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); 154c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 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); 163d635fbe2SEric Dumazet const u32 now = tcp_jiffies32; 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) 215589c49cbSGao 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) 220589c49cbSGao 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) { 237589c49cbSGao 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); 241589c49cbSGao 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 */ 254589c49cbSGao 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 29831770e34SFlorian Westphal /* If we advertise zero window, disable fast path. */ 2998e165e20SFlorian Westphal if (new_win == 0) { 30031770e34SFlorian Westphal 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; 31991b5b21cSLawrence Brakmo else if (tcp_ca_needs_ecn(sk) || 32091b5b21cSLawrence Brakmo tcp_bpf_ca_needs_ecn(sk)) 32130e502a3SDaniel Borkmann INET_ECN_xmit(sk); 322bdf1ee5dSIlpo Järvinen } 323bdf1ee5dSIlpo Järvinen 32467edfef7SAndi Kleen /* Packet ECN state for a SYN. */ 325735d3831SFlorian Westphal static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb) 326bdf1ee5dSIlpo Järvinen { 327bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 32891b5b21cSLawrence Brakmo bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk); 329f7b3bec6SFlorian Westphal bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 || 33091b5b21cSLawrence Brakmo tcp_ca_needs_ecn(sk) || bpf_needs_ecn; 331f7b3bec6SFlorian Westphal 332f7b3bec6SFlorian Westphal if (!use_ecn) { 333f7b3bec6SFlorian Westphal const struct dst_entry *dst = __sk_dst_get(sk); 334f7b3bec6SFlorian Westphal 335f7b3bec6SFlorian Westphal if (dst && dst_feature(dst, RTAX_FEATURE_ECN)) 336f7b3bec6SFlorian Westphal use_ecn = true; 337f7b3bec6SFlorian Westphal } 338bdf1ee5dSIlpo Järvinen 339bdf1ee5dSIlpo Järvinen tp->ecn_flags = 0; 340f7b3bec6SFlorian Westphal 341f7b3bec6SFlorian Westphal if (use_ecn) { 3424de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR; 343bdf1ee5dSIlpo Järvinen tp->ecn_flags = TCP_ECN_OK; 34491b5b21cSLawrence Brakmo if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn) 34530e502a3SDaniel Borkmann INET_ECN_xmit(sk); 346bdf1ee5dSIlpo Järvinen } 347bdf1ee5dSIlpo Järvinen } 348bdf1ee5dSIlpo Järvinen 34949213555SDaniel Borkmann static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb) 35049213555SDaniel Borkmann { 35149213555SDaniel Borkmann if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback) 35249213555SDaniel Borkmann /* tp->ecn_flags are cleared at a later point in time when 35349213555SDaniel Borkmann * SYN ACK is ultimatively being received. 35449213555SDaniel Borkmann */ 35549213555SDaniel Borkmann TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR); 35649213555SDaniel Borkmann } 35749213555SDaniel Borkmann 358735d3831SFlorian Westphal static void 3596ac705b1SEric Dumazet tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th) 360bdf1ee5dSIlpo Järvinen { 3616ac705b1SEric Dumazet if (inet_rsk(req)->ecn_ok) 362bdf1ee5dSIlpo Järvinen th->ece = 1; 363bdf1ee5dSIlpo Järvinen } 364bdf1ee5dSIlpo Järvinen 36567edfef7SAndi Kleen /* Set up ECN state for a packet on a ESTABLISHED socket that is about to 36667edfef7SAndi Kleen * be sent. 36767edfef7SAndi Kleen */ 368735d3831SFlorian Westphal static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb, 369ea1627c2SEric Dumazet struct tcphdr *th, int tcp_header_len) 370bdf1ee5dSIlpo Järvinen { 371bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 372bdf1ee5dSIlpo Järvinen 373bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_OK) { 374bdf1ee5dSIlpo Järvinen /* Not-retransmitted data segment: set ECT and inject CWR. */ 375bdf1ee5dSIlpo Järvinen if (skb->len != tcp_header_len && 376bdf1ee5dSIlpo Järvinen !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { 377bdf1ee5dSIlpo Järvinen INET_ECN_xmit(sk); 378bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) { 379bdf1ee5dSIlpo Järvinen tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; 380ea1627c2SEric Dumazet th->cwr = 1; 381bdf1ee5dSIlpo Järvinen skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; 382bdf1ee5dSIlpo Järvinen } 38330e502a3SDaniel Borkmann } else if (!tcp_ca_needs_ecn(sk)) { 384bdf1ee5dSIlpo Järvinen /* ACK or retransmitted segment: clear ECT|CE */ 385bdf1ee5dSIlpo Järvinen INET_ECN_dontxmit(sk); 386bdf1ee5dSIlpo Järvinen } 387bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) 388ea1627c2SEric Dumazet th->ece = 1; 389bdf1ee5dSIlpo Järvinen } 390bdf1ee5dSIlpo Järvinen } 391bdf1ee5dSIlpo Järvinen 392e870a8efSIlpo Järvinen /* Constructs common control bits of non-data skb. If SYN/FIN is present, 393e870a8efSIlpo Järvinen * auto increment end seqno. 394e870a8efSIlpo Järvinen */ 395e870a8efSIlpo Järvinen static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags) 396e870a8efSIlpo Järvinen { 3972e8e18efSDavid S. Miller skb->ip_summed = CHECKSUM_PARTIAL; 398e870a8efSIlpo Järvinen skb->csum = 0; 399e870a8efSIlpo Järvinen 4004de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags; 401e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->sacked = 0; 402e870a8efSIlpo Järvinen 403cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, 1); 404e870a8efSIlpo Järvinen 405e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->seq = seq; 406a3433f35SChangli Gao if (flags & (TCPHDR_SYN | TCPHDR_FIN)) 407e870a8efSIlpo Järvinen seq++; 408e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->end_seq = seq; 409e870a8efSIlpo Järvinen } 410e870a8efSIlpo Järvinen 411a2a385d6SEric Dumazet static inline bool tcp_urg_mode(const struct tcp_sock *tp) 41233f5f57eSIlpo Järvinen { 41333f5f57eSIlpo Järvinen return tp->snd_una != tp->snd_up; 41433f5f57eSIlpo Järvinen } 41533f5f57eSIlpo Järvinen 41633ad798cSAdam Langley #define OPTION_SACK_ADVERTISE (1 << 0) 41733ad798cSAdam Langley #define OPTION_TS (1 << 1) 41833ad798cSAdam Langley #define OPTION_MD5 (1 << 2) 41989e95a61SOri Finkelman #define OPTION_WSCALE (1 << 3) 4202100c8d2SYuchung Cheng #define OPTION_FAST_OPEN_COOKIE (1 << 8) 42133ad798cSAdam Langley 42233ad798cSAdam Langley struct tcp_out_options { 4232100c8d2SYuchung Cheng u16 options; /* bit field of OPTION_* */ 4242100c8d2SYuchung Cheng u16 mss; /* 0 to disable */ 42533ad798cSAdam Langley u8 ws; /* window scale, 0 to disable */ 42633ad798cSAdam Langley u8 num_sack_blocks; /* number of SACK blocks to include */ 427bd0388aeSWilliam Allen Simpson u8 hash_size; /* bytes in hash_location */ 428bd0388aeSWilliam Allen Simpson __u8 *hash_location; /* temporary pointer, overloaded */ 4292100c8d2SYuchung Cheng __u32 tsval, tsecr; /* need to include OPTION_TS */ 4302100c8d2SYuchung Cheng struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */ 43133ad798cSAdam Langley }; 43233ad798cSAdam Langley 43367edfef7SAndi Kleen /* Write previously computed TCP options to the packet. 43467edfef7SAndi Kleen * 43567edfef7SAndi Kleen * Beware: Something in the Internet is very sensitive to the ordering of 436fd6149d3SIlpo Järvinen * TCP options, we learned this through the hard way, so be careful here. 437fd6149d3SIlpo Järvinen * Luckily we can at least blame others for their non-compliance but from 4388e3bff96Sstephen hemminger * inter-operability perspective it seems that we're somewhat stuck with 439fd6149d3SIlpo Järvinen * the ordering which we have been using if we want to keep working with 440fd6149d3SIlpo Järvinen * those broken things (not that it currently hurts anybody as there isn't 441fd6149d3SIlpo Järvinen * particular reason why the ordering would need to be changed). 442fd6149d3SIlpo Järvinen * 443fd6149d3SIlpo Järvinen * At least SACK_PERM as the first option is known to lead to a disaster 444fd6149d3SIlpo Järvinen * (but it may well be that other scenarios fail similarly). 445fd6149d3SIlpo Järvinen */ 44633ad798cSAdam Langley static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp, 447bd0388aeSWilliam Allen Simpson struct tcp_out_options *opts) 448bd0388aeSWilliam Allen Simpson { 4492100c8d2SYuchung Cheng u16 options = opts->options; /* mungable copy */ 450bd0388aeSWilliam Allen Simpson 451bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_MD5 & options)) { 4521a2c6181SChristoph Paasch *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | 4531a2c6181SChristoph Paasch (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG); 454bd0388aeSWilliam Allen Simpson /* overload cookie hash location */ 455bd0388aeSWilliam Allen Simpson opts->hash_location = (__u8 *)ptr; 45633ad798cSAdam Langley ptr += 4; 45733ad798cSAdam Langley } 45833ad798cSAdam Langley 459fd6149d3SIlpo Järvinen if (unlikely(opts->mss)) { 460fd6149d3SIlpo Järvinen *ptr++ = htonl((TCPOPT_MSS << 24) | 461fd6149d3SIlpo Järvinen (TCPOLEN_MSS << 16) | 462fd6149d3SIlpo Järvinen opts->mss); 463fd6149d3SIlpo Järvinen } 464fd6149d3SIlpo Järvinen 465bd0388aeSWilliam Allen Simpson if (likely(OPTION_TS & options)) { 466bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 46733ad798cSAdam Langley *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | 46833ad798cSAdam Langley (TCPOLEN_SACK_PERM << 16) | 46933ad798cSAdam Langley (TCPOPT_TIMESTAMP << 8) | 47033ad798cSAdam Langley TCPOLEN_TIMESTAMP); 471bd0388aeSWilliam Allen Simpson options &= ~OPTION_SACK_ADVERTISE; 47233ad798cSAdam Langley } else { 473496c98dfSYOSHIFUJI Hideaki *ptr++ = htonl((TCPOPT_NOP << 24) | 47440efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 47540efc6faSStephen Hemminger (TCPOPT_TIMESTAMP << 8) | 47640efc6faSStephen Hemminger TCPOLEN_TIMESTAMP); 47740efc6faSStephen Hemminger } 47833ad798cSAdam Langley *ptr++ = htonl(opts->tsval); 47933ad798cSAdam Langley *ptr++ = htonl(opts->tsecr); 48033ad798cSAdam Langley } 48133ad798cSAdam Langley 482bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 48333ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 48433ad798cSAdam Langley (TCPOPT_NOP << 16) | 48533ad798cSAdam Langley (TCPOPT_SACK_PERM << 8) | 48633ad798cSAdam Langley TCPOLEN_SACK_PERM); 48733ad798cSAdam Langley } 48833ad798cSAdam Langley 489bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_WSCALE & options)) { 49033ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 49133ad798cSAdam Langley (TCPOPT_WINDOW << 16) | 49233ad798cSAdam Langley (TCPOLEN_WINDOW << 8) | 49333ad798cSAdam Langley opts->ws); 49433ad798cSAdam Langley } 49533ad798cSAdam Langley 49633ad798cSAdam Langley if (unlikely(opts->num_sack_blocks)) { 49733ad798cSAdam Langley struct tcp_sack_block *sp = tp->rx_opt.dsack ? 49833ad798cSAdam Langley tp->duplicate_sack : tp->selective_acks; 49940efc6faSStephen Hemminger int this_sack; 50040efc6faSStephen Hemminger 50140efc6faSStephen Hemminger *ptr++ = htonl((TCPOPT_NOP << 24) | 50240efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 50340efc6faSStephen Hemminger (TCPOPT_SACK << 8) | 50433ad798cSAdam Langley (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * 50540efc6faSStephen Hemminger TCPOLEN_SACK_PERBLOCK))); 5062de979bdSStephen Hemminger 50733ad798cSAdam Langley for (this_sack = 0; this_sack < opts->num_sack_blocks; 50833ad798cSAdam Langley ++this_sack) { 50940efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].start_seq); 51040efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].end_seq); 51140efc6faSStephen Hemminger } 5122de979bdSStephen Hemminger 51340efc6faSStephen Hemminger tp->rx_opt.dsack = 0; 51440efc6faSStephen Hemminger } 5152100c8d2SYuchung Cheng 5162100c8d2SYuchung Cheng if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) { 5172100c8d2SYuchung Cheng struct tcp_fastopen_cookie *foc = opts->fastopen_cookie; 5187f9b838bSDaniel Lee u8 *p = (u8 *)ptr; 5197f9b838bSDaniel Lee u32 len; /* Fast Open option length */ 5202100c8d2SYuchung Cheng 5217f9b838bSDaniel Lee if (foc->exp) { 5227f9b838bSDaniel Lee len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len; 5237f9b838bSDaniel Lee *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) | 5242100c8d2SYuchung Cheng TCPOPT_FASTOPEN_MAGIC); 5257f9b838bSDaniel Lee p += TCPOLEN_EXP_FASTOPEN_BASE; 5267f9b838bSDaniel Lee } else { 5277f9b838bSDaniel Lee len = TCPOLEN_FASTOPEN_BASE + foc->len; 5287f9b838bSDaniel Lee *p++ = TCPOPT_FASTOPEN; 5297f9b838bSDaniel Lee *p++ = len; 5302100c8d2SYuchung Cheng } 5317f9b838bSDaniel Lee 5327f9b838bSDaniel Lee memcpy(p, foc->val, foc->len); 5337f9b838bSDaniel Lee if ((len & 3) == 2) { 5347f9b838bSDaniel Lee p[foc->len] = TCPOPT_NOP; 5357f9b838bSDaniel Lee p[foc->len + 1] = TCPOPT_NOP; 5367f9b838bSDaniel Lee } 5377f9b838bSDaniel Lee ptr += (len + 3) >> 2; 5382100c8d2SYuchung Cheng } 53940efc6faSStephen Hemminger } 54040efc6faSStephen Hemminger 54167edfef7SAndi Kleen /* Compute TCP options for SYN packets. This is not the final 54267edfef7SAndi Kleen * network wire format yet. 54367edfef7SAndi Kleen */ 54495c96174SEric Dumazet static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb, 54533ad798cSAdam Langley struct tcp_out_options *opts, 546cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 547cf533ea5SEric Dumazet { 54833ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 54995c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 550783237e8SYuchung Cheng struct tcp_fastopen_request *fastopen = tp->fastopen_req; 55133ad798cSAdam Langley 552cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 55333ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 55433ad798cSAdam Langley if (*md5) { 55533ad798cSAdam Langley opts->options |= OPTION_MD5; 556bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 557cfb6eeb4SYOSHIFUJI Hideaki } 55833ad798cSAdam Langley #else 55933ad798cSAdam Langley *md5 = NULL; 560cfb6eeb4SYOSHIFUJI Hideaki #endif 56133ad798cSAdam Langley 56233ad798cSAdam Langley /* We always get an MSS option. The option bytes which will be seen in 56333ad798cSAdam Langley * normal data packets should timestamps be used, must be in the MSS 56433ad798cSAdam Langley * advertised. But we subtract them from tp->mss_cache so that 56533ad798cSAdam Langley * calculations in tcp_sendmsg are simpler etc. So account for this 56633ad798cSAdam Langley * fact here if necessary. If we don't do this correctly, as a 56733ad798cSAdam Langley * receiver we won't recognize data packets as being full sized when we 56833ad798cSAdam Langley * should, and thus we won't abide by the delayed ACK rules correctly. 56933ad798cSAdam Langley * SACKs don't matter, we never delay an ACK when we have any of those 57033ad798cSAdam Langley * going out. */ 57133ad798cSAdam Langley opts->mss = tcp_advertise_mss(sk); 572bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 57333ad798cSAdam Langley 5745d2ed052SEric Dumazet if (likely(sock_net(sk)->ipv4.sysctl_tcp_timestamps && !*md5)) { 57533ad798cSAdam Langley opts->options |= OPTION_TS; 5767faee5c0SEric Dumazet opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset; 57733ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 578bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 57933ad798cSAdam Langley } 5809bb37ef0SEric Dumazet if (likely(sock_net(sk)->ipv4.sysctl_tcp_window_scaling)) { 58133ad798cSAdam Langley opts->ws = tp->rx_opt.rcv_wscale; 58289e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 583bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 58433ad798cSAdam Langley } 585f9301034SEric Dumazet if (likely(sock_net(sk)->ipv4.sysctl_tcp_sack)) { 58633ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 587b32d1310SDavid S. Miller if (unlikely(!(OPTION_TS & opts->options))) 588bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 58933ad798cSAdam Langley } 59033ad798cSAdam Langley 591783237e8SYuchung Cheng if (fastopen && fastopen->cookie.len >= 0) { 5922646c831SDaniel Lee u32 need = fastopen->cookie.len; 5932646c831SDaniel Lee 5942646c831SDaniel Lee need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE : 5952646c831SDaniel Lee TCPOLEN_FASTOPEN_BASE; 596783237e8SYuchung Cheng need = (need + 3) & ~3U; /* Align to 32 bits */ 597783237e8SYuchung Cheng if (remaining >= need) { 598783237e8SYuchung Cheng opts->options |= OPTION_FAST_OPEN_COOKIE; 599783237e8SYuchung Cheng opts->fastopen_cookie = &fastopen->cookie; 600783237e8SYuchung Cheng remaining -= need; 601783237e8SYuchung Cheng tp->syn_fastopen = 1; 6022646c831SDaniel Lee tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0; 603783237e8SYuchung Cheng } 604783237e8SYuchung Cheng } 605bd0388aeSWilliam Allen Simpson 606bd0388aeSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 60733ad798cSAdam Langley } 60833ad798cSAdam Langley 60967edfef7SAndi Kleen /* Set up TCP options for SYN-ACKs. */ 61037bfbddaSEric Dumazet static unsigned int tcp_synack_options(struct request_sock *req, 61195c96174SEric Dumazet unsigned int mss, struct sk_buff *skb, 61233ad798cSAdam Langley struct tcp_out_options *opts, 61380f03e27SEric Dumazet const struct tcp_md5sig_key *md5, 6148336886fSJerry Chu struct tcp_fastopen_cookie *foc) 6154957faadSWilliam Allen Simpson { 61633ad798cSAdam Langley struct inet_request_sock *ireq = inet_rsk(req); 61795c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 61833ad798cSAdam Langley 61933ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 62080f03e27SEric Dumazet if (md5) { 62133ad798cSAdam Langley opts->options |= OPTION_MD5; 6224957faadSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 6234957faadSWilliam Allen Simpson 6244957faadSWilliam Allen Simpson /* We can't fit any SACK blocks in a packet with MD5 + TS 6254957faadSWilliam Allen Simpson * options. There was discussion about disabling SACK 6264957faadSWilliam Allen Simpson * rather than TS in order to fit in better with old, 6274957faadSWilliam Allen Simpson * buggy kernels, but that was deemed to be unnecessary. 6284957faadSWilliam Allen Simpson */ 629de213e5eSEric Dumazet ireq->tstamp_ok &= !ireq->sack_ok; 63033ad798cSAdam Langley } 63133ad798cSAdam Langley #endif 63233ad798cSAdam Langley 6334957faadSWilliam Allen Simpson /* We always send an MSS option. */ 63433ad798cSAdam Langley opts->mss = mss; 6354957faadSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 63633ad798cSAdam Langley 63733ad798cSAdam Langley if (likely(ireq->wscale_ok)) { 63833ad798cSAdam Langley opts->ws = ireq->rcv_wscale; 63989e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 6404957faadSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 64133ad798cSAdam Langley } 642de213e5eSEric Dumazet if (likely(ireq->tstamp_ok)) { 64333ad798cSAdam Langley opts->options |= OPTION_TS; 64495a22caeSFlorian Westphal opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off; 64533ad798cSAdam Langley opts->tsecr = req->ts_recent; 6464957faadSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 64733ad798cSAdam Langley } 64833ad798cSAdam Langley if (likely(ireq->sack_ok)) { 64933ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 650de213e5eSEric Dumazet if (unlikely(!ireq->tstamp_ok)) 6514957faadSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 65233ad798cSAdam Langley } 6537f9b838bSDaniel Lee if (foc != NULL && foc->len >= 0) { 6547f9b838bSDaniel Lee u32 need = foc->len; 6557f9b838bSDaniel Lee 6567f9b838bSDaniel Lee need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE : 6577f9b838bSDaniel Lee TCPOLEN_FASTOPEN_BASE; 6588336886fSJerry Chu need = (need + 3) & ~3U; /* Align to 32 bits */ 6598336886fSJerry Chu if (remaining >= need) { 6608336886fSJerry Chu opts->options |= OPTION_FAST_OPEN_COOKIE; 6618336886fSJerry Chu opts->fastopen_cookie = foc; 6628336886fSJerry Chu remaining -= need; 6638336886fSJerry Chu } 6648336886fSJerry Chu } 6654957faadSWilliam Allen Simpson 6664957faadSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 66733ad798cSAdam Langley } 66833ad798cSAdam Langley 66967edfef7SAndi Kleen /* Compute TCP options for ESTABLISHED sockets. This is not the 67067edfef7SAndi Kleen * final wire format yet. 67167edfef7SAndi Kleen */ 67295c96174SEric Dumazet static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb, 67333ad798cSAdam Langley struct tcp_out_options *opts, 674cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 675cf533ea5SEric Dumazet { 67633ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 67795c96174SEric Dumazet unsigned int size = 0; 678cabeccbdSIlpo Järvinen unsigned int eff_sacks; 67933ad798cSAdam Langley 6805843ef42SAndi Kleen opts->options = 0; 6815843ef42SAndi Kleen 68233ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 68333ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 68433ad798cSAdam Langley if (unlikely(*md5)) { 68533ad798cSAdam Langley opts->options |= OPTION_MD5; 68633ad798cSAdam Langley size += TCPOLEN_MD5SIG_ALIGNED; 68733ad798cSAdam Langley } 68833ad798cSAdam Langley #else 68933ad798cSAdam Langley *md5 = NULL; 69033ad798cSAdam Langley #endif 69133ad798cSAdam Langley 69233ad798cSAdam Langley if (likely(tp->rx_opt.tstamp_ok)) { 69333ad798cSAdam Langley opts->options |= OPTION_TS; 6947faee5c0SEric Dumazet opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0; 69533ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 69633ad798cSAdam Langley size += TCPOLEN_TSTAMP_ALIGNED; 69733ad798cSAdam Langley } 69833ad798cSAdam Langley 699cabeccbdSIlpo Järvinen eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; 700cabeccbdSIlpo Järvinen if (unlikely(eff_sacks)) { 70195c96174SEric Dumazet const unsigned int remaining = MAX_TCP_OPTION_SPACE - size; 70233ad798cSAdam Langley opts->num_sack_blocks = 70395c96174SEric Dumazet min_t(unsigned int, eff_sacks, 70433ad798cSAdam Langley (remaining - TCPOLEN_SACK_BASE_ALIGNED) / 70533ad798cSAdam Langley TCPOLEN_SACK_PERBLOCK); 70633ad798cSAdam Langley size += TCPOLEN_SACK_BASE_ALIGNED + 70733ad798cSAdam Langley opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; 70833ad798cSAdam Langley } 70933ad798cSAdam Langley 71033ad798cSAdam Langley return size; 71140efc6faSStephen Hemminger } 7121da177e4SLinus Torvalds 71346d3ceabSEric Dumazet 71446d3ceabSEric Dumazet /* TCP SMALL QUEUES (TSQ) 71546d3ceabSEric Dumazet * 71646d3ceabSEric Dumazet * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev) 71746d3ceabSEric Dumazet * to reduce RTT and bufferbloat. 71846d3ceabSEric Dumazet * We do this using a special skb destructor (tcp_wfree). 71946d3ceabSEric Dumazet * 72046d3ceabSEric Dumazet * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb 72146d3ceabSEric Dumazet * needs to be reallocated in a driver. 7228e3bff96Sstephen hemminger * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc 72346d3ceabSEric Dumazet * 72446d3ceabSEric Dumazet * Since transmit from skb destructor is forbidden, we use a tasklet 72546d3ceabSEric Dumazet * to process all sockets that eventually need to send more skbs. 72646d3ceabSEric Dumazet * We use one tasklet per cpu, with its own queue of sockets. 