Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1 | /* |
| 2 | * INET An implementation of the TCP/IP protocol suite for the LINUX |
| 3 | * operating system. INET is implemented using the BSD Socket |
| 4 | * interface as the means of communication with the user level. |
| 5 | * |
| 6 | * Implementation of the Transmission Control Protocol(TCP). |
| 7 | * |
Jesper Juhl | 02c30a8 | 2005-05-05 16:16:16 -0700 | [diff] [blame] | 8 | * Authors: Ross Biro |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 9 | * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> |
| 10 | * Mark Evans, <evansmp@uhura.aston.ac.uk> |
| 11 | * Corey Minyard <wf-rch!minyard@relay.EU.net> |
| 12 | * Florian La Roche, <flla@stud.uni-sb.de> |
| 13 | * Charles Hedrick, <hedrick@klinzhai.rutgers.edu> |
| 14 | * Linus Torvalds, <torvalds@cs.helsinki.fi> |
| 15 | * Alan Cox, <gw4pts@gw4pts.ampr.org> |
| 16 | * Matthew Dillon, <dillon@apollo.west.oic.com> |
| 17 | * Arnt Gulbrandsen, <agulbra@nvg.unit.no> |
| 18 | * Jorge Cwik, <jorge@laser.satlink.net> |
| 19 | */ |
| 20 | |
| 21 | /* |
| 22 | * Changes: Pedro Roque : Retransmit queue handled by TCP. |
| 23 | * : Fragmentation on mtu decrease |
| 24 | * : Segment collapse on retransmit |
| 25 | * : AF independence |
| 26 | * |
| 27 | * Linus Torvalds : send_delayed_ack |
| 28 | * David S. Miller : Charge memory using the right skb |
| 29 | * during syn/ack processing. |
| 30 | * David S. Miller : Output engine completely rewritten. |
| 31 | * Andrea Arcangeli: SYNACK carry ts_recent in tsecr. |
| 32 | * Cacophonix Gaul : draft-minshall-nagle-01 |
| 33 | * J Hadi Salim : ECN support |
| 34 | * |
| 35 | */ |
| 36 | |
Joe Perches | 91df42b | 2012-05-15 14:11:54 +0000 | [diff] [blame] | 37 | #define pr_fmt(fmt) "TCP: " fmt |
| 38 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 39 | #include <net/tcp.h> |
| 40 | |
| 41 | #include <linux/compiler.h> |
Tejun Heo | 5a0e3ad | 2010-03-24 17:04:11 +0900 | [diff] [blame] | 42 | #include <linux/gfp.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 43 | #include <linux/module.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 44 | |
| 45 | /* People can turn this off for buggy TCP's found in printers etc. */ |
Brian Haley | ab32ea5 | 2006-09-22 14:15:41 -0700 | [diff] [blame] | 46 | int sysctl_tcp_retrans_collapse __read_mostly = 1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 47 | |
Jianjun Kong | 09cb105 | 2008-11-03 00:27:11 -0800 | [diff] [blame] | 48 | /* People can turn this on to work with those rare, broken TCPs that |
Rick Jones | 15d99e0 | 2006-03-20 22:40:29 -0800 | [diff] [blame] | 49 | * interpret the window field as a signed quantity. |
| 50 | */ |
Brian Haley | ab32ea5 | 2006-09-22 14:15:41 -0700 | [diff] [blame] | 51 | int sysctl_tcp_workaround_signed_windows __read_mostly = 0; |
Rick Jones | 15d99e0 | 2006-03-20 22:40:29 -0800 | [diff] [blame] | 52 | |
Wei Liu | c39c4c6 | 2015-06-03 11:10:42 +0100 | [diff] [blame] | 53 | /* Default TSQ limit of four TSO segments */ |
| 54 | int sysctl_tcp_limit_output_bytes __read_mostly = 262144; |
Eric Dumazet | 46d3cea | 2012-07-11 05:50:31 +0000 | [diff] [blame] | 55 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 56 | /* This limits the percentage of the congestion window which we |
| 57 | * will allow a single TSO frame to consume. Building TSO frames |
| 58 | * which are too large can cause TCP streams to be bursty. |
| 59 | */ |
Brian Haley | ab32ea5 | 2006-09-22 14:15:41 -0700 | [diff] [blame] | 60 | int sysctl_tcp_tso_win_divisor __read_mostly = 3; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 61 | |
David S. Miller | 35089bb | 2006-06-13 22:33:04 -0700 | [diff] [blame] | 62 | /* By default, RFC2861 behavior. */ |
Brian Haley | ab32ea5 | 2006-09-22 14:15:41 -0700 | [diff] [blame] | 63 | int sysctl_tcp_slow_start_after_idle __read_mostly = 1; |
David S. Miller | 35089bb | 2006-06-13 22:33:04 -0700 | [diff] [blame] | 64 | |
Eric Dumazet | 46d3cea | 2012-07-11 05:50:31 +0000 | [diff] [blame] | 65 | static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, |
| 66 | int push_one, gfp_t gfp); |
William Allen Simpson | 519855c | 2009-12-02 18:14:19 +0000 | [diff] [blame] | 67 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 68 | /* Account for new data that has been sent to the network. */ |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 69 | static void tcp_event_new_data_sent(struct sock *sk, const struct sk_buff *skb) |
Ilpo Järvinen | 6ff03ac | 2007-08-24 22:44:06 -0700 | [diff] [blame] | 70 | { |
Nandita Dukkipati | 6ba8a3b | 2013-03-11 10:00:43 +0000 | [diff] [blame] | 71 | struct inet_connection_sock *icsk = inet_csk(sk); |
Ilpo Järvinen | 6ff03ac | 2007-08-24 22:44:06 -0700 | [diff] [blame] | 72 | struct tcp_sock *tp = tcp_sk(sk); |
Ilpo Järvinen | 66f5fe6 | 2007-12-31 04:43:57 -0800 | [diff] [blame] | 73 | unsigned int prior_packets = tp->packets_out; |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 74 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 75 | tcp_advance_send_head(sk, skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 76 | tp->snd_nxt = TCP_SKB_CB(skb)->end_seq; |
Ilpo Järvinen | 8512430 | 2007-11-26 20:17:38 +0800 | [diff] [blame] | 77 | |
Ilpo Järvinen | 66f5fe6 | 2007-12-31 04:43:57 -0800 | [diff] [blame] | 78 | tp->packets_out += tcp_skb_pcount(skb); |
Nandita Dukkipati | 6ba8a3b | 2013-03-11 10:00:43 +0000 | [diff] [blame] | 79 | if (!prior_packets || icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS || |
Eric Dumazet | 6a5dc9e | 2013-04-29 08:39:56 +0000 | [diff] [blame] | 80 | icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) { |
Yuchung Cheng | 750ea2b | 2012-05-02 13:30:04 +0000 | [diff] [blame] | 81 | tcp_rearm_rto(sk); |
Eric Dumazet | 6a5dc9e | 2013-04-29 08:39:56 +0000 | [diff] [blame] | 82 | } |
Yuchung Cheng | f19c29e | 2014-03-03 12:31:36 -0800 | [diff] [blame] | 83 | |
David S. Miller | f7324ac | 2014-03-06 15:03:17 -0500 | [diff] [blame] | 84 | NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT, |
| 85 | tcp_skb_pcount(skb)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 86 | } |
| 87 | |
| 88 | /* SND.NXT, if window was not shrunk. |
| 89 | * If window has been shrunk, what should we make? It is not clear at all. |
| 90 | * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-( |
| 91 | * Anything in between SND.UNA...SND.UNA+SND.WND also can be already |
| 92 | * invalid. OK, let's make this for now: |
| 93 | */ |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 94 | static inline __u32 tcp_acceptable_seq(const struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 95 | { |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 96 | const struct tcp_sock *tp = tcp_sk(sk); |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 97 | |
Ilpo Järvinen | 90840de | 2007-12-31 04:48:41 -0800 | [diff] [blame] | 98 | if (!before(tcp_wnd_end(tp), tp->snd_nxt)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 99 | return tp->snd_nxt; |
| 100 | else |
Ilpo Järvinen | 90840de | 2007-12-31 04:48:41 -0800 | [diff] [blame] | 101 | return tcp_wnd_end(tp); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 102 | } |
| 103 | |
| 104 | /* Calculate mss to advertise in SYN segment. |
| 105 | * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that: |
| 106 | * |
| 107 | * 1. It is independent of path mtu. |
| 108 | * 2. Ideally, it is maximal possible segment size i.e. 65535-40. |
| 109 | * 3. For IPv4 it is reasonable to calculate it from maximal MTU of |
| 110 | * attached devices, because some buggy hosts are confused by |
| 111 | * large MSS. |
| 112 | * 4. We do not make 3, we advertise MSS, calculated from first |
| 113 | * hop device mtu, but allow to raise it to ip_rt_min_advmss. |
| 114 | * This may be overridden via information stored in routing table. |
| 115 | * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible, |
| 116 | * probably even Jumbo". |
| 117 | */ |
| 118 | static __u16 tcp_advertise_mss(struct sock *sk) |
| 119 | { |
| 120 | struct tcp_sock *tp = tcp_sk(sk); |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 121 | const struct dst_entry *dst = __sk_dst_get(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 122 | int mss = tp->advmss; |
| 123 | |
David S. Miller | 0dbaee3 | 2010-12-13 12:52:14 -0800 | [diff] [blame] | 124 | if (dst) { |
| 125 | unsigned int metric = dst_metric_advmss(dst); |
| 126 | |
| 127 | if (metric < mss) { |
| 128 | mss = metric; |
| 129 | tp->advmss = mss; |
| 130 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 131 | } |
| 132 | |
| 133 | return (__u16)mss; |
| 134 | } |
| 135 | |
| 136 | /* RFC2861. Reset CWND after idle period longer RTO to "restart window". |
Eric Dumazet | 6f021c6 | 2015-08-21 12:30:00 -0700 | [diff] [blame] | 137 | * This is the first part of cwnd validation mechanism. |
| 138 | */ |
| 139 | void tcp_cwnd_restart(struct sock *sk, s32 delta) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 140 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 141 | struct tcp_sock *tp = tcp_sk(sk); |
Eric Dumazet | 6f021c6 | 2015-08-21 12:30:00 -0700 | [diff] [blame] | 142 | u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 143 | u32 cwnd = tp->snd_cwnd; |
| 144 | |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 145 | tcp_ca_event(sk, CA_EVENT_CWND_RESTART); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 146 | |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 147 | tp->snd_ssthresh = tcp_current_ssthresh(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 148 | restart_cwnd = min(restart_cwnd, cwnd); |
| 149 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 150 | while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 151 | cwnd >>= 1; |
| 152 | tp->snd_cwnd = max(cwnd, restart_cwnd); |
| 153 | tp->snd_cwnd_stamp = tcp_time_stamp; |
| 154 | tp->snd_cwnd_used = 0; |
| 155 | } |
| 156 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 157 | /* Congestion state accounting after a packet has been sent. */ |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 158 | static void tcp_event_data_sent(struct tcp_sock *tp, |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 159 | struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 160 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 161 | struct inet_connection_sock *icsk = inet_csk(sk); |
| 162 | const u32 now = tcp_time_stamp; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 163 | |
Neal Cardwell | 05c5a46 | 2015-09-09 21:54:37 -0700 | [diff] [blame] | 164 | if (tcp_packets_in_flight(tp) == 0) |
| 165 | tcp_ca_event(sk, CA_EVENT_TX_START); |
| 166 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 167 | tp->lsndtime = now; |
| 168 | |
| 169 | /* If it is a reply for ato after last received |
| 170 | * packet, enter pingpong mode. |
| 171 | */ |
Jon Maxwell | 2251ae4 | 2015-07-08 10:12:28 +1000 | [diff] [blame] | 172 | if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato) |
| 173 | icsk->icsk_ack.pingpong = 1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 174 | } |
| 175 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 176 | /* Account for an ACK we sent. */ |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 177 | static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 178 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 179 | tcp_dec_quickack_mode(sk, pkts); |
| 180 | inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 181 | } |
| 182 | |
Yuchung Cheng | 85f1652 | 2013-06-11 15:35:32 -0700 | [diff] [blame] | 183 | |
| 184 | u32 tcp_default_init_rwnd(u32 mss) |
| 185 | { |
| 186 | /* Initial receive window should be twice of TCP_INIT_CWND to |
Weiping Pan | 9ef71e0 | 2013-06-18 21:00:31 +0800 | [diff] [blame] | 187 | * enable proper sending of new unsent data during fast recovery |
Yuchung Cheng | 85f1652 | 2013-06-11 15:35:32 -0700 | [diff] [blame] | 188 | * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a |
| 189 | * limit when mss is larger than 1460. |
| 190 | */ |
| 191 | u32 init_rwnd = TCP_INIT_CWND * 2; |
| 192 | |
| 193 | if (mss > 1460) |
| 194 | init_rwnd = max((1460 * init_rwnd) / mss, 2U); |
| 195 | return init_rwnd; |
| 196 | } |
| 197 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 198 | /* Determine a window scaling and initial window to offer. |
| 199 | * Based on the assumption that the given amount of space |
| 200 | * will be offered. Store the results in the tp structure. |
| 201 | * NOTE: for smooth operation initial space offering should |
| 202 | * be a multiple of mss if possible. We assume here that mss >= 1. |
| 203 | * This MUST be enforced by all callers. |
| 204 | */ |
| 205 | void tcp_select_initial_window(int __space, __u32 mss, |
| 206 | __u32 *rcv_wnd, __u32 *window_clamp, |
laurent chavey | 31d1292 | 2009-12-15 11:15:28 +0000 | [diff] [blame] | 207 | int wscale_ok, __u8 *rcv_wscale, |
| 208 | __u32 init_rcv_wnd) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 209 | { |
| 210 | unsigned int space = (__space < 0 ? 0 : __space); |
| 211 | |
| 212 | /* If no clamp set the clamp to the max possible scaled window */ |
| 213 | if (*window_clamp == 0) |
| 214 | (*window_clamp) = (65535 << 14); |
| 215 | space = min(*window_clamp, space); |
| 216 | |
| 217 | /* Quantize space offering to a multiple of mss if possible. */ |
| 218 | if (space > mss) |
| 219 | space = (space / mss) * mss; |
| 220 | |
| 221 | /* NOTE: offering an initial window larger than 32767 |
Rick Jones | 15d99e0 | 2006-03-20 22:40:29 -0800 | [diff] [blame] | 222 | * will break some buggy TCP stacks. If the admin tells us |
| 223 | * it is likely we could be speaking with such a buggy stack |
| 224 | * we will truncate our initial window offering to 32K-1 |
| 225 | * unless the remote has sent us a window scaling option, |
| 226 | * which we interpret as a sign the remote TCP is not |
| 227 | * misinterpreting the window field as a signed quantity. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 228 | */ |
Rick Jones | 15d99e0 | 2006-03-20 22:40:29 -0800 | [diff] [blame] | 229 | if (sysctl_tcp_workaround_signed_windows) |
| 230 | (*rcv_wnd) = min(space, MAX_TCP_WINDOW); |
| 231 | else |
| 232 | (*rcv_wnd) = space; |
| 233 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 234 | (*rcv_wscale) = 0; |
| 235 | if (wscale_ok) { |
| 236 | /* Set window scaling on max possible window |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 237 | * See RFC1323 for an explanation of the limit to 14 |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 238 | */ |
Soheil Hassas Yeganeh | f626300 | 2016-07-29 09:34:02 -0400 | [diff] [blame] | 239 | space = max_t(u32, space, sysctl_tcp_rmem[2]); |
| 240 | space = max_t(u32, space, sysctl_rmem_max); |
Stephen Hemminger | 316c159 | 2006-08-22 00:06:11 -0700 | [diff] [blame] | 241 | space = min_t(u32, space, *window_clamp); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 242 | while (space > 65535 && (*rcv_wscale) < 14) { |
| 243 | space >>= 1; |
| 244 | (*rcv_wscale)++; |
| 245 | } |
| 246 | } |
| 247 | |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 248 | if (mss > (1 << *rcv_wscale)) { |
Yuchung Cheng | 85f1652 | 2013-06-11 15:35:32 -0700 | [diff] [blame] | 249 | if (!init_rcv_wnd) /* Use default unless specified otherwise */ |
| 250 | init_rcv_wnd = tcp_default_init_rwnd(mss); |
| 251 | *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 252 | } |
| 253 | |
| 254 | /* Set the clamp no higher than max representable value */ |
| 255 | (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp); |
| 256 | } |
Eric Dumazet | 4bc2f18 | 2010-07-09 21:22:10 +0000 | [diff] [blame] | 257 | EXPORT_SYMBOL(tcp_select_initial_window); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 258 | |
| 259 | /* Chose a new window to advertise, update state in tcp_sock for the |
| 260 | * socket, and return result with RFC1323 scaling applied. The return |
| 261 | * value can be stuffed directly into th->window for an outgoing |
| 262 | * frame. |
| 263 | */ |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 264 | static u16 tcp_select_window(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 265 | { |
| 266 | struct tcp_sock *tp = tcp_sk(sk); |
Florian Westphal | 8e165e2 | 2014-02-25 14:34:32 +0100 | [diff] [blame] | 267 | u32 old_win = tp->rcv_wnd; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 268 | u32 cur_win = tcp_receive_window(tp); |
| 269 | u32 new_win = __tcp_select_window(sk); |
| 270 | |
| 271 | /* Never shrink the offered window */ |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 272 | if (new_win < cur_win) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 273 | /* Danger Will Robinson! |
| 274 | * Don't update rcv_wup/rcv_wnd here or else |
| 275 | * we will not be able to advertise a zero |
| 276 | * window in time. --DaveM |
| 277 | * |
| 278 | * Relax Will Robinson. |
| 279 | */ |
Florian Westphal | 8e165e2 | 2014-02-25 14:34:32 +0100 | [diff] [blame] | 280 | if (new_win == 0) |
| 281 | NET_INC_STATS(sock_net(sk), |
| 282 | LINUX_MIB_TCPWANTZEROWINDOWADV); |
Patrick McHardy | 607bfbf | 2008-03-20 16:11:27 -0700 | [diff] [blame] | 283 | new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 284 | } |
| 285 | tp->rcv_wnd = new_win; |
| 286 | tp->rcv_wup = tp->rcv_nxt; |
| 287 | |
| 288 | /* Make sure we do not exceed the maximum possible |
| 289 | * scaled window. |
| 290 | */ |
Rick Jones | 15d99e0 | 2006-03-20 22:40:29 -0800 | [diff] [blame] | 291 | if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 292 | new_win = min(new_win, MAX_TCP_WINDOW); |
| 293 | else |
| 294 | new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); |
| 295 | |
| 296 | /* RFC1323 scaling applied */ |
| 297 | new_win >>= tp->rx_opt.rcv_wscale; |
| 298 | |
| 299 | /* If we advertise zero window, disable fast path. */ |
Florian Westphal | 8e165e2 | 2014-02-25 14:34:32 +0100 | [diff] [blame] | 300 | if (new_win == 0) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 301 | tp->pred_flags = 0; |
Florian Westphal | 8e165e2 | 2014-02-25 14:34:32 +0100 | [diff] [blame] | 302 | if (old_win) |
| 303 | NET_INC_STATS(sock_net(sk), |
| 304 | LINUX_MIB_TCPTOZEROWINDOWADV); |
| 305 | } else if (old_win == 0) { |
| 306 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV); |
| 307 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 308 | |
| 309 | return new_win; |
| 310 | } |
| 311 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 312 | /* Packet ECN state for a SYN-ACK */ |
Florian Westphal | 735d383 | 2014-09-29 13:08:30 +0200 | [diff] [blame] | 313 | static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb) |
Ilpo Järvinen | bdf1ee5 | 2007-05-27 02:04:16 -0700 | [diff] [blame] | 314 | { |
Daniel Borkmann | 30e502a | 2014-09-26 22:37:33 +0200 | [diff] [blame] | 315 | const struct tcp_sock *tp = tcp_sk(sk); |
| 316 | |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 317 | TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR; |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 318 | if (!(tp->ecn_flags & TCP_ECN_OK)) |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 319 | TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE; |
Daniel Borkmann | 30e502a | 2014-09-26 22:37:33 +0200 | [diff] [blame] | 320 | else if (tcp_ca_needs_ecn(sk)) |
| 321 | INET_ECN_xmit(sk); |
Ilpo Järvinen | bdf1ee5 | 2007-05-27 02:04:16 -0700 | [diff] [blame] | 322 | } |
| 323 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 324 | /* Packet ECN state for a SYN. */ |
Florian Westphal | 735d383 | 2014-09-29 13:08:30 +0200 | [diff] [blame] | 325 | static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb) |
Ilpo Järvinen | bdf1ee5 | 2007-05-27 02:04:16 -0700 | [diff] [blame] | 326 | { |
| 327 | struct tcp_sock *tp = tcp_sk(sk); |
Florian Westphal | f7b3bec | 2014-11-03 17:35:03 +0100 | [diff] [blame] | 328 | bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 || |
| 329 | tcp_ca_needs_ecn(sk); |
| 330 | |
| 331 | if (!use_ecn) { |
| 332 | const struct dst_entry *dst = __sk_dst_get(sk); |
| 333 | |
| 334 | if (dst && dst_feature(dst, RTAX_FEATURE_ECN)) |
| 335 | use_ecn = true; |
| 336 | } |
Ilpo Järvinen | bdf1ee5 | 2007-05-27 02:04:16 -0700 | [diff] [blame] | 337 | |
| 338 | tp->ecn_flags = 0; |
Florian Westphal | f7b3bec | 2014-11-03 17:35:03 +0100 | [diff] [blame] | 339 | |
| 340 | if (use_ecn) { |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 341 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR; |
Ilpo Järvinen | bdf1ee5 | 2007-05-27 02:04:16 -0700 | [diff] [blame] | 342 | tp->ecn_flags = TCP_ECN_OK; |
Daniel Borkmann | 30e502a | 2014-09-26 22:37:33 +0200 | [diff] [blame] | 343 | if (tcp_ca_needs_ecn(sk)) |
| 344 | INET_ECN_xmit(sk); |
Ilpo Järvinen | bdf1ee5 | 2007-05-27 02:04:16 -0700 | [diff] [blame] | 345 | } |
| 346 | } |
| 347 | |
Daniel Borkmann | 4921355 | 2015-05-19 21:04:22 +0200 | [diff] [blame] | 348 | static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb) |
| 349 | { |
| 350 | if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback) |
| 351 | /* tp->ecn_flags are cleared at a later point in time when |
| 352 | * SYN ACK is ultimatively being received. |
| 353 | */ |
| 354 | TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR); |
| 355 | } |
| 356 | |
Florian Westphal | 735d383 | 2014-09-29 13:08:30 +0200 | [diff] [blame] | 357 | static void |
Eric Dumazet | 6ac705b | 2015-09-25 07:39:18 -0700 | [diff] [blame] | 358 | tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th) |
Ilpo Järvinen | bdf1ee5 | 2007-05-27 02:04:16 -0700 | [diff] [blame] | 359 | { |
Eric Dumazet | 6ac705b | 2015-09-25 07:39:18 -0700 | [diff] [blame] | 360 | if (inet_rsk(req)->ecn_ok) |
Ilpo Järvinen | bdf1ee5 | 2007-05-27 02:04:16 -0700 | [diff] [blame] | 361 | th->ece = 1; |
| 362 | } |
| 363 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 364 | /* Set up ECN state for a packet on a ESTABLISHED socket that is about to |
| 365 | * be sent. |
| 366 | */ |
Florian Westphal | 735d383 | 2014-09-29 13:08:30 +0200 | [diff] [blame] | 367 | static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb, |
Eric Dumazet | ea1627c | 2016-05-13 09:16:40 -0700 | [diff] [blame] | 368 | struct tcphdr *th, int tcp_header_len) |
Ilpo Järvinen | bdf1ee5 | 2007-05-27 02:04:16 -0700 | [diff] [blame] | 369 | { |
| 370 | struct tcp_sock *tp = tcp_sk(sk); |
| 371 | |
| 372 | if (tp->ecn_flags & TCP_ECN_OK) { |
| 373 | /* Not-retransmitted data segment: set ECT and inject CWR. */ |
| 374 | if (skb->len != tcp_header_len && |
| 375 | !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { |
| 376 | INET_ECN_xmit(sk); |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 377 | if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) { |
Ilpo Järvinen | bdf1ee5 | 2007-05-27 02:04:16 -0700 | [diff] [blame] | 378 | tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; |
Eric Dumazet | ea1627c | 2016-05-13 09:16:40 -0700 | [diff] [blame] | 379 | th->cwr = 1; |
Ilpo Järvinen | bdf1ee5 | 2007-05-27 02:04:16 -0700 | [diff] [blame] | 380 | skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; |
| 381 | } |
Daniel Borkmann | 30e502a | 2014-09-26 22:37:33 +0200 | [diff] [blame] | 382 | } else if (!tcp_ca_needs_ecn(sk)) { |
Ilpo Järvinen | bdf1ee5 | 2007-05-27 02:04:16 -0700 | [diff] [blame] | 383 | /* ACK or retransmitted segment: clear ECT|CE */ |
| 384 | INET_ECN_dontxmit(sk); |
| 385 | } |
| 386 | if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) |
Eric Dumazet | ea1627c | 2016-05-13 09:16:40 -0700 | [diff] [blame] | 387 | th->ece = 1; |
Ilpo Järvinen | bdf1ee5 | 2007-05-27 02:04:16 -0700 | [diff] [blame] | 388 | } |
| 389 | } |
| 390 | |
Ilpo Järvinen | e870a8e | 2008-01-03 20:39:01 -0800 | [diff] [blame] | 391 | /* Constructs common control bits of non-data skb. If SYN/FIN is present, |
| 392 | * auto increment end seqno. |
| 393 | */ |
| 394 | static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags) |
| 395 | { |
David S. Miller | 2e8e18e | 2010-04-08 11:32:30 -0700 | [diff] [blame] | 396 | skb->ip_summed = CHECKSUM_PARTIAL; |
Ilpo Järvinen | e870a8e | 2008-01-03 20:39:01 -0800 | [diff] [blame] | 397 | skb->csum = 0; |
| 398 | |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 399 | TCP_SKB_CB(skb)->tcp_flags = flags; |
Ilpo Järvinen | e870a8e | 2008-01-03 20:39:01 -0800 | [diff] [blame] | 400 | TCP_SKB_CB(skb)->sacked = 0; |
| 401 | |
Eric Dumazet | cd7d849 | 2014-09-24 04:11:22 -0700 | [diff] [blame] | 402 | tcp_skb_pcount_set(skb, 1); |
Ilpo Järvinen | e870a8e | 2008-01-03 20:39:01 -0800 | [diff] [blame] | 403 | |
| 404 | TCP_SKB_CB(skb)->seq = seq; |
Changli Gao | a3433f3 | 2010-06-12 14:01:43 +0000 | [diff] [blame] | 405 | if (flags & (TCPHDR_SYN | TCPHDR_FIN)) |
Ilpo Järvinen | e870a8e | 2008-01-03 20:39:01 -0800 | [diff] [blame] | 406 | seq++; |
| 407 | TCP_SKB_CB(skb)->end_seq = seq; |
| 408 | } |
| 409 | |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 410 | static inline bool tcp_urg_mode(const struct tcp_sock *tp) |
Ilpo Järvinen | 33f5f57 | 2008-10-07 14:43:06 -0700 | [diff] [blame] | 411 | { |
| 412 | return tp->snd_una != tp->snd_up; |
| 413 | } |
| 414 | |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 415 | #define OPTION_SACK_ADVERTISE (1 << 0) |
| 416 | #define OPTION_TS (1 << 1) |
| 417 | #define OPTION_MD5 (1 << 2) |
Ori Finkelman | 89e95a6 | 2009-10-01 06:41:59 +0000 | [diff] [blame] | 418 | #define OPTION_WSCALE (1 << 3) |
Yuchung Cheng | 2100c8d | 2012-07-19 06:43:05 +0000 | [diff] [blame] | 419 | #define OPTION_FAST_OPEN_COOKIE (1 << 8) |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 420 | |
| 421 | struct tcp_out_options { |
Yuchung Cheng | 2100c8d | 2012-07-19 06:43:05 +0000 | [diff] [blame] | 422 | u16 options; /* bit field of OPTION_* */ |
| 423 | u16 mss; /* 0 to disable */ |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 424 | u8 ws; /* window scale, 0 to disable */ |
| 425 | u8 num_sack_blocks; /* number of SACK blocks to include */ |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 426 | u8 hash_size; /* bytes in hash_location */ |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 427 | __u8 *hash_location; /* temporary pointer, overloaded */ |
Yuchung Cheng | 2100c8d | 2012-07-19 06:43:05 +0000 | [diff] [blame] | 428 | __u32 tsval, tsecr; /* need to include OPTION_TS */ |
| 429 | struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */ |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 430 | }; |
| 431 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 432 | /* Write previously computed TCP options to the packet. |
| 433 | * |
| 434 | * Beware: Something in the Internet is very sensitive to the ordering of |
Ilpo Järvinen | fd6149d | 2008-10-23 14:06:35 -0700 | [diff] [blame] | 435 | * TCP options, we learned this through the hard way, so be careful here. |
| 436 | * Luckily we can at least blame others for their non-compliance but from |
stephen hemminger | 8e3bff9 | 2013-12-08 12:15:44 -0800 | [diff] [blame] | 437 | * inter-operability perspective it seems that we're somewhat stuck with |
Ilpo Järvinen | fd6149d | 2008-10-23 14:06:35 -0700 | [diff] [blame] | 438 | * the ordering which we have been using if we want to keep working with |
| 439 | * those broken things (not that it currently hurts anybody as there isn't |
| 440 | * particular reason why the ordering would need to be changed). |
| 441 | * |
| 442 | * At least SACK_PERM as the first option is known to lead to a disaster |
| 443 | * (but it may well be that other scenarios fail similarly). |
| 444 | */ |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 445 | static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp, |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 446 | struct tcp_out_options *opts) |
| 447 | { |
Yuchung Cheng | 2100c8d | 2012-07-19 06:43:05 +0000 | [diff] [blame] | 448 | u16 options = opts->options; /* mungable copy */ |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 449 | |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 450 | if (unlikely(OPTION_MD5 & options)) { |
Christoph Paasch | 1a2c618 | 2013-03-17 08:23:34 +0000 | [diff] [blame] | 451 | *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | |
| 452 | (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG); |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 453 | /* overload cookie hash location */ |
