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> |
| 43 | #include <linux/smp_lock.h> |
| 44 | |
| 45 | /* People can turn this off for buggy TCP's found in printers etc. */ |
Brian Haley | ab32ea5 | 2006-09-22 14:15:41 -0700 | [diff] [blame] | 46 | int sysctl_tcp_retrans_collapse __read_mostly = 1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 47 | |
Rick Jones | 15d99e0 | 2006-03-20 22:40:29 -0800 | [diff] [blame] | 48 | /* People can turn this on to work with those rare, broken TCPs that |
| 49 | * interpret the window field as a signed quantity. |
| 50 | */ |
Brian Haley | ab32ea5 | 2006-09-22 14:15:41 -0700 | [diff] [blame] | 51 | int sysctl_tcp_workaround_signed_windows __read_mostly = 0; |
Rick Jones | 15d99e0 | 2006-03-20 22:40:29 -0800 | [diff] [blame] | 52 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 53 | /* This limits the percentage of the congestion window which we |
| 54 | * will allow a single TSO frame to consume. Building TSO frames |
| 55 | * which are too large can cause TCP streams to be bursty. |
| 56 | */ |
Brian Haley | ab32ea5 | 2006-09-22 14:15:41 -0700 | [diff] [blame] | 57 | int sysctl_tcp_tso_win_divisor __read_mostly = 3; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 58 | |
Brian Haley | ab32ea5 | 2006-09-22 14:15:41 -0700 | [diff] [blame] | 59 | int sysctl_tcp_mtu_probing __read_mostly = 0; |
| 60 | int sysctl_tcp_base_mss __read_mostly = 512; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 61 | |
David S. Miller | 35089bb | 2006-06-13 22:33:04 -0700 | [diff] [blame] | 62 | /* By default, RFC2861 behavior. */ |
Brian Haley | ab32ea5 | 2006-09-22 14:15:41 -0700 | [diff] [blame] | 63 | int sysctl_tcp_slow_start_after_idle __read_mostly = 1; |
David S. Miller | 35089bb | 2006-06-13 22:33:04 -0700 | [diff] [blame] | 64 | |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 65 | static void update_send_head(struct sock *sk, struct sk_buff *skb) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 66 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 67 | struct tcp_sock *tp = tcp_sk(sk); |
| 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 | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 71 | tcp_packets_out_inc(sk, skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 72 | } |
| 73 | |
| 74 | /* SND.NXT, if window was not shrunk. |
| 75 | * If window has been shrunk, what should we make? It is not clear at all. |
| 76 | * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-( |
| 77 | * Anything in between SND.UNA...SND.UNA+SND.WND also can be already |
| 78 | * invalid. OK, let's make this for now: |
| 79 | */ |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 80 | static inline __u32 tcp_acceptable_seq(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 81 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 82 | struct tcp_sock *tp = tcp_sk(sk); |
| 83 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 84 | if (!before(tp->snd_una+tp->snd_wnd, tp->snd_nxt)) |
| 85 | return tp->snd_nxt; |
| 86 | else |
| 87 | return tp->snd_una+tp->snd_wnd; |
| 88 | } |
| 89 | |
| 90 | /* Calculate mss to advertise in SYN segment. |
| 91 | * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that: |
| 92 | * |
| 93 | * 1. It is independent of path mtu. |
| 94 | * 2. Ideally, it is maximal possible segment size i.e. 65535-40. |
| 95 | * 3. For IPv4 it is reasonable to calculate it from maximal MTU of |
| 96 | * attached devices, because some buggy hosts are confused by |
| 97 | * large MSS. |
| 98 | * 4. We do not make 3, we advertise MSS, calculated from first |
| 99 | * hop device mtu, but allow to raise it to ip_rt_min_advmss. |
| 100 | * This may be overridden via information stored in routing table. |
| 101 | * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible, |
| 102 | * probably even Jumbo". |
| 103 | */ |
| 104 | static __u16 tcp_advertise_mss(struct sock *sk) |
| 105 | { |
| 106 | struct tcp_sock *tp = tcp_sk(sk); |
| 107 | struct dst_entry *dst = __sk_dst_get(sk); |
| 108 | int mss = tp->advmss; |
| 109 | |
| 110 | if (dst && dst_metric(dst, RTAX_ADVMSS) < mss) { |
| 111 | mss = dst_metric(dst, RTAX_ADVMSS); |
| 112 | tp->advmss = mss; |
| 113 | } |
| 114 | |
| 115 | return (__u16)mss; |
| 116 | } |
| 117 | |
| 118 | /* RFC2861. Reset CWND after idle period longer RTO to "restart window". |
| 119 | * This is the first part of cwnd validation mechanism. */ |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 120 | static void tcp_cwnd_restart(struct sock *sk, struct dst_entry *dst) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 121 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 122 | struct tcp_sock *tp = tcp_sk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 123 | s32 delta = tcp_time_stamp - tp->lsndtime; |
| 124 | u32 restart_cwnd = tcp_init_cwnd(tp, dst); |
| 125 | u32 cwnd = tp->snd_cwnd; |
| 126 | |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 127 | tcp_ca_event(sk, CA_EVENT_CWND_RESTART); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 128 | |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 129 | tp->snd_ssthresh = tcp_current_ssthresh(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 130 | restart_cwnd = min(restart_cwnd, cwnd); |
| 131 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 132 | while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 133 | cwnd >>= 1; |
| 134 | tp->snd_cwnd = max(cwnd, restart_cwnd); |
| 135 | tp->snd_cwnd_stamp = tcp_time_stamp; |
| 136 | tp->snd_cwnd_used = 0; |
| 137 | } |
| 138 | |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 139 | static void tcp_event_data_sent(struct tcp_sock *tp, |
| 140 | struct sk_buff *skb, struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 141 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 142 | struct inet_connection_sock *icsk = inet_csk(sk); |
| 143 | const u32 now = tcp_time_stamp; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 144 | |
David S. Miller | 35089bb | 2006-06-13 22:33:04 -0700 | [diff] [blame] | 145 | if (sysctl_tcp_slow_start_after_idle && |
| 146 | (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto)) |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 147 | tcp_cwnd_restart(sk, __sk_dst_get(sk)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 148 | |
| 149 | tp->lsndtime = now; |
| 150 | |
| 151 | /* If it is a reply for ato after last received |
| 152 | * packet, enter pingpong mode. |
| 153 | */ |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 154 | if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato) |
| 155 | icsk->icsk_ack.pingpong = 1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 156 | } |
| 157 | |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 158 | 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] | 159 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 160 | tcp_dec_quickack_mode(sk, pkts); |
| 161 | inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 162 | } |
| 163 | |
| 164 | /* Determine a window scaling and initial window to offer. |
| 165 | * Based on the assumption that the given amount of space |
| 166 | * will be offered. Store the results in the tp structure. |
| 167 | * NOTE: for smooth operation initial space offering should |
| 168 | * be a multiple of mss if possible. We assume here that mss >= 1. |
| 169 | * This MUST be enforced by all callers. |
| 170 | */ |
| 171 | void tcp_select_initial_window(int __space, __u32 mss, |
| 172 | __u32 *rcv_wnd, __u32 *window_clamp, |
| 173 | int wscale_ok, __u8 *rcv_wscale) |
| 174 | { |
| 175 | unsigned int space = (__space < 0 ? 0 : __space); |
| 176 | |
| 177 | /* If no clamp set the clamp to the max possible scaled window */ |
| 178 | if (*window_clamp == 0) |
| 179 | (*window_clamp) = (65535 << 14); |
| 180 | space = min(*window_clamp, space); |
| 181 | |
| 182 | /* Quantize space offering to a multiple of mss if possible. */ |
| 183 | if (space > mss) |
| 184 | space = (space / mss) * mss; |
| 185 | |
| 186 | /* NOTE: offering an initial window larger than 32767 |
Rick Jones | 15d99e0 | 2006-03-20 22:40:29 -0800 | [diff] [blame] | 187 | * will break some buggy TCP stacks. If the admin tells us |
| 188 | * it is likely we could be speaking with such a buggy stack |
| 189 | * we will truncate our initial window offering to 32K-1 |
| 190 | * unless the remote has sent us a window scaling option, |
| 191 | * which we interpret as a sign the remote TCP is not |
| 192 | * misinterpreting the window field as a signed quantity. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 193 | */ |
Rick Jones | 15d99e0 | 2006-03-20 22:40:29 -0800 | [diff] [blame] | 194 | if (sysctl_tcp_workaround_signed_windows) |
| 195 | (*rcv_wnd) = min(space, MAX_TCP_WINDOW); |
| 196 | else |
| 197 | (*rcv_wnd) = space; |
| 198 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 199 | (*rcv_wscale) = 0; |
| 200 | if (wscale_ok) { |
| 201 | /* Set window scaling on max possible window |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 202 | * See RFC1323 for an explanation of the limit to 14 |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 203 | */ |
| 204 | space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max); |
Stephen Hemminger | 316c159 | 2006-08-22 00:06:11 -0700 | [diff] [blame] | 205 | space = min_t(u32, space, *window_clamp); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 206 | while (space > 65535 && (*rcv_wscale) < 14) { |
| 207 | space >>= 1; |
| 208 | (*rcv_wscale)++; |
| 209 | } |
| 210 | } |
| 211 | |
| 212 | /* Set initial window to value enough for senders, |
David S. Miller | 6b25185 | 2005-09-28 16:31:48 -0700 | [diff] [blame] | 213 | * following RFC2414. Senders, not following this RFC, |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 214 | * will be satisfied with 2. |
| 215 | */ |
| 216 | if (mss > (1<<*rcv_wscale)) { |
David S. Miller | 01ff367 | 2005-09-29 17:07:20 -0700 | [diff] [blame] | 217 | int init_cwnd = 4; |
| 218 | if (mss > 1460*3) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 219 | init_cwnd = 2; |
David S. Miller | 01ff367 | 2005-09-29 17:07:20 -0700 | [diff] [blame] | 220 | else if (mss > 1460) |
| 221 | init_cwnd = 3; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 222 | if (*rcv_wnd > init_cwnd*mss) |
| 223 | *rcv_wnd = init_cwnd*mss; |
| 224 | } |
| 225 | |
| 226 | /* Set the clamp no higher than max representable value */ |
| 227 | (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp); |
| 228 | } |
| 229 | |
| 230 | /* Chose a new window to advertise, update state in tcp_sock for the |
| 231 | * socket, and return result with RFC1323 scaling applied. The return |
| 232 | * value can be stuffed directly into th->window for an outgoing |
| 233 | * frame. |
| 234 | */ |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 235 | static u16 tcp_select_window(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 236 | { |
| 237 | struct tcp_sock *tp = tcp_sk(sk); |
| 238 | u32 cur_win = tcp_receive_window(tp); |
| 239 | u32 new_win = __tcp_select_window(sk); |
| 240 | |
| 241 | /* Never shrink the offered window */ |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 242 | if (new_win < cur_win) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 243 | /* Danger Will Robinson! |
| 244 | * Don't update rcv_wup/rcv_wnd here or else |
| 245 | * we will not be able to advertise a zero |
| 246 | * window in time. --DaveM |
| 247 | * |
| 248 | * Relax Will Robinson. |
| 249 | */ |
| 250 | new_win = cur_win; |
| 251 | } |
| 252 | tp->rcv_wnd = new_win; |
| 253 | tp->rcv_wup = tp->rcv_nxt; |
| 254 | |
| 255 | /* Make sure we do not exceed the maximum possible |
| 256 | * scaled window. |
| 257 | */ |
Rick Jones | 15d99e0 | 2006-03-20 22:40:29 -0800 | [diff] [blame] | 258 | if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 259 | new_win = min(new_win, MAX_TCP_WINDOW); |
| 260 | else |
| 261 | new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); |
| 262 | |
| 263 | /* RFC1323 scaling applied */ |
| 264 | new_win >>= tp->rx_opt.rcv_wscale; |
| 265 | |
| 266 | /* If we advertise zero window, disable fast path. */ |
| 267 | if (new_win == 0) |
| 268 | tp->pred_flags = 0; |
| 269 | |
| 270 | return new_win; |
| 271 | } |
| 272 | |
Al Viro | df7a3b0 | 2006-09-27 18:38:52 -0700 | [diff] [blame] | 273 | 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] | 274 | __u32 tstamp, __u8 **md5_hash) |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 275 | { |
| 276 | if (tp->rx_opt.tstamp_ok) { |
YOSHIFUJI Hideaki | 496c98d | 2006-10-10 19:41:21 -0700 | [diff] [blame] | 277 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
| 278 | (TCPOPT_NOP << 16) | |
| 279 | (TCPOPT_TIMESTAMP << 8) | |
| 280 | TCPOLEN_TIMESTAMP); |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 281 | *ptr++ = htonl(tstamp); |
| 282 | *ptr++ = htonl(tp->rx_opt.ts_recent); |
| 283 | } |
| 284 | if (tp->rx_opt.eff_sacks) { |
| 285 | struct tcp_sack_block *sp = tp->rx_opt.dsack ? tp->duplicate_sack : tp->selective_acks; |
| 286 | int this_sack; |
| 287 | |
| 288 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
| 289 | (TCPOPT_NOP << 16) | |
| 290 | (TCPOPT_SACK << 8) | |
| 291 | (TCPOLEN_SACK_BASE + (tp->rx_opt.eff_sacks * |
| 292 | TCPOLEN_SACK_PERBLOCK))); |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 293 | |
| 294 | 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] | 295 | *ptr++ = htonl(sp[this_sack].start_seq); |
| 296 | *ptr++ = htonl(sp[this_sack].end_seq); |
| 297 | } |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 298 | |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 299 | if (tp->rx_opt.dsack) { |
| 300 | tp->rx_opt.dsack = 0; |
| 301 | tp->rx_opt.eff_sacks--; |
| 302 | } |
| 303 | } |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 304 | #ifdef CONFIG_TCP_MD5SIG |
| 305 | if (md5_hash) { |
| 306 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
| 307 | (TCPOPT_NOP << 16) | |
| 308 | (TCPOPT_MD5SIG << 8) | |
| 309 | TCPOLEN_MD5SIG); |
| 310 | *md5_hash = (__u8 *)ptr; |
| 311 | } |
| 312 | #endif |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 313 | } |
| 314 | |
| 315 | /* Construct a tcp options header for a SYN or SYN_ACK packet. |
| 316 | * If this is every changed make sure to change the definition of |
| 317 | * MAX_SYN_SIZE to match the new maximum number of options that you |
| 318 | * can generate. |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 319 | * |
| 320 | * Note - that with the RFC2385 TCP option, we make room for the |
| 321 | * 16 byte MD5 hash. This will be filled in later, so the pointer for the |
| 322 | * location to be filled is passed back up. |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 323 | */ |
Al Viro | df7a3b0 | 2006-09-27 18:38:52 -0700 | [diff] [blame] | 324 | 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] | 325 | int offer_wscale, int wscale, __u32 tstamp, |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 326 | __u32 ts_recent, __u8 **md5_hash) |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 327 | { |
| 328 | /* We always get an MSS option. |
| 329 | * The option bytes which will be seen in normal data |
| 330 | * packets should timestamps be used, must be in the MSS |
| 331 | * advertised. But we subtract them from tp->mss_cache so |
| 332 | * that calculations in tcp_sendmsg are simpler etc. |
| 333 | * So account for this fact here if necessary. If we |
| 334 | * don't do this correctly, as a receiver we won't |
| 335 | * recognize data packets as being full sized when we |
| 336 | * should, and thus we won't abide by the delayed ACK |
| 337 | * rules correctly. |
| 338 | * SACKs don't matter, we never delay an ACK when we |
| 339 | * have any of those going out. |
| 340 | */ |
| 341 | *ptr++ = htonl((TCPOPT_MSS << 24) | (TCPOLEN_MSS << 16) | mss); |
| 342 | if (ts) { |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 343 | if (sack) |
YOSHIFUJI Hideaki | 496c98d | 2006-10-10 19:41:21 -0700 | [diff] [blame] | 344 | *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | |
| 345 | (TCPOLEN_SACK_PERM << 16) | |
| 346 | (TCPOPT_TIMESTAMP << 8) | |
| 347 | TCPOLEN_TIMESTAMP); |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 348 | else |
