221056834d9SIlpo Järvinen skb_put(buff, nsize), 12221da177e4SLinus Torvalds nsize, 0); 12231da177e4SLinus Torvalds 12241da177e4SLinus Torvalds skb_trim(skb, len); 12251da177e4SLinus Torvalds 12261da177e4SLinus Torvalds skb->csum = csum_block_sub(skb->csum, buff->csum, len); 12271da177e4SLinus Torvalds } else { 122884fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 12291da177e4SLinus Torvalds skb_split(skb, buff, len); 12301da177e4SLinus Torvalds } 12311da177e4SLinus Torvalds 12321da177e4SLinus Torvalds buff->ip_summed = skb->ip_summed; 12331da177e4SLinus Torvalds 1234a61bbcf2SPatrick McHardy buff->tstamp = skb->tstamp; 1235490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 12361da177e4SLinus Torvalds 12376475be16SDavid S. Miller old_factor = tcp_skb_pcount(skb); 12386475be16SDavid S. Miller 12391da177e4SLinus Torvalds /* Fix up tso_factor for both original and new SKB. */ 12405bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 12415bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 12421da177e4SLinus Torvalds 1243b9f64820SYuchung Cheng /* Update delivered info for the new segment */ 1244b9f64820SYuchung Cheng TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx; 1245b9f64820SYuchung Cheng 12466475be16SDavid S. Miller /* If this packet has been sent out already, we must 12476475be16SDavid S. Miller * adjust the various packet counters. 12486475be16SDavid S. Miller */ 1249cf0b450cSHerbert Xu if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 12506475be16SDavid S. Miller int diff = old_factor - tcp_skb_pcount(skb) - 12516475be16SDavid S. Miller tcp_skb_pcount(buff); 12521da177e4SLinus Torvalds 1253797108d1SIlpo Järvinen if (diff) 1254797108d1SIlpo Järvinen tcp_adjust_pcount(sk, skb, diff); 12551da177e4SLinus Torvalds } 12561da177e4SLinus Torvalds 12571da177e4SLinus Torvalds /* Link BUFF into the send queue. */ 1258f4a775d1SEric Dumazet __skb_header_release(buff); 1259fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 12601da177e4SLinus Torvalds 12611da177e4SLinus Torvalds return 0; 12621da177e4SLinus Torvalds } 12631da177e4SLinus Torvalds 12641da177e4SLinus Torvalds /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c 12651da177e4SLinus Torvalds * eventually). The difference is that pulled data not copied, but 12661da177e4SLinus Torvalds * immediately discarded. 12671da177e4SLinus Torvalds */ 1268f2911969SHerbert Xu ~{PmVHI~} static void __pskb_trim_head(struct sk_buff *skb, int len) 12691da177e4SLinus Torvalds { 12707b7fc97aSEric Dumazet struct skb_shared_info *shinfo; 12711da177e4SLinus Torvalds int i, k, eat; 12721da177e4SLinus Torvalds 12734fa48bf3SEric Dumazet eat = min_t(int, len, skb_headlen(skb)); 12744fa48bf3SEric Dumazet if (eat) { 12754fa48bf3SEric Dumazet __skb_pull(skb, eat); 12764fa48bf3SEric Dumazet len -= eat; 12774fa48bf3SEric Dumazet if (!len) 12784fa48bf3SEric Dumazet return; 12794fa48bf3SEric Dumazet } 12801da177e4SLinus Torvalds eat = len; 12811da177e4SLinus Torvalds k = 0; 12827b7fc97aSEric Dumazet shinfo = skb_shinfo(skb); 12837b7fc97aSEric Dumazet for (i = 0; i < shinfo->nr_frags; i++) { 12847b7fc97aSEric Dumazet int size = skb_frag_size(&shinfo->frags[i]); 12859e903e08SEric Dumazet 12869e903e08SEric Dumazet if (size <= eat) { 1287aff65da0SIan Campbell skb_frag_unref(skb, i); 12889e903e08SEric Dumazet eat -= size; 12891da177e4SLinus Torvalds } else { 12907b7fc97aSEric Dumazet shinfo->frags[k] = shinfo->frags[i]; 12911da177e4SLinus Torvalds if (eat) { 12927b7fc97aSEric Dumazet shinfo->frags[k].page_offset += eat; 12937b7fc97aSEric Dumazet skb_frag_size_sub(&shinfo->frags[k], eat); 12941da177e4SLinus Torvalds eat = 0; 12951da177e4SLinus Torvalds } 12961da177e4SLinus Torvalds k++; 12971da177e4SLinus Torvalds } 12981da177e4SLinus Torvalds } 12997b7fc97aSEric Dumazet shinfo->nr_frags = k; 13001da177e4SLinus Torvalds 130127a884dcSArnaldo Carvalho de Melo skb_reset_tail_pointer(skb); 13021da177e4SLinus Torvalds skb->data_len -= len; 13031da177e4SLinus Torvalds skb->len = skb->data_len; 13041da177e4SLinus Torvalds } 13051da177e4SLinus Torvalds 130667edfef7SAndi Kleen /* Remove acked data from a packet in the transmit queue. */ 13071da177e4SLinus Torvalds int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 13081da177e4SLinus Torvalds { 130914bbd6a5SPravin B Shelar if (skb_unclone(skb, GFP_ATOMIC)) 13101da177e4SLinus Torvalds return -ENOMEM; 13111da177e4SLinus Torvalds 13124fa48bf3SEric Dumazet __pskb_trim_head(skb, len); 13131da177e4SLinus Torvalds 13141da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq += len; 131584fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 13161da177e4SLinus Torvalds 13171da177e4SLinus Torvalds skb->truesize -= len; 13181da177e4SLinus Torvalds sk->sk_wmem_queued -= len; 13193ab224beSHideo Aoki sk_mem_uncharge(sk, len); 13201da177e4SLinus Torvalds sock_set_flag(sk, SOCK_QUEUE_SHRUNK); 13211da177e4SLinus Torvalds 13225b35e1e6SNeal Cardwell /* Any change of skb->len requires recalculation of tso factor. */ 13231da177e4SLinus Torvalds if (tcp_skb_pcount(skb) > 1) 13245bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb)); 13251da177e4SLinus Torvalds 13261da177e4SLinus Torvalds return 0; 13271da177e4SLinus Torvalds } 13281da177e4SLinus Torvalds 13291b63edd6SYuchung Cheng /* Calculate MSS not accounting any TCP options. */ 13301b63edd6SYuchung Cheng static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu) 13315d424d5aSJohn Heffner { 1332cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1333cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 13345d424d5aSJohn Heffner int mss_now; 13355d424d5aSJohn Heffner 13365d424d5aSJohn Heffner /* Calculate base mss without TCP options: 13375d424d5aSJohn Heffner It is MMS_S - sizeof(tcphdr) of rfc1122 13385d424d5aSJohn Heffner */ 13395d424d5aSJohn Heffner mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 13405d424d5aSJohn Heffner 134167469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 134267469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 134367469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 134467469601SEric Dumazet 134567469601SEric Dumazet if (dst && dst_allfrag(dst)) 134667469601SEric Dumazet mss_now -= icsk->icsk_af_ops->net_frag_header_len; 134767469601SEric Dumazet } 134867469601SEric Dumazet 13495d424d5aSJohn Heffner /* Clamp it (mss_clamp does not include tcp options) */ 13505d424d5aSJohn Heffner if (mss_now > tp->rx_opt.mss_clamp) 13515d424d5aSJohn Heffner mss_now = tp->rx_opt.mss_clamp; 13525d424d5aSJohn Heffner 13535d424d5aSJohn Heffner /* Now subtract optional transport overhead */ 13545d424d5aSJohn Heffner mss_now -= icsk->icsk_ext_hdr_len; 13555d424d5aSJohn Heffner 13565d424d5aSJohn Heffner /* Then reserve room for full set of TCP options and 8 bytes of data */ 13575d424d5aSJohn Heffner if (mss_now < 48) 13585d424d5aSJohn Heffner mss_now = 48; 13595d424d5aSJohn Heffner return mss_now; 13605d424d5aSJohn Heffner } 13615d424d5aSJohn Heffner 13621b63edd6SYuchung Cheng /* Calculate MSS. Not accounting for SACKs here. */ 13631b63edd6SYuchung Cheng int tcp_mtu_to_mss(struct sock *sk, int pmtu) 13641b63edd6SYuchung Cheng { 13651b63edd6SYuchung Cheng /* Subtract TCP options size, not including SACKs */ 13661b63edd6SYuchung Cheng return __tcp_mtu_to_mss(sk, pmtu) - 13671b63edd6SYuchung Cheng (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr)); 13681b63edd6SYuchung Cheng } 13691b63edd6SYuchung Cheng 13705d424d5aSJohn Heffner /* Inverse of above */ 137167469601SEric Dumazet int tcp_mss_to_mtu(struct sock *sk, int mss) 13725d424d5aSJohn Heffner { 1373cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1374cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 13755d424d5aSJohn Heffner int mtu; 13765d424d5aSJohn Heffner 13775d424d5aSJohn Heffner mtu = mss + 13785d424d5aSJohn Heffner tp->tcp_header_len + 13795d424d5aSJohn Heffner icsk->icsk_ext_hdr_len + 13805d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 13815d424d5aSJohn Heffner 138267469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 138367469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 138467469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 138567469601SEric Dumazet 138667469601SEric Dumazet if (dst && dst_allfrag(dst)) 138767469601SEric Dumazet mtu += icsk->icsk_af_ops->net_frag_header_len; 138867469601SEric Dumazet } 13895d424d5aSJohn Heffner return mtu; 13905d424d5aSJohn Heffner } 1391556c6b46SNeal Cardwell EXPORT_SYMBOL(tcp_mss_to_mtu); 13925d424d5aSJohn Heffner 139367edfef7SAndi Kleen /* MTU probing init per socket */ 13945d424d5aSJohn Heffner void tcp_mtup_init(struct sock *sk) 13955d424d5aSJohn Heffner { 13965d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 13975d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 1398b0f9ca53SFan Du struct net *net = sock_net(sk); 13995d424d5aSJohn Heffner 1400b0f9ca53SFan Du icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1; 14015d424d5aSJohn Heffner icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 14025d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 1403b0f9ca53SFan Du icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss); 14045d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 140505cbc0dbSFan Du if (icsk->icsk_mtup.enabled) 140605cbc0dbSFan Du icsk->icsk_mtup.probe_timestamp = tcp_time_stamp; 14075d424d5aSJohn Heffner } 14084bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_mtup_init); 14095d424d5aSJohn Heffner 14101da177e4SLinus Torvalds /* This function synchronize snd mss to current pmtu/exthdr set. 14111da177e4SLinus Torvalds 14121da177e4SLinus Torvalds tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 14131da177e4SLinus Torvalds for TCP options, but includes only bare TCP header. 14141da177e4SLinus Torvalds 14151da177e4SLinus Torvalds tp->rx_opt.mss_clamp is mss negotiated at connection setup. 1416caa20d9aSStephen Hemminger It is minimum of user_mss and mss received with SYN. 14171da177e4SLinus Torvalds It also does not include TCP options. 14181da177e4SLinus Torvalds 1419d83d8461SArnaldo Carvalho de Melo inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. 14201da177e4SLinus Torvalds 14211da177e4SLinus Torvalds tp->mss_cache is current effective sending mss, including 14221da177e4SLinus Torvalds all tcp options except for SACKs. It is evaluated, 14231da177e4SLinus Torvalds taking into account current pmtu, but never exceeds 14241da177e4SLinus Torvalds tp->rx_opt.mss_clamp. 14251da177e4SLinus Torvalds 14261da177e4SLinus Torvalds NOTE1. rfc1122 clearly states that advertised MSS 14271da177e4SLinus Torvalds DOES NOT include either tcp or ip options. 14281da177e4SLinus Torvalds 1429d83d8461SArnaldo Carvalho de Melo NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache 1430d83d8461SArnaldo Carvalho de Melo are READ ONLY outside this function. --ANK (980731) 14311da177e4SLinus Torvalds */ 14321da177e4SLinus Torvalds unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 14331da177e4SLinus Torvalds { 14341da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1435d83d8461SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 14365d424d5aSJohn Heffner int mss_now; 14371da177e4SLinus Torvalds 14385d424d5aSJohn Heffner if (icsk->icsk_mtup.search_high > pmtu) 14395d424d5aSJohn Heffner icsk->icsk_mtup.search_high = pmtu; 14401da177e4SLinus Torvalds 14415d424d5aSJohn Heffner mss_now = tcp_mtu_to_mss(sk, pmtu); 1442409d22b4SIlpo Järvinen mss_now = tcp_bound_to_half_wnd(tp, mss_now); 14431da177e4SLinus Torvalds 14441da177e4SLinus Torvalds /* And store cached results */ 1445d83d8461SArnaldo Carvalho de Melo icsk->icsk_pmtu_cookie = pmtu; 14465d424d5aSJohn Heffner if (icsk->icsk_mtup.enabled) 14475d424d5aSJohn Heffner mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); 1448c1b4a7e6SDavid S. Miller tp->mss_cache = mss_now; 14491da177e4SLinus Torvalds 14501da177e4SLinus Torvalds return mss_now; 14511da177e4SLinus Torvalds } 14524bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_sync_mss); 14531da177e4SLinus Torvalds 14541da177e4SLinus Torvalds /* Compute the current effective MSS, taking SACKs and IP options, 14551da177e4SLinus Torvalds * and even PMTU discovery events into account. 14561da177e4SLinus Torvalds */ 14570c54b85fSIlpo Järvinen unsigned int tcp_current_mss(struct sock *sk) 14581da177e4SLinus Torvalds { 1459cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1460cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1461c1b4a7e6SDavid S. Miller u32 mss_now; 146295c96174SEric Dumazet unsigned int header_len; 146333ad798cSAdam Langley struct tcp_out_options opts; 146433ad798cSAdam Langley struct tcp_md5sig_key *md5; 14651da177e4SLinus Torvalds 1466c1b4a7e6SDavid S. Miller mss_now = tp->mss_cache; 1467c1b4a7e6SDavid S. Miller 14681da177e4SLinus Torvalds if (dst) { 14691da177e4SLinus Torvalds u32 mtu = dst_mtu(dst); 1470d83d8461SArnaldo Carvalho de Melo if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 14711da177e4SLinus Torvalds mss_now = tcp_sync_mss(sk, mtu); 14721da177e4SLinus Torvalds } 14731da177e4SLinus Torvalds 147433ad798cSAdam Langley header_len = tcp_established_options(sk, NULL, &opts, &md5) + 147533ad798cSAdam Langley sizeof(struct tcphdr); 147633ad798cSAdam Langley /* The mss_cache is sized based on tp->tcp_header_len, which assumes 147733ad798cSAdam Langley * some common options. If this is an odd packet (because we have SACK 147833ad798cSAdam Langley * blocks etc) then our calculated header_len will be different, and 147933ad798cSAdam Langley * we have to adjust mss_now correspondingly */ 148033ad798cSAdam Langley if (header_len != tp->tcp_header_len) { 148133ad798cSAdam Langley int delta = (int) header_len - tp->tcp_header_len; 148233ad798cSAdam Langley mss_now -= delta; 148333ad798cSAdam Langley } 1484cfb6eeb4SYOSHIFUJI Hideaki 14851da177e4SLinus Torvalds return mss_now; 14861da177e4SLinus Torvalds } 14871da177e4SLinus Torvalds 148886fd14adSWeiping Pan /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. 148986fd14adSWeiping Pan * As additional protections, we do not touch cwnd in retransmission phases, 149086fd14adSWeiping Pan * and if application hit its sndbuf limit recently. 