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 | * |
| 8 | * Version: $Id: tcp_output.c,v 1.146 2002/02/01 22:01:04 davem Exp $ |
| 9 | * |
Jesper Juhl | 02c30a8 | 2005-05-05 16:16:16 -0700 | [diff] [blame] | 10 | * Authors: Ross Biro |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 11 | * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> |
| 12 | * Mark Evans, <evansmp@uhura.aston.ac.uk> |
| 13 | * Corey Minyard <wf-rch!minyard@relay.EU.net> |
| 14 | * Florian La Roche, <flla@stud.uni-sb.de> |
| 15 | * Charles Hedrick, <hedrick@klinzhai.rutgers.edu> |
| 16 | * Linus Torvalds, <torvalds@cs.helsinki.fi> |
| 17 | * Alan Cox, <gw4pts@gw4pts.ampr.org> |
| 18 | * Matthew Dillon, <dillon@apollo.west.oic.com> |
| 19 | * Arnt Gulbrandsen, <agulbra@nvg.unit.no> |
| 20 | * Jorge Cwik, <jorge@laser.satlink.net> |
| 21 | */ |
| 22 | |
| 23 | /* |
| 24 | * Changes: Pedro Roque : Retransmit queue handled by TCP. |
| 25 | * : Fragmentation on mtu decrease |
| 26 | * : Segment collapse on retransmit |
| 27 | * : AF independence |
| 28 | * |
| 29 | * Linus Torvalds : send_delayed_ack |
| 30 | * David S. Miller : Charge memory using the right skb |
| 31 | * during syn/ack processing. |
| 32 | * David S. Miller : Output engine completely rewritten. |
| 33 | * Andrea Arcangeli: SYNACK carry ts_recent in tsecr. |
| 34 | * Cacophonix Gaul : draft-minshall-nagle-01 |
| 35 | * J Hadi Salim : ECN support |
| 36 | * |
| 37 | */ |
| 38 | |
| 39 | #include <net/tcp.h> |
| 40 | |
| 41 | #include <linux/compiler.h> |
| 42 | #include <linux/module.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 43 | |
| 44 | /* 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] | 45 | int sysctl_tcp_retrans_collapse __read_mostly = 1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 46 | |
Rick Jones | 15d99e0 | 2006-03-20 22:40:29 -0800 | [diff] [blame] | 47 | /* People can turn this on to work with those rare, broken TCPs that |
| 48 | * interpret the window field as a signed quantity. |
| 49 | */ |
Brian Haley | ab32ea5 | 2006-09-22 14:15:41 -0700 | [diff] [blame] | 50 | int sysctl_tcp_workaround_signed_windows __read_mostly = 0; |
Rick Jones | 15d99e0 | 2006-03-20 22:40:29 -0800 | [diff] [blame] | 51 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 52 | /* This limits the percentage of the congestion window which we |
| 53 | * will allow a single TSO frame to consume. Building TSO frames |
| 54 | * which are too large can cause TCP streams to be bursty. |
| 55 | */ |
Brian Haley | ab32ea5 | 2006-09-22 14:15:41 -0700 | [diff] [blame] | 56 | int sysctl_tcp_tso_win_divisor __read_mostly = 3; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 57 | |
Brian Haley | ab32ea5 | 2006-09-22 14:15:41 -0700 | [diff] [blame] | 58 | int sysctl_tcp_mtu_probing __read_mostly = 0; |
| 59 | int sysctl_tcp_base_mss __read_mostly = 512; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 60 | |
David S. Miller | 35089bb | 2006-06-13 22:33:04 -0700 | [diff] [blame] | 61 | /* By default, RFC2861 behavior. */ |
Brian Haley | ab32ea5 | 2006-09-22 14:15:41 -0700 | [diff] [blame] | 62 | int sysctl_tcp_slow_start_after_idle __read_mostly = 1; |
David S. Miller | 35089bb | 2006-06-13 22:33:04 -0700 | [diff] [blame] | 63 | |
Ilpo Järvinen | 66f5fe6 | 2007-12-31 04:43:57 -0800 | [diff] [blame] | 64 | static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb) |
Ilpo Järvinen | 6ff03ac | 2007-08-24 22:44:06 -0700 | [diff] [blame] | 65 | { |
| 66 | struct tcp_sock *tp = tcp_sk(sk); |
Ilpo Järvinen | 66f5fe6 | 2007-12-31 04:43:57 -0800 | [diff] [blame] | 67 | unsigned int prior_packets = tp->packets_out; |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 68 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 69 | tcp_advance_send_head(sk, skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 70 | tp->snd_nxt = TCP_SKB_CB(skb)->end_seq; |
Ilpo Järvinen | 8512430 | 2007-11-26 20:17:38 +0800 | [diff] [blame] | 71 | |
| 72 | /* Don't override Nagle indefinately with F-RTO */ |
| 73 | if (tp->frto_counter == 2) |
| 74 | tp->frto_counter = 3; |
Ilpo Järvinen | 66f5fe6 | 2007-12-31 04:43:57 -0800 | [diff] [blame] | 75 | |
| 76 | tp->packets_out += tcp_skb_pcount(skb); |
| 77 | if (!prior_packets) |
| 78 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, |
| 79 | inet_csk(sk)->icsk_rto, TCP_RTO_MAX); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 80 | } |
| 81 | |
| 82 | /* SND.NXT, if window was not shrunk. |
| 83 | * If window has been shrunk, what should we make? It is not clear at all. |
| 84 | * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-( |
| 85 | * Anything in between SND.UNA...SND.UNA+SND.WND also can be already |
| 86 | * invalid. OK, let's make this for now: |
| 87 | */ |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 88 | static inline __u32 tcp_acceptable_seq(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 89 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 90 | struct tcp_sock *tp = tcp_sk(sk); |
| 91 | |
Ilpo Järvinen | 90840de | 2007-12-31 04:48:41 -0800 | [diff] [blame] | 92 | if (!before(tcp_wnd_end(tp), tp->snd_nxt)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 93 | return tp->snd_nxt; |
| 94 | else |
Ilpo Järvinen | 90840de | 2007-12-31 04:48:41 -0800 | [diff] [blame] | 95 | return tcp_wnd_end(tp); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 96 | } |
| 97 | |
| 98 | /* Calculate mss to advertise in SYN segment. |
| 99 | * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that: |
| 100 | * |
| 101 | * 1. It is independent of path mtu. |
| 102 | * 2. Ideally, it is maximal possible segment size i.e. 65535-40. |
| 103 | * 3. For IPv4 it is reasonable to calculate it from maximal MTU of |
| 104 | * attached devices, because some buggy hosts are confused by |
| 105 | * large MSS. |
| 106 | * 4. We do not make 3, we advertise MSS, calculated from first |
| 107 | * hop device mtu, but allow to raise it to ip_rt_min_advmss. |
| 108 | * This may be overridden via information stored in routing table. |
| 109 | * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible, |
| 110 | * probably even Jumbo". |
| 111 | */ |
| 112 | static __u16 tcp_advertise_mss(struct sock *sk) |
| 113 | { |
| 114 | struct tcp_sock *tp = tcp_sk(sk); |
| 115 | struct dst_entry *dst = __sk_dst_get(sk); |
| 116 | int mss = tp->advmss; |
| 117 | |
| 118 | if (dst && dst_metric(dst, RTAX_ADVMSS) < mss) { |
| 119 | mss = dst_metric(dst, RTAX_ADVMSS); |
| 120 | tp->advmss = mss; |
| 121 | } |
| 122 | |
| 123 | return (__u16)mss; |
| 124 | } |
| 125 | |
| 126 | /* RFC2861. Reset CWND after idle period longer RTO to "restart window". |
| 127 | * This is the first part of cwnd validation mechanism. */ |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 128 | static void tcp_cwnd_restart(struct sock *sk, struct dst_entry *dst) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 129 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 130 | struct tcp_sock *tp = tcp_sk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 131 | s32 delta = tcp_time_stamp - tp->lsndtime; |
| 132 | u32 restart_cwnd = tcp_init_cwnd(tp, dst); |
| 133 | u32 cwnd = tp->snd_cwnd; |
| 134 | |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 135 | tcp_ca_event(sk, CA_EVENT_CWND_RESTART); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 136 | |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 137 | tp->snd_ssthresh = tcp_current_ssthresh(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 138 | restart_cwnd = min(restart_cwnd, cwnd); |
| 139 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 140 | while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 141 | cwnd >>= 1; |
| 142 | tp->snd_cwnd = max(cwnd, restart_cwnd); |
| 143 | tp->snd_cwnd_stamp = tcp_time_stamp; |
| 144 | tp->snd_cwnd_used = 0; |
| 145 | } |
| 146 | |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 147 | static void tcp_event_data_sent(struct tcp_sock *tp, |
| 148 | struct sk_buff *skb, struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 149 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 150 | struct inet_connection_sock *icsk = inet_csk(sk); |
| 151 | const u32 now = tcp_time_stamp; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 152 | |
David S. Miller | 35089bb | 2006-06-13 22:33:04 -0700 | [diff] [blame] | 153 | if (sysctl_tcp_slow_start_after_idle && |
| 154 | (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto)) |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 155 | tcp_cwnd_restart(sk, __sk_dst_get(sk)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 156 | |
| 157 | tp->lsndtime = now; |
| 158 | |
| 159 | /* If it is a reply for ato after last received |
| 160 | * packet, enter pingpong mode. |
| 161 | */ |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 162 | if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato) |
| 163 | icsk->icsk_ack.pingpong = 1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 164 | } |
| 165 | |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 166 | 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] | 167 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 168 | tcp_dec_quickack_mode(sk, pkts); |
| 169 | inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 170 | } |
| 171 | |
| 172 | /* Determine a window scaling and initial window to offer. |
| 173 | * Based on the assumption that the given amount of space |
| 174 | * will be offered. Store the results in the tp structure. |
| 175 | * NOTE: for smooth operation initial space offering should |
| 176 | * be a multiple of mss if possible. We assume here that mss >= 1. |
| 177 | * This MUST be enforced by all callers. |
| 178 | */ |
| 179 | void tcp_select_initial_window(int __space, __u32 mss, |
| 180 | __u32 *rcv_wnd, __u32 *window_clamp, |
| 181 | int wscale_ok, __u8 *rcv_wscale) |
| 182 | { |
| 183 | unsigned int space = (__space < 0 ? 0 : __space); |
| 184 | |
| 185 | /* If no clamp set the clamp to the max possible scaled window */ |
| 186 | if (*window_clamp == 0) |
| 187 | (*window_clamp) = (65535 << 14); |
| 188 | space = min(*window_clamp, space); |
| 189 | |
| 190 | /* Quantize space offering to a multiple of mss if possible. */ |
| 191 | if (space > mss) |
| 192 | space = (space / mss) * mss; |
| 193 | |
| 194 | /* NOTE: offering an initial window larger than 32767 |
Rick Jones | 15d99e0 | 2006-03-20 22:40:29 -0800 | [diff] [blame] | 195 | * will break some buggy TCP stacks. If the admin tells us |
| 196 | * it is likely we could be speaking with such a buggy stack |
| 197 | * we will truncate our initial window offering to 32K-1 |
| 198 | * unless the remote has sent us a window scaling option, |
| 199 | * which we interpret as a sign the remote TCP is not |
| 200 | * misinterpreting the window field as a signed quantity. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 201 | */ |
Rick Jones | 15d99e0 | 2006-03-20 22:40:29 -0800 | [diff] [blame] | 202 | if (sysctl_tcp_workaround_signed_windows) |
| 203 | (*rcv_wnd) = min(space, MAX_TCP_WINDOW); |
| 204 | else |
| 205 | (*rcv_wnd) = space; |
| 206 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 207 | (*rcv_wscale) = 0; |
| 208 | if (wscale_ok) { |
| 209 | /* Set window scaling on max possible window |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 210 | * See RFC1323 for an explanation of the limit to 14 |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 211 | */ |
| 212 | space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max); |
Stephen Hemminger | 316c159 | 2006-08-22 00:06:11 -0700 | [diff] [blame] | 213 | space = min_t(u32, space, *window_clamp); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 214 | while (space > 65535 && (*rcv_wscale) < 14) { |
| 215 | space >>= 1; |
| 216 | (*rcv_wscale)++; |
| 217 | } |
| 218 | } |
| 219 | |
| 220 | /* Set initial window to value enough for senders, |
David S. Miller | 6b25185 | 2005-09-28 16:31:48 -0700 | [diff] [blame] | 221 | * following RFC2414. Senders, not following this RFC, |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 222 | * will be satisfied with 2. |
| 223 | */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 224 | if (mss > (1 << *rcv_wscale)) { |
David S. Miller | 01ff367 | 2005-09-29 17:07:20 -0700 | [diff] [blame] | 225 | int init_cwnd = 4; |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 226 | if (mss > 1460 * 3) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 227 | init_cwnd = 2; |
David S. Miller | 01ff367 | 2005-09-29 17:07:20 -0700 | [diff] [blame] | 228 | else if (mss > 1460) |
| 229 | init_cwnd = 3; |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 230 | if (*rcv_wnd > init_cwnd * mss) |
| 231 | *rcv_wnd = init_cwnd * mss; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 232 | } |
| 233 | |
| 234 | /* Set the clamp no higher than max representable value */ |
| 235 | (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp); |
| 236 | } |
| 237 | |
| 238 | /* Chose a new window to advertise, update state in tcp_sock for the |
| 239 | * socket, and return result with RFC1323 scaling applied. The return |
| 240 | * value can be stuffed directly into th->window for an outgoing |
| 241 | * frame. |
| 242 | */ |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 243 | static u16 tcp_select_window(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 244 | { |
| 245 | struct tcp_sock *tp = tcp_sk(sk); |
| 246 | u32 cur_win = tcp_receive_window(tp); |
| 247 | u32 new_win = __tcp_select_window(sk); |
| 248 | |
| 249 | /* Never shrink the offered window */ |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 250 | if (new_win < cur_win) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 251 | /* Danger Will Robinson! |
| 252 | * Don't update rcv_wup/rcv_wnd here or else |
| 253 | * we will not be able to advertise a zero |
| 254 | * window in time. --DaveM |
| 255 | * |
| 256 | * Relax Will Robinson. |
| 257 | */ |
Patrick McHardy | 607bfbf | 2008-03-20 16:11:27 -0700 | [diff] [blame] | 258 | new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 259 | } |
| 260 | tp->rcv_wnd = new_win; |
| 261 | tp->rcv_wup = tp->rcv_nxt; |
| 262 | |
| 263 | /* Make sure we do not exceed the maximum possible |
| 264 | * scaled window. |
| 265 | */ |
Rick Jones | 15d99e0 | 2006-03-20 22:40:29 -0800 | [diff] [blame] | 266 | if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 267 | new_win = min(new_win, MAX_TCP_WINDOW); |
| 268 | else |
| 269 | new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); |
| 270 | |
| 271 | /* RFC1323 scaling applied */ |
| 272 | new_win >>= tp->rx_opt.rcv_wscale; |
| 273 | |
| 274 | /* If we advertise zero window, disable fast path. */ |
| 275 | if (new_win == 0) |
| 276 | tp->pred_flags = 0; |
| 277 | |
| 278 | return new_win; |
| 279 | } |
| 280 | |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 281 | static inline void TCP_ECN_send_synack(struct tcp_sock *tp, struct sk_buff *skb) |
Ilpo Järvinen | bdf1ee5 | 2007-05-27 02:04:16 -0700 | [diff] [blame] | 282 | { |
| 283 | TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_CWR; |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 284 | if (!(tp->ecn_flags & TCP_ECN_OK)) |
Ilpo Järvinen | bdf1ee5 | 2007-05-27 02:04:16 -0700 | [diff] [blame] | 285 | TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_ECE; |
| 286 | } |
| 287 | |
| 288 | static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb) |
| 289 | { |
| 290 | struct tcp_sock *tp = tcp_sk(sk); |
| 291 | |
| 292 | tp->ecn_flags = 0; |
| 293 | if (sysctl_tcp_ecn) { |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 294 | TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ECE | TCPCB_FLAG_CWR; |
Ilpo Järvinen | bdf1ee5 | 2007-05-27 02:04:16 -0700 | [diff] [blame] | 295 | tp->ecn_flags = TCP_ECN_OK; |
| 296 | } |
| 297 | } |
| 298 | |
| 299 | static __inline__ void |
| 300 | TCP_ECN_make_synack(struct request_sock *req, struct tcphdr *th) |
| 301 | { |
| 302 | if (inet_rsk(req)->ecn_ok) |
| 303 | th->ece = 1; |
| 304 | } |
| 305 | |
| 306 | static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb, |
| 307 | int tcp_header_len) |
| 308 | { |
| 309 | struct tcp_sock *tp = tcp_sk(sk); |
| 310 | |
| 311 | if (tp->ecn_flags & TCP_ECN_OK) { |
| 312 | /* Not-retransmitted data segment: set ECT and inject CWR. */ |
| 313 | if (skb->len != tcp_header_len && |
| 314 | !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { |
| 315 | INET_ECN_xmit(sk); |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 316 | if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) { |
Ilpo Järvinen | bdf1ee5 | 2007-05-27 02:04:16 -0700 | [diff] [blame] | 317 | tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; |
| 318 | tcp_hdr(skb)->cwr = 1; |
| 319 | skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; |
| 320 | } |
| 321 | } else { |
| 322 | /* ACK or retransmitted segment: clear ECT|CE */ |
| 323 | INET_ECN_dontxmit(sk); |
| 324 | } |
| 325 | if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) |
| 326 | tcp_hdr(skb)->ece = 1; |
| 327 | } |
| 328 | } |
| 329 | |
Ilpo Järvinen | e870a8e | 2008-01-03 20:39:01 -0800 | [diff] [blame] | 330 | /* Constructs common control bits of non-data skb. If SYN/FIN is present, |
| 331 | * auto increment end seqno. |
| 332 | */ |
| 333 | static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags) |
| 334 | { |
| 335 | skb->csum = 0; |
| 336 | |
| 337 | TCP_SKB_CB(skb)->flags = flags; |
| 338 | TCP_SKB_CB(skb)->sacked = 0; |
| 339 | |
| 340 | skb_shinfo(skb)->gso_segs = 1; |
| 341 | skb_shinfo(skb)->gso_size = 0; |
| 342 | skb_shinfo(skb)->gso_type = 0; |
| 343 | |
| 344 | TCP_SKB_CB(skb)->seq = seq; |
| 345 | if (flags & (TCPCB_FLAG_SYN | TCPCB_FLAG_FIN)) |
| 346 | seq++; |
| 347 | TCP_SKB_CB(skb)->end_seq = seq; |
| 348 | } |
| 349 | |
