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 | |
Eric Dumazet | 46d3cea | 2012-07-11 05:50:31 +0000 | [diff] [blame] | 53 | /* Default TSQ limit of two TSO segments */ |
| 54 | int sysctl_tcp_limit_output_bytes __read_mostly = 131072; |
| 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 | |
Brian Haley | ab32ea5 | 2006-09-22 14:15:41 -0700 | [diff] [blame] | 62 | int sysctl_tcp_mtu_probing __read_mostly = 0; |
Shan Wei | 97b1ce25 | 2010-12-01 18:04:50 +0000 | [diff] [blame] | 63 | int sysctl_tcp_base_mss __read_mostly = TCP_BASE_MSS; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 64 | |
David S. Miller | 35089bb | 2006-06-13 22:33:04 -0700 | [diff] [blame] | 65 | /* By default, RFC2861 behavior. */ |
Brian Haley | ab32ea5 | 2006-09-22 14:15:41 -0700 | [diff] [blame] | 66 | int sysctl_tcp_slow_start_after_idle __read_mostly = 1; |
David S. Miller | 35089bb | 2006-06-13 22:33:04 -0700 | [diff] [blame] | 67 | |
Eric Dumazet | c9bee3b7 | 2013-07-22 20:27:07 -0700 | [diff] [blame] | 68 | unsigned int sysctl_tcp_notsent_lowat __read_mostly = UINT_MAX; |
| 69 | EXPORT_SYMBOL(sysctl_tcp_notsent_lowat); |
| 70 | |
Eric Dumazet | 46d3cea | 2012-07-11 05:50:31 +0000 | [diff] [blame] | 71 | static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, |
| 72 | int push_one, gfp_t gfp); |
William Allen Simpson | 519855c | 2009-12-02 18:14:19 +0000 | [diff] [blame] | 73 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 74 | /* Account for new data that has been sent to the network. */ |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 75 | 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] | 76 | { |
Nandita Dukkipati | 6ba8a3b | 2013-03-11 10:00:43 +0000 | [diff] [blame] | 77 | struct inet_connection_sock *icsk = inet_csk(sk); |
Ilpo Järvinen | 6ff03ac | 2007-08-24 22:44:06 -0700 | [diff] [blame] | 78 | struct tcp_sock *tp = tcp_sk(sk); |
Ilpo Järvinen | 66f5fe6 | 2007-12-31 04:43:57 -0800 | [diff] [blame] | 79 | unsigned int prior_packets = tp->packets_out; |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 80 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 81 | tcp_advance_send_head(sk, skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 82 | tp->snd_nxt = TCP_SKB_CB(skb)->end_seq; |
Ilpo Järvinen | 8512430 | 2007-11-26 20:17:38 +0800 | [diff] [blame] | 83 | |
Ilpo Järvinen | 66f5fe6 | 2007-12-31 04:43:57 -0800 | [diff] [blame] | 84 | tp->packets_out += tcp_skb_pcount(skb); |
Nandita Dukkipati | 6ba8a3b | 2013-03-11 10:00:43 +0000 | [diff] [blame] | 85 | if (!prior_packets || icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS || |
Eric Dumazet | 6a5dc9e | 2013-04-29 08:39:56 +0000 | [diff] [blame] | 86 | icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) { |
Yuchung Cheng | 750ea2b | 2012-05-02 13:30:04 +0000 | [diff] [blame] | 87 | tcp_rearm_rto(sk); |
Eric Dumazet | 6a5dc9e | 2013-04-29 08:39:56 +0000 | [diff] [blame] | 88 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 89 | } |
| 90 | |
| 91 | /* SND.NXT, if window was not shrunk. |
| 92 | * If window has been shrunk, what should we make? It is not clear at all. |
| 93 | * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-( |
| 94 | * Anything in between SND.UNA...SND.UNA+SND.WND also can be already |
| 95 | * invalid. OK, let's make this for now: |
| 96 | */ |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 97 | static inline __u32 tcp_acceptable_seq(const struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 98 | { |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 99 | const struct tcp_sock *tp = tcp_sk(sk); |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 100 | |
Ilpo Järvinen | 90840de | 2007-12-31 04:48:41 -0800 | [diff] [blame] | 101 | if (!before(tcp_wnd_end(tp), tp->snd_nxt)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 102 | return tp->snd_nxt; |
| 103 | else |
Ilpo Järvinen | 90840de | 2007-12-31 04:48:41 -0800 | [diff] [blame] | 104 | return tcp_wnd_end(tp); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 105 | } |
| 106 | |
| 107 | /* Calculate mss to advertise in SYN segment. |
| 108 | * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that: |
| 109 | * |
| 110 | * 1. It is independent of path mtu. |
| 111 | * 2. Ideally, it is maximal possible segment size i.e. 65535-40. |
| 112 | * 3. For IPv4 it is reasonable to calculate it from maximal MTU of |
| 113 | * attached devices, because some buggy hosts are confused by |
| 114 | * large MSS. |
| 115 | * 4. We do not make 3, we advertise MSS, calculated from first |
| 116 | * hop device mtu, but allow to raise it to ip_rt_min_advmss. |
| 117 | * This may be overridden via information stored in routing table. |
| 118 | * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible, |
| 119 | * probably even Jumbo". |
| 120 | */ |
| 121 | static __u16 tcp_advertise_mss(struct sock *sk) |
| 122 | { |
| 123 | struct tcp_sock *tp = tcp_sk(sk); |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 124 | const struct dst_entry *dst = __sk_dst_get(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 125 | int mss = tp->advmss; |
| 126 | |
David S. Miller | 0dbaee3 | 2010-12-13 12:52:14 -0800 | [diff] [blame] | 127 | if (dst) { |
| 128 | unsigned int metric = dst_metric_advmss(dst); |
| 129 | |
| 130 | if (metric < mss) { |
| 131 | mss = metric; |
| 132 | tp->advmss = mss; |
| 133 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 134 | } |
| 135 | |
| 136 | return (__u16)mss; |
| 137 | } |
| 138 | |
| 139 | /* RFC2861. Reset CWND after idle period longer RTO to "restart window". |
| 140 | * This is the first part of cwnd validation mechanism. */ |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 141 | static void tcp_cwnd_restart(struct sock *sk, const struct dst_entry *dst) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 142 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 143 | struct tcp_sock *tp = tcp_sk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 144 | s32 delta = tcp_time_stamp - tp->lsndtime; |
| 145 | u32 restart_cwnd = tcp_init_cwnd(tp, dst); |
| 146 | u32 cwnd = tp->snd_cwnd; |
| 147 | |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 148 | tcp_ca_event(sk, CA_EVENT_CWND_RESTART); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 149 | |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 150 | tp->snd_ssthresh = tcp_current_ssthresh(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 151 | restart_cwnd = min(restart_cwnd, cwnd); |
| 152 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 153 | while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 154 | cwnd >>= 1; |
| 155 | tp->snd_cwnd = max(cwnd, restart_cwnd); |
| 156 | tp->snd_cwnd_stamp = tcp_time_stamp; |
| 157 | tp->snd_cwnd_used = 0; |
| 158 | } |
| 159 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 160 | /* Congestion state accounting after a packet has been sent. */ |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 161 | static void tcp_event_data_sent(struct tcp_sock *tp, |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 162 | struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 163 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 164 | struct inet_connection_sock *icsk = inet_csk(sk); |
| 165 | const u32 now = tcp_time_stamp; |
Cong Wang | bcefe17 | 2013-06-15 09:39:18 +0800 | [diff] [blame] | 166 | const struct dst_entry *dst = __sk_dst_get(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 167 | |
David S. Miller | 35089bb | 2006-06-13 22:33:04 -0700 | [diff] [blame] | 168 | if (sysctl_tcp_slow_start_after_idle && |
| 169 | (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto)) |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 170 | tcp_cwnd_restart(sk, __sk_dst_get(sk)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 171 | |
| 172 | tp->lsndtime = now; |
| 173 | |
| 174 | /* If it is a reply for ato after last received |
| 175 | * packet, enter pingpong mode. |
| 176 | */ |
Cong Wang | bcefe17 | 2013-06-15 09:39:18 +0800 | [diff] [blame] | 177 | if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato && |
| 178 | (!dst || !dst_metric(dst, RTAX_QUICKACK))) |
| 179 | icsk->icsk_ack.pingpong = 1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 180 | } |
| 181 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 182 | /* Account for an ACK we sent. */ |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 183 | 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] | 184 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 185 | tcp_dec_quickack_mode(sk, pkts); |
| 186 | inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 187 | } |
| 188 | |
Yuchung Cheng | 85f1652 | 2013-06-11 15:35:32 -0700 | [diff] [blame] | 189 | |
| 190 | u32 tcp_default_init_rwnd(u32 mss) |
| 191 | { |
| 192 | /* Initial receive window should be twice of TCP_INIT_CWND to |
Weiping Pan | 9ef71e0 | 2013-06-18 21:00:31 +0800 | [diff] [blame] | 193 | * enable proper sending of new unsent data during fast recovery |
Yuchung Cheng | 85f1652 | 2013-06-11 15:35:32 -0700 | [diff] [blame] | 194 | * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a |
| 195 | * limit when mss is larger than 1460. |
| 196 | */ |
| 197 | u32 init_rwnd = TCP_INIT_CWND * 2; |
| 198 | |
| 199 | if (mss > 1460) |
| 200 | init_rwnd = max((1460 * init_rwnd) / mss, 2U); |
| 201 | return init_rwnd; |
| 202 | } |
| 203 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 204 | /* Determine a window scaling and initial window to offer. |
| 205 | * Based on the assumption that the given amount of space |
| 206 | * will be offered. Store the results in the tp structure. |
| 207 | * NOTE: for smooth operation initial space offering should |
| 208 | * be a multiple of mss if possible. We assume here that mss >= 1. |
| 209 | * This MUST be enforced by all callers. |
| 210 | */ |
| 211 | void tcp_select_initial_window(int __space, __u32 mss, |
| 212 | __u32 *rcv_wnd, __u32 *window_clamp, |
laurent chavey | 31d1292 | 2009-12-15 11:15:28 +0000 | [diff] [blame] | 213 | int wscale_ok, __u8 *rcv_wscale, |
| 214 | __u32 init_rcv_wnd) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 215 | { |
| 216 | unsigned int space = (__space < 0 ? 0 : __space); |
| 217 | |
| 218 | /* If no clamp set the clamp to the max possible scaled window */ |
| 219 | if (*window_clamp == 0) |
| 220 | (*window_clamp) = (65535 << 14); |
| 221 | space = min(*window_clamp, space); |
| 222 | |
| 223 | /* Quantize space offering to a multiple of mss if possible. */ |
| 224 | if (space > mss) |
| 225 | space = (space / mss) * mss; |
| 226 | |
| 227 | /* NOTE: offering an initial window larger than 32767 |
Rick Jones | 15d99e0 | 2006-03-20 22:40:29 -0800 | [diff] [blame] | 228 | * will break some buggy TCP stacks. If the admin tells us |
| 229 | * it is likely we could be speaking with such a buggy stack |
| 230 | * we will truncate our initial window offering to 32K-1 |
| 231 | * unless the remote has sent us a window scaling option, |
| 232 | * which we interpret as a sign the remote TCP is not |
| 233 | * misinterpreting the window field as a signed quantity. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 234 | */ |
Rick Jones | 15d99e0 | 2006-03-20 22:40:29 -0800 | [diff] [blame] | 235 | if (sysctl_tcp_workaround_signed_windows) |
| 236 | (*rcv_wnd) = min(space, MAX_TCP_WINDOW); |
| 237 | else |
| 238 | (*rcv_wnd) = space; |
| 239 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 240 | (*rcv_wscale) = 0; |
| 241 | if (wscale_ok) { |
| 242 | /* Set window scaling on max possible window |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 243 | * See RFC1323 for an explanation of the limit to 14 |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 244 | */ |
| 245 | space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max); |
Stephen Hemminger | 316c159 | 2006-08-22 00:06:11 -0700 | [diff] [blame] | 246 | space = min_t(u32, space, *window_clamp); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 247 | while (space > 65535 && (*rcv_wscale) < 14) { |
| 248 | space >>= 1; |
| 249 | (*rcv_wscale)++; |
| 250 | } |
| 251 | } |
| 252 | |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 253 | if (mss > (1 << *rcv_wscale)) { |
Yuchung Cheng | 85f1652 | 2013-06-11 15:35:32 -0700 | [diff] [blame] | 254 | if (!init_rcv_wnd) /* Use default unless specified otherwise */ |
| 255 | init_rcv_wnd = tcp_default_init_rwnd(mss); |
| 256 | *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 257 | } |
| 258 | |
| 259 | /* Set the clamp no higher than max representable value */ |
| 260 | (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp); |
| 261 | } |
Eric Dumazet | 4bc2f18 | 2010-07-09 21:22:10 +0000 | [diff] [blame] | 262 | EXPORT_SYMBOL(tcp_select_initial_window); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 263 | |
| 264 | /* Chose a new window to advertise, update state in tcp_sock for the |
| 265 | * socket, and return result with RFC1323 scaling applied. The return |
| 266 | * value can be stuffed directly into th->window for an outgoing |
| 267 | * frame. |
| 268 | */ |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 269 | static u16 tcp_select_window(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 270 | { |
| 271 | struct tcp_sock *tp = tcp_sk(sk); |
| 272 | u32 cur_win = tcp_receive_window(tp); |
| 273 | u32 new_win = __tcp_select_window(sk); |
| 274 | |
| 275 | /* Never shrink the offered window */ |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 276 | if (new_win < cur_win) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 277 | /* Danger Will Robinson! |
| 278 | * Don't update rcv_wup/rcv_wnd here or else |
| 279 | * we will not be able to advertise a zero |
| 280 | * window in time. --DaveM |
| 281 | * |
| 282 | * Relax Will Robinson. |
| 283 | */ |
Patrick McHardy | 607bfbf | 2008-03-20 16:11:27 -0700 | [diff] [blame] | 284 | new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 285 | } |
| 286 | tp->rcv_wnd = new_win; |
| 287 | tp->rcv_wup = tp->rcv_nxt; |
| 288 | |
| 289 | /* Make sure we do not exceed the maximum possible |
| 290 | * scaled window. |
| 291 | */ |
Rick Jones | 15d99e0 | 2006-03-20 22:40:29 -0800 | [diff] [blame] | 292 | if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 293 | new_win = min(new_win, MAX_TCP_WINDOW); |
| 294 | else |
| 295 | new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); |
| 296 | |
| 297 | /* RFC1323 scaling applied */ |
| 298 | new_win >>= tp->rx_opt.rcv_wscale; |
| 299 | |
| 300 | /* If we advertise zero window, disable fast path. */ |
| 301 | if (new_win == 0) |
| 302 | tp->pred_flags = 0; |
| 303 | |
| 304 | return new_win; |
| 305 | } |
| 306 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 307 | /* Packet ECN state for a SYN-ACK */ |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 308 | static inline void TCP_ECN_send_synack(const struct tcp_sock *tp, struct sk_buff *skb) |
Ilpo Järvinen | bdf1ee5 | 2007-05-27 02:04:16 -0700 | [diff] [blame] | 309 | { |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 310 | TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR; |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 311 | if (!(tp->ecn_flags & TCP_ECN_OK)) |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 312 | TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE; |
Ilpo Järvinen | bdf1ee5 | 2007-05-27 02:04:16 -0700 | [diff] [blame] | 313 | } |
| 314 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 315 | /* Packet ECN state for a SYN. */ |
Ilpo Järvinen | bdf1ee5 | 2007-05-27 02:04:16 -0700 | [diff] [blame] | 316 | static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb) |
| 317 | { |
| 318 | struct tcp_sock *tp = tcp_sk(sk); |
| 319 | |
| 320 | tp->ecn_flags = 0; |
Hannes Frederic Sowa | 5d134f1 | 2013-01-05 16:10:48 +0000 | [diff] [blame] | 321 | if (sock_net(sk)->ipv4.sysctl_tcp_ecn == 1) { |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 322 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR; |
Ilpo Järvinen | bdf1ee5 | 2007-05-27 02:04:16 -0700 | [diff] [blame] | 323 | tp->ecn_flags = TCP_ECN_OK; |
| 324 | } |
| 325 | } |
| 326 | |
| 327 | static __inline__ void |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 328 | TCP_ECN_make_synack(const struct request_sock *req, struct tcphdr *th) |
Ilpo Järvinen | bdf1ee5 | 2007-05-27 02:04:16 -0700 | [diff] [blame] | 329 | { |
| 330 | if (inet_rsk(req)->ecn_ok) |
| 331 | th->ece = 1; |
| 332 | } |
| 333 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 334 | /* Set up ECN state for a packet on a ESTABLISHED socket that is about to |
| 335 | * be sent. |
| 336 | */ |
Ilpo Järvinen | bdf1ee5 | 2007-05-27 02:04:16 -0700 | [diff] [blame] | 337 | static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb, |
| 338 | int tcp_header_len) |
| 339 | { |
| 340 | struct tcp_sock *tp = tcp_sk(sk); |
| 341 | |
| 342 | if (tp->ecn_flags & TCP_ECN_OK) { |
| 343 | /* Not-retransmitted data segment: set ECT and inject CWR. */ |
| 344 | if (skb->len != tcp_header_len && |
| 345 | !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { |
| 346 | INET_ECN_xmit(sk); |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 347 | if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) { |
Ilpo Järvinen | bdf1ee5 | 2007-05-27 02:04:16 -0700 | [diff] [blame] | 348 | tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; |
| 349 | tcp_hdr(skb)->cwr = 1; |
| 350 | skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; |
| 351 | } |
| 352 | } else { |
| 353 | /* ACK or retransmitted segment: clear ECT|CE */ |
| 354 | INET_ECN_dontxmit(sk); |
| 355 | } |
| 356 | if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) |
| 357 | tcp_hdr(skb)->ece = 1; |
| 358 | } |
| 359 | } |
| 360 | |
Ilpo Järvinen | e870a8e | 2008-01-03 20:39:01 -0800 | [diff] [blame] | 361 | /* Constructs common control bits of non-data skb. If SYN/FIN is present, |
| 362 | * auto increment end seqno. |
| 363 | */ |
| 364 | static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags) |
| 365 | { |
David S. Miller | 2e8e18e | 2010-04-08 11:32:30 -0700 | [diff] [blame] | 366 | skb->ip_summed = CHECKSUM_PARTIAL; |
Ilpo Järvinen | e870a8e | 2008-01-03 20:39:01 -0800 | [diff] [blame] | 367 | skb->csum = 0; |
| 368 | |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 369 | TCP_SKB_CB(skb)->tcp_flags = flags; |
Ilpo Järvinen | e870a8e | 2008-01-03 20:39:01 -0800 | [diff] [blame] | 370 | TCP_SKB_CB(skb)->sacked = 0; |
| 371 | |
| 372 | skb_shinfo(skb)->gso_segs = 1; |
| 373 | skb_shinfo(skb)->gso_size = 0; |
| 374 | skb_shinfo(skb)->gso_type = 0; |
| 375 | |
| 376 | TCP_SKB_CB(skb)->seq = seq; |
Changli Gao | a3433f3 | 2010-06-12 14:01:43 +0000 | [diff] [blame] | 377 | if (flags & (TCPHDR_SYN | TCPHDR_FIN)) |
Ilpo Järvinen | e870a8e | 2008-01-03 20:39:01 -0800 | [diff] [blame] | 378 | seq++; |
| 379 | TCP_SKB_CB(skb)->end_seq = seq; |
| 380 | } |
| 381 | |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 382 | static inline bool tcp_urg_mode(const struct tcp_sock *tp) |
Ilpo Järvinen | 33f5f57 | 2008-10-07 14:43:06 -0700 | [diff] [blame] | 383 | { |
| 384 | return tp->snd_una != tp->snd_up; |
| 385 | } |
| 386 | |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 387 | #define OPTION_SACK_ADVERTISE (1 << 0) |
| 388 | #define OPTION_TS (1 << 1) |
| 389 | #define OPTION_MD5 (1 << 2) |
Ori Finkelman | 89e95a6 | 2009-10-01 06:41:59 +0000 | [diff] [blame] | 390 | #define OPTION_WSCALE (1 << 3) |
Yuchung Cheng | 2100c8d | 2012-07-19 06:43:05 +0000 | [diff] [blame] | 391 | #define OPTION_FAST_OPEN_COOKIE (1 << 8) |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 392 | |
| 393 | struct tcp_out_options { |
Yuchung Cheng | 2100c8d | 2012-07-19 06:43:05 +0000 | [diff] [blame] | 394 | u16 options; /* bit field of OPTION_* */ |
| 395 | u16 mss; /* 0 to disable */ |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 396 | u8 ws; /* window scale, 0 to disable */ |
| 397 | u8 num_sack_blocks; /* number of SACK blocks to include */ |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 398 | u8 hash_size; /* bytes in hash_location */ |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 399 | __u8 *hash_location; /* temporary pointer, overloaded */ |
Yuchung Cheng | 2100c8d | 2012-07-19 06:43:05 +0000 | [diff] [blame] | 400 | __u32 tsval, tsecr; /* need to include OPTION_TS */ |
| 401 | struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */ |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 402 | }; |
| 403 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 404 | /* Write previously computed TCP options to the packet. |
| 405 | * |
| 406 | * Beware: Something in the Internet is very sensitive to the ordering of |
Ilpo Järvinen | fd6149d | 2008-10-23 14:06:35 -0700 | [diff] [blame] | 407 | * TCP options, we learned this through the hard way, so be careful here. |
| 408 | * Luckily we can at least blame others for their non-compliance but from |
| 409 | * inter-operatibility perspective it seems that we're somewhat stuck with |
| 410 | * the ordering which we have been using if we want to keep working with |
| 411 | * those broken things (not that it currently hurts anybody as there isn't |
| 412 | * particular reason why the ordering would need to be changed). |
| 413 | * |
| 414 | * At least SACK_PERM as the first option is known to lead to a disaster |
| 415 | * (but it may well be that other scenarios fail similarly). |
| 416 | */ |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 417 | static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp, |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 418 | struct tcp_out_options *opts) |
| 419 | { |
Yuchung Cheng | 2100c8d | 2012-07-19 06:43:05 +0000 | [diff] [blame] | 420 | u16 options = opts->options; /* mungable copy */ |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 421 | |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 422 | if (unlikely(OPTION_MD5 & options)) { |
Christoph Paasch | 1a2c618 | 2013-03-17 08:23:34 +0000 | [diff] [blame] | 423 | *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | |
| 424 | (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG); |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 425 | /* overload cookie hash location */ |
| 426 | opts->hash_location = (__u8 *)ptr; |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 427 | ptr += 4; |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 428 | } |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 429 | |
Ilpo Järvinen | fd6149d | 2008-10-23 14:06:35 -0700 | [diff] [blame] | 430 | if (unlikely(opts->mss)) { |
| 431 | *ptr++ = htonl((TCPOPT_MSS << 24) | |
| 432 | (TCPOLEN_MSS << 16) | |
| 433 | opts->mss); |
| 434 | } |
| 435 | |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 436 | if (likely(OPTION_TS & options)) { |
| 437 | if (unlikely(OPTION_SACK_ADVERTISE & options)) { |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 438 | *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | |
| 439 | (TCPOLEN_SACK_PERM << 16) | |
| 440 | (TCPOPT_TIMESTAMP << 8) | |
| 441 | TCPOLEN_TIMESTAMP); |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 442 | options &= ~OPTION_SACK_ADVERTISE; |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 443 | } else { |
| 444 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
| 445 | (TCPOPT_NOP << 16) | |
| 446 | (TCPOPT_TIMESTAMP << 8) | |