YOSHIFUJI Hideaki | 496c98d | 2006-10-10 19:41:21 -0700 | [diff] [blame] | 349 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
| 350 | (TCPOPT_NOP << 16) | |
| 351 | (TCPOPT_TIMESTAMP << 8) | |
| 352 | TCPOLEN_TIMESTAMP); |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 353 | *ptr++ = htonl(tstamp); /* TSVAL */ |
| 354 | *ptr++ = htonl(ts_recent); /* TSECR */ |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 355 | } else if (sack) |
YOSHIFUJI Hideaki | 496c98d | 2006-10-10 19:41:21 -0700 | [diff] [blame] | 356 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
| 357 | (TCPOPT_NOP << 16) | |
| 358 | (TCPOPT_SACK_PERM << 8) | |
| 359 | TCPOLEN_SACK_PERM); |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 360 | if (offer_wscale) |
YOSHIFUJI Hideaki | 496c98d | 2006-10-10 19:41:21 -0700 | [diff] [blame] | 361 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
| 362 | (TCPOPT_WINDOW << 16) | |
| 363 | (TCPOLEN_WINDOW << 8) | |
| 364 | (wscale)); |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 365 | #ifdef CONFIG_TCP_MD5SIG |
| 366 | /* |
| 367 | * If MD5 is enabled, then we set the option, and include the size |
| 368 | * (always 18). The actual MD5 hash is added just before the |
| 369 | * packet is sent. |
| 370 | */ |
| 371 | if (md5_hash) { |
| 372 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
| 373 | (TCPOPT_NOP << 16) | |
| 374 | (TCPOPT_MD5SIG << 8) | |
| 375 | TCPOLEN_MD5SIG); |
| 376 | *md5_hash = (__u8 *) ptr; |
| 377 | } |
| 378 | #endif |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 379 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 380 | |
| 381 | /* This routine actually transmits TCP packets queued in by |
| 382 | * tcp_do_sendmsg(). This is used by both the initial |
| 383 | * transmission and possible later retransmissions. |
| 384 | * All SKB's seen here are completely headerless. It is our |
| 385 | * job to build the TCP header, and pass the packet down to |
| 386 | * IP so it can do the same plus pass the packet off to the |
| 387 | * device. |
| 388 | * |
| 389 | * We are working here with either a clone of the original |
| 390 | * SKB, or a fresh unique copy made by the retransmit engine. |
| 391 | */ |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 392 | static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, gfp_t gfp_mask) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 393 | { |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 394 | const struct inet_connection_sock *icsk = inet_csk(sk); |
| 395 | struct inet_sock *inet; |
| 396 | struct tcp_sock *tp; |
| 397 | struct tcp_skb_cb *tcb; |
| 398 | int tcp_header_size; |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 399 | #ifdef CONFIG_TCP_MD5SIG |
| 400 | struct tcp_md5sig_key *md5; |
| 401 | __u8 *md5_hash_location; |
| 402 | #endif |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 403 | struct tcphdr *th; |
| 404 | int sysctl_flags; |
| 405 | int err; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 406 | |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 407 | BUG_ON(!skb || !tcp_skb_pcount(skb)); |
| 408 | |
| 409 | /* If congestion control is doing timestamping, we must |
| 410 | * take such a timestamp before we potentially clone/copy. |
| 411 | */ |
Stephen Hemminger | 164891a | 2007-04-23 22:26:16 -0700 | [diff] [blame] | 412 | if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP) |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 413 | __net_timestamp(skb); |
| 414 | |
| 415 | if (likely(clone_it)) { |
| 416 | if (unlikely(skb_cloned(skb))) |
| 417 | skb = pskb_copy(skb, gfp_mask); |
| 418 | else |
| 419 | skb = skb_clone(skb, gfp_mask); |
| 420 | if (unlikely(!skb)) |
| 421 | return -ENOBUFS; |
| 422 | } |
| 423 | |
| 424 | inet = inet_sk(sk); |
| 425 | tp = tcp_sk(sk); |
| 426 | tcb = TCP_SKB_CB(skb); |
| 427 | tcp_header_size = tp->tcp_header_len; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 428 | |
| 429 | #define SYSCTL_FLAG_TSTAMPS 0x1 |
| 430 | #define SYSCTL_FLAG_WSCALE 0x2 |
| 431 | #define SYSCTL_FLAG_SACK 0x4 |
| 432 | |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 433 | sysctl_flags = 0; |
| 434 | if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) { |
| 435 | tcp_header_size = sizeof(struct tcphdr) + TCPOLEN_MSS; |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 436 | if (sysctl_tcp_timestamps) { |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 437 | tcp_header_size += TCPOLEN_TSTAMP_ALIGNED; |
| 438 | sysctl_flags |= SYSCTL_FLAG_TSTAMPS; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 439 | } |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 440 | if (sysctl_tcp_window_scaling) { |
| 441 | tcp_header_size += TCPOLEN_WSCALE_ALIGNED; |
| 442 | sysctl_flags |= SYSCTL_FLAG_WSCALE; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 443 | } |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 444 | if (sysctl_tcp_sack) { |
| 445 | sysctl_flags |= SYSCTL_FLAG_SACK; |
| 446 | if (!(sysctl_flags & SYSCTL_FLAG_TSTAMPS)) |
| 447 | tcp_header_size += TCPOLEN_SACKPERM_ALIGNED; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 448 | } |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 449 | } else if (unlikely(tp->rx_opt.eff_sacks)) { |
| 450 | /* A SACK is 2 pad bytes, a 2 byte header, plus |
| 451 | * 2 32-bit sequence numbers for each SACK block. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 452 | */ |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 453 | tcp_header_size += (TCPOLEN_SACK_BASE_ALIGNED + |
| 454 | (tp->rx_opt.eff_sacks * |
| 455 | TCPOLEN_SACK_PERBLOCK)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 456 | } |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 457 | |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 458 | if (tcp_packets_in_flight(tp) == 0) |
| 459 | tcp_ca_event(sk, CA_EVENT_TX_START); |
| 460 | |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 461 | #ifdef CONFIG_TCP_MD5SIG |
| 462 | /* |
| 463 | * Are we doing MD5 on this segment? If so - make |
| 464 | * room for it. |
| 465 | */ |
| 466 | md5 = tp->af_specific->md5_lookup(sk, sk); |
| 467 | if (md5) |
| 468 | tcp_header_size += TCPOLEN_MD5SIG_ALIGNED; |
| 469 | #endif |
| 470 | |
Arnaldo Carvalho de Melo | aa8223c | 2007-04-10 21:04:22 -0700 | [diff] [blame] | 471 | skb_push(skb, tcp_header_size); |
| 472 | skb_reset_transport_header(skb); |
David S. Miller | e89862f | 2007-01-26 01:04:55 -0800 | [diff] [blame] | 473 | skb_set_owner_w(skb, sk); |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 474 | |
| 475 | /* Build TCP header and checksum it. */ |
Arnaldo Carvalho de Melo | aa8223c | 2007-04-10 21:04:22 -0700 | [diff] [blame] | 476 | th = tcp_hdr(skb); |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 477 | th->source = inet->sport; |
| 478 | th->dest = inet->dport; |
| 479 | th->seq = htonl(tcb->seq); |
| 480 | th->ack_seq = htonl(tp->rcv_nxt); |
Al Viro | df7a3b0 | 2006-09-27 18:38:52 -0700 | [diff] [blame] | 481 | *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 482 | tcb->flags); |
| 483 | |
| 484 | if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) { |
| 485 | /* RFC1323: The window in SYN & SYN/ACK segments |
| 486 | * is never scaled. |
| 487 | */ |
Ilpo Järvinen | 600ff0c | 2007-02-13 12:42:11 -0800 | [diff] [blame] | 488 | th->window = htons(min(tp->rcv_wnd, 65535U)); |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 489 | } else { |
| 490 | th->window = htons(tcp_select_window(sk)); |
| 491 | } |
| 492 | th->check = 0; |
| 493 | th->urg_ptr = 0; |
| 494 | |
| 495 | if (unlikely(tp->urg_mode && |
| 496 | between(tp->snd_up, tcb->seq+1, tcb->seq+0xFFFF))) { |
| 497 | th->urg_ptr = htons(tp->snd_up-tcb->seq); |
| 498 | th->urg = 1; |
| 499 | } |
| 500 | |
| 501 | if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) { |
Al Viro | df7a3b0 | 2006-09-27 18:38:52 -0700 | [diff] [blame] | 502 | tcp_syn_build_options((__be32 *)(th + 1), |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 503 | tcp_advertise_mss(sk), |
| 504 | (sysctl_flags & SYSCTL_FLAG_TSTAMPS), |
| 505 | (sysctl_flags & SYSCTL_FLAG_SACK), |
| 506 | (sysctl_flags & SYSCTL_FLAG_WSCALE), |
| 507 | tp->rx_opt.rcv_wscale, |
| 508 | tcb->when, |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 509 | tp->rx_opt.ts_recent, |
| 510 | |
| 511 | #ifdef CONFIG_TCP_MD5SIG |
| 512 | md5 ? &md5_hash_location : |
| 513 | #endif |
| 514 | NULL); |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 515 | } else { |
Al Viro | df7a3b0 | 2006-09-27 18:38:52 -0700 | [diff] [blame] | 516 | tcp_build_and_update_options((__be32 *)(th + 1), |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 517 | tp, tcb->when, |
| 518 | #ifdef CONFIG_TCP_MD5SIG |
| 519 | md5 ? &md5_hash_location : |
| 520 | #endif |
| 521 | NULL); |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 522 | TCP_ECN_send(sk, skb, tcp_header_size); |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 523 | } |
| 524 | |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 525 | #ifdef CONFIG_TCP_MD5SIG |
| 526 | /* Calculate the MD5 hash, as we have all we need now */ |
| 527 | if (md5) { |
| 528 | tp->af_specific->calc_md5_hash(md5_hash_location, |
| 529 | md5, |
| 530 | sk, NULL, NULL, |
Arnaldo Carvalho de Melo | aa8223c | 2007-04-10 21:04:22 -0700 | [diff] [blame] | 531 | tcp_hdr(skb), |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 532 | sk->sk_protocol, |
| 533 | skb->len); |
| 534 | } |
| 535 | #endif |
| 536 | |
Arnaldo Carvalho de Melo | 8292a17 | 2005-12-13 23:15:52 -0800 | [diff] [blame] | 537 | icsk->icsk_af_ops->send_check(sk, skb->len, skb); |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 538 | |
| 539 | if (likely(tcb->flags & TCPCB_FLAG_ACK)) |
| 540 | tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); |
| 541 | |
| 542 | if (skb->len != tcp_header_size) |
| 543 | tcp_event_data_sent(tp, skb, sk); |
| 544 | |
Wei Yongjun | bd37a08 | 2006-08-07 21:04:15 -0700 | [diff] [blame] | 545 | if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) |
| 546 | TCP_INC_STATS(TCP_MIB_OUTSEGS); |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 547 | |
David S. Miller | e89862f | 2007-01-26 01:04:55 -0800 | [diff] [blame] | 548 | err = icsk->icsk_af_ops->queue_xmit(skb, 0); |
Hua Zhong | 83de47c | 2006-04-28 15:26:50 -0700 | [diff] [blame] | 549 | if (likely(err <= 0)) |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 550 | return err; |
| 551 | |
Ilpo Järvinen | 3cfe3ba | 2007-02-27 10:09:49 -0800 | [diff] [blame] | 552 | tcp_enter_cwr(sk, 1); |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 553 | |
Gerrit Renker | b9df3cb | 2006-11-14 11:21:36 -0200 | [diff] [blame] | 554 | return net_xmit_eval(err); |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 555 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 556 | #undef SYSCTL_FLAG_TSTAMPS |
| 557 | #undef SYSCTL_FLAG_WSCALE |
| 558 | #undef SYSCTL_FLAG_SACK |
| 559 | } |
| 560 | |
| 561 | |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 562 | /* This routine just queue's the buffer |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 563 | * |
| 564 | * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, |
| 565 | * otherwise socket can stall. |
| 566 | */ |
| 567 | static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) |
| 568 | { |
| 569 | struct tcp_sock *tp = tcp_sk(sk); |
| 570 | |
| 571 | /* Advance write_seq and place onto the write_queue. */ |
| 572 | tp->write_seq = TCP_SKB_CB(skb)->end_seq; |
| 573 | skb_header_release(skb); |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 574 | tcp_add_write_queue_tail(sk, skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 575 | sk_charge_skb(sk, skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 576 | } |
| 577 | |
David S. Miller | 846998a | 2005-08-04 19:52:01 -0700 | [diff] [blame] | 578 | static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb, unsigned int mss_now) |
David S. Miller | f6302d1 | 2005-07-05 15:18:03 -0700 | [diff] [blame] | 579 | { |
Herbert Xu | bcd7611 | 2006-06-30 13:36:35 -0700 | [diff] [blame] | 580 | if (skb->len <= mss_now || !sk_can_gso(sk)) { |
David S. Miller | f6302d1 | 2005-07-05 15:18:03 -0700 | [diff] [blame] | 581 | /* Avoid the costly divide in the normal |
| 582 | * non-TSO case. |
| 583 | */ |
Herbert Xu | 7967168 | 2006-06-22 02:40:14 -0700 | [diff] [blame] | 584 | skb_shinfo(skb)->gso_segs = 1; |
| 585 | skb_shinfo(skb)->gso_size = 0; |
| 586 | skb_shinfo(skb)->gso_type = 0; |
David S. Miller | f6302d1 | 2005-07-05 15:18:03 -0700 | [diff] [blame] | 587 | } else { |
| 588 | unsigned int factor; |
| 589 | |
David S. Miller | 846998a | 2005-08-04 19:52:01 -0700 | [diff] [blame] | 590 | factor = skb->len + (mss_now - 1); |
| 591 | factor /= mss_now; |
Herbert Xu | 7967168 | 2006-06-22 02:40:14 -0700 | [diff] [blame] | 592 | skb_shinfo(skb)->gso_segs = factor; |
| 593 | skb_shinfo(skb)->gso_size = mss_now; |
Herbert Xu | bcd7611 | 2006-06-30 13:36:35 -0700 | [diff] [blame] | 594 | skb_shinfo(skb)->gso_type = sk->sk_gso_type; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 595 | } |
| 596 | } |
| 597 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 598 | /* Function to create two new TCP segments. Shrinks the given segment |
| 599 | * 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] | 600 | * packet to the list. This won't be called frequently, I hope. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 601 | * Remember, these are still headerless SKBs at this point. |
| 602 | */ |
David S. Miller | 6475be1 | 2005-09-01 22:47:01 -0700 | [diff] [blame] | 603 | int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, unsigned int mss_now) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 604 | { |
| 605 | struct tcp_sock *tp = tcp_sk(sk); |
| 606 | struct sk_buff *buff; |
David S. Miller | 6475be1 | 2005-09-01 22:47:01 -0700 | [diff] [blame] | 607 | int nsize, old_factor; |
Herbert Xu | b60b49e | 2006-04-19 21:35:00 -0700 | [diff] [blame] | 608 | int nlen; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 609 | u16 flags; |
| 610 | |
Herbert Xu | b2cc99f0 | 2005-10-20 17:13:13 -0200 | [diff] [blame] | 611 | BUG_ON(len > skb->len); |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 612 | |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 613 | clear_all_retrans_hints(tp); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 614 | nsize = skb_headlen(skb) - len; |
| 615 | if (nsize < 0) |
| 616 | nsize = 0; |
| 617 | |
| 618 | if (skb_cloned(skb) && |
| 619 | skb_is_nonlinear(skb) && |
| 620 | pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) |
| 621 | return -ENOMEM; |
| 622 | |
| 623 | /* Get a new skb... force flag on. */ |
| 624 | buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC); |
| 625 | if (buff == NULL) |
| 626 | return -ENOMEM; /* We'll just try again later. */ |
Herbert Xu | ef5cb97 | 2006-04-18 13:24:14 -0700 | [diff] [blame] | 627 | |
Herbert Xu | b60b49e | 2006-04-19 21:35:00 -0700 | [diff] [blame] | 628 | sk_charge_skb(sk, buff); |
| 629 | nlen = skb->len - len - nsize; |
| 630 | buff->truesize += nlen; |
| 631 | skb->truesize -= nlen; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 632 | |
| 633 | /* Correct the sequence numbers. */ |
| 634 | TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; |
| 635 | TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; |
| 636 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; |
| 637 | |
| 638 | /* PSH and FIN should only be set in the second packet. */ |
| 639 | flags = TCP_SKB_CB(skb)->flags; |
| 640 | TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH); |
| 641 | TCP_SKB_CB(buff)->flags = flags; |
Herbert Xu | e14c3ca | 2005-09-19 18:18:38 -0700 | [diff] [blame] | 642 | TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 643 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_AT_TAIL; |
| 644 | |
Patrick McHardy | 84fa793 | 2006-08-29 16:44:56 -0700 | [diff] [blame] | 645 | if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 646 | /* Copy and checksum data tail into the new buffer. */ |
| 647 | buff->csum = csum_partial_copy_nocheck(skb->data + len, skb_put(buff, nsize), |