Al Viro | df7a3b0 | 2006-09-27 18:38:52 -0700 | [diff] [blame] | 350 | static void tcp_build_and_update_options(__be32 *ptr, struct tcp_sock *tp, |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 351 | __u32 tstamp, __u8 **md5_hash) |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 352 | { |
| 353 | if (tp->rx_opt.tstamp_ok) { |
YOSHIFUJI Hideaki | 496c98d | 2006-10-10 19:41:21 -0700 | [diff] [blame] | 354 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
| 355 | (TCPOPT_NOP << 16) | |
| 356 | (TCPOPT_TIMESTAMP << 8) | |
| 357 | TCPOLEN_TIMESTAMP); |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 358 | *ptr++ = htonl(tstamp); |
| 359 | *ptr++ = htonl(tp->rx_opt.ts_recent); |
| 360 | } |
| 361 | if (tp->rx_opt.eff_sacks) { |
| 362 | struct tcp_sack_block *sp = tp->rx_opt.dsack ? tp->duplicate_sack : tp->selective_acks; |
| 363 | int this_sack; |
| 364 | |
| 365 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
| 366 | (TCPOPT_NOP << 16) | |
| 367 | (TCPOPT_SACK << 8) | |
| 368 | (TCPOLEN_SACK_BASE + (tp->rx_opt.eff_sacks * |
| 369 | TCPOLEN_SACK_PERBLOCK))); |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 370 | |
| 371 | for (this_sack = 0; this_sack < tp->rx_opt.eff_sacks; this_sack++) { |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 372 | *ptr++ = htonl(sp[this_sack].start_seq); |
| 373 | *ptr++ = htonl(sp[this_sack].end_seq); |
| 374 | } |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 375 | |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 376 | if (tp->rx_opt.dsack) { |
| 377 | tp->rx_opt.dsack = 0; |
| 378 | tp->rx_opt.eff_sacks--; |
| 379 | } |
| 380 | } |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 381 | #ifdef CONFIG_TCP_MD5SIG |
| 382 | if (md5_hash) { |
| 383 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
| 384 | (TCPOPT_NOP << 16) | |
| 385 | (TCPOPT_MD5SIG << 8) | |
| 386 | TCPOLEN_MD5SIG); |
| 387 | *md5_hash = (__u8 *)ptr; |
| 388 | } |
| 389 | #endif |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 390 | } |
| 391 | |
| 392 | /* Construct a tcp options header for a SYN or SYN_ACK packet. |
| 393 | * If this is every changed make sure to change the definition of |
| 394 | * MAX_SYN_SIZE to match the new maximum number of options that you |
| 395 | * can generate. |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 396 | * |
| 397 | * Note - that with the RFC2385 TCP option, we make room for the |
| 398 | * 16 byte MD5 hash. This will be filled in later, so the pointer for the |
| 399 | * location to be filled is passed back up. |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 400 | */ |
Al Viro | df7a3b0 | 2006-09-27 18:38:52 -0700 | [diff] [blame] | 401 | static void tcp_syn_build_options(__be32 *ptr, int mss, int ts, int sack, |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 402 | int offer_wscale, int wscale, __u32 tstamp, |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 403 | __u32 ts_recent, __u8 **md5_hash) |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 404 | { |
| 405 | /* We always get an MSS option. |
| 406 | * The option bytes which will be seen in normal data |
| 407 | * packets should timestamps be used, must be in the MSS |
| 408 | * advertised. But we subtract them from tp->mss_cache so |
| 409 | * that calculations in tcp_sendmsg are simpler etc. |
| 410 | * So account for this fact here if necessary. If we |
| 411 | * don't do this correctly, as a receiver we won't |
| 412 | * recognize data packets as being full sized when we |
| 413 | * should, and thus we won't abide by the delayed ACK |
| 414 | * rules correctly. |
| 415 | * SACKs don't matter, we never delay an ACK when we |
| 416 | * have any of those going out. |
| 417 | */ |
| 418 | *ptr++ = htonl((TCPOPT_MSS << 24) | (TCPOLEN_MSS << 16) | mss); |
| 419 | if (ts) { |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 420 | if (sack) |
YOSHIFUJI Hideaki | 496c98d | 2006-10-10 19:41:21 -0700 | [diff] [blame] | 421 | *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | |
| 422 | (TCPOLEN_SACK_PERM << 16) | |
| 423 | (TCPOPT_TIMESTAMP << 8) | |
| 424 | TCPOLEN_TIMESTAMP); |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 425 | else |
YOSHIFUJI Hideaki | 496c98d | 2006-10-10 19:41:21 -0700 | [diff] [blame] | 426 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
| 427 | (TCPOPT_NOP << 16) | |
| 428 | (TCPOPT_TIMESTAMP << 8) | |
| 429 | TCPOLEN_TIMESTAMP); |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 430 | *ptr++ = htonl(tstamp); /* TSVAL */ |
| 431 | *ptr++ = htonl(ts_recent); /* TSECR */ |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 432 | } else if (sack) |
YOSHIFUJI Hideaki | 496c98d | 2006-10-10 19:41:21 -0700 | [diff] [blame] | 433 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
| 434 | (TCPOPT_NOP << 16) | |
| 435 | (TCPOPT_SACK_PERM << 8) | |
| 436 | TCPOLEN_SACK_PERM); |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 437 | if (offer_wscale) |
YOSHIFUJI Hideaki | 496c98d | 2006-10-10 19:41:21 -0700 | [diff] [blame] | 438 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
| 439 | (TCPOPT_WINDOW << 16) | |
| 440 | (TCPOLEN_WINDOW << 8) | |
| 441 | (wscale)); |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 442 | #ifdef CONFIG_TCP_MD5SIG |
| 443 | /* |
| 444 | * If MD5 is enabled, then we set the option, and include the size |
| 445 | * (always 18). The actual MD5 hash is added just before the |
| 446 | * packet is sent. |
| 447 | */ |
| 448 | if (md5_hash) { |
| 449 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
| 450 | (TCPOPT_NOP << 16) | |
| 451 | (TCPOPT_MD5SIG << 8) | |
| 452 | TCPOLEN_MD5SIG); |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 453 | *md5_hash = (__u8 *)ptr; |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 454 | } |
| 455 | #endif |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 456 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 457 | |
| 458 | /* This routine actually transmits TCP packets queued in by |
| 459 | * tcp_do_sendmsg(). This is used by both the initial |
| 460 | * transmission and possible later retransmissions. |
| 461 | * All SKB's seen here are completely headerless. It is our |
| 462 | * job to build the TCP header, and pass the packet down to |
| 463 | * IP so it can do the same plus pass the packet off to the |
| 464 | * device. |
| 465 | * |
| 466 | * We are working here with either a clone of the original |
| 467 | * SKB, or a fresh unique copy made by the retransmit engine. |
| 468 | */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 469 | static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, |
| 470 | gfp_t gfp_mask) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 471 | { |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 472 | const struct inet_connection_sock *icsk = inet_csk(sk); |
| 473 | struct inet_sock *inet; |
| 474 | struct tcp_sock *tp; |
| 475 | struct tcp_skb_cb *tcb; |
| 476 | int tcp_header_size; |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 477 | #ifdef CONFIG_TCP_MD5SIG |
| 478 | struct tcp_md5sig_key *md5; |
| 479 | __u8 *md5_hash_location; |
| 480 | #endif |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 481 | struct tcphdr *th; |
| 482 | int sysctl_flags; |
| 483 | int err; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 484 | |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 485 | BUG_ON(!skb || !tcp_skb_pcount(skb)); |
| 486 | |
| 487 | /* If congestion control is doing timestamping, we must |
| 488 | * take such a timestamp before we potentially clone/copy. |
| 489 | */ |
Stephen Hemminger | 164891a | 2007-04-23 22:26:16 -0700 | [diff] [blame] | 490 | if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP) |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 491 | __net_timestamp(skb); |
| 492 | |
| 493 | if (likely(clone_it)) { |
| 494 | if (unlikely(skb_cloned(skb))) |
| 495 | skb = pskb_copy(skb, gfp_mask); |
| 496 | else |
| 497 | skb = skb_clone(skb, gfp_mask); |
| 498 | if (unlikely(!skb)) |
| 499 | return -ENOBUFS; |
| 500 | } |
| 501 | |
| 502 | inet = inet_sk(sk); |
| 503 | tp = tcp_sk(sk); |
| 504 | tcb = TCP_SKB_CB(skb); |
| 505 | tcp_header_size = tp->tcp_header_len; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 506 | |
| 507 | #define SYSCTL_FLAG_TSTAMPS 0x1 |
| 508 | #define SYSCTL_FLAG_WSCALE 0x2 |
| 509 | #define SYSCTL_FLAG_SACK 0x4 |
| 510 | |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 511 | sysctl_flags = 0; |
| 512 | if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) { |
| 513 | tcp_header_size = sizeof(struct tcphdr) + TCPOLEN_MSS; |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 514 | if (sysctl_tcp_timestamps) { |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 515 | tcp_header_size += TCPOLEN_TSTAMP_ALIGNED; |
| 516 | sysctl_flags |= SYSCTL_FLAG_TSTAMPS; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 517 | } |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 518 | if (sysctl_tcp_window_scaling) { |
| 519 | tcp_header_size += TCPOLEN_WSCALE_ALIGNED; |
| 520 | sysctl_flags |= SYSCTL_FLAG_WSCALE; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 521 | } |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 522 | if (sysctl_tcp_sack) { |
| 523 | sysctl_flags |= SYSCTL_FLAG_SACK; |
| 524 | if (!(sysctl_flags & SYSCTL_FLAG_TSTAMPS)) |
| 525 | tcp_header_size += TCPOLEN_SACKPERM_ALIGNED; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 526 | } |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 527 | } else if (unlikely(tp->rx_opt.eff_sacks)) { |
| 528 | /* A SACK is 2 pad bytes, a 2 byte header, plus |
| 529 | * 2 32-bit sequence numbers for each SACK block. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 530 | */ |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 531 | tcp_header_size += (TCPOLEN_SACK_BASE_ALIGNED + |
| 532 | (tp->rx_opt.eff_sacks * |
| 533 | TCPOLEN_SACK_PERBLOCK)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 534 | } |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 535 | |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 536 | if (tcp_packets_in_flight(tp) == 0) |
| 537 | tcp_ca_event(sk, CA_EVENT_TX_START); |
| 538 | |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 539 | #ifdef CONFIG_TCP_MD5SIG |
| 540 | /* |
| 541 | * Are we doing MD5 on this segment? If so - make |
| 542 | * room for it. |
| 543 | */ |
| 544 | md5 = tp->af_specific->md5_lookup(sk, sk); |
| 545 | if (md5) |
| 546 | tcp_header_size += TCPOLEN_MD5SIG_ALIGNED; |
| 547 | #endif |
| 548 | |
Arnaldo Carvalho de Melo | aa8223c | 2007-04-10 21:04:22 -0700 | [diff] [blame] | 549 | skb_push(skb, tcp_header_size); |
| 550 | skb_reset_transport_header(skb); |
David S. Miller | e89862f | 2007-01-26 01:04:55 -0800 | [diff] [blame] | 551 | skb_set_owner_w(skb, sk); |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 552 | |
| 553 | /* Build TCP header and checksum it. */ |
Arnaldo Carvalho de Melo | aa8223c | 2007-04-10 21:04:22 -0700 | [diff] [blame] | 554 | th = tcp_hdr(skb); |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 555 | th->source = inet->sport; |
| 556 | th->dest = inet->dport; |
| 557 | th->seq = htonl(tcb->seq); |
| 558 | th->ack_seq = htonl(tp->rcv_nxt); |
Al Viro | df7a3b0 | 2006-09-27 18:38:52 -0700 | [diff] [blame] | 559 | *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 560 | tcb->flags); |
| 561 | |
| 562 | if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) { |
| 563 | /* RFC1323: The window in SYN & SYN/ACK segments |
| 564 | * is never scaled. |
| 565 | */ |
Ilpo Järvinen | 600ff0c | 2007-02-13 12:42:11 -0800 | [diff] [blame] | 566 | th->window = htons(min(tp->rcv_wnd, 65535U)); |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 567 | } else { |
| 568 | th->window = htons(tcp_select_window(sk)); |
| 569 | } |
| 570 | th->check = 0; |
| 571 | th->urg_ptr = 0; |
| 572 | |
| 573 | if (unlikely(tp->urg_mode && |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 574 | between(tp->snd_up, tcb->seq + 1, tcb->seq + 0xFFFF))) { |
| 575 | th->urg_ptr = htons(tp->snd_up - tcb->seq); |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 576 | th->urg = 1; |
| 577 | } |
| 578 | |
| 579 | if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) { |
Al Viro | df7a3b0 | 2006-09-27 18:38:52 -0700 | [diff] [blame] | 580 | tcp_syn_build_options((__be32 *)(th + 1), |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 581 | tcp_advertise_mss(sk), |
| 582 | (sysctl_flags & SYSCTL_FLAG_TSTAMPS), |
| 583 | (sysctl_flags & SYSCTL_FLAG_SACK), |
| 584 | (sysctl_flags & SYSCTL_FLAG_WSCALE), |
| 585 | tp->rx_opt.rcv_wscale, |
| 586 | tcb->when, |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 587 | tp->rx_opt.ts_recent, |
| 588 | |
| 589 | #ifdef CONFIG_TCP_MD5SIG |
| 590 | md5 ? &md5_hash_location : |
| 591 | #endif |
| 592 | NULL); |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 593 | } else { |
Al Viro | df7a3b0 | 2006-09-27 18:38:52 -0700 | [diff] [blame] | 594 | tcp_build_and_update_options((__be32 *)(th + 1), |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 595 | tp, tcb->when, |
| 596 | #ifdef CONFIG_TCP_MD5SIG |
| 597 | md5 ? &md5_hash_location : |
| 598 | #endif |
| 599 | NULL); |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 600 | TCP_ECN_send(sk, skb, tcp_header_size); |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 601 | } |
| 602 | |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 603 | #ifdef CONFIG_TCP_MD5SIG |
| 604 | /* Calculate the MD5 hash, as we have all we need now */ |
| 605 | if (md5) { |
| 606 | tp->af_specific->calc_md5_hash(md5_hash_location, |
| 607 | md5, |
| 608 | sk, NULL, NULL, |
Arnaldo Carvalho de Melo | aa8223c | 2007-04-10 21:04:22 -0700 | [diff] [blame] | 609 | tcp_hdr(skb), |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 610 | sk->sk_protocol, |
| 611 | skb->len); |
| 612 | } |
| 613 | #endif |
| 614 | |
Arnaldo Carvalho de Melo | 8292a17 | 2005-12-13 23:15:52 -0800 | [diff] [blame] | 615 | icsk->icsk_af_ops->send_check(sk, skb->len, skb); |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 616 | |
| 617 | if (likely(tcb->flags & TCPCB_FLAG_ACK)) |
| 618 | tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); |
| 619 | |
| 620 | if (skb->len != tcp_header_size) |
| 621 | tcp_event_data_sent(tp, skb, sk); |
| 622 | |
Wei Yongjun | bd37a08 | 2006-08-07 21:04:15 -0700 | [diff] [blame] | 623 | if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) |
| 624 | TCP_INC_STATS(TCP_MIB_OUTSEGS); |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 625 | |
David S. Miller | e89862f | 2007-01-26 01:04:55 -0800 | [diff] [blame] | 626 | err = icsk->icsk_af_ops->queue_xmit(skb, 0); |
Hua Zhong | 83de47c | 2006-04-28 15:26:50 -0700 | [diff] [blame] | 627 | if (likely(err <= 0)) |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 628 | return err; |
| 629 | |
Ilpo Järvinen | 3cfe3ba | 2007-02-27 10:09:49 -0800 | [diff] [blame] | 630 | tcp_enter_cwr(sk, 1); |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 631 | |
Gerrit Renker | b9df3cb | 2006-11-14 11:21:36 -0200 | [diff] [blame] | 632 | return net_xmit_eval(err); |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 633 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 634 | #undef SYSCTL_FLAG_TSTAMPS |
| 635 | #undef SYSCTL_FLAG_WSCALE |
| 636 | #undef SYSCTL_FLAG_SACK |
| 637 | } |
| 638 | |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 639 | /* This routine just queue's the buffer |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 640 | * |
| 641 | * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, |
| 642 | * otherwise socket can stall. |
| 643 | */ |
| 644 | static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) |
| 645 | { |
| 646 | struct tcp_sock *tp = tcp_sk(sk); |
| 647 | |
| 648 | /* Advance write_seq and place onto the write_queue. */ |
| 649 | tp->write_seq = TCP_SKB_CB(skb)->end_seq; |
| 650 | skb_header_release(skb); |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 651 | tcp_add_write_queue_tail(sk, skb); |
Hideo Aoki | 3ab224b | 2007-12-31 00:11:19 -0800 | [diff] [blame] | 652 | sk->sk_wmem_queued += skb->truesize; |
| 653 | sk_mem_charge(sk, skb->truesize); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 654 | } |
| 655 | |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 656 | static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb, |
| 657 | unsigned int mss_now) |
David S. Miller | f6302d1 | 2005-07-05 15:18:03 -0700 | [diff] [blame] | 658 | { |
Herbert Xu | bcd7611 | 2006-06-30 13:36:35 -0700 | [diff] [blame] | 659 | if (skb->len <= mss_now || !sk_can_gso(sk)) { |
David S. Miller | f6302d1 | 2005-07-05 15:18:03 -0700 | [diff] [blame] | 660 | /* Avoid the costly divide in the normal |
| 661 | * non-TSO case. |
| 662 | */ |
Herbert Xu | 7967168 | 2006-06-22 02:40:14 -0700 | [diff] [blame] | 663 | skb_shinfo(skb)->gso_segs = 1; |
| 664 | skb_shinfo(skb)->gso_size = 0; |
| 665 | skb_shinfo(skb)->gso_type = 0; |
David S. Miller | f6302d1 | 2005-07-05 15:18:03 -0700 | [diff] [blame] | 666 | } else { |
Ilpo Järvinen | 356f89e | 2007-08-24 23:00:31 -0700 | [diff] [blame] | 667 | skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now); |
Herbert Xu | 7967168 | 2006-06-22 02:40:14 -0700 | [diff] [blame] | 668 | skb_shinfo(skb)->gso_size = mss_now; |
Herbert Xu | bcd7611 | 2006-06-30 13:36:35 -0700 | [diff] [blame] | 669 | skb_shinfo(skb)->gso_type = sk->sk_gso_type; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 670 | } |
| 671 | } |
| 672 | |
Ilpo Järvinen | 91fed7a | 2007-10-09 01:24:15 -0700 | [diff] [blame] | 673 | /* When a modification to fackets out becomes necessary, we need to check |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 674 | * skb is counted to fackets_out or not. |
Ilpo Järvinen | 91fed7a | 2007-10-09 01:24:15 -0700 | [diff] [blame] | 675 | */ |
Ilpo Järvinen | a47e5a9 | 2007-11-15 19:41:46 -0800 | [diff] [blame] | 676 | static void tcp_adjust_fackets_out(struct sock *sk, struct sk_buff *skb, |
Ilpo Järvinen | 91fed7a | 2007-10-09 01:24:15 -0700 | [diff] [blame] | 677 | int decr) |