| 447 | TCPOLEN_TIMESTAMP); |
| 448 | } |
| 449 | *ptr++ = htonl(opts->tsval); |
| 450 | *ptr++ = htonl(opts->tsecr); |
| 451 | } |
| 452 | |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 453 | if (unlikely(OPTION_SACK_ADVERTISE & options)) { |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 454 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
| 455 | (TCPOPT_NOP << 16) | |
| 456 | (TCPOPT_SACK_PERM << 8) | |
| 457 | TCPOLEN_SACK_PERM); |
| 458 | } |
| 459 | |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 460 | if (unlikely(OPTION_WSCALE & options)) { |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 461 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
| 462 | (TCPOPT_WINDOW << 16) | |
| 463 | (TCPOLEN_WINDOW << 8) | |
| 464 | opts->ws); |
| 465 | } |
| 466 | |
| 467 | if (unlikely(opts->num_sack_blocks)) { |
| 468 | struct tcp_sack_block *sp = tp->rx_opt.dsack ? |
| 469 | tp->duplicate_sack : tp->selective_acks; |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 470 | int this_sack; |
| 471 | |
| 472 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
| 473 | (TCPOPT_NOP << 16) | |
| 474 | (TCPOPT_SACK << 8) | |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 475 | (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 476 | TCPOLEN_SACK_PERBLOCK))); |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 477 | |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 478 | for (this_sack = 0; this_sack < opts->num_sack_blocks; |
| 479 | ++this_sack) { |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 480 | *ptr++ = htonl(sp[this_sack].start_seq); |
| 481 | *ptr++ = htonl(sp[this_sack].end_seq); |
| 482 | } |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 483 | |
Ilpo Järvinen | 5861f8e | 2009-03-14 14:23:01 +0000 | [diff] [blame] | 484 | tp->rx_opt.dsack = 0; |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 485 | } |
Yuchung Cheng | 2100c8d | 2012-07-19 06:43:05 +0000 | [diff] [blame] | 486 | |
| 487 | if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) { |
| 488 | struct tcp_fastopen_cookie *foc = opts->fastopen_cookie; |
| 489 | |
| 490 | *ptr++ = htonl((TCPOPT_EXP << 24) | |
| 491 | ((TCPOLEN_EXP_FASTOPEN_BASE + foc->len) << 16) | |
| 492 | TCPOPT_FASTOPEN_MAGIC); |
| 493 | |
| 494 | memcpy(ptr, foc->val, foc->len); |
| 495 | if ((foc->len & 3) == 2) { |
| 496 | u8 *align = ((u8 *)ptr) + foc->len; |
| 497 | align[0] = align[1] = TCPOPT_NOP; |
| 498 | } |
| 499 | ptr += (foc->len + 3) >> 2; |
| 500 | } |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 501 | } |
| 502 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 503 | /* Compute TCP options for SYN packets. This is not the final |
| 504 | * network wire format yet. |
| 505 | */ |
Eric Dumazet | 95c9617 | 2012-04-15 05:58:06 +0000 | [diff] [blame] | 506 | 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] | 507 | struct tcp_out_options *opts, |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 508 | struct tcp_md5sig_key **md5) |
| 509 | { |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 510 | struct tcp_sock *tp = tcp_sk(sk); |
Eric Dumazet | 95c9617 | 2012-04-15 05:58:06 +0000 | [diff] [blame] | 511 | unsigned int remaining = MAX_TCP_OPTION_SPACE; |
Yuchung Cheng | 783237e | 2012-07-19 06:43:07 +0000 | [diff] [blame] | 512 | struct tcp_fastopen_request *fastopen = tp->fastopen_req; |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 513 | |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 514 | #ifdef CONFIG_TCP_MD5SIG |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 515 | *md5 = tp->af_specific->md5_lookup(sk, sk); |
| 516 | if (*md5) { |
| 517 | opts->options |= OPTION_MD5; |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 518 | remaining -= TCPOLEN_MD5SIG_ALIGNED; |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 519 | } |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 520 | #else |
| 521 | *md5 = NULL; |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 522 | #endif |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 523 | |
| 524 | /* We always get an MSS option. The option bytes which will be seen in |
| 525 | * normal data packets should timestamps be used, must be in the MSS |
| 526 | * advertised. But we subtract them from tp->mss_cache so that |
| 527 | * calculations in tcp_sendmsg are simpler etc. So account for this |
| 528 | * fact here if necessary. If we don't do this correctly, as a |
| 529 | * receiver we won't recognize data packets as being full sized when we |
| 530 | * should, and thus we won't abide by the delayed ACK rules correctly. |
| 531 | * SACKs don't matter, we never delay an ACK when we have any of those |
| 532 | * going out. */ |
| 533 | opts->mss = tcp_advertise_mss(sk); |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 534 | remaining -= TCPOLEN_MSS_ALIGNED; |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 535 | |
David S. Miller | bb5b7c1 | 2009-12-15 20:56:42 -0800 | [diff] [blame] | 536 | if (likely(sysctl_tcp_timestamps && *md5 == NULL)) { |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 537 | opts->options |= OPTION_TS; |
Andrey Vagin | ee684b6 | 2013-02-11 05:50:19 +0000 | [diff] [blame] | 538 | opts->tsval = TCP_SKB_CB(skb)->when + tp->tsoffset; |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 539 | opts->tsecr = tp->rx_opt.ts_recent; |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 540 | remaining -= TCPOLEN_TSTAMP_ALIGNED; |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 541 | } |
David S. Miller | bb5b7c1 | 2009-12-15 20:56:42 -0800 | [diff] [blame] | 542 | if (likely(sysctl_tcp_window_scaling)) { |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 543 | opts->ws = tp->rx_opt.rcv_wscale; |
Ori Finkelman | 89e95a6 | 2009-10-01 06:41:59 +0000 | [diff] [blame] | 544 | opts->options |= OPTION_WSCALE; |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 545 | remaining -= TCPOLEN_WSCALE_ALIGNED; |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 546 | } |
David S. Miller | bb5b7c1 | 2009-12-15 20:56:42 -0800 | [diff] [blame] | 547 | if (likely(sysctl_tcp_sack)) { |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 548 | opts->options |= OPTION_SACK_ADVERTISE; |
David S. Miller | b32d131 | 2008-07-21 18:45:34 -0700 | [diff] [blame] | 549 | if (unlikely(!(OPTION_TS & opts->options))) |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 550 | remaining -= TCPOLEN_SACKPERM_ALIGNED; |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 551 | } |
| 552 | |
Yuchung Cheng | 783237e | 2012-07-19 06:43:07 +0000 | [diff] [blame] | 553 | if (fastopen && fastopen->cookie.len >= 0) { |
| 554 | u32 need = TCPOLEN_EXP_FASTOPEN_BASE + fastopen->cookie.len; |
| 555 | need = (need + 3) & ~3U; /* Align to 32 bits */ |
| 556 | if (remaining >= need) { |
| 557 | opts->options |= OPTION_FAST_OPEN_COOKIE; |
| 558 | opts->fastopen_cookie = &fastopen->cookie; |
| 559 | remaining -= need; |
| 560 | tp->syn_fastopen = 1; |
| 561 | } |
| 562 | } |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 563 | |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 564 | return MAX_TCP_OPTION_SPACE - remaining; |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 565 | } |
| 566 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 567 | /* Set up TCP options for SYN-ACKs. */ |
Eric Dumazet | 95c9617 | 2012-04-15 05:58:06 +0000 | [diff] [blame] | 568 | static unsigned int tcp_synack_options(struct sock *sk, |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 569 | struct request_sock *req, |
Eric Dumazet | 95c9617 | 2012-04-15 05:58:06 +0000 | [diff] [blame] | 570 | unsigned int mss, struct sk_buff *skb, |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 571 | struct tcp_out_options *opts, |
William Allen Simpson | 4957faade | 2009-12-02 18:25:27 +0000 | [diff] [blame] | 572 | struct tcp_md5sig_key **md5, |
Jerry Chu | 8336886 | 2012-08-31 12:29:12 +0000 | [diff] [blame] | 573 | struct tcp_fastopen_cookie *foc) |
William Allen Simpson | 4957faade | 2009-12-02 18:25:27 +0000 | [diff] [blame] | 574 | { |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 575 | struct inet_request_sock *ireq = inet_rsk(req); |
Eric Dumazet | 95c9617 | 2012-04-15 05:58:06 +0000 | [diff] [blame] | 576 | unsigned int remaining = MAX_TCP_OPTION_SPACE; |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 577 | |
| 578 | #ifdef CONFIG_TCP_MD5SIG |
| 579 | *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req); |
| 580 | if (*md5) { |
| 581 | opts->options |= OPTION_MD5; |
William Allen Simpson | 4957faade | 2009-12-02 18:25:27 +0000 | [diff] [blame] | 582 | remaining -= TCPOLEN_MD5SIG_ALIGNED; |
| 583 | |
| 584 | /* We can't fit any SACK blocks in a packet with MD5 + TS |
| 585 | * options. There was discussion about disabling SACK |
| 586 | * rather than TS in order to fit in better with old, |
| 587 | * buggy kernels, but that was deemed to be unnecessary. |
| 588 | */ |
Eric Dumazet | de213e5e | 2010-05-17 22:35:36 -0700 | [diff] [blame] | 589 | ireq->tstamp_ok &= !ireq->sack_ok; |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 590 | } |
| 591 | #else |
| 592 | *md5 = NULL; |
| 593 | #endif |
| 594 | |
William Allen Simpson | 4957faade | 2009-12-02 18:25:27 +0000 | [diff] [blame] | 595 | /* We always send an MSS option. */ |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 596 | opts->mss = mss; |
William Allen Simpson | 4957faade | 2009-12-02 18:25:27 +0000 | [diff] [blame] | 597 | remaining -= TCPOLEN_MSS_ALIGNED; |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 598 | |
| 599 | if (likely(ireq->wscale_ok)) { |
| 600 | opts->ws = ireq->rcv_wscale; |
Ori Finkelman | 89e95a6 | 2009-10-01 06:41:59 +0000 | [diff] [blame] | 601 | opts->options |= OPTION_WSCALE; |
William Allen Simpson | 4957faade | 2009-12-02 18:25:27 +0000 | [diff] [blame] | 602 | remaining -= TCPOLEN_WSCALE_ALIGNED; |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 603 | } |
Eric Dumazet | de213e5e | 2010-05-17 22:35:36 -0700 | [diff] [blame] | 604 | if (likely(ireq->tstamp_ok)) { |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 605 | opts->options |= OPTION_TS; |
| 606 | opts->tsval = TCP_SKB_CB(skb)->when; |
| 607 | opts->tsecr = req->ts_recent; |
William Allen Simpson | 4957faade | 2009-12-02 18:25:27 +0000 | [diff] [blame] | 608 | remaining -= TCPOLEN_TSTAMP_ALIGNED; |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 609 | } |
| 610 | if (likely(ireq->sack_ok)) { |
| 611 | opts->options |= OPTION_SACK_ADVERTISE; |
Eric Dumazet | de213e5e | 2010-05-17 22:35:36 -0700 | [diff] [blame] | 612 | if (unlikely(!ireq->tstamp_ok)) |
William Allen Simpson | 4957faade | 2009-12-02 18:25:27 +0000 | [diff] [blame] | 613 | remaining -= TCPOLEN_SACKPERM_ALIGNED; |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 614 | } |
Jerry Chu | 8336886 | 2012-08-31 12:29:12 +0000 | [diff] [blame] | 615 | if (foc != NULL) { |
| 616 | u32 need = TCPOLEN_EXP_FASTOPEN_BASE + foc->len; |
| 617 | need = (need + 3) & ~3U; /* Align to 32 bits */ |
| 618 | if (remaining >= need) { |
| 619 | opts->options |= OPTION_FAST_OPEN_COOKIE; |
| 620 | opts->fastopen_cookie = foc; |
| 621 | remaining -= need; |
| 622 | } |
| 623 | } |
William Allen Simpson | 4957faade | 2009-12-02 18:25:27 +0000 | [diff] [blame] | 624 | |
William Allen Simpson | 4957faade | 2009-12-02 18:25:27 +0000 | [diff] [blame] | 625 | return MAX_TCP_OPTION_SPACE - remaining; |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 626 | } |
| 627 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 628 | /* Compute TCP options for ESTABLISHED sockets. This is not the |
| 629 | * final wire format yet. |
| 630 | */ |
Eric Dumazet | 95c9617 | 2012-04-15 05:58:06 +0000 | [diff] [blame] | 631 | 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] | 632 | struct tcp_out_options *opts, |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 633 | struct tcp_md5sig_key **md5) |
| 634 | { |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 635 | struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL; |
| 636 | struct tcp_sock *tp = tcp_sk(sk); |
Eric Dumazet | 95c9617 | 2012-04-15 05:58:06 +0000 | [diff] [blame] | 637 | unsigned int size = 0; |
Ilpo Järvinen | cabeccb | 2009-02-28 04:44:38 +0000 | [diff] [blame] | 638 | unsigned int eff_sacks; |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 639 | |
| 640 | #ifdef CONFIG_TCP_MD5SIG |
| 641 | *md5 = tp->af_specific->md5_lookup(sk, sk); |
| 642 | if (unlikely(*md5)) { |
| 643 | opts->options |= OPTION_MD5; |
| 644 | size += TCPOLEN_MD5SIG_ALIGNED; |
| 645 | } |
| 646 | #else |
| 647 | *md5 = NULL; |
| 648 | #endif |
| 649 | |
| 650 | if (likely(tp->rx_opt.tstamp_ok)) { |
| 651 | opts->options |= OPTION_TS; |
Andrey Vagin | ee684b6 | 2013-02-11 05:50:19 +0000 | [diff] [blame] | 652 | opts->tsval = tcb ? tcb->when + tp->tsoffset : 0; |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 653 | opts->tsecr = tp->rx_opt.ts_recent; |
| 654 | size += TCPOLEN_TSTAMP_ALIGNED; |
| 655 | } |
| 656 | |
Ilpo Järvinen | cabeccb | 2009-02-28 04:44:38 +0000 | [diff] [blame] | 657 | eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; |
| 658 | if (unlikely(eff_sacks)) { |
Eric Dumazet | 95c9617 | 2012-04-15 05:58:06 +0000 | [diff] [blame] | 659 | const unsigned int remaining = MAX_TCP_OPTION_SPACE - size; |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 660 | opts->num_sack_blocks = |
Eric Dumazet | 95c9617 | 2012-04-15 05:58:06 +0000 | [diff] [blame] | 661 | min_t(unsigned int, eff_sacks, |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 662 | (remaining - TCPOLEN_SACK_BASE_ALIGNED) / |
| 663 | TCPOLEN_SACK_PERBLOCK); |
| 664 | size += TCPOLEN_SACK_BASE_ALIGNED + |
| 665 | opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; |
| 666 | } |
| 667 | |
| 668 | return size; |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 669 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 670 | |
Eric Dumazet | 46d3cea | 2012-07-11 05:50:31 +0000 | [diff] [blame] | 671 | |
| 672 | /* TCP SMALL QUEUES (TSQ) |
| 673 | * |
| 674 | * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev) |
| 675 | * to reduce RTT and bufferbloat. |
| 676 | * We do this using a special skb destructor (tcp_wfree). |
| 677 | * |
| 678 | * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb |
| 679 | * needs to be reallocated in a driver. |
| 680 | * The invariant being skb->truesize substracted from sk->sk_wmem_alloc |
| 681 | * |
| 682 | * Since transmit from skb destructor is forbidden, we use a tasklet |
| 683 | * to process all sockets that eventually need to send more skbs. |
| 684 | * We use one tasklet per cpu, with its own queue of sockets. |
| 685 | */ |
| 686 | struct tsq_tasklet { |
| 687 | struct tasklet_struct tasklet; |
| 688 | struct list_head head; /* queue of tcp sockets */ |
| 689 | }; |
| 690 | static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet); |
| 691 | |
Eric Dumazet | 6f458df | 2012-07-20 05:45:50 +0000 | [diff] [blame] | 692 | static void tcp_tsq_handler(struct sock *sk) |
| 693 | { |
| 694 | if ((1 << sk->sk_state) & |
| 695 | (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING | |
| 696 | TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) |
| 697 | tcp_write_xmit(sk, tcp_current_mss(sk), 0, 0, GFP_ATOMIC); |
| 698 | } |
Eric Dumazet | 46d3cea | 2012-07-11 05:50:31 +0000 | [diff] [blame] | 699 | /* |
| 700 | * One tasklest per cpu tries to send more skbs. |
| 701 | * We run in tasklet context but need to disable irqs when |
| 702 | * transfering tsq->head because tcp_wfree() might |
| 703 | * interrupt us (non NAPI drivers) |
| 704 | */ |
| 705 | static void tcp_tasklet_func(unsigned long data) |
| 706 | { |
| 707 | struct tsq_tasklet *tsq = (struct tsq_tasklet *)data; |
| 708 | LIST_HEAD(list); |
| 709 | unsigned long flags; |
| 710 | struct list_head *q, *n; |
| 711 | struct tcp_sock *tp; |
| 712 | struct sock *sk; |
| 713 | |
| 714 | local_irq_save(flags); |
| 715 | list_splice_init(&tsq->head, &list); |
| 716 | local_irq_restore(flags); |
| 717 | |
| 718 | list_for_each_safe(q, n, &list) { |
| 719 | tp = list_entry(q, struct tcp_sock, tsq_node); |
| 720 | list_del(&tp->tsq_node); |
| 721 | |
| 722 | sk = (struct sock *)tp; |
| 723 | bh_lock_sock(sk); |
| 724 | |
| 725 | if (!sock_owned_by_user(sk)) { |
Eric Dumazet | 6f458df | 2012-07-20 05:45:50 +0000 | [diff] [blame] | 726 | tcp_tsq_handler(sk); |
Eric Dumazet | 46d3cea | 2012-07-11 05:50:31 +0000 | [diff] [blame] | 727 | } else { |
| 728 | /* defer the work to tcp_release_cb() */ |
Eric Dumazet | 6f458df | 2012-07-20 05:45:50 +0000 | [diff] [blame] | 729 | set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags); |
Eric Dumazet | 46d3cea | 2012-07-11 05:50:31 +0000 | [diff] [blame] | 730 | } |
| 731 | bh_unlock_sock(sk); |
| 732 | |
| 733 | clear_bit(TSQ_QUEUED, &tp->tsq_flags); |
| 734 | sk_free(sk); |
| 735 | } |
| 736 | } |
| 737 | |
Eric Dumazet | 6f458df | 2012-07-20 05:45:50 +0000 | [diff] [blame] | 738 | #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \ |
| 739 | (1UL << TCP_WRITE_TIMER_DEFERRED) | \ |
Eric Dumazet | 563d34d | 2012-07-23 09:48:52 +0200 | [diff] [blame] | 740 | (1UL << TCP_DELACK_TIMER_DEFERRED) | \ |
| 741 | (1UL << TCP_MTU_REDUCED_DEFERRED)) |
Eric Dumazet | 46d3cea | 2012-07-11 05:50:31 +0000 | [diff] [blame] | 742 | /** |
| 743 | * tcp_release_cb - tcp release_sock() callback |
| 744 | * @sk: socket |
| 745 | * |
| 746 | * called from release_sock() to perform protocol dependent |
| 747 | * actions before socket release. |
| 748 | */ |
| 749 | void tcp_release_cb(struct sock *sk) |
| 750 | { |
| 751 | struct tcp_sock *tp = tcp_sk(sk); |
Eric Dumazet | 6f458df | 2012-07-20 05:45:50 +0000 | [diff] [blame] | 752 | unsigned long flags, nflags; |
Eric Dumazet | 46d3cea | 2012-07-11 05:50:31 +0000 | [diff] [blame] | 753 | |
Eric Dumazet | 6f458df | 2012-07-20 05:45:50 +0000 | [diff] [blame] | 754 | /* perform an atomic operation only if at least one flag is set */ |
| 755 | do { |
| 756 | flags = tp->tsq_flags; |
| 757 | if (!(flags & TCP_DEFERRED_ALL)) |
| 758 | return; |
| 759 | nflags = flags & ~TCP_DEFERRED_ALL; |
| 760 | } while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags); |
| 761 | |
| 762 | if (flags & (1UL << TCP_TSQ_DEFERRED)) |
| 763 | tcp_tsq_handler(sk); |
| 764 | |
Eric Dumazet | 144d56e | 2012-08-20 00:22:46 +0000 | [diff] [blame] | 765 | if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) { |
Eric Dumazet | 6f458df | 2012-07-20 05:45:50 +0000 | [diff] [blame] | 766 | tcp_write_timer_handler(sk); |
Eric Dumazet | 144d56e | 2012-08-20 00:22:46 +0000 | [diff] [blame] | 767 | __sock_put(sk); |
| 768 | } |
| 769 | if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) { |
Eric Dumazet | 6f458df | 2012-07-20 05:45:50 +0000 | [diff] [blame] | 770 | tcp_delack_timer_handler(sk); |
Eric Dumazet | 144d56e | 2012-08-20 00:22:46 +0000 | [diff] [blame] | 771 | __sock_put(sk); |
| 772 | } |
| 773 | if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) { |
Eric Dumazet | 563d34d | 2012-07-23 09:48:52 +0200 | [diff] [blame] | 774 | sk->sk_prot->mtu_reduced(sk); |
Eric Dumazet | 144d56e | 2012-08-20 00:22:46 +0000 | [diff] [blame] | 775 | __sock_put(sk); |
| 776 | } |
Eric Dumazet | 46d3cea | 2012-07-11 05:50:31 +0000 | [diff] [blame] | 777 | } |
| 778 | EXPORT_SYMBOL(tcp_release_cb); |
| 779 | |
| 780 | void __init tcp_tasklet_init(void) |
| 781 | { |
| 782 | int i; |
| 783 | |
| 784 | for_each_possible_cpu(i) { |
| 785 | struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i); |
| 786 | |
| 787 | INIT_LIST_HEAD(&tsq->head); |
| 788 | tasklet_init(&tsq->tasklet, |
| 789 | tcp_tasklet_func, |
| 790 | (unsigned long)tsq); |
| 791 | } |
| 792 | } |
| 793 | |
| 794 | /* |
| 795 | * Write buffer destructor automatically called from kfree_skb. |
| 796 | * We cant xmit new skbs from this context, as we might already |
| 797 | * hold qdisc lock. |
| 798 | */ |
Eric Dumazet | d6a4a10 | 2013-04-12 11:31:52 +0000 | [diff] [blame] | 799 | void tcp_wfree(struct sk_buff *skb) |
Eric Dumazet | 46d3cea | 2012-07-11 05:50:31 +0000 | [diff] [blame] | 800 | { |
| 801 | struct sock *sk = skb->sk; |
| 802 | struct tcp_sock *tp = tcp_sk(sk); |
| 803 | |
| 804 | if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) && |
| 805 | !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) { |
| 806 | unsigned long flags; |
| 807 | struct tsq_tasklet *tsq; |
| 808 | |
| 809 | /* Keep a ref on socket. |
| 810 | * This last ref will be released in tcp_tasklet_func() |
| 811 | */ |
| 812 | atomic_sub(skb->truesize - 1, &sk->sk_wmem_alloc); |
| 813 | |
| 814 | /* queue this socket to tasklet queue */ |
| 815 | local_irq_save(flags); |
| 816 | tsq = &__get_cpu_var(tsq_tasklet); |
| 817 | list_add(&tp->tsq_node, &tsq->head); |
| 818 | tasklet_schedule(&tsq->tasklet); |
| 819 | local_irq_restore(flags); |
| 820 | } else { |
| 821 | sock_wfree(skb); |
| 822 | } |
| 823 | } |
| 824 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 825 | /* This routine actually transmits TCP packets queued in by |
| 826 | * tcp_do_sendmsg(). This is used by both the initial |
| 827 | * transmission and possible later retransmissions. |
| 828 | * All SKB's seen here are completely headerless. It is our |
| 829 | * job to build the TCP header, and pass the packet down to |
| 830 | * IP so it can do the same plus pass the packet off to the |
| 831 | * device. |
| 832 | * |
| 833 | * We are working here with either a clone of the original |
| 834 | * SKB, or a fresh unique copy made by the retransmit engine. |
| 835 | */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 836 | static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, |
| 837 | gfp_t gfp_mask) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 838 | { |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 839 | const struct inet_connection_sock *icsk = inet_csk(sk); |
| 840 | struct inet_sock *inet; |
| 841 | struct tcp_sock *tp; |
| 842 | struct tcp_skb_cb *tcb; |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 843 | struct tcp_out_options opts; |
Eric Dumazet | 95c9617 | 2012-04-15 05:58:06 +0000 | [diff] [blame] | 844 | unsigned int tcp_options_size, tcp_header_size; |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 845 | struct tcp_md5sig_key *md5; |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 846 | struct tcphdr *th; |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 847 | int err; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 848 | |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 849 | BUG_ON(!skb || !tcp_skb_pcount(skb)); |
| 850 | |
| 851 | /* If congestion control is doing timestamping, we must |
| 852 | * take such a timestamp before we potentially clone/copy. |
| 853 | */ |
Stephen Hemminger | 164891a | 2007-04-23 22:26:16 -0700 | [diff] [blame] | 854 | if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP) |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 855 | __net_timestamp(skb); |
| 856 | |
| 857 | if (likely(clone_it)) { |
Eric Dumazet | 0e280af | 2013-04-18 06:52:51 +0000 | [diff] [blame] | 858 | const struct sk_buff *fclone = skb + 1; |
| 859 | |
| 860 | if (unlikely(skb->fclone == SKB_FCLONE_ORIG && |
| 861 | fclone->fclone == SKB_FCLONE_CLONE)) |
| 862 | NET_INC_STATS_BH(sock_net(sk), |
| 863 | LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES); |