| 648 | nsize, 0); |
| 649 | |
| 650 | skb_trim(skb, len); |
| 651 | |
| 652 | skb->csum = csum_block_sub(skb->csum, buff->csum, len); |
| 653 | } else { |
Patrick McHardy | 84fa793 | 2006-08-29 16:44:56 -0700 | [diff] [blame] | 654 | skb->ip_summed = CHECKSUM_PARTIAL; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 655 | skb_split(skb, buff, len); |
| 656 | } |
| 657 | |
| 658 | buff->ip_summed = skb->ip_summed; |
| 659 | |
| 660 | /* Looks stupid, but our code really uses when of |
| 661 | * skbs, which it never sent before. --ANK |
| 662 | */ |
| 663 | TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when; |
Patrick McHardy | a61bbcf | 2005-08-14 17:24:31 -0700 | [diff] [blame] | 664 | buff->tstamp = skb->tstamp; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 665 | |
David S. Miller | 6475be1 | 2005-09-01 22:47:01 -0700 | [diff] [blame] | 666 | old_factor = tcp_skb_pcount(skb); |
| 667 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 668 | /* Fix up tso_factor for both original and new SKB. */ |
David S. Miller | 846998a | 2005-08-04 19:52:01 -0700 | [diff] [blame] | 669 | tcp_set_skb_tso_segs(sk, skb, mss_now); |
| 670 | tcp_set_skb_tso_segs(sk, buff, mss_now); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 671 | |
David S. Miller | 6475be1 | 2005-09-01 22:47:01 -0700 | [diff] [blame] | 672 | /* If this packet has been sent out already, we must |
| 673 | * adjust the various packet counters. |
| 674 | */ |
Herbert Xu | cf0b450 | 2005-09-08 15:10:52 -0700 | [diff] [blame] | 675 | if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { |
David S. Miller | 6475be1 | 2005-09-01 22:47:01 -0700 | [diff] [blame] | 676 | int diff = old_factor - tcp_skb_pcount(skb) - |
| 677 | tcp_skb_pcount(buff); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 678 | |
David S. Miller | 6475be1 | 2005-09-01 22:47:01 -0700 | [diff] [blame] | 679 | tp->packets_out -= diff; |
Herbert Xu | e14c3ca | 2005-09-19 18:18:38 -0700 | [diff] [blame] | 680 | |
| 681 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) |
| 682 | tp->sacked_out -= diff; |
| 683 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) |
| 684 | tp->retrans_out -= diff; |
| 685 | |
David S. Miller | 6475be1 | 2005-09-01 22:47:01 -0700 | [diff] [blame] | 686 | if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) { |
| 687 | tp->lost_out -= diff; |
| 688 | tp->left_out -= diff; |
| 689 | } |
Herbert Xu | 83ca28b | 2005-09-22 23:32:56 -0700 | [diff] [blame] | 690 | |
David S. Miller | 6475be1 | 2005-09-01 22:47:01 -0700 | [diff] [blame] | 691 | if (diff > 0) { |
Herbert Xu | 83ca28b | 2005-09-22 23:32:56 -0700 | [diff] [blame] | 692 | /* Adjust Reno SACK estimate. */ |
| 693 | if (!tp->rx_opt.sack_ok) { |
| 694 | tp->sacked_out -= diff; |
| 695 | if ((int)tp->sacked_out < 0) |
| 696 | tp->sacked_out = 0; |
| 697 | tcp_sync_left_out(tp); |
| 698 | } |
| 699 | |
David S. Miller | 6475be1 | 2005-09-01 22:47:01 -0700 | [diff] [blame] | 700 | tp->fackets_out -= diff; |
| 701 | if ((int)tp->fackets_out < 0) |
| 702 | tp->fackets_out = 0; |
| 703 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 704 | } |
| 705 | |
| 706 | /* Link BUFF into the send queue. */ |
David S. Miller | f44b527 | 2005-07-05 15:18:34 -0700 | [diff] [blame] | 707 | skb_header_release(buff); |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 708 | tcp_insert_write_queue_after(skb, buff, sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 709 | |
| 710 | return 0; |
| 711 | } |
| 712 | |
| 713 | /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c |
| 714 | * eventually). The difference is that pulled data not copied, but |
| 715 | * immediately discarded. |
| 716 | */ |
Herbert Xu ~{PmVHI~} | f291196 | 2006-06-05 15:03:37 -0700 | [diff] [blame] | 717 | static void __pskb_trim_head(struct sk_buff *skb, int len) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 718 | { |
| 719 | int i, k, eat; |
| 720 | |
| 721 | eat = len; |
| 722 | k = 0; |
| 723 | for (i=0; i<skb_shinfo(skb)->nr_frags; i++) { |
| 724 | if (skb_shinfo(skb)->frags[i].size <= eat) { |
| 725 | put_page(skb_shinfo(skb)->frags[i].page); |
| 726 | eat -= skb_shinfo(skb)->frags[i].size; |
| 727 | } else { |
| 728 | skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i]; |
| 729 | if (eat) { |
| 730 | skb_shinfo(skb)->frags[k].page_offset += eat; |
| 731 | skb_shinfo(skb)->frags[k].size -= eat; |
| 732 | eat = 0; |
| 733 | } |
| 734 | k++; |
| 735 | } |
| 736 | } |
| 737 | skb_shinfo(skb)->nr_frags = k; |
| 738 | |
Arnaldo Carvalho de Melo | 27a884d | 2007-04-19 20:29:13 -0700 | [diff] [blame] | 739 | skb_reset_tail_pointer(skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 740 | skb->data_len -= len; |
| 741 | skb->len = skb->data_len; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 742 | } |
| 743 | |
| 744 | int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) |
| 745 | { |
| 746 | if (skb_cloned(skb) && |
| 747 | pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) |
| 748 | return -ENOMEM; |
| 749 | |
Herbert Xu ~{PmVHI~} | f291196 | 2006-06-05 15:03:37 -0700 | [diff] [blame] | 750 | /* If len == headlen, we avoid __skb_pull to preserve alignment. */ |
| 751 | if (unlikely(len < skb_headlen(skb))) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 752 | __skb_pull(skb, len); |
Herbert Xu ~{PmVHI~} | f291196 | 2006-06-05 15:03:37 -0700 | [diff] [blame] | 753 | else |
| 754 | __pskb_trim_head(skb, len - skb_headlen(skb)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 755 | |
| 756 | TCP_SKB_CB(skb)->seq += len; |
Patrick McHardy | 84fa793 | 2006-08-29 16:44:56 -0700 | [diff] [blame] | 757 | skb->ip_summed = CHECKSUM_PARTIAL; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 758 | |
| 759 | skb->truesize -= len; |
| 760 | sk->sk_wmem_queued -= len; |
| 761 | sk->sk_forward_alloc += len; |
| 762 | sock_set_flag(sk, SOCK_QUEUE_SHRUNK); |
| 763 | |
| 764 | /* Any change of skb->len requires recalculation of tso |
| 765 | * factor and mss. |
| 766 | */ |
| 767 | if (tcp_skb_pcount(skb) > 1) |
David S. Miller | 846998a | 2005-08-04 19:52:01 -0700 | [diff] [blame] | 768 | tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk, 1)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 769 | |
| 770 | return 0; |
| 771 | } |
| 772 | |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 773 | /* Not accounting for SACKs here. */ |
| 774 | int tcp_mtu_to_mss(struct sock *sk, int pmtu) |
| 775 | { |
| 776 | struct tcp_sock *tp = tcp_sk(sk); |
| 777 | struct inet_connection_sock *icsk = inet_csk(sk); |
| 778 | int mss_now; |
| 779 | |
| 780 | /* Calculate base mss without TCP options: |
| 781 | It is MMS_S - sizeof(tcphdr) of rfc1122 |
| 782 | */ |
| 783 | mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); |
| 784 | |
| 785 | /* Clamp it (mss_clamp does not include tcp options) */ |
| 786 | if (mss_now > tp->rx_opt.mss_clamp) |
| 787 | mss_now = tp->rx_opt.mss_clamp; |
| 788 | |
| 789 | /* Now subtract optional transport overhead */ |
| 790 | mss_now -= icsk->icsk_ext_hdr_len; |
| 791 | |
| 792 | /* Then reserve room for full set of TCP options and 8 bytes of data */ |
| 793 | if (mss_now < 48) |
| 794 | mss_now = 48; |
| 795 | |
| 796 | /* Now subtract TCP options size, not including SACKs */ |
| 797 | mss_now -= tp->tcp_header_len - sizeof(struct tcphdr); |
| 798 | |
| 799 | return mss_now; |
| 800 | } |
| 801 | |
| 802 | /* Inverse of above */ |
| 803 | int tcp_mss_to_mtu(struct sock *sk, int mss) |
| 804 | { |
| 805 | struct tcp_sock *tp = tcp_sk(sk); |
| 806 | struct inet_connection_sock *icsk = inet_csk(sk); |
| 807 | int mtu; |
| 808 | |
| 809 | mtu = mss + |
| 810 | tp->tcp_header_len + |
| 811 | icsk->icsk_ext_hdr_len + |
| 812 | icsk->icsk_af_ops->net_header_len; |
| 813 | |
| 814 | return mtu; |
| 815 | } |
| 816 | |
| 817 | void tcp_mtup_init(struct sock *sk) |
| 818 | { |
| 819 | struct tcp_sock *tp = tcp_sk(sk); |
| 820 | struct inet_connection_sock *icsk = inet_csk(sk); |
| 821 | |
| 822 | icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1; |
| 823 | 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] | 824 | icsk->icsk_af_ops->net_header_len; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 825 | icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss); |
| 826 | icsk->icsk_mtup.probe_size = 0; |
| 827 | } |
| 828 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 829 | /* This function synchronize snd mss to current pmtu/exthdr set. |
| 830 | |
| 831 | tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts |
| 832 | for TCP options, but includes only bare TCP header. |
| 833 | |
| 834 | tp->rx_opt.mss_clamp is mss negotiated at connection setup. |
Stephen Hemminger | caa20d9a | 2005-11-10 17:13:47 -0800 | [diff] [blame] | 835 | It is minimum of user_mss and mss received with SYN. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 836 | It also does not include TCP options. |
| 837 | |
Arnaldo Carvalho de Melo | d83d846 | 2005-12-13 23:26:10 -0800 | [diff] [blame] | 838 | 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] | 839 | |
| 840 | tp->mss_cache is current effective sending mss, including |
| 841 | all tcp options except for SACKs. It is evaluated, |
| 842 | taking into account current pmtu, but never exceeds |
| 843 | tp->rx_opt.mss_clamp. |
| 844 | |
| 845 | NOTE1. rfc1122 clearly states that advertised MSS |
| 846 | DOES NOT include either tcp or ip options. |
| 847 | |
Arnaldo Carvalho de Melo | d83d846 | 2005-12-13 23:26:10 -0800 | [diff] [blame] | 848 | NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache |
| 849 | are READ ONLY outside this function. --ANK (980731) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 850 | */ |
| 851 | |
| 852 | unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) |
| 853 | { |
| 854 | struct tcp_sock *tp = tcp_sk(sk); |
Arnaldo Carvalho de Melo | d83d846 | 2005-12-13 23:26:10 -0800 | [diff] [blame] | 855 | struct inet_connection_sock *icsk = inet_csk(sk); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 856 | int mss_now; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 857 | |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 858 | if (icsk->icsk_mtup.search_high > pmtu) |
| 859 | icsk->icsk_mtup.search_high = pmtu; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 860 | |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 861 | mss_now = tcp_mtu_to_mss(sk, pmtu); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 862 | |
| 863 | /* Bound mss with half of window */ |
| 864 | if (tp->max_window && mss_now > (tp->max_window>>1)) |
| 865 | mss_now = max((tp->max_window>>1), 68U - tp->tcp_header_len); |
| 866 | |
| 867 | /* And store cached results */ |
Arnaldo Carvalho de Melo | d83d846 | 2005-12-13 23:26:10 -0800 | [diff] [blame] | 868 | icsk->icsk_pmtu_cookie = pmtu; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 869 | if (icsk->icsk_mtup.enabled) |
| 870 | 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] | 871 | tp->mss_cache = mss_now; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 872 | |
| 873 | return mss_now; |
| 874 | } |
| 875 | |
| 876 | /* Compute the current effective MSS, taking SACKs and IP options, |
| 877 | * and even PMTU discovery events into account. |
| 878 | * |
| 879 | * LARGESEND note: !urg_mode is overkill, only frames up to snd_up |
| 880 | * cannot be large. However, taking into account rare use of URG, this |
| 881 | * is not a big flaw. |
| 882 | */ |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 883 | unsigned int tcp_current_mss(struct sock *sk, int large_allowed) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 884 | { |
| 885 | struct tcp_sock *tp = tcp_sk(sk); |
| 886 | struct dst_entry *dst = __sk_dst_get(sk); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 887 | u32 mss_now; |
| 888 | u16 xmit_size_goal; |
| 889 | int doing_tso = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 890 | |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 891 | mss_now = tp->mss_cache; |
| 892 | |
Herbert Xu | bcd7611 | 2006-06-30 13:36:35 -0700 | [diff] [blame] | 893 | if (large_allowed && sk_can_gso(sk) && !tp->urg_mode) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 894 | doing_tso = 1; |
| 895 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 896 | if (dst) { |
| 897 | u32 mtu = dst_mtu(dst); |
Arnaldo Carvalho de Melo | d83d846 | 2005-12-13 23:26:10 -0800 | [diff] [blame] | 898 | if (mtu != inet_csk(sk)->icsk_pmtu_cookie) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 899 | mss_now = tcp_sync_mss(sk, mtu); |
| 900 | } |
| 901 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 902 | if (tp->rx_opt.eff_sacks) |
| 903 | mss_now -= (TCPOLEN_SACK_BASE_ALIGNED + |
| 904 | (tp->rx_opt.eff_sacks * TCPOLEN_SACK_PERBLOCK)); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 905 | |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 906 | #ifdef CONFIG_TCP_MD5SIG |
| 907 | if (tp->af_specific->md5_lookup(sk, sk)) |
| 908 | mss_now -= TCPOLEN_MD5SIG_ALIGNED; |
| 909 | #endif |
| 910 | |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 911 | xmit_size_goal = mss_now; |
| 912 | |
| 913 | if (doing_tso) { |
Arnaldo Carvalho de Melo | 8292a17 | 2005-12-13 23:15:52 -0800 | [diff] [blame] | 914 | xmit_size_goal = (65535 - |
| 915 | inet_csk(sk)->icsk_af_ops->net_header_len - |
Arnaldo Carvalho de Melo | d83d846 | 2005-12-13 23:26:10 -0800 | [diff] [blame] | 916 | inet_csk(sk)->icsk_ext_hdr_len - |
| 917 | tp->tcp_header_len); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 918 | |
| 919 | if (tp->max_window && |
| 920 | (xmit_size_goal > (tp->max_window >> 1))) |
| 921 | xmit_size_goal = max((tp->max_window >> 1), |
| 922 | 68U - tp->tcp_header_len); |
| 923 | |
| 924 | xmit_size_goal -= (xmit_size_goal % mss_now); |
| 925 | } |
| 926 | tp->xmit_size_goal = xmit_size_goal; |
| 927 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 928 | return mss_now; |
| 929 | } |
| 930 | |
David S. Miller | a762a98 | 2005-07-05 15:18:51 -0700 | [diff] [blame] | 931 | /* Congestion window validation. (RFC2861) */ |
| 932 | |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 933 | static void tcp_cwnd_validate(struct sock *sk) |
David S. Miller | a762a98 | 2005-07-05 15:18:51 -0700 | [diff] [blame] | 934 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 935 | struct tcp_sock *tp = tcp_sk(sk); |
David S. Miller | a762a98 | 2005-07-05 15:18:51 -0700 | [diff] [blame] | 936 | __u32 packets_out = tp->packets_out; |
| 937 | |
| 938 | if (packets_out >= tp->snd_cwnd) { |
| 939 | /* Network is feed fully. */ |
| 940 | tp->snd_cwnd_used = 0; |
| 941 | tp->snd_cwnd_stamp = tcp_time_stamp; |
| 942 | } else { |
| 943 | /* Network starves. */ |
| 944 | if (tp->packets_out > tp->snd_cwnd_used) |
| 945 | tp->snd_cwnd_used = tp->packets_out; |
| 946 | |
David S. Miller | 15d33c0 | 2007-04-09 13:23:14 -0700 | [diff] [blame] | 947 | if (sysctl_tcp_slow_start_after_idle && |
| 948 | (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] | 949 | tcp_cwnd_application_limited(sk); |
| 950 | } |
| 951 | } |
| 952 | |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 953 | static unsigned int tcp_window_allows(struct tcp_sock *tp, struct sk_buff *skb, unsigned int mss_now, unsigned int cwnd) |
| 954 | { |
| 955 | u32 window, cwnd_len; |
| 956 | |
| 957 | window = (tp->snd_una + tp->snd_wnd - TCP_SKB_CB(skb)->seq); |
| 958 | cwnd_len = mss_now * cwnd; |
| 959 | return min(window, cwnd_len); |
| 960 | } |
| 961 | |
| 962 | /* Can at least one segment of SKB be sent right now, according to the |
| 963 | * congestion window rules? If so, return how many segments are allowed. |
| 964 | */ |
| 965 | static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp, struct sk_buff *skb) |
| 966 | { |
| 967 | u32 in_flight, cwnd; |
| 968 | |
| 969 | /* Don't be strict about the congestion window for the final FIN. */ |
John Heffner | 104439a | 2007-02-05 17:53:11 -0800 | [diff] [blame] | 970 | if ((TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) && |
| 971 | tcp_skb_pcount(skb) == 1) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 972 | return 1; |
| 973 | |
| 974 | in_flight = tcp_packets_in_flight(tp); |