| 454 | opts->hash_location = (__u8 *)ptr; |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 455 | ptr += 4; |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 456 | } |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 457 | |
Ilpo Järvinen | fd6149d | 2008-10-23 14:06:35 -0700 | [diff] [blame] | 458 | if (unlikely(opts->mss)) { |
| 459 | *ptr++ = htonl((TCPOPT_MSS << 24) | |
| 460 | (TCPOLEN_MSS << 16) | |
| 461 | opts->mss); |
| 462 | } |
| 463 | |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 464 | if (likely(OPTION_TS & options)) { |
| 465 | if (unlikely(OPTION_SACK_ADVERTISE & options)) { |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 466 | *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | |
| 467 | (TCPOLEN_SACK_PERM << 16) | |
| 468 | (TCPOPT_TIMESTAMP << 8) | |
| 469 | TCPOLEN_TIMESTAMP); |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 470 | options &= ~OPTION_SACK_ADVERTISE; |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 471 | } else { |
| 472 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
| 473 | (TCPOPT_NOP << 16) | |
| 474 | (TCPOPT_TIMESTAMP << 8) | |
| 475 | TCPOLEN_TIMESTAMP); |
| 476 | } |
| 477 | *ptr++ = htonl(opts->tsval); |
| 478 | *ptr++ = htonl(opts->tsecr); |
| 479 | } |
| 480 | |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 481 | if (unlikely(OPTION_SACK_ADVERTISE & options)) { |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 482 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
| 483 | (TCPOPT_NOP << 16) | |
| 484 | (TCPOPT_SACK_PERM << 8) | |
| 485 | TCPOLEN_SACK_PERM); |
| 486 | } |
| 487 | |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 488 | if (unlikely(OPTION_WSCALE & options)) { |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 489 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
| 490 | (TCPOPT_WINDOW << 16) | |
| 491 | (TCPOLEN_WINDOW << 8) | |
| 492 | opts->ws); |
| 493 | } |
| 494 | |
| 495 | if (unlikely(opts->num_sack_blocks)) { |
| 496 | struct tcp_sack_block *sp = tp->rx_opt.dsack ? |
| 497 | tp->duplicate_sack : tp->selective_acks; |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 498 | int this_sack; |
| 499 | |
| 500 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
| 501 | (TCPOPT_NOP << 16) | |
| 502 | (TCPOPT_SACK << 8) | |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 503 | (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 504 | TCPOLEN_SACK_PERBLOCK))); |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 505 | |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 506 | for (this_sack = 0; this_sack < opts->num_sack_blocks; |
| 507 | ++this_sack) { |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 508 | *ptr++ = htonl(sp[this_sack].start_seq); |
| 509 | *ptr++ = htonl(sp[this_sack].end_seq); |
| 510 | } |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 511 | |
Ilpo Järvinen | 5861f8e | 2009-03-14 14:23:01 +0000 | [diff] [blame] | 512 | tp->rx_opt.dsack = 0; |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 513 | } |
Yuchung Cheng | 2100c8d | 2012-07-19 06:43:05 +0000 | [diff] [blame] | 514 | |
| 515 | if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) { |
| 516 | struct tcp_fastopen_cookie *foc = opts->fastopen_cookie; |
Daniel Lee | 7f9b838 | 2015-04-06 14:37:26 -0700 | [diff] [blame] | 517 | u8 *p = (u8 *)ptr; |
| 518 | u32 len; /* Fast Open option length */ |
Yuchung Cheng | 2100c8d | 2012-07-19 06:43:05 +0000 | [diff] [blame] | 519 | |
Daniel Lee | 7f9b838 | 2015-04-06 14:37:26 -0700 | [diff] [blame] | 520 | if (foc->exp) { |
| 521 | len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len; |
| 522 | *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) | |
| 523 | TCPOPT_FASTOPEN_MAGIC); |
| 524 | p += TCPOLEN_EXP_FASTOPEN_BASE; |
| 525 | } else { |
| 526 | len = TCPOLEN_FASTOPEN_BASE + foc->len; |
| 527 | *p++ = TCPOPT_FASTOPEN; |
| 528 | *p++ = len; |
Yuchung Cheng | 2100c8d | 2012-07-19 06:43:05 +0000 | [diff] [blame] | 529 | } |
Daniel Lee | 7f9b838 | 2015-04-06 14:37:26 -0700 | [diff] [blame] | 530 | |
| 531 | memcpy(p, foc->val, foc->len); |
| 532 | if ((len & 3) == 2) { |
| 533 | p[foc->len] = TCPOPT_NOP; |
| 534 | p[foc->len + 1] = TCPOPT_NOP; |
| 535 | } |
| 536 | ptr += (len + 3) >> 2; |
Yuchung Cheng | 2100c8d | 2012-07-19 06:43:05 +0000 | [diff] [blame] | 537 | } |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 538 | } |
| 539 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 540 | /* Compute TCP options for SYN packets. This is not the final |
| 541 | * network wire format yet. |
| 542 | */ |
Eric Dumazet | 95c9617 | 2012-04-15 05:58:06 +0000 | [diff] [blame] | 543 | static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb, |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 544 | struct tcp_out_options *opts, |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 545 | struct tcp_md5sig_key **md5) |
| 546 | { |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 547 | struct tcp_sock *tp = tcp_sk(sk); |
Eric Dumazet | 95c9617 | 2012-04-15 05:58:06 +0000 | [diff] [blame] | 548 | unsigned int remaining = MAX_TCP_OPTION_SPACE; |
Yuchung Cheng | 783237e | 2012-07-19 06:43:07 +0000 | [diff] [blame] | 549 | struct tcp_fastopen_request *fastopen = tp->fastopen_req; |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 550 | |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 551 | #ifdef CONFIG_TCP_MD5SIG |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 552 | *md5 = tp->af_specific->md5_lookup(sk, sk); |
| 553 | if (*md5) { |
| 554 | opts->options |= OPTION_MD5; |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 555 | remaining -= TCPOLEN_MD5SIG_ALIGNED; |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 556 | } |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 557 | #else |
| 558 | *md5 = NULL; |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 559 | #endif |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 560 | |
| 561 | /* We always get an MSS option. The option bytes which will be seen in |
| 562 | * normal data packets should timestamps be used, must be in the MSS |
| 563 | * advertised. But we subtract them from tp->mss_cache so that |
| 564 | * calculations in tcp_sendmsg are simpler etc. So account for this |
| 565 | * fact here if necessary. If we don't do this correctly, as a |
| 566 | * receiver we won't recognize data packets as being full sized when we |
| 567 | * should, and thus we won't abide by the delayed ACK rules correctly. |
| 568 | * SACKs don't matter, we never delay an ACK when we have any of those |
| 569 | * going out. */ |
| 570 | opts->mss = tcp_advertise_mss(sk); |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 571 | remaining -= TCPOLEN_MSS_ALIGNED; |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 572 | |
Ian Morris | 51456b2 | 2015-04-03 09:17:26 +0100 | [diff] [blame] | 573 | if (likely(sysctl_tcp_timestamps && !*md5)) { |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 574 | opts->options |= OPTION_TS; |
Eric Dumazet | 7faee5c | 2014-09-05 15:33:33 -0700 | [diff] [blame] | 575 | opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset; |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 576 | opts->tsecr = tp->rx_opt.ts_recent; |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 577 | remaining -= TCPOLEN_TSTAMP_ALIGNED; |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 578 | } |
David S. Miller | bb5b7c1 | 2009-12-15 20:56:42 -0800 | [diff] [blame] | 579 | if (likely(sysctl_tcp_window_scaling)) { |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 580 | opts->ws = tp->rx_opt.rcv_wscale; |
Ori Finkelman | 89e95a6 | 2009-10-01 06:41:59 +0000 | [diff] [blame] | 581 | opts->options |= OPTION_WSCALE; |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 582 | remaining -= TCPOLEN_WSCALE_ALIGNED; |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 583 | } |
David S. Miller | bb5b7c1 | 2009-12-15 20:56:42 -0800 | [diff] [blame] | 584 | if (likely(sysctl_tcp_sack)) { |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 585 | opts->options |= OPTION_SACK_ADVERTISE; |
David S. Miller | b32d131 | 2008-07-21 18:45:34 -0700 | [diff] [blame] | 586 | if (unlikely(!(OPTION_TS & opts->options))) |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 587 | remaining -= TCPOLEN_SACKPERM_ALIGNED; |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 588 | } |
| 589 | |
Yuchung Cheng | 783237e | 2012-07-19 06:43:07 +0000 | [diff] [blame] | 590 | if (fastopen && fastopen->cookie.len >= 0) { |
Daniel Lee | 2646c83 | 2015-04-06 14:37:27 -0700 | [diff] [blame] | 591 | u32 need = fastopen->cookie.len; |
| 592 | |
| 593 | need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE : |
| 594 | TCPOLEN_FASTOPEN_BASE; |
Yuchung Cheng | 783237e | 2012-07-19 06:43:07 +0000 | [diff] [blame] | 595 | need = (need + 3) & ~3U; /* Align to 32 bits */ |
| 596 | if (remaining >= need) { |
| 597 | opts->options |= OPTION_FAST_OPEN_COOKIE; |
| 598 | opts->fastopen_cookie = &fastopen->cookie; |
| 599 | remaining -= need; |
| 600 | tp->syn_fastopen = 1; |
Daniel Lee | 2646c83 | 2015-04-06 14:37:27 -0700 | [diff] [blame] | 601 | tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0; |
Yuchung Cheng | 783237e | 2012-07-19 06:43:07 +0000 | [diff] [blame] | 602 | } |
| 603 | } |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 604 | |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 605 | return MAX_TCP_OPTION_SPACE - remaining; |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 606 | } |
| 607 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 608 | /* Set up TCP options for SYN-ACKs. */ |
Eric Dumazet | 37bfbdd | 2015-09-25 07:39:17 -0700 | [diff] [blame] | 609 | static unsigned int tcp_synack_options(struct request_sock *req, |
| 610 | unsigned int mss, struct sk_buff *skb, |
| 611 | struct tcp_out_options *opts, |
| 612 | const struct tcp_md5sig_key *md5, |
| 613 | struct tcp_fastopen_cookie *foc) |
William Allen Simpson | 4957faade | 2009-12-02 18:25:27 +0000 | [diff] [blame] | 614 | { |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 615 | struct inet_request_sock *ireq = inet_rsk(req); |
Eric Dumazet | 95c9617 | 2012-04-15 05:58:06 +0000 | [diff] [blame] | 616 | unsigned int remaining = MAX_TCP_OPTION_SPACE; |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 617 | |
| 618 | #ifdef CONFIG_TCP_MD5SIG |
Eric Dumazet | 80f03e2 | 2015-03-24 15:58:52 -0700 | [diff] [blame] | 619 | if (md5) { |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 620 | opts->options |= OPTION_MD5; |
William Allen Simpson | 4957faade | 2009-12-02 18:25:27 +0000 | [diff] [blame] | 621 | remaining -= TCPOLEN_MD5SIG_ALIGNED; |
| 622 | |
| 623 | /* We can't fit any SACK blocks in a packet with MD5 + TS |
| 624 | * options. There was discussion about disabling SACK |
| 625 | * rather than TS in order to fit in better with old, |
| 626 | * buggy kernels, but that was deemed to be unnecessary. |
| 627 | */ |
Eric Dumazet | de213e5e | 2010-05-17 22:35:36 -0700 | [diff] [blame] | 628 | ireq->tstamp_ok &= !ireq->sack_ok; |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 629 | } |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 630 | #endif |
| 631 | |
William Allen Simpson | 4957faade | 2009-12-02 18:25:27 +0000 | [diff] [blame] | 632 | /* We always send an MSS option. */ |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 633 | opts->mss = mss; |
William Allen Simpson | 4957faade | 2009-12-02 18:25:27 +0000 | [diff] [blame] | 634 | remaining -= TCPOLEN_MSS_ALIGNED; |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 635 | |
| 636 | if (likely(ireq->wscale_ok)) { |
| 637 | opts->ws = ireq->rcv_wscale; |
Ori Finkelman | 89e95a6 | 2009-10-01 06:41:59 +0000 | [diff] [blame] | 638 | opts->options |= OPTION_WSCALE; |
William Allen Simpson | 4957faade | 2009-12-02 18:25:27 +0000 | [diff] [blame] | 639 | remaining -= TCPOLEN_WSCALE_ALIGNED; |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 640 | } |
Eric Dumazet | de213e5e | 2010-05-17 22:35:36 -0700 | [diff] [blame] | 641 | if (likely(ireq->tstamp_ok)) { |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 642 | opts->options |= OPTION_TS; |
Eric Dumazet | 7faee5c | 2014-09-05 15:33:33 -0700 | [diff] [blame] | 643 | opts->tsval = tcp_skb_timestamp(skb); |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 644 | opts->tsecr = req->ts_recent; |
William Allen Simpson | 4957faade | 2009-12-02 18:25:27 +0000 | [diff] [blame] | 645 | remaining -= TCPOLEN_TSTAMP_ALIGNED; |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 646 | } |
| 647 | if (likely(ireq->sack_ok)) { |
| 648 | opts->options |= OPTION_SACK_ADVERTISE; |
Eric Dumazet | de213e5e | 2010-05-17 22:35:36 -0700 | [diff] [blame] | 649 | if (unlikely(!ireq->tstamp_ok)) |
William Allen Simpson | 4957faade | 2009-12-02 18:25:27 +0000 | [diff] [blame] | 650 | remaining -= TCPOLEN_SACKPERM_ALIGNED; |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 651 | } |
Daniel Lee | 7f9b838 | 2015-04-06 14:37:26 -0700 | [diff] [blame] | 652 | if (foc != NULL && foc->len >= 0) { |
| 653 | u32 need = foc->len; |
| 654 | |
| 655 | need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE : |
| 656 | TCPOLEN_FASTOPEN_BASE; |
Jerry Chu | 8336886 | 2012-08-31 12:29:12 +0000 | [diff] [blame] | 657 | need = (need + 3) & ~3U; /* Align to 32 bits */ |
| 658 | if (remaining >= need) { |
| 659 | opts->options |= OPTION_FAST_OPEN_COOKIE; |
| 660 | opts->fastopen_cookie = foc; |
| 661 | remaining -= need; |
| 662 | } |
| 663 | } |
William Allen Simpson | 4957faade | 2009-12-02 18:25:27 +0000 | [diff] [blame] | 664 | |
William Allen Simpson | 4957faade | 2009-12-02 18:25:27 +0000 | [diff] [blame] | 665 | return MAX_TCP_OPTION_SPACE - remaining; |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 666 | } |
| 667 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 668 | /* Compute TCP options for ESTABLISHED sockets. This is not the |
| 669 | * final wire format yet. |
| 670 | */ |
Eric Dumazet | 95c9617 | 2012-04-15 05:58:06 +0000 | [diff] [blame] | 671 | static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb, |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 672 | struct tcp_out_options *opts, |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 673 | struct tcp_md5sig_key **md5) |
| 674 | { |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 675 | struct tcp_sock *tp = tcp_sk(sk); |
Eric Dumazet | 95c9617 | 2012-04-15 05:58:06 +0000 | [diff] [blame] | 676 | unsigned int size = 0; |
Ilpo Järvinen | cabeccb | 2009-02-28 04:44:38 +0000 | [diff] [blame] | 677 | unsigned int eff_sacks; |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 678 | |
Andi Kleen | 5843ef4 | 2013-09-30 13:29:11 -0700 | [diff] [blame] | 679 | opts->options = 0; |
| 680 | |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 681 | #ifdef CONFIG_TCP_MD5SIG |
| 682 | *md5 = tp->af_specific->md5_lookup(sk, sk); |
| 683 | if (unlikely(*md5)) { |
| 684 | opts->options |= OPTION_MD5; |
| 685 | size += TCPOLEN_MD5SIG_ALIGNED; |
| 686 | } |
| 687 | #else |
| 688 | *md5 = NULL; |
| 689 | #endif |
| 690 | |
| 691 | if (likely(tp->rx_opt.tstamp_ok)) { |
| 692 | opts->options |= OPTION_TS; |
Eric Dumazet | 7faee5c | 2014-09-05 15:33:33 -0700 | [diff] [blame] | 693 | opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0; |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 694 | opts->tsecr = tp->rx_opt.ts_recent; |
| 695 | size += TCPOLEN_TSTAMP_ALIGNED; |
| 696 | } |
| 697 | |
Ilpo Järvinen | cabeccb | 2009-02-28 04:44:38 +0000 | [diff] [blame] | 698 | eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; |
| 699 | if (unlikely(eff_sacks)) { |
Eric Dumazet | 95c9617 | 2012-04-15 05:58:06 +0000 | [diff] [blame] | 700 | const unsigned int remaining = MAX_TCP_OPTION_SPACE - size; |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 701 | opts->num_sack_blocks = |
Eric Dumazet | 95c9617 | 2012-04-15 05:58:06 +0000 | [diff] [blame] | 702 | min_t(unsigned int, eff_sacks, |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 703 | (remaining - TCPOLEN_SACK_BASE_ALIGNED) / |
| 704 | TCPOLEN_SACK_PERBLOCK); |
| 705 | size += TCPOLEN_SACK_BASE_ALIGNED + |
| 706 | opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; |
| 707 | } |
| 708 | |
| 709 | return size; |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 710 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 711 | |
Eric Dumazet | 46d3cea | 2012-07-11 05:50:31 +0000 | [diff] [blame] | 712 | |
| 713 | /* TCP SMALL QUEUES (TSQ) |
| 714 | * |
| 715 | * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev) |
| 716 | * to reduce RTT and bufferbloat. |
| 717 | * We do this using a special skb destructor (tcp_wfree). |
| 718 | * |
| 719 | * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb |
| 720 | * needs to be reallocated in a driver. |
stephen hemminger | 8e3bff9 | 2013-12-08 12:15:44 -0800 | [diff] [blame] | 721 | * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc |
Eric Dumazet | 46d3cea | 2012-07-11 05:50:31 +0000 | [diff] [blame] | 722 | * |
| 723 | * Since transmit from skb destructor is forbidden, we use a tasklet |
| 724 | * to process all sockets that eventually need to send more skbs. |
| 725 | * We use one tasklet per cpu, with its own queue of sockets. |
| 726 | */ |
| 727 | struct tsq_tasklet { |
| 728 | struct tasklet_struct tasklet; |
| 729 | struct list_head head; /* queue of tcp sockets */ |
| 730 | }; |
| 731 | static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet); |
| 732 | |
Eric Dumazet | 6f458df | 2012-07-20 05:45:50 +0000 | [diff] [blame] | 733 | static void tcp_tsq_handler(struct sock *sk) |
| 734 | { |
| 735 | if ((1 << sk->sk_state) & |
| 736 | (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING | |
| 737 | TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) |
John Ogness | bf06200 | 2014-02-09 18:40:11 -0800 | [diff] [blame] | 738 | tcp_write_xmit(sk, tcp_current_mss(sk), tcp_sk(sk)->nonagle, |
| 739 | 0, GFP_ATOMIC); |
Eric Dumazet | 6f458df | 2012-07-20 05:45:50 +0000 | [diff] [blame] | 740 | } |
Eric Dumazet | 46d3cea | 2012-07-11 05:50:31 +0000 | [diff] [blame] | 741 | /* |
stephen hemminger | 8e3bff9 | 2013-12-08 12:15:44 -0800 | [diff] [blame] | 742 | * One tasklet per cpu tries to send more skbs. |
Eric Dumazet | 46d3cea | 2012-07-11 05:50:31 +0000 | [diff] [blame] | 743 | * We run in tasklet context but need to disable irqs when |
stephen hemminger | 8e3bff9 | 2013-12-08 12:15:44 -0800 | [diff] [blame] | 744 | * transferring tsq->head because tcp_wfree() might |
Eric Dumazet | 46d3cea | 2012-07-11 05:50:31 +0000 | [diff] [blame] | 745 | * interrupt us (non NAPI drivers) |
| 746 | */ |
| 747 | static void tcp_tasklet_func(unsigned long data) |
| 748 | { |
| 749 | struct tsq_tasklet *tsq = (struct tsq_tasklet *)data; |
| 750 | LIST_HEAD(list); |
| 751 | unsigned long flags; |
| 752 | struct list_head *q, *n; |
| 753 | struct tcp_sock *tp; |
| 754 | struct sock *sk; |
| 755 | |
| 756 | local_irq_save(flags); |
| 757 | list_splice_init(&tsq->head, &list); |
| 758 | local_irq_restore(flags); |
| 759 | |
| 760 | list_for_each_safe(q, n, &list) { |
| 761 | tp = list_entry(q, struct tcp_sock, tsq_node); |
| 762 | list_del(&tp->tsq_node); |
| 763 | |
| 764 | sk = (struct sock *)tp; |
| 765 | bh_lock_sock(sk); |
| 766 | |
| 767 | if (!sock_owned_by_user(sk)) { |
Eric Dumazet | 6f458df | 2012-07-20 05:45:50 +0000 | [diff] [blame] | 768 | tcp_tsq_handler(sk); |
Eric Dumazet | 46d3cea | 2012-07-11 05:50:31 +0000 | [diff] [blame] | 769 | } else { |
| 770 | /* defer the work to tcp_release_cb() */ |
Eric Dumazet | 6f458df | 2012-07-20 05:45:50 +0000 | [diff] [blame] | 771 | set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags); |
Eric Dumazet | 46d3cea | 2012-07-11 05:50:31 +0000 | [diff] [blame] | 772 | } |
| 773 | bh_unlock_sock(sk); |
| 774 | |
| 775 | clear_bit(TSQ_QUEUED, &tp->tsq_flags); |
| 776 | sk_free(sk); |
| 777 | } |
| 778 | } |
| 779 | |
Eric Dumazet | 6f458df | 2012-07-20 05:45:50 +0000 | [diff] [blame] | 780 | #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \ |
| 781 | (1UL << TCP_WRITE_TIMER_DEFERRED) | \ |
Eric Dumazet | 563d34d | 2012-07-23 09:48:52 +0200 | [diff] [blame] | 782 | (1UL << TCP_DELACK_TIMER_DEFERRED) | \ |
| 783 | (1UL << TCP_MTU_REDUCED_DEFERRED)) |
Eric Dumazet | 46d3cea | 2012-07-11 05:50:31 +0000 | [diff] [blame] | 784 | /** |
| 785 | * tcp_release_cb - tcp release_sock() callback |
| 786 | * @sk: socket |
| 787 | * |
| 788 | * called from release_sock() to perform protocol dependent |
| 789 | * actions before socket release. |
| 790 | */ |
| 791 | void tcp_release_cb(struct sock *sk) |
| 792 | { |
| 793 | struct tcp_sock *tp = tcp_sk(sk); |
Eric Dumazet | 6f458df | 2012-07-20 05:45:50 +0000 | [diff] [blame] | 794 | unsigned long flags, nflags; |
Eric Dumazet | 46d3cea | 2012-07-11 05:50:31 +0000 | [diff] [blame] | 795 | |
Eric Dumazet | 6f458df | 2012-07-20 05:45:50 +0000 | [diff] [blame] | 796 | /* perform an atomic operation only if at least one flag is set */ |
| 797 | do { |
| 798 | flags = tp->tsq_flags; |
| 799 | if (!(flags & TCP_DEFERRED_ALL)) |
| 800 | return; |
| 801 | nflags = flags & ~TCP_DEFERRED_ALL; |
| 802 | } while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags); |
| 803 | |
| 804 | if (flags & (1UL << TCP_TSQ_DEFERRED)) |
| 805 | tcp_tsq_handler(sk); |
| 806 | |
Eric Dumazet | c3f9b01 | 2014-03-10 09:50:11 -0700 | [diff] [blame] | 807 | /* Here begins the tricky part : |
| 808 | * We are called from release_sock() with : |
| 809 | * 1) BH disabled |
| 810 | * 2) sk_lock.slock spinlock held |
| 811 | * 3) socket owned by us (sk->sk_lock.owned == 1) |
| 812 | * |
| 813 | * But following code is meant to be called from BH handlers, |
| 814 | * so we should keep BH disabled, but early release socket ownership |
| 815 | */ |
| 816 | sock_release_ownership(sk); |
| 817 | |
Eric Dumazet | 144d56e | 2012-08-20 00:22:46 +0000 | [diff] [blame] | 818 | if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) { |
Eric Dumazet | 6f458df | 2012-07-20 05:45:50 +0000 | [diff] [blame] | 819 | tcp_write_timer_handler(sk); |
Eric Dumazet | 144d56e | 2012-08-20 00:22:46 +0000 | [diff] [blame] | 820 | __sock_put(sk); |
| 821 | } |
| 822 | if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) { |
Eric Dumazet | 6f458df | 2012-07-20 05:45:50 +0000 | [diff] [blame] | 823 | tcp_delack_timer_handler(sk); |
Eric Dumazet | 144d56e | 2012-08-20 00:22:46 +0000 | [diff] [blame] | 824 | __sock_put(sk); |
| 825 | } |
| 826 | if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) { |
Neal Cardwell | 4fab907 | 2014-08-14 12:40:05 -0400 | [diff] [blame] | 827 | inet_csk(sk)->icsk_af_ops->mtu_reduced(sk); |
Eric Dumazet | 144d56e | 2012-08-20 00:22:46 +0000 | [diff] [blame] | 828 | __sock_put(sk); |
| 829 | } |
Eric Dumazet | 46d3cea | 2012-07-11 05:50:31 +0000 | [diff] [blame] | 830 | } |
| 831 | EXPORT_SYMBOL(tcp_release_cb); |
| 832 | |
| 833 | void __init tcp_tasklet_init(void) |
| 834 | { |
| 835 | int i; |
| 836 | |
| 837 | for_each_possible_cpu(i) { |
| 838 | struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i); |
| 839 | |
| 840 | INIT_LIST_HEAD(&tsq->head); |
| 841 | tasklet_init(&tsq->tasklet, |
| 842 | tcp_tasklet_func, |
| 843 | (unsigned long)tsq); |
| 844 | } |
| 845 | } |
| 846 | |
| 847 | /* |
| 848 | * Write buffer destructor automatically called from kfree_skb. |
stephen hemminger | 8e3bff9 | 2013-12-08 12:15:44 -0800 | [diff] [blame] | 849 | * We can't xmit new skbs from this context, as we might already |
Eric Dumazet | 46d3cea | 2012-07-11 05:50:31 +0000 | [diff] [blame] | 850 | * hold qdisc lock. |
| 851 | */ |
Eric Dumazet | d6a4a10 | 2013-04-12 11:31:52 +0000 | [diff] [blame] | 852 | void tcp_wfree(struct sk_buff *skb) |
Eric Dumazet | 46d3cea | 2012-07-11 05:50:31 +0000 | [diff] [blame] | 853 | { |
| 854 | struct sock *sk = skb->sk; |
| 855 | struct tcp_sock *tp = tcp_sk(sk); |
Eric Dumazet | 9b462d0 | 2014-10-13 06:27:47 -0700 | [diff] [blame] | 856 | int wmem; |
| 857 | |
| 858 | /* Keep one reference on sk_wmem_alloc. |
| 859 | * Will be released by sk_free() from here or tcp_tasklet_func() |
| 860 | */ |
| 861 | wmem = atomic_sub_return(skb->truesize - 1, &sk->sk_wmem_alloc); |
| 862 | |
| 863 | /* If this softirq is serviced by ksoftirqd, we are likely under stress. |
| 864 | * Wait until our queues (qdisc + devices) are drained. |
| 865 | * This gives : |
| 866 | * - less callbacks to tcp_write_xmit(), reducing stress (batches) |
| 867 | * - chance for incoming ACK (processed by another cpu maybe) |
| 868 | * to migrate this flow (skb->ooo_okay will be eventually set) |
| 869 | */ |
| 870 | if (wmem >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current) |
| 871 | goto out; |
Eric Dumazet | 46d3cea | 2012-07-11 05:50:31 +0000 | [diff] [blame] | 872 | |
| 873 | if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) && |
| 874 | !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) { |
| 875 | unsigned long flags; |
| 876 | struct tsq_tasklet *tsq; |
| 877 | |
Eric Dumazet | 46d3cea | 2012-07-11 05:50:31 +0000 | [diff] [blame] | 878 | /* queue this socket to tasklet queue */ |
| 879 | local_irq_save(flags); |
Christoph Lameter | 903ceff | 2014-08-17 12:30:35 -0500 | [diff] [blame] | 880 | tsq = this_cpu_ptr(&tsq_tasklet); |
Eric Dumazet | 46d3cea | 2012-07-11 05:50:31 +0000 | [diff] [blame] | 881 | list_add(&tp->tsq_node, &tsq->head); |
| 882 | tasklet_schedule(&tsq->tasklet); |
| 883 | local_irq_restore(flags); |
Eric Dumazet | 9b462d0 | 2014-10-13 06:27:47 -0700 | [diff] [blame] | 884 | return; |
Eric Dumazet | 46d3cea | 2012-07-11 05:50:31 +0000 | [diff] [blame] | 885 | } |
Eric Dumazet | 9b462d0 | 2014-10-13 06:27:47 -0700 | [diff] [blame] | 886 | out: |
| 887 | sk_free(sk); |
Eric Dumazet | 46d3cea | 2012-07-11 05:50:31 +0000 | [diff] [blame] | 888 | } |
| 889 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 890 | /* This routine actually transmits TCP packets queued in by |
| 891 | * tcp_do_sendmsg(). This is used by both the initial |
| 892 | * transmission and possible later retransmissions. |
| 893 | * All SKB's seen here are completely headerless. It is our |
| 894 | * job to build the TCP header, and pass the packet down to |
| 895 | * IP so it can do the same plus pass the packet off to the |
| 896 | * device. |
| 897 | * |
| 898 | * We are working here with either a clone of the original |
| 899 | * SKB, or a fresh unique copy made by the retransmit engine. |
| 900 | */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 901 | static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, |
| 902 | gfp_t gfp_mask) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 903 | { |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 904 | const struct inet_connection_sock *icsk = inet_csk(sk); |
| 905 | struct inet_sock *inet; |
| 906 | struct tcp_sock *tp; |
| 907 | struct tcp_skb_cb *tcb; |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 908 | struct tcp_out_options opts; |
Eric Dumazet | 95c9617 | 2012-04-15 05:58:06 +0000 | [diff] [blame] | 909 | unsigned int tcp_options_size, tcp_header_size; |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 910 | struct tcp_md5sig_key *md5; |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 911 | struct tcphdr *th; |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 912 | int err; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 913 | |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 914 | BUG_ON(!skb || !tcp_skb_pcount(skb)); |
Lawrence Brakmo | 6f094b9 | 2016-06-08 21:16:44 -0700 | [diff] [blame] | 915 | tp = tcp_sk(sk); |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 916 | |
Eric Dumazet | ccdbb6e | 2013-10-10 08:43:00 -0700 | [diff] [blame] | 917 | if (clone_it) { |
Eric Dumazet | 740b0f1 | 2014-02-26 14:02:48 -0800 | [diff] [blame] | 918 | skb_mstamp_get(&skb->skb_mstamp); |
Lawrence Brakmo | 6f094b9 | 2016-06-08 21:16:44 -0700 | [diff] [blame] | 919 | TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq |
| 920 | - tp->snd_una; |
Eric Dumazet | ccdbb6e | 2013-10-10 08:43:00 -0700 | [diff] [blame] | 921 | |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 922 | if (unlikely(skb_cloned(skb))) |
| 923 | skb = pskb_copy(skb, gfp_mask); |
| 924 | else |
| 925 | skb = skb_clone(skb, gfp_mask); |
| 926 | if (unlikely(!skb)) |
| 927 | return -ENOBUFS; |
| 928 | } |
| 929 | |