| 864 | |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 865 | if (unlikely(skb_cloned(skb))) |
| 866 | skb = pskb_copy(skb, gfp_mask); |
| 867 | else |
| 868 | skb = skb_clone(skb, gfp_mask); |
| 869 | if (unlikely(!skb)) |
| 870 | return -ENOBUFS; |
| 871 | } |
| 872 | |
| 873 | inet = inet_sk(sk); |
| 874 | tp = tcp_sk(sk); |
| 875 | tcb = TCP_SKB_CB(skb); |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 876 | memset(&opts, 0, sizeof(opts)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 877 | |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 878 | if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 879 | tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); |
| 880 | else |
| 881 | tcp_options_size = tcp_established_options(sk, skb, &opts, |
| 882 | &md5); |
| 883 | tcp_header_size = tcp_options_size + sizeof(struct tcphdr); |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 884 | |
Eric Dumazet | 547669d | 2013-05-23 07:44:20 +0000 | [diff] [blame] | 885 | if (tcp_packets_in_flight(tp) == 0) |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 886 | tcp_ca_event(sk, CA_EVENT_TX_START); |
Eric Dumazet | 547669d | 2013-05-23 07:44:20 +0000 | [diff] [blame] | 887 | |
| 888 | /* if no packet is in qdisc/device queue, then allow XPS to select |
| 889 | * another queue. |
| 890 | */ |
| 891 | skb->ooo_okay = sk_wmem_alloc_get(sk) == 0; |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 892 | |
Arnaldo Carvalho de Melo | aa8223c | 2007-04-10 21:04:22 -0700 | [diff] [blame] | 893 | skb_push(skb, tcp_header_size); |
| 894 | skb_reset_transport_header(skb); |
Eric Dumazet | 46d3cea | 2012-07-11 05:50:31 +0000 | [diff] [blame] | 895 | |
| 896 | skb_orphan(skb); |
| 897 | skb->sk = sk; |
| 898 | skb->destructor = (sysctl_tcp_limit_output_bytes > 0) ? |
| 899 | tcp_wfree : sock_wfree; |
| 900 | atomic_add(skb->truesize, &sk->sk_wmem_alloc); |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 901 | |
| 902 | /* Build TCP header and checksum it. */ |
Arnaldo Carvalho de Melo | aa8223c | 2007-04-10 21:04:22 -0700 | [diff] [blame] | 903 | th = tcp_hdr(skb); |
Eric Dumazet | c720c7e | 2009-10-15 06:30:45 +0000 | [diff] [blame] | 904 | th->source = inet->inet_sport; |
| 905 | th->dest = inet->inet_dport; |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 906 | th->seq = htonl(tcb->seq); |
| 907 | th->ack_seq = htonl(tp->rcv_nxt); |
Al Viro | df7a3b0 | 2006-09-27 18:38:52 -0700 | [diff] [blame] | 908 | *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 909 | tcb->tcp_flags); |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 910 | |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 911 | if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) { |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 912 | /* RFC1323: The window in SYN & SYN/ACK segments |
| 913 | * is never scaled. |
| 914 | */ |
Ilpo Järvinen | 600ff0c | 2007-02-13 12:42:11 -0800 | [diff] [blame] | 915 | th->window = htons(min(tp->rcv_wnd, 65535U)); |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 916 | } else { |
| 917 | th->window = htons(tcp_select_window(sk)); |
| 918 | } |
| 919 | th->check = 0; |
| 920 | th->urg_ptr = 0; |
| 921 | |
Ilpo Järvinen | 33f5f57 | 2008-10-07 14:43:06 -0700 | [diff] [blame] | 922 | /* 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] | 923 | if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { |
| 924 | if (before(tp->snd_up, tcb->seq + 0x10000)) { |
| 925 | th->urg_ptr = htons(tp->snd_up - tcb->seq); |
| 926 | th->urg = 1; |
| 927 | } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { |
Eric Dumazet | 0eae88f | 2010-04-20 19:06:52 -0700 | [diff] [blame] | 928 | th->urg_ptr = htons(0xFFFF); |
Herbert Xu | 7691367 | 2009-02-21 23:52:29 -0800 | [diff] [blame] | 929 | th->urg = 1; |
| 930 | } |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 931 | } |
| 932 | |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 933 | tcp_options_write((__be32 *)(th + 1), tp, &opts); |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 934 | if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0)) |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 935 | TCP_ECN_send(sk, skb, tcp_header_size); |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 936 | |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 937 | #ifdef CONFIG_TCP_MD5SIG |
| 938 | /* Calculate the MD5 hash, as we have all we need now */ |
| 939 | if (md5) { |
Eric Dumazet | a465419 | 2010-05-16 00:36:33 -0700 | [diff] [blame] | 940 | sk_nocaps_add(sk, NETIF_F_GSO_MASK); |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 941 | tp->af_specific->calc_md5_hash(opts.hash_location, |
Adam Langley | 49a72df | 2008-07-19 00:01:42 -0700 | [diff] [blame] | 942 | md5, sk, NULL, skb); |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 943 | } |
| 944 | #endif |
| 945 | |
Herbert Xu | bb29624 | 2010-04-11 02:15:55 +0000 | [diff] [blame] | 946 | icsk->icsk_af_ops->send_check(sk, skb); |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 947 | |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 948 | if (likely(tcb->tcp_flags & TCPHDR_ACK)) |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 949 | tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); |
| 950 | |
| 951 | if (skb->len != tcp_header_size) |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 952 | tcp_event_data_sent(tp, sk); |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 953 | |
Wei Yongjun | bd37a08 | 2006-08-07 21:04:15 -0700 | [diff] [blame] | 954 | 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] | 955 | TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, |
| 956 | tcp_skb_pcount(skb)); |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 957 | |
David S. Miller | d9d8da8 | 2011-05-06 22:23:20 -0700 | [diff] [blame] | 958 | err = icsk->icsk_af_ops->queue_xmit(skb, &inet->cork.fl); |
Hua Zhong | 83de47c | 2006-04-28 15:26:50 -0700 | [diff] [blame] | 959 | if (likely(err <= 0)) |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 960 | return err; |
| 961 | |
Ilpo Järvinen | 3cfe3ba | 2007-02-27 10:09:49 -0800 | [diff] [blame] | 962 | tcp_enter_cwr(sk, 1); |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 963 | |
Gerrit Renker | b9df3cb | 2006-11-14 11:21:36 -0200 | [diff] [blame] | 964 | return net_xmit_eval(err); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 965 | } |
| 966 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 967 | /* This routine just queues the buffer for sending. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 968 | * |
| 969 | * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, |
| 970 | * otherwise socket can stall. |
| 971 | */ |
| 972 | static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) |
| 973 | { |
| 974 | struct tcp_sock *tp = tcp_sk(sk); |
| 975 | |
| 976 | /* Advance write_seq and place onto the write_queue. */ |
| 977 | tp->write_seq = TCP_SKB_CB(skb)->end_seq; |
| 978 | skb_header_release(skb); |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 979 | tcp_add_write_queue_tail(sk, skb); |
Hideo Aoki | 3ab224b | 2007-12-31 00:11:19 -0800 | [diff] [blame] | 980 | sk->sk_wmem_queued += skb->truesize; |
| 981 | sk_mem_charge(sk, skb->truesize); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 982 | } |
| 983 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 984 | /* Initialize TSO segments for a packet. */ |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 985 | static void tcp_set_skb_tso_segs(const struct sock *sk, struct sk_buff *skb, |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 986 | unsigned int mss_now) |
David S. Miller | f6302d1 | 2005-07-05 15:18:03 -0700 | [diff] [blame] | 987 | { |
Herbert Xu | 8e5b9dd | 2009-06-28 18:03:30 +0000 | [diff] [blame] | 988 | if (skb->len <= mss_now || !sk_can_gso(sk) || |
| 989 | skb->ip_summed == CHECKSUM_NONE) { |
David S. Miller | f6302d1 | 2005-07-05 15:18:03 -0700 | [diff] [blame] | 990 | /* Avoid the costly divide in the normal |
| 991 | * non-TSO case. |
| 992 | */ |
Herbert Xu | 7967168 | 2006-06-22 02:40:14 -0700 | [diff] [blame] | 993 | skb_shinfo(skb)->gso_segs = 1; |
| 994 | skb_shinfo(skb)->gso_size = 0; |
Pravin B Shelar | c9af6db | 2013-02-11 09:27:41 +0000 | [diff] [blame] | 995 | skb_shinfo(skb)->gso_type = 0; |
David S. Miller | f6302d1 | 2005-07-05 15:18:03 -0700 | [diff] [blame] | 996 | } else { |
Ilpo Järvinen | 356f89e | 2007-08-24 23:00:31 -0700 | [diff] [blame] | 997 | skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now); |
Herbert Xu | 7967168 | 2006-06-22 02:40:14 -0700 | [diff] [blame] | 998 | skb_shinfo(skb)->gso_size = mss_now; |
Pravin B Shelar | c9af6db | 2013-02-11 09:27:41 +0000 | [diff] [blame] | 999 | skb_shinfo(skb)->gso_type = sk->sk_gso_type; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1000 | } |
| 1001 | } |
| 1002 | |
Ilpo Järvinen | 91fed7a | 2007-10-09 01:24:15 -0700 | [diff] [blame] | 1003 | /* When a modification to fackets out becomes necessary, we need to check |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1004 | * skb is counted to fackets_out or not. |
Ilpo Järvinen | 91fed7a | 2007-10-09 01:24:15 -0700 | [diff] [blame] | 1005 | */ |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 1006 | 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] | 1007 | int decr) |
| 1008 | { |
Ilpo Järvinen | a47e5a9 | 2007-11-15 19:41:46 -0800 | [diff] [blame] | 1009 | struct tcp_sock *tp = tcp_sk(sk); |
| 1010 | |
Ilpo Järvinen | dc86967 | 2007-10-01 15:27:19 -0700 | [diff] [blame] | 1011 | if (!tp->sacked_out || tcp_is_reno(tp)) |
Ilpo Järvinen | 91fed7a | 2007-10-09 01:24:15 -0700 | [diff] [blame] | 1012 | return; |
| 1013 | |
Ilpo Järvinen | 6859d49 | 2007-12-02 00:48:06 +0200 | [diff] [blame] | 1014 | if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq)) |
Ilpo Järvinen | 91fed7a | 2007-10-09 01:24:15 -0700 | [diff] [blame] | 1015 | tp->fackets_out -= decr; |
Ilpo Järvinen | 91fed7a | 2007-10-09 01:24:15 -0700 | [diff] [blame] | 1016 | } |
| 1017 | |
Ilpo Järvinen | 797108d | 2009-04-01 23:15:17 +0000 | [diff] [blame] | 1018 | /* Pcount in the middle of the write queue got changed, we need to do various |
| 1019 | * tweaks to fix counters |
| 1020 | */ |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 1021 | 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] | 1022 | { |
| 1023 | struct tcp_sock *tp = tcp_sk(sk); |
| 1024 | |
| 1025 | tp->packets_out -= decr; |
| 1026 | |
| 1027 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) |
| 1028 | tp->sacked_out -= decr; |
| 1029 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) |
| 1030 | tp->retrans_out -= decr; |
| 1031 | if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) |
| 1032 | tp->lost_out -= decr; |
| 1033 | |
| 1034 | /* Reno case is special. Sigh... */ |
| 1035 | if (tcp_is_reno(tp) && decr > 0) |
| 1036 | tp->sacked_out -= min_t(u32, tp->sacked_out, decr); |
| 1037 | |
| 1038 | tcp_adjust_fackets_out(sk, skb, decr); |
| 1039 | |
| 1040 | if (tp->lost_skb_hint && |
| 1041 | 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] | 1042 | (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))) |
Ilpo Järvinen | 797108d | 2009-04-01 23:15:17 +0000 | [diff] [blame] | 1043 | tp->lost_cnt_hint -= decr; |
| 1044 | |
| 1045 | tcp_verify_left_out(tp); |
| 1046 | } |
| 1047 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1048 | /* Function to create two new TCP segments. Shrinks the given segment |
| 1049 | * 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] | 1050 | * packet to the list. This won't be called frequently, I hope. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1051 | * Remember, these are still headerless SKBs at this point. |
| 1052 | */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1053 | int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, |
| 1054 | unsigned int mss_now) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1055 | { |
| 1056 | struct tcp_sock *tp = tcp_sk(sk); |
| 1057 | struct sk_buff *buff; |
David S. Miller | 6475be1 | 2005-09-01 22:47:01 -0700 | [diff] [blame] | 1058 | int nsize, old_factor; |
Herbert Xu | b60b49e | 2006-04-19 21:35:00 -0700 | [diff] [blame] | 1059 | int nlen; |
Ilpo Järvinen | 9ce0146 | 2009-02-28 04:44:42 +0000 | [diff] [blame] | 1060 | u8 flags; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1061 | |
Ilpo Järvinen | 2fceec1 | 2011-04-01 21:47:41 -0700 | [diff] [blame] | 1062 | if (WARN_ON(len > skb->len)) |
| 1063 | return -EINVAL; |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 1064 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1065 | nsize = skb_headlen(skb) - len; |
| 1066 | if (nsize < 0) |
| 1067 | nsize = 0; |
| 1068 | |
| 1069 | if (skb_cloned(skb) && |
| 1070 | skb_is_nonlinear(skb) && |
| 1071 | pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) |
| 1072 | return -ENOMEM; |
| 1073 | |
| 1074 | /* Get a new skb... force flag on. */ |
| 1075 | buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC); |
| 1076 | if (buff == NULL) |
| 1077 | return -ENOMEM; /* We'll just try again later. */ |
Herbert Xu | ef5cb97 | 2006-04-18 13:24:14 -0700 | [diff] [blame] | 1078 | |
Hideo Aoki | 3ab224b | 2007-12-31 00:11:19 -0800 | [diff] [blame] | 1079 | sk->sk_wmem_queued += buff->truesize; |
| 1080 | sk_mem_charge(sk, buff->truesize); |
Herbert Xu | b60b49e | 2006-04-19 21:35:00 -0700 | [diff] [blame] | 1081 | nlen = skb->len - len - nsize; |
| 1082 | buff->truesize += nlen; |
| 1083 | skb->truesize -= nlen; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1084 | |
| 1085 | /* Correct the sequence numbers. */ |
| 1086 | TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; |
| 1087 | TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; |
| 1088 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; |
| 1089 | |
| 1090 | /* PSH and FIN should only be set in the second packet. */ |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 1091 | flags = TCP_SKB_CB(skb)->tcp_flags; |
| 1092 | TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); |
| 1093 | TCP_SKB_CB(buff)->tcp_flags = flags; |
Herbert Xu | e14c3ca | 2005-09-19 18:18:38 -0700 | [diff] [blame] | 1094 | TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1095 | |
Patrick McHardy | 84fa793 | 2006-08-29 16:44:56 -0700 | [diff] [blame] | 1096 | if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1097 | /* Copy and checksum data tail into the new buffer. */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1098 | buff->csum = csum_partial_copy_nocheck(skb->data + len, |
| 1099 | skb_put(buff, nsize), |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1100 | nsize, 0); |
| 1101 | |
| 1102 | skb_trim(skb, len); |
| 1103 | |
| 1104 | skb->csum = csum_block_sub(skb->csum, buff->csum, len); |
| 1105 | } else { |
Patrick McHardy | 84fa793 | 2006-08-29 16:44:56 -0700 | [diff] [blame] | 1106 | skb->ip_summed = CHECKSUM_PARTIAL; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1107 | skb_split(skb, buff, len); |
| 1108 | } |
| 1109 | |
| 1110 | buff->ip_summed = skb->ip_summed; |
| 1111 | |
| 1112 | /* Looks stupid, but our code really uses when of |
| 1113 | * skbs, which it never sent before. --ANK |
| 1114 | */ |
| 1115 | TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when; |
Patrick McHardy | a61bbcf | 2005-08-14 17:24:31 -0700 | [diff] [blame] | 1116 | buff->tstamp = skb->tstamp; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1117 | |
David S. Miller | 6475be1 | 2005-09-01 22:47:01 -0700 | [diff] [blame] | 1118 | old_factor = tcp_skb_pcount(skb); |
| 1119 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1120 | /* Fix up tso_factor for both original and new SKB. */ |
David S. Miller | 846998a | 2005-08-04 19:52:01 -0700 | [diff] [blame] | 1121 | tcp_set_skb_tso_segs(sk, skb, mss_now); |
| 1122 | tcp_set_skb_tso_segs(sk, buff, mss_now); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1123 | |
David S. Miller | 6475be1 | 2005-09-01 22:47:01 -0700 | [diff] [blame] | 1124 | /* If this packet has been sent out already, we must |
| 1125 | * adjust the various packet counters. |
| 1126 | */ |
Herbert Xu | cf0b450 | 2005-09-08 15:10:52 -0700 | [diff] [blame] | 1127 | if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { |
David S. Miller | 6475be1 | 2005-09-01 22:47:01 -0700 | [diff] [blame] | 1128 | int diff = old_factor - tcp_skb_pcount(skb) - |
| 1129 | tcp_skb_pcount(buff); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1130 | |
Ilpo Järvinen | 797108d | 2009-04-01 23:15:17 +0000 | [diff] [blame] | 1131 | if (diff) |
| 1132 | tcp_adjust_pcount(sk, skb, diff); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1133 | } |
| 1134 | |
| 1135 | /* Link BUFF into the send queue. */ |
David S. Miller | f44b527 | 2005-07-05 15:18:34 -0700 | [diff] [blame] | 1136 | skb_header_release(buff); |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1137 | tcp_insert_write_queue_after(skb, buff, sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1138 | |
| 1139 | return 0; |
| 1140 | } |
| 1141 | |
| 1142 | /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c |
| 1143 | * eventually). The difference is that pulled data not copied, but |
| 1144 | * immediately discarded. |
| 1145 | */ |
Herbert Xu ~{PmVHI~} | f291196 | 2006-06-05 15:03:37 -0700 | [diff] [blame] | 1146 | static void __pskb_trim_head(struct sk_buff *skb, int len) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1147 | { |
| 1148 | int i, k, eat; |
| 1149 | |
Eric Dumazet | 4fa48bf | 2011-12-04 08:51:08 +0000 | [diff] [blame] | 1150 | eat = min_t(int, len, skb_headlen(skb)); |
| 1151 | if (eat) { |
| 1152 | __skb_pull(skb, eat); |
| 1153 | len -= eat; |
| 1154 | if (!len) |
| 1155 | return; |
| 1156 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1157 | eat = len; |
| 1158 | k = 0; |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1159 | for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) { |
Eric Dumazet | 9e903e0 | 2011-10-18 21:00:24 +0000 | [diff] [blame] | 1160 | int size = skb_frag_size(&skb_shinfo(skb)->frags[i]); |
| 1161 | |
| 1162 | if (size <= eat) { |
Ian Campbell | aff65da | 2011-08-22 23:44:59 +0000 | [diff] [blame] | 1163 | skb_frag_unref(skb, i); |
Eric Dumazet | 9e903e0 | 2011-10-18 21:00:24 +0000 | [diff] [blame] | 1164 | eat -= size; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1165 | } else { |
| 1166 | skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i]; |
| 1167 | if (eat) { |
| 1168 | skb_shinfo(skb)->frags[k].page_offset += eat; |
Eric Dumazet | 9e903e0 | 2011-10-18 21:00:24 +0000 | [diff] [blame] | 1169 | skb_frag_size_sub(&skb_shinfo(skb)->frags[k], eat); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1170 | eat = 0; |
| 1171 | } |
| 1172 | k++; |
| 1173 | } |
| 1174 | } |
| 1175 | skb_shinfo(skb)->nr_frags = k; |
| 1176 | |
Arnaldo Carvalho de Melo | 27a884d | 2007-04-19 20:29:13 -0700 | [diff] [blame] | 1177 | skb_reset_tail_pointer(skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1178 | skb->data_len -= len; |
| 1179 | skb->len = skb->data_len; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1180 | } |
| 1181 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 1182 | /* Remove acked data from a packet in the transmit queue. */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1183 | int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) |
| 1184 | { |
Pravin B Shelar | 14bbd6a | 2013-02-14 09:44:49 +0000 | [diff] [blame] | 1185 | if (skb_unclone(skb, GFP_ATOMIC)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1186 | return -ENOMEM; |
| 1187 | |
Eric Dumazet | 4fa48bf | 2011-12-04 08:51:08 +0000 | [diff] [blame] | 1188 | __pskb_trim_head(skb, len); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1189 | |
| 1190 | TCP_SKB_CB(skb)->seq += len; |
Patrick McHardy | 84fa793 | 2006-08-29 16:44:56 -0700 | [diff] [blame] | 1191 | skb->ip_summed = CHECKSUM_PARTIAL; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1192 | |
| 1193 | skb->truesize -= len; |
| 1194 | sk->sk_wmem_queued -= len; |
Hideo Aoki | 3ab224b | 2007-12-31 00:11:19 -0800 | [diff] [blame] | 1195 | sk_mem_uncharge(sk, len); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1196 | sock_set_flag(sk, SOCK_QUEUE_SHRUNK); |
| 1197 | |
Neal Cardwell | 5b35e1e | 2012-01-28 17:29:46 +0000 | [diff] [blame] | 1198 | /* Any change of skb->len requires recalculation of tso factor. */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1199 | if (tcp_skb_pcount(skb) > 1) |
Neal Cardwell | 5b35e1e | 2012-01-28 17:29:46 +0000 | [diff] [blame] | 1200 | tcp_set_skb_tso_segs(sk, skb, tcp_skb_mss(skb)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1201 | |
| 1202 | return 0; |
| 1203 | } |
| 1204 | |
Yuchung Cheng | 1b63edd | 2013-02-22 08:59:06 +0000 | [diff] [blame] | 1205 | /* Calculate MSS not accounting any TCP options. */ |
| 1206 | static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu) |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1207 | { |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 1208 | const struct tcp_sock *tp = tcp_sk(sk); |
| 1209 | const struct inet_connection_sock *icsk = inet_csk(sk); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1210 | int mss_now; |
| 1211 | |
| 1212 | /* Calculate base mss without TCP options: |
| 1213 | It is MMS_S - sizeof(tcphdr) of rfc1122 |
| 1214 | */ |
| 1215 | mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); |
| 1216 | |
Eric Dumazet | 6746960 | 2012-04-24 07:37:38 +0000 | [diff] [blame] | 1217 | /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ |
| 1218 | if (icsk->icsk_af_ops->net_frag_header_len) { |
| 1219 | const struct dst_entry *dst = __sk_dst_get(sk); |
| 1220 | |
| 1221 | if (dst && dst_allfrag(dst)) |
| 1222 | mss_now -= icsk->icsk_af_ops->net_frag_header_len; |
| 1223 | } |
| 1224 | |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1225 | /* Clamp it (mss_clamp does not include tcp options) */ |
| 1226 | if (mss_now > tp->rx_opt.mss_clamp) |
| 1227 | mss_now = tp->rx_opt.mss_clamp; |
| 1228 | |
| 1229 | /* Now subtract optional transport overhead */ |
| 1230 | mss_now -= icsk->icsk_ext_hdr_len; |
| 1231 | |
| 1232 | /* Then reserve room for full set of TCP options and 8 bytes of data */ |
| 1233 | if (mss_now < 48) |
| 1234 | mss_now = 48; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1235 | return mss_now; |
| 1236 | } |
| 1237 | |
Yuchung Cheng | 1b63edd | 2013-02-22 08:59:06 +0000 | [diff] [blame] | 1238 | /* Calculate MSS. Not accounting for SACKs here. */ |
| 1239 | int tcp_mtu_to_mss(struct sock *sk, int pmtu) |
| 1240 | { |
| 1241 | /* Subtract TCP options size, not including SACKs */ |
| 1242 | return __tcp_mtu_to_mss(sk, pmtu) - |
| 1243 | (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr)); |
| 1244 | } |
| 1245 | |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1246 | /* Inverse of above */ |
Eric Dumazet | 6746960 | 2012-04-24 07:37:38 +0000 | [diff] [blame] | 1247 | int tcp_mss_to_mtu(struct sock *sk, int mss) |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1248 | { |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 1249 | const struct tcp_sock *tp = tcp_sk(sk); |
| 1250 | const struct inet_connection_sock *icsk = inet_csk(sk); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1251 | int mtu; |
| 1252 | |
| 1253 | mtu = mss + |
