11da177e4SLinus Torvalds /* 21da177e4SLinus Torvalds * INET An implementation of the TCP/IP protocol suite for the LINUX 31da177e4SLinus Torvalds * operating system. INET is implemented using the BSD Socket 41da177e4SLinus Torvalds * interface as the means of communication with the user level. 51da177e4SLinus Torvalds * 61da177e4SLinus Torvalds * Implementation of the Transmission Control Protocol(TCP). 71da177e4SLinus Torvalds * 802c30a84SJesper Juhl * Authors: Ross Biro 91da177e4SLinus Torvalds * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> 101da177e4SLinus Torvalds * Mark Evans, <evansmp@uhura.aston.ac.uk> 111da177e4SLinus Torvalds * Corey Minyard <wf-rch!minyard@relay.EU.net> 121da177e4SLinus Torvalds * Florian La Roche, <flla@stud.uni-sb.de> 131da177e4SLinus Torvalds * Charles Hedrick, <hedrick@klinzhai.rutgers.edu> 141da177e4SLinus Torvalds * Linus Torvalds, <torvalds@cs.helsinki.fi> 151da177e4SLinus Torvalds * Alan Cox, <gw4pts@gw4pts.ampr.org> 161da177e4SLinus Torvalds * Matthew Dillon, <dillon@apollo.west.oic.com> 171da177e4SLinus Torvalds * Arnt Gulbrandsen, <agulbra@nvg.unit.no> 181da177e4SLinus Torvalds * Jorge Cwik, <jorge@laser.satlink.net> 191da177e4SLinus Torvalds */ 201da177e4SLinus Torvalds 211da177e4SLinus Torvalds /* 221da177e4SLinus Torvalds * Changes: Pedro Roque : Retransmit queue handled by TCP. 231da177e4SLinus Torvalds * : Fragmentation on mtu decrease 241da177e4SLinus Torvalds * : Segment collapse on retransmit 251da177e4SLinus Torvalds * : AF independence 261da177e4SLinus Torvalds * 271da177e4SLinus Torvalds * Linus Torvalds : send_delayed_ack 281da177e4SLinus Torvalds * David S. Miller : Charge memory using the right skb 291da177e4SLinus Torvalds * during syn/ack processing. 301da177e4SLinus Torvalds * David S. Miller : Output engine completely rewritten. 311da177e4SLinus Torvalds * Andrea Arcangeli: SYNACK carry ts_recent in tsecr. 321da177e4SLinus Torvalds * Cacophonix Gaul : draft-minshall-nagle-01 331da177e4SLinus Torvalds * J Hadi Salim : ECN support 341da177e4SLinus Torvalds * 351da177e4SLinus Torvalds */ 361da177e4SLinus Torvalds 3791df42beSJoe Perches #define pr_fmt(fmt) "TCP: " fmt 3891df42beSJoe Perches 391da177e4SLinus Torvalds #include <net/tcp.h> 401da177e4SLinus Torvalds 411da177e4SLinus Torvalds #include <linux/compiler.h> 425a0e3ad6STejun Heo #include <linux/gfp.h> 431da177e4SLinus Torvalds #include <linux/module.h> 441da177e4SLinus Torvalds 451da177e4SLinus Torvalds /* People can turn this off for buggy TCP's found in printers etc. */ 46ab32ea5dSBrian Haley int sysctl_tcp_retrans_collapse __read_mostly = 1; 471da177e4SLinus Torvalds 4815d99e02SRick Jones /* People can turn this on to work with those rare, broken TCPs that 4915d99e02SRick Jones * interpret the window field as a signed quantity. 5015d99e02SRick Jones */ 51ab32ea5dSBrian Haley int sysctl_tcp_workaround_signed_windows __read_mostly = 0; 5215d99e02SRick Jones 53c39c4c6aSWei Liu /* Default TSQ limit of four TSO segments */ 54c39c4c6aSWei Liu int sysctl_tcp_limit_output_bytes __read_mostly = 262144; 5546d3ceabSEric Dumazet 561da177e4SLinus Torvalds /* This limits the percentage of the congestion window which we 571da177e4SLinus Torvalds * will allow a single TSO frame to consume. Building TSO frames 581da177e4SLinus Torvalds * which are too large can cause TCP streams to be bursty. 591da177e4SLinus Torvalds */ 60ab32ea5dSBrian Haley int sysctl_tcp_tso_win_divisor __read_mostly = 3; 611da177e4SLinus Torvalds 6235089bb2SDavid S. Miller /* By default, RFC2861 behavior. */ 63ab32ea5dSBrian Haley int sysctl_tcp_slow_start_after_idle __read_mostly = 1; 6435089bb2SDavid S. Miller 6546d3ceabSEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 6646d3ceabSEric Dumazet int push_one, gfp_t gfp); 67519855c5SWilliam Allen Simpson 6867edfef7SAndi Kleen /* Account for new data that has been sent to the network. */ 69cf533ea5SEric Dumazet static void tcp_event_new_data_sent(struct sock *sk, const struct sk_buff *skb) 706ff03ac3SIlpo Järvinen { 716ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 726ff03ac3SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 7366f5fe62SIlpo Järvinen unsigned int prior_packets = tp->packets_out; 749e412ba7SIlpo Järvinen 75fe067e8aSDavid S. Miller tcp_advance_send_head(sk, skb); 761da177e4SLinus Torvalds tp->snd_nxt = TCP_SKB_CB(skb)->end_seq; 778512430eSIlpo Järvinen 7866f5fe62SIlpo Järvinen tp->packets_out += tcp_skb_pcount(skb); 79bec41a11SYuchung Cheng if (!prior_packets || icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) 80750ea2baSYuchung Cheng tcp_rearm_rto(sk); 81f19c29e3SYuchung Cheng 82f7324acdSDavid S. Miller NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT, 83f19c29e3SYuchung Cheng tcp_skb_pcount(skb)); 846a5dc9e5SEric Dumazet } 851da177e4SLinus Torvalds 86a4ecb15aSCui, Cheng /* SND.NXT, if window was not shrunk or the amount of shrunk was less than one 87a4ecb15aSCui, Cheng * window scaling factor due to loss of precision. 881da177e4SLinus Torvalds * If window has been shrunk, what should we make? It is not clear at all. 891da177e4SLinus Torvalds * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-( 901da177e4SLinus Torvalds * Anything in between SND.UNA...SND.UNA+SND.WND also can be already 911da177e4SLinus Torvalds * invalid. OK, let's make this for now: 921da177e4SLinus Torvalds */ 93cf533ea5SEric Dumazet static inline __u32 tcp_acceptable_seq(const struct sock *sk) 941da177e4SLinus Torvalds { 95cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 969e412ba7SIlpo Järvinen 97a4ecb15aSCui, Cheng if (!before(tcp_wnd_end(tp), tp->snd_nxt) || 98a4ecb15aSCui, Cheng (tp->rx_opt.wscale_ok && 99a4ecb15aSCui, Cheng ((tp->snd_nxt - tcp_wnd_end(tp)) < (1 << tp->rx_opt.rcv_wscale)))) 1001da177e4SLinus Torvalds return tp->snd_nxt; 1011da177e4SLinus Torvalds else 10290840defSIlpo Järvinen return tcp_wnd_end(tp); 1031da177e4SLinus Torvalds } 1041da177e4SLinus Torvalds 1051da177e4SLinus Torvalds /* Calculate mss to advertise in SYN segment. 1061da177e4SLinus Torvalds * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that: 1071da177e4SLinus Torvalds * 1081da177e4SLinus Torvalds * 1. It is independent of path mtu. 1091da177e4SLinus Torvalds * 2. Ideally, it is maximal possible segment size i.e. 65535-40. 1101da177e4SLinus Torvalds * 3. For IPv4 it is reasonable to calculate it from maximal MTU of 1111da177e4SLinus Torvalds * attached devices, because some buggy hosts are confused by 1121da177e4SLinus Torvalds * large MSS. 1131da177e4SLinus Torvalds * 4. We do not make 3, we advertise MSS, calculated from first 1141da177e4SLinus Torvalds * hop device mtu, but allow to raise it to ip_rt_min_advmss. 1151da177e4SLinus Torvalds * This may be overridden via information stored in routing table. 1161da177e4SLinus Torvalds * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible, 1171da177e4SLinus Torvalds * probably even Jumbo". 1181da177e4SLinus Torvalds */ 1191da177e4SLinus Torvalds static __u16 tcp_advertise_mss(struct sock *sk) 1201da177e4SLinus Torvalds { 1211da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 122cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1231da177e4SLinus Torvalds int mss = tp->advmss; 1241da177e4SLinus Torvalds 1250dbaee3bSDavid S. Miller if (dst) { 1260dbaee3bSDavid S. Miller unsigned int metric = dst_metric_advmss(dst); 1270dbaee3bSDavid S. Miller 1280dbaee3bSDavid S. Miller if (metric < mss) { 1290dbaee3bSDavid S. Miller mss = metric; 1301da177e4SLinus Torvalds tp->advmss = mss; 1311da177e4SLinus Torvalds } 1320dbaee3bSDavid S. Miller } 1331da177e4SLinus Torvalds 1341da177e4SLinus Torvalds return (__u16)mss; 1351da177e4SLinus Torvalds } 1361da177e4SLinus Torvalds 1371da177e4SLinus Torvalds /* RFC2861. Reset CWND after idle period longer RTO to "restart window". 1386f021c62SEric Dumazet * This is the first part of cwnd validation mechanism. 1396f021c62SEric Dumazet */ 1406f021c62SEric Dumazet void tcp_cwnd_restart(struct sock *sk, s32 delta) 1411da177e4SLinus Torvalds { 142463c84b9SArnaldo Carvalho de Melo struct tcp_sock *tp = tcp_sk(sk); 1436f021c62SEric Dumazet u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk)); 1441da177e4SLinus Torvalds u32 cwnd = tp->snd_cwnd; 1451da177e4SLinus Torvalds 1466687e988SArnaldo Carvalho de Melo tcp_ca_event(sk, CA_EVENT_CWND_RESTART); 1471da177e4SLinus Torvalds 1486687e988SArnaldo Carvalho de Melo tp->snd_ssthresh = tcp_current_ssthresh(sk); 1491da177e4SLinus Torvalds restart_cwnd = min(restart_cwnd, cwnd); 1501da177e4SLinus Torvalds 151463c84b9SArnaldo Carvalho de Melo while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd) 1521da177e4SLinus Torvalds cwnd >>= 1; 1531da177e4SLinus Torvalds tp->snd_cwnd = max(cwnd, restart_cwnd); 154c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 1551da177e4SLinus Torvalds tp->snd_cwnd_used = 0; 1561da177e4SLinus Torvalds } 1571da177e4SLinus Torvalds 15867edfef7SAndi Kleen /* Congestion state accounting after a packet has been sent. */ 15940efc6faSStephen Hemminger static void tcp_event_data_sent(struct tcp_sock *tp, 160cf533ea5SEric Dumazet struct sock *sk) 1611da177e4SLinus Torvalds { 162463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 163d635fbe2SEric Dumazet const u32 now = tcp_jiffies32; 1641da177e4SLinus Torvalds 16505c5a46dSNeal Cardwell if (tcp_packets_in_flight(tp) == 0) 16605c5a46dSNeal Cardwell tcp_ca_event(sk, CA_EVENT_TX_START); 16705c5a46dSNeal Cardwell 1681da177e4SLinus Torvalds tp->lsndtime = now; 1691da177e4SLinus Torvalds 1701da177e4SLinus Torvalds /* If it is a reply for ato after last received 1711da177e4SLinus Torvalds * packet, enter pingpong mode. 1721da177e4SLinus Torvalds */ 1732251ae46SJon Maxwell if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato) 174463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pingpong = 1; 1751da177e4SLinus Torvalds } 1761da177e4SLinus Torvalds 17767edfef7SAndi Kleen /* Account for an ACK we sent. */ 17840efc6faSStephen Hemminger static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts) 1791da177e4SLinus Torvalds { 180463c84b9SArnaldo Carvalho de Melo tcp_dec_quickack_mode(sk, pkts); 181463c84b9SArnaldo Carvalho de Melo inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); 1821da177e4SLinus Torvalds } 1831da177e4SLinus Torvalds 18485f16525SYuchung Cheng 18585f16525SYuchung Cheng u32 tcp_default_init_rwnd(u32 mss) 18685f16525SYuchung Cheng { 18785f16525SYuchung Cheng /* Initial receive window should be twice of TCP_INIT_CWND to 1889ef71e0cSWeiping Pan * enable proper sending of new unsent data during fast recovery 18985f16525SYuchung Cheng * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a 19085f16525SYuchung Cheng * limit when mss is larger than 1460. 19185f16525SYuchung Cheng */ 19285f16525SYuchung Cheng u32 init_rwnd = TCP_INIT_CWND * 2; 19385f16525SYuchung Cheng 19485f16525SYuchung Cheng if (mss > 1460) 19585f16525SYuchung Cheng init_rwnd = max((1460 * init_rwnd) / mss, 2U); 19685f16525SYuchung Cheng return init_rwnd; 19785f16525SYuchung Cheng } 19885f16525SYuchung Cheng 1991da177e4SLinus Torvalds /* Determine a window scaling and initial window to offer. 2001da177e4SLinus Torvalds * Based on the assumption that the given amount of space 2011da177e4SLinus Torvalds * will be offered. Store the results in the tp structure. 2021da177e4SLinus Torvalds * NOTE: for smooth operation initial space offering should 2031da177e4SLinus Torvalds * be a multiple of mss if possible. We assume here that mss >= 1. 2041da177e4SLinus Torvalds * This MUST be enforced by all callers. 2051da177e4SLinus Torvalds */ 2061da177e4SLinus Torvalds void tcp_select_initial_window(int __space, __u32 mss, 2071da177e4SLinus Torvalds __u32 *rcv_wnd, __u32 *window_clamp, 20831d12926Slaurent chavey int wscale_ok, __u8 *rcv_wscale, 20931d12926Slaurent chavey __u32 init_rcv_wnd) 2101da177e4SLinus Torvalds { 2111da177e4SLinus Torvalds unsigned int space = (__space < 0 ? 0 : __space); 2121da177e4SLinus Torvalds 2131da177e4SLinus Torvalds /* If no clamp set the clamp to the max possible scaled window */ 2141da177e4SLinus Torvalds if (*window_clamp == 0) 215589c49cbSGao Feng (*window_clamp) = (U16_MAX << TCP_MAX_WSCALE); 2161da177e4SLinus Torvalds space = min(*window_clamp, space); 2171da177e4SLinus Torvalds 2181da177e4SLinus Torvalds /* Quantize space offering to a multiple of mss if possible. */ 2191da177e4SLinus Torvalds if (space > mss) 220589c49cbSGao Feng space = rounddown(space, mss); 2211da177e4SLinus Torvalds 2221da177e4SLinus Torvalds /* NOTE: offering an initial window larger than 32767 22315d99e02SRick Jones * will break some buggy TCP stacks. If the admin tells us 22415d99e02SRick Jones * it is likely we could be speaking with such a buggy stack 22515d99e02SRick Jones * we will truncate our initial window offering to 32K-1 22615d99e02SRick Jones * unless the remote has sent us a window scaling option, 22715d99e02SRick Jones * which we interpret as a sign the remote TCP is not 22815d99e02SRick Jones * misinterpreting the window field as a signed quantity. 2291da177e4SLinus Torvalds */ 23015d99e02SRick Jones if (sysctl_tcp_workaround_signed_windows) 2311da177e4SLinus Torvalds (*rcv_wnd) = min(space, MAX_TCP_WINDOW); 23215d99e02SRick Jones else 23315d99e02SRick Jones (*rcv_wnd) = space; 23415d99e02SRick Jones 2351da177e4SLinus Torvalds (*rcv_wscale) = 0; 2361da177e4SLinus Torvalds if (wscale_ok) { 237589c49cbSGao Feng /* Set window scaling on max possible window */ 238f626300aSSoheil Hassas Yeganeh space = max_t(u32, space, sysctl_tcp_rmem[2]); 239f626300aSSoheil Hassas Yeganeh space = max_t(u32, space, sysctl_rmem_max); 240316c1592SStephen Hemminger space = min_t(u32, space, *window_clamp); 241589c49cbSGao Feng while (space > U16_MAX && (*rcv_wscale) < TCP_MAX_WSCALE) { 2421da177e4SLinus Torvalds space >>= 1; 2431da177e4SLinus Torvalds (*rcv_wscale)++; 2441da177e4SLinus Torvalds } 2451da177e4SLinus Torvalds } 2461da177e4SLinus Torvalds 2471da177e4SLinus Torvalds if (mss > (1 << *rcv_wscale)) { 24885f16525SYuchung Cheng if (!init_rcv_wnd) /* Use default unless specified otherwise */ 24985f16525SYuchung Cheng init_rcv_wnd = tcp_default_init_rwnd(mss); 250b1afde60SNandita Dukkipati *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss); 2511da177e4SLinus Torvalds } 2521da177e4SLinus Torvalds 2531da177e4SLinus Torvalds /* Set the clamp no higher than max representable value */ 254589c49cbSGao Feng (*window_clamp) = min_t(__u32, U16_MAX << (*rcv_wscale), *window_clamp); 2551da177e4SLinus Torvalds } 2564bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_select_initial_window); 2571da177e4SLinus Torvalds 2581da177e4SLinus Torvalds /* Chose a new window to advertise, update state in tcp_sock for the 2591da177e4SLinus Torvalds * socket, and return result with RFC1323 scaling applied. The return 2601da177e4SLinus Torvalds * value can be stuffed directly into th->window for an outgoing 2611da177e4SLinus Torvalds * frame. 2621da177e4SLinus Torvalds */ 26340efc6faSStephen Hemminger static u16 tcp_select_window(struct sock *sk) 2641da177e4SLinus Torvalds { 2651da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2668e165e20SFlorian Westphal u32 old_win = tp->rcv_wnd; 2671da177e4SLinus Torvalds u32 cur_win = tcp_receive_window(tp); 2681da177e4SLinus Torvalds u32 new_win = __tcp_select_window(sk); 2691da177e4SLinus Torvalds 2701da177e4SLinus Torvalds /* Never shrink the offered window */ 2711da177e4SLinus Torvalds if (new_win < cur_win) { 2721da177e4SLinus Torvalds /* Danger Will Robinson! 2731da177e4SLinus Torvalds * Don't update rcv_wup/rcv_wnd here or else 2741da177e4SLinus Torvalds * we will not be able to advertise a zero 2751da177e4SLinus Torvalds * window in time. --DaveM 2761da177e4SLinus Torvalds * 2771da177e4SLinus Torvalds * Relax Will Robinson. 2781da177e4SLinus Torvalds */ 2798e165e20SFlorian Westphal if (new_win == 0) 2808e165e20SFlorian Westphal NET_INC_STATS(sock_net(sk), 2818e165e20SFlorian Westphal LINUX_MIB_TCPWANTZEROWINDOWADV); 282607bfbf2SPatrick McHardy new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale); 2831da177e4SLinus Torvalds } 2841da177e4SLinus Torvalds tp->rcv_wnd = new_win; 2851da177e4SLinus Torvalds tp->rcv_wup = tp->rcv_nxt; 2861da177e4SLinus Torvalds 2871da177e4SLinus Torvalds /* Make sure we do not exceed the maximum possible 2881da177e4SLinus Torvalds * scaled window. 2891da177e4SLinus Torvalds */ 29015d99e02SRick Jones if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows) 2911da177e4SLinus Torvalds new_win = min(new_win, MAX_TCP_WINDOW); 2921da177e4SLinus Torvalds else 2931da177e4SLinus Torvalds new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); 2941da177e4SLinus Torvalds 2951da177e4SLinus Torvalds /* RFC1323 scaling applied */ 2961da177e4SLinus Torvalds new_win >>= tp->rx_opt.rcv_wscale; 2971da177e4SLinus Torvalds 2981da177e4SLinus Torvalds /* If we advertise zero window, disable fast path. */ 2998e165e20SFlorian Westphal if (new_win == 0) { 3001da177e4SLinus Torvalds tp->pred_flags = 0; 3018e165e20SFlorian Westphal if (old_win) 3028e165e20SFlorian Westphal NET_INC_STATS(sock_net(sk), 3038e165e20SFlorian Westphal LINUX_MIB_TCPTOZEROWINDOWADV); 3048e165e20SFlorian Westphal } else if (old_win == 0) { 3058e165e20SFlorian Westphal NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV); 3068e165e20SFlorian Westphal } 3071da177e4SLinus Torvalds 3081da177e4SLinus Torvalds return new_win; 3091da177e4SLinus Torvalds } 3101da177e4SLinus Torvalds 31167edfef7SAndi Kleen /* Packet ECN state for a SYN-ACK */ 312735d3831SFlorian Westphal static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb) 313bdf1ee5dSIlpo Järvinen { 31430e502a3SDaniel Borkmann const struct tcp_sock *tp = tcp_sk(sk); 31530e502a3SDaniel Borkmann 3164de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR; 317bdf1ee5dSIlpo Järvinen if (!(tp->ecn_flags & TCP_ECN_OK)) 3184de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE; 31930e502a3SDaniel Borkmann else if (tcp_ca_needs_ecn(sk)) 32030e502a3SDaniel Borkmann INET_ECN_xmit(sk); 321bdf1ee5dSIlpo Järvinen } 322bdf1ee5dSIlpo Järvinen 32367edfef7SAndi Kleen /* Packet ECN state for a SYN. */ 324735d3831SFlorian Westphal static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb) 325bdf1ee5dSIlpo Järvinen { 326bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 327f7b3bec6SFlorian Westphal bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 || 328f7b3bec6SFlorian Westphal tcp_ca_needs_ecn(sk); 329f7b3bec6SFlorian Westphal 330f7b3bec6SFlorian Westphal if (!use_ecn) { 331f7b3bec6SFlorian Westphal const struct dst_entry *dst = __sk_dst_get(sk); 332f7b3bec6SFlorian Westphal 333f7b3bec6SFlorian Westphal if (dst && dst_feature(dst, RTAX_FEATURE_ECN)) 334f7b3bec6SFlorian Westphal use_ecn = true; 335f7b3bec6SFlorian Westphal } 336bdf1ee5dSIlpo Järvinen 337bdf1ee5dSIlpo Järvinen tp->ecn_flags = 0; 338f7b3bec6SFlorian Westphal 339f7b3bec6SFlorian Westphal if (use_ecn) { 3404de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR; 341bdf1ee5dSIlpo Järvinen tp->ecn_flags = TCP_ECN_OK; 34230e502a3SDaniel Borkmann if (tcp_ca_needs_ecn(sk)) 34330e502a3SDaniel Borkmann INET_ECN_xmit(sk); 344bdf1ee5dSIlpo Järvinen } 345bdf1ee5dSIlpo Järvinen } 346bdf1ee5dSIlpo Järvinen 34749213555SDaniel Borkmann static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb) 34849213555SDaniel Borkmann { 34949213555SDaniel Borkmann if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback) 35049213555SDaniel Borkmann /* tp->ecn_flags are cleared at a later point in time when 35149213555SDaniel Borkmann * SYN ACK is ultimatively being received. 35249213555SDaniel Borkmann */ 35349213555SDaniel Borkmann TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR); 35449213555SDaniel Borkmann } 35549213555SDaniel Borkmann 356735d3831SFlorian Westphal static void 3576ac705b1SEric Dumazet tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th) 358bdf1ee5dSIlpo Järvinen { 3596ac705b1SEric Dumazet if (inet_rsk(req)->ecn_ok) 360bdf1ee5dSIlpo Järvinen th->ece = 1; 361bdf1ee5dSIlpo Järvinen } 362bdf1ee5dSIlpo Järvinen 36367edfef7SAndi Kleen /* Set up ECN state for a packet on a ESTABLISHED socket that is about to 36467edfef7SAndi Kleen * be sent. 36567edfef7SAndi Kleen */ 366735d3831SFlorian Westphal static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb, 367ea1627c2SEric Dumazet struct tcphdr *th, int tcp_header_len) 368bdf1ee5dSIlpo Järvinen { 369bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 370bdf1ee5dSIlpo Järvinen 371bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_OK) { 372bdf1ee5dSIlpo Järvinen /* Not-retransmitted data segment: set ECT and inject CWR. */ 373bdf1ee5dSIlpo Järvinen if (skb->len != tcp_header_len && 374bdf1ee5dSIlpo Järvinen !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { 375bdf1ee5dSIlpo Järvinen INET_ECN_xmit(sk); 376bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) { 377bdf1ee5dSIlpo Järvinen tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; 378ea1627c2SEric Dumazet th->cwr = 1; 379bdf1ee5dSIlpo Järvinen skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; 380bdf1ee5dSIlpo Järvinen } 38130e502a3SDaniel Borkmann } else if (!tcp_ca_needs_ecn(sk)) { 382bdf1ee5dSIlpo Järvinen /* ACK or retransmitted segment: clear ECT|CE */ 383bdf1ee5dSIlpo Järvinen INET_ECN_dontxmit(sk); 384bdf1ee5dSIlpo Järvinen } 385bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) 386ea1627c2SEric Dumazet th->ece = 1; 387bdf1ee5dSIlpo Järvinen } 388bdf1ee5dSIlpo Järvinen } 389bdf1ee5dSIlpo Järvinen 390e870a8efSIlpo Järvinen /* Constructs common control bits of non-data skb. If SYN/FIN is present, 391e870a8efSIlpo Järvinen * auto increment end seqno. 392e870a8efSIlpo Järvinen */ 393e870a8efSIlpo Järvinen static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags) 394e870a8efSIlpo Järvinen { 3952e8e18efSDavid S. Miller skb->ip_summed = CHECKSUM_PARTIAL; 396e870a8efSIlpo Järvinen skb->csum = 0; 397e870a8efSIlpo Järvinen 3984de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags; 399e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->sacked = 0; 400e870a8efSIlpo Järvinen 401cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, 1); 402e870a8efSIlpo Järvinen 403e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->seq = seq; 404a3433f35SChangli Gao if (flags & (TCPHDR_SYN | TCPHDR_FIN)) 405e870a8efSIlpo Järvinen seq++; 406e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->end_seq = seq; 407e870a8efSIlpo Järvinen } 408e870a8efSIlpo Järvinen 409a2a385d6SEric Dumazet static inline bool tcp_urg_mode(const struct tcp_sock *tp) 41033f5f57eSIlpo Järvinen { 41133f5f57eSIlpo Järvinen return tp->snd_una != tp->snd_up; 41233f5f57eSIlpo Järvinen } 41333f5f57eSIlpo Järvinen 41433ad798cSAdam Langley #define OPTION_SACK_ADVERTISE (1 << 0) 41533ad798cSAdam Langley #define OPTION_TS (1 << 1) 41633ad798cSAdam Langley #define OPTION_MD5 (1 << 2) 41789e95a61SOri Finkelman #define OPTION_WSCALE (1 << 3) 4182100c8d2SYuchung Cheng #define OPTION_FAST_OPEN_COOKIE (1 << 8) 41933ad798cSAdam Langley 42033ad798cSAdam Langley struct tcp_out_options { 4212100c8d2SYuchung Cheng u16 options; /* bit field of OPTION_* */ 4222100c8d2SYuchung Cheng u16 mss; /* 0 to disable */ 42333ad798cSAdam Langley u8 ws; /* window scale, 0 to disable */ 42433ad798cSAdam Langley u8 num_sack_blocks; /* number of SACK blocks to include */ 425bd0388aeSWilliam Allen Simpson u8 hash_size; /* bytes in hash_location */ 426bd0388aeSWilliam Allen Simpson __u8 *hash_location; /* temporary pointer, overloaded */ 4272100c8d2SYuchung Cheng __u32 tsval, tsecr; /* need to include OPTION_TS */ 4282100c8d2SYuchung Cheng struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */ 42933ad798cSAdam Langley }; 43033ad798cSAdam Langley 43167edfef7SAndi Kleen /* Write previously computed TCP options to the packet. 43267edfef7SAndi Kleen * 43367edfef7SAndi Kleen * Beware: Something in the Internet is very sensitive to the ordering of 434fd6149d3SIlpo Järvinen * TCP options, we learned this through the hard way, so be careful here. 435fd6149d3SIlpo Järvinen * Luckily we can at least blame others for their non-compliance but from 4368e3bff96Sstephen hemminger * inter-operability perspective it seems that we're somewhat stuck with 437fd6149d3SIlpo Järvinen * the ordering which we have been using if we want to keep working with 438fd6149d3SIlpo Järvinen * those broken things (not that it currently hurts anybody as there isn't 439fd6149d3SIlpo Järvinen * particular reason why the ordering would need to be changed). 440fd6149d3SIlpo Järvinen * 441fd6149d3SIlpo Järvinen * At least SACK_PERM as the first option is known to lead to a disaster 442fd6149d3SIlpo Järvinen * (but it may well be that other scenarios fail similarly). 