| 678 | { |
Ilpo Järvinen | a47e5a9 | 2007-11-15 19:41:46 -0800 | [diff] [blame] | 679 | struct tcp_sock *tp = tcp_sk(sk); |
| 680 | |
Ilpo Järvinen | dc86967 | 2007-10-01 15:27:19 -0700 | [diff] [blame] | 681 | if (!tp->sacked_out || tcp_is_reno(tp)) |
Ilpo Järvinen | 91fed7a | 2007-10-09 01:24:15 -0700 | [diff] [blame] | 682 | return; |
| 683 | |
Ilpo Järvinen | 6859d49 | 2007-12-02 00:48:06 +0200 | [diff] [blame] | 684 | if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq)) |
Ilpo Järvinen | 91fed7a | 2007-10-09 01:24:15 -0700 | [diff] [blame] | 685 | tp->fackets_out -= decr; |
Ilpo Järvinen | 91fed7a | 2007-10-09 01:24:15 -0700 | [diff] [blame] | 686 | } |
| 687 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 688 | /* Function to create two new TCP segments. Shrinks the given segment |
| 689 | * 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] | 690 | * packet to the list. This won't be called frequently, I hope. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 691 | * Remember, these are still headerless SKBs at this point. |
| 692 | */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 693 | int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, |
| 694 | unsigned int mss_now) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 695 | { |
| 696 | struct tcp_sock *tp = tcp_sk(sk); |
| 697 | struct sk_buff *buff; |
David S. Miller | 6475be1 | 2005-09-01 22:47:01 -0700 | [diff] [blame] | 698 | int nsize, old_factor; |
Herbert Xu | b60b49e | 2006-04-19 21:35:00 -0700 | [diff] [blame] | 699 | int nlen; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 700 | u16 flags; |
| 701 | |
Herbert Xu | b2cc99f0 | 2005-10-20 17:13:13 -0200 | [diff] [blame] | 702 | BUG_ON(len > skb->len); |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 703 | |
Ilpo Järvinen | b768920 | 2007-09-20 11:37:19 -0700 | [diff] [blame] | 704 | tcp_clear_retrans_hints_partial(tp); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 705 | nsize = skb_headlen(skb) - len; |
| 706 | if (nsize < 0) |
| 707 | nsize = 0; |
| 708 | |
| 709 | if (skb_cloned(skb) && |
| 710 | skb_is_nonlinear(skb) && |
| 711 | pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) |
| 712 | return -ENOMEM; |
| 713 | |
| 714 | /* Get a new skb... force flag on. */ |
| 715 | buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC); |
| 716 | if (buff == NULL) |
| 717 | return -ENOMEM; /* We'll just try again later. */ |
Herbert Xu | ef5cb97 | 2006-04-18 13:24:14 -0700 | [diff] [blame] | 718 | |
Hideo Aoki | 3ab224b | 2007-12-31 00:11:19 -0800 | [diff] [blame] | 719 | sk->sk_wmem_queued += buff->truesize; |
| 720 | sk_mem_charge(sk, buff->truesize); |
Herbert Xu | b60b49e | 2006-04-19 21:35:00 -0700 | [diff] [blame] | 721 | nlen = skb->len - len - nsize; |
| 722 | buff->truesize += nlen; |
| 723 | skb->truesize -= nlen; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 724 | |
| 725 | /* Correct the sequence numbers. */ |
| 726 | TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; |
| 727 | TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; |
| 728 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; |
| 729 | |
| 730 | /* PSH and FIN should only be set in the second packet. */ |
| 731 | flags = TCP_SKB_CB(skb)->flags; |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 732 | TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN | TCPCB_FLAG_PSH); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 733 | TCP_SKB_CB(buff)->flags = flags; |
Herbert Xu | e14c3ca | 2005-09-19 18:18:38 -0700 | [diff] [blame] | 734 | TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 735 | |
Patrick McHardy | 84fa793 | 2006-08-29 16:44:56 -0700 | [diff] [blame] | 736 | if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 737 | /* Copy and checksum data tail into the new buffer. */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 738 | buff->csum = csum_partial_copy_nocheck(skb->data + len, |
| 739 | skb_put(buff, nsize), |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 740 | nsize, 0); |
| 741 | |
| 742 | skb_trim(skb, len); |
| 743 | |
| 744 | skb->csum = csum_block_sub(skb->csum, buff->csum, len); |
| 745 | } else { |
Patrick McHardy | 84fa793 | 2006-08-29 16:44:56 -0700 | [diff] [blame] | 746 | skb->ip_summed = CHECKSUM_PARTIAL; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 747 | skb_split(skb, buff, len); |
| 748 | } |
| 749 | |
| 750 | buff->ip_summed = skb->ip_summed; |
| 751 | |
| 752 | /* Looks stupid, but our code really uses when of |
| 753 | * skbs, which it never sent before. --ANK |
| 754 | */ |
| 755 | TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when; |
Patrick McHardy | a61bbcf | 2005-08-14 17:24:31 -0700 | [diff] [blame] | 756 | buff->tstamp = skb->tstamp; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 757 | |
David S. Miller | 6475be1 | 2005-09-01 22:47:01 -0700 | [diff] [blame] | 758 | old_factor = tcp_skb_pcount(skb); |
| 759 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 760 | /* Fix up tso_factor for both original and new SKB. */ |
David S. Miller | 846998a | 2005-08-04 19:52:01 -0700 | [diff] [blame] | 761 | tcp_set_skb_tso_segs(sk, skb, mss_now); |
| 762 | tcp_set_skb_tso_segs(sk, buff, mss_now); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 763 | |
David S. Miller | 6475be1 | 2005-09-01 22:47:01 -0700 | [diff] [blame] | 764 | /* If this packet has been sent out already, we must |
| 765 | * adjust the various packet counters. |
| 766 | */ |
Herbert Xu | cf0b450 | 2005-09-08 15:10:52 -0700 | [diff] [blame] | 767 | if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { |
David S. Miller | 6475be1 | 2005-09-01 22:47:01 -0700 | [diff] [blame] | 768 | int diff = old_factor - tcp_skb_pcount(skb) - |
| 769 | tcp_skb_pcount(buff); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 770 | |
David S. Miller | 6475be1 | 2005-09-01 22:47:01 -0700 | [diff] [blame] | 771 | tp->packets_out -= diff; |
Herbert Xu | e14c3ca | 2005-09-19 18:18:38 -0700 | [diff] [blame] | 772 | |
| 773 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) |
| 774 | tp->sacked_out -= diff; |
| 775 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) |
| 776 | tp->retrans_out -= diff; |
| 777 | |
Ilpo Järvinen | b5860bb | 2007-08-09 14:33:18 +0300 | [diff] [blame] | 778 | if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) |
David S. Miller | 6475be1 | 2005-09-01 22:47:01 -0700 | [diff] [blame] | 779 | tp->lost_out -= diff; |
Herbert Xu | 83ca28b | 2005-09-22 23:32:56 -0700 | [diff] [blame] | 780 | |
Ilpo Järvinen | 91fed7a | 2007-10-09 01:24:15 -0700 | [diff] [blame] | 781 | /* Adjust Reno SACK estimate. */ |
| 782 | if (tcp_is_reno(tp) && diff > 0) { |
| 783 | tcp_dec_pcount_approx_int(&tp->sacked_out, diff); |
| 784 | tcp_verify_left_out(tp); |
David S. Miller | 6475be1 | 2005-09-01 22:47:01 -0700 | [diff] [blame] | 785 | } |
Ilpo Järvinen | a47e5a9 | 2007-11-15 19:41:46 -0800 | [diff] [blame] | 786 | tcp_adjust_fackets_out(sk, skb, diff); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 787 | } |
| 788 | |
| 789 | /* Link BUFF into the send queue. */ |
David S. Miller | f44b527 | 2005-07-05 15:18:34 -0700 | [diff] [blame] | 790 | skb_header_release(buff); |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 791 | tcp_insert_write_queue_after(skb, buff, sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 792 | |
| 793 | return 0; |
| 794 | } |
| 795 | |
| 796 | /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c |
| 797 | * eventually). The difference is that pulled data not copied, but |
| 798 | * immediately discarded. |
| 799 | */ |
Herbert Xu ~{PmVHI~} | f291196 | 2006-06-05 15:03:37 -0700 | [diff] [blame] | 800 | static void __pskb_trim_head(struct sk_buff *skb, int len) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 801 | { |
| 802 | int i, k, eat; |
| 803 | |
| 804 | eat = len; |
| 805 | k = 0; |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 806 | for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 807 | if (skb_shinfo(skb)->frags[i].size <= eat) { |
| 808 | put_page(skb_shinfo(skb)->frags[i].page); |
| 809 | eat -= skb_shinfo(skb)->frags[i].size; |
| 810 | } else { |
| 811 | skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i]; |
| 812 | if (eat) { |
| 813 | skb_shinfo(skb)->frags[k].page_offset += eat; |
| 814 | skb_shinfo(skb)->frags[k].size -= eat; |
| 815 | eat = 0; |
| 816 | } |
| 817 | k++; |
| 818 | } |
| 819 | } |
| 820 | skb_shinfo(skb)->nr_frags = k; |
| 821 | |
Arnaldo Carvalho de Melo | 27a884d | 2007-04-19 20:29:13 -0700 | [diff] [blame] | 822 | skb_reset_tail_pointer(skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 823 | skb->data_len -= len; |
| 824 | skb->len = skb->data_len; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 825 | } |
| 826 | |
| 827 | int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) |
| 828 | { |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 829 | if (skb_cloned(skb) && pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 830 | return -ENOMEM; |
| 831 | |
Herbert Xu ~{PmVHI~} | f291196 | 2006-06-05 15:03:37 -0700 | [diff] [blame] | 832 | /* If len == headlen, we avoid __skb_pull to preserve alignment. */ |
| 833 | if (unlikely(len < skb_headlen(skb))) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 834 | __skb_pull(skb, len); |
Herbert Xu ~{PmVHI~} | f291196 | 2006-06-05 15:03:37 -0700 | [diff] [blame] | 835 | else |
| 836 | __pskb_trim_head(skb, len - skb_headlen(skb)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 837 | |
| 838 | TCP_SKB_CB(skb)->seq += len; |
Patrick McHardy | 84fa793 | 2006-08-29 16:44:56 -0700 | [diff] [blame] | 839 | skb->ip_summed = CHECKSUM_PARTIAL; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 840 | |
| 841 | skb->truesize -= len; |
| 842 | sk->sk_wmem_queued -= len; |
Hideo Aoki | 3ab224b | 2007-12-31 00:11:19 -0800 | [diff] [blame] | 843 | sk_mem_uncharge(sk, len); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 844 | sock_set_flag(sk, SOCK_QUEUE_SHRUNK); |
| 845 | |
| 846 | /* Any change of skb->len requires recalculation of tso |
| 847 | * factor and mss. |
| 848 | */ |
| 849 | if (tcp_skb_pcount(skb) > 1) |
David S. Miller | 846998a | 2005-08-04 19:52:01 -0700 | [diff] [blame] | 850 | tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk, 1)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 851 | |
| 852 | return 0; |
| 853 | } |
| 854 | |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 855 | /* Not accounting for SACKs here. */ |
| 856 | int tcp_mtu_to_mss(struct sock *sk, int pmtu) |
| 857 | { |
| 858 | struct tcp_sock *tp = tcp_sk(sk); |
| 859 | struct inet_connection_sock *icsk = inet_csk(sk); |
| 860 | int mss_now; |
| 861 | |
| 862 | /* Calculate base mss without TCP options: |
| 863 | It is MMS_S - sizeof(tcphdr) of rfc1122 |
| 864 | */ |
| 865 | mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); |
| 866 | |
| 867 | /* Clamp it (mss_clamp does not include tcp options) */ |
| 868 | if (mss_now > tp->rx_opt.mss_clamp) |
| 869 | mss_now = tp->rx_opt.mss_clamp; |
| 870 | |
| 871 | /* Now subtract optional transport overhead */ |
| 872 | mss_now -= icsk->icsk_ext_hdr_len; |
| 873 | |
| 874 | /* Then reserve room for full set of TCP options and 8 bytes of data */ |
| 875 | if (mss_now < 48) |
| 876 | mss_now = 48; |
| 877 | |
| 878 | /* Now subtract TCP options size, not including SACKs */ |
| 879 | mss_now -= tp->tcp_header_len - sizeof(struct tcphdr); |
| 880 | |
| 881 | return mss_now; |
| 882 | } |
| 883 | |
| 884 | /* Inverse of above */ |
| 885 | int tcp_mss_to_mtu(struct sock *sk, int mss) |
| 886 | { |
| 887 | struct tcp_sock *tp = tcp_sk(sk); |
| 888 | struct inet_connection_sock *icsk = inet_csk(sk); |
| 889 | int mtu; |
| 890 | |
| 891 | mtu = mss + |
| 892 | tp->tcp_header_len + |
| 893 | icsk->icsk_ext_hdr_len + |
| 894 | icsk->icsk_af_ops->net_header_len; |
| 895 | |
| 896 | return mtu; |
| 897 | } |
| 898 | |
| 899 | void tcp_mtup_init(struct sock *sk) |
| 900 | { |
| 901 | struct tcp_sock *tp = tcp_sk(sk); |
| 902 | struct inet_connection_sock *icsk = inet_csk(sk); |
| 903 | |
| 904 | icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1; |
| 905 | 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] | 906 | icsk->icsk_af_ops->net_header_len; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 907 | icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss); |
| 908 | icsk->icsk_mtup.probe_size = 0; |
| 909 | } |
| 910 | |
Ilpo Järvinen | 409d22b | 2007-12-31 14:57:40 -0800 | [diff] [blame] | 911 | /* Bound MSS / TSO packet size with the half of the window */ |
| 912 | static int tcp_bound_to_half_wnd(struct tcp_sock *tp, int pktsize) |
| 913 | { |
| 914 | if (tp->max_window && pktsize > (tp->max_window >> 1)) |
| 915 | return max(tp->max_window >> 1, 68U - tp->tcp_header_len); |
| 916 | else |
| 917 | return pktsize; |
| 918 | } |
| 919 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 920 | /* This function synchronize snd mss to current pmtu/exthdr set. |
| 921 | |
| 922 | tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts |
| 923 | for TCP options, but includes only bare TCP header. |
| 924 | |
| 925 | tp->rx_opt.mss_clamp is mss negotiated at connection setup. |
Stephen Hemminger | caa20d9a | 2005-11-10 17:13:47 -0800 | [diff] [blame] | 926 | It is minimum of user_mss and mss received with SYN. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 927 | It also does not include TCP options. |
| 928 | |
Arnaldo Carvalho de Melo | d83d846 | 2005-12-13 23:26:10 -0800 | [diff] [blame] | 929 | 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] | 930 | |
| 931 | tp->mss_cache is current effective sending mss, including |
| 932 | all tcp options except for SACKs. It is evaluated, |
| 933 | taking into account current pmtu, but never exceeds |
| 934 | tp->rx_opt.mss_clamp. |
| 935 | |
| 936 | NOTE1. rfc1122 clearly states that advertised MSS |
| 937 | DOES NOT include either tcp or ip options. |
| 938 | |
Arnaldo Carvalho de Melo | d83d846 | 2005-12-13 23:26:10 -0800 | [diff] [blame] | 939 | NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache |
| 940 | are READ ONLY outside this function. --ANK (980731) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 941 | */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 942 | unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) |
| 943 | { |
| 944 | struct tcp_sock *tp = tcp_sk(sk); |
Arnaldo Carvalho de Melo | d83d846 | 2005-12-13 23:26:10 -0800 | [diff] [blame] | 945 | struct inet_connection_sock *icsk = inet_csk(sk); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 946 | int mss_now; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 947 | |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 948 | if (icsk->icsk_mtup.search_high > pmtu) |
| 949 | icsk->icsk_mtup.search_high = pmtu; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 950 | |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 951 | mss_now = tcp_mtu_to_mss(sk, pmtu); |
Ilpo Järvinen | 409d22b | 2007-12-31 14:57:40 -0800 | [diff] [blame] | 952 | mss_now = tcp_bound_to_half_wnd(tp, mss_now); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 953 | |
| 954 | /* And store cached results */ |
Arnaldo Carvalho de Melo | d83d846 | 2005-12-13 23:26:10 -0800 | [diff] [blame] | 955 | icsk->icsk_pmtu_cookie = pmtu; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 956 | if (icsk->icsk_mtup.enabled) |
| 957 | 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] | 958 | tp->mss_cache = mss_now; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 959 | |
| 960 | return mss_now; |
| 961 | } |
| 962 | |
| 963 | /* Compute the current effective MSS, taking SACKs and IP options, |
| 964 | * and even PMTU discovery events into account. |
| 965 | * |
| 966 | * LARGESEND note: !urg_mode is overkill, only frames up to snd_up |
| 967 | * cannot be large. However, taking into account rare use of URG, this |
| 968 | * is not a big flaw. |
| 969 | */ |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 970 | unsigned int tcp_current_mss(struct sock *sk, int large_allowed) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 971 | { |
| 972 | struct tcp_sock *tp = tcp_sk(sk); |
| 973 | struct dst_entry *dst = __sk_dst_get(sk); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 974 | u32 mss_now; |
| 975 | u16 xmit_size_goal; |
| 976 | int doing_tso = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 977 | |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 978 | mss_now = tp->mss_cache; |
| 979 | |
Herbert Xu | bcd7611 | 2006-06-30 13:36:35 -0700 | [diff] [blame] | 980 | if (large_allowed && sk_can_gso(sk) && !tp->urg_mode) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 981 | doing_tso = 1; |
| 982 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 983 | if (dst) { |
| 984 | u32 mtu = dst_mtu(dst); |
Arnaldo Carvalho de Melo | d83d846 | 2005-12-13 23:26:10 -0800 | [diff] [blame] | 985 | if (mtu != inet_csk(sk)->icsk_pmtu_cookie) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 986 | mss_now = tcp_sync_mss(sk, mtu); |
| 987 | } |
| 988 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 989 | if (tp->rx_opt.eff_sacks) |
| 990 | mss_now -= (TCPOLEN_SACK_BASE_ALIGNED + |
| 991 | (tp->rx_opt.eff_sacks * TCPOLEN_SACK_PERBLOCK)); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 992 | |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 993 | #ifdef CONFIG_TCP_MD5SIG |
| 994 | if (tp->af_specific->md5_lookup(sk, sk)) |
| 995 | mss_now -= TCPOLEN_MD5SIG_ALIGNED; |
| 996 | #endif |
| 997 | |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 998 | xmit_size_goal = mss_now; |
| 999 | |
| 1000 | if (doing_tso) { |
Peter P Waskiewicz Jr | 82cc1a7 | 2008-03-21 03:43:19 -0700 | [diff] [blame] | 1001 | xmit_size_goal = ((sk->sk_gso_max_size - 1) - |
Arnaldo Carvalho de Melo | 8292a17 | 2005-12-13 23:15:52 -0800 | [diff] [blame] | 1002 | inet_csk(sk)->icsk_af_ops->net_header_len - |
