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> 4460e2a778SUrsula Braun #include <linux/static_key.h> 451da177e4SLinus Torvalds 46e086101bSCong Wang #include <trace/events/tcp.h> 47e086101bSCong Wang 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 6246d3ceabSEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 6346d3ceabSEric Dumazet int push_one, gfp_t gfp); 64519855c5SWilliam Allen Simpson 6567edfef7SAndi Kleen /* Account for new data that has been sent to the network. */ 6675c119afSEric Dumazet static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb) 676ff03ac3SIlpo Järvinen { 686ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 696ff03ac3SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 7066f5fe62SIlpo Järvinen unsigned int prior_packets = tp->packets_out; 719e412ba7SIlpo Järvinen 721da177e4SLinus Torvalds tp->snd_nxt = TCP_SKB_CB(skb)->end_seq; 738512430eSIlpo Järvinen 7475c119afSEric Dumazet __skb_unlink(skb, &sk->sk_write_queue); 7575c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, skb); 7675c119afSEric Dumazet 7766f5fe62SIlpo Järvinen tp->packets_out += tcp_skb_pcount(skb); 78bec41a11SYuchung Cheng if (!prior_packets || icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) 79750ea2baSYuchung Cheng tcp_rearm_rto(sk); 80f19c29e3SYuchung Cheng 81f7324acdSDavid S. Miller NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT, 82f19c29e3SYuchung Cheng tcp_skb_pcount(skb)); 836a5dc9e5SEric Dumazet } 841da177e4SLinus Torvalds 85a4ecb15aSCui, Cheng /* SND.NXT, if window was not shrunk or the amount of shrunk was less than one 86a4ecb15aSCui, Cheng * window scaling factor due to loss of precision. 871da177e4SLinus Torvalds * If window has been shrunk, what should we make? It is not clear at all. 881da177e4SLinus Torvalds * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-( 891da177e4SLinus Torvalds * Anything in between SND.UNA...SND.UNA+SND.WND also can be already 901da177e4SLinus Torvalds * invalid. OK, let's make this for now: 911da177e4SLinus Torvalds */ 92cf533ea5SEric Dumazet static inline __u32 tcp_acceptable_seq(const struct sock *sk) 931da177e4SLinus Torvalds { 94cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 959e412ba7SIlpo Järvinen 96a4ecb15aSCui, Cheng if (!before(tcp_wnd_end(tp), tp->snd_nxt) || 97a4ecb15aSCui, Cheng (tp->rx_opt.wscale_ok && 98a4ecb15aSCui, Cheng ((tp->snd_nxt - tcp_wnd_end(tp)) < (1 << tp->rx_opt.rcv_wscale)))) 991da177e4SLinus Torvalds return tp->snd_nxt; 1001da177e4SLinus Torvalds else 10190840defSIlpo Järvinen return tcp_wnd_end(tp); 1021da177e4SLinus Torvalds } 1031da177e4SLinus Torvalds 1041da177e4SLinus Torvalds /* Calculate mss to advertise in SYN segment. 1051da177e4SLinus Torvalds * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that: 1061da177e4SLinus Torvalds * 1071da177e4SLinus Torvalds * 1. It is independent of path mtu. 1081da177e4SLinus Torvalds * 2. Ideally, it is maximal possible segment size i.e. 65535-40. 1091da177e4SLinus Torvalds * 3. For IPv4 it is reasonable to calculate it from maximal MTU of 1101da177e4SLinus Torvalds * attached devices, because some buggy hosts are confused by 1111da177e4SLinus Torvalds * large MSS. 1121da177e4SLinus Torvalds * 4. We do not make 3, we advertise MSS, calculated from first 1131da177e4SLinus Torvalds * hop device mtu, but allow to raise it to ip_rt_min_advmss. 1141da177e4SLinus Torvalds * This may be overridden via information stored in routing table. 1151da177e4SLinus Torvalds * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible, 1161da177e4SLinus Torvalds * probably even Jumbo". 1171da177e4SLinus Torvalds */ 1181da177e4SLinus Torvalds static __u16 tcp_advertise_mss(struct sock *sk) 1191da177e4SLinus Torvalds { 1201da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 121cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1221da177e4SLinus Torvalds int mss = tp->advmss; 1231da177e4SLinus Torvalds 1240dbaee3bSDavid S. Miller if (dst) { 1250dbaee3bSDavid S. Miller unsigned int metric = dst_metric_advmss(dst); 1260dbaee3bSDavid S. Miller 1270dbaee3bSDavid S. Miller if (metric < mss) { 1280dbaee3bSDavid S. Miller mss = metric; 1291da177e4SLinus Torvalds tp->advmss = mss; 1301da177e4SLinus Torvalds } 1310dbaee3bSDavid S. Miller } 1321da177e4SLinus Torvalds 1331da177e4SLinus Torvalds return (__u16)mss; 1341da177e4SLinus Torvalds } 1351da177e4SLinus Torvalds 1361da177e4SLinus Torvalds /* RFC2861. Reset CWND after idle period longer RTO to "restart window". 1376f021c62SEric Dumazet * This is the first part of cwnd validation mechanism. 1386f021c62SEric Dumazet */ 1396f021c62SEric Dumazet void tcp_cwnd_restart(struct sock *sk, s32 delta) 1401da177e4SLinus Torvalds { 141463c84b9SArnaldo Carvalho de Melo struct tcp_sock *tp = tcp_sk(sk); 1426f021c62SEric Dumazet u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk)); 1431da177e4SLinus Torvalds u32 cwnd = tp->snd_cwnd; 1441da177e4SLinus Torvalds 1456687e988SArnaldo Carvalho de Melo tcp_ca_event(sk, CA_EVENT_CWND_RESTART); 1461da177e4SLinus Torvalds 1476687e988SArnaldo Carvalho de Melo tp->snd_ssthresh = tcp_current_ssthresh(sk); 1481da177e4SLinus Torvalds restart_cwnd = min(restart_cwnd, cwnd); 1491da177e4SLinus Torvalds 150463c84b9SArnaldo Carvalho de Melo while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd) 1511da177e4SLinus Torvalds cwnd >>= 1; 1521da177e4SLinus Torvalds tp->snd_cwnd = max(cwnd, restart_cwnd); 153c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 1541da177e4SLinus Torvalds tp->snd_cwnd_used = 0; 1551da177e4SLinus Torvalds } 1561da177e4SLinus Torvalds 15767edfef7SAndi Kleen /* Congestion state accounting after a packet has been sent. */ 15840efc6faSStephen Hemminger static void tcp_event_data_sent(struct tcp_sock *tp, 159cf533ea5SEric Dumazet struct sock *sk) 1601da177e4SLinus Torvalds { 161463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 162d635fbe2SEric Dumazet const u32 now = tcp_jiffies32; 1631da177e4SLinus Torvalds 16405c5a46dSNeal Cardwell if (tcp_packets_in_flight(tp) == 0) 16505c5a46dSNeal Cardwell tcp_ca_event(sk, CA_EVENT_TX_START); 16605c5a46dSNeal Cardwell 1671da177e4SLinus Torvalds tp->lsndtime = now; 1681da177e4SLinus Torvalds 1691da177e4SLinus Torvalds /* If it is a reply for ato after last received 1701da177e4SLinus Torvalds * packet, enter pingpong mode. 1711da177e4SLinus Torvalds */ 1722251ae46SJon Maxwell if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato) 173463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pingpong = 1; 1741da177e4SLinus Torvalds } 1751da177e4SLinus Torvalds 17667edfef7SAndi Kleen /* Account for an ACK we sent. */ 17740efc6faSStephen Hemminger static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts) 1781da177e4SLinus Torvalds { 179463c84b9SArnaldo Carvalho de Melo tcp_dec_quickack_mode(sk, pkts); 180463c84b9SArnaldo Carvalho de Melo inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); 1811da177e4SLinus Torvalds } 1821da177e4SLinus Torvalds 18385f16525SYuchung Cheng 18485f16525SYuchung Cheng u32 tcp_default_init_rwnd(u32 mss) 18585f16525SYuchung Cheng { 18685f16525SYuchung Cheng /* Initial receive window should be twice of TCP_INIT_CWND to 1879ef71e0cSWeiping Pan * enable proper sending of new unsent data during fast recovery 18885f16525SYuchung Cheng * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a 18985f16525SYuchung Cheng * limit when mss is larger than 1460. 19085f16525SYuchung Cheng */ 19185f16525SYuchung Cheng u32 init_rwnd = TCP_INIT_CWND * 2; 19285f16525SYuchung Cheng 19385f16525SYuchung Cheng if (mss > 1460) 19485f16525SYuchung Cheng init_rwnd = max((1460 * init_rwnd) / mss, 2U); 19585f16525SYuchung Cheng return init_rwnd; 19685f16525SYuchung Cheng } 19785f16525SYuchung Cheng 1981da177e4SLinus Torvalds /* Determine a window scaling and initial window to offer. 1991da177e4SLinus Torvalds * Based on the assumption that the given amount of space 2001da177e4SLinus Torvalds * will be offered. Store the results in the tp structure. 2011da177e4SLinus Torvalds * NOTE: for smooth operation initial space offering should 2021da177e4SLinus Torvalds * be a multiple of mss if possible. We assume here that mss >= 1. 2031da177e4SLinus Torvalds * This MUST be enforced by all callers. 2041da177e4SLinus Torvalds */ 2051da177e4SLinus Torvalds void tcp_select_initial_window(int __space, __u32 mss, 2061da177e4SLinus Torvalds __u32 *rcv_wnd, __u32 *window_clamp, 20731d12926Slaurent chavey int wscale_ok, __u8 *rcv_wscale, 20831d12926Slaurent chavey __u32 init_rcv_wnd) 2091da177e4SLinus Torvalds { 2101da177e4SLinus Torvalds unsigned int space = (__space < 0 ? 0 : __space); 2111da177e4SLinus Torvalds 2121da177e4SLinus Torvalds /* If no clamp set the clamp to the max possible scaled window */ 2131da177e4SLinus Torvalds if (*window_clamp == 0) 214589c49cbSGao Feng (*window_clamp) = (U16_MAX << TCP_MAX_WSCALE); 2151da177e4SLinus Torvalds space = min(*window_clamp, space); 2161da177e4SLinus Torvalds 2171da177e4SLinus Torvalds /* Quantize space offering to a multiple of mss if possible. */ 2181da177e4SLinus Torvalds if (space > mss) 219589c49cbSGao Feng space = rounddown(space, mss); 2201da177e4SLinus Torvalds 2211da177e4SLinus Torvalds /* NOTE: offering an initial window larger than 32767 22215d99e02SRick Jones * will break some buggy TCP stacks. If the admin tells us 22315d99e02SRick Jones * it is likely we could be speaking with such a buggy stack 22415d99e02SRick Jones * we will truncate our initial window offering to 32K-1 22515d99e02SRick Jones * unless the remote has sent us a window scaling option, 22615d99e02SRick Jones * which we interpret as a sign the remote TCP is not 22715d99e02SRick Jones * misinterpreting the window field as a signed quantity. 2281da177e4SLinus Torvalds */ 22915d99e02SRick Jones if (sysctl_tcp_workaround_signed_windows) 2301da177e4SLinus Torvalds (*rcv_wnd) = min(space, MAX_TCP_WINDOW); 23115d99e02SRick Jones else 23215d99e02SRick Jones (*rcv_wnd) = space; 23315d99e02SRick Jones 2341da177e4SLinus Torvalds (*rcv_wscale) = 0; 2351da177e4SLinus Torvalds if (wscale_ok) { 236589c49cbSGao Feng /* Set window scaling on max possible window */ 237f626300aSSoheil Hassas Yeganeh space = max_t(u32, space, sysctl_tcp_rmem[2]); 238f626300aSSoheil Hassas Yeganeh space = max_t(u32, space, sysctl_rmem_max); 239316c1592SStephen Hemminger space = min_t(u32, space, *window_clamp); 240589c49cbSGao Feng while (space > U16_MAX && (*rcv_wscale) < TCP_MAX_WSCALE) { 2411da177e4SLinus Torvalds space >>= 1; 2421da177e4SLinus Torvalds (*rcv_wscale)++; 2431da177e4SLinus Torvalds } 2441da177e4SLinus Torvalds } 2451da177e4SLinus Torvalds 2461da177e4SLinus Torvalds if (mss > (1 << *rcv_wscale)) { 24785f16525SYuchung Cheng if (!init_rcv_wnd) /* Use default unless specified otherwise */ 24885f16525SYuchung Cheng init_rcv_wnd = tcp_default_init_rwnd(mss); 249b1afde60SNandita Dukkipati *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss); 2501da177e4SLinus Torvalds } 2511da177e4SLinus Torvalds 2521da177e4SLinus Torvalds /* Set the clamp no higher than max representable value */ 253589c49cbSGao Feng (*window_clamp) = min_t(__u32, U16_MAX << (*rcv_wscale), *window_clamp); 2541da177e4SLinus Torvalds } 2554bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_select_initial_window); 2561da177e4SLinus Torvalds 2571da177e4SLinus Torvalds /* Chose a new window to advertise, update state in tcp_sock for the 2581da177e4SLinus Torvalds * socket, and return result with RFC1323 scaling applied. The return 2591da177e4SLinus Torvalds * value can be stuffed directly into th->window for an outgoing 2601da177e4SLinus Torvalds * frame. 2611da177e4SLinus Torvalds */ 26240efc6faSStephen Hemminger static u16 tcp_select_window(struct sock *sk) 2631da177e4SLinus Torvalds { 2641da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2658e165e20SFlorian Westphal u32 old_win = tp->rcv_wnd; 2661da177e4SLinus Torvalds u32 cur_win = tcp_receive_window(tp); 2671da177e4SLinus Torvalds u32 new_win = __tcp_select_window(sk); 2681da177e4SLinus Torvalds 2691da177e4SLinus Torvalds /* Never shrink the offered window */ 2701da177e4SLinus Torvalds if (new_win < cur_win) { 2711da177e4SLinus Torvalds /* Danger Will Robinson! 2721da177e4SLinus Torvalds * Don't update rcv_wup/rcv_wnd here or else 2731da177e4SLinus Torvalds * we will not be able to advertise a zero 2741da177e4SLinus Torvalds * window in time. --DaveM 2751da177e4SLinus Torvalds * 2761da177e4SLinus Torvalds * Relax Will Robinson. 2771da177e4SLinus Torvalds */ 2788e165e20SFlorian Westphal if (new_win == 0) 2798e165e20SFlorian Westphal NET_INC_STATS(sock_net(sk), 2808e165e20SFlorian Westphal LINUX_MIB_TCPWANTZEROWINDOWADV); 281607bfbf2SPatrick McHardy new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale); 2821da177e4SLinus Torvalds } 2831da177e4SLinus Torvalds tp->rcv_wnd = new_win; 2841da177e4SLinus Torvalds tp->rcv_wup = tp->rcv_nxt; 2851da177e4SLinus Torvalds 2861da177e4SLinus Torvalds /* Make sure we do not exceed the maximum possible 2871da177e4SLinus Torvalds * scaled window. 2881da177e4SLinus Torvalds */ 28915d99e02SRick Jones if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows) 2901da177e4SLinus Torvalds new_win = min(new_win, MAX_TCP_WINDOW); 2911da177e4SLinus Torvalds else 2921da177e4SLinus Torvalds new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); 2931da177e4SLinus Torvalds 2941da177e4SLinus Torvalds /* RFC1323 scaling applied */ 2951da177e4SLinus Torvalds new_win >>= tp->rx_opt.rcv_wscale; 2961da177e4SLinus Torvalds 29731770e34SFlorian Westphal /* If we advertise zero window, disable fast path. */ 2988e165e20SFlorian Westphal if (new_win == 0) { 29931770e34SFlorian Westphal tp->pred_flags = 0; 3008e165e20SFlorian Westphal if (old_win) 3018e165e20SFlorian Westphal NET_INC_STATS(sock_net(sk), 3028e165e20SFlorian Westphal LINUX_MIB_TCPTOZEROWINDOWADV); 3038e165e20SFlorian Westphal } else if (old_win == 0) { 3048e165e20SFlorian Westphal NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV); 3058e165e20SFlorian Westphal } 3061da177e4SLinus Torvalds 3071da177e4SLinus Torvalds return new_win; 3081da177e4SLinus Torvalds } 3091da177e4SLinus Torvalds 31067edfef7SAndi Kleen /* Packet ECN state for a SYN-ACK */ 311735d3831SFlorian Westphal static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb) 312bdf1ee5dSIlpo Järvinen { 31330e502a3SDaniel Borkmann const struct tcp_sock *tp = tcp_sk(sk); 31430e502a3SDaniel Borkmann 3154de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR; 316bdf1ee5dSIlpo Järvinen if (!(tp->ecn_flags & TCP_ECN_OK)) 3174de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE; 31891b5b21cSLawrence Brakmo else if (tcp_ca_needs_ecn(sk) || 31991b5b21cSLawrence Brakmo tcp_bpf_ca_needs_ecn(sk)) 32030e502a3SDaniel Borkmann INET_ECN_xmit(sk); 321bdf1ee5dSIlpo Järvinen } 322bdf1ee5dSIlpo Järvinen 32367edfef7SAndi Kleen /* Packet ECN state for a SYN. */ 324735d3831SFlorian Westphal static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb) 325bdf1ee5dSIlpo Järvinen { 326bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 32791b5b21cSLawrence Brakmo bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk); 328f7b3bec6SFlorian Westphal bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 || 32991b5b21cSLawrence Brakmo tcp_ca_needs_ecn(sk) || bpf_needs_ecn; 330f7b3bec6SFlorian Westphal 331f7b3bec6SFlorian Westphal if (!use_ecn) { 332f7b3bec6SFlorian Westphal const struct dst_entry *dst = __sk_dst_get(sk); 333f7b3bec6SFlorian Westphal 334f7b3bec6SFlorian Westphal if (dst && dst_feature(dst, RTAX_FEATURE_ECN)) 335f7b3bec6SFlorian Westphal use_ecn = true; 336f7b3bec6SFlorian Westphal } 337bdf1ee5dSIlpo Järvinen 338bdf1ee5dSIlpo Järvinen tp->ecn_flags = 0; 339f7b3bec6SFlorian Westphal 340f7b3bec6SFlorian Westphal if (use_ecn) { 3414de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR; 342bdf1ee5dSIlpo Järvinen tp->ecn_flags = TCP_ECN_OK; 34391b5b21cSLawrence Brakmo if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn) 34430e502a3SDaniel Borkmann INET_ECN_xmit(sk); 345bdf1ee5dSIlpo Järvinen } 346bdf1ee5dSIlpo Järvinen } 347bdf1ee5dSIlpo Järvinen 34849213555SDaniel Borkmann static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb) 34949213555SDaniel Borkmann { 35049213555SDaniel Borkmann if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback) 35149213555SDaniel Borkmann /* tp->ecn_flags are cleared at a later point in time when 35249213555SDaniel Borkmann * SYN ACK is ultimatively being received. 35349213555SDaniel Borkmann */ 35449213555SDaniel Borkmann TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR); 35549213555SDaniel Borkmann } 35649213555SDaniel Borkmann 357735d3831SFlorian Westphal static void 3586ac705b1SEric Dumazet tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th) 359bdf1ee5dSIlpo Järvinen { 3606ac705b1SEric Dumazet if (inet_rsk(req)->ecn_ok) 361bdf1ee5dSIlpo Järvinen th->ece = 1; 362bdf1ee5dSIlpo Järvinen } 363bdf1ee5dSIlpo Järvinen 36467edfef7SAndi Kleen /* Set up ECN state for a packet on a ESTABLISHED socket that is about to 36567edfef7SAndi Kleen * be sent. 36667edfef7SAndi Kleen */ 367735d3831SFlorian Westphal static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb, 368ea1627c2SEric Dumazet struct tcphdr *th, int tcp_header_len) 369bdf1ee5dSIlpo Järvinen { 370bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 371bdf1ee5dSIlpo Järvinen 372bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_OK) { 373bdf1ee5dSIlpo Järvinen /* Not-retransmitted data segment: set ECT and inject CWR. */ 374bdf1ee5dSIlpo Järvinen if (skb->len != tcp_header_len && 375bdf1ee5dSIlpo Järvinen !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { 376bdf1ee5dSIlpo Järvinen INET_ECN_xmit(sk); 377bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) { 378bdf1ee5dSIlpo Järvinen tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; 379ea1627c2SEric Dumazet th->cwr = 1; 380bdf1ee5dSIlpo Järvinen skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; 381bdf1ee5dSIlpo Järvinen } 38230e502a3SDaniel Borkmann } else if (!tcp_ca_needs_ecn(sk)) { 383bdf1ee5dSIlpo Järvinen /* ACK or retransmitted segment: clear ECT|CE */ 384bdf1ee5dSIlpo Järvinen INET_ECN_dontxmit(sk); 385bdf1ee5dSIlpo Järvinen } 386bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) 387ea1627c2SEric Dumazet th->ece = 1; 388bdf1ee5dSIlpo Järvinen } 389bdf1ee5dSIlpo Järvinen } 390bdf1ee5dSIlpo Järvinen 391e870a8efSIlpo Järvinen /* Constructs common control bits of non-data skb. If SYN/FIN is present, 392e870a8efSIlpo Järvinen * auto increment end seqno. 393e870a8efSIlpo Järvinen */ 394e870a8efSIlpo Järvinen static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags) 395e870a8efSIlpo Järvinen { 3962e8e18efSDavid S. Miller skb->ip_summed = CHECKSUM_PARTIAL; 397e870a8efSIlpo Järvinen skb->csum = 0; 398e870a8efSIlpo Järvinen 3994de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags; 400e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->sacked = 0; 401e870a8efSIlpo Järvinen 402cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, 1); 403e870a8efSIlpo Järvinen 404e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->seq = seq; 405a3433f35SChangli Gao if (flags & (TCPHDR_SYN | TCPHDR_FIN)) 406e870a8efSIlpo Järvinen seq++; 407e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->end_seq = seq; 408e870a8efSIlpo Järvinen } 409e870a8efSIlpo Järvinen 410a2a385d6SEric Dumazet static inline bool tcp_urg_mode(const struct tcp_sock *tp) 41133f5f57eSIlpo Järvinen { 41233f5f57eSIlpo Järvinen return tp->snd_una != tp->snd_up; 41333f5f57eSIlpo Järvinen } 41433f5f57eSIlpo Järvinen 41533ad798cSAdam Langley #define OPTION_SACK_ADVERTISE (1 << 0) 41633ad798cSAdam Langley #define OPTION_TS (1 << 1) 41733ad798cSAdam Langley #define OPTION_MD5 (1 << 2) 41889e95a61SOri Finkelman #define OPTION_WSCALE (1 << 3) 4192100c8d2SYuchung Cheng #define OPTION_FAST_OPEN_COOKIE (1 << 8) 42060e2a778SUrsula Braun #define OPTION_SMC (1 << 9) 42160e2a778SUrsula Braun 42260e2a778SUrsula Braun static void smc_options_write(__be32 *ptr, u16 *options) 42360e2a778SUrsula Braun { 42460e2a778SUrsula Braun #if IS_ENABLED(CONFIG_SMC) 42560e2a778SUrsula Braun if (static_branch_unlikely(&tcp_have_smc)) { 42660e2a778SUrsula Braun if (unlikely(OPTION_SMC & *options)) { 42760e2a778SUrsula Braun *ptr++ = htonl((TCPOPT_NOP << 24) | 42860e2a778SUrsula Braun (TCPOPT_NOP << 16) | 42960e2a778SUrsula Braun (TCPOPT_EXP << 8) | 43060e2a778SUrsula Braun (TCPOLEN_EXP_SMC_BASE)); 43160e2a778SUrsula Braun *ptr++ = htonl(TCPOPT_SMC_MAGIC); 43260e2a778SUrsula Braun } 43360e2a778SUrsula Braun } 43460e2a778SUrsula Braun #endif 43560e2a778SUrsula Braun } 43633ad798cSAdam Langley 43733ad798cSAdam Langley struct tcp_out_options { 4382100c8d2SYuchung Cheng u16 options; /* bit field of OPTION_* */ 4392100c8d2SYuchung Cheng u16 mss; /* 0 to disable */ 44033ad798cSAdam Langley u8 ws; /* window scale, 0 to disable */ 44133ad798cSAdam Langley u8 num_sack_blocks; /* number of SACK blocks to include */ 442bd0388aeSWilliam Allen Simpson u8 hash_size; /* bytes in hash_location */ 443bd0388aeSWilliam Allen Simpson __u8 *hash_location; /* temporary pointer, overloaded */ 4442100c8d2SYuchung Cheng __u32 tsval, tsecr; /* need to include OPTION_TS */ 4452100c8d2SYuchung Cheng struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */ 44633ad798cSAdam Langley }; 44733ad798cSAdam Langley 44867edfef7SAndi Kleen /* Write previously computed TCP options to the packet. 44967edfef7SAndi Kleen * 45067edfef7SAndi Kleen * Beware: Something in the Internet is very sensitive to the ordering of 451fd6149d3SIlpo Järvinen * TCP options, we learned this through the hard way, so be careful here. 452fd6149d3SIlpo Järvinen * Luckily we can at least blame others for their non-compliance but from 4538e3bff96Sstephen hemminger * inter-operability perspective it seems that we're somewhat stuck with 454fd6149d3SIlpo Järvinen * the ordering which we have been using if we want to keep working with 455fd6149d3SIlpo Järvinen * those broken things (not that it currently hurts anybody as there isn't 456fd6149d3SIlpo Järvinen * particular reason why the ordering would need to be changed). 457fd6149d3SIlpo Järvinen * 458fd6149d3SIlpo Järvinen * At least SACK_PERM as the first option is known to lead to a disaster 459fd6149d3SIlpo Järvinen * (but it may well be that other scenarios fail similarly). 460fd6149d3SIlpo Järvinen */ 46133ad798cSAdam Langley static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp, 462bd0388aeSWilliam Allen Simpson struct tcp_out_options *opts) 463bd0388aeSWilliam Allen Simpson { 4642100c8d2SYuchung Cheng u16 options = opts->options; /* mungable copy */ 465bd0388aeSWilliam Allen Simpson 466bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_MD5 & options)) { 4671a2c6181SChristoph Paasch *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | 4681a2c6181SChristoph Paasch (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG); 469bd0388aeSWilliam Allen Simpson /* overload cookie hash location */ 470bd0388aeSWilliam Allen Simpson opts->hash_location = (__u8 *)ptr; 47133ad798cSAdam Langley ptr += 4; 47233ad798cSAdam Langley } 47333ad798cSAdam Langley 474fd6149d3SIlpo Järvinen if (unlikely(opts->mss)) { 475fd6149d3SIlpo Järvinen *ptr++ = htonl((TCPOPT_MSS << 24) | 476fd6149d3SIlpo Järvinen (TCPOLEN_MSS << 16) | 477fd6149d3SIlpo Järvinen opts->mss); 478fd6149d3SIlpo Järvinen } 479fd6149d3SIlpo Järvinen 480bd0388aeSWilliam Allen Simpson if (likely(OPTION_TS & options)) { 481bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 48233ad798cSAdam Langley *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | 48333ad798cSAdam Langley (TCPOLEN_SACK_PERM << 16) | 48433ad798cSAdam Langley (TCPOPT_TIMESTAMP << 8) | 48533ad798cSAdam Langley TCPOLEN_TIMESTAMP); 486bd0388aeSWilliam Allen Simpson options &= ~OPTION_SACK_ADVERTISE; 48733ad798cSAdam Langley } else { 488496c98dfSYOSHIFUJI Hideaki *ptr++ = htonl((TCPOPT_NOP << 24) | 48940efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 49040efc6faSStephen Hemminger (TCPOPT_TIMESTAMP << 8) | 49140efc6faSStephen Hemminger TCPOLEN_TIMESTAMP); 49240efc6faSStephen Hemminger } 49333ad798cSAdam Langley *ptr++ = htonl(opts->tsval); 49433ad798cSAdam Langley *ptr++ = htonl(opts->tsecr); 49533ad798cSAdam Langley } 49633ad798cSAdam Langley 497bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 49833ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 49933ad798cSAdam Langley (TCPOPT_NOP << 16) | 50033ad798cSAdam Langley (TCPOPT_SACK_PERM << 8) | 50133ad798cSAdam Langley TCPOLEN_SACK_PERM); 50233ad798cSAdam Langley } 50333ad798cSAdam Langley 504bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_WSCALE & options)) { 50533ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 50633ad798cSAdam Langley (TCPOPT_WINDOW << 16) | 50733ad798cSAdam Langley (TCPOLEN_WINDOW << 8) | 50833ad798cSAdam Langley opts->ws); 50933ad798cSAdam Langley } 51033ad798cSAdam Langley 51133ad798cSAdam Langley if (unlikely(opts->num_sack_blocks)) { 51233ad798cSAdam Langley struct tcp_sack_block *sp = tp->rx_opt.dsack ? 