| 1254 | tp->tcp_header_len + |
| 1255 | icsk->icsk_ext_hdr_len + |
| 1256 | icsk->icsk_af_ops->net_header_len; |
| 1257 | |
Eric Dumazet | 6746960 | 2012-04-24 07:37:38 +0000 | [diff] [blame] | 1258 | /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ |
| 1259 | if (icsk->icsk_af_ops->net_frag_header_len) { |
| 1260 | const struct dst_entry *dst = __sk_dst_get(sk); |
| 1261 | |
| 1262 | if (dst && dst_allfrag(dst)) |
| 1263 | mtu += icsk->icsk_af_ops->net_frag_header_len; |
| 1264 | } |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1265 | return mtu; |
| 1266 | } |
| 1267 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 1268 | /* MTU probing init per socket */ |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1269 | void tcp_mtup_init(struct sock *sk) |
| 1270 | { |
| 1271 | struct tcp_sock *tp = tcp_sk(sk); |
| 1272 | struct inet_connection_sock *icsk = inet_csk(sk); |
| 1273 | |
| 1274 | icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1; |
| 1275 | 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] | 1276 | icsk->icsk_af_ops->net_header_len; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1277 | icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss); |
| 1278 | icsk->icsk_mtup.probe_size = 0; |
| 1279 | } |
Eric Dumazet | 4bc2f18 | 2010-07-09 21:22:10 +0000 | [diff] [blame] | 1280 | EXPORT_SYMBOL(tcp_mtup_init); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1281 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1282 | /* This function synchronize snd mss to current pmtu/exthdr set. |
| 1283 | |
| 1284 | tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts |
| 1285 | for TCP options, but includes only bare TCP header. |
| 1286 | |
| 1287 | tp->rx_opt.mss_clamp is mss negotiated at connection setup. |
Stephen Hemminger | caa20d9a | 2005-11-10 17:13:47 -0800 | [diff] [blame] | 1288 | It is minimum of user_mss and mss received with SYN. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1289 | It also does not include TCP options. |
| 1290 | |
Arnaldo Carvalho de Melo | d83d846 | 2005-12-13 23:26:10 -0800 | [diff] [blame] | 1291 | 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] | 1292 | |
| 1293 | tp->mss_cache is current effective sending mss, including |
| 1294 | all tcp options except for SACKs. It is evaluated, |
| 1295 | taking into account current pmtu, but never exceeds |
| 1296 | tp->rx_opt.mss_clamp. |
| 1297 | |
| 1298 | NOTE1. rfc1122 clearly states that advertised MSS |
| 1299 | DOES NOT include either tcp or ip options. |
| 1300 | |
Arnaldo Carvalho de Melo | d83d846 | 2005-12-13 23:26:10 -0800 | [diff] [blame] | 1301 | NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache |
| 1302 | are READ ONLY outside this function. --ANK (980731) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1303 | */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1304 | unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) |
| 1305 | { |
| 1306 | struct tcp_sock *tp = tcp_sk(sk); |
Arnaldo Carvalho de Melo | d83d846 | 2005-12-13 23:26:10 -0800 | [diff] [blame] | 1307 | struct inet_connection_sock *icsk = inet_csk(sk); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1308 | int mss_now; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1309 | |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1310 | if (icsk->icsk_mtup.search_high > pmtu) |
| 1311 | icsk->icsk_mtup.search_high = pmtu; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1312 | |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1313 | mss_now = tcp_mtu_to_mss(sk, pmtu); |
Ilpo Järvinen | 409d22b | 2007-12-31 14:57:40 -0800 | [diff] [blame] | 1314 | mss_now = tcp_bound_to_half_wnd(tp, mss_now); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1315 | |
| 1316 | /* And store cached results */ |
Arnaldo Carvalho de Melo | d83d846 | 2005-12-13 23:26:10 -0800 | [diff] [blame] | 1317 | icsk->icsk_pmtu_cookie = pmtu; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1318 | if (icsk->icsk_mtup.enabled) |
| 1319 | 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] | 1320 | tp->mss_cache = mss_now; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1321 | |
| 1322 | return mss_now; |
| 1323 | } |
Eric Dumazet | 4bc2f18 | 2010-07-09 21:22:10 +0000 | [diff] [blame] | 1324 | EXPORT_SYMBOL(tcp_sync_mss); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1325 | |
| 1326 | /* Compute the current effective MSS, taking SACKs and IP options, |
| 1327 | * and even PMTU discovery events into account. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1328 | */ |
Ilpo Järvinen | 0c54b85 | 2009-03-14 14:23:05 +0000 | [diff] [blame] | 1329 | unsigned int tcp_current_mss(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1330 | { |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 1331 | const struct tcp_sock *tp = tcp_sk(sk); |
| 1332 | const struct dst_entry *dst = __sk_dst_get(sk); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1333 | u32 mss_now; |
Eric Dumazet | 95c9617 | 2012-04-15 05:58:06 +0000 | [diff] [blame] | 1334 | unsigned int header_len; |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 1335 | struct tcp_out_options opts; |
| 1336 | struct tcp_md5sig_key *md5; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1337 | |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1338 | mss_now = tp->mss_cache; |
| 1339 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1340 | if (dst) { |
| 1341 | u32 mtu = dst_mtu(dst); |
Arnaldo Carvalho de Melo | d83d846 | 2005-12-13 23:26:10 -0800 | [diff] [blame] | 1342 | if (mtu != inet_csk(sk)->icsk_pmtu_cookie) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1343 | mss_now = tcp_sync_mss(sk, mtu); |
| 1344 | } |
| 1345 | |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 1346 | header_len = tcp_established_options(sk, NULL, &opts, &md5) + |
| 1347 | sizeof(struct tcphdr); |
| 1348 | /* The mss_cache is sized based on tp->tcp_header_len, which assumes |
| 1349 | * some common options. If this is an odd packet (because we have SACK |
| 1350 | * blocks etc) then our calculated header_len will be different, and |
| 1351 | * we have to adjust mss_now correspondingly */ |
| 1352 | if (header_len != tp->tcp_header_len) { |
| 1353 | int delta = (int) header_len - tp->tcp_header_len; |
| 1354 | mss_now -= delta; |
| 1355 | } |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 1356 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1357 | return mss_now; |
| 1358 | } |
| 1359 | |
David S. Miller | a762a98 | 2005-07-05 15:18:51 -0700 | [diff] [blame] | 1360 | /* Congestion window validation. (RFC2861) */ |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 1361 | static void tcp_cwnd_validate(struct sock *sk) |
David S. Miller | a762a98 | 2005-07-05 15:18:51 -0700 | [diff] [blame] | 1362 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 1363 | struct tcp_sock *tp = tcp_sk(sk); |
David S. Miller | a762a98 | 2005-07-05 15:18:51 -0700 | [diff] [blame] | 1364 | |
Ilpo Järvinen | d436d68 | 2007-12-31 14:58:00 -0800 | [diff] [blame] | 1365 | if (tp->packets_out >= tp->snd_cwnd) { |
David S. Miller | a762a98 | 2005-07-05 15:18:51 -0700 | [diff] [blame] | 1366 | /* Network is feed fully. */ |
| 1367 | tp->snd_cwnd_used = 0; |
| 1368 | tp->snd_cwnd_stamp = tcp_time_stamp; |
| 1369 | } else { |
| 1370 | /* Network starves. */ |
| 1371 | if (tp->packets_out > tp->snd_cwnd_used) |
| 1372 | tp->snd_cwnd_used = tp->packets_out; |
| 1373 | |
David S. Miller | 15d33c0 | 2007-04-09 13:23:14 -0700 | [diff] [blame] | 1374 | if (sysctl_tcp_slow_start_after_idle && |
| 1375 | (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] | 1376 | tcp_cwnd_application_limited(sk); |
| 1377 | } |
| 1378 | } |
| 1379 | |
Ilpo Järvinen | 0e3a480 | 2007-12-24 21:33:45 -0800 | [diff] [blame] | 1380 | /* Returns the portion of skb which can be sent right away without |
| 1381 | * introducing MSS oddities to segment boundaries. In rare cases where |
| 1382 | * mss_now != mss_cache, we will request caller to create a small skb |
| 1383 | * per input skb which could be mostly avoided here (if desired). |
Ilpo Järvinen | 5ea3a74 | 2008-03-11 17:55:27 -0700 | [diff] [blame] | 1384 | * |
| 1385 | * We explicitly want to create a request for splitting write queue tail |
| 1386 | * to a small skb for Nagle purposes while avoiding unnecessary modulos, |
| 1387 | * thus all the complexity (cwnd_len is always MSS multiple which we |
| 1388 | * return whenever allowed by the other factors). Basically we need the |
| 1389 | * modulo only when the receiver window alone is the limiting factor or |
| 1390 | * when we would be allowed to send the split-due-to-Nagle skb fully. |
Ilpo Järvinen | 0e3a480 | 2007-12-24 21:33:45 -0800 | [diff] [blame] | 1391 | */ |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 1392 | static unsigned int tcp_mss_split_point(const struct sock *sk, const struct sk_buff *skb, |
Ben Hutchings | 1485348 | 2012-07-30 16:11:42 +0000 | [diff] [blame] | 1393 | unsigned int mss_now, unsigned int max_segs) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1394 | { |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 1395 | const struct tcp_sock *tp = tcp_sk(sk); |
Ben Hutchings | 1485348 | 2012-07-30 16:11:42 +0000 | [diff] [blame] | 1396 | u32 needed, window, max_len; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1397 | |
Ilpo Järvinen | 90840de | 2007-12-31 04:48:41 -0800 | [diff] [blame] | 1398 | window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; |
Ben Hutchings | 1485348 | 2012-07-30 16:11:42 +0000 | [diff] [blame] | 1399 | max_len = mss_now * max_segs; |
Ilpo Järvinen | 0e3a480 | 2007-12-24 21:33:45 -0800 | [diff] [blame] | 1400 | |
Ben Hutchings | 1485348 | 2012-07-30 16:11:42 +0000 | [diff] [blame] | 1401 | if (likely(max_len <= window && skb != tcp_write_queue_tail(sk))) |
| 1402 | return max_len; |
Ilpo Järvinen | 0e3a480 | 2007-12-24 21:33:45 -0800 | [diff] [blame] | 1403 | |
Ilpo Järvinen | 5ea3a74 | 2008-03-11 17:55:27 -0700 | [diff] [blame] | 1404 | needed = min(skb->len, window); |
| 1405 | |
Ben Hutchings | 1485348 | 2012-07-30 16:11:42 +0000 | [diff] [blame] | 1406 | if (max_len <= needed) |
| 1407 | return max_len; |
Ilpo Järvinen | 0e3a480 | 2007-12-24 21:33:45 -0800 | [diff] [blame] | 1408 | |
Ilpo Järvinen | 0e3a480 | 2007-12-24 21:33:45 -0800 | [diff] [blame] | 1409 | return needed - needed % mss_now; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1410 | } |
| 1411 | |
| 1412 | /* Can at least one segment of SKB be sent right now, according to the |
| 1413 | * congestion window rules? If so, return how many segments are allowed. |
| 1414 | */ |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 1415 | static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, |
| 1416 | const struct sk_buff *skb) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1417 | { |
| 1418 | u32 in_flight, cwnd; |
| 1419 | |
| 1420 | /* Don't be strict about the congestion window for the final FIN. */ |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 1421 | if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && |
| 1422 | tcp_skb_pcount(skb) == 1) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1423 | return 1; |
| 1424 | |
| 1425 | in_flight = tcp_packets_in_flight(tp); |
| 1426 | cwnd = tp->snd_cwnd; |
| 1427 | if (in_flight < cwnd) |
| 1428 | return (cwnd - in_flight); |
| 1429 | |
| 1430 | return 0; |
| 1431 | } |
| 1432 | |
Uwe Kleine-König | b595076 | 2010-11-01 15:38:34 -0400 | [diff] [blame] | 1433 | /* Initialize TSO state of a skb. |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 1434 | * This must be invoked the first time we consider transmitting |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1435 | * SKB onto the wire. |
| 1436 | */ |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 1437 | static int tcp_init_tso_segs(const struct sock *sk, struct sk_buff *skb, |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1438 | unsigned int mss_now) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1439 | { |
| 1440 | int tso_segs = tcp_skb_pcount(skb); |
| 1441 | |
Ilpo Järvinen | f8269a4 | 2008-12-03 21:24:48 -0800 | [diff] [blame] | 1442 | if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { |
David S. Miller | 846998a | 2005-08-04 19:52:01 -0700 | [diff] [blame] | 1443 | tcp_set_skb_tso_segs(sk, skb, mss_now); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1444 | tso_segs = tcp_skb_pcount(skb); |
| 1445 | } |
| 1446 | return tso_segs; |
| 1447 | } |
| 1448 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 1449 | /* Minshall's variant of the Nagle send check. */ |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1450 | static inline bool tcp_minshall_check(const struct tcp_sock *tp) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1451 | { |
Jianjun Kong | 09cb105 | 2008-11-03 00:27:11 -0800 | [diff] [blame] | 1452 | return after(tp->snd_sml, tp->snd_una) && |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1453 | !after(tp->snd_sml, tp->snd_nxt); |
| 1454 | } |
| 1455 | |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1456 | /* Return false, if packet can be sent now without violation Nagle's rules: |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1457 | * 1. It is full sized. |
| 1458 | * 2. Or it contains FIN. (already checked by caller) |
Feng King | 6d67e9b | 2011-11-05 04:23:23 +0000 | [diff] [blame] | 1459 | * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1460 | * 4. Or TCP_CORK is not set, and all sent packets are ACKed. |
| 1461 | * With Minshall's modification: all sent small packets are ACKed. |
| 1462 | */ |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1463 | static inline bool tcp_nagle_check(const struct tcp_sock *tp, |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 1464 | const struct sk_buff *skb, |
Eric Dumazet | 95c9617 | 2012-04-15 05:58:06 +0000 | [diff] [blame] | 1465 | unsigned int mss_now, int nonagle) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1466 | { |
Eric Dumazet | a02cec2 | 2010-09-22 20:43:57 +0000 | [diff] [blame] | 1467 | return skb->len < mss_now && |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1468 | ((nonagle & TCP_NAGLE_CORK) || |
Eric Dumazet | a02cec2 | 2010-09-22 20:43:57 +0000 | [diff] [blame] | 1469 | (!nonagle && tp->packets_out && tcp_minshall_check(tp))); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1470 | } |
| 1471 | |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1472 | /* Return true if the Nagle test allows this packet to be |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1473 | * sent now. |
| 1474 | */ |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1475 | static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, |
| 1476 | unsigned int cur_mss, int nonagle) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1477 | { |
| 1478 | /* Nagle rule does not apply to frames, which sit in the middle of the |
| 1479 | * write_queue (they have no chances to get new data). |
| 1480 | * |
| 1481 | * This is implemented in the callers, where they modify the 'nonagle' |
| 1482 | * argument based upon the location of SKB in the send queue. |
| 1483 | */ |
| 1484 | if (nonagle & TCP_NAGLE_PUSH) |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1485 | return true; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1486 | |
Yuchung Cheng | 9b44190 | 2013-03-20 13:32:58 +0000 | [diff] [blame] | 1487 | /* Don't use the nagle rule for urgent data (or for the final FIN). */ |
| 1488 | 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] | 1489 | return true; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1490 | |
| 1491 | if (!tcp_nagle_check(tp, skb, cur_mss, nonagle)) |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1492 | return true; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1493 | |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1494 | return false; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1495 | } |
| 1496 | |
| 1497 | /* 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] | 1498 | static bool tcp_snd_wnd_test(const struct tcp_sock *tp, |
| 1499 | const struct sk_buff *skb, |
| 1500 | unsigned int cur_mss) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1501 | { |
| 1502 | u32 end_seq = TCP_SKB_CB(skb)->end_seq; |
| 1503 | |
| 1504 | if (skb->len > cur_mss) |
| 1505 | end_seq = TCP_SKB_CB(skb)->seq + cur_mss; |
| 1506 | |
Ilpo Järvinen | 90840de | 2007-12-31 04:48:41 -0800 | [diff] [blame] | 1507 | return !after(end_seq, tcp_wnd_end(tp)); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1508 | } |
| 1509 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1510 | /* 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] | 1511 | * should be put on the wire right now. If so, it returns the number of |
| 1512 | * packets allowed by the congestion window. |
| 1513 | */ |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 1514 | 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] | 1515 | unsigned int cur_mss, int nonagle) |
| 1516 | { |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 1517 | const struct tcp_sock *tp = tcp_sk(sk); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1518 | unsigned int cwnd_quota; |
| 1519 | |
David S. Miller | 846998a | 2005-08-04 19:52:01 -0700 | [diff] [blame] | 1520 | tcp_init_tso_segs(sk, skb, cur_mss); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1521 | |
| 1522 | if (!tcp_nagle_test(tp, skb, cur_mss, nonagle)) |
| 1523 | return 0; |
| 1524 | |
| 1525 | cwnd_quota = tcp_cwnd_test(tp, skb); |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1526 | if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss)) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1527 | cwnd_quota = 0; |
| 1528 | |
| 1529 | return cwnd_quota; |
| 1530 | } |
| 1531 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 1532 | /* Test if sending is allowed right now. */ |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1533 | bool tcp_may_send_now(struct sock *sk) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1534 | { |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 1535 | const struct tcp_sock *tp = tcp_sk(sk); |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1536 | struct sk_buff *skb = tcp_send_head(sk); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1537 | |
Eric Dumazet | a02cec2 | 2010-09-22 20:43:57 +0000 | [diff] [blame] | 1538 | return skb && |
Ilpo Järvinen | 0c54b85 | 2009-03-14 14:23:05 +0000 | [diff] [blame] | 1539 | tcp_snd_test(sk, skb, tcp_current_mss(sk), |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1540 | (tcp_skb_is_last(sk, skb) ? |
Eric Dumazet | a02cec2 | 2010-09-22 20:43:57 +0000 | [diff] [blame] | 1541 | tp->nonagle : TCP_NAGLE_PUSH)); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1542 | } |
| 1543 | |
| 1544 | /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet |
| 1545 | * which is put after SKB on the list. It is very much like |
| 1546 | * tcp_fragment() except that it may make several kinds of assumptions |
| 1547 | * in order to speed up the splitting operation. In particular, we |
| 1548 | * know that all the data is in scatter-gather pages, and that the |
| 1549 | * packet has never been sent out before (and thus is not cloned). |
| 1550 | */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1551 | 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] | 1552 | unsigned int mss_now, gfp_t gfp) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1553 | { |
| 1554 | struct sk_buff *buff; |
| 1555 | int nlen = skb->len - len; |
Ilpo Järvinen | 9ce0146 | 2009-02-28 04:44:42 +0000 | [diff] [blame] | 1556 | u8 flags; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1557 | |
| 1558 | /* All of a TSO frame must be composed of paged data. */ |
Herbert Xu | c8ac377 | 2005-08-16 20:43:40 -0700 | [diff] [blame] | 1559 | if (skb->len != skb->data_len) |
| 1560 | return tcp_fragment(sk, skb, len, mss_now); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1561 | |
Eric Dumazet | c4ead4c | 2010-06-24 01:00:22 +0000 | [diff] [blame] | 1562 | buff = sk_stream_alloc_skb(sk, 0, gfp); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1563 | if (unlikely(buff == NULL)) |
| 1564 | return -ENOMEM; |
| 1565 | |
Hideo Aoki | 3ab224b | 2007-12-31 00:11:19 -0800 | [diff] [blame] | 1566 | sk->sk_wmem_queued += buff->truesize; |
| 1567 | sk_mem_charge(sk, buff->truesize); |
Herbert Xu | b60b49e | 2006-04-19 21:35:00 -0700 | [diff] [blame] | 1568 | buff->truesize += nlen; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1569 | skb->truesize -= nlen; |
| 1570 | |
| 1571 | /* Correct the sequence numbers. */ |
| 1572 | TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; |
| 1573 | TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; |
| 1574 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; |
| 1575 | |
| 1576 | /* PSH and FIN should only be set in the second packet. */ |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 1577 | flags = TCP_SKB_CB(skb)->tcp_flags; |
| 1578 | TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); |
| 1579 | TCP_SKB_CB(buff)->tcp_flags = flags; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1580 | |
| 1581 | /* This packet was never sent out yet, so no SACK bits. */ |
| 1582 | TCP_SKB_CB(buff)->sacked = 0; |
| 1583 | |
Patrick McHardy | 84fa793 | 2006-08-29 16:44:56 -0700 | [diff] [blame] | 1584 | buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1585 | skb_split(skb, buff, len); |
| 1586 | |
| 1587 | /* Fix up tso_factor for both original and new SKB. */ |
David S. Miller | 846998a | 2005-08-04 19:52:01 -0700 | [diff] [blame] | 1588 | tcp_set_skb_tso_segs(sk, skb, mss_now); |
| 1589 | tcp_set_skb_tso_segs(sk, buff, mss_now); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1590 | |
| 1591 | /* Link BUFF into the send queue. */ |
| 1592 | skb_header_release(buff); |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1593 | tcp_insert_write_queue_after(skb, buff, sk); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1594 | |
| 1595 | return 0; |
| 1596 | } |
| 1597 | |
| 1598 | /* Try to defer sending, if possible, in order to minimize the amount |
| 1599 | * of TSO splitting we do. View it as a kind of TSO Nagle test. |
| 1600 | * |
| 1601 | * This algorithm is from John Heffner. |
| 1602 | */ |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1603 | static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1604 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 1605 | struct tcp_sock *tp = tcp_sk(sk); |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 1606 | const struct inet_connection_sock *icsk = inet_csk(sk); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1607 | u32 send_win, cong_win, limit, in_flight; |
Eric Dumazet | ad9f4f5 | 2010-12-07 12:03:55 +0000 | [diff] [blame] | 1608 | int win_divisor; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1609 | |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 1610 | if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) |
John Heffner | ae8064a | 2006-10-18 20:36:48 -0700 | [diff] [blame] | 1611 | goto send_now; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1612 | |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 1613 | if (icsk->icsk_ca_state != TCP_CA_Open) |
John Heffner | ae8064a | 2006-10-18 20:36:48 -0700 | [diff] [blame] | 1614 | goto send_now; |