443fd6149d3SIlpo Järvinen */ 44433ad798cSAdam Langley static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp, 445bd0388aeSWilliam Allen Simpson struct tcp_out_options *opts) 446bd0388aeSWilliam Allen Simpson { 4472100c8d2SYuchung Cheng u16 options = opts->options; /* mungable copy */ 448bd0388aeSWilliam Allen Simpson 449bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_MD5 & options)) { 4501a2c6181SChristoph Paasch *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | 4511a2c6181SChristoph Paasch (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG); 452bd0388aeSWilliam Allen Simpson /* overload cookie hash location */ 453bd0388aeSWilliam Allen Simpson opts->hash_location = (__u8 *)ptr; 45433ad798cSAdam Langley ptr += 4; 45533ad798cSAdam Langley } 45633ad798cSAdam Langley 457fd6149d3SIlpo Järvinen if (unlikely(opts->mss)) { 458fd6149d3SIlpo Järvinen *ptr++ = htonl((TCPOPT_MSS << 24) | 459fd6149d3SIlpo Järvinen (TCPOLEN_MSS << 16) | 460fd6149d3SIlpo Järvinen opts->mss); 461fd6149d3SIlpo Järvinen } 462fd6149d3SIlpo Järvinen 463bd0388aeSWilliam Allen Simpson if (likely(OPTION_TS & options)) { 464bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 46533ad798cSAdam Langley *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | 46633ad798cSAdam Langley (TCPOLEN_SACK_PERM << 16) | 46733ad798cSAdam Langley (TCPOPT_TIMESTAMP << 8) | 46833ad798cSAdam Langley TCPOLEN_TIMESTAMP); 469bd0388aeSWilliam Allen Simpson options &= ~OPTION_SACK_ADVERTISE; 47033ad798cSAdam Langley } else { 471496c98dfSYOSHIFUJI Hideaki *ptr++ = htonl((TCPOPT_NOP << 24) | 47240efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 47340efc6faSStephen Hemminger (TCPOPT_TIMESTAMP << 8) | 47440efc6faSStephen Hemminger TCPOLEN_TIMESTAMP); 47540efc6faSStephen Hemminger } 47633ad798cSAdam Langley *ptr++ = htonl(opts->tsval); 47733ad798cSAdam Langley *ptr++ = htonl(opts->tsecr); 47833ad798cSAdam Langley } 47933ad798cSAdam Langley 480bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 48133ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 48233ad798cSAdam Langley (TCPOPT_NOP << 16) | 48333ad798cSAdam Langley (TCPOPT_SACK_PERM << 8) | 48433ad798cSAdam Langley TCPOLEN_SACK_PERM); 48533ad798cSAdam Langley } 48633ad798cSAdam Langley 487bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_WSCALE & options)) { 48833ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 48933ad798cSAdam Langley (TCPOPT_WINDOW << 16) | 49033ad798cSAdam Langley (TCPOLEN_WINDOW << 8) | 49133ad798cSAdam Langley opts->ws); 49233ad798cSAdam Langley } 49333ad798cSAdam Langley 49433ad798cSAdam Langley if (unlikely(opts->num_sack_blocks)) { 49533ad798cSAdam Langley struct tcp_sack_block *sp = tp->rx_opt.dsack ? 49633ad798cSAdam Langley tp->duplicate_sack : tp->selective_acks; 49740efc6faSStephen Hemminger int this_sack; 49840efc6faSStephen Hemminger 49940efc6faSStephen Hemminger *ptr++ = htonl((TCPOPT_NOP << 24) | 50040efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 50140efc6faSStephen Hemminger (TCPOPT_SACK << 8) | 50233ad798cSAdam Langley (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * 50340efc6faSStephen Hemminger TCPOLEN_SACK_PERBLOCK))); 5042de979bdSStephen Hemminger 50533ad798cSAdam Langley for (this_sack = 0; this_sack < opts->num_sack_blocks; 50633ad798cSAdam Langley ++this_sack) { 50740efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].start_seq); 50840efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].end_seq); 50940efc6faSStephen Hemminger } 5102de979bdSStephen Hemminger 51140efc6faSStephen Hemminger tp->rx_opt.dsack = 0; 51240efc6faSStephen Hemminger } 5132100c8d2SYuchung Cheng 5142100c8d2SYuchung Cheng if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) { 5152100c8d2SYuchung Cheng struct tcp_fastopen_cookie *foc = opts->fastopen_cookie; 5167f9b838bSDaniel Lee u8 *p = (u8 *)ptr; 5177f9b838bSDaniel Lee u32 len; /* Fast Open option length */ 5182100c8d2SYuchung Cheng 5197f9b838bSDaniel Lee if (foc->exp) { 5207f9b838bSDaniel Lee len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len; 5217f9b838bSDaniel Lee *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) | 5222100c8d2SYuchung Cheng TCPOPT_FASTOPEN_MAGIC); 5237f9b838bSDaniel Lee p += TCPOLEN_EXP_FASTOPEN_BASE; 5247f9b838bSDaniel Lee } else { 5257f9b838bSDaniel Lee len = TCPOLEN_FASTOPEN_BASE + foc->len; 5267f9b838bSDaniel Lee *p++ = TCPOPT_FASTOPEN; 5277f9b838bSDaniel Lee *p++ = len; 5282100c8d2SYuchung Cheng } 5297f9b838bSDaniel Lee 5307f9b838bSDaniel Lee memcpy(p, foc->val, foc->len); 5317f9b838bSDaniel Lee if ((len & 3) == 2) { 5327f9b838bSDaniel Lee p[foc->len] = TCPOPT_NOP; 5337f9b838bSDaniel Lee p[foc->len + 1] = TCPOPT_NOP; 5347f9b838bSDaniel Lee } 5357f9b838bSDaniel Lee ptr += (len + 3) >> 2; 5362100c8d2SYuchung Cheng } 53740efc6faSStephen Hemminger } 53840efc6faSStephen Hemminger 53967edfef7SAndi Kleen /* Compute TCP options for SYN packets. This is not the final 54067edfef7SAndi Kleen * network wire format yet. 54167edfef7SAndi Kleen */ 54295c96174SEric Dumazet static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb, 54333ad798cSAdam Langley struct tcp_out_options *opts, 544cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 545cf533ea5SEric Dumazet { 54633ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 54795c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 548783237e8SYuchung Cheng struct tcp_fastopen_request *fastopen = tp->fastopen_req; 54933ad798cSAdam Langley 550cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 55133ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 55233ad798cSAdam Langley if (*md5) { 55333ad798cSAdam Langley opts->options |= OPTION_MD5; 554bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 555cfb6eeb4SYOSHIFUJI Hideaki } 55633ad798cSAdam Langley #else 55733ad798cSAdam Langley *md5 = NULL; 558cfb6eeb4SYOSHIFUJI Hideaki #endif 55933ad798cSAdam Langley 56033ad798cSAdam Langley /* We always get an MSS option. The option bytes which will be seen in 56133ad798cSAdam Langley * normal data packets should timestamps be used, must be in the MSS 56233ad798cSAdam Langley * advertised. But we subtract them from tp->mss_cache so that 56333ad798cSAdam Langley * calculations in tcp_sendmsg are simpler etc. So account for this 56433ad798cSAdam Langley * fact here if necessary. If we don't do this correctly, as a 56533ad798cSAdam Langley * receiver we won't recognize data packets as being full sized when we 56633ad798cSAdam Langley * should, and thus we won't abide by the delayed ACK rules correctly. 56733ad798cSAdam Langley * SACKs don't matter, we never delay an ACK when we have any of those 56833ad798cSAdam Langley * going out. */ 56933ad798cSAdam Langley opts->mss = tcp_advertise_mss(sk); 570bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 57133ad798cSAdam Langley 57251456b29SIan Morris if (likely(sysctl_tcp_timestamps && !*md5)) { 57333ad798cSAdam Langley opts->options |= OPTION_TS; 5747faee5c0SEric Dumazet opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset; 57533ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 576bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 57733ad798cSAdam Langley } 578bb5b7c11SDavid S. Miller if (likely(sysctl_tcp_window_scaling)) { 57933ad798cSAdam Langley opts->ws = tp->rx_opt.rcv_wscale; 58089e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 581bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 58233ad798cSAdam Langley } 583bb5b7c11SDavid S. Miller if (likely(sysctl_tcp_sack)) { 58433ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 585b32d1310SDavid S. Miller if (unlikely(!(OPTION_TS & opts->options))) 586bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 58733ad798cSAdam Langley } 58833ad798cSAdam Langley 589783237e8SYuchung Cheng if (fastopen && fastopen->cookie.len >= 0) { 5902646c831SDaniel Lee u32 need = fastopen->cookie.len; 5912646c831SDaniel Lee 5922646c831SDaniel Lee need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE : 5932646c831SDaniel Lee TCPOLEN_FASTOPEN_BASE; 594783237e8SYuchung Cheng need = (need + 3) & ~3U; /* Align to 32 bits */ 595783237e8SYuchung Cheng if (remaining >= need) { 596783237e8SYuchung Cheng opts->options |= OPTION_FAST_OPEN_COOKIE; 597783237e8SYuchung Cheng opts->fastopen_cookie = &fastopen->cookie; 598783237e8SYuchung Cheng remaining -= need; 599783237e8SYuchung Cheng tp->syn_fastopen = 1; 6002646c831SDaniel Lee tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0; 601783237e8SYuchung Cheng } 602783237e8SYuchung Cheng } 603bd0388aeSWilliam Allen Simpson 604bd0388aeSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 60533ad798cSAdam Langley } 60633ad798cSAdam Langley 60767edfef7SAndi Kleen /* Set up TCP options for SYN-ACKs. */ 60837bfbddaSEric Dumazet static unsigned int tcp_synack_options(struct request_sock *req, 60995c96174SEric Dumazet unsigned int mss, struct sk_buff *skb, 61033ad798cSAdam Langley struct tcp_out_options *opts, 61180f03e27SEric Dumazet const struct tcp_md5sig_key *md5, 6128336886fSJerry Chu struct tcp_fastopen_cookie *foc) 6134957faadSWilliam Allen Simpson { 61433ad798cSAdam Langley struct inet_request_sock *ireq = inet_rsk(req); 61595c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 61633ad798cSAdam Langley 61733ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 61880f03e27SEric Dumazet if (md5) { 61933ad798cSAdam Langley opts->options |= OPTION_MD5; 6204957faadSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 6214957faadSWilliam Allen Simpson 6224957faadSWilliam Allen Simpson /* We can't fit any SACK blocks in a packet with MD5 + TS 6234957faadSWilliam Allen Simpson * options. There was discussion about disabling SACK 6244957faadSWilliam Allen Simpson * rather than TS in order to fit in better with old, 6254957faadSWilliam Allen Simpson * buggy kernels, but that was deemed to be unnecessary. 6264957faadSWilliam Allen Simpson */ 627de213e5eSEric Dumazet ireq->tstamp_ok &= !ireq->sack_ok; 62833ad798cSAdam Langley } 62933ad798cSAdam Langley #endif 63033ad798cSAdam Langley 6314957faadSWilliam Allen Simpson /* We always send an MSS option. */ 63233ad798cSAdam Langley opts->mss = mss; 6334957faadSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 63433ad798cSAdam Langley 63533ad798cSAdam Langley if (likely(ireq->wscale_ok)) { 63633ad798cSAdam Langley opts->ws = ireq->rcv_wscale; 63789e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 6384957faadSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 63933ad798cSAdam Langley } 640de213e5eSEric Dumazet if (likely(ireq->tstamp_ok)) { 64133ad798cSAdam Langley opts->options |= OPTION_TS; 64295a22caeSFlorian Westphal opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off; 64333ad798cSAdam Langley opts->tsecr = req->ts_recent; 6444957faadSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 64533ad798cSAdam Langley } 64633ad798cSAdam Langley if (likely(ireq->sack_ok)) { 64733ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 648de213e5eSEric Dumazet if (unlikely(!ireq->tstamp_ok)) 6494957faadSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 65033ad798cSAdam Langley } 6517f9b838bSDaniel Lee if (foc != NULL && foc->len >= 0) { 6527f9b838bSDaniel Lee u32 need = foc->len; 6537f9b838bSDaniel Lee 6547f9b838bSDaniel Lee need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE : 6557f9b838bSDaniel Lee TCPOLEN_FASTOPEN_BASE; 6568336886fSJerry Chu need = (need + 3) & ~3U; /* Align to 32 bits */ 6578336886fSJerry Chu if (remaining >= need) { 6588336886fSJerry Chu opts->options |= OPTION_FAST_OPEN_COOKIE; 6598336886fSJerry Chu opts->fastopen_cookie = foc; 6608336886fSJerry Chu remaining -= need; 6618336886fSJerry Chu } 6628336886fSJerry Chu } 6634957faadSWilliam Allen Simpson 6644957faadSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 66533ad798cSAdam Langley } 66633ad798cSAdam Langley 66767edfef7SAndi Kleen /* Compute TCP options for ESTABLISHED sockets. This is not the 66867edfef7SAndi Kleen * final wire format yet. 66967edfef7SAndi Kleen */ 67095c96174SEric Dumazet static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb, 67133ad798cSAdam Langley struct tcp_out_options *opts, 672cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 673cf533ea5SEric Dumazet { 67433ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 67595c96174SEric Dumazet unsigned int size = 0; 676cabeccbdSIlpo Järvinen unsigned int eff_sacks; 67733ad798cSAdam Langley 6785843ef42SAndi Kleen opts->options = 0; 6795843ef42SAndi Kleen 68033ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 68133ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 68233ad798cSAdam Langley if (unlikely(*md5)) { 68333ad798cSAdam Langley opts->options |= OPTION_MD5; 68433ad798cSAdam Langley size += TCPOLEN_MD5SIG_ALIGNED; 68533ad798cSAdam Langley } 68633ad798cSAdam Langley #else 68733ad798cSAdam Langley *md5 = NULL; 68833ad798cSAdam Langley #endif 68933ad798cSAdam Langley 69033ad798cSAdam Langley if (likely(tp->rx_opt.tstamp_ok)) { 69133ad798cSAdam Langley opts->options |= OPTION_TS; 6927faee5c0SEric Dumazet opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0; 69333ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 69433ad798cSAdam Langley size += TCPOLEN_TSTAMP_ALIGNED; 69533ad798cSAdam Langley } 69633ad798cSAdam Langley 697cabeccbdSIlpo Järvinen eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; 698cabeccbdSIlpo Järvinen if (unlikely(eff_sacks)) { 69995c96174SEric Dumazet const unsigned int remaining = MAX_TCP_OPTION_SPACE - size; 70033ad798cSAdam Langley opts->num_sack_blocks = 70195c96174SEric Dumazet min_t(unsigned int, eff_sacks, 70233ad798cSAdam Langley (remaining - TCPOLEN_SACK_BASE_ALIGNED) / 70333ad798cSAdam Langley TCPOLEN_SACK_PERBLOCK); 70433ad798cSAdam Langley size += TCPOLEN_SACK_BASE_ALIGNED + 70533ad798cSAdam Langley opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; 70633ad798cSAdam Langley } 70733ad798cSAdam Langley 70833ad798cSAdam Langley return size; 70940efc6faSStephen Hemminger } 7101da177e4SLinus Torvalds 71146d3ceabSEric Dumazet 71246d3ceabSEric Dumazet /* TCP SMALL QUEUES (TSQ) 71346d3ceabSEric Dumazet * 71446d3ceabSEric Dumazet * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev) 71546d3ceabSEric Dumazet * to reduce RTT and bufferbloat. 71646d3ceabSEric Dumazet * We do this using a special skb destructor (tcp_wfree). 71746d3ceabSEric Dumazet * 71846d3ceabSEric Dumazet * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb 71946d3ceabSEric Dumazet * needs to be reallocated in a driver. 7208e3bff96Sstephen hemminger * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc 72146d3ceabSEric Dumazet * 72246d3ceabSEric Dumazet * Since transmit from skb destructor is forbidden, we use a tasklet 72346d3ceabSEric Dumazet * to process all sockets that eventually need to send more skbs. 72446d3ceabSEric Dumazet * We use one tasklet per cpu, with its own queue of sockets. 72546d3ceabSEric Dumazet */ 72646d3ceabSEric Dumazet struct tsq_tasklet { 72746d3ceabSEric Dumazet struct tasklet_struct tasklet; 72846d3ceabSEric Dumazet struct list_head head; /* queue of tcp sockets */ 72946d3ceabSEric Dumazet }; 73046d3ceabSEric Dumazet static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet); 73146d3ceabSEric Dumazet 7326f458dfbSEric Dumazet static void tcp_tsq_handler(struct sock *sk) 7336f458dfbSEric Dumazet { 7346f458dfbSEric Dumazet if ((1 << sk->sk_state) & 7356f458dfbSEric Dumazet (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING | 736f9616c35SEric Dumazet TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) { 737f9616c35SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 738f9616c35SEric Dumazet 739f9616c35SEric Dumazet if (tp->lost_out > tp->retrans_out && 740f9616c35SEric Dumazet tp->snd_cwnd > tcp_packets_in_flight(tp)) 741f9616c35SEric Dumazet tcp_xmit_retransmit_queue(sk); 742f9616c35SEric Dumazet 743f9616c35SEric Dumazet tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle, 744bf06200eSJohn Ogness 0, GFP_ATOMIC); 7456f458dfbSEric Dumazet } 746f9616c35SEric Dumazet } 74746d3ceabSEric Dumazet /* 7488e3bff96Sstephen hemminger * One tasklet per cpu tries to send more skbs. 74946d3ceabSEric Dumazet * We run in tasklet context but need to disable irqs when 7508e3bff96Sstephen hemminger * transferring tsq->head because tcp_wfree() might 75146d3ceabSEric Dumazet * interrupt us (non NAPI drivers) 75246d3ceabSEric Dumazet */ 75346d3ceabSEric Dumazet static void tcp_tasklet_func(unsigned long data) 75446d3ceabSEric Dumazet { 75546d3ceabSEric Dumazet struct tsq_tasklet *tsq = (struct tsq_tasklet *)data; 75646d3ceabSEric Dumazet LIST_HEAD(list); 75746d3ceabSEric Dumazet unsigned long flags; 75846d3ceabSEric Dumazet struct list_head *q, *n; 75946d3ceabSEric Dumazet struct tcp_sock *tp; 76046d3ceabSEric Dumazet struct sock *sk; 76146d3ceabSEric Dumazet 76246d3ceabSEric Dumazet local_irq_save(flags); 76346d3ceabSEric Dumazet list_splice_init(&tsq->head, &list); 76446d3ceabSEric Dumazet local_irq_restore(flags); 76546d3ceabSEric Dumazet 76646d3ceabSEric Dumazet list_for_each_safe(q, n, &list) { 76746d3ceabSEric Dumazet tp = list_entry(q, struct tcp_sock, tsq_node); 76846d3ceabSEric Dumazet list_del(&tp->tsq_node); 76946d3ceabSEric Dumazet 77046d3ceabSEric Dumazet sk = (struct sock *)tp; 7710a9648f1SEric Dumazet smp_mb__before_atomic(); 7727aa5470cSEric Dumazet clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags); 7737aa5470cSEric Dumazet 774b223feb9SEric Dumazet if (!sk->sk_lock.owned && 7757aa5470cSEric Dumazet test_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags)) { 77646d3ceabSEric Dumazet bh_lock_sock(sk); 77746d3ceabSEric Dumazet if (!sock_owned_by_user(sk)) { 7787aa5470cSEric Dumazet clear_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags); 7796f458dfbSEric Dumazet tcp_tsq_handler(sk); 78046d3ceabSEric Dumazet } 78146d3ceabSEric Dumazet bh_unlock_sock(sk); 782b223feb9SEric Dumazet } 78346d3ceabSEric Dumazet 78446d3ceabSEric Dumazet sk_free(sk); 78546d3ceabSEric Dumazet } 78646d3ceabSEric Dumazet } 78746d3ceabSEric Dumazet 78840fc3423SEric Dumazet #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \ 78940fc3423SEric Dumazet TCPF_WRITE_TIMER_DEFERRED | \ 79040fc3423SEric Dumazet TCPF_DELACK_TIMER_DEFERRED | \ 79140fc3423SEric Dumazet TCPF_MTU_REDUCED_DEFERRED) 79246d3ceabSEric Dumazet /** 79346d3ceabSEric Dumazet * tcp_release_cb - tcp release_sock() callback 79446d3ceabSEric Dumazet * @sk: socket 79546d3ceabSEric Dumazet * 79646d3ceabSEric Dumazet * called from release_sock() to perform protocol dependent 79746d3ceabSEric Dumazet * actions before socket release. 79846d3ceabSEric Dumazet */ 79946d3ceabSEric Dumazet void tcp_release_cb(struct sock *sk) 80046d3ceabSEric Dumazet { 8016f458dfbSEric Dumazet unsigned long flags, nflags; 80246d3ceabSEric Dumazet 8036f458dfbSEric Dumazet /* perform an atomic operation only if at least one flag is set */ 8046f458dfbSEric Dumazet do { 8057aa5470cSEric Dumazet flags = sk->sk_tsq_flags; 8066f458dfbSEric Dumazet if (!(flags & TCP_DEFERRED_ALL)) 8076f458dfbSEric Dumazet return; 8086f458dfbSEric Dumazet nflags = flags & ~TCP_DEFERRED_ALL; 8097aa5470cSEric Dumazet } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags); 8106f458dfbSEric Dumazet 81140fc3423SEric Dumazet if (flags & TCPF_TSQ_DEFERRED) 8126f458dfbSEric Dumazet tcp_tsq_handler(sk); 8136f458dfbSEric Dumazet 814c3f9b018SEric Dumazet /* Here begins the tricky part : 815c3f9b018SEric Dumazet * We are called from release_sock() with : 816c3f9b018SEric Dumazet * 1) BH disabled 817c3f9b018SEric Dumazet * 2) sk_lock.slock spinlock held 818c3f9b018SEric Dumazet * 3) socket owned by us (sk->sk_lock.owned == 1) 819c3f9b018SEric Dumazet * 820c3f9b018SEric Dumazet * But following code is meant to be called from BH handlers, 821c3f9b018SEric Dumazet * so we should keep BH disabled, but early release socket ownership 822c3f9b018SEric Dumazet */ 823c3f9b018SEric Dumazet sock_release_ownership(sk); 824c3f9b018SEric Dumazet 82540fc3423SEric Dumazet if (flags & TCPF_WRITE_TIMER_DEFERRED) { 8266f458dfbSEric Dumazet tcp_write_timer_handler(sk); 827144d56e9SEric Dumazet __sock_put(sk); 828144d56e9SEric Dumazet } 82940fc3423SEric Dumazet if (flags & TCPF_DELACK_TIMER_DEFERRED) { 8306f458dfbSEric Dumazet tcp_delack_timer_handler(sk); 831144d56e9SEric Dumazet __sock_put(sk); 832144d56e9SEric Dumazet } 83340fc3423SEric Dumazet if (flags & TCPF_MTU_REDUCED_DEFERRED) { 8344fab9071SNeal Cardwell inet_csk(sk)->icsk_af_ops->mtu_reduced(sk); 835144d56e9SEric Dumazet __sock_put(sk); 836144d56e9SEric Dumazet } 83746d3ceabSEric Dumazet } 83846d3ceabSEric Dumazet EXPORT_SYMBOL(tcp_release_cb); 83946d3ceabSEric Dumazet 84046d3ceabSEric Dumazet void __init tcp_tasklet_init(void) 84146d3ceabSEric Dumazet { 84246d3ceabSEric Dumazet int i; 84346d3ceabSEric Dumazet 84446d3ceabSEric Dumazet for_each_possible_cpu(i) { 84546d3ceabSEric Dumazet struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i); 84646d3ceabSEric Dumazet 84746d3ceabSEric Dumazet INIT_LIST_HEAD(&tsq->head); 84846d3ceabSEric Dumazet tasklet_init(&tsq->tasklet, 84946d3ceabSEric Dumazet tcp_tasklet_func, 85046d3ceabSEric Dumazet (unsigned long)tsq); 85146d3ceabSEric Dumazet } 85246d3ceabSEric Dumazet } 85346d3ceabSEric Dumazet 85446d3ceabSEric Dumazet /* 85546d3ceabSEric Dumazet * Write buffer destructor automatically called from kfree_skb. 8568e3bff96Sstephen hemminger * We can't xmit new skbs from this context, as we might already 85746d3ceabSEric Dumazet * hold qdisc lock. 85846d3ceabSEric Dumazet */ 859d6a4a104SEric Dumazet void tcp_wfree(struct sk_buff *skb) 86046d3ceabSEric Dumazet { 86146d3ceabSEric Dumazet struct sock *sk = skb->sk; 86246d3ceabSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 863408f0a6cSEric Dumazet unsigned long flags, nval, oval; 8649b462d02SEric Dumazet int wmem; 8659b462d02SEric Dumazet 8669b462d02SEric Dumazet /* Keep one reference on sk_wmem_alloc. 8679b462d02SEric Dumazet * Will be released by sk_free() from here or tcp_tasklet_func() 8689b462d02SEric Dumazet */ 8699b462d02SEric Dumazet wmem = atomic_sub_return(skb->truesize - 1, &sk->sk_wmem_alloc); 8709b462d02SEric Dumazet 8719b462d02SEric Dumazet /* If this softirq is serviced by ksoftirqd, we are likely under stress. 8729b462d02SEric Dumazet * Wait until our queues (qdisc + devices) are drained. 8739b462d02SEric Dumazet * This gives : 8749b462d02SEric Dumazet * - less callbacks to tcp_write_xmit(), reducing stress (batches) 8759b462d02SEric Dumazet * - chance for incoming ACK (processed by another cpu maybe) 8769b462d02SEric Dumazet * to migrate this flow (skb->ooo_okay will be eventually set) 8779b462d02SEric Dumazet */ 8789b462d02SEric Dumazet if (wmem >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current) 8799b462d02SEric Dumazet goto out; 88046d3ceabSEric Dumazet 8817aa5470cSEric Dumazet for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) { 88246d3ceabSEric Dumazet struct tsq_tasklet *tsq; 883a9b204d1SEric Dumazet bool empty; 88446d3ceabSEric Dumazet 885408f0a6cSEric Dumazet if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED)) 886408f0a6cSEric Dumazet goto out; 887408f0a6cSEric Dumazet 888b223feb9SEric Dumazet nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED | TCPF_TSQ_DEFERRED; 8897aa5470cSEric Dumazet nval = cmpxchg(&sk->sk_tsq_flags, oval, nval); 890408f0a6cSEric Dumazet if (nval != oval) 891408f0a6cSEric Dumazet continue; 892408f0a6cSEric Dumazet 89346d3ceabSEric Dumazet /* queue this socket to tasklet queue */ 89446d3ceabSEric Dumazet local_irq_save(flags); 895903ceff7SChristoph Lameter tsq = this_cpu_ptr(&tsq_tasklet); 896a9b204d1SEric Dumazet empty = list_empty(&tsq->head); 89746d3ceabSEric Dumazet list_add(&tp->tsq_node, &tsq->head); 898a9b204d1SEric Dumazet if (empty) 89946d3ceabSEric Dumazet tasklet_schedule(&tsq->tasklet); 90046d3ceabSEric Dumazet local_irq_restore(flags); 9019b462d02SEric Dumazet return; 90246d3ceabSEric Dumazet } 9039b462d02SEric Dumazet out: 9049b462d02SEric Dumazet sk_free(sk); 90546d3ceabSEric Dumazet } 90646d3ceabSEric Dumazet 907218af599SEric Dumazet /* Note: Called under hard irq. 908218af599SEric Dumazet * We can not call TCP stack right away. 909218af599SEric Dumazet */ 910218af599SEric Dumazet enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer) 911218af599SEric Dumazet { 912218af599SEric Dumazet struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer); 913218af599SEric Dumazet struct sock *sk = (struct sock *)tp; 914218af599SEric Dumazet unsigned long nval, oval; 915218af599SEric Dumazet 916218af599SEric Dumazet for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) { 917218af599SEric Dumazet struct tsq_tasklet *tsq; 918218af599SEric Dumazet bool empty; 919218af599SEric Dumazet 920218af599SEric Dumazet if (oval & TSQF_QUEUED) 921218af599SEric Dumazet break; 922218af599SEric Dumazet 923218af599SEric Dumazet nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED | TCPF_TSQ_DEFERRED; 924218af599SEric Dumazet nval = cmpxchg(&sk->sk_tsq_flags, oval, nval); 925218af599SEric Dumazet if (nval != oval) 926218af599SEric Dumazet continue; 927218af599SEric Dumazet 928218af599SEric Dumazet if (!atomic_inc_not_zero(&sk->sk_wmem_alloc)) 929218af599SEric Dumazet break; 930218af599SEric Dumazet /* queue this socket to tasklet queue */ 931218af599SEric Dumazet tsq = this_cpu_ptr(&tsq_tasklet); 932218af599SEric Dumazet empty = list_empty(&tsq->head); 933218af599SEric Dumazet list_add(&tp->tsq_node, &tsq->head); 934218af599SEric Dumazet if (empty) 935218af599SEric Dumazet tasklet_schedule(&tsq->tasklet); 936218af599SEric Dumazet break; 937218af599SEric Dumazet } 938218af599SEric Dumazet return HRTIMER_NORESTART; 939218af599SEric Dumazet } 940218af599SEric Dumazet 941218af599SEric Dumazet /* BBR congestion control needs pacing. 