| 930 | inet = inet_sk(sk); |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 931 | tcb = TCP_SKB_CB(skb); |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 932 | memset(&opts, 0, sizeof(opts)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 933 | |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 934 | if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 935 | tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); |
| 936 | else |
| 937 | tcp_options_size = tcp_established_options(sk, skb, &opts, |
| 938 | &md5); |
| 939 | tcp_header_size = tcp_options_size + sizeof(struct tcphdr); |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 940 | |
Eric Dumazet | 547669d | 2013-05-23 07:44:20 +0000 | [diff] [blame] | 941 | /* if no packet is in qdisc/device queue, then allow XPS to select |
Eric Dumazet | b2532eb | 2014-10-10 18:06:35 -0700 | [diff] [blame] | 942 | * another queue. We can be called from tcp_tsq_handler() |
| 943 | * which holds one reference to sk_wmem_alloc. |
| 944 | * |
| 945 | * TODO: Ideally, in-flight pure ACK packets should not matter here. |
| 946 | * One way to get this would be to set skb->truesize = 2 on them. |
Eric Dumazet | 547669d | 2013-05-23 07:44:20 +0000 | [diff] [blame] | 947 | */ |
Eric Dumazet | b2532eb | 2014-10-10 18:06:35 -0700 | [diff] [blame] | 948 | skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1); |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 949 | |
Arnaldo Carvalho de Melo | aa8223c | 2007-04-10 21:04:22 -0700 | [diff] [blame] | 950 | skb_push(skb, tcp_header_size); |
| 951 | skb_reset_transport_header(skb); |
Eric Dumazet | 46d3cea | 2012-07-11 05:50:31 +0000 | [diff] [blame] | 952 | |
| 953 | skb_orphan(skb); |
| 954 | skb->sk = sk; |
Eric Dumazet | 1d2077a | 2016-05-02 10:56:27 -0700 | [diff] [blame] | 955 | skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree; |
Tom Herbert | b73c3d0 | 2014-07-01 21:32:17 -0700 | [diff] [blame] | 956 | skb_set_hash_from_sk(skb, sk); |
Eric Dumazet | 46d3cea | 2012-07-11 05:50:31 +0000 | [diff] [blame] | 957 | atomic_add(skb->truesize, &sk->sk_wmem_alloc); |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 958 | |
| 959 | /* Build TCP header and checksum it. */ |
Eric Dumazet | ea1627c | 2016-05-13 09:16:40 -0700 | [diff] [blame] | 960 | th = (struct tcphdr *)skb->data; |
Eric Dumazet | c720c7e | 2009-10-15 06:30:45 +0000 | [diff] [blame] | 961 | th->source = inet->inet_sport; |
| 962 | th->dest = inet->inet_dport; |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 963 | th->seq = htonl(tcb->seq); |
| 964 | th->ack_seq = htonl(tp->rcv_nxt); |
Al Viro | df7a3b0 | 2006-09-27 18:38:52 -0700 | [diff] [blame] | 965 | *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 966 | tcb->tcp_flags); |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 967 | |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 968 | th->check = 0; |
| 969 | th->urg_ptr = 0; |
| 970 | |
Ilpo Järvinen | 33f5f57 | 2008-10-07 14:43:06 -0700 | [diff] [blame] | 971 | /* The urg_mode check is necessary during a below snd_una win probe */ |
Herbert Xu | 7691367 | 2009-02-21 23:52:29 -0800 | [diff] [blame] | 972 | if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { |
| 973 | if (before(tp->snd_up, tcb->seq + 0x10000)) { |
| 974 | th->urg_ptr = htons(tp->snd_up - tcb->seq); |
| 975 | th->urg = 1; |
| 976 | } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { |
Eric Dumazet | 0eae88f | 2010-04-20 19:06:52 -0700 | [diff] [blame] | 977 | th->urg_ptr = htons(0xFFFF); |
Herbert Xu | 7691367 | 2009-02-21 23:52:29 -0800 | [diff] [blame] | 978 | th->urg = 1; |
| 979 | } |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 980 | } |
| 981 | |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 982 | tcp_options_write((__be32 *)(th + 1), tp, &opts); |
Eric Dumazet | 51466a7 | 2015-06-11 09:15:16 -0700 | [diff] [blame] | 983 | skb_shinfo(skb)->gso_type = sk->sk_gso_type; |
Eric Dumazet | ea1627c | 2016-05-13 09:16:40 -0700 | [diff] [blame] | 984 | if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) { |
| 985 | th->window = htons(tcp_select_window(sk)); |
| 986 | tcp_ecn_send(sk, skb, th, tcp_header_size); |
| 987 | } else { |
| 988 | /* RFC1323: The window in SYN & SYN/ACK segments |
| 989 | * is never scaled. |
| 990 | */ |
| 991 | th->window = htons(min(tp->rcv_wnd, 65535U)); |
| 992 | } |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 993 | #ifdef CONFIG_TCP_MD5SIG |
| 994 | /* Calculate the MD5 hash, as we have all we need now */ |
| 995 | if (md5) { |
Eric Dumazet | a465419 | 2010-05-16 00:36:33 -0700 | [diff] [blame] | 996 | sk_nocaps_add(sk, NETIF_F_GSO_MASK); |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 997 | tp->af_specific->calc_md5_hash(opts.hash_location, |
Eric Dumazet | 39f8e58 | 2015-03-24 15:58:55 -0700 | [diff] [blame] | 998 | md5, sk, skb); |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 999 | } |
| 1000 | #endif |
| 1001 | |
Herbert Xu | bb29624 | 2010-04-11 02:15:55 +0000 | [diff] [blame] | 1002 | icsk->icsk_af_ops->send_check(sk, skb); |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 1003 | |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 1004 | if (likely(tcb->tcp_flags & TCPHDR_ACK)) |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 1005 | tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); |
| 1006 | |
Martin KaFai Lau | a44d6ea | 2016-03-14 10:52:15 -0700 | [diff] [blame] | 1007 | if (skb->len != tcp_header_size) { |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 1008 | tcp_event_data_sent(tp, sk); |
Martin KaFai Lau | a44d6ea | 2016-03-14 10:52:15 -0700 | [diff] [blame] | 1009 | tp->data_segs_out += tcp_skb_pcount(skb); |
| 1010 | } |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 1011 | |
Wei Yongjun | bd37a08 | 2006-08-07 21:04:15 -0700 | [diff] [blame] | 1012 | if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) |
Tom Herbert | aa2ea05 | 2010-04-22 07:00:24 +0000 | [diff] [blame] | 1013 | TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, |
| 1014 | tcp_skb_pcount(skb)); |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 1015 | |
Marcelo Ricardo Leitner | 2efd055 | 2015-05-20 16:35:41 -0700 | [diff] [blame] | 1016 | tp->segs_out += tcp_skb_pcount(skb); |
Eric Dumazet | f69ad29 | 2015-06-11 09:15:18 -0700 | [diff] [blame] | 1017 | /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */ |
Eric Dumazet | cd7d849 | 2014-09-24 04:11:22 -0700 | [diff] [blame] | 1018 | skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb); |
Eric Dumazet | f69ad29 | 2015-06-11 09:15:18 -0700 | [diff] [blame] | 1019 | skb_shinfo(skb)->gso_size = tcp_skb_mss(skb); |
Eric Dumazet | cd7d849 | 2014-09-24 04:11:22 -0700 | [diff] [blame] | 1020 | |
Eric Dumazet | 7faee5c | 2014-09-05 15:33:33 -0700 | [diff] [blame] | 1021 | /* Our usage of tstamp should remain private */ |
| 1022 | skb->tstamp.tv64 = 0; |
Eric Dumazet | 971f10e | 2014-09-27 09:50:57 -0700 | [diff] [blame] | 1023 | |
| 1024 | /* Cleanup our debris for IP stacks */ |
| 1025 | memset(skb->cb, 0, max(sizeof(struct inet_skb_parm), |
| 1026 | sizeof(struct inet6_skb_parm))); |
| 1027 | |
Eric Dumazet | b0270e9 | 2014-04-15 12:58:34 -0400 | [diff] [blame] | 1028 | err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl); |
Eric Dumazet | 7faee5c | 2014-09-05 15:33:33 -0700 | [diff] [blame] | 1029 | |
Hua Zhong | 83de47c | 2006-04-28 15:26:50 -0700 | [diff] [blame] | 1030 | if (likely(err <= 0)) |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 1031 | return err; |
| 1032 | |
Christoph Paasch | 5ee2c94 | 2014-07-14 16:58:32 +0200 | [diff] [blame] | 1033 | tcp_enter_cwr(sk); |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 1034 | |
Gerrit Renker | b9df3cb | 2006-11-14 11:21:36 -0200 | [diff] [blame] | 1035 | return net_xmit_eval(err); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1036 | } |
| 1037 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 1038 | /* This routine just queues the buffer for sending. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1039 | * |
| 1040 | * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, |
| 1041 | * otherwise socket can stall. |
| 1042 | */ |
| 1043 | static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) |
| 1044 | { |
| 1045 | struct tcp_sock *tp = tcp_sk(sk); |
| 1046 | |
| 1047 | /* Advance write_seq and place onto the write_queue. */ |
| 1048 | tp->write_seq = TCP_SKB_CB(skb)->end_seq; |
Eric Dumazet | f4a775d | 2014-09-22 16:29:32 -0700 | [diff] [blame] | 1049 | __skb_header_release(skb); |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1050 | tcp_add_write_queue_tail(sk, skb); |
Hideo Aoki | 3ab224b | 2007-12-31 00:11:19 -0800 | [diff] [blame] | 1051 | sk->sk_wmem_queued += skb->truesize; |
| 1052 | sk_mem_charge(sk, skb->truesize); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1053 | } |
| 1054 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 1055 | /* Initialize TSO segments for a packet. */ |
Eric Dumazet | 5bbb432 | 2015-06-11 09:15:17 -0700 | [diff] [blame] | 1056 | static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now) |
David S. Miller | f6302d1 | 2005-07-05 15:18:03 -0700 | [diff] [blame] | 1057 | { |
Eric Dumazet | 8f26fb1 | 2013-10-15 12:24:54 -0700 | [diff] [blame] | 1058 | if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) { |
David S. Miller | f6302d1 | 2005-07-05 15:18:03 -0700 | [diff] [blame] | 1059 | /* Avoid the costly divide in the normal |
| 1060 | * non-TSO case. |
| 1061 | */ |
Eric Dumazet | cd7d849 | 2014-09-24 04:11:22 -0700 | [diff] [blame] | 1062 | tcp_skb_pcount_set(skb, 1); |
Eric Dumazet | f69ad29 | 2015-06-11 09:15:18 -0700 | [diff] [blame] | 1063 | TCP_SKB_CB(skb)->tcp_gso_size = 0; |
David S. Miller | f6302d1 | 2005-07-05 15:18:03 -0700 | [diff] [blame] | 1064 | } else { |
Eric Dumazet | cd7d849 | 2014-09-24 04:11:22 -0700 | [diff] [blame] | 1065 | tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now)); |
Eric Dumazet | f69ad29 | 2015-06-11 09:15:18 -0700 | [diff] [blame] | 1066 | TCP_SKB_CB(skb)->tcp_gso_size = mss_now; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1067 | } |
| 1068 | } |
| 1069 | |
Ilpo Järvinen | 91fed7a | 2007-10-09 01:24:15 -0700 | [diff] [blame] | 1070 | /* When a modification to fackets out becomes necessary, we need to check |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1071 | * skb is counted to fackets_out or not. |
Ilpo Järvinen | 91fed7a | 2007-10-09 01:24:15 -0700 | [diff] [blame] | 1072 | */ |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 1073 | static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb, |
Ilpo Järvinen | 91fed7a | 2007-10-09 01:24:15 -0700 | [diff] [blame] | 1074 | int decr) |
| 1075 | { |
Ilpo Järvinen | a47e5a9 | 2007-11-15 19:41:46 -0800 | [diff] [blame] | 1076 | struct tcp_sock *tp = tcp_sk(sk); |
| 1077 | |
Ilpo Järvinen | dc86967 | 2007-10-01 15:27:19 -0700 | [diff] [blame] | 1078 | if (!tp->sacked_out || tcp_is_reno(tp)) |
Ilpo Järvinen | 91fed7a | 2007-10-09 01:24:15 -0700 | [diff] [blame] | 1079 | return; |
| 1080 | |
Ilpo Järvinen | 6859d49 | 2007-12-02 00:48:06 +0200 | [diff] [blame] | 1081 | if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq)) |
Ilpo Järvinen | 91fed7a | 2007-10-09 01:24:15 -0700 | [diff] [blame] | 1082 | tp->fackets_out -= decr; |
Ilpo Järvinen | 91fed7a | 2007-10-09 01:24:15 -0700 | [diff] [blame] | 1083 | } |
| 1084 | |
Ilpo Järvinen | 797108d | 2009-04-01 23:15:17 +0000 | [diff] [blame] | 1085 | /* Pcount in the middle of the write queue got changed, we need to do various |
| 1086 | * tweaks to fix counters |
| 1087 | */ |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 1088 | static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) |
Ilpo Järvinen | 797108d | 2009-04-01 23:15:17 +0000 | [diff] [blame] | 1089 | { |
| 1090 | struct tcp_sock *tp = tcp_sk(sk); |
| 1091 | |
| 1092 | tp->packets_out -= decr; |
| 1093 | |
| 1094 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) |
| 1095 | tp->sacked_out -= decr; |
| 1096 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) |
| 1097 | tp->retrans_out -= decr; |
| 1098 | if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) |
| 1099 | tp->lost_out -= decr; |
| 1100 | |
| 1101 | /* Reno case is special. Sigh... */ |
| 1102 | if (tcp_is_reno(tp) && decr > 0) |
| 1103 | tp->sacked_out -= min_t(u32, tp->sacked_out, decr); |
| 1104 | |
| 1105 | tcp_adjust_fackets_out(sk, skb, decr); |
| 1106 | |
| 1107 | if (tp->lost_skb_hint && |
| 1108 | before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && |
Ilpo Järvinen | 52cf3cc | 2009-04-18 05:48:48 +0000 | [diff] [blame] | 1109 | (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))) |
Ilpo Järvinen | 797108d | 2009-04-01 23:15:17 +0000 | [diff] [blame] | 1110 | tp->lost_cnt_hint -= decr; |
| 1111 | |
| 1112 | tcp_verify_left_out(tp); |
| 1113 | } |
| 1114 | |
Soheil Hassas Yeganeh | 0a2cf20 | 2016-04-27 23:39:01 -0400 | [diff] [blame] | 1115 | static bool tcp_has_tx_tstamp(const struct sk_buff *skb) |
| 1116 | { |
| 1117 | return TCP_SKB_CB(skb)->txstamp_ack || |
| 1118 | (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP); |
| 1119 | } |
| 1120 | |
Willem de Bruijn | 490cc7d | 2014-08-12 15:08:12 -0400 | [diff] [blame] | 1121 | static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2) |
| 1122 | { |
| 1123 | struct skb_shared_info *shinfo = skb_shinfo(skb); |
| 1124 | |
Soheil Hassas Yeganeh | 0a2cf20 | 2016-04-27 23:39:01 -0400 | [diff] [blame] | 1125 | if (unlikely(tcp_has_tx_tstamp(skb)) && |
Willem de Bruijn | 490cc7d | 2014-08-12 15:08:12 -0400 | [diff] [blame] | 1126 | !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) { |
| 1127 | struct skb_shared_info *shinfo2 = skb_shinfo(skb2); |
| 1128 | u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP; |
| 1129 | |
| 1130 | shinfo->tx_flags &= ~tsflags; |
| 1131 | shinfo2->tx_flags |= tsflags; |
| 1132 | swap(shinfo->tskey, shinfo2->tskey); |
Martin KaFai Lau | b51e13f | 2016-04-19 22:50:47 -0700 | [diff] [blame] | 1133 | TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack; |
| 1134 | TCP_SKB_CB(skb)->txstamp_ack = 0; |
Willem de Bruijn | 490cc7d | 2014-08-12 15:08:12 -0400 | [diff] [blame] | 1135 | } |
| 1136 | } |
| 1137 | |
Martin KaFai Lau | a166140 | 2016-04-25 14:44:50 -0700 | [diff] [blame] | 1138 | static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2) |
| 1139 | { |
| 1140 | TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor; |
| 1141 | TCP_SKB_CB(skb)->eor = 0; |
| 1142 | } |
| 1143 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1144 | /* Function to create two new TCP segments. Shrinks the given segment |
| 1145 | * to the specified size and appends a new segment with the rest of the |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 1146 | * packet to the list. This won't be called frequently, I hope. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1147 | * Remember, these are still headerless SKBs at this point. |
| 1148 | */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1149 | int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, |
Octavian Purdila | 6cc55e0 | 2014-06-06 17:32:37 +0300 | [diff] [blame] | 1150 | unsigned int mss_now, gfp_t gfp) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1151 | { |
| 1152 | struct tcp_sock *tp = tcp_sk(sk); |
| 1153 | struct sk_buff *buff; |
David S. Miller | 6475be1 | 2005-09-01 22:47:01 -0700 | [diff] [blame] | 1154 | int nsize, old_factor; |
Herbert Xu | b60b49e | 2006-04-19 21:35:00 -0700 | [diff] [blame] | 1155 | int nlen; |
Ilpo Järvinen | 9ce0146 | 2009-02-28 04:44:42 +0000 | [diff] [blame] | 1156 | u8 flags; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1157 | |
Ilpo Järvinen | 2fceec1 | 2011-04-01 21:47:41 -0700 | [diff] [blame] | 1158 | if (WARN_ON(len > skb->len)) |
| 1159 | return -EINVAL; |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 1160 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1161 | nsize = skb_headlen(skb) - len; |
| 1162 | if (nsize < 0) |
| 1163 | nsize = 0; |
| 1164 | |
Octavian Purdila | 6cc55e0 | 2014-06-06 17:32:37 +0300 | [diff] [blame] | 1165 | if (skb_unclone(skb, gfp)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1166 | return -ENOMEM; |
| 1167 | |
| 1168 | /* Get a new skb... force flag on. */ |
Eric Dumazet | eb93447 | 2015-05-19 13:26:55 -0700 | [diff] [blame] | 1169 | buff = sk_stream_alloc_skb(sk, nsize, gfp, true); |
Ian Morris | 51456b2 | 2015-04-03 09:17:26 +0100 | [diff] [blame] | 1170 | if (!buff) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1171 | return -ENOMEM; /* We'll just try again later. */ |
Herbert Xu | ef5cb97 | 2006-04-18 13:24:14 -0700 | [diff] [blame] | 1172 | |
Hideo Aoki | 3ab224b | 2007-12-31 00:11:19 -0800 | [diff] [blame] | 1173 | sk->sk_wmem_queued += buff->truesize; |
| 1174 | sk_mem_charge(sk, buff->truesize); |
Herbert Xu | b60b49e | 2006-04-19 21:35:00 -0700 | [diff] [blame] | 1175 | nlen = skb->len - len - nsize; |
| 1176 | buff->truesize += nlen; |
| 1177 | skb->truesize -= nlen; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1178 | |
| 1179 | /* Correct the sequence numbers. */ |
| 1180 | TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; |
| 1181 | TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; |
| 1182 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; |
| 1183 | |
| 1184 | /* PSH and FIN should only be set in the second packet. */ |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 1185 | flags = TCP_SKB_CB(skb)->tcp_flags; |
| 1186 | TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); |
| 1187 | TCP_SKB_CB(buff)->tcp_flags = flags; |
Herbert Xu | e14c3ca | 2005-09-19 18:18:38 -0700 | [diff] [blame] | 1188 | TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; |
Martin KaFai Lau | a166140 | 2016-04-25 14:44:50 -0700 | [diff] [blame] | 1189 | tcp_skb_fragment_eor(skb, buff); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1190 | |
Patrick McHardy | 84fa793 | 2006-08-29 16:44:56 -0700 | [diff] [blame] | 1191 | if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1192 | /* Copy and checksum data tail into the new buffer. */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1193 | buff->csum = csum_partial_copy_nocheck(skb->data + len, |
| 1194 | skb_put(buff, nsize), |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1195 | nsize, 0); |
| 1196 | |
| 1197 | skb_trim(skb, len); |
| 1198 | |
| 1199 | skb->csum = csum_block_sub(skb->csum, buff->csum, len); |
| 1200 | } else { |
Patrick McHardy | 84fa793 | 2006-08-29 16:44:56 -0700 | [diff] [blame] | 1201 | skb->ip_summed = CHECKSUM_PARTIAL; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1202 | skb_split(skb, buff, len); |
| 1203 | } |
| 1204 | |
| 1205 | buff->ip_summed = skb->ip_summed; |
| 1206 | |
Patrick McHardy | a61bbcf | 2005-08-14 17:24:31 -0700 | [diff] [blame] | 1207 | buff->tstamp = skb->tstamp; |
Willem de Bruijn | 490cc7d | 2014-08-12 15:08:12 -0400 | [diff] [blame] | 1208 | tcp_fragment_tstamp(skb, buff); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1209 | |
David S. Miller | 6475be1 | 2005-09-01 22:47:01 -0700 | [diff] [blame] | 1210 | old_factor = tcp_skb_pcount(skb); |
| 1211 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1212 | /* Fix up tso_factor for both original and new SKB. */ |
Eric Dumazet | 5bbb432 | 2015-06-11 09:15:17 -0700 | [diff] [blame] | 1213 | tcp_set_skb_tso_segs(skb, mss_now); |
| 1214 | tcp_set_skb_tso_segs(buff, mss_now); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1215 | |
David S. Miller | 6475be1 | 2005-09-01 22:47:01 -0700 | [diff] [blame] | 1216 | /* If this packet has been sent out already, we must |
| 1217 | * adjust the various packet counters. |
| 1218 | */ |
Herbert Xu | cf0b450 | 2005-09-08 15:10:52 -0700 | [diff] [blame] | 1219 | if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { |
David S. Miller | 6475be1 | 2005-09-01 22:47:01 -0700 | [diff] [blame] | 1220 | int diff = old_factor - tcp_skb_pcount(skb) - |
| 1221 | tcp_skb_pcount(buff); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1222 | |
Ilpo Järvinen | 797108d | 2009-04-01 23:15:17 +0000 | [diff] [blame] | 1223 | if (diff) |
| 1224 | tcp_adjust_pcount(sk, skb, diff); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1225 | } |
| 1226 | |
| 1227 | /* Link BUFF into the send queue. */ |
Eric Dumazet | f4a775d | 2014-09-22 16:29:32 -0700 | [diff] [blame] | 1228 | __skb_header_release(buff); |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1229 | tcp_insert_write_queue_after(skb, buff, sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1230 | |
| 1231 | return 0; |
| 1232 | } |
| 1233 | |
| 1234 | /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c |
| 1235 | * eventually). The difference is that pulled data not copied, but |
| 1236 | * immediately discarded. |
| 1237 | */ |
Herbert Xu ~{PmVHI~} | f291196 | 2006-06-05 15:03:37 -0700 | [diff] [blame] | 1238 | static void __pskb_trim_head(struct sk_buff *skb, int len) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1239 | { |
Eric Dumazet | 7b7fc97 | 2013-12-05 22:31:30 -0800 | [diff] [blame] | 1240 | struct skb_shared_info *shinfo; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1241 | int i, k, eat; |
| 1242 | |
Eric Dumazet | 4fa48bf | 2011-12-04 08:51:08 +0000 | [diff] [blame] | 1243 | eat = min_t(int, len, skb_headlen(skb)); |
| 1244 | if (eat) { |
| 1245 | __skb_pull(skb, eat); |
| 1246 | len -= eat; |
| 1247 | if (!len) |
| 1248 | return; |
| 1249 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1250 | eat = len; |
| 1251 | k = 0; |
Eric Dumazet | 7b7fc97 | 2013-12-05 22:31:30 -0800 | [diff] [blame] | 1252 | shinfo = skb_shinfo(skb); |
| 1253 | for (i = 0; i < shinfo->nr_frags; i++) { |
| 1254 | int size = skb_frag_size(&shinfo->frags[i]); |
Eric Dumazet | 9e903e0 | 2011-10-18 21:00:24 +0000 | [diff] [blame] | 1255 | |
| 1256 | if (size <= eat) { |
Ian Campbell | aff65da | 2011-08-22 23:44:59 +0000 | [diff] [blame] | 1257 | skb_frag_unref(skb, i); |
Eric Dumazet | 9e903e0 | 2011-10-18 21:00:24 +0000 | [diff] [blame] | 1258 | eat -= size; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1259 | } else { |
Eric Dumazet | 7b7fc97 | 2013-12-05 22:31:30 -0800 | [diff] [blame] | 1260 | shinfo->frags[k] = shinfo->frags[i]; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1261 | if (eat) { |
Eric Dumazet | 7b7fc97 | 2013-12-05 22:31:30 -0800 | [diff] [blame] | 1262 | shinfo->frags[k].page_offset += eat; |
| 1263 | skb_frag_size_sub(&shinfo->frags[k], eat); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1264 | eat = 0; |
| 1265 | } |
| 1266 | k++; |
| 1267 | } |
| 1268 | } |
Eric Dumazet | 7b7fc97 | 2013-12-05 22:31:30 -0800 | [diff] [blame] | 1269 | shinfo->nr_frags = k; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1270 | |
Arnaldo Carvalho de Melo | 27a884d | 2007-04-19 20:29:13 -0700 | [diff] [blame] | 1271 | skb_reset_tail_pointer(skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1272 | skb->data_len -= len; |
| 1273 | skb->len = skb->data_len; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1274 | } |
| 1275 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 1276 | /* Remove acked data from a packet in the transmit queue. */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1277 | int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) |
| 1278 | { |
Pravin B Shelar | 14bbd6a | 2013-02-14 09:44:49 +0000 | [diff] [blame] | 1279 | if (skb_unclone(skb, GFP_ATOMIC)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1280 | return -ENOMEM; |
| 1281 | |
Eric Dumazet | 4fa48bf | 2011-12-04 08:51:08 +0000 | [diff] [blame] | 1282 | __pskb_trim_head(skb, len); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1283 | |
| 1284 | TCP_SKB_CB(skb)->seq += len; |
Patrick McHardy | 84fa793 | 2006-08-29 16:44:56 -0700 | [diff] [blame] | 1285 | skb->ip_summed = CHECKSUM_PARTIAL; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1286 | |
| 1287 | skb->truesize -= len; |
| 1288 | sk->sk_wmem_queued -= len; |
Hideo Aoki | 3ab224b | 2007-12-31 00:11:19 -0800 | [diff] [blame] | 1289 | sk_mem_uncharge(sk, len); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1290 | sock_set_flag(sk, SOCK_QUEUE_SHRUNK); |
| 1291 | |
Neal Cardwell | 5b35e1e | 2012-01-28 17:29:46 +0000 | [diff] [blame] | 1292 | /* Any change of skb->len requires recalculation of tso factor. */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1293 | if (tcp_skb_pcount(skb) > 1) |
Eric Dumazet | 5bbb432 | 2015-06-11 09:15:17 -0700 | [diff] [blame] | 1294 | tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1295 | |
| 1296 | return 0; |
| 1297 | } |
| 1298 | |
Yuchung Cheng | 1b63edd | 2013-02-22 08:59:06 +0000 | [diff] [blame] | 1299 | /* Calculate MSS not accounting any TCP options. */ |
| 1300 | static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu) |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1301 | { |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 1302 | const struct tcp_sock *tp = tcp_sk(sk); |
| 1303 | const struct inet_connection_sock *icsk = inet_csk(sk); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1304 | int mss_now; |
| 1305 | |
| 1306 | /* Calculate base mss without TCP options: |
| 1307 | It is MMS_S - sizeof(tcphdr) of rfc1122 |
| 1308 | */ |
| 1309 | mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); |
| 1310 | |
Eric Dumazet | 6746960 | 2012-04-24 07:37:38 +0000 | [diff] [blame] | 1311 | /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ |
| 1312 | if (icsk->icsk_af_ops->net_frag_header_len) { |
| 1313 | const struct dst_entry *dst = __sk_dst_get(sk); |
| 1314 | |
| 1315 | if (dst && dst_allfrag(dst)) |
| 1316 | mss_now -= icsk->icsk_af_ops->net_frag_header_len; |
| 1317 | } |
| 1318 | |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1319 | /* Clamp it (mss_clamp does not include tcp options) */ |
| 1320 | if (mss_now > tp->rx_opt.mss_clamp) |
| 1321 | mss_now = tp->rx_opt.mss_clamp; |
| 1322 | |
| 1323 | /* Now subtract optional transport overhead */ |
| 1324 | mss_now -= icsk->icsk_ext_hdr_len; |
| 1325 | |
| 1326 | /* Then reserve room for full set of TCP options and 8 bytes of data */ |
| 1327 | if (mss_now < 48) |
| 1328 | mss_now = 48; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1329 | return mss_now; |
| 1330 | } |
| 1331 | |
Yuchung Cheng | 1b63edd | 2013-02-22 08:59:06 +0000 | [diff] [blame] | 1332 | /* Calculate MSS. Not accounting for SACKs here. */ |
| 1333 | int tcp_mtu_to_mss(struct sock *sk, int pmtu) |
| 1334 | { |
| 1335 | /* Subtract TCP options size, not including SACKs */ |
| 1336 | return __tcp_mtu_to_mss(sk, pmtu) - |
| 1337 | (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr)); |
| 1338 | } |
| 1339 | |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1340 | /* Inverse of above */ |
Eric Dumazet | 6746960 | 2012-04-24 07:37:38 +0000 | [diff] [blame] | 1341 | int tcp_mss_to_mtu(struct sock *sk, int mss) |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1342 | { |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 1343 | const struct tcp_sock *tp = tcp_sk(sk); |
| 1344 | const struct inet_connection_sock *icsk = inet_csk(sk); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1345 | int mtu; |
| 1346 | |
| 1347 | mtu = mss + |
| 1348 | tp->tcp_header_len + |
| 1349 | icsk->icsk_ext_hdr_len + |
| 1350 | icsk->icsk_af_ops->net_header_len; |
| 1351 | |
Eric Dumazet | 6746960 | 2012-04-24 07:37:38 +0000 | [diff] [blame] | 1352 | /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ |
| 1353 | if (icsk->icsk_af_ops->net_frag_header_len) { |
| 1354 | const struct dst_entry *dst = __sk_dst_get(sk); |
| 1355 | |
| 1356 | if (dst && dst_allfrag(dst)) |
| 1357 | mtu += icsk->icsk_af_ops->net_frag_header_len; |
| 1358 | } |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1359 | return mtu; |
| 1360 | } |
| 1361 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 1362 | /* MTU probing init per socket */ |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1363 | void tcp_mtup_init(struct sock *sk) |
| 1364 | { |
| 1365 | struct tcp_sock *tp = tcp_sk(sk); |
| 1366 | struct inet_connection_sock *icsk = inet_csk(sk); |
Fan Du | b0f9ca5 | 2015-02-10 09:53:16 +0800 | [diff] [blame] | 1367 | struct net *net = sock_net(sk); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1368 | |
Fan Du | b0f9ca5 | 2015-02-10 09:53:16 +0800 | [diff] [blame] | 1369 | icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1370 | icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 1371 | icsk->icsk_af_ops->net_header_len; |