149186fd14adSWeiping Pan */ 149286fd14adSWeiping Pan static void tcp_cwnd_application_limited(struct sock *sk) 1493a762a980SDavid S. Miller { 14949e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1495a762a980SDavid S. Miller 149686fd14adSWeiping Pan if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && 149786fd14adSWeiping Pan sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 149886fd14adSWeiping Pan /* Limited by application or receiver window. */ 149986fd14adSWeiping Pan u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); 150086fd14adSWeiping Pan u32 win_used = max(tp->snd_cwnd_used, init_win); 150186fd14adSWeiping Pan if (win_used < tp->snd_cwnd) { 150286fd14adSWeiping Pan tp->snd_ssthresh = tcp_current_ssthresh(sk); 150386fd14adSWeiping Pan tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1; 150486fd14adSWeiping Pan } 150586fd14adSWeiping Pan tp->snd_cwnd_used = 0; 150686fd14adSWeiping Pan } 150786fd14adSWeiping Pan tp->snd_cwnd_stamp = tcp_time_stamp; 150886fd14adSWeiping Pan } 150986fd14adSWeiping Pan 1510ca8a2263SNeal Cardwell static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited) 1511a762a980SDavid S. Miller { 1512a762a980SDavid S. Miller struct tcp_sock *tp = tcp_sk(sk); 1513a762a980SDavid S. Miller 1514ca8a2263SNeal Cardwell /* Track the maximum number of outstanding packets in each 1515ca8a2263SNeal Cardwell * window, and remember whether we were cwnd-limited then. 1516ca8a2263SNeal Cardwell */ 1517ca8a2263SNeal Cardwell if (!before(tp->snd_una, tp->max_packets_seq) || 1518ca8a2263SNeal Cardwell tp->packets_out > tp->max_packets_out) { 1519ca8a2263SNeal Cardwell tp->max_packets_out = tp->packets_out; 1520ca8a2263SNeal Cardwell tp->max_packets_seq = tp->snd_nxt; 1521ca8a2263SNeal Cardwell tp->is_cwnd_limited = is_cwnd_limited; 1522ca8a2263SNeal Cardwell } 1523e114a710SEric Dumazet 152424901551SEric Dumazet if (tcp_is_cwnd_limited(sk)) { 1525a762a980SDavid S. Miller /* Network is feed fully. */ 1526a762a980SDavid S. Miller tp->snd_cwnd_used = 0; 1527a762a980SDavid S. Miller tp->snd_cwnd_stamp = tcp_time_stamp; 1528a762a980SDavid S. Miller } else { 1529a762a980SDavid S. Miller /* Network starves. */ 1530a762a980SDavid S. Miller if (tp->packets_out > tp->snd_cwnd_used) 1531a762a980SDavid S. Miller tp->snd_cwnd_used = tp->packets_out; 1532a762a980SDavid S. Miller 153315d33c07SDavid S. Miller if (sysctl_tcp_slow_start_after_idle && 153415d33c07SDavid S. Miller (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto) 1535a762a980SDavid S. Miller tcp_cwnd_application_limited(sk); 1536b0f71bd3SFrancis Yan 1537b0f71bd3SFrancis Yan /* The following conditions together indicate the starvation 1538b0f71bd3SFrancis Yan * is caused by insufficient sender buffer: 1539b0f71bd3SFrancis Yan * 1) just sent some data (see tcp_write_xmit) 1540b0f71bd3SFrancis Yan * 2) not cwnd limited (this else condition) 1541b0f71bd3SFrancis Yan * 3) no more data to send (null tcp_send_head ) 1542b0f71bd3SFrancis Yan * 4) application is hitting buffer limit (SOCK_NOSPACE) 1543b0f71bd3SFrancis Yan */ 1544b0f71bd3SFrancis Yan if (!tcp_send_head(sk) && sk->sk_socket && 1545b0f71bd3SFrancis Yan test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) && 1546b0f71bd3SFrancis Yan (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) 1547b0f71bd3SFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED); 1548a762a980SDavid S. Miller } 1549a762a980SDavid S. Miller } 1550a762a980SDavid S. Miller 1551d4589926SEric Dumazet /* Minshall's variant of the Nagle send check. */ 1552d4589926SEric Dumazet static bool tcp_minshall_check(const struct tcp_sock *tp) 1553d4589926SEric Dumazet { 1554d4589926SEric Dumazet return after(tp->snd_sml, tp->snd_una) && 1555d4589926SEric Dumazet !after(tp->snd_sml, tp->snd_nxt); 1556d4589926SEric Dumazet } 1557d4589926SEric Dumazet 1558d4589926SEric Dumazet /* Update snd_sml if this skb is under mss 1559d4589926SEric Dumazet * Note that a TSO packet might end with a sub-mss segment 1560d4589926SEric Dumazet * The test is really : 1561d4589926SEric Dumazet * if ((skb->len % mss) != 0) 1562d4589926SEric Dumazet * tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1563d4589926SEric Dumazet * But we can avoid doing the divide again given we already have 1564d4589926SEric Dumazet * skb_pcount = skb->len / mss_now 15650e3a4803SIlpo Järvinen */ 1566d4589926SEric Dumazet static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now, 1567d4589926SEric Dumazet const struct sk_buff *skb) 1568d4589926SEric Dumazet { 1569d4589926SEric Dumazet if (skb->len < tcp_skb_pcount(skb) * mss_now) 1570d4589926SEric Dumazet tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1571d4589926SEric Dumazet } 1572d4589926SEric Dumazet 1573d4589926SEric Dumazet /* Return false, if packet can be sent now without violation Nagle's rules: 1574d4589926SEric Dumazet * 1. It is full sized. (provided by caller in %partial bool) 1575d4589926SEric Dumazet * 2. Or it contains FIN. (already checked by caller) 1576d4589926SEric Dumazet * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. 1577d4589926SEric Dumazet * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 1578d4589926SEric Dumazet * With Minshall's modification: all sent small packets are ACKed. 1579d4589926SEric Dumazet */ 1580d4589926SEric Dumazet static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp, 1581cc93fc51SPeter Pan(潘卫平) int nonagle) 1582d4589926SEric Dumazet { 1583d4589926SEric Dumazet return partial && 1584d4589926SEric Dumazet ((nonagle & TCP_NAGLE_CORK) || 1585d4589926SEric Dumazet (!nonagle && tp->packets_out && tcp_minshall_check(tp))); 1586d4589926SEric Dumazet } 1587605ad7f1SEric Dumazet 1588605ad7f1SEric Dumazet /* Return how many segs we'd like on a TSO packet, 1589605ad7f1SEric Dumazet * to send one TSO packet per ms 1590605ad7f1SEric Dumazet */ 15911b3878caSNeal Cardwell u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now, 15921b3878caSNeal Cardwell int min_tso_segs) 1593605ad7f1SEric Dumazet { 1594605ad7f1SEric Dumazet u32 bytes, segs; 1595605ad7f1SEric Dumazet 1596605ad7f1SEric Dumazet bytes = min(sk->sk_pacing_rate >> 10, 1597605ad7f1SEric Dumazet sk->sk_gso_max_size - 1 - MAX_TCP_HEADER); 1598605ad7f1SEric Dumazet 1599605ad7f1SEric Dumazet /* Goal is to send at least one packet per ms, 1600605ad7f1SEric Dumazet * not one big TSO packet every 100 ms. 1601605ad7f1SEric Dumazet * This preserves ACK clocking and is consistent 1602605ad7f1SEric Dumazet * with tcp_tso_should_defer() heuristic. 1603605ad7f1SEric Dumazet */ 16041b3878caSNeal Cardwell segs = max_t(u32, bytes / mss_now, min_tso_segs); 1605605ad7f1SEric Dumazet 1606605ad7f1SEric Dumazet return min_t(u32, segs, sk->sk_gso_max_segs); 1607605ad7f1SEric Dumazet } 16081b3878caSNeal Cardwell EXPORT_SYMBOL(tcp_tso_autosize); 1609605ad7f1SEric Dumazet 1610ed6e7268SNeal Cardwell /* Return the number of segments we want in the skb we are transmitting. 1611ed6e7268SNeal Cardwell * See if congestion control module wants to decide; otherwise, autosize. 1612ed6e7268SNeal Cardwell */ 1613ed6e7268SNeal Cardwell static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now) 1614ed6e7268SNeal Cardwell { 1615ed6e7268SNeal Cardwell const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 1616ed6e7268SNeal Cardwell u32 tso_segs = ca_ops->tso_segs_goal ? ca_ops->tso_segs_goal(sk) : 0; 1617ed6e7268SNeal Cardwell 16181b3878caSNeal Cardwell return tso_segs ? : 16191b3878caSNeal Cardwell tcp_tso_autosize(sk, mss_now, sysctl_tcp_min_tso_segs); 1620ed6e7268SNeal Cardwell } 1621ed6e7268SNeal Cardwell 1622d4589926SEric Dumazet /* Returns the portion of skb which can be sent right away */ 1623d4589926SEric Dumazet static unsigned int tcp_mss_split_point(const struct sock *sk, 1624d4589926SEric Dumazet const struct sk_buff *skb, 1625d4589926SEric Dumazet unsigned int mss_now, 1626d4589926SEric Dumazet unsigned int max_segs, 1627d4589926SEric Dumazet int nonagle) 1628c1b4a7e6SDavid S. Miller { 1629cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1630d4589926SEric Dumazet u32 partial, needed, window, max_len; 1631c1b4a7e6SDavid S. Miller 163290840defSIlpo Järvinen window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 16331485348dSBen Hutchings max_len = mss_now * max_segs; 16340e3a4803SIlpo Järvinen 16351485348dSBen Hutchings if (likely(max_len <= window && skb != tcp_write_queue_tail(sk))) 16361485348dSBen Hutchings return max_len; 16370e3a4803SIlpo Järvinen 16385ea3a748SIlpo Järvinen needed = min(skb->len, window); 16395ea3a748SIlpo Järvinen 16401485348dSBen Hutchings if (max_len <= needed) 16411485348dSBen Hutchings return max_len; 16420e3a4803SIlpo Järvinen 1643d4589926SEric Dumazet partial = needed % mss_now; 1644d4589926SEric Dumazet /* If last segment is not a full MSS, check if Nagle rules allow us 1645d4589926SEric Dumazet * to include this last segment in this skb. 1646d4589926SEric Dumazet * Otherwise, we'll split the skb at last MSS boundary 1647d4589926SEric Dumazet */ 1648cc93fc51SPeter Pan(潘卫平) if (tcp_nagle_check(partial != 0, tp, nonagle)) 1649d4589926SEric Dumazet return needed - partial; 1650d4589926SEric Dumazet 1651d4589926SEric Dumazet return needed; 1652c1b4a7e6SDavid S. Miller } 1653c1b4a7e6SDavid S. Miller 1654c1b4a7e6SDavid S. Miller /* Can at least one segment of SKB be sent right now, according to the 1655c1b4a7e6SDavid S. Miller * congestion window rules? If so, return how many segments are allowed. 1656c1b4a7e6SDavid S. Miller */ 1657cf533ea5SEric Dumazet static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, 1658cf533ea5SEric Dumazet const struct sk_buff *skb) 1659c1b4a7e6SDavid S. Miller { 1660d649a7a8SEric Dumazet u32 in_flight, cwnd, halfcwnd; 1661c1b4a7e6SDavid S. Miller 1662c1b4a7e6SDavid S. Miller /* Don't be strict about the congestion window for the final FIN. */ 16634de075e0SEric Dumazet if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && 16644de075e0SEric Dumazet tcp_skb_pcount(skb) == 1) 1665c1b4a7e6SDavid S. Miller return 1; 1666c1b4a7e6SDavid S. Miller 1667c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1668c1b4a7e6SDavid S. Miller cwnd = tp->snd_cwnd; 1669d649a7a8SEric Dumazet if (in_flight >= cwnd) 1670c1b4a7e6SDavid S. Miller return 0; 1671d649a7a8SEric Dumazet 1672d649a7a8SEric Dumazet /* For better scheduling, ensure we have at least 1673d649a7a8SEric Dumazet * 2 GSO packets in flight. 1674d649a7a8SEric Dumazet */ 1675d649a7a8SEric Dumazet halfcwnd = max(cwnd >> 1, 1U); 1676d649a7a8SEric Dumazet return min(halfcwnd, cwnd - in_flight); 1677c1b4a7e6SDavid S. Miller } 1678c1b4a7e6SDavid S. Miller 1679b595076aSUwe Kleine-König /* Initialize TSO state of a skb. 168067edfef7SAndi Kleen * This must be invoked the first time we consider transmitting 1681c1b4a7e6SDavid S. Miller * SKB onto the wire. 1682c1b4a7e6SDavid S. Miller */ 16835bbb432cSEric Dumazet static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now) 1684c1b4a7e6SDavid S. Miller { 1685c1b4a7e6SDavid S. Miller int tso_segs = tcp_skb_pcount(skb); 1686c1b4a7e6SDavid S. Miller 1687f8269a49SIlpo Järvinen if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { 16885bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 1689c1b4a7e6SDavid S. Miller tso_segs = tcp_skb_pcount(skb); 1690c1b4a7e6SDavid S. Miller } 1691c1b4a7e6SDavid S. Miller return tso_segs; 1692c1b4a7e6SDavid S. Miller } 1693c1b4a7e6SDavid S. Miller 1694c1b4a7e6SDavid S. Miller 1695a2a385d6SEric Dumazet /* Return true if the Nagle test allows this packet to be 1696c1b4a7e6SDavid S. Miller * sent now. 1697c1b4a7e6SDavid S. Miller */ 1698a2a385d6SEric Dumazet static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, 1699c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1700c1b4a7e6SDavid S. Miller { 1701c1b4a7e6SDavid S. Miller /* Nagle rule does not apply to frames, which sit in the middle of the 1702c1b4a7e6SDavid S. Miller * write_queue (they have no chances to get new data). 1703c1b4a7e6SDavid S. Miller * 1704c1b4a7e6SDavid S. Miller * This is implemented in the callers, where they modify the 'nonagle' 1705c1b4a7e6SDavid S. Miller * argument based upon the location of SKB in the send queue. 1706c1b4a7e6SDavid S. Miller */ 1707c1b4a7e6SDavid S. Miller if (nonagle & TCP_NAGLE_PUSH) 1708a2a385d6SEric Dumazet return true; 1709c1b4a7e6SDavid S. Miller 17109b44190dSYuchung Cheng /* Don't use the nagle rule for urgent data (or for the final FIN). */ 17119b44190dSYuchung Cheng if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) 1712a2a385d6SEric Dumazet return true; 1713c1b4a7e6SDavid S. Miller 1714cc93fc51SPeter Pan(潘卫平) if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle)) 1715a2a385d6SEric Dumazet return true; 1716c1b4a7e6SDavid S. Miller 1717a2a385d6SEric Dumazet return false; 1718c1b4a7e6SDavid S. Miller } 1719c1b4a7e6SDavid S. Miller 1720c1b4a7e6SDavid S. Miller /* Does at least the first segment of SKB fit into the send window? */ 1721a2a385d6SEric Dumazet static bool tcp_snd_wnd_test(const struct tcp_sock *tp, 1722a2a385d6SEric Dumazet const struct sk_buff *skb, 1723056834d9SIlpo Järvinen unsigned int cur_mss) 1724c1b4a7e6SDavid S. Miller { 1725c1b4a7e6SDavid S. Miller u32 end_seq = TCP_SKB_CB(skb)->end_seq; 1726c1b4a7e6SDavid S. Miller 1727c1b4a7e6SDavid S. Miller if (skb->len > cur_mss) 1728c1b4a7e6SDavid S. Miller end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 1729c1b4a7e6SDavid S. Miller 173090840defSIlpo Järvinen return !after(end_seq, tcp_wnd_end(tp)); 1731c1b4a7e6SDavid S. Miller } 1732c1b4a7e6SDavid S. Miller 1733fe067e8aSDavid S. Miller /* This checks if the data bearing packet SKB (usually tcp_send_head(sk)) 1734c1b4a7e6SDavid S. Miller * should be put on the wire right now. If so, it returns the number of 1735c1b4a7e6SDavid S. Miller * packets allowed by the congestion window. 1736c1b4a7e6SDavid S. Miller */ 1737cf533ea5SEric Dumazet static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb, 1738c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1739c1b4a7e6SDavid S. Miller { 1740cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1741c1b4a7e6SDavid S. Miller unsigned int cwnd_quota; 1742c1b4a7e6SDavid S. Miller 17435bbb432cSEric Dumazet tcp_init_tso_segs(skb, cur_mss); 1744c1b4a7e6SDavid S. Miller 1745c1b4a7e6SDavid S. Miller if (!tcp_nagle_test(tp, skb, cur_mss, nonagle)) 1746c1b4a7e6SDavid S. Miller return 0; 1747c1b4a7e6SDavid S. Miller 1748c1b4a7e6SDavid S. Miller cwnd_quota = tcp_cwnd_test(tp, skb); 1749056834d9SIlpo Järvinen if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss)) 1750c1b4a7e6SDavid S. Miller cwnd_quota = 0; 1751c1b4a7e6SDavid S. Miller 1752c1b4a7e6SDavid S. Miller return cwnd_quota; 1753c1b4a7e6SDavid S. Miller } 1754c1b4a7e6SDavid S. Miller 175567edfef7SAndi Kleen /* Test if sending is allowed right now. */ 1756a2a385d6SEric Dumazet bool tcp_may_send_now(struct sock *sk) 1757c1b4a7e6SDavid S. Miller { 1758cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1759fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 1760c1b4a7e6SDavid S. Miller 1761a02cec21SEric Dumazet return skb && 17620c54b85fSIlpo Järvinen tcp_snd_test(sk, skb, tcp_current_mss(sk), 1763c1b4a7e6SDavid S. Miller (tcp_skb_is_last(sk, skb) ? 1764a02cec21SEric Dumazet tp->nonagle : TCP_NAGLE_PUSH)); 1765c1b4a7e6SDavid S. Miller } 1766c1b4a7e6SDavid S. Miller 1767c1b4a7e6SDavid S. Miller /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 1768c1b4a7e6SDavid S. Miller * which is put after SKB on the list. It is very much like 1769c1b4a7e6SDavid S. Miller * tcp_fragment() except that it may make several kinds of assumptions 1770c1b4a7e6SDavid S. Miller * in order to speed up the splitting operation. In particular, we 1771c1b4a7e6SDavid S. Miller * know that all the data is in scatter-gather pages, and that the 1772c1b4a7e6SDavid S. Miller * packet has never been sent out before (and thus is not cloned). 1773c1b4a7e6SDavid S. Miller */ 1774056834d9SIlpo Järvinen static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, 1775c4ead4c5SEric Dumazet unsigned int mss_now, gfp_t gfp) 1776c1b4a7e6SDavid S. Miller { 1777c1b4a7e6SDavid S. Miller struct sk_buff *buff; 1778c1b4a7e6SDavid S. Miller int nlen = skb->len - len; 17799ce01461SIlpo Järvinen u8 flags; 1780c1b4a7e6SDavid S. Miller 1781c1b4a7e6SDavid S. Miller /* All of a TSO frame must be composed of paged data. */ 1782c8ac3774SHerbert Xu if (skb->len != skb->data_len) 17836cc55e09SOctavian Purdila return tcp_fragment(sk, skb, len, mss_now, gfp); 1784c1b4a7e6SDavid S. Miller 1785eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, gfp, true); 178651456b29SIan Morris if (unlikely(!buff)) 1787c1b4a7e6SDavid S. Miller return -ENOMEM; 1788c1b4a7e6SDavid S. Miller 17893ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 17903ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1791b60b49eaSHerbert Xu buff->truesize += nlen; 1792c1b4a7e6SDavid S. Miller skb->truesize -= nlen; 1793c1b4a7e6SDavid S. Miller 1794c1b4a7e6SDavid S. Miller /* Correct the sequence numbers. */ 1795c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 1796c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 1797c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 1798c1b4a7e6SDavid S. Miller 1799c1b4a7e6SDavid S. Miller /* PSH and FIN should only be set in the second packet. */ 18004de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 18014de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 18024de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1803c1b4a7e6SDavid S. Miller 1804c1b4a7e6SDavid S. Miller /* This packet was never sent out yet, so no SACK bits. */ 1805c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->sacked = 0; 1806c1b4a7e6SDavid S. Miller 1807a166140eSMartin KaFai Lau tcp_skb_fragment_eor(skb, buff); 1808a166140eSMartin KaFai Lau 180984fa7933SPatrick McHardy buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; 1810c1b4a7e6SDavid S. Miller skb_split(skb, buff, len); 1811490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 1812c1b4a7e6SDavid S. Miller 1813c1b4a7e6SDavid S. Miller /* Fix up tso_factor for both original and new SKB. */ 18145bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 18155bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 1816c1b4a7e6SDavid S. Miller 1817c1b4a7e6SDavid S. Miller /* Link BUFF into the send queue. */ 1818f4a775d1SEric Dumazet __skb_header_release(buff); 1819fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 1820c1b4a7e6SDavid S. Miller 1821c1b4a7e6SDavid S. Miller return 0; 1822c1b4a7e6SDavid S. Miller } 1823c1b4a7e6SDavid S. Miller 1824c1b4a7e6SDavid S. Miller /* Try to defer sending, if possible, in order to minimize the amount 1825c1b4a7e6SDavid S. Miller * of TSO splitting we do. View it as a kind of TSO Nagle test. 1826c1b4a7e6SDavid S. Miller * 1827c1b4a7e6SDavid S. Miller * This algorithm is from John Heffner. 