| 1615 | |
| 1616 | /* Defer for less than two clock ticks. */ |
Ilpo Järvinen | bd515c3 | 2007-12-20 20:36:03 -0800 | [diff] [blame] | 1617 | if (tp->tso_deferred && |
Ilpo Järvinen | a2acde0 | 2008-12-05 22:49:37 -0800 | [diff] [blame] | 1618 | (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1) |
John Heffner | ae8064a | 2006-10-18 20:36:48 -0700 | [diff] [blame] | 1619 | goto send_now; |
David S. Miller | 908a75c | 2005-07-05 15:43:58 -0700 | [diff] [blame] | 1620 | |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1621 | in_flight = tcp_packets_in_flight(tp); |
| 1622 | |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1623 | 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] | 1624 | |
Ilpo Järvinen | 90840de | 2007-12-31 04:48:41 -0800 | [diff] [blame] | 1625 | send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1626 | |
| 1627 | /* From in_flight test above, we know that cwnd > in_flight. */ |
| 1628 | cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; |
| 1629 | |
| 1630 | limit = min(send_win, cong_win); |
| 1631 | |
David S. Miller | ba244fe | 2006-03-11 18:51:49 -0800 | [diff] [blame] | 1632 | /* If a full-sized TSO skb can be sent, do it. */ |
Ben Hutchings | 1485348 | 2012-07-30 16:11:42 +0000 | [diff] [blame] | 1633 | if (limit >= min_t(unsigned int, sk->sk_gso_max_size, |
Eric Dumazet | 95bd09e | 2013-08-27 05:46:32 -0700 | [diff] [blame^] | 1634 | tp->xmit_size_goal_segs * tp->mss_cache)) |
John Heffner | ae8064a | 2006-10-18 20:36:48 -0700 | [diff] [blame] | 1635 | goto send_now; |
David S. Miller | ba244fe | 2006-03-11 18:51:49 -0800 | [diff] [blame] | 1636 | |
Ilpo Järvinen | 62ad276 | 2009-02-28 04:44:29 +0000 | [diff] [blame] | 1637 | /* Middle in queue won't get any more data, full sendable already? */ |
| 1638 | if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) |
| 1639 | goto send_now; |
| 1640 | |
Eric Dumazet | ad9f4f5 | 2010-12-07 12:03:55 +0000 | [diff] [blame] | 1641 | win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor); |
| 1642 | if (win_divisor) { |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1643 | u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); |
| 1644 | |
| 1645 | /* If at least some fraction of a window is available, |
| 1646 | * just use it. |
| 1647 | */ |
Eric Dumazet | ad9f4f5 | 2010-12-07 12:03:55 +0000 | [diff] [blame] | 1648 | chunk /= win_divisor; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1649 | if (limit >= chunk) |
John Heffner | ae8064a | 2006-10-18 20:36:48 -0700 | [diff] [blame] | 1650 | goto send_now; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1651 | } else { |
| 1652 | /* Different approach, try not to defer past a single |
| 1653 | * ACK. Receiver should ACK every other full sized |
| 1654 | * frame, so if we have space for more than 3 frames |
| 1655 | * then send now. |
| 1656 | */ |
Neal Cardwell | 6b5a5c0 | 2011-11-21 17:15:14 +0000 | [diff] [blame] | 1657 | if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) |
John Heffner | ae8064a | 2006-10-18 20:36:48 -0700 | [diff] [blame] | 1658 | goto send_now; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1659 | } |
| 1660 | |
Eric Dumazet | f4541d6 | 2013-03-21 17:36:09 +0000 | [diff] [blame] | 1661 | /* Ok, it looks like it is advisable to defer. |
| 1662 | * Do not rearm the timer if already set to not break TCP ACK clocking. |
| 1663 | */ |
| 1664 | if (!tp->tso_deferred) |
| 1665 | tp->tso_deferred = 1 | (jiffies << 1); |
John Heffner | ae8064a | 2006-10-18 20:36:48 -0700 | [diff] [blame] | 1666 | |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1667 | return true; |
John Heffner | ae8064a | 2006-10-18 20:36:48 -0700 | [diff] [blame] | 1668 | |
| 1669 | send_now: |
| 1670 | tp->tso_deferred = 0; |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1671 | return false; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1672 | } |
| 1673 | |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1674 | /* Create a new MTU probe if we are ready. |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 1675 | * MTU probe is regularly attempting to increase the path MTU by |
| 1676 | * deliberately sending larger packets. This discovers routing |
| 1677 | * changes resulting in larger path MTUs. |
| 1678 | * |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1679 | * Returns 0 if we should wait to probe (no cwnd available), |
| 1680 | * 1 if a probe was sent, |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1681 | * -1 otherwise |
| 1682 | */ |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1683 | static int tcp_mtu_probe(struct sock *sk) |
| 1684 | { |
| 1685 | struct tcp_sock *tp = tcp_sk(sk); |
| 1686 | struct inet_connection_sock *icsk = inet_csk(sk); |
| 1687 | struct sk_buff *skb, *nskb, *next; |
| 1688 | int len; |
| 1689 | int probe_size; |
Ilpo Järvinen | 91cc17c | 2007-11-23 19:08:16 +0800 | [diff] [blame] | 1690 | int size_needed; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1691 | int copy; |
| 1692 | int mss_now; |
| 1693 | |
| 1694 | /* Not currently probing/verifying, |
| 1695 | * not in recovery, |
| 1696 | * have enough cwnd, and |
| 1697 | * not SACKing (the variable headers throw things off) */ |
| 1698 | if (!icsk->icsk_mtup.enabled || |
| 1699 | icsk->icsk_mtup.probe_size || |
| 1700 | inet_csk(sk)->icsk_ca_state != TCP_CA_Open || |
| 1701 | tp->snd_cwnd < 11 || |
Ilpo Järvinen | cabeccb | 2009-02-28 04:44:38 +0000 | [diff] [blame] | 1702 | tp->rx_opt.num_sacks || tp->rx_opt.dsack) |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1703 | return -1; |
| 1704 | |
| 1705 | /* Very simple search strategy: just double the MSS. */ |
Ilpo Järvinen | 0c54b85 | 2009-03-14 14:23:05 +0000 | [diff] [blame] | 1706 | mss_now = tcp_current_mss(sk); |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1707 | probe_size = 2 * tp->mss_cache; |
Ilpo Järvinen | 91cc17c | 2007-11-23 19:08:16 +0800 | [diff] [blame] | 1708 | size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1709 | if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) { |
| 1710 | /* TODO: set timer for probe_converge_event */ |
| 1711 | return -1; |
| 1712 | } |
| 1713 | |
| 1714 | /* Have enough data in the send queue to probe? */ |
Ilpo Järvinen | 7f9c33e | 2007-11-23 19:10:56 +0800 | [diff] [blame] | 1715 | if (tp->write_seq - tp->snd_nxt < size_needed) |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1716 | return -1; |
| 1717 | |
Ilpo Järvinen | 91cc17c | 2007-11-23 19:08:16 +0800 | [diff] [blame] | 1718 | if (tp->snd_wnd < size_needed) |
| 1719 | return -1; |
Ilpo Järvinen | 90840de | 2007-12-31 04:48:41 -0800 | [diff] [blame] | 1720 | if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) |
Ilpo Järvinen | 91cc17c | 2007-11-23 19:08:16 +0800 | [diff] [blame] | 1721 | return 0; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1722 | |
Ilpo Järvinen | d67c58e | 2007-12-02 00:48:01 +0200 | [diff] [blame] | 1723 | /* Do we need to wait to drain cwnd? With none in flight, don't stall */ |
| 1724 | if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { |
| 1725 | if (!tcp_packets_in_flight(tp)) |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1726 | return -1; |
| 1727 | else |
| 1728 | return 0; |
| 1729 | } |
| 1730 | |
| 1731 | /* We're allowed to probe. Build it now. */ |
| 1732 | if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL) |
| 1733 | return -1; |
Hideo Aoki | 3ab224b | 2007-12-31 00:11:19 -0800 | [diff] [blame] | 1734 | sk->sk_wmem_queued += nskb->truesize; |
| 1735 | sk_mem_charge(sk, nskb->truesize); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1736 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1737 | skb = tcp_send_head(sk); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1738 | |
| 1739 | TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; |
| 1740 | 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] | 1741 | TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1742 | TCP_SKB_CB(nskb)->sacked = 0; |
| 1743 | nskb->csum = 0; |
Patrick McHardy | 84fa793 | 2006-08-29 16:44:56 -0700 | [diff] [blame] | 1744 | nskb->ip_summed = skb->ip_summed; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1745 | |
Ilpo Järvinen | 50c4817 | 2007-12-02 00:48:00 +0200 | [diff] [blame] | 1746 | tcp_insert_write_queue_before(nskb, skb, sk); |
| 1747 | |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1748 | len = 0; |
Ilpo Järvinen | 234b686 | 2007-12-02 00:48:02 +0200 | [diff] [blame] | 1749 | tcp_for_write_queue_from_safe(skb, next, sk) { |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1750 | copy = min_t(int, skb->len, probe_size - len); |
| 1751 | if (nskb->ip_summed) |
| 1752 | skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); |
| 1753 | else |
| 1754 | nskb->csum = skb_copy_and_csum_bits(skb, 0, |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1755 | skb_put(nskb, copy), |
| 1756 | copy, nskb->csum); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1757 | |
| 1758 | if (skb->len <= copy) { |
| 1759 | /* We've eaten all the data from this skb. |
| 1760 | * Throw it away. */ |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 1761 | 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] | 1762 | tcp_unlink_write_queue(skb, sk); |
Hideo Aoki | 3ab224b | 2007-12-31 00:11:19 -0800 | [diff] [blame] | 1763 | sk_wmem_free_skb(sk, skb); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1764 | } else { |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 1765 | TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & |
Changli Gao | a3433f3 | 2010-06-12 14:01:43 +0000 | [diff] [blame] | 1766 | ~(TCPHDR_FIN|TCPHDR_PSH); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1767 | if (!skb_shinfo(skb)->nr_frags) { |
| 1768 | skb_pull(skb, copy); |
Patrick McHardy | 84fa793 | 2006-08-29 16:44:56 -0700 | [diff] [blame] | 1769 | if (skb->ip_summed != CHECKSUM_PARTIAL) |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1770 | skb->csum = csum_partial(skb->data, |
| 1771 | skb->len, 0); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1772 | } else { |
| 1773 | __pskb_trim_head(skb, copy); |
| 1774 | tcp_set_skb_tso_segs(sk, skb, mss_now); |
| 1775 | } |
| 1776 | TCP_SKB_CB(skb)->seq += copy; |
| 1777 | } |
| 1778 | |
| 1779 | len += copy; |
Ilpo Järvinen | 234b686 | 2007-12-02 00:48:02 +0200 | [diff] [blame] | 1780 | |
| 1781 | if (len >= probe_size) |
| 1782 | break; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1783 | } |
| 1784 | tcp_init_tso_segs(sk, nskb, nskb->len); |
| 1785 | |
| 1786 | /* We're ready to send. If this fails, the probe will |
| 1787 | * be resegmented into mss-sized pieces by tcp_write_xmit(). */ |
| 1788 | TCP_SKB_CB(nskb)->when = tcp_time_stamp; |
| 1789 | if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { |
| 1790 | /* Decrement cwnd here because we are sending |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1791 | * effectively two packets. */ |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1792 | tp->snd_cwnd--; |
Ilpo Järvinen | 66f5fe6 | 2007-12-31 04:43:57 -0800 | [diff] [blame] | 1793 | tcp_event_new_data_sent(sk, nskb); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1794 | |
| 1795 | icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); |
John Heffner | 0e7b136 | 2006-03-20 21:32:58 -0800 | [diff] [blame] | 1796 | tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; |
| 1797 | tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1798 | |
| 1799 | return 1; |
| 1800 | } |
| 1801 | |
| 1802 | return -1; |
| 1803 | } |
| 1804 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1805 | /* This routine writes packets to the network. It advances the |
| 1806 | * send_head. This happens as incoming acks open up the remote |
| 1807 | * window for us. |
| 1808 | * |
Ilpo Järvinen | f8269a4 | 2008-12-03 21:24:48 -0800 | [diff] [blame] | 1809 | * LARGESEND note: !tcp_urg_mode is overkill, only frames between |
| 1810 | * snd_up-64k-mss .. snd_up cannot be large. However, taking into |
| 1811 | * account rare use of URG, this is not a big flaw. |
| 1812 | * |
Nandita Dukkipati | 6ba8a3b | 2013-03-11 10:00:43 +0000 | [diff] [blame] | 1813 | * Send at most one packet when push_one > 0. Temporarily ignore |
| 1814 | * cwnd limit to force at most one packet out when push_one == 2. |
| 1815 | |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1816 | * Returns true, if no segments are in flight and we have queued segments, |
| 1817 | * but cannot send anything now because of SWS or another problem. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1818 | */ |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1819 | static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, |
| 1820 | int push_one, gfp_t gfp) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1821 | { |
| 1822 | struct tcp_sock *tp = tcp_sk(sk); |
David S. Miller | 92df7b5 | 2005-07-05 15:19:06 -0700 | [diff] [blame] | 1823 | struct sk_buff *skb; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1824 | unsigned int tso_segs, sent_pkts; |
| 1825 | int cwnd_quota; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1826 | int result; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1827 | |
David S. Miller | 92df7b5 | 2005-07-05 15:19:06 -0700 | [diff] [blame] | 1828 | sent_pkts = 0; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1829 | |
Ilpo Järvinen | d5dd917 | 2008-12-05 22:48:55 -0800 | [diff] [blame] | 1830 | if (!push_one) { |
| 1831 | /* Do MTU probing. */ |
| 1832 | result = tcp_mtu_probe(sk); |
| 1833 | if (!result) { |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1834 | return false; |
Ilpo Järvinen | d5dd917 | 2008-12-05 22:48:55 -0800 | [diff] [blame] | 1835 | } else if (result > 0) { |
| 1836 | sent_pkts = 1; |
| 1837 | } |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1838 | } |
| 1839 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1840 | while ((skb = tcp_send_head(sk))) { |
Herbert Xu | c8ac377 | 2005-08-16 20:43:40 -0700 | [diff] [blame] | 1841 | unsigned int limit; |
| 1842 | |
Eric Dumazet | 46d3cea | 2012-07-11 05:50:31 +0000 | [diff] [blame] | 1843 | |
Herbert Xu | b68e9f8 | 2005-08-04 19:52:02 -0700 | [diff] [blame] | 1844 | tso_segs = tcp_init_tso_segs(sk, skb, mss_now); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1845 | BUG_ON(!tso_segs); |
David S. Miller | aa93466 | 2005-07-05 15:20:09 -0700 | [diff] [blame] | 1846 | |
Andrew Vagin | ec34232 | 2012-11-15 04:03:17 +0000 | [diff] [blame] | 1847 | if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) |
| 1848 | goto repair; /* Skip network transmission */ |
| 1849 | |
Herbert Xu | b68e9f8 | 2005-08-04 19:52:02 -0700 | [diff] [blame] | 1850 | cwnd_quota = tcp_cwnd_test(tp, skb); |
Nandita Dukkipati | 6ba8a3b | 2013-03-11 10:00:43 +0000 | [diff] [blame] | 1851 | if (!cwnd_quota) { |
| 1852 | if (push_one == 2) |
| 1853 | /* Force out a loss probe pkt. */ |
| 1854 | cwnd_quota = 1; |
| 1855 | else |
| 1856 | break; |
| 1857 | } |
Herbert Xu | b68e9f8 | 2005-08-04 19:52:02 -0700 | [diff] [blame] | 1858 | |
| 1859 | if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) |
| 1860 | break; |
| 1861 | |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1862 | if (tso_segs == 1) { |
| 1863 | if (unlikely(!tcp_nagle_test(tp, skb, mss_now, |
| 1864 | (tcp_skb_is_last(sk, skb) ? |
| 1865 | nonagle : TCP_NAGLE_PUSH)))) |
| 1866 | break; |
| 1867 | } else { |
Ilpo Järvinen | d5dd917 | 2008-12-05 22:48:55 -0800 | [diff] [blame] | 1868 | if (!push_one && tcp_tso_should_defer(sk, skb)) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1869 | break; |
| 1870 | } |
David S. Miller | aa93466 | 2005-07-05 15:20:09 -0700 | [diff] [blame] | 1871 | |
Eric Dumazet | 46d3cea | 2012-07-11 05:50:31 +0000 | [diff] [blame] | 1872 | /* TSQ : sk_wmem_alloc accounts skb truesize, |
| 1873 | * including skb overhead. But thats OK. |
| 1874 | */ |
| 1875 | if (atomic_read(&sk->sk_wmem_alloc) >= sysctl_tcp_limit_output_bytes) { |
| 1876 | set_bit(TSQ_THROTTLED, &tp->tsq_flags); |
| 1877 | break; |
| 1878 | } |
Herbert Xu | c8ac377 | 2005-08-16 20:43:40 -0700 | [diff] [blame] | 1879 | limit = mss_now; |
Ilpo Järvinen | f8269a4 | 2008-12-03 21:24:48 -0800 | [diff] [blame] | 1880 | if (tso_segs > 1 && !tcp_urg_mode(tp)) |
Ilpo Järvinen | 0e3a480 | 2007-12-24 21:33:45 -0800 | [diff] [blame] | 1881 | limit = tcp_mss_split_point(sk, skb, mss_now, |
Ben Hutchings | 1485348 | 2012-07-30 16:11:42 +0000 | [diff] [blame] | 1882 | min_t(unsigned int, |
| 1883 | cwnd_quota, |
| 1884 | sk->sk_gso_max_segs)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1885 | |
Herbert Xu | c8ac377 | 2005-08-16 20:43:40 -0700 | [diff] [blame] | 1886 | if (skb->len > limit && |
Eric Dumazet | c4ead4c | 2010-06-24 01:00:22 +0000 | [diff] [blame] | 1887 | unlikely(tso_fragment(sk, skb, limit, mss_now, gfp))) |
Herbert Xu | c8ac377 | 2005-08-16 20:43:40 -0700 | [diff] [blame] | 1888 | break; |
| 1889 | |
David S. Miller | 92df7b5 | 2005-07-05 15:19:06 -0700 | [diff] [blame] | 1890 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1891 | |
Ilpo Järvinen | d5dd917 | 2008-12-05 22:48:55 -0800 | [diff] [blame] | 1892 | if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) |
David S. Miller | 92df7b5 | 2005-07-05 15:19:06 -0700 | [diff] [blame] | 1893 | break; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1894 | |
Andrew Vagin | ec34232 | 2012-11-15 04:03:17 +0000 | [diff] [blame] | 1895 | repair: |
David S. Miller | 92df7b5 | 2005-07-05 15:19:06 -0700 | [diff] [blame] | 1896 | /* Advance the send_head. This one is sent out. |
| 1897 | * This call will increment packets_out. |
| 1898 | */ |
Ilpo Järvinen | 66f5fe6 | 2007-12-31 04:43:57 -0800 | [diff] [blame] | 1899 | tcp_event_new_data_sent(sk, skb); |
David S. Miller | 92df7b5 | 2005-07-05 15:19:06 -0700 | [diff] [blame] | 1900 | |
| 1901 | tcp_minshall_update(tp, mss_now, skb); |
Nandita Dukkipati | a262f0c | 2011-08-21 20:21:57 +0000 | [diff] [blame] | 1902 | sent_pkts += tcp_skb_pcount(skb); |
Ilpo Järvinen | d5dd917 | 2008-12-05 22:48:55 -0800 | [diff] [blame] | 1903 | |
| 1904 | if (push_one) |
| 1905 | break; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1906 | } |
David S. Miller | 92df7b5 | 2005-07-05 15:19:06 -0700 | [diff] [blame] | 1907 | |
David S. Miller | aa93466 | 2005-07-05 15:20:09 -0700 | [diff] [blame] | 1908 | if (likely(sent_pkts)) { |
Yuchung Cheng | 684bad1 | 2012-09-02 17:38:04 +0000 | [diff] [blame] | 1909 | if (tcp_in_cwnd_reduction(sk)) |
| 1910 | tp->prr_out += sent_pkts; |
Nandita Dukkipati | 6ba8a3b | 2013-03-11 10:00:43 +0000 | [diff] [blame] | 1911 | |
| 1912 | /* Send one loss probe per tail loss episode. */ |
| 1913 | if (push_one != 2) |
| 1914 | tcp_schedule_loss_probe(sk); |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 1915 | tcp_cwnd_validate(sk); |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1916 | return false; |
David S. Miller | 92df7b5 | 2005-07-05 15:19:06 -0700 | [diff] [blame] | 1917 | } |
Nandita Dukkipati | 6ba8a3b | 2013-03-11 10:00:43 +0000 | [diff] [blame] | 1918 | return (push_one == 2) || (!tp->packets_out && tcp_send_head(sk)); |
| 1919 | } |
| 1920 | |
| 1921 | bool tcp_schedule_loss_probe(struct sock *sk) |
| 1922 | { |
| 1923 | struct inet_connection_sock *icsk = inet_csk(sk); |
| 1924 | struct tcp_sock *tp = tcp_sk(sk); |
| 1925 | u32 timeout, tlp_time_stamp, rto_time_stamp; |
| 1926 | u32 rtt = tp->srtt >> 3; |
| 1927 | |
| 1928 | if (WARN_ON(icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS)) |
| 1929 | return false; |
| 1930 | /* No consecutive loss probes. */ |
| 1931 | if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) { |
| 1932 | tcp_rearm_rto(sk); |
| 1933 | return false; |
| 1934 | } |
| 1935 | /* Don't do any loss probe on a Fast Open connection before 3WHS |
| 1936 | * finishes. |
| 1937 | */ |
| 1938 | if (sk->sk_state == TCP_SYN_RECV) |
| 1939 | return false; |
| 1940 | |
| 1941 | /* TLP is only scheduled when next timer event is RTO. */ |
| 1942 | if (icsk->icsk_pending != ICSK_TIME_RETRANS) |
| 1943 | return false; |
| 1944 | |
| 1945 | /* Schedule a loss probe in 2*RTT for SACK capable connections |
| 1946 | * in Open state, that are either limited by cwnd or application. |
| 1947 | */ |
| 1948 | if (sysctl_tcp_early_retrans < 3 || !rtt || !tp->packets_out || |
| 1949 | !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open) |
| 1950 | return false; |
| 1951 | |
| 1952 | if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) && |
| 1953 | tcp_send_head(sk)) |
| 1954 | return false; |
| 1955 | |
| 1956 | /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account |
| 1957 | * for delayed ack when there's one outstanding packet. |
| 1958 | */ |
| 1959 | timeout = rtt << 1; |
| 1960 | if (tp->packets_out == 1) |
| 1961 | timeout = max_t(u32, timeout, |
| 1962 | (rtt + (rtt >> 1) + TCP_DELACK_MAX)); |
| 1963 | timeout = max_t(u32, timeout, msecs_to_jiffies(10)); |
| 1964 | |
| 1965 | /* If RTO is shorter, just schedule TLP in its place. */ |
| 1966 | tlp_time_stamp = tcp_time_stamp + timeout; |
| 1967 | rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout; |
| 1968 | if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) { |
| 1969 | s32 delta = rto_time_stamp - tcp_time_stamp; |
| 1970 | if (delta > 0) |
| 1971 | timeout = delta; |
| 1972 | } |
| 1973 | |
| 1974 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, |
| 1975 | TCP_RTO_MAX); |
| 1976 | return true; |
| 1977 | } |
| 1978 | |
| 1979 | /* When probe timeout (PTO) fires, send a new segment if one exists, else |
| 1980 | * retransmit the last segment. |
| 1981 | */ |
| 1982 | void tcp_send_loss_probe(struct sock *sk) |
| 1983 | { |
Nandita Dukkipati | 9b717a8 | 2013-03-11 10:00:44 +0000 | [diff] [blame] | 1984 | struct tcp_sock *tp = tcp_sk(sk); |
Nandita Dukkipati | 6ba8a3b | 2013-03-11 10:00:43 +0000 | [diff] [blame] | 1985 | struct sk_buff *skb; |
| 1986 | int pcount; |
| 1987 | int mss = tcp_current_mss(sk); |
| 1988 | int err = -1; |
| 1989 | |
| 1990 | if (tcp_send_head(sk) != NULL) { |
| 1991 | err = tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC); |
| 1992 | goto rearm_timer; |
| 1993 | } |
| 1994 | |
Nandita Dukkipati | 9b717a8 | 2013-03-11 10:00:44 +0000 | [diff] [blame] | 1995 | /* At most one outstanding TLP retransmission. */ |
| 1996 | if (tp->tlp_high_seq) |
| 1997 | goto rearm_timer; |
| 1998 | |
Nandita Dukkipati | 6ba8a3b | 2013-03-11 10:00:43 +0000 | [diff] [blame] | 1999 | /* Retransmit last segment. */ |
| 2000 | skb = tcp_write_queue_tail(sk); |
| 2001 | if (WARN_ON(!skb)) |
| 2002 | goto rearm_timer; |
| 2003 | |
| 2004 | pcount = tcp_skb_pcount(skb); |
| 2005 | if (WARN_ON(!pcount)) |
| 2006 | goto rearm_timer; |
| 2007 | |
| 2008 | if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) { |
| 2009 | if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss))) |
| 2010 | goto rearm_timer; |
| 2011 | skb = tcp_write_queue_tail(sk); |
| 2012 | } |
| 2013 | |
| 2014 | if (WARN_ON(!skb || !tcp_skb_pcount(skb))) |
| 2015 | goto rearm_timer; |
| 2016 | |
| 2017 | /* Probe with zero data doesn't trigger fast recovery. */ |
| 2018 | if (skb->len > 0) |
| 2019 | err = __tcp_retransmit_skb(sk, skb); |
| 2020 | |
Nandita Dukkipati | 9b717a8 | 2013-03-11 10:00:44 +0000 | [diff] [blame] | 2021 | /* Record snd_nxt for loss detection. */ |