Fan Du | b0f9ca5 | 2015-02-10 09:53:16 +0800 | [diff] [blame] | 1372 | icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1373 | icsk->icsk_mtup.probe_size = 0; |
Fan Du | 05cbc0d | 2015-03-06 11:18:24 +0800 | [diff] [blame] | 1374 | if (icsk->icsk_mtup.enabled) |
| 1375 | icsk->icsk_mtup.probe_timestamp = tcp_time_stamp; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1376 | } |
Eric Dumazet | 4bc2f18 | 2010-07-09 21:22:10 +0000 | [diff] [blame] | 1377 | EXPORT_SYMBOL(tcp_mtup_init); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1378 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1379 | /* This function synchronize snd mss to current pmtu/exthdr set. |
| 1380 | |
| 1381 | tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts |
| 1382 | for TCP options, but includes only bare TCP header. |
| 1383 | |
| 1384 | tp->rx_opt.mss_clamp is mss negotiated at connection setup. |
Stephen Hemminger | caa20d9a | 2005-11-10 17:13:47 -0800 | [diff] [blame] | 1385 | It is minimum of user_mss and mss received with SYN. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1386 | It also does not include TCP options. |
| 1387 | |
Arnaldo Carvalho de Melo | d83d846 | 2005-12-13 23:26:10 -0800 | [diff] [blame] | 1388 | inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1389 | |
| 1390 | tp->mss_cache is current effective sending mss, including |
| 1391 | all tcp options except for SACKs. It is evaluated, |
| 1392 | taking into account current pmtu, but never exceeds |
| 1393 | tp->rx_opt.mss_clamp. |
| 1394 | |
| 1395 | NOTE1. rfc1122 clearly states that advertised MSS |
| 1396 | DOES NOT include either tcp or ip options. |
| 1397 | |
Arnaldo Carvalho de Melo | d83d846 | 2005-12-13 23:26:10 -0800 | [diff] [blame] | 1398 | NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache |
| 1399 | are READ ONLY outside this function. --ANK (980731) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1400 | */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1401 | unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) |
| 1402 | { |
| 1403 | struct tcp_sock *tp = tcp_sk(sk); |
Arnaldo Carvalho de Melo | d83d846 | 2005-12-13 23:26:10 -0800 | [diff] [blame] | 1404 | struct inet_connection_sock *icsk = inet_csk(sk); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1405 | int mss_now; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1406 | |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1407 | if (icsk->icsk_mtup.search_high > pmtu) |
| 1408 | icsk->icsk_mtup.search_high = pmtu; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1409 | |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1410 | mss_now = tcp_mtu_to_mss(sk, pmtu); |
Ilpo Järvinen | 409d22b | 2007-12-31 14:57:40 -0800 | [diff] [blame] | 1411 | mss_now = tcp_bound_to_half_wnd(tp, mss_now); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1412 | |
| 1413 | /* And store cached results */ |
Arnaldo Carvalho de Melo | d83d846 | 2005-12-13 23:26:10 -0800 | [diff] [blame] | 1414 | icsk->icsk_pmtu_cookie = pmtu; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1415 | if (icsk->icsk_mtup.enabled) |
| 1416 | mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1417 | tp->mss_cache = mss_now; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1418 | |
| 1419 | return mss_now; |
| 1420 | } |
Eric Dumazet | 4bc2f18 | 2010-07-09 21:22:10 +0000 | [diff] [blame] | 1421 | EXPORT_SYMBOL(tcp_sync_mss); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1422 | |
| 1423 | /* Compute the current effective MSS, taking SACKs and IP options, |
| 1424 | * and even PMTU discovery events into account. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1425 | */ |
Ilpo Järvinen | 0c54b85 | 2009-03-14 14:23:05 +0000 | [diff] [blame] | 1426 | unsigned int tcp_current_mss(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1427 | { |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 1428 | const struct tcp_sock *tp = tcp_sk(sk); |
| 1429 | const struct dst_entry *dst = __sk_dst_get(sk); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1430 | u32 mss_now; |
Eric Dumazet | 95c9617 | 2012-04-15 05:58:06 +0000 | [diff] [blame] | 1431 | unsigned int header_len; |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 1432 | struct tcp_out_options opts; |
| 1433 | struct tcp_md5sig_key *md5; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1434 | |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1435 | mss_now = tp->mss_cache; |
| 1436 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1437 | if (dst) { |
| 1438 | u32 mtu = dst_mtu(dst); |
Arnaldo Carvalho de Melo | d83d846 | 2005-12-13 23:26:10 -0800 | [diff] [blame] | 1439 | if (mtu != inet_csk(sk)->icsk_pmtu_cookie) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1440 | mss_now = tcp_sync_mss(sk, mtu); |
| 1441 | } |
| 1442 | |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 1443 | header_len = tcp_established_options(sk, NULL, &opts, &md5) + |
| 1444 | sizeof(struct tcphdr); |
| 1445 | /* The mss_cache is sized based on tp->tcp_header_len, which assumes |
| 1446 | * some common options. If this is an odd packet (because we have SACK |
| 1447 | * blocks etc) then our calculated header_len will be different, and |
| 1448 | * we have to adjust mss_now correspondingly */ |
| 1449 | if (header_len != tp->tcp_header_len) { |
| 1450 | int delta = (int) header_len - tp->tcp_header_len; |
| 1451 | mss_now -= delta; |
| 1452 | } |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 1453 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1454 | return mss_now; |
| 1455 | } |
| 1456 | |
Weiping Pan | 86fd14a | 2014-04-18 12:27:46 +0800 | [diff] [blame] | 1457 | /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. |
| 1458 | * As additional protections, we do not touch cwnd in retransmission phases, |
| 1459 | * and if application hit its sndbuf limit recently. |
| 1460 | */ |
| 1461 | static void tcp_cwnd_application_limited(struct sock *sk) |
David S. Miller | a762a98 | 2005-07-05 15:18:51 -0700 | [diff] [blame] | 1462 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 1463 | struct tcp_sock *tp = tcp_sk(sk); |
David S. Miller | a762a98 | 2005-07-05 15:18:51 -0700 | [diff] [blame] | 1464 | |
Weiping Pan | 86fd14a | 2014-04-18 12:27:46 +0800 | [diff] [blame] | 1465 | if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && |
| 1466 | sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { |
| 1467 | /* Limited by application or receiver window. */ |
| 1468 | u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); |
| 1469 | u32 win_used = max(tp->snd_cwnd_used, init_win); |
| 1470 | if (win_used < tp->snd_cwnd) { |
| 1471 | tp->snd_ssthresh = tcp_current_ssthresh(sk); |
| 1472 | tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1; |
| 1473 | } |
| 1474 | tp->snd_cwnd_used = 0; |
| 1475 | } |
| 1476 | tp->snd_cwnd_stamp = tcp_time_stamp; |
| 1477 | } |
| 1478 | |
Neal Cardwell | ca8a226 | 2014-05-22 10:41:08 -0400 | [diff] [blame] | 1479 | static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited) |
David S. Miller | a762a98 | 2005-07-05 15:18:51 -0700 | [diff] [blame] | 1480 | { |
| 1481 | struct tcp_sock *tp = tcp_sk(sk); |
| 1482 | |
Neal Cardwell | ca8a226 | 2014-05-22 10:41:08 -0400 | [diff] [blame] | 1483 | /* Track the maximum number of outstanding packets in each |
| 1484 | * window, and remember whether we were cwnd-limited then. |
| 1485 | */ |
| 1486 | if (!before(tp->snd_una, tp->max_packets_seq) || |
| 1487 | tp->packets_out > tp->max_packets_out) { |
| 1488 | tp->max_packets_out = tp->packets_out; |
| 1489 | tp->max_packets_seq = tp->snd_nxt; |
| 1490 | tp->is_cwnd_limited = is_cwnd_limited; |
| 1491 | } |
Eric Dumazet | e114a71 | 2014-04-30 11:58:13 -0700 | [diff] [blame] | 1492 | |
Eric Dumazet | 2490155 | 2014-05-02 21:18:05 -0700 | [diff] [blame] | 1493 | if (tcp_is_cwnd_limited(sk)) { |
David S. Miller | a762a98 | 2005-07-05 15:18:51 -0700 | [diff] [blame] | 1494 | /* Network is feed fully. */ |
| 1495 | tp->snd_cwnd_used = 0; |
| 1496 | tp->snd_cwnd_stamp = tcp_time_stamp; |
| 1497 | } else { |
| 1498 | /* Network starves. */ |
| 1499 | if (tp->packets_out > tp->snd_cwnd_used) |
| 1500 | tp->snd_cwnd_used = tp->packets_out; |
| 1501 | |
David S. Miller | 15d33c0 | 2007-04-09 13:23:14 -0700 | [diff] [blame] | 1502 | if (sysctl_tcp_slow_start_after_idle && |
| 1503 | (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto) |
David S. Miller | a762a98 | 2005-07-05 15:18:51 -0700 | [diff] [blame] | 1504 | tcp_cwnd_application_limited(sk); |
| 1505 | } |
| 1506 | } |
| 1507 | |
Eric Dumazet | d458992 | 2013-12-13 13:51:23 -0800 | [diff] [blame] | 1508 | /* Minshall's variant of the Nagle send check. */ |
| 1509 | static bool tcp_minshall_check(const struct tcp_sock *tp) |
| 1510 | { |
| 1511 | return after(tp->snd_sml, tp->snd_una) && |
| 1512 | !after(tp->snd_sml, tp->snd_nxt); |
| 1513 | } |
| 1514 | |
| 1515 | /* Update snd_sml if this skb is under mss |
| 1516 | * Note that a TSO packet might end with a sub-mss segment |
| 1517 | * The test is really : |
| 1518 | * if ((skb->len % mss) != 0) |
| 1519 | * tp->snd_sml = TCP_SKB_CB(skb)->end_seq; |
| 1520 | * But we can avoid doing the divide again given we already have |
| 1521 | * skb_pcount = skb->len / mss_now |
Ilpo Järvinen | 0e3a480 | 2007-12-24 21:33:45 -0800 | [diff] [blame] | 1522 | */ |
Eric Dumazet | d458992 | 2013-12-13 13:51:23 -0800 | [diff] [blame] | 1523 | static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now, |
| 1524 | const struct sk_buff *skb) |
| 1525 | { |
| 1526 | if (skb->len < tcp_skb_pcount(skb) * mss_now) |
| 1527 | tp->snd_sml = TCP_SKB_CB(skb)->end_seq; |
| 1528 | } |
| 1529 | |
| 1530 | /* Return false, if packet can be sent now without violation Nagle's rules: |
| 1531 | * 1. It is full sized. (provided by caller in %partial bool) |
| 1532 | * 2. Or it contains FIN. (already checked by caller) |
| 1533 | * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. |
| 1534 | * 4. Or TCP_CORK is not set, and all sent packets are ACKed. |
| 1535 | * With Minshall's modification: all sent small packets are ACKed. |
| 1536 | */ |
| 1537 | static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp, |
Peter Pan(潘卫平) | cc93fc5 | 2014-03-24 14:49:34 +0800 | [diff] [blame] | 1538 | int nonagle) |
Eric Dumazet | d458992 | 2013-12-13 13:51:23 -0800 | [diff] [blame] | 1539 | { |
| 1540 | return partial && |
| 1541 | ((nonagle & TCP_NAGLE_CORK) || |
| 1542 | (!nonagle && tp->packets_out && tcp_minshall_check(tp))); |
| 1543 | } |
Eric Dumazet | 605ad7f | 2014-12-07 12:22:18 -0800 | [diff] [blame] | 1544 | |
| 1545 | /* Return how many segs we'd like on a TSO packet, |
| 1546 | * to send one TSO packet per ms |
| 1547 | */ |
| 1548 | static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now) |
| 1549 | { |
| 1550 | u32 bytes, segs; |
| 1551 | |
| 1552 | bytes = min(sk->sk_pacing_rate >> 10, |
| 1553 | sk->sk_gso_max_size - 1 - MAX_TCP_HEADER); |
| 1554 | |
| 1555 | /* Goal is to send at least one packet per ms, |
| 1556 | * not one big TSO packet every 100 ms. |
| 1557 | * This preserves ACK clocking and is consistent |
| 1558 | * with tcp_tso_should_defer() heuristic. |
| 1559 | */ |
| 1560 | segs = max_t(u32, bytes / mss_now, sysctl_tcp_min_tso_segs); |
| 1561 | |
| 1562 | return min_t(u32, segs, sk->sk_gso_max_segs); |
| 1563 | } |
| 1564 | |
Eric Dumazet | d458992 | 2013-12-13 13:51:23 -0800 | [diff] [blame] | 1565 | /* Returns the portion of skb which can be sent right away */ |
| 1566 | static unsigned int tcp_mss_split_point(const struct sock *sk, |
| 1567 | const struct sk_buff *skb, |
| 1568 | unsigned int mss_now, |
| 1569 | unsigned int max_segs, |
| 1570 | int nonagle) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1571 | { |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 1572 | const struct tcp_sock *tp = tcp_sk(sk); |
Eric Dumazet | d458992 | 2013-12-13 13:51:23 -0800 | [diff] [blame] | 1573 | u32 partial, needed, window, max_len; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1574 | |
Ilpo Järvinen | 90840de | 2007-12-31 04:48:41 -0800 | [diff] [blame] | 1575 | window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; |
Ben Hutchings | 1485348 | 2012-07-30 16:11:42 +0000 | [diff] [blame] | 1576 | max_len = mss_now * max_segs; |
Ilpo Järvinen | 0e3a480 | 2007-12-24 21:33:45 -0800 | [diff] [blame] | 1577 | |
Ben Hutchings | 1485348 | 2012-07-30 16:11:42 +0000 | [diff] [blame] | 1578 | if (likely(max_len <= window && skb != tcp_write_queue_tail(sk))) |
| 1579 | return max_len; |
Ilpo Järvinen | 0e3a480 | 2007-12-24 21:33:45 -0800 | [diff] [blame] | 1580 | |
Ilpo Järvinen | 5ea3a74 | 2008-03-11 17:55:27 -0700 | [diff] [blame] | 1581 | needed = min(skb->len, window); |
| 1582 | |
Ben Hutchings | 1485348 | 2012-07-30 16:11:42 +0000 | [diff] [blame] | 1583 | if (max_len <= needed) |
| 1584 | return max_len; |
Ilpo Järvinen | 0e3a480 | 2007-12-24 21:33:45 -0800 | [diff] [blame] | 1585 | |
Eric Dumazet | d458992 | 2013-12-13 13:51:23 -0800 | [diff] [blame] | 1586 | partial = needed % mss_now; |
| 1587 | /* If last segment is not a full MSS, check if Nagle rules allow us |
| 1588 | * to include this last segment in this skb. |
| 1589 | * Otherwise, we'll split the skb at last MSS boundary |
| 1590 | */ |
Peter Pan(潘卫平) | cc93fc5 | 2014-03-24 14:49:34 +0800 | [diff] [blame] | 1591 | if (tcp_nagle_check(partial != 0, tp, nonagle)) |
Eric Dumazet | d458992 | 2013-12-13 13:51:23 -0800 | [diff] [blame] | 1592 | return needed - partial; |
| 1593 | |
| 1594 | return needed; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1595 | } |
| 1596 | |
| 1597 | /* Can at least one segment of SKB be sent right now, according to the |
| 1598 | * congestion window rules? If so, return how many segments are allowed. |
| 1599 | */ |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 1600 | static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, |
| 1601 | const struct sk_buff *skb) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1602 | { |
Eric Dumazet | d649a7a | 2014-11-13 09:45:22 -0800 | [diff] [blame] | 1603 | u32 in_flight, cwnd, halfcwnd; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1604 | |
| 1605 | /* Don't be strict about the congestion window for the final FIN. */ |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 1606 | if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && |
| 1607 | tcp_skb_pcount(skb) == 1) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1608 | return 1; |
| 1609 | |
| 1610 | in_flight = tcp_packets_in_flight(tp); |
| 1611 | cwnd = tp->snd_cwnd; |
Eric Dumazet | d649a7a | 2014-11-13 09:45:22 -0800 | [diff] [blame] | 1612 | if (in_flight >= cwnd) |
| 1613 | return 0; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1614 | |
Eric Dumazet | d649a7a | 2014-11-13 09:45:22 -0800 | [diff] [blame] | 1615 | /* For better scheduling, ensure we have at least |
| 1616 | * 2 GSO packets in flight. |
| 1617 | */ |
| 1618 | halfcwnd = max(cwnd >> 1, 1U); |
| 1619 | return min(halfcwnd, cwnd - in_flight); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1620 | } |
| 1621 | |
Uwe Kleine-König | b595076 | 2010-11-01 15:38:34 -0400 | [diff] [blame] | 1622 | /* Initialize TSO state of a skb. |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 1623 | * This must be invoked the first time we consider transmitting |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1624 | * SKB onto the wire. |
| 1625 | */ |
Eric Dumazet | 5bbb432 | 2015-06-11 09:15:17 -0700 | [diff] [blame] | 1626 | static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1627 | { |
| 1628 | int tso_segs = tcp_skb_pcount(skb); |
| 1629 | |
Ilpo Järvinen | f8269a4 | 2008-12-03 21:24:48 -0800 | [diff] [blame] | 1630 | if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { |
Eric Dumazet | 5bbb432 | 2015-06-11 09:15:17 -0700 | [diff] [blame] | 1631 | tcp_set_skb_tso_segs(skb, mss_now); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1632 | tso_segs = tcp_skb_pcount(skb); |
| 1633 | } |
| 1634 | return tso_segs; |
| 1635 | } |
| 1636 | |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1637 | |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1638 | /* Return true if the Nagle test allows this packet to be |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1639 | * sent now. |
| 1640 | */ |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1641 | static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, |
| 1642 | unsigned int cur_mss, int nonagle) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1643 | { |
| 1644 | /* Nagle rule does not apply to frames, which sit in the middle of the |
| 1645 | * write_queue (they have no chances to get new data). |
| 1646 | * |
| 1647 | * This is implemented in the callers, where they modify the 'nonagle' |
| 1648 | * argument based upon the location of SKB in the send queue. |
| 1649 | */ |
| 1650 | if (nonagle & TCP_NAGLE_PUSH) |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1651 | return true; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1652 | |
Yuchung Cheng | 9b44190 | 2013-03-20 13:32:58 +0000 | [diff] [blame] | 1653 | /* Don't use the nagle rule for urgent data (or for the final FIN). */ |
| 1654 | if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1655 | return true; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1656 | |
Peter Pan(潘卫平) | cc93fc5 | 2014-03-24 14:49:34 +0800 | [diff] [blame] | 1657 | if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle)) |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1658 | return true; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1659 | |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1660 | return false; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1661 | } |
| 1662 | |
| 1663 | /* Does at least the first segment of SKB fit into the send window? */ |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1664 | static bool tcp_snd_wnd_test(const struct tcp_sock *tp, |
| 1665 | const struct sk_buff *skb, |
| 1666 | unsigned int cur_mss) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1667 | { |
| 1668 | u32 end_seq = TCP_SKB_CB(skb)->end_seq; |
| 1669 | |
| 1670 | if (skb->len > cur_mss) |
| 1671 | end_seq = TCP_SKB_CB(skb)->seq + cur_mss; |
| 1672 | |
Ilpo Järvinen | 90840de | 2007-12-31 04:48:41 -0800 | [diff] [blame] | 1673 | return !after(end_seq, tcp_wnd_end(tp)); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1674 | } |
| 1675 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1676 | /* This checks if the data bearing packet SKB (usually tcp_send_head(sk)) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1677 | * should be put on the wire right now. If so, it returns the number of |
| 1678 | * packets allowed by the congestion window. |
| 1679 | */ |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 1680 | static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb, |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1681 | unsigned int cur_mss, int nonagle) |
| 1682 | { |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 1683 | const struct tcp_sock *tp = tcp_sk(sk); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1684 | unsigned int cwnd_quota; |
| 1685 | |
Eric Dumazet | 5bbb432 | 2015-06-11 09:15:17 -0700 | [diff] [blame] | 1686 | tcp_init_tso_segs(skb, cur_mss); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1687 | |
| 1688 | if (!tcp_nagle_test(tp, skb, cur_mss, nonagle)) |
| 1689 | return 0; |
| 1690 | |
| 1691 | cwnd_quota = tcp_cwnd_test(tp, skb); |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1692 | if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss)) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1693 | cwnd_quota = 0; |
| 1694 | |
| 1695 | return cwnd_quota; |
| 1696 | } |
| 1697 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 1698 | /* Test if sending is allowed right now. */ |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1699 | bool tcp_may_send_now(struct sock *sk) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1700 | { |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 1701 | const struct tcp_sock *tp = tcp_sk(sk); |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1702 | struct sk_buff *skb = tcp_send_head(sk); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1703 | |
Eric Dumazet | a02cec2 | 2010-09-22 20:43:57 +0000 | [diff] [blame] | 1704 | return skb && |
Ilpo Järvinen | 0c54b85 | 2009-03-14 14:23:05 +0000 | [diff] [blame] | 1705 | tcp_snd_test(sk, skb, tcp_current_mss(sk), |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1706 | (tcp_skb_is_last(sk, skb) ? |
Eric Dumazet | a02cec2 | 2010-09-22 20:43:57 +0000 | [diff] [blame] | 1707 | tp->nonagle : TCP_NAGLE_PUSH)); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1708 | } |
| 1709 | |
| 1710 | /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet |
| 1711 | * which is put after SKB on the list. It is very much like |
| 1712 | * tcp_fragment() except that it may make several kinds of assumptions |
| 1713 | * in order to speed up the splitting operation. In particular, we |
| 1714 | * know that all the data is in scatter-gather pages, and that the |
| 1715 | * packet has never been sent out before (and thus is not cloned). |
| 1716 | */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1717 | static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, |
Eric Dumazet | c4ead4c | 2010-06-24 01:00:22 +0000 | [diff] [blame] | 1718 | unsigned int mss_now, gfp_t gfp) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1719 | { |
| 1720 | struct sk_buff *buff; |
| 1721 | int nlen = skb->len - len; |
Ilpo Järvinen | 9ce0146 | 2009-02-28 04:44:42 +0000 | [diff] [blame] | 1722 | u8 flags; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1723 | |
| 1724 | /* All of a TSO frame must be composed of paged data. */ |
Herbert Xu | c8ac377 | 2005-08-16 20:43:40 -0700 | [diff] [blame] | 1725 | if (skb->len != skb->data_len) |
Octavian Purdila | 6cc55e0 | 2014-06-06 17:32:37 +0300 | [diff] [blame] | 1726 | return tcp_fragment(sk, skb, len, mss_now, gfp); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1727 | |
Eric Dumazet | eb93447 | 2015-05-19 13:26:55 -0700 | [diff] [blame] | 1728 | buff = sk_stream_alloc_skb(sk, 0, gfp, true); |
Ian Morris | 51456b2 | 2015-04-03 09:17:26 +0100 | [diff] [blame] | 1729 | if (unlikely(!buff)) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1730 | return -ENOMEM; |
| 1731 | |
Hideo Aoki | 3ab224b | 2007-12-31 00:11:19 -0800 | [diff] [blame] | 1732 | sk->sk_wmem_queued += buff->truesize; |
| 1733 | sk_mem_charge(sk, buff->truesize); |
Herbert Xu | b60b49e | 2006-04-19 21:35:00 -0700 | [diff] [blame] | 1734 | buff->truesize += nlen; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1735 | skb->truesize -= nlen; |
| 1736 | |
| 1737 | /* Correct the sequence numbers. */ |
| 1738 | TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; |
| 1739 | TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; |
| 1740 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; |
| 1741 | |
| 1742 | /* PSH and FIN should only be set in the second packet. */ |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 1743 | flags = TCP_SKB_CB(skb)->tcp_flags; |
| 1744 | TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); |
| 1745 | TCP_SKB_CB(buff)->tcp_flags = flags; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1746 | |
| 1747 | /* This packet was never sent out yet, so no SACK bits. */ |
| 1748 | TCP_SKB_CB(buff)->sacked = 0; |
| 1749 | |
Martin KaFai Lau | a166140 | 2016-04-25 14:44:50 -0700 | [diff] [blame] | 1750 | tcp_skb_fragment_eor(skb, buff); |
| 1751 | |
Patrick McHardy | 84fa793 | 2006-08-29 16:44:56 -0700 | [diff] [blame] | 1752 | buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1753 | skb_split(skb, buff, len); |
Willem de Bruijn | 490cc7d | 2014-08-12 15:08:12 -0400 | [diff] [blame] | 1754 | tcp_fragment_tstamp(skb, buff); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1755 | |
| 1756 | /* Fix up tso_factor for both original and new SKB. */ |
Eric Dumazet | 5bbb432 | 2015-06-11 09:15:17 -0700 | [diff] [blame] | 1757 | tcp_set_skb_tso_segs(skb, mss_now); |
| 1758 | tcp_set_skb_tso_segs(buff, mss_now); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1759 | |
| 1760 | /* Link BUFF into the send queue. */ |
Eric Dumazet | f4a775d | 2014-09-22 16:29:32 -0700 | [diff] [blame] | 1761 | __skb_header_release(buff); |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1762 | tcp_insert_write_queue_after(skb, buff, sk); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1763 | |
| 1764 | return 0; |
| 1765 | } |
| 1766 | |
| 1767 | /* Try to defer sending, if possible, in order to minimize the amount |
| 1768 | * of TSO splitting we do. View it as a kind of TSO Nagle test. |
| 1769 | * |
| 1770 | * This algorithm is from John Heffner. |
| 1771 | */ |
Neal Cardwell | ca8a226 | 2014-05-22 10:41:08 -0400 | [diff] [blame] | 1772 | static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb, |
Eric Dumazet | 605ad7f | 2014-12-07 12:22:18 -0800 | [diff] [blame] | 1773 | bool *is_cwnd_limited, u32 max_segs) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1774 | { |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 1775 | const struct inet_connection_sock *icsk = inet_csk(sk); |
Eric Dumazet | 50c8339 | 2015-02-26 14:10:19 -0800 | [diff] [blame] | 1776 | u32 age, send_win, cong_win, limit, in_flight; |
| 1777 | struct tcp_sock *tp = tcp_sk(sk); |
| 1778 | struct skb_mstamp now; |
| 1779 | struct sk_buff *head; |
Eric Dumazet | ad9f4f5 | 2010-12-07 12:03:55 +0000 | [diff] [blame] | 1780 | int win_divisor; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1781 | |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 1782 | if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) |
John Heffner | ae8064a | 2006-10-18 20:36:48 -0700 | [diff] [blame] | 1783 | goto send_now; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1784 | |
Eric Dumazet | 99d7662 | 2015-07-26 09:45:24 +0200 | [diff] [blame] | 1785 | if (icsk->icsk_ca_state >= TCP_CA_Recovery) |
John Heffner | ae8064a | 2006-10-18 20:36:48 -0700 | [diff] [blame] | 1786 | goto send_now; |
| 1787 | |
Eric Dumazet | 5f852eb | 2015-02-26 14:10:18 -0800 | [diff] [blame] | 1788 | /* Avoid bursty behavior by allowing defer |
| 1789 | * only if the last write was recent. |
| 1790 | */ |
| 1791 | if ((s32)(tcp_time_stamp - tp->lsndtime) > 0) |
John Heffner | ae8064a | 2006-10-18 20:36:48 -0700 | [diff] [blame] | 1792 | goto send_now; |
David S. Miller | 908a75c | 2005-07-05 15:43:58 -0700 | [diff] [blame] | 1793 | |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1794 | in_flight = tcp_packets_in_flight(tp); |
| 1795 | |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1796 | BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight)); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1797 | |
Ilpo Järvinen | 90840de | 2007-12-31 04:48:41 -0800 | [diff] [blame] | 1798 | send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1799 | |
| 1800 | /* From in_flight test above, we know that cwnd > in_flight. */ |
| 1801 | cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; |
| 1802 | |
| 1803 | limit = min(send_win, cong_win); |
| 1804 | |
David S. Miller | ba244fe | 2006-03-11 18:51:49 -0800 | [diff] [blame] | 1805 | /* If a full-sized TSO skb can be sent, do it. */ |
Eric Dumazet | 605ad7f | 2014-12-07 12:22:18 -0800 | [diff] [blame] | 1806 | if (limit >= max_segs * tp->mss_cache) |
John Heffner | ae8064a | 2006-10-18 20:36:48 -0700 | [diff] [blame] | 1807 | goto send_now; |
David S. Miller | ba244fe | 2006-03-11 18:51:49 -0800 | [diff] [blame] | 1808 | |
Ilpo Järvinen | 62ad276 | 2009-02-28 04:44:29 +0000 | [diff] [blame] | 1809 | /* Middle in queue won't get any more data, full sendable already? */ |
| 1810 | if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) |
| 1811 | goto send_now; |
| 1812 | |
Eric Dumazet | ad9f4f5 | 2010-12-07 12:03:55 +0000 | [diff] [blame] | 1813 | win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor); |
| 1814 | if (win_divisor) { |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1815 | u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); |
| 1816 | |
| 1817 | /* If at least some fraction of a window is available, |
| 1818 | * just use it. |
| 1819 | */ |
Eric Dumazet | ad9f4f5 | 2010-12-07 12:03:55 +0000 | [diff] [blame] | 1820 | chunk /= win_divisor; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1821 | if (limit >= chunk) |