| 975 | cwnd = tp->snd_cwnd; |
| 976 | if (in_flight < cwnd) |
| 977 | return (cwnd - in_flight); |
| 978 | |
| 979 | return 0; |
| 980 | } |
| 981 | |
| 982 | /* This must be invoked the first time we consider transmitting |
| 983 | * SKB onto the wire. |
| 984 | */ |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 985 | static int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb, unsigned int mss_now) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 986 | { |
| 987 | int tso_segs = tcp_skb_pcount(skb); |
| 988 | |
David S. Miller | 846998a | 2005-08-04 19:52:01 -0700 | [diff] [blame] | 989 | if (!tso_segs || |
| 990 | (tso_segs > 1 && |
Herbert Xu | 7967168 | 2006-06-22 02:40:14 -0700 | [diff] [blame] | 991 | tcp_skb_mss(skb) != mss_now)) { |
David S. Miller | 846998a | 2005-08-04 19:52:01 -0700 | [diff] [blame] | 992 | tcp_set_skb_tso_segs(sk, skb, mss_now); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 993 | tso_segs = tcp_skb_pcount(skb); |
| 994 | } |
| 995 | return tso_segs; |
| 996 | } |
| 997 | |
| 998 | static inline int tcp_minshall_check(const struct tcp_sock *tp) |
| 999 | { |
| 1000 | return after(tp->snd_sml,tp->snd_una) && |
| 1001 | !after(tp->snd_sml, tp->snd_nxt); |
| 1002 | } |
| 1003 | |
| 1004 | /* Return 0, if packet can be sent now without violation Nagle's rules: |
| 1005 | * 1. It is full sized. |
| 1006 | * 2. Or it contains FIN. (already checked by caller) |
| 1007 | * 3. Or TCP_NODELAY was set. |
| 1008 | * 4. Or TCP_CORK is not set, and all sent packets are ACKed. |
| 1009 | * With Minshall's modification: all sent small packets are ACKed. |
| 1010 | */ |
| 1011 | |
| 1012 | static inline int tcp_nagle_check(const struct tcp_sock *tp, |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 1013 | const struct sk_buff *skb, |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1014 | unsigned mss_now, int nonagle) |
| 1015 | { |
| 1016 | return (skb->len < mss_now && |
| 1017 | ((nonagle&TCP_NAGLE_CORK) || |
| 1018 | (!nonagle && |
| 1019 | tp->packets_out && |
| 1020 | tcp_minshall_check(tp)))); |
| 1021 | } |
| 1022 | |
| 1023 | /* Return non-zero if the Nagle test allows this packet to be |
| 1024 | * sent now. |
| 1025 | */ |
| 1026 | static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb, |
| 1027 | unsigned int cur_mss, int nonagle) |
| 1028 | { |
| 1029 | /* Nagle rule does not apply to frames, which sit in the middle of the |
| 1030 | * write_queue (they have no chances to get new data). |
| 1031 | * |
| 1032 | * This is implemented in the callers, where they modify the 'nonagle' |
| 1033 | * argument based upon the location of SKB in the send queue. |
| 1034 | */ |
| 1035 | if (nonagle & TCP_NAGLE_PUSH) |
| 1036 | return 1; |
| 1037 | |
Ilpo Järvinen | d551e45 | 2007-04-30 00:42:20 -0700 | [diff] [blame] | 1038 | /* Don't use the nagle rule for urgent data (or for the final FIN). |
| 1039 | * Nagle can be ignored during F-RTO too (see RFC4138). |
| 1040 | */ |
| 1041 | if (tp->urg_mode || (tp->frto_counter == 2) || |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1042 | (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)) |
| 1043 | return 1; |
| 1044 | |
| 1045 | if (!tcp_nagle_check(tp, skb, cur_mss, nonagle)) |
| 1046 | return 1; |
| 1047 | |
| 1048 | return 0; |
| 1049 | } |
| 1050 | |
| 1051 | /* Does at least the first segment of SKB fit into the send window? */ |
| 1052 | static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb, unsigned int cur_mss) |
| 1053 | { |
| 1054 | u32 end_seq = TCP_SKB_CB(skb)->end_seq; |
| 1055 | |
| 1056 | if (skb->len > cur_mss) |
| 1057 | end_seq = TCP_SKB_CB(skb)->seq + cur_mss; |
| 1058 | |
| 1059 | return !after(end_seq, tp->snd_una + tp->snd_wnd); |
| 1060 | } |
| 1061 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1062 | /* 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] | 1063 | * should be put on the wire right now. If so, it returns the number of |
| 1064 | * packets allowed by the congestion window. |
| 1065 | */ |
| 1066 | static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb, |
| 1067 | unsigned int cur_mss, int nonagle) |
| 1068 | { |
| 1069 | struct tcp_sock *tp = tcp_sk(sk); |
| 1070 | unsigned int cwnd_quota; |
| 1071 | |
David S. Miller | 846998a | 2005-08-04 19:52:01 -0700 | [diff] [blame] | 1072 | tcp_init_tso_segs(sk, skb, cur_mss); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1073 | |
| 1074 | if (!tcp_nagle_test(tp, skb, cur_mss, nonagle)) |
| 1075 | return 0; |
| 1076 | |
| 1077 | cwnd_quota = tcp_cwnd_test(tp, skb); |
| 1078 | if (cwnd_quota && |
| 1079 | !tcp_snd_wnd_test(tp, skb, cur_mss)) |
| 1080 | cwnd_quota = 0; |
| 1081 | |
| 1082 | return cwnd_quota; |
| 1083 | } |
| 1084 | |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 1085 | int tcp_may_send_now(struct sock *sk) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1086 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 1087 | struct tcp_sock *tp = tcp_sk(sk); |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1088 | struct sk_buff *skb = tcp_send_head(sk); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1089 | |
| 1090 | return (skb && |
| 1091 | tcp_snd_test(sk, skb, tcp_current_mss(sk, 1), |
| 1092 | (tcp_skb_is_last(sk, skb) ? |
| 1093 | TCP_NAGLE_PUSH : |
| 1094 | tp->nonagle))); |
| 1095 | } |
| 1096 | |
| 1097 | /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet |
| 1098 | * which is put after SKB on the list. It is very much like |
| 1099 | * tcp_fragment() except that it may make several kinds of assumptions |
| 1100 | * in order to speed up the splitting operation. In particular, we |
| 1101 | * know that all the data is in scatter-gather pages, and that the |
| 1102 | * packet has never been sent out before (and thus is not cloned). |
| 1103 | */ |
David S. Miller | 846998a | 2005-08-04 19:52:01 -0700 | [diff] [blame] | 1104 | static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, unsigned int mss_now) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1105 | { |
| 1106 | struct sk_buff *buff; |
| 1107 | int nlen = skb->len - len; |
| 1108 | u16 flags; |
| 1109 | |
| 1110 | /* All of a TSO frame must be composed of paged data. */ |
Herbert Xu | c8ac377 | 2005-08-16 20:43:40 -0700 | [diff] [blame] | 1111 | if (skb->len != skb->data_len) |
| 1112 | return tcp_fragment(sk, skb, len, mss_now); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1113 | |
| 1114 | buff = sk_stream_alloc_pskb(sk, 0, 0, GFP_ATOMIC); |
| 1115 | if (unlikely(buff == NULL)) |
| 1116 | return -ENOMEM; |
| 1117 | |
Herbert Xu | b60b49e | 2006-04-19 21:35:00 -0700 | [diff] [blame] | 1118 | sk_charge_skb(sk, buff); |
| 1119 | buff->truesize += nlen; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1120 | skb->truesize -= nlen; |
| 1121 | |
| 1122 | /* Correct the sequence numbers. */ |
| 1123 | TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; |
| 1124 | TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; |
| 1125 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; |
| 1126 | |
| 1127 | /* PSH and FIN should only be set in the second packet. */ |
| 1128 | flags = TCP_SKB_CB(skb)->flags; |
| 1129 | TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH); |
| 1130 | TCP_SKB_CB(buff)->flags = flags; |
| 1131 | |
| 1132 | /* This packet was never sent out yet, so no SACK bits. */ |
| 1133 | TCP_SKB_CB(buff)->sacked = 0; |
| 1134 | |
Patrick McHardy | 84fa793 | 2006-08-29 16:44:56 -0700 | [diff] [blame] | 1135 | buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1136 | skb_split(skb, buff, len); |
| 1137 | |
| 1138 | /* Fix up tso_factor for both original and new SKB. */ |
David S. Miller | 846998a | 2005-08-04 19:52:01 -0700 | [diff] [blame] | 1139 | tcp_set_skb_tso_segs(sk, skb, mss_now); |
| 1140 | tcp_set_skb_tso_segs(sk, buff, mss_now); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1141 | |
| 1142 | /* Link BUFF into the send queue. */ |
| 1143 | skb_header_release(buff); |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1144 | tcp_insert_write_queue_after(skb, buff, sk); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1145 | |
| 1146 | return 0; |
| 1147 | } |
| 1148 | |
| 1149 | /* Try to defer sending, if possible, in order to minimize the amount |
| 1150 | * of TSO splitting we do. View it as a kind of TSO Nagle test. |
| 1151 | * |
| 1152 | * This algorithm is from John Heffner. |
| 1153 | */ |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 1154 | 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] | 1155 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 1156 | struct tcp_sock *tp = tcp_sk(sk); |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 1157 | const struct inet_connection_sock *icsk = inet_csk(sk); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1158 | u32 send_win, cong_win, limit, in_flight; |
| 1159 | |
| 1160 | if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) |
John Heffner | ae8064a | 2006-10-18 20:36:48 -0700 | [diff] [blame] | 1161 | goto send_now; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1162 | |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 1163 | if (icsk->icsk_ca_state != TCP_CA_Open) |
John Heffner | ae8064a | 2006-10-18 20:36:48 -0700 | [diff] [blame] | 1164 | goto send_now; |
| 1165 | |
| 1166 | /* Defer for less than two clock ticks. */ |
| 1167 | if (!tp->tso_deferred && ((jiffies<<1)>>1) - (tp->tso_deferred>>1) > 1) |
| 1168 | goto send_now; |
David S. Miller | 908a75c | 2005-07-05 15:43:58 -0700 | [diff] [blame] | 1169 | |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1170 | in_flight = tcp_packets_in_flight(tp); |
| 1171 | |
| 1172 | BUG_ON(tcp_skb_pcount(skb) <= 1 || |
| 1173 | (tp->snd_cwnd <= in_flight)); |
| 1174 | |
| 1175 | send_win = (tp->snd_una + tp->snd_wnd) - TCP_SKB_CB(skb)->seq; |
| 1176 | |
| 1177 | /* From in_flight test above, we know that cwnd > in_flight. */ |
| 1178 | cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; |
| 1179 | |
| 1180 | limit = min(send_win, cong_win); |
| 1181 | |
David S. Miller | ba244fe | 2006-03-11 18:51:49 -0800 | [diff] [blame] | 1182 | /* If a full-sized TSO skb can be sent, do it. */ |
| 1183 | if (limit >= 65536) |
John Heffner | ae8064a | 2006-10-18 20:36:48 -0700 | [diff] [blame] | 1184 | goto send_now; |
David S. Miller | ba244fe | 2006-03-11 18:51:49 -0800 | [diff] [blame] | 1185 | |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1186 | if (sysctl_tcp_tso_win_divisor) { |
| 1187 | u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); |
| 1188 | |
| 1189 | /* If at least some fraction of a window is available, |
| 1190 | * just use it. |
| 1191 | */ |
| 1192 | chunk /= sysctl_tcp_tso_win_divisor; |
| 1193 | if (limit >= chunk) |
John Heffner | ae8064a | 2006-10-18 20:36:48 -0700 | [diff] [blame] | 1194 | goto send_now; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1195 | } else { |
| 1196 | /* Different approach, try not to defer past a single |
| 1197 | * ACK. Receiver should ACK every other full sized |
| 1198 | * frame, so if we have space for more than 3 frames |
| 1199 | * then send now. |
| 1200 | */ |
| 1201 | if (limit > tcp_max_burst(tp) * tp->mss_cache) |
John Heffner | ae8064a | 2006-10-18 20:36:48 -0700 | [diff] [blame] | 1202 | goto send_now; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1203 | } |
| 1204 | |
| 1205 | /* Ok, it looks like it is advisable to defer. */ |
John Heffner | ae8064a | 2006-10-18 20:36:48 -0700 | [diff] [blame] | 1206 | tp->tso_deferred = 1 | (jiffies<<1); |
| 1207 | |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1208 | return 1; |
John Heffner | ae8064a | 2006-10-18 20:36:48 -0700 | [diff] [blame] | 1209 | |
| 1210 | send_now: |
| 1211 | tp->tso_deferred = 0; |
| 1212 | return 0; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1213 | } |
| 1214 | |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1215 | /* Create a new MTU probe if we are ready. |
| 1216 | * Returns 0 if we should wait to probe (no cwnd available), |
| 1217 | * 1 if a probe was sent, |
| 1218 | * -1 otherwise */ |
| 1219 | static int tcp_mtu_probe(struct sock *sk) |
| 1220 | { |
| 1221 | struct tcp_sock *tp = tcp_sk(sk); |
| 1222 | struct inet_connection_sock *icsk = inet_csk(sk); |
| 1223 | struct sk_buff *skb, *nskb, *next; |
| 1224 | int len; |
| 1225 | int probe_size; |
| 1226 | unsigned int pif; |
| 1227 | int copy; |
| 1228 | int mss_now; |
| 1229 | |
| 1230 | /* Not currently probing/verifying, |
| 1231 | * not in recovery, |
| 1232 | * have enough cwnd, and |
| 1233 | * not SACKing (the variable headers throw things off) */ |
| 1234 | if (!icsk->icsk_mtup.enabled || |
| 1235 | icsk->icsk_mtup.probe_size || |
| 1236 | inet_csk(sk)->icsk_ca_state != TCP_CA_Open || |
| 1237 | tp->snd_cwnd < 11 || |
| 1238 | tp->rx_opt.eff_sacks) |
| 1239 | return -1; |
| 1240 | |
| 1241 | /* Very simple search strategy: just double the MSS. */ |
| 1242 | mss_now = tcp_current_mss(sk, 0); |
| 1243 | probe_size = 2*tp->mss_cache; |
| 1244 | if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) { |
| 1245 | /* TODO: set timer for probe_converge_event */ |
| 1246 | return -1; |
| 1247 | } |
| 1248 | |
| 1249 | /* Have enough data in the send queue to probe? */ |
| 1250 | len = 0; |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1251 | if ((skb = tcp_send_head(sk)) == NULL) |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1252 | return -1; |
| 1253 | while ((len += skb->len) < probe_size && !tcp_skb_is_last(sk, skb)) |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1254 | skb = tcp_write_queue_next(sk, skb); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1255 | if (len < probe_size) |
| 1256 | return -1; |
| 1257 | |
| 1258 | /* Receive window check. */ |
| 1259 | if (after(TCP_SKB_CB(skb)->seq + probe_size, tp->snd_una + tp->snd_wnd)) { |
| 1260 | if (tp->snd_wnd < probe_size) |
| 1261 | return -1; |
| 1262 | else |
| 1263 | return 0; |
| 1264 | } |
| 1265 | |
| 1266 | /* Do we need to wait to drain cwnd? */ |
| 1267 | pif = tcp_packets_in_flight(tp); |
| 1268 | if (pif + 2 > tp->snd_cwnd) { |
| 1269 | /* With no packets in flight, don't stall. */ |
| 1270 | if (pif == 0) |
| 1271 | return -1; |
| 1272 | else |
| 1273 | return 0; |
| 1274 | } |
| 1275 | |
| 1276 | /* We're allowed to probe. Build it now. */ |
| 1277 | if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL) |
| 1278 | return -1; |
| 1279 | sk_charge_skb(sk, nskb); |
| 1280 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1281 | skb = tcp_send_head(sk); |
| 1282 | tcp_insert_write_queue_before(nskb, skb, sk); |
| 1283 | tcp_advance_send_head(sk, skb); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1284 | |
| 1285 | TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; |
| 1286 | TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; |
| 1287 | TCP_SKB_CB(nskb)->flags = TCPCB_FLAG_ACK; |
| 1288 | TCP_SKB_CB(nskb)->sacked = 0; |
| 1289 | nskb->csum = 0; |
Patrick McHardy | 84fa793 | 2006-08-29 16:44:56 -0700 | [diff] [blame] | 1290 | nskb->ip_summed = skb->ip_summed; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1291 | |
| 1292 | len = 0; |
| 1293 | while (len < probe_size) { |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1294 | next = tcp_write_queue_next(sk, skb); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1295 | |
| 1296 | copy = min_t(int, skb->len, probe_size - len); |
| 1297 | if (nskb->ip_summed) |
| 1298 | skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); |
| 1299 | else |
| 1300 | nskb->csum = skb_copy_and_csum_bits(skb, 0, |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 1301 | skb_put(nskb, copy), copy, nskb->csum); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1302 | |
| 1303 | if (skb->len <= copy) { |
| 1304 | /* We've eaten all the data from this skb. |
| 1305 | * Throw it away. */ |
| 1306 | TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags; |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1307 | tcp_unlink_write_queue(skb, sk); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1308 | sk_stream_free_skb(sk, skb); |
| 1309 | } else { |