| 2022 | if (likely(!err)) |
| 2023 | tp->tlp_high_seq = tp->snd_nxt; |
| 2024 | |
Nandita Dukkipati | 6ba8a3b | 2013-03-11 10:00:43 +0000 | [diff] [blame] | 2025 | rearm_timer: |
| 2026 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, |
| 2027 | inet_csk(sk)->icsk_rto, |
| 2028 | TCP_RTO_MAX); |
| 2029 | |
| 2030 | if (likely(!err)) |
| 2031 | NET_INC_STATS_BH(sock_net(sk), |
| 2032 | LINUX_MIB_TCPLOSSPROBES); |
| 2033 | return; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2034 | } |
| 2035 | |
David S. Miller | a762a98 | 2005-07-05 15:18:51 -0700 | [diff] [blame] | 2036 | /* Push out any pending frames which were held back due to |
| 2037 | * TCP_CORK or attempt at coalescing tiny packets. |
| 2038 | * The socket must be locked by the caller. |
| 2039 | */ |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2040 | void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, |
| 2041 | int nonagle) |
David S. Miller | a762a98 | 2005-07-05 15:18:51 -0700 | [diff] [blame] | 2042 | { |
Ilpo Järvinen | 726e07a | 2008-12-05 22:43:56 -0800 | [diff] [blame] | 2043 | /* If we are closed, the bytes will have to remain here. |
| 2044 | * In time closedown will finish, we empty the write queue and |
| 2045 | * all will be happy. |
| 2046 | */ |
| 2047 | if (unlikely(sk->sk_state == TCP_CLOSE)) |
| 2048 | return; |
| 2049 | |
Mel Gorman | 99a1dec | 2012-07-31 16:44:14 -0700 | [diff] [blame] | 2050 | if (tcp_write_xmit(sk, cur_mss, nonagle, 0, |
| 2051 | sk_gfp_atomic(sk, GFP_ATOMIC))) |
Ilpo Järvinen | 726e07a | 2008-12-05 22:43:56 -0800 | [diff] [blame] | 2052 | tcp_check_probe_timer(sk); |
David S. Miller | a762a98 | 2005-07-05 15:18:51 -0700 | [diff] [blame] | 2053 | } |
| 2054 | |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 2055 | /* Send _single_ skb sitting at the send head. This function requires |
| 2056 | * true push pending frames to setup probe timer etc. |
| 2057 | */ |
| 2058 | void tcp_push_one(struct sock *sk, unsigned int mss_now) |
| 2059 | { |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2060 | struct sk_buff *skb = tcp_send_head(sk); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 2061 | |
| 2062 | BUG_ON(!skb || skb->len < mss_now); |
| 2063 | |
Ilpo Järvinen | d5dd917 | 2008-12-05 22:48:55 -0800 | [diff] [blame] | 2064 | 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] | 2065 | } |
| 2066 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2067 | /* This function returns the amount that we can raise the |
| 2068 | * usable window based on the following constraints |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2069 | * |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2070 | * 1. The window can never be shrunk once it is offered (RFC 793) |
| 2071 | * 2. We limit memory per socket |
| 2072 | * |
| 2073 | * RFC 1122: |
| 2074 | * "the suggested [SWS] avoidance algorithm for the receiver is to keep |
| 2075 | * RECV.NEXT + RCV.WIN fixed until: |
| 2076 | * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" |
| 2077 | * |
| 2078 | * i.e. don't raise the right edge of the window until you can raise |
| 2079 | * it at least MSS bytes. |
| 2080 | * |
| 2081 | * Unfortunately, the recommended algorithm breaks header prediction, |
| 2082 | * since header prediction assumes th->window stays fixed. |
| 2083 | * |
| 2084 | * Strictly speaking, keeping th->window fixed violates the receiver |
| 2085 | * side SWS prevention criteria. The problem is that under this rule |
| 2086 | * a stream of single byte packets will cause the right side of the |
| 2087 | * window to always advance by a single byte. |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2088 | * |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2089 | * Of course, if the sender implements sender side SWS prevention |
| 2090 | * then this will not be a problem. |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2091 | * |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2092 | * BSD seems to make the following compromise: |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2093 | * |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2094 | * If the free space is less than the 1/4 of the maximum |
| 2095 | * space available and the free space is less than 1/2 mss, |
| 2096 | * then set the window to 0. |
| 2097 | * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] |
| 2098 | * Otherwise, just prevent the window from shrinking |
| 2099 | * and from being larger than the largest representable value. |
| 2100 | * |
| 2101 | * This prevents incremental opening of the window in the regime |
| 2102 | * where TCP is limited by the speed of the reader side taking |
| 2103 | * data out of the TCP receive queue. It does nothing about |
| 2104 | * those cases where the window is constrained on the sender side |
| 2105 | * because the pipeline is full. |
| 2106 | * |
| 2107 | * BSD also seems to "accidentally" limit itself to windows that are a |
| 2108 | * multiple of MSS, at least until the free space gets quite small. |
| 2109 | * This would appear to be a side effect of the mbuf implementation. |
| 2110 | * Combining these two algorithms results in the observed behavior |
| 2111 | * of having a fixed window size at almost all times. |
| 2112 | * |
| 2113 | * Below we obtain similar behavior by forcing the offered window to |
| 2114 | * a multiple of the mss when it is feasible to do so. |
| 2115 | * |
| 2116 | * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. |
| 2117 | * Regular options like TIMESTAMP are taken into account. |
| 2118 | */ |
| 2119 | u32 __tcp_select_window(struct sock *sk) |
| 2120 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2121 | struct inet_connection_sock *icsk = inet_csk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2122 | struct tcp_sock *tp = tcp_sk(sk); |
Stephen Hemminger | caa20d9a | 2005-11-10 17:13:47 -0800 | [diff] [blame] | 2123 | /* MSS for the peer's data. Previous versions used mss_clamp |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2124 | * here. I don't know if the value based on our guesses |
| 2125 | * of peer's MSS is better for the performance. It's more correct |
| 2126 | * but may be worse for the performance because of rcv_mss |
| 2127 | * fluctuations. --SAW 1998/11/1 |
| 2128 | */ |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2129 | int mss = icsk->icsk_ack.rcv_mss; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2130 | int free_space = tcp_space(sk); |
| 2131 | int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk)); |
| 2132 | int window; |
| 2133 | |
| 2134 | if (mss > full_space) |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2135 | mss = full_space; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2136 | |
Eric Dumazet | b92edbe | 2007-12-20 21:48:32 -0800 | [diff] [blame] | 2137 | if (free_space < (full_space >> 1)) { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2138 | icsk->icsk_ack.quick = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2139 | |
Glauber Costa | 180d8cd | 2011-12-11 21:47:02 +0000 | [diff] [blame] | 2140 | if (sk_under_memory_pressure(sk)) |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 2141 | tp->rcv_ssthresh = min(tp->rcv_ssthresh, |
| 2142 | 4U * tp->advmss); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2143 | |
| 2144 | if (free_space < mss) |
| 2145 | return 0; |
| 2146 | } |
| 2147 | |
| 2148 | if (free_space > tp->rcv_ssthresh) |
| 2149 | free_space = tp->rcv_ssthresh; |
| 2150 | |
| 2151 | /* Don't do rounding if we are using window scaling, since the |
| 2152 | * scaled window will not line up with the MSS boundary anyway. |
| 2153 | */ |
| 2154 | window = tp->rcv_wnd; |
| 2155 | if (tp->rx_opt.rcv_wscale) { |
| 2156 | window = free_space; |
| 2157 | |
| 2158 | /* Advertise enough space so that it won't get scaled away. |
| 2159 | * Import case: prevent zero window announcement if |
| 2160 | * 1<<rcv_wscale > mss. |
| 2161 | */ |
| 2162 | if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window) |
| 2163 | window = (((window >> tp->rx_opt.rcv_wscale) + 1) |
| 2164 | << tp->rx_opt.rcv_wscale); |
| 2165 | } else { |
| 2166 | /* Get the largest window that is a nice multiple of mss. |
| 2167 | * Window clamp already applied above. |
| 2168 | * If our current window offering is within 1 mss of the |
| 2169 | * free space we just keep it. This prevents the divide |
| 2170 | * and multiply from happening most of the time. |
| 2171 | * We also don't do any window rounding when the free space |
| 2172 | * is too small. |
| 2173 | */ |
| 2174 | if (window <= free_space - mss || window > free_space) |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 2175 | window = (free_space / mss) * mss; |
John Heffner | 8456507 | 2007-04-02 13:56:32 -0700 | [diff] [blame] | 2176 | else if (mss == full_space && |
Eric Dumazet | b92edbe | 2007-12-20 21:48:32 -0800 | [diff] [blame] | 2177 | free_space > window + (full_space >> 1)) |
John Heffner | 8456507 | 2007-04-02 13:56:32 -0700 | [diff] [blame] | 2178 | window = free_space; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2179 | } |
| 2180 | |
| 2181 | return window; |
| 2182 | } |
| 2183 | |
Ilpo Järvinen | 4a17fc3 | 2008-11-24 21:03:43 -0800 | [diff] [blame] | 2184 | /* Collapses two adjacent SKB's during retransmission. */ |
| 2185 | static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2186 | { |
| 2187 | struct tcp_sock *tp = tcp_sk(sk); |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2188 | struct sk_buff *next_skb = tcp_write_queue_next(sk, skb); |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2189 | int skb_size, next_skb_size; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2190 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2191 | skb_size = skb->len; |
| 2192 | next_skb_size = next_skb->len; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2193 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2194 | 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] | 2195 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2196 | tcp_highest_sack_combine(sk, next_skb, skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2197 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2198 | tcp_unlink_write_queue(next_skb, sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2199 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2200 | skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size), |
| 2201 | next_skb_size); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2202 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2203 | if (next_skb->ip_summed == CHECKSUM_PARTIAL) |
| 2204 | skb->ip_summed = CHECKSUM_PARTIAL; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2205 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2206 | if (skb->ip_summed != CHECKSUM_PARTIAL) |
| 2207 | skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2208 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2209 | /* Update sequence range on original skb. */ |
| 2210 | 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] | 2211 | |
Ilpo Järvinen | e6c7d08 | 2009-02-28 04:44:35 +0000 | [diff] [blame] | 2212 | /* Merge over control information. This moves PSH/FIN etc. over */ |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 2213 | 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] | 2214 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2215 | /* All done, get rid of second SKB and account for it so |
| 2216 | * packet counting does not break. |
| 2217 | */ |
| 2218 | TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; |
Ilpo Järvinen | b768920 | 2007-09-20 11:37:19 -0700 | [diff] [blame] | 2219 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2220 | /* changed transmit queue under us so clear hints */ |
Ilpo Järvinen | ef9da47 | 2008-09-20 21:25:15 -0700 | [diff] [blame] | 2221 | tcp_clear_retrans_hints_partial(tp); |
| 2222 | if (next_skb == tp->retransmit_skb_hint) |
| 2223 | tp->retransmit_skb_hint = skb; |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2224 | |
Ilpo Järvinen | 797108d | 2009-04-01 23:15:17 +0000 | [diff] [blame] | 2225 | tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); |
| 2226 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2227 | sk_wmem_free_skb(sk, next_skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2228 | } |
| 2229 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 2230 | /* Check if coalescing SKBs is legal. */ |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 2231 | 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] | 2232 | { |
| 2233 | if (tcp_skb_pcount(skb) > 1) |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 2234 | return false; |
Ilpo Järvinen | 4a17fc3 | 2008-11-24 21:03:43 -0800 | [diff] [blame] | 2235 | /* TODO: SACK collapsing could be used to remove this condition */ |
| 2236 | if (skb_shinfo(skb)->nr_frags != 0) |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 2237 | return false; |
Ilpo Järvinen | 4a17fc3 | 2008-11-24 21:03:43 -0800 | [diff] [blame] | 2238 | if (skb_cloned(skb)) |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 2239 | return false; |
Ilpo Järvinen | 4a17fc3 | 2008-11-24 21:03:43 -0800 | [diff] [blame] | 2240 | if (skb == tcp_send_head(sk)) |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 2241 | return false; |
Ilpo Järvinen | 4a17fc3 | 2008-11-24 21:03:43 -0800 | [diff] [blame] | 2242 | /* Some heurestics for collapsing over SACK'd could be invented */ |
| 2243 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 2244 | return false; |
Ilpo Järvinen | 4a17fc3 | 2008-11-24 21:03:43 -0800 | [diff] [blame] | 2245 | |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 2246 | return true; |
Ilpo Järvinen | 4a17fc3 | 2008-11-24 21:03:43 -0800 | [diff] [blame] | 2247 | } |
| 2248 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 2249 | /* Collapse packets in the retransmit queue to make to create |
| 2250 | * less packets on the wire. This is only done on retransmission. |
| 2251 | */ |
Ilpo Järvinen | 4a17fc3 | 2008-11-24 21:03:43 -0800 | [diff] [blame] | 2252 | static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, |
| 2253 | int space) |
| 2254 | { |
| 2255 | struct tcp_sock *tp = tcp_sk(sk); |
| 2256 | struct sk_buff *skb = to, *tmp; |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 2257 | bool first = true; |
Ilpo Järvinen | 4a17fc3 | 2008-11-24 21:03:43 -0800 | [diff] [blame] | 2258 | |
| 2259 | if (!sysctl_tcp_retrans_collapse) |
| 2260 | return; |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 2261 | if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) |
Ilpo Järvinen | 4a17fc3 | 2008-11-24 21:03:43 -0800 | [diff] [blame] | 2262 | return; |
| 2263 | |
| 2264 | tcp_for_write_queue_from_safe(skb, tmp, sk) { |
| 2265 | if (!tcp_can_collapse(sk, skb)) |
| 2266 | break; |
| 2267 | |
| 2268 | space -= skb->len; |
| 2269 | |
| 2270 | if (first) { |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 2271 | first = false; |
Ilpo Järvinen | 4a17fc3 | 2008-11-24 21:03:43 -0800 | [diff] [blame] | 2272 | continue; |
| 2273 | } |
| 2274 | |
| 2275 | if (space < 0) |
| 2276 | break; |
| 2277 | /* Punt if not enough space exists in the first SKB for |
| 2278 | * the data in the second |
| 2279 | */ |
Eric Dumazet | a21d457 | 2012-04-10 20:30:48 +0000 | [diff] [blame] | 2280 | if (skb->len > skb_availroom(to)) |
Ilpo Järvinen | 4a17fc3 | 2008-11-24 21:03:43 -0800 | [diff] [blame] | 2281 | break; |
| 2282 | |
| 2283 | if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) |
| 2284 | break; |
| 2285 | |
| 2286 | tcp_collapse_retrans(sk, to); |
| 2287 | } |
| 2288 | } |
| 2289 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2290 | /* This retransmits one SKB. Policy decisions and retransmit queue |
| 2291 | * state updates are done by the caller. Returns non-zero if an |
| 2292 | * error occurred which prevented the send. |
| 2293 | */ |
Yuchung Cheng | 93b174a | 2012-12-06 08:45:32 +0000 | [diff] [blame] | 2294 | int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2295 | { |
| 2296 | struct tcp_sock *tp = tcp_sk(sk); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 2297 | struct inet_connection_sock *icsk = inet_csk(sk); |
Sridhar Samudrala | 7d227cd2 | 2008-05-21 16:42:20 -0700 | [diff] [blame] | 2298 | unsigned int cur_mss; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2299 | |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 2300 | /* Inconslusive MTU probe */ |
| 2301 | if (icsk->icsk_mtup.probe_size) { |
| 2302 | icsk->icsk_mtup.probe_size = 0; |
| 2303 | } |
| 2304 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2305 | /* 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] | 2306 | * copying overhead: fragmentation, tunneling, mangling etc. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2307 | */ |
| 2308 | if (atomic_read(&sk->sk_wmem_alloc) > |
| 2309 | min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf)) |
| 2310 | return -EAGAIN; |
| 2311 | |
| 2312 | if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { |
| 2313 | if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) |
| 2314 | BUG(); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2315 | if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) |
| 2316 | return -ENOMEM; |
| 2317 | } |
| 2318 | |
Sridhar Samudrala | 7d227cd2 | 2008-05-21 16:42:20 -0700 | [diff] [blame] | 2319 | if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) |
| 2320 | return -EHOSTUNREACH; /* Routing failure or similar. */ |
| 2321 | |
Ilpo Järvinen | 0c54b85 | 2009-03-14 14:23:05 +0000 | [diff] [blame] | 2322 | cur_mss = tcp_current_mss(sk); |
Sridhar Samudrala | 7d227cd2 | 2008-05-21 16:42:20 -0700 | [diff] [blame] | 2323 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2324 | /* If receiver has shrunk his window, and skb is out of |
| 2325 | * new window, do not retransmit it. The exception is the |
| 2326 | * case, when window is shrunk to zero. In this case |
| 2327 | * our retransmit serves as a zero window probe. |
| 2328 | */ |
Joe Perches | 9d4fb27 | 2009-11-23 10:41:23 -0800 | [diff] [blame] | 2329 | if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && |
| 2330 | TCP_SKB_CB(skb)->seq != tp->snd_una) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2331 | return -EAGAIN; |
| 2332 | |
| 2333 | if (skb->len > cur_mss) { |
David S. Miller | 846998a | 2005-08-04 19:52:01 -0700 | [diff] [blame] | 2334 | if (tcp_fragment(sk, skb, cur_mss, cur_mss)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2335 | return -ENOMEM; /* We'll try again later. */ |
Ilpo Järvinen | 02276f3 | 2009-02-28 04:44:31 +0000 | [diff] [blame] | 2336 | } else { |
Ilpo Järvinen | 9eb9362 | 2009-04-01 23:18:20 +0000 | [diff] [blame] | 2337 | int oldpcount = tcp_skb_pcount(skb); |
| 2338 | |
| 2339 | if (unlikely(oldpcount > 1)) { |
| 2340 | tcp_init_tso_segs(sk, skb, cur_mss); |
| 2341 | tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb)); |
| 2342 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2343 | } |
| 2344 | |
Ilpo Järvinen | 4a17fc3 | 2008-11-24 21:03:43 -0800 | [diff] [blame] | 2345 | tcp_retrans_try_collapse(sk, skb, cur_mss); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2346 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2347 | /* Some Solaris stacks overoptimize and ignore the FIN on a |
| 2348 | * retransmit when old data is attached. So strip it off |
| 2349 | * since it is cheap to do so and saves bytes on the network. |
| 2350 | */ |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 2351 | if (skb->len > 0 && |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 2352 | (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 2353 | tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2354 | if (!pskb_trim(skb, 0)) { |
Ilpo Järvinen | e870a8e | 2008-01-03 20:39:01 -0800 | [diff] [blame] | 2355 | /* Reuse, even though it does some unnecessary work */ |
| 2356 | tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1, |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 2357 | TCP_SKB_CB(skb)->tcp_flags); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2358 | skb->ip_summed = CHECKSUM_NONE; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2359 | } |
| 2360 | } |
| 2361 | |
| 2362 | /* Make a copy, if the first transmission SKB clone we made |
| 2363 | * is still in somebody's hands, else make a clone. |
| 2364 | */ |
| 2365 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2366 | |
Thomas Graf | 50bceae | 2013-04-11 10:57:18 +0000 | [diff] [blame] | 2367 | /* make sure skb->data is aligned on arches that require it |
| 2368 | * and check if ack-trimming & collapsing extended the headroom |
| 2369 | * beyond what csum_start can cover. |
| 2370 | */ |
| 2371 | if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) || |
| 2372 | skb_headroom(skb) >= 0xFFFF)) { |
Eric Dumazet | 117632e | 2011-12-03 21:39:53 +0000 | [diff] [blame] | 2373 | struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER, |
| 2374 | GFP_ATOMIC); |
Yuchung Cheng | 93b174a | 2012-12-06 08:45:32 +0000 | [diff] [blame] | 2375 | return nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) : |
| 2376 | -ENOBUFS; |
Eric Dumazet | 117632e | 2011-12-03 21:39:53 +0000 | [diff] [blame] | 2377 | } else { |
Yuchung Cheng | 93b174a | 2012-12-06 08:45:32 +0000 | [diff] [blame] | 2378 | return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); |
Eric Dumazet | 117632e | 2011-12-03 21:39:53 +0000 | [diff] [blame] | 2379 | } |
Yuchung Cheng | 93b174a | 2012-12-06 08:45:32 +0000 | [diff] [blame] | 2380 | } |
| 2381 | |
| 2382 | int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) |
| 2383 | { |
| 2384 | struct tcp_sock *tp = tcp_sk(sk); |
| 2385 | int err = __tcp_retransmit_skb(sk, skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2386 | |
| 2387 | if (err == 0) { |
| 2388 | /* Update global TCP statistics. */ |
Pavel Emelyanov | 81cc8a7 | 2008-07-16 20:22:04 -0700 | [diff] [blame] | 2389 | TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2390 | |
| 2391 | tp->total_retrans++; |
| 2392 | |
| 2393 | #if FASTRETRANS_DEBUG > 0 |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 2394 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { |
Joe Perches | e87cc47 | 2012-05-13 21:56:26 +0000 | [diff] [blame] | 2395 | net_dbg_ratelimited("retrans_out leaked\n"); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2396 | } |
| 2397 | #endif |
Ilpo Järvinen | b08d6cb | 2007-10-11 17:36:13 -0700 | [diff] [blame] | 2398 | if (!tp->retrans_out) |
| 2399 | tp->lost_retrans_low = tp->snd_nxt; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2400 | TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; |
| 2401 | tp->retrans_out += tcp_skb_pcount(skb); |
| 2402 | |
| 2403 | /* Save stamp of the first retransmit. */ |
| 2404 | if (!tp->retrans_stamp) |
| 2405 | tp->retrans_stamp = TCP_SKB_CB(skb)->when; |
| 2406 | |
Yuchung Cheng | c24f691 | 2011-02-07 12:57:04 +0000 | [diff] [blame] | 2407 | tp->undo_retrans += tcp_skb_pcount(skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2408 | |
| 2409 | /* snd_nxt is stored to detect loss of retransmitted segment, |
| 2410 | * see tcp_input.c tcp_sacktag_write_queue(). |
| 2411 | */ |
| 2412 | TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt; |
Yuchung Cheng | 24ab6be | 2013-07-12 11:33:04 -0700 | [diff] [blame] | 2413 | } else { |