1828c1b4a7e6SDavid S. Miller */ 1829ca8a2263SNeal Cardwell static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb, 1830605ad7f1SEric Dumazet bool *is_cwnd_limited, u32 max_segs) 1831c1b4a7e6SDavid S. Miller { 18326687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 183350c8339eSEric Dumazet u32 age, send_win, cong_win, limit, in_flight; 183450c8339eSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 183550c8339eSEric Dumazet struct skb_mstamp now; 183650c8339eSEric Dumazet struct sk_buff *head; 1837ad9f4f50SEric Dumazet int win_divisor; 1838c1b4a7e6SDavid S. Miller 18394de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 1840ae8064acSJohn Heffner goto send_now; 1841c1b4a7e6SDavid S. Miller 184299d7662aSEric Dumazet if (icsk->icsk_ca_state >= TCP_CA_Recovery) 1843ae8064acSJohn Heffner goto send_now; 1844ae8064acSJohn Heffner 18455f852eb5SEric Dumazet /* Avoid bursty behavior by allowing defer 18465f852eb5SEric Dumazet * only if the last write was recent. 18475f852eb5SEric Dumazet */ 18485f852eb5SEric Dumazet if ((s32)(tcp_time_stamp - tp->lsndtime) > 0) 1849ae8064acSJohn Heffner goto send_now; 1850908a75c1SDavid S. Miller 1851c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1852c1b4a7e6SDavid S. Miller 1853056834d9SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight)); 1854c1b4a7e6SDavid S. Miller 185590840defSIlpo Järvinen send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1856c1b4a7e6SDavid S. Miller 1857c1b4a7e6SDavid S. Miller /* From in_flight test above, we know that cwnd > in_flight. */ 1858c1b4a7e6SDavid S. Miller cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; 1859c1b4a7e6SDavid S. Miller 1860c1b4a7e6SDavid S. Miller limit = min(send_win, cong_win); 1861c1b4a7e6SDavid S. Miller 1862ba244fe9SDavid S. Miller /* If a full-sized TSO skb can be sent, do it. */ 1863605ad7f1SEric Dumazet if (limit >= max_segs * tp->mss_cache) 1864ae8064acSJohn Heffner goto send_now; 1865ba244fe9SDavid S. Miller 186662ad2761SIlpo Järvinen /* Middle in queue won't get any more data, full sendable already? */ 186762ad2761SIlpo Järvinen if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) 186862ad2761SIlpo Järvinen goto send_now; 186962ad2761SIlpo Järvinen 1870ad9f4f50SEric Dumazet win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor); 1871ad9f4f50SEric Dumazet if (win_divisor) { 1872c1b4a7e6SDavid S. Miller u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); 1873c1b4a7e6SDavid S. Miller 1874c1b4a7e6SDavid S. Miller /* If at least some fraction of a window is available, 1875c1b4a7e6SDavid S. Miller * just use it. 1876c1b4a7e6SDavid S. Miller */ 1877ad9f4f50SEric Dumazet chunk /= win_divisor; 1878c1b4a7e6SDavid S. Miller if (limit >= chunk) 1879ae8064acSJohn Heffner goto send_now; 1880c1b4a7e6SDavid S. Miller } else { 1881c1b4a7e6SDavid S. Miller /* Different approach, try not to defer past a single 1882c1b4a7e6SDavid S. Miller * ACK. Receiver should ACK every other full sized 1883c1b4a7e6SDavid S. Miller * frame, so if we have space for more than 3 frames 1884c1b4a7e6SDavid S. Miller * then send now. 1885c1b4a7e6SDavid S. Miller */ 18866b5a5c0dSNeal Cardwell if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) 1887ae8064acSJohn Heffner goto send_now; 1888c1b4a7e6SDavid S. Miller } 1889c1b4a7e6SDavid S. Miller 189050c8339eSEric Dumazet head = tcp_write_queue_head(sk); 189150c8339eSEric Dumazet skb_mstamp_get(&now); 189250c8339eSEric Dumazet age = skb_mstamp_us_delta(&now, &head->skb_mstamp); 189350c8339eSEric Dumazet /* If next ACK is likely to come too late (half srtt), do not defer */ 189450c8339eSEric Dumazet if (age < (tp->srtt_us >> 4)) 189550c8339eSEric Dumazet goto send_now; 189650c8339eSEric Dumazet 18975f852eb5SEric Dumazet /* Ok, it looks like it is advisable to defer. */ 1898ae8064acSJohn Heffner 1899d2e1339fSBendik Rønning Opstad if (cong_win < send_win && cong_win <= skb->len) 1900ca8a2263SNeal Cardwell *is_cwnd_limited = true; 1901ca8a2263SNeal Cardwell 1902a2a385d6SEric Dumazet return true; 1903ae8064acSJohn Heffner 1904ae8064acSJohn Heffner send_now: 1905a2a385d6SEric Dumazet return false; 1906c1b4a7e6SDavid S. Miller } 1907c1b4a7e6SDavid S. Miller 190805cbc0dbSFan Du static inline void tcp_mtu_check_reprobe(struct sock *sk) 190905cbc0dbSFan Du { 191005cbc0dbSFan Du struct inet_connection_sock *icsk = inet_csk(sk); 191105cbc0dbSFan Du struct tcp_sock *tp = tcp_sk(sk); 191205cbc0dbSFan Du struct net *net = sock_net(sk); 191305cbc0dbSFan Du u32 interval; 191405cbc0dbSFan Du s32 delta; 191505cbc0dbSFan Du 191605cbc0dbSFan Du interval = net->ipv4.sysctl_tcp_probe_interval; 191705cbc0dbSFan Du delta = tcp_time_stamp - icsk->icsk_mtup.probe_timestamp; 191805cbc0dbSFan Du if (unlikely(delta >= interval * HZ)) { 191905cbc0dbSFan Du int mss = tcp_current_mss(sk); 192005cbc0dbSFan Du 192105cbc0dbSFan Du /* Update current search range */ 192205cbc0dbSFan Du icsk->icsk_mtup.probe_size = 0; 192305cbc0dbSFan Du icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + 192405cbc0dbSFan Du sizeof(struct tcphdr) + 192505cbc0dbSFan Du icsk->icsk_af_ops->net_header_len; 192605cbc0dbSFan Du icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss); 192705cbc0dbSFan Du 192805cbc0dbSFan Du /* Update probe time stamp */ 192905cbc0dbSFan Du icsk->icsk_mtup.probe_timestamp = tcp_time_stamp; 193005cbc0dbSFan Du } 193105cbc0dbSFan Du } 193205cbc0dbSFan Du 19335d424d5aSJohn Heffner /* Create a new MTU probe if we are ready. 193467edfef7SAndi Kleen * MTU probe is regularly attempting to increase the path MTU by 193567edfef7SAndi Kleen * deliberately sending larger packets. This discovers routing 193667edfef7SAndi Kleen * changes resulting in larger path MTUs. 193767edfef7SAndi Kleen * 19385d424d5aSJohn Heffner * Returns 0 if we should wait to probe (no cwnd available), 19395d424d5aSJohn Heffner * 1 if a probe was sent, 1940056834d9SIlpo Järvinen * -1 otherwise 1941056834d9SIlpo Järvinen */ 19425d424d5aSJohn Heffner static int tcp_mtu_probe(struct sock *sk) 19435d424d5aSJohn Heffner { 19445d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 194512a59abcSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 19465d424d5aSJohn Heffner struct sk_buff *skb, *nskb, *next; 19476b58e0a5SFan Du struct net *net = sock_net(sk); 19485d424d5aSJohn Heffner int probe_size; 194991cc17c0SIlpo Järvinen int size_needed; 195012a59abcSEric Dumazet int copy, len; 19515d424d5aSJohn Heffner int mss_now; 19526b58e0a5SFan Du int interval; 19535d424d5aSJohn Heffner 19545d424d5aSJohn Heffner /* Not currently probing/verifying, 19555d424d5aSJohn Heffner * not in recovery, 19565d424d5aSJohn Heffner * have enough cwnd, and 195712a59abcSEric Dumazet * not SACKing (the variable headers throw things off) 195812a59abcSEric Dumazet */ 195912a59abcSEric Dumazet if (likely(!icsk->icsk_mtup.enabled || 19605d424d5aSJohn Heffner icsk->icsk_mtup.probe_size || 19615d424d5aSJohn Heffner inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 19625d424d5aSJohn Heffner tp->snd_cwnd < 11 || 196312a59abcSEric Dumazet tp->rx_opt.num_sacks || tp->rx_opt.dsack)) 19645d424d5aSJohn Heffner return -1; 19655d424d5aSJohn Heffner 19666b58e0a5SFan Du /* Use binary search for probe_size between tcp_mss_base, 19676b58e0a5SFan Du * and current mss_clamp. if (search_high - search_low) 19686b58e0a5SFan Du * smaller than a threshold, backoff from probing. 19696b58e0a5SFan Du */ 19700c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 19716b58e0a5SFan Du probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high + 19726b58e0a5SFan Du icsk->icsk_mtup.search_low) >> 1); 197391cc17c0SIlpo Järvinen size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; 19746b58e0a5SFan Du interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low; 197505cbc0dbSFan Du /* When misfortune happens, we are reprobing actively, 197605cbc0dbSFan Du * and then reprobe timer has expired. We stick with current 197705cbc0dbSFan Du * probing process by not resetting search range to its orignal. 197805cbc0dbSFan Du */ 19796b58e0a5SFan Du if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) || 198005cbc0dbSFan Du interval < net->ipv4.sysctl_tcp_probe_threshold) { 198105cbc0dbSFan Du /* Check whether enough time has elaplased for 198205cbc0dbSFan Du * another round of probing. 198305cbc0dbSFan Du */ 198405cbc0dbSFan Du tcp_mtu_check_reprobe(sk); 19855d424d5aSJohn Heffner return -1; 19865d424d5aSJohn Heffner } 19875d424d5aSJohn Heffner 19885d424d5aSJohn Heffner /* Have enough data in the send queue to probe? */ 19897f9c33e5SIlpo Järvinen if (tp->write_seq - tp->snd_nxt < size_needed) 19905d424d5aSJohn Heffner return -1; 19915d424d5aSJohn Heffner 199291cc17c0SIlpo Järvinen if (tp->snd_wnd < size_needed) 19935d424d5aSJohn Heffner return -1; 199490840defSIlpo Järvinen if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) 19955d424d5aSJohn Heffner return 0; 19965d424d5aSJohn Heffner 1997d67c58e9SIlpo Järvinen /* Do we need to wait to drain cwnd? With none in flight, don't stall */ 1998d67c58e9SIlpo Järvinen if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { 1999d67c58e9SIlpo Järvinen if (!tcp_packets_in_flight(tp)) 20005d424d5aSJohn Heffner return -1; 20015d424d5aSJohn Heffner else 20025d424d5aSJohn Heffner return 0; 20035d424d5aSJohn Heffner } 20045d424d5aSJohn Heffner 20055d424d5aSJohn Heffner /* We're allowed to probe. Build it now. */ 2006eb934478SEric Dumazet nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false); 200751456b29SIan Morris if (!nskb) 20085d424d5aSJohn Heffner return -1; 20093ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 20103ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 20115d424d5aSJohn Heffner 2012fe067e8aSDavid S. Miller skb = tcp_send_head(sk); 20135d424d5aSJohn Heffner 20145d424d5aSJohn Heffner TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 20155d424d5aSJohn Heffner TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 20164de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; 20175d424d5aSJohn Heffner TCP_SKB_CB(nskb)->sacked = 0; 20185d424d5aSJohn Heffner nskb->csum = 0; 201984fa7933SPatrick McHardy nskb->ip_summed = skb->ip_summed; 20205d424d5aSJohn Heffner 202150c4817eSIlpo Järvinen tcp_insert_write_queue_before(nskb, skb, sk); 202250c4817eSIlpo Järvinen 20235d424d5aSJohn Heffner len = 0; 2024234b6860SIlpo Järvinen tcp_for_write_queue_from_safe(skb, next, sk) { 20255d424d5aSJohn Heffner copy = min_t(int, skb->len, probe_size - len); 20262fe664f1SDouglas Caetano dos Santos if (nskb->ip_summed) { 20275d424d5aSJohn Heffner skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); 20282fe664f1SDouglas Caetano dos Santos } else { 20292fe664f1SDouglas Caetano dos Santos __wsum csum = skb_copy_and_csum_bits(skb, 0, 2030056834d9SIlpo Järvinen skb_put(nskb, copy), 20312fe664f1SDouglas Caetano dos Santos copy, 0); 20322fe664f1SDouglas Caetano dos Santos nskb->csum = csum_block_add(nskb->csum, csum, len); 20332fe664f1SDouglas Caetano dos Santos } 20345d424d5aSJohn Heffner 20355d424d5aSJohn Heffner if (skb->len <= copy) { 20365d424d5aSJohn Heffner /* We've eaten all the data from this skb. 20375d424d5aSJohn Heffner * Throw it away. */ 20384de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 2039fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 20403ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 20415d424d5aSJohn Heffner } else { 20424de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & 2043a3433f35SChangli Gao ~(TCPHDR_FIN|TCPHDR_PSH); 20445d424d5aSJohn Heffner if (!skb_shinfo(skb)->nr_frags) { 20455d424d5aSJohn Heffner skb_pull(skb, copy); 204684fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 2047056834d9SIlpo Järvinen skb->csum = csum_partial(skb->data, 2048056834d9SIlpo Järvinen skb->len, 0); 20495d424d5aSJohn Heffner } else { 20505d424d5aSJohn Heffner __pskb_trim_head(skb, copy); 20515bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 20525d424d5aSJohn Heffner } 20535d424d5aSJohn Heffner TCP_SKB_CB(skb)->seq += copy; 20545d424d5aSJohn Heffner } 20555d424d5aSJohn Heffner 20565d424d5aSJohn Heffner len += copy; 2057234b6860SIlpo Järvinen 2058234b6860SIlpo Järvinen if (len >= probe_size) 2059234b6860SIlpo Järvinen break; 20605d424d5aSJohn Heffner } 20615bbb432cSEric Dumazet tcp_init_tso_segs(nskb, nskb->len); 20625d424d5aSJohn Heffner 20635d424d5aSJohn Heffner /* We're ready to send. If this fails, the probe will 20647faee5c0SEric Dumazet * be resegmented into mss-sized pieces by tcp_write_xmit(). 20657faee5c0SEric Dumazet */ 20665d424d5aSJohn Heffner if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 20675d424d5aSJohn Heffner /* Decrement cwnd here because we are sending 20685d424d5aSJohn Heffner * effectively two packets. */ 20695d424d5aSJohn Heffner tp->snd_cwnd--; 207066f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, nskb); 20715d424d5aSJohn Heffner 20725d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 20730e7b1368SJohn Heffner tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 20740e7b1368SJohn Heffner tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 20755d424d5aSJohn Heffner 20765d424d5aSJohn Heffner return 1; 20775d424d5aSJohn Heffner } 20785d424d5aSJohn Heffner 20795d424d5aSJohn Heffner return -1; 20805d424d5aSJohn Heffner } 20815d424d5aSJohn Heffner 2082f9616c35SEric Dumazet /* TCP Small Queues : 2083f9616c35SEric Dumazet * Control number of packets in qdisc/devices to two packets / or ~1 ms. 2084f9616c35SEric Dumazet * (These limits are doubled for retransmits) 2085f9616c35SEric Dumazet * This allows for : 2086f9616c35SEric Dumazet * - better RTT estimation and ACK scheduling 2087f9616c35SEric Dumazet * - faster recovery 2088f9616c35SEric Dumazet * - high rates 2089f9616c35SEric Dumazet * Alas, some drivers / subsystems require a fair amount 2090f9616c35SEric Dumazet * of queued bytes to ensure line rate. 2091f9616c35SEric Dumazet * One example is wifi aggregation (802.11 AMPDU) 2092f9616c35SEric Dumazet */ 2093f9616c35SEric Dumazet static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb, 2094f9616c35SEric Dumazet unsigned int factor) 2095f9616c35SEric Dumazet { 2096f9616c35SEric Dumazet unsigned int limit; 2097f9616c35SEric Dumazet 2098f9616c35SEric Dumazet limit = max(2 * skb->truesize, sk->sk_pacing_rate >> 10); 2099f9616c35SEric Dumazet limit = min_t(u32, limit, sysctl_tcp_limit_output_bytes); 2100f9616c35SEric Dumazet limit <<= factor; 2101f9616c35SEric Dumazet 2102f9616c35SEric Dumazet if (atomic_read(&sk->sk_wmem_alloc) > limit) { 210375eefc6cSEric Dumazet /* Always send the 1st or 2nd skb in write queue. 210475eefc6cSEric Dumazet * No need to wait for TX completion to call us back, 210575eefc6cSEric Dumazet * after softirq/tasklet schedule. 210675eefc6cSEric Dumazet * This helps when TX completions are delayed too much. 210775eefc6cSEric Dumazet */ 210875eefc6cSEric Dumazet if (skb == sk->sk_write_queue.next || 210975eefc6cSEric Dumazet skb->prev == sk->sk_write_queue.next) 211075eefc6cSEric Dumazet return false; 211175eefc6cSEric Dumazet 21127aa5470cSEric Dumazet set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags); 2113f9616c35SEric Dumazet /* It is possible TX completion already happened 2114f9616c35SEric Dumazet * before we set TSQ_THROTTLED, so we must 2115f9616c35SEric Dumazet * test again the condition. 2116f9616c35SEric Dumazet */ 2117f9616c35SEric Dumazet smp_mb__after_atomic(); 2118f9616c35SEric Dumazet if (atomic_read(&sk->sk_wmem_alloc) > limit) 2119f9616c35SEric Dumazet return true; 2120f9616c35SEric Dumazet } 2121f9616c35SEric Dumazet return false; 2122f9616c35SEric Dumazet } 2123f9616c35SEric Dumazet 212405b055e8SFrancis Yan static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new) 212505b055e8SFrancis Yan { 212605b055e8SFrancis Yan const u32 now = tcp_time_stamp; 212705b055e8SFrancis Yan 212805b055e8SFrancis Yan if (tp->chrono_type > TCP_CHRONO_UNSPEC) 212905b055e8SFrancis Yan tp->chrono_stat[tp->chrono_type - 1] += now - tp->chrono_start; 213005b055e8SFrancis Yan tp->chrono_start = now; 213105b055e8SFrancis Yan tp->chrono_type = new; 213205b055e8SFrancis Yan } 213305b055e8SFrancis Yan 213405b055e8SFrancis Yan void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type) 213505b055e8SFrancis Yan { 213605b055e8SFrancis Yan struct tcp_sock *tp = tcp_sk(sk); 213705b055e8SFrancis Yan 213805b055e8SFrancis Yan /* If there are multiple conditions worthy of tracking in a 21390f87230dSFrancis Yan * chronograph then the highest priority enum takes precedence 21400f87230dSFrancis Yan * over the other conditions. So that if something "more interesting" 214105b055e8SFrancis Yan * starts happening, stop the previous chrono and start a new one. 214205b055e8SFrancis Yan */ 214305b055e8SFrancis Yan if (type > tp->chrono_type) 214405b055e8SFrancis Yan tcp_chrono_set(tp, type); 214505b055e8SFrancis Yan } 214605b055e8SFrancis Yan 214705b055e8SFrancis Yan void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type) 214805b055e8SFrancis Yan { 214905b055e8SFrancis Yan struct tcp_sock *tp = tcp_sk(sk); 215005b055e8SFrancis Yan 21510f87230dSFrancis Yan 21520f87230dSFrancis Yan /* There are multiple conditions worthy of tracking in a 21530f87230dSFrancis Yan * chronograph, so that the highest priority enum takes 21540f87230dSFrancis Yan * precedence over the other conditions (see tcp_chrono_start). 