Arnaldo Carvalho de Melo | d83d846 | 2005-12-13 23:26:10 -0800 | [diff] [blame] | 1003 | inet_csk(sk)->icsk_ext_hdr_len - |
| 1004 | tp->tcp_header_len); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1005 | |
Ilpo Järvinen | 409d22b | 2007-12-31 14:57:40 -0800 | [diff] [blame] | 1006 | xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1007 | xmit_size_goal -= (xmit_size_goal % mss_now); |
| 1008 | } |
| 1009 | tp->xmit_size_goal = xmit_size_goal; |
| 1010 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1011 | return mss_now; |
| 1012 | } |
| 1013 | |
David S. Miller | a762a98 | 2005-07-05 15:18:51 -0700 | [diff] [blame] | 1014 | /* Congestion window validation. (RFC2861) */ |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 1015 | static void tcp_cwnd_validate(struct sock *sk) |
David S. Miller | a762a98 | 2005-07-05 15:18:51 -0700 | [diff] [blame] | 1016 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 1017 | struct tcp_sock *tp = tcp_sk(sk); |
David S. Miller | a762a98 | 2005-07-05 15:18:51 -0700 | [diff] [blame] | 1018 | |
Ilpo Järvinen | d436d68 | 2007-12-31 14:58:00 -0800 | [diff] [blame] | 1019 | if (tp->packets_out >= tp->snd_cwnd) { |
David S. Miller | a762a98 | 2005-07-05 15:18:51 -0700 | [diff] [blame] | 1020 | /* Network is feed fully. */ |
| 1021 | tp->snd_cwnd_used = 0; |
| 1022 | tp->snd_cwnd_stamp = tcp_time_stamp; |
| 1023 | } else { |
| 1024 | /* Network starves. */ |
| 1025 | if (tp->packets_out > tp->snd_cwnd_used) |
| 1026 | tp->snd_cwnd_used = tp->packets_out; |
| 1027 | |
David S. Miller | 15d33c0 | 2007-04-09 13:23:14 -0700 | [diff] [blame] | 1028 | if (sysctl_tcp_slow_start_after_idle && |
| 1029 | (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] | 1030 | tcp_cwnd_application_limited(sk); |
| 1031 | } |
| 1032 | } |
| 1033 | |
Ilpo Järvinen | 0e3a480 | 2007-12-24 21:33:45 -0800 | [diff] [blame] | 1034 | /* Returns the portion of skb which can be sent right away without |
| 1035 | * introducing MSS oddities to segment boundaries. In rare cases where |
| 1036 | * mss_now != mss_cache, we will request caller to create a small skb |
| 1037 | * per input skb which could be mostly avoided here (if desired). |
Ilpo Järvinen | 5ea3a74 | 2008-03-11 17:55:27 -0700 | [diff] [blame] | 1038 | * |
| 1039 | * We explicitly want to create a request for splitting write queue tail |
| 1040 | * to a small skb for Nagle purposes while avoiding unnecessary modulos, |
| 1041 | * thus all the complexity (cwnd_len is always MSS multiple which we |
| 1042 | * return whenever allowed by the other factors). Basically we need the |
| 1043 | * modulo only when the receiver window alone is the limiting factor or |
| 1044 | * 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] | 1045 | */ |
| 1046 | static unsigned int tcp_mss_split_point(struct sock *sk, struct sk_buff *skb, |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1047 | unsigned int mss_now, unsigned int cwnd) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1048 | { |
Ilpo Järvinen | 0e3a480 | 2007-12-24 21:33:45 -0800 | [diff] [blame] | 1049 | struct tcp_sock *tp = tcp_sk(sk); |
| 1050 | u32 needed, window, cwnd_len; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1051 | |
Ilpo Järvinen | 90840de | 2007-12-31 04:48:41 -0800 | [diff] [blame] | 1052 | window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1053 | cwnd_len = mss_now * cwnd; |
Ilpo Järvinen | 0e3a480 | 2007-12-24 21:33:45 -0800 | [diff] [blame] | 1054 | |
| 1055 | if (likely(cwnd_len <= window && skb != tcp_write_queue_tail(sk))) |
| 1056 | return cwnd_len; |
| 1057 | |
Ilpo Järvinen | 5ea3a74 | 2008-03-11 17:55:27 -0700 | [diff] [blame] | 1058 | needed = min(skb->len, window); |
| 1059 | |
Ilpo Järvinen | 1751540 | 2008-04-15 20:36:55 -0700 | [diff] [blame] | 1060 | if (cwnd_len <= needed) |
Ilpo Järvinen | 0e3a480 | 2007-12-24 21:33:45 -0800 | [diff] [blame] | 1061 | return cwnd_len; |
| 1062 | |
Ilpo Järvinen | 0e3a480 | 2007-12-24 21:33:45 -0800 | [diff] [blame] | 1063 | return needed - needed % mss_now; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1064 | } |
| 1065 | |
| 1066 | /* Can at least one segment of SKB be sent right now, according to the |
| 1067 | * congestion window rules? If so, return how many segments are allowed. |
| 1068 | */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1069 | static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp, |
| 1070 | struct sk_buff *skb) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1071 | { |
| 1072 | u32 in_flight, cwnd; |
| 1073 | |
| 1074 | /* Don't be strict about the congestion window for the final FIN. */ |
John Heffner | 104439a | 2007-02-05 17:53:11 -0800 | [diff] [blame] | 1075 | if ((TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) && |
| 1076 | tcp_skb_pcount(skb) == 1) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1077 | return 1; |
| 1078 | |
| 1079 | in_flight = tcp_packets_in_flight(tp); |
| 1080 | cwnd = tp->snd_cwnd; |
| 1081 | if (in_flight < cwnd) |
| 1082 | return (cwnd - in_flight); |
| 1083 | |
| 1084 | return 0; |
| 1085 | } |
| 1086 | |
| 1087 | /* This must be invoked the first time we consider transmitting |
| 1088 | * SKB onto the wire. |
| 1089 | */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1090 | static int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb, |
| 1091 | unsigned int mss_now) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1092 | { |
| 1093 | int tso_segs = tcp_skb_pcount(skb); |
| 1094 | |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1095 | 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] | 1096 | tcp_set_skb_tso_segs(sk, skb, mss_now); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1097 | tso_segs = tcp_skb_pcount(skb); |
| 1098 | } |
| 1099 | return tso_segs; |
| 1100 | } |
| 1101 | |
| 1102 | static inline int tcp_minshall_check(const struct tcp_sock *tp) |
| 1103 | { |
| 1104 | return after(tp->snd_sml,tp->snd_una) && |
| 1105 | !after(tp->snd_sml, tp->snd_nxt); |
| 1106 | } |
| 1107 | |
| 1108 | /* Return 0, if packet can be sent now without violation Nagle's rules: |
| 1109 | * 1. It is full sized. |
| 1110 | * 2. Or it contains FIN. (already checked by caller) |
| 1111 | * 3. Or TCP_NODELAY was set. |
| 1112 | * 4. Or TCP_CORK is not set, and all sent packets are ACKed. |
| 1113 | * With Minshall's modification: all sent small packets are ACKed. |
| 1114 | */ |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1115 | static inline int tcp_nagle_check(const struct tcp_sock *tp, |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 1116 | const struct sk_buff *skb, |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1117 | unsigned mss_now, int nonagle) |
| 1118 | { |
| 1119 | return (skb->len < mss_now && |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1120 | ((nonagle & TCP_NAGLE_CORK) || |
| 1121 | (!nonagle && tp->packets_out && tcp_minshall_check(tp)))); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1122 | } |
| 1123 | |
| 1124 | /* Return non-zero if the Nagle test allows this packet to be |
| 1125 | * sent now. |
| 1126 | */ |
| 1127 | static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb, |
| 1128 | unsigned int cur_mss, int nonagle) |
| 1129 | { |
| 1130 | /* Nagle rule does not apply to frames, which sit in the middle of the |
| 1131 | * write_queue (they have no chances to get new data). |
| 1132 | * |
| 1133 | * This is implemented in the callers, where they modify the 'nonagle' |
| 1134 | * argument based upon the location of SKB in the send queue. |
| 1135 | */ |
| 1136 | if (nonagle & TCP_NAGLE_PUSH) |
| 1137 | return 1; |
| 1138 | |
Ilpo Järvinen | d551e45 | 2007-04-30 00:42:20 -0700 | [diff] [blame] | 1139 | /* Don't use the nagle rule for urgent data (or for the final FIN). |
| 1140 | * Nagle can be ignored during F-RTO too (see RFC4138). |
| 1141 | */ |
| 1142 | if (tp->urg_mode || (tp->frto_counter == 2) || |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1143 | (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)) |
| 1144 | return 1; |
| 1145 | |
| 1146 | if (!tcp_nagle_check(tp, skb, cur_mss, nonagle)) |
| 1147 | return 1; |
| 1148 | |
| 1149 | return 0; |
| 1150 | } |
| 1151 | |
| 1152 | /* Does at least the first segment of SKB fit into the send window? */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1153 | static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb, |
| 1154 | unsigned int cur_mss) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1155 | { |
| 1156 | u32 end_seq = TCP_SKB_CB(skb)->end_seq; |
| 1157 | |
| 1158 | if (skb->len > cur_mss) |
| 1159 | end_seq = TCP_SKB_CB(skb)->seq + cur_mss; |
| 1160 | |
Ilpo Järvinen | 90840de | 2007-12-31 04:48:41 -0800 | [diff] [blame] | 1161 | return !after(end_seq, tcp_wnd_end(tp)); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1162 | } |
| 1163 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1164 | /* 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] | 1165 | * should be put on the wire right now. If so, it returns the number of |
| 1166 | * packets allowed by the congestion window. |
| 1167 | */ |
| 1168 | static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb, |
| 1169 | unsigned int cur_mss, int nonagle) |
| 1170 | { |
| 1171 | struct tcp_sock *tp = tcp_sk(sk); |
| 1172 | unsigned int cwnd_quota; |
| 1173 | |
David S. Miller | 846998a | 2005-08-04 19:52:01 -0700 | [diff] [blame] | 1174 | tcp_init_tso_segs(sk, skb, cur_mss); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1175 | |
| 1176 | if (!tcp_nagle_test(tp, skb, cur_mss, nonagle)) |
| 1177 | return 0; |
| 1178 | |
| 1179 | cwnd_quota = tcp_cwnd_test(tp, skb); |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1180 | if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss)) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1181 | cwnd_quota = 0; |
| 1182 | |
| 1183 | return cwnd_quota; |
| 1184 | } |
| 1185 | |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 1186 | int tcp_may_send_now(struct sock *sk) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1187 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 1188 | struct tcp_sock *tp = tcp_sk(sk); |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1189 | struct sk_buff *skb = tcp_send_head(sk); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1190 | |
| 1191 | return (skb && |
| 1192 | tcp_snd_test(sk, skb, tcp_current_mss(sk, 1), |
| 1193 | (tcp_skb_is_last(sk, skb) ? |
Ilpo Järvinen | 4e67d87 | 2007-12-05 02:25:32 -0800 | [diff] [blame] | 1194 | tp->nonagle : TCP_NAGLE_PUSH))); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1195 | } |
| 1196 | |
| 1197 | /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet |
| 1198 | * which is put after SKB on the list. It is very much like |
| 1199 | * tcp_fragment() except that it may make several kinds of assumptions |
| 1200 | * in order to speed up the splitting operation. In particular, we |
| 1201 | * know that all the data is in scatter-gather pages, and that the |
| 1202 | * packet has never been sent out before (and thus is not cloned). |
| 1203 | */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1204 | static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, |
| 1205 | unsigned int mss_now) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1206 | { |
| 1207 | struct sk_buff *buff; |
| 1208 | int nlen = skb->len - len; |
| 1209 | u16 flags; |
| 1210 | |
| 1211 | /* All of a TSO frame must be composed of paged data. */ |
Herbert Xu | c8ac377 | 2005-08-16 20:43:40 -0700 | [diff] [blame] | 1212 | if (skb->len != skb->data_len) |
| 1213 | return tcp_fragment(sk, skb, len, mss_now); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1214 | |
Pavel Emelyanov | df97c70 | 2007-11-29 21:22:33 +1100 | [diff] [blame] | 1215 | buff = sk_stream_alloc_skb(sk, 0, GFP_ATOMIC); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1216 | if (unlikely(buff == NULL)) |
| 1217 | return -ENOMEM; |
| 1218 | |
Hideo Aoki | 3ab224b | 2007-12-31 00:11:19 -0800 | [diff] [blame] | 1219 | sk->sk_wmem_queued += buff->truesize; |
| 1220 | sk_mem_charge(sk, buff->truesize); |
Herbert Xu | b60b49e | 2006-04-19 21:35:00 -0700 | [diff] [blame] | 1221 | buff->truesize += nlen; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1222 | skb->truesize -= nlen; |
| 1223 | |
| 1224 | /* Correct the sequence numbers. */ |
| 1225 | TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; |
| 1226 | TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; |
| 1227 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; |
| 1228 | |
| 1229 | /* PSH and FIN should only be set in the second packet. */ |
| 1230 | flags = TCP_SKB_CB(skb)->flags; |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1231 | TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN | TCPCB_FLAG_PSH); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1232 | TCP_SKB_CB(buff)->flags = flags; |
| 1233 | |
| 1234 | /* This packet was never sent out yet, so no SACK bits. */ |
| 1235 | TCP_SKB_CB(buff)->sacked = 0; |
| 1236 | |
Patrick McHardy | 84fa793 | 2006-08-29 16:44:56 -0700 | [diff] [blame] | 1237 | buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1238 | skb_split(skb, buff, len); |
| 1239 | |
| 1240 | /* Fix up tso_factor for both original and new SKB. */ |
David S. Miller | 846998a | 2005-08-04 19:52:01 -0700 | [diff] [blame] | 1241 | tcp_set_skb_tso_segs(sk, skb, mss_now); |
| 1242 | tcp_set_skb_tso_segs(sk, buff, mss_now); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1243 | |
| 1244 | /* Link BUFF into the send queue. */ |
| 1245 | skb_header_release(buff); |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1246 | tcp_insert_write_queue_after(skb, buff, sk); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1247 | |
| 1248 | return 0; |
| 1249 | } |
| 1250 | |
| 1251 | /* Try to defer sending, if possible, in order to minimize the amount |
| 1252 | * of TSO splitting we do. View it as a kind of TSO Nagle test. |
| 1253 | * |
| 1254 | * This algorithm is from John Heffner. |
| 1255 | */ |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 1256 | static int tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1257 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 1258 | struct tcp_sock *tp = tcp_sk(sk); |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 1259 | const struct inet_connection_sock *icsk = inet_csk(sk); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1260 | u32 send_win, cong_win, limit, in_flight; |
| 1261 | |
| 1262 | if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) |
John Heffner | ae8064a | 2006-10-18 20:36:48 -0700 | [diff] [blame] | 1263 | goto send_now; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1264 | |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 1265 | if (icsk->icsk_ca_state != TCP_CA_Open) |
John Heffner | ae8064a | 2006-10-18 20:36:48 -0700 | [diff] [blame] | 1266 | goto send_now; |
| 1267 | |
| 1268 | /* Defer for less than two clock ticks. */ |
Ilpo Järvinen | bd515c3 | 2007-12-20 20:36:03 -0800 | [diff] [blame] | 1269 | if (tp->tso_deferred && |
| 1270 | ((jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1) |
John Heffner | ae8064a | 2006-10-18 20:36:48 -0700 | [diff] [blame] | 1271 | goto send_now; |
David S. Miller | 908a75c | 2005-07-05 15:43:58 -0700 | [diff] [blame] | 1272 | |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1273 | in_flight = tcp_packets_in_flight(tp); |
| 1274 | |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1275 | 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] | 1276 | |
Ilpo Järvinen | 90840de | 2007-12-31 04:48:41 -0800 | [diff] [blame] | 1277 | send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1278 | |
| 1279 | /* From in_flight test above, we know that cwnd > in_flight. */ |
| 1280 | cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; |
| 1281 | |
| 1282 | limit = min(send_win, cong_win); |
| 1283 | |
David S. Miller | ba244fe | 2006-03-11 18:51:49 -0800 | [diff] [blame] | 1284 | /* If a full-sized TSO skb can be sent, do it. */ |
Peter P Waskiewicz Jr | 82cc1a7 | 2008-03-21 03:43:19 -0700 | [diff] [blame] | 1285 | if (limit >= sk->sk_gso_max_size) |
John Heffner | ae8064a | 2006-10-18 20:36:48 -0700 | [diff] [blame] | 1286 | goto send_now; |
David S. Miller | ba244fe | 2006-03-11 18:51:49 -0800 | [diff] [blame] | 1287 | |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1288 | if (sysctl_tcp_tso_win_divisor) { |
| 1289 | u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); |
| 1290 | |
| 1291 | /* If at least some fraction of a window is available, |
| 1292 | * just use it. |
| 1293 | */ |
| 1294 | chunk /= sysctl_tcp_tso_win_divisor; |
| 1295 | if (limit >= chunk) |
John Heffner | ae8064a | 2006-10-18 20:36:48 -0700 | [diff] [blame] | 1296 | goto send_now; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1297 | } else { |
| 1298 | /* Different approach, try not to defer past a single |
| 1299 | * ACK. Receiver should ACK every other full sized |
| 1300 | * frame, so if we have space for more than 3 frames |
| 1301 | * then send now. |
| 1302 | */ |
| 1303 | if (limit > tcp_max_burst(tp) * tp->mss_cache) |
John Heffner | ae8064a | 2006-10-18 20:36:48 -0700 | [diff] [blame] | 1304 | goto send_now; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1305 | } |
| 1306 | |
| 1307 | /* Ok, it looks like it is advisable to defer. */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1308 | tp->tso_deferred = 1 | (jiffies << 1); |
John Heffner | ae8064a | 2006-10-18 20:36:48 -0700 | [diff] [blame] | 1309 | |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1310 | return 1; |
John Heffner | ae8064a | 2006-10-18 20:36:48 -0700 | [diff] [blame] | 1311 | |
| 1312 | send_now: |
| 1313 | tp->tso_deferred = 0; |
| 1314 | return 0; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1315 | } |
| 1316 | |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1317 | /* Create a new MTU probe if we are ready. |