| 1310 | TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags & |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 1311 | ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1312 | if (!skb_shinfo(skb)->nr_frags) { |
| 1313 | skb_pull(skb, copy); |
Patrick McHardy | 84fa793 | 2006-08-29 16:44:56 -0700 | [diff] [blame] | 1314 | if (skb->ip_summed != CHECKSUM_PARTIAL) |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1315 | skb->csum = csum_partial(skb->data, skb->len, 0); |
| 1316 | } else { |
| 1317 | __pskb_trim_head(skb, copy); |
| 1318 | tcp_set_skb_tso_segs(sk, skb, mss_now); |
| 1319 | } |
| 1320 | TCP_SKB_CB(skb)->seq += copy; |
| 1321 | } |
| 1322 | |
| 1323 | len += copy; |
| 1324 | skb = next; |
| 1325 | } |
| 1326 | tcp_init_tso_segs(sk, nskb, nskb->len); |
| 1327 | |
| 1328 | /* We're ready to send. If this fails, the probe will |
| 1329 | * be resegmented into mss-sized pieces by tcp_write_xmit(). */ |
| 1330 | TCP_SKB_CB(nskb)->when = tcp_time_stamp; |
| 1331 | if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { |
| 1332 | /* Decrement cwnd here because we are sending |
| 1333 | * effectively two packets. */ |
| 1334 | tp->snd_cwnd--; |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 1335 | update_send_head(sk, nskb); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1336 | |
| 1337 | icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); |
John Heffner | 0e7b136 | 2006-03-20 21:32:58 -0800 | [diff] [blame] | 1338 | tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; |
| 1339 | tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1340 | |
| 1341 | return 1; |
| 1342 | } |
| 1343 | |
| 1344 | return -1; |
| 1345 | } |
| 1346 | |
| 1347 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1348 | /* This routine writes packets to the network. It advances the |
| 1349 | * send_head. This happens as incoming acks open up the remote |
| 1350 | * window for us. |
| 1351 | * |
| 1352 | * Returns 1, if no segments are in flight and we have queued segments, but |
| 1353 | * cannot send anything now because of SWS or another problem. |
| 1354 | */ |
David S. Miller | a2e2a59 | 2005-07-05 15:19:23 -0700 | [diff] [blame] | 1355 | 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] | 1356 | { |
| 1357 | struct tcp_sock *tp = tcp_sk(sk); |
David S. Miller | 92df7b5 | 2005-07-05 15:19:06 -0700 | [diff] [blame] | 1358 | struct sk_buff *skb; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1359 | unsigned int tso_segs, sent_pkts; |
| 1360 | int cwnd_quota; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1361 | int result; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1362 | |
| 1363 | /* If we are closed, the bytes will have to remain here. |
| 1364 | * In time closedown will finish, we empty the write queue and all |
| 1365 | * will be happy. |
| 1366 | */ |
David S. Miller | 92df7b5 | 2005-07-05 15:19:06 -0700 | [diff] [blame] | 1367 | if (unlikely(sk->sk_state == TCP_CLOSE)) |
| 1368 | return 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1369 | |
David S. Miller | 92df7b5 | 2005-07-05 15:19:06 -0700 | [diff] [blame] | 1370 | sent_pkts = 0; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1371 | |
| 1372 | /* Do MTU probing. */ |
| 1373 | if ((result = tcp_mtu_probe(sk)) == 0) { |
| 1374 | return 0; |
| 1375 | } else if (result > 0) { |
| 1376 | sent_pkts = 1; |
| 1377 | } |
| 1378 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1379 | while ((skb = tcp_send_head(sk))) { |
Herbert Xu | c8ac377 | 2005-08-16 20:43:40 -0700 | [diff] [blame] | 1380 | unsigned int limit; |
| 1381 | |
Herbert Xu | b68e9f8 | 2005-08-04 19:52:02 -0700 | [diff] [blame] | 1382 | tso_segs = tcp_init_tso_segs(sk, skb, mss_now); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1383 | BUG_ON(!tso_segs); |
David S. Miller | aa93466 | 2005-07-05 15:20:09 -0700 | [diff] [blame] | 1384 | |
Herbert Xu | b68e9f8 | 2005-08-04 19:52:02 -0700 | [diff] [blame] | 1385 | cwnd_quota = tcp_cwnd_test(tp, skb); |
| 1386 | if (!cwnd_quota) |
| 1387 | break; |
| 1388 | |
| 1389 | if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) |
| 1390 | break; |
| 1391 | |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1392 | if (tso_segs == 1) { |
| 1393 | if (unlikely(!tcp_nagle_test(tp, skb, mss_now, |
| 1394 | (tcp_skb_is_last(sk, skb) ? |
| 1395 | nonagle : TCP_NAGLE_PUSH)))) |
| 1396 | break; |
| 1397 | } else { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 1398 | if (tcp_tso_should_defer(sk, skb)) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1399 | break; |
| 1400 | } |
David S. Miller | aa93466 | 2005-07-05 15:20:09 -0700 | [diff] [blame] | 1401 | |
Herbert Xu | c8ac377 | 2005-08-16 20:43:40 -0700 | [diff] [blame] | 1402 | limit = mss_now; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1403 | if (tso_segs > 1) { |
Herbert Xu | c8ac377 | 2005-08-16 20:43:40 -0700 | [diff] [blame] | 1404 | limit = tcp_window_allows(tp, skb, |
| 1405 | mss_now, cwnd_quota); |
David S. Miller | aa93466 | 2005-07-05 15:20:09 -0700 | [diff] [blame] | 1406 | |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1407 | if (skb->len < limit) { |
| 1408 | unsigned int trim = skb->len % mss_now; |
| 1409 | |
| 1410 | if (trim) |
| 1411 | limit = skb->len - trim; |
| 1412 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1413 | } |
| 1414 | |
Herbert Xu | c8ac377 | 2005-08-16 20:43:40 -0700 | [diff] [blame] | 1415 | if (skb->len > limit && |
| 1416 | unlikely(tso_fragment(sk, skb, limit, mss_now))) |
| 1417 | break; |
| 1418 | |
David S. Miller | 92df7b5 | 2005-07-05 15:19:06 -0700 | [diff] [blame] | 1419 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1420 | |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 1421 | if (unlikely(tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC))) |
David S. Miller | 92df7b5 | 2005-07-05 15:19:06 -0700 | [diff] [blame] | 1422 | break; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1423 | |
David S. Miller | 92df7b5 | 2005-07-05 15:19:06 -0700 | [diff] [blame] | 1424 | /* Advance the send_head. This one is sent out. |
| 1425 | * This call will increment packets_out. |
| 1426 | */ |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 1427 | update_send_head(sk, skb); |
David S. Miller | 92df7b5 | 2005-07-05 15:19:06 -0700 | [diff] [blame] | 1428 | |
| 1429 | tcp_minshall_update(tp, mss_now, skb); |
David S. Miller | aa93466 | 2005-07-05 15:20:09 -0700 | [diff] [blame] | 1430 | sent_pkts++; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1431 | } |
David S. Miller | 92df7b5 | 2005-07-05 15:19:06 -0700 | [diff] [blame] | 1432 | |
David S. Miller | aa93466 | 2005-07-05 15:20:09 -0700 | [diff] [blame] | 1433 | if (likely(sent_pkts)) { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 1434 | tcp_cwnd_validate(sk); |
David S. Miller | 92df7b5 | 2005-07-05 15:19:06 -0700 | [diff] [blame] | 1435 | return 0; |
| 1436 | } |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1437 | return !tp->packets_out && tcp_send_head(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1438 | } |
| 1439 | |
David S. Miller | a762a98 | 2005-07-05 15:18:51 -0700 | [diff] [blame] | 1440 | /* Push out any pending frames which were held back due to |
| 1441 | * TCP_CORK or attempt at coalescing tiny packets. |
| 1442 | * The socket must be locked by the caller. |
| 1443 | */ |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 1444 | void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, |
| 1445 | int nonagle) |
David S. Miller | a762a98 | 2005-07-05 15:18:51 -0700 | [diff] [blame] | 1446 | { |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1447 | struct sk_buff *skb = tcp_send_head(sk); |
David S. Miller | a762a98 | 2005-07-05 15:18:51 -0700 | [diff] [blame] | 1448 | |
| 1449 | if (skb) { |
David S. Miller | 55c97f3 | 2005-07-05 15:19:38 -0700 | [diff] [blame] | 1450 | if (tcp_write_xmit(sk, cur_mss, nonagle)) |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 1451 | tcp_check_probe_timer(sk); |
David S. Miller | a762a98 | 2005-07-05 15:18:51 -0700 | [diff] [blame] | 1452 | } |
| 1453 | } |
| 1454 | |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1455 | /* Send _single_ skb sitting at the send head. This function requires |
| 1456 | * true push pending frames to setup probe timer etc. |
| 1457 | */ |
| 1458 | void tcp_push_one(struct sock *sk, unsigned int mss_now) |
| 1459 | { |
| 1460 | struct tcp_sock *tp = tcp_sk(sk); |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1461 | struct sk_buff *skb = tcp_send_head(sk); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1462 | unsigned int tso_segs, cwnd_quota; |
| 1463 | |
| 1464 | BUG_ON(!skb || skb->len < mss_now); |
| 1465 | |
David S. Miller | 846998a | 2005-08-04 19:52:01 -0700 | [diff] [blame] | 1466 | tso_segs = tcp_init_tso_segs(sk, skb, mss_now); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1467 | cwnd_quota = tcp_snd_test(sk, skb, mss_now, TCP_NAGLE_PUSH); |
| 1468 | |
| 1469 | if (likely(cwnd_quota)) { |
Herbert Xu | c8ac377 | 2005-08-16 20:43:40 -0700 | [diff] [blame] | 1470 | unsigned int limit; |
| 1471 | |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1472 | BUG_ON(!tso_segs); |
| 1473 | |
Herbert Xu | c8ac377 | 2005-08-16 20:43:40 -0700 | [diff] [blame] | 1474 | limit = mss_now; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1475 | if (tso_segs > 1) { |
Herbert Xu | c8ac377 | 2005-08-16 20:43:40 -0700 | [diff] [blame] | 1476 | limit = tcp_window_allows(tp, skb, |
| 1477 | mss_now, cwnd_quota); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1478 | |
| 1479 | if (skb->len < limit) { |
| 1480 | unsigned int trim = skb->len % mss_now; |
| 1481 | |
| 1482 | if (trim) |
| 1483 | limit = skb->len - trim; |
| 1484 | } |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1485 | } |
| 1486 | |
Herbert Xu | c8ac377 | 2005-08-16 20:43:40 -0700 | [diff] [blame] | 1487 | if (skb->len > limit && |
| 1488 | unlikely(tso_fragment(sk, skb, limit, mss_now))) |
| 1489 | return; |
| 1490 | |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1491 | /* Send it out now. */ |
| 1492 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
| 1493 | |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 1494 | if (likely(!tcp_transmit_skb(sk, skb, 1, sk->sk_allocation))) { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 1495 | update_send_head(sk, skb); |
| 1496 | tcp_cwnd_validate(sk); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1497 | return; |
| 1498 | } |
| 1499 | } |
| 1500 | } |
| 1501 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1502 | /* This function returns the amount that we can raise the |
| 1503 | * usable window based on the following constraints |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 1504 | * |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1505 | * 1. The window can never be shrunk once it is offered (RFC 793) |
| 1506 | * 2. We limit memory per socket |
| 1507 | * |
| 1508 | * RFC 1122: |
| 1509 | * "the suggested [SWS] avoidance algorithm for the receiver is to keep |
| 1510 | * RECV.NEXT + RCV.WIN fixed until: |
| 1511 | * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" |
| 1512 | * |
| 1513 | * i.e. don't raise the right edge of the window until you can raise |
| 1514 | * it at least MSS bytes. |
| 1515 | * |
| 1516 | * Unfortunately, the recommended algorithm breaks header prediction, |
| 1517 | * since header prediction assumes th->window stays fixed. |
| 1518 | * |
| 1519 | * Strictly speaking, keeping th->window fixed violates the receiver |
| 1520 | * side SWS prevention criteria. The problem is that under this rule |
| 1521 | * a stream of single byte packets will cause the right side of the |
| 1522 | * window to always advance by a single byte. |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 1523 | * |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1524 | * Of course, if the sender implements sender side SWS prevention |
| 1525 | * then this will not be a problem. |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 1526 | * |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1527 | * BSD seems to make the following compromise: |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 1528 | * |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1529 | * If the free space is less than the 1/4 of the maximum |
| 1530 | * space available and the free space is less than 1/2 mss, |
| 1531 | * then set the window to 0. |
| 1532 | * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] |
| 1533 | * Otherwise, just prevent the window from shrinking |
| 1534 | * and from being larger than the largest representable value. |
| 1535 | * |
| 1536 | * This prevents incremental opening of the window in the regime |
| 1537 | * where TCP is limited by the speed of the reader side taking |
| 1538 | * data out of the TCP receive queue. It does nothing about |
| 1539 | * those cases where the window is constrained on the sender side |
| 1540 | * because the pipeline is full. |
| 1541 | * |
| 1542 | * BSD also seems to "accidentally" limit itself to windows that are a |
| 1543 | * multiple of MSS, at least until the free space gets quite small. |
| 1544 | * This would appear to be a side effect of the mbuf implementation. |
| 1545 | * Combining these two algorithms results in the observed behavior |
| 1546 | * of having a fixed window size at almost all times. |
| 1547 | * |
| 1548 | * Below we obtain similar behavior by forcing the offered window to |
| 1549 | * a multiple of the mss when it is feasible to do so. |
| 1550 | * |
| 1551 | * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. |
| 1552 | * Regular options like TIMESTAMP are taken into account. |
| 1553 | */ |
| 1554 | u32 __tcp_select_window(struct sock *sk) |
| 1555 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 1556 | struct inet_connection_sock *icsk = inet_csk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1557 | struct tcp_sock *tp = tcp_sk(sk); |
Stephen Hemminger | caa20d9a | 2005-11-10 17:13:47 -0800 | [diff] [blame] | 1558 | /* MSS for the peer's data. Previous versions used mss_clamp |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1559 | * here. I don't know if the value based on our guesses |
| 1560 | * of peer's MSS is better for the performance. It's more correct |
| 1561 | * but may be worse for the performance because of rcv_mss |
| 1562 | * fluctuations. --SAW 1998/11/1 |
| 1563 | */ |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 1564 | int mss = icsk->icsk_ack.rcv_mss; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1565 | int free_space = tcp_space(sk); |
| 1566 | int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk)); |
| 1567 | int window; |
| 1568 | |
| 1569 | if (mss > full_space) |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 1570 | mss = full_space; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1571 | |
| 1572 | if (free_space < full_space/2) { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 1573 | icsk->icsk_ack.quick = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1574 | |
| 1575 | if (tcp_memory_pressure) |
| 1576 | tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U*tp->advmss); |
| 1577 | |
| 1578 | if (free_space < mss) |
| 1579 | return 0; |
| 1580 | } |
| 1581 | |
| 1582 | if (free_space > tp->rcv_ssthresh) |
| 1583 | free_space = tp->rcv_ssthresh; |
| 1584 | |
| 1585 | /* Don't do rounding if we are using window scaling, since the |
| 1586 | * scaled window will not line up with the MSS boundary anyway. |
| 1587 | */ |
| 1588 | window = tp->rcv_wnd; |
| 1589 | if (tp->rx_opt.rcv_wscale) { |
| 1590 | window = free_space; |
| 1591 | |
| 1592 | /* Advertise enough space so that it won't get scaled away. |
| 1593 | * Import case: prevent zero window announcement if |
| 1594 | * 1<<rcv_wscale > mss. |
| 1595 | */ |
| 1596 | if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window) |
| 1597 | window = (((window >> tp->rx_opt.rcv_wscale) + 1) |
| 1598 | << tp->rx_opt.rcv_wscale); |
| 1599 | } else { |
| 1600 | /* Get the largest window that is a nice multiple of mss. |
| 1601 | * Window clamp already applied above. |
| 1602 | * If our current window offering is within 1 mss of the |
| 1603 | * free space we just keep it. This prevents the divide |