21550f87230dSFrancis Yan * If a condition stops, we only stop chrono tracking if 21560f87230dSFrancis Yan * it's the "most interesting" or current chrono we are 21570f87230dSFrancis Yan * tracking and starts busy chrono if we have pending data. 21580f87230dSFrancis Yan */ 21590f87230dSFrancis Yan if (tcp_write_queue_empty(sk)) 216005b055e8SFrancis Yan tcp_chrono_set(tp, TCP_CHRONO_UNSPEC); 21610f87230dSFrancis Yan else if (type == tp->chrono_type) 21620f87230dSFrancis Yan tcp_chrono_set(tp, TCP_CHRONO_BUSY); 216305b055e8SFrancis Yan } 216405b055e8SFrancis Yan 21651da177e4SLinus Torvalds /* This routine writes packets to the network. It advances the 21661da177e4SLinus Torvalds * send_head. This happens as incoming acks open up the remote 21671da177e4SLinus Torvalds * window for us. 21681da177e4SLinus Torvalds * 2169f8269a49SIlpo Järvinen * LARGESEND note: !tcp_urg_mode is overkill, only frames between 2170f8269a49SIlpo Järvinen * snd_up-64k-mss .. snd_up cannot be large. However, taking into 2171f8269a49SIlpo Järvinen * account rare use of URG, this is not a big flaw. 2172f8269a49SIlpo Järvinen * 21736ba8a3b1SNandita Dukkipati * Send at most one packet when push_one > 0. Temporarily ignore 21746ba8a3b1SNandita Dukkipati * cwnd limit to force at most one packet out when push_one == 2. 21756ba8a3b1SNandita Dukkipati 2176a2a385d6SEric Dumazet * Returns true, if no segments are in flight and we have queued segments, 2177a2a385d6SEric Dumazet * but cannot send anything now because of SWS or another problem. 21781da177e4SLinus Torvalds */ 2179a2a385d6SEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 2180d5dd9175SIlpo Järvinen int push_one, gfp_t gfp) 21811da177e4SLinus Torvalds { 21821da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 218392df7b51SDavid S. Miller struct sk_buff *skb; 2184c1b4a7e6SDavid S. Miller unsigned int tso_segs, sent_pkts; 2185c1b4a7e6SDavid S. Miller int cwnd_quota; 21865d424d5aSJohn Heffner int result; 21875615f886SFrancis Yan bool is_cwnd_limited = false, is_rwnd_limited = false; 2188605ad7f1SEric Dumazet u32 max_segs; 21891da177e4SLinus Torvalds 2190c1b4a7e6SDavid S. Miller sent_pkts = 0; 21915d424d5aSJohn Heffner 2192d5dd9175SIlpo Järvinen if (!push_one) { 21935d424d5aSJohn Heffner /* Do MTU probing. */ 2194d5dd9175SIlpo Järvinen result = tcp_mtu_probe(sk); 2195d5dd9175SIlpo Järvinen if (!result) { 2196a2a385d6SEric Dumazet return false; 21975d424d5aSJohn Heffner } else if (result > 0) { 21985d424d5aSJohn Heffner sent_pkts = 1; 21995d424d5aSJohn Heffner } 2200d5dd9175SIlpo Järvinen } 22015d424d5aSJohn Heffner 2202ed6e7268SNeal Cardwell max_segs = tcp_tso_segs(sk, mss_now); 2203fe067e8aSDavid S. Miller while ((skb = tcp_send_head(sk))) { 2204c8ac3774SHerbert Xu unsigned int limit; 2205c8ac3774SHerbert Xu 22065bbb432cSEric Dumazet tso_segs = tcp_init_tso_segs(skb, mss_now); 2207c1b4a7e6SDavid S. Miller BUG_ON(!tso_segs); 2208c1b4a7e6SDavid S. Miller 22099d186cacSAndrey Vagin if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) { 22107faee5c0SEric Dumazet /* "skb_mstamp" is used as a start point for the retransmit timer */ 22117faee5c0SEric Dumazet skb_mstamp_get(&skb->skb_mstamp); 2212ec342325SAndrew Vagin goto repair; /* Skip network transmission */ 22139d186cacSAndrey Vagin } 2214ec342325SAndrew Vagin 2215b68e9f85SHerbert Xu cwnd_quota = tcp_cwnd_test(tp, skb); 22166ba8a3b1SNandita Dukkipati if (!cwnd_quota) { 22176ba8a3b1SNandita Dukkipati if (push_one == 2) 22186ba8a3b1SNandita Dukkipati /* Force out a loss probe pkt. */ 22196ba8a3b1SNandita Dukkipati cwnd_quota = 1; 22206ba8a3b1SNandita Dukkipati else 2221b68e9f85SHerbert Xu break; 22226ba8a3b1SNandita Dukkipati } 2223b68e9f85SHerbert Xu 22245615f886SFrancis Yan if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) { 22255615f886SFrancis Yan is_rwnd_limited = true; 2226b68e9f85SHerbert Xu break; 22275615f886SFrancis Yan } 2228b68e9f85SHerbert Xu 2229d6a4e26aSEric Dumazet if (tso_segs == 1) { 2230aa93466bSDavid S. Miller if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 2231aa93466bSDavid S. Miller (tcp_skb_is_last(sk, skb) ? 2232aa93466bSDavid S. Miller nonagle : TCP_NAGLE_PUSH)))) 2233aa93466bSDavid S. Miller break; 2234c1b4a7e6SDavid S. Miller } else { 2235ca8a2263SNeal Cardwell if (!push_one && 2236605ad7f1SEric Dumazet tcp_tso_should_defer(sk, skb, &is_cwnd_limited, 2237605ad7f1SEric Dumazet max_segs)) 2238aa93466bSDavid S. Miller break; 2239c1b4a7e6SDavid S. Miller } 2240aa93466bSDavid S. Miller 2241605ad7f1SEric Dumazet limit = mss_now; 2242d6a4e26aSEric Dumazet if (tso_segs > 1 && !tcp_urg_mode(tp)) 2243605ad7f1SEric Dumazet limit = tcp_mss_split_point(sk, skb, mss_now, 2244605ad7f1SEric Dumazet min_t(unsigned int, 2245605ad7f1SEric Dumazet cwnd_quota, 2246605ad7f1SEric Dumazet max_segs), 2247605ad7f1SEric Dumazet nonagle); 2248605ad7f1SEric Dumazet 2249605ad7f1SEric Dumazet if (skb->len > limit && 2250605ad7f1SEric Dumazet unlikely(tso_fragment(sk, skb, limit, mss_now, gfp))) 2251605ad7f1SEric Dumazet break; 2252605ad7f1SEric Dumazet 22537aa5470cSEric Dumazet if (test_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags)) 22547aa5470cSEric Dumazet clear_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags); 2255f9616c35SEric Dumazet if (tcp_small_queue_check(sk, skb, 0)) 225646d3ceabSEric Dumazet break; 2257c9eeec26SEric Dumazet 2258d5dd9175SIlpo Järvinen if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) 22591da177e4SLinus Torvalds break; 22601da177e4SLinus Torvalds 2261ec342325SAndrew Vagin repair: 22621da177e4SLinus Torvalds /* Advance the send_head. This one is sent out. 22631da177e4SLinus Torvalds * This call will increment packets_out. 22641da177e4SLinus Torvalds */ 226566f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 22661da177e4SLinus Torvalds 22671da177e4SLinus Torvalds tcp_minshall_update(tp, mss_now, skb); 2268a262f0cdSNandita Dukkipati sent_pkts += tcp_skb_pcount(skb); 2269d5dd9175SIlpo Järvinen 2270d5dd9175SIlpo Järvinen if (push_one) 2271d5dd9175SIlpo Järvinen break; 22721da177e4SLinus Torvalds } 22731da177e4SLinus Torvalds 22745615f886SFrancis Yan if (is_rwnd_limited) 22755615f886SFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED); 22765615f886SFrancis Yan else 22775615f886SFrancis Yan tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED); 22785615f886SFrancis Yan 2279aa93466bSDavid S. Miller if (likely(sent_pkts)) { 2280684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 2281684bad11SYuchung Cheng tp->prr_out += sent_pkts; 22826ba8a3b1SNandita Dukkipati 22836ba8a3b1SNandita Dukkipati /* Send one loss probe per tail loss episode. */ 22846ba8a3b1SNandita Dukkipati if (push_one != 2) 22856ba8a3b1SNandita Dukkipati tcp_schedule_loss_probe(sk); 2286d2e1339fSBendik Rønning Opstad is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd); 2287ca8a2263SNeal Cardwell tcp_cwnd_validate(sk, is_cwnd_limited); 2288a2a385d6SEric Dumazet return false; 22891da177e4SLinus Torvalds } 2290b340b264SYuchung Cheng return !tp->packets_out && tcp_send_head(sk); 22916ba8a3b1SNandita Dukkipati } 22926ba8a3b1SNandita Dukkipati 22936ba8a3b1SNandita Dukkipati bool tcp_schedule_loss_probe(struct sock *sk) 22946ba8a3b1SNandita Dukkipati { 22956ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 22966ba8a3b1SNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 22976ba8a3b1SNandita Dukkipati u32 timeout, tlp_time_stamp, rto_time_stamp; 2298740b0f18SEric Dumazet u32 rtt = usecs_to_jiffies(tp->srtt_us >> 3); 22996ba8a3b1SNandita Dukkipati 23006ba8a3b1SNandita Dukkipati /* No consecutive loss probes. */ 23016ba8a3b1SNandita Dukkipati if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) { 23026ba8a3b1SNandita Dukkipati tcp_rearm_rto(sk); 23036ba8a3b1SNandita Dukkipati return false; 23046ba8a3b1SNandita Dukkipati } 23056ba8a3b1SNandita Dukkipati /* Don't do any loss probe on a Fast Open connection before 3WHS 23066ba8a3b1SNandita Dukkipati * finishes. 23076ba8a3b1SNandita Dukkipati */ 2308f9b99582SYuchung Cheng if (tp->fastopen_rsk) 23096ba8a3b1SNandita Dukkipati return false; 23106ba8a3b1SNandita Dukkipati 23116ba8a3b1SNandita Dukkipati /* TLP is only scheduled when next timer event is RTO. */ 23126ba8a3b1SNandita Dukkipati if (icsk->icsk_pending != ICSK_TIME_RETRANS) 23136ba8a3b1SNandita Dukkipati return false; 23146ba8a3b1SNandita Dukkipati 23156ba8a3b1SNandita Dukkipati /* Schedule a loss probe in 2*RTT for SACK capable connections 23166ba8a3b1SNandita Dukkipati * in Open state, that are either limited by cwnd or application. 23176ba8a3b1SNandita Dukkipati */ 2318bec41a11SYuchung Cheng if ((sysctl_tcp_early_retrans != 3 && sysctl_tcp_early_retrans != 4) || 2319bec41a11SYuchung Cheng !tp->packets_out || !tcp_is_sack(tp) || 2320bec41a11SYuchung Cheng icsk->icsk_ca_state != TCP_CA_Open) 23216ba8a3b1SNandita Dukkipati return false; 23226ba8a3b1SNandita Dukkipati 23236ba8a3b1SNandita Dukkipati if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) && 23246ba8a3b1SNandita Dukkipati tcp_send_head(sk)) 23256ba8a3b1SNandita Dukkipati return false; 23266ba8a3b1SNandita Dukkipati 23276ba8a3b1SNandita Dukkipati /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account 2328f9b99582SYuchung Cheng * for delayed ack when there's one outstanding packet. If no RTT 2329f9b99582SYuchung Cheng * sample is available then probe after TCP_TIMEOUT_INIT. 23306ba8a3b1SNandita Dukkipati */ 2331f9b99582SYuchung Cheng timeout = rtt << 1 ? : TCP_TIMEOUT_INIT; 23326ba8a3b1SNandita Dukkipati if (tp->packets_out == 1) 23336ba8a3b1SNandita Dukkipati timeout = max_t(u32, timeout, 23346ba8a3b1SNandita Dukkipati (rtt + (rtt >> 1) + TCP_DELACK_MAX)); 23356ba8a3b1SNandita Dukkipati timeout = max_t(u32, timeout, msecs_to_jiffies(10)); 23366ba8a3b1SNandita Dukkipati 23376ba8a3b1SNandita Dukkipati /* If RTO is shorter, just schedule TLP in its place. */ 23386ba8a3b1SNandita Dukkipati tlp_time_stamp = tcp_time_stamp + timeout; 23396ba8a3b1SNandita Dukkipati rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout; 23406ba8a3b1SNandita Dukkipati if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) { 23416ba8a3b1SNandita Dukkipati s32 delta = rto_time_stamp - tcp_time_stamp; 23426ba8a3b1SNandita Dukkipati if (delta > 0) 23436ba8a3b1SNandita Dukkipati timeout = delta; 23446ba8a3b1SNandita Dukkipati } 23456ba8a3b1SNandita Dukkipati 23466ba8a3b1SNandita Dukkipati inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, 23476ba8a3b1SNandita Dukkipati TCP_RTO_MAX); 23486ba8a3b1SNandita Dukkipati return true; 23496ba8a3b1SNandita Dukkipati } 23506ba8a3b1SNandita Dukkipati 23511f3279aeSEric Dumazet /* Thanks to skb fast clones, we can detect if a prior transmit of 23521f3279aeSEric Dumazet * a packet is still in a qdisc or driver queue. 23531f3279aeSEric Dumazet * In this case, there is very little point doing a retransmit ! 23541f3279aeSEric Dumazet */ 23551f3279aeSEric Dumazet static bool skb_still_in_host_queue(const struct sock *sk, 23561f3279aeSEric Dumazet const struct sk_buff *skb) 23571f3279aeSEric Dumazet { 235839bb5e62SEric Dumazet if (unlikely(skb_fclone_busy(sk, skb))) { 2359c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), 23601f3279aeSEric Dumazet LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES); 23611f3279aeSEric Dumazet return true; 23621f3279aeSEric Dumazet } 23631f3279aeSEric Dumazet return false; 23641f3279aeSEric Dumazet } 23651f3279aeSEric Dumazet 2366b340b264SYuchung Cheng /* When probe timeout (PTO) fires, try send a new segment if possible, else 23676ba8a3b1SNandita Dukkipati * retransmit the last segment. 23686ba8a3b1SNandita Dukkipati */ 23696ba8a3b1SNandita Dukkipati void tcp_send_loss_probe(struct sock *sk) 23706ba8a3b1SNandita Dukkipati { 23719b717a8dSNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 23726ba8a3b1SNandita Dukkipati struct sk_buff *skb; 23736ba8a3b1SNandita Dukkipati int pcount; 23746ba8a3b1SNandita Dukkipati int mss = tcp_current_mss(sk); 23756ba8a3b1SNandita Dukkipati 2376b340b264SYuchung Cheng skb = tcp_send_head(sk); 2377b340b264SYuchung Cheng if (skb) { 2378b340b264SYuchung Cheng if (tcp_snd_wnd_test(tp, skb, mss)) { 2379b340b264SYuchung Cheng pcount = tp->packets_out; 2380b340b264SYuchung Cheng tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC); 2381b340b264SYuchung Cheng if (tp->packets_out > pcount) 2382b340b264SYuchung Cheng goto probe_sent; 23836ba8a3b1SNandita Dukkipati goto rearm_timer; 23846ba8a3b1SNandita Dukkipati } 2385b340b264SYuchung Cheng skb = tcp_write_queue_prev(sk, skb); 2386b340b264SYuchung Cheng } else { 2387b340b264SYuchung Cheng skb = tcp_write_queue_tail(sk); 2388b340b264SYuchung Cheng } 23896ba8a3b1SNandita Dukkipati 23909b717a8dSNandita Dukkipati /* At most one outstanding TLP retransmission. */ 23919b717a8dSNandita Dukkipati if (tp->tlp_high_seq) 23929b717a8dSNandita Dukkipati goto rearm_timer; 23939b717a8dSNandita Dukkipati 23946ba8a3b1SNandita Dukkipati /* Retransmit last segment. */ 23956ba8a3b1SNandita Dukkipati if (WARN_ON(!skb)) 23966ba8a3b1SNandita Dukkipati goto rearm_timer; 23976ba8a3b1SNandita Dukkipati 23981f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 23991f3279aeSEric Dumazet goto rearm_timer; 24001f3279aeSEric Dumazet 24016ba8a3b1SNandita Dukkipati pcount = tcp_skb_pcount(skb); 24026ba8a3b1SNandita Dukkipati if (WARN_ON(!pcount)) 24036ba8a3b1SNandita Dukkipati goto rearm_timer; 24046ba8a3b1SNandita Dukkipati 24056ba8a3b1SNandita Dukkipati if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) { 24066cc55e09SOctavian Purdila if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss, 24076cc55e09SOctavian Purdila GFP_ATOMIC))) 24086ba8a3b1SNandita Dukkipati goto rearm_timer; 2409b340b264SYuchung Cheng skb = tcp_write_queue_next(sk, skb); 24106ba8a3b1SNandita Dukkipati } 24116ba8a3b1SNandita Dukkipati 24126ba8a3b1SNandita Dukkipati if (WARN_ON(!skb || !tcp_skb_pcount(skb))) 24136ba8a3b1SNandita Dukkipati goto rearm_timer; 24146ba8a3b1SNandita Dukkipati 241510d3be56SEric Dumazet if (__tcp_retransmit_skb(sk, skb, 1)) 2416b340b264SYuchung Cheng goto rearm_timer; 24176ba8a3b1SNandita Dukkipati 24189b717a8dSNandita Dukkipati /* Record snd_nxt for loss detection. */ 24199b717a8dSNandita Dukkipati tp->tlp_high_seq = tp->snd_nxt; 24209b717a8dSNandita Dukkipati 2421b340b264SYuchung Cheng probe_sent: 2422c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES); 2423fcd16c0aSYuchung Cheng /* Reset s.t. tcp_rearm_rto will restart timer from now */ 2424fcd16c0aSYuchung Cheng inet_csk(sk)->icsk_pending = 0; 2425b340b264SYuchung Cheng rearm_timer: 2426fcd16c0aSYuchung Cheng tcp_rearm_rto(sk); 24271da177e4SLinus Torvalds } 24281da177e4SLinus Torvalds 2429a762a980SDavid S. Miller /* Push out any pending frames which were held back due to 2430a762a980SDavid S. Miller * TCP_CORK or attempt at coalescing tiny packets. 2431a762a980SDavid S. Miller * The socket must be locked by the caller. 2432a762a980SDavid S. Miller */ 24339e412ba7SIlpo Järvinen void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 24349e412ba7SIlpo Järvinen int nonagle) 2435a762a980SDavid S. Miller { 2436726e07a8SIlpo Järvinen /* If we are closed, the bytes will have to remain here. 2437726e07a8SIlpo Järvinen * In time closedown will finish, we empty the write queue and 2438726e07a8SIlpo Järvinen * all will be happy. 