72746d3ceabSEric Dumazet */ 72846d3ceabSEric Dumazet struct tsq_tasklet { 72946d3ceabSEric Dumazet struct tasklet_struct tasklet; 73046d3ceabSEric Dumazet struct list_head head; /* queue of tcp sockets */ 73146d3ceabSEric Dumazet }; 73246d3ceabSEric Dumazet static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet); 73346d3ceabSEric Dumazet 7346f458dfbSEric Dumazet static void tcp_tsq_handler(struct sock *sk) 7356f458dfbSEric Dumazet { 7366f458dfbSEric Dumazet if ((1 << sk->sk_state) & 7376f458dfbSEric Dumazet (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING | 738f9616c35SEric Dumazet TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) { 739f9616c35SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 740f9616c35SEric Dumazet 741f9616c35SEric Dumazet if (tp->lost_out > tp->retrans_out && 742f9616c35SEric Dumazet tp->snd_cwnd > tcp_packets_in_flight(tp)) 743f9616c35SEric Dumazet tcp_xmit_retransmit_queue(sk); 744f9616c35SEric Dumazet 745f9616c35SEric Dumazet tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle, 746bf06200eSJohn Ogness 0, GFP_ATOMIC); 7476f458dfbSEric Dumazet } 748f9616c35SEric Dumazet } 74946d3ceabSEric Dumazet /* 7508e3bff96Sstephen hemminger * One tasklet per cpu tries to send more skbs. 75146d3ceabSEric Dumazet * We run in tasklet context but need to disable irqs when 7528e3bff96Sstephen hemminger * transferring tsq->head because tcp_wfree() might 75346d3ceabSEric Dumazet * interrupt us (non NAPI drivers) 75446d3ceabSEric Dumazet */ 75546d3ceabSEric Dumazet static void tcp_tasklet_func(unsigned long data) 75646d3ceabSEric Dumazet { 75746d3ceabSEric Dumazet struct tsq_tasklet *tsq = (struct tsq_tasklet *)data; 75846d3ceabSEric Dumazet LIST_HEAD(list); 75946d3ceabSEric Dumazet unsigned long flags; 76046d3ceabSEric Dumazet struct list_head *q, *n; 76146d3ceabSEric Dumazet struct tcp_sock *tp; 76246d3ceabSEric Dumazet struct sock *sk; 76346d3ceabSEric Dumazet 76446d3ceabSEric Dumazet local_irq_save(flags); 76546d3ceabSEric Dumazet list_splice_init(&tsq->head, &list); 76646d3ceabSEric Dumazet local_irq_restore(flags); 76746d3ceabSEric Dumazet 76846d3ceabSEric Dumazet list_for_each_safe(q, n, &list) { 76946d3ceabSEric Dumazet tp = list_entry(q, struct tcp_sock, tsq_node); 77046d3ceabSEric Dumazet list_del(&tp->tsq_node); 77146d3ceabSEric Dumazet 77246d3ceabSEric Dumazet sk = (struct sock *)tp; 7730a9648f1SEric Dumazet smp_mb__before_atomic(); 7747aa5470cSEric Dumazet clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags); 7757aa5470cSEric Dumazet 776b223feb9SEric Dumazet if (!sk->sk_lock.owned && 7777aa5470cSEric Dumazet test_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags)) { 77846d3ceabSEric Dumazet bh_lock_sock(sk); 77946d3ceabSEric Dumazet if (!sock_owned_by_user(sk)) { 7807aa5470cSEric Dumazet clear_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags); 7816f458dfbSEric Dumazet tcp_tsq_handler(sk); 78246d3ceabSEric Dumazet } 78346d3ceabSEric Dumazet bh_unlock_sock(sk); 784b223feb9SEric Dumazet } 78546d3ceabSEric Dumazet 78646d3ceabSEric Dumazet sk_free(sk); 78746d3ceabSEric Dumazet } 78846d3ceabSEric Dumazet } 78946d3ceabSEric Dumazet 79040fc3423SEric Dumazet #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \ 79140fc3423SEric Dumazet TCPF_WRITE_TIMER_DEFERRED | \ 79240fc3423SEric Dumazet TCPF_DELACK_TIMER_DEFERRED | \ 79340fc3423SEric Dumazet TCPF_MTU_REDUCED_DEFERRED) 79446d3ceabSEric Dumazet /** 79546d3ceabSEric Dumazet * tcp_release_cb - tcp release_sock() callback 79646d3ceabSEric Dumazet * @sk: socket 79746d3ceabSEric Dumazet * 79846d3ceabSEric Dumazet * called from release_sock() to perform protocol dependent 79946d3ceabSEric Dumazet * actions before socket release. 80046d3ceabSEric Dumazet */ 80146d3ceabSEric Dumazet void tcp_release_cb(struct sock *sk) 80246d3ceabSEric Dumazet { 8036f458dfbSEric Dumazet unsigned long flags, nflags; 80446d3ceabSEric Dumazet 8056f458dfbSEric Dumazet /* perform an atomic operation only if at least one flag is set */ 8066f458dfbSEric Dumazet do { 8077aa5470cSEric Dumazet flags = sk->sk_tsq_flags; 8086f458dfbSEric Dumazet if (!(flags & TCP_DEFERRED_ALL)) 8096f458dfbSEric Dumazet return; 8106f458dfbSEric Dumazet nflags = flags & ~TCP_DEFERRED_ALL; 8117aa5470cSEric Dumazet } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags); 8126f458dfbSEric Dumazet 81340fc3423SEric Dumazet if (flags & TCPF_TSQ_DEFERRED) 8146f458dfbSEric Dumazet tcp_tsq_handler(sk); 8156f458dfbSEric Dumazet 816c3f9b018SEric Dumazet /* Here begins the tricky part : 817c3f9b018SEric Dumazet * We are called from release_sock() with : 818c3f9b018SEric Dumazet * 1) BH disabled 819c3f9b018SEric Dumazet * 2) sk_lock.slock spinlock held 820c3f9b018SEric Dumazet * 3) socket owned by us (sk->sk_lock.owned == 1) 821c3f9b018SEric Dumazet * 822c3f9b018SEric Dumazet * But following code is meant to be called from BH handlers, 823c3f9b018SEric Dumazet * so we should keep BH disabled, but early release socket ownership 824c3f9b018SEric Dumazet */ 825c3f9b018SEric Dumazet sock_release_ownership(sk); 826c3f9b018SEric Dumazet 82740fc3423SEric Dumazet if (flags & TCPF_WRITE_TIMER_DEFERRED) { 8286f458dfbSEric Dumazet tcp_write_timer_handler(sk); 829144d56e9SEric Dumazet __sock_put(sk); 830144d56e9SEric Dumazet } 83140fc3423SEric Dumazet if (flags & TCPF_DELACK_TIMER_DEFERRED) { 8326f458dfbSEric Dumazet tcp_delack_timer_handler(sk); 833144d56e9SEric Dumazet __sock_put(sk); 834144d56e9SEric Dumazet } 83540fc3423SEric Dumazet if (flags & TCPF_MTU_REDUCED_DEFERRED) { 8364fab9071SNeal Cardwell inet_csk(sk)->icsk_af_ops->mtu_reduced(sk); 837144d56e9SEric Dumazet __sock_put(sk); 838144d56e9SEric Dumazet } 83946d3ceabSEric Dumazet } 84046d3ceabSEric Dumazet EXPORT_SYMBOL(tcp_release_cb); 84146d3ceabSEric Dumazet 84246d3ceabSEric Dumazet void __init tcp_tasklet_init(void) 84346d3ceabSEric Dumazet { 84446d3ceabSEric Dumazet int i; 84546d3ceabSEric Dumazet 84646d3ceabSEric Dumazet for_each_possible_cpu(i) { 84746d3ceabSEric Dumazet struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i); 84846d3ceabSEric Dumazet 84946d3ceabSEric Dumazet INIT_LIST_HEAD(&tsq->head); 85046d3ceabSEric Dumazet tasklet_init(&tsq->tasklet, 85146d3ceabSEric Dumazet tcp_tasklet_func, 85246d3ceabSEric Dumazet (unsigned long)tsq); 85346d3ceabSEric Dumazet } 85446d3ceabSEric Dumazet } 85546d3ceabSEric Dumazet 85646d3ceabSEric Dumazet /* 85746d3ceabSEric Dumazet * Write buffer destructor automatically called from kfree_skb. 8588e3bff96Sstephen hemminger * We can't xmit new skbs from this context, as we might already 85946d3ceabSEric Dumazet * hold qdisc lock. 86046d3ceabSEric Dumazet */ 861d6a4a104SEric Dumazet void tcp_wfree(struct sk_buff *skb) 86246d3ceabSEric Dumazet { 86346d3ceabSEric Dumazet struct sock *sk = skb->sk; 86446d3ceabSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 865408f0a6cSEric Dumazet unsigned long flags, nval, oval; 8669b462d02SEric Dumazet 8679b462d02SEric Dumazet /* Keep one reference on sk_wmem_alloc. 8689b462d02SEric Dumazet * Will be released by sk_free() from here or tcp_tasklet_func() 8699b462d02SEric Dumazet */ 87014afee4bSReshetova, Elena WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc)); 8719b462d02SEric Dumazet 8729b462d02SEric Dumazet /* If this softirq is serviced by ksoftirqd, we are likely under stress. 8739b462d02SEric Dumazet * Wait until our queues (qdisc + devices) are drained. 8749b462d02SEric Dumazet * This gives : 8759b462d02SEric Dumazet * - less callbacks to tcp_write_xmit(), reducing stress (batches) 8769b462d02SEric Dumazet * - chance for incoming ACK (processed by another cpu maybe) 8779b462d02SEric Dumazet * to migrate this flow (skb->ooo_okay will be eventually set) 8789b462d02SEric Dumazet */ 87914afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current) 8809b462d02SEric Dumazet goto out; 88146d3ceabSEric Dumazet 8827aa5470cSEric Dumazet for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) { 88346d3ceabSEric Dumazet struct tsq_tasklet *tsq; 884a9b204d1SEric Dumazet bool empty; 88546d3ceabSEric Dumazet 886408f0a6cSEric Dumazet if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED)) 887408f0a6cSEric Dumazet goto out; 888408f0a6cSEric Dumazet 889b223feb9SEric Dumazet nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED | TCPF_TSQ_DEFERRED; 8907aa5470cSEric Dumazet nval = cmpxchg(&sk->sk_tsq_flags, oval, nval); 891408f0a6cSEric Dumazet if (nval != oval) 892408f0a6cSEric Dumazet continue; 893408f0a6cSEric Dumazet 89446d3ceabSEric Dumazet /* queue this socket to tasklet queue */ 89546d3ceabSEric Dumazet local_irq_save(flags); 896903ceff7SChristoph Lameter tsq = this_cpu_ptr(&tsq_tasklet); 897a9b204d1SEric Dumazet empty = list_empty(&tsq->head); 89846d3ceabSEric Dumazet list_add(&tp->tsq_node, &tsq->head); 899a9b204d1SEric Dumazet if (empty) 90046d3ceabSEric Dumazet tasklet_schedule(&tsq->tasklet); 90146d3ceabSEric Dumazet local_irq_restore(flags); 9029b462d02SEric Dumazet return; 90346d3ceabSEric Dumazet } 9049b462d02SEric Dumazet out: 9059b462d02SEric Dumazet sk_free(sk); 90646d3ceabSEric Dumazet } 90746d3ceabSEric Dumazet 908218af599SEric Dumazet /* Note: Called under hard irq. 909218af599SEric Dumazet * We can not call TCP stack right away. 910218af599SEric Dumazet */ 911218af599SEric Dumazet enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer) 912218af599SEric Dumazet { 913218af599SEric Dumazet struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer); 914218af599SEric Dumazet struct sock *sk = (struct sock *)tp; 915218af599SEric Dumazet unsigned long nval, oval; 916218af599SEric Dumazet 917218af599SEric Dumazet for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) { 918218af599SEric Dumazet struct tsq_tasklet *tsq; 919218af599SEric Dumazet bool empty; 920218af599SEric Dumazet 921218af599SEric Dumazet if (oval & TSQF_QUEUED) 922218af599SEric Dumazet break; 923218af599SEric Dumazet 924218af599SEric Dumazet nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED | TCPF_TSQ_DEFERRED; 925218af599SEric Dumazet nval = cmpxchg(&sk->sk_tsq_flags, oval, nval); 926218af599SEric Dumazet if (nval != oval) 927218af599SEric Dumazet continue; 928218af599SEric Dumazet 92914afee4bSReshetova, Elena if (!refcount_inc_not_zero(&sk->sk_wmem_alloc)) 930218af599SEric Dumazet break; 931218af599SEric Dumazet /* queue this socket to tasklet queue */ 932218af599SEric Dumazet tsq = this_cpu_ptr(&tsq_tasklet); 933218af599SEric Dumazet empty = list_empty(&tsq->head); 934218af599SEric Dumazet list_add(&tp->tsq_node, &tsq->head); 935218af599SEric Dumazet if (empty) 936218af599SEric Dumazet tasklet_schedule(&tsq->tasklet); 937218af599SEric Dumazet break; 938218af599SEric Dumazet } 939218af599SEric Dumazet return HRTIMER_NORESTART; 940218af599SEric Dumazet } 941218af599SEric Dumazet 942218af599SEric Dumazet /* BBR congestion control needs pacing. 943218af599SEric Dumazet * Same remark for SO_MAX_PACING_RATE. 944218af599SEric Dumazet * sch_fq packet scheduler is efficiently handling pacing, 945218af599SEric Dumazet * but is not always installed/used. 946218af599SEric Dumazet * Return true if TCP stack should pace packets itself. 947218af599SEric Dumazet */ 948218af599SEric Dumazet static bool tcp_needs_internal_pacing(const struct sock *sk) 949218af599SEric Dumazet { 950218af599SEric Dumazet return smp_load_acquire(&sk->sk_pacing_status) == SK_PACING_NEEDED; 951218af599SEric Dumazet } 952218af599SEric Dumazet 953218af599SEric Dumazet static void tcp_internal_pacing(struct sock *sk, const struct sk_buff *skb) 954218af599SEric Dumazet { 955218af599SEric Dumazet u64 len_ns; 956218af599SEric Dumazet u32 rate; 957218af599SEric Dumazet 958218af599SEric Dumazet if (!tcp_needs_internal_pacing(sk)) 959218af599SEric Dumazet return; 960218af599SEric Dumazet rate = sk->sk_pacing_rate; 961218af599SEric Dumazet if (!rate || rate == ~0U) 962218af599SEric Dumazet return; 963218af599SEric Dumazet 964218af599SEric Dumazet /* Should account for header sizes as sch_fq does, 965218af599SEric Dumazet * but lets make things simple. 966218af599SEric Dumazet */ 967218af599SEric Dumazet len_ns = (u64)skb->len * NSEC_PER_SEC; 968218af599SEric Dumazet do_div(len_ns, rate); 969218af599SEric Dumazet hrtimer_start(&tcp_sk(sk)->pacing_timer, 970218af599SEric Dumazet ktime_add_ns(ktime_get(), len_ns), 971218af599SEric Dumazet HRTIMER_MODE_ABS_PINNED); 972218af599SEric Dumazet } 973218af599SEric Dumazet 974*e2080072SEric Dumazet static void tcp_update_skb_after_send(struct tcp_sock *tp, struct sk_buff *skb) 975*e2080072SEric Dumazet { 976*e2080072SEric Dumazet skb->skb_mstamp = tp->tcp_mstamp; 977*e2080072SEric Dumazet list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue); 978*e2080072SEric Dumazet } 979*e2080072SEric Dumazet 9801da177e4SLinus Torvalds /* This routine actually transmits TCP packets queued in by 9811da177e4SLinus Torvalds * tcp_do_sendmsg(). This is used by both the initial 9821da177e4SLinus Torvalds * transmission and possible later retransmissions. 9831da177e4SLinus Torvalds * All SKB's seen here are completely headerless. It is our 9841da177e4SLinus Torvalds * job to build the TCP header, and pass the packet down to 9851da177e4SLinus Torvalds * IP so it can do the same plus pass the packet off to the 9861da177e4SLinus Torvalds * device. 9871da177e4SLinus Torvalds * 9881da177e4SLinus Torvalds * We are working here with either a clone of the original 9891da177e4SLinus Torvalds * SKB, or a fresh unique copy made by the retransmit engine. 9901da177e4SLinus Torvalds */ 991056834d9SIlpo Järvinen static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, 992056834d9SIlpo Järvinen gfp_t gfp_mask) 9931da177e4SLinus Torvalds { 9946687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 995dfb4b9dcSDavid S. Miller struct inet_sock *inet; 996dfb4b9dcSDavid S. Miller struct tcp_sock *tp; 997dfb4b9dcSDavid S. Miller struct tcp_skb_cb *tcb; 99833ad798cSAdam Langley struct tcp_out_options opts; 99995c96174SEric Dumazet unsigned int tcp_options_size, tcp_header_size; 10008c72c65bSEric Dumazet struct sk_buff *oskb = NULL; 1001cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 10021da177e4SLinus Torvalds struct tcphdr *th; 10031da177e4SLinus Torvalds int err; 10041da177e4SLinus Torvalds 1005dfb4b9dcSDavid S. Miller BUG_ON(!skb || !tcp_skb_pcount(skb)); 10066f094b9eSLawrence Brakmo tp = tcp_sk(sk); 1007dfb4b9dcSDavid S. Miller 1008ccdbb6e9SEric Dumazet if (clone_it) { 10096f094b9eSLawrence Brakmo TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq 10106f094b9eSLawrence Brakmo - tp->snd_una; 10118c72c65bSEric Dumazet oskb = skb; 1012*e2080072SEric Dumazet 1013*e2080072SEric Dumazet tcp_skb_tsorted_save(oskb) { 1014*e2080072SEric Dumazet if (unlikely(skb_cloned(oskb))) 1015*e2080072SEric Dumazet skb = pskb_copy(oskb, gfp_mask); 1016dfb4b9dcSDavid S. Miller else 1017*e2080072SEric Dumazet skb = skb_clone(oskb, gfp_mask); 1018*e2080072SEric Dumazet } tcp_skb_tsorted_restore(oskb); 1019*e2080072SEric Dumazet 1020dfb4b9dcSDavid S. Miller if (unlikely(!skb)) 1021dfb4b9dcSDavid S. Miller return -ENOBUFS; 1022dfb4b9dcSDavid S. Miller } 10238c72c65bSEric Dumazet skb->skb_mstamp = tp->tcp_mstamp; 1024dfb4b9dcSDavid S. Miller 1025dfb4b9dcSDavid S. Miller inet = inet_sk(sk); 1026dfb4b9dcSDavid S. Miller tcb = TCP_SKB_CB(skb); 102733ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 10281da177e4SLinus Torvalds 10294de075e0SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) 103033ad798cSAdam Langley tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); 103133ad798cSAdam Langley else 103233ad798cSAdam Langley tcp_options_size = tcp_established_options(sk, skb, &opts, 103333ad798cSAdam Langley &md5); 103433ad798cSAdam Langley tcp_header_size = tcp_options_size + sizeof(struct tcphdr); 10351da177e4SLinus Torvalds 1036547669d4SEric Dumazet /* if no packet is in qdisc/device queue, then allow XPS to select 1037b2532eb9SEric Dumazet * another queue. We can be called from tcp_tsq_handler() 1038b2532eb9SEric Dumazet * which holds one reference to sk_wmem_alloc. 1039b2532eb9SEric Dumazet * 1040b2532eb9SEric Dumazet * TODO: Ideally, in-flight pure ACK packets should not matter here. 1041b2532eb9SEric Dumazet * One way to get this would be to set skb->truesize = 2 on them. 1042547669d4SEric Dumazet */ 1043b2532eb9SEric Dumazet skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1); 10441da177e4SLinus Torvalds 104538ab52e8SEric Dumazet /* If we had to use memory reserve to allocate this skb, 104638ab52e8SEric Dumazet * this might cause drops if packet is looped back : 104738ab52e8SEric Dumazet * Other socket might not have SOCK_MEMALLOC. 104838ab52e8SEric Dumazet * Packets not looped back do not care about pfmemalloc. 104938ab52e8SEric Dumazet */ 105038ab52e8SEric Dumazet skb->pfmemalloc = 0; 105138ab52e8SEric Dumazet 1052aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 1053aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 105446d3ceabSEric Dumazet 105546d3ceabSEric Dumazet skb_orphan(skb); 105646d3ceabSEric Dumazet skb->sk = sk; 10571d2077acSEric Dumazet skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree; 1058b73c3d0eSTom Herbert skb_set_hash_from_sk(skb, sk); 105914afee4bSReshetova, Elena refcount_add(skb->truesize, &sk->sk_wmem_alloc); 10601da177e4SLinus Torvalds 1061c3a2e837SJulian Anastasov skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm); 1062c3a2e837SJulian Anastasov 10631da177e4SLinus Torvalds /* Build TCP header and checksum it. */ 1064ea1627c2SEric Dumazet th = (struct tcphdr *)skb->data; 1065c720c7e8SEric Dumazet th->source = inet->inet_sport; 1066c720c7e8SEric Dumazet th->dest = inet->inet_dport; 10671da177e4SLinus Torvalds th->seq = htonl(tcb->seq); 10681da177e4SLinus Torvalds th->ack_seq = htonl(tp->rcv_nxt); 1069df7a3b07SAl Viro *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | 10704de075e0SEric Dumazet tcb->tcp_flags); 1071dfb4b9dcSDavid S. Miller 10721da177e4SLinus Torvalds th->check = 0; 10731da177e4SLinus Torvalds th->urg_ptr = 0; 10741da177e4SLinus Torvalds 107533f5f57eSIlpo Järvinen /* The urg_mode check is necessary during a below snd_una win probe */ 10767691367dSHerbert Xu if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { 10777691367dSHerbert Xu if (before(tp->snd_up, tcb->seq + 0x10000)) { 10781da177e4SLinus Torvalds th->urg_ptr = htons(tp->snd_up - tcb->seq); 10791da177e4SLinus Torvalds th->urg = 1; 10807691367dSHerbert Xu } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { 10810eae88f3SEric Dumazet th->urg_ptr = htons(0xFFFF); 10827691367dSHerbert Xu th->urg = 1; 10837691367dSHerbert Xu } 10841da177e4SLinus Torvalds } 10851da177e4SLinus Torvalds 1086bd0388aeSWilliam Allen Simpson tcp_options_write((__be32 *)(th + 1), tp, &opts); 108751466a75SEric Dumazet skb_shinfo(skb)->gso_type = sk->sk_gso_type; 1088ea1627c2SEric Dumazet if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) { 1089ea1627c2SEric Dumazet th->window = htons(tcp_select_window(sk)); 1090ea1627c2SEric Dumazet tcp_ecn_send(sk, skb, th, tcp_header_size); 1091ea1627c2SEric Dumazet } else { 1092ea1627c2SEric Dumazet /* RFC1323: The window in SYN & SYN/ACK segments 1093ea1627c2SEric Dumazet * is never scaled. 1094ea1627c2SEric Dumazet */ 1095ea1627c2SEric Dumazet th->window = htons(min(tp->rcv_wnd, 65535U)); 1096ea1627c2SEric Dumazet } 1097cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 1098cfb6eeb4SYOSHIFUJI Hideaki /* Calculate the MD5 hash, as we have all we need now */ 1099cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 1100a465419bSEric Dumazet sk_nocaps_add(sk, NETIF_F_GSO_MASK); 1101bd0388aeSWilliam Allen Simpson tp->af_specific->calc_md5_hash(opts.hash_location, 110239f8e58eSEric Dumazet md5, sk, skb); 1103cfb6eeb4SYOSHIFUJI Hideaki } 1104cfb6eeb4SYOSHIFUJI Hideaki #endif 1105cfb6eeb4SYOSHIFUJI Hideaki 1106bb296246SHerbert Xu icsk->icsk_af_ops->send_check(sk, skb); 11071da177e4SLinus Torvalds 11084de075e0SEric Dumazet if (likely(tcb->tcp_flags & TCPHDR_ACK)) 1109fc6415bcSDavid S. Miller tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); 11101da177e4SLinus Torvalds 1111a44d6eacSMartin KaFai Lau if (skb->len != tcp_header_size) { 1112cf533ea5SEric Dumazet tcp_event_data_sent(tp, sk); 1113a44d6eacSMartin KaFai Lau tp->data_segs_out += tcp_skb_pcount(skb); 1114218af599SEric Dumazet tcp_internal_pacing(sk, skb); 1115a44d6eacSMartin KaFai Lau } 11161da177e4SLinus Torvalds 1117bd37a088SWei Yongjun if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) 1118aa2ea058STom Herbert TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, 1119aa2ea058STom Herbert tcp_skb_pcount(skb)); 11201da177e4SLinus Torvalds 11212efd055cSMarcelo Ricardo Leitner tp->segs_out += tcp_skb_pcount(skb); 1122f69ad292SEric Dumazet /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */ 1123cd7d8498SEric Dumazet skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb); 1124f69ad292SEric Dumazet skb_shinfo(skb)->gso_size = tcp_skb_mss(skb); 1125cd7d8498SEric Dumazet 11267faee5c0SEric Dumazet /* Our usage of tstamp should remain private */ 11272456e855SThomas Gleixner skb->tstamp = 0; 1128971f10ecSEric Dumazet 1129971f10ecSEric Dumazet /* Cleanup our debris for IP stacks */ 1130971f10ecSEric Dumazet memset(skb->cb, 0, max(sizeof(struct inet_skb_parm), 1131971f10ecSEric Dumazet sizeof(struct inet6_skb_parm))); 1132971f10ecSEric Dumazet 1133b0270e91SEric Dumazet err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl); 11347faee5c0SEric Dumazet 11358c72c65bSEric Dumazet if (unlikely(err > 0)) { 11365ee2c941SChristoph Paasch tcp_enter_cwr(sk); 11378c72c65bSEric Dumazet err = net_xmit_eval(err); 11388c72c65bSEric Dumazet } 1139fc225799SEric Dumazet if (!err && oskb) { 1140*e2080072SEric Dumazet tcp_update_skb_after_send(tp, oskb); 1141fc225799SEric Dumazet tcp_rate_skb_sent(sk, oskb); 1142fc225799SEric Dumazet } 11438c72c65bSEric Dumazet return err; 11441da177e4SLinus Torvalds } 11451da177e4SLinus Torvalds 114667edfef7SAndi Kleen /* This routine just queues the buffer for sending. 11471da177e4SLinus Torvalds * 11481da177e4SLinus Torvalds * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, 11491da177e4SLinus Torvalds * otherwise socket can stall. 