John Heffner | ae8064a | 2006-10-18 20:36:48 -0700 | [diff] [blame] | 1822 | goto send_now; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1823 | } else { |
| 1824 | /* Different approach, try not to defer past a single |
| 1825 | * ACK. Receiver should ACK every other full sized |
| 1826 | * frame, so if we have space for more than 3 frames |
| 1827 | * then send now. |
| 1828 | */ |
Neal Cardwell | 6b5a5c0 | 2011-11-21 17:15:14 +0000 | [diff] [blame] | 1829 | if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) |
John Heffner | ae8064a | 2006-10-18 20:36:48 -0700 | [diff] [blame] | 1830 | goto send_now; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1831 | } |
| 1832 | |
Eric Dumazet | 50c8339 | 2015-02-26 14:10:19 -0800 | [diff] [blame] | 1833 | head = tcp_write_queue_head(sk); |
| 1834 | skb_mstamp_get(&now); |
| 1835 | age = skb_mstamp_us_delta(&now, &head->skb_mstamp); |
| 1836 | /* If next ACK is likely to come too late (half srtt), do not defer */ |
| 1837 | if (age < (tp->srtt_us >> 4)) |
| 1838 | goto send_now; |
| 1839 | |
Eric Dumazet | 5f852eb | 2015-02-26 14:10:18 -0800 | [diff] [blame] | 1840 | /* Ok, it looks like it is advisable to defer. */ |
John Heffner | ae8064a | 2006-10-18 20:36:48 -0700 | [diff] [blame] | 1841 | |
Bendik Rønning Opstad | d2e1339 | 2015-09-23 18:49:53 +0200 | [diff] [blame] | 1842 | if (cong_win < send_win && cong_win <= skb->len) |
Neal Cardwell | ca8a226 | 2014-05-22 10:41:08 -0400 | [diff] [blame] | 1843 | *is_cwnd_limited = true; |
| 1844 | |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1845 | return true; |
John Heffner | ae8064a | 2006-10-18 20:36:48 -0700 | [diff] [blame] | 1846 | |
| 1847 | send_now: |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1848 | return false; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1849 | } |
| 1850 | |
Fan Du | 05cbc0d | 2015-03-06 11:18:24 +0800 | [diff] [blame] | 1851 | static inline void tcp_mtu_check_reprobe(struct sock *sk) |
| 1852 | { |
| 1853 | struct inet_connection_sock *icsk = inet_csk(sk); |
| 1854 | struct tcp_sock *tp = tcp_sk(sk); |
| 1855 | struct net *net = sock_net(sk); |
| 1856 | u32 interval; |
| 1857 | s32 delta; |
| 1858 | |
| 1859 | interval = net->ipv4.sysctl_tcp_probe_interval; |
| 1860 | delta = tcp_time_stamp - icsk->icsk_mtup.probe_timestamp; |
| 1861 | if (unlikely(delta >= interval * HZ)) { |
| 1862 | int mss = tcp_current_mss(sk); |
| 1863 | |
| 1864 | /* Update current search range */ |
| 1865 | icsk->icsk_mtup.probe_size = 0; |
| 1866 | icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + |
| 1867 | sizeof(struct tcphdr) + |
| 1868 | icsk->icsk_af_ops->net_header_len; |
| 1869 | icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss); |
| 1870 | |
| 1871 | /* Update probe time stamp */ |
| 1872 | icsk->icsk_mtup.probe_timestamp = tcp_time_stamp; |
| 1873 | } |
| 1874 | } |
| 1875 | |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1876 | /* Create a new MTU probe if we are ready. |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 1877 | * MTU probe is regularly attempting to increase the path MTU by |
| 1878 | * deliberately sending larger packets. This discovers routing |
| 1879 | * changes resulting in larger path MTUs. |
| 1880 | * |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1881 | * Returns 0 if we should wait to probe (no cwnd available), |
| 1882 | * 1 if a probe was sent, |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1883 | * -1 otherwise |
| 1884 | */ |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1885 | static int tcp_mtu_probe(struct sock *sk) |
| 1886 | { |
| 1887 | struct tcp_sock *tp = tcp_sk(sk); |
| 1888 | struct inet_connection_sock *icsk = inet_csk(sk); |
| 1889 | struct sk_buff *skb, *nskb, *next; |
Fan Du | 6b58e0a | 2015-03-06 11:18:23 +0800 | [diff] [blame] | 1890 | struct net *net = sock_net(sk); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1891 | int len; |
| 1892 | int probe_size; |
Ilpo Järvinen | 91cc17c | 2007-11-23 19:08:16 +0800 | [diff] [blame] | 1893 | int size_needed; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1894 | int copy; |
| 1895 | int mss_now; |
Fan Du | 6b58e0a | 2015-03-06 11:18:23 +0800 | [diff] [blame] | 1896 | int interval; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1897 | |
| 1898 | /* Not currently probing/verifying, |
| 1899 | * not in recovery, |
| 1900 | * have enough cwnd, and |
| 1901 | * not SACKing (the variable headers throw things off) */ |
| 1902 | if (!icsk->icsk_mtup.enabled || |
| 1903 | icsk->icsk_mtup.probe_size || |
| 1904 | inet_csk(sk)->icsk_ca_state != TCP_CA_Open || |
| 1905 | tp->snd_cwnd < 11 || |
Ilpo Järvinen | cabeccb | 2009-02-28 04:44:38 +0000 | [diff] [blame] | 1906 | tp->rx_opt.num_sacks || tp->rx_opt.dsack) |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1907 | return -1; |
| 1908 | |
Fan Du | 6b58e0a | 2015-03-06 11:18:23 +0800 | [diff] [blame] | 1909 | /* Use binary search for probe_size between tcp_mss_base, |
| 1910 | * and current mss_clamp. if (search_high - search_low) |
| 1911 | * smaller than a threshold, backoff from probing. |
| 1912 | */ |
Ilpo Järvinen | 0c54b85 | 2009-03-14 14:23:05 +0000 | [diff] [blame] | 1913 | mss_now = tcp_current_mss(sk); |
Fan Du | 6b58e0a | 2015-03-06 11:18:23 +0800 | [diff] [blame] | 1914 | probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high + |
| 1915 | icsk->icsk_mtup.search_low) >> 1); |
Ilpo Järvinen | 91cc17c | 2007-11-23 19:08:16 +0800 | [diff] [blame] | 1916 | size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; |
Fan Du | 6b58e0a | 2015-03-06 11:18:23 +0800 | [diff] [blame] | 1917 | interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low; |
Fan Du | 05cbc0d | 2015-03-06 11:18:24 +0800 | [diff] [blame] | 1918 | /* When misfortune happens, we are reprobing actively, |
| 1919 | * and then reprobe timer has expired. We stick with current |
| 1920 | * probing process by not resetting search range to its orignal. |
| 1921 | */ |
Fan Du | 6b58e0a | 2015-03-06 11:18:23 +0800 | [diff] [blame] | 1922 | if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) || |
Fan Du | 05cbc0d | 2015-03-06 11:18:24 +0800 | [diff] [blame] | 1923 | interval < net->ipv4.sysctl_tcp_probe_threshold) { |
| 1924 | /* Check whether enough time has elaplased for |
| 1925 | * another round of probing. |
| 1926 | */ |
| 1927 | tcp_mtu_check_reprobe(sk); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1928 | return -1; |
| 1929 | } |
| 1930 | |
| 1931 | /* Have enough data in the send queue to probe? */ |
Ilpo Järvinen | 7f9c33e | 2007-11-23 19:10:56 +0800 | [diff] [blame] | 1932 | if (tp->write_seq - tp->snd_nxt < size_needed) |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1933 | return -1; |
| 1934 | |
Ilpo Järvinen | 91cc17c | 2007-11-23 19:08:16 +0800 | [diff] [blame] | 1935 | if (tp->snd_wnd < size_needed) |
| 1936 | return -1; |
Ilpo Järvinen | 90840de | 2007-12-31 04:48:41 -0800 | [diff] [blame] | 1937 | if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) |
Ilpo Järvinen | 91cc17c | 2007-11-23 19:08:16 +0800 | [diff] [blame] | 1938 | return 0; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1939 | |
Ilpo Järvinen | d67c58e | 2007-12-02 00:48:01 +0200 | [diff] [blame] | 1940 | /* Do we need to wait to drain cwnd? With none in flight, don't stall */ |
| 1941 | if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { |
| 1942 | if (!tcp_packets_in_flight(tp)) |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1943 | return -1; |
| 1944 | else |
| 1945 | return 0; |
| 1946 | } |
| 1947 | |
| 1948 | /* We're allowed to probe. Build it now. */ |
Eric Dumazet | eb93447 | 2015-05-19 13:26:55 -0700 | [diff] [blame] | 1949 | nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false); |
Ian Morris | 51456b2 | 2015-04-03 09:17:26 +0100 | [diff] [blame] | 1950 | if (!nskb) |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1951 | return -1; |
Hideo Aoki | 3ab224b | 2007-12-31 00:11:19 -0800 | [diff] [blame] | 1952 | sk->sk_wmem_queued += nskb->truesize; |
| 1953 | sk_mem_charge(sk, nskb->truesize); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1954 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1955 | skb = tcp_send_head(sk); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1956 | |
| 1957 | TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; |
| 1958 | TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 1959 | TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1960 | TCP_SKB_CB(nskb)->sacked = 0; |
| 1961 | nskb->csum = 0; |
Patrick McHardy | 84fa793 | 2006-08-29 16:44:56 -0700 | [diff] [blame] | 1962 | nskb->ip_summed = skb->ip_summed; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1963 | |
Ilpo Järvinen | 50c4817 | 2007-12-02 00:48:00 +0200 | [diff] [blame] | 1964 | tcp_insert_write_queue_before(nskb, skb, sk); |
| 1965 | |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1966 | len = 0; |
Ilpo Järvinen | 234b686 | 2007-12-02 00:48:02 +0200 | [diff] [blame] | 1967 | tcp_for_write_queue_from_safe(skb, next, sk) { |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1968 | copy = min_t(int, skb->len, probe_size - len); |
| 1969 | if (nskb->ip_summed) |
| 1970 | skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); |
| 1971 | else |
| 1972 | nskb->csum = skb_copy_and_csum_bits(skb, 0, |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1973 | skb_put(nskb, copy), |
| 1974 | copy, nskb->csum); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1975 | |
| 1976 | if (skb->len <= copy) { |
| 1977 | /* We've eaten all the data from this skb. |
| 1978 | * Throw it away. */ |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 1979 | TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1980 | tcp_unlink_write_queue(skb, sk); |
Hideo Aoki | 3ab224b | 2007-12-31 00:11:19 -0800 | [diff] [blame] | 1981 | sk_wmem_free_skb(sk, skb); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1982 | } else { |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 1983 | TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & |
Changli Gao | a3433f3 | 2010-06-12 14:01:43 +0000 | [diff] [blame] | 1984 | ~(TCPHDR_FIN|TCPHDR_PSH); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1985 | if (!skb_shinfo(skb)->nr_frags) { |
| 1986 | skb_pull(skb, copy); |
Patrick McHardy | 84fa793 | 2006-08-29 16:44:56 -0700 | [diff] [blame] | 1987 | if (skb->ip_summed != CHECKSUM_PARTIAL) |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1988 | skb->csum = csum_partial(skb->data, |
| 1989 | skb->len, 0); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1990 | } else { |
| 1991 | __pskb_trim_head(skb, copy); |
Eric Dumazet | 5bbb432 | 2015-06-11 09:15:17 -0700 | [diff] [blame] | 1992 | tcp_set_skb_tso_segs(skb, mss_now); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1993 | } |
| 1994 | TCP_SKB_CB(skb)->seq += copy; |
| 1995 | } |
| 1996 | |
| 1997 | len += copy; |
Ilpo Järvinen | 234b686 | 2007-12-02 00:48:02 +0200 | [diff] [blame] | 1998 | |
| 1999 | if (len >= probe_size) |
| 2000 | break; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 2001 | } |
Eric Dumazet | 5bbb432 | 2015-06-11 09:15:17 -0700 | [diff] [blame] | 2002 | tcp_init_tso_segs(nskb, nskb->len); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 2003 | |
| 2004 | /* We're ready to send. If this fails, the probe will |
Eric Dumazet | 7faee5c | 2014-09-05 15:33:33 -0700 | [diff] [blame] | 2005 | * be resegmented into mss-sized pieces by tcp_write_xmit(). |
| 2006 | */ |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 2007 | if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { |
| 2008 | /* Decrement cwnd here because we are sending |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 2009 | * effectively two packets. */ |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 2010 | tp->snd_cwnd--; |
Ilpo Järvinen | 66f5fe6 | 2007-12-31 04:43:57 -0800 | [diff] [blame] | 2011 | tcp_event_new_data_sent(sk, nskb); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 2012 | |
| 2013 | icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); |
John Heffner | 0e7b136 | 2006-03-20 21:32:58 -0800 | [diff] [blame] | 2014 | tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; |
| 2015 | tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 2016 | |
| 2017 | return 1; |
| 2018 | } |
| 2019 | |
| 2020 | return -1; |
| 2021 | } |
| 2022 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2023 | /* This routine writes packets to the network. It advances the |
| 2024 | * send_head. This happens as incoming acks open up the remote |
| 2025 | * window for us. |
| 2026 | * |
Ilpo Järvinen | f8269a4 | 2008-12-03 21:24:48 -0800 | [diff] [blame] | 2027 | * LARGESEND note: !tcp_urg_mode is overkill, only frames between |
| 2028 | * snd_up-64k-mss .. snd_up cannot be large. However, taking into |
| 2029 | * account rare use of URG, this is not a big flaw. |
| 2030 | * |
Nandita Dukkipati | 6ba8a3b | 2013-03-11 10:00:43 +0000 | [diff] [blame] | 2031 | * Send at most one packet when push_one > 0. Temporarily ignore |
| 2032 | * cwnd limit to force at most one packet out when push_one == 2. |
| 2033 | |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 2034 | * Returns true, if no segments are in flight and we have queued segments, |
| 2035 | * but cannot send anything now because of SWS or another problem. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2036 | */ |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 2037 | static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, |
| 2038 | int push_one, gfp_t gfp) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2039 | { |
| 2040 | struct tcp_sock *tp = tcp_sk(sk); |
David S. Miller | 92df7b5 | 2005-07-05 15:19:06 -0700 | [diff] [blame] | 2041 | struct sk_buff *skb; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 2042 | unsigned int tso_segs, sent_pkts; |
| 2043 | int cwnd_quota; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 2044 | int result; |
Neal Cardwell | ca8a226 | 2014-05-22 10:41:08 -0400 | [diff] [blame] | 2045 | bool is_cwnd_limited = false; |
Eric Dumazet | 605ad7f | 2014-12-07 12:22:18 -0800 | [diff] [blame] | 2046 | u32 max_segs; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2047 | |
David S. Miller | 92df7b5 | 2005-07-05 15:19:06 -0700 | [diff] [blame] | 2048 | sent_pkts = 0; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 2049 | |
Ilpo Järvinen | d5dd917 | 2008-12-05 22:48:55 -0800 | [diff] [blame] | 2050 | if (!push_one) { |
| 2051 | /* Do MTU probing. */ |
| 2052 | result = tcp_mtu_probe(sk); |
| 2053 | if (!result) { |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 2054 | return false; |
Ilpo Järvinen | d5dd917 | 2008-12-05 22:48:55 -0800 | [diff] [blame] | 2055 | } else if (result > 0) { |
| 2056 | sent_pkts = 1; |
| 2057 | } |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 2058 | } |
| 2059 | |
Eric Dumazet | 605ad7f | 2014-12-07 12:22:18 -0800 | [diff] [blame] | 2060 | max_segs = tcp_tso_autosize(sk, mss_now); |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2061 | while ((skb = tcp_send_head(sk))) { |
Herbert Xu | c8ac377 | 2005-08-16 20:43:40 -0700 | [diff] [blame] | 2062 | unsigned int limit; |
| 2063 | |
Eric Dumazet | 5bbb432 | 2015-06-11 09:15:17 -0700 | [diff] [blame] | 2064 | tso_segs = tcp_init_tso_segs(skb, mss_now); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 2065 | BUG_ON(!tso_segs); |
David S. Miller | aa93466 | 2005-07-05 15:20:09 -0700 | [diff] [blame] | 2066 | |
Andrey Vagin | 9d186ca | 2014-08-13 16:03:10 +0400 | [diff] [blame] | 2067 | if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) { |
Eric Dumazet | 7faee5c | 2014-09-05 15:33:33 -0700 | [diff] [blame] | 2068 | /* "skb_mstamp" is used as a start point for the retransmit timer */ |
| 2069 | skb_mstamp_get(&skb->skb_mstamp); |
Andrew Vagin | ec34232 | 2012-11-15 04:03:17 +0000 | [diff] [blame] | 2070 | goto repair; /* Skip network transmission */ |
Andrey Vagin | 9d186ca | 2014-08-13 16:03:10 +0400 | [diff] [blame] | 2071 | } |
Andrew Vagin | ec34232 | 2012-11-15 04:03:17 +0000 | [diff] [blame] | 2072 | |
Herbert Xu | b68e9f8 | 2005-08-04 19:52:02 -0700 | [diff] [blame] | 2073 | cwnd_quota = tcp_cwnd_test(tp, skb); |
Nandita Dukkipati | 6ba8a3b | 2013-03-11 10:00:43 +0000 | [diff] [blame] | 2074 | if (!cwnd_quota) { |
| 2075 | if (push_one == 2) |
| 2076 | /* Force out a loss probe pkt. */ |
| 2077 | cwnd_quota = 1; |
| 2078 | else |
| 2079 | break; |
| 2080 | } |
Herbert Xu | b68e9f8 | 2005-08-04 19:52:02 -0700 | [diff] [blame] | 2081 | |
| 2082 | if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) |
| 2083 | break; |
| 2084 | |
Eric Dumazet | d6a4e26 | 2015-05-26 08:55:28 -0700 | [diff] [blame] | 2085 | if (tso_segs == 1) { |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 2086 | if (unlikely(!tcp_nagle_test(tp, skb, mss_now, |
| 2087 | (tcp_skb_is_last(sk, skb) ? |
| 2088 | nonagle : TCP_NAGLE_PUSH)))) |
| 2089 | break; |
| 2090 | } else { |
Neal Cardwell | ca8a226 | 2014-05-22 10:41:08 -0400 | [diff] [blame] | 2091 | if (!push_one && |
Eric Dumazet | 605ad7f | 2014-12-07 12:22:18 -0800 | [diff] [blame] | 2092 | tcp_tso_should_defer(sk, skb, &is_cwnd_limited, |
| 2093 | max_segs)) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 2094 | break; |
| 2095 | } |
David S. Miller | aa93466 | 2005-07-05 15:20:09 -0700 | [diff] [blame] | 2096 | |
Eric Dumazet | 605ad7f | 2014-12-07 12:22:18 -0800 | [diff] [blame] | 2097 | limit = mss_now; |
Eric Dumazet | d6a4e26 | 2015-05-26 08:55:28 -0700 | [diff] [blame] | 2098 | if (tso_segs > 1 && !tcp_urg_mode(tp)) |
Eric Dumazet | 605ad7f | 2014-12-07 12:22:18 -0800 | [diff] [blame] | 2099 | limit = tcp_mss_split_point(sk, skb, mss_now, |
| 2100 | min_t(unsigned int, |
| 2101 | cwnd_quota, |
| 2102 | max_segs), |
| 2103 | nonagle); |
| 2104 | |
| 2105 | if (skb->len > limit && |
| 2106 | unlikely(tso_fragment(sk, skb, limit, mss_now, gfp))) |
| 2107 | break; |
| 2108 | |
Eric Dumazet | c9eeec2 | 2013-09-27 03:28:54 -0700 | [diff] [blame] | 2109 | /* TCP Small Queues : |
| 2110 | * Control number of packets in qdisc/devices to two packets / or ~1 ms. |
| 2111 | * This allows for : |
| 2112 | * - better RTT estimation and ACK scheduling |
| 2113 | * - faster recovery |
| 2114 | * - high rates |
Eric Dumazet | 98e0938 | 2013-11-13 06:32:54 -0800 | [diff] [blame] | 2115 | * Alas, some drivers / subsystems require a fair amount |
| 2116 | * of queued bytes to ensure line rate. |
| 2117 | * One example is wifi aggregation (802.11 AMPDU) |
Eric Dumazet | 46d3cea | 2012-07-11 05:50:31 +0000 | [diff] [blame] | 2118 | */ |
Eric Dumazet | 605ad7f | 2014-12-07 12:22:18 -0800 | [diff] [blame] | 2119 | limit = max(2 * skb->truesize, sk->sk_pacing_rate >> 10); |
| 2120 | limit = min_t(u32, limit, sysctl_tcp_limit_output_bytes); |
Eric Dumazet | c9eeec2 | 2013-09-27 03:28:54 -0700 | [diff] [blame] | 2121 | |
| 2122 | if (atomic_read(&sk->sk_wmem_alloc) > limit) { |
Eric Dumazet | 46d3cea | 2012-07-11 05:50:31 +0000 | [diff] [blame] | 2123 | set_bit(TSQ_THROTTLED, &tp->tsq_flags); |
John Ogness | bf06200 | 2014-02-09 18:40:11 -0800 | [diff] [blame] | 2124 | /* It is possible TX completion already happened |
| 2125 | * before we set TSQ_THROTTLED, so we must |
| 2126 | * test again the condition. |
John Ogness | bf06200 | 2014-02-09 18:40:11 -0800 | [diff] [blame] | 2127 | */ |
Peter Zijlstra | 4e857c5 | 2014-03-17 18:06:10 +0100 | [diff] [blame] | 2128 | smp_mb__after_atomic(); |
John Ogness | bf06200 | 2014-02-09 18:40:11 -0800 | [diff] [blame] | 2129 | if (atomic_read(&sk->sk_wmem_alloc) > limit) |
| 2130 | break; |
Eric Dumazet | 46d3cea | 2012-07-11 05:50:31 +0000 | [diff] [blame] | 2131 | } |
Eric Dumazet | c9eeec2 | 2013-09-27 03:28:54 -0700 | [diff] [blame] | 2132 | |
Ilpo Järvinen | d5dd917 | 2008-12-05 22:48:55 -0800 | [diff] [blame] | 2133 | if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) |
David S. Miller | 92df7b5 | 2005-07-05 15:19:06 -0700 | [diff] [blame] | 2134 | break; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2135 | |
Andrew Vagin | ec34232 | 2012-11-15 04:03:17 +0000 | [diff] [blame] | 2136 | repair: |
David S. Miller | 92df7b5 | 2005-07-05 15:19:06 -0700 | [diff] [blame] | 2137 | /* Advance the send_head. This one is sent out. |
| 2138 | * This call will increment packets_out. |
| 2139 | */ |
Ilpo Järvinen | 66f5fe6 | 2007-12-31 04:43:57 -0800 | [diff] [blame] | 2140 | tcp_event_new_data_sent(sk, skb); |
David S. Miller | 92df7b5 | 2005-07-05 15:19:06 -0700 | [diff] [blame] | 2141 | |
| 2142 | tcp_minshall_update(tp, mss_now, skb); |
Nandita Dukkipati | a262f0c | 2011-08-21 20:21:57 +0000 | [diff] [blame] | 2143 | sent_pkts += tcp_skb_pcount(skb); |
Ilpo Järvinen | d5dd917 | 2008-12-05 22:48:55 -0800 | [diff] [blame] | 2144 | |
| 2145 | if (push_one) |
| 2146 | break; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2147 | } |
David S. Miller | 92df7b5 | 2005-07-05 15:19:06 -0700 | [diff] [blame] | 2148 | |
David S. Miller | aa93466 | 2005-07-05 15:20:09 -0700 | [diff] [blame] | 2149 | if (likely(sent_pkts)) { |
Yuchung Cheng | 684bad1 | 2012-09-02 17:38:04 +0000 | [diff] [blame] | 2150 | if (tcp_in_cwnd_reduction(sk)) |
| 2151 | tp->prr_out += sent_pkts; |
Nandita Dukkipati | 6ba8a3b | 2013-03-11 10:00:43 +0000 | [diff] [blame] | 2152 | |
| 2153 | /* Send one loss probe per tail loss episode. */ |
| 2154 | if (push_one != 2) |
| 2155 | tcp_schedule_loss_probe(sk); |
Bendik Rønning Opstad | d2e1339 | 2015-09-23 18:49:53 +0200 | [diff] [blame] | 2156 | is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd); |
Neal Cardwell | ca8a226 | 2014-05-22 10:41:08 -0400 | [diff] [blame] | 2157 | tcp_cwnd_validate(sk, is_cwnd_limited); |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 2158 | return false; |
David S. Miller | 92df7b5 | 2005-07-05 15:19:06 -0700 | [diff] [blame] | 2159 | } |
Yuchung Cheng | b340b26 | 2015-08-12 11:18:19 -0700 | [diff] [blame] | 2160 | return !tp->packets_out && tcp_send_head(sk); |
Nandita Dukkipati | 6ba8a3b | 2013-03-11 10:00:43 +0000 | [diff] [blame] | 2161 | } |
| 2162 | |
| 2163 | bool tcp_schedule_loss_probe(struct sock *sk) |
| 2164 | { |
| 2165 | struct inet_connection_sock *icsk = inet_csk(sk); |
| 2166 | struct tcp_sock *tp = tcp_sk(sk); |
| 2167 | u32 timeout, tlp_time_stamp, rto_time_stamp; |
Eric Dumazet | 740b0f1 | 2014-02-26 14:02:48 -0800 | [diff] [blame] | 2168 | u32 rtt = usecs_to_jiffies(tp->srtt_us >> 3); |
Nandita Dukkipati | 6ba8a3b | 2013-03-11 10:00:43 +0000 | [diff] [blame] | 2169 | |
| 2170 | if (WARN_ON(icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS)) |
| 2171 | return false; |
| 2172 | /* No consecutive loss probes. */ |
| 2173 | if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) { |
| 2174 | tcp_rearm_rto(sk); |
| 2175 | return false; |
| 2176 | } |
| 2177 | /* Don't do any loss probe on a Fast Open connection before 3WHS |
| 2178 | * finishes. |
| 2179 | */ |
Yuchung Cheng | f9b9958 | 2015-09-18 11:40:33 -0700 | [diff] [blame] | 2180 | if (tp->fastopen_rsk) |
Nandita Dukkipati | 6ba8a3b | 2013-03-11 10:00:43 +0000 | [diff] [blame] | 2181 | return false; |
| 2182 | |
| 2183 | /* TLP is only scheduled when next timer event is RTO. */ |
| 2184 | if (icsk->icsk_pending != ICSK_TIME_RETRANS) |
| 2185 | return false; |
| 2186 | |
| 2187 | /* Schedule a loss probe in 2*RTT for SACK capable connections |
| 2188 | * in Open state, that are either limited by cwnd or application. |
| 2189 | */ |
Yuchung Cheng | f9b9958 | 2015-09-18 11:40:33 -0700 | [diff] [blame] | 2190 | if (sysctl_tcp_early_retrans < 3 || !tp->packets_out || |
Nandita Dukkipati | 6ba8a3b | 2013-03-11 10:00:43 +0000 | [diff] [blame] | 2191 | !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open) |
| 2192 | return false; |
| 2193 | |
| 2194 | if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) && |
| 2195 | tcp_send_head(sk)) |
| 2196 | return false; |
| 2197 | |
| 2198 | /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account |
Yuchung Cheng | f9b9958 | 2015-09-18 11:40:33 -0700 | [diff] [blame] | 2199 | * for delayed ack when there's one outstanding packet. If no RTT |
| 2200 | * sample is available then probe after TCP_TIMEOUT_INIT. |
Nandita Dukkipati | 6ba8a3b | 2013-03-11 10:00:43 +0000 | [diff] [blame] | 2201 | */ |
Yuchung Cheng | f9b9958 | 2015-09-18 11:40:33 -0700 | [diff] [blame] | 2202 | timeout = rtt << 1 ? : TCP_TIMEOUT_INIT; |
Nandita Dukkipati | 6ba8a3b | 2013-03-11 10:00:43 +0000 | [diff] [blame] | 2203 | if (tp->packets_out == 1) |
| 2204 | timeout = max_t(u32, timeout, |
| 2205 | (rtt + (rtt >> 1) + TCP_DELACK_MAX)); |
| 2206 | timeout = max_t(u32, timeout, msecs_to_jiffies(10)); |
| 2207 | |
| 2208 | /* If RTO is shorter, just schedule TLP in its place. */ |
| 2209 | tlp_time_stamp = tcp_time_stamp + timeout; |
| 2210 | rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout; |
| 2211 | if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) { |
| 2212 | s32 delta = rto_time_stamp - tcp_time_stamp; |
| 2213 | if (delta > 0) |
| 2214 | timeout = delta; |
| 2215 | } |
| 2216 | |
| 2217 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, |
| 2218 | TCP_RTO_MAX); |
| 2219 | return true; |
| 2220 | } |
| 2221 | |
Eric Dumazet | 1f3279a | 2014-04-20 17:58:17 -0700 | [diff] [blame] | 2222 | /* Thanks to skb fast clones, we can detect if a prior transmit of |
| 2223 | * a packet is still in a qdisc or driver queue. |
| 2224 | * In this case, there is very little point doing a retransmit ! |
Eric Dumazet | 1f3279a | 2014-04-20 17:58:17 -0700 | [diff] [blame] | 2225 | */ |
| 2226 | static bool skb_still_in_host_queue(const struct sock *sk, |
| 2227 | const struct sk_buff *skb) |
| 2228 | { |
Eric Dumazet | 39bb5e6 | 2014-10-30 10:32:34 -0700 | [diff] [blame] | 2229 | if (unlikely(skb_fclone_busy(sk, skb))) { |
Eric Dumazet | c10d931 | 2016-04-29 14:16:47 -0700 | [diff] [blame] | 2230 | NET_INC_STATS(sock_net(sk), |
| 2231 | LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES); |
Eric Dumazet | 1f3279a | 2014-04-20 17:58:17 -0700 | [diff] [blame] | 2232 | return true; |
| 2233 | } |
| 2234 | return false; |
| 2235 | } |
| 2236 | |
Yuchung Cheng | b340b26 | 2015-08-12 11:18:19 -0700 | [diff] [blame] | 2237 | /* When probe timeout (PTO) fires, try send a new segment if possible, else |
Nandita Dukkipati | 6ba8a3b | 2013-03-11 10:00:43 +0000 | [diff] [blame] | 2238 | * retransmit the last segment. |
| 2239 | */ |
| 2240 | void tcp_send_loss_probe(struct sock *sk) |
| 2241 | { |
Nandita Dukkipati | 9b717a8 | 2013-03-11 10:00:44 +0000 | [diff] [blame] | 2242 | struct tcp_sock *tp = tcp_sk(sk); |
Nandita Dukkipati | 6ba8a3b | 2013-03-11 10:00:43 +0000 | [diff] [blame] | 2243 | struct sk_buff *skb; |
| 2244 | int pcount; |
| 2245 | int mss = tcp_current_mss(sk); |
Nandita Dukkipati | 6ba8a3b | 2013-03-11 10:00:43 +0000 | [diff] [blame] | 2246 | |
Yuchung Cheng | b340b26 | 2015-08-12 11:18:19 -0700 | [diff] [blame] | 2247 | skb = tcp_send_head(sk); |
| 2248 | if (skb) { |
| 2249 | if (tcp_snd_wnd_test(tp, skb, mss)) { |
| 2250 | pcount = tp->packets_out; |
| 2251 | tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC); |
| 2252 | if (tp->packets_out > pcount) |
| 2253 | goto probe_sent; |
| 2254 | goto rearm_timer; |
| 2255 | } |
| 2256 | skb = tcp_write_queue_prev(sk, skb); |
| 2257 | } else { |
| 2258 | skb = tcp_write_queue_tail(sk); |
Nandita Dukkipati | 6ba8a3b | 2013-03-11 10:00:43 +0000 | [diff] [blame] | 2259 | } |
| 2260 | |
Nandita Dukkipati | 9b717a8 | 2013-03-11 10:00:44 +0000 | [diff] [blame] | 2261 | /* At most one outstanding TLP retransmission. */ |
| 2262 | if (tp->tlp_high_seq) |
| 2263 | goto rearm_timer; |
| 2264 | |
Nandita Dukkipati | 6ba8a3b | 2013-03-11 10:00:43 +0000 | [diff] [blame] | 2265 | /* Retransmit last segment. */ |
Nandita Dukkipati | 6ba8a3b | 2013-03-11 10:00:43 +0000 | [diff] [blame] | 2266 | if (WARN_ON(!skb)) |
| 2267 | goto rearm_timer; |
| 2268 | |
Eric Dumazet | 1f3279a | 2014-04-20 17:58:17 -0700 | [diff] [blame] | 2269 | if (skb_still_in_host_queue(sk, skb)) |
| 2270 | goto rearm_timer; |
| 2271 | |
Nandita Dukkipati | 6ba8a3b | 2013-03-11 10:00:43 +0000 | [diff] [blame] | 2272 | pcount = tcp_skb_pcount(skb); |
| 2273 | if (WARN_ON(!pcount)) |
| 2274 | goto rearm_timer; |
| 2275 | |
| 2276 | if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) { |