2439726e07a8SIlpo Järvinen */ 2440726e07a8SIlpo Järvinen if (unlikely(sk->sk_state == TCP_CLOSE)) 2441726e07a8SIlpo Järvinen return; 2442726e07a8SIlpo Järvinen 244399a1dec7SMel Gorman if (tcp_write_xmit(sk, cur_mss, nonagle, 0, 24447450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC))) 24459e412ba7SIlpo Järvinen tcp_check_probe_timer(sk); 2446a762a980SDavid S. Miller } 2447a762a980SDavid S. Miller 2448c1b4a7e6SDavid S. Miller /* Send _single_ skb sitting at the send head. This function requires 2449c1b4a7e6SDavid S. Miller * true push pending frames to setup probe timer etc. 2450c1b4a7e6SDavid S. Miller */ 2451c1b4a7e6SDavid S. Miller void tcp_push_one(struct sock *sk, unsigned int mss_now) 2452c1b4a7e6SDavid S. Miller { 2453fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 2454c1b4a7e6SDavid S. Miller 2455c1b4a7e6SDavid S. Miller BUG_ON(!skb || skb->len < mss_now); 2456c1b4a7e6SDavid S. Miller 2457d5dd9175SIlpo Järvinen tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); 2458c1b4a7e6SDavid S. Miller } 2459c1b4a7e6SDavid S. Miller 24601da177e4SLinus Torvalds /* This function returns the amount that we can raise the 24611da177e4SLinus Torvalds * usable window based on the following constraints 24621da177e4SLinus Torvalds * 24631da177e4SLinus Torvalds * 1. The window can never be shrunk once it is offered (RFC 793) 24641da177e4SLinus Torvalds * 2. We limit memory per socket 24651da177e4SLinus Torvalds * 24661da177e4SLinus Torvalds * RFC 1122: 24671da177e4SLinus Torvalds * "the suggested [SWS] avoidance algorithm for the receiver is to keep 24681da177e4SLinus Torvalds * RECV.NEXT + RCV.WIN fixed until: 24691da177e4SLinus Torvalds * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 24701da177e4SLinus Torvalds * 24711da177e4SLinus Torvalds * i.e. don't raise the right edge of the window until you can raise 24721da177e4SLinus Torvalds * it at least MSS bytes. 24731da177e4SLinus Torvalds * 24741da177e4SLinus Torvalds * Unfortunately, the recommended algorithm breaks header prediction, 24751da177e4SLinus Torvalds * since header prediction assumes th->window stays fixed. 24761da177e4SLinus Torvalds * 24771da177e4SLinus Torvalds * Strictly speaking, keeping th->window fixed violates the receiver 24781da177e4SLinus Torvalds * side SWS prevention criteria. The problem is that under this rule 24791da177e4SLinus Torvalds * a stream of single byte packets will cause the right side of the 24801da177e4SLinus Torvalds * window to always advance by a single byte. 24811da177e4SLinus Torvalds * 24821da177e4SLinus Torvalds * Of course, if the sender implements sender side SWS prevention 24831da177e4SLinus Torvalds * then this will not be a problem. 24841da177e4SLinus Torvalds * 24851da177e4SLinus Torvalds * BSD seems to make the following compromise: 24861da177e4SLinus Torvalds * 24871da177e4SLinus Torvalds * If the free space is less than the 1/4 of the maximum 24881da177e4SLinus Torvalds * space available and the free space is less than 1/2 mss, 24891da177e4SLinus Torvalds * then set the window to 0. 24901da177e4SLinus Torvalds * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 24911da177e4SLinus Torvalds * Otherwise, just prevent the window from shrinking 24921da177e4SLinus Torvalds * and from being larger than the largest representable value. 24931da177e4SLinus Torvalds * 24941da177e4SLinus Torvalds * This prevents incremental opening of the window in the regime 24951da177e4SLinus Torvalds * where TCP is limited by the speed of the reader side taking 24961da177e4SLinus Torvalds * data out of the TCP receive queue. It does nothing about 24971da177e4SLinus Torvalds * those cases where the window is constrained on the sender side 24981da177e4SLinus Torvalds * because the pipeline is full. 24991da177e4SLinus Torvalds * 25001da177e4SLinus Torvalds * BSD also seems to "accidentally" limit itself to windows that are a 25011da177e4SLinus Torvalds * multiple of MSS, at least until the free space gets quite small. 25021da177e4SLinus Torvalds * This would appear to be a side effect of the mbuf implementation. 25031da177e4SLinus Torvalds * Combining these two algorithms results in the observed behavior 25041da177e4SLinus Torvalds * of having a fixed window size at almost all times. 25051da177e4SLinus Torvalds * 25061da177e4SLinus Torvalds * Below we obtain similar behavior by forcing the offered window to 25071da177e4SLinus Torvalds * a multiple of the mss when it is feasible to do so. 25081da177e4SLinus Torvalds * 25091da177e4SLinus Torvalds * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 25101da177e4SLinus Torvalds * Regular options like TIMESTAMP are taken into account. 25111da177e4SLinus Torvalds */ 25121da177e4SLinus Torvalds u32 __tcp_select_window(struct sock *sk) 25131da177e4SLinus Torvalds { 2514463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 25151da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2516caa20d9aSStephen Hemminger /* MSS for the peer's data. Previous versions used mss_clamp 25171da177e4SLinus Torvalds * here. I don't know if the value based on our guesses 25181da177e4SLinus Torvalds * of peer's MSS is better for the performance. It's more correct 25191da177e4SLinus Torvalds * but may be worse for the performance because of rcv_mss 25201da177e4SLinus Torvalds * fluctuations. --SAW 1998/11/1 25211da177e4SLinus Torvalds */ 2522463c84b9SArnaldo Carvalho de Melo int mss = icsk->icsk_ack.rcv_mss; 25231da177e4SLinus Torvalds int free_space = tcp_space(sk); 252486c1a045SFlorian Westphal int allowed_space = tcp_full_space(sk); 252586c1a045SFlorian Westphal int full_space = min_t(int, tp->window_clamp, allowed_space); 25261da177e4SLinus Torvalds int window; 25271da177e4SLinus Torvalds 252806425c30SEric Dumazet if (unlikely(mss > full_space)) { 25291da177e4SLinus Torvalds mss = full_space; 253006425c30SEric Dumazet if (mss <= 0) 253106425c30SEric Dumazet return 0; 253206425c30SEric Dumazet } 2533b92edbe0SEric Dumazet if (free_space < (full_space >> 1)) { 2534463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.quick = 0; 25351da177e4SLinus Torvalds 2536b8da51ebSEric Dumazet if (tcp_under_memory_pressure(sk)) 2537056834d9SIlpo Järvinen tp->rcv_ssthresh = min(tp->rcv_ssthresh, 2538056834d9SIlpo Järvinen 4U * tp->advmss); 25391da177e4SLinus Torvalds 254086c1a045SFlorian Westphal /* free_space might become our new window, make sure we don't 254186c1a045SFlorian Westphal * increase it due to wscale. 254286c1a045SFlorian Westphal */ 254386c1a045SFlorian Westphal free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale); 254486c1a045SFlorian Westphal 254586c1a045SFlorian Westphal /* if free space is less than mss estimate, or is below 1/16th 254686c1a045SFlorian Westphal * of the maximum allowed, try to move to zero-window, else 254786c1a045SFlorian Westphal * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and 254886c1a045SFlorian Westphal * new incoming data is dropped due to memory limits. 254986c1a045SFlorian Westphal * With large window, mss test triggers way too late in order 255086c1a045SFlorian Westphal * to announce zero window in time before rmem limit kicks in. 255186c1a045SFlorian Westphal */ 255286c1a045SFlorian Westphal if (free_space < (allowed_space >> 4) || free_space < mss) 25531da177e4SLinus Torvalds return 0; 25541da177e4SLinus Torvalds } 25551da177e4SLinus Torvalds 25561da177e4SLinus Torvalds if (free_space > tp->rcv_ssthresh) 25571da177e4SLinus Torvalds free_space = tp->rcv_ssthresh; 25581da177e4SLinus Torvalds 25591da177e4SLinus Torvalds /* Don't do rounding if we are using window scaling, since the 25601da177e4SLinus Torvalds * scaled window will not line up with the MSS boundary anyway. 25611da177e4SLinus Torvalds */ 25621da177e4SLinus Torvalds if (tp->rx_opt.rcv_wscale) { 25631da177e4SLinus Torvalds window = free_space; 25641da177e4SLinus Torvalds 25651da177e4SLinus Torvalds /* Advertise enough space so that it won't get scaled away. 25661da177e4SLinus Torvalds * Import case: prevent zero window announcement if 25671da177e4SLinus Torvalds * 1<<rcv_wscale > mss. 25681da177e4SLinus Torvalds */ 25691935299dSGao Feng window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale)); 25701da177e4SLinus Torvalds } else { 25711935299dSGao Feng window = tp->rcv_wnd; 25721da177e4SLinus Torvalds /* Get the largest window that is a nice multiple of mss. 25731da177e4SLinus Torvalds * Window clamp already applied above. 25741da177e4SLinus Torvalds * If our current window offering is within 1 mss of the 25751da177e4SLinus Torvalds * free space we just keep it. This prevents the divide 25761da177e4SLinus Torvalds * and multiply from happening most of the time. 25771da177e4SLinus Torvalds * We also don't do any window rounding when the free space 25781da177e4SLinus Torvalds * is too small. 25791da177e4SLinus Torvalds */ 25801da177e4SLinus Torvalds if (window <= free_space - mss || window > free_space) 25811935299dSGao Feng window = rounddown(free_space, mss); 258284565070SJohn Heffner else if (mss == full_space && 2583b92edbe0SEric Dumazet free_space > window + (full_space >> 1)) 258484565070SJohn Heffner window = free_space; 25851da177e4SLinus Torvalds } 25861da177e4SLinus Torvalds 25871da177e4SLinus Torvalds return window; 25881da177e4SLinus Torvalds } 25891da177e4SLinus Torvalds 2590cfea5a68SMartin KaFai Lau void tcp_skb_collapse_tstamp(struct sk_buff *skb, 2591082ac2d5SMartin KaFai Lau const struct sk_buff *next_skb) 2592082ac2d5SMartin KaFai Lau { 25930a2cf20cSSoheil Hassas Yeganeh if (unlikely(tcp_has_tx_tstamp(next_skb))) { 25940a2cf20cSSoheil Hassas Yeganeh const struct skb_shared_info *next_shinfo = 25950a2cf20cSSoheil Hassas Yeganeh skb_shinfo(next_skb); 2596082ac2d5SMartin KaFai Lau struct skb_shared_info *shinfo = skb_shinfo(skb); 2597082ac2d5SMartin KaFai Lau 25980a2cf20cSSoheil Hassas Yeganeh shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP; 2599082ac2d5SMartin KaFai Lau shinfo->tskey = next_shinfo->tskey; 26002de8023eSMartin KaFai Lau TCP_SKB_CB(skb)->txstamp_ack |= 26012de8023eSMartin KaFai Lau TCP_SKB_CB(next_skb)->txstamp_ack; 2602082ac2d5SMartin KaFai Lau } 2603082ac2d5SMartin KaFai Lau } 2604082ac2d5SMartin KaFai Lau 26054a17fc3aSIlpo Järvinen /* Collapses two adjacent SKB's during retransmission. */ 2606f8071cdeSEric Dumazet static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) 26071da177e4SLinus Torvalds { 26081da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2609fe067e8aSDavid S. Miller struct sk_buff *next_skb = tcp_write_queue_next(sk, skb); 2610058dc334SIlpo Järvinen int skb_size, next_skb_size; 26111da177e4SLinus Torvalds 2612058dc334SIlpo Järvinen skb_size = skb->len; 2613058dc334SIlpo Järvinen next_skb_size = next_skb->len; 26141da177e4SLinus Torvalds 2615058dc334SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); 26161da177e4SLinus Torvalds 2617f8071cdeSEric Dumazet if (next_skb_size) { 2618f8071cdeSEric Dumazet if (next_skb_size <= skb_availroom(skb)) 2619f8071cdeSEric Dumazet skb_copy_bits(next_skb, 0, skb_put(skb, next_skb_size), 2620f8071cdeSEric Dumazet next_skb_size); 2621f8071cdeSEric Dumazet else if (!skb_shift(skb, next_skb, next_skb_size)) 2622f8071cdeSEric Dumazet return false; 2623f8071cdeSEric Dumazet } 26246859d494SIlpo Järvinen tcp_highest_sack_combine(sk, next_skb, skb); 2625a6963a6bSIlpo Järvinen 2626fe067e8aSDavid S. Miller tcp_unlink_write_queue(next_skb, sk); 26271da177e4SLinus Torvalds 262852d570aaSJarek Poplawski if (next_skb->ip_summed == CHECKSUM_PARTIAL) 262952d570aaSJarek Poplawski skb->ip_summed = CHECKSUM_PARTIAL; 26301da177e4SLinus Torvalds 263184fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 26321da177e4SLinus Torvalds skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); 26331da177e4SLinus Torvalds 26341da177e4SLinus Torvalds /* Update sequence range on original skb. */ 26351da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 26361da177e4SLinus Torvalds 2637e6c7d085SIlpo Järvinen /* Merge over control information. This moves PSH/FIN etc. over */ 26384de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; 26391da177e4SLinus Torvalds 26401da177e4SLinus Torvalds /* All done, get rid of second SKB and account for it so 26411da177e4SLinus Torvalds * packet counting does not break. 26421da177e4SLinus Torvalds */ 26434828e7f4SIlpo Järvinen TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; 2644a643b5d4SMartin KaFai Lau TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor; 2645b7689205SIlpo Järvinen 2646b7689205SIlpo Järvinen /* changed transmit queue under us so clear hints */ 2647ef9da47cSIlpo Järvinen tcp_clear_retrans_hints_partial(tp); 2648ef9da47cSIlpo Järvinen if (next_skb == tp->retransmit_skb_hint) 2649ef9da47cSIlpo Järvinen tp->retransmit_skb_hint = skb; 2650b7689205SIlpo Järvinen 2651797108d1SIlpo Järvinen tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); 2652797108d1SIlpo Järvinen 2653082ac2d5SMartin KaFai Lau tcp_skb_collapse_tstamp(skb, next_skb); 2654082ac2d5SMartin KaFai Lau 26553ab224beSHideo Aoki sk_wmem_free_skb(sk, next_skb); 2656f8071cdeSEric Dumazet return true; 26571da177e4SLinus Torvalds } 26581da177e4SLinus Torvalds 265967edfef7SAndi Kleen /* Check if coalescing SKBs is legal. */ 2660a2a385d6SEric Dumazet static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) 26614a17fc3aSIlpo Järvinen { 26624a17fc3aSIlpo Järvinen if (tcp_skb_pcount(skb) > 1) 2663a2a385d6SEric Dumazet return false; 26644a17fc3aSIlpo Järvinen if (skb_cloned(skb)) 2665a2a385d6SEric Dumazet return false; 26664a17fc3aSIlpo Järvinen if (skb == tcp_send_head(sk)) 2667a2a385d6SEric Dumazet return false; 26682331ccc5SEric Dumazet /* Some heuristics for collapsing over SACK'd could be invented */ 26694a17fc3aSIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 2670a2a385d6SEric Dumazet return false; 26714a17fc3aSIlpo Järvinen 2672a2a385d6SEric Dumazet return true; 26734a17fc3aSIlpo Järvinen } 26744a17fc3aSIlpo Järvinen 267567edfef7SAndi Kleen /* Collapse packets in the retransmit queue to make to create 267667edfef7SAndi Kleen * less packets on the wire. This is only done on retransmission. 267767edfef7SAndi Kleen */ 26784a17fc3aSIlpo Järvinen static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, 26794a17fc3aSIlpo Järvinen int space) 26804a17fc3aSIlpo Järvinen { 26814a17fc3aSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 26824a17fc3aSIlpo Järvinen struct sk_buff *skb = to, *tmp; 2683a2a385d6SEric Dumazet bool first = true; 26844a17fc3aSIlpo Järvinen 26854a17fc3aSIlpo Järvinen if (!sysctl_tcp_retrans_collapse) 26864a17fc3aSIlpo Järvinen return; 26874de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 26884a17fc3aSIlpo Järvinen return; 26894a17fc3aSIlpo Järvinen 26904a17fc3aSIlpo Järvinen tcp_for_write_queue_from_safe(skb, tmp, sk) { 26914a17fc3aSIlpo Järvinen if (!tcp_can_collapse(sk, skb)) 26924a17fc3aSIlpo Järvinen break; 26934a17fc3aSIlpo Järvinen 2694a643b5d4SMartin KaFai Lau if (!tcp_skb_can_collapse_to(to)) 2695a643b5d4SMartin KaFai Lau break; 2696a643b5d4SMartin KaFai Lau 26974a17fc3aSIlpo Järvinen space -= skb->len; 26984a17fc3aSIlpo Järvinen 26994a17fc3aSIlpo Järvinen if (first) { 2700a2a385d6SEric Dumazet first = false; 27014a17fc3aSIlpo Järvinen continue; 27024a17fc3aSIlpo Järvinen } 27034a17fc3aSIlpo Järvinen 27044a17fc3aSIlpo Järvinen if (space < 0) 27054a17fc3aSIlpo Järvinen break; 27064a17fc3aSIlpo Järvinen 27074a17fc3aSIlpo Järvinen if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) 27084a17fc3aSIlpo Järvinen break; 27094a17fc3aSIlpo Järvinen 2710f8071cdeSEric Dumazet if (!tcp_collapse_retrans(sk, to)) 2711f8071cdeSEric Dumazet break; 27124a17fc3aSIlpo Järvinen } 27134a17fc3aSIlpo Järvinen } 27144a17fc3aSIlpo Järvinen 27151da177e4SLinus Torvalds /* This retransmits one SKB. Policy decisions and retransmit queue 27161da177e4SLinus Torvalds * state updates are done by the caller. Returns non-zero if an 27171da177e4SLinus Torvalds * error occurred which prevented the send. 27181da177e4SLinus Torvalds */ 271910d3be56SEric Dumazet int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) 27201da177e4SLinus Torvalds { 27215d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 272210d3be56SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 27237d227cd2SSridhar Samudrala unsigned int cur_mss; 272410d3be56SEric Dumazet int diff, len, err; 27251da177e4SLinus Torvalds 272610d3be56SEric Dumazet 272710d3be56SEric Dumazet /* Inconclusive MTU probe */ 272810d3be56SEric Dumazet if (icsk->icsk_mtup.probe_size) 27295d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 27305d424d5aSJohn Heffner 27311da177e4SLinus Torvalds /* Do not sent more than we queued. 