| 1318 | * Returns 0 if we should wait to probe (no cwnd available), |
| 1319 | * 1 if a probe was sent, |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1320 | * -1 otherwise |
| 1321 | */ |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1322 | static int tcp_mtu_probe(struct sock *sk) |
| 1323 | { |
| 1324 | struct tcp_sock *tp = tcp_sk(sk); |
| 1325 | struct inet_connection_sock *icsk = inet_csk(sk); |
| 1326 | struct sk_buff *skb, *nskb, *next; |
| 1327 | int len; |
| 1328 | int probe_size; |
Ilpo Järvinen | 91cc17c | 2007-11-23 19:08:16 +0800 | [diff] [blame] | 1329 | int size_needed; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1330 | int copy; |
| 1331 | int mss_now; |
| 1332 | |
| 1333 | /* Not currently probing/verifying, |
| 1334 | * not in recovery, |
| 1335 | * have enough cwnd, and |
| 1336 | * not SACKing (the variable headers throw things off) */ |
| 1337 | if (!icsk->icsk_mtup.enabled || |
| 1338 | icsk->icsk_mtup.probe_size || |
| 1339 | inet_csk(sk)->icsk_ca_state != TCP_CA_Open || |
| 1340 | tp->snd_cwnd < 11 || |
| 1341 | tp->rx_opt.eff_sacks) |
| 1342 | return -1; |
| 1343 | |
| 1344 | /* Very simple search strategy: just double the MSS. */ |
| 1345 | mss_now = tcp_current_mss(sk, 0); |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1346 | probe_size = 2 * tp->mss_cache; |
Ilpo Järvinen | 91cc17c | 2007-11-23 19:08:16 +0800 | [diff] [blame] | 1347 | size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1348 | if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) { |
| 1349 | /* TODO: set timer for probe_converge_event */ |
| 1350 | return -1; |
| 1351 | } |
| 1352 | |
| 1353 | /* Have enough data in the send queue to probe? */ |
Ilpo Järvinen | 7f9c33e | 2007-11-23 19:10:56 +0800 | [diff] [blame] | 1354 | if (tp->write_seq - tp->snd_nxt < size_needed) |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1355 | return -1; |
| 1356 | |
Ilpo Järvinen | 91cc17c | 2007-11-23 19:08:16 +0800 | [diff] [blame] | 1357 | if (tp->snd_wnd < size_needed) |
| 1358 | return -1; |
Ilpo Järvinen | 90840de | 2007-12-31 04:48:41 -0800 | [diff] [blame] | 1359 | if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) |
Ilpo Järvinen | 91cc17c | 2007-11-23 19:08:16 +0800 | [diff] [blame] | 1360 | return 0; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1361 | |
Ilpo Järvinen | d67c58e | 2007-12-02 00:48:01 +0200 | [diff] [blame] | 1362 | /* Do we need to wait to drain cwnd? With none in flight, don't stall */ |
| 1363 | if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { |
| 1364 | if (!tcp_packets_in_flight(tp)) |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1365 | return -1; |
| 1366 | else |
| 1367 | return 0; |
| 1368 | } |
| 1369 | |
| 1370 | /* We're allowed to probe. Build it now. */ |
| 1371 | if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL) |
| 1372 | return -1; |
Hideo Aoki | 3ab224b | 2007-12-31 00:11:19 -0800 | [diff] [blame] | 1373 | sk->sk_wmem_queued += nskb->truesize; |
| 1374 | sk_mem_charge(sk, nskb->truesize); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1375 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1376 | skb = tcp_send_head(sk); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1377 | |
| 1378 | TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; |
| 1379 | TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; |
| 1380 | TCP_SKB_CB(nskb)->flags = TCPCB_FLAG_ACK; |
| 1381 | TCP_SKB_CB(nskb)->sacked = 0; |
| 1382 | nskb->csum = 0; |
Patrick McHardy | 84fa793 | 2006-08-29 16:44:56 -0700 | [diff] [blame] | 1383 | nskb->ip_summed = skb->ip_summed; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1384 | |
Ilpo Järvinen | 50c4817 | 2007-12-02 00:48:00 +0200 | [diff] [blame] | 1385 | tcp_insert_write_queue_before(nskb, skb, sk); |
| 1386 | |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1387 | len = 0; |
Ilpo Järvinen | 234b686 | 2007-12-02 00:48:02 +0200 | [diff] [blame] | 1388 | tcp_for_write_queue_from_safe(skb, next, sk) { |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1389 | copy = min_t(int, skb->len, probe_size - len); |
| 1390 | if (nskb->ip_summed) |
| 1391 | skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); |
| 1392 | else |
| 1393 | nskb->csum = skb_copy_and_csum_bits(skb, 0, |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1394 | skb_put(nskb, copy), |
| 1395 | copy, nskb->csum); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1396 | |
| 1397 | if (skb->len <= copy) { |
| 1398 | /* We've eaten all the data from this skb. |
| 1399 | * Throw it away. */ |
| 1400 | TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags; |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1401 | tcp_unlink_write_queue(skb, sk); |
Hideo Aoki | 3ab224b | 2007-12-31 00:11:19 -0800 | [diff] [blame] | 1402 | sk_wmem_free_skb(sk, skb); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1403 | } else { |
| 1404 | TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags & |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 1405 | ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1406 | if (!skb_shinfo(skb)->nr_frags) { |
| 1407 | skb_pull(skb, copy); |
Patrick McHardy | 84fa793 | 2006-08-29 16:44:56 -0700 | [diff] [blame] | 1408 | if (skb->ip_summed != CHECKSUM_PARTIAL) |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1409 | skb->csum = csum_partial(skb->data, |
| 1410 | skb->len, 0); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1411 | } else { |
| 1412 | __pskb_trim_head(skb, copy); |
| 1413 | tcp_set_skb_tso_segs(sk, skb, mss_now); |
| 1414 | } |
| 1415 | TCP_SKB_CB(skb)->seq += copy; |
| 1416 | } |
| 1417 | |
| 1418 | len += copy; |
Ilpo Järvinen | 234b686 | 2007-12-02 00:48:02 +0200 | [diff] [blame] | 1419 | |
| 1420 | if (len >= probe_size) |
| 1421 | break; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1422 | } |
| 1423 | tcp_init_tso_segs(sk, nskb, nskb->len); |
| 1424 | |
| 1425 | /* We're ready to send. If this fails, the probe will |
| 1426 | * be resegmented into mss-sized pieces by tcp_write_xmit(). */ |
| 1427 | TCP_SKB_CB(nskb)->when = tcp_time_stamp; |
| 1428 | if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { |
| 1429 | /* Decrement cwnd here because we are sending |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1430 | * effectively two packets. */ |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1431 | tp->snd_cwnd--; |
Ilpo Järvinen | 66f5fe6 | 2007-12-31 04:43:57 -0800 | [diff] [blame] | 1432 | tcp_event_new_data_sent(sk, nskb); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1433 | |
| 1434 | icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); |
John Heffner | 0e7b136 | 2006-03-20 21:32:58 -0800 | [diff] [blame] | 1435 | tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; |
| 1436 | tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1437 | |
| 1438 | return 1; |
| 1439 | } |
| 1440 | |
| 1441 | return -1; |
| 1442 | } |
| 1443 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1444 | /* This routine writes packets to the network. It advances the |
| 1445 | * send_head. This happens as incoming acks open up the remote |
| 1446 | * window for us. |
| 1447 | * |
| 1448 | * Returns 1, if no segments are in flight and we have queued segments, but |
| 1449 | * cannot send anything now because of SWS or another problem. |
| 1450 | */ |
David S. Miller | a2e2a59 | 2005-07-05 15:19:23 -0700 | [diff] [blame] | 1451 | static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1452 | { |
| 1453 | struct tcp_sock *tp = tcp_sk(sk); |
David S. Miller | 92df7b5 | 2005-07-05 15:19:06 -0700 | [diff] [blame] | 1454 | struct sk_buff *skb; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1455 | unsigned int tso_segs, sent_pkts; |
| 1456 | int cwnd_quota; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1457 | int result; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1458 | |
| 1459 | /* If we are closed, the bytes will have to remain here. |
| 1460 | * In time closedown will finish, we empty the write queue and all |
| 1461 | * will be happy. |
| 1462 | */ |
David S. Miller | 92df7b5 | 2005-07-05 15:19:06 -0700 | [diff] [blame] | 1463 | if (unlikely(sk->sk_state == TCP_CLOSE)) |
| 1464 | return 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1465 | |
David S. Miller | 92df7b5 | 2005-07-05 15:19:06 -0700 | [diff] [blame] | 1466 | sent_pkts = 0; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1467 | |
| 1468 | /* Do MTU probing. */ |
| 1469 | if ((result = tcp_mtu_probe(sk)) == 0) { |
| 1470 | return 0; |
| 1471 | } else if (result > 0) { |
| 1472 | sent_pkts = 1; |
| 1473 | } |
| 1474 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1475 | while ((skb = tcp_send_head(sk))) { |
Herbert Xu | c8ac377 | 2005-08-16 20:43:40 -0700 | [diff] [blame] | 1476 | unsigned int limit; |
| 1477 | |
Herbert Xu | b68e9f8 | 2005-08-04 19:52:02 -0700 | [diff] [blame] | 1478 | tso_segs = tcp_init_tso_segs(sk, skb, mss_now); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1479 | BUG_ON(!tso_segs); |
David S. Miller | aa93466 | 2005-07-05 15:20:09 -0700 | [diff] [blame] | 1480 | |
Herbert Xu | b68e9f8 | 2005-08-04 19:52:02 -0700 | [diff] [blame] | 1481 | cwnd_quota = tcp_cwnd_test(tp, skb); |
| 1482 | if (!cwnd_quota) |
| 1483 | break; |
| 1484 | |
| 1485 | if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) |
| 1486 | break; |
| 1487 | |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1488 | if (tso_segs == 1) { |
| 1489 | if (unlikely(!tcp_nagle_test(tp, skb, mss_now, |
| 1490 | (tcp_skb_is_last(sk, skb) ? |
| 1491 | nonagle : TCP_NAGLE_PUSH)))) |
| 1492 | break; |
| 1493 | } else { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 1494 | if (tcp_tso_should_defer(sk, skb)) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1495 | break; |
| 1496 | } |
David S. Miller | aa93466 | 2005-07-05 15:20:09 -0700 | [diff] [blame] | 1497 | |
Herbert Xu | c8ac377 | 2005-08-16 20:43:40 -0700 | [diff] [blame] | 1498 | limit = mss_now; |
Ilpo Järvinen | 0e3a480 | 2007-12-24 21:33:45 -0800 | [diff] [blame] | 1499 | if (tso_segs > 1) |
| 1500 | limit = tcp_mss_split_point(sk, skb, mss_now, |
| 1501 | cwnd_quota); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1502 | |
Herbert Xu | c8ac377 | 2005-08-16 20:43:40 -0700 | [diff] [blame] | 1503 | if (skb->len > limit && |
| 1504 | unlikely(tso_fragment(sk, skb, limit, mss_now))) |
| 1505 | break; |
| 1506 | |
David S. Miller | 92df7b5 | 2005-07-05 15:19:06 -0700 | [diff] [blame] | 1507 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1508 | |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 1509 | if (unlikely(tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC))) |
David S. Miller | 92df7b5 | 2005-07-05 15:19:06 -0700 | [diff] [blame] | 1510 | break; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1511 | |
David S. Miller | 92df7b5 | 2005-07-05 15:19:06 -0700 | [diff] [blame] | 1512 | /* Advance the send_head. This one is sent out. |
| 1513 | * This call will increment packets_out. |
| 1514 | */ |
Ilpo Järvinen | 66f5fe6 | 2007-12-31 04:43:57 -0800 | [diff] [blame] | 1515 | tcp_event_new_data_sent(sk, skb); |
David S. Miller | 92df7b5 | 2005-07-05 15:19:06 -0700 | [diff] [blame] | 1516 | |
| 1517 | tcp_minshall_update(tp, mss_now, skb); |
David S. Miller | aa93466 | 2005-07-05 15:20:09 -0700 | [diff] [blame] | 1518 | sent_pkts++; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1519 | } |
David S. Miller | 92df7b5 | 2005-07-05 15:19:06 -0700 | [diff] [blame] | 1520 | |
David S. Miller | aa93466 | 2005-07-05 15:20:09 -0700 | [diff] [blame] | 1521 | if (likely(sent_pkts)) { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 1522 | tcp_cwnd_validate(sk); |
David S. Miller | 92df7b5 | 2005-07-05 15:19:06 -0700 | [diff] [blame] | 1523 | return 0; |
| 1524 | } |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1525 | return !tp->packets_out && tcp_send_head(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1526 | } |
| 1527 | |
David S. Miller | a762a98 | 2005-07-05 15:18:51 -0700 | [diff] [blame] | 1528 | /* Push out any pending frames which were held back due to |
| 1529 | * TCP_CORK or attempt at coalescing tiny packets. |
| 1530 | * The socket must be locked by the caller. |
| 1531 | */ |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 1532 | void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, |
| 1533 | int nonagle) |
David S. Miller | a762a98 | 2005-07-05 15:18:51 -0700 | [diff] [blame] | 1534 | { |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1535 | struct sk_buff *skb = tcp_send_head(sk); |
David S. Miller | a762a98 | 2005-07-05 15:18:51 -0700 | [diff] [blame] | 1536 | |
| 1537 | if (skb) { |
David S. Miller | 55c97f3 | 2005-07-05 15:19:38 -0700 | [diff] [blame] | 1538 | if (tcp_write_xmit(sk, cur_mss, nonagle)) |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 1539 | tcp_check_probe_timer(sk); |
David S. Miller | a762a98 | 2005-07-05 15:18:51 -0700 | [diff] [blame] | 1540 | } |
| 1541 | } |
| 1542 | |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1543 | /* Send _single_ skb sitting at the send head. This function requires |
| 1544 | * true push pending frames to setup probe timer etc. |
| 1545 | */ |
| 1546 | void tcp_push_one(struct sock *sk, unsigned int mss_now) |
| 1547 | { |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1548 | struct sk_buff *skb = tcp_send_head(sk); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1549 | unsigned int tso_segs, cwnd_quota; |
| 1550 | |
| 1551 | BUG_ON(!skb || skb->len < mss_now); |
| 1552 | |
David S. Miller | 846998a | 2005-08-04 19:52:01 -0700 | [diff] [blame] | 1553 | tso_segs = tcp_init_tso_segs(sk, skb, mss_now); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1554 | cwnd_quota = tcp_snd_test(sk, skb, mss_now, TCP_NAGLE_PUSH); |
| 1555 | |
| 1556 | if (likely(cwnd_quota)) { |
Herbert Xu | c8ac377 | 2005-08-16 20:43:40 -0700 | [diff] [blame] | 1557 | unsigned int limit; |
| 1558 | |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1559 | BUG_ON(!tso_segs); |
| 1560 | |
Herbert Xu | c8ac377 | 2005-08-16 20:43:40 -0700 | [diff] [blame] | 1561 | limit = mss_now; |
Ilpo Järvinen | 0e3a480 | 2007-12-24 21:33:45 -0800 | [diff] [blame] | 1562 | if (tso_segs > 1) |
| 1563 | limit = tcp_mss_split_point(sk, skb, mss_now, |
| 1564 | cwnd_quota); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1565 | |
Herbert Xu | c8ac377 | 2005-08-16 20:43:40 -0700 | [diff] [blame] | 1566 | if (skb->len > limit && |
| 1567 | unlikely(tso_fragment(sk, skb, limit, mss_now))) |
| 1568 | return; |
| 1569 | |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1570 | /* Send it out now. */ |
| 1571 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
| 1572 | |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 1573 | if (likely(!tcp_transmit_skb(sk, skb, 1, sk->sk_allocation))) { |
Ilpo Järvinen | 66f5fe6 | 2007-12-31 04:43:57 -0800 | [diff] [blame] | 1574 | tcp_event_new_data_sent(sk, skb); |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 1575 | tcp_cwnd_validate(sk); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1576 | return; |
| 1577 | } |
| 1578 | } |
| 1579 | } |
| 1580 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1581 | /* This function returns the amount that we can raise the |
| 1582 | * usable window based on the following constraints |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 1583 | * |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1584 | * 1. The window can never be shrunk once it is offered (RFC 793) |
| 1585 | * 2. We limit memory per socket |
| 1586 | * |
| 1587 | * RFC 1122: |
| 1588 | * "the suggested [SWS] avoidance algorithm for the receiver is to keep |
| 1589 | * RECV.NEXT + RCV.WIN fixed until: |
| 1590 | * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" |
| 1591 | * |
| 1592 | * i.e. don't raise the right edge of the window until you can raise |
| 1593 | * it at least MSS bytes. |
| 1594 | * |
| 1595 | * Unfortunately, the recommended algorithm breaks header prediction, |
| 1596 | * since header prediction assumes th->window stays fixed. |
| 1597 | * |
| 1598 | * Strictly speaking, keeping th->window fixed violates the receiver |
| 1599 | * side SWS prevention criteria. The problem is that under this rule |
| 1600 | * a stream of single byte packets will cause the right side of the |
| 1601 | * window to always advance by a single byte. |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 1602 | * |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1603 | * Of course, if the sender implements sender side SWS prevention |
| 1604 | * then this will not be a problem. |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 1605 | * |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1606 | * BSD seems to make the following compromise: |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 1607 | * |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1608 | * If the free space is less than the 1/4 of the maximum |
| 1609 | * space available and the free space is less than 1/2 mss, |
| 1610 | * then set the window to 0. |
| 1611 | * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] |
| 1612 | * Otherwise, just prevent the window from shrinking |
| 1613 | * and from being larger than the largest representable value. |
| 1614 | * |
| 1615 | * This prevents incremental opening of the window in the regime |
| 1616 | * where TCP is limited by the speed of the reader side taking |
| 1617 | * data out of the TCP receive queue. It does nothing about |
| 1618 | * those cases where the window is constrained on the sender side |
| 1619 | * because the pipeline is full. |
| 1620 | * |
| 1621 | * BSD also seems to "accidentally" limit itself to windows that are a |
| 1622 | * multiple of MSS, at least until the free space gets quite small. |
| 1623 | * This would appear to be a side effect of the mbuf implementation. |