942218af599SEric Dumazet * Same remark for SO_MAX_PACING_RATE. 943218af599SEric Dumazet * sch_fq packet scheduler is efficiently handling pacing, 944218af599SEric Dumazet * but is not always installed/used. 945218af599SEric Dumazet * Return true if TCP stack should pace packets itself. 946218af599SEric Dumazet */ 947218af599SEric Dumazet static bool tcp_needs_internal_pacing(const struct sock *sk) 948218af599SEric Dumazet { 949218af599SEric Dumazet return smp_load_acquire(&sk->sk_pacing_status) == SK_PACING_NEEDED; 950218af599SEric Dumazet } 951218af599SEric Dumazet 952218af599SEric Dumazet static void tcp_internal_pacing(struct sock *sk, const struct sk_buff *skb) 953218af599SEric Dumazet { 954218af599SEric Dumazet u64 len_ns; 955218af599SEric Dumazet u32 rate; 956218af599SEric Dumazet 957218af599SEric Dumazet if (!tcp_needs_internal_pacing(sk)) 958218af599SEric Dumazet return; 959218af599SEric Dumazet rate = sk->sk_pacing_rate; 960218af599SEric Dumazet if (!rate || rate == ~0U) 961218af599SEric Dumazet return; 962218af599SEric Dumazet 963218af599SEric Dumazet /* Should account for header sizes as sch_fq does, 964218af599SEric Dumazet * but lets make things simple. 965218af599SEric Dumazet */ 966218af599SEric Dumazet len_ns = (u64)skb->len * NSEC_PER_SEC; 967218af599SEric Dumazet do_div(len_ns, rate); 968218af599SEric Dumazet hrtimer_start(&tcp_sk(sk)->pacing_timer, 969218af599SEric Dumazet ktime_add_ns(ktime_get(), len_ns), 970218af599SEric Dumazet HRTIMER_MODE_ABS_PINNED); 971218af599SEric Dumazet } 972218af599SEric Dumazet 9731da177e4SLinus Torvalds /* This routine actually transmits TCP packets queued in by 9741da177e4SLinus Torvalds * tcp_do_sendmsg(). This is used by both the initial 9751da177e4SLinus Torvalds * transmission and possible later retransmissions. 9761da177e4SLinus Torvalds * All SKB's seen here are completely headerless. It is our 9771da177e4SLinus Torvalds * job to build the TCP header, and pass the packet down to 9781da177e4SLinus Torvalds * IP so it can do the same plus pass the packet off to the 9791da177e4SLinus Torvalds * device. 9801da177e4SLinus Torvalds * 9811da177e4SLinus Torvalds * We are working here with either a clone of the original 9821da177e4SLinus Torvalds * SKB, or a fresh unique copy made by the retransmit engine. 9831da177e4SLinus Torvalds */ 984056834d9SIlpo Järvinen static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, 985056834d9SIlpo Järvinen gfp_t gfp_mask) 9861da177e4SLinus Torvalds { 9876687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 988dfb4b9dcSDavid S. Miller struct inet_sock *inet; 989dfb4b9dcSDavid S. Miller struct tcp_sock *tp; 990dfb4b9dcSDavid S. Miller struct tcp_skb_cb *tcb; 99133ad798cSAdam Langley struct tcp_out_options opts; 99295c96174SEric Dumazet unsigned int tcp_options_size, tcp_header_size; 993cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 9941da177e4SLinus Torvalds struct tcphdr *th; 9951da177e4SLinus Torvalds int err; 9961da177e4SLinus Torvalds 997dfb4b9dcSDavid S. Miller BUG_ON(!skb || !tcp_skb_pcount(skb)); 9986f094b9eSLawrence Brakmo tp = tcp_sk(sk); 999dfb4b9dcSDavid S. Miller 1000385e2070SEric Dumazet skb->skb_mstamp = tp->tcp_mstamp; 1001ccdbb6e9SEric Dumazet if (clone_it) { 10026f094b9eSLawrence Brakmo TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq 10036f094b9eSLawrence Brakmo - tp->snd_una; 1004b9f64820SYuchung Cheng tcp_rate_skb_sent(sk, skb); 1005dfb4b9dcSDavid S. Miller 1006dfb4b9dcSDavid S. Miller if (unlikely(skb_cloned(skb))) 1007dfb4b9dcSDavid S. Miller skb = pskb_copy(skb, gfp_mask); 1008dfb4b9dcSDavid S. Miller else 1009dfb4b9dcSDavid S. Miller skb = skb_clone(skb, gfp_mask); 1010dfb4b9dcSDavid S. Miller if (unlikely(!skb)) 1011dfb4b9dcSDavid S. Miller return -ENOBUFS; 1012dfb4b9dcSDavid S. Miller } 1013dfb4b9dcSDavid S. Miller 1014dfb4b9dcSDavid S. Miller inet = inet_sk(sk); 1015dfb4b9dcSDavid S. Miller tcb = TCP_SKB_CB(skb); 101633ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 10171da177e4SLinus Torvalds 10184de075e0SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) 101933ad798cSAdam Langley tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); 102033ad798cSAdam Langley else 102133ad798cSAdam Langley tcp_options_size = tcp_established_options(sk, skb, &opts, 102233ad798cSAdam Langley &md5); 102333ad798cSAdam Langley tcp_header_size = tcp_options_size + sizeof(struct tcphdr); 10241da177e4SLinus Torvalds 1025547669d4SEric Dumazet /* if no packet is in qdisc/device queue, then allow XPS to select 1026b2532eb9SEric Dumazet * another queue. We can be called from tcp_tsq_handler() 1027b2532eb9SEric Dumazet * which holds one reference to sk_wmem_alloc. 1028b2532eb9SEric Dumazet * 1029b2532eb9SEric Dumazet * TODO: Ideally, in-flight pure ACK packets should not matter here. 1030b2532eb9SEric Dumazet * One way to get this would be to set skb->truesize = 2 on them. 1031547669d4SEric Dumazet */ 1032b2532eb9SEric Dumazet skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1); 10331da177e4SLinus Torvalds 103438ab52e8SEric Dumazet /* If we had to use memory reserve to allocate this skb, 103538ab52e8SEric Dumazet * this might cause drops if packet is looped back : 103638ab52e8SEric Dumazet * Other socket might not have SOCK_MEMALLOC. 103738ab52e8SEric Dumazet * Packets not looped back do not care about pfmemalloc. 103838ab52e8SEric Dumazet */ 103938ab52e8SEric Dumazet skb->pfmemalloc = 0; 104038ab52e8SEric Dumazet 1041aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 1042aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 104346d3ceabSEric Dumazet 104446d3ceabSEric Dumazet skb_orphan(skb); 104546d3ceabSEric Dumazet skb->sk = sk; 10461d2077acSEric Dumazet skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree; 1047b73c3d0eSTom Herbert skb_set_hash_from_sk(skb, sk); 104846d3ceabSEric Dumazet atomic_add(skb->truesize, &sk->sk_wmem_alloc); 10491da177e4SLinus Torvalds 1050c3a2e837SJulian Anastasov skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm); 1051c3a2e837SJulian Anastasov 10521da177e4SLinus Torvalds /* Build TCP header and checksum it. */ 1053ea1627c2SEric Dumazet th = (struct tcphdr *)skb->data; 1054c720c7e8SEric Dumazet th->source = inet->inet_sport; 1055c720c7e8SEric Dumazet th->dest = inet->inet_dport; 10561da177e4SLinus Torvalds th->seq = htonl(tcb->seq); 10571da177e4SLinus Torvalds th->ack_seq = htonl(tp->rcv_nxt); 1058df7a3b07SAl Viro *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | 10594de075e0SEric Dumazet tcb->tcp_flags); 1060dfb4b9dcSDavid S. Miller 10611da177e4SLinus Torvalds th->check = 0; 10621da177e4SLinus Torvalds th->urg_ptr = 0; 10631da177e4SLinus Torvalds 106433f5f57eSIlpo Järvinen /* The urg_mode check is necessary during a below snd_una win probe */ 10657691367dSHerbert Xu if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { 10667691367dSHerbert Xu if (before(tp->snd_up, tcb->seq + 0x10000)) { 10671da177e4SLinus Torvalds th->urg_ptr = htons(tp->snd_up - tcb->seq); 10681da177e4SLinus Torvalds th->urg = 1; 10697691367dSHerbert Xu } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { 10700eae88f3SEric Dumazet th->urg_ptr = htons(0xFFFF); 10717691367dSHerbert Xu th->urg = 1; 10727691367dSHerbert Xu } 10731da177e4SLinus Torvalds } 10741da177e4SLinus Torvalds 1075bd0388aeSWilliam Allen Simpson tcp_options_write((__be32 *)(th + 1), tp, &opts); 107651466a75SEric Dumazet skb_shinfo(skb)->gso_type = sk->sk_gso_type; 1077ea1627c2SEric Dumazet if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) { 1078ea1627c2SEric Dumazet th->window = htons(tcp_select_window(sk)); 1079ea1627c2SEric Dumazet tcp_ecn_send(sk, skb, th, tcp_header_size); 1080ea1627c2SEric Dumazet } else { 1081ea1627c2SEric Dumazet /* RFC1323: The window in SYN & SYN/ACK segments 1082ea1627c2SEric Dumazet * is never scaled. 1083ea1627c2SEric Dumazet */ 1084ea1627c2SEric Dumazet th->window = htons(min(tp->rcv_wnd, 65535U)); 1085ea1627c2SEric Dumazet } 1086cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 1087cfb6eeb4SYOSHIFUJI Hideaki /* Calculate the MD5 hash, as we have all we need now */ 1088cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 1089a465419bSEric Dumazet sk_nocaps_add(sk, NETIF_F_GSO_MASK); 1090bd0388aeSWilliam Allen Simpson tp->af_specific->calc_md5_hash(opts.hash_location, 109139f8e58eSEric Dumazet md5, sk, skb); 1092cfb6eeb4SYOSHIFUJI Hideaki } 1093cfb6eeb4SYOSHIFUJI Hideaki #endif 1094cfb6eeb4SYOSHIFUJI Hideaki 1095bb296246SHerbert Xu icsk->icsk_af_ops->send_check(sk, skb); 10961da177e4SLinus Torvalds 10974de075e0SEric Dumazet if (likely(tcb->tcp_flags & TCPHDR_ACK)) 1098fc6415bcSDavid S. Miller tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); 10991da177e4SLinus Torvalds 1100a44d6eacSMartin KaFai Lau if (skb->len != tcp_header_size) { 1101cf533ea5SEric Dumazet tcp_event_data_sent(tp, sk); 1102a44d6eacSMartin KaFai Lau tp->data_segs_out += tcp_skb_pcount(skb); 1103218af599SEric Dumazet tcp_internal_pacing(sk, skb); 1104a44d6eacSMartin KaFai Lau } 11051da177e4SLinus Torvalds 1106bd37a088SWei Yongjun if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) 1107aa2ea058STom Herbert TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, 1108aa2ea058STom Herbert tcp_skb_pcount(skb)); 11091da177e4SLinus Torvalds 11102efd055cSMarcelo Ricardo Leitner tp->segs_out += tcp_skb_pcount(skb); 1111f69ad292SEric Dumazet /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */ 1112cd7d8498SEric Dumazet skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb); 1113f69ad292SEric Dumazet skb_shinfo(skb)->gso_size = tcp_skb_mss(skb); 1114cd7d8498SEric Dumazet 11157faee5c0SEric Dumazet /* Our usage of tstamp should remain private */ 11162456e855SThomas Gleixner skb->tstamp = 0; 1117971f10ecSEric Dumazet 1118971f10ecSEric Dumazet /* Cleanup our debris for IP stacks */ 1119971f10ecSEric Dumazet memset(skb->cb, 0, max(sizeof(struct inet_skb_parm), 1120971f10ecSEric Dumazet sizeof(struct inet6_skb_parm))); 1121971f10ecSEric Dumazet 1122b0270e91SEric Dumazet err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl); 11237faee5c0SEric Dumazet 112483de47cdSHua Zhong if (likely(err <= 0)) 11251da177e4SLinus Torvalds return err; 11261da177e4SLinus Torvalds 11275ee2c941SChristoph Paasch tcp_enter_cwr(sk); 11281da177e4SLinus Torvalds 1129b9df3cb8SGerrit Renker return net_xmit_eval(err); 11301da177e4SLinus Torvalds } 11311da177e4SLinus Torvalds 113267edfef7SAndi Kleen /* This routine just queues the buffer for sending. 11331da177e4SLinus Torvalds * 11341da177e4SLinus Torvalds * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, 11351da177e4SLinus Torvalds * otherwise socket can stall. 11361da177e4SLinus Torvalds */ 11371da177e4SLinus Torvalds static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) 11381da177e4SLinus Torvalds { 11391da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 11401da177e4SLinus Torvalds 11411da177e4SLinus Torvalds /* Advance write_seq and place onto the write_queue. */ 11421da177e4SLinus Torvalds tp->write_seq = TCP_SKB_CB(skb)->end_seq; 1143f4a775d1SEric Dumazet __skb_header_release(skb); 1144fe067e8aSDavid S. Miller tcp_add_write_queue_tail(sk, skb); 11453ab224beSHideo Aoki sk->sk_wmem_queued += skb->truesize; 11463ab224beSHideo Aoki sk_mem_charge(sk, skb->truesize); 11471da177e4SLinus Torvalds } 11481da177e4SLinus Torvalds 114967edfef7SAndi Kleen /* Initialize TSO segments for a packet. */ 11505bbb432cSEric Dumazet static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now) 1151f6302d1dSDavid S. Miller { 11528f26fb1cSEric Dumazet if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) { 1153f6302d1dSDavid S. Miller /* Avoid the costly divide in the normal 1154f6302d1dSDavid S. Miller * non-TSO case. 1155f6302d1dSDavid S. Miller */ 1156cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, 1); 1157f69ad292SEric Dumazet TCP_SKB_CB(skb)->tcp_gso_size = 0; 1158f6302d1dSDavid S. Miller } else { 1159cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now)); 1160f69ad292SEric Dumazet TCP_SKB_CB(skb)->tcp_gso_size = mss_now; 11611da177e4SLinus Torvalds } 11621da177e4SLinus Torvalds } 11631da177e4SLinus Torvalds 116491fed7a1SIlpo Järvinen /* When a modification to fackets out becomes necessary, we need to check 116568f8353bSIlpo Järvinen * skb is counted to fackets_out or not. 116691fed7a1SIlpo Järvinen */ 1167cf533ea5SEric Dumazet static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb, 116891fed7a1SIlpo Järvinen int decr) 116991fed7a1SIlpo Järvinen { 1170a47e5a98SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1171a47e5a98SIlpo Järvinen 1172dc86967bSIlpo Järvinen if (!tp->sacked_out || tcp_is_reno(tp)) 117391fed7a1SIlpo Järvinen return; 117491fed7a1SIlpo Järvinen 11756859d494SIlpo Järvinen if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq)) 117691fed7a1SIlpo Järvinen tp->fackets_out -= decr; 117791fed7a1SIlpo Järvinen } 117891fed7a1SIlpo Järvinen 1179797108d1SIlpo Järvinen /* Pcount in the middle of the write queue got changed, we need to do various 1180797108d1SIlpo Järvinen * tweaks to fix counters 1181797108d1SIlpo Järvinen */ 1182cf533ea5SEric Dumazet static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) 1183797108d1SIlpo Järvinen { 1184797108d1SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1185797108d1SIlpo Järvinen 1186797108d1SIlpo Järvinen tp->packets_out -= decr; 1187797108d1SIlpo Järvinen 1188797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 1189797108d1SIlpo Järvinen tp->sacked_out -= decr; 1190797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 1191797108d1SIlpo Järvinen tp->retrans_out -= decr; 1192797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 1193797108d1SIlpo Järvinen tp->lost_out -= decr; 1194797108d1SIlpo Järvinen 1195797108d1SIlpo Järvinen /* Reno case is special. Sigh... */ 1196797108d1SIlpo Järvinen if (tcp_is_reno(tp) && decr > 0) 1197797108d1SIlpo Järvinen tp->sacked_out -= min_t(u32, tp->sacked_out, decr); 1198797108d1SIlpo Järvinen 1199797108d1SIlpo Järvinen tcp_adjust_fackets_out(sk, skb, decr); 1200797108d1SIlpo Järvinen 1201797108d1SIlpo Järvinen if (tp->lost_skb_hint && 1202797108d1SIlpo Järvinen before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && 120352cf3cc8SIlpo Järvinen (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))) 1204797108d1SIlpo Järvinen tp->lost_cnt_hint -= decr; 1205797108d1SIlpo Järvinen 1206797108d1SIlpo Järvinen tcp_verify_left_out(tp); 1207797108d1SIlpo Järvinen } 1208797108d1SIlpo Järvinen 12090a2cf20cSSoheil Hassas Yeganeh static bool tcp_has_tx_tstamp(const struct sk_buff *skb) 12100a2cf20cSSoheil Hassas Yeganeh { 12110a2cf20cSSoheil Hassas Yeganeh return TCP_SKB_CB(skb)->txstamp_ack || 12120a2cf20cSSoheil Hassas Yeganeh (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP); 12130a2cf20cSSoheil Hassas Yeganeh } 12140a2cf20cSSoheil Hassas Yeganeh 1215490cc7d0SWillem de Bruijn static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2) 1216490cc7d0SWillem de Bruijn { 1217490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo = skb_shinfo(skb); 1218490cc7d0SWillem de Bruijn 12190a2cf20cSSoheil Hassas Yeganeh if (unlikely(tcp_has_tx_tstamp(skb)) && 1220490cc7d0SWillem de Bruijn !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) { 1221490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo2 = skb_shinfo(skb2); 1222490cc7d0SWillem de Bruijn u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP; 1223490cc7d0SWillem de Bruijn 1224490cc7d0SWillem de Bruijn shinfo->tx_flags &= ~tsflags; 1225490cc7d0SWillem de Bruijn shinfo2->tx_flags |= tsflags; 1226490cc7d0SWillem de Bruijn swap(shinfo->tskey, shinfo2->tskey); 1227b51e13faSMartin KaFai Lau TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack; 1228b51e13faSMartin KaFai Lau TCP_SKB_CB(skb)->txstamp_ack = 0; 1229490cc7d0SWillem de Bruijn } 1230490cc7d0SWillem de Bruijn } 1231490cc7d0SWillem de Bruijn 1232a166140eSMartin KaFai Lau static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2) 1233a166140eSMartin KaFai Lau { 1234a166140eSMartin KaFai Lau TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor; 1235a166140eSMartin KaFai Lau TCP_SKB_CB(skb)->eor = 0; 1236a166140eSMartin KaFai Lau } 1237a166140eSMartin KaFai Lau 12381da177e4SLinus Torvalds /* Function to create two new TCP segments. Shrinks the given segment 12391da177e4SLinus Torvalds * to the specified size and appends a new segment with the rest of the 12401da177e4SLinus Torvalds * packet to the list. This won't be called frequently, I hope. 12411da177e4SLinus Torvalds * Remember, these are still headerless SKBs at this point. 12421da177e4SLinus Torvalds */ 1243056834d9SIlpo Järvinen int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, 12446cc55e09SOctavian Purdila unsigned int mss_now, gfp_t gfp) 12451da177e4SLinus Torvalds { 12461da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 12471da177e4SLinus Torvalds struct sk_buff *buff; 12486475be16SDavid S. Miller int nsize, old_factor; 1249b60b49eaSHerbert Xu int nlen; 12509ce01461SIlpo Järvinen u8 flags; 12511da177e4SLinus Torvalds 12522fceec13SIlpo Järvinen if (WARN_ON(len > skb->len)) 12532fceec13SIlpo Järvinen return -EINVAL; 12546a438bbeSStephen Hemminger 12551da177e4SLinus Torvalds nsize = skb_headlen(skb) - len; 12561da177e4SLinus Torvalds if (nsize < 0) 12571da177e4SLinus Torvalds nsize = 0; 12581da177e4SLinus Torvalds 12596cc55e09SOctavian Purdila if (skb_unclone(skb, gfp)) 12601da177e4SLinus Torvalds return -ENOMEM; 12611da177e4SLinus Torvalds 12621da177e4SLinus Torvalds /* Get a new skb... force flag on. */ 1263eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, nsize, gfp, true); 126451456b29SIan Morris if (!buff) 12651da177e4SLinus Torvalds return -ENOMEM; /* We'll just try again later. */ 1266ef5cb973SHerbert Xu 12673ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 12683ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1269b60b49eaSHerbert Xu nlen = skb->len - len - nsize; 1270b60b49eaSHerbert Xu buff->truesize += nlen; 1271b60b49eaSHerbert Xu skb->truesize -= nlen; 12721da177e4SLinus Torvalds 12731da177e4SLinus Torvalds /* Correct the sequence numbers. */ 12741da177e4SLinus Torvalds TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 12751da177e4SLinus Torvalds TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 12761da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 12771da177e4SLinus Torvalds 12781da177e4SLinus Torvalds /* PSH and FIN should only be set in the second packet. */ 12794de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 12804de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 12814de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1282e14c3cafSHerbert Xu TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; 1283a166140eSMartin KaFai Lau tcp_skb_fragment_eor(skb, buff); 12841da177e4SLinus Torvalds 128584fa7933SPatrick McHardy if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { 12861da177e4SLinus Torvalds /* Copy and checksum data tail into the new buffer. */ 1287056834d9SIlpo Järvinen buff->csum = csum_partial_copy_nocheck(skb->data + len, 1288056834d9SIlpo Järvinen skb_put(buff, nsize), 12891da177e4SLinus Torvalds nsize, 0); 12901da177e4SLinus Torvalds 12911da177e4SLinus Torvalds skb_trim(skb, len); 12921da177e4SLinus Torvalds 12931da177e4SLinus Torvalds skb->csum = csum_block_sub(skb->csum, buff->csum, len); 12941da177e4SLinus Torvalds } else { 129584fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 12961da177e4SLinus Torvalds skb_split(skb, buff, len); 12971da177e4SLinus Torvalds } 12981da177e4SLinus Torvalds 12991da177e4SLinus Torvalds buff->ip_summed = skb->ip_summed; 13001da177e4SLinus Torvalds 1301a61bbcf2SPatrick McHardy buff->tstamp = skb->tstamp; 1302490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 13031da177e4SLinus Torvalds 13046475be16SDavid S. Miller old_factor = tcp_skb_pcount(skb); 13056475be16SDavid S. Miller 13061da177e4SLinus Torvalds /* Fix up tso_factor for both original and new SKB. */ 13075bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 13085bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 13091da177e4SLinus Torvalds 1310b9f64820SYuchung Cheng /* Update delivered info for the new segment */ 1311b9f64820SYuchung Cheng TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx; 1312b9f64820SYuchung Cheng 13136475be16SDavid S. Miller /* If this packet has been sent out already, we must 13146475be16SDavid S. Miller * adjust the various packet counters. 13156475be16SDavid S. Miller */ 1316cf0b450cSHerbert Xu if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 13176475be16SDavid S. Miller int diff = old_factor - tcp_skb_pcount(skb) - 13186475be16SDavid S. Miller tcp_skb_pcount(buff); 13191da177e4SLinus Torvalds 1320797108d1SIlpo Järvinen if (diff) 1321797108d1SIlpo Järvinen tcp_adjust_pcount(sk, skb, diff); 13221da177e4SLinus Torvalds } 13231da177e4SLinus Torvalds 13241da177e4SLinus Torvalds /* Link BUFF into the send queue. */ 1325f4a775d1SEric Dumazet __skb_header_release(buff); 1326fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 13271da177e4SLinus Torvalds 13281da177e4SLinus Torvalds return 0; 13291da177e4SLinus Torvalds } 13301da177e4SLinus Torvalds 13311da177e4SLinus Torvalds /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c 13321da177e4SLinus Torvalds * eventually). The difference is that pulled data not copied, but 13331da177e4SLinus Torvalds * immediately discarded. 13341da177e4SLinus Torvalds */ 13357162fb24SEric Dumazet static int __pskb_trim_head(struct sk_buff *skb, int len) 13361da177e4SLinus Torvalds { 13377b7fc97aSEric Dumazet struct skb_shared_info *shinfo; 13381da177e4SLinus Torvalds int i, k, eat; 13391da177e4SLinus Torvalds 13404fa48bf3SEric Dumazet eat = min_t(int, len, skb_headlen(skb)); 13414fa48bf3SEric Dumazet if (eat) { 13424fa48bf3SEric Dumazet __skb_pull(skb, eat); 13434fa48bf3SEric Dumazet len -= eat; 13444fa48bf3SEric Dumazet if (!len) 13457162fb24SEric Dumazet return 0; 13464fa48bf3SEric Dumazet } 13471da177e4SLinus Torvalds eat = len; 13481da177e4SLinus Torvalds k = 0; 13497b7fc97aSEric Dumazet shinfo = skb_shinfo(skb); 13507b7fc97aSEric Dumazet for (i = 0; i < shinfo->nr_frags; i++) { 13517b7fc97aSEric Dumazet int size = skb_frag_size(&shinfo->frags[i]); 13529e903e08SEric Dumazet 13539e903e08SEric Dumazet if (size <= eat) { 1354aff65da0SIan Campbell skb_frag_unref(skb, i); 13559e903e08SEric Dumazet eat -= size; 13561da177e4SLinus Torvalds } else { 13577b7fc97aSEric Dumazet shinfo->frags[k] = shinfo->frags[i]; 13581da177e4SLinus Torvalds if (eat) { 13597b7fc97aSEric Dumazet shinfo->frags[k].page_offset += eat; 13607b7fc97aSEric Dumazet skb_frag_size_sub(&shinfo->frags[k], eat); 13611da177e4SLinus Torvalds eat = 0; 13621da177e4SLinus Torvalds } 13631da177e4SLinus Torvalds k++; 13641da177e4SLinus Torvalds } 13651da177e4SLinus Torvalds } 13667b7fc97aSEric Dumazet shinfo->nr_frags = k; 13671da177e4SLinus Torvalds 136827a884dcSArnaldo Carvalho de Melo skb_reset_tail_pointer(skb); 13691da177e4SLinus Torvalds skb->data_len -= len; 13701da177e4SLinus Torvalds skb->len = skb->data_len; 13717162fb24SEric Dumazet return len; 13721da177e4SLinus Torvalds } 13731da177e4SLinus Torvalds 137467edfef7SAndi Kleen /* Remove acked data from a packet in the transmit queue. */ 13751da177e4SLinus Torvalds int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 13761da177e4SLinus Torvalds { 13777162fb24SEric Dumazet u32 delta_truesize; 13787162fb24SEric Dumazet 137914bbd6a5SPravin B Shelar if (skb_unclone(skb, GFP_ATOMIC)) 13801da177e4SLinus Torvalds return -ENOMEM; 13811da177e4SLinus Torvalds 13827162fb24SEric Dumazet delta_truesize = __pskb_trim_head(skb, len); 13831da177e4SLinus Torvalds 13841da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq += len; 138584fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 13861da177e4SLinus Torvalds 13877162fb24SEric Dumazet if (delta_truesize) { 13887162fb24SEric Dumazet skb->truesize -= delta_truesize; 13897162fb24SEric Dumazet sk->sk_wmem_queued -= delta_truesize; 13907162fb24SEric Dumazet sk_mem_uncharge(sk, delta_truesize); 13911da177e4SLinus Torvalds sock_set_flag(sk, SOCK_QUEUE_SHRUNK); 13927162fb24SEric Dumazet } 13931da177e4SLinus Torvalds 13945b35e1e6SNeal Cardwell /* Any change of skb->len requires recalculation of tso factor. */ 13951da177e4SLinus Torvalds if (tcp_skb_pcount(skb) > 1) 13965bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb)); 13971da177e4SLinus Torvalds 13981da177e4SLinus Torvalds return 0; 13991da177e4SLinus Torvalds } 14001da177e4SLinus Torvalds 14011b63edd6SYuchung Cheng /* Calculate MSS not accounting any TCP options. */ 14021b63edd6SYuchung Cheng static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu) 14035d424d5aSJohn Heffner { 1404cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1405cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 14065d424d5aSJohn Heffner int mss_now; 14075d424d5aSJohn Heffner 14085d424d5aSJohn Heffner /* Calculate base mss without TCP options: 14095d424d5aSJohn Heffner It is MMS_S - sizeof(tcphdr) of rfc1122 14105d424d5aSJohn Heffner */ 14115d424d5aSJohn Heffner mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 14125d424d5aSJohn Heffner 141367469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 141467469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 141567469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 141667469601SEric Dumazet 141767469601SEric Dumazet if (dst && dst_allfrag(dst)) 141867469601SEric Dumazet mss_now -= icsk->icsk_af_ops->net_frag_header_len; 141967469601SEric Dumazet } 142067469601SEric Dumazet 14215d424d5aSJohn Heffner /* Clamp it (mss_clamp does not include tcp options) */ 14225d424d5aSJohn Heffner if (mss_now > tp->rx_opt.mss_clamp) 14235d424d5aSJohn Heffner mss_now = tp->rx_opt.mss_clamp; 14245d424d5aSJohn Heffner 14255d424d5aSJohn Heffner /* Now subtract optional transport overhead */ 14265d424d5aSJohn Heffner mss_now -= icsk->icsk_ext_hdr_len; 14275d424d5aSJohn Heffner 14285d424d5aSJohn Heffner /* Then reserve room for full set of TCP options and 8 bytes of data */ 14295d424d5aSJohn Heffner if (mss_now < 48) 14305d424d5aSJohn Heffner mss_now = 48; 14315d424d5aSJohn Heffner return mss_now; 14325d424d5aSJohn Heffner } 14335d424d5aSJohn Heffner 14341b63edd6SYuchung Cheng /* Calculate MSS. Not accounting for SACKs here. */ 14351b63edd6SYuchung Cheng int tcp_mtu_to_mss(struct sock *sk, int pmtu) 14361b63edd6SYuchung Cheng { 14371b63edd6SYuchung Cheng /* Subtract TCP options size, not including SACKs */ 14381b63edd6SYuchung Cheng return __tcp_mtu_to_mss(sk, pmtu) - 14391b63edd6SYuchung Cheng (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr)); 14401b63edd6SYuchung Cheng } 14411b63edd6SYuchung Cheng 14425d424d5aSJohn Heffner /* Inverse of above */ 144367469601SEric Dumazet int tcp_mss_to_mtu(struct sock *sk, int mss) 14445d424d5aSJohn Heffner { 1445cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1446cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 14475d424d5aSJohn Heffner int mtu; 14485d424d5aSJohn Heffner 14495d424d5aSJohn Heffner mtu = mss + 14505d424d5aSJohn Heffner tp->tcp_header_len + 14515d424d5aSJohn Heffner icsk->icsk_ext_hdr_len + 14525d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 14535d424d5aSJohn Heffner 145467469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 145567469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 145667469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 145767469601SEric Dumazet 145867469601SEric Dumazet if (dst && dst_allfrag(dst)) 145967469601SEric Dumazet mtu += icsk->icsk_af_ops->net_frag_header_len; 146067469601SEric Dumazet } 14615d424d5aSJohn Heffner return mtu; 14625d424d5aSJohn Heffner } 1463556c6b46SNeal Cardwell EXPORT_SYMBOL(tcp_mss_to_mtu); 14645d424d5aSJohn Heffner 146567edfef7SAndi Kleen /* MTU probing init per socket */ 14665d424d5aSJohn Heffner void tcp_mtup_init(struct sock *sk) 14675d424d5aSJohn Heffner { 14685d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 14695d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 1470b0f9ca53SFan Du struct net *net = sock_net(sk); 14715d424d5aSJohn Heffner 1472b0f9ca53SFan Du icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1; 14735d424d5aSJohn Heffner icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 14745d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 1475b0f9ca53SFan Du icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss); 14765d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 147705cbc0dbSFan Du if (icsk->icsk_mtup.enabled) 147805cbc0dbSFan Du icsk->icsk_mtup.probe_timestamp = tcp_time_stamp; 14795d424d5aSJohn Heffner } 14804bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_mtup_init); 14815d424d5aSJohn Heffner 14821da177e4SLinus Torvalds /* This function synchronize snd mss to current pmtu/exthdr set. 14831da177e4SLinus Torvalds 14841da177e4SLinus Torvalds tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 14851da177e4SLinus Torvalds for TCP options, but includes only bare TCP header. 