11501da177e4SLinus Torvalds */ 11511da177e4SLinus Torvalds static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) 11521da177e4SLinus Torvalds { 11531da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 11541da177e4SLinus Torvalds 11551da177e4SLinus Torvalds /* Advance write_seq and place onto the write_queue. */ 11561da177e4SLinus Torvalds tp->write_seq = TCP_SKB_CB(skb)->end_seq; 1157f4a775d1SEric Dumazet __skb_header_release(skb); 1158fe067e8aSDavid S. Miller tcp_add_write_queue_tail(sk, skb); 11593ab224beSHideo Aoki sk->sk_wmem_queued += skb->truesize; 11603ab224beSHideo Aoki sk_mem_charge(sk, skb->truesize); 11611da177e4SLinus Torvalds } 11621da177e4SLinus Torvalds 116367edfef7SAndi Kleen /* Initialize TSO segments for a packet. */ 11645bbb432cSEric Dumazet static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now) 1165f6302d1dSDavid S. Miller { 11668f26fb1cSEric Dumazet if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) { 1167f6302d1dSDavid S. Miller /* Avoid the costly divide in the normal 1168f6302d1dSDavid S. Miller * non-TSO case. 1169f6302d1dSDavid S. Miller */ 1170cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, 1); 1171f69ad292SEric Dumazet TCP_SKB_CB(skb)->tcp_gso_size = 0; 1172f6302d1dSDavid S. Miller } else { 1173cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now)); 1174f69ad292SEric Dumazet TCP_SKB_CB(skb)->tcp_gso_size = mss_now; 11751da177e4SLinus Torvalds } 11761da177e4SLinus Torvalds } 11771da177e4SLinus Torvalds 117891fed7a1SIlpo Järvinen /* When a modification to fackets out becomes necessary, we need to check 117968f8353bSIlpo Järvinen * skb is counted to fackets_out or not. 118091fed7a1SIlpo Järvinen */ 1181cf533ea5SEric Dumazet static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb, 118291fed7a1SIlpo Järvinen int decr) 118391fed7a1SIlpo Järvinen { 1184a47e5a98SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1185a47e5a98SIlpo Järvinen 1186dc86967bSIlpo Järvinen if (!tp->sacked_out || tcp_is_reno(tp)) 118791fed7a1SIlpo Järvinen return; 118891fed7a1SIlpo Järvinen 11896859d494SIlpo Järvinen if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq)) 119091fed7a1SIlpo Järvinen tp->fackets_out -= decr; 119191fed7a1SIlpo Järvinen } 119291fed7a1SIlpo Järvinen 1193797108d1SIlpo Järvinen /* Pcount in the middle of the write queue got changed, we need to do various 1194797108d1SIlpo Järvinen * tweaks to fix counters 1195797108d1SIlpo Järvinen */ 1196cf533ea5SEric Dumazet static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) 1197797108d1SIlpo Järvinen { 1198797108d1SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1199797108d1SIlpo Järvinen 1200797108d1SIlpo Järvinen tp->packets_out -= decr; 1201797108d1SIlpo Järvinen 1202797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 1203797108d1SIlpo Järvinen tp->sacked_out -= decr; 1204797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 1205797108d1SIlpo Järvinen tp->retrans_out -= decr; 1206797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 1207797108d1SIlpo Järvinen tp->lost_out -= decr; 1208797108d1SIlpo Järvinen 1209797108d1SIlpo Järvinen /* Reno case is special. Sigh... */ 1210797108d1SIlpo Järvinen if (tcp_is_reno(tp) && decr > 0) 1211797108d1SIlpo Järvinen tp->sacked_out -= min_t(u32, tp->sacked_out, decr); 1212797108d1SIlpo Järvinen 1213797108d1SIlpo Järvinen tcp_adjust_fackets_out(sk, skb, decr); 1214797108d1SIlpo Järvinen 1215797108d1SIlpo Järvinen if (tp->lost_skb_hint && 1216797108d1SIlpo Järvinen before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && 121752cf3cc8SIlpo Järvinen (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))) 1218797108d1SIlpo Järvinen tp->lost_cnt_hint -= decr; 1219797108d1SIlpo Järvinen 1220797108d1SIlpo Järvinen tcp_verify_left_out(tp); 1221797108d1SIlpo Järvinen } 1222797108d1SIlpo Järvinen 12230a2cf20cSSoheil Hassas Yeganeh static bool tcp_has_tx_tstamp(const struct sk_buff *skb) 12240a2cf20cSSoheil Hassas Yeganeh { 12250a2cf20cSSoheil Hassas Yeganeh return TCP_SKB_CB(skb)->txstamp_ack || 12260a2cf20cSSoheil Hassas Yeganeh (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP); 12270a2cf20cSSoheil Hassas Yeganeh } 12280a2cf20cSSoheil Hassas Yeganeh 1229490cc7d0SWillem de Bruijn static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2) 1230490cc7d0SWillem de Bruijn { 1231490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo = skb_shinfo(skb); 1232490cc7d0SWillem de Bruijn 12330a2cf20cSSoheil Hassas Yeganeh if (unlikely(tcp_has_tx_tstamp(skb)) && 1234490cc7d0SWillem de Bruijn !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) { 1235490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo2 = skb_shinfo(skb2); 1236490cc7d0SWillem de Bruijn u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP; 1237490cc7d0SWillem de Bruijn 1238490cc7d0SWillem de Bruijn shinfo->tx_flags &= ~tsflags; 1239490cc7d0SWillem de Bruijn shinfo2->tx_flags |= tsflags; 1240490cc7d0SWillem de Bruijn swap(shinfo->tskey, shinfo2->tskey); 1241b51e13faSMartin KaFai Lau TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack; 1242b51e13faSMartin KaFai Lau TCP_SKB_CB(skb)->txstamp_ack = 0; 1243490cc7d0SWillem de Bruijn } 1244490cc7d0SWillem de Bruijn } 1245490cc7d0SWillem de Bruijn 1246a166140eSMartin KaFai Lau static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2) 1247a166140eSMartin KaFai Lau { 1248a166140eSMartin KaFai Lau TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor; 1249a166140eSMartin KaFai Lau TCP_SKB_CB(skb)->eor = 0; 1250a166140eSMartin KaFai Lau } 1251a166140eSMartin KaFai Lau 12521da177e4SLinus Torvalds /* Function to create two new TCP segments. Shrinks the given segment 12531da177e4SLinus Torvalds * to the specified size and appends a new segment with the rest of the 12541da177e4SLinus Torvalds * packet to the list. This won't be called frequently, I hope. 12551da177e4SLinus Torvalds * Remember, these are still headerless SKBs at this point. 12561da177e4SLinus Torvalds */ 1257056834d9SIlpo Järvinen int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, 12586cc55e09SOctavian Purdila unsigned int mss_now, gfp_t gfp) 12591da177e4SLinus Torvalds { 12601da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 12611da177e4SLinus Torvalds struct sk_buff *buff; 12626475be16SDavid S. Miller int nsize, old_factor; 1263b60b49eaSHerbert Xu int nlen; 12649ce01461SIlpo Järvinen u8 flags; 12651da177e4SLinus Torvalds 12662fceec13SIlpo Järvinen if (WARN_ON(len > skb->len)) 12672fceec13SIlpo Järvinen return -EINVAL; 12686a438bbeSStephen Hemminger 12691da177e4SLinus Torvalds nsize = skb_headlen(skb) - len; 12701da177e4SLinus Torvalds if (nsize < 0) 12711da177e4SLinus Torvalds nsize = 0; 12721da177e4SLinus Torvalds 12736cc55e09SOctavian Purdila if (skb_unclone(skb, gfp)) 12741da177e4SLinus Torvalds return -ENOMEM; 12751da177e4SLinus Torvalds 12761da177e4SLinus Torvalds /* Get a new skb... force flag on. */ 1277eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, nsize, gfp, true); 127851456b29SIan Morris if (!buff) 12791da177e4SLinus Torvalds return -ENOMEM; /* We'll just try again later. */ 1280ef5cb973SHerbert Xu 12813ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 12823ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1283b60b49eaSHerbert Xu nlen = skb->len - len - nsize; 1284b60b49eaSHerbert Xu buff->truesize += nlen; 1285b60b49eaSHerbert Xu skb->truesize -= nlen; 12861da177e4SLinus Torvalds 12871da177e4SLinus Torvalds /* Correct the sequence numbers. */ 12881da177e4SLinus Torvalds TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 12891da177e4SLinus Torvalds TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 12901da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 12911da177e4SLinus Torvalds 12921da177e4SLinus Torvalds /* PSH and FIN should only be set in the second packet. */ 12934de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 12944de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 12954de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1296e14c3cafSHerbert Xu TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; 1297a166140eSMartin KaFai Lau tcp_skb_fragment_eor(skb, buff); 12981da177e4SLinus Torvalds 129984fa7933SPatrick McHardy if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { 13001da177e4SLinus Torvalds /* Copy and checksum data tail into the new buffer. */ 1301056834d9SIlpo Järvinen buff->csum = csum_partial_copy_nocheck(skb->data + len, 1302056834d9SIlpo Järvinen skb_put(buff, nsize), 13031da177e4SLinus Torvalds nsize, 0); 13041da177e4SLinus Torvalds 13051da177e4SLinus Torvalds skb_trim(skb, len); 13061da177e4SLinus Torvalds 13071da177e4SLinus Torvalds skb->csum = csum_block_sub(skb->csum, buff->csum, len); 13081da177e4SLinus Torvalds } else { 130984fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 13101da177e4SLinus Torvalds skb_split(skb, buff, len); 13111da177e4SLinus Torvalds } 13121da177e4SLinus Torvalds 13131da177e4SLinus Torvalds buff->ip_summed = skb->ip_summed; 13141da177e4SLinus Torvalds 1315a61bbcf2SPatrick McHardy buff->tstamp = skb->tstamp; 1316490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 13171da177e4SLinus Torvalds 13186475be16SDavid S. Miller old_factor = tcp_skb_pcount(skb); 13196475be16SDavid S. Miller 13201da177e4SLinus Torvalds /* Fix up tso_factor for both original and new SKB. */ 13215bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 13225bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 13231da177e4SLinus Torvalds 1324b9f64820SYuchung Cheng /* Update delivered info for the new segment */ 1325b9f64820SYuchung Cheng TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx; 1326b9f64820SYuchung Cheng 13276475be16SDavid S. Miller /* If this packet has been sent out already, we must 13286475be16SDavid S. Miller * adjust the various packet counters. 13296475be16SDavid S. Miller */ 1330cf0b450cSHerbert Xu if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 13316475be16SDavid S. Miller int diff = old_factor - tcp_skb_pcount(skb) - 13326475be16SDavid S. Miller tcp_skb_pcount(buff); 13331da177e4SLinus Torvalds 1334797108d1SIlpo Järvinen if (diff) 1335797108d1SIlpo Järvinen tcp_adjust_pcount(sk, skb, diff); 13361da177e4SLinus Torvalds } 13371da177e4SLinus Torvalds 13381da177e4SLinus Torvalds /* Link BUFF into the send queue. */ 1339f4a775d1SEric Dumazet __skb_header_release(buff); 1340fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 1341*e2080072SEric Dumazet list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor); 13421da177e4SLinus Torvalds 13431da177e4SLinus Torvalds return 0; 13441da177e4SLinus Torvalds } 13451da177e4SLinus Torvalds 1346f4d01666SEric Dumazet /* This is similar to __pskb_pull_tail(). The difference is that pulled 1347f4d01666SEric Dumazet * data is not copied, but immediately discarded. 13481da177e4SLinus Torvalds */ 13497162fb24SEric Dumazet static int __pskb_trim_head(struct sk_buff *skb, int len) 13501da177e4SLinus Torvalds { 13517b7fc97aSEric Dumazet struct skb_shared_info *shinfo; 13521da177e4SLinus Torvalds int i, k, eat; 13531da177e4SLinus Torvalds 13544fa48bf3SEric Dumazet eat = min_t(int, len, skb_headlen(skb)); 13554fa48bf3SEric Dumazet if (eat) { 13564fa48bf3SEric Dumazet __skb_pull(skb, eat); 13574fa48bf3SEric Dumazet len -= eat; 13584fa48bf3SEric Dumazet if (!len) 13597162fb24SEric Dumazet return 0; 13604fa48bf3SEric Dumazet } 13611da177e4SLinus Torvalds eat = len; 13621da177e4SLinus Torvalds k = 0; 13637b7fc97aSEric Dumazet shinfo = skb_shinfo(skb); 13647b7fc97aSEric Dumazet for (i = 0; i < shinfo->nr_frags; i++) { 13657b7fc97aSEric Dumazet int size = skb_frag_size(&shinfo->frags[i]); 13669e903e08SEric Dumazet 13679e903e08SEric Dumazet if (size <= eat) { 1368aff65da0SIan Campbell skb_frag_unref(skb, i); 13699e903e08SEric Dumazet eat -= size; 13701da177e4SLinus Torvalds } else { 13717b7fc97aSEric Dumazet shinfo->frags[k] = shinfo->frags[i]; 13721da177e4SLinus Torvalds if (eat) { 13737b7fc97aSEric Dumazet shinfo->frags[k].page_offset += eat; 13747b7fc97aSEric Dumazet skb_frag_size_sub(&shinfo->frags[k], eat); 13751da177e4SLinus Torvalds eat = 0; 13761da177e4SLinus Torvalds } 13771da177e4SLinus Torvalds k++; 13781da177e4SLinus Torvalds } 13791da177e4SLinus Torvalds } 13807b7fc97aSEric Dumazet shinfo->nr_frags = k; 13811da177e4SLinus Torvalds 13821da177e4SLinus Torvalds skb->data_len -= len; 13831da177e4SLinus Torvalds skb->len = skb->data_len; 13847162fb24SEric Dumazet return len; 13851da177e4SLinus Torvalds } 13861da177e4SLinus Torvalds 138767edfef7SAndi Kleen /* Remove acked data from a packet in the transmit queue. */ 13881da177e4SLinus Torvalds int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 13891da177e4SLinus Torvalds { 13907162fb24SEric Dumazet u32 delta_truesize; 13917162fb24SEric Dumazet 139214bbd6a5SPravin B Shelar if (skb_unclone(skb, GFP_ATOMIC)) 13931da177e4SLinus Torvalds return -ENOMEM; 13941da177e4SLinus Torvalds 13957162fb24SEric Dumazet delta_truesize = __pskb_trim_head(skb, len); 13961da177e4SLinus Torvalds 13971da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq += len; 139884fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 13991da177e4SLinus Torvalds 14007162fb24SEric Dumazet if (delta_truesize) { 14017162fb24SEric Dumazet skb->truesize -= delta_truesize; 14027162fb24SEric Dumazet sk->sk_wmem_queued -= delta_truesize; 14037162fb24SEric Dumazet sk_mem_uncharge(sk, delta_truesize); 14041da177e4SLinus Torvalds sock_set_flag(sk, SOCK_QUEUE_SHRUNK); 14057162fb24SEric Dumazet } 14061da177e4SLinus Torvalds 14075b35e1e6SNeal Cardwell /* Any change of skb->len requires recalculation of tso factor. */ 14081da177e4SLinus Torvalds if (tcp_skb_pcount(skb) > 1) 14095bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb)); 14101da177e4SLinus Torvalds 14111da177e4SLinus Torvalds return 0; 14121da177e4SLinus Torvalds } 14131da177e4SLinus Torvalds 14141b63edd6SYuchung Cheng /* Calculate MSS not accounting any TCP options. */ 14151b63edd6SYuchung Cheng static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu) 14165d424d5aSJohn Heffner { 1417cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1418cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 14195d424d5aSJohn Heffner int mss_now; 14205d424d5aSJohn Heffner 14215d424d5aSJohn Heffner /* Calculate base mss without TCP options: 14225d424d5aSJohn Heffner It is MMS_S - sizeof(tcphdr) of rfc1122 14235d424d5aSJohn Heffner */ 14245d424d5aSJohn Heffner mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 14255d424d5aSJohn Heffner 142667469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 142767469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 142867469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 142967469601SEric Dumazet 143067469601SEric Dumazet if (dst && dst_allfrag(dst)) 143167469601SEric Dumazet mss_now -= icsk->icsk_af_ops->net_frag_header_len; 143267469601SEric Dumazet } 143367469601SEric Dumazet 14345d424d5aSJohn Heffner /* Clamp it (mss_clamp does not include tcp options) */ 14355d424d5aSJohn Heffner if (mss_now > tp->rx_opt.mss_clamp) 14365d424d5aSJohn Heffner mss_now = tp->rx_opt.mss_clamp; 14375d424d5aSJohn Heffner 14385d424d5aSJohn Heffner /* Now subtract optional transport overhead */ 14395d424d5aSJohn Heffner mss_now -= icsk->icsk_ext_hdr_len; 14405d424d5aSJohn Heffner 14415d424d5aSJohn Heffner /* Then reserve room for full set of TCP options and 8 bytes of data */ 14425d424d5aSJohn Heffner if (mss_now < 48) 14435d424d5aSJohn Heffner mss_now = 48; 14445d424d5aSJohn Heffner return mss_now; 14455d424d5aSJohn Heffner } 14465d424d5aSJohn Heffner 14471b63edd6SYuchung Cheng /* Calculate MSS. Not accounting for SACKs here. */ 14481b63edd6SYuchung Cheng int tcp_mtu_to_mss(struct sock *sk, int pmtu) 14491b63edd6SYuchung Cheng { 14501b63edd6SYuchung Cheng /* Subtract TCP options size, not including SACKs */ 14511b63edd6SYuchung Cheng return __tcp_mtu_to_mss(sk, pmtu) - 14521b63edd6SYuchung Cheng (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr)); 14531b63edd6SYuchung Cheng } 14541b63edd6SYuchung Cheng 14555d424d5aSJohn Heffner /* Inverse of above */ 145667469601SEric Dumazet int tcp_mss_to_mtu(struct sock *sk, int mss) 14575d424d5aSJohn Heffner { 1458cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1459cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 14605d424d5aSJohn Heffner int mtu; 14615d424d5aSJohn Heffner 14625d424d5aSJohn Heffner mtu = mss + 14635d424d5aSJohn Heffner tp->tcp_header_len + 14645d424d5aSJohn Heffner icsk->icsk_ext_hdr_len + 14655d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 14665d424d5aSJohn Heffner 146767469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 146867469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 146967469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 147067469601SEric Dumazet 147167469601SEric Dumazet if (dst && dst_allfrag(dst)) 147267469601SEric Dumazet mtu += icsk->icsk_af_ops->net_frag_header_len; 147367469601SEric Dumazet } 14745d424d5aSJohn Heffner return mtu; 14755d424d5aSJohn Heffner } 1476556c6b46SNeal Cardwell EXPORT_SYMBOL(tcp_mss_to_mtu); 14775d424d5aSJohn Heffner 147867edfef7SAndi Kleen /* MTU probing init per socket */ 14795d424d5aSJohn Heffner void tcp_mtup_init(struct sock *sk) 14805d424d5aSJohn Heffner { 14815d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 14825d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 1483b0f9ca53SFan Du struct net *net = sock_net(sk); 14845d424d5aSJohn Heffner 1485b0f9ca53SFan Du icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1; 14865d424d5aSJohn Heffner icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 14875d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 1488b0f9ca53SFan Du icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss); 14895d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 149005cbc0dbSFan Du if (icsk->icsk_mtup.enabled) 1491c74df29aSEric Dumazet icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; 14925d424d5aSJohn Heffner } 14934bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_mtup_init); 14945d424d5aSJohn Heffner 14951da177e4SLinus Torvalds /* This function synchronize snd mss to current pmtu/exthdr set. 14961da177e4SLinus Torvalds 14971da177e4SLinus Torvalds tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 14981da177e4SLinus Torvalds for TCP options, but includes only bare TCP header. 14991da177e4SLinus Torvalds 15001da177e4SLinus Torvalds tp->rx_opt.mss_clamp is mss negotiated at connection setup. 1501caa20d9aSStephen Hemminger It is minimum of user_mss and mss received with SYN. 15021da177e4SLinus Torvalds It also does not include TCP options. 15031da177e4SLinus Torvalds 1504d83d8461SArnaldo Carvalho de Melo inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. 15051da177e4SLinus Torvalds 15061da177e4SLinus Torvalds tp->mss_cache is current effective sending mss, including 15071da177e4SLinus Torvalds all tcp options except for SACKs. It is evaluated, 15081da177e4SLinus Torvalds taking into account current pmtu, but never exceeds 15091da177e4SLinus Torvalds tp->rx_opt.mss_clamp. 15101da177e4SLinus Torvalds 15111da177e4SLinus Torvalds NOTE1. rfc1122 clearly states that advertised MSS 15121da177e4SLinus Torvalds DOES NOT include either tcp or ip options. 15131da177e4SLinus Torvalds 1514d83d8461SArnaldo Carvalho de Melo NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache 1515d83d8461SArnaldo Carvalho de Melo are READ ONLY outside this function. --ANK (980731) 15161da177e4SLinus Torvalds */ 15171da177e4SLinus Torvalds unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 15181da177e4SLinus Torvalds { 15191da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1520d83d8461SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 15215d424d5aSJohn Heffner int mss_now; 15221da177e4SLinus Torvalds 15235d424d5aSJohn Heffner if (icsk->icsk_mtup.search_high > pmtu) 15245d424d5aSJohn Heffner icsk->icsk_mtup.search_high = pmtu; 15251da177e4SLinus Torvalds 15265d424d5aSJohn Heffner mss_now = tcp_mtu_to_mss(sk, pmtu); 1527409d22b4SIlpo Järvinen mss_now = tcp_bound_to_half_wnd(tp, mss_now); 15281da177e4SLinus Torvalds 15291da177e4SLinus Torvalds /* And store cached results */ 1530d83d8461SArnaldo Carvalho de Melo icsk->icsk_pmtu_cookie = pmtu; 15315d424d5aSJohn Heffner if (icsk->icsk_mtup.enabled) 15325d424d5aSJohn Heffner mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); 1533c1b4a7e6SDavid S. Miller tp->mss_cache = mss_now; 15341da177e4SLinus Torvalds 15351da177e4SLinus Torvalds return mss_now; 15361da177e4SLinus Torvalds } 15374bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_sync_mss); 15381da177e4SLinus Torvalds 15391da177e4SLinus Torvalds /* Compute the current effective MSS, taking SACKs and IP options, 15401da177e4SLinus Torvalds * and even PMTU discovery events into account. 15411da177e4SLinus Torvalds */ 15420c54b85fSIlpo Järvinen unsigned int tcp_current_mss(struct sock *sk) 15431da177e4SLinus Torvalds { 1544cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1545cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1546c1b4a7e6SDavid S. Miller u32 mss_now; 154795c96174SEric Dumazet unsigned int header_len; 154833ad798cSAdam Langley struct tcp_out_options opts; 154933ad798cSAdam Langley struct tcp_md5sig_key *md5; 15501da177e4SLinus Torvalds 1551c1b4a7e6SDavid S. Miller mss_now = tp->mss_cache; 1552c1b4a7e6SDavid S. Miller 15531da177e4SLinus Torvalds if (dst) { 15541da177e4SLinus Torvalds u32 mtu = dst_mtu(dst); 1555d83d8461SArnaldo Carvalho de Melo if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 15561da177e4SLinus Torvalds mss_now = tcp_sync_mss(sk, mtu); 15571da177e4SLinus Torvalds } 15581da177e4SLinus Torvalds 155933ad798cSAdam Langley header_len = tcp_established_options(sk, NULL, &opts, &md5) + 156033ad798cSAdam Langley sizeof(struct tcphdr); 156133ad798cSAdam Langley /* The mss_cache is sized based on tp->tcp_header_len, which assumes 156233ad798cSAdam Langley * some common options. If this is an odd packet (because we have SACK 156333ad798cSAdam Langley * blocks etc) then our calculated header_len will be different, and 156433ad798cSAdam Langley * we have to adjust mss_now correspondingly */ 156533ad798cSAdam Langley if (header_len != tp->tcp_header_len) { 156633ad798cSAdam Langley int delta = (int) header_len - tp->tcp_header_len; 156733ad798cSAdam Langley mss_now -= delta; 156833ad798cSAdam Langley } 1569cfb6eeb4SYOSHIFUJI Hideaki 15701da177e4SLinus Torvalds return mss_now; 15711da177e4SLinus Torvalds } 15721da177e4SLinus Torvalds 157386fd14adSWeiping Pan /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. 157486fd14adSWeiping Pan * As additional protections, we do not touch cwnd in retransmission phases, 157586fd14adSWeiping Pan * and if application hit its sndbuf limit recently. 