51333ad798cSAdam Langley tp->duplicate_sack : tp->selective_acks; 51440efc6faSStephen Hemminger int this_sack; 51540efc6faSStephen Hemminger 51640efc6faSStephen Hemminger *ptr++ = htonl((TCPOPT_NOP << 24) | 51740efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 51840efc6faSStephen Hemminger (TCPOPT_SACK << 8) | 51933ad798cSAdam Langley (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * 52040efc6faSStephen Hemminger TCPOLEN_SACK_PERBLOCK))); 5212de979bdSStephen Hemminger 52233ad798cSAdam Langley for (this_sack = 0; this_sack < opts->num_sack_blocks; 52333ad798cSAdam Langley ++this_sack) { 52440efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].start_seq); 52540efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].end_seq); 52640efc6faSStephen Hemminger } 5272de979bdSStephen Hemminger 52840efc6faSStephen Hemminger tp->rx_opt.dsack = 0; 52940efc6faSStephen Hemminger } 5302100c8d2SYuchung Cheng 5312100c8d2SYuchung Cheng if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) { 5322100c8d2SYuchung Cheng struct tcp_fastopen_cookie *foc = opts->fastopen_cookie; 5337f9b838bSDaniel Lee u8 *p = (u8 *)ptr; 5347f9b838bSDaniel Lee u32 len; /* Fast Open option length */ 5352100c8d2SYuchung Cheng 5367f9b838bSDaniel Lee if (foc->exp) { 5377f9b838bSDaniel Lee len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len; 5387f9b838bSDaniel Lee *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) | 5392100c8d2SYuchung Cheng TCPOPT_FASTOPEN_MAGIC); 5407f9b838bSDaniel Lee p += TCPOLEN_EXP_FASTOPEN_BASE; 5417f9b838bSDaniel Lee } else { 5427f9b838bSDaniel Lee len = TCPOLEN_FASTOPEN_BASE + foc->len; 5437f9b838bSDaniel Lee *p++ = TCPOPT_FASTOPEN; 5447f9b838bSDaniel Lee *p++ = len; 5452100c8d2SYuchung Cheng } 5467f9b838bSDaniel Lee 5477f9b838bSDaniel Lee memcpy(p, foc->val, foc->len); 5487f9b838bSDaniel Lee if ((len & 3) == 2) { 5497f9b838bSDaniel Lee p[foc->len] = TCPOPT_NOP; 5507f9b838bSDaniel Lee p[foc->len + 1] = TCPOPT_NOP; 5517f9b838bSDaniel Lee } 5527f9b838bSDaniel Lee ptr += (len + 3) >> 2; 5532100c8d2SYuchung Cheng } 55460e2a778SUrsula Braun 55560e2a778SUrsula Braun smc_options_write(ptr, &options); 55660e2a778SUrsula Braun } 55760e2a778SUrsula Braun 55860e2a778SUrsula Braun static void smc_set_option(const struct tcp_sock *tp, 55960e2a778SUrsula Braun struct tcp_out_options *opts, 56060e2a778SUrsula Braun unsigned int *remaining) 56160e2a778SUrsula Braun { 56260e2a778SUrsula Braun #if IS_ENABLED(CONFIG_SMC) 56360e2a778SUrsula Braun if (static_branch_unlikely(&tcp_have_smc)) { 56460e2a778SUrsula Braun if (tp->syn_smc) { 56560e2a778SUrsula Braun if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) { 56660e2a778SUrsula Braun opts->options |= OPTION_SMC; 56760e2a778SUrsula Braun *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED; 56860e2a778SUrsula Braun } 56960e2a778SUrsula Braun } 57060e2a778SUrsula Braun } 57160e2a778SUrsula Braun #endif 57260e2a778SUrsula Braun } 57360e2a778SUrsula Braun 57460e2a778SUrsula Braun static void smc_set_option_cond(const struct tcp_sock *tp, 57560e2a778SUrsula Braun const struct inet_request_sock *ireq, 57660e2a778SUrsula Braun struct tcp_out_options *opts, 57760e2a778SUrsula Braun unsigned int *remaining) 57860e2a778SUrsula Braun { 57960e2a778SUrsula Braun #if IS_ENABLED(CONFIG_SMC) 58060e2a778SUrsula Braun if (static_branch_unlikely(&tcp_have_smc)) { 58160e2a778SUrsula Braun if (tp->syn_smc && ireq->smc_ok) { 58260e2a778SUrsula Braun if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) { 58360e2a778SUrsula Braun opts->options |= OPTION_SMC; 58460e2a778SUrsula Braun *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED; 58560e2a778SUrsula Braun } 58660e2a778SUrsula Braun } 58760e2a778SUrsula Braun } 58860e2a778SUrsula Braun #endif 58940efc6faSStephen Hemminger } 59040efc6faSStephen Hemminger 59167edfef7SAndi Kleen /* Compute TCP options for SYN packets. This is not the final 59267edfef7SAndi Kleen * network wire format yet. 59367edfef7SAndi Kleen */ 59495c96174SEric Dumazet static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb, 59533ad798cSAdam Langley struct tcp_out_options *opts, 596cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 597cf533ea5SEric Dumazet { 59833ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 59995c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 600783237e8SYuchung Cheng struct tcp_fastopen_request *fastopen = tp->fastopen_req; 60133ad798cSAdam Langley 602cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 60333ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 60433ad798cSAdam Langley if (*md5) { 60533ad798cSAdam Langley opts->options |= OPTION_MD5; 606bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 607cfb6eeb4SYOSHIFUJI Hideaki } 60833ad798cSAdam Langley #else 60933ad798cSAdam Langley *md5 = NULL; 610cfb6eeb4SYOSHIFUJI Hideaki #endif 61133ad798cSAdam Langley 61233ad798cSAdam Langley /* We always get an MSS option. The option bytes which will be seen in 61333ad798cSAdam Langley * normal data packets should timestamps be used, must be in the MSS 61433ad798cSAdam Langley * advertised. But we subtract them from tp->mss_cache so that 61533ad798cSAdam Langley * calculations in tcp_sendmsg are simpler etc. So account for this 61633ad798cSAdam Langley * fact here if necessary. If we don't do this correctly, as a 61733ad798cSAdam Langley * receiver we won't recognize data packets as being full sized when we 61833ad798cSAdam Langley * should, and thus we won't abide by the delayed ACK rules correctly. 61933ad798cSAdam Langley * SACKs don't matter, we never delay an ACK when we have any of those 62033ad798cSAdam Langley * going out. */ 62133ad798cSAdam Langley opts->mss = tcp_advertise_mss(sk); 622bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 62333ad798cSAdam Langley 6245d2ed052SEric Dumazet if (likely(sock_net(sk)->ipv4.sysctl_tcp_timestamps && !*md5)) { 62533ad798cSAdam Langley opts->options |= OPTION_TS; 6267faee5c0SEric Dumazet opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset; 62733ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 628bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 62933ad798cSAdam Langley } 6309bb37ef0SEric Dumazet if (likely(sock_net(sk)->ipv4.sysctl_tcp_window_scaling)) { 63133ad798cSAdam Langley opts->ws = tp->rx_opt.rcv_wscale; 63289e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 633bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 63433ad798cSAdam Langley } 635f9301034SEric Dumazet if (likely(sock_net(sk)->ipv4.sysctl_tcp_sack)) { 63633ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 637b32d1310SDavid S. Miller if (unlikely(!(OPTION_TS & opts->options))) 638bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 63933ad798cSAdam Langley } 64033ad798cSAdam Langley 641783237e8SYuchung Cheng if (fastopen && fastopen->cookie.len >= 0) { 6422646c831SDaniel Lee u32 need = fastopen->cookie.len; 6432646c831SDaniel Lee 6442646c831SDaniel Lee need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE : 6452646c831SDaniel Lee TCPOLEN_FASTOPEN_BASE; 646783237e8SYuchung Cheng need = (need + 3) & ~3U; /* Align to 32 bits */ 647783237e8SYuchung Cheng if (remaining >= need) { 648783237e8SYuchung Cheng opts->options |= OPTION_FAST_OPEN_COOKIE; 649783237e8SYuchung Cheng opts->fastopen_cookie = &fastopen->cookie; 650783237e8SYuchung Cheng remaining -= need; 651783237e8SYuchung Cheng tp->syn_fastopen = 1; 6522646c831SDaniel Lee tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0; 653783237e8SYuchung Cheng } 654783237e8SYuchung Cheng } 655bd0388aeSWilliam Allen Simpson 65660e2a778SUrsula Braun smc_set_option(tp, opts, &remaining); 65760e2a778SUrsula Braun 658bd0388aeSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 65933ad798cSAdam Langley } 66033ad798cSAdam Langley 66167edfef7SAndi Kleen /* Set up TCP options for SYN-ACKs. */ 66260e2a778SUrsula Braun static unsigned int tcp_synack_options(const struct sock *sk, 66360e2a778SUrsula Braun struct request_sock *req, 66495c96174SEric Dumazet unsigned int mss, struct sk_buff *skb, 66533ad798cSAdam Langley struct tcp_out_options *opts, 66680f03e27SEric Dumazet const struct tcp_md5sig_key *md5, 6678336886fSJerry Chu struct tcp_fastopen_cookie *foc) 6684957faadSWilliam Allen Simpson { 66933ad798cSAdam Langley struct inet_request_sock *ireq = inet_rsk(req); 67095c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 67133ad798cSAdam Langley 67233ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 67380f03e27SEric Dumazet if (md5) { 67433ad798cSAdam Langley opts->options |= OPTION_MD5; 6754957faadSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 6764957faadSWilliam Allen Simpson 6774957faadSWilliam Allen Simpson /* We can't fit any SACK blocks in a packet with MD5 + TS 6784957faadSWilliam Allen Simpson * options. There was discussion about disabling SACK 6794957faadSWilliam Allen Simpson * rather than TS in order to fit in better with old, 6804957faadSWilliam Allen Simpson * buggy kernels, but that was deemed to be unnecessary. 6814957faadSWilliam Allen Simpson */ 682de213e5eSEric Dumazet ireq->tstamp_ok &= !ireq->sack_ok; 68333ad798cSAdam Langley } 68433ad798cSAdam Langley #endif 68533ad798cSAdam Langley 6864957faadSWilliam Allen Simpson /* We always send an MSS option. */ 68733ad798cSAdam Langley opts->mss = mss; 6884957faadSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 68933ad798cSAdam Langley 69033ad798cSAdam Langley if (likely(ireq->wscale_ok)) { 69133ad798cSAdam Langley opts->ws = ireq->rcv_wscale; 69289e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 6934957faadSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 69433ad798cSAdam Langley } 695de213e5eSEric Dumazet if (likely(ireq->tstamp_ok)) { 69633ad798cSAdam Langley opts->options |= OPTION_TS; 69795a22caeSFlorian Westphal opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off; 69833ad798cSAdam Langley opts->tsecr = req->ts_recent; 6994957faadSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 70033ad798cSAdam Langley } 70133ad798cSAdam Langley if (likely(ireq->sack_ok)) { 70233ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 703de213e5eSEric Dumazet if (unlikely(!ireq->tstamp_ok)) 7044957faadSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 70533ad798cSAdam Langley } 7067f9b838bSDaniel Lee if (foc != NULL && foc->len >= 0) { 7077f9b838bSDaniel Lee u32 need = foc->len; 7087f9b838bSDaniel Lee 7097f9b838bSDaniel Lee need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE : 7107f9b838bSDaniel Lee TCPOLEN_FASTOPEN_BASE; 7118336886fSJerry Chu need = (need + 3) & ~3U; /* Align to 32 bits */ 7128336886fSJerry Chu if (remaining >= need) { 7138336886fSJerry Chu opts->options |= OPTION_FAST_OPEN_COOKIE; 7148336886fSJerry Chu opts->fastopen_cookie = foc; 7158336886fSJerry Chu remaining -= need; 7168336886fSJerry Chu } 7178336886fSJerry Chu } 7184957faadSWilliam Allen Simpson 71960e2a778SUrsula Braun smc_set_option_cond(tcp_sk(sk), ireq, opts, &remaining); 72060e2a778SUrsula Braun 7214957faadSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 72233ad798cSAdam Langley } 72333ad798cSAdam Langley 72467edfef7SAndi Kleen /* Compute TCP options for ESTABLISHED sockets. This is not the 72567edfef7SAndi Kleen * final wire format yet. 72667edfef7SAndi Kleen */ 72795c96174SEric Dumazet static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb, 72833ad798cSAdam Langley struct tcp_out_options *opts, 729cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 730cf533ea5SEric Dumazet { 73133ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 73295c96174SEric Dumazet unsigned int size = 0; 733cabeccbdSIlpo Järvinen unsigned int eff_sacks; 73433ad798cSAdam Langley 7355843ef42SAndi Kleen opts->options = 0; 7365843ef42SAndi Kleen 73733ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 73833ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 73933ad798cSAdam Langley if (unlikely(*md5)) { 74033ad798cSAdam Langley opts->options |= OPTION_MD5; 74133ad798cSAdam Langley size += TCPOLEN_MD5SIG_ALIGNED; 74233ad798cSAdam Langley } 74333ad798cSAdam Langley #else 74433ad798cSAdam Langley *md5 = NULL; 74533ad798cSAdam Langley #endif 74633ad798cSAdam Langley 74733ad798cSAdam Langley if (likely(tp->rx_opt.tstamp_ok)) { 74833ad798cSAdam Langley opts->options |= OPTION_TS; 7497faee5c0SEric Dumazet opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0; 75033ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 75133ad798cSAdam Langley size += TCPOLEN_TSTAMP_ALIGNED; 75233ad798cSAdam Langley } 75333ad798cSAdam Langley 754cabeccbdSIlpo Järvinen eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; 755cabeccbdSIlpo Järvinen if (unlikely(eff_sacks)) { 75695c96174SEric Dumazet const unsigned int remaining = MAX_TCP_OPTION_SPACE - size; 75733ad798cSAdam Langley opts->num_sack_blocks = 75895c96174SEric Dumazet min_t(unsigned int, eff_sacks, 75933ad798cSAdam Langley (remaining - TCPOLEN_SACK_BASE_ALIGNED) / 76033ad798cSAdam Langley TCPOLEN_SACK_PERBLOCK); 76133ad798cSAdam Langley size += TCPOLEN_SACK_BASE_ALIGNED + 76233ad798cSAdam Langley opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; 76333ad798cSAdam Langley } 76433ad798cSAdam Langley 76533ad798cSAdam Langley return size; 76640efc6faSStephen Hemminger } 7671da177e4SLinus Torvalds 76846d3ceabSEric Dumazet 76946d3ceabSEric Dumazet /* TCP SMALL QUEUES (TSQ) 77046d3ceabSEric Dumazet * 77146d3ceabSEric Dumazet * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev) 77246d3ceabSEric Dumazet * to reduce RTT and bufferbloat. 77346d3ceabSEric Dumazet * We do this using a special skb destructor (tcp_wfree). 77446d3ceabSEric Dumazet * 77546d3ceabSEric Dumazet * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb 77646d3ceabSEric Dumazet * needs to be reallocated in a driver. 7778e3bff96Sstephen hemminger * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc 77846d3ceabSEric Dumazet * 77946d3ceabSEric Dumazet * Since transmit from skb destructor is forbidden, we use a tasklet 78046d3ceabSEric Dumazet * to process all sockets that eventually need to send more skbs. 78146d3ceabSEric Dumazet * We use one tasklet per cpu, with its own queue of sockets. 78246d3ceabSEric Dumazet */ 78346d3ceabSEric Dumazet struct tsq_tasklet { 78446d3ceabSEric Dumazet struct tasklet_struct tasklet; 78546d3ceabSEric Dumazet struct list_head head; /* queue of tcp sockets */ 78646d3ceabSEric Dumazet }; 78746d3ceabSEric Dumazet static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet); 78846d3ceabSEric Dumazet 7896f458dfbSEric Dumazet static void tcp_tsq_handler(struct sock *sk) 7906f458dfbSEric Dumazet { 7916f458dfbSEric Dumazet if ((1 << sk->sk_state) & 7926f458dfbSEric Dumazet (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING | 793f9616c35SEric Dumazet TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) { 794f9616c35SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 795f9616c35SEric Dumazet 796f9616c35SEric Dumazet if (tp->lost_out > tp->retrans_out && 797f9616c35SEric Dumazet tp->snd_cwnd > tcp_packets_in_flight(tp)) 798f9616c35SEric Dumazet tcp_xmit_retransmit_queue(sk); 799f9616c35SEric Dumazet 800f9616c35SEric Dumazet tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle, 801bf06200eSJohn Ogness 0, GFP_ATOMIC); 8026f458dfbSEric Dumazet } 803f9616c35SEric Dumazet } 80446d3ceabSEric Dumazet /* 8058e3bff96Sstephen hemminger * One tasklet per cpu tries to send more skbs. 80646d3ceabSEric Dumazet * We run in tasklet context but need to disable irqs when 8078e3bff96Sstephen hemminger * transferring tsq->head because tcp_wfree() might 80846d3ceabSEric Dumazet * interrupt us (non NAPI drivers) 80946d3ceabSEric Dumazet */ 81046d3ceabSEric Dumazet static void tcp_tasklet_func(unsigned long data) 81146d3ceabSEric Dumazet { 81246d3ceabSEric Dumazet struct tsq_tasklet *tsq = (struct tsq_tasklet *)data; 81346d3ceabSEric Dumazet LIST_HEAD(list); 81446d3ceabSEric Dumazet unsigned long flags; 81546d3ceabSEric Dumazet struct list_head *q, *n; 81646d3ceabSEric Dumazet struct tcp_sock *tp; 81746d3ceabSEric Dumazet struct sock *sk; 81846d3ceabSEric Dumazet 81946d3ceabSEric Dumazet local_irq_save(flags); 82046d3ceabSEric Dumazet list_splice_init(&tsq->head, &list); 82146d3ceabSEric Dumazet local_irq_restore(flags); 82246d3ceabSEric Dumazet 82346d3ceabSEric Dumazet list_for_each_safe(q, n, &list) { 82446d3ceabSEric Dumazet tp = list_entry(q, struct tcp_sock, tsq_node); 82546d3ceabSEric Dumazet list_del(&tp->tsq_node); 82646d3ceabSEric Dumazet 82746d3ceabSEric Dumazet sk = (struct sock *)tp; 8280a9648f1SEric Dumazet smp_mb__before_atomic(); 8297aa5470cSEric Dumazet clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags); 8307aa5470cSEric Dumazet 831b223feb9SEric Dumazet if (!sk->sk_lock.owned && 8327aa5470cSEric Dumazet test_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags)) { 83346d3ceabSEric Dumazet bh_lock_sock(sk); 83446d3ceabSEric Dumazet if (!sock_owned_by_user(sk)) { 8357aa5470cSEric Dumazet clear_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags); 8366f458dfbSEric Dumazet tcp_tsq_handler(sk); 83746d3ceabSEric Dumazet } 83846d3ceabSEric Dumazet bh_unlock_sock(sk); 839b223feb9SEric Dumazet } 84046d3ceabSEric Dumazet 84146d3ceabSEric Dumazet sk_free(sk); 84246d3ceabSEric Dumazet } 84346d3ceabSEric Dumazet } 84446d3ceabSEric Dumazet 84540fc3423SEric Dumazet #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \ 84640fc3423SEric Dumazet TCPF_WRITE_TIMER_DEFERRED | \ 84740fc3423SEric Dumazet TCPF_DELACK_TIMER_DEFERRED | \ 84840fc3423SEric Dumazet TCPF_MTU_REDUCED_DEFERRED) 84946d3ceabSEric Dumazet /** 85046d3ceabSEric Dumazet * tcp_release_cb - tcp release_sock() callback 85146d3ceabSEric Dumazet * @sk: socket 85246d3ceabSEric Dumazet * 85346d3ceabSEric Dumazet * called from release_sock() to perform protocol dependent 85446d3ceabSEric Dumazet * actions before socket release. 85546d3ceabSEric Dumazet */ 85646d3ceabSEric Dumazet void tcp_release_cb(struct sock *sk) 85746d3ceabSEric Dumazet { 8586f458dfbSEric Dumazet unsigned long flags, nflags; 85946d3ceabSEric Dumazet 8606f458dfbSEric Dumazet /* perform an atomic operation only if at least one flag is set */ 8616f458dfbSEric Dumazet do { 8627aa5470cSEric Dumazet flags = sk->sk_tsq_flags; 8636f458dfbSEric Dumazet if (!(flags & TCP_DEFERRED_ALL)) 8646f458dfbSEric Dumazet return; 8656f458dfbSEric Dumazet nflags = flags & ~TCP_DEFERRED_ALL; 8667aa5470cSEric Dumazet } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags); 8676f458dfbSEric Dumazet 86840fc3423SEric Dumazet if (flags & TCPF_TSQ_DEFERRED) 8696f458dfbSEric Dumazet tcp_tsq_handler(sk); 8706f458dfbSEric Dumazet 871c3f9b018SEric Dumazet /* Here begins the tricky part : 872c3f9b018SEric Dumazet * We are called from release_sock() with : 873c3f9b018SEric Dumazet * 1) BH disabled 874c3f9b018SEric Dumazet * 2) sk_lock.slock spinlock held 875c3f9b018SEric Dumazet * 3) socket owned by us (sk->sk_lock.owned == 1) 876c3f9b018SEric Dumazet * 877c3f9b018SEric Dumazet * But following code is meant to be called from BH handlers, 878c3f9b018SEric Dumazet * so we should keep BH disabled, but early release socket ownership 879c3f9b018SEric Dumazet */ 880c3f9b018SEric Dumazet sock_release_ownership(sk); 881c3f9b018SEric Dumazet 88240fc3423SEric Dumazet if (flags & TCPF_WRITE_TIMER_DEFERRED) { 8836f458dfbSEric Dumazet tcp_write_timer_handler(sk); 884144d56e9SEric Dumazet __sock_put(sk); 885144d56e9SEric Dumazet } 88640fc3423SEric Dumazet if (flags & TCPF_DELACK_TIMER_DEFERRED) { 8876f458dfbSEric Dumazet tcp_delack_timer_handler(sk); 888144d56e9SEric Dumazet __sock_put(sk); 889144d56e9SEric Dumazet } 89040fc3423SEric Dumazet if (flags & TCPF_MTU_REDUCED_DEFERRED) { 8914fab9071SNeal Cardwell inet_csk(sk)->icsk_af_ops->mtu_reduced(sk); 892144d56e9SEric Dumazet __sock_put(sk); 893144d56e9SEric Dumazet } 89446d3ceabSEric Dumazet } 89546d3ceabSEric Dumazet EXPORT_SYMBOL(tcp_release_cb); 89646d3ceabSEric Dumazet 89746d3ceabSEric Dumazet void __init tcp_tasklet_init(void) 89846d3ceabSEric Dumazet { 89946d3ceabSEric Dumazet int i; 90046d3ceabSEric Dumazet 90146d3ceabSEric Dumazet for_each_possible_cpu(i) { 90246d3ceabSEric Dumazet struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i); 90346d3ceabSEric Dumazet 90446d3ceabSEric Dumazet INIT_LIST_HEAD(&tsq->head); 90546d3ceabSEric Dumazet tasklet_init(&tsq->tasklet, 90646d3ceabSEric Dumazet tcp_tasklet_func, 90746d3ceabSEric Dumazet (unsigned long)tsq); 90846d3ceabSEric Dumazet } 90946d3ceabSEric Dumazet } 91046d3ceabSEric Dumazet 91146d3ceabSEric Dumazet /* 91246d3ceabSEric Dumazet * Write buffer destructor automatically called from kfree_skb. 9138e3bff96Sstephen hemminger * We can't xmit new skbs from this context, as we might already 91446d3ceabSEric Dumazet * hold qdisc lock. 91546d3ceabSEric Dumazet */ 916d6a4a104SEric Dumazet void tcp_wfree(struct sk_buff *skb) 91746d3ceabSEric Dumazet { 91846d3ceabSEric Dumazet struct sock *sk = skb->sk; 91946d3ceabSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 920408f0a6cSEric Dumazet unsigned long flags, nval, oval; 9219b462d02SEric Dumazet 9229b462d02SEric Dumazet /* Keep one reference on sk_wmem_alloc. 9239b462d02SEric Dumazet * Will be released by sk_free() from here or tcp_tasklet_func() 9249b462d02SEric Dumazet */ 92514afee4bSReshetova, Elena WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc)); 9269b462d02SEric Dumazet 9279b462d02SEric Dumazet /* If this softirq is serviced by ksoftirqd, we are likely under stress. 9289b462d02SEric Dumazet * Wait until our queues (qdisc + devices) are drained. 9299b462d02SEric Dumazet * This gives : 9309b462d02SEric Dumazet * - less callbacks to tcp_write_xmit(), reducing stress (batches) 9319b462d02SEric Dumazet * - chance for incoming ACK (processed by another cpu maybe) 9329b462d02SEric Dumazet * to migrate this flow (skb->ooo_okay will be eventually set) 9339b462d02SEric Dumazet */ 93414afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current) 9359b462d02SEric Dumazet goto out; 93646d3ceabSEric Dumazet 9377aa5470cSEric Dumazet for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) { 93846d3ceabSEric Dumazet struct tsq_tasklet *tsq; 939a9b204d1SEric Dumazet bool empty; 94046d3ceabSEric Dumazet 941408f0a6cSEric Dumazet if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED)) 942408f0a6cSEric Dumazet goto out; 943408f0a6cSEric Dumazet 944b223feb9SEric Dumazet nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED | TCPF_TSQ_DEFERRED; 9457aa5470cSEric Dumazet nval = cmpxchg(&sk->sk_tsq_flags, oval, nval); 946408f0a6cSEric Dumazet if (nval != oval) 947408f0a6cSEric Dumazet continue; 948408f0a6cSEric Dumazet 94946d3ceabSEric Dumazet /* queue this socket to tasklet queue */ 95046d3ceabSEric Dumazet local_irq_save(flags); 951903ceff7SChristoph Lameter tsq = this_cpu_ptr(&tsq_tasklet); 952a9b204d1SEric Dumazet empty = list_empty(&tsq->head); 95346d3ceabSEric Dumazet list_add(&tp->tsq_node, &tsq->head); 954a9b204d1SEric Dumazet if (empty) 95546d3ceabSEric Dumazet tasklet_schedule(&tsq->tasklet); 95646d3ceabSEric Dumazet local_irq_restore(flags); 9579b462d02SEric Dumazet return; 95846d3ceabSEric Dumazet } 9599b462d02SEric Dumazet out: 9609b462d02SEric Dumazet sk_free(sk); 96146d3ceabSEric Dumazet } 96246d3ceabSEric Dumazet 963218af599SEric Dumazet /* Note: Called under hard irq. 964218af599SEric Dumazet * We can not call TCP stack right away. 965218af599SEric Dumazet */ 966218af599SEric Dumazet enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer) 967218af599SEric Dumazet { 968218af599SEric Dumazet struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer); 969218af599SEric Dumazet struct sock *sk = (struct sock *)tp; 970218af599SEric Dumazet unsigned long nval, oval; 971218af599SEric Dumazet 972218af599SEric Dumazet for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) { 973218af599SEric Dumazet struct tsq_tasklet *tsq; 974218af599SEric Dumazet bool empty; 975218af599SEric Dumazet 976218af599SEric Dumazet if (oval & TSQF_QUEUED) 977218af599SEric Dumazet break; 978218af599SEric Dumazet 979218af599SEric Dumazet nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED | TCPF_TSQ_DEFERRED; 980218af599SEric Dumazet nval = cmpxchg(&sk->sk_tsq_flags, oval, nval); 981218af599SEric Dumazet if (nval != oval) 982218af599SEric Dumazet continue; 983218af599SEric Dumazet 98414afee4bSReshetova, Elena if (!refcount_inc_not_zero(&sk->sk_wmem_alloc)) 985218af599SEric Dumazet break; 986218af599SEric Dumazet /* queue this socket to tasklet queue */ 987218af599SEric Dumazet tsq = this_cpu_ptr(&tsq_tasklet); 988218af599SEric Dumazet empty = list_empty(&tsq->head); 989218af599SEric Dumazet list_add(&tp->tsq_node, &tsq->head); 990218af599SEric Dumazet if (empty) 991218af599SEric Dumazet tasklet_schedule(&tsq->tasklet); 992218af599SEric Dumazet break; 993218af599SEric Dumazet } 994218af599SEric Dumazet return HRTIMER_NORESTART; 995218af599SEric Dumazet } 996218af599SEric Dumazet 997218af599SEric Dumazet /* BBR congestion control needs pacing. 998218af599SEric Dumazet * Same remark for SO_MAX_PACING_RATE. 999218af599SEric Dumazet * sch_fq packet scheduler is efficiently handling pacing, 1000218af599SEric Dumazet * but is not always installed/used. 1001218af599SEric Dumazet * Return true if TCP stack should pace packets itself. 