14861da177e4SLinus Torvalds 14871da177e4SLinus Torvalds tp->rx_opt.mss_clamp is mss negotiated at connection setup. 1488caa20d9aSStephen Hemminger It is minimum of user_mss and mss received with SYN. 14891da177e4SLinus Torvalds It also does not include TCP options. 14901da177e4SLinus Torvalds 1491d83d8461SArnaldo Carvalho de Melo inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. 14921da177e4SLinus Torvalds 14931da177e4SLinus Torvalds tp->mss_cache is current effective sending mss, including 14941da177e4SLinus Torvalds all tcp options except for SACKs. It is evaluated, 14951da177e4SLinus Torvalds taking into account current pmtu, but never exceeds 14961da177e4SLinus Torvalds tp->rx_opt.mss_clamp. 14971da177e4SLinus Torvalds 14981da177e4SLinus Torvalds NOTE1. rfc1122 clearly states that advertised MSS 14991da177e4SLinus Torvalds DOES NOT include either tcp or ip options. 15001da177e4SLinus Torvalds 1501d83d8461SArnaldo Carvalho de Melo NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache 1502d83d8461SArnaldo Carvalho de Melo are READ ONLY outside this function. --ANK (980731) 15031da177e4SLinus Torvalds */ 15041da177e4SLinus Torvalds unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 15051da177e4SLinus Torvalds { 15061da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1507d83d8461SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 15085d424d5aSJohn Heffner int mss_now; 15091da177e4SLinus Torvalds 15105d424d5aSJohn Heffner if (icsk->icsk_mtup.search_high > pmtu) 15115d424d5aSJohn Heffner icsk->icsk_mtup.search_high = pmtu; 15121da177e4SLinus Torvalds 15135d424d5aSJohn Heffner mss_now = tcp_mtu_to_mss(sk, pmtu); 1514409d22b4SIlpo Järvinen mss_now = tcp_bound_to_half_wnd(tp, mss_now); 15151da177e4SLinus Torvalds 15161da177e4SLinus Torvalds /* And store cached results */ 1517d83d8461SArnaldo Carvalho de Melo icsk->icsk_pmtu_cookie = pmtu; 15185d424d5aSJohn Heffner if (icsk->icsk_mtup.enabled) 15195d424d5aSJohn Heffner mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); 1520c1b4a7e6SDavid S. Miller tp->mss_cache = mss_now; 15211da177e4SLinus Torvalds 15221da177e4SLinus Torvalds return mss_now; 15231da177e4SLinus Torvalds } 15244bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_sync_mss); 15251da177e4SLinus Torvalds 15261da177e4SLinus Torvalds /* Compute the current effective MSS, taking SACKs and IP options, 15271da177e4SLinus Torvalds * and even PMTU discovery events into account. 15281da177e4SLinus Torvalds */ 15290c54b85fSIlpo Järvinen unsigned int tcp_current_mss(struct sock *sk) 15301da177e4SLinus Torvalds { 1531cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1532cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1533c1b4a7e6SDavid S. Miller u32 mss_now; 153495c96174SEric Dumazet unsigned int header_len; 153533ad798cSAdam Langley struct tcp_out_options opts; 153633ad798cSAdam Langley struct tcp_md5sig_key *md5; 15371da177e4SLinus Torvalds 1538c1b4a7e6SDavid S. Miller mss_now = tp->mss_cache; 1539c1b4a7e6SDavid S. Miller 15401da177e4SLinus Torvalds if (dst) { 15411da177e4SLinus Torvalds u32 mtu = dst_mtu(dst); 1542d83d8461SArnaldo Carvalho de Melo if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 15431da177e4SLinus Torvalds mss_now = tcp_sync_mss(sk, mtu); 15441da177e4SLinus Torvalds } 15451da177e4SLinus Torvalds 154633ad798cSAdam Langley header_len = tcp_established_options(sk, NULL, &opts, &md5) + 154733ad798cSAdam Langley sizeof(struct tcphdr); 154833ad798cSAdam Langley /* The mss_cache is sized based on tp->tcp_header_len, which assumes 154933ad798cSAdam Langley * some common options. If this is an odd packet (because we have SACK 155033ad798cSAdam Langley * blocks etc) then our calculated header_len will be different, and 155133ad798cSAdam Langley * we have to adjust mss_now correspondingly */ 155233ad798cSAdam Langley if (header_len != tp->tcp_header_len) { 155333ad798cSAdam Langley int delta = (int) header_len - tp->tcp_header_len; 155433ad798cSAdam Langley mss_now -= delta; 155533ad798cSAdam Langley } 1556cfb6eeb4SYOSHIFUJI Hideaki 15571da177e4SLinus Torvalds return mss_now; 15581da177e4SLinus Torvalds } 15591da177e4SLinus Torvalds 156086fd14adSWeiping Pan /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. 156186fd14adSWeiping Pan * As additional protections, we do not touch cwnd in retransmission phases, 156286fd14adSWeiping Pan * and if application hit its sndbuf limit recently. 156386fd14adSWeiping Pan */ 156486fd14adSWeiping Pan static void tcp_cwnd_application_limited(struct sock *sk) 1565a762a980SDavid S. Miller { 15669e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1567a762a980SDavid S. Miller 156886fd14adSWeiping Pan if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && 156986fd14adSWeiping Pan sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 157086fd14adSWeiping Pan /* Limited by application or receiver window. */ 157186fd14adSWeiping Pan u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); 157286fd14adSWeiping Pan u32 win_used = max(tp->snd_cwnd_used, init_win); 157386fd14adSWeiping Pan if (win_used < tp->snd_cwnd) { 157486fd14adSWeiping Pan tp->snd_ssthresh = tcp_current_ssthresh(sk); 157586fd14adSWeiping Pan tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1; 157686fd14adSWeiping Pan } 157786fd14adSWeiping Pan tp->snd_cwnd_used = 0; 157886fd14adSWeiping Pan } 1579c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 158086fd14adSWeiping Pan } 158186fd14adSWeiping Pan 1582ca8a2263SNeal Cardwell static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited) 1583a762a980SDavid S. Miller { 15841b1fc3fdSWei Wang const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 1585a762a980SDavid S. Miller struct tcp_sock *tp = tcp_sk(sk); 1586a762a980SDavid S. Miller 1587ca8a2263SNeal Cardwell /* Track the maximum number of outstanding packets in each 1588ca8a2263SNeal Cardwell * window, and remember whether we were cwnd-limited then. 1589ca8a2263SNeal Cardwell */ 1590ca8a2263SNeal Cardwell if (!before(tp->snd_una, tp->max_packets_seq) || 1591ca8a2263SNeal Cardwell tp->packets_out > tp->max_packets_out) { 1592ca8a2263SNeal Cardwell tp->max_packets_out = tp->packets_out; 1593ca8a2263SNeal Cardwell tp->max_packets_seq = tp->snd_nxt; 1594ca8a2263SNeal Cardwell tp->is_cwnd_limited = is_cwnd_limited; 1595ca8a2263SNeal Cardwell } 1596e114a710SEric Dumazet 159724901551SEric Dumazet if (tcp_is_cwnd_limited(sk)) { 1598a762a980SDavid S. Miller /* Network is feed fully. */ 1599a762a980SDavid S. Miller tp->snd_cwnd_used = 0; 1600c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 1601a762a980SDavid S. Miller } else { 1602a762a980SDavid S. Miller /* Network starves. */ 1603a762a980SDavid S. Miller if (tp->packets_out > tp->snd_cwnd_used) 1604a762a980SDavid S. Miller tp->snd_cwnd_used = tp->packets_out; 1605a762a980SDavid S. Miller 160615d33c07SDavid S. Miller if (sysctl_tcp_slow_start_after_idle && 1607c2203cf7SEric Dumazet (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto && 16081b1fc3fdSWei Wang !ca_ops->cong_control) 1609a762a980SDavid S. Miller tcp_cwnd_application_limited(sk); 1610b0f71bd3SFrancis Yan 1611b0f71bd3SFrancis Yan /* The following conditions together indicate the starvation 1612b0f71bd3SFrancis Yan * is caused by insufficient sender buffer: 1613b0f71bd3SFrancis Yan * 1) just sent some data (see tcp_write_xmit) 1614b0f71bd3SFrancis Yan * 2) not cwnd limited (this else condition) 1615b0f71bd3SFrancis Yan * 3) no more data to send (null tcp_send_head ) 1616b0f71bd3SFrancis Yan * 4) application is hitting buffer limit (SOCK_NOSPACE) 1617b0f71bd3SFrancis Yan */ 1618b0f71bd3SFrancis Yan if (!tcp_send_head(sk) && sk->sk_socket && 1619b0f71bd3SFrancis Yan test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) && 1620b0f71bd3SFrancis Yan (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) 1621b0f71bd3SFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED); 1622a762a980SDavid S. Miller } 1623a762a980SDavid S. Miller } 1624a762a980SDavid S. Miller 1625d4589926SEric Dumazet /* Minshall's variant of the Nagle send check. */ 1626d4589926SEric Dumazet static bool tcp_minshall_check(const struct tcp_sock *tp) 1627d4589926SEric Dumazet { 1628d4589926SEric Dumazet return after(tp->snd_sml, tp->snd_una) && 1629d4589926SEric Dumazet !after(tp->snd_sml, tp->snd_nxt); 1630d4589926SEric Dumazet } 1631d4589926SEric Dumazet 1632d4589926SEric Dumazet /* Update snd_sml if this skb is under mss 1633d4589926SEric Dumazet * Note that a TSO packet might end with a sub-mss segment 1634d4589926SEric Dumazet * The test is really : 1635d4589926SEric Dumazet * if ((skb->len % mss) != 0) 1636d4589926SEric Dumazet * tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1637d4589926SEric Dumazet * But we can avoid doing the divide again given we already have 1638d4589926SEric Dumazet * skb_pcount = skb->len / mss_now 16390e3a4803SIlpo Järvinen */ 1640d4589926SEric Dumazet static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now, 1641d4589926SEric Dumazet const struct sk_buff *skb) 1642d4589926SEric Dumazet { 1643d4589926SEric Dumazet if (skb->len < tcp_skb_pcount(skb) * mss_now) 1644d4589926SEric Dumazet tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1645d4589926SEric Dumazet } 1646d4589926SEric Dumazet 1647d4589926SEric Dumazet /* Return false, if packet can be sent now without violation Nagle's rules: 1648d4589926SEric Dumazet * 1. It is full sized. (provided by caller in %partial bool) 1649d4589926SEric Dumazet * 2. Or it contains FIN. (already checked by caller) 1650d4589926SEric Dumazet * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. 1651d4589926SEric Dumazet * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 1652d4589926SEric Dumazet * With Minshall's modification: all sent small packets are ACKed. 1653d4589926SEric Dumazet */ 1654d4589926SEric Dumazet static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp, 1655cc93fc51SPeter Pan(潘卫平) int nonagle) 1656d4589926SEric Dumazet { 1657d4589926SEric Dumazet return partial && 1658d4589926SEric Dumazet ((nonagle & TCP_NAGLE_CORK) || 1659d4589926SEric Dumazet (!nonagle && tp->packets_out && tcp_minshall_check(tp))); 1660d4589926SEric Dumazet } 1661605ad7f1SEric Dumazet 1662605ad7f1SEric Dumazet /* Return how many segs we'd like on a TSO packet, 1663605ad7f1SEric Dumazet * to send one TSO packet per ms 1664605ad7f1SEric Dumazet */ 16651b3878caSNeal Cardwell u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now, 16661b3878caSNeal Cardwell int min_tso_segs) 1667605ad7f1SEric Dumazet { 1668605ad7f1SEric Dumazet u32 bytes, segs; 1669605ad7f1SEric Dumazet 1670605ad7f1SEric Dumazet bytes = min(sk->sk_pacing_rate >> 10, 1671605ad7f1SEric Dumazet sk->sk_gso_max_size - 1 - MAX_TCP_HEADER); 1672605ad7f1SEric Dumazet 1673605ad7f1SEric Dumazet /* Goal is to send at least one packet per ms, 1674605ad7f1SEric Dumazet * not one big TSO packet every 100 ms. 1675605ad7f1SEric Dumazet * This preserves ACK clocking and is consistent 1676605ad7f1SEric Dumazet * with tcp_tso_should_defer() heuristic. 1677605ad7f1SEric Dumazet */ 16781b3878caSNeal Cardwell segs = max_t(u32, bytes / mss_now, min_tso_segs); 1679605ad7f1SEric Dumazet 1680605ad7f1SEric Dumazet return min_t(u32, segs, sk->sk_gso_max_segs); 1681605ad7f1SEric Dumazet } 16821b3878caSNeal Cardwell EXPORT_SYMBOL(tcp_tso_autosize); 1683605ad7f1SEric Dumazet 1684ed6e7268SNeal Cardwell /* Return the number of segments we want in the skb we are transmitting. 1685ed6e7268SNeal Cardwell * See if congestion control module wants to decide; otherwise, autosize. 1686ed6e7268SNeal Cardwell */ 1687ed6e7268SNeal Cardwell static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now) 1688ed6e7268SNeal Cardwell { 1689ed6e7268SNeal Cardwell const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 1690ed6e7268SNeal Cardwell u32 tso_segs = ca_ops->tso_segs_goal ? ca_ops->tso_segs_goal(sk) : 0; 1691ed6e7268SNeal Cardwell 16921b3878caSNeal Cardwell return tso_segs ? : 16931b3878caSNeal Cardwell tcp_tso_autosize(sk, mss_now, sysctl_tcp_min_tso_segs); 1694ed6e7268SNeal Cardwell } 1695ed6e7268SNeal Cardwell 1696d4589926SEric Dumazet /* Returns the portion of skb which can be sent right away */ 1697d4589926SEric Dumazet static unsigned int tcp_mss_split_point(const struct sock *sk, 1698d4589926SEric Dumazet const struct sk_buff *skb, 1699d4589926SEric Dumazet unsigned int mss_now, 1700d4589926SEric Dumazet unsigned int max_segs, 1701d4589926SEric Dumazet int nonagle) 1702c1b4a7e6SDavid S. Miller { 1703cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1704d4589926SEric Dumazet u32 partial, needed, window, max_len; 1705c1b4a7e6SDavid S. Miller 170690840defSIlpo Järvinen window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 17071485348dSBen Hutchings max_len = mss_now * max_segs; 17080e3a4803SIlpo Järvinen 17091485348dSBen Hutchings if (likely(max_len <= window && skb != tcp_write_queue_tail(sk))) 17101485348dSBen Hutchings return max_len; 17110e3a4803SIlpo Järvinen 17125ea3a748SIlpo Järvinen needed = min(skb->len, window); 17135ea3a748SIlpo Järvinen 17141485348dSBen Hutchings if (max_len <= needed) 17151485348dSBen Hutchings return max_len; 17160e3a4803SIlpo Järvinen 1717d4589926SEric Dumazet partial = needed % mss_now; 1718d4589926SEric Dumazet /* If last segment is not a full MSS, check if Nagle rules allow us 1719d4589926SEric Dumazet * to include this last segment in this skb. 1720d4589926SEric Dumazet * Otherwise, we'll split the skb at last MSS boundary 1721d4589926SEric Dumazet */ 1722cc93fc51SPeter Pan(潘卫平) if (tcp_nagle_check(partial != 0, tp, nonagle)) 1723d4589926SEric Dumazet return needed - partial; 1724d4589926SEric Dumazet 1725d4589926SEric Dumazet return needed; 1726c1b4a7e6SDavid S. Miller } 1727c1b4a7e6SDavid S. Miller 1728c1b4a7e6SDavid S. Miller /* Can at least one segment of SKB be sent right now, according to the 1729c1b4a7e6SDavid S. Miller * congestion window rules? If so, return how many segments are allowed. 1730c1b4a7e6SDavid S. Miller */ 1731cf533ea5SEric Dumazet static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, 1732cf533ea5SEric Dumazet const struct sk_buff *skb) 1733c1b4a7e6SDavid S. Miller { 1734d649a7a8SEric Dumazet u32 in_flight, cwnd, halfcwnd; 1735c1b4a7e6SDavid S. Miller 1736c1b4a7e6SDavid S. Miller /* Don't be strict about the congestion window for the final FIN. */ 17374de075e0SEric Dumazet if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && 17384de075e0SEric Dumazet tcp_skb_pcount(skb) == 1) 1739c1b4a7e6SDavid S. Miller return 1; 1740c1b4a7e6SDavid S. Miller 1741c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1742c1b4a7e6SDavid S. Miller cwnd = tp->snd_cwnd; 1743d649a7a8SEric Dumazet if (in_flight >= cwnd) 1744c1b4a7e6SDavid S. Miller return 0; 1745d649a7a8SEric Dumazet 1746d649a7a8SEric Dumazet /* For better scheduling, ensure we have at least 1747d649a7a8SEric Dumazet * 2 GSO packets in flight. 1748d649a7a8SEric Dumazet */ 1749d649a7a8SEric Dumazet halfcwnd = max(cwnd >> 1, 1U); 1750d649a7a8SEric Dumazet return min(halfcwnd, cwnd - in_flight); 1751c1b4a7e6SDavid S. Miller } 1752c1b4a7e6SDavid S. Miller 1753b595076aSUwe Kleine-König /* Initialize TSO state of a skb. 175467edfef7SAndi Kleen * This must be invoked the first time we consider transmitting 1755c1b4a7e6SDavid S. Miller * SKB onto the wire. 1756c1b4a7e6SDavid S. Miller */ 17575bbb432cSEric Dumazet static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now) 1758c1b4a7e6SDavid S. Miller { 1759c1b4a7e6SDavid S. Miller int tso_segs = tcp_skb_pcount(skb); 1760c1b4a7e6SDavid S. Miller 1761f8269a49SIlpo Järvinen if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { 17625bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 1763c1b4a7e6SDavid S. Miller tso_segs = tcp_skb_pcount(skb); 1764c1b4a7e6SDavid S. Miller } 1765c1b4a7e6SDavid S. Miller return tso_segs; 1766c1b4a7e6SDavid S. Miller } 1767c1b4a7e6SDavid S. Miller 1768c1b4a7e6SDavid S. Miller 1769a2a385d6SEric Dumazet /* Return true if the Nagle test allows this packet to be 1770c1b4a7e6SDavid S. Miller * sent now. 1771c1b4a7e6SDavid S. Miller */ 1772a2a385d6SEric Dumazet static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, 1773c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1774c1b4a7e6SDavid S. Miller { 1775c1b4a7e6SDavid S. Miller /* Nagle rule does not apply to frames, which sit in the middle of the 1776c1b4a7e6SDavid S. Miller * write_queue (they have no chances to get new data). 1777c1b4a7e6SDavid S. Miller * 1778c1b4a7e6SDavid S. Miller * This is implemented in the callers, where they modify the 'nonagle' 1779c1b4a7e6SDavid S. Miller * argument based upon the location of SKB in the send queue. 1780c1b4a7e6SDavid S. Miller */ 1781c1b4a7e6SDavid S. Miller if (nonagle & TCP_NAGLE_PUSH) 1782a2a385d6SEric Dumazet return true; 1783c1b4a7e6SDavid S. Miller 17849b44190dSYuchung Cheng /* Don't use the nagle rule for urgent data (or for the final FIN). */ 17859b44190dSYuchung Cheng if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) 1786a2a385d6SEric Dumazet return true; 1787c1b4a7e6SDavid S. Miller 1788cc93fc51SPeter Pan(潘卫平) if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle)) 1789a2a385d6SEric Dumazet return true; 1790c1b4a7e6SDavid S. Miller 1791a2a385d6SEric Dumazet return false; 1792c1b4a7e6SDavid S. Miller } 1793c1b4a7e6SDavid S. Miller 1794c1b4a7e6SDavid S. Miller /* Does at least the first segment of SKB fit into the send window? */ 1795a2a385d6SEric Dumazet static bool tcp_snd_wnd_test(const struct tcp_sock *tp, 1796a2a385d6SEric Dumazet const struct sk_buff *skb, 1797056834d9SIlpo Järvinen unsigned int cur_mss) 1798c1b4a7e6SDavid S. Miller { 1799c1b4a7e6SDavid S. Miller u32 end_seq = TCP_SKB_CB(skb)->end_seq; 1800c1b4a7e6SDavid S. Miller 1801c1b4a7e6SDavid S. Miller if (skb->len > cur_mss) 1802c1b4a7e6SDavid S. Miller end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 1803c1b4a7e6SDavid S. Miller 180490840defSIlpo Järvinen return !after(end_seq, tcp_wnd_end(tp)); 1805c1b4a7e6SDavid S. Miller } 1806c1b4a7e6SDavid S. Miller 1807fe067e8aSDavid S. Miller /* This checks if the data bearing packet SKB (usually tcp_send_head(sk)) 1808c1b4a7e6SDavid S. Miller * should be put on the wire right now. If so, it returns the number of 1809c1b4a7e6SDavid S. Miller * packets allowed by the congestion window. 1810c1b4a7e6SDavid S. Miller */ 1811cf533ea5SEric Dumazet static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb, 1812c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1813c1b4a7e6SDavid S. Miller { 1814cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1815c1b4a7e6SDavid S. Miller unsigned int cwnd_quota; 1816c1b4a7e6SDavid S. Miller 18175bbb432cSEric Dumazet tcp_init_tso_segs(skb, cur_mss); 1818c1b4a7e6SDavid S. Miller 1819c1b4a7e6SDavid S. Miller if (!tcp_nagle_test(tp, skb, cur_mss, nonagle)) 1820c1b4a7e6SDavid S. Miller return 0; 1821c1b4a7e6SDavid S. Miller 1822c1b4a7e6SDavid S. Miller cwnd_quota = tcp_cwnd_test(tp, skb); 1823056834d9SIlpo Järvinen if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss)) 1824c1b4a7e6SDavid S. Miller cwnd_quota = 0; 1825c1b4a7e6SDavid S. Miller 1826c1b4a7e6SDavid S. Miller return cwnd_quota; 1827c1b4a7e6SDavid S. Miller } 1828c1b4a7e6SDavid S. Miller 182967edfef7SAndi Kleen /* Test if sending is allowed right now. */ 1830a2a385d6SEric Dumazet bool tcp_may_send_now(struct sock *sk) 1831c1b4a7e6SDavid S. Miller { 1832cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1833fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 1834c1b4a7e6SDavid S. Miller 1835a02cec21SEric Dumazet return skb && 18360c54b85fSIlpo Järvinen tcp_snd_test(sk, skb, tcp_current_mss(sk), 1837c1b4a7e6SDavid S. Miller (tcp_skb_is_last(sk, skb) ? 1838a02cec21SEric Dumazet tp->nonagle : TCP_NAGLE_PUSH)); 1839c1b4a7e6SDavid S. Miller } 1840c1b4a7e6SDavid S. Miller 1841c1b4a7e6SDavid S. Miller /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 1842c1b4a7e6SDavid S. Miller * which is put after SKB on the list. It is very much like 1843c1b4a7e6SDavid S. Miller * tcp_fragment() except that it may make several kinds of assumptions 1844c1b4a7e6SDavid S. Miller * in order to speed up the splitting operation. In particular, we 1845c1b4a7e6SDavid S. Miller * know that all the data is in scatter-gather pages, and that the 1846c1b4a7e6SDavid S. Miller * packet has never been sent out before (and thus is not cloned). 