157686fd14adSWeiping Pan */ 157786fd14adSWeiping Pan static void tcp_cwnd_application_limited(struct sock *sk) 1578a762a980SDavid S. Miller { 15799e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1580a762a980SDavid S. Miller 158186fd14adSWeiping Pan if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && 158286fd14adSWeiping Pan sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 158386fd14adSWeiping Pan /* Limited by application or receiver window. */ 158486fd14adSWeiping Pan u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); 158586fd14adSWeiping Pan u32 win_used = max(tp->snd_cwnd_used, init_win); 158686fd14adSWeiping Pan if (win_used < tp->snd_cwnd) { 158786fd14adSWeiping Pan tp->snd_ssthresh = tcp_current_ssthresh(sk); 158886fd14adSWeiping Pan tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1; 158986fd14adSWeiping Pan } 159086fd14adSWeiping Pan tp->snd_cwnd_used = 0; 159186fd14adSWeiping Pan } 1592c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 159386fd14adSWeiping Pan } 159486fd14adSWeiping Pan 1595ca8a2263SNeal Cardwell static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited) 1596a762a980SDavid S. Miller { 15971b1fc3fdSWei Wang const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 1598a762a980SDavid S. Miller struct tcp_sock *tp = tcp_sk(sk); 1599a762a980SDavid S. Miller 1600ca8a2263SNeal Cardwell /* Track the maximum number of outstanding packets in each 1601ca8a2263SNeal Cardwell * window, and remember whether we were cwnd-limited then. 1602ca8a2263SNeal Cardwell */ 1603ca8a2263SNeal Cardwell if (!before(tp->snd_una, tp->max_packets_seq) || 1604ca8a2263SNeal Cardwell tp->packets_out > tp->max_packets_out) { 1605ca8a2263SNeal Cardwell tp->max_packets_out = tp->packets_out; 1606ca8a2263SNeal Cardwell tp->max_packets_seq = tp->snd_nxt; 1607ca8a2263SNeal Cardwell tp->is_cwnd_limited = is_cwnd_limited; 1608ca8a2263SNeal Cardwell } 1609e114a710SEric Dumazet 161024901551SEric Dumazet if (tcp_is_cwnd_limited(sk)) { 1611a762a980SDavid S. Miller /* Network is feed fully. */ 1612a762a980SDavid S. Miller tp->snd_cwnd_used = 0; 1613c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 1614a762a980SDavid S. Miller } else { 1615a762a980SDavid S. Miller /* Network starves. */ 1616a762a980SDavid S. Miller if (tp->packets_out > tp->snd_cwnd_used) 1617a762a980SDavid S. Miller tp->snd_cwnd_used = tp->packets_out; 1618a762a980SDavid S. Miller 161915d33c07SDavid S. Miller if (sysctl_tcp_slow_start_after_idle && 1620c2203cf7SEric Dumazet (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto && 16211b1fc3fdSWei Wang !ca_ops->cong_control) 1622a762a980SDavid S. Miller tcp_cwnd_application_limited(sk); 1623b0f71bd3SFrancis Yan 1624b0f71bd3SFrancis Yan /* The following conditions together indicate the starvation 1625b0f71bd3SFrancis Yan * is caused by insufficient sender buffer: 1626b0f71bd3SFrancis Yan * 1) just sent some data (see tcp_write_xmit) 1627b0f71bd3SFrancis Yan * 2) not cwnd limited (this else condition) 1628b0f71bd3SFrancis Yan * 3) no more data to send (null tcp_send_head ) 1629b0f71bd3SFrancis Yan * 4) application is hitting buffer limit (SOCK_NOSPACE) 1630b0f71bd3SFrancis Yan */ 1631b0f71bd3SFrancis Yan if (!tcp_send_head(sk) && sk->sk_socket && 1632b0f71bd3SFrancis Yan test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) && 1633b0f71bd3SFrancis Yan (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) 1634b0f71bd3SFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED); 1635a762a980SDavid S. Miller } 1636a762a980SDavid S. Miller } 1637a762a980SDavid S. Miller 1638d4589926SEric Dumazet /* Minshall's variant of the Nagle send check. */ 1639d4589926SEric Dumazet static bool tcp_minshall_check(const struct tcp_sock *tp) 1640d4589926SEric Dumazet { 1641d4589926SEric Dumazet return after(tp->snd_sml, tp->snd_una) && 1642d4589926SEric Dumazet !after(tp->snd_sml, tp->snd_nxt); 1643d4589926SEric Dumazet } 1644d4589926SEric Dumazet 1645d4589926SEric Dumazet /* Update snd_sml if this skb is under mss 1646d4589926SEric Dumazet * Note that a TSO packet might end with a sub-mss segment 1647d4589926SEric Dumazet * The test is really : 1648d4589926SEric Dumazet * if ((skb->len % mss) != 0) 1649d4589926SEric Dumazet * tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1650d4589926SEric Dumazet * But we can avoid doing the divide again given we already have 1651d4589926SEric Dumazet * skb_pcount = skb->len / mss_now 16520e3a4803SIlpo Järvinen */ 1653d4589926SEric Dumazet static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now, 1654d4589926SEric Dumazet const struct sk_buff *skb) 1655d4589926SEric Dumazet { 1656d4589926SEric Dumazet if (skb->len < tcp_skb_pcount(skb) * mss_now) 1657d4589926SEric Dumazet tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1658d4589926SEric Dumazet } 1659d4589926SEric Dumazet 1660d4589926SEric Dumazet /* Return false, if packet can be sent now without violation Nagle's rules: 1661d4589926SEric Dumazet * 1. It is full sized. (provided by caller in %partial bool) 1662d4589926SEric Dumazet * 2. Or it contains FIN. (already checked by caller) 1663d4589926SEric Dumazet * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. 1664d4589926SEric Dumazet * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 1665d4589926SEric Dumazet * With Minshall's modification: all sent small packets are ACKed. 1666d4589926SEric Dumazet */ 1667d4589926SEric Dumazet static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp, 1668cc93fc51SPeter Pan(潘卫平) int nonagle) 1669d4589926SEric Dumazet { 1670d4589926SEric Dumazet return partial && 1671d4589926SEric Dumazet ((nonagle & TCP_NAGLE_CORK) || 1672d4589926SEric Dumazet (!nonagle && tp->packets_out && tcp_minshall_check(tp))); 1673d4589926SEric Dumazet } 1674605ad7f1SEric Dumazet 1675605ad7f1SEric Dumazet /* Return how many segs we'd like on a TSO packet, 1676605ad7f1SEric Dumazet * to send one TSO packet per ms 1677605ad7f1SEric Dumazet */ 16781b3878caSNeal Cardwell u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now, 16791b3878caSNeal Cardwell int min_tso_segs) 1680605ad7f1SEric Dumazet { 1681605ad7f1SEric Dumazet u32 bytes, segs; 1682605ad7f1SEric Dumazet 1683605ad7f1SEric Dumazet bytes = min(sk->sk_pacing_rate >> 10, 1684605ad7f1SEric Dumazet sk->sk_gso_max_size - 1 - MAX_TCP_HEADER); 1685605ad7f1SEric Dumazet 1686605ad7f1SEric Dumazet /* Goal is to send at least one packet per ms, 1687605ad7f1SEric Dumazet * not one big TSO packet every 100 ms. 1688605ad7f1SEric Dumazet * This preserves ACK clocking and is consistent 1689605ad7f1SEric Dumazet * with tcp_tso_should_defer() heuristic. 1690605ad7f1SEric Dumazet */ 16911b3878caSNeal Cardwell segs = max_t(u32, bytes / mss_now, min_tso_segs); 1692605ad7f1SEric Dumazet 1693605ad7f1SEric Dumazet return min_t(u32, segs, sk->sk_gso_max_segs); 1694605ad7f1SEric Dumazet } 16951b3878caSNeal Cardwell EXPORT_SYMBOL(tcp_tso_autosize); 1696605ad7f1SEric Dumazet 1697ed6e7268SNeal Cardwell /* Return the number of segments we want in the skb we are transmitting. 1698ed6e7268SNeal Cardwell * See if congestion control module wants to decide; otherwise, autosize. 1699ed6e7268SNeal Cardwell */ 1700ed6e7268SNeal Cardwell static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now) 1701ed6e7268SNeal Cardwell { 1702ed6e7268SNeal Cardwell const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 1703ed6e7268SNeal Cardwell u32 tso_segs = ca_ops->tso_segs_goal ? ca_ops->tso_segs_goal(sk) : 0; 1704ed6e7268SNeal Cardwell 17051b3878caSNeal Cardwell return tso_segs ? : 17061b3878caSNeal Cardwell tcp_tso_autosize(sk, mss_now, sysctl_tcp_min_tso_segs); 1707ed6e7268SNeal Cardwell } 1708ed6e7268SNeal Cardwell 1709d4589926SEric Dumazet /* Returns the portion of skb which can be sent right away */ 1710d4589926SEric Dumazet static unsigned int tcp_mss_split_point(const struct sock *sk, 1711d4589926SEric Dumazet const struct sk_buff *skb, 1712d4589926SEric Dumazet unsigned int mss_now, 1713d4589926SEric Dumazet unsigned int max_segs, 1714d4589926SEric Dumazet int nonagle) 1715c1b4a7e6SDavid S. Miller { 1716cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1717d4589926SEric Dumazet u32 partial, needed, window, max_len; 1718c1b4a7e6SDavid S. Miller 171990840defSIlpo Järvinen window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 17201485348dSBen Hutchings max_len = mss_now * max_segs; 17210e3a4803SIlpo Järvinen 17221485348dSBen Hutchings if (likely(max_len <= window && skb != tcp_write_queue_tail(sk))) 17231485348dSBen Hutchings return max_len; 17240e3a4803SIlpo Järvinen 17255ea3a748SIlpo Järvinen needed = min(skb->len, window); 17265ea3a748SIlpo Järvinen 17271485348dSBen Hutchings if (max_len <= needed) 17281485348dSBen Hutchings return max_len; 17290e3a4803SIlpo Järvinen 1730d4589926SEric Dumazet partial = needed % mss_now; 1731d4589926SEric Dumazet /* If last segment is not a full MSS, check if Nagle rules allow us 1732d4589926SEric Dumazet * to include this last segment in this skb. 1733d4589926SEric Dumazet * Otherwise, we'll split the skb at last MSS boundary 1734d4589926SEric Dumazet */ 1735cc93fc51SPeter Pan(潘卫平) if (tcp_nagle_check(partial != 0, tp, nonagle)) 1736d4589926SEric Dumazet return needed - partial; 1737d4589926SEric Dumazet 1738d4589926SEric Dumazet return needed; 1739c1b4a7e6SDavid S. Miller } 1740c1b4a7e6SDavid S. Miller 1741c1b4a7e6SDavid S. Miller /* Can at least one segment of SKB be sent right now, according to the 1742c1b4a7e6SDavid S. Miller * congestion window rules? If so, return how many segments are allowed. 1743c1b4a7e6SDavid S. Miller */ 1744cf533ea5SEric Dumazet static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, 1745cf533ea5SEric Dumazet const struct sk_buff *skb) 1746c1b4a7e6SDavid S. Miller { 1747d649a7a8SEric Dumazet u32 in_flight, cwnd, halfcwnd; 1748c1b4a7e6SDavid S. Miller 1749c1b4a7e6SDavid S. Miller /* Don't be strict about the congestion window for the final FIN. */ 17504de075e0SEric Dumazet if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && 17514de075e0SEric Dumazet tcp_skb_pcount(skb) == 1) 1752c1b4a7e6SDavid S. Miller return 1; 1753c1b4a7e6SDavid S. Miller 1754c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1755c1b4a7e6SDavid S. Miller cwnd = tp->snd_cwnd; 1756d649a7a8SEric Dumazet if (in_flight >= cwnd) 1757c1b4a7e6SDavid S. Miller return 0; 1758d649a7a8SEric Dumazet 1759d649a7a8SEric Dumazet /* For better scheduling, ensure we have at least 1760d649a7a8SEric Dumazet * 2 GSO packets in flight. 1761d649a7a8SEric Dumazet */ 1762d649a7a8SEric Dumazet halfcwnd = max(cwnd >> 1, 1U); 1763d649a7a8SEric Dumazet return min(halfcwnd, cwnd - in_flight); 1764c1b4a7e6SDavid S. Miller } 1765c1b4a7e6SDavid S. Miller 1766b595076aSUwe Kleine-König /* Initialize TSO state of a skb. 176767edfef7SAndi Kleen * This must be invoked the first time we consider transmitting 1768c1b4a7e6SDavid S. Miller * SKB onto the wire. 1769c1b4a7e6SDavid S. Miller */ 17705bbb432cSEric Dumazet static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now) 1771c1b4a7e6SDavid S. Miller { 1772c1b4a7e6SDavid S. Miller int tso_segs = tcp_skb_pcount(skb); 1773c1b4a7e6SDavid S. Miller 1774f8269a49SIlpo Järvinen if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { 17755bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 1776c1b4a7e6SDavid S. Miller tso_segs = tcp_skb_pcount(skb); 1777c1b4a7e6SDavid S. Miller } 1778c1b4a7e6SDavid S. Miller return tso_segs; 1779c1b4a7e6SDavid S. Miller } 1780c1b4a7e6SDavid S. Miller 1781c1b4a7e6SDavid S. Miller 1782a2a385d6SEric Dumazet /* Return true if the Nagle test allows this packet to be 1783c1b4a7e6SDavid S. Miller * sent now. 1784c1b4a7e6SDavid S. Miller */ 1785a2a385d6SEric Dumazet static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, 1786c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1787c1b4a7e6SDavid S. Miller { 1788c1b4a7e6SDavid S. Miller /* Nagle rule does not apply to frames, which sit in the middle of the 1789c1b4a7e6SDavid S. Miller * write_queue (they have no chances to get new data). 1790c1b4a7e6SDavid S. Miller * 1791c1b4a7e6SDavid S. Miller * This is implemented in the callers, where they modify the 'nonagle' 1792c1b4a7e6SDavid S. Miller * argument based upon the location of SKB in the send queue. 1793c1b4a7e6SDavid S. Miller */ 1794c1b4a7e6SDavid S. Miller if (nonagle & TCP_NAGLE_PUSH) 1795a2a385d6SEric Dumazet return true; 1796c1b4a7e6SDavid S. Miller 17979b44190dSYuchung Cheng /* Don't use the nagle rule for urgent data (or for the final FIN). */ 17989b44190dSYuchung Cheng if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) 1799a2a385d6SEric Dumazet return true; 1800c1b4a7e6SDavid S. Miller 1801cc93fc51SPeter Pan(潘卫平) if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle)) 1802a2a385d6SEric Dumazet return true; 1803c1b4a7e6SDavid S. Miller 1804a2a385d6SEric Dumazet return false; 1805c1b4a7e6SDavid S. Miller } 1806c1b4a7e6SDavid S. Miller 1807c1b4a7e6SDavid S. Miller /* Does at least the first segment of SKB fit into the send window? */ 1808a2a385d6SEric Dumazet static bool tcp_snd_wnd_test(const struct tcp_sock *tp, 1809a2a385d6SEric Dumazet const struct sk_buff *skb, 1810056834d9SIlpo Järvinen unsigned int cur_mss) 1811c1b4a7e6SDavid S. Miller { 1812c1b4a7e6SDavid S. Miller u32 end_seq = TCP_SKB_CB(skb)->end_seq; 1813c1b4a7e6SDavid S. Miller 1814c1b4a7e6SDavid S. Miller if (skb->len > cur_mss) 1815c1b4a7e6SDavid S. Miller end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 1816c1b4a7e6SDavid S. Miller 181790840defSIlpo Järvinen return !after(end_seq, tcp_wnd_end(tp)); 1818c1b4a7e6SDavid S. Miller } 1819c1b4a7e6SDavid S. Miller 1820c1b4a7e6SDavid S. Miller /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 1821c1b4a7e6SDavid S. Miller * which is put after SKB on the list. It is very much like 1822c1b4a7e6SDavid S. Miller * tcp_fragment() except that it may make several kinds of assumptions 1823c1b4a7e6SDavid S. Miller * in order to speed up the splitting operation. In particular, we 1824c1b4a7e6SDavid S. Miller * know that all the data is in scatter-gather pages, and that the 1825c1b4a7e6SDavid S. Miller * packet has never been sent out before (and thus is not cloned). 1826c1b4a7e6SDavid S. Miller */ 1827056834d9SIlpo Järvinen static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, 1828c4ead4c5SEric Dumazet unsigned int mss_now, gfp_t gfp) 1829c1b4a7e6SDavid S. Miller { 1830c1b4a7e6SDavid S. Miller struct sk_buff *buff; 1831c1b4a7e6SDavid S. Miller int nlen = skb->len - len; 18329ce01461SIlpo Järvinen u8 flags; 1833c1b4a7e6SDavid S. Miller 1834c1b4a7e6SDavid S. Miller /* All of a TSO frame must be composed of paged data. */ 1835c8ac3774SHerbert Xu if (skb->len != skb->data_len) 18366cc55e09SOctavian Purdila return tcp_fragment(sk, skb, len, mss_now, gfp); 1837c1b4a7e6SDavid S. Miller 1838eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, gfp, true); 183951456b29SIan Morris if (unlikely(!buff)) 1840c1b4a7e6SDavid S. Miller return -ENOMEM; 1841c1b4a7e6SDavid S. Miller 18423ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 18433ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1844b60b49eaSHerbert Xu buff->truesize += nlen; 1845c1b4a7e6SDavid S. Miller skb->truesize -= nlen; 1846c1b4a7e6SDavid S. Miller 1847c1b4a7e6SDavid S. Miller /* Correct the sequence numbers. */ 1848c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 1849c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 1850c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 1851c1b4a7e6SDavid S. Miller 1852c1b4a7e6SDavid S. Miller /* PSH and FIN should only be set in the second packet. */ 18534de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 18544de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 18554de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1856c1b4a7e6SDavid S. Miller 1857c1b4a7e6SDavid S. Miller /* This packet was never sent out yet, so no SACK bits. */ 1858c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->sacked = 0; 1859c1b4a7e6SDavid S. Miller 1860a166140eSMartin KaFai Lau tcp_skb_fragment_eor(skb, buff); 1861a166140eSMartin KaFai Lau 186284fa7933SPatrick McHardy buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; 1863c1b4a7e6SDavid S. Miller skb_split(skb, buff, len); 1864490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 1865c1b4a7e6SDavid S. Miller 1866c1b4a7e6SDavid S. Miller /* Fix up tso_factor for both original and new SKB. */ 18675bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 18685bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 1869c1b4a7e6SDavid S. Miller 1870c1b4a7e6SDavid S. Miller /* Link BUFF into the send queue. */ 1871f4a775d1SEric Dumazet __skb_header_release(buff); 1872fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 1873c1b4a7e6SDavid S. Miller 1874c1b4a7e6SDavid S. Miller return 0; 1875c1b4a7e6SDavid S. Miller } 1876c1b4a7e6SDavid S. Miller 1877c1b4a7e6SDavid S. Miller /* Try to defer sending, if possible, in order to minimize the amount 1878c1b4a7e6SDavid S. Miller * of TSO splitting we do. View it as a kind of TSO Nagle test. 1879c1b4a7e6SDavid S. Miller * 1880c1b4a7e6SDavid S. Miller * This algorithm is from John Heffner. 1881c1b4a7e6SDavid S. Miller */ 1882ca8a2263SNeal Cardwell static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb, 1883605ad7f1SEric Dumazet bool *is_cwnd_limited, u32 max_segs) 1884c1b4a7e6SDavid S. Miller { 18856687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 188650c8339eSEric Dumazet u32 age, send_win, cong_win, limit, in_flight; 188750c8339eSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 188850c8339eSEric Dumazet struct sk_buff *head; 1889ad9f4f50SEric Dumazet int win_divisor; 1890c1b4a7e6SDavid S. Miller 18914de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 1892ae8064acSJohn Heffner goto send_now; 1893c1b4a7e6SDavid S. Miller 189499d7662aSEric Dumazet if (icsk->icsk_ca_state >= TCP_CA_Recovery) 1895ae8064acSJohn Heffner goto send_now; 1896ae8064acSJohn Heffner 18975f852eb5SEric Dumazet /* Avoid bursty behavior by allowing defer 18985f852eb5SEric Dumazet * only if the last write was recent. 18995f852eb5SEric Dumazet */ 1900d635fbe2SEric Dumazet if ((s32)(tcp_jiffies32 - tp->lsndtime) > 0) 1901ae8064acSJohn Heffner goto send_now; 1902908a75c1SDavid S. Miller 1903c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1904c1b4a7e6SDavid S. Miller 1905056834d9SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight)); 1906c1b4a7e6SDavid S. Miller 190790840defSIlpo Järvinen send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1908c1b4a7e6SDavid S. Miller 1909c1b4a7e6SDavid S. Miller /* From in_flight test above, we know that cwnd > in_flight. */ 1910c1b4a7e6SDavid S. Miller cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; 1911c1b4a7e6SDavid S. Miller 1912c1b4a7e6SDavid S. Miller limit = min(send_win, cong_win); 1913c1b4a7e6SDavid S. Miller 1914ba244fe9SDavid S. Miller /* If a full-sized TSO skb can be sent, do it. */ 1915605ad7f1SEric Dumazet if (limit >= max_segs * tp->mss_cache) 1916ae8064acSJohn Heffner goto send_now; 1917ba244fe9SDavid S. Miller 191862ad2761SIlpo Järvinen /* Middle in queue won't get any more data, full sendable already? */ 191962ad2761SIlpo Järvinen if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) 192062ad2761SIlpo Järvinen goto send_now; 192162ad2761SIlpo Järvinen 1922ad9f4f50SEric Dumazet win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor); 1923ad9f4f50SEric Dumazet if (win_divisor) { 1924c1b4a7e6SDavid S. Miller u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); 1925c1b4a7e6SDavid S. Miller 1926c1b4a7e6SDavid S. Miller /* If at least some fraction of a window is available, 1927c1b4a7e6SDavid S. Miller * just use it. 1928c1b4a7e6SDavid S. Miller */ 1929ad9f4f50SEric Dumazet chunk /= win_divisor; 1930c1b4a7e6SDavid S. Miller if (limit >= chunk) 1931ae8064acSJohn Heffner goto send_now; 1932c1b4a7e6SDavid S. Miller } else { 1933c1b4a7e6SDavid S. Miller /* Different approach, try not to defer past a single 1934c1b4a7e6SDavid S. Miller * ACK. Receiver should ACK every other full sized 1935c1b4a7e6SDavid S. Miller * frame, so if we have space for more than 3 frames 1936c1b4a7e6SDavid S. Miller * then send now. 1937c1b4a7e6SDavid S. Miller */ 19386b5a5c0dSNeal Cardwell if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) 1939ae8064acSJohn Heffner goto send_now; 1940c1b4a7e6SDavid S. Miller } 1941c1b4a7e6SDavid S. Miller 194250c8339eSEric Dumazet head = tcp_write_queue_head(sk); 1943385e2070SEric Dumazet 19449a568de4SEric Dumazet age = tcp_stamp_us_delta(tp->tcp_mstamp, head->skb_mstamp); 194550c8339eSEric Dumazet /* If next ACK is likely to come too late (half srtt), do not defer */ 194650c8339eSEric Dumazet if (age < (tp->srtt_us >> 4)) 194750c8339eSEric Dumazet goto send_now; 194850c8339eSEric Dumazet 19495f852eb5SEric Dumazet /* Ok, it looks like it is advisable to defer. */ 1950ae8064acSJohn Heffner 1951d2e1339fSBendik Rønning Opstad if (cong_win < send_win && cong_win <= skb->len) 1952ca8a2263SNeal Cardwell *is_cwnd_limited = true; 1953ca8a2263SNeal Cardwell 1954a2a385d6SEric Dumazet return true; 1955ae8064acSJohn Heffner 1956ae8064acSJohn Heffner send_now: 1957a2a385d6SEric Dumazet return false; 1958c1b4a7e6SDavid S. Miller } 1959c1b4a7e6SDavid S. Miller 196005cbc0dbSFan Du static inline void tcp_mtu_check_reprobe(struct sock *sk) 196105cbc0dbSFan Du { 196205cbc0dbSFan Du struct inet_connection_sock *icsk = inet_csk(sk); 196305cbc0dbSFan Du struct tcp_sock *tp = tcp_sk(sk); 196405cbc0dbSFan Du struct net *net = sock_net(sk); 196505cbc0dbSFan Du u32 interval; 196605cbc0dbSFan Du s32 delta; 196705cbc0dbSFan Du 196805cbc0dbSFan Du interval = net->ipv4.sysctl_tcp_probe_interval; 1969c74df29aSEric Dumazet delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp; 197005cbc0dbSFan Du if (unlikely(delta >= interval * HZ)) { 197105cbc0dbSFan Du int mss = tcp_current_mss(sk); 197205cbc0dbSFan Du 197305cbc0dbSFan Du /* Update current search range */ 197405cbc0dbSFan Du icsk->icsk_mtup.probe_size = 0; 197505cbc0dbSFan Du icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + 197605cbc0dbSFan Du sizeof(struct tcphdr) + 197705cbc0dbSFan Du icsk->icsk_af_ops->net_header_len; 197805cbc0dbSFan Du icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss); 197905cbc0dbSFan Du 198005cbc0dbSFan Du /* Update probe time stamp */ 1981c74df29aSEric Dumazet icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; 198205cbc0dbSFan Du } 198305cbc0dbSFan Du } 198405cbc0dbSFan Du 19855d424d5aSJohn Heffner /* Create a new MTU probe if we are ready. 198667edfef7SAndi Kleen * MTU probe is regularly attempting to increase the path MTU by 198767edfef7SAndi Kleen * deliberately sending larger packets. This discovers routing 198867edfef7SAndi Kleen * changes resulting in larger path MTUs. 198967edfef7SAndi Kleen * 19905d424d5aSJohn Heffner * Returns 0 if we should wait to probe (no cwnd available), 19915d424d5aSJohn Heffner * 1 if a probe was sent, 1992056834d9SIlpo Järvinen * -1 otherwise 1993056834d9SIlpo Järvinen */ 19945d424d5aSJohn Heffner static int tcp_mtu_probe(struct sock *sk) 19955d424d5aSJohn Heffner { 19965d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 199712a59abcSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 19985d424d5aSJohn Heffner struct sk_buff *skb, *nskb, *next; 19996b58e0a5SFan Du struct net *net = sock_net(sk); 20005d424d5aSJohn Heffner int probe_size; 200191cc17c0SIlpo Järvinen int size_needed; 200212a59abcSEric Dumazet int copy, len; 20035d424d5aSJohn Heffner int mss_now; 20046b58e0a5SFan Du int interval; 20055d424d5aSJohn Heffner 20065d424d5aSJohn Heffner /* Not currently probing/verifying, 20075d424d5aSJohn Heffner * not in recovery, 20085d424d5aSJohn Heffner * have enough cwnd, and 200912a59abcSEric Dumazet * not SACKing (the variable headers throw things off) 201012a59abcSEric Dumazet */ 201112a59abcSEric Dumazet if (likely(!icsk->icsk_mtup.enabled || 20125d424d5aSJohn Heffner icsk->icsk_mtup.probe_size || 20135d424d5aSJohn Heffner inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 20145d424d5aSJohn Heffner tp->snd_cwnd < 11 || 201512a59abcSEric Dumazet tp->rx_opt.num_sacks || tp->rx_opt.dsack)) 20165d424d5aSJohn Heffner return -1; 20175d424d5aSJohn Heffner 20186b58e0a5SFan Du /* Use binary search for probe_size between tcp_mss_base, 20196b58e0a5SFan Du * and current mss_clamp. if (search_high - search_low) 20206b58e0a5SFan Du * smaller than a threshold, backoff from probing. 20216b58e0a5SFan Du */ 20220c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 20236b58e0a5SFan Du probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high + 20246b58e0a5SFan Du icsk->icsk_mtup.search_low) >> 1); 202591cc17c0SIlpo Järvinen size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; 20266b58e0a5SFan Du interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low; 202705cbc0dbSFan Du /* When misfortune happens, we are reprobing actively, 202805cbc0dbSFan Du * and then reprobe timer has expired. We stick with current 202905cbc0dbSFan Du * probing process by not resetting search range to its orignal. 