| 1624 | * Combining these two algorithms results in the observed behavior |
| 1625 | * of having a fixed window size at almost all times. |
| 1626 | * |
| 1627 | * Below we obtain similar behavior by forcing the offered window to |
| 1628 | * a multiple of the mss when it is feasible to do so. |
| 1629 | * |
| 1630 | * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. |
| 1631 | * Regular options like TIMESTAMP are taken into account. |
| 1632 | */ |
| 1633 | u32 __tcp_select_window(struct sock *sk) |
| 1634 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 1635 | struct inet_connection_sock *icsk = inet_csk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1636 | struct tcp_sock *tp = tcp_sk(sk); |
Stephen Hemminger | caa20d9a | 2005-11-10 17:13:47 -0800 | [diff] [blame] | 1637 | /* MSS for the peer's data. Previous versions used mss_clamp |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1638 | * here. I don't know if the value based on our guesses |
| 1639 | * of peer's MSS is better for the performance. It's more correct |
| 1640 | * but may be worse for the performance because of rcv_mss |
| 1641 | * fluctuations. --SAW 1998/11/1 |
| 1642 | */ |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 1643 | int mss = icsk->icsk_ack.rcv_mss; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1644 | int free_space = tcp_space(sk); |
| 1645 | int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk)); |
| 1646 | int window; |
| 1647 | |
| 1648 | if (mss > full_space) |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 1649 | mss = full_space; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1650 | |
Eric Dumazet | b92edbe | 2007-12-20 21:48:32 -0800 | [diff] [blame] | 1651 | if (free_space < (full_space >> 1)) { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 1652 | icsk->icsk_ack.quick = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1653 | |
| 1654 | if (tcp_memory_pressure) |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1655 | tp->rcv_ssthresh = min(tp->rcv_ssthresh, |
| 1656 | 4U * tp->advmss); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1657 | |
| 1658 | if (free_space < mss) |
| 1659 | return 0; |
| 1660 | } |
| 1661 | |
| 1662 | if (free_space > tp->rcv_ssthresh) |
| 1663 | free_space = tp->rcv_ssthresh; |
| 1664 | |
| 1665 | /* Don't do rounding if we are using window scaling, since the |
| 1666 | * scaled window will not line up with the MSS boundary anyway. |
| 1667 | */ |
| 1668 | window = tp->rcv_wnd; |
| 1669 | if (tp->rx_opt.rcv_wscale) { |
| 1670 | window = free_space; |
| 1671 | |
| 1672 | /* Advertise enough space so that it won't get scaled away. |
| 1673 | * Import case: prevent zero window announcement if |
| 1674 | * 1<<rcv_wscale > mss. |
| 1675 | */ |
| 1676 | if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window) |
| 1677 | window = (((window >> tp->rx_opt.rcv_wscale) + 1) |
| 1678 | << tp->rx_opt.rcv_wscale); |
| 1679 | } else { |
| 1680 | /* Get the largest window that is a nice multiple of mss. |
| 1681 | * Window clamp already applied above. |
| 1682 | * If our current window offering is within 1 mss of the |
| 1683 | * free space we just keep it. This prevents the divide |
| 1684 | * and multiply from happening most of the time. |
| 1685 | * We also don't do any window rounding when the free space |
| 1686 | * is too small. |
| 1687 | */ |
| 1688 | if (window <= free_space - mss || window > free_space) |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1689 | window = (free_space / mss) * mss; |
John Heffner | 8456507 | 2007-04-02 13:56:32 -0700 | [diff] [blame] | 1690 | else if (mss == full_space && |
Eric Dumazet | b92edbe | 2007-12-20 21:48:32 -0800 | [diff] [blame] | 1691 | free_space > window + (full_space >> 1)) |
John Heffner | 8456507 | 2007-04-02 13:56:32 -0700 | [diff] [blame] | 1692 | window = free_space; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1693 | } |
| 1694 | |
| 1695 | return window; |
| 1696 | } |
| 1697 | |
| 1698 | /* Attempt to collapse two adjacent SKB's during retransmission. */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1699 | static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *skb, |
| 1700 | int mss_now) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1701 | { |
| 1702 | struct tcp_sock *tp = tcp_sk(sk); |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1703 | struct sk_buff *next_skb = tcp_write_queue_next(sk, skb); |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 1704 | int skb_size, next_skb_size; |
| 1705 | u16 flags; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1706 | |
| 1707 | /* The first test we must make is that neither of these two |
| 1708 | * SKB's are still referenced by someone else. |
| 1709 | */ |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 1710 | if (skb_cloned(skb) || skb_cloned(next_skb)) |
| 1711 | return; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1712 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 1713 | skb_size = skb->len; |
| 1714 | next_skb_size = next_skb->len; |
| 1715 | flags = TCP_SKB_CB(skb)->flags; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1716 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 1717 | /* Also punt if next skb has been SACK'd. */ |
| 1718 | if (TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_ACKED) |
| 1719 | return; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1720 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 1721 | /* Next skb is out of window. */ |
| 1722 | if (after(TCP_SKB_CB(next_skb)->end_seq, tcp_wnd_end(tp))) |
| 1723 | return; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1724 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 1725 | /* Punt if not enough space exists in the first SKB for |
| 1726 | * the data in the second, or the total combined payload |
| 1727 | * would exceed the MSS. |
| 1728 | */ |
| 1729 | if ((next_skb_size > skb_tailroom(skb)) || |
| 1730 | ((skb_size + next_skb_size) > mss_now)) |
| 1731 | return; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1732 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 1733 | 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] | 1734 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 1735 | tcp_highest_sack_combine(sk, next_skb, skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1736 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 1737 | /* Ok. We will be able to collapse the packet. */ |
| 1738 | tcp_unlink_write_queue(next_skb, sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1739 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 1740 | skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size), |
| 1741 | next_skb_size); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1742 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 1743 | if (next_skb->ip_summed == CHECKSUM_PARTIAL) |
| 1744 | skb->ip_summed = CHECKSUM_PARTIAL; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1745 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 1746 | if (skb->ip_summed != CHECKSUM_PARTIAL) |
| 1747 | skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1748 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 1749 | /* Update sequence range on original skb. */ |
| 1750 | 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] | 1751 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 1752 | /* Merge over control information. */ |
| 1753 | flags |= TCP_SKB_CB(next_skb)->flags; /* This moves PSH/FIN etc. over */ |
| 1754 | TCP_SKB_CB(skb)->flags = flags; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1755 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 1756 | /* All done, get rid of second SKB and account for it so |
| 1757 | * packet counting does not break. |
| 1758 | */ |
| 1759 | TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; |
| 1760 | if (TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_RETRANS) |
| 1761 | tp->retrans_out -= tcp_skb_pcount(next_skb); |
| 1762 | if (TCP_SKB_CB(next_skb)->sacked & TCPCB_LOST) |
| 1763 | tp->lost_out -= tcp_skb_pcount(next_skb); |
| 1764 | /* Reno case is special. Sigh... */ |
| 1765 | if (tcp_is_reno(tp) && tp->sacked_out) |
| 1766 | tcp_dec_pcount_approx(&tp->sacked_out, next_skb); |
Ilpo Järvinen | b768920 | 2007-09-20 11:37:19 -0700 | [diff] [blame] | 1767 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 1768 | tcp_adjust_fackets_out(sk, next_skb, tcp_skb_pcount(next_skb)); |
| 1769 | tp->packets_out -= tcp_skb_pcount(next_skb); |
Ilpo Järvinen | b768920 | 2007-09-20 11:37:19 -0700 | [diff] [blame] | 1770 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 1771 | /* changed transmit queue under us so clear hints */ |
| 1772 | tcp_clear_retrans_hints_partial(tp); |
| 1773 | |
| 1774 | sk_wmem_free_skb(sk, next_skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1775 | } |
| 1776 | |
| 1777 | /* Do a simple retransmit without using the backoff mechanisms in |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 1778 | * tcp_timer. This is used for path mtu discovery. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1779 | * The socket is already locked here. |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 1780 | */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1781 | void tcp_simple_retransmit(struct sock *sk) |
| 1782 | { |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 1783 | const struct inet_connection_sock *icsk = inet_csk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1784 | struct tcp_sock *tp = tcp_sk(sk); |
| 1785 | struct sk_buff *skb; |
| 1786 | unsigned int mss = tcp_current_mss(sk, 0); |
| 1787 | int lost = 0; |
| 1788 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1789 | tcp_for_write_queue(skb, sk) { |
| 1790 | if (skb == tcp_send_head(sk)) |
| 1791 | break; |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 1792 | if (skb->len > mss && |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1793 | !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) { |
| 1794 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1795 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; |
| 1796 | tp->retrans_out -= tcp_skb_pcount(skb); |
| 1797 | } |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1798 | if (!(TCP_SKB_CB(skb)->sacked & TCPCB_LOST)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1799 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; |
| 1800 | tp->lost_out += tcp_skb_pcount(skb); |
| 1801 | lost = 1; |
| 1802 | } |
| 1803 | } |
| 1804 | } |
| 1805 | |
Ilpo Järvinen | 5af4ec2 | 2007-09-20 11:30:48 -0700 | [diff] [blame] | 1806 | tcp_clear_all_retrans_hints(tp); |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 1807 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1808 | if (!lost) |
| 1809 | return; |
| 1810 | |
Ilpo Järvinen | 882beba | 2008-04-07 22:33:07 -0700 | [diff] [blame] | 1811 | if (tcp_is_reno(tp)) |
| 1812 | tcp_limit_reno_sacked(tp); |
| 1813 | |
Ilpo Järvinen | 005903b | 2007-08-09 14:44:16 +0300 | [diff] [blame] | 1814 | tcp_verify_left_out(tp); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1815 | |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 1816 | /* Don't muck with the congestion window here. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1817 | * Reason is that we do not increase amount of _data_ |
| 1818 | * in network, but units changed and effective |
| 1819 | * cwnd/ssthresh really reduced now. |
| 1820 | */ |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 1821 | if (icsk->icsk_ca_state != TCP_CA_Loss) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1822 | tp->high_seq = tp->snd_nxt; |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 1823 | tp->snd_ssthresh = tcp_current_ssthresh(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1824 | tp->prior_ssthresh = 0; |
| 1825 | tp->undo_marker = 0; |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 1826 | tcp_set_ca_state(sk, TCP_CA_Loss); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1827 | } |
| 1828 | tcp_xmit_retransmit_queue(sk); |
| 1829 | } |
| 1830 | |
| 1831 | /* This retransmits one SKB. Policy decisions and retransmit queue |
| 1832 | * state updates are done by the caller. Returns non-zero if an |
| 1833 | * error occurred which prevented the send. |
| 1834 | */ |
| 1835 | int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) |
| 1836 | { |
| 1837 | struct tcp_sock *tp = tcp_sk(sk); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1838 | struct inet_connection_sock *icsk = inet_csk(sk); |
Sridhar Samudrala | 7d227cd2 | 2008-05-21 16:42:20 -0700 | [diff] [blame] | 1839 | unsigned int cur_mss; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1840 | int err; |
| 1841 | |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1842 | /* Inconslusive MTU probe */ |
| 1843 | if (icsk->icsk_mtup.probe_size) { |
| 1844 | icsk->icsk_mtup.probe_size = 0; |
| 1845 | } |
| 1846 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1847 | /* 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] | 1848 | * copying overhead: fragmentation, tunneling, mangling etc. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1849 | */ |
| 1850 | if (atomic_read(&sk->sk_wmem_alloc) > |
| 1851 | min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf)) |
| 1852 | return -EAGAIN; |
| 1853 | |
| 1854 | if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { |
| 1855 | if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) |
| 1856 | BUG(); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1857 | if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) |
| 1858 | return -ENOMEM; |
| 1859 | } |
| 1860 | |
Sridhar Samudrala | 7d227cd2 | 2008-05-21 16:42:20 -0700 | [diff] [blame] | 1861 | if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) |
| 1862 | return -EHOSTUNREACH; /* Routing failure or similar. */ |
| 1863 | |
| 1864 | cur_mss = tcp_current_mss(sk, 0); |
| 1865 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1866 | /* If receiver has shrunk his window, and skb is out of |
| 1867 | * new window, do not retransmit it. The exception is the |
| 1868 | * case, when window is shrunk to zero. In this case |
| 1869 | * our retransmit serves as a zero window probe. |
| 1870 | */ |
Ilpo Järvinen | 90840de | 2007-12-31 04:48:41 -0800 | [diff] [blame] | 1871 | if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1872 | && TCP_SKB_CB(skb)->seq != tp->snd_una) |
| 1873 | return -EAGAIN; |
| 1874 | |
| 1875 | if (skb->len > cur_mss) { |
David S. Miller | 846998a | 2005-08-04 19:52:01 -0700 | [diff] [blame] | 1876 | if (tcp_fragment(sk, skb, cur_mss, cur_mss)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1877 | return -ENOMEM; /* We'll try again later. */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1878 | } |
| 1879 | |
| 1880 | /* Collapse two adjacent packets if worthwhile and we can. */ |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 1881 | if (!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN) && |
| 1882 | (skb->len < (cur_mss >> 1)) && |
| 1883 | (tcp_write_queue_next(sk, skb) != tcp_send_head(sk)) && |
| 1884 | (!tcp_skb_is_last(sk, skb)) && |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1885 | (skb_shinfo(skb)->nr_frags == 0 && |
| 1886 | skb_shinfo(tcp_write_queue_next(sk, skb))->nr_frags == 0) && |
| 1887 | (tcp_skb_pcount(skb) == 1 && |
| 1888 | tcp_skb_pcount(tcp_write_queue_next(sk, skb)) == 1) && |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 1889 | (sysctl_tcp_retrans_collapse != 0)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1890 | tcp_retrans_try_collapse(sk, skb, cur_mss); |
| 1891 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1892 | /* Some Solaris stacks overoptimize and ignore the FIN on a |
| 1893 | * retransmit when old data is attached. So strip it off |
| 1894 | * since it is cheap to do so and saves bytes on the network. |
| 1895 | */ |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 1896 | if (skb->len > 0 && |
| 1897 | (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) && |
| 1898 | tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1899 | if (!pskb_trim(skb, 0)) { |
Ilpo Järvinen | e870a8e | 2008-01-03 20:39:01 -0800 | [diff] [blame] | 1900 | /* Reuse, even though it does some unnecessary work */ |
| 1901 | tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1, |
| 1902 | TCP_SKB_CB(skb)->flags); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1903 | skb->ip_summed = CHECKSUM_NONE; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1904 | } |
| 1905 | } |
| 1906 | |
| 1907 | /* Make a copy, if the first transmission SKB clone we made |
| 1908 | * is still in somebody's hands, else make a clone. |
| 1909 | */ |
| 1910 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1911 | |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 1912 | err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1913 | |
| 1914 | if (err == 0) { |
| 1915 | /* Update global TCP statistics. */ |
| 1916 | TCP_INC_STATS(TCP_MIB_RETRANSSEGS); |
| 1917 | |
| 1918 | tp->total_retrans++; |
| 1919 | |
| 1920 | #if FASTRETRANS_DEBUG > 0 |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1921 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1922 | if (net_ratelimit()) |
| 1923 | printk(KERN_DEBUG "retrans_out leaked.\n"); |
| 1924 | } |
| 1925 | #endif |
Ilpo Järvinen | b08d6cb | 2007-10-11 17:36:13 -0700 | [diff] [blame] | 1926 | if (!tp->retrans_out) |
| 1927 | tp->lost_retrans_low = tp->snd_nxt; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1928 | TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; |
| 1929 | tp->retrans_out += tcp_skb_pcount(skb); |
| 1930 | |
| 1931 | /* Save stamp of the first retransmit. */ |
| 1932 | if (!tp->retrans_stamp) |
| 1933 | tp->retrans_stamp = TCP_SKB_CB(skb)->when; |
| 1934 | |
| 1935 | tp->undo_retrans++; |
| 1936 | |
| 1937 | /* snd_nxt is stored to detect loss of retransmitted segment, |