| 1604 | * and multiply from happening most of the time. |
| 1605 | * We also don't do any window rounding when the free space |
| 1606 | * is too small. |
| 1607 | */ |
| 1608 | if (window <= free_space - mss || window > free_space) |
| 1609 | window = (free_space/mss)*mss; |
John Heffner | 8456507 | 2007-04-02 13:56:32 -0700 | [diff] [blame] | 1610 | else if (mss == full_space && |
| 1611 | free_space > window + full_space/2) |
| 1612 | window = free_space; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1613 | } |
| 1614 | |
| 1615 | return window; |
| 1616 | } |
| 1617 | |
| 1618 | /* Attempt to collapse two adjacent SKB's during retransmission. */ |
| 1619 | static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *skb, int mss_now) |
| 1620 | { |
| 1621 | struct tcp_sock *tp = tcp_sk(sk); |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1622 | struct sk_buff *next_skb = tcp_write_queue_next(sk, skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1623 | |
| 1624 | /* The first test we must make is that neither of these two |
| 1625 | * SKB's are still referenced by someone else. |
| 1626 | */ |
| 1627 | if (!skb_cloned(skb) && !skb_cloned(next_skb)) { |
| 1628 | int skb_size = skb->len, next_skb_size = next_skb->len; |
| 1629 | u16 flags = TCP_SKB_CB(skb)->flags; |
| 1630 | |
| 1631 | /* Also punt if next skb has been SACK'd. */ |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 1632 | if (TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_ACKED) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1633 | return; |
| 1634 | |
| 1635 | /* Next skb is out of window. */ |
| 1636 | if (after(TCP_SKB_CB(next_skb)->end_seq, tp->snd_una+tp->snd_wnd)) |
| 1637 | return; |
| 1638 | |
| 1639 | /* Punt if not enough space exists in the first SKB for |
| 1640 | * the data in the second, or the total combined payload |
| 1641 | * would exceed the MSS. |
| 1642 | */ |
| 1643 | if ((next_skb_size > skb_tailroom(skb)) || |
| 1644 | ((skb_size + next_skb_size) > mss_now)) |
| 1645 | return; |
| 1646 | |
| 1647 | BUG_ON(tcp_skb_pcount(skb) != 1 || |
| 1648 | tcp_skb_pcount(next_skb) != 1); |
| 1649 | |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 1650 | /* changing transmit queue under us so clear hints */ |
| 1651 | clear_all_retrans_hints(tp); |
| 1652 | |
| 1653 | /* Ok. We will be able to collapse the packet. */ |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1654 | tcp_unlink_write_queue(next_skb, sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1655 | |
Arnaldo Carvalho de Melo | 1a4e2d0 | 2007-03-31 11:55:45 -0300 | [diff] [blame] | 1656 | skb_copy_from_linear_data(next_skb, |
| 1657 | skb_put(skb, next_skb_size), |
| 1658 | next_skb_size); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1659 | |
Jarek Poplawski | 52d570a | 2007-01-23 22:07:12 -0800 | [diff] [blame] | 1660 | if (next_skb->ip_summed == CHECKSUM_PARTIAL) |
| 1661 | skb->ip_summed = CHECKSUM_PARTIAL; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1662 | |
Patrick McHardy | 84fa793 | 2006-08-29 16:44:56 -0700 | [diff] [blame] | 1663 | if (skb->ip_summed != CHECKSUM_PARTIAL) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1664 | skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); |
| 1665 | |
| 1666 | /* Update sequence range on original skb. */ |
| 1667 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; |
| 1668 | |
| 1669 | /* Merge over control information. */ |
| 1670 | flags |= TCP_SKB_CB(next_skb)->flags; /* This moves PSH/FIN etc. over */ |
| 1671 | TCP_SKB_CB(skb)->flags = flags; |
| 1672 | |
| 1673 | /* All done, get rid of second SKB and account for it so |
| 1674 | * packet counting does not break. |
| 1675 | */ |
| 1676 | TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked&(TCPCB_EVER_RETRANS|TCPCB_AT_TAIL); |
| 1677 | if (TCP_SKB_CB(next_skb)->sacked&TCPCB_SACKED_RETRANS) |
| 1678 | tp->retrans_out -= tcp_skb_pcount(next_skb); |
| 1679 | if (TCP_SKB_CB(next_skb)->sacked&TCPCB_LOST) { |
| 1680 | tp->lost_out -= tcp_skb_pcount(next_skb); |
| 1681 | tp->left_out -= tcp_skb_pcount(next_skb); |
| 1682 | } |
| 1683 | /* Reno case is special. Sigh... */ |
| 1684 | if (!tp->rx_opt.sack_ok && tp->sacked_out) { |
| 1685 | tcp_dec_pcount_approx(&tp->sacked_out, next_skb); |
| 1686 | tp->left_out -= tcp_skb_pcount(next_skb); |
| 1687 | } |
| 1688 | |
| 1689 | /* Not quite right: it can be > snd.fack, but |
| 1690 | * it is better to underestimate fackets. |
| 1691 | */ |
| 1692 | tcp_dec_pcount_approx(&tp->fackets_out, next_skb); |
| 1693 | tcp_packets_out_dec(tp, next_skb); |
| 1694 | sk_stream_free_skb(sk, next_skb); |
| 1695 | } |
| 1696 | } |
| 1697 | |
| 1698 | /* Do a simple retransmit without using the backoff mechanisms in |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 1699 | * tcp_timer. This is used for path mtu discovery. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1700 | * The socket is already locked here. |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 1701 | */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1702 | void tcp_simple_retransmit(struct sock *sk) |
| 1703 | { |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 1704 | const struct inet_connection_sock *icsk = inet_csk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1705 | struct tcp_sock *tp = tcp_sk(sk); |
| 1706 | struct sk_buff *skb; |
| 1707 | unsigned int mss = tcp_current_mss(sk, 0); |
| 1708 | int lost = 0; |
| 1709 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1710 | tcp_for_write_queue(skb, sk) { |
| 1711 | if (skb == tcp_send_head(sk)) |
| 1712 | break; |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 1713 | if (skb->len > mss && |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1714 | !(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) { |
| 1715 | if (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) { |
| 1716 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; |
| 1717 | tp->retrans_out -= tcp_skb_pcount(skb); |
| 1718 | } |
| 1719 | if (!(TCP_SKB_CB(skb)->sacked&TCPCB_LOST)) { |
| 1720 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; |
| 1721 | tp->lost_out += tcp_skb_pcount(skb); |
| 1722 | lost = 1; |
| 1723 | } |
| 1724 | } |
| 1725 | } |
| 1726 | |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 1727 | clear_all_retrans_hints(tp); |
| 1728 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1729 | if (!lost) |
| 1730 | return; |
| 1731 | |
| 1732 | tcp_sync_left_out(tp); |
| 1733 | |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 1734 | /* Don't muck with the congestion window here. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1735 | * Reason is that we do not increase amount of _data_ |
| 1736 | * in network, but units changed and effective |
| 1737 | * cwnd/ssthresh really reduced now. |
| 1738 | */ |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 1739 | if (icsk->icsk_ca_state != TCP_CA_Loss) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1740 | tp->high_seq = tp->snd_nxt; |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 1741 | tp->snd_ssthresh = tcp_current_ssthresh(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1742 | tp->prior_ssthresh = 0; |
| 1743 | tp->undo_marker = 0; |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 1744 | tcp_set_ca_state(sk, TCP_CA_Loss); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1745 | } |
| 1746 | tcp_xmit_retransmit_queue(sk); |
| 1747 | } |
| 1748 | |
| 1749 | /* This retransmits one SKB. Policy decisions and retransmit queue |
| 1750 | * state updates are done by the caller. Returns non-zero if an |
| 1751 | * error occurred which prevented the send. |
| 1752 | */ |
| 1753 | int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) |
| 1754 | { |
| 1755 | struct tcp_sock *tp = tcp_sk(sk); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1756 | struct inet_connection_sock *icsk = inet_csk(sk); |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 1757 | unsigned int cur_mss = tcp_current_mss(sk, 0); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1758 | int err; |
| 1759 | |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1760 | /* Inconslusive MTU probe */ |
| 1761 | if (icsk->icsk_mtup.probe_size) { |
| 1762 | icsk->icsk_mtup.probe_size = 0; |
| 1763 | } |
| 1764 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1765 | /* 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] | 1766 | * copying overhead: fragmentation, tunneling, mangling etc. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1767 | */ |
| 1768 | if (atomic_read(&sk->sk_wmem_alloc) > |
| 1769 | min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf)) |
| 1770 | return -EAGAIN; |
| 1771 | |
| 1772 | if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { |
| 1773 | if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) |
| 1774 | BUG(); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1775 | if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) |
| 1776 | return -ENOMEM; |
| 1777 | } |
| 1778 | |
| 1779 | /* If receiver has shrunk his window, and skb is out of |
| 1780 | * new window, do not retransmit it. The exception is the |
| 1781 | * case, when window is shrunk to zero. In this case |
| 1782 | * our retransmit serves as a zero window probe. |
| 1783 | */ |
| 1784 | if (!before(TCP_SKB_CB(skb)->seq, tp->snd_una+tp->snd_wnd) |
| 1785 | && TCP_SKB_CB(skb)->seq != tp->snd_una) |
| 1786 | return -EAGAIN; |
| 1787 | |
| 1788 | if (skb->len > cur_mss) { |
David S. Miller | 846998a | 2005-08-04 19:52:01 -0700 | [diff] [blame] | 1789 | if (tcp_fragment(sk, skb, cur_mss, cur_mss)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1790 | return -ENOMEM; /* We'll try again later. */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1791 | } |
| 1792 | |
| 1793 | /* Collapse two adjacent packets if worthwhile and we can. */ |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 1794 | if (!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN) && |
| 1795 | (skb->len < (cur_mss >> 1)) && |
| 1796 | (tcp_write_queue_next(sk, skb) != tcp_send_head(sk)) && |
| 1797 | (!tcp_skb_is_last(sk, skb)) && |
| 1798 | (skb_shinfo(skb)->nr_frags == 0 && skb_shinfo(tcp_write_queue_next(sk, skb))->nr_frags == 0) && |
| 1799 | (tcp_skb_pcount(skb) == 1 && tcp_skb_pcount(tcp_write_queue_next(sk, skb)) == 1) && |
| 1800 | (sysctl_tcp_retrans_collapse != 0)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1801 | tcp_retrans_try_collapse(sk, skb, cur_mss); |
| 1802 | |
Arnaldo Carvalho de Melo | 8292a17 | 2005-12-13 23:15:52 -0800 | [diff] [blame] | 1803 | if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1804 | return -EHOSTUNREACH; /* Routing failure or similar. */ |
| 1805 | |
| 1806 | /* Some Solaris stacks overoptimize and ignore the FIN on a |
| 1807 | * retransmit when old data is attached. So strip it off |
| 1808 | * since it is cheap to do so and saves bytes on the network. |
| 1809 | */ |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 1810 | if (skb->len > 0 && |
| 1811 | (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) && |
| 1812 | tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1813 | if (!pskb_trim(skb, 0)) { |
| 1814 | TCP_SKB_CB(skb)->seq = TCP_SKB_CB(skb)->end_seq - 1; |
Herbert Xu | 7967168 | 2006-06-22 02:40:14 -0700 | [diff] [blame] | 1815 | skb_shinfo(skb)->gso_segs = 1; |
| 1816 | skb_shinfo(skb)->gso_size = 0; |
| 1817 | skb_shinfo(skb)->gso_type = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1818 | skb->ip_summed = CHECKSUM_NONE; |
| 1819 | skb->csum = 0; |
| 1820 | } |
| 1821 | } |
| 1822 | |
| 1823 | /* Make a copy, if the first transmission SKB clone we made |
| 1824 | * is still in somebody's hands, else make a clone. |
| 1825 | */ |
| 1826 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1827 | |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 1828 | err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1829 | |
| 1830 | if (err == 0) { |
| 1831 | /* Update global TCP statistics. */ |
| 1832 | TCP_INC_STATS(TCP_MIB_RETRANSSEGS); |
| 1833 | |
| 1834 | tp->total_retrans++; |
| 1835 | |
| 1836 | #if FASTRETRANS_DEBUG > 0 |
| 1837 | if (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) { |
| 1838 | if (net_ratelimit()) |
| 1839 | printk(KERN_DEBUG "retrans_out leaked.\n"); |
| 1840 | } |
| 1841 | #endif |
| 1842 | TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; |
| 1843 | tp->retrans_out += tcp_skb_pcount(skb); |
| 1844 | |
| 1845 | /* Save stamp of the first retransmit. */ |
| 1846 | if (!tp->retrans_stamp) |
| 1847 | tp->retrans_stamp = TCP_SKB_CB(skb)->when; |
| 1848 | |
| 1849 | tp->undo_retrans++; |
| 1850 | |
| 1851 | /* snd_nxt is stored to detect loss of retransmitted segment, |
| 1852 | * see tcp_input.c tcp_sacktag_write_queue(). |
| 1853 | */ |
| 1854 | TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt; |
| 1855 | } |
| 1856 | return err; |
| 1857 | } |
| 1858 | |
| 1859 | /* This gets called after a retransmit timeout, and the initially |
| 1860 | * retransmitted data is acknowledged. It tries to continue |
| 1861 | * resending the rest of the retransmit queue, until either |
| 1862 | * we've sent it all or the congestion window limit is reached. |
| 1863 | * If doing SACK, the first ACK which comes back for a timeout |
| 1864 | * based retransmit packet might feed us FACK information again. |
| 1865 | * If so, we use it to avoid unnecessarily retransmissions. |
| 1866 | */ |
| 1867 | void tcp_xmit_retransmit_queue(struct sock *sk) |
| 1868 | { |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 1869 | const struct inet_connection_sock *icsk = inet_csk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1870 | struct tcp_sock *tp = tcp_sk(sk); |
| 1871 | struct sk_buff *skb; |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 1872 | int packet_cnt; |
| 1873 | |
| 1874 | if (tp->retransmit_skb_hint) { |
| 1875 | skb = tp->retransmit_skb_hint; |
| 1876 | packet_cnt = tp->retransmit_cnt_hint; |
| 1877 | }else{ |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1878 | skb = tcp_write_queue_head(sk); |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 1879 | packet_cnt = 0; |
| 1880 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1881 | |
| 1882 | /* First pass: retransmit lost packets. */ |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 1883 | if (tp->lost_out) { |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1884 | tcp_for_write_queue_from(skb, sk) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1885 | __u8 sacked = TCP_SKB_CB(skb)->sacked; |
| 1886 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1887 | if (skb == tcp_send_head(sk)) |
| 1888 | break; |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 1889 | /* we could do better than to assign each time */ |
| 1890 | tp->retransmit_skb_hint = skb; |
| 1891 | tp->retransmit_cnt_hint = packet_cnt; |
| 1892 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1893 | /* Assume this retransmit will generate |
| 1894 | * only one packet for congestion window |
| 1895 | * calculation purposes. This works because |
| 1896 | * tcp_retransmit_skb() will chop up the |
| 1897 | * packet to be MSS sized and all the |
| 1898 | * packet counting works out. |
| 1899 | */ |
| 1900 | if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) |
| 1901 | return; |
| 1902 | |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 1903 | if (sacked & TCPCB_LOST) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1904 | if (!(sacked&(TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))) { |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 1905 | if (tcp_retransmit_skb(sk, skb)) { |
| 1906 | tp->retransmit_skb_hint = NULL; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1907 | return; |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 1908 | } |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 1909 | if (icsk->icsk_ca_state != TCP_CA_Loss) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1910 | NET_INC_STATS_BH(LINUX_MIB_TCPFASTRETRANS); |
| 1911 | else |
| 1912 | NET_INC_STATS_BH(LINUX_MIB_TCPSLOWSTARTRETRANS); |