203005cbc0dbSFan Du */ 20316b58e0a5SFan Du if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) || 203205cbc0dbSFan Du interval < net->ipv4.sysctl_tcp_probe_threshold) { 203305cbc0dbSFan Du /* Check whether enough time has elaplased for 203405cbc0dbSFan Du * another round of probing. 203505cbc0dbSFan Du */ 203605cbc0dbSFan Du tcp_mtu_check_reprobe(sk); 20375d424d5aSJohn Heffner return -1; 20385d424d5aSJohn Heffner } 20395d424d5aSJohn Heffner 20405d424d5aSJohn Heffner /* Have enough data in the send queue to probe? */ 20417f9c33e5SIlpo Järvinen if (tp->write_seq - tp->snd_nxt < size_needed) 20425d424d5aSJohn Heffner return -1; 20435d424d5aSJohn Heffner 204491cc17c0SIlpo Järvinen if (tp->snd_wnd < size_needed) 20455d424d5aSJohn Heffner return -1; 204690840defSIlpo Järvinen if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) 20475d424d5aSJohn Heffner return 0; 20485d424d5aSJohn Heffner 2049d67c58e9SIlpo Järvinen /* Do we need to wait to drain cwnd? With none in flight, don't stall */ 2050d67c58e9SIlpo Järvinen if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { 2051d67c58e9SIlpo Järvinen if (!tcp_packets_in_flight(tp)) 20525d424d5aSJohn Heffner return -1; 20535d424d5aSJohn Heffner else 20545d424d5aSJohn Heffner return 0; 20555d424d5aSJohn Heffner } 20565d424d5aSJohn Heffner 20575d424d5aSJohn Heffner /* We're allowed to probe. Build it now. */ 2058eb934478SEric Dumazet nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false); 205951456b29SIan Morris if (!nskb) 20605d424d5aSJohn Heffner return -1; 20613ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 20623ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 20635d424d5aSJohn Heffner 2064fe067e8aSDavid S. Miller skb = tcp_send_head(sk); 20655d424d5aSJohn Heffner 20665d424d5aSJohn Heffner TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 20675d424d5aSJohn Heffner TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 20684de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; 20695d424d5aSJohn Heffner TCP_SKB_CB(nskb)->sacked = 0; 20705d424d5aSJohn Heffner nskb->csum = 0; 207184fa7933SPatrick McHardy nskb->ip_summed = skb->ip_summed; 20725d424d5aSJohn Heffner 207350c4817eSIlpo Järvinen tcp_insert_write_queue_before(nskb, skb, sk); 207450c4817eSIlpo Järvinen 20755d424d5aSJohn Heffner len = 0; 2076234b6860SIlpo Järvinen tcp_for_write_queue_from_safe(skb, next, sk) { 20775d424d5aSJohn Heffner copy = min_t(int, skb->len, probe_size - len); 20782fe664f1SDouglas Caetano dos Santos if (nskb->ip_summed) { 20795d424d5aSJohn Heffner skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); 20802fe664f1SDouglas Caetano dos Santos } else { 20812fe664f1SDouglas Caetano dos Santos __wsum csum = skb_copy_and_csum_bits(skb, 0, 2082056834d9SIlpo Järvinen skb_put(nskb, copy), 20832fe664f1SDouglas Caetano dos Santos copy, 0); 20842fe664f1SDouglas Caetano dos Santos nskb->csum = csum_block_add(nskb->csum, csum, len); 20852fe664f1SDouglas Caetano dos Santos } 20865d424d5aSJohn Heffner 20875d424d5aSJohn Heffner if (skb->len <= copy) { 20885d424d5aSJohn Heffner /* We've eaten all the data from this skb. 20895d424d5aSJohn Heffner * Throw it away. */ 20904de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 2091fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 20923ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 20935d424d5aSJohn Heffner } else { 20944de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & 2095a3433f35SChangli Gao ~(TCPHDR_FIN|TCPHDR_PSH); 20965d424d5aSJohn Heffner if (!skb_shinfo(skb)->nr_frags) { 20975d424d5aSJohn Heffner skb_pull(skb, copy); 209884fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 2099056834d9SIlpo Järvinen skb->csum = csum_partial(skb->data, 2100056834d9SIlpo Järvinen skb->len, 0); 21015d424d5aSJohn Heffner } else { 21025d424d5aSJohn Heffner __pskb_trim_head(skb, copy); 21035bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 21045d424d5aSJohn Heffner } 21055d424d5aSJohn Heffner TCP_SKB_CB(skb)->seq += copy; 21065d424d5aSJohn Heffner } 21075d424d5aSJohn Heffner 21085d424d5aSJohn Heffner len += copy; 2109234b6860SIlpo Järvinen 2110234b6860SIlpo Järvinen if (len >= probe_size) 2111234b6860SIlpo Järvinen break; 21125d424d5aSJohn Heffner } 21135bbb432cSEric Dumazet tcp_init_tso_segs(nskb, nskb->len); 21145d424d5aSJohn Heffner 21155d424d5aSJohn Heffner /* We're ready to send. If this fails, the probe will 21167faee5c0SEric Dumazet * be resegmented into mss-sized pieces by tcp_write_xmit(). 21177faee5c0SEric Dumazet */ 21185d424d5aSJohn Heffner if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 21195d424d5aSJohn Heffner /* Decrement cwnd here because we are sending 21205d424d5aSJohn Heffner * effectively two packets. */ 21215d424d5aSJohn Heffner tp->snd_cwnd--; 212266f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, nskb); 21235d424d5aSJohn Heffner 21245d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 21250e7b1368SJohn Heffner tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 21260e7b1368SJohn Heffner tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 21275d424d5aSJohn Heffner 21285d424d5aSJohn Heffner return 1; 21295d424d5aSJohn Heffner } 21305d424d5aSJohn Heffner 21315d424d5aSJohn Heffner return -1; 21325d424d5aSJohn Heffner } 21335d424d5aSJohn Heffner 2134218af599SEric Dumazet static bool tcp_pacing_check(const struct sock *sk) 2135218af599SEric Dumazet { 2136218af599SEric Dumazet return tcp_needs_internal_pacing(sk) && 2137218af599SEric Dumazet hrtimer_active(&tcp_sk(sk)->pacing_timer); 2138218af599SEric Dumazet } 2139218af599SEric Dumazet 2140f9616c35SEric Dumazet /* TCP Small Queues : 2141f9616c35SEric Dumazet * Control number of packets in qdisc/devices to two packets / or ~1 ms. 2142f9616c35SEric Dumazet * (These limits are doubled for retransmits) 2143f9616c35SEric Dumazet * This allows for : 2144f9616c35SEric Dumazet * - better RTT estimation and ACK scheduling 2145f9616c35SEric Dumazet * - faster recovery 2146f9616c35SEric Dumazet * - high rates 2147f9616c35SEric Dumazet * Alas, some drivers / subsystems require a fair amount 2148f9616c35SEric Dumazet * of queued bytes to ensure line rate. 2149f9616c35SEric Dumazet * One example is wifi aggregation (802.11 AMPDU) 2150f9616c35SEric Dumazet */ 2151f9616c35SEric Dumazet static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb, 2152f9616c35SEric Dumazet unsigned int factor) 2153f9616c35SEric Dumazet { 2154f9616c35SEric Dumazet unsigned int limit; 2155f9616c35SEric Dumazet 2156f9616c35SEric Dumazet limit = max(2 * skb->truesize, sk->sk_pacing_rate >> 10); 2157f9616c35SEric Dumazet limit = min_t(u32, limit, sysctl_tcp_limit_output_bytes); 2158f9616c35SEric Dumazet limit <<= factor; 2159f9616c35SEric Dumazet 216014afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) > limit) { 216175eefc6cSEric Dumazet /* Always send the 1st or 2nd skb in write queue. 216275eefc6cSEric Dumazet * No need to wait for TX completion to call us back, 216375eefc6cSEric Dumazet * after softirq/tasklet schedule. 216475eefc6cSEric Dumazet * This helps when TX completions are delayed too much. 216575eefc6cSEric Dumazet */ 216675eefc6cSEric Dumazet if (skb == sk->sk_write_queue.next || 216775eefc6cSEric Dumazet skb->prev == sk->sk_write_queue.next) 216875eefc6cSEric Dumazet return false; 216975eefc6cSEric Dumazet 21707aa5470cSEric Dumazet set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags); 2171f9616c35SEric Dumazet /* It is possible TX completion already happened 2172f9616c35SEric Dumazet * before we set TSQ_THROTTLED, so we must 2173f9616c35SEric Dumazet * test again the condition. 2174f9616c35SEric Dumazet */ 2175f9616c35SEric Dumazet smp_mb__after_atomic(); 217614afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) > limit) 2177f9616c35SEric Dumazet return true; 2178f9616c35SEric Dumazet } 2179f9616c35SEric Dumazet return false; 2180f9616c35SEric Dumazet } 2181f9616c35SEric Dumazet 218205b055e8SFrancis Yan static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new) 218305b055e8SFrancis Yan { 2184628174ccSEric Dumazet const u32 now = tcp_jiffies32; 2185efe967cdSArnd Bergmann enum tcp_chrono old = tp->chrono_type; 218605b055e8SFrancis Yan 2187efe967cdSArnd Bergmann if (old > TCP_CHRONO_UNSPEC) 2188efe967cdSArnd Bergmann tp->chrono_stat[old - 1] += now - tp->chrono_start; 218905b055e8SFrancis Yan tp->chrono_start = now; 219005b055e8SFrancis Yan tp->chrono_type = new; 219105b055e8SFrancis Yan } 219205b055e8SFrancis Yan 219305b055e8SFrancis Yan void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type) 219405b055e8SFrancis Yan { 219505b055e8SFrancis Yan struct tcp_sock *tp = tcp_sk(sk); 219605b055e8SFrancis Yan 219705b055e8SFrancis Yan /* If there are multiple conditions worthy of tracking in a 21980f87230dSFrancis Yan * chronograph then the highest priority enum takes precedence 21990f87230dSFrancis Yan * over the other conditions. So that if something "more interesting" 220005b055e8SFrancis Yan * starts happening, stop the previous chrono and start a new one. 220105b055e8SFrancis Yan */ 220205b055e8SFrancis Yan if (type > tp->chrono_type) 220305b055e8SFrancis Yan tcp_chrono_set(tp, type); 220405b055e8SFrancis Yan } 220505b055e8SFrancis Yan 220605b055e8SFrancis Yan void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type) 220705b055e8SFrancis Yan { 220805b055e8SFrancis Yan struct tcp_sock *tp = tcp_sk(sk); 220905b055e8SFrancis Yan 22100f87230dSFrancis Yan 22110f87230dSFrancis Yan /* There are multiple conditions worthy of tracking in a 22120f87230dSFrancis Yan * chronograph, so that the highest priority enum takes 22130f87230dSFrancis Yan * precedence over the other conditions (see tcp_chrono_start). 22140f87230dSFrancis Yan * If a condition stops, we only stop chrono tracking if 22150f87230dSFrancis Yan * it's the "most interesting" or current chrono we are 22160f87230dSFrancis Yan * tracking and starts busy chrono if we have pending data. 22170f87230dSFrancis Yan */ 22180f87230dSFrancis Yan if (tcp_write_queue_empty(sk)) 221905b055e8SFrancis Yan tcp_chrono_set(tp, TCP_CHRONO_UNSPEC); 22200f87230dSFrancis Yan else if (type == tp->chrono_type) 22210f87230dSFrancis Yan tcp_chrono_set(tp, TCP_CHRONO_BUSY); 222205b055e8SFrancis Yan } 222305b055e8SFrancis Yan 22241da177e4SLinus Torvalds /* This routine writes packets to the network. It advances the 22251da177e4SLinus Torvalds * send_head. This happens as incoming acks open up the remote 22261da177e4SLinus Torvalds * window for us. 22271da177e4SLinus Torvalds * 2228f8269a49SIlpo Järvinen * LARGESEND note: !tcp_urg_mode is overkill, only frames between 2229f8269a49SIlpo Järvinen * snd_up-64k-mss .. snd_up cannot be large. However, taking into 2230f8269a49SIlpo Järvinen * account rare use of URG, this is not a big flaw. 2231f8269a49SIlpo Järvinen * 22326ba8a3b1SNandita Dukkipati * Send at most one packet when push_one > 0. Temporarily ignore 22336ba8a3b1SNandita Dukkipati * cwnd limit to force at most one packet out when push_one == 2. 22346ba8a3b1SNandita Dukkipati 2235a2a385d6SEric Dumazet * Returns true, if no segments are in flight and we have queued segments, 2236a2a385d6SEric Dumazet * but cannot send anything now because of SWS or another problem. 22371da177e4SLinus Torvalds */ 2238a2a385d6SEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 2239d5dd9175SIlpo Järvinen int push_one, gfp_t gfp) 22401da177e4SLinus Torvalds { 22411da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 224292df7b51SDavid S. Miller struct sk_buff *skb; 2243c1b4a7e6SDavid S. Miller unsigned int tso_segs, sent_pkts; 2244c1b4a7e6SDavid S. Miller int cwnd_quota; 22455d424d5aSJohn Heffner int result; 22465615f886SFrancis Yan bool is_cwnd_limited = false, is_rwnd_limited = false; 2247605ad7f1SEric Dumazet u32 max_segs; 22481da177e4SLinus Torvalds 2249c1b4a7e6SDavid S. Miller sent_pkts = 0; 22505d424d5aSJohn Heffner 2251d5dd9175SIlpo Järvinen if (!push_one) { 22525d424d5aSJohn Heffner /* Do MTU probing. */ 2253d5dd9175SIlpo Järvinen result = tcp_mtu_probe(sk); 2254d5dd9175SIlpo Järvinen if (!result) { 2255a2a385d6SEric Dumazet return false; 22565d424d5aSJohn Heffner } else if (result > 0) { 22575d424d5aSJohn Heffner sent_pkts = 1; 22585d424d5aSJohn Heffner } 2259d5dd9175SIlpo Järvinen } 22605d424d5aSJohn Heffner 2261ed6e7268SNeal Cardwell max_segs = tcp_tso_segs(sk, mss_now); 22629a568de4SEric Dumazet tcp_mstamp_refresh(tp); 2263fe067e8aSDavid S. Miller while ((skb = tcp_send_head(sk))) { 2264c8ac3774SHerbert Xu unsigned int limit; 2265c8ac3774SHerbert Xu 2266218af599SEric Dumazet if (tcp_pacing_check(sk)) 2267218af599SEric Dumazet break; 2268218af599SEric Dumazet 22695bbb432cSEric Dumazet tso_segs = tcp_init_tso_segs(skb, mss_now); 2270c1b4a7e6SDavid S. Miller BUG_ON(!tso_segs); 2271c1b4a7e6SDavid S. Miller 22729d186cacSAndrey Vagin if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) { 22737faee5c0SEric Dumazet /* "skb_mstamp" is used as a start point for the retransmit timer */ 2274*e2080072SEric Dumazet tcp_update_skb_after_send(tp, skb); 2275ec342325SAndrew Vagin goto repair; /* Skip network transmission */ 22769d186cacSAndrey Vagin } 2277ec342325SAndrew Vagin 2278b68e9f85SHerbert Xu cwnd_quota = tcp_cwnd_test(tp, skb); 22796ba8a3b1SNandita Dukkipati if (!cwnd_quota) { 22806ba8a3b1SNandita Dukkipati if (push_one == 2) 22816ba8a3b1SNandita Dukkipati /* Force out a loss probe pkt. */ 22826ba8a3b1SNandita Dukkipati cwnd_quota = 1; 22836ba8a3b1SNandita Dukkipati else 2284b68e9f85SHerbert Xu break; 22856ba8a3b1SNandita Dukkipati } 2286b68e9f85SHerbert Xu 22875615f886SFrancis Yan if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) { 22885615f886SFrancis Yan is_rwnd_limited = true; 2289b68e9f85SHerbert Xu break; 22905615f886SFrancis Yan } 2291b68e9f85SHerbert Xu 2292d6a4e26aSEric Dumazet if (tso_segs == 1) { 2293aa93466bSDavid S. Miller if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 2294aa93466bSDavid S. Miller (tcp_skb_is_last(sk, skb) ? 2295aa93466bSDavid S. Miller nonagle : TCP_NAGLE_PUSH)))) 2296aa93466bSDavid S. Miller break; 2297c1b4a7e6SDavid S. Miller } else { 2298ca8a2263SNeal Cardwell if (!push_one && 2299605ad7f1SEric Dumazet tcp_tso_should_defer(sk, skb, &is_cwnd_limited, 2300605ad7f1SEric Dumazet max_segs)) 2301aa93466bSDavid S. Miller break; 2302c1b4a7e6SDavid S. Miller } 2303aa93466bSDavid S. Miller 2304605ad7f1SEric Dumazet limit = mss_now; 2305d6a4e26aSEric Dumazet if (tso_segs > 1 && !tcp_urg_mode(tp)) 2306605ad7f1SEric Dumazet limit = tcp_mss_split_point(sk, skb, mss_now, 2307605ad7f1SEric Dumazet min_t(unsigned int, 2308605ad7f1SEric Dumazet cwnd_quota, 2309605ad7f1SEric Dumazet max_segs), 2310605ad7f1SEric Dumazet nonagle); 2311605ad7f1SEric Dumazet 2312605ad7f1SEric Dumazet if (skb->len > limit && 2313605ad7f1SEric Dumazet unlikely(tso_fragment(sk, skb, limit, mss_now, gfp))) 2314605ad7f1SEric Dumazet break; 2315605ad7f1SEric Dumazet 23167aa5470cSEric Dumazet if (test_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags)) 23177aa5470cSEric Dumazet clear_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags); 2318f9616c35SEric Dumazet if (tcp_small_queue_check(sk, skb, 0)) 231946d3ceabSEric Dumazet break; 2320c9eeec26SEric Dumazet 2321d5dd9175SIlpo Järvinen if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) 23221da177e4SLinus Torvalds break; 23231da177e4SLinus Torvalds 2324ec342325SAndrew Vagin repair: 23251da177e4SLinus Torvalds /* Advance the send_head. This one is sent out. 23261da177e4SLinus Torvalds * This call will increment packets_out. 23271da177e4SLinus Torvalds */ 232866f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 23291da177e4SLinus Torvalds 23301da177e4SLinus Torvalds tcp_minshall_update(tp, mss_now, skb); 2331a262f0cdSNandita Dukkipati sent_pkts += tcp_skb_pcount(skb); 2332d5dd9175SIlpo Järvinen 2333d5dd9175SIlpo Järvinen if (push_one) 2334d5dd9175SIlpo Järvinen break; 23351da177e4SLinus Torvalds } 23361da177e4SLinus Torvalds 23375615f886SFrancis Yan if (is_rwnd_limited) 23385615f886SFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED); 23395615f886SFrancis Yan else 23405615f886SFrancis Yan tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED); 23415615f886SFrancis Yan 2342aa93466bSDavid S. Miller if (likely(sent_pkts)) { 2343684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 2344684bad11SYuchung Cheng tp->prr_out += sent_pkts; 23456ba8a3b1SNandita Dukkipati 23466ba8a3b1SNandita Dukkipati /* Send one loss probe per tail loss episode. */ 23476ba8a3b1SNandita Dukkipati if (push_one != 2) 23486ba8a3b1SNandita Dukkipati tcp_schedule_loss_probe(sk); 2349d2e1339fSBendik Rønning Opstad is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd); 2350ca8a2263SNeal Cardwell tcp_cwnd_validate(sk, is_cwnd_limited); 2351a2a385d6SEric Dumazet return false; 23521da177e4SLinus Torvalds } 2353b340b264SYuchung Cheng return !tp->packets_out && tcp_send_head(sk); 23546ba8a3b1SNandita Dukkipati } 23556ba8a3b1SNandita Dukkipati 23566ba8a3b1SNandita Dukkipati bool tcp_schedule_loss_probe(struct sock *sk) 23576ba8a3b1SNandita Dukkipati { 23586ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 23596ba8a3b1SNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 2360a2815817SNeal Cardwell u32 timeout, rto_delta_us; 23616ba8a3b1SNandita Dukkipati 23626ba8a3b1SNandita Dukkipati /* Don't do any loss probe on a Fast Open connection before 3WHS 23636ba8a3b1SNandita Dukkipati * finishes. 23646ba8a3b1SNandita Dukkipati */ 2365f9b99582SYuchung Cheng if (tp->fastopen_rsk) 23666ba8a3b1SNandita Dukkipati return false; 23676ba8a3b1SNandita Dukkipati 23686ba8a3b1SNandita Dukkipati /* Schedule a loss probe in 2*RTT for SACK capable connections 23696ba8a3b1SNandita Dukkipati * in Open state, that are either limited by cwnd or application. 23706ba8a3b1SNandita Dukkipati */ 2371bec41a11SYuchung Cheng if ((sysctl_tcp_early_retrans != 3 && sysctl_tcp_early_retrans != 4) || 2372bec41a11SYuchung Cheng !tp->packets_out || !tcp_is_sack(tp) || 2373bec41a11SYuchung Cheng icsk->icsk_ca_state != TCP_CA_Open) 23746ba8a3b1SNandita Dukkipati return false; 23756ba8a3b1SNandita Dukkipati 23766ba8a3b1SNandita Dukkipati if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) && 23776ba8a3b1SNandita Dukkipati tcp_send_head(sk)) 23786ba8a3b1SNandita Dukkipati return false; 23796ba8a3b1SNandita Dukkipati 2380bb4d991aSYuchung Cheng /* Probe timeout is 2*rtt. Add minimum RTO to account 2381f9b99582SYuchung Cheng * for delayed ack when there's one outstanding packet. If no RTT 2382f9b99582SYuchung Cheng * sample is available then probe after TCP_TIMEOUT_INIT. 23836ba8a3b1SNandita Dukkipati */ 2384bb4d991aSYuchung Cheng if (tp->srtt_us) { 2385bb4d991aSYuchung Cheng timeout = usecs_to_jiffies(tp->srtt_us >> 2); 23866ba8a3b1SNandita Dukkipati if (tp->packets_out == 1) 2387bb4d991aSYuchung Cheng timeout += TCP_RTO_MIN; 2388bb4d991aSYuchung Cheng else 2389bb4d991aSYuchung Cheng timeout += TCP_TIMEOUT_MIN; 2390bb4d991aSYuchung Cheng } else { 2391bb4d991aSYuchung Cheng timeout = TCP_TIMEOUT_INIT; 2392bb4d991aSYuchung Cheng } 23936ba8a3b1SNandita Dukkipati 2394a2815817SNeal Cardwell /* If the RTO formula yields an earlier time, then use that time. */ 2395a2815817SNeal Cardwell rto_delta_us = tcp_rto_delta_us(sk); /* How far in future is RTO? */ 2396a2815817SNeal Cardwell if (rto_delta_us > 0) 2397a2815817SNeal Cardwell timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us)); 23986ba8a3b1SNandita Dukkipati 23996ba8a3b1SNandita Dukkipati inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, 24006ba8a3b1SNandita Dukkipati TCP_RTO_MAX); 24016ba8a3b1SNandita Dukkipati return true; 24026ba8a3b1SNandita Dukkipati } 24036ba8a3b1SNandita Dukkipati 24041f3279aeSEric Dumazet /* Thanks to skb fast clones, we can detect if a prior transmit of 24051f3279aeSEric Dumazet * a packet is still in a qdisc or driver queue. 24061f3279aeSEric Dumazet * In this case, there is very little point doing a retransmit ! 24071f3279aeSEric Dumazet */ 24081f3279aeSEric Dumazet static bool skb_still_in_host_queue(const struct sock *sk, 24091f3279aeSEric Dumazet const struct sk_buff *skb) 24101f3279aeSEric Dumazet { 241139bb5e62SEric Dumazet if (unlikely(skb_fclone_busy(sk, skb))) { 2412c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), 24131f3279aeSEric Dumazet LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES); 24141f3279aeSEric Dumazet return true; 24151f3279aeSEric Dumazet } 24161f3279aeSEric Dumazet return false; 24171f3279aeSEric Dumazet } 24181f3279aeSEric Dumazet 2419b340b264SYuchung Cheng /* When probe timeout (PTO) fires, try send a new segment if possible, else 24206ba8a3b1SNandita Dukkipati * retransmit the last segment. 24216ba8a3b1SNandita Dukkipati */ 24226ba8a3b1SNandita Dukkipati void tcp_send_loss_probe(struct sock *sk) 24236ba8a3b1SNandita Dukkipati { 24249b717a8dSNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 24256ba8a3b1SNandita Dukkipati struct sk_buff *skb; 24266ba8a3b1SNandita Dukkipati int pcount; 24276ba8a3b1SNandita Dukkipati int mss = tcp_current_mss(sk); 24286ba8a3b1SNandita Dukkipati 2429b340b264SYuchung Cheng skb = tcp_send_head(sk); 2430b340b264SYuchung Cheng if (skb) { 2431b340b264SYuchung Cheng if (tcp_snd_wnd_test(tp, skb, mss)) { 2432b340b264SYuchung Cheng pcount = tp->packets_out; 2433b340b264SYuchung Cheng tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC); 2434b340b264SYuchung Cheng if (tp->packets_out > pcount) 2435b340b264SYuchung Cheng goto probe_sent; 24366ba8a3b1SNandita Dukkipati goto rearm_timer; 24376ba8a3b1SNandita Dukkipati } 2438b340b264SYuchung Cheng skb = tcp_write_queue_prev(sk, skb); 2439b340b264SYuchung Cheng } else { 2440b340b264SYuchung Cheng skb = tcp_write_queue_tail(sk); 2441b340b264SYuchung Cheng } 24426ba8a3b1SNandita Dukkipati 24439b717a8dSNandita Dukkipati /* At most one outstanding TLP retransmission. */ 24449b717a8dSNandita Dukkipati if (tp->tlp_high_seq) 24459b717a8dSNandita Dukkipati goto rearm_timer; 24469b717a8dSNandita Dukkipati 24476ba8a3b1SNandita Dukkipati /* Retransmit last segment. */ 24486ba8a3b1SNandita Dukkipati if (WARN_ON(!skb)) 24496ba8a3b1SNandita Dukkipati goto rearm_timer; 24506ba8a3b1SNandita Dukkipati 24511f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 24521f3279aeSEric Dumazet goto rearm_timer; 24531f3279aeSEric Dumazet 24546ba8a3b1SNandita Dukkipati pcount = tcp_skb_pcount(skb); 24556ba8a3b1SNandita Dukkipati if (WARN_ON(!pcount)) 24566ba8a3b1SNandita Dukkipati goto rearm_timer; 24576ba8a3b1SNandita Dukkipati 24586ba8a3b1SNandita Dukkipati if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) { 24596cc55e09SOctavian Purdila if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss, 24606cc55e09SOctavian Purdila GFP_ATOMIC))) 24616ba8a3b1SNandita Dukkipati goto rearm_timer; 2462b340b264SYuchung Cheng skb = tcp_write_queue_next(sk, skb); 24636ba8a3b1SNandita Dukkipati } 24646ba8a3b1SNandita Dukkipati 24656ba8a3b1SNandita Dukkipati if (WARN_ON(!skb || !tcp_skb_pcount(skb))) 24666ba8a3b1SNandita Dukkipati goto rearm_timer; 24676ba8a3b1SNandita Dukkipati 246810d3be56SEric Dumazet if (__tcp_retransmit_skb(sk, skb, 1)) 2469b340b264SYuchung Cheng goto rearm_timer; 24706ba8a3b1SNandita Dukkipati 24719b717a8dSNandita Dukkipati /* Record snd_nxt for loss detection. */ 24729b717a8dSNandita Dukkipati tp->tlp_high_seq = tp->snd_nxt; 24739b717a8dSNandita Dukkipati 2474b340b264SYuchung Cheng probe_sent: 2475c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES); 2476fcd16c0aSYuchung Cheng /* Reset s.t. tcp_rearm_rto will restart timer from now */ 2477fcd16c0aSYuchung Cheng inet_csk(sk)->icsk_pending = 0; 2478b340b264SYuchung Cheng rearm_timer: 2479fcd16c0aSYuchung Cheng tcp_rearm_rto(sk); 24801da177e4SLinus Torvalds } 24811da177e4SLinus Torvalds 2482a762a980SDavid S. Miller /* Push out any pending frames which were held back due to 2483a762a980SDavid S. Miller * TCP_CORK or attempt at coalescing tiny packets. 