1002218af599SEric Dumazet */ 1003218af599SEric Dumazet static bool tcp_needs_internal_pacing(const struct sock *sk) 1004218af599SEric Dumazet { 1005218af599SEric Dumazet return smp_load_acquire(&sk->sk_pacing_status) == SK_PACING_NEEDED; 1006218af599SEric Dumazet } 1007218af599SEric Dumazet 1008218af599SEric Dumazet static void tcp_internal_pacing(struct sock *sk, const struct sk_buff *skb) 1009218af599SEric Dumazet { 1010218af599SEric Dumazet u64 len_ns; 1011218af599SEric Dumazet u32 rate; 1012218af599SEric Dumazet 1013218af599SEric Dumazet if (!tcp_needs_internal_pacing(sk)) 1014218af599SEric Dumazet return; 1015218af599SEric Dumazet rate = sk->sk_pacing_rate; 1016218af599SEric Dumazet if (!rate || rate == ~0U) 1017218af599SEric Dumazet return; 1018218af599SEric Dumazet 1019218af599SEric Dumazet /* Should account for header sizes as sch_fq does, 1020218af599SEric Dumazet * but lets make things simple. 1021218af599SEric Dumazet */ 1022218af599SEric Dumazet len_ns = (u64)skb->len * NSEC_PER_SEC; 1023218af599SEric Dumazet do_div(len_ns, rate); 1024218af599SEric Dumazet hrtimer_start(&tcp_sk(sk)->pacing_timer, 1025218af599SEric Dumazet ktime_add_ns(ktime_get(), len_ns), 1026218af599SEric Dumazet HRTIMER_MODE_ABS_PINNED); 1027218af599SEric Dumazet } 1028218af599SEric Dumazet 1029e2080072SEric Dumazet static void tcp_update_skb_after_send(struct tcp_sock *tp, struct sk_buff *skb) 1030e2080072SEric Dumazet { 1031e2080072SEric Dumazet skb->skb_mstamp = tp->tcp_mstamp; 1032e2080072SEric Dumazet list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue); 1033e2080072SEric Dumazet } 1034e2080072SEric Dumazet 10351da177e4SLinus Torvalds /* This routine actually transmits TCP packets queued in by 10361da177e4SLinus Torvalds * tcp_do_sendmsg(). This is used by both the initial 10371da177e4SLinus Torvalds * transmission and possible later retransmissions. 10381da177e4SLinus Torvalds * All SKB's seen here are completely headerless. It is our 10391da177e4SLinus Torvalds * job to build the TCP header, and pass the packet down to 10401da177e4SLinus Torvalds * IP so it can do the same plus pass the packet off to the 10411da177e4SLinus Torvalds * device. 10421da177e4SLinus Torvalds * 10431da177e4SLinus Torvalds * We are working here with either a clone of the original 10441da177e4SLinus Torvalds * SKB, or a fresh unique copy made by the retransmit engine. 10451da177e4SLinus Torvalds */ 1046056834d9SIlpo Järvinen static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, 1047056834d9SIlpo Järvinen gfp_t gfp_mask) 10481da177e4SLinus Torvalds { 10496687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 1050dfb4b9dcSDavid S. Miller struct inet_sock *inet; 1051dfb4b9dcSDavid S. Miller struct tcp_sock *tp; 1052dfb4b9dcSDavid S. Miller struct tcp_skb_cb *tcb; 105333ad798cSAdam Langley struct tcp_out_options opts; 105495c96174SEric Dumazet unsigned int tcp_options_size, tcp_header_size; 10558c72c65bSEric Dumazet struct sk_buff *oskb = NULL; 1056cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 10571da177e4SLinus Torvalds struct tcphdr *th; 10581da177e4SLinus Torvalds int err; 10591da177e4SLinus Torvalds 1060dfb4b9dcSDavid S. Miller BUG_ON(!skb || !tcp_skb_pcount(skb)); 10616f094b9eSLawrence Brakmo tp = tcp_sk(sk); 1062dfb4b9dcSDavid S. Miller 1063ccdbb6e9SEric Dumazet if (clone_it) { 10646f094b9eSLawrence Brakmo TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq 10656f094b9eSLawrence Brakmo - tp->snd_una; 10668c72c65bSEric Dumazet oskb = skb; 1067e2080072SEric Dumazet 1068e2080072SEric Dumazet tcp_skb_tsorted_save(oskb) { 1069e2080072SEric Dumazet if (unlikely(skb_cloned(oskb))) 1070e2080072SEric Dumazet skb = pskb_copy(oskb, gfp_mask); 1071dfb4b9dcSDavid S. Miller else 1072e2080072SEric Dumazet skb = skb_clone(oskb, gfp_mask); 1073e2080072SEric Dumazet } tcp_skb_tsorted_restore(oskb); 1074e2080072SEric Dumazet 1075dfb4b9dcSDavid S. Miller if (unlikely(!skb)) 1076dfb4b9dcSDavid S. Miller return -ENOBUFS; 1077dfb4b9dcSDavid S. Miller } 10788c72c65bSEric Dumazet skb->skb_mstamp = tp->tcp_mstamp; 1079dfb4b9dcSDavid S. Miller 1080dfb4b9dcSDavid S. Miller inet = inet_sk(sk); 1081dfb4b9dcSDavid S. Miller tcb = TCP_SKB_CB(skb); 108233ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 10831da177e4SLinus Torvalds 10844de075e0SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) 108533ad798cSAdam Langley tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); 108633ad798cSAdam Langley else 108733ad798cSAdam Langley tcp_options_size = tcp_established_options(sk, skb, &opts, 108833ad798cSAdam Langley &md5); 108933ad798cSAdam Langley tcp_header_size = tcp_options_size + sizeof(struct tcphdr); 10901da177e4SLinus Torvalds 1091547669d4SEric Dumazet /* if no packet is in qdisc/device queue, then allow XPS to select 1092b2532eb9SEric Dumazet * another queue. We can be called from tcp_tsq_handler() 1093b2532eb9SEric Dumazet * which holds one reference to sk_wmem_alloc. 1094b2532eb9SEric Dumazet * 1095b2532eb9SEric Dumazet * TODO: Ideally, in-flight pure ACK packets should not matter here. 1096b2532eb9SEric Dumazet * One way to get this would be to set skb->truesize = 2 on them. 1097547669d4SEric Dumazet */ 1098b2532eb9SEric Dumazet skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1); 10991da177e4SLinus Torvalds 110038ab52e8SEric Dumazet /* If we had to use memory reserve to allocate this skb, 110138ab52e8SEric Dumazet * this might cause drops if packet is looped back : 110238ab52e8SEric Dumazet * Other socket might not have SOCK_MEMALLOC. 110338ab52e8SEric Dumazet * Packets not looped back do not care about pfmemalloc. 110438ab52e8SEric Dumazet */ 110538ab52e8SEric Dumazet skb->pfmemalloc = 0; 110638ab52e8SEric Dumazet 1107aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 1108aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 110946d3ceabSEric Dumazet 111046d3ceabSEric Dumazet skb_orphan(skb); 111146d3ceabSEric Dumazet skb->sk = sk; 11121d2077acSEric Dumazet skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree; 1113b73c3d0eSTom Herbert skb_set_hash_from_sk(skb, sk); 111414afee4bSReshetova, Elena refcount_add(skb->truesize, &sk->sk_wmem_alloc); 11151da177e4SLinus Torvalds 1116c3a2e837SJulian Anastasov skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm); 1117c3a2e837SJulian Anastasov 11181da177e4SLinus Torvalds /* Build TCP header and checksum it. */ 1119ea1627c2SEric Dumazet th = (struct tcphdr *)skb->data; 1120c720c7e8SEric Dumazet th->source = inet->inet_sport; 1121c720c7e8SEric Dumazet th->dest = inet->inet_dport; 11221da177e4SLinus Torvalds th->seq = htonl(tcb->seq); 11231da177e4SLinus Torvalds th->ack_seq = htonl(tp->rcv_nxt); 1124df7a3b07SAl Viro *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | 11254de075e0SEric Dumazet tcb->tcp_flags); 1126dfb4b9dcSDavid S. Miller 11271da177e4SLinus Torvalds th->check = 0; 11281da177e4SLinus Torvalds th->urg_ptr = 0; 11291da177e4SLinus Torvalds 113033f5f57eSIlpo Järvinen /* The urg_mode check is necessary during a below snd_una win probe */ 11317691367dSHerbert Xu if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { 11327691367dSHerbert Xu if (before(tp->snd_up, tcb->seq + 0x10000)) { 11331da177e4SLinus Torvalds th->urg_ptr = htons(tp->snd_up - tcb->seq); 11341da177e4SLinus Torvalds th->urg = 1; 11357691367dSHerbert Xu } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { 11360eae88f3SEric Dumazet th->urg_ptr = htons(0xFFFF); 11377691367dSHerbert Xu th->urg = 1; 11387691367dSHerbert Xu } 11391da177e4SLinus Torvalds } 11401da177e4SLinus Torvalds 1141bd0388aeSWilliam Allen Simpson tcp_options_write((__be32 *)(th + 1), tp, &opts); 114251466a75SEric Dumazet skb_shinfo(skb)->gso_type = sk->sk_gso_type; 1143ea1627c2SEric Dumazet if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) { 1144ea1627c2SEric Dumazet th->window = htons(tcp_select_window(sk)); 1145ea1627c2SEric Dumazet tcp_ecn_send(sk, skb, th, tcp_header_size); 1146ea1627c2SEric Dumazet } else { 1147ea1627c2SEric Dumazet /* RFC1323: The window in SYN & SYN/ACK segments 1148ea1627c2SEric Dumazet * is never scaled. 1149ea1627c2SEric Dumazet */ 1150ea1627c2SEric Dumazet th->window = htons(min(tp->rcv_wnd, 65535U)); 1151ea1627c2SEric Dumazet } 1152cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 1153cfb6eeb4SYOSHIFUJI Hideaki /* Calculate the MD5 hash, as we have all we need now */ 1154cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 1155a465419bSEric Dumazet sk_nocaps_add(sk, NETIF_F_GSO_MASK); 1156bd0388aeSWilliam Allen Simpson tp->af_specific->calc_md5_hash(opts.hash_location, 115739f8e58eSEric Dumazet md5, sk, skb); 1158cfb6eeb4SYOSHIFUJI Hideaki } 1159cfb6eeb4SYOSHIFUJI Hideaki #endif 1160cfb6eeb4SYOSHIFUJI Hideaki 1161bb296246SHerbert Xu icsk->icsk_af_ops->send_check(sk, skb); 11621da177e4SLinus Torvalds 11634de075e0SEric Dumazet if (likely(tcb->tcp_flags & TCPHDR_ACK)) 1164fc6415bcSDavid S. Miller tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); 11651da177e4SLinus Torvalds 1166a44d6eacSMartin KaFai Lau if (skb->len != tcp_header_size) { 1167cf533ea5SEric Dumazet tcp_event_data_sent(tp, sk); 1168a44d6eacSMartin KaFai Lau tp->data_segs_out += tcp_skb_pcount(skb); 1169218af599SEric Dumazet tcp_internal_pacing(sk, skb); 1170a44d6eacSMartin KaFai Lau } 11711da177e4SLinus Torvalds 1172bd37a088SWei Yongjun if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) 1173aa2ea058STom Herbert TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, 1174aa2ea058STom Herbert tcp_skb_pcount(skb)); 11751da177e4SLinus Torvalds 11762efd055cSMarcelo Ricardo Leitner tp->segs_out += tcp_skb_pcount(skb); 1177f69ad292SEric Dumazet /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */ 1178cd7d8498SEric Dumazet skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb); 1179f69ad292SEric Dumazet skb_shinfo(skb)->gso_size = tcp_skb_mss(skb); 1180cd7d8498SEric Dumazet 11817faee5c0SEric Dumazet /* Our usage of tstamp should remain private */ 11822456e855SThomas Gleixner skb->tstamp = 0; 1183971f10ecSEric Dumazet 1184971f10ecSEric Dumazet /* Cleanup our debris for IP stacks */ 1185971f10ecSEric Dumazet memset(skb->cb, 0, max(sizeof(struct inet_skb_parm), 1186971f10ecSEric Dumazet sizeof(struct inet6_skb_parm))); 1187971f10ecSEric Dumazet 1188b0270e91SEric Dumazet err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl); 11897faee5c0SEric Dumazet 11908c72c65bSEric Dumazet if (unlikely(err > 0)) { 11915ee2c941SChristoph Paasch tcp_enter_cwr(sk); 11928c72c65bSEric Dumazet err = net_xmit_eval(err); 11938c72c65bSEric Dumazet } 1194fc225799SEric Dumazet if (!err && oskb) { 1195e2080072SEric Dumazet tcp_update_skb_after_send(tp, oskb); 1196fc225799SEric Dumazet tcp_rate_skb_sent(sk, oskb); 1197fc225799SEric Dumazet } 11988c72c65bSEric Dumazet return err; 11991da177e4SLinus Torvalds } 12001da177e4SLinus Torvalds 120167edfef7SAndi Kleen /* This routine just queues the buffer for sending. 12021da177e4SLinus Torvalds * 12031da177e4SLinus Torvalds * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, 12041da177e4SLinus Torvalds * otherwise socket can stall. 12051da177e4SLinus Torvalds */ 12061da177e4SLinus Torvalds static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) 12071da177e4SLinus Torvalds { 12081da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 12091da177e4SLinus Torvalds 12101da177e4SLinus Torvalds /* Advance write_seq and place onto the write_queue. */ 12111da177e4SLinus Torvalds tp->write_seq = TCP_SKB_CB(skb)->end_seq; 1212f4a775d1SEric Dumazet __skb_header_release(skb); 1213fe067e8aSDavid S. Miller tcp_add_write_queue_tail(sk, skb); 12143ab224beSHideo Aoki sk->sk_wmem_queued += skb->truesize; 12153ab224beSHideo Aoki sk_mem_charge(sk, skb->truesize); 12161da177e4SLinus Torvalds } 12171da177e4SLinus Torvalds 121867edfef7SAndi Kleen /* Initialize TSO segments for a packet. */ 12195bbb432cSEric Dumazet static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now) 1220f6302d1dSDavid S. Miller { 12218f26fb1cSEric Dumazet if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) { 1222f6302d1dSDavid S. Miller /* Avoid the costly divide in the normal 1223f6302d1dSDavid S. Miller * non-TSO case. 1224f6302d1dSDavid S. Miller */ 1225cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, 1); 1226f69ad292SEric Dumazet TCP_SKB_CB(skb)->tcp_gso_size = 0; 1227f6302d1dSDavid S. Miller } else { 1228cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now)); 1229f69ad292SEric Dumazet TCP_SKB_CB(skb)->tcp_gso_size = mss_now; 12301da177e4SLinus Torvalds } 12311da177e4SLinus Torvalds } 12321da177e4SLinus Torvalds 123391fed7a1SIlpo Järvinen /* When a modification to fackets out becomes necessary, we need to check 123468f8353bSIlpo Järvinen * skb is counted to fackets_out or not. 123591fed7a1SIlpo Järvinen */ 1236cf533ea5SEric Dumazet static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb, 123791fed7a1SIlpo Järvinen int decr) 123891fed7a1SIlpo Järvinen { 1239a47e5a98SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1240a47e5a98SIlpo Järvinen 1241dc86967bSIlpo Järvinen if (!tp->sacked_out || tcp_is_reno(tp)) 124291fed7a1SIlpo Järvinen return; 124391fed7a1SIlpo Järvinen 12446859d494SIlpo Järvinen if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq)) 124591fed7a1SIlpo Järvinen tp->fackets_out -= decr; 124691fed7a1SIlpo Järvinen } 124791fed7a1SIlpo Järvinen 1248797108d1SIlpo Järvinen /* Pcount in the middle of the write queue got changed, we need to do various 1249797108d1SIlpo Järvinen * tweaks to fix counters 1250797108d1SIlpo Järvinen */ 1251cf533ea5SEric Dumazet static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) 1252797108d1SIlpo Järvinen { 1253797108d1SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1254797108d1SIlpo Järvinen 1255797108d1SIlpo Järvinen tp->packets_out -= decr; 1256797108d1SIlpo Järvinen 1257797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 1258797108d1SIlpo Järvinen tp->sacked_out -= decr; 1259797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 1260797108d1SIlpo Järvinen tp->retrans_out -= decr; 1261797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 1262797108d1SIlpo Järvinen tp->lost_out -= decr; 1263797108d1SIlpo Järvinen 1264797108d1SIlpo Järvinen /* Reno case is special. Sigh... */ 1265797108d1SIlpo Järvinen if (tcp_is_reno(tp) && decr > 0) 1266797108d1SIlpo Järvinen tp->sacked_out -= min_t(u32, tp->sacked_out, decr); 1267797108d1SIlpo Järvinen 1268797108d1SIlpo Järvinen tcp_adjust_fackets_out(sk, skb, decr); 1269797108d1SIlpo Järvinen 1270797108d1SIlpo Järvinen if (tp->lost_skb_hint && 1271797108d1SIlpo Järvinen before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && 127252cf3cc8SIlpo Järvinen (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))) 1273797108d1SIlpo Järvinen tp->lost_cnt_hint -= decr; 1274797108d1SIlpo Järvinen 1275797108d1SIlpo Järvinen tcp_verify_left_out(tp); 1276797108d1SIlpo Järvinen } 1277797108d1SIlpo Järvinen 12780a2cf20cSSoheil Hassas Yeganeh static bool tcp_has_tx_tstamp(const struct sk_buff *skb) 12790a2cf20cSSoheil Hassas Yeganeh { 12800a2cf20cSSoheil Hassas Yeganeh return TCP_SKB_CB(skb)->txstamp_ack || 12810a2cf20cSSoheil Hassas Yeganeh (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP); 12820a2cf20cSSoheil Hassas Yeganeh } 12830a2cf20cSSoheil Hassas Yeganeh 1284490cc7d0SWillem de Bruijn static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2) 1285490cc7d0SWillem de Bruijn { 1286490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo = skb_shinfo(skb); 1287490cc7d0SWillem de Bruijn 12880a2cf20cSSoheil Hassas Yeganeh if (unlikely(tcp_has_tx_tstamp(skb)) && 1289490cc7d0SWillem de Bruijn !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) { 1290490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo2 = skb_shinfo(skb2); 1291490cc7d0SWillem de Bruijn u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP; 1292490cc7d0SWillem de Bruijn 1293490cc7d0SWillem de Bruijn shinfo->tx_flags &= ~tsflags; 1294490cc7d0SWillem de Bruijn shinfo2->tx_flags |= tsflags; 1295490cc7d0SWillem de Bruijn swap(shinfo->tskey, shinfo2->tskey); 1296b51e13faSMartin KaFai Lau TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack; 1297b51e13faSMartin KaFai Lau TCP_SKB_CB(skb)->txstamp_ack = 0; 1298490cc7d0SWillem de Bruijn } 1299490cc7d0SWillem de Bruijn } 1300490cc7d0SWillem de Bruijn 1301a166140eSMartin KaFai Lau static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2) 1302a166140eSMartin KaFai Lau { 1303a166140eSMartin KaFai Lau TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor; 1304a166140eSMartin KaFai Lau TCP_SKB_CB(skb)->eor = 0; 1305a166140eSMartin KaFai Lau } 1306a166140eSMartin KaFai Lau 130775c119afSEric Dumazet /* Insert buff after skb on the write or rtx queue of sk. */ 130875c119afSEric Dumazet static void tcp_insert_write_queue_after(struct sk_buff *skb, 130975c119afSEric Dumazet struct sk_buff *buff, 131075c119afSEric Dumazet struct sock *sk, 131175c119afSEric Dumazet enum tcp_queue tcp_queue) 131275c119afSEric Dumazet { 131375c119afSEric Dumazet if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE) 131475c119afSEric Dumazet __skb_queue_after(&sk->sk_write_queue, skb, buff); 131575c119afSEric Dumazet else 131675c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, buff); 131775c119afSEric Dumazet } 131875c119afSEric Dumazet 13191da177e4SLinus Torvalds /* Function to create two new TCP segments. Shrinks the given segment 13201da177e4SLinus Torvalds * to the specified size and appends a new segment with the rest of the 13211da177e4SLinus Torvalds * packet to the list. This won't be called frequently, I hope. 13221da177e4SLinus Torvalds * Remember, these are still headerless SKBs at this point. 13231da177e4SLinus Torvalds */ 132475c119afSEric Dumazet int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue, 132575c119afSEric Dumazet struct sk_buff *skb, u32 len, 13266cc55e09SOctavian Purdila unsigned int mss_now, gfp_t gfp) 13271da177e4SLinus Torvalds { 13281da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 13291da177e4SLinus Torvalds struct sk_buff *buff; 13306475be16SDavid S. Miller int nsize, old_factor; 1331b60b49eaSHerbert Xu int nlen; 13329ce01461SIlpo Järvinen u8 flags; 13331da177e4SLinus Torvalds 13342fceec13SIlpo Järvinen if (WARN_ON(len > skb->len)) 13352fceec13SIlpo Järvinen return -EINVAL; 13366a438bbeSStephen Hemminger 13371da177e4SLinus Torvalds nsize = skb_headlen(skb) - len; 13381da177e4SLinus Torvalds if (nsize < 0) 13391da177e4SLinus Torvalds nsize = 0; 13401da177e4SLinus Torvalds 13416cc55e09SOctavian Purdila if (skb_unclone(skb, gfp)) 13421da177e4SLinus Torvalds return -ENOMEM; 13431da177e4SLinus Torvalds 13441da177e4SLinus Torvalds /* Get a new skb... force flag on. */ 1345eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, nsize, gfp, true); 134651456b29SIan Morris if (!buff) 13471da177e4SLinus Torvalds return -ENOMEM; /* We'll just try again later. */ 1348ef5cb973SHerbert Xu 13493ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 13503ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1351b60b49eaSHerbert Xu nlen = skb->len - len - nsize; 1352b60b49eaSHerbert Xu buff->truesize += nlen; 1353b60b49eaSHerbert Xu skb->truesize -= nlen; 13541da177e4SLinus Torvalds 13551da177e4SLinus Torvalds /* Correct the sequence numbers. */ 13561da177e4SLinus Torvalds TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 13571da177e4SLinus Torvalds TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 13581da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 13591da177e4SLinus Torvalds 13601da177e4SLinus Torvalds /* PSH and FIN should only be set in the second packet. */ 13614de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 13624de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 13634de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1364e14c3cafSHerbert Xu TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; 1365a166140eSMartin KaFai Lau tcp_skb_fragment_eor(skb, buff); 13661da177e4SLinus Torvalds 136784fa7933SPatrick McHardy if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { 13681da177e4SLinus Torvalds /* Copy and checksum data tail into the new buffer. */ 1369056834d9SIlpo Järvinen buff->csum = csum_partial_copy_nocheck(skb->data + len, 1370056834d9SIlpo Järvinen skb_put(buff, nsize), 13711da177e4SLinus Torvalds nsize, 0); 13721da177e4SLinus Torvalds 13731da177e4SLinus Torvalds skb_trim(skb, len); 13741da177e4SLinus Torvalds 13751da177e4SLinus Torvalds skb->csum = csum_block_sub(skb->csum, buff->csum, len); 13761da177e4SLinus Torvalds } else { 137784fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 13781da177e4SLinus Torvalds skb_split(skb, buff, len); 13791da177e4SLinus Torvalds } 13801da177e4SLinus Torvalds 13811da177e4SLinus Torvalds buff->ip_summed = skb->ip_summed; 13821da177e4SLinus Torvalds 1383a61bbcf2SPatrick McHardy buff->tstamp = skb->tstamp; 1384490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 13851da177e4SLinus Torvalds 13866475be16SDavid S. Miller old_factor = tcp_skb_pcount(skb); 13876475be16SDavid S. Miller 13881da177e4SLinus Torvalds /* Fix up tso_factor for both original and new SKB. */ 13895bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 13905bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 13911da177e4SLinus Torvalds 1392b9f64820SYuchung Cheng /* Update delivered info for the new segment */ 1393b9f64820SYuchung Cheng TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx; 1394b9f64820SYuchung Cheng 13956475be16SDavid S. Miller /* If this packet has been sent out already, we must 13966475be16SDavid S. Miller * adjust the various packet counters. 13976475be16SDavid S. Miller */ 1398cf0b450cSHerbert Xu if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 13996475be16SDavid S. Miller int diff = old_factor - tcp_skb_pcount(skb) - 14006475be16SDavid S. Miller tcp_skb_pcount(buff); 14011da177e4SLinus Torvalds 1402797108d1SIlpo Järvinen if (diff) 1403797108d1SIlpo Järvinen tcp_adjust_pcount(sk, skb, diff); 14041da177e4SLinus Torvalds } 14051da177e4SLinus Torvalds 14061da177e4SLinus Torvalds /* Link BUFF into the send queue. */ 1407f4a775d1SEric Dumazet __skb_header_release(buff); 140875c119afSEric Dumazet tcp_insert_write_queue_after(skb, buff, sk, tcp_queue); 1409e2080072SEric Dumazet list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor); 14101da177e4SLinus Torvalds 14111da177e4SLinus Torvalds return 0; 14121da177e4SLinus Torvalds } 14131da177e4SLinus Torvalds 1414f4d01666SEric Dumazet /* This is similar to __pskb_pull_tail(). The difference is that pulled 1415f4d01666SEric Dumazet * data is not copied, but immediately discarded. 