| 2414 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2415 | } |
| 2416 | return err; |
| 2417 | } |
| 2418 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 2419 | /* Check if we forward retransmits are possible in the current |
| 2420 | * window/congestion state. |
| 2421 | */ |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 2422 | static bool tcp_can_forward_retransmit(struct sock *sk) |
Ilpo Järvinen | b5afe7b | 2008-09-20 21:21:54 -0700 | [diff] [blame] | 2423 | { |
| 2424 | const struct inet_connection_sock *icsk = inet_csk(sk); |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 2425 | const struct tcp_sock *tp = tcp_sk(sk); |
Ilpo Järvinen | b5afe7b | 2008-09-20 21:21:54 -0700 | [diff] [blame] | 2426 | |
| 2427 | /* Forward retransmissions are possible only during Recovery. */ |
| 2428 | if (icsk->icsk_ca_state != TCP_CA_Recovery) |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 2429 | return false; |
Ilpo Järvinen | b5afe7b | 2008-09-20 21:21:54 -0700 | [diff] [blame] | 2430 | |
| 2431 | /* No forward retransmissions in Reno are possible. */ |
| 2432 | if (tcp_is_reno(tp)) |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 2433 | return false; |
Ilpo Järvinen | b5afe7b | 2008-09-20 21:21:54 -0700 | [diff] [blame] | 2434 | |
| 2435 | /* Yeah, we have to make difficult choice between forward transmission |
| 2436 | * and retransmission... Both ways have their merits... |
| 2437 | * |
| 2438 | * For now we do not retransmit anything, while we have some new |
| 2439 | * segments to send. In the other cases, follow rule 3 for |
| 2440 | * NextSeg() specified in RFC3517. |
| 2441 | */ |
| 2442 | |
| 2443 | if (tcp_may_send_now(sk)) |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 2444 | return false; |
Ilpo Järvinen | b5afe7b | 2008-09-20 21:21:54 -0700 | [diff] [blame] | 2445 | |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 2446 | return true; |
Ilpo Järvinen | b5afe7b | 2008-09-20 21:21:54 -0700 | [diff] [blame] | 2447 | } |
| 2448 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2449 | /* This gets called after a retransmit timeout, and the initially |
| 2450 | * retransmitted data is acknowledged. It tries to continue |
| 2451 | * resending the rest of the retransmit queue, until either |
| 2452 | * we've sent it all or the congestion window limit is reached. |
| 2453 | * If doing SACK, the first ACK which comes back for a timeout |
| 2454 | * based retransmit packet might feed us FACK information again. |
| 2455 | * If so, we use it to avoid unnecessarily retransmissions. |
| 2456 | */ |
| 2457 | void tcp_xmit_retransmit_queue(struct sock *sk) |
| 2458 | { |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2459 | const struct inet_connection_sock *icsk = inet_csk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2460 | struct tcp_sock *tp = tcp_sk(sk); |
| 2461 | struct sk_buff *skb; |
Ilpo Järvinen | 0e1c54c | 2008-09-20 21:24:21 -0700 | [diff] [blame] | 2462 | struct sk_buff *hole = NULL; |
Ilpo Järvinen | 618d9f2 | 2008-09-20 21:26:22 -0700 | [diff] [blame] | 2463 | u32 last_lost; |
Ilpo Järvinen | 61eb55f | 2008-09-20 21:22:59 -0700 | [diff] [blame] | 2464 | int mib_idx; |
Ilpo Järvinen | 0e1c54c | 2008-09-20 21:24:21 -0700 | [diff] [blame] | 2465 | int fwd_rexmitting = 0; |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 2466 | |
Ilpo Järvinen | 45e77d3 | 2010-07-19 01:16:18 +0000 | [diff] [blame] | 2467 | if (!tp->packets_out) |
| 2468 | return; |
| 2469 | |
Ilpo Järvinen | 08ebd17 | 2008-09-20 21:23:49 -0700 | [diff] [blame] | 2470 | if (!tp->lost_out) |
| 2471 | tp->retransmit_high = tp->snd_una; |
| 2472 | |
Ilpo Järvinen | 618d9f2 | 2008-09-20 21:26:22 -0700 | [diff] [blame] | 2473 | if (tp->retransmit_skb_hint) { |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 2474 | skb = tp->retransmit_skb_hint; |
Ilpo Järvinen | 618d9f2 | 2008-09-20 21:26:22 -0700 | [diff] [blame] | 2475 | last_lost = TCP_SKB_CB(skb)->end_seq; |
| 2476 | if (after(last_lost, tp->retransmit_high)) |
| 2477 | last_lost = tp->retransmit_high; |
| 2478 | } else { |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2479 | skb = tcp_write_queue_head(sk); |
Ilpo Järvinen | 618d9f2 | 2008-09-20 21:26:22 -0700 | [diff] [blame] | 2480 | last_lost = tp->snd_una; |
| 2481 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2482 | |
Ilpo Järvinen | 08ebd17 | 2008-09-20 21:23:49 -0700 | [diff] [blame] | 2483 | tcp_for_write_queue_from(skb, sk) { |
| 2484 | __u8 sacked = TCP_SKB_CB(skb)->sacked; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2485 | |
Ilpo Järvinen | 08ebd17 | 2008-09-20 21:23:49 -0700 | [diff] [blame] | 2486 | if (skb == tcp_send_head(sk)) |
| 2487 | break; |
| 2488 | /* we could do better than to assign each time */ |
Ilpo Järvinen | 0e1c54c | 2008-09-20 21:24:21 -0700 | [diff] [blame] | 2489 | if (hole == NULL) |
| 2490 | tp->retransmit_skb_hint = skb; |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 2491 | |
Ilpo Järvinen | 08ebd17 | 2008-09-20 21:23:49 -0700 | [diff] [blame] | 2492 | /* Assume this retransmit will generate |
| 2493 | * only one packet for congestion window |
| 2494 | * calculation purposes. This works because |
| 2495 | * tcp_retransmit_skb() will chop up the |
| 2496 | * packet to be MSS sized and all the |
| 2497 | * packet counting works out. |
| 2498 | */ |
| 2499 | if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) |
| 2500 | return; |
Ilpo Järvinen | 0e1c54c | 2008-09-20 21:24:21 -0700 | [diff] [blame] | 2501 | |
| 2502 | if (fwd_rexmitting) { |
| 2503 | begin_fwd: |
| 2504 | if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp))) |
| 2505 | break; |
| 2506 | mib_idx = LINUX_MIB_TCPFORWARDRETRANS; |
| 2507 | |
| 2508 | } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) { |
Ilpo Järvinen | 618d9f2 | 2008-09-20 21:26:22 -0700 | [diff] [blame] | 2509 | tp->retransmit_high = last_lost; |
Ilpo Järvinen | 0e1c54c | 2008-09-20 21:24:21 -0700 | [diff] [blame] | 2510 | if (!tcp_can_forward_retransmit(sk)) |
| 2511 | break; |
| 2512 | /* Backtrack if necessary to non-L'ed skb */ |
| 2513 | if (hole != NULL) { |
| 2514 | skb = hole; |
| 2515 | hole = NULL; |
| 2516 | } |
| 2517 | fwd_rexmitting = 1; |
| 2518 | goto begin_fwd; |
| 2519 | |
| 2520 | } else if (!(sacked & TCPCB_LOST)) { |
Ilpo Järvinen | ac11ba7 | 2009-02-28 04:44:27 +0000 | [diff] [blame] | 2521 | if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) |
Ilpo Järvinen | 0e1c54c | 2008-09-20 21:24:21 -0700 | [diff] [blame] | 2522 | hole = skb; |
Ilpo Järvinen | 08ebd17 | 2008-09-20 21:23:49 -0700 | [diff] [blame] | 2523 | continue; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2524 | |
Ilpo Järvinen | 0e1c54c | 2008-09-20 21:24:21 -0700 | [diff] [blame] | 2525 | } else { |
Ilpo Järvinen | 618d9f2 | 2008-09-20 21:26:22 -0700 | [diff] [blame] | 2526 | last_lost = TCP_SKB_CB(skb)->end_seq; |
Ilpo Järvinen | 0e1c54c | 2008-09-20 21:24:21 -0700 | [diff] [blame] | 2527 | if (icsk->icsk_ca_state != TCP_CA_Loss) |
| 2528 | mib_idx = LINUX_MIB_TCPFASTRETRANS; |
| 2529 | else |
| 2530 | mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; |
| 2531 | } |
| 2532 | |
| 2533 | if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) |
Ilpo Järvinen | 08ebd17 | 2008-09-20 21:23:49 -0700 | [diff] [blame] | 2534 | continue; |
Pavel Emelyanov | 40b215e | 2008-07-03 01:05:41 -0700 | [diff] [blame] | 2535 | |
Yuchung Cheng | 24ab6be | 2013-07-12 11:33:04 -0700 | [diff] [blame] | 2536 | if (tcp_retransmit_skb(sk, skb)) |
Ilpo Järvinen | 08ebd17 | 2008-09-20 21:23:49 -0700 | [diff] [blame] | 2537 | return; |
Yuchung Cheng | 24ab6be | 2013-07-12 11:33:04 -0700 | [diff] [blame] | 2538 | |
Ilpo Järvinen | 08ebd17 | 2008-09-20 21:23:49 -0700 | [diff] [blame] | 2539 | NET_INC_STATS_BH(sock_net(sk), mib_idx); |
| 2540 | |
Yuchung Cheng | 684bad1 | 2012-09-02 17:38:04 +0000 | [diff] [blame] | 2541 | if (tcp_in_cwnd_reduction(sk)) |
Nandita Dukkipati | a262f0c | 2011-08-21 20:21:57 +0000 | [diff] [blame] | 2542 | tp->prr_out += tcp_skb_pcount(skb); |
| 2543 | |
Ilpo Järvinen | 08ebd17 | 2008-09-20 21:23:49 -0700 | [diff] [blame] | 2544 | if (skb == tcp_write_queue_head(sk)) |
| 2545 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, |
| 2546 | inet_csk(sk)->icsk_rto, |
| 2547 | TCP_RTO_MAX); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2548 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2549 | } |
| 2550 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2551 | /* Send a fin. The caller locks the socket for us. This cannot be |
| 2552 | * allowed to fail queueing a FIN frame under any circumstances. |
| 2553 | */ |
| 2554 | void tcp_send_fin(struct sock *sk) |
| 2555 | { |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2556 | struct tcp_sock *tp = tcp_sk(sk); |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2557 | struct sk_buff *skb = tcp_write_queue_tail(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2558 | int mss_now; |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2559 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2560 | /* Optimization, tack on the FIN if we have a queue of |
| 2561 | * unsent frames. But be careful about outgoing SACKS |
| 2562 | * and IP options. |
| 2563 | */ |
Ilpo Järvinen | 0c54b85 | 2009-03-14 14:23:05 +0000 | [diff] [blame] | 2564 | mss_now = tcp_current_mss(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2565 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2566 | if (tcp_send_head(sk) != NULL) { |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 2567 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_FIN; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2568 | TCP_SKB_CB(skb)->end_seq++; |
| 2569 | tp->write_seq++; |
| 2570 | } else { |
| 2571 | /* Socket is locked, keep trying until memory is available. */ |
| 2572 | for (;;) { |
Wu Fengguang | aa13307 | 2009-09-02 23:45:45 -0700 | [diff] [blame] | 2573 | skb = alloc_skb_fclone(MAX_TCP_HEADER, |
| 2574 | sk->sk_allocation); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2575 | if (skb) |
| 2576 | break; |
| 2577 | yield(); |
| 2578 | } |
| 2579 | |
| 2580 | /* Reserve space for headers and prepare control bits. */ |
| 2581 | skb_reserve(skb, MAX_TCP_HEADER); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2582 | /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ |
Ilpo Järvinen | e870a8e | 2008-01-03 20:39:01 -0800 | [diff] [blame] | 2583 | tcp_init_nondata_skb(skb, tp->write_seq, |
Changli Gao | a3433f3 | 2010-06-12 14:01:43 +0000 | [diff] [blame] | 2584 | TCPHDR_ACK | TCPHDR_FIN); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2585 | tcp_queue_skb(sk, skb); |
| 2586 | } |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2587 | __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2588 | } |
| 2589 | |
| 2590 | /* We get here when a process closes a file descriptor (either due to |
| 2591 | * an explicit close() or as a byproduct of exit()'ing) and there |
| 2592 | * was unread data in the receive queue. This behavior is recommended |
Gerrit Renker | 65bb723 | 2007-04-28 21:21:46 -0700 | [diff] [blame] | 2593 | * by RFC 2525, section 2.17. -DaveM |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2594 | */ |
Al Viro | dd0fc66 | 2005-10-07 07:46:04 +0100 | [diff] [blame] | 2595 | void tcp_send_active_reset(struct sock *sk, gfp_t priority) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2596 | { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2597 | struct sk_buff *skb; |
| 2598 | |
| 2599 | /* NOTE: No TCP options attached and we never retransmit this. */ |
| 2600 | skb = alloc_skb(MAX_TCP_HEADER, priority); |
| 2601 | if (!skb) { |
Pavel Emelyanov | 4e67344 | 2008-07-16 20:30:14 -0700 | [diff] [blame] | 2602 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2603 | return; |
| 2604 | } |
| 2605 | |
| 2606 | /* Reserve space for headers and prepare control bits. */ |
| 2607 | skb_reserve(skb, MAX_TCP_HEADER); |
Ilpo Järvinen | e870a8e | 2008-01-03 20:39:01 -0800 | [diff] [blame] | 2608 | tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), |
Changli Gao | a3433f3 | 2010-06-12 14:01:43 +0000 | [diff] [blame] | 2609 | TCPHDR_ACK | TCPHDR_RST); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2610 | /* Send it off. */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2611 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 2612 | if (tcp_transmit_skb(sk, skb, 0, priority)) |
Pavel Emelyanov | 4e67344 | 2008-07-16 20:30:14 -0700 | [diff] [blame] | 2613 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); |
Sridhar Samudrala | 26af65c | 2008-06-04 15:19:35 -0700 | [diff] [blame] | 2614 | |
Pavel Emelyanov | 81cc8a7 | 2008-07-16 20:22:04 -0700 | [diff] [blame] | 2615 | TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2616 | } |
| 2617 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 2618 | /* Send a crossed SYN-ACK during socket establishment. |
| 2619 | * WARNING: This routine must only be called when we have already sent |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2620 | * a SYN packet that crossed the incoming SYN that caused this routine |
| 2621 | * to get called. If this assumption fails then the initial rcv_wnd |
| 2622 | * and rcv_wscale values will not be correct. |
| 2623 | */ |
| 2624 | int tcp_send_synack(struct sock *sk) |
| 2625 | { |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 2626 | struct sk_buff *skb; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2627 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2628 | skb = tcp_write_queue_head(sk); |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 2629 | if (skb == NULL || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { |
Joe Perches | 91df42b | 2012-05-15 14:11:54 +0000 | [diff] [blame] | 2630 | pr_debug("%s: wrong queue state\n", __func__); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2631 | return -EFAULT; |
| 2632 | } |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 2633 | if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2634 | if (skb_cloned(skb)) { |
| 2635 | struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC); |
| 2636 | if (nskb == NULL) |
| 2637 | return -ENOMEM; |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2638 | tcp_unlink_write_queue(skb, sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2639 | skb_header_release(nskb); |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2640 | __tcp_add_write_queue_head(sk, nskb); |
Hideo Aoki | 3ab224b | 2007-12-31 00:11:19 -0800 | [diff] [blame] | 2641 | sk_wmem_free_skb(sk, skb); |
| 2642 | sk->sk_wmem_queued += nskb->truesize; |
| 2643 | sk_mem_charge(sk, nskb->truesize); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2644 | skb = nskb; |
| 2645 | } |
| 2646 | |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 2647 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2648 | TCP_ECN_send_synack(tcp_sk(sk), skb); |
| 2649 | } |
| 2650 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 2651 | return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2652 | } |
| 2653 | |
Eric Dumazet | 4aea39c | 2012-06-03 20:33:21 +0000 | [diff] [blame] | 2654 | /** |
| 2655 | * tcp_make_synack - Prepare a SYN-ACK. |
| 2656 | * sk: listener socket |
| 2657 | * dst: dst entry attached to the SYNACK |
| 2658 | * req: request_sock pointer |
Eric Dumazet | 4aea39c | 2012-06-03 20:33:21 +0000 | [diff] [blame] | 2659 | * |
| 2660 | * Allocate one skb and build a SYNACK packet. |
| 2661 | * @dst is consumed : Caller should not use it again. |
| 2662 | */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 2663 | struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst, |
William Allen Simpson | e6b4d11 | 2009-12-02 18:07:39 +0000 | [diff] [blame] | 2664 | struct request_sock *req, |
Jerry Chu | 8336886 | 2012-08-31 12:29:12 +0000 | [diff] [blame] | 2665 | struct tcp_fastopen_cookie *foc) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2666 | { |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 2667 | struct tcp_out_options opts; |
Arnaldo Carvalho de Melo | 2e6599c | 2005-06-18 22:46:52 -0700 | [diff] [blame] | 2668 | struct inet_request_sock *ireq = inet_rsk(req); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2669 | struct tcp_sock *tp = tcp_sk(sk); |
| 2670 | struct tcphdr *th; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2671 | struct sk_buff *skb; |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 2672 | struct tcp_md5sig_key *md5; |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 2673 | int tcp_header_size; |
Tom Quetchenbach | f5fff5d | 2008-09-21 00:21:51 -0700 | [diff] [blame] | 2674 | int mss; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2675 | |
Christoph Paasch | 1a2c618 | 2013-03-17 08:23:34 +0000 | [diff] [blame] | 2676 | skb = alloc_skb(MAX_TCP_HEADER + 15, sk_gfp_atomic(sk, GFP_ATOMIC)); |
Eric Dumazet | 4aea39c | 2012-06-03 20:33:21 +0000 | [diff] [blame] | 2677 | if (unlikely(!skb)) { |
| 2678 | dst_release(dst); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2679 | return NULL; |
Eric Dumazet | 4aea39c | 2012-06-03 20:33:21 +0000 | [diff] [blame] | 2680 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2681 | /* Reserve space for headers. */ |
| 2682 | skb_reserve(skb, MAX_TCP_HEADER); |
| 2683 | |
Eric Dumazet | 4aea39c | 2012-06-03 20:33:21 +0000 | [diff] [blame] | 2684 | skb_dst_set(skb, dst); |
Eric Dumazet | ca10b9e | 2013-04-08 17:58:11 +0000 | [diff] [blame] | 2685 | security_skb_owned_by(skb, sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2686 | |
David S. Miller | 0dbaee3 | 2010-12-13 12:52:14 -0800 | [diff] [blame] | 2687 | mss = dst_metric_advmss(dst); |
Tom Quetchenbach | f5fff5d | 2008-09-21 00:21:51 -0700 | [diff] [blame] | 2688 | if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss) |
| 2689 | mss = tp->rx_opt.user_mss; |
| 2690 | |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 2691 | if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */ |
| 2692 | __u8 rcv_wscale; |
| 2693 | /* Set this up on the first call only */ |
| 2694 | req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW); |
Hagen Paul Pfeifer | e88c64f | 2010-08-19 06:33:05 +0000 | [diff] [blame] | 2695 | |
| 2696 | /* limit the window selection if the user enforce a smaller rx buffer */ |
| 2697 | if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && |
| 2698 | (req->window_clamp > tcp_full_space(sk) || req->window_clamp == 0)) |
| 2699 | req->window_clamp = tcp_full_space(sk); |
| 2700 | |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 2701 | /* tcp_full_space because it is guaranteed to be the first packet */ |
| 2702 | tcp_select_initial_window(tcp_full_space(sk), |
Tom Quetchenbach | f5fff5d | 2008-09-21 00:21:51 -0700 | [diff] [blame] | 2703 | mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0), |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 2704 | &req->rcv_wnd, |
| 2705 | &req->window_clamp, |
| 2706 | ireq->wscale_ok, |
laurent chavey | 31d1292 | 2009-12-15 11:15:28 +0000 | [diff] [blame] | 2707 | &rcv_wscale, |
| 2708 | dst_metric(dst, RTAX_INITRWND)); |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 2709 | ireq->rcv_wscale = rcv_wscale; |
| 2710 | } |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 2711 | |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 2712 | memset(&opts, 0, sizeof(opts)); |
Florian Westphal | 8b5f12d | 2008-10-26 23:10:12 -0700 | [diff] [blame] | 2713 | #ifdef CONFIG_SYN_COOKIES |
| 2714 | if (unlikely(req->cookie_ts)) |
| 2715 | TCP_SKB_CB(skb)->when = cookie_init_timestamp(req); |
| 2716 | else |
| 2717 | #endif |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 2718 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
Christoph Paasch | 1a2c618 | 2013-03-17 08:23:34 +0000 | [diff] [blame] | 2719 | tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, &md5, |
| 2720 | foc) + sizeof(*th); |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 2721 | |
Arnaldo Carvalho de Melo | aa8223c | 2007-04-10 21:04:22 -0700 | [diff] [blame] | 2722 | skb_push(skb, tcp_header_size); |
| 2723 | skb_reset_transport_header(skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2724 | |
Arnaldo Carvalho de Melo | aa8223c | 2007-04-10 21:04:22 -0700 | [diff] [blame] | 2725 | th = tcp_hdr(skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2726 | memset(th, 0, sizeof(struct tcphdr)); |
| 2727 | th->syn = 1; |
| 2728 | th->ack = 1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2729 | TCP_ECN_make_synack(req, th); |
KOVACS Krisztian | a3116ac5 | 2008-10-01 07:46:49 -0700 | [diff] [blame] | 2730 | th->source = ireq->loc_port; |
Arnaldo Carvalho de Melo | 2e6599c | 2005-06-18 22:46:52 -0700 | [diff] [blame] | 2731 | th->dest = ireq->rmt_port; |
Ilpo Järvinen | e870a8e | 2008-01-03 20:39:01 -0800 | [diff] [blame] | 2732 | /* Setting of flags are superfluous here for callers (and ECE is |
| 2733 | * not even correctly set) |
| 2734 | */ |
| 2735 | tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn, |
Changli Gao | a3433f3 | 2010-06-12 14:01:43 +0000 | [diff] [blame] | 2736 | TCPHDR_SYN | TCPHDR_ACK); |
William Allen Simpson | 4957faade | 2009-12-02 18:25:27 +0000 | [diff] [blame] | 2737 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2738 | th->seq = htonl(TCP_SKB_CB(skb)->seq); |
Jerry Chu | 8336886 | 2012-08-31 12:29:12 +0000 | [diff] [blame] | 2739 | /* XXX data is queued and acked as is. No buffer/window check */ |
| 2740 | th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2741 | |
| 2742 | /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ |
Ilpo Järvinen | 600ff0c | 2007-02-13 12:42:11 -0800 | [diff] [blame] | 2743 | th->window = htons(min(req->rcv_wnd, 65535U)); |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 2744 | tcp_options_write((__be32 *)(th + 1), tp, &opts); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2745 | th->doff = (tcp_header_size >> 2); |
Tom Herbert | aa2ea05 | 2010-04-22 07:00:24 +0000 | [diff] [blame] | 2746 | TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, tcp_skb_pcount(skb)); |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 2747 | |
| 2748 | #ifdef CONFIG_TCP_MD5SIG |
| 2749 | /* Okay, we have all we need - do the md5 hash if needed */ |
| 2750 | if (md5) { |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 2751 | tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, |
Adam Langley | 49a72df | 2008-07-19 00:01:42 -0700 | [diff] [blame] | 2752 | md5, NULL, req, skb); |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 2753 | } |
| 2754 | #endif |
| 2755 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2756 | return skb; |
| 2757 | } |
Eric Dumazet | 4bc2f18 | 2010-07-09 21:22:10 +0000 | [diff] [blame] | 2758 | EXPORT_SYMBOL(tcp_make_synack); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2759 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 2760 | /* Do all connect socket setups that can be done AF independent. */ |
Pavel Emelyanov | 370816a | 2012-04-19 03:40:01 +0000 | [diff] [blame] | 2761 | void tcp_connect_init(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2762 | { |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 2763 | const struct dst_entry *dst = __sk_dst_get(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2764 | struct tcp_sock *tp = tcp_sk(sk); |
| 2765 | __u8 rcv_wscale; |
| 2766 | |