2484a762a980SDavid S. Miller * The socket must be locked by the caller. 2485a762a980SDavid S. Miller */ 24869e412ba7SIlpo Järvinen void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 24879e412ba7SIlpo Järvinen int nonagle) 2488a762a980SDavid S. Miller { 2489726e07a8SIlpo Järvinen /* If we are closed, the bytes will have to remain here. 2490726e07a8SIlpo Järvinen * In time closedown will finish, we empty the write queue and 2491726e07a8SIlpo Järvinen * all will be happy. 2492726e07a8SIlpo Järvinen */ 2493726e07a8SIlpo Järvinen if (unlikely(sk->sk_state == TCP_CLOSE)) 2494726e07a8SIlpo Järvinen return; 2495726e07a8SIlpo Järvinen 249699a1dec7SMel Gorman if (tcp_write_xmit(sk, cur_mss, nonagle, 0, 24977450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC))) 24989e412ba7SIlpo Järvinen tcp_check_probe_timer(sk); 2499a762a980SDavid S. Miller } 2500a762a980SDavid S. Miller 2501c1b4a7e6SDavid S. Miller /* Send _single_ skb sitting at the send head. This function requires 2502c1b4a7e6SDavid S. Miller * true push pending frames to setup probe timer etc. 2503c1b4a7e6SDavid S. Miller */ 2504c1b4a7e6SDavid S. Miller void tcp_push_one(struct sock *sk, unsigned int mss_now) 2505c1b4a7e6SDavid S. Miller { 2506fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 2507c1b4a7e6SDavid S. Miller 2508c1b4a7e6SDavid S. Miller BUG_ON(!skb || skb->len < mss_now); 2509c1b4a7e6SDavid S. Miller 2510d5dd9175SIlpo Järvinen tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); 2511c1b4a7e6SDavid S. Miller } 2512c1b4a7e6SDavid S. Miller 25131da177e4SLinus Torvalds /* This function returns the amount that we can raise the 25141da177e4SLinus Torvalds * usable window based on the following constraints 25151da177e4SLinus Torvalds * 25161da177e4SLinus Torvalds * 1. The window can never be shrunk once it is offered (RFC 793) 25171da177e4SLinus Torvalds * 2. We limit memory per socket 25181da177e4SLinus Torvalds * 25191da177e4SLinus Torvalds * RFC 1122: 25201da177e4SLinus Torvalds * "the suggested [SWS] avoidance algorithm for the receiver is to keep 25211da177e4SLinus Torvalds * RECV.NEXT + RCV.WIN fixed until: 25221da177e4SLinus Torvalds * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 25231da177e4SLinus Torvalds * 25241da177e4SLinus Torvalds * i.e. don't raise the right edge of the window until you can raise 25251da177e4SLinus Torvalds * it at least MSS bytes. 25261da177e4SLinus Torvalds * 25271da177e4SLinus Torvalds * Unfortunately, the recommended algorithm breaks header prediction, 25281da177e4SLinus Torvalds * since header prediction assumes th->window stays fixed. 25291da177e4SLinus Torvalds * 25301da177e4SLinus Torvalds * Strictly speaking, keeping th->window fixed violates the receiver 25311da177e4SLinus Torvalds * side SWS prevention criteria. The problem is that under this rule 25321da177e4SLinus Torvalds * a stream of single byte packets will cause the right side of the 25331da177e4SLinus Torvalds * window to always advance by a single byte. 25341da177e4SLinus Torvalds * 25351da177e4SLinus Torvalds * Of course, if the sender implements sender side SWS prevention 25361da177e4SLinus Torvalds * then this will not be a problem. 25371da177e4SLinus Torvalds * 25381da177e4SLinus Torvalds * BSD seems to make the following compromise: 25391da177e4SLinus Torvalds * 25401da177e4SLinus Torvalds * If the free space is less than the 1/4 of the maximum 25411da177e4SLinus Torvalds * space available and the free space is less than 1/2 mss, 25421da177e4SLinus Torvalds * then set the window to 0. 25431da177e4SLinus Torvalds * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 25441da177e4SLinus Torvalds * Otherwise, just prevent the window from shrinking 25451da177e4SLinus Torvalds * and from being larger than the largest representable value. 25461da177e4SLinus Torvalds * 25471da177e4SLinus Torvalds * This prevents incremental opening of the window in the regime 25481da177e4SLinus Torvalds * where TCP is limited by the speed of the reader side taking 25491da177e4SLinus Torvalds * data out of the TCP receive queue. It does nothing about 25501da177e4SLinus Torvalds * those cases where the window is constrained on the sender side 25511da177e4SLinus Torvalds * because the pipeline is full. 25521da177e4SLinus Torvalds * 25531da177e4SLinus Torvalds * BSD also seems to "accidentally" limit itself to windows that are a 25541da177e4SLinus Torvalds * multiple of MSS, at least until the free space gets quite small. 25551da177e4SLinus Torvalds * This would appear to be a side effect of the mbuf implementation. 25561da177e4SLinus Torvalds * Combining these two algorithms results in the observed behavior 25571da177e4SLinus Torvalds * of having a fixed window size at almost all times. 25581da177e4SLinus Torvalds * 25591da177e4SLinus Torvalds * Below we obtain similar behavior by forcing the offered window to 25601da177e4SLinus Torvalds * a multiple of the mss when it is feasible to do so. 25611da177e4SLinus Torvalds * 25621da177e4SLinus Torvalds * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 25631da177e4SLinus Torvalds * Regular options like TIMESTAMP are taken into account. 25641da177e4SLinus Torvalds */ 25651da177e4SLinus Torvalds u32 __tcp_select_window(struct sock *sk) 25661da177e4SLinus Torvalds { 2567463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 25681da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2569caa20d9aSStephen Hemminger /* MSS for the peer's data. Previous versions used mss_clamp 25701da177e4SLinus Torvalds * here. I don't know if the value based on our guesses 25711da177e4SLinus Torvalds * of peer's MSS is better for the performance. It's more correct 25721da177e4SLinus Torvalds * but may be worse for the performance because of rcv_mss 25731da177e4SLinus Torvalds * fluctuations. --SAW 1998/11/1 25741da177e4SLinus Torvalds */ 2575463c84b9SArnaldo Carvalho de Melo int mss = icsk->icsk_ack.rcv_mss; 25761da177e4SLinus Torvalds int free_space = tcp_space(sk); 257786c1a045SFlorian Westphal int allowed_space = tcp_full_space(sk); 257886c1a045SFlorian Westphal int full_space = min_t(int, tp->window_clamp, allowed_space); 25791da177e4SLinus Torvalds int window; 25801da177e4SLinus Torvalds 258106425c30SEric Dumazet if (unlikely(mss > full_space)) { 25821da177e4SLinus Torvalds mss = full_space; 258306425c30SEric Dumazet if (mss <= 0) 258406425c30SEric Dumazet return 0; 258506425c30SEric Dumazet } 2586b92edbe0SEric Dumazet if (free_space < (full_space >> 1)) { 2587463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.quick = 0; 25881da177e4SLinus Torvalds 2589b8da51ebSEric Dumazet if (tcp_under_memory_pressure(sk)) 2590056834d9SIlpo Järvinen tp->rcv_ssthresh = min(tp->rcv_ssthresh, 2591056834d9SIlpo Järvinen 4U * tp->advmss); 25921da177e4SLinus Torvalds 259386c1a045SFlorian Westphal /* free_space might become our new window, make sure we don't 259486c1a045SFlorian Westphal * increase it due to wscale. 259586c1a045SFlorian Westphal */ 259686c1a045SFlorian Westphal free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale); 259786c1a045SFlorian Westphal 259886c1a045SFlorian Westphal /* if free space is less than mss estimate, or is below 1/16th 259986c1a045SFlorian Westphal * of the maximum allowed, try to move to zero-window, else 260086c1a045SFlorian Westphal * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and 260186c1a045SFlorian Westphal * new incoming data is dropped due to memory limits. 260286c1a045SFlorian Westphal * With large window, mss test triggers way too late in order 260386c1a045SFlorian Westphal * to announce zero window in time before rmem limit kicks in. 260486c1a045SFlorian Westphal */ 260586c1a045SFlorian Westphal if (free_space < (allowed_space >> 4) || free_space < mss) 26061da177e4SLinus Torvalds return 0; 26071da177e4SLinus Torvalds } 26081da177e4SLinus Torvalds 26091da177e4SLinus Torvalds if (free_space > tp->rcv_ssthresh) 26101da177e4SLinus Torvalds free_space = tp->rcv_ssthresh; 26111da177e4SLinus Torvalds 26121da177e4SLinus Torvalds /* Don't do rounding if we are using window scaling, since the 26131da177e4SLinus Torvalds * scaled window will not line up with the MSS boundary anyway. 26141da177e4SLinus Torvalds */ 26151da177e4SLinus Torvalds if (tp->rx_opt.rcv_wscale) { 26161da177e4SLinus Torvalds window = free_space; 26171da177e4SLinus Torvalds 26181da177e4SLinus Torvalds /* Advertise enough space so that it won't get scaled away. 26191da177e4SLinus Torvalds * Import case: prevent zero window announcement if 26201da177e4SLinus Torvalds * 1<<rcv_wscale > mss. 26211da177e4SLinus Torvalds */ 26221935299dSGao Feng window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale)); 26231da177e4SLinus Torvalds } else { 26241935299dSGao Feng window = tp->rcv_wnd; 26251da177e4SLinus Torvalds /* Get the largest window that is a nice multiple of mss. 26261da177e4SLinus Torvalds * Window clamp already applied above. 26271da177e4SLinus Torvalds * If our current window offering is within 1 mss of the 26281da177e4SLinus Torvalds * free space we just keep it. This prevents the divide 26291da177e4SLinus Torvalds * and multiply from happening most of the time. 26301da177e4SLinus Torvalds * We also don't do any window rounding when the free space 26311da177e4SLinus Torvalds * is too small. 26321da177e4SLinus Torvalds */ 26331da177e4SLinus Torvalds if (window <= free_space - mss || window > free_space) 26341935299dSGao Feng window = rounddown(free_space, mss); 263584565070SJohn Heffner else if (mss == full_space && 2636b92edbe0SEric Dumazet free_space > window + (full_space >> 1)) 263784565070SJohn Heffner window = free_space; 26381da177e4SLinus Torvalds } 26391da177e4SLinus Torvalds 26401da177e4SLinus Torvalds return window; 26411da177e4SLinus Torvalds } 26421da177e4SLinus Torvalds 2643cfea5a68SMartin KaFai Lau void tcp_skb_collapse_tstamp(struct sk_buff *skb, 2644082ac2d5SMartin KaFai Lau const struct sk_buff *next_skb) 2645082ac2d5SMartin KaFai Lau { 26460a2cf20cSSoheil Hassas Yeganeh if (unlikely(tcp_has_tx_tstamp(next_skb))) { 26470a2cf20cSSoheil Hassas Yeganeh const struct skb_shared_info *next_shinfo = 26480a2cf20cSSoheil Hassas Yeganeh skb_shinfo(next_skb); 2649082ac2d5SMartin KaFai Lau struct skb_shared_info *shinfo = skb_shinfo(skb); 2650082ac2d5SMartin KaFai Lau 26510a2cf20cSSoheil Hassas Yeganeh shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP; 2652082ac2d5SMartin KaFai Lau shinfo->tskey = next_shinfo->tskey; 26532de8023eSMartin KaFai Lau TCP_SKB_CB(skb)->txstamp_ack |= 26542de8023eSMartin KaFai Lau TCP_SKB_CB(next_skb)->txstamp_ack; 2655082ac2d5SMartin KaFai Lau } 2656082ac2d5SMartin KaFai Lau } 2657082ac2d5SMartin KaFai Lau 26584a17fc3aSIlpo Järvinen /* Collapses two adjacent SKB's during retransmission. */ 2659f8071cdeSEric Dumazet static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) 26601da177e4SLinus Torvalds { 26611da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2662fe067e8aSDavid S. Miller struct sk_buff *next_skb = tcp_write_queue_next(sk, skb); 2663058dc334SIlpo Järvinen int skb_size, next_skb_size; 26641da177e4SLinus Torvalds 2665058dc334SIlpo Järvinen skb_size = skb->len; 2666058dc334SIlpo Järvinen next_skb_size = next_skb->len; 26671da177e4SLinus Torvalds 2668058dc334SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); 26691da177e4SLinus Torvalds 2670f8071cdeSEric Dumazet if (next_skb_size) { 2671f8071cdeSEric Dumazet if (next_skb_size <= skb_availroom(skb)) 2672f8071cdeSEric Dumazet skb_copy_bits(next_skb, 0, skb_put(skb, next_skb_size), 2673f8071cdeSEric Dumazet next_skb_size); 2674f8071cdeSEric Dumazet else if (!skb_shift(skb, next_skb, next_skb_size)) 2675f8071cdeSEric Dumazet return false; 2676f8071cdeSEric Dumazet } 26776859d494SIlpo Järvinen tcp_highest_sack_combine(sk, next_skb, skb); 2678a6963a6bSIlpo Järvinen 2679fe067e8aSDavid S. Miller tcp_unlink_write_queue(next_skb, sk); 26801da177e4SLinus Torvalds 268152d570aaSJarek Poplawski if (next_skb->ip_summed == CHECKSUM_PARTIAL) 268252d570aaSJarek Poplawski skb->ip_summed = CHECKSUM_PARTIAL; 26831da177e4SLinus Torvalds 268484fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 26851da177e4SLinus Torvalds skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); 26861da177e4SLinus Torvalds 26871da177e4SLinus Torvalds /* Update sequence range on original skb. */ 26881da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 26891da177e4SLinus Torvalds 2690e6c7d085SIlpo Järvinen /* Merge over control information. This moves PSH/FIN etc. over */ 26914de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; 26921da177e4SLinus Torvalds 26931da177e4SLinus Torvalds /* All done, get rid of second SKB and account for it so 26941da177e4SLinus Torvalds * packet counting does not break. 26951da177e4SLinus Torvalds */ 26964828e7f4SIlpo Järvinen TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; 2697a643b5d4SMartin KaFai Lau TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor; 2698b7689205SIlpo Järvinen 2699b7689205SIlpo Järvinen /* changed transmit queue under us so clear hints */ 2700ef9da47cSIlpo Järvinen tcp_clear_retrans_hints_partial(tp); 2701ef9da47cSIlpo Järvinen if (next_skb == tp->retransmit_skb_hint) 2702ef9da47cSIlpo Järvinen tp->retransmit_skb_hint = skb; 2703b7689205SIlpo Järvinen 2704797108d1SIlpo Järvinen tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); 2705797108d1SIlpo Järvinen 2706082ac2d5SMartin KaFai Lau tcp_skb_collapse_tstamp(skb, next_skb); 2707082ac2d5SMartin KaFai Lau 27083ab224beSHideo Aoki sk_wmem_free_skb(sk, next_skb); 2709f8071cdeSEric Dumazet return true; 27101da177e4SLinus Torvalds } 27111da177e4SLinus Torvalds 271267edfef7SAndi Kleen /* Check if coalescing SKBs is legal. */ 2713a2a385d6SEric Dumazet static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) 27144a17fc3aSIlpo Järvinen { 27154a17fc3aSIlpo Järvinen if (tcp_skb_pcount(skb) > 1) 2716a2a385d6SEric Dumazet return false; 27174a17fc3aSIlpo Järvinen if (skb_cloned(skb)) 2718a2a385d6SEric Dumazet return false; 27194a17fc3aSIlpo Järvinen if (skb == tcp_send_head(sk)) 2720a2a385d6SEric Dumazet return false; 27212331ccc5SEric Dumazet /* Some heuristics for collapsing over SACK'd could be invented */ 27224a17fc3aSIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 2723a2a385d6SEric Dumazet return false; 27244a17fc3aSIlpo Järvinen 2725a2a385d6SEric Dumazet return true; 27264a17fc3aSIlpo Järvinen } 27274a17fc3aSIlpo Järvinen 272867edfef7SAndi Kleen /* Collapse packets in the retransmit queue to make to create 272967edfef7SAndi Kleen * less packets on the wire. This is only done on retransmission. 273067edfef7SAndi Kleen */ 27314a17fc3aSIlpo Järvinen static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, 27324a17fc3aSIlpo Järvinen int space) 27334a17fc3aSIlpo Järvinen { 27344a17fc3aSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 27354a17fc3aSIlpo Järvinen struct sk_buff *skb = to, *tmp; 2736a2a385d6SEric Dumazet bool first = true; 27374a17fc3aSIlpo Järvinen 27384a17fc3aSIlpo Järvinen if (!sysctl_tcp_retrans_collapse) 27394a17fc3aSIlpo Järvinen return; 27404de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 27414a17fc3aSIlpo Järvinen return; 27424a17fc3aSIlpo Järvinen 27434a17fc3aSIlpo Järvinen tcp_for_write_queue_from_safe(skb, tmp, sk) { 27444a17fc3aSIlpo Järvinen if (!tcp_can_collapse(sk, skb)) 27454a17fc3aSIlpo Järvinen break; 27464a17fc3aSIlpo Järvinen 2747a643b5d4SMartin KaFai Lau if (!tcp_skb_can_collapse_to(to)) 2748a643b5d4SMartin KaFai Lau break; 2749a643b5d4SMartin KaFai Lau 27504a17fc3aSIlpo Järvinen space -= skb->len; 27514a17fc3aSIlpo Järvinen 27524a17fc3aSIlpo Järvinen if (first) { 2753a2a385d6SEric Dumazet first = false; 27544a17fc3aSIlpo Järvinen continue; 27554a17fc3aSIlpo Järvinen } 27564a17fc3aSIlpo Järvinen 27574a17fc3aSIlpo Järvinen if (space < 0) 27584a17fc3aSIlpo Järvinen break; 27594a17fc3aSIlpo Järvinen 27604a17fc3aSIlpo Järvinen if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) 27614a17fc3aSIlpo Järvinen break; 27624a17fc3aSIlpo Järvinen 2763f8071cdeSEric Dumazet if (!tcp_collapse_retrans(sk, to)) 2764f8071cdeSEric Dumazet break; 27654a17fc3aSIlpo Järvinen } 27664a17fc3aSIlpo Järvinen } 27674a17fc3aSIlpo Järvinen 27681da177e4SLinus Torvalds /* This retransmits one SKB. Policy decisions and retransmit queue 27691da177e4SLinus Torvalds * state updates are done by the caller. Returns non-zero if an 27701da177e4SLinus Torvalds * error occurred which prevented the send. 27711da177e4SLinus Torvalds */ 277210d3be56SEric Dumazet int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) 27731da177e4SLinus Torvalds { 27745d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 277510d3be56SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 27767d227cd2SSridhar Samudrala unsigned int cur_mss; 277710d3be56SEric Dumazet int diff, len, err; 27781da177e4SLinus Torvalds 277910d3be56SEric Dumazet 278010d3be56SEric Dumazet /* Inconclusive MTU probe */ 278110d3be56SEric Dumazet if (icsk->icsk_mtup.probe_size) 27825d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 27835d424d5aSJohn Heffner 27841da177e4SLinus Torvalds /* Do not sent more than we queued. 1/4 is reserved for possible 2785caa20d9aSStephen Hemminger * copying overhead: fragmentation, tunneling, mangling etc. 27861da177e4SLinus Torvalds */ 278714afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) > 2788ffb4d6c8SEric Dumazet min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), 2789ffb4d6c8SEric Dumazet sk->sk_sndbuf)) 27901da177e4SLinus Torvalds return -EAGAIN; 27911da177e4SLinus Torvalds 27921f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 27931f3279aeSEric Dumazet return -EBUSY; 27941f3279aeSEric Dumazet 27951da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 27961da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 27971da177e4SLinus Torvalds BUG(); 27981da177e4SLinus Torvalds if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 27991da177e4SLinus Torvalds return -ENOMEM; 28001da177e4SLinus Torvalds } 28011da177e4SLinus Torvalds 28027d227cd2SSridhar Samudrala if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 28037d227cd2SSridhar Samudrala return -EHOSTUNREACH; /* Routing failure or similar. */ 28047d227cd2SSridhar Samudrala 28050c54b85fSIlpo Järvinen cur_mss = tcp_current_mss(sk); 28067d227cd2SSridhar Samudrala 28071da177e4SLinus Torvalds /* If receiver has shrunk his window, and skb is out of 28081da177e4SLinus Torvalds * new window, do not retransmit it. The exception is the 28091da177e4SLinus Torvalds * case, when window is shrunk to zero. In this case 28101da177e4SLinus Torvalds * our retransmit serves as a zero window probe. 28111da177e4SLinus Torvalds */ 28129d4fb27dSJoe Perches if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && 28139d4fb27dSJoe Perches TCP_SKB_CB(skb)->seq != tp->snd_una) 28141da177e4SLinus Torvalds return -EAGAIN; 28151da177e4SLinus Torvalds 281610d3be56SEric Dumazet len = cur_mss * segs; 281710d3be56SEric Dumazet if (skb->len > len) { 281810d3be56SEric Dumazet if (tcp_fragment(sk, skb, len, cur_mss, GFP_ATOMIC)) 28191da177e4SLinus Torvalds return -ENOMEM; /* We'll try again later. */ 282002276f3cSIlpo Järvinen } else { 2821c52e2421SEric Dumazet if (skb_unclone(skb, GFP_ATOMIC)) 2822c52e2421SEric Dumazet return -ENOMEM; 282310d3be56SEric Dumazet 282410d3be56SEric Dumazet diff = tcp_skb_pcount(skb); 282510d3be56SEric Dumazet tcp_set_skb_tso_segs(skb, cur_mss); 282610d3be56SEric Dumazet diff -= tcp_skb_pcount(skb); 282710d3be56SEric Dumazet if (diff) 282810d3be56SEric Dumazet tcp_adjust_pcount(sk, skb, diff); 282910d3be56SEric Dumazet if (skb->len < cur_mss) 283010d3be56SEric Dumazet tcp_retrans_try_collapse(sk, skb, cur_mss); 28311da177e4SLinus Torvalds } 28321da177e4SLinus Torvalds 283349213555SDaniel Borkmann /* RFC3168, section 6.1.1.1. ECN fallback */ 283449213555SDaniel Borkmann if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN) 283549213555SDaniel Borkmann tcp_ecn_clear_syn(sk, skb); 283649213555SDaniel Borkmann 2837678550c6SYuchung Cheng /* Update global and local TCP statistics. */ 2838678550c6SYuchung Cheng segs = tcp_skb_pcount(skb); 2839678550c6SYuchung Cheng TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs); 2840678550c6SYuchung Cheng if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 2841678550c6SYuchung Cheng __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 2842678550c6SYuchung Cheng tp->total_retrans += segs; 2843678550c6SYuchung Cheng 284450bceae9SThomas Graf /* make sure skb->data is aligned on arches that require it 284550bceae9SThomas Graf * and check if ack-trimming & collapsing extended the headroom 284650bceae9SThomas Graf * beyond what csum_start can cover. 284750bceae9SThomas Graf */ 284850bceae9SThomas Graf if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) || 284950bceae9SThomas Graf skb_headroom(skb) >= 0xFFFF)) { 285010a81980SEric Dumazet struct sk_buff *nskb; 285110a81980SEric Dumazet 2852*e2080072SEric Dumazet tcp_skb_tsorted_save(skb) { 285310a81980SEric Dumazet nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC); 2854c84a5711SYuchung Cheng err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) : 2855117632e6SEric Dumazet -ENOBUFS; 2856*e2080072SEric Dumazet } tcp_skb_tsorted_restore(skb); 2857*e2080072SEric Dumazet 28588c72c65bSEric Dumazet if (!err) 2859*e2080072SEric Dumazet tcp_update_skb_after_send(tp, skb); 2860117632e6SEric Dumazet } else { 2861c84a5711SYuchung Cheng err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 2862117632e6SEric Dumazet } 2863c84a5711SYuchung Cheng 2864fc9f3501SEric Dumazet if (likely(!err)) { 2865c84a5711SYuchung Cheng TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS; 2866678550c6SYuchung Cheng } else if (err != -EBUSY) { 2867678550c6SYuchung Cheng NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL); 2868fc9f3501SEric Dumazet } 2869c84a5711SYuchung Cheng return err; 287093b174adSYuchung Cheng } 287193b174adSYuchung Cheng 287210d3be56SEric Dumazet int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) 287393b174adSYuchung Cheng { 287493b174adSYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 287510d3be56SEric Dumazet int err = __tcp_retransmit_skb(sk, skb, segs); 28761da177e4SLinus Torvalds 28771da177e4SLinus Torvalds if (err == 0) { 28781da177e4SLinus Torvalds #if FASTRETRANS_DEBUG > 0 28791da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 2880e87cc472SJoe Perches net_dbg_ratelimited("retrans_out leaked\n"); 28811da177e4SLinus Torvalds } 28821da177e4SLinus Torvalds #endif 28831da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 28841da177e4SLinus Torvalds tp->retrans_out += tcp_skb_pcount(skb); 28851da177e4SLinus Torvalds 28861da177e4SLinus Torvalds /* Save stamp of the first retransmit. */ 28871da177e4SLinus Torvalds if (!tp->retrans_stamp) 28887faee5c0SEric Dumazet tp->retrans_stamp = tcp_skb_timestamp(skb); 28891da177e4SLinus Torvalds 28901da177e4SLinus Torvalds } 28916e08d5e3SYuchung Cheng 28926e08d5e3SYuchung Cheng if (tp->undo_retrans < 0) 28936e08d5e3SYuchung Cheng tp->undo_retrans = 0; 28946e08d5e3SYuchung Cheng tp->undo_retrans += tcp_skb_pcount(skb); 28951da177e4SLinus Torvalds return err; 28961da177e4SLinus Torvalds } 28971da177e4SLinus Torvalds 28981da177e4SLinus Torvalds /* This gets called after a retransmit timeout, and the initially 28991da177e4SLinus Torvalds * retransmitted data is acknowledged. It tries to continue 29001da177e4SLinus Torvalds * resending the rest of the retransmit queue, until either 29011da177e4SLinus Torvalds * we've sent it all or the congestion window limit is reached. 