1847c1b4a7e6SDavid S. Miller */ 1848056834d9SIlpo Järvinen static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, 1849c4ead4c5SEric Dumazet unsigned int mss_now, gfp_t gfp) 1850c1b4a7e6SDavid S. Miller { 1851c1b4a7e6SDavid S. Miller struct sk_buff *buff; 1852c1b4a7e6SDavid S. Miller int nlen = skb->len - len; 18539ce01461SIlpo Järvinen u8 flags; 1854c1b4a7e6SDavid S. Miller 1855c1b4a7e6SDavid S. Miller /* All of a TSO frame must be composed of paged data. */ 1856c8ac3774SHerbert Xu if (skb->len != skb->data_len) 18576cc55e09SOctavian Purdila return tcp_fragment(sk, skb, len, mss_now, gfp); 1858c1b4a7e6SDavid S. Miller 1859eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, gfp, true); 186051456b29SIan Morris if (unlikely(!buff)) 1861c1b4a7e6SDavid S. Miller return -ENOMEM; 1862c1b4a7e6SDavid S. Miller 18633ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 18643ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1865b60b49eaSHerbert Xu buff->truesize += nlen; 1866c1b4a7e6SDavid S. Miller skb->truesize -= nlen; 1867c1b4a7e6SDavid S. Miller 1868c1b4a7e6SDavid S. Miller /* Correct the sequence numbers. */ 1869c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 1870c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 1871c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 1872c1b4a7e6SDavid S. Miller 1873c1b4a7e6SDavid S. Miller /* PSH and FIN should only be set in the second packet. */ 18744de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 18754de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 18764de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1877c1b4a7e6SDavid S. Miller 1878c1b4a7e6SDavid S. Miller /* This packet was never sent out yet, so no SACK bits. */ 1879c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->sacked = 0; 1880c1b4a7e6SDavid S. Miller 1881a166140eSMartin KaFai Lau tcp_skb_fragment_eor(skb, buff); 1882a166140eSMartin KaFai Lau 188384fa7933SPatrick McHardy buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; 1884c1b4a7e6SDavid S. Miller skb_split(skb, buff, len); 1885490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 1886c1b4a7e6SDavid S. Miller 1887c1b4a7e6SDavid S. Miller /* Fix up tso_factor for both original and new SKB. */ 18885bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 18895bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 1890c1b4a7e6SDavid S. Miller 1891c1b4a7e6SDavid S. Miller /* Link BUFF into the send queue. */ 1892f4a775d1SEric Dumazet __skb_header_release(buff); 1893fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 1894c1b4a7e6SDavid S. Miller 1895c1b4a7e6SDavid S. Miller return 0; 1896c1b4a7e6SDavid S. Miller } 1897c1b4a7e6SDavid S. Miller 1898c1b4a7e6SDavid S. Miller /* Try to defer sending, if possible, in order to minimize the amount 1899c1b4a7e6SDavid S. Miller * of TSO splitting we do. View it as a kind of TSO Nagle test. 1900c1b4a7e6SDavid S. Miller * 1901c1b4a7e6SDavid S. Miller * This algorithm is from John Heffner. 1902c1b4a7e6SDavid S. Miller */ 1903ca8a2263SNeal Cardwell static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb, 1904605ad7f1SEric Dumazet bool *is_cwnd_limited, u32 max_segs) 1905c1b4a7e6SDavid S. Miller { 19066687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 190750c8339eSEric Dumazet u32 age, send_win, cong_win, limit, in_flight; 190850c8339eSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 190950c8339eSEric Dumazet struct sk_buff *head; 1910ad9f4f50SEric Dumazet int win_divisor; 1911c1b4a7e6SDavid S. Miller 19124de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 1913ae8064acSJohn Heffner goto send_now; 1914c1b4a7e6SDavid S. Miller 191599d7662aSEric Dumazet if (icsk->icsk_ca_state >= TCP_CA_Recovery) 1916ae8064acSJohn Heffner goto send_now; 1917ae8064acSJohn Heffner 19185f852eb5SEric Dumazet /* Avoid bursty behavior by allowing defer 19195f852eb5SEric Dumazet * only if the last write was recent. 19205f852eb5SEric Dumazet */ 1921d635fbe2SEric Dumazet if ((s32)(tcp_jiffies32 - tp->lsndtime) > 0) 1922ae8064acSJohn Heffner goto send_now; 1923908a75c1SDavid S. Miller 1924c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1925c1b4a7e6SDavid S. Miller 1926056834d9SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight)); 1927c1b4a7e6SDavid S. Miller 192890840defSIlpo Järvinen send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1929c1b4a7e6SDavid S. Miller 1930c1b4a7e6SDavid S. Miller /* From in_flight test above, we know that cwnd > in_flight. */ 1931c1b4a7e6SDavid S. Miller cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; 1932c1b4a7e6SDavid S. Miller 1933c1b4a7e6SDavid S. Miller limit = min(send_win, cong_win); 1934c1b4a7e6SDavid S. Miller 1935ba244fe9SDavid S. Miller /* If a full-sized TSO skb can be sent, do it. */ 1936605ad7f1SEric Dumazet if (limit >= max_segs * tp->mss_cache) 1937ae8064acSJohn Heffner goto send_now; 1938ba244fe9SDavid S. Miller 193962ad2761SIlpo Järvinen /* Middle in queue won't get any more data, full sendable already? */ 194062ad2761SIlpo Järvinen if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) 194162ad2761SIlpo Järvinen goto send_now; 194262ad2761SIlpo Järvinen 1943ad9f4f50SEric Dumazet win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor); 1944ad9f4f50SEric Dumazet if (win_divisor) { 1945c1b4a7e6SDavid S. Miller u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); 1946c1b4a7e6SDavid S. Miller 1947c1b4a7e6SDavid S. Miller /* If at least some fraction of a window is available, 1948c1b4a7e6SDavid S. Miller * just use it. 1949c1b4a7e6SDavid S. Miller */ 1950ad9f4f50SEric Dumazet chunk /= win_divisor; 1951c1b4a7e6SDavid S. Miller if (limit >= chunk) 1952ae8064acSJohn Heffner goto send_now; 1953c1b4a7e6SDavid S. Miller } else { 1954c1b4a7e6SDavid S. Miller /* Different approach, try not to defer past a single 1955c1b4a7e6SDavid S. Miller * ACK. Receiver should ACK every other full sized 1956c1b4a7e6SDavid S. Miller * frame, so if we have space for more than 3 frames 1957c1b4a7e6SDavid S. Miller * then send now. 1958c1b4a7e6SDavid S. Miller */ 19596b5a5c0dSNeal Cardwell if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) 1960ae8064acSJohn Heffner goto send_now; 1961c1b4a7e6SDavid S. Miller } 1962c1b4a7e6SDavid S. Miller 196350c8339eSEric Dumazet head = tcp_write_queue_head(sk); 1964385e2070SEric Dumazet 1965385e2070SEric Dumazet age = skb_mstamp_us_delta(&tp->tcp_mstamp, &head->skb_mstamp); 196650c8339eSEric Dumazet /* If next ACK is likely to come too late (half srtt), do not defer */ 196750c8339eSEric Dumazet if (age < (tp->srtt_us >> 4)) 196850c8339eSEric Dumazet goto send_now; 196950c8339eSEric Dumazet 19705f852eb5SEric Dumazet /* Ok, it looks like it is advisable to defer. */ 1971ae8064acSJohn Heffner 1972d2e1339fSBendik Rønning Opstad if (cong_win < send_win && cong_win <= skb->len) 1973ca8a2263SNeal Cardwell *is_cwnd_limited = true; 1974ca8a2263SNeal Cardwell 1975a2a385d6SEric Dumazet return true; 1976ae8064acSJohn Heffner 1977ae8064acSJohn Heffner send_now: 1978a2a385d6SEric Dumazet return false; 1979c1b4a7e6SDavid S. Miller } 1980c1b4a7e6SDavid S. Miller 198105cbc0dbSFan Du static inline void tcp_mtu_check_reprobe(struct sock *sk) 198205cbc0dbSFan Du { 198305cbc0dbSFan Du struct inet_connection_sock *icsk = inet_csk(sk); 198405cbc0dbSFan Du struct tcp_sock *tp = tcp_sk(sk); 198505cbc0dbSFan Du struct net *net = sock_net(sk); 198605cbc0dbSFan Du u32 interval; 198705cbc0dbSFan Du s32 delta; 198805cbc0dbSFan Du 198905cbc0dbSFan Du interval = net->ipv4.sysctl_tcp_probe_interval; 199005cbc0dbSFan Du delta = tcp_time_stamp - icsk->icsk_mtup.probe_timestamp; 199105cbc0dbSFan Du if (unlikely(delta >= interval * HZ)) { 199205cbc0dbSFan Du int mss = tcp_current_mss(sk); 199305cbc0dbSFan Du 199405cbc0dbSFan Du /* Update current search range */ 199505cbc0dbSFan Du icsk->icsk_mtup.probe_size = 0; 199605cbc0dbSFan Du icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + 199705cbc0dbSFan Du sizeof(struct tcphdr) + 199805cbc0dbSFan Du icsk->icsk_af_ops->net_header_len; 199905cbc0dbSFan Du icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss); 200005cbc0dbSFan Du 200105cbc0dbSFan Du /* Update probe time stamp */ 200205cbc0dbSFan Du icsk->icsk_mtup.probe_timestamp = tcp_time_stamp; 200305cbc0dbSFan Du } 200405cbc0dbSFan Du } 200505cbc0dbSFan Du 20065d424d5aSJohn Heffner /* Create a new MTU probe if we are ready. 200767edfef7SAndi Kleen * MTU probe is regularly attempting to increase the path MTU by 200867edfef7SAndi Kleen * deliberately sending larger packets. This discovers routing 200967edfef7SAndi Kleen * changes resulting in larger path MTUs. 201067edfef7SAndi Kleen * 20115d424d5aSJohn Heffner * Returns 0 if we should wait to probe (no cwnd available), 20125d424d5aSJohn Heffner * 1 if a probe was sent, 2013056834d9SIlpo Järvinen * -1 otherwise 2014056834d9SIlpo Järvinen */ 20155d424d5aSJohn Heffner static int tcp_mtu_probe(struct sock *sk) 20165d424d5aSJohn Heffner { 20175d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 201812a59abcSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 20195d424d5aSJohn Heffner struct sk_buff *skb, *nskb, *next; 20206b58e0a5SFan Du struct net *net = sock_net(sk); 20215d424d5aSJohn Heffner int probe_size; 202291cc17c0SIlpo Järvinen int size_needed; 202312a59abcSEric Dumazet int copy, len; 20245d424d5aSJohn Heffner int mss_now; 20256b58e0a5SFan Du int interval; 20265d424d5aSJohn Heffner 20275d424d5aSJohn Heffner /* Not currently probing/verifying, 20285d424d5aSJohn Heffner * not in recovery, 20295d424d5aSJohn Heffner * have enough cwnd, and 203012a59abcSEric Dumazet * not SACKing (the variable headers throw things off) 203112a59abcSEric Dumazet */ 203212a59abcSEric Dumazet if (likely(!icsk->icsk_mtup.enabled || 20335d424d5aSJohn Heffner icsk->icsk_mtup.probe_size || 20345d424d5aSJohn Heffner inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 20355d424d5aSJohn Heffner tp->snd_cwnd < 11 || 203612a59abcSEric Dumazet tp->rx_opt.num_sacks || tp->rx_opt.dsack)) 20375d424d5aSJohn Heffner return -1; 20385d424d5aSJohn Heffner 20396b58e0a5SFan Du /* Use binary search for probe_size between tcp_mss_base, 20406b58e0a5SFan Du * and current mss_clamp. if (search_high - search_low) 20416b58e0a5SFan Du * smaller than a threshold, backoff from probing. 20426b58e0a5SFan Du */ 20430c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 20446b58e0a5SFan Du probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high + 20456b58e0a5SFan Du icsk->icsk_mtup.search_low) >> 1); 204691cc17c0SIlpo Järvinen size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; 20476b58e0a5SFan Du interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low; 204805cbc0dbSFan Du /* When misfortune happens, we are reprobing actively, 204905cbc0dbSFan Du * and then reprobe timer has expired. We stick with current 205005cbc0dbSFan Du * probing process by not resetting search range to its orignal. 205105cbc0dbSFan Du */ 20526b58e0a5SFan Du if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) || 205305cbc0dbSFan Du interval < net->ipv4.sysctl_tcp_probe_threshold) { 205405cbc0dbSFan Du /* Check whether enough time has elaplased for 205505cbc0dbSFan Du * another round of probing. 205605cbc0dbSFan Du */ 205705cbc0dbSFan Du tcp_mtu_check_reprobe(sk); 20585d424d5aSJohn Heffner return -1; 20595d424d5aSJohn Heffner } 20605d424d5aSJohn Heffner 20615d424d5aSJohn Heffner /* Have enough data in the send queue to probe? */ 20627f9c33e5SIlpo Järvinen if (tp->write_seq - tp->snd_nxt < size_needed) 20635d424d5aSJohn Heffner return -1; 20645d424d5aSJohn Heffner 206591cc17c0SIlpo Järvinen if (tp->snd_wnd < size_needed) 20665d424d5aSJohn Heffner return -1; 206790840defSIlpo Järvinen if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) 20685d424d5aSJohn Heffner return 0; 20695d424d5aSJohn Heffner 2070d67c58e9SIlpo Järvinen /* Do we need to wait to drain cwnd? With none in flight, don't stall */ 2071d67c58e9SIlpo Järvinen if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { 2072d67c58e9SIlpo Järvinen if (!tcp_packets_in_flight(tp)) 20735d424d5aSJohn Heffner return -1; 20745d424d5aSJohn Heffner else 20755d424d5aSJohn Heffner return 0; 20765d424d5aSJohn Heffner } 20775d424d5aSJohn Heffner 20785d424d5aSJohn Heffner /* We're allowed to probe. Build it now. */ 2079eb934478SEric Dumazet nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false); 208051456b29SIan Morris if (!nskb) 20815d424d5aSJohn Heffner return -1; 20823ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 20833ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 20845d424d5aSJohn Heffner 2085fe067e8aSDavid S. Miller skb = tcp_send_head(sk); 20865d424d5aSJohn Heffner 20875d424d5aSJohn Heffner TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 20885d424d5aSJohn Heffner TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 20894de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; 20905d424d5aSJohn Heffner TCP_SKB_CB(nskb)->sacked = 0; 20915d424d5aSJohn Heffner nskb->csum = 0; 209284fa7933SPatrick McHardy nskb->ip_summed = skb->ip_summed; 20935d424d5aSJohn Heffner 209450c4817eSIlpo Järvinen tcp_insert_write_queue_before(nskb, skb, sk); 209550c4817eSIlpo Järvinen 20965d424d5aSJohn Heffner len = 0; 2097234b6860SIlpo Järvinen tcp_for_write_queue_from_safe(skb, next, sk) { 20985d424d5aSJohn Heffner copy = min_t(int, skb->len, probe_size - len); 20992fe664f1SDouglas Caetano dos Santos if (nskb->ip_summed) { 21005d424d5aSJohn Heffner skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); 21012fe664f1SDouglas Caetano dos Santos } else { 21022fe664f1SDouglas Caetano dos Santos __wsum csum = skb_copy_and_csum_bits(skb, 0, 2103056834d9SIlpo Järvinen skb_put(nskb, copy), 21042fe664f1SDouglas Caetano dos Santos copy, 0); 21052fe664f1SDouglas Caetano dos Santos nskb->csum = csum_block_add(nskb->csum, csum, len); 21062fe664f1SDouglas Caetano dos Santos } 21075d424d5aSJohn Heffner 21085d424d5aSJohn Heffner if (skb->len <= copy) { 21095d424d5aSJohn Heffner /* We've eaten all the data from this skb. 21105d424d5aSJohn Heffner * Throw it away. */ 21114de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 2112fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 21133ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 21145d424d5aSJohn Heffner } else { 21154de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & 2116a3433f35SChangli Gao ~(TCPHDR_FIN|TCPHDR_PSH); 21175d424d5aSJohn Heffner if (!skb_shinfo(skb)->nr_frags) { 21185d424d5aSJohn Heffner skb_pull(skb, copy); 211984fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 2120056834d9SIlpo Järvinen skb->csum = csum_partial(skb->data, 2121056834d9SIlpo Järvinen skb->len, 0); 21225d424d5aSJohn Heffner } else { 21235d424d5aSJohn Heffner __pskb_trim_head(skb, copy); 21245bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 21255d424d5aSJohn Heffner } 21265d424d5aSJohn Heffner TCP_SKB_CB(skb)->seq += copy; 21275d424d5aSJohn Heffner } 21285d424d5aSJohn Heffner 21295d424d5aSJohn Heffner len += copy; 2130234b6860SIlpo Järvinen 2131234b6860SIlpo Järvinen if (len >= probe_size) 2132234b6860SIlpo Järvinen break; 21335d424d5aSJohn Heffner } 21345bbb432cSEric Dumazet tcp_init_tso_segs(nskb, nskb->len); 21355d424d5aSJohn Heffner 21365d424d5aSJohn Heffner /* We're ready to send. If this fails, the probe will 21377faee5c0SEric Dumazet * be resegmented into mss-sized pieces by tcp_write_xmit(). 21387faee5c0SEric Dumazet */ 21395d424d5aSJohn Heffner if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 21405d424d5aSJohn Heffner /* Decrement cwnd here because we are sending 21415d424d5aSJohn Heffner * effectively two packets. */ 21425d424d5aSJohn Heffner tp->snd_cwnd--; 214366f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, nskb); 21445d424d5aSJohn Heffner 21455d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 21460e7b1368SJohn Heffner tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 21470e7b1368SJohn Heffner tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 21485d424d5aSJohn Heffner 21495d424d5aSJohn Heffner return 1; 21505d424d5aSJohn Heffner } 21515d424d5aSJohn Heffner 21525d424d5aSJohn Heffner return -1; 21535d424d5aSJohn Heffner } 21545d424d5aSJohn Heffner 2155218af599SEric Dumazet static bool tcp_pacing_check(const struct sock *sk) 2156218af599SEric Dumazet { 2157218af599SEric Dumazet return tcp_needs_internal_pacing(sk) && 2158218af599SEric Dumazet hrtimer_active(&tcp_sk(sk)->pacing_timer); 2159218af599SEric Dumazet } 2160218af599SEric Dumazet 2161f9616c35SEric Dumazet /* TCP Small Queues : 2162f9616c35SEric Dumazet * Control number of packets in qdisc/devices to two packets / or ~1 ms. 2163f9616c35SEric Dumazet * (These limits are doubled for retransmits) 2164f9616c35SEric Dumazet * This allows for : 2165f9616c35SEric Dumazet * - better RTT estimation and ACK scheduling 2166f9616c35SEric Dumazet * - faster recovery 2167f9616c35SEric Dumazet * - high rates 2168f9616c35SEric Dumazet * Alas, some drivers / subsystems require a fair amount 2169f9616c35SEric Dumazet * of queued bytes to ensure line rate. 2170f9616c35SEric Dumazet * One example is wifi aggregation (802.11 AMPDU) 2171f9616c35SEric Dumazet */ 2172f9616c35SEric Dumazet static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb, 2173f9616c35SEric Dumazet unsigned int factor) 2174f9616c35SEric Dumazet { 2175f9616c35SEric Dumazet unsigned int limit; 2176f9616c35SEric Dumazet 2177f9616c35SEric Dumazet limit = max(2 * skb->truesize, sk->sk_pacing_rate >> 10); 2178f9616c35SEric Dumazet limit = min_t(u32, limit, sysctl_tcp_limit_output_bytes); 2179f9616c35SEric Dumazet limit <<= factor; 2180f9616c35SEric Dumazet 2181f9616c35SEric Dumazet if (atomic_read(&sk->sk_wmem_alloc) > limit) { 218275eefc6cSEric Dumazet /* Always send the 1st or 2nd skb in write queue. 218375eefc6cSEric Dumazet * No need to wait for TX completion to call us back, 218475eefc6cSEric Dumazet * after softirq/tasklet schedule. 218575eefc6cSEric Dumazet * This helps when TX completions are delayed too much. 218675eefc6cSEric Dumazet */ 218775eefc6cSEric Dumazet if (skb == sk->sk_write_queue.next || 218875eefc6cSEric Dumazet skb->prev == sk->sk_write_queue.next) 218975eefc6cSEric Dumazet return false; 219075eefc6cSEric Dumazet 21917aa5470cSEric Dumazet set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags); 2192f9616c35SEric Dumazet /* It is possible TX completion already happened 2193f9616c35SEric Dumazet * before we set TSQ_THROTTLED, so we must 2194f9616c35SEric Dumazet * test again the condition. 2195f9616c35SEric Dumazet */ 2196f9616c35SEric Dumazet smp_mb__after_atomic(); 2197f9616c35SEric Dumazet if (atomic_read(&sk->sk_wmem_alloc) > limit) 2198f9616c35SEric Dumazet return true; 2199f9616c35SEric Dumazet } 2200f9616c35SEric Dumazet return false; 2201f9616c35SEric Dumazet } 2202f9616c35SEric Dumazet 220305b055e8SFrancis Yan static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new) 220405b055e8SFrancis Yan { 220505b055e8SFrancis Yan const u32 now = tcp_time_stamp; 220605b055e8SFrancis Yan 220705b055e8SFrancis Yan if (tp->chrono_type > TCP_CHRONO_UNSPEC) 220805b055e8SFrancis Yan tp->chrono_stat[tp->chrono_type - 1] += now - tp->chrono_start; 220905b055e8SFrancis Yan tp->chrono_start = now; 221005b055e8SFrancis Yan tp->chrono_type = new; 221105b055e8SFrancis Yan } 221205b055e8SFrancis Yan 221305b055e8SFrancis Yan void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type) 221405b055e8SFrancis Yan { 221505b055e8SFrancis Yan struct tcp_sock *tp = tcp_sk(sk); 221605b055e8SFrancis Yan 221705b055e8SFrancis Yan /* If there are multiple conditions worthy of tracking in a 22180f87230dSFrancis Yan * chronograph then the highest priority enum takes precedence 22190f87230dSFrancis Yan * over the other conditions. So that if something "more interesting" 222005b055e8SFrancis Yan * starts happening, stop the previous chrono and start a new one. 222105b055e8SFrancis Yan */ 222205b055e8SFrancis Yan if (type > tp->chrono_type) 222305b055e8SFrancis Yan tcp_chrono_set(tp, type); 222405b055e8SFrancis Yan } 222505b055e8SFrancis Yan 222605b055e8SFrancis Yan void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type) 222705b055e8SFrancis Yan { 222805b055e8SFrancis Yan struct tcp_sock *tp = tcp_sk(sk); 222905b055e8SFrancis Yan 22300f87230dSFrancis Yan 22310f87230dSFrancis Yan /* There are multiple conditions worthy of tracking in a 22320f87230dSFrancis Yan * chronograph, so that the highest priority enum takes 22330f87230dSFrancis Yan * precedence over the other conditions (see tcp_chrono_start). 22340f87230dSFrancis Yan * If a condition stops, we only stop chrono tracking if 22350f87230dSFrancis Yan * it's the "most interesting" or current chrono we are 22360f87230dSFrancis Yan * tracking and starts busy chrono if we have pending data. 22370f87230dSFrancis Yan */ 22380f87230dSFrancis Yan if (tcp_write_queue_empty(sk)) 223905b055e8SFrancis Yan tcp_chrono_set(tp, TCP_CHRONO_UNSPEC); 22400f87230dSFrancis Yan else if (type == tp->chrono_type) 22410f87230dSFrancis Yan tcp_chrono_set(tp, TCP_CHRONO_BUSY); 224205b055e8SFrancis Yan } 224305b055e8SFrancis Yan 22441da177e4SLinus Torvalds /* This routine writes packets to the network. It advances the 22451da177e4SLinus Torvalds * send_head. This happens as incoming acks open up the remote 22461da177e4SLinus Torvalds * window for us. 22471da177e4SLinus Torvalds * 2248f8269a49SIlpo Järvinen * LARGESEND note: !tcp_urg_mode is overkill, only frames between 2249f8269a49SIlpo Järvinen * snd_up-64k-mss .. snd_up cannot be large. However, taking into 2250f8269a49SIlpo Järvinen * account rare use of URG, this is not a big flaw. 2251f8269a49SIlpo Järvinen * 22526ba8a3b1SNandita Dukkipati * Send at most one packet when push_one > 0. Temporarily ignore 22536ba8a3b1SNandita Dukkipati * cwnd limit to force at most one packet out when push_one == 2. 22546ba8a3b1SNandita Dukkipati 2255a2a385d6SEric Dumazet * Returns true, if no segments are in flight and we have queued segments, 2256a2a385d6SEric Dumazet * but cannot send anything now because of SWS or another problem. 22571da177e4SLinus Torvalds */ 2258a2a385d6SEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 2259d5dd9175SIlpo Järvinen int push_one, gfp_t gfp) 22601da177e4SLinus Torvalds { 22611da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 226292df7b51SDavid S. Miller struct sk_buff *skb; 2263c1b4a7e6SDavid S. Miller unsigned int tso_segs, sent_pkts; 2264c1b4a7e6SDavid S. Miller int cwnd_quota; 22655d424d5aSJohn Heffner int result; 22665615f886SFrancis Yan bool is_cwnd_limited = false, is_rwnd_limited = false; 2267605ad7f1SEric Dumazet u32 max_segs; 22681da177e4SLinus Torvalds 2269c1b4a7e6SDavid S. Miller sent_pkts = 0; 22705d424d5aSJohn Heffner 2271d5dd9175SIlpo Järvinen if (!push_one) { 22725d424d5aSJohn Heffner /* Do MTU probing. */ 2273d5dd9175SIlpo Järvinen result = tcp_mtu_probe(sk); 2274d5dd9175SIlpo Järvinen if (!result) { 2275a2a385d6SEric Dumazet return false; 22765d424d5aSJohn Heffner } else if (result > 0) { 22775d424d5aSJohn Heffner sent_pkts = 1; 22785d424d5aSJohn Heffner } 2279d5dd9175SIlpo Järvinen } 22805d424d5aSJohn Heffner 2281ed6e7268SNeal Cardwell max_segs = tcp_tso_segs(sk, mss_now); 2282385e2070SEric Dumazet skb_mstamp_get(&tp->tcp_mstamp); 2283fe067e8aSDavid S. Miller while ((skb = tcp_send_head(sk))) { 2284c8ac3774SHerbert Xu unsigned int limit; 2285c8ac3774SHerbert Xu 2286218af599SEric Dumazet if (tcp_pacing_check(sk)) 2287218af599SEric Dumazet break; 2288218af599SEric Dumazet 22895bbb432cSEric Dumazet tso_segs = tcp_init_tso_segs(skb, mss_now); 2290c1b4a7e6SDavid S. Miller BUG_ON(!tso_segs); 2291c1b4a7e6SDavid S. Miller 22929d186cacSAndrey Vagin if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) { 22937faee5c0SEric Dumazet /* "skb_mstamp" is used as a start point for the retransmit timer */ 2294385e2070SEric Dumazet skb->skb_mstamp = tp->tcp_mstamp; 2295ec342325SAndrew Vagin goto repair; /* Skip network transmission */ 22969d186cacSAndrey Vagin } 2297ec342325SAndrew Vagin 2298b68e9f85SHerbert Xu cwnd_quota = tcp_cwnd_test(tp, skb); 22996ba8a3b1SNandita Dukkipati if (!cwnd_quota) { 23006ba8a3b1SNandita Dukkipati if (push_one == 2) 23016ba8a3b1SNandita Dukkipati /* Force out a loss probe pkt. */ 23026ba8a3b1SNandita Dukkipati cwnd_quota = 1; 23036ba8a3b1SNandita Dukkipati else 2304b68e9f85SHerbert Xu break; 23056ba8a3b1SNandita Dukkipati } 2306b68e9f85SHerbert Xu 23075615f886SFrancis Yan if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) { 23085615f886SFrancis Yan is_rwnd_limited = true; 2309b68e9f85SHerbert Xu break; 23105615f886SFrancis Yan } 2311b68e9f85SHerbert Xu 2312d6a4e26aSEric Dumazet if (tso_segs == 1) { 2313aa93466bSDavid S. Miller if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 2314aa93466bSDavid S. Miller (tcp_skb_is_last(sk, skb) ? 