| 1938 | * see tcp_input.c tcp_sacktag_write_queue(). |
| 1939 | */ |
| 1940 | TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt; |
| 1941 | } |
| 1942 | return err; |
| 1943 | } |
| 1944 | |
| 1945 | /* This gets called after a retransmit timeout, and the initially |
| 1946 | * retransmitted data is acknowledged. It tries to continue |
| 1947 | * resending the rest of the retransmit queue, until either |
| 1948 | * we've sent it all or the congestion window limit is reached. |
| 1949 | * If doing SACK, the first ACK which comes back for a timeout |
| 1950 | * based retransmit packet might feed us FACK information again. |
| 1951 | * If so, we use it to avoid unnecessarily retransmissions. |
| 1952 | */ |
| 1953 | void tcp_xmit_retransmit_queue(struct sock *sk) |
| 1954 | { |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 1955 | const struct inet_connection_sock *icsk = inet_csk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1956 | struct tcp_sock *tp = tcp_sk(sk); |
| 1957 | struct sk_buff *skb; |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 1958 | int packet_cnt; |
| 1959 | |
| 1960 | if (tp->retransmit_skb_hint) { |
| 1961 | skb = tp->retransmit_skb_hint; |
| 1962 | packet_cnt = tp->retransmit_cnt_hint; |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1963 | } else { |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1964 | skb = tcp_write_queue_head(sk); |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 1965 | packet_cnt = 0; |
| 1966 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1967 | |
| 1968 | /* First pass: retransmit lost packets. */ |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 1969 | if (tp->lost_out) { |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1970 | tcp_for_write_queue_from(skb, sk) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1971 | __u8 sacked = TCP_SKB_CB(skb)->sacked; |
| 1972 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1973 | if (skb == tcp_send_head(sk)) |
| 1974 | break; |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 1975 | /* we could do better than to assign each time */ |
| 1976 | tp->retransmit_skb_hint = skb; |
| 1977 | tp->retransmit_cnt_hint = packet_cnt; |
| 1978 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1979 | /* Assume this retransmit will generate |
| 1980 | * only one packet for congestion window |
| 1981 | * calculation purposes. This works because |
| 1982 | * tcp_retransmit_skb() will chop up the |
| 1983 | * packet to be MSS sized and all the |
| 1984 | * packet counting works out. |
| 1985 | */ |
| 1986 | if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) |
| 1987 | return; |
| 1988 | |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 1989 | if (sacked & TCPCB_LOST) { |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1990 | if (!(sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))) { |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 1991 | if (tcp_retransmit_skb(sk, skb)) { |
| 1992 | tp->retransmit_skb_hint = NULL; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1993 | return; |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 1994 | } |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 1995 | if (icsk->icsk_ca_state != TCP_CA_Loss) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1996 | NET_INC_STATS_BH(LINUX_MIB_TCPFASTRETRANS); |
| 1997 | else |
| 1998 | NET_INC_STATS_BH(LINUX_MIB_TCPSLOWSTARTRETRANS); |
| 1999 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2000 | if (skb == tcp_write_queue_head(sk)) |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2001 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, |
Arnaldo Carvalho de Melo | 3f421ba | 2005-08-09 20:11:08 -0700 | [diff] [blame] | 2002 | inet_csk(sk)->icsk_rto, |
| 2003 | TCP_RTO_MAX); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2004 | } |
| 2005 | |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 2006 | packet_cnt += tcp_skb_pcount(skb); |
| 2007 | if (packet_cnt >= tp->lost_out) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2008 | break; |
| 2009 | } |
| 2010 | } |
| 2011 | } |
| 2012 | |
| 2013 | /* OK, demanded retransmission is finished. */ |
| 2014 | |
| 2015 | /* Forward retransmissions are possible only during Recovery. */ |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2016 | if (icsk->icsk_ca_state != TCP_CA_Recovery) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2017 | return; |
| 2018 | |
| 2019 | /* No forward retransmissions in Reno are possible. */ |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 2020 | if (tcp_is_reno(tp)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2021 | return; |
| 2022 | |
| 2023 | /* Yeah, we have to make difficult choice between forward transmission |
| 2024 | * and retransmission... Both ways have their merits... |
| 2025 | * |
| 2026 | * For now we do not retransmit anything, while we have some new |
Ilpo Järvinen | 539d243 | 2007-05-27 02:03:20 -0700 | [diff] [blame] | 2027 | * segments to send. In the other cases, follow rule 3 for |
| 2028 | * NextSeg() specified in RFC3517. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2029 | */ |
| 2030 | |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2031 | if (tcp_may_send_now(sk)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2032 | return; |
| 2033 | |
Ilpo Järvinen | 539d243 | 2007-05-27 02:03:20 -0700 | [diff] [blame] | 2034 | /* If nothing is SACKed, highest_sack in the loop won't be valid */ |
| 2035 | if (!tp->sacked_out) |
| 2036 | return; |
| 2037 | |
| 2038 | if (tp->forward_skb_hint) |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 2039 | skb = tp->forward_skb_hint; |
Ilpo Järvinen | 539d243 | 2007-05-27 02:03:20 -0700 | [diff] [blame] | 2040 | else |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2041 | skb = tcp_write_queue_head(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2042 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2043 | tcp_for_write_queue_from(skb, sk) { |
| 2044 | if (skb == tcp_send_head(sk)) |
| 2045 | break; |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 2046 | tp->forward_skb_hint = skb; |
| 2047 | |
Ilpo Järvinen | 6859d49 | 2007-12-02 00:48:06 +0200 | [diff] [blame] | 2048 | if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp))) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2049 | break; |
| 2050 | |
| 2051 | if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) |
| 2052 | break; |
| 2053 | |
| 2054 | if (TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) |
| 2055 | continue; |
| 2056 | |
| 2057 | /* Ok, retransmit it. */ |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 2058 | if (tcp_retransmit_skb(sk, skb)) { |
| 2059 | tp->forward_skb_hint = NULL; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2060 | break; |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 2061 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2062 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2063 | if (skb == tcp_write_queue_head(sk)) |
Arnaldo Carvalho de Melo | 3f421ba | 2005-08-09 20:11:08 -0700 | [diff] [blame] | 2064 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, |
| 2065 | inet_csk(sk)->icsk_rto, |
| 2066 | TCP_RTO_MAX); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2067 | |
| 2068 | NET_INC_STATS_BH(LINUX_MIB_TCPFORWARDRETRANS); |
| 2069 | } |
| 2070 | } |
| 2071 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2072 | /* Send a fin. The caller locks the socket for us. This cannot be |
| 2073 | * allowed to fail queueing a FIN frame under any circumstances. |
| 2074 | */ |
| 2075 | void tcp_send_fin(struct sock *sk) |
| 2076 | { |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2077 | struct tcp_sock *tp = tcp_sk(sk); |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2078 | struct sk_buff *skb = tcp_write_queue_tail(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2079 | int mss_now; |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2080 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2081 | /* Optimization, tack on the FIN if we have a queue of |
| 2082 | * unsent frames. But be careful about outgoing SACKS |
| 2083 | * and IP options. |
| 2084 | */ |
| 2085 | mss_now = tcp_current_mss(sk, 1); |
| 2086 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2087 | if (tcp_send_head(sk) != NULL) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2088 | TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_FIN; |
| 2089 | TCP_SKB_CB(skb)->end_seq++; |
| 2090 | tp->write_seq++; |
| 2091 | } else { |
| 2092 | /* Socket is locked, keep trying until memory is available. */ |
| 2093 | for (;;) { |
David S. Miller | d179cd1 | 2005-08-17 14:57:30 -0700 | [diff] [blame] | 2094 | skb = alloc_skb_fclone(MAX_TCP_HEADER, GFP_KERNEL); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2095 | if (skb) |
| 2096 | break; |
| 2097 | yield(); |
| 2098 | } |
| 2099 | |
| 2100 | /* Reserve space for headers and prepare control bits. */ |
| 2101 | skb_reserve(skb, MAX_TCP_HEADER); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2102 | /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ |
Ilpo Järvinen | e870a8e | 2008-01-03 20:39:01 -0800 | [diff] [blame] | 2103 | tcp_init_nondata_skb(skb, tp->write_seq, |
| 2104 | TCPCB_FLAG_ACK | TCPCB_FLAG_FIN); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2105 | tcp_queue_skb(sk, skb); |
| 2106 | } |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2107 | __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2108 | } |
| 2109 | |
| 2110 | /* We get here when a process closes a file descriptor (either due to |
| 2111 | * an explicit close() or as a byproduct of exit()'ing) and there |
| 2112 | * was unread data in the receive queue. This behavior is recommended |
Gerrit Renker | 65bb723 | 2007-04-28 21:21:46 -0700 | [diff] [blame] | 2113 | * by RFC 2525, section 2.17. -DaveM |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2114 | */ |
Al Viro | dd0fc66 | 2005-10-07 07:46:04 +0100 | [diff] [blame] | 2115 | void tcp_send_active_reset(struct sock *sk, gfp_t priority) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2116 | { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2117 | struct sk_buff *skb; |
| 2118 | |
| 2119 | /* NOTE: No TCP options attached and we never retransmit this. */ |
| 2120 | skb = alloc_skb(MAX_TCP_HEADER, priority); |
| 2121 | if (!skb) { |
| 2122 | NET_INC_STATS(LINUX_MIB_TCPABORTFAILED); |
| 2123 | return; |
| 2124 | } |
| 2125 | |
| 2126 | /* Reserve space for headers and prepare control bits. */ |
| 2127 | skb_reserve(skb, MAX_TCP_HEADER); |
Ilpo Järvinen | e870a8e | 2008-01-03 20:39:01 -0800 | [diff] [blame] | 2128 | tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), |
| 2129 | TCPCB_FLAG_ACK | TCPCB_FLAG_RST); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2130 | /* Send it off. */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2131 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 2132 | if (tcp_transmit_skb(sk, skb, 0, priority)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2133 | NET_INC_STATS(LINUX_MIB_TCPABORTFAILED); |
Sridhar Samudrala | 26af65c | 2008-06-04 15:19:35 -0700 | [diff] [blame] | 2134 | |
| 2135 | TCP_INC_STATS(TCP_MIB_OUTRSTS); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2136 | } |
| 2137 | |
| 2138 | /* WARNING: This routine must only be called when we have already sent |
| 2139 | * a SYN packet that crossed the incoming SYN that caused this routine |
| 2140 | * to get called. If this assumption fails then the initial rcv_wnd |
| 2141 | * and rcv_wscale values will not be correct. |
| 2142 | */ |
| 2143 | int tcp_send_synack(struct sock *sk) |
| 2144 | { |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 2145 | struct sk_buff *skb; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2146 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2147 | skb = tcp_write_queue_head(sk); |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 2148 | if (skb == NULL || !(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2149 | printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n"); |
| 2150 | return -EFAULT; |
| 2151 | } |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 2152 | if (!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_ACK)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2153 | if (skb_cloned(skb)) { |
| 2154 | struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC); |
| 2155 | if (nskb == NULL) |
| 2156 | return -ENOMEM; |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2157 | tcp_unlink_write_queue(skb, sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2158 | skb_header_release(nskb); |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2159 | __tcp_add_write_queue_head(sk, nskb); |
Hideo Aoki | 3ab224b | 2007-12-31 00:11:19 -0800 | [diff] [blame] | 2160 | sk_wmem_free_skb(sk, skb); |
| 2161 | sk->sk_wmem_queued += nskb->truesize; |
| 2162 | sk_mem_charge(sk, nskb->truesize); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2163 | skb = nskb; |
| 2164 | } |
| 2165 | |
| 2166 | TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ACK; |
| 2167 | TCP_ECN_send_synack(tcp_sk(sk), skb); |
| 2168 | } |
| 2169 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 2170 | return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2171 | } |
| 2172 | |
| 2173 | /* |
| 2174 | * Prepare a SYN-ACK. |
| 2175 | */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 2176 | struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst, |
| 2177 | struct request_sock *req) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2178 | { |
Arnaldo Carvalho de Melo | 2e6599c | 2005-06-18 22:46:52 -0700 | [diff] [blame] | 2179 | struct inet_request_sock *ireq = inet_rsk(req); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2180 | struct tcp_sock *tp = tcp_sk(sk); |
| 2181 | struct tcphdr *th; |
| 2182 | int tcp_header_size; |
| 2183 | struct sk_buff *skb; |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 2184 | #ifdef CONFIG_TCP_MD5SIG |
| 2185 | struct tcp_md5sig_key *md5; |
| 2186 | __u8 *md5_hash_location; |
| 2187 | #endif |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2188 | |
| 2189 | skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC); |
| 2190 | if (skb == NULL) |
| 2191 | return NULL; |
| 2192 | |
| 2193 | /* Reserve space for headers. */ |
| 2194 | skb_reserve(skb, MAX_TCP_HEADER); |
| 2195 | |
| 2196 | skb->dst = dst_clone(dst); |
| 2197 | |
| 2198 | tcp_header_size = (sizeof(struct tcphdr) + TCPOLEN_MSS + |
Arnaldo Carvalho de Melo | 2e6599c | 2005-06-18 22:46:52 -0700 | [diff] [blame] | 2199 | (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0) + |
| 2200 | (ireq->wscale_ok ? TCPOLEN_WSCALE_ALIGNED : 0) + |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2201 | /* SACK_PERM is in the place of NOP NOP of TS */ |
Arnaldo Carvalho de Melo | 2e6599c | 2005-06-18 22:46:52 -0700 | [diff] [blame] | 2202 | ((ireq->sack_ok && !ireq->tstamp_ok) ? TCPOLEN_SACKPERM_ALIGNED : 0)); |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 2203 | |
| 2204 | #ifdef CONFIG_TCP_MD5SIG |
| 2205 | /* Are we doing MD5 on this segment? If so - make room for it */ |
| 2206 | md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req); |
| 2207 | if (md5) |
| 2208 | tcp_header_size += TCPOLEN_MD5SIG_ALIGNED; |
| 2209 | #endif |
Arnaldo Carvalho de Melo | aa8223c | 2007-04-10 21:04:22 -0700 | [diff] [blame] | 2210 | skb_push(skb, tcp_header_size); |
| 2211 | skb_reset_transport_header(skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2212 | |
Arnaldo Carvalho de Melo | aa8223c | 2007-04-10 21:04:22 -0700 | [diff] [blame] | 2213 | th = tcp_hdr(skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2214 | memset(th, 0, sizeof(struct tcphdr)); |
| 2215 | th->syn = 1; |
| 2216 | th->ack = 1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2217 | TCP_ECN_make_synack(req, th); |
| 2218 | th->source = inet_sk(sk)->sport; |
Arnaldo Carvalho de Melo | 2e6599c | 2005-06-18 22:46:52 -0700 | [diff] [blame] | 2219 | th->dest = ireq->rmt_port; |
Ilpo Järvinen | e870a8e | 2008-01-03 20:39:01 -0800 | [diff] [blame] | 2220 | /* Setting of flags are superfluous here for callers (and ECE is |
| 2221 | * not even correctly set) |
| 2222 | */ |
| 2223 | tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn, |
| 2224 | TCPCB_FLAG_SYN | TCPCB_FLAG_ACK); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2225 | th->seq = htonl(TCP_SKB_CB(skb)->seq); |
Arnaldo Carvalho de Melo | 2e6599c | 2005-06-18 22:46:52 -0700 | [diff] [blame] | 2226 | th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2227 | if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */ |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2228 | __u8 rcv_wscale; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2229 | /* Set this up on the first call only */ |
| 2230 | req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW); |
| 2231 | /* tcp_full_space because it is guaranteed to be the first packet */ |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2232 | tcp_select_initial_window(tcp_full_space(sk), |
Arnaldo Carvalho de Melo | 2e6599c | 2005-06-18 22:46:52 -0700 | [diff] [blame] | 2233 | dst_metric(dst, RTAX_ADVMSS) - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0), |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2234 | &req->rcv_wnd, |
| 2235 | &req->window_clamp, |
Arnaldo Carvalho de Melo | 2e6599c | 2005-06-18 22:46:52 -0700 | [diff] [blame] | 2236 | ireq->wscale_ok, |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2237 | &rcv_wscale); |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2238 | ireq->rcv_wscale = rcv_wscale; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2239 | } |
| 2240 | |
| 2241 | /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ |
Ilpo Järvinen | 600ff0c | 2007-02-13 12:42:11 -0800 | [diff] [blame] | 2242 | th->window = htons(min(req->rcv_wnd, 65535U)); |