Octavian Purdila | 6cc55e0 | 2014-06-06 17:32:37 +0300 | [diff] [blame] | 2277 | if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss, |
| 2278 | GFP_ATOMIC))) |
Nandita Dukkipati | 6ba8a3b | 2013-03-11 10:00:43 +0000 | [diff] [blame] | 2279 | goto rearm_timer; |
Yuchung Cheng | b340b26 | 2015-08-12 11:18:19 -0700 | [diff] [blame] | 2280 | skb = tcp_write_queue_next(sk, skb); |
Nandita Dukkipati | 6ba8a3b | 2013-03-11 10:00:43 +0000 | [diff] [blame] | 2281 | } |
| 2282 | |
| 2283 | if (WARN_ON(!skb || !tcp_skb_pcount(skb))) |
| 2284 | goto rearm_timer; |
| 2285 | |
Eric Dumazet | 10d3be5 | 2016-04-21 10:55:23 -0700 | [diff] [blame] | 2286 | if (__tcp_retransmit_skb(sk, skb, 1)) |
Yuchung Cheng | b340b26 | 2015-08-12 11:18:19 -0700 | [diff] [blame] | 2287 | goto rearm_timer; |
Nandita Dukkipati | 6ba8a3b | 2013-03-11 10:00:43 +0000 | [diff] [blame] | 2288 | |
Nandita Dukkipati | 9b717a8 | 2013-03-11 10:00:44 +0000 | [diff] [blame] | 2289 | /* Record snd_nxt for loss detection. */ |
Yuchung Cheng | b340b26 | 2015-08-12 11:18:19 -0700 | [diff] [blame] | 2290 | tp->tlp_high_seq = tp->snd_nxt; |
Nandita Dukkipati | 9b717a8 | 2013-03-11 10:00:44 +0000 | [diff] [blame] | 2291 | |
Yuchung Cheng | b340b26 | 2015-08-12 11:18:19 -0700 | [diff] [blame] | 2292 | probe_sent: |
Eric Dumazet | c10d931 | 2016-04-29 14:16:47 -0700 | [diff] [blame] | 2293 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES); |
Yuchung Cheng | b340b26 | 2015-08-12 11:18:19 -0700 | [diff] [blame] | 2294 | /* Reset s.t. tcp_rearm_rto will restart timer from now */ |
| 2295 | inet_csk(sk)->icsk_pending = 0; |
Nandita Dukkipati | 6ba8a3b | 2013-03-11 10:00:43 +0000 | [diff] [blame] | 2296 | rearm_timer: |
Yuchung Cheng | fcd16c0 | 2015-08-12 11:18:18 -0700 | [diff] [blame] | 2297 | tcp_rearm_rto(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2298 | } |
| 2299 | |
David S. Miller | a762a98 | 2005-07-05 15:18:51 -0700 | [diff] [blame] | 2300 | /* Push out any pending frames which were held back due to |
| 2301 | * TCP_CORK or attempt at coalescing tiny packets. |
| 2302 | * The socket must be locked by the caller. |
| 2303 | */ |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2304 | void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, |
| 2305 | int nonagle) |
David S. Miller | a762a98 | 2005-07-05 15:18:51 -0700 | [diff] [blame] | 2306 | { |
Ilpo Järvinen | 726e07a | 2008-12-05 22:43:56 -0800 | [diff] [blame] | 2307 | /* If we are closed, the bytes will have to remain here. |
| 2308 | * In time closedown will finish, we empty the write queue and |
| 2309 | * all will be happy. |
| 2310 | */ |
| 2311 | if (unlikely(sk->sk_state == TCP_CLOSE)) |
| 2312 | return; |
| 2313 | |
Mel Gorman | 99a1dec | 2012-07-31 16:44:14 -0700 | [diff] [blame] | 2314 | if (tcp_write_xmit(sk, cur_mss, nonagle, 0, |
Eric Dumazet | 7450aaf | 2015-11-30 08:57:28 -0800 | [diff] [blame] | 2315 | sk_gfp_mask(sk, GFP_ATOMIC))) |
Ilpo Järvinen | 726e07a | 2008-12-05 22:43:56 -0800 | [diff] [blame] | 2316 | tcp_check_probe_timer(sk); |
David S. Miller | a762a98 | 2005-07-05 15:18:51 -0700 | [diff] [blame] | 2317 | } |
| 2318 | |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 2319 | /* Send _single_ skb sitting at the send head. This function requires |
| 2320 | * true push pending frames to setup probe timer etc. |
| 2321 | */ |
| 2322 | void tcp_push_one(struct sock *sk, unsigned int mss_now) |
| 2323 | { |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2324 | struct sk_buff *skb = tcp_send_head(sk); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 2325 | |
| 2326 | BUG_ON(!skb || skb->len < mss_now); |
| 2327 | |
Ilpo Järvinen | d5dd917 | 2008-12-05 22:48:55 -0800 | [diff] [blame] | 2328 | tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 2329 | } |
| 2330 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2331 | /* This function returns the amount that we can raise the |
| 2332 | * usable window based on the following constraints |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2333 | * |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2334 | * 1. The window can never be shrunk once it is offered (RFC 793) |
| 2335 | * 2. We limit memory per socket |
| 2336 | * |
| 2337 | * RFC 1122: |
| 2338 | * "the suggested [SWS] avoidance algorithm for the receiver is to keep |
| 2339 | * RECV.NEXT + RCV.WIN fixed until: |
| 2340 | * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" |
| 2341 | * |
| 2342 | * i.e. don't raise the right edge of the window until you can raise |
| 2343 | * it at least MSS bytes. |
| 2344 | * |
| 2345 | * Unfortunately, the recommended algorithm breaks header prediction, |
| 2346 | * since header prediction assumes th->window stays fixed. |
| 2347 | * |
| 2348 | * Strictly speaking, keeping th->window fixed violates the receiver |
| 2349 | * side SWS prevention criteria. The problem is that under this rule |
| 2350 | * a stream of single byte packets will cause the right side of the |
| 2351 | * window to always advance by a single byte. |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2352 | * |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2353 | * Of course, if the sender implements sender side SWS prevention |
| 2354 | * then this will not be a problem. |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2355 | * |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2356 | * BSD seems to make the following compromise: |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2357 | * |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2358 | * If the free space is less than the 1/4 of the maximum |
| 2359 | * space available and the free space is less than 1/2 mss, |
| 2360 | * then set the window to 0. |
| 2361 | * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] |
| 2362 | * Otherwise, just prevent the window from shrinking |
| 2363 | * and from being larger than the largest representable value. |
| 2364 | * |
| 2365 | * This prevents incremental opening of the window in the regime |
| 2366 | * where TCP is limited by the speed of the reader side taking |
| 2367 | * data out of the TCP receive queue. It does nothing about |
| 2368 | * those cases where the window is constrained on the sender side |
| 2369 | * because the pipeline is full. |
| 2370 | * |
| 2371 | * BSD also seems to "accidentally" limit itself to windows that are a |
| 2372 | * multiple of MSS, at least until the free space gets quite small. |
| 2373 | * This would appear to be a side effect of the mbuf implementation. |
| 2374 | * Combining these two algorithms results in the observed behavior |
| 2375 | * of having a fixed window size at almost all times. |
| 2376 | * |
| 2377 | * Below we obtain similar behavior by forcing the offered window to |
| 2378 | * a multiple of the mss when it is feasible to do so. |
| 2379 | * |
| 2380 | * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. |
| 2381 | * Regular options like TIMESTAMP are taken into account. |
| 2382 | */ |
| 2383 | u32 __tcp_select_window(struct sock *sk) |
| 2384 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2385 | struct inet_connection_sock *icsk = inet_csk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2386 | struct tcp_sock *tp = tcp_sk(sk); |
Stephen Hemminger | caa20d9a | 2005-11-10 17:13:47 -0800 | [diff] [blame] | 2387 | /* MSS for the peer's data. Previous versions used mss_clamp |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2388 | * here. I don't know if the value based on our guesses |
| 2389 | * of peer's MSS is better for the performance. It's more correct |
| 2390 | * but may be worse for the performance because of rcv_mss |
| 2391 | * fluctuations. --SAW 1998/11/1 |
| 2392 | */ |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2393 | int mss = icsk->icsk_ack.rcv_mss; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2394 | int free_space = tcp_space(sk); |
Florian Westphal | 86c1a04 | 2014-02-19 12:51:10 +0100 | [diff] [blame] | 2395 | int allowed_space = tcp_full_space(sk); |
| 2396 | int full_space = min_t(int, tp->window_clamp, allowed_space); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2397 | int window; |
| 2398 | |
| 2399 | if (mss > full_space) |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2400 | mss = full_space; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2401 | |
Eric Dumazet | b92edbe | 2007-12-20 21:48:32 -0800 | [diff] [blame] | 2402 | if (free_space < (full_space >> 1)) { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2403 | icsk->icsk_ack.quick = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2404 | |
Eric Dumazet | b8da51e | 2015-05-15 12:39:27 -0700 | [diff] [blame] | 2405 | if (tcp_under_memory_pressure(sk)) |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 2406 | tp->rcv_ssthresh = min(tp->rcv_ssthresh, |
| 2407 | 4U * tp->advmss); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2408 | |
Florian Westphal | 86c1a04 | 2014-02-19 12:51:10 +0100 | [diff] [blame] | 2409 | /* free_space might become our new window, make sure we don't |
| 2410 | * increase it due to wscale. |
| 2411 | */ |
| 2412 | free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale); |
| 2413 | |
| 2414 | /* if free space is less than mss estimate, or is below 1/16th |
| 2415 | * of the maximum allowed, try to move to zero-window, else |
| 2416 | * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and |
| 2417 | * new incoming data is dropped due to memory limits. |
| 2418 | * With large window, mss test triggers way too late in order |
| 2419 | * to announce zero window in time before rmem limit kicks in. |
| 2420 | */ |
| 2421 | if (free_space < (allowed_space >> 4) || free_space < mss) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2422 | return 0; |
| 2423 | } |
| 2424 | |
| 2425 | if (free_space > tp->rcv_ssthresh) |
| 2426 | free_space = tp->rcv_ssthresh; |
| 2427 | |
| 2428 | /* Don't do rounding if we are using window scaling, since the |
| 2429 | * scaled window will not line up with the MSS boundary anyway. |
| 2430 | */ |
| 2431 | window = tp->rcv_wnd; |
| 2432 | if (tp->rx_opt.rcv_wscale) { |
| 2433 | window = free_space; |
| 2434 | |
| 2435 | /* Advertise enough space so that it won't get scaled away. |
| 2436 | * Import case: prevent zero window announcement if |
| 2437 | * 1<<rcv_wscale > mss. |
| 2438 | */ |
| 2439 | if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window) |
| 2440 | window = (((window >> tp->rx_opt.rcv_wscale) + 1) |
| 2441 | << tp->rx_opt.rcv_wscale); |
| 2442 | } else { |
| 2443 | /* Get the largest window that is a nice multiple of mss. |
| 2444 | * Window clamp already applied above. |
| 2445 | * If our current window offering is within 1 mss of the |
| 2446 | * free space we just keep it. This prevents the divide |
| 2447 | * and multiply from happening most of the time. |
| 2448 | * We also don't do any window rounding when the free space |
| 2449 | * is too small. |
| 2450 | */ |
| 2451 | if (window <= free_space - mss || window > free_space) |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 2452 | window = (free_space / mss) * mss; |
John Heffner | 8456507 | 2007-04-02 13:56:32 -0700 | [diff] [blame] | 2453 | else if (mss == full_space && |
Eric Dumazet | b92edbe | 2007-12-20 21:48:32 -0800 | [diff] [blame] | 2454 | free_space > window + (full_space >> 1)) |
John Heffner | 8456507 | 2007-04-02 13:56:32 -0700 | [diff] [blame] | 2455 | window = free_space; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2456 | } |
| 2457 | |
| 2458 | return window; |
| 2459 | } |
| 2460 | |
Martin KaFai Lau | cfea5a6 | 2016-04-19 22:39:29 -0700 | [diff] [blame] | 2461 | void tcp_skb_collapse_tstamp(struct sk_buff *skb, |
| 2462 | const struct sk_buff *next_skb) |
Martin KaFai Lau | 082ac2d | 2016-04-19 22:39:28 -0700 | [diff] [blame] | 2463 | { |
Soheil Hassas Yeganeh | 0a2cf20 | 2016-04-27 23:39:01 -0400 | [diff] [blame] | 2464 | if (unlikely(tcp_has_tx_tstamp(next_skb))) { |
| 2465 | const struct skb_shared_info *next_shinfo = |
| 2466 | skb_shinfo(next_skb); |
Martin KaFai Lau | 082ac2d | 2016-04-19 22:39:28 -0700 | [diff] [blame] | 2467 | struct skb_shared_info *shinfo = skb_shinfo(skb); |
| 2468 | |
Soheil Hassas Yeganeh | 0a2cf20 | 2016-04-27 23:39:01 -0400 | [diff] [blame] | 2469 | shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP; |
Martin KaFai Lau | 082ac2d | 2016-04-19 22:39:28 -0700 | [diff] [blame] | 2470 | shinfo->tskey = next_shinfo->tskey; |
Martin KaFai Lau | 2de8023 | 2016-04-19 22:50:48 -0700 | [diff] [blame] | 2471 | TCP_SKB_CB(skb)->txstamp_ack |= |
| 2472 | TCP_SKB_CB(next_skb)->txstamp_ack; |
Martin KaFai Lau | 082ac2d | 2016-04-19 22:39:28 -0700 | [diff] [blame] | 2473 | } |
| 2474 | } |
| 2475 | |
Ilpo Järvinen | 4a17fc3 | 2008-11-24 21:03:43 -0800 | [diff] [blame] | 2476 | /* Collapses two adjacent SKB's during retransmission. */ |
| 2477 | static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2478 | { |
| 2479 | struct tcp_sock *tp = tcp_sk(sk); |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2480 | struct sk_buff *next_skb = tcp_write_queue_next(sk, skb); |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2481 | int skb_size, next_skb_size; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2482 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2483 | skb_size = skb->len; |
| 2484 | next_skb_size = next_skb->len; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2485 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2486 | BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); |
Ilpo Järvinen | a6963a6 | 2007-09-25 22:44:14 -0700 | [diff] [blame] | 2487 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2488 | tcp_highest_sack_combine(sk, next_skb, skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2489 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2490 | tcp_unlink_write_queue(next_skb, sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2491 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2492 | skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size), |
| 2493 | next_skb_size); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2494 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2495 | if (next_skb->ip_summed == CHECKSUM_PARTIAL) |
| 2496 | skb->ip_summed = CHECKSUM_PARTIAL; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2497 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2498 | if (skb->ip_summed != CHECKSUM_PARTIAL) |
| 2499 | skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2500 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2501 | /* Update sequence range on original skb. */ |
| 2502 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2503 | |
Ilpo Järvinen | e6c7d08 | 2009-02-28 04:44:35 +0000 | [diff] [blame] | 2504 | /* Merge over control information. This moves PSH/FIN etc. over */ |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 2505 | TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2506 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2507 | /* All done, get rid of second SKB and account for it so |
| 2508 | * packet counting does not break. |
| 2509 | */ |
| 2510 | TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; |
Martin KaFai Lau | a643b5d | 2016-04-25 14:44:49 -0700 | [diff] [blame] | 2511 | TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor; |
Ilpo Järvinen | b768920 | 2007-09-20 11:37:19 -0700 | [diff] [blame] | 2512 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2513 | /* changed transmit queue under us so clear hints */ |
Ilpo Järvinen | ef9da47 | 2008-09-20 21:25:15 -0700 | [diff] [blame] | 2514 | tcp_clear_retrans_hints_partial(tp); |
| 2515 | if (next_skb == tp->retransmit_skb_hint) |
| 2516 | tp->retransmit_skb_hint = skb; |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2517 | |
Ilpo Järvinen | 797108d | 2009-04-01 23:15:17 +0000 | [diff] [blame] | 2518 | tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); |
| 2519 | |
Martin KaFai Lau | 082ac2d | 2016-04-19 22:39:28 -0700 | [diff] [blame] | 2520 | tcp_skb_collapse_tstamp(skb, next_skb); |
| 2521 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2522 | sk_wmem_free_skb(sk, next_skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2523 | } |
| 2524 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 2525 | /* Check if coalescing SKBs is legal. */ |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 2526 | static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) |
Ilpo Järvinen | 4a17fc3 | 2008-11-24 21:03:43 -0800 | [diff] [blame] | 2527 | { |
| 2528 | if (tcp_skb_pcount(skb) > 1) |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 2529 | return false; |
Ilpo Järvinen | 4a17fc3 | 2008-11-24 21:03:43 -0800 | [diff] [blame] | 2530 | /* TODO: SACK collapsing could be used to remove this condition */ |
| 2531 | if (skb_shinfo(skb)->nr_frags != 0) |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 2532 | return false; |
Ilpo Järvinen | 4a17fc3 | 2008-11-24 21:03:43 -0800 | [diff] [blame] | 2533 | if (skb_cloned(skb)) |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 2534 | return false; |
Ilpo Järvinen | 4a17fc3 | 2008-11-24 21:03:43 -0800 | [diff] [blame] | 2535 | if (skb == tcp_send_head(sk)) |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 2536 | return false; |
Ilpo Järvinen | 4a17fc3 | 2008-11-24 21:03:43 -0800 | [diff] [blame] | 2537 | /* Some heurestics for collapsing over SACK'd could be invented */ |
| 2538 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 2539 | return false; |
Ilpo Järvinen | 4a17fc3 | 2008-11-24 21:03:43 -0800 | [diff] [blame] | 2540 | |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 2541 | return true; |
Ilpo Järvinen | 4a17fc3 | 2008-11-24 21:03:43 -0800 | [diff] [blame] | 2542 | } |
| 2543 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 2544 | /* Collapse packets in the retransmit queue to make to create |
| 2545 | * less packets on the wire. This is only done on retransmission. |
| 2546 | */ |
Ilpo Järvinen | 4a17fc3 | 2008-11-24 21:03:43 -0800 | [diff] [blame] | 2547 | static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, |
| 2548 | int space) |
| 2549 | { |
| 2550 | struct tcp_sock *tp = tcp_sk(sk); |
| 2551 | struct sk_buff *skb = to, *tmp; |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 2552 | bool first = true; |
Ilpo Järvinen | 4a17fc3 | 2008-11-24 21:03:43 -0800 | [diff] [blame] | 2553 | |
| 2554 | if (!sysctl_tcp_retrans_collapse) |
| 2555 | return; |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 2556 | if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) |
Ilpo Järvinen | 4a17fc3 | 2008-11-24 21:03:43 -0800 | [diff] [blame] | 2557 | return; |
| 2558 | |
| 2559 | tcp_for_write_queue_from_safe(skb, tmp, sk) { |
| 2560 | if (!tcp_can_collapse(sk, skb)) |
| 2561 | break; |
| 2562 | |
Martin KaFai Lau | a643b5d | 2016-04-25 14:44:49 -0700 | [diff] [blame] | 2563 | if (!tcp_skb_can_collapse_to(to)) |
| 2564 | break; |
| 2565 | |
Ilpo Järvinen | 4a17fc3 | 2008-11-24 21:03:43 -0800 | [diff] [blame] | 2566 | space -= skb->len; |
| 2567 | |
| 2568 | if (first) { |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 2569 | first = false; |
Ilpo Järvinen | 4a17fc3 | 2008-11-24 21:03:43 -0800 | [diff] [blame] | 2570 | continue; |
| 2571 | } |
| 2572 | |
| 2573 | if (space < 0) |
| 2574 | break; |
| 2575 | /* Punt if not enough space exists in the first SKB for |
| 2576 | * the data in the second |
| 2577 | */ |
Eric Dumazet | a21d457 | 2012-04-10 20:30:48 +0000 | [diff] [blame] | 2578 | if (skb->len > skb_availroom(to)) |
Ilpo Järvinen | 4a17fc3 | 2008-11-24 21:03:43 -0800 | [diff] [blame] | 2579 | break; |
| 2580 | |
| 2581 | if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) |
| 2582 | break; |
| 2583 | |
| 2584 | tcp_collapse_retrans(sk, to); |
| 2585 | } |
| 2586 | } |
| 2587 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2588 | /* This retransmits one SKB. Policy decisions and retransmit queue |
| 2589 | * state updates are done by the caller. Returns non-zero if an |
| 2590 | * error occurred which prevented the send. |
| 2591 | */ |
Eric Dumazet | 10d3be5 | 2016-04-21 10:55:23 -0700 | [diff] [blame] | 2592 | int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2593 | { |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 2594 | struct inet_connection_sock *icsk = inet_csk(sk); |
Eric Dumazet | 10d3be5 | 2016-04-21 10:55:23 -0700 | [diff] [blame] | 2595 | struct tcp_sock *tp = tcp_sk(sk); |
Sridhar Samudrala | 7d227cd2 | 2008-05-21 16:42:20 -0700 | [diff] [blame] | 2596 | unsigned int cur_mss; |
Eric Dumazet | 10d3be5 | 2016-04-21 10:55:23 -0700 | [diff] [blame] | 2597 | int diff, len, err; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2598 | |
Eric Dumazet | 10d3be5 | 2016-04-21 10:55:23 -0700 | [diff] [blame] | 2599 | |
| 2600 | /* Inconclusive MTU probe */ |
| 2601 | if (icsk->icsk_mtup.probe_size) |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 2602 | icsk->icsk_mtup.probe_size = 0; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 2603 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2604 | /* Do not sent more than we queued. 1/4 is reserved for possible |
Stephen Hemminger | caa20d9a | 2005-11-10 17:13:47 -0800 | [diff] [blame] | 2605 | * copying overhead: fragmentation, tunneling, mangling etc. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2606 | */ |
| 2607 | if (atomic_read(&sk->sk_wmem_alloc) > |
| 2608 | min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf)) |
| 2609 | return -EAGAIN; |
| 2610 | |
Eric Dumazet | 1f3279a | 2014-04-20 17:58:17 -0700 | [diff] [blame] | 2611 | if (skb_still_in_host_queue(sk, skb)) |
| 2612 | return -EBUSY; |
| 2613 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2614 | if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { |
| 2615 | if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) |
| 2616 | BUG(); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2617 | if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) |
| 2618 | return -ENOMEM; |
| 2619 | } |
| 2620 | |
Sridhar Samudrala | 7d227cd2 | 2008-05-21 16:42:20 -0700 | [diff] [blame] | 2621 | if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) |
| 2622 | return -EHOSTUNREACH; /* Routing failure or similar. */ |
| 2623 | |
Ilpo Järvinen | 0c54b85 | 2009-03-14 14:23:05 +0000 | [diff] [blame] | 2624 | cur_mss = tcp_current_mss(sk); |
Sridhar Samudrala | 7d227cd2 | 2008-05-21 16:42:20 -0700 | [diff] [blame] | 2625 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2626 | /* If receiver has shrunk his window, and skb is out of |
| 2627 | * new window, do not retransmit it. The exception is the |
| 2628 | * case, when window is shrunk to zero. In this case |
| 2629 | * our retransmit serves as a zero window probe. |
| 2630 | */ |
Joe Perches | 9d4fb27 | 2009-11-23 10:41:23 -0800 | [diff] [blame] | 2631 | if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && |
| 2632 | TCP_SKB_CB(skb)->seq != tp->snd_una) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2633 | return -EAGAIN; |
| 2634 | |
Eric Dumazet | 10d3be5 | 2016-04-21 10:55:23 -0700 | [diff] [blame] | 2635 | len = cur_mss * segs; |
| 2636 | if (skb->len > len) { |
| 2637 | if (tcp_fragment(sk, skb, len, cur_mss, GFP_ATOMIC)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2638 | return -ENOMEM; /* We'll try again later. */ |
Ilpo Järvinen | 02276f3 | 2009-02-28 04:44:31 +0000 | [diff] [blame] | 2639 | } else { |
Eric Dumazet | 10d3be5 | 2016-04-21 10:55:23 -0700 | [diff] [blame] | 2640 | if (skb_unclone(skb, GFP_ATOMIC)) |
| 2641 | return -ENOMEM; |
Ilpo Järvinen | 9eb9362 | 2009-04-01 23:18:20 +0000 | [diff] [blame] | 2642 | |
Eric Dumazet | 10d3be5 | 2016-04-21 10:55:23 -0700 | [diff] [blame] | 2643 | diff = tcp_skb_pcount(skb); |
| 2644 | tcp_set_skb_tso_segs(skb, cur_mss); |
| 2645 | diff -= tcp_skb_pcount(skb); |
| 2646 | if (diff) |
| 2647 | tcp_adjust_pcount(sk, skb, diff); |
| 2648 | if (skb->len < cur_mss) |
| 2649 | tcp_retrans_try_collapse(sk, skb, cur_mss); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2650 | } |
| 2651 | |
Daniel Borkmann | 4921355 | 2015-05-19 21:04:22 +0200 | [diff] [blame] | 2652 | /* RFC3168, section 6.1.1.1. ECN fallback */ |
| 2653 | if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN) |
| 2654 | tcp_ecn_clear_syn(sk, skb); |
| 2655 | |
Thomas Graf | 50bceae | 2013-04-11 10:57:18 +0000 | [diff] [blame] | 2656 | /* make sure skb->data is aligned on arches that require it |
| 2657 | * and check if ack-trimming & collapsing extended the headroom |
| 2658 | * beyond what csum_start can cover. |
| 2659 | */ |
| 2660 | if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) || |
| 2661 | skb_headroom(skb) >= 0xFFFF)) { |
Eric Dumazet | 10a8198 | 2016-05-09 20:55:16 -0700 | [diff] [blame] | 2662 | struct sk_buff *nskb; |
| 2663 | |
| 2664 | skb_mstamp_get(&skb->skb_mstamp); |
| 2665 | nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC); |
Yuchung Cheng | c84a571 | 2014-02-28 16:42:26 -0800 | [diff] [blame] | 2666 | err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) : |
| 2667 | -ENOBUFS; |
Eric Dumazet | 117632e | 2011-12-03 21:39:53 +0000 | [diff] [blame] | 2668 | } else { |
Yuchung Cheng | c84a571 | 2014-02-28 16:42:26 -0800 | [diff] [blame] | 2669 | err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); |
Eric Dumazet | 117632e | 2011-12-03 21:39:53 +0000 | [diff] [blame] | 2670 | } |
Yuchung Cheng | c84a571 | 2014-02-28 16:42:26 -0800 | [diff] [blame] | 2671 | |
Eric Dumazet | fc9f350 | 2014-04-28 22:00:29 -0700 | [diff] [blame] | 2672 | if (likely(!err)) { |
Eric Dumazet | 10d3be5 | 2016-04-21 10:55:23 -0700 | [diff] [blame] | 2673 | segs = tcp_skb_pcount(skb); |
| 2674 | |
Yuchung Cheng | c84a571 | 2014-02-28 16:42:26 -0800 | [diff] [blame] | 2675 | TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS; |
Eric Dumazet | fc9f350 | 2014-04-28 22:00:29 -0700 | [diff] [blame] | 2676 | /* Update global TCP statistics. */ |
Eric Dumazet | 10d3be5 | 2016-04-21 10:55:23 -0700 | [diff] [blame] | 2677 | TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs); |
Eric Dumazet | fc9f350 | 2014-04-28 22:00:29 -0700 | [diff] [blame] | 2678 | if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) |
Eric Dumazet | 02a1d6e | 2016-04-27 16:44:39 -0700 | [diff] [blame] | 2679 | __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); |
Eric Dumazet | 10d3be5 | 2016-04-21 10:55:23 -0700 | [diff] [blame] | 2680 | tp->total_retrans += segs; |
Eric Dumazet | fc9f350 | 2014-04-28 22:00:29 -0700 | [diff] [blame] | 2681 | } |
Yuchung Cheng | c84a571 | 2014-02-28 16:42:26 -0800 | [diff] [blame] | 2682 | return err; |
Yuchung Cheng | 93b174a | 2012-12-06 08:45:32 +0000 | [diff] [blame] | 2683 | } |
| 2684 | |
Eric Dumazet | 10d3be5 | 2016-04-21 10:55:23 -0700 | [diff] [blame] | 2685 | int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) |
Yuchung Cheng | 93b174a | 2012-12-06 08:45:32 +0000 | [diff] [blame] | 2686 | { |
| 2687 | struct tcp_sock *tp = tcp_sk(sk); |
Eric Dumazet | 10d3be5 | 2016-04-21 10:55:23 -0700 | [diff] [blame] | 2688 | int err = __tcp_retransmit_skb(sk, skb, segs); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2689 | |
| 2690 | if (err == 0) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2691 | #if FASTRETRANS_DEBUG > 0 |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 2692 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { |
Joe Perches | e87cc47 | 2012-05-13 21:56:26 +0000 | [diff] [blame] | 2693 | net_dbg_ratelimited("retrans_out leaked\n"); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2694 | } |
| 2695 | #endif |
| 2696 | TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; |
| 2697 | tp->retrans_out += tcp_skb_pcount(skb); |
| 2698 | |
| 2699 | /* Save stamp of the first retransmit. */ |
| 2700 | if (!tp->retrans_stamp) |
Eric Dumazet | 7faee5c | 2014-09-05 15:33:33 -0700 | [diff] [blame] | 2701 | tp->retrans_stamp = tcp_skb_timestamp(skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2702 | |
Eric Dumazet | 1f3279a | 2014-04-20 17:58:17 -0700 | [diff] [blame] | 2703 | } else if (err != -EBUSY) { |
Eric Dumazet | c10d931 | 2016-04-29 14:16:47 -0700 | [diff] [blame] | 2704 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2705 | } |
Yuchung Cheng | 6e08d5e | 2014-07-02 12:07:16 -0700 | [diff] [blame] | 2706 | |
| 2707 | if (tp->undo_retrans < 0) |
| 2708 | tp->undo_retrans = 0; |
| 2709 | tp->undo_retrans += tcp_skb_pcount(skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2710 | return err; |