29021da177e4SLinus Torvalds * If doing SACK, the first ACK which comes back for a timeout 29031da177e4SLinus Torvalds * based retransmit packet might feed us FACK information again. 29041da177e4SLinus Torvalds * If so, we use it to avoid unnecessarily retransmissions. 29051da177e4SLinus Torvalds */ 29061da177e4SLinus Torvalds void tcp_xmit_retransmit_queue(struct sock *sk) 29071da177e4SLinus Torvalds { 29086687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 29091da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 29101da177e4SLinus Torvalds struct sk_buff *skb; 29110e1c54c2SIlpo Järvinen struct sk_buff *hole = NULL; 2912840a3cbeSYuchung Cheng u32 max_segs; 291361eb55f4SIlpo Järvinen int mib_idx; 29146a438bbeSStephen Hemminger 291545e77d31SIlpo Järvinen if (!tp->packets_out) 291645e77d31SIlpo Järvinen return; 291745e77d31SIlpo Järvinen 2918618d9f25SIlpo Järvinen if (tp->retransmit_skb_hint) { 29196a438bbeSStephen Hemminger skb = tp->retransmit_skb_hint; 2920618d9f25SIlpo Järvinen } else { 2921fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 2922618d9f25SIlpo Järvinen } 29231da177e4SLinus Torvalds 2924ed6e7268SNeal Cardwell max_segs = tcp_tso_segs(sk, tcp_current_mss(sk)); 2925fe067e8aSDavid S. Miller tcp_for_write_queue_from(skb, sk) { 2926dca0aaf8SEric Dumazet __u8 sacked; 292710d3be56SEric Dumazet int segs; 29281da177e4SLinus Torvalds 2929fe067e8aSDavid S. Miller if (skb == tcp_send_head(sk)) 2930fe067e8aSDavid S. Miller break; 2931218af599SEric Dumazet 2932218af599SEric Dumazet if (tcp_pacing_check(sk)) 2933218af599SEric Dumazet break; 2934218af599SEric Dumazet 29356a438bbeSStephen Hemminger /* we could do better than to assign each time */ 293651456b29SIan Morris if (!hole) 29376a438bbeSStephen Hemminger tp->retransmit_skb_hint = skb; 29386a438bbeSStephen Hemminger 293910d3be56SEric Dumazet segs = tp->snd_cwnd - tcp_packets_in_flight(tp); 294010d3be56SEric Dumazet if (segs <= 0) 29411da177e4SLinus Torvalds return; 2942dca0aaf8SEric Dumazet sacked = TCP_SKB_CB(skb)->sacked; 2943a3d2e9f8SEric Dumazet /* In case tcp_shift_skb_data() have aggregated large skbs, 2944a3d2e9f8SEric Dumazet * we need to make sure not sending too bigs TSO packets 2945a3d2e9f8SEric Dumazet */ 2946a3d2e9f8SEric Dumazet segs = min_t(int, segs, max_segs); 29470e1c54c2SIlpo Järvinen 2948840a3cbeSYuchung Cheng if (tp->retrans_out >= tp->lost_out) { 2949006f582cSIlpo Järvinen break; 29500e1c54c2SIlpo Järvinen } else if (!(sacked & TCPCB_LOST)) { 295151456b29SIan Morris if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) 29520e1c54c2SIlpo Järvinen hole = skb; 295361eb55f4SIlpo Järvinen continue; 29541da177e4SLinus Torvalds 29550e1c54c2SIlpo Järvinen } else { 29560e1c54c2SIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Loss) 29570e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFASTRETRANS; 29580e1c54c2SIlpo Järvinen else 29590e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; 29600e1c54c2SIlpo Järvinen } 29610e1c54c2SIlpo Järvinen 29620e1c54c2SIlpo Järvinen if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) 296361eb55f4SIlpo Järvinen continue; 296440b215e5SPavel Emelyanov 2965f9616c35SEric Dumazet if (tcp_small_queue_check(sk, skb, 1)) 2966f9616c35SEric Dumazet return; 2967f9616c35SEric Dumazet 296810d3be56SEric Dumazet if (tcp_retransmit_skb(sk, skb, segs)) 29691da177e4SLinus Torvalds return; 297024ab6becSYuchung Cheng 2971de1d6578SYuchung Cheng NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb)); 29721da177e4SLinus Torvalds 2973684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 2974a262f0cdSNandita Dukkipati tp->prr_out += tcp_skb_pcount(skb); 2975a262f0cdSNandita Dukkipati 297657dde7f7SYuchung Cheng if (skb == tcp_write_queue_head(sk) && 297757dde7f7SYuchung Cheng icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT) 2978463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 29793f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, 29803f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 29811da177e4SLinus Torvalds } 29821da177e4SLinus Torvalds } 29831da177e4SLinus Torvalds 2984d83769a5SEric Dumazet /* We allow to exceed memory limits for FIN packets to expedite 2985d83769a5SEric Dumazet * connection tear down and (memory) recovery. 2986845704a5SEric Dumazet * Otherwise tcp_send_fin() could be tempted to either delay FIN 2987845704a5SEric Dumazet * or even be forced to close flow without any FIN. 2988a6c5ea4cSEric Dumazet * In general, we want to allow one skb per socket to avoid hangs 2989a6c5ea4cSEric Dumazet * with edge trigger epoll() 2990d83769a5SEric Dumazet */ 2991a6c5ea4cSEric Dumazet void sk_forced_mem_schedule(struct sock *sk, int size) 2992d83769a5SEric Dumazet { 2993e805605cSJohannes Weiner int amt; 2994d83769a5SEric Dumazet 2995d83769a5SEric Dumazet if (size <= sk->sk_forward_alloc) 2996d83769a5SEric Dumazet return; 2997d83769a5SEric Dumazet amt = sk_mem_pages(size); 2998d83769a5SEric Dumazet sk->sk_forward_alloc += amt * SK_MEM_QUANTUM; 2999e805605cSJohannes Weiner sk_memory_allocated_add(sk, amt); 3000e805605cSJohannes Weiner 3001baac50bbSJohannes Weiner if (mem_cgroup_sockets_enabled && sk->sk_memcg) 3002baac50bbSJohannes Weiner mem_cgroup_charge_skmem(sk->sk_memcg, amt); 3003d83769a5SEric Dumazet } 3004d83769a5SEric Dumazet 3005845704a5SEric Dumazet /* Send a FIN. The caller locks the socket for us. 3006845704a5SEric Dumazet * We should try to send a FIN packet really hard, but eventually give up. 30071da177e4SLinus Torvalds */ 30081da177e4SLinus Torvalds void tcp_send_fin(struct sock *sk) 30091da177e4SLinus Torvalds { 3010845704a5SEric Dumazet struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk); 30111da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 30121da177e4SLinus Torvalds 3013845704a5SEric Dumazet /* Optimization, tack on the FIN if we have one skb in write queue and 3014845704a5SEric Dumazet * this skb was not yet sent, or we are under memory pressure. 3015845704a5SEric Dumazet * Note: in the latter case, FIN packet will be sent after a timeout, 3016845704a5SEric Dumazet * as TCP stack thinks it has already been transmitted. 30171da177e4SLinus Torvalds */ 3018b8da51ebSEric Dumazet if (tskb && (tcp_send_head(sk) || tcp_under_memory_pressure(sk))) { 3019845704a5SEric Dumazet coalesce: 3020845704a5SEric Dumazet TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN; 3021845704a5SEric Dumazet TCP_SKB_CB(tskb)->end_seq++; 30221da177e4SLinus Torvalds tp->write_seq++; 3023845704a5SEric Dumazet if (!tcp_send_head(sk)) { 3024845704a5SEric Dumazet /* This means tskb was already sent. 3025845704a5SEric Dumazet * Pretend we included the FIN on previous transmit. 3026845704a5SEric Dumazet * We need to set tp->snd_nxt to the value it would have 3027845704a5SEric Dumazet * if FIN had been sent. This is because retransmit path 3028845704a5SEric Dumazet * does not change tp->snd_nxt. 3029845704a5SEric Dumazet */ 3030845704a5SEric Dumazet tp->snd_nxt++; 3031845704a5SEric Dumazet return; 3032845704a5SEric Dumazet } 30331da177e4SLinus Torvalds } else { 3034845704a5SEric Dumazet skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation); 3035845704a5SEric Dumazet if (unlikely(!skb)) { 3036845704a5SEric Dumazet if (tskb) 3037845704a5SEric Dumazet goto coalesce; 3038845704a5SEric Dumazet return; 30391da177e4SLinus Torvalds } 3040*e2080072SEric Dumazet INIT_LIST_HEAD(&skb->tcp_tsorted_anchor); 3041d83769a5SEric Dumazet skb_reserve(skb, MAX_TCP_HEADER); 3042a6c5ea4cSEric Dumazet sk_forced_mem_schedule(sk, skb->truesize); 30431da177e4SLinus Torvalds /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 3044e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tp->write_seq, 3045a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_FIN); 30461da177e4SLinus Torvalds tcp_queue_skb(sk, skb); 30471da177e4SLinus Torvalds } 3048845704a5SEric Dumazet __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF); 30491da177e4SLinus Torvalds } 30501da177e4SLinus Torvalds 30511da177e4SLinus Torvalds /* We get here when a process closes a file descriptor (either due to 30521da177e4SLinus Torvalds * an explicit close() or as a byproduct of exit()'ing) and there 30531da177e4SLinus Torvalds * was unread data in the receive queue. This behavior is recommended 305465bb723cSGerrit Renker * by RFC 2525, section 2.17. -DaveM 30551da177e4SLinus Torvalds */ 3056dd0fc66fSAl Viro void tcp_send_active_reset(struct sock *sk, gfp_t priority) 30571da177e4SLinus Torvalds { 30581da177e4SLinus Torvalds struct sk_buff *skb; 30591da177e4SLinus Torvalds 30607cc2b043SGao Feng TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); 30617cc2b043SGao Feng 30621da177e4SLinus Torvalds /* NOTE: No TCP options attached and we never retransmit this. */ 30631da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, priority); 30641da177e4SLinus Torvalds if (!skb) { 30654e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 30661da177e4SLinus Torvalds return; 30671da177e4SLinus Torvalds } 30681da177e4SLinus Torvalds 30691da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 30701da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 3071e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), 3072a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_RST); 30739a568de4SEric Dumazet tcp_mstamp_refresh(tcp_sk(sk)); 30741da177e4SLinus Torvalds /* Send it off. */ 3075dfb4b9dcSDavid S. Miller if (tcp_transmit_skb(sk, skb, 0, priority)) 30764e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 30771da177e4SLinus Torvalds } 30781da177e4SLinus Torvalds 307967edfef7SAndi Kleen /* Send a crossed SYN-ACK during socket establishment. 308067edfef7SAndi Kleen * WARNING: This routine must only be called when we have already sent 30811da177e4SLinus Torvalds * a SYN packet that crossed the incoming SYN that caused this routine 30821da177e4SLinus Torvalds * to get called. If this assumption fails then the initial rcv_wnd 30831da177e4SLinus Torvalds * and rcv_wscale values will not be correct. 30841da177e4SLinus Torvalds */ 30851da177e4SLinus Torvalds int tcp_send_synack(struct sock *sk) 30861da177e4SLinus Torvalds { 30871da177e4SLinus Torvalds struct sk_buff *skb; 30881da177e4SLinus Torvalds 3089fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 309051456b29SIan Morris if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 309191df42beSJoe Perches pr_debug("%s: wrong queue state\n", __func__); 30921da177e4SLinus Torvalds return -EFAULT; 30931da177e4SLinus Torvalds } 30944de075e0SEric Dumazet if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { 30951da177e4SLinus Torvalds if (skb_cloned(skb)) { 3096*e2080072SEric Dumazet struct sk_buff *nskb; 3097*e2080072SEric Dumazet 3098*e2080072SEric Dumazet tcp_skb_tsorted_save(skb) { 3099*e2080072SEric Dumazet nskb = skb_copy(skb, GFP_ATOMIC); 3100*e2080072SEric Dumazet } tcp_skb_tsorted_restore(skb); 310151456b29SIan Morris if (!nskb) 31021da177e4SLinus Torvalds return -ENOMEM; 3103*e2080072SEric Dumazet INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor); 3104fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 3105f4a775d1SEric Dumazet __skb_header_release(nskb); 3106fe067e8aSDavid S. Miller __tcp_add_write_queue_head(sk, nskb); 31073ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 31083ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 31093ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 31101da177e4SLinus Torvalds skb = nskb; 31111da177e4SLinus Torvalds } 31121da177e4SLinus Torvalds 31134de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; 3114735d3831SFlorian Westphal tcp_ecn_send_synack(sk, skb); 31151da177e4SLinus Torvalds } 3116dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 31171da177e4SLinus Torvalds } 31181da177e4SLinus Torvalds 31194aea39c1SEric Dumazet /** 31204aea39c1SEric Dumazet * tcp_make_synack - Prepare a SYN-ACK. 31214aea39c1SEric Dumazet * sk: listener socket 31224aea39c1SEric Dumazet * dst: dst entry attached to the SYNACK 31234aea39c1SEric Dumazet * req: request_sock pointer 31244aea39c1SEric Dumazet * 31254aea39c1SEric Dumazet * Allocate one skb and build a SYNACK packet. 31264aea39c1SEric Dumazet * @dst is consumed : Caller should not use it again. 31274aea39c1SEric Dumazet */ 31285d062de7SEric Dumazet struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst, 3129e6b4d113SWilliam Allen Simpson struct request_sock *req, 3130ca6fb065SEric Dumazet struct tcp_fastopen_cookie *foc, 3131b3d05147SEric Dumazet enum tcp_synack_type synack_type) 31321da177e4SLinus Torvalds { 31332e6599cbSArnaldo Carvalho de Melo struct inet_request_sock *ireq = inet_rsk(req); 31345d062de7SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 313580f03e27SEric Dumazet struct tcp_md5sig_key *md5 = NULL; 31365d062de7SEric Dumazet struct tcp_out_options opts; 31375d062de7SEric Dumazet struct sk_buff *skb; 3138bd0388aeSWilliam Allen Simpson int tcp_header_size; 31395d062de7SEric Dumazet struct tcphdr *th; 3140f5fff5dcSTom Quetchenbach int mss; 31411da177e4SLinus Torvalds 3142ca6fb065SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 31434aea39c1SEric Dumazet if (unlikely(!skb)) { 31444aea39c1SEric Dumazet dst_release(dst); 31451da177e4SLinus Torvalds return NULL; 31464aea39c1SEric Dumazet } 31471da177e4SLinus Torvalds /* Reserve space for headers. */ 31481da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 31491da177e4SLinus Torvalds 3150b3d05147SEric Dumazet switch (synack_type) { 3151b3d05147SEric Dumazet case TCP_SYNACK_NORMAL: 31529e17f8a4SEric Dumazet skb_set_owner_w(skb, req_to_sk(req)); 3153b3d05147SEric Dumazet break; 3154b3d05147SEric Dumazet case TCP_SYNACK_COOKIE: 3155b3d05147SEric Dumazet /* Under synflood, we do not attach skb to a socket, 3156b3d05147SEric Dumazet * to avoid false sharing. 3157b3d05147SEric Dumazet */ 3158b3d05147SEric Dumazet break; 3159b3d05147SEric Dumazet case TCP_SYNACK_FASTOPEN: 3160ca6fb065SEric Dumazet /* sk is a const pointer, because we want to express multiple 3161ca6fb065SEric Dumazet * cpu might call us concurrently. 3162ca6fb065SEric Dumazet * sk->sk_wmem_alloc in an atomic, we can promote to rw. 3163ca6fb065SEric Dumazet */ 3164ca6fb065SEric Dumazet skb_set_owner_w(skb, (struct sock *)sk); 3165b3d05147SEric Dumazet break; 3166ca6fb065SEric Dumazet } 31674aea39c1SEric Dumazet skb_dst_set(skb, dst); 31681da177e4SLinus Torvalds 31693541f9e8SEric Dumazet mss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 3170f5fff5dcSTom Quetchenbach 317133ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 31728b5f12d0SFlorian Westphal #ifdef CONFIG_SYN_COOKIES 31738b5f12d0SFlorian Westphal if (unlikely(req->cookie_ts)) 31749a568de4SEric Dumazet skb->skb_mstamp = cookie_init_timestamp(req); 31758b5f12d0SFlorian Westphal else 31768b5f12d0SFlorian Westphal #endif 31779a568de4SEric Dumazet skb->skb_mstamp = tcp_clock_us(); 317880f03e27SEric Dumazet 317980f03e27SEric Dumazet #ifdef CONFIG_TCP_MD5SIG 318080f03e27SEric Dumazet rcu_read_lock(); 3181fd3a154aSEric Dumazet md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req)); 318280f03e27SEric Dumazet #endif 318358d607d3SEric Dumazet skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4); 318437bfbddaSEric Dumazet tcp_header_size = tcp_synack_options(req, mss, skb, &opts, md5, foc) + 318537bfbddaSEric Dumazet sizeof(*th); 318633ad798cSAdam Langley 3187aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 3188aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 31891da177e4SLinus Torvalds 3190ea1627c2SEric Dumazet th = (struct tcphdr *)skb->data; 31911da177e4SLinus Torvalds memset(th, 0, sizeof(struct tcphdr)); 31921da177e4SLinus Torvalds th->syn = 1; 31931da177e4SLinus Torvalds th->ack = 1; 31946ac705b1SEric Dumazet tcp_ecn_make_synack(req, th); 3195b44084c2SEric Dumazet th->source = htons(ireq->ir_num); 3196634fb979SEric Dumazet th->dest = ireq->ir_rmt_port; 3197e05a90ecSJamal Hadi Salim skb->mark = ireq->ir_mark; 3198e870a8efSIlpo Järvinen /* Setting of flags are superfluous here for callers (and ECE is 3199e870a8efSIlpo Järvinen * not even correctly set) 3200e870a8efSIlpo Järvinen */ 3201e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn, 3202a3433f35SChangli Gao TCPHDR_SYN | TCPHDR_ACK); 32034957faadSWilliam Allen Simpson 32041da177e4SLinus Torvalds th->seq = htonl(TCP_SKB_CB(skb)->seq); 32058336886fSJerry Chu /* XXX data is queued and acked as is. No buffer/window check */ 32068336886fSJerry Chu th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt); 32071da177e4SLinus Torvalds 32081da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 3209ed53d0abSEric Dumazet th->window = htons(min(req->rsk_rcv_wnd, 65535U)); 32105d062de7SEric Dumazet tcp_options_write((__be32 *)(th + 1), NULL, &opts); 32111da177e4SLinus Torvalds th->doff = (tcp_header_size >> 2); 321290bbcc60SEric Dumazet __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS); 3213cfb6eeb4SYOSHIFUJI Hideaki 3214cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 3215cfb6eeb4SYOSHIFUJI Hideaki /* Okay, we have all we need - do the md5 hash if needed */ 321680f03e27SEric Dumazet if (md5) 3217bd0388aeSWilliam Allen Simpson tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, 321839f8e58eSEric Dumazet md5, req_to_sk(req), skb); 321980f03e27SEric Dumazet rcu_read_unlock(); 3220cfb6eeb4SYOSHIFUJI Hideaki #endif 3221cfb6eeb4SYOSHIFUJI Hideaki 3222b50edd78SEric Dumazet /* Do not fool tcpdump (if any), clean our debris */ 32232456e855SThomas Gleixner skb->tstamp = 0; 32241da177e4SLinus Torvalds return skb; 32251da177e4SLinus Torvalds } 32264bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_make_synack); 32271da177e4SLinus Torvalds 322881164413SDaniel Borkmann static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst) 322981164413SDaniel Borkmann { 323081164413SDaniel Borkmann struct inet_connection_sock *icsk = inet_csk(sk); 323181164413SDaniel Borkmann const struct tcp_congestion_ops *ca; 323281164413SDaniel Borkmann u32 ca_key = dst_metric(dst, RTAX_CC_ALGO); 323381164413SDaniel Borkmann 323481164413SDaniel Borkmann if (ca_key == TCP_CA_UNSPEC) 323581164413SDaniel Borkmann return; 323681164413SDaniel Borkmann 323781164413SDaniel Borkmann rcu_read_lock(); 323881164413SDaniel Borkmann ca = tcp_ca_find_key(ca_key); 323981164413SDaniel Borkmann if (likely(ca && try_module_get(ca->owner))) { 324081164413SDaniel Borkmann module_put(icsk->icsk_ca_ops->owner); 324181164413SDaniel Borkmann icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst); 324281164413SDaniel Borkmann icsk->icsk_ca_ops = ca; 324381164413SDaniel Borkmann } 324481164413SDaniel Borkmann rcu_read_unlock(); 324581164413SDaniel Borkmann } 324681164413SDaniel Borkmann 324767edfef7SAndi Kleen /* Do all connect socket setups that can be done AF independent. */ 3248f7e56a76Sstephen hemminger static void tcp_connect_init(struct sock *sk) 32491da177e4SLinus Torvalds { 3250cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 32511da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 32521da177e4SLinus Torvalds __u8 rcv_wscale; 325313d3b1ebSLawrence Brakmo u32 rcv_wnd; 32541da177e4SLinus Torvalds 32551da177e4SLinus Torvalds /* We'll fix this up when we get a response from the other end. 32561da177e4SLinus Torvalds * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 32571da177e4SLinus Torvalds */ 32585d2ed052SEric Dumazet tp->tcp_header_len = sizeof(struct tcphdr); 32595d2ed052SEric Dumazet if (sock_net(sk)->ipv4.sysctl_tcp_timestamps) 32605d2ed052SEric Dumazet tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED; 32611da177e4SLinus Torvalds 3262cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 326300db4124SIan Morris if (tp->af_specific->md5_lookup(sk, sk)) 3264cfb6eeb4SYOSHIFUJI Hideaki tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 3265cfb6eeb4SYOSHIFUJI Hideaki #endif 3266cfb6eeb4SYOSHIFUJI Hideaki 32671da177e4SLinus Torvalds /* If user gave his TCP_MAXSEG, record it to clamp */ 32681da177e4SLinus Torvalds if (tp->rx_opt.user_mss) 32691da177e4SLinus Torvalds tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 32701da177e4SLinus Torvalds tp->max_window = 0; 32715d424d5aSJohn Heffner tcp_mtup_init(sk); 32721da177e4SLinus Torvalds tcp_sync_mss(sk, dst_mtu(dst)); 32731da177e4SLinus Torvalds 327481164413SDaniel Borkmann tcp_ca_dst_init(sk, dst); 327581164413SDaniel Borkmann 32761da177e4SLinus Torvalds if (!tp->window_clamp) 32771da177e4SLinus Torvalds tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 32783541f9e8SEric Dumazet tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 3279f5fff5dcSTom Quetchenbach 32801da177e4SLinus Torvalds tcp_initialize_rcv_mss(sk); 32811da177e4SLinus Torvalds 3282e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 3283e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 3284e88c64f0SHagen Paul Pfeifer (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) 3285e88c64f0SHagen Paul Pfeifer tp->window_clamp = tcp_full_space(sk); 3286e88c64f0SHagen Paul Pfeifer 328713d3b1ebSLawrence Brakmo rcv_wnd = tcp_rwnd_init_bpf(sk); 328813d3b1ebSLawrence Brakmo if (rcv_wnd == 0) 328913d3b1ebSLawrence Brakmo rcv_wnd = dst_metric(dst, RTAX_INITRWND); 329013d3b1ebSLawrence Brakmo 32911da177e4SLinus Torvalds tcp_select_initial_window(tcp_full_space(sk), 32921da177e4SLinus Torvalds tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 32931da177e4SLinus Torvalds &tp->rcv_wnd, 32941da177e4SLinus Torvalds &tp->window_clamp, 32959bb37ef0SEric Dumazet sock_net(sk)->ipv4.sysctl_tcp_window_scaling, 329631d12926Slaurent chavey &rcv_wscale, 329713d3b1ebSLawrence Brakmo rcv_wnd); 32981da177e4SLinus Torvalds 32991da177e4SLinus Torvalds tp->rx_opt.rcv_wscale = rcv_wscale; 33001da177e4SLinus Torvalds tp->rcv_ssthresh = tp->rcv_wnd; 33011da177e4SLinus Torvalds 33021da177e4SLinus Torvalds sk->sk_err = 0; 33031da177e4SLinus Torvalds sock_reset_flag(sk, SOCK_DONE); 33041da177e4SLinus Torvalds tp->snd_wnd = 0; 3305ee7537b6SHantzis Fotis tcp_init_wl(tp, 0); 33061da177e4SLinus Torvalds tp->snd_una = tp->write_seq; 33071da177e4SLinus Torvalds tp->snd_sml = tp->write_seq; 330833f5f57eSIlpo Järvinen tp->snd_up = tp->write_seq; 3309370816aeSPavel Emelyanov tp->snd_nxt = tp->write_seq; 3310ee995283SPavel Emelyanov 3311ee995283SPavel Emelyanov if (likely(!tp->repair)) 33121da177e4SLinus Torvalds tp->rcv_nxt = 0; 3313c7781a6eSAndrew Vagin else 331470eabf0eSEric Dumazet tp->rcv_tstamp = tcp_jiffies32; 3315ee995283SPavel Emelyanov tp->rcv_wup = tp->rcv_nxt; 3316ee995283SPavel Emelyanov tp->copied_seq = tp->rcv_nxt; 33171da177e4SLinus Torvalds 33188550f328SLawrence Brakmo inet_csk(sk)->icsk_rto = tcp_timeout_init(sk); 3319463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_retransmits = 0; 33201da177e4SLinus Torvalds tcp_clear_retrans(tp); 33211da177e4SLinus Torvalds } 33221da177e4SLinus Torvalds 3323783237e8SYuchung Cheng static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) 3324783237e8SYuchung Cheng { 3325783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3326783237e8SYuchung Cheng struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 3327783237e8SYuchung Cheng 3328783237e8SYuchung Cheng tcb->end_seq += skb->len; 3329f4a775d1SEric Dumazet __skb_header_release(skb); 3330783237e8SYuchung Cheng __tcp_add_write_queue_tail(sk, skb); 3331783237e8SYuchung Cheng sk->sk_wmem_queued += skb->truesize; 3332783237e8SYuchung Cheng sk_mem_charge(sk, skb->truesize); 3333783237e8SYuchung Cheng tp->write_seq = tcb->end_seq; 3334783237e8SYuchung Cheng tp->packets_out += tcp_skb_pcount(skb); 3335783237e8SYuchung Cheng } 3336783237e8SYuchung Cheng 3337783237e8SYuchung Cheng /* Build and send a SYN with data and (cached) Fast Open cookie. However, 3338783237e8SYuchung Cheng * queue a data-only packet after the regular SYN, such that regular SYNs 3339783237e8SYuchung Cheng * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges 3340783237e8SYuchung Cheng * only the SYN sequence, the data are retransmitted in the first ACK. 3341783237e8SYuchung Cheng * If cookie is not cached or other error occurs, falls back to send a 3342783237e8SYuchung Cheng * regular SYN with Fast Open cookie request option. 