14161da177e4SLinus Torvalds */ 14177162fb24SEric Dumazet static int __pskb_trim_head(struct sk_buff *skb, int len) 14181da177e4SLinus Torvalds { 14197b7fc97aSEric Dumazet struct skb_shared_info *shinfo; 14201da177e4SLinus Torvalds int i, k, eat; 14211da177e4SLinus Torvalds 14224fa48bf3SEric Dumazet eat = min_t(int, len, skb_headlen(skb)); 14234fa48bf3SEric Dumazet if (eat) { 14244fa48bf3SEric Dumazet __skb_pull(skb, eat); 14254fa48bf3SEric Dumazet len -= eat; 14264fa48bf3SEric Dumazet if (!len) 14277162fb24SEric Dumazet return 0; 14284fa48bf3SEric Dumazet } 14291da177e4SLinus Torvalds eat = len; 14301da177e4SLinus Torvalds k = 0; 14317b7fc97aSEric Dumazet shinfo = skb_shinfo(skb); 14327b7fc97aSEric Dumazet for (i = 0; i < shinfo->nr_frags; i++) { 14337b7fc97aSEric Dumazet int size = skb_frag_size(&shinfo->frags[i]); 14349e903e08SEric Dumazet 14359e903e08SEric Dumazet if (size <= eat) { 1436aff65da0SIan Campbell skb_frag_unref(skb, i); 14379e903e08SEric Dumazet eat -= size; 14381da177e4SLinus Torvalds } else { 14397b7fc97aSEric Dumazet shinfo->frags[k] = shinfo->frags[i]; 14401da177e4SLinus Torvalds if (eat) { 14417b7fc97aSEric Dumazet shinfo->frags[k].page_offset += eat; 14427b7fc97aSEric Dumazet skb_frag_size_sub(&shinfo->frags[k], eat); 14431da177e4SLinus Torvalds eat = 0; 14441da177e4SLinus Torvalds } 14451da177e4SLinus Torvalds k++; 14461da177e4SLinus Torvalds } 14471da177e4SLinus Torvalds } 14487b7fc97aSEric Dumazet shinfo->nr_frags = k; 14491da177e4SLinus Torvalds 14501da177e4SLinus Torvalds skb->data_len -= len; 14511da177e4SLinus Torvalds skb->len = skb->data_len; 14527162fb24SEric Dumazet return len; 14531da177e4SLinus Torvalds } 14541da177e4SLinus Torvalds 145567edfef7SAndi Kleen /* Remove acked data from a packet in the transmit queue. */ 14561da177e4SLinus Torvalds int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 14571da177e4SLinus Torvalds { 14587162fb24SEric Dumazet u32 delta_truesize; 14597162fb24SEric Dumazet 146014bbd6a5SPravin B Shelar if (skb_unclone(skb, GFP_ATOMIC)) 14611da177e4SLinus Torvalds return -ENOMEM; 14621da177e4SLinus Torvalds 14637162fb24SEric Dumazet delta_truesize = __pskb_trim_head(skb, len); 14641da177e4SLinus Torvalds 14651da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq += len; 146684fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 14671da177e4SLinus Torvalds 14687162fb24SEric Dumazet if (delta_truesize) { 14697162fb24SEric Dumazet skb->truesize -= delta_truesize; 14707162fb24SEric Dumazet sk->sk_wmem_queued -= delta_truesize; 14717162fb24SEric Dumazet sk_mem_uncharge(sk, delta_truesize); 14721da177e4SLinus Torvalds sock_set_flag(sk, SOCK_QUEUE_SHRUNK); 14737162fb24SEric Dumazet } 14741da177e4SLinus Torvalds 14755b35e1e6SNeal Cardwell /* Any change of skb->len requires recalculation of tso factor. */ 14761da177e4SLinus Torvalds if (tcp_skb_pcount(skb) > 1) 14775bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb)); 14781da177e4SLinus Torvalds 14791da177e4SLinus Torvalds return 0; 14801da177e4SLinus Torvalds } 14811da177e4SLinus Torvalds 14821b63edd6SYuchung Cheng /* Calculate MSS not accounting any TCP options. */ 14831b63edd6SYuchung Cheng static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu) 14845d424d5aSJohn Heffner { 1485cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1486cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 14875d424d5aSJohn Heffner int mss_now; 14885d424d5aSJohn Heffner 14895d424d5aSJohn Heffner /* Calculate base mss without TCP options: 14905d424d5aSJohn Heffner It is MMS_S - sizeof(tcphdr) of rfc1122 14915d424d5aSJohn Heffner */ 14925d424d5aSJohn Heffner mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 14935d424d5aSJohn Heffner 149467469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 149567469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 149667469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 149767469601SEric Dumazet 149867469601SEric Dumazet if (dst && dst_allfrag(dst)) 149967469601SEric Dumazet mss_now -= icsk->icsk_af_ops->net_frag_header_len; 150067469601SEric Dumazet } 150167469601SEric Dumazet 15025d424d5aSJohn Heffner /* Clamp it (mss_clamp does not include tcp options) */ 15035d424d5aSJohn Heffner if (mss_now > tp->rx_opt.mss_clamp) 15045d424d5aSJohn Heffner mss_now = tp->rx_opt.mss_clamp; 15055d424d5aSJohn Heffner 15065d424d5aSJohn Heffner /* Now subtract optional transport overhead */ 15075d424d5aSJohn Heffner mss_now -= icsk->icsk_ext_hdr_len; 15085d424d5aSJohn Heffner 15095d424d5aSJohn Heffner /* Then reserve room for full set of TCP options and 8 bytes of data */ 15105d424d5aSJohn Heffner if (mss_now < 48) 15115d424d5aSJohn Heffner mss_now = 48; 15125d424d5aSJohn Heffner return mss_now; 15135d424d5aSJohn Heffner } 15145d424d5aSJohn Heffner 15151b63edd6SYuchung Cheng /* Calculate MSS. Not accounting for SACKs here. */ 15161b63edd6SYuchung Cheng int tcp_mtu_to_mss(struct sock *sk, int pmtu) 15171b63edd6SYuchung Cheng { 15181b63edd6SYuchung Cheng /* Subtract TCP options size, not including SACKs */ 15191b63edd6SYuchung Cheng return __tcp_mtu_to_mss(sk, pmtu) - 15201b63edd6SYuchung Cheng (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr)); 15211b63edd6SYuchung Cheng } 15221b63edd6SYuchung Cheng 15235d424d5aSJohn Heffner /* Inverse of above */ 152467469601SEric Dumazet int tcp_mss_to_mtu(struct sock *sk, int mss) 15255d424d5aSJohn Heffner { 1526cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1527cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 15285d424d5aSJohn Heffner int mtu; 15295d424d5aSJohn Heffner 15305d424d5aSJohn Heffner mtu = mss + 15315d424d5aSJohn Heffner tp->tcp_header_len + 15325d424d5aSJohn Heffner icsk->icsk_ext_hdr_len + 15335d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 15345d424d5aSJohn Heffner 153567469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 153667469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 153767469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 153867469601SEric Dumazet 153967469601SEric Dumazet if (dst && dst_allfrag(dst)) 154067469601SEric Dumazet mtu += icsk->icsk_af_ops->net_frag_header_len; 154167469601SEric Dumazet } 15425d424d5aSJohn Heffner return mtu; 15435d424d5aSJohn Heffner } 1544556c6b46SNeal Cardwell EXPORT_SYMBOL(tcp_mss_to_mtu); 15455d424d5aSJohn Heffner 154667edfef7SAndi Kleen /* MTU probing init per socket */ 15475d424d5aSJohn Heffner void tcp_mtup_init(struct sock *sk) 15485d424d5aSJohn Heffner { 15495d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 15505d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 1551b0f9ca53SFan Du struct net *net = sock_net(sk); 15525d424d5aSJohn Heffner 1553b0f9ca53SFan Du icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1; 15545d424d5aSJohn Heffner icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 15555d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 1556b0f9ca53SFan Du icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss); 15575d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 155805cbc0dbSFan Du if (icsk->icsk_mtup.enabled) 1559c74df29aSEric Dumazet icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; 15605d424d5aSJohn Heffner } 15614bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_mtup_init); 15625d424d5aSJohn Heffner 15631da177e4SLinus Torvalds /* This function synchronize snd mss to current pmtu/exthdr set. 15641da177e4SLinus Torvalds 15651da177e4SLinus Torvalds tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 15661da177e4SLinus Torvalds for TCP options, but includes only bare TCP header. 15671da177e4SLinus Torvalds 15681da177e4SLinus Torvalds tp->rx_opt.mss_clamp is mss negotiated at connection setup. 1569caa20d9aSStephen Hemminger It is minimum of user_mss and mss received with SYN. 15701da177e4SLinus Torvalds It also does not include TCP options. 15711da177e4SLinus Torvalds 1572d83d8461SArnaldo Carvalho de Melo inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. 15731da177e4SLinus Torvalds 15741da177e4SLinus Torvalds tp->mss_cache is current effective sending mss, including 15751da177e4SLinus Torvalds all tcp options except for SACKs. It is evaluated, 15761da177e4SLinus Torvalds taking into account current pmtu, but never exceeds 15771da177e4SLinus Torvalds tp->rx_opt.mss_clamp. 15781da177e4SLinus Torvalds 15791da177e4SLinus Torvalds NOTE1. rfc1122 clearly states that advertised MSS 15801da177e4SLinus Torvalds DOES NOT include either tcp or ip options. 15811da177e4SLinus Torvalds 1582d83d8461SArnaldo Carvalho de Melo NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache 1583d83d8461SArnaldo Carvalho de Melo are READ ONLY outside this function. --ANK (980731) 15841da177e4SLinus Torvalds */ 15851da177e4SLinus Torvalds unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 15861da177e4SLinus Torvalds { 15871da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1588d83d8461SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 15895d424d5aSJohn Heffner int mss_now; 15901da177e4SLinus Torvalds 15915d424d5aSJohn Heffner if (icsk->icsk_mtup.search_high > pmtu) 15925d424d5aSJohn Heffner icsk->icsk_mtup.search_high = pmtu; 15931da177e4SLinus Torvalds 15945d424d5aSJohn Heffner mss_now = tcp_mtu_to_mss(sk, pmtu); 1595409d22b4SIlpo Järvinen mss_now = tcp_bound_to_half_wnd(tp, mss_now); 15961da177e4SLinus Torvalds 15971da177e4SLinus Torvalds /* And store cached results */ 1598d83d8461SArnaldo Carvalho de Melo icsk->icsk_pmtu_cookie = pmtu; 15995d424d5aSJohn Heffner if (icsk->icsk_mtup.enabled) 16005d424d5aSJohn Heffner mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); 1601c1b4a7e6SDavid S. Miller tp->mss_cache = mss_now; 16021da177e4SLinus Torvalds 16031da177e4SLinus Torvalds return mss_now; 16041da177e4SLinus Torvalds } 16054bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_sync_mss); 16061da177e4SLinus Torvalds 16071da177e4SLinus Torvalds /* Compute the current effective MSS, taking SACKs and IP options, 16081da177e4SLinus Torvalds * and even PMTU discovery events into account. 16091da177e4SLinus Torvalds */ 16100c54b85fSIlpo Järvinen unsigned int tcp_current_mss(struct sock *sk) 16111da177e4SLinus Torvalds { 1612cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1613cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1614c1b4a7e6SDavid S. Miller u32 mss_now; 161595c96174SEric Dumazet unsigned int header_len; 161633ad798cSAdam Langley struct tcp_out_options opts; 161733ad798cSAdam Langley struct tcp_md5sig_key *md5; 16181da177e4SLinus Torvalds 1619c1b4a7e6SDavid S. Miller mss_now = tp->mss_cache; 1620c1b4a7e6SDavid S. Miller 16211da177e4SLinus Torvalds if (dst) { 16221da177e4SLinus Torvalds u32 mtu = dst_mtu(dst); 1623d83d8461SArnaldo Carvalho de Melo if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 16241da177e4SLinus Torvalds mss_now = tcp_sync_mss(sk, mtu); 16251da177e4SLinus Torvalds } 16261da177e4SLinus Torvalds 162733ad798cSAdam Langley header_len = tcp_established_options(sk, NULL, &opts, &md5) + 162833ad798cSAdam Langley sizeof(struct tcphdr); 162933ad798cSAdam Langley /* The mss_cache is sized based on tp->tcp_header_len, which assumes 163033ad798cSAdam Langley * some common options. If this is an odd packet (because we have SACK 163133ad798cSAdam Langley * blocks etc) then our calculated header_len will be different, and 163233ad798cSAdam Langley * we have to adjust mss_now correspondingly */ 163333ad798cSAdam Langley if (header_len != tp->tcp_header_len) { 163433ad798cSAdam Langley int delta = (int) header_len - tp->tcp_header_len; 163533ad798cSAdam Langley mss_now -= delta; 163633ad798cSAdam Langley } 1637cfb6eeb4SYOSHIFUJI Hideaki 16381da177e4SLinus Torvalds return mss_now; 16391da177e4SLinus Torvalds } 16401da177e4SLinus Torvalds 164186fd14adSWeiping Pan /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. 164286fd14adSWeiping Pan * As additional protections, we do not touch cwnd in retransmission phases, 164386fd14adSWeiping Pan * and if application hit its sndbuf limit recently. 164486fd14adSWeiping Pan */ 164586fd14adSWeiping Pan static void tcp_cwnd_application_limited(struct sock *sk) 1646a762a980SDavid S. Miller { 16479e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1648a762a980SDavid S. Miller 164986fd14adSWeiping Pan if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && 165086fd14adSWeiping Pan sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 165186fd14adSWeiping Pan /* Limited by application or receiver window. */ 165286fd14adSWeiping Pan u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); 165386fd14adSWeiping Pan u32 win_used = max(tp->snd_cwnd_used, init_win); 165486fd14adSWeiping Pan if (win_used < tp->snd_cwnd) { 165586fd14adSWeiping Pan tp->snd_ssthresh = tcp_current_ssthresh(sk); 165686fd14adSWeiping Pan tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1; 165786fd14adSWeiping Pan } 165886fd14adSWeiping Pan tp->snd_cwnd_used = 0; 165986fd14adSWeiping Pan } 1660c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 166186fd14adSWeiping Pan } 166286fd14adSWeiping Pan 1663ca8a2263SNeal Cardwell static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited) 1664a762a980SDavid S. Miller { 16651b1fc3fdSWei Wang const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 1666a762a980SDavid S. Miller struct tcp_sock *tp = tcp_sk(sk); 1667a762a980SDavid S. Miller 1668ca8a2263SNeal Cardwell /* Track the maximum number of outstanding packets in each 1669ca8a2263SNeal Cardwell * window, and remember whether we were cwnd-limited then. 1670ca8a2263SNeal Cardwell */ 1671ca8a2263SNeal Cardwell if (!before(tp->snd_una, tp->max_packets_seq) || 1672ca8a2263SNeal Cardwell tp->packets_out > tp->max_packets_out) { 1673ca8a2263SNeal Cardwell tp->max_packets_out = tp->packets_out; 1674ca8a2263SNeal Cardwell tp->max_packets_seq = tp->snd_nxt; 1675ca8a2263SNeal Cardwell tp->is_cwnd_limited = is_cwnd_limited; 1676ca8a2263SNeal Cardwell } 1677e114a710SEric Dumazet 167824901551SEric Dumazet if (tcp_is_cwnd_limited(sk)) { 1679a762a980SDavid S. Miller /* Network is feed fully. */ 1680a762a980SDavid S. Miller tp->snd_cwnd_used = 0; 1681c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 1682a762a980SDavid S. Miller } else { 1683a762a980SDavid S. Miller /* Network starves. */ 1684a762a980SDavid S. Miller if (tp->packets_out > tp->snd_cwnd_used) 1685a762a980SDavid S. Miller tp->snd_cwnd_used = tp->packets_out; 1686a762a980SDavid S. Miller 1687b510f0d2SEric Dumazet if (sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle && 1688c2203cf7SEric Dumazet (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto && 16891b1fc3fdSWei Wang !ca_ops->cong_control) 1690a762a980SDavid S. Miller tcp_cwnd_application_limited(sk); 1691b0f71bd3SFrancis Yan 1692b0f71bd3SFrancis Yan /* The following conditions together indicate the starvation 1693b0f71bd3SFrancis Yan * is caused by insufficient sender buffer: 1694b0f71bd3SFrancis Yan * 1) just sent some data (see tcp_write_xmit) 1695b0f71bd3SFrancis Yan * 2) not cwnd limited (this else condition) 169675c119afSEric Dumazet * 3) no more data to send (tcp_write_queue_empty()) 1697b0f71bd3SFrancis Yan * 4) application is hitting buffer limit (SOCK_NOSPACE) 1698b0f71bd3SFrancis Yan */ 169975c119afSEric Dumazet if (tcp_write_queue_empty(sk) && sk->sk_socket && 1700b0f71bd3SFrancis Yan test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) && 1701b0f71bd3SFrancis Yan (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) 1702b0f71bd3SFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED); 1703a762a980SDavid S. Miller } 1704a762a980SDavid S. Miller } 1705a762a980SDavid S. Miller 1706d4589926SEric Dumazet /* Minshall's variant of the Nagle send check. */ 1707d4589926SEric Dumazet static bool tcp_minshall_check(const struct tcp_sock *tp) 1708d4589926SEric Dumazet { 1709d4589926SEric Dumazet return after(tp->snd_sml, tp->snd_una) && 1710d4589926SEric Dumazet !after(tp->snd_sml, tp->snd_nxt); 1711d4589926SEric Dumazet } 1712d4589926SEric Dumazet 1713d4589926SEric Dumazet /* Update snd_sml if this skb is under mss 1714d4589926SEric Dumazet * Note that a TSO packet might end with a sub-mss segment 1715d4589926SEric Dumazet * The test is really : 1716d4589926SEric Dumazet * if ((skb->len % mss) != 0) 1717d4589926SEric Dumazet * tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1718d4589926SEric Dumazet * But we can avoid doing the divide again given we already have 1719d4589926SEric Dumazet * skb_pcount = skb->len / mss_now 17200e3a4803SIlpo Järvinen */ 1721d4589926SEric Dumazet static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now, 1722d4589926SEric Dumazet const struct sk_buff *skb) 1723d4589926SEric Dumazet { 1724d4589926SEric Dumazet if (skb->len < tcp_skb_pcount(skb) * mss_now) 1725d4589926SEric Dumazet tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1726d4589926SEric Dumazet } 1727d4589926SEric Dumazet 1728d4589926SEric Dumazet /* Return false, if packet can be sent now without violation Nagle's rules: 1729d4589926SEric Dumazet * 1. It is full sized. (provided by caller in %partial bool) 1730d4589926SEric Dumazet * 2. Or it contains FIN. (already checked by caller) 1731d4589926SEric Dumazet * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. 1732d4589926SEric Dumazet * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 1733d4589926SEric Dumazet * With Minshall's modification: all sent small packets are ACKed. 1734d4589926SEric Dumazet */ 1735d4589926SEric Dumazet static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp, 1736cc93fc51SPeter Pan(潘卫平) int nonagle) 1737d4589926SEric Dumazet { 1738d4589926SEric Dumazet return partial && 1739d4589926SEric Dumazet ((nonagle & TCP_NAGLE_CORK) || 1740d4589926SEric Dumazet (!nonagle && tp->packets_out && tcp_minshall_check(tp))); 1741d4589926SEric Dumazet } 1742605ad7f1SEric Dumazet 1743605ad7f1SEric Dumazet /* Return how many segs we'd like on a TSO packet, 1744605ad7f1SEric Dumazet * to send one TSO packet per ms 1745605ad7f1SEric Dumazet */ 17461b3878caSNeal Cardwell u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now, 17471b3878caSNeal Cardwell int min_tso_segs) 1748605ad7f1SEric Dumazet { 1749605ad7f1SEric Dumazet u32 bytes, segs; 1750605ad7f1SEric Dumazet 1751605ad7f1SEric Dumazet bytes = min(sk->sk_pacing_rate >> 10, 1752605ad7f1SEric Dumazet sk->sk_gso_max_size - 1 - MAX_TCP_HEADER); 1753605ad7f1SEric Dumazet 1754605ad7f1SEric Dumazet /* Goal is to send at least one packet per ms, 1755605ad7f1SEric Dumazet * not one big TSO packet every 100 ms. 1756605ad7f1SEric Dumazet * This preserves ACK clocking and is consistent 1757605ad7f1SEric Dumazet * with tcp_tso_should_defer() heuristic. 1758605ad7f1SEric Dumazet */ 17591b3878caSNeal Cardwell segs = max_t(u32, bytes / mss_now, min_tso_segs); 1760605ad7f1SEric Dumazet 1761605ad7f1SEric Dumazet return min_t(u32, segs, sk->sk_gso_max_segs); 1762605ad7f1SEric Dumazet } 17631b3878caSNeal Cardwell EXPORT_SYMBOL(tcp_tso_autosize); 1764605ad7f1SEric Dumazet 1765ed6e7268SNeal Cardwell /* Return the number of segments we want in the skb we are transmitting. 1766ed6e7268SNeal Cardwell * See if congestion control module wants to decide; otherwise, autosize. 1767ed6e7268SNeal Cardwell */ 1768ed6e7268SNeal Cardwell static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now) 1769ed6e7268SNeal Cardwell { 1770ed6e7268SNeal Cardwell const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 1771ed6e7268SNeal Cardwell u32 tso_segs = ca_ops->tso_segs_goal ? ca_ops->tso_segs_goal(sk) : 0; 1772ed6e7268SNeal Cardwell 17731b3878caSNeal Cardwell return tso_segs ? : 17741b3878caSNeal Cardwell tcp_tso_autosize(sk, mss_now, sysctl_tcp_min_tso_segs); 1775ed6e7268SNeal Cardwell } 1776ed6e7268SNeal Cardwell 1777d4589926SEric Dumazet /* Returns the portion of skb which can be sent right away */ 1778d4589926SEric Dumazet static unsigned int tcp_mss_split_point(const struct sock *sk, 1779d4589926SEric Dumazet const struct sk_buff *skb, 1780d4589926SEric Dumazet unsigned int mss_now, 1781d4589926SEric Dumazet unsigned int max_segs, 1782d4589926SEric Dumazet int nonagle) 1783c1b4a7e6SDavid S. Miller { 1784cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1785d4589926SEric Dumazet u32 partial, needed, window, max_len; 1786c1b4a7e6SDavid S. Miller 178790840defSIlpo Järvinen window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 17881485348dSBen Hutchings max_len = mss_now * max_segs; 17890e3a4803SIlpo Järvinen 17901485348dSBen Hutchings if (likely(max_len <= window && skb != tcp_write_queue_tail(sk))) 17911485348dSBen Hutchings return max_len; 17920e3a4803SIlpo Järvinen 17935ea3a748SIlpo Järvinen needed = min(skb->len, window); 17945ea3a748SIlpo Järvinen 17951485348dSBen Hutchings if (max_len <= needed) 17961485348dSBen Hutchings return max_len; 17970e3a4803SIlpo Järvinen 1798d4589926SEric Dumazet partial = needed % mss_now; 1799d4589926SEric Dumazet /* If last segment is not a full MSS, check if Nagle rules allow us 1800d4589926SEric Dumazet * to include this last segment in this skb. 1801d4589926SEric Dumazet * Otherwise, we'll split the skb at last MSS boundary 1802d4589926SEric Dumazet */ 1803cc93fc51SPeter Pan(潘卫平) if (tcp_nagle_check(partial != 0, tp, nonagle)) 1804d4589926SEric Dumazet return needed - partial; 1805d4589926SEric Dumazet 1806d4589926SEric Dumazet return needed; 1807c1b4a7e6SDavid S. Miller } 1808c1b4a7e6SDavid S. Miller 1809c1b4a7e6SDavid S. Miller /* Can at least one segment of SKB be sent right now, according to the 1810c1b4a7e6SDavid S. Miller * congestion window rules? If so, return how many segments are allowed. 1811c1b4a7e6SDavid S. Miller */ 1812cf533ea5SEric Dumazet static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, 1813cf533ea5SEric Dumazet const struct sk_buff *skb) 1814c1b4a7e6SDavid S. Miller { 1815d649a7a8SEric Dumazet u32 in_flight, cwnd, halfcwnd; 1816c1b4a7e6SDavid S. Miller 1817c1b4a7e6SDavid S. Miller /* Don't be strict about the congestion window for the final FIN. */ 18184de075e0SEric Dumazet if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && 18194de075e0SEric Dumazet tcp_skb_pcount(skb) == 1) 1820c1b4a7e6SDavid S. Miller return 1; 1821c1b4a7e6SDavid S. Miller 1822c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1823c1b4a7e6SDavid S. Miller cwnd = tp->snd_cwnd; 1824d649a7a8SEric Dumazet if (in_flight >= cwnd) 1825c1b4a7e6SDavid S. Miller return 0; 1826d649a7a8SEric Dumazet 1827d649a7a8SEric Dumazet /* For better scheduling, ensure we have at least 1828d649a7a8SEric Dumazet * 2 GSO packets in flight. 1829d649a7a8SEric Dumazet */ 1830d649a7a8SEric Dumazet halfcwnd = max(cwnd >> 1, 1U); 1831d649a7a8SEric Dumazet return min(halfcwnd, cwnd - in_flight); 1832c1b4a7e6SDavid S. Miller } 1833c1b4a7e6SDavid S. Miller 1834b595076aSUwe Kleine-König /* Initialize TSO state of a skb. 183567edfef7SAndi Kleen * This must be invoked the first time we consider transmitting 1836c1b4a7e6SDavid S. Miller * SKB onto the wire. 1837c1b4a7e6SDavid S. Miller */ 18385bbb432cSEric Dumazet static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now) 1839c1b4a7e6SDavid S. Miller { 1840c1b4a7e6SDavid S. Miller int tso_segs = tcp_skb_pcount(skb); 1841c1b4a7e6SDavid S. Miller 1842f8269a49SIlpo Järvinen if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { 18435bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 1844c1b4a7e6SDavid S. Miller tso_segs = tcp_skb_pcount(skb); 1845c1b4a7e6SDavid S. Miller } 1846c1b4a7e6SDavid S. Miller return tso_segs; 1847c1b4a7e6SDavid S. Miller } 1848c1b4a7e6SDavid S. Miller 1849c1b4a7e6SDavid S. Miller 1850a2a385d6SEric Dumazet /* Return true if the Nagle test allows this packet to be 1851c1b4a7e6SDavid S. Miller * sent now. 1852c1b4a7e6SDavid S. Miller */ 1853a2a385d6SEric Dumazet static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, 1854c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1855c1b4a7e6SDavid S. Miller { 1856c1b4a7e6SDavid S. Miller /* Nagle rule does not apply to frames, which sit in the middle of the 1857c1b4a7e6SDavid S. Miller * write_queue (they have no chances to get new data). 1858c1b4a7e6SDavid S. Miller * 1859c1b4a7e6SDavid S. Miller * This is implemented in the callers, where they modify the 'nonagle' 1860c1b4a7e6SDavid S. Miller * argument based upon the location of SKB in the send queue. 1861c1b4a7e6SDavid S. Miller */ 1862c1b4a7e6SDavid S. Miller if (nonagle & TCP_NAGLE_PUSH) 1863a2a385d6SEric Dumazet return true; 1864c1b4a7e6SDavid S. Miller 18659b44190dSYuchung Cheng /* Don't use the nagle rule for urgent data (or for the final FIN). */ 18669b44190dSYuchung Cheng if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) 1867a2a385d6SEric Dumazet return true; 1868c1b4a7e6SDavid S. Miller 1869cc93fc51SPeter Pan(潘卫平) if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle)) 1870a2a385d6SEric Dumazet return true; 1871c1b4a7e6SDavid S. Miller 1872a2a385d6SEric Dumazet return false; 1873c1b4a7e6SDavid S. Miller } 1874c1b4a7e6SDavid S. Miller 1875c1b4a7e6SDavid S. Miller /* Does at least the first segment of SKB fit into the send window? */ 1876a2a385d6SEric Dumazet static bool tcp_snd_wnd_test(const struct tcp_sock *tp, 1877a2a385d6SEric Dumazet const struct sk_buff *skb, 1878056834d9SIlpo Järvinen unsigned int cur_mss) 1879c1b4a7e6SDavid S. Miller { 1880c1b4a7e6SDavid S. Miller u32 end_seq = TCP_SKB_CB(skb)->end_seq; 1881c1b4a7e6SDavid S. Miller 1882c1b4a7e6SDavid S. Miller if (skb->len > cur_mss) 1883c1b4a7e6SDavid S. Miller end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 1884c1b4a7e6SDavid S. Miller 188590840defSIlpo Järvinen return !after(end_seq, tcp_wnd_end(tp)); 1886c1b4a7e6SDavid S. Miller } 1887c1b4a7e6SDavid S. Miller 1888c1b4a7e6SDavid S. Miller /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 1889c1b4a7e6SDavid S. Miller * which is put after SKB on the list. It is very much like 1890c1b4a7e6SDavid S. Miller * tcp_fragment() except that it may make several kinds of assumptions 1891c1b4a7e6SDavid S. Miller * in order to speed up the splitting operation. In particular, we 1892c1b4a7e6SDavid S. Miller * know that all the data is in scatter-gather pages, and that the 1893c1b4a7e6SDavid S. Miller * packet has never been sent out before (and thus is not cloned). 1894c1b4a7e6SDavid S. Miller */ 189575c119afSEric Dumazet static int tso_fragment(struct sock *sk, enum tcp_queue tcp_queue, 189675c119afSEric Dumazet struct sk_buff *skb, unsigned int len, 1897c4ead4c5SEric Dumazet unsigned int mss_now, gfp_t gfp) 1898c1b4a7e6SDavid S. Miller { 1899c1b4a7e6SDavid S. Miller struct sk_buff *buff; 1900c1b4a7e6SDavid S. Miller int nlen = skb->len - len; 19019ce01461SIlpo Järvinen u8 flags; 1902c1b4a7e6SDavid S. Miller 1903c1b4a7e6SDavid S. Miller /* All of a TSO frame must be composed of paged data. */ 1904c8ac3774SHerbert Xu if (skb->len != skb->data_len) 190575c119afSEric Dumazet return tcp_fragment(sk, tcp_queue, skb, len, mss_now, gfp); 1906c1b4a7e6SDavid S. Miller 1907eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, gfp, true); 190851456b29SIan Morris if (unlikely(!buff)) 1909c1b4a7e6SDavid S. Miller return -ENOMEM; 1910c1b4a7e6SDavid S. Miller 19113ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 19123ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1913b60b49eaSHerbert Xu buff->truesize += nlen; 1914c1b4a7e6SDavid S. Miller skb->truesize -= nlen; 1915c1b4a7e6SDavid S. Miller 1916c1b4a7e6SDavid S. Miller /* Correct the sequence numbers. */ 1917c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 1918c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 1919c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 1920c1b4a7e6SDavid S. Miller 1921c1b4a7e6SDavid S. Miller /* PSH and FIN should only be set in the second packet. */ 19224de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 19234de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 19244de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1925c1b4a7e6SDavid S. Miller 1926c1b4a7e6SDavid S. Miller /* This packet was never sent out yet, so no SACK bits. */ 1927c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->sacked = 0; 1928c1b4a7e6SDavid S. Miller 1929a166140eSMartin KaFai Lau tcp_skb_fragment_eor(skb, buff); 1930a166140eSMartin KaFai Lau 193184fa7933SPatrick McHardy buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; 1932c1b4a7e6SDavid S. Miller skb_split(skb, buff, len); 1933490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 1934c1b4a7e6SDavid S. Miller 1935c1b4a7e6SDavid S. Miller /* Fix up tso_factor for both original and new SKB. */ 19365bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 19375bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 1938c1b4a7e6SDavid S. Miller 1939c1b4a7e6SDavid S. Miller /* Link BUFF into the send queue. */ 1940f4a775d1SEric Dumazet __skb_header_release(buff); 194175c119afSEric Dumazet tcp_insert_write_queue_after(skb, buff, sk, tcp_queue); 1942c1b4a7e6SDavid S. Miller 1943c1b4a7e6SDavid S. Miller return 0; 1944c1b4a7e6SDavid S. Miller } 1945c1b4a7e6SDavid S. Miller 1946c1b4a7e6SDavid S. Miller /* Try to defer sending, if possible, in order to minimize the amount 1947c1b4a7e6SDavid S. Miller * of TSO splitting we do. View it as a kind of TSO Nagle test. 1948c1b4a7e6SDavid S. Miller * 1949c1b4a7e6SDavid S. Miller * This algorithm is from John Heffner. 