1/4 is reserved for possible 2732caa20d9aSStephen Hemminger * copying overhead: fragmentation, tunneling, mangling etc. 27331da177e4SLinus Torvalds */ 27341da177e4SLinus Torvalds if (atomic_read(&sk->sk_wmem_alloc) > 2735ffb4d6c8SEric Dumazet min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), 2736ffb4d6c8SEric Dumazet sk->sk_sndbuf)) 27371da177e4SLinus Torvalds return -EAGAIN; 27381da177e4SLinus Torvalds 27391f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 27401f3279aeSEric Dumazet return -EBUSY; 27411f3279aeSEric Dumazet 27421da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 27431da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 27441da177e4SLinus Torvalds BUG(); 27451da177e4SLinus Torvalds if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 27461da177e4SLinus Torvalds return -ENOMEM; 27471da177e4SLinus Torvalds } 27481da177e4SLinus Torvalds 27497d227cd2SSridhar Samudrala if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 27507d227cd2SSridhar Samudrala return -EHOSTUNREACH; /* Routing failure or similar. */ 27517d227cd2SSridhar Samudrala 27520c54b85fSIlpo Järvinen cur_mss = tcp_current_mss(sk); 27537d227cd2SSridhar Samudrala 27541da177e4SLinus Torvalds /* If receiver has shrunk his window, and skb is out of 27551da177e4SLinus Torvalds * new window, do not retransmit it. The exception is the 27561da177e4SLinus Torvalds * case, when window is shrunk to zero. In this case 27571da177e4SLinus Torvalds * our retransmit serves as a zero window probe. 27581da177e4SLinus Torvalds */ 27599d4fb27dSJoe Perches if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && 27609d4fb27dSJoe Perches TCP_SKB_CB(skb)->seq != tp->snd_una) 27611da177e4SLinus Torvalds return -EAGAIN; 27621da177e4SLinus Torvalds 276310d3be56SEric Dumazet len = cur_mss * segs; 276410d3be56SEric Dumazet if (skb->len > len) { 276510d3be56SEric Dumazet if (tcp_fragment(sk, skb, len, cur_mss, GFP_ATOMIC)) 27661da177e4SLinus Torvalds return -ENOMEM; /* We'll try again later. */ 276702276f3cSIlpo Järvinen } else { 2768c52e2421SEric Dumazet if (skb_unclone(skb, GFP_ATOMIC)) 2769c52e2421SEric Dumazet return -ENOMEM; 277010d3be56SEric Dumazet 277110d3be56SEric Dumazet diff = tcp_skb_pcount(skb); 277210d3be56SEric Dumazet tcp_set_skb_tso_segs(skb, cur_mss); 277310d3be56SEric Dumazet diff -= tcp_skb_pcount(skb); 277410d3be56SEric Dumazet if (diff) 277510d3be56SEric Dumazet tcp_adjust_pcount(sk, skb, diff); 277610d3be56SEric Dumazet if (skb->len < cur_mss) 277710d3be56SEric Dumazet tcp_retrans_try_collapse(sk, skb, cur_mss); 27781da177e4SLinus Torvalds } 27791da177e4SLinus Torvalds 278049213555SDaniel Borkmann /* RFC3168, section 6.1.1.1. ECN fallback */ 278149213555SDaniel Borkmann if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN) 278249213555SDaniel Borkmann tcp_ecn_clear_syn(sk, skb); 278349213555SDaniel Borkmann 2784678550c6SYuchung Cheng /* Update global and local TCP statistics. */ 2785678550c6SYuchung Cheng segs = tcp_skb_pcount(skb); 2786678550c6SYuchung Cheng TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs); 2787678550c6SYuchung Cheng if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 2788678550c6SYuchung Cheng __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 2789678550c6SYuchung Cheng tp->total_retrans += segs; 2790678550c6SYuchung Cheng 279150bceae9SThomas Graf /* make sure skb->data is aligned on arches that require it 279250bceae9SThomas Graf * and check if ack-trimming & collapsing extended the headroom 279350bceae9SThomas Graf * beyond what csum_start can cover. 279450bceae9SThomas Graf */ 279550bceae9SThomas Graf if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) || 279650bceae9SThomas Graf skb_headroom(skb) >= 0xFFFF)) { 279710a81980SEric Dumazet struct sk_buff *nskb; 279810a81980SEric Dumazet 279910a81980SEric Dumazet skb_mstamp_get(&skb->skb_mstamp); 280010a81980SEric Dumazet nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC); 2801c84a5711SYuchung Cheng err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) : 2802117632e6SEric Dumazet -ENOBUFS; 2803117632e6SEric Dumazet } else { 2804c84a5711SYuchung Cheng err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 2805117632e6SEric Dumazet } 2806c84a5711SYuchung Cheng 2807fc9f3501SEric Dumazet if (likely(!err)) { 2808c84a5711SYuchung Cheng TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS; 2809678550c6SYuchung Cheng } else if (err != -EBUSY) { 2810678550c6SYuchung Cheng NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL); 2811fc9f3501SEric Dumazet } 2812c84a5711SYuchung Cheng return err; 281393b174adSYuchung Cheng } 281493b174adSYuchung Cheng 281510d3be56SEric Dumazet int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) 281693b174adSYuchung Cheng { 281793b174adSYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 281810d3be56SEric Dumazet int err = __tcp_retransmit_skb(sk, skb, segs); 28191da177e4SLinus Torvalds 28201da177e4SLinus Torvalds if (err == 0) { 28211da177e4SLinus Torvalds #if FASTRETRANS_DEBUG > 0 28221da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 2823e87cc472SJoe Perches net_dbg_ratelimited("retrans_out leaked\n"); 28241da177e4SLinus Torvalds } 28251da177e4SLinus Torvalds #endif 28261da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 28271da177e4SLinus Torvalds tp->retrans_out += tcp_skb_pcount(skb); 28281da177e4SLinus Torvalds 28291da177e4SLinus Torvalds /* Save stamp of the first retransmit. */ 28301da177e4SLinus Torvalds if (!tp->retrans_stamp) 28317faee5c0SEric Dumazet tp->retrans_stamp = tcp_skb_timestamp(skb); 28321da177e4SLinus Torvalds 28331da177e4SLinus Torvalds } 28346e08d5e3SYuchung Cheng 28356e08d5e3SYuchung Cheng if (tp->undo_retrans < 0) 28366e08d5e3SYuchung Cheng tp->undo_retrans = 0; 28376e08d5e3SYuchung Cheng tp->undo_retrans += tcp_skb_pcount(skb); 28381da177e4SLinus Torvalds return err; 28391da177e4SLinus Torvalds } 28401da177e4SLinus Torvalds 28411da177e4SLinus Torvalds /* This gets called after a retransmit timeout, and the initially 28421da177e4SLinus Torvalds * retransmitted data is acknowledged. It tries to continue 28431da177e4SLinus Torvalds * resending the rest of the retransmit queue, until either 28441da177e4SLinus Torvalds * we've sent it all or the congestion window limit is reached. 28451da177e4SLinus Torvalds * If doing SACK, the first ACK which comes back for a timeout 28461da177e4SLinus Torvalds * based retransmit packet might feed us FACK information again. 28471da177e4SLinus Torvalds * If so, we use it to avoid unnecessarily retransmissions. 28481da177e4SLinus Torvalds */ 28491da177e4SLinus Torvalds void tcp_xmit_retransmit_queue(struct sock *sk) 28501da177e4SLinus Torvalds { 28516687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 28521da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 28531da177e4SLinus Torvalds struct sk_buff *skb; 28540e1c54c2SIlpo Järvinen struct sk_buff *hole = NULL; 2855840a3cbeSYuchung Cheng u32 max_segs; 285661eb55f4SIlpo Järvinen int mib_idx; 28576a438bbeSStephen Hemminger 285845e77d31SIlpo Järvinen if (!tp->packets_out) 285945e77d31SIlpo Järvinen return; 286045e77d31SIlpo Järvinen 2861618d9f25SIlpo Järvinen if (tp->retransmit_skb_hint) { 28626a438bbeSStephen Hemminger skb = tp->retransmit_skb_hint; 2863618d9f25SIlpo Järvinen } else { 2864fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 2865618d9f25SIlpo Järvinen } 28661da177e4SLinus Torvalds 2867ed6e7268SNeal Cardwell max_segs = tcp_tso_segs(sk, tcp_current_mss(sk)); 2868fe067e8aSDavid S. Miller tcp_for_write_queue_from(skb, sk) { 2869dca0aaf8SEric Dumazet __u8 sacked; 287010d3be56SEric Dumazet int segs; 28711da177e4SLinus Torvalds 2872fe067e8aSDavid S. Miller if (skb == tcp_send_head(sk)) 2873fe067e8aSDavid S. Miller break; 28746a438bbeSStephen Hemminger /* we could do better than to assign each time */ 287551456b29SIan Morris if (!hole) 28766a438bbeSStephen Hemminger tp->retransmit_skb_hint = skb; 28776a438bbeSStephen Hemminger 287810d3be56SEric Dumazet segs = tp->snd_cwnd - tcp_packets_in_flight(tp); 287910d3be56SEric Dumazet if (segs <= 0) 28801da177e4SLinus Torvalds return; 2881dca0aaf8SEric Dumazet sacked = TCP_SKB_CB(skb)->sacked; 2882a3d2e9f8SEric Dumazet /* In case tcp_shift_skb_data() have aggregated large skbs, 2883a3d2e9f8SEric Dumazet * we need to make sure not sending too bigs TSO packets 2884a3d2e9f8SEric Dumazet */ 2885a3d2e9f8SEric Dumazet segs = min_t(int, segs, max_segs); 28860e1c54c2SIlpo Järvinen 2887840a3cbeSYuchung Cheng if (tp->retrans_out >= tp->lost_out) { 2888006f582cSIlpo Järvinen break; 28890e1c54c2SIlpo Järvinen } else if (!(sacked & TCPCB_LOST)) { 289051456b29SIan Morris if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) 28910e1c54c2SIlpo Järvinen hole = skb; 289261eb55f4SIlpo Järvinen continue; 28931da177e4SLinus Torvalds 28940e1c54c2SIlpo Järvinen } else { 28950e1c54c2SIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Loss) 28960e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFASTRETRANS; 28970e1c54c2SIlpo Järvinen else 28980e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; 28990e1c54c2SIlpo Järvinen } 29000e1c54c2SIlpo Järvinen 29010e1c54c2SIlpo Järvinen if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) 290261eb55f4SIlpo Järvinen continue; 290340b215e5SPavel Emelyanov 2904f9616c35SEric Dumazet if (tcp_small_queue_check(sk, skb, 1)) 2905f9616c35SEric Dumazet return; 2906f9616c35SEric Dumazet 290710d3be56SEric Dumazet if (tcp_retransmit_skb(sk, skb, segs)) 29081da177e4SLinus Torvalds return; 290924ab6becSYuchung Cheng 2910de1d6578SYuchung Cheng NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb)); 29111da177e4SLinus Torvalds 2912684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 2913a262f0cdSNandita Dukkipati tp->prr_out += tcp_skb_pcount(skb); 2914a262f0cdSNandita Dukkipati 291557dde7f7SYuchung Cheng if (skb == tcp_write_queue_head(sk) && 291657dde7f7SYuchung Cheng icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT) 2917463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 29183f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, 29193f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 29201da177e4SLinus Torvalds } 29211da177e4SLinus Torvalds } 29221da177e4SLinus Torvalds 2923d83769a5SEric Dumazet /* We allow to exceed memory limits for FIN packets to expedite 2924d83769a5SEric Dumazet * connection tear down and (memory) recovery. 2925845704a5SEric Dumazet * Otherwise tcp_send_fin() could be tempted to either delay FIN 2926845704a5SEric Dumazet * or even be forced to close flow without any FIN. 2927a6c5ea4cSEric Dumazet * In general, we want to allow one skb per socket to avoid hangs 2928a6c5ea4cSEric Dumazet * with edge trigger epoll() 2929d83769a5SEric Dumazet */ 2930a6c5ea4cSEric Dumazet void sk_forced_mem_schedule(struct sock *sk, int size) 2931d83769a5SEric Dumazet { 2932e805605cSJohannes Weiner int amt; 2933d83769a5SEric Dumazet 2934d83769a5SEric Dumazet if (size <= sk->sk_forward_alloc) 2935d83769a5SEric Dumazet return; 2936d83769a5SEric Dumazet amt = sk_mem_pages(size); 2937d83769a5SEric Dumazet sk->sk_forward_alloc += amt * SK_MEM_QUANTUM; 2938e805605cSJohannes Weiner sk_memory_allocated_add(sk, amt); 2939e805605cSJohannes Weiner 2940baac50bbSJohannes Weiner if (mem_cgroup_sockets_enabled && sk->sk_memcg) 2941baac50bbSJohannes Weiner mem_cgroup_charge_skmem(sk->sk_memcg, amt); 2942d83769a5SEric Dumazet } 2943d83769a5SEric Dumazet 2944845704a5SEric Dumazet /* Send a FIN. The caller locks the socket for us. 2945845704a5SEric Dumazet * We should try to send a FIN packet really hard, but eventually give up. 29461da177e4SLinus Torvalds */ 29471da177e4SLinus Torvalds void tcp_send_fin(struct sock *sk) 29481da177e4SLinus Torvalds { 2949845704a5SEric Dumazet struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk); 29501da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 29511da177e4SLinus Torvalds 2952845704a5SEric Dumazet /* Optimization, tack on the FIN if we have one skb in write queue and 2953845704a5SEric Dumazet * this skb was not yet sent, or we are under memory pressure. 2954845704a5SEric Dumazet * Note: in the latter case, FIN packet will be sent after a timeout, 2955845704a5SEric Dumazet * as TCP stack thinks it has already been transmitted. 29561da177e4SLinus Torvalds */ 2957b8da51ebSEric Dumazet if (tskb && (tcp_send_head(sk) || tcp_under_memory_pressure(sk))) { 2958845704a5SEric Dumazet coalesce: 2959845704a5SEric Dumazet TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN; 2960845704a5SEric Dumazet TCP_SKB_CB(tskb)->end_seq++; 29611da177e4SLinus Torvalds tp->write_seq++; 2962845704a5SEric Dumazet if (!tcp_send_head(sk)) { 2963845704a5SEric Dumazet /* This means tskb was already sent. 2964845704a5SEric Dumazet * Pretend we included the FIN on previous transmit. 2965845704a5SEric Dumazet * We need to set tp->snd_nxt to the value it would have 2966845704a5SEric Dumazet * if FIN had been sent. This is because retransmit path 2967845704a5SEric Dumazet * does not change tp->snd_nxt. 2968845704a5SEric Dumazet */ 2969845704a5SEric Dumazet tp->snd_nxt++; 2970845704a5SEric Dumazet return; 2971845704a5SEric Dumazet } 29721da177e4SLinus Torvalds } else { 2973845704a5SEric Dumazet skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation); 2974845704a5SEric Dumazet if (unlikely(!skb)) { 2975845704a5SEric Dumazet if (tskb) 2976845704a5SEric Dumazet goto coalesce; 2977845704a5SEric Dumazet return; 29781da177e4SLinus Torvalds } 2979d83769a5SEric Dumazet skb_reserve(skb, MAX_TCP_HEADER); 2980a6c5ea4cSEric Dumazet sk_forced_mem_schedule(sk, skb->truesize); 29811da177e4SLinus Torvalds /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 2982e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tp->write_seq, 2983a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_FIN); 29841da177e4SLinus Torvalds tcp_queue_skb(sk, skb); 29851da177e4SLinus Torvalds } 2986845704a5SEric Dumazet __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF); 29871da177e4SLinus Torvalds } 29881da177e4SLinus Torvalds 29891da177e4SLinus Torvalds /* We get here when a process closes a file descriptor (either due to 29901da177e4SLinus Torvalds * an explicit close() or as a byproduct of exit()'ing) and there 29911da177e4SLinus Torvalds * was unread data in the receive queue. This behavior is recommended 299265bb723cSGerrit Renker * by RFC 2525, section 2.17. -DaveM 29931da177e4SLinus Torvalds */ 2994dd0fc66fSAl Viro void tcp_send_active_reset(struct sock *sk, gfp_t priority) 29951da177e4SLinus Torvalds { 29961da177e4SLinus Torvalds struct sk_buff *skb; 29971da177e4SLinus Torvalds 29981da177e4SLinus Torvalds /* NOTE: No TCP options attached and we never retransmit this. */ 29991da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, priority); 30001da177e4SLinus Torvalds if (!skb) { 30014e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 30021da177e4SLinus Torvalds return; 30031da177e4SLinus Torvalds } 30041da177e4SLinus Torvalds 30051da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 30061da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 3007e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), 3008a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_RST); 3009675ee231SEric Dumazet skb_mstamp_get(&skb->skb_mstamp); 30101da177e4SLinus Torvalds /* Send it off. */ 3011dfb4b9dcSDavid S. Miller if (tcp_transmit_skb(sk, skb, 0, priority)) 30124e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 301326af65cbSSridhar Samudrala 301481cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); 30151da177e4SLinus Torvalds } 30161da177e4SLinus Torvalds 301767edfef7SAndi Kleen /* Send a crossed SYN-ACK during socket establishment. 301867edfef7SAndi Kleen * WARNING: This routine must only be called when we have already sent 30191da177e4SLinus Torvalds * a SYN packet that crossed the incoming SYN that caused this routine 30201da177e4SLinus Torvalds * to get called. If this assumption fails then the initial rcv_wnd 30211da177e4SLinus Torvalds * and rcv_wscale values will not be correct. 