2315aa93466bSDavid S. Miller nonagle : TCP_NAGLE_PUSH)))) 2316aa93466bSDavid S. Miller break; 2317c1b4a7e6SDavid S. Miller } else { 2318ca8a2263SNeal Cardwell if (!push_one && 2319605ad7f1SEric Dumazet tcp_tso_should_defer(sk, skb, &is_cwnd_limited, 2320605ad7f1SEric Dumazet max_segs)) 2321aa93466bSDavid S. Miller break; 2322c1b4a7e6SDavid S. Miller } 2323aa93466bSDavid S. Miller 2324605ad7f1SEric Dumazet limit = mss_now; 2325d6a4e26aSEric Dumazet if (tso_segs > 1 && !tcp_urg_mode(tp)) 2326605ad7f1SEric Dumazet limit = tcp_mss_split_point(sk, skb, mss_now, 2327605ad7f1SEric Dumazet min_t(unsigned int, 2328605ad7f1SEric Dumazet cwnd_quota, 2329605ad7f1SEric Dumazet max_segs), 2330605ad7f1SEric Dumazet nonagle); 2331605ad7f1SEric Dumazet 2332605ad7f1SEric Dumazet if (skb->len > limit && 2333605ad7f1SEric Dumazet unlikely(tso_fragment(sk, skb, limit, mss_now, gfp))) 2334605ad7f1SEric Dumazet break; 2335605ad7f1SEric Dumazet 23367aa5470cSEric Dumazet if (test_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags)) 23377aa5470cSEric Dumazet clear_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags); 2338f9616c35SEric Dumazet if (tcp_small_queue_check(sk, skb, 0)) 233946d3ceabSEric Dumazet break; 2340c9eeec26SEric Dumazet 2341d5dd9175SIlpo Järvinen if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) 23421da177e4SLinus Torvalds break; 23431da177e4SLinus Torvalds 2344ec342325SAndrew Vagin repair: 23451da177e4SLinus Torvalds /* Advance the send_head. This one is sent out. 23461da177e4SLinus Torvalds * This call will increment packets_out. 23471da177e4SLinus Torvalds */ 234866f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 23491da177e4SLinus Torvalds 23501da177e4SLinus Torvalds tcp_minshall_update(tp, mss_now, skb); 2351a262f0cdSNandita Dukkipati sent_pkts += tcp_skb_pcount(skb); 2352d5dd9175SIlpo Järvinen 2353d5dd9175SIlpo Järvinen if (push_one) 2354d5dd9175SIlpo Järvinen break; 23551da177e4SLinus Torvalds } 23561da177e4SLinus Torvalds 23575615f886SFrancis Yan if (is_rwnd_limited) 23585615f886SFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED); 23595615f886SFrancis Yan else 23605615f886SFrancis Yan tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED); 23615615f886SFrancis Yan 2362aa93466bSDavid S. Miller if (likely(sent_pkts)) { 2363684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 2364684bad11SYuchung Cheng tp->prr_out += sent_pkts; 23656ba8a3b1SNandita Dukkipati 23666ba8a3b1SNandita Dukkipati /* Send one loss probe per tail loss episode. */ 23676ba8a3b1SNandita Dukkipati if (push_one != 2) 23686ba8a3b1SNandita Dukkipati tcp_schedule_loss_probe(sk); 2369d2e1339fSBendik Rønning Opstad is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd); 2370ca8a2263SNeal Cardwell tcp_cwnd_validate(sk, is_cwnd_limited); 2371a2a385d6SEric Dumazet return false; 23721da177e4SLinus Torvalds } 2373b340b264SYuchung Cheng return !tp->packets_out && tcp_send_head(sk); 23746ba8a3b1SNandita Dukkipati } 23756ba8a3b1SNandita Dukkipati 23766ba8a3b1SNandita Dukkipati bool tcp_schedule_loss_probe(struct sock *sk) 23776ba8a3b1SNandita Dukkipati { 23786ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 23796ba8a3b1SNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 23806ba8a3b1SNandita Dukkipati u32 timeout, tlp_time_stamp, rto_time_stamp; 2381740b0f18SEric Dumazet u32 rtt = usecs_to_jiffies(tp->srtt_us >> 3); 23826ba8a3b1SNandita Dukkipati 23836ba8a3b1SNandita Dukkipati /* No consecutive loss probes. */ 23846ba8a3b1SNandita Dukkipati if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) { 23856ba8a3b1SNandita Dukkipati tcp_rearm_rto(sk); 23866ba8a3b1SNandita Dukkipati return false; 23876ba8a3b1SNandita Dukkipati } 23886ba8a3b1SNandita Dukkipati /* Don't do any loss probe on a Fast Open connection before 3WHS 23896ba8a3b1SNandita Dukkipati * finishes. 23906ba8a3b1SNandita Dukkipati */ 2391f9b99582SYuchung Cheng if (tp->fastopen_rsk) 23926ba8a3b1SNandita Dukkipati return false; 23936ba8a3b1SNandita Dukkipati 23946ba8a3b1SNandita Dukkipati /* TLP is only scheduled when next timer event is RTO. */ 23956ba8a3b1SNandita Dukkipati if (icsk->icsk_pending != ICSK_TIME_RETRANS) 23966ba8a3b1SNandita Dukkipati return false; 23976ba8a3b1SNandita Dukkipati 23986ba8a3b1SNandita Dukkipati /* Schedule a loss probe in 2*RTT for SACK capable connections 23996ba8a3b1SNandita Dukkipati * in Open state, that are either limited by cwnd or application. 24006ba8a3b1SNandita Dukkipati */ 2401bec41a11SYuchung Cheng if ((sysctl_tcp_early_retrans != 3 && sysctl_tcp_early_retrans != 4) || 2402bec41a11SYuchung Cheng !tp->packets_out || !tcp_is_sack(tp) || 2403bec41a11SYuchung Cheng icsk->icsk_ca_state != TCP_CA_Open) 24046ba8a3b1SNandita Dukkipati return false; 24056ba8a3b1SNandita Dukkipati 24066ba8a3b1SNandita Dukkipati if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) && 24076ba8a3b1SNandita Dukkipati tcp_send_head(sk)) 24086ba8a3b1SNandita Dukkipati return false; 24096ba8a3b1SNandita Dukkipati 24106ba8a3b1SNandita Dukkipati /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account 2411f9b99582SYuchung Cheng * for delayed ack when there's one outstanding packet. If no RTT 2412f9b99582SYuchung Cheng * sample is available then probe after TCP_TIMEOUT_INIT. 24136ba8a3b1SNandita Dukkipati */ 2414f9b99582SYuchung Cheng timeout = rtt << 1 ? : TCP_TIMEOUT_INIT; 24156ba8a3b1SNandita Dukkipati if (tp->packets_out == 1) 24166ba8a3b1SNandita Dukkipati timeout = max_t(u32, timeout, 24176ba8a3b1SNandita Dukkipati (rtt + (rtt >> 1) + TCP_DELACK_MAX)); 24186ba8a3b1SNandita Dukkipati timeout = max_t(u32, timeout, msecs_to_jiffies(10)); 24196ba8a3b1SNandita Dukkipati 24206ba8a3b1SNandita Dukkipati /* If RTO is shorter, just schedule TLP in its place. */ 24216ba8a3b1SNandita Dukkipati tlp_time_stamp = tcp_time_stamp + timeout; 24226ba8a3b1SNandita Dukkipati rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout; 24236ba8a3b1SNandita Dukkipati if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) { 24246ba8a3b1SNandita Dukkipati s32 delta = rto_time_stamp - tcp_time_stamp; 24256ba8a3b1SNandita Dukkipati if (delta > 0) 24266ba8a3b1SNandita Dukkipati timeout = delta; 24276ba8a3b1SNandita Dukkipati } 24286ba8a3b1SNandita Dukkipati 24296ba8a3b1SNandita Dukkipati inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, 24306ba8a3b1SNandita Dukkipati TCP_RTO_MAX); 24316ba8a3b1SNandita Dukkipati return true; 24326ba8a3b1SNandita Dukkipati } 24336ba8a3b1SNandita Dukkipati 24341f3279aeSEric Dumazet /* Thanks to skb fast clones, we can detect if a prior transmit of 24351f3279aeSEric Dumazet * a packet is still in a qdisc or driver queue. 24361f3279aeSEric Dumazet * In this case, there is very little point doing a retransmit ! 24371f3279aeSEric Dumazet */ 24381f3279aeSEric Dumazet static bool skb_still_in_host_queue(const struct sock *sk, 24391f3279aeSEric Dumazet const struct sk_buff *skb) 24401f3279aeSEric Dumazet { 244139bb5e62SEric Dumazet if (unlikely(skb_fclone_busy(sk, skb))) { 2442c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), 24431f3279aeSEric Dumazet LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES); 24441f3279aeSEric Dumazet return true; 24451f3279aeSEric Dumazet } 24461f3279aeSEric Dumazet return false; 24471f3279aeSEric Dumazet } 24481f3279aeSEric Dumazet 2449b340b264SYuchung Cheng /* When probe timeout (PTO) fires, try send a new segment if possible, else 24506ba8a3b1SNandita Dukkipati * retransmit the last segment. 24516ba8a3b1SNandita Dukkipati */ 24526ba8a3b1SNandita Dukkipati void tcp_send_loss_probe(struct sock *sk) 24536ba8a3b1SNandita Dukkipati { 24549b717a8dSNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 24556ba8a3b1SNandita Dukkipati struct sk_buff *skb; 24566ba8a3b1SNandita Dukkipati int pcount; 24576ba8a3b1SNandita Dukkipati int mss = tcp_current_mss(sk); 24586ba8a3b1SNandita Dukkipati 2459b340b264SYuchung Cheng skb = tcp_send_head(sk); 2460b340b264SYuchung Cheng if (skb) { 2461b340b264SYuchung Cheng if (tcp_snd_wnd_test(tp, skb, mss)) { 2462b340b264SYuchung Cheng pcount = tp->packets_out; 2463b340b264SYuchung Cheng tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC); 2464b340b264SYuchung Cheng if (tp->packets_out > pcount) 2465b340b264SYuchung Cheng goto probe_sent; 24666ba8a3b1SNandita Dukkipati goto rearm_timer; 24676ba8a3b1SNandita Dukkipati } 2468b340b264SYuchung Cheng skb = tcp_write_queue_prev(sk, skb); 2469b340b264SYuchung Cheng } else { 2470b340b264SYuchung Cheng skb = tcp_write_queue_tail(sk); 2471b340b264SYuchung Cheng } 24726ba8a3b1SNandita Dukkipati 24739b717a8dSNandita Dukkipati /* At most one outstanding TLP retransmission. */ 24749b717a8dSNandita Dukkipati if (tp->tlp_high_seq) 24759b717a8dSNandita Dukkipati goto rearm_timer; 24769b717a8dSNandita Dukkipati 24776ba8a3b1SNandita Dukkipati /* Retransmit last segment. */ 24786ba8a3b1SNandita Dukkipati if (WARN_ON(!skb)) 24796ba8a3b1SNandita Dukkipati goto rearm_timer; 24806ba8a3b1SNandita Dukkipati 24811f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 24821f3279aeSEric Dumazet goto rearm_timer; 24831f3279aeSEric Dumazet 24846ba8a3b1SNandita Dukkipati pcount = tcp_skb_pcount(skb); 24856ba8a3b1SNandita Dukkipati if (WARN_ON(!pcount)) 24866ba8a3b1SNandita Dukkipati goto rearm_timer; 24876ba8a3b1SNandita Dukkipati 24886ba8a3b1SNandita Dukkipati if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) { 24896cc55e09SOctavian Purdila if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss, 24906cc55e09SOctavian Purdila GFP_ATOMIC))) 24916ba8a3b1SNandita Dukkipati goto rearm_timer; 2492b340b264SYuchung Cheng skb = tcp_write_queue_next(sk, skb); 24936ba8a3b1SNandita Dukkipati } 24946ba8a3b1SNandita Dukkipati 24956ba8a3b1SNandita Dukkipati if (WARN_ON(!skb || !tcp_skb_pcount(skb))) 24966ba8a3b1SNandita Dukkipati goto rearm_timer; 24976ba8a3b1SNandita Dukkipati 249810d3be56SEric Dumazet if (__tcp_retransmit_skb(sk, skb, 1)) 2499b340b264SYuchung Cheng goto rearm_timer; 25006ba8a3b1SNandita Dukkipati 25019b717a8dSNandita Dukkipati /* Record snd_nxt for loss detection. */ 25029b717a8dSNandita Dukkipati tp->tlp_high_seq = tp->snd_nxt; 25039b717a8dSNandita Dukkipati 2504b340b264SYuchung Cheng probe_sent: 2505c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES); 2506fcd16c0aSYuchung Cheng /* Reset s.t. tcp_rearm_rto will restart timer from now */ 2507fcd16c0aSYuchung Cheng inet_csk(sk)->icsk_pending = 0; 2508b340b264SYuchung Cheng rearm_timer: 2509fcd16c0aSYuchung Cheng tcp_rearm_rto(sk); 25101da177e4SLinus Torvalds } 25111da177e4SLinus Torvalds 2512a762a980SDavid S. Miller /* Push out any pending frames which were held back due to 2513a762a980SDavid S. Miller * TCP_CORK or attempt at coalescing tiny packets. 2514a762a980SDavid S. Miller * The socket must be locked by the caller. 2515a762a980SDavid S. Miller */ 25169e412ba7SIlpo Järvinen void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 25179e412ba7SIlpo Järvinen int nonagle) 2518a762a980SDavid S. Miller { 2519726e07a8SIlpo Järvinen /* If we are closed, the bytes will have to remain here. 2520726e07a8SIlpo Järvinen * In time closedown will finish, we empty the write queue and 2521726e07a8SIlpo Järvinen * all will be happy. 2522726e07a8SIlpo Järvinen */ 2523726e07a8SIlpo Järvinen if (unlikely(sk->sk_state == TCP_CLOSE)) 2524726e07a8SIlpo Järvinen return; 2525726e07a8SIlpo Järvinen 252699a1dec7SMel Gorman if (tcp_write_xmit(sk, cur_mss, nonagle, 0, 25277450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC))) 25289e412ba7SIlpo Järvinen tcp_check_probe_timer(sk); 2529a762a980SDavid S. Miller } 2530a762a980SDavid S. Miller 2531c1b4a7e6SDavid S. Miller /* Send _single_ skb sitting at the send head. This function requires 2532c1b4a7e6SDavid S. Miller * true push pending frames to setup probe timer etc. 2533c1b4a7e6SDavid S. Miller */ 2534c1b4a7e6SDavid S. Miller void tcp_push_one(struct sock *sk, unsigned int mss_now) 2535c1b4a7e6SDavid S. Miller { 2536fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 2537c1b4a7e6SDavid S. Miller 2538c1b4a7e6SDavid S. Miller BUG_ON(!skb || skb->len < mss_now); 2539c1b4a7e6SDavid S. Miller 2540d5dd9175SIlpo Järvinen tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); 2541c1b4a7e6SDavid S. Miller } 2542c1b4a7e6SDavid S. Miller 25431da177e4SLinus Torvalds /* This function returns the amount that we can raise the 25441da177e4SLinus Torvalds * usable window based on the following constraints 25451da177e4SLinus Torvalds * 25461da177e4SLinus Torvalds * 1. The window can never be shrunk once it is offered (RFC 793) 25471da177e4SLinus Torvalds * 2. We limit memory per socket 25481da177e4SLinus Torvalds * 25491da177e4SLinus Torvalds * RFC 1122: 25501da177e4SLinus Torvalds * "the suggested [SWS] avoidance algorithm for the receiver is to keep 25511da177e4SLinus Torvalds * RECV.NEXT + RCV.WIN fixed until: 25521da177e4SLinus Torvalds * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 25531da177e4SLinus Torvalds * 25541da177e4SLinus Torvalds * i.e. don't raise the right edge of the window until you can raise 25551da177e4SLinus Torvalds * it at least MSS bytes. 25561da177e4SLinus Torvalds * 25571da177e4SLinus Torvalds * Unfortunately, the recommended algorithm breaks header prediction, 25581da177e4SLinus Torvalds * since header prediction assumes th->window stays fixed. 25591da177e4SLinus Torvalds * 25601da177e4SLinus Torvalds * Strictly speaking, keeping th->window fixed violates the receiver 25611da177e4SLinus Torvalds * side SWS prevention criteria. The problem is that under this rule 25621da177e4SLinus Torvalds * a stream of single byte packets will cause the right side of the 25631da177e4SLinus Torvalds * window to always advance by a single byte. 25641da177e4SLinus Torvalds * 25651da177e4SLinus Torvalds * Of course, if the sender implements sender side SWS prevention 25661da177e4SLinus Torvalds * then this will not be a problem. 25671da177e4SLinus Torvalds * 25681da177e4SLinus Torvalds * BSD seems to make the following compromise: 25691da177e4SLinus Torvalds * 25701da177e4SLinus Torvalds * If the free space is less than the 1/4 of the maximum 25711da177e4SLinus Torvalds * space available and the free space is less than 1/2 mss, 25721da177e4SLinus Torvalds * then set the window to 0. 25731da177e4SLinus Torvalds * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 25741da177e4SLinus Torvalds * Otherwise, just prevent the window from shrinking 25751da177e4SLinus Torvalds * and from being larger than the largest representable value. 25761da177e4SLinus Torvalds * 25771da177e4SLinus Torvalds * This prevents incremental opening of the window in the regime 25781da177e4SLinus Torvalds * where TCP is limited by the speed of the reader side taking 25791da177e4SLinus Torvalds * data out of the TCP receive queue. It does nothing about 25801da177e4SLinus Torvalds * those cases where the window is constrained on the sender side 25811da177e4SLinus Torvalds * because the pipeline is full. 25821da177e4SLinus Torvalds * 25831da177e4SLinus Torvalds * BSD also seems to "accidentally" limit itself to windows that are a 25841da177e4SLinus Torvalds * multiple of MSS, at least until the free space gets quite small. 25851da177e4SLinus Torvalds * This would appear to be a side effect of the mbuf implementation. 25861da177e4SLinus Torvalds * Combining these two algorithms results in the observed behavior 25871da177e4SLinus Torvalds * of having a fixed window size at almost all times. 25881da177e4SLinus Torvalds * 25891da177e4SLinus Torvalds * Below we obtain similar behavior by forcing the offered window to 25901da177e4SLinus Torvalds * a multiple of the mss when it is feasible to do so. 25911da177e4SLinus Torvalds * 25921da177e4SLinus Torvalds * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 25931da177e4SLinus Torvalds * Regular options like TIMESTAMP are taken into account. 25941da177e4SLinus Torvalds */ 25951da177e4SLinus Torvalds u32 __tcp_select_window(struct sock *sk) 25961da177e4SLinus Torvalds { 2597463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 25981da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2599caa20d9aSStephen Hemminger /* MSS for the peer's data. Previous versions used mss_clamp 26001da177e4SLinus Torvalds * here. I don't know if the value based on our guesses 26011da177e4SLinus Torvalds * of peer's MSS is better for the performance. It's more correct 26021da177e4SLinus Torvalds * but may be worse for the performance because of rcv_mss 26031da177e4SLinus Torvalds * fluctuations. --SAW 1998/11/1 26041da177e4SLinus Torvalds */ 2605463c84b9SArnaldo Carvalho de Melo int mss = icsk->icsk_ack.rcv_mss; 26061da177e4SLinus Torvalds int free_space = tcp_space(sk); 260786c1a045SFlorian Westphal int allowed_space = tcp_full_space(sk); 260886c1a045SFlorian Westphal int full_space = min_t(int, tp->window_clamp, allowed_space); 26091da177e4SLinus Torvalds int window; 26101da177e4SLinus Torvalds 261106425c30SEric Dumazet if (unlikely(mss > full_space)) { 26121da177e4SLinus Torvalds mss = full_space; 261306425c30SEric Dumazet if (mss <= 0) 261406425c30SEric Dumazet return 0; 261506425c30SEric Dumazet } 2616b92edbe0SEric Dumazet if (free_space < (full_space >> 1)) { 2617463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.quick = 0; 26181da177e4SLinus Torvalds 2619b8da51ebSEric Dumazet if (tcp_under_memory_pressure(sk)) 2620056834d9SIlpo Järvinen tp->rcv_ssthresh = min(tp->rcv_ssthresh, 2621056834d9SIlpo Järvinen 4U * tp->advmss); 26221da177e4SLinus Torvalds 262386c1a045SFlorian Westphal /* free_space might become our new window, make sure we don't 262486c1a045SFlorian Westphal * increase it due to wscale. 262586c1a045SFlorian Westphal */ 262686c1a045SFlorian Westphal free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale); 262786c1a045SFlorian Westphal 262886c1a045SFlorian Westphal /* if free space is less than mss estimate, or is below 1/16th 262986c1a045SFlorian Westphal * of the maximum allowed, try to move to zero-window, else 263086c1a045SFlorian Westphal * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and 263186c1a045SFlorian Westphal * new incoming data is dropped due to memory limits. 263286c1a045SFlorian Westphal * With large window, mss test triggers way too late in order 263386c1a045SFlorian Westphal * to announce zero window in time before rmem limit kicks in. 263486c1a045SFlorian Westphal */ 263586c1a045SFlorian Westphal if (free_space < (allowed_space >> 4) || free_space < mss) 26361da177e4SLinus Torvalds return 0; 26371da177e4SLinus Torvalds } 26381da177e4SLinus Torvalds 26391da177e4SLinus Torvalds if (free_space > tp->rcv_ssthresh) 26401da177e4SLinus Torvalds free_space = tp->rcv_ssthresh; 26411da177e4SLinus Torvalds 26421da177e4SLinus Torvalds /* Don't do rounding if we are using window scaling, since the 26431da177e4SLinus Torvalds * scaled window will not line up with the MSS boundary anyway. 26441da177e4SLinus Torvalds */ 26451da177e4SLinus Torvalds if (tp->rx_opt.rcv_wscale) { 26461da177e4SLinus Torvalds window = free_space; 26471da177e4SLinus Torvalds 26481da177e4SLinus Torvalds /* Advertise enough space so that it won't get scaled away. 26491da177e4SLinus Torvalds * Import case: prevent zero window announcement if 26501da177e4SLinus Torvalds * 1<<rcv_wscale > mss. 26511da177e4SLinus Torvalds */ 26521935299dSGao Feng window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale)); 26531da177e4SLinus Torvalds } else { 26541935299dSGao Feng window = tp->rcv_wnd; 26551da177e4SLinus Torvalds /* Get the largest window that is a nice multiple of mss. 26561da177e4SLinus Torvalds * Window clamp already applied above. 26571da177e4SLinus Torvalds * If our current window offering is within 1 mss of the 26581da177e4SLinus Torvalds * free space we just keep it. This prevents the divide 26591da177e4SLinus Torvalds * and multiply from happening most of the time. 26601da177e4SLinus Torvalds * We also don't do any window rounding when the free space 26611da177e4SLinus Torvalds * is too small. 26621da177e4SLinus Torvalds */ 26631da177e4SLinus Torvalds if (window <= free_space - mss || window > free_space) 26641935299dSGao Feng window = rounddown(free_space, mss); 266584565070SJohn Heffner else if (mss == full_space && 2666b92edbe0SEric Dumazet free_space > window + (full_space >> 1)) 266784565070SJohn Heffner window = free_space; 26681da177e4SLinus Torvalds } 26691da177e4SLinus Torvalds 26701da177e4SLinus Torvalds return window; 26711da177e4SLinus Torvalds } 26721da177e4SLinus Torvalds 2673cfea5a68SMartin KaFai Lau void tcp_skb_collapse_tstamp(struct sk_buff *skb, 2674082ac2d5SMartin KaFai Lau const struct sk_buff *next_skb) 2675082ac2d5SMartin KaFai Lau { 26760a2cf20cSSoheil Hassas Yeganeh if (unlikely(tcp_has_tx_tstamp(next_skb))) { 26770a2cf20cSSoheil Hassas Yeganeh const struct skb_shared_info *next_shinfo = 26780a2cf20cSSoheil Hassas Yeganeh skb_shinfo(next_skb); 2679082ac2d5SMartin KaFai Lau struct skb_shared_info *shinfo = skb_shinfo(skb); 2680082ac2d5SMartin KaFai Lau 26810a2cf20cSSoheil Hassas Yeganeh shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP; 2682082ac2d5SMartin KaFai Lau shinfo->tskey = next_shinfo->tskey; 26832de8023eSMartin KaFai Lau TCP_SKB_CB(skb)->txstamp_ack |= 26842de8023eSMartin KaFai Lau TCP_SKB_CB(next_skb)->txstamp_ack; 2685082ac2d5SMartin KaFai Lau } 2686082ac2d5SMartin KaFai Lau } 2687082ac2d5SMartin KaFai Lau 26884a17fc3aSIlpo Järvinen /* Collapses two adjacent SKB's during retransmission. */ 2689f8071cdeSEric Dumazet static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) 26901da177e4SLinus Torvalds { 26911da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2692fe067e8aSDavid S. Miller struct sk_buff *next_skb = tcp_write_queue_next(sk, skb); 2693058dc334SIlpo Järvinen int skb_size, next_skb_size; 26941da177e4SLinus Torvalds 2695058dc334SIlpo Järvinen skb_size = skb->len; 2696058dc334SIlpo Järvinen next_skb_size = next_skb->len; 26971da177e4SLinus Torvalds 2698058dc334SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); 26991da177e4SLinus Torvalds 2700f8071cdeSEric Dumazet if (next_skb_size) { 2701f8071cdeSEric Dumazet if (next_skb_size <= skb_availroom(skb)) 2702f8071cdeSEric Dumazet skb_copy_bits(next_skb, 0, skb_put(skb, next_skb_size), 2703f8071cdeSEric Dumazet next_skb_size); 2704f8071cdeSEric Dumazet else if (!skb_shift(skb, next_skb, next_skb_size)) 2705f8071cdeSEric Dumazet return false; 2706f8071cdeSEric Dumazet } 27076859d494SIlpo Järvinen tcp_highest_sack_combine(sk, next_skb, skb); 2708a6963a6bSIlpo Järvinen 2709fe067e8aSDavid S. Miller tcp_unlink_write_queue(next_skb, sk); 27101da177e4SLinus Torvalds 271152d570aaSJarek Poplawski if (next_skb->ip_summed == CHECKSUM_PARTIAL) 271252d570aaSJarek Poplawski skb->ip_summed = CHECKSUM_PARTIAL; 27131da177e4SLinus Torvalds 271484fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 27151da177e4SLinus Torvalds skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); 27161da177e4SLinus Torvalds 27171da177e4SLinus Torvalds /* Update sequence range on original skb. */ 27181da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 27191da177e4SLinus Torvalds 2720e6c7d085SIlpo Järvinen /* Merge over control information. This moves PSH/FIN etc. over */ 27214de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; 27221da177e4SLinus Torvalds 27231da177e4SLinus Torvalds /* All done, get rid of second SKB and account for it so 27241da177e4SLinus Torvalds * packet counting does not break. 27251da177e4SLinus Torvalds */ 27264828e7f4SIlpo Järvinen TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; 2727a643b5d4SMartin KaFai Lau TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor; 2728b7689205SIlpo Järvinen 2729b7689205SIlpo Järvinen /* changed transmit queue under us so clear hints */ 2730ef9da47cSIlpo Järvinen tcp_clear_retrans_hints_partial(tp); 2731ef9da47cSIlpo Järvinen if (next_skb == tp->retransmit_skb_hint) 2732ef9da47cSIlpo Järvinen tp->retransmit_skb_hint = skb; 2733b7689205SIlpo Järvinen 2734797108d1SIlpo Järvinen tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); 2735797108d1SIlpo Järvinen 2736082ac2d5SMartin KaFai Lau tcp_skb_collapse_tstamp(skb, next_skb); 2737082ac2d5SMartin KaFai Lau 27383ab224beSHideo Aoki sk_wmem_free_skb(sk, next_skb); 2739f8071cdeSEric Dumazet return true; 27401da177e4SLinus Torvalds } 27411da177e4SLinus Torvalds 274267edfef7SAndi Kleen /* Check if coalescing SKBs is legal. */ 2743a2a385d6SEric Dumazet static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) 27444a17fc3aSIlpo Järvinen { 27454a17fc3aSIlpo Järvinen if (tcp_skb_pcount(skb) > 1) 2746a2a385d6SEric Dumazet return false; 27474a17fc3aSIlpo Järvinen if (skb_cloned(skb)) 2748a2a385d6SEric Dumazet return false; 27494a17fc3aSIlpo Järvinen if (skb == tcp_send_head(sk)) 2750a2a385d6SEric Dumazet return false; 27512331ccc5SEric Dumazet /* Some heuristics for collapsing over SACK'd could be invented */ 27524a17fc3aSIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 2753a2a385d6SEric Dumazet return false; 27544a17fc3aSIlpo Järvinen 2755a2a385d6SEric Dumazet return true; 27564a17fc3aSIlpo Järvinen } 27574a17fc3aSIlpo Järvinen 275867edfef7SAndi Kleen /* Collapse packets in the retransmit queue to make to create 275967edfef7SAndi Kleen * less packets on the wire. This is only done on retransmission. 276067edfef7SAndi Kleen */ 27614a17fc3aSIlpo Järvinen static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, 27624a17fc3aSIlpo Järvinen int space) 27634a17fc3aSIlpo Järvinen { 27644a17fc3aSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 27654a17fc3aSIlpo Järvinen struct sk_buff *skb = to, *tmp; 2766a2a385d6SEric Dumazet bool first = true; 27674a17fc3aSIlpo Järvinen 27684a17fc3aSIlpo Järvinen if (!sysctl_tcp_retrans_collapse) 27694a17fc3aSIlpo Järvinen return; 27704de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 27714a17fc3aSIlpo Järvinen return; 27724a17fc3aSIlpo Järvinen 27734a17fc3aSIlpo Järvinen tcp_for_write_queue_from_safe(skb, tmp, sk) { 27744a17fc3aSIlpo Järvinen if (!tcp_can_collapse(sk, skb)) 27754a17fc3aSIlpo Järvinen break; 27764a17fc3aSIlpo Järvinen 2777a643b5d4SMartin KaFai Lau if (!tcp_skb_can_collapse_to(to)) 2778a643b5d4SMartin KaFai Lau break; 2779a643b5d4SMartin KaFai Lau 27804a17fc3aSIlpo Järvinen space -= skb->len; 27814a17fc3aSIlpo Järvinen 27824a17fc3aSIlpo Järvinen if (first) { 2783a2a385d6SEric Dumazet first = false; 27844a17fc3aSIlpo Järvinen continue; 27854a17fc3aSIlpo Järvinen } 27864a17fc3aSIlpo Järvinen 27874a17fc3aSIlpo Järvinen if (space < 0) 27884a17fc3aSIlpo Järvinen break; 27894a17fc3aSIlpo Järvinen 27904a17fc3aSIlpo Järvinen if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) 27914a17fc3aSIlpo Järvinen break; 27924a17fc3aSIlpo Järvinen 2793f8071cdeSEric Dumazet if (!tcp_collapse_retrans(sk, to)) 2794f8071cdeSEric Dumazet break; 27954a17fc3aSIlpo Järvinen } 27964a17fc3aSIlpo Järvinen } 27974a17fc3aSIlpo Järvinen 27981da177e4SLinus Torvalds /* This retransmits one SKB. Policy decisions and retransmit queue 27991da177e4SLinus Torvalds * state updates are done by the caller. Returns non-zero if an 28001da177e4SLinus Torvalds * error occurred which prevented the send. 28011da177e4SLinus Torvalds */ 280210d3be56SEric Dumazet int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) 28031da177e4SLinus Torvalds { 28045d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 280510d3be56SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 28067d227cd2SSridhar Samudrala unsigned int cur_mss; 280710d3be56SEric Dumazet int diff, len, err; 28081da177e4SLinus Torvalds 280910d3be56SEric Dumazet 281010d3be56SEric Dumazet /* Inconclusive MTU probe */ 281110d3be56SEric Dumazet if (icsk->icsk_mtup.probe_size) 28125d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 28135d424d5aSJohn Heffner 28141da177e4SLinus Torvalds /* Do not sent more than we queued. 