1950c1b4a7e6SDavid S. Miller */ 1951ca8a2263SNeal Cardwell static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb, 1952605ad7f1SEric Dumazet bool *is_cwnd_limited, u32 max_segs) 1953c1b4a7e6SDavid S. Miller { 19546687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 195550c8339eSEric Dumazet u32 age, send_win, cong_win, limit, in_flight; 195650c8339eSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 195750c8339eSEric Dumazet struct sk_buff *head; 1958ad9f4f50SEric Dumazet int win_divisor; 1959c1b4a7e6SDavid S. Miller 19604de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 1961ae8064acSJohn Heffner goto send_now; 1962c1b4a7e6SDavid S. Miller 196399d7662aSEric Dumazet if (icsk->icsk_ca_state >= TCP_CA_Recovery) 1964ae8064acSJohn Heffner goto send_now; 1965ae8064acSJohn Heffner 19665f852eb5SEric Dumazet /* Avoid bursty behavior by allowing defer 19675f852eb5SEric Dumazet * only if the last write was recent. 19685f852eb5SEric Dumazet */ 1969d635fbe2SEric Dumazet if ((s32)(tcp_jiffies32 - tp->lsndtime) > 0) 1970ae8064acSJohn Heffner goto send_now; 1971908a75c1SDavid S. Miller 1972c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1973c1b4a7e6SDavid S. Miller 1974056834d9SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight)); 1975c1b4a7e6SDavid S. Miller 197690840defSIlpo Järvinen send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1977c1b4a7e6SDavid S. Miller 1978c1b4a7e6SDavid S. Miller /* From in_flight test above, we know that cwnd > in_flight. */ 1979c1b4a7e6SDavid S. Miller cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; 1980c1b4a7e6SDavid S. Miller 1981c1b4a7e6SDavid S. Miller limit = min(send_win, cong_win); 1982c1b4a7e6SDavid S. Miller 1983ba244fe9SDavid S. Miller /* If a full-sized TSO skb can be sent, do it. */ 1984605ad7f1SEric Dumazet if (limit >= max_segs * tp->mss_cache) 1985ae8064acSJohn Heffner goto send_now; 1986ba244fe9SDavid S. Miller 198762ad2761SIlpo Järvinen /* Middle in queue won't get any more data, full sendable already? */ 198862ad2761SIlpo Järvinen if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) 198962ad2761SIlpo Järvinen goto send_now; 199062ad2761SIlpo Järvinen 1991ad9f4f50SEric Dumazet win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor); 1992ad9f4f50SEric Dumazet if (win_divisor) { 1993c1b4a7e6SDavid S. Miller u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); 1994c1b4a7e6SDavid S. Miller 1995c1b4a7e6SDavid S. Miller /* If at least some fraction of a window is available, 1996c1b4a7e6SDavid S. Miller * just use it. 1997c1b4a7e6SDavid S. Miller */ 1998ad9f4f50SEric Dumazet chunk /= win_divisor; 1999c1b4a7e6SDavid S. Miller if (limit >= chunk) 2000ae8064acSJohn Heffner goto send_now; 2001c1b4a7e6SDavid S. Miller } else { 2002c1b4a7e6SDavid S. Miller /* Different approach, try not to defer past a single 2003c1b4a7e6SDavid S. Miller * ACK. Receiver should ACK every other full sized 2004c1b4a7e6SDavid S. Miller * frame, so if we have space for more than 3 frames 2005c1b4a7e6SDavid S. Miller * then send now. 2006c1b4a7e6SDavid S. Miller */ 20076b5a5c0dSNeal Cardwell if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) 2008ae8064acSJohn Heffner goto send_now; 2009c1b4a7e6SDavid S. Miller } 2010c1b4a7e6SDavid S. Miller 201175c119afSEric Dumazet /* TODO : use tsorted_sent_queue ? */ 201275c119afSEric Dumazet head = tcp_rtx_queue_head(sk); 201375c119afSEric Dumazet if (!head) 201475c119afSEric Dumazet goto send_now; 20159a568de4SEric Dumazet age = tcp_stamp_us_delta(tp->tcp_mstamp, head->skb_mstamp); 201650c8339eSEric Dumazet /* If next ACK is likely to come too late (half srtt), do not defer */ 201750c8339eSEric Dumazet if (age < (tp->srtt_us >> 4)) 201850c8339eSEric Dumazet goto send_now; 201950c8339eSEric Dumazet 20205f852eb5SEric Dumazet /* Ok, it looks like it is advisable to defer. */ 2021ae8064acSJohn Heffner 2022d2e1339fSBendik Rønning Opstad if (cong_win < send_win && cong_win <= skb->len) 2023ca8a2263SNeal Cardwell *is_cwnd_limited = true; 2024ca8a2263SNeal Cardwell 2025a2a385d6SEric Dumazet return true; 2026ae8064acSJohn Heffner 2027ae8064acSJohn Heffner send_now: 2028a2a385d6SEric Dumazet return false; 2029c1b4a7e6SDavid S. Miller } 2030c1b4a7e6SDavid S. Miller 203105cbc0dbSFan Du static inline void tcp_mtu_check_reprobe(struct sock *sk) 203205cbc0dbSFan Du { 203305cbc0dbSFan Du struct inet_connection_sock *icsk = inet_csk(sk); 203405cbc0dbSFan Du struct tcp_sock *tp = tcp_sk(sk); 203505cbc0dbSFan Du struct net *net = sock_net(sk); 203605cbc0dbSFan Du u32 interval; 203705cbc0dbSFan Du s32 delta; 203805cbc0dbSFan Du 203905cbc0dbSFan Du interval = net->ipv4.sysctl_tcp_probe_interval; 2040c74df29aSEric Dumazet delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp; 204105cbc0dbSFan Du if (unlikely(delta >= interval * HZ)) { 204205cbc0dbSFan Du int mss = tcp_current_mss(sk); 204305cbc0dbSFan Du 204405cbc0dbSFan Du /* Update current search range */ 204505cbc0dbSFan Du icsk->icsk_mtup.probe_size = 0; 204605cbc0dbSFan Du icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + 204705cbc0dbSFan Du sizeof(struct tcphdr) + 204805cbc0dbSFan Du icsk->icsk_af_ops->net_header_len; 204905cbc0dbSFan Du icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss); 205005cbc0dbSFan Du 205105cbc0dbSFan Du /* Update probe time stamp */ 2052c74df29aSEric Dumazet icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; 205305cbc0dbSFan Du } 205405cbc0dbSFan Du } 205505cbc0dbSFan Du 20565d424d5aSJohn Heffner /* Create a new MTU probe if we are ready. 205767edfef7SAndi Kleen * MTU probe is regularly attempting to increase the path MTU by 205867edfef7SAndi Kleen * deliberately sending larger packets. This discovers routing 205967edfef7SAndi Kleen * changes resulting in larger path MTUs. 206067edfef7SAndi Kleen * 20615d424d5aSJohn Heffner * Returns 0 if we should wait to probe (no cwnd available), 20625d424d5aSJohn Heffner * 1 if a probe was sent, 2063056834d9SIlpo Järvinen * -1 otherwise 2064056834d9SIlpo Järvinen */ 20655d424d5aSJohn Heffner static int tcp_mtu_probe(struct sock *sk) 20665d424d5aSJohn Heffner { 20675d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 206812a59abcSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 20695d424d5aSJohn Heffner struct sk_buff *skb, *nskb, *next; 20706b58e0a5SFan Du struct net *net = sock_net(sk); 20715d424d5aSJohn Heffner int probe_size; 207291cc17c0SIlpo Järvinen int size_needed; 207312a59abcSEric Dumazet int copy, len; 20745d424d5aSJohn Heffner int mss_now; 20756b58e0a5SFan Du int interval; 20765d424d5aSJohn Heffner 20775d424d5aSJohn Heffner /* Not currently probing/verifying, 20785d424d5aSJohn Heffner * not in recovery, 20795d424d5aSJohn Heffner * have enough cwnd, and 208012a59abcSEric Dumazet * not SACKing (the variable headers throw things off) 208112a59abcSEric Dumazet */ 208212a59abcSEric Dumazet if (likely(!icsk->icsk_mtup.enabled || 20835d424d5aSJohn Heffner icsk->icsk_mtup.probe_size || 20845d424d5aSJohn Heffner inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 20855d424d5aSJohn Heffner tp->snd_cwnd < 11 || 208612a59abcSEric Dumazet tp->rx_opt.num_sacks || tp->rx_opt.dsack)) 20875d424d5aSJohn Heffner return -1; 20885d424d5aSJohn Heffner 20896b58e0a5SFan Du /* Use binary search for probe_size between tcp_mss_base, 20906b58e0a5SFan Du * and current mss_clamp. if (search_high - search_low) 20916b58e0a5SFan Du * smaller than a threshold, backoff from probing. 20926b58e0a5SFan Du */ 20930c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 20946b58e0a5SFan Du probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high + 20956b58e0a5SFan Du icsk->icsk_mtup.search_low) >> 1); 209691cc17c0SIlpo Järvinen size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; 20976b58e0a5SFan Du interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low; 209805cbc0dbSFan Du /* When misfortune happens, we are reprobing actively, 209905cbc0dbSFan Du * and then reprobe timer has expired. We stick with current 210005cbc0dbSFan Du * probing process by not resetting search range to its orignal. 210105cbc0dbSFan Du */ 21026b58e0a5SFan Du if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) || 210305cbc0dbSFan Du interval < net->ipv4.sysctl_tcp_probe_threshold) { 210405cbc0dbSFan Du /* Check whether enough time has elaplased for 210505cbc0dbSFan Du * another round of probing. 210605cbc0dbSFan Du */ 210705cbc0dbSFan Du tcp_mtu_check_reprobe(sk); 21085d424d5aSJohn Heffner return -1; 21095d424d5aSJohn Heffner } 21105d424d5aSJohn Heffner 21115d424d5aSJohn Heffner /* Have enough data in the send queue to probe? */ 21127f9c33e5SIlpo Järvinen if (tp->write_seq - tp->snd_nxt < size_needed) 21135d424d5aSJohn Heffner return -1; 21145d424d5aSJohn Heffner 211591cc17c0SIlpo Järvinen if (tp->snd_wnd < size_needed) 21165d424d5aSJohn Heffner return -1; 211790840defSIlpo Järvinen if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) 21185d424d5aSJohn Heffner return 0; 21195d424d5aSJohn Heffner 2120d67c58e9SIlpo Järvinen /* Do we need to wait to drain cwnd? With none in flight, don't stall */ 2121d67c58e9SIlpo Järvinen if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { 2122d67c58e9SIlpo Järvinen if (!tcp_packets_in_flight(tp)) 21235d424d5aSJohn Heffner return -1; 21245d424d5aSJohn Heffner else 21255d424d5aSJohn Heffner return 0; 21265d424d5aSJohn Heffner } 21275d424d5aSJohn Heffner 21285d424d5aSJohn Heffner /* We're allowed to probe. Build it now. */ 2129eb934478SEric Dumazet nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false); 213051456b29SIan Morris if (!nskb) 21315d424d5aSJohn Heffner return -1; 21323ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 21333ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 21345d424d5aSJohn Heffner 2135fe067e8aSDavid S. Miller skb = tcp_send_head(sk); 21365d424d5aSJohn Heffner 21375d424d5aSJohn Heffner TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 21385d424d5aSJohn Heffner TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 21394de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; 21405d424d5aSJohn Heffner TCP_SKB_CB(nskb)->sacked = 0; 21415d424d5aSJohn Heffner nskb->csum = 0; 214284fa7933SPatrick McHardy nskb->ip_summed = skb->ip_summed; 21435d424d5aSJohn Heffner 214450c4817eSIlpo Järvinen tcp_insert_write_queue_before(nskb, skb, sk); 214550c4817eSIlpo Järvinen 21465d424d5aSJohn Heffner len = 0; 2147234b6860SIlpo Järvinen tcp_for_write_queue_from_safe(skb, next, sk) { 21485d424d5aSJohn Heffner copy = min_t(int, skb->len, probe_size - len); 21492fe664f1SDouglas Caetano dos Santos if (nskb->ip_summed) { 21505d424d5aSJohn Heffner skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); 21512fe664f1SDouglas Caetano dos Santos } else { 21522fe664f1SDouglas Caetano dos Santos __wsum csum = skb_copy_and_csum_bits(skb, 0, 2153056834d9SIlpo Järvinen skb_put(nskb, copy), 21542fe664f1SDouglas Caetano dos Santos copy, 0); 21552fe664f1SDouglas Caetano dos Santos nskb->csum = csum_block_add(nskb->csum, csum, len); 21562fe664f1SDouglas Caetano dos Santos } 21575d424d5aSJohn Heffner 21585d424d5aSJohn Heffner if (skb->len <= copy) { 21595d424d5aSJohn Heffner /* We've eaten all the data from this skb. 21605d424d5aSJohn Heffner * Throw it away. */ 21614de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 2162fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 21633ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 21645d424d5aSJohn Heffner } else { 21654de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & 2166a3433f35SChangli Gao ~(TCPHDR_FIN|TCPHDR_PSH); 21675d424d5aSJohn Heffner if (!skb_shinfo(skb)->nr_frags) { 21685d424d5aSJohn Heffner skb_pull(skb, copy); 216984fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 2170056834d9SIlpo Järvinen skb->csum = csum_partial(skb->data, 2171056834d9SIlpo Järvinen skb->len, 0); 21725d424d5aSJohn Heffner } else { 21735d424d5aSJohn Heffner __pskb_trim_head(skb, copy); 21745bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 21755d424d5aSJohn Heffner } 21765d424d5aSJohn Heffner TCP_SKB_CB(skb)->seq += copy; 21775d424d5aSJohn Heffner } 21785d424d5aSJohn Heffner 21795d424d5aSJohn Heffner len += copy; 2180234b6860SIlpo Järvinen 2181234b6860SIlpo Järvinen if (len >= probe_size) 2182234b6860SIlpo Järvinen break; 21835d424d5aSJohn Heffner } 21845bbb432cSEric Dumazet tcp_init_tso_segs(nskb, nskb->len); 21855d424d5aSJohn Heffner 21865d424d5aSJohn Heffner /* We're ready to send. If this fails, the probe will 21877faee5c0SEric Dumazet * be resegmented into mss-sized pieces by tcp_write_xmit(). 21887faee5c0SEric Dumazet */ 21895d424d5aSJohn Heffner if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 21905d424d5aSJohn Heffner /* Decrement cwnd here because we are sending 21915d424d5aSJohn Heffner * effectively two packets. */ 21925d424d5aSJohn Heffner tp->snd_cwnd--; 219366f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, nskb); 21945d424d5aSJohn Heffner 21955d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 21960e7b1368SJohn Heffner tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 21970e7b1368SJohn Heffner tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 21985d424d5aSJohn Heffner 21995d424d5aSJohn Heffner return 1; 22005d424d5aSJohn Heffner } 22015d424d5aSJohn Heffner 22025d424d5aSJohn Heffner return -1; 22035d424d5aSJohn Heffner } 22045d424d5aSJohn Heffner 2205218af599SEric Dumazet static bool tcp_pacing_check(const struct sock *sk) 2206218af599SEric Dumazet { 2207218af599SEric Dumazet return tcp_needs_internal_pacing(sk) && 2208218af599SEric Dumazet hrtimer_active(&tcp_sk(sk)->pacing_timer); 2209218af599SEric Dumazet } 2210218af599SEric Dumazet 2211f9616c35SEric Dumazet /* TCP Small Queues : 2212f9616c35SEric Dumazet * Control number of packets in qdisc/devices to two packets / or ~1 ms. 2213f9616c35SEric Dumazet * (These limits are doubled for retransmits) 2214f9616c35SEric Dumazet * This allows for : 2215f9616c35SEric Dumazet * - better RTT estimation and ACK scheduling 2216f9616c35SEric Dumazet * - faster recovery 2217f9616c35SEric Dumazet * - high rates 2218f9616c35SEric Dumazet * Alas, some drivers / subsystems require a fair amount 2219f9616c35SEric Dumazet * of queued bytes to ensure line rate. 2220f9616c35SEric Dumazet * One example is wifi aggregation (802.11 AMPDU) 2221f9616c35SEric Dumazet */ 2222f9616c35SEric Dumazet static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb, 2223f9616c35SEric Dumazet unsigned int factor) 2224f9616c35SEric Dumazet { 2225f9616c35SEric Dumazet unsigned int limit; 2226f9616c35SEric Dumazet 2227f9616c35SEric Dumazet limit = max(2 * skb->truesize, sk->sk_pacing_rate >> 10); 2228f9616c35SEric Dumazet limit = min_t(u32, limit, sysctl_tcp_limit_output_bytes); 2229f9616c35SEric Dumazet limit <<= factor; 2230f9616c35SEric Dumazet 223114afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) > limit) { 223275c119afSEric Dumazet /* Always send skb if rtx queue is empty. 223375eefc6cSEric Dumazet * No need to wait for TX completion to call us back, 223475eefc6cSEric Dumazet * after softirq/tasklet schedule. 223575eefc6cSEric Dumazet * This helps when TX completions are delayed too much. 223675eefc6cSEric Dumazet */ 223775c119afSEric Dumazet if (tcp_rtx_queue_empty(sk)) 223875eefc6cSEric Dumazet return false; 223975eefc6cSEric Dumazet 22407aa5470cSEric Dumazet set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags); 2241f9616c35SEric Dumazet /* It is possible TX completion already happened 2242f9616c35SEric Dumazet * before we set TSQ_THROTTLED, so we must 2243f9616c35SEric Dumazet * test again the condition. 2244f9616c35SEric Dumazet */ 2245f9616c35SEric Dumazet smp_mb__after_atomic(); 224614afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) > limit) 2247f9616c35SEric Dumazet return true; 2248f9616c35SEric Dumazet } 2249f9616c35SEric Dumazet return false; 2250f9616c35SEric Dumazet } 2251f9616c35SEric Dumazet 225205b055e8SFrancis Yan static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new) 225305b055e8SFrancis Yan { 2254628174ccSEric Dumazet const u32 now = tcp_jiffies32; 2255efe967cdSArnd Bergmann enum tcp_chrono old = tp->chrono_type; 225605b055e8SFrancis Yan 2257efe967cdSArnd Bergmann if (old > TCP_CHRONO_UNSPEC) 2258efe967cdSArnd Bergmann tp->chrono_stat[old - 1] += now - tp->chrono_start; 225905b055e8SFrancis Yan tp->chrono_start = now; 226005b055e8SFrancis Yan tp->chrono_type = new; 226105b055e8SFrancis Yan } 226205b055e8SFrancis Yan 226305b055e8SFrancis Yan void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type) 226405b055e8SFrancis Yan { 226505b055e8SFrancis Yan struct tcp_sock *tp = tcp_sk(sk); 226605b055e8SFrancis Yan 226705b055e8SFrancis Yan /* If there are multiple conditions worthy of tracking in a 22680f87230dSFrancis Yan * chronograph then the highest priority enum takes precedence 22690f87230dSFrancis Yan * over the other conditions. So that if something "more interesting" 227005b055e8SFrancis Yan * starts happening, stop the previous chrono and start a new one. 227105b055e8SFrancis Yan */ 227205b055e8SFrancis Yan if (type > tp->chrono_type) 227305b055e8SFrancis Yan tcp_chrono_set(tp, type); 227405b055e8SFrancis Yan } 227505b055e8SFrancis Yan 227605b055e8SFrancis Yan void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type) 227705b055e8SFrancis Yan { 227805b055e8SFrancis Yan struct tcp_sock *tp = tcp_sk(sk); 227905b055e8SFrancis Yan 22800f87230dSFrancis Yan 22810f87230dSFrancis Yan /* There are multiple conditions worthy of tracking in a 22820f87230dSFrancis Yan * chronograph, so that the highest priority enum takes 22830f87230dSFrancis Yan * precedence over the other conditions (see tcp_chrono_start). 22840f87230dSFrancis Yan * If a condition stops, we only stop chrono tracking if 22850f87230dSFrancis Yan * it's the "most interesting" or current chrono we are 22860f87230dSFrancis Yan * tracking and starts busy chrono if we have pending data. 22870f87230dSFrancis Yan */ 228875c119afSEric Dumazet if (tcp_rtx_and_write_queues_empty(sk)) 228905b055e8SFrancis Yan tcp_chrono_set(tp, TCP_CHRONO_UNSPEC); 22900f87230dSFrancis Yan else if (type == tp->chrono_type) 22910f87230dSFrancis Yan tcp_chrono_set(tp, TCP_CHRONO_BUSY); 229205b055e8SFrancis Yan } 229305b055e8SFrancis Yan 22941da177e4SLinus Torvalds /* This routine writes packets to the network. It advances the 22951da177e4SLinus Torvalds * send_head. This happens as incoming acks open up the remote 22961da177e4SLinus Torvalds * window for us. 22971da177e4SLinus Torvalds * 2298f8269a49SIlpo Järvinen * LARGESEND note: !tcp_urg_mode is overkill, only frames between 2299f8269a49SIlpo Järvinen * snd_up-64k-mss .. snd_up cannot be large. However, taking into 2300f8269a49SIlpo Järvinen * account rare use of URG, this is not a big flaw. 2301f8269a49SIlpo Järvinen * 23026ba8a3b1SNandita Dukkipati * Send at most one packet when push_one > 0. Temporarily ignore 23036ba8a3b1SNandita Dukkipati * cwnd limit to force at most one packet out when push_one == 2. 23046ba8a3b1SNandita Dukkipati 2305a2a385d6SEric Dumazet * Returns true, if no segments are in flight and we have queued segments, 2306a2a385d6SEric Dumazet * but cannot send anything now because of SWS or another problem. 23071da177e4SLinus Torvalds */ 2308a2a385d6SEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 2309d5dd9175SIlpo Järvinen int push_one, gfp_t gfp) 23101da177e4SLinus Torvalds { 23111da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 231292df7b51SDavid S. Miller struct sk_buff *skb; 2313c1b4a7e6SDavid S. Miller unsigned int tso_segs, sent_pkts; 2314c1b4a7e6SDavid S. Miller int cwnd_quota; 23155d424d5aSJohn Heffner int result; 23165615f886SFrancis Yan bool is_cwnd_limited = false, is_rwnd_limited = false; 2317605ad7f1SEric Dumazet u32 max_segs; 23181da177e4SLinus Torvalds 2319c1b4a7e6SDavid S. Miller sent_pkts = 0; 23205d424d5aSJohn Heffner 2321d5dd9175SIlpo Järvinen if (!push_one) { 23225d424d5aSJohn Heffner /* Do MTU probing. */ 2323d5dd9175SIlpo Järvinen result = tcp_mtu_probe(sk); 2324d5dd9175SIlpo Järvinen if (!result) { 2325a2a385d6SEric Dumazet return false; 23265d424d5aSJohn Heffner } else if (result > 0) { 23275d424d5aSJohn Heffner sent_pkts = 1; 23285d424d5aSJohn Heffner } 2329d5dd9175SIlpo Järvinen } 23305d424d5aSJohn Heffner 2331ed6e7268SNeal Cardwell max_segs = tcp_tso_segs(sk, mss_now); 23329a568de4SEric Dumazet tcp_mstamp_refresh(tp); 2333fe067e8aSDavid S. Miller while ((skb = tcp_send_head(sk))) { 2334c8ac3774SHerbert Xu unsigned int limit; 2335c8ac3774SHerbert Xu 2336218af599SEric Dumazet if (tcp_pacing_check(sk)) 2337218af599SEric Dumazet break; 2338218af599SEric Dumazet 23395bbb432cSEric Dumazet tso_segs = tcp_init_tso_segs(skb, mss_now); 2340c1b4a7e6SDavid S. Miller BUG_ON(!tso_segs); 2341c1b4a7e6SDavid S. Miller 23429d186cacSAndrey Vagin if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) { 23437faee5c0SEric Dumazet /* "skb_mstamp" is used as a start point for the retransmit timer */ 2344e2080072SEric Dumazet tcp_update_skb_after_send(tp, skb); 2345ec342325SAndrew Vagin goto repair; /* Skip network transmission */ 23469d186cacSAndrey Vagin } 2347ec342325SAndrew Vagin 2348b68e9f85SHerbert Xu cwnd_quota = tcp_cwnd_test(tp, skb); 23496ba8a3b1SNandita Dukkipati if (!cwnd_quota) { 23506ba8a3b1SNandita Dukkipati if (push_one == 2) 23516ba8a3b1SNandita Dukkipati /* Force out a loss probe pkt. */ 23526ba8a3b1SNandita Dukkipati cwnd_quota = 1; 23536ba8a3b1SNandita Dukkipati else 2354b68e9f85SHerbert Xu break; 23556ba8a3b1SNandita Dukkipati } 2356b68e9f85SHerbert Xu 23575615f886SFrancis Yan if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) { 23585615f886SFrancis Yan is_rwnd_limited = true; 2359b68e9f85SHerbert Xu break; 23605615f886SFrancis Yan } 2361b68e9f85SHerbert Xu 2362d6a4e26aSEric Dumazet if (tso_segs == 1) { 2363aa93466bSDavid S. Miller if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 2364aa93466bSDavid S. Miller (tcp_skb_is_last(sk, skb) ? 2365aa93466bSDavid S. Miller nonagle : TCP_NAGLE_PUSH)))) 2366aa93466bSDavid S. Miller break; 2367c1b4a7e6SDavid S. Miller } else { 2368ca8a2263SNeal Cardwell if (!push_one && 2369605ad7f1SEric Dumazet tcp_tso_should_defer(sk, skb, &is_cwnd_limited, 2370605ad7f1SEric Dumazet max_segs)) 2371aa93466bSDavid S. Miller break; 2372c1b4a7e6SDavid S. Miller } 2373aa93466bSDavid S. Miller 2374605ad7f1SEric Dumazet limit = mss_now; 2375d6a4e26aSEric Dumazet if (tso_segs > 1 && !tcp_urg_mode(tp)) 2376605ad7f1SEric Dumazet limit = tcp_mss_split_point(sk, skb, mss_now, 2377605ad7f1SEric Dumazet min_t(unsigned int, 2378605ad7f1SEric Dumazet cwnd_quota, 2379605ad7f1SEric Dumazet max_segs), 2380605ad7f1SEric Dumazet nonagle); 2381605ad7f1SEric Dumazet 2382605ad7f1SEric Dumazet if (skb->len > limit && 238375c119afSEric Dumazet unlikely(tso_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE, 238475c119afSEric Dumazet skb, limit, mss_now, gfp))) 2385605ad7f1SEric Dumazet break; 2386605ad7f1SEric Dumazet 23877aa5470cSEric Dumazet if (test_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags)) 23887aa5470cSEric Dumazet clear_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags); 2389f9616c35SEric Dumazet if (tcp_small_queue_check(sk, skb, 0)) 239046d3ceabSEric Dumazet break; 2391c9eeec26SEric Dumazet 2392d5dd9175SIlpo Järvinen if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) 23931da177e4SLinus Torvalds break; 23941da177e4SLinus Torvalds 2395ec342325SAndrew Vagin repair: 23961da177e4SLinus Torvalds /* Advance the send_head. This one is sent out. 23971da177e4SLinus Torvalds * This call will increment packets_out. 23981da177e4SLinus Torvalds */ 239966f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 24001da177e4SLinus Torvalds 24011da177e4SLinus Torvalds tcp_minshall_update(tp, mss_now, skb); 2402a262f0cdSNandita Dukkipati sent_pkts += tcp_skb_pcount(skb); 2403d5dd9175SIlpo Järvinen 2404d5dd9175SIlpo Järvinen if (push_one) 2405d5dd9175SIlpo Järvinen break; 24061da177e4SLinus Torvalds } 24071da177e4SLinus Torvalds 24085615f886SFrancis Yan if (is_rwnd_limited) 24095615f886SFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED); 24105615f886SFrancis Yan else 24115615f886SFrancis Yan tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED); 24125615f886SFrancis Yan 2413aa93466bSDavid S. Miller if (likely(sent_pkts)) { 2414684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 2415684bad11SYuchung Cheng tp->prr_out += sent_pkts; 24166ba8a3b1SNandita Dukkipati 24176ba8a3b1SNandita Dukkipati /* Send one loss probe per tail loss episode. */ 24186ba8a3b1SNandita Dukkipati if (push_one != 2) 24196ba8a3b1SNandita Dukkipati tcp_schedule_loss_probe(sk); 2420d2e1339fSBendik Rønning Opstad is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd); 2421ca8a2263SNeal Cardwell tcp_cwnd_validate(sk, is_cwnd_limited); 2422a2a385d6SEric Dumazet return false; 24231da177e4SLinus Torvalds } 242475c119afSEric Dumazet return !tp->packets_out && !tcp_write_queue_empty(sk); 24256ba8a3b1SNandita Dukkipati } 24266ba8a3b1SNandita Dukkipati 24276ba8a3b1SNandita Dukkipati bool tcp_schedule_loss_probe(struct sock *sk) 24286ba8a3b1SNandita Dukkipati { 24296ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 24306ba8a3b1SNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 2431a2815817SNeal Cardwell u32 timeout, rto_delta_us; 24322ae21cf5SEric Dumazet int early_retrans; 24336ba8a3b1SNandita Dukkipati 24346ba8a3b1SNandita Dukkipati /* Don't do any loss probe on a Fast Open connection before 3WHS 24356ba8a3b1SNandita Dukkipati * finishes. 