| 1913 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1914 | if (skb == tcp_write_queue_head(sk)) |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 1915 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, |
Arnaldo Carvalho de Melo | 3f421ba | 2005-08-09 20:11:08 -0700 | [diff] [blame] | 1916 | inet_csk(sk)->icsk_rto, |
| 1917 | TCP_RTO_MAX); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1918 | } |
| 1919 | |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 1920 | packet_cnt += tcp_skb_pcount(skb); |
| 1921 | if (packet_cnt >= tp->lost_out) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1922 | break; |
| 1923 | } |
| 1924 | } |
| 1925 | } |
| 1926 | |
| 1927 | /* OK, demanded retransmission is finished. */ |
| 1928 | |
| 1929 | /* Forward retransmissions are possible only during Recovery. */ |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 1930 | if (icsk->icsk_ca_state != TCP_CA_Recovery) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1931 | return; |
| 1932 | |
| 1933 | /* No forward retransmissions in Reno are possible. */ |
| 1934 | if (!tp->rx_opt.sack_ok) |
| 1935 | return; |
| 1936 | |
| 1937 | /* Yeah, we have to make difficult choice between forward transmission |
| 1938 | * and retransmission... Both ways have their merits... |
| 1939 | * |
| 1940 | * For now we do not retransmit anything, while we have some new |
| 1941 | * segments to send. |
| 1942 | */ |
| 1943 | |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 1944 | if (tcp_may_send_now(sk)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1945 | return; |
| 1946 | |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 1947 | if (tp->forward_skb_hint) { |
| 1948 | skb = tp->forward_skb_hint; |
| 1949 | packet_cnt = tp->forward_cnt_hint; |
| 1950 | } else{ |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1951 | skb = tcp_write_queue_head(sk); |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 1952 | packet_cnt = 0; |
| 1953 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1954 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1955 | tcp_for_write_queue_from(skb, sk) { |
| 1956 | if (skb == tcp_send_head(sk)) |
| 1957 | break; |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 1958 | tp->forward_cnt_hint = packet_cnt; |
| 1959 | tp->forward_skb_hint = skb; |
| 1960 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1961 | /* Similar to the retransmit loop above we |
| 1962 | * can pretend that the retransmitted SKB |
| 1963 | * we send out here will be composed of one |
| 1964 | * real MSS sized packet because tcp_retransmit_skb() |
| 1965 | * will fragment it if necessary. |
| 1966 | */ |
| 1967 | if (++packet_cnt > tp->fackets_out) |
| 1968 | break; |
| 1969 | |
| 1970 | if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) |
| 1971 | break; |
| 1972 | |
| 1973 | if (TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) |
| 1974 | continue; |
| 1975 | |
| 1976 | /* Ok, retransmit it. */ |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 1977 | if (tcp_retransmit_skb(sk, skb)) { |
| 1978 | tp->forward_skb_hint = NULL; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1979 | break; |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 1980 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1981 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1982 | if (skb == tcp_write_queue_head(sk)) |
Arnaldo Carvalho de Melo | 3f421ba | 2005-08-09 20:11:08 -0700 | [diff] [blame] | 1983 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, |
| 1984 | inet_csk(sk)->icsk_rto, |
| 1985 | TCP_RTO_MAX); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1986 | |
| 1987 | NET_INC_STATS_BH(LINUX_MIB_TCPFORWARDRETRANS); |
| 1988 | } |
| 1989 | } |
| 1990 | |
| 1991 | |
| 1992 | /* Send a fin. The caller locks the socket for us. This cannot be |
| 1993 | * allowed to fail queueing a FIN frame under any circumstances. |
| 1994 | */ |
| 1995 | void tcp_send_fin(struct sock *sk) |
| 1996 | { |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 1997 | struct tcp_sock *tp = tcp_sk(sk); |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1998 | struct sk_buff *skb = tcp_write_queue_tail(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1999 | int mss_now; |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2000 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2001 | /* Optimization, tack on the FIN if we have a queue of |
| 2002 | * unsent frames. But be careful about outgoing SACKS |
| 2003 | * and IP options. |
| 2004 | */ |
| 2005 | mss_now = tcp_current_mss(sk, 1); |
| 2006 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2007 | if (tcp_send_head(sk) != NULL) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2008 | TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_FIN; |
| 2009 | TCP_SKB_CB(skb)->end_seq++; |
| 2010 | tp->write_seq++; |
| 2011 | } else { |
| 2012 | /* Socket is locked, keep trying until memory is available. */ |
| 2013 | for (;;) { |
David S. Miller | d179cd1 | 2005-08-17 14:57:30 -0700 | [diff] [blame] | 2014 | skb = alloc_skb_fclone(MAX_TCP_HEADER, GFP_KERNEL); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2015 | if (skb) |
| 2016 | break; |
| 2017 | yield(); |
| 2018 | } |
| 2019 | |
| 2020 | /* Reserve space for headers and prepare control bits. */ |
| 2021 | skb_reserve(skb, MAX_TCP_HEADER); |
| 2022 | skb->csum = 0; |
| 2023 | TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_FIN); |
| 2024 | TCP_SKB_CB(skb)->sacked = 0; |
Herbert Xu | 7967168 | 2006-06-22 02:40:14 -0700 | [diff] [blame] | 2025 | skb_shinfo(skb)->gso_segs = 1; |
| 2026 | skb_shinfo(skb)->gso_size = 0; |
| 2027 | skb_shinfo(skb)->gso_type = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2028 | |
| 2029 | /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ |
| 2030 | TCP_SKB_CB(skb)->seq = tp->write_seq; |
| 2031 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1; |
| 2032 | tcp_queue_skb(sk, skb); |
| 2033 | } |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2034 | __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2035 | } |
| 2036 | |
| 2037 | /* We get here when a process closes a file descriptor (either due to |
| 2038 | * an explicit close() or as a byproduct of exit()'ing) and there |
| 2039 | * was unread data in the receive queue. This behavior is recommended |
Gerrit Renker | 65bb723 | 2007-04-28 21:21:46 -0700 | [diff] [blame] | 2040 | * by RFC 2525, section 2.17. -DaveM |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2041 | */ |
Al Viro | dd0fc66 | 2005-10-07 07:46:04 +0100 | [diff] [blame] | 2042 | void tcp_send_active_reset(struct sock *sk, gfp_t priority) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2043 | { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2044 | struct sk_buff *skb; |
| 2045 | |
| 2046 | /* NOTE: No TCP options attached and we never retransmit this. */ |
| 2047 | skb = alloc_skb(MAX_TCP_HEADER, priority); |
| 2048 | if (!skb) { |
| 2049 | NET_INC_STATS(LINUX_MIB_TCPABORTFAILED); |
| 2050 | return; |
| 2051 | } |
| 2052 | |
| 2053 | /* Reserve space for headers and prepare control bits. */ |
| 2054 | skb_reserve(skb, MAX_TCP_HEADER); |
| 2055 | skb->csum = 0; |
| 2056 | TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_RST); |
| 2057 | TCP_SKB_CB(skb)->sacked = 0; |
Herbert Xu | 7967168 | 2006-06-22 02:40:14 -0700 | [diff] [blame] | 2058 | skb_shinfo(skb)->gso_segs = 1; |
| 2059 | skb_shinfo(skb)->gso_size = 0; |
| 2060 | skb_shinfo(skb)->gso_type = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2061 | |
| 2062 | /* Send it off. */ |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2063 | TCP_SKB_CB(skb)->seq = tcp_acceptable_seq(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2064 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq; |
| 2065 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 2066 | if (tcp_transmit_skb(sk, skb, 0, priority)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2067 | NET_INC_STATS(LINUX_MIB_TCPABORTFAILED); |
| 2068 | } |
| 2069 | |
| 2070 | /* WARNING: This routine must only be called when we have already sent |
| 2071 | * a SYN packet that crossed the incoming SYN that caused this routine |
| 2072 | * to get called. If this assumption fails then the initial rcv_wnd |
| 2073 | * and rcv_wscale values will not be correct. |
| 2074 | */ |
| 2075 | int tcp_send_synack(struct sock *sk) |
| 2076 | { |
| 2077 | struct sk_buff* skb; |
| 2078 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2079 | skb = tcp_write_queue_head(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2080 | if (skb == NULL || !(TCP_SKB_CB(skb)->flags&TCPCB_FLAG_SYN)) { |
| 2081 | printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n"); |
| 2082 | return -EFAULT; |
| 2083 | } |
| 2084 | if (!(TCP_SKB_CB(skb)->flags&TCPCB_FLAG_ACK)) { |
| 2085 | if (skb_cloned(skb)) { |
| 2086 | struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC); |
| 2087 | if (nskb == NULL) |
| 2088 | return -ENOMEM; |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2089 | tcp_unlink_write_queue(skb, sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2090 | skb_header_release(nskb); |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2091 | __tcp_add_write_queue_head(sk, nskb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2092 | sk_stream_free_skb(sk, skb); |
| 2093 | sk_charge_skb(sk, nskb); |
| 2094 | skb = nskb; |
| 2095 | } |
| 2096 | |
| 2097 | TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ACK; |
| 2098 | TCP_ECN_send_synack(tcp_sk(sk), skb); |
| 2099 | } |
| 2100 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 2101 | return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2102 | } |
| 2103 | |
| 2104 | /* |
| 2105 | * Prepare a SYN-ACK. |
| 2106 | */ |
| 2107 | struct sk_buff * tcp_make_synack(struct sock *sk, struct dst_entry *dst, |
Arnaldo Carvalho de Melo | 60236fd | 2005-06-18 22:47:21 -0700 | [diff] [blame] | 2108 | struct request_sock *req) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2109 | { |
Arnaldo Carvalho de Melo | 2e6599c | 2005-06-18 22:46:52 -0700 | [diff] [blame] | 2110 | struct inet_request_sock *ireq = inet_rsk(req); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2111 | struct tcp_sock *tp = tcp_sk(sk); |
| 2112 | struct tcphdr *th; |
| 2113 | int tcp_header_size; |
| 2114 | struct sk_buff *skb; |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 2115 | #ifdef CONFIG_TCP_MD5SIG |
| 2116 | struct tcp_md5sig_key *md5; |
| 2117 | __u8 *md5_hash_location; |
| 2118 | #endif |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2119 | |
| 2120 | skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC); |
| 2121 | if (skb == NULL) |
| 2122 | return NULL; |
| 2123 | |
| 2124 | /* Reserve space for headers. */ |
| 2125 | skb_reserve(skb, MAX_TCP_HEADER); |
| 2126 | |
| 2127 | skb->dst = dst_clone(dst); |
| 2128 | |
| 2129 | tcp_header_size = (sizeof(struct tcphdr) + TCPOLEN_MSS + |
Arnaldo Carvalho de Melo | 2e6599c | 2005-06-18 22:46:52 -0700 | [diff] [blame] | 2130 | (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0) + |
| 2131 | (ireq->wscale_ok ? TCPOLEN_WSCALE_ALIGNED : 0) + |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2132 | /* 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] | 2133 | ((ireq->sack_ok && !ireq->tstamp_ok) ? TCPOLEN_SACKPERM_ALIGNED : 0)); |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 2134 | |
| 2135 | #ifdef CONFIG_TCP_MD5SIG |
| 2136 | /* Are we doing MD5 on this segment? If so - make room for it */ |
| 2137 | md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req); |
| 2138 | if (md5) |
| 2139 | tcp_header_size += TCPOLEN_MD5SIG_ALIGNED; |
| 2140 | #endif |
Arnaldo Carvalho de Melo | aa8223c | 2007-04-10 21:04:22 -0700 | [diff] [blame] | 2141 | skb_push(skb, tcp_header_size); |
| 2142 | skb_reset_transport_header(skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2143 | |
Arnaldo Carvalho de Melo | aa8223c | 2007-04-10 21:04:22 -0700 | [diff] [blame] | 2144 | th = tcp_hdr(skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2145 | memset(th, 0, sizeof(struct tcphdr)); |
| 2146 | th->syn = 1; |
| 2147 | th->ack = 1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2148 | TCP_ECN_make_synack(req, th); |
| 2149 | th->source = inet_sk(sk)->sport; |
Arnaldo Carvalho de Melo | 2e6599c | 2005-06-18 22:46:52 -0700 | [diff] [blame] | 2150 | th->dest = ireq->rmt_port; |
| 2151 | TCP_SKB_CB(skb)->seq = tcp_rsk(req)->snt_isn; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2152 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1; |
| 2153 | TCP_SKB_CB(skb)->sacked = 0; |
Herbert Xu | 7967168 | 2006-06-22 02:40:14 -0700 | [diff] [blame] | 2154 | skb_shinfo(skb)->gso_segs = 1; |
| 2155 | skb_shinfo(skb)->gso_size = 0; |
| 2156 | skb_shinfo(skb)->gso_type = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2157 | th->seq = htonl(TCP_SKB_CB(skb)->seq); |
Arnaldo Carvalho de Melo | 2e6599c | 2005-06-18 22:46:52 -0700 | [diff] [blame] | 2158 | th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2159 | if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */ |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2160 | __u8 rcv_wscale; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2161 | /* Set this up on the first call only */ |
| 2162 | req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW); |
| 2163 | /* tcp_full_space because it is guaranteed to be the first packet */ |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2164 | tcp_select_initial_window(tcp_full_space(sk), |
Arnaldo Carvalho de Melo | 2e6599c | 2005-06-18 22:46:52 -0700 | [diff] [blame] | 2165 | dst_metric(dst, RTAX_ADVMSS) - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0), |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2166 | &req->rcv_wnd, |
| 2167 | &req->window_clamp, |
Arnaldo Carvalho de Melo | 2e6599c | 2005-06-18 22:46:52 -0700 | [diff] [blame] | 2168 | ireq->wscale_ok, |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2169 | &rcv_wscale); |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2170 | ireq->rcv_wscale = rcv_wscale; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2171 | } |
| 2172 | |
| 2173 | /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ |
Ilpo Järvinen | 600ff0c | 2007-02-13 12:42:11 -0800 | [diff] [blame] | 2174 | th->window = htons(min(req->rcv_wnd, 65535U)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2175 | |
| 2176 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
Al Viro | df7a3b0 | 2006-09-27 18:38:52 -0700 | [diff] [blame] | 2177 | 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] | 2178 | ireq->sack_ok, ireq->wscale_ok, ireq->rcv_wscale, |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2179 | TCP_SKB_CB(skb)->when, |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 2180 | req->ts_recent, |
| 2181 | ( |
| 2182 | #ifdef CONFIG_TCP_MD5SIG |
| 2183 | md5 ? &md5_hash_location : |
| 2184 | #endif |
| 2185 | NULL) |
| 2186 | ); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2187 | |
| 2188 | skb->csum = 0; |
| 2189 | th->doff = (tcp_header_size >> 2); |
| 2190 | TCP_INC_STATS(TCP_MIB_OUTSEGS); |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 2191 | |
| 2192 | #ifdef CONFIG_TCP_MD5SIG |
| 2193 | /* Okay, we have all we need - do the md5 hash if needed */ |
| 2194 | if (md5) { |
| 2195 | tp->af_specific->calc_md5_hash(md5_hash_location, |
| 2196 | md5, |
| 2197 | NULL, dst, req, |
Arnaldo Carvalho de Melo | aa8223c | 2007-04-10 21:04:22 -0700 | [diff] [blame] | 2198 | tcp_hdr(skb), sk->sk_protocol, |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 2199 | skb->len); |
| 2200 | } |
| 2201 | #endif |
| 2202 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2203 | return skb; |
| 2204 | } |
| 2205 | |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2206 | /* |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2207 | * Do all connect socket setups that can be done AF independent. |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2208 | */ |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 2209 | static void tcp_connect_init(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2210 | { |
| 2211 | struct dst_entry *dst = __sk_dst_get(sk); |
| 2212 | struct tcp_sock *tp = tcp_sk(sk); |
| 2213 | __u8 rcv_wscale; |