1/4 is reserved for possible 2815caa20d9aSStephen Hemminger * copying overhead: fragmentation, tunneling, mangling etc. 28161da177e4SLinus Torvalds */ 28171da177e4SLinus Torvalds if (atomic_read(&sk->sk_wmem_alloc) > 2818ffb4d6c8SEric Dumazet min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), 2819ffb4d6c8SEric Dumazet sk->sk_sndbuf)) 28201da177e4SLinus Torvalds return -EAGAIN; 28211da177e4SLinus Torvalds 28221f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 28231f3279aeSEric Dumazet return -EBUSY; 28241f3279aeSEric Dumazet 28251da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 28261da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 28271da177e4SLinus Torvalds BUG(); 28281da177e4SLinus Torvalds if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 28291da177e4SLinus Torvalds return -ENOMEM; 28301da177e4SLinus Torvalds } 28311da177e4SLinus Torvalds 28327d227cd2SSridhar Samudrala if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 28337d227cd2SSridhar Samudrala return -EHOSTUNREACH; /* Routing failure or similar. */ 28347d227cd2SSridhar Samudrala 28350c54b85fSIlpo Järvinen cur_mss = tcp_current_mss(sk); 28367d227cd2SSridhar Samudrala 28371da177e4SLinus Torvalds /* If receiver has shrunk his window, and skb is out of 28381da177e4SLinus Torvalds * new window, do not retransmit it. The exception is the 28391da177e4SLinus Torvalds * case, when window is shrunk to zero. In this case 28401da177e4SLinus Torvalds * our retransmit serves as a zero window probe. 28411da177e4SLinus Torvalds */ 28429d4fb27dSJoe Perches if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && 28439d4fb27dSJoe Perches TCP_SKB_CB(skb)->seq != tp->snd_una) 28441da177e4SLinus Torvalds return -EAGAIN; 28451da177e4SLinus Torvalds 284610d3be56SEric Dumazet len = cur_mss * segs; 284710d3be56SEric Dumazet if (skb->len > len) { 284810d3be56SEric Dumazet if (tcp_fragment(sk, skb, len, cur_mss, GFP_ATOMIC)) 28491da177e4SLinus Torvalds return -ENOMEM; /* We'll try again later. */ 285002276f3cSIlpo Järvinen } else { 2851c52e2421SEric Dumazet if (skb_unclone(skb, GFP_ATOMIC)) 2852c52e2421SEric Dumazet return -ENOMEM; 285310d3be56SEric Dumazet 285410d3be56SEric Dumazet diff = tcp_skb_pcount(skb); 285510d3be56SEric Dumazet tcp_set_skb_tso_segs(skb, cur_mss); 285610d3be56SEric Dumazet diff -= tcp_skb_pcount(skb); 285710d3be56SEric Dumazet if (diff) 285810d3be56SEric Dumazet tcp_adjust_pcount(sk, skb, diff); 285910d3be56SEric Dumazet if (skb->len < cur_mss) 286010d3be56SEric Dumazet tcp_retrans_try_collapse(sk, skb, cur_mss); 28611da177e4SLinus Torvalds } 28621da177e4SLinus Torvalds 286349213555SDaniel Borkmann /* RFC3168, section 6.1.1.1. ECN fallback */ 286449213555SDaniel Borkmann if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN) 286549213555SDaniel Borkmann tcp_ecn_clear_syn(sk, skb); 286649213555SDaniel Borkmann 2867678550c6SYuchung Cheng /* Update global and local TCP statistics. */ 2868678550c6SYuchung Cheng segs = tcp_skb_pcount(skb); 2869678550c6SYuchung Cheng TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs); 2870678550c6SYuchung Cheng if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 2871678550c6SYuchung Cheng __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 2872678550c6SYuchung Cheng tp->total_retrans += segs; 2873678550c6SYuchung Cheng 287450bceae9SThomas Graf /* make sure skb->data is aligned on arches that require it 287550bceae9SThomas Graf * and check if ack-trimming & collapsing extended the headroom 287650bceae9SThomas Graf * beyond what csum_start can cover. 287750bceae9SThomas Graf */ 287850bceae9SThomas Graf if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) || 287950bceae9SThomas Graf skb_headroom(skb) >= 0xFFFF)) { 288010a81980SEric Dumazet struct sk_buff *nskb; 288110a81980SEric Dumazet 2882385e2070SEric Dumazet skb->skb_mstamp = tp->tcp_mstamp; 288310a81980SEric Dumazet nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC); 2884c84a5711SYuchung Cheng err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) : 2885117632e6SEric Dumazet -ENOBUFS; 2886117632e6SEric Dumazet } else { 2887c84a5711SYuchung Cheng err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 2888117632e6SEric Dumazet } 2889c84a5711SYuchung Cheng 2890fc9f3501SEric Dumazet if (likely(!err)) { 2891c84a5711SYuchung Cheng TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS; 2892678550c6SYuchung Cheng } else if (err != -EBUSY) { 2893678550c6SYuchung Cheng NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL); 2894fc9f3501SEric Dumazet } 2895c84a5711SYuchung Cheng return err; 289693b174adSYuchung Cheng } 289793b174adSYuchung Cheng 289810d3be56SEric Dumazet int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) 289993b174adSYuchung Cheng { 290093b174adSYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 290110d3be56SEric Dumazet int err = __tcp_retransmit_skb(sk, skb, segs); 29021da177e4SLinus Torvalds 29031da177e4SLinus Torvalds if (err == 0) { 29041da177e4SLinus Torvalds #if FASTRETRANS_DEBUG > 0 29051da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 2906e87cc472SJoe Perches net_dbg_ratelimited("retrans_out leaked\n"); 29071da177e4SLinus Torvalds } 29081da177e4SLinus Torvalds #endif 29091da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 29101da177e4SLinus Torvalds tp->retrans_out += tcp_skb_pcount(skb); 29111da177e4SLinus Torvalds 29121da177e4SLinus Torvalds /* Save stamp of the first retransmit. */ 29131da177e4SLinus Torvalds if (!tp->retrans_stamp) 29147faee5c0SEric Dumazet tp->retrans_stamp = tcp_skb_timestamp(skb); 29151da177e4SLinus Torvalds 29161da177e4SLinus Torvalds } 29176e08d5e3SYuchung Cheng 29186e08d5e3SYuchung Cheng if (tp->undo_retrans < 0) 29196e08d5e3SYuchung Cheng tp->undo_retrans = 0; 29206e08d5e3SYuchung Cheng tp->undo_retrans += tcp_skb_pcount(skb); 29211da177e4SLinus Torvalds return err; 29221da177e4SLinus Torvalds } 29231da177e4SLinus Torvalds 29241da177e4SLinus Torvalds /* This gets called after a retransmit timeout, and the initially 29251da177e4SLinus Torvalds * retransmitted data is acknowledged. It tries to continue 29261da177e4SLinus Torvalds * resending the rest of the retransmit queue, until either 29271da177e4SLinus Torvalds * we've sent it all or the congestion window limit is reached. 29281da177e4SLinus Torvalds * If doing SACK, the first ACK which comes back for a timeout 29291da177e4SLinus Torvalds * based retransmit packet might feed us FACK information again. 29301da177e4SLinus Torvalds * If so, we use it to avoid unnecessarily retransmissions. 29311da177e4SLinus Torvalds */ 29321da177e4SLinus Torvalds void tcp_xmit_retransmit_queue(struct sock *sk) 29331da177e4SLinus Torvalds { 29346687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 29351da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 29361da177e4SLinus Torvalds struct sk_buff *skb; 29370e1c54c2SIlpo Järvinen struct sk_buff *hole = NULL; 2938840a3cbeSYuchung Cheng u32 max_segs; 293961eb55f4SIlpo Järvinen int mib_idx; 29406a438bbeSStephen Hemminger 294145e77d31SIlpo Järvinen if (!tp->packets_out) 294245e77d31SIlpo Järvinen return; 294345e77d31SIlpo Järvinen 2944618d9f25SIlpo Järvinen if (tp->retransmit_skb_hint) { 29456a438bbeSStephen Hemminger skb = tp->retransmit_skb_hint; 2946618d9f25SIlpo Järvinen } else { 2947fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 2948618d9f25SIlpo Järvinen } 29491da177e4SLinus Torvalds 2950ed6e7268SNeal Cardwell max_segs = tcp_tso_segs(sk, tcp_current_mss(sk)); 2951fe067e8aSDavid S. Miller tcp_for_write_queue_from(skb, sk) { 2952dca0aaf8SEric Dumazet __u8 sacked; 295310d3be56SEric Dumazet int segs; 29541da177e4SLinus Torvalds 2955fe067e8aSDavid S. Miller if (skb == tcp_send_head(sk)) 2956fe067e8aSDavid S. Miller break; 2957218af599SEric Dumazet 2958218af599SEric Dumazet if (tcp_pacing_check(sk)) 2959218af599SEric Dumazet break; 2960218af599SEric Dumazet 29616a438bbeSStephen Hemminger /* we could do better than to assign each time */ 296251456b29SIan Morris if (!hole) 29636a438bbeSStephen Hemminger tp->retransmit_skb_hint = skb; 29646a438bbeSStephen Hemminger 296510d3be56SEric Dumazet segs = tp->snd_cwnd - tcp_packets_in_flight(tp); 296610d3be56SEric Dumazet if (segs <= 0) 29671da177e4SLinus Torvalds return; 2968dca0aaf8SEric Dumazet sacked = TCP_SKB_CB(skb)->sacked; 2969a3d2e9f8SEric Dumazet /* In case tcp_shift_skb_data() have aggregated large skbs, 2970a3d2e9f8SEric Dumazet * we need to make sure not sending too bigs TSO packets 2971a3d2e9f8SEric Dumazet */ 2972a3d2e9f8SEric Dumazet segs = min_t(int, segs, max_segs); 29730e1c54c2SIlpo Järvinen 2974840a3cbeSYuchung Cheng if (tp->retrans_out >= tp->lost_out) { 2975006f582cSIlpo Järvinen break; 29760e1c54c2SIlpo Järvinen } else if (!(sacked & TCPCB_LOST)) { 297751456b29SIan Morris if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) 29780e1c54c2SIlpo Järvinen hole = skb; 297961eb55f4SIlpo Järvinen continue; 29801da177e4SLinus Torvalds 29810e1c54c2SIlpo Järvinen } else { 29820e1c54c2SIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Loss) 29830e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFASTRETRANS; 29840e1c54c2SIlpo Järvinen else 29850e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; 29860e1c54c2SIlpo Järvinen } 29870e1c54c2SIlpo Järvinen 29880e1c54c2SIlpo Järvinen if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) 298961eb55f4SIlpo Järvinen continue; 299040b215e5SPavel Emelyanov 2991f9616c35SEric Dumazet if (tcp_small_queue_check(sk, skb, 1)) 2992f9616c35SEric Dumazet return; 2993f9616c35SEric Dumazet 299410d3be56SEric Dumazet if (tcp_retransmit_skb(sk, skb, segs)) 29951da177e4SLinus Torvalds return; 299624ab6becSYuchung Cheng 2997de1d6578SYuchung Cheng NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb)); 29981da177e4SLinus Torvalds 2999684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 3000a262f0cdSNandita Dukkipati tp->prr_out += tcp_skb_pcount(skb); 3001a262f0cdSNandita Dukkipati 300257dde7f7SYuchung Cheng if (skb == tcp_write_queue_head(sk) && 300357dde7f7SYuchung Cheng icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT) 3004463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 30053f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, 30063f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 30071da177e4SLinus Torvalds } 30081da177e4SLinus Torvalds } 30091da177e4SLinus Torvalds 3010d83769a5SEric Dumazet /* We allow to exceed memory limits for FIN packets to expedite 3011d83769a5SEric Dumazet * connection tear down and (memory) recovery. 3012845704a5SEric Dumazet * Otherwise tcp_send_fin() could be tempted to either delay FIN 3013845704a5SEric Dumazet * or even be forced to close flow without any FIN. 3014a6c5ea4cSEric Dumazet * In general, we want to allow one skb per socket to avoid hangs 3015a6c5ea4cSEric Dumazet * with edge trigger epoll() 3016d83769a5SEric Dumazet */ 3017a6c5ea4cSEric Dumazet void sk_forced_mem_schedule(struct sock *sk, int size) 3018d83769a5SEric Dumazet { 3019e805605cSJohannes Weiner int amt; 3020d83769a5SEric Dumazet 3021d83769a5SEric Dumazet if (size <= sk->sk_forward_alloc) 3022d83769a5SEric Dumazet return; 3023d83769a5SEric Dumazet amt = sk_mem_pages(size); 3024d83769a5SEric Dumazet sk->sk_forward_alloc += amt * SK_MEM_QUANTUM; 3025e805605cSJohannes Weiner sk_memory_allocated_add(sk, amt); 3026e805605cSJohannes Weiner 3027baac50bbSJohannes Weiner if (mem_cgroup_sockets_enabled && sk->sk_memcg) 3028baac50bbSJohannes Weiner mem_cgroup_charge_skmem(sk->sk_memcg, amt); 3029d83769a5SEric Dumazet } 3030d83769a5SEric Dumazet 3031845704a5SEric Dumazet /* Send a FIN. The caller locks the socket for us. 3032845704a5SEric Dumazet * We should try to send a FIN packet really hard, but eventually give up. 30331da177e4SLinus Torvalds */ 30341da177e4SLinus Torvalds void tcp_send_fin(struct sock *sk) 30351da177e4SLinus Torvalds { 3036845704a5SEric Dumazet struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk); 30371da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 30381da177e4SLinus Torvalds 3039845704a5SEric Dumazet /* Optimization, tack on the FIN if we have one skb in write queue and 3040845704a5SEric Dumazet * this skb was not yet sent, or we are under memory pressure. 3041845704a5SEric Dumazet * Note: in the latter case, FIN packet will be sent after a timeout, 3042845704a5SEric Dumazet * as TCP stack thinks it has already been transmitted. 30431da177e4SLinus Torvalds */ 3044b8da51ebSEric Dumazet if (tskb && (tcp_send_head(sk) || tcp_under_memory_pressure(sk))) { 3045845704a5SEric Dumazet coalesce: 3046845704a5SEric Dumazet TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN; 3047845704a5SEric Dumazet TCP_SKB_CB(tskb)->end_seq++; 30481da177e4SLinus Torvalds tp->write_seq++; 3049845704a5SEric Dumazet if (!tcp_send_head(sk)) { 3050845704a5SEric Dumazet /* This means tskb was already sent. 3051845704a5SEric Dumazet * Pretend we included the FIN on previous transmit. 3052845704a5SEric Dumazet * We need to set tp->snd_nxt to the value it would have 3053845704a5SEric Dumazet * if FIN had been sent. This is because retransmit path 3054845704a5SEric Dumazet * does not change tp->snd_nxt. 3055845704a5SEric Dumazet */ 3056845704a5SEric Dumazet tp->snd_nxt++; 3057845704a5SEric Dumazet return; 3058845704a5SEric Dumazet } 30591da177e4SLinus Torvalds } else { 3060845704a5SEric Dumazet skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation); 3061845704a5SEric Dumazet if (unlikely(!skb)) { 3062845704a5SEric Dumazet if (tskb) 3063845704a5SEric Dumazet goto coalesce; 3064845704a5SEric Dumazet return; 30651da177e4SLinus Torvalds } 3066d83769a5SEric Dumazet skb_reserve(skb, MAX_TCP_HEADER); 3067a6c5ea4cSEric Dumazet sk_forced_mem_schedule(sk, skb->truesize); 30681da177e4SLinus Torvalds /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 3069e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tp->write_seq, 3070a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_FIN); 30711da177e4SLinus Torvalds tcp_queue_skb(sk, skb); 30721da177e4SLinus Torvalds } 3073845704a5SEric Dumazet __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF); 30741da177e4SLinus Torvalds } 30751da177e4SLinus Torvalds 30761da177e4SLinus Torvalds /* We get here when a process closes a file descriptor (either due to 30771da177e4SLinus Torvalds * an explicit close() or as a byproduct of exit()'ing) and there 30781da177e4SLinus Torvalds * was unread data in the receive queue. This behavior is recommended 307965bb723cSGerrit Renker * by RFC 2525, section 2.17. -DaveM 30801da177e4SLinus Torvalds */ 3081dd0fc66fSAl Viro void tcp_send_active_reset(struct sock *sk, gfp_t priority) 30821da177e4SLinus Torvalds { 30831da177e4SLinus Torvalds struct sk_buff *skb; 30841da177e4SLinus Torvalds 30857cc2b043SGao Feng TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); 30867cc2b043SGao Feng 30871da177e4SLinus Torvalds /* NOTE: No TCP options attached and we never retransmit this. */ 30881da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, priority); 30891da177e4SLinus Torvalds if (!skb) { 30904e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 30911da177e4SLinus Torvalds return; 30921da177e4SLinus Torvalds } 30931da177e4SLinus Torvalds 30941da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 30951da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 3096e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), 3097a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_RST); 3098385e2070SEric Dumazet skb_mstamp_get(&tcp_sk(sk)->tcp_mstamp); 30991da177e4SLinus Torvalds /* Send it off. */ 3100dfb4b9dcSDavid S. Miller if (tcp_transmit_skb(sk, skb, 0, priority)) 31014e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 31021da177e4SLinus Torvalds } 31031da177e4SLinus Torvalds 310467edfef7SAndi Kleen /* Send a crossed SYN-ACK during socket establishment. 310567edfef7SAndi Kleen * WARNING: This routine must only be called when we have already sent 31061da177e4SLinus Torvalds * a SYN packet that crossed the incoming SYN that caused this routine 31071da177e4SLinus Torvalds * to get called. If this assumption fails then the initial rcv_wnd 31081da177e4SLinus Torvalds * and rcv_wscale values will not be correct. 31091da177e4SLinus Torvalds */ 31101da177e4SLinus Torvalds int tcp_send_synack(struct sock *sk) 31111da177e4SLinus Torvalds { 31121da177e4SLinus Torvalds struct sk_buff *skb; 31131da177e4SLinus Torvalds 3114fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 311551456b29SIan Morris if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 311691df42beSJoe Perches pr_debug("%s: wrong queue state\n", __func__); 31171da177e4SLinus Torvalds return -EFAULT; 31181da177e4SLinus Torvalds } 31194de075e0SEric Dumazet if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { 31201da177e4SLinus Torvalds if (skb_cloned(skb)) { 31211da177e4SLinus Torvalds struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC); 312251456b29SIan Morris if (!nskb) 31231da177e4SLinus Torvalds return -ENOMEM; 3124fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 3125f4a775d1SEric Dumazet __skb_header_release(nskb); 3126fe067e8aSDavid S. Miller __tcp_add_write_queue_head(sk, nskb); 31273ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 31283ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 31293ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 31301da177e4SLinus Torvalds skb = nskb; 31311da177e4SLinus Torvalds } 31321da177e4SLinus Torvalds 31334de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; 3134735d3831SFlorian Westphal tcp_ecn_send_synack(sk, skb); 31351da177e4SLinus Torvalds } 3136dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 31371da177e4SLinus Torvalds } 31381da177e4SLinus Torvalds 31394aea39c1SEric Dumazet /** 31404aea39c1SEric Dumazet * tcp_make_synack - Prepare a SYN-ACK. 31414aea39c1SEric Dumazet * sk: listener socket 31424aea39c1SEric Dumazet * dst: dst entry attached to the SYNACK 31434aea39c1SEric Dumazet * req: request_sock pointer 31444aea39c1SEric Dumazet * 31454aea39c1SEric Dumazet * Allocate one skb and build a SYNACK packet. 31464aea39c1SEric Dumazet * @dst is consumed : Caller should not use it again. 31474aea39c1SEric Dumazet */ 31485d062de7SEric Dumazet struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst, 3149e6b4d113SWilliam Allen Simpson struct request_sock *req, 3150ca6fb065SEric Dumazet struct tcp_fastopen_cookie *foc, 3151b3d05147SEric Dumazet enum tcp_synack_type synack_type) 31521da177e4SLinus Torvalds { 31532e6599cbSArnaldo Carvalho de Melo struct inet_request_sock *ireq = inet_rsk(req); 31545d062de7SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 315580f03e27SEric Dumazet struct tcp_md5sig_key *md5 = NULL; 31565d062de7SEric Dumazet struct tcp_out_options opts; 31575d062de7SEric Dumazet struct sk_buff *skb; 3158bd0388aeSWilliam Allen Simpson int tcp_header_size; 31595d062de7SEric Dumazet struct tcphdr *th; 3160f5fff5dcSTom Quetchenbach int mss; 31611da177e4SLinus Torvalds 3162ca6fb065SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 31634aea39c1SEric Dumazet if (unlikely(!skb)) { 31644aea39c1SEric Dumazet dst_release(dst); 31651da177e4SLinus Torvalds return NULL; 31664aea39c1SEric Dumazet } 31671da177e4SLinus Torvalds /* Reserve space for headers. */ 31681da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 31691da177e4SLinus Torvalds 3170b3d05147SEric Dumazet switch (synack_type) { 3171b3d05147SEric Dumazet case TCP_SYNACK_NORMAL: 31729e17f8a4SEric Dumazet skb_set_owner_w(skb, req_to_sk(req)); 3173b3d05147SEric Dumazet break; 3174b3d05147SEric Dumazet case TCP_SYNACK_COOKIE: 3175b3d05147SEric Dumazet /* Under synflood, we do not attach skb to a socket, 3176b3d05147SEric Dumazet * to avoid false sharing. 3177b3d05147SEric Dumazet */ 3178b3d05147SEric Dumazet break; 3179b3d05147SEric Dumazet case TCP_SYNACK_FASTOPEN: 3180ca6fb065SEric Dumazet /* sk is a const pointer, because we want to express multiple 3181ca6fb065SEric Dumazet * cpu might call us concurrently. 3182ca6fb065SEric Dumazet * sk->sk_wmem_alloc in an atomic, we can promote to rw. 3183ca6fb065SEric Dumazet */ 3184ca6fb065SEric Dumazet skb_set_owner_w(skb, (struct sock *)sk); 3185b3d05147SEric Dumazet break; 3186ca6fb065SEric Dumazet } 31874aea39c1SEric Dumazet skb_dst_set(skb, dst); 31881da177e4SLinus Torvalds 31893541f9e8SEric Dumazet mss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 3190f5fff5dcSTom Quetchenbach 319133ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 31928b5f12d0SFlorian Westphal #ifdef CONFIG_SYN_COOKIES 31938b5f12d0SFlorian Westphal if (unlikely(req->cookie_ts)) 31947faee5c0SEric Dumazet skb->skb_mstamp.stamp_jiffies = cookie_init_timestamp(req); 31958b5f12d0SFlorian Westphal else 31968b5f12d0SFlorian Westphal #endif 31977faee5c0SEric Dumazet skb_mstamp_get(&skb->skb_mstamp); 319880f03e27SEric Dumazet 319980f03e27SEric Dumazet #ifdef CONFIG_TCP_MD5SIG 320080f03e27SEric Dumazet rcu_read_lock(); 3201fd3a154aSEric Dumazet md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req)); 320280f03e27SEric Dumazet #endif 320358d607d3SEric Dumazet skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4); 320437bfbddaSEric Dumazet tcp_header_size = tcp_synack_options(req, mss, skb, &opts, md5, foc) + 320537bfbddaSEric Dumazet sizeof(*th); 320633ad798cSAdam Langley 3207aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 3208aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 32091da177e4SLinus Torvalds 3210ea1627c2SEric Dumazet th = (struct tcphdr *)skb->data; 32111da177e4SLinus Torvalds memset(th, 0, sizeof(struct tcphdr)); 32121da177e4SLinus Torvalds th->syn = 1; 32131da177e4SLinus Torvalds th->ack = 1; 32146ac705b1SEric Dumazet tcp_ecn_make_synack(req, th); 3215b44084c2SEric Dumazet th->source = htons(ireq->ir_num); 3216634fb979SEric Dumazet th->dest = ireq->ir_rmt_port; 3217e870a8efSIlpo Järvinen /* Setting of flags are superfluous here for callers (and ECE is 3218e870a8efSIlpo Järvinen * not even correctly set) 3219e870a8efSIlpo Järvinen */ 3220e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn, 3221a3433f35SChangli Gao TCPHDR_SYN | TCPHDR_ACK); 32224957faadSWilliam Allen Simpson 32231da177e4SLinus Torvalds th->seq = htonl(TCP_SKB_CB(skb)->seq); 32248336886fSJerry Chu /* XXX data is queued and acked as is. No buffer/window check */ 32258336886fSJerry Chu th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt); 32261da177e4SLinus Torvalds 32271da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 3228ed53d0abSEric Dumazet th->window = htons(min(req->rsk_rcv_wnd, 65535U)); 32295d062de7SEric Dumazet tcp_options_write((__be32 *)(th + 1), NULL, &opts); 32301da177e4SLinus Torvalds th->doff = (tcp_header_size >> 2); 323190bbcc60SEric Dumazet __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS); 3232cfb6eeb4SYOSHIFUJI Hideaki 3233cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 3234cfb6eeb4SYOSHIFUJI Hideaki /* Okay, we have all we need - do the md5 hash if needed */ 323580f03e27SEric Dumazet if (md5) 3236bd0388aeSWilliam Allen Simpson tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, 323739f8e58eSEric Dumazet md5, req_to_sk(req), skb); 323880f03e27SEric Dumazet rcu_read_unlock(); 3239cfb6eeb4SYOSHIFUJI Hideaki #endif 3240cfb6eeb4SYOSHIFUJI Hideaki 3241b50edd78SEric Dumazet /* Do not fool tcpdump (if any), clean our debris */ 32422456e855SThomas Gleixner skb->tstamp = 0; 32431da177e4SLinus Torvalds return skb; 32441da177e4SLinus Torvalds } 32454bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_make_synack); 32461da177e4SLinus Torvalds 324781164413SDaniel Borkmann static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst) 324881164413SDaniel Borkmann { 324981164413SDaniel Borkmann struct inet_connection_sock *icsk = inet_csk(sk); 325081164413SDaniel Borkmann const struct tcp_congestion_ops *ca; 325181164413SDaniel Borkmann u32 ca_key = dst_metric(dst, RTAX_CC_ALGO); 325281164413SDaniel Borkmann 325381164413SDaniel Borkmann if (ca_key == TCP_CA_UNSPEC) 325481164413SDaniel Borkmann return; 325581164413SDaniel Borkmann 325681164413SDaniel Borkmann rcu_read_lock(); 325781164413SDaniel Borkmann ca = tcp_ca_find_key(ca_key); 325881164413SDaniel Borkmann if (likely(ca && try_module_get(ca->owner))) { 325981164413SDaniel Borkmann module_put(icsk->icsk_ca_ops->owner); 326081164413SDaniel Borkmann icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst); 326181164413SDaniel Borkmann icsk->icsk_ca_ops = ca; 326281164413SDaniel Borkmann } 326381164413SDaniel Borkmann rcu_read_unlock(); 