24366ba8a3b1SNandita Dukkipati */ 2437f9b99582SYuchung Cheng if (tp->fastopen_rsk) 24386ba8a3b1SNandita Dukkipati return false; 24396ba8a3b1SNandita Dukkipati 24402ae21cf5SEric Dumazet early_retrans = sock_net(sk)->ipv4.sysctl_tcp_early_retrans; 24416ba8a3b1SNandita Dukkipati /* Schedule a loss probe in 2*RTT for SACK capable connections 24426ba8a3b1SNandita Dukkipati * in Open state, that are either limited by cwnd or application. 24436ba8a3b1SNandita Dukkipati */ 24442ae21cf5SEric Dumazet if ((early_retrans != 3 && early_retrans != 4) || 2445bec41a11SYuchung Cheng !tp->packets_out || !tcp_is_sack(tp) || 2446bec41a11SYuchung Cheng icsk->icsk_ca_state != TCP_CA_Open) 24476ba8a3b1SNandita Dukkipati return false; 24486ba8a3b1SNandita Dukkipati 24496ba8a3b1SNandita Dukkipati if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) && 245075c119afSEric Dumazet !tcp_write_queue_empty(sk)) 24516ba8a3b1SNandita Dukkipati return false; 24526ba8a3b1SNandita Dukkipati 2453bb4d991aSYuchung Cheng /* Probe timeout is 2*rtt. Add minimum RTO to account 2454f9b99582SYuchung Cheng * for delayed ack when there's one outstanding packet. If no RTT 2455f9b99582SYuchung Cheng * sample is available then probe after TCP_TIMEOUT_INIT. 24566ba8a3b1SNandita Dukkipati */ 2457bb4d991aSYuchung Cheng if (tp->srtt_us) { 2458bb4d991aSYuchung Cheng timeout = usecs_to_jiffies(tp->srtt_us >> 2); 24596ba8a3b1SNandita Dukkipati if (tp->packets_out == 1) 2460bb4d991aSYuchung Cheng timeout += TCP_RTO_MIN; 2461bb4d991aSYuchung Cheng else 2462bb4d991aSYuchung Cheng timeout += TCP_TIMEOUT_MIN; 2463bb4d991aSYuchung Cheng } else { 2464bb4d991aSYuchung Cheng timeout = TCP_TIMEOUT_INIT; 2465bb4d991aSYuchung Cheng } 24666ba8a3b1SNandita Dukkipati 2467a2815817SNeal Cardwell /* If the RTO formula yields an earlier time, then use that time. */ 2468a2815817SNeal Cardwell rto_delta_us = tcp_rto_delta_us(sk); /* How far in future is RTO? */ 2469a2815817SNeal Cardwell if (rto_delta_us > 0) 2470a2815817SNeal Cardwell timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us)); 24716ba8a3b1SNandita Dukkipati 24726ba8a3b1SNandita Dukkipati inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, 24736ba8a3b1SNandita Dukkipati TCP_RTO_MAX); 24746ba8a3b1SNandita Dukkipati return true; 24756ba8a3b1SNandita Dukkipati } 24766ba8a3b1SNandita Dukkipati 24771f3279aeSEric Dumazet /* Thanks to skb fast clones, we can detect if a prior transmit of 24781f3279aeSEric Dumazet * a packet is still in a qdisc or driver queue. 24791f3279aeSEric Dumazet * In this case, there is very little point doing a retransmit ! 24801f3279aeSEric Dumazet */ 24811f3279aeSEric Dumazet static bool skb_still_in_host_queue(const struct sock *sk, 24821f3279aeSEric Dumazet const struct sk_buff *skb) 24831f3279aeSEric Dumazet { 248439bb5e62SEric Dumazet if (unlikely(skb_fclone_busy(sk, skb))) { 2485c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), 24861f3279aeSEric Dumazet LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES); 24871f3279aeSEric Dumazet return true; 24881f3279aeSEric Dumazet } 24891f3279aeSEric Dumazet return false; 24901f3279aeSEric Dumazet } 24911f3279aeSEric Dumazet 2492b340b264SYuchung Cheng /* When probe timeout (PTO) fires, try send a new segment if possible, else 24936ba8a3b1SNandita Dukkipati * retransmit the last segment. 24946ba8a3b1SNandita Dukkipati */ 24956ba8a3b1SNandita Dukkipati void tcp_send_loss_probe(struct sock *sk) 24966ba8a3b1SNandita Dukkipati { 24979b717a8dSNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 24986ba8a3b1SNandita Dukkipati struct sk_buff *skb; 24996ba8a3b1SNandita Dukkipati int pcount; 25006ba8a3b1SNandita Dukkipati int mss = tcp_current_mss(sk); 25016ba8a3b1SNandita Dukkipati 2502b340b264SYuchung Cheng skb = tcp_send_head(sk); 250375c119afSEric Dumazet if (skb && tcp_snd_wnd_test(tp, skb, mss)) { 2504b340b264SYuchung Cheng pcount = tp->packets_out; 2505b340b264SYuchung Cheng tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC); 2506b340b264SYuchung Cheng if (tp->packets_out > pcount) 2507b340b264SYuchung Cheng goto probe_sent; 25086ba8a3b1SNandita Dukkipati goto rearm_timer; 25096ba8a3b1SNandita Dukkipati } 251075c119afSEric Dumazet skb = skb_rb_last(&sk->tcp_rtx_queue); 25116ba8a3b1SNandita Dukkipati 25129b717a8dSNandita Dukkipati /* At most one outstanding TLP retransmission. */ 25139b717a8dSNandita Dukkipati if (tp->tlp_high_seq) 25149b717a8dSNandita Dukkipati goto rearm_timer; 25159b717a8dSNandita Dukkipati 25166ba8a3b1SNandita Dukkipati /* Retransmit last segment. */ 25176ba8a3b1SNandita Dukkipati if (WARN_ON(!skb)) 25186ba8a3b1SNandita Dukkipati goto rearm_timer; 25196ba8a3b1SNandita Dukkipati 25201f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 25211f3279aeSEric Dumazet goto rearm_timer; 25221f3279aeSEric Dumazet 25236ba8a3b1SNandita Dukkipati pcount = tcp_skb_pcount(skb); 25246ba8a3b1SNandita Dukkipati if (WARN_ON(!pcount)) 25256ba8a3b1SNandita Dukkipati goto rearm_timer; 25266ba8a3b1SNandita Dukkipati 25276ba8a3b1SNandita Dukkipati if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) { 252875c119afSEric Dumazet if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, 252975c119afSEric Dumazet (pcount - 1) * mss, mss, 25306cc55e09SOctavian Purdila GFP_ATOMIC))) 25316ba8a3b1SNandita Dukkipati goto rearm_timer; 253275c119afSEric Dumazet skb = skb_rb_next(skb); 25336ba8a3b1SNandita Dukkipati } 25346ba8a3b1SNandita Dukkipati 25356ba8a3b1SNandita Dukkipati if (WARN_ON(!skb || !tcp_skb_pcount(skb))) 25366ba8a3b1SNandita Dukkipati goto rearm_timer; 25376ba8a3b1SNandita Dukkipati 253810d3be56SEric Dumazet if (__tcp_retransmit_skb(sk, skb, 1)) 2539b340b264SYuchung Cheng goto rearm_timer; 25406ba8a3b1SNandita Dukkipati 25419b717a8dSNandita Dukkipati /* Record snd_nxt for loss detection. */ 25429b717a8dSNandita Dukkipati tp->tlp_high_seq = tp->snd_nxt; 25439b717a8dSNandita Dukkipati 2544b340b264SYuchung Cheng probe_sent: 2545c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES); 2546fcd16c0aSYuchung Cheng /* Reset s.t. tcp_rearm_rto will restart timer from now */ 2547fcd16c0aSYuchung Cheng inet_csk(sk)->icsk_pending = 0; 2548b340b264SYuchung Cheng rearm_timer: 2549fcd16c0aSYuchung Cheng tcp_rearm_rto(sk); 25501da177e4SLinus Torvalds } 25511da177e4SLinus Torvalds 2552a762a980SDavid S. Miller /* Push out any pending frames which were held back due to 2553a762a980SDavid S. Miller * TCP_CORK or attempt at coalescing tiny packets. 2554a762a980SDavid S. Miller * The socket must be locked by the caller. 2555a762a980SDavid S. Miller */ 25569e412ba7SIlpo Järvinen void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 25579e412ba7SIlpo Järvinen int nonagle) 2558a762a980SDavid S. Miller { 2559726e07a8SIlpo Järvinen /* If we are closed, the bytes will have to remain here. 2560726e07a8SIlpo Järvinen * In time closedown will finish, we empty the write queue and 2561726e07a8SIlpo Järvinen * all will be happy. 2562726e07a8SIlpo Järvinen */ 2563726e07a8SIlpo Järvinen if (unlikely(sk->sk_state == TCP_CLOSE)) 2564726e07a8SIlpo Järvinen return; 2565726e07a8SIlpo Järvinen 256699a1dec7SMel Gorman if (tcp_write_xmit(sk, cur_mss, nonagle, 0, 25677450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC))) 25689e412ba7SIlpo Järvinen tcp_check_probe_timer(sk); 2569a762a980SDavid S. Miller } 2570a762a980SDavid S. Miller 2571c1b4a7e6SDavid S. Miller /* Send _single_ skb sitting at the send head. This function requires 2572c1b4a7e6SDavid S. Miller * true push pending frames to setup probe timer etc. 2573c1b4a7e6SDavid S. Miller */ 2574c1b4a7e6SDavid S. Miller void tcp_push_one(struct sock *sk, unsigned int mss_now) 2575c1b4a7e6SDavid S. Miller { 2576fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 2577c1b4a7e6SDavid S. Miller 2578c1b4a7e6SDavid S. Miller BUG_ON(!skb || skb->len < mss_now); 2579c1b4a7e6SDavid S. Miller 2580d5dd9175SIlpo Järvinen tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); 2581c1b4a7e6SDavid S. Miller } 2582c1b4a7e6SDavid S. Miller 25831da177e4SLinus Torvalds /* This function returns the amount that we can raise the 25841da177e4SLinus Torvalds * usable window based on the following constraints 25851da177e4SLinus Torvalds * 25861da177e4SLinus Torvalds * 1. The window can never be shrunk once it is offered (RFC 793) 25871da177e4SLinus Torvalds * 2. We limit memory per socket 25881da177e4SLinus Torvalds * 25891da177e4SLinus Torvalds * RFC 1122: 25901da177e4SLinus Torvalds * "the suggested [SWS] avoidance algorithm for the receiver is to keep 25911da177e4SLinus Torvalds * RECV.NEXT + RCV.WIN fixed until: 25921da177e4SLinus Torvalds * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 25931da177e4SLinus Torvalds * 25941da177e4SLinus Torvalds * i.e. don't raise the right edge of the window until you can raise 25951da177e4SLinus Torvalds * it at least MSS bytes. 25961da177e4SLinus Torvalds * 25971da177e4SLinus Torvalds * Unfortunately, the recommended algorithm breaks header prediction, 25981da177e4SLinus Torvalds * since header prediction assumes th->window stays fixed. 25991da177e4SLinus Torvalds * 26001da177e4SLinus Torvalds * Strictly speaking, keeping th->window fixed violates the receiver 26011da177e4SLinus Torvalds * side SWS prevention criteria. The problem is that under this rule 26021da177e4SLinus Torvalds * a stream of single byte packets will cause the right side of the 26031da177e4SLinus Torvalds * window to always advance by a single byte. 26041da177e4SLinus Torvalds * 26051da177e4SLinus Torvalds * Of course, if the sender implements sender side SWS prevention 26061da177e4SLinus Torvalds * then this will not be a problem. 26071da177e4SLinus Torvalds * 26081da177e4SLinus Torvalds * BSD seems to make the following compromise: 26091da177e4SLinus Torvalds * 26101da177e4SLinus Torvalds * If the free space is less than the 1/4 of the maximum 26111da177e4SLinus Torvalds * space available and the free space is less than 1/2 mss, 26121da177e4SLinus Torvalds * then set the window to 0. 26131da177e4SLinus Torvalds * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 26141da177e4SLinus Torvalds * Otherwise, just prevent the window from shrinking 26151da177e4SLinus Torvalds * and from being larger than the largest representable value. 26161da177e4SLinus Torvalds * 26171da177e4SLinus Torvalds * This prevents incremental opening of the window in the regime 26181da177e4SLinus Torvalds * where TCP is limited by the speed of the reader side taking 26191da177e4SLinus Torvalds * data out of the TCP receive queue. It does nothing about 26201da177e4SLinus Torvalds * those cases where the window is constrained on the sender side 26211da177e4SLinus Torvalds * because the pipeline is full. 26221da177e4SLinus Torvalds * 26231da177e4SLinus Torvalds * BSD also seems to "accidentally" limit itself to windows that are a 26241da177e4SLinus Torvalds * multiple of MSS, at least until the free space gets quite small. 26251da177e4SLinus Torvalds * This would appear to be a side effect of the mbuf implementation. 26261da177e4SLinus Torvalds * Combining these two algorithms results in the observed behavior 26271da177e4SLinus Torvalds * of having a fixed window size at almost all times. 26281da177e4SLinus Torvalds * 26291da177e4SLinus Torvalds * Below we obtain similar behavior by forcing the offered window to 26301da177e4SLinus Torvalds * a multiple of the mss when it is feasible to do so. 26311da177e4SLinus Torvalds * 26321da177e4SLinus Torvalds * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 26331da177e4SLinus Torvalds * Regular options like TIMESTAMP are taken into account. 26341da177e4SLinus Torvalds */ 26351da177e4SLinus Torvalds u32 __tcp_select_window(struct sock *sk) 26361da177e4SLinus Torvalds { 2637463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 26381da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2639caa20d9aSStephen Hemminger /* MSS for the peer's data. Previous versions used mss_clamp 26401da177e4SLinus Torvalds * here. I don't know if the value based on our guesses 26411da177e4SLinus Torvalds * of peer's MSS is better for the performance. It's more correct 26421da177e4SLinus Torvalds * but may be worse for the performance because of rcv_mss 26431da177e4SLinus Torvalds * fluctuations. --SAW 1998/11/1 26441da177e4SLinus Torvalds */ 2645463c84b9SArnaldo Carvalho de Melo int mss = icsk->icsk_ack.rcv_mss; 26461da177e4SLinus Torvalds int free_space = tcp_space(sk); 264786c1a045SFlorian Westphal int allowed_space = tcp_full_space(sk); 264886c1a045SFlorian Westphal int full_space = min_t(int, tp->window_clamp, allowed_space); 26491da177e4SLinus Torvalds int window; 26501da177e4SLinus Torvalds 265106425c30SEric Dumazet if (unlikely(mss > full_space)) { 26521da177e4SLinus Torvalds mss = full_space; 265306425c30SEric Dumazet if (mss <= 0) 265406425c30SEric Dumazet return 0; 265506425c30SEric Dumazet } 2656b92edbe0SEric Dumazet if (free_space < (full_space >> 1)) { 2657463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.quick = 0; 26581da177e4SLinus Torvalds 2659b8da51ebSEric Dumazet if (tcp_under_memory_pressure(sk)) 2660056834d9SIlpo Järvinen tp->rcv_ssthresh = min(tp->rcv_ssthresh, 2661056834d9SIlpo Järvinen 4U * tp->advmss); 26621da177e4SLinus Torvalds 266386c1a045SFlorian Westphal /* free_space might become our new window, make sure we don't 266486c1a045SFlorian Westphal * increase it due to wscale. 266586c1a045SFlorian Westphal */ 266686c1a045SFlorian Westphal free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale); 266786c1a045SFlorian Westphal 266886c1a045SFlorian Westphal /* if free space is less than mss estimate, or is below 1/16th 266986c1a045SFlorian Westphal * of the maximum allowed, try to move to zero-window, else 267086c1a045SFlorian Westphal * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and 267186c1a045SFlorian Westphal * new incoming data is dropped due to memory limits. 267286c1a045SFlorian Westphal * With large window, mss test triggers way too late in order 267386c1a045SFlorian Westphal * to announce zero window in time before rmem limit kicks in. 267486c1a045SFlorian Westphal */ 267586c1a045SFlorian Westphal if (free_space < (allowed_space >> 4) || free_space < mss) 26761da177e4SLinus Torvalds return 0; 26771da177e4SLinus Torvalds } 26781da177e4SLinus Torvalds 26791da177e4SLinus Torvalds if (free_space > tp->rcv_ssthresh) 26801da177e4SLinus Torvalds free_space = tp->rcv_ssthresh; 26811da177e4SLinus Torvalds 26821da177e4SLinus Torvalds /* Don't do rounding if we are using window scaling, since the 26831da177e4SLinus Torvalds * scaled window will not line up with the MSS boundary anyway. 26841da177e4SLinus Torvalds */ 26851da177e4SLinus Torvalds if (tp->rx_opt.rcv_wscale) { 26861da177e4SLinus Torvalds window = free_space; 26871da177e4SLinus Torvalds 26881da177e4SLinus Torvalds /* Advertise enough space so that it won't get scaled away. 26891da177e4SLinus Torvalds * Import case: prevent zero window announcement if 26901da177e4SLinus Torvalds * 1<<rcv_wscale > mss. 26911da177e4SLinus Torvalds */ 26921935299dSGao Feng window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale)); 26931da177e4SLinus Torvalds } else { 26941935299dSGao Feng window = tp->rcv_wnd; 26951da177e4SLinus Torvalds /* Get the largest window that is a nice multiple of mss. 26961da177e4SLinus Torvalds * Window clamp already applied above. 26971da177e4SLinus Torvalds * If our current window offering is within 1 mss of the 26981da177e4SLinus Torvalds * free space we just keep it. This prevents the divide 26991da177e4SLinus Torvalds * and multiply from happening most of the time. 27001da177e4SLinus Torvalds * We also don't do any window rounding when the free space 27011da177e4SLinus Torvalds * is too small. 27021da177e4SLinus Torvalds */ 27031da177e4SLinus Torvalds if (window <= free_space - mss || window > free_space) 27041935299dSGao Feng window = rounddown(free_space, mss); 270584565070SJohn Heffner else if (mss == full_space && 2706b92edbe0SEric Dumazet free_space > window + (full_space >> 1)) 270784565070SJohn Heffner window = free_space; 27081da177e4SLinus Torvalds } 27091da177e4SLinus Torvalds 27101da177e4SLinus Torvalds return window; 27111da177e4SLinus Torvalds } 27121da177e4SLinus Torvalds 2713cfea5a68SMartin KaFai Lau void tcp_skb_collapse_tstamp(struct sk_buff *skb, 2714082ac2d5SMartin KaFai Lau const struct sk_buff *next_skb) 2715082ac2d5SMartin KaFai Lau { 27160a2cf20cSSoheil Hassas Yeganeh if (unlikely(tcp_has_tx_tstamp(next_skb))) { 27170a2cf20cSSoheil Hassas Yeganeh const struct skb_shared_info *next_shinfo = 27180a2cf20cSSoheil Hassas Yeganeh skb_shinfo(next_skb); 2719082ac2d5SMartin KaFai Lau struct skb_shared_info *shinfo = skb_shinfo(skb); 2720082ac2d5SMartin KaFai Lau 27210a2cf20cSSoheil Hassas Yeganeh shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP; 2722082ac2d5SMartin KaFai Lau shinfo->tskey = next_shinfo->tskey; 27232de8023eSMartin KaFai Lau TCP_SKB_CB(skb)->txstamp_ack |= 27242de8023eSMartin KaFai Lau TCP_SKB_CB(next_skb)->txstamp_ack; 2725082ac2d5SMartin KaFai Lau } 2726082ac2d5SMartin KaFai Lau } 2727082ac2d5SMartin KaFai Lau 27284a17fc3aSIlpo Järvinen /* Collapses two adjacent SKB's during retransmission. */ 2729f8071cdeSEric Dumazet static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) 27301da177e4SLinus Torvalds { 27311da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 273275c119afSEric Dumazet struct sk_buff *next_skb = skb_rb_next(skb); 2733058dc334SIlpo Järvinen int skb_size, next_skb_size; 27341da177e4SLinus Torvalds 2735058dc334SIlpo Järvinen skb_size = skb->len; 2736058dc334SIlpo Järvinen next_skb_size = next_skb->len; 27371da177e4SLinus Torvalds 2738058dc334SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); 27391da177e4SLinus Torvalds 2740f8071cdeSEric Dumazet if (next_skb_size) { 2741f8071cdeSEric Dumazet if (next_skb_size <= skb_availroom(skb)) 2742f8071cdeSEric Dumazet skb_copy_bits(next_skb, 0, skb_put(skb, next_skb_size), 2743f8071cdeSEric Dumazet next_skb_size); 2744f8071cdeSEric Dumazet else if (!skb_shift(skb, next_skb, next_skb_size)) 2745f8071cdeSEric Dumazet return false; 2746f8071cdeSEric Dumazet } 27476859d494SIlpo Järvinen tcp_highest_sack_combine(sk, next_skb, skb); 2748a6963a6bSIlpo Järvinen 274952d570aaSJarek Poplawski if (next_skb->ip_summed == CHECKSUM_PARTIAL) 275052d570aaSJarek Poplawski skb->ip_summed = CHECKSUM_PARTIAL; 27511da177e4SLinus Torvalds 275284fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 27531da177e4SLinus Torvalds skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); 27541da177e4SLinus Torvalds 27551da177e4SLinus Torvalds /* Update sequence range on original skb. */ 27561da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 27571da177e4SLinus Torvalds 2758e6c7d085SIlpo Järvinen /* Merge over control information. This moves PSH/FIN etc. over */ 27594de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; 27601da177e4SLinus Torvalds 27611da177e4SLinus Torvalds /* All done, get rid of second SKB and account for it so 27621da177e4SLinus Torvalds * packet counting does not break. 27631da177e4SLinus Torvalds */ 27644828e7f4SIlpo Järvinen TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; 2765a643b5d4SMartin KaFai Lau TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor; 2766b7689205SIlpo Järvinen 2767b7689205SIlpo Järvinen /* changed transmit queue under us so clear hints */ 2768ef9da47cSIlpo Järvinen tcp_clear_retrans_hints_partial(tp); 2769ef9da47cSIlpo Järvinen if (next_skb == tp->retransmit_skb_hint) 2770ef9da47cSIlpo Järvinen tp->retransmit_skb_hint = skb; 2771b7689205SIlpo Järvinen 2772797108d1SIlpo Järvinen tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); 2773797108d1SIlpo Järvinen 2774082ac2d5SMartin KaFai Lau tcp_skb_collapse_tstamp(skb, next_skb); 2775082ac2d5SMartin KaFai Lau 277675c119afSEric Dumazet tcp_rtx_queue_unlink_and_free(next_skb, sk); 2777f8071cdeSEric Dumazet return true; 27781da177e4SLinus Torvalds } 27791da177e4SLinus Torvalds 278067edfef7SAndi Kleen /* Check if coalescing SKBs is legal. */ 2781a2a385d6SEric Dumazet static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) 27824a17fc3aSIlpo Järvinen { 27834a17fc3aSIlpo Järvinen if (tcp_skb_pcount(skb) > 1) 2784a2a385d6SEric Dumazet return false; 27854a17fc3aSIlpo Järvinen if (skb_cloned(skb)) 2786a2a385d6SEric Dumazet return false; 27872331ccc5SEric Dumazet /* Some heuristics for collapsing over SACK'd could be invented */ 27884a17fc3aSIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 2789a2a385d6SEric Dumazet return false; 27904a17fc3aSIlpo Järvinen 2791a2a385d6SEric Dumazet return true; 27924a17fc3aSIlpo Järvinen } 27934a17fc3aSIlpo Järvinen 279467edfef7SAndi Kleen /* Collapse packets in the retransmit queue to make to create 279567edfef7SAndi Kleen * less packets on the wire. This is only done on retransmission. 279667edfef7SAndi Kleen */ 27974a17fc3aSIlpo Järvinen static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, 27984a17fc3aSIlpo Järvinen int space) 27994a17fc3aSIlpo Järvinen { 28004a17fc3aSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 28014a17fc3aSIlpo Järvinen struct sk_buff *skb = to, *tmp; 2802a2a385d6SEric Dumazet bool first = true; 28034a17fc3aSIlpo Järvinen 2804*e0a1e5b5SEric Dumazet if (!sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse) 28054a17fc3aSIlpo Järvinen return; 28064de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 28074a17fc3aSIlpo Järvinen return; 28084a17fc3aSIlpo Järvinen 280975c119afSEric Dumazet skb_rbtree_walk_from_safe(skb, tmp) { 28104a17fc3aSIlpo Järvinen if (!tcp_can_collapse(sk, skb)) 28114a17fc3aSIlpo Järvinen break; 28124a17fc3aSIlpo Järvinen 2813a643b5d4SMartin KaFai Lau if (!tcp_skb_can_collapse_to(to)) 2814a643b5d4SMartin KaFai Lau break; 2815a643b5d4SMartin KaFai Lau 28164a17fc3aSIlpo Järvinen space -= skb->len; 28174a17fc3aSIlpo Järvinen 28184a17fc3aSIlpo Järvinen if (first) { 2819a2a385d6SEric Dumazet first = false; 28204a17fc3aSIlpo Järvinen continue; 28214a17fc3aSIlpo Järvinen } 28224a17fc3aSIlpo Järvinen 28234a17fc3aSIlpo Järvinen if (space < 0) 28244a17fc3aSIlpo Järvinen break; 28254a17fc3aSIlpo Järvinen 28264a17fc3aSIlpo Järvinen if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) 28274a17fc3aSIlpo Järvinen break; 28284a17fc3aSIlpo Järvinen 2829f8071cdeSEric Dumazet if (!tcp_collapse_retrans(sk, to)) 2830f8071cdeSEric Dumazet break; 28314a17fc3aSIlpo Järvinen } 28324a17fc3aSIlpo Järvinen } 28334a17fc3aSIlpo Järvinen 28341da177e4SLinus Torvalds /* This retransmits one SKB. Policy decisions and retransmit queue 28351da177e4SLinus Torvalds * state updates are done by the caller. Returns non-zero if an 28361da177e4SLinus Torvalds * error occurred which prevented the send. 28371da177e4SLinus Torvalds */ 283810d3be56SEric Dumazet int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) 28391da177e4SLinus Torvalds { 28405d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 284110d3be56SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 28427d227cd2SSridhar Samudrala unsigned int cur_mss; 284310d3be56SEric Dumazet int diff, len, err; 28441da177e4SLinus Torvalds 284510d3be56SEric Dumazet 284610d3be56SEric Dumazet /* Inconclusive MTU probe */ 284710d3be56SEric Dumazet if (icsk->icsk_mtup.probe_size) 28485d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 28495d424d5aSJohn Heffner 28501da177e4SLinus Torvalds /* Do not sent more than we queued. 1/4 is reserved for possible 2851caa20d9aSStephen Hemminger * copying overhead: fragmentation, tunneling, mangling etc. 28521da177e4SLinus Torvalds */ 285314afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) > 2854ffb4d6c8SEric Dumazet min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), 2855ffb4d6c8SEric Dumazet sk->sk_sndbuf)) 28561da177e4SLinus Torvalds return -EAGAIN; 28571da177e4SLinus Torvalds 28581f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 28591f3279aeSEric Dumazet return -EBUSY; 28601f3279aeSEric Dumazet 28611da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 28621da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 28631da177e4SLinus Torvalds BUG(); 28641da177e4SLinus Torvalds if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 28651da177e4SLinus Torvalds return -ENOMEM; 28661da177e4SLinus Torvalds } 28671da177e4SLinus Torvalds 28687d227cd2SSridhar Samudrala if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 28697d227cd2SSridhar Samudrala return -EHOSTUNREACH; /* Routing failure or similar. */ 28707d227cd2SSridhar Samudrala 28710c54b85fSIlpo Järvinen cur_mss = tcp_current_mss(sk); 28727d227cd2SSridhar Samudrala 28731da177e4SLinus Torvalds /* If receiver has shrunk his window, and skb is out of 28741da177e4SLinus Torvalds * new window, do not retransmit it. The exception is the 28751da177e4SLinus Torvalds * case, when window is shrunk to zero. In this case 28761da177e4SLinus Torvalds * our retransmit serves as a zero window probe. 