30221da177e4SLinus Torvalds */ 30231da177e4SLinus Torvalds int tcp_send_synack(struct sock *sk) 30241da177e4SLinus Torvalds { 30251da177e4SLinus Torvalds struct sk_buff *skb; 30261da177e4SLinus Torvalds 3027fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 302851456b29SIan Morris if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 302991df42beSJoe Perches pr_debug("%s: wrong queue state\n", __func__); 30301da177e4SLinus Torvalds return -EFAULT; 30311da177e4SLinus Torvalds } 30324de075e0SEric Dumazet if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { 30331da177e4SLinus Torvalds if (skb_cloned(skb)) { 30341da177e4SLinus Torvalds struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC); 303551456b29SIan Morris if (!nskb) 30361da177e4SLinus Torvalds return -ENOMEM; 3037fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 3038f4a775d1SEric Dumazet __skb_header_release(nskb); 3039fe067e8aSDavid S. Miller __tcp_add_write_queue_head(sk, nskb); 30403ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 30413ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 30423ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 30431da177e4SLinus Torvalds skb = nskb; 30441da177e4SLinus Torvalds } 30451da177e4SLinus Torvalds 30464de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; 3047735d3831SFlorian Westphal tcp_ecn_send_synack(sk, skb); 30481da177e4SLinus Torvalds } 3049dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 30501da177e4SLinus Torvalds } 30511da177e4SLinus Torvalds 30524aea39c1SEric Dumazet /** 30534aea39c1SEric Dumazet * tcp_make_synack - Prepare a SYN-ACK. 30544aea39c1SEric Dumazet * sk: listener socket 30554aea39c1SEric Dumazet * dst: dst entry attached to the SYNACK 30564aea39c1SEric Dumazet * req: request_sock pointer 30574aea39c1SEric Dumazet * 30584aea39c1SEric Dumazet * Allocate one skb and build a SYNACK packet. 30594aea39c1SEric Dumazet * @dst is consumed : Caller should not use it again. 30604aea39c1SEric Dumazet */ 30615d062de7SEric Dumazet struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst, 3062e6b4d113SWilliam Allen Simpson struct request_sock *req, 3063ca6fb065SEric Dumazet struct tcp_fastopen_cookie *foc, 3064b3d05147SEric Dumazet enum tcp_synack_type synack_type) 30651da177e4SLinus Torvalds { 30662e6599cbSArnaldo Carvalho de Melo struct inet_request_sock *ireq = inet_rsk(req); 30675d062de7SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 306880f03e27SEric Dumazet struct tcp_md5sig_key *md5 = NULL; 30695d062de7SEric Dumazet struct tcp_out_options opts; 30705d062de7SEric Dumazet struct sk_buff *skb; 3071bd0388aeSWilliam Allen Simpson int tcp_header_size; 30725d062de7SEric Dumazet struct tcphdr *th; 3073f5fff5dcSTom Quetchenbach int mss; 30741da177e4SLinus Torvalds 3075ca6fb065SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 30764aea39c1SEric Dumazet if (unlikely(!skb)) { 30774aea39c1SEric Dumazet dst_release(dst); 30781da177e4SLinus Torvalds return NULL; 30794aea39c1SEric Dumazet } 30801da177e4SLinus Torvalds /* Reserve space for headers. */ 30811da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 30821da177e4SLinus Torvalds 3083b3d05147SEric Dumazet switch (synack_type) { 3084b3d05147SEric Dumazet case TCP_SYNACK_NORMAL: 30859e17f8a4SEric Dumazet skb_set_owner_w(skb, req_to_sk(req)); 3086b3d05147SEric Dumazet break; 3087b3d05147SEric Dumazet case TCP_SYNACK_COOKIE: 3088b3d05147SEric Dumazet /* Under synflood, we do not attach skb to a socket, 3089b3d05147SEric Dumazet * to avoid false sharing. 3090b3d05147SEric Dumazet */ 3091b3d05147SEric Dumazet break; 3092b3d05147SEric Dumazet case TCP_SYNACK_FASTOPEN: 3093ca6fb065SEric Dumazet /* sk is a const pointer, because we want to express multiple 3094ca6fb065SEric Dumazet * cpu might call us concurrently. 3095ca6fb065SEric Dumazet * sk->sk_wmem_alloc in an atomic, we can promote to rw. 3096ca6fb065SEric Dumazet */ 3097ca6fb065SEric Dumazet skb_set_owner_w(skb, (struct sock *)sk); 3098b3d05147SEric Dumazet break; 3099ca6fb065SEric Dumazet } 31004aea39c1SEric Dumazet skb_dst_set(skb, dst); 31011da177e4SLinus Torvalds 31023541f9e8SEric Dumazet mss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 3103f5fff5dcSTom Quetchenbach 310433ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 31058b5f12d0SFlorian Westphal #ifdef CONFIG_SYN_COOKIES 31068b5f12d0SFlorian Westphal if (unlikely(req->cookie_ts)) 31077faee5c0SEric Dumazet skb->skb_mstamp.stamp_jiffies = cookie_init_timestamp(req); 31088b5f12d0SFlorian Westphal else 31098b5f12d0SFlorian Westphal #endif 31107faee5c0SEric Dumazet skb_mstamp_get(&skb->skb_mstamp); 311180f03e27SEric Dumazet 311280f03e27SEric Dumazet #ifdef CONFIG_TCP_MD5SIG 311380f03e27SEric Dumazet rcu_read_lock(); 3114fd3a154aSEric Dumazet md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req)); 311580f03e27SEric Dumazet #endif 311658d607d3SEric Dumazet skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4); 311737bfbddaSEric Dumazet tcp_header_size = tcp_synack_options(req, mss, skb, &opts, md5, foc) + 311837bfbddaSEric Dumazet sizeof(*th); 311933ad798cSAdam Langley 3120aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 3121aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 31221da177e4SLinus Torvalds 3123ea1627c2SEric Dumazet th = (struct tcphdr *)skb->data; 31241da177e4SLinus Torvalds memset(th, 0, sizeof(struct tcphdr)); 31251da177e4SLinus Torvalds th->syn = 1; 31261da177e4SLinus Torvalds th->ack = 1; 31276ac705b1SEric Dumazet tcp_ecn_make_synack(req, th); 3128b44084c2SEric Dumazet th->source = htons(ireq->ir_num); 3129634fb979SEric Dumazet th->dest = ireq->ir_rmt_port; 3130e870a8efSIlpo Järvinen /* Setting of flags are superfluous here for callers (and ECE is 3131e870a8efSIlpo Järvinen * not even correctly set) 3132e870a8efSIlpo Järvinen */ 3133e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn, 3134a3433f35SChangli Gao TCPHDR_SYN | TCPHDR_ACK); 31354957faadSWilliam Allen Simpson 31361da177e4SLinus Torvalds th->seq = htonl(TCP_SKB_CB(skb)->seq); 31378336886fSJerry Chu /* XXX data is queued and acked as is. No buffer/window check */ 31388336886fSJerry Chu th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt); 31391da177e4SLinus Torvalds 31401da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 3141ed53d0abSEric Dumazet th->window = htons(min(req->rsk_rcv_wnd, 65535U)); 31425d062de7SEric Dumazet tcp_options_write((__be32 *)(th + 1), NULL, &opts); 31431da177e4SLinus Torvalds th->doff = (tcp_header_size >> 2); 314490bbcc60SEric Dumazet __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS); 3145cfb6eeb4SYOSHIFUJI Hideaki 3146cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 3147cfb6eeb4SYOSHIFUJI Hideaki /* Okay, we have all we need - do the md5 hash if needed */ 314880f03e27SEric Dumazet if (md5) 3149bd0388aeSWilliam Allen Simpson tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, 315039f8e58eSEric Dumazet md5, req_to_sk(req), skb); 315180f03e27SEric Dumazet rcu_read_unlock(); 3152cfb6eeb4SYOSHIFUJI Hideaki #endif 3153cfb6eeb4SYOSHIFUJI Hideaki 3154b50edd78SEric Dumazet /* Do not fool tcpdump (if any), clean our debris */ 31552456e855SThomas Gleixner skb->tstamp = 0; 31561da177e4SLinus Torvalds return skb; 31571da177e4SLinus Torvalds } 31584bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_make_synack); 31591da177e4SLinus Torvalds 316081164413SDaniel Borkmann static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst) 316181164413SDaniel Borkmann { 316281164413SDaniel Borkmann struct inet_connection_sock *icsk = inet_csk(sk); 316381164413SDaniel Borkmann const struct tcp_congestion_ops *ca; 316481164413SDaniel Borkmann u32 ca_key = dst_metric(dst, RTAX_CC_ALGO); 316581164413SDaniel Borkmann 316681164413SDaniel Borkmann if (ca_key == TCP_CA_UNSPEC) 316781164413SDaniel Borkmann return; 316881164413SDaniel Borkmann 316981164413SDaniel Borkmann rcu_read_lock(); 317081164413SDaniel Borkmann ca = tcp_ca_find_key(ca_key); 317181164413SDaniel Borkmann if (likely(ca && try_module_get(ca->owner))) { 317281164413SDaniel Borkmann module_put(icsk->icsk_ca_ops->owner); 317381164413SDaniel Borkmann icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst); 317481164413SDaniel Borkmann icsk->icsk_ca_ops = ca; 317581164413SDaniel Borkmann } 317681164413SDaniel Borkmann rcu_read_unlock(); 317781164413SDaniel Borkmann } 317881164413SDaniel Borkmann 317967edfef7SAndi Kleen /* Do all connect socket setups that can be done AF independent. */ 3180f7e56a76Sstephen hemminger static void tcp_connect_init(struct sock *sk) 31811da177e4SLinus Torvalds { 3182cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 31831da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 31841da177e4SLinus Torvalds __u8 rcv_wscale; 31851da177e4SLinus Torvalds 31861da177e4SLinus Torvalds /* We'll fix this up when we get a response from the other end. 31871da177e4SLinus Torvalds * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 31881da177e4SLinus Torvalds */ 31891da177e4SLinus Torvalds tp->tcp_header_len = sizeof(struct tcphdr) + 3190bb5b7c11SDavid S. Miller (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0); 31911da177e4SLinus Torvalds 3192cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 319300db4124SIan Morris if (tp->af_specific->md5_lookup(sk, sk)) 3194cfb6eeb4SYOSHIFUJI Hideaki tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 3195cfb6eeb4SYOSHIFUJI Hideaki #endif 3196cfb6eeb4SYOSHIFUJI Hideaki 31971da177e4SLinus Torvalds /* If user gave his TCP_MAXSEG, record it to clamp */ 31981da177e4SLinus Torvalds if (tp->rx_opt.user_mss) 31991da177e4SLinus Torvalds tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 32001da177e4SLinus Torvalds tp->max_window = 0; 32015d424d5aSJohn Heffner tcp_mtup_init(sk); 32021da177e4SLinus Torvalds tcp_sync_mss(sk, dst_mtu(dst)); 32031da177e4SLinus Torvalds 320481164413SDaniel Borkmann tcp_ca_dst_init(sk, dst); 320581164413SDaniel Borkmann 32061da177e4SLinus Torvalds if (!tp->window_clamp) 32071da177e4SLinus Torvalds tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 32083541f9e8SEric Dumazet tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 3209f5fff5dcSTom Quetchenbach 32101da177e4SLinus Torvalds tcp_initialize_rcv_mss(sk); 32111da177e4SLinus Torvalds 3212e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 3213e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 3214e88c64f0SHagen Paul Pfeifer (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) 3215e88c64f0SHagen Paul Pfeifer tp->window_clamp = tcp_full_space(sk); 3216e88c64f0SHagen Paul Pfeifer 32171da177e4SLinus Torvalds tcp_select_initial_window(tcp_full_space(sk), 32181da177e4SLinus Torvalds tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 32191da177e4SLinus Torvalds &tp->rcv_wnd, 32201da177e4SLinus Torvalds &tp->window_clamp, 3221bb5b7c11SDavid S. Miller sysctl_tcp_window_scaling, 322231d12926Slaurent chavey &rcv_wscale, 322331d12926Slaurent chavey dst_metric(dst, RTAX_INITRWND)); 32241da177e4SLinus Torvalds 32251da177e4SLinus Torvalds tp->rx_opt.rcv_wscale = rcv_wscale; 32261da177e4SLinus Torvalds tp->rcv_ssthresh = tp->rcv_wnd; 32271da177e4SLinus Torvalds 32281da177e4SLinus Torvalds sk->sk_err = 0; 32291da177e4SLinus Torvalds sock_reset_flag(sk, SOCK_DONE); 32301da177e4SLinus Torvalds tp->snd_wnd = 0; 3231ee7537b6SHantzis Fotis tcp_init_wl(tp, 0); 32321da177e4SLinus Torvalds tp->snd_una = tp->write_seq; 32331da177e4SLinus Torvalds tp->snd_sml = tp->write_seq; 323433f5f57eSIlpo Järvinen tp->snd_up = tp->write_seq; 3235370816aeSPavel Emelyanov tp->snd_nxt = tp->write_seq; 3236ee995283SPavel Emelyanov 3237ee995283SPavel Emelyanov if (likely(!tp->repair)) 32381da177e4SLinus Torvalds tp->rcv_nxt = 0; 3239c7781a6eSAndrew Vagin else 3240c7781a6eSAndrew Vagin tp->rcv_tstamp = tcp_time_stamp; 3241ee995283SPavel Emelyanov tp->rcv_wup = tp->rcv_nxt; 3242ee995283SPavel Emelyanov tp->copied_seq = tp->rcv_nxt; 32431da177e4SLinus Torvalds 3244463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; 3245463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_retransmits = 0; 32461da177e4SLinus Torvalds tcp_clear_retrans(tp); 32471da177e4SLinus Torvalds } 32481da177e4SLinus Torvalds 3249783237e8SYuchung Cheng static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) 3250783237e8SYuchung Cheng { 3251783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3252783237e8SYuchung Cheng struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 3253783237e8SYuchung Cheng 3254783237e8SYuchung Cheng tcb->end_seq += skb->len; 3255f4a775d1SEric Dumazet __skb_header_release(skb); 3256783237e8SYuchung Cheng __tcp_add_write_queue_tail(sk, skb); 3257783237e8SYuchung Cheng sk->sk_wmem_queued += skb->truesize; 3258783237e8SYuchung Cheng sk_mem_charge(sk, skb->truesize); 3259783237e8SYuchung Cheng tp->write_seq = tcb->end_seq; 3260783237e8SYuchung Cheng tp->packets_out += tcp_skb_pcount(skb); 3261783237e8SYuchung Cheng } 3262783237e8SYuchung Cheng 3263783237e8SYuchung Cheng /* Build and send a SYN with data and (cached) Fast Open cookie. However, 3264783237e8SYuchung Cheng * queue a data-only packet after the regular SYN, such that regular SYNs 3265783237e8SYuchung Cheng * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges 3266783237e8SYuchung Cheng * only the SYN sequence, the data are retransmitted in the first ACK. 3267783237e8SYuchung Cheng * If cookie is not cached or other error occurs, falls back to send a 3268783237e8SYuchung Cheng * regular SYN with Fast Open cookie request option. 3269783237e8SYuchung Cheng */ 3270783237e8SYuchung Cheng static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) 3271783237e8SYuchung Cheng { 3272783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3273783237e8SYuchung Cheng struct tcp_fastopen_request *fo = tp->fastopen_req; 3274065263f4SWei Wang int space, err = 0; 3275355a901eSEric Dumazet struct sk_buff *syn_data; 3276783237e8SYuchung Cheng 327767da22d2SYuchung Cheng tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ 3278065263f4SWei Wang if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie)) 3279783237e8SYuchung Cheng goto fallback; 3280783237e8SYuchung Cheng 3281783237e8SYuchung Cheng /* MSS for SYN-data is based on cached MSS and bounded by PMTU and 3282783237e8SYuchung Cheng * user-MSS. Reserve maximum option space for middleboxes that add 3283783237e8SYuchung Cheng * private TCP options. The cost is reduced data space in SYN :( 3284783237e8SYuchung Cheng */ 32853541f9e8SEric Dumazet tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp); 32863541f9e8SEric Dumazet 32871b63edd6SYuchung Cheng space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) - 3288783237e8SYuchung Cheng MAX_TCP_OPTION_SPACE; 3289783237e8SYuchung Cheng 3290f5ddcbbbSEric Dumazet space = min_t(size_t, space, fo->size); 3291f5ddcbbbSEric Dumazet 3292f5ddcbbbSEric Dumazet /* limit to order-0 allocations */ 3293f5ddcbbbSEric Dumazet space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER)); 3294f5ddcbbbSEric Dumazet 3295eb934478SEric Dumazet syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false); 3296355a901eSEric Dumazet if (!syn_data) 3297783237e8SYuchung Cheng goto fallback; 3298355a901eSEric Dumazet syn_data->ip_summed = CHECKSUM_PARTIAL; 3299355a901eSEric Dumazet memcpy(syn_data->cb, syn->cb, sizeof(syn->cb)); 330007e100f9SEric Dumazet if (space) { 330107e100f9SEric Dumazet int copied = copy_from_iter(skb_put(syn_data, space), space, 330257be5bdaSAl Viro &fo->data->msg_iter); 330357be5bdaSAl Viro if (unlikely(!copied)) { 3304355a901eSEric Dumazet kfree_skb(syn_data); 3305783237e8SYuchung Cheng goto fallback; 3306783237e8SYuchung Cheng } 330757be5bdaSAl Viro if (copied != space) { 330857be5bdaSAl Viro skb_trim(syn_data, copied); 330957be5bdaSAl Viro space = copied; 331057be5bdaSAl Viro } 331107e100f9SEric Dumazet } 3312355a901eSEric Dumazet /* No more data pending in inet_wait_for_connect() */ 3313355a901eSEric Dumazet if (space == fo->size) 3314355a901eSEric Dumazet fo->data = NULL; 3315355a901eSEric Dumazet fo->copied = space; 3316783237e8SYuchung Cheng 3317355a901eSEric Dumazet tcp_connect_queue_skb(sk, syn_data); 33180f87230dSFrancis Yan if (syn_data->len) 33190f87230dSFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_BUSY); 3320355a901eSEric Dumazet 3321355a901eSEric Dumazet err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation); 3322355a901eSEric Dumazet 3323355a901eSEric Dumazet syn->skb_mstamp = syn_data->skb_mstamp; 3324355a901eSEric Dumazet 3325355a901eSEric Dumazet /* Now full SYN+DATA was cloned and sent (or not), 3326355a901eSEric Dumazet * remove the SYN from the original skb (syn_data) 3327355a901eSEric Dumazet * we keep in write queue in case of a retransmit, as we 3328355a901eSEric Dumazet * also have the SYN packet (with no data) in the same queue. 