3343783237e8SYuchung Cheng */ 3344783237e8SYuchung Cheng static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) 3345783237e8SYuchung Cheng { 3346783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3347783237e8SYuchung Cheng struct tcp_fastopen_request *fo = tp->fastopen_req; 3348065263f4SWei Wang int space, err = 0; 3349355a901eSEric Dumazet struct sk_buff *syn_data; 3350783237e8SYuchung Cheng 335167da22d2SYuchung Cheng tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ 3352065263f4SWei Wang if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie)) 3353783237e8SYuchung Cheng goto fallback; 3354783237e8SYuchung Cheng 3355783237e8SYuchung Cheng /* MSS for SYN-data is based on cached MSS and bounded by PMTU and 3356783237e8SYuchung Cheng * user-MSS. Reserve maximum option space for middleboxes that add 3357783237e8SYuchung Cheng * private TCP options. The cost is reduced data space in SYN :( 3358783237e8SYuchung Cheng */ 33593541f9e8SEric Dumazet tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp); 33603541f9e8SEric Dumazet 33611b63edd6SYuchung Cheng space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) - 3362783237e8SYuchung Cheng MAX_TCP_OPTION_SPACE; 3363783237e8SYuchung Cheng 3364f5ddcbbbSEric Dumazet space = min_t(size_t, space, fo->size); 3365f5ddcbbbSEric Dumazet 3366f5ddcbbbSEric Dumazet /* limit to order-0 allocations */ 3367f5ddcbbbSEric Dumazet space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER)); 3368f5ddcbbbSEric Dumazet 3369eb934478SEric Dumazet syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false); 3370355a901eSEric Dumazet if (!syn_data) 3371783237e8SYuchung Cheng goto fallback; 3372355a901eSEric Dumazet syn_data->ip_summed = CHECKSUM_PARTIAL; 3373355a901eSEric Dumazet memcpy(syn_data->cb, syn->cb, sizeof(syn->cb)); 337407e100f9SEric Dumazet if (space) { 337507e100f9SEric Dumazet int copied = copy_from_iter(skb_put(syn_data, space), space, 337657be5bdaSAl Viro &fo->data->msg_iter); 337757be5bdaSAl Viro if (unlikely(!copied)) { 3378355a901eSEric Dumazet kfree_skb(syn_data); 3379783237e8SYuchung Cheng goto fallback; 3380783237e8SYuchung Cheng } 338157be5bdaSAl Viro if (copied != space) { 338257be5bdaSAl Viro skb_trim(syn_data, copied); 338357be5bdaSAl Viro space = copied; 338457be5bdaSAl Viro } 338507e100f9SEric Dumazet } 3386355a901eSEric Dumazet /* No more data pending in inet_wait_for_connect() */ 3387355a901eSEric Dumazet if (space == fo->size) 3388355a901eSEric Dumazet fo->data = NULL; 3389355a901eSEric Dumazet fo->copied = space; 3390783237e8SYuchung Cheng 3391355a901eSEric Dumazet tcp_connect_queue_skb(sk, syn_data); 33920f87230dSFrancis Yan if (syn_data->len) 33930f87230dSFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_BUSY); 3394355a901eSEric Dumazet 3395355a901eSEric Dumazet err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation); 3396355a901eSEric Dumazet 3397355a901eSEric Dumazet syn->skb_mstamp = syn_data->skb_mstamp; 3398355a901eSEric Dumazet 3399355a901eSEric Dumazet /* Now full SYN+DATA was cloned and sent (or not), 3400355a901eSEric Dumazet * remove the SYN from the original skb (syn_data) 3401355a901eSEric Dumazet * we keep in write queue in case of a retransmit, as we 3402355a901eSEric Dumazet * also have the SYN packet (with no data) in the same queue. 3403431a9124SEric Dumazet */ 3404355a901eSEric Dumazet TCP_SKB_CB(syn_data)->seq++; 3405355a901eSEric Dumazet TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH; 3406355a901eSEric Dumazet if (!err) { 340767da22d2SYuchung Cheng tp->syn_data = (fo->copied > 0); 3408f19c29e3SYuchung Cheng NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT); 3409783237e8SYuchung Cheng goto done; 3410783237e8SYuchung Cheng } 3411783237e8SYuchung Cheng 3412b5b7db8dSEric Dumazet /* data was not sent, this is our new send_head */ 3413b5b7db8dSEric Dumazet sk->sk_send_head = syn_data; 3414b5b7db8dSEric Dumazet tp->packets_out -= tcp_skb_pcount(syn_data); 3415b5b7db8dSEric Dumazet 3416783237e8SYuchung Cheng fallback: 3417783237e8SYuchung Cheng /* Send a regular SYN with Fast Open cookie request option */ 3418783237e8SYuchung Cheng if (fo->cookie.len > 0) 3419783237e8SYuchung Cheng fo->cookie.len = 0; 3420783237e8SYuchung Cheng err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); 3421783237e8SYuchung Cheng if (err) 3422783237e8SYuchung Cheng tp->syn_fastopen = 0; 3423783237e8SYuchung Cheng done: 3424783237e8SYuchung Cheng fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ 3425783237e8SYuchung Cheng return err; 3426783237e8SYuchung Cheng } 3427783237e8SYuchung Cheng 342867edfef7SAndi Kleen /* Build a SYN and send it off. */ 34291da177e4SLinus Torvalds int tcp_connect(struct sock *sk) 34301da177e4SLinus Torvalds { 34311da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 34321da177e4SLinus Torvalds struct sk_buff *buff; 3433ee586811SEric Paris int err; 34341da177e4SLinus Torvalds 34359872a4bdSLawrence Brakmo tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB); 34368ba60924SEric Dumazet 34378ba60924SEric Dumazet if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 34388ba60924SEric Dumazet return -EHOSTUNREACH; /* Routing failure or similar. */ 34398ba60924SEric Dumazet 34401da177e4SLinus Torvalds tcp_connect_init(sk); 34411da177e4SLinus Torvalds 34422b916477SAndrey Vagin if (unlikely(tp->repair)) { 34432b916477SAndrey Vagin tcp_finish_connect(sk, NULL); 34442b916477SAndrey Vagin return 0; 34452b916477SAndrey Vagin } 34462b916477SAndrey Vagin 3447eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true); 3448355a901eSEric Dumazet if (unlikely(!buff)) 34491da177e4SLinus Torvalds return -ENOBUFS; 34501da177e4SLinus Torvalds 3451a3433f35SChangli Gao tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); 34529a568de4SEric Dumazet tcp_mstamp_refresh(tp); 34539a568de4SEric Dumazet tp->retrans_stamp = tcp_time_stamp(tp); 3454783237e8SYuchung Cheng tcp_connect_queue_skb(sk, buff); 3455735d3831SFlorian Westphal tcp_ecn_send_syn(sk, buff); 34561da177e4SLinus Torvalds 3457783237e8SYuchung Cheng /* Send off SYN; include data in Fast Open. */ 3458783237e8SYuchung Cheng err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : 3459783237e8SYuchung Cheng tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); 3460ee586811SEric Paris if (err == -ECONNREFUSED) 3461ee586811SEric Paris return err; 3462bd37a088SWei Yongjun 3463bd37a088SWei Yongjun /* We change tp->snd_nxt after the tcp_transmit_skb() call 3464bd37a088SWei Yongjun * in order to make this packet get counted in tcpOutSegs. 3465bd37a088SWei Yongjun */ 3466bd37a088SWei Yongjun tp->snd_nxt = tp->write_seq; 3467bd37a088SWei Yongjun tp->pushed_seq = tp->write_seq; 3468b5b7db8dSEric Dumazet buff = tcp_send_head(sk); 3469b5b7db8dSEric Dumazet if (unlikely(buff)) { 3470b5b7db8dSEric Dumazet tp->snd_nxt = TCP_SKB_CB(buff)->seq; 3471b5b7db8dSEric Dumazet tp->pushed_seq = TCP_SKB_CB(buff)->seq; 3472b5b7db8dSEric Dumazet } 347381cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); 34741da177e4SLinus Torvalds 34751da177e4SLinus Torvalds /* Timer for repeating the SYN until an answer. */ 34763f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 34773f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 34781da177e4SLinus Torvalds return 0; 34791da177e4SLinus Torvalds } 34804bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_connect); 34811da177e4SLinus Torvalds 34821da177e4SLinus Torvalds /* Send out a delayed ack, the caller does the policy checking 34831da177e4SLinus Torvalds * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 34841da177e4SLinus Torvalds * for details. 34851da177e4SLinus Torvalds */ 34861da177e4SLinus Torvalds void tcp_send_delayed_ack(struct sock *sk) 34871da177e4SLinus Torvalds { 3488463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 3489463c84b9SArnaldo Carvalho de Melo int ato = icsk->icsk_ack.ato; 34901da177e4SLinus Torvalds unsigned long timeout; 34911da177e4SLinus Torvalds 34929890092eSFlorian Westphal tcp_ca_event(sk, CA_EVENT_DELAYED_ACK); 34939890092eSFlorian Westphal 34941da177e4SLinus Torvalds if (ato > TCP_DELACK_MIN) { 3495463c84b9SArnaldo Carvalho de Melo const struct tcp_sock *tp = tcp_sk(sk); 34961da177e4SLinus Torvalds int max_ato = HZ / 2; 34971da177e4SLinus Torvalds 3498056834d9SIlpo Järvinen if (icsk->icsk_ack.pingpong || 3499056834d9SIlpo Järvinen (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 35001da177e4SLinus Torvalds max_ato = TCP_DELACK_MAX; 35011da177e4SLinus Torvalds 35021da177e4SLinus Torvalds /* Slow path, intersegment interval is "high". */ 35031da177e4SLinus Torvalds 35041da177e4SLinus Torvalds /* If some rtt estimate is known, use it to bound delayed ack. 3505463c84b9SArnaldo Carvalho de Melo * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 35061da177e4SLinus Torvalds * directly. 35071da177e4SLinus Torvalds */ 3508740b0f18SEric Dumazet if (tp->srtt_us) { 3509740b0f18SEric Dumazet int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3), 3510740b0f18SEric Dumazet TCP_DELACK_MIN); 35111da177e4SLinus Torvalds 35121da177e4SLinus Torvalds if (rtt < max_ato) 35131da177e4SLinus Torvalds max_ato = rtt; 35141da177e4SLinus Torvalds } 35151da177e4SLinus Torvalds 35161da177e4SLinus Torvalds ato = min(ato, max_ato); 35171da177e4SLinus Torvalds } 35181da177e4SLinus Torvalds 35191da177e4SLinus Torvalds /* Stay within the limit we were given */ 35201da177e4SLinus Torvalds timeout = jiffies + ato; 35211da177e4SLinus Torvalds 35221da177e4SLinus Torvalds /* Use new timeout only if there wasn't a older one earlier. */ 3523463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 35241da177e4SLinus Torvalds /* If delack timer was blocked or is about to expire, 35251da177e4SLinus Torvalds * send ACK now. 35261da177e4SLinus Torvalds */ 3527463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.blocked || 3528463c84b9SArnaldo Carvalho de Melo time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 35291da177e4SLinus Torvalds tcp_send_ack(sk); 35301da177e4SLinus Torvalds return; 35311da177e4SLinus Torvalds } 35321da177e4SLinus Torvalds 3533463c84b9SArnaldo Carvalho de Melo if (!time_before(timeout, icsk->icsk_ack.timeout)) 3534463c84b9SArnaldo Carvalho de Melo timeout = icsk->icsk_ack.timeout; 35351da177e4SLinus Torvalds } 3536463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 3537463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.timeout = timeout; 3538463c84b9SArnaldo Carvalho de Melo sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 35391da177e4SLinus Torvalds } 35401da177e4SLinus Torvalds 35411da177e4SLinus Torvalds /* This routine sends an ack and also updates the window. */ 35421da177e4SLinus Torvalds void tcp_send_ack(struct sock *sk) 35431da177e4SLinus Torvalds { 35441da177e4SLinus Torvalds struct sk_buff *buff; 35451da177e4SLinus Torvalds 3546058dc334SIlpo Järvinen /* If we have been reset, we may not send again. */ 3547058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3548058dc334SIlpo Järvinen return; 3549058dc334SIlpo Järvinen 35509890092eSFlorian Westphal tcp_ca_event(sk, CA_EVENT_NON_DELAYED_ACK); 35519890092eSFlorian Westphal 35521da177e4SLinus Torvalds /* We are not putting this on the write queue, so 35531da177e4SLinus Torvalds * tcp_transmit_skb() will set the ownership to this 35541da177e4SLinus Torvalds * sock. 35551da177e4SLinus Torvalds */ 35567450aaf6SEric Dumazet buff = alloc_skb(MAX_TCP_HEADER, 35577450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 35587450aaf6SEric Dumazet if (unlikely(!buff)) { 3559463c84b9SArnaldo Carvalho de Melo inet_csk_schedule_ack(sk); 3560463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; 35613f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 35623f421baaSArnaldo Carvalho de Melo TCP_DELACK_MAX, TCP_RTO_MAX); 35631da177e4SLinus Torvalds return; 35641da177e4SLinus Torvalds } 35651da177e4SLinus Torvalds 35661da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 35671da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 3568a3433f35SChangli Gao tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); 35691da177e4SLinus Torvalds 357098781965SEric Dumazet /* We do not want pure acks influencing TCP Small Queues or fq/pacing 357198781965SEric Dumazet * too much. 357298781965SEric Dumazet * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784 357398781965SEric Dumazet */ 357498781965SEric Dumazet skb_set_tcp_pure_ack(buff); 357598781965SEric Dumazet 35761da177e4SLinus Torvalds /* Send it off, this clears delayed acks for us. */ 35777450aaf6SEric Dumazet tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0); 35781da177e4SLinus Torvalds } 3579e3118e83SDaniel Borkmann EXPORT_SYMBOL_GPL(tcp_send_ack); 35801da177e4SLinus Torvalds 35811da177e4SLinus Torvalds /* This routine sends a packet with an out of date sequence 35821da177e4SLinus Torvalds * number. It assumes the other end will try to ack it. 35831da177e4SLinus Torvalds * 35841da177e4SLinus Torvalds * Question: what should we make while urgent mode? 35851da177e4SLinus Torvalds * 4.4BSD forces sending single byte of data. We cannot send 35861da177e4SLinus Torvalds * out of window data, because we have SND.NXT==SND.MAX... 35871da177e4SLinus Torvalds * 35881da177e4SLinus Torvalds * Current solution: to send TWO zero-length segments in urgent mode: 35891da177e4SLinus Torvalds * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 35901da177e4SLinus Torvalds * out-of-date with SND.UNA-1 to probe window. 35911da177e4SLinus Torvalds */ 3592e520af48SEric Dumazet static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib) 35931da177e4SLinus Torvalds { 35941da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 35951da177e4SLinus Torvalds struct sk_buff *skb; 35961da177e4SLinus Torvalds 35971da177e4SLinus Torvalds /* We don't queue it, tcp_transmit_skb() sets ownership. */ 35987450aaf6SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, 35997450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 360051456b29SIan Morris if (!skb) 36011da177e4SLinus Torvalds return -1; 36021da177e4SLinus Torvalds 36031da177e4SLinus Torvalds /* Reserve space for headers and set control bits. */ 36041da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 36051da177e4SLinus Torvalds /* Use a previous sequence. This should cause the other 36061da177e4SLinus Torvalds * end to send an ack. Don't queue or clone SKB, just 36071da177e4SLinus Torvalds * send it. 36081da177e4SLinus Torvalds */ 3609a3433f35SChangli Gao tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); 3610e2e8009fSRenato Westphal NET_INC_STATS(sock_net(sk), mib); 36117450aaf6SEric Dumazet return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0); 36121da177e4SLinus Torvalds } 36131da177e4SLinus Torvalds 3614385e2070SEric Dumazet /* Called from setsockopt( ... TCP_REPAIR ) */ 3615ee995283SPavel Emelyanov void tcp_send_window_probe(struct sock *sk) 3616ee995283SPavel Emelyanov { 3617ee995283SPavel Emelyanov if (sk->sk_state == TCP_ESTABLISHED) { 3618ee995283SPavel Emelyanov tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; 36199a568de4SEric Dumazet tcp_mstamp_refresh(tcp_sk(sk)); 3620e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE); 3621ee995283SPavel Emelyanov } 3622ee995283SPavel Emelyanov } 3623ee995283SPavel Emelyanov 362467edfef7SAndi Kleen /* Initiate keepalive or window probe from timer. */ 3625e520af48SEric Dumazet int tcp_write_wakeup(struct sock *sk, int mib) 36261da177e4SLinus Torvalds { 36271da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 36281da177e4SLinus Torvalds struct sk_buff *skb; 36291da177e4SLinus Torvalds 3630058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3631058dc334SIlpo Järvinen return -1; 3632058dc334SIlpo Järvinen 363300db4124SIan Morris skb = tcp_send_head(sk); 363400db4124SIan Morris if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { 36351da177e4SLinus Torvalds int err; 36360c54b85fSIlpo Järvinen unsigned int mss = tcp_current_mss(sk); 363790840defSIlpo Järvinen unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 36381da177e4SLinus Torvalds 36391da177e4SLinus Torvalds if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 36401da177e4SLinus Torvalds tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 36411da177e4SLinus Torvalds 36421da177e4SLinus Torvalds /* We are probing the opening of a window 36431da177e4SLinus Torvalds * but the window size is != 0 36441da177e4SLinus Torvalds * must have been a result SWS avoidance ( sender ) 36451da177e4SLinus Torvalds */ 36461da177e4SLinus Torvalds if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 36471da177e4SLinus Torvalds skb->len > mss) { 36481da177e4SLinus Torvalds seg_size = min(seg_size, mss); 36494de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 36506cc55e09SOctavian Purdila if (tcp_fragment(sk, skb, seg_size, mss, GFP_ATOMIC)) 36511da177e4SLinus Torvalds return -1; 36521da177e4SLinus Torvalds } else if (!tcp_skb_pcount(skb)) 36535bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss); 36541da177e4SLinus Torvalds 36554de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 3656dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 365766f5fe62SIlpo Järvinen if (!err) 365866f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 36591da177e4SLinus Torvalds return err; 36601da177e4SLinus Torvalds } else { 366133f5f57eSIlpo Järvinen if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) 3662e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 1, mib); 3663e520af48SEric Dumazet return tcp_xmit_probe_skb(sk, 0, mib); 36641da177e4SLinus Torvalds } 36651da177e4SLinus Torvalds } 36661da177e4SLinus Torvalds 36671da177e4SLinus Torvalds /* A window probe timeout has occurred. If window is not closed send 36681da177e4SLinus Torvalds * a partial packet else a zero probe. 36691da177e4SLinus Torvalds */ 36701da177e4SLinus Torvalds void tcp_send_probe0(struct sock *sk) 36711da177e4SLinus Torvalds { 3672463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 36731da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 3674c6214a97SNikolay Borisov struct net *net = sock_net(sk); 3675fcdd1cf4SEric Dumazet unsigned long probe_max; 36761da177e4SLinus Torvalds int err; 36771da177e4SLinus Torvalds 3678e520af48SEric Dumazet err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE); 36791da177e4SLinus Torvalds 3680fe067e8aSDavid S. Miller if (tp->packets_out || !tcp_send_head(sk)) { 36811da177e4SLinus Torvalds /* Cancel probe timer, if it is not required. */ 36826687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 0; 3683463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff = 0; 36841da177e4SLinus Torvalds return; 36851da177e4SLinus Torvalds } 36861da177e4SLinus Torvalds 36871da177e4SLinus Torvalds if (err <= 0) { 3688c6214a97SNikolay Borisov if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2) 3689463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff++; 36906687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out++; 3691fcdd1cf4SEric Dumazet probe_max = TCP_RTO_MAX; 36921da177e4SLinus Torvalds } else { 36931da177e4SLinus Torvalds /* If packet was not sent due to local congestion, 36946687e988SArnaldo Carvalho de Melo * do not backoff and do not remember icsk_probes_out. 36951da177e4SLinus Torvalds * Let local senders to fight for local resources. 36961da177e4SLinus Torvalds * 36971da177e4SLinus Torvalds * Use accumulated backoff yet. 36981da177e4SLinus Torvalds */ 36996687e988SArnaldo Carvalho de Melo if (!icsk->icsk_probes_out) 37006687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 1; 3701fcdd1cf4SEric Dumazet probe_max = TCP_RESOURCE_PROBE_INTERVAL; 37021da177e4SLinus Torvalds } 3703fcdd1cf4SEric Dumazet inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 370421c8fe99SEric Dumazet tcp_probe0_when(sk, probe_max), 3705fcdd1cf4SEric Dumazet TCP_RTO_MAX); 37061da177e4SLinus Torvalds } 37075db92c99SOctavian Purdila 3708ea3bea3aSEric Dumazet int tcp_rtx_synack(const struct sock *sk, struct request_sock *req) 37095db92c99SOctavian Purdila { 37105db92c99SOctavian Purdila const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific; 37115db92c99SOctavian Purdila struct flowi fl; 37125db92c99SOctavian Purdila int res; 37135db92c99SOctavian Purdila 371458d607d3SEric Dumazet tcp_rsk(req)->txhash = net_tx_rndhash(); 3715b3d05147SEric Dumazet res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL); 37165db92c99SOctavian Purdila if (!res) { 371790bbcc60SEric Dumazet __TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); 371802a1d6e7SEric Dumazet __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 37197e32b443SYuchung Cheng if (unlikely(tcp_passive_fastopen(sk))) 37207e32b443SYuchung Cheng tcp_sk(sk)->total_retrans++; 37215db92c99SOctavian Purdila } 37225db92c99SOctavian Purdila return res; 37235db92c99SOctavian Purdila } 37245db92c99SOctavian Purdila EXPORT_SYMBOL(tcp_rtx_synack); 3725