28771da177e4SLinus Torvalds */ 28789d4fb27dSJoe Perches if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && 28799d4fb27dSJoe Perches TCP_SKB_CB(skb)->seq != tp->snd_una) 28801da177e4SLinus Torvalds return -EAGAIN; 28811da177e4SLinus Torvalds 288210d3be56SEric Dumazet len = cur_mss * segs; 288310d3be56SEric Dumazet if (skb->len > len) { 288475c119afSEric Dumazet if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len, 288575c119afSEric Dumazet cur_mss, GFP_ATOMIC)) 28861da177e4SLinus Torvalds return -ENOMEM; /* We'll try again later. */ 288702276f3cSIlpo Järvinen } else { 2888c52e2421SEric Dumazet if (skb_unclone(skb, GFP_ATOMIC)) 2889c52e2421SEric Dumazet return -ENOMEM; 289010d3be56SEric Dumazet 289110d3be56SEric Dumazet diff = tcp_skb_pcount(skb); 289210d3be56SEric Dumazet tcp_set_skb_tso_segs(skb, cur_mss); 289310d3be56SEric Dumazet diff -= tcp_skb_pcount(skb); 289410d3be56SEric Dumazet if (diff) 289510d3be56SEric Dumazet tcp_adjust_pcount(sk, skb, diff); 289610d3be56SEric Dumazet if (skb->len < cur_mss) 289710d3be56SEric Dumazet tcp_retrans_try_collapse(sk, skb, cur_mss); 28981da177e4SLinus Torvalds } 28991da177e4SLinus Torvalds 290049213555SDaniel Borkmann /* RFC3168, section 6.1.1.1. ECN fallback */ 290149213555SDaniel Borkmann if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN) 290249213555SDaniel Borkmann tcp_ecn_clear_syn(sk, skb); 290349213555SDaniel Borkmann 2904678550c6SYuchung Cheng /* Update global and local TCP statistics. */ 2905678550c6SYuchung Cheng segs = tcp_skb_pcount(skb); 2906678550c6SYuchung Cheng TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs); 2907678550c6SYuchung Cheng if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 2908678550c6SYuchung Cheng __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 2909678550c6SYuchung Cheng tp->total_retrans += segs; 2910678550c6SYuchung Cheng 291150bceae9SThomas Graf /* make sure skb->data is aligned on arches that require it 291250bceae9SThomas Graf * and check if ack-trimming & collapsing extended the headroom 291350bceae9SThomas Graf * beyond what csum_start can cover. 291450bceae9SThomas Graf */ 291550bceae9SThomas Graf if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) || 291650bceae9SThomas Graf skb_headroom(skb) >= 0xFFFF)) { 291710a81980SEric Dumazet struct sk_buff *nskb; 291810a81980SEric Dumazet 2919e2080072SEric Dumazet tcp_skb_tsorted_save(skb) { 292010a81980SEric Dumazet nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC); 2921c84a5711SYuchung Cheng err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) : 2922117632e6SEric Dumazet -ENOBUFS; 2923e2080072SEric Dumazet } tcp_skb_tsorted_restore(skb); 2924e2080072SEric Dumazet 29258c72c65bSEric Dumazet if (!err) 2926e2080072SEric Dumazet tcp_update_skb_after_send(tp, skb); 2927117632e6SEric Dumazet } else { 2928c84a5711SYuchung Cheng err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 2929117632e6SEric Dumazet } 2930c84a5711SYuchung Cheng 2931fc9f3501SEric Dumazet if (likely(!err)) { 2932c84a5711SYuchung Cheng TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS; 2933e086101bSCong Wang trace_tcp_retransmit_skb(sk, skb); 2934678550c6SYuchung Cheng } else if (err != -EBUSY) { 2935678550c6SYuchung Cheng NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL); 2936fc9f3501SEric Dumazet } 2937c84a5711SYuchung Cheng return err; 293893b174adSYuchung Cheng } 293993b174adSYuchung Cheng 294010d3be56SEric Dumazet int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) 294193b174adSYuchung Cheng { 294293b174adSYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 294310d3be56SEric Dumazet int err = __tcp_retransmit_skb(sk, skb, segs); 29441da177e4SLinus Torvalds 29451da177e4SLinus Torvalds if (err == 0) { 29461da177e4SLinus Torvalds #if FASTRETRANS_DEBUG > 0 29471da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 2948e87cc472SJoe Perches net_dbg_ratelimited("retrans_out leaked\n"); 29491da177e4SLinus Torvalds } 29501da177e4SLinus Torvalds #endif 29511da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 29521da177e4SLinus Torvalds tp->retrans_out += tcp_skb_pcount(skb); 29531da177e4SLinus Torvalds 29541da177e4SLinus Torvalds /* Save stamp of the first retransmit. */ 29551da177e4SLinus Torvalds if (!tp->retrans_stamp) 29567faee5c0SEric Dumazet tp->retrans_stamp = tcp_skb_timestamp(skb); 29571da177e4SLinus Torvalds 29581da177e4SLinus Torvalds } 29596e08d5e3SYuchung Cheng 29606e08d5e3SYuchung Cheng if (tp->undo_retrans < 0) 29616e08d5e3SYuchung Cheng tp->undo_retrans = 0; 29626e08d5e3SYuchung Cheng tp->undo_retrans += tcp_skb_pcount(skb); 29631da177e4SLinus Torvalds return err; 29641da177e4SLinus Torvalds } 29651da177e4SLinus Torvalds 29661da177e4SLinus Torvalds /* This gets called after a retransmit timeout, and the initially 29671da177e4SLinus Torvalds * retransmitted data is acknowledged. It tries to continue 29681da177e4SLinus Torvalds * resending the rest of the retransmit queue, until either 29691da177e4SLinus Torvalds * we've sent it all or the congestion window limit is reached. 29701da177e4SLinus Torvalds * If doing SACK, the first ACK which comes back for a timeout 29711da177e4SLinus Torvalds * based retransmit packet might feed us FACK information again. 29721da177e4SLinus Torvalds * If so, we use it to avoid unnecessarily retransmissions. 29731da177e4SLinus Torvalds */ 29741da177e4SLinus Torvalds void tcp_xmit_retransmit_queue(struct sock *sk) 29751da177e4SLinus Torvalds { 29766687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 2977b9f1f1ceSEric Dumazet struct sk_buff *skb, *rtx_head, *hole = NULL; 29781da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2979840a3cbeSYuchung Cheng u32 max_segs; 298061eb55f4SIlpo Järvinen int mib_idx; 29816a438bbeSStephen Hemminger 298245e77d31SIlpo Järvinen if (!tp->packets_out) 298345e77d31SIlpo Järvinen return; 298445e77d31SIlpo Järvinen 298575c119afSEric Dumazet rtx_head = tcp_rtx_queue_head(sk); 2986b9f1f1ceSEric Dumazet skb = tp->retransmit_skb_hint ?: rtx_head; 2987ed6e7268SNeal Cardwell max_segs = tcp_tso_segs(sk, tcp_current_mss(sk)); 298875c119afSEric Dumazet skb_rbtree_walk_from(skb) { 2989dca0aaf8SEric Dumazet __u8 sacked; 299010d3be56SEric Dumazet int segs; 29911da177e4SLinus Torvalds 2992218af599SEric Dumazet if (tcp_pacing_check(sk)) 2993218af599SEric Dumazet break; 2994218af599SEric Dumazet 29956a438bbeSStephen Hemminger /* we could do better than to assign each time */ 299651456b29SIan Morris if (!hole) 29976a438bbeSStephen Hemminger tp->retransmit_skb_hint = skb; 29986a438bbeSStephen Hemminger 299910d3be56SEric Dumazet segs = tp->snd_cwnd - tcp_packets_in_flight(tp); 300010d3be56SEric Dumazet if (segs <= 0) 30011da177e4SLinus Torvalds return; 3002dca0aaf8SEric Dumazet sacked = TCP_SKB_CB(skb)->sacked; 3003a3d2e9f8SEric Dumazet /* In case tcp_shift_skb_data() have aggregated large skbs, 3004a3d2e9f8SEric Dumazet * we need to make sure not sending too bigs TSO packets 3005a3d2e9f8SEric Dumazet */ 3006a3d2e9f8SEric Dumazet segs = min_t(int, segs, max_segs); 30070e1c54c2SIlpo Järvinen 3008840a3cbeSYuchung Cheng if (tp->retrans_out >= tp->lost_out) { 3009006f582cSIlpo Järvinen break; 30100e1c54c2SIlpo Järvinen } else if (!(sacked & TCPCB_LOST)) { 301151456b29SIan Morris if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) 30120e1c54c2SIlpo Järvinen hole = skb; 301361eb55f4SIlpo Järvinen continue; 30141da177e4SLinus Torvalds 30150e1c54c2SIlpo Järvinen } else { 30160e1c54c2SIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Loss) 30170e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFASTRETRANS; 30180e1c54c2SIlpo Järvinen else 30190e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; 30200e1c54c2SIlpo Järvinen } 30210e1c54c2SIlpo Järvinen 30220e1c54c2SIlpo Järvinen if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) 302361eb55f4SIlpo Järvinen continue; 302440b215e5SPavel Emelyanov 3025f9616c35SEric Dumazet if (tcp_small_queue_check(sk, skb, 1)) 3026f9616c35SEric Dumazet return; 3027f9616c35SEric Dumazet 302810d3be56SEric Dumazet if (tcp_retransmit_skb(sk, skb, segs)) 30291da177e4SLinus Torvalds return; 303024ab6becSYuchung Cheng 3031de1d6578SYuchung Cheng NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb)); 30321da177e4SLinus Torvalds 3033684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 3034a262f0cdSNandita Dukkipati tp->prr_out += tcp_skb_pcount(skb); 3035a262f0cdSNandita Dukkipati 303675c119afSEric Dumazet if (skb == rtx_head && 303757dde7f7SYuchung Cheng icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT) 3038463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 30393f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, 30403f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 30411da177e4SLinus Torvalds } 30421da177e4SLinus Torvalds } 30431da177e4SLinus Torvalds 3044d83769a5SEric Dumazet /* We allow to exceed memory limits for FIN packets to expedite 3045d83769a5SEric Dumazet * connection tear down and (memory) recovery. 3046845704a5SEric Dumazet * Otherwise tcp_send_fin() could be tempted to either delay FIN 3047845704a5SEric Dumazet * or even be forced to close flow without any FIN. 3048a6c5ea4cSEric Dumazet * In general, we want to allow one skb per socket to avoid hangs 3049a6c5ea4cSEric Dumazet * with edge trigger epoll() 3050d83769a5SEric Dumazet */ 3051a6c5ea4cSEric Dumazet void sk_forced_mem_schedule(struct sock *sk, int size) 3052d83769a5SEric Dumazet { 3053e805605cSJohannes Weiner int amt; 3054d83769a5SEric Dumazet 3055d83769a5SEric Dumazet if (size <= sk->sk_forward_alloc) 3056d83769a5SEric Dumazet return; 3057d83769a5SEric Dumazet amt = sk_mem_pages(size); 3058d83769a5SEric Dumazet sk->sk_forward_alloc += amt * SK_MEM_QUANTUM; 3059e805605cSJohannes Weiner sk_memory_allocated_add(sk, amt); 3060e805605cSJohannes Weiner 3061baac50bbSJohannes Weiner if (mem_cgroup_sockets_enabled && sk->sk_memcg) 3062baac50bbSJohannes Weiner mem_cgroup_charge_skmem(sk->sk_memcg, amt); 3063d83769a5SEric Dumazet } 3064d83769a5SEric Dumazet 3065845704a5SEric Dumazet /* Send a FIN. The caller locks the socket for us. 3066845704a5SEric Dumazet * We should try to send a FIN packet really hard, but eventually give up. 30671da177e4SLinus Torvalds */ 30681da177e4SLinus Torvalds void tcp_send_fin(struct sock *sk) 30691da177e4SLinus Torvalds { 3070845704a5SEric Dumazet struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk); 30711da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 30721da177e4SLinus Torvalds 3073845704a5SEric Dumazet /* Optimization, tack on the FIN if we have one skb in write queue and 3074845704a5SEric Dumazet * this skb was not yet sent, or we are under memory pressure. 3075845704a5SEric Dumazet * Note: in the latter case, FIN packet will be sent after a timeout, 3076845704a5SEric Dumazet * as TCP stack thinks it has already been transmitted. 30771da177e4SLinus Torvalds */ 307875c119afSEric Dumazet if (!tskb && tcp_under_memory_pressure(sk)) 307975c119afSEric Dumazet tskb = skb_rb_last(&sk->tcp_rtx_queue); 308075c119afSEric Dumazet 308175c119afSEric Dumazet if (tskb) { 3082845704a5SEric Dumazet coalesce: 3083845704a5SEric Dumazet TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN; 3084845704a5SEric Dumazet TCP_SKB_CB(tskb)->end_seq++; 30851da177e4SLinus Torvalds tp->write_seq++; 308675c119afSEric Dumazet if (tcp_write_queue_empty(sk)) { 3087845704a5SEric Dumazet /* This means tskb was already sent. 3088845704a5SEric Dumazet * Pretend we included the FIN on previous transmit. 3089845704a5SEric Dumazet * We need to set tp->snd_nxt to the value it would have 3090845704a5SEric Dumazet * if FIN had been sent. This is because retransmit path 3091845704a5SEric Dumazet * does not change tp->snd_nxt. 3092845704a5SEric Dumazet */ 3093845704a5SEric Dumazet tp->snd_nxt++; 3094845704a5SEric Dumazet return; 3095845704a5SEric Dumazet } 30961da177e4SLinus Torvalds } else { 3097845704a5SEric Dumazet skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation); 3098845704a5SEric Dumazet if (unlikely(!skb)) { 3099845704a5SEric Dumazet if (tskb) 3100845704a5SEric Dumazet goto coalesce; 3101845704a5SEric Dumazet return; 31021da177e4SLinus Torvalds } 3103e2080072SEric Dumazet INIT_LIST_HEAD(&skb->tcp_tsorted_anchor); 3104d83769a5SEric Dumazet skb_reserve(skb, MAX_TCP_HEADER); 3105a6c5ea4cSEric Dumazet sk_forced_mem_schedule(sk, skb->truesize); 31061da177e4SLinus Torvalds /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 3107e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tp->write_seq, 3108a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_FIN); 31091da177e4SLinus Torvalds tcp_queue_skb(sk, skb); 31101da177e4SLinus Torvalds } 3111845704a5SEric Dumazet __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF); 31121da177e4SLinus Torvalds } 31131da177e4SLinus Torvalds 31141da177e4SLinus Torvalds /* We get here when a process closes a file descriptor (either due to 31151da177e4SLinus Torvalds * an explicit close() or as a byproduct of exit()'ing) and there 31161da177e4SLinus Torvalds * was unread data in the receive queue. This behavior is recommended 311765bb723cSGerrit Renker * by RFC 2525, section 2.17. -DaveM 31181da177e4SLinus Torvalds */ 3119dd0fc66fSAl Viro void tcp_send_active_reset(struct sock *sk, gfp_t priority) 31201da177e4SLinus Torvalds { 31211da177e4SLinus Torvalds struct sk_buff *skb; 31221da177e4SLinus Torvalds 31237cc2b043SGao Feng TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); 31247cc2b043SGao Feng 31251da177e4SLinus Torvalds /* NOTE: No TCP options attached and we never retransmit this. */ 31261da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, priority); 31271da177e4SLinus Torvalds if (!skb) { 31284e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 31291da177e4SLinus Torvalds return; 31301da177e4SLinus Torvalds } 31311da177e4SLinus Torvalds 31321da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 31331da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 3134e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), 3135a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_RST); 31369a568de4SEric Dumazet tcp_mstamp_refresh(tcp_sk(sk)); 31371da177e4SLinus Torvalds /* Send it off. */ 3138dfb4b9dcSDavid S. Miller if (tcp_transmit_skb(sk, skb, 0, priority)) 31394e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 3140c24b14c4SSong Liu 3141c24b14c4SSong Liu /* skb of trace_tcp_send_reset() keeps the skb that caused RST, 3142c24b14c4SSong Liu * skb here is different to the troublesome skb, so use NULL 3143c24b14c4SSong Liu */ 3144c24b14c4SSong Liu trace_tcp_send_reset(sk, NULL); 31451da177e4SLinus Torvalds } 31461da177e4SLinus Torvalds 314767edfef7SAndi Kleen /* Send a crossed SYN-ACK during socket establishment. 314867edfef7SAndi Kleen * WARNING: This routine must only be called when we have already sent 31491da177e4SLinus Torvalds * a SYN packet that crossed the incoming SYN that caused this routine 31501da177e4SLinus Torvalds * to get called. If this assumption fails then the initial rcv_wnd 31511da177e4SLinus Torvalds * and rcv_wscale values will not be correct. 31521da177e4SLinus Torvalds */ 31531da177e4SLinus Torvalds int tcp_send_synack(struct sock *sk) 31541da177e4SLinus Torvalds { 31551da177e4SLinus Torvalds struct sk_buff *skb; 31561da177e4SLinus Torvalds 315775c119afSEric Dumazet skb = tcp_rtx_queue_head(sk); 315851456b29SIan Morris if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 315975c119afSEric Dumazet pr_err("%s: wrong queue state\n", __func__); 31601da177e4SLinus Torvalds return -EFAULT; 31611da177e4SLinus Torvalds } 31624de075e0SEric Dumazet if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { 31631da177e4SLinus Torvalds if (skb_cloned(skb)) { 3164e2080072SEric Dumazet struct sk_buff *nskb; 3165e2080072SEric Dumazet 3166e2080072SEric Dumazet tcp_skb_tsorted_save(skb) { 3167e2080072SEric Dumazet nskb = skb_copy(skb, GFP_ATOMIC); 3168e2080072SEric Dumazet } tcp_skb_tsorted_restore(skb); 316951456b29SIan Morris if (!nskb) 31701da177e4SLinus Torvalds return -ENOMEM; 3171e2080072SEric Dumazet INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor); 317275c119afSEric Dumazet tcp_rtx_queue_unlink_and_free(skb, sk); 3173f4a775d1SEric Dumazet __skb_header_release(nskb); 317475c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb); 31753ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 31763ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 31771da177e4SLinus Torvalds skb = nskb; 31781da177e4SLinus Torvalds } 31791da177e4SLinus Torvalds 31804de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; 3181735d3831SFlorian Westphal tcp_ecn_send_synack(sk, skb); 31821da177e4SLinus Torvalds } 3183dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 31841da177e4SLinus Torvalds } 31851da177e4SLinus Torvalds 31864aea39c1SEric Dumazet /** 31874aea39c1SEric Dumazet * tcp_make_synack - Prepare a SYN-ACK. 31884aea39c1SEric Dumazet * sk: listener socket 31894aea39c1SEric Dumazet * dst: dst entry attached to the SYNACK 31904aea39c1SEric Dumazet * req: request_sock pointer 31914aea39c1SEric Dumazet * 31924aea39c1SEric Dumazet * Allocate one skb and build a SYNACK packet. 31934aea39c1SEric Dumazet * @dst is consumed : Caller should not use it again. 31944aea39c1SEric Dumazet */ 31955d062de7SEric Dumazet struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst, 3196e6b4d113SWilliam Allen Simpson struct request_sock *req, 3197ca6fb065SEric Dumazet struct tcp_fastopen_cookie *foc, 3198b3d05147SEric Dumazet enum tcp_synack_type synack_type) 31991da177e4SLinus Torvalds { 32002e6599cbSArnaldo Carvalho de Melo struct inet_request_sock *ireq = inet_rsk(req); 32015d062de7SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 320280f03e27SEric Dumazet struct tcp_md5sig_key *md5 = NULL; 32035d062de7SEric Dumazet struct tcp_out_options opts; 32045d062de7SEric Dumazet struct sk_buff *skb; 3205bd0388aeSWilliam Allen Simpson int tcp_header_size; 32065d062de7SEric Dumazet struct tcphdr *th; 3207f5fff5dcSTom Quetchenbach int mss; 32081da177e4SLinus Torvalds 3209ca6fb065SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 32104aea39c1SEric Dumazet if (unlikely(!skb)) { 32114aea39c1SEric Dumazet dst_release(dst); 32121da177e4SLinus Torvalds return NULL; 32134aea39c1SEric Dumazet } 32141da177e4SLinus Torvalds /* Reserve space for headers. */ 32151da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 32161da177e4SLinus Torvalds 3217b3d05147SEric Dumazet switch (synack_type) { 3218b3d05147SEric Dumazet case TCP_SYNACK_NORMAL: 32199e17f8a4SEric Dumazet skb_set_owner_w(skb, req_to_sk(req)); 3220b3d05147SEric Dumazet break; 3221b3d05147SEric Dumazet case TCP_SYNACK_COOKIE: 3222b3d05147SEric Dumazet /* Under synflood, we do not attach skb to a socket, 3223b3d05147SEric Dumazet * to avoid false sharing. 3224b3d05147SEric Dumazet */ 3225b3d05147SEric Dumazet break; 3226b3d05147SEric Dumazet case TCP_SYNACK_FASTOPEN: 3227ca6fb065SEric Dumazet /* sk is a const pointer, because we want to express multiple 3228ca6fb065SEric Dumazet * cpu might call us concurrently. 3229ca6fb065SEric Dumazet * sk->sk_wmem_alloc in an atomic, we can promote to rw. 3230ca6fb065SEric Dumazet */ 3231ca6fb065SEric Dumazet skb_set_owner_w(skb, (struct sock *)sk); 3232b3d05147SEric Dumazet break; 3233ca6fb065SEric Dumazet } 32344aea39c1SEric Dumazet skb_dst_set(skb, dst); 32351da177e4SLinus Torvalds 32363541f9e8SEric Dumazet mss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 3237f5fff5dcSTom Quetchenbach 323833ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 32398b5f12d0SFlorian Westphal #ifdef CONFIG_SYN_COOKIES 32408b5f12d0SFlorian Westphal if (unlikely(req->cookie_ts)) 32419a568de4SEric Dumazet skb->skb_mstamp = cookie_init_timestamp(req); 32428b5f12d0SFlorian Westphal else 32438b5f12d0SFlorian Westphal #endif 32449a568de4SEric Dumazet skb->skb_mstamp = tcp_clock_us(); 324580f03e27SEric Dumazet 324680f03e27SEric Dumazet #ifdef CONFIG_TCP_MD5SIG 324780f03e27SEric Dumazet rcu_read_lock(); 3248fd3a154aSEric Dumazet md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req)); 324980f03e27SEric Dumazet #endif 325058d607d3SEric Dumazet skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4); 325160e2a778SUrsula Braun tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5, 325260e2a778SUrsula Braun foc) + sizeof(*th); 325333ad798cSAdam Langley 3254aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 3255aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 32561da177e4SLinus Torvalds 3257ea1627c2SEric Dumazet th = (struct tcphdr *)skb->data; 32581da177e4SLinus Torvalds memset(th, 0, sizeof(struct tcphdr)); 32591da177e4SLinus Torvalds th->syn = 1; 32601da177e4SLinus Torvalds th->ack = 1; 32616ac705b1SEric Dumazet tcp_ecn_make_synack(req, th); 3262b44084c2SEric Dumazet th->source = htons(ireq->ir_num); 3263634fb979SEric Dumazet th->dest = ireq->ir_rmt_port; 3264e05a90ecSJamal Hadi Salim skb->mark = ireq->ir_mark; 3265e870a8efSIlpo Järvinen /* Setting of flags are superfluous here for callers (and ECE is 3266e870a8efSIlpo Järvinen * not even correctly set) 3267e870a8efSIlpo Järvinen */ 3268e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn, 3269a3433f35SChangli Gao TCPHDR_SYN | TCPHDR_ACK); 32704957faadSWilliam Allen Simpson 32711da177e4SLinus Torvalds th->seq = htonl(TCP_SKB_CB(skb)->seq); 32728336886fSJerry Chu /* XXX data is queued and acked as is. No buffer/window check */ 32738336886fSJerry Chu th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt); 32741da177e4SLinus Torvalds 32751da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 3276ed53d0abSEric Dumazet th->window = htons(min(req->rsk_rcv_wnd, 65535U)); 32775d062de7SEric Dumazet tcp_options_write((__be32 *)(th + 1), NULL, &opts); 32781da177e4SLinus Torvalds th->doff = (tcp_header_size >> 2); 327990bbcc60SEric Dumazet __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS); 3280cfb6eeb4SYOSHIFUJI Hideaki 3281cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 3282cfb6eeb4SYOSHIFUJI Hideaki /* Okay, we have all we need - do the md5 hash if needed */ 328380f03e27SEric Dumazet if (md5) 3284bd0388aeSWilliam Allen Simpson tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, 328539f8e58eSEric Dumazet md5, req_to_sk(req), skb); 328680f03e27SEric Dumazet rcu_read_unlock(); 3287cfb6eeb4SYOSHIFUJI Hideaki #endif 3288cfb6eeb4SYOSHIFUJI Hideaki 3289b50edd78SEric Dumazet /* Do not fool tcpdump (if any), clean our debris */ 32902456e855SThomas Gleixner skb->tstamp = 0; 32911da177e4SLinus Torvalds return skb; 32921da177e4SLinus Torvalds } 32934bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_make_synack); 32941da177e4SLinus Torvalds 329581164413SDaniel Borkmann static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst) 329681164413SDaniel Borkmann { 329781164413SDaniel Borkmann struct inet_connection_sock *icsk = inet_csk(sk); 329881164413SDaniel Borkmann const struct tcp_congestion_ops *ca; 329981164413SDaniel Borkmann u32 ca_key = dst_metric(dst, RTAX_CC_ALGO); 330081164413SDaniel Borkmann 330181164413SDaniel Borkmann if (ca_key == TCP_CA_UNSPEC) 330281164413SDaniel Borkmann return; 330381164413SDaniel Borkmann 330481164413SDaniel Borkmann rcu_read_lock(); 330581164413SDaniel Borkmann ca = tcp_ca_find_key(ca_key); 330681164413SDaniel Borkmann if (likely(ca && try_module_get(ca->owner))) { 330781164413SDaniel Borkmann module_put(icsk->icsk_ca_ops->owner); 330881164413SDaniel Borkmann icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst); 330981164413SDaniel Borkmann icsk->icsk_ca_ops = ca; 331081164413SDaniel Borkmann } 331181164413SDaniel Borkmann rcu_read_unlock(); 331281164413SDaniel Borkmann } 331381164413SDaniel Borkmann 331467edfef7SAndi Kleen /* Do all connect socket setups that can be done AF independent. */ 3315f7e56a76Sstephen hemminger static void tcp_connect_init(struct sock *sk) 33161da177e4SLinus Torvalds { 3317cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 33181da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 33191da177e4SLinus Torvalds __u8 rcv_wscale; 332013d3b1ebSLawrence Brakmo u32 rcv_wnd; 33211da177e4SLinus Torvalds 33221da177e4SLinus Torvalds /* We'll fix this up when we get a response from the other end. 33231da177e4SLinus Torvalds * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 33241da177e4SLinus Torvalds */ 33255d2ed052SEric Dumazet tp->tcp_header_len = sizeof(struct tcphdr); 33265d2ed052SEric Dumazet if (sock_net(sk)->ipv4.sysctl_tcp_timestamps) 33275d2ed052SEric Dumazet tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED; 33281da177e4SLinus Torvalds 3329cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 333000db4124SIan Morris if (tp->af_specific->md5_lookup(sk, sk)) 3331cfb6eeb4SYOSHIFUJI Hideaki tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 3332cfb6eeb4SYOSHIFUJI Hideaki #endif 3333cfb6eeb4SYOSHIFUJI Hideaki 