3329431a9124SEric Dumazet */ 3330355a901eSEric Dumazet TCP_SKB_CB(syn_data)->seq++; 3331355a901eSEric Dumazet TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH; 3332355a901eSEric Dumazet if (!err) { 333367da22d2SYuchung Cheng tp->syn_data = (fo->copied > 0); 3334f19c29e3SYuchung Cheng NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT); 3335783237e8SYuchung Cheng goto done; 3336783237e8SYuchung Cheng } 3337783237e8SYuchung Cheng 3338783237e8SYuchung Cheng fallback: 3339783237e8SYuchung Cheng /* Send a regular SYN with Fast Open cookie request option */ 3340783237e8SYuchung Cheng if (fo->cookie.len > 0) 3341783237e8SYuchung Cheng fo->cookie.len = 0; 3342783237e8SYuchung Cheng err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); 3343783237e8SYuchung Cheng if (err) 3344783237e8SYuchung Cheng tp->syn_fastopen = 0; 3345783237e8SYuchung Cheng done: 3346783237e8SYuchung Cheng fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ 3347783237e8SYuchung Cheng return err; 3348783237e8SYuchung Cheng } 3349783237e8SYuchung Cheng 335067edfef7SAndi Kleen /* Build a SYN and send it off. */ 33511da177e4SLinus Torvalds int tcp_connect(struct sock *sk) 33521da177e4SLinus Torvalds { 33531da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 33541da177e4SLinus Torvalds struct sk_buff *buff; 3355ee586811SEric Paris int err; 33561da177e4SLinus Torvalds 33571da177e4SLinus Torvalds tcp_connect_init(sk); 33581da177e4SLinus Torvalds 33592b916477SAndrey Vagin if (unlikely(tp->repair)) { 33602b916477SAndrey Vagin tcp_finish_connect(sk, NULL); 33612b916477SAndrey Vagin return 0; 33622b916477SAndrey Vagin } 33632b916477SAndrey Vagin 3364eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true); 3365355a901eSEric Dumazet if (unlikely(!buff)) 33661da177e4SLinus Torvalds return -ENOBUFS; 33671da177e4SLinus Torvalds 3368a3433f35SChangli Gao tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); 33697faee5c0SEric Dumazet tp->retrans_stamp = tcp_time_stamp; 3370783237e8SYuchung Cheng tcp_connect_queue_skb(sk, buff); 3371735d3831SFlorian Westphal tcp_ecn_send_syn(sk, buff); 33721da177e4SLinus Torvalds 3373783237e8SYuchung Cheng /* Send off SYN; include data in Fast Open. */ 3374783237e8SYuchung Cheng err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : 3375783237e8SYuchung Cheng tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); 3376ee586811SEric Paris if (err == -ECONNREFUSED) 3377ee586811SEric Paris return err; 3378bd37a088SWei Yongjun 3379bd37a088SWei Yongjun /* We change tp->snd_nxt after the tcp_transmit_skb() call 3380bd37a088SWei Yongjun * in order to make this packet get counted in tcpOutSegs. 3381bd37a088SWei Yongjun */ 3382bd37a088SWei Yongjun tp->snd_nxt = tp->write_seq; 3383bd37a088SWei Yongjun tp->pushed_seq = tp->write_seq; 338481cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); 33851da177e4SLinus Torvalds 33861da177e4SLinus Torvalds /* Timer for repeating the SYN until an answer. */ 33873f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 33883f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 33891da177e4SLinus Torvalds return 0; 33901da177e4SLinus Torvalds } 33914bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_connect); 33921da177e4SLinus Torvalds 33931da177e4SLinus Torvalds /* Send out a delayed ack, the caller does the policy checking 33941da177e4SLinus Torvalds * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 33951da177e4SLinus Torvalds * for details. 33961da177e4SLinus Torvalds */ 33971da177e4SLinus Torvalds void tcp_send_delayed_ack(struct sock *sk) 33981da177e4SLinus Torvalds { 3399463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 3400463c84b9SArnaldo Carvalho de Melo int ato = icsk->icsk_ack.ato; 34011da177e4SLinus Torvalds unsigned long timeout; 34021da177e4SLinus Torvalds 34039890092eSFlorian Westphal tcp_ca_event(sk, CA_EVENT_DELAYED_ACK); 34049890092eSFlorian Westphal 34051da177e4SLinus Torvalds if (ato > TCP_DELACK_MIN) { 3406463c84b9SArnaldo Carvalho de Melo const struct tcp_sock *tp = tcp_sk(sk); 34071da177e4SLinus Torvalds int max_ato = HZ / 2; 34081da177e4SLinus Torvalds 3409056834d9SIlpo Järvinen if (icsk->icsk_ack.pingpong || 3410056834d9SIlpo Järvinen (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 34111da177e4SLinus Torvalds max_ato = TCP_DELACK_MAX; 34121da177e4SLinus Torvalds 34131da177e4SLinus Torvalds /* Slow path, intersegment interval is "high". */ 34141da177e4SLinus Torvalds 34151da177e4SLinus Torvalds /* If some rtt estimate is known, use it to bound delayed ack. 3416463c84b9SArnaldo Carvalho de Melo * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 34171da177e4SLinus Torvalds * directly. 34181da177e4SLinus Torvalds */ 3419740b0f18SEric Dumazet if (tp->srtt_us) { 3420740b0f18SEric Dumazet int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3), 3421740b0f18SEric Dumazet TCP_DELACK_MIN); 34221da177e4SLinus Torvalds 34231da177e4SLinus Torvalds if (rtt < max_ato) 34241da177e4SLinus Torvalds max_ato = rtt; 34251da177e4SLinus Torvalds } 34261da177e4SLinus Torvalds 34271da177e4SLinus Torvalds ato = min(ato, max_ato); 34281da177e4SLinus Torvalds } 34291da177e4SLinus Torvalds 34301da177e4SLinus Torvalds /* Stay within the limit we were given */ 34311da177e4SLinus Torvalds timeout = jiffies + ato; 34321da177e4SLinus Torvalds 34331da177e4SLinus Torvalds /* Use new timeout only if there wasn't a older one earlier. */ 3434463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 34351da177e4SLinus Torvalds /* If delack timer was blocked or is about to expire, 34361da177e4SLinus Torvalds * send ACK now. 34371da177e4SLinus Torvalds */ 3438463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.blocked || 3439463c84b9SArnaldo Carvalho de Melo time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 34401da177e4SLinus Torvalds tcp_send_ack(sk); 34411da177e4SLinus Torvalds return; 34421da177e4SLinus Torvalds } 34431da177e4SLinus Torvalds 3444463c84b9SArnaldo Carvalho de Melo if (!time_before(timeout, icsk->icsk_ack.timeout)) 3445463c84b9SArnaldo Carvalho de Melo timeout = icsk->icsk_ack.timeout; 34461da177e4SLinus Torvalds } 3447463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 3448463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.timeout = timeout; 3449463c84b9SArnaldo Carvalho de Melo sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 34501da177e4SLinus Torvalds } 34511da177e4SLinus Torvalds 34521da177e4SLinus Torvalds /* This routine sends an ack and also updates the window. */ 34531da177e4SLinus Torvalds void tcp_send_ack(struct sock *sk) 34541da177e4SLinus Torvalds { 34551da177e4SLinus Torvalds struct sk_buff *buff; 34561da177e4SLinus Torvalds 3457058dc334SIlpo Järvinen /* If we have been reset, we may not send again. */ 3458058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3459058dc334SIlpo Järvinen return; 3460058dc334SIlpo Järvinen 34619890092eSFlorian Westphal tcp_ca_event(sk, CA_EVENT_NON_DELAYED_ACK); 34629890092eSFlorian Westphal 34631da177e4SLinus Torvalds /* We are not putting this on the write queue, so 34641da177e4SLinus Torvalds * tcp_transmit_skb() will set the ownership to this 34651da177e4SLinus Torvalds * sock. 34661da177e4SLinus Torvalds */ 34677450aaf6SEric Dumazet buff = alloc_skb(MAX_TCP_HEADER, 34687450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 34697450aaf6SEric Dumazet if (unlikely(!buff)) { 3470463c84b9SArnaldo Carvalho de Melo inet_csk_schedule_ack(sk); 3471463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; 34723f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 34733f421baaSArnaldo Carvalho de Melo TCP_DELACK_MAX, TCP_RTO_MAX); 34741da177e4SLinus Torvalds return; 34751da177e4SLinus Torvalds } 34761da177e4SLinus Torvalds 34771da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 34781da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 3479a3433f35SChangli Gao tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); 34801da177e4SLinus Torvalds 348198781965SEric Dumazet /* We do not want pure acks influencing TCP Small Queues or fq/pacing 348298781965SEric Dumazet * too much. 348398781965SEric Dumazet * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784 348498781965SEric Dumazet */ 348598781965SEric Dumazet skb_set_tcp_pure_ack(buff); 348698781965SEric Dumazet 34871da177e4SLinus Torvalds /* Send it off, this clears delayed acks for us. */ 34887faee5c0SEric Dumazet skb_mstamp_get(&buff->skb_mstamp); 34897450aaf6SEric Dumazet tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0); 34901da177e4SLinus Torvalds } 3491e3118e83SDaniel Borkmann EXPORT_SYMBOL_GPL(tcp_send_ack); 34921da177e4SLinus Torvalds 34931da177e4SLinus Torvalds /* This routine sends a packet with an out of date sequence 34941da177e4SLinus Torvalds * number. It assumes the other end will try to ack it. 34951da177e4SLinus Torvalds * 34961da177e4SLinus Torvalds * Question: what should we make while urgent mode? 34971da177e4SLinus Torvalds * 4.4BSD forces sending single byte of data. We cannot send 34981da177e4SLinus Torvalds * out of window data, because we have SND.NXT==SND.MAX... 34991da177e4SLinus Torvalds * 35001da177e4SLinus Torvalds * Current solution: to send TWO zero-length segments in urgent mode: 35011da177e4SLinus Torvalds * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 35021da177e4SLinus Torvalds * out-of-date with SND.UNA-1 to probe window. 35031da177e4SLinus Torvalds */ 3504e520af48SEric Dumazet static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib) 35051da177e4SLinus Torvalds { 35061da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 35071da177e4SLinus Torvalds struct sk_buff *skb; 35081da177e4SLinus Torvalds 35091da177e4SLinus Torvalds /* We don't queue it, tcp_transmit_skb() sets ownership. */ 35107450aaf6SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, 35117450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 351251456b29SIan Morris if (!skb) 35131da177e4SLinus Torvalds return -1; 35141da177e4SLinus Torvalds 35151da177e4SLinus Torvalds /* Reserve space for headers and set control bits. */ 35161da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 35171da177e4SLinus Torvalds /* Use a previous sequence. This should cause the other 35181da177e4SLinus Torvalds * end to send an ack. Don't queue or clone SKB, just 35191da177e4SLinus Torvalds * send it. 35201da177e4SLinus Torvalds */ 3521a3433f35SChangli Gao tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); 35227faee5c0SEric Dumazet skb_mstamp_get(&skb->skb_mstamp); 3523e2e8009fSRenato Westphal NET_INC_STATS(sock_net(sk), mib); 35247450aaf6SEric Dumazet return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0); 35251da177e4SLinus Torvalds } 35261da177e4SLinus Torvalds 3527ee995283SPavel Emelyanov void tcp_send_window_probe(struct sock *sk) 3528ee995283SPavel Emelyanov { 3529ee995283SPavel Emelyanov if (sk->sk_state == TCP_ESTABLISHED) { 3530ee995283SPavel Emelyanov tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; 3531e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE); 3532ee995283SPavel Emelyanov } 3533ee995283SPavel Emelyanov } 3534ee995283SPavel Emelyanov 353567edfef7SAndi Kleen /* Initiate keepalive or window probe from timer. */ 3536e520af48SEric Dumazet int tcp_write_wakeup(struct sock *sk, int mib) 35371da177e4SLinus Torvalds { 35381da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 35391da177e4SLinus Torvalds struct sk_buff *skb; 35401da177e4SLinus Torvalds 3541058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3542058dc334SIlpo Järvinen return -1; 3543058dc334SIlpo Järvinen 354400db4124SIan Morris skb = tcp_send_head(sk); 354500db4124SIan Morris if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { 35461da177e4SLinus Torvalds int err; 35470c54b85fSIlpo Järvinen unsigned int mss = tcp_current_mss(sk); 354890840defSIlpo Järvinen unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 35491da177e4SLinus Torvalds 35501da177e4SLinus Torvalds if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 35511da177e4SLinus Torvalds tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 35521da177e4SLinus Torvalds 35531da177e4SLinus Torvalds /* We are probing the opening of a window 35541da177e4SLinus Torvalds * but the window size is != 0 35551da177e4SLinus Torvalds * must have been a result SWS avoidance ( sender ) 35561da177e4SLinus Torvalds */ 35571da177e4SLinus Torvalds if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 35581da177e4SLinus Torvalds skb->len > mss) { 35591da177e4SLinus Torvalds seg_size = min(seg_size, mss); 35604de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 35616cc55e09SOctavian Purdila if (tcp_fragment(sk, skb, seg_size, mss, GFP_ATOMIC)) 35621da177e4SLinus Torvalds return -1; 35631da177e4SLinus Torvalds } else if (!tcp_skb_pcount(skb)) 35645bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss); 35651da177e4SLinus Torvalds 35664de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 3567dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 356866f5fe62SIlpo Järvinen if (!err) 356966f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 35701da177e4SLinus Torvalds return err; 35711da177e4SLinus Torvalds } else { 357233f5f57eSIlpo Järvinen if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) 3573e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 1, mib); 3574e520af48SEric Dumazet return tcp_xmit_probe_skb(sk, 0, mib); 35751da177e4SLinus Torvalds } 35761da177e4SLinus Torvalds } 35771da177e4SLinus Torvalds 35781da177e4SLinus Torvalds /* A window probe timeout has occurred. If window is not closed send 35791da177e4SLinus Torvalds * a partial packet else a zero probe. 35801da177e4SLinus Torvalds */ 35811da177e4SLinus Torvalds void tcp_send_probe0(struct sock *sk) 35821da177e4SLinus Torvalds { 3583463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 35841da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 3585c6214a97SNikolay Borisov struct net *net = sock_net(sk); 3586fcdd1cf4SEric Dumazet unsigned long probe_max; 35871da177e4SLinus Torvalds int err; 35881da177e4SLinus Torvalds 3589e520af48SEric Dumazet err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE); 35901da177e4SLinus Torvalds 3591fe067e8aSDavid S. Miller if (tp->packets_out || !tcp_send_head(sk)) { 35921da177e4SLinus Torvalds /* Cancel probe timer, if it is not required. */ 35936687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 0; 3594463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff = 0; 35951da177e4SLinus Torvalds return; 35961da177e4SLinus Torvalds } 35971da177e4SLinus Torvalds 35981da177e4SLinus Torvalds if (err <= 0) { 3599c6214a97SNikolay Borisov if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2) 3600463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff++; 36016687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out++; 3602fcdd1cf4SEric Dumazet probe_max = TCP_RTO_MAX; 36031da177e4SLinus Torvalds } else { 36041da177e4SLinus Torvalds /* If packet was not sent due to local congestion, 36056687e988SArnaldo Carvalho de Melo * do not backoff and do not remember icsk_probes_out. 36061da177e4SLinus Torvalds * Let local senders to fight for local resources. 36071da177e4SLinus Torvalds * 36081da177e4SLinus Torvalds * Use accumulated backoff yet. 36091da177e4SLinus Torvalds */ 36106687e988SArnaldo Carvalho de Melo if (!icsk->icsk_probes_out) 36116687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 1; 3612fcdd1cf4SEric Dumazet probe_max = TCP_RESOURCE_PROBE_INTERVAL; 36131da177e4SLinus Torvalds } 3614fcdd1cf4SEric Dumazet inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 361521c8fe99SEric Dumazet tcp_probe0_when(sk, probe_max), 3616fcdd1cf4SEric Dumazet TCP_RTO_MAX); 36171da177e4SLinus Torvalds } 36185db92c99SOctavian Purdila 3619ea3bea3aSEric Dumazet int tcp_rtx_synack(const struct sock *sk, struct request_sock *req) 36205db92c99SOctavian Purdila { 36215db92c99SOctavian Purdila const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific; 36225db92c99SOctavian Purdila struct flowi fl; 36235db92c99SOctavian Purdila int res; 36245db92c99SOctavian Purdila 362558d607d3SEric Dumazet tcp_rsk(req)->txhash = net_tx_rndhash(); 3626b3d05147SEric Dumazet res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL); 36275db92c99SOctavian Purdila if (!res) { 362890bbcc60SEric Dumazet __TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); 362902a1d6e7SEric Dumazet __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 36307e32b443SYuchung Cheng if (unlikely(tcp_passive_fastopen(sk))) 36317e32b443SYuchung Cheng tcp_sk(sk)->total_retrans++; 36325db92c99SOctavian Purdila } 36335db92c99SOctavian Purdila return res; 36345db92c99SOctavian Purdila } 36355db92c99SOctavian Purdila EXPORT_SYMBOL(tcp_rtx_synack); 3636