| 2214 | |
| 2215 | /* We'll fix this up when we get a response from the other end. |
| 2216 | * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. |
| 2217 | */ |
| 2218 | tp->tcp_header_len = sizeof(struct tcphdr) + |
| 2219 | (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0); |
| 2220 | |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 2221 | #ifdef CONFIG_TCP_MD5SIG |
| 2222 | if (tp->af_specific->md5_lookup(sk, sk) != NULL) |
| 2223 | tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; |
| 2224 | #endif |
| 2225 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2226 | /* If user gave his TCP_MAXSEG, record it to clamp */ |
| 2227 | if (tp->rx_opt.user_mss) |
| 2228 | tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; |
| 2229 | tp->max_window = 0; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 2230 | tcp_mtup_init(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2231 | tcp_sync_mss(sk, dst_mtu(dst)); |
| 2232 | |
| 2233 | if (!tp->window_clamp) |
| 2234 | tp->window_clamp = dst_metric(dst, RTAX_WINDOW); |
| 2235 | tp->advmss = dst_metric(dst, RTAX_ADVMSS); |
| 2236 | tcp_initialize_rcv_mss(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2237 | |
| 2238 | tcp_select_initial_window(tcp_full_space(sk), |
| 2239 | tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), |
| 2240 | &tp->rcv_wnd, |
| 2241 | &tp->window_clamp, |
| 2242 | sysctl_tcp_window_scaling, |
| 2243 | &rcv_wscale); |
| 2244 | |
| 2245 | tp->rx_opt.rcv_wscale = rcv_wscale; |
| 2246 | tp->rcv_ssthresh = tp->rcv_wnd; |
| 2247 | |
| 2248 | sk->sk_err = 0; |
| 2249 | sock_reset_flag(sk, SOCK_DONE); |
| 2250 | tp->snd_wnd = 0; |
| 2251 | tcp_init_wl(tp, tp->write_seq, 0); |
| 2252 | tp->snd_una = tp->write_seq; |
| 2253 | tp->snd_sml = tp->write_seq; |
| 2254 | tp->rcv_nxt = 0; |
| 2255 | tp->rcv_wup = 0; |
| 2256 | tp->copied_seq = 0; |
| 2257 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2258 | inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; |
| 2259 | inet_csk(sk)->icsk_retransmits = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2260 | tcp_clear_retrans(tp); |
| 2261 | } |
| 2262 | |
| 2263 | /* |
| 2264 | * Build a SYN and send it off. |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2265 | */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2266 | int tcp_connect(struct sock *sk) |
| 2267 | { |
| 2268 | struct tcp_sock *tp = tcp_sk(sk); |
| 2269 | struct sk_buff *buff; |
| 2270 | |
| 2271 | tcp_connect_init(sk); |
| 2272 | |
David S. Miller | d179cd1 | 2005-08-17 14:57:30 -0700 | [diff] [blame] | 2273 | buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2274 | if (unlikely(buff == NULL)) |
| 2275 | return -ENOBUFS; |
| 2276 | |
| 2277 | /* Reserve space for headers. */ |
| 2278 | skb_reserve(buff, MAX_TCP_HEADER); |
| 2279 | |
| 2280 | TCP_SKB_CB(buff)->flags = TCPCB_FLAG_SYN; |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2281 | TCP_ECN_send_syn(sk, buff); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2282 | TCP_SKB_CB(buff)->sacked = 0; |
Herbert Xu | 7967168 | 2006-06-22 02:40:14 -0700 | [diff] [blame] | 2283 | skb_shinfo(buff)->gso_segs = 1; |
| 2284 | skb_shinfo(buff)->gso_size = 0; |
| 2285 | skb_shinfo(buff)->gso_type = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2286 | buff->csum = 0; |
Wei Yongjun | bd37a08 | 2006-08-07 21:04:15 -0700 | [diff] [blame] | 2287 | tp->snd_nxt = tp->write_seq; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2288 | TCP_SKB_CB(buff)->seq = tp->write_seq++; |
| 2289 | TCP_SKB_CB(buff)->end_seq = tp->write_seq; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2290 | |
| 2291 | /* Send it off. */ |
| 2292 | TCP_SKB_CB(buff)->when = tcp_time_stamp; |
| 2293 | tp->retrans_stamp = TCP_SKB_CB(buff)->when; |
| 2294 | skb_header_release(buff); |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2295 | __tcp_add_write_queue_tail(sk, buff); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2296 | sk_charge_skb(sk, buff); |
| 2297 | tp->packets_out += tcp_skb_pcount(buff); |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 2298 | tcp_transmit_skb(sk, buff, 1, GFP_KERNEL); |
Wei Yongjun | bd37a08 | 2006-08-07 21:04:15 -0700 | [diff] [blame] | 2299 | |
| 2300 | /* We change tp->snd_nxt after the tcp_transmit_skb() call |
| 2301 | * in order to make this packet get counted in tcpOutSegs. |
| 2302 | */ |
| 2303 | tp->snd_nxt = tp->write_seq; |
| 2304 | tp->pushed_seq = tp->write_seq; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2305 | TCP_INC_STATS(TCP_MIB_ACTIVEOPENS); |
| 2306 | |
| 2307 | /* Timer for repeating the SYN until an answer. */ |
Arnaldo Carvalho de Melo | 3f421ba | 2005-08-09 20:11:08 -0700 | [diff] [blame] | 2308 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, |
| 2309 | inet_csk(sk)->icsk_rto, TCP_RTO_MAX); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2310 | return 0; |
| 2311 | } |
| 2312 | |
| 2313 | /* Send out a delayed ack, the caller does the policy checking |
| 2314 | * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() |
| 2315 | * for details. |
| 2316 | */ |
| 2317 | void tcp_send_delayed_ack(struct sock *sk) |
| 2318 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2319 | struct inet_connection_sock *icsk = inet_csk(sk); |
| 2320 | int ato = icsk->icsk_ack.ato; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2321 | unsigned long timeout; |
| 2322 | |
| 2323 | if (ato > TCP_DELACK_MIN) { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2324 | const struct tcp_sock *tp = tcp_sk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2325 | int max_ato = HZ/2; |
| 2326 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2327 | if (icsk->icsk_ack.pingpong || (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2328 | max_ato = TCP_DELACK_MAX; |
| 2329 | |
| 2330 | /* Slow path, intersegment interval is "high". */ |
| 2331 | |
| 2332 | /* 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] | 2333 | * 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] | 2334 | * directly. |
| 2335 | */ |
| 2336 | if (tp->srtt) { |
| 2337 | int rtt = max(tp->srtt>>3, TCP_DELACK_MIN); |
| 2338 | |
| 2339 | if (rtt < max_ato) |
| 2340 | max_ato = rtt; |
| 2341 | } |
| 2342 | |
| 2343 | ato = min(ato, max_ato); |
| 2344 | } |
| 2345 | |
| 2346 | /* Stay within the limit we were given */ |
| 2347 | timeout = jiffies + ato; |
| 2348 | |
| 2349 | /* 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] | 2350 | if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2351 | /* If delack timer was blocked or is about to expire, |
| 2352 | * send ACK now. |
| 2353 | */ |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2354 | if (icsk->icsk_ack.blocked || |
| 2355 | time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2356 | tcp_send_ack(sk); |
| 2357 | return; |
| 2358 | } |
| 2359 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2360 | if (!time_before(timeout, icsk->icsk_ack.timeout)) |
| 2361 | timeout = icsk->icsk_ack.timeout; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2362 | } |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2363 | icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; |
| 2364 | icsk->icsk_ack.timeout = timeout; |
| 2365 | sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2366 | } |
| 2367 | |
| 2368 | /* This routine sends an ack and also updates the window. */ |
| 2369 | void tcp_send_ack(struct sock *sk) |
| 2370 | { |
| 2371 | /* If we have been reset, we may not send again. */ |
| 2372 | if (sk->sk_state != TCP_CLOSE) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2373 | struct sk_buff *buff; |
| 2374 | |
| 2375 | /* We are not putting this on the write queue, so |
| 2376 | * tcp_transmit_skb() will set the ownership to this |
| 2377 | * sock. |
| 2378 | */ |
| 2379 | buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); |
| 2380 | if (buff == NULL) { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2381 | inet_csk_schedule_ack(sk); |
| 2382 | inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; |
Arnaldo Carvalho de Melo | 3f421ba | 2005-08-09 20:11:08 -0700 | [diff] [blame] | 2383 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, |
| 2384 | TCP_DELACK_MAX, TCP_RTO_MAX); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2385 | return; |
| 2386 | } |
| 2387 | |
| 2388 | /* Reserve space for headers and prepare control bits. */ |
| 2389 | skb_reserve(buff, MAX_TCP_HEADER); |
| 2390 | buff->csum = 0; |
| 2391 | TCP_SKB_CB(buff)->flags = TCPCB_FLAG_ACK; |
| 2392 | TCP_SKB_CB(buff)->sacked = 0; |
Herbert Xu | 7967168 | 2006-06-22 02:40:14 -0700 | [diff] [blame] | 2393 | skb_shinfo(buff)->gso_segs = 1; |
| 2394 | skb_shinfo(buff)->gso_size = 0; |
| 2395 | skb_shinfo(buff)->gso_type = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2396 | |
| 2397 | /* Send it off, this clears delayed acks for us. */ |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2398 | TCP_SKB_CB(buff)->seq = TCP_SKB_CB(buff)->end_seq = tcp_acceptable_seq(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2399 | TCP_SKB_CB(buff)->when = tcp_time_stamp; |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 2400 | tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2401 | } |
| 2402 | } |
| 2403 | |
| 2404 | /* This routine sends a packet with an out of date sequence |
| 2405 | * number. It assumes the other end will try to ack it. |
| 2406 | * |
| 2407 | * Question: what should we make while urgent mode? |
| 2408 | * 4.4BSD forces sending single byte of data. We cannot send |
| 2409 | * out of window data, because we have SND.NXT==SND.MAX... |
| 2410 | * |
| 2411 | * Current solution: to send TWO zero-length segments in urgent mode: |
| 2412 | * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is |
| 2413 | * out-of-date with SND.UNA-1 to probe window. |
| 2414 | */ |
| 2415 | static int tcp_xmit_probe_skb(struct sock *sk, int urgent) |
| 2416 | { |
| 2417 | struct tcp_sock *tp = tcp_sk(sk); |
| 2418 | struct sk_buff *skb; |
| 2419 | |
| 2420 | /* We don't queue it, tcp_transmit_skb() sets ownership. */ |
| 2421 | skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2422 | if (skb == NULL) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2423 | return -1; |
| 2424 | |
| 2425 | /* Reserve space for headers and set control bits. */ |
| 2426 | skb_reserve(skb, MAX_TCP_HEADER); |
| 2427 | skb->csum = 0; |
| 2428 | TCP_SKB_CB(skb)->flags = TCPCB_FLAG_ACK; |
| 2429 | TCP_SKB_CB(skb)->sacked = urgent; |
Herbert Xu | 7967168 | 2006-06-22 02:40:14 -0700 | [diff] [blame] | 2430 | skb_shinfo(skb)->gso_segs = 1; |
| 2431 | skb_shinfo(skb)->gso_size = 0; |
| 2432 | skb_shinfo(skb)->gso_type = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2433 | |
| 2434 | /* Use a previous sequence. This should cause the other |
| 2435 | * end to send an ack. Don't queue or clone SKB, just |
| 2436 | * send it. |
| 2437 | */ |
| 2438 | TCP_SKB_CB(skb)->seq = urgent ? tp->snd_una : tp->snd_una - 1; |
| 2439 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq; |
| 2440 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 2441 | return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2442 | } |
| 2443 | |
| 2444 | int tcp_write_wakeup(struct sock *sk) |
| 2445 | { |
| 2446 | if (sk->sk_state != TCP_CLOSE) { |
| 2447 | struct tcp_sock *tp = tcp_sk(sk); |
| 2448 | struct sk_buff *skb; |
| 2449 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2450 | if ((skb = tcp_send_head(sk)) != NULL && |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2451 | before(TCP_SKB_CB(skb)->seq, tp->snd_una+tp->snd_wnd)) { |
| 2452 | int err; |
| 2453 | unsigned int mss = tcp_current_mss(sk, 0); |
| 2454 | unsigned int seg_size = tp->snd_una+tp->snd_wnd-TCP_SKB_CB(skb)->seq; |
| 2455 | |
| 2456 | if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) |
| 2457 | tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; |
| 2458 | |
| 2459 | /* We are probing the opening of a window |
| 2460 | * but the window size is != 0 |
| 2461 | * must have been a result SWS avoidance ( sender ) |
| 2462 | */ |
| 2463 | if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || |
| 2464 | skb->len > mss) { |
| 2465 | seg_size = min(seg_size, mss); |
| 2466 | TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH; |
David S. Miller | 846998a | 2005-08-04 19:52:01 -0700 | [diff] [blame] | 2467 | if (tcp_fragment(sk, skb, seg_size, mss)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2468 | return -1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2469 | } else if (!tcp_skb_pcount(skb)) |
David S. Miller | 846998a | 2005-08-04 19:52:01 -0700 | [diff] [blame] | 2470 | tcp_set_skb_tso_segs(sk, skb, mss); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2471 | |
| 2472 | TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH; |
| 2473 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 2474 | err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2475 | if (!err) { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2476 | update_send_head(sk, skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2477 | } |
| 2478 | return err; |
| 2479 | } else { |
| 2480 | if (tp->urg_mode && |
| 2481 | between(tp->snd_up, tp->snd_una+1, tp->snd_una+0xFFFF)) |
| 2482 | tcp_xmit_probe_skb(sk, TCPCB_URG); |
| 2483 | return tcp_xmit_probe_skb(sk, 0); |
| 2484 | } |
| 2485 | } |
| 2486 | return -1; |
| 2487 | } |
| 2488 | |
| 2489 | /* A window probe timeout has occurred. If window is not closed send |
| 2490 | * a partial packet else a zero probe. |
| 2491 | */ |
| 2492 | void tcp_send_probe0(struct sock *sk) |
| 2493 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2494 | struct inet_connection_sock *icsk = inet_csk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2495 | struct tcp_sock *tp = tcp_sk(sk); |
| 2496 | int err; |
| 2497 | |
| 2498 | err = tcp_write_wakeup(sk); |
| 2499 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2500 | if (tp->packets_out || !tcp_send_head(sk)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2501 | /* Cancel probe timer, if it is not required. */ |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2502 | icsk->icsk_probes_out = 0; |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2503 | icsk->icsk_backoff = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2504 | return; |
| 2505 | } |
| 2506 | |
| 2507 | if (err <= 0) { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2508 | if (icsk->icsk_backoff < sysctl_tcp_retries2) |
| 2509 | icsk->icsk_backoff++; |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2510 | icsk->icsk_probes_out++; |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2511 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, |
Arnaldo Carvalho de Melo | 3f421ba | 2005-08-09 20:11:08 -0700 | [diff] [blame] | 2512 | min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), |
| 2513 | TCP_RTO_MAX); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2514 | } else { |
| 2515 | /* If packet was not sent due to local congestion, |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2516 | * do not backoff and do not remember icsk_probes_out. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2517 | * Let local senders to fight for local resources. |
| 2518 | * |
| 2519 | * Use accumulated backoff yet. |
| 2520 | */ |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2521 | if (!icsk->icsk_probes_out) |
| 2522 | icsk->icsk_probes_out = 1; |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2523 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2524 | min(icsk->icsk_rto << icsk->icsk_backoff, |
Arnaldo Carvalho de Melo | 3f421ba | 2005-08-09 20:11:08 -0700 | [diff] [blame] | 2525 | TCP_RESOURCE_PROBE_INTERVAL), |
| 2526 | TCP_RTO_MAX); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2527 | } |
| 2528 | } |
| 2529 | |
| 2530 | EXPORT_SYMBOL(tcp_connect); |
| 2531 | EXPORT_SYMBOL(tcp_make_synack); |
| 2532 | EXPORT_SYMBOL(tcp_simple_retransmit); |
| 2533 | EXPORT_SYMBOL(tcp_sync_mss); |
Stephen Hemminger | f4805ed | 2005-11-10 16:53:30 -0800 | [diff] [blame] | 2534 | EXPORT_SYMBOL(sysctl_tcp_tso_win_divisor); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 2535 | EXPORT_SYMBOL(tcp_mtup_init); |