| 2711 | } |
| 2712 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 2713 | /* Check if we forward retransmits are possible in the current |
| 2714 | * window/congestion state. |
| 2715 | */ |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 2716 | static bool tcp_can_forward_retransmit(struct sock *sk) |
Ilpo Järvinen | b5afe7b | 2008-09-20 21:21:54 -0700 | [diff] [blame] | 2717 | { |
| 2718 | const struct inet_connection_sock *icsk = inet_csk(sk); |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 2719 | const struct tcp_sock *tp = tcp_sk(sk); |
Ilpo Järvinen | b5afe7b | 2008-09-20 21:21:54 -0700 | [diff] [blame] | 2720 | |
| 2721 | /* Forward retransmissions are possible only during Recovery. */ |
| 2722 | if (icsk->icsk_ca_state != TCP_CA_Recovery) |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 2723 | return false; |
Ilpo Järvinen | b5afe7b | 2008-09-20 21:21:54 -0700 | [diff] [blame] | 2724 | |
| 2725 | /* No forward retransmissions in Reno are possible. */ |
| 2726 | if (tcp_is_reno(tp)) |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 2727 | return false; |
Ilpo Järvinen | b5afe7b | 2008-09-20 21:21:54 -0700 | [diff] [blame] | 2728 | |
| 2729 | /* Yeah, we have to make difficult choice between forward transmission |
| 2730 | * and retransmission... Both ways have their merits... |
| 2731 | * |
| 2732 | * For now we do not retransmit anything, while we have some new |
| 2733 | * segments to send. In the other cases, follow rule 3 for |
| 2734 | * NextSeg() specified in RFC3517. |
| 2735 | */ |
| 2736 | |
| 2737 | if (tcp_may_send_now(sk)) |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 2738 | return false; |
Ilpo Järvinen | b5afe7b | 2008-09-20 21:21:54 -0700 | [diff] [blame] | 2739 | |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 2740 | return true; |
Ilpo Järvinen | b5afe7b | 2008-09-20 21:21:54 -0700 | [diff] [blame] | 2741 | } |
| 2742 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2743 | /* This gets called after a retransmit timeout, and the initially |
| 2744 | * retransmitted data is acknowledged. It tries to continue |
| 2745 | * resending the rest of the retransmit queue, until either |
| 2746 | * we've sent it all or the congestion window limit is reached. |
| 2747 | * If doing SACK, the first ACK which comes back for a timeout |
| 2748 | * based retransmit packet might feed us FACK information again. |
| 2749 | * If so, we use it to avoid unnecessarily retransmissions. |
| 2750 | */ |
| 2751 | void tcp_xmit_retransmit_queue(struct sock *sk) |
| 2752 | { |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2753 | const struct inet_connection_sock *icsk = inet_csk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2754 | struct tcp_sock *tp = tcp_sk(sk); |
| 2755 | struct sk_buff *skb; |
Ilpo Järvinen | 0e1c54c | 2008-09-20 21:24:21 -0700 | [diff] [blame] | 2756 | struct sk_buff *hole = NULL; |
Eric Dumazet | a3d2e9f | 2016-06-27 17:38:50 +0200 | [diff] [blame] | 2757 | u32 max_segs, last_lost; |
Ilpo Järvinen | 61eb55f | 2008-09-20 21:22:59 -0700 | [diff] [blame] | 2758 | int mib_idx; |
Ilpo Järvinen | 0e1c54c | 2008-09-20 21:24:21 -0700 | [diff] [blame] | 2759 | int fwd_rexmitting = 0; |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 2760 | |
Ilpo Järvinen | 45e77d3 | 2010-07-19 01:16:18 +0000 | [diff] [blame] | 2761 | if (!tp->packets_out) |
| 2762 | return; |
| 2763 | |
Ilpo Järvinen | 08ebd17 | 2008-09-20 21:23:49 -0700 | [diff] [blame] | 2764 | if (!tp->lost_out) |
| 2765 | tp->retransmit_high = tp->snd_una; |
| 2766 | |
Ilpo Järvinen | 618d9f2 | 2008-09-20 21:26:22 -0700 | [diff] [blame] | 2767 | if (tp->retransmit_skb_hint) { |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 2768 | skb = tp->retransmit_skb_hint; |
Ilpo Järvinen | 618d9f2 | 2008-09-20 21:26:22 -0700 | [diff] [blame] | 2769 | last_lost = TCP_SKB_CB(skb)->end_seq; |
| 2770 | if (after(last_lost, tp->retransmit_high)) |
| 2771 | last_lost = tp->retransmit_high; |
| 2772 | } else { |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2773 | skb = tcp_write_queue_head(sk); |
Ilpo Järvinen | 618d9f2 | 2008-09-20 21:26:22 -0700 | [diff] [blame] | 2774 | last_lost = tp->snd_una; |
| 2775 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2776 | |
Eric Dumazet | a3d2e9f | 2016-06-27 17:38:50 +0200 | [diff] [blame] | 2777 | max_segs = tcp_tso_autosize(sk, tcp_current_mss(sk)); |
Ilpo Järvinen | 08ebd17 | 2008-09-20 21:23:49 -0700 | [diff] [blame] | 2778 | tcp_for_write_queue_from(skb, sk) { |
| 2779 | __u8 sacked = TCP_SKB_CB(skb)->sacked; |
Eric Dumazet | 10d3be5 | 2016-04-21 10:55:23 -0700 | [diff] [blame] | 2780 | int segs; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2781 | |
Ilpo Järvinen | 08ebd17 | 2008-09-20 21:23:49 -0700 | [diff] [blame] | 2782 | if (skb == tcp_send_head(sk)) |
| 2783 | break; |
| 2784 | /* we could do better than to assign each time */ |
Ian Morris | 51456b2 | 2015-04-03 09:17:26 +0100 | [diff] [blame] | 2785 | if (!hole) |
Ilpo Järvinen | 0e1c54c | 2008-09-20 21:24:21 -0700 | [diff] [blame] | 2786 | tp->retransmit_skb_hint = skb; |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 2787 | |
Eric Dumazet | 10d3be5 | 2016-04-21 10:55:23 -0700 | [diff] [blame] | 2788 | segs = tp->snd_cwnd - tcp_packets_in_flight(tp); |
| 2789 | if (segs <= 0) |
Ilpo Järvinen | 08ebd17 | 2008-09-20 21:23:49 -0700 | [diff] [blame] | 2790 | return; |
Eric Dumazet | a3d2e9f | 2016-06-27 17:38:50 +0200 | [diff] [blame] | 2791 | /* In case tcp_shift_skb_data() have aggregated large skbs, |
| 2792 | * we need to make sure not sending too bigs TSO packets |
| 2793 | */ |
| 2794 | segs = min_t(int, segs, max_segs); |
Ilpo Järvinen | 0e1c54c | 2008-09-20 21:24:21 -0700 | [diff] [blame] | 2795 | |
| 2796 | if (fwd_rexmitting) { |
| 2797 | begin_fwd: |
| 2798 | if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp))) |
| 2799 | break; |
| 2800 | mib_idx = LINUX_MIB_TCPFORWARDRETRANS; |
| 2801 | |
| 2802 | } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) { |
Ilpo Järvinen | 618d9f2 | 2008-09-20 21:26:22 -0700 | [diff] [blame] | 2803 | tp->retransmit_high = last_lost; |
Ilpo Järvinen | 0e1c54c | 2008-09-20 21:24:21 -0700 | [diff] [blame] | 2804 | if (!tcp_can_forward_retransmit(sk)) |
| 2805 | break; |
| 2806 | /* Backtrack if necessary to non-L'ed skb */ |
Ian Morris | 00db412 | 2015-04-03 09:17:27 +0100 | [diff] [blame] | 2807 | if (hole) { |
Ilpo Järvinen | 0e1c54c | 2008-09-20 21:24:21 -0700 | [diff] [blame] | 2808 | skb = hole; |
| 2809 | hole = NULL; |
| 2810 | } |
| 2811 | fwd_rexmitting = 1; |
| 2812 | goto begin_fwd; |
| 2813 | |
| 2814 | } else if (!(sacked & TCPCB_LOST)) { |
Ian Morris | 51456b2 | 2015-04-03 09:17:26 +0100 | [diff] [blame] | 2815 | if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) |
Ilpo Järvinen | 0e1c54c | 2008-09-20 21:24:21 -0700 | [diff] [blame] | 2816 | hole = skb; |
Ilpo Järvinen | 08ebd17 | 2008-09-20 21:23:49 -0700 | [diff] [blame] | 2817 | continue; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2818 | |
Ilpo Järvinen | 0e1c54c | 2008-09-20 21:24:21 -0700 | [diff] [blame] | 2819 | } else { |
Ilpo Järvinen | 618d9f2 | 2008-09-20 21:26:22 -0700 | [diff] [blame] | 2820 | last_lost = TCP_SKB_CB(skb)->end_seq; |
Ilpo Järvinen | 0e1c54c | 2008-09-20 21:24:21 -0700 | [diff] [blame] | 2821 | if (icsk->icsk_ca_state != TCP_CA_Loss) |
| 2822 | mib_idx = LINUX_MIB_TCPFASTRETRANS; |
| 2823 | else |
| 2824 | mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; |
| 2825 | } |
| 2826 | |
| 2827 | if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) |
Ilpo Järvinen | 08ebd17 | 2008-09-20 21:23:49 -0700 | [diff] [blame] | 2828 | continue; |
Pavel Emelyanov | 40b215e | 2008-07-03 01:05:41 -0700 | [diff] [blame] | 2829 | |
Eric Dumazet | 10d3be5 | 2016-04-21 10:55:23 -0700 | [diff] [blame] | 2830 | if (tcp_retransmit_skb(sk, skb, segs)) |
Ilpo Järvinen | 08ebd17 | 2008-09-20 21:23:49 -0700 | [diff] [blame] | 2831 | return; |
Yuchung Cheng | 24ab6be | 2013-07-12 11:33:04 -0700 | [diff] [blame] | 2832 | |
Eric Dumazet | c10d931 | 2016-04-29 14:16:47 -0700 | [diff] [blame] | 2833 | NET_INC_STATS(sock_net(sk), mib_idx); |
Ilpo Järvinen | 08ebd17 | 2008-09-20 21:23:49 -0700 | [diff] [blame] | 2834 | |
Yuchung Cheng | 684bad1 | 2012-09-02 17:38:04 +0000 | [diff] [blame] | 2835 | if (tcp_in_cwnd_reduction(sk)) |
Nandita Dukkipati | a262f0c | 2011-08-21 20:21:57 +0000 | [diff] [blame] | 2836 | tp->prr_out += tcp_skb_pcount(skb); |
| 2837 | |
Ilpo Järvinen | 08ebd17 | 2008-09-20 21:23:49 -0700 | [diff] [blame] | 2838 | if (skb == tcp_write_queue_head(sk)) |
| 2839 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, |
| 2840 | inet_csk(sk)->icsk_rto, |
| 2841 | TCP_RTO_MAX); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2842 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2843 | } |
| 2844 | |
Eric Dumazet | d83769a | 2015-04-21 18:32:24 -0700 | [diff] [blame] | 2845 | /* We allow to exceed memory limits for FIN packets to expedite |
| 2846 | * connection tear down and (memory) recovery. |
Eric Dumazet | 845704a | 2015-04-23 10:42:39 -0700 | [diff] [blame] | 2847 | * Otherwise tcp_send_fin() could be tempted to either delay FIN |
| 2848 | * or even be forced to close flow without any FIN. |
Eric Dumazet | a6c5ea4 | 2015-05-15 12:39:26 -0700 | [diff] [blame] | 2849 | * In general, we want to allow one skb per socket to avoid hangs |
| 2850 | * with edge trigger epoll() |
Eric Dumazet | d83769a | 2015-04-21 18:32:24 -0700 | [diff] [blame] | 2851 | */ |
Eric Dumazet | a6c5ea4 | 2015-05-15 12:39:26 -0700 | [diff] [blame] | 2852 | void sk_forced_mem_schedule(struct sock *sk, int size) |
Eric Dumazet | d83769a | 2015-04-21 18:32:24 -0700 | [diff] [blame] | 2853 | { |
Johannes Weiner | e805605 | 2016-01-14 15:21:14 -0800 | [diff] [blame] | 2854 | int amt; |
Eric Dumazet | d83769a | 2015-04-21 18:32:24 -0700 | [diff] [blame] | 2855 | |
| 2856 | if (size <= sk->sk_forward_alloc) |
| 2857 | return; |
| 2858 | amt = sk_mem_pages(size); |
| 2859 | sk->sk_forward_alloc += amt * SK_MEM_QUANTUM; |
Johannes Weiner | e805605 | 2016-01-14 15:21:14 -0800 | [diff] [blame] | 2860 | sk_memory_allocated_add(sk, amt); |
| 2861 | |
Johannes Weiner | baac50b | 2016-01-14 15:21:17 -0800 | [diff] [blame] | 2862 | if (mem_cgroup_sockets_enabled && sk->sk_memcg) |
| 2863 | mem_cgroup_charge_skmem(sk->sk_memcg, amt); |
Eric Dumazet | d83769a | 2015-04-21 18:32:24 -0700 | [diff] [blame] | 2864 | } |
| 2865 | |
Eric Dumazet | 845704a | 2015-04-23 10:42:39 -0700 | [diff] [blame] | 2866 | /* Send a FIN. The caller locks the socket for us. |
| 2867 | * We should try to send a FIN packet really hard, but eventually give up. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2868 | */ |
| 2869 | void tcp_send_fin(struct sock *sk) |
| 2870 | { |
Eric Dumazet | 845704a | 2015-04-23 10:42:39 -0700 | [diff] [blame] | 2871 | struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk); |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2872 | struct tcp_sock *tp = tcp_sk(sk); |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2873 | |
Eric Dumazet | 845704a | 2015-04-23 10:42:39 -0700 | [diff] [blame] | 2874 | /* Optimization, tack on the FIN if we have one skb in write queue and |
| 2875 | * this skb was not yet sent, or we are under memory pressure. |
| 2876 | * Note: in the latter case, FIN packet will be sent after a timeout, |
| 2877 | * as TCP stack thinks it has already been transmitted. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2878 | */ |
Eric Dumazet | b8da51e | 2015-05-15 12:39:27 -0700 | [diff] [blame] | 2879 | if (tskb && (tcp_send_head(sk) || tcp_under_memory_pressure(sk))) { |
Eric Dumazet | 845704a | 2015-04-23 10:42:39 -0700 | [diff] [blame] | 2880 | coalesce: |
| 2881 | TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN; |
| 2882 | TCP_SKB_CB(tskb)->end_seq++; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2883 | tp->write_seq++; |
Eric Dumazet | 845704a | 2015-04-23 10:42:39 -0700 | [diff] [blame] | 2884 | if (!tcp_send_head(sk)) { |
| 2885 | /* This means tskb was already sent. |
| 2886 | * Pretend we included the FIN on previous transmit. |
| 2887 | * We need to set tp->snd_nxt to the value it would have |
| 2888 | * if FIN had been sent. This is because retransmit path |
| 2889 | * does not change tp->snd_nxt. |
| 2890 | */ |
| 2891 | tp->snd_nxt++; |
| 2892 | return; |
| 2893 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2894 | } else { |
Eric Dumazet | 845704a | 2015-04-23 10:42:39 -0700 | [diff] [blame] | 2895 | skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation); |
| 2896 | if (unlikely(!skb)) { |
| 2897 | if (tskb) |
| 2898 | goto coalesce; |
| 2899 | return; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2900 | } |
Eric Dumazet | d83769a | 2015-04-21 18:32:24 -0700 | [diff] [blame] | 2901 | skb_reserve(skb, MAX_TCP_HEADER); |
Eric Dumazet | a6c5ea4 | 2015-05-15 12:39:26 -0700 | [diff] [blame] | 2902 | sk_forced_mem_schedule(sk, skb->truesize); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2903 | /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ |
Ilpo Järvinen | e870a8e | 2008-01-03 20:39:01 -0800 | [diff] [blame] | 2904 | tcp_init_nondata_skb(skb, tp->write_seq, |
Changli Gao | a3433f3 | 2010-06-12 14:01:43 +0000 | [diff] [blame] | 2905 | TCPHDR_ACK | TCPHDR_FIN); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2906 | tcp_queue_skb(sk, skb); |
| 2907 | } |
Eric Dumazet | 845704a | 2015-04-23 10:42:39 -0700 | [diff] [blame] | 2908 | __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2909 | } |
| 2910 | |
| 2911 | /* We get here when a process closes a file descriptor (either due to |
| 2912 | * an explicit close() or as a byproduct of exit()'ing) and there |
| 2913 | * was unread data in the receive queue. This behavior is recommended |
Gerrit Renker | 65bb723 | 2007-04-28 21:21:46 -0700 | [diff] [blame] | 2914 | * by RFC 2525, section 2.17. -DaveM |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2915 | */ |
Al Viro | dd0fc66 | 2005-10-07 07:46:04 +0100 | [diff] [blame] | 2916 | void tcp_send_active_reset(struct sock *sk, gfp_t priority) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2917 | { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2918 | struct sk_buff *skb; |
| 2919 | |
| 2920 | /* NOTE: No TCP options attached and we never retransmit this. */ |
| 2921 | skb = alloc_skb(MAX_TCP_HEADER, priority); |
| 2922 | if (!skb) { |
Pavel Emelyanov | 4e67344 | 2008-07-16 20:30:14 -0700 | [diff] [blame] | 2923 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2924 | return; |
| 2925 | } |
| 2926 | |
| 2927 | /* Reserve space for headers and prepare control bits. */ |
| 2928 | skb_reserve(skb, MAX_TCP_HEADER); |
Ilpo Järvinen | e870a8e | 2008-01-03 20:39:01 -0800 | [diff] [blame] | 2929 | tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), |
Changli Gao | a3433f3 | 2010-06-12 14:01:43 +0000 | [diff] [blame] | 2930 | TCPHDR_ACK | TCPHDR_RST); |
Eric Dumazet | 675ee23 | 2015-09-23 14:00:21 -0700 | [diff] [blame] | 2931 | skb_mstamp_get(&skb->skb_mstamp); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2932 | /* Send it off. */ |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 2933 | if (tcp_transmit_skb(sk, skb, 0, priority)) |
Pavel Emelyanov | 4e67344 | 2008-07-16 20:30:14 -0700 | [diff] [blame] | 2934 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); |
Sridhar Samudrala | 26af65c | 2008-06-04 15:19:35 -0700 | [diff] [blame] | 2935 | |
Pavel Emelyanov | 81cc8a7 | 2008-07-16 20:22:04 -0700 | [diff] [blame] | 2936 | TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2937 | } |
| 2938 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 2939 | /* Send a crossed SYN-ACK during socket establishment. |
| 2940 | * WARNING: This routine must only be called when we have already sent |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2941 | * a SYN packet that crossed the incoming SYN that caused this routine |
| 2942 | * to get called. If this assumption fails then the initial rcv_wnd |
| 2943 | * and rcv_wscale values will not be correct. |
| 2944 | */ |
| 2945 | int tcp_send_synack(struct sock *sk) |
| 2946 | { |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 2947 | struct sk_buff *skb; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2948 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2949 | skb = tcp_write_queue_head(sk); |
Ian Morris | 51456b2 | 2015-04-03 09:17:26 +0100 | [diff] [blame] | 2950 | if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { |
Joe Perches | 91df42b | 2012-05-15 14:11:54 +0000 | [diff] [blame] | 2951 | pr_debug("%s: wrong queue state\n", __func__); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2952 | return -EFAULT; |
| 2953 | } |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 2954 | if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2955 | if (skb_cloned(skb)) { |
| 2956 | struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC); |
Ian Morris | 51456b2 | 2015-04-03 09:17:26 +0100 | [diff] [blame] | 2957 | if (!nskb) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2958 | return -ENOMEM; |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2959 | tcp_unlink_write_queue(skb, sk); |
Eric Dumazet | f4a775d | 2014-09-22 16:29:32 -0700 | [diff] [blame] | 2960 | __skb_header_release(nskb); |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2961 | __tcp_add_write_queue_head(sk, nskb); |
Hideo Aoki | 3ab224b | 2007-12-31 00:11:19 -0800 | [diff] [blame] | 2962 | sk_wmem_free_skb(sk, skb); |
| 2963 | sk->sk_wmem_queued += nskb->truesize; |
| 2964 | sk_mem_charge(sk, nskb->truesize); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2965 | skb = nskb; |
| 2966 | } |
| 2967 | |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 2968 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; |
Florian Westphal | 735d383 | 2014-09-29 13:08:30 +0200 | [diff] [blame] | 2969 | tcp_ecn_send_synack(sk, skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2970 | } |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 2971 | return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2972 | } |
| 2973 | |
Eric Dumazet | 4aea39c | 2012-06-03 20:33:21 +0000 | [diff] [blame] | 2974 | /** |
| 2975 | * tcp_make_synack - Prepare a SYN-ACK. |
| 2976 | * sk: listener socket |
| 2977 | * dst: dst entry attached to the SYNACK |
| 2978 | * req: request_sock pointer |
Eric Dumazet | 4aea39c | 2012-06-03 20:33:21 +0000 | [diff] [blame] | 2979 | * |
| 2980 | * Allocate one skb and build a SYNACK packet. |
| 2981 | * @dst is consumed : Caller should not use it again. |
| 2982 | */ |
Eric Dumazet | 5d062de | 2015-09-25 07:39:19 -0700 | [diff] [blame] | 2983 | struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst, |
William Allen Simpson | e6b4d11 | 2009-12-02 18:07:39 +0000 | [diff] [blame] | 2984 | struct request_sock *req, |
Eric Dumazet | ca6fb06 | 2015-10-02 11:43:35 -0700 | [diff] [blame] | 2985 | struct tcp_fastopen_cookie *foc, |
Eric Dumazet | b3d0514 | 2016-04-13 22:05:39 -0700 | [diff] [blame] | 2986 | enum tcp_synack_type synack_type) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2987 | { |
Arnaldo Carvalho de Melo | 2e6599c | 2005-06-18 22:46:52 -0700 | [diff] [blame] | 2988 | struct inet_request_sock *ireq = inet_rsk(req); |
Eric Dumazet | 5d062de | 2015-09-25 07:39:19 -0700 | [diff] [blame] | 2989 | const struct tcp_sock *tp = tcp_sk(sk); |
Eric Dumazet | 80f03e2 | 2015-03-24 15:58:52 -0700 | [diff] [blame] | 2990 | struct tcp_md5sig_key *md5 = NULL; |
Eric Dumazet | 5d062de | 2015-09-25 07:39:19 -0700 | [diff] [blame] | 2991 | struct tcp_out_options opts; |
| 2992 | struct sk_buff *skb; |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 2993 | int tcp_header_size; |
Eric Dumazet | 5d062de | 2015-09-25 07:39:19 -0700 | [diff] [blame] | 2994 | struct tcphdr *th; |
| 2995 | u16 user_mss; |
Tom Quetchenbach | f5fff5d | 2008-09-21 00:21:51 -0700 | [diff] [blame] | 2996 | int mss; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2997 | |
Eric Dumazet | ca6fb06 | 2015-10-02 11:43:35 -0700 | [diff] [blame] | 2998 | skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); |
Eric Dumazet | 4aea39c | 2012-06-03 20:33:21 +0000 | [diff] [blame] | 2999 | if (unlikely(!skb)) { |
| 3000 | dst_release(dst); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3001 | return NULL; |
Eric Dumazet | 4aea39c | 2012-06-03 20:33:21 +0000 | [diff] [blame] | 3002 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3003 | /* Reserve space for headers. */ |
| 3004 | skb_reserve(skb, MAX_TCP_HEADER); |
| 3005 | |
Eric Dumazet | b3d0514 | 2016-04-13 22:05:39 -0700 | [diff] [blame] | 3006 | switch (synack_type) { |
| 3007 | case TCP_SYNACK_NORMAL: |
Eric Dumazet | 9e17f8a | 2015-11-01 15:36:55 -0800 | [diff] [blame] | 3008 | skb_set_owner_w(skb, req_to_sk(req)); |
Eric Dumazet | b3d0514 | 2016-04-13 22:05:39 -0700 | [diff] [blame] | 3009 | break; |
| 3010 | case TCP_SYNACK_COOKIE: |
| 3011 | /* Under synflood, we do not attach skb to a socket, |
| 3012 | * to avoid false sharing. |
| 3013 | */ |
| 3014 | break; |
| 3015 | case TCP_SYNACK_FASTOPEN: |
Eric Dumazet | ca6fb06 | 2015-10-02 11:43:35 -0700 | [diff] [blame] | 3016 | /* sk is a const pointer, because we want to express multiple |
| 3017 | * cpu might call us concurrently. |
| 3018 | * sk->sk_wmem_alloc in an atomic, we can promote to rw. |
| 3019 | */ |
| 3020 | skb_set_owner_w(skb, (struct sock *)sk); |
Eric Dumazet | b3d0514 | 2016-04-13 22:05:39 -0700 | [diff] [blame] | 3021 | break; |
Eric Dumazet | ca6fb06 | 2015-10-02 11:43:35 -0700 | [diff] [blame] | 3022 | } |
Eric Dumazet | 4aea39c | 2012-06-03 20:33:21 +0000 | [diff] [blame] | 3023 | skb_dst_set(skb, dst); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3024 | |
David S. Miller | 0dbaee3 | 2010-12-13 12:52:14 -0800 | [diff] [blame] | 3025 | mss = dst_metric_advmss(dst); |
Eric Dumazet | 5d062de | 2015-09-25 07:39:19 -0700 | [diff] [blame] | 3026 | user_mss = READ_ONCE(tp->rx_opt.user_mss); |
| 3027 | if (user_mss && user_mss < mss) |
| 3028 | mss = user_mss; |
Tom Quetchenbach | f5fff5d | 2008-09-21 00:21:51 -0700 | [diff] [blame] | 3029 | |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 3030 | memset(&opts, 0, sizeof(opts)); |
Florian Westphal | 8b5f12d | 2008-10-26 23:10:12 -0700 | [diff] [blame] | 3031 | #ifdef CONFIG_SYN_COOKIES |
| 3032 | if (unlikely(req->cookie_ts)) |
Eric Dumazet | 7faee5c | 2014-09-05 15:33:33 -0700 | [diff] [blame] | 3033 | skb->skb_mstamp.stamp_jiffies = cookie_init_timestamp(req); |
Florian Westphal | 8b5f12d | 2008-10-26 23:10:12 -0700 | [diff] [blame] | 3034 | else |
| 3035 | #endif |
Eric Dumazet | 7faee5c | 2014-09-05 15:33:33 -0700 | [diff] [blame] | 3036 | skb_mstamp_get(&skb->skb_mstamp); |
Eric Dumazet | 80f03e2 | 2015-03-24 15:58:52 -0700 | [diff] [blame] | 3037 | |
| 3038 | #ifdef CONFIG_TCP_MD5SIG |
| 3039 | rcu_read_lock(); |
Eric Dumazet | fd3a154 | 2015-03-24 15:58:56 -0700 | [diff] [blame] | 3040 | md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req)); |
Eric Dumazet | 80f03e2 | 2015-03-24 15:58:52 -0700 | [diff] [blame] | 3041 | #endif |
Eric Dumazet | 58d607d | 2015-09-15 15:24:20 -0700 | [diff] [blame] | 3042 | skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4); |
Eric Dumazet | 37bfbdd | 2015-09-25 07:39:17 -0700 | [diff] [blame] | 3043 | tcp_header_size = tcp_synack_options(req, mss, skb, &opts, md5, foc) + |
| 3044 | sizeof(*th); |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 3045 | |
Arnaldo Carvalho de Melo | aa8223c | 2007-04-10 21:04:22 -0700 | [diff] [blame] | 3046 | skb_push(skb, tcp_header_size); |
| 3047 | skb_reset_transport_header(skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3048 | |
Eric Dumazet | ea1627c | 2016-05-13 09:16:40 -0700 | [diff] [blame] | 3049 | th = (struct tcphdr *)skb->data; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3050 | memset(th, 0, sizeof(struct tcphdr)); |
| 3051 | th->syn = 1; |
| 3052 | th->ack = 1; |
Eric Dumazet | 6ac705b | 2015-09-25 07:39:18 -0700 | [diff] [blame] | 3053 | tcp_ecn_make_synack(req, th); |
Eric Dumazet | b44084c | 2013-10-10 00:04:37 -0700 | [diff] [blame] | 3054 | th->source = htons(ireq->ir_num); |
Eric Dumazet | 634fb979 | 2013-10-09 15:21:29 -0700 | [diff] [blame] | 3055 | th->dest = ireq->ir_rmt_port; |
Ilpo Järvinen | e870a8e | 2008-01-03 20:39:01 -0800 | [diff] [blame] | 3056 | /* Setting of flags are superfluous here for callers (and ECE is |
| 3057 | * not even correctly set) |
| 3058 | */ |
| 3059 | tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn, |
Changli Gao | a3433f3 | 2010-06-12 14:01:43 +0000 | [diff] [blame] | 3060 | TCPHDR_SYN | TCPHDR_ACK); |
William Allen Simpson | 4957faade | 2009-12-02 18:25:27 +0000 | [diff] [blame] | 3061 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3062 | th->seq = htonl(TCP_SKB_CB(skb)->seq); |
Jerry Chu | 8336886 | 2012-08-31 12:29:12 +0000 | [diff] [blame] | 3063 | /* XXX data is queued and acked as is. No buffer/window check */ |
| 3064 | th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3065 | |
| 3066 | /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ |
Eric Dumazet | ed53d0a | 2015-10-08 19:33:23 -0700 | [diff] [blame] | 3067 | th->window = htons(min(req->rsk_rcv_wnd, 65535U)); |
Eric Dumazet | 5d062de | 2015-09-25 07:39:19 -0700 | [diff] [blame] | 3068 | tcp_options_write((__be32 *)(th + 1), NULL, &opts); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3069 | th->doff = (tcp_header_size >> 2); |
Eric Dumazet | 90bbcc6 | 2016-04-27 16:44:32 -0700 | [diff] [blame] | 3070 | __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS); |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 3071 | |
| 3072 | #ifdef CONFIG_TCP_MD5SIG |
| 3073 | /* Okay, we have all we need - do the md5 hash if needed */ |
Eric Dumazet | 80f03e2 | 2015-03-24 15:58:52 -0700 | [diff] [blame] | 3074 | if (md5) |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 3075 | tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, |
Eric Dumazet | 39f8e58 | 2015-03-24 15:58:55 -0700 | [diff] [blame] | 3076 | md5, req_to_sk(req), skb); |
Eric Dumazet | 80f03e2 | 2015-03-24 15:58:52 -0700 | [diff] [blame] | 3077 | rcu_read_unlock(); |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 3078 | #endif |
| 3079 | |
Eric Dumazet | b50edd7 | 2015-04-09 13:31:56 -0700 | [diff] [blame] | 3080 | /* Do not fool tcpdump (if any), clean our debris */ |
| 3081 | skb->tstamp.tv64 = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3082 | return skb; |
| 3083 | } |
Eric Dumazet | 4bc2f18 | 2010-07-09 21:22:10 +0000 | [diff] [blame] | 3084 | EXPORT_SYMBOL(tcp_make_synack); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3085 | |
Daniel Borkmann | 8116441 | 2015-01-05 23:57:48 +0100 | [diff] [blame] | 3086 | static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst) |
| 3087 | { |
| 3088 | struct inet_connection_sock *icsk = inet_csk(sk); |
| 3089 | const struct tcp_congestion_ops *ca; |
| 3090 | u32 ca_key = dst_metric(dst, RTAX_CC_ALGO); |
| 3091 | |
| 3092 | if (ca_key == TCP_CA_UNSPEC) |
| 3093 | return; |
| 3094 | |
| 3095 | rcu_read_lock(); |
| 3096 | ca = tcp_ca_find_key(ca_key); |
| 3097 | if (likely(ca && try_module_get(ca->owner))) { |
| 3098 | module_put(icsk->icsk_ca_ops->owner); |
| 3099 | icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst); |
| 3100 | icsk->icsk_ca_ops = ca; |
| 3101 | } |
| 3102 | rcu_read_unlock(); |
| 3103 | } |
| 3104 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 3105 | /* Do all connect socket setups that can be done AF independent. */ |
stephen hemminger | f7e56a7 | 2013-12-29 11:39:51 -0800 | [diff] [blame] | 3106 | static void tcp_connect_init(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3107 | { |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 3108 | const struct dst_entry *dst = __sk_dst_get(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3109 | struct tcp_sock *tp = tcp_sk(sk); |
| 3110 | __u8 rcv_wscale; |
| 3111 | |
| 3112 | /* We'll fix this up when we get a response from the other end. |
| 3113 | * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. |
| 3114 | */ |
| 3115 | tp->tcp_header_len = sizeof(struct tcphdr) + |
David S. Miller | bb5b7c1 | 2009-12-15 20:56:42 -0800 | [diff] [blame] | 3116 | (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3117 | |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 3118 | #ifdef CONFIG_TCP_MD5SIG |