33341da177e4SLinus Torvalds /* If user gave his TCP_MAXSEG, record it to clamp */ 33351da177e4SLinus Torvalds if (tp->rx_opt.user_mss) 33361da177e4SLinus Torvalds tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 33371da177e4SLinus Torvalds tp->max_window = 0; 33385d424d5aSJohn Heffner tcp_mtup_init(sk); 33391da177e4SLinus Torvalds tcp_sync_mss(sk, dst_mtu(dst)); 33401da177e4SLinus Torvalds 334181164413SDaniel Borkmann tcp_ca_dst_init(sk, dst); 334281164413SDaniel Borkmann 33431da177e4SLinus Torvalds if (!tp->window_clamp) 33441da177e4SLinus Torvalds tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 33453541f9e8SEric Dumazet tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 3346f5fff5dcSTom Quetchenbach 33471da177e4SLinus Torvalds tcp_initialize_rcv_mss(sk); 33481da177e4SLinus Torvalds 3349e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 3350e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 3351e88c64f0SHagen Paul Pfeifer (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) 3352e88c64f0SHagen Paul Pfeifer tp->window_clamp = tcp_full_space(sk); 3353e88c64f0SHagen Paul Pfeifer 335413d3b1ebSLawrence Brakmo rcv_wnd = tcp_rwnd_init_bpf(sk); 335513d3b1ebSLawrence Brakmo if (rcv_wnd == 0) 335613d3b1ebSLawrence Brakmo rcv_wnd = dst_metric(dst, RTAX_INITRWND); 335713d3b1ebSLawrence Brakmo 33581da177e4SLinus Torvalds tcp_select_initial_window(tcp_full_space(sk), 33591da177e4SLinus Torvalds tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 33601da177e4SLinus Torvalds &tp->rcv_wnd, 33611da177e4SLinus Torvalds &tp->window_clamp, 33629bb37ef0SEric Dumazet sock_net(sk)->ipv4.sysctl_tcp_window_scaling, 336331d12926Slaurent chavey &rcv_wscale, 336413d3b1ebSLawrence Brakmo rcv_wnd); 33651da177e4SLinus Torvalds 33661da177e4SLinus Torvalds tp->rx_opt.rcv_wscale = rcv_wscale; 33671da177e4SLinus Torvalds tp->rcv_ssthresh = tp->rcv_wnd; 33681da177e4SLinus Torvalds 33691da177e4SLinus Torvalds sk->sk_err = 0; 33701da177e4SLinus Torvalds sock_reset_flag(sk, SOCK_DONE); 33711da177e4SLinus Torvalds tp->snd_wnd = 0; 3372ee7537b6SHantzis Fotis tcp_init_wl(tp, 0); 33731da177e4SLinus Torvalds tp->snd_una = tp->write_seq; 33741da177e4SLinus Torvalds tp->snd_sml = tp->write_seq; 337533f5f57eSIlpo Järvinen tp->snd_up = tp->write_seq; 3376370816aeSPavel Emelyanov tp->snd_nxt = tp->write_seq; 3377ee995283SPavel Emelyanov 3378ee995283SPavel Emelyanov if (likely(!tp->repair)) 33791da177e4SLinus Torvalds tp->rcv_nxt = 0; 3380c7781a6eSAndrew Vagin else 338170eabf0eSEric Dumazet tp->rcv_tstamp = tcp_jiffies32; 3382ee995283SPavel Emelyanov tp->rcv_wup = tp->rcv_nxt; 3383ee995283SPavel Emelyanov tp->copied_seq = tp->rcv_nxt; 33841da177e4SLinus Torvalds 33858550f328SLawrence Brakmo inet_csk(sk)->icsk_rto = tcp_timeout_init(sk); 3386463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_retransmits = 0; 33871da177e4SLinus Torvalds tcp_clear_retrans(tp); 33881da177e4SLinus Torvalds } 33891da177e4SLinus Torvalds 3390783237e8SYuchung Cheng static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) 3391783237e8SYuchung Cheng { 3392783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3393783237e8SYuchung Cheng struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 3394783237e8SYuchung Cheng 3395783237e8SYuchung Cheng tcb->end_seq += skb->len; 3396f4a775d1SEric Dumazet __skb_header_release(skb); 3397783237e8SYuchung Cheng sk->sk_wmem_queued += skb->truesize; 3398783237e8SYuchung Cheng sk_mem_charge(sk, skb->truesize); 3399783237e8SYuchung Cheng tp->write_seq = tcb->end_seq; 3400783237e8SYuchung Cheng tp->packets_out += tcp_skb_pcount(skb); 3401783237e8SYuchung Cheng } 3402783237e8SYuchung Cheng 3403783237e8SYuchung Cheng /* Build and send a SYN with data and (cached) Fast Open cookie. However, 3404783237e8SYuchung Cheng * queue a data-only packet after the regular SYN, such that regular SYNs 3405783237e8SYuchung Cheng * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges 3406783237e8SYuchung Cheng * only the SYN sequence, the data are retransmitted in the first ACK. 3407783237e8SYuchung Cheng * If cookie is not cached or other error occurs, falls back to send a 3408783237e8SYuchung Cheng * regular SYN with Fast Open cookie request option. 3409783237e8SYuchung Cheng */ 3410783237e8SYuchung Cheng static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) 3411783237e8SYuchung Cheng { 3412783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3413783237e8SYuchung Cheng struct tcp_fastopen_request *fo = tp->fastopen_req; 3414065263f4SWei Wang int space, err = 0; 3415355a901eSEric Dumazet struct sk_buff *syn_data; 3416783237e8SYuchung Cheng 341767da22d2SYuchung Cheng tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ 3418065263f4SWei Wang if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie)) 3419783237e8SYuchung Cheng goto fallback; 3420783237e8SYuchung Cheng 3421783237e8SYuchung Cheng /* MSS for SYN-data is based on cached MSS and bounded by PMTU and 3422783237e8SYuchung Cheng * user-MSS. Reserve maximum option space for middleboxes that add 3423783237e8SYuchung Cheng * private TCP options. The cost is reduced data space in SYN :( 3424783237e8SYuchung Cheng */ 34253541f9e8SEric Dumazet tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp); 34263541f9e8SEric Dumazet 34271b63edd6SYuchung Cheng space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) - 3428783237e8SYuchung Cheng MAX_TCP_OPTION_SPACE; 3429783237e8SYuchung Cheng 3430f5ddcbbbSEric Dumazet space = min_t(size_t, space, fo->size); 3431f5ddcbbbSEric Dumazet 3432f5ddcbbbSEric Dumazet /* limit to order-0 allocations */ 3433f5ddcbbbSEric Dumazet space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER)); 3434f5ddcbbbSEric Dumazet 3435eb934478SEric Dumazet syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false); 3436355a901eSEric Dumazet if (!syn_data) 3437783237e8SYuchung Cheng goto fallback; 3438355a901eSEric Dumazet syn_data->ip_summed = CHECKSUM_PARTIAL; 3439355a901eSEric Dumazet memcpy(syn_data->cb, syn->cb, sizeof(syn->cb)); 344007e100f9SEric Dumazet if (space) { 344107e100f9SEric Dumazet int copied = copy_from_iter(skb_put(syn_data, space), space, 344257be5bdaSAl Viro &fo->data->msg_iter); 344357be5bdaSAl Viro if (unlikely(!copied)) { 3444ba233b34SEric Dumazet tcp_skb_tsorted_anchor_cleanup(syn_data); 3445355a901eSEric Dumazet kfree_skb(syn_data); 3446783237e8SYuchung Cheng goto fallback; 3447783237e8SYuchung Cheng } 344857be5bdaSAl Viro if (copied != space) { 344957be5bdaSAl Viro skb_trim(syn_data, copied); 345057be5bdaSAl Viro space = copied; 345157be5bdaSAl Viro } 345207e100f9SEric Dumazet } 3453355a901eSEric Dumazet /* No more data pending in inet_wait_for_connect() */ 3454355a901eSEric Dumazet if (space == fo->size) 3455355a901eSEric Dumazet fo->data = NULL; 3456355a901eSEric Dumazet fo->copied = space; 3457783237e8SYuchung Cheng 3458355a901eSEric Dumazet tcp_connect_queue_skb(sk, syn_data); 34590f87230dSFrancis Yan if (syn_data->len) 34600f87230dSFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_BUSY); 3461355a901eSEric Dumazet 3462355a901eSEric Dumazet err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation); 3463355a901eSEric Dumazet 3464355a901eSEric Dumazet syn->skb_mstamp = syn_data->skb_mstamp; 3465355a901eSEric Dumazet 3466355a901eSEric Dumazet /* Now full SYN+DATA was cloned and sent (or not), 3467355a901eSEric Dumazet * remove the SYN from the original skb (syn_data) 3468355a901eSEric Dumazet * we keep in write queue in case of a retransmit, as we 3469355a901eSEric Dumazet * also have the SYN packet (with no data) in the same queue. 3470431a9124SEric Dumazet */ 3471355a901eSEric Dumazet TCP_SKB_CB(syn_data)->seq++; 3472355a901eSEric Dumazet TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH; 3473355a901eSEric Dumazet if (!err) { 347467da22d2SYuchung Cheng tp->syn_data = (fo->copied > 0); 347575c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data); 3476f19c29e3SYuchung Cheng NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT); 3477783237e8SYuchung Cheng goto done; 3478783237e8SYuchung Cheng } 3479783237e8SYuchung Cheng 348075c119afSEric Dumazet /* data was not sent, put it in write_queue */ 348175c119afSEric Dumazet __skb_queue_tail(&sk->sk_write_queue, syn_data); 3482b5b7db8dSEric Dumazet tp->packets_out -= tcp_skb_pcount(syn_data); 3483b5b7db8dSEric Dumazet 3484783237e8SYuchung Cheng fallback: 3485783237e8SYuchung Cheng /* Send a regular SYN with Fast Open cookie request option */ 3486783237e8SYuchung Cheng if (fo->cookie.len > 0) 3487783237e8SYuchung Cheng fo->cookie.len = 0; 3488783237e8SYuchung Cheng err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); 3489783237e8SYuchung Cheng if (err) 3490783237e8SYuchung Cheng tp->syn_fastopen = 0; 3491783237e8SYuchung Cheng done: 3492783237e8SYuchung Cheng fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ 3493783237e8SYuchung Cheng return err; 3494783237e8SYuchung Cheng } 3495783237e8SYuchung Cheng 349667edfef7SAndi Kleen /* Build a SYN and send it off. */ 34971da177e4SLinus Torvalds int tcp_connect(struct sock *sk) 34981da177e4SLinus Torvalds { 34991da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 35001da177e4SLinus Torvalds struct sk_buff *buff; 3501ee586811SEric Paris int err; 35021da177e4SLinus Torvalds 35039872a4bdSLawrence Brakmo tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB); 35048ba60924SEric Dumazet 35058ba60924SEric Dumazet if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 35068ba60924SEric Dumazet return -EHOSTUNREACH; /* Routing failure or similar. */ 35078ba60924SEric Dumazet 35081da177e4SLinus Torvalds tcp_connect_init(sk); 35091da177e4SLinus Torvalds 35102b916477SAndrey Vagin if (unlikely(tp->repair)) { 35112b916477SAndrey Vagin tcp_finish_connect(sk, NULL); 35122b916477SAndrey Vagin return 0; 35132b916477SAndrey Vagin } 35142b916477SAndrey Vagin 3515eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true); 3516355a901eSEric Dumazet if (unlikely(!buff)) 35171da177e4SLinus Torvalds return -ENOBUFS; 35181da177e4SLinus Torvalds 3519a3433f35SChangli Gao tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); 35209a568de4SEric Dumazet tcp_mstamp_refresh(tp); 35219a568de4SEric Dumazet tp->retrans_stamp = tcp_time_stamp(tp); 3522783237e8SYuchung Cheng tcp_connect_queue_skb(sk, buff); 3523735d3831SFlorian Westphal tcp_ecn_send_syn(sk, buff); 352475c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, buff); 35251da177e4SLinus Torvalds 3526783237e8SYuchung Cheng /* Send off SYN; include data in Fast Open. */ 3527783237e8SYuchung Cheng err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : 3528783237e8SYuchung Cheng tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); 3529ee586811SEric Paris if (err == -ECONNREFUSED) 3530ee586811SEric Paris return err; 3531bd37a088SWei Yongjun 3532bd37a088SWei Yongjun /* We change tp->snd_nxt after the tcp_transmit_skb() call 3533bd37a088SWei Yongjun * in order to make this packet get counted in tcpOutSegs. 3534bd37a088SWei Yongjun */ 3535bd37a088SWei Yongjun tp->snd_nxt = tp->write_seq; 3536bd37a088SWei Yongjun tp->pushed_seq = tp->write_seq; 3537b5b7db8dSEric Dumazet buff = tcp_send_head(sk); 3538b5b7db8dSEric Dumazet if (unlikely(buff)) { 3539b5b7db8dSEric Dumazet tp->snd_nxt = TCP_SKB_CB(buff)->seq; 3540b5b7db8dSEric Dumazet tp->pushed_seq = TCP_SKB_CB(buff)->seq; 3541b5b7db8dSEric Dumazet } 354281cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); 35431da177e4SLinus Torvalds 35441da177e4SLinus Torvalds /* Timer for repeating the SYN until an answer. */ 35453f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 35463f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 35471da177e4SLinus Torvalds return 0; 35481da177e4SLinus Torvalds } 35494bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_connect); 35501da177e4SLinus Torvalds 35511da177e4SLinus Torvalds /* Send out a delayed ack, the caller does the policy checking 35521da177e4SLinus Torvalds * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 35531da177e4SLinus Torvalds * for details. 35541da177e4SLinus Torvalds */ 35551da177e4SLinus Torvalds void tcp_send_delayed_ack(struct sock *sk) 35561da177e4SLinus Torvalds { 3557463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 3558463c84b9SArnaldo Carvalho de Melo int ato = icsk->icsk_ack.ato; 35591da177e4SLinus Torvalds unsigned long timeout; 35601da177e4SLinus Torvalds 35619890092eSFlorian Westphal tcp_ca_event(sk, CA_EVENT_DELAYED_ACK); 35629890092eSFlorian Westphal 35631da177e4SLinus Torvalds if (ato > TCP_DELACK_MIN) { 3564463c84b9SArnaldo Carvalho de Melo const struct tcp_sock *tp = tcp_sk(sk); 35651da177e4SLinus Torvalds int max_ato = HZ / 2; 35661da177e4SLinus Torvalds 3567056834d9SIlpo Järvinen if (icsk->icsk_ack.pingpong || 3568056834d9SIlpo Järvinen (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 35691da177e4SLinus Torvalds max_ato = TCP_DELACK_MAX; 35701da177e4SLinus Torvalds 35711da177e4SLinus Torvalds /* Slow path, intersegment interval is "high". */ 35721da177e4SLinus Torvalds 35731da177e4SLinus Torvalds /* If some rtt estimate is known, use it to bound delayed ack. 3574463c84b9SArnaldo Carvalho de Melo * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 35751da177e4SLinus Torvalds * directly. 35761da177e4SLinus Torvalds */ 3577740b0f18SEric Dumazet if (tp->srtt_us) { 3578740b0f18SEric Dumazet int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3), 3579740b0f18SEric Dumazet TCP_DELACK_MIN); 35801da177e4SLinus Torvalds 35811da177e4SLinus Torvalds if (rtt < max_ato) 35821da177e4SLinus Torvalds max_ato = rtt; 35831da177e4SLinus Torvalds } 35841da177e4SLinus Torvalds 35851da177e4SLinus Torvalds ato = min(ato, max_ato); 35861da177e4SLinus Torvalds } 35871da177e4SLinus Torvalds 35881da177e4SLinus Torvalds /* Stay within the limit we were given */ 35891da177e4SLinus Torvalds timeout = jiffies + ato; 35901da177e4SLinus Torvalds 35911da177e4SLinus Torvalds /* Use new timeout only if there wasn't a older one earlier. */ 3592463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 35931da177e4SLinus Torvalds /* If delack timer was blocked or is about to expire, 35941da177e4SLinus Torvalds * send ACK now. 35951da177e4SLinus Torvalds */ 3596463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.blocked || 3597463c84b9SArnaldo Carvalho de Melo time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 35981da177e4SLinus Torvalds tcp_send_ack(sk); 35991da177e4SLinus Torvalds return; 36001da177e4SLinus Torvalds } 36011da177e4SLinus Torvalds 3602463c84b9SArnaldo Carvalho de Melo if (!time_before(timeout, icsk->icsk_ack.timeout)) 3603463c84b9SArnaldo Carvalho de Melo timeout = icsk->icsk_ack.timeout; 36041da177e4SLinus Torvalds } 3605463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 3606463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.timeout = timeout; 3607463c84b9SArnaldo Carvalho de Melo sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 36081da177e4SLinus Torvalds } 36091da177e4SLinus Torvalds 36101da177e4SLinus Torvalds /* This routine sends an ack and also updates the window. */ 36111da177e4SLinus Torvalds void tcp_send_ack(struct sock *sk) 36121da177e4SLinus Torvalds { 36131da177e4SLinus Torvalds struct sk_buff *buff; 36141da177e4SLinus Torvalds 3615058dc334SIlpo Järvinen /* If we have been reset, we may not send again. */ 3616058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3617058dc334SIlpo Järvinen return; 3618058dc334SIlpo Järvinen 36199890092eSFlorian Westphal tcp_ca_event(sk, CA_EVENT_NON_DELAYED_ACK); 36209890092eSFlorian Westphal 36211da177e4SLinus Torvalds /* We are not putting this on the write queue, so 36221da177e4SLinus Torvalds * tcp_transmit_skb() will set the ownership to this 36231da177e4SLinus Torvalds * sock. 36241da177e4SLinus Torvalds */ 36257450aaf6SEric Dumazet buff = alloc_skb(MAX_TCP_HEADER, 36267450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 36277450aaf6SEric Dumazet if (unlikely(!buff)) { 3628463c84b9SArnaldo Carvalho de Melo inet_csk_schedule_ack(sk); 3629463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; 36303f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 36313f421baaSArnaldo Carvalho de Melo TCP_DELACK_MAX, TCP_RTO_MAX); 36321da177e4SLinus Torvalds return; 36331da177e4SLinus Torvalds } 36341da177e4SLinus Torvalds 36351da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 36361da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 3637a3433f35SChangli Gao tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); 36381da177e4SLinus Torvalds 363998781965SEric Dumazet /* We do not want pure acks influencing TCP Small Queues or fq/pacing 364098781965SEric Dumazet * too much. 364198781965SEric Dumazet * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784 364298781965SEric Dumazet */ 364398781965SEric Dumazet skb_set_tcp_pure_ack(buff); 364498781965SEric Dumazet 36451da177e4SLinus Torvalds /* Send it off, this clears delayed acks for us. */ 36467450aaf6SEric Dumazet tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0); 36471da177e4SLinus Torvalds } 3648e3118e83SDaniel Borkmann EXPORT_SYMBOL_GPL(tcp_send_ack); 36491da177e4SLinus Torvalds 36501da177e4SLinus Torvalds /* This routine sends a packet with an out of date sequence 36511da177e4SLinus Torvalds * number. It assumes the other end will try to ack it. 36521da177e4SLinus Torvalds * 36531da177e4SLinus Torvalds * Question: what should we make while urgent mode? 36541da177e4SLinus Torvalds * 4.4BSD forces sending single byte of data. We cannot send 36551da177e4SLinus Torvalds * out of window data, because we have SND.NXT==SND.MAX... 36561da177e4SLinus Torvalds * 36571da177e4SLinus Torvalds * Current solution: to send TWO zero-length segments in urgent mode: 36581da177e4SLinus Torvalds * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 36591da177e4SLinus Torvalds * out-of-date with SND.UNA-1 to probe window. 36601da177e4SLinus Torvalds */ 3661e520af48SEric Dumazet static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib) 36621da177e4SLinus Torvalds { 36631da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 36641da177e4SLinus Torvalds struct sk_buff *skb; 36651da177e4SLinus Torvalds 36661da177e4SLinus Torvalds /* We don't queue it, tcp_transmit_skb() sets ownership. */ 36677450aaf6SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, 36687450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 366951456b29SIan Morris if (!skb) 36701da177e4SLinus Torvalds return -1; 36711da177e4SLinus Torvalds 36721da177e4SLinus Torvalds /* Reserve space for headers and set control bits. */ 36731da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 36741da177e4SLinus Torvalds /* Use a previous sequence. This should cause the other 36751da177e4SLinus Torvalds * end to send an ack. Don't queue or clone SKB, just 36761da177e4SLinus Torvalds * send it. 36771da177e4SLinus Torvalds */ 3678a3433f35SChangli Gao tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); 3679e2e8009fSRenato Westphal NET_INC_STATS(sock_net(sk), mib); 36807450aaf6SEric Dumazet return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0); 36811da177e4SLinus Torvalds } 36821da177e4SLinus Torvalds 3683385e2070SEric Dumazet /* Called from setsockopt( ... TCP_REPAIR ) */ 3684ee995283SPavel Emelyanov void tcp_send_window_probe(struct sock *sk) 3685ee995283SPavel Emelyanov { 3686ee995283SPavel Emelyanov if (sk->sk_state == TCP_ESTABLISHED) { 3687ee995283SPavel Emelyanov tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; 36889a568de4SEric Dumazet tcp_mstamp_refresh(tcp_sk(sk)); 3689e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE); 3690ee995283SPavel Emelyanov } 3691ee995283SPavel Emelyanov } 3692ee995283SPavel Emelyanov 369367edfef7SAndi Kleen /* Initiate keepalive or window probe from timer. */ 3694e520af48SEric Dumazet int tcp_write_wakeup(struct sock *sk, int mib) 36951da177e4SLinus Torvalds { 36961da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 36971da177e4SLinus Torvalds struct sk_buff *skb; 36981da177e4SLinus Torvalds 3699058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3700058dc334SIlpo Järvinen return -1; 3701058dc334SIlpo Järvinen 370200db4124SIan Morris skb = tcp_send_head(sk); 370300db4124SIan Morris if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { 37041da177e4SLinus Torvalds int err; 37050c54b85fSIlpo Järvinen unsigned int mss = tcp_current_mss(sk); 370690840defSIlpo Järvinen unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 37071da177e4SLinus Torvalds 37081da177e4SLinus Torvalds if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 37091da177e4SLinus Torvalds tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 37101da177e4SLinus Torvalds 37111da177e4SLinus Torvalds /* We are probing the opening of a window 37121da177e4SLinus Torvalds * but the window size is != 0 37131da177e4SLinus Torvalds * must have been a result SWS avoidance ( sender ) 37141da177e4SLinus Torvalds */ 37151da177e4SLinus Torvalds if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 37161da177e4SLinus Torvalds skb->len > mss) { 37171da177e4SLinus Torvalds seg_size = min(seg_size, mss); 37184de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 371975c119afSEric Dumazet if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE, 372075c119afSEric Dumazet skb, seg_size, mss, GFP_ATOMIC)) 37211da177e4SLinus Torvalds return -1; 37221da177e4SLinus Torvalds } else if (!tcp_skb_pcount(skb)) 37235bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss); 37241da177e4SLinus Torvalds 37254de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 3726dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 372766f5fe62SIlpo Järvinen if (!err) 372866f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 37291da177e4SLinus Torvalds return err; 37301da177e4SLinus Torvalds } else { 373133f5f57eSIlpo Järvinen if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) 3732e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 1, mib); 3733e520af48SEric Dumazet return tcp_xmit_probe_skb(sk, 0, mib); 37341da177e4SLinus Torvalds } 37351da177e4SLinus Torvalds } 37361da177e4SLinus Torvalds 37371da177e4SLinus Torvalds /* A window probe timeout has occurred. If window is not closed send 37381da177e4SLinus Torvalds * a partial packet else a zero probe. 37391da177e4SLinus Torvalds */ 37401da177e4SLinus Torvalds void tcp_send_probe0(struct sock *sk) 37411da177e4SLinus Torvalds { 3742463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 37431da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 3744c6214a97SNikolay Borisov struct net *net = sock_net(sk); 3745fcdd1cf4SEric Dumazet unsigned long probe_max; 37461da177e4SLinus Torvalds int err; 37471da177e4SLinus Torvalds 3748e520af48SEric Dumazet err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE); 37491da177e4SLinus Torvalds 375075c119afSEric Dumazet if (tp->packets_out || tcp_write_queue_empty(sk)) { 37511da177e4SLinus Torvalds /* Cancel probe timer, if it is not required. */ 37526687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 0; 3753463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff = 0; 37541da177e4SLinus Torvalds return; 37551da177e4SLinus Torvalds } 37561da177e4SLinus Torvalds 37571da177e4SLinus Torvalds if (err <= 0) { 3758c6214a97SNikolay Borisov if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2) 3759463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff++; 37606687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out++; 3761fcdd1cf4SEric Dumazet probe_max = TCP_RTO_MAX; 37621da177e4SLinus Torvalds } else { 37631da177e4SLinus Torvalds /* If packet was not sent due to local congestion, 37646687e988SArnaldo Carvalho de Melo * do not backoff and do not remember icsk_probes_out. 37651da177e4SLinus Torvalds * Let local senders to fight for local resources. 37661da177e4SLinus Torvalds * 37671da177e4SLinus Torvalds * Use accumulated backoff yet. 37681da177e4SLinus Torvalds */ 37696687e988SArnaldo Carvalho de Melo if (!icsk->icsk_probes_out) 37706687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 1; 3771fcdd1cf4SEric Dumazet probe_max = TCP_RESOURCE_PROBE_INTERVAL; 37721da177e4SLinus Torvalds } 3773fcdd1cf4SEric Dumazet inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 377421c8fe99SEric Dumazet tcp_probe0_when(sk, probe_max), 3775fcdd1cf4SEric Dumazet TCP_RTO_MAX); 37761da177e4SLinus Torvalds } 37775db92c99SOctavian Purdila 3778ea3bea3aSEric Dumazet int tcp_rtx_synack(const struct sock *sk, struct request_sock *req) 37795db92c99SOctavian Purdila { 37805db92c99SOctavian Purdila const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific; 37815db92c99SOctavian Purdila struct flowi fl; 37825db92c99SOctavian Purdila int res; 37835db92c99SOctavian Purdila 378458d607d3SEric Dumazet tcp_rsk(req)->txhash = net_tx_rndhash(); 3785b3d05147SEric Dumazet res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL); 37865db92c99SOctavian Purdila if (!res) { 378790bbcc60SEric Dumazet __TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); 378802a1d6e7SEric Dumazet __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 37897e32b443SYuchung Cheng if (unlikely(tcp_passive_fastopen(sk))) 37907e32b443SYuchung Cheng tcp_sk(sk)->total_retrans++; 37915db92c99SOctavian Purdila } 37925db92c99SOctavian Purdila return res; 37935db92c99SOctavian Purdila } 37945db92c99SOctavian Purdila EXPORT_SYMBOL(tcp_rtx_synack); 3795