| 2767 | /* We'll fix this up when we get a response from the other end. |
| 2768 | * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. |
| 2769 | */ |
| 2770 | tp->tcp_header_len = sizeof(struct tcphdr) + |
David S. Miller | bb5b7c1 | 2009-12-15 20:56:42 -0800 | [diff] [blame] | 2771 | (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2772 | |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 2773 | #ifdef CONFIG_TCP_MD5SIG |
| 2774 | if (tp->af_specific->md5_lookup(sk, sk) != NULL) |
| 2775 | tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; |
| 2776 | #endif |
| 2777 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2778 | /* If user gave his TCP_MAXSEG, record it to clamp */ |
| 2779 | if (tp->rx_opt.user_mss) |
| 2780 | tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; |
| 2781 | tp->max_window = 0; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 2782 | tcp_mtup_init(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2783 | tcp_sync_mss(sk, dst_mtu(dst)); |
| 2784 | |
| 2785 | if (!tp->window_clamp) |
| 2786 | tp->window_clamp = dst_metric(dst, RTAX_WINDOW); |
David S. Miller | 0dbaee3 | 2010-12-13 12:52:14 -0800 | [diff] [blame] | 2787 | tp->advmss = dst_metric_advmss(dst); |
Tom Quetchenbach | f5fff5d | 2008-09-21 00:21:51 -0700 | [diff] [blame] | 2788 | if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss) |
| 2789 | tp->advmss = tp->rx_opt.user_mss; |
| 2790 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2791 | tcp_initialize_rcv_mss(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2792 | |
Hagen Paul Pfeifer | e88c64f | 2010-08-19 06:33:05 +0000 | [diff] [blame] | 2793 | /* limit the window selection if the user enforce a smaller rx buffer */ |
| 2794 | if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && |
| 2795 | (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) |
| 2796 | tp->window_clamp = tcp_full_space(sk); |
| 2797 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2798 | tcp_select_initial_window(tcp_full_space(sk), |
| 2799 | tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), |
| 2800 | &tp->rcv_wnd, |
| 2801 | &tp->window_clamp, |
David S. Miller | bb5b7c1 | 2009-12-15 20:56:42 -0800 | [diff] [blame] | 2802 | sysctl_tcp_window_scaling, |
laurent chavey | 31d1292 | 2009-12-15 11:15:28 +0000 | [diff] [blame] | 2803 | &rcv_wscale, |
| 2804 | dst_metric(dst, RTAX_INITRWND)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2805 | |
| 2806 | tp->rx_opt.rcv_wscale = rcv_wscale; |
| 2807 | tp->rcv_ssthresh = tp->rcv_wnd; |
| 2808 | |
| 2809 | sk->sk_err = 0; |
| 2810 | sock_reset_flag(sk, SOCK_DONE); |
| 2811 | tp->snd_wnd = 0; |
Hantzis Fotis | ee7537b | 2009-03-02 22:42:02 -0800 | [diff] [blame] | 2812 | tcp_init_wl(tp, 0); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2813 | tp->snd_una = tp->write_seq; |
| 2814 | tp->snd_sml = tp->write_seq; |
Ilpo Järvinen | 33f5f57 | 2008-10-07 14:43:06 -0700 | [diff] [blame] | 2815 | tp->snd_up = tp->write_seq; |
Pavel Emelyanov | 370816a | 2012-04-19 03:40:01 +0000 | [diff] [blame] | 2816 | tp->snd_nxt = tp->write_seq; |
Pavel Emelyanov | ee99528 | 2012-04-19 03:40:39 +0000 | [diff] [blame] | 2817 | |
| 2818 | if (likely(!tp->repair)) |
| 2819 | tp->rcv_nxt = 0; |
| 2820 | tp->rcv_wup = tp->rcv_nxt; |
| 2821 | tp->copied_seq = tp->rcv_nxt; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2822 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2823 | inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; |
| 2824 | inet_csk(sk)->icsk_retransmits = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2825 | tcp_clear_retrans(tp); |
| 2826 | } |
| 2827 | |
Yuchung Cheng | 783237e | 2012-07-19 06:43:07 +0000 | [diff] [blame] | 2828 | static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) |
| 2829 | { |
| 2830 | struct tcp_sock *tp = tcp_sk(sk); |
| 2831 | struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); |
| 2832 | |
| 2833 | tcb->end_seq += skb->len; |
| 2834 | skb_header_release(skb); |
| 2835 | __tcp_add_write_queue_tail(sk, skb); |
| 2836 | sk->sk_wmem_queued += skb->truesize; |
| 2837 | sk_mem_charge(sk, skb->truesize); |
| 2838 | tp->write_seq = tcb->end_seq; |
| 2839 | tp->packets_out += tcp_skb_pcount(skb); |
| 2840 | } |
| 2841 | |
| 2842 | /* Build and send a SYN with data and (cached) Fast Open cookie. However, |
| 2843 | * queue a data-only packet after the regular SYN, such that regular SYNs |
| 2844 | * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges |
| 2845 | * only the SYN sequence, the data are retransmitted in the first ACK. |
| 2846 | * If cookie is not cached or other error occurs, falls back to send a |
| 2847 | * regular SYN with Fast Open cookie request option. |
| 2848 | */ |
| 2849 | static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) |
| 2850 | { |
| 2851 | struct tcp_sock *tp = tcp_sk(sk); |
| 2852 | struct tcp_fastopen_request *fo = tp->fastopen_req; |
Yuchung Cheng | aab4874 | 2012-07-19 06:43:10 +0000 | [diff] [blame] | 2853 | int syn_loss = 0, space, i, err = 0, iovlen = fo->data->msg_iovlen; |
Yuchung Cheng | 783237e | 2012-07-19 06:43:07 +0000 | [diff] [blame] | 2854 | struct sk_buff *syn_data = NULL, *data; |
Yuchung Cheng | aab4874 | 2012-07-19 06:43:10 +0000 | [diff] [blame] | 2855 | unsigned long last_syn_loss = 0; |
Yuchung Cheng | 783237e | 2012-07-19 06:43:07 +0000 | [diff] [blame] | 2856 | |
Yuchung Cheng | 67da22d | 2012-07-19 06:43:11 +0000 | [diff] [blame] | 2857 | tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ |
Yuchung Cheng | aab4874 | 2012-07-19 06:43:10 +0000 | [diff] [blame] | 2858 | tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie, |
| 2859 | &syn_loss, &last_syn_loss); |
| 2860 | /* Recurring FO SYN losses: revert to regular handshake temporarily */ |
| 2861 | if (syn_loss > 1 && |
| 2862 | time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) { |
| 2863 | fo->cookie.len = -1; |
| 2864 | goto fallback; |
| 2865 | } |
| 2866 | |
Yuchung Cheng | 67da22d | 2012-07-19 06:43:11 +0000 | [diff] [blame] | 2867 | if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE) |
| 2868 | fo->cookie.len = -1; |
| 2869 | else if (fo->cookie.len <= 0) |
Yuchung Cheng | 783237e | 2012-07-19 06:43:07 +0000 | [diff] [blame] | 2870 | goto fallback; |
| 2871 | |
| 2872 | /* MSS for SYN-data is based on cached MSS and bounded by PMTU and |
| 2873 | * user-MSS. Reserve maximum option space for middleboxes that add |
| 2874 | * private TCP options. The cost is reduced data space in SYN :( |
| 2875 | */ |
| 2876 | if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp) |
| 2877 | tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; |
Yuchung Cheng | 1b63edd | 2013-02-22 08:59:06 +0000 | [diff] [blame] | 2878 | space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) - |
Yuchung Cheng | 783237e | 2012-07-19 06:43:07 +0000 | [diff] [blame] | 2879 | MAX_TCP_OPTION_SPACE; |
| 2880 | |
| 2881 | syn_data = skb_copy_expand(syn, skb_headroom(syn), space, |
| 2882 | sk->sk_allocation); |
| 2883 | if (syn_data == NULL) |
| 2884 | goto fallback; |
| 2885 | |
| 2886 | for (i = 0; i < iovlen && syn_data->len < space; ++i) { |
| 2887 | struct iovec *iov = &fo->data->msg_iov[i]; |
| 2888 | unsigned char __user *from = iov->iov_base; |
| 2889 | int len = iov->iov_len; |
| 2890 | |
| 2891 | if (syn_data->len + len > space) |
| 2892 | len = space - syn_data->len; |
| 2893 | else if (i + 1 == iovlen) |
| 2894 | /* No more data pending in inet_wait_for_connect() */ |
| 2895 | fo->data = NULL; |
| 2896 | |
| 2897 | if (skb_add_data(syn_data, from, len)) |
| 2898 | goto fallback; |
| 2899 | } |
| 2900 | |
| 2901 | /* Queue a data-only packet after the regular SYN for retransmission */ |
| 2902 | data = pskb_copy(syn_data, sk->sk_allocation); |
| 2903 | if (data == NULL) |
| 2904 | goto fallback; |
| 2905 | TCP_SKB_CB(data)->seq++; |
| 2906 | TCP_SKB_CB(data)->tcp_flags &= ~TCPHDR_SYN; |
| 2907 | TCP_SKB_CB(data)->tcp_flags = (TCPHDR_ACK|TCPHDR_PSH); |
| 2908 | tcp_connect_queue_skb(sk, data); |
| 2909 | fo->copied = data->len; |
| 2910 | |
| 2911 | if (tcp_transmit_skb(sk, syn_data, 0, sk->sk_allocation) == 0) { |
Yuchung Cheng | 67da22d | 2012-07-19 06:43:11 +0000 | [diff] [blame] | 2912 | tp->syn_data = (fo->copied > 0); |
Yuchung Cheng | 783237e | 2012-07-19 06:43:07 +0000 | [diff] [blame] | 2913 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE); |
| 2914 | goto done; |
| 2915 | } |
| 2916 | syn_data = NULL; |
| 2917 | |
| 2918 | fallback: |
| 2919 | /* Send a regular SYN with Fast Open cookie request option */ |
| 2920 | if (fo->cookie.len > 0) |
| 2921 | fo->cookie.len = 0; |
| 2922 | err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); |
| 2923 | if (err) |
| 2924 | tp->syn_fastopen = 0; |
| 2925 | kfree_skb(syn_data); |
| 2926 | done: |
| 2927 | fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ |
| 2928 | return err; |
| 2929 | } |
| 2930 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 2931 | /* Build a SYN and send it off. */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2932 | int tcp_connect(struct sock *sk) |
| 2933 | { |
| 2934 | struct tcp_sock *tp = tcp_sk(sk); |
| 2935 | struct sk_buff *buff; |
Eric Paris | ee58681 | 2010-11-16 11:52:49 +0000 | [diff] [blame] | 2936 | int err; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2937 | |
| 2938 | tcp_connect_init(sk); |
| 2939 | |
Andrey Vagin | 2b91647 | 2012-11-22 01:13:58 +0000 | [diff] [blame] | 2940 | if (unlikely(tp->repair)) { |
| 2941 | tcp_finish_connect(sk, NULL); |
| 2942 | return 0; |
| 2943 | } |
| 2944 | |
David S. Miller | d179cd1 | 2005-08-17 14:57:30 -0700 | [diff] [blame] | 2945 | buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2946 | if (unlikely(buff == NULL)) |
| 2947 | return -ENOBUFS; |
| 2948 | |
| 2949 | /* Reserve space for headers. */ |
| 2950 | skb_reserve(buff, MAX_TCP_HEADER); |
| 2951 | |
Changli Gao | a3433f3 | 2010-06-12 14:01:43 +0000 | [diff] [blame] | 2952 | tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); |
Yuchung Cheng | 783237e | 2012-07-19 06:43:07 +0000 | [diff] [blame] | 2953 | tp->retrans_stamp = TCP_SKB_CB(buff)->when = tcp_time_stamp; |
| 2954 | tcp_connect_queue_skb(sk, buff); |
Ilpo Järvinen | e870a8e | 2008-01-03 20:39:01 -0800 | [diff] [blame] | 2955 | TCP_ECN_send_syn(sk, buff); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2956 | |
Yuchung Cheng | 783237e | 2012-07-19 06:43:07 +0000 | [diff] [blame] | 2957 | /* Send off SYN; include data in Fast Open. */ |
| 2958 | err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : |
| 2959 | tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); |
Eric Paris | ee58681 | 2010-11-16 11:52:49 +0000 | [diff] [blame] | 2960 | if (err == -ECONNREFUSED) |
| 2961 | return err; |
Wei Yongjun | bd37a08 | 2006-08-07 21:04:15 -0700 | [diff] [blame] | 2962 | |
| 2963 | /* We change tp->snd_nxt after the tcp_transmit_skb() call |
| 2964 | * in order to make this packet get counted in tcpOutSegs. |
| 2965 | */ |
| 2966 | tp->snd_nxt = tp->write_seq; |
| 2967 | tp->pushed_seq = tp->write_seq; |
Pavel Emelyanov | 81cc8a7 | 2008-07-16 20:22:04 -0700 | [diff] [blame] | 2968 | TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2969 | |
| 2970 | /* Timer for repeating the SYN until an answer. */ |
Arnaldo Carvalho de Melo | 3f421ba | 2005-08-09 20:11:08 -0700 | [diff] [blame] | 2971 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, |
| 2972 | inet_csk(sk)->icsk_rto, TCP_RTO_MAX); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2973 | return 0; |
| 2974 | } |
Eric Dumazet | 4bc2f18 | 2010-07-09 21:22:10 +0000 | [diff] [blame] | 2975 | EXPORT_SYMBOL(tcp_connect); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2976 | |
| 2977 | /* Send out a delayed ack, the caller does the policy checking |
| 2978 | * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() |
| 2979 | * for details. |
| 2980 | */ |
| 2981 | void tcp_send_delayed_ack(struct sock *sk) |
| 2982 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2983 | struct inet_connection_sock *icsk = inet_csk(sk); |
| 2984 | int ato = icsk->icsk_ack.ato; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2985 | unsigned long timeout; |
| 2986 | |
| 2987 | if (ato > TCP_DELACK_MIN) { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2988 | const struct tcp_sock *tp = tcp_sk(sk); |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 2989 | int max_ato = HZ / 2; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2990 | |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 2991 | if (icsk->icsk_ack.pingpong || |
| 2992 | (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2993 | max_ato = TCP_DELACK_MAX; |
| 2994 | |
| 2995 | /* Slow path, intersegment interval is "high". */ |
| 2996 | |
| 2997 | /* 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] | 2998 | * 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] | 2999 | * directly. |
| 3000 | */ |
| 3001 | if (tp->srtt) { |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 3002 | int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3003 | |
| 3004 | if (rtt < max_ato) |
| 3005 | max_ato = rtt; |
| 3006 | } |
| 3007 | |
| 3008 | ato = min(ato, max_ato); |
| 3009 | } |
| 3010 | |
| 3011 | /* Stay within the limit we were given */ |
| 3012 | timeout = jiffies + ato; |
| 3013 | |
| 3014 | /* 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] | 3015 | if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3016 | /* If delack timer was blocked or is about to expire, |
| 3017 | * send ACK now. |
| 3018 | */ |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 3019 | if (icsk->icsk_ack.blocked || |
| 3020 | time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3021 | tcp_send_ack(sk); |
| 3022 | return; |
| 3023 | } |
| 3024 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 3025 | if (!time_before(timeout, icsk->icsk_ack.timeout)) |
| 3026 | timeout = icsk->icsk_ack.timeout; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3027 | } |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 3028 | icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; |
| 3029 | icsk->icsk_ack.timeout = timeout; |
| 3030 | sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3031 | } |
| 3032 | |
| 3033 | /* This routine sends an ack and also updates the window. */ |
| 3034 | void tcp_send_ack(struct sock *sk) |
| 3035 | { |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 3036 | struct sk_buff *buff; |
| 3037 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3038 | /* If we have been reset, we may not send again. */ |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 3039 | if (sk->sk_state == TCP_CLOSE) |
| 3040 | return; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3041 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 3042 | /* We are not putting this on the write queue, so |
| 3043 | * tcp_transmit_skb() will set the ownership to this |
| 3044 | * sock. |
| 3045 | */ |
Mel Gorman | 99a1dec | 2012-07-31 16:44:14 -0700 | [diff] [blame] | 3046 | buff = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC)); |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 3047 | if (buff == NULL) { |
| 3048 | inet_csk_schedule_ack(sk); |
| 3049 | inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; |
| 3050 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, |
| 3051 | TCP_DELACK_MAX, TCP_RTO_MAX); |
| 3052 | return; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3053 | } |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 3054 | |
| 3055 | /* Reserve space for headers and prepare control bits. */ |
| 3056 | skb_reserve(buff, MAX_TCP_HEADER); |
Changli Gao | a3433f3 | 2010-06-12 14:01:43 +0000 | [diff] [blame] | 3057 | tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 3058 | |
| 3059 | /* Send it off, this clears delayed acks for us. */ |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 3060 | TCP_SKB_CB(buff)->when = tcp_time_stamp; |
Mel Gorman | 99a1dec | 2012-07-31 16:44:14 -0700 | [diff] [blame] | 3061 | tcp_transmit_skb(sk, buff, 0, sk_gfp_atomic(sk, GFP_ATOMIC)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3062 | } |
| 3063 | |
| 3064 | /* This routine sends a packet with an out of date sequence |
| 3065 | * number. It assumes the other end will try to ack it. |
| 3066 | * |
| 3067 | * Question: what should we make while urgent mode? |
| 3068 | * 4.4BSD forces sending single byte of data. We cannot send |
| 3069 | * out of window data, because we have SND.NXT==SND.MAX... |
| 3070 | * |
| 3071 | * Current solution: to send TWO zero-length segments in urgent mode: |
| 3072 | * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is |
| 3073 | * out-of-date with SND.UNA-1 to probe window. |
| 3074 | */ |
| 3075 | static int tcp_xmit_probe_skb(struct sock *sk, int urgent) |
| 3076 | { |
| 3077 | struct tcp_sock *tp = tcp_sk(sk); |
| 3078 | struct sk_buff *skb; |
| 3079 | |
| 3080 | /* We don't queue it, tcp_transmit_skb() sets ownership. */ |
Mel Gorman | 99a1dec | 2012-07-31 16:44:14 -0700 | [diff] [blame] | 3081 | skb = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC)); |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 3082 | if (skb == NULL) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3083 | return -1; |
| 3084 | |
| 3085 | /* Reserve space for headers and set control bits. */ |
| 3086 | skb_reserve(skb, MAX_TCP_HEADER); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3087 | /* Use a previous sequence. This should cause the other |
| 3088 | * end to send an ack. Don't queue or clone SKB, just |
| 3089 | * send it. |
| 3090 | */ |
Changli Gao | a3433f3 | 2010-06-12 14:01:43 +0000 | [diff] [blame] | 3091 | tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3092 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 3093 | return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3094 | } |
| 3095 | |
Pavel Emelyanov | ee99528 | 2012-04-19 03:40:39 +0000 | [diff] [blame] | 3096 | void tcp_send_window_probe(struct sock *sk) |
| 3097 | { |
| 3098 | if (sk->sk_state == TCP_ESTABLISHED) { |
| 3099 | tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; |
Pavel Emelyanov | c0e88ff | 2012-04-19 03:41:01 +0000 | [diff] [blame] | 3100 | tcp_sk(sk)->snd_nxt = tcp_sk(sk)->write_seq; |
Pavel Emelyanov | ee99528 | 2012-04-19 03:40:39 +0000 | [diff] [blame] | 3101 | tcp_xmit_probe_skb(sk, 0); |
| 3102 | } |
| 3103 | } |
| 3104 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 3105 | /* Initiate keepalive or window probe from timer. */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3106 | int tcp_write_wakeup(struct sock *sk) |
| 3107 | { |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 3108 | struct tcp_sock *tp = tcp_sk(sk); |
| 3109 | struct sk_buff *skb; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3110 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 3111 | if (sk->sk_state == TCP_CLOSE) |
| 3112 | return -1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3113 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 3114 | if ((skb = tcp_send_head(sk)) != NULL && |
| 3115 | before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { |
| 3116 | int err; |
Ilpo Järvinen | 0c54b85 | 2009-03-14 14:23:05 +0000 | [diff] [blame] | 3117 | unsigned int mss = tcp_current_mss(sk); |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 3118 | 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] | 3119 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 3120 | if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) |
| 3121 | tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3122 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 3123 | /* We are probing the opening of a window |
| 3124 | * but the window size is != 0 |
| 3125 | * must have been a result SWS avoidance ( sender ) |
| 3126 | */ |
| 3127 | if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || |
| 3128 | skb->len > mss) { |
| 3129 | seg_size = min(seg_size, mss); |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 3130 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 3131 | if (tcp_fragment(sk, skb, seg_size, mss)) |
| 3132 | return -1; |
| 3133 | } else if (!tcp_skb_pcount(skb)) |
| 3134 | tcp_set_skb_tso_segs(sk, skb, mss); |
| 3135 | |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 3136 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 3137 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
| 3138 | err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); |
| 3139 | if (!err) |
| 3140 | tcp_event_new_data_sent(sk, skb); |
| 3141 | return err; |
| 3142 | } else { |
Ilpo Järvinen | 33f5f57 | 2008-10-07 14:43:06 -0700 | [diff] [blame] | 3143 | if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 3144 | tcp_xmit_probe_skb(sk, 1); |
| 3145 | return tcp_xmit_probe_skb(sk, 0); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3146 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3147 | } |
| 3148 | |
| 3149 | /* A window probe timeout has occurred. If window is not closed send |
| 3150 | * a partial packet else a zero probe. |
| 3151 | */ |
| 3152 | void tcp_send_probe0(struct sock *sk) |
| 3153 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 3154 | struct inet_connection_sock *icsk = inet_csk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3155 | struct tcp_sock *tp = tcp_sk(sk); |
| 3156 | int err; |
| 3157 | |
| 3158 | err = tcp_write_wakeup(sk); |
| 3159 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 3160 | if (tp->packets_out || !tcp_send_head(sk)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3161 | /* Cancel probe timer, if it is not required. */ |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 3162 | icsk->icsk_probes_out = 0; |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 3163 | icsk->icsk_backoff = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3164 | return; |
| 3165 | } |
| 3166 | |
| 3167 | if (err <= 0) { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 3168 | if (icsk->icsk_backoff < sysctl_tcp_retries2) |
| 3169 | icsk->icsk_backoff++; |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 3170 | icsk->icsk_probes_out++; |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 3171 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, |
Arnaldo Carvalho de Melo | 3f421ba | 2005-08-09 20:11:08 -0700 | [diff] [blame] | 3172 | min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), |
| 3173 | TCP_RTO_MAX); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3174 | } else { |
| 3175 | /* If packet was not sent due to local congestion, |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 3176 | * do not backoff and do not remember icsk_probes_out. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3177 | * Let local senders to fight for local resources. |
| 3178 | * |
| 3179 | * Use accumulated backoff yet. |
| 3180 | */ |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 3181 | if (!icsk->icsk_probes_out) |
| 3182 | icsk->icsk_probes_out = 1; |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 3183 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 3184 | min(icsk->icsk_rto << icsk->icsk_backoff, |
Arnaldo Carvalho de Melo | 3f421ba | 2005-08-09 20:11:08 -0700 | [diff] [blame] | 3185 | TCP_RESOURCE_PROBE_INTERVAL), |
| 3186 | TCP_RTO_MAX); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3187 | } |
| 3188 | } |