Florian Westphal | 4dfc281 | 2008-04-10 03:12:40 -0700 | [diff] [blame] | 2243 | #ifdef CONFIG_SYN_COOKIES |
| 2244 | if (unlikely(req->cookie_ts)) |
| 2245 | TCP_SKB_CB(skb)->when = cookie_init_timestamp(req); |
| 2246 | else |
| 2247 | #endif |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2248 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
Al Viro | df7a3b0 | 2006-09-27 18:38:52 -0700 | [diff] [blame] | 2249 | tcp_syn_build_options((__be32 *)(th + 1), dst_metric(dst, RTAX_ADVMSS), ireq->tstamp_ok, |
Arnaldo Carvalho de Melo | 2e6599c | 2005-06-18 22:46:52 -0700 | [diff] [blame] | 2250 | ireq->sack_ok, ireq->wscale_ok, ireq->rcv_wscale, |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2251 | TCP_SKB_CB(skb)->when, |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 2252 | req->ts_recent, |
| 2253 | ( |
| 2254 | #ifdef CONFIG_TCP_MD5SIG |
| 2255 | md5 ? &md5_hash_location : |
| 2256 | #endif |
| 2257 | NULL) |
| 2258 | ); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2259 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2260 | th->doff = (tcp_header_size >> 2); |
| 2261 | TCP_INC_STATS(TCP_MIB_OUTSEGS); |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 2262 | |
| 2263 | #ifdef CONFIG_TCP_MD5SIG |
| 2264 | /* Okay, we have all we need - do the md5 hash if needed */ |
| 2265 | if (md5) { |
| 2266 | tp->af_specific->calc_md5_hash(md5_hash_location, |
| 2267 | md5, |
| 2268 | NULL, dst, req, |
Arnaldo Carvalho de Melo | aa8223c | 2007-04-10 21:04:22 -0700 | [diff] [blame] | 2269 | tcp_hdr(skb), sk->sk_protocol, |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 2270 | skb->len); |
| 2271 | } |
| 2272 | #endif |
| 2273 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2274 | return skb; |
| 2275 | } |
| 2276 | |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2277 | /* |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2278 | * Do all connect socket setups that can be done AF independent. |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2279 | */ |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 2280 | static void tcp_connect_init(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2281 | { |
| 2282 | struct dst_entry *dst = __sk_dst_get(sk); |
| 2283 | struct tcp_sock *tp = tcp_sk(sk); |
| 2284 | __u8 rcv_wscale; |
| 2285 | |
| 2286 | /* We'll fix this up when we get a response from the other end. |
| 2287 | * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. |
| 2288 | */ |
| 2289 | tp->tcp_header_len = sizeof(struct tcphdr) + |
| 2290 | (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0); |
| 2291 | |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 2292 | #ifdef CONFIG_TCP_MD5SIG |
| 2293 | if (tp->af_specific->md5_lookup(sk, sk) != NULL) |
| 2294 | tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; |
| 2295 | #endif |
| 2296 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2297 | /* If user gave his TCP_MAXSEG, record it to clamp */ |
| 2298 | if (tp->rx_opt.user_mss) |
| 2299 | tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; |
| 2300 | tp->max_window = 0; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 2301 | tcp_mtup_init(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2302 | tcp_sync_mss(sk, dst_mtu(dst)); |
| 2303 | |
| 2304 | if (!tp->window_clamp) |
| 2305 | tp->window_clamp = dst_metric(dst, RTAX_WINDOW); |
| 2306 | tp->advmss = dst_metric(dst, RTAX_ADVMSS); |
| 2307 | tcp_initialize_rcv_mss(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2308 | |
| 2309 | tcp_select_initial_window(tcp_full_space(sk), |
| 2310 | tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), |
| 2311 | &tp->rcv_wnd, |
| 2312 | &tp->window_clamp, |
| 2313 | sysctl_tcp_window_scaling, |
| 2314 | &rcv_wscale); |
| 2315 | |
| 2316 | tp->rx_opt.rcv_wscale = rcv_wscale; |
| 2317 | tp->rcv_ssthresh = tp->rcv_wnd; |
| 2318 | |
| 2319 | sk->sk_err = 0; |
| 2320 | sock_reset_flag(sk, SOCK_DONE); |
| 2321 | tp->snd_wnd = 0; |
| 2322 | tcp_init_wl(tp, tp->write_seq, 0); |
| 2323 | tp->snd_una = tp->write_seq; |
| 2324 | tp->snd_sml = tp->write_seq; |
| 2325 | tp->rcv_nxt = 0; |
| 2326 | tp->rcv_wup = 0; |
| 2327 | tp->copied_seq = 0; |
| 2328 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2329 | inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; |
| 2330 | inet_csk(sk)->icsk_retransmits = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2331 | tcp_clear_retrans(tp); |
| 2332 | } |
| 2333 | |
| 2334 | /* |
| 2335 | * Build a SYN and send it off. |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2336 | */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2337 | int tcp_connect(struct sock *sk) |
| 2338 | { |
| 2339 | struct tcp_sock *tp = tcp_sk(sk); |
| 2340 | struct sk_buff *buff; |
| 2341 | |
| 2342 | tcp_connect_init(sk); |
| 2343 | |
David S. Miller | d179cd1 | 2005-08-17 14:57:30 -0700 | [diff] [blame] | 2344 | buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2345 | if (unlikely(buff == NULL)) |
| 2346 | return -ENOBUFS; |
| 2347 | |
| 2348 | /* Reserve space for headers. */ |
| 2349 | skb_reserve(buff, MAX_TCP_HEADER); |
| 2350 | |
Wei Yongjun | bd37a08 | 2006-08-07 21:04:15 -0700 | [diff] [blame] | 2351 | tp->snd_nxt = tp->write_seq; |
Ilpo Järvinen | e870a8e | 2008-01-03 20:39:01 -0800 | [diff] [blame] | 2352 | tcp_init_nondata_skb(buff, tp->write_seq++, TCPCB_FLAG_SYN); |
| 2353 | TCP_ECN_send_syn(sk, buff); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2354 | |
| 2355 | /* Send it off. */ |
| 2356 | TCP_SKB_CB(buff)->when = tcp_time_stamp; |
| 2357 | tp->retrans_stamp = TCP_SKB_CB(buff)->when; |
| 2358 | skb_header_release(buff); |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2359 | __tcp_add_write_queue_tail(sk, buff); |
Hideo Aoki | 3ab224b | 2007-12-31 00:11:19 -0800 | [diff] [blame] | 2360 | sk->sk_wmem_queued += buff->truesize; |
| 2361 | sk_mem_charge(sk, buff->truesize); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2362 | tp->packets_out += tcp_skb_pcount(buff); |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 2363 | tcp_transmit_skb(sk, buff, 1, GFP_KERNEL); |
Wei Yongjun | bd37a08 | 2006-08-07 21:04:15 -0700 | [diff] [blame] | 2364 | |
| 2365 | /* We change tp->snd_nxt after the tcp_transmit_skb() call |
| 2366 | * in order to make this packet get counted in tcpOutSegs. |
| 2367 | */ |
| 2368 | tp->snd_nxt = tp->write_seq; |
| 2369 | tp->pushed_seq = tp->write_seq; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2370 | TCP_INC_STATS(TCP_MIB_ACTIVEOPENS); |
| 2371 | |
| 2372 | /* Timer for repeating the SYN until an answer. */ |
Arnaldo Carvalho de Melo | 3f421ba | 2005-08-09 20:11:08 -0700 | [diff] [blame] | 2373 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, |
| 2374 | inet_csk(sk)->icsk_rto, TCP_RTO_MAX); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2375 | return 0; |
| 2376 | } |
| 2377 | |
| 2378 | /* Send out a delayed ack, the caller does the policy checking |
| 2379 | * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() |
| 2380 | * for details. |
| 2381 | */ |
| 2382 | void tcp_send_delayed_ack(struct sock *sk) |
| 2383 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2384 | struct inet_connection_sock *icsk = inet_csk(sk); |
| 2385 | int ato = icsk->icsk_ack.ato; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2386 | unsigned long timeout; |
| 2387 | |
| 2388 | if (ato > TCP_DELACK_MIN) { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2389 | const struct tcp_sock *tp = tcp_sk(sk); |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 2390 | int max_ato = HZ / 2; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2391 | |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 2392 | if (icsk->icsk_ack.pingpong || |
| 2393 | (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2394 | max_ato = TCP_DELACK_MAX; |
| 2395 | |
| 2396 | /* Slow path, intersegment interval is "high". */ |
| 2397 | |
| 2398 | /* 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] | 2399 | * 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] | 2400 | * directly. |
| 2401 | */ |
| 2402 | if (tp->srtt) { |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 2403 | int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2404 | |
| 2405 | if (rtt < max_ato) |
| 2406 | max_ato = rtt; |
| 2407 | } |
| 2408 | |
| 2409 | ato = min(ato, max_ato); |
| 2410 | } |
| 2411 | |
| 2412 | /* Stay within the limit we were given */ |
| 2413 | timeout = jiffies + ato; |
| 2414 | |
| 2415 | /* 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] | 2416 | if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2417 | /* If delack timer was blocked or is about to expire, |
| 2418 | * send ACK now. |
| 2419 | */ |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2420 | if (icsk->icsk_ack.blocked || |
| 2421 | time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2422 | tcp_send_ack(sk); |
| 2423 | return; |
| 2424 | } |
| 2425 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2426 | if (!time_before(timeout, icsk->icsk_ack.timeout)) |
| 2427 | timeout = icsk->icsk_ack.timeout; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2428 | } |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2429 | icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; |
| 2430 | icsk->icsk_ack.timeout = timeout; |
| 2431 | sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2432 | } |
| 2433 | |
| 2434 | /* This routine sends an ack and also updates the window. */ |
| 2435 | void tcp_send_ack(struct sock *sk) |
| 2436 | { |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2437 | struct sk_buff *buff; |
| 2438 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2439 | /* If we have been reset, we may not send again. */ |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2440 | if (sk->sk_state == TCP_CLOSE) |
| 2441 | return; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2442 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2443 | /* We are not putting this on the write queue, so |
| 2444 | * tcp_transmit_skb() will set the ownership to this |
| 2445 | * sock. |
| 2446 | */ |
| 2447 | buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); |
| 2448 | if (buff == NULL) { |
| 2449 | inet_csk_schedule_ack(sk); |
| 2450 | inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; |
| 2451 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, |
| 2452 | TCP_DELACK_MAX, TCP_RTO_MAX); |
| 2453 | return; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2454 | } |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2455 | |
| 2456 | /* Reserve space for headers and prepare control bits. */ |
| 2457 | skb_reserve(buff, MAX_TCP_HEADER); |
Ilpo Järvinen | e870a8e | 2008-01-03 20:39:01 -0800 | [diff] [blame] | 2458 | tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPCB_FLAG_ACK); |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2459 | |
| 2460 | /* Send it off, this clears delayed acks for us. */ |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2461 | TCP_SKB_CB(buff)->when = tcp_time_stamp; |
| 2462 | tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2463 | } |
| 2464 | |
| 2465 | /* This routine sends a packet with an out of date sequence |
| 2466 | * number. It assumes the other end will try to ack it. |
| 2467 | * |
| 2468 | * Question: what should we make while urgent mode? |
| 2469 | * 4.4BSD forces sending single byte of data. We cannot send |
| 2470 | * out of window data, because we have SND.NXT==SND.MAX... |
| 2471 | * |
| 2472 | * Current solution: to send TWO zero-length segments in urgent mode: |
| 2473 | * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is |
| 2474 | * out-of-date with SND.UNA-1 to probe window. |
| 2475 | */ |
| 2476 | static int tcp_xmit_probe_skb(struct sock *sk, int urgent) |
| 2477 | { |
| 2478 | struct tcp_sock *tp = tcp_sk(sk); |
| 2479 | struct sk_buff *skb; |
| 2480 | |
| 2481 | /* We don't queue it, tcp_transmit_skb() sets ownership. */ |
| 2482 | skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2483 | if (skb == NULL) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2484 | return -1; |
| 2485 | |
| 2486 | /* Reserve space for headers and set control bits. */ |
| 2487 | skb_reserve(skb, MAX_TCP_HEADER); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2488 | /* Use a previous sequence. This should cause the other |
| 2489 | * end to send an ack. Don't queue or clone SKB, just |
| 2490 | * send it. |
| 2491 | */ |
Ilpo Järvinen | e870a8e | 2008-01-03 20:39:01 -0800 | [diff] [blame] | 2492 | tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPCB_FLAG_ACK); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2493 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 2494 | return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2495 | } |
| 2496 | |
| 2497 | int tcp_write_wakeup(struct sock *sk) |
| 2498 | { |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2499 | struct tcp_sock *tp = tcp_sk(sk); |
| 2500 | struct sk_buff *skb; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2501 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2502 | if (sk->sk_state == TCP_CLOSE) |
| 2503 | return -1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2504 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2505 | if ((skb = tcp_send_head(sk)) != NULL && |
| 2506 | before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { |
| 2507 | int err; |
| 2508 | unsigned int mss = tcp_current_mss(sk, 0); |
| 2509 | 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] | 2510 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2511 | if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) |
| 2512 | tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2513 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2514 | /* We are probing the opening of a window |
| 2515 | * but the window size is != 0 |
| 2516 | * must have been a result SWS avoidance ( sender ) |
| 2517 | */ |
| 2518 | if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || |
| 2519 | skb->len > mss) { |
| 2520 | seg_size = min(seg_size, mss); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2521 | TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH; |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2522 | if (tcp_fragment(sk, skb, seg_size, mss)) |
| 2523 | return -1; |
| 2524 | } else if (!tcp_skb_pcount(skb)) |
| 2525 | tcp_set_skb_tso_segs(sk, skb, mss); |
| 2526 | |
| 2527 | TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH; |
| 2528 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
| 2529 | err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); |
| 2530 | if (!err) |
| 2531 | tcp_event_new_data_sent(sk, skb); |
| 2532 | return err; |
| 2533 | } else { |
| 2534 | if (tp->urg_mode && |
| 2535 | between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) |
| 2536 | tcp_xmit_probe_skb(sk, 1); |
| 2537 | return tcp_xmit_probe_skb(sk, 0); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2538 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2539 | } |
| 2540 | |
| 2541 | /* A window probe timeout has occurred. If window is not closed send |
| 2542 | * a partial packet else a zero probe. |
| 2543 | */ |
| 2544 | void tcp_send_probe0(struct sock *sk) |
| 2545 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2546 | struct inet_connection_sock *icsk = inet_csk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2547 | struct tcp_sock *tp = tcp_sk(sk); |
| 2548 | int err; |
| 2549 | |
| 2550 | err = tcp_write_wakeup(sk); |
| 2551 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2552 | if (tp->packets_out || !tcp_send_head(sk)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2553 | /* Cancel probe timer, if it is not required. */ |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2554 | icsk->icsk_probes_out = 0; |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2555 | icsk->icsk_backoff = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2556 | return; |
| 2557 | } |
| 2558 | |
| 2559 | if (err <= 0) { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2560 | if (icsk->icsk_backoff < sysctl_tcp_retries2) |
| 2561 | icsk->icsk_backoff++; |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2562 | icsk->icsk_probes_out++; |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2563 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, |
Arnaldo Carvalho de Melo | 3f421ba | 2005-08-09 20:11:08 -0700 | [diff] [blame] | 2564 | min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), |
| 2565 | TCP_RTO_MAX); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2566 | } else { |
| 2567 | /* If packet was not sent due to local congestion, |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2568 | * do not backoff and do not remember icsk_probes_out. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2569 | * Let local senders to fight for local resources. |
| 2570 | * |
| 2571 | * Use accumulated backoff yet. |
| 2572 | */ |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2573 | if (!icsk->icsk_probes_out) |
| 2574 | icsk->icsk_probes_out = 1; |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2575 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2576 | min(icsk->icsk_rto << icsk->icsk_backoff, |
Arnaldo Carvalho de Melo | 3f421ba | 2005-08-09 20:11:08 -0700 | [diff] [blame] | 2577 | TCP_RESOURCE_PROBE_INTERVAL), |
| 2578 | TCP_RTO_MAX); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2579 | } |
| 2580 | } |
| 2581 | |
Glenn Griffin | c6aefaf | 2008-02-07 21:49:26 -0800 | [diff] [blame] | 2582 | EXPORT_SYMBOL(tcp_select_initial_window); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2583 | EXPORT_SYMBOL(tcp_connect); |
| 2584 | EXPORT_SYMBOL(tcp_make_synack); |
| 2585 | EXPORT_SYMBOL(tcp_simple_retransmit); |
| 2586 | EXPORT_SYMBOL(tcp_sync_mss); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 2587 | EXPORT_SYMBOL(tcp_mtup_init); |