326481164413SDaniel Borkmann } 326581164413SDaniel Borkmann 326667edfef7SAndi Kleen /* Do all connect socket setups that can be done AF independent. */ 3267f7e56a76Sstephen hemminger static void tcp_connect_init(struct sock *sk) 32681da177e4SLinus Torvalds { 3269cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 32701da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 32711da177e4SLinus Torvalds __u8 rcv_wscale; 32721da177e4SLinus Torvalds 32731da177e4SLinus Torvalds /* We'll fix this up when we get a response from the other end. 32741da177e4SLinus Torvalds * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 32751da177e4SLinus Torvalds */ 32761da177e4SLinus Torvalds tp->tcp_header_len = sizeof(struct tcphdr) + 3277bb5b7c11SDavid S. Miller (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0); 32781da177e4SLinus Torvalds 3279cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 328000db4124SIan Morris if (tp->af_specific->md5_lookup(sk, sk)) 3281cfb6eeb4SYOSHIFUJI Hideaki tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 3282cfb6eeb4SYOSHIFUJI Hideaki #endif 3283cfb6eeb4SYOSHIFUJI Hideaki 32841da177e4SLinus Torvalds /* If user gave his TCP_MAXSEG, record it to clamp */ 32851da177e4SLinus Torvalds if (tp->rx_opt.user_mss) 32861da177e4SLinus Torvalds tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 32871da177e4SLinus Torvalds tp->max_window = 0; 32885d424d5aSJohn Heffner tcp_mtup_init(sk); 32891da177e4SLinus Torvalds tcp_sync_mss(sk, dst_mtu(dst)); 32901da177e4SLinus Torvalds 329181164413SDaniel Borkmann tcp_ca_dst_init(sk, dst); 329281164413SDaniel Borkmann 32931da177e4SLinus Torvalds if (!tp->window_clamp) 32941da177e4SLinus Torvalds tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 32953541f9e8SEric Dumazet tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 3296f5fff5dcSTom Quetchenbach 32971da177e4SLinus Torvalds tcp_initialize_rcv_mss(sk); 32981da177e4SLinus Torvalds 3299e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 3300e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 3301e88c64f0SHagen Paul Pfeifer (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) 3302e88c64f0SHagen Paul Pfeifer tp->window_clamp = tcp_full_space(sk); 3303e88c64f0SHagen Paul Pfeifer 33041da177e4SLinus Torvalds tcp_select_initial_window(tcp_full_space(sk), 33051da177e4SLinus Torvalds tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 33061da177e4SLinus Torvalds &tp->rcv_wnd, 33071da177e4SLinus Torvalds &tp->window_clamp, 3308bb5b7c11SDavid S. Miller sysctl_tcp_window_scaling, 330931d12926Slaurent chavey &rcv_wscale, 331031d12926Slaurent chavey dst_metric(dst, RTAX_INITRWND)); 33111da177e4SLinus Torvalds 33121da177e4SLinus Torvalds tp->rx_opt.rcv_wscale = rcv_wscale; 33131da177e4SLinus Torvalds tp->rcv_ssthresh = tp->rcv_wnd; 33141da177e4SLinus Torvalds 33151da177e4SLinus Torvalds sk->sk_err = 0; 33161da177e4SLinus Torvalds sock_reset_flag(sk, SOCK_DONE); 33171da177e4SLinus Torvalds tp->snd_wnd = 0; 3318ee7537b6SHantzis Fotis tcp_init_wl(tp, 0); 33191da177e4SLinus Torvalds tp->snd_una = tp->write_seq; 33201da177e4SLinus Torvalds tp->snd_sml = tp->write_seq; 332133f5f57eSIlpo Järvinen tp->snd_up = tp->write_seq; 3322370816aeSPavel Emelyanov tp->snd_nxt = tp->write_seq; 3323ee995283SPavel Emelyanov 3324ee995283SPavel Emelyanov if (likely(!tp->repair)) 33251da177e4SLinus Torvalds tp->rcv_nxt = 0; 3326c7781a6eSAndrew Vagin else 3327*70eabf0eSEric Dumazet tp->rcv_tstamp = tcp_jiffies32; 3328ee995283SPavel Emelyanov tp->rcv_wup = tp->rcv_nxt; 3329ee995283SPavel Emelyanov tp->copied_seq = tp->rcv_nxt; 33301da177e4SLinus Torvalds 3331463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; 3332463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_retransmits = 0; 33331da177e4SLinus Torvalds tcp_clear_retrans(tp); 33341da177e4SLinus Torvalds } 33351da177e4SLinus Torvalds 3336783237e8SYuchung Cheng static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) 3337783237e8SYuchung Cheng { 3338783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3339783237e8SYuchung Cheng struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 3340783237e8SYuchung Cheng 3341783237e8SYuchung Cheng tcb->end_seq += skb->len; 3342f4a775d1SEric Dumazet __skb_header_release(skb); 3343783237e8SYuchung Cheng __tcp_add_write_queue_tail(sk, skb); 3344783237e8SYuchung Cheng sk->sk_wmem_queued += skb->truesize; 3345783237e8SYuchung Cheng sk_mem_charge(sk, skb->truesize); 3346783237e8SYuchung Cheng tp->write_seq = tcb->end_seq; 3347783237e8SYuchung Cheng tp->packets_out += tcp_skb_pcount(skb); 3348783237e8SYuchung Cheng } 3349783237e8SYuchung Cheng 3350783237e8SYuchung Cheng /* Build and send a SYN with data and (cached) Fast Open cookie. However, 3351783237e8SYuchung Cheng * queue a data-only packet after the regular SYN, such that regular SYNs 3352783237e8SYuchung Cheng * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges 3353783237e8SYuchung Cheng * only the SYN sequence, the data are retransmitted in the first ACK. 3354783237e8SYuchung Cheng * If cookie is not cached or other error occurs, falls back to send a 3355783237e8SYuchung Cheng * regular SYN with Fast Open cookie request option. 3356783237e8SYuchung Cheng */ 3357783237e8SYuchung Cheng static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) 3358783237e8SYuchung Cheng { 3359783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3360783237e8SYuchung Cheng struct tcp_fastopen_request *fo = tp->fastopen_req; 3361065263f4SWei Wang int space, err = 0; 3362355a901eSEric Dumazet struct sk_buff *syn_data; 3363783237e8SYuchung Cheng 336467da22d2SYuchung Cheng tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ 3365065263f4SWei Wang if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie)) 3366783237e8SYuchung Cheng goto fallback; 3367783237e8SYuchung Cheng 3368783237e8SYuchung Cheng /* MSS for SYN-data is based on cached MSS and bounded by PMTU and 3369783237e8SYuchung Cheng * user-MSS. Reserve maximum option space for middleboxes that add 3370783237e8SYuchung Cheng * private TCP options. The cost is reduced data space in SYN :( 3371783237e8SYuchung Cheng */ 33723541f9e8SEric Dumazet tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp); 33733541f9e8SEric Dumazet 33741b63edd6SYuchung Cheng space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) - 3375783237e8SYuchung Cheng MAX_TCP_OPTION_SPACE; 3376783237e8SYuchung Cheng 3377f5ddcbbbSEric Dumazet space = min_t(size_t, space, fo->size); 3378f5ddcbbbSEric Dumazet 3379f5ddcbbbSEric Dumazet /* limit to order-0 allocations */ 3380f5ddcbbbSEric Dumazet space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER)); 3381f5ddcbbbSEric Dumazet 3382eb934478SEric Dumazet syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false); 3383355a901eSEric Dumazet if (!syn_data) 3384783237e8SYuchung Cheng goto fallback; 3385355a901eSEric Dumazet syn_data->ip_summed = CHECKSUM_PARTIAL; 3386355a901eSEric Dumazet memcpy(syn_data->cb, syn->cb, sizeof(syn->cb)); 338707e100f9SEric Dumazet if (space) { 338807e100f9SEric Dumazet int copied = copy_from_iter(skb_put(syn_data, space), space, 338957be5bdaSAl Viro &fo->data->msg_iter); 339057be5bdaSAl Viro if (unlikely(!copied)) { 3391355a901eSEric Dumazet kfree_skb(syn_data); 3392783237e8SYuchung Cheng goto fallback; 3393783237e8SYuchung Cheng } 339457be5bdaSAl Viro if (copied != space) { 339557be5bdaSAl Viro skb_trim(syn_data, copied); 339657be5bdaSAl Viro space = copied; 339757be5bdaSAl Viro } 339807e100f9SEric Dumazet } 3399355a901eSEric Dumazet /* No more data pending in inet_wait_for_connect() */ 3400355a901eSEric Dumazet if (space == fo->size) 3401355a901eSEric Dumazet fo->data = NULL; 3402355a901eSEric Dumazet fo->copied = space; 3403783237e8SYuchung Cheng 3404355a901eSEric Dumazet tcp_connect_queue_skb(sk, syn_data); 34050f87230dSFrancis Yan if (syn_data->len) 34060f87230dSFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_BUSY); 3407355a901eSEric Dumazet 3408355a901eSEric Dumazet err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation); 3409355a901eSEric Dumazet 3410355a901eSEric Dumazet syn->skb_mstamp = syn_data->skb_mstamp; 3411355a901eSEric Dumazet 3412355a901eSEric Dumazet /* Now full SYN+DATA was cloned and sent (or not), 3413355a901eSEric Dumazet * remove the SYN from the original skb (syn_data) 3414355a901eSEric Dumazet * we keep in write queue in case of a retransmit, as we 3415355a901eSEric Dumazet * also have the SYN packet (with no data) in the same queue. 3416431a9124SEric Dumazet */ 3417355a901eSEric Dumazet TCP_SKB_CB(syn_data)->seq++; 3418355a901eSEric Dumazet TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH; 3419355a901eSEric Dumazet if (!err) { 342067da22d2SYuchung Cheng tp->syn_data = (fo->copied > 0); 3421f19c29e3SYuchung Cheng NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT); 3422783237e8SYuchung Cheng goto done; 3423783237e8SYuchung Cheng } 3424783237e8SYuchung Cheng 3425783237e8SYuchung Cheng fallback: 3426783237e8SYuchung Cheng /* Send a regular SYN with Fast Open cookie request option */ 3427783237e8SYuchung Cheng if (fo->cookie.len > 0) 3428783237e8SYuchung Cheng fo->cookie.len = 0; 3429783237e8SYuchung Cheng err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); 3430783237e8SYuchung Cheng if (err) 3431783237e8SYuchung Cheng tp->syn_fastopen = 0; 3432783237e8SYuchung Cheng done: 3433783237e8SYuchung Cheng fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ 3434783237e8SYuchung Cheng return err; 3435783237e8SYuchung Cheng } 3436783237e8SYuchung Cheng 343767edfef7SAndi Kleen /* Build a SYN and send it off. */ 34381da177e4SLinus Torvalds int tcp_connect(struct sock *sk) 34391da177e4SLinus Torvalds { 34401da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 34411da177e4SLinus Torvalds struct sk_buff *buff; 3442ee586811SEric Paris int err; 34431da177e4SLinus Torvalds 34441da177e4SLinus Torvalds tcp_connect_init(sk); 34451da177e4SLinus Torvalds 34462b916477SAndrey Vagin if (unlikely(tp->repair)) { 34472b916477SAndrey Vagin tcp_finish_connect(sk, NULL); 34482b916477SAndrey Vagin return 0; 34492b916477SAndrey Vagin } 34502b916477SAndrey Vagin 3451eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true); 3452355a901eSEric Dumazet if (unlikely(!buff)) 34531da177e4SLinus Torvalds return -ENOBUFS; 34541da177e4SLinus Torvalds 3455a3433f35SChangli Gao tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); 3456385e2070SEric Dumazet skb_mstamp_get(&tp->tcp_mstamp); 3457385e2070SEric Dumazet tp->retrans_stamp = tp->tcp_mstamp.stamp_jiffies; 3458783237e8SYuchung Cheng tcp_connect_queue_skb(sk, buff); 3459735d3831SFlorian Westphal tcp_ecn_send_syn(sk, buff); 34601da177e4SLinus Torvalds 3461783237e8SYuchung Cheng /* Send off SYN; include data in Fast Open. */ 3462783237e8SYuchung Cheng err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : 3463783237e8SYuchung Cheng tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); 3464ee586811SEric Paris if (err == -ECONNREFUSED) 3465ee586811SEric Paris return err; 3466bd37a088SWei Yongjun 3467bd37a088SWei Yongjun /* We change tp->snd_nxt after the tcp_transmit_skb() call 3468bd37a088SWei Yongjun * in order to make this packet get counted in tcpOutSegs. 3469bd37a088SWei Yongjun */ 3470bd37a088SWei Yongjun tp->snd_nxt = tp->write_seq; 3471bd37a088SWei Yongjun tp->pushed_seq = tp->write_seq; 347281cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); 34731da177e4SLinus Torvalds 34741da177e4SLinus Torvalds /* Timer for repeating the SYN until an answer. */ 34753f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 34763f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 34771da177e4SLinus Torvalds return 0; 34781da177e4SLinus Torvalds } 34794bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_connect); 34801da177e4SLinus Torvalds 34811da177e4SLinus Torvalds /* Send out a delayed ack, the caller does the policy checking 34821da177e4SLinus Torvalds * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 34831da177e4SLinus Torvalds * for details. 34841da177e4SLinus Torvalds */ 34851da177e4SLinus Torvalds void tcp_send_delayed_ack(struct sock *sk) 34861da177e4SLinus Torvalds { 3487463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 3488463c84b9SArnaldo Carvalho de Melo int ato = icsk->icsk_ack.ato; 34891da177e4SLinus Torvalds unsigned long timeout; 34901da177e4SLinus Torvalds 34919890092eSFlorian Westphal tcp_ca_event(sk, CA_EVENT_DELAYED_ACK); 34929890092eSFlorian Westphal 34931da177e4SLinus Torvalds if (ato > TCP_DELACK_MIN) { 3494463c84b9SArnaldo Carvalho de Melo const struct tcp_sock *tp = tcp_sk(sk); 34951da177e4SLinus Torvalds int max_ato = HZ / 2; 34961da177e4SLinus Torvalds 3497056834d9SIlpo Järvinen if (icsk->icsk_ack.pingpong || 3498056834d9SIlpo Järvinen (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 34991da177e4SLinus Torvalds max_ato = TCP_DELACK_MAX; 35001da177e4SLinus Torvalds 35011da177e4SLinus Torvalds /* Slow path, intersegment interval is "high". */ 35021da177e4SLinus Torvalds 35031da177e4SLinus Torvalds /* If some rtt estimate is known, use it to bound delayed ack. 3504463c84b9SArnaldo Carvalho de Melo * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 35051da177e4SLinus Torvalds * directly. 35061da177e4SLinus Torvalds */ 3507740b0f18SEric Dumazet if (tp->srtt_us) { 3508740b0f18SEric Dumazet int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3), 3509740b0f18SEric Dumazet TCP_DELACK_MIN); 35101da177e4SLinus Torvalds 35111da177e4SLinus Torvalds if (rtt < max_ato) 35121da177e4SLinus Torvalds max_ato = rtt; 35131da177e4SLinus Torvalds } 35141da177e4SLinus Torvalds 35151da177e4SLinus Torvalds ato = min(ato, max_ato); 35161da177e4SLinus Torvalds } 35171da177e4SLinus Torvalds 35181da177e4SLinus Torvalds /* Stay within the limit we were given */ 35191da177e4SLinus Torvalds timeout = jiffies + ato; 35201da177e4SLinus Torvalds 35211da177e4SLinus Torvalds /* Use new timeout only if there wasn't a older one earlier. */ 3522463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 35231da177e4SLinus Torvalds /* If delack timer was blocked or is about to expire, 35241da177e4SLinus Torvalds * send ACK now. 35251da177e4SLinus Torvalds */ 3526463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.blocked || 3527463c84b9SArnaldo Carvalho de Melo time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 35281da177e4SLinus Torvalds tcp_send_ack(sk); 35291da177e4SLinus Torvalds return; 35301da177e4SLinus Torvalds } 35311da177e4SLinus Torvalds 3532463c84b9SArnaldo Carvalho de Melo if (!time_before(timeout, icsk->icsk_ack.timeout)) 3533463c84b9SArnaldo Carvalho de Melo timeout = icsk->icsk_ack.timeout; 35341da177e4SLinus Torvalds } 3535463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 3536463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.timeout = timeout; 3537463c84b9SArnaldo Carvalho de Melo sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 35381da177e4SLinus Torvalds } 35391da177e4SLinus Torvalds 35401da177e4SLinus Torvalds /* This routine sends an ack and also updates the window. */ 35411da177e4SLinus Torvalds void tcp_send_ack(struct sock *sk) 35421da177e4SLinus Torvalds { 35431da177e4SLinus Torvalds struct sk_buff *buff; 35441da177e4SLinus Torvalds 3545058dc334SIlpo Järvinen /* If we have been reset, we may not send again. */ 3546058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3547058dc334SIlpo Järvinen return; 3548058dc334SIlpo Järvinen 35499890092eSFlorian Westphal tcp_ca_event(sk, CA_EVENT_NON_DELAYED_ACK); 35509890092eSFlorian Westphal 35511da177e4SLinus Torvalds /* We are not putting this on the write queue, so 35521da177e4SLinus Torvalds * tcp_transmit_skb() will set the ownership to this 35531da177e4SLinus Torvalds * sock. 35541da177e4SLinus Torvalds */ 35557450aaf6SEric Dumazet buff = alloc_skb(MAX_TCP_HEADER, 35567450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 35577450aaf6SEric Dumazet if (unlikely(!buff)) { 3558463c84b9SArnaldo Carvalho de Melo inet_csk_schedule_ack(sk); 3559463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; 35603f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 35613f421baaSArnaldo Carvalho de Melo TCP_DELACK_MAX, TCP_RTO_MAX); 35621da177e4SLinus Torvalds return; 35631da177e4SLinus Torvalds } 35641da177e4SLinus Torvalds 35651da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 35661da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 3567a3433f35SChangli Gao tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); 35681da177e4SLinus Torvalds 356998781965SEric Dumazet /* We do not want pure acks influencing TCP Small Queues or fq/pacing 357098781965SEric Dumazet * too much. 357198781965SEric Dumazet * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784 357298781965SEric Dumazet */ 357398781965SEric Dumazet skb_set_tcp_pure_ack(buff); 357498781965SEric Dumazet 35751da177e4SLinus Torvalds /* Send it off, this clears delayed acks for us. */ 35767450aaf6SEric Dumazet tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0); 35771da177e4SLinus Torvalds } 3578e3118e83SDaniel Borkmann EXPORT_SYMBOL_GPL(tcp_send_ack); 35791da177e4SLinus Torvalds 35801da177e4SLinus Torvalds /* This routine sends a packet with an out of date sequence 35811da177e4SLinus Torvalds * number. It assumes the other end will try to ack it. 35821da177e4SLinus Torvalds * 35831da177e4SLinus Torvalds * Question: what should we make while urgent mode? 35841da177e4SLinus Torvalds * 4.4BSD forces sending single byte of data. We cannot send 35851da177e4SLinus Torvalds * out of window data, because we have SND.NXT==SND.MAX... 35861da177e4SLinus Torvalds * 35871da177e4SLinus Torvalds * Current solution: to send TWO zero-length segments in urgent mode: 35881da177e4SLinus Torvalds * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 35891da177e4SLinus Torvalds * out-of-date with SND.UNA-1 to probe window. 35901da177e4SLinus Torvalds */ 3591e520af48SEric Dumazet static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib) 35921da177e4SLinus Torvalds { 35931da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 35941da177e4SLinus Torvalds struct sk_buff *skb; 35951da177e4SLinus Torvalds 35961da177e4SLinus Torvalds /* We don't queue it, tcp_transmit_skb() sets ownership. */ 35977450aaf6SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, 35987450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 359951456b29SIan Morris if (!skb) 36001da177e4SLinus Torvalds return -1; 36011da177e4SLinus Torvalds 36021da177e4SLinus Torvalds /* Reserve space for headers and set control bits. */ 36031da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 36041da177e4SLinus Torvalds /* Use a previous sequence. This should cause the other 36051da177e4SLinus Torvalds * end to send an ack. Don't queue or clone SKB, just 36061da177e4SLinus Torvalds * send it. 36071da177e4SLinus Torvalds */ 3608a3433f35SChangli Gao tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); 3609e2e8009fSRenato Westphal NET_INC_STATS(sock_net(sk), mib); 36107450aaf6SEric Dumazet return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0); 36111da177e4SLinus Torvalds } 36121da177e4SLinus Torvalds 3613385e2070SEric Dumazet /* Called from setsockopt( ... TCP_REPAIR ) */ 3614ee995283SPavel Emelyanov void tcp_send_window_probe(struct sock *sk) 3615ee995283SPavel Emelyanov { 3616ee995283SPavel Emelyanov if (sk->sk_state == TCP_ESTABLISHED) { 3617ee995283SPavel Emelyanov tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; 3618385e2070SEric Dumazet skb_mstamp_get(&tcp_sk(sk)->tcp_mstamp); 3619e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE); 3620ee995283SPavel Emelyanov } 3621ee995283SPavel Emelyanov } 3622ee995283SPavel Emelyanov 362367edfef7SAndi Kleen /* Initiate keepalive or window probe from timer. */ 3624e520af48SEric Dumazet int tcp_write_wakeup(struct sock *sk, int mib) 36251da177e4SLinus Torvalds { 36261da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 36271da177e4SLinus Torvalds struct sk_buff *skb; 36281da177e4SLinus Torvalds 3629058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3630058dc334SIlpo Järvinen return -1; 3631058dc334SIlpo Järvinen 363200db4124SIan Morris skb = tcp_send_head(sk); 363300db4124SIan Morris if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { 36341da177e4SLinus Torvalds int err; 36350c54b85fSIlpo Järvinen unsigned int mss = tcp_current_mss(sk); 363690840defSIlpo Järvinen unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 36371da177e4SLinus Torvalds 36381da177e4SLinus Torvalds if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 36391da177e4SLinus Torvalds tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 36401da177e4SLinus Torvalds 36411da177e4SLinus Torvalds /* We are probing the opening of a window 36421da177e4SLinus Torvalds * but the window size is != 0 36431da177e4SLinus Torvalds * must have been a result SWS avoidance ( sender ) 36441da177e4SLinus Torvalds */ 36451da177e4SLinus Torvalds if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 36461da177e4SLinus Torvalds skb->len > mss) { 36471da177e4SLinus Torvalds seg_size = min(seg_size, mss); 36484de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 36496cc55e09SOctavian Purdila if (tcp_fragment(sk, skb, seg_size, mss, GFP_ATOMIC)) 36501da177e4SLinus Torvalds return -1; 36511da177e4SLinus Torvalds } else if (!tcp_skb_pcount(skb)) 36525bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss); 36531da177e4SLinus Torvalds 36544de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 3655dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 365666f5fe62SIlpo Järvinen if (!err) 365766f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 36581da177e4SLinus Torvalds return err; 36591da177e4SLinus Torvalds } else { 366033f5f57eSIlpo Järvinen if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) 3661e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 1, mib); 3662e520af48SEric Dumazet return tcp_xmit_probe_skb(sk, 0, mib); 36631da177e4SLinus Torvalds } 36641da177e4SLinus Torvalds } 36651da177e4SLinus Torvalds 36661da177e4SLinus Torvalds /* A window probe timeout has occurred. If window is not closed send 36671da177e4SLinus Torvalds * a partial packet else a zero probe. 36681da177e4SLinus Torvalds */ 36691da177e4SLinus Torvalds void tcp_send_probe0(struct sock *sk) 36701da177e4SLinus Torvalds { 3671463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 36721da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 3673c6214a97SNikolay Borisov struct net *net = sock_net(sk); 3674fcdd1cf4SEric Dumazet unsigned long probe_max; 36751da177e4SLinus Torvalds int err; 36761da177e4SLinus Torvalds 3677e520af48SEric Dumazet err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE); 36781da177e4SLinus Torvalds 3679fe067e8aSDavid S. Miller if (tp->packets_out || !tcp_send_head(sk)) { 36801da177e4SLinus Torvalds /* Cancel probe timer, if it is not required. */ 36816687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 0; 3682463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff = 0; 36831da177e4SLinus Torvalds return; 36841da177e4SLinus Torvalds } 36851da177e4SLinus Torvalds 36861da177e4SLinus Torvalds if (err <= 0) { 3687c6214a97SNikolay Borisov if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2) 3688463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff++; 36896687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out++; 3690fcdd1cf4SEric Dumazet probe_max = TCP_RTO_MAX; 36911da177e4SLinus Torvalds } else { 36921da177e4SLinus Torvalds /* If packet was not sent due to local congestion, 36936687e988SArnaldo Carvalho de Melo * do not backoff and do not remember icsk_probes_out. 36941da177e4SLinus Torvalds * Let local senders to fight for local resources. 36951da177e4SLinus Torvalds * 36961da177e4SLinus Torvalds * Use accumulated backoff yet. 36971da177e4SLinus Torvalds */ 36986687e988SArnaldo Carvalho de Melo if (!icsk->icsk_probes_out) 36996687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 1; 3700fcdd1cf4SEric Dumazet probe_max = TCP_RESOURCE_PROBE_INTERVAL; 37011da177e4SLinus Torvalds } 3702fcdd1cf4SEric Dumazet inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 370321c8fe99SEric Dumazet tcp_probe0_when(sk, probe_max), 3704fcdd1cf4SEric Dumazet TCP_RTO_MAX); 37051da177e4SLinus Torvalds } 37065db92c99SOctavian Purdila 3707ea3bea3aSEric Dumazet int tcp_rtx_synack(const struct sock *sk, struct request_sock *req) 37085db92c99SOctavian Purdila { 37095db92c99SOctavian Purdila const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific; 37105db92c99SOctavian Purdila struct flowi fl; 37115db92c99SOctavian Purdila int res; 37125db92c99SOctavian Purdila 371358d607d3SEric Dumazet tcp_rsk(req)->txhash = net_tx_rndhash(); 3714b3d05147SEric Dumazet res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL); 37155db92c99SOctavian Purdila if (!res) { 371690bbcc60SEric Dumazet __TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); 371702a1d6e7SEric Dumazet __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 37187e32b443SYuchung Cheng if (unlikely(tcp_passive_fastopen(sk))) 37197e32b443SYuchung Cheng tcp_sk(sk)->total_retrans++; 37205db92c99SOctavian Purdila } 37215db92c99SOctavian Purdila return res; 37225db92c99SOctavian Purdila } 37235db92c99SOctavian Purdila EXPORT_SYMBOL(tcp_rtx_synack); 3724