Ian Morris | 00db412 | 2015-04-03 09:17:27 +0100 | [diff] [blame] | 3119 | if (tp->af_specific->md5_lookup(sk, sk)) |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 3120 | tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; |
| 3121 | #endif |
| 3122 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3123 | /* If user gave his TCP_MAXSEG, record it to clamp */ |
| 3124 | if (tp->rx_opt.user_mss) |
| 3125 | tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; |
| 3126 | tp->max_window = 0; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 3127 | tcp_mtup_init(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3128 | tcp_sync_mss(sk, dst_mtu(dst)); |
| 3129 | |
Daniel Borkmann | 8116441 | 2015-01-05 23:57:48 +0100 | [diff] [blame] | 3130 | tcp_ca_dst_init(sk, dst); |
| 3131 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3132 | if (!tp->window_clamp) |
| 3133 | tp->window_clamp = dst_metric(dst, RTAX_WINDOW); |
David S. Miller | 0dbaee3 | 2010-12-13 12:52:14 -0800 | [diff] [blame] | 3134 | tp->advmss = dst_metric_advmss(dst); |
Tom Quetchenbach | f5fff5d | 2008-09-21 00:21:51 -0700 | [diff] [blame] | 3135 | if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss) |
| 3136 | tp->advmss = tp->rx_opt.user_mss; |
| 3137 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3138 | tcp_initialize_rcv_mss(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3139 | |
Hagen Paul Pfeifer | e88c64f | 2010-08-19 06:33:05 +0000 | [diff] [blame] | 3140 | /* limit the window selection if the user enforce a smaller rx buffer */ |
| 3141 | if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && |
| 3142 | (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) |
| 3143 | tp->window_clamp = tcp_full_space(sk); |
| 3144 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3145 | tcp_select_initial_window(tcp_full_space(sk), |
| 3146 | tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), |
| 3147 | &tp->rcv_wnd, |
| 3148 | &tp->window_clamp, |
David S. Miller | bb5b7c1 | 2009-12-15 20:56:42 -0800 | [diff] [blame] | 3149 | sysctl_tcp_window_scaling, |
laurent chavey | 31d1292 | 2009-12-15 11:15:28 +0000 | [diff] [blame] | 3150 | &rcv_wscale, |
| 3151 | dst_metric(dst, RTAX_INITRWND)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3152 | |
| 3153 | tp->rx_opt.rcv_wscale = rcv_wscale; |
| 3154 | tp->rcv_ssthresh = tp->rcv_wnd; |
| 3155 | |
| 3156 | sk->sk_err = 0; |
| 3157 | sock_reset_flag(sk, SOCK_DONE); |
| 3158 | tp->snd_wnd = 0; |
Hantzis Fotis | ee7537b | 2009-03-02 22:42:02 -0800 | [diff] [blame] | 3159 | tcp_init_wl(tp, 0); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3160 | tp->snd_una = tp->write_seq; |
| 3161 | tp->snd_sml = tp->write_seq; |
Ilpo Järvinen | 33f5f57 | 2008-10-07 14:43:06 -0700 | [diff] [blame] | 3162 | tp->snd_up = tp->write_seq; |
Pavel Emelyanov | 370816a | 2012-04-19 03:40:01 +0000 | [diff] [blame] | 3163 | tp->snd_nxt = tp->write_seq; |
Pavel Emelyanov | ee99528 | 2012-04-19 03:40:39 +0000 | [diff] [blame] | 3164 | |
| 3165 | if (likely(!tp->repair)) |
| 3166 | tp->rcv_nxt = 0; |
Andrew Vagin | c7781a6 | 2013-08-27 12:20:40 +0400 | [diff] [blame] | 3167 | else |
| 3168 | tp->rcv_tstamp = tcp_time_stamp; |
Pavel Emelyanov | ee99528 | 2012-04-19 03:40:39 +0000 | [diff] [blame] | 3169 | tp->rcv_wup = tp->rcv_nxt; |
| 3170 | tp->copied_seq = tp->rcv_nxt; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3171 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 3172 | inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; |
| 3173 | inet_csk(sk)->icsk_retransmits = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3174 | tcp_clear_retrans(tp); |
| 3175 | } |
| 3176 | |
Yuchung Cheng | 783237e | 2012-07-19 06:43:07 +0000 | [diff] [blame] | 3177 | static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) |
| 3178 | { |
| 3179 | struct tcp_sock *tp = tcp_sk(sk); |
| 3180 | struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); |
| 3181 | |
| 3182 | tcb->end_seq += skb->len; |
Eric Dumazet | f4a775d | 2014-09-22 16:29:32 -0700 | [diff] [blame] | 3183 | __skb_header_release(skb); |
Yuchung Cheng | 783237e | 2012-07-19 06:43:07 +0000 | [diff] [blame] | 3184 | __tcp_add_write_queue_tail(sk, skb); |
| 3185 | sk->sk_wmem_queued += skb->truesize; |
| 3186 | sk_mem_charge(sk, skb->truesize); |
| 3187 | tp->write_seq = tcb->end_seq; |
| 3188 | tp->packets_out += tcp_skb_pcount(skb); |
| 3189 | } |
| 3190 | |
| 3191 | /* Build and send a SYN with data and (cached) Fast Open cookie. However, |
| 3192 | * queue a data-only packet after the regular SYN, such that regular SYNs |
| 3193 | * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges |
| 3194 | * only the SYN sequence, the data are retransmitted in the first ACK. |
| 3195 | * If cookie is not cached or other error occurs, falls back to send a |
| 3196 | * regular SYN with Fast Open cookie request option. |
| 3197 | */ |
| 3198 | static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) |
| 3199 | { |
| 3200 | struct tcp_sock *tp = tcp_sk(sk); |
| 3201 | struct tcp_fastopen_request *fo = tp->fastopen_req; |
Eric Dumazet | 07e100f | 2015-12-16 13:53:10 -0800 | [diff] [blame] | 3202 | int syn_loss = 0, space, err = 0; |
Yuchung Cheng | aab4874 | 2012-07-19 06:43:10 +0000 | [diff] [blame] | 3203 | unsigned long last_syn_loss = 0; |
Eric Dumazet | 355a901 | 2014-11-17 23:06:20 -0800 | [diff] [blame] | 3204 | struct sk_buff *syn_data; |
Yuchung Cheng | 783237e | 2012-07-19 06:43:07 +0000 | [diff] [blame] | 3205 | |
Yuchung Cheng | 67da22d | 2012-07-19 06:43:11 +0000 | [diff] [blame] | 3206 | tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ |
Yuchung Cheng | aab4874 | 2012-07-19 06:43:10 +0000 | [diff] [blame] | 3207 | tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie, |
| 3208 | &syn_loss, &last_syn_loss); |
| 3209 | /* Recurring FO SYN losses: revert to regular handshake temporarily */ |
| 3210 | if (syn_loss > 1 && |
| 3211 | time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) { |
| 3212 | fo->cookie.len = -1; |
| 3213 | goto fallback; |
| 3214 | } |
| 3215 | |
Yuchung Cheng | 67da22d | 2012-07-19 06:43:11 +0000 | [diff] [blame] | 3216 | if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE) |
| 3217 | fo->cookie.len = -1; |
| 3218 | else if (fo->cookie.len <= 0) |
Yuchung Cheng | 783237e | 2012-07-19 06:43:07 +0000 | [diff] [blame] | 3219 | goto fallback; |
| 3220 | |
| 3221 | /* MSS for SYN-data is based on cached MSS and bounded by PMTU and |
| 3222 | * user-MSS. Reserve maximum option space for middleboxes that add |
| 3223 | * private TCP options. The cost is reduced data space in SYN :( |
| 3224 | */ |
| 3225 | if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp) |
| 3226 | tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; |
Yuchung Cheng | 1b63edd | 2013-02-22 08:59:06 +0000 | [diff] [blame] | 3227 | space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) - |
Yuchung Cheng | 783237e | 2012-07-19 06:43:07 +0000 | [diff] [blame] | 3228 | MAX_TCP_OPTION_SPACE; |
| 3229 | |
Eric Dumazet | f5ddcbb | 2014-02-20 10:09:18 -0800 | [diff] [blame] | 3230 | space = min_t(size_t, space, fo->size); |
| 3231 | |
| 3232 | /* limit to order-0 allocations */ |
| 3233 | space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER)); |
| 3234 | |
Eric Dumazet | eb93447 | 2015-05-19 13:26:55 -0700 | [diff] [blame] | 3235 | syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false); |
Eric Dumazet | 355a901 | 2014-11-17 23:06:20 -0800 | [diff] [blame] | 3236 | if (!syn_data) |
Yuchung Cheng | 783237e | 2012-07-19 06:43:07 +0000 | [diff] [blame] | 3237 | goto fallback; |
Eric Dumazet | 355a901 | 2014-11-17 23:06:20 -0800 | [diff] [blame] | 3238 | syn_data->ip_summed = CHECKSUM_PARTIAL; |
| 3239 | memcpy(syn_data->cb, syn->cb, sizeof(syn->cb)); |
Eric Dumazet | 07e100f | 2015-12-16 13:53:10 -0800 | [diff] [blame] | 3240 | if (space) { |
| 3241 | int copied = copy_from_iter(skb_put(syn_data, space), space, |
| 3242 | &fo->data->msg_iter); |
| 3243 | if (unlikely(!copied)) { |
| 3244 | kfree_skb(syn_data); |
| 3245 | goto fallback; |
| 3246 | } |
| 3247 | if (copied != space) { |
| 3248 | skb_trim(syn_data, copied); |
| 3249 | space = copied; |
| 3250 | } |
Yuchung Cheng | 783237e | 2012-07-19 06:43:07 +0000 | [diff] [blame] | 3251 | } |
Eric Dumazet | 355a901 | 2014-11-17 23:06:20 -0800 | [diff] [blame] | 3252 | /* No more data pending in inet_wait_for_connect() */ |
| 3253 | if (space == fo->size) |
| 3254 | fo->data = NULL; |
| 3255 | fo->copied = space; |
Yuchung Cheng | 783237e | 2012-07-19 06:43:07 +0000 | [diff] [blame] | 3256 | |
Eric Dumazet | 355a901 | 2014-11-17 23:06:20 -0800 | [diff] [blame] | 3257 | tcp_connect_queue_skb(sk, syn_data); |
| 3258 | |
| 3259 | err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation); |
| 3260 | |
| 3261 | syn->skb_mstamp = syn_data->skb_mstamp; |
| 3262 | |
| 3263 | /* Now full SYN+DATA was cloned and sent (or not), |
| 3264 | * remove the SYN from the original skb (syn_data) |
| 3265 | * we keep in write queue in case of a retransmit, as we |
| 3266 | * also have the SYN packet (with no data) in the same queue. |
Eric Dumazet | 431a912 | 2014-03-09 17:36:02 -0700 | [diff] [blame] | 3267 | */ |
Eric Dumazet | 355a901 | 2014-11-17 23:06:20 -0800 | [diff] [blame] | 3268 | TCP_SKB_CB(syn_data)->seq++; |
| 3269 | TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH; |
| 3270 | if (!err) { |
Yuchung Cheng | 67da22d | 2012-07-19 06:43:11 +0000 | [diff] [blame] | 3271 | tp->syn_data = (fo->copied > 0); |
Yuchung Cheng | f19c29e | 2014-03-03 12:31:36 -0800 | [diff] [blame] | 3272 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT); |
Yuchung Cheng | 783237e | 2012-07-19 06:43:07 +0000 | [diff] [blame] | 3273 | goto done; |
| 3274 | } |
Yuchung Cheng | 783237e | 2012-07-19 06:43:07 +0000 | [diff] [blame] | 3275 | |
| 3276 | fallback: |
| 3277 | /* Send a regular SYN with Fast Open cookie request option */ |
| 3278 | if (fo->cookie.len > 0) |
| 3279 | fo->cookie.len = 0; |
| 3280 | err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); |
| 3281 | if (err) |
| 3282 | tp->syn_fastopen = 0; |
Yuchung Cheng | 783237e | 2012-07-19 06:43:07 +0000 | [diff] [blame] | 3283 | done: |
| 3284 | fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ |
| 3285 | return err; |
| 3286 | } |
| 3287 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 3288 | /* Build a SYN and send it off. */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3289 | int tcp_connect(struct sock *sk) |
| 3290 | { |
| 3291 | struct tcp_sock *tp = tcp_sk(sk); |
| 3292 | struct sk_buff *buff; |
Eric Paris | ee58681 | 2010-11-16 11:52:49 +0000 | [diff] [blame] | 3293 | int err; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3294 | |
| 3295 | tcp_connect_init(sk); |
| 3296 | |
Andrey Vagin | 2b91647 | 2012-11-22 01:13:58 +0000 | [diff] [blame] | 3297 | if (unlikely(tp->repair)) { |
| 3298 | tcp_finish_connect(sk, NULL); |
| 3299 | return 0; |
| 3300 | } |
| 3301 | |
Eric Dumazet | eb93447 | 2015-05-19 13:26:55 -0700 | [diff] [blame] | 3302 | buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true); |
Eric Dumazet | 355a901 | 2014-11-17 23:06:20 -0800 | [diff] [blame] | 3303 | if (unlikely(!buff)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3304 | return -ENOBUFS; |
| 3305 | |
Changli Gao | a3433f3 | 2010-06-12 14:01:43 +0000 | [diff] [blame] | 3306 | tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); |
Eric Dumazet | 7faee5c | 2014-09-05 15:33:33 -0700 | [diff] [blame] | 3307 | tp->retrans_stamp = tcp_time_stamp; |
Yuchung Cheng | 783237e | 2012-07-19 06:43:07 +0000 | [diff] [blame] | 3308 | tcp_connect_queue_skb(sk, buff); |
Florian Westphal | 735d383 | 2014-09-29 13:08:30 +0200 | [diff] [blame] | 3309 | tcp_ecn_send_syn(sk, buff); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3310 | |
Yuchung Cheng | 783237e | 2012-07-19 06:43:07 +0000 | [diff] [blame] | 3311 | /* Send off SYN; include data in Fast Open. */ |
| 3312 | err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : |
| 3313 | tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); |
Eric Paris | ee58681 | 2010-11-16 11:52:49 +0000 | [diff] [blame] | 3314 | if (err == -ECONNREFUSED) |
| 3315 | return err; |
Wei Yongjun | bd37a08 | 2006-08-07 21:04:15 -0700 | [diff] [blame] | 3316 | |
| 3317 | /* We change tp->snd_nxt after the tcp_transmit_skb() call |
| 3318 | * in order to make this packet get counted in tcpOutSegs. |
| 3319 | */ |
| 3320 | tp->snd_nxt = tp->write_seq; |
| 3321 | tp->pushed_seq = tp->write_seq; |
Pavel Emelyanov | 81cc8a7 | 2008-07-16 20:22:04 -0700 | [diff] [blame] | 3322 | TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3323 | |
| 3324 | /* Timer for repeating the SYN until an answer. */ |
Arnaldo Carvalho de Melo | 3f421ba | 2005-08-09 20:11:08 -0700 | [diff] [blame] | 3325 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, |
| 3326 | inet_csk(sk)->icsk_rto, TCP_RTO_MAX); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3327 | return 0; |
| 3328 | } |
Eric Dumazet | 4bc2f18 | 2010-07-09 21:22:10 +0000 | [diff] [blame] | 3329 | EXPORT_SYMBOL(tcp_connect); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3330 | |
| 3331 | /* Send out a delayed ack, the caller does the policy checking |
| 3332 | * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() |
| 3333 | * for details. |
| 3334 | */ |
| 3335 | void tcp_send_delayed_ack(struct sock *sk) |
| 3336 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 3337 | struct inet_connection_sock *icsk = inet_csk(sk); |
| 3338 | int ato = icsk->icsk_ack.ato; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3339 | unsigned long timeout; |
| 3340 | |
Florian Westphal | 9890092 | 2014-09-26 22:37:35 +0200 | [diff] [blame] | 3341 | tcp_ca_event(sk, CA_EVENT_DELAYED_ACK); |
| 3342 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3343 | if (ato > TCP_DELACK_MIN) { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 3344 | const struct tcp_sock *tp = tcp_sk(sk); |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 3345 | int max_ato = HZ / 2; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3346 | |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 3347 | if (icsk->icsk_ack.pingpong || |
| 3348 | (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3349 | max_ato = TCP_DELACK_MAX; |
| 3350 | |
| 3351 | /* Slow path, intersegment interval is "high". */ |
| 3352 | |
| 3353 | /* If some rtt estimate is known, use it to bound delayed ack. |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 3354 | * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3355 | * directly. |
| 3356 | */ |
Eric Dumazet | 740b0f1 | 2014-02-26 14:02:48 -0800 | [diff] [blame] | 3357 | if (tp->srtt_us) { |
| 3358 | int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3), |
| 3359 | TCP_DELACK_MIN); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3360 | |
| 3361 | if (rtt < max_ato) |
| 3362 | max_ato = rtt; |
| 3363 | } |
| 3364 | |
| 3365 | ato = min(ato, max_ato); |
| 3366 | } |
| 3367 | |
| 3368 | /* Stay within the limit we were given */ |
| 3369 | timeout = jiffies + ato; |
| 3370 | |
| 3371 | /* Use new timeout only if there wasn't a older one earlier. */ |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 3372 | if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3373 | /* If delack timer was blocked or is about to expire, |
| 3374 | * send ACK now. |
| 3375 | */ |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 3376 | if (icsk->icsk_ack.blocked || |
| 3377 | time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3378 | tcp_send_ack(sk); |
| 3379 | return; |
| 3380 | } |
| 3381 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 3382 | if (!time_before(timeout, icsk->icsk_ack.timeout)) |
| 3383 | timeout = icsk->icsk_ack.timeout; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3384 | } |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 3385 | icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; |
| 3386 | icsk->icsk_ack.timeout = timeout; |
| 3387 | sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3388 | } |
| 3389 | |
| 3390 | /* This routine sends an ack and also updates the window. */ |
| 3391 | void tcp_send_ack(struct sock *sk) |
| 3392 | { |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 3393 | struct sk_buff *buff; |
| 3394 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3395 | /* If we have been reset, we may not send again. */ |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 3396 | if (sk->sk_state == TCP_CLOSE) |
| 3397 | return; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3398 | |
Florian Westphal | 9890092 | 2014-09-26 22:37:35 +0200 | [diff] [blame] | 3399 | tcp_ca_event(sk, CA_EVENT_NON_DELAYED_ACK); |
| 3400 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 3401 | /* We are not putting this on the write queue, so |
| 3402 | * tcp_transmit_skb() will set the ownership to this |
| 3403 | * sock. |
| 3404 | */ |
Eric Dumazet | 7450aaf | 2015-11-30 08:57:28 -0800 | [diff] [blame] | 3405 | buff = alloc_skb(MAX_TCP_HEADER, |
| 3406 | sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); |
| 3407 | if (unlikely(!buff)) { |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 3408 | inet_csk_schedule_ack(sk); |
| 3409 | inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; |
| 3410 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, |
| 3411 | TCP_DELACK_MAX, TCP_RTO_MAX); |
| 3412 | return; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3413 | } |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 3414 | |
| 3415 | /* Reserve space for headers and prepare control bits. */ |
| 3416 | skb_reserve(buff, MAX_TCP_HEADER); |
Changli Gao | a3433f3 | 2010-06-12 14:01:43 +0000 | [diff] [blame] | 3417 | tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 3418 | |
Eric Dumazet | 9878196 | 2015-02-03 18:31:53 -0800 | [diff] [blame] | 3419 | /* We do not want pure acks influencing TCP Small Queues or fq/pacing |
| 3420 | * too much. |
| 3421 | * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784 |
| 3422 | * We also avoid tcp_wfree() overhead (cache line miss accessing |
| 3423 | * tp->tsq_flags) by using regular sock_wfree() |
| 3424 | */ |
| 3425 | skb_set_tcp_pure_ack(buff); |
| 3426 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 3427 | /* Send it off, this clears delayed acks for us. */ |
Eric Dumazet | 7faee5c | 2014-09-05 15:33:33 -0700 | [diff] [blame] | 3428 | skb_mstamp_get(&buff->skb_mstamp); |
Eric Dumazet | 7450aaf | 2015-11-30 08:57:28 -0800 | [diff] [blame] | 3429 | tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3430 | } |
Daniel Borkmann | e3118e8 | 2014-09-26 22:37:36 +0200 | [diff] [blame] | 3431 | EXPORT_SYMBOL_GPL(tcp_send_ack); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3432 | |
| 3433 | /* This routine sends a packet with an out of date sequence |
| 3434 | * number. It assumes the other end will try to ack it. |
| 3435 | * |
| 3436 | * Question: what should we make while urgent mode? |
| 3437 | * 4.4BSD forces sending single byte of data. We cannot send |
| 3438 | * out of window data, because we have SND.NXT==SND.MAX... |
| 3439 | * |
| 3440 | * Current solution: to send TWO zero-length segments in urgent mode: |
| 3441 | * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is |
| 3442 | * out-of-date with SND.UNA-1 to probe window. |
| 3443 | */ |
Eric Dumazet | e520af4 | 2015-05-06 14:26:25 -0700 | [diff] [blame] | 3444 | static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3445 | { |
| 3446 | struct tcp_sock *tp = tcp_sk(sk); |
| 3447 | struct sk_buff *skb; |
| 3448 | |
| 3449 | /* We don't queue it, tcp_transmit_skb() sets ownership. */ |
Eric Dumazet | 7450aaf | 2015-11-30 08:57:28 -0800 | [diff] [blame] | 3450 | skb = alloc_skb(MAX_TCP_HEADER, |
| 3451 | sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); |
Ian Morris | 51456b2 | 2015-04-03 09:17:26 +0100 | [diff] [blame] | 3452 | if (!skb) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3453 | return -1; |
| 3454 | |
| 3455 | /* Reserve space for headers and set control bits. */ |
| 3456 | skb_reserve(skb, MAX_TCP_HEADER); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3457 | /* Use a previous sequence. This should cause the other |
| 3458 | * end to send an ack. Don't queue or clone SKB, just |
| 3459 | * send it. |
| 3460 | */ |
Changli Gao | a3433f3 | 2010-06-12 14:01:43 +0000 | [diff] [blame] | 3461 | tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); |
Eric Dumazet | 7faee5c | 2014-09-05 15:33:33 -0700 | [diff] [blame] | 3462 | skb_mstamp_get(&skb->skb_mstamp); |
Renato Westphal | e2e8009 | 2015-10-19 18:51:34 -0200 | [diff] [blame] | 3463 | NET_INC_STATS(sock_net(sk), mib); |
Eric Dumazet | 7450aaf | 2015-11-30 08:57:28 -0800 | [diff] [blame] | 3464 | return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3465 | } |
| 3466 | |
Pavel Emelyanov | ee99528 | 2012-04-19 03:40:39 +0000 | [diff] [blame] | 3467 | void tcp_send_window_probe(struct sock *sk) |
| 3468 | { |
| 3469 | if (sk->sk_state == TCP_ESTABLISHED) { |
| 3470 | tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; |
Eric Dumazet | e520af4 | 2015-05-06 14:26:25 -0700 | [diff] [blame] | 3471 | tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE); |
Pavel Emelyanov | ee99528 | 2012-04-19 03:40:39 +0000 | [diff] [blame] | 3472 | } |
| 3473 | } |
| 3474 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 3475 | /* Initiate keepalive or window probe from timer. */ |
Eric Dumazet | e520af4 | 2015-05-06 14:26:25 -0700 | [diff] [blame] | 3476 | int tcp_write_wakeup(struct sock *sk, int mib) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3477 | { |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 3478 | struct tcp_sock *tp = tcp_sk(sk); |
| 3479 | struct sk_buff *skb; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3480 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 3481 | if (sk->sk_state == TCP_CLOSE) |
| 3482 | return -1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3483 | |
Ian Morris | 00db412 | 2015-04-03 09:17:27 +0100 | [diff] [blame] | 3484 | skb = tcp_send_head(sk); |
| 3485 | if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 3486 | int err; |
Ilpo Järvinen | 0c54b85 | 2009-03-14 14:23:05 +0000 | [diff] [blame] | 3487 | unsigned int mss = tcp_current_mss(sk); |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 3488 | unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3489 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 3490 | if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) |
| 3491 | tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3492 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 3493 | /* We are probing the opening of a window |
| 3494 | * but the window size is != 0 |
| 3495 | * must have been a result SWS avoidance ( sender ) |
| 3496 | */ |
| 3497 | if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || |
| 3498 | skb->len > mss) { |
| 3499 | seg_size = min(seg_size, mss); |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 3500 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; |
Octavian Purdila | 6cc55e0 | 2014-06-06 17:32:37 +0300 | [diff] [blame] | 3501 | if (tcp_fragment(sk, skb, seg_size, mss, GFP_ATOMIC)) |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 3502 | return -1; |
| 3503 | } else if (!tcp_skb_pcount(skb)) |
Eric Dumazet | 5bbb432 | 2015-06-11 09:15:17 -0700 | [diff] [blame] | 3504 | tcp_set_skb_tso_segs(skb, mss); |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 3505 | |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 3506 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 3507 | err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); |
| 3508 | if (!err) |
| 3509 | tcp_event_new_data_sent(sk, skb); |
| 3510 | return err; |
| 3511 | } else { |
Ilpo Järvinen | 33f5f57 | 2008-10-07 14:43:06 -0700 | [diff] [blame] | 3512 | if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) |
Eric Dumazet | e520af4 | 2015-05-06 14:26:25 -0700 | [diff] [blame] | 3513 | tcp_xmit_probe_skb(sk, 1, mib); |
| 3514 | return tcp_xmit_probe_skb(sk, 0, mib); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3515 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3516 | } |
| 3517 | |
| 3518 | /* A window probe timeout has occurred. If window is not closed send |
| 3519 | * a partial packet else a zero probe. |
| 3520 | */ |
| 3521 | void tcp_send_probe0(struct sock *sk) |
| 3522 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 3523 | struct inet_connection_sock *icsk = inet_csk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3524 | struct tcp_sock *tp = tcp_sk(sk); |
Nikolay Borisov | c6214a9 | 2016-02-03 09:46:54 +0200 | [diff] [blame] | 3525 | struct net *net = sock_net(sk); |
Eric Dumazet | fcdd1cf | 2014-09-22 13:19:44 -0700 | [diff] [blame] | 3526 | unsigned long probe_max; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3527 | int err; |
| 3528 | |
Eric Dumazet | e520af4 | 2015-05-06 14:26:25 -0700 | [diff] [blame] | 3529 | err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3530 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 3531 | if (tp->packets_out || !tcp_send_head(sk)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3532 | /* Cancel probe timer, if it is not required. */ |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 3533 | icsk->icsk_probes_out = 0; |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 3534 | icsk->icsk_backoff = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3535 | return; |
| 3536 | } |
| 3537 | |
| 3538 | if (err <= 0) { |
Nikolay Borisov | c6214a9 | 2016-02-03 09:46:54 +0200 | [diff] [blame] | 3539 | if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2) |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 3540 | icsk->icsk_backoff++; |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 3541 | icsk->icsk_probes_out++; |
Eric Dumazet | fcdd1cf | 2014-09-22 13:19:44 -0700 | [diff] [blame] | 3542 | probe_max = TCP_RTO_MAX; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3543 | } else { |
| 3544 | /* If packet was not sent due to local congestion, |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 3545 | * do not backoff and do not remember icsk_probes_out. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3546 | * Let local senders to fight for local resources. |
| 3547 | * |
| 3548 | * Use accumulated backoff yet. |
| 3549 | */ |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 3550 | if (!icsk->icsk_probes_out) |
| 3551 | icsk->icsk_probes_out = 1; |
Eric Dumazet | fcdd1cf | 2014-09-22 13:19:44 -0700 | [diff] [blame] | 3552 | probe_max = TCP_RESOURCE_PROBE_INTERVAL; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3553 | } |
Eric Dumazet | fcdd1cf | 2014-09-22 13:19:44 -0700 | [diff] [blame] | 3554 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, |
Eric Dumazet | 21c8fe9 | 2015-05-06 14:26:24 -0700 | [diff] [blame] | 3555 | tcp_probe0_when(sk, probe_max), |
Eric Dumazet | fcdd1cf | 2014-09-22 13:19:44 -0700 | [diff] [blame] | 3556 | TCP_RTO_MAX); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3557 | } |
Octavian Purdila | 5db92c9 | 2014-06-25 17:09:59 +0300 | [diff] [blame] | 3558 | |
Eric Dumazet | ea3bea3 | 2015-09-25 07:39:23 -0700 | [diff] [blame] | 3559 | int tcp_rtx_synack(const struct sock *sk, struct request_sock *req) |
Octavian Purdila | 5db92c9 | 2014-06-25 17:09:59 +0300 | [diff] [blame] | 3560 | { |
| 3561 | const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific; |
| 3562 | struct flowi fl; |
| 3563 | int res; |
| 3564 | |
Eric Dumazet | 58d607d | 2015-09-15 15:24:20 -0700 | [diff] [blame] | 3565 | tcp_rsk(req)->txhash = net_tx_rndhash(); |
Eric Dumazet | b3d0514 | 2016-04-13 22:05:39 -0700 | [diff] [blame] | 3566 | res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL); |
Octavian Purdila | 5db92c9 | 2014-06-25 17:09:59 +0300 | [diff] [blame] | 3567 | if (!res) { |
Eric Dumazet | 90bbcc6 | 2016-04-27 16:44:32 -0700 | [diff] [blame] | 3568 | __TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); |
Eric Dumazet | 02a1d6e | 2016-04-27 16:44:39 -0700 | [diff] [blame] | 3569 | __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); |
Octavian Purdila | 5db92c9 | 2014-06-25 17:09:59 +0300 | [diff] [blame] | 3570 | } |
| 3571 | return res; |
| 3572 | } |
| 3573 | EXPORT_SYMBOL(tcp_rtx_synack); |